diff --git a/.gitattributes b/.gitattributes index 68af34794a8a3ab640673c94d5c6ad83faa3c7de..4588eaaaa44f21659d0725ba4a3c3ace2e6bb47a 100644 --- a/.gitattributes +++ b/.gitattributes @@ -2087,3 +2087,67 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text 2025/Seeing[[:space:]]the[[:space:]]Abstract_[[:space:]]Translating[[:space:]]the[[:space:]]Abstract[[:space:]]Language[[:space:]]for[[:space:]]Vision[[:space:]]Language[[:space:]]Models/d6183534-ef29-4f52-b58a-e7e56ad32a4f_origin.pdf filter=lfs diff=lfs merge=lfs -text 2025/Seek[[:space:]]Common[[:space:]]Ground[[:space:]]While[[:space:]]Reserving[[:space:]]Differences_[[:space:]]Semi-Supervised[[:space:]]Image-Text[[:space:]]Sentiment[[:space:]]Recognition/129ec86b-7981-45f7-94dd-dd117a01e0d2_origin.pdf filter=lfs diff=lfs merge=lfs -text 2025/Seeking[[:space:]]Consistent[[:space:]]Flat[[:space:]]Minima[[:space:]]for[[:space:]]Better[[:space:]]Domain[[:space:]]Generalization[[:space:]]via[[:space:]]Refining[[:space:]]Loss[[:space:]]Landscapes/eabf1cb1-416d-4833-b397-be038a9ae636_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Simpler[[:space:]]Diffusion_[[:space:]]1.5[[:space:]]FID[[:space:]]on[[:space:]]ImageNet512[[:space:]]with[[:space:]]Pixel-space[[:space:]]Diffusion/3cf16242-5862-4aa6-864f-5aba8865f5f9_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Simplification[[:space:]]Is[[:space:]]All[[:space:]]You[[:space:]]Need[[:space:]]against[[:space:]]Out-of-Distribution[[:space:]]Overconfidence/c4dc8f46-aeb5-4477-a94f-497bd92530cf_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Simulator[[:space:]]HC_[[:space:]]Regression-based[[:space:]]Online[[:space:]]Simulation[[:space:]]of[[:space:]]Starting[[:space:]]Problem-Solution[[:space:]]Pairs[[:space:]]for[[:space:]]Homotopy[[:space:]]Continuation[[:space:]]in[[:space:]]Geometric[[:space:]]Vision/d90c2114-786f-423a-b72c-4f3391aa43d5_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/SinGS_[[:space:]]Animatable[[:space:]]Single-Image[[:space:]]Human[[:space:]]Gaussian[[:space:]]Splats[[:space:]]with[[:space:]]Kinematic[[:space:]]Priors/a1907643-3cec-4712-b625-aae92326d106_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Single[[:space:]]Domain[[:space:]]Generalization[[:space:]]for[[:space:]]Few-Shot[[:space:]]Counting[[:space:]]via[[:space:]]Universal[[:space:]]Representation[[:space:]]Matching/ef6b53d9-c810-43e9-a737-a386bc2ad915_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Six-CD_[[:space:]]Benchmarking[[:space:]]Concept[[:space:]]Removals[[:space:]]for[[:space:]]Text-to-image[[:space:]]Diffusion[[:space:]]Models/17b4b8e3-2cbd-49ce-875b-c3c7f736c51d_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Sketch[[:space:]]Down[[:space:]]the[[:space:]]FLOPs_[[:space:]]Towards[[:space:]]Efficient[[:space:]]Networks[[:space:]]for[[:space:]]Human[[:space:]]Sketch/d51d8c32-98f5-44d4-9674-ee9bc3d4af26_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/SketchAgent_[[:space:]]Language-Driven[[:space:]]Sequential[[:space:]]Sketch[[:space:]]Generation/76b8d485-0aca-43aa-b3b2-9c94133cc069_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/SketchFusion_[[:space:]]Learning[[:space:]]Universal[[:space:]]Sketch[[:space:]]Features[[:space:]]through[[:space:]]Fusing[[:space:]]Foundation[[:space:]]Models/1fc738f6-a4cf-4a97-bcd3-ea7682cab3c6_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/SketchVideo_[[:space:]]Sketch-based[[:space:]]Video[[:space:]]Generation[[:space:]]and[[:space:]]Editing/86ad126c-969d-47b0-8162-a7cc53739116_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Sketchtopia_[[:space:]]A[[:space:]]Dataset[[:space:]]and[[:space:]]Foundational[[:space:]]Agents[[:space:]]for[[:space:]]Benchmarking[[:space:]]Asynchronous[[:space:]]Multimodal[[:space:]]Communication[[:space:]]with[[:space:]]Iconic[[:space:]]Feedback/fff37445-0c8d-4c24-a3bf-32cf9fcbe8ea_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Sketchy[[:space:]]Bounding-box[[:space:]]Supervision[[:space:]]for[[:space:]]3D[[:space:]]Instance[[:space:]]Segmentation/fed9c100-3d53-4c29-809d-ba6f2288846b_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/SkillMimic_[[:space:]]Learning[[:space:]]Basketball[[:space:]]Interaction[[:space:]]Skills[[:space:]]from[[:space:]]Demonstrations/70ea78a5-d7df-4937-ac48-382497691b79_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Skip[[:space:]]Tuning_[[:space:]]Pre-trained[[:space:]]Vision-Language[[:space:]]Models[[:space:]]are[[:space:]]Effective[[:space:]]and[[:space:]]Efficient[[:space:]]Adapters[[:space:]]Themselves/e5453c9b-356b-43d9-9454-788a3f8b346d_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/SkySense-O_[[:space:]]Towards[[:space:]]Open-World[[:space:]]Remote[[:space:]]Sensing[[:space:]]Interpretation[[:space:]]with[[:space:]]Vision-Centric[[:space:]]Visual-Language[[:space:]]Modeling/2415dd78-f506-4cd3-a12c-42283424590a_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/SleeperMark_[[:space:]]Towards[[:space:]]Robust[[:space:]]Watermark[[:space:]]against[[:space:]]Fine-Tuning[[:space:]]Text-to-image[[:space:]]Diffusion[[:space:]]Models/2530599b-741a-485a-addd-c98f8e367f9b_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/SlideChat_[[:space:]]A[[:space:]]Large[[:space:]]Vision-Language[[:space:]]Assistant[[:space:]]for[[:space:]]Whole-Slide[[:space:]]Pathology[[:space:]]Image[[:space:]]Understanding/5542bef9-2edf-4ac5-b8df-cf3f63aeaa6c_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/SmartCLIP_[[:space:]]Modular[[:space:]]Vision-language[[:space:]]Alignment[[:space:]]with[[:space:]]Identification[[:space:]]Guarantees/28f50741-407e-4ef2-b9fa-f2bf47df9fd3_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/SmartEraser_[[:space:]]Remove[[:space:]]Anything[[:space:]]from[[:space:]]Images[[:space:]]using[[:space:]]Masked-Region[[:space:]]Guidance/ab3bbba3-1da3-4883-858e-f047afa69964_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/SnapGen-V_[[:space:]]Generating[[:space:]]a[[:space:]]Five-Second[[:space:]]Video[[:space:]]within[[:space:]]Five[[:space:]]Seconds[[:space:]]on[[:space:]]a[[:space:]]Mobile[[:space:]]Device/2089e6e3-1b02-4449-b418-985fa5c0c175_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/SnapGen_[[:space:]]Taming[[:space:]]High-Resolution[[:space:]]Text-to-Image[[:space:]]Models[[:space:]]for[[:space:]]Mobile[[:space:]]Devices[[:space:]]with[[:space:]]Efficient[[:space:]]Architectures[[:space:]]and[[:space:]]Training/d6336171-7440-4517-9fbd-fdd52a399434_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/SnowMaster_[[:space:]]Comprehensive[[:space:]]Real-world[[:space:]]Image[[:space:]]Desnowing[[:space:]]via[[:space:]]MLLM[[:space:]]with[[:space:]]Multi-Model[[:space:]]Feedback[[:space:]]Optimization/01611a60-cb24-4ca8-93fd-c47c38025543_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/SoMA_[[:space:]]Singular[[:space:]]Value[[:space:]]Decomposed[[:space:]]Minor[[:space:]]Components[[:space:]]Adaptation[[:space:]]for[[:space:]]Domain[[:space:]]Generalizable[[:space:]]Representation[[:space:]]Learning/282504c5-c667-4202-94ca-c92386507635_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/SocialGesture_[[:space:]]Delving[[:space:]]into[[:space:]]Multi-person[[:space:]]Gesture[[:space:]]Understanding/17e9a440-617f-4e5f-9712-f768185c9d5d_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/SocialMOIF_[[:space:]]Multi-Order[[:space:]]Intention[[:space:]]Fusion[[:space:]]for[[:space:]]Pedestrian[[:space:]]Trajectory[[:space:]]Prediction/74c8547f-5c13-42b2-b27c-77ef8f52058e_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Soft[[:space:]]Self-labeling[[:space:]]and[[:space:]]Potts[[:space:]]Relaxations[[:space:]]for[[:space:]]Weakly-supervised[[:space:]]Segmentation/4b6c17bc-bca2-47a9-b3a7-cd9f02074980_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/SoftShadow_[[:space:]]Leveraging[[:space:]]Soft[[:space:]]Masks[[:space:]]for[[:space:]]Penumbra-Aware[[:space:]]Shadow[[:space:]]Removal/1a57ac9e-bdab-4cc5-a82c-bea51caf54d0_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/SoftVQ-VAE_[[:space:]]Efficient[[:space:]]1-Dimensional[[:space:]]Continuous[[:space:]]Tokenizer/315fadef-93cf-49d1-961b-c20538f1f4de_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Solving[[:space:]]Instance[[:space:]]Detection[[:space:]]from[[:space:]]an[[:space:]]Open-World[[:space:]]Perspective/ab6d1961-27f9-4a61-af19-064b80505b12_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Sonata_[[:space:]]Self-Supervised[[:space:]]Learning[[:space:]]of[[:space:]]Reliable[[:space:]]Point[[:space:]]Representations/3e06b74b-3b4f-4789-9e14-04119c218bb3_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Sonic_[[:space:]]Shifting[[:space:]]Focus[[:space:]]to[[:space:]]Global[[:space:]]Audio[[:space:]]Perception[[:space:]]in[[:space:]]Portrait[[:space:]]Animation/108dd369-6bdf-4c82-9e76-1c8b1e3be7c1_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Sound[[:space:]]Bridge_[[:space:]]Associating[[:space:]]Egocentric[[:space:]]and[[:space:]]Exocentric[[:space:]]Videos[[:space:]]via[[:space:]]Audio[[:space:]]Cues/37c295df-9911-4189-b46a-8795fdcf6a6f_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/SoundVista_[[:space:]]Novel-View[[:space:]]Ambient[[:space:]]Sound[[:space:]]Synthesis[[:space:]]via[[:space:]]Visual-Acoustic[[:space:]]Binding/9c309271-ab78-4453-a939-59e8aa02fe41_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Sparse[[:space:]]Point[[:space:]]Cloud[[:space:]]Patches[[:space:]]Rendering[[:space:]]via[[:space:]]Splitting[[:space:]]2D[[:space:]]Gaussians/dd278d9d-da45-4b1e-bcc2-33b0924a1fa1_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Sparse[[:space:]]Voxels[[:space:]]Rasterization_[[:space:]]Real-time[[:space:]]High-fidelity[[:space:]]Radiance[[:space:]]Field[[:space:]]Rendering/a6bd5a3d-4547-4194-a227-b9c1f40e864c_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Sparse2DGS_[[:space:]]Geometry-Prioritized[[:space:]]Gaussian[[:space:]]Splatting[[:space:]]for[[:space:]]Surface[[:space:]]Reconstruction[[:space:]]from[[:space:]]Sparse[[:space:]]Views/3c5dfe86-080c-4ec1-9ca9-65bf2b3453a3_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/SparseAlign_[[:space:]]a[[:space:]]Fully[[:space:]]Sparse[[:space:]]Framework[[:space:]]for[[:space:]]Cooperative[[:space:]]Object[[:space:]]Detection/e9f3f305-e62d-4aad-9255-49e76e43c9cf_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Spatial[[:space:]]Transport[[:space:]]Optimization[[:space:]]by[[:space:]]Repositioning[[:space:]]Attention[[:space:]]Map[[:space:]]for[[:space:]]Training-Free[[:space:]]Text-to-Image[[:space:]]Synthesis/68174923-d215-4833-8530-d036bb11b7fd_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Spatial-Temporal[[:space:]]Graph[[:space:]]Diffusion[[:space:]]Policy[[:space:]]with[[:space:]]Kinematic[[:space:]]Modeling[[:space:]]for[[:space:]]Bimanual[[:space:]]Robotic[[:space:]]Manipulation/8300c221-08df-4f6a-8aec-775be3434ff5_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Spatial457_[[:space:]]A[[:space:]]Diagnostic[[:space:]]Benchmark[[:space:]]for[[:space:]]6D[[:space:]]Spatial[[:space:]]Reasoning[[:space:]]of[[:space:]]Large[[:space:]]Mutimodal[[:space:]]Models/96f4562f-945b-497d-9468-cd476fb5f769_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/SpatialCLIP_[[:space:]]Learning[[:space:]]3D-aware[[:space:]]Image[[:space:]]Representations[[:space:]]from[[:space:]]Spatially[[:space:]]Discriminative[[:space:]]Language/dad5650f-d216-4b30-a0c7-dd7ca6658593_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/SpatialDreamer_[[:space:]]Self-supervised[[:space:]]Stereo[[:space:]]Video[[:space:]]Synthesis[[:space:]]from[[:space:]]Monocular[[:space:]]Input/24175650-223e-41c7-b804-6cc94c2c5191_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/SpatialLLM_[[:space:]]A[[:space:]]Compound[[:space:]]3D-Informed[[:space:]]Design[[:space:]]towards[[:space:]]Spatially-Intelligent[[:space:]]Large[[:space:]]Multimodal[[:space:]]Models/cf47bc9d-8b67-4bc0-a3fe-ba2e3465761d_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Spatiotemporal[[:space:]]Decoupling[[:space:]]for[[:space:]]Efficient[[:space:]]Vision-Based[[:space:]]Occupancy[[:space:]]Forecasting/9ae564d7-647a-4ca8-a4dc-af9f575b1e21_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Spatiotemporal[[:space:]]Skip[[:space:]]Guidance[[:space:]]for[[:space:]]Enhanced[[:space:]]Video[[:space:]]Diffusion[[:space:]]Sampling/4ec7da94-3031-45c8-a8d6-19bf226e3cba_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/SpecTRe-GS_[[:space:]]Modeling[[:space:]]Highly[[:space:]]Specular[[:space:]]Surfaces[[:space:]]with[[:space:]]Reflected[[:space:]]Nearby[[:space:]]Objects[[:space:]]by[[:space:]]Tracing[[:space:]]Rays[[:space:]]in[[:space:]]3D[[:space:]]Gaussian[[:space:]]Splatting/05871852-d282-4279-bc18-a4b518c95c0a_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Spectral[[:space:]]Informed[[:space:]]Mamba[[:space:]]for[[:space:]]Robust[[:space:]]Point[[:space:]]Cloud[[:space:]]Processing/9acc5c2f-4c09-4911-b4a0-485ec3edcc53_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Spectral[[:space:]]State[[:space:]]Space[[:space:]]Model[[:space:]]for[[:space:]]Rotation-Invariant[[:space:]]Visual[[:space:]]Representation[[:space:]]Learning/da806762-c7b4-4ed7-80ed-cc0683ab510c_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/SpectroMotion_[[:space:]]Dynamic[[:space:]]3D[[:space:]]Reconstruction[[:space:]]of[[:space:]]Specular[[:space:]]Scenes/5055bce6-6fec-402b-adcc-413e1a1becdd_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Speedy-Splat_[[:space:]]Fast[[:space:]]3D[[:space:]]Gaussian[[:space:]]Splatting[[:space:]]with[[:space:]]Sparse[[:space:]]Pixels[[:space:]]and[[:space:]]Sparse[[:space:]]Primitives/08166209-021b-4f8f-8611-9f7790e8bc2b_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/SphereUFormer_[[:space:]]A[[:space:]]U-Shaped[[:space:]]Transformer[[:space:]]for[[:space:]]Spherical[[:space:]]360[[:space:]]Perception/d3dc320a-7961-421e-ae82-56124645a3a0_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Spherical[[:space:]]Manifold[[:space:]]Guided[[:space:]]Diffusion[[:space:]]Model[[:space:]]for[[:space:]]Panoramic[[:space:]]Image[[:space:]]Generation/25742454-f26d-401b-a2ab-0a30d46da30a_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Spiking[[:space:]]Transformer[[:space:]]with[[:space:]]Spatial-Temporal[[:space:]]Attention/de8548d3-fd04-4e18-9d13-559b7231ddb6_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Spiking[[:space:]]Transformer_[[:space:]]Introducing[[:space:]]Accurate[[:space:]]Addition-Only[[:space:]]Spiking[[:space:]]Self-Attention[[:space:]]for[[:space:]]Transformer/24b0ab2b-a4ff-49fc-865c-b8d7799e49bd_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/SpiritSight[[:space:]]Agent_[[:space:]]Advanced[[:space:]]GUI[[:space:]]Agent[[:space:]]with[[:space:]]One[[:space:]]Look/c38ba9da-d4d0-4e2a-a186-fdd0cc0687b1_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Spk2SRImgNet_[[:space:]]Super-Resolve[[:space:]]Dynamic[[:space:]]Scene[[:space:]]from[[:space:]]Spike[[:space:]]Stream[[:space:]]via[[:space:]]Motion[[:space:]]Aligned[[:space:]]Collaborative[[:space:]]Filtering/e0bcc965-cf40-409b-a203-db12626ad68b_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/SplatAD_[[:space:]]Real-Time[[:space:]]Lidar[[:space:]]and[[:space:]]Camera[[:space:]]Rendering[[:space:]]with[[:space:]]3D[[:space:]]Gaussian[[:space:]]Splatting[[:space:]]for[[:space:]]Autonomous[[:space:]]Driving/58aed1a8-44d5-4314-9f37-1456921ac406_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/SplatFlow_[[:space:]]Multi-View[[:space:]]Rectified[[:space:]]Flow[[:space:]]Model[[:space:]]for[[:space:]]3D[[:space:]]Gaussian[[:space:]]Splatting[[:space:]]Synthesis/a1da8541-9ef2-4901-8219-e22e7fdfc1fc_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/SplatFlow_[[:space:]]Self-Supervised[[:space:]]Dynamic[[:space:]]Gaussian[[:space:]]Splatting[[:space:]]in[[:space:]]Neural[[:space:]]Motion[[:space:]]Flow[[:space:]]Field[[:space:]]for[[:space:]]Autonomous[[:space:]]Driving/8f50ebde-225c-4574-a537-44bd7c2cfecd_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Splatter-360_[[:space:]]Generalizable[[:space:]]360[[:space:]]Gaussian[[:space:]]Splatting[[:space:]]for[[:space:]]Wide-baseline[[:space:]]Panoramic[[:space:]]Images/695c9e66-9d36-4a85-a2d0-e99f9431d57c_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/SplineGS_[[:space:]]Robust[[:space:]]Motion-Adaptive[[:space:]]Spline[[:space:]]for[[:space:]]Real-Time[[:space:]]Dynamic[[:space:]]3D[[:space:]]Gaussians[[:space:]]from[[:space:]]Monocular[[:space:]]Video/c20079e6-f30b-47d1-a98c-7f27a87d2799_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Split[[:space:]]Adaptation[[:space:]]for[[:space:]]Pre-trained[[:space:]]Vision[[:space:]]Transformers/b5d6478d-8866-4dba-8ddb-9d84bd75792a_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Spotting[[:space:]]the[[:space:]]Unexpected[[:space:]](STU)_[[:space:]]A[[:space:]]3D[[:space:]]LiDAR[[:space:]]Dataset[[:space:]]for[[:space:]]Anomaly[[:space:]]Segmentation[[:space:]]in[[:space:]]Autonomous[[:space:]]Driving/7019444b-0523-423c-8391-a6c740f37e7c_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Stabilizing[[:space:]]and[[:space:]]Accelerating[[:space:]]Autofocus[[:space:]]with[[:space:]]Expert[[:space:]]Trajectory[[:space:]]Regularized[[:space:]]Deep[[:space:]]Reinforcement[[:space:]]Learning/a699e1d4-e683-4371-bc56-144538826827_origin.pdf filter=lfs diff=lfs merge=lfs -text diff --git a/2025/Simpler Diffusion_ 1.5 FID on ImageNet512 with Pixel-space Diffusion/3cf16242-5862-4aa6-864f-5aba8865f5f9_content_list.json b/2025/Simpler Diffusion_ 1.5 FID on ImageNet512 with Pixel-space Diffusion/3cf16242-5862-4aa6-864f-5aba8865f5f9_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..421e162222a6e793e2526c6dc9b8fca71bb1ac11 --- /dev/null +++ b/2025/Simpler Diffusion_ 1.5 FID on ImageNet512 with Pixel-space Diffusion/3cf16242-5862-4aa6-864f-5aba8865f5f9_content_list.json @@ -0,0 +1,1596 @@ +[ + { + "type": "text", + "text": "Simpler Diffusion: 1.5 FID on ImageNet512 with pixel-space diffusion", + "text_level": 1, + "bbox": [ + 145, + 130, + 851, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Emiel Hoogeboom", + "bbox": [ + 192, + 181, + 346, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Google Deepmind", + "bbox": [ + 192, + 199, + 344, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "emielh@google.com", + "bbox": [ + 183, + 217, + 361, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Thomas Mensink", + "bbox": [ + 431, + 181, + 573, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Google Deepmind", + "bbox": [ + 431, + 199, + 576, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "mensink@google.com", + "bbox": [ + 408, + 217, + 598, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jonathan Heek", + "bbox": [ + 669, + 181, + 790, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Google Deepmind", + "bbox": [ + 656, + 199, + 805, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "jheek@google.com", + "bbox": [ + 648, + 217, + 815, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Kay Lamarigts", + "bbox": [ + 202, + 260, + 325, + 277 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Google Research", + "bbox": [ + 196, + 277, + 334, + 294 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "kaylamerigts@google.com", + "bbox": [ + 143, + 295, + 385, + 311 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ruiqi Gao", + "bbox": [ + 473, + 260, + 558, + 277 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Google Deepmind", + "bbox": [ + 442, + 277, + 589, + 294 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ruiqig@google.com", + "bbox": [ + 428, + 295, + 606, + 311 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Tim Salimans", + "bbox": [ + 692, + 260, + 805, + 276 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Google Deepmind", + "bbox": [ + 674, + 277, + 823, + 294 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "salimans@google.com", + "bbox": [ + 648, + 295, + 846, + 311 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 364, + 326, + 380 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Latent diffusion models have become the popular choice for scaling up diffusion models for high resolution image synthesis. Compared to pixel-space models that are trained end-to-end, latent models are perceived to be more efficient and to produce higher image quality at high resolution. Here we challenge these notions, and show that pixel-space models can be very competitive to latent models both in quality and efficiency, achieving 1.5 FID on ImageNet512 and new SOTA results on ImageNet128, ImageNet256 and Kinetics600.", + "bbox": [ + 89, + 395, + 483, + 530 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We present a simple recipe for scaling end-to-end pixel-space diffusion models to high resolutions. 1: Use the sigmoid loss-weighting [25] with our prescribed hyperparameters. 2: Use our simplified memory-efficient architecture with fewer skip-connections. 3: Scale the model to favor processing the image at a high resolution with fewer parameters, rather than using more parameters at a lower resolution. Combining these with guidance intervals, we obtain a family of pixel-space diffusion models we call Simpler Diffusion (SiD2).", + "bbox": [ + 88, + 532, + 483, + 683 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 709, + 220, + 724 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Diffusion models have been tremendously successful in domains such as image, video and audio generation. For high resolution data, latent diffusion models [32] have been particularly successful: Through encoding the data in a lower dimensional space, these models sidestep the difficulties faced by more traditional pixel-space diffusion models in scaling up to high resolution. This makes latent diffusion models also scale better in training compute than other approaches, as shown by [20].", + "bbox": [ + 89, + 734, + 483, + 869 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "However, latent models have their own challenges: It can be difficult to train the autoencoders that produce the latents,", + "bbox": [ + 89, + 869, + 483, + 900 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8b5a7ae5161b82f15e43113268a9298406651cafd4317c76bb573e36806124bf.jpg", + "image_caption": [ + "Figure 1. End-to-end pixel diffusion performance on ImageNet512." + ], + "image_footnote": [], + "bbox": [ + 542, + 364, + 882, + 467 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "which in practice requires careful tuning of regularization hyperparameters [32]. Even when trained well, the decoder of a latent diffusion model can produce artifacts or otherwise imperfect reconstructions, especially when it is applied to latents that are out of distribution compared to the training set. A single-stage pixel-space diffusion model does not have these issues, and so we ask the question: Can pixel-space diffusion models scale as well as latent models with the right architecture and hyperparameters?", + "bbox": [ + 509, + 505, + 905, + 641 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Over the years, end-to-end pixel-space diffusion has improved considerably (Figure 1). The main problem was that the standard schedule and loss-weighting [11] work well for small resolution images, like $32^2$ or $64^2$ , but not at higher resolution [3, 14]. It was further shown [25] that it can be helpful to disentangle loss weighting and time sampling for improved control over both, leading to better performance at higher resolution. However, in previous works pixels-space models still did not match the best latent models on standard benchmarks like ImageNet512.", + "bbox": [ + 509, + 642, + 906, + 792 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The main contribution of this paper is to improve pixel-space diffusion on high resolution images while keeping as close to the original formulation of diffusion models as possible. Specifically, we analyze and tune the sigmoid loss weighting from [25], and show why it should be preferred over their favored EDM monotonic. We further propose architectural improvements that lead to even better perfor", + "bbox": [ + 509, + 795, + 908, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "18062", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b18a0f4d018046d9ab7155f067e3df4b7341ad15d635e2a00a6cbc6332bde9c2.jpg", + "image_caption": [ + "Figure 2. Some text-to-image samples of $512 \\times 512$ that have been generated by SiD2." + ], + "image_footnote": [], + "bbox": [ + 91, + 87, + 290, + 241 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/4673782a530a2ffca3a5a3a1f6c3ddfa3d65cdd0850dc8041f8a3798bc066625.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 295, + 87, + 495, + 241 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/9b3e67cdffaed8395cf0f73b870516817a0a09d2e7eec74fb72d26e6b38056ce.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 88, + 700, + 241 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/2e50436e0cea30003bd97c6bc97008729d7dedcd626988278a1604a13b1e4d6b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 707, + 88, + 903, + 241 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "mance, namely flop heavy scaling and removing block skip-connections (Residual U-ViTs). Most notably, we improve the state-of-the-art of pixel-space diffusion from 2.65 to 1.48 on ImageNet512 (see also Figure 1). Further, we also achieve overall state-of-the-art on ImageNet128, ImageNet256 and Kinetics600 without heavy tuning.", + "bbox": [ + 89, + 292, + 483, + 383 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1.1. Background", + "text_level": 1, + "bbox": [ + 89, + 393, + 220, + 409 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "A diffusion process gradually destroys data over a conceptual time $t$ , typically with Gaussian noise as follows:", + "bbox": [ + 89, + 415, + 483, + 445 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {z} _ {t} = \\alpha_ {t} \\boldsymbol {x} + \\sigma_ {t} \\boldsymbol {\\epsilon}, \\text {w h e r e} \\boldsymbol {\\epsilon} \\sim \\mathcal {N} (0, \\mathbf {I}), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 158, + 458, + 480, + 474 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "which defines a distribution $q(\\pmb{z}_t|\\pmb{x})$ such that $\\pmb{z}_0$ is close to clean data and $\\pmb{z}_1$ is close to Gaussian noise. In this paper we assume a variance-preserving [41] schedule, which is defined as $\\alpha_{t} = \\sqrt{\\sigma(\\lambda_{t})}$ and $\\sigma_t = \\sqrt{\\sigma(-\\lambda_t)}$ , with $\\lambda_{t} = \\log \\left(\\alpha_{t}^{2} / \\sigma_{t}^{2}\\right)$ being the log signal-to-noise ratio, which monotonically decreases with time as $t\\to 1$ , and $\\sigma (\\cdot)$ being the sigmoid function. It can be shown that other types of schedules such as the variance-exploding schedule or the flow matching schedule [28] can always be converted to a variance-preserving schedule with a linear scaling of $\\pmb{z}_{t}$ [25]. The key defining property of a noise schedule is the log signal-to-noise ratio $\\lambda_{t}$ . A neural network is used to predict $\\hat{\\pmb{x}} = \\hat{\\pmb{x}}_{\\theta}(\\pmb{z}_{t},t)$ that effectively estimates $\\mathbb{E}[\\pmb {x}|\\pmb {z}_t]$ , which can be done by minimizing a weighted mean squared error loss:", + "bbox": [ + 89, + 486, + 483, + 698 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\nL (\\boldsymbol {x}) = \\mathbb {E} _ {t \\sim \\mathcal {U} (0, 1)} - \\frac {\\mathrm {d} \\lambda_ {t}}{\\mathrm {d} t} w \\left(\\lambda_ {t}\\right) \\| \\boldsymbol {x} - \\hat {\\boldsymbol {x}} \\| ^ {2}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 709, + 480, + 738 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "where $w(\\lambda_t)$ is the weighting function. $-\\mathrm{d}\\lambda_{t} / \\mathrm{d}t$ cancels out the effect of the noise schedule $\\lambda_{t}$ on the loss. In other words, $L(\\mathbf{x})$ can be rewritten as $\\int -w(\\lambda)\\| \\pmb {x} - \\hat{\\pmb{x}}\\| ^2 d\\lambda$ that is invariant to the choice of $\\lambda_{t}$ . However, $\\lambda_{t}$ still impacts the sampling efficiency of the Monte Carlo estimator of the expected loss, and thus affects the training dynamics of diffusion models. This decouples noise schedule $\\lambda_{t}$ and weighting function $w(\\lambda)$ as a function of $\\lambda$ [25].", + "bbox": [ + 89, + 750, + 482, + 869 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The above squared error loss can be derived from the variational interpretation of diffusion models [26]. Given", + "bbox": [ + 89, + 869, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "two time steps $s < t$ , we define the denoising distribution $p(\\boldsymbol{z}_s|\\boldsymbol{z}_t)$ in the same formulation as the true posterior $q(\\boldsymbol{z}_s|\\boldsymbol{z}_t,\\boldsymbol{x})$ , but replace the ground truth $\\boldsymbol{x}$ with the learned approximation $\\hat{\\boldsymbol{x}} \\approx \\mathbb{E}[\\boldsymbol{x}|\\boldsymbol{z}_t]$ :", + "bbox": [ + 511, + 292, + 906, + 354 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\np \\left(\\boldsymbol {z} _ {s} \\mid \\boldsymbol {z} _ {t}\\right) = q \\left(\\boldsymbol {z} _ {s} \\mid \\boldsymbol {z} _ {t}, \\boldsymbol {x} = \\hat {\\boldsymbol {x}}\\right) = \\mathcal {N} \\left(\\boldsymbol {z} _ {s} \\mid \\mu_ {t \\rightarrow s}, \\sigma_ {t \\rightarrow s}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 549, + 364, + 903, + 381 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "where $\\sigma_{t\\rightarrow s}^2 = \\left(1 / \\sigma_s^2 +\\alpha_{t|s}^2 /\\sigma_{t|s}^2\\right)^{-1}$ and $\\mu_{t\\to s} = \\sigma_{t\\to s}^{2}\\Bigl (\\alpha_{t|s} / \\sigma_{t|s}^{2}\\pmb {z}_{t} + \\alpha_{s} / \\sigma_{s}^{2}\\pmb {x}\\Bigr)$ with $\\alpha_{t|s} = \\alpha_t / \\alpha_s$ and $\\sigma_{t|s}^2 = \\sigma_t^2 -\\alpha_{t|s}\\sigma_s^2$ . Note that during sampling, the standard deviation for $p(z_{s}|z_{t})$ is chosen to be a log-linear interpolation between $\\sigma_{t|s}$ and $\\sigma_{t\\rightarrow s}$ of the form $\\sigma_{t\\rightarrow s}^{\\gamma}\\sigma_{t|s}^{1 - \\gamma}$ where $\\gamma = 0.3$ was set empirically, which is important as FID is sensitive to over-smooth or grainy images.", + "bbox": [ + 511, + 393, + 905, + 517 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Simpler Diffusion", + "text_level": 1, + "bbox": [ + 511, + 532, + 687, + 549 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Revisiting the sigmoid loss", + "text_level": 1, + "bbox": [ + 511, + 558, + 751, + 574 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recall that one of the proposed losses in VDM++ [25] is the sigmoid loss. They expressed the loss with respect to $\\epsilon$ -mse: $L(\\pmb{x}) = \\mathbb{E}_{t \\sim \\mathcal{U}(0,1)} - \\frac{\\mathrm{d}\\lambda_t}{\\mathrm{d}t} \\tilde{w}(\\lambda_t) \\| \\pmb{\\epsilon} - \\hat{\\pmb{\\epsilon}} \\|^2$ , which is connected to $\\pmb{x}$ -mse through $\\| \\pmb{\\epsilon} - \\hat{\\pmb{\\epsilon}} \\|^2 = \\exp(\\lambda_t) \\| \\pmb{x} - \\hat{\\pmb{x}} \\|^2$ and $\\tilde{w}(\\lambda_t) = \\exp(\\lambda_t) w(\\lambda_t)$ . In VDM++, viewing the loss in $\\epsilon$ -mse enables a natural connection between the weighted loss and the ELBO and showed that losses should be monotonically decreasing on $\\epsilon$ -mse. Hence, they propose the sigmoid loss: $\\tilde{w}(\\lambda_t) = \\sigma(b - \\lambda_t)$ , where $b$ denotes the bias or shift and $\\sigma$ is the sigmoid function. However, they ultimately opted for another loss-weighting named EDM-monotonic for high resolution image generation, see Figure 3a.", + "bbox": [ + 511, + 580, + 906, + 763 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Enforcing weighting monotonicity in $x$ Revisiting the sigmoid loss, we derive its $x$ -mse form, which turns out to conveniently be defined by a mirrored sigmoid (albeit multiplied by a constant factor):", + "bbox": [ + 511, + 785, + 906, + 845 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\sigma \\left(b - \\lambda_ {t}\\right) \\| \\boldsymbol {\\epsilon} - \\hat {\\boldsymbol {\\epsilon}} \\| ^ {2} = \\sigma \\left(b - \\lambda_ {t}\\right) \\exp \\left(\\lambda_ {t}\\right) \\| \\boldsymbol {x} - \\hat {\\boldsymbol {x}} \\| ^ {2} \\tag {4} \\\\ = \\exp (b) \\sigma \\left(\\lambda_ {t} - b\\right) \\| \\boldsymbol {x} - \\hat {\\boldsymbol {x}} \\| ^ {2}. \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 531, + 858, + 903, + 896 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "18063", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/12bc20bf77af3c4d4311affe192124ed53e769feeff62a5c95055d6236936202.jpg", + "image_caption": [ + "(a) Weighting over $\\epsilon$ -mse" + ], + "image_footnote": [], + "bbox": [ + 145, + 104, + 439, + 246 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/2c131284de18ff2945917ee8d0833c6474829f5f2450df0baff16c2b498b7be2.jpg", + "image_caption": [ + "(b) Weighting over $\\pmb{x}$ -mse", + "Figure 3. Weighting of different losses in both $\\epsilon$ and $x$ space. Note that the weightings are scaled to have a maximum of one." + ], + "image_footnote": [], + "bbox": [ + 545, + 104, + 836, + 246 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This reveals that the sigmoid loss is not only monotonically decreasing in $\\epsilon$ -mse, it is also monotonically increasing in $x$ -mse. We claim that a reasonable weighting function $w(\\lambda_t)$ should be monotonically increasing over $\\lambda_t$ . Observe that with less noise in $z_t$ , there is more information available and it should be strictly easier to predict the original data, meaning that the $x$ -mse generally decreases for higher $\\lambda_t$ . This phenomenon is also observed in practice. If the loss weighting in $x$ -mse also decreases, the model would be able to make worse predictions while obtaining less noisy (better) conditioning information and still achieve a lower loss. Simply put: Why put less weight on an easier problem?", + "bbox": [ + 88, + 291, + 482, + 472 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Studying both the form of the sigmoid and EDM-monotonic weighting functions in Figure 3, we observe that both weighting functions are monotonically decreasing in $\\epsilon$ -mse, but only the sigmoid weighting is monotonically increasing in $x$ -mse. Therefore, we expect the sigmoid weighting to be a better candidate to scale to high resolution generation than the EDM-monotonic version.", + "bbox": [ + 89, + 472, + 483, + 580 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Loss shifting Recall that for a resolution that is two times higher in both height and width, an average (low pass) filter would see $\\mathrm{Var}[(x_{11} + x_{12} + x_{21} + x_{22}) / 4] = 1 / 4$ of the noise, while signal strength stays the same [14, 25]. Hence, its effective logsnr is expected to increase by about $2\\log 2\\approx 1.386\\approx 1.5$ when viewed through a low-pass filter. Suppose that the optimal setting for images of $128\\times 128$ is a sigmoid weighting with $b\\approx 0$ , then one would expect the optimal setting for images of $512\\times 512$ to be roughly $b = -3$ .1 Empirically we find that slightly more aggressive shifts perform even better.", + "bbox": [ + 89, + 595, + 483, + 763 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Flop heavy scaling", + "text_level": 1, + "bbox": [ + 89, + 771, + 269, + 787 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "When one deals with a limited data budget, it can be difficult to balance scaling of models and regularization. Our small model version at $512 \\times 512$ resolution starts out with $4 \\times 4$ patching, and subsequently runs a U-ViT [14] on the $128 \\times$", + "bbox": [ + 89, + 792, + 482, + 854 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "$128 \\times 48$ input (assuming the raw input is of 3 channels). To make the model more expressive, one of the most common options is to double the number of channels for every layer, leading to $2 \\times$ model size. However, our experiments show that this results in diminishing returns. We hypothesize that it is because the model suffers from over-fitting when trained with limited data budget such as ImageNet. Indeed, when we train using distribution augmentation [18], the best performance improves from 1.9 to 1.6. This however suggests another way to scale the model: Keep the architecture and parameter count almost the same and instead run on $2 \\times 2$ patched input (so $256 \\times 256 \\times 12$ for 512 inputs). Similar to $2 \\times$ model size, reducing the patching size also increases FLOPS by roughly a factor of 4. However, we found this model performs even better to about 1.5 without the need of distribution augmentation. We call this model flop heavy, because it has a much higher flop over parameter ratio.", + "bbox": [ + 511, + 291, + 906, + 547 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Flop heavy scaling has another interesting property. Suppose we have a small model trained to fit input of size $256 \\times 256$ with $2 \\times 2$ patching. Using the flop heavy scaling at $512 \\times 512$ resolution enables us to directly finetune from the small model trained at $256 \\times 256$ with exactly the same amount of parameters, whereas its counterpart version at $512 \\times 512$ resolution and with $4 \\times 4$ patching needs to introduce additional parameters to handle the more input channels. This method of finetuning to higher resolutions has already been popularized in the context of text-to-image generation [32]. Here we find that flop heavy scaling is a more favorable way to scale on smaller datasets such as ImageNet. Of course, in principle one could also train a model with more channels and $4 \\times 4$ patching on a lower resolution, but this is not done in practice to train 'small' models.", + "bbox": [ + 511, + 549, + 908, + 776 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Residual U-ViTs: Removing blockwise skipconnections", + "text_level": 1, + "bbox": [ + 511, + 786, + 906, + 816 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "U-Nets require that sub-sampled levels (for example via average pooling) are bypassed by skip-connections. However, this makes the architecture very different from large-scale language models such as transformers, where these skip-connections are not used and residual connections suffice.", + "bbox": [ + 511, + 824, + 908, + 898 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "This holds under the condition that the most important information is already contained by the $128 \\times 128$ image, and this may for example not hold for images without global structure, e.g. that only contain textures.", + "bbox": [ + 89, + 862, + 482, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "18064", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this paper we use the U-ViT model of [14], which is a variant of U-Nets with self-attention layers (with convolutional layers) being replaced by transformer blocks. Hence, that part of the architecture is closer to standard transformers.", + "bbox": [ + 89, + 90, + 483, + 151 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this paper, we further simplify model definition with Residual U-ViTs, see Figure 4. The blockwise skip-connections are removed, and in their place only a single skip-connection is used per downsampling operation. To be precise, let $f_{d}, f_{u}$ denote a down and up level of a U-ViT and let $f_{m}$ denote the middle stage. This two-stage Residual U-ViT can now be described by:", + "bbox": [ + 89, + 152, + 483, + 258 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nf (x) = f _ {u} \\left(\\mathrm {U} \\left(f _ {m} (\\mathrm {D} (h)) - \\mathrm {D} (h)\\right) + h\\right) \\text {w h e r e} h = f _ {d} (x) \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 271, + 483, + 309 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $D$ is typically an average pooling and linear and $U$ is typically a linear and nearest neighbour upsampling. Defining this network to preserve an identity function at initialization is now much easier: As long as $f_{d}$ , $f_{u}$ and $f_{m}$ are initialized to be identity, the entire $f$ is also identity at initialization. The generalization to any-depth U-ViTs is also straightforward: we can define another two levels of U-ViTs in place of $f_{m}$ recursively, where the nested U-ViTs can also be guaranteed to be initialized as an identity function.", + "bbox": [ + 89, + 309, + 483, + 445 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For small networks, it could be helpful to keep the blockwise skip-connections, and expect to incur a slight performance degradation from removing the blockwise skip-connections. Small networks have limited capacity, so limiting skip-connections may create bottleneck issues. Experimentally, this is confirmed for the small architectures, where we observe 0.2 FID performance degradation on ImageNet512 and 0.06 FID performance degradation on ImageNet256. Because of the simplified definition, we choose to accept this performance penalty for small models. Moreover, we find that larger models perform on par or even benefit from removing the blockwise skip-connections.", + "bbox": [ + 89, + 446, + 483, + 628 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The final advantage is memory consumption at evaluation. For example, a feature map at $256^{2}$ resolution with 128 channels requires $256^{2} \\cdot 128 \\cdot 2 / 1024^{2} = 16\\mathrm{MB}$ . With 3 blocks per level, this requires holding an extra $3 \\cdot 16 = 48\\mathrm{MB}$ per example for the skip-connections, whereas the Residual U-ViT only requires holding $h$ and $D(h)$ (the latter could be cheaply recomputed), so $16 + 8 = 24\\mathrm{MB}$ per example while computing the lower levels. For a hypothetical super-resolution model at $4096^{2}$ resolution, this leads to a difference between $3 \\cdot 4 = 12\\mathrm{GB}$ and only $4 + 2 = 6\\mathrm{GB}$ . During training this advantage is usually negligible, because feature maps also need to be retained for the backward pass. A final advantage of Residual U-ViTs is that it is more straightforward to define asymmetric U-ViTs, i.e., with a different number of blocks in the downsampling (encoding) stage from the one in the upsampling (decoding) stage. In certain cases, moving blocks from the downsampling stage to the upsampling stage can increase performance.", + "bbox": [ + 89, + 628, + 485, + 900 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/db47be7ec0de83b6399d13043bf84e5cd0faa4c5cc3431cd3fa9f8131560132e.jpg", + "image_caption": [ + "Figure 4. Residual U-ViT with levelwise skip-connections (instead of blockwise). When the network blocks are initialized to be identity, the entire network is also an identity function at initialization." + ], + "image_footnote": [], + "bbox": [ + 517, + 88, + 905, + 191 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.4. Distillation", + "text_level": 1, + "bbox": [ + 513, + 263, + 635, + 277 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Sampling from a pixel-space diffusion model naively often requires more function evaluations than sampling from a latent-space diffusion model. For example, the popular Stable Diffusion model [32] is usually sampled using 20-50 sampling steps, while pixel-space models are sampled with $100+$ steps [14]. To reduce the number of sampling steps, we could use either more advanced solvers or distillation. Recently distillation methods show that pixel-space models can be adapted into few-step samplers that maintain high sample quality [23, 29, 36, 44, 45, 47]. In this work, we use moment matching distillation [37] which is currently one of the best distillation methods available in the few step regime. We distill our pixel-space models for use with 8-16 sampling steps, which empirically we find maintains or even improves the sample quality of our models.", + "bbox": [ + 511, + 284, + 906, + 512 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3. Related Work", + "text_level": 1, + "bbox": [ + 513, + 523, + 653, + 540 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The first end-to-end diffusion work that directly generated high resolution images was [4]. Given the poor performance of end-to-end models, two popular alternatives were introduced: cascaded diffusion [12] and latent diffusion [32].", + "bbox": [ + 511, + 549, + 906, + 609 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": " Cascaded diffusion approaches have been shown to scale to large datasets and models to generate high quality and high resolution imagery, albeit in multiple stages [12, 40, 42, 52], or with fine-tuning [43]. Although the focus of this paper has been single stage generation up to a resolution of $10^{24^2}$ , the loss weightings may also be used for cascaded models.", + "bbox": [ + 511, + 609, + 906, + 700 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Significant improvements were made to end-to-end diffusion by shifting noise schedules [3, 14] and subsequently loss weightings [25]. Other approaches such as multi-resolution losses (and possibly architecture adaptions) [6, 7, 14] added marginal gains. However, the current best known latent model EDM2 [20] outperforms these prior end-to-end approaches by a large margin. In contrast, our models are competitive with latent models, outperforming DiT-XL/2 [31] and only performing slightly worse than EDM2.", + "bbox": [ + 511, + 700, + 908, + 837 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Skip connections The U-Net architecture is a fully convolutional network with long-range skip-connections introduced for (medical) image segmentation [33, 39]. The main", + "bbox": [ + 511, + 854, + 908, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "18065", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/3e8b1ffcef67cfe66bf5c5c04d334ca80c64b63852d2acf792d78168f86a8358.jpg", + "image_caption": [ + "Figure 5. Effect of weighting functions in the loss on the FID score of ImageNet512 over the training iterations. The proposed sigmoid weighting with $b = -3$ shows best performance consistent across all training iterations." + ], + "image_footnote": [], + "bbox": [ + 91, + 85, + 491, + 272 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "idea of U-Net is to combine a contracting path (which encodes the image semantics) with long skip-connections at the same resolution to enable precise localization.", + "bbox": [ + 89, + 349, + 485, + 393 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "With the introduction of ViTs [5] as an alternative for convolutional layers, the U-Net has been converted to U-ViTs using a similar network but replacing (some) ResNet blocks with ViTs, either with spatial down-sampling [14] or without [2]. Long-range skip-connections play an pivotal role in these networks, for example for fine-tuning controllable diffusion models [17] and for the stability and training speed [15, 30]. In these latter works, it is shown that multiplying the skip-connections with a (constant) scalar $(< 1)$ improves training stability and speed, because of the propagation of noise throughout the network. In our work, we reduce the number of skip-connections significantly and initialise the network such that it resembles an identify function.", + "bbox": [ + 89, + 395, + 485, + 592 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Flop heavy fine-tuning Fine-tuning on a higher resolution at large scale was effectively done in Stable Diffusion [32] to limit the training cost, which is a form of flop heavy scaling over different resolutions. In this paper we show that flop heavy scaling is also preferred from a regularization perspective, which can be done either by finetuning from a lower resolution, or by lowering the patching size.", + "bbox": [ + 89, + 611, + 483, + 717 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 732, + 222, + 750 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Sigmoid weighting versus others", + "text_level": 1, + "bbox": [ + 89, + 757, + 375, + 773 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Loss weighting In this experiment, we want to understand the effect of loss weighting on the final image quality of a model. For the noise schedule, we use an interpolated cosine schedule because it has relatively broad coverage over logsnr values. We compare the weightings corresponding to the original $\\epsilon$ -mse [11] and the multiscale $\\epsilon$ -mse [14], the EDM shifted weightings [19], including the original and monotonic versions [19, 26], and the proposed sigmoid weighting", + "bbox": [ + 89, + 779, + 485, + 902 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/0a20f14d695a3b2c1dc90f07e7e80a7af439384ddd39d007754bda786705f538.jpg", + "image_caption": [ + "Figure 6. Effect of training noise schedules on the FID score of ImageNet512 over the training iterations. Fortunately, with our proposed sigmoid weighting $(b = -3)$ , the exact choice of the noise schedules has little effect on the training process." + ], + "image_footnote": [], + "bbox": [ + 500, + 85, + 901, + 272 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/efb9313d8f94b8d94214c99f76f5cc0c88c6879119bcc7ddafe1adbca754cbf4.jpg", + "table_caption": [ + "Table 1. Sigmoid weighting performance (left) with the flop heavy architecture. Larger resolutions require more negative bias. Power loss performance (right) on ImageNet512 with the small model. Increasing downsampling has a similar effect as decreasing bias, hence we prefer the more simple sigmoid weighting." + ], + "table_footnote": [], + "table_body": "
(a) Sigmoid weighting(b) Power loss weighting
biasresolutionbiasdownsample levels
12825651210240123
-5---1.82-42.192.583.625.11
-4--1.631.75-32.192.322.854.02
-3-1.471.482.32-22.292.152.343.13
-21.491.421.552.17-12.642.322.252.54
-11.351.381.75-03.222.742.072.22
01.301.461.94-1-2.692.292.16
11.261.56--2-3.262.532.19
", + "bbox": [ + 517, + 428, + 895, + 563 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "with shift $b = -3$ . As can be seen in Figure 5, the sigmoid weighting outperforms all other weightings, with EDM monotonic shifted being the closest alternative. Appendix C contains additional experiments, where we also evaluate the sigmoid weighting against the closest competitor without guidance. Here again sigmoid weighting is consistently better. We also evaluate a time-shifted variant, where the bias depends on the progress in training.", + "bbox": [ + 511, + 579, + 908, + 702 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training efficiency of noise schedules Since the training noise schedule can be decoupled from the loss weighting and it only affects the training efficiency but not the final performance, one can measure the convergence speed over the training process caused by leveraging particular noise schedules. In this experiment we test the cosine schedule [11], the shifted cosine and the interpolated cosine schedules [14] (Figure 6). For fair comparison, we fix the sampling noise schedule as the cosine interpolated schedule. Interestingly, when the sigmoid weighting is used, the choice of noise schedule does not affect the training efficiency significantly. Arguably, the shifted schedule is a bit more sample effi", + "bbox": [ + 511, + 719, + 910, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "18066", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/cfba554f0499d140f421d739d32d5229f4c5d7630c45e389a8d625935e959e56.jpg", + "image_caption": [ + "Figure 7. FID versus training cost on ImageNet512. Note that the EDM-2 training curves are without guidance, but the best performance with guidance intervals is reported with a separate marker, the figure source adapted has been adapted from EDM2 [20]. Following their assumptions, one training iteration is three times as expensive as evaluating the model. Our Flop Heavy (finetune) is trained by first training a small model on ImageNet256 for 400k iterations and then finetuning the same model on ImageNet512." + ], + "image_footnote": [], + "bbox": [ + 119, + 87, + 880, + 280 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/584f7bccdcdfe72473e5875f5ad05299e881bc000dab70b530254646605fc7cc.jpg", + "image_caption": [ + "Figure 8. The effect of scaling either channels or tokens, compared to the small model variant. FLOPS between model size $2\\mathrm{x}$ and flop heavy are similar." + ], + "image_footnote": [], + "bbox": [ + 91, + 357, + 359, + 481 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c32f83e922fb55d924c8bf41c4c5064257dcb4a5573c063f4ece679f81f58f5a.jpg", + "image_caption": [ + "Figure 9. The effect of Residual-UViTs (removing block skip connections). Although smaller models can be effected, larger models are unaffected by the simplification." + ], + "image_footnote": [], + "bbox": [ + 364, + 357, + 632, + 481 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/21d9702abf3eda288b9e9681b14556e5f7071b8f2e50a14b2017ca07968d5fd0.jpg", + "image_caption": [ + "Figure 10. Changing the encoder decoder architecture. Decoder blocks are more important than encoder blocks, and symmetric architectures have decent performance." + ], + "image_footnote": [], + "bbox": [ + 638, + 352, + 906, + 481 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "cient initially but all schedules perform similarly towards the end. This suggests that there is a lot of transfer across different logsnr values. We also tried the EDM training noise schedule with the sigma power sampling noise schedule [19] (not plotted), with both shifted or the original version, but its narrower logsnr sampling schedule led to performance degradation, achieving about 9 FID at best. This suggests that for high resolution, a wide logsnr range is crucial to cover the important noise levels.", + "bbox": [ + 88, + 573, + 482, + 708 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Bias over resolutions In Table 1a, we show the effect of the bias term $b$ on the final performance when trained using a flop heavy model. All models use the same guidance and similar guidance intervals, where the lower bound of the guidance interval has been shifted according to resolution (again approximately 1.5 per 2x resolution change), see Appendix B for more details.", + "bbox": [ + 89, + 729, + 483, + 835 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Sigmoid weighting versus the power loss In literature, high resolution generation can greatly benefit from having a multiscale loss [6, 7, 14]. To that end we create a multiscale", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "variant that we name the power loss. In addition to the sigmoid weighting that effectively upweights low frequency components of the data and downweights high frequency components, the power loss aims to amplify the loss of low sub-band (smooth) signals in the spatial domain via Haar wavelets. For details on this loss see Appendix B.2.", + "bbox": [ + 511, + 573, + 906, + 662 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In Table 1b we see that the bias should be adjusted based on the number of Haar transform downsampling levels. Despite being competitive, the performance of power loss is roughly on par with the one with sigmoid weighting but without multiscaling (i.e. downsample levels $= 0$ ), if we tune the bias of the sigmoid weighting more aggressively. Interestingly, lowering the bias has almost the same effect as increasing the scale. Therefore, building the multiscale variant is not necessary with the sigmoid weighting.", + "bbox": [ + 511, + 665, + 908, + 801 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Scaling channels vs tokens on ImageNet", + "text_level": 1, + "bbox": [ + 511, + 816, + 856, + 832 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section we show the difference between scaling a model 4x FLOPS, either via channels or by creating larger resolution feature maps (or put differently, feature maps with a higher token count). Note that our small model on", + "bbox": [ + 511, + 839, + 906, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "18067", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ImageNet512 is patched $4 \\times 4$ . This model is scaled either by increasing channels by two (named model size $2x$ ), or by reducing patching to $2 \\times 2$ (named flop heavy, because FLOPS over params is much higher). Indeed as can be seen in Figure 8, the flop heavy model performs much better than the model size $2x$ , when trained in a standard way on ImageNet512. Note that unlike most experiments, in these experiments it was necessary to use very different optimal guidance settings, more details in Appendix B.", + "bbox": [ + 89, + 90, + 480, + 227 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We confirm this is an issue related to overfitting, by training the model size 2x with distribution augmentation [18] the model size 2x variant performs much better, with an FID of 1.6 instead of 1.9 (latter in Figure 8). Still, without any augmentation beyond the standard horizontal flips, the flop heavy model achieves a score of 1.54.", + "bbox": [ + 89, + 229, + 482, + 320 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Removing block skip connection", + "text_level": 1, + "bbox": [ + 89, + 339, + 375, + 354 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section we study the difference between the residual skip-connections and the blockwise skip-connections. As hypothesized, the small model performance drops a bit because of this. For ImageNet512 a deficit of about 0.2 FID (in Figure 9) and for Imagenet256 a deficit of about 0.05 FID. That the task of fitting ImageNet256 is less affected by removing skip-connections can be explained by the task being easier: The task of fitting ImageNet256 is a subset of fitting ImageNet512.", + "bbox": [ + 89, + 363, + 482, + 500 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For larger models, removing blockwise skip-connections in favor of a single residual connection does not degrade performance, and may even lead to slightly increased performance (see Figure 9). This is somewhat unexpected, and the phenomenon persists with different guidance settings as well. The FID curve reminds of the model size $2\\mathrm{x}$ curve, which at first improves but flattens out over iterations. Therefore our best hypothesis is that the block skip (flop) baseline may have similar problems such as overfitting. Regardless, it shows that Residual U-ViTs scale well.", + "bbox": [ + 89, + 503, + 482, + 654 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Asymmetric U-ViT Since our U-ViT models encompass only a single skip connection for each level, we can use a different number of ResNet or ViT blocks per level. This allows to move compute from the encoder stage (the down sampling half of the model) to the decoder stage (the upsampling half of the model). From Figure 10 we observe that (a) for a fixed compute budget, it is better to add more parameters to the decoder, the encoder heavy model (E5-D1) is significantly worse than the baseline (E3-D3) and the decoder heavy model (E1-D5); (b) there needs to be some balance between the encoder and the decoder, E1-D7 works less good than E3-D5. (c) by a slightly heavier model (E3-D7, only $10\\%$ more parameters than our small (E3-D3) model, the performance significantly increases.", + "bbox": [ + 89, + 688, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/ef128d1d326e3d3a214cf0c61296f4451a59a5987594436ff51fc59603ea22ff.jpg", + "table_caption": [ + "Table 2. Literature Comparison on ImageNet on FID. Note that even though certain NFEs are high (e.g. 4100 for CDM) typically these suffer minor performance degradation for lower NFEs and should not be taken too literally." + ], + "table_footnote": [], + "table_body": "
MethodNFEFID
Imagenet 128 x 128
RIN [16]10002.75
SD, U-ViT-L [14]5121.94
StyleGAN-XL [38]11.81
VDM+++ [25]2501.75
DisCo-Diff [46]4141.73
PaGoDA [22]11.48
SiD2, Flop Heavy (ours)5121.26
Imagenet 256 x 256
ADM-G [4]2504.59
ADM-U [4]2503.94
CDM [12]41004.88
RIN [3, 16]10003.52
SD, U-ViT-L [14]5122.44
MDM [7]2503.51
StyleGAN-XL [38]12.30
DiT-XL/2 [31]2502.27
VDM++ [25]2502.12
RDM [42]3561.89
PaGoDA [22]11.56
SiD2, small (ours)5121.72
SiD2, Flop Heavy (ours)5121.38
Imagenet 512 x 512
ADM-G [4]2507.72
RIN [3, 16]10003.95
ADM-U [4]2503.85
LDM-4-G [32]2503.60
SD, U-ViT-L [14]5123.02
DiffiT [9]2502.67
StyleGAN-XL [38]12.40
DiT-XL/2 [31]2503.04
VDM++ [25]2502.65
MAGVIT-v2 [51]641.91
PaGoDA [22]11.80
EDM2-S (interval) [27]631.68
EDM2-XXL (interval) [27]631.40
SiD2, small (ours)5122.19
SiD2, Flop Heavy (ours)5121.48
SiD2, Flop Heavy distilled (ours)161.50
Imagenet 1024 x 1024
RIN [16]10008.72
SiD2, Flop Heavy (ours)5121.75
", + "bbox": [ + 535, + 146, + 883, + 868 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "18068", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/5d6987a48d055fb4e712dfadd36587f8b0d1bf41716e0491a2aae199d07621da.jpg", + "table_caption": [ + "Table 3. Zero-shot text-to-image performance evaluated on MSCOCO in $\\mathrm{FID}_{30\\mathrm{k}}$ . Our distilled model, evaluated on the more difficult 512 setting, outperforming other works." + ], + "table_footnote": [], + "table_body": "
MethodSizeNFEFID30k
DeepFloyd [40]2566.7
eDiff-I [1]2566.9
Parti [49]2567.2
Imagen [35]2567.3
simple diffusion [14]2568.3
simple diffusion [14]5122569.6
SDv1.5 (eval DMD2) [32]5122007.2
DMD2 [48]51248.3
SiD2 (ours, bias=-3)5122568.1
SiD2 16-step distilled (ours)512166.7
", + "bbox": [ + 133, + 142, + 442, + 313 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Literature comparison", + "text_level": 1, + "bbox": [ + 89, + 329, + 302, + 345 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Here we compare different models from the literature. In the first section end-to-end diffusion models are compared, and in the second section we compare between all types of generative models. In Table 2 it can be seen that our SiD2 is very competitive with existing generative models: On ImageNet128 and ImageNet256 it achieves state-of-the-art performance, on ImageNet512 it is a little worse but still competitive with the best latent diffusion model and outperforms all other models, pixel-based and latent.", + "bbox": [ + 89, + 349, + 485, + 488 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Comparing diffusion models on compute The training cost SiD2 is another helpful way to see how the model scales, as depicted in Figure 7). Importantly, the performance of our pixel-based approach outperforms other significantly. Observe for example the performance of our small variant (using no guidance interval to keep this comparison fair), which roughly the same performance as the VDM++ while only using a third of the training cost, and a fourth of the FLOPS in a single forward pass (see Figure 16 in Appendix C). It is also competitive with the best currently known latent diffusion approach EDM2 [20], but in fairness EDM2 scales a little bit better. These models attain even better performance for a smaller training cost, and a similar story holds for forward pass complexity. End-to-end diffusion can work much better than previously demonstrated, but it is still somewhat outperformed by latent diffusion on quality versus compute comparisons for both sampling and training.", + "bbox": [ + 89, + 505, + 486, + 763 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Text to image generation Finally, we show that our method can be used for end-to-end training of text-to-image models. Here, evaluation is more difficult as MSCOCO zero-shot performance depends a lot on the dataset used and is known to not necessarily correspond to actual perceived performance. Nevertheless, it is important that the model achieves performance in the right ballpark. Indeed, when we train SiD2 it performs comparable to other models in the", + "bbox": [ + 89, + 780, + 485, + 902 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/d5236b904016a7ed46178509b5d43b738cfd980c3ca046b1e34ffde478850fc2.jpg", + "table_caption": [ + "Table 4. Conditional Video generation performance on Kinetics600 in FVD (resized to 64). For video generation the loss specification is important to obtain good sample quality (see ablations)." + ], + "table_footnote": [], + "table_body": "
MethodFVD50k
pixel diffusion
Video Diffusion Models [13]16.6
RIN [16]10.8
Rolling Diffusion [34]5.2
latent diffusion
MAGVIT [50]9.9
MAGVITv2 [51]4.3
W.A.L.T [8]3.3
SiD2 (ours, sigmoid bias = 0)2.3
ablations
eps-mse5.7
sigmoid bias = +23.2
sigmoid bias = -22.5
", + "bbox": [ + 596, + 142, + 826, + 345 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "literature on MSCOCO $\\mathrm{FID}_{30\\mathrm{K}}$ , and detailed samples have only a small number of artifacts (Figure 2). Furthermore, the moment matching distilled version using [37] outperforms the baseline model as previously observed. This model also outperforms all other zero-shot approaches at 512 resolution as shown in Table 3.", + "bbox": [ + 511, + 361, + 908, + 452 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Video Generation To test whether our approach transfers to video, we test our model on Kinetics600 following [13], by conditioning on 5 frames and generating (sometimes referred to as predicting) the subsequent 11 frames. We then compute the FVD between the training set and the generated samples both on the subsequent 11 frames. Even though our method generates at $128 \\times 128$ , we follow [8] and resize samples to have resolution $64 \\times 64$ . Our model achieves SOTA performance and it turns out that tuning the loss is very important to achieve this as shown in Table 4. Note that the latent approaches ([8, 50, 51]) use much less training compute than our models in pixel-space, which we discuss in more detail in the Appendix B.", + "bbox": [ + 511, + 469, + 908, + 667 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Discussion", + "text_level": 1, + "bbox": [ + 511, + 679, + 627, + 694 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper we present Simpler Diffusion (SiD2) and show how one can train end-to-end diffusion and achieve competitive performance with latent models: (1) use the sigmoid loss, (2) with a simpler U-ViT architecture, and (3) favour a flop heavy model design. As a result, our method outperforms existing end-to-end models by a large margin. Nevertheless, at this moment latent diffusion methods still seem to have slightly better scaling properties and it remains to be seen whether this gap can be closed any further. Importantly, for many applications training a separate autoencoder may not be either practical or desired. Therefore, we believe this paper is an important contribution that allows end-to-end modelling directly in the pixel-space.", + "bbox": [ + 511, + 703, + 910, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "18069", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Ji-aming Song, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, Bryan Catanzaro, Tero Karras, and Ming-Yu Liu. ediffi: Text-to-image diffusion models with an ensemble of expert denoisers. CoRR, abs/2211.01324, 2022. 8", + "[2] Fan Bao, Shen Nie, Kaiwen Xue, Yue Cao, Chongxuan Li, Hang Su, and Jun Zhu. All are worth words: A vit backbone for diffusion models. In CVPR, 2023. 5", + "[3] Ting Chen. On the importance of noise scheduling for diffusion models. arxiv, 2023. 1, 4, 7", + "[4] Prafulla Dhariwal and Alex Nichol. Diffusion models beat gans on image synthesis. CoRR, abs/2105.05233, 2021. 4, 7", + "[5] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In 9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7, 2021. OpenReview.net, 2021. 5", + "[6] Jiatao Gu, Shuangfei Zhai, Yizhe Zhang, Miguel Angel Bautista, and Josh M. Susskind. f-dm: A multi-stage diffusion model via progressive signal transformation. CoRR, abs/2210.04955, 2022. 4, 6", + "[7] Jiatao Gu, Shuangfei Zhai, Yizhe Zhang, Josh M. Susskind, and Navdeep Jaitly. Matryoshka diffusion models. CoRR, abs/2310.15111, 2023. 4, 6, 7", + "[8] Agrim Gupta, Lijun Yu, Kihyuk Sohn, Xiuye Gu, Meera Hahn, Fei-Fei Li, Irfan Essa, Lu Jiang, and José Lezama. Photorealistic video generation with diffusion models. In Computer Vision - ECCV 2024 - 18th European Conference, Milan, Italy, September 29-October 4, 2024, Proceedings, Part LXXIX, pages 393-411. Springer, 2024. 8", + "[9] Ali Hatamizadeh, Jiaming Song, Guilin Liu, Jan Kautz, and Arash Vahdat. Diffit: Diffusion vision transformers for image generation. CoRR, abs/2312.02139, 2023. 7", + "[10] Jonathan Heek, Emiel Hoogeboom, and Tim Salimans. Multistep consistency models. CoRR, abs/2403.06807, 2024. 3", + "[11] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In Advances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, NeurIPS, 2020. 1, 5", + "[12] Jonathan Ho, Chitwan Sahara, William Chan, David J. Fleet, Mohammad Norouzi, and Tim Salimans. Cascaded diffusion models for high fidelity image generation. J. Mach. Learn. Res., 23:47:1-47:33, 2022. 4, 7", + "[13] Jonathan Ho, Tim Salimans, Alexey A. Gritsenko, William Chan, Mohammad Norouzi, and David J. Fleet. Video diffusion models. In Advances in Neural Information Processing Systems 35: Annual Conference on Neural Information Processing Systems 2022, NeurIPS, 2022. 8", + "[14] Emiel Hoogeboom, Jonathan Heek, and Tim Salimans. simple diffusion: End-to-end diffusion for high resolution images. In International Conference on Machine Learning, ICML, pages 13213-13232. PMLR, 2023. 1, 3, 4, 5, 6, 7, 8" + ], + "bbox": [ + 93, + 114, + 485, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] Zhongzhan Huang, Zhou Pan, Shuicheng Yan, and Liang Lin. Scalelong: Towards more stable training of diffusion model via scaling network long skip connection. In NeurIPS, 2023. 5", + "[16] Allan Jabri, David J. Fleet, and Ting Chen. Scalable adaptive computation for iterative generation. CoRR, abs/2212.11972, 2022. 7, 8", + "[17] Zeyinzi Jiang, Chaojie Mao, Yulin Pan, Zhen Han, and Jingfeng Zhang. Scedit: Efficient and controllable image diffusion generation via skip connection editing. Technical Report 2312.11392, arXiv, 2023. 5", + "[18] Heewoo Jun, Rewon Child, Mark Chen, John Schulman, Aditya Ramesh, Alec Radford, and Ilya Sutskever. Distribution augmentation for generative modeling. In Proceedings of the 37th International Conference on Machine Learning, ICML, 2020. 3, 7", + "[19] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. In Advances in Neural Information Processing Systems 35: Annual Conference on Neural Information Processing Systems 2022, NeurIPS, 2022. 5, 6", + "[20] Tero Karras, Miika Aittala, Jaakko Lehtinen, Janne Hellsten, Timo Aila, and Samuli Laine. Analyzing and improving the training dynamics of diffusion models. CoRR, abs/2312.02696, 2023. 1, 4, 6, 8, 2, 5", + "[21] Tero Karras, Miika Aittala, Tuomas Kynkänniemi, Jaakko Lehtinen, Timo Aila, and Samuli Laine. Guiding a diffusion model with a bad version of itself. CoRR, abs/2406.02507, 2024. 2", + "[22] Dongjun Kim, Chieh-Hsin Lai, Wei-Hsiang Liao, Yuhta Takida, Naoki Murata, Toshimitsu Uesaka, Yuki Mitsufuji, and Stefano Ermon. Pagoda: Progressive growing of a one-step generator from a low-resolution diffusion teacher. CoRR, abs/2405.14822, 2024. 7, 5", + "[23] Dongjun Kim, Chieh-Hsin Lai, Wei-Hsiang Liao, Naoki Murata, Yuhta Takida, Toshimitsu Uesaka, Yutong He, Yuki Mitsufuji, and Stefano Ermon. Consistency trajectory models: Learning probability flow ODE trajectory of diffusion. In The Twelfth International Conference on Learning Representations, 2024. 4", + "[24] Durk P Kingma and Prafulla Dhariwal. Glow: Generative flow with invertible 1x1 convolutions. Advances in neural information processing systems, 31, 2018. 1", + "[25] Diederik P. Kingma and Ruiqi Gao. Understanding the diffusion objective as a weighted integral of ellbods. CoRR, abs/2303.00848, 2023. 1, 2, 3, 4, 7", + "[26] Diederik P. Kingma, Tim Salimans, Ben Poole, and Jonathan Ho. Variational diffusion models. CoRR, abs/2107.00630, 2021. 2, 5", + "[27] Tuomas Kynkänniemi, Miika Aittala, Tero Karras, Samuli Laine, Timo Aila, and Jaakko Lehtinen. Applying guidance in a limited interval improves sample and distribution quality in diffusion models. CoRR, abs/2404.07724, 2024. 7, 2", + "[28] Yaron Lipman, Ricky T. Q. Chen, Heli Ben-Hamu, Maximilian Nickel, and Matthew Le. Flow matching for generative modeling. In The Eleventh International Conference on Learning Representations, ICLR 2023, Kigali, Rwanda. OpenReview.net, 2023. 2" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "18070", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[29] Weijian Luo, Tianyang Hu, Shifeng Zhang, Jiacheng Sun, Zhenguo Li, and Zhihua Zhang. Diff-instruct: A universal approach for transferring knowledge from pre-trained diffusion models. Advances in Neural Information Processing Systems, 36, 2024. 4", + "[30] Jiajun Ma, Shuchen Xue, Tianyang Hu, Wenjia Wang, Zhaoqiang Liu, Zhenguo Li, Zhi-Ming Ma, and Kenji Kawaguchi. The surprising effectiveness of skip-tuning in diffusion sampling. In Proceedings of the 41st International Conference on Machine Learning, pages 34053-34074, 2024. 5", + "[31] William Peebles and Saining Xie. Scalable diffusion models with transformers. CoRR, abs/2212.09748, 2022. 4, 7, 5", + "[32] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 10674-10685. IEEE, 2022. 1, 3, 4, 5, 7, 8", + "[33] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. Technical report, ArXiV, 2015. 4", + "[34] David Ruhe, Jonathan Heek, Tim Salimans, and Emiel Hoogeboom. Rolling diffusion models. In *Forty-first International Conference on Machine Learning*, ICML 2024, Vienna, Austria, July 21-27, 2024. OpenReview.net, 2024. 8", + "[35] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S. Sara Mahdavi, Rapha Gontijo Lopes, Tim Salimans, Jonathan Ho, David J. Fleet, and Mohammad Norouzi. Photorealistic text-to-image diffusion models with deep language understanding. CoRR, abs/2205.11487, 2022. 8", + "[36] Tim Salimans and Jonathan Ho. Progressive distillation for fast sampling of diffusion models. In *The Tenth International Conference on Learning Representations*, ICLR. OpenReview.net, 2022. 4", + "[37] Tim Salimans, Thomas Mensink, Jonathan Heek, and Emiel Hoogeboom. Multistep distillation of diffusion models via moment matching. arXiv preprint arXiv:2406.04103, 2024. 4, 8", + "[38] Axel Sauer, Katja Schwarz, and Andreas Geiger. Stylegan-xl: Scaling stylegan to large diverse datasets. In SIGGRAPH '22: Special Interest Group on Computer Graphics and Interactive Techniques Conference, pages 49:1-49:10. ACM, 2022. 7, 5", + "[39] Evan Shelhamer, Jonathan Long, and Trevor Darrell. Fully convolutional networks for semantic segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 39(4):640-651, 2016. 4", + "[40] Alex Shonenkov, Misha Konstantinov, Daria Bakshandaeva, Christoph Schuhmann, Ksenia Ivanova, and Nadija Klokova. Deepfloyd if: A text-to-image model with a high degree of photorealism and language understanding, 2023. 4, 8", + "[41] Yang Song, Jascha Sohl-Dickstein, Diederik P. Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In 9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7, 2021. OpenReview.net, 2021. 2" + ], + "bbox": [ + 91, + 90, + 485, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[42] Jiayan Teng, Wendi Zheng, Ming Ding, Wenyi Hong, Jianqiao Wangni, Zhuoyi Yang, and Jie Tang. Relay diffusion: Unifying diffusion process across resolutions for image synthesis. In The Twelfth International Conference on Learning Representations, ICLR. OpenReview.net, 2024. 4, 7", + "[43] Haoning Wu, Shaocheng Shen, Qiang Hu, Xiaoyun Zhang, Ya Zhang, and Yanfeng Wang. Megafusion: Extend diffusion models towards higher-resolution image generation without further tuning. CoRR, abs/2408.11001, 2024. 4", + "[44] Zhisheng Xiao, Karsten Kreis, and Arash Vahdat. Tackling the generative learning trilemma with denoising diffusion gans. In International Conference on Learning Representations, 2021. 4", + "[45] Yanwu Xu, Yang Zhao, Zhisheng Xiao, and Tingbo Hou. Ufogen: You forward once large scale text-to-image generation via diffusion gans. arXiv preprint arXiv:2311.09257, 2023. 4", + "[46] Yilun Xu, Gabriele Corso, Tommi S. Jaakkola, Arash Vahdat, and Karsten Kreis. Disco-diff: Enhancing continuous diffusion models with discrete latents. In *Forty-first International Conference on Machine Learning*, ICML 2024, Vienna, Austria, July 21-27, 2024. OpenReview.net, 2024. 7", + "[47] Tianwei Yin, Michael Gharbi, Richard Zhang, Eli Shechtman, Fredo Durand, William T Freeman, and Taesung Park. One-step diffusion with distribution matching distillation. arXiv preprint arXiv:2311.18828, 2023. 4", + "[48] Tianwei Yin, Michael Gharbi, Taesung Park, Richard Zhang, Eli Shechtman, Fredo Durand, and William T. Freeman. Improved distribution matching distillation for fast image synthesis. In NeurIPS, 2024. 8", + "[49] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, Ben Hutchinson, Wei Han, Zarana Parekh, Xin Li, Han Zhang, Jason Baldridge, and Yonghui Wu. Scaling autoregressive models for content-rich text-to-image generation. CoRR, abs/2206.10789, 2022. 8", + "[50] Lijun Yu, Yong Cheng, Kihyuk Sohn, José Lezama, Han Zhang, Huiwen Chang, Alexander G. Hauptmann, Ming-Hsuan Yang, Yuan Hao, Irfan Essa, and Lu Jiang. MAGVIT: masked generative video transformer. CoRR, abs/2212.05199, 2022. 8", + "[51] Lijun Yu, José Lezama, Nitesh Bharadwaj Gundavarapu, Luca Versari, Kihyuk Sohn, David Minnen, Yong Cheng, Agrim Gupta, Xiuye Gu, Alexander G. Hauptmann, Boqing Gong, Ming-Hsuan Yang, Irfan Essa, David A. Ross, and Lu Jiang. Language model beats diffusion - tokenizer is key to visual generation. In The Twelfth International Conference on Learning Representations, ICLR. OpenReview.net, 2024. 7, 8", + "[52] Wendi Zheng, Jiayan Teng, Zhuoyi Yang, Weihan Wang, Jidong Chen, Xiaotao Gu, Yuxiao Dong, Ming Ding, and Jie Tang. Cogview3: Finer and faster text-to-image generation via relay diffusion. In Computer Vision - ECCV Proceedings, Part LXXVII, pages 1-22. Springer, 2024. 4" + ], + "bbox": [ + 516, + 90, + 906, + 825 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "18071", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Simpler Diffusion_ 1.5 FID on ImageNet512 with Pixel-space Diffusion/3cf16242-5862-4aa6-864f-5aba8865f5f9_model.json b/2025/Simpler Diffusion_ 1.5 FID on ImageNet512 with Pixel-space Diffusion/3cf16242-5862-4aa6-864f-5aba8865f5f9_model.json new file mode 100644 index 0000000000000000000000000000000000000000..129b04dab8698a3621e9fef73c78c579040c7502 --- /dev/null +++ b/2025/Simpler Diffusion_ 1.5 FID on ImageNet512 with Pixel-space Diffusion/3cf16242-5862-4aa6-864f-5aba8865f5f9_model.json @@ -0,0 +1,2211 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.146, + 0.131, + 0.852, + 0.154 + ], + "angle": 0, + "content": "Simpler Diffusion: 1.5 FID on ImageNet512 with pixel-space diffusion" + }, + { + "type": "text", + "bbox": [ + 0.194, + 0.182, + 0.348, + 0.199 + ], + "angle": 0, + "content": "Emiel Hoogeboom" + }, + { + "type": "text", + "bbox": [ + 0.194, + 0.2, + 0.346, + 0.217 + ], + "angle": 0, + "content": "Google Deepmind" + }, + { + "type": "text", + "bbox": [ + 0.184, + 0.218, + 0.362, + 0.234 + ], + "angle": 0, + "content": "emielh@google.com" + }, + { + "type": "text", + "bbox": [ + 0.433, + 0.182, + 0.575, + 0.198 + ], + "angle": 0, + "content": "Thomas Mensink" + }, + { + "type": "text", + "bbox": [ + 0.433, + 0.2, + 0.578, + 0.217 + ], + "angle": 0, + "content": "Google Deepmind" + }, + { + "type": "text", + "bbox": [ + 0.409, + 0.218, + 0.599, + 0.234 + ], + "angle": 0, + "content": "mensink@google.com" + }, + { + "type": "text", + "bbox": [ + 0.671, + 0.182, + 0.791, + 0.198 + ], + "angle": 0, + "content": "Jonathan Heek" + }, + { + "type": "text", + "bbox": [ + 0.657, + 0.2, + 0.806, + 0.217 + ], + "angle": 0, + "content": "Google Deepmind" + }, + { + "type": "text", + "bbox": [ + 0.649, + 0.218, + 0.816, + 0.234 + ], + "angle": 0, + "content": "jheek@google.com" + }, + { + "type": "text", + "bbox": [ + 0.204, + 0.261, + 0.326, + 0.278 + ], + "angle": 0, + "content": "Kay Lamarigts" + }, + { + "type": "text", + "bbox": [ + 0.197, + 0.279, + 0.335, + 0.295 + ], + "angle": 0, + "content": "Google Research" + }, + { + "type": "text", + "bbox": [ + 0.145, + 0.296, + 0.386, + 0.312 + ], + "angle": 0, + "content": "kaylamerigts@google.com" + }, + { + "type": "text", + "bbox": [ + 0.474, + 0.261, + 0.559, + 0.278 + ], + "angle": 0, + "content": "Ruiqi Gao" + }, + { + "type": "text", + "bbox": [ + 0.444, + 0.279, + 0.59, + 0.295 + ], + "angle": 0, + "content": "Google Deepmind" + }, + { + "type": "text", + "bbox": [ + 0.429, + 0.296, + 0.607, + 0.312 + ], + "angle": 0, + "content": "ruiqig@google.com" + }, + { + "type": "text", + "bbox": [ + 0.694, + 0.261, + 0.806, + 0.277 + ], + "angle": 0, + "content": "Tim Salimans" + }, + { + "type": "text", + "bbox": [ + 0.676, + 0.278, + 0.825, + 0.295 + ], + "angle": 0, + "content": "Google Deepmind" + }, + { + "type": "text", + "bbox": [ + 0.65, + 0.296, + 0.848, + 0.312 + ], + "angle": 0, + "content": "salimans@google.com" + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.365, + 0.327, + 0.381 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.396, + 0.484, + 0.531 + ], + "angle": 0, + "content": "Latent diffusion models have become the popular choice for scaling up diffusion models for high resolution image synthesis. Compared to pixel-space models that are trained end-to-end, latent models are perceived to be more efficient and to produce higher image quality at high resolution. Here we challenge these notions, and show that pixel-space models can be very competitive to latent models both in quality and efficiency, achieving 1.5 FID on ImageNet512 and new SOTA results on ImageNet128, ImageNet256 and Kinetics600." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.533, + 0.484, + 0.684 + ], + "angle": 0, + "content": "We present a simple recipe for scaling end-to-end pixel-space diffusion models to high resolutions. 1: Use the sigmoid loss-weighting [25] with our prescribed hyperparameters. 2: Use our simplified memory-efficient architecture with fewer skip-connections. 3: Scale the model to favor processing the image at a high resolution with fewer parameters, rather than using more parameters at a lower resolution. Combining these with guidance intervals, we obtain a family of pixel-space diffusion models we call Simpler Diffusion (SiD2)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.71, + 0.222, + 0.725 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.484, + 0.87 + ], + "angle": 0, + "content": "Diffusion models have been tremendously successful in domains such as image, video and audio generation. For high resolution data, latent diffusion models [32] have been particularly successful: Through encoding the data in a lower dimensional space, these models sidestep the difficulties faced by more traditional pixel-space diffusion models in scaling up to high resolution. This makes latent diffusion models also scale better in training compute than other approaches, as shown by [20]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.901 + ], + "angle": 0, + "content": "However, latent models have their own challenges: It can be difficult to train the autoencoders that produce the latents," + }, + { + "type": "image", + "bbox": [ + 0.543, + 0.366, + 0.883, + 0.468 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.48, + 0.905, + 0.495 + ], + "angle": 0, + "content": "Figure 1. End-to-end pixel diffusion performance on ImageNet512." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.506, + 0.906, + 0.642 + ], + "angle": 0, + "content": "which in practice requires careful tuning of regularization hyperparameters [32]. Even when trained well, the decoder of a latent diffusion model can produce artifacts or otherwise imperfect reconstructions, especially when it is applied to latents that are out of distribution compared to the training set. A single-stage pixel-space diffusion model does not have these issues, and so we ask the question: Can pixel-space diffusion models scale as well as latent models with the right architecture and hyperparameters?" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.643, + 0.907, + 0.794 + ], + "angle": 0, + "content": "Over the years, end-to-end pixel-space diffusion has improved considerably (Figure 1). The main problem was that the standard schedule and loss-weighting [11] work well for small resolution images, like \\(32^2\\) or \\(64^2\\), but not at higher resolution [3, 14]. It was further shown [25] that it can be helpful to disentangle loss weighting and time sampling for improved control over both, leading to better performance at higher resolution. However, in previous works pixels-space models still did not match the best latent models on standard benchmarks like ImageNet512." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.796, + 0.909, + 0.901 + ], + "angle": 0, + "content": "The main contribution of this paper is to improve pixel-space diffusion on high resolution images while keeping as close to the original formulation of diffusion models as possible. Specifically, we analyze and tune the sigmoid loss weighting from [25], and show why it should be preferred over their favored EDM monotonic. We further propose architectural improvements that lead to even better perfor" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18062" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.093, + 0.088, + 0.291, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.296, + 0.088, + 0.496, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.089, + 0.701, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.708, + 0.089, + 0.905, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.239, + 0.252, + 0.757, + 0.267 + ], + "angle": 0, + "content": "Figure 2. Some text-to-image samples of \\(512 \\times 512\\) that have been generated by SiD2." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.293, + 0.484, + 0.384 + ], + "angle": 0, + "content": "mance, namely flop heavy scaling and removing block skip-connections (Residual U-ViTs). Most notably, we improve the state-of-the-art of pixel-space diffusion from 2.65 to 1.48 on ImageNet512 (see also Figure 1). Further, we also achieve overall state-of-the-art on ImageNet128, ImageNet256 and Kinetics600 without heavy tuning." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.394, + 0.222, + 0.41 + ], + "angle": 0, + "content": "1.1. Background" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.416, + 0.484, + 0.446 + ], + "angle": 0, + "content": "A diffusion process gradually destroys data over a conceptual time \\(t\\), typically with Gaussian noise as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.16, + 0.459, + 0.482, + 0.475 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {z} _ {t} = \\alpha_ {t} \\boldsymbol {x} + \\sigma_ {t} \\boldsymbol {\\epsilon}, \\text {w h e r e} \\boldsymbol {\\epsilon} \\sim \\mathcal {N} (0, \\mathbf {I}), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.487, + 0.484, + 0.699 + ], + "angle": 0, + "content": "which defines a distribution \\(q(\\pmb{z}_t|\\pmb{x})\\) such that \\(\\pmb{z}_0\\) is close to clean data and \\(\\pmb{z}_1\\) is close to Gaussian noise. In this paper we assume a variance-preserving [41] schedule, which is defined as \\(\\alpha_{t} = \\sqrt{\\sigma(\\lambda_{t})}\\) and \\(\\sigma_t = \\sqrt{\\sigma(-\\lambda_t)}\\), with \\(\\lambda_{t} = \\log \\left(\\alpha_{t}^{2} / \\sigma_{t}^{2}\\right)\\) being the log signal-to-noise ratio, which monotonically decreases with time as \\(t\\to 1\\), and \\(\\sigma (\\cdot)\\) being the sigmoid function. It can be shown that other types of schedules such as the variance-exploding schedule or the flow matching schedule [28] can always be converted to a variance-preserving schedule with a linear scaling of \\(\\pmb{z}_{t}\\) [25]. The key defining property of a noise schedule is the log signal-to-noise ratio \\(\\lambda_{t}\\). A neural network is used to predict \\(\\hat{\\pmb{x}} = \\hat{\\pmb{x}}_{\\theta}(\\pmb{z}_{t},t)\\) that effectively estimates \\(\\mathbb{E}[\\pmb {x}|\\pmb {z}_t]\\), which can be done by minimizing a weighted mean squared error loss:" + }, + { + "type": "equation", + "bbox": [ + 0.149, + 0.71, + 0.482, + 0.739 + ], + "angle": 0, + "content": "\\[\nL (\\boldsymbol {x}) = \\mathbb {E} _ {t \\sim \\mathcal {U} (0, 1)} - \\frac {\\mathrm {d} \\lambda_ {t}}{\\mathrm {d} t} w \\left(\\lambda_ {t}\\right) \\| \\boldsymbol {x} - \\hat {\\boldsymbol {x}} \\| ^ {2}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.87 + ], + "angle": 0, + "content": "where \\( w(\\lambda_t) \\) is the weighting function. \\( -\\mathrm{d}\\lambda_{t} / \\mathrm{d}t \\) cancels out the effect of the noise schedule \\( \\lambda_{t} \\) on the loss. In other words, \\( L(\\mathbf{x}) \\) can be rewritten as \\( \\int -w(\\lambda)\\| \\pmb {x} - \\hat{\\pmb{x}}\\| ^2 d\\lambda \\) that is invariant to the choice of \\( \\lambda_{t} \\). However, \\( \\lambda_{t} \\) still impacts the sampling efficiency of the Monte Carlo estimator of the expected loss, and thus affects the training dynamics of diffusion models. This decouples noise schedule \\( \\lambda_{t} \\) and weighting function \\( w(\\lambda) \\) as a function of \\( \\lambda \\) [25]." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.483, + 0.901 + ], + "angle": 0, + "content": "The above squared error loss can be derived from the variational interpretation of diffusion models [26]. Given" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.293, + 0.908, + 0.355 + ], + "angle": 0, + "content": "two time steps \\(s < t\\), we define the denoising distribution \\(p(\\boldsymbol{z}_s|\\boldsymbol{z}_t)\\) in the same formulation as the true posterior \\(q(\\boldsymbol{z}_s|\\boldsymbol{z}_t,\\boldsymbol{x})\\), but replace the ground truth \\(\\boldsymbol{x}\\) with the learned approximation \\(\\hat{\\boldsymbol{x}} \\approx \\mathbb{E}[\\boldsymbol{x}|\\boldsymbol{z}_t]\\):" + }, + { + "type": "equation", + "bbox": [ + 0.55, + 0.366, + 0.905, + 0.382 + ], + "angle": 0, + "content": "\\[\np \\left(\\boldsymbol {z} _ {s} \\mid \\boldsymbol {z} _ {t}\\right) = q \\left(\\boldsymbol {z} _ {s} \\mid \\boldsymbol {z} _ {t}, \\boldsymbol {x} = \\hat {\\boldsymbol {x}}\\right) = \\mathcal {N} \\left(\\boldsymbol {z} _ {s} \\mid \\mu_ {t \\rightarrow s}, \\sigma_ {t \\rightarrow s}\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.394, + 0.906, + 0.518 + ], + "angle": 0, + "content": "where \\(\\sigma_{t\\rightarrow s}^2 = \\left(1 / \\sigma_s^2 +\\alpha_{t|s}^2 /\\sigma_{t|s}^2\\right)^{-1}\\) and \\(\\mu_{t\\to s} = \\sigma_{t\\to s}^{2}\\Bigl (\\alpha_{t|s} / \\sigma_{t|s}^{2}\\pmb {z}_{t} + \\alpha_{s} / \\sigma_{s}^{2}\\pmb {x}\\Bigr)\\) with \\(\\alpha_{t|s} = \\alpha_t / \\alpha_s\\) and \\(\\sigma_{t|s}^2 = \\sigma_t^2 -\\alpha_{t|s}\\sigma_s^2\\) . Note that during sampling, the standard deviation for \\(p(z_{s}|z_{t})\\) is chosen to be a log-linear interpolation between \\(\\sigma_{t|s}\\) and \\(\\sigma_{t\\rightarrow s}\\) of the form \\(\\sigma_{t\\rightarrow s}^{\\gamma}\\sigma_{t|s}^{1 - \\gamma}\\) where \\(\\gamma = 0.3\\) was set empirically, which is important as FID is sensitive to over-smooth or grainy images." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.533, + 0.688, + 0.55 + ], + "angle": 0, + "content": "2. Simpler Diffusion" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.559, + 0.753, + 0.575 + ], + "angle": 0, + "content": "2.1. Revisiting the sigmoid loss" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.581, + 0.908, + 0.764 + ], + "angle": 0, + "content": "Recall that one of the proposed losses in VDM++ [25] is the sigmoid loss. They expressed the loss with respect to \\(\\epsilon\\)-mse: \\(L(\\pmb{x}) = \\mathbb{E}_{t \\sim \\mathcal{U}(0,1)} - \\frac{\\mathrm{d}\\lambda_t}{\\mathrm{d}t} \\tilde{w}(\\lambda_t) \\| \\pmb{\\epsilon} - \\hat{\\pmb{\\epsilon}} \\|^2\\), which is connected to \\(\\pmb{x}\\)-mse through \\(\\| \\pmb{\\epsilon} - \\hat{\\pmb{\\epsilon}} \\|^2 = \\exp(\\lambda_t) \\| \\pmb{x} - \\hat{\\pmb{x}} \\|^2\\) and \\(\\tilde{w}(\\lambda_t) = \\exp(\\lambda_t) w(\\lambda_t)\\). In VDM++, viewing the loss in \\(\\epsilon\\)-mse enables a natural connection between the weighted loss and the ELBO and showed that losses should be monotonically decreasing on \\(\\epsilon\\)-mse. Hence, they propose the sigmoid loss: \\(\\tilde{w}(\\lambda_t) = \\sigma(b - \\lambda_t)\\), where \\(b\\) denotes the bias or shift and \\(\\sigma\\) is the sigmoid function. However, they ultimately opted for another loss-weighting named EDM-monotonic for high resolution image generation, see Figure 3a." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.786, + 0.908, + 0.847 + ], + "angle": 0, + "content": "Enforcing weighting monotonicity in \\( x \\) Revisiting the sigmoid loss, we derive its \\( x \\)-mse form, which turns out to conveniently be defined by a mirrored sigmoid (albeit multiplied by a constant factor):" + }, + { + "type": "equation", + "bbox": [ + 0.532, + 0.859, + 0.905, + 0.897 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\sigma \\left(b - \\lambda_ {t}\\right) \\| \\boldsymbol {\\epsilon} - \\hat {\\boldsymbol {\\epsilon}} \\| ^ {2} = \\sigma \\left(b - \\lambda_ {t}\\right) \\exp \\left(\\lambda_ {t}\\right) \\| \\boldsymbol {x} - \\hat {\\boldsymbol {x}} \\| ^ {2} \\tag {4} \\\\ = \\exp (b) \\sigma \\left(\\lambda_ {t} - b\\right) \\| \\boldsymbol {x} - \\hat {\\boldsymbol {x}} \\| ^ {2}. \\\\ \\end{array}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "18063" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.147, + 0.106, + 0.44, + 0.247 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.239, + 0.247, + 0.359, + 0.257 + ], + "angle": 0, + "content": "(a) Weighting over \\(\\epsilon\\)-mse" + }, + { + "type": "image", + "bbox": [ + 0.547, + 0.106, + 0.838, + 0.247 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.638, + 0.247, + 0.76, + 0.257 + ], + "angle": 0, + "content": "(b) Weighting over \\( \\pmb{x} \\)-mse" + }, + { + "type": "image_caption", + "bbox": [ + 0.128, + 0.269, + 0.868, + 0.283 + ], + "angle": 0, + "content": "Figure 3. Weighting of different losses in both \\(\\epsilon\\) and \\(x\\) space. Note that the weightings are scaled to have a maximum of one." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.292, + 0.483, + 0.473 + ], + "angle": 0, + "content": "This reveals that the sigmoid loss is not only monotonically decreasing in \\(\\epsilon\\)-mse, it is also monotonically increasing in \\(x\\)-mse. We claim that a reasonable weighting function \\(w(\\lambda_t)\\) should be monotonically increasing over \\(\\lambda_t\\). Observe that with less noise in \\(z_t\\), there is more information available and it should be strictly easier to predict the original data, meaning that the \\(x\\)-mse generally decreases for higher \\(\\lambda_t\\). This phenomenon is also observed in practice. If the loss weighting in \\(x\\)-mse also decreases, the model would be able to make worse predictions while obtaining less noisy (better) conditioning information and still achieve a lower loss. Simply put: Why put less weight on an easier problem?" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.473, + 0.484, + 0.581 + ], + "angle": 0, + "content": "Studying both the form of the sigmoid and EDM-monotonic weighting functions in Figure 3, we observe that both weighting functions are monotonically decreasing in \\(\\epsilon\\)-mse, but only the sigmoid weighting is monotonically increasing in \\(x\\)-mse. Therefore, we expect the sigmoid weighting to be a better candidate to scale to high resolution generation than the EDM-monotonic version." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.597, + 0.484, + 0.765 + ], + "angle": 0, + "content": "Loss shifting Recall that for a resolution that is two times higher in both height and width, an average (low pass) filter would see \\(\\mathrm{Var}[(x_{11} + x_{12} + x_{21} + x_{22}) / 4] = 1 / 4\\) of the noise, while signal strength stays the same [14, 25]. Hence, its effective logsnr is expected to increase by about \\(2\\log 2\\approx 1.386\\approx 1.5\\) when viewed through a low-pass filter. Suppose that the optimal setting for images of \\(128\\times 128\\) is a sigmoid weighting with \\(b\\approx 0\\), then one would expect the optimal setting for images of \\(512\\times 512\\) to be roughly \\(b = -3\\).1 Empirically we find that slightly more aggressive shifts perform even better." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.772, + 0.27, + 0.789 + ], + "angle": 0, + "content": "2.2. Flop heavy scaling" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.794, + 0.483, + 0.856 + ], + "angle": 0, + "content": "When one deals with a limited data budget, it can be difficult to balance scaling of models and regularization. Our small model version at \\(512 \\times 512\\) resolution starts out with \\(4 \\times 4\\) patching, and subsequently runs a U-ViT [14] on the \\(128 \\times\\)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.292, + 0.908, + 0.549 + ], + "angle": 0, + "content": "\\(128 \\times 48\\) input (assuming the raw input is of 3 channels). To make the model more expressive, one of the most common options is to double the number of channels for every layer, leading to \\(2 \\times\\) model size. However, our experiments show that this results in diminishing returns. We hypothesize that it is because the model suffers from over-fitting when trained with limited data budget such as ImageNet. Indeed, when we train using distribution augmentation [18], the best performance improves from 1.9 to 1.6. This however suggests another way to scale the model: Keep the architecture and parameter count almost the same and instead run on \\(2 \\times 2\\) patched input (so \\(256 \\times 256 \\times 12\\) for 512 inputs). Similar to \\(2 \\times\\) model size, reducing the patching size also increases FLOPS by roughly a factor of 4. However, we found this model performs even better to about 1.5 without the need of distribution augmentation. We call this model flop heavy, because it has a much higher flop over parameter ratio." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.55, + 0.909, + 0.777 + ], + "angle": 0, + "content": "Flop heavy scaling has another interesting property. Suppose we have a small model trained to fit input of size \\(256 \\times 256\\) with \\(2 \\times 2\\) patching. Using the flop heavy scaling at \\(512 \\times 512\\) resolution enables us to directly finetune from the small model trained at \\(256 \\times 256\\) with exactly the same amount of parameters, whereas its counterpart version at \\(512 \\times 512\\) resolution and with \\(4 \\times 4\\) patching needs to introduce additional parameters to handle the more input channels. This method of finetuning to higher resolutions has already been popularized in the context of text-to-image generation [32]. Here we find that flop heavy scaling is a more favorable way to scale on smaller datasets such as ImageNet. Of course, in principle one could also train a model with more channels and \\(4 \\times 4\\) patching on a lower resolution, but this is not done in practice to train 'small' models." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.787, + 0.907, + 0.817 + ], + "angle": 0, + "content": "2.3. Residual U-ViTs: Removing blockwise skipconnections" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.909, + 0.9 + ], + "angle": 0, + "content": "U-Nets require that sub-sampled levels (for example via average pooling) are bypassed by skip-connections. However, this makes the architecture very different from large-scale language models such as transformers, where these skip-connections are not used and residual connections suffice." + }, + { + "type": "page_footnote", + "bbox": [ + 0.09, + 0.863, + 0.483, + 0.901 + ], + "angle": 0, + "content": "This holds under the condition that the most important information is already contained by the \\(128 \\times 128\\) image, and this may for example not hold for images without global structure, e.g. that only contain textures." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18064" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.152 + ], + "angle": 0, + "content": "In this paper we use the U-ViT model of [14], which is a variant of U-Nets with self-attention layers (with convolutional layers) being replaced by transformer blocks. Hence, that part of the architecture is closer to standard transformers." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.153, + 0.485, + 0.259 + ], + "angle": 0, + "content": "In this paper, we further simplify model definition with Residual U-ViTs, see Figure 4. The blockwise skip-connections are removed, and in their place only a single skip-connection is used per downsampling operation. To be precise, let \\( f_{d}, f_{u} \\) denote a down and up level of a U-ViT and let \\( f_{m} \\) denote the middle stage. This two-stage Residual U-ViT can now be described by:" + }, + { + "type": "equation", + "bbox": [ + 0.09, + 0.272, + 0.484, + 0.31 + ], + "angle": 0, + "content": "\\[\nf (x) = f _ {u} \\left(\\mathrm {U} \\left(f _ {m} (\\mathrm {D} (h)) - \\mathrm {D} (h)\\right) + h\\right) \\text {w h e r e} h = f _ {d} (x) \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.31, + 0.484, + 0.446 + ], + "angle": 0, + "content": "where \\( D \\) is typically an average pooling and linear and \\( U \\) is typically a linear and nearest neighbour upsampling. Defining this network to preserve an identity function at initialization is now much easier: As long as \\( f_{d} \\), \\( f_{u} \\) and \\( f_{m} \\) are initialized to be identity, the entire \\( f \\) is also identity at initialization. The generalization to any-depth U-ViTs is also straightforward: we can define another two levels of U-ViTs in place of \\( f_{m} \\) recursively, where the nested U-ViTs can also be guaranteed to be initialized as an identity function." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.447, + 0.485, + 0.629 + ], + "angle": 0, + "content": "For small networks, it could be helpful to keep the blockwise skip-connections, and expect to incur a slight performance degradation from removing the blockwise skip-connections. Small networks have limited capacity, so limiting skip-connections may create bottleneck issues. Experimentally, this is confirmed for the small architectures, where we observe 0.2 FID performance degradation on ImageNet512 and 0.06 FID performance degradation on ImageNet256. Because of the simplified definition, we choose to accept this performance penalty for small models. Moreover, we find that larger models perform on par or even benefit from removing the blockwise skip-connections." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.63, + 0.486, + 0.901 + ], + "angle": 0, + "content": "The final advantage is memory consumption at evaluation. For example, a feature map at \\(256^{2}\\) resolution with 128 channels requires \\(256^{2} \\cdot 128 \\cdot 2 / 1024^{2} = 16\\mathrm{MB}\\). With 3 blocks per level, this requires holding an extra \\(3 \\cdot 16 = 48\\mathrm{MB}\\) per example for the skip-connections, whereas the Residual U-ViT only requires holding \\(h\\) and \\(D(h)\\) (the latter could be cheaply recomputed), so \\(16 + 8 = 24\\mathrm{MB}\\) per example while computing the lower levels. For a hypothetical super-resolution model at \\(4096^{2}\\) resolution, this leads to a difference between \\(3 \\cdot 4 = 12\\mathrm{GB}\\) and only \\(4 + 2 = 6\\mathrm{GB}\\). During training this advantage is usually negligible, because feature maps also need to be retained for the backward pass. A final advantage of Residual U-ViTs is that it is more straightforward to define asymmetric U-ViTs, i.e., with a different number of blocks in the downsampling (encoding) stage from the one in the upsampling (decoding) stage. In certain cases, moving blocks from the downsampling stage to the upsampling stage can increase performance." + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.089, + 0.906, + 0.193 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.203, + 0.908, + 0.245 + ], + "angle": 0, + "content": "Figure 4. Residual U-ViT with levelwise skip-connections (instead of blockwise). When the network blocks are initialized to be identity, the entire network is also an identity function at initialization." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.264, + 0.637, + 0.278 + ], + "angle": 0, + "content": "2.4. Distillation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.285, + 0.908, + 0.513 + ], + "angle": 0, + "content": "Sampling from a pixel-space diffusion model naively often requires more function evaluations than sampling from a latent-space diffusion model. For example, the popular Stable Diffusion model [32] is usually sampled using 20-50 sampling steps, while pixel-space models are sampled with \\(100+\\) steps [14]. To reduce the number of sampling steps, we could use either more advanced solvers or distillation. Recently distillation methods show that pixel-space models can be adapted into few-step samplers that maintain high sample quality [23, 29, 36, 44, 45, 47]. In this work, we use moment matching distillation [37] which is currently one of the best distillation methods available in the few step regime. We distill our pixel-space models for use with 8-16 sampling steps, which empirically we find maintains or even improves the sample quality of our models." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.525, + 0.655, + 0.541 + ], + "angle": 0, + "content": "3. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.55, + 0.908, + 0.611 + ], + "angle": 0, + "content": "The first end-to-end diffusion work that directly generated high resolution images was [4]. Given the poor performance of end-to-end models, two popular alternatives were introduced: cascaded diffusion [12] and latent diffusion [32]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.611, + 0.908, + 0.701 + ], + "angle": 0, + "content": " Cascaded diffusion approaches have been shown to scale to large datasets and models to generate high quality and high resolution imagery, albeit in multiple stages [12, 40, 42, 52], or with fine-tuning [43]. Although the focus of this paper has been single stage generation up to a resolution of \\(10^{24^2}\\), the loss weightings may also be used for cascaded models." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.702, + 0.909, + 0.838 + ], + "angle": 0, + "content": "Significant improvements were made to end-to-end diffusion by shifting noise schedules [3, 14] and subsequently loss weightings [25]. Other approaches such as multi-resolution losses (and possibly architecture adaptions) [6, 7, 14] added marginal gains. However, the current best known latent model EDM2 [20] outperforms these prior end-to-end approaches by a large margin. In contrast, our models are competitive with latent models, outperforming DiT-XL/2 [31] and only performing slightly worse than EDM2." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.91, + 0.901 + ], + "angle": 0, + "content": "Skip connections The U-Net architecture is a fully convolutional network with long-range skip-connections introduced for (medical) image segmentation [33, 39]. The main" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18065" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.087, + 0.493, + 0.273 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.281, + 0.493, + 0.338 + ], + "angle": 0, + "content": "Figure 5. Effect of weighting functions in the loss on the FID score of ImageNet512 over the training iterations. The proposed sigmoid weighting with \\( b = -3 \\) shows best performance consistent across all training iterations." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.35, + 0.486, + 0.395 + ], + "angle": 0, + "content": "idea of U-Net is to combine a contracting path (which encodes the image semantics) with long skip-connections at the same resolution to enable precise localization." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.396, + 0.486, + 0.593 + ], + "angle": 0, + "content": "With the introduction of ViTs [5] as an alternative for convolutional layers, the U-Net has been converted to U-ViTs using a similar network but replacing (some) ResNet blocks with ViTs, either with spatial down-sampling [14] or without [2]. Long-range skip-connections play an pivotal role in these networks, for example for fine-tuning controllable diffusion models [17] and for the stability and training speed [15, 30]. In these latter works, it is shown that multiplying the skip-connections with a (constant) scalar \\((< 1)\\) improves training stability and speed, because of the propagation of noise throughout the network. In our work, we reduce the number of skip-connections significantly and initialise the network such that it resembles an identify function." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.612, + 0.484, + 0.718 + ], + "angle": 0, + "content": "Flop heavy fine-tuning Fine-tuning on a higher resolution at large scale was effectively done in Stable Diffusion [32] to limit the training cost, which is a form of flop heavy scaling over different resolutions. In this paper we show that flop heavy scaling is also preferred from a regularization perspective, which can be done either by finetuning from a lower resolution, or by lowering the patching size." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.733, + 0.223, + 0.75 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.758, + 0.376, + 0.774 + ], + "angle": 0, + "content": "4.1. Sigmoid weighting versus others" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.486, + 0.903 + ], + "angle": 0, + "content": "Loss weighting In this experiment, we want to understand the effect of loss weighting on the final image quality of a model. For the noise schedule, we use an interpolated cosine schedule because it has relatively broad coverage over logsnr values. We compare the weightings corresponding to the original \\(\\epsilon\\)-mse [11] and the multiscale \\(\\epsilon\\)-mse [14], the EDM shifted weightings [19], including the original and monotonic versions [19, 26], and the proposed sigmoid weighting" + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.087, + 0.903, + 0.273 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.281, + 0.905, + 0.338 + ], + "angle": 0, + "content": "Figure 6. Effect of training noise schedules on the FID score of ImageNet512 over the training iterations. Fortunately, with our proposed sigmoid weighting \\((b = -3)\\), the exact choice of the noise schedules has little effect on the training process." + }, + { + "type": "table_caption", + "bbox": [ + 0.512, + 0.348, + 0.908, + 0.418 + ], + "angle": 0, + "content": "Table 1. Sigmoid weighting performance (left) with the flop heavy architecture. Larger resolutions require more negative bias. Power loss performance (right) on ImageNet512 with the small model. Increasing downsampling has a similar effect as decreasing bias, hence we prefer the more simple sigmoid weighting." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.429, + 0.897, + 0.564 + ], + "angle": 0, + "content": "
(a) Sigmoid weighting(b) Power loss weighting
biasresolutionbiasdownsample levels
12825651210240123
-5---1.82-42.192.583.625.11
-4--1.631.75-32.192.322.854.02
-3-1.471.482.32-22.292.152.343.13
-21.491.421.552.17-12.642.322.252.54
-11.351.381.75-03.222.742.072.22
01.301.461.94-1-2.692.292.16
11.261.56--2-3.262.532.19
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.58, + 0.909, + 0.703 + ], + "angle": 0, + "content": "with shift \\( b = -3 \\). As can be seen in Figure 5, the sigmoid weighting outperforms all other weightings, with EDM monotonic shifted being the closest alternative. Appendix C contains additional experiments, where we also evaluate the sigmoid weighting against the closest competitor without guidance. Here again sigmoid weighting is consistently better. We also evaluate a time-shifted variant, where the bias depends on the progress in training." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.911, + 0.902 + ], + "angle": 0, + "content": "Training efficiency of noise schedules Since the training noise schedule can be decoupled from the loss weighting and it only affects the training efficiency but not the final performance, one can measure the convergence speed over the training process caused by leveraging particular noise schedules. In this experiment we test the cosine schedule [11], the shifted cosine and the interpolated cosine schedules [14] (Figure 6). For fair comparison, we fix the sampling noise schedule as the cosine interpolated schedule. Interestingly, when the sigmoid weighting is used, the choice of noise schedule does not affect the training efficiency significantly. Arguably, the shifted schedule is a bit more sample effi" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18066" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.12, + 0.088, + 0.882, + 0.281 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.285, + 0.908, + 0.342 + ], + "angle": 0, + "content": "Figure 7. FID versus training cost on ImageNet512. Note that the EDM-2 training curves are without guidance, but the best performance with guidance intervals is reported with a separate marker, the figure source adapted has been adapted from EDM2 [20]. Following their assumptions, one training iteration is three times as expensive as evaluating the model. Our Flop Heavy (finetune) is trained by first training a small model on ImageNet256 for 400k iterations and then finetuning the same model on ImageNet512." + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.358, + 0.36, + 0.482 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.366, + 0.358, + 0.633, + 0.482 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.639, + 0.353, + 0.907, + 0.482 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.491, + 0.36, + 0.547 + ], + "angle": 0, + "content": "Figure 8. The effect of scaling either channels or tokens, compared to the small model variant. FLOPS between model size \\(2\\mathrm{x}\\) and flop heavy are similar." + }, + { + "type": "image_caption", + "bbox": [ + 0.364, + 0.491, + 0.633, + 0.547 + ], + "angle": 0, + "content": "Figure 9. The effect of Residual-UViTs (removing block skip connections). Although smaller models can be effected, larger models are unaffected by the simplification." + }, + { + "type": "image_caption", + "bbox": [ + 0.639, + 0.491, + 0.908, + 0.547 + ], + "angle": 0, + "content": "Figure 10. Changing the encoder decoder architecture. Decoder blocks are more important than encoder blocks, and symmetric architectures have decent performance." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.574, + 0.483, + 0.709 + ], + "angle": 0, + "content": "cient initially but all schedules perform similarly towards the end. This suggests that there is a lot of transfer across different logsnr values. We also tried the EDM training noise schedule with the sigma power sampling noise schedule [19] (not plotted), with both shifted or the original version, but its narrower logsnr sampling schedule led to performance degradation, achieving about 9 FID at best. This suggests that for high resolution, a wide logsnr range is crucial to cover the important noise levels." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.73, + 0.484, + 0.837 + ], + "angle": 0, + "content": "Bias over resolutions In Table 1a, we show the effect of the bias term \\( b \\) on the final performance when trained using a flop heavy model. All models use the same guidance and similar guidance intervals, where the lower bound of the guidance interval has been shifted according to resolution (again approximately 1.5 per 2x resolution change), see Appendix B for more details." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Sigmoid weighting versus the power loss In literature, high resolution generation can greatly benefit from having a multiscale loss [6, 7, 14]. To that end we create a multiscale" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.574, + 0.907, + 0.664 + ], + "angle": 0, + "content": "variant that we name the power loss. In addition to the sigmoid weighting that effectively upweights low frequency components of the data and downweights high frequency components, the power loss aims to amplify the loss of low sub-band (smooth) signals in the spatial domain via Haar wavelets. For details on this loss see Appendix B.2." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.666, + 0.909, + 0.803 + ], + "angle": 0, + "content": "In Table 1b we see that the bias should be adjusted based on the number of Haar transform downsampling levels. Despite being competitive, the performance of power loss is roughly on par with the one with sigmoid weighting but without multiscaling (i.e. downsample levels \\(= 0\\)), if we tune the bias of the sigmoid weighting more aggressively. Interestingly, lowering the bias has almost the same effect as increasing the scale. Therefore, building the multiscale variant is not necessary with the sigmoid weighting." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.817, + 0.857, + 0.833 + ], + "angle": 0, + "content": "4.2. Scaling channels vs tokens on ImageNet" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.908, + 0.901 + ], + "angle": 0, + "content": "In this section we show the difference between scaling a model 4x FLOPS, either via channels or by creating larger resolution feature maps (or put differently, feature maps with a higher token count). Note that our small model on" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18067" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.091, + 0.482, + 0.228 + ], + "angle": 0, + "content": "ImageNet512 is patched \\(4 \\times 4\\). This model is scaled either by increasing channels by two (named model size \\(2x\\)), or by reducing patching to \\(2 \\times 2\\) (named flop heavy, because FLOPS over params is much higher). Indeed as can be seen in Figure 8, the flop heavy model performs much better than the model size \\(2x\\), when trained in a standard way on ImageNet512. Note that unlike most experiments, in these experiments it was necessary to use very different optimal guidance settings, more details in Appendix B." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.23, + 0.483, + 0.321 + ], + "angle": 0, + "content": "We confirm this is an issue related to overfitting, by training the model size 2x with distribution augmentation [18] the model size 2x variant performs much better, with an FID of 1.6 instead of 1.9 (latter in Figure 8). Still, without any augmentation beyond the standard horizontal flips, the flop heavy model achieves a score of 1.54." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.34, + 0.376, + 0.356 + ], + "angle": 0, + "content": "4.3. Removing block skip connection" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.364, + 0.483, + 0.501 + ], + "angle": 0, + "content": "In this section we study the difference between the residual skip-connections and the blockwise skip-connections. As hypothesized, the small model performance drops a bit because of this. For ImageNet512 a deficit of about 0.2 FID (in Figure 9) and for Imagenet256 a deficit of about 0.05 FID. That the task of fitting ImageNet256 is less affected by removing skip-connections can be explained by the task being easier: The task of fitting ImageNet256 is a subset of fitting ImageNet512." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.504, + 0.483, + 0.655 + ], + "angle": 0, + "content": "For larger models, removing blockwise skip-connections in favor of a single residual connection does not degrade performance, and may even lead to slightly increased performance (see Figure 9). This is somewhat unexpected, and the phenomenon persists with different guidance settings as well. The FID curve reminds of the model size \\(2\\mathrm{x}\\) curve, which at first improves but flattens out over iterations. Therefore our best hypothesis is that the block skip (flop) baseline may have similar problems such as overfitting. Regardless, it shows that Residual U-ViTs scale well." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Asymmetric U-ViT Since our U-ViT models encompass only a single skip connection for each level, we can use a different number of ResNet or ViT blocks per level. This allows to move compute from the encoder stage (the down sampling half of the model) to the decoder stage (the upsampling half of the model). From Figure 10 we observe that (a) for a fixed compute budget, it is better to add more parameters to the decoder, the encoder heavy model (E5-D1) is significantly worse than the baseline (E3-D3) and the decoder heavy model (E1-D5); (b) there needs to be some balance between the encoder and the decoder, E1-D7 works less good than E3-D5. (c) by a slightly heavier model (E3-D7, only \\(10\\%\\) more parameters than our small (E3-D3) model, the performance significantly increases." + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.09, + 0.905, + 0.145 + ], + "angle": 0, + "content": "Table 2. Literature Comparison on ImageNet on FID. Note that even though certain NFEs are high (e.g. 4100 for CDM) typically these suffer minor performance degradation for lower NFEs and should not be taken too literally." + }, + { + "type": "table", + "bbox": [ + 0.536, + 0.147, + 0.885, + 0.869 + ], + "angle": 0, + "content": "
MethodNFEFID
Imagenet 128 x 128
RIN [16]10002.75
SD, U-ViT-L [14]5121.94
StyleGAN-XL [38]11.81
VDM+++ [25]2501.75
DisCo-Diff [46]4141.73
PaGoDA [22]11.48
SiD2, Flop Heavy (ours)5121.26
Imagenet 256 x 256
ADM-G [4]2504.59
ADM-U [4]2503.94
CDM [12]41004.88
RIN [3, 16]10003.52
SD, U-ViT-L [14]5122.44
MDM [7]2503.51
StyleGAN-XL [38]12.30
DiT-XL/2 [31]2502.27
VDM++ [25]2502.12
RDM [42]3561.89
PaGoDA [22]11.56
SiD2, small (ours)5121.72
SiD2, Flop Heavy (ours)5121.38
Imagenet 512 x 512
ADM-G [4]2507.72
RIN [3, 16]10003.95
ADM-U [4]2503.85
LDM-4-G [32]2503.60
SD, U-ViT-L [14]5123.02
DiffiT [9]2502.67
StyleGAN-XL [38]12.40
DiT-XL/2 [31]2503.04
VDM++ [25]2502.65
MAGVIT-v2 [51]641.91
PaGoDA [22]11.80
EDM2-S (interval) [27]631.68
EDM2-XXL (interval) [27]631.40
SiD2, small (ours)5122.19
SiD2, Flop Heavy (ours)5121.48
SiD2, Flop Heavy distilled (ours)161.50
Imagenet 1024 x 1024
RIN [16]10008.72
SiD2, Flop Heavy (ours)5121.75
" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "18068" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.486, + 0.133 + ], + "angle": 0, + "content": "Table 3. Zero-shot text-to-image performance evaluated on MSCOCO in \\(\\mathrm{FID}_{30\\mathrm{k}}\\). Our distilled model, evaluated on the more difficult 512 setting, outperforming other works." + }, + { + "type": "table", + "bbox": [ + 0.134, + 0.143, + 0.444, + 0.314 + ], + "angle": 0, + "content": "
MethodSizeNFEFID30k
DeepFloyd [40]2566.7
eDiff-I [1]2566.9
Parti [49]2567.2
Imagen [35]2567.3
simple diffusion [14]2568.3
simple diffusion [14]5122569.6
SDv1.5 (eval DMD2) [32]5122007.2
DMD2 [48]51248.3
SiD2 (ours, bias=-3)5122568.1
SiD2 16-step distilled (ours)512166.7
" + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.33, + 0.303, + 0.346 + ], + "angle": 0, + "content": "4.4. Literature comparison" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.351, + 0.486, + 0.489 + ], + "angle": 0, + "content": "Here we compare different models from the literature. In the first section end-to-end diffusion models are compared, and in the second section we compare between all types of generative models. In Table 2 it can be seen that our SiD2 is very competitive with existing generative models: On ImageNet128 and ImageNet256 it achieves state-of-the-art performance, on ImageNet512 it is a little worse but still competitive with the best latent diffusion model and outperforms all other models, pixel-based and latent." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.506, + 0.487, + 0.765 + ], + "angle": 0, + "content": "Comparing diffusion models on compute The training cost SiD2 is another helpful way to see how the model scales, as depicted in Figure 7). Importantly, the performance of our pixel-based approach outperforms other significantly. Observe for example the performance of our small variant (using no guidance interval to keep this comparison fair), which roughly the same performance as the VDM++ while only using a third of the training cost, and a fourth of the FLOPS in a single forward pass (see Figure 16 in Appendix C). It is also competitive with the best currently known latent diffusion approach EDM2 [20], but in fairness EDM2 scales a little bit better. These models attain even better performance for a smaller training cost, and a similar story holds for forward pass complexity. End-to-end diffusion can work much better than previously demonstrated, but it is still somewhat outperformed by latent diffusion on quality versus compute comparisons for both sampling and training." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.486, + 0.903 + ], + "angle": 0, + "content": "Text to image generation Finally, we show that our method can be used for end-to-end training of text-to-image models. Here, evaluation is more difficult as MSCOCO zero-shot performance depends a lot on the dataset used and is known to not necessarily correspond to actual perceived performance. Nevertheless, it is important that the model achieves performance in the right ballpark. Indeed, when we train SiD2 it performs comparable to other models in the" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.089, + 0.909, + 0.133 + ], + "angle": 0, + "content": "Table 4. Conditional Video generation performance on Kinetics600 in FVD (resized to 64). For video generation the loss specification is important to obtain good sample quality (see ablations)." + }, + { + "type": "table", + "bbox": [ + 0.597, + 0.143, + 0.828, + 0.347 + ], + "angle": 0, + "content": "
MethodFVD50k
pixel diffusion
Video Diffusion Models [13]16.6
RIN [16]10.8
Rolling Diffusion [34]5.2
latent diffusion
MAGVIT [50]9.9
MAGVITv2 [51]4.3
W.A.L.T [8]3.3
SiD2 (ours, sigmoid bias = 0)2.3
ablations
eps-mse5.7
sigmoid bias = +23.2
sigmoid bias = -22.5
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.362, + 0.909, + 0.453 + ], + "angle": 0, + "content": "literature on MSCOCO \\(\\mathrm{FID}_{30\\mathrm{K}}\\), and detailed samples have only a small number of artifacts (Figure 2). Furthermore, the moment matching distilled version using [37] outperforms the baseline model as previously observed. This model also outperforms all other zero-shot approaches at 512 resolution as shown in Table 3." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.47, + 0.909, + 0.669 + ], + "angle": 0, + "content": "Video Generation To test whether our approach transfers to video, we test our model on Kinetics600 following [13], by conditioning on 5 frames and generating (sometimes referred to as predicting) the subsequent 11 frames. We then compute the FVD between the training set and the generated samples both on the subsequent 11 frames. Even though our method generates at \\(128 \\times 128\\), we follow [8] and resize samples to have resolution \\(64 \\times 64\\). Our model achieves SOTA performance and it turns out that tuning the loss is very important to achieve this as shown in Table 4. Note that the latent approaches ([8, 50, 51]) use much less training compute than our models in pixel-space, which we discuss in more detail in the Appendix B." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.68, + 0.628, + 0.695 + ], + "angle": 0, + "content": "5. Discussion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.704, + 0.911, + 0.903 + ], + "angle": 0, + "content": "In this paper we present Simpler Diffusion (SiD2) and show how one can train end-to-end diffusion and achieve competitive performance with latent models: (1) use the sigmoid loss, (2) with a simpler U-ViT architecture, and (3) favour a flop heavy model design. As a result, our method outperforms existing end-to-end models by a large margin. Nevertheless, at this moment latent diffusion methods still seem to have slightly better scaling properties and it remains to be seen whether this gap can be closed any further. Importantly, for many applications training a separate autoencoder may not be either practical or desired. Therefore, we believe this paper is an important contribution that allows end-to-end modelling directly in the pixel-space." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "18069" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.486, + 0.184 + ], + "angle": 0, + "content": "[1] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Ji-aming Song, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, Bryan Catanzaro, Tero Karras, and Ming-Yu Liu. ediffi: Text-to-image diffusion models with an ensemble of expert denoisers. CoRR, abs/2211.01324, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.187, + 0.484, + 0.227 + ], + "angle": 0, + "content": "[2] Fan Bao, Shen Nie, Kaiwen Xue, Yue Cao, Chongxuan Li, Hang Su, and Jun Zhu. All are worth words: A vit backbone for diffusion models. In CVPR, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.23, + 0.484, + 0.257 + ], + "angle": 0, + "content": "[3] Ting Chen. On the importance of noise scheduling for diffusion models. arxiv, 2023. 1, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.259, + 0.484, + 0.287 + ], + "angle": 0, + "content": "[4] Prafulla Dhariwal and Alex Nichol. Diffusion models beat gans on image synthesis. CoRR, abs/2105.05233, 2021. 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.289, + 0.484, + 0.399 + ], + "angle": 0, + "content": "[5] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In 9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7, 2021. OpenReview.net, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.402, + 0.484, + 0.457 + ], + "angle": 0, + "content": "[6] Jiatao Gu, Shuangfei Zhai, Yizhe Zhang, Miguel Angel Bautista, and Josh M. Susskind. f-dm: A multi-stage diffusion model via progressive signal transformation. CoRR, abs/2210.04955, 2022. 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.459, + 0.484, + 0.5 + ], + "angle": 0, + "content": "[7] Jiatao Gu, Shuangfei Zhai, Yizhe Zhang, Josh M. Susskind, and Navdeep Jaitly. Matryoshka diffusion models. CoRR, abs/2310.15111, 2023. 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.502, + 0.484, + 0.585 + ], + "angle": 0, + "content": "[8] Agrim Gupta, Lijun Yu, Kihyuk Sohn, Xiuye Gu, Meera Hahn, Fei-Fei Li, Irfan Essa, Lu Jiang, and José Lezama. Photorealistic video generation with diffusion models. In Computer Vision - ECCV 2024 - 18th European Conference, Milan, Italy, September 29-October 4, 2024, Proceedings, Part LXXIX, pages 393-411. Springer, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.588, + 0.484, + 0.628 + ], + "angle": 0, + "content": "[9] Ali Hatamizadeh, Jiaming Song, Guilin Liu, Jan Kautz, and Arash Vahdat. Diffit: Diffusion vision transformers for image generation. CoRR, abs/2312.02139, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.631, + 0.484, + 0.658 + ], + "angle": 0, + "content": "[10] Jonathan Heek, Emiel Hoogeboom, and Tim Salimans. Multistep consistency models. CoRR, abs/2403.06807, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.661, + 0.484, + 0.715 + ], + "angle": 0, + "content": "[11] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In Advances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, NeurIPS, 2020. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.718, + 0.484, + 0.771 + ], + "angle": 0, + "content": "[12] Jonathan Ho, Chitwan Sahara, William Chan, David J. Fleet, Mohammad Norouzi, and Tim Salimans. Cascaded diffusion models for high fidelity image generation. J. Mach. Learn. Res., 23:47:1-47:33, 2022. 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.775, + 0.484, + 0.843 + ], + "angle": 0, + "content": "[13] Jonathan Ho, Tim Salimans, Alexey A. Gritsenko, William Chan, Mohammad Norouzi, and David J. Fleet. Video diffusion models. In Advances in Neural Information Processing Systems 35: Annual Conference on Neural Information Processing Systems 2022, NeurIPS, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.484, + 0.9 + ], + "angle": 0, + "content": "[14] Emiel Hoogeboom, Jonathan Heek, and Tim Salimans. simple diffusion: End-to-end diffusion for high resolution images. In International Conference on Machine Learning, ICML, pages 13213-13232. PMLR, 2023. 1, 3, 4, 5, 6, 7, 8" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.486, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.146 + ], + "angle": 0, + "content": "[15] Zhongzhan Huang, Zhou Pan, Shuicheng Yan, and Liang Lin. Scalelong: Towards more stable training of diffusion model via scaling network long skip connection. In NeurIPS, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.149, + 0.907, + 0.188 + ], + "angle": 0, + "content": "[16] Allan Jabri, David J. Fleet, and Ting Chen. Scalable adaptive computation for iterative generation. CoRR, abs/2212.11972, 2022. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.191, + 0.907, + 0.245 + ], + "angle": 0, + "content": "[17] Zeyinzi Jiang, Chaojie Mao, Yulin Pan, Zhen Han, and Jingfeng Zhang. Scedit: Efficient and controllable image diffusion generation via skip connection editing. Technical Report 2312.11392, arXiv, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.247, + 0.908, + 0.314 + ], + "angle": 0, + "content": "[18] Heewoo Jun, Rewon Child, Mark Chen, John Schulman, Aditya Ramesh, Alec Radford, and Ilya Sutskever. Distribution augmentation for generative modeling. In Proceedings of the 37th International Conference on Machine Learning, ICML, 2020. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.317, + 0.908, + 0.384 + ], + "angle": 0, + "content": "[19] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. In Advances in Neural Information Processing Systems 35: Annual Conference on Neural Information Processing Systems 2022, NeurIPS, 2022. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.386, + 0.908, + 0.44 + ], + "angle": 0, + "content": "[20] Tero Karras, Miika Aittala, Jaakko Lehtinen, Janne Hellsten, Timo Aila, and Samuli Laine. Analyzing and improving the training dynamics of diffusion models. CoRR, abs/2312.02696, 2023. 1, 4, 6, 8, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.442, + 0.907, + 0.495 + ], + "angle": 0, + "content": "[21] Tero Karras, Miika Aittala, Tuomas Kynkänniemi, Jaakko Lehtinen, Timo Aila, and Samuli Laine. Guiding a diffusion model with a bad version of itself. CoRR, abs/2406.02507, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.498, + 0.908, + 0.565 + ], + "angle": 0, + "content": "[22] Dongjun Kim, Chieh-Hsin Lai, Wei-Hsiang Liao, Yuhta Takida, Naoki Murata, Toshimitsu Uesaka, Yuki Mitsufuji, and Stefano Ermon. Pagoda: Progressive growing of a one-step generator from a low-resolution diffusion teacher. CoRR, abs/2405.14822, 2024. 7, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.568, + 0.908, + 0.648 + ], + "angle": 0, + "content": "[23] Dongjun Kim, Chieh-Hsin Lai, Wei-Hsiang Liao, Naoki Murata, Yuhta Takida, Toshimitsu Uesaka, Yutong He, Yuki Mitsufuji, and Stefano Ermon. Consistency trajectory models: Learning probability flow ODE trajectory of diffusion. In The Twelfth International Conference on Learning Representations, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.651, + 0.907, + 0.692 + ], + "angle": 0, + "content": "[24] Durk P Kingma and Prafulla Dhariwal. Glow: Generative flow with invertible 1x1 convolutions. Advances in neural information processing systems, 31, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.693, + 0.908, + 0.733 + ], + "angle": 0, + "content": "[25] Diederik P. Kingma and Ruiqi Gao. Understanding the diffusion objective as a weighted integral of ellbods. CoRR, abs/2303.00848, 2023. 1, 2, 3, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.735, + 0.907, + 0.774 + ], + "angle": 0, + "content": "[26] Diederik P. Kingma, Tim Salimans, Ben Poole, and Jonathan Ho. Variational diffusion models. CoRR, abs/2107.00630, 2021. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.777, + 0.907, + 0.83 + ], + "angle": 0, + "content": "[27] Tuomas Kynkänniemi, Miika Aittala, Tero Karras, Samuli Laine, Timo Aila, and Jaakko Lehtinen. Applying guidance in a limited interval improves sample and distribution quality in diffusion models. CoRR, abs/2404.07724, 2024. 7, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.832, + 0.908, + 0.901 + ], + "angle": 0, + "content": "[28] Yaron Lipman, Ricky T. Q. Chen, Heli Ben-Hamu, Maximilian Nickel, and Matthew Le. Flow matching for generative modeling. In The Eleventh International Conference on Learning Representations, ICLR 2023, Kigali, Rwanda. OpenReview.net, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18070" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.161 + ], + "angle": 0, + "content": "[29] Weijian Luo, Tianyang Hu, Shifeng Zhang, Jiacheng Sun, Zhenguo Li, and Zhihua Zhang. Diff-instruct: A universal approach for transferring knowledge from pre-trained diffusion models. Advances in Neural Information Processing Systems, 36, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.162, + 0.485, + 0.232 + ], + "angle": 0, + "content": "[30] Jiajun Ma, Shuchen Xue, Tianyang Hu, Wenjia Wang, Zhaoqiang Liu, Zhenguo Li, Zhi-Ming Ma, and Kenji Kawaguchi. The surprising effectiveness of skip-tuning in diffusion sampling. In Proceedings of the 41st International Conference on Machine Learning, pages 34053-34074, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.233, + 0.483, + 0.259 + ], + "angle": 0, + "content": "[31] William Peebles and Saining Xie. Scalable diffusion models with transformers. CoRR, abs/2212.09748, 2022. 4, 7, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.261, + 0.485, + 0.342 + ], + "angle": 0, + "content": "[32] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 10674-10685. IEEE, 2022. 1, 3, 4, 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.344, + 0.485, + 0.385 + ], + "angle": 0, + "content": "[33] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. Technical report, ArXiV, 2015. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.386, + 0.485, + 0.44 + ], + "angle": 0, + "content": "[34] David Ruhe, Jonathan Heek, Tim Salimans, and Emiel Hoogeboom. Rolling diffusion models. In *Forty-first International Conference on Machine Learning*, ICML 2024, Vienna, Austria, July 21-27, 2024. OpenReview.net, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.442, + 0.486, + 0.536 + ], + "angle": 0, + "content": "[35] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S. Sara Mahdavi, Rapha Gontijo Lopes, Tim Salimans, Jonathan Ho, David J. Fleet, and Mohammad Norouzi. Photorealistic text-to-image diffusion models with deep language understanding. CoRR, abs/2205.11487, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.538, + 0.485, + 0.594 + ], + "angle": 0, + "content": "[36] Tim Salimans and Jonathan Ho. Progressive distillation for fast sampling of diffusion models. In *The Tenth International Conference on Learning Representations*, ICLR. OpenReview.net, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.595, + 0.485, + 0.649 + ], + "angle": 0, + "content": "[37] Tim Salimans, Thomas Mensink, Jonathan Heek, and Emiel Hoogeboom. Multistep distillation of diffusion models via moment matching. arXiv preprint arXiv:2406.04103, 2024. 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.65, + 0.485, + 0.706 + ], + "angle": 0, + "content": "[38] Axel Sauer, Katja Schwarz, and Andreas Geiger. Stylegan-xl: Scaling stylegan to large diverse datasets. In SIGGRAPH '22: Special Interest Group on Computer Graphics and Interactive Techniques Conference, pages 49:1-49:10. ACM, 2022. 7, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.707, + 0.485, + 0.761 + ], + "angle": 0, + "content": "[39] Evan Shelhamer, Jonathan Long, and Trevor Darrell. Fully convolutional networks for semantic segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 39(4):640-651, 2016. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.762, + 0.485, + 0.817 + ], + "angle": 0, + "content": "[40] Alex Shonenkov, Misha Konstantinov, Daria Bakshandaeva, Christoph Schuhmann, Ksenia Ivanova, and Nadija Klokova. Deepfloyd if: A text-to-image model with a high degree of photorealism and language understanding, 2023. 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.818, + 0.486, + 0.9 + ], + "angle": 0, + "content": "[41] Yang Song, Jascha Sohl-Dickstein, Diederik P. Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In 9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7, 2021. OpenReview.net, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.908, + 0.162 + ], + "angle": 0, + "content": "[42] Jiayan Teng, Wendi Zheng, Ming Ding, Wenyi Hong, Jianqiao Wangni, Zhuoyi Yang, and Jie Tang. Relay diffusion: Unifying diffusion process across resolutions for image synthesis. In The Twelfth International Conference on Learning Representations, ICLR. OpenReview.net, 2024. 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.163, + 0.907, + 0.219 + ], + "angle": 0, + "content": "[43] Haoning Wu, Shaocheng Shen, Qiang Hu, Xiaoyun Zhang, Ya Zhang, and Yanfeng Wang. Megafusion: Extend diffusion models towards higher-resolution image generation without further tuning. CoRR, abs/2408.11001, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.22, + 0.908, + 0.275 + ], + "angle": 0, + "content": "[44] Zhisheng Xiao, Karsten Kreis, and Arash Vahdat. Tackling the generative learning trilemma with denoising diffusion gans. In International Conference on Learning Representations, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.277, + 0.907, + 0.319 + ], + "angle": 0, + "content": "[45] Yanwu Xu, Yang Zhao, Zhisheng Xiao, and Tingbo Hou. Ufogen: You forward once large scale text-to-image generation via diffusion gans. arXiv preprint arXiv:2311.09257, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.32, + 0.908, + 0.388 + ], + "angle": 0, + "content": "[46] Yilun Xu, Gabriele Corso, Tommi S. Jaakkola, Arash Vahdat, and Karsten Kreis. Disco-diff: Enhancing continuous diffusion models with discrete latents. In *Forty-first International Conference on Machine Learning*, ICML 2024, Vienna, Austria, July 21-27, 2024. OpenReview.net, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.39, + 0.907, + 0.445 + ], + "angle": 0, + "content": "[47] Tianwei Yin, Michael Gharbi, Richard Zhang, Eli Shechtman, Fredo Durand, William T Freeman, and Taesung Park. One-step diffusion with distribution matching distillation. arXiv preprint arXiv:2311.18828, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.446, + 0.908, + 0.502 + ], + "angle": 0, + "content": "[48] Tianwei Yin, Michael Gharbi, Taesung Park, Richard Zhang, Eli Shechtman, Fredo Durand, and William T. Freeman. Improved distribution matching distillation for fast image synthesis. In NeurIPS, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.503, + 0.908, + 0.586 + ], + "angle": 0, + "content": "[49] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, Ben Hutchinson, Wei Han, Zarana Parekh, Xin Li, Han Zhang, Jason Baldridge, and Yonghui Wu. Scaling autoregressive models for content-rich text-to-image generation. CoRR, abs/2206.10789, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.587, + 0.908, + 0.655 + ], + "angle": 0, + "content": "[50] Lijun Yu, Yong Cheng, Kihyuk Sohn, José Lezama, Han Zhang, Huiwen Chang, Alexander G. Hauptmann, Ming-Hsuan Yang, Yuan Hao, Irfan Essa, and Lu Jiang. MAGVIT: masked generative video transformer. CoRR, abs/2212.05199, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.657, + 0.908, + 0.755 + ], + "angle": 0, + "content": "[51] Lijun Yu, José Lezama, Nitesh Bharadwaj Gundavarapu, Luca Versari, Kihyuk Sohn, David Minnen, Yong Cheng, Agrim Gupta, Xiuye Gu, Alexander G. Hauptmann, Boqing Gong, Ming-Hsuan Yang, Irfan Essa, David A. Ross, and Lu Jiang. Language model beats diffusion - tokenizer is key to visual generation. In The Twelfth International Conference on Learning Representations, ICLR. OpenReview.net, 2024. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.756, + 0.908, + 0.826 + ], + "angle": 0, + "content": "[52] Wendi Zheng, Jiayan Teng, Zhuoyi Yang, Weihan Wang, Jidong Chen, Xiaotao Gu, Yuxiao Dong, Ming Ding, and Jie Tang. Cogview3: Finer and faster text-to-image generation via relay diffusion. In Computer Vision - ECCV Proceedings, Part LXXVII, pages 1-22. Springer, 2024. 4" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.908, + 0.826 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.956 + ], + "angle": 0, + "content": "18071" + } + ] +] \ No newline at end of file diff --git a/2025/Simpler Diffusion_ 1.5 FID on ImageNet512 with Pixel-space Diffusion/3cf16242-5862-4aa6-864f-5aba8865f5f9_origin.pdf b/2025/Simpler Diffusion_ 1.5 FID on ImageNet512 with Pixel-space Diffusion/3cf16242-5862-4aa6-864f-5aba8865f5f9_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..6136fbc9cae262e154cfa2a84f2be7fd475f1683 --- /dev/null +++ b/2025/Simpler Diffusion_ 1.5 FID on ImageNet512 with Pixel-space Diffusion/3cf16242-5862-4aa6-864f-5aba8865f5f9_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f634b9ca3f235e5f066d037f7b49bd186045676415ec12d12911315702070c04 +size 1126235 diff --git a/2025/Simpler Diffusion_ 1.5 FID on ImageNet512 with Pixel-space Diffusion/full.md b/2025/Simpler Diffusion_ 1.5 FID on ImageNet512 with Pixel-space Diffusion/full.md new file mode 100644 index 0000000000000000000000000000000000000000..545ad9a6817a24899c1c4801448caf09d1320f61 --- /dev/null +++ b/2025/Simpler Diffusion_ 1.5 FID on ImageNet512 with Pixel-space Diffusion/full.md @@ -0,0 +1,313 @@ +# Simpler Diffusion: 1.5 FID on ImageNet512 with pixel-space diffusion + +Emiel Hoogeboom + +Google Deepmind + +emielh@google.com + +Thomas Mensink + +Google Deepmind + +mensink@google.com + +Jonathan Heek + +Google Deepmind + +jheek@google.com + +Kay Lamarigts + +Google Research + +kaylamerigts@google.com + +Ruiqi Gao + +Google Deepmind + +ruiqig@google.com + +Tim Salimans + +Google Deepmind + +salimans@google.com + +# Abstract + +Latent diffusion models have become the popular choice for scaling up diffusion models for high resolution image synthesis. Compared to pixel-space models that are trained end-to-end, latent models are perceived to be more efficient and to produce higher image quality at high resolution. Here we challenge these notions, and show that pixel-space models can be very competitive to latent models both in quality and efficiency, achieving 1.5 FID on ImageNet512 and new SOTA results on ImageNet128, ImageNet256 and Kinetics600. + +We present a simple recipe for scaling end-to-end pixel-space diffusion models to high resolutions. 1: Use the sigmoid loss-weighting [25] with our prescribed hyperparameters. 2: Use our simplified memory-efficient architecture with fewer skip-connections. 3: Scale the model to favor processing the image at a high resolution with fewer parameters, rather than using more parameters at a lower resolution. Combining these with guidance intervals, we obtain a family of pixel-space diffusion models we call Simpler Diffusion (SiD2). + +# 1. Introduction + +Diffusion models have been tremendously successful in domains such as image, video and audio generation. For high resolution data, latent diffusion models [32] have been particularly successful: Through encoding the data in a lower dimensional space, these models sidestep the difficulties faced by more traditional pixel-space diffusion models in scaling up to high resolution. This makes latent diffusion models also scale better in training compute than other approaches, as shown by [20]. + +However, latent models have their own challenges: It can be difficult to train the autoencoders that produce the latents, + +![](images/8b5a7ae5161b82f15e43113268a9298406651cafd4317c76bb573e36806124bf.jpg) +Figure 1. End-to-end pixel diffusion performance on ImageNet512. + +which in practice requires careful tuning of regularization hyperparameters [32]. Even when trained well, the decoder of a latent diffusion model can produce artifacts or otherwise imperfect reconstructions, especially when it is applied to latents that are out of distribution compared to the training set. A single-stage pixel-space diffusion model does not have these issues, and so we ask the question: Can pixel-space diffusion models scale as well as latent models with the right architecture and hyperparameters? + +Over the years, end-to-end pixel-space diffusion has improved considerably (Figure 1). The main problem was that the standard schedule and loss-weighting [11] work well for small resolution images, like $32^2$ or $64^2$ , but not at higher resolution [3, 14]. It was further shown [25] that it can be helpful to disentangle loss weighting and time sampling for improved control over both, leading to better performance at higher resolution. However, in previous works pixels-space models still did not match the best latent models on standard benchmarks like ImageNet512. + +The main contribution of this paper is to improve pixel-space diffusion on high resolution images while keeping as close to the original formulation of diffusion models as possible. Specifically, we analyze and tune the sigmoid loss weighting from [25], and show why it should be preferred over their favored EDM monotonic. We further propose architectural improvements that lead to even better perfor + +![](images/b18a0f4d018046d9ab7155f067e3df4b7341ad15d635e2a00a6cbc6332bde9c2.jpg) +Figure 2. Some text-to-image samples of $512 \times 512$ that have been generated by SiD2. + +![](images/4673782a530a2ffca3a5a3a1f6c3ddfa3d65cdd0850dc8041f8a3798bc066625.jpg) + +![](images/9b3e67cdffaed8395cf0f73b870516817a0a09d2e7eec74fb72d26e6b38056ce.jpg) + +![](images/2e50436e0cea30003bd97c6bc97008729d7dedcd626988278a1604a13b1e4d6b.jpg) + +mance, namely flop heavy scaling and removing block skip-connections (Residual U-ViTs). Most notably, we improve the state-of-the-art of pixel-space diffusion from 2.65 to 1.48 on ImageNet512 (see also Figure 1). Further, we also achieve overall state-of-the-art on ImageNet128, ImageNet256 and Kinetics600 without heavy tuning. + +# 1.1. Background + +A diffusion process gradually destroys data over a conceptual time $t$ , typically with Gaussian noise as follows: + +$$ +\boldsymbol {z} _ {t} = \alpha_ {t} \boldsymbol {x} + \sigma_ {t} \boldsymbol {\epsilon}, \text {w h e r e} \boldsymbol {\epsilon} \sim \mathcal {N} (0, \mathbf {I}), \tag {1} +$$ + +which defines a distribution $q(\pmb{z}_t|\pmb{x})$ such that $\pmb{z}_0$ is close to clean data and $\pmb{z}_1$ is close to Gaussian noise. In this paper we assume a variance-preserving [41] schedule, which is defined as $\alpha_{t} = \sqrt{\sigma(\lambda_{t})}$ and $\sigma_t = \sqrt{\sigma(-\lambda_t)}$ , with $\lambda_{t} = \log \left(\alpha_{t}^{2} / \sigma_{t}^{2}\right)$ being the log signal-to-noise ratio, which monotonically decreases with time as $t\to 1$ , and $\sigma (\cdot)$ being the sigmoid function. It can be shown that other types of schedules such as the variance-exploding schedule or the flow matching schedule [28] can always be converted to a variance-preserving schedule with a linear scaling of $\pmb{z}_{t}$ [25]. The key defining property of a noise schedule is the log signal-to-noise ratio $\lambda_{t}$ . A neural network is used to predict $\hat{\pmb{x}} = \hat{\pmb{x}}_{\theta}(\pmb{z}_{t},t)$ that effectively estimates $\mathbb{E}[\pmb {x}|\pmb {z}_t]$ , which can be done by minimizing a weighted mean squared error loss: + +$$ +L (\boldsymbol {x}) = \mathbb {E} _ {t \sim \mathcal {U} (0, 1)} - \frac {\mathrm {d} \lambda_ {t}}{\mathrm {d} t} w \left(\lambda_ {t}\right) \| \boldsymbol {x} - \hat {\boldsymbol {x}} \| ^ {2}, \tag {2} +$$ + +where $w(\lambda_t)$ is the weighting function. $-\mathrm{d}\lambda_{t} / \mathrm{d}t$ cancels out the effect of the noise schedule $\lambda_{t}$ on the loss. In other words, $L(\mathbf{x})$ can be rewritten as $\int -w(\lambda)\| \pmb {x} - \hat{\pmb{x}}\| ^2 d\lambda$ that is invariant to the choice of $\lambda_{t}$ . However, $\lambda_{t}$ still impacts the sampling efficiency of the Monte Carlo estimator of the expected loss, and thus affects the training dynamics of diffusion models. This decouples noise schedule $\lambda_{t}$ and weighting function $w(\lambda)$ as a function of $\lambda$ [25]. + +The above squared error loss can be derived from the variational interpretation of diffusion models [26]. Given + +two time steps $s < t$ , we define the denoising distribution $p(\boldsymbol{z}_s|\boldsymbol{z}_t)$ in the same formulation as the true posterior $q(\boldsymbol{z}_s|\boldsymbol{z}_t,\boldsymbol{x})$ , but replace the ground truth $\boldsymbol{x}$ with the learned approximation $\hat{\boldsymbol{x}} \approx \mathbb{E}[\boldsymbol{x}|\boldsymbol{z}_t]$ : + +$$ +p \left(\boldsymbol {z} _ {s} \mid \boldsymbol {z} _ {t}\right) = q \left(\boldsymbol {z} _ {s} \mid \boldsymbol {z} _ {t}, \boldsymbol {x} = \hat {\boldsymbol {x}}\right) = \mathcal {N} \left(\boldsymbol {z} _ {s} \mid \mu_ {t \rightarrow s}, \sigma_ {t \rightarrow s}\right), \tag {3} +$$ + +where $\sigma_{t\rightarrow s}^2 = \left(1 / \sigma_s^2 +\alpha_{t|s}^2 /\sigma_{t|s}^2\right)^{-1}$ and $\mu_{t\to s} = \sigma_{t\to s}^{2}\Bigl (\alpha_{t|s} / \sigma_{t|s}^{2}\pmb {z}_{t} + \alpha_{s} / \sigma_{s}^{2}\pmb {x}\Bigr)$ with $\alpha_{t|s} = \alpha_t / \alpha_s$ and $\sigma_{t|s}^2 = \sigma_t^2 -\alpha_{t|s}\sigma_s^2$ . Note that during sampling, the standard deviation for $p(z_{s}|z_{t})$ is chosen to be a log-linear interpolation between $\sigma_{t|s}$ and $\sigma_{t\rightarrow s}$ of the form $\sigma_{t\rightarrow s}^{\gamma}\sigma_{t|s}^{1 - \gamma}$ where $\gamma = 0.3$ was set empirically, which is important as FID is sensitive to over-smooth or grainy images. + +# 2. Simpler Diffusion + +# 2.1. Revisiting the sigmoid loss + +Recall that one of the proposed losses in VDM++ [25] is the sigmoid loss. They expressed the loss with respect to $\epsilon$ -mse: $L(\pmb{x}) = \mathbb{E}_{t \sim \mathcal{U}(0,1)} - \frac{\mathrm{d}\lambda_t}{\mathrm{d}t} \tilde{w}(\lambda_t) \| \pmb{\epsilon} - \hat{\pmb{\epsilon}} \|^2$ , which is connected to $\pmb{x}$ -mse through $\| \pmb{\epsilon} - \hat{\pmb{\epsilon}} \|^2 = \exp(\lambda_t) \| \pmb{x} - \hat{\pmb{x}} \|^2$ and $\tilde{w}(\lambda_t) = \exp(\lambda_t) w(\lambda_t)$ . In VDM++, viewing the loss in $\epsilon$ -mse enables a natural connection between the weighted loss and the ELBO and showed that losses should be monotonically decreasing on $\epsilon$ -mse. Hence, they propose the sigmoid loss: $\tilde{w}(\lambda_t) = \sigma(b - \lambda_t)$ , where $b$ denotes the bias or shift and $\sigma$ is the sigmoid function. However, they ultimately opted for another loss-weighting named EDM-monotonic for high resolution image generation, see Figure 3a. + +Enforcing weighting monotonicity in $x$ Revisiting the sigmoid loss, we derive its $x$ -mse form, which turns out to conveniently be defined by a mirrored sigmoid (albeit multiplied by a constant factor): + +$$ +\begin{array}{l} \sigma \left(b - \lambda_ {t}\right) \| \boldsymbol {\epsilon} - \hat {\boldsymbol {\epsilon}} \| ^ {2} = \sigma \left(b - \lambda_ {t}\right) \exp \left(\lambda_ {t}\right) \| \boldsymbol {x} - \hat {\boldsymbol {x}} \| ^ {2} \tag {4} \\ = \exp (b) \sigma \left(\lambda_ {t} - b\right) \| \boldsymbol {x} - \hat {\boldsymbol {x}} \| ^ {2}. \\ \end{array} +$$ + +![](images/12bc20bf77af3c4d4311affe192124ed53e769feeff62a5c95055d6236936202.jpg) +(a) Weighting over $\epsilon$ -mse + +![](images/2c131284de18ff2945917ee8d0833c6474829f5f2450df0baff16c2b498b7be2.jpg) +(b) Weighting over $\pmb{x}$ -mse +Figure 3. Weighting of different losses in both $\epsilon$ and $x$ space. Note that the weightings are scaled to have a maximum of one. + +This reveals that the sigmoid loss is not only monotonically decreasing in $\epsilon$ -mse, it is also monotonically increasing in $x$ -mse. We claim that a reasonable weighting function $w(\lambda_t)$ should be monotonically increasing over $\lambda_t$ . Observe that with less noise in $z_t$ , there is more information available and it should be strictly easier to predict the original data, meaning that the $x$ -mse generally decreases for higher $\lambda_t$ . This phenomenon is also observed in practice. If the loss weighting in $x$ -mse also decreases, the model would be able to make worse predictions while obtaining less noisy (better) conditioning information and still achieve a lower loss. Simply put: Why put less weight on an easier problem? + +Studying both the form of the sigmoid and EDM-monotonic weighting functions in Figure 3, we observe that both weighting functions are monotonically decreasing in $\epsilon$ -mse, but only the sigmoid weighting is monotonically increasing in $x$ -mse. Therefore, we expect the sigmoid weighting to be a better candidate to scale to high resolution generation than the EDM-monotonic version. + +Loss shifting Recall that for a resolution that is two times higher in both height and width, an average (low pass) filter would see $\mathrm{Var}[(x_{11} + x_{12} + x_{21} + x_{22}) / 4] = 1 / 4$ of the noise, while signal strength stays the same [14, 25]. Hence, its effective logsnr is expected to increase by about $2\log 2\approx 1.386\approx 1.5$ when viewed through a low-pass filter. Suppose that the optimal setting for images of $128\times 128$ is a sigmoid weighting with $b\approx 0$ , then one would expect the optimal setting for images of $512\times 512$ to be roughly $b = -3$ .1 Empirically we find that slightly more aggressive shifts perform even better. + +# 2.2. Flop heavy scaling + +When one deals with a limited data budget, it can be difficult to balance scaling of models and regularization. Our small model version at $512 \times 512$ resolution starts out with $4 \times 4$ patching, and subsequently runs a U-ViT [14] on the $128 \times$ + +$128 \times 48$ input (assuming the raw input is of 3 channels). To make the model more expressive, one of the most common options is to double the number of channels for every layer, leading to $2 \times$ model size. However, our experiments show that this results in diminishing returns. We hypothesize that it is because the model suffers from over-fitting when trained with limited data budget such as ImageNet. Indeed, when we train using distribution augmentation [18], the best performance improves from 1.9 to 1.6. This however suggests another way to scale the model: Keep the architecture and parameter count almost the same and instead run on $2 \times 2$ patched input (so $256 \times 256 \times 12$ for 512 inputs). Similar to $2 \times$ model size, reducing the patching size also increases FLOPS by roughly a factor of 4. However, we found this model performs even better to about 1.5 without the need of distribution augmentation. We call this model flop heavy, because it has a much higher flop over parameter ratio. + +Flop heavy scaling has another interesting property. Suppose we have a small model trained to fit input of size $256 \times 256$ with $2 \times 2$ patching. Using the flop heavy scaling at $512 \times 512$ resolution enables us to directly finetune from the small model trained at $256 \times 256$ with exactly the same amount of parameters, whereas its counterpart version at $512 \times 512$ resolution and with $4 \times 4$ patching needs to introduce additional parameters to handle the more input channels. This method of finetuning to higher resolutions has already been popularized in the context of text-to-image generation [32]. Here we find that flop heavy scaling is a more favorable way to scale on smaller datasets such as ImageNet. Of course, in principle one could also train a model with more channels and $4 \times 4$ patching on a lower resolution, but this is not done in practice to train 'small' models. + +# 2.3. Residual U-ViTs: Removing blockwise skipconnections + +U-Nets require that sub-sampled levels (for example via average pooling) are bypassed by skip-connections. However, this makes the architecture very different from large-scale language models such as transformers, where these skip-connections are not used and residual connections suffice. + +In this paper we use the U-ViT model of [14], which is a variant of U-Nets with self-attention layers (with convolutional layers) being replaced by transformer blocks. Hence, that part of the architecture is closer to standard transformers. + +In this paper, we further simplify model definition with Residual U-ViTs, see Figure 4. The blockwise skip-connections are removed, and in their place only a single skip-connection is used per downsampling operation. To be precise, let $f_{d}, f_{u}$ denote a down and up level of a U-ViT and let $f_{m}$ denote the middle stage. This two-stage Residual U-ViT can now be described by: + +$$ +f (x) = f _ {u} \left(\mathrm {U} \left(f _ {m} (\mathrm {D} (h)) - \mathrm {D} (h)\right) + h\right) \text {w h e r e} h = f _ {d} (x) \tag {5} +$$ + +where $D$ is typically an average pooling and linear and $U$ is typically a linear and nearest neighbour upsampling. Defining this network to preserve an identity function at initialization is now much easier: As long as $f_{d}$ , $f_{u}$ and $f_{m}$ are initialized to be identity, the entire $f$ is also identity at initialization. The generalization to any-depth U-ViTs is also straightforward: we can define another two levels of U-ViTs in place of $f_{m}$ recursively, where the nested U-ViTs can also be guaranteed to be initialized as an identity function. + +For small networks, it could be helpful to keep the blockwise skip-connections, and expect to incur a slight performance degradation from removing the blockwise skip-connections. Small networks have limited capacity, so limiting skip-connections may create bottleneck issues. Experimentally, this is confirmed for the small architectures, where we observe 0.2 FID performance degradation on ImageNet512 and 0.06 FID performance degradation on ImageNet256. Because of the simplified definition, we choose to accept this performance penalty for small models. Moreover, we find that larger models perform on par or even benefit from removing the blockwise skip-connections. + +The final advantage is memory consumption at evaluation. For example, a feature map at $256^{2}$ resolution with 128 channels requires $256^{2} \cdot 128 \cdot 2 / 1024^{2} = 16\mathrm{MB}$ . With 3 blocks per level, this requires holding an extra $3 \cdot 16 = 48\mathrm{MB}$ per example for the skip-connections, whereas the Residual U-ViT only requires holding $h$ and $D(h)$ (the latter could be cheaply recomputed), so $16 + 8 = 24\mathrm{MB}$ per example while computing the lower levels. For a hypothetical super-resolution model at $4096^{2}$ resolution, this leads to a difference between $3 \cdot 4 = 12\mathrm{GB}$ and only $4 + 2 = 6\mathrm{GB}$ . During training this advantage is usually negligible, because feature maps also need to be retained for the backward pass. A final advantage of Residual U-ViTs is that it is more straightforward to define asymmetric U-ViTs, i.e., with a different number of blocks in the downsampling (encoding) stage from the one in the upsampling (decoding) stage. In certain cases, moving blocks from the downsampling stage to the upsampling stage can increase performance. + +![](images/db47be7ec0de83b6399d13043bf84e5cd0faa4c5cc3431cd3fa9f8131560132e.jpg) +Figure 4. Residual U-ViT with levelwise skip-connections (instead of blockwise). When the network blocks are initialized to be identity, the entire network is also an identity function at initialization. + +# 2.4. Distillation + +Sampling from a pixel-space diffusion model naively often requires more function evaluations than sampling from a latent-space diffusion model. For example, the popular Stable Diffusion model [32] is usually sampled using 20-50 sampling steps, while pixel-space models are sampled with $100+$ steps [14]. To reduce the number of sampling steps, we could use either more advanced solvers or distillation. Recently distillation methods show that pixel-space models can be adapted into few-step samplers that maintain high sample quality [23, 29, 36, 44, 45, 47]. In this work, we use moment matching distillation [37] which is currently one of the best distillation methods available in the few step regime. We distill our pixel-space models for use with 8-16 sampling steps, which empirically we find maintains or even improves the sample quality of our models. + +# 3. Related Work + +The first end-to-end diffusion work that directly generated high resolution images was [4]. Given the poor performance of end-to-end models, two popular alternatives were introduced: cascaded diffusion [12] and latent diffusion [32]. + +Cascaded diffusion approaches have been shown to scale to large datasets and models to generate high quality and high resolution imagery, albeit in multiple stages [12, 40, 42, 52], or with fine-tuning [43]. Although the focus of this paper has been single stage generation up to a resolution of $10^{24^2}$ , the loss weightings may also be used for cascaded models. + +Significant improvements were made to end-to-end diffusion by shifting noise schedules [3, 14] and subsequently loss weightings [25]. Other approaches such as multi-resolution losses (and possibly architecture adaptions) [6, 7, 14] added marginal gains. However, the current best known latent model EDM2 [20] outperforms these prior end-to-end approaches by a large margin. In contrast, our models are competitive with latent models, outperforming DiT-XL/2 [31] and only performing slightly worse than EDM2. + +Skip connections The U-Net architecture is a fully convolutional network with long-range skip-connections introduced for (medical) image segmentation [33, 39]. The main + +![](images/3e8b1ffcef67cfe66bf5c5c04d334ca80c64b63852d2acf792d78168f86a8358.jpg) +Figure 5. Effect of weighting functions in the loss on the FID score of ImageNet512 over the training iterations. The proposed sigmoid weighting with $b = -3$ shows best performance consistent across all training iterations. + +idea of U-Net is to combine a contracting path (which encodes the image semantics) with long skip-connections at the same resolution to enable precise localization. + +With the introduction of ViTs [5] as an alternative for convolutional layers, the U-Net has been converted to U-ViTs using a similar network but replacing (some) ResNet blocks with ViTs, either with spatial down-sampling [14] or without [2]. Long-range skip-connections play an pivotal role in these networks, for example for fine-tuning controllable diffusion models [17] and for the stability and training speed [15, 30]. In these latter works, it is shown that multiplying the skip-connections with a (constant) scalar $(< 1)$ improves training stability and speed, because of the propagation of noise throughout the network. In our work, we reduce the number of skip-connections significantly and initialise the network such that it resembles an identify function. + +Flop heavy fine-tuning Fine-tuning on a higher resolution at large scale was effectively done in Stable Diffusion [32] to limit the training cost, which is a form of flop heavy scaling over different resolutions. In this paper we show that flop heavy scaling is also preferred from a regularization perspective, which can be done either by finetuning from a lower resolution, or by lowering the patching size. + +# 4. Experiments + +# 4.1. Sigmoid weighting versus others + +Loss weighting In this experiment, we want to understand the effect of loss weighting on the final image quality of a model. For the noise schedule, we use an interpolated cosine schedule because it has relatively broad coverage over logsnr values. We compare the weightings corresponding to the original $\epsilon$ -mse [11] and the multiscale $\epsilon$ -mse [14], the EDM shifted weightings [19], including the original and monotonic versions [19, 26], and the proposed sigmoid weighting + +![](images/0a20f14d695a3b2c1dc90f07e7e80a7af439384ddd39d007754bda786705f538.jpg) +Figure 6. Effect of training noise schedules on the FID score of ImageNet512 over the training iterations. Fortunately, with our proposed sigmoid weighting $(b = -3)$ , the exact choice of the noise schedules has little effect on the training process. + +Table 1. Sigmoid weighting performance (left) with the flop heavy architecture. Larger resolutions require more negative bias. Power loss performance (right) on ImageNet512 with the small model. Increasing downsampling has a similar effect as decreasing bias, hence we prefer the more simple sigmoid weighting. + +
(a) Sigmoid weighting(b) Power loss weighting
biasresolutionbiasdownsample levels
12825651210240123
-5---1.82-42.192.583.625.11
-4--1.631.75-32.192.322.854.02
-3-1.471.482.32-22.292.152.343.13
-21.491.421.552.17-12.642.322.252.54
-11.351.381.75-03.222.742.072.22
01.301.461.94-1-2.692.292.16
11.261.56--2-3.262.532.19
+ +with shift $b = -3$ . As can be seen in Figure 5, the sigmoid weighting outperforms all other weightings, with EDM monotonic shifted being the closest alternative. Appendix C contains additional experiments, where we also evaluate the sigmoid weighting against the closest competitor without guidance. Here again sigmoid weighting is consistently better. We also evaluate a time-shifted variant, where the bias depends on the progress in training. + +Training efficiency of noise schedules Since the training noise schedule can be decoupled from the loss weighting and it only affects the training efficiency but not the final performance, one can measure the convergence speed over the training process caused by leveraging particular noise schedules. In this experiment we test the cosine schedule [11], the shifted cosine and the interpolated cosine schedules [14] (Figure 6). For fair comparison, we fix the sampling noise schedule as the cosine interpolated schedule. Interestingly, when the sigmoid weighting is used, the choice of noise schedule does not affect the training efficiency significantly. Arguably, the shifted schedule is a bit more sample effi + +![](images/cfba554f0499d140f421d739d32d5229f4c5d7630c45e389a8d625935e959e56.jpg) +Figure 7. FID versus training cost on ImageNet512. Note that the EDM-2 training curves are without guidance, but the best performance with guidance intervals is reported with a separate marker, the figure source adapted has been adapted from EDM2 [20]. Following their assumptions, one training iteration is three times as expensive as evaluating the model. Our Flop Heavy (finetune) is trained by first training a small model on ImageNet256 for 400k iterations and then finetuning the same model on ImageNet512. + +![](images/584f7bccdcdfe72473e5875f5ad05299e881bc000dab70b530254646605fc7cc.jpg) +Figure 8. The effect of scaling either channels or tokens, compared to the small model variant. FLOPS between model size $2\mathrm{x}$ and flop heavy are similar. + +![](images/c32f83e922fb55d924c8bf41c4c5064257dcb4a5573c063f4ece679f81f58f5a.jpg) +Figure 9. The effect of Residual-UViTs (removing block skip connections). Although smaller models can be effected, larger models are unaffected by the simplification. + +![](images/21d9702abf3eda288b9e9681b14556e5f7071b8f2e50a14b2017ca07968d5fd0.jpg) +Figure 10. Changing the encoder decoder architecture. Decoder blocks are more important than encoder blocks, and symmetric architectures have decent performance. + +cient initially but all schedules perform similarly towards the end. This suggests that there is a lot of transfer across different logsnr values. We also tried the EDM training noise schedule with the sigma power sampling noise schedule [19] (not plotted), with both shifted or the original version, but its narrower logsnr sampling schedule led to performance degradation, achieving about 9 FID at best. This suggests that for high resolution, a wide logsnr range is crucial to cover the important noise levels. + +Bias over resolutions In Table 1a, we show the effect of the bias term $b$ on the final performance when trained using a flop heavy model. All models use the same guidance and similar guidance intervals, where the lower bound of the guidance interval has been shifted according to resolution (again approximately 1.5 per 2x resolution change), see Appendix B for more details. + +Sigmoid weighting versus the power loss In literature, high resolution generation can greatly benefit from having a multiscale loss [6, 7, 14]. To that end we create a multiscale + +variant that we name the power loss. In addition to the sigmoid weighting that effectively upweights low frequency components of the data and downweights high frequency components, the power loss aims to amplify the loss of low sub-band (smooth) signals in the spatial domain via Haar wavelets. For details on this loss see Appendix B.2. + +In Table 1b we see that the bias should be adjusted based on the number of Haar transform downsampling levels. Despite being competitive, the performance of power loss is roughly on par with the one with sigmoid weighting but without multiscaling (i.e. downsample levels $= 0$ ), if we tune the bias of the sigmoid weighting more aggressively. Interestingly, lowering the bias has almost the same effect as increasing the scale. Therefore, building the multiscale variant is not necessary with the sigmoid weighting. + +# 4.2. Scaling channels vs tokens on ImageNet + +In this section we show the difference between scaling a model 4x FLOPS, either via channels or by creating larger resolution feature maps (or put differently, feature maps with a higher token count). Note that our small model on + +ImageNet512 is patched $4 \times 4$ . This model is scaled either by increasing channels by two (named model size $2x$ ), or by reducing patching to $2 \times 2$ (named flop heavy, because FLOPS over params is much higher). Indeed as can be seen in Figure 8, the flop heavy model performs much better than the model size $2x$ , when trained in a standard way on ImageNet512. Note that unlike most experiments, in these experiments it was necessary to use very different optimal guidance settings, more details in Appendix B. + +We confirm this is an issue related to overfitting, by training the model size 2x with distribution augmentation [18] the model size 2x variant performs much better, with an FID of 1.6 instead of 1.9 (latter in Figure 8). Still, without any augmentation beyond the standard horizontal flips, the flop heavy model achieves a score of 1.54. + +# 4.3. Removing block skip connection + +In this section we study the difference between the residual skip-connections and the blockwise skip-connections. As hypothesized, the small model performance drops a bit because of this. For ImageNet512 a deficit of about 0.2 FID (in Figure 9) and for Imagenet256 a deficit of about 0.05 FID. That the task of fitting ImageNet256 is less affected by removing skip-connections can be explained by the task being easier: The task of fitting ImageNet256 is a subset of fitting ImageNet512. + +For larger models, removing blockwise skip-connections in favor of a single residual connection does not degrade performance, and may even lead to slightly increased performance (see Figure 9). This is somewhat unexpected, and the phenomenon persists with different guidance settings as well. The FID curve reminds of the model size $2\mathrm{x}$ curve, which at first improves but flattens out over iterations. Therefore our best hypothesis is that the block skip (flop) baseline may have similar problems such as overfitting. Regardless, it shows that Residual U-ViTs scale well. + +Asymmetric U-ViT Since our U-ViT models encompass only a single skip connection for each level, we can use a different number of ResNet or ViT blocks per level. This allows to move compute from the encoder stage (the down sampling half of the model) to the decoder stage (the upsampling half of the model). From Figure 10 we observe that (a) for a fixed compute budget, it is better to add more parameters to the decoder, the encoder heavy model (E5-D1) is significantly worse than the baseline (E3-D3) and the decoder heavy model (E1-D5); (b) there needs to be some balance between the encoder and the decoder, E1-D7 works less good than E3-D5. (c) by a slightly heavier model (E3-D7, only $10\%$ more parameters than our small (E3-D3) model, the performance significantly increases. + +Table 2. Literature Comparison on ImageNet on FID. Note that even though certain NFEs are high (e.g. 4100 for CDM) typically these suffer minor performance degradation for lower NFEs and should not be taken too literally. + +
MethodNFEFID
Imagenet 128 x 128
RIN [16]10002.75
SD, U-ViT-L [14]5121.94
StyleGAN-XL [38]11.81
VDM+++ [25]2501.75
DisCo-Diff [46]4141.73
PaGoDA [22]11.48
SiD2, Flop Heavy (ours)5121.26
Imagenet 256 x 256
ADM-G [4]2504.59
ADM-U [4]2503.94
CDM [12]41004.88
RIN [3, 16]10003.52
SD, U-ViT-L [14]5122.44
MDM [7]2503.51
StyleGAN-XL [38]12.30
DiT-XL/2 [31]2502.27
VDM++ [25]2502.12
RDM [42]3561.89
PaGoDA [22]11.56
SiD2, small (ours)5121.72
SiD2, Flop Heavy (ours)5121.38
Imagenet 512 x 512
ADM-G [4]2507.72
RIN [3, 16]10003.95
ADM-U [4]2503.85
LDM-4-G [32]2503.60
SD, U-ViT-L [14]5123.02
DiffiT [9]2502.67
StyleGAN-XL [38]12.40
DiT-XL/2 [31]2503.04
VDM++ [25]2502.65
MAGVIT-v2 [51]641.91
PaGoDA [22]11.80
EDM2-S (interval) [27]631.68
EDM2-XXL (interval) [27]631.40
SiD2, small (ours)5122.19
SiD2, Flop Heavy (ours)5121.48
SiD2, Flop Heavy distilled (ours)161.50
Imagenet 1024 x 1024
RIN [16]10008.72
SiD2, Flop Heavy (ours)5121.75
+ +Table 3. Zero-shot text-to-image performance evaluated on MSCOCO in $\mathrm{FID}_{30\mathrm{k}}$ . Our distilled model, evaluated on the more difficult 512 setting, outperforming other works. + +
MethodSizeNFEFID30k
DeepFloyd [40]2566.7
eDiff-I [1]2566.9
Parti [49]2567.2
Imagen [35]2567.3
simple diffusion [14]2568.3
simple diffusion [14]5122569.6
SDv1.5 (eval DMD2) [32]5122007.2
DMD2 [48]51248.3
SiD2 (ours, bias=-3)5122568.1
SiD2 16-step distilled (ours)512166.7
+ +# 4.4. Literature comparison + +Here we compare different models from the literature. In the first section end-to-end diffusion models are compared, and in the second section we compare between all types of generative models. In Table 2 it can be seen that our SiD2 is very competitive with existing generative models: On ImageNet128 and ImageNet256 it achieves state-of-the-art performance, on ImageNet512 it is a little worse but still competitive with the best latent diffusion model and outperforms all other models, pixel-based and latent. + +Comparing diffusion models on compute The training cost SiD2 is another helpful way to see how the model scales, as depicted in Figure 7). Importantly, the performance of our pixel-based approach outperforms other significantly. Observe for example the performance of our small variant (using no guidance interval to keep this comparison fair), which roughly the same performance as the VDM++ while only using a third of the training cost, and a fourth of the FLOPS in a single forward pass (see Figure 16 in Appendix C). It is also competitive with the best currently known latent diffusion approach EDM2 [20], but in fairness EDM2 scales a little bit better. These models attain even better performance for a smaller training cost, and a similar story holds for forward pass complexity. End-to-end diffusion can work much better than previously demonstrated, but it is still somewhat outperformed by latent diffusion on quality versus compute comparisons for both sampling and training. + +Text to image generation Finally, we show that our method can be used for end-to-end training of text-to-image models. Here, evaluation is more difficult as MSCOCO zero-shot performance depends a lot on the dataset used and is known to not necessarily correspond to actual perceived performance. Nevertheless, it is important that the model achieves performance in the right ballpark. Indeed, when we train SiD2 it performs comparable to other models in the + +Table 4. Conditional Video generation performance on Kinetics600 in FVD (resized to 64). For video generation the loss specification is important to obtain good sample quality (see ablations). + +
MethodFVD50k
pixel diffusion
Video Diffusion Models [13]16.6
RIN [16]10.8
Rolling Diffusion [34]5.2
latent diffusion
MAGVIT [50]9.9
MAGVITv2 [51]4.3
W.A.L.T [8]3.3
SiD2 (ours, sigmoid bias = 0)2.3
ablations
eps-mse5.7
sigmoid bias = +23.2
sigmoid bias = -22.5
+ +literature on MSCOCO $\mathrm{FID}_{30\mathrm{K}}$ , and detailed samples have only a small number of artifacts (Figure 2). Furthermore, the moment matching distilled version using [37] outperforms the baseline model as previously observed. This model also outperforms all other zero-shot approaches at 512 resolution as shown in Table 3. + +Video Generation To test whether our approach transfers to video, we test our model on Kinetics600 following [13], by conditioning on 5 frames and generating (sometimes referred to as predicting) the subsequent 11 frames. We then compute the FVD between the training set and the generated samples both on the subsequent 11 frames. Even though our method generates at $128 \times 128$ , we follow [8] and resize samples to have resolution $64 \times 64$ . Our model achieves SOTA performance and it turns out that tuning the loss is very important to achieve this as shown in Table 4. Note that the latent approaches ([8, 50, 51]) use much less training compute than our models in pixel-space, which we discuss in more detail in the Appendix B. + +# 5. Discussion + +In this paper we present Simpler Diffusion (SiD2) and show how one can train end-to-end diffusion and achieve competitive performance with latent models: (1) use the sigmoid loss, (2) with a simpler U-ViT architecture, and (3) favour a flop heavy model design. As a result, our method outperforms existing end-to-end models by a large margin. Nevertheless, at this moment latent diffusion methods still seem to have slightly better scaling properties and it remains to be seen whether this gap can be closed any further. Importantly, for many applications training a separate autoencoder may not be either practical or desired. Therefore, we believe this paper is an important contribution that allows end-to-end modelling directly in the pixel-space. + +# References + +[1] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Ji-aming Song, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, Bryan Catanzaro, Tero Karras, and Ming-Yu Liu. ediffi: Text-to-image diffusion models with an ensemble of expert denoisers. CoRR, abs/2211.01324, 2022. 8 +[2] Fan Bao, Shen Nie, Kaiwen Xue, Yue Cao, Chongxuan Li, Hang Su, and Jun Zhu. All are worth words: A vit backbone for diffusion models. In CVPR, 2023. 5 +[3] Ting Chen. On the importance of noise scheduling for diffusion models. arxiv, 2023. 1, 4, 7 +[4] Prafulla Dhariwal and Alex Nichol. Diffusion models beat gans on image synthesis. CoRR, abs/2105.05233, 2021. 4, 7 +[5] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In 9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7, 2021. OpenReview.net, 2021. 5 +[6] Jiatao Gu, Shuangfei Zhai, Yizhe Zhang, Miguel Angel Bautista, and Josh M. Susskind. f-dm: A multi-stage diffusion model via progressive signal transformation. CoRR, abs/2210.04955, 2022. 4, 6 +[7] Jiatao Gu, Shuangfei Zhai, Yizhe Zhang, Josh M. Susskind, and Navdeep Jaitly. Matryoshka diffusion models. CoRR, abs/2310.15111, 2023. 4, 6, 7 +[8] Agrim Gupta, Lijun Yu, Kihyuk Sohn, Xiuye Gu, Meera Hahn, Fei-Fei Li, Irfan Essa, Lu Jiang, and José Lezama. Photorealistic video generation with diffusion models. In Computer Vision - ECCV 2024 - 18th European Conference, Milan, Italy, September 29-October 4, 2024, Proceedings, Part LXXIX, pages 393-411. Springer, 2024. 8 +[9] Ali Hatamizadeh, Jiaming Song, Guilin Liu, Jan Kautz, and Arash Vahdat. Diffit: Diffusion vision transformers for image generation. CoRR, abs/2312.02139, 2023. 7 +[10] Jonathan Heek, Emiel Hoogeboom, and Tim Salimans. Multistep consistency models. CoRR, abs/2403.06807, 2024. 3 +[11] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In Advances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, NeurIPS, 2020. 1, 5 +[12] Jonathan Ho, Chitwan Sahara, William Chan, David J. Fleet, Mohammad Norouzi, and Tim Salimans. Cascaded diffusion models for high fidelity image generation. J. Mach. Learn. Res., 23:47:1-47:33, 2022. 4, 7 +[13] Jonathan Ho, Tim Salimans, Alexey A. Gritsenko, William Chan, Mohammad Norouzi, and David J. Fleet. Video diffusion models. In Advances in Neural Information Processing Systems 35: Annual Conference on Neural Information Processing Systems 2022, NeurIPS, 2022. 8 +[14] Emiel Hoogeboom, Jonathan Heek, and Tim Salimans. simple diffusion: End-to-end diffusion for high resolution images. In International Conference on Machine Learning, ICML, pages 13213-13232. PMLR, 2023. 1, 3, 4, 5, 6, 7, 8 + +[15] Zhongzhan Huang, Zhou Pan, Shuicheng Yan, and Liang Lin. Scalelong: Towards more stable training of diffusion model via scaling network long skip connection. In NeurIPS, 2023. 5 +[16] Allan Jabri, David J. Fleet, and Ting Chen. Scalable adaptive computation for iterative generation. CoRR, abs/2212.11972, 2022. 7, 8 +[17] Zeyinzi Jiang, Chaojie Mao, Yulin Pan, Zhen Han, and Jingfeng Zhang. Scedit: Efficient and controllable image diffusion generation via skip connection editing. Technical Report 2312.11392, arXiv, 2023. 5 +[18] Heewoo Jun, Rewon Child, Mark Chen, John Schulman, Aditya Ramesh, Alec Radford, and Ilya Sutskever. Distribution augmentation for generative modeling. In Proceedings of the 37th International Conference on Machine Learning, ICML, 2020. 3, 7 +[19] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. In Advances in Neural Information Processing Systems 35: Annual Conference on Neural Information Processing Systems 2022, NeurIPS, 2022. 5, 6 +[20] Tero Karras, Miika Aittala, Jaakko Lehtinen, Janne Hellsten, Timo Aila, and Samuli Laine. Analyzing and improving the training dynamics of diffusion models. CoRR, abs/2312.02696, 2023. 1, 4, 6, 8, 2, 5 +[21] Tero Karras, Miika Aittala, Tuomas Kynkänniemi, Jaakko Lehtinen, Timo Aila, and Samuli Laine. Guiding a diffusion model with a bad version of itself. CoRR, abs/2406.02507, 2024. 2 +[22] Dongjun Kim, Chieh-Hsin Lai, Wei-Hsiang Liao, Yuhta Takida, Naoki Murata, Toshimitsu Uesaka, Yuki Mitsufuji, and Stefano Ermon. Pagoda: Progressive growing of a one-step generator from a low-resolution diffusion teacher. CoRR, abs/2405.14822, 2024. 7, 5 +[23] Dongjun Kim, Chieh-Hsin Lai, Wei-Hsiang Liao, Naoki Murata, Yuhta Takida, Toshimitsu Uesaka, Yutong He, Yuki Mitsufuji, and Stefano Ermon. Consistency trajectory models: Learning probability flow ODE trajectory of diffusion. In The Twelfth International Conference on Learning Representations, 2024. 4 +[24] Durk P Kingma and Prafulla Dhariwal. Glow: Generative flow with invertible 1x1 convolutions. Advances in neural information processing systems, 31, 2018. 1 +[25] Diederik P. Kingma and Ruiqi Gao. Understanding the diffusion objective as a weighted integral of ellbods. CoRR, abs/2303.00848, 2023. 1, 2, 3, 4, 7 +[26] Diederik P. Kingma, Tim Salimans, Ben Poole, and Jonathan Ho. Variational diffusion models. CoRR, abs/2107.00630, 2021. 2, 5 +[27] Tuomas Kynkänniemi, Miika Aittala, Tero Karras, Samuli Laine, Timo Aila, and Jaakko Lehtinen. Applying guidance in a limited interval improves sample and distribution quality in diffusion models. CoRR, abs/2404.07724, 2024. 7, 2 +[28] Yaron Lipman, Ricky T. Q. Chen, Heli Ben-Hamu, Maximilian Nickel, and Matthew Le. Flow matching for generative modeling. In The Eleventh International Conference on Learning Representations, ICLR 2023, Kigali, Rwanda. OpenReview.net, 2023. 2 + +[29] Weijian Luo, Tianyang Hu, Shifeng Zhang, Jiacheng Sun, Zhenguo Li, and Zhihua Zhang. Diff-instruct: A universal approach for transferring knowledge from pre-trained diffusion models. Advances in Neural Information Processing Systems, 36, 2024. 4 +[30] Jiajun Ma, Shuchen Xue, Tianyang Hu, Wenjia Wang, Zhaoqiang Liu, Zhenguo Li, Zhi-Ming Ma, and Kenji Kawaguchi. The surprising effectiveness of skip-tuning in diffusion sampling. In Proceedings of the 41st International Conference on Machine Learning, pages 34053-34074, 2024. 5 +[31] William Peebles and Saining Xie. Scalable diffusion models with transformers. CoRR, abs/2212.09748, 2022. 4, 7, 5 +[32] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 10674-10685. IEEE, 2022. 1, 3, 4, 5, 7, 8 +[33] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. Technical report, ArXiV, 2015. 4 +[34] David Ruhe, Jonathan Heek, Tim Salimans, and Emiel Hoogeboom. Rolling diffusion models. In *Forty-first International Conference on Machine Learning*, ICML 2024, Vienna, Austria, July 21-27, 2024. OpenReview.net, 2024. 8 +[35] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S. Sara Mahdavi, Rapha Gontijo Lopes, Tim Salimans, Jonathan Ho, David J. Fleet, and Mohammad Norouzi. Photorealistic text-to-image diffusion models with deep language understanding. CoRR, abs/2205.11487, 2022. 8 +[36] Tim Salimans and Jonathan Ho. Progressive distillation for fast sampling of diffusion models. In *The Tenth International Conference on Learning Representations*, ICLR. OpenReview.net, 2022. 4 +[37] Tim Salimans, Thomas Mensink, Jonathan Heek, and Emiel Hoogeboom. Multistep distillation of diffusion models via moment matching. arXiv preprint arXiv:2406.04103, 2024. 4, 8 +[38] Axel Sauer, Katja Schwarz, and Andreas Geiger. Stylegan-xl: Scaling stylegan to large diverse datasets. In SIGGRAPH '22: Special Interest Group on Computer Graphics and Interactive Techniques Conference, pages 49:1-49:10. ACM, 2022. 7, 5 +[39] Evan Shelhamer, Jonathan Long, and Trevor Darrell. Fully convolutional networks for semantic segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 39(4):640-651, 2016. 4 +[40] Alex Shonenkov, Misha Konstantinov, Daria Bakshandaeva, Christoph Schuhmann, Ksenia Ivanova, and Nadija Klokova. Deepfloyd if: A text-to-image model with a high degree of photorealism and language understanding, 2023. 4, 8 +[41] Yang Song, Jascha Sohl-Dickstein, Diederik P. Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In 9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7, 2021. OpenReview.net, 2021. 2 + +[42] Jiayan Teng, Wendi Zheng, Ming Ding, Wenyi Hong, Jianqiao Wangni, Zhuoyi Yang, and Jie Tang. Relay diffusion: Unifying diffusion process across resolutions for image synthesis. In The Twelfth International Conference on Learning Representations, ICLR. OpenReview.net, 2024. 4, 7 +[43] Haoning Wu, Shaocheng Shen, Qiang Hu, Xiaoyun Zhang, Ya Zhang, and Yanfeng Wang. Megafusion: Extend diffusion models towards higher-resolution image generation without further tuning. CoRR, abs/2408.11001, 2024. 4 +[44] Zhisheng Xiao, Karsten Kreis, and Arash Vahdat. Tackling the generative learning trilemma with denoising diffusion gans. In International Conference on Learning Representations, 2021. 4 +[45] Yanwu Xu, Yang Zhao, Zhisheng Xiao, and Tingbo Hou. Ufogen: You forward once large scale text-to-image generation via diffusion gans. arXiv preprint arXiv:2311.09257, 2023. 4 +[46] Yilun Xu, Gabriele Corso, Tommi S. Jaakkola, Arash Vahdat, and Karsten Kreis. Disco-diff: Enhancing continuous diffusion models with discrete latents. In *Forty-first International Conference on Machine Learning*, ICML 2024, Vienna, Austria, July 21-27, 2024. OpenReview.net, 2024. 7 +[47] Tianwei Yin, Michael Gharbi, Richard Zhang, Eli Shechtman, Fredo Durand, William T Freeman, and Taesung Park. One-step diffusion with distribution matching distillation. arXiv preprint arXiv:2311.18828, 2023. 4 +[48] Tianwei Yin, Michael Gharbi, Taesung Park, Richard Zhang, Eli Shechtman, Fredo Durand, and William T. Freeman. Improved distribution matching distillation for fast image synthesis. In NeurIPS, 2024. 8 +[49] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, Ben Hutchinson, Wei Han, Zarana Parekh, Xin Li, Han Zhang, Jason Baldridge, and Yonghui Wu. Scaling autoregressive models for content-rich text-to-image generation. CoRR, abs/2206.10789, 2022. 8 +[50] Lijun Yu, Yong Cheng, Kihyuk Sohn, José Lezama, Han Zhang, Huiwen Chang, Alexander G. Hauptmann, Ming-Hsuan Yang, Yuan Hao, Irfan Essa, and Lu Jiang. MAGVIT: masked generative video transformer. CoRR, abs/2212.05199, 2022. 8 +[51] Lijun Yu, José Lezama, Nitesh Bharadwaj Gundavarapu, Luca Versari, Kihyuk Sohn, David Minnen, Yong Cheng, Agrim Gupta, Xiuye Gu, Alexander G. Hauptmann, Boqing Gong, Ming-Hsuan Yang, Irfan Essa, David A. Ross, and Lu Jiang. Language model beats diffusion - tokenizer is key to visual generation. In The Twelfth International Conference on Learning Representations, ICLR. OpenReview.net, 2024. 7, 8 +[52] Wendi Zheng, Jiayan Teng, Zhuoyi Yang, Weihan Wang, Jidong Chen, Xiaotao Gu, Yuxiao Dong, Ming Ding, and Jie Tang. Cogview3: Finer and faster text-to-image generation via relay diffusion. In Computer Vision - ECCV Proceedings, Part LXXVII, pages 1-22. Springer, 2024. 4 \ No newline at end of file diff --git a/2025/Simpler Diffusion_ 1.5 FID on ImageNet512 with Pixel-space Diffusion/images.zip b/2025/Simpler Diffusion_ 1.5 FID on ImageNet512 with Pixel-space Diffusion/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..69e70a267eb2e21c7ba232c4144b168604764cca --- /dev/null +++ b/2025/Simpler Diffusion_ 1.5 FID on ImageNet512 with Pixel-space Diffusion/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9e0c1072392663deb03205df51d5705fd640e5f57ee512a19ab94ce1ba0eb137 +size 508421 diff --git a/2025/Simpler Diffusion_ 1.5 FID on ImageNet512 with Pixel-space Diffusion/layout.json b/2025/Simpler Diffusion_ 1.5 FID on ImageNet512 with Pixel-space Diffusion/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..18b7ac4fbd1d55d13ccc4375b5d3905a80e395b7 --- /dev/null +++ b/2025/Simpler Diffusion_ 1.5 FID on ImageNet512 with Pixel-space Diffusion/layout.json @@ -0,0 +1,9352 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 89, + 103, + 521, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 103, + 521, + 121 + ], + "spans": [ + { + "bbox": [ + 89, + 103, + 521, + 121 + ], + "type": "text", + "content": "Simpler Diffusion: 1.5 FID on ImageNet512 with pixel-space diffusion" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 118, + 144, + 212, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 144, + 212, + 157 + ], + "spans": [ + { + "bbox": [ + 118, + 144, + 212, + 157 + ], + "type": "text", + "content": "Emiel Hoogeboom" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 118, + 158, + 211, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 158, + 211, + 171 + ], + "spans": [ + { + "bbox": [ + 118, + 158, + 211, + 171 + ], + "type": "text", + "content": "Google Deepmind" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 112, + 172, + 221, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 112, + 172, + 221, + 185 + ], + "spans": [ + { + "bbox": [ + 112, + 172, + 221, + 185 + ], + "type": "text", + "content": "emielh@google.com" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 264, + 144, + 351, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 264, + 144, + 351, + 156 + ], + "spans": [ + { + "bbox": [ + 264, + 144, + 351, + 156 + ], + "type": "text", + "content": "Thomas Mensink" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 264, + 158, + 353, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 264, + 158, + 353, + 171 + ], + "spans": [ + { + "bbox": [ + 264, + 158, + 353, + 171 + ], + "type": "text", + "content": "Google Deepmind" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 250, + 172, + 366, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 250, + 172, + 366, + 185 + ], + "spans": [ + { + "bbox": [ + 250, + 172, + 366, + 185 + ], + "type": "text", + "content": "mensink@google.com" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 410, + 144, + 484, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 410, + 144, + 484, + 156 + ], + "spans": [ + { + "bbox": [ + 410, + 144, + 484, + 156 + ], + "type": "text", + "content": "Jonathan Heek" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 402, + 158, + 493, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 402, + 158, + 493, + 171 + ], + "spans": [ + { + "bbox": [ + 402, + 158, + 493, + 171 + ], + "type": "text", + "content": "Google Deepmind" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 397, + 172, + 499, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 397, + 172, + 499, + 185 + ], + "spans": [ + { + "bbox": [ + 397, + 172, + 499, + 185 + ], + "type": "text", + "content": "jheek@google.com" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 124, + 206, + 199, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 124, + 206, + 199, + 220 + ], + "spans": [ + { + "bbox": [ + 124, + 206, + 199, + 220 + ], + "type": "text", + "content": "Kay Lamarigts" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 120, + 220, + 205, + 233 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 120, + 220, + 205, + 233 + ], + "spans": [ + { + "bbox": [ + 120, + 220, + 205, + 233 + ], + "type": "text", + "content": "Google Research" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 88, + 234, + 236, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 234, + 236, + 247 + ], + "spans": [ + { + "bbox": [ + 88, + 234, + 236, + 247 + ], + "type": "text", + "content": "kaylamerigts@google.com" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 290, + 206, + 342, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 290, + 206, + 342, + 220 + ], + "spans": [ + { + "bbox": [ + 290, + 206, + 342, + 220 + ], + "type": "text", + "content": "Ruiqi Gao" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 271, + 220, + 361, + 233 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 271, + 220, + 361, + 233 + ], + "spans": [ + { + "bbox": [ + 271, + 220, + 361, + 233 + ], + "type": "text", + "content": "Google Deepmind" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 262, + 234, + 371, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 262, + 234, + 371, + 247 + ], + "spans": [ + { + "bbox": [ + 262, + 234, + 371, + 247 + ], + "type": "text", + "content": "ruiqig@google.com" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 424, + 206, + 493, + 219 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 424, + 206, + 493, + 219 + ], + "spans": [ + { + "bbox": [ + 424, + 206, + 493, + 219 + ], + "type": "text", + "content": "Tim Salimans" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 413, + 220, + 504, + 233 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 413, + 220, + 504, + 233 + ], + "spans": [ + { + "bbox": [ + 413, + 220, + 504, + 233 + ], + "type": "text", + "content": "Google Deepmind" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 397, + 234, + 518, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 397, + 234, + 518, + 247 + ], + "spans": [ + { + "bbox": [ + 397, + 234, + 518, + 247 + ], + "type": "text", + "content": "salimans@google.com" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 152, + 289, + 200, + 301 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 289, + 200, + 301 + ], + "spans": [ + { + "bbox": [ + 152, + 289, + 200, + 301 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 55, + 313, + 296, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 313, + 296, + 420 + ], + "spans": [ + { + "bbox": [ + 55, + 313, + 296, + 420 + ], + "type": "text", + "content": "Latent diffusion models have become the popular choice for scaling up diffusion models for high resolution image synthesis. Compared to pixel-space models that are trained end-to-end, latent models are perceived to be more efficient and to produce higher image quality at high resolution. Here we challenge these notions, and show that pixel-space models can be very competitive to latent models both in quality and efficiency, achieving 1.5 FID on ImageNet512 and new SOTA results on ImageNet128, ImageNet256 and Kinetics600." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 54, + 422, + 296, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 422, + 296, + 541 + ], + "spans": [ + { + "bbox": [ + 54, + 422, + 296, + 541 + ], + "type": "text", + "content": "We present a simple recipe for scaling end-to-end pixel-space diffusion models to high resolutions. 1: Use the sigmoid loss-weighting [25] with our prescribed hyperparameters. 2: Use our simplified memory-efficient architecture with fewer skip-connections. 3: Scale the model to favor processing the image at a high resolution with fewer parameters, rather than using more parameters at a lower resolution. Combining these with guidance intervals, we obtain a family of pixel-space diffusion models we call Simpler Diffusion (SiD2)." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 55, + 562, + 135, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 562, + 135, + 574 + ], + "spans": [ + { + "bbox": [ + 55, + 562, + 135, + 574 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 55, + 582, + 296, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 296, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 296, + 689 + ], + "type": "text", + "content": "Diffusion models have been tremendously successful in domains such as image, video and audio generation. For high resolution data, latent diffusion models [32] have been particularly successful: Through encoding the data in a lower dimensional space, these models sidestep the difficulties faced by more traditional pixel-space diffusion models in scaling up to high resolution. This makes latent diffusion models also scale better in training compute than other approaches, as shown by [20]." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": "However, latent models have their own challenges: It can be difficult to train the autoencoders that produce the latents," + } + ] + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 332, + 289, + 540, + 370 + ], + "blocks": [ + { + "bbox": [ + 332, + 289, + 540, + 370 + ], + "lines": [ + { + "bbox": [ + 332, + 289, + 540, + 370 + ], + "spans": [ + { + "bbox": [ + 332, + 289, + 540, + 370 + ], + "type": "image", + "image_path": "8b5a7ae5161b82f15e43113268a9298406651cafd4317c76bb573e36806124bf.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 380, + 553, + 392 + ], + "lines": [ + { + "bbox": [ + 313, + 380, + 553, + 392 + ], + "spans": [ + { + "bbox": [ + 313, + 380, + 553, + 392 + ], + "type": "text", + "content": "Figure 1. End-to-end pixel diffusion performance on ImageNet512." + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + } + ], + "index": 27 + }, + { + "bbox": [ + 312, + 400, + 554, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 400, + 554, + 508 + ], + "spans": [ + { + "bbox": [ + 312, + 400, + 554, + 508 + ], + "type": "text", + "content": "which in practice requires careful tuning of regularization hyperparameters [32]. Even when trained well, the decoder of a latent diffusion model can produce artifacts or otherwise imperfect reconstructions, especially when it is applied to latents that are out of distribution compared to the training set. A single-stage pixel-space diffusion model does not have these issues, and so we ask the question: Can pixel-space diffusion models scale as well as latent models with the right architecture and hyperparameters?" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 312, + 509, + 555, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 509, + 555, + 628 + ], + "spans": [ + { + "bbox": [ + 312, + 509, + 555, + 628 + ], + "type": "text", + "content": "Over the years, end-to-end pixel-space diffusion has improved considerably (Figure 1). The main problem was that the standard schedule and loss-weighting [11] work well for small resolution images, like " + }, + { + "bbox": [ + 312, + 509, + 555, + 628 + ], + "type": "inline_equation", + "content": "32^2" + }, + { + "bbox": [ + 312, + 509, + 555, + 628 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 312, + 509, + 555, + 628 + ], + "type": "inline_equation", + "content": "64^2" + }, + { + "bbox": [ + 312, + 509, + 555, + 628 + ], + "type": "text", + "content": ", but not at higher resolution [3, 14]. It was further shown [25] that it can be helpful to disentangle loss weighting and time sampling for improved control over both, leading to better performance at higher resolution. However, in previous works pixels-space models still did not match the best latent models on standard benchmarks like ImageNet512." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 312, + 630, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 630, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 312, + 630, + 556, + 713 + ], + "type": "text", + "content": "The main contribution of this paper is to improve pixel-space diffusion on high resolution images while keeping as close to the original formulation of diffusion models as possible. Specifically, we analyze and tune the sigmoid loss weighting from [25], and show why it should be preferred over their favored EDM monotonic. We further propose architectural improvements that lead to even better perfor" + } + ] + } + ], + "index": 31 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18062" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 69, + 178, + 191 + ], + "blocks": [ + { + "bbox": [ + 56, + 69, + 178, + 191 + ], + "lines": [ + { + "bbox": [ + 56, + 69, + 178, + 191 + ], + "spans": [ + { + "bbox": [ + 56, + 69, + 178, + 191 + ], + "type": "image", + "image_path": "b18a0f4d018046d9ab7155f067e3df4b7341ad15d635e2a00a6cbc6332bde9c2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 146, + 199, + 463, + 211 + ], + "lines": [ + { + "bbox": [ + 146, + 199, + 463, + 211 + ], + "spans": [ + { + "bbox": [ + 146, + 199, + 463, + 211 + ], + "type": "text", + "content": "Figure 2. Some text-to-image samples of " + }, + { + "bbox": [ + 146, + 199, + 463, + 211 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 146, + 199, + 463, + 211 + ], + "type": "text", + "content": " that have been generated by SiD2." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 181, + 69, + 303, + 191 + ], + "blocks": [ + { + "bbox": [ + 181, + 69, + 303, + 191 + ], + "lines": [ + { + "bbox": [ + 181, + 69, + 303, + 191 + ], + "spans": [ + { + "bbox": [ + 181, + 69, + 303, + 191 + ], + "type": "image", + "image_path": "4673782a530a2ffca3a5a3a1f6c3ddfa3d65cdd0850dc8041f8a3798bc066625.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 307, + 70, + 429, + 191 + ], + "blocks": [ + { + "bbox": [ + 307, + 70, + 429, + 191 + ], + "lines": [ + { + "bbox": [ + 307, + 70, + 429, + 191 + ], + "spans": [ + { + "bbox": [ + 307, + 70, + 429, + 191 + ], + "type": "image", + "image_path": "9b3e67cdffaed8395cf0f73b870516817a0a09d2e7eec74fb72d26e6b38056ce.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 433, + 70, + 553, + 191 + ], + "blocks": [ + { + "bbox": [ + 433, + 70, + 553, + 191 + ], + "lines": [ + { + "bbox": [ + 433, + 70, + 553, + 191 + ], + "spans": [ + { + "bbox": [ + 433, + 70, + 553, + 191 + ], + "type": "image", + "image_path": "2e50436e0cea30003bd97c6bc97008729d7dedcd626988278a1604a13b1e4d6b.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 232, + 296, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 232, + 296, + 304 + ], + "spans": [ + { + "bbox": [ + 55, + 232, + 296, + 304 + ], + "type": "text", + "content": "mance, namely flop heavy scaling and removing block skip-connections (Residual U-ViTs). Most notably, we improve the state-of-the-art of pixel-space diffusion from 2.65 to 1.48 on ImageNet512 (see also Figure 1). Further, we also achieve overall state-of-the-art on ImageNet128, ImageNet256 and Kinetics600 without heavy tuning." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 312, + 135, + 324 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 312, + 135, + 324 + ], + "spans": [ + { + "bbox": [ + 55, + 312, + 135, + 324 + ], + "type": "text", + "content": "1.1. Background" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 329, + 296, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 329, + 296, + 353 + ], + "spans": [ + { + "bbox": [ + 55, + 329, + 296, + 353 + ], + "type": "text", + "content": "A diffusion process gradually destroys data over a conceptual time " + }, + { + "bbox": [ + 55, + 329, + 296, + 353 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 329, + 296, + 353 + ], + "type": "text", + "content": ", typically with Gaussian noise as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 97, + 363, + 294, + 376 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 363, + 294, + 376 + ], + "spans": [ + { + "bbox": [ + 97, + 363, + 294, + 376 + ], + "type": "interline_equation", + "content": "\\boldsymbol {z} _ {t} = \\alpha_ {t} \\boldsymbol {x} + \\sigma_ {t} \\boldsymbol {\\epsilon}, \\text {w h e r e} \\boldsymbol {\\epsilon} \\sim \\mathcal {N} (0, \\mathbf {I}), \\tag {1}", + "image_path": "022a722ca49ae4501aafe676f8bae04ef71dc4a873e0f6cd48d4c2c2e8a19169.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 385, + 296, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 385, + 296, + 553 + ], + "spans": [ + { + "bbox": [ + 55, + 385, + 296, + 553 + ], + "type": "text", + "content": "which defines a distribution " + }, + { + "bbox": [ + 55, + 385, + 296, + 553 + ], + "type": "inline_equation", + "content": "q(\\pmb{z}_t|\\pmb{x})" + }, + { + "bbox": [ + 55, + 385, + 296, + 553 + ], + "type": "text", + "content": " such that " + }, + { + "bbox": [ + 55, + 385, + 296, + 553 + ], + "type": "inline_equation", + "content": "\\pmb{z}_0" + }, + { + "bbox": [ + 55, + 385, + 296, + 553 + ], + "type": "text", + "content": " is close to clean data and " + }, + { + "bbox": [ + 55, + 385, + 296, + 553 + ], + "type": "inline_equation", + "content": "\\pmb{z}_1" + }, + { + "bbox": [ + 55, + 385, + 296, + 553 + ], + "type": "text", + "content": " is close to Gaussian noise. In this paper we assume a variance-preserving [41] schedule, which is defined as " + }, + { + "bbox": [ + 55, + 385, + 296, + 553 + ], + "type": "inline_equation", + "content": "\\alpha_{t} = \\sqrt{\\sigma(\\lambda_{t})}" + }, + { + "bbox": [ + 55, + 385, + 296, + 553 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 385, + 296, + 553 + ], + "type": "inline_equation", + "content": "\\sigma_t = \\sqrt{\\sigma(-\\lambda_t)}" + }, + { + "bbox": [ + 55, + 385, + 296, + 553 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 55, + 385, + 296, + 553 + ], + "type": "inline_equation", + "content": "\\lambda_{t} = \\log \\left(\\alpha_{t}^{2} / \\sigma_{t}^{2}\\right)" + }, + { + "bbox": [ + 55, + 385, + 296, + 553 + ], + "type": "text", + "content": " being the log signal-to-noise ratio, which monotonically decreases with time as " + }, + { + "bbox": [ + 55, + 385, + 296, + 553 + ], + "type": "inline_equation", + "content": "t\\to 1" + }, + { + "bbox": [ + 55, + 385, + 296, + 553 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 385, + 296, + 553 + ], + "type": "inline_equation", + "content": "\\sigma (\\cdot)" + }, + { + "bbox": [ + 55, + 385, + 296, + 553 + ], + "type": "text", + "content": " being the sigmoid function. It can be shown that other types of schedules such as the variance-exploding schedule or the flow matching schedule [28] can always be converted to a variance-preserving schedule with a linear scaling of " + }, + { + "bbox": [ + 55, + 385, + 296, + 553 + ], + "type": "inline_equation", + "content": "\\pmb{z}_{t}" + }, + { + "bbox": [ + 55, + 385, + 296, + 553 + ], + "type": "text", + "content": " [25]. The key defining property of a noise schedule is the log signal-to-noise ratio " + }, + { + "bbox": [ + 55, + 385, + 296, + 553 + ], + "type": "inline_equation", + "content": "\\lambda_{t}" + }, + { + "bbox": [ + 55, + 385, + 296, + 553 + ], + "type": "text", + "content": ". A neural network is used to predict " + }, + { + "bbox": [ + 55, + 385, + 296, + 553 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{x}} = \\hat{\\pmb{x}}_{\\theta}(\\pmb{z}_{t},t)" + }, + { + "bbox": [ + 55, + 385, + 296, + 553 + ], + "type": "text", + "content": " that effectively estimates " + }, + { + "bbox": [ + 55, + 385, + 296, + 553 + ], + "type": "inline_equation", + "content": "\\mathbb{E}[\\pmb {x}|\\pmb {z}_t]" + }, + { + "bbox": [ + 55, + 385, + 296, + 553 + ], + "type": "text", + "content": ", which can be done by minimizing a weighted mean squared error loss:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 91, + 562, + 294, + 585 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 562, + 294, + 585 + ], + "spans": [ + { + "bbox": [ + 91, + 562, + 294, + 585 + ], + "type": "interline_equation", + "content": "L (\\boldsymbol {x}) = \\mathbb {E} _ {t \\sim \\mathcal {U} (0, 1)} - \\frac {\\mathrm {d} \\lambda_ {t}}{\\mathrm {d} t} w \\left(\\lambda_ {t}\\right) \\| \\boldsymbol {x} - \\hat {\\boldsymbol {x}} \\| ^ {2}, \\tag {2}", + "image_path": "fd5e575aa8bbc0bad257e6103d3e8b574d052087dec8662e0146a2a9145ffa84.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 594, + 295, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 689 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 594, + 295, + 689 + ], + "type": "inline_equation", + "content": "w(\\lambda_t)" + }, + { + "bbox": [ + 55, + 594, + 295, + 689 + ], + "type": "text", + "content": " is the weighting function. " + }, + { + "bbox": [ + 55, + 594, + 295, + 689 + ], + "type": "inline_equation", + "content": "-\\mathrm{d}\\lambda_{t} / \\mathrm{d}t" + }, + { + "bbox": [ + 55, + 594, + 295, + 689 + ], + "type": "text", + "content": " cancels out the effect of the noise schedule " + }, + { + "bbox": [ + 55, + 594, + 295, + 689 + ], + "type": "inline_equation", + "content": "\\lambda_{t}" + }, + { + "bbox": [ + 55, + 594, + 295, + 689 + ], + "type": "text", + "content": " on the loss. In other words, " + }, + { + "bbox": [ + 55, + 594, + 295, + 689 + ], + "type": "inline_equation", + "content": "L(\\mathbf{x})" + }, + { + "bbox": [ + 55, + 594, + 295, + 689 + ], + "type": "text", + "content": " can be rewritten as " + }, + { + "bbox": [ + 55, + 594, + 295, + 689 + ], + "type": "inline_equation", + "content": "\\int -w(\\lambda)\\| \\pmb {x} - \\hat{\\pmb{x}}\\| ^2 d\\lambda" + }, + { + "bbox": [ + 55, + 594, + 295, + 689 + ], + "type": "text", + "content": " that is invariant to the choice of " + }, + { + "bbox": [ + 55, + 594, + 295, + 689 + ], + "type": "inline_equation", + "content": "\\lambda_{t}" + }, + { + "bbox": [ + 55, + 594, + 295, + 689 + ], + "type": "text", + "content": ". However, " + }, + { + "bbox": [ + 55, + 594, + 295, + 689 + ], + "type": "inline_equation", + "content": "\\lambda_{t}" + }, + { + "bbox": [ + 55, + 594, + 295, + 689 + ], + "type": "text", + "content": " still impacts the sampling efficiency of the Monte Carlo estimator of the expected loss, and thus affects the training dynamics of diffusion models. This decouples noise schedule " + }, + { + "bbox": [ + 55, + 594, + 295, + 689 + ], + "type": "inline_equation", + "content": "\\lambda_{t}" + }, + { + "bbox": [ + 55, + 594, + 295, + 689 + ], + "type": "text", + "content": " and weighting function " + }, + { + "bbox": [ + 55, + 594, + 295, + 689 + ], + "type": "inline_equation", + "content": "w(\\lambda)" + }, + { + "bbox": [ + 55, + 594, + 295, + 689 + ], + "type": "text", + "content": " as a function of " + }, + { + "bbox": [ + 55, + 594, + 295, + 689 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 594, + 295, + 689 + ], + "type": "text", + "content": " [25]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "content": "The above squared error loss can be derived from the variational interpretation of diffusion models [26]. Given" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 232, + 555, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 232, + 555, + 281 + ], + "spans": [ + { + "bbox": [ + 313, + 232, + 555, + 281 + ], + "type": "text", + "content": "two time steps " + }, + { + "bbox": [ + 313, + 232, + 555, + 281 + ], + "type": "inline_equation", + "content": "s < t" + }, + { + "bbox": [ + 313, + 232, + 555, + 281 + ], + "type": "text", + "content": ", we define the denoising distribution " + }, + { + "bbox": [ + 313, + 232, + 555, + 281 + ], + "type": "inline_equation", + "content": "p(\\boldsymbol{z}_s|\\boldsymbol{z}_t)" + }, + { + "bbox": [ + 313, + 232, + 555, + 281 + ], + "type": "text", + "content": " in the same formulation as the true posterior " + }, + { + "bbox": [ + 313, + 232, + 555, + 281 + ], + "type": "inline_equation", + "content": "q(\\boldsymbol{z}_s|\\boldsymbol{z}_t,\\boldsymbol{x})" + }, + { + "bbox": [ + 313, + 232, + 555, + 281 + ], + "type": "text", + "content": ", but replace the ground truth " + }, + { + "bbox": [ + 313, + 232, + 555, + 281 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}" + }, + { + "bbox": [ + 313, + 232, + 555, + 281 + ], + "type": "text", + "content": " with the learned approximation " + }, + { + "bbox": [ + 313, + 232, + 555, + 281 + ], + "type": "inline_equation", + "content": "\\hat{\\boldsymbol{x}} \\approx \\mathbb{E}[\\boldsymbol{x}|\\boldsymbol{z}_t]" + }, + { + "bbox": [ + 313, + 232, + 555, + 281 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 336, + 289, + 553, + 302 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 336, + 289, + 553, + 302 + ], + "spans": [ + { + "bbox": [ + 336, + 289, + 553, + 302 + ], + "type": "interline_equation", + "content": "p \\left(\\boldsymbol {z} _ {s} \\mid \\boldsymbol {z} _ {t}\\right) = q \\left(\\boldsymbol {z} _ {s} \\mid \\boldsymbol {z} _ {t}, \\boldsymbol {x} = \\hat {\\boldsymbol {x}}\\right) = \\mathcal {N} \\left(\\boldsymbol {z} _ {s} \\mid \\mu_ {t \\rightarrow s}, \\sigma_ {t \\rightarrow s}\\right), \\tag {3}", + "image_path": "60815a642b975a0e673239a2dd600683dd527b6014f5c7a1bc7b6b42e5ec4bee.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 312, + 554, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 312, + 554, + 410 + ], + "spans": [ + { + "bbox": [ + 313, + 312, + 554, + 410 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 312, + 554, + 410 + ], + "type": "inline_equation", + "content": "\\sigma_{t\\rightarrow s}^2 = \\left(1 / \\sigma_s^2 +\\alpha_{t|s}^2 /\\sigma_{t|s}^2\\right)^{-1}" + }, + { + "bbox": [ + 313, + 312, + 554, + 410 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 312, + 554, + 410 + ], + "type": "inline_equation", + "content": "\\mu_{t\\to s} = \\sigma_{t\\to s}^{2}\\Bigl (\\alpha_{t|s} / \\sigma_{t|s}^{2}\\pmb {z}_{t} + \\alpha_{s} / \\sigma_{s}^{2}\\pmb {x}\\Bigr)" + }, + { + "bbox": [ + 313, + 312, + 554, + 410 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 313, + 312, + 554, + 410 + ], + "type": "inline_equation", + "content": "\\alpha_{t|s} = \\alpha_t / \\alpha_s" + }, + { + "bbox": [ + 313, + 312, + 554, + 410 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 312, + 554, + 410 + ], + "type": "inline_equation", + "content": "\\sigma_{t|s}^2 = \\sigma_t^2 -\\alpha_{t|s}\\sigma_s^2" + }, + { + "bbox": [ + 313, + 312, + 554, + 410 + ], + "type": "text", + "content": " . Note that during sampling, the standard deviation for " + }, + { + "bbox": [ + 313, + 312, + 554, + 410 + ], + "type": "inline_equation", + "content": "p(z_{s}|z_{t})" + }, + { + "bbox": [ + 313, + 312, + 554, + 410 + ], + "type": "text", + "content": " is chosen to be a log-linear interpolation between " + }, + { + "bbox": [ + 313, + 312, + 554, + 410 + ], + "type": "inline_equation", + "content": "\\sigma_{t|s}" + }, + { + "bbox": [ + 313, + 312, + 554, + 410 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 312, + 554, + 410 + ], + "type": "inline_equation", + "content": "\\sigma_{t\\rightarrow s}" + }, + { + "bbox": [ + 313, + 312, + 554, + 410 + ], + "type": "text", + "content": " of the form " + }, + { + "bbox": [ + 313, + 312, + 554, + 410 + ], + "type": "inline_equation", + "content": "\\sigma_{t\\rightarrow s}^{\\gamma}\\sigma_{t|s}^{1 - \\gamma}" + }, + { + "bbox": [ + 313, + 312, + 554, + 410 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 313, + 312, + 554, + 410 + ], + "type": "inline_equation", + "content": "\\gamma = 0.3" + }, + { + "bbox": [ + 313, + 312, + 554, + 410 + ], + "type": "text", + "content": " was set empirically, which is important as FID is sensitive to over-smooth or grainy images." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 422, + 421, + 435 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 422, + 421, + 435 + ], + "spans": [ + { + "bbox": [ + 313, + 422, + 421, + 435 + ], + "type": "text", + "content": "2. Simpler Diffusion" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 442, + 460, + 455 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 442, + 460, + 455 + ], + "spans": [ + { + "bbox": [ + 313, + 442, + 460, + 455 + ], + "type": "text", + "content": "2.1. Revisiting the sigmoid loss" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 460, + 555, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 460, + 555, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 460, + 555, + 605 + ], + "type": "text", + "content": "Recall that one of the proposed losses in VDM++ [25] is the sigmoid loss. They expressed the loss with respect to " + }, + { + "bbox": [ + 313, + 460, + 555, + 605 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 313, + 460, + 555, + 605 + ], + "type": "text", + "content": "-mse: " + }, + { + "bbox": [ + 313, + 460, + 555, + 605 + ], + "type": "inline_equation", + "content": "L(\\pmb{x}) = \\mathbb{E}_{t \\sim \\mathcal{U}(0,1)} - \\frac{\\mathrm{d}\\lambda_t}{\\mathrm{d}t} \\tilde{w}(\\lambda_t) \\| \\pmb{\\epsilon} - \\hat{\\pmb{\\epsilon}} \\|^2" + }, + { + "bbox": [ + 313, + 460, + 555, + 605 + ], + "type": "text", + "content": ", which is connected to " + }, + { + "bbox": [ + 313, + 460, + 555, + 605 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 313, + 460, + 555, + 605 + ], + "type": "text", + "content": "-mse through " + }, + { + "bbox": [ + 313, + 460, + 555, + 605 + ], + "type": "inline_equation", + "content": "\\| \\pmb{\\epsilon} - \\hat{\\pmb{\\epsilon}} \\|^2 = \\exp(\\lambda_t) \\| \\pmb{x} - \\hat{\\pmb{x}} \\|^2" + }, + { + "bbox": [ + 313, + 460, + 555, + 605 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 460, + 555, + 605 + ], + "type": "inline_equation", + "content": "\\tilde{w}(\\lambda_t) = \\exp(\\lambda_t) w(\\lambda_t)" + }, + { + "bbox": [ + 313, + 460, + 555, + 605 + ], + "type": "text", + "content": ". In VDM++, viewing the loss in " + }, + { + "bbox": [ + 313, + 460, + 555, + 605 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 313, + 460, + 555, + 605 + ], + "type": "text", + "content": "-mse enables a natural connection between the weighted loss and the ELBO and showed that losses should be monotonically decreasing on " + }, + { + "bbox": [ + 313, + 460, + 555, + 605 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 313, + 460, + 555, + 605 + ], + "type": "text", + "content": "-mse. Hence, they propose the sigmoid loss: " + }, + { + "bbox": [ + 313, + 460, + 555, + 605 + ], + "type": "inline_equation", + "content": "\\tilde{w}(\\lambda_t) = \\sigma(b - \\lambda_t)" + }, + { + "bbox": [ + 313, + 460, + 555, + 605 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 460, + 555, + 605 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 313, + 460, + 555, + 605 + ], + "type": "text", + "content": " denotes the bias or shift and " + }, + { + "bbox": [ + 313, + 460, + 555, + 605 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 313, + 460, + 555, + 605 + ], + "type": "text", + "content": " is the sigmoid function. However, they ultimately opted for another loss-weighting named EDM-monotonic for high resolution image generation, see Figure 3a." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 622, + 555, + 670 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 622, + 555, + 670 + ], + "spans": [ + { + "bbox": [ + 313, + 622, + 555, + 670 + ], + "type": "text", + "content": "Enforcing weighting monotonicity in " + }, + { + "bbox": [ + 313, + 622, + 555, + 670 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 622, + 555, + 670 + ], + "type": "text", + "content": " Revisiting the sigmoid loss, we derive its " + }, + { + "bbox": [ + 313, + 622, + 555, + 670 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 622, + 555, + 670 + ], + "type": "text", + "content": "-mse form, which turns out to conveniently be defined by a mirrored sigmoid (albeit multiplied by a constant factor):" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 325, + 680, + 553, + 710 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 680, + 553, + 710 + ], + "spans": [ + { + "bbox": [ + 325, + 680, + 553, + 710 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\sigma \\left(b - \\lambda_ {t}\\right) \\| \\boldsymbol {\\epsilon} - \\hat {\\boldsymbol {\\epsilon}} \\| ^ {2} = \\sigma \\left(b - \\lambda_ {t}\\right) \\exp \\left(\\lambda_ {t}\\right) \\| \\boldsymbol {x} - \\hat {\\boldsymbol {x}} \\| ^ {2} \\tag {4} \\\\ = \\exp (b) \\sigma \\left(\\lambda_ {t} - b\\right) \\| \\boldsymbol {x} - \\hat {\\boldsymbol {x}} \\| ^ {2}. \\\\ \\end{array}", + "image_path": "04ff4eb9d020bdd4f121870c7a7a3d3881e338dfd90511f31953fdf944ff26a1.jpg" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18063" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 89, + 83, + 269, + 195 + ], + "blocks": [ + { + "bbox": [ + 89, + 83, + 269, + 195 + ], + "lines": [ + { + "bbox": [ + 89, + 83, + 269, + 195 + ], + "spans": [ + { + "bbox": [ + 89, + 83, + 269, + 195 + ], + "type": "image", + "image_path": "12bc20bf77af3c4d4311affe192124ed53e769feeff62a5c95055d6236936202.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 146, + 195, + 219, + 203 + ], + "lines": [ + { + "bbox": [ + 146, + 195, + 219, + 203 + ], + "spans": [ + { + "bbox": [ + 146, + 195, + 219, + 203 + ], + "type": "text", + "content": "(a) Weighting over " + }, + { + "bbox": [ + 146, + 195, + 219, + 203 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 146, + 195, + 219, + 203 + ], + "type": "text", + "content": "-mse" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 334, + 83, + 512, + 195 + ], + "blocks": [ + { + "bbox": [ + 334, + 83, + 512, + 195 + ], + "lines": [ + { + "bbox": [ + 334, + 83, + 512, + 195 + ], + "spans": [ + { + "bbox": [ + 334, + 83, + 512, + 195 + ], + "type": "image", + "image_path": "2c131284de18ff2945917ee8d0833c6474829f5f2450df0baff16c2b498b7be2.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 390, + 195, + 465, + 203 + ], + "lines": [ + { + "bbox": [ + 390, + 195, + 465, + 203 + ], + "spans": [ + { + "bbox": [ + 390, + 195, + 465, + 203 + ], + "type": "text", + "content": "(b) Weighting over " + }, + { + "bbox": [ + 390, + 195, + 465, + 203 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 390, + 195, + 465, + 203 + ], + "type": "text", + "content": "-mse" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 78, + 213, + 531, + 224 + ], + "lines": [ + { + "bbox": [ + 78, + 213, + 531, + 224 + ], + "spans": [ + { + "bbox": [ + 78, + 213, + 531, + 224 + ], + "type": "text", + "content": "Figure 3. Weighting of different losses in both " + }, + { + "bbox": [ + 78, + 213, + 531, + 224 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 78, + 213, + 531, + 224 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 78, + 213, + 531, + 224 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 78, + 213, + 531, + 224 + ], + "type": "text", + "content": " space. Note that the weightings are scaled to have a maximum of one." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 231, + 295, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 231, + 295, + 374 + ], + "spans": [ + { + "bbox": [ + 54, + 231, + 295, + 374 + ], + "type": "text", + "content": "This reveals that the sigmoid loss is not only monotonically decreasing in " + }, + { + "bbox": [ + 54, + 231, + 295, + 374 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 54, + 231, + 295, + 374 + ], + "type": "text", + "content": "-mse, it is also monotonically increasing in " + }, + { + "bbox": [ + 54, + 231, + 295, + 374 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 54, + 231, + 295, + 374 + ], + "type": "text", + "content": "-mse. We claim that a reasonable weighting function " + }, + { + "bbox": [ + 54, + 231, + 295, + 374 + ], + "type": "inline_equation", + "content": "w(\\lambda_t)" + }, + { + "bbox": [ + 54, + 231, + 295, + 374 + ], + "type": "text", + "content": " should be monotonically increasing over " + }, + { + "bbox": [ + 54, + 231, + 295, + 374 + ], + "type": "inline_equation", + "content": "\\lambda_t" + }, + { + "bbox": [ + 54, + 231, + 295, + 374 + ], + "type": "text", + "content": ". Observe that with less noise in " + }, + { + "bbox": [ + 54, + 231, + 295, + 374 + ], + "type": "inline_equation", + "content": "z_t" + }, + { + "bbox": [ + 54, + 231, + 295, + 374 + ], + "type": "text", + "content": ", there is more information available and it should be strictly easier to predict the original data, meaning that the " + }, + { + "bbox": [ + 54, + 231, + 295, + 374 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 54, + 231, + 295, + 374 + ], + "type": "text", + "content": "-mse generally decreases for higher " + }, + { + "bbox": [ + 54, + 231, + 295, + 374 + ], + "type": "inline_equation", + "content": "\\lambda_t" + }, + { + "bbox": [ + 54, + 231, + 295, + 374 + ], + "type": "text", + "content": ". This phenomenon is also observed in practice. If the loss weighting in " + }, + { + "bbox": [ + 54, + 231, + 295, + 374 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 54, + 231, + 295, + 374 + ], + "type": "text", + "content": "-mse also decreases, the model would be able to make worse predictions while obtaining less noisy (better) conditioning information and still achieve a lower loss. Simply put: Why put less weight on an easier problem?" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 374, + 296, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 374, + 296, + 460 + ], + "spans": [ + { + "bbox": [ + 55, + 374, + 296, + 460 + ], + "type": "text", + "content": "Studying both the form of the sigmoid and EDM-monotonic weighting functions in Figure 3, we observe that both weighting functions are monotonically decreasing in " + }, + { + "bbox": [ + 55, + 374, + 296, + 460 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 55, + 374, + 296, + 460 + ], + "type": "text", + "content": "-mse, but only the sigmoid weighting is monotonically increasing in " + }, + { + "bbox": [ + 55, + 374, + 296, + 460 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 374, + 296, + 460 + ], + "type": "text", + "content": "-mse. Therefore, we expect the sigmoid weighting to be a better candidate to scale to high resolution generation than the EDM-monotonic version." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 472, + 296, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 472, + 296, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 472, + 296, + 605 + ], + "type": "text", + "content": "Loss shifting Recall that for a resolution that is two times higher in both height and width, an average (low pass) filter would see " + }, + { + "bbox": [ + 55, + 472, + 296, + 605 + ], + "type": "inline_equation", + "content": "\\mathrm{Var}[(x_{11} + x_{12} + x_{21} + x_{22}) / 4] = 1 / 4" + }, + { + "bbox": [ + 55, + 472, + 296, + 605 + ], + "type": "text", + "content": " of the noise, while signal strength stays the same [14, 25]. Hence, its effective logsnr is expected to increase by about " + }, + { + "bbox": [ + 55, + 472, + 296, + 605 + ], + "type": "inline_equation", + "content": "2\\log 2\\approx 1.386\\approx 1.5" + }, + { + "bbox": [ + 55, + 472, + 296, + 605 + ], + "type": "text", + "content": " when viewed through a low-pass filter. Suppose that the optimal setting for images of " + }, + { + "bbox": [ + 55, + 472, + 296, + 605 + ], + "type": "inline_equation", + "content": "128\\times 128" + }, + { + "bbox": [ + 55, + 472, + 296, + 605 + ], + "type": "text", + "content": " is a sigmoid weighting with " + }, + { + "bbox": [ + 55, + 472, + 296, + 605 + ], + "type": "inline_equation", + "content": "b\\approx 0" + }, + { + "bbox": [ + 55, + 472, + 296, + 605 + ], + "type": "text", + "content": ", then one would expect the optimal setting for images of " + }, + { + "bbox": [ + 55, + 472, + 296, + 605 + ], + "type": "inline_equation", + "content": "512\\times 512" + }, + { + "bbox": [ + 55, + 472, + 296, + 605 + ], + "type": "text", + "content": " to be roughly " + }, + { + "bbox": [ + 55, + 472, + 296, + 605 + ], + "type": "inline_equation", + "content": "b = -3" + }, + { + "bbox": [ + 55, + 472, + 296, + 605 + ], + "type": "text", + "content": ".1 Empirically we find that slightly more aggressive shifts perform even better." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 611, + 165, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 611, + 165, + 624 + ], + "spans": [ + { + "bbox": [ + 55, + 611, + 165, + 624 + ], + "type": "text", + "content": "2.2. Flop heavy scaling" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 628, + 295, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 628, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 628, + 295, + 677 + ], + "type": "text", + "content": "When one deals with a limited data budget, it can be difficult to balance scaling of models and regularization. Our small model version at " + }, + { + "bbox": [ + 55, + 628, + 295, + 677 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 55, + 628, + 295, + 677 + ], + "type": "text", + "content": " resolution starts out with " + }, + { + "bbox": [ + 55, + 628, + 295, + 677 + ], + "type": "inline_equation", + "content": "4 \\times 4" + }, + { + "bbox": [ + 55, + 628, + 295, + 677 + ], + "type": "text", + "content": " patching, and subsequently runs a U-ViT [14] on the " + }, + { + "bbox": [ + 55, + 628, + 295, + 677 + ], + "type": "inline_equation", + "content": "128 \\times" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 231, + 555, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 231, + 555, + 434 + ], + "spans": [ + { + "bbox": [ + 313, + 231, + 555, + 434 + ], + "type": "inline_equation", + "content": "128 \\times 48" + }, + { + "bbox": [ + 313, + 231, + 555, + 434 + ], + "type": "text", + "content": " input (assuming the raw input is of 3 channels). To make the model more expressive, one of the most common options is to double the number of channels for every layer, leading to " + }, + { + "bbox": [ + 313, + 231, + 555, + 434 + ], + "type": "inline_equation", + "content": "2 \\times" + }, + { + "bbox": [ + 313, + 231, + 555, + 434 + ], + "type": "text", + "content": " model size. However, our experiments show that this results in diminishing returns. We hypothesize that it is because the model suffers from over-fitting when trained with limited data budget such as ImageNet. Indeed, when we train using distribution augmentation [18], the best performance improves from 1.9 to 1.6. This however suggests another way to scale the model: Keep the architecture and parameter count almost the same and instead run on " + }, + { + "bbox": [ + 313, + 231, + 555, + 434 + ], + "type": "inline_equation", + "content": "2 \\times 2" + }, + { + "bbox": [ + 313, + 231, + 555, + 434 + ], + "type": "text", + "content": " patched input (so " + }, + { + "bbox": [ + 313, + 231, + 555, + 434 + ], + "type": "inline_equation", + "content": "256 \\times 256 \\times 12" + }, + { + "bbox": [ + 313, + 231, + 555, + 434 + ], + "type": "text", + "content": " for 512 inputs). Similar to " + }, + { + "bbox": [ + 313, + 231, + 555, + 434 + ], + "type": "inline_equation", + "content": "2 \\times" + }, + { + "bbox": [ + 313, + 231, + 555, + 434 + ], + "type": "text", + "content": " model size, reducing the patching size also increases FLOPS by roughly a factor of 4. However, we found this model performs even better to about 1.5 without the need of distribution augmentation. We call this model flop heavy, because it has a much higher flop over parameter ratio." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 435, + 556, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 435, + 556, + 615 + ], + "spans": [ + { + "bbox": [ + 313, + 435, + 556, + 615 + ], + "type": "text", + "content": "Flop heavy scaling has another interesting property. Suppose we have a small model trained to fit input of size " + }, + { + "bbox": [ + 313, + 435, + 556, + 615 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 313, + 435, + 556, + 615 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 313, + 435, + 556, + 615 + ], + "type": "inline_equation", + "content": "2 \\times 2" + }, + { + "bbox": [ + 313, + 435, + 556, + 615 + ], + "type": "text", + "content": " patching. Using the flop heavy scaling at " + }, + { + "bbox": [ + 313, + 435, + 556, + 615 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 313, + 435, + 556, + 615 + ], + "type": "text", + "content": " resolution enables us to directly finetune from the small model trained at " + }, + { + "bbox": [ + 313, + 435, + 556, + 615 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 313, + 435, + 556, + 615 + ], + "type": "text", + "content": " with exactly the same amount of parameters, whereas its counterpart version at " + }, + { + "bbox": [ + 313, + 435, + 556, + 615 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 313, + 435, + 556, + 615 + ], + "type": "text", + "content": " resolution and with " + }, + { + "bbox": [ + 313, + 435, + 556, + 615 + ], + "type": "inline_equation", + "content": "4 \\times 4" + }, + { + "bbox": [ + 313, + 435, + 556, + 615 + ], + "type": "text", + "content": " patching needs to introduce additional parameters to handle the more input channels. This method of finetuning to higher resolutions has already been popularized in the context of text-to-image generation [32]. Here we find that flop heavy scaling is a more favorable way to scale on smaller datasets such as ImageNet. Of course, in principle one could also train a model with more channels and " + }, + { + "bbox": [ + 313, + 435, + 556, + 615 + ], + "type": "inline_equation", + "content": "4 \\times 4" + }, + { + "bbox": [ + 313, + 435, + 556, + 615 + ], + "type": "text", + "content": " patching on a lower resolution, but this is not done in practice to train 'small' models." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 623, + 555, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 623, + 555, + 647 + ], + "spans": [ + { + "bbox": [ + 313, + 623, + 555, + 647 + ], + "type": "text", + "content": "2.3. Residual U-ViTs: Removing blockwise skipconnections" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 653, + 556, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 556, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 556, + 712 + ], + "type": "text", + "content": "U-Nets require that sub-sampled levels (for example via average pooling) are bypassed by skip-connections. However, this makes the architecture very different from large-scale language models such as transformers, where these skip-connections are not used and residual connections suffice." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 55, + 683, + 295, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 683, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 683, + 295, + 713 + ], + "type": "text", + "content": "This holds under the condition that the most important information is already contained by the " + }, + { + "bbox": [ + 55, + 683, + 295, + 713 + ], + "type": "inline_equation", + "content": "128 \\times 128" + }, + { + "bbox": [ + 55, + 683, + 295, + 713 + ], + "type": "text", + "content": " image, and this may for example not hold for images without global structure, e.g. that only contain textures." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18064" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "text", + "content": "In this paper we use the U-ViT model of [14], which is a variant of U-Nets with self-attention layers (with convolutional layers) being replaced by transformer blocks. Hence, that part of the architecture is closer to standard transformers." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 121, + 296, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 121, + 296, + 205 + ], + "spans": [ + { + "bbox": [ + 55, + 121, + 296, + 205 + ], + "type": "text", + "content": "In this paper, we further simplify model definition with Residual U-ViTs, see Figure 4. The blockwise skip-connections are removed, and in their place only a single skip-connection is used per downsampling operation. To be precise, let " + }, + { + "bbox": [ + 55, + 121, + 296, + 205 + ], + "type": "inline_equation", + "content": "f_{d}, f_{u}" + }, + { + "bbox": [ + 55, + 121, + 296, + 205 + ], + "type": "text", + "content": " denote a down and up level of a U-ViT and let " + }, + { + "bbox": [ + 55, + 121, + 296, + 205 + ], + "type": "inline_equation", + "content": "f_{m}" + }, + { + "bbox": [ + 55, + 121, + 296, + 205 + ], + "type": "text", + "content": " denote the middle stage. This two-stage Residual U-ViT can now be described by:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 215, + 296, + 245 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 215, + 296, + 245 + ], + "spans": [ + { + "bbox": [ + 55, + 215, + 296, + 245 + ], + "type": "interline_equation", + "content": "f (x) = f _ {u} \\left(\\mathrm {U} \\left(f _ {m} (\\mathrm {D} (h)) - \\mathrm {D} (h)\\right) + h\\right) \\text {w h e r e} h = f _ {d} (x) \\tag {5}", + "image_path": "7efeeec561b04e6707a3d818ab5ceb770a5314c97c4eb1294487e04ac09b0ae4.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 245, + 296, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 245, + 296, + 353 + ], + "spans": [ + { + "bbox": [ + 55, + 245, + 296, + 353 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 245, + 296, + 353 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 55, + 245, + 296, + 353 + ], + "type": "text", + "content": " is typically an average pooling and linear and " + }, + { + "bbox": [ + 55, + 245, + 296, + 353 + ], + "type": "inline_equation", + "content": "U" + }, + { + "bbox": [ + 55, + 245, + 296, + 353 + ], + "type": "text", + "content": " is typically a linear and nearest neighbour upsampling. Defining this network to preserve an identity function at initialization is now much easier: As long as " + }, + { + "bbox": [ + 55, + 245, + 296, + 353 + ], + "type": "inline_equation", + "content": "f_{d}" + }, + { + "bbox": [ + 55, + 245, + 296, + 353 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 245, + 296, + 353 + ], + "type": "inline_equation", + "content": "f_{u}" + }, + { + "bbox": [ + 55, + 245, + 296, + 353 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 245, + 296, + 353 + ], + "type": "inline_equation", + "content": "f_{m}" + }, + { + "bbox": [ + 55, + 245, + 296, + 353 + ], + "type": "text", + "content": " are initialized to be identity, the entire " + }, + { + "bbox": [ + 55, + 245, + 296, + 353 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 55, + 245, + 296, + 353 + ], + "type": "text", + "content": " is also identity at initialization. The generalization to any-depth U-ViTs is also straightforward: we can define another two levels of U-ViTs in place of " + }, + { + "bbox": [ + 55, + 245, + 296, + 353 + ], + "type": "inline_equation", + "content": "f_{m}" + }, + { + "bbox": [ + 55, + 245, + 296, + 353 + ], + "type": "text", + "content": " recursively, where the nested U-ViTs can also be guaranteed to be initialized as an identity function." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 354, + 296, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 354, + 296, + 498 + ], + "spans": [ + { + "bbox": [ + 55, + 354, + 296, + 498 + ], + "type": "text", + "content": "For small networks, it could be helpful to keep the blockwise skip-connections, and expect to incur a slight performance degradation from removing the blockwise skip-connections. Small networks have limited capacity, so limiting skip-connections may create bottleneck issues. Experimentally, this is confirmed for the small architectures, where we observe 0.2 FID performance degradation on ImageNet512 and 0.06 FID performance degradation on ImageNet256. Because of the simplified definition, we choose to accept this performance penalty for small models. Moreover, we find that larger models perform on par or even benefit from removing the blockwise skip-connections." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 498, + 297, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 297, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 297, + 713 + ], + "type": "text", + "content": "The final advantage is memory consumption at evaluation. For example, a feature map at " + }, + { + "bbox": [ + 55, + 498, + 297, + 713 + ], + "type": "inline_equation", + "content": "256^{2}" + }, + { + "bbox": [ + 55, + 498, + 297, + 713 + ], + "type": "text", + "content": " resolution with 128 channels requires " + }, + { + "bbox": [ + 55, + 498, + 297, + 713 + ], + "type": "inline_equation", + "content": "256^{2} \\cdot 128 \\cdot 2 / 1024^{2} = 16\\mathrm{MB}" + }, + { + "bbox": [ + 55, + 498, + 297, + 713 + ], + "type": "text", + "content": ". With 3 blocks per level, this requires holding an extra " + }, + { + "bbox": [ + 55, + 498, + 297, + 713 + ], + "type": "inline_equation", + "content": "3 \\cdot 16 = 48\\mathrm{MB}" + }, + { + "bbox": [ + 55, + 498, + 297, + 713 + ], + "type": "text", + "content": " per example for the skip-connections, whereas the Residual U-ViT only requires holding " + }, + { + "bbox": [ + 55, + 498, + 297, + 713 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 55, + 498, + 297, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 498, + 297, + 713 + ], + "type": "inline_equation", + "content": "D(h)" + }, + { + "bbox": [ + 55, + 498, + 297, + 713 + ], + "type": "text", + "content": " (the latter could be cheaply recomputed), so " + }, + { + "bbox": [ + 55, + 498, + 297, + 713 + ], + "type": "inline_equation", + "content": "16 + 8 = 24\\mathrm{MB}" + }, + { + "bbox": [ + 55, + 498, + 297, + 713 + ], + "type": "text", + "content": " per example while computing the lower levels. For a hypothetical super-resolution model at " + }, + { + "bbox": [ + 55, + 498, + 297, + 713 + ], + "type": "inline_equation", + "content": "4096^{2}" + }, + { + "bbox": [ + 55, + 498, + 297, + 713 + ], + "type": "text", + "content": " resolution, this leads to a difference between " + }, + { + "bbox": [ + 55, + 498, + 297, + 713 + ], + "type": "inline_equation", + "content": "3 \\cdot 4 = 12\\mathrm{GB}" + }, + { + "bbox": [ + 55, + 498, + 297, + 713 + ], + "type": "text", + "content": " and only " + }, + { + "bbox": [ + 55, + 498, + 297, + 713 + ], + "type": "inline_equation", + "content": "4 + 2 = 6\\mathrm{GB}" + }, + { + "bbox": [ + 55, + 498, + 297, + 713 + ], + "type": "text", + "content": ". During training this advantage is usually negligible, because feature maps also need to be retained for the backward pass. A final advantage of Residual U-ViTs is that it is more straightforward to define asymmetric U-ViTs, i.e., with a different number of blocks in the downsampling (encoding) stage from the one in the upsampling (decoding) stage. In certain cases, moving blocks from the downsampling stage to the upsampling stage can increase performance." + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 317, + 70, + 554, + 152 + ], + "blocks": [ + { + "bbox": [ + 317, + 70, + 554, + 152 + ], + "lines": [ + { + "bbox": [ + 317, + 70, + 554, + 152 + ], + "spans": [ + { + "bbox": [ + 317, + 70, + 554, + 152 + ], + "type": "image", + "image_path": "db47be7ec0de83b6399d13043bf84e5cd0faa4c5cc3431cd3fa9f8131560132e.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 160, + 555, + 194 + ], + "lines": [ + { + "bbox": [ + 313, + 160, + 555, + 194 + ], + "spans": [ + { + "bbox": [ + 313, + 160, + 555, + 194 + ], + "type": "text", + "content": "Figure 4. Residual U-ViT with levelwise skip-connections (instead of blockwise). When the network blocks are initialized to be identity, the entire network is also an identity function at initialization." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 209, + 389, + 220 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 209, + 389, + 220 + ], + "spans": [ + { + "bbox": [ + 314, + 209, + 389, + 220 + ], + "type": "text", + "content": "2.4. Distillation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 225, + 555, + 406 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 225, + 555, + 406 + ], + "spans": [ + { + "bbox": [ + 313, + 225, + 555, + 406 + ], + "type": "text", + "content": "Sampling from a pixel-space diffusion model naively often requires more function evaluations than sampling from a latent-space diffusion model. For example, the popular Stable Diffusion model [32] is usually sampled using 20-50 sampling steps, while pixel-space models are sampled with " + }, + { + "bbox": [ + 313, + 225, + 555, + 406 + ], + "type": "inline_equation", + "content": "100+" + }, + { + "bbox": [ + 313, + 225, + 555, + 406 + ], + "type": "text", + "content": " steps [14]. To reduce the number of sampling steps, we could use either more advanced solvers or distillation. Recently distillation methods show that pixel-space models can be adapted into few-step samplers that maintain high sample quality [23, 29, 36, 44, 45, 47]. In this work, we use moment matching distillation [37] which is currently one of the best distillation methods available in the few step regime. We distill our pixel-space models for use with 8-16 sampling steps, which empirically we find maintains or even improves the sample quality of our models." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 415, + 400, + 428 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 415, + 400, + 428 + ], + "spans": [ + { + "bbox": [ + 314, + 415, + 400, + 428 + ], + "type": "text", + "content": "3. Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 435, + 555, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 435, + 555, + 483 + ], + "spans": [ + { + "bbox": [ + 313, + 435, + 555, + 483 + ], + "type": "text", + "content": "The first end-to-end diffusion work that directly generated high resolution images was [4]. Given the poor performance of end-to-end models, two popular alternatives were introduced: cascaded diffusion [12] and latent diffusion [32]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 483, + 555, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 483, + 555, + 555 + ], + "spans": [ + { + "bbox": [ + 313, + 483, + 555, + 555 + ], + "type": "text", + "content": " Cascaded diffusion approaches have been shown to scale to large datasets and models to generate high quality and high resolution imagery, albeit in multiple stages [12, 40, 42, 52], or with fine-tuning [43]. Although the focus of this paper has been single stage generation up to a resolution of " + }, + { + "bbox": [ + 313, + 483, + 555, + 555 + ], + "type": "inline_equation", + "content": "10^{24^2}" + }, + { + "bbox": [ + 313, + 483, + 555, + 555 + ], + "type": "text", + "content": ", the loss weightings may also be used for cascaded models." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 555, + 556, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 555, + 556, + 663 + ], + "spans": [ + { + "bbox": [ + 313, + 555, + 556, + 663 + ], + "type": "text", + "content": "Significant improvements were made to end-to-end diffusion by shifting noise schedules [3, 14] and subsequently loss weightings [25]. Other approaches such as multi-resolution losses (and possibly architecture adaptions) [6, 7, 14] added marginal gains. However, the current best known latent model EDM2 [20] outperforms these prior end-to-end approaches by a large margin. In contrast, our models are competitive with latent models, outperforming DiT-XL/2 [31] and only performing slightly worse than EDM2." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 677, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 556, + 713 + ], + "type": "text", + "content": "Skip connections The U-Net architecture is a fully convolutional network with long-range skip-connections introduced for (medical) image segmentation [33, 39]. The main" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18065" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 68, + 301, + 216 + ], + "blocks": [ + { + "bbox": [ + 56, + 68, + 301, + 216 + ], + "lines": [ + { + "bbox": [ + 56, + 68, + 301, + 216 + ], + "spans": [ + { + "bbox": [ + 56, + 68, + 301, + 216 + ], + "type": "image", + "image_path": "3e8b1ffcef67cfe66bf5c5c04d334ca80c64b63852d2acf792d78168f86a8358.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 222, + 301, + 267 + ], + "lines": [ + { + "bbox": [ + 55, + 222, + 301, + 267 + ], + "spans": [ + { + "bbox": [ + 55, + 222, + 301, + 267 + ], + "type": "text", + "content": "Figure 5. Effect of weighting functions in the loss on the FID score of ImageNet512 over the training iterations. The proposed sigmoid weighting with " + }, + { + "bbox": [ + 55, + 222, + 301, + 267 + ], + "type": "inline_equation", + "content": "b = -3" + }, + { + "bbox": [ + 55, + 222, + 301, + 267 + ], + "type": "text", + "content": " shows best performance consistent across all training iterations." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 277, + 297, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 277, + 297, + 312 + ], + "spans": [ + { + "bbox": [ + 55, + 277, + 297, + 312 + ], + "type": "text", + "content": "idea of U-Net is to combine a contracting path (which encodes the image semantics) with long skip-connections at the same resolution to enable precise localization." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 313, + 297, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 313, + 297, + 469 + ], + "spans": [ + { + "bbox": [ + 55, + 313, + 297, + 469 + ], + "type": "text", + "content": "With the introduction of ViTs [5] as an alternative for convolutional layers, the U-Net has been converted to U-ViTs using a similar network but replacing (some) ResNet blocks with ViTs, either with spatial down-sampling [14] or without [2]. Long-range skip-connections play an pivotal role in these networks, for example for fine-tuning controllable diffusion models [17] and for the stability and training speed [15, 30]. In these latter works, it is shown that multiplying the skip-connections with a (constant) scalar " + }, + { + "bbox": [ + 55, + 313, + 297, + 469 + ], + "type": "inline_equation", + "content": "(< 1)" + }, + { + "bbox": [ + 55, + 313, + 297, + 469 + ], + "type": "text", + "content": " improves training stability and speed, because of the propagation of noise throughout the network. In our work, we reduce the number of skip-connections significantly and initialise the network such that it resembles an identify function." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 484, + 296, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 484, + 296, + 568 + ], + "spans": [ + { + "bbox": [ + 55, + 484, + 296, + 568 + ], + "type": "text", + "content": "Flop heavy fine-tuning Fine-tuning on a higher resolution at large scale was effectively done in Stable Diffusion [32] to limit the training cost, which is a form of flop heavy scaling over different resolutions. In this paper we show that flop heavy scaling is also preferred from a regularization perspective, which can be done either by finetuning from a lower resolution, or by lowering the patching size." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 580, + 136, + 594 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 580, + 136, + 594 + ], + "spans": [ + { + "bbox": [ + 55, + 580, + 136, + 594 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 600, + 230, + 613 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 600, + 230, + 613 + ], + "spans": [ + { + "bbox": [ + 55, + 600, + 230, + 613 + ], + "type": "text", + "content": "4.1. Sigmoid weighting versus others" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 617, + 297, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 297, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 297, + 715 + ], + "type": "text", + "content": "Loss weighting In this experiment, we want to understand the effect of loss weighting on the final image quality of a model. For the noise schedule, we use an interpolated cosine schedule because it has relatively broad coverage over logsnr values. We compare the weightings corresponding to the original " + }, + { + "bbox": [ + 55, + 617, + 297, + 715 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 55, + 617, + 297, + 715 + ], + "type": "text", + "content": "-mse [11] and the multiscale " + }, + { + "bbox": [ + 55, + 617, + 297, + 715 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 55, + 617, + 297, + 715 + ], + "type": "text", + "content": "-mse [14], the EDM shifted weightings [19], including the original and monotonic versions [19, 26], and the proposed sigmoid weighting" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 306, + 68, + 552, + 216 + ], + "blocks": [ + { + "bbox": [ + 306, + 68, + 552, + 216 + ], + "lines": [ + { + "bbox": [ + 306, + 68, + 552, + 216 + ], + "spans": [ + { + "bbox": [ + 306, + 68, + 552, + 216 + ], + "type": "image", + "image_path": "0a20f14d695a3b2c1dc90f07e7e80a7af439384ddd39d007754bda786705f538.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 222, + 553, + 267 + ], + "lines": [ + { + "bbox": [ + 304, + 222, + 553, + 267 + ], + "spans": [ + { + "bbox": [ + 304, + 222, + 553, + 267 + ], + "type": "text", + "content": "Figure 6. Effect of training noise schedules on the FID score of ImageNet512 over the training iterations. Fortunately, with our proposed sigmoid weighting " + }, + { + "bbox": [ + 304, + 222, + 553, + 267 + ], + "type": "inline_equation", + "content": "(b = -3)" + }, + { + "bbox": [ + 304, + 222, + 553, + 267 + ], + "type": "text", + "content": ", the exact choice of the noise schedules has little effect on the training process." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 317, + 339, + 548, + 446 + ], + "blocks": [ + { + "bbox": [ + 313, + 275, + 555, + 331 + ], + "lines": [ + { + "bbox": [ + 313, + 275, + 555, + 331 + ], + "spans": [ + { + "bbox": [ + 313, + 275, + 555, + 331 + ], + "type": "text", + "content": "Table 1. Sigmoid weighting performance (left) with the flop heavy architecture. Larger resolutions require more negative bias. Power loss performance (right) on ImageNet512 with the small model. Increasing downsampling has a similar effect as decreasing bias, hence we prefer the more simple sigmoid weighting." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 339, + 548, + 446 + ], + "lines": [ + { + "bbox": [ + 317, + 339, + 548, + 446 + ], + "spans": [ + { + "bbox": [ + 317, + 339, + 548, + 446 + ], + "type": "table", + "html": "
(a) Sigmoid weighting(b) Power loss weighting
biasresolutionbiasdownsample levels
12825651210240123
-5---1.82-42.192.583.625.11
-4--1.631.75-32.192.322.854.02
-3-1.471.482.32-22.292.152.343.13
-21.491.421.552.17-12.642.322.252.54
-11.351.381.75-03.222.742.072.22
01.301.461.94-1-2.692.292.16
11.261.56--2-3.262.532.19
", + "image_path": "efb9313d8f94b8d94214c99f76f5cc0c88c6879119bcc7ddafe1adbca754cbf4.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 459, + 556, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 459, + 556, + 556 + ], + "spans": [ + { + "bbox": [ + 313, + 459, + 556, + 556 + ], + "type": "text", + "content": "with shift " + }, + { + "bbox": [ + 313, + 459, + 556, + 556 + ], + "type": "inline_equation", + "content": "b = -3" + }, + { + "bbox": [ + 313, + 459, + 556, + 556 + ], + "type": "text", + "content": ". As can be seen in Figure 5, the sigmoid weighting outperforms all other weightings, with EDM monotonic shifted being the closest alternative. Appendix C contains additional experiments, where we also evaluate the sigmoid weighting against the closest competitor without guidance. Here again sigmoid weighting is consistently better. We also evaluate a time-shifted variant, where the bias depends on the progress in training." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 570, + 557, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 557, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 557, + 714 + ], + "type": "text", + "content": "Training efficiency of noise schedules Since the training noise schedule can be decoupled from the loss weighting and it only affects the training efficiency but not the final performance, one can measure the convergence speed over the training process caused by leveraging particular noise schedules. In this experiment we test the cosine schedule [11], the shifted cosine and the interpolated cosine schedules [14] (Figure 6). For fair comparison, we fix the sampling noise schedule as the cosine interpolated schedule. Interestingly, when the sigmoid weighting is used, the choice of noise schedule does not affect the training efficiency significantly. Arguably, the shifted schedule is a bit more sample effi" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18066" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 73, + 69, + 539, + 222 + ], + "blocks": [ + { + "bbox": [ + 73, + 69, + 539, + 222 + ], + "lines": [ + { + "bbox": [ + 73, + 69, + 539, + 222 + ], + "spans": [ + { + "bbox": [ + 73, + 69, + 539, + 222 + ], + "type": "image", + "image_path": "cfba554f0499d140f421d739d32d5229f4c5d7630c45e389a8d625935e959e56.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 225, + 555, + 270 + ], + "lines": [ + { + "bbox": [ + 54, + 225, + 555, + 270 + ], + "spans": [ + { + "bbox": [ + 54, + 225, + 555, + 270 + ], + "type": "text", + "content": "Figure 7. FID versus training cost on ImageNet512. Note that the EDM-2 training curves are without guidance, but the best performance with guidance intervals is reported with a separate marker, the figure source adapted has been adapted from EDM2 [20]. Following their assumptions, one training iteration is three times as expensive as evaluating the model. Our Flop Heavy (finetune) is trained by first training a small model on ImageNet256 for 400k iterations and then finetuning the same model on ImageNet512." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 56, + 283, + 220, + 381 + ], + "blocks": [ + { + "bbox": [ + 56, + 283, + 220, + 381 + ], + "lines": [ + { + "bbox": [ + 56, + 283, + 220, + 381 + ], + "spans": [ + { + "bbox": [ + 56, + 283, + 220, + 381 + ], + "type": "image", + "image_path": "584f7bccdcdfe72473e5875f5ad05299e881bc000dab70b530254646605fc7cc.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 388, + 220, + 433 + ], + "lines": [ + { + "bbox": [ + 55, + 388, + 220, + 433 + ], + "spans": [ + { + "bbox": [ + 55, + 388, + 220, + 433 + ], + "type": "text", + "content": "Figure 8. The effect of scaling either channels or tokens, compared to the small model variant. FLOPS between model size " + }, + { + "bbox": [ + 55, + 388, + 220, + 433 + ], + "type": "inline_equation", + "content": "2\\mathrm{x}" + }, + { + "bbox": [ + 55, + 388, + 220, + 433 + ], + "type": "text", + "content": " and flop heavy are similar." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 223, + 283, + 387, + 381 + ], + "blocks": [ + { + "bbox": [ + 223, + 283, + 387, + 381 + ], + "lines": [ + { + "bbox": [ + 223, + 283, + 387, + 381 + ], + "spans": [ + { + "bbox": [ + 223, + 283, + 387, + 381 + ], + "type": "image", + "image_path": "c32f83e922fb55d924c8bf41c4c5064257dcb4a5573c063f4ece679f81f58f5a.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 222, + 388, + 387, + 433 + ], + "lines": [ + { + "bbox": [ + 222, + 388, + 387, + 433 + ], + "spans": [ + { + "bbox": [ + 222, + 388, + 387, + 433 + ], + "type": "text", + "content": "Figure 9. The effect of Residual-UViTs (removing block skip connections). Although smaller models can be effected, larger models are unaffected by the simplification." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 391, + 279, + 555, + 381 + ], + "blocks": [ + { + "bbox": [ + 391, + 279, + 555, + 381 + ], + "lines": [ + { + "bbox": [ + 391, + 279, + 555, + 381 + ], + "spans": [ + { + "bbox": [ + 391, + 279, + 555, + 381 + ], + "type": "image", + "image_path": "21d9702abf3eda288b9e9681b14556e5f7071b8f2e50a14b2017ca07968d5fd0.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 391, + 388, + 555, + 433 + ], + "lines": [ + { + "bbox": [ + 391, + 388, + 555, + 433 + ], + "spans": [ + { + "bbox": [ + 391, + 388, + 555, + 433 + ], + "type": "text", + "content": "Figure 10. Changing the encoder decoder architecture. Decoder blocks are more important than encoder blocks, and symmetric architectures have decent performance." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 454, + 295, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 454, + 295, + 561 + ], + "spans": [ + { + "bbox": [ + 54, + 454, + 295, + 561 + ], + "type": "text", + "content": "cient initially but all schedules perform similarly towards the end. This suggests that there is a lot of transfer across different logsnr values. We also tried the EDM training noise schedule with the sigma power sampling noise schedule [19] (not plotted), with both shifted or the original version, but its narrower logsnr sampling schedule led to performance degradation, achieving about 9 FID at best. This suggests that for high resolution, a wide logsnr range is crucial to cover the important noise levels." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 578, + 296, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 578, + 296, + 662 + ], + "spans": [ + { + "bbox": [ + 55, + 578, + 296, + 662 + ], + "type": "text", + "content": "Bias over resolutions In Table 1a, we show the effect of the bias term " + }, + { + "bbox": [ + 55, + 578, + 296, + 662 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 55, + 578, + 296, + 662 + ], + "type": "text", + "content": " on the final performance when trained using a flop heavy model. All models use the same guidance and similar guidance intervals, where the lower bound of the guidance interval has been shifted according to resolution (again approximately 1.5 per 2x resolution change), see Appendix B for more details." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "Sigmoid weighting versus the power loss In literature, high resolution generation can greatly benefit from having a multiscale loss [6, 7, 14]. To that end we create a multiscale" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 454, + 555, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 454, + 555, + 525 + ], + "spans": [ + { + "bbox": [ + 313, + 454, + 555, + 525 + ], + "type": "text", + "content": "variant that we name the power loss. In addition to the sigmoid weighting that effectively upweights low frequency components of the data and downweights high frequency components, the power loss aims to amplify the loss of low sub-band (smooth) signals in the spatial domain via Haar wavelets. For details on this loss see Appendix B.2." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 527, + 556, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 527, + 556, + 635 + ], + "spans": [ + { + "bbox": [ + 313, + 527, + 556, + 635 + ], + "type": "text", + "content": "In Table 1b we see that the bias should be adjusted based on the number of Haar transform downsampling levels. Despite being competitive, the performance of power loss is roughly on par with the one with sigmoid weighting but without multiscaling (i.e. downsample levels " + }, + { + "bbox": [ + 313, + 527, + 556, + 635 + ], + "type": "inline_equation", + "content": "= 0" + }, + { + "bbox": [ + 313, + 527, + 556, + 635 + ], + "type": "text", + "content": "), if we tune the bias of the sigmoid weighting more aggressively. Interestingly, lowering the bias has almost the same effect as increasing the scale. Therefore, building the multiscale variant is not necessary with the sigmoid weighting." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 647, + 524, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 647, + 524, + 659 + ], + "spans": [ + { + "bbox": [ + 313, + 647, + 524, + 659 + ], + "type": "text", + "content": "4.2. Scaling channels vs tokens on ImageNet" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "type": "text", + "content": "In this section we show the difference between scaling a model 4x FLOPS, either via channels or by creating larger resolution feature maps (or put differently, feature maps with a higher token count). Note that our small model on" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18067" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 180 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 180 + ], + "type": "text", + "content": "ImageNet512 is patched " + }, + { + "bbox": [ + 55, + 72, + 294, + 180 + ], + "type": "inline_equation", + "content": "4 \\times 4" + }, + { + "bbox": [ + 55, + 72, + 294, + 180 + ], + "type": "text", + "content": ". This model is scaled either by increasing channels by two (named model size " + }, + { + "bbox": [ + 55, + 72, + 294, + 180 + ], + "type": "inline_equation", + "content": "2x" + }, + { + "bbox": [ + 55, + 72, + 294, + 180 + ], + "type": "text", + "content": "), or by reducing patching to " + }, + { + "bbox": [ + 55, + 72, + 294, + 180 + ], + "type": "inline_equation", + "content": "2 \\times 2" + }, + { + "bbox": [ + 55, + 72, + 294, + 180 + ], + "type": "text", + "content": " (named flop heavy, because FLOPS over params is much higher). Indeed as can be seen in Figure 8, the flop heavy model performs much better than the model size " + }, + { + "bbox": [ + 55, + 72, + 294, + 180 + ], + "type": "inline_equation", + "content": "2x" + }, + { + "bbox": [ + 55, + 72, + 294, + 180 + ], + "type": "text", + "content": ", when trained in a standard way on ImageNet512. Note that unlike most experiments, in these experiments it was necessary to use very different optimal guidance settings, more details in Appendix B." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 182, + 295, + 254 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 182, + 295, + 254 + ], + "spans": [ + { + "bbox": [ + 55, + 182, + 295, + 254 + ], + "type": "text", + "content": "We confirm this is an issue related to overfitting, by training the model size 2x with distribution augmentation [18] the model size 2x variant performs much better, with an FID of 1.6 instead of 1.9 (latter in Figure 8). Still, without any augmentation beyond the standard horizontal flips, the flop heavy model achieves a score of 1.54." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 269, + 230, + 281 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 269, + 230, + 281 + ], + "spans": [ + { + "bbox": [ + 55, + 269, + 230, + 281 + ], + "type": "text", + "content": "4.3. Removing block skip connection" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 288, + 295, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 288, + 295, + 396 + ], + "spans": [ + { + "bbox": [ + 55, + 288, + 295, + 396 + ], + "type": "text", + "content": "In this section we study the difference between the residual skip-connections and the blockwise skip-connections. As hypothesized, the small model performance drops a bit because of this. For ImageNet512 a deficit of about 0.2 FID (in Figure 9) and for Imagenet256 a deficit of about 0.05 FID. That the task of fitting ImageNet256 is less affected by removing skip-connections can be explained by the task being easier: The task of fitting ImageNet256 is a subset of fitting ImageNet512." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 399, + 295, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 399, + 295, + 518 + ], + "spans": [ + { + "bbox": [ + 55, + 399, + 295, + 518 + ], + "type": "text", + "content": "For larger models, removing blockwise skip-connections in favor of a single residual connection does not degrade performance, and may even lead to slightly increased performance (see Figure 9). This is somewhat unexpected, and the phenomenon persists with different guidance settings as well. The FID curve reminds of the model size " + }, + { + "bbox": [ + 55, + 399, + 295, + 518 + ], + "type": "inline_equation", + "content": "2\\mathrm{x}" + }, + { + "bbox": [ + 55, + 399, + 295, + 518 + ], + "type": "text", + "content": " curve, which at first improves but flattens out over iterations. Therefore our best hypothesis is that the block skip (flop) baseline may have similar problems such as overfitting. Regardless, it shows that Residual U-ViTs scale well." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "text", + "content": "Asymmetric U-ViT Since our U-ViT models encompass only a single skip connection for each level, we can use a different number of ResNet or ViT blocks per level. This allows to move compute from the encoder stage (the down sampling half of the model) to the decoder stage (the upsampling half of the model). From Figure 10 we observe that (a) for a fixed compute budget, it is better to add more parameters to the decoder, the encoder heavy model (E5-D1) is significantly worse than the baseline (E3-D3) and the decoder heavy model (E1-D5); (b) there needs to be some balance between the encoder and the decoder, E1-D7 works less good than E3-D5. (c) by a slightly heavier model (E3-D7, only " + }, + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "text", + "content": " more parameters than our small (E3-D3) model, the performance significantly increases." + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 328, + 116, + 541, + 688 + ], + "blocks": [ + { + "bbox": [ + 313, + 71, + 553, + 114 + ], + "lines": [ + { + "bbox": [ + 313, + 71, + 553, + 114 + ], + "spans": [ + { + "bbox": [ + 313, + 71, + 553, + 114 + ], + "type": "text", + "content": "Table 2. Literature Comparison on ImageNet on FID. Note that even though certain NFEs are high (e.g. 4100 for CDM) typically these suffer minor performance degradation for lower NFEs and should not be taken too literally." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 328, + 116, + 541, + 688 + ], + "lines": [ + { + "bbox": [ + 328, + 116, + 541, + 688 + ], + "spans": [ + { + "bbox": [ + 328, + 116, + 541, + 688 + ], + "type": "table", + "html": "
MethodNFEFID
Imagenet 128 x 128
RIN [16]10002.75
SD, U-ViT-L [14]5121.94
StyleGAN-XL [38]11.81
VDM+++ [25]2501.75
DisCo-Diff [46]4141.73
PaGoDA [22]11.48
SiD2, Flop Heavy (ours)5121.26
Imagenet 256 x 256
ADM-G [4]2504.59
ADM-U [4]2503.94
CDM [12]41004.88
RIN [3, 16]10003.52
SD, U-ViT-L [14]5122.44
MDM [7]2503.51
StyleGAN-XL [38]12.30
DiT-XL/2 [31]2502.27
VDM++ [25]2502.12
RDM [42]3561.89
PaGoDA [22]11.56
SiD2, small (ours)5121.72
SiD2, Flop Heavy (ours)5121.38
Imagenet 512 x 512
ADM-G [4]2507.72
RIN [3, 16]10003.95
ADM-U [4]2503.85
LDM-4-G [32]2503.60
SD, U-ViT-L [14]5123.02
DiffiT [9]2502.67
StyleGAN-XL [38]12.40
DiT-XL/2 [31]2503.04
VDM++ [25]2502.65
MAGVIT-v2 [51]641.91
PaGoDA [22]11.80
EDM2-S (interval) [27]631.68
EDM2-XXL (interval) [27]631.40
SiD2, small (ours)5122.19
SiD2, Flop Heavy (ours)5121.48
SiD2, Flop Heavy distilled (ours)161.50
Imagenet 1024 x 1024
RIN [16]10008.72
SiD2, Flop Heavy (ours)5121.75
", + "image_path": "ef128d1d326e3d3a214cf0c61296f4451a59a5987594436ff51fc59603ea22ff.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18068" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 82, + 113, + 271, + 248 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 297, + 105 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 297, + 105 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 297, + 105 + ], + "type": "text", + "content": "Table 3. Zero-shot text-to-image performance evaluated on MSCOCO in " + }, + { + "bbox": [ + 55, + 70, + 297, + 105 + ], + "type": "inline_equation", + "content": "\\mathrm{FID}_{30\\mathrm{k}}" + }, + { + "bbox": [ + 55, + 70, + 297, + 105 + ], + "type": "text", + "content": ". Our distilled model, evaluated on the more difficult 512 setting, outperforming other works." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 82, + 113, + 271, + 248 + ], + "lines": [ + { + "bbox": [ + 82, + 113, + 271, + 248 + ], + "spans": [ + { + "bbox": [ + 82, + 113, + 271, + 248 + ], + "type": "table", + "html": "
MethodSizeNFEFID30k
DeepFloyd [40]2566.7
eDiff-I [1]2566.9
Parti [49]2567.2
Imagen [35]2567.3
simple diffusion [14]2568.3
simple diffusion [14]5122569.6
SDv1.5 (eval DMD2) [32]5122007.2
DMD2 [48]51248.3
SiD2 (ours, bias=-3)5122568.1
SiD2 16-step distilled (ours)512166.7
", + "image_path": "5d6987a48d055fb4e712dfadd36587f8b0d1bf41716e0491a2aae199d07621da.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 261, + 185, + 274 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 261, + 185, + 274 + ], + "spans": [ + { + "bbox": [ + 55, + 261, + 185, + 274 + ], + "type": "text", + "content": "4.4. Literature comparison" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 277, + 297, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 277, + 297, + 387 + ], + "spans": [ + { + "bbox": [ + 55, + 277, + 297, + 387 + ], + "type": "text", + "content": "Here we compare different models from the literature. In the first section end-to-end diffusion models are compared, and in the second section we compare between all types of generative models. In Table 2 it can be seen that our SiD2 is very competitive with existing generative models: On ImageNet128 and ImageNet256 it achieves state-of-the-art performance, on ImageNet512 it is a little worse but still competitive with the best latent diffusion model and outperforms all other models, pixel-based and latent." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 400, + 298, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 400, + 298, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 400, + 298, + 605 + ], + "type": "text", + "content": "Comparing diffusion models on compute The training cost SiD2 is another helpful way to see how the model scales, as depicted in Figure 7). Importantly, the performance of our pixel-based approach outperforms other significantly. Observe for example the performance of our small variant (using no guidance interval to keep this comparison fair), which roughly the same performance as the VDM++ while only using a third of the training cost, and a fourth of the FLOPS in a single forward pass (see Figure 16 in Appendix C). It is also competitive with the best currently known latent diffusion approach EDM2 [20], but in fairness EDM2 scales a little bit better. These models attain even better performance for a smaller training cost, and a similar story holds for forward pass complexity. End-to-end diffusion can work much better than previously demonstrated, but it is still somewhat outperformed by latent diffusion on quality versus compute comparisons for both sampling and training." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 618, + 297, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 297, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 297, + 715 + ], + "type": "text", + "content": "Text to image generation Finally, we show that our method can be used for end-to-end training of text-to-image models. Here, evaluation is more difficult as MSCOCO zero-shot performance depends a lot on the dataset used and is known to not necessarily correspond to actual perceived performance. Nevertheless, it is important that the model achieves performance in the right ballpark. Indeed, when we train SiD2 it performs comparable to other models in the" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 365, + 113, + 506, + 274 + ], + "blocks": [ + { + "bbox": [ + 313, + 70, + 556, + 105 + ], + "lines": [ + { + "bbox": [ + 313, + 70, + 556, + 105 + ], + "spans": [ + { + "bbox": [ + 313, + 70, + 556, + 105 + ], + "type": "text", + "content": "Table 4. Conditional Video generation performance on Kinetics600 in FVD (resized to 64). For video generation the loss specification is important to obtain good sample quality (see ablations)." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 365, + 113, + 506, + 274 + ], + "lines": [ + { + "bbox": [ + 365, + 113, + 506, + 274 + ], + "spans": [ + { + "bbox": [ + 365, + 113, + 506, + 274 + ], + "type": "table", + "html": "
MethodFVD50k
pixel diffusion
Video Diffusion Models [13]16.6
RIN [16]10.8
Rolling Diffusion [34]5.2
latent diffusion
MAGVIT [50]9.9
MAGVITv2 [51]4.3
W.A.L.T [8]3.3
SiD2 (ours, sigmoid bias = 0)2.3
ablations
eps-mse5.7
sigmoid bias = +23.2
sigmoid bias = -22.5
", + "image_path": "d5236b904016a7ed46178509b5d43b738cfd980c3ca046b1e34ffde478850fc2.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 286, + 556, + 358 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 286, + 556, + 358 + ], + "spans": [ + { + "bbox": [ + 313, + 286, + 556, + 358 + ], + "type": "text", + "content": "literature on MSCOCO " + }, + { + "bbox": [ + 313, + 286, + 556, + 358 + ], + "type": "inline_equation", + "content": "\\mathrm{FID}_{30\\mathrm{K}}" + }, + { + "bbox": [ + 313, + 286, + 556, + 358 + ], + "type": "text", + "content": ", and detailed samples have only a small number of artifacts (Figure 2). Furthermore, the moment matching distilled version using [37] outperforms the baseline model as previously observed. This model also outperforms all other zero-shot approaches at 512 resolution as shown in Table 3." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 372, + 556, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 372, + 556, + 529 + ], + "spans": [ + { + "bbox": [ + 313, + 372, + 556, + 529 + ], + "type": "text", + "content": "Video Generation To test whether our approach transfers to video, we test our model on Kinetics600 following [13], by conditioning on 5 frames and generating (sometimes referred to as predicting) the subsequent 11 frames. We then compute the FVD between the training set and the generated samples both on the subsequent 11 frames. Even though our method generates at " + }, + { + "bbox": [ + 313, + 372, + 556, + 529 + ], + "type": "inline_equation", + "content": "128 \\times 128" + }, + { + "bbox": [ + 313, + 372, + 556, + 529 + ], + "type": "text", + "content": ", we follow [8] and resize samples to have resolution " + }, + { + "bbox": [ + 313, + 372, + 556, + 529 + ], + "type": "inline_equation", + "content": "64 \\times 64" + }, + { + "bbox": [ + 313, + 372, + 556, + 529 + ], + "type": "text", + "content": ". Our model achieves SOTA performance and it turns out that tuning the loss is very important to achieve this as shown in Table 4. Note that the latent approaches ([8, 50, 51]) use much less training compute than our models in pixel-space, which we discuss in more detail in the Appendix B." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 538, + 384, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 538, + 384, + 550 + ], + "spans": [ + { + "bbox": [ + 313, + 538, + 384, + 550 + ], + "type": "text", + "content": "5. Discussion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 557, + 557, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 557, + 557, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 557, + 557, + 715 + ], + "type": "text", + "content": "In this paper we present Simpler Diffusion (SiD2) and show how one can train end-to-end diffusion and achieve competitive performance with latent models: (1) use the sigmoid loss, (2) with a simpler U-ViT architecture, and (3) favour a flop heavy model design. As a result, our method outperforms existing end-to-end models by a large margin. Nevertheless, at this moment latent diffusion methods still seem to have slightly better scaling properties and it remains to be seen whether this gap can be closed any further. Importantly, for many applications training a separate autoencoder may not be either practical or desired. Therefore, we believe this paper is an important contribution that allows end-to-end modelling directly in the pixel-space." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "18069" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 297, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 91, + 297, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 297, + 145 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 297, + 145 + ], + "type": "text", + "content": "[1] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Ji-aming Song, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, Bryan Catanzaro, Tero Karras, and Ming-Yu Liu. ediffi: Text-to-image diffusion models with an ensemble of expert denoisers. CoRR, abs/2211.01324, 2022. 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 148, + 296, + 179 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 148, + 296, + 179 + ], + "spans": [ + { + "bbox": [ + 61, + 148, + 296, + 179 + ], + "type": "text", + "content": "[2] Fan Bao, Shen Nie, Kaiwen Xue, Yue Cao, Chongxuan Li, Hang Su, and Jun Zhu. All are worth words: A vit backbone for diffusion models. In CVPR, 2023. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 182, + 296, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 182, + 296, + 203 + ], + "spans": [ + { + "bbox": [ + 62, + 182, + 296, + 203 + ], + "type": "text", + "content": "[3] Ting Chen. On the importance of noise scheduling for diffusion models. arxiv, 2023. 1, 4, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 205, + 296, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 205, + 296, + 227 + ], + "spans": [ + { + "bbox": [ + 62, + 205, + 296, + 227 + ], + "type": "text", + "content": "[4] Prafulla Dhariwal and Alex Nichol. Diffusion models beat gans on image synthesis. CoRR, abs/2105.05233, 2021. 4, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 228, + 296, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 228, + 296, + 316 + ], + "spans": [ + { + "bbox": [ + 62, + 228, + 296, + 316 + ], + "type": "text", + "content": "[5] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In 9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7, 2021. OpenReview.net, 2021. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 318, + 296, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 318, + 296, + 361 + ], + "spans": [ + { + "bbox": [ + 62, + 318, + 296, + 361 + ], + "type": "text", + "content": "[6] Jiatao Gu, Shuangfei Zhai, Yizhe Zhang, Miguel Angel Bautista, and Josh M. Susskind. f-dm: A multi-stage diffusion model via progressive signal transformation. CoRR, abs/2210.04955, 2022. 4, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 363, + 296, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 363, + 296, + 396 + ], + "spans": [ + { + "bbox": [ + 62, + 363, + 296, + 396 + ], + "type": "text", + "content": "[7] Jiatao Gu, Shuangfei Zhai, Yizhe Zhang, Josh M. Susskind, and Navdeep Jaitly. Matryoshka diffusion models. CoRR, abs/2310.15111, 2023. 4, 6, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 397, + 296, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 397, + 296, + 463 + ], + "spans": [ + { + "bbox": [ + 62, + 397, + 296, + 463 + ], + "type": "text", + "content": "[8] Agrim Gupta, Lijun Yu, Kihyuk Sohn, Xiuye Gu, Meera Hahn, Fei-Fei Li, Irfan Essa, Lu Jiang, and José Lezama. Photorealistic video generation with diffusion models. In Computer Vision - ECCV 2024 - 18th European Conference, Milan, Italy, September 29-October 4, 2024, Proceedings, Part LXXIX, pages 393-411. Springer, 2024. 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 465, + 296, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 465, + 296, + 497 + ], + "spans": [ + { + "bbox": [ + 62, + 465, + 296, + 497 + ], + "type": "text", + "content": "[9] Ali Hatamizadeh, Jiaming Song, Guilin Liu, Jan Kautz, and Arash Vahdat. Diffit: Diffusion vision transformers for image generation. CoRR, abs/2312.02139, 2023. 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 499, + 296, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 499, + 296, + 521 + ], + "spans": [ + { + "bbox": [ + 57, + 499, + 296, + 521 + ], + "type": "text", + "content": "[10] Jonathan Heek, Emiel Hoogeboom, and Tim Salimans. Multistep consistency models. CoRR, abs/2403.06807, 2024. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 523, + 296, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 523, + 296, + 566 + ], + "spans": [ + { + "bbox": [ + 57, + 523, + 296, + 566 + ], + "type": "text", + "content": "[11] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In Advances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, NeurIPS, 2020. 1, 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 568, + 296, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 568, + 296, + 610 + ], + "spans": [ + { + "bbox": [ + 57, + 568, + 296, + 610 + ], + "type": "text", + "content": "[12] Jonathan Ho, Chitwan Sahara, William Chan, David J. Fleet, Mohammad Norouzi, and Tim Salimans. Cascaded diffusion models for high fidelity image generation. J. Mach. Learn. Res., 23:47:1-47:33, 2022. 4, 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 613, + 296, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 613, + 296, + 667 + ], + "spans": [ + { + "bbox": [ + 57, + 613, + 296, + 667 + ], + "type": "text", + "content": "[13] Jonathan Ho, Tim Salimans, Alexey A. Gritsenko, William Chan, Mohammad Norouzi, and David J. Fleet. Video diffusion models. In Advances in Neural Information Processing Systems 35: Annual Conference on Neural Information Processing Systems 2022, NeurIPS, 2022. 8" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 670, + 296, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 296, + 712 + ], + "type": "text", + "content": "[14] Emiel Hoogeboom, Jonathan Heek, and Tim Salimans. simple diffusion: End-to-end diffusion for high resolution images. In International Conference on Machine Learning, ICML, pages 13213-13232. PMLR, 2023. 1, 3, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 316, + 73, + 555, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 555, + 115 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 555, + 115 + ], + "type": "text", + "content": "[15] Zhongzhan Huang, Zhou Pan, Shuicheng Yan, and Liang Lin. Scalelong: Towards more stable training of diffusion model via scaling network long skip connection. In NeurIPS, 2023. 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 118, + 555, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 118, + 555, + 148 + ], + "spans": [ + { + "bbox": [ + 317, + 118, + 555, + 148 + ], + "type": "text", + "content": "[16] Allan Jabri, David J. Fleet, and Ting Chen. Scalable adaptive computation for iterative generation. CoRR, abs/2212.11972, 2022. 7, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 151, + 555, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 151, + 555, + 194 + ], + "spans": [ + { + "bbox": [ + 317, + 151, + 555, + 194 + ], + "type": "text", + "content": "[17] Zeyinzi Jiang, Chaojie Mao, Yulin Pan, Zhen Han, and Jingfeng Zhang. Scedit: Efficient and controllable image diffusion generation via skip connection editing. Technical Report 2312.11392, arXiv, 2023. 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 195, + 555, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 195, + 555, + 248 + ], + "spans": [ + { + "bbox": [ + 317, + 195, + 555, + 248 + ], + "type": "text", + "content": "[18] Heewoo Jun, Rewon Child, Mark Chen, John Schulman, Aditya Ramesh, Alec Radford, and Ilya Sutskever. Distribution augmentation for generative modeling. In Proceedings of the 37th International Conference on Machine Learning, ICML, 2020. 3, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 251, + 555, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 251, + 555, + 304 + ], + "spans": [ + { + "bbox": [ + 316, + 251, + 555, + 304 + ], + "type": "text", + "content": "[19] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. In Advances in Neural Information Processing Systems 35: Annual Conference on Neural Information Processing Systems 2022, NeurIPS, 2022. 5, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 305, + 555, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 305, + 555, + 348 + ], + "spans": [ + { + "bbox": [ + 316, + 305, + 555, + 348 + ], + "type": "text", + "content": "[20] Tero Karras, Miika Aittala, Jaakko Lehtinen, Janne Hellsten, Timo Aila, and Samuli Laine. Analyzing and improving the training dynamics of diffusion models. CoRR, abs/2312.02696, 2023. 1, 4, 6, 8, 2, 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 350, + 555, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 350, + 555, + 392 + ], + "spans": [ + { + "bbox": [ + 316, + 350, + 555, + 392 + ], + "type": "text", + "content": "[21] Tero Karras, Miika Aittala, Tuomas Kynkänniemi, Jaakko Lehtinen, Timo Aila, and Samuli Laine. Guiding a diffusion model with a bad version of itself. CoRR, abs/2406.02507, 2024. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 394, + 555, + 447 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 394, + 555, + 447 + ], + "spans": [ + { + "bbox": [ + 316, + 394, + 555, + 447 + ], + "type": "text", + "content": "[22] Dongjun Kim, Chieh-Hsin Lai, Wei-Hsiang Liao, Yuhta Takida, Naoki Murata, Toshimitsu Uesaka, Yuki Mitsufuji, and Stefano Ermon. Pagoda: Progressive growing of a one-step generator from a low-resolution diffusion teacher. CoRR, abs/2405.14822, 2024. 7, 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 449, + 555, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 449, + 555, + 513 + ], + "spans": [ + { + "bbox": [ + 317, + 449, + 555, + 513 + ], + "type": "text", + "content": "[23] Dongjun Kim, Chieh-Hsin Lai, Wei-Hsiang Liao, Naoki Murata, Yuhta Takida, Toshimitsu Uesaka, Yutong He, Yuki Mitsufuji, and Stefano Ermon. Consistency trajectory models: Learning probability flow ODE trajectory of diffusion. In The Twelfth International Conference on Learning Representations, 2024. 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 515, + 555, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 515, + 555, + 548 + ], + "spans": [ + { + "bbox": [ + 316, + 515, + 555, + 548 + ], + "type": "text", + "content": "[24] Durk P Kingma and Prafulla Dhariwal. Glow: Generative flow with invertible 1x1 convolutions. Advances in neural information processing systems, 31, 2018. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 548, + 555, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 548, + 555, + 580 + ], + "spans": [ + { + "bbox": [ + 316, + 548, + 555, + 580 + ], + "type": "text", + "content": "[25] Diederik P. Kingma and Ruiqi Gao. Understanding the diffusion objective as a weighted integral of ellbods. CoRR, abs/2303.00848, 2023. 1, 2, 3, 4, 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 582, + 555, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 582, + 555, + 613 + ], + "spans": [ + { + "bbox": [ + 316, + 582, + 555, + 613 + ], + "type": "text", + "content": "[26] Diederik P. Kingma, Tim Salimans, Ben Poole, and Jonathan Ho. Variational diffusion models. CoRR, abs/2107.00630, 2021. 2, 5" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 615, + 555, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 615, + 555, + 657 + ], + "spans": [ + { + "bbox": [ + 317, + 615, + 555, + 657 + ], + "type": "text", + "content": "[27] Tuomas Kynkänniemi, Miika Aittala, Tero Karras, Samuli Laine, Timo Aila, and Jaakko Lehtinen. Applying guidance in a limited interval improves sample and distribution quality in diffusion models. CoRR, abs/2404.07724, 2024. 7, 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 317, + 658, + 555, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 658, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 658, + 555, + 713 + ], + "type": "text", + "content": "[28] Yaron Lipman, Ricky T. Q. Chen, Heli Ben-Hamu, Maximilian Nickel, and Matthew Le. Flow matching for generative modeling. In The Eleventh International Conference on Learning Representations, ICLR 2023, Kigali, Rwanda. OpenReview.net, 2023. 2" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18070" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 297, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 297, + 127 + ], + "type": "text", + "content": "[29] Weijian Luo, Tianyang Hu, Shifeng Zhang, Jiacheng Sun, Zhenguo Li, and Zhihua Zhang. Diff-instruct: A universal approach for transferring knowledge from pre-trained diffusion models. Advances in Neural Information Processing Systems, 36, 2024. 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 128, + 296, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 128, + 296, + 183 + ], + "spans": [ + { + "bbox": [ + 56, + 128, + 296, + 183 + ], + "type": "text", + "content": "[30] Jiajun Ma, Shuchen Xue, Tianyang Hu, Wenjia Wang, Zhaoqiang Liu, Zhenguo Li, Zhi-Ming Ma, and Kenji Kawaguchi. The surprising effectiveness of skip-tuning in diffusion sampling. In Proceedings of the 41st International Conference on Machine Learning, pages 34053-34074, 2024. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 184, + 295, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 184, + 295, + 205 + ], + "spans": [ + { + "bbox": [ + 56, + 184, + 295, + 205 + ], + "type": "text", + "content": "[31] William Peebles and Saining Xie. Scalable diffusion models with transformers. CoRR, abs/2212.09748, 2022. 4, 7, 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 206, + 296, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 206, + 296, + 270 + ], + "spans": [ + { + "bbox": [ + 56, + 206, + 296, + 270 + ], + "type": "text", + "content": "[32] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 10674-10685. IEEE, 2022. 1, 3, 4, 5, 7, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 272, + 296, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 272, + 296, + 304 + ], + "spans": [ + { + "bbox": [ + 56, + 272, + 296, + 304 + ], + "type": "text", + "content": "[33] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. Technical report, ArXiV, 2015. 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 305, + 296, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 305, + 296, + 348 + ], + "spans": [ + { + "bbox": [ + 56, + 305, + 296, + 348 + ], + "type": "text", + "content": "[34] David Ruhe, Jonathan Heek, Tim Salimans, and Emiel Hoogeboom. Rolling diffusion models. In *Forty-first International Conference on Machine Learning*, ICML 2024, Vienna, Austria, July 21-27, 2024. OpenReview.net, 2024. 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 350, + 297, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 350, + 297, + 424 + ], + "spans": [ + { + "bbox": [ + 56, + 350, + 297, + 424 + ], + "type": "text", + "content": "[35] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S. Sara Mahdavi, Rapha Gontijo Lopes, Tim Salimans, Jonathan Ho, David J. Fleet, and Mohammad Norouzi. Photorealistic text-to-image diffusion models with deep language understanding. CoRR, abs/2205.11487, 2022. 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 426, + 296, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 426, + 296, + 470 + ], + "spans": [ + { + "bbox": [ + 56, + 426, + 296, + 470 + ], + "type": "text", + "content": "[36] Tim Salimans and Jonathan Ho. Progressive distillation for fast sampling of diffusion models. In *The Tenth International Conference on Learning Representations*, ICLR. OpenReview.net, 2022. 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 471, + 296, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 471, + 296, + 514 + ], + "spans": [ + { + "bbox": [ + 56, + 471, + 296, + 514 + ], + "type": "text", + "content": "[37] Tim Salimans, Thomas Mensink, Jonathan Heek, and Emiel Hoogeboom. Multistep distillation of diffusion models via moment matching. arXiv preprint arXiv:2406.04103, 2024. 4, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 514, + 296, + 559 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 514, + 296, + 559 + ], + "spans": [ + { + "bbox": [ + 56, + 514, + 296, + 559 + ], + "type": "text", + "content": "[38] Axel Sauer, Katja Schwarz, and Andreas Geiger. Stylegan-xl: Scaling stylegan to large diverse datasets. In SIGGRAPH '22: Special Interest Group on Computer Graphics and Interactive Techniques Conference, pages 49:1-49:10. ACM, 2022. 7, 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 559, + 296, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 559, + 296, + 602 + ], + "spans": [ + { + "bbox": [ + 56, + 559, + 296, + 602 + ], + "type": "text", + "content": "[39] Evan Shelhamer, Jonathan Long, and Trevor Darrell. Fully convolutional networks for semantic segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 39(4):640-651, 2016. 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 603, + 296, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 603, + 296, + 647 + ], + "spans": [ + { + "bbox": [ + 56, + 603, + 296, + 647 + ], + "type": "text", + "content": "[40] Alex Shonenkov, Misha Konstantinov, Daria Bakshandaeva, Christoph Schuhmann, Ksenia Ivanova, and Nadija Klokova. Deepfloyd if: A text-to-image model with a high degree of photorealism and language understanding, 2023. 4, 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 647, + 297, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 297, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 297, + 712 + ], + "type": "text", + "content": "[41] Yang Song, Jascha Sohl-Dickstein, Diederik P. Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In 9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7, 2021. OpenReview.net, 2021. 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 555, + 654 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 316, + 72, + 555, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 555, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 555, + 128 + ], + "type": "text", + "content": "[42] Jiayan Teng, Wendi Zheng, Ming Ding, Wenyi Hong, Jianqiao Wangni, Zhuoyi Yang, and Jie Tang. Relay diffusion: Unifying diffusion process across resolutions for image synthesis. In The Twelfth International Conference on Learning Representations, ICLR. OpenReview.net, 2024. 4, 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 129, + 555, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 555, + 173 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 555, + 173 + ], + "type": "text", + "content": "[43] Haoning Wu, Shaocheng Shen, Qiang Hu, Xiaoyun Zhang, Ya Zhang, and Yanfeng Wang. Megafusion: Extend diffusion models towards higher-resolution image generation without further tuning. CoRR, abs/2408.11001, 2024. 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 174, + 555, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 174, + 555, + 217 + ], + "spans": [ + { + "bbox": [ + 316, + 174, + 555, + 217 + ], + "type": "text", + "content": "[44] Zhisheng Xiao, Karsten Kreis, and Arash Vahdat. Tackling the generative learning trilemma with denoising diffusion gans. In International Conference on Learning Representations, 2021. 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 219, + 555, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 219, + 555, + 252 + ], + "spans": [ + { + "bbox": [ + 316, + 219, + 555, + 252 + ], + "type": "text", + "content": "[45] Yanwu Xu, Yang Zhao, Zhisheng Xiao, and Tingbo Hou. Ufogen: You forward once large scale text-to-image generation via diffusion gans. arXiv preprint arXiv:2311.09257, 2023. 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 253, + 555, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 253, + 555, + 307 + ], + "spans": [ + { + "bbox": [ + 316, + 253, + 555, + 307 + ], + "type": "text", + "content": "[46] Yilun Xu, Gabriele Corso, Tommi S. Jaakkola, Arash Vahdat, and Karsten Kreis. Disco-diff: Enhancing continuous diffusion models with discrete latents. In *Forty-first International Conference on Machine Learning*, ICML 2024, Vienna, Austria, July 21-27, 2024. OpenReview.net, 2024. 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 308, + 555, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 308, + 555, + 352 + ], + "spans": [ + { + "bbox": [ + 316, + 308, + 555, + 352 + ], + "type": "text", + "content": "[47] Tianwei Yin, Michael Gharbi, Richard Zhang, Eli Shechtman, Fredo Durand, William T Freeman, and Taesung Park. One-step diffusion with distribution matching distillation. arXiv preprint arXiv:2311.18828, 2023. 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 353, + 555, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 353, + 555, + 397 + ], + "spans": [ + { + "bbox": [ + 316, + 353, + 555, + 397 + ], + "type": "text", + "content": "[48] Tianwei Yin, Michael Gharbi, Taesung Park, Richard Zhang, Eli Shechtman, Fredo Durand, and William T. Freeman. Improved distribution matching distillation for fast image synthesis. In NeurIPS, 2024. 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 398, + 555, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 398, + 555, + 464 + ], + "spans": [ + { + "bbox": [ + 316, + 398, + 555, + 464 + ], + "type": "text", + "content": "[49] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, Ben Hutchinson, Wei Han, Zarana Parekh, Xin Li, Han Zhang, Jason Baldridge, and Yonghui Wu. Scaling autoregressive models for content-rich text-to-image generation. CoRR, abs/2206.10789, 2022. 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 464, + 555, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 464, + 555, + 518 + ], + "spans": [ + { + "bbox": [ + 316, + 464, + 555, + 518 + ], + "type": "text", + "content": "[50] Lijun Yu, Yong Cheng, Kihyuk Sohn, José Lezama, Han Zhang, Huiwen Chang, Alexander G. Hauptmann, Ming-Hsuan Yang, Yuan Hao, Irfan Essa, and Lu Jiang. MAGVIT: masked generative video transformer. CoRR, abs/2212.05199, 2022. 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 520, + 555, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 520, + 555, + 597 + ], + "spans": [ + { + "bbox": [ + 316, + 520, + 555, + 597 + ], + "type": "text", + "content": "[51] Lijun Yu, José Lezama, Nitesh Bharadwaj Gundavarapu, Luca Versari, Kihyuk Sohn, David Minnen, Yong Cheng, Agrim Gupta, Xiuye Gu, Alexander G. Hauptmann, Boqing Gong, Ming-Hsuan Yang, Irfan Essa, David A. Ross, and Lu Jiang. Language model beats diffusion - tokenizer is key to visual generation. In The Twelfth International Conference on Learning Representations, ICLR. OpenReview.net, 2024. 7, 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 598, + 555, + 654 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 598, + 555, + 654 + ], + "spans": [ + { + "bbox": [ + 316, + 598, + 555, + 654 + ], + "type": "text", + "content": "[52] Wendi Zheng, Jiayan Teng, Zhuoyi Yang, Weihan Wang, Jidong Chen, Xiaotao Gu, Yuxiao Dong, Ming Ding, and Jie Tang. Cogview3: Finer and faster text-to-image generation via relay diffusion. In Computer Vision - ECCV Proceedings, Part LXXVII, pages 1-22. Springer, 2024. 4" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18071" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Simplification Is All You Need against Out-of-Distribution Overconfidence/c4dc8f46-aeb5-4477-a94f-497bd92530cf_content_list.json b/2025/Simplification Is All You Need against Out-of-Distribution Overconfidence/c4dc8f46-aeb5-4477-a94f-497bd92530cf_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..339414367d49796af8830d774f9a6d51e3a27eab --- /dev/null +++ b/2025/Simplification Is All You Need against Out-of-Distribution Overconfidence/c4dc8f46-aeb5-4477-a94f-497bd92530cf_content_list.json @@ -0,0 +1,1567 @@ +[ + { + "type": "text", + "text": "Simplification Is All You Need against Out-of-Distribution Overconfidence", + "text_level": 1, + "bbox": [ + 124, + 128, + 875, + 154 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keke Tang $^{1*}$ , Chao Hou $^{1*}$ , Weilong Peng $^{1\\dagger}$ , Xiang Fang $^{2}$ , Zhize Wu $^{3}$ , Yongwei Nie $^{4}$ , Wenping Wang $^{5}$ , Zhihong Tian $^{1}$", + "bbox": [ + 261, + 191, + 733, + 229 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Guangzhou University $^{2}$ Nanyang Technological University $^{3}$ Hefei University $^{4}$ South China University of Technology $^{5}$ Texas A&M University", + "bbox": [ + 156, + 239, + 841, + 277 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tangbohutbh@gmail.com, houchaohk@gmail.com, wlpeng@tju.edu.cn, xfang9508@gmail.com, wuzz@hfuu.edu.cn, nieyongwei@scut.edu.cn, wenping@amu.edu, tianzhihong@gzhu.edu.cn", + "bbox": [ + 129, + 279, + 867, + 311 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 354, + 328, + 372 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Deep neural networks (DNNs) often exhibit out-of-distribution (OOD) overconfidence, producing overly confident predictions on OOD samples. We attribute this issue to the inherent over-complexity of DNNs and investigate two key aspects: capacity and nonlinearity. First, we demonstrate that reducing model capacity through knowledge distillation can effectively mitigate OOD overconfidence. Second, we show that selectively reducing nonlinearity by removing ReLU operations further alleviates the issue. Building on these findings, we present a practical guide to model simplification, combining both strategies to significantly reduce OOD overconfidence. Extensive experiments validate the effectiveness of this approach in mitigating OOD overconfidence and demonstrate its superiority over state-of-the-art methods. Additionally, our simplification strategies can be combined with existing OOD detection techniques to further enhance OOD detection performance.", + "bbox": [ + 88, + 387, + 485, + 643 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 671, + 223, + 686 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Deep neural networks (DNNs) have revolutionized image classification tasks, consistently achieving state-of-the-art performance [8, 15, 25, 45, 46]. Despite this success, DNN classifiers are inherently designed for static, closed-world environments [2], limiting their ability to adapt to novel, unforeseen concepts encountered in real-world scenarios, known as out-of-distribution (OOD) samples. Indeed, many recent studies demonstrate that DNN classifiers suffer from OOD overconfidence, generating overly confident predictions on OOD samples [13, 34]. This issue significantly impacts the reliability and safety of DNNs in real-world applications, highlighting an urgent need for effective solutions.", + "bbox": [ + 89, + 696, + 483, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5e6a4a1f5e3b5d07fabd2929ba9082d24079e99165782dac9f651cebbe002a7e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 526, + 354, + 705, + 446 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/dce8fa53a43848f5d9336be0b0e3c2b2f9d31368c92f5d54078bbf5d6f2c43ca.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 707, + 354, + 888, + 446 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/916a7cfce59a808608cd0e609f3941a41b1aa0e793d6375a2c3ee3c3bae2dec2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 524, + 448, + 707, + 540 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5f9ef8a82b0e767c3e2a2dca39b494acf7a0e0ac697dcbde74f8bfb892ec4ef0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 709, + 448, + 888, + 556 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ff7334fac2b18610984438b898d30b9be4fc31f3283c45235f6b64b71e572110.jpg", + "image_caption": [ + "Figure 1. Histograms (logarithmic scale) of maximum confidence values for the teacher network (ResNet-50), the distilled student network (ResNet-18), and its variant with additional ReLU removal. All models are trained on CIFAR-100 and evaluated on OOD datasets: (a) SVHN, (b) LSUN-R, (c) iSUN, and (d) TinyImageNet-R. The $y$ -axis represents the frequency of samples, and the $x$ -axis represents confidence scores." + ], + "image_footnote": [], + "bbox": [ + 563, + 542, + 885, + 556 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Extensive research has been dedicated to addressing the OOD overconfidence issue. One common approach involves calibrating outputs for both in-distribution (ID) and OOD samples to improve discrimination [20], enabling OOD detectors to proactively reject OOD inputs. While these methods can reduce the impact of overconfidence, they do not fundamentally resolve the issue within DNN classifiers. Moreover, their effectiveness heavily depends on the classifiers' ability to adequately separate predictive distributions; they fail when the classifiers cannot achieve this separation [26]. As a more intrinsic solution, another approach trains DNN classifiers to suppress predictions on a subset of typical OOD samples, with the expectation that this behavior will generalize to other OOD data [18, 20].", + "bbox": [ + 511, + 689, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Joint first authors. †Corresponding author.", + "bbox": [ + 114, + 886, + 351, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "5030", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "However, given the vast diversity of potential OOD samples, the effectiveness of this method remains limited.", + "bbox": [ + 89, + 90, + 480, + 119 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "DNNs can indeed be regarded as complex systems, often characterized by over-parameterization, in which the number of parameters exceeds the amount of available training data [66]. This excessive complexity can cause DNNs to exhibit emergent behaviors, leading to unpredictable outcomes, as observed in other complex systems [31]. Consequently, when presented with certain OOD inputs, DNNs may unexpectedly but confidently misclassify them as ID classes. Therefore, investigating the relationship between model complexity and OOD overconfidence presents a promising path toward effectively addressing this issue.", + "bbox": [ + 88, + 121, + 480, + 286 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we explore two key aspects of model complexity: capacity and nonlinearity, and their roles in influencing OOD overconfidence. For capacity, we use knowledge distillation to examine how varying the student network's capacity affects OOD overconfidence, finding that an appropriately reduced capacity can significantly alleviate this issue (see Fig. 1). For nonlinearity, we investigate selective ReLU removal at different ratios and locations within the network, observing that controlled reductions in nonlinearity similarly mitigate OOD overconfidence (see Fig. 1). Based on these insights, we present a practical guide to model simplification that combines these two strategies to mitigate the OOD overconfidence issue in DNNs more effectively, achieving superior results compared to state-of-the-art methods. Furthermore, we demonstrate that our approach can be effectively combined with existing OOD detection techniques to further enhance performance.", + "bbox": [ + 91, + 287, + 483, + 542 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Overall, our contribution is summarized as follows:", + "bbox": [ + 109, + 544, + 450, + 558 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We are the first to attribute the OOD overconfidence issue in DNNs to their over-complexity, identifying model simplification as a crucial and effective solution.", + "- We systematically investigate two fundamental aspects of complexity: capacity and nonlinearity, analyzing their effects on OOD overconfidence through knowledge distillation and selective ReLU removal.", + "- We present a practical guide for model simplification and demonstrate through extensive experiments its effectiveness in significantly mitigating OOD overconfidence." + ], + "bbox": [ + 89, + 561, + 482, + 712 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 724, + 232, + 739 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Mitigating OOD Overconfidence. The issue of OOD overconfidence, where DNNs assign excessively high confidence scores to OOD samples, presents a substantial security risk for real-world applications [18, 42, 56]. Current approaches often incorporate auxiliary data to improve OOD sensitivity, using techniques such as adding random noise and image shuffling [26], leveraging large volumes of unlabeled data [51, 55, 60], relying on external OOD datasets [20], or generating synthetic OOD samples via models like Chamfer GAN [41]. Additionally, some", + "bbox": [ + 89, + 750, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "methods, including VOS [9] and NPOS [47], create virtual features in feature space rather than explicit OOD instances. However, these approaches typically require substantial volumes of OOD data.", + "bbox": [ + 511, + 90, + 903, + 150 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In contrast, our approach addresses OOD overconfidence by reducing model complexity, thus eliminating the need for auxiliary OOD data. Through extensive experiments, we systematically explore the effects of capacity reduction and selective activation removal, presenting a data-independent alternative for mitigating OOD overconfidence.", + "bbox": [ + 511, + 152, + 903, + 241 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "OD Detection. Another approach to managing OOD overconfidence is to develop scoring or normalization techniques that increase the distinction between ID and OOD samples [11, 17, 36, 43]. A foundational method, maximum softmax probability, introduced by Hendrycks and Gimpel [19], demonstrates that classifiers typically assign lower confidence to OOD samples. Follow-up methods, including ODIN [28] with temperature scaling and input perturbation, Energy Score [29], and density-based scores [27, 40], further enhance this ID-OOD separation. To improve robustness, several techniques refine feature representations: feature activation rectification [39, 62], feature selection [67], neuron pruning [3], and methods like CORES [44], which identify layer-specific relevant kernels for OOD detection. In contrast, our approach inherently addresses OOD overconfidence through model simplification. By tackling the root of the overconfidence issue, our method naturally enhances OOD detection performance as a result.", + "bbox": [ + 511, + 242, + 903, + 513 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Knowledge Distillation. Knowledge distillation transfers knowledge from a larger teacher model to a smaller student model, enabling the student to approximate the teacher's performance [22]. It typically utilizes response-based knowledge (soft logits) [1, 22, 32], feature-based knowledge (intermediate feature maps) [37, 50, 53], and relation-based knowledge (layer/sample relationships) [57, 58]. We use distillation to reduce model capacity and mitigate OOD overconfidence.", + "bbox": [ + 511, + 513, + 903, + 648 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ReLU Activation. ReLU $(f(x) = \\max(0, x))$ introduces nonlinearity while maintaining efficiency, making it widely used in deep learning [10, 12]. However, ReLU networks are prone to OOD overconfidence [18]. We address this by selectively removing ReLU activations to mitigate OOD overconfidence.", + "bbox": [ + 511, + 648, + 903, + 739 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. How Complexity Affects Overconfidence?", + "text_level": 1, + "bbox": [ + 511, + 753, + 883, + 771 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The complexity of a neural network is influenced by various factors. To investigate whether model complexity affects the degree of OOD overconfidence in DNNs, we focus on two key aspects: capacity and nonlinearity. In this section, we examine each of these factors individually to better understand their impact on OOD overconfidence.", + "bbox": [ + 511, + 779, + 903, + 869 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Evaluation Metrics. To assess the severity of OOD overconfidence across different DNNs, we use the mean max", + "bbox": [ + 511, + 869, + 903, + 898 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "5031", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 1 + }, + { + "type": "table", + "img_path": "images/75ec9c1ad56e1ef40ee44c7598cc1dadae5d9442f505313699ce8ac33cf9ef87.jpg", + "table_caption": [ + "Table 1. Comparison of MMC and OOD detection performance for student models (ResNet-18) distilled from a teacher model (ResNet-50) using different knowledge distillation methods on CIFAR datasets. Results are averaged across multiple OOD datasets." + ], + "table_footnote": [], + "table_body": "
IDCIFAR-10CIFAR-100
MethodMMC↓FPR95↓AUROC↑MMC↓FPR95↓AUROC↑
Teacher79.5854.2391.7872.1283.6774.73
KD [22]76.8149.8192.4262.6979.7776.31
FitNet [37]79.4650.5892.2660.6077.5377.78
AT [64]77.9048.0992.6964.1182.1475.51
SP [48]78.8150.5292.2258.7577.1280.01
", + "bbox": [ + 94, + 148, + 478, + 241 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "imum confidence (MMC) on OOD samples as our evaluation metric. Additionally, to analyze how OOD overconfidence impacts the model's ability for ID/OOD discrimination, we also report the area under the receiver operating characteristic curve (AUROC) [6], the false positive rate at a true positive rate of $95\\%$ (FPR95) [28], and the area under the precision-recall curve, with ID samples specified as positives (AUPRin) [38].", + "bbox": [ + 89, + 260, + 483, + 383 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Capacity", + "text_level": 1, + "bbox": [ + 89, + 392, + 197, + 409 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To investigate whether model capacity affects the severity of OOD overconfidence, we employ knowledge distillation to transfer knowledge from a large-capacity teacher network to a smaller-capacity student network. Specifically, we examine how different knowledge distillation strategies and varying student network capacities influence OOD overconfidence. Since DNN capacity is influenced by various factors, such as depth, width, and number of parameters, we ensure a fair comparison by testing networks with the same architecture but different depths.", + "bbox": [ + 89, + 415, + 483, + 566 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Effectiveness of Knowledge Distillation. To validate the effectiveness of knowledge distillation in mitigating OOD overconfidence, we select ResNet-50 as the teacher network and ResNet-18 as the student network, applying four distinct knowledge distillation methods: KD [22], FitNet [37], AT [64], and SP [48]. Experiments are conducted using CIFAR-10 and CIFAR-100 as ID datasets, with SVHN, LSUN-R, iSUN, Textures, and TinyImageNet-R as OOD datasets. The results, shown in Tab. 1, indicate that all four distillation methods lead to reductions in MMC and enhancements in AUROC and FPR95 metrics, demonstrating that knowledge distillation effectively mitigates the OOD overconfidence issue.", + "bbox": [ + 89, + 568, + 483, + 763 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Analysis of Student Networks with Varying Capacities. We conduct experiments on two datasets using KD [22] with student networks of varying capacities. For CIFAR-100, we set ResNet-101 as the teacher and evaluate student networks with capacities ranging from ResNet-50, ResNet-34, ResNet-18, to ResNet-10. For ImageNet-100, we use ResNet-152 as the teacher, with ResNet-101, ResNet-50, ResNet-34, and ResNet-18 as student models. As shown in Fig. 2, on CIFAR-100, reducing the student", + "bbox": [ + 89, + 763, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/4ffa6e7764a95fff13f73b17cd996d9e5cf871addff07cfe74f627858ec654f7.jpg", + "table_caption": [ + "Table 2. Comparison of MMC and OOD detection performance for different ReLU removal proportions using ResNet-101 as the classifier. Results are averaged across multiple OOD datasets." + ], + "table_footnote": [], + "table_body": "
IDCIFAR-10CIFAR-100
RatioMMC↓FPR95↓AUROC↑MMC↓FPR95↓AUROC↑
0%83.4651.9292.5373.8484.4173.03
1%84.4657.5890.4869.1482.4376.22
2%77.1843.0093.6972.2683.0775.25
3%77.9146.9393.0669.3778.7478.12
5%83.0756.5390.7168.7178.5779.43
10%79.6050.0692.2571.2480.9477.70
20%80.7850.3192.3073.5482.7374.81
50%79.2248.3492.7370.4082.2275.40
", + "bbox": [ + 522, + 135, + 895, + 252 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "network capacity to ResNet-18 improves OOD overconfidence and enhances ID/OOD discrimination. However, further reducing capacity to ResNet-10 results in degraded performance, likely due to insufficient capacity to effectively handle CIFAR-100 classification. A similar trend is observed on ImageNet-100, where reducing the student network capacity to ResNet-50 continues to yield improvements, but further reductions impair performance, likely due to the dataset's increased complexity. These findings suggest that reducing network capacity to a moderate level can effectively alleviate the OOD overconfidence issue.", + "bbox": [ + 511, + 268, + 906, + 436 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Nonlinearity", + "text_level": 1, + "bbox": [ + 513, + 445, + 648, + 460 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Nonlinearity in DNNs arises from activation functions, e.g., ReLU and Sigmoid. The more activation operations are applied, the greater the degree of nonlinearity. In this subsection, we investigate whether the degree of nonlinearity affects the severity of OOD overconfidence. Given that ReLU is one of the most widely used activation functions, we focus on examining the effect of removing ReLU operations.", + "bbox": [ + 511, + 468, + 905, + 574 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Specifically, we examine the impact of removing ReLU operations at different proportions and locations, similar to the approach in [4]. as we will detail in Sec. 4, to evaluate their effect on mitigating OOD overconfidence. Additionally, we assess the influence of initiating ReLU removal at various stages of training. For these experiments, CIFAR-10 and CIFAR-100 are used as the ID datasets, with SVHN, LSUN-R, iSUN, Textures, and TinyImageNet-R as the OOD datasets, using ResNet-18/101 as the classifier.", + "bbox": [ + 511, + 574, + 905, + 710 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Analysis of Different Removal Proportions. To assess the impact of varying removal proportions, we introduce the Negative Response Ratio (NRO), which measures the proportion of negative responses that are passed to ReLU operations:", + "bbox": [ + 511, + 710, + 903, + 786 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nN R O = \\frac {\\text {N u m b e r o f n e g a t i v e c o n v o l u t i o n a l r e s p o n s e s}}{\\text {T o t a l n u m b e r o f c o n v o l u t i o n a l r e s p o n s e s}}. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 527, + 795, + 903, + 840 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "A smaller NRO indicates fewer activations requiring truncation, suggesting that the corresponding ReLU layers contribute less to overall nonlinearity. Based on NRO values, we selectively remove ReLU operations with the low", + "bbox": [ + 511, + 840, + 906, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "5032", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/19135528d01fbdfee79fa48fbdcda4f0bbc6b063ba79534135e0e1344c342092.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 114, + 88, + 495, + 224 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/b8889e50737340affa915d527572fc1e1bb309e3feca8bf34d0bc30c4f98eaf6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 496, + 88, + 880, + 226 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/b7bbd30e25a50659fc0a95cd265e3c0a4fa6a07ed765e9a686a75683f3190403.jpg", + "image_caption": [ + "Figure 2. Comparison of MMC and OOD detection metrics, including FPR95, AUROC, and AUPRin, for student models with varying capacities (e.g., ResNet models) across datasets: (a) CIFAR-100 and (b) ImageNet-100." + ], + "image_footnote": [], + "bbox": [ + 112, + 220, + 305, + 354 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/0b695f5d3704bfe92f4f6a91eb4ec4f0903c73dcfe1cfa4ecbbaa9cafc7e32d6.jpg", + "image_caption": [], + "image_footnote": [ + "Teacher Student" + ], + "bbox": [ + 307, + 229, + 495, + 354 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/301ad9c7ffc67ad80272d3b6901be0ab800881fe239a99c2e8aa6d17f490666a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 496, + 229, + 687, + 354 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/521aa6f536e610a027968e764225eecf6777f02685bbb4831f6844b2bd32fd44.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 689, + 229, + 879, + 354 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/01f77a8c33e4dc7cae2ae8746efb3e495b5e1a150cbfd698a4aca7d73c6ef79a.jpg", + "image_caption": [ + "Figure 3. Comparison of MMC and FPR95 for ResNet-101 with ReLU removal applied at different network locations." + ], + "image_footnote": [], + "bbox": [ + 114, + 422, + 454, + 563 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "est NRO scores at proportions of $1\\%$ , $2\\%$ , $3\\%$ , $5\\%$ , $10\\%$ , $20\\%$ , and $50\\%$ , starting at the 40th epoch and continuing through the 100th epoch. As shown in Tab. 2, removing a certain proportion of ReLU operations reduces OOD overconfidence and enhances OOD detection. For example, removing $2\\%$ of ReLU operations on CIFAR-10 and $5\\%$ on CIFAR-100 yields the best results.", + "bbox": [ + 89, + 622, + 482, + 728 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Analysis of Different Removal Locations. To identify which ReLU operations are most suitable for removal, we compare the effects of removing ReLU operations from shallow, middle, and deep layers, as well as from layers with the lowest NRO values. Following previous findings, we remove $2\\%$ of ReLU operations for CIFAR-10 and $5\\%$ for CIFAR-100, initiating removal at the 40th epoch. As shown in Fig. 3, each removal strategy effectively reduces OOD overconfidence, with the best results achieved by removing ReLU operations based on NRO values.", + "bbox": [ + 89, + 731, + 482, + 882 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Impact of ReLU Removal Initiation Points. We also", + "bbox": [ + 89, + 885, + 482, + 900 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/09125d787ec7e270736996341582f23c8db509209b4a91dacf06d71b306f5a76.jpg", + "image_caption": [ + "Figure 4. Comparison of MMC and FPR95 for ResNet-18 with ReLU removal initiated at different training epochs." + ], + "image_footnote": [], + "bbox": [ + 542, + 428, + 879, + 526 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "analyze the effects of initiating ReLU removal at different stages during training. As illustrated in Fig. 4, the specific stage at which ReLU removal begins has only a moderate impact on overall performance. The best results are achieved when removal is initiated around the middle of the training process, such as at the 40th epoch.", + "bbox": [ + 511, + 587, + 906, + 679 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. A Practical Guide to Model Simplification", + "text_level": 1, + "bbox": [ + 511, + 693, + 888, + 710 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Based on our findings reported in Sec. 3, we present a practical guide to model simplification for addressing the OOD overconfidence issue in DNNs. For models that exhibit high OOD overconfidence, we suggest reducing complexity through a combination of knowledge distillation and selective ReLU removal. Please refer to Fig. 5 for demonstration. Knowledge Distillation with a Moderately-Sized Student. The first step is to choose a small student model with a capacity suited to the task requirements. This student model should be sufficiently compact to reduce complexity but still capable of performing well on the original task. We then apply knowledge distillation (e.g., KD) to transfer es", + "bbox": [ + 511, + 719, + 906, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "5033", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/bf8fc906eb6da835625877dfe77bf98c2395802d61b870b8320876d358dc070b.jpg", + "image_caption": [ + "Figure 5. Illustration of our model simplification pipeline. Given a complex teacher model with high capacity and nonlinearity, we use knowledge distillation to transfer information to a low-capacity student model. During distillation, ReLU operations are selectively replaced with Leaky ReLU, and the negative slope is iteratively adjusted from 0 to 1, ultimately transforming ReLU into an identity function." + ], + "image_footnote": [], + "bbox": [ + 163, + 87, + 836, + 314 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "sential knowledge from the larger teacher model to the student. This process enables the student to approximate the teacher's performance while operating with a reduced capacity, thereby alleviating the OOD overconfidence issue.", + "bbox": [ + 88, + 387, + 480, + 446 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Selective ReLU Removal During Distillation. To further streamline the model, we selectively remove ReLU operations. Using the NRO metric defined in Eq. (1), we identify ReLU operations that contribute minimally to the model's responses. Following the approach in [4], we replace these selected ReLU operations with Leaky ReLU, initially setting the negative slope to 0 and gradually increasing it to 1 midway through the distillation process. This controlled transition effectively transforms ReLU into an identity function, thereby reducing the model's nonlinearity.", + "bbox": [ + 88, + 448, + 482, + 598 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This practical guide, combining knowledge distillation for capacity reduction with selective ReLU removal for nonlinearity reduction, provides an accessible approach to mitigate OOD overconfidence issue in DNNs.", + "bbox": [ + 89, + 599, + 482, + 659 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Experimental Results", + "text_level": 1, + "bbox": [ + 89, + 671, + 295, + 688 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 89, + 696, + 282, + 713 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation Details. We employ KD [22] to transfer knowledge from teacher to student models, with a weight ratio of 1.5:1 between KD loss and cross-entropy (CE) loss. All student models are trained for 100 epochs, with selective ReLU removal starting at the 40th epoch and continuing through the end of training, similar to the approach in [4]. We remove $2\\%$ of ReLU operations for models trained on CIFAR-10, $5\\%$ for models trained on CIFAR-100, and $5\\%$ for models trained on ImageNet-1k [7]. All experiments are implemented in PyTorch [35] and conducted on a workstation equipped with dual 2.40 GHz CPUs, 128 GB RAM, and eight NVIDIA RTX 3090Ti GPUs.", + "bbox": [ + 89, + 719, + 482, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Setup. For (1) Mitigating OOD Overconfidence, we use CIFAR-10 and CIFAR-100 [24] as ID datasets, with SVHN [33], LSUN-R [59], iSUN [54], Textures [5], and TINR, a resized variant of TinyImageNet uniformly scaled to $32 \\times 32$ , as the OOD datasets. For (2) OOD Detection, we use the same CIFAR-based configuration as in (1). For large-scale evaluations, we use ImageNet-1k [7] as the ID dataset, with SUN [52], Places [65], Textures [5], and OpenImage-O [49] as the OOD datasets. We train ResNet-50 and ResNet-101 [16] on CIFAR-10 and CIFAR-100, and train DenseNet-201 [23] and WideResNet-101-2 (WRN-101) [63] on ImageNet-1k [7], following the training protocols specified in their respective original papers.", + "bbox": [ + 511, + 387, + 906, + 584 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Setting of Our Approach. We designate ResNet-18 and WideResNet-16-10 (WRN-16) as student networks for ResNet-50, and ResNet-34 and WideResNet-28-10 (WRN-28) as students for ResNet-101 on CIFAR datasets. For the ImageNet-1k dataset, DenseNet-161 serves as the student for DenseNet-201, and WideResNet-50-2 (WRN-50) for WRN-101. All student models are trained using the SGD optimizer with a momentum of 0.9 and an initial learning rate of 0.1. On the ImageNet-1k benchmark, the learning rate is reduced by a factor of 10 every 30 epochs, while on the CIFAR benchmarks, the learning rate is adjusted using the CosineAnnealingLR scheduler. For simplicity, we refer to our method as \"Ours (student)\" in subsequent sections to indicate the specific student network employed.", + "bbox": [ + 511, + 590, + 908, + 801 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Baselines. For (1) Mitigating OOD Overconfidence, we select three state-of-the-art baselines: ACET [18], CEDA [18], and CCUs [30]. For (2) OOD Detection, we benchmark our approach against five post-hoc methods: maximum softmax probability (MSP) [19], MaxLogit [21], Energy [29], ReAct [39], and FeatureNorm [61].", + "bbox": [ + 511, + 809, + 908, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5034", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/a91d1c4df56b66c6db7db4329e9c32041400b2e2012c7a1d597cb46e2c4d64fc.jpg", + "table_caption": [ + "Table 3. Comparison of MMC on OOD datasets for ResNet-50 and ResNet-101 models trained on CIFAR-10 and CIFAR-100, evaluated using various OOD overconfidence mitigation techniques. Underlined values denote the best performance for a single method, while bolded values indicate the best performance across all methods, including both individual and combined methods." + ], + "table_footnote": [], + "table_body": "
ModelMethodID: CIFAR-10AverageID: CIFAR-100Average
SVHNLSUN-RiSUNTexturesTINRSVHNLSUN-RiSUNTexturesTINR
ResNet-50—75.3479.6479.8681.5381.5379.5872.5072.0772.2072.7471.1172.12
CEDA74.7370.9172.0273.3778.6473.9355.6454.7756.2562.8555.4256.99
ACET74.2967.3567.7270.3769.8769.9263.6862.4962.1964.4461.1562.79
CCUs69.5663.5163.1068.6468.9266.7557.3352.3054.7161.2554.8956.09
Ours (ResNet-18)78.2473.2974.3779.5176.9876.4854.7543.6542.2761.4039.6948.35
Ours (WRN-16)63.6566.3268.4977.0970.1869.1543.0651.7651.9957.9652.3751.43
CEDA + Ours (WRN-16)61.3167.4069.6470.0973.5468.4047.8459.4060.6959.2460.0057.43
ACET + Ours (WRN-16)67.4461.2862.3670.0565.4865.3252.5459.4561.5960.8260.4858.98
CCUs + Ours (WRN-16)65.9758.8161.3264.8465.9663.3850.8849.0451.8558.4051.8252.40
ResNet-101—83.4482.1383.0284.5684.1783.4665.2577.2178.0373.3875.3273.84
CEDA70.0768.2668.4274.4274.0271.0458.8258.1859.0066.2358.4760.14
ACET75.8253.6955.0064.5059.9861.8063.6259.5758.4460.3458.0160.00
CCUs73.5665.7264.9866.9469.8368.2158.4554.5356.5259.8056.3257.12
Ours (ResNet-34)75.2274.8175.9482.0878.8677.3863.6258.4260.1366.6959.8961.75
Ours (WRN-28)71.2174.7376.0381.8077.8976.3355.6947.8451.2063.6954.0954.50
CEDA + Ours (WRN-28)67.1568.2170.5976.7773.8371.3156.5165.1266.5962.5166.7763.50
ACET + Ours (WRN-28)59.3352.7053.0766.9558.0558.0256.1062.4863.9458.3463.8660.95
CCUs + Ours (WRN-28)67.7961.7464.2065.9568.8065.7053.7150.9653.5857.8654.5254.13
", + "bbox": [ + 158, + 133, + 841, + 417 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/2a70a23a637a5e8fe3f704e446e6090892e643e0f8fa1c4cf8befca0eb7c2718.jpg", + "image_caption": [ + "(a)", + "(b)" + ], + "image_footnote": [], + "bbox": [ + 125, + 441, + 272, + 512 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5450420c4e84847d0faa06662c292c27f4cd299cf0f3c2763cd43382beb50d1f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 125, + 518, + 272, + 593 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/154d7c306de00cb139e4e86ecf243644f310051d2d3e7abe9102816e300c8e5b.jpg", + "image_caption": [ + "ACET" + ], + "image_footnote": [], + "bbox": [ + 274, + 440, + 419, + 513 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d683d1fef2149b66067bae02fe792dedf427a55d07dac875bffc6d84f9c63b0a.jpg", + "image_caption": [ + "CEDA" + ], + "image_footnote": [], + "bbox": [ + 421, + 440, + 568, + 513 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c58ce1c74130b039c4d9056eca9d0802ac214005ae37fab257fa844c8d247a26.jpg", + "image_caption": [ + "CCUs" + ], + "image_footnote": [], + "bbox": [ + 570, + 440, + 715, + 513 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/195006fd2e1c97a2d69f08dae8f6f38e157e8e06aeab95d519a2cb87a577b88b.jpg", + "image_caption": [ + "Ours (ResNet-18)" + ], + "image_footnote": [], + "bbox": [ + 718, + 440, + 864, + 513 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e81ef09903ff89a34b2893ed9cc7fe4e1630ab89cf858bd580e0130fe071d9ec.jpg", + "image_caption": [ + "(c)", + "(d)" + ], + "image_footnote": [], + "bbox": [ + 125, + 609, + 272, + 684 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c0844b9b5ba1f2864d2361bc5eb81364dc23e04710ccf26e08a3dae0c41f836d.jpg", + "image_caption": [ + "Figure 6. Histograms (logarithmic scale) of MMC values for (a, b) ResNet-50 and (c, d) ResNet-101, w/ and w/o applying mitigation techniques: ACET, CEDA, CCUs, and Ours. All models are trained on CIFAR-100 and evaluated on OOD datasets: (a, c) LSUN-R and (d) TinyImageNet-R. In our approach, ResNet-18 and WRN-28 serve as students for ResNet-50 and ResNet-101, respectively." + ], + "image_footnote": [], + "bbox": [ + 125, + 688, + 272, + 762 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5f3f12b6ef1143b59a059d7d12e983c1526e867a2d63a304238575ebc2ac6b72.jpg", + "image_caption": [ + "ACET" + ], + "image_footnote": [], + "bbox": [ + 274, + 609, + 419, + 684 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ff89139269c01983fdda6fe307d4f08114af96d0a9f4a0bcb690b1f1e5430d4c.jpg", + "image_caption": [ + "CEDA" + ], + "image_footnote": [], + "bbox": [ + 421, + 609, + 568, + 683 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/011606af33bf67277ae955b252a1ece844b99209b093b7d7ddd7cd2b207b2f9c.jpg", + "image_caption": [ + "CCUs" + ], + "image_footnote": [], + "bbox": [ + 570, + 609, + 715, + 683 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/196b6591331d6b3a402901ff99e39d9e570fed47804d5134bef54eed362cd2e0.jpg", + "image_caption": [ + "Ours (WRN-28)" + ], + "image_footnote": [], + "bbox": [ + 718, + 611, + 864, + 685 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Results of Mitigating OOD Overconfidence", + "text_level": 1, + "bbox": [ + 125, + 821, + 455, + 838 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Comparison with State-of-the-art Methods. To assess the effectiveness and superiority of our approach, we conduct comparative experiments evaluating OOD overconfi", + "bbox": [ + 89, + 854, + 480, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "dence for ResNet-50 and ResNet-101 trained on CIFAR-10 and CIFAR-100. The results in Tab. 3 demonstrate that our approach consistently reduces the mean maximum confidence (MMC) across different student networks in all cases, validating its effectiveness. In comparison, methods such", + "bbox": [ + 511, + 823, + 906, + 898 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "5035", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/5fe2cce378dda2cce1d573c8a6201f09a59af1e5463fa9acc29c6edd94c89e71.jpg", + "table_caption": [ + "Table 4. Comparison of OOD detection performance for various methods on the CIFAR-10 and CIFAR-100 datasets. Underlined values indicate results enhanced by applying our approach to the baseline, while bolded values denote the best performance." + ], + "table_footnote": [], + "table_body": "
IDModelMethodOODAverage
SVHNLSUN-RiSUNTexturesTINR
FPR95↓AUROC↑FPR95↓AUROC↑FPR95↓AUROC↑FPR95↓AUROC↑FPR95↓AUROC↑FPR95↓AUROC↑
CIFAR-10ResNet-50MSP43.7594.1954.3291.9655.4691.6659.0190.3158.6190.8054.2391.78
MSP + Ours (ResNet-18)51.8293.2741.9393.8944.8693.2755.6490.5750.1891.9748.8992.59
MSP + Ours (WRN-16)28.1896.1235.7495.0139.7894.3457.0089.7942.3593.7540.6193.80
MaxLogit22.0796.3132.0394.9534.9594.5346.3391.3944.9692.5736.0793.95
MaxLogit + Ours (ResNet-18)30.2594.9425.0695.3427.9194.7344.5091.7136.7592.7332.8993.89
MaxLogit + Ours (WRN-16)17.6196.9618.5496.9221.6996.3947.7089.5827.8195.5526.6795.08
Energy20.3596.4029.7995.1132.8094.6745.4191.4443.7692.6434.4294.05
Energy + Ours (ResNet-18)28.5395.0723.9995.4326.5594.8344.1389.7035.9292.7931.8293.56
Energy + Ours (WRN-16)18.8096.8818.0797.0220.9196.5047.8989.5427.6395.6126.6695.11
ReAct37.6691.1016.2496.6419.6595.8446.7286.2930.7993.4630.2192.67
ReAct + Ours (ResNet-18)31.5994.6223.8795.5126.7894.8643.6090.1935.8792.9232.3493.62
ReAct + Ours (WRN-16)19.0796.8417.9097.0320.6696.5147.4589.8427.2495.6826.4695.18
FeatureNorm9.2597.9777.3482.0770.2983.9648.9784.1667.7184.4954.7186.53
FeatureNorm + Ours (ResNet-18)2.7399.3524.3495.7521.0496.2822.8294.8929.5294.3620.0996.13
FeatureNorm + Ours (WRN-16)3.3799.2622.0396.4321.5496.4753.6778.8529.3594.5025.9993.10
CIFAR-100ResNet-101MSP76.3081.1486.6769.7688.1169.0685.6973.7385.2771.4784.4173.03
MSP + Ours (ResNet-34)77.4878.9272.7483.0974.8382.1482.5077.2674.5282.2076.4180.72
MSP + Ours (WRN-28)74.8781.3962.5785.7166.9483.9383.2876.3369.4481.7071.4281.81
MaxLogit68.0586.7782.8474.3484.9473.4087.2574.8181.9275.0581.0076.87
MaxLogit + Ours (ResNet-34)71.6983.4763.7487.2665.6986.5481.7478.4666.8886.4469.9584.43
MaxLogit + Ours (WRN-28)73.8482.6154.3089.1160.2987.2683.6277.0664.1284.9867.2384.20
Energy68.3886.9482.3874.6484.2173.6788.0374.6281.4075.2580.8877.02
Energy + Ours (ResNet-34)69.2084.1458.8388.0561.1087.3781.6878.5162.8987.1866.7485.05
Energy + Ours (WRN-28)74.1782.6148.6090.0755.0588.1284.1077.0160.4185.7164.4784.70
ReAct65.6888.1486.4469.1886.9769.0983.4480.1383.8071.5981.2775.63
ReAct + Ours (ResNet-34)60.2688.3057.9486.4958.1186.5263.0785.7860.0385.6959.8886.56
ReAct + Ours (WRN-28)71.8986.7347.3489.2053.6687.4680.5781.4858.9784.6662.4985.91
FeatureNorm31.8894.5999.5434.1199.1039.0253.4282.3298.2242.3276.4358.47
FeatureNorm + Ours (ResNet-34)41.4790.4792.4671.4991.4172.8965.1880.3389.9969.0076.1076.84
FeatureNorm + Ours (WRN-28)54.1789.0880.8679.5280.6979.9171.4974.3778.0080.7473.0480.72
", + "bbox": [ + 96, + 122, + 906, + 556 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "as CEDA, ACET, and CCUs, which employ specific techniques to suppress confidence scores during training, also significantly mitigate OOD overconfidence in DNNs. However, our approach, which does not incorporate specific confidence suppression measures, still outperforms these state-of-the-art methods, particularly in the more challenging CIFAR-100 scenario. This further highlights the robustness and effectiveness of our strategy.", + "bbox": [ + 89, + 574, + 482, + 696 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "On CIFAR-10, while our base method may not show a pronounced advantage over others, combining our approach with confidence suppression techniques provides additional improvements, surpassing the performance of these state-of-the-art methods. These results reinforce the effectiveness of our approach in addressing OOD overconfidence.", + "bbox": [ + 89, + 699, + 482, + 791 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Visualization. We visualize the maximum confidence scores predicted by various DNN classifiers as histograms (logarithmic scale) in Fig. 6. As shown, applying OOD overconfidence mitigation solutions shifts the confidence distributions of the Baseline to the left, substantially reducing the number of samples with high confidence scores. Notably, our approach achieves the most pronounced shift, un", + "bbox": [ + 89, + 794, + 482, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "derscoring its effectiveness and validating its superiority in mitigating OOD overconfidence.", + "bbox": [ + 511, + 574, + 906, + 606 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Results of Enhancing OOD Detection", + "text_level": 1, + "bbox": [ + 511, + 617, + 834, + 633 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "CIFAR Experiments. We evaluate the effectiveness of our approach in enhancing OOD detection. The results in Tab. 4 demonstrate that applying our method to existing OOD detection approaches leads to significant performance improvements. Even when directly applied to the classifier without any advanced post-hoc processing, i.e., MSP, the performance enhancement is still noticeable. These improvements are consistent across different student model configurations, datasets, and classifiers, validating the effectiveness of our approach in boosting OOD detection.", + "bbox": [ + 511, + 642, + 906, + 792 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ImageNet Experiments. We further validate the effectiveness of our approach in enhancing OOD detection on the ImageNet-1k dataset. The results in Tab. 5 indicate that, similar to the CIFAR experiments, applying our approach significantly improves OOD detection performance across methods, demonstrating the robustness and scalability of our approach in large-scale settings.", + "bbox": [ + 511, + 795, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "5036", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/6c83316610385ef32e2e9572156dfe65487b7f6951758bec3c7e77f4fa41d3d3.jpg", + "table_caption": [ + "Table 5. Comparison of OOD detection performance for various methods on the ImageNet-1k dataset." + ], + "table_footnote": [], + "table_body": "
ModelMethodOODAverage
SUNPlacesTexturesOpenImage-O
FPR95↓AUROC↑FPR95↓AUROC↑FPR95↓AUROC↑FPR95↓AUROC↑FPR95↓AUROC↑
DenseNet-201MSP65.7682.2469.5381.1467.8279.3664.9284.4767.0181.80
MSP + Ours (DenseNet-161)66.3181.7268.6181.3266.9779.5562.1285.1066.0081.92
MaxLogit55.4486.0661.8983.9958.4883.4559.6486.9458.8685.11
MaxLogit + Ours (DenseNet-161)55.3387.6759.9586.2655.8085.8460.0388.6057.7887.09
Energy52.8786.0460.4683.7156.1383.3559.8786.6557.3384.94
Energy + Ours (DenseNet-161)50.8888.2556.9086.6052.2586.4558.7688.6054.7087.48
ReAct46.0090.8157.4686.1155.1484.1969.3274.4356.9883.89
ReAct + Ours (DenseNet-161)49.2188.5555.0286.9450.9986.7557.6188.6453.2187.72
FeatureNorm43.0690.0555.7185.4250.4679.9079.9767.2257.3080.65
FeatureNorm + Ours (DenseNet-161)36.5291.7949.2487.5056.7378.2982.2065.4956.1780.77
WRN-101MSP66.9382.4470.4881.3565.1480.9159.8786.1365.6182.71
MSP + Ours (WRN-50)67.7582.8069.0781.8166.3880.9660.9186.7866.0383.09
MaxLogit63.2584.7568.4082.6957.3984.3855.8488.4361.2285.06
MaxLogit + Ours (WRN-50)61.9286.1864.3484.4056.7485.6055.6989.2559.6786.36
Energy65.0684.4670.3282.2757.1184.3758.0488.0262.6384.78
Energy + Ours (WRN-50)63.1885.9666.0684.0055.9085.6857.8788.7760.7586.10
ReAct52.2288.1158.7685.4757.0984.9949.6387.8254.4386.60
ReAct + Ours (WRN-50)37.0292.2448.5488.7050.9087.3654.4985.9247.7488.56
FeatureNorm70.0778.8978.6173.2320.9894.7975.0974.8161.1980.43
FeatureNorm + Ours (WRN-50)60.2283.6774.0175.9617.7196.0969.9478.9055.4783.66
", + "bbox": [ + 99, + 108, + 897, + 382 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/fde2ce04aaca16c76111726c9f84627990c92ae2e4070b9d447cfb428d1867da.jpg", + "table_caption": [ + "Table 6. Comparison of MMC and OOD detection performance for ResNet-50 enhanced by Ours (ResNet-18) and its variants w/o knowledge distillation (KD) and ReLU removal (RR), using CIFAR-10 as the ID dataset, and SVHN, LSUN-R, iSUN, Textures, and TINR as the OOD datasets. Results are averaged." + ], + "table_footnote": [], + "table_body": "
KDRRMMC↓FPR95↓AUROC↑
79.5854.2391.78
✓76.8149.8192.42
✓78.4650.5091.76
✓✓76.4848.8992.59
", + "bbox": [ + 169, + 476, + 406, + 547 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.4. Ablation Studies and More Analysis", + "text_level": 1, + "bbox": [ + 89, + 573, + 403, + 589 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Importance of Knowledge Distillation and ReLU Removal. We conduct comprehensive ablation studies to evaluate the impact of knowledge distillation and ReLU removal, the two essential components of our approach. As shown in Tab. 6, excluding either component notably diminishes the effectiveness of our method in mitigating OOD overconfidence and complicates the accuracy of OOD detection. These findings validate the critical importance of both knowledge distillation and selective ReLU removal in achieving optimal performance with our approach.", + "bbox": [ + 89, + 598, + 482, + 748 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Knowledge Distillation vs. Network Pruning for Capacity Reduction. We further investigate an alternative capacity reduction method for our approach by applying network pruning. To facilitate a fair comparison with knowledge distillation, we evaluate ResNet-50 and ResNet-101 models trained on CIFAR-10, applying two strategies: (1) using KD [22] to reduce ResNet-50 to ResNet-18 and ResNet-101 to WRN-28, and (2) applying Prune [14] to ResNet-50 and ResNet-101, reducing them to parameter counts comparable to ResNet-18 and WRN-28, respectively. The re", + "bbox": [ + 89, + 750, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/60035a89324ec8924f96fb8ab61cb275242cae4d322d35e32b3d47d66f281b7f.jpg", + "table_caption": [ + "Table 7. Comparison of MMC and OOD detection performance for ResNet-50 and ResNet-101 reduced to comparable parameter counts via KD [22] and Prune [14]. CIFAR-10 serves as the ID dataset, with SVHN, LSUN-R, iSUN, Textures, and TINR as OOD datasets. Results are averaged across all five OOD datasets" + ], + "table_footnote": [], + "table_body": "
ModelCapacity ReductionMMC↓FPR95↓AUROC↑
ResNet-50—79.5854.2391.78
Prune77.3452.5592.27
KD (ResNet-18)76.8149.8192.42
ResNet-101—83.4651.9292.53
Prune78.2050.9291.61
KD (WRN-28)77.0141.0992.56
", + "bbox": [ + 529, + 476, + 890, + 561 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "sults in Tab. 7 show that both methods effectively reduce OOD overconfidence and improve OOD detection; however, knowledge distillation consistently delivers a more substantial improvement than network pruning. Additionally, knowledge distillation offers greater stability and flexibility during training, making it a more suitable choice for capacity reduction in our approach.", + "bbox": [ + 511, + 588, + 905, + 694 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion and Future Work", + "text_level": 1, + "bbox": [ + 511, + 709, + 782, + 724 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper has presented a comprehensive investigation into mitigating OOD overconfidence in DNNs through model simplification. By reducing model capacity via knowledge distillation and decreasing nonlinearity through selective ReLU removal, we demonstrate that OOD overconfidence in DNN classifiers can be significantly mitigated. Furthermore, these strategies can be effectively combined with OOD detection methods to enhance overall performance. In future work, we plan to study the effects of model simplification on OOD overconfidence in greater depth and from a theoretical perspective.", + "bbox": [ + 511, + 734, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "5037", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 91, + 90, + 259, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We sincerely thank the anonymous reviewers for their valuable comments. This work was supported in part by the National Natural Science Foundation of China (62472117, U2436208, 62406095), the Guangdong Basic and Applied Basic Research Foundation (2025A1515010157, 2024A1515012064), the Science and Technology Projects in Guangzhou (2025A03J0137, SL2022A04J01112), and the Academician Bixing Fang's Specialized Class.", + "bbox": [ + 89, + 114, + 483, + 236 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 247, + 187, + 263 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Lei Jimmy Ba and Rich Caruana. Do deep nets really need to be deep? In NeurIPS, pages 2654-2662, 2014. 2", + "[2] Abhijit Bendale and Terrance Boult. Towards open world recognition. In CVPR, pages 1893-1902, 2015. 1", + "[3] Chao Chen, Zhihang Fu, Kai Liu, Ze Chen, Mingyuan Tao, and Jieping Ye. Optimal parameter and neuron pruning for out-of-distribution detection. In NeurIPS, 2023. 2", + "[4] Hanting Chen, Yunhe Wang, Jianyuan Guo, and Dacheng Tao. Vanillanet: the power of minimalism in deep learning. In NeurIPS, 2024. 3, 5", + "[5] Mircea Cimpoi, Subhransu Maji, Iasonas Kokkinos, Sammy Mohamed, and Andrea Vedaldi. Describing textures in the wild. In CVPR, pages 3606-3613, 2014. 5", + "[6] Jesse Davis and Mark Goadrich. The relationship between precision-recall and roc curves. In ICML, pages 233-240, 2006. 3", + "[7] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, pages 248-255, 2009. 5", + "[8] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 1", + "[9] Xuefeng Du, Zhaoning Wang, Mu Cai, and Yixuan Li. Vos: Learning what you don't know by virtual outlier synthesis. In ICLR, 2021. 2", + "[10] Shiv Ram Dubey, Satish Kumar Singh, and Bidyut Baran Chaudhuri. Activation functions in deep learning: A comprehensive survey and benchmark. Neurocomputing, 503: 92-108, 2022. 2", + "[11] Xiang Fang, Arvind Easwaran, Blaise Genest, and Ponnuthurai Nagaratnam Suganthan. Your data is not perfect: Towards cross-domain out-of-distribution detection in class-imbalanced data. ESWA, 2024. 2", + "[12] Xavier Glorot, Antoine Bordes, and Yoshua Bengio. Deep sparse rectifier neural networks. In AISTATS, pages 315-323, 2011. 2", + "[13] Ian J. Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. In ICLR, 2015. 1", + "[14] Song Han, Jeff Pool, John Tran, and William J Dally. Learning both weights and connections for efficient neural networks. In NeurIPS, pages 1135-1143, 2015. 8" + ], + "bbox": [ + 93, + 272, + 482, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Delving deep into rectifiers: Surpassing human-level performance on imagenet classification. In ICCV, pages 1026-1034, 2015. 1", + "[16] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 5", + "[17] Xu He, Keke Tang, Yawen Shi, Yin Li, Weilong Peng, and Peican Zhu. Are deep point cloud classifiers suffer from out-of-distribution overconfidence issue? In SMC, pages 2620-2627. IEEE, 2023. 2", + "[18] Matthias Hein, Maksym Andriushchenko, and Julian Bitterwolf. Why relu networks yield high-confidence predictions far away from the training data and how to mitigate the problem. In CVPR, pages 41-50, 2019. 1, 2, 5", + "[19] Dan Hendrycks and Kevin Gimpel. A baseline for detecting misclassified and out-of-distribution examples in neural networks. In ICLR, 2016. 2, 5", + "[20] Dan Hendrycks, Mantas Mazeika, and Thomas Dietterich. Deep anomaly detection with outlier exposure. In ICLR, 2019. 1, 2", + "[21] Dan Hendrycks, Steven Basart, Mantas Mazeika, Andy Zou, Joseph Kwon, Mohammadreza Mostajabi, Jacob Steinhardt, and Dawn Song. Scaling out-of-distribution detection for real-world settings. In ICML, pages 8759-8773, 2022. 5", + "[22] Geoffrey Hinton. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. 2, 3, 5, 8", + "[23] Gao Huang, Zhuang Liu, Laurens Van Der Maaten, and Kilian Q Weinberger. Densely connected convolutional networks. In CVPR, pages 4700-4708, 2017. 5", + "[24] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 5", + "[25] Yann LeCun, Yoshua Bengio, and Geoffrey Hinton. Deep learning. Nature, 521(7553):436-444, 2015. 1", + "[26] Kimin Lee, Honglak Lee, Kibok Lee, and Jinwoo Shin. Training confidence-calibrated classifiers for detecting out-of-distribution samples. In ICLR, 2018. 1, 2", + "[27] Kimin Lee, Kibok Lee, Honglak Lee, and Jinwoo Shin. A simple unified framework for detecting out-of-distribution samples and adversarial attacks. In NeurIPS, pages 7167-7177, 2018. 2", + "[28] Shiyu Liang, Yixuan Li, and R Srikant. Enhancing the reliability of out-of-distribution image detection in neural networks. In ICLR, 2018. 2, 3", + "[29] Weitang Liu, Xiaoyun Wang, John D Owens, and Yixuan Li. Energy-based out-of-distribution detection. In NeurIPS, pages 21464-21475, 2020. 2, 5", + "[30] Alexander Meinke and Matthias Hein. Towards neural networks that provably know when they don't know. In ICLR, 2020. 5", + "[31] Melanie Mitchell. Complexity: A guided tour. Oxford University Press, 2009. 2", + "[32] Rafael Müller, Simon Kornblith, and Geoffrey Hinton. When does label smoothing help? In NeurIPS, pages 4694-4703, 2019. 2", + "[33] Yuval Netzer, Tao Wang, Adam Coates, Alessandro Bissacco, Bo Wu, and Andrew Y Ng. Reading digits in natural images with unsupervised feature learning. In NeurIPS" + ], + "bbox": [ + 514, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "5038", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Workshop on Deep Learning and Unsupervised Feature Learning, 2011. 5", + "[34] Anh Nguyen, Jason Yosinski, and Jeff Clune. Deep neural networks are easily fooled: High confidence predictions for unrecognizable images. In CVPR, pages 427-436, 2015. 1", + "[35] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Köpf, Edward Yang, Zach DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. Pytorch: An imperative style, high-performance deep learning library. In NeurIPS, pages 8026-8037, 2019. 5", + "[36] Jiacheng Qiao, Chengzhi Zhong, Peican Zhu, and Keke Tang. Advscod: Bayesian-based out-of-distribution detection via curvature sketching and adversarial sample enrichment. Mathematics, 11(3):692, 2023. 2", + "[37] Adriana Romero, Nicolas Ballas, Samira Ebrahimi Kahou, Antoine Chassang, Carlo Gatta, and Yoshua Bengio. Fitnets: Hints for thin deep nets. arXiv preprint arXiv:1412.6550, 2014. 2, 3", + "[38] K Simonyan and A Zisserman. Very deep convolutional networks for large-scale image recognition. In ICLR, 2015. 3", + "[39] Yiyou Sun, Chuan Guo, and Yixuan Li. React: Out-of-distribution detection with rectified activations. In NeurIPS, 2021. 2, 5", + "[40] Yiyou Sun, Yifei Ming, Xiaojin Zhu, and Yixuan Li. Out-of-distribution detection with deep nearest neighbors. In ICML, 2022. 2", + "[41] Keke Tang, Dingruibo Miao, Weilong Peng, Jianpeng Wu, Yawen Shi, Zhaoquan Gu, Zhihong Tian, and Wenping Wang. Codes: Chamfer out-of-distribution examples against overconfidence issue. In ICCV, pages 1153-1162, 2021. 2", + "[42] Keke Tang, Xujian Cai, Weilong Peng, Shudong Li, and Wenping Wang. Ood attack: Generating overconfident out-of-distribution examples to fool deep neural classifiers. In ICIP, pages 1260-1264, 2023. 2", + "[43] Keke Tang, Xujian Cai, Weilong Peng, Daizong Liu, Peican Zhu, Pan Zhou, Zhihong Tian, and Wenping Wang. Matching words for out-of-distribution detection. In ICDM, pages 578-587, 2023. 2", + "[44] Keke Tang, Chao Hou, Weilong Peng, Runnan Chen, Peican Zhu, Wenping Wang, and Zhihong Tian. Cores: Convolutional response-based score for out-of-distribution detection. In CVPR, pages 10916-10925, 2024. 2", + "[45] Keke Tang, Wenyu Zhao, Weilong Peng, Xiang Fang, Xiaodong Cui, Peican Zhu, and Zhihong Tian. Reparaterization head for efficient multi-input networks. In ICASSP, pages 6190-6194, 2024. 1", + "[46] Keke Tang, Yuexin Ma, Dingruibo Miao, Peng Song, Zhaoquan Gu, Zhihong Tian, and Wenping Wang. Decision fusion networks for image classification. IEEE TNNLS, 36(3): 3890-3903, 2025. 1", + "[47] Leitian Tao, Xuefeng Du, Jerry Zhu, and Yixuan Li. Nonparametric outlier synthesis. In ICLR, 2022. 2", + "[48] Frederick Tung and Greg Mori. Similarity-preserving knowledge distillation. In ICCV, pages 1365-1374, 2019. 3" + ], + "bbox": [ + 91, + 92, + 483, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[49] Haoqi Wang, Zhizhong Li, Litong Feng, and Wayne Zhang. Vim: Out-of-distribution with virtual-logit matching. In CVPR, pages 4921-4930, 2022. 5", + "[50] Xiaobo Wang, Tianyu Fu, Shengcai Liao, Shuo Wang, Zhen Lei, and Tao Mei. Exclusivity-consistency regularized knowledge distillation for face recognition. In ECCV, pages 325-342, 2020. 2", + "[51] Zhi-Fan Wu, Tong Wei, Jianwen Jiang, Chaojie Mao, Mingqian Tang, and Yu-Feng Li. Ngc: A unified framework for learning with open-world noisy data. In ICCV, pages 62-71, 2021. 2", + "[52] Jianxiong Xiao, James Hays, Krista A Ehinger, Aude Oliva, and Antonio Torralba. Sun database: Large-scale scene recognition from abbey to zoo. In CVPR, pages 3485-3492, 2010. 5", + "[53] Kunran Xu, Lai Rui, Yishi Li, and Lin Gu. Feature normalized knowledge distillation for image classification. In ECCV, pages 664-680, 2020. 2", + "[54] Pingmei Xu, Krista A Ehinger, Yinda Zhang, Adam Finkelstein, Sanjeev R Kulkarni, and Jianxiong Xiao. Turkergaze: Crowdsourcing saliency with webcam based eye tracking. arXiv preprint arXiv:1504.06755, 2015. 5", + "[55] Jingkang Yang, Haoqi Wang, Litong Feng, Xiaopeng Yan, Huabin Zheng, Wayne Zhang, and Ziwei Liu. Semantically coherent out-of-distribution detection. In ICCV, pages 8301-8309, 2021. 2", + "[56] Jingkang Yang, Kaiyang Zhou, Yixuan Li, and Ziwei Liu. Generalized out-of-distribution detection: A survey. arXiv preprint arXiv:2110.11334, 2021. 2", + "[57] Junho Yim, Donggyu Joo, Jihoon Bae, and Junmo Kim. A gift from knowledge distillation: Fast optimization, network minimization and transfer learning. In CVPR, pages 4133-4141, 2017. 2", + "[58] Shan You, Chang Xu, Chao Xu, and Dacheng Tao. Learning from multiple teacher networks. In SIGKDD, pages 1285-1294, 2017. 2", + "[59] Fisher Yu, Ari Seff, Yinda Zhang, Shuran Song, Thomas Funkhouser, and Jianxiong Xiao. Lsun: Construction of a large-scale image dataset using deep learning with humans in the loop. arXiv preprint arXiv:1506.03365, 2015. 5", + "[60] Qing Yu and Kiyoharu Aizawa. Unsupervised out-of-distribution detection by maximum classifier discrepancy. In ICCV, pages 9518-9526, 2019. 2", + "[61] Yeonguk Yu, Sungho Shin, Seongju Lee, Changyun Jun, and Kyoobin Lee. Block selection method for using feature norm in out-of-distribution detection. In CVPR, pages 15701-15711, 2023. 5", + "[62] Yue Yuan, Rundong He, Zhongyi Han, and Yilong Yin. Lhact: Rectifying extremely low and high activations for out-of-distribution detection. In MM, pages 8105-8113, 2023. 2", + "[63] Sergey Zagoruyko and Nikos Komodakis. Wide residual networks. In BMVC, 2016. 5", + "[64] Sergey Zagoruyko and Nikos Komodakis. Paying more attention to attention: Improving the performance of convolutional neural networks via attention transfer. In ICLR, 2022. 3" + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "5039", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[65] Bolei Zhou, Agata Lapedriza, Aditya Khosla, Aude Oliva, and Antonio Torralba. Places: A 10 million image database for scene recognition. TPAMI, 40(6):1452-1464, 2017. 5", + "[66] Zhi-Hua Zhou. Why over-parameterization of deep neural networks does not overfit. Science China Information Sciences, 64(1):1-3, 2021. 2", + "[67] Yao Zhu, YueFeng Chen, Chuanlong Xie, Xiaodan Li, Rong Zhang, Xiang Tian, Yaowu Chen, et al. Boosting out-of-distribution detection with typical features. In NeurIPS, 2022. 2" + ], + "bbox": [ + 91, + 92, + 480, + 231 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "5040", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Simplification Is All You Need against Out-of-Distribution Overconfidence/c4dc8f46-aeb5-4477-a94f-497bd92530cf_model.json b/2025/Simplification Is All You Need against Out-of-Distribution Overconfidence/c4dc8f46-aeb5-4477-a94f-497bd92530cf_model.json new file mode 100644 index 0000000000000000000000000000000000000000..62d599a0a25616e3abe4b6ae9115048f6bf52d74 --- /dev/null +++ b/2025/Simplification Is All You Need against Out-of-Distribution Overconfidence/c4dc8f46-aeb5-4477-a94f-497bd92530cf_model.json @@ -0,0 +1,2433 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.125, + 0.13, + 0.877, + 0.155 + ], + "angle": 0, + "content": "Simplification Is All You Need against Out-of-Distribution Overconfidence" + }, + { + "type": "text", + "bbox": [ + 0.262, + 0.193, + 0.734, + 0.231 + ], + "angle": 0, + "content": "Keke Tang\\(^{1*}\\), Chao Hou\\(^{1*}\\), Weilong Peng\\(^{1\\dagger}\\), Xiang Fang\\(^{2}\\), Zhize Wu\\(^{3}\\), Yongwei Nie\\(^{4}\\), Wenping Wang\\(^{5}\\), Zhihong Tian\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.157, + 0.241, + 0.842, + 0.278 + ], + "angle": 0, + "content": "\\(^{1}\\)Guangzhou University \\(^{2}\\)Nanyang Technological University \\(^{3}\\)Hefei University \\(^{4}\\)South China University of Technology \\(^{5}\\)Texas A&M University" + }, + { + "type": "text", + "bbox": [ + 0.13, + 0.28, + 0.868, + 0.313 + ], + "angle": 0, + "content": "tangbohutbh@gmail.com, houchaohk@gmail.com, wlpeng@tju.edu.cn, xfang9508@gmail.com, wuzz@hfuu.edu.cn, nieyongwei@scut.edu.cn, wenping@amu.edu, tianzhihong@gzhu.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.356, + 0.329, + 0.373 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.388, + 0.486, + 0.645 + ], + "angle": 0, + "content": "Deep neural networks (DNNs) often exhibit out-of-distribution (OOD) overconfidence, producing overly confident predictions on OOD samples. We attribute this issue to the inherent over-complexity of DNNs and investigate two key aspects: capacity and nonlinearity. First, we demonstrate that reducing model capacity through knowledge distillation can effectively mitigate OOD overconfidence. Second, we show that selectively reducing nonlinearity by removing ReLU operations further alleviates the issue. Building on these findings, we present a practical guide to model simplification, combining both strategies to significantly reduce OOD overconfidence. Extensive experiments validate the effectiveness of this approach in mitigating OOD overconfidence and demonstrate its superiority over state-of-the-art methods. Additionally, our simplification strategies can be combined with existing OOD detection techniques to further enhance OOD detection performance." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.672, + 0.224, + 0.687 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.697, + 0.485, + 0.879 + ], + "angle": 0, + "content": "Deep neural networks (DNNs) have revolutionized image classification tasks, consistently achieving state-of-the-art performance [8, 15, 25, 45, 46]. Despite this success, DNN classifiers are inherently designed for static, closed-world environments [2], limiting their ability to adapt to novel, unforeseen concepts encountered in real-world scenarios, known as out-of-distribution (OOD) samples. Indeed, many recent studies demonstrate that DNN classifiers suffer from OOD overconfidence, generating overly confident predictions on OOD samples [13, 34]. This issue significantly impacts the reliability and safety of DNNs in real-world applications, highlighting an urgent need for effective solutions." + }, + { + "type": "image", + "bbox": [ + 0.527, + 0.355, + 0.707, + 0.448 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.709, + 0.355, + 0.89, + 0.448 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.526, + 0.449, + 0.709, + 0.541 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.71, + 0.449, + 0.89, + 0.557 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.564, + 0.543, + 0.887, + 0.557 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.563, + 0.908, + 0.66 + ], + "angle": 0, + "content": "Figure 1. Histograms (logarithmic scale) of maximum confidence values for the teacher network (ResNet-50), the distilled student network (ResNet-18), and its variant with additional ReLU removal. All models are trained on CIFAR-100 and evaluated on OOD datasets: (a) SVHN, (b) LSUN-R, (c) iSUN, and (d) TinyImageNet-R. The \\(y\\)-axis represents the frequency of samples, and the \\(x\\)-axis represents confidence scores." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.69, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Extensive research has been dedicated to addressing the OOD overconfidence issue. One common approach involves calibrating outputs for both in-distribution (ID) and OOD samples to improve discrimination [20], enabling OOD detectors to proactively reject OOD inputs. While these methods can reduce the impact of overconfidence, they do not fundamentally resolve the issue within DNN classifiers. Moreover, their effectiveness heavily depends on the classifiers' ability to adequately separate predictive distributions; they fail when the classifiers cannot achieve this separation [26]. As a more intrinsic solution, another approach trains DNN classifiers to suppress predictions on a subset of typical OOD samples, with the expectation that this behavior will generalize to other OOD data [18, 20]." + }, + { + "type": "page_footnote", + "bbox": [ + 0.115, + 0.887, + 0.352, + 0.901 + ], + "angle": 0, + "content": "*Joint first authors. †Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "5030" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.121 + ], + "angle": 0, + "content": "However, given the vast diversity of potential OOD samples, the effectiveness of this method remains limited." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.122, + 0.482, + 0.287 + ], + "angle": 0, + "content": "DNNs can indeed be regarded as complex systems, often characterized by over-parameterization, in which the number of parameters exceeds the amount of available training data [66]. This excessive complexity can cause DNNs to exhibit emergent behaviors, leading to unpredictable outcomes, as observed in other complex systems [31]. Consequently, when presented with certain OOD inputs, DNNs may unexpectedly but confidently misclassify them as ID classes. Therefore, investigating the relationship between model complexity and OOD overconfidence presents a promising path toward effectively addressing this issue." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.288, + 0.485, + 0.544 + ], + "angle": 0, + "content": "In this paper, we explore two key aspects of model complexity: capacity and nonlinearity, and their roles in influencing OOD overconfidence. For capacity, we use knowledge distillation to examine how varying the student network's capacity affects OOD overconfidence, finding that an appropriately reduced capacity can significantly alleviate this issue (see Fig. 1). For nonlinearity, we investigate selective ReLU removal at different ratios and locations within the network, observing that controlled reductions in nonlinearity similarly mitigate OOD overconfidence (see Fig. 1). Based on these insights, we present a practical guide to model simplification that combines these two strategies to mitigate the OOD overconfidence issue in DNNs more effectively, achieving superior results compared to state-of-the-art methods. Furthermore, we demonstrate that our approach can be effectively combined with existing OOD detection techniques to further enhance performance." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.545, + 0.451, + 0.559 + ], + "angle": 0, + "content": "Overall, our contribution is summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.562, + 0.483, + 0.606 + ], + "angle": 0, + "content": "- We are the first to attribute the OOD overconfidence issue in DNNs to their over-complexity, identifying model simplification as a crucial and effective solution." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.607, + 0.483, + 0.667 + ], + "angle": 0, + "content": "- We systematically investigate two fundamental aspects of complexity: capacity and nonlinearity, analyzing their effects on OOD overconfidence through knowledge distillation and selective ReLU removal." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.668, + 0.483, + 0.713 + ], + "angle": 0, + "content": "- We present a practical guide for model simplification and demonstrate through extensive experiments its effectiveness in significantly mitigating OOD overconfidence." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.562, + 0.483, + 0.713 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.725, + 0.233, + 0.74 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Mitigating OOD Overconfidence. The issue of OOD overconfidence, where DNNs assign excessively high confidence scores to OOD samples, presents a substantial security risk for real-world applications [18, 42, 56]. Current approaches often incorporate auxiliary data to improve OOD sensitivity, using techniques such as adding random noise and image shuffling [26], leveraging large volumes of unlabeled data [51, 55, 60], relying on external OOD datasets [20], or generating synthetic OOD samples via models like Chamfer GAN [41]. Additionally, some" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.151 + ], + "angle": 0, + "content": "methods, including VOS [9] and NPOS [47], create virtual features in feature space rather than explicit OOD instances. However, these approaches typically require substantial volumes of OOD data." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.153, + 0.905, + 0.242 + ], + "angle": 0, + "content": "In contrast, our approach addresses OOD overconfidence by reducing model complexity, thus eliminating the need for auxiliary OOD data. Through extensive experiments, we systematically explore the effects of capacity reduction and selective activation removal, presenting a data-independent alternative for mitigating OOD overconfidence." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.243, + 0.905, + 0.514 + ], + "angle": 0, + "content": "OD Detection. Another approach to managing OOD overconfidence is to develop scoring or normalization techniques that increase the distinction between ID and OOD samples [11, 17, 36, 43]. A foundational method, maximum softmax probability, introduced by Hendrycks and Gimpel [19], demonstrates that classifiers typically assign lower confidence to OOD samples. Follow-up methods, including ODIN [28] with temperature scaling and input perturbation, Energy Score [29], and density-based scores [27, 40], further enhance this ID-OOD separation. To improve robustness, several techniques refine feature representations: feature activation rectification [39, 62], feature selection [67], neuron pruning [3], and methods like CORES [44], which identify layer-specific relevant kernels for OOD detection. In contrast, our approach inherently addresses OOD overconfidence through model simplification. By tackling the root of the overconfidence issue, our method naturally enhances OOD detection performance as a result." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.515, + 0.905, + 0.649 + ], + "angle": 0, + "content": "Knowledge Distillation. Knowledge distillation transfers knowledge from a larger teacher model to a smaller student model, enabling the student to approximate the teacher's performance [22]. It typically utilizes response-based knowledge (soft logits) [1, 22, 32], feature-based knowledge (intermediate feature maps) [37, 50, 53], and relation-based knowledge (layer/sample relationships) [57, 58]. We use distillation to reduce model capacity and mitigate OOD overconfidence." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.65, + 0.905, + 0.74 + ], + "angle": 0, + "content": "ReLU Activation. ReLU \\((f(x) = \\max(0, x))\\) introduces nonlinearity while maintaining efficiency, making it widely used in deep learning [10, 12]. However, ReLU networks are prone to OOD overconfidence [18]. We address this by selectively removing ReLU activations to mitigate OOD overconfidence." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.755, + 0.885, + 0.772 + ], + "angle": 0, + "content": "3. How Complexity Affects Overconfidence?" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.905, + 0.87 + ], + "angle": 0, + "content": "The complexity of a neural network is influenced by various factors. To investigate whether model complexity affects the degree of OOD overconfidence in DNNs, we focus on two key aspects: capacity and nonlinearity. In this section, we examine each of these factors individually to better understand their impact on OOD overconfidence." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.905, + 0.9 + ], + "angle": 0, + "content": "Evaluation Metrics. To assess the severity of OOD overconfidence across different DNNs, we use the mean max" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "5031" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.089, + 0.486, + 0.146 + ], + "angle": 0, + "content": "Table 1. Comparison of MMC and OOD detection performance for student models (ResNet-18) distilled from a teacher model (ResNet-50) using different knowledge distillation methods on CIFAR datasets. Results are averaged across multiple OOD datasets." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.15, + 0.48, + 0.242 + ], + "angle": 0, + "content": "
IDCIFAR-10CIFAR-100
MethodMMC↓FPR95↓AUROC↑MMC↓FPR95↓AUROC↑
Teacher79.5854.2391.7872.1283.6774.73
KD [22]76.8149.8192.4262.6979.7776.31
FitNet [37]79.4650.5892.2660.6077.5377.78
AT [64]77.9048.0992.6964.1182.1475.51
SP [48]78.8150.5292.2258.7577.1280.01
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.261, + 0.484, + 0.385 + ], + "angle": 0, + "content": "imum confidence (MMC) on OOD samples as our evaluation metric. Additionally, to analyze how OOD overconfidence impacts the model's ability for ID/OOD discrimination, we also report the area under the receiver operating characteristic curve (AUROC) [6], the false positive rate at a true positive rate of \\(95\\%\\) (FPR95) [28], and the area under the precision-recall curve, with ID samples specified as positives (AUPRin) [38]." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.393, + 0.199, + 0.41 + ], + "angle": 0, + "content": "3.1. Capacity" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.416, + 0.484, + 0.568 + ], + "angle": 0, + "content": "To investigate whether model capacity affects the severity of OOD overconfidence, we employ knowledge distillation to transfer knowledge from a large-capacity teacher network to a smaller-capacity student network. Specifically, we examine how different knowledge distillation strategies and varying student network capacities influence OOD overconfidence. Since DNN capacity is influenced by various factors, such as depth, width, and number of parameters, we ensure a fair comparison by testing networks with the same architecture but different depths." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.569, + 0.484, + 0.764 + ], + "angle": 0, + "content": "Effectiveness of Knowledge Distillation. To validate the effectiveness of knowledge distillation in mitigating OOD overconfidence, we select ResNet-50 as the teacher network and ResNet-18 as the student network, applying four distinct knowledge distillation methods: KD [22], FitNet [37], AT [64], and SP [48]. Experiments are conducted using CIFAR-10 and CIFAR-100 as ID datasets, with SVHN, LSUN-R, iSUN, Textures, and TinyImageNet-R as OOD datasets. The results, shown in Tab. 1, indicate that all four distillation methods lead to reductions in MMC and enhancements in AUROC and FPR95 metrics, demonstrating that knowledge distillation effectively mitigates the OOD overconfidence issue." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Analysis of Student Networks with Varying Capacities. We conduct experiments on two datasets using KD [22] with student networks of varying capacities. For CIFAR-100, we set ResNet-101 as the teacher and evaluate student networks with capacities ranging from ResNet-50, ResNet-34, ResNet-18, to ResNet-10. For ImageNet-100, we use ResNet-152 as the teacher, with ResNet-101, ResNet-50, ResNet-34, and ResNet-18 as student models. As shown in Fig. 2, on CIFAR-100, reducing the student" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.089, + 0.907, + 0.133 + ], + "angle": 0, + "content": "Table 2. Comparison of MMC and OOD detection performance for different ReLU removal proportions using ResNet-101 as the classifier. Results are averaged across multiple OOD datasets." + }, + { + "type": "table", + "bbox": [ + 0.524, + 0.136, + 0.897, + 0.253 + ], + "angle": 0, + "content": "
IDCIFAR-10CIFAR-100
RatioMMC↓FPR95↓AUROC↑MMC↓FPR95↓AUROC↑
0%83.4651.9292.5373.8484.4173.03
1%84.4657.5890.4869.1482.4376.22
2%77.1843.0093.6972.2683.0775.25
3%77.9146.9393.0669.3778.7478.12
5%83.0756.5390.7168.7178.5779.43
10%79.6050.0692.2571.2480.9477.70
20%80.7850.3192.3073.5482.7374.81
50%79.2248.3492.7370.4082.2275.40
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.27, + 0.907, + 0.438 + ], + "angle": 0, + "content": "network capacity to ResNet-18 improves OOD overconfidence and enhances ID/OOD discrimination. However, further reducing capacity to ResNet-10 results in degraded performance, likely due to insufficient capacity to effectively handle CIFAR-100 classification. A similar trend is observed on ImageNet-100, where reducing the student network capacity to ResNet-50 continues to yield improvements, but further reductions impair performance, likely due to the dataset's increased complexity. These findings suggest that reducing network capacity to a moderate level can effectively alleviate the OOD overconfidence issue." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.446, + 0.649, + 0.461 + ], + "angle": 0, + "content": "3.2. Nonlinearity" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.469, + 0.906, + 0.575 + ], + "angle": 0, + "content": "Nonlinearity in DNNs arises from activation functions, e.g., ReLU and Sigmoid. The more activation operations are applied, the greater the degree of nonlinearity. In this subsection, we investigate whether the degree of nonlinearity affects the severity of OOD overconfidence. Given that ReLU is one of the most widely used activation functions, we focus on examining the effect of removing ReLU operations." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.575, + 0.906, + 0.711 + ], + "angle": 0, + "content": "Specifically, we examine the impact of removing ReLU operations at different proportions and locations, similar to the approach in [4]. as we will detail in Sec. 4, to evaluate their effect on mitigating OOD overconfidence. Additionally, we assess the influence of initiating ReLU removal at various stages of training. For these experiments, CIFAR-10 and CIFAR-100 are used as the ID datasets, with SVHN, LSUN-R, iSUN, Textures, and TinyImageNet-R as the OOD datasets, using ResNet-18/101 as the classifier." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.711, + 0.905, + 0.787 + ], + "angle": 0, + "content": "Analysis of Different Removal Proportions. To assess the impact of varying removal proportions, we introduce the Negative Response Ratio (NRO), which measures the proportion of negative responses that are passed to ReLU operations:" + }, + { + "type": "equation", + "bbox": [ + 0.529, + 0.796, + 0.905, + 0.841 + ], + "angle": 0, + "content": "\\[\nN R O = \\frac {\\text {N u m b e r o f n e g a t i v e c o n v o l u t i o n a l r e s p o n s e s}}{\\text {T o t a l n u m b e r o f c o n v o l u t i o n a l r e s p o n s e s}}. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.841, + 0.907, + 0.903 + ], + "angle": 0, + "content": "A smaller NRO indicates fewer activations requiring truncation, suggesting that the corresponding ReLU layers contribute less to overall nonlinearity. Based on NRO values, we selectively remove ReLU operations with the low" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5032" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.115, + 0.089, + 0.496, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.498, + 0.089, + 0.881, + 0.227 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.114, + 0.222, + 0.306, + 0.356 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.308, + 0.23, + 0.496, + 0.356 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.498, + 0.23, + 0.688, + 0.356 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.69, + 0.23, + 0.88, + 0.356 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.402, + 0.36, + 0.608, + 0.376 + ], + "angle": 0, + "content": "Teacher Student" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.38, + 0.908, + 0.409 + ], + "angle": 0, + "content": "Figure 2. Comparison of MMC and OOD detection metrics, including FPR95, AUROC, and AUPRin, for student models with varying capacities (e.g., ResNet models) across datasets: (a) CIFAR-100 and (b) ImageNet-100." + }, + { + "type": "image", + "bbox": [ + 0.116, + 0.424, + 0.455, + 0.564 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.091, + 0.569, + 0.483, + 0.598 + ], + "angle": 0, + "content": "Figure 3. Comparison of MMC and FPR95 for ResNet-101 with ReLU removal applied at different network locations." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.623, + 0.483, + 0.729 + ], + "angle": 0, + "content": "est NRO scores at proportions of \\(1\\%\\), \\(2\\%\\), \\(3\\%\\), \\(5\\%\\), \\(10\\%\\), \\(20\\%\\), and \\(50\\%\\), starting at the 40th epoch and continuing through the 100th epoch. As shown in Tab. 2, removing a certain proportion of ReLU operations reduces OOD overconfidence and enhances OOD detection. For example, removing \\(2\\%\\) of ReLU operations on CIFAR-10 and \\(5\\%\\) on CIFAR-100 yields the best results." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.732, + 0.483, + 0.883 + ], + "angle": 0, + "content": "Analysis of Different Removal Locations. To identify which ReLU operations are most suitable for removal, we compare the effects of removing ReLU operations from shallow, middle, and deep layers, as well as from layers with the lowest NRO values. Following previous findings, we remove \\(2\\%\\) of ReLU operations for CIFAR-10 and \\(5\\%\\) for CIFAR-100, initiating removal at the 40th epoch. As shown in Fig. 3, each removal strategy effectively reduces OOD overconfidence, with the best results achieved by removing ReLU operations based on NRO values." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.886, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Impact of ReLU Removal Initiation Points. We also" + }, + { + "type": "image", + "bbox": [ + 0.543, + 0.429, + 0.88, + 0.527 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.533, + 0.906, + 0.562 + ], + "angle": 0, + "content": "Figure 4. Comparison of MMC and FPR95 for ResNet-18 with ReLU removal initiated at different training epochs." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.588, + 0.907, + 0.68 + ], + "angle": 0, + "content": "analyze the effects of initiating ReLU removal at different stages during training. As illustrated in Fig. 4, the specific stage at which ReLU removal begins has only a moderate impact on overall performance. The best results are achieved when removal is initiated around the middle of the training process, such as at the 40th epoch." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.694, + 0.89, + 0.712 + ], + "angle": 0, + "content": "4. A Practical Guide to Model Simplification" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Based on our findings reported in Sec. 3, we present a practical guide to model simplification for addressing the OOD overconfidence issue in DNNs. For models that exhibit high OOD overconfidence, we suggest reducing complexity through a combination of knowledge distillation and selective ReLU removal. Please refer to Fig. 5 for demonstration. Knowledge Distillation with a Moderately-Sized Student. The first step is to choose a small student model with a capacity suited to the task requirements. This student model should be sufficiently compact to reduce complexity but still capable of performing well on the original task. We then apply knowledge distillation (e.g., KD) to transfer es" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5033" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.164, + 0.088, + 0.837, + 0.315 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.319, + 0.908, + 0.364 + ], + "angle": 0, + "content": "Figure 5. Illustration of our model simplification pipeline. Given a complex teacher model with high capacity and nonlinearity, we use knowledge distillation to transfer information to a low-capacity student model. During distillation, ReLU operations are selectively replaced with Leaky ReLU, and the negative slope is iteratively adjusted from 0 to 1, ultimately transforming ReLU into an identity function." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.388, + 0.482, + 0.448 + ], + "angle": 0, + "content": "sential knowledge from the larger teacher model to the student. This process enables the student to approximate the teacher's performance while operating with a reduced capacity, thereby alleviating the OOD overconfidence issue." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.449, + 0.483, + 0.599 + ], + "angle": 0, + "content": "Selective ReLU Removal During Distillation. To further streamline the model, we selectively remove ReLU operations. Using the NRO metric defined in Eq. (1), we identify ReLU operations that contribute minimally to the model's responses. Following the approach in [4], we replace these selected ReLU operations with Leaky ReLU, initially setting the negative slope to 0 and gradually increasing it to 1 midway through the distillation process. This controlled transition effectively transforms ReLU into an identity function, thereby reducing the model's nonlinearity." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.6, + 0.483, + 0.66 + ], + "angle": 0, + "content": "This practical guide, combining knowledge distillation for capacity reduction with selective ReLU removal for nonlinearity reduction, provides an accessible approach to mitigate OOD overconfidence issue in DNNs." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.672, + 0.296, + 0.689 + ], + "angle": 0, + "content": "5. Experimental Results" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.697, + 0.283, + 0.714 + ], + "angle": 0, + "content": "5.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Implementation Details. We employ KD [22] to transfer knowledge from teacher to student models, with a weight ratio of 1.5:1 between KD loss and cross-entropy (CE) loss. All student models are trained for 100 epochs, with selective ReLU removal starting at the 40th epoch and continuing through the end of training, similar to the approach in [4]. We remove \\(2\\%\\) of ReLU operations for models trained on CIFAR-10, \\(5\\%\\) for models trained on CIFAR-100, and \\(5\\%\\) for models trained on ImageNet-1k [7]. All experiments are implemented in PyTorch [35] and conducted on a workstation equipped with dual 2.40 GHz CPUs, 128 GB RAM, and eight NVIDIA RTX 3090Ti GPUs." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.388, + 0.907, + 0.585 + ], + "angle": 0, + "content": "Setup. For (1) Mitigating OOD Overconfidence, we use CIFAR-10 and CIFAR-100 [24] as ID datasets, with SVHN [33], LSUN-R [59], iSUN [54], Textures [5], and TINR, a resized variant of TinyImageNet uniformly scaled to \\(32 \\times 32\\), as the OOD datasets. For (2) OOD Detection, we use the same CIFAR-based configuration as in (1). For large-scale evaluations, we use ImageNet-1k [7] as the ID dataset, with SUN [52], Places [65], Textures [5], and OpenImage-O [49] as the OOD datasets. We train ResNet-50 and ResNet-101 [16] on CIFAR-10 and CIFAR-100, and train DenseNet-201 [23] and WideResNet-101-2 (WRN-101) [63] on ImageNet-1k [7], following the training protocols specified in their respective original papers." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.592, + 0.909, + 0.803 + ], + "angle": 0, + "content": "Setting of Our Approach. We designate ResNet-18 and WideResNet-16-10 (WRN-16) as student networks for ResNet-50, and ResNet-34 and WideResNet-28-10 (WRN-28) as students for ResNet-101 on CIFAR datasets. For the ImageNet-1k dataset, DenseNet-161 serves as the student for DenseNet-201, and WideResNet-50-2 (WRN-50) for WRN-101. All student models are trained using the SGD optimizer with a momentum of 0.9 and an initial learning rate of 0.1. On the ImageNet-1k benchmark, the learning rate is reduced by a factor of 10 every 30 epochs, while on the CIFAR benchmarks, the learning rate is adjusted using the CosineAnnealingLR scheduler. For simplicity, we refer to our method as \"Ours (student)\" in subsequent sections to indicate the specific student network employed." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Baselines. For (1) Mitigating OOD Overconfidence, we select three state-of-the-art baselines: ACET [18], CEDA [18], and CCUs [30]. For (2) OOD Detection, we benchmark our approach against five post-hoc methods: maximum softmax probability (MSP) [19], MaxLogit [21], Energy [29], ReAct [39], and FeatureNorm [61]." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5034" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.908, + 0.131 + ], + "angle": 0, + "content": "Table 3. Comparison of MMC on OOD datasets for ResNet-50 and ResNet-101 models trained on CIFAR-10 and CIFAR-100, evaluated using various OOD overconfidence mitigation techniques. Underlined values denote the best performance for a single method, while bolded values indicate the best performance across all methods, including both individual and combined methods." + }, + { + "type": "table", + "bbox": [ + 0.16, + 0.134, + 0.843, + 0.419 + ], + "angle": 0, + "content": "
ModelMethodID: CIFAR-10AverageID: CIFAR-100Average
SVHNLSUN-RiSUNTexturesTINRSVHNLSUN-RiSUNTexturesTINR
ResNet-50—75.3479.6479.8681.5381.5379.5872.5072.0772.2072.7471.1172.12
CEDA74.7370.9172.0273.3778.6473.9355.6454.7756.2562.8555.4256.99
ACET74.2967.3567.7270.3769.8769.9263.6862.4962.1964.4461.1562.79
CCUs69.5663.5163.1068.6468.9266.7557.3352.3054.7161.2554.8956.09
Ours (ResNet-18)78.2473.2974.3779.5176.9876.4854.7543.6542.2761.4039.6948.35
Ours (WRN-16)63.6566.3268.4977.0970.1869.1543.0651.7651.9957.9652.3751.43
CEDA + Ours (WRN-16)61.3167.4069.6470.0973.5468.4047.8459.4060.6959.2460.0057.43
ACET + Ours (WRN-16)67.4461.2862.3670.0565.4865.3252.5459.4561.5960.8260.4858.98
CCUs + Ours (WRN-16)65.9758.8161.3264.8465.9663.3850.8849.0451.8558.4051.8252.40
ResNet-101—83.4482.1383.0284.5684.1783.4665.2577.2178.0373.3875.3273.84
CEDA70.0768.2668.4274.4274.0271.0458.8258.1859.0066.2358.4760.14
ACET75.8253.6955.0064.5059.9861.8063.6259.5758.4460.3458.0160.00
CCUs73.5665.7264.9866.9469.8368.2158.4554.5356.5259.8056.3257.12
Ours (ResNet-34)75.2274.8175.9482.0878.8677.3863.6258.4260.1366.6959.8961.75
Ours (WRN-28)71.2174.7376.0381.8077.8976.3355.6947.8451.2063.6954.0954.50
CEDA + Ours (WRN-28)67.1568.2170.5976.7773.8371.3156.5165.1266.5962.5166.7763.50
ACET + Ours (WRN-28)59.3352.7053.0766.9558.0558.0256.1062.4863.9458.3463.8660.95
CCUs + Ours (WRN-28)67.7961.7464.2065.9568.8065.7053.7150.9653.5857.8654.5254.13
" + }, + { + "type": "image_caption", + "bbox": [ + 0.128, + 0.432, + 0.143, + 0.442 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.127, + 0.442, + 0.273, + 0.513 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.129, + 0.512, + 0.143, + 0.52 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.127, + 0.52, + 0.273, + 0.594 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.351, + 0.433, + 0.374, + 0.441 + ], + "angle": 0, + "content": "ACET" + }, + { + "type": "image", + "bbox": [ + 0.276, + 0.441, + 0.421, + 0.514 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.433, + 0.522, + 0.441 + ], + "angle": 0, + "content": "CEDA" + }, + { + "type": "image", + "bbox": [ + 0.423, + 0.441, + 0.569, + 0.515 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.646, + 0.433, + 0.67, + 0.441 + ], + "angle": 0, + "content": "CCUs" + }, + { + "type": "image", + "bbox": [ + 0.571, + 0.441, + 0.717, + 0.515 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.773, + 0.433, + 0.838, + 0.441 + ], + "angle": 0, + "content": "Ours (ResNet-18)" + }, + { + "type": "image", + "bbox": [ + 0.719, + 0.441, + 0.865, + 0.515 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.129, + 0.602, + 0.143, + 0.611 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image", + "bbox": [ + 0.127, + 0.611, + 0.273, + 0.685 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.129, + 0.681, + 0.143, + 0.69 + ], + "angle": 0, + "content": "(d)" + }, + { + "type": "image", + "bbox": [ + 0.127, + 0.689, + 0.274, + 0.763 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.351, + 0.602, + 0.385, + 0.61 + ], + "angle": 0, + "content": "ACET" + }, + { + "type": "image", + "bbox": [ + 0.276, + 0.61, + 0.421, + 0.685 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.497, + 0.602, + 0.521, + 0.61 + ], + "angle": 0, + "content": "CEDA" + }, + { + "type": "image", + "bbox": [ + 0.423, + 0.61, + 0.569, + 0.684 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.646, + 0.603, + 0.67, + 0.61 + ], + "angle": 0, + "content": "CCUs" + }, + { + "type": "image", + "bbox": [ + 0.571, + 0.61, + 0.717, + 0.684 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.778, + 0.604, + 0.834, + 0.612 + ], + "angle": 0, + "content": "Ours (WRN-28)" + }, + { + "type": "image", + "bbox": [ + 0.719, + 0.612, + 0.865, + 0.686 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.128, + 0.767, + 0.907, + 0.809 + ], + "angle": 0, + "content": "Figure 6. Histograms (logarithmic scale) of MMC values for (a, b) ResNet-50 and (c, d) ResNet-101, w/ and w/o applying mitigation techniques: ACET, CEDA, CCUs, and Ours. All models are trained on CIFAR-100 and evaluated on OOD datasets: (a, c) LSUN-R and (d) TinyImageNet-R. In our approach, ResNet-18 and WRN-28 serve as students for ResNet-50 and ResNet-101, respectively." + }, + { + "type": "title", + "bbox": [ + 0.127, + 0.823, + 0.457, + 0.839 + ], + "angle": 0, + "content": "Results of Mitigating OOD Overconfidence" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.482, + 0.901 + ], + "angle": 0, + "content": "Comparison with State-of-the-art Methods. To assess the effectiveness and superiority of our approach, we conduct comparative experiments evaluating OOD overconfi" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.824, + 0.907, + 0.899 + ], + "angle": 0, + "content": "dence for ResNet-50 and ResNet-101 trained on CIFAR-10 and CIFAR-100. The results in Tab. 3 demonstrate that our approach consistently reduces the mean maximum confidence (MMC) across different student networks in all cases, validating its effectiveness. In comparison, methods such" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5035" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.09, + 0.907, + 0.119 + ], + "angle": 0, + "content": "Table 4. Comparison of OOD detection performance for various methods on the CIFAR-10 and CIFAR-100 datasets. Underlined values indicate results enhanced by applying our approach to the baseline, while bolded values denote the best performance." + }, + { + "type": "table", + "bbox": [ + 0.097, + 0.123, + 0.907, + 0.557 + ], + "angle": 0, + "content": "
IDModelMethodOODAverage
SVHNLSUN-RiSUNTexturesTINR
FPR95↓AUROC↑FPR95↓AUROC↑FPR95↓AUROC↑FPR95↓AUROC↑FPR95↓AUROC↑FPR95↓AUROC↑
CIFAR-10ResNet-50MSP43.7594.1954.3291.9655.4691.6659.0190.3158.6190.8054.2391.78
MSP + Ours (ResNet-18)51.8293.2741.9393.8944.8693.2755.6490.5750.1891.9748.8992.59
MSP + Ours (WRN-16)28.1896.1235.7495.0139.7894.3457.0089.7942.3593.7540.6193.80
MaxLogit22.0796.3132.0394.9534.9594.5346.3391.3944.9692.5736.0793.95
MaxLogit + Ours (ResNet-18)30.2594.9425.0695.3427.9194.7344.5091.7136.7592.7332.8993.89
MaxLogit + Ours (WRN-16)17.6196.9618.5496.9221.6996.3947.7089.5827.8195.5526.6795.08
Energy20.3596.4029.7995.1132.8094.6745.4191.4443.7692.6434.4294.05
Energy + Ours (ResNet-18)28.5395.0723.9995.4326.5594.8344.1389.7035.9292.7931.8293.56
Energy + Ours (WRN-16)18.8096.8818.0797.0220.9196.5047.8989.5427.6395.6126.6695.11
ReAct37.6691.1016.2496.6419.6595.8446.7286.2930.7993.4630.2192.67
ReAct + Ours (ResNet-18)31.5994.6223.8795.5126.7894.8643.6090.1935.8792.9232.3493.62
ReAct + Ours (WRN-16)19.0796.8417.9097.0320.6696.5147.4589.8427.2495.6826.4695.18
FeatureNorm9.2597.9777.3482.0770.2983.9648.9784.1667.7184.4954.7186.53
FeatureNorm + Ours (ResNet-18)2.7399.3524.3495.7521.0496.2822.8294.8929.5294.3620.0996.13
FeatureNorm + Ours (WRN-16)3.3799.2622.0396.4321.5496.4753.6778.8529.3594.5025.9993.10
CIFAR-100ResNet-101MSP76.3081.1486.6769.7688.1169.0685.6973.7385.2771.4784.4173.03
MSP + Ours (ResNet-34)77.4878.9272.7483.0974.8382.1482.5077.2674.5282.2076.4180.72
MSP + Ours (WRN-28)74.8781.3962.5785.7166.9483.9383.2876.3369.4481.7071.4281.81
MaxLogit68.0586.7782.8474.3484.9473.4087.2574.8181.9275.0581.0076.87
MaxLogit + Ours (ResNet-34)71.6983.4763.7487.2665.6986.5481.7478.4666.8886.4469.9584.43
MaxLogit + Ours (WRN-28)73.8482.6154.3089.1160.2987.2683.6277.0664.1284.9867.2384.20
Energy68.3886.9482.3874.6484.2173.6788.0374.6281.4075.2580.8877.02
Energy + Ours (ResNet-34)69.2084.1458.8388.0561.1087.3781.6878.5162.8987.1866.7485.05
Energy + Ours (WRN-28)74.1782.6148.6090.0755.0588.1284.1077.0160.4185.7164.4784.70
ReAct65.6888.1486.4469.1886.9769.0983.4480.1383.8071.5981.2775.63
ReAct + Ours (ResNet-34)60.2688.3057.9486.4958.1186.5263.0785.7860.0385.6959.8886.56
ReAct + Ours (WRN-28)71.8986.7347.3489.2053.6687.4680.5781.4858.9784.6662.4985.91
FeatureNorm31.8894.5999.5434.1199.1039.0253.4282.3298.2242.3276.4358.47
FeatureNorm + Ours (ResNet-34)41.4790.4792.4671.4991.4172.8965.1880.3389.9969.0076.1076.84
FeatureNorm + Ours (WRN-28)54.1789.0880.8679.5280.6979.9171.4974.3778.0080.7473.0480.72
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.575, + 0.483, + 0.697 + ], + "angle": 0, + "content": "as CEDA, ACET, and CCUs, which employ specific techniques to suppress confidence scores during training, also significantly mitigate OOD overconfidence in DNNs. However, our approach, which does not incorporate specific confidence suppression measures, still outperforms these state-of-the-art methods, particularly in the more challenging CIFAR-100 scenario. This further highlights the robustness and effectiveness of our strategy." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.7, + 0.483, + 0.792 + ], + "angle": 0, + "content": "On CIFAR-10, while our base method may not show a pronounced advantage over others, combining our approach with confidence suppression techniques provides additional improvements, surpassing the performance of these state-of-the-art methods. These results reinforce the effectiveness of our approach in addressing OOD overconfidence." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.483, + 0.903 + ], + "angle": 0, + "content": "Visualization. We visualize the maximum confidence scores predicted by various DNN classifiers as histograms (logarithmic scale) in Fig. 6. As shown, applying OOD overconfidence mitigation solutions shifts the confidence distributions of the Baseline to the left, substantially reducing the number of samples with high confidence scores. Notably, our approach achieves the most pronounced shift, un" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.575, + 0.907, + 0.607 + ], + "angle": 0, + "content": "derscoring its effectiveness and validating its superiority in mitigating OOD overconfidence." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.618, + 0.836, + 0.635 + ], + "angle": 0, + "content": "5.3. Results of Enhancing OOD Detection" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.643, + 0.907, + 0.794 + ], + "angle": 0, + "content": "CIFAR Experiments. We evaluate the effectiveness of our approach in enhancing OOD detection. The results in Tab. 4 demonstrate that applying our method to existing OOD detection approaches leads to significant performance improvements. Even when directly applied to the classifier without any advanced post-hoc processing, i.e., MSP, the performance enhancement is still noticeable. These improvements are consistent across different student model configurations, datasets, and classifiers, validating the effectiveness of our approach in boosting OOD detection." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.909, + 0.903 + ], + "angle": 0, + "content": "ImageNet Experiments. We further validate the effectiveness of our approach in enhancing OOD detection on the ImageNet-1k dataset. The results in Tab. 5 indicate that, similar to the CIFAR experiments, applying our approach significantly improves OOD detection performance across methods, demonstrating the robustness and scalability of our approach in large-scale settings." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.956 + ], + "angle": 0, + "content": "5036" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.195, + 0.09, + 0.799, + 0.104 + ], + "angle": 0, + "content": "Table 5. Comparison of OOD detection performance for various methods on the ImageNet-1k dataset." + }, + { + "type": "table", + "bbox": [ + 0.1, + 0.109, + 0.898, + 0.383 + ], + "angle": 0, + "content": "
ModelMethodOODAverage
SUNPlacesTexturesOpenImage-O
FPR95↓AUROC↑FPR95↓AUROC↑FPR95↓AUROC↑FPR95↓AUROC↑FPR95↓AUROC↑
DenseNet-201MSP65.7682.2469.5381.1467.8279.3664.9284.4767.0181.80
MSP + Ours (DenseNet-161)66.3181.7268.6181.3266.9779.5562.1285.1066.0081.92
MaxLogit55.4486.0661.8983.9958.4883.4559.6486.9458.8685.11
MaxLogit + Ours (DenseNet-161)55.3387.6759.9586.2655.8085.8460.0388.6057.7887.09
Energy52.8786.0460.4683.7156.1383.3559.8786.6557.3384.94
Energy + Ours (DenseNet-161)50.8888.2556.9086.6052.2586.4558.7688.6054.7087.48
ReAct46.0090.8157.4686.1155.1484.1969.3274.4356.9883.89
ReAct + Ours (DenseNet-161)49.2188.5555.0286.9450.9986.7557.6188.6453.2187.72
FeatureNorm43.0690.0555.7185.4250.4679.9079.9767.2257.3080.65
FeatureNorm + Ours (DenseNet-161)36.5291.7949.2487.5056.7378.2982.2065.4956.1780.77
WRN-101MSP66.9382.4470.4881.3565.1480.9159.8786.1365.6182.71
MSP + Ours (WRN-50)67.7582.8069.0781.8166.3880.9660.9186.7866.0383.09
MaxLogit63.2584.7568.4082.6957.3984.3855.8488.4361.2285.06
MaxLogit + Ours (WRN-50)61.9286.1864.3484.4056.7485.6055.6989.2559.6786.36
Energy65.0684.4670.3282.2757.1184.3758.0488.0262.6384.78
Energy + Ours (WRN-50)63.1885.9666.0684.0055.9085.6857.8788.7760.7586.10
ReAct52.2288.1158.7685.4757.0984.9949.6387.8254.4386.60
ReAct + Ours (WRN-50)37.0292.2448.5488.7050.9087.3654.4985.9247.7488.56
FeatureNorm70.0778.8978.6173.2320.9894.7975.0974.8161.1980.43
FeatureNorm + Ours (WRN-50)60.2283.6774.0175.9617.7196.0969.9478.9055.4783.66
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.403, + 0.482, + 0.472 + ], + "angle": 0, + "content": "Table 6. Comparison of MMC and OOD detection performance for ResNet-50 enhanced by Ours (ResNet-18) and its variants w/o knowledge distillation (KD) and ReLU removal (RR), using CIFAR-10 as the ID dataset, and SVHN, LSUN-R, iSUN, Textures, and TINR as the OOD datasets. Results are averaged." + }, + { + "type": "table", + "bbox": [ + 0.171, + 0.477, + 0.408, + 0.549 + ], + "angle": 0, + "content": "
KDRRMMC↓FPR95↓AUROC↑
79.5854.2391.78
✓76.8149.8192.42
✓78.4650.5091.76
✓✓76.4848.8992.59
" + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.574, + 0.404, + 0.59 + ], + "angle": 0, + "content": "5.4. Ablation Studies and More Analysis" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.599, + 0.483, + 0.749 + ], + "angle": 0, + "content": "Importance of Knowledge Distillation and ReLU Removal. We conduct comprehensive ablation studies to evaluate the impact of knowledge distillation and ReLU removal, the two essential components of our approach. As shown in Tab. 6, excluding either component notably diminishes the effectiveness of our method in mitigating OOD overconfidence and complicates the accuracy of OOD detection. These findings validate the critical importance of both knowledge distillation and selective ReLU removal in achieving optimal performance with our approach." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Knowledge Distillation vs. Network Pruning for Capacity Reduction. We further investigate an alternative capacity reduction method for our approach by applying network pruning. To facilitate a fair comparison with knowledge distillation, we evaluate ResNet-50 and ResNet-101 models trained on CIFAR-10, applying two strategies: (1) using KD [22] to reduce ResNet-50 to ResNet-18 and ResNet-101 to WRN-28, and (2) applying Prune [14] to ResNet-50 and ResNet-101, reducing them to parameter counts comparable to ResNet-18 and WRN-28, respectively. The re" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.403, + 0.907, + 0.472 + ], + "angle": 0, + "content": "Table 7. Comparison of MMC and OOD detection performance for ResNet-50 and ResNet-101 reduced to comparable parameter counts via KD [22] and Prune [14]. CIFAR-10 serves as the ID dataset, with SVHN, LSUN-R, iSUN, Textures, and TINR as OOD datasets. Results are averaged across all five OOD datasets" + }, + { + "type": "table", + "bbox": [ + 0.531, + 0.477, + 0.892, + 0.563 + ], + "angle": 0, + "content": "
ModelCapacity ReductionMMC↓FPR95↓AUROC↑
ResNet-50—79.5854.2391.78
Prune77.3452.5592.27
KD (ResNet-18)76.8149.8192.42
ResNet-101—83.4651.9292.53
Prune78.2050.9291.61
KD (WRN-28)77.0141.0992.56
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.589, + 0.906, + 0.695 + ], + "angle": 0, + "content": "sults in Tab. 7 show that both methods effectively reduce OOD overconfidence and improve OOD detection; however, knowledge distillation consistently delivers a more substantial improvement than network pruning. Additionally, knowledge distillation offers greater stability and flexibility during training, making it a more suitable choice for capacity reduction in our approach." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.71, + 0.783, + 0.725 + ], + "angle": 0, + "content": "6. Conclusion and Future Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.907, + 0.901 + ], + "angle": 0, + "content": "This paper has presented a comprehensive investigation into mitigating OOD overconfidence in DNNs through model simplification. By reducing model capacity via knowledge distillation and decreasing nonlinearity through selective ReLU removal, we demonstrate that OOD overconfidence in DNN classifiers can be significantly mitigated. Furthermore, these strategies can be effectively combined with OOD detection methods to enhance overall performance. In future work, we plan to study the effects of model simplification on OOD overconfidence in greater depth and from a theoretical perspective." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5037" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.26, + 0.108 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.484, + 0.237 + ], + "angle": 0, + "content": "We sincerely thank the anonymous reviewers for their valuable comments. This work was supported in part by the National Natural Science Foundation of China (62472117, U2436208, 62406095), the Guangdong Basic and Applied Basic Research Foundation (2025A1515010157, 2024A1515012064), the Science and Technology Projects in Guangzhou (2025A03J0137, SL2022A04J01112), and the Academician Bixing Fang's Specialized Class." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.248, + 0.188, + 0.264 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.273, + 0.483, + 0.3 + ], + "angle": 0, + "content": "[1] Lei Jimmy Ba and Rich Caruana. Do deep nets really need to be deep? In NeurIPS, pages 2654-2662, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.302, + 0.483, + 0.328 + ], + "angle": 0, + "content": "[2] Abhijit Bendale and Terrance Boult. Towards open world recognition. In CVPR, pages 1893-1902, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.33, + 0.483, + 0.37 + ], + "angle": 0, + "content": "[3] Chao Chen, Zhihang Fu, Kai Liu, Ze Chen, Mingyuan Tao, and Jieping Ye. Optimal parameter and neuron pruning for out-of-distribution detection. In NeurIPS, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.372, + 0.482, + 0.412 + ], + "angle": 0, + "content": "[4] Hanting Chen, Yunhe Wang, Jianyuan Guo, and Dacheng Tao. Vanillanet: the power of minimalism in deep learning. In NeurIPS, 2024. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.414, + 0.482, + 0.455 + ], + "angle": 0, + "content": "[5] Mircea Cimpoi, Subhransu Maji, Iasonas Kokkinos, Sammy Mohamed, and Andrea Vedaldi. Describing textures in the wild. In CVPR, pages 3606-3613, 2014. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.457, + 0.482, + 0.495 + ], + "angle": 0, + "content": "[6] Jesse Davis and Mark Goadrich. The relationship between precision-recall and roc curves. In ICML, pages 233-240, 2006. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.497, + 0.482, + 0.538 + ], + "angle": 0, + "content": "[7] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, pages 248-255, 2009. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.54, + 0.482, + 0.62 + ], + "angle": 0, + "content": "[8] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.623, + 0.482, + 0.662 + ], + "angle": 0, + "content": "[9] Xuefeng Du, Zhaoning Wang, Mu Cai, and Yixuan Li. Vos: Learning what you don't know by virtual outlier synthesis. In ICLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.665, + 0.482, + 0.718 + ], + "angle": 0, + "content": "[10] Shiv Ram Dubey, Satish Kumar Singh, and Bidyut Baran Chaudhuri. Activation functions in deep learning: A comprehensive survey and benchmark. Neurocomputing, 503: 92-108, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.721, + 0.482, + 0.774 + ], + "angle": 0, + "content": "[11] Xiang Fang, Arvind Easwaran, Blaise Genest, and Ponnuthurai Nagaratnam Suganthan. Your data is not perfect: Towards cross-domain out-of-distribution detection in class-imbalanced data. ESWA, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.776, + 0.482, + 0.816 + ], + "angle": 0, + "content": "[12] Xavier Glorot, Antoine Bordes, and Yoshua Bengio. Deep sparse rectifier neural networks. In AISTATS, pages 315-323, 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.818, + 0.482, + 0.857 + ], + "angle": 0, + "content": "[13] Ian J. Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. In ICLR, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.86, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[14] Song Han, Jeff Pool, John Tran, and William J Dally. Learning both weights and connections for efficient neural networks. In NeurIPS, pages 1135-1143, 2015. 8" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.273, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[15] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Delving deep into rectifiers: Surpassing human-level performance on imagenet classification. In ICCV, pages 1026-1034, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.149, + 0.905, + 0.189 + ], + "angle": 0, + "content": "[16] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.191, + 0.905, + 0.243 + ], + "angle": 0, + "content": "[17] Xu He, Keke Tang, Yawen Shi, Yin Li, Weilong Peng, and Peican Zhu. Are deep point cloud classifiers suffer from out-of-distribution overconfidence issue? In SMC, pages 2620-2627. IEEE, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.246, + 0.905, + 0.3 + ], + "angle": 0, + "content": "[18] Matthias Hein, Maksym Andriushchenko, and Julian Bitterwolf. Why relu networks yield high-confidence predictions far away from the training data and how to mitigate the problem. In CVPR, pages 41-50, 2019. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.302, + 0.905, + 0.342 + ], + "angle": 0, + "content": "[19] Dan Hendrycks and Kevin Gimpel. A baseline for detecting misclassified and out-of-distribution examples in neural networks. In ICLR, 2016. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.344, + 0.904, + 0.383 + ], + "angle": 0, + "content": "[20] Dan Hendrycks, Mantas Mazeika, and Thomas Dietterich. Deep anomaly detection with outlier exposure. In ICLR, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.385, + 0.905, + 0.44 + ], + "angle": 0, + "content": "[21] Dan Hendrycks, Steven Basart, Mantas Mazeika, Andy Zou, Joseph Kwon, Mohammadreza Mostajabi, Jacob Steinhardt, and Dawn Song. Scaling out-of-distribution detection for real-world settings. In ICML, pages 8759-8773, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.442, + 0.904, + 0.468 + ], + "angle": 0, + "content": "[22] Geoffrey Hinton. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. 2, 3, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.47, + 0.904, + 0.509 + ], + "angle": 0, + "content": "[23] Gao Huang, Zhuang Liu, Laurens Van Der Maaten, and Kilian Q Weinberger. Densely connected convolutional networks. In CVPR, pages 4700-4708, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.511, + 0.905, + 0.538 + ], + "angle": 0, + "content": "[24] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.54, + 0.905, + 0.566 + ], + "angle": 0, + "content": "[25] Yann LeCun, Yoshua Bengio, and Geoffrey Hinton. Deep learning. Nature, 521(7553):436-444, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.567, + 0.905, + 0.607 + ], + "angle": 0, + "content": "[26] Kimin Lee, Honglak Lee, Kibok Lee, and Jinwoo Shin. Training confidence-calibrated classifiers for detecting out-of-distribution samples. In ICLR, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.609, + 0.905, + 0.662 + ], + "angle": 0, + "content": "[27] Kimin Lee, Kibok Lee, Honglak Lee, and Jinwoo Shin. A simple unified framework for detecting out-of-distribution samples and adversarial attacks. In NeurIPS, pages 7167-7177, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.665, + 0.904, + 0.704 + ], + "angle": 0, + "content": "[28] Shiyu Liang, Yixuan Li, and R Srikant. Enhancing the reliability of out-of-distribution image detection in neural networks. In ICLR, 2018. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.707, + 0.905, + 0.747 + ], + "angle": 0, + "content": "[29] Weitang Liu, Xiaoyun Wang, John D Owens, and Yixuan Li. Energy-based out-of-distribution detection. In NeurIPS, pages 21464-21475, 2020. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.749, + 0.905, + 0.787 + ], + "angle": 0, + "content": "[30] Alexander Meinke and Matthias Hein. Towards neural networks that provably know when they don't know. In ICLR, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.79, + 0.905, + 0.816 + ], + "angle": 0, + "content": "[31] Melanie Mitchell. Complexity: A guided tour. Oxford University Press, 2009. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.818, + 0.905, + 0.857 + ], + "angle": 0, + "content": "[32] Rafael Müller, Simon Kornblith, and Geoffrey Hinton. When does label smoothing help? In NeurIPS, pages 4694-4703, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[33] Yuval Netzer, Tao Wang, Adam Coates, Alessandro Bissacco, Bo Wu, and Andrew Y Ng. Reading digits in natural images with unsupervised feature learning. In NeurIPS" + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "5038" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.126, + 0.093, + 0.483, + 0.12 + ], + "angle": 0, + "content": "Workshop on Deep Learning and Unsupervised Feature Learning, 2011. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.484, + 0.164 + ], + "angle": 0, + "content": "[34] Anh Nguyen, Jason Yosinski, and Jeff Clune. Deep neural networks are easily fooled: High confidence predictions for unrecognizable images. In CVPR, pages 427-436, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.165, + 0.483, + 0.275 + ], + "angle": 0, + "content": "[35] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Köpf, Edward Yang, Zach DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. Pytorch: An imperative style, high-performance deep learning library. In NeurIPS, pages 8026-8037, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.277, + 0.483, + 0.332 + ], + "angle": 0, + "content": "[36] Jiacheng Qiao, Chengzhi Zhong, Peican Zhu, and Keke Tang. Advscod: Bayesian-based out-of-distribution detection via curvature sketching and adversarial sample enrichment. Mathematics, 11(3):692, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.334, + 0.483, + 0.387 + ], + "angle": 0, + "content": "[37] Adriana Romero, Nicolas Ballas, Samira Ebrahimi Kahou, Antoine Chassang, Carlo Gatta, and Yoshua Bengio. Fitnets: Hints for thin deep nets. arXiv preprint arXiv:1412.6550, 2014. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.39, + 0.483, + 0.418 + ], + "angle": 0, + "content": "[38] K Simonyan and A Zisserman. Very deep convolutional networks for large-scale image recognition. In ICLR, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.42, + 0.483, + 0.46 + ], + "angle": 0, + "content": "[39] Yiyou Sun, Chuan Guo, and Yixuan Li. React: Out-of-distribution detection with rectified activations. In NeurIPS, 2021. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.462, + 0.483, + 0.502 + ], + "angle": 0, + "content": "[40] Yiyou Sun, Yifei Ming, Xiaojin Zhu, and Yixuan Li. Out-of-distribution detection with deep nearest neighbors. In ICML, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.504, + 0.483, + 0.56 + ], + "angle": 0, + "content": "[41] Keke Tang, Dingruibo Miao, Weilong Peng, Jianpeng Wu, Yawen Shi, Zhaoquan Gu, Zhihong Tian, and Wenping Wang. Codes: Chamfer out-of-distribution examples against overconfidence issue. In ICCV, pages 1153-1162, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.562, + 0.483, + 0.616 + ], + "angle": 0, + "content": "[42] Keke Tang, Xujian Cai, Weilong Peng, Shudong Li, and Wenping Wang. Ood attack: Generating overconfident out-of-distribution examples to fool deep neural classifiers. In ICIP, pages 1260-1264, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.618, + 0.483, + 0.672 + ], + "angle": 0, + "content": "[43] Keke Tang, Xujian Cai, Weilong Peng, Daizong Liu, Peican Zhu, Pan Zhou, Zhihong Tian, and Wenping Wang. Matching words for out-of-distribution detection. In ICDM, pages 578-587, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.675, + 0.483, + 0.73 + ], + "angle": 0, + "content": "[44] Keke Tang, Chao Hou, Weilong Peng, Runnan Chen, Peican Zhu, Wenping Wang, and Zhihong Tian. Cores: Convolutional response-based score for out-of-distribution detection. In CVPR, pages 10916-10925, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.732, + 0.483, + 0.787 + ], + "angle": 0, + "content": "[45] Keke Tang, Wenyu Zhao, Weilong Peng, Xiang Fang, Xiaodong Cui, Peican Zhu, and Zhihong Tian. Reparaterization head for efficient multi-input networks. In ICASSP, pages 6190-6194, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.789, + 0.483, + 0.842 + ], + "angle": 0, + "content": "[46] Keke Tang, Yuexin Ma, Dingruibo Miao, Peng Song, Zhaoquan Gu, Zhihong Tian, and Wenping Wang. Decision fusion networks for image classification. IEEE TNNLS, 36(3): 3890-3903, 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.845, + 0.483, + 0.872 + ], + "angle": 0, + "content": "[47] Leitian Tao, Xuefeng Du, Jerry Zhu, and Yixuan Li. Nonparametric outlier synthesis. In ICLR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.874, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[48] Frederick Tung and Greg Mori. Similarity-preserving knowledge distillation. In ICCV, pages 1365-1374, 2019. 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.135 + ], + "angle": 0, + "content": "[49] Haoqi Wang, Zhizhong Li, Litong Feng, and Wayne Zhang. Vim: Out-of-distribution with virtual-logit matching. In CVPR, pages 4921-4930, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.137, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[50] Xiaobo Wang, Tianyu Fu, Shengcai Liao, Shuo Wang, Zhen Lei, and Tao Mei. Exclusivity-consistency regularized knowledge distillation for face recognition. In ECCV, pages 325-342, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.193, + 0.905, + 0.247 + ], + "angle": 0, + "content": "[51] Zhi-Fan Wu, Tong Wei, Jianwen Jiang, Chaojie Mao, Mingqian Tang, and Yu-Feng Li. Ngc: A unified framework for learning with open-world noisy data. In ICCV, pages 62-71, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.249, + 0.905, + 0.303 + ], + "angle": 0, + "content": "[52] Jianxiong Xiao, James Hays, Krista A Ehinger, Aude Oliva, and Antonio Torralba. Sun database: Large-scale scene recognition from abbey to zoo. In CVPR, pages 3485-3492, 2010. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.306, + 0.905, + 0.348 + ], + "angle": 0, + "content": "[53] Kunran Xu, Lai Rui, Yishi Li, and Lin Gu. Feature normalized knowledge distillation for image classification. In ECCV, pages 664-680, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.349, + 0.905, + 0.404 + ], + "angle": 0, + "content": "[54] Pingmei Xu, Krista A Ehinger, Yinda Zhang, Adam Finkelstein, Sanjeev R Kulkarni, and Jianxiong Xiao. Turkergaze: Crowdsourcing saliency with webcam based eye tracking. arXiv preprint arXiv:1504.06755, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.406, + 0.905, + 0.46 + ], + "angle": 0, + "content": "[55] Jingkang Yang, Haoqi Wang, Litong Feng, Xiaopeng Yan, Huabin Zheng, Wayne Zhang, and Ziwei Liu. Semantically coherent out-of-distribution detection. In ICCV, pages 8301-8309, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.462, + 0.905, + 0.504 + ], + "angle": 0, + "content": "[56] Jingkang Yang, Kaiyang Zhou, Yixuan Li, and Ziwei Liu. Generalized out-of-distribution detection: A survey. arXiv preprint arXiv:2110.11334, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.505, + 0.905, + 0.559 + ], + "angle": 0, + "content": "[57] Junho Yim, Donggyu Joo, Jihoon Bae, and Junmo Kim. A gift from knowledge distillation: Fast optimization, network minimization and transfer learning. In CVPR, pages 4133-4141, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.562, + 0.905, + 0.602 + ], + "angle": 0, + "content": "[58] Shan You, Chang Xu, Chao Xu, and Dacheng Tao. Learning from multiple teacher networks. In SIGKDD, pages 1285-1294, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.604, + 0.905, + 0.659 + ], + "angle": 0, + "content": "[59] Fisher Yu, Ari Seff, Yinda Zhang, Shuran Song, Thomas Funkhouser, and Jianxiong Xiao. Lsun: Construction of a large-scale image dataset using deep learning with humans in the loop. arXiv preprint arXiv:1506.03365, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.661, + 0.905, + 0.702 + ], + "angle": 0, + "content": "[60] Qing Yu and Kiyoharu Aizawa. Unsupervised out-of-distribution detection by maximum classifier discrepancy. In ICCV, pages 9518-9526, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.704, + 0.905, + 0.758 + ], + "angle": 0, + "content": "[61] Yeonguk Yu, Sungho Shin, Seongju Lee, Changyun Jun, and Kyoobin Lee. Block selection method for using feature norm in out-of-distribution detection. In CVPR, pages 15701-15711, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.76, + 0.905, + 0.814 + ], + "angle": 0, + "content": "[62] Yue Yuan, Rundong He, Zhongyi Han, and Yilong Yin. Lhact: Rectifying extremely low and high activations for out-of-distribution detection. In MM, pages 8105-8113, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.817, + 0.905, + 0.843 + ], + "angle": 0, + "content": "[63] Sergey Zagoruyko and Nikos Komodakis. Wide residual networks. In BMVC, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.905, + 0.899 + ], + "angle": 0, + "content": "[64] Sergey Zagoruyko and Nikos Komodakis. Paying more attention to attention: Improving the performance of convolutional neural networks via attention transfer. In ICLR, 2022. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "5039" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.134 + ], + "angle": 0, + "content": "[65] Bolei Zhou, Agata Lapedriza, Aditya Khosla, Aude Oliva, and Antonio Torralba. Places: A 10 million image database for scene recognition. TPAMI, 40(6):1452-1464, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.482, + 0.177 + ], + "angle": 0, + "content": "[66] Zhi-Hua Zhou. Why over-parameterization of deep neural networks does not overfit. Science China Information Sciences, 64(1):1-3, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.179, + 0.482, + 0.232 + ], + "angle": 0, + "content": "[67] Yao Zhu, YueFeng Chen, Chuanlong Xie, Xiaodan Li, Rong Zhang, Xiang Tian, Yaowu Chen, et al. Boosting out-of-distribution detection with typical features. In NeurIPS, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.956 + ], + "angle": 0, + "content": "5040" + } + ] +] \ No newline at end of file diff --git a/2025/Simplification Is All You Need against Out-of-Distribution Overconfidence/c4dc8f46-aeb5-4477-a94f-497bd92530cf_origin.pdf b/2025/Simplification Is All You Need against Out-of-Distribution Overconfidence/c4dc8f46-aeb5-4477-a94f-497bd92530cf_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..b4901dc87eb68e1e1cc7ecb12904c9719562b9ba --- /dev/null +++ b/2025/Simplification Is All You Need against Out-of-Distribution Overconfidence/c4dc8f46-aeb5-4477-a94f-497bd92530cf_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d05fa0aa577b2936b50060c0dd6696fe10e43932c4e5ea35cf2ff92f64b8d77c +size 9174068 diff --git a/2025/Simplification Is All You Need against Out-of-Distribution Overconfidence/full.md b/2025/Simplification Is All You Need against Out-of-Distribution Overconfidence/full.md new file mode 100644 index 0000000000000000000000000000000000000000..59782e63b1fae3d5ee4039ebf179bc62789d9ff3 --- /dev/null +++ b/2025/Simplification Is All You Need against Out-of-Distribution Overconfidence/full.md @@ -0,0 +1,315 @@ +# Simplification Is All You Need against Out-of-Distribution Overconfidence + +Keke Tang $^{1*}$ , Chao Hou $^{1*}$ , Weilong Peng $^{1\dagger}$ , Xiang Fang $^{2}$ , Zhize Wu $^{3}$ , Yongwei Nie $^{4}$ , Wenping Wang $^{5}$ , Zhihong Tian $^{1}$ + +$^{1}$ Guangzhou University $^{2}$ Nanyang Technological University $^{3}$ Hefei University $^{4}$ South China University of Technology $^{5}$ Texas A&M University + +tangbohutbh@gmail.com, houchaohk@gmail.com, wlpeng@tju.edu.cn, xfang9508@gmail.com, wuzz@hfuu.edu.cn, nieyongwei@scut.edu.cn, wenping@amu.edu, tianzhihong@gzhu.edu.cn + +# Abstract + +Deep neural networks (DNNs) often exhibit out-of-distribution (OOD) overconfidence, producing overly confident predictions on OOD samples. We attribute this issue to the inherent over-complexity of DNNs and investigate two key aspects: capacity and nonlinearity. First, we demonstrate that reducing model capacity through knowledge distillation can effectively mitigate OOD overconfidence. Second, we show that selectively reducing nonlinearity by removing ReLU operations further alleviates the issue. Building on these findings, we present a practical guide to model simplification, combining both strategies to significantly reduce OOD overconfidence. Extensive experiments validate the effectiveness of this approach in mitigating OOD overconfidence and demonstrate its superiority over state-of-the-art methods. Additionally, our simplification strategies can be combined with existing OOD detection techniques to further enhance OOD detection performance. + +# 1. Introduction + +Deep neural networks (DNNs) have revolutionized image classification tasks, consistently achieving state-of-the-art performance [8, 15, 25, 45, 46]. Despite this success, DNN classifiers are inherently designed for static, closed-world environments [2], limiting their ability to adapt to novel, unforeseen concepts encountered in real-world scenarios, known as out-of-distribution (OOD) samples. Indeed, many recent studies demonstrate that DNN classifiers suffer from OOD overconfidence, generating overly confident predictions on OOD samples [13, 34]. This issue significantly impacts the reliability and safety of DNNs in real-world applications, highlighting an urgent need for effective solutions. + +![](images/5e6a4a1f5e3b5d07fabd2929ba9082d24079e99165782dac9f651cebbe002a7e.jpg) + +![](images/dce8fa53a43848f5d9336be0b0e3c2b2f9d31368c92f5d54078bbf5d6f2c43ca.jpg) + +![](images/916a7cfce59a808608cd0e609f3941a41b1aa0e793d6375a2c3ee3c3bae2dec2.jpg) + +![](images/5f9ef8a82b0e767c3e2a2dca39b494acf7a0e0ac697dcbde74f8bfb892ec4ef0.jpg) + +![](images/ff7334fac2b18610984438b898d30b9be4fc31f3283c45235f6b64b71e572110.jpg) +Figure 1. Histograms (logarithmic scale) of maximum confidence values for the teacher network (ResNet-50), the distilled student network (ResNet-18), and its variant with additional ReLU removal. All models are trained on CIFAR-100 and evaluated on OOD datasets: (a) SVHN, (b) LSUN-R, (c) iSUN, and (d) TinyImageNet-R. The $y$ -axis represents the frequency of samples, and the $x$ -axis represents confidence scores. + +Extensive research has been dedicated to addressing the OOD overconfidence issue. One common approach involves calibrating outputs for both in-distribution (ID) and OOD samples to improve discrimination [20], enabling OOD detectors to proactively reject OOD inputs. While these methods can reduce the impact of overconfidence, they do not fundamentally resolve the issue within DNN classifiers. Moreover, their effectiveness heavily depends on the classifiers' ability to adequately separate predictive distributions; they fail when the classifiers cannot achieve this separation [26]. As a more intrinsic solution, another approach trains DNN classifiers to suppress predictions on a subset of typical OOD samples, with the expectation that this behavior will generalize to other OOD data [18, 20]. + +However, given the vast diversity of potential OOD samples, the effectiveness of this method remains limited. + +DNNs can indeed be regarded as complex systems, often characterized by over-parameterization, in which the number of parameters exceeds the amount of available training data [66]. This excessive complexity can cause DNNs to exhibit emergent behaviors, leading to unpredictable outcomes, as observed in other complex systems [31]. Consequently, when presented with certain OOD inputs, DNNs may unexpectedly but confidently misclassify them as ID classes. Therefore, investigating the relationship between model complexity and OOD overconfidence presents a promising path toward effectively addressing this issue. + +In this paper, we explore two key aspects of model complexity: capacity and nonlinearity, and their roles in influencing OOD overconfidence. For capacity, we use knowledge distillation to examine how varying the student network's capacity affects OOD overconfidence, finding that an appropriately reduced capacity can significantly alleviate this issue (see Fig. 1). For nonlinearity, we investigate selective ReLU removal at different ratios and locations within the network, observing that controlled reductions in nonlinearity similarly mitigate OOD overconfidence (see Fig. 1). Based on these insights, we present a practical guide to model simplification that combines these two strategies to mitigate the OOD overconfidence issue in DNNs more effectively, achieving superior results compared to state-of-the-art methods. Furthermore, we demonstrate that our approach can be effectively combined with existing OOD detection techniques to further enhance performance. + +Overall, our contribution is summarized as follows: + +- We are the first to attribute the OOD overconfidence issue in DNNs to their over-complexity, identifying model simplification as a crucial and effective solution. +- We systematically investigate two fundamental aspects of complexity: capacity and nonlinearity, analyzing their effects on OOD overconfidence through knowledge distillation and selective ReLU removal. +- We present a practical guide for model simplification and demonstrate through extensive experiments its effectiveness in significantly mitigating OOD overconfidence. + +# 2. Related Work + +Mitigating OOD Overconfidence. The issue of OOD overconfidence, where DNNs assign excessively high confidence scores to OOD samples, presents a substantial security risk for real-world applications [18, 42, 56]. Current approaches often incorporate auxiliary data to improve OOD sensitivity, using techniques such as adding random noise and image shuffling [26], leveraging large volumes of unlabeled data [51, 55, 60], relying on external OOD datasets [20], or generating synthetic OOD samples via models like Chamfer GAN [41]. Additionally, some + +methods, including VOS [9] and NPOS [47], create virtual features in feature space rather than explicit OOD instances. However, these approaches typically require substantial volumes of OOD data. + +In contrast, our approach addresses OOD overconfidence by reducing model complexity, thus eliminating the need for auxiliary OOD data. Through extensive experiments, we systematically explore the effects of capacity reduction and selective activation removal, presenting a data-independent alternative for mitigating OOD overconfidence. + +OD Detection. Another approach to managing OOD overconfidence is to develop scoring or normalization techniques that increase the distinction between ID and OOD samples [11, 17, 36, 43]. A foundational method, maximum softmax probability, introduced by Hendrycks and Gimpel [19], demonstrates that classifiers typically assign lower confidence to OOD samples. Follow-up methods, including ODIN [28] with temperature scaling and input perturbation, Energy Score [29], and density-based scores [27, 40], further enhance this ID-OOD separation. To improve robustness, several techniques refine feature representations: feature activation rectification [39, 62], feature selection [67], neuron pruning [3], and methods like CORES [44], which identify layer-specific relevant kernels for OOD detection. In contrast, our approach inherently addresses OOD overconfidence through model simplification. By tackling the root of the overconfidence issue, our method naturally enhances OOD detection performance as a result. + +Knowledge Distillation. Knowledge distillation transfers knowledge from a larger teacher model to a smaller student model, enabling the student to approximate the teacher's performance [22]. It typically utilizes response-based knowledge (soft logits) [1, 22, 32], feature-based knowledge (intermediate feature maps) [37, 50, 53], and relation-based knowledge (layer/sample relationships) [57, 58]. We use distillation to reduce model capacity and mitigate OOD overconfidence. + +ReLU Activation. ReLU $(f(x) = \max(0, x))$ introduces nonlinearity while maintaining efficiency, making it widely used in deep learning [10, 12]. However, ReLU networks are prone to OOD overconfidence [18]. We address this by selectively removing ReLU activations to mitigate OOD overconfidence. + +# 3. How Complexity Affects Overconfidence? + +The complexity of a neural network is influenced by various factors. To investigate whether model complexity affects the degree of OOD overconfidence in DNNs, we focus on two key aspects: capacity and nonlinearity. In this section, we examine each of these factors individually to better understand their impact on OOD overconfidence. + +Evaluation Metrics. To assess the severity of OOD overconfidence across different DNNs, we use the mean max + +Table 1. Comparison of MMC and OOD detection performance for student models (ResNet-18) distilled from a teacher model (ResNet-50) using different knowledge distillation methods on CIFAR datasets. Results are averaged across multiple OOD datasets. + +
IDCIFAR-10CIFAR-100
MethodMMC↓FPR95↓AUROC↑MMC↓FPR95↓AUROC↑
Teacher79.5854.2391.7872.1283.6774.73
KD [22]76.8149.8192.4262.6979.7776.31
FitNet [37]79.4650.5892.2660.6077.5377.78
AT [64]77.9048.0992.6964.1182.1475.51
SP [48]78.8150.5292.2258.7577.1280.01
+ +imum confidence (MMC) on OOD samples as our evaluation metric. Additionally, to analyze how OOD overconfidence impacts the model's ability for ID/OOD discrimination, we also report the area under the receiver operating characteristic curve (AUROC) [6], the false positive rate at a true positive rate of $95\%$ (FPR95) [28], and the area under the precision-recall curve, with ID samples specified as positives (AUPRin) [38]. + +# 3.1. Capacity + +To investigate whether model capacity affects the severity of OOD overconfidence, we employ knowledge distillation to transfer knowledge from a large-capacity teacher network to a smaller-capacity student network. Specifically, we examine how different knowledge distillation strategies and varying student network capacities influence OOD overconfidence. Since DNN capacity is influenced by various factors, such as depth, width, and number of parameters, we ensure a fair comparison by testing networks with the same architecture but different depths. + +Effectiveness of Knowledge Distillation. To validate the effectiveness of knowledge distillation in mitigating OOD overconfidence, we select ResNet-50 as the teacher network and ResNet-18 as the student network, applying four distinct knowledge distillation methods: KD [22], FitNet [37], AT [64], and SP [48]. Experiments are conducted using CIFAR-10 and CIFAR-100 as ID datasets, with SVHN, LSUN-R, iSUN, Textures, and TinyImageNet-R as OOD datasets. The results, shown in Tab. 1, indicate that all four distillation methods lead to reductions in MMC and enhancements in AUROC and FPR95 metrics, demonstrating that knowledge distillation effectively mitigates the OOD overconfidence issue. + +Analysis of Student Networks with Varying Capacities. We conduct experiments on two datasets using KD [22] with student networks of varying capacities. For CIFAR-100, we set ResNet-101 as the teacher and evaluate student networks with capacities ranging from ResNet-50, ResNet-34, ResNet-18, to ResNet-10. For ImageNet-100, we use ResNet-152 as the teacher, with ResNet-101, ResNet-50, ResNet-34, and ResNet-18 as student models. As shown in Fig. 2, on CIFAR-100, reducing the student + +Table 2. Comparison of MMC and OOD detection performance for different ReLU removal proportions using ResNet-101 as the classifier. Results are averaged across multiple OOD datasets. + +
IDCIFAR-10CIFAR-100
RatioMMC↓FPR95↓AUROC↑MMC↓FPR95↓AUROC↑
0%83.4651.9292.5373.8484.4173.03
1%84.4657.5890.4869.1482.4376.22
2%77.1843.0093.6972.2683.0775.25
3%77.9146.9393.0669.3778.7478.12
5%83.0756.5390.7168.7178.5779.43
10%79.6050.0692.2571.2480.9477.70
20%80.7850.3192.3073.5482.7374.81
50%79.2248.3492.7370.4082.2275.40
+ +network capacity to ResNet-18 improves OOD overconfidence and enhances ID/OOD discrimination. However, further reducing capacity to ResNet-10 results in degraded performance, likely due to insufficient capacity to effectively handle CIFAR-100 classification. A similar trend is observed on ImageNet-100, where reducing the student network capacity to ResNet-50 continues to yield improvements, but further reductions impair performance, likely due to the dataset's increased complexity. These findings suggest that reducing network capacity to a moderate level can effectively alleviate the OOD overconfidence issue. + +# 3.2. Nonlinearity + +Nonlinearity in DNNs arises from activation functions, e.g., ReLU and Sigmoid. The more activation operations are applied, the greater the degree of nonlinearity. In this subsection, we investigate whether the degree of nonlinearity affects the severity of OOD overconfidence. Given that ReLU is one of the most widely used activation functions, we focus on examining the effect of removing ReLU operations. + +Specifically, we examine the impact of removing ReLU operations at different proportions and locations, similar to the approach in [4]. as we will detail in Sec. 4, to evaluate their effect on mitigating OOD overconfidence. Additionally, we assess the influence of initiating ReLU removal at various stages of training. For these experiments, CIFAR-10 and CIFAR-100 are used as the ID datasets, with SVHN, LSUN-R, iSUN, Textures, and TinyImageNet-R as the OOD datasets, using ResNet-18/101 as the classifier. + +Analysis of Different Removal Proportions. To assess the impact of varying removal proportions, we introduce the Negative Response Ratio (NRO), which measures the proportion of negative responses that are passed to ReLU operations: + +$$ +N R O = \frac {\text {N u m b e r o f n e g a t i v e c o n v o l u t i o n a l r e s p o n s e s}}{\text {T o t a l n u m b e r o f c o n v o l u t i o n a l r e s p o n s e s}}. \tag {1} +$$ + +A smaller NRO indicates fewer activations requiring truncation, suggesting that the corresponding ReLU layers contribute less to overall nonlinearity. Based on NRO values, we selectively remove ReLU operations with the low + +![](images/19135528d01fbdfee79fa48fbdcda4f0bbc6b063ba79534135e0e1344c342092.jpg) + +![](images/b8889e50737340affa915d527572fc1e1bb309e3feca8bf34d0bc30c4f98eaf6.jpg) + +![](images/b7bbd30e25a50659fc0a95cd265e3c0a4fa6a07ed765e9a686a75683f3190403.jpg) +Figure 2. Comparison of MMC and OOD detection metrics, including FPR95, AUROC, and AUPRin, for student models with varying capacities (e.g., ResNet models) across datasets: (a) CIFAR-100 and (b) ImageNet-100. + +![](images/0b695f5d3704bfe92f4f6a91eb4ec4f0903c73dcfe1cfa4ecbbaa9cafc7e32d6.jpg) +Teacher Student + +![](images/301ad9c7ffc67ad80272d3b6901be0ab800881fe239a99c2e8aa6d17f490666a.jpg) + +![](images/521aa6f536e610a027968e764225eecf6777f02685bbb4831f6844b2bd32fd44.jpg) + +![](images/01f77a8c33e4dc7cae2ae8746efb3e495b5e1a150cbfd698a4aca7d73c6ef79a.jpg) +Figure 3. Comparison of MMC and FPR95 for ResNet-101 with ReLU removal applied at different network locations. + +est NRO scores at proportions of $1\%$ , $2\%$ , $3\%$ , $5\%$ , $10\%$ , $20\%$ , and $50\%$ , starting at the 40th epoch and continuing through the 100th epoch. As shown in Tab. 2, removing a certain proportion of ReLU operations reduces OOD overconfidence and enhances OOD detection. For example, removing $2\%$ of ReLU operations on CIFAR-10 and $5\%$ on CIFAR-100 yields the best results. + +Analysis of Different Removal Locations. To identify which ReLU operations are most suitable for removal, we compare the effects of removing ReLU operations from shallow, middle, and deep layers, as well as from layers with the lowest NRO values. Following previous findings, we remove $2\%$ of ReLU operations for CIFAR-10 and $5\%$ for CIFAR-100, initiating removal at the 40th epoch. As shown in Fig. 3, each removal strategy effectively reduces OOD overconfidence, with the best results achieved by removing ReLU operations based on NRO values. + +Impact of ReLU Removal Initiation Points. We also + +![](images/09125d787ec7e270736996341582f23c8db509209b4a91dacf06d71b306f5a76.jpg) +Figure 4. Comparison of MMC and FPR95 for ResNet-18 with ReLU removal initiated at different training epochs. + +analyze the effects of initiating ReLU removal at different stages during training. As illustrated in Fig. 4, the specific stage at which ReLU removal begins has only a moderate impact on overall performance. The best results are achieved when removal is initiated around the middle of the training process, such as at the 40th epoch. + +# 4. A Practical Guide to Model Simplification + +Based on our findings reported in Sec. 3, we present a practical guide to model simplification for addressing the OOD overconfidence issue in DNNs. For models that exhibit high OOD overconfidence, we suggest reducing complexity through a combination of knowledge distillation and selective ReLU removal. Please refer to Fig. 5 for demonstration. Knowledge Distillation with a Moderately-Sized Student. The first step is to choose a small student model with a capacity suited to the task requirements. This student model should be sufficiently compact to reduce complexity but still capable of performing well on the original task. We then apply knowledge distillation (e.g., KD) to transfer es + +![](images/bf8fc906eb6da835625877dfe77bf98c2395802d61b870b8320876d358dc070b.jpg) +Figure 5. Illustration of our model simplification pipeline. Given a complex teacher model with high capacity and nonlinearity, we use knowledge distillation to transfer information to a low-capacity student model. During distillation, ReLU operations are selectively replaced with Leaky ReLU, and the negative slope is iteratively adjusted from 0 to 1, ultimately transforming ReLU into an identity function. + +sential knowledge from the larger teacher model to the student. This process enables the student to approximate the teacher's performance while operating with a reduced capacity, thereby alleviating the OOD overconfidence issue. + +Selective ReLU Removal During Distillation. To further streamline the model, we selectively remove ReLU operations. Using the NRO metric defined in Eq. (1), we identify ReLU operations that contribute minimally to the model's responses. Following the approach in [4], we replace these selected ReLU operations with Leaky ReLU, initially setting the negative slope to 0 and gradually increasing it to 1 midway through the distillation process. This controlled transition effectively transforms ReLU into an identity function, thereby reducing the model's nonlinearity. + +This practical guide, combining knowledge distillation for capacity reduction with selective ReLU removal for nonlinearity reduction, provides an accessible approach to mitigate OOD overconfidence issue in DNNs. + +# 5. Experimental Results + +# 5.1. Experimental Setup + +Implementation Details. We employ KD [22] to transfer knowledge from teacher to student models, with a weight ratio of 1.5:1 between KD loss and cross-entropy (CE) loss. All student models are trained for 100 epochs, with selective ReLU removal starting at the 40th epoch and continuing through the end of training, similar to the approach in [4]. We remove $2\%$ of ReLU operations for models trained on CIFAR-10, $5\%$ for models trained on CIFAR-100, and $5\%$ for models trained on ImageNet-1k [7]. All experiments are implemented in PyTorch [35] and conducted on a workstation equipped with dual 2.40 GHz CPUs, 128 GB RAM, and eight NVIDIA RTX 3090Ti GPUs. + +Setup. For (1) Mitigating OOD Overconfidence, we use CIFAR-10 and CIFAR-100 [24] as ID datasets, with SVHN [33], LSUN-R [59], iSUN [54], Textures [5], and TINR, a resized variant of TinyImageNet uniformly scaled to $32 \times 32$ , as the OOD datasets. For (2) OOD Detection, we use the same CIFAR-based configuration as in (1). For large-scale evaluations, we use ImageNet-1k [7] as the ID dataset, with SUN [52], Places [65], Textures [5], and OpenImage-O [49] as the OOD datasets. We train ResNet-50 and ResNet-101 [16] on CIFAR-10 and CIFAR-100, and train DenseNet-201 [23] and WideResNet-101-2 (WRN-101) [63] on ImageNet-1k [7], following the training protocols specified in their respective original papers. + +Setting of Our Approach. We designate ResNet-18 and WideResNet-16-10 (WRN-16) as student networks for ResNet-50, and ResNet-34 and WideResNet-28-10 (WRN-28) as students for ResNet-101 on CIFAR datasets. For the ImageNet-1k dataset, DenseNet-161 serves as the student for DenseNet-201, and WideResNet-50-2 (WRN-50) for WRN-101. All student models are trained using the SGD optimizer with a momentum of 0.9 and an initial learning rate of 0.1. On the ImageNet-1k benchmark, the learning rate is reduced by a factor of 10 every 30 epochs, while on the CIFAR benchmarks, the learning rate is adjusted using the CosineAnnealingLR scheduler. For simplicity, we refer to our method as "Ours (student)" in subsequent sections to indicate the specific student network employed. + +Baselines. For (1) Mitigating OOD Overconfidence, we select three state-of-the-art baselines: ACET [18], CEDA [18], and CCUs [30]. For (2) OOD Detection, we benchmark our approach against five post-hoc methods: maximum softmax probability (MSP) [19], MaxLogit [21], Energy [29], ReAct [39], and FeatureNorm [61]. + +Table 3. Comparison of MMC on OOD datasets for ResNet-50 and ResNet-101 models trained on CIFAR-10 and CIFAR-100, evaluated using various OOD overconfidence mitigation techniques. Underlined values denote the best performance for a single method, while bolded values indicate the best performance across all methods, including both individual and combined methods. + +
ModelMethodID: CIFAR-10AverageID: CIFAR-100Average
SVHNLSUN-RiSUNTexturesTINRSVHNLSUN-RiSUNTexturesTINR
ResNet-50—75.3479.6479.8681.5381.5379.5872.5072.0772.2072.7471.1172.12
CEDA74.7370.9172.0273.3778.6473.9355.6454.7756.2562.8555.4256.99
ACET74.2967.3567.7270.3769.8769.9263.6862.4962.1964.4461.1562.79
CCUs69.5663.5163.1068.6468.9266.7557.3352.3054.7161.2554.8956.09
Ours (ResNet-18)78.2473.2974.3779.5176.9876.4854.7543.6542.2761.4039.6948.35
Ours (WRN-16)63.6566.3268.4977.0970.1869.1543.0651.7651.9957.9652.3751.43
CEDA + Ours (WRN-16)61.3167.4069.6470.0973.5468.4047.8459.4060.6959.2460.0057.43
ACET + Ours (WRN-16)67.4461.2862.3670.0565.4865.3252.5459.4561.5960.8260.4858.98
CCUs + Ours (WRN-16)65.9758.8161.3264.8465.9663.3850.8849.0451.8558.4051.8252.40
ResNet-101—83.4482.1383.0284.5684.1783.4665.2577.2178.0373.3875.3273.84
CEDA70.0768.2668.4274.4274.0271.0458.8258.1859.0066.2358.4760.14
ACET75.8253.6955.0064.5059.9861.8063.6259.5758.4460.3458.0160.00
CCUs73.5665.7264.9866.9469.8368.2158.4554.5356.5259.8056.3257.12
Ours (ResNet-34)75.2274.8175.9482.0878.8677.3863.6258.4260.1366.6959.8961.75
Ours (WRN-28)71.2174.7376.0381.8077.8976.3355.6947.8451.2063.6954.0954.50
CEDA + Ours (WRN-28)67.1568.2170.5976.7773.8371.3156.5165.1266.5962.5166.7763.50
ACET + Ours (WRN-28)59.3352.7053.0766.9558.0558.0256.1062.4863.9458.3463.8660.95
CCUs + Ours (WRN-28)67.7961.7464.2065.9568.8065.7053.7150.9653.5857.8654.5254.13
+ +![](images/2a70a23a637a5e8fe3f704e446e6090892e643e0f8fa1c4cf8befca0eb7c2718.jpg) +(a) +(b) + +![](images/5450420c4e84847d0faa06662c292c27f4cd299cf0f3c2763cd43382beb50d1f.jpg) + +![](images/154d7c306de00cb139e4e86ecf243644f310051d2d3e7abe9102816e300c8e5b.jpg) +ACET + +![](images/d683d1fef2149b66067bae02fe792dedf427a55d07dac875bffc6d84f9c63b0a.jpg) +CEDA + +![](images/c58ce1c74130b039c4d9056eca9d0802ac214005ae37fab257fa844c8d247a26.jpg) +CCUs + +![](images/195006fd2e1c97a2d69f08dae8f6f38e157e8e06aeab95d519a2cb87a577b88b.jpg) +Ours (ResNet-18) + +![](images/e81ef09903ff89a34b2893ed9cc7fe4e1630ab89cf858bd580e0130fe071d9ec.jpg) +(c) +(d) + +![](images/c0844b9b5ba1f2864d2361bc5eb81364dc23e04710ccf26e08a3dae0c41f836d.jpg) +Figure 6. Histograms (logarithmic scale) of MMC values for (a, b) ResNet-50 and (c, d) ResNet-101, w/ and w/o applying mitigation techniques: ACET, CEDA, CCUs, and Ours. All models are trained on CIFAR-100 and evaluated on OOD datasets: (a, c) LSUN-R and (d) TinyImageNet-R. In our approach, ResNet-18 and WRN-28 serve as students for ResNet-50 and ResNet-101, respectively. + +![](images/5f3f12b6ef1143b59a059d7d12e983c1526e867a2d63a304238575ebc2ac6b72.jpg) +ACET + +![](images/ff89139269c01983fdda6fe307d4f08114af96d0a9f4a0bcb690b1f1e5430d4c.jpg) +CEDA + +![](images/011606af33bf67277ae955b252a1ece844b99209b093b7d7ddd7cd2b207b2f9c.jpg) +CCUs + +![](images/196b6591331d6b3a402901ff99e39d9e570fed47804d5134bef54eed362cd2e0.jpg) +Ours (WRN-28) + +# Results of Mitigating OOD Overconfidence + +Comparison with State-of-the-art Methods. To assess the effectiveness and superiority of our approach, we conduct comparative experiments evaluating OOD overconfi + +dence for ResNet-50 and ResNet-101 trained on CIFAR-10 and CIFAR-100. The results in Tab. 3 demonstrate that our approach consistently reduces the mean maximum confidence (MMC) across different student networks in all cases, validating its effectiveness. In comparison, methods such + +Table 4. Comparison of OOD detection performance for various methods on the CIFAR-10 and CIFAR-100 datasets. Underlined values indicate results enhanced by applying our approach to the baseline, while bolded values denote the best performance. + +
IDModelMethodOODAverage
SVHNLSUN-RiSUNTexturesTINR
FPR95↓AUROC↑FPR95↓AUROC↑FPR95↓AUROC↑FPR95↓AUROC↑FPR95↓AUROC↑FPR95↓AUROC↑
CIFAR-10ResNet-50MSP43.7594.1954.3291.9655.4691.6659.0190.3158.6190.8054.2391.78
MSP + Ours (ResNet-18)51.8293.2741.9393.8944.8693.2755.6490.5750.1891.9748.8992.59
MSP + Ours (WRN-16)28.1896.1235.7495.0139.7894.3457.0089.7942.3593.7540.6193.80
MaxLogit22.0796.3132.0394.9534.9594.5346.3391.3944.9692.5736.0793.95
MaxLogit + Ours (ResNet-18)30.2594.9425.0695.3427.9194.7344.5091.7136.7592.7332.8993.89
MaxLogit + Ours (WRN-16)17.6196.9618.5496.9221.6996.3947.7089.5827.8195.5526.6795.08
Energy20.3596.4029.7995.1132.8094.6745.4191.4443.7692.6434.4294.05
Energy + Ours (ResNet-18)28.5395.0723.9995.4326.5594.8344.1389.7035.9292.7931.8293.56
Energy + Ours (WRN-16)18.8096.8818.0797.0220.9196.5047.8989.5427.6395.6126.6695.11
ReAct37.6691.1016.2496.6419.6595.8446.7286.2930.7993.4630.2192.67
ReAct + Ours (ResNet-18)31.5994.6223.8795.5126.7894.8643.6090.1935.8792.9232.3493.62
ReAct + Ours (WRN-16)19.0796.8417.9097.0320.6696.5147.4589.8427.2495.6826.4695.18
FeatureNorm9.2597.9777.3482.0770.2983.9648.9784.1667.7184.4954.7186.53
FeatureNorm + Ours (ResNet-18)2.7399.3524.3495.7521.0496.2822.8294.8929.5294.3620.0996.13
FeatureNorm + Ours (WRN-16)3.3799.2622.0396.4321.5496.4753.6778.8529.3594.5025.9993.10
CIFAR-100ResNet-101MSP76.3081.1486.6769.7688.1169.0685.6973.7385.2771.4784.4173.03
MSP + Ours (ResNet-34)77.4878.9272.7483.0974.8382.1482.5077.2674.5282.2076.4180.72
MSP + Ours (WRN-28)74.8781.3962.5785.7166.9483.9383.2876.3369.4481.7071.4281.81
MaxLogit68.0586.7782.8474.3484.9473.4087.2574.8181.9275.0581.0076.87
MaxLogit + Ours (ResNet-34)71.6983.4763.7487.2665.6986.5481.7478.4666.8886.4469.9584.43
MaxLogit + Ours (WRN-28)73.8482.6154.3089.1160.2987.2683.6277.0664.1284.9867.2384.20
Energy68.3886.9482.3874.6484.2173.6788.0374.6281.4075.2580.8877.02
Energy + Ours (ResNet-34)69.2084.1458.8388.0561.1087.3781.6878.5162.8987.1866.7485.05
Energy + Ours (WRN-28)74.1782.6148.6090.0755.0588.1284.1077.0160.4185.7164.4784.70
ReAct65.6888.1486.4469.1886.9769.0983.4480.1383.8071.5981.2775.63
ReAct + Ours (ResNet-34)60.2688.3057.9486.4958.1186.5263.0785.7860.0385.6959.8886.56
ReAct + Ours (WRN-28)71.8986.7347.3489.2053.6687.4680.5781.4858.9784.6662.4985.91
FeatureNorm31.8894.5999.5434.1199.1039.0253.4282.3298.2242.3276.4358.47
FeatureNorm + Ours (ResNet-34)41.4790.4792.4671.4991.4172.8965.1880.3389.9969.0076.1076.84
FeatureNorm + Ours (WRN-28)54.1789.0880.8679.5280.6979.9171.4974.3778.0080.7473.0480.72
+ +as CEDA, ACET, and CCUs, which employ specific techniques to suppress confidence scores during training, also significantly mitigate OOD overconfidence in DNNs. However, our approach, which does not incorporate specific confidence suppression measures, still outperforms these state-of-the-art methods, particularly in the more challenging CIFAR-100 scenario. This further highlights the robustness and effectiveness of our strategy. + +On CIFAR-10, while our base method may not show a pronounced advantage over others, combining our approach with confidence suppression techniques provides additional improvements, surpassing the performance of these state-of-the-art methods. These results reinforce the effectiveness of our approach in addressing OOD overconfidence. + +Visualization. We visualize the maximum confidence scores predicted by various DNN classifiers as histograms (logarithmic scale) in Fig. 6. As shown, applying OOD overconfidence mitigation solutions shifts the confidence distributions of the Baseline to the left, substantially reducing the number of samples with high confidence scores. Notably, our approach achieves the most pronounced shift, un + +derscoring its effectiveness and validating its superiority in mitigating OOD overconfidence. + +# 5.3. Results of Enhancing OOD Detection + +CIFAR Experiments. We evaluate the effectiveness of our approach in enhancing OOD detection. The results in Tab. 4 demonstrate that applying our method to existing OOD detection approaches leads to significant performance improvements. Even when directly applied to the classifier without any advanced post-hoc processing, i.e., MSP, the performance enhancement is still noticeable. These improvements are consistent across different student model configurations, datasets, and classifiers, validating the effectiveness of our approach in boosting OOD detection. + +ImageNet Experiments. We further validate the effectiveness of our approach in enhancing OOD detection on the ImageNet-1k dataset. The results in Tab. 5 indicate that, similar to the CIFAR experiments, applying our approach significantly improves OOD detection performance across methods, demonstrating the robustness and scalability of our approach in large-scale settings. + +Table 5. Comparison of OOD detection performance for various methods on the ImageNet-1k dataset. + +
ModelMethodOODAverage
SUNPlacesTexturesOpenImage-O
FPR95↓AUROC↑FPR95↓AUROC↑FPR95↓AUROC↑FPR95↓AUROC↑FPR95↓AUROC↑
DenseNet-201MSP65.7682.2469.5381.1467.8279.3664.9284.4767.0181.80
MSP + Ours (DenseNet-161)66.3181.7268.6181.3266.9779.5562.1285.1066.0081.92
MaxLogit55.4486.0661.8983.9958.4883.4559.6486.9458.8685.11
MaxLogit + Ours (DenseNet-161)55.3387.6759.9586.2655.8085.8460.0388.6057.7887.09
Energy52.8786.0460.4683.7156.1383.3559.8786.6557.3384.94
Energy + Ours (DenseNet-161)50.8888.2556.9086.6052.2586.4558.7688.6054.7087.48
ReAct46.0090.8157.4686.1155.1484.1969.3274.4356.9883.89
ReAct + Ours (DenseNet-161)49.2188.5555.0286.9450.9986.7557.6188.6453.2187.72
FeatureNorm43.0690.0555.7185.4250.4679.9079.9767.2257.3080.65
FeatureNorm + Ours (DenseNet-161)36.5291.7949.2487.5056.7378.2982.2065.4956.1780.77
WRN-101MSP66.9382.4470.4881.3565.1480.9159.8786.1365.6182.71
MSP + Ours (WRN-50)67.7582.8069.0781.8166.3880.9660.9186.7866.0383.09
MaxLogit63.2584.7568.4082.6957.3984.3855.8488.4361.2285.06
MaxLogit + Ours (WRN-50)61.9286.1864.3484.4056.7485.6055.6989.2559.6786.36
Energy65.0684.4670.3282.2757.1184.3758.0488.0262.6384.78
Energy + Ours (WRN-50)63.1885.9666.0684.0055.9085.6857.8788.7760.7586.10
ReAct52.2288.1158.7685.4757.0984.9949.6387.8254.4386.60
ReAct + Ours (WRN-50)37.0292.2448.5488.7050.9087.3654.4985.9247.7488.56
FeatureNorm70.0778.8978.6173.2320.9894.7975.0974.8161.1980.43
FeatureNorm + Ours (WRN-50)60.2283.6774.0175.9617.7196.0969.9478.9055.4783.66
+ +Table 6. Comparison of MMC and OOD detection performance for ResNet-50 enhanced by Ours (ResNet-18) and its variants w/o knowledge distillation (KD) and ReLU removal (RR), using CIFAR-10 as the ID dataset, and SVHN, LSUN-R, iSUN, Textures, and TINR as the OOD datasets. Results are averaged. + +
KDRRMMC↓FPR95↓AUROC↑
79.5854.2391.78
✓76.8149.8192.42
✓78.4650.5091.76
✓✓76.4848.8992.59
+ +# 5.4. Ablation Studies and More Analysis + +Importance of Knowledge Distillation and ReLU Removal. We conduct comprehensive ablation studies to evaluate the impact of knowledge distillation and ReLU removal, the two essential components of our approach. As shown in Tab. 6, excluding either component notably diminishes the effectiveness of our method in mitigating OOD overconfidence and complicates the accuracy of OOD detection. These findings validate the critical importance of both knowledge distillation and selective ReLU removal in achieving optimal performance with our approach. + +Knowledge Distillation vs. Network Pruning for Capacity Reduction. We further investigate an alternative capacity reduction method for our approach by applying network pruning. To facilitate a fair comparison with knowledge distillation, we evaluate ResNet-50 and ResNet-101 models trained on CIFAR-10, applying two strategies: (1) using KD [22] to reduce ResNet-50 to ResNet-18 and ResNet-101 to WRN-28, and (2) applying Prune [14] to ResNet-50 and ResNet-101, reducing them to parameter counts comparable to ResNet-18 and WRN-28, respectively. The re + +Table 7. Comparison of MMC and OOD detection performance for ResNet-50 and ResNet-101 reduced to comparable parameter counts via KD [22] and Prune [14]. CIFAR-10 serves as the ID dataset, with SVHN, LSUN-R, iSUN, Textures, and TINR as OOD datasets. Results are averaged across all five OOD datasets + +
ModelCapacity ReductionMMC↓FPR95↓AUROC↑
ResNet-50—79.5854.2391.78
Prune77.3452.5592.27
KD (ResNet-18)76.8149.8192.42
ResNet-101—83.4651.9292.53
Prune78.2050.9291.61
KD (WRN-28)77.0141.0992.56
+ +sults in Tab. 7 show that both methods effectively reduce OOD overconfidence and improve OOD detection; however, knowledge distillation consistently delivers a more substantial improvement than network pruning. Additionally, knowledge distillation offers greater stability and flexibility during training, making it a more suitable choice for capacity reduction in our approach. + +# 6. Conclusion and Future Work + +This paper has presented a comprehensive investigation into mitigating OOD overconfidence in DNNs through model simplification. By reducing model capacity via knowledge distillation and decreasing nonlinearity through selective ReLU removal, we demonstrate that OOD overconfidence in DNN classifiers can be significantly mitigated. Furthermore, these strategies can be effectively combined with OOD detection methods to enhance overall performance. In future work, we plan to study the effects of model simplification on OOD overconfidence in greater depth and from a theoretical perspective. + +# Acknowledgements + +We sincerely thank the anonymous reviewers for their valuable comments. This work was supported in part by the National Natural Science Foundation of China (62472117, U2436208, 62406095), the Guangdong Basic and Applied Basic Research Foundation (2025A1515010157, 2024A1515012064), the Science and Technology Projects in Guangzhou (2025A03J0137, SL2022A04J01112), and the Academician Bixing Fang's Specialized Class. + +# References + +[1] Lei Jimmy Ba and Rich Caruana. Do deep nets really need to be deep? In NeurIPS, pages 2654-2662, 2014. 2 +[2] Abhijit Bendale and Terrance Boult. Towards open world recognition. In CVPR, pages 1893-1902, 2015. 1 +[3] Chao Chen, Zhihang Fu, Kai Liu, Ze Chen, Mingyuan Tao, and Jieping Ye. Optimal parameter and neuron pruning for out-of-distribution detection. In NeurIPS, 2023. 2 +[4] Hanting Chen, Yunhe Wang, Jianyuan Guo, and Dacheng Tao. Vanillanet: the power of minimalism in deep learning. In NeurIPS, 2024. 3, 5 +[5] Mircea Cimpoi, Subhransu Maji, Iasonas Kokkinos, Sammy Mohamed, and Andrea Vedaldi. Describing textures in the wild. In CVPR, pages 3606-3613, 2014. 5 +[6] Jesse Davis and Mark Goadrich. The relationship between precision-recall and roc curves. In ICML, pages 233-240, 2006. 3 +[7] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, pages 248-255, 2009. 5 +[8] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 1 +[9] Xuefeng Du, Zhaoning Wang, Mu Cai, and Yixuan Li. Vos: Learning what you don't know by virtual outlier synthesis. In ICLR, 2021. 2 +[10] Shiv Ram Dubey, Satish Kumar Singh, and Bidyut Baran Chaudhuri. Activation functions in deep learning: A comprehensive survey and benchmark. Neurocomputing, 503: 92-108, 2022. 2 +[11] Xiang Fang, Arvind Easwaran, Blaise Genest, and Ponnuthurai Nagaratnam Suganthan. Your data is not perfect: Towards cross-domain out-of-distribution detection in class-imbalanced data. ESWA, 2024. 2 +[12] Xavier Glorot, Antoine Bordes, and Yoshua Bengio. Deep sparse rectifier neural networks. In AISTATS, pages 315-323, 2011. 2 +[13] Ian J. Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. In ICLR, 2015. 1 +[14] Song Han, Jeff Pool, John Tran, and William J Dally. Learning both weights and connections for efficient neural networks. In NeurIPS, pages 1135-1143, 2015. 8 + +[15] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Delving deep into rectifiers: Surpassing human-level performance on imagenet classification. In ICCV, pages 1026-1034, 2015. 1 +[16] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 5 +[17] Xu He, Keke Tang, Yawen Shi, Yin Li, Weilong Peng, and Peican Zhu. Are deep point cloud classifiers suffer from out-of-distribution overconfidence issue? In SMC, pages 2620-2627. IEEE, 2023. 2 +[18] Matthias Hein, Maksym Andriushchenko, and Julian Bitterwolf. Why relu networks yield high-confidence predictions far away from the training data and how to mitigate the problem. In CVPR, pages 41-50, 2019. 1, 2, 5 +[19] Dan Hendrycks and Kevin Gimpel. A baseline for detecting misclassified and out-of-distribution examples in neural networks. In ICLR, 2016. 2, 5 +[20] Dan Hendrycks, Mantas Mazeika, and Thomas Dietterich. Deep anomaly detection with outlier exposure. In ICLR, 2019. 1, 2 +[21] Dan Hendrycks, Steven Basart, Mantas Mazeika, Andy Zou, Joseph Kwon, Mohammadreza Mostajabi, Jacob Steinhardt, and Dawn Song. Scaling out-of-distribution detection for real-world settings. In ICML, pages 8759-8773, 2022. 5 +[22] Geoffrey Hinton. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. 2, 3, 5, 8 +[23] Gao Huang, Zhuang Liu, Laurens Van Der Maaten, and Kilian Q Weinberger. Densely connected convolutional networks. In CVPR, pages 4700-4708, 2017. 5 +[24] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 5 +[25] Yann LeCun, Yoshua Bengio, and Geoffrey Hinton. Deep learning. Nature, 521(7553):436-444, 2015. 1 +[26] Kimin Lee, Honglak Lee, Kibok Lee, and Jinwoo Shin. Training confidence-calibrated classifiers for detecting out-of-distribution samples. In ICLR, 2018. 1, 2 +[27] Kimin Lee, Kibok Lee, Honglak Lee, and Jinwoo Shin. A simple unified framework for detecting out-of-distribution samples and adversarial attacks. In NeurIPS, pages 7167-7177, 2018. 2 +[28] Shiyu Liang, Yixuan Li, and R Srikant. Enhancing the reliability of out-of-distribution image detection in neural networks. In ICLR, 2018. 2, 3 +[29] Weitang Liu, Xiaoyun Wang, John D Owens, and Yixuan Li. Energy-based out-of-distribution detection. In NeurIPS, pages 21464-21475, 2020. 2, 5 +[30] Alexander Meinke and Matthias Hein. Towards neural networks that provably know when they don't know. In ICLR, 2020. 5 +[31] Melanie Mitchell. Complexity: A guided tour. Oxford University Press, 2009. 2 +[32] Rafael Müller, Simon Kornblith, and Geoffrey Hinton. When does label smoothing help? In NeurIPS, pages 4694-4703, 2019. 2 +[33] Yuval Netzer, Tao Wang, Adam Coates, Alessandro Bissacco, Bo Wu, and Andrew Y Ng. Reading digits in natural images with unsupervised feature learning. In NeurIPS + +Workshop on Deep Learning and Unsupervised Feature Learning, 2011. 5 +[34] Anh Nguyen, Jason Yosinski, and Jeff Clune. Deep neural networks are easily fooled: High confidence predictions for unrecognizable images. In CVPR, pages 427-436, 2015. 1 +[35] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Köpf, Edward Yang, Zach DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. Pytorch: An imperative style, high-performance deep learning library. In NeurIPS, pages 8026-8037, 2019. 5 +[36] Jiacheng Qiao, Chengzhi Zhong, Peican Zhu, and Keke Tang. Advscod: Bayesian-based out-of-distribution detection via curvature sketching and adversarial sample enrichment. Mathematics, 11(3):692, 2023. 2 +[37] Adriana Romero, Nicolas Ballas, Samira Ebrahimi Kahou, Antoine Chassang, Carlo Gatta, and Yoshua Bengio. Fitnets: Hints for thin deep nets. arXiv preprint arXiv:1412.6550, 2014. 2, 3 +[38] K Simonyan and A Zisserman. Very deep convolutional networks for large-scale image recognition. In ICLR, 2015. 3 +[39] Yiyou Sun, Chuan Guo, and Yixuan Li. React: Out-of-distribution detection with rectified activations. In NeurIPS, 2021. 2, 5 +[40] Yiyou Sun, Yifei Ming, Xiaojin Zhu, and Yixuan Li. Out-of-distribution detection with deep nearest neighbors. In ICML, 2022. 2 +[41] Keke Tang, Dingruibo Miao, Weilong Peng, Jianpeng Wu, Yawen Shi, Zhaoquan Gu, Zhihong Tian, and Wenping Wang. Codes: Chamfer out-of-distribution examples against overconfidence issue. In ICCV, pages 1153-1162, 2021. 2 +[42] Keke Tang, Xujian Cai, Weilong Peng, Shudong Li, and Wenping Wang. Ood attack: Generating overconfident out-of-distribution examples to fool deep neural classifiers. In ICIP, pages 1260-1264, 2023. 2 +[43] Keke Tang, Xujian Cai, Weilong Peng, Daizong Liu, Peican Zhu, Pan Zhou, Zhihong Tian, and Wenping Wang. Matching words for out-of-distribution detection. In ICDM, pages 578-587, 2023. 2 +[44] Keke Tang, Chao Hou, Weilong Peng, Runnan Chen, Peican Zhu, Wenping Wang, and Zhihong Tian. Cores: Convolutional response-based score for out-of-distribution detection. In CVPR, pages 10916-10925, 2024. 2 +[45] Keke Tang, Wenyu Zhao, Weilong Peng, Xiang Fang, Xiaodong Cui, Peican Zhu, and Zhihong Tian. Reparaterization head for efficient multi-input networks. In ICASSP, pages 6190-6194, 2024. 1 +[46] Keke Tang, Yuexin Ma, Dingruibo Miao, Peng Song, Zhaoquan Gu, Zhihong Tian, and Wenping Wang. Decision fusion networks for image classification. IEEE TNNLS, 36(3): 3890-3903, 2025. 1 +[47] Leitian Tao, Xuefeng Du, Jerry Zhu, and Yixuan Li. Nonparametric outlier synthesis. In ICLR, 2022. 2 +[48] Frederick Tung and Greg Mori. Similarity-preserving knowledge distillation. In ICCV, pages 1365-1374, 2019. 3 + +[49] Haoqi Wang, Zhizhong Li, Litong Feng, and Wayne Zhang. Vim: Out-of-distribution with virtual-logit matching. In CVPR, pages 4921-4930, 2022. 5 +[50] Xiaobo Wang, Tianyu Fu, Shengcai Liao, Shuo Wang, Zhen Lei, and Tao Mei. Exclusivity-consistency regularized knowledge distillation for face recognition. In ECCV, pages 325-342, 2020. 2 +[51] Zhi-Fan Wu, Tong Wei, Jianwen Jiang, Chaojie Mao, Mingqian Tang, and Yu-Feng Li. Ngc: A unified framework for learning with open-world noisy data. In ICCV, pages 62-71, 2021. 2 +[52] Jianxiong Xiao, James Hays, Krista A Ehinger, Aude Oliva, and Antonio Torralba. Sun database: Large-scale scene recognition from abbey to zoo. In CVPR, pages 3485-3492, 2010. 5 +[53] Kunran Xu, Lai Rui, Yishi Li, and Lin Gu. Feature normalized knowledge distillation for image classification. In ECCV, pages 664-680, 2020. 2 +[54] Pingmei Xu, Krista A Ehinger, Yinda Zhang, Adam Finkelstein, Sanjeev R Kulkarni, and Jianxiong Xiao. Turkergaze: Crowdsourcing saliency with webcam based eye tracking. arXiv preprint arXiv:1504.06755, 2015. 5 +[55] Jingkang Yang, Haoqi Wang, Litong Feng, Xiaopeng Yan, Huabin Zheng, Wayne Zhang, and Ziwei Liu. Semantically coherent out-of-distribution detection. In ICCV, pages 8301-8309, 2021. 2 +[56] Jingkang Yang, Kaiyang Zhou, Yixuan Li, and Ziwei Liu. Generalized out-of-distribution detection: A survey. arXiv preprint arXiv:2110.11334, 2021. 2 +[57] Junho Yim, Donggyu Joo, Jihoon Bae, and Junmo Kim. A gift from knowledge distillation: Fast optimization, network minimization and transfer learning. In CVPR, pages 4133-4141, 2017. 2 +[58] Shan You, Chang Xu, Chao Xu, and Dacheng Tao. Learning from multiple teacher networks. In SIGKDD, pages 1285-1294, 2017. 2 +[59] Fisher Yu, Ari Seff, Yinda Zhang, Shuran Song, Thomas Funkhouser, and Jianxiong Xiao. Lsun: Construction of a large-scale image dataset using deep learning with humans in the loop. arXiv preprint arXiv:1506.03365, 2015. 5 +[60] Qing Yu and Kiyoharu Aizawa. Unsupervised out-of-distribution detection by maximum classifier discrepancy. In ICCV, pages 9518-9526, 2019. 2 +[61] Yeonguk Yu, Sungho Shin, Seongju Lee, Changyun Jun, and Kyoobin Lee. Block selection method for using feature norm in out-of-distribution detection. In CVPR, pages 15701-15711, 2023. 5 +[62] Yue Yuan, Rundong He, Zhongyi Han, and Yilong Yin. Lhact: Rectifying extremely low and high activations for out-of-distribution detection. In MM, pages 8105-8113, 2023. 2 +[63] Sergey Zagoruyko and Nikos Komodakis. Wide residual networks. In BMVC, 2016. 5 +[64] Sergey Zagoruyko and Nikos Komodakis. Paying more attention to attention: Improving the performance of convolutional neural networks via attention transfer. In ICLR, 2022. 3 + +[65] Bolei Zhou, Agata Lapedriza, Aditya Khosla, Aude Oliva, and Antonio Torralba. Places: A 10 million image database for scene recognition. TPAMI, 40(6):1452-1464, 2017. 5 +[66] Zhi-Hua Zhou. Why over-parameterization of deep neural networks does not overfit. Science China Information Sciences, 64(1):1-3, 2021. 2 +[67] Yao Zhu, YueFeng Chen, Chuanlong Xie, Xiaodan Li, Rong Zhang, Xiang Tian, Yaowu Chen, et al. Boosting out-of-distribution detection with typical features. In NeurIPS, 2022. 2 \ No newline at end of file diff --git a/2025/Simplification Is All You Need against Out-of-Distribution Overconfidence/images.zip b/2025/Simplification Is All You Need against Out-of-Distribution Overconfidence/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..a910cbbd82a59dcde59b6d57e0f324334551b939 --- /dev/null +++ b/2025/Simplification Is All You Need against Out-of-Distribution Overconfidence/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fdf3386738ba5bc3be647a5fcbb5f38df0fd97e1dfbbb84edeb6629492b4c9f6 +size 1030796 diff --git a/2025/Simplification Is All You Need against Out-of-Distribution Overconfidence/layout.json b/2025/Simplification Is All You Need against Out-of-Distribution Overconfidence/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..ceeaca0f1013c654ff8a997d3743e9289944513b --- /dev/null +++ b/2025/Simplification Is All You Need against Out-of-Distribution Overconfidence/layout.json @@ -0,0 +1,8264 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 76, + 102, + 536, + 122 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 102, + 536, + 122 + ], + "spans": [ + { + "bbox": [ + 76, + 102, + 536, + 122 + ], + "type": "text", + "content": "Simplification Is All You Need against Out-of-Distribution Overconfidence" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 160, + 152, + 449, + 182 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 152, + 449, + 182 + ], + "spans": [ + { + "bbox": [ + 160, + 152, + 449, + 182 + ], + "type": "text", + "content": "Keke Tang" + }, + { + "bbox": [ + 160, + 152, + 449, + 182 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 160, + 152, + 449, + 182 + ], + "type": "text", + "content": ", Chao Hou" + }, + { + "bbox": [ + 160, + 152, + 449, + 182 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 160, + 152, + 449, + 182 + ], + "type": "text", + "content": ", Weilong Peng" + }, + { + "bbox": [ + 160, + 152, + 449, + 182 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 160, + 152, + 449, + 182 + ], + "type": "text", + "content": ", Xiang Fang" + }, + { + "bbox": [ + 160, + 152, + 449, + 182 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 160, + 152, + 449, + 182 + ], + "type": "text", + "content": ", Zhize Wu" + }, + { + "bbox": [ + 160, + 152, + 449, + 182 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 160, + 152, + 449, + 182 + ], + "type": "text", + "content": ", Yongwei Nie" + }, + { + "bbox": [ + 160, + 152, + 449, + 182 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 160, + 152, + 449, + 182 + ], + "type": "text", + "content": ", Wenping Wang" + }, + { + "bbox": [ + 160, + 152, + 449, + 182 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 160, + 152, + 449, + 182 + ], + "type": "text", + "content": ", Zhihong Tian" + }, + { + "bbox": [ + 160, + 152, + 449, + 182 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 96, + 190, + 515, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 190, + 515, + 220 + ], + "spans": [ + { + "bbox": [ + 96, + 190, + 515, + 220 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 96, + 190, + 515, + 220 + ], + "type": "text", + "content": "Guangzhou University " + }, + { + "bbox": [ + 96, + 190, + 515, + 220 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 96, + 190, + 515, + 220 + ], + "type": "text", + "content": "Nanyang Technological University " + }, + { + "bbox": [ + 96, + 190, + 515, + 220 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 96, + 190, + 515, + 220 + ], + "type": "text", + "content": "Hefei University " + }, + { + "bbox": [ + 96, + 190, + 515, + 220 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 96, + 190, + 515, + 220 + ], + "type": "text", + "content": "South China University of Technology " + }, + { + "bbox": [ + 96, + 190, + 515, + 220 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 96, + 190, + 515, + 220 + ], + "type": "text", + "content": "Texas A&M University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 79, + 221, + 531, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 221, + 531, + 247 + ], + "spans": [ + { + "bbox": [ + 79, + 221, + 531, + 247 + ], + "type": "text", + "content": "tangbohutbh@gmail.com, houchaohk@gmail.com, wlpeng@tju.edu.cn, xfang9508@gmail.com, wuzz@hfuu.edu.cn, nieyongwei@scut.edu.cn, wenping@amu.edu, tianzhihong@gzhu.edu.cn" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 151, + 281, + 201, + 295 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 281, + 201, + 295 + ], + "spans": [ + { + "bbox": [ + 151, + 281, + 201, + 295 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 307, + 297, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 307, + 297, + 510 + ], + "spans": [ + { + "bbox": [ + 54, + 307, + 297, + 510 + ], + "type": "text", + "content": "Deep neural networks (DNNs) often exhibit out-of-distribution (OOD) overconfidence, producing overly confident predictions on OOD samples. We attribute this issue to the inherent over-complexity of DNNs and investigate two key aspects: capacity and nonlinearity. First, we demonstrate that reducing model capacity through knowledge distillation can effectively mitigate OOD overconfidence. Second, we show that selectively reducing nonlinearity by removing ReLU operations further alleviates the issue. Building on these findings, we present a practical guide to model simplification, combining both strategies to significantly reduce OOD overconfidence. Extensive experiments validate the effectiveness of this approach in mitigating OOD overconfidence and demonstrate its superiority over state-of-the-art methods. Additionally, our simplification strategies can be combined with existing OOD detection techniques to further enhance OOD detection performance." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 532, + 137, + 544 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 532, + 137, + 544 + ], + "spans": [ + { + "bbox": [ + 55, + 532, + 137, + 544 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 552, + 296, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 552, + 296, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 552, + 296, + 696 + ], + "type": "text", + "content": "Deep neural networks (DNNs) have revolutionized image classification tasks, consistently achieving state-of-the-art performance [8, 15, 25, 45, 46]. Despite this success, DNN classifiers are inherently designed for static, closed-world environments [2], limiting their ability to adapt to novel, unforeseen concepts encountered in real-world scenarios, known as out-of-distribution (OOD) samples. Indeed, many recent studies demonstrate that DNN classifiers suffer from OOD overconfidence, generating overly confident predictions on OOD samples [13, 34]. This issue significantly impacts the reliability and safety of DNNs in real-world applications, highlighting an urgent need for effective solutions." + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 322, + 281, + 432, + 354 + ], + "blocks": [ + { + "bbox": [ + 322, + 281, + 432, + 354 + ], + "lines": [ + { + "bbox": [ + 322, + 281, + 432, + 354 + ], + "spans": [ + { + "bbox": [ + 322, + 281, + 432, + 354 + ], + "type": "image", + "image_path": "5e6a4a1f5e3b5d07fabd2929ba9082d24079e99165782dac9f651cebbe002a7e.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 433, + 281, + 544, + 354 + ], + "blocks": [ + { + "bbox": [ + 433, + 281, + 544, + 354 + ], + "lines": [ + { + "bbox": [ + 433, + 281, + 544, + 354 + ], + "spans": [ + { + "bbox": [ + 433, + 281, + 544, + 354 + ], + "type": "image", + "image_path": "dce8fa53a43848f5d9336be0b0e3c2b2f9d31368c92f5d54078bbf5d6f2c43ca.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 321, + 355, + 433, + 428 + ], + "blocks": [ + { + "bbox": [ + 321, + 355, + 433, + 428 + ], + "lines": [ + { + "bbox": [ + 321, + 355, + 433, + 428 + ], + "spans": [ + { + "bbox": [ + 321, + 355, + 433, + 428 + ], + "type": "image", + "image_path": "916a7cfce59a808608cd0e609f3941a41b1aa0e793d6375a2c3ee3c3bae2dec2.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 434, + 355, + 544, + 441 + ], + "blocks": [ + { + "bbox": [ + 434, + 355, + 544, + 441 + ], + "lines": [ + { + "bbox": [ + 434, + 355, + 544, + 441 + ], + "spans": [ + { + "bbox": [ + 434, + 355, + 544, + 441 + ], + "type": "image", + "image_path": "5f9ef8a82b0e767c3e2a2dca39b494acf7a0e0ac697dcbde74f8bfb892ec4ef0.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 345, + 430, + 542, + 441 + ], + "blocks": [ + { + "bbox": [ + 345, + 430, + 542, + 441 + ], + "lines": [ + { + "bbox": [ + 345, + 430, + 542, + 441 + ], + "spans": [ + { + "bbox": [ + 345, + 430, + 542, + 441 + ], + "type": "image", + "image_path": "ff7334fac2b18610984438b898d30b9be4fc31f3283c45235f6b64b71e572110.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 445, + 555, + 522 + ], + "lines": [ + { + "bbox": [ + 313, + 445, + 555, + 522 + ], + "spans": [ + { + "bbox": [ + 313, + 445, + 555, + 522 + ], + "type": "text", + "content": "Figure 1. Histograms (logarithmic scale) of maximum confidence values for the teacher network (ResNet-50), the distilled student network (ResNet-18), and its variant with additional ReLU removal. All models are trained on CIFAR-100 and evaluated on OOD datasets: (a) SVHN, (b) LSUN-R, (c) iSUN, and (d) TinyImageNet-R. The " + }, + { + "bbox": [ + 313, + 445, + 555, + 522 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 313, + 445, + 555, + 522 + ], + "type": "text", + "content": "-axis represents the frequency of samples, and the " + }, + { + "bbox": [ + 313, + 445, + 555, + 522 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 445, + 555, + 522 + ], + "type": "text", + "content": "-axis represents confidence scores." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 546, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 546, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 546, + 556, + 715 + ], + "type": "text", + "content": "Extensive research has been dedicated to addressing the OOD overconfidence issue. One common approach involves calibrating outputs for both in-distribution (ID) and OOD samples to improve discrimination [20], enabling OOD detectors to proactively reject OOD inputs. While these methods can reduce the impact of overconfidence, they do not fundamentally resolve the issue within DNN classifiers. Moreover, their effectiveness heavily depends on the classifiers' ability to adequately separate predictive distributions; they fail when the classifiers cannot achieve this separation [26]. As a more intrinsic solution, another approach trains DNN classifiers to suppress predictions on a subset of typical OOD samples, with the expectation that this behavior will generalize to other OOD data [18, 20]." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 70, + 702, + 215, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 702, + 215, + 713 + ], + "spans": [ + { + "bbox": [ + 70, + 702, + 215, + 713 + ], + "type": "text", + "content": "*Joint first authors. †Corresponding author." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "5030" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 95 + ], + "type": "text", + "content": "However, given the vast diversity of potential OOD samples, the effectiveness of this method remains limited." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 96, + 294, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 96, + 294, + 227 + ], + "spans": [ + { + "bbox": [ + 54, + 96, + 294, + 227 + ], + "type": "text", + "content": "DNNs can indeed be regarded as complex systems, often characterized by over-parameterization, in which the number of parameters exceeds the amount of available training data [66]. This excessive complexity can cause DNNs to exhibit emergent behaviors, leading to unpredictable outcomes, as observed in other complex systems [31]. Consequently, when presented with certain OOD inputs, DNNs may unexpectedly but confidently misclassify them as ID classes. Therefore, investigating the relationship between model complexity and OOD overconfidence presents a promising path toward effectively addressing this issue." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 228, + 296, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 228, + 296, + 430 + ], + "spans": [ + { + "bbox": [ + 56, + 228, + 296, + 430 + ], + "type": "text", + "content": "In this paper, we explore two key aspects of model complexity: capacity and nonlinearity, and their roles in influencing OOD overconfidence. For capacity, we use knowledge distillation to examine how varying the student network's capacity affects OOD overconfidence, finding that an appropriately reduced capacity can significantly alleviate this issue (see Fig. 1). For nonlinearity, we investigate selective ReLU removal at different ratios and locations within the network, observing that controlled reductions in nonlinearity similarly mitigate OOD overconfidence (see Fig. 1). Based on these insights, we present a practical guide to model simplification that combines these two strategies to mitigate the OOD overconfidence issue in DNNs more effectively, achieving superior results compared to state-of-the-art methods. Furthermore, we demonstrate that our approach can be effectively combined with existing OOD detection techniques to further enhance performance." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 431, + 276, + 442 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 431, + 276, + 442 + ], + "spans": [ + { + "bbox": [ + 67, + 431, + 276, + 442 + ], + "type": "text", + "content": "Overall, our contribution is summarized as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 445, + 295, + 564 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 55, + 445, + 295, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 445, + 295, + 479 + ], + "spans": [ + { + "bbox": [ + 55, + 445, + 295, + 479 + ], + "type": "text", + "content": "- We are the first to attribute the OOD overconfidence issue in DNNs to their over-complexity, identifying model simplification as a crucial and effective solution." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 480, + 295, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 480, + 295, + 528 + ], + "spans": [ + { + "bbox": [ + 55, + 480, + 295, + 528 + ], + "type": "text", + "content": "- We systematically investigate two fundamental aspects of complexity: capacity and nonlinearity, analyzing their effects on OOD overconfidence through knowledge distillation and selective ReLU removal." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 529, + 295, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 529, + 295, + 564 + ], + "spans": [ + { + "bbox": [ + 55, + 529, + 295, + 564 + ], + "type": "text", + "content": "- We present a practical guide for model simplification and demonstrate through extensive experiments its effectiveness in significantly mitigating OOD overconfidence." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 574, + 142, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 574, + 142, + 586 + ], + "spans": [ + { + "bbox": [ + 55, + 574, + 142, + 586 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "text", + "content": "Mitigating OOD Overconfidence. The issue of OOD overconfidence, where DNNs assign excessively high confidence scores to OOD samples, presents a substantial security risk for real-world applications [18, 42, 56]. Current approaches often incorporate auxiliary data to improve OOD sensitivity, using techniques such as adding random noise and image shuffling [26], leveraging large volumes of unlabeled data [51, 55, 60], relying on external OOD datasets [20], or generating synthetic OOD samples via models like Chamfer GAN [41]. Additionally, some" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 553, + 119 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 119 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 119 + ], + "type": "text", + "content": "methods, including VOS [9] and NPOS [47], create virtual features in feature space rather than explicit OOD instances. However, these approaches typically require substantial volumes of OOD data." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 121, + 553, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 121, + 553, + 191 + ], + "spans": [ + { + "bbox": [ + 313, + 121, + 553, + 191 + ], + "type": "text", + "content": "In contrast, our approach addresses OOD overconfidence by reducing model complexity, thus eliminating the need for auxiliary OOD data. Through extensive experiments, we systematically explore the effects of capacity reduction and selective activation removal, presenting a data-independent alternative for mitigating OOD overconfidence." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 192, + 553, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 192, + 553, + 407 + ], + "spans": [ + { + "bbox": [ + 313, + 192, + 553, + 407 + ], + "type": "text", + "content": "OD Detection. Another approach to managing OOD overconfidence is to develop scoring or normalization techniques that increase the distinction between ID and OOD samples [11, 17, 36, 43]. A foundational method, maximum softmax probability, introduced by Hendrycks and Gimpel [19], demonstrates that classifiers typically assign lower confidence to OOD samples. Follow-up methods, including ODIN [28] with temperature scaling and input perturbation, Energy Score [29], and density-based scores [27, 40], further enhance this ID-OOD separation. To improve robustness, several techniques refine feature representations: feature activation rectification [39, 62], feature selection [67], neuron pruning [3], and methods like CORES [44], which identify layer-specific relevant kernels for OOD detection. In contrast, our approach inherently addresses OOD overconfidence through model simplification. By tackling the root of the overconfidence issue, our method naturally enhances OOD detection performance as a result." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 407, + 553, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 407, + 553, + 514 + ], + "spans": [ + { + "bbox": [ + 313, + 407, + 553, + 514 + ], + "type": "text", + "content": "Knowledge Distillation. Knowledge distillation transfers knowledge from a larger teacher model to a smaller student model, enabling the student to approximate the teacher's performance [22]. It typically utilizes response-based knowledge (soft logits) [1, 22, 32], feature-based knowledge (intermediate feature maps) [37, 50, 53], and relation-based knowledge (layer/sample relationships) [57, 58]. We use distillation to reduce model capacity and mitigate OOD overconfidence." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 514, + 553, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 514, + 553, + 586 + ], + "spans": [ + { + "bbox": [ + 313, + 514, + 553, + 586 + ], + "type": "text", + "content": "ReLU Activation. ReLU " + }, + { + "bbox": [ + 313, + 514, + 553, + 586 + ], + "type": "inline_equation", + "content": "(f(x) = \\max(0, x))" + }, + { + "bbox": [ + 313, + 514, + 553, + 586 + ], + "type": "text", + "content": " introduces nonlinearity while maintaining efficiency, making it widely used in deep learning [10, 12]. However, ReLU networks are prone to OOD overconfidence [18]. We address this by selectively removing ReLU activations to mitigate OOD overconfidence." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 597, + 541, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 597, + 541, + 611 + ], + "spans": [ + { + "bbox": [ + 313, + 597, + 541, + 611 + ], + "type": "text", + "content": "3. How Complexity Affects Overconfidence?" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 617, + 553, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 553, + 689 + ], + "type": "text", + "content": "The complexity of a neural network is influenced by various factors. To investigate whether model complexity affects the degree of OOD overconfidence in DNNs, we focus on two key aspects: capacity and nonlinearity. In this section, we examine each of these factors individually to better understand their impact on OOD overconfidence." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 689, + 553, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 553, + 712 + ], + "type": "text", + "content": "Evaluation Metrics. To assess the severity of OOD overconfidence across different DNNs, we use the mean max" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "5031" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 118, + 293, + 191 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 297, + 115 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 297, + 115 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 297, + 115 + ], + "type": "text", + "content": "Table 1. Comparison of MMC and OOD detection performance for student models (ResNet-18) distilled from a teacher model (ResNet-50) using different knowledge distillation methods on CIFAR datasets. Results are averaged across multiple OOD datasets." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 118, + 293, + 191 + ], + "lines": [ + { + "bbox": [ + 58, + 118, + 293, + 191 + ], + "spans": [ + { + "bbox": [ + 58, + 118, + 293, + 191 + ], + "type": "table", + "html": "
IDCIFAR-10CIFAR-100
MethodMMC↓FPR95↓AUROC↑MMC↓FPR95↓AUROC↑
Teacher79.5854.2391.7872.1283.6774.73
KD [22]76.8149.8192.4262.6979.7776.31
FitNet [37]79.4650.5892.2660.6077.5377.78
AT [64]77.9048.0992.6964.1182.1475.51
SP [48]78.8150.5292.2258.7577.1280.01
", + "image_path": "75ec9c1ad56e1ef40ee44c7598cc1dadae5d9442f505313699ce8ac33cf9ef87.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 206, + 296, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 206, + 296, + 304 + ], + "spans": [ + { + "bbox": [ + 55, + 206, + 296, + 304 + ], + "type": "text", + "content": "imum confidence (MMC) on OOD samples as our evaluation metric. Additionally, to analyze how OOD overconfidence impacts the model's ability for ID/OOD discrimination, we also report the area under the receiver operating characteristic curve (AUROC) [6], the false positive rate at a true positive rate of " + }, + { + "bbox": [ + 55, + 206, + 296, + 304 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 55, + 206, + 296, + 304 + ], + "type": "text", + "content": " (FPR95) [28], and the area under the precision-recall curve, with ID samples specified as positives (AUPRin) [38]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 311, + 121, + 324 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 311, + 121, + 324 + ], + "spans": [ + { + "bbox": [ + 55, + 311, + 121, + 324 + ], + "type": "text", + "content": "3.1. Capacity" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 329, + 296, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 329, + 296, + 449 + ], + "spans": [ + { + "bbox": [ + 55, + 329, + 296, + 449 + ], + "type": "text", + "content": "To investigate whether model capacity affects the severity of OOD overconfidence, we employ knowledge distillation to transfer knowledge from a large-capacity teacher network to a smaller-capacity student network. Specifically, we examine how different knowledge distillation strategies and varying student network capacities influence OOD overconfidence. Since DNN capacity is influenced by various factors, such as depth, width, and number of parameters, we ensure a fair comparison by testing networks with the same architecture but different depths." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 450, + 296, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 450, + 296, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 450, + 296, + 605 + ], + "type": "text", + "content": "Effectiveness of Knowledge Distillation. To validate the effectiveness of knowledge distillation in mitigating OOD overconfidence, we select ResNet-50 as the teacher network and ResNet-18 as the student network, applying four distinct knowledge distillation methods: KD [22], FitNet [37], AT [64], and SP [48]. Experiments are conducted using CIFAR-10 and CIFAR-100 as ID datasets, with SVHN, LSUN-R, iSUN, Textures, and TinyImageNet-R as OOD datasets. The results, shown in Tab. 1, indicate that all four distillation methods lead to reductions in MMC and enhancements in AUROC and FPR95 metrics, demonstrating that knowledge distillation effectively mitigates the OOD overconfidence issue." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "text", + "content": "Analysis of Student Networks with Varying Capacities. We conduct experiments on two datasets using KD [22] with student networks of varying capacities. For CIFAR-100, we set ResNet-101 as the teacher and evaluate student networks with capacities ranging from ResNet-50, ResNet-34, ResNet-18, to ResNet-10. For ImageNet-100, we use ResNet-152 as the teacher, with ResNet-101, ResNet-50, ResNet-34, and ResNet-18 as student models. As shown in Fig. 2, on CIFAR-100, reducing the student" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 320, + 107, + 548, + 200 + ], + "blocks": [ + { + "bbox": [ + 313, + 70, + 555, + 105 + ], + "lines": [ + { + "bbox": [ + 313, + 70, + 555, + 105 + ], + "spans": [ + { + "bbox": [ + 313, + 70, + 555, + 105 + ], + "type": "text", + "content": "Table 2. Comparison of MMC and OOD detection performance for different ReLU removal proportions using ResNet-101 as the classifier. Results are averaged across multiple OOD datasets." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 320, + 107, + 548, + 200 + ], + "lines": [ + { + "bbox": [ + 320, + 107, + 548, + 200 + ], + "spans": [ + { + "bbox": [ + 320, + 107, + 548, + 200 + ], + "type": "table", + "html": "
IDCIFAR-10CIFAR-100
RatioMMC↓FPR95↓AUROC↑MMC↓FPR95↓AUROC↑
0%83.4651.9292.5373.8484.4173.03
1%84.4657.5890.4869.1482.4376.22
2%77.1843.0093.6972.2683.0775.25
3%77.9146.9393.0669.3778.7478.12
5%83.0756.5390.7168.7178.5779.43
10%79.6050.0692.2571.2480.9477.70
20%80.7850.3192.3073.5482.7374.81
50%79.2248.3492.7370.4082.2275.40
", + "image_path": "4ffa6e7764a95fff13f73b17cd996d9e5cf871addff07cfe74f627858ec654f7.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 213, + 555, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 213, + 555, + 346 + ], + "spans": [ + { + "bbox": [ + 313, + 213, + 555, + 346 + ], + "type": "text", + "content": "network capacity to ResNet-18 improves OOD overconfidence and enhances ID/OOD discrimination. However, further reducing capacity to ResNet-10 results in degraded performance, likely due to insufficient capacity to effectively handle CIFAR-100 classification. A similar trend is observed on ImageNet-100, where reducing the student network capacity to ResNet-50 continues to yield improvements, but further reductions impair performance, likely due to the dataset's increased complexity. These findings suggest that reducing network capacity to a moderate level can effectively alleviate the OOD overconfidence issue." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 353, + 397, + 365 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 353, + 397, + 365 + ], + "spans": [ + { + "bbox": [ + 314, + 353, + 397, + 365 + ], + "type": "text", + "content": "3.2. Nonlinearity" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 371, + 554, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 371, + 554, + 455 + ], + "spans": [ + { + "bbox": [ + 313, + 371, + 554, + 455 + ], + "type": "text", + "content": "Nonlinearity in DNNs arises from activation functions, e.g., ReLU and Sigmoid. The more activation operations are applied, the greater the degree of nonlinearity. In this subsection, we investigate whether the degree of nonlinearity affects the severity of OOD overconfidence. Given that ReLU is one of the most widely used activation functions, we focus on examining the effect of removing ReLU operations." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 455, + 554, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 455, + 554, + 563 + ], + "spans": [ + { + "bbox": [ + 313, + 455, + 554, + 563 + ], + "type": "text", + "content": "Specifically, we examine the impact of removing ReLU operations at different proportions and locations, similar to the approach in [4]. as we will detail in Sec. 4, to evaluate their effect on mitigating OOD overconfidence. Additionally, we assess the influence of initiating ReLU removal at various stages of training. For these experiments, CIFAR-10 and CIFAR-100 are used as the ID datasets, with SVHN, LSUN-R, iSUN, Textures, and TinyImageNet-R as the OOD datasets, using ResNet-18/101 as the classifier." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 563, + 553, + 623 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 563, + 553, + 623 + ], + "spans": [ + { + "bbox": [ + 313, + 563, + 553, + 623 + ], + "type": "text", + "content": "Analysis of Different Removal Proportions. To assess the impact of varying removal proportions, we introduce the Negative Response Ratio (NRO), which measures the proportion of negative responses that are passed to ReLU operations:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 323, + 630, + 553, + 666 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 630, + 553, + 666 + ], + "spans": [ + { + "bbox": [ + 323, + 630, + 553, + 666 + ], + "type": "interline_equation", + "content": "N R O = \\frac {\\text {N u m b e r o f n e g a t i v e c o n v o l u t i o n a l r e s p o n s e s}}{\\text {T o t a l n u m b e r o f c o n v o l u t i o n a l r e s p o n s e s}}. \\tag {1}", + "image_path": "8912a333c3586f7f6f814317cce67607f2eece803b2b3592f1f6267aa7c772dc.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 666, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 666, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 666, + 555, + 715 + ], + "type": "text", + "content": "A smaller NRO indicates fewer activations requiring truncation, suggesting that the corresponding ReLU layers contribute less to overall nonlinearity. Based on NRO values, we selectively remove ReLU operations with the low" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5032" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 70, + 70, + 303, + 178 + ], + "blocks": [ + { + "bbox": [ + 70, + 70, + 303, + 178 + ], + "lines": [ + { + "bbox": [ + 70, + 70, + 303, + 178 + ], + "spans": [ + { + "bbox": [ + 70, + 70, + 303, + 178 + ], + "type": "image", + "image_path": "19135528d01fbdfee79fa48fbdcda4f0bbc6b063ba79534135e0e1344c342092.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 304, + 70, + 539, + 179 + ], + "blocks": [ + { + "bbox": [ + 304, + 70, + 539, + 179 + ], + "lines": [ + { + "bbox": [ + 304, + 70, + 539, + 179 + ], + "spans": [ + { + "bbox": [ + 304, + 70, + 539, + 179 + ], + "type": "image", + "image_path": "b8889e50737340affa915d527572fc1e1bb309e3feca8bf34d0bc30c4f98eaf6.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 69, + 175, + 187, + 281 + ], + "blocks": [ + { + "bbox": [ + 69, + 175, + 187, + 281 + ], + "lines": [ + { + "bbox": [ + 69, + 175, + 187, + 281 + ], + "spans": [ + { + "bbox": [ + 69, + 175, + 187, + 281 + ], + "type": "image", + "image_path": "b7bbd30e25a50659fc0a95cd265e3c0a4fa6a07ed765e9a686a75683f3190403.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 300, + 555, + 323 + ], + "lines": [ + { + "bbox": [ + 55, + 300, + 555, + 323 + ], + "spans": [ + { + "bbox": [ + 55, + 300, + 555, + 323 + ], + "type": "text", + "content": "Figure 2. Comparison of MMC and OOD detection metrics, including FPR95, AUROC, and AUPRin, for student models with varying capacities (e.g., ResNet models) across datasets: (a) CIFAR-100 and (b) ImageNet-100." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 188, + 182, + 303, + 281 + ], + "blocks": [ + { + "bbox": [ + 188, + 182, + 303, + 281 + ], + "lines": [ + { + "bbox": [ + 188, + 182, + 303, + 281 + ], + "spans": [ + { + "bbox": [ + 188, + 182, + 303, + 281 + ], + "type": "image", + "image_path": "0b695f5d3704bfe92f4f6a91eb4ec4f0903c73dcfe1cfa4ecbbaa9cafc7e32d6.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 246, + 285, + 372, + 297 + ], + "lines": [ + { + "bbox": [ + 246, + 285, + 372, + 297 + ], + "spans": [ + { + "bbox": [ + 246, + 285, + 372, + 297 + ], + "type": "text", + "content": "Teacher Student" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 304, + 182, + 421, + 281 + ], + "blocks": [ + { + "bbox": [ + 304, + 182, + 421, + 281 + ], + "lines": [ + { + "bbox": [ + 304, + 182, + 421, + 281 + ], + "spans": [ + { + "bbox": [ + 304, + 182, + 421, + 281 + ], + "type": "image", + "image_path": "301ad9c7ffc67ad80272d3b6901be0ab800881fe239a99c2e8aa6d17f490666a.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 422, + 182, + 538, + 281 + ], + "blocks": [ + { + "bbox": [ + 422, + 182, + 538, + 281 + ], + "lines": [ + { + "bbox": [ + 422, + 182, + 538, + 281 + ], + "spans": [ + { + "bbox": [ + 422, + 182, + 538, + 281 + ], + "type": "image", + "image_path": "521aa6f536e610a027968e764225eecf6777f02685bbb4831f6844b2bd32fd44.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 70, + 335, + 278, + 446 + ], + "blocks": [ + { + "bbox": [ + 70, + 335, + 278, + 446 + ], + "lines": [ + { + "bbox": [ + 70, + 335, + 278, + 446 + ], + "spans": [ + { + "bbox": [ + 70, + 335, + 278, + 446 + ], + "type": "image", + "image_path": "01f77a8c33e4dc7cae2ae8746efb3e495b5e1a150cbfd698a4aca7d73c6ef79a.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 450, + 295, + 473 + ], + "lines": [ + { + "bbox": [ + 55, + 450, + 295, + 473 + ], + "spans": [ + { + "bbox": [ + 55, + 450, + 295, + 473 + ], + "type": "text", + "content": "Figure 3. Comparison of MMC and FPR95 for ResNet-101 with ReLU removal applied at different network locations." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 493, + 295, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 493, + 295, + 577 + ], + "spans": [ + { + "bbox": [ + 55, + 493, + 295, + 577 + ], + "type": "text", + "content": "est NRO scores at proportions of " + }, + { + "bbox": [ + 55, + 493, + 295, + 577 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 55, + 493, + 295, + 577 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 493, + 295, + 577 + ], + "type": "inline_equation", + "content": "2\\%" + }, + { + "bbox": [ + 55, + 493, + 295, + 577 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 493, + 295, + 577 + ], + "type": "inline_equation", + "content": "3\\%" + }, + { + "bbox": [ + 55, + 493, + 295, + 577 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 493, + 295, + 577 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 55, + 493, + 295, + 577 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 493, + 295, + 577 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 55, + 493, + 295, + 577 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 493, + 295, + 577 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 55, + 493, + 295, + 577 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 493, + 295, + 577 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 55, + 493, + 295, + 577 + ], + "type": "text", + "content": ", starting at the 40th epoch and continuing through the 100th epoch. As shown in Tab. 2, removing a certain proportion of ReLU operations reduces OOD overconfidence and enhances OOD detection. For example, removing " + }, + { + "bbox": [ + 55, + 493, + 295, + 577 + ], + "type": "inline_equation", + "content": "2\\%" + }, + { + "bbox": [ + 55, + 493, + 295, + 577 + ], + "type": "text", + "content": " of ReLU operations on CIFAR-10 and " + }, + { + "bbox": [ + 55, + 493, + 295, + 577 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 55, + 493, + 295, + 577 + ], + "type": "text", + "content": " on CIFAR-100 yields the best results." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 579, + 295, + 699 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 579, + 295, + 699 + ], + "spans": [ + { + "bbox": [ + 55, + 579, + 295, + 699 + ], + "type": "text", + "content": "Analysis of Different Removal Locations. To identify which ReLU operations are most suitable for removal, we compare the effects of removing ReLU operations from shallow, middle, and deep layers, as well as from layers with the lowest NRO values. Following previous findings, we remove " + }, + { + "bbox": [ + 55, + 579, + 295, + 699 + ], + "type": "inline_equation", + "content": "2\\%" + }, + { + "bbox": [ + 55, + 579, + 295, + 699 + ], + "type": "text", + "content": " of ReLU operations for CIFAR-10 and " + }, + { + "bbox": [ + 55, + 579, + 295, + 699 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 55, + 579, + 295, + 699 + ], + "type": "text", + "content": " for CIFAR-100, initiating removal at the 40th epoch. As shown in Fig. 3, each removal strategy effectively reduces OOD overconfidence, with the best results achieved by removing ReLU operations based on NRO values." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 701, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 701, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 701, + 295, + 713 + ], + "type": "text", + "content": "Impact of ReLU Removal Initiation Points. We also" + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 332, + 339, + 538, + 417 + ], + "blocks": [ + { + "bbox": [ + 332, + 339, + 538, + 417 + ], + "lines": [ + { + "bbox": [ + 332, + 339, + 538, + 417 + ], + "spans": [ + { + "bbox": [ + 332, + 339, + 538, + 417 + ], + "type": "image", + "image_path": "09125d787ec7e270736996341582f23c8db509209b4a91dacf06d71b306f5a76.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 422, + 554, + 445 + ], + "lines": [ + { + "bbox": [ + 313, + 422, + 554, + 445 + ], + "spans": [ + { + "bbox": [ + 313, + 422, + 554, + 445 + ], + "type": "text", + "content": "Figure 4. Comparison of MMC and FPR95 for ResNet-18 with ReLU removal initiated at different training epochs." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 465, + 555, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 465, + 555, + 538 + ], + "spans": [ + { + "bbox": [ + 313, + 465, + 555, + 538 + ], + "type": "text", + "content": "analyze the effects of initiating ReLU removal at different stages during training. As illustrated in Fig. 4, the specific stage at which ReLU removal begins has only a moderate impact on overall performance. The best results are achieved when removal is initiated around the middle of the training process, such as at the 40th epoch." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 549, + 544, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 549, + 544, + 563 + ], + "spans": [ + { + "bbox": [ + 313, + 549, + 544, + 563 + ], + "type": "text", + "content": "4. A Practical Guide to Model Simplification" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "content": "Based on our findings reported in Sec. 3, we present a practical guide to model simplification for addressing the OOD overconfidence issue in DNNs. For models that exhibit high OOD overconfidence, we suggest reducing complexity through a combination of knowledge distillation and selective ReLU removal. Please refer to Fig. 5 for demonstration. Knowledge Distillation with a Moderately-Sized Student. The first step is to choose a small student model with a capacity suited to the task requirements. This student model should be sufficiently compact to reduce complexity but still capable of performing well on the original task. We then apply knowledge distillation (e.g., KD) to transfer es" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5033" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 100, + 69, + 512, + 249 + ], + "blocks": [ + { + "bbox": [ + 100, + 69, + 512, + 249 + ], + "lines": [ + { + "bbox": [ + 100, + 69, + 512, + 249 + ], + "spans": [ + { + "bbox": [ + 100, + 69, + 512, + 249 + ], + "type": "image", + "image_path": "bf8fc906eb6da835625877dfe77bf98c2395802d61b870b8320876d358dc070b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 252, + 555, + 288 + ], + "lines": [ + { + "bbox": [ + 54, + 252, + 555, + 288 + ], + "spans": [ + { + "bbox": [ + 54, + 252, + 555, + 288 + ], + "type": "text", + "content": "Figure 5. Illustration of our model simplification pipeline. Given a complex teacher model with high capacity and nonlinearity, we use knowledge distillation to transfer information to a low-capacity student model. During distillation, ReLU operations are selectively replaced with Leaky ReLU, and the negative slope is iteratively adjusted from 0 to 1, ultimately transforming ReLU into an identity function." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 307, + 294, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 307, + 294, + 354 + ], + "spans": [ + { + "bbox": [ + 54, + 307, + 294, + 354 + ], + "type": "text", + "content": "sential knowledge from the larger teacher model to the student. This process enables the student to approximate the teacher's performance while operating with a reduced capacity, thereby alleviating the OOD overconfidence issue." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 355, + 295, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 355, + 295, + 474 + ], + "spans": [ + { + "bbox": [ + 54, + 355, + 295, + 474 + ], + "type": "text", + "content": "Selective ReLU Removal During Distillation. To further streamline the model, we selectively remove ReLU operations. Using the NRO metric defined in Eq. (1), we identify ReLU operations that contribute minimally to the model's responses. Following the approach in [4], we replace these selected ReLU operations with Leaky ReLU, initially setting the negative slope to 0 and gradually increasing it to 1 midway through the distillation process. This controlled transition effectively transforms ReLU into an identity function, thereby reducing the model's nonlinearity." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 475, + 295, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 475, + 295, + 522 + ], + "spans": [ + { + "bbox": [ + 55, + 475, + 295, + 522 + ], + "type": "text", + "content": "This practical guide, combining knowledge distillation for capacity reduction with selective ReLU removal for nonlinearity reduction, provides an accessible approach to mitigate OOD overconfidence issue in DNNs." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 532, + 181, + 545 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 532, + 181, + 545 + ], + "spans": [ + { + "bbox": [ + 55, + 532, + 181, + 545 + ], + "type": "text", + "content": "5. Experimental Results" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 552, + 173, + 565 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 552, + 173, + 565 + ], + "spans": [ + { + "bbox": [ + 55, + 552, + 173, + 565 + ], + "type": "text", + "content": "5.1. Experimental Setup" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 570, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 295, + 713 + ], + "type": "text", + "content": "Implementation Details. We employ KD [22] to transfer knowledge from teacher to student models, with a weight ratio of 1.5:1 between KD loss and cross-entropy (CE) loss. All student models are trained for 100 epochs, with selective ReLU removal starting at the 40th epoch and continuing through the end of training, similar to the approach in [4]. We remove " + }, + { + "bbox": [ + 55, + 570, + 295, + 713 + ], + "type": "inline_equation", + "content": "2\\%" + }, + { + "bbox": [ + 55, + 570, + 295, + 713 + ], + "type": "text", + "content": " of ReLU operations for models trained on CIFAR-10, " + }, + { + "bbox": [ + 55, + 570, + 295, + 713 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 55, + 570, + 295, + 713 + ], + "type": "text", + "content": " for models trained on CIFAR-100, and " + }, + { + "bbox": [ + 55, + 570, + 295, + 713 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 55, + 570, + 295, + 713 + ], + "type": "text", + "content": " for models trained on ImageNet-1k [7]. All experiments are implemented in PyTorch [35] and conducted on a workstation equipped with dual 2.40 GHz CPUs, 128 GB RAM, and eight NVIDIA RTX 3090Ti GPUs." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 307, + 555, + 463 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 307, + 555, + 463 + ], + "spans": [ + { + "bbox": [ + 313, + 307, + 555, + 463 + ], + "type": "text", + "content": "Setup. For (1) Mitigating OOD Overconfidence, we use CIFAR-10 and CIFAR-100 [24] as ID datasets, with SVHN [33], LSUN-R [59], iSUN [54], Textures [5], and TINR, a resized variant of TinyImageNet uniformly scaled to " + }, + { + "bbox": [ + 313, + 307, + 555, + 463 + ], + "type": "inline_equation", + "content": "32 \\times 32" + }, + { + "bbox": [ + 313, + 307, + 555, + 463 + ], + "type": "text", + "content": ", as the OOD datasets. For (2) OOD Detection, we use the same CIFAR-based configuration as in (1). For large-scale evaluations, we use ImageNet-1k [7] as the ID dataset, with SUN [52], Places [65], Textures [5], and OpenImage-O [49] as the OOD datasets. We train ResNet-50 and ResNet-101 [16] on CIFAR-10 and CIFAR-100, and train DenseNet-201 [23] and WideResNet-101-2 (WRN-101) [63] on ImageNet-1k [7], following the training protocols specified in their respective original papers." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 468, + 556, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 468, + 556, + 635 + ], + "spans": [ + { + "bbox": [ + 313, + 468, + 556, + 635 + ], + "type": "text", + "content": "Setting of Our Approach. We designate ResNet-18 and WideResNet-16-10 (WRN-16) as student networks for ResNet-50, and ResNet-34 and WideResNet-28-10 (WRN-28) as students for ResNet-101 on CIFAR datasets. For the ImageNet-1k dataset, DenseNet-161 serves as the student for DenseNet-201, and WideResNet-50-2 (WRN-50) for WRN-101. All student models are trained using the SGD optimizer with a momentum of 0.9 and an initial learning rate of 0.1. On the ImageNet-1k benchmark, the learning rate is reduced by a factor of 10 every 30 epochs, while on the CIFAR benchmarks, the learning rate is adjusted using the CosineAnnealingLR scheduler. For simplicity, we refer to our method as \"Ours (student)\" in subsequent sections to indicate the specific student network employed." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "text", + "content": "Baselines. For (1) Mitigating OOD Overconfidence, we select three state-of-the-art baselines: ACET [18], CEDA [18], and CCUs [30]. For (2) OOD Detection, we benchmark our approach against five post-hoc methods: maximum softmax probability (MSP) [19], MaxLogit [21], Energy [29], ReAct [39], and FeatureNorm [61]." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5034" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 97, + 106, + 515, + 331 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 103 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 103 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 103 + ], + "type": "text", + "content": "Table 3. Comparison of MMC on OOD datasets for ResNet-50 and ResNet-101 models trained on CIFAR-10 and CIFAR-100, evaluated using various OOD overconfidence mitigation techniques. Underlined values denote the best performance for a single method, while bolded values indicate the best performance across all methods, including both individual and combined methods." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 97, + 106, + 515, + 331 + ], + "lines": [ + { + "bbox": [ + 97, + 106, + 515, + 331 + ], + "spans": [ + { + "bbox": [ + 97, + 106, + 515, + 331 + ], + "type": "table", + "html": "
ModelMethodID: CIFAR-10AverageID: CIFAR-100Average
SVHNLSUN-RiSUNTexturesTINRSVHNLSUN-RiSUNTexturesTINR
ResNet-50—75.3479.6479.8681.5381.5379.5872.5072.0772.2072.7471.1172.12
CEDA74.7370.9172.0273.3778.6473.9355.6454.7756.2562.8555.4256.99
ACET74.2967.3567.7270.3769.8769.9263.6862.4962.1964.4461.1562.79
CCUs69.5663.5163.1068.6468.9266.7557.3352.3054.7161.2554.8956.09
Ours (ResNet-18)78.2473.2974.3779.5176.9876.4854.7543.6542.2761.4039.6948.35
Ours (WRN-16)63.6566.3268.4977.0970.1869.1543.0651.7651.9957.9652.3751.43
CEDA + Ours (WRN-16)61.3167.4069.6470.0973.5468.4047.8459.4060.6959.2460.0057.43
ACET + Ours (WRN-16)67.4461.2862.3670.0565.4865.3252.5459.4561.5960.8260.4858.98
CCUs + Ours (WRN-16)65.9758.8161.3264.8465.9663.3850.8849.0451.8558.4051.8252.40
ResNet-101—83.4482.1383.0284.5684.1783.4665.2577.2178.0373.3875.3273.84
CEDA70.0768.2668.4274.4274.0271.0458.8258.1859.0066.2358.4760.14
ACET75.8253.6955.0064.5059.9861.8063.6259.5758.4460.3458.0160.00
CCUs73.5665.7264.9866.9469.8368.2158.4554.5356.5259.8056.3257.12
Ours (ResNet-34)75.2274.8175.9482.0878.8677.3863.6258.4260.1366.6959.8961.75
Ours (WRN-28)71.2174.7376.0381.8077.8976.3355.6947.8451.2063.6954.0954.50
CEDA + Ours (WRN-28)67.1568.2170.5976.7773.8371.3156.5165.1266.5962.5166.7763.50
ACET + Ours (WRN-28)59.3352.7053.0766.9558.0558.0256.1062.4863.9458.3463.8660.95
CCUs + Ours (WRN-28)67.7961.7464.2065.9568.8065.7053.7150.9653.5857.8654.5254.13
", + "image_path": "a91d1c4df56b66c6db7db4329e9c32041400b2e2012c7a1d597cb46e2c4d64fc.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 77, + 350, + 167, + 406 + ], + "blocks": [ + { + "bbox": [ + 78, + 342, + 87, + 350 + ], + "lines": [ + { + "bbox": [ + 78, + 342, + 87, + 350 + ], + "spans": [ + { + "bbox": [ + 78, + 342, + 87, + 350 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 77, + 350, + 167, + 406 + ], + "lines": [ + { + "bbox": [ + 77, + 350, + 167, + 406 + ], + "spans": [ + { + "bbox": [ + 77, + 350, + 167, + 406 + ], + "type": "image", + "image_path": "2a70a23a637a5e8fe3f704e446e6090892e643e0f8fa1c4cf8befca0eb7c2718.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 78, + 405, + 87, + 411 + ], + "lines": [ + { + "bbox": [ + 78, + 405, + 87, + 411 + ], + "spans": [ + { + "bbox": [ + 78, + 405, + 87, + 411 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 77, + 411, + 167, + 470 + ], + "blocks": [ + { + "bbox": [ + 77, + 411, + 167, + 470 + ], + "lines": [ + { + "bbox": [ + 77, + 411, + 167, + 470 + ], + "spans": [ + { + "bbox": [ + 77, + 411, + 167, + 470 + ], + "type": "image", + "image_path": "5450420c4e84847d0faa06662c292c27f4cd299cf0f3c2763cd43382beb50d1f.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 168, + 349, + 257, + 407 + ], + "blocks": [ + { + "bbox": [ + 214, + 342, + 228, + 349 + ], + "lines": [ + { + "bbox": [ + 214, + 342, + 228, + 349 + ], + "spans": [ + { + "bbox": [ + 214, + 342, + 228, + 349 + ], + "type": "text", + "content": "ACET" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 168, + 349, + 257, + 407 + ], + "lines": [ + { + "bbox": [ + 168, + 349, + 257, + 407 + ], + "spans": [ + { + "bbox": [ + 168, + 349, + 257, + 407 + ], + "type": "image", + "image_path": "154d7c306de00cb139e4e86ecf243644f310051d2d3e7abe9102816e300c8e5b.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 258, + 349, + 348, + 407 + ], + "blocks": [ + { + "bbox": [ + 304, + 342, + 319, + 349 + ], + "lines": [ + { + "bbox": [ + 304, + 342, + 319, + 349 + ], + "spans": [ + { + "bbox": [ + 304, + 342, + 319, + 349 + ], + "type": "text", + "content": "CEDA" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 258, + 349, + 348, + 407 + ], + "lines": [ + { + "bbox": [ + 258, + 349, + 348, + 407 + ], + "spans": [ + { + "bbox": [ + 258, + 349, + 348, + 407 + ], + "type": "image", + "image_path": "d683d1fef2149b66067bae02fe792dedf427a55d07dac875bffc6d84f9c63b0a.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 349, + 349, + 438, + 407 + ], + "blocks": [ + { + "bbox": [ + 395, + 342, + 410, + 349 + ], + "lines": [ + { + "bbox": [ + 395, + 342, + 410, + 349 + ], + "spans": [ + { + "bbox": [ + 395, + 342, + 410, + 349 + ], + "type": "text", + "content": "CCUs" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 349, + 349, + 438, + 407 + ], + "lines": [ + { + "bbox": [ + 349, + 349, + 438, + 407 + ], + "spans": [ + { + "bbox": [ + 349, + 349, + 438, + 407 + ], + "type": "image", + "image_path": "c58ce1c74130b039c4d9056eca9d0802ac214005ae37fab257fa844c8d247a26.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 440, + 349, + 529, + 407 + ], + "blocks": [ + { + "bbox": [ + 473, + 342, + 512, + 349 + ], + "lines": [ + { + "bbox": [ + 473, + 342, + 512, + 349 + ], + "spans": [ + { + "bbox": [ + 473, + 342, + 512, + 349 + ], + "type": "text", + "content": "Ours (ResNet-18)" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 440, + 349, + 529, + 407 + ], + "lines": [ + { + "bbox": [ + 440, + 349, + 529, + 407 + ], + "spans": [ + { + "bbox": [ + 440, + 349, + 529, + 407 + ], + "type": "image", + "image_path": "195006fd2e1c97a2d69f08dae8f6f38e157e8e06aeab95d519a2cb87a577b88b.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 77, + 483, + 167, + 542 + ], + "blocks": [ + { + "bbox": [ + 78, + 476, + 87, + 483 + ], + "lines": [ + { + "bbox": [ + 78, + 476, + 87, + 483 + ], + "spans": [ + { + "bbox": [ + 78, + 476, + 87, + 483 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 77, + 483, + 167, + 542 + ], + "lines": [ + { + "bbox": [ + 77, + 483, + 167, + 542 + ], + "spans": [ + { + "bbox": [ + 77, + 483, + 167, + 542 + ], + "type": "image", + "image_path": "e81ef09903ff89a34b2893ed9cc7fe4e1630ab89cf858bd580e0130fe071d9ec.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 78, + 539, + 87, + 546 + ], + "lines": [ + { + "bbox": [ + 78, + 539, + 87, + 546 + ], + "spans": [ + { + "bbox": [ + 78, + 539, + 87, + 546 + ], + "type": "text", + "content": "(d)" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 77, + 545, + 167, + 604 + ], + "blocks": [ + { + "bbox": [ + 77, + 545, + 167, + 604 + ], + "lines": [ + { + "bbox": [ + 77, + 545, + 167, + 604 + ], + "spans": [ + { + "bbox": [ + 77, + 545, + 167, + 604 + ], + "type": "image", + "image_path": "c0844b9b5ba1f2864d2361bc5eb81364dc23e04710ccf26e08a3dae0c41f836d.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 78, + 607, + 555, + 640 + ], + "lines": [ + { + "bbox": [ + 78, + 607, + 555, + 640 + ], + "spans": [ + { + "bbox": [ + 78, + 607, + 555, + 640 + ], + "type": "text", + "content": "Figure 6. Histograms (logarithmic scale) of MMC values for (a, b) ResNet-50 and (c, d) ResNet-101, w/ and w/o applying mitigation techniques: ACET, CEDA, CCUs, and Ours. All models are trained on CIFAR-100 and evaluated on OOD datasets: (a, c) LSUN-R and (d) TinyImageNet-R. In our approach, ResNet-18 and WRN-28 serve as students for ResNet-50 and ResNet-101, respectively." + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 168, + 483, + 257, + 542 + ], + "blocks": [ + { + "bbox": [ + 214, + 476, + 235, + 483 + ], + "lines": [ + { + "bbox": [ + 214, + 476, + 235, + 483 + ], + "spans": [ + { + "bbox": [ + 214, + 476, + 235, + 483 + ], + "type": "text", + "content": "ACET" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 168, + 483, + 257, + 542 + ], + "lines": [ + { + "bbox": [ + 168, + 483, + 257, + 542 + ], + "spans": [ + { + "bbox": [ + 168, + 483, + 257, + 542 + ], + "type": "image", + "image_path": "5f3f12b6ef1143b59a059d7d12e983c1526e867a2d63a304238575ebc2ac6b72.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 258, + 483, + 348, + 541 + ], + "blocks": [ + { + "bbox": [ + 304, + 476, + 318, + 483 + ], + "lines": [ + { + "bbox": [ + 304, + 476, + 318, + 483 + ], + "spans": [ + { + "bbox": [ + 304, + 476, + 318, + 483 + ], + "type": "text", + "content": "CEDA" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 258, + 483, + 348, + 541 + ], + "lines": [ + { + "bbox": [ + 258, + 483, + 348, + 541 + ], + "spans": [ + { + "bbox": [ + 258, + 483, + 348, + 541 + ], + "type": "image", + "image_path": "ff89139269c01983fdda6fe307d4f08114af96d0a9f4a0bcb690b1f1e5430d4c.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 349, + 483, + 438, + 541 + ], + "blocks": [ + { + "bbox": [ + 395, + 477, + 410, + 483 + ], + "lines": [ + { + "bbox": [ + 395, + 477, + 410, + 483 + ], + "spans": [ + { + "bbox": [ + 395, + 477, + 410, + 483 + ], + "type": "text", + "content": "CCUs" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 349, + 483, + 438, + 541 + ], + "lines": [ + { + "bbox": [ + 349, + 483, + 438, + 541 + ], + "spans": [ + { + "bbox": [ + 349, + 483, + 438, + 541 + ], + "type": "image", + "image_path": "011606af33bf67277ae955b252a1ece844b99209b093b7d7ddd7cd2b207b2f9c.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 440, + 484, + 529, + 543 + ], + "blocks": [ + { + "bbox": [ + 476, + 478, + 510, + 484 + ], + "lines": [ + { + "bbox": [ + 476, + 478, + 510, + 484 + ], + "spans": [ + { + "bbox": [ + 476, + 478, + 510, + 484 + ], + "type": "text", + "content": "Ours (WRN-28)" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 440, + 484, + 529, + 543 + ], + "lines": [ + { + "bbox": [ + 440, + 484, + 529, + 543 + ], + "spans": [ + { + "bbox": [ + 440, + 484, + 529, + 543 + ], + "type": "image", + "image_path": "196b6591331d6b3a402901ff99e39d9e570fed47804d5134bef54eed362cd2e0.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "bbox": [ + 77, + 651, + 279, + 664 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 651, + 279, + 664 + ], + "spans": [ + { + "bbox": [ + 77, + 651, + 279, + 664 + ], + "type": "text", + "content": "Results of Mitigating OOD Overconfidence" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 55, + 677, + 294, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 294, + 713 + ], + "type": "text", + "content": "Comparison with State-of-the-art Methods. To assess the effectiveness and superiority of our approach, we conduct comparative experiments evaluating OOD overconfi" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 313, + 652, + 555, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 652, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 652, + 555, + 712 + ], + "type": "text", + "content": "dence for ResNet-50 and ResNet-101 trained on CIFAR-10 and CIFAR-100. The results in Tab. 3 demonstrate that our approach consistently reduces the mean maximum confidence (MMC) across different student networks in all cases, validating its effectiveness. In comparison, methods such" + } + ] + } + ], + "index": 29 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "5035" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 59, + 97, + 555, + 441 + ], + "blocks": [ + { + "bbox": [ + 55, + 71, + 555, + 94 + ], + "lines": [ + { + "bbox": [ + 55, + 71, + 555, + 94 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 555, + 94 + ], + "type": "text", + "content": "Table 4. Comparison of OOD detection performance for various methods on the CIFAR-10 and CIFAR-100 datasets. Underlined values indicate results enhanced by applying our approach to the baseline, while bolded values denote the best performance." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 59, + 97, + 555, + 441 + ], + "lines": [ + { + "bbox": [ + 59, + 97, + 555, + 441 + ], + "spans": [ + { + "bbox": [ + 59, + 97, + 555, + 441 + ], + "type": "table", + "html": "
IDModelMethodOODAverage
SVHNLSUN-RiSUNTexturesTINR
FPR95↓AUROC↑FPR95↓AUROC↑FPR95↓AUROC↑FPR95↓AUROC↑FPR95↓AUROC↑FPR95↓AUROC↑
CIFAR-10ResNet-50MSP43.7594.1954.3291.9655.4691.6659.0190.3158.6190.8054.2391.78
MSP + Ours (ResNet-18)51.8293.2741.9393.8944.8693.2755.6490.5750.1891.9748.8992.59
MSP + Ours (WRN-16)28.1896.1235.7495.0139.7894.3457.0089.7942.3593.7540.6193.80
MaxLogit22.0796.3132.0394.9534.9594.5346.3391.3944.9692.5736.0793.95
MaxLogit + Ours (ResNet-18)30.2594.9425.0695.3427.9194.7344.5091.7136.7592.7332.8993.89
MaxLogit + Ours (WRN-16)17.6196.9618.5496.9221.6996.3947.7089.5827.8195.5526.6795.08
Energy20.3596.4029.7995.1132.8094.6745.4191.4443.7692.6434.4294.05
Energy + Ours (ResNet-18)28.5395.0723.9995.4326.5594.8344.1389.7035.9292.7931.8293.56
Energy + Ours (WRN-16)18.8096.8818.0797.0220.9196.5047.8989.5427.6395.6126.6695.11
ReAct37.6691.1016.2496.6419.6595.8446.7286.2930.7993.4630.2192.67
ReAct + Ours (ResNet-18)31.5994.6223.8795.5126.7894.8643.6090.1935.8792.9232.3493.62
ReAct + Ours (WRN-16)19.0796.8417.9097.0320.6696.5147.4589.8427.2495.6826.4695.18
FeatureNorm9.2597.9777.3482.0770.2983.9648.9784.1667.7184.4954.7186.53
FeatureNorm + Ours (ResNet-18)2.7399.3524.3495.7521.0496.2822.8294.8929.5294.3620.0996.13
FeatureNorm + Ours (WRN-16)3.3799.2622.0396.4321.5496.4753.6778.8529.3594.5025.9993.10
CIFAR-100ResNet-101MSP76.3081.1486.6769.7688.1169.0685.6973.7385.2771.4784.4173.03
MSP + Ours (ResNet-34)77.4878.9272.7483.0974.8382.1482.5077.2674.5282.2076.4180.72
MSP + Ours (WRN-28)74.8781.3962.5785.7166.9483.9383.2876.3369.4481.7071.4281.81
MaxLogit68.0586.7782.8474.3484.9473.4087.2574.8181.9275.0581.0076.87
MaxLogit + Ours (ResNet-34)71.6983.4763.7487.2665.6986.5481.7478.4666.8886.4469.9584.43
MaxLogit + Ours (WRN-28)73.8482.6154.3089.1160.2987.2683.6277.0664.1284.9867.2384.20
Energy68.3886.9482.3874.6484.2173.6788.0374.6281.4075.2580.8877.02
Energy + Ours (ResNet-34)69.2084.1458.8388.0561.1087.3781.6878.5162.8987.1866.7485.05
Energy + Ours (WRN-28)74.1782.6148.6090.0755.0588.1284.1077.0160.4185.7164.4784.70
ReAct65.6888.1486.4469.1886.9769.0983.4480.1383.8071.5981.2775.63
ReAct + Ours (ResNet-34)60.2688.3057.9486.4958.1186.5263.0785.7860.0385.6959.8886.56
ReAct + Ours (WRN-28)71.8986.7347.3489.2053.6687.4680.5781.4858.9784.6662.4985.91
FeatureNorm31.8894.5999.5434.1199.1039.0253.4282.3298.2242.3276.4358.47
FeatureNorm + Ours (ResNet-34)41.4790.4792.4671.4991.4172.8965.1880.3389.9969.0076.1076.84
FeatureNorm + Ours (WRN-28)54.1789.0880.8679.5280.6979.9171.4974.3778.0080.7473.0480.72
", + "image_path": "5fe2cce378dda2cce1d573c8a6201f09a59af1e5463fa9acc29c6edd94c89e71.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 455, + 295, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 455, + 295, + 552 + ], + "spans": [ + { + "bbox": [ + 55, + 455, + 295, + 552 + ], + "type": "text", + "content": "as CEDA, ACET, and CCUs, which employ specific techniques to suppress confidence scores during training, also significantly mitigate OOD overconfidence in DNNs. However, our approach, which does not incorporate specific confidence suppression measures, still outperforms these state-of-the-art methods, particularly in the more challenging CIFAR-100 scenario. This further highlights the robustness and effectiveness of our strategy." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 554, + 295, + 627 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 554, + 295, + 627 + ], + "spans": [ + { + "bbox": [ + 55, + 554, + 295, + 627 + ], + "type": "text", + "content": "On CIFAR-10, while our base method may not show a pronounced advantage over others, combining our approach with confidence suppression techniques provides additional improvements, surpassing the performance of these state-of-the-art methods. These results reinforce the effectiveness of our approach in addressing OOD overconfidence." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 629, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 295, + 715 + ], + "type": "text", + "content": "Visualization. We visualize the maximum confidence scores predicted by various DNN classifiers as histograms (logarithmic scale) in Fig. 6. As shown, applying OOD overconfidence mitigation solutions shifts the confidence distributions of the Baseline to the left, substantially reducing the number of samples with high confidence scores. Notably, our approach achieves the most pronounced shift, un" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 455, + 555, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 455, + 555, + 480 + ], + "spans": [ + { + "bbox": [ + 313, + 455, + 555, + 480 + ], + "type": "text", + "content": "derscoring its effectiveness and validating its superiority in mitigating OOD overconfidence." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 489, + 511, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 489, + 511, + 502 + ], + "spans": [ + { + "bbox": [ + 313, + 489, + 511, + 502 + ], + "type": "text", + "content": "5.3. Results of Enhancing OOD Detection" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 509, + 555, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 509, + 555, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 509, + 555, + 628 + ], + "type": "text", + "content": "CIFAR Experiments. We evaluate the effectiveness of our approach in enhancing OOD detection. The results in Tab. 4 demonstrate that applying our method to existing OOD detection approaches leads to significant performance improvements. Even when directly applied to the classifier without any advanced post-hoc processing, i.e., MSP, the performance enhancement is still noticeable. These improvements are consistent across different student model configurations, datasets, and classifiers, validating the effectiveness of our approach in boosting OOD detection." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "content": "ImageNet Experiments. We further validate the effectiveness of our approach in enhancing OOD detection on the ImageNet-1k dataset. The results in Tab. 5 indicate that, similar to the CIFAR experiments, applying our approach significantly improves OOD detection performance across methods, demonstrating the robustness and scalability of our approach in large-scale settings." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "5036" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 61, + 86, + 549, + 303 + ], + "blocks": [ + { + "bbox": [ + 119, + 71, + 488, + 82 + ], + "lines": [ + { + "bbox": [ + 119, + 71, + 488, + 82 + ], + "spans": [ + { + "bbox": [ + 119, + 71, + 488, + 82 + ], + "type": "text", + "content": "Table 5. Comparison of OOD detection performance for various methods on the ImageNet-1k dataset." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 61, + 86, + 549, + 303 + ], + "lines": [ + { + "bbox": [ + 61, + 86, + 549, + 303 + ], + "spans": [ + { + "bbox": [ + 61, + 86, + 549, + 303 + ], + "type": "table", + "html": "
ModelMethodOODAverage
SUNPlacesTexturesOpenImage-O
FPR95↓AUROC↑FPR95↓AUROC↑FPR95↓AUROC↑FPR95↓AUROC↑FPR95↓AUROC↑
DenseNet-201MSP65.7682.2469.5381.1467.8279.3664.9284.4767.0181.80
MSP + Ours (DenseNet-161)66.3181.7268.6181.3266.9779.5562.1285.1066.0081.92
MaxLogit55.4486.0661.8983.9958.4883.4559.6486.9458.8685.11
MaxLogit + Ours (DenseNet-161)55.3387.6759.9586.2655.8085.8460.0388.6057.7887.09
Energy52.8786.0460.4683.7156.1383.3559.8786.6557.3384.94
Energy + Ours (DenseNet-161)50.8888.2556.9086.6052.2586.4558.7688.6054.7087.48
ReAct46.0090.8157.4686.1155.1484.1969.3274.4356.9883.89
ReAct + Ours (DenseNet-161)49.2188.5555.0286.9450.9986.7557.6188.6453.2187.72
FeatureNorm43.0690.0555.7185.4250.4679.9079.9767.2257.3080.65
FeatureNorm + Ours (DenseNet-161)36.5291.7949.2487.5056.7378.2982.2065.4956.1780.77
WRN-101MSP66.9382.4470.4881.3565.1480.9159.8786.1365.6182.71
MSP + Ours (WRN-50)67.7582.8069.0781.8166.3880.9660.9186.7866.0383.09
MaxLogit63.2584.7568.4082.6957.3984.3855.8488.4361.2285.06
MaxLogit + Ours (WRN-50)61.9286.1864.3484.4056.7485.6055.6989.2559.6786.36
Energy65.0684.4670.3282.2757.1184.3758.0488.0262.6384.78
Energy + Ours (WRN-50)63.1885.9666.0684.0055.9085.6857.8788.7760.7586.10
ReAct52.2288.1158.7685.4757.0984.9949.6387.8254.4386.60
ReAct + Ours (WRN-50)37.0292.2448.5488.7050.9087.3654.4985.9247.7488.56
FeatureNorm70.0778.8978.6173.2320.9894.7975.0974.8161.1980.43
FeatureNorm + Ours (WRN-50)60.2283.6774.0175.9617.7196.0969.9478.9055.4783.66
", + "image_path": "6c83316610385ef32e2e9572156dfe65487b7f6951758bec3c7e77f4fa41d3d3.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 104, + 377, + 249, + 434 + ], + "blocks": [ + { + "bbox": [ + 55, + 319, + 294, + 373 + ], + "lines": [ + { + "bbox": [ + 55, + 319, + 294, + 373 + ], + "spans": [ + { + "bbox": [ + 55, + 319, + 294, + 373 + ], + "type": "text", + "content": "Table 6. Comparison of MMC and OOD detection performance for ResNet-50 enhanced by Ours (ResNet-18) and its variants w/o knowledge distillation (KD) and ReLU removal (RR), using CIFAR-10 as the ID dataset, and SVHN, LSUN-R, iSUN, Textures, and TINR as the OOD datasets. Results are averaged." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 104, + 377, + 249, + 434 + ], + "lines": [ + { + "bbox": [ + 104, + 377, + 249, + 434 + ], + "spans": [ + { + "bbox": [ + 104, + 377, + 249, + 434 + ], + "type": "table", + "html": "
KDRRMMC↓FPR95↓AUROC↑
79.5854.2391.78
✓76.8149.8192.42
✓78.4650.5091.76
✓✓76.4848.8992.59
", + "image_path": "fde2ce04aaca16c76111726c9f84627990c92ae2e4070b9d447cfb428d1867da.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 454, + 247, + 467 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 454, + 247, + 467 + ], + "spans": [ + { + "bbox": [ + 55, + 454, + 247, + 467 + ], + "type": "text", + "content": "5.4. Ablation Studies and More Analysis" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 474, + 295, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 474, + 295, + 593 + ], + "spans": [ + { + "bbox": [ + 55, + 474, + 295, + 593 + ], + "type": "text", + "content": "Importance of Knowledge Distillation and ReLU Removal. We conduct comprehensive ablation studies to evaluate the impact of knowledge distillation and ReLU removal, the two essential components of our approach. As shown in Tab. 6, excluding either component notably diminishes the effectiveness of our method in mitigating OOD overconfidence and complicates the accuracy of OOD detection. These findings validate the critical importance of both knowledge distillation and selective ReLU removal in achieving optimal performance with our approach." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "text", + "content": "Knowledge Distillation vs. Network Pruning for Capacity Reduction. We further investigate an alternative capacity reduction method for our approach by applying network pruning. To facilitate a fair comparison with knowledge distillation, we evaluate ResNet-50 and ResNet-101 models trained on CIFAR-10, applying two strategies: (1) using KD [22] to reduce ResNet-50 to ResNet-18 and ResNet-101 to WRN-28, and (2) applying Prune [14] to ResNet-50 and ResNet-101, reducing them to parameter counts comparable to ResNet-18 and WRN-28, respectively. The re" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 324, + 377, + 545, + 445 + ], + "blocks": [ + { + "bbox": [ + 313, + 319, + 555, + 373 + ], + "lines": [ + { + "bbox": [ + 313, + 319, + 555, + 373 + ], + "spans": [ + { + "bbox": [ + 313, + 319, + 555, + 373 + ], + "type": "text", + "content": "Table 7. Comparison of MMC and OOD detection performance for ResNet-50 and ResNet-101 reduced to comparable parameter counts via KD [22] and Prune [14]. CIFAR-10 serves as the ID dataset, with SVHN, LSUN-R, iSUN, Textures, and TINR as OOD datasets. Results are averaged across all five OOD datasets" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 324, + 377, + 545, + 445 + ], + "lines": [ + { + "bbox": [ + 324, + 377, + 545, + 445 + ], + "spans": [ + { + "bbox": [ + 324, + 377, + 545, + 445 + ], + "type": "table", + "html": "
ModelCapacity ReductionMMC↓FPR95↓AUROC↑
ResNet-50—79.5854.2391.78
Prune77.3452.5592.27
KD (ResNet-18)76.8149.8192.42
ResNet-101—83.4651.9292.53
Prune78.2050.9291.61
KD (WRN-28)77.0141.0992.56
", + "image_path": "60035a89324ec8924f96fb8ab61cb275242cae4d322d35e32b3d47d66f281b7f.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 466, + 554, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 466, + 554, + 550 + ], + "spans": [ + { + "bbox": [ + 313, + 466, + 554, + 550 + ], + "type": "text", + "content": "sults in Tab. 7 show that both methods effectively reduce OOD overconfidence and improve OOD detection; however, knowledge distillation consistently delivers a more substantial improvement than network pruning. Additionally, knowledge distillation offers greater stability and flexibility during training, making it a more suitable choice for capacity reduction in our approach." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 562, + 479, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 562, + 479, + 574 + ], + "spans": [ + { + "bbox": [ + 313, + 562, + 479, + 574 + ], + "type": "text", + "content": "6. Conclusion and Future Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "content": "This paper has presented a comprehensive investigation into mitigating OOD overconfidence in DNNs through model simplification. By reducing model capacity via knowledge distillation and decreasing nonlinearity through selective ReLU removal, we demonstrate that OOD overconfidence in DNN classifiers can be significantly mitigated. Furthermore, these strategies can be effectively combined with OOD detection methods to enhance overall performance. In future work, we plan to study the effects of model simplification on OOD overconfidence in greater depth and from a theoretical perspective." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5037" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 159, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 159, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 159, + 85 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 187 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 187 + ], + "type": "text", + "content": "We sincerely thank the anonymous reviewers for their valuable comments. This work was supported in part by the National Natural Science Foundation of China (62472117, U2436208, 62406095), the Guangdong Basic and Applied Basic Research Foundation (2025A1515010157, 2024A1515012064), the Science and Technology Projects in Guangzhou (2025A03J0137, SL2022A04J01112), and the Academician Bixing Fang's Specialized Class." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 196, + 115, + 209 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 196, + 115, + 209 + ], + "spans": [ + { + "bbox": [ + 56, + 196, + 115, + 209 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 216, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 61, + 216, + 295, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 216, + 295, + 237 + ], + "spans": [ + { + "bbox": [ + 61, + 216, + 295, + 237 + ], + "type": "text", + "content": "[1] Lei Jimmy Ba and Rich Caruana. Do deep nets really need to be deep? In NeurIPS, pages 2654-2662, 2014. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 239, + 295, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 239, + 295, + 259 + ], + "spans": [ + { + "bbox": [ + 61, + 239, + 295, + 259 + ], + "type": "text", + "content": "[2] Abhijit Bendale and Terrance Boult. Towards open world recognition. In CVPR, pages 1893-1902, 2015. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 261, + 295, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 261, + 295, + 293 + ], + "spans": [ + { + "bbox": [ + 62, + 261, + 295, + 293 + ], + "type": "text", + "content": "[3] Chao Chen, Zhihang Fu, Kai Liu, Ze Chen, Mingyuan Tao, and Jieping Ye. Optimal parameter and neuron pruning for out-of-distribution detection. In NeurIPS, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 294, + 294, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 294, + 294, + 326 + ], + "spans": [ + { + "bbox": [ + 62, + 294, + 294, + 326 + ], + "type": "text", + "content": "[4] Hanting Chen, Yunhe Wang, Jianyuan Guo, and Dacheng Tao. Vanillanet: the power of minimalism in deep learning. In NeurIPS, 2024. 3, 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 327, + 294, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 327, + 294, + 360 + ], + "spans": [ + { + "bbox": [ + 62, + 327, + 294, + 360 + ], + "type": "text", + "content": "[5] Mircea Cimpoi, Subhransu Maji, Iasonas Kokkinos, Sammy Mohamed, and Andrea Vedaldi. Describing textures in the wild. In CVPR, pages 3606-3613, 2014. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 361, + 294, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 361, + 294, + 392 + ], + "spans": [ + { + "bbox": [ + 62, + 361, + 294, + 392 + ], + "type": "text", + "content": "[6] Jesse Davis and Mark Goadrich. The relationship between precision-recall and roc curves. In ICML, pages 233-240, 2006. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 393, + 294, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 393, + 294, + 426 + ], + "spans": [ + { + "bbox": [ + 62, + 393, + 294, + 426 + ], + "type": "text", + "content": "[7] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, pages 248-255, 2009. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 427, + 294, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 427, + 294, + 491 + ], + "spans": [ + { + "bbox": [ + 62, + 427, + 294, + 491 + ], + "type": "text", + "content": "[8] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 493, + 294, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 493, + 294, + 524 + ], + "spans": [ + { + "bbox": [ + 62, + 493, + 294, + 524 + ], + "type": "text", + "content": "[9] Xuefeng Du, Zhaoning Wang, Mu Cai, and Yixuan Li. Vos: Learning what you don't know by virtual outlier synthesis. In ICLR, 2021. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 526, + 294, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 526, + 294, + 568 + ], + "spans": [ + { + "bbox": [ + 57, + 526, + 294, + 568 + ], + "type": "text", + "content": "[10] Shiv Ram Dubey, Satish Kumar Singh, and Bidyut Baran Chaudhuri. Activation functions in deep learning: A comprehensive survey and benchmark. Neurocomputing, 503: 92-108, 2022. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 571, + 294, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 571, + 294, + 613 + ], + "spans": [ + { + "bbox": [ + 57, + 571, + 294, + 613 + ], + "type": "text", + "content": "[11] Xiang Fang, Arvind Easwaran, Blaise Genest, and Ponnuthurai Nagaratnam Suganthan. Your data is not perfect: Towards cross-domain out-of-distribution detection in class-imbalanced data. ESWA, 2024. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 614, + 294, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 614, + 294, + 646 + ], + "spans": [ + { + "bbox": [ + 57, + 614, + 294, + 646 + ], + "type": "text", + "content": "[12] Xavier Glorot, Antoine Bordes, and Yoshua Bengio. Deep sparse rectifier neural networks. In AISTATS, pages 315-323, 2011. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 647, + 294, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 647, + 294, + 678 + ], + "spans": [ + { + "bbox": [ + 57, + 647, + 294, + 678 + ], + "type": "text", + "content": "[13] Ian J. Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. In ICLR, 2015. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 57, + 681, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 681, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 681, + 294, + 713 + ], + "type": "text", + "content": "[14] Song Han, Jeff Pool, John Tran, and William J Dally. Learning both weights and connections for efficient neural networks. In NeurIPS, pages 1135-1143, 2015. 8" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 37, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "text", + "content": "[15] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Delving deep into rectifiers: Surpassing human-level performance on imagenet classification. In ICCV, pages 1026-1034, 2015. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 118, + 553, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 118, + 553, + 149 + ], + "spans": [ + { + "bbox": [ + 317, + 118, + 553, + 149 + ], + "type": "text", + "content": "[16] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 151, + 553, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 151, + 553, + 192 + ], + "spans": [ + { + "bbox": [ + 317, + 151, + 553, + 192 + ], + "type": "text", + "content": "[17] Xu He, Keke Tang, Yawen Shi, Yin Li, Weilong Peng, and Peican Zhu. Are deep point cloud classifiers suffer from out-of-distribution overconfidence issue? In SMC, pages 2620-2627. IEEE, 2023. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 194, + 553, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 194, + 553, + 237 + ], + "spans": [ + { + "bbox": [ + 316, + 194, + 553, + 237 + ], + "type": "text", + "content": "[18] Matthias Hein, Maksym Andriushchenko, and Julian Bitterwolf. Why relu networks yield high-confidence predictions far away from the training data and how to mitigate the problem. In CVPR, pages 41-50, 2019. 1, 2, 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 239, + 553, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 239, + 553, + 270 + ], + "spans": [ + { + "bbox": [ + 316, + 239, + 553, + 270 + ], + "type": "text", + "content": "[19] Dan Hendrycks and Kevin Gimpel. A baseline for detecting misclassified and out-of-distribution examples in neural networks. In ICLR, 2016. 2, 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 272, + 553, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 272, + 553, + 303 + ], + "spans": [ + { + "bbox": [ + 316, + 272, + 553, + 303 + ], + "type": "text", + "content": "[20] Dan Hendrycks, Mantas Mazeika, and Thomas Dietterich. Deep anomaly detection with outlier exposure. In ICLR, 2019. 1, 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 304, + 553, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 304, + 553, + 348 + ], + "spans": [ + { + "bbox": [ + 316, + 304, + 553, + 348 + ], + "type": "text", + "content": "[21] Dan Hendrycks, Steven Basart, Mantas Mazeika, Andy Zou, Joseph Kwon, Mohammadreza Mostajabi, Jacob Steinhardt, and Dawn Song. Scaling out-of-distribution detection for real-world settings. In ICML, pages 8759-8773, 2022. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 350, + 553, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 350, + 553, + 370 + ], + "spans": [ + { + "bbox": [ + 316, + 350, + 553, + 370 + ], + "type": "text", + "content": "[22] Geoffrey Hinton. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. 2, 3, 5, 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 372, + 553, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 372, + 553, + 403 + ], + "spans": [ + { + "bbox": [ + 316, + 372, + 553, + 403 + ], + "type": "text", + "content": "[23] Gao Huang, Zhuang Liu, Laurens Van Der Maaten, and Kilian Q Weinberger. Densely connected convolutional networks. In CVPR, pages 4700-4708, 2017. 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 404, + 553, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 404, + 553, + 426 + ], + "spans": [ + { + "bbox": [ + 316, + 404, + 553, + 426 + ], + "type": "text", + "content": "[24] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 5" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 427, + 553, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 427, + 553, + 448 + ], + "spans": [ + { + "bbox": [ + 316, + 427, + 553, + 448 + ], + "type": "text", + "content": "[25] Yann LeCun, Yoshua Bengio, and Geoffrey Hinton. Deep learning. Nature, 521(7553):436-444, 2015. 1" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 449, + 553, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 449, + 553, + 480 + ], + "spans": [ + { + "bbox": [ + 316, + 449, + 553, + 480 + ], + "type": "text", + "content": "[26] Kimin Lee, Honglak Lee, Kibok Lee, and Jinwoo Shin. Training confidence-calibrated classifiers for detecting out-of-distribution samples. In ICLR, 2018. 1, 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 317, + 482, + 553, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 482, + 553, + 524 + ], + "spans": [ + { + "bbox": [ + 317, + 482, + 553, + 524 + ], + "type": "text", + "content": "[27] Kimin Lee, Kibok Lee, Honglak Lee, and Jinwoo Shin. A simple unified framework for detecting out-of-distribution samples and adversarial attacks. In NeurIPS, pages 7167-7177, 2018. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 526, + 553, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 526, + 553, + 557 + ], + "spans": [ + { + "bbox": [ + 316, + 526, + 553, + 557 + ], + "type": "text", + "content": "[28] Shiyu Liang, Yixuan Li, and R Srikant. Enhancing the reliability of out-of-distribution image detection in neural networks. In ICLR, 2018. 2, 3" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 316, + 559, + 553, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 559, + 553, + 591 + ], + "spans": [ + { + "bbox": [ + 316, + 559, + 553, + 591 + ], + "type": "text", + "content": "[29] Weitang Liu, Xiaoyun Wang, John D Owens, and Yixuan Li. Energy-based out-of-distribution detection. In NeurIPS, pages 21464-21475, 2020. 2, 5" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 316, + 593, + 553, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 593, + 553, + 623 + ], + "spans": [ + { + "bbox": [ + 316, + 593, + 553, + 623 + ], + "type": "text", + "content": "[30] Alexander Meinke and Matthias Hein. Towards neural networks that provably know when they don't know. In ICLR, 2020. 5" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 315, + 625, + 553, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 625, + 553, + 646 + ], + "spans": [ + { + "bbox": [ + 315, + 625, + 553, + 646 + ], + "type": "text", + "content": "[31] Melanie Mitchell. Complexity: A guided tour. Oxford University Press, 2009. 2" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 316, + 647, + 553, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 678 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 678 + ], + "type": "text", + "content": "[32] Rafael Müller, Simon Kornblith, and Geoffrey Hinton. When does label smoothing help? In NeurIPS, pages 4694-4703, 2019. 2" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "text", + "content": "[33] Yuval Netzer, Tao Wang, Adam Coates, Alessandro Bissacco, Bo Wu, and Andrew Y Ng. Reading digits in natural images with unsupervised feature learning. In NeurIPS" + } + ] + } + ], + "index": 36 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5038" + } + ] + } + ], + "index": 38 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 77, + 73, + 295, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 73, + 295, + 95 + ], + "spans": [ + { + "bbox": [ + 77, + 73, + 295, + 95 + ], + "type": "text", + "content": "Workshop on Deep Learning and Unsupervised Feature Learning, 2011. 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 296, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 296, + 129 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 296, + 129 + ], + "type": "text", + "content": "[34] Anh Nguyen, Jason Yosinski, and Jeff Clune. Deep neural networks are easily fooled: High confidence predictions for unrecognizable images. In CVPR, pages 427-436, 2015. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 130, + 295, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 130, + 295, + 217 + ], + "spans": [ + { + "bbox": [ + 56, + 130, + 295, + 217 + ], + "type": "text", + "content": "[35] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Köpf, Edward Yang, Zach DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. Pytorch: An imperative style, high-performance deep learning library. In NeurIPS, pages 8026-8037, 2019. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 219, + 295, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 219, + 295, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 219, + 295, + 262 + ], + "type": "text", + "content": "[36] Jiacheng Qiao, Chengzhi Zhong, Peican Zhu, and Keke Tang. Advscod: Bayesian-based out-of-distribution detection via curvature sketching and adversarial sample enrichment. Mathematics, 11(3):692, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 264, + 295, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 264, + 295, + 306 + ], + "spans": [ + { + "bbox": [ + 56, + 264, + 295, + 306 + ], + "type": "text", + "content": "[37] Adriana Romero, Nicolas Ballas, Samira Ebrahimi Kahou, Antoine Chassang, Carlo Gatta, and Yoshua Bengio. Fitnets: Hints for thin deep nets. arXiv preprint arXiv:1412.6550, 2014. 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 308, + 295, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 308, + 295, + 331 + ], + "spans": [ + { + "bbox": [ + 56, + 308, + 295, + 331 + ], + "type": "text", + "content": "[38] K Simonyan and A Zisserman. Very deep convolutional networks for large-scale image recognition. In ICLR, 2015. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 332, + 295, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 332, + 295, + 364 + ], + "spans": [ + { + "bbox": [ + 56, + 332, + 295, + 364 + ], + "type": "text", + "content": "[39] Yiyou Sun, Chuan Guo, and Yixuan Li. React: Out-of-distribution detection with rectified activations. In NeurIPS, 2021. 2, 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 365, + 295, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 365, + 295, + 397 + ], + "spans": [ + { + "bbox": [ + 56, + 365, + 295, + 397 + ], + "type": "text", + "content": "[40] Yiyou Sun, Yifei Ming, Xiaojin Zhu, and Yixuan Li. Out-of-distribution detection with deep nearest neighbors. In ICML, 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 399, + 295, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 399, + 295, + 443 + ], + "spans": [ + { + "bbox": [ + 56, + 399, + 295, + 443 + ], + "type": "text", + "content": "[41] Keke Tang, Dingruibo Miao, Weilong Peng, Jianpeng Wu, Yawen Shi, Zhaoquan Gu, Zhihong Tian, and Wenping Wang. Codes: Chamfer out-of-distribution examples against overconfidence issue. In ICCV, pages 1153-1162, 2021. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 445, + 295, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 445, + 295, + 487 + ], + "spans": [ + { + "bbox": [ + 56, + 445, + 295, + 487 + ], + "type": "text", + "content": "[42] Keke Tang, Xujian Cai, Weilong Peng, Shudong Li, and Wenping Wang. Ood attack: Generating overconfident out-of-distribution examples to fool deep neural classifiers. In ICIP, pages 1260-1264, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 489, + 295, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 489, + 295, + 532 + ], + "spans": [ + { + "bbox": [ + 56, + 489, + 295, + 532 + ], + "type": "text", + "content": "[43] Keke Tang, Xujian Cai, Weilong Peng, Daizong Liu, Peican Zhu, Pan Zhou, Zhihong Tian, and Wenping Wang. Matching words for out-of-distribution detection. In ICDM, pages 578-587, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 534, + 295, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 534, + 295, + 578 + ], + "spans": [ + { + "bbox": [ + 56, + 534, + 295, + 578 + ], + "type": "text", + "content": "[44] Keke Tang, Chao Hou, Weilong Peng, Runnan Chen, Peican Zhu, Wenping Wang, and Zhihong Tian. Cores: Convolutional response-based score for out-of-distribution detection. In CVPR, pages 10916-10925, 2024. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 579, + 295, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 579, + 295, + 623 + ], + "spans": [ + { + "bbox": [ + 56, + 579, + 295, + 623 + ], + "type": "text", + "content": "[45] Keke Tang, Wenyu Zhao, Weilong Peng, Xiang Fang, Xiaodong Cui, Peican Zhu, and Zhihong Tian. Reparaterization head for efficient multi-input networks. In ICASSP, pages 6190-6194, 2024. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 624, + 295, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 295, + 666 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 295, + 666 + ], + "type": "text", + "content": "[46] Keke Tang, Yuexin Ma, Dingruibo Miao, Peng Song, Zhaoquan Gu, Zhihong Tian, and Wenping Wang. Decision fusion networks for image classification. IEEE TNNLS, 36(3): 3890-3903, 2025. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 669, + 295, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 669, + 295, + 690 + ], + "spans": [ + { + "bbox": [ + 56, + 669, + 295, + 690 + ], + "type": "text", + "content": "[47] Leitian Tao, Xuefeng Du, Jerry Zhu, and Yixuan Li. Nonparametric outlier synthesis. In ICLR, 2022. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 56, + 692, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 692, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 692, + 295, + 713 + ], + "type": "text", + "content": "[48] Frederick Tung and Greg Mori. Similarity-preserving knowledge distillation. In ICCV, pages 1365-1374, 2019. 3" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "text", + "content": "[49] Haoqi Wang, Zhizhong Li, Litong Feng, and Wayne Zhang. Vim: Out-of-distribution with virtual-logit matching. In CVPR, pages 4921-4930, 2022. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 108, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 108, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 108, + 553, + 150 + ], + "type": "text", + "content": "[50] Xiaobo Wang, Tianyu Fu, Shengcai Liao, Shuo Wang, Zhen Lei, and Tao Mei. Exclusivity-consistency regularized knowledge distillation for face recognition. In ECCV, pages 325-342, 2020. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 152, + 553, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 553, + 195 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 553, + 195 + ], + "type": "text", + "content": "[51] Zhi-Fan Wu, Tong Wei, Jianwen Jiang, Chaojie Mao, Mingqian Tang, and Yu-Feng Li. Ngc: A unified framework for learning with open-world noisy data. In ICCV, pages 62-71, 2021. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 197, + 553, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 197, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 316, + 197, + 553, + 239 + ], + "type": "text", + "content": "[52] Jianxiong Xiao, James Hays, Krista A Ehinger, Aude Oliva, and Antonio Torralba. Sun database: Large-scale scene recognition from abbey to zoo. In CVPR, pages 3485-3492, 2010. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 242, + 553, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 242, + 553, + 275 + ], + "spans": [ + { + "bbox": [ + 316, + 242, + 553, + 275 + ], + "type": "text", + "content": "[53] Kunran Xu, Lai Rui, Yishi Li, and Lin Gu. Feature normalized knowledge distillation for image classification. In ECCV, pages 664-680, 2020. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 276, + 553, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 276, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 316, + 276, + 553, + 319 + ], + "type": "text", + "content": "[54] Pingmei Xu, Krista A Ehinger, Yinda Zhang, Adam Finkelstein, Sanjeev R Kulkarni, and Jianxiong Xiao. Turkergaze: Crowdsourcing saliency with webcam based eye tracking. arXiv preprint arXiv:1504.06755, 2015. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 321, + 553, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 321, + 553, + 364 + ], + "spans": [ + { + "bbox": [ + 316, + 321, + 553, + 364 + ], + "type": "text", + "content": "[55] Jingkang Yang, Haoqi Wang, Litong Feng, Xiaopeng Yan, Huabin Zheng, Wayne Zhang, and Ziwei Liu. Semantically coherent out-of-distribution detection. In ICCV, pages 8301-8309, 2021. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 365, + 553, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 365, + 553, + 399 + ], + "spans": [ + { + "bbox": [ + 316, + 365, + 553, + 399 + ], + "type": "text", + "content": "[56] Jingkang Yang, Kaiyang Zhou, Yixuan Li, and Ziwei Liu. Generalized out-of-distribution detection: A survey. arXiv preprint arXiv:2110.11334, 2021. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 399, + 553, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 399, + 553, + 442 + ], + "spans": [ + { + "bbox": [ + 316, + 399, + 553, + 442 + ], + "type": "text", + "content": "[57] Junho Yim, Donggyu Joo, Jihoon Bae, and Junmo Kim. A gift from knowledge distillation: Fast optimization, network minimization and transfer learning. In CVPR, pages 4133-4141, 2017. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 445, + 553, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 445, + 553, + 476 + ], + "spans": [ + { + "bbox": [ + 316, + 445, + 553, + 476 + ], + "type": "text", + "content": "[58] Shan You, Chang Xu, Chao Xu, and Dacheng Tao. Learning from multiple teacher networks. In SIGKDD, pages 1285-1294, 2017. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 478, + 553, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 478, + 553, + 521 + ], + "spans": [ + { + "bbox": [ + 316, + 478, + 553, + 521 + ], + "type": "text", + "content": "[59] Fisher Yu, Ari Seff, Yinda Zhang, Shuran Song, Thomas Funkhouser, and Jianxiong Xiao. Lsun: Construction of a large-scale image dataset using deep learning with humans in the loop. arXiv preprint arXiv:1506.03365, 2015. 5" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 523, + 553, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 523, + 553, + 555 + ], + "spans": [ + { + "bbox": [ + 316, + 523, + 553, + 555 + ], + "type": "text", + "content": "[60] Qing Yu and Kiyoharu Aizawa. Unsupervised out-of-distribution detection by maximum classifier discrepancy. In ICCV, pages 9518-9526, 2019. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 557, + 553, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 557, + 553, + 600 + ], + "spans": [ + { + "bbox": [ + 316, + 557, + 553, + 600 + ], + "type": "text", + "content": "[61] Yeonguk Yu, Sungho Shin, Seongju Lee, Changyun Jun, and Kyoobin Lee. Block selection method for using feature norm in out-of-distribution detection. In CVPR, pages 15701-15711, 2023. 5" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 601, + 553, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 601, + 553, + 644 + ], + "spans": [ + { + "bbox": [ + 316, + 601, + 553, + 644 + ], + "type": "text", + "content": "[62] Yue Yuan, Rundong He, Zhongyi Han, and Yilong Yin. Lhact: Rectifying extremely low and high activations for out-of-distribution detection. In MM, pages 8105-8113, 2023. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 647, + 553, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 667 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 667 + ], + "type": "text", + "content": "[63] Sergey Zagoruyko and Nikos Komodakis. Wide residual networks. In BMVC, 2016. 5" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "text", + "content": "[64] Sergey Zagoruyko and Nikos Komodakis. Paying more attention to attention: Improving the performance of convolutional neural networks via attention transfer. In ICLR, 2022. 3" + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5039" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 183 + ], + "type": "list", + "angle": 0, + "index": 3, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "type": "text", + "content": "[65] Bolei Zhou, Agata Lapedriza, Aditya Khosla, Aude Oliva, and Antonio Torralba. Places: A 10 million image database for scene recognition. TPAMI, 40(6):1452-1464, 2017. 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 294, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 294, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 294, + 140 + ], + "type": "text", + "content": "[66] Zhi-Hua Zhou. Why over-parameterization of deep neural networks does not overfit. Science China Information Sciences, 64(1):1-3, 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 141, + 294, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 141, + 294, + 183 + ], + "spans": [ + { + "bbox": [ + 56, + 141, + 294, + 183 + ], + "type": "text", + "content": "[67] Yao Zhu, YueFeng Chen, Chuanlong Xie, Xiaodan Li, Rong Zhang, Xiang Tian, Yaowu Chen, et al. Boosting out-of-distribution detection with typical features. In NeurIPS, 2022. 2" + } + ] + } + ], + "index": 2 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "5040" + } + ] + } + ], + "index": 4 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Simulator HC_ Regression-based Online Simulation of Starting Problem-Solution Pairs for Homotopy Continuation in Geometric Vision/d90c2114-786f-423a-b72c-4f3391aa43d5_content_list.json b/2025/Simulator HC_ Regression-based Online Simulation of Starting Problem-Solution Pairs for Homotopy Continuation in Geometric Vision/d90c2114-786f-423a-b72c-4f3391aa43d5_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..b303c3fde49d48aa3d108d4eec8df00fc632f2c5 --- /dev/null +++ b/2025/Simulator HC_ Regression-based Online Simulation of Starting Problem-Solution Pairs for Homotopy Continuation in Geometric Vision/d90c2114-786f-423a-b72c-4f3391aa43d5_content_list.json @@ -0,0 +1,1650 @@ +[ + { + "type": "text", + "text": "Simulator HC: Regression-based Online Simulation of Starting Problem-Solution Pairs for Homotopy Continuation in Geometric Vision", + "text_level": 1, + "bbox": [ + 119, + 130, + 849, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xinyue Zhang", + "bbox": [ + 186, + 202, + 310, + 222 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zijia Dai", + "bbox": [ + 362, + 202, + 444, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Wanting $\\mathbf{X}\\mathbf{u}^{1}$", + "bbox": [ + 496, + 202, + 602, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Laurent Kneip1", + "bbox": [ + 656, + 203, + 779, + 222 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Mobile Perception Lab, ShanghaiTech University, China", + "bbox": [ + 269, + 244, + 725, + 263 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 297, + 313, + 314 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "While automatically generated polynomial elimination templates have sparked great progress in the field of 3D computer vision, there remain many problems for which the degree of the constraints or the number of unknowns leads to intractability. In recent years, homotopy continuation has been introduced as a plausible alternative. However, the method currently depends on expensive parallel tracking of all possible solutions in the complex domain, or a classification network for starting problem-solution pairs trained over a limited set of real-world examples. Our innovation lies in a novel approach to finding solution-problem pairs, where we only need to predict a rough initial solution, with the corresponding problem generated by an online simulator. Subsequently, homotopy continuation is applied to track that single solution back to the original problem. We apply this elegant combination to generalized camera resection, and also introduce a new solution to the challenging generalized relative pose and scale problem. As demonstrated, the proposed method successfully compensates the raw error committed by the regressor alone, and leads to state-of-the-art efficiency and success rates.", + "bbox": [ + 76, + 330, + 472, + 650 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 662, + 209, + 678 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The solution of geometric camera calibration problems is a crucial step in many Structure-from-Motion (SfM) [21, 52, 54, 55], Visual Odometry (VO) [46], and visual Simultaneous Localization And Mapping (SLAM) [43] frameworks. Solvers developed over the years address camera resectioning [23, 29, 31, 65], two-view relative pose [30, 44, 58], pose estimation from lines [15, 41], partially calibrated cameras [64], generalized cameras [31, 41, 57, 60], rolling shutter cameras [10, 51], and many other scenarios. The problems often appear in the form of a polynomial equation system, and special techniques from the field of algebraic geometry have been used to solve them. Once an efficient solver is found, it is typically embedded into a random sampling and consensus scheme [19, 48] in order to gain robust", + "bbox": [ + 76, + 688, + 470, + 901 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ness in the presence of outliers. Hence, there is a requirement for such solvers to be efficient and embeddable into iterative schemes.", + "bbox": [ + 496, + 300, + 890, + 344 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The dominant approaches to efficiently solve systems of polynomial equations are given by the Gröbner basis method and polynomial resultants [9, 59]. Based on the Gröbner basis theory, Stewénius et al. [58] and Kukelova et al. [33] demonstrate that the calculations required to obtain a minimal set of ideal generators can be effectuated efficiently at the hand of a fixed elimination template. Kukelova et al. [33] in particular propose a solver generator to automatically discover such elimination templates. Later on, sparse resultant-based solvers [16] have been demonstrated as a powerful alternative sometimes leading to smaller elimination templates.", + "bbox": [ + 496, + 349, + 892, + 531 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "While polynomial elimination techniques have led to significant progress in the solution of geometric vision problems, the methods are limited as the complexity of the template search grows uncontrollably with the degree and dimensionality of the problem at hand. Furthermore, owing to their failure to respect inequality constraints and rule out practically infeasible solutions, high degrees or dimensionalities naturally lead to an elevated number of algebraically possible solutions and—assuming that it can be found at all—large elimination templates [13, 14, 45, 63–65]. This in turn causes problems in terms of numerical stability [63]. In recent years, a different solution strategy has therefore gained popularity: Homotopy Continuation (HC) [4, 12, 61]. The method proceeds by starting from a known problem-solution pair, and then tracks its roots to the target problem at hand during a step-wise interpolation of polynomial coefficients. The method has been used to successfully solve challenging tri-focal relative pose problems [17]. While promising, HC remains expensive as it generally requires the parallel continuation of many roots in the complex domain in order to identify all real solutions to a target problem, followed up by disambiguation.", + "bbox": [ + 496, + 534, + 893, + 867 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, Hruby et al. [25] propose an efficient, learning-based extension to the pure HC paradigm, which serves as", + "bbox": [ + 498, + 869, + 890, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "27103", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "our primary motivation. They propose the introduction of a classifier to pick a good problem-solution pair trained over a large set of known pairs obtained from SfM. This enables a highly efficient application of HC, as only a single root of interest needs to be continued. While still presenting a margin for improvement in terms of success rate, the method is demonstrated to have unprecedented computational efficiency on the challenging three-view-four-point problem.", + "bbox": [ + 75, + 90, + 472, + 212 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The method proposed in this work is similar to the work of Hruby et al. [25] in that it leverages a learning-based approach to produce a starting problem-solution pair that enables single-root-tracking. However, we note a couple of important differences:", + "bbox": [ + 76, + 219, + 470, + 297 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We do not train a classifier on a fixed set of problem-solution pairs taken from a real-world dataset. Instead, we propose to employ a solution regression network trained over an arbitrarily large set of simulated input correspondences for the considered geometric problem. The network is hence not limited by a finite set of candidate problems, and aims at being a general solution approximator for any instance of the polynomial problem. What's more, learning a solution is much easier than learning a proper problem-solution pair.", + "- The bridge to HC is formed by appending an online simulator that uses the input correspondences and the predicted solution to generate a complete, consistent problem-solution pair. Given only moderate regression accuracy, the produced problem-solution pair is sufficiently close to the original problem for successful single-root tracking via HC.", + "- We argue that the prediction of a single solution (or the unambiguous classification of a single problem-solution pair) from a minimal set may not always be possible as for complicated geometric problems, there may indeed be multiple geometrically feasible solutions. We therefore add one additional correspondence, which leads to a performance gain during both the regression and the continuation stage." + ], + "bbox": [ + 76, + 303, + 468, + 681 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The newly proposed simulate-and-solve paradigm is summarized in Figure 1. We demonstrate that this strategy can lead to high success-rate and efficient CPU-based solvers for two problems: generalized camera resection, and generalized relative pose and scale. Owing to its challenging nature, the latter problem has thus far only seen an optimization-based solution [32]. The solvers are compared against a purely learning-based alternative, a regressor followed by simple local refinement, as well as the exhaustive application of homotopy continuation over all roots. The proposed method is efficient, generalizable, and highly successful, and should thus be considered as an interesting alternative for the ongoing development of complicated geometric problem solvers.", + "bbox": [ + 75, + 689, + 468, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Literature Review", + "text_level": 1, + "bbox": [ + 500, + 90, + 678, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Polynomial elimination theory: Our contribution can be regarded as a novel alternative to traditional polynomial elimination theory [9, 59]. Stewenius et al. [58] manually derive one of the first Gröbner basis solvers in the field of computer vision. Automatic solver generators and important extensions improving performance and numerical accuracy are later introduced by Kukelova et al. [33] and Larsson et al. [34, 35, 37, 38]. Zheng and Kneip propose the use of the parametric Sylvester resultant [64], while Hesch and Roumeliotis make use of the Macaulay resultant [23]. The more general use of sparse resultants as a powerful alternative to the Gröbner basis method is analyzed by Emiris [16], Heikkilä [22], and—most recently—Bhayani et al. [5]. While having lead to significant progress, complicated geometric problems with as many as 64 [57], 272 [45], 81 [65] (before symmetry-removal), 210 [64], $>10000$ [13, 14], and 512 [63] solutions remain hard or impossible to be solved by polynomial elimination. The latter work, in particular, presents one of the largest successful elimination templates in the literature ( $96413 \\times 96879$ ) by making use of octuple machine precision. Recently Evgenity et al. [40] developed a new method to reduce the size of the elimination template. Yet, for a polynomial system, a large number of solutions in practice still leads to reduced computational efficiency.", + "bbox": [ + 496, + 114, + 893, + 492 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Homotopy continuation: Homotopy continuation implementations exist as independent programs [61] or as part of Macaulay2 [39], Bertini [4], and Julia [6]. Duff et al. [12] study the application of the method to geometric vision problems, and several related studies focus specifically on the tri-focal relative pose problem [2, 20, 24, 27, 47]. However, off-the-shelf implementations are often too general and not optimized for real-time applications. Progress is recently achieved by Fabbri et al. [17], who propose an efficient custom $\\mathrm{C + + }$ implementation of the Macaulay2 solver [39] for a tri-focal relative pose problem. More recently, Ding et al. [11] propose a fast GPU-based implementation of homotopy continuation [7, 8] for a minimal solution to the three-view-four-point problem which has the same performance as Julia [6] but is significantly faster. The method leverages massive parallelization to efficiently track all solutions of a small, fixed set of starting problem-solution pairs. Our contribution can be regarded as a problem-solution pair initialization scheme similar to Hruby et al. [25] for successful single solution tracking.", + "bbox": [ + 496, + 492, + 893, + 794 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Learning for polynomial problem solving: One of the first methods to successfully employ learning in polynomial solvers is presented by Xu et al. [62], who proposed to increase the numerical accuracy of polynomial elimination templates by learning the best permutation in permutation-invariant polynomial problems. Recently, Hruby et al. [25] propose the addition of a classifier for initial problem-", + "bbox": [ + 496, + 795, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "27104", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/b58d09d84c67abdaa48c10197923d8f9c25a0631bf30a7fa8b9ea55410f0fc1b.jpg", + "image_caption": [ + "Figure 1. Overview of the proposed geometric problem solution scheme. Given input correspondences, a regression network is utilized to approximate a solution. A subsequent online simulator generates a new set of correspondences that is consistent with the regression output. The obtained problem-solution pair is finally used to bootstrap homotopy continuation. The final solution is found efficiently by tracking a single root." + ], + "image_footnote": [], + "bbox": [ + 124, + 88, + 851, + 232 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/64d65d11bb933751be00e4ce3a349d00fa86c3601e492f430ed260a303b9fe92.jpg", + "image_caption": [ + "(a) HC multiple solution curves" + ], + "image_footnote": [], + "bbox": [ + 84, + 306, + 254, + 395 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/bdbe15f1c825887877d15210209f0c4a797e9486a1dfd5898d384d072adf9f77.jpg", + "image_caption": [ + "(b) HC solution tacking", + "Figure 2. Graphical illustration on homotopy continuation(HC) solution curves tracking. For a polynomial system with multiple solutions, each solution curve can be tracked independently as shown in Figure 2a. For each solution curve tracking, HC utilizes a \"prediction-correction\" scheme to approximate one of the solutions for the target system." + ], + "image_footnote": [], + "bbox": [ + 292, + 305, + 464, + 395 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "solution pair picking in homotopy continuation. In recent years, several works have proposed to directly apply neural networks to regress the pose of a camera, either directly from images [26, 49] or—aiming at replacing traditional algebraic-geometric solvers—from correspondence coordinates [53]. While such networks use algorithm unrolling [42] and differential optimization [3] for accurate end-to-end trainability, a successful application to more complicated problems remains yet to be demonstrated. Regressors can be regarded as general non-linear function approximators, and the challenge of using them to produce highly accurate solutions is easily explained by the potential discontinuity of root locations for continuous variations of input polynomial coefficients. Our contribution can be regarded as a new post-regression refinement scheme for polynomial problems enabling successful application to complicated problems for which it is hard to apply traditional polynomial solvers.", + "bbox": [ + 75, + 513, + 468, + 787 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Polynomial Solver: Homotopy Continuation", + "text_level": 1, + "bbox": [ + 76, + 799, + 468, + 816 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We start with a brief summary of HC. We refer to an instance of a certain polynomial system with specific coefficients as a problem. HC is a powerful tool that can numerically find the solutions of a problem. Each solution is found by tracking a curve independently, as illustrated in", + "bbox": [ + 75, + 824, + 468, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Figure 2a. When tracking the solution curves, the method solves a series of problems of similar form. It starts from a trivial problem whose solutions are simply known, and then gradually deforms that problem to the target problem that we are interested in but for which we do not know the solutions. It is hoped that—by continuing the known solutions throughout the problem sequence—they can lead to the solutions of the target, input problem. Two questions arise: 1) How to track the solution curves?, and 2) What is a good starting problem leading to the solutions we desire?", + "bbox": [ + 496, + 306, + 890, + 458 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The answer to the first question is given by HC. Formally, suppose we have a polynomial system $\\mathbf{F}(\\mathbf{x}) = \\mathbf{0}$ and we are interested in its solutions. Suppose further that we have a polynomial system of similar form $\\mathbf{G}(\\mathbf{x}) = \\mathbf{0}$ whose solutions are known. HC solves a series of polynomial systems of the form", + "bbox": [ + 496, + 458, + 890, + 549 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {H} (\\mathbf {x}, t) = (1 - t) \\mathbf {G} (\\mathbf {x}) + t \\mathbf {F} (\\mathbf {x}), \\quad t \\in [ 0, 1 ]. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 540, + 558, + 890, + 575 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As illustrated in Figure 2b, each $\\mathbf{H}(\\mathbf{x},t) = \\mathbf{0}$ is solved while sliding $t$ from 0 to 1. For $t = 0$ , the solution is known as $\\mathbf{H}(\\mathbf{x},0) = \\mathbf{G}(\\mathbf{x})$ . A solution of $\\mathbf{F}(\\mathbf{x}) = \\mathbf{0}$ is finally obtained as for $t = 1$ , $\\mathbf{H}(\\mathbf{x},1) = \\mathbf{F}(\\mathbf{x})$ . The problems are not actually solved but the solutions are simply tracked from problem to problem. Taking the partial derivative to both sides of $\\mathbf{H}(\\mathbf{x},t)$ , the solution curve $\\mathbf{x}(t)$ can be formulated as the following ordinary differential equation", + "bbox": [ + 496, + 585, + 890, + 707 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {J} _ {\\mathbf {H}} (\\mathbf {x}) \\frac {\\partial \\mathbf {x}}{\\partial t} + \\frac {\\partial \\mathbf {H} (\\mathbf {x} , t)}{\\partial t} = \\mathbf {0} \\Longleftrightarrow \\frac {\\partial \\mathbf {x}}{\\partial t} = - \\mathbf {J} _ {\\mathbf {H}} (\\mathbf {x}) ^ {\\dagger} \\frac {\\partial \\mathbf {H} (\\mathbf {x} , t)}{\\partial t}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 498, + 714, + 903, + 756 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "A \"prediction-correction\" scheme—as widely utilized in numerical analysis—is typically employed to trace the solution curve $\\mathbf{x}(t)$ .", + "bbox": [ + 496, + 757, + 890, + 801 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Prediction. Denote the stepsize of $t$ as $\\delta t$ . Euler's predictor can be used to update along the tangent direction", + "bbox": [ + 496, + 801, + 890, + 833 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {x} _ {i + 1} = \\mathbf {x} _ {i} + \\delta t \\frac {\\partial \\mathbf {x}}{\\partial t} \\bigg | _ {\\mathbf {x} _ {i}, t _ {i}}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 609, + 840, + 890, + 878 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Practically speaking, a higher-order generalized method is", + "bbox": [ + 500, + 885, + 890, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "27105", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "commonly used for better accuracy and stability with larger stepsize.", + "bbox": [ + 75, + 90, + 468, + 121 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Correction. Gauss-Newton steps are adopted to correct the prediction. In each iteration of Gauss-Newton's method, the solution is updated by", + "bbox": [ + 75, + 126, + 468, + 172 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {x} _ {j} = \\mathbf {x} _ {j - 1} - \\mathbf {J} _ {\\mathbf {H}} \\left(\\mathbf {x} _ {j - 1}\\right) ^ {\\dagger} \\mathbf {H} \\left(\\mathbf {x} _ {j - 1}, t\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 145, + 193, + 468, + 212 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{J}_{\\mathbf{H}}^{\\dagger}$ is the pseudo inverse of the jacobian $\\mathbf{J}_{\\mathbf{H}}$ , and $t = t_i + \\delta t$ . Figure 2b illustrates a 1D solution curve tracking, where the blue arrow represents the prediction update and the red arrow represents the correction update.", + "bbox": [ + 75, + 234, + 468, + 297 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The remainder of this paper—our key contribution—addresses the second question of how to find a good starting system. Traditional methods (e.g. Ab initial [17], monodromy [6, 12]) rely on algebraic geometry theories developed for coefficient-parameter homotopies, and tracking all roots from a small set of fixed, known problem-solution pairs in the complex domain may often enable a solution [56]. However, tracking all the isolated paths is expensive and in practice multiple solutions cost more time to select the best. A recent study proposes the use of a classifier to pick a good starting pair with as high as possible success probability upon tracking only a single root [25], which is much faster. Similarly, in this work, we propose a novel geometry-informed strategy that combines a neural regressor with an online simulator to produce a problem-solution pair and then track a single root by HC. Even on challenging problems for which no traditional solver has been found, it can find the solution with a success rate of over $95\\%$ .", + "bbox": [ + 75, + 301, + 468, + 571 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Remark (Homotopy Types). Since being introduced in 1980s, a variety of homotopies have been developed to serve the same purpose: how to solve the problem exactly as expected. The representative one would be coefficient-parameter homotopy. Its formulation provides the insight that if we can find all the solutions to a general coefficient-parameter polynomial, then we can find all the other members of that polynomial family. In fact, as pointed out in [56], in engineering problems (e.g. geometry vision problems) we are dealing with natural coefficient-parameter polynomials, that is coefficients of polynomials derived from visual features. Since the scope of this paper lies on a novel starting problem-solution pair finding and it is not highly dependent on the path tracking, we kindly refer the reader to [56] for an in-depth discussion of other homotopy types. As shown in the later applications, straight segment homotopies provide near-perfect success rates when working with our simulator. We also want to note that all the other homotopy types could be equally applied in conjunction with our simulator.", + "bbox": [ + 75, + 598, + 470, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Find a Good Starting Solution: Regression Network", + "text_level": 1, + "bbox": [ + 500, + 89, + 890, + 122 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Following recent developments in deep learning, multiple attempts have been made to, for example, learn camera pose using a neural network [26, 49, 53]. The first two methods extract features from images and points and use them in a regression network to predict the pose. The third method is a more preliminary approach and aims at predicting pose directly from geometric correspondences, which is also the aim of the present work. The bottleneck in regressing pose from correspondences arises from two concerns: 1) For simple geometric problems, we often have traditional methods that are efficient and accurate, and 2) For hard geometric problems, regression networks demonstrate a poor ability to predict accurate solutions. Our key insight is that while the predicted pose may not be sufficiently accurate, it can still serve as a good starting solution for HC.", + "bbox": [ + 496, + 133, + 890, + 359 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our pose regression networks are designed with multiple 1D convolutional layers and each layer is followed by batch normalization and ReLU activations. The loss function is given by a combination of relative pose errors with learnable weights [26].", + "bbox": [ + 496, + 361, + 890, + 436 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "5. Find a Consistent Starting Problem: Online Simulator", + "text_level": 1, + "bbox": [ + 498, + 450, + 890, + 484 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We recall the main idea of HC. In order to continue a solution to the target problem, an appropriate starting problem with a known solution is required. Given the approximate solution predicted by the pose regression network, our goal is to derive a consistent polynomial system for which one of the solutions matches the predicted one. We denote this step an online simulator, and—though problem specific—effective realizations can almost always be found by simple geometric rules. For example, in camera resectioning, one can simply use the input 3D points and reproject them into a camera view under the predicted pose. This will lead to new 2D points, and thus a new set of correspondences that is perfectly consistent with the predicted solution, i.e. our sought starting problem-solution pair. While not yet made use of in the present work, it is worthwhile noting that—owing to their algebraic simplicity—such an online simulator often appears to be a differentiable module.", + "bbox": [ + 496, + 494, + 890, + 752 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "6. Generalized Perspective-n-Points", + "text_level": 1, + "bbox": [ + 498, + 768, + 802, + 786 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our first concrete, practical application targets generalized camera pose estimation from 2D-3D correspondences. We use the Unified Perspective-n-Point (UPnP) algorithm [31] which formulates a polynomial system corresponding to the first-order optimality conditions of the sum of squared object space distances. The method can be used to solve both central and non-central camera resectioning problems from", + "bbox": [ + 496, + 794, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "27106", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/3a96df52960666700862ffecdc5039b230a9d1c53e40b69a24c2f285efa06d4f.jpg", + "image_caption": [ + "(a) Generalized Absolute Pose" + ], + "image_footnote": [], + "bbox": [ + 81, + 89, + 254, + 186 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/1288a391fc920f245307f7bfeb8287a9244de2ba1a51b68bec847baa32e9fdd9.jpg", + "image_caption": [ + "(b) Generalized Relative Pose and Scale", + "Figure 3. The geometry of the generalized camera problems we propose to solve by online simulator HC. $w$ represents the world frame, and $c$ and $c'$ represent camera frames in different views respectively." + ], + "image_footnote": [], + "bbox": [ + 292, + 90, + 467, + 186 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "an arbitrary number of points (including minimal). It can be efficiently solved by the Gröbner Bases method since—after variable elimination—it only involves 4 unknowns. In this section, we solve UPnP by our newly proposed method and compare it against conventional approaches, a raw regression network, as well as a parallel HC implementation tracking all roots.", + "bbox": [ + 75, + 282, + 468, + 388 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "6.1. Problem Formulation", + "text_level": 1, + "bbox": [ + 76, + 396, + 279, + 410 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Consider a 3D-2D generalized correspondence $(\\mathbf{p}_i, \\mathbf{f}_i, \\mathbf{v}_i)$ originating from the non-central camera resectioning model illustrated in Figure 3a, where $\\mathbf{p}_i \\in \\mathbb{R}^3$ , $\\mathbf{f}_i \\in \\mathbb{S}^2$ and $\\mathbf{v}_i \\in \\mathbb{R}^3$ denote the 3D world point, image ray and ray origin, respectively. The correspondence follows an underlying generalized absolute pose transformation depending on the rotation $\\mathbf{R} \\in SO(3)$ , the position $\\mathbf{t} \\in \\mathbb{R}^3$ and an unknown latent depth $\\alpha_i$ as follows:", + "bbox": [ + 75, + 417, + 468, + 540 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\alpha_ {i} \\mathbf {f} _ {i} + \\mathbf {v} _ {i} = \\mathbf {R p} _ {i} + \\mathbf {t}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 196, + 547, + 468, + 564 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Let $\\mathbf{M} \\in \\mathbb{R}^{11 \\times 11}$ denote a coefficient matrix that embeds all known variables (point measurements and ray offsets), $\\mathbf{q} = [q_0, q_1, q_2, q_3]^\\top$ denote the quaternion representation of $\\mathbf{R}$ , and $\\mathbf{s} \\in \\mathbb{R}^{11}$ denote a vector whose elements are all second and zeroth order monomials of the quaternion variables. After translation and depth elimination, $\\mathbf{q}$ can be constrained by first-order optimality conditions appearing in the form of the polynomial system1", + "bbox": [ + 75, + 571, + 468, + 691 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{ \\begin{array}{l l} \\mathbf {s} ^ {\\top} \\mathbf {M} \\frac {\\partial \\mathbf {s}}{\\partial q _ {i}} & = 0, \\quad i = 0, \\dots , 3 \\\\ \\mathbf {q} ^ {\\top} \\mathbf {q} - 1 & = 0. \\end{array} \\right. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 161, + 699, + 468, + 736 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "6.2. Regressor and online simulator", + "text_level": 1, + "bbox": [ + 76, + 736, + 352, + 750 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To provide a proper problem-solution pair for HC, we train a pose regression network from synthetic data. Its input is given by row-wise stacking of the correspondences, i.e.", + "bbox": [ + 75, + 758, + 468, + 804 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left[ \\begin{array}{c c c} \\mathbf {p} _ {1} ^ {\\top} & \\mathbf {f} _ {1} ^ {\\top} & \\mathbf {v} _ {1} ^ {\\top} \\\\ \\vdots & \\vdots & \\vdots \\\\ \\mathbf {p} _ {N} ^ {\\top} & \\mathbf {f} _ {N} ^ {\\top} & \\mathbf {v} _ {N} ^ {\\top} \\end{array} \\right], \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 210, + 809, + 468, + 869 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Algorithm 1 Online Simulator HC for generalized PnP", + "text_level": 1, + "bbox": [ + 501, + 104, + 864, + 119 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1: Input: correspondences $(\\mathbf{p}_i,\\mathbf{f}_i,\\mathbf{v}_i)_{i = 1}^M$ , stepsize $\\delta t$", + "2: Output: poses $\\hat{\\mathbf{R}},\\hat{\\mathbf{t}}$", + "3: $\\tilde{\\mathbf{R}},\\tilde{\\mathbf{t}}\\gets$ Trained Regression Network((pi,fi,v)M)", + "4: $(\\hat{\\mathbf{f}}_i)_{i=1}^M \\leftarrow$ Reprojection $(\\hat{\\mathbf{R}}, \\hat{\\mathbf{t}}; (\\mathbf{p}_i, \\mathbf{v}_i)_{i=1}^M)$ $\\triangleright$ Online Simulator", + "5: Starting problem $\\mathbf{G}(\\mathbf{x})\\gets (\\mathbf{p}_i,\\hat{\\mathbf{f}}_i,\\mathbf{v}_i)_{i = 1}^M$ with solution $\\hat{\\mathbf{R}},\\hat{\\mathbf{t}}$", + "6: Target problem $\\mathbf{F}(\\mathbf{x})\\gets (\\mathbf{p}_i,\\mathbf{f}_i,\\mathbf{v}_i)_{i = 1}^M$", + "7: $\\mathbf{q}_{HC} \\gets$ Homotopy Continuation $t\\mathbf{F}(\\mathbf{x}) + (1 - t)\\mathbf{G}(\\mathbf{x})$", + "8: $\\bar{\\mathbf{R}},\\hat{\\mathbf{t}}\\gets$ decompose $\\mathbf{q}_{HC}$" + ], + "bbox": [ + 506, + 126, + 890, + 233 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $N$ is the number of correspondences used, and the output is the quaternion vector. The network consists of 4 1D convolutional layers and a fully connected layer at the end to convert the output to a quaternion vector with the unit norm. Given the rotation predicted by the network, the translation can be recovered from correspondences and estimated rotation. We refer the reader to [31] for the details. For the online simulator, we can get reprojected $\\hat{\\mathbf{f}}_i$ from (5). Algorithm 1 summarizes all stages, including the final HC over (6). Note that in line 5 of Algorithm 1, the starting problem $\\mathbf{G}(\\mathbf{x})$ is constructed by replacing the coefficient matrix $\\mathbf{M}$ computed by $(\\mathbf{p}_i,\\mathbf{f}_i,\\mathbf{v}_i)$ in (6) with $\\hat{\\mathbf{M}}$ computed by $(\\mathbf{p}_i,\\hat{\\mathbf{f}}_i,\\mathbf{v}_i)$ .", + "bbox": [ + 496, + 255, + 892, + 455 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "6.3. Experiments", + "text_level": 1, + "bbox": [ + 500, + 464, + 633, + 479 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Data. We randomly generate 4 cameras in a cube $[-1,1]^3$ and randomly sample 16 3D points in the box $[-1,1]^2 \\times [4,8]$ . 16 samples would not be a good choice for a RANSAC solver [18], but in the present context the idea is to set a moderate and notably equal challenge to all compared solvers. Furthermore, for UPnP the number of constraints appearing in the polynomial system remains unchanged as it analyses first-order optimality rather than individual incidence relations. Elements of the rotation angles are randomly generated in the interval $[- \\pi /2, \\pi /2]$ , respectively, and the translation is randomly generated in the unit cube. We then generate the correspondences from rays that point to a 3D point. The noise level is set to 2 pixels. To train the pose regression network, we simulate a dataset of 2400 samples, which is sufficient for reliable Simulator HC initialization. We use a 2.70GHz Intel(R) Xeon(R) CPU and an NVIDIA GeForce RTX 4090 for network training and evaluation.", + "bbox": [ + 496, + 487, + 890, + 758 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Methods. For HC, the stepsize $\\delta t$ and the maximum number of iterations of Newton's method are set to 0.02 and 5, respectively. We compare the Gröbner Bases solver [31] and the popular local method Levenberg-Marquardt (LM) [1]. We also test the state-of-the-art parallel HC solver developed by [7, 8], and use both its CPU and $\\mathrm{GPU}^2$ version. Given that parallel HC tracks all solutions,", + "bbox": [ + 496, + 758, + 890, + 864 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "1We kindly refer the reader to the UPnP paper for detailed derivations [31].", + "bbox": [ + 75, + 875, + 468, + 898 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "2Note that GPU warmup overhead is ignored when averaging over multiple runs.", + "bbox": [ + 500, + 875, + 888, + 898 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "27107", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/60782c54be04b24b9ebfd90ca87518a28bd2b337f6cc66f93d7ad26d154fa750.jpg", + "table_caption": [ + "Table 1. Performance comparison of simulator HC, a traditional method, and parallel exhaustive tracking HC on noisy synthetic data. All results are averaged over 1000 trials. The success rate is computed for rotation errors with a threshold of 2 degree." + ], + "table_footnote": [], + "table_body": "
MethodSucc. (%)ER(deg)Et(%)Time(ms)
Gröbner Bases1000.0310.4891.254
CPU HC1000.0320.4999.796
GPU HC1000.0320.4991.002
Pred Initial327.204320.031
Pred Initial + LM5428.7222560.528
Rand Initial + Simulator HC2291.682280.231
Pred Initial + Simulator HC1000.0320.5130.316
", + "bbox": [ + 76, + 156, + 475, + 268 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "we manually pick the one with the smallest error.", + "bbox": [ + 76, + 280, + 403, + 294 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Results. The rotation error is measured by $\\mathcal{E}_{\\mathbf{R}} = 2\\arcsin (\\| \\hat{\\mathbf{R}} -\\mathbf{R}_{gt}\\| /(2\\sqrt{2}))$ in degrees as suggested by [11] and the translation error is measured by the $\\mathcal{E}_{\\mathrm{t}} = \\| \\hat{\\mathbf{t}} -\\mathbf{t}_{gt}\\| /\\| \\mathbf{t}_{gt}\\| \\times 100$ in percentage. Table 1 compares the traditional method, parallel exhaustive tracking HC, and different types of initial solutions for our simulator HC. We present the success rate as the percentage of rotation errors smaller than 2 degrees. It can be observed that on this simple task, except for the local method, traditional methods, as well as simulator HC using predicted initials, succeed in $100\\%$ of trials. Meanwhile, online simulator HC benefits from the single-track strategy and turns out to be the fastest one. We also perform an ablation study on the proposed simulator HC, using both predicted and random initializations. Although the pose learned by the regressor is not highly accurate and only slightly outperforms random poses, it serves as a reasonable starting solution for our simulator HC. In contrast, LM as a local method is unable to leverage the potential of the predicted initializations, as the simulator HC does.", + "bbox": [ + 76, + 296, + 470, + 599 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "7. Generalized Relative Pose and Scale", + "text_level": 1, + "bbox": [ + 76, + 617, + 403, + 632 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our final application targets a hard minimal problem: Generalized Relative Pose and Scale (GRPS) estimation from 2D-2D correspondences [32]. The GRPS problem is a natural extension of the generalized relative pose problem and arises in point-less registration of partial view-graphs in structure-from-motion. Owing to the fact that SfM results are generally up to scale, an inherent scale factor needs to be resolved. Such problem was first introduced by Kneip et al. [32]. It has 140 solutions if a unique rotation representation is used, and can be solved by multi-start optimization to overcome local minima. Later on, a traditional polynomial solver based on the Gröbner Bases method has been proposed [36, 40]. With a $144 \\times 284$ elimination template, the solver finds all 140 rotations within reasonable time. In this section, we solve GRPS by our proposed method and compare it against the existing approach, as well as a parallel HC implementation tracking all roots.", + "bbox": [ + 75, + 643, + 468, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "7.1. Problem Formulation", + "text_level": 1, + "bbox": [ + 500, + 90, + 702, + 104 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Consider a 2D-2D generalized correspondence $(\\mathbf{f}_i,\\mathbf{v}_i) - (\\mathbf{f}_i',\\mathbf{v}_i')$ observed from two non-central camera views as illustrated in Figure 3b. $\\mathbf{f}_i\\in \\mathbb{S}^2,\\mathbf{v}_i\\in \\mathbb{R}^3$ denote the image ray and ray origin in the first view, and $\\mathbf{f}_i'\\in \\mathbb{S}^2,\\mathbf{v}_i'\\in \\mathbb{R}^3$ are the corresponding entities in the second view. The correspondence follows an underlying generalized relative pose and scale incidence relation depending on the rotation $\\mathbf{R}\\in SO(3)$ , the position $\\mathbf{t}\\in \\mathbb{R}^3$ , the scale $s$ , and unknown latent depths $\\alpha_{i}$ and $\\alpha_{i}'$ as follows:", + "bbox": [ + 496, + 112, + 890, + 250 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\alpha_ {i} \\mathbf {f} _ {i} + \\mathbf {v} _ {i} = \\mathbf {R} \\left(\\alpha_ {i} ^ {\\prime} \\mathbf {f} _ {i} ^ {\\prime} + s \\mathbf {v} _ {i} ^ {\\prime}\\right) + \\mathbf {t}. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 583, + 260, + 890, + 277 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "With the help of Plücker coordinates and the scalar triple product $^3$ , the unknown depths get eliminated which results in a polynomial constraint that only involves the unknown poses", + "bbox": [ + 496, + 287, + 890, + 348 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {t} ^ {\\top} \\left(\\mathbf {f} _ {i} \\times \\mathbf {R} \\mathbf {f} _ {i} ^ {\\prime}\\right) - s \\mathbf {f} _ {i} ^ {\\top} \\mathbf {R} \\left[ \\mathbf {v} _ {i} ^ {\\prime} \\right] _ {\\times} \\mathbf {f} _ {i} ^ {\\prime} + \\mathbf {f} _ {i} ^ {\\top} \\left[ \\mathbf {v} _ {i} \\right] _ {\\times} \\mathbf {R} \\mathbf {f} _ {i} ^ {\\prime} = 0. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 517, + 357, + 890, + 377 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Therefore, a minimal formulation for the GRPS problem can be devised by a polynomial system formed by stacking 7 constraints in the form (9), each one depending on one correspondence.", + "bbox": [ + 496, + 387, + 890, + 446 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "7.2. Regressor and online simulator", + "text_level": 1, + "bbox": [ + 500, + 455, + 774, + 472 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As introduced in section 6.2, we continue to use a pose regression network that inputs a matrix with row-wise stacked correspondences. The output is given by $\\mathbf{R}$ , $\\mathbf{t}$ , $s$ . Note that the rotation is predicted in continuous 6D representation which is more suitable for neural networks [66]. The network consists of 11 1D convolutional layers, in the end, it branches to 3 sub-branches for $\\mathbf{R}$ , $\\mathbf{t}$ and $s$ . Each sub-branch consists of 2 1D convolutional layers and 1 fully connected layers for adjust output. Given the poses predicted by the network, we still need an estimation of the depths before simulating the reprojected image ray. With estimated $\\hat{\\mathbf{R}}$ , $\\hat{\\mathbf{t}}$ and $\\hat{s}$ , the corresponding depths can be obtained from (8) by following", + "bbox": [ + 496, + 479, + 890, + 676 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\left[ \\mathbf {f} _ {i} - \\hat {\\mathbf {R}} \\mathbf {f} _ {i} ^ {\\prime} \\right] \\left[ \\begin{array}{l} \\alpha_ {i} \\\\ \\alpha_ {i} ^ {\\prime} \\end{array} \\right] = - \\left(\\mathbf {v} _ {i} - \\hat {s} \\hat {\\mathbf {R}} \\mathbf {v} _ {i} ^ {\\prime} - \\hat {\\mathbf {t}}\\right), \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 560, + 686, + 890, + 720 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "which can be simply solved by linear least squares for each input correspondence. Finally, we simulate a pose-consistent reprojected image ray in the first view using", + "bbox": [ + 496, + 729, + 890, + 773 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {f}} _ {i} = \\frac {1}{\\hat {\\alpha} _ {i}} \\hat {\\mathbf {R}} \\left(\\hat {\\alpha} _ {i} ^ {\\prime} \\mathbf {f} _ {i} ^ {\\prime} + \\hat {s} \\mathbf {v} _ {i} ^ {\\prime}\\right) + \\hat {\\mathbf {t}} - \\mathbf {v} _ {i}. \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 578, + 784, + 890, + 814 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "7.3. Simulation Experiments", + "text_level": 1, + "bbox": [ + 500, + 824, + 723, + 840 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Data. We adopt a similar data generation method as in section 6.3. We generate 7 3D points in a $[-1,1]^2 \\times [2,20]$", + "bbox": [ + 496, + 847, + 890, + 878 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "3We kindly refer the reader to [28, 32] for detailed derivations", + "bbox": [ + 514, + 886, + 846, + 898 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "27108", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "volume. Two view frames are generated with origins picked from $[-1, 1]^3$ . Within each view frame, we then sample individual camera offsets from the volume $[-1, 1]^3$ . The scale is randomly generated in $[0.1, 5.0]$ , and all other parameters related to noise addition are left unchanged. Note that the ground truth relative pose needs to be extracted from the simulated pair of view poses.", + "bbox": [ + 75, + 90, + 468, + 196 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Methods. The stepsize for homotopy continuation is set to 0.05 in this scenario. We compare against the Eigen solver [32], Gröbner Bases solver [40] and the CPU and GPU versions of parallel HC which tracks 140 solutions found by Monodromy [6]. Note that to measure numerical errors, we just pick the best solution with the smallest error. In practice, traditional HC needs to evaluate additional information to select a single solution (e.g. euclidean distance). For a complicated problem like GRPS, evaluating 140 solutions would still induce non-neglectable time-consumption. Since GRPS is a hard problem, we train the regressor using a larger simulated dataset of size 64000.", + "bbox": [ + 75, + 196, + 470, + 378 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Metric. We measure the rotation error in degrees and the relative translation as in Sec. 6.3 and scale error in percentage using", + "bbox": [ + 75, + 378, + 468, + 422 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {E} _ {s} = \\left| \\hat {s} - s _ {g t} \\right| / s _ {g t} \\times 1 0 0. \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 183, + 422, + 468, + 440 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To calculate the success rate, for rotation errors the threshold is set to 2 deg, and for relative errors, it is set to $5\\%$ .", + "bbox": [ + 75, + 446, + 468, + 477 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Numerical Stability. We randomly generate 1000 trials to test the numerical stability of all solvers. During the analysis of the performance of simulator HC using the minimum number of correspondences, we observed that our method has around $70\\%$ success rate. By checking the failure cases, we found that sometimes simulator HC tracked another real solution that was different from the ground truth pose. To tackle that issue, we evaluate the addition of one more correspondence and solve for an overdetermined polynomial system. By using 8 correspondences, simulator HC reaches a success rate of $95\\%$ . Again, since CPU/GPU-HC is designed to handle square systems, only the minimum number of correspondences can be tested for other HC methods.", + "bbox": [ + 75, + 478, + 468, + 672 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Since the errors in rotation, translation and scale have similar density results, we only present the rotation error in the main paper and the rest in the supplementary. As illustrated in Figure 4a, the minimal HC solvers all present similar error distribution and around $75\\%$ of trials are successful. While Gröbner Bases shows $100\\%$ success rate, Simulator HC over 8pts presents similar numerical errors.", + "bbox": [ + 75, + 674, + 468, + 779 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Time Comparison. The main computational benefit of the proposed online simulator HC comes from single solution-tracking, which can be effectuated efficiently even on the CPU. Figure 4b demonstrates the efficiency gain of our 8pts simulator HC over the eigen solver and parallel HC variants. As expected, by tracking only one solution curve, we are significantly faster than parallel exhaustive tracking HC, even if the latter runs on GPU. The eigen solver", + "bbox": [ + 75, + 779, + 470, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "is an optimization-based solver with high complexity and ends up being slowest. The Gröbner Bases solver dealing with the large elimination template still costs a considerable time, even for just computing solutions for rotation, and not translation or scale.", + "bbox": [ + 496, + 90, + 890, + 167 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Remark. We would like to note that Gröbner Bases is solving a different formulation from Eq. (9). In order to make it easy to be solved by a smaller elimination template, it eliminates translation and scale [57]. But the resulting coefficient matrix is $130 \\times 275$ and still needs a large number of operations. Although the formulation of Eq. (9) with eliminated depths faces a slightly lower success rate, its number of operations is much smaller than Gröbner Bases. On the other hand, the proposed 8pt simulator HC provides a comparably high success rate as Gröbner Bases and is still $17 \\times$ faster.", + "bbox": [ + 496, + 172, + 893, + 339 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. RANSAC experiments for the GRPS problem using Gröbner Bases | simulator HC. Image points are randomly perturbed by 2-pixel uniform noise. We use vanilla RANSAC with at most 200 iterations. As can be observed, outlier ratios of up to $40\\%$ are easily handled within 200 iterations. And regressor-based simulator HC is consistently faster than Gröbner Bases.", + "bbox": [ + 496, + 356, + 890, + 441 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/3e6c19a93dc0dd4bb330eaea0303c84eabae815caffb52385d802105e89b0a0e.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
RatioSucc. (%)ER(deg)# ItersTime(s)
10%99.51000.100.09931252.330.85
20%99.41000.110.101711934.371.31
30%99.41000.130.142002005.101.35
40%98.9910.210.292002004.591.35
", + "bbox": [ + 527, + 452, + 861, + 520 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Noise Resilience. Figure 4c presents the performance of the proposed 8pts simulator HC with different numbers of cameras under increasing image noise levels ranging from 0.2 pixel to 1.0 pixel. The error behaves gracefully as a function of noise, and it can be observed that a higher number of cameras leads to slightly inferior noise. Furthermore, we also observe that rotation errors appear to be less sensitive to image noise than translation and scale errors.", + "bbox": [ + 496, + 539, + 890, + 659 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Evaluation inside RANSAC. To conclude, we embed simulator HC into RANSAC to test its practical usefulness. We synthesize 200 correspondences between 5 cameras and replace part of them with outliers. Uniformly sampled pixel noise from $[-2,2]$ is added to both coordinates of the image points by again assuming a virtual focal length of 800 px. We terminate RANSAC if it reaches $99\\%$ confidence or 200 iterations, and the inlier threshold is set to 0.01 for the distance measured by the normalized reprojection error. Given that Gröbner Bases solver estimates 140 solutions, we keep the real ones and select the best solution using reprojection errors in parallel to save time for it. Table 2 shows the results. The success rate is measured by a rotation error threshold of 2 degrees and the indicated errors are", + "bbox": [ + 496, + 661, + 890, + 872 + ], + "page_idx": 6 + }, + { + "type": "page_footnote", + "text": "${}^{4}$ shown in supplementary.", + "bbox": [ + 517, + 886, + 656, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "27109", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/45ed5eb77a4209965faf887b0475a13238408c8420f2cc0948dd2bceb9812326.jpg", + "image_caption": [ + "(a) Error distribution over noise-free data." + ], + "image_footnote": [], + "bbox": [ + 81, + 92, + 339, + 227 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a77d11a9c7f3b97eb41fb553e927e3e14c55e7b36f9afddcdc3d7665d63ae342.jpg", + "image_caption": [ + "(b) The boxplot of running time comparison." + ], + "image_footnote": [], + "bbox": [ + 348, + 92, + 620, + 228 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6c5e8c963e11700af2d2282864c9fd2eb43e1d380071f90354aed29cbf794556.jpg", + "image_caption": [ + "(c) Error statistics of simulator HC.", + "Figure 4. Experimental results on GRPS. Fig. 4a Error distribution over 1000 trials on noise-free data. The camera number is set to be 3. Considering the rotation error, except for Gröbner Bases, the other minimal solvers all present a similar error distribution with around $70\\%$ success rate, and the eigen solver presents only around $60\\%$ success rate. The proposed 8pts simulator HC presents $96.3\\%$ success rate where the Gröbner Bases has $100\\%$ success rate. Fig. 4b The boxplot of running time comparison for the proposed simulator HC and other methods, where the number represents the median time. As expected, on average the proposed online simulator HC running on CPU is the fastest. It is about $5\\times$ faster than GPU-HC since we only track one solution, $17\\times$ faster than Gröbner Bases and about $40\\times$ faster than CPU-HC. Fig. 4c Error statistics of simulator HC with respect to different noise levels ranging from 0.2 pixel to 1.0 pixel." + ], + "image_footnote": [], + "bbox": [ + 630, + 90, + 890, + 229 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "the average accuracy of successful cases, results on translation and scale are left in supplementary. Overall, simulator HC embedded into RANSAC can easily handle up to $40\\%$ outliers as Gröbner Bases and much faster, which suggests high practicality. Note that Gröbner Bases provides around 46 real solutions each time, which requires a significant amount of additional time for solution selection. Meanwhile, simulator HC does not require this extra step.", + "bbox": [ + 75, + 368, + 468, + 489 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7.4. Real World Experiments", + "text_level": 1, + "bbox": [ + 76, + 502, + 303, + 518 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We further explore the performance of simulator HC on the real dataset LaMAR [50]. As illustrated in Fig. 5, the provided HoloLens data is a natural application of GRPS with identity scale, which can be viewed as a multi-camera rig with 4 cameras in the front and sides. We extract the correspondences by consecutively taking a view-pair from the trajectories and ensuring the distance between each view is in $[0.2m, 0.5m]$ with at least 100 matches. The", + "bbox": [ + 75, + 527, + 468, + 650 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/18ecec1634171e8a7d5ba99f673cc4d24c41f8c6ed96af9e97ed16933d20d1e3.jpg", + "image_caption": [ + "Figure 5. LaMAR dataset multi-camera rig visualization." + ], + "image_footnote": [], + "bbox": [ + 81, + 674, + 468, + 729 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/ec7a3fd451ed54ad5f193d56b3538ff7bc13023c8351467efcde27eefb586c9f.jpg", + "table_caption": [ + "Table 3. LaMAR dataset RANSAC-simulatorHC Median results." + ], + "table_footnote": [], + "table_body": "
seq.ER(deg)Es(%)Et(deg)#itersTime(s)Succ. (%)
Gröbner BasesCAB1.1856.145.701372.8360
HGE0.6262.484.39591.2567
LIN0.6042.043.40310.6572
simulator HCCAB1.0543.209.163102.0868
HGE0.5641.527.191340.9084
LIN0.5632.194.32660.4486
", + "bbox": [ + 78, + 796, + 485, + 896 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "resulting number of pairs is 1794 (CAB), 2852 (HGE) and 739 (LIN) in total. The correspondences are matched with SIFT features using nearest neighbour matching. We also present the angular error in translation, as in practice, the scale of translation and overall scale could be unobservable, e.g. the degenerate scenario when correspondences remain on the frame captured by the same camera.", + "bbox": [ + 496, + 368, + 890, + 474 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3 shows the median results over all three sequences of LaMAR dataset. We adopted a similar setting for RANSAC as the above section with 500 maximum iterations. The proposed solver still presents consistent performances, where simulator HC has higher success rate with faster running time than Gröbner Bases.", + "bbox": [ + 496, + 491, + 890, + 582 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "8. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 601, + 617, + 617 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The present paper introduces a new paradigm for correspondence-based geometric problem solving using an elegant three-stage combination of a regression network, an online correspondence simulator, and homotopy continuation. The regression network acts as a general solution approximator and can be trained in simulation only. While pure regression accuracy is often insufficient on its own, a key message conveyed through our results is that the accuracy is generally still good enough to simulate a consistent starting problem-solution pair that successfully enables single-solution continuation. As demonstrated on two generalized camera problems, the strategy enables high success rates on common computing hardware, and the resulting solvers prove to be a viable solution if utilized as part of a random sample and consensus scheme. In particular, we propose a successful solution to the generalized relative pose and scale problem, which is around $17 \\times$ faster than traditional polynomial elimination techniques.", + "bbox": [ + 496, + 628, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "27110", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "9. Acknowledgments", + "text_level": 1, + "bbox": [ + 76, + 90, + 256, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The authors would like to acknowledge the funding support provided by projects 22DZ1201900, 22ZR1441300, and DFYJBJ-1 by the Natural Science Foundation of Shanghai.", + "bbox": [ + 76, + 114, + 468, + 161 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 174, + 174, + 189 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Sameer Agarwal, Keir Mierle, and Others. Ceres solver. http://ceres-solver.org, 2022.5", + "[2] C. Aholt and L. Oeding. The ideal of the trifocal variety. Math. Comput., 83(289):2553-2574, 2014. 2", + "[3] B. Amos and J. Z. Kolter. Optnet: Differentiable optimization as a layer in neural networks. In International Conference on Machine Learning, pages 136-145. PMLR, 2017. 3", + "[4] D. J. Bates, A. J. Sommese, J. D. Hauenstein, and C. W. Wampler. Numerically solving polynomial systems with Bertini. Software, environments, tools, SIAM, 25, 2013. 1, 2", + "[5] S. Bhayani, Z. Kukelova, and J. Heikkilä. A sparse resultant based method for efficient minimal solvers. In CVPR, 2020. 2", + "[6] P. Breiding and S. Timme. Homotopycontinuation.jl: A package for homotopy continuation in Julia. In International Conference on Mathematical Software - ICMS, 2018. 2, 4, 7", + "[7] C.-H. Chien, H. Fan, A. Abdelfattah, E. Tsigaridas, S. Tomov, and B. Kimia. GPU-based homotopy continuation for minimal problems in computer vision. In CVPR, 2022. 2, 5", + "[8] C.-H. Chien, H. Fan, A. Abdelfattah, E. Tsigaridas, S. Tomov, and B. Kimia. Parallel path tracking for homotopy continuation usinggpu. In Proceedings of the International Symposium on Symbolic and Algebraic Computation, 2022. 2, 5", + "[9] D. Cox, J. Little, and D. O'Shea. Using Algebraic Geometry. Springer, 1998. 1, 2", + "[10] Y. Dai, H. Li, and L. Kneip. Rolling shutter camera relative pose: Generalized epipolar geometry. In CVPR, 2016. 1", + "[11] Y. Ding, C.-H. Chien, V. Larsson, K. Åström, and B. Kimia. Minimal solutions to generalized three-view relative pose problem. In ICCV, 2023. 2, 6", + "[12] T. Duff, C. Hill, A. Jensen, K. Lee, A. Leykin, and J. Sommars. Solving polynomial systems via homotopy continuation and monodromy. IMA Journal of Numerical Analysis, 2018. 1, 2, 4", + "[13] T. Duff, K. Kohn, A. Leykin, and T. Pajdla. PLMP - point-line minimal problems in complete multi-view visibility. In ICCV, pages 1675-1684, 2019. 1, 2", + "[14] T. Duff, K. Kohn, A. Leykin, and T. Pajdla. PL1P - point-line minimal problems under partial visibility in three views. In ECCV, 2020. 1, 2", + "[15] A. Elquresh and A. M. Elgammal. Line-based relative pose estimation. In CVPR, 2011. 1", + "[16] I. Z. Emiris. A general solver based on sparse resultants, 2012. 1, 2", + "[17] R. Fabbri, T. Duff, H. Fan, M. H. Regan, D. da Costa de Pinho, E. P. Tsigaridas, C. W. Wampler, J. D. Hauenstein, P. J. Giblin, A. Kimia, B. B. Leykin, and T. Pajdla. TR-PLP" + ], + "bbox": [ + 78, + 199, + 467, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "- Trifocal relative pose from lines at points. In CVPR, 2020. 1, 2, 4", + "[18] Martin A Fischler and Robert C Bolles. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM, 24(6):381-395, 1981. 5", + "[19] M. A. Fischler and R. C. Bolles. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Commun. ACM, 24(6): 381-395, 1985. 1", + "[20] R. M. Haralick, C.-N. Lee, K. Ottenberg, and M. Nölle. Review and analysis of solutions of the three point perspective pose estimation problem. *IJCV*, 13(3):331--356, 1994. 2", + "[21] R. Hartley and A. Zisserman. Multiple View Geometry in Computer Vision. Cambridge University Press, 2nd edition, 2004. 1", + "[22] J. Heikkilä. Using sparse elimination for solving minimal problems in computer vision. In ICCV, 2017. 2", + "[23] J. Hesch and S. I. Roumeliotis. A Direct Least-Squares (DLS) method for PnP. In ICCV, 2011. 1, 2", + "[24] R. J. Holt and A. N. Netravali. Uniqueness of solutions to three perspective views of four points. IEEE TPAMI, 17(3): 303--307, 1995. 2", + "[25] P. Hruby, T. Duff, A. Leykin, and T. Pajdla. Learning to solve hard minimal problems. In CVPR, 2022. 1, 2, 4", + "[26] A. Kendall and R. Cipolla. Geometric loss functions for camera pose regression with deep learning. In CVPR, pages 5974-5983, 2017. 3, 4", + "[27] J. Kileel. Minimal problems for the calibrated trifocal variety. SIAM Journal on Applied Algebra and Geometry, 1(1): 575-598, 2017. 2", + "[28] Laurent Kneip and Hongdong Li. Efficient computation of relative pose for multi-camera systems. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 446-453, 2014. 6", + "[29] L. Kneip, D. Scaramuzza, and R. Siegwart. A novel parametrization of the perspective-three-point problem for a direct computation of absolute camera position and orientation. In CVPR, 2011. 1", + "[30] L. Kneip, R. Siegwart, and M. Pollefeys. Finding the exact rotation between two images independently of the translation. In ECCV, 2012. 1", + "[31] L. Kneip, H. Li, and Y. Seo. UPnP: An optimal O(n) solution to the absolute pose problem with universal applicability. In ECCV, 2014. 1, 4, 5", + "[32] L. Kneip, C. Sweeney, and R. Hartley. The generalized relative pose and scale problem: View-graph fusion via 2d-2d registration. 2016. 2, 6, 7", + "[33] Z. Kukelova, M. Bujnak, and T. Pajdla. Automatic generator of minimal problem solvers. In ECCV, 2008. 1, 2", + "[34] V. Larsson, K. Aström, and M. Oskarsson. Efficient solvers for minimal problems by syzygy-based reduction. In CVPR, 2017. 2", + "[35] V. Larsson, K. Aström, and M. Oskarsson. Polynomial solvers for saturated ideals. In ICCV, 2017. 2" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "27111", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[36] Viktor Larsson, Kalle Astrom, and Magnus Oskarsson. Efficient solvers for minimal problems by syzygy-based reduction. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 820-829, 2017. 6", + "[37] V. Larsson, Z. Kukelova, and Y. Zheng. Making minimal solvers for absolute pose estimation compact and robust. In ICCV, 2017. 2", + "[38] V. Larsson, M. Oskarsson, K. Aström, A. Wallis, Z. Kukelova, and T. Pajdla. Beyond gröbner bases: Basis selection for minimal solvers. In CVPR, 2018. 2", + "[39] A. Leykin. Numerical algebraic geometry. Journal of Software for Algebra and Geometry, 3(1):5-10, 2011. 2", + "[40] Evgeniy Martyushev, Jana Vrablikova, and Tomas Pajdla. Optimizing elimination templates by greedy parameter search. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15754-15764, 2022. 2, 6, 7", + "[41] P. Miraldo, T. Dias, and S. Ramalingam. A minimal closed-form solution for multi-perspective pose estimation using points and lines. In ECCV, 2018. 1", + "[42] V. Monga, Y. Li, and Y. C. Eldar. Algorithm unrolling: Interpretable, efficient deep learning for signal and image processing. IEEE Signal Processing Magazine, 38(2):18-44, 2021. 3", + "[43] R. Mur-Artal, J. M. M. Montiel, and Tardós. J. D. ORB-SLAM: A Versatile and Accurate Monocular SLAM System. 31(5), 2015. 1", + "[44] D. Niste'r. An efficient solution to the five-point relative pose problem. IEEE TPAMI, 26(6):756-770, 2004. 1", + "[45] D. Niste'r and F. Schaffalitzky. Four points in two or three calibrated views: Theory and practice. *IJCV*, 67(2):211–231, 2006. 1, 2", + "[46] D. Nistér, O. Naroditsky, and J. Bergen. Visual odometry. In CVPR, 2004. 1", + "[47] L. Quan, B. Triggs, and B. Mourrain. Some results on minimal euclidean reconstruction from four points. 24(3):341-348, 2006. 2", + "[48] R. Raguram, O. Chum, M. Pollefeys, J. Matas, and J.-M. Frahm. SAC: A universal framework for random sample consensus. IEEE TPAMI, 35(8):2022-2038, 2013. 1", + "[49] P.-E. Sarlin, A. Unagar, M. Larsson, H. Germain, C. Toft, V. Larsson, M. Pollefeys, V. Lepetit, L. Hammarstrand, F. Kahl, et al. Back to the feature: Learning robust camera localization from pixels to pose. In CVPR, 2021. 3, 4", + "[50] Paul-Edouard Sarlin, Mihai Dusmanu, Johannes L. Schonberger, Pablo Speciale, Lukas Gruber, Viktor Larsson, Ondrej Miksik, and Marc Pollefeys. LaMAR: Benchmarking Localization and Mapping for Augmented Reality. In ECCV, 2022. 8", + "[51] O. Saurer, M. Pollefeys, and G. H. Lee. A minimal solution to the rolling shutter pose estimation problem. 2015. 1", + "[52] J. Schonberger and J.-M. Frahm. Structure-from-motion revisited. In CVPR, 2016. 1", + "[53] R. Sheffer and A. Wiesel. PnP-Net: A hybrid Perspective-n-Point Network, 2020. 3, 4", + "[54] N. Snavely, S. M. Seitz, and R. Szeliski. Photo tourism: exploring photo collections in 3D. 2006. 1" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[55] N. Snavely, S. M. Seitz, and R. Szeliski. Modeling the world from internet photo collections. pages 189-210, 2008. 1", + "[56] Andrew J Sommese, Charles W Wampler, et al. The Numerical solution of systems of polynomials arising in engineering and science. World Scientific, 2005. 4", + "[57] H. Stewenius, D. Nister, M. Oskarsson, and K. Aström. Solutions to minimal generalized relative pose problems. In Workshop on Omnidirectional Vision in conjunction with ICCV, 2005. 1, 2, 7", + "[58] H. Stewenius, C. Engels, and D. Niste'r. Recent developments on direct relative orientation. ISPRS J. of Photogrammetry and Remote Sensing, 60:284-294, 2006. 1, 2", + "[59] B. Sturmfels. Solving systems of polynomial equations, 2002. 1, 2", + "[60] J. Ventura, C. Arth, and V. Lepetit. An efficient minimal solution for multi-camera motion. In ICCV, 2015. 1", + "[61] J. Verschelde. Polynomial homotopy continuation with phc-pack. ACM Commun. Comput. Algebra, 44(3/4):217—220, 2010. 1, 2", + "[62] W. Xu, H. Lan, M. C. Tsakiris, and L. Kneip. Online stability improvement of gröbner basis solvers using deep learning. 2019. 2", + "[63] J. Zhao, L. Kneip, Y. He, and J. Ma. Minimal case relative pose computation using ray-point-ray features. IEEE TPAMI, 42:1176-1190, 2020. 1, 2", + "[64] Y. Zheng and L. Kneip. A direct least-squares solution to the PnP problem with unknown focal length. In CVPR, 2016. 1, 2", + "[65] Y. Zheng, Y. Kuang, S. Sugimoto, K. Astrom, and M. Okutomi. Revisiting the pnp problem: A fast, general and optimal solution. In ICCV, 2013. 1, 2", + "[66] Yi Zhou, Connelly Barnes, Jingwan Lu, Jimei Yang, and Hao Li. On the continuity of rotation representations in neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5745-5753, 2019. 6" + ], + "bbox": [ + 501, + 90, + 890, + 603 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "27112", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Simulator HC_ Regression-based Online Simulation of Starting Problem-Solution Pairs for Homotopy Continuation in Geometric Vision/d90c2114-786f-423a-b72c-4f3391aa43d5_model.json b/2025/Simulator HC_ Regression-based Online Simulation of Starting Problem-Solution Pairs for Homotopy Continuation in Geometric Vision/d90c2114-786f-423a-b72c-4f3391aa43d5_model.json new file mode 100644 index 0000000000000000000000000000000000000000..05ee7c4021a2c04513b33c60282cc1417df5f22d --- /dev/null +++ b/2025/Simulator HC_ Regression-based Online Simulation of Starting Problem-Solution Pairs for Homotopy Continuation in Geometric Vision/d90c2114-786f-423a-b72c-4f3391aa43d5_model.json @@ -0,0 +1,2508 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.12, + 0.131, + 0.851, + 0.177 + ], + "angle": 0, + "content": "Simulator HC: Regression-based Online Simulation of Starting Problem-Solution Pairs for Homotopy Continuation in Geometric Vision" + }, + { + "type": "text", + "bbox": [ + 0.187, + 0.203, + 0.312, + 0.223 + ], + "angle": 0, + "content": "Xinyue Zhang" + }, + { + "type": "text", + "bbox": [ + 0.364, + 0.203, + 0.445, + 0.222 + ], + "angle": 0, + "content": "Zijia Dai" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.203, + 0.603, + 0.222 + ], + "angle": 0, + "content": "Wanting \\(\\mathbf{X}\\mathbf{u}^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.657, + 0.204, + 0.78, + 0.223 + ], + "angle": 0, + "content": "Laurent Kneip1" + }, + { + "type": "text", + "bbox": [ + 0.27, + 0.246, + 0.726, + 0.265 + ], + "angle": 0, + "content": "\\(^{1}\\)Mobile Perception Lab, ShanghaiTech University, China" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.299, + 0.314, + 0.315 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.332, + 0.473, + 0.651 + ], + "angle": 0, + "content": "While automatically generated polynomial elimination templates have sparked great progress in the field of 3D computer vision, there remain many problems for which the degree of the constraints or the number of unknowns leads to intractability. In recent years, homotopy continuation has been introduced as a plausible alternative. However, the method currently depends on expensive parallel tracking of all possible solutions in the complex domain, or a classification network for starting problem-solution pairs trained over a limited set of real-world examples. Our innovation lies in a novel approach to finding solution-problem pairs, where we only need to predict a rough initial solution, with the corresponding problem generated by an online simulator. Subsequently, homotopy continuation is applied to track that single solution back to the original problem. We apply this elegant combination to generalized camera resection, and also introduce a new solution to the challenging generalized relative pose and scale problem. As demonstrated, the proposed method successfully compensates the raw error committed by the regressor alone, and leads to state-of-the-art efficiency and success rates." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.663, + 0.21, + 0.679 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.689, + 0.471, + 0.902 + ], + "angle": 0, + "content": "The solution of geometric camera calibration problems is a crucial step in many Structure-from-Motion (SfM) [21, 52, 54, 55], Visual Odometry (VO) [46], and visual Simultaneous Localization And Mapping (SLAM) [43] frameworks. Solvers developed over the years address camera resectioning [23, 29, 31, 65], two-view relative pose [30, 44, 58], pose estimation from lines [15, 41], partially calibrated cameras [64], generalized cameras [31, 41, 57, 60], rolling shutter cameras [10, 51], and many other scenarios. The problems often appear in the form of a polynomial equation system, and special techniques from the field of algebraic geometry have been used to solve them. Once an efficient solver is found, it is typically embedded into a random sampling and consensus scheme [19, 48] in order to gain robust" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.301, + 0.892, + 0.345 + ], + "angle": 0, + "content": "ness in the presence of outliers. Hence, there is a requirement for such solvers to be efficient and embeddable into iterative schemes." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.35, + 0.893, + 0.532 + ], + "angle": 0, + "content": "The dominant approaches to efficiently solve systems of polynomial equations are given by the Gröbner basis method and polynomial resultants [9, 59]. Based on the Gröbner basis theory, Stewénius et al. [58] and Kukelova et al. [33] demonstrate that the calculations required to obtain a minimal set of ideal generators can be effectuated efficiently at the hand of a fixed elimination template. Kukelova et al. [33] in particular propose a solver generator to automatically discover such elimination templates. Later on, sparse resultant-based solvers [16] have been demonstrated as a powerful alternative sometimes leading to smaller elimination templates." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.535, + 0.895, + 0.868 + ], + "angle": 0, + "content": "While polynomial elimination techniques have led to significant progress in the solution of geometric vision problems, the methods are limited as the complexity of the template search grows uncontrollably with the degree and dimensionality of the problem at hand. Furthermore, owing to their failure to respect inequality constraints and rule out practically infeasible solutions, high degrees or dimensionalities naturally lead to an elevated number of algebraically possible solutions and—assuming that it can be found at all—large elimination templates [13, 14, 45, 63–65]. This in turn causes problems in terms of numerical stability [63]. In recent years, a different solution strategy has therefore gained popularity: Homotopy Continuation (HC) [4, 12, 61]. The method proceeds by starting from a known problem-solution pair, and then tracks its roots to the target problem at hand during a step-wise interpolation of polynomial coefficients. The method has been used to successfully solve challenging tri-focal relative pose problems [17]. While promising, HC remains expensive as it generally requires the parallel continuation of many roots in the complex domain in order to identify all real solutions to a target problem, followed up by disambiguation." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Recently, Hruby et al. [25] propose an efficient, learning-based extension to the pure HC paradigm, which serves as" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "27103" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.213 + ], + "angle": 0, + "content": "our primary motivation. They propose the introduction of a classifier to pick a good problem-solution pair trained over a large set of known pairs obtained from SfM. This enables a highly efficient application of HC, as only a single root of interest needs to be continued. While still presenting a margin for improvement in terms of success rate, the method is demonstrated to have unprecedented computational efficiency on the challenging three-view-four-point problem." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.22, + 0.472, + 0.298 + ], + "angle": 0, + "content": "The method proposed in this work is similar to the work of Hruby et al. [25] in that it leverages a learning-based approach to produce a starting problem-solution pair that enables single-root-tracking. However, we note a couple of important differences:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.304, + 0.47, + 0.456 + ], + "angle": 0, + "content": "- We do not train a classifier on a fixed set of problem-solution pairs taken from a real-world dataset. Instead, we propose to employ a solution regression network trained over an arbitrarily large set of simulated input correspondences for the considered geometric problem. The network is hence not limited by a finite set of candidate problems, and aims at being a general solution approximator for any instance of the polynomial problem. What's more, learning a solution is much easier than learning a proper problem-solution pair." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.455, + 0.47, + 0.56 + ], + "angle": 0, + "content": "- The bridge to HC is formed by appending an online simulator that uses the input correspondences and the predicted solution to generate a complete, consistent problem-solution pair. Given only moderate regression accuracy, the produced problem-solution pair is sufficiently close to the original problem for successful single-root tracking via HC." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.561, + 0.47, + 0.682 + ], + "angle": 0, + "content": "- We argue that the prediction of a single solution (or the unambiguous classification of a single problem-solution pair) from a minimal set may not always be possible as for complicated geometric problems, there may indeed be multiple geometrically feasible solutions. We therefore add one additional correspondence, which leads to a performance gain during both the regression and the continuation stage." + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.304, + 0.47, + 0.682 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.69, + 0.47, + 0.901 + ], + "angle": 0, + "content": "The newly proposed simulate-and-solve paradigm is summarized in Figure 1. We demonstrate that this strategy can lead to high success-rate and efficient CPU-based solvers for two problems: generalized camera resection, and generalized relative pose and scale. Owing to its challenging nature, the latter problem has thus far only seen an optimization-based solution [32]. The solvers are compared against a purely learning-based alternative, a regressor followed by simple local refinement, as well as the exhaustive application of homotopy continuation over all roots. The proposed method is efficient, generalizable, and highly successful, and should thus be considered as an interesting alternative for the ongoing development of complicated geometric problem solvers." + }, + { + "type": "title", + "bbox": [ + 0.501, + 0.091, + 0.679, + 0.106 + ], + "angle": 0, + "content": "2. Literature Review" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.116, + 0.895, + 0.493 + ], + "angle": 0, + "content": "Polynomial elimination theory: Our contribution can be regarded as a novel alternative to traditional polynomial elimination theory [9, 59]. Stewenius et al. [58] manually derive one of the first Gröbner basis solvers in the field of computer vision. Automatic solver generators and important extensions improving performance and numerical accuracy are later introduced by Kukelova et al. [33] and Larsson et al. [34, 35, 37, 38]. Zheng and Kneip propose the use of the parametric Sylvester resultant [64], while Hesch and Roumeliotis make use of the Macaulay resultant [23]. The more general use of sparse resultants as a powerful alternative to the Gröbner basis method is analyzed by Emiris [16], Heikkilä [22], and—most recently—Bhayani et al. [5]. While having lead to significant progress, complicated geometric problems with as many as 64 [57], 272 [45], 81 [65] (before symmetry-removal), 210 [64], \\(>10000\\) [13, 14], and 512 [63] solutions remain hard or impossible to be solved by polynomial elimination. The latter work, in particular, presents one of the largest successful elimination templates in the literature (\\(96413 \\times 96879\\)) by making use of octuple machine precision. Recently Evgenity et al. [40] developed a new method to reduce the size of the elimination template. Yet, for a polynomial system, a large number of solutions in practice still leads to reduced computational efficiency." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.493, + 0.895, + 0.795 + ], + "angle": 0, + "content": "Homotopy continuation: Homotopy continuation implementations exist as independent programs [61] or as part of Macaulay2 [39], Bertini [4], and Julia [6]. Duff et al. [12] study the application of the method to geometric vision problems, and several related studies focus specifically on the tri-focal relative pose problem [2, 20, 24, 27, 47]. However, off-the-shelf implementations are often too general and not optimized for real-time applications. Progress is recently achieved by Fabbri et al. [17], who propose an efficient custom \\(\\mathrm{C + + }\\) implementation of the Macaulay2 solver [39] for a tri-focal relative pose problem. More recently, Ding et al. [11] propose a fast GPU-based implementation of homotopy continuation [7, 8] for a minimal solution to the three-view-four-point problem which has the same performance as Julia [6] but is significantly faster. The method leverages massive parallelization to efficiently track all solutions of a small, fixed set of starting problem-solution pairs. Our contribution can be regarded as a problem-solution pair initialization scheme similar to Hruby et al. [25] for successful single solution tracking." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.796, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Learning for polynomial problem solving: One of the first methods to successfully employ learning in polynomial solvers is presented by Xu et al. [62], who proposed to increase the numerical accuracy of polynomial elimination templates by learning the best permutation in permutation-invariant polynomial problems. Recently, Hruby et al. [25] propose the addition of a classifier for initial problem-" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "27104" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.125, + 0.089, + 0.852, + 0.233 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.243, + 0.895, + 0.3 + ], + "angle": 0, + "content": "Figure 1. Overview of the proposed geometric problem solution scheme. Given input correspondences, a regression network is utilized to approximate a solution. A subsequent online simulator generates a new set of correspondences that is consistent with the regression output. The obtained problem-solution pair is finally used to bootstrap homotopy continuation. The final solution is found efficiently by tracking a single root." + }, + { + "type": "image", + "bbox": [ + 0.086, + 0.308, + 0.255, + 0.396 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.085, + 0.401, + 0.253, + 0.413 + ], + "angle": 0, + "content": "(a) HC multiple solution curves" + }, + { + "type": "image", + "bbox": [ + 0.293, + 0.306, + 0.465, + 0.396 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.316, + 0.401, + 0.443, + 0.413 + ], + "angle": 0, + "content": "(b) HC solution tacking" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.424, + 0.471, + 0.509 + ], + "angle": 0, + "content": "Figure 2. Graphical illustration on homotopy continuation(HC) solution curves tracking. For a polynomial system with multiple solutions, each solution curve can be tracked independently as shown in Figure 2a. For each solution curve tracking, HC utilizes a \"prediction-correction\" scheme to approximate one of the solutions for the target system." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.515, + 0.47, + 0.789 + ], + "angle": 0, + "content": "solution pair picking in homotopy continuation. In recent years, several works have proposed to directly apply neural networks to regress the pose of a camera, either directly from images [26, 49] or—aiming at replacing traditional algebraic-geometric solvers—from correspondence coordinates [53]. While such networks use algorithm unrolling [42] and differential optimization [3] for accurate end-to-end trainability, a successful application to more complicated problems remains yet to be demonstrated. Regressors can be regarded as general non-linear function approximators, and the challenge of using them to produce highly accurate solutions is easily explained by the potential discontinuity of root locations for continuous variations of input polynomial coefficients. Our contribution can be regarded as a new post-regression refinement scheme for polynomial problems enabling successful application to complicated problems for which it is hard to apply traditional polynomial solvers." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.8, + 0.469, + 0.818 + ], + "angle": 0, + "content": "3. Polynomial Solver: Homotopy Continuation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.47, + 0.902 + ], + "angle": 0, + "content": "We start with a brief summary of HC. We refer to an instance of a certain polynomial system with specific coefficients as a problem. HC is a powerful tool that can numerically find the solutions of a problem. Each solution is found by tracking a curve independently, as illustrated in" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.308, + 0.892, + 0.459 + ], + "angle": 0, + "content": "Figure 2a. When tracking the solution curves, the method solves a series of problems of similar form. It starts from a trivial problem whose solutions are simply known, and then gradually deforms that problem to the target problem that we are interested in but for which we do not know the solutions. It is hoped that—by continuing the known solutions throughout the problem sequence—they can lead to the solutions of the target, input problem. Two questions arise: 1) How to track the solution curves?, and 2) What is a good starting problem leading to the solutions we desire?" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.459, + 0.892, + 0.55 + ], + "angle": 0, + "content": "The answer to the first question is given by HC. Formally, suppose we have a polynomial system \\(\\mathbf{F}(\\mathbf{x}) = \\mathbf{0}\\) and we are interested in its solutions. Suppose further that we have a polynomial system of similar form \\(\\mathbf{G}(\\mathbf{x}) = \\mathbf{0}\\) whose solutions are known. HC solves a series of polynomial systems of the form" + }, + { + "type": "equation", + "bbox": [ + 0.542, + 0.559, + 0.892, + 0.577 + ], + "angle": 0, + "content": "\\[\n\\mathbf {H} (\\mathbf {x}, t) = (1 - t) \\mathbf {G} (\\mathbf {x}) + t \\mathbf {F} (\\mathbf {x}), \\quad t \\in [ 0, 1 ]. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.586, + 0.892, + 0.708 + ], + "angle": 0, + "content": "As illustrated in Figure 2b, each \\(\\mathbf{H}(\\mathbf{x},t) = \\mathbf{0}\\) is solved while sliding \\(t\\) from 0 to 1. For \\(t = 0\\), the solution is known as \\(\\mathbf{H}(\\mathbf{x},0) = \\mathbf{G}(\\mathbf{x})\\). A solution of \\(\\mathbf{F}(\\mathbf{x}) = \\mathbf{0}\\) is finally obtained as for \\(t = 1\\), \\(\\mathbf{H}(\\mathbf{x},1) = \\mathbf{F}(\\mathbf{x})\\). The problems are not actually solved but the solutions are simply tracked from problem to problem. Taking the partial derivative to both sides of \\(\\mathbf{H}(\\mathbf{x},t)\\), the solution curve \\(\\mathbf{x}(t)\\) can be formulated as the following ordinary differential equation" + }, + { + "type": "equation", + "bbox": [ + 0.499, + 0.715, + 0.905, + 0.757 + ], + "angle": 0, + "content": "\\[\n\\mathbf {J} _ {\\mathbf {H}} (\\mathbf {x}) \\frac {\\partial \\mathbf {x}}{\\partial t} + \\frac {\\partial \\mathbf {H} (\\mathbf {x} , t)}{\\partial t} = \\mathbf {0} \\Longleftrightarrow \\frac {\\partial \\mathbf {x}}{\\partial t} = - \\mathbf {J} _ {\\mathbf {H}} (\\mathbf {x}) ^ {\\dagger} \\frac {\\partial \\mathbf {H} (\\mathbf {x} , t)}{\\partial t}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.758, + 0.892, + 0.802 + ], + "angle": 0, + "content": "A \"prediction-correction\" scheme—as widely utilized in numerical analysis—is typically employed to trace the solution curve \\(\\mathbf{x}(t)\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.803, + 0.892, + 0.834 + ], + "angle": 0, + "content": "Prediction. Denote the stepsize of \\( t \\) as \\( \\delta t \\). Euler's predictor can be used to update along the tangent direction" + }, + { + "type": "equation", + "bbox": [ + 0.611, + 0.842, + 0.892, + 0.879 + ], + "angle": 0, + "content": "\\[\n\\mathbf {x} _ {i + 1} = \\mathbf {x} _ {i} + \\delta t \\frac {\\partial \\mathbf {x}}{\\partial t} \\bigg | _ {\\mathbf {x} _ {i}, t _ {i}}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.886, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Practically speaking, a higher-order generalized method is" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "27105" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.122 + ], + "angle": 0, + "content": "commonly used for better accuracy and stability with larger stepsize." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.127, + 0.47, + 0.173 + ], + "angle": 0, + "content": "Correction. Gauss-Newton steps are adopted to correct the prediction. In each iteration of Gauss-Newton's method, the solution is updated by" + }, + { + "type": "equation", + "bbox": [ + 0.147, + 0.194, + 0.47, + 0.213 + ], + "angle": 0, + "content": "\\[\n\\mathbf {x} _ {j} = \\mathbf {x} _ {j - 1} - \\mathbf {J} _ {\\mathbf {H}} \\left(\\mathbf {x} _ {j - 1}\\right) ^ {\\dagger} \\mathbf {H} \\left(\\mathbf {x} _ {j - 1}, t\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.235, + 0.469, + 0.298 + ], + "angle": 0, + "content": "where \\(\\mathbf{J}_{\\mathbf{H}}^{\\dagger}\\) is the pseudo inverse of the jacobian \\(\\mathbf{J}_{\\mathbf{H}}\\), and \\(t = t_i + \\delta t\\). Figure 2b illustrates a 1D solution curve tracking, where the blue arrow represents the prediction update and the red arrow represents the correction update." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.302, + 0.47, + 0.573 + ], + "angle": 0, + "content": "The remainder of this paper—our key contribution—addresses the second question of how to find a good starting system. Traditional methods (e.g. Ab initial [17], monodromy [6, 12]) rely on algebraic geometry theories developed for coefficient-parameter homotopies, and tracking all roots from a small set of fixed, known problem-solution pairs in the complex domain may often enable a solution [56]. However, tracking all the isolated paths is expensive and in practice multiple solutions cost more time to select the best. A recent study proposes the use of a classifier to pick a good starting pair with as high as possible success probability upon tracking only a single root [25], which is much faster. Similarly, in this work, we propose a novel geometry-informed strategy that combines a neural regressor with an online simulator to produce a problem-solution pair and then track a single root by HC. Even on challenging problems for which no traditional solver has been found, it can find the solution with a success rate of over \\(95\\%\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.599, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Remark (Homotopy Types). Since being introduced in 1980s, a variety of homotopies have been developed to serve the same purpose: how to solve the problem exactly as expected. The representative one would be coefficient-parameter homotopy. Its formulation provides the insight that if we can find all the solutions to a general coefficient-parameter polynomial, then we can find all the other members of that polynomial family. In fact, as pointed out in [56], in engineering problems (e.g. geometry vision problems) we are dealing with natural coefficient-parameter polynomials, that is coefficients of polynomials derived from visual features. Since the scope of this paper lies on a novel starting problem-solution pair finding and it is not highly dependent on the path tracking, we kindly refer the reader to [56] for an in-depth discussion of other homotopy types. As shown in the later applications, straight segment homotopies provide near-perfect success rates when working with our simulator. We also want to note that all the other homotopy types could be equally applied in conjunction with our simulator." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.09, + 0.892, + 0.123 + ], + "angle": 0, + "content": "4. Find a Good Starting Solution: Regression Network" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.134, + 0.892, + 0.36 + ], + "angle": 0, + "content": "Following recent developments in deep learning, multiple attempts have been made to, for example, learn camera pose using a neural network [26, 49, 53]. The first two methods extract features from images and points and use them in a regression network to predict the pose. The third method is a more preliminary approach and aims at predicting pose directly from geometric correspondences, which is also the aim of the present work. The bottleneck in regressing pose from correspondences arises from two concerns: 1) For simple geometric problems, we often have traditional methods that are efficient and accurate, and 2) For hard geometric problems, regression networks demonstrate a poor ability to predict accurate solutions. Our key insight is that while the predicted pose may not be sufficiently accurate, it can still serve as a good starting solution for HC." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.362, + 0.892, + 0.437 + ], + "angle": 0, + "content": "Our pose regression networks are designed with multiple 1D convolutional layers and each layer is followed by batch normalization and ReLU activations. The loss function is given by a combination of relative pose errors with learnable weights [26]." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.452, + 0.892, + 0.486 + ], + "angle": 0, + "content": "5. Find a Consistent Starting Problem: Online Simulator" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.496, + 0.892, + 0.753 + ], + "angle": 0, + "content": "We recall the main idea of HC. In order to continue a solution to the target problem, an appropriate starting problem with a known solution is required. Given the approximate solution predicted by the pose regression network, our goal is to derive a consistent polynomial system for which one of the solutions matches the predicted one. We denote this step an online simulator, and—though problem specific—effective realizations can almost always be found by simple geometric rules. For example, in camera resectioning, one can simply use the input 3D points and reproject them into a camera view under the predicted pose. This will lead to new 2D points, and thus a new set of correspondences that is perfectly consistent with the predicted solution, i.e. our sought starting problem-solution pair. While not yet made use of in the present work, it is worthwhile noting that—owing to their algebraic simplicity—such an online simulator often appears to be a differentiable module." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.769, + 0.803, + 0.787 + ], + "angle": 0, + "content": "6. Generalized Perspective-n-Points" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.795, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Our first concrete, practical application targets generalized camera pose estimation from 2D-3D correspondences. We use the Unified Perspective-n-Point (UPnP) algorithm [31] which formulates a polynomial system corresponding to the first-order optimality conditions of the sum of squared object space distances. The method can be used to solve both central and non-central camera resectioning problems from" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "27106" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.09, + 0.255, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.085, + 0.189, + 0.248, + 0.201 + ], + "angle": 0, + "content": "(a) Generalized Absolute Pose" + }, + { + "type": "image", + "bbox": [ + 0.293, + 0.092, + 0.468, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.291, + 0.189, + 0.469, + 0.213 + ], + "angle": 0, + "content": "(b) Generalized Relative Pose and Scale" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.226, + 0.47, + 0.282 + ], + "angle": 0, + "content": "Figure 3. The geometry of the generalized camera problems we propose to solve by online simulator HC. \\( w \\) represents the world frame, and \\( c \\) and \\( c' \\) represent camera frames in different views respectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.283, + 0.469, + 0.389 + ], + "angle": 0, + "content": "an arbitrary number of points (including minimal). It can be efficiently solved by the Gröbner Bases method since—after variable elimination—it only involves 4 unknowns. In this section, we solve UPnP by our newly proposed method and compare it against conventional approaches, a raw regression network, as well as a parallel HC implementation tracking all roots." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.397, + 0.281, + 0.411 + ], + "angle": 0, + "content": "6.1. Problem Formulation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.419, + 0.47, + 0.541 + ], + "angle": 0, + "content": "Consider a 3D-2D generalized correspondence \\((\\mathbf{p}_i, \\mathbf{f}_i, \\mathbf{v}_i)\\) originating from the non-central camera resectioning model illustrated in Figure 3a, where \\(\\mathbf{p}_i \\in \\mathbb{R}^3\\), \\(\\mathbf{f}_i \\in \\mathbb{S}^2\\) and \\(\\mathbf{v}_i \\in \\mathbb{R}^3\\) denote the 3D world point, image ray and ray origin, respectively. The correspondence follows an underlying generalized absolute pose transformation depending on the rotation \\(\\mathbf{R} \\in SO(3)\\), the position \\(\\mathbf{t} \\in \\mathbb{R}^3\\) and an unknown latent depth \\(\\alpha_i\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.197, + 0.549, + 0.469, + 0.565 + ], + "angle": 0, + "content": "\\[\n\\alpha_ {i} \\mathbf {f} _ {i} + \\mathbf {v} _ {i} = \\mathbf {R p} _ {i} + \\mathbf {t}. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.572, + 0.469, + 0.693 + ], + "angle": 0, + "content": "Let \\(\\mathbf{M} \\in \\mathbb{R}^{11 \\times 11}\\) denote a coefficient matrix that embeds all known variables (point measurements and ray offsets), \\(\\mathbf{q} = [q_0, q_1, q_2, q_3]^\\top\\) denote the quaternion representation of \\(\\mathbf{R}\\), and \\(\\mathbf{s} \\in \\mathbb{R}^{11}\\) denote a vector whose elements are all second and zeroth order monomials of the quaternion variables. After translation and depth elimination, \\(\\mathbf{q}\\) can be constrained by first-order optimality conditions appearing in the form of the polynomial system1" + }, + { + "type": "equation", + "bbox": [ + 0.163, + 0.7, + 0.469, + 0.737 + ], + "angle": 0, + "content": "\\[\n\\left\\{ \\begin{array}{l l} \\mathbf {s} ^ {\\top} \\mathbf {M} \\frac {\\partial \\mathbf {s}}{\\partial q _ {i}} & = 0, \\quad i = 0, \\dots , 3 \\\\ \\mathbf {q} ^ {\\top} \\mathbf {q} - 1 & = 0. \\end{array} \\right. \\tag {6}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.737, + 0.354, + 0.751 + ], + "angle": 0, + "content": "6.2. Regressor and online simulator" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.759, + 0.469, + 0.805 + ], + "angle": 0, + "content": "To provide a proper problem-solution pair for HC, we train a pose regression network from synthetic data. Its input is given by row-wise stacking of the correspondences, i.e." + }, + { + "type": "equation", + "bbox": [ + 0.212, + 0.81, + 0.469, + 0.87 + ], + "angle": 0, + "content": "\\[\n\\left[ \\begin{array}{c c c} \\mathbf {p} _ {1} ^ {\\top} & \\mathbf {f} _ {1} ^ {\\top} & \\mathbf {v} _ {1} ^ {\\top} \\\\ \\vdots & \\vdots & \\vdots \\\\ \\mathbf {p} _ {N} ^ {\\top} & \\mathbf {f} _ {N} ^ {\\top} & \\mathbf {v} _ {N} ^ {\\top} \\end{array} \\right], \\tag {7}\n\\]" + }, + { + "type": "page_footnote", + "bbox": [ + 0.076, + 0.875, + 0.469, + 0.9 + ], + "angle": 0, + "content": "1We kindly refer the reader to the UPnP paper for detailed derivations [31]." + }, + { + "type": "title", + "bbox": [ + 0.502, + 0.106, + 0.866, + 0.121 + ], + "angle": 0, + "content": "Algorithm 1 Online Simulator HC for generalized PnP" + }, + { + "type": "text", + "bbox": [ + 0.507, + 0.127, + 0.8, + 0.141 + ], + "angle": 0, + "content": "1: Input: correspondences \\((\\mathbf{p}_i,\\mathbf{f}_i,\\mathbf{v}_i)_{i = 1}^M\\) , stepsize \\(\\delta t\\)" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.142, + 0.632, + 0.153 + ], + "angle": 0, + "content": "2: Output: poses \\(\\hat{\\mathbf{R}},\\hat{\\mathbf{t}}\\)" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.153, + 0.813, + 0.166 + ], + "angle": 0, + "content": "3: \\(\\tilde{\\mathbf{R}},\\tilde{\\mathbf{t}}\\gets\\) Trained Regression Network((pi,fi,v)M)" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.166, + 0.891, + 0.181 + ], + "angle": 0, + "content": "4: \\((\\hat{\\mathbf{f}}_i)_{i=1}^M \\leftarrow\\) Reprojection \\((\\hat{\\mathbf{R}}, \\hat{\\mathbf{t}}; (\\mathbf{p}_i, \\mathbf{v}_i)_{i=1}^M)\\) \\(\\triangleright\\) Online Simulator" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.181, + 0.852, + 0.195 + ], + "angle": 0, + "content": "5: Starting problem \\(\\mathbf{G}(\\mathbf{x})\\gets (\\mathbf{p}_i,\\hat{\\mathbf{f}}_i,\\mathbf{v}_i)_{i = 1}^M\\) with solution \\(\\hat{\\mathbf{R}},\\hat{\\mathbf{t}}\\)" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.195, + 0.74, + 0.208 + ], + "angle": 0, + "content": "6: Target problem \\(\\mathbf{F}(\\mathbf{x})\\gets (\\mathbf{p}_i,\\mathbf{f}_i,\\mathbf{v}_i)_{i = 1}^M\\)" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.208, + 0.833, + 0.22 + ], + "angle": 0, + "content": "7: \\(\\mathbf{q}_{HC} \\gets\\) Homotopy Continuation \\(t\\mathbf{F}(\\mathbf{x}) + (1 - t)\\mathbf{G}(\\mathbf{x})\\)" + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.221, + 0.663, + 0.234 + ], + "angle": 0, + "content": "8: \\(\\bar{\\mathbf{R}},\\hat{\\mathbf{t}}\\gets\\) decompose \\(\\mathbf{q}_{HC}\\)" + }, + { + "type": "list", + "bbox": [ + 0.507, + 0.127, + 0.891, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.256, + 0.893, + 0.457 + ], + "angle": 0, + "content": "where \\(N\\) is the number of correspondences used, and the output is the quaternion vector. The network consists of 4 1D convolutional layers and a fully connected layer at the end to convert the output to a quaternion vector with the unit norm. Given the rotation predicted by the network, the translation can be recovered from correspondences and estimated rotation. We refer the reader to [31] for the details. For the online simulator, we can get reprojected \\(\\hat{\\mathbf{f}}_i\\) from (5). Algorithm 1 summarizes all stages, including the final HC over (6). Note that in line 5 of Algorithm 1, the starting problem \\(\\mathbf{G}(\\mathbf{x})\\) is constructed by replacing the coefficient matrix \\(\\mathbf{M}\\) computed by \\((\\mathbf{p}_i,\\mathbf{f}_i,\\mathbf{v}_i)\\) in (6) with \\(\\hat{\\mathbf{M}}\\) computed by \\((\\mathbf{p}_i,\\hat{\\mathbf{f}}_i,\\mathbf{v}_i)\\)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.465, + 0.635, + 0.481 + ], + "angle": 0, + "content": "6.3. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.488, + 0.892, + 0.759 + ], + "angle": 0, + "content": "Data. We randomly generate 4 cameras in a cube \\([-1,1]^3\\) and randomly sample 16 3D points in the box \\([-1,1]^2 \\times [4,8]\\). 16 samples would not be a good choice for a RANSAC solver [18], but in the present context the idea is to set a moderate and notably equal challenge to all compared solvers. Furthermore, for UPnP the number of constraints appearing in the polynomial system remains unchanged as it analyses first-order optimality rather than individual incidence relations. Elements of the rotation angles are randomly generated in the interval \\([- \\pi /2, \\pi /2]\\), respectively, and the translation is randomly generated in the unit cube. We then generate the correspondences from rays that point to a 3D point. The noise level is set to 2 pixels. To train the pose regression network, we simulate a dataset of 2400 samples, which is sufficient for reliable Simulator HC initialization. We use a 2.70GHz Intel(R) Xeon(R) CPU and an NVIDIA GeForce RTX 4090 for network training and evaluation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.76, + 0.892, + 0.866 + ], + "angle": 0, + "content": "Methods. For HC, the stepsize \\(\\delta t\\) and the maximum number of iterations of Newton's method are set to 0.02 and 5, respectively. We compare the Gröbner Bases solver [31] and the popular local method Levenberg-Marquardt (LM) [1]. We also test the state-of-the-art parallel HC solver developed by [7, 8], and use both its CPU and \\(\\mathrm{GPU}^2\\) version. Given that parallel HC tracks all solutions," + }, + { + "type": "page_footnote", + "bbox": [ + 0.5, + 0.875, + 0.89, + 0.9 + ], + "angle": 0, + "content": "2Note that GPU warmup overhead is ignored when averaging over multiple runs." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "27107" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.089, + 0.472, + 0.146 + ], + "angle": 0, + "content": "Table 1. Performance comparison of simulator HC, a traditional method, and parallel exhaustive tracking HC on noisy synthetic data. All results are averaged over 1000 trials. The success rate is computed for rotation errors with a threshold of 2 degree." + }, + { + "type": "table", + "bbox": [ + 0.078, + 0.157, + 0.477, + 0.269 + ], + "angle": 0, + "content": "
MethodSucc. (%)ER(deg)Et(%)Time(ms)
Gröbner Bases1000.0310.4891.254
CPU HC1000.0320.4999.796
GPU HC1000.0320.4991.002
Pred Initial327.204320.031
Pred Initial + LM5428.7222560.528
Rand Initial + Simulator HC2291.682280.231
Pred Initial + Simulator HC1000.0320.5130.316
" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.281, + 0.404, + 0.295 + ], + "angle": 0, + "content": "we manually pick the one with the smallest error." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.297, + 0.471, + 0.6 + ], + "angle": 0, + "content": "Results. The rotation error is measured by \\(\\mathcal{E}_{\\mathbf{R}} = 2\\arcsin (\\| \\hat{\\mathbf{R}} -\\mathbf{R}_{gt}\\| /(2\\sqrt{2}))\\) in degrees as suggested by [11] and the translation error is measured by the \\(\\mathcal{E}_{\\mathrm{t}} = \\| \\hat{\\mathbf{t}} -\\mathbf{t}_{gt}\\| /\\| \\mathbf{t}_{gt}\\| \\times 100\\) in percentage. Table 1 compares the traditional method, parallel exhaustive tracking HC, and different types of initial solutions for our simulator HC. We present the success rate as the percentage of rotation errors smaller than 2 degrees. It can be observed that on this simple task, except for the local method, traditional methods, as well as simulator HC using predicted initials, succeed in \\(100\\%\\) of trials. Meanwhile, online simulator HC benefits from the single-track strategy and turns out to be the fastest one. We also perform an ablation study on the proposed simulator HC, using both predicted and random initializations. Although the pose learned by the regressor is not highly accurate and only slightly outperforms random poses, it serves as a reasonable starting solution for our simulator HC. In contrast, LM as a local method is unable to leverage the potential of the predicted initializations, as the simulator HC does." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.618, + 0.404, + 0.633 + ], + "angle": 0, + "content": "7. Generalized Relative Pose and Scale" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.644, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Our final application targets a hard minimal problem: Generalized Relative Pose and Scale (GRPS) estimation from 2D-2D correspondences [32]. The GRPS problem is a natural extension of the generalized relative pose problem and arises in point-less registration of partial view-graphs in structure-from-motion. Owing to the fact that SfM results are generally up to scale, an inherent scale factor needs to be resolved. Such problem was first introduced by Kneip et al. [32]. It has 140 solutions if a unique rotation representation is used, and can be solved by multi-start optimization to overcome local minima. Later on, a traditional polynomial solver based on the Gröbner Bases method has been proposed [36, 40]. With a \\(144 \\times 284\\) elimination template, the solver finds all 140 rotations within reasonable time. In this section, we solve GRPS by our proposed method and compare it against the existing approach, as well as a parallel HC implementation tracking all roots." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.091, + 0.704, + 0.105 + ], + "angle": 0, + "content": "7.1. Problem Formulation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.113, + 0.892, + 0.251 + ], + "angle": 0, + "content": "Consider a 2D-2D generalized correspondence \\((\\mathbf{f}_i,\\mathbf{v}_i) - (\\mathbf{f}_i',\\mathbf{v}_i')\\) observed from two non-central camera views as illustrated in Figure 3b. \\(\\mathbf{f}_i\\in \\mathbb{S}^2,\\mathbf{v}_i\\in \\mathbb{R}^3\\) denote the image ray and ray origin in the first view, and \\(\\mathbf{f}_i'\\in \\mathbb{S}^2,\\mathbf{v}_i'\\in \\mathbb{R}^3\\) are the corresponding entities in the second view. The correspondence follows an underlying generalized relative pose and scale incidence relation depending on the rotation \\(\\mathbf{R}\\in SO(3)\\), the position \\(\\mathbf{t}\\in \\mathbb{R}^3\\), the scale \\(s\\), and unknown latent depths \\(\\alpha_{i}\\) and \\(\\alpha_{i}'\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.584, + 0.261, + 0.892, + 0.278 + ], + "angle": 0, + "content": "\\[\n\\alpha_ {i} \\mathbf {f} _ {i} + \\mathbf {v} _ {i} = \\mathbf {R} \\left(\\alpha_ {i} ^ {\\prime} \\mathbf {f} _ {i} ^ {\\prime} + s \\mathbf {v} _ {i} ^ {\\prime}\\right) + \\mathbf {t}. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.289, + 0.892, + 0.349 + ], + "angle": 0, + "content": "With the help of Plücker coordinates and the scalar triple product \\(^3\\), the unknown depths get eliminated which results in a polynomial constraint that only involves the unknown poses" + }, + { + "type": "equation", + "bbox": [ + 0.518, + 0.358, + 0.892, + 0.378 + ], + "angle": 0, + "content": "\\[\n\\mathbf {t} ^ {\\top} \\left(\\mathbf {f} _ {i} \\times \\mathbf {R} \\mathbf {f} _ {i} ^ {\\prime}\\right) - s \\mathbf {f} _ {i} ^ {\\top} \\mathbf {R} \\left[ \\mathbf {v} _ {i} ^ {\\prime} \\right] _ {\\times} \\mathbf {f} _ {i} ^ {\\prime} + \\mathbf {f} _ {i} ^ {\\top} \\left[ \\mathbf {v} _ {i} \\right] _ {\\times} \\mathbf {R} \\mathbf {f} _ {i} ^ {\\prime} = 0. \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.388, + 0.892, + 0.448 + ], + "angle": 0, + "content": "Therefore, a minimal formulation for the GRPS problem can be devised by a polynomial system formed by stacking 7 constraints in the form (9), each one depending on one correspondence." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.457, + 0.776, + 0.473 + ], + "angle": 0, + "content": "7.2. Regressor and online simulator" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.481, + 0.892, + 0.678 + ], + "angle": 0, + "content": "As introduced in section 6.2, we continue to use a pose regression network that inputs a matrix with row-wise stacked correspondences. The output is given by \\(\\mathbf{R}\\), \\(\\mathbf{t}\\), \\(s\\). Note that the rotation is predicted in continuous 6D representation which is more suitable for neural networks [66]. The network consists of 11 1D convolutional layers, in the end, it branches to 3 sub-branches for \\(\\mathbf{R}\\), \\(\\mathbf{t}\\) and \\(s\\). Each sub-branch consists of 2 1D convolutional layers and 1 fully connected layers for adjust output. Given the poses predicted by the network, we still need an estimation of the depths before simulating the reprojected image ray. With estimated \\(\\hat{\\mathbf{R}}\\), \\(\\hat{\\mathbf{t}}\\) and \\(\\hat{s}\\), the corresponding depths can be obtained from (8) by following" + }, + { + "type": "equation", + "bbox": [ + 0.561, + 0.687, + 0.892, + 0.721 + ], + "angle": 0, + "content": "\\[\n\\left[ \\mathbf {f} _ {i} - \\hat {\\mathbf {R}} \\mathbf {f} _ {i} ^ {\\prime} \\right] \\left[ \\begin{array}{l} \\alpha_ {i} \\\\ \\alpha_ {i} ^ {\\prime} \\end{array} \\right] = - \\left(\\mathbf {v} _ {i} - \\hat {s} \\hat {\\mathbf {R}} \\mathbf {v} _ {i} ^ {\\prime} - \\hat {\\mathbf {t}}\\right), \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.73, + 0.892, + 0.775 + ], + "angle": 0, + "content": "which can be simply solved by linear least squares for each input correspondence. Finally, we simulate a pose-consistent reprojected image ray in the first view using" + }, + { + "type": "equation", + "bbox": [ + 0.579, + 0.785, + 0.892, + 0.815 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {f}} _ {i} = \\frac {1}{\\hat {\\alpha} _ {i}} \\hat {\\mathbf {R}} \\left(\\hat {\\alpha} _ {i} ^ {\\prime} \\mathbf {f} _ {i} ^ {\\prime} + \\hat {s} \\mathbf {v} _ {i} ^ {\\prime}\\right) + \\hat {\\mathbf {t}} - \\mathbf {v} _ {i}. \\tag {11}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.825, + 0.724, + 0.841 + ], + "angle": 0, + "content": "7.3. Simulation Experiments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.848, + 0.892, + 0.88 + ], + "angle": 0, + "content": "Data. We adopt a similar data generation method as in section 6.3. We generate 7 3D points in a \\([-1,1]^2 \\times [2,20]\\)" + }, + { + "type": "page_footnote", + "bbox": [ + 0.516, + 0.887, + 0.848, + 0.9 + ], + "angle": 0, + "content": "3We kindly refer the reader to [28, 32] for detailed derivations" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "27108" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.198 + ], + "angle": 0, + "content": "volume. Two view frames are generated with origins picked from \\([-1, 1]^3\\). Within each view frame, we then sample individual camera offsets from the volume \\([-1, 1]^3\\). The scale is randomly generated in \\([0.1, 5.0]\\), and all other parameters related to noise addition are left unchanged. Note that the ground truth relative pose needs to be extracted from the simulated pair of view poses." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.198, + 0.471, + 0.379 + ], + "angle": 0, + "content": "Methods. The stepsize for homotopy continuation is set to 0.05 in this scenario. We compare against the Eigen solver [32], Gröbner Bases solver [40] and the CPU and GPU versions of parallel HC which tracks 140 solutions found by Monodromy [6]. Note that to measure numerical errors, we just pick the best solution with the smallest error. In practice, traditional HC needs to evaluate additional information to select a single solution (e.g. euclidean distance). For a complicated problem like GRPS, evaluating 140 solutions would still induce non-neglectable time-consumption. Since GRPS is a hard problem, we train the regressor using a larger simulated dataset of size 64000." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.379, + 0.47, + 0.424 + ], + "angle": 0, + "content": "Metric. We measure the rotation error in degrees and the relative translation as in Sec. 6.3 and scale error in percentage using" + }, + { + "type": "equation", + "bbox": [ + 0.184, + 0.424, + 0.469, + 0.441 + ], + "angle": 0, + "content": "\\[\n\\mathcal {E} _ {s} = \\left| \\hat {s} - s _ {g t} \\right| / s _ {g t} \\times 1 0 0. \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.448, + 0.469, + 0.478 + ], + "angle": 0, + "content": "To calculate the success rate, for rotation errors the threshold is set to 2 deg, and for relative errors, it is set to \\(5\\%\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.479, + 0.47, + 0.674 + ], + "angle": 0, + "content": "Numerical Stability. We randomly generate 1000 trials to test the numerical stability of all solvers. During the analysis of the performance of simulator HC using the minimum number of correspondences, we observed that our method has around \\(70\\%\\) success rate. By checking the failure cases, we found that sometimes simulator HC tracked another real solution that was different from the ground truth pose. To tackle that issue, we evaluate the addition of one more correspondence and solve for an overdetermined polynomial system. By using 8 correspondences, simulator HC reaches a success rate of \\(95\\%\\). Again, since CPU/GPU-HC is designed to handle square systems, only the minimum number of correspondences can be tested for other HC methods." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.675, + 0.47, + 0.78 + ], + "angle": 0, + "content": "Since the errors in rotation, translation and scale have similar density results, we only present the rotation error in the main paper and the rest in the supplementary. As illustrated in Figure 4a, the minimal HC solvers all present similar error distribution and around \\(75\\%\\) of trials are successful. While Gröbner Bases shows \\(100\\%\\) success rate, Simulator HC over 8pts presents similar numerical errors." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Time Comparison. The main computational benefit of the proposed online simulator HC comes from single solution-tracking, which can be effectuated efficiently even on the CPU. Figure 4b demonstrates the efficiency gain of our 8pts simulator HC over the eigen solver and parallel HC variants. As expected, by tracking only one solution curve, we are significantly faster than parallel exhaustive tracking HC, even if the latter runs on GPU. The eigen solver" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.168 + ], + "angle": 0, + "content": "is an optimization-based solver with high complexity and ends up being slowest. The Gröbner Bases solver dealing with the large elimination template still costs a considerable time, even for just computing solutions for rotation, and not translation or scale." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.173, + 0.895, + 0.34 + ], + "angle": 0, + "content": "Remark. We would like to note that Gröbner Bases is solving a different formulation from Eq. (9). In order to make it easy to be solved by a smaller elimination template, it eliminates translation and scale [57]. But the resulting coefficient matrix is \\(130 \\times 275\\) and still needs a large number of operations. Although the formulation of Eq. (9) with eliminated depths faces a slightly lower success rate, its number of operations is much smaller than Gröbner Bases. On the other hand, the proposed 8pt simulator HC provides a comparably high success rate as Gröbner Bases and is still \\(17 \\times\\) faster." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.357, + 0.892, + 0.442 + ], + "angle": 0, + "content": "Table 2. RANSAC experiments for the GRPS problem using Gröbner Bases | simulator HC. Image points are randomly perturbed by 2-pixel uniform noise. We use vanilla RANSAC with at most 200 iterations. As can be observed, outlier ratios of up to \\(40\\%\\) are easily handled within 200 iterations. And regressor-based simulator HC is consistently faster than Gröbner Bases." + }, + { + "type": "table", + "bbox": [ + 0.528, + 0.453, + 0.862, + 0.521 + ], + "angle": 0, + "content": "
RatioSucc. (%)ER(deg)# ItersTime(s)
10%99.51000.100.09931252.330.85
20%99.41000.110.101711934.371.31
30%99.41000.130.142002005.101.35
40%98.9910.210.292002004.591.35
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.54, + 0.892, + 0.66 + ], + "angle": 0, + "content": "Noise Resilience. Figure 4c presents the performance of the proposed 8pts simulator HC with different numbers of cameras under increasing image noise levels ranging from 0.2 pixel to 1.0 pixel. The error behaves gracefully as a function of noise, and it can be observed that a higher number of cameras leads to slightly inferior noise. Furthermore, we also observe that rotation errors appear to be less sensitive to image noise than translation and scale errors." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.662, + 0.892, + 0.873 + ], + "angle": 0, + "content": "Evaluation inside RANSAC. To conclude, we embed simulator HC into RANSAC to test its practical usefulness. We synthesize 200 correspondences between 5 cameras and replace part of them with outliers. Uniformly sampled pixel noise from \\([-2,2]\\) is added to both coordinates of the image points by again assuming a virtual focal length of 800 px. We terminate RANSAC if it reaches \\(99\\%\\) confidence or 200 iterations, and the inlier threshold is set to 0.01 for the distance measured by the normalized reprojection error. Given that Gröbner Bases solver estimates 140 solutions, we keep the real ones and select the best solution using reprojection errors in parallel to save time for it. Table 2 shows the results. The success rate is measured by a rotation error threshold of 2 degrees and the indicated errors are" + }, + { + "type": "page_footnote", + "bbox": [ + 0.518, + 0.887, + 0.658, + 0.901 + ], + "angle": 0, + "content": "\\( {}^{4} \\) shown in supplementary." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "27109" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.093, + 0.34, + 0.228 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.1, + 0.235, + 0.321, + 0.247 + ], + "angle": 0, + "content": "(a) Error distribution over noise-free data." + }, + { + "type": "image", + "bbox": [ + 0.349, + 0.093, + 0.621, + 0.229 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.367, + 0.235, + 0.603, + 0.248 + ], + "angle": 0, + "content": "(b) The boxplot of running time comparison." + }, + { + "type": "image", + "bbox": [ + 0.632, + 0.092, + 0.891, + 0.23 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.667, + 0.235, + 0.854, + 0.247 + ], + "angle": 0, + "content": "(c) Error statistics of simulator HC." + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.259, + 0.894, + 0.357 + ], + "angle": 0, + "content": "Figure 4. Experimental results on GRPS. Fig. 4a Error distribution over 1000 trials on noise-free data. The camera number is set to be 3. Considering the rotation error, except for Gröbner Bases, the other minimal solvers all present a similar error distribution with around \\(70\\%\\) success rate, and the eigen solver presents only around \\(60\\%\\) success rate. The proposed 8pts simulator HC presents \\(96.3\\%\\) success rate where the Gröbner Bases has \\(100\\%\\) success rate. Fig. 4b The boxplot of running time comparison for the proposed simulator HC and other methods, where the number represents the median time. As expected, on average the proposed online simulator HC running on CPU is the fastest. It is about \\(5\\times\\) faster than GPU-HC since we only track one solution, \\(17\\times\\) faster than Gröbner Bases and about \\(40\\times\\) faster than CPU-HC. Fig. 4c Error statistics of simulator HC with respect to different noise levels ranging from 0.2 pixel to 1.0 pixel." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.369, + 0.47, + 0.491 + ], + "angle": 0, + "content": "the average accuracy of successful cases, results on translation and scale are left in supplementary. Overall, simulator HC embedded into RANSAC can easily handle up to \\(40\\%\\) outliers as Gröbner Bases and much faster, which suggests high practicality. Note that Gröbner Bases provides around 46 real solutions each time, which requires a significant amount of additional time for solution selection. Meanwhile, simulator HC does not require this extra step." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.503, + 0.304, + 0.52 + ], + "angle": 0, + "content": "7.4. Real World Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.528, + 0.47, + 0.651 + ], + "angle": 0, + "content": "We further explore the performance of simulator HC on the real dataset LaMAR [50]. As illustrated in Fig. 5, the provided HoloLens data is a natural application of GRPS with identity scale, which can be viewed as a multi-camera rig with 4 cameras in the front and sides. We extract the correspondences by consecutively taking a view-pair from the trajectories and ensuring the distance between each view is in \\([0.2m, 0.5m]\\) with at least 100 matches. The" + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.675, + 0.47, + 0.73 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.103, + 0.74, + 0.443, + 0.754 + ], + "angle": 0, + "content": "Figure 5. LaMAR dataset multi-camera rig visualization." + }, + { + "type": "table_caption", + "bbox": [ + 0.08, + 0.771, + 0.466, + 0.785 + ], + "angle": 0, + "content": "Table 3. LaMAR dataset RANSAC-simulatorHC Median results." + }, + { + "type": "table", + "bbox": [ + 0.079, + 0.797, + 0.486, + 0.897 + ], + "angle": 0, + "content": "
seq.ER(deg)Es(%)Et(deg)#itersTime(s)Succ. (%)
Gröbner BasesCAB1.1856.145.701372.8360
HGE0.6262.484.39591.2567
LIN0.6042.043.40310.6572
simulator HCCAB1.0543.209.163102.0868
HGE0.5641.527.191340.9084
LIN0.5632.194.32660.4486
" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.369, + 0.892, + 0.476 + ], + "angle": 0, + "content": "resulting number of pairs is 1794 (CAB), 2852 (HGE) and 739 (LIN) in total. The correspondences are matched with SIFT features using nearest neighbour matching. We also present the angular error in translation, as in practice, the scale of translation and overall scale could be unobservable, e.g. the degenerate scenario when correspondences remain on the frame captured by the same camera." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.492, + 0.892, + 0.583 + ], + "angle": 0, + "content": "Table 3 shows the median results over all three sequences of LaMAR dataset. We adopted a similar setting for RANSAC as the above section with 500 maximum iterations. The proposed solver still presents consistent performances, where simulator HC has higher success rate with faster running time than Gröbner Bases." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.602, + 0.619, + 0.618 + ], + "angle": 0, + "content": "8. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.629, + 0.892, + 0.901 + ], + "angle": 0, + "content": "The present paper introduces a new paradigm for correspondence-based geometric problem solving using an elegant three-stage combination of a regression network, an online correspondence simulator, and homotopy continuation. The regression network acts as a general solution approximator and can be trained in simulation only. While pure regression accuracy is often insufficient on its own, a key message conveyed through our results is that the accuracy is generally still good enough to simulate a consistent starting problem-solution pair that successfully enables single-solution continuation. As demonstrated on two generalized camera problems, the strategy enables high success rates on common computing hardware, and the resulting solvers prove to be a viable solution if utilized as part of a random sample and consensus scheme. In particular, we propose a successful solution to the generalized relative pose and scale problem, which is around \\(17 \\times\\) faster than traditional polynomial elimination techniques." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "27110" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.078, + 0.091, + 0.258, + 0.108 + ], + "angle": 0, + "content": "9. Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.116, + 0.47, + 0.162 + ], + "angle": 0, + "content": "The authors would like to acknowledge the funding support provided by projects 22DZ1201900, 22ZR1441300, and DFYJBJ-1 by the Natural Science Foundation of Shanghai." + }, + { + "type": "title", + "bbox": [ + 0.08, + 0.175, + 0.175, + 0.19 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.2, + 0.468, + 0.227 + ], + "angle": 0, + "content": "[1] Sameer Agarwal, Keir Mierle, and Others. Ceres solver. http://ceres-solver.org, 2022.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.229, + 0.468, + 0.257 + ], + "angle": 0, + "content": "[2] C. Aholt and L. Oeding. The ideal of the trifocal variety. Math. Comput., 83(289):2553-2574, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.259, + 0.468, + 0.313 + ], + "angle": 0, + "content": "[3] B. Amos and J. Z. Kolter. Optnet: Differentiable optimization as a layer in neural networks. In International Conference on Machine Learning, pages 136-145. PMLR, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.315, + 0.468, + 0.357 + ], + "angle": 0, + "content": "[4] D. J. Bates, A. J. Sommese, J. D. Hauenstein, and C. W. Wampler. Numerically solving polynomial systems with Bertini. Software, environments, tools, SIAM, 25, 2013. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.359, + 0.468, + 0.397 + ], + "angle": 0, + "content": "[5] S. Bhayani, Z. Kukelova, and J. Heikkilä. A sparse resultant based method for efficient minimal solvers. In CVPR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.401, + 0.468, + 0.442 + ], + "angle": 0, + "content": "[6] P. Breiding and S. Timme. Homotopycontinuation.jl: A package for homotopy continuation in Julia. In International Conference on Mathematical Software - ICMS, 2018. 2, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.444, + 0.468, + 0.486 + ], + "angle": 0, + "content": "[7] C.-H. Chien, H. Fan, A. Abdelfattah, E. Tsigaridas, S. Tomov, and B. Kimia. GPU-based homotopy continuation for minimal problems in computer vision. In CVPR, 2022. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.488, + 0.468, + 0.555 + ], + "angle": 0, + "content": "[8] C.-H. Chien, H. Fan, A. Abdelfattah, E. Tsigaridas, S. Tomov, and B. Kimia. Parallel path tracking for homotopy continuation usinggpu. In Proceedings of the International Symposium on Symbolic and Algebraic Computation, 2022. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.558, + 0.468, + 0.585 + ], + "angle": 0, + "content": "[9] D. Cox, J. Little, and D. O'Shea. Using Algebraic Geometry. Springer, 1998. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.587, + 0.468, + 0.614 + ], + "angle": 0, + "content": "[10] Y. Dai, H. Li, and L. Kneip. Rolling shutter camera relative pose: Generalized epipolar geometry. In CVPR, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.616, + 0.468, + 0.657 + ], + "angle": 0, + "content": "[11] Y. Ding, C.-H. Chien, V. Larsson, K. Åström, and B. Kimia. Minimal solutions to generalized three-view relative pose problem. In ICCV, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.659, + 0.468, + 0.713 + ], + "angle": 0, + "content": "[12] T. Duff, C. Hill, A. Jensen, K. Lee, A. Leykin, and J. Sommars. Solving polynomial systems via homotopy continuation and monodromy. IMA Journal of Numerical Analysis, 2018. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.716, + 0.468, + 0.757 + ], + "angle": 0, + "content": "[13] T. Duff, K. Kohn, A. Leykin, and T. Pajdla. PLMP - point-line minimal problems in complete multi-view visibility. In ICCV, pages 1675-1684, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.759, + 0.468, + 0.799 + ], + "angle": 0, + "content": "[14] T. Duff, K. Kohn, A. Leykin, and T. Pajdla. PL1P - point-line minimal problems under partial visibility in three views. In ECCV, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.802, + 0.468, + 0.828 + ], + "angle": 0, + "content": "[15] A. Elquresh and A. M. Elgammal. Line-based relative pose estimation. In CVPR, 2011. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.831, + 0.468, + 0.857 + ], + "angle": 0, + "content": "[16] I. Z. Emiris. A general solver based on sparse resultants, 2012. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.859, + 0.468, + 0.901 + ], + "angle": 0, + "content": "[17] R. Fabbri, T. Duff, H. Fan, M. H. Regan, D. da Costa de Pinho, E. P. Tsigaridas, C. W. Wampler, J. D. Hauenstein, P. J. Giblin, A. Kimia, B. B. Leykin, and T. Pajdla. TR-PLP" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.2, + 0.468, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "- Trifocal relative pose from lines at points. In CVPR, 2020. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.123, + 0.892, + 0.178 + ], + "angle": 0, + "content": "[18] Martin A Fischler and Robert C Bolles. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM, 24(6):381-395, 1981. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.181, + 0.892, + 0.235 + ], + "angle": 0, + "content": "[19] M. A. Fischler and R. C. Bolles. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Commun. ACM, 24(6): 381-395, 1985. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.239, + 0.892, + 0.28 + ], + "angle": 0, + "content": "[20] R. M. Haralick, C.-N. Lee, K. Ottenberg, and M. Nölle. Review and analysis of solutions of the three point perspective pose estimation problem. *IJCV*, 13(3):331--356, 1994. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.283, + 0.892, + 0.323 + ], + "angle": 0, + "content": "[21] R. Hartley and A. Zisserman. Multiple View Geometry in Computer Vision. Cambridge University Press, 2nd edition, 2004. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.327, + 0.892, + 0.354 + ], + "angle": 0, + "content": "[22] J. Heikkilä. Using sparse elimination for solving minimal problems in computer vision. In ICCV, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.357, + 0.892, + 0.385 + ], + "angle": 0, + "content": "[23] J. Hesch and S. I. Roumeliotis. A Direct Least-Squares (DLS) method for PnP. In ICCV, 2011. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.388, + 0.892, + 0.429 + ], + "angle": 0, + "content": "[24] R. J. Holt and A. N. Netravali. Uniqueness of solutions to three perspective views of four points. IEEE TPAMI, 17(3): 303--307, 1995. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.432, + 0.892, + 0.459 + ], + "angle": 0, + "content": "[25] P. Hruby, T. Duff, A. Leykin, and T. Pajdla. Learning to solve hard minimal problems. In CVPR, 2022. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.462, + 0.892, + 0.503 + ], + "angle": 0, + "content": "[26] A. Kendall and R. Cipolla. Geometric loss functions for camera pose regression with deep learning. In CVPR, pages 5974-5983, 2017. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.506, + 0.892, + 0.546 + ], + "angle": 0, + "content": "[27] J. Kileel. Minimal problems for the calibrated trifocal variety. SIAM Journal on Applied Algebra and Geometry, 1(1): 575-598, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.551, + 0.892, + 0.605 + ], + "angle": 0, + "content": "[28] Laurent Kneip and Hongdong Li. Efficient computation of relative pose for multi-camera systems. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 446-453, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.608, + 0.892, + 0.662 + ], + "angle": 0, + "content": "[29] L. Kneip, D. Scaramuzza, and R. Siegwart. A novel parametrization of the perspective-three-point problem for a direct computation of absolute camera position and orientation. In CVPR, 2011. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.666, + 0.892, + 0.707 + ], + "angle": 0, + "content": "[30] L. Kneip, R. Siegwart, and M. Pollefeys. Finding the exact rotation between two images independently of the translation. In ECCV, 2012. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.711, + 0.892, + 0.751 + ], + "angle": 0, + "content": "[31] L. Kneip, H. Li, and Y. Seo. UPnP: An optimal O(n) solution to the absolute pose problem with universal applicability. In ECCV, 2014. 1, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.755, + 0.892, + 0.796 + ], + "angle": 0, + "content": "[32] L. Kneip, C. Sweeney, and R. Hartley. The generalized relative pose and scale problem: View-graph fusion via 2d-2d registration. 2016. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.799, + 0.892, + 0.826 + ], + "angle": 0, + "content": "[33] Z. Kukelova, M. Bujnak, and T. Pajdla. Automatic generator of minimal problem solvers. In ECCV, 2008. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.829, + 0.892, + 0.869 + ], + "angle": 0, + "content": "[34] V. Larsson, K. Aström, and M. Oskarsson. Efficient solvers for minimal problems by syzygy-based reduction. In CVPR, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.873, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[35] V. Larsson, K. Aström, and M. Oskarsson. Polynomial solvers for saturated ideals. In ICCV, 2017. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "27111" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.148 + ], + "angle": 0, + "content": "[36] Viktor Larsson, Kalle Astrom, and Magnus Oskarsson. Efficient solvers for minimal problems by syzygy-based reduction. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 820-829, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.149, + 0.469, + 0.19 + ], + "angle": 0, + "content": "[37] V. Larsson, Z. Kukelova, and Y. Zheng. Making minimal solvers for absolute pose estimation compact and robust. In ICCV, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.191, + 0.469, + 0.233 + ], + "angle": 0, + "content": "[38] V. Larsson, M. Oskarsson, K. Aström, A. Wallis, Z. Kukelova, and T. Pajdla. Beyond gröbner bases: Basis selection for minimal solvers. In CVPR, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.234, + 0.469, + 0.262 + ], + "angle": 0, + "content": "[39] A. Leykin. Numerical algebraic geometry. Journal of Software for Algebra and Geometry, 3(1):5-10, 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.263, + 0.469, + 0.331 + ], + "angle": 0, + "content": "[40] Evgeniy Martyushev, Jana Vrablikova, and Tomas Pajdla. Optimizing elimination templates by greedy parameter search. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15754-15764, 2022. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.333, + 0.469, + 0.375 + ], + "angle": 0, + "content": "[41] P. Miraldo, T. Dias, and S. Ramalingam. A minimal closed-form solution for multi-perspective pose estimation using points and lines. In ECCV, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.376, + 0.469, + 0.431 + ], + "angle": 0, + "content": "[42] V. Monga, Y. Li, and Y. C. Eldar. Algorithm unrolling: Interpretable, efficient deep learning for signal and image processing. IEEE Signal Processing Magazine, 38(2):18-44, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.432, + 0.469, + 0.473 + ], + "angle": 0, + "content": "[43] R. Mur-Artal, J. M. M. Montiel, and Tardós. J. D. ORB-SLAM: A Versatile and Accurate Monocular SLAM System. 31(5), 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.475, + 0.469, + 0.503 + ], + "angle": 0, + "content": "[44] D. Niste'r. An efficient solution to the five-point relative pose problem. IEEE TPAMI, 26(6):756-770, 2004. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.504, + 0.469, + 0.544 + ], + "angle": 0, + "content": "[45] D. Niste'r and F. Schaffalitzky. Four points in two or three calibrated views: Theory and practice. *IJCV*, 67(2):211–231, 2006. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.546, + 0.469, + 0.573 + ], + "angle": 0, + "content": "[46] D. Nistér, O. Naroditsky, and J. Bergen. Visual odometry. In CVPR, 2004. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.575, + 0.469, + 0.616 + ], + "angle": 0, + "content": "[47] L. Quan, B. Triggs, and B. Mourrain. Some results on minimal euclidean reconstruction from four points. 24(3):341-348, 2006. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.617, + 0.469, + 0.658 + ], + "angle": 0, + "content": "[48] R. Raguram, O. Chum, M. Pollefeys, J. Matas, and J.-M. Frahm. SAC: A universal framework for random sample consensus. IEEE TPAMI, 35(8):2022-2038, 2013. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.66, + 0.469, + 0.715 + ], + "angle": 0, + "content": "[49] P.-E. Sarlin, A. Unagar, M. Larsson, H. Germain, C. Toft, V. Larsson, M. Pollefeys, V. Lepetit, L. Hammarstrand, F. Kahl, et al. Back to the feature: Learning robust camera localization from pixels to pose. In CVPR, 2021. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.716, + 0.469, + 0.785 + ], + "angle": 0, + "content": "[50] Paul-Edouard Sarlin, Mihai Dusmanu, Johannes L. Schonberger, Pablo Speciale, Lukas Gruber, Viktor Larsson, Ondrej Miksik, and Marc Pollefeys. LaMAR: Benchmarking Localization and Mapping for Augmented Reality. In ECCV, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.787, + 0.469, + 0.815 + ], + "angle": 0, + "content": "[51] O. Saurer, M. Pollefeys, and G. H. Lee. A minimal solution to the rolling shutter pose estimation problem. 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.816, + 0.469, + 0.843 + ], + "angle": 0, + "content": "[52] J. Schonberger and J.-M. Frahm. Structure-from-motion revisited. In CVPR, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.845, + 0.469, + 0.872 + ], + "angle": 0, + "content": "[53] R. Sheffer and A. Wiesel. PnP-Net: A hybrid Perspective-n-Point Network, 2020. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[54] N. Snavely, S. M. Seitz, and R. Szeliski. Photo tourism: exploring photo collections in 3D. 2006. 1" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.092, + 0.892, + 0.121 + ], + "angle": 0, + "content": "[55] N. Snavely, S. M. Seitz, and R. Szeliski. Modeling the world from internet photo collections. pages 189-210, 2008. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.122, + 0.892, + 0.162 + ], + "angle": 0, + "content": "[56] Andrew J Sommese, Charles W Wampler, et al. The Numerical solution of systems of polynomials arising in engineering and science. World Scientific, 2005. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.164, + 0.892, + 0.219 + ], + "angle": 0, + "content": "[57] H. Stewenius, D. Nister, M. Oskarsson, and K. Aström. Solutions to minimal generalized relative pose problems. In Workshop on Omnidirectional Vision in conjunction with ICCV, 2005. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.221, + 0.892, + 0.262 + ], + "angle": 0, + "content": "[58] H. Stewenius, C. Engels, and D. Niste'r. Recent developments on direct relative orientation. ISPRS J. of Photogrammetry and Remote Sensing, 60:284-294, 2006. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.263, + 0.892, + 0.29 + ], + "angle": 0, + "content": "[59] B. Sturmfels. Solving systems of polynomial equations, 2002. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.293, + 0.892, + 0.32 + ], + "angle": 0, + "content": "[60] J. Ventura, C. Arth, and V. Lepetit. An efficient minimal solution for multi-camera motion. In ICCV, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.322, + 0.892, + 0.361 + ], + "angle": 0, + "content": "[61] J. Verschelde. Polynomial homotopy continuation with phc-pack. ACM Commun. Comput. Algebra, 44(3/4):217—220, 2010. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.364, + 0.892, + 0.404 + ], + "angle": 0, + "content": "[62] W. Xu, H. Lan, M. C. Tsakiris, and L. Kneip. Online stability improvement of gröbner basis solvers using deep learning. 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.407, + 0.892, + 0.448 + ], + "angle": 0, + "content": "[63] J. Zhao, L. Kneip, Y. He, and J. Ma. Minimal case relative pose computation using ray-point-ray features. IEEE TPAMI, 42:1176-1190, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.45, + 0.892, + 0.49 + ], + "angle": 0, + "content": "[64] Y. Zheng and L. Kneip. A direct least-squares solution to the PnP problem with unknown focal length. In CVPR, 2016. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.493, + 0.892, + 0.533 + ], + "angle": 0, + "content": "[65] Y. Zheng, Y. Kuang, S. Sugimoto, K. Astrom, and M. Okutomi. Revisiting the pnp problem: A fast, general and optimal solution. In ICCV, 2013. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.535, + 0.892, + 0.604 + ], + "angle": 0, + "content": "[66] Yi Zhou, Connelly Barnes, Jingwan Lu, Jimei Yang, and Hao Li. On the continuity of rotation representations in neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5745-5753, 2019. 6" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.092, + 0.892, + 0.604 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "27112" + } + ] +] \ No newline at end of file diff --git a/2025/Simulator HC_ Regression-based Online Simulation of Starting Problem-Solution Pairs for Homotopy Continuation in Geometric Vision/d90c2114-786f-423a-b72c-4f3391aa43d5_origin.pdf b/2025/Simulator HC_ Regression-based Online Simulation of Starting Problem-Solution Pairs for Homotopy Continuation in Geometric Vision/d90c2114-786f-423a-b72c-4f3391aa43d5_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..7982515e5df60d55156dcd8bb942b1747249c972 --- /dev/null +++ b/2025/Simulator HC_ Regression-based Online Simulation of Starting Problem-Solution Pairs for Homotopy Continuation in Geometric Vision/d90c2114-786f-423a-b72c-4f3391aa43d5_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:57eaed78fdaab28bde805e9a2177b0970477cba979f6d6ed99e4105f5ffb2a9b +size 1300665 diff --git a/2025/Simulator HC_ Regression-based Online Simulation of Starting Problem-Solution Pairs for Homotopy Continuation in Geometric Vision/full.md b/2025/Simulator HC_ Regression-based Online Simulation of Starting Problem-Solution Pairs for Homotopy Continuation in Geometric Vision/full.md new file mode 100644 index 0000000000000000000000000000000000000000..190df1938221dbdc00d6ec8249c4b765a1e3cb99 --- /dev/null +++ b/2025/Simulator HC_ Regression-based Online Simulation of Starting Problem-Solution Pairs for Homotopy Continuation in Geometric Vision/full.md @@ -0,0 +1,347 @@ +# Simulator HC: Regression-based Online Simulation of Starting Problem-Solution Pairs for Homotopy Continuation in Geometric Vision + +Xinyue Zhang + +Zijia Dai + +Wanting $\mathbf{X}\mathbf{u}^{1}$ + +Laurent Kneip1 + +$^{1}$ Mobile Perception Lab, ShanghaiTech University, China + +# Abstract + +While automatically generated polynomial elimination templates have sparked great progress in the field of 3D computer vision, there remain many problems for which the degree of the constraints or the number of unknowns leads to intractability. In recent years, homotopy continuation has been introduced as a plausible alternative. However, the method currently depends on expensive parallel tracking of all possible solutions in the complex domain, or a classification network for starting problem-solution pairs trained over a limited set of real-world examples. Our innovation lies in a novel approach to finding solution-problem pairs, where we only need to predict a rough initial solution, with the corresponding problem generated by an online simulator. Subsequently, homotopy continuation is applied to track that single solution back to the original problem. We apply this elegant combination to generalized camera resection, and also introduce a new solution to the challenging generalized relative pose and scale problem. As demonstrated, the proposed method successfully compensates the raw error committed by the regressor alone, and leads to state-of-the-art efficiency and success rates. + +# 1. Introduction + +The solution of geometric camera calibration problems is a crucial step in many Structure-from-Motion (SfM) [21, 52, 54, 55], Visual Odometry (VO) [46], and visual Simultaneous Localization And Mapping (SLAM) [43] frameworks. Solvers developed over the years address camera resectioning [23, 29, 31, 65], two-view relative pose [30, 44, 58], pose estimation from lines [15, 41], partially calibrated cameras [64], generalized cameras [31, 41, 57, 60], rolling shutter cameras [10, 51], and many other scenarios. The problems often appear in the form of a polynomial equation system, and special techniques from the field of algebraic geometry have been used to solve them. Once an efficient solver is found, it is typically embedded into a random sampling and consensus scheme [19, 48] in order to gain robust + +ness in the presence of outliers. Hence, there is a requirement for such solvers to be efficient and embeddable into iterative schemes. + +The dominant approaches to efficiently solve systems of polynomial equations are given by the Gröbner basis method and polynomial resultants [9, 59]. Based on the Gröbner basis theory, Stewénius et al. [58] and Kukelova et al. [33] demonstrate that the calculations required to obtain a minimal set of ideal generators can be effectuated efficiently at the hand of a fixed elimination template. Kukelova et al. [33] in particular propose a solver generator to automatically discover such elimination templates. Later on, sparse resultant-based solvers [16] have been demonstrated as a powerful alternative sometimes leading to smaller elimination templates. + +While polynomial elimination techniques have led to significant progress in the solution of geometric vision problems, the methods are limited as the complexity of the template search grows uncontrollably with the degree and dimensionality of the problem at hand. Furthermore, owing to their failure to respect inequality constraints and rule out practically infeasible solutions, high degrees or dimensionalities naturally lead to an elevated number of algebraically possible solutions and—assuming that it can be found at all—large elimination templates [13, 14, 45, 63–65]. This in turn causes problems in terms of numerical stability [63]. In recent years, a different solution strategy has therefore gained popularity: Homotopy Continuation (HC) [4, 12, 61]. The method proceeds by starting from a known problem-solution pair, and then tracks its roots to the target problem at hand during a step-wise interpolation of polynomial coefficients. The method has been used to successfully solve challenging tri-focal relative pose problems [17]. While promising, HC remains expensive as it generally requires the parallel continuation of many roots in the complex domain in order to identify all real solutions to a target problem, followed up by disambiguation. + +Recently, Hruby et al. [25] propose an efficient, learning-based extension to the pure HC paradigm, which serves as + +our primary motivation. They propose the introduction of a classifier to pick a good problem-solution pair trained over a large set of known pairs obtained from SfM. This enables a highly efficient application of HC, as only a single root of interest needs to be continued. While still presenting a margin for improvement in terms of success rate, the method is demonstrated to have unprecedented computational efficiency on the challenging three-view-four-point problem. + +The method proposed in this work is similar to the work of Hruby et al. [25] in that it leverages a learning-based approach to produce a starting problem-solution pair that enables single-root-tracking. However, we note a couple of important differences: + +- We do not train a classifier on a fixed set of problem-solution pairs taken from a real-world dataset. Instead, we propose to employ a solution regression network trained over an arbitrarily large set of simulated input correspondences for the considered geometric problem. The network is hence not limited by a finite set of candidate problems, and aims at being a general solution approximator for any instance of the polynomial problem. What's more, learning a solution is much easier than learning a proper problem-solution pair. +- The bridge to HC is formed by appending an online simulator that uses the input correspondences and the predicted solution to generate a complete, consistent problem-solution pair. Given only moderate regression accuracy, the produced problem-solution pair is sufficiently close to the original problem for successful single-root tracking via HC. +- We argue that the prediction of a single solution (or the unambiguous classification of a single problem-solution pair) from a minimal set may not always be possible as for complicated geometric problems, there may indeed be multiple geometrically feasible solutions. We therefore add one additional correspondence, which leads to a performance gain during both the regression and the continuation stage. + +The newly proposed simulate-and-solve paradigm is summarized in Figure 1. We demonstrate that this strategy can lead to high success-rate and efficient CPU-based solvers for two problems: generalized camera resection, and generalized relative pose and scale. Owing to its challenging nature, the latter problem has thus far only seen an optimization-based solution [32]. The solvers are compared against a purely learning-based alternative, a regressor followed by simple local refinement, as well as the exhaustive application of homotopy continuation over all roots. The proposed method is efficient, generalizable, and highly successful, and should thus be considered as an interesting alternative for the ongoing development of complicated geometric problem solvers. + +# 2. Literature Review + +Polynomial elimination theory: Our contribution can be regarded as a novel alternative to traditional polynomial elimination theory [9, 59]. Stewenius et al. [58] manually derive one of the first Gröbner basis solvers in the field of computer vision. Automatic solver generators and important extensions improving performance and numerical accuracy are later introduced by Kukelova et al. [33] and Larsson et al. [34, 35, 37, 38]. Zheng and Kneip propose the use of the parametric Sylvester resultant [64], while Hesch and Roumeliotis make use of the Macaulay resultant [23]. The more general use of sparse resultants as a powerful alternative to the Gröbner basis method is analyzed by Emiris [16], Heikkilä [22], and—most recently—Bhayani et al. [5]. While having lead to significant progress, complicated geometric problems with as many as 64 [57], 272 [45], 81 [65] (before symmetry-removal), 210 [64], $>10000$ [13, 14], and 512 [63] solutions remain hard or impossible to be solved by polynomial elimination. The latter work, in particular, presents one of the largest successful elimination templates in the literature ( $96413 \times 96879$ ) by making use of octuple machine precision. Recently Evgenity et al. [40] developed a new method to reduce the size of the elimination template. Yet, for a polynomial system, a large number of solutions in practice still leads to reduced computational efficiency. + +Homotopy continuation: Homotopy continuation implementations exist as independent programs [61] or as part of Macaulay2 [39], Bertini [4], and Julia [6]. Duff et al. [12] study the application of the method to geometric vision problems, and several related studies focus specifically on the tri-focal relative pose problem [2, 20, 24, 27, 47]. However, off-the-shelf implementations are often too general and not optimized for real-time applications. Progress is recently achieved by Fabbri et al. [17], who propose an efficient custom $\mathrm{C + + }$ implementation of the Macaulay2 solver [39] for a tri-focal relative pose problem. More recently, Ding et al. [11] propose a fast GPU-based implementation of homotopy continuation [7, 8] for a minimal solution to the three-view-four-point problem which has the same performance as Julia [6] but is significantly faster. The method leverages massive parallelization to efficiently track all solutions of a small, fixed set of starting problem-solution pairs. Our contribution can be regarded as a problem-solution pair initialization scheme similar to Hruby et al. [25] for successful single solution tracking. + +Learning for polynomial problem solving: One of the first methods to successfully employ learning in polynomial solvers is presented by Xu et al. [62], who proposed to increase the numerical accuracy of polynomial elimination templates by learning the best permutation in permutation-invariant polynomial problems. Recently, Hruby et al. [25] propose the addition of a classifier for initial problem- + +![](images/b58d09d84c67abdaa48c10197923d8f9c25a0631bf30a7fa8b9ea55410f0fc1b.jpg) +Figure 1. Overview of the proposed geometric problem solution scheme. Given input correspondences, a regression network is utilized to approximate a solution. A subsequent online simulator generates a new set of correspondences that is consistent with the regression output. The obtained problem-solution pair is finally used to bootstrap homotopy continuation. The final solution is found efficiently by tracking a single root. + +![](images/64d65d11bb933751be00e4ce3a349d00fa86c3601e492f430ed260a303b9fe92.jpg) +(a) HC multiple solution curves + +![](images/bdbe15f1c825887877d15210209f0c4a797e9486a1dfd5898d384d072adf9f77.jpg) +(b) HC solution tacking +Figure 2. Graphical illustration on homotopy continuation(HC) solution curves tracking. For a polynomial system with multiple solutions, each solution curve can be tracked independently as shown in Figure 2a. For each solution curve tracking, HC utilizes a "prediction-correction" scheme to approximate one of the solutions for the target system. + +solution pair picking in homotopy continuation. In recent years, several works have proposed to directly apply neural networks to regress the pose of a camera, either directly from images [26, 49] or—aiming at replacing traditional algebraic-geometric solvers—from correspondence coordinates [53]. While such networks use algorithm unrolling [42] and differential optimization [3] for accurate end-to-end trainability, a successful application to more complicated problems remains yet to be demonstrated. Regressors can be regarded as general non-linear function approximators, and the challenge of using them to produce highly accurate solutions is easily explained by the potential discontinuity of root locations for continuous variations of input polynomial coefficients. Our contribution can be regarded as a new post-regression refinement scheme for polynomial problems enabling successful application to complicated problems for which it is hard to apply traditional polynomial solvers. + +# 3. Polynomial Solver: Homotopy Continuation + +We start with a brief summary of HC. We refer to an instance of a certain polynomial system with specific coefficients as a problem. HC is a powerful tool that can numerically find the solutions of a problem. Each solution is found by tracking a curve independently, as illustrated in + +Figure 2a. When tracking the solution curves, the method solves a series of problems of similar form. It starts from a trivial problem whose solutions are simply known, and then gradually deforms that problem to the target problem that we are interested in but for which we do not know the solutions. It is hoped that—by continuing the known solutions throughout the problem sequence—they can lead to the solutions of the target, input problem. Two questions arise: 1) How to track the solution curves?, and 2) What is a good starting problem leading to the solutions we desire? + +The answer to the first question is given by HC. Formally, suppose we have a polynomial system $\mathbf{F}(\mathbf{x}) = \mathbf{0}$ and we are interested in its solutions. Suppose further that we have a polynomial system of similar form $\mathbf{G}(\mathbf{x}) = \mathbf{0}$ whose solutions are known. HC solves a series of polynomial systems of the form + +$$ +\mathbf {H} (\mathbf {x}, t) = (1 - t) \mathbf {G} (\mathbf {x}) + t \mathbf {F} (\mathbf {x}), \quad t \in [ 0, 1 ]. \tag {1} +$$ + +As illustrated in Figure 2b, each $\mathbf{H}(\mathbf{x},t) = \mathbf{0}$ is solved while sliding $t$ from 0 to 1. For $t = 0$ , the solution is known as $\mathbf{H}(\mathbf{x},0) = \mathbf{G}(\mathbf{x})$ . A solution of $\mathbf{F}(\mathbf{x}) = \mathbf{0}$ is finally obtained as for $t = 1$ , $\mathbf{H}(\mathbf{x},1) = \mathbf{F}(\mathbf{x})$ . The problems are not actually solved but the solutions are simply tracked from problem to problem. Taking the partial derivative to both sides of $\mathbf{H}(\mathbf{x},t)$ , the solution curve $\mathbf{x}(t)$ can be formulated as the following ordinary differential equation + +$$ +\mathbf {J} _ {\mathbf {H}} (\mathbf {x}) \frac {\partial \mathbf {x}}{\partial t} + \frac {\partial \mathbf {H} (\mathbf {x} , t)}{\partial t} = \mathbf {0} \Longleftrightarrow \frac {\partial \mathbf {x}}{\partial t} = - \mathbf {J} _ {\mathbf {H}} (\mathbf {x}) ^ {\dagger} \frac {\partial \mathbf {H} (\mathbf {x} , t)}{\partial t}. \tag {2} +$$ + +A "prediction-correction" scheme—as widely utilized in numerical analysis—is typically employed to trace the solution curve $\mathbf{x}(t)$ . + +Prediction. Denote the stepsize of $t$ as $\delta t$ . Euler's predictor can be used to update along the tangent direction + +$$ +\mathbf {x} _ {i + 1} = \mathbf {x} _ {i} + \delta t \frac {\partial \mathbf {x}}{\partial t} \bigg | _ {\mathbf {x} _ {i}, t _ {i}}. \tag {3} +$$ + +Practically speaking, a higher-order generalized method is + +commonly used for better accuracy and stability with larger stepsize. + +Correction. Gauss-Newton steps are adopted to correct the prediction. In each iteration of Gauss-Newton's method, the solution is updated by + +$$ +\mathbf {x} _ {j} = \mathbf {x} _ {j - 1} - \mathbf {J} _ {\mathbf {H}} \left(\mathbf {x} _ {j - 1}\right) ^ {\dagger} \mathbf {H} \left(\mathbf {x} _ {j - 1}, t\right), \tag {4} +$$ + +where $\mathbf{J}_{\mathbf{H}}^{\dagger}$ is the pseudo inverse of the jacobian $\mathbf{J}_{\mathbf{H}}$ , and $t = t_i + \delta t$ . Figure 2b illustrates a 1D solution curve tracking, where the blue arrow represents the prediction update and the red arrow represents the correction update. + +The remainder of this paper—our key contribution—addresses the second question of how to find a good starting system. Traditional methods (e.g. Ab initial [17], monodromy [6, 12]) rely on algebraic geometry theories developed for coefficient-parameter homotopies, and tracking all roots from a small set of fixed, known problem-solution pairs in the complex domain may often enable a solution [56]. However, tracking all the isolated paths is expensive and in practice multiple solutions cost more time to select the best. A recent study proposes the use of a classifier to pick a good starting pair with as high as possible success probability upon tracking only a single root [25], which is much faster. Similarly, in this work, we propose a novel geometry-informed strategy that combines a neural regressor with an online simulator to produce a problem-solution pair and then track a single root by HC. Even on challenging problems for which no traditional solver has been found, it can find the solution with a success rate of over $95\%$ . + +Remark (Homotopy Types). Since being introduced in 1980s, a variety of homotopies have been developed to serve the same purpose: how to solve the problem exactly as expected. The representative one would be coefficient-parameter homotopy. Its formulation provides the insight that if we can find all the solutions to a general coefficient-parameter polynomial, then we can find all the other members of that polynomial family. In fact, as pointed out in [56], in engineering problems (e.g. geometry vision problems) we are dealing with natural coefficient-parameter polynomials, that is coefficients of polynomials derived from visual features. Since the scope of this paper lies on a novel starting problem-solution pair finding and it is not highly dependent on the path tracking, we kindly refer the reader to [56] for an in-depth discussion of other homotopy types. As shown in the later applications, straight segment homotopies provide near-perfect success rates when working with our simulator. We also want to note that all the other homotopy types could be equally applied in conjunction with our simulator. + +# 4. Find a Good Starting Solution: Regression Network + +Following recent developments in deep learning, multiple attempts have been made to, for example, learn camera pose using a neural network [26, 49, 53]. The first two methods extract features from images and points and use them in a regression network to predict the pose. The third method is a more preliminary approach and aims at predicting pose directly from geometric correspondences, which is also the aim of the present work. The bottleneck in regressing pose from correspondences arises from two concerns: 1) For simple geometric problems, we often have traditional methods that are efficient and accurate, and 2) For hard geometric problems, regression networks demonstrate a poor ability to predict accurate solutions. Our key insight is that while the predicted pose may not be sufficiently accurate, it can still serve as a good starting solution for HC. + +Our pose regression networks are designed with multiple 1D convolutional layers and each layer is followed by batch normalization and ReLU activations. The loss function is given by a combination of relative pose errors with learnable weights [26]. + +# 5. Find a Consistent Starting Problem: Online Simulator + +We recall the main idea of HC. In order to continue a solution to the target problem, an appropriate starting problem with a known solution is required. Given the approximate solution predicted by the pose regression network, our goal is to derive a consistent polynomial system for which one of the solutions matches the predicted one. We denote this step an online simulator, and—though problem specific—effective realizations can almost always be found by simple geometric rules. For example, in camera resectioning, one can simply use the input 3D points and reproject them into a camera view under the predicted pose. This will lead to new 2D points, and thus a new set of correspondences that is perfectly consistent with the predicted solution, i.e. our sought starting problem-solution pair. While not yet made use of in the present work, it is worthwhile noting that—owing to their algebraic simplicity—such an online simulator often appears to be a differentiable module. + +# 6. Generalized Perspective-n-Points + +Our first concrete, practical application targets generalized camera pose estimation from 2D-3D correspondences. We use the Unified Perspective-n-Point (UPnP) algorithm [31] which formulates a polynomial system corresponding to the first-order optimality conditions of the sum of squared object space distances. The method can be used to solve both central and non-central camera resectioning problems from + +![](images/3a96df52960666700862ffecdc5039b230a9d1c53e40b69a24c2f285efa06d4f.jpg) +(a) Generalized Absolute Pose + +![](images/1288a391fc920f245307f7bfeb8287a9244de2ba1a51b68bec847baa32e9fdd9.jpg) +(b) Generalized Relative Pose and Scale +Figure 3. The geometry of the generalized camera problems we propose to solve by online simulator HC. $w$ represents the world frame, and $c$ and $c'$ represent camera frames in different views respectively. + +an arbitrary number of points (including minimal). It can be efficiently solved by the Gröbner Bases method since—after variable elimination—it only involves 4 unknowns. In this section, we solve UPnP by our newly proposed method and compare it against conventional approaches, a raw regression network, as well as a parallel HC implementation tracking all roots. + +# 6.1. Problem Formulation + +Consider a 3D-2D generalized correspondence $(\mathbf{p}_i, \mathbf{f}_i, \mathbf{v}_i)$ originating from the non-central camera resectioning model illustrated in Figure 3a, where $\mathbf{p}_i \in \mathbb{R}^3$ , $\mathbf{f}_i \in \mathbb{S}^2$ and $\mathbf{v}_i \in \mathbb{R}^3$ denote the 3D world point, image ray and ray origin, respectively. The correspondence follows an underlying generalized absolute pose transformation depending on the rotation $\mathbf{R} \in SO(3)$ , the position $\mathbf{t} \in \mathbb{R}^3$ and an unknown latent depth $\alpha_i$ as follows: + +$$ +\alpha_ {i} \mathbf {f} _ {i} + \mathbf {v} _ {i} = \mathbf {R p} _ {i} + \mathbf {t}. \tag {5} +$$ + +Let $\mathbf{M} \in \mathbb{R}^{11 \times 11}$ denote a coefficient matrix that embeds all known variables (point measurements and ray offsets), $\mathbf{q} = [q_0, q_1, q_2, q_3]^\top$ denote the quaternion representation of $\mathbf{R}$ , and $\mathbf{s} \in \mathbb{R}^{11}$ denote a vector whose elements are all second and zeroth order monomials of the quaternion variables. After translation and depth elimination, $\mathbf{q}$ can be constrained by first-order optimality conditions appearing in the form of the polynomial system1 + +$$ +\left\{ \begin{array}{l l} \mathbf {s} ^ {\top} \mathbf {M} \frac {\partial \mathbf {s}}{\partial q _ {i}} & = 0, \quad i = 0, \dots , 3 \\ \mathbf {q} ^ {\top} \mathbf {q} - 1 & = 0. \end{array} \right. \tag {6} +$$ + +# 6.2. Regressor and online simulator + +To provide a proper problem-solution pair for HC, we train a pose regression network from synthetic data. Its input is given by row-wise stacking of the correspondences, i.e. + +$$ +\left[ \begin{array}{c c c} \mathbf {p} _ {1} ^ {\top} & \mathbf {f} _ {1} ^ {\top} & \mathbf {v} _ {1} ^ {\top} \\ \vdots & \vdots & \vdots \\ \mathbf {p} _ {N} ^ {\top} & \mathbf {f} _ {N} ^ {\top} & \mathbf {v} _ {N} ^ {\top} \end{array} \right], \tag {7} +$$ + +# Algorithm 1 Online Simulator HC for generalized PnP + +1: Input: correspondences $(\mathbf{p}_i,\mathbf{f}_i,\mathbf{v}_i)_{i = 1}^M$ , stepsize $\delta t$ +2: Output: poses $\hat{\mathbf{R}},\hat{\mathbf{t}}$ +3: $\tilde{\mathbf{R}},\tilde{\mathbf{t}}\gets$ Trained Regression Network((pi,fi,v)M) +4: $(\hat{\mathbf{f}}_i)_{i=1}^M \leftarrow$ Reprojection $(\hat{\mathbf{R}}, \hat{\mathbf{t}}; (\mathbf{p}_i, \mathbf{v}_i)_{i=1}^M)$ $\triangleright$ Online Simulator +5: Starting problem $\mathbf{G}(\mathbf{x})\gets (\mathbf{p}_i,\hat{\mathbf{f}}_i,\mathbf{v}_i)_{i = 1}^M$ with solution $\hat{\mathbf{R}},\hat{\mathbf{t}}$ +6: Target problem $\mathbf{F}(\mathbf{x})\gets (\mathbf{p}_i,\mathbf{f}_i,\mathbf{v}_i)_{i = 1}^M$ +7: $\mathbf{q}_{HC} \gets$ Homotopy Continuation $t\mathbf{F}(\mathbf{x}) + (1 - t)\mathbf{G}(\mathbf{x})$ +8: $\bar{\mathbf{R}},\hat{\mathbf{t}}\gets$ decompose $\mathbf{q}_{HC}$ + +where $N$ is the number of correspondences used, and the output is the quaternion vector. The network consists of 4 1D convolutional layers and a fully connected layer at the end to convert the output to a quaternion vector with the unit norm. Given the rotation predicted by the network, the translation can be recovered from correspondences and estimated rotation. We refer the reader to [31] for the details. For the online simulator, we can get reprojected $\hat{\mathbf{f}}_i$ from (5). Algorithm 1 summarizes all stages, including the final HC over (6). Note that in line 5 of Algorithm 1, the starting problem $\mathbf{G}(\mathbf{x})$ is constructed by replacing the coefficient matrix $\mathbf{M}$ computed by $(\mathbf{p}_i,\mathbf{f}_i,\mathbf{v}_i)$ in (6) with $\hat{\mathbf{M}}$ computed by $(\mathbf{p}_i,\hat{\mathbf{f}}_i,\mathbf{v}_i)$ . + +# 6.3. Experiments + +Data. We randomly generate 4 cameras in a cube $[-1,1]^3$ and randomly sample 16 3D points in the box $[-1,1]^2 \times [4,8]$ . 16 samples would not be a good choice for a RANSAC solver [18], but in the present context the idea is to set a moderate and notably equal challenge to all compared solvers. Furthermore, for UPnP the number of constraints appearing in the polynomial system remains unchanged as it analyses first-order optimality rather than individual incidence relations. Elements of the rotation angles are randomly generated in the interval $[- \pi /2, \pi /2]$ , respectively, and the translation is randomly generated in the unit cube. We then generate the correspondences from rays that point to a 3D point. The noise level is set to 2 pixels. To train the pose regression network, we simulate a dataset of 2400 samples, which is sufficient for reliable Simulator HC initialization. We use a 2.70GHz Intel(R) Xeon(R) CPU and an NVIDIA GeForce RTX 4090 for network training and evaluation. + +Methods. For HC, the stepsize $\delta t$ and the maximum number of iterations of Newton's method are set to 0.02 and 5, respectively. We compare the Gröbner Bases solver [31] and the popular local method Levenberg-Marquardt (LM) [1]. We also test the state-of-the-art parallel HC solver developed by [7, 8], and use both its CPU and $\mathrm{GPU}^2$ version. Given that parallel HC tracks all solutions, + +Table 1. Performance comparison of simulator HC, a traditional method, and parallel exhaustive tracking HC on noisy synthetic data. All results are averaged over 1000 trials. The success rate is computed for rotation errors with a threshold of 2 degree. + +
MethodSucc. (%)ER(deg)Et(%)Time(ms)
Gröbner Bases1000.0310.4891.254
CPU HC1000.0320.4999.796
GPU HC1000.0320.4991.002
Pred Initial327.204320.031
Pred Initial + LM5428.7222560.528
Rand Initial + Simulator HC2291.682280.231
Pred Initial + Simulator HC1000.0320.5130.316
+ +we manually pick the one with the smallest error. + +Results. The rotation error is measured by $\mathcal{E}_{\mathbf{R}} = 2\arcsin (\| \hat{\mathbf{R}} -\mathbf{R}_{gt}\| /(2\sqrt{2}))$ in degrees as suggested by [11] and the translation error is measured by the $\mathcal{E}_{\mathrm{t}} = \| \hat{\mathbf{t}} -\mathbf{t}_{gt}\| /\| \mathbf{t}_{gt}\| \times 100$ in percentage. Table 1 compares the traditional method, parallel exhaustive tracking HC, and different types of initial solutions for our simulator HC. We present the success rate as the percentage of rotation errors smaller than 2 degrees. It can be observed that on this simple task, except for the local method, traditional methods, as well as simulator HC using predicted initials, succeed in $100\%$ of trials. Meanwhile, online simulator HC benefits from the single-track strategy and turns out to be the fastest one. We also perform an ablation study on the proposed simulator HC, using both predicted and random initializations. Although the pose learned by the regressor is not highly accurate and only slightly outperforms random poses, it serves as a reasonable starting solution for our simulator HC. In contrast, LM as a local method is unable to leverage the potential of the predicted initializations, as the simulator HC does. + +# 7. Generalized Relative Pose and Scale + +Our final application targets a hard minimal problem: Generalized Relative Pose and Scale (GRPS) estimation from 2D-2D correspondences [32]. The GRPS problem is a natural extension of the generalized relative pose problem and arises in point-less registration of partial view-graphs in structure-from-motion. Owing to the fact that SfM results are generally up to scale, an inherent scale factor needs to be resolved. Such problem was first introduced by Kneip et al. [32]. It has 140 solutions if a unique rotation representation is used, and can be solved by multi-start optimization to overcome local minima. Later on, a traditional polynomial solver based on the Gröbner Bases method has been proposed [36, 40]. With a $144 \times 284$ elimination template, the solver finds all 140 rotations within reasonable time. In this section, we solve GRPS by our proposed method and compare it against the existing approach, as well as a parallel HC implementation tracking all roots. + +# 7.1. Problem Formulation + +Consider a 2D-2D generalized correspondence $(\mathbf{f}_i,\mathbf{v}_i) - (\mathbf{f}_i',\mathbf{v}_i')$ observed from two non-central camera views as illustrated in Figure 3b. $\mathbf{f}_i\in \mathbb{S}^2,\mathbf{v}_i\in \mathbb{R}^3$ denote the image ray and ray origin in the first view, and $\mathbf{f}_i'\in \mathbb{S}^2,\mathbf{v}_i'\in \mathbb{R}^3$ are the corresponding entities in the second view. The correspondence follows an underlying generalized relative pose and scale incidence relation depending on the rotation $\mathbf{R}\in SO(3)$ , the position $\mathbf{t}\in \mathbb{R}^3$ , the scale $s$ , and unknown latent depths $\alpha_{i}$ and $\alpha_{i}'$ as follows: + +$$ +\alpha_ {i} \mathbf {f} _ {i} + \mathbf {v} _ {i} = \mathbf {R} \left(\alpha_ {i} ^ {\prime} \mathbf {f} _ {i} ^ {\prime} + s \mathbf {v} _ {i} ^ {\prime}\right) + \mathbf {t}. \tag {8} +$$ + +With the help of Plücker coordinates and the scalar triple product $^3$ , the unknown depths get eliminated which results in a polynomial constraint that only involves the unknown poses + +$$ +\mathbf {t} ^ {\top} \left(\mathbf {f} _ {i} \times \mathbf {R} \mathbf {f} _ {i} ^ {\prime}\right) - s \mathbf {f} _ {i} ^ {\top} \mathbf {R} \left[ \mathbf {v} _ {i} ^ {\prime} \right] _ {\times} \mathbf {f} _ {i} ^ {\prime} + \mathbf {f} _ {i} ^ {\top} \left[ \mathbf {v} _ {i} \right] _ {\times} \mathbf {R} \mathbf {f} _ {i} ^ {\prime} = 0. \tag {9} +$$ + +Therefore, a minimal formulation for the GRPS problem can be devised by a polynomial system formed by stacking 7 constraints in the form (9), each one depending on one correspondence. + +# 7.2. Regressor and online simulator + +As introduced in section 6.2, we continue to use a pose regression network that inputs a matrix with row-wise stacked correspondences. The output is given by $\mathbf{R}$ , $\mathbf{t}$ , $s$ . Note that the rotation is predicted in continuous 6D representation which is more suitable for neural networks [66]. The network consists of 11 1D convolutional layers, in the end, it branches to 3 sub-branches for $\mathbf{R}$ , $\mathbf{t}$ and $s$ . Each sub-branch consists of 2 1D convolutional layers and 1 fully connected layers for adjust output. Given the poses predicted by the network, we still need an estimation of the depths before simulating the reprojected image ray. With estimated $\hat{\mathbf{R}}$ , $\hat{\mathbf{t}}$ and $\hat{s}$ , the corresponding depths can be obtained from (8) by following + +$$ +\left[ \mathbf {f} _ {i} - \hat {\mathbf {R}} \mathbf {f} _ {i} ^ {\prime} \right] \left[ \begin{array}{l} \alpha_ {i} \\ \alpha_ {i} ^ {\prime} \end{array} \right] = - \left(\mathbf {v} _ {i} - \hat {s} \hat {\mathbf {R}} \mathbf {v} _ {i} ^ {\prime} - \hat {\mathbf {t}}\right), \tag {10} +$$ + +which can be simply solved by linear least squares for each input correspondence. Finally, we simulate a pose-consistent reprojected image ray in the first view using + +$$ +\hat {\mathbf {f}} _ {i} = \frac {1}{\hat {\alpha} _ {i}} \hat {\mathbf {R}} \left(\hat {\alpha} _ {i} ^ {\prime} \mathbf {f} _ {i} ^ {\prime} + \hat {s} \mathbf {v} _ {i} ^ {\prime}\right) + \hat {\mathbf {t}} - \mathbf {v} _ {i}. \tag {11} +$$ + +# 7.3. Simulation Experiments + +Data. We adopt a similar data generation method as in section 6.3. We generate 7 3D points in a $[-1,1]^2 \times [2,20]$ + +volume. Two view frames are generated with origins picked from $[-1, 1]^3$ . Within each view frame, we then sample individual camera offsets from the volume $[-1, 1]^3$ . The scale is randomly generated in $[0.1, 5.0]$ , and all other parameters related to noise addition are left unchanged. Note that the ground truth relative pose needs to be extracted from the simulated pair of view poses. + +Methods. The stepsize for homotopy continuation is set to 0.05 in this scenario. We compare against the Eigen solver [32], Gröbner Bases solver [40] and the CPU and GPU versions of parallel HC which tracks 140 solutions found by Monodromy [6]. Note that to measure numerical errors, we just pick the best solution with the smallest error. In practice, traditional HC needs to evaluate additional information to select a single solution (e.g. euclidean distance). For a complicated problem like GRPS, evaluating 140 solutions would still induce non-neglectable time-consumption. Since GRPS is a hard problem, we train the regressor using a larger simulated dataset of size 64000. + +Metric. We measure the rotation error in degrees and the relative translation as in Sec. 6.3 and scale error in percentage using + +$$ +\mathcal {E} _ {s} = \left| \hat {s} - s _ {g t} \right| / s _ {g t} \times 1 0 0. \tag {12} +$$ + +To calculate the success rate, for rotation errors the threshold is set to 2 deg, and for relative errors, it is set to $5\%$ . + +Numerical Stability. We randomly generate 1000 trials to test the numerical stability of all solvers. During the analysis of the performance of simulator HC using the minimum number of correspondences, we observed that our method has around $70\%$ success rate. By checking the failure cases, we found that sometimes simulator HC tracked another real solution that was different from the ground truth pose. To tackle that issue, we evaluate the addition of one more correspondence and solve for an overdetermined polynomial system. By using 8 correspondences, simulator HC reaches a success rate of $95\%$ . Again, since CPU/GPU-HC is designed to handle square systems, only the minimum number of correspondences can be tested for other HC methods. + +Since the errors in rotation, translation and scale have similar density results, we only present the rotation error in the main paper and the rest in the supplementary. As illustrated in Figure 4a, the minimal HC solvers all present similar error distribution and around $75\%$ of trials are successful. While Gröbner Bases shows $100\%$ success rate, Simulator HC over 8pts presents similar numerical errors. + +Time Comparison. The main computational benefit of the proposed online simulator HC comes from single solution-tracking, which can be effectuated efficiently even on the CPU. Figure 4b demonstrates the efficiency gain of our 8pts simulator HC over the eigen solver and parallel HC variants. As expected, by tracking only one solution curve, we are significantly faster than parallel exhaustive tracking HC, even if the latter runs on GPU. The eigen solver + +is an optimization-based solver with high complexity and ends up being slowest. The Gröbner Bases solver dealing with the large elimination template still costs a considerable time, even for just computing solutions for rotation, and not translation or scale. + +Remark. We would like to note that Gröbner Bases is solving a different formulation from Eq. (9). In order to make it easy to be solved by a smaller elimination template, it eliminates translation and scale [57]. But the resulting coefficient matrix is $130 \times 275$ and still needs a large number of operations. Although the formulation of Eq. (9) with eliminated depths faces a slightly lower success rate, its number of operations is much smaller than Gröbner Bases. On the other hand, the proposed 8pt simulator HC provides a comparably high success rate as Gröbner Bases and is still $17 \times$ faster. + +Table 2. RANSAC experiments for the GRPS problem using Gröbner Bases | simulator HC. Image points are randomly perturbed by 2-pixel uniform noise. We use vanilla RANSAC with at most 200 iterations. As can be observed, outlier ratios of up to $40\%$ are easily handled within 200 iterations. And regressor-based simulator HC is consistently faster than Gröbner Bases. + +
RatioSucc. (%)ER(deg)# ItersTime(s)
10%99.51000.100.09931252.330.85
20%99.41000.110.101711934.371.31
30%99.41000.130.142002005.101.35
40%98.9910.210.292002004.591.35
+ +Noise Resilience. Figure 4c presents the performance of the proposed 8pts simulator HC with different numbers of cameras under increasing image noise levels ranging from 0.2 pixel to 1.0 pixel. The error behaves gracefully as a function of noise, and it can be observed that a higher number of cameras leads to slightly inferior noise. Furthermore, we also observe that rotation errors appear to be less sensitive to image noise than translation and scale errors. + +Evaluation inside RANSAC. To conclude, we embed simulator HC into RANSAC to test its practical usefulness. We synthesize 200 correspondences between 5 cameras and replace part of them with outliers. Uniformly sampled pixel noise from $[-2,2]$ is added to both coordinates of the image points by again assuming a virtual focal length of 800 px. We terminate RANSAC if it reaches $99\%$ confidence or 200 iterations, and the inlier threshold is set to 0.01 for the distance measured by the normalized reprojection error. Given that Gröbner Bases solver estimates 140 solutions, we keep the real ones and select the best solution using reprojection errors in parallel to save time for it. Table 2 shows the results. The success rate is measured by a rotation error threshold of 2 degrees and the indicated errors are + +![](images/45ed5eb77a4209965faf887b0475a13238408c8420f2cc0948dd2bceb9812326.jpg) +(a) Error distribution over noise-free data. + +![](images/a77d11a9c7f3b97eb41fb553e927e3e14c55e7b36f9afddcdc3d7665d63ae342.jpg) +(b) The boxplot of running time comparison. + +![](images/6c5e8c963e11700af2d2282864c9fd2eb43e1d380071f90354aed29cbf794556.jpg) +(c) Error statistics of simulator HC. +Figure 4. Experimental results on GRPS. Fig. 4a Error distribution over 1000 trials on noise-free data. The camera number is set to be 3. Considering the rotation error, except for Gröbner Bases, the other minimal solvers all present a similar error distribution with around $70\%$ success rate, and the eigen solver presents only around $60\%$ success rate. The proposed 8pts simulator HC presents $96.3\%$ success rate where the Gröbner Bases has $100\%$ success rate. Fig. 4b The boxplot of running time comparison for the proposed simulator HC and other methods, where the number represents the median time. As expected, on average the proposed online simulator HC running on CPU is the fastest. It is about $5\times$ faster than GPU-HC since we only track one solution, $17\times$ faster than Gröbner Bases and about $40\times$ faster than CPU-HC. Fig. 4c Error statistics of simulator HC with respect to different noise levels ranging from 0.2 pixel to 1.0 pixel. + +the average accuracy of successful cases, results on translation and scale are left in supplementary. Overall, simulator HC embedded into RANSAC can easily handle up to $40\%$ outliers as Gröbner Bases and much faster, which suggests high practicality. Note that Gröbner Bases provides around 46 real solutions each time, which requires a significant amount of additional time for solution selection. Meanwhile, simulator HC does not require this extra step. + +# 7.4. Real World Experiments + +We further explore the performance of simulator HC on the real dataset LaMAR [50]. As illustrated in Fig. 5, the provided HoloLens data is a natural application of GRPS with identity scale, which can be viewed as a multi-camera rig with 4 cameras in the front and sides. We extract the correspondences by consecutively taking a view-pair from the trajectories and ensuring the distance between each view is in $[0.2m, 0.5m]$ with at least 100 matches. The + +![](images/18ecec1634171e8a7d5ba99f673cc4d24c41f8c6ed96af9e97ed16933d20d1e3.jpg) +Figure 5. LaMAR dataset multi-camera rig visualization. + +Table 3. LaMAR dataset RANSAC-simulatorHC Median results. + +
seq.ER(deg)Es(%)Et(deg)#itersTime(s)Succ. (%)
Gröbner BasesCAB1.1856.145.701372.8360
HGE0.6262.484.39591.2567
LIN0.6042.043.40310.6572
simulator HCCAB1.0543.209.163102.0868
HGE0.5641.527.191340.9084
LIN0.5632.194.32660.4486
+ +resulting number of pairs is 1794 (CAB), 2852 (HGE) and 739 (LIN) in total. The correspondences are matched with SIFT features using nearest neighbour matching. We also present the angular error in translation, as in practice, the scale of translation and overall scale could be unobservable, e.g. the degenerate scenario when correspondences remain on the frame captured by the same camera. + +Table 3 shows the median results over all three sequences of LaMAR dataset. We adopted a similar setting for RANSAC as the above section with 500 maximum iterations. The proposed solver still presents consistent performances, where simulator HC has higher success rate with faster running time than Gröbner Bases. + +# 8. Conclusion + +The present paper introduces a new paradigm for correspondence-based geometric problem solving using an elegant three-stage combination of a regression network, an online correspondence simulator, and homotopy continuation. The regression network acts as a general solution approximator and can be trained in simulation only. While pure regression accuracy is often insufficient on its own, a key message conveyed through our results is that the accuracy is generally still good enough to simulate a consistent starting problem-solution pair that successfully enables single-solution continuation. As demonstrated on two generalized camera problems, the strategy enables high success rates on common computing hardware, and the resulting solvers prove to be a viable solution if utilized as part of a random sample and consensus scheme. In particular, we propose a successful solution to the generalized relative pose and scale problem, which is around $17 \times$ faster than traditional polynomial elimination techniques. + +# 9. Acknowledgments + +The authors would like to acknowledge the funding support provided by projects 22DZ1201900, 22ZR1441300, and DFYJBJ-1 by the Natural Science Foundation of Shanghai. + +# References + +[1] Sameer Agarwal, Keir Mierle, and Others. Ceres solver. http://ceres-solver.org, 2022.5 +[2] C. Aholt and L. Oeding. The ideal of the trifocal variety. Math. Comput., 83(289):2553-2574, 2014. 2 +[3] B. Amos and J. Z. Kolter. Optnet: Differentiable optimization as a layer in neural networks. In International Conference on Machine Learning, pages 136-145. PMLR, 2017. 3 +[4] D. J. Bates, A. J. Sommese, J. D. Hauenstein, and C. W. Wampler. Numerically solving polynomial systems with Bertini. Software, environments, tools, SIAM, 25, 2013. 1, 2 +[5] S. Bhayani, Z. Kukelova, and J. Heikkilä. A sparse resultant based method for efficient minimal solvers. In CVPR, 2020. 2 +[6] P. Breiding and S. Timme. Homotopycontinuation.jl: A package for homotopy continuation in Julia. In International Conference on Mathematical Software - ICMS, 2018. 2, 4, 7 +[7] C.-H. Chien, H. Fan, A. Abdelfattah, E. Tsigaridas, S. Tomov, and B. Kimia. GPU-based homotopy continuation for minimal problems in computer vision. In CVPR, 2022. 2, 5 +[8] C.-H. Chien, H. Fan, A. Abdelfattah, E. Tsigaridas, S. Tomov, and B. Kimia. Parallel path tracking for homotopy continuation usinggpu. In Proceedings of the International Symposium on Symbolic and Algebraic Computation, 2022. 2, 5 +[9] D. Cox, J. Little, and D. O'Shea. Using Algebraic Geometry. Springer, 1998. 1, 2 +[10] Y. Dai, H. Li, and L. Kneip. Rolling shutter camera relative pose: Generalized epipolar geometry. In CVPR, 2016. 1 +[11] Y. Ding, C.-H. Chien, V. Larsson, K. Åström, and B. Kimia. Minimal solutions to generalized three-view relative pose problem. In ICCV, 2023. 2, 6 +[12] T. Duff, C. Hill, A. Jensen, K. Lee, A. Leykin, and J. Sommars. Solving polynomial systems via homotopy continuation and monodromy. IMA Journal of Numerical Analysis, 2018. 1, 2, 4 +[13] T. Duff, K. Kohn, A. Leykin, and T. Pajdla. PLMP - point-line minimal problems in complete multi-view visibility. In ICCV, pages 1675-1684, 2019. 1, 2 +[14] T. Duff, K. Kohn, A. Leykin, and T. Pajdla. PL1P - point-line minimal problems under partial visibility in three views. In ECCV, 2020. 1, 2 +[15] A. Elquresh and A. M. Elgammal. Line-based relative pose estimation. In CVPR, 2011. 1 +[16] I. Z. Emiris. A general solver based on sparse resultants, 2012. 1, 2 +[17] R. Fabbri, T. Duff, H. Fan, M. H. Regan, D. da Costa de Pinho, E. P. Tsigaridas, C. W. Wampler, J. D. Hauenstein, P. J. Giblin, A. Kimia, B. B. Leykin, and T. Pajdla. TR-PLP + +- Trifocal relative pose from lines at points. In CVPR, 2020. 1, 2, 4 +[18] Martin A Fischler and Robert C Bolles. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM, 24(6):381-395, 1981. 5 +[19] M. A. Fischler and R. C. Bolles. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Commun. ACM, 24(6): 381-395, 1985. 1 +[20] R. M. Haralick, C.-N. Lee, K. Ottenberg, and M. Nölle. Review and analysis of solutions of the three point perspective pose estimation problem. *IJCV*, 13(3):331--356, 1994. 2 +[21] R. Hartley and A. Zisserman. Multiple View Geometry in Computer Vision. Cambridge University Press, 2nd edition, 2004. 1 +[22] J. Heikkilä. Using sparse elimination for solving minimal problems in computer vision. In ICCV, 2017. 2 +[23] J. Hesch and S. I. Roumeliotis. A Direct Least-Squares (DLS) method for PnP. In ICCV, 2011. 1, 2 +[24] R. J. Holt and A. N. Netravali. Uniqueness of solutions to three perspective views of four points. IEEE TPAMI, 17(3): 303--307, 1995. 2 +[25] P. Hruby, T. Duff, A. Leykin, and T. Pajdla. Learning to solve hard minimal problems. In CVPR, 2022. 1, 2, 4 +[26] A. Kendall and R. Cipolla. Geometric loss functions for camera pose regression with deep learning. In CVPR, pages 5974-5983, 2017. 3, 4 +[27] J. Kileel. Minimal problems for the calibrated trifocal variety. SIAM Journal on Applied Algebra and Geometry, 1(1): 575-598, 2017. 2 +[28] Laurent Kneip and Hongdong Li. Efficient computation of relative pose for multi-camera systems. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 446-453, 2014. 6 +[29] L. Kneip, D. Scaramuzza, and R. Siegwart. A novel parametrization of the perspective-three-point problem for a direct computation of absolute camera position and orientation. In CVPR, 2011. 1 +[30] L. Kneip, R. Siegwart, and M. Pollefeys. Finding the exact rotation between two images independently of the translation. In ECCV, 2012. 1 +[31] L. Kneip, H. Li, and Y. Seo. UPnP: An optimal O(n) solution to the absolute pose problem with universal applicability. In ECCV, 2014. 1, 4, 5 +[32] L. Kneip, C. Sweeney, and R. Hartley. The generalized relative pose and scale problem: View-graph fusion via 2d-2d registration. 2016. 2, 6, 7 +[33] Z. Kukelova, M. Bujnak, and T. Pajdla. Automatic generator of minimal problem solvers. In ECCV, 2008. 1, 2 +[34] V. Larsson, K. Aström, and M. Oskarsson. Efficient solvers for minimal problems by syzygy-based reduction. In CVPR, 2017. 2 +[35] V. Larsson, K. Aström, and M. Oskarsson. Polynomial solvers for saturated ideals. In ICCV, 2017. 2 + +[36] Viktor Larsson, Kalle Astrom, and Magnus Oskarsson. Efficient solvers for minimal problems by syzygy-based reduction. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 820-829, 2017. 6 +[37] V. Larsson, Z. Kukelova, and Y. Zheng. Making minimal solvers for absolute pose estimation compact and robust. In ICCV, 2017. 2 +[38] V. Larsson, M. Oskarsson, K. Aström, A. Wallis, Z. Kukelova, and T. Pajdla. Beyond gröbner bases: Basis selection for minimal solvers. In CVPR, 2018. 2 +[39] A. Leykin. Numerical algebraic geometry. Journal of Software for Algebra and Geometry, 3(1):5-10, 2011. 2 +[40] Evgeniy Martyushev, Jana Vrablikova, and Tomas Pajdla. Optimizing elimination templates by greedy parameter search. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15754-15764, 2022. 2, 6, 7 +[41] P. Miraldo, T. Dias, and S. Ramalingam. A minimal closed-form solution for multi-perspective pose estimation using points and lines. In ECCV, 2018. 1 +[42] V. Monga, Y. Li, and Y. C. Eldar. Algorithm unrolling: Interpretable, efficient deep learning for signal and image processing. IEEE Signal Processing Magazine, 38(2):18-44, 2021. 3 +[43] R. Mur-Artal, J. M. M. Montiel, and Tardós. J. D. ORB-SLAM: A Versatile and Accurate Monocular SLAM System. 31(5), 2015. 1 +[44] D. Niste'r. An efficient solution to the five-point relative pose problem. IEEE TPAMI, 26(6):756-770, 2004. 1 +[45] D. Niste'r and F. Schaffalitzky. Four points in two or three calibrated views: Theory and practice. *IJCV*, 67(2):211–231, 2006. 1, 2 +[46] D. Nistér, O. Naroditsky, and J. Bergen. Visual odometry. In CVPR, 2004. 1 +[47] L. Quan, B. Triggs, and B. Mourrain. Some results on minimal euclidean reconstruction from four points. 24(3):341-348, 2006. 2 +[48] R. Raguram, O. Chum, M. Pollefeys, J. Matas, and J.-M. Frahm. SAC: A universal framework for random sample consensus. IEEE TPAMI, 35(8):2022-2038, 2013. 1 +[49] P.-E. Sarlin, A. Unagar, M. Larsson, H. Germain, C. Toft, V. Larsson, M. Pollefeys, V. Lepetit, L. Hammarstrand, F. Kahl, et al. Back to the feature: Learning robust camera localization from pixels to pose. In CVPR, 2021. 3, 4 +[50] Paul-Edouard Sarlin, Mihai Dusmanu, Johannes L. Schonberger, Pablo Speciale, Lukas Gruber, Viktor Larsson, Ondrej Miksik, and Marc Pollefeys. LaMAR: Benchmarking Localization and Mapping for Augmented Reality. In ECCV, 2022. 8 +[51] O. Saurer, M. Pollefeys, and G. H. Lee. A minimal solution to the rolling shutter pose estimation problem. 2015. 1 +[52] J. Schonberger and J.-M. Frahm. Structure-from-motion revisited. In CVPR, 2016. 1 +[53] R. Sheffer and A. Wiesel. PnP-Net: A hybrid Perspective-n-Point Network, 2020. 3, 4 +[54] N. Snavely, S. M. Seitz, and R. Szeliski. Photo tourism: exploring photo collections in 3D. 2006. 1 + +[55] N. Snavely, S. M. Seitz, and R. Szeliski. Modeling the world from internet photo collections. pages 189-210, 2008. 1 +[56] Andrew J Sommese, Charles W Wampler, et al. The Numerical solution of systems of polynomials arising in engineering and science. World Scientific, 2005. 4 +[57] H. Stewenius, D. Nister, M. Oskarsson, and K. Aström. Solutions to minimal generalized relative pose problems. In Workshop on Omnidirectional Vision in conjunction with ICCV, 2005. 1, 2, 7 +[58] H. Stewenius, C. Engels, and D. Niste'r. Recent developments on direct relative orientation. ISPRS J. of Photogrammetry and Remote Sensing, 60:284-294, 2006. 1, 2 +[59] B. Sturmfels. Solving systems of polynomial equations, 2002. 1, 2 +[60] J. Ventura, C. Arth, and V. Lepetit. An efficient minimal solution for multi-camera motion. In ICCV, 2015. 1 +[61] J. Verschelde. Polynomial homotopy continuation with phc-pack. ACM Commun. Comput. Algebra, 44(3/4):217—220, 2010. 1, 2 +[62] W. Xu, H. Lan, M. C. Tsakiris, and L. Kneip. Online stability improvement of gröbner basis solvers using deep learning. 2019. 2 +[63] J. Zhao, L. Kneip, Y. He, and J. Ma. Minimal case relative pose computation using ray-point-ray features. IEEE TPAMI, 42:1176-1190, 2020. 1, 2 +[64] Y. Zheng and L. Kneip. A direct least-squares solution to the PnP problem with unknown focal length. In CVPR, 2016. 1, 2 +[65] Y. Zheng, Y. Kuang, S. Sugimoto, K. Astrom, and M. Okutomi. Revisiting the pnp problem: A fast, general and optimal solution. In ICCV, 2013. 1, 2 +[66] Yi Zhou, Connelly Barnes, Jingwan Lu, Jimei Yang, and Hao Li. On the continuity of rotation representations in neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5745-5753, 2019. 6 \ No newline at end of file diff --git a/2025/Simulator HC_ Regression-based Online Simulation of Starting Problem-Solution Pairs for Homotopy Continuation in Geometric Vision/images.zip b/2025/Simulator HC_ Regression-based Online Simulation of Starting Problem-Solution Pairs for Homotopy Continuation in Geometric Vision/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..e5e58fe8220f8bca1dbdd118f9a975b1f2a21008 --- /dev/null +++ b/2025/Simulator HC_ Regression-based Online Simulation of Starting Problem-Solution Pairs for Homotopy Continuation in Geometric Vision/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:71ff65ae874ab5348685f8e5771cd79012b8a67847ccc2522792cecb7dc18dde +size 301090 diff --git a/2025/Simulator HC_ Regression-based Online Simulation of Starting Problem-Solution Pairs for Homotopy Continuation in Geometric Vision/layout.json b/2025/Simulator HC_ Regression-based Online Simulation of Starting Problem-Solution Pairs for Homotopy Continuation in Geometric Vision/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..acc0718ff81c290b0d82f59a00911388b23ae2c7 --- /dev/null +++ b/2025/Simulator HC_ Regression-based Online Simulation of Starting Problem-Solution Pairs for Homotopy Continuation in Geometric Vision/layout.json @@ -0,0 +1,9729 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 73, + 103, + 520, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 103, + 520, + 140 + ], + "spans": [ + { + "bbox": [ + 73, + 103, + 520, + 140 + ], + "type": "text", + "content": "Simulator HC: Regression-based Online Simulation of Starting Problem-Solution Pairs for Homotopy Continuation in Geometric Vision" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 114, + 160, + 190, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 114, + 160, + 190, + 176 + ], + "spans": [ + { + "bbox": [ + 114, + 160, + 190, + 176 + ], + "type": "text", + "content": "Xinyue Zhang" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 222, + 160, + 272, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 222, + 160, + 272, + 175 + ], + "spans": [ + { + "bbox": [ + 222, + 160, + 272, + 175 + ], + "type": "text", + "content": "Zijia Dai" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 160, + 369, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 160, + 369, + 175 + ], + "spans": [ + { + "bbox": [ + 304, + 160, + 369, + 175 + ], + "type": "text", + "content": "Wanting " + }, + { + "bbox": [ + 304, + 160, + 369, + 175 + ], + "type": "inline_equation", + "content": "\\mathbf{X}\\mathbf{u}^{1}" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 402, + 161, + 477, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 402, + 161, + 477, + 176 + ], + "spans": [ + { + "bbox": [ + 402, + 161, + 477, + 176 + ], + "type": "text", + "content": "Laurent Kneip1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 165, + 194, + 444, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 194, + 444, + 209 + ], + "spans": [ + { + "bbox": [ + 165, + 194, + 444, + 209 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 165, + 194, + 444, + 209 + ], + "type": "text", + "content": "Mobile Perception Lab, ShanghaiTech University, China" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 143, + 236, + 192, + 249 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 236, + 192, + 249 + ], + "spans": [ + { + "bbox": [ + 143, + 236, + 192, + 249 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 47, + 262, + 289, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 262, + 289, + 515 + ], + "spans": [ + { + "bbox": [ + 47, + 262, + 289, + 515 + ], + "type": "text", + "content": "While automatically generated polynomial elimination templates have sparked great progress in the field of 3D computer vision, there remain many problems for which the degree of the constraints or the number of unknowns leads to intractability. In recent years, homotopy continuation has been introduced as a plausible alternative. However, the method currently depends on expensive parallel tracking of all possible solutions in the complex domain, or a classification network for starting problem-solution pairs trained over a limited set of real-world examples. Our innovation lies in a novel approach to finding solution-problem pairs, where we only need to predict a rough initial solution, with the corresponding problem generated by an online simulator. Subsequently, homotopy continuation is applied to track that single solution back to the original problem. We apply this elegant combination to generalized camera resection, and also introduce a new solution to the challenging generalized relative pose and scale problem. As demonstrated, the proposed method successfully compensates the raw error committed by the regressor alone, and leads to state-of-the-art efficiency and success rates." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 525, + 128, + 537 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 525, + 128, + 537 + ], + "spans": [ + { + "bbox": [ + 47, + 525, + 128, + 537 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 545, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 545, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 545, + 288, + 714 + ], + "type": "text", + "content": "The solution of geometric camera calibration problems is a crucial step in many Structure-from-Motion (SfM) [21, 52, 54, 55], Visual Odometry (VO) [46], and visual Simultaneous Localization And Mapping (SLAM) [43] frameworks. Solvers developed over the years address camera resectioning [23, 29, 31, 65], two-view relative pose [30, 44, 58], pose estimation from lines [15, 41], partially calibrated cameras [64], generalized cameras [31, 41, 57, 60], rolling shutter cameras [10, 51], and many other scenarios. The problems often appear in the form of a polynomial equation system, and special techniques from the field of algebraic geometry have been used to solve them. Once an efficient solver is found, it is typically embedded into a random sampling and consensus scheme [19, 48] in order to gain robust" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 238, + 545, + 273 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 238, + 545, + 273 + ], + "spans": [ + { + "bbox": [ + 304, + 238, + 545, + 273 + ], + "type": "text", + "content": "ness in the presence of outliers. Hence, there is a requirement for such solvers to be efficient and embeddable into iterative schemes." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 277, + 546, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 277, + 546, + 421 + ], + "spans": [ + { + "bbox": [ + 304, + 277, + 546, + 421 + ], + "type": "text", + "content": "The dominant approaches to efficiently solve systems of polynomial equations are given by the Gröbner basis method and polynomial resultants [9, 59]. Based on the Gröbner basis theory, Stewénius et al. [58] and Kukelova et al. [33] demonstrate that the calculations required to obtain a minimal set of ideal generators can be effectuated efficiently at the hand of a fixed elimination template. Kukelova et al. [33] in particular propose a solver generator to automatically discover such elimination templates. Later on, sparse resultant-based solvers [16] have been demonstrated as a powerful alternative sometimes leading to smaller elimination templates." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 423, + 547, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 423, + 547, + 687 + ], + "spans": [ + { + "bbox": [ + 304, + 423, + 547, + 687 + ], + "type": "text", + "content": "While polynomial elimination techniques have led to significant progress in the solution of geometric vision problems, the methods are limited as the complexity of the template search grows uncontrollably with the degree and dimensionality of the problem at hand. Furthermore, owing to their failure to respect inequality constraints and rule out practically infeasible solutions, high degrees or dimensionalities naturally lead to an elevated number of algebraically possible solutions and—assuming that it can be found at all—large elimination templates [13, 14, 45, 63–65]. This in turn causes problems in terms of numerical stability [63]. In recent years, a different solution strategy has therefore gained popularity: Homotopy Continuation (HC) [4, 12, 61]. The method proceeds by starting from a known problem-solution pair, and then tracks its roots to the target problem at hand during a step-wise interpolation of polynomial coefficients. The method has been used to successfully solve challenging tri-focal relative pose problems [17]. While promising, HC remains expensive as it generally requires the parallel continuation of many roots in the complex domain in order to identify all real solutions to a target problem, followed up by disambiguation." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": "Recently, Hruby et al. [25] propose an efficient, learning-based extension to the pure HC paradigm, which serves as" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "27103" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 168 + ], + "type": "text", + "content": "our primary motivation. They propose the introduction of a classifier to pick a good problem-solution pair trained over a large set of known pairs obtained from SfM. This enables a highly efficient application of HC, as only a single root of interest needs to be continued. While still presenting a margin for improvement in terms of success rate, the method is demonstrated to have unprecedented computational efficiency on the challenging three-view-four-point problem." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 174, + 288, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 174, + 288, + 236 + ], + "spans": [ + { + "bbox": [ + 47, + 174, + 288, + 236 + ], + "type": "text", + "content": "The method proposed in this work is similar to the work of Hruby et al. [25] in that it leverages a learning-based approach to produce a starting problem-solution pair that enables single-root-tracking. However, we note a couple of important differences:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 240, + 287, + 540 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 47, + 240, + 287, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 240, + 287, + 361 + ], + "spans": [ + { + "bbox": [ + 47, + 240, + 287, + 361 + ], + "type": "text", + "content": "- We do not train a classifier on a fixed set of problem-solution pairs taken from a real-world dataset. Instead, we propose to employ a solution regression network trained over an arbitrarily large set of simulated input correspondences for the considered geometric problem. The network is hence not limited by a finite set of candidate problems, and aims at being a general solution approximator for any instance of the polynomial problem. What's more, learning a solution is much easier than learning a proper problem-solution pair." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 360, + 287, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 360, + 287, + 443 + ], + "spans": [ + { + "bbox": [ + 47, + 360, + 287, + 443 + ], + "type": "text", + "content": "- The bridge to HC is formed by appending an online simulator that uses the input correspondences and the predicted solution to generate a complete, consistent problem-solution pair. Given only moderate regression accuracy, the produced problem-solution pair is sufficiently close to the original problem for successful single-root tracking via HC." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 444, + 287, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 444, + 287, + 540 + ], + "spans": [ + { + "bbox": [ + 47, + 444, + 287, + 540 + ], + "type": "text", + "content": "- We argue that the prediction of a single solution (or the unambiguous classification of a single problem-solution pair) from a minimal set may not always be possible as for complicated geometric problems, there may indeed be multiple geometrically feasible solutions. We therefore add one additional correspondence, which leads to a performance gain during both the regression and the continuation stage." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 46, + 546, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 546, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 546, + 287, + 713 + ], + "type": "text", + "content": "The newly proposed simulate-and-solve paradigm is summarized in Figure 1. We demonstrate that this strategy can lead to high success-rate and efficient CPU-based solvers for two problems: generalized camera resection, and generalized relative pose and scale. Owing to its challenging nature, the latter problem has thus far only seen an optimization-based solution [32]. The solvers are compared against a purely learning-based alternative, a regressor followed by simple local refinement, as well as the exhaustive application of homotopy continuation over all roots. The proposed method is efficient, generalizable, and highly successful, and should thus be considered as an interesting alternative for the ongoing development of complicated geometric problem solvers." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 72, + 415, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 415, + 83 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 415, + 83 + ], + "type": "text", + "content": "2. Literature Review" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 91, + 547, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 91, + 547, + 390 + ], + "spans": [ + { + "bbox": [ + 304, + 91, + 547, + 390 + ], + "type": "text", + "content": "Polynomial elimination theory: Our contribution can be regarded as a novel alternative to traditional polynomial elimination theory [9, 59]. Stewenius et al. [58] manually derive one of the first Gröbner basis solvers in the field of computer vision. Automatic solver generators and important extensions improving performance and numerical accuracy are later introduced by Kukelova et al. [33] and Larsson et al. [34, 35, 37, 38]. Zheng and Kneip propose the use of the parametric Sylvester resultant [64], while Hesch and Roumeliotis make use of the Macaulay resultant [23]. The more general use of sparse resultants as a powerful alternative to the Gröbner basis method is analyzed by Emiris [16], Heikkilä [22], and—most recently—Bhayani et al. [5]. While having lead to significant progress, complicated geometric problems with as many as 64 [57], 272 [45], 81 [65] (before symmetry-removal), 210 [64], " + }, + { + "bbox": [ + 304, + 91, + 547, + 390 + ], + "type": "inline_equation", + "content": ">10000" + }, + { + "bbox": [ + 304, + 91, + 547, + 390 + ], + "type": "text", + "content": " [13, 14], and 512 [63] solutions remain hard or impossible to be solved by polynomial elimination. The latter work, in particular, presents one of the largest successful elimination templates in the literature (" + }, + { + "bbox": [ + 304, + 91, + 547, + 390 + ], + "type": "inline_equation", + "content": "96413 \\times 96879" + }, + { + "bbox": [ + 304, + 91, + 547, + 390 + ], + "type": "text", + "content": ") by making use of octuple machine precision. Recently Evgenity et al. [40] developed a new method to reduce the size of the elimination template. Yet, for a polynomial system, a large number of solutions in practice still leads to reduced computational efficiency." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 390, + 547, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 390, + 547, + 629 + ], + "spans": [ + { + "bbox": [ + 304, + 390, + 547, + 629 + ], + "type": "text", + "content": "Homotopy continuation: Homotopy continuation implementations exist as independent programs [61] or as part of Macaulay2 [39], Bertini [4], and Julia [6]. Duff et al. [12] study the application of the method to geometric vision problems, and several related studies focus specifically on the tri-focal relative pose problem [2, 20, 24, 27, 47]. However, off-the-shelf implementations are often too general and not optimized for real-time applications. Progress is recently achieved by Fabbri et al. [17], who propose an efficient custom " + }, + { + "bbox": [ + 304, + 390, + 547, + 629 + ], + "type": "inline_equation", + "content": "\\mathrm{C + + }" + }, + { + "bbox": [ + 304, + 390, + 547, + 629 + ], + "type": "text", + "content": " implementation of the Macaulay2 solver [39] for a tri-focal relative pose problem. More recently, Ding et al. [11] propose a fast GPU-based implementation of homotopy continuation [7, 8] for a minimal solution to the three-view-four-point problem which has the same performance as Julia [6] but is significantly faster. The method leverages massive parallelization to efficiently track all solutions of a small, fixed set of starting problem-solution pairs. Our contribution can be regarded as a problem-solution pair initialization scheme similar to Hruby et al. [25] for successful single solution tracking." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 630, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 630, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 630, + 547, + 715 + ], + "type": "text", + "content": "Learning for polynomial problem solving: One of the first methods to successfully employ learning in polynomial solvers is presented by Xu et al. [62], who proposed to increase the numerical accuracy of polynomial elimination templates by learning the best permutation in permutation-invariant polynomial problems. Recently, Hruby et al. [25] propose the addition of a classifier for initial problem-" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "27104" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 76, + 70, + 521, + 184 + ], + "blocks": [ + { + "bbox": [ + 76, + 70, + 521, + 184 + ], + "lines": [ + { + "bbox": [ + 76, + 70, + 521, + 184 + ], + "spans": [ + { + "bbox": [ + 76, + 70, + 521, + 184 + ], + "type": "image", + "image_path": "b58d09d84c67abdaa48c10197923d8f9c25a0631bf30a7fa8b9ea55410f0fc1b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 192, + 547, + 237 + ], + "lines": [ + { + "bbox": [ + 46, + 192, + 547, + 237 + ], + "spans": [ + { + "bbox": [ + 46, + 192, + 547, + 237 + ], + "type": "text", + "content": "Figure 1. Overview of the proposed geometric problem solution scheme. Given input correspondences, a regression network is utilized to approximate a solution. A subsequent online simulator generates a new set of correspondences that is consistent with the regression output. The obtained problem-solution pair is finally used to bootstrap homotopy continuation. The final solution is found efficiently by tracking a single root." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 52, + 243, + 156, + 313 + ], + "blocks": [ + { + "bbox": [ + 52, + 243, + 156, + 313 + ], + "lines": [ + { + "bbox": [ + 52, + 243, + 156, + 313 + ], + "spans": [ + { + "bbox": [ + 52, + 243, + 156, + 313 + ], + "type": "image", + "image_path": "64d65d11bb933751be00e4ce3a349d00fa86c3601e492f430ed260a303b9fe92.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 52, + 317, + 154, + 327 + ], + "lines": [ + { + "bbox": [ + 52, + 317, + 154, + 327 + ], + "spans": [ + { + "bbox": [ + 52, + 317, + 154, + 327 + ], + "type": "text", + "content": "(a) HC multiple solution curves" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 179, + 242, + 284, + 313 + ], + "blocks": [ + { + "bbox": [ + 179, + 242, + 284, + 313 + ], + "lines": [ + { + "bbox": [ + 179, + 242, + 284, + 313 + ], + "spans": [ + { + "bbox": [ + 179, + 242, + 284, + 313 + ], + "type": "image", + "image_path": "bdbe15f1c825887877d15210209f0c4a797e9486a1dfd5898d384d072adf9f77.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 193, + 317, + 271, + 327 + ], + "lines": [ + { + "bbox": [ + 193, + 317, + 271, + 327 + ], + "spans": [ + { + "bbox": [ + 193, + 317, + 271, + 327 + ], + "type": "text", + "content": "(b) HC solution tacking" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 335, + 288, + 403 + ], + "lines": [ + { + "bbox": [ + 46, + 335, + 288, + 403 + ], + "spans": [ + { + "bbox": [ + 46, + 335, + 288, + 403 + ], + "type": "text", + "content": "Figure 2. Graphical illustration on homotopy continuation(HC) solution curves tracking. For a polynomial system with multiple solutions, each solution curve can be tracked independently as shown in Figure 2a. For each solution curve tracking, HC utilizes a \"prediction-correction\" scheme to approximate one of the solutions for the target system." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 407, + 287, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 407, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 46, + 407, + 287, + 624 + ], + "type": "text", + "content": "solution pair picking in homotopy continuation. In recent years, several works have proposed to directly apply neural networks to regress the pose of a camera, either directly from images [26, 49] or—aiming at replacing traditional algebraic-geometric solvers—from correspondence coordinates [53]. While such networks use algorithm unrolling [42] and differential optimization [3] for accurate end-to-end trainability, a successful application to more complicated problems remains yet to be demonstrated. Regressors can be regarded as general non-linear function approximators, and the challenge of using them to produce highly accurate solutions is easily explained by the potential discontinuity of root locations for continuous variations of input polynomial coefficients. Our contribution can be regarded as a new post-regression refinement scheme for polynomial problems enabling successful application to complicated problems for which it is hard to apply traditional polynomial solvers." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 633, + 287, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 633, + 287, + 647 + ], + "spans": [ + { + "bbox": [ + 47, + 633, + 287, + 647 + ], + "type": "text", + "content": "3. Polynomial Solver: Homotopy Continuation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 653, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 714 + ], + "type": "text", + "content": "We start with a brief summary of HC. We refer to an instance of a certain polynomial system with specific coefficients as a problem. HC is a powerful tool that can numerically find the solutions of a problem. Each solution is found by tracking a curve independently, as illustrated in" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 243, + 545, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 243, + 545, + 363 + ], + "spans": [ + { + "bbox": [ + 304, + 243, + 545, + 363 + ], + "type": "text", + "content": "Figure 2a. When tracking the solution curves, the method solves a series of problems of similar form. It starts from a trivial problem whose solutions are simply known, and then gradually deforms that problem to the target problem that we are interested in but for which we do not know the solutions. It is hoped that—by continuing the known solutions throughout the problem sequence—they can lead to the solutions of the target, input problem. Two questions arise: 1) How to track the solution curves?, and 2) What is a good starting problem leading to the solutions we desire?" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 363, + 545, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 363, + 545, + 435 + ], + "spans": [ + { + "bbox": [ + 304, + 363, + 545, + 435 + ], + "type": "text", + "content": "The answer to the first question is given by HC. Formally, suppose we have a polynomial system " + }, + { + "bbox": [ + 304, + 363, + 545, + 435 + ], + "type": "inline_equation", + "content": "\\mathbf{F}(\\mathbf{x}) = \\mathbf{0}" + }, + { + "bbox": [ + 304, + 363, + 545, + 435 + ], + "type": "text", + "content": " and we are interested in its solutions. Suppose further that we have a polynomial system of similar form " + }, + { + "bbox": [ + 304, + 363, + 545, + 435 + ], + "type": "inline_equation", + "content": "\\mathbf{G}(\\mathbf{x}) = \\mathbf{0}" + }, + { + "bbox": [ + 304, + 363, + 545, + 435 + ], + "type": "text", + "content": " whose solutions are known. HC solves a series of polynomial systems of the form" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 331, + 442, + 545, + 456 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 442, + 545, + 456 + ], + "spans": [ + { + "bbox": [ + 331, + 442, + 545, + 456 + ], + "type": "interline_equation", + "content": "\\mathbf {H} (\\mathbf {x}, t) = (1 - t) \\mathbf {G} (\\mathbf {x}) + t \\mathbf {F} (\\mathbf {x}), \\quad t \\in [ 0, 1 ]. \\tag {1}", + "image_path": "5b50757140d05fa5195c3afac9010df3e495bf0b90c05ac9b2c0c48bc7ab9a9f.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 464, + 545, + 560 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 464, + 545, + 560 + ], + "spans": [ + { + "bbox": [ + 304, + 464, + 545, + 560 + ], + "type": "text", + "content": "As illustrated in Figure 2b, each " + }, + { + "bbox": [ + 304, + 464, + 545, + 560 + ], + "type": "inline_equation", + "content": "\\mathbf{H}(\\mathbf{x},t) = \\mathbf{0}" + }, + { + "bbox": [ + 304, + 464, + 545, + 560 + ], + "type": "text", + "content": " is solved while sliding " + }, + { + "bbox": [ + 304, + 464, + 545, + 560 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 464, + 545, + 560 + ], + "type": "text", + "content": " from 0 to 1. For " + }, + { + "bbox": [ + 304, + 464, + 545, + 560 + ], + "type": "inline_equation", + "content": "t = 0" + }, + { + "bbox": [ + 304, + 464, + 545, + 560 + ], + "type": "text", + "content": ", the solution is known as " + }, + { + "bbox": [ + 304, + 464, + 545, + 560 + ], + "type": "inline_equation", + "content": "\\mathbf{H}(\\mathbf{x},0) = \\mathbf{G}(\\mathbf{x})" + }, + { + "bbox": [ + 304, + 464, + 545, + 560 + ], + "type": "text", + "content": ". A solution of " + }, + { + "bbox": [ + 304, + 464, + 545, + 560 + ], + "type": "inline_equation", + "content": "\\mathbf{F}(\\mathbf{x}) = \\mathbf{0}" + }, + { + "bbox": [ + 304, + 464, + 545, + 560 + ], + "type": "text", + "content": " is finally obtained as for " + }, + { + "bbox": [ + 304, + 464, + 545, + 560 + ], + "type": "inline_equation", + "content": "t = 1" + }, + { + "bbox": [ + 304, + 464, + 545, + 560 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 464, + 545, + 560 + ], + "type": "inline_equation", + "content": "\\mathbf{H}(\\mathbf{x},1) = \\mathbf{F}(\\mathbf{x})" + }, + { + "bbox": [ + 304, + 464, + 545, + 560 + ], + "type": "text", + "content": ". The problems are not actually solved but the solutions are simply tracked from problem to problem. Taking the partial derivative to both sides of " + }, + { + "bbox": [ + 304, + 464, + 545, + 560 + ], + "type": "inline_equation", + "content": "\\mathbf{H}(\\mathbf{x},t)" + }, + { + "bbox": [ + 304, + 464, + 545, + 560 + ], + "type": "text", + "content": ", the solution curve " + }, + { + "bbox": [ + 304, + 464, + 545, + 560 + ], + "type": "inline_equation", + "content": "\\mathbf{x}(t)" + }, + { + "bbox": [ + 304, + 464, + 545, + 560 + ], + "type": "text", + "content": " can be formulated as the following ordinary differential equation" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 566, + 553, + 599 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 566, + 553, + 599 + ], + "spans": [ + { + "bbox": [ + 305, + 566, + 553, + 599 + ], + "type": "interline_equation", + "content": "\\mathbf {J} _ {\\mathbf {H}} (\\mathbf {x}) \\frac {\\partial \\mathbf {x}}{\\partial t} + \\frac {\\partial \\mathbf {H} (\\mathbf {x} , t)}{\\partial t} = \\mathbf {0} \\Longleftrightarrow \\frac {\\partial \\mathbf {x}}{\\partial t} = - \\mathbf {J} _ {\\mathbf {H}} (\\mathbf {x}) ^ {\\dagger} \\frac {\\partial \\mathbf {H} (\\mathbf {x} , t)}{\\partial t}. \\tag {2}", + "image_path": "a4a4a58bcfd3734084b90554db679595ccedd6dee86b2bba69f10410463354bb.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 600, + 545, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 600, + 545, + 635 + ], + "spans": [ + { + "bbox": [ + 304, + 600, + 545, + 635 + ], + "type": "text", + "content": "A \"prediction-correction\" scheme—as widely utilized in numerical analysis—is typically employed to trace the solution curve " + }, + { + "bbox": [ + 304, + 600, + 545, + 635 + ], + "type": "inline_equation", + "content": "\\mathbf{x}(t)" + }, + { + "bbox": [ + 304, + 600, + 545, + 635 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 635, + 545, + 660 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 635, + 545, + 660 + ], + "spans": [ + { + "bbox": [ + 304, + 635, + 545, + 660 + ], + "type": "text", + "content": "Prediction. Denote the stepsize of " + }, + { + "bbox": [ + 304, + 635, + 545, + 660 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 635, + 545, + 660 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 304, + 635, + 545, + 660 + ], + "type": "inline_equation", + "content": "\\delta t" + }, + { + "bbox": [ + 304, + 635, + 545, + 660 + ], + "type": "text", + "content": ". Euler's predictor can be used to update along the tangent direction" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 373, + 666, + 545, + 696 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 373, + 666, + 545, + 696 + ], + "spans": [ + { + "bbox": [ + 373, + 666, + 545, + 696 + ], + "type": "interline_equation", + "content": "\\mathbf {x} _ {i + 1} = \\mathbf {x} _ {i} + \\delta t \\frac {\\partial \\mathbf {x}}{\\partial t} \\bigg | _ {\\mathbf {x} _ {i}, t _ {i}}. \\tag {3}", + "image_path": "0f87d9e8bb57ae0f8d7ee0e5977fad743d6e24e3522ba996faa708bcd102c2e6.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 306, + 701, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 701, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 306, + 701, + 545, + 714 + ], + "type": "text", + "content": "Practically speaking, a higher-order generalized method is" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27105" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 96 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 96 + ], + "type": "text", + "content": "commonly used for better accuracy and stability with larger stepsize." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 100, + 287, + 137 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 100, + 287, + 137 + ], + "spans": [ + { + "bbox": [ + 46, + 100, + 287, + 137 + ], + "type": "text", + "content": "Correction. Gauss-Newton steps are adopted to correct the prediction. In each iteration of Gauss-Newton's method, the solution is updated by" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 89, + 153, + 287, + 168 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 153, + 287, + 168 + ], + "spans": [ + { + "bbox": [ + 89, + 153, + 287, + 168 + ], + "type": "interline_equation", + "content": "\\mathbf {x} _ {j} = \\mathbf {x} _ {j - 1} - \\mathbf {J} _ {\\mathbf {H}} \\left(\\mathbf {x} _ {j - 1}\\right) ^ {\\dagger} \\mathbf {H} \\left(\\mathbf {x} _ {j - 1}, t\\right), \\tag {4}", + "image_path": "c6bc2da91ae25c318e973fa65f46add04812ca07d1069c8aef36e53aac28088b.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 186, + 287, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 186, + 287, + 236 + ], + "spans": [ + { + "bbox": [ + 46, + 186, + 287, + 236 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 186, + 287, + 236 + ], + "type": "inline_equation", + "content": "\\mathbf{J}_{\\mathbf{H}}^{\\dagger}" + }, + { + "bbox": [ + 46, + 186, + 287, + 236 + ], + "type": "text", + "content": " is the pseudo inverse of the jacobian " + }, + { + "bbox": [ + 46, + 186, + 287, + 236 + ], + "type": "inline_equation", + "content": "\\mathbf{J}_{\\mathbf{H}}" + }, + { + "bbox": [ + 46, + 186, + 287, + 236 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 186, + 287, + 236 + ], + "type": "inline_equation", + "content": "t = t_i + \\delta t" + }, + { + "bbox": [ + 46, + 186, + 287, + 236 + ], + "type": "text", + "content": ". Figure 2b illustrates a 1D solution curve tracking, where the blue arrow represents the prediction update and the red arrow represents the correction update." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 239, + 287, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 239, + 287, + 453 + ], + "spans": [ + { + "bbox": [ + 46, + 239, + 287, + 453 + ], + "type": "text", + "content": "The remainder of this paper—our key contribution—addresses the second question of how to find a good starting system. Traditional methods (e.g. Ab initial [17], monodromy [6, 12]) rely on algebraic geometry theories developed for coefficient-parameter homotopies, and tracking all roots from a small set of fixed, known problem-solution pairs in the complex domain may often enable a solution [56]. However, tracking all the isolated paths is expensive and in practice multiple solutions cost more time to select the best. A recent study proposes the use of a classifier to pick a good starting pair with as high as possible success probability upon tracking only a single root [25], which is much faster. Similarly, in this work, we propose a novel geometry-informed strategy that combines a neural regressor with an online simulator to produce a problem-solution pair and then track a single root by HC. Even on challenging problems for which no traditional solver has been found, it can find the solution with a success rate of over " + }, + { + "bbox": [ + 46, + 239, + 287, + 453 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 46, + 239, + 287, + 453 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 474, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 474, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 474, + 288, + 713 + ], + "type": "text", + "content": "Remark (Homotopy Types). Since being introduced in 1980s, a variety of homotopies have been developed to serve the same purpose: how to solve the problem exactly as expected. The representative one would be coefficient-parameter homotopy. Its formulation provides the insight that if we can find all the solutions to a general coefficient-parameter polynomial, then we can find all the other members of that polynomial family. In fact, as pointed out in [56], in engineering problems (e.g. geometry vision problems) we are dealing with natural coefficient-parameter polynomials, that is coefficients of polynomials derived from visual features. Since the scope of this paper lies on a novel starting problem-solution pair finding and it is not highly dependent on the path tracking, we kindly refer the reader to [56] for an in-depth discussion of other homotopy types. As shown in the later applications, straight segment homotopies provide near-perfect success rates when working with our simulator. We also want to note that all the other homotopy types could be equally applied in conjunction with our simulator." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 71, + 545, + 97 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 71, + 545, + 97 + ], + "spans": [ + { + "bbox": [ + 306, + 71, + 545, + 97 + ], + "type": "text", + "content": "4. Find a Good Starting Solution: Regression Network" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 106, + 545, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 106, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 304, + 106, + 545, + 285 + ], + "type": "text", + "content": "Following recent developments in deep learning, multiple attempts have been made to, for example, learn camera pose using a neural network [26, 49, 53]. The first two methods extract features from images and points and use them in a regression network to predict the pose. The third method is a more preliminary approach and aims at predicting pose directly from geometric correspondences, which is also the aim of the present work. The bottleneck in regressing pose from correspondences arises from two concerns: 1) For simple geometric problems, we often have traditional methods that are efficient and accurate, and 2) For hard geometric problems, regression networks demonstrate a poor ability to predict accurate solutions. Our key insight is that while the predicted pose may not be sufficiently accurate, it can still serve as a good starting solution for HC." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 286, + 545, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 286, + 545, + 346 + ], + "spans": [ + { + "bbox": [ + 304, + 286, + 545, + 346 + ], + "type": "text", + "content": "Our pose regression networks are designed with multiple 1D convolutional layers and each layer is followed by batch normalization and ReLU activations. The loss function is given by a combination of relative pose errors with learnable weights [26]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 357, + 545, + 384 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 357, + 545, + 384 + ], + "spans": [ + { + "bbox": [ + 305, + 357, + 545, + 384 + ], + "type": "text", + "content": "5. Find a Consistent Starting Problem: Online Simulator" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 392, + 545, + 596 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 392, + 545, + 596 + ], + "spans": [ + { + "bbox": [ + 304, + 392, + 545, + 596 + ], + "type": "text", + "content": "We recall the main idea of HC. In order to continue a solution to the target problem, an appropriate starting problem with a known solution is required. Given the approximate solution predicted by the pose regression network, our goal is to derive a consistent polynomial system for which one of the solutions matches the predicted one. We denote this step an online simulator, and—though problem specific—effective realizations can almost always be found by simple geometric rules. For example, in camera resectioning, one can simply use the input 3D points and reproject them into a camera view under the predicted pose. This will lead to new 2D points, and thus a new set of correspondences that is perfectly consistent with the predicted solution, i.e. our sought starting problem-solution pair. While not yet made use of in the present work, it is worthwhile noting that—owing to their algebraic simplicity—such an online simulator often appears to be a differentiable module." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 609, + 491, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 609, + 491, + 623 + ], + "spans": [ + { + "bbox": [ + 305, + 609, + 491, + 623 + ], + "type": "text", + "content": "6. Generalized Perspective-n-Points" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 629, + 545, + 713 + ], + "type": "text", + "content": "Our first concrete, practical application targets generalized camera pose estimation from 2D-3D correspondences. We use the Unified Perspective-n-Point (UPnP) algorithm [31] which formulates a polynomial system corresponding to the first-order optimality conditions of the sum of squared object space distances. The method can be used to solve both central and non-central camera resectioning problems from" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "27106" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 71, + 156, + 148 + ], + "blocks": [ + { + "bbox": [ + 50, + 71, + 156, + 148 + ], + "lines": [ + { + "bbox": [ + 50, + 71, + 156, + 148 + ], + "spans": [ + { + "bbox": [ + 50, + 71, + 156, + 148 + ], + "type": "image", + "image_path": "3a96df52960666700862ffecdc5039b230a9d1c53e40b69a24c2f285efa06d4f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 52, + 149, + 151, + 159 + ], + "lines": [ + { + "bbox": [ + 52, + 149, + 151, + 159 + ], + "spans": [ + { + "bbox": [ + 52, + 149, + 151, + 159 + ], + "type": "text", + "content": "(a) Generalized Absolute Pose" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 179, + 72, + 286, + 148 + ], + "blocks": [ + { + "bbox": [ + 179, + 72, + 286, + 148 + ], + "lines": [ + { + "bbox": [ + 179, + 72, + 286, + 148 + ], + "spans": [ + { + "bbox": [ + 179, + 72, + 286, + 148 + ], + "type": "image", + "image_path": "1288a391fc920f245307f7bfeb8287a9244de2ba1a51b68bec847baa32e9fdd9.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 178, + 149, + 287, + 168 + ], + "lines": [ + { + "bbox": [ + 178, + 149, + 287, + 168 + ], + "spans": [ + { + "bbox": [ + 178, + 149, + 287, + 168 + ], + "type": "text", + "content": "(b) Generalized Relative Pose and Scale" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 178, + 287, + 223 + ], + "lines": [ + { + "bbox": [ + 46, + 178, + 287, + 223 + ], + "spans": [ + { + "bbox": [ + 46, + 178, + 287, + 223 + ], + "type": "text", + "content": "Figure 3. The geometry of the generalized camera problems we propose to solve by online simulator HC. " + }, + { + "bbox": [ + 46, + 178, + 287, + 223 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 46, + 178, + 287, + 223 + ], + "type": "text", + "content": " represents the world frame, and " + }, + { + "bbox": [ + 46, + 178, + 287, + 223 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 46, + 178, + 287, + 223 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 178, + 287, + 223 + ], + "type": "inline_equation", + "content": "c'" + }, + { + "bbox": [ + 46, + 178, + 287, + 223 + ], + "type": "text", + "content": " represent camera frames in different views respectively." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 224, + 287, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 224, + 287, + 308 + ], + "spans": [ + { + "bbox": [ + 46, + 224, + 287, + 308 + ], + "type": "text", + "content": "an arbitrary number of points (including minimal). It can be efficiently solved by the Gröbner Bases method since—after variable elimination—it only involves 4 unknowns. In this section, we solve UPnP by our newly proposed method and compare it against conventional approaches, a raw regression network, as well as a parallel HC implementation tracking all roots." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 314, + 171, + 325 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 314, + 171, + 325 + ], + "spans": [ + { + "bbox": [ + 47, + 314, + 171, + 325 + ], + "type": "text", + "content": "6.1. Problem Formulation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 331, + 287, + 428 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 331, + 287, + 428 + ], + "spans": [ + { + "bbox": [ + 46, + 331, + 287, + 428 + ], + "type": "text", + "content": "Consider a 3D-2D generalized correspondence " + }, + { + "bbox": [ + 46, + 331, + 287, + 428 + ], + "type": "inline_equation", + "content": "(\\mathbf{p}_i, \\mathbf{f}_i, \\mathbf{v}_i)" + }, + { + "bbox": [ + 46, + 331, + 287, + 428 + ], + "type": "text", + "content": " originating from the non-central camera resectioning model illustrated in Figure 3a, where " + }, + { + "bbox": [ + 46, + 331, + 287, + 428 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_i \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 46, + 331, + 287, + 428 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 331, + 287, + 428 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_i \\in \\mathbb{S}^2" + }, + { + "bbox": [ + 46, + 331, + 287, + 428 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 331, + 287, + 428 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_i \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 46, + 331, + 287, + 428 + ], + "type": "text", + "content": " denote the 3D world point, image ray and ray origin, respectively. The correspondence follows an underlying generalized absolute pose transformation depending on the rotation " + }, + { + "bbox": [ + 46, + 331, + 287, + 428 + ], + "type": "inline_equation", + "content": "\\mathbf{R} \\in SO(3)" + }, + { + "bbox": [ + 46, + 331, + 287, + 428 + ], + "type": "text", + "content": ", the position " + }, + { + "bbox": [ + 46, + 331, + 287, + 428 + ], + "type": "inline_equation", + "content": "\\mathbf{t} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 46, + 331, + 287, + 428 + ], + "type": "text", + "content": " and an unknown latent depth " + }, + { + "bbox": [ + 46, + 331, + 287, + 428 + ], + "type": "inline_equation", + "content": "\\alpha_i" + }, + { + "bbox": [ + 46, + 331, + 287, + 428 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 120, + 434, + 287, + 447 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 120, + 434, + 287, + 447 + ], + "spans": [ + { + "bbox": [ + 120, + 434, + 287, + 447 + ], + "type": "interline_equation", + "content": "\\alpha_ {i} \\mathbf {f} _ {i} + \\mathbf {v} _ {i} = \\mathbf {R p} _ {i} + \\mathbf {t}. \\tag {5}", + "image_path": "9f8f6b158429d7daf9b05056556f74ca65b2cb78485904b30d29fe1acd8fec60.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 453, + 287, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 453, + 287, + 548 + ], + "spans": [ + { + "bbox": [ + 46, + 453, + 287, + 548 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 46, + 453, + 287, + 548 + ], + "type": "inline_equation", + "content": "\\mathbf{M} \\in \\mathbb{R}^{11 \\times 11}" + }, + { + "bbox": [ + 46, + 453, + 287, + 548 + ], + "type": "text", + "content": " denote a coefficient matrix that embeds all known variables (point measurements and ray offsets), " + }, + { + "bbox": [ + 46, + 453, + 287, + 548 + ], + "type": "inline_equation", + "content": "\\mathbf{q} = [q_0, q_1, q_2, q_3]^\\top" + }, + { + "bbox": [ + 46, + 453, + 287, + 548 + ], + "type": "text", + "content": " denote the quaternion representation of " + }, + { + "bbox": [ + 46, + 453, + 287, + 548 + ], + "type": "inline_equation", + "content": "\\mathbf{R}" + }, + { + "bbox": [ + 46, + 453, + 287, + 548 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 453, + 287, + 548 + ], + "type": "inline_equation", + "content": "\\mathbf{s} \\in \\mathbb{R}^{11}" + }, + { + "bbox": [ + 46, + 453, + 287, + 548 + ], + "type": "text", + "content": " denote a vector whose elements are all second and zeroth order monomials of the quaternion variables. After translation and depth elimination, " + }, + { + "bbox": [ + 46, + 453, + 287, + 548 + ], + "type": "inline_equation", + "content": "\\mathbf{q}" + }, + { + "bbox": [ + 46, + 453, + 287, + 548 + ], + "type": "text", + "content": " can be constrained by first-order optimality conditions appearing in the form of the polynomial system1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 99, + 554, + 287, + 583 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 554, + 287, + 583 + ], + "spans": [ + { + "bbox": [ + 99, + 554, + 287, + 583 + ], + "type": "interline_equation", + "content": "\\left\\{ \\begin{array}{l l} \\mathbf {s} ^ {\\top} \\mathbf {M} \\frac {\\partial \\mathbf {s}}{\\partial q _ {i}} & = 0, \\quad i = 0, \\dots , 3 \\\\ \\mathbf {q} ^ {\\top} \\mathbf {q} - 1 & = 0. \\end{array} \\right. \\tag {6}", + "image_path": "e58f9f3739dadd9192fb2da841be853c3af2f2a2489f089293a4685f473a78df.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 583, + 216, + 594 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 583, + 216, + 594 + ], + "spans": [ + { + "bbox": [ + 47, + 583, + 216, + 594 + ], + "type": "text", + "content": "6.2. Regressor and online simulator" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 601, + 287, + 637 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 601, + 287, + 637 + ], + "spans": [ + { + "bbox": [ + 46, + 601, + 287, + 637 + ], + "type": "text", + "content": "To provide a proper problem-solution pair for HC, we train a pose regression network from synthetic data. Its input is given by row-wise stacking of the correspondences, i.e." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 129, + 641, + 287, + 689 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 641, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 129, + 641, + 287, + 689 + ], + "type": "interline_equation", + "content": "\\left[ \\begin{array}{c c c} \\mathbf {p} _ {1} ^ {\\top} & \\mathbf {f} _ {1} ^ {\\top} & \\mathbf {v} _ {1} ^ {\\top} \\\\ \\vdots & \\vdots & \\vdots \\\\ \\mathbf {p} _ {N} ^ {\\top} & \\mathbf {f} _ {N} ^ {\\top} & \\mathbf {v} _ {N} ^ {\\top} \\end{array} \\right], \\tag {7}", + "image_path": "10c33d07a3505311c9826e36b7688f6cd8279cedd26cf0d006048e38bf52ba67.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 307, + 83, + 529, + 95 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 83, + 529, + 95 + ], + "spans": [ + { + "bbox": [ + 307, + 83, + 529, + 95 + ], + "type": "text", + "content": "Algorithm 1 Online Simulator HC for generalized PnP" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 310, + 100, + 545, + 185 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 310, + 100, + 489, + 111 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 100, + 489, + 111 + ], + "spans": [ + { + "bbox": [ + 310, + 100, + 489, + 111 + ], + "type": "text", + "content": "1: Input: correspondences " + }, + { + "bbox": [ + 310, + 100, + 489, + 111 + ], + "type": "inline_equation", + "content": "(\\mathbf{p}_i,\\mathbf{f}_i,\\mathbf{v}_i)_{i = 1}^M" + }, + { + "bbox": [ + 310, + 100, + 489, + 111 + ], + "type": "text", + "content": " , stepsize " + }, + { + "bbox": [ + 310, + 100, + 489, + 111 + ], + "type": "inline_equation", + "content": "\\delta t" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 310, + 112, + 386, + 121 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 112, + 386, + 121 + ], + "spans": [ + { + "bbox": [ + 310, + 112, + 386, + 121 + ], + "type": "text", + "content": "2: Output: poses " + }, + { + "bbox": [ + 310, + 112, + 386, + 121 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{R}},\\hat{\\mathbf{t}}" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 310, + 121, + 497, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 121, + 497, + 131 + ], + "spans": [ + { + "bbox": [ + 310, + 121, + 497, + 131 + ], + "type": "text", + "content": "3: " + }, + { + "bbox": [ + 310, + 121, + 497, + 131 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{R}},\\tilde{\\mathbf{t}}\\gets" + }, + { + "bbox": [ + 310, + 121, + 497, + 131 + ], + "type": "text", + "content": " Trained Regression Network((pi,fi,v)M)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 310, + 131, + 545, + 143 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 131, + 545, + 143 + ], + "spans": [ + { + "bbox": [ + 310, + 131, + 545, + 143 + ], + "type": "text", + "content": "4: " + }, + { + "bbox": [ + 310, + 131, + 545, + 143 + ], + "type": "inline_equation", + "content": "(\\hat{\\mathbf{f}}_i)_{i=1}^M \\leftarrow" + }, + { + "bbox": [ + 310, + 131, + 545, + 143 + ], + "type": "text", + "content": " Reprojection " + }, + { + "bbox": [ + 310, + 131, + 545, + 143 + ], + "type": "inline_equation", + "content": "(\\hat{\\mathbf{R}}, \\hat{\\mathbf{t}}; (\\mathbf{p}_i, \\mathbf{v}_i)_{i=1}^M)" + }, + { + "bbox": [ + 310, + 131, + 545, + 143 + ], + "type": "inline_equation", + "content": "\\triangleright" + }, + { + "bbox": [ + 310, + 131, + 545, + 143 + ], + "type": "text", + "content": " Online Simulator" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 310, + 143, + 521, + 154 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 143, + 521, + 154 + ], + "spans": [ + { + "bbox": [ + 310, + 143, + 521, + 154 + ], + "type": "text", + "content": "5: Starting problem " + }, + { + "bbox": [ + 310, + 143, + 521, + 154 + ], + "type": "inline_equation", + "content": "\\mathbf{G}(\\mathbf{x})\\gets (\\mathbf{p}_i,\\hat{\\mathbf{f}}_i,\\mathbf{v}_i)_{i = 1}^M" + }, + { + "bbox": [ + 310, + 143, + 521, + 154 + ], + "type": "text", + "content": " with solution " + }, + { + "bbox": [ + 310, + 143, + 521, + 154 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{R}},\\hat{\\mathbf{t}}" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 310, + 154, + 452, + 164 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 154, + 452, + 164 + ], + "spans": [ + { + "bbox": [ + 310, + 154, + 452, + 164 + ], + "type": "text", + "content": "6: Target problem " + }, + { + "bbox": [ + 310, + 154, + 452, + 164 + ], + "type": "inline_equation", + "content": "\\mathbf{F}(\\mathbf{x})\\gets (\\mathbf{p}_i,\\mathbf{f}_i,\\mathbf{v}_i)_{i = 1}^M" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 310, + 164, + 509, + 174 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 164, + 509, + 174 + ], + "spans": [ + { + "bbox": [ + 310, + 164, + 509, + 174 + ], + "type": "text", + "content": "7: " + }, + { + "bbox": [ + 310, + 164, + 509, + 174 + ], + "type": "inline_equation", + "content": "\\mathbf{q}_{HC} \\gets" + }, + { + "bbox": [ + 310, + 164, + 509, + 174 + ], + "type": "text", + "content": " Homotopy Continuation " + }, + { + "bbox": [ + 310, + 164, + 509, + 174 + ], + "type": "inline_equation", + "content": "t\\mathbf{F}(\\mathbf{x}) + (1 - t)\\mathbf{G}(\\mathbf{x})" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 310, + 175, + 405, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 175, + 405, + 185 + ], + "spans": [ + { + "bbox": [ + 310, + 175, + 405, + 185 + ], + "type": "text", + "content": "8: " + }, + { + "bbox": [ + 310, + 175, + 405, + 185 + ], + "type": "inline_equation", + "content": "\\bar{\\mathbf{R}},\\hat{\\mathbf{t}}\\gets" + }, + { + "bbox": [ + 310, + 175, + 405, + 185 + ], + "type": "text", + "content": " decompose " + }, + { + "bbox": [ + 310, + 175, + 405, + 185 + ], + "type": "inline_equation", + "content": "\\mathbf{q}_{HC}" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 304, + 202, + 546, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 202, + 546, + 361 + ], + "spans": [ + { + "bbox": [ + 304, + 202, + 546, + 361 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 202, + 546, + 361 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 202, + 546, + 361 + ], + "type": "text", + "content": " is the number of correspondences used, and the output is the quaternion vector. The network consists of 4 1D convolutional layers and a fully connected layer at the end to convert the output to a quaternion vector with the unit norm. Given the rotation predicted by the network, the translation can be recovered from correspondences and estimated rotation. We refer the reader to [31] for the details. For the online simulator, we can get reprojected " + }, + { + "bbox": [ + 304, + 202, + 546, + 361 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{f}}_i" + }, + { + "bbox": [ + 304, + 202, + 546, + 361 + ], + "type": "text", + "content": " from (5). Algorithm 1 summarizes all stages, including the final HC over (6). Note that in line 5 of Algorithm 1, the starting problem " + }, + { + "bbox": [ + 304, + 202, + 546, + 361 + ], + "type": "inline_equation", + "content": "\\mathbf{G}(\\mathbf{x})" + }, + { + "bbox": [ + 304, + 202, + 546, + 361 + ], + "type": "text", + "content": " is constructed by replacing the coefficient matrix " + }, + { + "bbox": [ + 304, + 202, + 546, + 361 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 304, + 202, + 546, + 361 + ], + "type": "text", + "content": " computed by " + }, + { + "bbox": [ + 304, + 202, + 546, + 361 + ], + "type": "inline_equation", + "content": "(\\mathbf{p}_i,\\mathbf{f}_i,\\mathbf{v}_i)" + }, + { + "bbox": [ + 304, + 202, + 546, + 361 + ], + "type": "text", + "content": " in (6) with " + }, + { + "bbox": [ + 304, + 202, + 546, + 361 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{M}}" + }, + { + "bbox": [ + 304, + 202, + 546, + 361 + ], + "type": "text", + "content": " computed by " + }, + { + "bbox": [ + 304, + 202, + 546, + 361 + ], + "type": "inline_equation", + "content": "(\\mathbf{p}_i,\\hat{\\mathbf{f}}_i,\\mathbf{v}_i)" + }, + { + "bbox": [ + 304, + 202, + 546, + 361 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 306, + 368, + 388, + 380 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 368, + 388, + 380 + ], + "spans": [ + { + "bbox": [ + 306, + 368, + 388, + 380 + ], + "type": "text", + "content": "6.3. Experiments" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 304, + 386, + 545, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 386, + 545, + 601 + ], + "spans": [ + { + "bbox": [ + 304, + 386, + 545, + 601 + ], + "type": "text", + "content": "Data. We randomly generate 4 cameras in a cube " + }, + { + "bbox": [ + 304, + 386, + 545, + 601 + ], + "type": "inline_equation", + "content": "[-1,1]^3" + }, + { + "bbox": [ + 304, + 386, + 545, + 601 + ], + "type": "text", + "content": " and randomly sample 16 3D points in the box " + }, + { + "bbox": [ + 304, + 386, + 545, + 601 + ], + "type": "inline_equation", + "content": "[-1,1]^2 \\times [4,8]" + }, + { + "bbox": [ + 304, + 386, + 545, + 601 + ], + "type": "text", + "content": ". 16 samples would not be a good choice for a RANSAC solver [18], but in the present context the idea is to set a moderate and notably equal challenge to all compared solvers. Furthermore, for UPnP the number of constraints appearing in the polynomial system remains unchanged as it analyses first-order optimality rather than individual incidence relations. Elements of the rotation angles are randomly generated in the interval " + }, + { + "bbox": [ + 304, + 386, + 545, + 601 + ], + "type": "inline_equation", + "content": "[- \\pi /2, \\pi /2]" + }, + { + "bbox": [ + 304, + 386, + 545, + 601 + ], + "type": "text", + "content": ", respectively, and the translation is randomly generated in the unit cube. We then generate the correspondences from rays that point to a 3D point. The noise level is set to 2 pixels. To train the pose regression network, we simulate a dataset of 2400 samples, which is sufficient for reliable Simulator HC initialization. We use a 2.70GHz Intel(R) Xeon(R) CPU and an NVIDIA GeForce RTX 4090 for network training and evaluation." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 304, + 601, + 545, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 601, + 545, + 685 + ], + "spans": [ + { + "bbox": [ + 304, + 601, + 545, + 685 + ], + "type": "text", + "content": "Methods. For HC, the stepsize " + }, + { + "bbox": [ + 304, + 601, + 545, + 685 + ], + "type": "inline_equation", + "content": "\\delta t" + }, + { + "bbox": [ + 304, + 601, + 545, + 685 + ], + "type": "text", + "content": " and the maximum number of iterations of Newton's method are set to 0.02 and 5, respectively. We compare the Gröbner Bases solver [31] and the popular local method Levenberg-Marquardt (LM) [1]. We also test the state-of-the-art parallel HC solver developed by [7, 8], and use both its CPU and " + }, + { + "bbox": [ + 304, + 601, + 545, + 685 + ], + "type": "inline_equation", + "content": "\\mathrm{GPU}^2" + }, + { + "bbox": [ + 304, + 601, + 545, + 685 + ], + "type": "text", + "content": " version. Given that parallel HC tracks all solutions," + } + ] + } + ], + "index": 28 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 46, + 693, + 287, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 693, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 693, + 287, + 712 + ], + "type": "text", + "content": "1We kindly refer the reader to the UPnP paper for detailed derivations [31]." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 693, + 544, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 693, + 544, + 712 + ], + "spans": [ + { + "bbox": [ + 306, + 693, + 544, + 712 + ], + "type": "text", + "content": "2Note that GPU warmup overhead is ignored when averaging over multiple runs." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27107" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 47, + 124, + 291, + 213 + ], + "blocks": [ + { + "bbox": [ + 47, + 70, + 288, + 115 + ], + "lines": [ + { + "bbox": [ + 47, + 70, + 288, + 115 + ], + "spans": [ + { + "bbox": [ + 47, + 70, + 288, + 115 + ], + "type": "text", + "content": "Table 1. Performance comparison of simulator HC, a traditional method, and parallel exhaustive tracking HC on noisy synthetic data. All results are averaged over 1000 trials. The success rate is computed for rotation errors with a threshold of 2 degree." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 47, + 124, + 291, + 213 + ], + "lines": [ + { + "bbox": [ + 47, + 124, + 291, + 213 + ], + "spans": [ + { + "bbox": [ + 47, + 124, + 291, + 213 + ], + "type": "table", + "html": "
MethodSucc. (%)ER(deg)Et(%)Time(ms)
Gröbner Bases1000.0310.4891.254
CPU HC1000.0320.4999.796
GPU HC1000.0320.4991.002
Pred Initial327.204320.031
Pred Initial + LM5428.7222560.528
Rand Initial + Simulator HC2291.682280.231
Pred Initial + Simulator HC1000.0320.5130.316
", + "image_path": "60782c54be04b24b9ebfd90ca87518a28bd2b337f6cc66f93d7ad26d154fa750.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 222, + 247, + 233 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 222, + 247, + 233 + ], + "spans": [ + { + "bbox": [ + 47, + 222, + 247, + 233 + ], + "type": "text", + "content": "we manually pick the one with the smallest error." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 235, + 288, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 235, + 288, + 475 + ], + "spans": [ + { + "bbox": [ + 47, + 235, + 288, + 475 + ], + "type": "text", + "content": "Results. The rotation error is measured by " + }, + { + "bbox": [ + 47, + 235, + 288, + 475 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{\\mathbf{R}} = 2\\arcsin (\\| \\hat{\\mathbf{R}} -\\mathbf{R}_{gt}\\| /(2\\sqrt{2}))" + }, + { + "bbox": [ + 47, + 235, + 288, + 475 + ], + "type": "text", + "content": " in degrees as suggested by [11] and the translation error is measured by the " + }, + { + "bbox": [ + 47, + 235, + 288, + 475 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_{\\mathrm{t}} = \\| \\hat{\\mathbf{t}} -\\mathbf{t}_{gt}\\| /\\| \\mathbf{t}_{gt}\\| \\times 100" + }, + { + "bbox": [ + 47, + 235, + 288, + 475 + ], + "type": "text", + "content": " in percentage. Table 1 compares the traditional method, parallel exhaustive tracking HC, and different types of initial solutions for our simulator HC. We present the success rate as the percentage of rotation errors smaller than 2 degrees. It can be observed that on this simple task, except for the local method, traditional methods, as well as simulator HC using predicted initials, succeed in " + }, + { + "bbox": [ + 47, + 235, + 288, + 475 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 47, + 235, + 288, + 475 + ], + "type": "text", + "content": " of trials. Meanwhile, online simulator HC benefits from the single-track strategy and turns out to be the fastest one. We also perform an ablation study on the proposed simulator HC, using both predicted and random initializations. Although the pose learned by the regressor is not highly accurate and only slightly outperforms random poses, it serves as a reasonable starting solution for our simulator HC. In contrast, LM as a local method is unable to leverage the potential of the predicted initializations, as the simulator HC does." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 489, + 247, + 501 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 489, + 247, + 501 + ], + "spans": [ + { + "bbox": [ + 47, + 489, + 247, + 501 + ], + "type": "text", + "content": "7. Generalized Relative Pose and Scale" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 510, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 510, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 510, + 287, + 713 + ], + "type": "text", + "content": "Our final application targets a hard minimal problem: Generalized Relative Pose and Scale (GRPS) estimation from 2D-2D correspondences [32]. The GRPS problem is a natural extension of the generalized relative pose problem and arises in point-less registration of partial view-graphs in structure-from-motion. Owing to the fact that SfM results are generally up to scale, an inherent scale factor needs to be resolved. Such problem was first introduced by Kneip et al. [32]. It has 140 solutions if a unique rotation representation is used, and can be solved by multi-start optimization to overcome local minima. Later on, a traditional polynomial solver based on the Gröbner Bases method has been proposed [36, 40]. With a " + }, + { + "bbox": [ + 46, + 510, + 287, + 713 + ], + "type": "inline_equation", + "content": "144 \\times 284" + }, + { + "bbox": [ + 46, + 510, + 287, + 713 + ], + "type": "text", + "content": " elimination template, the solver finds all 140 rotations within reasonable time. In this section, we solve GRPS by our proposed method and compare it against the existing approach, as well as a parallel HC implementation tracking all roots." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 72, + 430, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 430, + 83 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 430, + 83 + ], + "type": "text", + "content": "7.1. Problem Formulation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 89, + 545, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 89, + 545, + 198 + ], + "spans": [ + { + "bbox": [ + 304, + 89, + 545, + 198 + ], + "type": "text", + "content": "Consider a 2D-2D generalized correspondence " + }, + { + "bbox": [ + 304, + 89, + 545, + 198 + ], + "type": "inline_equation", + "content": "(\\mathbf{f}_i,\\mathbf{v}_i) - (\\mathbf{f}_i',\\mathbf{v}_i')" + }, + { + "bbox": [ + 304, + 89, + 545, + 198 + ], + "type": "text", + "content": " observed from two non-central camera views as illustrated in Figure 3b. " + }, + { + "bbox": [ + 304, + 89, + 545, + 198 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_i\\in \\mathbb{S}^2,\\mathbf{v}_i\\in \\mathbb{R}^3" + }, + { + "bbox": [ + 304, + 89, + 545, + 198 + ], + "type": "text", + "content": " denote the image ray and ray origin in the first view, and " + }, + { + "bbox": [ + 304, + 89, + 545, + 198 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_i'\\in \\mathbb{S}^2,\\mathbf{v}_i'\\in \\mathbb{R}^3" + }, + { + "bbox": [ + 304, + 89, + 545, + 198 + ], + "type": "text", + "content": " are the corresponding entities in the second view. The correspondence follows an underlying generalized relative pose and scale incidence relation depending on the rotation " + }, + { + "bbox": [ + 304, + 89, + 545, + 198 + ], + "type": "inline_equation", + "content": "\\mathbf{R}\\in SO(3)" + }, + { + "bbox": [ + 304, + 89, + 545, + 198 + ], + "type": "text", + "content": ", the position " + }, + { + "bbox": [ + 304, + 89, + 545, + 198 + ], + "type": "inline_equation", + "content": "\\mathbf{t}\\in \\mathbb{R}^3" + }, + { + "bbox": [ + 304, + 89, + 545, + 198 + ], + "type": "text", + "content": ", the scale " + }, + { + "bbox": [ + 304, + 89, + 545, + 198 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 304, + 89, + 545, + 198 + ], + "type": "text", + "content": ", and unknown latent depths " + }, + { + "bbox": [ + 304, + 89, + 545, + 198 + ], + "type": "inline_equation", + "content": "\\alpha_{i}" + }, + { + "bbox": [ + 304, + 89, + 545, + 198 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 89, + 545, + 198 + ], + "type": "inline_equation", + "content": "\\alpha_{i}'" + }, + { + "bbox": [ + 304, + 89, + 545, + 198 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 357, + 206, + 545, + 220 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 357, + 206, + 545, + 220 + ], + "spans": [ + { + "bbox": [ + 357, + 206, + 545, + 220 + ], + "type": "interline_equation", + "content": "\\alpha_ {i} \\mathbf {f} _ {i} + \\mathbf {v} _ {i} = \\mathbf {R} \\left(\\alpha_ {i} ^ {\\prime} \\mathbf {f} _ {i} ^ {\\prime} + s \\mathbf {v} _ {i} ^ {\\prime}\\right) + \\mathbf {t}. \\tag {8}", + "image_path": "76625660ce053c041906f1df0821c24fb1871e8ecd9167c34211b88d26730783.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 228, + 545, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 228, + 545, + 276 + ], + "spans": [ + { + "bbox": [ + 304, + 228, + 545, + 276 + ], + "type": "text", + "content": "With the help of Plücker coordinates and the scalar triple product " + }, + { + "bbox": [ + 304, + 228, + 545, + 276 + ], + "type": "inline_equation", + "content": "^3" + }, + { + "bbox": [ + 304, + 228, + 545, + 276 + ], + "type": "text", + "content": ", the unknown depths get eliminated which results in a polynomial constraint that only involves the unknown poses" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 317, + 283, + 545, + 299 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 283, + 545, + 299 + ], + "spans": [ + { + "bbox": [ + 317, + 283, + 545, + 299 + ], + "type": "interline_equation", + "content": "\\mathbf {t} ^ {\\top} \\left(\\mathbf {f} _ {i} \\times \\mathbf {R} \\mathbf {f} _ {i} ^ {\\prime}\\right) - s \\mathbf {f} _ {i} ^ {\\top} \\mathbf {R} \\left[ \\mathbf {v} _ {i} ^ {\\prime} \\right] _ {\\times} \\mathbf {f} _ {i} ^ {\\prime} + \\mathbf {f} _ {i} ^ {\\top} \\left[ \\mathbf {v} _ {i} \\right] _ {\\times} \\mathbf {R} \\mathbf {f} _ {i} ^ {\\prime} = 0. \\tag {9}", + "image_path": "6689289f78376a3a37b5e46b19e325fba67f8833b4e3b45fd555fe83b56da2f0.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 307, + 545, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 307, + 545, + 354 + ], + "spans": [ + { + "bbox": [ + 304, + 307, + 545, + 354 + ], + "type": "text", + "content": "Therefore, a minimal formulation for the GRPS problem can be devised by a polynomial system formed by stacking 7 constraints in the form (9), each one depending on one correspondence." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 361, + 474, + 374 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 361, + 474, + 374 + ], + "spans": [ + { + "bbox": [ + 306, + 361, + 474, + 374 + ], + "type": "text", + "content": "7.2. Regressor and online simulator" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 380, + 545, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 380, + 545, + 536 + ], + "spans": [ + { + "bbox": [ + 304, + 380, + 545, + 536 + ], + "type": "text", + "content": "As introduced in section 6.2, we continue to use a pose regression network that inputs a matrix with row-wise stacked correspondences. The output is given by " + }, + { + "bbox": [ + 304, + 380, + 545, + 536 + ], + "type": "inline_equation", + "content": "\\mathbf{R}" + }, + { + "bbox": [ + 304, + 380, + 545, + 536 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 380, + 545, + 536 + ], + "type": "inline_equation", + "content": "\\mathbf{t}" + }, + { + "bbox": [ + 304, + 380, + 545, + 536 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 380, + 545, + 536 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 304, + 380, + 545, + 536 + ], + "type": "text", + "content": ". Note that the rotation is predicted in continuous 6D representation which is more suitable for neural networks [66]. The network consists of 11 1D convolutional layers, in the end, it branches to 3 sub-branches for " + }, + { + "bbox": [ + 304, + 380, + 545, + 536 + ], + "type": "inline_equation", + "content": "\\mathbf{R}" + }, + { + "bbox": [ + 304, + 380, + 545, + 536 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 380, + 545, + 536 + ], + "type": "inline_equation", + "content": "\\mathbf{t}" + }, + { + "bbox": [ + 304, + 380, + 545, + 536 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 380, + 545, + 536 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 304, + 380, + 545, + 536 + ], + "type": "text", + "content": ". Each sub-branch consists of 2 1D convolutional layers and 1 fully connected layers for adjust output. Given the poses predicted by the network, we still need an estimation of the depths before simulating the reprojected image ray. With estimated " + }, + { + "bbox": [ + 304, + 380, + 545, + 536 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{R}}" + }, + { + "bbox": [ + 304, + 380, + 545, + 536 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 380, + 545, + 536 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{t}}" + }, + { + "bbox": [ + 304, + 380, + 545, + 536 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 380, + 545, + 536 + ], + "type": "inline_equation", + "content": "\\hat{s}" + }, + { + "bbox": [ + 304, + 380, + 545, + 536 + ], + "type": "text", + "content": ", the corresponding depths can be obtained from (8) by following" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 343, + 544, + 545, + 571 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 343, + 544, + 545, + 571 + ], + "spans": [ + { + "bbox": [ + 343, + 544, + 545, + 571 + ], + "type": "interline_equation", + "content": "\\left[ \\mathbf {f} _ {i} - \\hat {\\mathbf {R}} \\mathbf {f} _ {i} ^ {\\prime} \\right] \\left[ \\begin{array}{l} \\alpha_ {i} \\\\ \\alpha_ {i} ^ {\\prime} \\end{array} \\right] = - \\left(\\mathbf {v} _ {i} - \\hat {s} \\hat {\\mathbf {R}} \\mathbf {v} _ {i} ^ {\\prime} - \\hat {\\mathbf {t}}\\right), \\tag {10}", + "image_path": "03ed5f080f38e58f293ae6a8c9fa87da7cb7706bdfe9d0ce6c1592b8128214f9.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 578, + 545, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 578, + 545, + 613 + ], + "spans": [ + { + "bbox": [ + 304, + 578, + 545, + 613 + ], + "type": "text", + "content": "which can be simply solved by linear least squares for each input correspondence. Finally, we simulate a pose-consistent reprojected image ray in the first view using" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 354, + 621, + 545, + 645 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 621, + 545, + 645 + ], + "spans": [ + { + "bbox": [ + 354, + 621, + 545, + 645 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {f}} _ {i} = \\frac {1}{\\hat {\\alpha} _ {i}} \\hat {\\mathbf {R}} \\left(\\hat {\\alpha} _ {i} ^ {\\prime} \\mathbf {f} _ {i} ^ {\\prime} + \\hat {s} \\mathbf {v} _ {i} ^ {\\prime}\\right) + \\hat {\\mathbf {t}} - \\mathbf {v} _ {i}. \\tag {11}", + "image_path": "a28b97a4d04a86d10aaf434ba985931692097c43d735cc760fb845b34fbcd1f3.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 653, + 443, + 666 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 653, + 443, + 666 + ], + "spans": [ + { + "bbox": [ + 306, + 653, + 443, + 666 + ], + "type": "text", + "content": "7.3. Simulation Experiments" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 671, + 545, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 671, + 545, + 696 + ], + "spans": [ + { + "bbox": [ + 304, + 671, + 545, + 696 + ], + "type": "text", + "content": "Data. We adopt a similar data generation method as in section 6.3. We generate 7 3D points in a " + }, + { + "bbox": [ + 304, + 671, + 545, + 696 + ], + "type": "inline_equation", + "content": "[-1,1]^2 \\times [2,20]" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 315, + 702, + 518, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 702, + 518, + 712 + ], + "spans": [ + { + "bbox": [ + 315, + 702, + 518, + 712 + ], + "type": "text", + "content": "3We kindly refer the reader to [28, 32] for detailed derivations" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27108" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "text", + "content": "volume. Two view frames are generated with origins picked from " + }, + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "inline_equation", + "content": "[-1, 1]^3" + }, + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "text", + "content": ". Within each view frame, we then sample individual camera offsets from the volume " + }, + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "inline_equation", + "content": "[-1, 1]^3" + }, + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "text", + "content": ". The scale is randomly generated in " + }, + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "inline_equation", + "content": "[0.1, 5.0]" + }, + { + "bbox": [ + 46, + 72, + 287, + 156 + ], + "type": "text", + "content": ", and all other parameters related to noise addition are left unchanged. Note that the ground truth relative pose needs to be extracted from the simulated pair of view poses." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 156, + 288, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 156, + 288, + 300 + ], + "spans": [ + { + "bbox": [ + 46, + 156, + 288, + 300 + ], + "type": "text", + "content": "Methods. The stepsize for homotopy continuation is set to 0.05 in this scenario. We compare against the Eigen solver [32], Gröbner Bases solver [40] and the CPU and GPU versions of parallel HC which tracks 140 solutions found by Monodromy [6]. Note that to measure numerical errors, we just pick the best solution with the smallest error. In practice, traditional HC needs to evaluate additional information to select a single solution (e.g. euclidean distance). For a complicated problem like GRPS, evaluating 140 solutions would still induce non-neglectable time-consumption. Since GRPS is a hard problem, we train the regressor using a larger simulated dataset of size 64000." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 300, + 287, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 300, + 287, + 335 + ], + "spans": [ + { + "bbox": [ + 46, + 300, + 287, + 335 + ], + "type": "text", + "content": "Metric. We measure the rotation error in degrees and the relative translation as in Sec. 6.3 and scale error in percentage using" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 112, + 335, + 287, + 349 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 112, + 335, + 287, + 349 + ], + "spans": [ + { + "bbox": [ + 112, + 335, + 287, + 349 + ], + "type": "interline_equation", + "content": "\\mathcal {E} _ {s} = \\left| \\hat {s} - s _ {g t} \\right| / s _ {g t} \\times 1 0 0. \\tag {12}", + "image_path": "5ada4082ddadd4410e8bf3baba5b37e2f02327f5be93b7f2d557a0da8e88518f.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 354, + 287, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 354, + 287, + 378 + ], + "spans": [ + { + "bbox": [ + 46, + 354, + 287, + 378 + ], + "type": "text", + "content": "To calculate the success rate, for rotation errors the threshold is set to 2 deg, and for relative errors, it is set to " + }, + { + "bbox": [ + 46, + 354, + 287, + 378 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 46, + 354, + 287, + 378 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 379, + 287, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 379, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 46, + 379, + 287, + 533 + ], + "type": "text", + "content": "Numerical Stability. We randomly generate 1000 trials to test the numerical stability of all solvers. During the analysis of the performance of simulator HC using the minimum number of correspondences, we observed that our method has around " + }, + { + "bbox": [ + 46, + 379, + 287, + 533 + ], + "type": "inline_equation", + "content": "70\\%" + }, + { + "bbox": [ + 46, + 379, + 287, + 533 + ], + "type": "text", + "content": " success rate. By checking the failure cases, we found that sometimes simulator HC tracked another real solution that was different from the ground truth pose. To tackle that issue, we evaluate the addition of one more correspondence and solve for an overdetermined polynomial system. By using 8 correspondences, simulator HC reaches a success rate of " + }, + { + "bbox": [ + 46, + 379, + 287, + 533 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 46, + 379, + 287, + 533 + ], + "type": "text", + "content": ". Again, since CPU/GPU-HC is designed to handle square systems, only the minimum number of correspondences can be tested for other HC methods." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 534, + 287, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 534, + 287, + 617 + ], + "spans": [ + { + "bbox": [ + 46, + 534, + 287, + 617 + ], + "type": "text", + "content": "Since the errors in rotation, translation and scale have similar density results, we only present the rotation error in the main paper and the rest in the supplementary. As illustrated in Figure 4a, the minimal HC solvers all present similar error distribution and around " + }, + { + "bbox": [ + 46, + 534, + 287, + 617 + ], + "type": "inline_equation", + "content": "75\\%" + }, + { + "bbox": [ + 46, + 534, + 287, + 617 + ], + "type": "text", + "content": " of trials are successful. While Gröbner Bases shows " + }, + { + "bbox": [ + 46, + 534, + 287, + 617 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 46, + 534, + 287, + 617 + ], + "type": "text", + "content": " success rate, Simulator HC over 8pts presents similar numerical errors." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 288, + 714 + ], + "type": "text", + "content": "Time Comparison. The main computational benefit of the proposed online simulator HC comes from single solution-tracking, which can be effectuated efficiently even on the CPU. Figure 4b demonstrates the efficiency gain of our 8pts simulator HC over the eigen solver and parallel HC variants. As expected, by tracking only one solution curve, we are significantly faster than parallel exhaustive tracking HC, even if the latter runs on GPU. The eigen solver" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 133 + ], + "type": "text", + "content": "is an optimization-based solver with high complexity and ends up being slowest. The Gröbner Bases solver dealing with the large elimination template still costs a considerable time, even for just computing solutions for rotation, and not translation or scale." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 137, + 547, + 269 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 137, + 547, + 269 + ], + "spans": [ + { + "bbox": [ + 304, + 137, + 547, + 269 + ], + "type": "text", + "content": "Remark. We would like to note that Gröbner Bases is solving a different formulation from Eq. (9). In order to make it easy to be solved by a smaller elimination template, it eliminates translation and scale [57]. But the resulting coefficient matrix is " + }, + { + "bbox": [ + 304, + 137, + 547, + 269 + ], + "type": "inline_equation", + "content": "130 \\times 275" + }, + { + "bbox": [ + 304, + 137, + 547, + 269 + ], + "type": "text", + "content": " and still needs a large number of operations. Although the formulation of Eq. (9) with eliminated depths faces a slightly lower success rate, its number of operations is much smaller than Gröbner Bases. On the other hand, the proposed 8pt simulator HC provides a comparably high success rate as Gröbner Bases and is still " + }, + { + "bbox": [ + 304, + 137, + 547, + 269 + ], + "type": "inline_equation", + "content": "17 \\times" + }, + { + "bbox": [ + 304, + 137, + 547, + 269 + ], + "type": "text", + "content": " faster." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 282, + 545, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 282, + 545, + 350 + ], + "spans": [ + { + "bbox": [ + 304, + 282, + 545, + 350 + ], + "type": "text", + "content": "Table 2. RANSAC experiments for the GRPS problem using Gröbner Bases | simulator HC. Image points are randomly perturbed by 2-pixel uniform noise. We use vanilla RANSAC with at most 200 iterations. As can be observed, outlier ratios of up to " + }, + { + "bbox": [ + 304, + 282, + 545, + 350 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 304, + 282, + 545, + 350 + ], + "type": "text", + "content": " are easily handled within 200 iterations. And regressor-based simulator HC is consistently faster than Gröbner Bases." + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 323, + 358, + 527, + 412 + ], + "blocks": [ + { + "bbox": [ + 323, + 358, + 527, + 412 + ], + "lines": [ + { + "bbox": [ + 323, + 358, + 527, + 412 + ], + "spans": [ + { + "bbox": [ + 323, + 358, + 527, + 412 + ], + "type": "table", + "html": "
RatioSucc. (%)ER(deg)# ItersTime(s)
10%99.51000.100.09931252.330.85
20%99.41000.110.101711934.371.31
30%99.41000.130.142002005.101.35
40%98.9910.210.292002004.591.35
", + "image_path": "3e6c19a93dc0dd4bb330eaea0303c84eabae815caffb52385d802105e89b0a0e.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 427, + 545, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 427, + 545, + 522 + ], + "spans": [ + { + "bbox": [ + 304, + 427, + 545, + 522 + ], + "type": "text", + "content": "Noise Resilience. Figure 4c presents the performance of the proposed 8pts simulator HC with different numbers of cameras under increasing image noise levels ranging from 0.2 pixel to 1.0 pixel. The error behaves gracefully as a function of noise, and it can be observed that a higher number of cameras leads to slightly inferior noise. Furthermore, we also observe that rotation errors appear to be less sensitive to image noise than translation and scale errors." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 524, + 545, + 691 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 524, + 545, + 691 + ], + "spans": [ + { + "bbox": [ + 304, + 524, + 545, + 691 + ], + "type": "text", + "content": "Evaluation inside RANSAC. To conclude, we embed simulator HC into RANSAC to test its practical usefulness. We synthesize 200 correspondences between 5 cameras and replace part of them with outliers. Uniformly sampled pixel noise from " + }, + { + "bbox": [ + 304, + 524, + 545, + 691 + ], + "type": "inline_equation", + "content": "[-2,2]" + }, + { + "bbox": [ + 304, + 524, + 545, + 691 + ], + "type": "text", + "content": " is added to both coordinates of the image points by again assuming a virtual focal length of 800 px. We terminate RANSAC if it reaches " + }, + { + "bbox": [ + 304, + 524, + 545, + 691 + ], + "type": "inline_equation", + "content": "99\\%" + }, + { + "bbox": [ + 304, + 524, + 545, + 691 + ], + "type": "text", + "content": " confidence or 200 iterations, and the inlier threshold is set to 0.01 for the distance measured by the normalized reprojection error. Given that Gröbner Bases solver estimates 140 solutions, we keep the real ones and select the best solution using reprojection errors in parallel to save time for it. Table 2 shows the results. The success rate is measured by a rotation error threshold of 2 degrees and the indicated errors are" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 317, + 702, + 402, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 702, + 402, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 702, + 402, + 713 + ], + "type": "inline_equation", + "content": "{}^{4}" + }, + { + "bbox": [ + 317, + 702, + 402, + 713 + ], + "type": "text", + "content": " shown in supplementary." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27109" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 73, + 208, + 180 + ], + "blocks": [ + { + "bbox": [ + 50, + 73, + 208, + 180 + ], + "lines": [ + { + "bbox": [ + 50, + 73, + 208, + 180 + ], + "spans": [ + { + "bbox": [ + 50, + 73, + 208, + 180 + ], + "type": "image", + "image_path": "45ed5eb77a4209965faf887b0475a13238408c8420f2cc0948dd2bceb9812326.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 61, + 186, + 196, + 195 + ], + "lines": [ + { + "bbox": [ + 61, + 186, + 196, + 195 + ], + "spans": [ + { + "bbox": [ + 61, + 186, + 196, + 195 + ], + "type": "text", + "content": "(a) Error distribution over noise-free data." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 213, + 73, + 380, + 181 + ], + "blocks": [ + { + "bbox": [ + 213, + 73, + 380, + 181 + ], + "lines": [ + { + "bbox": [ + 213, + 73, + 380, + 181 + ], + "spans": [ + { + "bbox": [ + 213, + 73, + 380, + 181 + ], + "type": "image", + "image_path": "a77d11a9c7f3b97eb41fb553e927e3e14c55e7b36f9afddcdc3d7665d63ae342.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 224, + 186, + 369, + 196 + ], + "lines": [ + { + "bbox": [ + 224, + 186, + 369, + 196 + ], + "spans": [ + { + "bbox": [ + 224, + 186, + 369, + 196 + ], + "type": "text", + "content": "(b) The boxplot of running time comparison." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 386, + 72, + 545, + 182 + ], + "blocks": [ + { + "bbox": [ + 386, + 72, + 545, + 182 + ], + "lines": [ + { + "bbox": [ + 386, + 72, + 545, + 182 + ], + "spans": [ + { + "bbox": [ + 386, + 72, + 545, + 182 + ], + "type": "image", + "image_path": "6c5e8c963e11700af2d2282864c9fd2eb43e1d380071f90354aed29cbf794556.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 408, + 186, + 522, + 195 + ], + "lines": [ + { + "bbox": [ + 408, + 186, + 522, + 195 + ], + "spans": [ + { + "bbox": [ + 408, + 186, + 522, + 195 + ], + "type": "text", + "content": "(c) Error statistics of simulator HC." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 46, + 205, + 547, + 282 + ], + "lines": [ + { + "bbox": [ + 46, + 205, + 547, + 282 + ], + "spans": [ + { + "bbox": [ + 46, + 205, + 547, + 282 + ], + "type": "text", + "content": "Figure 4. Experimental results on GRPS. Fig. 4a Error distribution over 1000 trials on noise-free data. The camera number is set to be 3. Considering the rotation error, except for Gröbner Bases, the other minimal solvers all present a similar error distribution with around " + }, + { + "bbox": [ + 46, + 205, + 547, + 282 + ], + "type": "inline_equation", + "content": "70\\%" + }, + { + "bbox": [ + 46, + 205, + 547, + 282 + ], + "type": "text", + "content": " success rate, and the eigen solver presents only around " + }, + { + "bbox": [ + 46, + 205, + 547, + 282 + ], + "type": "inline_equation", + "content": "60\\%" + }, + { + "bbox": [ + 46, + 205, + 547, + 282 + ], + "type": "text", + "content": " success rate. The proposed 8pts simulator HC presents " + }, + { + "bbox": [ + 46, + 205, + 547, + 282 + ], + "type": "inline_equation", + "content": "96.3\\%" + }, + { + "bbox": [ + 46, + 205, + 547, + 282 + ], + "type": "text", + "content": " success rate where the Gröbner Bases has " + }, + { + "bbox": [ + 46, + 205, + 547, + 282 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 46, + 205, + 547, + 282 + ], + "type": "text", + "content": " success rate. Fig. 4b The boxplot of running time comparison for the proposed simulator HC and other methods, where the number represents the median time. As expected, on average the proposed online simulator HC running on CPU is the fastest. It is about " + }, + { + "bbox": [ + 46, + 205, + 547, + 282 + ], + "type": "inline_equation", + "content": "5\\times" + }, + { + "bbox": [ + 46, + 205, + 547, + 282 + ], + "type": "text", + "content": " faster than GPU-HC since we only track one solution, " + }, + { + "bbox": [ + 46, + 205, + 547, + 282 + ], + "type": "inline_equation", + "content": "17\\times" + }, + { + "bbox": [ + 46, + 205, + 547, + 282 + ], + "type": "text", + "content": " faster than Gröbner Bases and about " + }, + { + "bbox": [ + 46, + 205, + 547, + 282 + ], + "type": "inline_equation", + "content": "40\\times" + }, + { + "bbox": [ + 46, + 205, + 547, + 282 + ], + "type": "text", + "content": " faster than CPU-HC. Fig. 4c Error statistics of simulator HC with respect to different noise levels ranging from 0.2 pixel to 1.0 pixel." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 292, + 287, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 292, + 287, + 388 + ], + "spans": [ + { + "bbox": [ + 46, + 292, + 287, + 388 + ], + "type": "text", + "content": "the average accuracy of successful cases, results on translation and scale are left in supplementary. Overall, simulator HC embedded into RANSAC can easily handle up to " + }, + { + "bbox": [ + 46, + 292, + 287, + 388 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 46, + 292, + 287, + 388 + ], + "type": "text", + "content": " outliers as Gröbner Bases and much faster, which suggests high practicality. Note that Gröbner Bases provides around 46 real solutions each time, which requires a significant amount of additional time for solution selection. Meanwhile, simulator HC does not require this extra step." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 398, + 186, + 411 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 398, + 186, + 411 + ], + "spans": [ + { + "bbox": [ + 47, + 398, + 186, + 411 + ], + "type": "text", + "content": "7.4. Real World Experiments" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 418, + 287, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 418, + 287, + 515 + ], + "spans": [ + { + "bbox": [ + 46, + 418, + 287, + 515 + ], + "type": "text", + "content": "We further explore the performance of simulator HC on the real dataset LaMAR [50]. As illustrated in Fig. 5, the provided HoloLens data is a natural application of GRPS with identity scale, which can be viewed as a multi-camera rig with 4 cameras in the front and sides. We extract the correspondences by consecutively taking a view-pair from the trajectories and ensuring the distance between each view is in " + }, + { + "bbox": [ + 46, + 418, + 287, + 515 + ], + "type": "inline_equation", + "content": "[0.2m, 0.5m]" + }, + { + "bbox": [ + 46, + 418, + 287, + 515 + ], + "type": "text", + "content": " with at least 100 matches. The" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 50, + 534, + 287, + 578 + ], + "blocks": [ + { + "bbox": [ + 50, + 534, + 287, + 578 + ], + "lines": [ + { + "bbox": [ + 50, + 534, + 287, + 578 + ], + "spans": [ + { + "bbox": [ + 50, + 534, + 287, + 578 + ], + "type": "image", + "image_path": "18ecec1634171e8a7d5ba99f673cc4d24c41f8c6ed96af9e97ed16933d20d1e3.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 63, + 586, + 271, + 597 + ], + "lines": [ + { + "bbox": [ + 63, + 586, + 271, + 597 + ], + "spans": [ + { + "bbox": [ + 63, + 586, + 271, + 597 + ], + "type": "text", + "content": "Figure 5. LaMAR dataset multi-camera rig visualization." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 48, + 631, + 297, + 710 + ], + "blocks": [ + { + "bbox": [ + 48, + 610, + 285, + 621 + ], + "lines": [ + { + "bbox": [ + 48, + 610, + 285, + 621 + ], + "spans": [ + { + "bbox": [ + 48, + 610, + 285, + 621 + ], + "type": "text", + "content": "Table 3. LaMAR dataset RANSAC-simulatorHC Median results." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 631, + 297, + 710 + ], + "lines": [ + { + "bbox": [ + 48, + 631, + 297, + 710 + ], + "spans": [ + { + "bbox": [ + 48, + 631, + 297, + 710 + ], + "type": "table", + "html": "
seq.ER(deg)Es(%)Et(deg)#itersTime(s)Succ. (%)
Gröbner BasesCAB1.1856.145.701372.8360
HGE0.6262.484.39591.2567
LIN0.6042.043.40310.6572
simulator HCCAB1.0543.209.163102.0868
HGE0.5641.527.191340.9084
LIN0.5632.194.32660.4486
", + "image_path": "ec7a3fd451ed54ad5f193d56b3538ff7bc13023c8351467efcde27eefb586c9f.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 292, + 545, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 292, + 545, + 376 + ], + "spans": [ + { + "bbox": [ + 304, + 292, + 545, + 376 + ], + "type": "text", + "content": "resulting number of pairs is 1794 (CAB), 2852 (HGE) and 739 (LIN) in total. The correspondences are matched with SIFT features using nearest neighbour matching. We also present the angular error in translation, as in practice, the scale of translation and overall scale could be unobservable, e.g. the degenerate scenario when correspondences remain on the frame captured by the same camera." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 389, + 545, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 389, + 545, + 461 + ], + "spans": [ + { + "bbox": [ + 304, + 389, + 545, + 461 + ], + "type": "text", + "content": "Table 3 shows the median results over all three sequences of LaMAR dataset. We adopted a similar setting for RANSAC as the above section with 500 maximum iterations. The proposed solver still presents consistent performances, where simulator HC has higher success rate with faster running time than Gröbner Bases." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 476, + 378, + 489 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 476, + 378, + 489 + ], + "spans": [ + { + "bbox": [ + 306, + 476, + 378, + 489 + ], + "type": "text", + "content": "8. Conclusion" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "type": "text", + "content": "The present paper introduces a new paradigm for correspondence-based geometric problem solving using an elegant three-stage combination of a regression network, an online correspondence simulator, and homotopy continuation. The regression network acts as a general solution approximator and can be trained in simulation only. While pure regression accuracy is often insufficient on its own, a key message conveyed through our results is that the accuracy is generally still good enough to simulate a consistent starting problem-solution pair that successfully enables single-solution continuation. As demonstrated on two generalized camera problems, the strategy enables high success rates on common computing hardware, and the resulting solvers prove to be a viable solution if utilized as part of a random sample and consensus scheme. In particular, we propose a successful solution to the generalized relative pose and scale problem, which is around " + }, + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "type": "inline_equation", + "content": "17 \\times" + }, + { + "bbox": [ + 304, + 498, + 545, + 713 + ], + "type": "text", + "content": " faster than traditional polynomial elimination techniques." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "27110" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 157, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 157, + 85 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 157, + 85 + ], + "type": "text", + "content": "9. Acknowledgments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 91, + 287, + 128 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 91, + 287, + 128 + ], + "spans": [ + { + "bbox": [ + 47, + 91, + 287, + 128 + ], + "type": "text", + "content": "The authors would like to acknowledge the funding support provided by projects 22DZ1201900, 22ZR1441300, and DFYJBJ-1 by the Natural Science Foundation of Shanghai." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 138, + 107, + 150 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 138, + 107, + 150 + ], + "spans": [ + { + "bbox": [ + 48, + 138, + 107, + 150 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 158, + 286, + 713 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 53, + 158, + 286, + 179 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 158, + 286, + 179 + ], + "spans": [ + { + "bbox": [ + 53, + 158, + 286, + 179 + ], + "type": "text", + "content": "[1] Sameer Agarwal, Keir Mierle, and Others. Ceres solver. http://ceres-solver.org, 2022.5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 181, + 286, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 181, + 286, + 203 + ], + "spans": [ + { + "bbox": [ + 53, + 181, + 286, + 203 + ], + "type": "text", + "content": "[2] C. Aholt and L. Oeding. The ideal of the trifocal variety. Math. Comput., 83(289):2553-2574, 2014. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 205, + 286, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 205, + 286, + 247 + ], + "spans": [ + { + "bbox": [ + 53, + 205, + 286, + 247 + ], + "type": "text", + "content": "[3] B. Amos and J. Z. Kolter. Optnet: Differentiable optimization as a layer in neural networks. In International Conference on Machine Learning, pages 136-145. PMLR, 2017. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 249, + 286, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 249, + 286, + 282 + ], + "spans": [ + { + "bbox": [ + 53, + 249, + 286, + 282 + ], + "type": "text", + "content": "[4] D. J. Bates, A. J. Sommese, J. D. Hauenstein, and C. W. Wampler. Numerically solving polynomial systems with Bertini. Software, environments, tools, SIAM, 25, 2013. 1, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 284, + 286, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 284, + 286, + 314 + ], + "spans": [ + { + "bbox": [ + 53, + 284, + 286, + 314 + ], + "type": "text", + "content": "[5] S. Bhayani, Z. Kukelova, and J. Heikkilä. A sparse resultant based method for efficient minimal solvers. In CVPR, 2020. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 317, + 286, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 317, + 286, + 350 + ], + "spans": [ + { + "bbox": [ + 53, + 317, + 286, + 350 + ], + "type": "text", + "content": "[6] P. Breiding and S. Timme. Homotopycontinuation.jl: A package for homotopy continuation in Julia. In International Conference on Mathematical Software - ICMS, 2018. 2, 4, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 351, + 286, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 351, + 286, + 384 + ], + "spans": [ + { + "bbox": [ + 53, + 351, + 286, + 384 + ], + "type": "text", + "content": "[7] C.-H. Chien, H. Fan, A. Abdelfattah, E. Tsigaridas, S. Tomov, and B. Kimia. GPU-based homotopy continuation for minimal problems in computer vision. In CVPR, 2022. 2, 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 386, + 286, + 439 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 386, + 286, + 439 + ], + "spans": [ + { + "bbox": [ + 53, + 386, + 286, + 439 + ], + "type": "text", + "content": "[8] C.-H. Chien, H. Fan, A. Abdelfattah, E. Tsigaridas, S. Tomov, and B. Kimia. Parallel path tracking for homotopy continuation usinggpu. In Proceedings of the International Symposium on Symbolic and Algebraic Computation, 2022. 2, 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 441, + 286, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 441, + 286, + 463 + ], + "spans": [ + { + "bbox": [ + 53, + 441, + 286, + 463 + ], + "type": "text", + "content": "[9] D. Cox, J. Little, and D. O'Shea. Using Algebraic Geometry. Springer, 1998. 1, 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 464, + 286, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 464, + 286, + 486 + ], + "spans": [ + { + "bbox": [ + 48, + 464, + 286, + 486 + ], + "type": "text", + "content": "[10] Y. Dai, H. Li, and L. Kneip. Rolling shutter camera relative pose: Generalized epipolar geometry. In CVPR, 2016. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 487, + 286, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 487, + 286, + 520 + ], + "spans": [ + { + "bbox": [ + 48, + 487, + 286, + 520 + ], + "type": "text", + "content": "[11] Y. Ding, C.-H. Chien, V. Larsson, K. Åström, and B. Kimia. Minimal solutions to generalized three-view relative pose problem. In ICCV, 2023. 2, 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 521, + 286, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 521, + 286, + 564 + ], + "spans": [ + { + "bbox": [ + 48, + 521, + 286, + 564 + ], + "type": "text", + "content": "[12] T. Duff, C. Hill, A. Jensen, K. Lee, A. Leykin, and J. Sommars. Solving polynomial systems via homotopy continuation and monodromy. IMA Journal of Numerical Analysis, 2018. 1, 2, 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 567, + 286, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 567, + 286, + 599 + ], + "spans": [ + { + "bbox": [ + 48, + 567, + 286, + 599 + ], + "type": "text", + "content": "[13] T. Duff, K. Kohn, A. Leykin, and T. Pajdla. PLMP - point-line minimal problems in complete multi-view visibility. In ICCV, pages 1675-1684, 2019. 1, 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 601, + 286, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 601, + 286, + 632 + ], + "spans": [ + { + "bbox": [ + 48, + 601, + 286, + 632 + ], + "type": "text", + "content": "[14] T. Duff, K. Kohn, A. Leykin, and T. Pajdla. PL1P - point-line minimal problems under partial visibility in three views. In ECCV, 2020. 1, 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 48, + 635, + 286, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 286, + 655 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 286, + 655 + ], + "type": "text", + "content": "[15] A. Elquresh and A. M. Elgammal. Line-based relative pose estimation. In CVPR, 2011. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 48, + 658, + 286, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 286, + 678 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 286, + 678 + ], + "type": "text", + "content": "[16] I. Z. Emiris. A general solver based on sparse resultants, 2012. 1, 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 48, + 680, + 286, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 680, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 680, + 286, + 713 + ], + "type": "text", + "content": "[17] R. Fabbri, T. Duff, H. Fan, M. H. Regan, D. da Costa de Pinho, E. P. Tsigaridas, C. W. Wampler, J. D. Hauenstein, P. J. Giblin, A. Kimia, B. B. Leykin, and T. Pajdla. TR-PLP" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 40, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 95 + ], + "type": "text", + "content": "- Trifocal relative pose from lines at points. In CVPR, 2020. 1, 2, 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 97, + 545, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 97, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 307, + 97, + 545, + 140 + ], + "type": "text", + "content": "[18] Martin A Fischler and Robert C Bolles. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM, 24(6):381-395, 1981. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 143, + 545, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 143, + 545, + 186 + ], + "spans": [ + { + "bbox": [ + 307, + 143, + 545, + 186 + ], + "type": "text", + "content": "[19] M. A. Fischler and R. C. Bolles. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Commun. ACM, 24(6): 381-395, 1985. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 189, + 545, + 221 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 189, + 545, + 221 + ], + "spans": [ + { + "bbox": [ + 307, + 189, + 545, + 221 + ], + "type": "text", + "content": "[20] R. M. Haralick, C.-N. Lee, K. Ottenberg, and M. Nölle. Review and analysis of solutions of the three point perspective pose estimation problem. *IJCV*, 13(3):331--356, 1994. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 224, + 545, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 224, + 545, + 255 + ], + "spans": [ + { + "bbox": [ + 307, + 224, + 545, + 255 + ], + "type": "text", + "content": "[21] R. Hartley and A. Zisserman. Multiple View Geometry in Computer Vision. Cambridge University Press, 2nd edition, 2004. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 258, + 545, + 280 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 258, + 545, + 280 + ], + "spans": [ + { + "bbox": [ + 307, + 258, + 545, + 280 + ], + "type": "text", + "content": "[22] J. Heikkilä. Using sparse elimination for solving minimal problems in computer vision. In ICCV, 2017. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 282, + 545, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 282, + 545, + 304 + ], + "spans": [ + { + "bbox": [ + 307, + 282, + 545, + 304 + ], + "type": "text", + "content": "[23] J. Hesch and S. I. Roumeliotis. A Direct Least-Squares (DLS) method for PnP. In ICCV, 2011. 1, 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 307, + 545, + 339 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 307, + 545, + 339 + ], + "spans": [ + { + "bbox": [ + 307, + 307, + 545, + 339 + ], + "type": "text", + "content": "[24] R. J. Holt and A. N. Netravali. Uniqueness of solutions to three perspective views of four points. IEEE TPAMI, 17(3): 303--307, 1995. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 342, + 545, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 342, + 545, + 363 + ], + "spans": [ + { + "bbox": [ + 307, + 342, + 545, + 363 + ], + "type": "text", + "content": "[25] P. Hruby, T. Duff, A. Leykin, and T. Pajdla. Learning to solve hard minimal problems. In CVPR, 2022. 1, 2, 4" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 365, + 545, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 365, + 545, + 398 + ], + "spans": [ + { + "bbox": [ + 307, + 365, + 545, + 398 + ], + "type": "text", + "content": "[26] A. Kendall and R. Cipolla. Geometric loss functions for camera pose regression with deep learning. In CVPR, pages 5974-5983, 2017. 3, 4" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 400, + 545, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 400, + 545, + 432 + ], + "spans": [ + { + "bbox": [ + 307, + 400, + 545, + 432 + ], + "type": "text", + "content": "[27] J. Kileel. Minimal problems for the calibrated trifocal variety. SIAM Journal on Applied Algebra and Geometry, 1(1): 575-598, 2017. 2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 307, + 436, + 545, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 436, + 545, + 479 + ], + "spans": [ + { + "bbox": [ + 307, + 436, + 545, + 479 + ], + "type": "text", + "content": "[28] Laurent Kneip and Hongdong Li. Efficient computation of relative pose for multi-camera systems. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 446-453, 2014. 6" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 307, + 481, + 545, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 481, + 545, + 524 + ], + "spans": [ + { + "bbox": [ + 307, + 481, + 545, + 524 + ], + "type": "text", + "content": "[29] L. Kneip, D. Scaramuzza, and R. Siegwart. A novel parametrization of the perspective-three-point problem for a direct computation of absolute camera position and orientation. In CVPR, 2011. 1" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 307, + 527, + 545, + 559 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 527, + 545, + 559 + ], + "spans": [ + { + "bbox": [ + 307, + 527, + 545, + 559 + ], + "type": "text", + "content": "[30] L. Kneip, R. Siegwart, and M. Pollefeys. Finding the exact rotation between two images independently of the translation. In ECCV, 2012. 1" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 307, + 563, + 545, + 594 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 563, + 545, + 594 + ], + "spans": [ + { + "bbox": [ + 307, + 563, + 545, + 594 + ], + "type": "text", + "content": "[31] L. Kneip, H. Li, and Y. Seo. UPnP: An optimal O(n) solution to the absolute pose problem with universal applicability. In ECCV, 2014. 1, 4, 5" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 307, + 597, + 545, + 630 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 597, + 545, + 630 + ], + "spans": [ + { + "bbox": [ + 307, + 597, + 545, + 630 + ], + "type": "text", + "content": "[32] L. Kneip, C. Sweeney, and R. Hartley. The generalized relative pose and scale problem: View-graph fusion via 2d-2d registration. 2016. 2, 6, 7" + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 307, + 632, + 545, + 654 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 632, + 545, + 654 + ], + "spans": [ + { + "bbox": [ + 307, + 632, + 545, + 654 + ], + "type": "text", + "content": "[33] Z. Kukelova, M. Bujnak, and T. Pajdla. Automatic generator of minimal problem solvers. In ECCV, 2008. 1, 2" + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 307, + 656, + 545, + 688 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 656, + 545, + 688 + ], + "spans": [ + { + "bbox": [ + 307, + 656, + 545, + 688 + ], + "type": "text", + "content": "[34] V. Larsson, K. Aström, and M. Oskarsson. Efficient solvers for minimal problems by syzygy-based reduction. In CVPR, 2017. 2" + } + ] + } + ], + "index": 38 + }, + { + "bbox": [ + 307, + 691, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 691, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 307, + 691, + 545, + 712 + ], + "type": "text", + "content": "[35] V. Larsson, K. Aström, and M. Oskarsson. Polynomial solvers for saturated ideals. In ICCV, 2017. 2" + } + ] + } + ], + "index": 39 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27111" + } + ] + } + ], + "index": 41 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 117 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 117 + ], + "type": "text", + "content": "[36] Viktor Larsson, Kalle Astrom, and Magnus Oskarsson. Efficient solvers for minimal problems by syzygy-based reduction. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 820-829, 2017. 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 118, + 287, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 118, + 287, + 150 + ], + "spans": [ + { + "bbox": [ + 49, + 118, + 287, + 150 + ], + "type": "text", + "content": "[37] V. Larsson, Z. Kukelova, and Y. Zheng. Making minimal solvers for absolute pose estimation compact and robust. In ICCV, 2017. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 151, + 287, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 151, + 287, + 184 + ], + "spans": [ + { + "bbox": [ + 48, + 151, + 287, + 184 + ], + "type": "text", + "content": "[38] V. Larsson, M. Oskarsson, K. Aström, A. Wallis, Z. Kukelova, and T. Pajdla. Beyond gröbner bases: Basis selection for minimal solvers. In CVPR, 2018. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 185, + 287, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 185, + 287, + 207 + ], + "spans": [ + { + "bbox": [ + 48, + 185, + 287, + 207 + ], + "type": "text", + "content": "[39] A. Leykin. Numerical algebraic geometry. Journal of Software for Algebra and Geometry, 3(1):5-10, 2011. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 208, + 287, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 208, + 287, + 262 + ], + "spans": [ + { + "bbox": [ + 48, + 208, + 287, + 262 + ], + "type": "text", + "content": "[40] Evgeniy Martyushev, Jana Vrablikova, and Tomas Pajdla. Optimizing elimination templates by greedy parameter search. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15754-15764, 2022. 2, 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 263, + 287, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 263, + 287, + 297 + ], + "spans": [ + { + "bbox": [ + 48, + 263, + 287, + 297 + ], + "type": "text", + "content": "[41] P. Miraldo, T. Dias, and S. Ramalingam. A minimal closed-form solution for multi-perspective pose estimation using points and lines. In ECCV, 2018. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 297, + 287, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 297, + 287, + 341 + ], + "spans": [ + { + "bbox": [ + 48, + 297, + 287, + 341 + ], + "type": "text", + "content": "[42] V. Monga, Y. Li, and Y. C. Eldar. Algorithm unrolling: Interpretable, efficient deep learning for signal and image processing. IEEE Signal Processing Magazine, 38(2):18-44, 2021. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 342, + 287, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 342, + 287, + 374 + ], + "spans": [ + { + "bbox": [ + 48, + 342, + 287, + 374 + ], + "type": "text", + "content": "[43] R. Mur-Artal, J. M. M. Montiel, and Tardós. J. D. ORB-SLAM: A Versatile and Accurate Monocular SLAM System. 31(5), 2015. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 376, + 287, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 376, + 287, + 398 + ], + "spans": [ + { + "bbox": [ + 48, + 376, + 287, + 398 + ], + "type": "text", + "content": "[44] D. Niste'r. An efficient solution to the five-point relative pose problem. IEEE TPAMI, 26(6):756-770, 2004. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 399, + 287, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 399, + 287, + 430 + ], + "spans": [ + { + "bbox": [ + 48, + 399, + 287, + 430 + ], + "type": "text", + "content": "[45] D. Niste'r and F. Schaffalitzky. Four points in two or three calibrated views: Theory and practice. *IJCV*, 67(2):211–231, 2006. 1, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 432, + 287, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 432, + 287, + 453 + ], + "spans": [ + { + "bbox": [ + 48, + 432, + 287, + 453 + ], + "type": "text", + "content": "[46] D. Nistér, O. Naroditsky, and J. Bergen. Visual odometry. In CVPR, 2004. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 455, + 287, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 455, + 287, + 487 + ], + "spans": [ + { + "bbox": [ + 48, + 455, + 287, + 487 + ], + "type": "text", + "content": "[47] L. Quan, B. Triggs, and B. Mourrain. Some results on minimal euclidean reconstruction from four points. 24(3):341-348, 2006. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 488, + 287, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 488, + 287, + 521 + ], + "spans": [ + { + "bbox": [ + 48, + 488, + 287, + 521 + ], + "type": "text", + "content": "[48] R. Raguram, O. Chum, M. Pollefeys, J. Matas, and J.-M. Frahm. SAC: A universal framework for random sample consensus. IEEE TPAMI, 35(8):2022-2038, 2013. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 522, + 287, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 522, + 287, + 566 + ], + "spans": [ + { + "bbox": [ + 48, + 522, + 287, + 566 + ], + "type": "text", + "content": "[49] P.-E. Sarlin, A. Unagar, M. Larsson, H. Germain, C. Toft, V. Larsson, M. Pollefeys, V. Lepetit, L. Hammarstrand, F. Kahl, et al. Back to the feature: Learning robust camera localization from pixels to pose. In CVPR, 2021. 3, 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 567, + 287, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 567, + 287, + 621 + ], + "spans": [ + { + "bbox": [ + 48, + 567, + 287, + 621 + ], + "type": "text", + "content": "[50] Paul-Edouard Sarlin, Mihai Dusmanu, Johannes L. Schonberger, Pablo Speciale, Lukas Gruber, Viktor Larsson, Ondrej Miksik, and Marc Pollefeys. LaMAR: Benchmarking Localization and Mapping for Augmented Reality. In ECCV, 2022. 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 623, + 287, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 623, + 287, + 645 + ], + "spans": [ + { + "bbox": [ + 48, + 623, + 287, + 645 + ], + "type": "text", + "content": "[51] O. Saurer, M. Pollefeys, and G. H. Lee. A minimal solution to the rolling shutter pose estimation problem. 2015. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 646, + 287, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 646, + 287, + 667 + ], + "spans": [ + { + "bbox": [ + 48, + 646, + 287, + 667 + ], + "type": "text", + "content": "[52] J. Schonberger and J.-M. Frahm. Structure-from-motion revisited. In CVPR, 2016. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 48, + 669, + 287, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 669, + 287, + 690 + ], + "spans": [ + { + "bbox": [ + 48, + 669, + 287, + 690 + ], + "type": "text", + "content": "[53] R. Sheffer and A. Wiesel. PnP-Net: A hybrid Perspective-n-Point Network, 2020. 3, 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "text", + "content": "[54] N. Snavely, S. M. Seitz, and R. Szeliski. Photo tourism: exploring photo collections in 3D. 2006. 1" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 72, + 545, + 478 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 307, + 72, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 72, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 545, + 95 + ], + "type": "text", + "content": "[55] N. Snavely, S. M. Seitz, and R. Szeliski. Modeling the world from internet photo collections. pages 189-210, 2008. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 96, + 545, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 96, + 545, + 128 + ], + "spans": [ + { + "bbox": [ + 307, + 96, + 545, + 128 + ], + "type": "text", + "content": "[56] Andrew J Sommese, Charles W Wampler, et al. The Numerical solution of systems of polynomials arising in engineering and science. World Scientific, 2005. 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 129, + 545, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 129, + 545, + 173 + ], + "spans": [ + { + "bbox": [ + 307, + 129, + 545, + 173 + ], + "type": "text", + "content": "[57] H. Stewenius, D. Nister, M. Oskarsson, and K. Aström. Solutions to minimal generalized relative pose problems. In Workshop on Omnidirectional Vision in conjunction with ICCV, 2005. 1, 2, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 175, + 545, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 175, + 545, + 207 + ], + "spans": [ + { + "bbox": [ + 307, + 175, + 545, + 207 + ], + "type": "text", + "content": "[58] H. Stewenius, C. Engels, and D. Niste'r. Recent developments on direct relative orientation. ISPRS J. of Photogrammetry and Remote Sensing, 60:284-294, 2006. 1, 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 208, + 545, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 208, + 545, + 229 + ], + "spans": [ + { + "bbox": [ + 307, + 208, + 545, + 229 + ], + "type": "text", + "content": "[59] B. Sturmfels. Solving systems of polynomial equations, 2002. 1, 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 232, + 545, + 253 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 232, + 545, + 253 + ], + "spans": [ + { + "bbox": [ + 307, + 232, + 545, + 253 + ], + "type": "text", + "content": "[60] J. Ventura, C. Arth, and V. Lepetit. An efficient minimal solution for multi-camera motion. In ICCV, 2015. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 255, + 545, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 255, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 307, + 255, + 545, + 285 + ], + "type": "text", + "content": "[61] J. Verschelde. Polynomial homotopy continuation with phc-pack. ACM Commun. Comput. Algebra, 44(3/4):217—220, 2010. 1, 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 288, + 545, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 288, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 307, + 288, + 545, + 319 + ], + "type": "text", + "content": "[62] W. Xu, H. Lan, M. C. Tsakiris, and L. Kneip. Online stability improvement of gröbner basis solvers using deep learning. 2019. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 322, + 545, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 322, + 545, + 354 + ], + "spans": [ + { + "bbox": [ + 307, + 322, + 545, + 354 + ], + "type": "text", + "content": "[63] J. Zhao, L. Kneip, Y. He, and J. Ma. Minimal case relative pose computation using ray-point-ray features. IEEE TPAMI, 42:1176-1190, 2020. 1, 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 356, + 545, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 356, + 545, + 388 + ], + "spans": [ + { + "bbox": [ + 307, + 356, + 545, + 388 + ], + "type": "text", + "content": "[64] Y. Zheng and L. Kneip. A direct least-squares solution to the PnP problem with unknown focal length. In CVPR, 2016. 1, 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 307, + 390, + 545, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 390, + 545, + 422 + ], + "spans": [ + { + "bbox": [ + 307, + 390, + 545, + 422 + ], + "type": "text", + "content": "[65] Y. Zheng, Y. Kuang, S. Sugimoto, K. Astrom, and M. Okutomi. Revisiting the pnp problem: A fast, general and optimal solution. In ICCV, 2013. 1, 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 307, + 423, + 545, + 478 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 423, + 545, + 478 + ], + "spans": [ + { + "bbox": [ + 307, + 423, + 545, + 478 + ], + "type": "text", + "content": "[66] Yi Zhou, Connelly Barnes, Jingwan Lu, Jimei Yang, and Hao Li. On the continuity of rotation representations in neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5745-5753, 2019. 6" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "27112" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/SinGS_ Animatable Single-Image Human Gaussian Splats with Kinematic Priors/a1907643-3cec-4712-b625-aae92326d106_content_list.json b/2025/SinGS_ Animatable Single-Image Human Gaussian Splats with Kinematic Priors/a1907643-3cec-4712-b625-aae92326d106_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..e3367e89506f77033a6efffee09ded38215ec53f --- /dev/null +++ b/2025/SinGS_ Animatable Single-Image Human Gaussian Splats with Kinematic Priors/a1907643-3cec-4712-b625-aae92326d106_content_list.json @@ -0,0 +1,1196 @@ +[ + { + "type": "text", + "text": "SinGS: Animatable Single-Image Human Gaussian Splats with Kinematic Priors", + "text_level": 1, + "bbox": [ + 76, + 130, + 890, + 151 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yufan Wu $^{1*}$ Xuanhong Chen $^{1\\dagger}$ Wen Li $^{1,3}$ Shunran Jia $^{2}$ Hualiang Wei $^{1*}$ Kairui Feng $^{4}$ \nJialiang Chen $^{1}$ Yuhan Li $^{1}$ Ang He $^{1*}$ Weimin Zhang $^{1}$ Bingbing Ni $^{1}$ Wenjun Zhang $^{1}$", + "bbox": [ + 109, + 179, + 856, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Shanghai Jiao Tong University, Shanghai, China $^{2}$ DreamX Inc $^{3}$ Akool Research $^{4}$ The National Key Laboratory of Autonomous Intelligent Unmanned Systems, Shanghai, China", + "bbox": [ + 107, + 215, + 861, + 251 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{chen19910528,wmzhang,nibingbing,zhangwenjun}@sjtu.edu.cn, elvinfkr@tongji.edu.cn", + "bbox": [ + 119, + 253, + 834, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 304, + 312, + 319 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite significant advances in accurately estimating geometry in contemporary single-image 3D human reconstruction, creating a high-quality, efficient, and animatable 3D avatar remains an open challenge. Two key obstacles persist: incomplete observation and inconsistent 3D priors. To address these challenges, we propose SinGS, aiming to achieve high-quality and efficient animatable 3D avatar reconstruction. At the heart of SinGS are two key components: Kinematic Human Diffusion and Geometry-Preserving 3D Gaussain Splatting. The former is a foundational human model that samples within pose space to generate a highly 3D-consistent and high-quality sequence of human images, inferring unseen viewpoints and providing kinematic priors. The latter is a system that reconstructs a compact, high-quality 3D avatar even under imperfect priors, achieved through a novel semantic Laplacian regularization and a geometry-preserving density control strategy that enable precise and compact assembly of 3D primitives. Extensive experiments demonstrate that SinGS enables life-like, animatable human reconstructions, maintaining both high quality and inference efficiency (up to 70FPS).", + "bbox": [ + 75, + 337, + 473, + 654 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 685, + 207, + 700 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Single-image 3D human reconstruction (i.e., SIHR) is an emerging modeling technique that aims to create a lifelike 3D avatar from just one image, offering a highly cost-effective and promising approach to 3D modeling. Recently, 3D Gaussian Splatting (3DGS)[14] has made significant breakthroughs in rendering efficiency and quality, establishing itself as a mainstream approach for 3D avatar reconstruction. However, current 3DGS-based methods [23, 26] for single-image human 3D avatar reconstruction suffer from critical limitations in animation, as most are", + "bbox": [ + 75, + 710, + 468, + 862 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9676d7bbff1e85aed3304a363715f481c4e2de0f9887b0cac347845423235752.jpg", + "image_caption": [ + "Figure 1. Existing single-image 3D avatar reconstruction methods are limited to sampling prior information from camera space. In contrast, SinGS proposes sampling from pose space, allowing us to obtain both camera and kinematic priors simultaneously. Combined with our Geometry-Preserving 3D Gaussian Splitting, this enables highly efficient animation rendering." + ], + "image_footnote": [], + "bbox": [ + 506, + 305, + 888, + 609 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "designed primarily for static assets. This paper aims to develop an animation reconstruction framework that achieves a high balance between efficiency and rendering quality.", + "bbox": [ + 496, + 718, + 890, + 763 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Currently, the vast majority of SIHR methods [7, 10, 23, 24, 26, 40, 41, 46, 49, 50] focus solely on optimizing geometry while neglecting to enhance animation capabilities. As a result, they are often incapable of animation or require third-party tools [7] to achieve approximate motion. Unlike digital assets, the primary requirement for a human avatar is its ability to animate, making animation the core goal of SIHR [9]. Achieving animatable single-image human reconstruction (i.e., Animatable SIHR) is particularly", + "bbox": [ + 496, + 763, + 892, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Work done during an internship at Shanghai Jiao Tong University.", + "bbox": [ + 94, + 875, + 449, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author: Xuanhong Chen.", + "bbox": [ + 96, + 887, + 315, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "5571", + "bbox": [ + 482, + 945, + 513, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "challenging for two main reasons: Incomplete Observation: A single image lacks comprehensive viewpoint information, offering an extremely limited, partial observation that cannot address occlusion issues. Current methods [7, 10, 47] mainly rely on fine-tuned diffusion models to synthesize (i.e., 3D hallucination) missing viewpoint information (such as front and back perspectives). However, this approach does not resolve the absence of motion information. While synthesized views can alleviate some occlusion issues, the human body's non-rigid structure requires more than static viewpoints to capture its complex kinematic patterns effectively. Inconsistent 3D Prior: Current mainstream methods [5, 7, 10, 47] use Score Distillation Sampling (SDS) [25] to leverage diffusion models as supervisory modules for learning 3D priors. However, the high generative diversity of these models often leads to viewpoint inconsistencies, resulting in implausible imagery (e.g., poor-quality hidden views and inconsistencies with the input image). HumanRef [47] attempts to mitigate the instability of SDS through ReferenceNet [8], but observable inconsistencies remain widespread in the results of ReferenceNet-based methods [8, 33], indicating the difficulty in fully resolving this issue. Consequently, 3D representations should be robust enough to handle imperfect 3D priors, though this perspective has largely been overlooked in current single-image avatar frameworks.", + "bbox": [ + 76, + 90, + 472, + 484 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address these challenges, this paper introduces a novel single-image 3D human avatar reconstruction framework, named SinGS. Unlike existing methods [23, 26], SinGS focuses on directly constructing an animatable 3DGS avatar. At the core of our SinGS framework are two key components: Kinematic Human Diffusion and a highly Geometry-Preserving 3D Gaussian Splatting. Specifically, Kinematic Human Diffusion (KHD) is designed to sample a highly 3D-consistent human sequence within the pose space, based solely on a single input image prompt. This approach enables high-quality and animatable 3D avatar reconstruction. To train KHD, we constructed a dataset with over ten thousand rotation videos, allowing KHD to learn a rich variety of human structures and texture distributions from different angles, while also gaining extensive knowledge of human kinematics. Unlike existing frameworks [5, 7, 10, 47] that rely on costly and scarce multi-view human data, our approach requires only more affordable and abundant video data, significantly reducing KHD's training costs and better aligning with scaling laws. Geometry-Preserving 3D Gaussian Splatting (i.e., GPGS) tackles artifact and hole issues under imperfect human priors (e.g., minor jitter and 3D inconsistencies) through two key designs: Semantic Laplacian Regularization addresses floating Gaussian spheres caused by 3D inconsistencies. It adjusts the constraint strength based on body parts, allowing more flexibility for areas like hair and tighter constraints", + "bbox": [ + 76, + 493, + 472, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "for facial details, reducing floating Gaussian spheres while maintaining rendering quality. Geometry-Preserving Density Control resolves over-deletion and incorrect deletion of spheres in the training process. By managing the Gaussian sphere density via mesh edges, it ensures the spheres remain around the mesh, preserving geometry and preventing large holes. Extensive experiments have validated that our SinGS can generate high-quality and animatable 3D avatars at ultra-fast inference speeds (up to 70 FPS) with efficient resource usage (requiring only 3.8GB of GPU memory).", + "bbox": [ + 496, + 90, + 893, + 244 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 253, + 640, + 271 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Diffusion-based Human Generative Models", + "text_level": 1, + "bbox": [ + 500, + 279, + 867, + 294 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recent advances in diffusion-based human generative models, especially those conditioned on pose, have improved model controllability. Frameworks like ControlNet [48], IP Adapter [44], and ReferenceNet [8], along with better human attribute detection (2D/3D pose estimation, SMPL parameters [18]), enable more precise and consistent results. In image synthesis, models like PIDM [1] and LFDM [22] enhance pose transfer and image adaptation, while LEO [38] uses flow maps for motion representation. For video, DreamPose [13] and DisCo [35] improve pose-to-image translation and control separation. Models likeAnimateAnyone [8], MagicAnimate [42], and Magic-Pose [4] refine pose conditioning with ReferenceNet. Despite these improvements, these models remain resource-heavy and struggle with stability when generating images of complex movements, such as head turns or rotations.", + "bbox": [ + 496, + 301, + 893, + 544 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Single-Image Human Reconstruction", + "text_level": 1, + "bbox": [ + 500, + 551, + 821, + 569 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Monocular video-based human reconstruction has reduced the need for specialized equipment, but single-image reconstruction remains challenging due to incomplete data. Early methods like PIFu [29] and SiCloPe [21] used pixel-aligned features and image-to-image translation, with PIFuHD [30] enhancing detail via normal maps. Later methods, such as ARCH [11] and ICON [40], incorporated body priors for animatable avatars, and ECON [41] improved model completeness by separately reconstructing front and back surfaces. Recent approaches like SiTH [7] and HumanRef [47] use generative models to fill in missing data, while SIFU [50] and Diffusion-FOF approximate multi-view features. R-Cyclic Diffuser [3] addresses limitations in Zero-1-to-3 [17], even with human-specific datasets, advancing single-image human reconstruction through generative priors and viewpoint conditioning.", + "bbox": [ + 496, + 575, + 893, + 819 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 500, + 830, + 635, + 847 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Given a target image $I_r$ , the objective of the SinGS architecture is to reconstruct a fully animatable 3D avatar $\\mathbf{A}_t$ based on $I_r$ . To address the challenge of limited viewpoint and", + "bbox": [ + 496, + 854, + 893, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "5572", + "bbox": [ + 482, + 944, + 516, + 957 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/e7a026c2e6e41bf24d06eb4e3d6269db01b5c71d11ad0ad2a3e81164ccea5dfd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 76, + 87, + 496, + 323 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/952abdf13459951d7c01846d40b42bfdd5e314026774502637291596384d9a15.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 496, + 88, + 890, + 321 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/08ab15398cc7fd693842e859d7b8824d277dbc012ecfe7c884450a93e98e4c62.jpg", + "image_caption": [ + "Figure 2. Overall framework of our method." + ], + "image_footnote": [], + "bbox": [ + 78, + 324, + 890, + 416 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "kinematic information inherent in single-image reconstruction, $I_{t}$ is first input into the Kinematic Human Diffusion (i.e., KHD) module to complete missing information, generating an augmented sequence of rendered images $S_{t} = \\{I_{p_{1}},\\dots ,I_{p_{i}},\\dots ,I_{p_{N}}\\}$ , where $N$ represents the total number of poses generated, and $p_i$ indicates each corresponding 3D human pose. KHD reframes the less-constrained, underdetermined 3D reconstruction problem as a generative task, thus producing a fully-constrained, fully-observed 3D reconstruction formulation. Finally, $S_{t}$ is fed into our compact 3DGS module to reconstruct an animatable 3D human avatar. The compact 3DGS module is capable of reconstructing a watertight, high-quality human model even with imperfect 3D priors. The distilled avatar is high-quality, computationally efficient, and well-suited for promising applications in AR, VR, and similar fields.", + "bbox": [ + 75, + 454, + 473, + 696 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Kinematic Priors", + "text_level": 1, + "bbox": [ + 76, + 712, + 246, + 726 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1.1 Kinematic Human Diffusion", + "text_level": 1, + "bbox": [ + 76, + 737, + 328, + 750 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Unlike existing hallucination architectures [7, 10, 50], our Kinematic Human Diffusion does not rely on camera input, which is prone to estimation errors and has the added challenges of multi-view data scarcity and high costs (e.g., multi-view human data typically requires capture with specialized setups like a light stage). Our approach instead posits that sampling within the motion space provides complete priors for viewpoint and kinematic features. Additionally, our KHD does not adopt a ReferenceNet structure [8];", + "bbox": [ + 75, + 763, + 468, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "while ReferenceNet has been highly successful, it requires maintaining two UNet models, significantly increasing the framework's resource demands and preventing integration of longer contexts (e.g., 24 frames). A shorter contextual window cannot provide an image sequence with a sufficient diversity of motion samples (e.g., 150 frames). Considering the above two aspects, our KHD is designed as a diffusion structure, as illustrated in Figure 2. KHD $\\mathcal{G}$ takes a single image $I_{r}$ as input and, constrained by a randomly sampled sequence of poses $\\{p_1,p_2,\\dots ,p_N\\}$ , augments the input image into a sequence $S_{t} = \\{I_{p_{1}},I_{p_{2}},\\ldots ,I_{p_{N}}\\}$ . This process can be formulated as follows:", + "bbox": [ + 496, + 454, + 890, + 635 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nS _ {t} = \\mathcal {G} \\left(I _ {r}; p _ {1}, p _ {2}, \\dots , p _ {N}\\right). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 607, + 647, + 890, + 662 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Model Architecture. Inspired by UniAnimate [36], our KHD also removes the ReferenceNet and directly inputs the reference image $I_r$ and the driving sequence into the UNet. The reference image $I_r$ is encoded by the VAE encoder of the diffusion model into a latent code $z_r$ , serving as an appearance reference throughout KHD. Additionally, $I_r$ is encoded by the CLIP image encoder [27] into features $f_r$ which are injected into the network through cross-attention by replacing the original text encoder's features in the diffusion model. This enables high-fidelity preservation of the subject's appearance characteristics. Existing works typically use 2D poses (e.g., OpenPose [2], DW-Pose [43]) as the driving signal. However, our experiments reveal that 2D poses fail to guide the diffusion model accurately in scenarios with extensive occlusions, such as when the subject", + "bbox": [ + 496, + 672, + 892, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "5573", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "is turning. Moreover, 2D poses cannot reliably differentiate between different viewpoints of the human body. Therefore, we use SMPL refined by ScoreHMR [31] as the driving signal $\\{p_1,p_2,\\dots,p_N\\}$ . It is worth noting that Champ [51] also uses SMPL as guidance; however, their SMPL estimations lack accuracy, and they rely on overly strict guidance inputs, such as depth and normal vectors. This approach leads to suboptimal results, performing even worse than 2D poses in practice, for instance, by generating inconsistent body shapes. The SMPL parameters are converted into a mesh model in the specified pose using Linear Blend Skinning (LBS) [18]. This mesh is rendered as an RGB image based on the fixed camera pose maintained throughout the process. The rendered image is then encoded by a 3D Pose Encoder, a network composed of several convolutional layers, into a latent code $z_{p_i}$ with the same dimensionality as $z_{r}$ . All $z_{p_i}$ codes are first concatenated with the input noise along the channel dimension, and then sequentially along the temporal dimension to obtain the final representation $z_{P} \\in \\mathbb{R}^{T \\times C \\times H \\times W}$ . Simultaneously, $I_{r}$ is processed to estimate and encode its pose latent code $z_{p_r}$ , which is then concatenated with $z_{P}$ along the temporal dimension before being fed into the diffusion model's backbone. Following most video generation models [36, 37], the UNet component of the diffusion model is also designed as a 3D UNet to ensure 3D consistency in the generated output.", + "bbox": [ + 76, + 90, + 472, + 484 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Training Dataset and Strategy. Human image animation [8, 36] primarily focuses on generating diverse frontal human movements (e.g., dancing). As a result, their training data (e.g., TikTok [12]) is largely composed of performance videos of dancers, typically featuring frontal views and wide movement ranges. In our experiments, we found that models trained on this type of data provide limited support for our 3D reconstruction tasks. These models often fail to generate results with high consistency, exhibit significant flickering in fine details, and display abnormal motion in clothing dynamics. In response to this, we collected over 6,000 high-definition videos (each around 20 seconds long) from YouTube, featuring activities such as model showcases and dancing, with a focus on segments that include turning motions. Additionally, we incorporated 600 clips from the UBC Fashion dataset [45], primarily showcasing model runway turns. As a key part of our dataset, we also integrated frontal-view data from MVHumanNet [39]. We used SAM2 [28] to segment human figures from the background in MVHumanNet, then merged the results with a white background to create videos with a consistent background, effectively eliminating the background variability caused by bounding boxes in MVHumanNet. These three sources together form a dataset of over 10,000 training videos, which we call RotationHuman. Notably, compared to multi-view data, this type of data collection is significantly more cost-effective and thus better", + "bbox": [ + 76, + 493, + 472, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "suited for building large-scale datasets. During training, our KHD generates 64 pose-driven images in a single pass, setting the context window to 64 frames (i.e., $N = 64$ ). In the inference stage, we directly generate 128 frames in one pass. The advantage of one-pass inference is that it avoids consistency jitter in overlapping transition frames. KHD training requires only one phase, where the objective is to reconstruct the ground truth images corresponding to each pose, with the training loss formulated as follows:", + "bbox": [ + 496, + 90, + 893, + 227 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathbb {E} _ {\\theta} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} \\left(z _ {t}, f _ {r}, t\\right) \\| ^ {2} \\right], \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 599, + 241, + 890, + 258 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where, $t$ denotes the denoising step, and $z_{t}$ represents the noise at step $t$ . Note that $z_{t}$ in KHD contains diffusion states for $N$ different driving poses. The feature $f_{r}$ is the CLIP image encoding of the reference image $I_{r}$ , used to ensure ID consistency in the generated results. $\\theta$ represents the parameters of the 3D UNet and the 3D Pose Encoder. The overall training process involves learning to reconstruct the target video based on $I_{r}$ as input and the pose set $\\{p_1, p_2, \\dots, p_N\\}$ as conditioning. The generated sequence $S_{t}$ represents the sampling results of the subject in the input image $I_{r}$ within the pose space, and it can be directly used for 3D avatar reconstruction.", + "bbox": [ + 496, + 263, + 893, + 444 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. 3D Distillation with Geometry-Preserving 3D Gaussian Splitting", + "text_level": 1, + "bbox": [ + 498, + 454, + 890, + 486 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Once we obtain the sequence of pose images sampled from KHD, the goal is to distill a animatable and real-time rendered human avatar using Geometry-Preserving 3D Gaussian Splatting (GPGS). This challenge can be addressed based on mainstream approaches [16, 32, 52] in monocular 3D reconstruction. As illustrated in the Figure 2 on the right, we initialize the positions of Gaussians using the human template mesh that has undergone loop subdivision. The initial positions are set to the vertex locations, with a default of 2 upsampling iterations. This typically results in approximately $110.6K$ Gaussians. To model the avatar representation, all vertices are fed into a learnable triplane [52] to extract vertex-wise features. These features are subsequently passed to MLPs to predict the geometry and appearance attributes of the Gaussians, respectively. The entire modeling process occurs in canonical space, and we finally map this canonical space into posed space using Linear Blend Skinning (LBS) based on the SMPL framework.", + "bbox": [ + 496, + 493, + 893, + 765 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.1 Semantic Laplacian Regularization", + "text_level": 1, + "bbox": [ + 498, + 785, + 802, + 800 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Even during the reconstruction of real video, it is challenging to recover perfect static 3D consistency. Therefore, while the generated videos may appear visually plausible, every tiny ambiguity during lifting the 2D outcomes into 3D space potentially leads to catastrophic issues. Early experimental results indicated that common reconstruction meth", + "bbox": [ + 496, + 809, + 893, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "5574", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/79c841c208901a493b99c45a5bd7271df7ea2ff9553df4900229e904929fd7b1.jpg", + "image_caption": [ + "Figure 3. Visual comparison of our method with TeCH [10] and SiTH [7] for single-image reconstruction. We reposed the reconstructed results to A Pose, showcasing both animation and reconstruction capabilities. We highly recommend checking the suppl. for more results. We have provided a large number of animation result images and videos in the suppl." + ], + "image_footnote": [], + "bbox": [ + 78, + 92, + 890, + 523 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "ods produced results that lacked smoothness and compactness, resulting in existing disconnected and floating gaus-. sians. Following [20], we found that applying Laplacian regularization on the canonical mesh, both for geometry and appearance, effectively mitigated the presence of gaus-. sians out of control. However, it also constrained offsets. In other words, isometrical Laplacian constraint on the whole body overly relied on the accuracy of the SMPL mesh, also making it difficult to fit areas such as hair and clothing that deviate significantly from the SMPL surface. Considering the complexity of the human body and the impossibility to obtain extremely accurate SMPL estimation, we repartition all gaussian vertices into $L = 15$ distinct semantic parts and dynamically adjust the smoothing strength. Besides, we impose the constraint on the anchors points which are closer to the rendering surface than gausicians center. The Semantic Laplacian Regularization (SLR) is defined as follows:", + "bbox": [ + 75, + 592, + 472, + 849 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {S L R} = \\sum_ {u \\in \\Omega} \\lambda_ {u} \\left\\| a _ {u, v} - \\frac {1}{| M (u , v) |} \\sum_ {v \\in M (u, v)} a _ {u, v} \\right\\|, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 101, + 859, + 468, + 905 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\Omega$ represents the set of semantic regions, $a$ denotes the surface anchor point of the $v$ -th Gaussian vertex in the $u$ -th semantic region, and $M(u,v)$ refers to the set of adjacent points of the Gaussian vertex $a_{u,v}$ located within the $u$ -th semantic region. Experiments demonstrate that the Semantic Laplacian Regularization significantly impacts the handling of more intricate situations, particularly when dealing with errors between the SMPL mesh and the actual character, which can arise from both inaccurate estimations and complex clothing. We apply Laplacians in both position and color spaces; the former enhances the smoothness of the rendering surface, while the latter ensures a more continuous variation in color space.", + "bbox": [ + 496, + 592, + 893, + 789 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2.2 Geometry-Preserving Density Control Strategy", + "text_level": 1, + "bbox": [ + 498, + 814, + 880, + 830 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The strategies in vanilla 3DGS were not originally build for complex, dynamic structures like the human body. While works [20] have adapted the approach, such as initializing Gaussians based on SMPL meshes, these efforts pri", + "bbox": [ + 496, + 839, + 893, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5575", + "bbox": [ + 482, + 944, + 514, + 957 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/490809444f4498540e92eacd9069a893f3627339c7656ddc5b98ba9e7f2916ce.jpg", + "table_caption": [ + "Table 1. Quantitative comparisons with existing methods on UBC fashion dataset [45]. Note that metrics in the table are measure with the background-removed video. Some methods are unable to perform animation, and therefore, these metrics are not available for those methods. Bold indicates the best results, while Underlined represents the second-best results." + ], + "table_footnote": [], + "table_body": "
MethodPrior TypePSNR ↑LPIPS ↓FVD ↓Context. ↓CLIP Score ↑Speed↑
DisCo [35] (CVPR24)-30.980.259167.72.13482.4%1FPS
Animate Anyone [8] (CVPR24)-31.330.238148.81.90186.6%2FPS
Champ [51] (ECCV24)-31.210.251155.12.03884.3%1FPS
UniAnimate [36] (ArXiv24)-32.310.211141.31.81187.1%2FPS
SHERF [9] (ICCV23)-27.390.301202.32.99870.1%
SIFU [50] (CVPR24)Camera---2.89172.1%5FPS
GST [26] (ArXiv24)Camera---3.01171.7%50FPS
TeCH [10] (3DV24)Camera29.010.271188.12.33179.3%6FPS
SiTH [7] (CVPR24)Camera---2.77674.7%<1FPS
HumanRef [47] (CVPR24)Camera---2.16781.5%3FPS
SHERT [46] (CVPR2024)Inpainting29.310.266179.82.31978.8%10FPS
UniAnimate + GPGS (ours)Kinematics29.970.251168.71.99282.1%70FPS
KHD (ours)-32.810.221134.41.76892.3%2FPS
SinGS (ours)Kinematics31.330.240156.31.84788.2%70FPS
", + "bbox": [ + 99, + 141, + 867, + 354 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "marily lay a solid foundation for the optimization process but fail to maintain this advantages throughout. Incompatibilities still arise in optimization stage, prompting us to propose a novel density control scheme that is inherently compatible with hybrid representations, designed for adaptive adjustment body-wide density. This scheme draws on mesh operation algorithms [6] to better address the unique challenges of human body. For primitives that locate in under-expression mesh areas, we subdivide these regions by adding gaussians on every edge of the selected geometric triangles, which creates much smaller triangles. And we skip the trivial post-smoothing for vertices as the position of Gaussians will continue to be optimized in the network. For candidate Gaussians that are selected due to excessively large scales, as well as all newly added Gaussians, we perform rescaling to adjust them to an average size. In the case of anomalous gaussians, especially those are almost transparent, we collapse relevant edges and replace them with single vertex, resulting in decimated adjacent faces. And the priority of selected edges is decided by gaussian attributes and edge length. This entire pruning process preserves the principal structure of the human body. Our strategy result in better reconstruction and animation as the optimization was conducted under connectivity and compactness of the entire human body rather than producing holes and stacks. Moreover, this approach synergistically enhances the effectiveness of our hybrid representation.", + "bbox": [ + 75, + 364, + 472, + 773 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 789, + 209, + 806 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 76, + 815, + 267, + 832 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details. Our KHD primarily utilizes the RotationHuman dataset we collected, which focuses on human pose image sequences involving turning motions. The specific structure of the 3D Pose Encoder is provided in", + "bbox": [ + 75, + 839, + 470, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "the supplementary materials. KHD was trained using 8 NVIDIA A100 80GB GPUs, with DeepSpeed and ZeRO2 employed to reduce memory usage. Similar toAnimate Anyone, we resized the video frames to a spatial dimension of $768 \\times 512$ . We used a context window size of 64 frames, and to enhance model scalability post-training, we implemented a random context window jitter strategy [36]. During training, the model randomly samples window sizes from 64 to 32 frames in an uniformly distributed manner. The training process utilized the AdamW optimizer [15], with a learning rate set to $5e - 5$ and DDPM with 1000 steps for noise sampling. During inference, DDIM with 30 steps was used to reduce inference time. For the 3DGS component, the Adam optimizer [15] was used with a learning rate of $1e - 3$ and a cosine annealing learning scheduler was applied to accelerate convergence. The input image sequence size for 3DGS was also $768 \\times 512$ . 3DGS was trained using a single NVIDIA 4090 24GB GPU.", + "bbox": [ + 496, + 364, + 890, + 635 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baseline Models. Our baseline models consist of two parts: the first part includes human image animation models, such as DisCo [35], Animation Anyone [8] $^{1}$ , Champ [51], and UniAnimate [36]. These models are used to compare our animation capability and the superiority of KHD. The second part consists of single-image 3D avatar reconstruction models, including SHERF [9], SIFU [50], GST [26], TeCH [10], SiTH [7], HumanRef [47], and SHERT [46].", + "bbox": [ + 496, + 636, + 890, + 757 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Metrics. We employ both PSNR and LPIPS to assess the quality of video reconstruction. To evaluate the animation capability of our model, we also use FVD [34] to measure the distribution discrepancy between the generated video and the original video, assessing the video generation performance. Furthermore, following HumanRef, we introduce contextual distance [19] and CLIP score [27] to evaluate the texture and semantic similarity between the rendered", + "bbox": [ + 496, + 757, + 890, + 878 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "1https://github.com/MooreThreads/Moore-AnimateAnyone", + "bbox": [ + 517, + 886, + 831, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "5576", + "bbox": [ + 482, + 944, + 514, + 957 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/3a42d13bfec4a738751303ba79faedc3dfb2230b70f53d76348c9b242b679bee.jpg", + "image_caption": [ + "Figure 4. Comparison between KHD and other SOTA human image animation methods. Please zoom in for better observation." + ], + "image_footnote": [], + "bbox": [ + 93, + 89, + 460, + 287 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "output and the input reference image.", + "bbox": [ + 76, + 338, + 326, + 354 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Qualitative Analysis", + "text_level": 1, + "bbox": [ + 76, + 364, + 269, + 382 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Singe-Image Reconstruction. We compare SinGS with two methods from the literature that claim to support animation (TeCH, SiTH). It is worth noting that during our experiments, we found that single-image animation with these methods was extremely difficult to execute, and in some cases, nearly impossible. We had to use third-party code ${}^{2}$ to achieve the reposing for some of these methods. We animate their reconstructed outputs to a common pose (A Pose) for comparison, which allows us to evaluate both reconstruction and animation capabilities. The comparison results are shown in Figure 3, SinGS generates results with better details and maintains high-quality animation, particularly on the in-the-wild data, where SinGS's visual performance significantly outperforms the other methods. This demonstrates the high feasibility and effectiveness of the kinematic prior.", + "bbox": [ + 75, + 388, + 468, + 631 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Kinematic Human Diffusion. KHD, as the core component of our pipeline, is compared with state-of-the-art human image animation methods, includingAnimate Anyone and UniAnimate. As shown in Figure 4, we drive the same character with all three methods, focusing on comparing the quality of their generated results, particularly for the turning pose. From the figure, it is clear that both UniAnimate and our KHD maintain higher consistency thanAnimate Anyone after the character turns, with KHD surpassing UniAnimate in detail. The details marked by the red arrows in Figure 4 reveals that UniAnimate generates incorrect details, a common issue when using 2D poses, as they often struggle to distinguish body orientation during back-facing poses, leading the model to generate erroneous images. This is detrimental to reconstruction, while KHD generates videos with much higher consistency.", + "bbox": [ + 75, + 632, + 470, + 875 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8571cb3aae5ca73f739c8b03c6c01711d6540f1a88c71bf7ebf6b1602c55dd38.jpg", + "image_caption": [ + "Figure 5. Fifty users rated the video quality, with UniAnimate scoring highest, followed closely by our model, which outperformed others. Our model is more efficient, using only 3.5 GB of memory and achieving 70 FPS, while UniAnimate requires over 12GB and 0.5 seconds per image." + ], + "image_footnote": [], + "bbox": [ + 517, + 87, + 874, + 181 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Geometry-Preserving 3D Gaussian Splatting. Although KHD provides high-consistency pose sampling results, it inevitably has some consistency flaws, meaning it can only offer imperfect prior information. Our geometry-preserving GS method is specifically designed to address this challenge. To validate the superiority of our Geometry-Preserving 3D Gaussian Splatting (GPGS) over general human modeling methods, we compared it with recent state-of-the-art (i.e., SOTA) methods, HAHA [32] and HUGS [16]. We generate a pose sequence using KHD and then performed reconstruction with all 3 methods. The reconstruction results are shown in Figure 7. As seen, both HAHA and HUGS fail to produce a compact 3D avatar, with significant holes or floating artifacts appearing in various areas, particularly during the turning motion. In contrast, our GPGS method is able to reconstruct a compact, watertight 3D avatar model.", + "bbox": [ + 496, + 275, + 893, + 532 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Quantitative Analysis", + "text_level": 1, + "bbox": [ + 500, + 544, + 702, + 560 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We selected 100 test videos from the UBC dataset for our quantitative analysis. For each video, we chose the first frame as the reference image and used the video's pose sequence as the driving signal. The performance of all methods was assessed by comparing the animation results with the original video to evaluate the differences. Video reconstruction metrics (including PSNR, LPIPS, and FVD) reflect the model's animation capabilities, while Contextual Distance and CLIP score allow for comparison with methods that do not support animation. The quantitative comparison results are shown in Table 1. Human image animation methods (i.e., the second to fifth rows in the table, including our KHD) outperform the reconstruction-based methods across the board. Among them, KHD surpasses the other methods in most metrics, and UniAnimate also demonstrates superior performance, which is expected due to the scale of training data and model size. While SinGS lags behind UniAnimate, it shows a significant advantage over other methods in some metrics. However, SinGS significantly outperforms diffusion-based animation methods in terms of inference efficiency and resource consumption. Our SinGS achieves 70 FPS during inference, with only", + "bbox": [ + 496, + 568, + 893, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_footnote", + "text": "$^{2}$ https://github.com/custom-humans/editable-humans", + "bbox": [ + 94, + 886, + 377, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "5577", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/0a5847d3fcf1cdcf6872f676ffbe0506b521a418c0fb7b16bc3e7fa5705b251f.jpg", + "table_caption": [ + "Table 2. The results of the ablation experiments on Semantic Laplacian Regularization (i.e., Semantic Lap. Reg.) and Geometry-Preserving Density Control (i.e., GPDC) in SinGS." + ], + "table_footnote": [], + "table_body": "
SettingPSNRLPIPSFVD
with Semantic Lap. Reg.31.330.240156.3
with KNN Lap. Reg.30.880.250169.9
w/o Lap. Reg.30.340.272181.3
with GPDC31.010.249166.1
naive Density Control [14]30.220.276190.2
", + "bbox": [ + 112, + 141, + 431, + 218 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.8 GB of memory usage (tested on a single NVIDIA 4090 GPU), while UniAnimate, Champ, andAnimate Anyone each require more than 0.4 seconds per image and consume over 12 GB of memory. These quantitative results strongly demonstrate that SinGS is highly effective and feasible. We believe that using KHD, a specialized foundational model, will greatly enhance 3D reconstruction performance while significantly reducing the stringent data requirements.", + "bbox": [ + 75, + 229, + 467, + 351 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "User Study. We invited 50 users to subjectively evaluate the animation results of SinGS. The evaluation was based on a scale from 0 to 10, with 10 being the highest score and 0 being the lowest. The results are visualized in Figure 5.", + "bbox": [ + 75, + 351, + 467, + 412 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Ablation Study", + "text_level": 1, + "bbox": [ + 76, + 422, + 228, + 439 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Kinematic Priors Analysis. The number of poses sampled by KHD has a significant impact on the 3D distillation process. We analyzed how the number of poses affects the animation capability of the final reconstructed model, and the results are shown in Figure 6. We incrementally increased the number of poses from 2 (which corresponds to the typical approach of generating front and rear views, as used by most methods) up to 150. We also tested the influence of different prior models (i.e., KHD, UniAnimate, and Animate Anyone) on the results. Additionally, we examined how different 3DGS reconstruction methods handle the use of prior knowledge. As shown, performance improves rapidly with an increase in the number of poses, but the rate of improvement slows significantly after reaching around 100 poses. Moore-AA + HAHA shows the slowest improvement, while KHD + GPGS exhibits the slowest decline in performance. This suggests that increasing the number of poses does not always lead to improved performance for two main reasons: first, the prior models have already provided sufficient information; and second, none of the prior models provide perfectly consistent priors. Excessive sampling can introduce a lot of inconsistent noise, which ultimately harms performance.", + "bbox": [ + 75, + 446, + 467, + 792 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Geometry-Preserving 3D Gaussian Splitting Analysis. We conducted ablation experiments on the two key designs in GPGS: Semantic Laplacian Regularization and GPDC, with the results shown in Table 2. By replacing and removing the Laplacian regularization, we observed that Semantic Laplacian consistently outperforms in terms of reconstruction metrics. Additionally, GPDC demonstrates a perfor", + "bbox": [ + 75, + 795, + 467, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b842dcf943409a52bdf762189b7e88ee69294668cd25f9d0b1b64457ba6422f0.jpg", + "image_caption": [ + "Figure 6. The comparison of the impact of augmented pose numbers on the performance of the 3D avatar models." + ], + "image_footnote": [], + "bbox": [ + 540, + 88, + 851, + 226 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b6dc8cd903053ead99689f69599e83ad168ab21e6ef0529d1e35b52fc60f9414.jpg", + "image_caption": [ + "Figure 7. Here is a comparison of our GPGS with two SOTA 3DGS avatar methods, HAHA [32] and HUGS [16]. All methods use images sampled by KHD as body prior." + ], + "image_footnote": [], + "bbox": [ + 501, + 292, + 885, + 459 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "mance advantage over the original density control strategy. The original strategy does not account for dynamic properties and geometric constraints, which makes it ineffective in addressing issues like floating and holes caused by data inconsistencies.", + "bbox": [ + 498, + 532, + 890, + 607 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 630, + 617, + 645 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we propose a new paradigm for single-image 3D avatar reconstruction. Unlike traditional methods that complete missing information in camera space, we sample in human kinematic space, enabling efficient capture of both viewpoints and motion traits. To achieve this, we introduce Kinematic Human Diffusion (KHD) for consistent pose-space generation, supported by our RotationHuman dataset, which focuses on turning motions. Additionally, we propose Geometry-Preserving 3D Gaussian Splating that reconstructs high-quality, animatable 3D avatars under imperfect human priors. This includes two key innovations: Semantic Laplacian regularization for compact, detailed constraints, and a geometry-preserving density control strategy to maintain stable human geometry. Together, these innovations allow SinGS to achieve high-quality, animatable 3D avatar reconstruction from a single image.", + "bbox": [ + 496, + 659, + 890, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "5578", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Acknowledgment", + "text_level": 1, + "bbox": [ + 76, + 90, + 246, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work supported by Shanghai Post-doctoral Excellence Programe (2024290). This work was also partly by Shanghai Municipal Bureau of Culture and Tourism and Meta-verse Technology Innovation Center of Shanghai Municipal Bureau of Culture and Tourism. Kairui Feng thanks to Xiaomi foundation.", + "bbox": [ + 76, + 114, + 468, + 204 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 217, + 173, + 233 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Ankan Kumar Bhunia, Salman H. Khan, Hisham Cholakkal, Rao Muhammad Anwer, Jorma Laaksonen, Mubarak Shah, and Fahad Shahbaz Khan. Person image synthesis via denoising diffusion model. In CVPR 2023, pages 5968-5976. IEEE, 2023. 2", + "[2] Zhe Cao, Gines Hidalgo, Tomas Simon, Shih-En Wei, and Yaser Sheikh. Openpose: Realtime multi-person 2d pose estimation using part affinity fields. IEEE Trans. Pattern Anal. Mach. Intell., 43(1):172-186, 2021. 3", + "[3] Kennard Yanting Chan, Fayao Liu, Guosheng Lin, Chuan Sheng Foo, and Weisi Lin. R-cyclic diffuser: Reductive and cyclic latent diffusion for 3d clothed human digitalization. In CVPR 2024, pages 10304-10313. IEEE, 2024. 2", + "[4] Di Chang, Yichun Shi, Quankai Gao, Hongyi Xu, Jessica Fu, Guoxian Song, Qing Yan, Yizhe Zhu, Xiao Yang, and Mohammad Soleymani. Magicpose: Realistic human poses and facial expressions retargeting with identity-aware diffusion. In ICML 2024. OpenReview.net, 2024. 2", + "[5] Mingjin Chen, Junhao Chen, Xiaojun Ye, Huan-ang Gao, Xiaoxue Chen, Zhaoxin Fan, and Hao Zhao. Ultraman: Single image 3d human reconstruction with ultra speed and detail. CoRR, abs/2403.12028, 2024. 2", + "[6] Michael Garland and Paul S. Heckbert. Surface simplification using quadric error metrics. In SIGGRAPH 1997, pages 209-216. ACM, 1997. 6", + "[7] Hsuan-I Ho, Jie Song, and Otmar Hilliges. Sith: Single-view textured human reconstruction with image-conditioned diffusion. In CVPR 2024, pages 538-549. IEEE, 2024. 1, 2, 3, 5, 6", + "[8] Li Hu. Animate anyone: Consistent and controllable imagery-to-video synthesis for character animation. In CVPR 2024, pages 8153-8163. IEEE, 2024. 2, 3, 4, 6", + "[9] Shoukang Hu, Fangzhou Hong, Liang Pan, Haiyi Mei, Lei Yang, and Ziwei Liu. SHERF: generalizable human nef from a single image. In ICCV 2023, pages 9318-9330. IEEE, 2023. 1, 6", + "[10] Yangyi Huang, Hongwei Yi, Yuliang Xiu, Tingting Liao, Ji-axiang Tang, Deng Cai, and Justus Thies. Tech: Text-guided reconstruction of lifelike clothed humans. In 3DV 2024, pages 1531–1542. IEEE, 2024. 1, 2, 3, 5, 6", + "[11] Zeng Huang, Yuanlu Xu, Christoph Lassner, Hao Li, and Tony Tung. ARCH: animatable reconstruction of clothed humans. In CVPR 2020, pages 3090–3099. Computer Vision Foundation / IEEE, 2020. 2", + "[12] Yasamin Jafarian and Hyun Soo Park. Learning high fidelity depths of dressed humans by watching social media dance" + ], + "bbox": [ + 78, + 243, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "videos. In CVPR 2021, pages 12753-12762. Computer Vision Foundation / IEEE, 2021. 4", + "[13] Johanna Karras, Aleksander Holynski, Ting-Chun Wang, and Ira Kemelmacher-Shlizerman. Dreampose: Fashion image-to-video synthesis via stable diffusion. In ICCV 2023, pages 22623-22633. IEEE, 2023. 2", + "[14] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4):139:1-139:14, 2023. 1, 8", + "[15] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR 2015, 2015. 6", + "[16] Muhammed Kocabas, Jen-Hao Rick Chang, James Gabriel, Oncel Tuzel, and Anurag Ranjan. HUGS: human gaussian splats. In CVPR 2024, pages 505-515. IEEE, 2024. 4, 7, 8", + "[17] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3d object. In ICCV 2023, pages 9264–9275. IEEE, 2023. 2", + "[18] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J. Black. SMPL: a skinned multiperson linear model. ACM Trans. Graph., 34(6):248:1-248:16, 2015. 2, 4", + "[19] Roey Mechrez, Itamar Talmi, and Lihi Zelnik-Manor. The contextual loss for image transformation with non-aligned data. In ECCV 2018, pages 800-815. Springer, 2018. 6", + "[20] Gyeongsik Moon, Takaaki Shiratori, and Shunsuke Saito. Expressive whole-body 3d gaussian avatar. CoRR, abs/2407.21686, 2024. 5", + "[21] Ryota Natsume, Shunsuke Saito, Zeng Huang, Weikai Chen, Chongyang Ma, Hao Li, and Shigeo Morishima. Siclope: Silhouette-based clothed people. In CVPR 2019, pages 4480-4490. Computer Vision Foundation / IEEE, 2019. 2", + "[22] Haomiao Ni, Changhao Shi, Kai Li, Sharon X. Huang, and Martin Renqiang Min. Conditional image-to-video generation with latent flow diffusion models. In CVPR 2023, pages 18444-18455. IEEE, 2023. 2", + "[23] Panwang Pan, Zhuo Su, Chenguo Lin, Zhen Fan, Yongjie Zhang, Zeming Li, Tingting Shen, Yadong Mu, and Yebin Liu. Humansplat: Generalizable single-image human gaussian splatting with structure priors. CoRR, abs/2406.12459, 2024. 1, 2", + "[24] Marco Pesavento, Yuanlu Xu, Nikolaos Sarafianos, Robert Maier, Ziyan Wang, Chun-Han Yao, Marco Volino, Edmond Boyer, Adrian Hilton, and Tony Tung. ANIM: accurate neural implicit model for human reconstruction from a single RGB-D image. In CVPR 2024, pages 5448-5458. IEEE, 2024. 1", + "[25] Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. In ICLR 2023. OpenReview.net, 2023. 2", + "[26] Lorenza Prospero, Abdullah Hamdi, João F. Henriques, and Christian Rupprecht. GST: precise 3d human body from a single image with gaussian splatting transformers. CoRR, abs/2409.04196, 2024. 1, 2, 6", + "[27] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry," + ], + "bbox": [ + 503, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "5579", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML 2021, pages 8748-8763. PMLR, 2021. 3, 6", + "[28] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, Eric Mintun, Junting Pan, Kalyan Vasudev Alwala, Nicolas Carion, Chao-Yuan Wu, Ross B. Girshick, Piotr Dollar, and Christoph Feichtenhofer. SAM 2: Segment anything in images and videos. CoRR, abs/2408.00714, 2024. 4", + "[29] Shunsuke Saito, Zeng Huang, Ryota Natsume, Shigeo Morishima, Hao Li, and Angjoo Kanazawa. Pifu: Pixel-aligned implicit function for high-resolution clothed human digitization. In ICCV 2019, pages 2304-2314. IEEE, 2019. 2", + "[30] Shunsuke Saito, Tomas Simon, Jason M. Saragih, and Hanbyul Joo. Pifuhd: Multi-level pixel-aligned implicit function for high-resolution 3d human digitization. In CVPR 2020, pages 81–90. Computer Vision Foundation / IEEE, 2020. 2", + "[31] Anastasis Stathopoulos, Ligong Han, and Dimitris N. Metaxas. Score-guided diffusion for 3d human recovery. In CVPR 2024, pages 906-915. IEEE, 2024. 4", + "[32] David Svitov, Pietro Morerio, Lourdes Agapito, and Alessio Del Bue. HAHA: highly articulated gaussian human avatars with textured mesh prior. CoRR, abs/2404.01053, 2024. 4, 7, 8", + "[33] Linrui Tian, Qi Wang, Bang Zhang, and Liefeng Bo. EMO: emote portrait alive - generating expressive portrait videos with audio2video diffusion model under weak conditions. CoRR, abs/2402.17485, 2024. 2", + "[34] Thomas Unterthiner, Sjoerd van Steenkiste, Karol Kurach, Raphaël Marinier, Marcin Michalski, and Sylvain Gelly. Towards accurate generative models of video: A new metric & challenges. CoRR, abs/1812.01717, 2018. 6", + "[35] Tan Wang, Linjie Li, Kevin Lin, Chung-Ching Lin, Zhengyuan Yang, Hanwang Zhang, Zicheng Liu, and Lijuan Wang. Disco: Disentangled control for referring human dance generation in real world. CoRR, abs/2307.00040, 2023. 2, 6", + "[36] Xiang Wang, Shiwei Zhang, Changxin Gao, Jiayu Wang, Xiaogiang Zhou, Yingya Zhang, Luxin Yan, and Nong Sang. Unianimate: Taming unified video diffusion models for consistent human image animation. CoRR, abs/2406.01188, 2024. 3, 4, 6", + "[37] Xiang Wang, Shiwei Zhang, Hangjie Yuan, Zhiwu Qing, Biao Gong, Yingya Zhang, Yujun Shen, Changxin Gao, and Nong Sang. A recipe for scaling up text-to-video generation with text-free videos. In CVPR 2024, pages 6572-6582. IEEE, 2024. 4", + "[38] Yaohui Wang, Xin Ma, Xinyuan Chen, Antitza Dantcheva, Bo Dai, and Yu Qiao. LEO: generative latent image animatior for human video synthesis. CoRR, abs/2305.03989, 2023. 2", + "[39] Zhangyang Xiong, Chenghong Li, Kenkun Liu, Hongjie Liao, Jianqiao Hu, Junyi Zhu, Shuliang Ning, Lingteng Qiu, Chongjie Wang, Shijie Wang, Shuguang Cui, and Xiaoguang Han. Mvhumannet: A large-scale dataset of multi-view daily dressing human captures. In CVPR 2024, pages 19801-19811. IEEE, 2024. 4" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[40] Yuliang Xiu, Jinlong Yang, Dimitrios Tzionas, and Michael J. Black. ICON: Implicit Clothed humans Obtained from Normals. In CVPR 2022, pages 13296-13306, 2022. 1, 2", + "[41] Yuliang Xiu, Jinlong Yang, Xu Cao, Dimitrios Tzionas, and Michael J. Black. ECON: Explicit Clothed humans Optimized via Normal integration. In CVPR 2023, 2023. 1, 2", + "[42] Zhongcong Xu, Jianfeng Zhang, Jun Hao Liew, Hanshu Yan, Jia-Wei Liu, Chenxu Zhang, Jiashi Feng, and Mike Zheng Shou. Magicanimate: Temporally consistent human image animation using diffusion model. In CVPR 2024, pages 1481–1490. IEEE, 2024. 2", + "[43] Zhendong Yang, Ailing Zeng, Chun Yuan, and Yu Li. Effective whole-body pose estimation with two-stages distillation. In ICCV 2023 - Workshops, pages 4212-4222. IEEE, 2023. 3", + "[44] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. CoRR, abs/2308.06721, 2023. 2", + "[45] Polina Zablotskaia, Aliaksandr Siarohin, Bo Zhao, and Leonid Sigal. Dwnet: Dense warp-based network for pose-guided human video generation. In BMVC 2019, page 51. BMVA Press, 2019. 4, 6", + "[46] Xiaoyu Zhan, Jianxin Yang, Yuanqi Li, Jie Guo, Yanwen Guo, and Wenping Wang. Semantic human mesh reconstruction with textures. In CVPR 2024, pages 142-152. IEEE, 2024. 1, 6", + "[47] Jingbo Zhang, Xiaoyu Li, Qi Zhang, Yanpei Cao, Ying Shan, and Jing Liao. Humanref: Single image to 3d human generation via reference-guided diffusion. In CVPR 2024, pages 1844-1854. IEEE, 2024. 2, 6", + "[48] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV 2023, pages 3813-3824. IEEE, 2023. 2", + "[49] Zechuan Zhang, Li Sun, Zongxin Yang, Ling Chen, and Yi Yang. Global-correlated 3d-decoupling transformer for clothed avatar reconstruction. In NeurIPS 2023, 2023. 1", + "[50] Zechuan Zhang, Zongxin Yang, and Yi Yang. SIFU: side-view conditioned implicit function for real-world usable clothed human reconstruction. In CVPR 2024, pages 9936-9947. IEEE, 2024. 1, 2, 3, 6", + "[51] Shenhao Zhu, Junming Leo Chen, Zuozhuo Dai, Yinghui Xu, Xun Cao, Yao Yao, Hao Zhu, and Siyu Zhu. Champ: Controllable and consistent human image animation with 3d parametric guidance. CoRR, abs/2403.14781, 2024. 4, 6", + "[52] Zi-Xin Zou, Zhipeng Yu, Yuan-Chen Guo, Yangguang Li, Ding Liang, Yan-Pei Cao, and Song-Hai Zhang. Triplane meets gaussian splatting: Fast and generalizable single-view 3d reconstruction with transformers. In CVPR 2024, pages 10324-10335. IEEE, 2024. 4" + ], + "bbox": [ + 501, + 92, + 890, + 797 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "5580", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/SinGS_ Animatable Single-Image Human Gaussian Splats with Kinematic Priors/a1907643-3cec-4712-b625-aae92326d106_model.json b/2025/SinGS_ Animatable Single-Image Human Gaussian Splats with Kinematic Priors/a1907643-3cec-4712-b625-aae92326d106_model.json new file mode 100644 index 0000000000000000000000000000000000000000..8e285f4fc11fcd8ac677cc08d45598deb728c7d1 --- /dev/null +++ b/2025/SinGS_ Animatable Single-Image Human Gaussian Splats with Kinematic Priors/a1907643-3cec-4712-b625-aae92326d106_model.json @@ -0,0 +1,1793 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.131, + 0.892, + 0.152 + ], + "angle": 0, + "content": "SinGS: Animatable Single-Image Human Gaussian Splats with Kinematic Priors" + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.18, + 0.857, + 0.216 + ], + "angle": 0, + "content": "Yufan Wu\\(^{1*}\\) Xuanhong Chen\\(^{1\\dagger}\\) Wen Li\\(^{1,3}\\) Shunran Jia\\(^{2}\\) Hualiang Wei\\(^{1*}\\) Kairui Feng\\(^{4}\\) \nJialiang Chen\\(^{1}\\) Yuhan Li\\(^{1}\\) Ang He\\(^{1*}\\) Weimin Zhang\\(^{1}\\) Bingbing Ni\\(^{1}\\) Wenjun Zhang\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.108, + 0.217, + 0.862, + 0.252 + ], + "angle": 0, + "content": "\\(^{1}\\)Shanghai Jiao Tong University, Shanghai, China \\(^{2}\\)DreamX Inc \\(^{3}\\)Akool Research \\(^{4}\\)The National Key Laboratory of Autonomous Intelligent Unmanned Systems, Shanghai, China" + }, + { + "type": "text", + "bbox": [ + 0.12, + 0.255, + 0.835, + 0.27 + ], + "angle": 0, + "content": "{chen19910528,wmzhang,nibingbing,zhangwenjun}@sjtu.edu.cn, elvinfkr@tongji.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.305, + 0.313, + 0.32 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.338, + 0.474, + 0.655 + ], + "angle": 0, + "content": "Despite significant advances in accurately estimating geometry in contemporary single-image 3D human reconstruction, creating a high-quality, efficient, and animatable 3D avatar remains an open challenge. Two key obstacles persist: incomplete observation and inconsistent 3D priors. To address these challenges, we propose SinGS, aiming to achieve high-quality and efficient animatable 3D avatar reconstruction. At the heart of SinGS are two key components: Kinematic Human Diffusion and Geometry-Preserving 3D Gaussain Splatting. The former is a foundational human model that samples within pose space to generate a highly 3D-consistent and high-quality sequence of human images, inferring unseen viewpoints and providing kinematic priors. The latter is a system that reconstructs a compact, high-quality 3D avatar even under imperfect priors, achieved through a novel semantic Laplacian regularization and a geometry-preserving density control strategy that enable precise and compact assembly of 3D primitives. Extensive experiments demonstrate that SinGS enables life-like, animatable human reconstructions, maintaining both high quality and inference efficiency (up to 70FPS)." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.686, + 0.208, + 0.701 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.712, + 0.47, + 0.863 + ], + "angle": 0, + "content": "Single-image 3D human reconstruction (i.e., SIHR) is an emerging modeling technique that aims to create a lifelike 3D avatar from just one image, offering a highly cost-effective and promising approach to 3D modeling. Recently, 3D Gaussian Splatting (3DGS)[14] has made significant breakthroughs in rendering efficiency and quality, establishing itself as a mainstream approach for 3D avatar reconstruction. However, current 3DGS-based methods [23, 26] for single-image human 3D avatar reconstruction suffer from critical limitations in animation, as most are" + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.306, + 0.889, + 0.611 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.625, + 0.892, + 0.709 + ], + "angle": 0, + "content": "Figure 1. Existing single-image 3D avatar reconstruction methods are limited to sampling prior information from camera space. In contrast, SinGS proposes sampling from pose space, allowing us to obtain both camera and kinematic priors simultaneously. Combined with our Geometry-Preserving 3D Gaussian Splitting, this enables highly efficient animation rendering." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.719, + 0.892, + 0.764 + ], + "angle": 0, + "content": "designed primarily for static assets. This paper aims to develop an animation reconstruction framework that achieves a high balance between efficiency and rendering quality." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.893, + 0.902 + ], + "angle": 0, + "content": "Currently, the vast majority of SIHR methods [7, 10, 23, 24, 26, 40, 41, 46, 49, 50] focus solely on optimizing geometry while neglecting to enhance animation capabilities. As a result, they are often incapable of animation or require third-party tools [7] to achieve approximate motion. Unlike digital assets, the primary requirement for a human avatar is its ability to animate, making animation the core goal of SIHR [9]. Achieving animatable single-image human reconstruction (i.e., Animatable SIHR) is particularly" + }, + { + "type": "page_footnote", + "bbox": [ + 0.095, + 0.875, + 0.45, + 0.888 + ], + "angle": 0, + "content": "*Work done during an internship at Shanghai Jiao Tong University." + }, + { + "type": "page_footnote", + "bbox": [ + 0.098, + 0.888, + 0.316, + 0.9 + ], + "angle": 0, + "content": "† Corresponding author: Xuanhong Chen." + }, + { + "type": "list", + "bbox": [ + 0.095, + 0.875, + 0.45, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.514, + 0.957 + ], + "angle": 0, + "content": "5571" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.485 + ], + "angle": 0, + "content": "challenging for two main reasons: Incomplete Observation: A single image lacks comprehensive viewpoint information, offering an extremely limited, partial observation that cannot address occlusion issues. Current methods [7, 10, 47] mainly rely on fine-tuned diffusion models to synthesize (i.e., 3D hallucination) missing viewpoint information (such as front and back perspectives). However, this approach does not resolve the absence of motion information. While synthesized views can alleviate some occlusion issues, the human body's non-rigid structure requires more than static viewpoints to capture its complex kinematic patterns effectively. Inconsistent 3D Prior: Current mainstream methods [5, 7, 10, 47] use Score Distillation Sampling (SDS) [25] to leverage diffusion models as supervisory modules for learning 3D priors. However, the high generative diversity of these models often leads to viewpoint inconsistencies, resulting in implausible imagery (e.g., poor-quality hidden views and inconsistencies with the input image). HumanRef [47] attempts to mitigate the instability of SDS through ReferenceNet [8], but observable inconsistencies remain widespread in the results of ReferenceNet-based methods [8, 33], indicating the difficulty in fully resolving this issue. Consequently, 3D representations should be robust enough to handle imperfect 3D priors, though this perspective has largely been overlooked in current single-image avatar frameworks." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.494, + 0.473, + 0.904 + ], + "angle": 0, + "content": "To address these challenges, this paper introduces a novel single-image 3D human avatar reconstruction framework, named SinGS. Unlike existing methods [23, 26], SinGS focuses on directly constructing an animatable 3DGS avatar. At the core of our SinGS framework are two key components: Kinematic Human Diffusion and a highly Geometry-Preserving 3D Gaussian Splatting. Specifically, Kinematic Human Diffusion (KHD) is designed to sample a highly 3D-consistent human sequence within the pose space, based solely on a single input image prompt. This approach enables high-quality and animatable 3D avatar reconstruction. To train KHD, we constructed a dataset with over ten thousand rotation videos, allowing KHD to learn a rich variety of human structures and texture distributions from different angles, while also gaining extensive knowledge of human kinematics. Unlike existing frameworks [5, 7, 10, 47] that rely on costly and scarce multi-view human data, our approach requires only more affordable and abundant video data, significantly reducing KHD's training costs and better aligning with scaling laws. Geometry-Preserving 3D Gaussian Splatting (i.e., GPGS) tackles artifact and hole issues under imperfect human priors (e.g., minor jitter and 3D inconsistencies) through two key designs: Semantic Laplacian Regularization addresses floating Gaussian spheres caused by 3D inconsistencies. It adjusts the constraint strength based on body parts, allowing more flexibility for areas like hair and tighter constraints" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.245 + ], + "angle": 0, + "content": "for facial details, reducing floating Gaussian spheres while maintaining rendering quality. Geometry-Preserving Density Control resolves over-deletion and incorrect deletion of spheres in the training process. By managing the Gaussian sphere density via mesh edges, it ensures the spheres remain around the mesh, preserving geometry and preventing large holes. Extensive experiments have validated that our SinGS can generate high-quality and animatable 3D avatars at ultra-fast inference speeds (up to 70 FPS) with efficient resource usage (requiring only 3.8GB of GPU memory)." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.255, + 0.642, + 0.272 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.28, + 0.869, + 0.295 + ], + "angle": 0, + "content": "2.1. Diffusion-based Human Generative Models" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.303, + 0.895, + 0.545 + ], + "angle": 0, + "content": "Recent advances in diffusion-based human generative models, especially those conditioned on pose, have improved model controllability. Frameworks like ControlNet [48], IP Adapter [44], and ReferenceNet [8], along with better human attribute detection (2D/3D pose estimation, SMPL parameters [18]), enable more precise and consistent results. In image synthesis, models like PIDM [1] and LFDM [22] enhance pose transfer and image adaptation, while LEO [38] uses flow maps for motion representation. For video, DreamPose [13] and DisCo [35] improve pose-to-image translation and control separation. Models likeAnimateAnyone [8], MagicAnimate [42], and Magic-Pose [4] refine pose conditioning with ReferenceNet. Despite these improvements, these models remain resource-heavy and struggle with stability when generating images of complex movements, such as head turns or rotations." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.553, + 0.822, + 0.57 + ], + "angle": 0, + "content": "2.2. Single-Image Human Reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.576, + 0.895, + 0.82 + ], + "angle": 0, + "content": "Monocular video-based human reconstruction has reduced the need for specialized equipment, but single-image reconstruction remains challenging due to incomplete data. Early methods like PIFu [29] and SiCloPe [21] used pixel-aligned features and image-to-image translation, with PIFuHD [30] enhancing detail via normal maps. Later methods, such as ARCH [11] and ICON [40], incorporated body priors for animatable avatars, and ECON [41] improved model completeness by separately reconstructing front and back surfaces. Recent approaches like SiTH [7] and HumanRef [47] use generative models to fill in missing data, while SIFU [50] and Diffusion-FOF approximate multi-view features. R-Cyclic Diffuser [3] addresses limitations in Zero-1-to-3 [17], even with human-specific datasets, advancing single-image human reconstruction through generative priors and viewpoint conditioning." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.831, + 0.637, + 0.848 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Given a target image \\( I_r \\), the objective of the SinGS architecture is to reconstruct a fully animatable 3D avatar \\( \\mathbf{A}_t \\) based on \\( I_r \\). To address the challenge of limited viewpoint and" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.958 + ], + "angle": 0, + "content": "5572" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.078, + 0.088, + 0.498, + 0.324 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.498, + 0.089, + 0.892, + 0.323 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.079, + 0.325, + 0.891, + 0.417 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.352, + 0.429, + 0.616, + 0.443 + ], + "angle": 0, + "content": "Figure 2. Overall framework of our method." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.455, + 0.474, + 0.698 + ], + "angle": 0, + "content": "kinematic information inherent in single-image reconstruction, \\( I_{t} \\) is first input into the Kinematic Human Diffusion (i.e., KHD) module to complete missing information, generating an augmented sequence of rendered images \\( S_{t} = \\{I_{p_{1}},\\dots ,I_{p_{i}},\\dots ,I_{p_{N}}\\} \\), where \\( N \\) represents the total number of poses generated, and \\( p_i \\) indicates each corresponding 3D human pose. KHD reframes the less-constrained, underdetermined 3D reconstruction problem as a generative task, thus producing a fully-constrained, fully-observed 3D reconstruction formulation. Finally, \\( S_{t} \\) is fed into our compact 3DGS module to reconstruct an animatable 3D human avatar. The compact 3DGS module is capable of reconstructing a watertight, high-quality human model even with imperfect 3D priors. The distilled avatar is high-quality, computationally efficient, and well-suited for promising applications in AR, VR, and similar fields." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.713, + 0.248, + 0.727 + ], + "angle": 0, + "content": "3.1. Kinematic Priors" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.738, + 0.33, + 0.751 + ], + "angle": 0, + "content": "3.1.1 Kinematic Human Diffusion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Unlike existing hallucination architectures [7, 10, 50], our Kinematic Human Diffusion does not rely on camera input, which is prone to estimation errors and has the added challenges of multi-view data scarcity and high costs (e.g., multi-view human data typically requires capture with specialized setups like a light stage). Our approach instead posits that sampling within the motion space provides complete priors for viewpoint and kinematic features. Additionally, our KHD does not adopt a ReferenceNet structure [8];" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.455, + 0.892, + 0.636 + ], + "angle": 0, + "content": "while ReferenceNet has been highly successful, it requires maintaining two UNet models, significantly increasing the framework's resource demands and preventing integration of longer contexts (e.g., 24 frames). A shorter contextual window cannot provide an image sequence with a sufficient diversity of motion samples (e.g., 150 frames). Considering the above two aspects, our KHD is designed as a diffusion structure, as illustrated in Figure 2. KHD \\(\\mathcal{G}\\) takes a single image \\(I_{r}\\) as input and, constrained by a randomly sampled sequence of poses \\(\\{p_1,p_2,\\dots ,p_N\\}\\), augments the input image into a sequence \\(S_{t} = \\{I_{p_{1}},I_{p_{2}},\\ldots ,I_{p_{N}}\\}\\). This process can be formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.608, + 0.648, + 0.892, + 0.663 + ], + "angle": 0, + "content": "\\[\nS _ {t} = \\mathcal {G} \\left(I _ {r}; p _ {1}, p _ {2}, \\dots , p _ {N}\\right). \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.674, + 0.893, + 0.902 + ], + "angle": 0, + "content": "Model Architecture. Inspired by UniAnimate [36], our KHD also removes the ReferenceNet and directly inputs the reference image \\( I_r \\) and the driving sequence into the UNet. The reference image \\( I_r \\) is encoded by the VAE encoder of the diffusion model into a latent code \\( z_r \\), serving as an appearance reference throughout KHD. Additionally, \\( I_r \\) is encoded by the CLIP image encoder [27] into features \\( f_r \\) which are injected into the network through cross-attention by replacing the original text encoder's features in the diffusion model. This enables high-fidelity preservation of the subject's appearance characteristics. Existing works typically use 2D poses (e.g., OpenPose [2], DW-Pose [43]) as the driving signal. However, our experiments reveal that 2D poses fail to guide the diffusion model accurately in scenarios with extensive occlusions, such as when the subject" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5573" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.473, + 0.486 + ], + "angle": 0, + "content": "is turning. Moreover, 2D poses cannot reliably differentiate between different viewpoints of the human body. Therefore, we use SMPL refined by ScoreHMR [31] as the driving signal \\(\\{p_1,p_2,\\dots,p_N\\}\\). It is worth noting that Champ [51] also uses SMPL as guidance; however, their SMPL estimations lack accuracy, and they rely on overly strict guidance inputs, such as depth and normal vectors. This approach leads to suboptimal results, performing even worse than 2D poses in practice, for instance, by generating inconsistent body shapes. The SMPL parameters are converted into a mesh model in the specified pose using Linear Blend Skinning (LBS) [18]. This mesh is rendered as an RGB image based on the fixed camera pose maintained throughout the process. The rendered image is then encoded by a 3D Pose Encoder, a network composed of several convolutional layers, into a latent code \\(z_{p_i}\\) with the same dimensionality as \\(z_{r}\\). All \\(z_{p_i}\\) codes are first concatenated with the input noise along the channel dimension, and then sequentially along the temporal dimension to obtain the final representation \\(z_{P} \\in \\mathbb{R}^{T \\times C \\times H \\times W}\\). Simultaneously, \\(I_{r}\\) is processed to estimate and encode its pose latent code \\(z_{p_r}\\), which is then concatenated with \\(z_{P}\\) along the temporal dimension before being fed into the diffusion model's backbone. Following most video generation models [36, 37], the UNet component of the diffusion model is also designed as a 3D UNet to ensure 3D consistency in the generated output." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.494, + 0.473, + 0.902 + ], + "angle": 0, + "content": "Training Dataset and Strategy. Human image animation [8, 36] primarily focuses on generating diverse frontal human movements (e.g., dancing). As a result, their training data (e.g., TikTok [12]) is largely composed of performance videos of dancers, typically featuring frontal views and wide movement ranges. In our experiments, we found that models trained on this type of data provide limited support for our 3D reconstruction tasks. These models often fail to generate results with high consistency, exhibit significant flickering in fine details, and display abnormal motion in clothing dynamics. In response to this, we collected over 6,000 high-definition videos (each around 20 seconds long) from YouTube, featuring activities such as model showcases and dancing, with a focus on segments that include turning motions. Additionally, we incorporated 600 clips from the UBC Fashion dataset [45], primarily showcasing model runway turns. As a key part of our dataset, we also integrated frontal-view data from MVHumanNet [39]. We used SAM2 [28] to segment human figures from the background in MVHumanNet, then merged the results with a white background to create videos with a consistent background, effectively eliminating the background variability caused by bounding boxes in MVHumanNet. These three sources together form a dataset of over 10,000 training videos, which we call RotationHuman. Notably, compared to multi-view data, this type of data collection is significantly more cost-effective and thus better" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.228 + ], + "angle": 0, + "content": "suited for building large-scale datasets. During training, our KHD generates 64 pose-driven images in a single pass, setting the context window to 64 frames (i.e., \\( N = 64 \\)). In the inference stage, we directly generate 128 frames in one pass. The advantage of one-pass inference is that it avoids consistency jitter in overlapping transition frames. KHD training requires only one phase, where the objective is to reconstruct the ground truth images corresponding to each pose, with the training loss formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.6, + 0.242, + 0.892, + 0.26 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathbb {E} _ {\\theta} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} \\left(z _ {t}, f _ {r}, t\\right) \\| ^ {2} \\right], \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.264, + 0.895, + 0.445 + ], + "angle": 0, + "content": "where, \\( t \\) denotes the denoising step, and \\( z_{t} \\) represents the noise at step \\( t \\). Note that \\( z_{t} \\) in KHD contains diffusion states for \\( N \\) different driving poses. The feature \\( f_{r} \\) is the CLIP image encoding of the reference image \\( I_{r} \\), used to ensure ID consistency in the generated results. \\( \\theta \\) represents the parameters of the 3D UNet and the 3D Pose Encoder. The overall training process involves learning to reconstruct the target video based on \\( I_{r} \\) as input and the pose set \\( \\{p_1, p_2, \\dots, p_N\\} \\) as conditioning. The generated sequence \\( S_{t} \\) represents the sampling results of the subject in the input image \\( I_{r} \\) within the pose space, and it can be directly used for 3D avatar reconstruction." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.455, + 0.892, + 0.487 + ], + "angle": 0, + "content": "3.2. 3D Distillation with Geometry-Preserving 3D Gaussian Splitting" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.494, + 0.895, + 0.766 + ], + "angle": 0, + "content": "Once we obtain the sequence of pose images sampled from KHD, the goal is to distill a animatable and real-time rendered human avatar using Geometry-Preserving 3D Gaussian Splatting (GPGS). This challenge can be addressed based on mainstream approaches [16, 32, 52] in monocular 3D reconstruction. As illustrated in the Figure 2 on the right, we initialize the positions of Gaussians using the human template mesh that has undergone loop subdivision. The initial positions are set to the vertex locations, with a default of 2 upsampling iterations. This typically results in approximately \\(110.6K\\) Gaussians. To model the avatar representation, all vertices are fed into a learnable triplane [52] to extract vertex-wise features. These features are subsequently passed to MLPs to predict the geometry and appearance attributes of the Gaussians, respectively. The entire modeling process occurs in canonical space, and we finally map this canonical space into posed space using Linear Blend Skinning (LBS) based on the SMPL framework." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.786, + 0.803, + 0.801 + ], + "angle": 0, + "content": "3.2.1 Semantic Laplacian Regularization" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.894, + 0.901 + ], + "angle": 0, + "content": "Even during the reconstruction of real video, it is challenging to recover perfect static 3D consistency. Therefore, while the generated videos may appear visually plausible, every tiny ambiguity during lifting the 2D outcomes into 3D space potentially leads to catastrophic issues. Early experimental results indicated that common reconstruction meth" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5574" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.08, + 0.093, + 0.891, + 0.525 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.538, + 0.893, + 0.582 + ], + "angle": 0, + "content": "Figure 3. Visual comparison of our method with TeCH [10] and SiTH [7] for single-image reconstruction. We reposed the reconstructed results to A Pose, showcasing both animation and reconstruction capabilities. We highly recommend checking the suppl. for more results. We have provided a large number of animation result images and videos in the suppl." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.593, + 0.473, + 0.851 + ], + "angle": 0, + "content": "ods produced results that lacked smoothness and compactness, resulting in existing disconnected and floating gaus-. sians. Following [20], we found that applying Laplacian regularization on the canonical mesh, both for geometry and appearance, effectively mitigated the presence of gaus-. sians out of control. However, it also constrained offsets. In other words, isometrical Laplacian constraint on the whole body overly relied on the accuracy of the SMPL mesh, also making it difficult to fit areas such as hair and clothing that deviate significantly from the SMPL surface. Considering the complexity of the human body and the impossibility to obtain extremely accurate SMPL estimation, we repartition all gaussian vertices into \\(L = 15\\) distinct semantic parts and dynamically adjust the smoothing strength. Besides, we impose the constraint on the anchors points which are closer to the rendering surface than gausicians center. The Semantic Laplacian Regularization (SLR) is defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.102, + 0.86, + 0.47, + 0.906 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {S L R} = \\sum_ {u \\in \\Omega} \\lambda_ {u} \\left\\| a _ {u, v} - \\frac {1}{| M (u , v) |} \\sum_ {v \\in M (u, v)} a _ {u, v} \\right\\|, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.593, + 0.895, + 0.79 + ], + "angle": 0, + "content": "where \\(\\Omega\\) represents the set of semantic regions, \\(a\\) denotes the surface anchor point of the \\(v\\)-th Gaussian vertex in the \\(u\\)-th semantic region, and \\(M(u,v)\\) refers to the set of adjacent points of the Gaussian vertex \\(a_{u,v}\\) located within the \\(u\\)-th semantic region. Experiments demonstrate that the Semantic Laplacian Regularization significantly impacts the handling of more intricate situations, particularly when dealing with errors between the SMPL mesh and the actual character, which can arise from both inaccurate estimations and complex clothing. We apply Laplacians in both position and color spaces; the former enhances the smoothness of the rendering surface, while the latter ensures a more continuous variation in color space." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.815, + 0.882, + 0.831 + ], + "angle": 0, + "content": "3.2.2 Geometry-Preserving Density Control Strategy" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.894, + 0.903 + ], + "angle": 0, + "content": "The strategies in vanilla 3DGS were not originally build for complex, dynamic structures like the human body. While works [20] have adapted the approach, such as initializing Gaussians based on SMPL meshes, these efforts pri" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.958 + ], + "angle": 0, + "content": "5575" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.089, + 0.895, + 0.133 + ], + "angle": 0, + "content": "Table 1. Quantitative comparisons with existing methods on UBC fashion dataset [45]. Note that metrics in the table are measure with the background-removed video. Some methods are unable to perform animation, and therefore, these metrics are not available for those methods. Bold indicates the best results, while Underlined represents the second-best results." + }, + { + "type": "table", + "bbox": [ + 0.101, + 0.142, + 0.868, + 0.356 + ], + "angle": 0, + "content": "
MethodPrior TypePSNR ↑LPIPS ↓FVD ↓Context. ↓CLIP Score ↑Speed↑
DisCo [35] (CVPR24)-30.980.259167.72.13482.4%1FPS
Animate Anyone [8] (CVPR24)-31.330.238148.81.90186.6%2FPS
Champ [51] (ECCV24)-31.210.251155.12.03884.3%1FPS
UniAnimate [36] (ArXiv24)-32.310.211141.31.81187.1%2FPS
SHERF [9] (ICCV23)-27.390.301202.32.99870.1%
SIFU [50] (CVPR24)Camera---2.89172.1%5FPS
GST [26] (ArXiv24)Camera---3.01171.7%50FPS
TeCH [10] (3DV24)Camera29.010.271188.12.33179.3%6FPS
SiTH [7] (CVPR24)Camera---2.77674.7%<1FPS
HumanRef [47] (CVPR24)Camera---2.16781.5%3FPS
SHERT [46] (CVPR2024)Inpainting29.310.266179.82.31978.8%10FPS
UniAnimate + GPGS (ours)Kinematics29.970.251168.71.99282.1%70FPS
KHD (ours)-32.810.221134.41.76892.3%2FPS
SinGS (ours)Kinematics31.330.240156.31.84788.2%70FPS
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.366, + 0.473, + 0.775 + ], + "angle": 0, + "content": "marily lay a solid foundation for the optimization process but fail to maintain this advantages throughout. Incompatibilities still arise in optimization stage, prompting us to propose a novel density control scheme that is inherently compatible with hybrid representations, designed for adaptive adjustment body-wide density. This scheme draws on mesh operation algorithms [6] to better address the unique challenges of human body. For primitives that locate in under-expression mesh areas, we subdivide these regions by adding gaussians on every edge of the selected geometric triangles, which creates much smaller triangles. And we skip the trivial post-smoothing for vertices as the position of Gaussians will continue to be optimized in the network. For candidate Gaussians that are selected due to excessively large scales, as well as all newly added Gaussians, we perform rescaling to adjust them to an average size. In the case of anomalous gaussians, especially those are almost transparent, we collapse relevant edges and replace them with single vertex, resulting in decimated adjacent faces. And the priority of selected edges is decided by gaussian attributes and edge length. This entire pruning process preserves the principal structure of the human body. Our strategy result in better reconstruction and animation as the optimization was conducted under connectivity and compactness of the entire human body rather than producing holes and stacks. Moreover, this approach synergistically enhances the effectiveness of our hybrid representation." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.79, + 0.21, + 0.807 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.816, + 0.268, + 0.833 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.472, + 0.903 + ], + "angle": 0, + "content": "Implementation Details. Our KHD primarily utilizes the RotationHuman dataset we collected, which focuses on human pose image sequences involving turning motions. The specific structure of the 3D Pose Encoder is provided in" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.366, + 0.892, + 0.636 + ], + "angle": 0, + "content": "the supplementary materials. KHD was trained using 8 NVIDIA A100 80GB GPUs, with DeepSpeed and ZeRO2 employed to reduce memory usage. Similar toAnimate Anyone, we resized the video frames to a spatial dimension of \\(768 \\times 512\\). We used a context window size of 64 frames, and to enhance model scalability post-training, we implemented a random context window jitter strategy [36]. During training, the model randomly samples window sizes from 64 to 32 frames in an uniformly distributed manner. The training process utilized the AdamW optimizer [15], with a learning rate set to \\(5e - 5\\) and DDPM with 1000 steps for noise sampling. During inference, DDIM with 30 steps was used to reduce inference time. For the 3DGS component, the Adam optimizer [15] was used with a learning rate of \\(1e - 3\\) and a cosine annealing learning scheduler was applied to accelerate convergence. The input image sequence size for 3DGS was also \\(768 \\times 512\\). 3DGS was trained using a single NVIDIA 4090 24GB GPU." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.637, + 0.892, + 0.758 + ], + "angle": 0, + "content": "Baseline Models. Our baseline models consist of two parts: the first part includes human image animation models, such as DisCo [35], Animation Anyone [8] \\(^{1}\\), Champ [51], and UniAnimate [36]. These models are used to compare our animation capability and the superiority of KHD. The second part consists of single-image 3D avatar reconstruction models, including SHERF [9], SIFU [50], GST [26], TeCH [10], SiTH [7], HumanRef [47], and SHERT [46]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.758, + 0.892, + 0.879 + ], + "angle": 0, + "content": "Metrics. We employ both PSNR and LPIPS to assess the quality of video reconstruction. To evaluate the animation capability of our model, we also use FVD [34] to measure the distribution discrepancy between the generated video and the original video, assessing the video generation performance. Furthermore, following HumanRef, we introduce contextual distance [19] and CLIP score [27] to evaluate the texture and semantic similarity between the rendered" + }, + { + "type": "page_footnote", + "bbox": [ + 0.518, + 0.887, + 0.833, + 0.902 + ], + "angle": 0, + "content": "1https://github.com/MooreThreads/Moore-AnimateAnyone" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.958 + ], + "angle": 0, + "content": "5576" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.09, + 0.462, + 0.288 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.296, + 0.47, + 0.325 + ], + "angle": 0, + "content": "Figure 4. Comparison between KHD and other SOTA human image animation methods. Please zoom in for better observation." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.339, + 0.327, + 0.355 + ], + "angle": 0, + "content": "output and the input reference image." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.366, + 0.27, + 0.383 + ], + "angle": 0, + "content": "4.2. Qualitative Analysis" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.39, + 0.47, + 0.632 + ], + "angle": 0, + "content": "Singe-Image Reconstruction. We compare SinGS with two methods from the literature that claim to support animation (TeCH, SiTH). It is worth noting that during our experiments, we found that single-image animation with these methods was extremely difficult to execute, and in some cases, nearly impossible. We had to use third-party code \\( {}^{2} \\) to achieve the reposing for some of these methods. We animate their reconstructed outputs to a common pose (A Pose) for comparison, which allows us to evaluate both reconstruction and animation capabilities. The comparison results are shown in Figure 3, SinGS generates results with better details and maintains high-quality animation, particularly on the in-the-wild data, where SinGS's visual performance significantly outperforms the other methods. This demonstrates the high feasibility and effectiveness of the kinematic prior." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.633, + 0.471, + 0.875 + ], + "angle": 0, + "content": "Kinematic Human Diffusion. KHD, as the core component of our pipeline, is compared with state-of-the-art human image animation methods, includingAnimate Anyone and UniAnimate. As shown in Figure 4, we drive the same character with all three methods, focusing on comparing the quality of their generated results, particularly for the turning pose. From the figure, it is clear that both UniAnimate and our KHD maintain higher consistency thanAnimate Anyone after the character turns, with KHD surpassing UniAnimate in detail. The details marked by the red arrows in Figure 4 reveals that UniAnimate generates incorrect details, a common issue when using 2D poses, as they often struggle to distinguish body orientation during back-facing poses, leading the model to generate erroneous images. This is detrimental to reconstruction, while KHD generates videos with much higher consistency." + }, + { + "type": "image", + "bbox": [ + 0.519, + 0.088, + 0.875, + 0.183 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.192, + 0.895, + 0.263 + ], + "angle": 0, + "content": "Figure 5. Fifty users rated the video quality, with UniAnimate scoring highest, followed closely by our model, which outperformed others. Our model is more efficient, using only 3.5 GB of memory and achieving 70 FPS, while UniAnimate requires over 12GB and 0.5 seconds per image." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.276, + 0.895, + 0.533 + ], + "angle": 0, + "content": "Geometry-Preserving 3D Gaussian Splatting. Although KHD provides high-consistency pose sampling results, it inevitably has some consistency flaws, meaning it can only offer imperfect prior information. Our geometry-preserving GS method is specifically designed to address this challenge. To validate the superiority of our Geometry-Preserving 3D Gaussian Splatting (GPGS) over general human modeling methods, we compared it with recent state-of-the-art (i.e., SOTA) methods, HAHA [32] and HUGS [16]. We generate a pose sequence using KHD and then performed reconstruction with all 3 methods. The reconstruction results are shown in Figure 7. As seen, both HAHA and HUGS fail to produce a compact 3D avatar, with significant holes or floating artifacts appearing in various areas, particularly during the turning motion. In contrast, our GPGS method is able to reconstruct a compact, watertight 3D avatar model." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.545, + 0.704, + 0.561 + ], + "angle": 0, + "content": "4.3. Quantitative Analysis" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.569, + 0.894, + 0.903 + ], + "angle": 0, + "content": "We selected 100 test videos from the UBC dataset for our quantitative analysis. For each video, we chose the first frame as the reference image and used the video's pose sequence as the driving signal. The performance of all methods was assessed by comparing the animation results with the original video to evaluate the differences. Video reconstruction metrics (including PSNR, LPIPS, and FVD) reflect the model's animation capabilities, while Contextual Distance and CLIP score allow for comparison with methods that do not support animation. The quantitative comparison results are shown in Table 1. Human image animation methods (i.e., the second to fifth rows in the table, including our KHD) outperform the reconstruction-based methods across the board. Among them, KHD surpasses the other methods in most metrics, and UniAnimate also demonstrates superior performance, which is expected due to the scale of training data and model size. While SinGS lags behind UniAnimate, it shows a significant advantage over other methods in some metrics. However, SinGS significantly outperforms diffusion-based animation methods in terms of inference efficiency and resource consumption. Our SinGS achieves 70 FPS during inference, with only" + }, + { + "type": "page_footnote", + "bbox": [ + 0.095, + 0.887, + 0.378, + 0.901 + ], + "angle": 0, + "content": "\\(^{2}\\)https://github.com/custom-humans/editable-humans" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5577" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.09, + 0.47, + 0.132 + ], + "angle": 0, + "content": "Table 2. The results of the ablation experiments on Semantic Laplacian Regularization (i.e., Semantic Lap. Reg.) and Geometry-Preserving Density Control (i.e., GPDC) in SinGS." + }, + { + "type": "table", + "bbox": [ + 0.113, + 0.142, + 0.432, + 0.219 + ], + "angle": 0, + "content": "
SettingPSNRLPIPSFVD
with Semantic Lap. Reg.31.330.240156.3
with KNN Lap. Reg.30.880.250169.9
w/o Lap. Reg.30.340.272181.3
with GPDC31.010.249166.1
naive Density Control [14]30.220.276190.2
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.231, + 0.468, + 0.352 + ], + "angle": 0, + "content": "3.8 GB of memory usage (tested on a single NVIDIA 4090 GPU), while UniAnimate, Champ, andAnimate Anyone each require more than 0.4 seconds per image and consume over 12 GB of memory. These quantitative results strongly demonstrate that SinGS is highly effective and feasible. We believe that using KHD, a specialized foundational model, will greatly enhance 3D reconstruction performance while significantly reducing the stringent data requirements." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.352, + 0.468, + 0.413 + ], + "angle": 0, + "content": "User Study. We invited 50 users to subjectively evaluate the animation results of SinGS. The evaluation was based on a scale from 0 to 10, with 10 being the highest score and 0 being the lowest. The results are visualized in Figure 5." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.424, + 0.23, + 0.44 + ], + "angle": 0, + "content": "4.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.448, + 0.468, + 0.794 + ], + "angle": 0, + "content": "Kinematic Priors Analysis. The number of poses sampled by KHD has a significant impact on the 3D distillation process. We analyzed how the number of poses affects the animation capability of the final reconstructed model, and the results are shown in Figure 6. We incrementally increased the number of poses from 2 (which corresponds to the typical approach of generating front and rear views, as used by most methods) up to 150. We also tested the influence of different prior models (i.e., KHD, UniAnimate, and Animate Anyone) on the results. Additionally, we examined how different 3DGS reconstruction methods handle the use of prior knowledge. As shown, performance improves rapidly with an increase in the number of poses, but the rate of improvement slows significantly after reaching around 100 poses. Moore-AA + HAHA shows the slowest improvement, while KHD + GPGS exhibits the slowest decline in performance. This suggests that increasing the number of poses does not always lead to improved performance for two main reasons: first, the prior models have already provided sufficient information; and second, none of the prior models provide perfectly consistent priors. Excessive sampling can introduce a lot of inconsistent noise, which ultimately harms performance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.468, + 0.901 + ], + "angle": 0, + "content": "Geometry-Preserving 3D Gaussian Splitting Analysis. We conducted ablation experiments on the two key designs in GPGS: Semantic Laplacian Regularization and GPDC, with the results shown in Table 2. By replacing and removing the Laplacian regularization, we observed that Semantic Laplacian consistently outperforms in terms of reconstruction metrics. Additionally, GPDC demonstrates a perfor" + }, + { + "type": "image", + "bbox": [ + 0.541, + 0.089, + 0.852, + 0.227 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.238, + 0.892, + 0.266 + ], + "angle": 0, + "content": "Figure 6. The comparison of the impact of augmented pose numbers on the performance of the 3D avatar models." + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.293, + 0.887, + 0.46 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.473, + 0.892, + 0.516 + ], + "angle": 0, + "content": "Figure 7. Here is a comparison of our GPGS with two SOTA 3DGS avatar methods, HAHA [32] and HUGS [16]. All methods use images sampled by KHD as body prior." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.534, + 0.892, + 0.608 + ], + "angle": 0, + "content": "mance advantage over the original density control strategy. The original strategy does not account for dynamic properties and geometric constraints, which makes it ineffective in addressing issues like floating and holes caused by data inconsistencies." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.631, + 0.619, + 0.646 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.66, + 0.892, + 0.901 + ], + "angle": 0, + "content": "In this paper, we propose a new paradigm for single-image 3D avatar reconstruction. Unlike traditional methods that complete missing information in camera space, we sample in human kinematic space, enabling efficient capture of both viewpoints and motion traits. To achieve this, we introduce Kinematic Human Diffusion (KHD) for consistent pose-space generation, supported by our RotationHuman dataset, which focuses on turning motions. Additionally, we propose Geometry-Preserving 3D Gaussian Splating that reconstructs high-quality, animatable 3D avatars under imperfect human priors. This includes two key innovations: Semantic Laplacian regularization for compact, detailed constraints, and a geometry-preserving density control strategy to maintain stable human geometry. Together, these innovations allow SinGS to achieve high-quality, animatable 3D avatar reconstruction from a single image." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5578" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.078, + 0.091, + 0.248, + 0.108 + ], + "angle": 0, + "content": "6. Acknowledgment" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.115, + 0.47, + 0.205 + ], + "angle": 0, + "content": "This work supported by Shanghai Post-doctoral Excellence Programe (2024290). This work was also partly by Shanghai Municipal Bureau of Culture and Tourism and Meta-verse Technology Innovation Center of Shanghai Municipal Bureau of Culture and Tourism. Kairui Feng thanks to Xiaomi foundation." + }, + { + "type": "title", + "bbox": [ + 0.08, + 0.218, + 0.174, + 0.234 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.244, + 0.47, + 0.311 + ], + "angle": 0, + "content": "[1] Ankan Kumar Bhunia, Salman H. Khan, Hisham Cholakkal, Rao Muhammad Anwer, Jorma Laaksonen, Mubarak Shah, and Fahad Shahbaz Khan. Person image synthesis via denoising diffusion model. In CVPR 2023, pages 5968-5976. IEEE, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.314, + 0.472, + 0.368 + ], + "angle": 0, + "content": "[2] Zhe Cao, Gines Hidalgo, Tomas Simon, Shih-En Wei, and Yaser Sheikh. Openpose: Realtime multi-person 2d pose estimation using part affinity fields. IEEE Trans. Pattern Anal. Mach. Intell., 43(1):172-186, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.37, + 0.47, + 0.436 + ], + "angle": 0, + "content": "[3] Kennard Yanting Chan, Fayao Liu, Guosheng Lin, Chuan Sheng Foo, and Weisi Lin. R-cyclic diffuser: Reductive and cyclic latent diffusion for 3d clothed human digitalization. In CVPR 2024, pages 10304-10313. IEEE, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.439, + 0.469, + 0.508 + ], + "angle": 0, + "content": "[4] Di Chang, Yichun Shi, Quankai Gao, Hongyi Xu, Jessica Fu, Guoxian Song, Qing Yan, Yizhe Zhu, Xiao Yang, and Mohammad Soleymani. Magicpose: Realistic human poses and facial expressions retargeting with identity-aware diffusion. In ICML 2024. OpenReview.net, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.51, + 0.469, + 0.564 + ], + "angle": 0, + "content": "[5] Mingjin Chen, Junhao Chen, Xiaojun Ye, Huan-ang Gao, Xiaoxue Chen, Zhaoxin Fan, and Hao Zhao. Ultraman: Single image 3d human reconstruction with ultra speed and detail. CoRR, abs/2403.12028, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.565, + 0.469, + 0.606 + ], + "angle": 0, + "content": "[6] Michael Garland and Paul S. Heckbert. Surface simplification using quadric error metrics. In SIGGRAPH 1997, pages 209-216. ACM, 1997. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.608, + 0.469, + 0.661 + ], + "angle": 0, + "content": "[7] Hsuan-I Ho, Jie Song, and Otmar Hilliges. Sith: Single-view textured human reconstruction with image-conditioned diffusion. In CVPR 2024, pages 538-549. IEEE, 2024. 1, 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.664, + 0.469, + 0.704 + ], + "angle": 0, + "content": "[8] Li Hu. Animate anyone: Consistent and controllable imagery-to-video synthesis for character animation. In CVPR 2024, pages 8153-8163. IEEE, 2024. 2, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.705, + 0.469, + 0.759 + ], + "angle": 0, + "content": "[9] Shoukang Hu, Fangzhou Hong, Liang Pan, Haiyi Mei, Lei Yang, and Ziwei Liu. SHERF: generalizable human nef from a single image. In ICCV 2023, pages 9318-9330. IEEE, 2023. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.761, + 0.469, + 0.816 + ], + "angle": 0, + "content": "[10] Yangyi Huang, Hongwei Yi, Yuliang Xiu, Tingting Liao, Ji-axiang Tang, Deng Cai, and Justus Thies. Tech: Text-guided reconstruction of lifelike clothed humans. In 3DV 2024, pages 1531–1542. IEEE, 2024. 1, 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.818, + 0.469, + 0.871 + ], + "angle": 0, + "content": "[11] Zeng Huang, Yuanlu Xu, Christoph Lassner, Hao Li, and Tony Tung. ARCH: animatable reconstruction of clothed humans. In CVPR 2020, pages 3090–3099. Computer Vision Foundation / IEEE, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.469, + 0.901 + ], + "angle": 0, + "content": "[12] Yasamin Jafarian and Hyun Soo Park. Learning high fidelity depths of dressed humans by watching social media dance" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.244, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "videos. In CVPR 2021, pages 12753-12762. Computer Vision Foundation / IEEE, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.892, + 0.177 + ], + "angle": 0, + "content": "[13] Johanna Karras, Aleksander Holynski, Ting-Chun Wang, and Ira Kemelmacher-Shlizerman. Dreampose: Fashion image-to-video synthesis via stable diffusion. In ICCV 2023, pages 22623-22633. IEEE, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.179, + 0.892, + 0.233 + ], + "angle": 0, + "content": "[14] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4):139:1-139:14, 2023. 1, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.235, + 0.892, + 0.262 + ], + "angle": 0, + "content": "[15] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR 2015, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.264, + 0.892, + 0.305 + ], + "angle": 0, + "content": "[16] Muhammed Kocabas, Jen-Hao Rick Chang, James Gabriel, Oncel Tuzel, and Anurag Ranjan. HUGS: human gaussian splats. In CVPR 2024, pages 505-515. IEEE, 2024. 4, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.307, + 0.892, + 0.361 + ], + "angle": 0, + "content": "[17] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3d object. In ICCV 2023, pages 9264–9275. IEEE, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.364, + 0.892, + 0.417 + ], + "angle": 0, + "content": "[18] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J. Black. SMPL: a skinned multiperson linear model. ACM Trans. Graph., 34(6):248:1-248:16, 2015. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.42, + 0.892, + 0.462 + ], + "angle": 0, + "content": "[19] Roey Mechrez, Itamar Talmi, and Lihi Zelnik-Manor. The contextual loss for image transformation with non-aligned data. In ECCV 2018, pages 800-815. Springer, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.464, + 0.892, + 0.504 + ], + "angle": 0, + "content": "[20] Gyeongsik Moon, Takaaki Shiratori, and Shunsuke Saito. Expressive whole-body 3d gaussian avatar. CoRR, abs/2407.21686, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.506, + 0.892, + 0.56 + ], + "angle": 0, + "content": "[21] Ryota Natsume, Shunsuke Saito, Zeng Huang, Weikai Chen, Chongyang Ma, Hao Li, and Shigeo Morishima. Siclope: Silhouette-based clothed people. In CVPR 2019, pages 4480-4490. Computer Vision Foundation / IEEE, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.563, + 0.892, + 0.616 + ], + "angle": 0, + "content": "[22] Haomiao Ni, Changhao Shi, Kai Li, Sharon X. Huang, and Martin Renqiang Min. Conditional image-to-video generation with latent flow diffusion models. In CVPR 2023, pages 18444-18455. IEEE, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.619, + 0.892, + 0.687 + ], + "angle": 0, + "content": "[23] Panwang Pan, Zhuo Su, Chenguo Lin, Zhen Fan, Yongjie Zhang, Zeming Li, Tingting Shen, Yadong Mu, and Yebin Liu. Humansplat: Generalizable single-image human gaussian splatting with structure priors. CoRR, abs/2406.12459, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.689, + 0.892, + 0.771 + ], + "angle": 0, + "content": "[24] Marco Pesavento, Yuanlu Xu, Nikolaos Sarafianos, Robert Maier, Ziyan Wang, Chun-Han Yao, Marco Volino, Edmond Boyer, Adrian Hilton, and Tony Tung. ANIM: accurate neural implicit model for human reconstruction from a single RGB-D image. In CVPR 2024, pages 5448-5458. IEEE, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.774, + 0.892, + 0.815 + ], + "angle": 0, + "content": "[25] Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. In ICLR 2023. OpenReview.net, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.817, + 0.892, + 0.87 + ], + "angle": 0, + "content": "[26] Lorenza Prospero, Abdullah Hamdi, João F. Henriques, and Christian Rupprecht. GST: precise 3d human body from a single image with gaussian splatting transformers. CoRR, abs/2409.04196, 2024. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.873, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[27] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry," + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5579" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.469, + 0.147 + ], + "angle": 0, + "content": "Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML 2021, pages 8748-8763. PMLR, 2021. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.149, + 0.469, + 0.244 + ], + "angle": 0, + "content": "[28] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, Eric Mintun, Junting Pan, Kalyan Vasudev Alwala, Nicolas Carion, Chao-Yuan Wu, Ross B. Girshick, Piotr Dollar, and Christoph Feichtenhofer. SAM 2: Segment anything in images and videos. CoRR, abs/2408.00714, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.247, + 0.469, + 0.3 + ], + "angle": 0, + "content": "[29] Shunsuke Saito, Zeng Huang, Ryota Natsume, Shigeo Morishima, Hao Li, and Angjoo Kanazawa. Pifu: Pixel-aligned implicit function for high-resolution clothed human digitization. In ICCV 2019, pages 2304-2314. IEEE, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.302, + 0.469, + 0.357 + ], + "angle": 0, + "content": "[30] Shunsuke Saito, Tomas Simon, Jason M. Saragih, and Hanbyul Joo. Pifuhd: Multi-level pixel-aligned implicit function for high-resolution 3d human digitization. In CVPR 2020, pages 81–90. Computer Vision Foundation / IEEE, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.359, + 0.469, + 0.398 + ], + "angle": 0, + "content": "[31] Anastasis Stathopoulos, Ligong Han, and Dimitris N. Metaxas. Score-guided diffusion for 3d human recovery. In CVPR 2024, pages 906-915. IEEE, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.401, + 0.469, + 0.453 + ], + "angle": 0, + "content": "[32] David Svitov, Pietro Morerio, Lourdes Agapito, and Alessio Del Bue. HAHA: highly articulated gaussian human avatars with textured mesh prior. CoRR, abs/2404.01053, 2024. 4, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.456, + 0.469, + 0.509 + ], + "angle": 0, + "content": "[33] Linrui Tian, Qi Wang, Bang Zhang, and Liefeng Bo. EMO: emote portrait alive - generating expressive portrait videos with audio2video diffusion model under weak conditions. CoRR, abs/2402.17485, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.512, + 0.469, + 0.566 + ], + "angle": 0, + "content": "[34] Thomas Unterthiner, Sjoerd van Steenkiste, Karol Kurach, Raphaël Marinier, Marcin Michalski, and Sylvain Gelly. Towards accurate generative models of video: A new metric & challenges. CoRR, abs/1812.01717, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.568, + 0.469, + 0.635 + ], + "angle": 0, + "content": "[35] Tan Wang, Linjie Li, Kevin Lin, Chung-Ching Lin, Zhengyuan Yang, Hanwang Zhang, Zicheng Liu, and Lijuan Wang. Disco: Disentangled control for referring human dance generation in real world. CoRR, abs/2307.00040, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.638, + 0.469, + 0.704 + ], + "angle": 0, + "content": "[36] Xiang Wang, Shiwei Zhang, Changxin Gao, Jiayu Wang, Xiaogiang Zhou, Yingya Zhang, Luxin Yan, and Nong Sang. Unianimate: Taming unified video diffusion models for consistent human image animation. CoRR, abs/2406.01188, 2024. 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.707, + 0.469, + 0.773 + ], + "angle": 0, + "content": "[37] Xiang Wang, Shiwei Zhang, Hangjie Yuan, Zhiwu Qing, Biao Gong, Yingya Zhang, Yujun Shen, Changxin Gao, and Nong Sang. A recipe for scaling up text-to-video generation with text-free videos. In CVPR 2024, pages 6572-6582. IEEE, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.777, + 0.469, + 0.816 + ], + "angle": 0, + "content": "[38] Yaohui Wang, Xin Ma, Xinyuan Chen, Antitza Dantcheva, Bo Dai, and Yu Qiao. LEO: generative latent image animatior for human video synthesis. CoRR, abs/2305.03989, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.819, + 0.469, + 0.899 + ], + "angle": 0, + "content": "[39] Zhangyang Xiong, Chenghong Li, Kenkun Liu, Hongjie Liao, Jianqiao Hu, Junyi Zhu, Shuliang Ning, Lingteng Qiu, Chongjie Wang, Shijie Wang, Shuguang Cui, and Xiaoguang Han. Mvhumannet: A large-scale dataset of multi-view daily dressing human captures. In CVPR 2024, pages 19801-19811. IEEE, 2024. 4" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.469, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[40] Yuliang Xiu, Jinlong Yang, Dimitrios Tzionas, and Michael J. Black. ICON: Implicit Clothed humans Obtained from Normals. In CVPR 2022, pages 13296-13306, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.15, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[41] Yuliang Xiu, Jinlong Yang, Xu Cao, Dimitrios Tzionas, and Michael J. Black. ECON: Explicit Clothed humans Optimized via Normal integration. In CVPR 2023, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.193, + 0.892, + 0.26 + ], + "angle": 0, + "content": "[42] Zhongcong Xu, Jianfeng Zhang, Jun Hao Liew, Hanshu Yan, Jia-Wei Liu, Chenxu Zhang, Jiashi Feng, and Mike Zheng Shou. Magicanimate: Temporally consistent human image animation using diffusion model. In CVPR 2024, pages 1481–1490. IEEE, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.263, + 0.892, + 0.316 + ], + "angle": 0, + "content": "[43] Zhendong Yang, Ailing Zeng, Chun Yuan, and Yu Li. Effective whole-body pose estimation with two-stages distillation. In ICCV 2023 - Workshops, pages 4212-4222. IEEE, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.32, + 0.892, + 0.36 + ], + "angle": 0, + "content": "[44] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. CoRR, abs/2308.06721, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.363, + 0.892, + 0.416 + ], + "angle": 0, + "content": "[45] Polina Zablotskaia, Aliaksandr Siarohin, Bo Zhao, and Leonid Sigal. Dwnet: Dense warp-based network for pose-guided human video generation. In BMVC 2019, page 51. BMVA Press, 2019. 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.419, + 0.892, + 0.473 + ], + "angle": 0, + "content": "[46] Xiaoyu Zhan, Jianxin Yang, Yuanqi Li, Jie Guo, Yanwen Guo, and Wenping Wang. Semantic human mesh reconstruction with textures. In CVPR 2024, pages 142-152. IEEE, 2024. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.476, + 0.892, + 0.529 + ], + "angle": 0, + "content": "[47] Jingbo Zhang, Xiaoyu Li, Qi Zhang, Yanpei Cao, Ying Shan, and Jing Liao. Humanref: Single image to 3d human generation via reference-guided diffusion. In CVPR 2024, pages 1844-1854. IEEE, 2024. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.532, + 0.892, + 0.572 + ], + "angle": 0, + "content": "[48] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV 2023, pages 3813-3824. IEEE, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.575, + 0.892, + 0.614 + ], + "angle": 0, + "content": "[49] Zechuan Zhang, Li Sun, Zongxin Yang, Ling Chen, and Yi Yang. Global-correlated 3d-decoupling transformer for clothed avatar reconstruction. In NeurIPS 2023, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.618, + 0.892, + 0.671 + ], + "angle": 0, + "content": "[50] Zechuan Zhang, Zongxin Yang, and Yi Yang. SIFU: side-view conditioned implicit function for real-world usable clothed human reconstruction. In CVPR 2024, pages 9936-9947. IEEE, 2024. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.674, + 0.892, + 0.73 + ], + "angle": 0, + "content": "[51] Shenhao Zhu, Junming Leo Chen, Zuozhuo Dai, Yinghui Xu, Xun Cao, Yao Yao, Hao Zhu, and Siyu Zhu. Champ: Controllable and consistent human image animation with 3d parametric guidance. CoRR, abs/2403.14781, 2024. 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.732, + 0.892, + 0.798 + ], + "angle": 0, + "content": "[52] Zi-Xin Zou, Zhipeng Yu, Yuan-Chen Guo, Yangguang Li, Ding Liang, Yan-Pei Cao, and Song-Hai Zhang. Triplane meets gaussian splatting: Fast and generalizable single-view 3d reconstruction with transformers. In CVPR 2024, pages 10324-10335. IEEE, 2024. 4" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.798 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5580" + } + ] +] \ No newline at end of file diff --git a/2025/SinGS_ Animatable Single-Image Human Gaussian Splats with Kinematic Priors/a1907643-3cec-4712-b625-aae92326d106_origin.pdf b/2025/SinGS_ Animatable Single-Image Human Gaussian Splats with Kinematic Priors/a1907643-3cec-4712-b625-aae92326d106_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..32ab6b3cca2b6a6d28167e74bd498af476f27ed9 --- /dev/null +++ b/2025/SinGS_ Animatable Single-Image Human Gaussian Splats with Kinematic Priors/a1907643-3cec-4712-b625-aae92326d106_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8cc48d3418c10af5cebd3f3d08e7721201d13245e63f18b29df3103dd3ba6539 +size 1911059 diff --git a/2025/SinGS_ Animatable Single-Image Human Gaussian Splats with Kinematic Priors/full.md b/2025/SinGS_ Animatable Single-Image Human Gaussian Splats with Kinematic Priors/full.md new file mode 100644 index 0000000000000000000000000000000000000000..1a3ce501537b6e51d736d3dcef2176fe6cca580d --- /dev/null +++ b/2025/SinGS_ Animatable Single-Image Human Gaussian Splats with Kinematic Priors/full.md @@ -0,0 +1,229 @@ +# SinGS: Animatable Single-Image Human Gaussian Splats with Kinematic Priors + +Yufan Wu $^{1*}$ Xuanhong Chen $^{1\dagger}$ Wen Li $^{1,3}$ Shunran Jia $^{2}$ Hualiang Wei $^{1*}$ Kairui Feng $^{4}$ +Jialiang Chen $^{1}$ Yuhan Li $^{1}$ Ang He $^{1*}$ Weimin Zhang $^{1}$ Bingbing Ni $^{1}$ Wenjun Zhang $^{1}$ + +$^{1}$ Shanghai Jiao Tong University, Shanghai, China $^{2}$ DreamX Inc $^{3}$ Akool Research $^{4}$ The National Key Laboratory of Autonomous Intelligent Unmanned Systems, Shanghai, China + +{chen19910528,wmzhang,nibingbing,zhangwenjun}@sjtu.edu.cn, elvinfkr@tongji.edu.cn + +# Abstract + +Despite significant advances in accurately estimating geometry in contemporary single-image 3D human reconstruction, creating a high-quality, efficient, and animatable 3D avatar remains an open challenge. Two key obstacles persist: incomplete observation and inconsistent 3D priors. To address these challenges, we propose SinGS, aiming to achieve high-quality and efficient animatable 3D avatar reconstruction. At the heart of SinGS are two key components: Kinematic Human Diffusion and Geometry-Preserving 3D Gaussain Splatting. The former is a foundational human model that samples within pose space to generate a highly 3D-consistent and high-quality sequence of human images, inferring unseen viewpoints and providing kinematic priors. The latter is a system that reconstructs a compact, high-quality 3D avatar even under imperfect priors, achieved through a novel semantic Laplacian regularization and a geometry-preserving density control strategy that enable precise and compact assembly of 3D primitives. Extensive experiments demonstrate that SinGS enables life-like, animatable human reconstructions, maintaining both high quality and inference efficiency (up to 70FPS). + +# 1. Introduction + +Single-image 3D human reconstruction (i.e., SIHR) is an emerging modeling technique that aims to create a lifelike 3D avatar from just one image, offering a highly cost-effective and promising approach to 3D modeling. Recently, 3D Gaussian Splatting (3DGS)[14] has made significant breakthroughs in rendering efficiency and quality, establishing itself as a mainstream approach for 3D avatar reconstruction. However, current 3DGS-based methods [23, 26] for single-image human 3D avatar reconstruction suffer from critical limitations in animation, as most are + +![](images/9676d7bbff1e85aed3304a363715f481c4e2de0f9887b0cac347845423235752.jpg) +Figure 1. Existing single-image 3D avatar reconstruction methods are limited to sampling prior information from camera space. In contrast, SinGS proposes sampling from pose space, allowing us to obtain both camera and kinematic priors simultaneously. Combined with our Geometry-Preserving 3D Gaussian Splitting, this enables highly efficient animation rendering. + +designed primarily for static assets. This paper aims to develop an animation reconstruction framework that achieves a high balance between efficiency and rendering quality. + +Currently, the vast majority of SIHR methods [7, 10, 23, 24, 26, 40, 41, 46, 49, 50] focus solely on optimizing geometry while neglecting to enhance animation capabilities. As a result, they are often incapable of animation or require third-party tools [7] to achieve approximate motion. Unlike digital assets, the primary requirement for a human avatar is its ability to animate, making animation the core goal of SIHR [9]. Achieving animatable single-image human reconstruction (i.e., Animatable SIHR) is particularly + +challenging for two main reasons: Incomplete Observation: A single image lacks comprehensive viewpoint information, offering an extremely limited, partial observation that cannot address occlusion issues. Current methods [7, 10, 47] mainly rely on fine-tuned diffusion models to synthesize (i.e., 3D hallucination) missing viewpoint information (such as front and back perspectives). However, this approach does not resolve the absence of motion information. While synthesized views can alleviate some occlusion issues, the human body's non-rigid structure requires more than static viewpoints to capture its complex kinematic patterns effectively. Inconsistent 3D Prior: Current mainstream methods [5, 7, 10, 47] use Score Distillation Sampling (SDS) [25] to leverage diffusion models as supervisory modules for learning 3D priors. However, the high generative diversity of these models often leads to viewpoint inconsistencies, resulting in implausible imagery (e.g., poor-quality hidden views and inconsistencies with the input image). HumanRef [47] attempts to mitigate the instability of SDS through ReferenceNet [8], but observable inconsistencies remain widespread in the results of ReferenceNet-based methods [8, 33], indicating the difficulty in fully resolving this issue. Consequently, 3D representations should be robust enough to handle imperfect 3D priors, though this perspective has largely been overlooked in current single-image avatar frameworks. + +To address these challenges, this paper introduces a novel single-image 3D human avatar reconstruction framework, named SinGS. Unlike existing methods [23, 26], SinGS focuses on directly constructing an animatable 3DGS avatar. At the core of our SinGS framework are two key components: Kinematic Human Diffusion and a highly Geometry-Preserving 3D Gaussian Splatting. Specifically, Kinematic Human Diffusion (KHD) is designed to sample a highly 3D-consistent human sequence within the pose space, based solely on a single input image prompt. This approach enables high-quality and animatable 3D avatar reconstruction. To train KHD, we constructed a dataset with over ten thousand rotation videos, allowing KHD to learn a rich variety of human structures and texture distributions from different angles, while also gaining extensive knowledge of human kinematics. Unlike existing frameworks [5, 7, 10, 47] that rely on costly and scarce multi-view human data, our approach requires only more affordable and abundant video data, significantly reducing KHD's training costs and better aligning with scaling laws. Geometry-Preserving 3D Gaussian Splatting (i.e., GPGS) tackles artifact and hole issues under imperfect human priors (e.g., minor jitter and 3D inconsistencies) through two key designs: Semantic Laplacian Regularization addresses floating Gaussian spheres caused by 3D inconsistencies. It adjusts the constraint strength based on body parts, allowing more flexibility for areas like hair and tighter constraints + +for facial details, reducing floating Gaussian spheres while maintaining rendering quality. Geometry-Preserving Density Control resolves over-deletion and incorrect deletion of spheres in the training process. By managing the Gaussian sphere density via mesh edges, it ensures the spheres remain around the mesh, preserving geometry and preventing large holes. Extensive experiments have validated that our SinGS can generate high-quality and animatable 3D avatars at ultra-fast inference speeds (up to 70 FPS) with efficient resource usage (requiring only 3.8GB of GPU memory). + +# 2. Related Work + +# 2.1. Diffusion-based Human Generative Models + +Recent advances in diffusion-based human generative models, especially those conditioned on pose, have improved model controllability. Frameworks like ControlNet [48], IP Adapter [44], and ReferenceNet [8], along with better human attribute detection (2D/3D pose estimation, SMPL parameters [18]), enable more precise and consistent results. In image synthesis, models like PIDM [1] and LFDM [22] enhance pose transfer and image adaptation, while LEO [38] uses flow maps for motion representation. For video, DreamPose [13] and DisCo [35] improve pose-to-image translation and control separation. Models likeAnimateAnyone [8], MagicAnimate [42], and Magic-Pose [4] refine pose conditioning with ReferenceNet. Despite these improvements, these models remain resource-heavy and struggle with stability when generating images of complex movements, such as head turns or rotations. + +# 2.2. Single-Image Human Reconstruction + +Monocular video-based human reconstruction has reduced the need for specialized equipment, but single-image reconstruction remains challenging due to incomplete data. Early methods like PIFu [29] and SiCloPe [21] used pixel-aligned features and image-to-image translation, with PIFuHD [30] enhancing detail via normal maps. Later methods, such as ARCH [11] and ICON [40], incorporated body priors for animatable avatars, and ECON [41] improved model completeness by separately reconstructing front and back surfaces. Recent approaches like SiTH [7] and HumanRef [47] use generative models to fill in missing data, while SIFU [50] and Diffusion-FOF approximate multi-view features. R-Cyclic Diffuser [3] addresses limitations in Zero-1-to-3 [17], even with human-specific datasets, advancing single-image human reconstruction through generative priors and viewpoint conditioning. + +# 3. Methodology + +Given a target image $I_r$ , the objective of the SinGS architecture is to reconstruct a fully animatable 3D avatar $\mathbf{A}_t$ based on $I_r$ . To address the challenge of limited viewpoint and + +![](images/e7a026c2e6e41bf24d06eb4e3d6269db01b5c71d11ad0ad2a3e81164ccea5dfd.jpg) + +![](images/952abdf13459951d7c01846d40b42bfdd5e314026774502637291596384d9a15.jpg) + +![](images/08ab15398cc7fd693842e859d7b8824d277dbc012ecfe7c884450a93e98e4c62.jpg) +Figure 2. Overall framework of our method. + +kinematic information inherent in single-image reconstruction, $I_{t}$ is first input into the Kinematic Human Diffusion (i.e., KHD) module to complete missing information, generating an augmented sequence of rendered images $S_{t} = \{I_{p_{1}},\dots ,I_{p_{i}},\dots ,I_{p_{N}}\}$ , where $N$ represents the total number of poses generated, and $p_i$ indicates each corresponding 3D human pose. KHD reframes the less-constrained, underdetermined 3D reconstruction problem as a generative task, thus producing a fully-constrained, fully-observed 3D reconstruction formulation. Finally, $S_{t}$ is fed into our compact 3DGS module to reconstruct an animatable 3D human avatar. The compact 3DGS module is capable of reconstructing a watertight, high-quality human model even with imperfect 3D priors. The distilled avatar is high-quality, computationally efficient, and well-suited for promising applications in AR, VR, and similar fields. + +# 3.1. Kinematic Priors + +# 3.1.1 Kinematic Human Diffusion + +Unlike existing hallucination architectures [7, 10, 50], our Kinematic Human Diffusion does not rely on camera input, which is prone to estimation errors and has the added challenges of multi-view data scarcity and high costs (e.g., multi-view human data typically requires capture with specialized setups like a light stage). Our approach instead posits that sampling within the motion space provides complete priors for viewpoint and kinematic features. Additionally, our KHD does not adopt a ReferenceNet structure [8]; + +while ReferenceNet has been highly successful, it requires maintaining two UNet models, significantly increasing the framework's resource demands and preventing integration of longer contexts (e.g., 24 frames). A shorter contextual window cannot provide an image sequence with a sufficient diversity of motion samples (e.g., 150 frames). Considering the above two aspects, our KHD is designed as a diffusion structure, as illustrated in Figure 2. KHD $\mathcal{G}$ takes a single image $I_{r}$ as input and, constrained by a randomly sampled sequence of poses $\{p_1,p_2,\dots ,p_N\}$ , augments the input image into a sequence $S_{t} = \{I_{p_{1}},I_{p_{2}},\ldots ,I_{p_{N}}\}$ . This process can be formulated as follows: + +$$ +S _ {t} = \mathcal {G} \left(I _ {r}; p _ {1}, p _ {2}, \dots , p _ {N}\right). \tag {1} +$$ + +Model Architecture. Inspired by UniAnimate [36], our KHD also removes the ReferenceNet and directly inputs the reference image $I_r$ and the driving sequence into the UNet. The reference image $I_r$ is encoded by the VAE encoder of the diffusion model into a latent code $z_r$ , serving as an appearance reference throughout KHD. Additionally, $I_r$ is encoded by the CLIP image encoder [27] into features $f_r$ which are injected into the network through cross-attention by replacing the original text encoder's features in the diffusion model. This enables high-fidelity preservation of the subject's appearance characteristics. Existing works typically use 2D poses (e.g., OpenPose [2], DW-Pose [43]) as the driving signal. However, our experiments reveal that 2D poses fail to guide the diffusion model accurately in scenarios with extensive occlusions, such as when the subject + +is turning. Moreover, 2D poses cannot reliably differentiate between different viewpoints of the human body. Therefore, we use SMPL refined by ScoreHMR [31] as the driving signal $\{p_1,p_2,\dots,p_N\}$ . It is worth noting that Champ [51] also uses SMPL as guidance; however, their SMPL estimations lack accuracy, and they rely on overly strict guidance inputs, such as depth and normal vectors. This approach leads to suboptimal results, performing even worse than 2D poses in practice, for instance, by generating inconsistent body shapes. The SMPL parameters are converted into a mesh model in the specified pose using Linear Blend Skinning (LBS) [18]. This mesh is rendered as an RGB image based on the fixed camera pose maintained throughout the process. The rendered image is then encoded by a 3D Pose Encoder, a network composed of several convolutional layers, into a latent code $z_{p_i}$ with the same dimensionality as $z_{r}$ . All $z_{p_i}$ codes are first concatenated with the input noise along the channel dimension, and then sequentially along the temporal dimension to obtain the final representation $z_{P} \in \mathbb{R}^{T \times C \times H \times W}$ . Simultaneously, $I_{r}$ is processed to estimate and encode its pose latent code $z_{p_r}$ , which is then concatenated with $z_{P}$ along the temporal dimension before being fed into the diffusion model's backbone. Following most video generation models [36, 37], the UNet component of the diffusion model is also designed as a 3D UNet to ensure 3D consistency in the generated output. + +Training Dataset and Strategy. Human image animation [8, 36] primarily focuses on generating diverse frontal human movements (e.g., dancing). As a result, their training data (e.g., TikTok [12]) is largely composed of performance videos of dancers, typically featuring frontal views and wide movement ranges. In our experiments, we found that models trained on this type of data provide limited support for our 3D reconstruction tasks. These models often fail to generate results with high consistency, exhibit significant flickering in fine details, and display abnormal motion in clothing dynamics. In response to this, we collected over 6,000 high-definition videos (each around 20 seconds long) from YouTube, featuring activities such as model showcases and dancing, with a focus on segments that include turning motions. Additionally, we incorporated 600 clips from the UBC Fashion dataset [45], primarily showcasing model runway turns. As a key part of our dataset, we also integrated frontal-view data from MVHumanNet [39]. We used SAM2 [28] to segment human figures from the background in MVHumanNet, then merged the results with a white background to create videos with a consistent background, effectively eliminating the background variability caused by bounding boxes in MVHumanNet. These three sources together form a dataset of over 10,000 training videos, which we call RotationHuman. Notably, compared to multi-view data, this type of data collection is significantly more cost-effective and thus better + +suited for building large-scale datasets. During training, our KHD generates 64 pose-driven images in a single pass, setting the context window to 64 frames (i.e., $N = 64$ ). In the inference stage, we directly generate 128 frames in one pass. The advantage of one-pass inference is that it avoids consistency jitter in overlapping transition frames. KHD training requires only one phase, where the objective is to reconstruct the ground truth images corresponding to each pose, with the training loss formulated as follows: + +$$ +\mathcal {L} = \mathbb {E} _ {\theta} \left[ \| \epsilon - \epsilon_ {\theta} \left(z _ {t}, f _ {r}, t\right) \| ^ {2} \right], \tag {2} +$$ + +where, $t$ denotes the denoising step, and $z_{t}$ represents the noise at step $t$ . Note that $z_{t}$ in KHD contains diffusion states for $N$ different driving poses. The feature $f_{r}$ is the CLIP image encoding of the reference image $I_{r}$ , used to ensure ID consistency in the generated results. $\theta$ represents the parameters of the 3D UNet and the 3D Pose Encoder. The overall training process involves learning to reconstruct the target video based on $I_{r}$ as input and the pose set $\{p_1, p_2, \dots, p_N\}$ as conditioning. The generated sequence $S_{t}$ represents the sampling results of the subject in the input image $I_{r}$ within the pose space, and it can be directly used for 3D avatar reconstruction. + +# 3.2. 3D Distillation with Geometry-Preserving 3D Gaussian Splitting + +Once we obtain the sequence of pose images sampled from KHD, the goal is to distill a animatable and real-time rendered human avatar using Geometry-Preserving 3D Gaussian Splatting (GPGS). This challenge can be addressed based on mainstream approaches [16, 32, 52] in monocular 3D reconstruction. As illustrated in the Figure 2 on the right, we initialize the positions of Gaussians using the human template mesh that has undergone loop subdivision. The initial positions are set to the vertex locations, with a default of 2 upsampling iterations. This typically results in approximately $110.6K$ Gaussians. To model the avatar representation, all vertices are fed into a learnable triplane [52] to extract vertex-wise features. These features are subsequently passed to MLPs to predict the geometry and appearance attributes of the Gaussians, respectively. The entire modeling process occurs in canonical space, and we finally map this canonical space into posed space using Linear Blend Skinning (LBS) based on the SMPL framework. + +# 3.2.1 Semantic Laplacian Regularization + +Even during the reconstruction of real video, it is challenging to recover perfect static 3D consistency. Therefore, while the generated videos may appear visually plausible, every tiny ambiguity during lifting the 2D outcomes into 3D space potentially leads to catastrophic issues. Early experimental results indicated that common reconstruction meth + +![](images/79c841c208901a493b99c45a5bd7271df7ea2ff9553df4900229e904929fd7b1.jpg) +Figure 3. Visual comparison of our method with TeCH [10] and SiTH [7] for single-image reconstruction. We reposed the reconstructed results to A Pose, showcasing both animation and reconstruction capabilities. We highly recommend checking the suppl. for more results. We have provided a large number of animation result images and videos in the suppl. + +ods produced results that lacked smoothness and compactness, resulting in existing disconnected and floating gaus-. sians. Following [20], we found that applying Laplacian regularization on the canonical mesh, both for geometry and appearance, effectively mitigated the presence of gaus-. sians out of control. However, it also constrained offsets. In other words, isometrical Laplacian constraint on the whole body overly relied on the accuracy of the SMPL mesh, also making it difficult to fit areas such as hair and clothing that deviate significantly from the SMPL surface. Considering the complexity of the human body and the impossibility to obtain extremely accurate SMPL estimation, we repartition all gaussian vertices into $L = 15$ distinct semantic parts and dynamically adjust the smoothing strength. Besides, we impose the constraint on the anchors points which are closer to the rendering surface than gausicians center. The Semantic Laplacian Regularization (SLR) is defined as follows: + +$$ +\mathcal {L} _ {S L R} = \sum_ {u \in \Omega} \lambda_ {u} \left\| a _ {u, v} - \frac {1}{| M (u , v) |} \sum_ {v \in M (u, v)} a _ {u, v} \right\|, \tag {3} +$$ + +where $\Omega$ represents the set of semantic regions, $a$ denotes the surface anchor point of the $v$ -th Gaussian vertex in the $u$ -th semantic region, and $M(u,v)$ refers to the set of adjacent points of the Gaussian vertex $a_{u,v}$ located within the $u$ -th semantic region. Experiments demonstrate that the Semantic Laplacian Regularization significantly impacts the handling of more intricate situations, particularly when dealing with errors between the SMPL mesh and the actual character, which can arise from both inaccurate estimations and complex clothing. We apply Laplacians in both position and color spaces; the former enhances the smoothness of the rendering surface, while the latter ensures a more continuous variation in color space. + +# 3.2.2 Geometry-Preserving Density Control Strategy + +The strategies in vanilla 3DGS were not originally build for complex, dynamic structures like the human body. While works [20] have adapted the approach, such as initializing Gaussians based on SMPL meshes, these efforts pri + +Table 1. Quantitative comparisons with existing methods on UBC fashion dataset [45]. Note that metrics in the table are measure with the background-removed video. Some methods are unable to perform animation, and therefore, these metrics are not available for those methods. Bold indicates the best results, while Underlined represents the second-best results. + +
MethodPrior TypePSNR ↑LPIPS ↓FVD ↓Context. ↓CLIP Score ↑Speed↑
DisCo [35] (CVPR24)-30.980.259167.72.13482.4%1FPS
Animate Anyone [8] (CVPR24)-31.330.238148.81.90186.6%2FPS
Champ [51] (ECCV24)-31.210.251155.12.03884.3%1FPS
UniAnimate [36] (ArXiv24)-32.310.211141.31.81187.1%2FPS
SHERF [9] (ICCV23)-27.390.301202.32.99870.1%
SIFU [50] (CVPR24)Camera---2.89172.1%5FPS
GST [26] (ArXiv24)Camera---3.01171.7%50FPS
TeCH [10] (3DV24)Camera29.010.271188.12.33179.3%6FPS
SiTH [7] (CVPR24)Camera---2.77674.7%<1FPS
HumanRef [47] (CVPR24)Camera---2.16781.5%3FPS
SHERT [46] (CVPR2024)Inpainting29.310.266179.82.31978.8%10FPS
UniAnimate + GPGS (ours)Kinematics29.970.251168.71.99282.1%70FPS
KHD (ours)-32.810.221134.41.76892.3%2FPS
SinGS (ours)Kinematics31.330.240156.31.84788.2%70FPS
+ +marily lay a solid foundation for the optimization process but fail to maintain this advantages throughout. Incompatibilities still arise in optimization stage, prompting us to propose a novel density control scheme that is inherently compatible with hybrid representations, designed for adaptive adjustment body-wide density. This scheme draws on mesh operation algorithms [6] to better address the unique challenges of human body. For primitives that locate in under-expression mesh areas, we subdivide these regions by adding gaussians on every edge of the selected geometric triangles, which creates much smaller triangles. And we skip the trivial post-smoothing for vertices as the position of Gaussians will continue to be optimized in the network. For candidate Gaussians that are selected due to excessively large scales, as well as all newly added Gaussians, we perform rescaling to adjust them to an average size. In the case of anomalous gaussians, especially those are almost transparent, we collapse relevant edges and replace them with single vertex, resulting in decimated adjacent faces. And the priority of selected edges is decided by gaussian attributes and edge length. This entire pruning process preserves the principal structure of the human body. Our strategy result in better reconstruction and animation as the optimization was conducted under connectivity and compactness of the entire human body rather than producing holes and stacks. Moreover, this approach synergistically enhances the effectiveness of our hybrid representation. + +# 4. Experiments + +# 4.1. Experimental Setup + +Implementation Details. Our KHD primarily utilizes the RotationHuman dataset we collected, which focuses on human pose image sequences involving turning motions. The specific structure of the 3D Pose Encoder is provided in + +the supplementary materials. KHD was trained using 8 NVIDIA A100 80GB GPUs, with DeepSpeed and ZeRO2 employed to reduce memory usage. Similar toAnimate Anyone, we resized the video frames to a spatial dimension of $768 \times 512$ . We used a context window size of 64 frames, and to enhance model scalability post-training, we implemented a random context window jitter strategy [36]. During training, the model randomly samples window sizes from 64 to 32 frames in an uniformly distributed manner. The training process utilized the AdamW optimizer [15], with a learning rate set to $5e - 5$ and DDPM with 1000 steps for noise sampling. During inference, DDIM with 30 steps was used to reduce inference time. For the 3DGS component, the Adam optimizer [15] was used with a learning rate of $1e - 3$ and a cosine annealing learning scheduler was applied to accelerate convergence. The input image sequence size for 3DGS was also $768 \times 512$ . 3DGS was trained using a single NVIDIA 4090 24GB GPU. + +Baseline Models. Our baseline models consist of two parts: the first part includes human image animation models, such as DisCo [35], Animation Anyone [8] $^{1}$ , Champ [51], and UniAnimate [36]. These models are used to compare our animation capability and the superiority of KHD. The second part consists of single-image 3D avatar reconstruction models, including SHERF [9], SIFU [50], GST [26], TeCH [10], SiTH [7], HumanRef [47], and SHERT [46]. + +Metrics. We employ both PSNR and LPIPS to assess the quality of video reconstruction. To evaluate the animation capability of our model, we also use FVD [34] to measure the distribution discrepancy between the generated video and the original video, assessing the video generation performance. Furthermore, following HumanRef, we introduce contextual distance [19] and CLIP score [27] to evaluate the texture and semantic similarity between the rendered + +![](images/3a42d13bfec4a738751303ba79faedc3dfb2230b70f53d76348c9b242b679bee.jpg) +Figure 4. Comparison between KHD and other SOTA human image animation methods. Please zoom in for better observation. + +output and the input reference image. + +# 4.2. Qualitative Analysis + +Singe-Image Reconstruction. We compare SinGS with two methods from the literature that claim to support animation (TeCH, SiTH). It is worth noting that during our experiments, we found that single-image animation with these methods was extremely difficult to execute, and in some cases, nearly impossible. We had to use third-party code ${}^{2}$ to achieve the reposing for some of these methods. We animate their reconstructed outputs to a common pose (A Pose) for comparison, which allows us to evaluate both reconstruction and animation capabilities. The comparison results are shown in Figure 3, SinGS generates results with better details and maintains high-quality animation, particularly on the in-the-wild data, where SinGS's visual performance significantly outperforms the other methods. This demonstrates the high feasibility and effectiveness of the kinematic prior. + +Kinematic Human Diffusion. KHD, as the core component of our pipeline, is compared with state-of-the-art human image animation methods, includingAnimate Anyone and UniAnimate. As shown in Figure 4, we drive the same character with all three methods, focusing on comparing the quality of their generated results, particularly for the turning pose. From the figure, it is clear that both UniAnimate and our KHD maintain higher consistency thanAnimate Anyone after the character turns, with KHD surpassing UniAnimate in detail. The details marked by the red arrows in Figure 4 reveals that UniAnimate generates incorrect details, a common issue when using 2D poses, as they often struggle to distinguish body orientation during back-facing poses, leading the model to generate erroneous images. This is detrimental to reconstruction, while KHD generates videos with much higher consistency. + +![](images/8571cb3aae5ca73f739c8b03c6c01711d6540f1a88c71bf7ebf6b1602c55dd38.jpg) +Figure 5. Fifty users rated the video quality, with UniAnimate scoring highest, followed closely by our model, which outperformed others. Our model is more efficient, using only 3.5 GB of memory and achieving 70 FPS, while UniAnimate requires over 12GB and 0.5 seconds per image. + +Geometry-Preserving 3D Gaussian Splatting. Although KHD provides high-consistency pose sampling results, it inevitably has some consistency flaws, meaning it can only offer imperfect prior information. Our geometry-preserving GS method is specifically designed to address this challenge. To validate the superiority of our Geometry-Preserving 3D Gaussian Splatting (GPGS) over general human modeling methods, we compared it with recent state-of-the-art (i.e., SOTA) methods, HAHA [32] and HUGS [16]. We generate a pose sequence using KHD and then performed reconstruction with all 3 methods. The reconstruction results are shown in Figure 7. As seen, both HAHA and HUGS fail to produce a compact 3D avatar, with significant holes or floating artifacts appearing in various areas, particularly during the turning motion. In contrast, our GPGS method is able to reconstruct a compact, watertight 3D avatar model. + +# 4.3. Quantitative Analysis + +We selected 100 test videos from the UBC dataset for our quantitative analysis. For each video, we chose the first frame as the reference image and used the video's pose sequence as the driving signal. The performance of all methods was assessed by comparing the animation results with the original video to evaluate the differences. Video reconstruction metrics (including PSNR, LPIPS, and FVD) reflect the model's animation capabilities, while Contextual Distance and CLIP score allow for comparison with methods that do not support animation. The quantitative comparison results are shown in Table 1. Human image animation methods (i.e., the second to fifth rows in the table, including our KHD) outperform the reconstruction-based methods across the board. Among them, KHD surpasses the other methods in most metrics, and UniAnimate also demonstrates superior performance, which is expected due to the scale of training data and model size. While SinGS lags behind UniAnimate, it shows a significant advantage over other methods in some metrics. However, SinGS significantly outperforms diffusion-based animation methods in terms of inference efficiency and resource consumption. Our SinGS achieves 70 FPS during inference, with only + +Table 2. The results of the ablation experiments on Semantic Laplacian Regularization (i.e., Semantic Lap. Reg.) and Geometry-Preserving Density Control (i.e., GPDC) in SinGS. + +
SettingPSNRLPIPSFVD
with Semantic Lap. Reg.31.330.240156.3
with KNN Lap. Reg.30.880.250169.9
w/o Lap. Reg.30.340.272181.3
with GPDC31.010.249166.1
naive Density Control [14]30.220.276190.2
+ +3.8 GB of memory usage (tested on a single NVIDIA 4090 GPU), while UniAnimate, Champ, andAnimate Anyone each require more than 0.4 seconds per image and consume over 12 GB of memory. These quantitative results strongly demonstrate that SinGS is highly effective and feasible. We believe that using KHD, a specialized foundational model, will greatly enhance 3D reconstruction performance while significantly reducing the stringent data requirements. + +User Study. We invited 50 users to subjectively evaluate the animation results of SinGS. The evaluation was based on a scale from 0 to 10, with 10 being the highest score and 0 being the lowest. The results are visualized in Figure 5. + +# 4.4. Ablation Study + +Kinematic Priors Analysis. The number of poses sampled by KHD has a significant impact on the 3D distillation process. We analyzed how the number of poses affects the animation capability of the final reconstructed model, and the results are shown in Figure 6. We incrementally increased the number of poses from 2 (which corresponds to the typical approach of generating front and rear views, as used by most methods) up to 150. We also tested the influence of different prior models (i.e., KHD, UniAnimate, and Animate Anyone) on the results. Additionally, we examined how different 3DGS reconstruction methods handle the use of prior knowledge. As shown, performance improves rapidly with an increase in the number of poses, but the rate of improvement slows significantly after reaching around 100 poses. Moore-AA + HAHA shows the slowest improvement, while KHD + GPGS exhibits the slowest decline in performance. This suggests that increasing the number of poses does not always lead to improved performance for two main reasons: first, the prior models have already provided sufficient information; and second, none of the prior models provide perfectly consistent priors. Excessive sampling can introduce a lot of inconsistent noise, which ultimately harms performance. + +Geometry-Preserving 3D Gaussian Splitting Analysis. We conducted ablation experiments on the two key designs in GPGS: Semantic Laplacian Regularization and GPDC, with the results shown in Table 2. By replacing and removing the Laplacian regularization, we observed that Semantic Laplacian consistently outperforms in terms of reconstruction metrics. Additionally, GPDC demonstrates a perfor + +![](images/b842dcf943409a52bdf762189b7e88ee69294668cd25f9d0b1b64457ba6422f0.jpg) +Figure 6. The comparison of the impact of augmented pose numbers on the performance of the 3D avatar models. + +![](images/b6dc8cd903053ead99689f69599e83ad168ab21e6ef0529d1e35b52fc60f9414.jpg) +Figure 7. Here is a comparison of our GPGS with two SOTA 3DGS avatar methods, HAHA [32] and HUGS [16]. All methods use images sampled by KHD as body prior. + +mance advantage over the original density control strategy. The original strategy does not account for dynamic properties and geometric constraints, which makes it ineffective in addressing issues like floating and holes caused by data inconsistencies. + +# 5. Conclusion + +In this paper, we propose a new paradigm for single-image 3D avatar reconstruction. Unlike traditional methods that complete missing information in camera space, we sample in human kinematic space, enabling efficient capture of both viewpoints and motion traits. To achieve this, we introduce Kinematic Human Diffusion (KHD) for consistent pose-space generation, supported by our RotationHuman dataset, which focuses on turning motions. Additionally, we propose Geometry-Preserving 3D Gaussian Splating that reconstructs high-quality, animatable 3D avatars under imperfect human priors. This includes two key innovations: Semantic Laplacian regularization for compact, detailed constraints, and a geometry-preserving density control strategy to maintain stable human geometry. Together, these innovations allow SinGS to achieve high-quality, animatable 3D avatar reconstruction from a single image. + +# 6. Acknowledgment + +This work supported by Shanghai Post-doctoral Excellence Programe (2024290). This work was also partly by Shanghai Municipal Bureau of Culture and Tourism and Meta-verse Technology Innovation Center of Shanghai Municipal Bureau of Culture and Tourism. Kairui Feng thanks to Xiaomi foundation. + +# References + +[1] Ankan Kumar Bhunia, Salman H. Khan, Hisham Cholakkal, Rao Muhammad Anwer, Jorma Laaksonen, Mubarak Shah, and Fahad Shahbaz Khan. Person image synthesis via denoising diffusion model. In CVPR 2023, pages 5968-5976. IEEE, 2023. 2 +[2] Zhe Cao, Gines Hidalgo, Tomas Simon, Shih-En Wei, and Yaser Sheikh. Openpose: Realtime multi-person 2d pose estimation using part affinity fields. IEEE Trans. Pattern Anal. Mach. Intell., 43(1):172-186, 2021. 3 +[3] Kennard Yanting Chan, Fayao Liu, Guosheng Lin, Chuan Sheng Foo, and Weisi Lin. R-cyclic diffuser: Reductive and cyclic latent diffusion for 3d clothed human digitalization. In CVPR 2024, pages 10304-10313. IEEE, 2024. 2 +[4] Di Chang, Yichun Shi, Quankai Gao, Hongyi Xu, Jessica Fu, Guoxian Song, Qing Yan, Yizhe Zhu, Xiao Yang, and Mohammad Soleymani. Magicpose: Realistic human poses and facial expressions retargeting with identity-aware diffusion. In ICML 2024. OpenReview.net, 2024. 2 +[5] Mingjin Chen, Junhao Chen, Xiaojun Ye, Huan-ang Gao, Xiaoxue Chen, Zhaoxin Fan, and Hao Zhao. Ultraman: Single image 3d human reconstruction with ultra speed and detail. CoRR, abs/2403.12028, 2024. 2 +[6] Michael Garland and Paul S. Heckbert. Surface simplification using quadric error metrics. In SIGGRAPH 1997, pages 209-216. ACM, 1997. 6 +[7] Hsuan-I Ho, Jie Song, and Otmar Hilliges. Sith: Single-view textured human reconstruction with image-conditioned diffusion. In CVPR 2024, pages 538-549. IEEE, 2024. 1, 2, 3, 5, 6 +[8] Li Hu. Animate anyone: Consistent and controllable imagery-to-video synthesis for character animation. In CVPR 2024, pages 8153-8163. IEEE, 2024. 2, 3, 4, 6 +[9] Shoukang Hu, Fangzhou Hong, Liang Pan, Haiyi Mei, Lei Yang, and Ziwei Liu. SHERF: generalizable human nef from a single image. In ICCV 2023, pages 9318-9330. IEEE, 2023. 1, 6 +[10] Yangyi Huang, Hongwei Yi, Yuliang Xiu, Tingting Liao, Ji-axiang Tang, Deng Cai, and Justus Thies. Tech: Text-guided reconstruction of lifelike clothed humans. In 3DV 2024, pages 1531–1542. IEEE, 2024. 1, 2, 3, 5, 6 +[11] Zeng Huang, Yuanlu Xu, Christoph Lassner, Hao Li, and Tony Tung. ARCH: animatable reconstruction of clothed humans. In CVPR 2020, pages 3090–3099. Computer Vision Foundation / IEEE, 2020. 2 +[12] Yasamin Jafarian and Hyun Soo Park. Learning high fidelity depths of dressed humans by watching social media dance + +videos. In CVPR 2021, pages 12753-12762. Computer Vision Foundation / IEEE, 2021. 4 +[13] Johanna Karras, Aleksander Holynski, Ting-Chun Wang, and Ira Kemelmacher-Shlizerman. Dreampose: Fashion image-to-video synthesis via stable diffusion. In ICCV 2023, pages 22623-22633. IEEE, 2023. 2 +[14] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4):139:1-139:14, 2023. 1, 8 +[15] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR 2015, 2015. 6 +[16] Muhammed Kocabas, Jen-Hao Rick Chang, James Gabriel, Oncel Tuzel, and Anurag Ranjan. HUGS: human gaussian splats. In CVPR 2024, pages 505-515. IEEE, 2024. 4, 7, 8 +[17] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3d object. In ICCV 2023, pages 9264–9275. IEEE, 2023. 2 +[18] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J. Black. SMPL: a skinned multiperson linear model. ACM Trans. Graph., 34(6):248:1-248:16, 2015. 2, 4 +[19] Roey Mechrez, Itamar Talmi, and Lihi Zelnik-Manor. The contextual loss for image transformation with non-aligned data. In ECCV 2018, pages 800-815. Springer, 2018. 6 +[20] Gyeongsik Moon, Takaaki Shiratori, and Shunsuke Saito. Expressive whole-body 3d gaussian avatar. CoRR, abs/2407.21686, 2024. 5 +[21] Ryota Natsume, Shunsuke Saito, Zeng Huang, Weikai Chen, Chongyang Ma, Hao Li, and Shigeo Morishima. Siclope: Silhouette-based clothed people. In CVPR 2019, pages 4480-4490. Computer Vision Foundation / IEEE, 2019. 2 +[22] Haomiao Ni, Changhao Shi, Kai Li, Sharon X. Huang, and Martin Renqiang Min. Conditional image-to-video generation with latent flow diffusion models. In CVPR 2023, pages 18444-18455. IEEE, 2023. 2 +[23] Panwang Pan, Zhuo Su, Chenguo Lin, Zhen Fan, Yongjie Zhang, Zeming Li, Tingting Shen, Yadong Mu, and Yebin Liu. Humansplat: Generalizable single-image human gaussian splatting with structure priors. CoRR, abs/2406.12459, 2024. 1, 2 +[24] Marco Pesavento, Yuanlu Xu, Nikolaos Sarafianos, Robert Maier, Ziyan Wang, Chun-Han Yao, Marco Volino, Edmond Boyer, Adrian Hilton, and Tony Tung. ANIM: accurate neural implicit model for human reconstruction from a single RGB-D image. In CVPR 2024, pages 5448-5458. IEEE, 2024. 1 +[25] Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. In ICLR 2023. OpenReview.net, 2023. 2 +[26] Lorenza Prospero, Abdullah Hamdi, João F. Henriques, and Christian Rupprecht. GST: precise 3d human body from a single image with gaussian splatting transformers. CoRR, abs/2409.04196, 2024. 1, 2, 6 +[27] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, + +Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML 2021, pages 8748-8763. PMLR, 2021. 3, 6 +[28] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, Eric Mintun, Junting Pan, Kalyan Vasudev Alwala, Nicolas Carion, Chao-Yuan Wu, Ross B. Girshick, Piotr Dollar, and Christoph Feichtenhofer. SAM 2: Segment anything in images and videos. CoRR, abs/2408.00714, 2024. 4 +[29] Shunsuke Saito, Zeng Huang, Ryota Natsume, Shigeo Morishima, Hao Li, and Angjoo Kanazawa. Pifu: Pixel-aligned implicit function for high-resolution clothed human digitization. In ICCV 2019, pages 2304-2314. IEEE, 2019. 2 +[30] Shunsuke Saito, Tomas Simon, Jason M. Saragih, and Hanbyul Joo. Pifuhd: Multi-level pixel-aligned implicit function for high-resolution 3d human digitization. In CVPR 2020, pages 81–90. Computer Vision Foundation / IEEE, 2020. 2 +[31] Anastasis Stathopoulos, Ligong Han, and Dimitris N. Metaxas. Score-guided diffusion for 3d human recovery. In CVPR 2024, pages 906-915. IEEE, 2024. 4 +[32] David Svitov, Pietro Morerio, Lourdes Agapito, and Alessio Del Bue. HAHA: highly articulated gaussian human avatars with textured mesh prior. CoRR, abs/2404.01053, 2024. 4, 7, 8 +[33] Linrui Tian, Qi Wang, Bang Zhang, and Liefeng Bo. EMO: emote portrait alive - generating expressive portrait videos with audio2video diffusion model under weak conditions. CoRR, abs/2402.17485, 2024. 2 +[34] Thomas Unterthiner, Sjoerd van Steenkiste, Karol Kurach, Raphaël Marinier, Marcin Michalski, and Sylvain Gelly. Towards accurate generative models of video: A new metric & challenges. CoRR, abs/1812.01717, 2018. 6 +[35] Tan Wang, Linjie Li, Kevin Lin, Chung-Ching Lin, Zhengyuan Yang, Hanwang Zhang, Zicheng Liu, and Lijuan Wang. Disco: Disentangled control for referring human dance generation in real world. CoRR, abs/2307.00040, 2023. 2, 6 +[36] Xiang Wang, Shiwei Zhang, Changxin Gao, Jiayu Wang, Xiaogiang Zhou, Yingya Zhang, Luxin Yan, and Nong Sang. Unianimate: Taming unified video diffusion models for consistent human image animation. CoRR, abs/2406.01188, 2024. 3, 4, 6 +[37] Xiang Wang, Shiwei Zhang, Hangjie Yuan, Zhiwu Qing, Biao Gong, Yingya Zhang, Yujun Shen, Changxin Gao, and Nong Sang. A recipe for scaling up text-to-video generation with text-free videos. In CVPR 2024, pages 6572-6582. IEEE, 2024. 4 +[38] Yaohui Wang, Xin Ma, Xinyuan Chen, Antitza Dantcheva, Bo Dai, and Yu Qiao. LEO: generative latent image animatior for human video synthesis. CoRR, abs/2305.03989, 2023. 2 +[39] Zhangyang Xiong, Chenghong Li, Kenkun Liu, Hongjie Liao, Jianqiao Hu, Junyi Zhu, Shuliang Ning, Lingteng Qiu, Chongjie Wang, Shijie Wang, Shuguang Cui, and Xiaoguang Han. Mvhumannet: A large-scale dataset of multi-view daily dressing human captures. In CVPR 2024, pages 19801-19811. IEEE, 2024. 4 + +[40] Yuliang Xiu, Jinlong Yang, Dimitrios Tzionas, and Michael J. Black. ICON: Implicit Clothed humans Obtained from Normals. In CVPR 2022, pages 13296-13306, 2022. 1, 2 +[41] Yuliang Xiu, Jinlong Yang, Xu Cao, Dimitrios Tzionas, and Michael J. Black. ECON: Explicit Clothed humans Optimized via Normal integration. In CVPR 2023, 2023. 1, 2 +[42] Zhongcong Xu, Jianfeng Zhang, Jun Hao Liew, Hanshu Yan, Jia-Wei Liu, Chenxu Zhang, Jiashi Feng, and Mike Zheng Shou. Magicanimate: Temporally consistent human image animation using diffusion model. In CVPR 2024, pages 1481–1490. IEEE, 2024. 2 +[43] Zhendong Yang, Ailing Zeng, Chun Yuan, and Yu Li. Effective whole-body pose estimation with two-stages distillation. In ICCV 2023 - Workshops, pages 4212-4222. IEEE, 2023. 3 +[44] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. CoRR, abs/2308.06721, 2023. 2 +[45] Polina Zablotskaia, Aliaksandr Siarohin, Bo Zhao, and Leonid Sigal. Dwnet: Dense warp-based network for pose-guided human video generation. In BMVC 2019, page 51. BMVA Press, 2019. 4, 6 +[46] Xiaoyu Zhan, Jianxin Yang, Yuanqi Li, Jie Guo, Yanwen Guo, and Wenping Wang. Semantic human mesh reconstruction with textures. In CVPR 2024, pages 142-152. IEEE, 2024. 1, 6 +[47] Jingbo Zhang, Xiaoyu Li, Qi Zhang, Yanpei Cao, Ying Shan, and Jing Liao. Humanref: Single image to 3d human generation via reference-guided diffusion. In CVPR 2024, pages 1844-1854. IEEE, 2024. 2, 6 +[48] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV 2023, pages 3813-3824. IEEE, 2023. 2 +[49] Zechuan Zhang, Li Sun, Zongxin Yang, Ling Chen, and Yi Yang. Global-correlated 3d-decoupling transformer for clothed avatar reconstruction. In NeurIPS 2023, 2023. 1 +[50] Zechuan Zhang, Zongxin Yang, and Yi Yang. SIFU: side-view conditioned implicit function for real-world usable clothed human reconstruction. In CVPR 2024, pages 9936-9947. IEEE, 2024. 1, 2, 3, 6 +[51] Shenhao Zhu, Junming Leo Chen, Zuozhuo Dai, Yinghui Xu, Xun Cao, Yao Yao, Hao Zhu, and Siyu Zhu. Champ: Controllable and consistent human image animation with 3d parametric guidance. CoRR, abs/2403.14781, 2024. 4, 6 +[52] Zi-Xin Zou, Zhipeng Yu, Yuan-Chen Guo, Yangguang Li, Ding Liang, Yan-Pei Cao, and Song-Hai Zhang. Triplane meets gaussian splatting: Fast and generalizable single-view 3d reconstruction with transformers. In CVPR 2024, pages 10324-10335. IEEE, 2024. 4 \ No newline at end of file diff --git a/2025/SinGS_ Animatable Single-Image Human Gaussian Splats with Kinematic Priors/images.zip b/2025/SinGS_ Animatable Single-Image Human Gaussian Splats with Kinematic Priors/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..501f1e9b668c6f6f7fc15cf36fc5da15bcf50f02 --- /dev/null +++ b/2025/SinGS_ Animatable Single-Image Human Gaussian Splats with Kinematic Priors/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:39d7813321c092b925c2b2a016912a662bd176e6f0d888ee4ce25cdd36a9bd3c +size 561245 diff --git a/2025/SinGS_ Animatable Single-Image Human Gaussian Splats with Kinematic Priors/layout.json b/2025/SinGS_ Animatable Single-Image Human Gaussian Splats with Kinematic Priors/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..d8f0f068436dbbabb5d160959398587474831b37 --- /dev/null +++ b/2025/SinGS_ Animatable Single-Image Human Gaussian Splats with Kinematic Priors/layout.json @@ -0,0 +1,6827 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 47, + 103, + 545, + 120 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 103, + 545, + 120 + ], + "spans": [ + { + "bbox": [ + 47, + 103, + 545, + 120 + ], + "type": "text", + "content": "SinGS: Animatable Single-Image Human Gaussian Splats with Kinematic Priors" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 142, + 524, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 142, + 524, + 171 + ], + "spans": [ + { + "bbox": [ + 67, + 142, + 524, + 171 + ], + "type": "text", + "content": "Yufan Wu" + }, + { + "bbox": [ + 67, + 142, + 524, + 171 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 67, + 142, + 524, + 171 + ], + "type": "text", + "content": " Xuanhong Chen" + }, + { + "bbox": [ + 67, + 142, + 524, + 171 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 67, + 142, + 524, + 171 + ], + "type": "text", + "content": " Wen Li" + }, + { + "bbox": [ + 67, + 142, + 524, + 171 + ], + "type": "inline_equation", + "content": "^{1,3}" + }, + { + "bbox": [ + 67, + 142, + 524, + 171 + ], + "type": "text", + "content": " Shunran Jia" + }, + { + "bbox": [ + 67, + 142, + 524, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 67, + 142, + 524, + 171 + ], + "type": "text", + "content": " Hualiang Wei" + }, + { + "bbox": [ + 67, + 142, + 524, + 171 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 67, + 142, + 524, + 171 + ], + "type": "text", + "content": " Kairui Feng" + }, + { + "bbox": [ + 67, + 142, + 524, + 171 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 67, + 142, + 524, + 171 + ], + "type": "text", + "content": " \nJialiang Chen" + }, + { + "bbox": [ + 67, + 142, + 524, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 67, + 142, + 524, + 171 + ], + "type": "text", + "content": " Yuhan Li" + }, + { + "bbox": [ + 67, + 142, + 524, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 67, + 142, + 524, + 171 + ], + "type": "text", + "content": " Ang He" + }, + { + "bbox": [ + 67, + 142, + 524, + 171 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 67, + 142, + 524, + 171 + ], + "type": "text", + "content": " Weimin Zhang" + }, + { + "bbox": [ + 67, + 142, + 524, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 67, + 142, + 524, + 171 + ], + "type": "text", + "content": " Bingbing Ni" + }, + { + "bbox": [ + 67, + 142, + 524, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 67, + 142, + 524, + 171 + ], + "type": "text", + "content": " Wenjun Zhang" + }, + { + "bbox": [ + 67, + 142, + 524, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 66, + 171, + 527, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 171, + 527, + 199 + ], + "spans": [ + { + "bbox": [ + 66, + 171, + 527, + 199 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 66, + 171, + 527, + 199 + ], + "type": "text", + "content": "Shanghai Jiao Tong University, Shanghai, China " + }, + { + "bbox": [ + 66, + 171, + 527, + 199 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 66, + 171, + 527, + 199 + ], + "type": "text", + "content": "DreamX Inc " + }, + { + "bbox": [ + 66, + 171, + 527, + 199 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 66, + 171, + 527, + 199 + ], + "type": "text", + "content": "Akool Research " + }, + { + "bbox": [ + 66, + 171, + 527, + 199 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 66, + 171, + 527, + 199 + ], + "type": "text", + "content": "The National Key Laboratory of Autonomous Intelligent Unmanned Systems, Shanghai, China" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 73, + 201, + 511, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 201, + 511, + 213 + ], + "spans": [ + { + "bbox": [ + 73, + 201, + 511, + 213 + ], + "type": "text", + "content": "{chen19910528,wmzhang,nibingbing,zhangwenjun}@sjtu.edu.cn, elvinfkr@tongji.edu.cn" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 143, + 241, + 191, + 253 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 241, + 191, + 253 + ], + "spans": [ + { + "bbox": [ + 143, + 241, + 191, + 253 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 267, + 290, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 267, + 290, + 518 + ], + "spans": [ + { + "bbox": [ + 46, + 267, + 290, + 518 + ], + "type": "text", + "content": "Despite significant advances in accurately estimating geometry in contemporary single-image 3D human reconstruction, creating a high-quality, efficient, and animatable 3D avatar remains an open challenge. Two key obstacles persist: incomplete observation and inconsistent 3D priors. To address these challenges, we propose SinGS, aiming to achieve high-quality and efficient animatable 3D avatar reconstruction. At the heart of SinGS are two key components: Kinematic Human Diffusion and Geometry-Preserving 3D Gaussain Splatting. The former is a foundational human model that samples within pose space to generate a highly 3D-consistent and high-quality sequence of human images, inferring unseen viewpoints and providing kinematic priors. The latter is a system that reconstructs a compact, high-quality 3D avatar even under imperfect priors, achieved through a novel semantic Laplacian regularization and a geometry-preserving density control strategy that enable precise and compact assembly of 3D primitives. Extensive experiments demonstrate that SinGS enables life-like, animatable human reconstructions, maintaining both high quality and inference efficiency (up to 70FPS)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 543, + 127, + 555 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 543, + 127, + 555 + ], + "spans": [ + { + "bbox": [ + 47, + 543, + 127, + 555 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 563, + 287, + 683 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 563, + 287, + 683 + ], + "spans": [ + { + "bbox": [ + 46, + 563, + 287, + 683 + ], + "type": "text", + "content": "Single-image 3D human reconstruction (i.e., SIHR) is an emerging modeling technique that aims to create a lifelike 3D avatar from just one image, offering a highly cost-effective and promising approach to 3D modeling. Recently, 3D Gaussian Splatting (3DGS)[14] has made significant breakthroughs in rendering efficiency and quality, establishing itself as a mainstream approach for 3D avatar reconstruction. However, current 3DGS-based methods [23, 26] for single-image human 3D avatar reconstruction suffer from critical limitations in animation, as most are" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 310, + 242, + 544, + 483 + ], + "blocks": [ + { + "bbox": [ + 310, + 242, + 544, + 483 + ], + "lines": [ + { + "bbox": [ + 310, + 242, + 544, + 483 + ], + "spans": [ + { + "bbox": [ + 310, + 242, + 544, + 483 + ], + "type": "image", + "image_path": "9676d7bbff1e85aed3304a363715f481c4e2de0f9887b0cac347845423235752.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 495, + 545, + 561 + ], + "lines": [ + { + "bbox": [ + 304, + 495, + 545, + 561 + ], + "spans": [ + { + "bbox": [ + 304, + 495, + 545, + 561 + ], + "type": "text", + "content": "Figure 1. Existing single-image 3D avatar reconstruction methods are limited to sampling prior information from camera space. In contrast, SinGS proposes sampling from pose space, allowing us to obtain both camera and kinematic priors simultaneously. Combined with our Geometry-Preserving 3D Gaussian Splitting, this enables highly efficient animation rendering." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 569, + 545, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 569, + 545, + 605 + ], + "spans": [ + { + "bbox": [ + 304, + 569, + 545, + 605 + ], + "type": "text", + "content": "designed primarily for static assets. This paper aims to develop an animation reconstruction framework that achieves a high balance between efficiency and rendering quality." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 605, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 546, + 714 + ], + "type": "text", + "content": "Currently, the vast majority of SIHR methods [7, 10, 23, 24, 26, 40, 41, 46, 49, 50] focus solely on optimizing geometry while neglecting to enhance animation capabilities. As a result, they are often incapable of animation or require third-party tools [7] to achieve approximate motion. Unlike digital assets, the primary requirement for a human avatar is its ability to animate, making animation the core goal of SIHR [9]. Achieving animatable single-image human reconstruction (i.e., Animatable SIHR) is particularly" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 693, + 275, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 693, + 275, + 703 + ], + "spans": [ + { + "bbox": [ + 58, + 693, + 275, + 703 + ], + "type": "text", + "content": "*Work done during an internship at Shanghai Jiao Tong University." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 59, + 703, + 193, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 703, + 193, + 712 + ], + "spans": [ + { + "bbox": [ + 59, + 703, + 193, + 712 + ], + "type": "text", + "content": "† Corresponding author: Xuanhong Chen." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "text", + "content": "5571" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 384 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 384 + ], + "type": "text", + "content": "challenging for two main reasons: Incomplete Observation: A single image lacks comprehensive viewpoint information, offering an extremely limited, partial observation that cannot address occlusion issues. Current methods [7, 10, 47] mainly rely on fine-tuned diffusion models to synthesize (i.e., 3D hallucination) missing viewpoint information (such as front and back perspectives). However, this approach does not resolve the absence of motion information. While synthesized views can alleviate some occlusion issues, the human body's non-rigid structure requires more than static viewpoints to capture its complex kinematic patterns effectively. Inconsistent 3D Prior: Current mainstream methods [5, 7, 10, 47] use Score Distillation Sampling (SDS) [25] to leverage diffusion models as supervisory modules for learning 3D priors. However, the high generative diversity of these models often leads to viewpoint inconsistencies, resulting in implausible imagery (e.g., poor-quality hidden views and inconsistencies with the input image). HumanRef [47] attempts to mitigate the instability of SDS through ReferenceNet [8], but observable inconsistencies remain widespread in the results of ReferenceNet-based methods [8, 33], indicating the difficulty in fully resolving this issue. Consequently, 3D representations should be robust enough to handle imperfect 3D priors, though this perspective has largely been overlooked in current single-image avatar frameworks." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 391, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 391, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 391, + 289, + 715 + ], + "type": "text", + "content": "To address these challenges, this paper introduces a novel single-image 3D human avatar reconstruction framework, named SinGS. Unlike existing methods [23, 26], SinGS focuses on directly constructing an animatable 3DGS avatar. At the core of our SinGS framework are two key components: Kinematic Human Diffusion and a highly Geometry-Preserving 3D Gaussian Splatting. Specifically, Kinematic Human Diffusion (KHD) is designed to sample a highly 3D-consistent human sequence within the pose space, based solely on a single input image prompt. This approach enables high-quality and animatable 3D avatar reconstruction. To train KHD, we constructed a dataset with over ten thousand rotation videos, allowing KHD to learn a rich variety of human structures and texture distributions from different angles, while also gaining extensive knowledge of human kinematics. Unlike existing frameworks [5, 7, 10, 47] that rely on costly and scarce multi-view human data, our approach requires only more affordable and abundant video data, significantly reducing KHD's training costs and better aligning with scaling laws. Geometry-Preserving 3D Gaussian Splatting (i.e., GPGS) tackles artifact and hole issues under imperfect human priors (e.g., minor jitter and 3D inconsistencies) through two key designs: Semantic Laplacian Regularization addresses floating Gaussian spheres caused by 3D inconsistencies. It adjusts the constraint strength based on body parts, allowing more flexibility for areas like hair and tighter constraints" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 304, + 72, + 547, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 194 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 194 + ], + "type": "text", + "content": "for facial details, reducing floating Gaussian spheres while maintaining rendering quality. Geometry-Preserving Density Control resolves over-deletion and incorrect deletion of spheres in the training process. By managing the Gaussian sphere density via mesh edges, it ensures the spheres remain around the mesh, preserving geometry and preventing large holes. Extensive experiments have validated that our SinGS can generate high-quality and animatable 3D avatars at ultra-fast inference speeds (up to 70 FPS) with efficient resource usage (requiring only 3.8GB of GPU memory)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 306, + 201, + 392, + 215 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 201, + 392, + 215 + ], + "spans": [ + { + "bbox": [ + 306, + 201, + 392, + 215 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 306, + 221, + 531, + 233 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 221, + 531, + 233 + ], + "spans": [ + { + "bbox": [ + 306, + 221, + 531, + 233 + ], + "type": "text", + "content": "2.1. Diffusion-based Human Generative Models" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 239, + 547, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 239, + 547, + 431 + ], + "spans": [ + { + "bbox": [ + 304, + 239, + 547, + 431 + ], + "type": "text", + "content": "Recent advances in diffusion-based human generative models, especially those conditioned on pose, have improved model controllability. Frameworks like ControlNet [48], IP Adapter [44], and ReferenceNet [8], along with better human attribute detection (2D/3D pose estimation, SMPL parameters [18]), enable more precise and consistent results. In image synthesis, models like PIDM [1] and LFDM [22] enhance pose transfer and image adaptation, while LEO [38] uses flow maps for motion representation. For video, DreamPose [13] and DisCo [35] improve pose-to-image translation and control separation. Models likeAnimateAnyone [8], MagicAnimate [42], and Magic-Pose [4] refine pose conditioning with ReferenceNet. Despite these improvements, these models remain resource-heavy and struggle with stability when generating images of complex movements, such as head turns or rotations." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 306, + 437, + 503, + 451 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 437, + 503, + 451 + ], + "spans": [ + { + "bbox": [ + 306, + 437, + 503, + 451 + ], + "type": "text", + "content": "2.2. Single-Image Human Reconstruction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 456, + 547, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 456, + 547, + 649 + ], + "spans": [ + { + "bbox": [ + 304, + 456, + 547, + 649 + ], + "type": "text", + "content": "Monocular video-based human reconstruction has reduced the need for specialized equipment, but single-image reconstruction remains challenging due to incomplete data. Early methods like PIFu [29] and SiCloPe [21] used pixel-aligned features and image-to-image translation, with PIFuHD [30] enhancing detail via normal maps. Later methods, such as ARCH [11] and ICON [40], incorporated body priors for animatable avatars, and ECON [41] improved model completeness by separately reconstructing front and back surfaces. Recent approaches like SiTH [7] and HumanRef [47] use generative models to fill in missing data, while SIFU [50] and Diffusion-FOF approximate multi-view features. R-Cyclic Diffuser [3] addresses limitations in Zero-1-to-3 [17], even with human-specific datasets, advancing single-image human reconstruction through generative priors and viewpoint conditioning." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 658, + 389, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 658, + 389, + 671 + ], + "spans": [ + { + "bbox": [ + 306, + 658, + 389, + 671 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "type": "text", + "content": "Given a target image " + }, + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "type": "inline_equation", + "content": "I_r" + }, + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "type": "text", + "content": ", the objective of the SinGS architecture is to reconstruct a fully animatable 3D avatar " + }, + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_t" + }, + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "type": "text", + "content": " based on " + }, + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "type": "inline_equation", + "content": "I_r" + }, + { + "bbox": [ + 304, + 677, + 547, + 715 + ], + "type": "text", + "content": ". To address the challenge of limited viewpoint and" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 758 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 758 + ], + "type": "text", + "content": "5572" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 47, + 69, + 304, + 256 + ], + "blocks": [ + { + "bbox": [ + 47, + 69, + 304, + 256 + ], + "lines": [ + { + "bbox": [ + 47, + 69, + 304, + 256 + ], + "spans": [ + { + "bbox": [ + 47, + 69, + 304, + 256 + ], + "type": "image", + "image_path": "e7a026c2e6e41bf24d06eb4e3d6269db01b5c71d11ad0ad2a3e81164ccea5dfd.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 304, + 70, + 545, + 255 + ], + "blocks": [ + { + "bbox": [ + 304, + 70, + 545, + 255 + ], + "lines": [ + { + "bbox": [ + 304, + 70, + 545, + 255 + ], + "spans": [ + { + "bbox": [ + 304, + 70, + 545, + 255 + ], + "type": "image", + "image_path": "952abdf13459951d7c01846d40b42bfdd5e314026774502637291596384d9a15.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 48, + 257, + 545, + 330 + ], + "blocks": [ + { + "bbox": [ + 48, + 257, + 545, + 330 + ], + "lines": [ + { + "bbox": [ + 48, + 257, + 545, + 330 + ], + "spans": [ + { + "bbox": [ + 48, + 257, + 545, + 330 + ], + "type": "image", + "image_path": "08ab15398cc7fd693842e859d7b8824d277dbc012ecfe7c884450a93e98e4c62.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 215, + 339, + 376, + 350 + ], + "lines": [ + { + "bbox": [ + 215, + 339, + 376, + 350 + ], + "spans": [ + { + "bbox": [ + 215, + 339, + 376, + 350 + ], + "type": "text", + "content": "Figure 2. Overall framework of our method." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 360, + 290, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 360, + 290, + 552 + ], + "spans": [ + { + "bbox": [ + 46, + 360, + 290, + 552 + ], + "type": "text", + "content": "kinematic information inherent in single-image reconstruction, " + }, + { + "bbox": [ + 46, + 360, + 290, + 552 + ], + "type": "inline_equation", + "content": "I_{t}" + }, + { + "bbox": [ + 46, + 360, + 290, + 552 + ], + "type": "text", + "content": " is first input into the Kinematic Human Diffusion (i.e., KHD) module to complete missing information, generating an augmented sequence of rendered images " + }, + { + "bbox": [ + 46, + 360, + 290, + 552 + ], + "type": "inline_equation", + "content": "S_{t} = \\{I_{p_{1}},\\dots ,I_{p_{i}},\\dots ,I_{p_{N}}\\}" + }, + { + "bbox": [ + 46, + 360, + 290, + 552 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 360, + 290, + 552 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 360, + 290, + 552 + ], + "type": "text", + "content": " represents the total number of poses generated, and " + }, + { + "bbox": [ + 46, + 360, + 290, + 552 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 46, + 360, + 290, + 552 + ], + "type": "text", + "content": " indicates each corresponding 3D human pose. KHD reframes the less-constrained, underdetermined 3D reconstruction problem as a generative task, thus producing a fully-constrained, fully-observed 3D reconstruction formulation. Finally, " + }, + { + "bbox": [ + 46, + 360, + 290, + 552 + ], + "type": "inline_equation", + "content": "S_{t}" + }, + { + "bbox": [ + 46, + 360, + 290, + 552 + ], + "type": "text", + "content": " is fed into our compact 3DGS module to reconstruct an animatable 3D human avatar. The compact 3DGS module is capable of reconstructing a watertight, high-quality human model even with imperfect 3D priors. The distilled avatar is high-quality, computationally efficient, and well-suited for promising applications in AR, VR, and similar fields." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 564, + 151, + 575 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 564, + 151, + 575 + ], + "spans": [ + { + "bbox": [ + 47, + 564, + 151, + 575 + ], + "type": "text", + "content": "3.1. Kinematic Priors" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 584, + 201, + 594 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 584, + 201, + 594 + ], + "spans": [ + { + "bbox": [ + 47, + 584, + 201, + 594 + ], + "type": "text", + "content": "3.1.1 Kinematic Human Diffusion" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "type": "text", + "content": "Unlike existing hallucination architectures [7, 10, 50], our Kinematic Human Diffusion does not rely on camera input, which is prone to estimation errors and has the added challenges of multi-view data scarcity and high costs (e.g., multi-view human data typically requires capture with specialized setups like a light stage). Our approach instead posits that sampling within the motion space provides complete priors for viewpoint and kinematic features. Additionally, our KHD does not adopt a ReferenceNet structure [8];" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 360, + 545, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 360, + 545, + 503 + ], + "spans": [ + { + "bbox": [ + 304, + 360, + 545, + 503 + ], + "type": "text", + "content": "while ReferenceNet has been highly successful, it requires maintaining two UNet models, significantly increasing the framework's resource demands and preventing integration of longer contexts (e.g., 24 frames). A shorter contextual window cannot provide an image sequence with a sufficient diversity of motion samples (e.g., 150 frames). Considering the above two aspects, our KHD is designed as a diffusion structure, as illustrated in Figure 2. KHD " + }, + { + "bbox": [ + 304, + 360, + 545, + 503 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 304, + 360, + 545, + 503 + ], + "type": "text", + "content": " takes a single image " + }, + { + "bbox": [ + 304, + 360, + 545, + 503 + ], + "type": "inline_equation", + "content": "I_{r}" + }, + { + "bbox": [ + 304, + 360, + 545, + 503 + ], + "type": "text", + "content": " as input and, constrained by a randomly sampled sequence of poses " + }, + { + "bbox": [ + 304, + 360, + 545, + 503 + ], + "type": "inline_equation", + "content": "\\{p_1,p_2,\\dots ,p_N\\}" + }, + { + "bbox": [ + 304, + 360, + 545, + 503 + ], + "type": "text", + "content": ", augments the input image into a sequence " + }, + { + "bbox": [ + 304, + 360, + 545, + 503 + ], + "type": "inline_equation", + "content": "S_{t} = \\{I_{p_{1}},I_{p_{2}},\\ldots ,I_{p_{N}}\\}" + }, + { + "bbox": [ + 304, + 360, + 545, + 503 + ], + "type": "text", + "content": ". This process can be formulated as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 372, + 513, + 545, + 525 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 513, + 545, + 525 + ], + "spans": [ + { + "bbox": [ + 372, + 513, + 545, + 525 + ], + "type": "interline_equation", + "content": "S _ {t} = \\mathcal {G} \\left(I _ {r}; p _ {1}, p _ {2}, \\dots , p _ {N}\\right). \\tag {1}", + "image_path": "74f44482bb38a97279082d18b72dcfa300efedd0462d0fc7ee1996aee29c1729.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 533, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 533, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 533, + 546, + 714 + ], + "type": "text", + "content": "Model Architecture. Inspired by UniAnimate [36], our KHD also removes the ReferenceNet and directly inputs the reference image " + }, + { + "bbox": [ + 304, + 533, + 546, + 714 + ], + "type": "inline_equation", + "content": "I_r" + }, + { + "bbox": [ + 304, + 533, + 546, + 714 + ], + "type": "text", + "content": " and the driving sequence into the UNet. The reference image " + }, + { + "bbox": [ + 304, + 533, + 546, + 714 + ], + "type": "inline_equation", + "content": "I_r" + }, + { + "bbox": [ + 304, + 533, + 546, + 714 + ], + "type": "text", + "content": " is encoded by the VAE encoder of the diffusion model into a latent code " + }, + { + "bbox": [ + 304, + 533, + 546, + 714 + ], + "type": "inline_equation", + "content": "z_r" + }, + { + "bbox": [ + 304, + 533, + 546, + 714 + ], + "type": "text", + "content": ", serving as an appearance reference throughout KHD. Additionally, " + }, + { + "bbox": [ + 304, + 533, + 546, + 714 + ], + "type": "inline_equation", + "content": "I_r" + }, + { + "bbox": [ + 304, + 533, + 546, + 714 + ], + "type": "text", + "content": " is encoded by the CLIP image encoder [27] into features " + }, + { + "bbox": [ + 304, + 533, + 546, + 714 + ], + "type": "inline_equation", + "content": "f_r" + }, + { + "bbox": [ + 304, + 533, + 546, + 714 + ], + "type": "text", + "content": " which are injected into the network through cross-attention by replacing the original text encoder's features in the diffusion model. This enables high-fidelity preservation of the subject's appearance characteristics. Existing works typically use 2D poses (e.g., OpenPose [2], DW-Pose [43]) as the driving signal. However, our experiments reveal that 2D poses fail to guide the diffusion model accurately in scenarios with extensive occlusions, such as when the subject" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5573" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 289, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 289, + 384 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 289, + 384 + ], + "type": "text", + "content": "is turning. Moreover, 2D poses cannot reliably differentiate between different viewpoints of the human body. Therefore, we use SMPL refined by ScoreHMR [31] as the driving signal " + }, + { + "bbox": [ + 47, + 72, + 289, + 384 + ], + "type": "inline_equation", + "content": "\\{p_1,p_2,\\dots,p_N\\}" + }, + { + "bbox": [ + 47, + 72, + 289, + 384 + ], + "type": "text", + "content": ". It is worth noting that Champ [51] also uses SMPL as guidance; however, their SMPL estimations lack accuracy, and they rely on overly strict guidance inputs, such as depth and normal vectors. This approach leads to suboptimal results, performing even worse than 2D poses in practice, for instance, by generating inconsistent body shapes. The SMPL parameters are converted into a mesh model in the specified pose using Linear Blend Skinning (LBS) [18]. This mesh is rendered as an RGB image based on the fixed camera pose maintained throughout the process. The rendered image is then encoded by a 3D Pose Encoder, a network composed of several convolutional layers, into a latent code " + }, + { + "bbox": [ + 47, + 72, + 289, + 384 + ], + "type": "inline_equation", + "content": "z_{p_i}" + }, + { + "bbox": [ + 47, + 72, + 289, + 384 + ], + "type": "text", + "content": " with the same dimensionality as " + }, + { + "bbox": [ + 47, + 72, + 289, + 384 + ], + "type": "inline_equation", + "content": "z_{r}" + }, + { + "bbox": [ + 47, + 72, + 289, + 384 + ], + "type": "text", + "content": ". All " + }, + { + "bbox": [ + 47, + 72, + 289, + 384 + ], + "type": "inline_equation", + "content": "z_{p_i}" + }, + { + "bbox": [ + 47, + 72, + 289, + 384 + ], + "type": "text", + "content": " codes are first concatenated with the input noise along the channel dimension, and then sequentially along the temporal dimension to obtain the final representation " + }, + { + "bbox": [ + 47, + 72, + 289, + 384 + ], + "type": "inline_equation", + "content": "z_{P} \\in \\mathbb{R}^{T \\times C \\times H \\times W}" + }, + { + "bbox": [ + 47, + 72, + 289, + 384 + ], + "type": "text", + "content": ". Simultaneously, " + }, + { + "bbox": [ + 47, + 72, + 289, + 384 + ], + "type": "inline_equation", + "content": "I_{r}" + }, + { + "bbox": [ + 47, + 72, + 289, + 384 + ], + "type": "text", + "content": " is processed to estimate and encode its pose latent code " + }, + { + "bbox": [ + 47, + 72, + 289, + 384 + ], + "type": "inline_equation", + "content": "z_{p_r}" + }, + { + "bbox": [ + 47, + 72, + 289, + 384 + ], + "type": "text", + "content": ", which is then concatenated with " + }, + { + "bbox": [ + 47, + 72, + 289, + 384 + ], + "type": "inline_equation", + "content": "z_{P}" + }, + { + "bbox": [ + 47, + 72, + 289, + 384 + ], + "type": "text", + "content": " along the temporal dimension before being fed into the diffusion model's backbone. Following most video generation models [36, 37], the UNet component of the diffusion model is also designed as a 3D UNet to ensure 3D consistency in the generated output." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 391, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 391, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 391, + 289, + 714 + ], + "type": "text", + "content": "Training Dataset and Strategy. Human image animation [8, 36] primarily focuses on generating diverse frontal human movements (e.g., dancing). As a result, their training data (e.g., TikTok [12]) is largely composed of performance videos of dancers, typically featuring frontal views and wide movement ranges. In our experiments, we found that models trained on this type of data provide limited support for our 3D reconstruction tasks. These models often fail to generate results with high consistency, exhibit significant flickering in fine details, and display abnormal motion in clothing dynamics. In response to this, we collected over 6,000 high-definition videos (each around 20 seconds long) from YouTube, featuring activities such as model showcases and dancing, with a focus on segments that include turning motions. Additionally, we incorporated 600 clips from the UBC Fashion dataset [45], primarily showcasing model runway turns. As a key part of our dataset, we also integrated frontal-view data from MVHumanNet [39]. We used SAM2 [28] to segment human figures from the background in MVHumanNet, then merged the results with a white background to create videos with a consistent background, effectively eliminating the background variability caused by bounding boxes in MVHumanNet. These three sources together form a dataset of over 10,000 training videos, which we call RotationHuman. Notably, compared to multi-view data, this type of data collection is significantly more cost-effective and thus better" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 304, + 72, + 547, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 180 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 180 + ], + "type": "text", + "content": "suited for building large-scale datasets. During training, our KHD generates 64 pose-driven images in a single pass, setting the context window to 64 frames (i.e., " + }, + { + "bbox": [ + 304, + 72, + 547, + 180 + ], + "type": "inline_equation", + "content": "N = 64" + }, + { + "bbox": [ + 304, + 72, + 547, + 180 + ], + "type": "text", + "content": "). In the inference stage, we directly generate 128 frames in one pass. The advantage of one-pass inference is that it avoids consistency jitter in overlapping transition frames. KHD training requires only one phase, where the objective is to reconstruct the ground truth images corresponding to each pose, with the training loss formulated as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 367, + 191, + 545, + 205 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 191, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 367, + 191, + 545, + 205 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathbb {E} _ {\\theta} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} \\left(z _ {t}, f _ {r}, t\\right) \\| ^ {2} \\right], \\tag {2}", + "image_path": "14c1a09aae6586a833d08a7aa323aa6d800274749db5b53bae3562c4b186032d.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 209, + 547, + 352 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 209, + 547, + 352 + ], + "spans": [ + { + "bbox": [ + 304, + 209, + 547, + 352 + ], + "type": "text", + "content": "where, " + }, + { + "bbox": [ + 304, + 209, + 547, + 352 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 209, + 547, + 352 + ], + "type": "text", + "content": " denotes the denoising step, and " + }, + { + "bbox": [ + 304, + 209, + 547, + 352 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 304, + 209, + 547, + 352 + ], + "type": "text", + "content": " represents the noise at step " + }, + { + "bbox": [ + 304, + 209, + 547, + 352 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 209, + 547, + 352 + ], + "type": "text", + "content": ". Note that " + }, + { + "bbox": [ + 304, + 209, + 547, + 352 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 304, + 209, + 547, + 352 + ], + "type": "text", + "content": " in KHD contains diffusion states for " + }, + { + "bbox": [ + 304, + 209, + 547, + 352 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 209, + 547, + 352 + ], + "type": "text", + "content": " different driving poses. The feature " + }, + { + "bbox": [ + 304, + 209, + 547, + 352 + ], + "type": "inline_equation", + "content": "f_{r}" + }, + { + "bbox": [ + 304, + 209, + 547, + 352 + ], + "type": "text", + "content": " is the CLIP image encoding of the reference image " + }, + { + "bbox": [ + 304, + 209, + 547, + 352 + ], + "type": "inline_equation", + "content": "I_{r}" + }, + { + "bbox": [ + 304, + 209, + 547, + 352 + ], + "type": "text", + "content": ", used to ensure ID consistency in the generated results. " + }, + { + "bbox": [ + 304, + 209, + 547, + 352 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 304, + 209, + 547, + 352 + ], + "type": "text", + "content": " represents the parameters of the 3D UNet and the 3D Pose Encoder. The overall training process involves learning to reconstruct the target video based on " + }, + { + "bbox": [ + 304, + 209, + 547, + 352 + ], + "type": "inline_equation", + "content": "I_{r}" + }, + { + "bbox": [ + 304, + 209, + 547, + 352 + ], + "type": "text", + "content": " as input and the pose set " + }, + { + "bbox": [ + 304, + 209, + 547, + 352 + ], + "type": "inline_equation", + "content": "\\{p_1, p_2, \\dots, p_N\\}" + }, + { + "bbox": [ + 304, + 209, + 547, + 352 + ], + "type": "text", + "content": " as conditioning. The generated sequence " + }, + { + "bbox": [ + 304, + 209, + 547, + 352 + ], + "type": "inline_equation", + "content": "S_{t}" + }, + { + "bbox": [ + 304, + 209, + 547, + 352 + ], + "type": "text", + "content": " represents the sampling results of the subject in the input image " + }, + { + "bbox": [ + 304, + 209, + 547, + 352 + ], + "type": "inline_equation", + "content": "I_{r}" + }, + { + "bbox": [ + 304, + 209, + 547, + 352 + ], + "type": "text", + "content": " within the pose space, and it can be directly used for 3D avatar reconstruction." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 305, + 360, + 545, + 385 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 360, + 545, + 385 + ], + "spans": [ + { + "bbox": [ + 305, + 360, + 545, + 385 + ], + "type": "text", + "content": "3.2. 3D Distillation with Geometry-Preserving 3D Gaussian Splitting" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 391, + 547, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 391, + 547, + 606 + ], + "spans": [ + { + "bbox": [ + 304, + 391, + 547, + 606 + ], + "type": "text", + "content": "Once we obtain the sequence of pose images sampled from KHD, the goal is to distill a animatable and real-time rendered human avatar using Geometry-Preserving 3D Gaussian Splatting (GPGS). This challenge can be addressed based on mainstream approaches [16, 32, 52] in monocular 3D reconstruction. As illustrated in the Figure 2 on the right, we initialize the positions of Gaussians using the human template mesh that has undergone loop subdivision. The initial positions are set to the vertex locations, with a default of 2 upsampling iterations. This typically results in approximately " + }, + { + "bbox": [ + 304, + 391, + 547, + 606 + ], + "type": "inline_equation", + "content": "110.6K" + }, + { + "bbox": [ + 304, + 391, + 547, + 606 + ], + "type": "text", + "content": " Gaussians. To model the avatar representation, all vertices are fed into a learnable triplane [52] to extract vertex-wise features. These features are subsequently passed to MLPs to predict the geometry and appearance attributes of the Gaussians, respectively. The entire modeling process occurs in canonical space, and we finally map this canonical space into posed space using Linear Blend Skinning (LBS) based on the SMPL framework." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 622, + 491, + 634 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 622, + 491, + 634 + ], + "spans": [ + { + "bbox": [ + 305, + 622, + 491, + 634 + ], + "type": "text", + "content": "3.2.1 Semantic Laplacian Regularization" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 641, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 547, + 713 + ], + "type": "text", + "content": "Even during the reconstruction of real video, it is challenging to recover perfect static 3D consistency. Therefore, while the generated videos may appear visually plausible, every tiny ambiguity during lifting the 2D outcomes into 3D space potentially leads to catastrophic issues. Early experimental results indicated that common reconstruction meth" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5574" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 48, + 73, + 545, + 415 + ], + "blocks": [ + { + "bbox": [ + 48, + 73, + 545, + 415 + ], + "lines": [ + { + "bbox": [ + 48, + 73, + 545, + 415 + ], + "spans": [ + { + "bbox": [ + 48, + 73, + 545, + 415 + ], + "type": "image", + "image_path": "79c841c208901a493b99c45a5bd7271df7ea2ff9553df4900229e904929fd7b1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 426, + 546, + 460 + ], + "lines": [ + { + "bbox": [ + 46, + 426, + 546, + 460 + ], + "spans": [ + { + "bbox": [ + 46, + 426, + 546, + 460 + ], + "type": "text", + "content": "Figure 3. Visual comparison of our method with TeCH [10] and SiTH [7] for single-image reconstruction. We reposed the reconstructed results to A Pose, showcasing both animation and reconstruction capabilities. We highly recommend checking the suppl. for more results. We have provided a large number of animation result images and videos in the suppl." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 469, + 289, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 469, + 289, + 673 + ], + "spans": [ + { + "bbox": [ + 46, + 469, + 289, + 673 + ], + "type": "text", + "content": "ods produced results that lacked smoothness and compactness, resulting in existing disconnected and floating gaus-. sians. Following [20], we found that applying Laplacian regularization on the canonical mesh, both for geometry and appearance, effectively mitigated the presence of gaus-. sians out of control. However, it also constrained offsets. In other words, isometrical Laplacian constraint on the whole body overly relied on the accuracy of the SMPL mesh, also making it difficult to fit areas such as hair and clothing that deviate significantly from the SMPL surface. Considering the complexity of the human body and the impossibility to obtain extremely accurate SMPL estimation, we repartition all gaussian vertices into " + }, + { + "bbox": [ + 46, + 469, + 289, + 673 + ], + "type": "inline_equation", + "content": "L = 15" + }, + { + "bbox": [ + 46, + 469, + 289, + 673 + ], + "type": "text", + "content": " distinct semantic parts and dynamically adjust the smoothing strength. Besides, we impose the constraint on the anchors points which are closer to the rendering surface than gausicians center. The Semantic Laplacian Regularization (SLR) is defined as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 681, + 287, + 717 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 681, + 287, + 717 + ], + "spans": [ + { + "bbox": [ + 62, + 681, + 287, + 717 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {S L R} = \\sum_ {u \\in \\Omega} \\lambda_ {u} \\left\\| a _ {u, v} - \\frac {1}{| M (u , v) |} \\sum_ {v \\in M (u, v)} a _ {u, v} \\right\\|, \\tag {3}", + "image_path": "a042049e2900a7bf1168ed13d8b74786f2fc7d1aacee4c80d5f4ca4281743764.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 304, + 469, + 547, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 469, + 547, + 625 + ], + "spans": [ + { + "bbox": [ + 304, + 469, + 547, + 625 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 469, + 547, + 625 + ], + "type": "inline_equation", + "content": "\\Omega" + }, + { + "bbox": [ + 304, + 469, + 547, + 625 + ], + "type": "text", + "content": " represents the set of semantic regions, " + }, + { + "bbox": [ + 304, + 469, + 547, + 625 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 304, + 469, + 547, + 625 + ], + "type": "text", + "content": " denotes the surface anchor point of the " + }, + { + "bbox": [ + 304, + 469, + 547, + 625 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 304, + 469, + 547, + 625 + ], + "type": "text", + "content": "-th Gaussian vertex in the " + }, + { + "bbox": [ + 304, + 469, + 547, + 625 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 304, + 469, + 547, + 625 + ], + "type": "text", + "content": "-th semantic region, and " + }, + { + "bbox": [ + 304, + 469, + 547, + 625 + ], + "type": "inline_equation", + "content": "M(u,v)" + }, + { + "bbox": [ + 304, + 469, + 547, + 625 + ], + "type": "text", + "content": " refers to the set of adjacent points of the Gaussian vertex " + }, + { + "bbox": [ + 304, + 469, + 547, + 625 + ], + "type": "inline_equation", + "content": "a_{u,v}" + }, + { + "bbox": [ + 304, + 469, + 547, + 625 + ], + "type": "text", + "content": " located within the " + }, + { + "bbox": [ + 304, + 469, + 547, + 625 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 304, + 469, + 547, + 625 + ], + "type": "text", + "content": "-th semantic region. Experiments demonstrate that the Semantic Laplacian Regularization significantly impacts the handling of more intricate situations, particularly when dealing with errors between the SMPL mesh and the actual character, which can arise from both inaccurate estimations and complex clothing. We apply Laplacians in both position and color spaces; the former enhances the smoothness of the rendering surface, while the latter ensures a more continuous variation in color space." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 305, + 645, + 539, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 645, + 539, + 658 + ], + "spans": [ + { + "bbox": [ + 305, + 645, + 539, + 658 + ], + "type": "text", + "content": "3.2.2 Geometry-Preserving Density Control Strategy" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "content": "The strategies in vanilla 3DGS were not originally build for complex, dynamic structures like the human body. While works [20] have adapted the approach, such as initializing Gaussians based on SMPL meshes, these efforts pri" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "type": "text", + "content": "5575" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 61, + 112, + 531, + 281 + ], + "blocks": [ + { + "bbox": [ + 46, + 70, + 547, + 105 + ], + "lines": [ + { + "bbox": [ + 46, + 70, + 547, + 105 + ], + "spans": [ + { + "bbox": [ + 46, + 70, + 547, + 105 + ], + "type": "text", + "content": "Table 1. Quantitative comparisons with existing methods on UBC fashion dataset [45]. Note that metrics in the table are measure with the background-removed video. Some methods are unable to perform animation, and therefore, these metrics are not available for those methods. Bold indicates the best results, while Underlined represents the second-best results." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 61, + 112, + 531, + 281 + ], + "lines": [ + { + "bbox": [ + 61, + 112, + 531, + 281 + ], + "spans": [ + { + "bbox": [ + 61, + 112, + 531, + 281 + ], + "type": "table", + "html": "
MethodPrior TypePSNR ↑LPIPS ↓FVD ↓Context. ↓CLIP Score ↑Speed↑
DisCo [35] (CVPR24)-30.980.259167.72.13482.4%1FPS
Animate Anyone [8] (CVPR24)-31.330.238148.81.90186.6%2FPS
Champ [51] (ECCV24)-31.210.251155.12.03884.3%1FPS
UniAnimate [36] (ArXiv24)-32.310.211141.31.81187.1%2FPS
SHERF [9] (ICCV23)-27.390.301202.32.99870.1%
SIFU [50] (CVPR24)Camera---2.89172.1%5FPS
GST [26] (ArXiv24)Camera---3.01171.7%50FPS
TeCH [10] (3DV24)Camera29.010.271188.12.33179.3%6FPS
SiTH [7] (CVPR24)Camera---2.77674.7%<1FPS
HumanRef [47] (CVPR24)Camera---2.16781.5%3FPS
SHERT [46] (CVPR2024)Inpainting29.310.266179.82.31978.8%10FPS
UniAnimate + GPGS (ours)Kinematics29.970.251168.71.99282.1%70FPS
KHD (ours)-32.810.221134.41.76892.3%2FPS
SinGS (ours)Kinematics31.330.240156.31.84788.2%70FPS
", + "image_path": "490809444f4498540e92eacd9069a893f3627339c7656ddc5b98ba9e7f2916ce.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 289, + 289, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 289, + 289, + 613 + ], + "spans": [ + { + "bbox": [ + 46, + 289, + 289, + 613 + ], + "type": "text", + "content": "marily lay a solid foundation for the optimization process but fail to maintain this advantages throughout. Incompatibilities still arise in optimization stage, prompting us to propose a novel density control scheme that is inherently compatible with hybrid representations, designed for adaptive adjustment body-wide density. This scheme draws on mesh operation algorithms [6] to better address the unique challenges of human body. For primitives that locate in under-expression mesh areas, we subdivide these regions by adding gaussians on every edge of the selected geometric triangles, which creates much smaller triangles. And we skip the trivial post-smoothing for vertices as the position of Gaussians will continue to be optimized in the network. For candidate Gaussians that are selected due to excessively large scales, as well as all newly added Gaussians, we perform rescaling to adjust them to an average size. In the case of anomalous gaussians, especially those are almost transparent, we collapse relevant edges and replace them with single vertex, resulting in decimated adjacent faces. And the priority of selected edges is decided by gaussian attributes and edge length. This entire pruning process preserves the principal structure of the human body. Our strategy result in better reconstruction and animation as the optimization was conducted under connectivity and compactness of the entire human body rather than producing holes and stacks. Moreover, this approach synergistically enhances the effectiveness of our hybrid representation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 625, + 128, + 639 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 625, + 128, + 639 + ], + "spans": [ + { + "bbox": [ + 47, + 625, + 128, + 639 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 646, + 164, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 646, + 164, + 659 + ], + "spans": [ + { + "bbox": [ + 47, + 646, + 164, + 659 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 288, + 715 + ], + "type": "text", + "content": "Implementation Details. Our KHD primarily utilizes the RotationHuman dataset we collected, which focuses on human pose image sequences involving turning motions. The specific structure of the 3D Pose Encoder is provided in" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 289, + 545, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 289, + 545, + 503 + ], + "spans": [ + { + "bbox": [ + 304, + 289, + 545, + 503 + ], + "type": "text", + "content": "the supplementary materials. KHD was trained using 8 NVIDIA A100 80GB GPUs, with DeepSpeed and ZeRO2 employed to reduce memory usage. Similar toAnimate Anyone, we resized the video frames to a spatial dimension of " + }, + { + "bbox": [ + 304, + 289, + 545, + 503 + ], + "type": "inline_equation", + "content": "768 \\times 512" + }, + { + "bbox": [ + 304, + 289, + 545, + 503 + ], + "type": "text", + "content": ". We used a context window size of 64 frames, and to enhance model scalability post-training, we implemented a random context window jitter strategy [36]. During training, the model randomly samples window sizes from 64 to 32 frames in an uniformly distributed manner. The training process utilized the AdamW optimizer [15], with a learning rate set to " + }, + { + "bbox": [ + 304, + 289, + 545, + 503 + ], + "type": "inline_equation", + "content": "5e - 5" + }, + { + "bbox": [ + 304, + 289, + 545, + 503 + ], + "type": "text", + "content": " and DDPM with 1000 steps for noise sampling. During inference, DDIM with 30 steps was used to reduce inference time. For the 3DGS component, the Adam optimizer [15] was used with a learning rate of " + }, + { + "bbox": [ + 304, + 289, + 545, + 503 + ], + "type": "inline_equation", + "content": "1e - 3" + }, + { + "bbox": [ + 304, + 289, + 545, + 503 + ], + "type": "text", + "content": " and a cosine annealing learning scheduler was applied to accelerate convergence. The input image sequence size for 3DGS was also " + }, + { + "bbox": [ + 304, + 289, + 545, + 503 + ], + "type": "inline_equation", + "content": "768 \\times 512" + }, + { + "bbox": [ + 304, + 289, + 545, + 503 + ], + "type": "text", + "content": ". 3DGS was trained using a single NVIDIA 4090 24GB GPU." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 504, + 545, + 600 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 504, + 545, + 600 + ], + "spans": [ + { + "bbox": [ + 304, + 504, + 545, + 600 + ], + "type": "text", + "content": "Baseline Models. Our baseline models consist of two parts: the first part includes human image animation models, such as DisCo [35], Animation Anyone [8] " + }, + { + "bbox": [ + 304, + 504, + 545, + 600 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 304, + 504, + 545, + 600 + ], + "type": "text", + "content": ", Champ [51], and UniAnimate [36]. These models are used to compare our animation capability and the superiority of KHD. The second part consists of single-image 3D avatar reconstruction models, including SHERF [9], SIFU [50], GST [26], TeCH [10], SiTH [7], HumanRef [47], and SHERT [46]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 600, + 545, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 600, + 545, + 696 + ], + "spans": [ + { + "bbox": [ + 304, + 600, + 545, + 696 + ], + "type": "text", + "content": "Metrics. We employ both PSNR and LPIPS to assess the quality of video reconstruction. To evaluate the animation capability of our model, we also use FVD [34] to measure the distribution discrepancy between the generated video and the original video, assessing the video generation performance. Furthermore, following HumanRef, we introduce contextual distance [19] and CLIP score [27] to evaluate the texture and semantic similarity between the rendered" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 317, + 702, + 509, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 702, + 509, + 714 + ], + "spans": [ + { + "bbox": [ + 317, + 702, + 509, + 714 + ], + "type": "text", + "content": "1https://github.com/MooreThreads/Moore-AnimateAnyone" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "type": "text", + "content": "5576" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 71, + 282, + 228 + ], + "blocks": [ + { + "bbox": [ + 57, + 71, + 282, + 228 + ], + "lines": [ + { + "bbox": [ + 57, + 71, + 282, + 228 + ], + "spans": [ + { + "bbox": [ + 57, + 71, + 282, + 228 + ], + "type": "image", + "image_path": "3a42d13bfec4a738751303ba79faedc3dfb2230b70f53d76348c9b242b679bee.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 234, + 287, + 257 + ], + "lines": [ + { + "bbox": [ + 46, + 234, + 287, + 257 + ], + "spans": [ + { + "bbox": [ + 46, + 234, + 287, + 257 + ], + "type": "text", + "content": "Figure 4. Comparison between KHD and other SOTA human image animation methods. Please zoom in for better observation." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 268, + 200, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 268, + 200, + 281 + ], + "spans": [ + { + "bbox": [ + 47, + 268, + 200, + 281 + ], + "type": "text", + "content": "output and the input reference image." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 289, + 165, + 303 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 289, + 165, + 303 + ], + "spans": [ + { + "bbox": [ + 47, + 289, + 165, + 303 + ], + "type": "text", + "content": "4.2. Qualitative Analysis" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 308, + 287, + 500 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 308, + 287, + 500 + ], + "spans": [ + { + "bbox": [ + 46, + 308, + 287, + 500 + ], + "type": "text", + "content": "Singe-Image Reconstruction. We compare SinGS with two methods from the literature that claim to support animation (TeCH, SiTH). It is worth noting that during our experiments, we found that single-image animation with these methods was extremely difficult to execute, and in some cases, nearly impossible. We had to use third-party code " + }, + { + "bbox": [ + 46, + 308, + 287, + 500 + ], + "type": "inline_equation", + "content": "{}^{2}" + }, + { + "bbox": [ + 46, + 308, + 287, + 500 + ], + "type": "text", + "content": " to achieve the reposing for some of these methods. We animate their reconstructed outputs to a common pose (A Pose) for comparison, which allows us to evaluate both reconstruction and animation capabilities. The comparison results are shown in Figure 3, SinGS generates results with better details and maintains high-quality animation, particularly on the in-the-wild data, where SinGS's visual performance significantly outperforms the other methods. This demonstrates the high feasibility and effectiveness of the kinematic prior." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 501, + 288, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 501, + 288, + 693 + ], + "spans": [ + { + "bbox": [ + 46, + 501, + 288, + 693 + ], + "type": "text", + "content": "Kinematic Human Diffusion. KHD, as the core component of our pipeline, is compared with state-of-the-art human image animation methods, includingAnimate Anyone and UniAnimate. As shown in Figure 4, we drive the same character with all three methods, focusing on comparing the quality of their generated results, particularly for the turning pose. From the figure, it is clear that both UniAnimate and our KHD maintain higher consistency thanAnimate Anyone after the character turns, with KHD surpassing UniAnimate in detail. The details marked by the red arrows in Figure 4 reveals that UniAnimate generates incorrect details, a common issue when using 2D poses, as they often struggle to distinguish body orientation during back-facing poses, leading the model to generate erroneous images. This is detrimental to reconstruction, while KHD generates videos with much higher consistency." + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 317, + 69, + 535, + 144 + ], + "blocks": [ + { + "bbox": [ + 317, + 69, + 535, + 144 + ], + "lines": [ + { + "bbox": [ + 317, + 69, + 535, + 144 + ], + "spans": [ + { + "bbox": [ + 317, + 69, + 535, + 144 + ], + "type": "image", + "image_path": "8571cb3aae5ca73f739c8b03c6c01711d6540f1a88c71bf7ebf6b1602c55dd38.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 152, + 547, + 208 + ], + "lines": [ + { + "bbox": [ + 305, + 152, + 547, + 208 + ], + "spans": [ + { + "bbox": [ + 305, + 152, + 547, + 208 + ], + "type": "text", + "content": "Figure 5. Fifty users rated the video quality, with UniAnimate scoring highest, followed closely by our model, which outperformed others. Our model is more efficient, using only 3.5 GB of memory and achieving 70 FPS, while UniAnimate requires over 12GB and 0.5 seconds per image." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 218, + 547, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 218, + 547, + 422 + ], + "spans": [ + { + "bbox": [ + 304, + 218, + 547, + 422 + ], + "type": "text", + "content": "Geometry-Preserving 3D Gaussian Splatting. Although KHD provides high-consistency pose sampling results, it inevitably has some consistency flaws, meaning it can only offer imperfect prior information. Our geometry-preserving GS method is specifically designed to address this challenge. To validate the superiority of our Geometry-Preserving 3D Gaussian Splatting (GPGS) over general human modeling methods, we compared it with recent state-of-the-art (i.e., SOTA) methods, HAHA [32] and HUGS [16]. We generate a pose sequence using KHD and then performed reconstruction with all 3 methods. The reconstruction results are shown in Figure 7. As seen, both HAHA and HUGS fail to produce a compact 3D avatar, with significant holes or floating artifacts appearing in various areas, particularly during the turning motion. In contrast, our GPGS method is able to reconstruct a compact, watertight 3D avatar model." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 431, + 430, + 444 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 431, + 430, + 444 + ], + "spans": [ + { + "bbox": [ + 306, + 431, + 430, + 444 + ], + "type": "text", + "content": "4.3. Quantitative Analysis" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 450, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 450, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 450, + 547, + 715 + ], + "type": "text", + "content": "We selected 100 test videos from the UBC dataset for our quantitative analysis. For each video, we chose the first frame as the reference image and used the video's pose sequence as the driving signal. The performance of all methods was assessed by comparing the animation results with the original video to evaluate the differences. Video reconstruction metrics (including PSNR, LPIPS, and FVD) reflect the model's animation capabilities, while Contextual Distance and CLIP score allow for comparison with methods that do not support animation. The quantitative comparison results are shown in Table 1. Human image animation methods (i.e., the second to fifth rows in the table, including our KHD) outperform the reconstruction-based methods across the board. Among them, KHD surpasses the other methods in most metrics, and UniAnimate also demonstrates superior performance, which is expected due to the scale of training data and model size. While SinGS lags behind UniAnimate, it shows a significant advantage over other methods in some metrics. However, SinGS significantly outperforms diffusion-based animation methods in terms of inference efficiency and resource consumption. Our SinGS achieves 70 FPS during inference, with only" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 58, + 702, + 231, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 702, + 231, + 713 + ], + "spans": [ + { + "bbox": [ + 58, + 702, + 231, + 713 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 58, + 702, + 231, + 713 + ], + "type": "text", + "content": "https://github.com/custom-humans/editable-humans" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5577" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 69, + 112, + 264, + 173 + ], + "blocks": [ + { + "bbox": [ + 47, + 71, + 287, + 104 + ], + "lines": [ + { + "bbox": [ + 47, + 71, + 287, + 104 + ], + "spans": [ + { + "bbox": [ + 47, + 71, + 287, + 104 + ], + "type": "text", + "content": "Table 2. The results of the ablation experiments on Semantic Laplacian Regularization (i.e., Semantic Lap. Reg.) and Geometry-Preserving Density Control (i.e., GPDC) in SinGS." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 69, + 112, + 264, + 173 + ], + "lines": [ + { + "bbox": [ + 69, + 112, + 264, + 173 + ], + "spans": [ + { + "bbox": [ + 69, + 112, + 264, + 173 + ], + "type": "table", + "html": "
SettingPSNRLPIPSFVD
with Semantic Lap. Reg.31.330.240156.3
with KNN Lap. Reg.30.880.250169.9
w/o Lap. Reg.30.340.272181.3
with GPDC31.010.249166.1
naive Density Control [14]30.220.276190.2
", + "image_path": "0a5847d3fcf1cdcf6872f676ffbe0506b521a418c0fb7b16bc3e7fa5705b251f.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 182, + 286, + 278 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 182, + 286, + 278 + ], + "spans": [ + { + "bbox": [ + 46, + 182, + 286, + 278 + ], + "type": "text", + "content": "3.8 GB of memory usage (tested on a single NVIDIA 4090 GPU), while UniAnimate, Champ, andAnimate Anyone each require more than 0.4 seconds per image and consume over 12 GB of memory. These quantitative results strongly demonstrate that SinGS is highly effective and feasible. We believe that using KHD, a specialized foundational model, will greatly enhance 3D reconstruction performance while significantly reducing the stringent data requirements." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 278, + 286, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 278, + 286, + 327 + ], + "spans": [ + { + "bbox": [ + 46, + 278, + 286, + 327 + ], + "type": "text", + "content": "User Study. We invited 50 users to subjectively evaluate the animation results of SinGS. The evaluation was based on a scale from 0 to 10, with 10 being the highest score and 0 being the lowest. The results are visualized in Figure 5." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 335, + 140, + 348 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 335, + 140, + 348 + ], + "spans": [ + { + "bbox": [ + 47, + 335, + 140, + 348 + ], + "type": "text", + "content": "4.4. Ablation Study" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 354, + 286, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 354, + 286, + 628 + ], + "spans": [ + { + "bbox": [ + 46, + 354, + 286, + 628 + ], + "type": "text", + "content": "Kinematic Priors Analysis. The number of poses sampled by KHD has a significant impact on the 3D distillation process. We analyzed how the number of poses affects the animation capability of the final reconstructed model, and the results are shown in Figure 6. We incrementally increased the number of poses from 2 (which corresponds to the typical approach of generating front and rear views, as used by most methods) up to 150. We also tested the influence of different prior models (i.e., KHD, UniAnimate, and Animate Anyone) on the results. Additionally, we examined how different 3DGS reconstruction methods handle the use of prior knowledge. As shown, performance improves rapidly with an increase in the number of poses, but the rate of improvement slows significantly after reaching around 100 poses. Moore-AA + HAHA shows the slowest improvement, while KHD + GPGS exhibits the slowest decline in performance. This suggests that increasing the number of poses does not always lead to improved performance for two main reasons: first, the prior models have already provided sufficient information; and second, none of the prior models provide perfectly consistent priors. Excessive sampling can introduce a lot of inconsistent noise, which ultimately harms performance." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 630, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 286, + 713 + ], + "type": "text", + "content": "Geometry-Preserving 3D Gaussian Splitting Analysis. We conducted ablation experiments on the two key designs in GPGS: Semantic Laplacian Regularization and GPDC, with the results shown in Table 2. By replacing and removing the Laplacian regularization, we observed that Semantic Laplacian consistently outperforms in terms of reconstruction metrics. Additionally, GPDC demonstrates a perfor" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 331, + 70, + 521, + 179 + ], + "blocks": [ + { + "bbox": [ + 331, + 70, + 521, + 179 + ], + "lines": [ + { + "bbox": [ + 331, + 70, + 521, + 179 + ], + "spans": [ + { + "bbox": [ + 331, + 70, + 521, + 179 + ], + "type": "image", + "image_path": "b842dcf943409a52bdf762189b7e88ee69294668cd25f9d0b1b64457ba6422f0.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 188, + 545, + 210 + ], + "lines": [ + { + "bbox": [ + 306, + 188, + 545, + 210 + ], + "spans": [ + { + "bbox": [ + 306, + 188, + 545, + 210 + ], + "type": "text", + "content": "Figure 6. The comparison of the impact of augmented pose numbers on the performance of the 3D avatar models." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 307, + 232, + 542, + 364 + ], + "blocks": [ + { + "bbox": [ + 307, + 232, + 542, + 364 + ], + "lines": [ + { + "bbox": [ + 307, + 232, + 542, + 364 + ], + "spans": [ + { + "bbox": [ + 307, + 232, + 542, + 364 + ], + "type": "image", + "image_path": "b6dc8cd903053ead99689f69599e83ad168ab21e6ef0529d1e35b52fc60f9414.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 374, + 545, + 408 + ], + "lines": [ + { + "bbox": [ + 305, + 374, + 545, + 408 + ], + "spans": [ + { + "bbox": [ + 305, + 374, + 545, + 408 + ], + "type": "text", + "content": "Figure 7. Here is a comparison of our GPGS with two SOTA 3DGS avatar methods, HAHA [32] and HUGS [16]. All methods use images sampled by KHD as body prior." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 422, + 545, + 481 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 422, + 545, + 481 + ], + "spans": [ + { + "bbox": [ + 305, + 422, + 545, + 481 + ], + "type": "text", + "content": "mance advantage over the original density control strategy. The original strategy does not account for dynamic properties and geometric constraints, which makes it ineffective in addressing issues like floating and holes caused by data inconsistencies." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 499, + 378, + 511 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 499, + 378, + 511 + ], + "spans": [ + { + "bbox": [ + 306, + 499, + 378, + 511 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 522, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 522, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 522, + 545, + 713 + ], + "type": "text", + "content": "In this paper, we propose a new paradigm for single-image 3D avatar reconstruction. Unlike traditional methods that complete missing information in camera space, we sample in human kinematic space, enabling efficient capture of both viewpoints and motion traits. To achieve this, we introduce Kinematic Human Diffusion (KHD) for consistent pose-space generation, supported by our RotationHuman dataset, which focuses on turning motions. Additionally, we propose Geometry-Preserving 3D Gaussian Splating that reconstructs high-quality, animatable 3D avatars under imperfect human priors. This includes two key innovations: Semantic Laplacian regularization for compact, detailed constraints, and a geometry-preserving density control strategy to maintain stable human geometry. Together, these innovations allow SinGS to achieve high-quality, animatable 3D avatar reconstruction from a single image." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "5578" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 151, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 151, + 85 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 151, + 85 + ], + "type": "text", + "content": "6. Acknowledgment" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 91, + 287, + 162 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 91, + 287, + 162 + ], + "spans": [ + { + "bbox": [ + 47, + 91, + 287, + 162 + ], + "type": "text", + "content": "This work supported by Shanghai Post-doctoral Excellence Programe (2024290). This work was also partly by Shanghai Municipal Bureau of Culture and Tourism and Meta-verse Technology Innovation Center of Shanghai Municipal Bureau of Culture and Tourism. Kairui Feng thanks to Xiaomi foundation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 172, + 106, + 185 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 172, + 106, + 185 + ], + "spans": [ + { + "bbox": [ + 48, + 172, + 106, + 185 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 193, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 193, + 287, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 193, + 287, + 246 + ], + "spans": [ + { + "bbox": [ + 53, + 193, + 287, + 246 + ], + "type": "text", + "content": "[1] Ankan Kumar Bhunia, Salman H. Khan, Hisham Cholakkal, Rao Muhammad Anwer, Jorma Laaksonen, Mubarak Shah, and Fahad Shahbaz Khan. Person image synthesis via denoising diffusion model. In CVPR 2023, pages 5968-5976. IEEE, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 248, + 288, + 291 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 248, + 288, + 291 + ], + "spans": [ + { + "bbox": [ + 53, + 248, + 288, + 291 + ], + "type": "text", + "content": "[2] Zhe Cao, Gines Hidalgo, Tomas Simon, Shih-En Wei, and Yaser Sheikh. Openpose: Realtime multi-person 2d pose estimation using part affinity fields. IEEE Trans. Pattern Anal. Mach. Intell., 43(1):172-186, 2021. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 293, + 287, + 345 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 293, + 287, + 345 + ], + "spans": [ + { + "bbox": [ + 53, + 293, + 287, + 345 + ], + "type": "text", + "content": "[3] Kennard Yanting Chan, Fayao Liu, Guosheng Lin, Chuan Sheng Foo, and Weisi Lin. R-cyclic diffuser: Reductive and cyclic latent diffusion for 3d clothed human digitalization. In CVPR 2024, pages 10304-10313. IEEE, 2024. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 347, + 287, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 347, + 287, + 402 + ], + "spans": [ + { + "bbox": [ + 53, + 347, + 287, + 402 + ], + "type": "text", + "content": "[4] Di Chang, Yichun Shi, Quankai Gao, Hongyi Xu, Jessica Fu, Guoxian Song, Qing Yan, Yizhe Zhu, Xiao Yang, and Mohammad Soleymani. Magicpose: Realistic human poses and facial expressions retargeting with identity-aware diffusion. In ICML 2024. OpenReview.net, 2024. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 403, + 287, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 403, + 287, + 446 + ], + "spans": [ + { + "bbox": [ + 53, + 403, + 287, + 446 + ], + "type": "text", + "content": "[5] Mingjin Chen, Junhao Chen, Xiaojun Ye, Huan-ang Gao, Xiaoxue Chen, Zhaoxin Fan, and Hao Zhao. Ultraman: Single image 3d human reconstruction with ultra speed and detail. CoRR, abs/2403.12028, 2024. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 447, + 287, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 447, + 287, + 479 + ], + "spans": [ + { + "bbox": [ + 53, + 447, + 287, + 479 + ], + "type": "text", + "content": "[6] Michael Garland and Paul S. Heckbert. Surface simplification using quadric error metrics. In SIGGRAPH 1997, pages 209-216. ACM, 1997. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 481, + 287, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 481, + 287, + 523 + ], + "spans": [ + { + "bbox": [ + 53, + 481, + 287, + 523 + ], + "type": "text", + "content": "[7] Hsuan-I Ho, Jie Song, and Otmar Hilliges. Sith: Single-view textured human reconstruction with image-conditioned diffusion. In CVPR 2024, pages 538-549. IEEE, 2024. 1, 2, 3, 5, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 525, + 287, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 525, + 287, + 557 + ], + "spans": [ + { + "bbox": [ + 53, + 525, + 287, + 557 + ], + "type": "text", + "content": "[8] Li Hu. Animate anyone: Consistent and controllable imagery-to-video synthesis for character animation. In CVPR 2024, pages 8153-8163. IEEE, 2024. 2, 3, 4, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 558, + 287, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 558, + 287, + 601 + ], + "spans": [ + { + "bbox": [ + 53, + 558, + 287, + 601 + ], + "type": "text", + "content": "[9] Shoukang Hu, Fangzhou Hong, Liang Pan, Haiyi Mei, Lei Yang, and Ziwei Liu. SHERF: generalizable human nef from a single image. In ICCV 2023, pages 9318-9330. IEEE, 2023. 1, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 602, + 287, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 602, + 287, + 646 + ], + "spans": [ + { + "bbox": [ + 48, + 602, + 287, + 646 + ], + "type": "text", + "content": "[10] Yangyi Huang, Hongwei Yi, Yuliang Xiu, Tingting Liao, Ji-axiang Tang, Deng Cai, and Justus Thies. Tech: Text-guided reconstruction of lifelike clothed humans. In 3DV 2024, pages 1531–1542. IEEE, 2024. 1, 2, 3, 5, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 689 + ], + "type": "text", + "content": "[11] Zeng Huang, Yuanlu Xu, Christoph Lassner, Hao Li, and Tony Tung. ARCH: animatable reconstruction of clothed humans. In CVPR 2020, pages 3090–3099. Computer Vision Foundation / IEEE, 2020. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "text", + "content": "[12] Yasamin Jafarian and Hyun Soo Park. Learning high fidelity depths of dressed humans by watching social media dance" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "videos. In CVPR 2021, pages 12753-12762. Computer Vision Foundation / IEEE, 2021. 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 96, + 545, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 545, + 140 + ], + "type": "text", + "content": "[13] Johanna Karras, Aleksander Holynski, Ting-Chun Wang, and Ira Kemelmacher-Shlizerman. Dreampose: Fashion image-to-video synthesis via stable diffusion. In ICCV 2023, pages 22623-22633. IEEE, 2023. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 141, + 545, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 141, + 545, + 184 + ], + "spans": [ + { + "bbox": [ + 308, + 141, + 545, + 184 + ], + "type": "text", + "content": "[14] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4):139:1-139:14, 2023. 1, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 186, + 545, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 186, + 545, + 207 + ], + "spans": [ + { + "bbox": [ + 308, + 186, + 545, + 207 + ], + "type": "text", + "content": "[15] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR 2015, 2015. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 209, + 545, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 209, + 545, + 241 + ], + "spans": [ + { + "bbox": [ + 308, + 209, + 545, + 241 + ], + "type": "text", + "content": "[16] Muhammed Kocabas, Jen-Hao Rick Chang, James Gabriel, Oncel Tuzel, and Anurag Ranjan. HUGS: human gaussian splats. In CVPR 2024, pages 505-515. IEEE, 2024. 4, 7, 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 243, + 545, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 243, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 308, + 243, + 545, + 285 + ], + "type": "text", + "content": "[17] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3d object. In ICCV 2023, pages 9264–9275. IEEE, 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 288, + 545, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 288, + 545, + 330 + ], + "spans": [ + { + "bbox": [ + 308, + 288, + 545, + 330 + ], + "type": "text", + "content": "[18] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J. Black. SMPL: a skinned multiperson linear model. ACM Trans. Graph., 34(6):248:1-248:16, 2015. 2, 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 332, + 545, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 332, + 545, + 365 + ], + "spans": [ + { + "bbox": [ + 308, + 332, + 545, + 365 + ], + "type": "text", + "content": "[19] Roey Mechrez, Itamar Talmi, and Lihi Zelnik-Manor. The contextual loss for image transformation with non-aligned data. In ECCV 2018, pages 800-815. Springer, 2018. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 367, + 545, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 367, + 545, + 399 + ], + "spans": [ + { + "bbox": [ + 308, + 367, + 545, + 399 + ], + "type": "text", + "content": "[20] Gyeongsik Moon, Takaaki Shiratori, and Shunsuke Saito. Expressive whole-body 3d gaussian avatar. CoRR, abs/2407.21686, 2024. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 400, + 545, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 400, + 545, + 443 + ], + "spans": [ + { + "bbox": [ + 308, + 400, + 545, + 443 + ], + "type": "text", + "content": "[21] Ryota Natsume, Shunsuke Saito, Zeng Huang, Weikai Chen, Chongyang Ma, Hao Li, and Shigeo Morishima. Siclope: Silhouette-based clothed people. In CVPR 2019, pages 4480-4490. Computer Vision Foundation / IEEE, 2019. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 445, + 545, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 445, + 545, + 487 + ], + "spans": [ + { + "bbox": [ + 308, + 445, + 545, + 487 + ], + "type": "text", + "content": "[22] Haomiao Ni, Changhao Shi, Kai Li, Sharon X. Huang, and Martin Renqiang Min. Conditional image-to-video generation with latent flow diffusion models. In CVPR 2023, pages 18444-18455. IEEE, 2023. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 490, + 545, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 490, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 308, + 490, + 545, + 544 + ], + "type": "text", + "content": "[23] Panwang Pan, Zhuo Su, Chenguo Lin, Zhen Fan, Yongjie Zhang, Zeming Li, Tingting Shen, Yadong Mu, and Yebin Liu. Humansplat: Generalizable single-image human gaussian splatting with structure priors. CoRR, abs/2406.12459, 2024. 1, 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 545, + 545, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 545, + 545, + 610 + ], + "spans": [ + { + "bbox": [ + 308, + 545, + 545, + 610 + ], + "type": "text", + "content": "[24] Marco Pesavento, Yuanlu Xu, Nikolaos Sarafianos, Robert Maier, Ziyan Wang, Chun-Han Yao, Marco Volino, Edmond Boyer, Adrian Hilton, and Tony Tung. ANIM: accurate neural implicit model for human reconstruction from a single RGB-D image. In CVPR 2024, pages 5448-5458. IEEE, 2024. 1" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 613, + 545, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 613, + 545, + 645 + ], + "spans": [ + { + "bbox": [ + 308, + 613, + 545, + 645 + ], + "type": "text", + "content": "[25] Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. In ICLR 2023. OpenReview.net, 2023. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 647, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 647, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 308, + 647, + 545, + 689 + ], + "type": "text", + "content": "[26] Lorenza Prospero, Abdullah Hamdi, João F. Henriques, and Christian Rupprecht. GST: precise 3d human body from a single image with gaussian splatting transformers. CoRR, abs/2409.04196, 2024. 1, 2, 6" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 691, + 545, + 713 + ], + "type": "text", + "content": "[27] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry," + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "5579" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 67, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 287, + 116 + ], + "type": "text", + "content": "Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML 2021, pages 8748-8763. PMLR, 2021. 3, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 118, + 287, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 118, + 287, + 193 + ], + "spans": [ + { + "bbox": [ + 48, + 118, + 287, + 193 + ], + "type": "text", + "content": "[28] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, Eric Mintun, Junting Pan, Kalyan Vasudev Alwala, Nicolas Carion, Chao-Yuan Wu, Ross B. Girshick, Piotr Dollar, and Christoph Feichtenhofer. SAM 2: Segment anything in images and videos. CoRR, abs/2408.00714, 2024. 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 49, + 195, + 287, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 195, + 287, + 237 + ], + "spans": [ + { + "bbox": [ + 49, + 195, + 287, + 237 + ], + "type": "text", + "content": "[29] Shunsuke Saito, Zeng Huang, Ryota Natsume, Shigeo Morishima, Hao Li, and Angjoo Kanazawa. Pifu: Pixel-aligned implicit function for high-resolution clothed human digitization. In ICCV 2019, pages 2304-2314. IEEE, 2019. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 49, + 239, + 287, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 239, + 287, + 282 + ], + "spans": [ + { + "bbox": [ + 49, + 239, + 287, + 282 + ], + "type": "text", + "content": "[30] Shunsuke Saito, Tomas Simon, Jason M. Saragih, and Hanbyul Joo. Pifuhd: Multi-level pixel-aligned implicit function for high-resolution 3d human digitization. In CVPR 2020, pages 81–90. Computer Vision Foundation / IEEE, 2020. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 49, + 284, + 287, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 284, + 287, + 315 + ], + "spans": [ + { + "bbox": [ + 49, + 284, + 287, + 315 + ], + "type": "text", + "content": "[31] Anastasis Stathopoulos, Ligong Han, and Dimitris N. Metaxas. Score-guided diffusion for 3d human recovery. In CVPR 2024, pages 906-915. IEEE, 2024. 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 49, + 317, + 287, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 317, + 287, + 358 + ], + "spans": [ + { + "bbox": [ + 49, + 317, + 287, + 358 + ], + "type": "text", + "content": "[32] David Svitov, Pietro Morerio, Lourdes Agapito, and Alessio Del Bue. HAHA: highly articulated gaussian human avatars with textured mesh prior. CoRR, abs/2404.01053, 2024. 4, 7, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 49, + 361, + 287, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 361, + 287, + 403 + ], + "spans": [ + { + "bbox": [ + 49, + 361, + 287, + 403 + ], + "type": "text", + "content": "[33] Linrui Tian, Qi Wang, Bang Zhang, and Liefeng Bo. EMO: emote portrait alive - generating expressive portrait videos with audio2video diffusion model under weak conditions. CoRR, abs/2402.17485, 2024. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 49, + 405, + 287, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 405, + 287, + 448 + ], + "spans": [ + { + "bbox": [ + 49, + 405, + 287, + 448 + ], + "type": "text", + "content": "[34] Thomas Unterthiner, Sjoerd van Steenkiste, Karol Kurach, Raphaël Marinier, Marcin Michalski, and Sylvain Gelly. Towards accurate generative models of video: A new metric & challenges. CoRR, abs/1812.01717, 2018. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 49, + 449, + 287, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 449, + 287, + 502 + ], + "spans": [ + { + "bbox": [ + 49, + 449, + 287, + 502 + ], + "type": "text", + "content": "[35] Tan Wang, Linjie Li, Kevin Lin, Chung-Ching Lin, Zhengyuan Yang, Hanwang Zhang, Zicheng Liu, and Lijuan Wang. Disco: Disentangled control for referring human dance generation in real world. CoRR, abs/2307.00040, 2023. 2, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 49, + 505, + 287, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 505, + 287, + 557 + ], + "spans": [ + { + "bbox": [ + 49, + 505, + 287, + 557 + ], + "type": "text", + "content": "[36] Xiang Wang, Shiwei Zhang, Changxin Gao, Jiayu Wang, Xiaogiang Zhou, Yingya Zhang, Luxin Yan, and Nong Sang. Unianimate: Taming unified video diffusion models for consistent human image animation. CoRR, abs/2406.01188, 2024. 3, 4, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 49, + 559, + 287, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 559, + 287, + 612 + ], + "spans": [ + { + "bbox": [ + 49, + 559, + 287, + 612 + ], + "type": "text", + "content": "[37] Xiang Wang, Shiwei Zhang, Hangjie Yuan, Zhiwu Qing, Biao Gong, Yingya Zhang, Yujun Shen, Changxin Gao, and Nong Sang. A recipe for scaling up text-to-video generation with text-free videos. In CVPR 2024, pages 6572-6582. IEEE, 2024. 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 49, + 615, + 287, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 615, + 287, + 646 + ], + "spans": [ + { + "bbox": [ + 49, + 615, + 287, + 646 + ], + "type": "text", + "content": "[38] Yaohui Wang, Xin Ma, Xinyuan Chen, Antitza Dantcheva, Bo Dai, and Yu Qiao. LEO: generative latent image animatior for human video synthesis. CoRR, abs/2305.03989, 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 49, + 648, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 648, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 49, + 648, + 287, + 712 + ], + "type": "text", + "content": "[39] Zhangyang Xiong, Chenghong Li, Kenkun Liu, Hongjie Liao, Jianqiao Hu, Junyi Zhu, Shuliang Ning, Lingteng Qiu, Chongjie Wang, Shijie Wang, Shuguang Cui, and Xiaoguang Han. Mvhumannet: A large-scale dataset of multi-view daily dressing human captures. In CVPR 2024, pages 19801-19811. IEEE, 2024. 4" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 632 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 116 + ], + "type": "text", + "content": "[40] Yuliang Xiu, Jinlong Yang, Dimitrios Tzionas, and Michael J. Black. ICON: Implicit Clothed humans Obtained from Normals. In CVPR 2022, pages 13296-13306, 2022. 1, 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 118, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 150 + ], + "type": "text", + "content": "[41] Yuliang Xiu, Jinlong Yang, Xu Cao, Dimitrios Tzionas, and Michael J. Black. ECON: Explicit Clothed humans Optimized via Normal integration. In CVPR 2023, 2023. 1, 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 152, + 545, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 205 + ], + "type": "text", + "content": "[42] Zhongcong Xu, Jianfeng Zhang, Jun Hao Liew, Hanshu Yan, Jia-Wei Liu, Chenxu Zhang, Jiashi Feng, and Mike Zheng Shou. Magicanimate: Temporally consistent human image animation using diffusion model. In CVPR 2024, pages 1481–1490. IEEE, 2024. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 208, + 545, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 208, + 545, + 250 + ], + "spans": [ + { + "bbox": [ + 307, + 208, + 545, + 250 + ], + "type": "text", + "content": "[43] Zhendong Yang, Ailing Zeng, Chun Yuan, and Yu Li. Effective whole-body pose estimation with two-stages distillation. In ICCV 2023 - Workshops, pages 4212-4222. IEEE, 2023. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 253, + 545, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 253, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 307, + 253, + 545, + 285 + ], + "type": "text", + "content": "[44] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. CoRR, abs/2308.06721, 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 287, + 545, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 287, + 545, + 329 + ], + "spans": [ + { + "bbox": [ + 307, + 287, + 545, + 329 + ], + "type": "text", + "content": "[45] Polina Zablotskaia, Aliaksandr Siarohin, Bo Zhao, and Leonid Sigal. Dwnet: Dense warp-based network for pose-guided human video generation. In BMVC 2019, page 51. BMVA Press, 2019. 4, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 331, + 545, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 331, + 545, + 374 + ], + "spans": [ + { + "bbox": [ + 307, + 331, + 545, + 374 + ], + "type": "text", + "content": "[46] Xiaoyu Zhan, Jianxin Yang, Yuanqi Li, Jie Guo, Yanwen Guo, and Wenping Wang. Semantic human mesh reconstruction with textures. In CVPR 2024, pages 142-152. IEEE, 2024. 1, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 376, + 545, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 376, + 545, + 418 + ], + "spans": [ + { + "bbox": [ + 307, + 376, + 545, + 418 + ], + "type": "text", + "content": "[47] Jingbo Zhang, Xiaoyu Li, Qi Zhang, Yanpei Cao, Ying Shan, and Jing Liao. Humanref: Single image to 3d human generation via reference-guided diffusion. In CVPR 2024, pages 1844-1854. IEEE, 2024. 2, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 421, + 545, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 421, + 545, + 453 + ], + "spans": [ + { + "bbox": [ + 307, + 421, + 545, + 453 + ], + "type": "text", + "content": "[48] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV 2023, pages 3813-3824. IEEE, 2023. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 455, + 545, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 455, + 545, + 486 + ], + "spans": [ + { + "bbox": [ + 307, + 455, + 545, + 486 + ], + "type": "text", + "content": "[49] Zechuan Zhang, Li Sun, Zongxin Yang, Ling Chen, and Yi Yang. Global-correlated 3d-decoupling transformer for clothed avatar reconstruction. In NeurIPS 2023, 2023. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 489, + 545, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 489, + 545, + 531 + ], + "spans": [ + { + "bbox": [ + 307, + 489, + 545, + 531 + ], + "type": "text", + "content": "[50] Zechuan Zhang, Zongxin Yang, and Yi Yang. SIFU: side-view conditioned implicit function for real-world usable clothed human reconstruction. In CVPR 2024, pages 9936-9947. IEEE, 2024. 1, 2, 3, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 533, + 545, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 533, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 307, + 533, + 545, + 578 + ], + "type": "text", + "content": "[51] Shenhao Zhu, Junming Leo Chen, Zuozhuo Dai, Yinghui Xu, Xun Cao, Yao Yao, Hao Zhu, and Siyu Zhu. Champ: Controllable and consistent human image animation with 3d parametric guidance. CoRR, abs/2403.14781, 2024. 4, 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 579, + 545, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 579, + 545, + 632 + ], + "spans": [ + { + "bbox": [ + 307, + 579, + 545, + 632 + ], + "type": "text", + "content": "[52] Zi-Xin Zou, Zhipeng Yu, Yuan-Chen Guo, Yangguang Li, Ding Liang, Yan-Pei Cao, and Song-Hai Zhang. Triplane meets gaussian splatting: Fast and generalizable single-view 3d reconstruction with transformers. In CVPR 2024, pages 10324-10335. IEEE, 2024. 4" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "5580" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Single Domain Generalization for Few-Shot Counting via Universal Representation Matching/ef6b53d9-c810-43e9-a737-a386bc2ad915_content_list.json b/2025/Single Domain Generalization for Few-Shot Counting via Universal Representation Matching/ef6b53d9-c810-43e9-a737-a386bc2ad915_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..c9c1d4cf013fedb4bcd2e462c6614bd6d9c54976 --- /dev/null +++ b/2025/Single Domain Generalization for Few-Shot Counting via Universal Representation Matching/ef6b53d9-c810-43e9-a737-a386bc2ad915_content_list.json @@ -0,0 +1,1627 @@ +[ + { + "type": "text", + "text": "Single Domain Generalization for Few-Shot Counting via Universal Representation Matching", + "text_level": 1, + "bbox": [ + 214, + 130, + 756, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xianing Chen Si Huo Borui Jiang Hailin $\\mathbf{H}\\mathbf{u}^{*}$ Xinghao Chen\\*", + "bbox": [ + 205, + 215, + 769, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Huawei Noah's Ark Lab.", + "bbox": [ + 401, + 236, + 566, + 250 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "sdxianing@gmail.com, {huosi,jiangborui,hailin.hu,xinghao.chen}@huawei.com", + "bbox": [ + 161, + 251, + 807, + 266 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 311, + 313, + 328 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Few-shot counting estimates the number of target objects in an image using only a few annotated exemplars. However, domain shift severely hinders existing methods to generalize to unseen scenarios. This falls into the realm of single domain generalization that remains unexplored in few-shot counting. To solve this problem, we begin by analyzing the main limitations of current methods, which typically follow a standard pipeline that extract the object prototypes from exemplars and then match them with image feature to construct the correlation map. We argue that existing methods overlook the significance of learning highly generalized prototypes. Building on this insight, we propose the first single domain generalization few-shot counting model, Universal Representation Matching, termed URM. Our primary contribution is the discovery that incorporating universal vision-language representations distilled from a large scale pretrained vision-language model into the correlation construction process substantially improves robustness to domain shifts without compromising in domain performance. As a result, URM achieves state-of-the-art performance on both in domain and the newly introduced domain generalization setting.", + "bbox": [ + 75, + 343, + 473, + 676 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 704, + 209, + 719 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Few-shot counting (FSC) [46] is the task of estimating the number of target objects in an image using only a few annotated exemplars or without any exemplar at all. Although many proposed approaches [40, 65] have achieved satisfactory results, they experience severe performance degradation when tested on data from unseen scenarios in real-world deployment due to the domain shift [24] problem. This shift arises from the deviation from the source domain training data to unseen target domain testing data, such as", + "bbox": [ + 75, + 729, + 468, + 864 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/83f5c2b727645c1d7d481fe0dd7079105ad0d0982261f73b293682159b70057d.jpg", + "image_caption": [ + "(a) domain generalization for few-shot counting" + ], + "image_footnote": [], + "bbox": [ + 500, + 314, + 908, + 444 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/265a0917b498bc31622b123b775e1bf2394dab383896a788adf3f80b8e69ff2d.jpg", + "image_caption": [ + "(b) extract-then-match" + ], + "image_footnote": [], + "bbox": [ + 511, + 465, + 885, + 564 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/df92af04a82a6cebad38c03e75f74a19bd2103afcc689fa64e305cfc45fc6d02.jpg", + "image_caption": [ + "(c) universal representation matching", + "Figure 1. Illustration of (a) domain generalization for few-shot counting, (b) the vanilla extract-then-match pipeline, and (c) our proposed universal representation matching." + ], + "image_footnote": [], + "bbox": [ + 504, + 584, + 908, + 758 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "changes in camera position and weather condition [59, 74], as well as biases introduced during data collection. For instance, images in the FSC147 [46] show each target object clearly while scenes in FSCD-LIVS [35] feature more oc", + "bbox": [ + 496, + 839, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding authors.", + "bbox": [ + 91, + 875, + 227, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "**Code is available at https://github.com/jbr97/URM.", + "bbox": [ + 91, + 887, + 433, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "4639", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "clusions and quality degradation as discussed in [35].", + "bbox": [ + 76, + 90, + 429, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Although domain adaptation (DA) [55] has been proposed to adapt the source domain distribution to that of the target domain by finetuning models with target domain data, the process is often laborious, and target domain data is not always available in practice, especially in the class-agnostic setting of FSC [46]. In contrast, domain generalization (DG) [73] seeks to generalize the model to any unseen target domain without requiring target domain data and model updates as shown in Figure 1a, that is, to learn a generalized model by only source domain with a limited distribution. We consider in this work the unexplored and challenging field of single DG for few-shot counting to address the aforementioned challenges, where only one source domain—likely of narrow distribution—is used for training.", + "bbox": [ + 75, + 108, + 468, + 319 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We start by analyzing why existing methods that follow the common extract-then-match pipeline [10] illustrated in Figure 1b fail. In this pipeline, exemplars prototypes are first extracted from the image feature. Next, candidates are generated by analyzing the correlation between prototypes and the image feature. Finally, the resulting correlation map is regressed into an object density map.", + "bbox": [ + 75, + 321, + 468, + 426 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The primary distinction among these methods lies in their approaches to constructing correlation, such as attention [17, 31, 32], Siamese similarity [10, 44, 46], and so on [51, 65]. However, they all overlook the importance of prototypes in the construction process. Prototypes learned from the source domain with a narrow distribution inherently exhibit a similarly narrow distribution. Intuitively, matching with such a limited representation in the unobserved target domain with a rather broad distribution inherently leads to sub-optimal performance. Therefore, enhancing the generalization of prototypes is essential for learning a domain-general few-shot counter.", + "bbox": [ + 73, + 429, + 470, + 609 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The contrastive language-image pre-training model (CLIP) [43] and its extensions [7, 18, 66, 72] have demonstrated remarkable generalization capacity in open-world down-stream tasks. The representation from CLIP can serve as universal representation with domain-invariant information which demonstrate extraordinary performance across several distributions owing to the vast diversity of distributions seen during training [36]. Our finding indicates that matching with the universal vision-language representation distilled from the frozen CLIP model during the correlation construction process exhibit remarkable resilience to domain shift while without in domain performance degradation. The operation is illustrated in Figure 1c.", + "bbox": [ + 75, + 611, + 468, + 808 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Building on the discovery, we introduce the first domain-general few-shot counting model, Universal Representation Matching, termed URM. Specifically, we introduce universal vision and language prototypes to act as knowledge distiller [53] to learn universal knowledge from the large-scale pre-trained CLIP model motivated by prompt knowledge", + "bbox": [ + 75, + 810, + 468, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "distillation [28], aiming to extend the narrow distribution of these prototypes. To adapt the V-L representations from CLIP for local objects matching, we minimally adjust the CLIP encoder while keeping its pre-trained weights frozen.", + "bbox": [ + 496, + 90, + 890, + 151 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- For the visual representation: the class token focuses more on global image properties which may not suitable for capturing localized details [11], and the image tokens are weak in local discriminability [25]. Thus, we follow MaskCLIP [72] to obtain class-specific segmentation, allowing us to extract locality object-level representations [47] which are then used to distill region-specific information into the visual prototypes.", + "- For the language representation: we follow CuPL [42] to generate diverse customized prompts that leverage knowledge contained in LLMs [1] as well as hand-written templates, to produce numerous descriptive sentences that capture important discriminative characteristics of the object categories. These prompts are subsequently encoded by the language tower to obtain universal language embeddings for distillation." + ], + "bbox": [ + 500, + 155, + 890, + 397 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The learned universal V-L prototypes are then matched with the image feature through cross attention to construct the correlation map. As the distillation operation is conducted only during the training phase, URM remains as efficient as other methods that do not require external knowledge at inference time.", + "bbox": [ + 496, + 400, + 890, + 491 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To validate URM, we conduct extensive experiments on both cross and in domain setting on FSC147 [46] and FSCD-LVIS [35] datasets in both few-shot and zero-shot scenarios. The results demonstrate that URM performs excellently on both intra-dataset and cross domain settings.", + "bbox": [ + 496, + 492, + 890, + 566 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The main contributions of our work are as follows:", + "bbox": [ + 517, + 568, + 854, + 582 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce a new challenging setting of the single domain generalized few-shot counting task. And we point out that existing paradigm's inability to accurately count objects in new scenes stems from the learned narrow distribution prototypes.", + "- We introduce URM, the first narrow source domain—general few-shot counting model. Specifically, we propose to match universal V-L representations transferred from CLIP through knowledge distillation to construct the correlation map, a novel approach not explored in previous works.", + "- Our proposed URM outperforms state-of-the-art methods by a large margin in both the introduced narrow source domain general setting as well as in domain benchmark." + ], + "bbox": [ + 500, + 587, + 890, + 797 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 500, + 814, + 648, + 830 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Few-shot Counting. Object counting refers to the task of estimating the number of specific objects present in an image, typically limited to a few predefined categories, e.g., human heads [29, 69], cells [54], cars [60], and polyps [67].", + "bbox": [ + 496, + 839, + 890, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "4640", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Few-shot counting addresses this limitation by given only a few annotated exemplars or without any exemplar at all. FamNet [46] proposed a pipeline that extracts appearance-based queries from exemplars and matches them with image feature to infer the object counts. Then, BMNet+ [51] enhanced localization by jointly learning representation with a non-linear similarity metric. CounTR [32] introduced cross attention-based interaction module to fuse image and exemplar features. SAFE-Count [65] incorporated a feature enhancement module. LOCA [10] further refined this pipeline by iteratively adapting object prototypes based on exemplar appearance and shape. DAVE [40] proposed a post-process method for LOCA by generating a high-recall detection set and verifying the results to identify and remove outliers. All of these methods follow the same extract-then-match paradigm and struggle with domain generalization.", + "bbox": [ + 75, + 90, + 472, + 333 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "CLIP-guided Counting. Some studies also explored to count by leveraging CLIP. For instance, [37] extended CLIP's capabilities to counting but still failed when objects number is large. CrowdCLIP [30] mapped crowd patches to count text by constructing ranking-based text prompts to match the size-sorted crowd patches when category names were provided. Other works [2, 21, 22, 58, 62, 76] explored enumerating objects given class names, creating a link between object categories and visual representations. Besides, some open-set methods [20, 49, 64] achieve competitive results by leveraging CLIP as a bridge between V-L representations for interactive object detection. However, they require much more computational power and data for training, and their training classes cover the test classes.", + "bbox": [ + 75, + 352, + 472, + 564 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Single Domain Generalization. Since domain shift problem in FSC remains unexplored, we mainly focus on related approaches for crowd counting. Although some domain adaptation techniques exist [61, 74, 75], they all inadequate for real-world application due to the lack of target data [4, 13, 33, 59, 74]. Domain generalization methods are primarily built on meta-learning algorithm MLDG [27], which employs meta-train and meta-test set to simulate the domain shift between training and testing phases in reality to learn domain-invariant features. For instance, DCCUS [12] proposed a dynamic sub-domain division scheme, dividing the source domain into multiple sub-domains. Whitening-based methods [9, 38, 39] remove domain-specific style information by eliminating feature correlation to ensure each feature has unit variance. However, these algorithms require a sufficiently broad source domain distribution. MPCount [41] proposed storing diverse density values for regression and reconstruct domain-invariant features using memory bank, eliminating the need for sub-domain partitioning. However, MPCount is still limited by the narrow source training data without accessing external knowledge.", + "bbox": [ + 75, + 583, + 472, + 902 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/23eea1e33e71f99ff99956f4f1862073a3c957cc97e35929cf6f2945c8903480.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 501, + 89, + 629, + 186 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/404c35f5480f00a15ab4815d4f843d1324be47abe3b69028de41f125a9e7f46e.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 633, + 89, + 756, + 186 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/0a72f5c49f4abd72a16e393c8477a8c144ae88ab1b94f0dbb155ba090938a79f.jpg", + "image_caption": [ + "(c)", + "Figure 2. t-SNE visualization of the prototypes feature space for different categories from FSC147 by (a) the vanilla paradigm trained on FSC147, (b) the vanilla paradigm trained on FSCD-LVIS, and (c) our proposed URM trained on FSCD-LVIS. Note that the visualization is conduct on the test set where the object categories are disjoint from the train set. Best viewed in color." + ], + "image_footnote": [], + "bbox": [ + 763, + 90, + 888, + 186 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Knowledge Distillation (KD). [16] proposed to optimize a student model under the effective information transfer and supervision of a teacher model or ensemble. Although several methods have been proposed to learn knowledge from teacher network with high fidelity [14, 53, 71], we simply follow CLIP-KD [63] to use feature mimicry learning [5] which has been proved to be the most effective approach for CLIP knowledge distillation.", + "bbox": [ + 496, + 324, + 893, + 446 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 500, + 459, + 591, + 474 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we first recap the preliminaries of CLIP, and then analyze why prior methods fail to generalize. Afterward, we introduce our proposed universal representation matching model.", + "bbox": [ + 496, + 484, + 893, + 546 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Preliminary. CLIP [43], trained on large-scale image-text pairs by image-level contrastive learning, has well demonstrated the potential for learning transferable knowledge and open-set visual concepts. CLIP consists of a pair of image encoder $\\mathcal{V}(\\cdot)$ and text encoder $\\mathcal{T}(\\cdot)$ , both of which are jointly trained to map input image and text into an unified representation space. However, since CLIP is designed for image-level prediction, extracting local patch-level predictions from CLIP is non-trivial [50]. MaskCLIP [72] modifies the image encoder by removing the last query and key embedding layers, as well as reformulating the value-embedding layer and the last linear layer into two respective convolutional layers, while keeping the text encoder unchanged. The text encoder then processes language prompts containing target classes, and the resulting text embedding for each class is used as a classifier to produce pixel-level mask predictions.", + "bbox": [ + 496, + 564, + 893, + 821 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Motivation. Since the existing extract-then-match paradigm fails to generalize, we begin by analyzing the reason behind this undesirable phenomenon, which motivates our proposed solution. After training on the source", + "bbox": [ + 496, + 839, + 893, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "4641", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/fa450d996fe984d55ea204b8cd57752a267250f1f811ca57cf551549890aee8c.jpg", + "image_caption": [ + "Figure 3. The framework of our proposed URM. The inference architecture is depicted in the gray part, where the learned prototypes are matched with the image feature through cross attention. The yellow part illustrates the universal V-L representations obtained from CLIP, which are distilled into the prototypes exclusively during the training phase." + ], + "image_footnote": [], + "bbox": [ + 114, + 89, + 857, + 419 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "domain, we use t-SNE [56] to visualize the feature space of prototypes for different categories on the test set of FSC147 learned using both vanilla paradigm and URM. One noticeable observation is the clear boundary in Figure 2a using vanilla paradigm trained on FSC147, which corresponds to its well performance in the intra-dataset setting, even when the categories are disjoint from those in the training set. In contrary, Figure 2b shows an unclear boundary on the target domain when the traditional paradigm is trained on dataset with large doamin gap, aligning with a significant drop in accuracy in the cross-domain setting. The phenomenon suggests that simply learning prototypes within a narrow distribution is suboptimal for generalization. The key, we believe, is that the matching process should occur in a feature space with well-defined open-set visual concept boundaries. As shown in Figure 2c, prototypes from our URM—transferred from CLIP representation—better separate objects with unseen categories in unseen scenarios, supporting our hypothesis.", + "bbox": [ + 75, + 497, + 472, + 785 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Universal Representation Matching", + "text_level": 1, + "bbox": [ + 76, + 800, + 388, + 816 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Overview. The overall framework of our proposed URM is shown in Figure 3. The input image $\\mathbf{I} \\in \\mathbb{R}^{H \\times W \\times 3}$ is first encoded by the backbone network. Multi-scale features are extracted from the last three blocks, then resized to the same size of $h \\times w$ and reduced to $d$ channels using a", + "bbox": [ + 75, + 824, + 470, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "linear layer after concatenation. Next, a set of $n$ selected exemplar bounding boxes are encoded into the image feature $\\mathbf{f} \\in \\mathbb{R}^{h \\times w \\times d}$ via a prompt encoder. After that, the proposed universal vision and language prototypes $p_v$ and $p_l \\in \\mathbb{R}^{n \\times d}$ are initialized as learnable embeddings, and iteratively updated through interaction with the encoded image feature $f^E \\in \\mathbb{R}^{hw \\times d}$ using $N_1$ cross attention layers [57]. During this process, the resulting prototypes are distilled from the universal V-L representations obtained from the frozen CLIP model, respectively. Finally, the concatenated universal prototypes are correlated with $f^E$ by $N_2$ cross attention layers to obtain a correlation map, which is then regressed into a density map $\\mathbf{R} \\in \\mathbb{R}^{H \\times W}$ , whose number is summed to the count estimation, i.e., $c = \\text{sum}(\\mathbf{R})$ .", + "bbox": [ + 496, + 497, + 893, + 709 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Prompt Encoder. Given the extracted image feature $\\mathbf{f}$ , we use a two-way attention mechanism [23] to encode exemplar prompts into $\\mathbf{f}$ , to indicate the object category to be counted. The exemplars are represented as a concatenated embedding composed of three elements: the height and width $R^{n\\times 2}$ of bounding boxes $R^{n\\times 4}$ which are mapped to shape embedding $R^{n\\times d}$ by three linear layers with ReLU activation follow [10], a learnable embedding group $R^{n\\times d}$ , and appearance tokens extracted from $\\mathbf{f}$ using ROI Pooling [48]. As illustrated in Figure 4, cross attention operations are then applied iteratively to update both the em", + "bbox": [ + 496, + 734, + 893, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4642", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/b00d49cefa4f5f9a36d7ed8dcb784fbfa5ef30127c964656e91f30fadce26279.jpg", + "image_caption": [ + "Figure 4. Illustration of the prompt encoder." + ], + "image_footnote": [], + "bbox": [ + 117, + 93, + 436, + 251 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "bedding and image feature, resulting in the category-aware image feature $f^{E}$ .", + "bbox": [ + 75, + 311, + 467, + 340 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Universal Vision Representation Learning. Since CLIP trained with image-level supervision lacks precise localization of objects and includes noise on non-relevant areas in its visual representation, we follow MaskCLIP [72] to reformulate the value-embedding layer of image encoder $\\mathcal{V}(\\cdot)$ to produce a dense attentive segmentation $\\mathcal{M} \\in \\mathbb{R}^{h \\times w}$ .", + "bbox": [ + 75, + 362, + 467, + 453 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Then, the universal vision representation $r_v$ containing fine-grained local image semantics is obtained by applying Mask Pooling [6] between the 2D visual embeddings of the image output from the last layer of CLIP vision encoder and the class-aware mask $\\mathcal{M}$ :", + "bbox": [ + 75, + 454, + 467, + 527 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nr _ {v} = \\operatorname {M a s k P o o l} (\\mathcal {V} (\\mathbf {I}), \\mathcal {M}) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 181, + 547, + 467, + 564 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Additionally, the vision prototype $p_v$ initialized as a learnable embedding $\\mathbb{R}^{n \\times d}$ is iteratively updated through $N_1$ cross attention layers with the encoded image feature, where $q = p_v$ , $k = v = f^E$ .", + "bbox": [ + 76, + 571, + 467, + 632 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Since distribution of the prototype learned from a narrow source domain is also narrow, we adopt feature mimicry loss following CLIP-KD [63] to transfer universal representation knowledge into the visual prototypes:", + "bbox": [ + 75, + 633, + 467, + 694 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {V - K D} = \\frac {1}{| \\mathcal {B} |} \\sum_ {k = 1} ^ {| \\mathcal {B} |} \\| r _ {v k} - p _ {v k} \\| \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 161, + 707, + 467, + 750 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{B}$ is the training batch, $r_v$ is the universal vision representation, and $p_v$ is the universal vision prototype. For simplicity, the linear projection layer for matching the dimension size of the CLIP embedding is omitted.", + "bbox": [ + 75, + 758, + 467, + 818 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Universal Language Representation Learning. Handwritten templates like “A photo of {}” completed with a category name are limited in describing details for fine-grained embedding [42]. Besides, writing high-performing", + "bbox": [ + 75, + 839, + 467, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "prompt templates is labor-intensive and requires prior information about dataset contents. For example, ImageNet hand-written prompts [43] like “A toy {}” and “A sculpture of {}” do not generalize well to other datasets like FSC-147.", + "bbox": [ + 498, + 90, + 890, + 151 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We thus follow CuPL [42] to further extend templates with customized prompts generated by LLM that contain important discriminating characteristics which can best represent an object category in natural language. The templates and prompts for LLM, i.e., GPT-4 [1] in this work, are detailed in Appendix.", + "bbox": [ + 498, + 151, + 890, + 241 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The generated prompts for each category are then encoded by CLIP's text encoder $\\mathcal{T}(\\cdot)$ . The corresponding text embeddings are averaged to produce the resulting universal text representation $r_t$ .", + "bbox": [ + 498, + 242, + 890, + 301 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Similar to vision prototype learning, the language prototype $p_l$ is iteratively updated through interactions with $f^E$ using the same cross attention layers. And the language distillation loss is given by:", + "bbox": [ + 498, + 301, + 890, + 362 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {L - K D} = \\frac {1}{\\mathcal {B}} \\sum_ {k = 1} ^ {\\mathcal {B}} \\| r _ {l k} - p _ {l k} \\| \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 593, + 373, + 890, + 414 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $r_l$ is the universal language representation, $p_l$ is the universal language prototype. The linear projection layer is also applied to the language prototype.", + "bbox": [ + 498, + 417, + 890, + 463 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Universal Representation Matching. After obtaining the universal V-L prototypes for the object category in the image, we can construct a correlation between the concatenated prototypes and image feature. Although various matching approaches have been proposed as discussed, we focus on the essential matching elements within the correlation construction process. Specifically, we apply $N_{2}$ cross attention layers with $\\mathbf{q} = f^{E}, \\mathbf{k} = \\mathbf{v} = \\mathrm{concat}(p_{v}, p_{l})$ to obtain the correlation map $\\mathbb{R}^{hw \\times d}$ .", + "bbox": [ + 498, + 479, + 890, + 616 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Finally, a regression head predict the final 2D density map $\\mathbf{R} \\in \\mathbb{R}^{H \\times W}$ from the correlation map.", + "bbox": [ + 498, + 616, + 890, + 646 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Loss Function. The model is trained with $\\ell_2$ loss between the predicted density map $\\mathbf{R}$ and the ground-truth map $\\mathbf{G}$ , normalized by the number of objects:", + "bbox": [ + 498, + 662, + 890, + 709 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {d e n s i t y}} = \\frac {1}{2 \\mathcal {B}} \\sum_ {k = 1} ^ {\\mathcal {B}} \\frac {1}{N _ {k}} \\| \\boldsymbol {G} _ {k} - \\boldsymbol {R} _ {k} \\| _ {2} ^ {2} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 570, + 720, + 890, + 762 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $N_{k}$ represents the number of objects in image $k$ . The overall training loss is thus a combination of the density and the distillation losses from Equation (2) (3):", + "bbox": [ + 498, + 763, + 890, + 810 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {\\text {d e n s i t y}} + \\alpha \\mathcal {L} _ {V - K D} + (1 - \\alpha) \\mathcal {L} _ {L - K D} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 542, + 832, + 890, + 849 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\alpha$ is the distillation weight parameter to balance the contributions of the vision and language representations in the universal prototypes.", + "bbox": [ + 498, + 854, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "4643", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/af14cdc47c1442473edf412f0833844589fa15844d020cdc6a63473597e7c041.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
SettingFew-shotZero-shot
Source → TargetA → BB → AA → BB → A
MetricMAERMSEMAERMSEMAERMSEMAERMSE
Vanilla Methods
FamNet (CVPR21) [46]29.9848.9542.74118.47////
BMNet+ (CVPR22) [51]37.8668.9436.53116.02////
CounTR (BMVC22) [32]34.7654.3933.24107.34////
LOCA (ICCV23) [10]30.6344.3329.1388.0631.1746.9834.13113.72
DAVE (CVPR24) [40]29.6745.2829.5186.1330.4247.0335.40115.81
Domain Generalization Methods
ISW (CVPR21) [9]35.9551.8336.41124.09////
GS (ICCV21) [34]30.5547.5729.8398.42////
DCCUS (AAAI23) [12]29.0243.5528.4683.3230.3944.8734.29109.20
MPCount (CVPR24) [41]25.1141.3222.0780.1727.6844.5830.99102.18
URM (Ours)21.8738.4221.1773.4223.5439.9327.4994.57
", + "bbox": [ + 148, + 87, + 823, + 323 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Comparison with the state-of-the-art methods on FSC-147 (A) and FSCD-LVIS (B).", + "bbox": [ + 199, + 332, + 769, + 347 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 372, + 209, + 388 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we evaluate the effectiveness of URM, demonstrating how universal representation matching enhances the generalization and performance of few-shot counting models.", + "bbox": [ + 75, + 397, + 470, + 458 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experimental setting", + "text_level": 1, + "bbox": [ + 76, + 465, + 274, + 483 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation details. We first resize input images to $H = W = 512$ and apply standard data augmentations such as tiling, horizontal flipping and color jitter. The model uses an Imagenet pre-trained Resnet50 backbone, with features from the final three blocks upsampled to $h = w = 64$ . The backbone network parameters are frozen follow [10, 46], while all other parameters are trained for 150 epochs. The resulting features are projected into $d = 256$ channels via an linear layer. The distillation weight $\\alpha$ in Equation (5) is set to 0.9. The number of cross attention layers is $N_{1} = N_{2} = 3$ . The regression head consists of three $3 \\times 3$ convolutional layers with 128, 64 and 32 feature channels, each followed by a Leaky ReLU, a $2 \\times$ bilinear upsampling, and a linear layer, consistent with the LOCA configuration. We use AdamW optimizer with the fixed learning rate $10^{-4}$ and a weight decay of $10^{-2}$ . Gradient clipping with maximum norm of 0.1 is applied. URM is trained on a single NVIDIA V100 GPU with batch size of 4.", + "bbox": [ + 75, + 489, + 470, + 762 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. We evaluate our method on two mainstream datasets: FSC-147 [46] and FSCD-LVIS [35]. Specifically, FSC-147 contains 6135 images of 147 object categories, split into training, validation and test sets consisting of 3659, 1286 and 1190 images respectively. Images in FSC-147 are collected from the Internet. The sets of object categories present in each split are disjoint. FSCD-LVIS contains 6195 images spanning 372 classes, which are", + "bbox": [ + 75, + 780, + 470, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "extracted from the LVIS dataset [15]. Each image annotation consists of three bounding boxes for exemplar objects within each category.", + "bbox": [ + 496, + 373, + 890, + 419 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Let $S \\to \\mathcal{T}$ denote the case where $S$ is the source domain and $\\mathcal{T}$ is the target domain. The datasets are utilized in experiments under two cross-domain settings: the full training dataset of one is regarded as the source domain, while other entire test dataset is regarded as the target domain, i.e., FSC-147 $\\rightarrow$ FSCD-LVIS and FSCD-LVIS $\\rightarrow$ FSC-147.", + "bbox": [ + 496, + 419, + 890, + 508 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation metrics. We evaluate our method using two commonly employed metrics in object counting: mean absolute error (MAE) and root mean square error (RMSE), between estimated and labeled counts for each image:", + "bbox": [ + 496, + 530, + 893, + 592 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nM A E = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\left| c _ {i} - \\hat {c} _ {i} \\right|, R M S E = \\sqrt {\\frac {1}{N} \\sum_ {i = 1} ^ {N} \\left(c _ {i} - \\hat {c} _ {i}\\right) ^ {2}} \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 501, + 617, + 890, + 676 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $N$ is the number of test images, $c_{i}$ is the ground truth count for the $i$ -th image, and $\\hat{c}_i$ is the predicted count. Lower values of both metrics indicate better performance.", + "bbox": [ + 496, + 678, + 890, + 724 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Comparison with the state of the art", + "text_level": 1, + "bbox": [ + 498, + 733, + 815, + 750 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we compare our URM with state-of-the-art methods on different benchmarks. The selected baselines can be grouped into two categories:", + "bbox": [ + 496, + 757, + 890, + 803 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Vanilla Methods: FamNet [46], BMNet+ [51], CounTR [32], LOCA [10], and DAVE [40].", + "- Domain Generalization Methods: ISW [9, 41], GS [34], DCCUS [12], and MPCount [41]." + ], + "bbox": [ + 500, + 806, + 890, + 867 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "All these DG methods are originally designed or adapted for crowd counting, we add the prompt encoder after feature", + "bbox": [ + 500, + 869, + 890, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "4644", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/adccebd21a50c00a5613b4b1b4104d5ec79a8b000eebcf26a3d7b0524e853af7.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodMAERMSE
Baseline30.1548.04
Language Distillation
Naive Template25.4442.59
Prompt Generator23.8341.03
Vision Distillation
[CLS] Token23.5341.12
Global Pooling22.9439.44
Mask Pooling21.8738.42
", + "bbox": [ + 114, + 88, + 433, + 229 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "extraction based on their implementation. Besides, GS is a combination of Gradient Surgery [34] and LOCA [10].", + "bbox": [ + 75, + 284, + 468, + 315 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 1 presents the results of domain generalized few-shot counting. Our URM surpasses all the methods significantly, providing compelling evidence for the effectiveness of our designation. In contrast, sub-domain division methods like DCCUS and GS fail to achieve satisfactory results, as they struggle to produce meaningful sub-domains when the source domain distribution is narrow and the domain gap is large. Additionally, the whitening-based method ISW is even harmful because useful information is eliminated through the feature statistics normalization operation. Although MPCount introduced a single memory bank for regression does not require sub-domain partitioning, its generalization ability remains far inferior to our method.", + "bbox": [ + 75, + 316, + 468, + 512 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We further evaluate our method on the zero-shot setting with no user annotation, the results are shown in the right part of Table 1. We follow [10] to take a minor modification of the prompt encoder by using only the learnable embedding to represent the exemplar prompt. The results confirm that our proposed URM effectively enhances the generalization ability of prototypes, enabling accurate count estimation even in the challenging case without manually annotated exemplar.", + "bbox": [ + 75, + 513, + 468, + 650 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Analysis and Ablations", + "text_level": 1, + "bbox": [ + 76, + 662, + 290, + 679 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Ablation of the designed elements. We first ablate the important elements of our designation under the FSC-147 (A) $\\rightarrow$ FSCD-LVIS (B) few-shot setting. The baseline model shares the same architecture as URM in the inference phase but without knowledge distillation. The result shown in the first row of Table 2 demonstrates that simply altering architecture, i.e., matching with learned prototypes that lack knowledge transfer and thus have a narrow distribution, introduces no benefit for generalization.", + "bbox": [ + 75, + 686, + 468, + 823 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In contrast, the method can achieve comparable performance to the SOTA method MPCount by using only language knowledge from CLIP with a naive template. Since the naive template \"A photo of {}\" is too limited to fully represent a category, adding more templates and customized", + "bbox": [ + 75, + 824, + 468, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/635da56a77bc969e2e0eebdad5042e157a50af304bcaf75d49fec1ea94f93131.jpg", + "table_caption": [ + "Table 2. Ablation of the designed elements." + ], + "table_footnote": [], + "table_body": "
α00.250.50.750.91
MAE26.5423.4423.3022.3721.8723.83
RMSE42.8239.8739.4838.9238.4241.03
", + "bbox": [ + 504, + 88, + 887, + 137 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/579c645fa7eb2fd2989426a5c24dff50a4919d77528eaa93df1264eeb67aaae7.jpg", + "table_caption": [ + "Table 3. Ablation of the distillation weight." + ], + "table_footnote": [], + "table_body": "
N123334
N232343
MAE22.7722.2321.8723.5823.45
RMSE39.6939.1238.4241.2239.38
", + "bbox": [ + 531, + 176, + 861, + 243 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. Ablation of the number of layers, where $N_{1}$ is the universal prototypes learning layers, and $N_{2}$ is the universal representation matching layers.", + "bbox": [ + 498, + 253, + 890, + 295 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "language prompts by the prompt generator further improves the model performance.", + "bbox": [ + 498, + 323, + 890, + 352 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Subsequently, incorporating visual representation via global average pooling from the 2D vision tokens output from the last layer improve performance, demonstrating that the vision representation is complementary to language knowledge. Using the [CLS] token which captures global description results in lower performance. Finally, by introducing more precise locality representation as well as noise removal through mask pooling, our model reaches a MAE of 21.87 as shown in the last row of Table 2, outperforming the baseline method with $27.5\\%$ .", + "bbox": [ + 496, + 354, + 890, + 503 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Ablation of distillation weight. The distillation weight $\\alpha$ in Equation (5) controls the balance between visual and language knowledge contributions. In this part, we conduct an ablation study on the weight to fully understand its impact. As shown in Table 3, when $\\alpha = 0$ or $\\alpha = 1$ , where only vision or language knowledge is introduced respectively, the performance is inferior to other settings, demonstrating that V-L representations complement each other. Setting $\\alpha = 0.9$ yields the best performance, indicating that language representation from CLIP, which has been aligned to visual space, plays an influential role.", + "bbox": [ + 496, + 525, + 890, + 691 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Ablation of the number layers. We then ablate the hyper-parameters related to the number of layers. The universal prototypes learning layer number $N_{1}$ determines how many layers are required to absorb the universal knowledge transferred from CLIP. And the universal representation matching layer number $N_{2}$ defines how many layers are needed to learn an effective correlation map during matching in the hidden space. The results in Table 4 show that $N_{1} = N_{2} = 3$ achieves the best performance.", + "bbox": [ + 496, + 712, + 890, + 849 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Ablation of different representation sources. In this part, we investigate the effect of different representation", + "bbox": [ + 498, + 869, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "4645", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/4fd50233c42c4e6e289ffdc726568345c856eb72c371b1f9cdb195be165cded2.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelArchitectureACCMAERMSE
OpenAI CLIP [43]ViT-B/1668.321.8738.42
OpenAI CLIP [43]ViT-L/1475.522.6338.95
OpenAI CLIP [43]RN50×1670.523.1539.63
OpenCLIP [8]ConvNeXt-B71.522.7839.10
SigLIP [68]ViT-B/1673.421.3337.85
EVA-02-CLIP [52]ViT-B/1674.721.8638.12
", + "bbox": [ + 78, + 88, + 470, + 198 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "learning sources in Table 5. The first row presents the results of our default setting which uses the OpenAI ViT-B/16 CLIP model. Increasing the model size leads to a performance drop due to the large dimension gap between the CLIP model and URM as seen in the second row. Besides, models with convolutional backbones, such as ResNet50 and ConvNeXt, perform poorly in comparison despite achieving higher ImageNet accuracy, suggesting that ViT architecture is better suited for representation learning in our task. Finally, the results in the last two lines indicate that using a stronger CLIP teacher model can further enhance our model's performance. However, this improvement also does not directly correlate with ImageNet accuracy, aligning with the observations in [26].", + "bbox": [ + 75, + 277, + 468, + 488 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Performance on in domain benchmark", + "text_level": 1, + "bbox": [ + 76, + 497, + 410, + 512 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this section, URM is compared with the state-of-the-art methods on in domain benchmark. The three-shot evaluation results on FSC147 dataset are presented in Table 6. While demonstrating strong generalization, URM outperforms state-of-the-art few-shot counting models in terms of in-domain performance. Additionally, DAVE [40] proposed a detect-and-verify paradigm for the existing regression-based methods like LOCA to mitigate false activations on objects of other categories. Since DAVE is a general method for merging the benefits of both density and detection-based approaches, we also extend our method by incorporating this paradigm. Although inductive biases introduced by DAVE are insufficient to handle complex distribution gap in domain generalization setting (as show in Table 1), the resulting method still surpasses other methods on the traditional in domain benchmark, as demonstrated in the last row of Table 6.", + "bbox": [ + 75, + 520, + 468, + 777 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Generating language prompts with LVLMs", + "text_level": 1, + "bbox": [ + 76, + 786, + 444, + 803 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this section, we explore the use of large vision-language models to generate language prompts, enabling our method to train without predefined category names. However, existing proprietary and public LVLMs struggle with image classification especially in open-world setting where the class list is unknown [70], making them difficult to", + "bbox": [ + 75, + 809, + 468, + 902 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/c7f391d148839015a26215e9486f1468afbacaa4b87c7cb61d994a8c4063a757.jpg", + "table_caption": [ + "Table 5. Ablation of different universal V-L representation sources, where ACC represents the ImageNet zero-shot classification performance of the CLIP model." + ], + "table_footnote": [], + "table_body": "
SplitValidation setTest set
MethodMAERMSEMAERMSE
FamNet (CVPR21) [46]23.7569.0722.0899.54
BMNet+ (CVPR22) [51]15.7458.5314.6291.83
VCN (CVPR22) [45]19.3860.1518.1795.60
CountTR (BMVC22) [32]13.1349.8311.9591.23
SAFECount (WACV23) [65]15.2847.2014.3285.54
PseCo (CVPR24) [19]15.3168.3413.05112.86
LOCA (ICCV23) [10]10.2432.5610.7956.97
DAVE (CVPR24) [40]8.9128.088.6632.36
URM (Ours)10.6332.0110.1554.83
URM+DAVE [40]8.4726.518.3232.05
", + "bbox": [ + 501, + 88, + 903, + 275 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/1fe44c5a5da46a1f1bb2c5037a3a42e58349df751877a1c32dd8050f6be30a99.jpg", + "table_caption": [ + "Table 6. Performance comparison on in domain benchmark." + ], + "table_footnote": [], + "table_body": "
LVLMMAERMSE
Claude3 [3]22.5838.96
GPT4 [1]22.4538.75
", + "bbox": [ + 584, + 313, + 805, + 363 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 7. Results of generating language prompts using LVLMs when category names are not given.", + "bbox": [ + 500, + 373, + 890, + 402 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "accurately fill templates in the prompt generator. Fortunately, customized descriptive prompts with details are effective [42] as shown in Section 4.3 and Appendix. As a result, we choose Claude3-Opus [3] and GPT4 [1] to generate language prompts which have demonstrated the best open-world classification performance among existing VLMs [70]. We replace “{}” in our template prompts with “The category of the most numerous objects in the image” to prompt the VLMs. For example, “Describe what $a(n)$ {} looks like?” becomes “Describe what the category of the most numerous objects in the image looks like?”. The results in Table 7 show that our method can successfully generalize to scenarios where category names are not provided during training.", + "bbox": [ + 496, + 429, + 893, + 641 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 654, + 619, + 669 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper proposes URM to address single domain generalization for few-shot counting with narrow source distribution. URM is motivated by the unique challenge that existing methods struggle to learn high generalization prototypes. To address this issue, we propose a universal representation matching paradigm that incorporating universal vision-language prototypes distilled from the universal representations from CLIP model, which has seen vast diversity of distributions during training. Matching with these universal prototypes leads to high generalization and superior performance. Extensive experiments on both cross and in domain settings show that URM significantly outperform the state-of-the-art.", + "bbox": [ + 496, + 679, + 893, + 875 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "4646", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] OpenAI Josh Achiam, Steven Adler, Sandhini Agarwal, and et al. Gpt-4 technical report. 2023. 2, 5, 8", + "[2] Niki Amini-Naeni, Kiana Amini-Naeni, Tengda Han, and Andrew Zisserman. Open-world text-specified object counting. arXiv preprint arXiv:2306.01851, 2023. 3", + "[3] anthropic. Introducing the next generation of claude. 2024. 8", + "[4] Yiqing Cai, Lianggangxu Chen, Haoyue Guan, Shaohui Lin, Changhong Lu, Changbo Wang, and Gaoqi He. Explicit invariant feature induced cross-domain crowd counting. In AAAI Conference on Artificial Intelligence, 2023. 3", + "[5] Xianing Chen, Qiong Cao, Yujie Zhong, Jing Zhang, Shenghua Gao, and Dacheng Tao. Dearkd: Data-efficient early knowledge distillation for vision transformers. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12042-12052, 2022. 3", + "[6] Xi Chen, Shuang Li, Ser Nam Lim, Antonio Torralba, and Hengshuang Zhao. Open-vocabulary panoptic segmentation with embedding modulation. 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 1141-1150, 2023. 5", + "[7] Junting Cheng, Dong Liang, and Shan Tan. Transfer clip for generalizable image denoising. ArXiv, abs/2403.15132, 2024. 2", + "[8] Mehdi Cherti, Romain Beaumont, Ross Wightman, Mitchell Wortsman, Gabriel Ilharco, Cade Gordon, Christoph Schuhmann, Ludwig Schmidt, and Jenia Jitsev. Reproducible scaling laws for contrastive language-image learning. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2818-2829, 2022. 8", + "[9] Sungha Choi, Sanghun Jung, Huiwon Yun, Joanne Taery Kim, Seungryong Kim, and Jaegul Choo. Robustnet: Improving domain generalization in urban-scene segmentation via instance selective whitening. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11575-11585, 2021. 3, 6", + "[10] Nikola Djukic, Alan Luke, Vitjan Zavrtanik, and Matej Kristan. A low-shot object counting network with iterative prototype adaptation. 2023 IEEE/CVF International Conference on Computer Vision (ICCV), 2022. 2, 3, 4, 6, 7, 8", + "[11] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. ArXiv, abs/2010.11929, 2020. 2", + "[12] Zhipeng Du, Jiankang Deng, and Miaojing Shi. Domain-general crowd counting in unseen scenarios. ArXiv, abs/2212.02573, 2022. 3, 6", + "[13] Junyuan Gao, Tao Han, Yuan Yuan, and Qi Wang. Domain-adaptive crowd counting via high-quality image translation and density reconstruction. IEEE Transactions on Neural Networks and Learning Systems, 34:4803-4815, 2019. 3", + "[14] Jianping Gou, B. Yu, Stephen J. Maybank, and Dacheng Tao. Knowledge distillation: A survey. International Journal of Computer Vision, 129:1789 - 1819, 2020. 3" + ], + "bbox": [ + 78, + 114, + 468, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] Agrim Gupta, Piotr Dólár, and Ross B. Girshick. Lvis: A dataset for large vocabulary instance segmentation. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5351-5359, 2019. 6", + "[16] Geoffrey E. Hinton, Oriol Vinyals, and Jeffrey Dean. Distilling the knowledge in a neural network. ArXiv, abs/1503.02531, 2015. 3", + "[17] Michael A. Hobley and Victor Adrian Prisacariu. Learning to count anything: Reference-less class-agnostic counting with weak supervision. ArXiv, abs/2205.10203, 2022. 2", + "[18] Xiaohui Huang, Hao Zhou, Kun Yao, and Kai Han. Froster: Frozen clip is a strong teacher for open-vocabulary action recognition. ArXiv, abs/2402.03241, 2024. 2", + "[19] Zhizhong Huang, Mingliang Dai, Yi Zhang, Junping Zhang, and Hongming Shan. Point, segment and count: A generalized framework for object counting. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17067-17076, 2023. 8", + "[20] Qing Jiang, Feng Li, Zhaoyang Zeng, Tianhe Ren, Shilong Liu, and Lei Zhang. T-rex2: Towards generic object detection via text-visual prompt synergy. In European Conference on Computer Vision, pages 38-57. Springer, 2024. 3", + "[21] Ruixiang Jiang, Lingbo Liu, and Changwen Chen. Clipcount: Towards text-guided zero-shot object counting. In Proceedings of the 31st ACM International Conference on Multimedia, pages 4535-4545, 2023. 3", + "[22] Seunggu Kang, WonJun Moon, Euiyeon Kim, and Jae-Pil Heo. Vlcounter: Text-aware visual representation for zero-shot object counting. In AAAI Conference on Artificial Intelligence, 2023. 3", + "[23] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dólár, and Ross B. Girshick. Segment anything. 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 3992-4003, 2023. 4", + "[24] Wouter M. Kouw. An introduction to domain adaptation and transfer learning. ArXiv, abs/1812.11806, 2018. 1", + "[25] Mengcheng Lan, Chaofeng Chen, Yiping Ke, Xinjiang Wang, Litong Feng, and Wayne Zhang. Clearclip: Decomposing clip representations for dense vision-language inference. ArXiv, abs/2407.12442, 2024. 2", + "[26] Hugo Laurençon, Léo Tronchon, Matthieu Cord, and Victor Sanh. What matters when building vision-language models? ArXiv, abs/2405.02246, 2024. 8", + "[27] Da Li, Yongxin Yang, Yi-Zhe Song, and Timothy M. Hospedales. Learning to generalize: Meta-learning for domain generalization. In AAAI Conference on Artificial Intelligence, 2017. 3", + "[28] Zheng Li, Xiang Li, Xinyi Fu, Xing Zhang, Weiqiang Wang, Shuo Chen, and Jian Yang. Promptkd: Unsupervised prompt distillation for vision-language models. ArXiv, abs/2403.02781, 2024. 2", + "[29] Dongze Lian, Xianing Chen, Jing Li, Weixin Luo, and Shenghua Gao. Locating and counting heads in crowds with a depth prior. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44:9056-9072, 2021. 2" + ], + "bbox": [ + 501, + 92, + 890, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "4647", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[30] Dingkang Liang, Jiahao Xie, Zhikang Zou, Xiaoqing Ye, Wei Xu, and Xiang Bai. Crowdclip: Unsupervised crowd counting via vision-language model. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2893-2903, 2023. 3", + "[31] Hui Lin, Xiaopeng Hong, and Yabin Wang. Object counting: You only need to look at one. ArXiv, abs/2112.05993, 2021. 2", + "[32] Chang Liu, Yujie Zhong, Andrew Zisserman, and Weidi Xie. Countr: Transformer-based generalised visual counting. ArXiv, abs/2208.13721, 2022. 2, 3, 6, 8", + "[33] Yuting Liu, Zheng Wang, Miaojing Shi, Shin'ichi Satoh, Qijun Zhao, and Hongyu Yang. Towards unsupervised crowd counting via regression-detection bi-knowledge transfer. Proceedings of the 28th ACM International Conference on Multimedia, 2020. 3", + "[34] Lucas Mansilla, Rodrigo Echeveste, Diego H. Milone, and Enzo Ferrante. Domain generalization via gradient surgery. 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 6610-6618, 2021. 6, 7", + "[35] Thanh Thoi Nguyen, Chau Khoa Pham, Khoi Duc Minh Nguyen, and Minh Hoai. Few-shot object counting and detection. ArXiv, abs/2207.10988, 2022. 1, 2, 6", + "[36] Utkarsh Ojha, Yuheng Li, and Yong Jae Lee. Towards universal fake image detectors that generalize across generative models. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 24480-24489, 2023. 2", + "[37] Roni Paiss, Ariel Ephrat, Omer Tov, Shiran Zada, Inbar Mosseri, Michal Irani, and Tali Dekel. Teaching clip to count to ten. 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 3147-3157, 2023. 3", + "[38] Xingang Pan, Ping Luo, Jianping Shi, and Xiaou Tang. Two at once: Enhancing learning and generalization capacities via ibn-net. In European Conference on Computer Vision, 2018. 3", + "[39] Xingang Pan, Xiaohang Zhan, Jianping Shi, Xiaou Tang, and Ping Luo. Switchable whitening for deep representation learning. 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 1863-1871, 2019. 3", + "[40] Jer Pelhan, Alan Luke, Vitjan Zavrtanik, and Matej Kristan. Dave - a detect-and-verify paradigm for low-shot counting. ArXiv, abs/2404.16622, 2024. 1, 3, 6, 8", + "[41] Zhuoxuan Peng and S.-H. Gary Chan. Single domain generalization for crowd counting. ArXiv, abs/2403.09124, 2024. 3, 6", + "[42] Sarah Pratt, Rosanne Liu, and Ali Farhadi. What does a platypus look like? generating customized prompts for zero-shot image classification. 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 15645-15655, 2022. 2, 5, 8", + "[43] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, 2021. 2, 3, 5, 8" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[44] Viresh Ranjan and Minh Hoai. Exemplar free class agnostic counting. In Asian Conference on Computer Vision, 2022. 2", + "[45] Viresh Ranjan and Minh Hoai. Vicinal counting networks. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 4220-4229, 2022. 8", + "[46] Viresh Ranjan, Udbhav Sharma, Thua Nguyen, and Minh Hoai. Learning to count everything. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3393-3402, 2021. 1, 2, 3, 6, 8", + "[47] Hanoona Abdul Rasheed, Muhammad Maaz, Muhammad Uzair Khattak, Salman H. Khan, and Fahad Shahbaz Khan. Bridging the gap between object and image-level representations for open-vocabulary detection. ArXiv, abs/2207.03482, 2022. 2", + "[48] Shaoqing Ren, Kaiming He, Ross B. Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. IEEE Transactions on Pattern Analysis and Machine Intelligence, 39:1137-1149, 2015. 4", + "[49] Tianhe Ren, Qing Jiang, Shilong Liu, Zhaoyang Zeng, Wenlong Liu, Han Gao, Hongjie Huang, Zhengyu Ma, Xiaoke Jiang, Yihao Chen, et al. Grounding dino 1.5: Advance the\" edge\" of open-set object detection. arXiv preprint arXiv:2405.10300, 2024. 3", + "[50] Cheng Shi and Sibei Yang. Edadet: Open-vocabulary object detection using early dense alignment. 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 15678-15688, 2023. 3", + "[51] Min Shi, Hao Lu, Chen Feng, Chengxin Liu, and ZHIGUO CAO. Represent, compare, and learn: A similarity-aware framework for class-agnostic counting. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9519–9528, 2022. 2, 3, 6, 8", + "[52] Quan Sun, Yuxin Fang, Ledell Yu Wu, Xinlong Wang, and Yue Cao. Eva-clip: Improved training techniques for clip at scale. ArXiv, abs/2303.15389, 2023. 8", + "[53] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Herv'e J'egou. Training data-efficient image transformers & distillation through attention. In International Conference on Machine Learning, 2020. 2, 3", + "[54] Aayushi Tyagi, Chirag Mohapatra, Prasenjit Das, Govind Kumar Makharia, Lalita Mehra, A. P. Prathosh, Musam, and Iit Delhi. Degpr: Deep guided posterior regularization for multi-class cell detection and counting. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 23913-23923, 2023. 2", + "[55] Eric Tzeng, Judy Hoffman, Kate Saenko, and Trevor Darrell. Adversarial discriminative domain adaptation. 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2962-2971, 2017. 2", + "[56] Laurens van der Maaten and Geoffrey E. Hinton. Visualizing data using t-sne. Journal of Machine Learning Research, 9: 2579-2605, 2008. 4", + "[57] Ashish Vaswani, Noam M. Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Neural Information Processing Systems, 2017. 4" + ], + "bbox": [ + 501, + 92, + 890, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "4648", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[58] Mingjie Wang, Song Yuan, Zhuohang Li, Longlong Zhu, Eric Buys, and Minglun Gong. Language-guided zero-shot object counting. 2024 IEEE International Conference on Multimedia and Expo Workshops (ICMEW), pages 1-6, 2024. 3", + "[59] Qi Wang, Junyu Gao, Wei Lin, and Yuan Yuan. Learning from synthetic data for crowd counting in the wild. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8190-8199, 2019. 1, 3", + "[60] Longyin Wen, Dawei Du, Pengfei Zhu, Qinghua Hu, Qilong Wang, Liefeng Bo, and Siwei Lyu. Detection, tracking, and counting meets drones in crowds: A benchmark. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7808-7817, 2021. 2", + "[61] Qiangqiang Wu, Jia Wan, and Antoni B. Chan. Dynamic momentum adaptation for zero-shot cross-domain crowd counting. Proceedings of the 29th ACM International Conference on Multimedia, 2021. 3", + "[62] Jingyi Xu, Hieu M. Le, Vu Nguyen, Viresh Ranjan, and Dimitris Samaras. Zero-shot object counting. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15548-15557, 2023. 3", + "[63] Chuanguang Yang, Zhulin An, Libo Huang, Junyu Bi, Xinqiang Yu, Hansheng Yang, Boyu Diao, and Yongjun Xu. Clip-kd: An empirical study of clip model distillation. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15952-15962, 2023. 3, 5", + "[64] Lewei Yao, Renjie Pi, Jianhua Han, Xiaodan Liang, Hang Xu, Wei Zhang, Zhenguo Li, and Dan Xu. Detclipv3: Towards versatile generative open-vocabulary object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 27391-27401, 2024. 3", + "[65] Zhiyuan You, Yujun Shen, Kai Yang, Wenhan Luo, X. Lu, Lei Cui, and Xinyi Le. Few-shot object counting with similarity-aware feature enhancement. 2023 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 6304-6313, 2022. 1, 2, 3, 8", + "[66] Wenwen Yu, Yuliang Liu, Wei Hua, Deqiang Jiang, Bo Ren, and Xiang Bai. Turning a clip model into a scene text detector. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6978-6988, 2023. 2", + "[67] Vitjan Zavrtanik, Martin Vodopivec, and Matej Kristan. A segmentation-based approach for polyp counting in the wild. Eng. Appl. Artif. Intell., 88, 2020. 2", + "[68] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 11941-11952, 2023. 8", + "[69] Yingying Zhang, Desen Zhou, Siqin Chen, Shenghua Gao, and Yi Ma. Single-image crowd counting via multi-column convolutional neural network. 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 589-597, 2016. 2", + "[70] Yuhui Zhang, Alyssa Unell, Xiaohan Wang, Dhruba Ghosh, Yuchang Su, Ludwig Schmidt, and Serena Yeung-Levy. Why are visually-grounded language models bad at image classification? ArXiv, abs/2405.18415, 2024. 8" + ], + "bbox": [ + 78, + 90, + 468, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[71] Borui Zhao, Quan Cui, Renjie Song, Yiyu Qiu, and Jiajun Liang. Decoupled knowledge distillation. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11943-11952, 2022. 3", + "[72] Chong Zhou, Chen Change Loy, and Bo Dai. Extract free dense labels from clip. In European Conference on Computer Vision, 2021. 2, 3, 5", + "[73] Kaiyang Zhou, Ziwei Liu, Yu Qiao, Tao Xiang, and Chen Change Loy. Domain generalization: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45:4396-4415, 2021. 2", + "[74] Huilin Zhu, Jingling Yuan, Zhengwei Yang, Xian Zhong, and Zheng Wang. Fine-grained fragment diffusion for cross-domain crowd counting. Proceedings of the 30th ACM International Conference on Multimedia, 2022. 1, 3", + "[75] Huilin Zhu, Jingling Yuan, Xian Zhong, Zhengwei Yang, Zheng Wang, and Shengfeng He. Daot: Domain-agnostically aligned optimal transport for domain-adaptive crowd counting. Proceedings of the 31st ACM International Conference on Multimedia, 2023. 3", + "[76] Huilin Zhu, Jingling Yuan, Zhengwei Yang, Yu Guo, Zheng Wang, Xian Zhong, and Shengfeng He. Zero-shot object counting with good exemplars. ArXiv, abs/2407.04948, 2024.3" + ], + "bbox": [ + 501, + 90, + 892, + 429 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "4649", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Single Domain Generalization for Few-Shot Counting via Universal Representation Matching/ef6b53d9-c810-43e9-a737-a386bc2ad915_model.json b/2025/Single Domain Generalization for Few-Shot Counting via Universal Representation Matching/ef6b53d9-c810-43e9-a737-a386bc2ad915_model.json new file mode 100644 index 0000000000000000000000000000000000000000..526eb1205ae4b90073e5a5b39870d49afda61100 --- /dev/null +++ b/2025/Single Domain Generalization for Few-Shot Counting via Universal Representation Matching/ef6b53d9-c810-43e9-a737-a386bc2ad915_model.json @@ -0,0 +1,2543 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.215, + 0.131, + 0.758, + 0.177 + ], + "angle": 0, + "content": "Single Domain Generalization for Few-Shot Counting via Universal Representation Matching" + }, + { + "type": "text", + "bbox": [ + 0.207, + 0.216, + 0.77, + 0.235 + ], + "angle": 0, + "content": "Xianing Chen Si Huo Borui Jiang Hailin \\(\\mathbf{H}\\mathbf{u}^{*}\\) Xinghao Chen\\*" + }, + { + "type": "text", + "bbox": [ + 0.402, + 0.237, + 0.568, + 0.251 + ], + "angle": 0, + "content": "Huawei Noah's Ark Lab." + }, + { + "type": "text", + "bbox": [ + 0.163, + 0.252, + 0.808, + 0.267 + ], + "angle": 0, + "content": "sdxianing@gmail.com, {huosi,jiangborui,hailin.hu,xinghao.chen}@huawei.com" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.312, + 0.314, + 0.329 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.344, + 0.474, + 0.678 + ], + "angle": 0, + "content": "Few-shot counting estimates the number of target objects in an image using only a few annotated exemplars. However, domain shift severely hinders existing methods to generalize to unseen scenarios. This falls into the realm of single domain generalization that remains unexplored in few-shot counting. To solve this problem, we begin by analyzing the main limitations of current methods, which typically follow a standard pipeline that extract the object prototypes from exemplars and then match them with image feature to construct the correlation map. We argue that existing methods overlook the significance of learning highly generalized prototypes. Building on this insight, we propose the first single domain generalization few-shot counting model, Universal Representation Matching, termed URM. Our primary contribution is the discovery that incorporating universal vision-language representations distilled from a large scale pretrained vision-language model into the correlation construction process substantially improves robustness to domain shifts without compromising in domain performance. As a result, URM achieves state-of-the-art performance on both in domain and the newly introduced domain generalization setting." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.705, + 0.21, + 0.72 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.73, + 0.47, + 0.866 + ], + "angle": 0, + "content": "Few-shot counting (FSC) [46] is the task of estimating the number of target objects in an image using only a few annotated exemplars or without any exemplar at all. Although many proposed approaches [40, 65] have achieved satisfactory results, they experience severe performance degradation when tested on data from unseen scenarios in real-world deployment due to the domain shift [24] problem. This shift arises from the deviation from the source domain training data to unseen target domain testing data, such as" + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.315, + 0.909, + 0.445 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.571, + 0.447, + 0.839, + 0.46 + ], + "angle": 0, + "content": "(a) domain generalization for few-shot counting" + }, + { + "type": "image", + "bbox": [ + 0.512, + 0.467, + 0.886, + 0.565 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.633, + 0.569, + 0.762, + 0.581 + ], + "angle": 0, + "content": "(b) extract-then-match" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.585, + 0.909, + 0.759 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.598, + 0.763, + 0.812, + 0.776 + ], + "angle": 0, + "content": "(c) universal representation matching" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.79, + 0.895, + 0.832 + ], + "angle": 0, + "content": "Figure 1. Illustration of (a) domain generalization for few-shot counting, (b) the vanilla extract-then-match pipeline, and (c) our proposed universal representation matching." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.895, + 0.903 + ], + "angle": 0, + "content": "changes in camera position and weather condition [59, 74], as well as biases introduced during data collection. For instance, images in the FSC147 [46] show each target object clearly while scenes in FSCD-LIVS [35] feature more oc" + }, + { + "type": "page_footnote", + "bbox": [ + 0.093, + 0.875, + 0.228, + 0.888 + ], + "angle": 0, + "content": "*Corresponding authors." + }, + { + "type": "page_footnote", + "bbox": [ + 0.093, + 0.888, + 0.434, + 0.9 + ], + "angle": 0, + "content": "**Code is available at https://github.com/jbr97/URM." + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.875, + 0.434, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4639" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.431, + 0.106 + ], + "angle": 0, + "content": "clusions and quality degradation as discussed in [35]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.109, + 0.47, + 0.32 + ], + "angle": 0, + "content": "Although domain adaptation (DA) [55] has been proposed to adapt the source domain distribution to that of the target domain by finetuning models with target domain data, the process is often laborious, and target domain data is not always available in practice, especially in the class-agnostic setting of FSC [46]. In contrast, domain generalization (DG) [73] seeks to generalize the model to any unseen target domain without requiring target domain data and model updates as shown in Figure 1a, that is, to learn a generalized model by only source domain with a limited distribution. We consider in this work the unexplored and challenging field of single DG for few-shot counting to address the aforementioned challenges, where only one source domain—likely of narrow distribution—is used for training." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.322, + 0.47, + 0.428 + ], + "angle": 0, + "content": "We start by analyzing why existing methods that follow the common extract-then-match pipeline [10] illustrated in Figure 1b fail. In this pipeline, exemplars prototypes are first extracted from the image feature. Next, candidates are generated by analyzing the correlation between prototypes and the image feature. Finally, the resulting correlation map is regressed into an object density map." + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.43, + 0.471, + 0.61 + ], + "angle": 0, + "content": "The primary distinction among these methods lies in their approaches to constructing correlation, such as attention [17, 31, 32], Siamese similarity [10, 44, 46], and so on [51, 65]. However, they all overlook the importance of prototypes in the construction process. Prototypes learned from the source domain with a narrow distribution inherently exhibit a similarly narrow distribution. Intuitively, matching with such a limited representation in the unobserved target domain with a rather broad distribution inherently leads to sub-optimal performance. Therefore, enhancing the generalization of prototypes is essential for learning a domain-general few-shot counter." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.612, + 0.47, + 0.809 + ], + "angle": 0, + "content": "The contrastive language-image pre-training model (CLIP) [43] and its extensions [7, 18, 66, 72] have demonstrated remarkable generalization capacity in open-world down-stream tasks. The representation from CLIP can serve as universal representation with domain-invariant information which demonstrate extraordinary performance across several distributions owing to the vast diversity of distributions seen during training [36]. Our finding indicates that matching with the universal vision-language representation distilled from the frozen CLIP model during the correlation construction process exhibit remarkable resilience to domain shift while without in domain performance degradation. The operation is illustrated in Figure 1c." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Building on the discovery, we introduce the first domain-general few-shot counting model, Universal Representation Matching, termed URM. Specifically, we introduce universal vision and language prototypes to act as knowledge distiller [53] to learn universal knowledge from the large-scale pre-trained CLIP model motivated by prompt knowledge" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.892, + 0.152 + ], + "angle": 0, + "content": "distillation [28], aiming to extend the narrow distribution of these prototypes. To adapt the V-L representations from CLIP for local objects matching, we minimally adjust the CLIP encoder while keeping its pre-trained weights frozen." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.156, + 0.892, + 0.277 + ], + "angle": 0, + "content": "- For the visual representation: the class token focuses more on global image properties which may not suitable for capturing localized details [11], and the image tokens are weak in local discriminability [25]. Thus, we follow MaskCLIP [72] to obtain class-specific segmentation, allowing us to extract locality object-level representations [47] which are then used to distill region-specific information into the visual prototypes." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.277, + 0.892, + 0.398 + ], + "angle": 0, + "content": "- For the language representation: we follow CuPL [42] to generate diverse customized prompts that leverage knowledge contained in LLMs [1] as well as hand-written templates, to produce numerous descriptive sentences that capture important discriminative characteristics of the object categories. These prompts are subsequently encoded by the language tower to obtain universal language embeddings for distillation." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.156, + 0.892, + 0.398 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.401, + 0.892, + 0.492 + ], + "angle": 0, + "content": "The learned universal V-L prototypes are then matched with the image feature through cross attention to construct the correlation map. As the distillation operation is conducted only during the training phase, URM remains as efficient as other methods that do not require external knowledge at inference time." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.493, + 0.892, + 0.568 + ], + "angle": 0, + "content": "To validate URM, we conduct extensive experiments on both cross and in domain setting on FSC147 [46] and FSCD-LVIS [35] datasets in both few-shot and zero-shot scenarios. The results demonstrate that URM performs excellently on both intra-dataset and cross domain settings." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.569, + 0.856, + 0.583 + ], + "angle": 0, + "content": "The main contributions of our work are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.588, + 0.892, + 0.663 + ], + "angle": 0, + "content": "- We introduce a new challenging setting of the single domain generalized few-shot counting task. And we point out that existing paradigm's inability to accurately count objects in new scenes stems from the learned narrow distribution prototypes." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.664, + 0.892, + 0.752 + ], + "angle": 0, + "content": "- We introduce URM, the first narrow source domain—general few-shot counting model. Specifically, we propose to match universal V-L representations transferred from CLIP through knowledge distillation to construct the correlation map, a novel approach not explored in previous works." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.754, + 0.892, + 0.799 + ], + "angle": 0, + "content": "- Our proposed URM outperforms state-of-the-art methods by a large margin in both the introduced narrow source domain general setting as well as in domain benchmark." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.588, + 0.892, + 0.799 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.815, + 0.65, + 0.831 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Few-shot Counting. Object counting refers to the task of estimating the number of specific objects present in an image, typically limited to a few predefined categories, e.g., human heads [29, 69], cells [54], cars [60], and polyps [67]." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4640" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.334 + ], + "angle": 0, + "content": "Few-shot counting addresses this limitation by given only a few annotated exemplars or without any exemplar at all. FamNet [46] proposed a pipeline that extracts appearance-based queries from exemplars and matches them with image feature to infer the object counts. Then, BMNet+ [51] enhanced localization by jointly learning representation with a non-linear similarity metric. CounTR [32] introduced cross attention-based interaction module to fuse image and exemplar features. SAFE-Count [65] incorporated a feature enhancement module. LOCA [10] further refined this pipeline by iteratively adapting object prototypes based on exemplar appearance and shape. DAVE [40] proposed a post-process method for LOCA by generating a high-recall detection set and verifying the results to identify and remove outliers. All of these methods follow the same extract-then-match paradigm and struggle with domain generalization." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.353, + 0.473, + 0.565 + ], + "angle": 0, + "content": "CLIP-guided Counting. Some studies also explored to count by leveraging CLIP. For instance, [37] extended CLIP's capabilities to counting but still failed when objects number is large. CrowdCLIP [30] mapped crowd patches to count text by constructing ranking-based text prompts to match the size-sorted crowd patches when category names were provided. Other works [2, 21, 22, 58, 62, 76] explored enumerating objects given class names, creating a link between object categories and visual representations. Besides, some open-set methods [20, 49, 64] achieve competitive results by leveraging CLIP as a bridge between V-L representations for interactive object detection. However, they require much more computational power and data for training, and their training classes cover the test classes." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.584, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Single Domain Generalization. Since domain shift problem in FSC remains unexplored, we mainly focus on related approaches for crowd counting. Although some domain adaptation techniques exist [61, 74, 75], they all inadequate for real-world application due to the lack of target data [4, 13, 33, 59, 74]. Domain generalization methods are primarily built on meta-learning algorithm MLDG [27], which employs meta-train and meta-test set to simulate the domain shift between training and testing phases in reality to learn domain-invariant features. For instance, DCCUS [12] proposed a dynamic sub-domain division scheme, dividing the source domain into multiple sub-domains. Whitening-based methods [9, 38, 39] remove domain-specific style information by eliminating feature correlation to ensure each feature has unit variance. However, these algorithms require a sufficiently broad source domain distribution. MPCount [41] proposed storing diverse density values for regression and reconstruct domain-invariant features using memory bank, eliminating the need for sub-domain partitioning. However, MPCount is still limited by the narrow source training data without accessing external knowledge." + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.09, + 0.63, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.559, + 0.19, + 0.576, + 0.202 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.635, + 0.09, + 0.758, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.688, + 0.19, + 0.706, + 0.201 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.764, + 0.091, + 0.889, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.819, + 0.19, + 0.835, + 0.201 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.214, + 0.895, + 0.299 + ], + "angle": 0, + "content": "Figure 2. t-SNE visualization of the prototypes feature space for different categories from FSC147 by (a) the vanilla paradigm trained on FSC147, (b) the vanilla paradigm trained on FSCD-LVIS, and (c) our proposed URM trained on FSCD-LVIS. Note that the visualization is conduct on the test set where the object categories are disjoint from the train set. Best viewed in color." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.325, + 0.895, + 0.448 + ], + "angle": 0, + "content": "Knowledge Distillation (KD). [16] proposed to optimize a student model under the effective information transfer and supervision of a teacher model or ensemble. Although several methods have been proposed to learn knowledge from teacher network with high fidelity [14, 53, 71], we simply follow CLIP-KD [63] to use feature mimicry learning [5] which has been proved to be the most effective approach for CLIP knowledge distillation." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.46, + 0.593, + 0.475 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.485, + 0.895, + 0.547 + ], + "angle": 0, + "content": "In this section, we first recap the preliminaries of CLIP, and then analyze why prior methods fail to generalize. Afterward, we introduce our proposed universal representation matching model." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.565, + 0.895, + 0.822 + ], + "angle": 0, + "content": "Preliminary. CLIP [43], trained on large-scale image-text pairs by image-level contrastive learning, has well demonstrated the potential for learning transferable knowledge and open-set visual concepts. CLIP consists of a pair of image encoder \\(\\mathcal{V}(\\cdot)\\) and text encoder \\(\\mathcal{T}(\\cdot)\\), both of which are jointly trained to map input image and text into an unified representation space. However, since CLIP is designed for image-level prediction, extracting local patch-level predictions from CLIP is non-trivial [50]. MaskCLIP [72] modifies the image encoder by removing the last query and key embedding layers, as well as reformulating the value-embedding layer and the last linear layer into two respective convolutional layers, while keeping the text encoder unchanged. The text encoder then processes language prompts containing target classes, and the resulting text embedding for each class is used as a classifier to produce pixel-level mask predictions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Motivation. Since the existing extract-then-match paradigm fails to generalize, we begin by analyzing the reason behind this undesirable phenomenon, which motivates our proposed solution. After training on the source" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "4641" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.115, + 0.09, + 0.859, + 0.42 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.43, + 0.896, + 0.473 + ], + "angle": 0, + "content": "Figure 3. The framework of our proposed URM. The inference architecture is depicted in the gray part, where the learned prototypes are matched with the image feature through cross attention. The yellow part illustrates the universal V-L representations obtained from CLIP, which are distilled into the prototypes exclusively during the training phase." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.498, + 0.473, + 0.786 + ], + "angle": 0, + "content": "domain, we use t-SNE [56] to visualize the feature space of prototypes for different categories on the test set of FSC147 learned using both vanilla paradigm and URM. One noticeable observation is the clear boundary in Figure 2a using vanilla paradigm trained on FSC147, which corresponds to its well performance in the intra-dataset setting, even when the categories are disjoint from those in the training set. In contrary, Figure 2b shows an unclear boundary on the target domain when the traditional paradigm is trained on dataset with large doamin gap, aligning with a significant drop in accuracy in the cross-domain setting. The phenomenon suggests that simply learning prototypes within a narrow distribution is suboptimal for generalization. The key, we believe, is that the matching process should occur in a feature space with well-defined open-set visual concept boundaries. As shown in Figure 2c, prototypes from our URM—transferred from CLIP representation—better separate objects with unseen categories in unseen scenarios, supporting our hypothesis." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.801, + 0.389, + 0.817 + ], + "angle": 0, + "content": "3.1. Universal Representation Matching" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.471, + 0.903 + ], + "angle": 0, + "content": "Overview. The overall framework of our proposed URM is shown in Figure 3. The input image \\(\\mathbf{I} \\in \\mathbb{R}^{H \\times W \\times 3}\\) is first encoded by the backbone network. Multi-scale features are extracted from the last three blocks, then resized to the same size of \\(h \\times w\\) and reduced to \\(d\\) channels using a" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.498, + 0.895, + 0.71 + ], + "angle": 0, + "content": "linear layer after concatenation. Next, a set of \\(n\\) selected exemplar bounding boxes are encoded into the image feature \\(\\mathbf{f} \\in \\mathbb{R}^{h \\times w \\times d}\\) via a prompt encoder. After that, the proposed universal vision and language prototypes \\(p_v\\) and \\(p_l \\in \\mathbb{R}^{n \\times d}\\) are initialized as learnable embeddings, and iteratively updated through interaction with the encoded image feature \\(f^E \\in \\mathbb{R}^{hw \\times d}\\) using \\(N_1\\) cross attention layers [57]. During this process, the resulting prototypes are distilled from the universal V-L representations obtained from the frozen CLIP model, respectively. Finally, the concatenated universal prototypes are correlated with \\(f^E\\) by \\(N_2\\) cross attention layers to obtain a correlation map, which is then regressed into a density map \\(\\mathbf{R} \\in \\mathbb{R}^{H \\times W}\\), whose number is summed to the count estimation, i.e., \\(c = \\text{sum}(\\mathbf{R})\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Prompt Encoder. Given the extracted image feature \\(\\mathbf{f}\\), we use a two-way attention mechanism [23] to encode exemplar prompts into \\(\\mathbf{f}\\), to indicate the object category to be counted. The exemplars are represented as a concatenated embedding composed of three elements: the height and width \\(R^{n\\times 2}\\) of bounding boxes \\(R^{n\\times 4}\\) which are mapped to shape embedding \\(R^{n\\times d}\\) by three linear layers with ReLU activation follow [10], a learnable embedding group \\(R^{n\\times d}\\), and appearance tokens extracted from \\(\\mathbf{f}\\) using ROI Pooling [48]. As illustrated in Figure 4, cross attention operations are then applied iteratively to update both the em" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4642" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.118, + 0.094, + 0.437, + 0.252 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.133, + 0.269, + 0.411, + 0.284 + ], + "angle": 0, + "content": "Figure 4. Illustration of the prompt encoder." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.312, + 0.468, + 0.342 + ], + "angle": 0, + "content": "bedding and image feature, resulting in the category-aware image feature \\( f^{E} \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.363, + 0.468, + 0.454 + ], + "angle": 0, + "content": "Universal Vision Representation Learning. Since CLIP trained with image-level supervision lacks precise localization of objects and includes noise on non-relevant areas in its visual representation, we follow MaskCLIP [72] to reformulate the value-embedding layer of image encoder \\(\\mathcal{V}(\\cdot)\\) to produce a dense attentive segmentation \\(\\mathcal{M} \\in \\mathbb{R}^{h \\times w}\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.455, + 0.468, + 0.529 + ], + "angle": 0, + "content": "Then, the universal vision representation \\( r_v \\) containing fine-grained local image semantics is obtained by applying Mask Pooling [6] between the 2D visual embeddings of the image output from the last layer of CLIP vision encoder and the class-aware mask \\( \\mathcal{M} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.182, + 0.548, + 0.468, + 0.565 + ], + "angle": 0, + "content": "\\[\nr _ {v} = \\operatorname {M a s k P o o l} (\\mathcal {V} (\\mathbf {I}), \\mathcal {M}) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.573, + 0.468, + 0.633 + ], + "angle": 0, + "content": "Additionally, the vision prototype \\( p_v \\) initialized as a learnable embedding \\( \\mathbb{R}^{n \\times d} \\) is iteratively updated through \\( N_1 \\) cross attention layers with the encoded image feature, where \\( q = p_v \\), \\( k = v = f^E \\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.634, + 0.468, + 0.695 + ], + "angle": 0, + "content": "Since distribution of the prototype learned from a narrow source domain is also narrow, we adopt feature mimicry loss following CLIP-KD [63] to transfer universal representation knowledge into the visual prototypes:" + }, + { + "type": "equation", + "bbox": [ + 0.163, + 0.708, + 0.468, + 0.751 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {V - K D} = \\frac {1}{| \\mathcal {B} |} \\sum_ {k = 1} ^ {| \\mathcal {B} |} \\| r _ {v k} - p _ {v k} \\| \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.759, + 0.468, + 0.819 + ], + "angle": 0, + "content": "where \\(\\mathcal{B}\\) is the training batch, \\(r_v\\) is the universal vision representation, and \\(p_v\\) is the universal vision prototype. For simplicity, the linear projection layer for matching the dimension size of the CLIP embedding is omitted." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.84, + 0.468, + 0.902 + ], + "angle": 0, + "content": "Universal Language Representation Learning. Handwritten templates like “A photo of {}” completed with a category name are limited in describing details for fine-grained embedding [42]. Besides, writing high-performing" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.152 + ], + "angle": 0, + "content": "prompt templates is labor-intensive and requires prior information about dataset contents. For example, ImageNet hand-written prompts [43] like “A toy {}” and “A sculpture of {}” do not generalize well to other datasets like FSC-147." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.152, + 0.892, + 0.242 + ], + "angle": 0, + "content": "We thus follow CuPL [42] to further extend templates with customized prompts generated by LLM that contain important discriminating characteristics which can best represent an object category in natural language. The templates and prompts for LLM, i.e., GPT-4 [1] in this work, are detailed in Appendix." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.243, + 0.892, + 0.302 + ], + "angle": 0, + "content": "The generated prompts for each category are then encoded by CLIP's text encoder \\(\\mathcal{T}(\\cdot)\\). The corresponding text embeddings are averaged to produce the resulting universal text representation \\(r_t\\)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.303, + 0.892, + 0.363 + ], + "angle": 0, + "content": "Similar to vision prototype learning, the language prototype \\( p_l \\) is iteratively updated through interactions with \\( f^E \\) using the same cross attention layers. And the language distillation loss is given by:" + }, + { + "type": "equation", + "bbox": [ + 0.594, + 0.374, + 0.891, + 0.415 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {L - K D} = \\frac {1}{\\mathcal {B}} \\sum_ {k = 1} ^ {\\mathcal {B}} \\| r _ {l k} - p _ {l k} \\| \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.419, + 0.892, + 0.464 + ], + "angle": 0, + "content": "where \\( r_l \\) is the universal language representation, \\( p_l \\) is the universal language prototype. The linear projection layer is also applied to the language prototype." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.481, + 0.892, + 0.617 + ], + "angle": 0, + "content": "Universal Representation Matching. After obtaining the universal V-L prototypes for the object category in the image, we can construct a correlation between the concatenated prototypes and image feature. Although various matching approaches have been proposed as discussed, we focus on the essential matching elements within the correlation construction process. Specifically, we apply \\( N_{2} \\) cross attention layers with \\( \\mathbf{q} = f^{E}, \\mathbf{k} = \\mathbf{v} = \\mathrm{concat}(p_{v}, p_{l}) \\) to obtain the correlation map \\( \\mathbb{R}^{hw \\times d} \\)." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.617, + 0.891, + 0.647 + ], + "angle": 0, + "content": "Finally, a regression head predict the final 2D density map \\(\\mathbf{R} \\in \\mathbb{R}^{H \\times W}\\) from the correlation map." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.664, + 0.891, + 0.71 + ], + "angle": 0, + "content": "Loss Function. The model is trained with \\(\\ell_2\\) loss between the predicted density map \\(\\mathbf{R}\\) and the ground-truth map \\(\\mathbf{G}\\), normalized by the number of objects:" + }, + { + "type": "equation", + "bbox": [ + 0.571, + 0.721, + 0.891, + 0.763 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {d e n s i t y}} = \\frac {1}{2 \\mathcal {B}} \\sum_ {k = 1} ^ {\\mathcal {B}} \\frac {1}{N _ {k}} \\| \\boldsymbol {G} _ {k} - \\boldsymbol {R} _ {k} \\| _ {2} ^ {2} \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.765, + 0.891, + 0.811 + ], + "angle": 0, + "content": "where \\(N_{k}\\) represents the number of objects in image \\(k\\). The overall training loss is thus a combination of the density and the distillation losses from Equation (2) (3):" + }, + { + "type": "equation", + "bbox": [ + 0.543, + 0.833, + 0.891, + 0.85 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {\\text {d e n s i t y}} + \\alpha \\mathcal {L} _ {V - K D} + (1 - \\alpha) \\mathcal {L} _ {L - K D} \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.856, + 0.891, + 0.901 + ], + "angle": 0, + "content": "where \\(\\alpha\\) is the distillation weight parameter to balance the contributions of the vision and language representations in the universal prototypes." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.515, + 0.956 + ], + "angle": 0, + "content": "4643" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.149, + 0.088, + 0.825, + 0.324 + ], + "angle": 0, + "content": "
SettingFew-shotZero-shot
Source → TargetA → BB → AA → BB → A
MetricMAERMSEMAERMSEMAERMSEMAERMSE
Vanilla Methods
FamNet (CVPR21) [46]29.9848.9542.74118.47////
BMNet+ (CVPR22) [51]37.8668.9436.53116.02////
CounTR (BMVC22) [32]34.7654.3933.24107.34////
LOCA (ICCV23) [10]30.6344.3329.1388.0631.1746.9834.13113.72
DAVE (CVPR24) [40]29.6745.2829.5186.1330.4247.0335.40115.81
Domain Generalization Methods
ISW (CVPR21) [9]35.9551.8336.41124.09////
GS (ICCV21) [34]30.5547.5729.8398.42////
DCCUS (AAAI23) [12]29.0243.5528.4683.3230.3944.8734.29109.20
MPCount (CVPR24) [41]25.1141.3222.0780.1727.6844.5830.99102.18
URM (Ours)21.8738.4221.1773.4223.5439.9327.4994.57
" + }, + { + "type": "table_caption", + "bbox": [ + 0.2, + 0.333, + 0.771, + 0.348 + ], + "angle": 0, + "content": "Table 1. Comparison with the state-of-the-art methods on FSC-147 (A) and FSCD-LVIS (B)." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.373, + 0.21, + 0.39 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.398, + 0.471, + 0.459 + ], + "angle": 0, + "content": "In this section, we evaluate the effectiveness of URM, demonstrating how universal representation matching enhances the generalization and performance of few-shot counting models." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.467, + 0.276, + 0.484 + ], + "angle": 0, + "content": "4.1. Experimental setting" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.49, + 0.471, + 0.763 + ], + "angle": 0, + "content": "Implementation details. We first resize input images to \\( H = W = 512 \\) and apply standard data augmentations such as tiling, horizontal flipping and color jitter. The model uses an Imagenet pre-trained Resnet50 backbone, with features from the final three blocks upsampled to \\( h = w = 64 \\). The backbone network parameters are frozen follow [10, 46], while all other parameters are trained for 150 epochs. The resulting features are projected into \\( d = 256 \\) channels via an linear layer. The distillation weight \\( \\alpha \\) in Equation (5) is set to 0.9. The number of cross attention layers is \\( N_{1} = N_{2} = 3 \\). The regression head consists of three \\( 3 \\times 3 \\) convolutional layers with 128, 64 and 32 feature channels, each followed by a Leaky ReLU, a \\( 2 \\times \\) bilinear upsampling, and a linear layer, consistent with the LOCA configuration. We use AdamW optimizer with the fixed learning rate \\( 10^{-4} \\) and a weight decay of \\( 10^{-2} \\). Gradient clipping with maximum norm of 0.1 is applied. URM is trained on a single NVIDIA V100 GPU with batch size of 4." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.781, + 0.471, + 0.902 + ], + "angle": 0, + "content": "Datasets. We evaluate our method on two mainstream datasets: FSC-147 [46] and FSCD-LVIS [35]. Specifically, FSC-147 contains 6135 images of 147 object categories, split into training, validation and test sets consisting of 3659, 1286 and 1190 images respectively. Images in FSC-147 are collected from the Internet. The sets of object categories present in each split are disjoint. FSCD-LVIS contains 6195 images spanning 372 classes, which are" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.374, + 0.892, + 0.42 + ], + "angle": 0, + "content": "extracted from the LVIS dataset [15]. Each image annotation consists of three bounding boxes for exemplar objects within each category." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.42, + 0.892, + 0.51 + ], + "angle": 0, + "content": "Let \\(S \\to \\mathcal{T}\\) denote the case where \\(S\\) is the source domain and \\(\\mathcal{T}\\) is the target domain. The datasets are utilized in experiments under two cross-domain settings: the full training dataset of one is regarded as the source domain, while other entire test dataset is regarded as the target domain, i.e., FSC-147 \\(\\rightarrow\\) FSCD-LVIS and FSCD-LVIS \\(\\rightarrow\\) FSC-147." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.531, + 0.894, + 0.593 + ], + "angle": 0, + "content": "Evaluation metrics. We evaluate our method using two commonly employed metrics in object counting: mean absolute error (MAE) and root mean square error (RMSE), between estimated and labeled counts for each image:" + }, + { + "type": "equation", + "bbox": [ + 0.502, + 0.618, + 0.892, + 0.678 + ], + "angle": 0, + "content": "\\[\nM A E = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\left| c _ {i} - \\hat {c} _ {i} \\right|, R M S E = \\sqrt {\\frac {1}{N} \\sum_ {i = 1} ^ {N} \\left(c _ {i} - \\hat {c} _ {i}\\right) ^ {2}} \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.679, + 0.892, + 0.725 + ], + "angle": 0, + "content": "where \\(N\\) is the number of test images, \\(c_{i}\\) is the ground truth count for the \\(i\\)-th image, and \\(\\hat{c}_i\\) is the predicted count. Lower values of both metrics indicate better performance." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.734, + 0.816, + 0.751 + ], + "angle": 0, + "content": "4.2. Comparison with the state of the art" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.758, + 0.892, + 0.804 + ], + "angle": 0, + "content": "In this section, we compare our URM with state-of-the-art methods on different benchmarks. The selected baselines can be grouped into two categories:" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.807, + 0.891, + 0.837 + ], + "angle": 0, + "content": "- Vanilla Methods: FamNet [46], BMNet+ [51], CounTR [32], LOCA [10], and DAVE [40]." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.838, + 0.891, + 0.868 + ], + "angle": 0, + "content": "- Domain Generalization Methods: ISW [9, 41], GS [34], DCCUS [12], and MPCount [41]." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.807, + 0.891, + 0.868 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.871, + 0.892, + 0.902 + ], + "angle": 0, + "content": "All these DG methods are originally designed or adapted for crowd counting, we add the prompt encoder after feature" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4644" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.115, + 0.089, + 0.434, + 0.231 + ], + "angle": 0, + "content": "
MethodMAERMSE
Baseline30.1548.04
Language Distillation
Naive Template25.4442.59
Prompt Generator23.8341.03
Vision Distillation
[CLS] Token23.5341.12
Global Pooling22.9439.44
Mask Pooling21.8738.42
" + }, + { + "type": "table_caption", + "bbox": [ + 0.138, + 0.242, + 0.408, + 0.256 + ], + "angle": 0, + "content": "Table 2. Ablation of the designed elements." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.285, + 0.469, + 0.316 + ], + "angle": 0, + "content": "extraction based on their implementation. Besides, GS is a combination of Gradient Surgery [34] and LOCA [10]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.317, + 0.469, + 0.513 + ], + "angle": 0, + "content": "Table 1 presents the results of domain generalized few-shot counting. Our URM surpasses all the methods significantly, providing compelling evidence for the effectiveness of our designation. In contrast, sub-domain division methods like DCCUS and GS fail to achieve satisfactory results, as they struggle to produce meaningful sub-domains when the source domain distribution is narrow and the domain gap is large. Additionally, the whitening-based method ISW is even harmful because useful information is eliminated through the feature statistics normalization operation. Although MPCount introduced a single memory bank for regression does not require sub-domain partitioning, its generalization ability remains far inferior to our method." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.514, + 0.469, + 0.651 + ], + "angle": 0, + "content": "We further evaluate our method on the zero-shot setting with no user annotation, the results are shown in the right part of Table 1. We follow [10] to take a minor modification of the prompt encoder by using only the learnable embedding to represent the exemplar prompt. The results confirm that our proposed URM effectively enhances the generalization ability of prototypes, enabling accurate count estimation even in the challenging case without manually annotated exemplar." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.664, + 0.291, + 0.68 + ], + "angle": 0, + "content": "4.3. Analysis and Ablations" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.688, + 0.469, + 0.824 + ], + "angle": 0, + "content": "Ablation of the designed elements. We first ablate the important elements of our designation under the FSC-147 (A) \\(\\rightarrow\\) FSCD-LVIS (B) few-shot setting. The baseline model shares the same architecture as URM in the inference phase but without knowledge distillation. The result shown in the first row of Table 2 demonstrates that simply altering architecture, i.e., matching with learned prototypes that lack knowledge transfer and thus have a narrow distribution, introduces no benefit for generalization." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.825, + 0.469, + 0.901 + ], + "angle": 0, + "content": "In contrast, the method can achieve comparable performance to the SOTA method MPCount by using only language knowledge from CLIP with a naive template. Since the naive template \"A photo of {}\" is too limited to fully represent a category, adding more templates and customized" + }, + { + "type": "table", + "bbox": [ + 0.506, + 0.089, + 0.888, + 0.138 + ], + "angle": 0, + "content": "
α00.250.50.750.91
MAE26.5423.4423.3022.3721.8723.83
RMSE42.8239.8739.4838.9238.4241.03
" + }, + { + "type": "table_caption", + "bbox": [ + 0.561, + 0.149, + 0.831, + 0.163 + ], + "angle": 0, + "content": "Table 3. Ablation of the distillation weight." + }, + { + "type": "table", + "bbox": [ + 0.532, + 0.178, + 0.862, + 0.244 + ], + "angle": 0, + "content": "
N123334
N232343
MAE22.7722.2321.8723.5823.45
RMSE39.6939.1238.4241.2239.38
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.254, + 0.892, + 0.296 + ], + "angle": 0, + "content": "Table 4. Ablation of the number of layers, where \\( N_{1} \\) is the universal prototypes learning layers, and \\( N_{2} \\) is the universal representation matching layers." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.324, + 0.891, + 0.353 + ], + "angle": 0, + "content": "language prompts by the prompt generator further improves the model performance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.355, + 0.892, + 0.504 + ], + "angle": 0, + "content": "Subsequently, incorporating visual representation via global average pooling from the 2D vision tokens output from the last layer improve performance, demonstrating that the vision representation is complementary to language knowledge. Using the [CLS] token which captures global description results in lower performance. Finally, by introducing more precise locality representation as well as noise removal through mask pooling, our model reaches a MAE of 21.87 as shown in the last row of Table 2, outperforming the baseline method with \\(27.5\\%\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.526, + 0.892, + 0.693 + ], + "angle": 0, + "content": "Ablation of distillation weight. The distillation weight \\(\\alpha\\) in Equation (5) controls the balance between visual and language knowledge contributions. In this part, we conduct an ablation study on the weight to fully understand its impact. As shown in Table 3, when \\(\\alpha = 0\\) or \\(\\alpha = 1\\), where only vision or language knowledge is introduced respectively, the performance is inferior to other settings, demonstrating that V-L representations complement each other. Setting \\(\\alpha = 0.9\\) yields the best performance, indicating that language representation from CLIP, which has been aligned to visual space, plays an influential role." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.713, + 0.892, + 0.85 + ], + "angle": 0, + "content": "Ablation of the number layers. We then ablate the hyper-parameters related to the number of layers. The universal prototypes learning layer number \\( N_{1} \\) determines how many layers are required to absorb the universal knowledge transferred from CLIP. And the universal representation matching layer number \\( N_{2} \\) defines how many layers are needed to learn an effective correlation map during matching in the hidden space. The results in Table 4 show that \\( N_{1} = N_{2} = 3 \\) achieves the best performance." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Ablation of different representation sources. In this part, we investigate the effect of different representation" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4645" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.08, + 0.089, + 0.472, + 0.199 + ], + "angle": 0, + "content": "
ModelArchitectureACCMAERMSE
OpenAI CLIP [43]ViT-B/1668.321.8738.42
OpenAI CLIP [43]ViT-L/1475.522.6338.95
OpenAI CLIP [43]RN50×1670.523.1539.63
OpenCLIP [8]ConvNeXt-B71.522.7839.10
SigLIP [68]ViT-B/1673.421.3337.85
EVA-02-CLIP [52]ViT-B/1674.721.8638.12
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.209, + 0.471, + 0.252 + ], + "angle": 0, + "content": "Table 5. Ablation of different universal V-L representation sources, where ACC represents the ImageNet zero-shot classification performance of the CLIP model." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.278, + 0.47, + 0.489 + ], + "angle": 0, + "content": "learning sources in Table 5. The first row presents the results of our default setting which uses the OpenAI ViT-B/16 CLIP model. Increasing the model size leads to a performance drop due to the large dimension gap between the CLIP model and URM as seen in the second row. Besides, models with convolutional backbones, such as ResNet50 and ConvNeXt, perform poorly in comparison despite achieving higher ImageNet accuracy, suggesting that ViT architecture is better suited for representation learning in our task. Finally, the results in the last two lines indicate that using a stronger CLIP teacher model can further enhance our model's performance. However, this improvement also does not directly correlate with ImageNet accuracy, aligning with the observations in [26]." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.498, + 0.411, + 0.513 + ], + "angle": 0, + "content": "4.4. Performance on in domain benchmark" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.521, + 0.47, + 0.778 + ], + "angle": 0, + "content": "In this section, URM is compared with the state-of-the-art methods on in domain benchmark. The three-shot evaluation results on FSC147 dataset are presented in Table 6. While demonstrating strong generalization, URM outperforms state-of-the-art few-shot counting models in terms of in-domain performance. Additionally, DAVE [40] proposed a detect-and-verify paradigm for the existing regression-based methods like LOCA to mitigate false activations on objects of other categories. Since DAVE is a general method for merging the benefits of both density and detection-based approaches, we also extend our method by incorporating this paradigm. Although inductive biases introduced by DAVE are insufficient to handle complex distribution gap in domain generalization setting (as show in Table 1), the resulting method still surpasses other methods on the traditional in domain benchmark, as demonstrated in the last row of Table 6." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.787, + 0.445, + 0.804 + ], + "angle": 0, + "content": "4.5. Generating language prompts with LVLMs" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.81, + 0.47, + 0.903 + ], + "angle": 0, + "content": "In this section, we explore the use of large vision-language models to generate language prompts, enabling our method to train without predefined category names. However, existing proprietary and public LVLMs struggle with image classification especially in open-world setting where the class list is unknown [70], making them difficult to" + }, + { + "type": "table", + "bbox": [ + 0.503, + 0.089, + 0.905, + 0.276 + ], + "angle": 0, + "content": "
SplitValidation setTest set
MethodMAERMSEMAERMSE
FamNet (CVPR21) [46]23.7569.0722.0899.54
BMNet+ (CVPR22) [51]15.7458.5314.6291.83
VCN (CVPR22) [45]19.3860.1518.1795.60
CountTR (BMVC22) [32]13.1349.8311.9591.23
SAFECount (WACV23) [65]15.2847.2014.3285.54
PseCo (CVPR24) [19]15.3168.3413.05112.86
LOCA (ICCV23) [10]10.2432.5610.7956.97
DAVE (CVPR24) [40]8.9128.088.6632.36
URM (Ours)10.6332.0110.1554.83
URM+DAVE [40]8.4726.518.3232.05
" + }, + { + "type": "table_caption", + "bbox": [ + 0.506, + 0.286, + 0.885, + 0.3 + ], + "angle": 0, + "content": "Table 6. Performance comparison on in domain benchmark." + }, + { + "type": "table", + "bbox": [ + 0.586, + 0.314, + 0.807, + 0.364 + ], + "angle": 0, + "content": "
LVLMMAERMSE
Claude3 [3]22.5838.96
GPT4 [1]22.4538.75
" + }, + { + "type": "table_caption", + "bbox": [ + 0.5, + 0.374, + 0.892, + 0.403 + ], + "angle": 0, + "content": "Table 7. Results of generating language prompts using LVLMs when category names are not given." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.43, + 0.895, + 0.642 + ], + "angle": 0, + "content": "accurately fill templates in the prompt generator. Fortunately, customized descriptive prompts with details are effective [42] as shown in Section 4.3 and Appendix. As a result, we choose Claude3-Opus [3] and GPT4 [1] to generate language prompts which have demonstrated the best open-world classification performance among existing VLMs [70]. We replace “{}” in our template prompts with “The category of the most numerous objects in the image” to prompt the VLMs. For example, “Describe what \\(a(n)\\) {} looks like?” becomes “Describe what the category of the most numerous objects in the image looks like?”. The results in Table 7 show that our method can successfully generalize to scenarios where category names are not provided during training." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.655, + 0.62, + 0.67 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.68, + 0.894, + 0.876 + ], + "angle": 0, + "content": "This paper proposes URM to address single domain generalization for few-shot counting with narrow source distribution. URM is motivated by the unique challenge that existing methods struggle to learn high generalization prototypes. To address this issue, we propose a universal representation matching paradigm that incorporating universal vision-language prototypes distilled from the universal representations from CLIP model, which has seen vast diversity of distributions during training. Matching with these universal prototypes leads to high generalization and superior performance. Extensive experiments on both cross and in domain settings show that URM significantly outperform the state-of-the-art." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4646" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.468, + 0.143 + ], + "angle": 0, + "content": "[1] OpenAI Josh Achiam, Steven Adler, Sandhini Agarwal, and et al. Gpt-4 technical report. 2023. 2, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.145, + 0.469, + 0.186 + ], + "angle": 0, + "content": "[2] Niki Amini-Naeni, Kiana Amini-Naeni, Tengda Han, and Andrew Zisserman. Open-world text-specified object counting. arXiv preprint arXiv:2306.01851, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.187, + 0.468, + 0.213 + ], + "angle": 0, + "content": "[3] anthropic. Introducing the next generation of claude. 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.215, + 0.469, + 0.27 + ], + "angle": 0, + "content": "[4] Yiqing Cai, Lianggangxu Chen, Haoyue Guan, Shaohui Lin, Changhong Lu, Changbo Wang, and Gaoqi He. Explicit invariant feature induced cross-domain crowd counting. In AAAI Conference on Artificial Intelligence, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.271, + 0.469, + 0.34 + ], + "angle": 0, + "content": "[5] Xianing Chen, Qiong Cao, Yujie Zhong, Jing Zhang, Shenghua Gao, and Dacheng Tao. Dearkd: Data-efficient early knowledge distillation for vision transformers. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12042-12052, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.342, + 0.469, + 0.408 + ], + "angle": 0, + "content": "[6] Xi Chen, Shuang Li, Ser Nam Lim, Antonio Torralba, and Hengshuang Zhao. Open-vocabulary panoptic segmentation with embedding modulation. 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 1141-1150, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.411, + 0.469, + 0.45 + ], + "angle": 0, + "content": "[7] Junting Cheng, Dong Liang, and Shan Tan. Transfer clip for generalizable image denoising. ArXiv, abs/2403.15132, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.453, + 0.469, + 0.536 + ], + "angle": 0, + "content": "[8] Mehdi Cherti, Romain Beaumont, Ross Wightman, Mitchell Wortsman, Gabriel Ilharco, Cade Gordon, Christoph Schuhmann, Ludwig Schmidt, and Jenia Jitsev. Reproducible scaling laws for contrastive language-image learning. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2818-2829, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.538, + 0.469, + 0.62 + ], + "angle": 0, + "content": "[9] Sungha Choi, Sanghun Jung, Huiwon Yun, Joanne Taery Kim, Seungryong Kim, and Jaegul Choo. Robustnet: Improving domain generalization in urban-scene segmentation via instance selective whitening. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11575-11585, 2021. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.622, + 0.469, + 0.676 + ], + "angle": 0, + "content": "[10] Nikola Djukic, Alan Luke, Vitjan Zavrtanik, and Matej Kristan. A low-shot object counting network with iterative prototype adaptation. 2023 IEEE/CVF International Conference on Computer Vision (ICCV), 2022. 2, 3, 4, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.677, + 0.469, + 0.759 + ], + "angle": 0, + "content": "[11] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. ArXiv, abs/2010.11929, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.761, + 0.469, + 0.802 + ], + "angle": 0, + "content": "[12] Zhipeng Du, Jiankang Deng, and Miaojing Shi. Domain-general crowd counting in unseen scenarios. ArXiv, abs/2212.02573, 2022. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.804, + 0.469, + 0.858 + ], + "angle": 0, + "content": "[13] Junyuan Gao, Tao Han, Yuan Yuan, and Qi Wang. Domain-adaptive crowd counting via high-quality image translation and density reconstruction. IEEE Transactions on Neural Networks and Learning Systems, 34:4803-4815, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.86, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[14] Jianping Gou, B. Yu, Stephen J. Maybank, and Dacheng Tao. Knowledge distillation: A survey. International Journal of Computer Vision, 129:1789 - 1819, 2020. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.148 + ], + "angle": 0, + "content": "[15] Agrim Gupta, Piotr Dólár, and Ross B. Girshick. Lvis: A dataset for large vocabulary instance segmentation. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5351-5359, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.19 + ], + "angle": 0, + "content": "[16] Geoffrey E. Hinton, Oriol Vinyals, and Jeffrey Dean. Distilling the knowledge in a neural network. ArXiv, abs/1503.02531, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.192, + 0.892, + 0.234 + ], + "angle": 0, + "content": "[17] Michael A. Hobley and Victor Adrian Prisacariu. Learning to count anything: Reference-less class-agnostic counting with weak supervision. ArXiv, abs/2205.10203, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.235, + 0.892, + 0.277 + ], + "angle": 0, + "content": "[18] Xiaohui Huang, Hao Zhou, Kun Yao, and Kai Han. Froster: Frozen clip is a strong teacher for open-vocabulary action recognition. ArXiv, abs/2402.03241, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.278, + 0.892, + 0.347 + ], + "angle": 0, + "content": "[19] Zhizhong Huang, Mingliang Dai, Yi Zhang, Junping Zhang, and Hongming Shan. Point, segment and count: A generalized framework for object counting. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17067-17076, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.349, + 0.892, + 0.404 + ], + "angle": 0, + "content": "[20] Qing Jiang, Feng Li, Zhaoyang Zeng, Tianhe Ren, Shilong Liu, and Lei Zhang. T-rex2: Towards generic object detection via text-visual prompt synergy. In European Conference on Computer Vision, pages 38-57. Springer, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.406, + 0.892, + 0.461 + ], + "angle": 0, + "content": "[21] Ruixiang Jiang, Lingbo Liu, and Changwen Chen. Clipcount: Towards text-guided zero-shot object counting. In Proceedings of the 31st ACM International Conference on Multimedia, pages 4535-4545, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.462, + 0.892, + 0.517 + ], + "angle": 0, + "content": "[22] Seunggu Kang, WonJun Moon, Euiyeon Kim, and Jae-Pil Heo. Vlcounter: Text-aware visual representation for zero-shot object counting. In AAAI Conference on Artificial Intelligence, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.519, + 0.892, + 0.601 + ], + "angle": 0, + "content": "[23] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dólár, and Ross B. Girshick. Segment anything. 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 3992-4003, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.603, + 0.892, + 0.631 + ], + "angle": 0, + "content": "[24] Wouter M. Kouw. An introduction to domain adaptation and transfer learning. ArXiv, abs/1812.11806, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.633, + 0.892, + 0.687 + ], + "angle": 0, + "content": "[25] Mengcheng Lan, Chaofeng Chen, Yiping Ke, Xinjiang Wang, Litong Feng, and Wayne Zhang. Clearclip: Decomposing clip representations for dense vision-language inference. ArXiv, abs/2407.12442, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.689, + 0.892, + 0.731 + ], + "angle": 0, + "content": "[26] Hugo Laurençon, Léo Tronchon, Matthieu Cord, and Victor Sanh. What matters when building vision-language models? ArXiv, abs/2405.02246, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.732, + 0.892, + 0.787 + ], + "angle": 0, + "content": "[27] Da Li, Yongxin Yang, Yi-Zhe Song, and Timothy M. Hospedales. Learning to generalize: Meta-learning for domain generalization. In AAAI Conference on Artificial Intelligence, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.789, + 0.892, + 0.843 + ], + "angle": 0, + "content": "[28] Zheng Li, Xiang Li, Xinyi Fu, Xing Zhang, Weiqiang Wang, Shuo Chen, and Jian Yang. Promptkd: Unsupervised prompt distillation for vision-language models. ArXiv, abs/2403.02781, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.845, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[29] Dongze Lian, Xianing Chen, Jing Li, Weixin Luo, and Shenghua Gao. Locating and counting heads in crowds with a depth prior. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44:9056-9072, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4647" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[30] Dingkang Liang, Jiahao Xie, Zhikang Zou, Xiaoqing Ye, Wei Xu, and Xiang Bai. Crowdclip: Unsupervised crowd counting via vision-language model. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2893-2903, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.163, + 0.469, + 0.203 + ], + "angle": 0, + "content": "[31] Hui Lin, Xiaopeng Hong, and Yabin Wang. Object counting: You only need to look at one. ArXiv, abs/2112.05993, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.206, + 0.469, + 0.248 + ], + "angle": 0, + "content": "[32] Chang Liu, Yujie Zhong, Andrew Zisserman, and Weidi Xie. Countr: Transformer-based generalised visual counting. ArXiv, abs/2208.13721, 2022. 2, 3, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.25, + 0.469, + 0.317 + ], + "angle": 0, + "content": "[33] Yuting Liu, Zheng Wang, Miaojing Shi, Shin'ichi Satoh, Qijun Zhao, and Hongyu Yang. Towards unsupervised crowd counting via regression-detection bi-knowledge transfer. Proceedings of the 28th ACM International Conference on Multimedia, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.32, + 0.469, + 0.375 + ], + "angle": 0, + "content": "[34] Lucas Mansilla, Rodrigo Echeveste, Diego H. Milone, and Enzo Ferrante. Domain generalization via gradient surgery. 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 6610-6618, 2021. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.377, + 0.469, + 0.418 + ], + "angle": 0, + "content": "[35] Thanh Thoi Nguyen, Chau Khoa Pham, Khoi Duc Minh Nguyen, and Minh Hoai. Few-shot object counting and detection. ArXiv, abs/2207.10988, 2022. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.42, + 0.469, + 0.487 + ], + "angle": 0, + "content": "[36] Utkarsh Ojha, Yuheng Li, and Yong Jae Lee. Towards universal fake image detectors that generalize across generative models. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 24480-24489, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.49, + 0.469, + 0.546 + ], + "angle": 0, + "content": "[37] Roni Paiss, Ariel Ephrat, Omer Tov, Shiran Zada, Inbar Mosseri, Michal Irani, and Tali Dekel. Teaching clip to count to ten. 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 3147-3157, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.548, + 0.469, + 0.601 + ], + "angle": 0, + "content": "[38] Xingang Pan, Ping Luo, Jianping Shi, and Xiaou Tang. Two at once: Enhancing learning and generalization capacities via ibn-net. In European Conference on Computer Vision, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.604, + 0.469, + 0.659 + ], + "angle": 0, + "content": "[39] Xingang Pan, Xiaohang Zhan, Jianping Shi, Xiaou Tang, and Ping Luo. Switchable whitening for deep representation learning. 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 1863-1871, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.661, + 0.469, + 0.702 + ], + "angle": 0, + "content": "[40] Jer Pelhan, Alan Luke, Vitjan Zavrtanik, and Matej Kristan. Dave - a detect-and-verify paradigm for low-shot counting. ArXiv, abs/2404.16622, 2024. 1, 3, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.704, + 0.469, + 0.745 + ], + "angle": 0, + "content": "[41] Zhuoxuan Peng and S.-H. Gary Chan. Single domain generalization for crowd counting. ArXiv, abs/2403.09124, 2024. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.747, + 0.469, + 0.815 + ], + "angle": 0, + "content": "[42] Sarah Pratt, Rosanne Liu, and Ali Farhadi. What does a platypus look like? generating customized prompts for zero-shot image classification. 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 15645-15655, 2022. 2, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.818, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[43] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, 2021. 2, 3, 5, 8" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "[44] Viresh Ranjan and Minh Hoai. Exemplar free class agnostic counting. In Asian Conference on Computer Vision, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.121, + 0.892, + 0.175 + ], + "angle": 0, + "content": "[45] Viresh Ranjan and Minh Hoai. Vicinal counting networks. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 4220-4229, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.177, + 0.892, + 0.232 + ], + "angle": 0, + "content": "[46] Viresh Ranjan, Udbhav Sharma, Thua Nguyen, and Minh Hoai. Learning to count everything. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3393-3402, 2021. 1, 2, 3, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.233, + 0.892, + 0.3 + ], + "angle": 0, + "content": "[47] Hanoona Abdul Rasheed, Muhammad Maaz, Muhammad Uzair Khattak, Salman H. Khan, and Fahad Shahbaz Khan. Bridging the gap between object and image-level representations for open-vocabulary detection. ArXiv, abs/2207.03482, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.302, + 0.892, + 0.357 + ], + "angle": 0, + "content": "[48] Shaoqing Ren, Kaiming He, Ross B. Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. IEEE Transactions on Pattern Analysis and Machine Intelligence, 39:1137-1149, 2015. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.358, + 0.892, + 0.426 + ], + "angle": 0, + "content": "[49] Tianhe Ren, Qing Jiang, Shilong Liu, Zhaoyang Zeng, Wenlong Liu, Han Gao, Hongjie Huang, Zhengyu Ma, Xiaoke Jiang, Yihao Chen, et al. Grounding dino 1.5: Advance the\" edge\" of open-set object detection. arXiv preprint arXiv:2405.10300, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.428, + 0.892, + 0.482 + ], + "angle": 0, + "content": "[50] Cheng Shi and Sibei Yang. Edadet: Open-vocabulary object detection using early dense alignment. 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 15678-15688, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.484, + 0.892, + 0.552 + ], + "angle": 0, + "content": "[51] Min Shi, Hao Lu, Chen Feng, Chengxin Liu, and ZHIGUO CAO. Represent, compare, and learn: A similarity-aware framework for class-agnostic counting. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9519–9528, 2022. 2, 3, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.554, + 0.892, + 0.594 + ], + "angle": 0, + "content": "[52] Quan Sun, Yuxin Fang, Ledell Yu Wu, Xinlong Wang, and Yue Cao. Eva-clip: Improved training techniques for clip at scale. ArXiv, abs/2303.15389, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.595, + 0.892, + 0.662 + ], + "angle": 0, + "content": "[53] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Herv'e J'egou. Training data-efficient image transformers & distillation through attention. In International Conference on Machine Learning, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.664, + 0.892, + 0.747 + ], + "angle": 0, + "content": "[54] Aayushi Tyagi, Chirag Mohapatra, Prasenjit Das, Govind Kumar Makharia, Lalita Mehra, A. P. Prathosh, Musam, and Iit Delhi. Degpr: Deep guided posterior regularization for multi-class cell detection and counting. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 23913-23923, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.748, + 0.892, + 0.803 + ], + "angle": 0, + "content": "[55] Eric Tzeng, Judy Hoffman, Kate Saenko, and Trevor Darrell. Adversarial discriminative domain adaptation. 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2962-2971, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.804, + 0.892, + 0.844 + ], + "angle": 0, + "content": "[56] Laurens van der Maaten and Geoffrey E. Hinton. Visualizing data using t-sne. Journal of Machine Learning Research, 9: 2579-2605, 2008. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.845, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[57] Ashish Vaswani, Noam M. Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Neural Information Processing Systems, 2017. 4" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4648" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.16 + ], + "angle": 0, + "content": "[58] Mingjie Wang, Song Yuan, Zhuohang Li, Longlong Zhu, Eric Buys, and Minglun Gong. Language-guided zero-shot object counting. 2024 IEEE International Conference on Multimedia and Expo Workshops (ICMEW), pages 1-6, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.161, + 0.47, + 0.218 + ], + "angle": 0, + "content": "[59] Qi Wang, Junyu Gao, Wei Lin, and Yuan Yuan. Learning from synthetic data for crowd counting in the wild. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8190-8199, 2019. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.218, + 0.47, + 0.287 + ], + "angle": 0, + "content": "[60] Longyin Wen, Dawei Du, Pengfei Zhu, Qinghua Hu, Qilong Wang, Liefeng Bo, and Siwei Lyu. Detection, tracking, and counting meets drones in crowds: A benchmark. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7808-7817, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.288, + 0.47, + 0.342 + ], + "angle": 0, + "content": "[61] Qiangqiang Wu, Jia Wan, and Antoni B. Chan. Dynamic momentum adaptation for zero-shot cross-domain crowd counting. Proceedings of the 29th ACM International Conference on Multimedia, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.343, + 0.47, + 0.399 + ], + "angle": 0, + "content": "[62] Jingyi Xu, Hieu M. Le, Vu Nguyen, Viresh Ranjan, and Dimitris Samaras. Zero-shot object counting. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15548-15557, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.4, + 0.47, + 0.469 + ], + "angle": 0, + "content": "[63] Chuanguang Yang, Zhulin An, Libo Huang, Junyu Bi, Xinqiang Yu, Hansheng Yang, Boyu Diao, and Yongjun Xu. Clip-kd: An empirical study of clip model distillation. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15952-15962, 2023. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.469, + 0.47, + 0.55 + ], + "angle": 0, + "content": "[64] Lewei Yao, Renjie Pi, Jianhua Han, Xiaodan Liang, Hang Xu, Wei Zhang, Zhenguo Li, and Dan Xu. Detclipv3: Towards versatile generative open-vocabulary object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 27391-27401, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.552, + 0.469, + 0.621 + ], + "angle": 0, + "content": "[65] Zhiyuan You, Yujun Shen, Kai Yang, Wenhan Luo, X. Lu, Lei Cui, and Xinyi Le. Few-shot object counting with similarity-aware feature enhancement. 2023 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 6304-6313, 2022. 1, 2, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.622, + 0.47, + 0.678 + ], + "angle": 0, + "content": "[66] Wenwen Yu, Yuliang Liu, Wei Hua, Deqiang Jiang, Bo Ren, and Xiang Bai. Turning a clip model into a scene text detector. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6978-6988, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.678, + 0.469, + 0.718 + ], + "angle": 0, + "content": "[67] Vitjan Zavrtanik, Martin Vodopivec, and Matej Kristan. A segmentation-based approach for polyp counting in the wild. Eng. Appl. Artif. Intell., 88, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.72, + 0.47, + 0.775 + ], + "angle": 0, + "content": "[68] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 11941-11952, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.776, + 0.47, + 0.844 + ], + "angle": 0, + "content": "[69] Yingying Zhang, Desen Zhou, Siqin Chen, Shenghua Gao, and Yi Ma. Single-image crowd counting via multi-column convolutional neural network. 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 589-597, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.845, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[70] Yuhui Zhang, Alyssa Unell, Xiaohan Wang, Dhruba Ghosh, Yuchang Su, Ludwig Schmidt, and Serena Yeung-Levy. Why are visually-grounded language models bad at image classification? ArXiv, abs/2405.18415, 2024. 8" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.092, + 0.893, + 0.148 + ], + "angle": 0, + "content": "[71] Borui Zhao, Quan Cui, Renjie Song, Yiyu Qiu, and Jiajun Liang. Decoupled knowledge distillation. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11943-11952, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.893, + 0.19 + ], + "angle": 0, + "content": "[72] Chong Zhou, Chen Change Loy, and Bo Dai. Extract free dense labels from clip. In European Conference on Computer Vision, 2021. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.191, + 0.893, + 0.247 + ], + "angle": 0, + "content": "[73] Kaiyang Zhou, Ziwei Liu, Yu Qiao, Tao Xiang, and Chen Change Loy. Domain generalization: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45:4396-4415, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.248, + 0.893, + 0.303 + ], + "angle": 0, + "content": "[74] Huilin Zhu, Jingling Yuan, Zhengwei Yang, Xian Zhong, and Zheng Wang. Fine-grained fragment diffusion for cross-domain crowd counting. Proceedings of the 30th ACM International Conference on Multimedia, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.304, + 0.893, + 0.374 + ], + "angle": 0, + "content": "[75] Huilin Zhu, Jingling Yuan, Xian Zhong, Zhengwei Yang, Zheng Wang, and Shengfeng He. Daot: Domain-agnostically aligned optimal transport for domain-adaptive crowd counting. Proceedings of the 31st ACM International Conference on Multimedia, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.375, + 0.893, + 0.43 + ], + "angle": 0, + "content": "[76] Huilin Zhu, Jingling Yuan, Zhengwei Yang, Yu Guo, Zheng Wang, Xian Zhong, and Shengfeng He. Zero-shot object counting with good exemplars. ArXiv, abs/2407.04948, 2024.3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.092, + 0.893, + 0.43 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.956 + ], + "angle": 0, + "content": "4649" + } + ] +] \ No newline at end of file diff --git a/2025/Single Domain Generalization for Few-Shot Counting via Universal Representation Matching/ef6b53d9-c810-43e9-a737-a386bc2ad915_origin.pdf b/2025/Single Domain Generalization for Few-Shot Counting via Universal Representation Matching/ef6b53d9-c810-43e9-a737-a386bc2ad915_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..679c5118746763165a6e400723ddd48689b2a868 --- /dev/null +++ b/2025/Single Domain Generalization for Few-Shot Counting via Universal Representation Matching/ef6b53d9-c810-43e9-a737-a386bc2ad915_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:438aeb4f0b989b769af686f3a2bacf648645da322a504f139a8ac919d750bdcc +size 1713696 diff --git a/2025/Single Domain Generalization for Few-Shot Counting via Universal Representation Matching/full.md b/2025/Single Domain Generalization for Few-Shot Counting via Universal Representation Matching/full.md new file mode 100644 index 0000000000000000000000000000000000000000..c2100f7da15ddac8cba5157595b642b6280f25b7 --- /dev/null +++ b/2025/Single Domain Generalization for Few-Shot Counting via Universal Representation Matching/full.md @@ -0,0 +1,335 @@ +# Single Domain Generalization for Few-Shot Counting via Universal Representation Matching + +Xianing Chen Si Huo Borui Jiang Hailin $\mathbf{H}\mathbf{u}^{*}$ Xinghao Chen\* + +Huawei Noah's Ark Lab. + +sdxianing@gmail.com, {huosi,jiangborui,hailin.hu,xinghao.chen}@huawei.com + +# Abstract + +Few-shot counting estimates the number of target objects in an image using only a few annotated exemplars. However, domain shift severely hinders existing methods to generalize to unseen scenarios. This falls into the realm of single domain generalization that remains unexplored in few-shot counting. To solve this problem, we begin by analyzing the main limitations of current methods, which typically follow a standard pipeline that extract the object prototypes from exemplars and then match them with image feature to construct the correlation map. We argue that existing methods overlook the significance of learning highly generalized prototypes. Building on this insight, we propose the first single domain generalization few-shot counting model, Universal Representation Matching, termed URM. Our primary contribution is the discovery that incorporating universal vision-language representations distilled from a large scale pretrained vision-language model into the correlation construction process substantially improves robustness to domain shifts without compromising in domain performance. As a result, URM achieves state-of-the-art performance on both in domain and the newly introduced domain generalization setting. + +# 1. Introduction + +Few-shot counting (FSC) [46] is the task of estimating the number of target objects in an image using only a few annotated exemplars or without any exemplar at all. Although many proposed approaches [40, 65] have achieved satisfactory results, they experience severe performance degradation when tested on data from unseen scenarios in real-world deployment due to the domain shift [24] problem. This shift arises from the deviation from the source domain training data to unseen target domain testing data, such as + +![](images/83f5c2b727645c1d7d481fe0dd7079105ad0d0982261f73b293682159b70057d.jpg) +(a) domain generalization for few-shot counting + +![](images/265a0917b498bc31622b123b775e1bf2394dab383896a788adf3f80b8e69ff2d.jpg) +(b) extract-then-match + +![](images/df92af04a82a6cebad38c03e75f74a19bd2103afcc689fa64e305cfc45fc6d02.jpg) +(c) universal representation matching +Figure 1. Illustration of (a) domain generalization for few-shot counting, (b) the vanilla extract-then-match pipeline, and (c) our proposed universal representation matching. + +changes in camera position and weather condition [59, 74], as well as biases introduced during data collection. For instance, images in the FSC147 [46] show each target object clearly while scenes in FSCD-LIVS [35] feature more oc + +clusions and quality degradation as discussed in [35]. + +Although domain adaptation (DA) [55] has been proposed to adapt the source domain distribution to that of the target domain by finetuning models with target domain data, the process is often laborious, and target domain data is not always available in practice, especially in the class-agnostic setting of FSC [46]. In contrast, domain generalization (DG) [73] seeks to generalize the model to any unseen target domain without requiring target domain data and model updates as shown in Figure 1a, that is, to learn a generalized model by only source domain with a limited distribution. We consider in this work the unexplored and challenging field of single DG for few-shot counting to address the aforementioned challenges, where only one source domain—likely of narrow distribution—is used for training. + +We start by analyzing why existing methods that follow the common extract-then-match pipeline [10] illustrated in Figure 1b fail. In this pipeline, exemplars prototypes are first extracted from the image feature. Next, candidates are generated by analyzing the correlation between prototypes and the image feature. Finally, the resulting correlation map is regressed into an object density map. + +The primary distinction among these methods lies in their approaches to constructing correlation, such as attention [17, 31, 32], Siamese similarity [10, 44, 46], and so on [51, 65]. However, they all overlook the importance of prototypes in the construction process. Prototypes learned from the source domain with a narrow distribution inherently exhibit a similarly narrow distribution. Intuitively, matching with such a limited representation in the unobserved target domain with a rather broad distribution inherently leads to sub-optimal performance. Therefore, enhancing the generalization of prototypes is essential for learning a domain-general few-shot counter. + +The contrastive language-image pre-training model (CLIP) [43] and its extensions [7, 18, 66, 72] have demonstrated remarkable generalization capacity in open-world down-stream tasks. The representation from CLIP can serve as universal representation with domain-invariant information which demonstrate extraordinary performance across several distributions owing to the vast diversity of distributions seen during training [36]. Our finding indicates that matching with the universal vision-language representation distilled from the frozen CLIP model during the correlation construction process exhibit remarkable resilience to domain shift while without in domain performance degradation. The operation is illustrated in Figure 1c. + +Building on the discovery, we introduce the first domain-general few-shot counting model, Universal Representation Matching, termed URM. Specifically, we introduce universal vision and language prototypes to act as knowledge distiller [53] to learn universal knowledge from the large-scale pre-trained CLIP model motivated by prompt knowledge + +distillation [28], aiming to extend the narrow distribution of these prototypes. To adapt the V-L representations from CLIP for local objects matching, we minimally adjust the CLIP encoder while keeping its pre-trained weights frozen. + +- For the visual representation: the class token focuses more on global image properties which may not suitable for capturing localized details [11], and the image tokens are weak in local discriminability [25]. Thus, we follow MaskCLIP [72] to obtain class-specific segmentation, allowing us to extract locality object-level representations [47] which are then used to distill region-specific information into the visual prototypes. +- For the language representation: we follow CuPL [42] to generate diverse customized prompts that leverage knowledge contained in LLMs [1] as well as hand-written templates, to produce numerous descriptive sentences that capture important discriminative characteristics of the object categories. These prompts are subsequently encoded by the language tower to obtain universal language embeddings for distillation. + +The learned universal V-L prototypes are then matched with the image feature through cross attention to construct the correlation map. As the distillation operation is conducted only during the training phase, URM remains as efficient as other methods that do not require external knowledge at inference time. + +To validate URM, we conduct extensive experiments on both cross and in domain setting on FSC147 [46] and FSCD-LVIS [35] datasets in both few-shot and zero-shot scenarios. The results demonstrate that URM performs excellently on both intra-dataset and cross domain settings. + +The main contributions of our work are as follows: + +- We introduce a new challenging setting of the single domain generalized few-shot counting task. And we point out that existing paradigm's inability to accurately count objects in new scenes stems from the learned narrow distribution prototypes. +- We introduce URM, the first narrow source domain—general few-shot counting model. Specifically, we propose to match universal V-L representations transferred from CLIP through knowledge distillation to construct the correlation map, a novel approach not explored in previous works. +- Our proposed URM outperforms state-of-the-art methods by a large margin in both the introduced narrow source domain general setting as well as in domain benchmark. + +# 2. Related Works + +Few-shot Counting. Object counting refers to the task of estimating the number of specific objects present in an image, typically limited to a few predefined categories, e.g., human heads [29, 69], cells [54], cars [60], and polyps [67]. + +Few-shot counting addresses this limitation by given only a few annotated exemplars or without any exemplar at all. FamNet [46] proposed a pipeline that extracts appearance-based queries from exemplars and matches them with image feature to infer the object counts. Then, BMNet+ [51] enhanced localization by jointly learning representation with a non-linear similarity metric. CounTR [32] introduced cross attention-based interaction module to fuse image and exemplar features. SAFE-Count [65] incorporated a feature enhancement module. LOCA [10] further refined this pipeline by iteratively adapting object prototypes based on exemplar appearance and shape. DAVE [40] proposed a post-process method for LOCA by generating a high-recall detection set and verifying the results to identify and remove outliers. All of these methods follow the same extract-then-match paradigm and struggle with domain generalization. + +CLIP-guided Counting. Some studies also explored to count by leveraging CLIP. For instance, [37] extended CLIP's capabilities to counting but still failed when objects number is large. CrowdCLIP [30] mapped crowd patches to count text by constructing ranking-based text prompts to match the size-sorted crowd patches when category names were provided. Other works [2, 21, 22, 58, 62, 76] explored enumerating objects given class names, creating a link between object categories and visual representations. Besides, some open-set methods [20, 49, 64] achieve competitive results by leveraging CLIP as a bridge between V-L representations for interactive object detection. However, they require much more computational power and data for training, and their training classes cover the test classes. + +Single Domain Generalization. Since domain shift problem in FSC remains unexplored, we mainly focus on related approaches for crowd counting. Although some domain adaptation techniques exist [61, 74, 75], they all inadequate for real-world application due to the lack of target data [4, 13, 33, 59, 74]. Domain generalization methods are primarily built on meta-learning algorithm MLDG [27], which employs meta-train and meta-test set to simulate the domain shift between training and testing phases in reality to learn domain-invariant features. For instance, DCCUS [12] proposed a dynamic sub-domain division scheme, dividing the source domain into multiple sub-domains. Whitening-based methods [9, 38, 39] remove domain-specific style information by eliminating feature correlation to ensure each feature has unit variance. However, these algorithms require a sufficiently broad source domain distribution. MPCount [41] proposed storing diverse density values for regression and reconstruct domain-invariant features using memory bank, eliminating the need for sub-domain partitioning. However, MPCount is still limited by the narrow source training data without accessing external knowledge. + +![](images/23eea1e33e71f99ff99956f4f1862073a3c957cc97e35929cf6f2945c8903480.jpg) +(a) + +![](images/404c35f5480f00a15ab4815d4f843d1324be47abe3b69028de41f125a9e7f46e.jpg) +(b) + +![](images/0a72f5c49f4abd72a16e393c8477a8c144ae88ab1b94f0dbb155ba090938a79f.jpg) +(c) +Figure 2. t-SNE visualization of the prototypes feature space for different categories from FSC147 by (a) the vanilla paradigm trained on FSC147, (b) the vanilla paradigm trained on FSCD-LVIS, and (c) our proposed URM trained on FSCD-LVIS. Note that the visualization is conduct on the test set where the object categories are disjoint from the train set. Best viewed in color. + +Knowledge Distillation (KD). [16] proposed to optimize a student model under the effective information transfer and supervision of a teacher model or ensemble. Although several methods have been proposed to learn knowledge from teacher network with high fidelity [14, 53, 71], we simply follow CLIP-KD [63] to use feature mimicry learning [5] which has been proved to be the most effective approach for CLIP knowledge distillation. + +# 3. Method + +In this section, we first recap the preliminaries of CLIP, and then analyze why prior methods fail to generalize. Afterward, we introduce our proposed universal representation matching model. + +Preliminary. CLIP [43], trained on large-scale image-text pairs by image-level contrastive learning, has well demonstrated the potential for learning transferable knowledge and open-set visual concepts. CLIP consists of a pair of image encoder $\mathcal{V}(\cdot)$ and text encoder $\mathcal{T}(\cdot)$ , both of which are jointly trained to map input image and text into an unified representation space. However, since CLIP is designed for image-level prediction, extracting local patch-level predictions from CLIP is non-trivial [50]. MaskCLIP [72] modifies the image encoder by removing the last query and key embedding layers, as well as reformulating the value-embedding layer and the last linear layer into two respective convolutional layers, while keeping the text encoder unchanged. The text encoder then processes language prompts containing target classes, and the resulting text embedding for each class is used as a classifier to produce pixel-level mask predictions. + +Motivation. Since the existing extract-then-match paradigm fails to generalize, we begin by analyzing the reason behind this undesirable phenomenon, which motivates our proposed solution. After training on the source + +![](images/fa450d996fe984d55ea204b8cd57752a267250f1f811ca57cf551549890aee8c.jpg) +Figure 3. The framework of our proposed URM. The inference architecture is depicted in the gray part, where the learned prototypes are matched with the image feature through cross attention. The yellow part illustrates the universal V-L representations obtained from CLIP, which are distilled into the prototypes exclusively during the training phase. + +domain, we use t-SNE [56] to visualize the feature space of prototypes for different categories on the test set of FSC147 learned using both vanilla paradigm and URM. One noticeable observation is the clear boundary in Figure 2a using vanilla paradigm trained on FSC147, which corresponds to its well performance in the intra-dataset setting, even when the categories are disjoint from those in the training set. In contrary, Figure 2b shows an unclear boundary on the target domain when the traditional paradigm is trained on dataset with large doamin gap, aligning with a significant drop in accuracy in the cross-domain setting. The phenomenon suggests that simply learning prototypes within a narrow distribution is suboptimal for generalization. The key, we believe, is that the matching process should occur in a feature space with well-defined open-set visual concept boundaries. As shown in Figure 2c, prototypes from our URM—transferred from CLIP representation—better separate objects with unseen categories in unseen scenarios, supporting our hypothesis. + +# 3.1. Universal Representation Matching + +Overview. The overall framework of our proposed URM is shown in Figure 3. The input image $\mathbf{I} \in \mathbb{R}^{H \times W \times 3}$ is first encoded by the backbone network. Multi-scale features are extracted from the last three blocks, then resized to the same size of $h \times w$ and reduced to $d$ channels using a + +linear layer after concatenation. Next, a set of $n$ selected exemplar bounding boxes are encoded into the image feature $\mathbf{f} \in \mathbb{R}^{h \times w \times d}$ via a prompt encoder. After that, the proposed universal vision and language prototypes $p_v$ and $p_l \in \mathbb{R}^{n \times d}$ are initialized as learnable embeddings, and iteratively updated through interaction with the encoded image feature $f^E \in \mathbb{R}^{hw \times d}$ using $N_1$ cross attention layers [57]. During this process, the resulting prototypes are distilled from the universal V-L representations obtained from the frozen CLIP model, respectively. Finally, the concatenated universal prototypes are correlated with $f^E$ by $N_2$ cross attention layers to obtain a correlation map, which is then regressed into a density map $\mathbf{R} \in \mathbb{R}^{H \times W}$ , whose number is summed to the count estimation, i.e., $c = \text{sum}(\mathbf{R})$ . + +Prompt Encoder. Given the extracted image feature $\mathbf{f}$ , we use a two-way attention mechanism [23] to encode exemplar prompts into $\mathbf{f}$ , to indicate the object category to be counted. The exemplars are represented as a concatenated embedding composed of three elements: the height and width $R^{n\times 2}$ of bounding boxes $R^{n\times 4}$ which are mapped to shape embedding $R^{n\times d}$ by three linear layers with ReLU activation follow [10], a learnable embedding group $R^{n\times d}$ , and appearance tokens extracted from $\mathbf{f}$ using ROI Pooling [48]. As illustrated in Figure 4, cross attention operations are then applied iteratively to update both the em + +![](images/b00d49cefa4f5f9a36d7ed8dcb784fbfa5ef30127c964656e91f30fadce26279.jpg) +Figure 4. Illustration of the prompt encoder. + +bedding and image feature, resulting in the category-aware image feature $f^{E}$ . + +Universal Vision Representation Learning. Since CLIP trained with image-level supervision lacks precise localization of objects and includes noise on non-relevant areas in its visual representation, we follow MaskCLIP [72] to reformulate the value-embedding layer of image encoder $\mathcal{V}(\cdot)$ to produce a dense attentive segmentation $\mathcal{M} \in \mathbb{R}^{h \times w}$ . + +Then, the universal vision representation $r_v$ containing fine-grained local image semantics is obtained by applying Mask Pooling [6] between the 2D visual embeddings of the image output from the last layer of CLIP vision encoder and the class-aware mask $\mathcal{M}$ : + +$$ +r _ {v} = \operatorname {M a s k P o o l} (\mathcal {V} (\mathbf {I}), \mathcal {M}) \tag {1} +$$ + +Additionally, the vision prototype $p_v$ initialized as a learnable embedding $\mathbb{R}^{n \times d}$ is iteratively updated through $N_1$ cross attention layers with the encoded image feature, where $q = p_v$ , $k = v = f^E$ . + +Since distribution of the prototype learned from a narrow source domain is also narrow, we adopt feature mimicry loss following CLIP-KD [63] to transfer universal representation knowledge into the visual prototypes: + +$$ +\mathcal {L} _ {V - K D} = \frac {1}{| \mathcal {B} |} \sum_ {k = 1} ^ {| \mathcal {B} |} \| r _ {v k} - p _ {v k} \| \tag {2} +$$ + +where $\mathcal{B}$ is the training batch, $r_v$ is the universal vision representation, and $p_v$ is the universal vision prototype. For simplicity, the linear projection layer for matching the dimension size of the CLIP embedding is omitted. + +Universal Language Representation Learning. Handwritten templates like “A photo of {}” completed with a category name are limited in describing details for fine-grained embedding [42]. Besides, writing high-performing + +prompt templates is labor-intensive and requires prior information about dataset contents. For example, ImageNet hand-written prompts [43] like “A toy {}” and “A sculpture of {}” do not generalize well to other datasets like FSC-147. + +We thus follow CuPL [42] to further extend templates with customized prompts generated by LLM that contain important discriminating characteristics which can best represent an object category in natural language. The templates and prompts for LLM, i.e., GPT-4 [1] in this work, are detailed in Appendix. + +The generated prompts for each category are then encoded by CLIP's text encoder $\mathcal{T}(\cdot)$ . The corresponding text embeddings are averaged to produce the resulting universal text representation $r_t$ . + +Similar to vision prototype learning, the language prototype $p_l$ is iteratively updated through interactions with $f^E$ using the same cross attention layers. And the language distillation loss is given by: + +$$ +\mathcal {L} _ {L - K D} = \frac {1}{\mathcal {B}} \sum_ {k = 1} ^ {\mathcal {B}} \| r _ {l k} - p _ {l k} \| \tag {3} +$$ + +where $r_l$ is the universal language representation, $p_l$ is the universal language prototype. The linear projection layer is also applied to the language prototype. + +Universal Representation Matching. After obtaining the universal V-L prototypes for the object category in the image, we can construct a correlation between the concatenated prototypes and image feature. Although various matching approaches have been proposed as discussed, we focus on the essential matching elements within the correlation construction process. Specifically, we apply $N_{2}$ cross attention layers with $\mathbf{q} = f^{E}, \mathbf{k} = \mathbf{v} = \mathrm{concat}(p_{v}, p_{l})$ to obtain the correlation map $\mathbb{R}^{hw \times d}$ . + +Finally, a regression head predict the final 2D density map $\mathbf{R} \in \mathbb{R}^{H \times W}$ from the correlation map. + +Loss Function. The model is trained with $\ell_2$ loss between the predicted density map $\mathbf{R}$ and the ground-truth map $\mathbf{G}$ , normalized by the number of objects: + +$$ +\mathcal {L} _ {\text {d e n s i t y}} = \frac {1}{2 \mathcal {B}} \sum_ {k = 1} ^ {\mathcal {B}} \frac {1}{N _ {k}} \| \boldsymbol {G} _ {k} - \boldsymbol {R} _ {k} \| _ {2} ^ {2} \tag {4} +$$ + +where $N_{k}$ represents the number of objects in image $k$ . The overall training loss is thus a combination of the density and the distillation losses from Equation (2) (3): + +$$ +\mathcal {L} = \mathcal {L} _ {\text {d e n s i t y}} + \alpha \mathcal {L} _ {V - K D} + (1 - \alpha) \mathcal {L} _ {L - K D} \tag {5} +$$ + +where $\alpha$ is the distillation weight parameter to balance the contributions of the vision and language representations in the universal prototypes. + +
SettingFew-shotZero-shot
Source → TargetA → BB → AA → BB → A
MetricMAERMSEMAERMSEMAERMSEMAERMSE
Vanilla Methods
FamNet (CVPR21) [46]29.9848.9542.74118.47////
BMNet+ (CVPR22) [51]37.8668.9436.53116.02////
CounTR (BMVC22) [32]34.7654.3933.24107.34////
LOCA (ICCV23) [10]30.6344.3329.1388.0631.1746.9834.13113.72
DAVE (CVPR24) [40]29.6745.2829.5186.1330.4247.0335.40115.81
Domain Generalization Methods
ISW (CVPR21) [9]35.9551.8336.41124.09////
GS (ICCV21) [34]30.5547.5729.8398.42////
DCCUS (AAAI23) [12]29.0243.5528.4683.3230.3944.8734.29109.20
MPCount (CVPR24) [41]25.1141.3222.0780.1727.6844.5830.99102.18
URM (Ours)21.8738.4221.1773.4223.5439.9327.4994.57
+ +Table 1. Comparison with the state-of-the-art methods on FSC-147 (A) and FSCD-LVIS (B). + +# 4. Experiments + +In this section, we evaluate the effectiveness of URM, demonstrating how universal representation matching enhances the generalization and performance of few-shot counting models. + +# 4.1. Experimental setting + +Implementation details. We first resize input images to $H = W = 512$ and apply standard data augmentations such as tiling, horizontal flipping and color jitter. The model uses an Imagenet pre-trained Resnet50 backbone, with features from the final three blocks upsampled to $h = w = 64$ . The backbone network parameters are frozen follow [10, 46], while all other parameters are trained for 150 epochs. The resulting features are projected into $d = 256$ channels via an linear layer. The distillation weight $\alpha$ in Equation (5) is set to 0.9. The number of cross attention layers is $N_{1} = N_{2} = 3$ . The regression head consists of three $3 \times 3$ convolutional layers with 128, 64 and 32 feature channels, each followed by a Leaky ReLU, a $2 \times$ bilinear upsampling, and a linear layer, consistent with the LOCA configuration. We use AdamW optimizer with the fixed learning rate $10^{-4}$ and a weight decay of $10^{-2}$ . Gradient clipping with maximum norm of 0.1 is applied. URM is trained on a single NVIDIA V100 GPU with batch size of 4. + +Datasets. We evaluate our method on two mainstream datasets: FSC-147 [46] and FSCD-LVIS [35]. Specifically, FSC-147 contains 6135 images of 147 object categories, split into training, validation and test sets consisting of 3659, 1286 and 1190 images respectively. Images in FSC-147 are collected from the Internet. The sets of object categories present in each split are disjoint. FSCD-LVIS contains 6195 images spanning 372 classes, which are + +extracted from the LVIS dataset [15]. Each image annotation consists of three bounding boxes for exemplar objects within each category. + +Let $S \to \mathcal{T}$ denote the case where $S$ is the source domain and $\mathcal{T}$ is the target domain. The datasets are utilized in experiments under two cross-domain settings: the full training dataset of one is regarded as the source domain, while other entire test dataset is regarded as the target domain, i.e., FSC-147 $\rightarrow$ FSCD-LVIS and FSCD-LVIS $\rightarrow$ FSC-147. + +Evaluation metrics. We evaluate our method using two commonly employed metrics in object counting: mean absolute error (MAE) and root mean square error (RMSE), between estimated and labeled counts for each image: + +$$ +M A E = \frac {1}{N} \sum_ {i = 1} ^ {N} \left| c _ {i} - \hat {c} _ {i} \right|, R M S E = \sqrt {\frac {1}{N} \sum_ {i = 1} ^ {N} \left(c _ {i} - \hat {c} _ {i}\right) ^ {2}} \tag {6} +$$ + +where $N$ is the number of test images, $c_{i}$ is the ground truth count for the $i$ -th image, and $\hat{c}_i$ is the predicted count. Lower values of both metrics indicate better performance. + +# 4.2. Comparison with the state of the art + +In this section, we compare our URM with state-of-the-art methods on different benchmarks. The selected baselines can be grouped into two categories: + +- Vanilla Methods: FamNet [46], BMNet+ [51], CounTR [32], LOCA [10], and DAVE [40]. +- Domain Generalization Methods: ISW [9, 41], GS [34], DCCUS [12], and MPCount [41]. + +All these DG methods are originally designed or adapted for crowd counting, we add the prompt encoder after feature + +
MethodMAERMSE
Baseline30.1548.04
Language Distillation
Naive Template25.4442.59
Prompt Generator23.8341.03
Vision Distillation
[CLS] Token23.5341.12
Global Pooling22.9439.44
Mask Pooling21.8738.42
+ +extraction based on their implementation. Besides, GS is a combination of Gradient Surgery [34] and LOCA [10]. + +Table 1 presents the results of domain generalized few-shot counting. Our URM surpasses all the methods significantly, providing compelling evidence for the effectiveness of our designation. In contrast, sub-domain division methods like DCCUS and GS fail to achieve satisfactory results, as they struggle to produce meaningful sub-domains when the source domain distribution is narrow and the domain gap is large. Additionally, the whitening-based method ISW is even harmful because useful information is eliminated through the feature statistics normalization operation. Although MPCount introduced a single memory bank for regression does not require sub-domain partitioning, its generalization ability remains far inferior to our method. + +We further evaluate our method on the zero-shot setting with no user annotation, the results are shown in the right part of Table 1. We follow [10] to take a minor modification of the prompt encoder by using only the learnable embedding to represent the exemplar prompt. The results confirm that our proposed URM effectively enhances the generalization ability of prototypes, enabling accurate count estimation even in the challenging case without manually annotated exemplar. + +# 4.3. Analysis and Ablations + +Ablation of the designed elements. We first ablate the important elements of our designation under the FSC-147 (A) $\rightarrow$ FSCD-LVIS (B) few-shot setting. The baseline model shares the same architecture as URM in the inference phase but without knowledge distillation. The result shown in the first row of Table 2 demonstrates that simply altering architecture, i.e., matching with learned prototypes that lack knowledge transfer and thus have a narrow distribution, introduces no benefit for generalization. + +In contrast, the method can achieve comparable performance to the SOTA method MPCount by using only language knowledge from CLIP with a naive template. Since the naive template "A photo of {}" is too limited to fully represent a category, adding more templates and customized + +Table 2. Ablation of the designed elements. + +
α00.250.50.750.91
MAE26.5423.4423.3022.3721.8723.83
RMSE42.8239.8739.4838.9238.4241.03
+ +Table 3. Ablation of the distillation weight. + +
N123334
N232343
MAE22.7722.2321.8723.5823.45
RMSE39.6939.1238.4241.2239.38
+ +Table 4. Ablation of the number of layers, where $N_{1}$ is the universal prototypes learning layers, and $N_{2}$ is the universal representation matching layers. + +language prompts by the prompt generator further improves the model performance. + +Subsequently, incorporating visual representation via global average pooling from the 2D vision tokens output from the last layer improve performance, demonstrating that the vision representation is complementary to language knowledge. Using the [CLS] token which captures global description results in lower performance. Finally, by introducing more precise locality representation as well as noise removal through mask pooling, our model reaches a MAE of 21.87 as shown in the last row of Table 2, outperforming the baseline method with $27.5\%$ . + +Ablation of distillation weight. The distillation weight $\alpha$ in Equation (5) controls the balance between visual and language knowledge contributions. In this part, we conduct an ablation study on the weight to fully understand its impact. As shown in Table 3, when $\alpha = 0$ or $\alpha = 1$ , where only vision or language knowledge is introduced respectively, the performance is inferior to other settings, demonstrating that V-L representations complement each other. Setting $\alpha = 0.9$ yields the best performance, indicating that language representation from CLIP, which has been aligned to visual space, plays an influential role. + +Ablation of the number layers. We then ablate the hyper-parameters related to the number of layers. The universal prototypes learning layer number $N_{1}$ determines how many layers are required to absorb the universal knowledge transferred from CLIP. And the universal representation matching layer number $N_{2}$ defines how many layers are needed to learn an effective correlation map during matching in the hidden space. The results in Table 4 show that $N_{1} = N_{2} = 3$ achieves the best performance. + +Ablation of different representation sources. In this part, we investigate the effect of different representation + +
ModelArchitectureACCMAERMSE
OpenAI CLIP [43]ViT-B/1668.321.8738.42
OpenAI CLIP [43]ViT-L/1475.522.6338.95
OpenAI CLIP [43]RN50×1670.523.1539.63
OpenCLIP [8]ConvNeXt-B71.522.7839.10
SigLIP [68]ViT-B/1673.421.3337.85
EVA-02-CLIP [52]ViT-B/1674.721.8638.12
+ +learning sources in Table 5. The first row presents the results of our default setting which uses the OpenAI ViT-B/16 CLIP model. Increasing the model size leads to a performance drop due to the large dimension gap between the CLIP model and URM as seen in the second row. Besides, models with convolutional backbones, such as ResNet50 and ConvNeXt, perform poorly in comparison despite achieving higher ImageNet accuracy, suggesting that ViT architecture is better suited for representation learning in our task. Finally, the results in the last two lines indicate that using a stronger CLIP teacher model can further enhance our model's performance. However, this improvement also does not directly correlate with ImageNet accuracy, aligning with the observations in [26]. + +# 4.4. Performance on in domain benchmark + +In this section, URM is compared with the state-of-the-art methods on in domain benchmark. The three-shot evaluation results on FSC147 dataset are presented in Table 6. While demonstrating strong generalization, URM outperforms state-of-the-art few-shot counting models in terms of in-domain performance. Additionally, DAVE [40] proposed a detect-and-verify paradigm for the existing regression-based methods like LOCA to mitigate false activations on objects of other categories. Since DAVE is a general method for merging the benefits of both density and detection-based approaches, we also extend our method by incorporating this paradigm. Although inductive biases introduced by DAVE are insufficient to handle complex distribution gap in domain generalization setting (as show in Table 1), the resulting method still surpasses other methods on the traditional in domain benchmark, as demonstrated in the last row of Table 6. + +# 4.5. Generating language prompts with LVLMs + +In this section, we explore the use of large vision-language models to generate language prompts, enabling our method to train without predefined category names. However, existing proprietary and public LVLMs struggle with image classification especially in open-world setting where the class list is unknown [70], making them difficult to + +Table 5. Ablation of different universal V-L representation sources, where ACC represents the ImageNet zero-shot classification performance of the CLIP model. + +
SplitValidation setTest set
MethodMAERMSEMAERMSE
FamNet (CVPR21) [46]23.7569.0722.0899.54
BMNet+ (CVPR22) [51]15.7458.5314.6291.83
VCN (CVPR22) [45]19.3860.1518.1795.60
CountTR (BMVC22) [32]13.1349.8311.9591.23
SAFECount (WACV23) [65]15.2847.2014.3285.54
PseCo (CVPR24) [19]15.3168.3413.05112.86
LOCA (ICCV23) [10]10.2432.5610.7956.97
DAVE (CVPR24) [40]8.9128.088.6632.36
URM (Ours)10.6332.0110.1554.83
URM+DAVE [40]8.4726.518.3232.05
+ +Table 6. Performance comparison on in domain benchmark. + +
LVLMMAERMSE
Claude3 [3]22.5838.96
GPT4 [1]22.4538.75
+ +Table 7. Results of generating language prompts using LVLMs when category names are not given. + +accurately fill templates in the prompt generator. Fortunately, customized descriptive prompts with details are effective [42] as shown in Section 4.3 and Appendix. As a result, we choose Claude3-Opus [3] and GPT4 [1] to generate language prompts which have demonstrated the best open-world classification performance among existing VLMs [70]. We replace “{}” in our template prompts with “The category of the most numerous objects in the image” to prompt the VLMs. For example, “Describe what $a(n)$ {} looks like?” becomes “Describe what the category of the most numerous objects in the image looks like?”. The results in Table 7 show that our method can successfully generalize to scenarios where category names are not provided during training. + +# 5. Conclusion + +This paper proposes URM to address single domain generalization for few-shot counting with narrow source distribution. URM is motivated by the unique challenge that existing methods struggle to learn high generalization prototypes. To address this issue, we propose a universal representation matching paradigm that incorporating universal vision-language prototypes distilled from the universal representations from CLIP model, which has seen vast diversity of distributions during training. Matching with these universal prototypes leads to high generalization and superior performance. Extensive experiments on both cross and in domain settings show that URM significantly outperform the state-of-the-art. + +# References + +[1] OpenAI Josh Achiam, Steven Adler, Sandhini Agarwal, and et al. Gpt-4 technical report. 2023. 2, 5, 8 +[2] Niki Amini-Naeni, Kiana Amini-Naeni, Tengda Han, and Andrew Zisserman. Open-world text-specified object counting. arXiv preprint arXiv:2306.01851, 2023. 3 +[3] anthropic. Introducing the next generation of claude. 2024. 8 +[4] Yiqing Cai, Lianggangxu Chen, Haoyue Guan, Shaohui Lin, Changhong Lu, Changbo Wang, and Gaoqi He. Explicit invariant feature induced cross-domain crowd counting. In AAAI Conference on Artificial Intelligence, 2023. 3 +[5] Xianing Chen, Qiong Cao, Yujie Zhong, Jing Zhang, Shenghua Gao, and Dacheng Tao. Dearkd: Data-efficient early knowledge distillation for vision transformers. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12042-12052, 2022. 3 +[6] Xi Chen, Shuang Li, Ser Nam Lim, Antonio Torralba, and Hengshuang Zhao. Open-vocabulary panoptic segmentation with embedding modulation. 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 1141-1150, 2023. 5 +[7] Junting Cheng, Dong Liang, and Shan Tan. Transfer clip for generalizable image denoising. ArXiv, abs/2403.15132, 2024. 2 +[8] Mehdi Cherti, Romain Beaumont, Ross Wightman, Mitchell Wortsman, Gabriel Ilharco, Cade Gordon, Christoph Schuhmann, Ludwig Schmidt, and Jenia Jitsev. Reproducible scaling laws for contrastive language-image learning. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2818-2829, 2022. 8 +[9] Sungha Choi, Sanghun Jung, Huiwon Yun, Joanne Taery Kim, Seungryong Kim, and Jaegul Choo. Robustnet: Improving domain generalization in urban-scene segmentation via instance selective whitening. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11575-11585, 2021. 3, 6 +[10] Nikola Djukic, Alan Luke, Vitjan Zavrtanik, and Matej Kristan. A low-shot object counting network with iterative prototype adaptation. 2023 IEEE/CVF International Conference on Computer Vision (ICCV), 2022. 2, 3, 4, 6, 7, 8 +[11] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. ArXiv, abs/2010.11929, 2020. 2 +[12] Zhipeng Du, Jiankang Deng, and Miaojing Shi. Domain-general crowd counting in unseen scenarios. ArXiv, abs/2212.02573, 2022. 3, 6 +[13] Junyuan Gao, Tao Han, Yuan Yuan, and Qi Wang. Domain-adaptive crowd counting via high-quality image translation and density reconstruction. IEEE Transactions on Neural Networks and Learning Systems, 34:4803-4815, 2019. 3 +[14] Jianping Gou, B. Yu, Stephen J. Maybank, and Dacheng Tao. Knowledge distillation: A survey. International Journal of Computer Vision, 129:1789 - 1819, 2020. 3 + +[15] Agrim Gupta, Piotr Dólár, and Ross B. Girshick. Lvis: A dataset for large vocabulary instance segmentation. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5351-5359, 2019. 6 +[16] Geoffrey E. Hinton, Oriol Vinyals, and Jeffrey Dean. Distilling the knowledge in a neural network. ArXiv, abs/1503.02531, 2015. 3 +[17] Michael A. Hobley and Victor Adrian Prisacariu. Learning to count anything: Reference-less class-agnostic counting with weak supervision. ArXiv, abs/2205.10203, 2022. 2 +[18] Xiaohui Huang, Hao Zhou, Kun Yao, and Kai Han. Froster: Frozen clip is a strong teacher for open-vocabulary action recognition. ArXiv, abs/2402.03241, 2024. 2 +[19] Zhizhong Huang, Mingliang Dai, Yi Zhang, Junping Zhang, and Hongming Shan. Point, segment and count: A generalized framework for object counting. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17067-17076, 2023. 8 +[20] Qing Jiang, Feng Li, Zhaoyang Zeng, Tianhe Ren, Shilong Liu, and Lei Zhang. T-rex2: Towards generic object detection via text-visual prompt synergy. In European Conference on Computer Vision, pages 38-57. Springer, 2024. 3 +[21] Ruixiang Jiang, Lingbo Liu, and Changwen Chen. Clipcount: Towards text-guided zero-shot object counting. In Proceedings of the 31st ACM International Conference on Multimedia, pages 4535-4545, 2023. 3 +[22] Seunggu Kang, WonJun Moon, Euiyeon Kim, and Jae-Pil Heo. Vlcounter: Text-aware visual representation for zero-shot object counting. In AAAI Conference on Artificial Intelligence, 2023. 3 +[23] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dólár, and Ross B. Girshick. Segment anything. 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 3992-4003, 2023. 4 +[24] Wouter M. Kouw. An introduction to domain adaptation and transfer learning. ArXiv, abs/1812.11806, 2018. 1 +[25] Mengcheng Lan, Chaofeng Chen, Yiping Ke, Xinjiang Wang, Litong Feng, and Wayne Zhang. Clearclip: Decomposing clip representations for dense vision-language inference. ArXiv, abs/2407.12442, 2024. 2 +[26] Hugo Laurençon, Léo Tronchon, Matthieu Cord, and Victor Sanh. What matters when building vision-language models? ArXiv, abs/2405.02246, 2024. 8 +[27] Da Li, Yongxin Yang, Yi-Zhe Song, and Timothy M. Hospedales. Learning to generalize: Meta-learning for domain generalization. In AAAI Conference on Artificial Intelligence, 2017. 3 +[28] Zheng Li, Xiang Li, Xinyi Fu, Xing Zhang, Weiqiang Wang, Shuo Chen, and Jian Yang. Promptkd: Unsupervised prompt distillation for vision-language models. ArXiv, abs/2403.02781, 2024. 2 +[29] Dongze Lian, Xianing Chen, Jing Li, Weixin Luo, and Shenghua Gao. Locating and counting heads in crowds with a depth prior. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44:9056-9072, 2021. 2 + +[30] Dingkang Liang, Jiahao Xie, Zhikang Zou, Xiaoqing Ye, Wei Xu, and Xiang Bai. Crowdclip: Unsupervised crowd counting via vision-language model. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2893-2903, 2023. 3 +[31] Hui Lin, Xiaopeng Hong, and Yabin Wang. Object counting: You only need to look at one. ArXiv, abs/2112.05993, 2021. 2 +[32] Chang Liu, Yujie Zhong, Andrew Zisserman, and Weidi Xie. Countr: Transformer-based generalised visual counting. ArXiv, abs/2208.13721, 2022. 2, 3, 6, 8 +[33] Yuting Liu, Zheng Wang, Miaojing Shi, Shin'ichi Satoh, Qijun Zhao, and Hongyu Yang. Towards unsupervised crowd counting via regression-detection bi-knowledge transfer. Proceedings of the 28th ACM International Conference on Multimedia, 2020. 3 +[34] Lucas Mansilla, Rodrigo Echeveste, Diego H. Milone, and Enzo Ferrante. Domain generalization via gradient surgery. 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 6610-6618, 2021. 6, 7 +[35] Thanh Thoi Nguyen, Chau Khoa Pham, Khoi Duc Minh Nguyen, and Minh Hoai. Few-shot object counting and detection. ArXiv, abs/2207.10988, 2022. 1, 2, 6 +[36] Utkarsh Ojha, Yuheng Li, and Yong Jae Lee. Towards universal fake image detectors that generalize across generative models. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 24480-24489, 2023. 2 +[37] Roni Paiss, Ariel Ephrat, Omer Tov, Shiran Zada, Inbar Mosseri, Michal Irani, and Tali Dekel. Teaching clip to count to ten. 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 3147-3157, 2023. 3 +[38] Xingang Pan, Ping Luo, Jianping Shi, and Xiaou Tang. Two at once: Enhancing learning and generalization capacities via ibn-net. In European Conference on Computer Vision, 2018. 3 +[39] Xingang Pan, Xiaohang Zhan, Jianping Shi, Xiaou Tang, and Ping Luo. Switchable whitening for deep representation learning. 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 1863-1871, 2019. 3 +[40] Jer Pelhan, Alan Luke, Vitjan Zavrtanik, and Matej Kristan. Dave - a detect-and-verify paradigm for low-shot counting. ArXiv, abs/2404.16622, 2024. 1, 3, 6, 8 +[41] Zhuoxuan Peng and S.-H. Gary Chan. Single domain generalization for crowd counting. ArXiv, abs/2403.09124, 2024. 3, 6 +[42] Sarah Pratt, Rosanne Liu, and Ali Farhadi. What does a platypus look like? generating customized prompts for zero-shot image classification. 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 15645-15655, 2022. 2, 5, 8 +[43] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, 2021. 2, 3, 5, 8 + +[44] Viresh Ranjan and Minh Hoai. Exemplar free class agnostic counting. In Asian Conference on Computer Vision, 2022. 2 +[45] Viresh Ranjan and Minh Hoai. Vicinal counting networks. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 4220-4229, 2022. 8 +[46] Viresh Ranjan, Udbhav Sharma, Thua Nguyen, and Minh Hoai. Learning to count everything. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3393-3402, 2021. 1, 2, 3, 6, 8 +[47] Hanoona Abdul Rasheed, Muhammad Maaz, Muhammad Uzair Khattak, Salman H. Khan, and Fahad Shahbaz Khan. Bridging the gap between object and image-level representations for open-vocabulary detection. ArXiv, abs/2207.03482, 2022. 2 +[48] Shaoqing Ren, Kaiming He, Ross B. Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. IEEE Transactions on Pattern Analysis and Machine Intelligence, 39:1137-1149, 2015. 4 +[49] Tianhe Ren, Qing Jiang, Shilong Liu, Zhaoyang Zeng, Wenlong Liu, Han Gao, Hongjie Huang, Zhengyu Ma, Xiaoke Jiang, Yihao Chen, et al. Grounding dino 1.5: Advance the" edge" of open-set object detection. arXiv preprint arXiv:2405.10300, 2024. 3 +[50] Cheng Shi and Sibei Yang. Edadet: Open-vocabulary object detection using early dense alignment. 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 15678-15688, 2023. 3 +[51] Min Shi, Hao Lu, Chen Feng, Chengxin Liu, and ZHIGUO CAO. Represent, compare, and learn: A similarity-aware framework for class-agnostic counting. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9519–9528, 2022. 2, 3, 6, 8 +[52] Quan Sun, Yuxin Fang, Ledell Yu Wu, Xinlong Wang, and Yue Cao. Eva-clip: Improved training techniques for clip at scale. ArXiv, abs/2303.15389, 2023. 8 +[53] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Herv'e J'egou. Training data-efficient image transformers & distillation through attention. In International Conference on Machine Learning, 2020. 2, 3 +[54] Aayushi Tyagi, Chirag Mohapatra, Prasenjit Das, Govind Kumar Makharia, Lalita Mehra, A. P. Prathosh, Musam, and Iit Delhi. Degpr: Deep guided posterior regularization for multi-class cell detection and counting. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 23913-23923, 2023. 2 +[55] Eric Tzeng, Judy Hoffman, Kate Saenko, and Trevor Darrell. Adversarial discriminative domain adaptation. 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2962-2971, 2017. 2 +[56] Laurens van der Maaten and Geoffrey E. Hinton. Visualizing data using t-sne. Journal of Machine Learning Research, 9: 2579-2605, 2008. 4 +[57] Ashish Vaswani, Noam M. Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Neural Information Processing Systems, 2017. 4 + +[58] Mingjie Wang, Song Yuan, Zhuohang Li, Longlong Zhu, Eric Buys, and Minglun Gong. Language-guided zero-shot object counting. 2024 IEEE International Conference on Multimedia and Expo Workshops (ICMEW), pages 1-6, 2024. 3 +[59] Qi Wang, Junyu Gao, Wei Lin, and Yuan Yuan. Learning from synthetic data for crowd counting in the wild. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8190-8199, 2019. 1, 3 +[60] Longyin Wen, Dawei Du, Pengfei Zhu, Qinghua Hu, Qilong Wang, Liefeng Bo, and Siwei Lyu. Detection, tracking, and counting meets drones in crowds: A benchmark. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7808-7817, 2021. 2 +[61] Qiangqiang Wu, Jia Wan, and Antoni B. Chan. Dynamic momentum adaptation for zero-shot cross-domain crowd counting. Proceedings of the 29th ACM International Conference on Multimedia, 2021. 3 +[62] Jingyi Xu, Hieu M. Le, Vu Nguyen, Viresh Ranjan, and Dimitris Samaras. Zero-shot object counting. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15548-15557, 2023. 3 +[63] Chuanguang Yang, Zhulin An, Libo Huang, Junyu Bi, Xinqiang Yu, Hansheng Yang, Boyu Diao, and Yongjun Xu. Clip-kd: An empirical study of clip model distillation. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15952-15962, 2023. 3, 5 +[64] Lewei Yao, Renjie Pi, Jianhua Han, Xiaodan Liang, Hang Xu, Wei Zhang, Zhenguo Li, and Dan Xu. Detclipv3: Towards versatile generative open-vocabulary object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 27391-27401, 2024. 3 +[65] Zhiyuan You, Yujun Shen, Kai Yang, Wenhan Luo, X. Lu, Lei Cui, and Xinyi Le. Few-shot object counting with similarity-aware feature enhancement. 2023 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 6304-6313, 2022. 1, 2, 3, 8 +[66] Wenwen Yu, Yuliang Liu, Wei Hua, Deqiang Jiang, Bo Ren, and Xiang Bai. Turning a clip model into a scene text detector. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6978-6988, 2023. 2 +[67] Vitjan Zavrtanik, Martin Vodopivec, and Matej Kristan. A segmentation-based approach for polyp counting in the wild. Eng. Appl. Artif. Intell., 88, 2020. 2 +[68] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 11941-11952, 2023. 8 +[69] Yingying Zhang, Desen Zhou, Siqin Chen, Shenghua Gao, and Yi Ma. Single-image crowd counting via multi-column convolutional neural network. 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 589-597, 2016. 2 +[70] Yuhui Zhang, Alyssa Unell, Xiaohan Wang, Dhruba Ghosh, Yuchang Su, Ludwig Schmidt, and Serena Yeung-Levy. Why are visually-grounded language models bad at image classification? ArXiv, abs/2405.18415, 2024. 8 + +[71] Borui Zhao, Quan Cui, Renjie Song, Yiyu Qiu, and Jiajun Liang. Decoupled knowledge distillation. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11943-11952, 2022. 3 +[72] Chong Zhou, Chen Change Loy, and Bo Dai. Extract free dense labels from clip. In European Conference on Computer Vision, 2021. 2, 3, 5 +[73] Kaiyang Zhou, Ziwei Liu, Yu Qiao, Tao Xiang, and Chen Change Loy. Domain generalization: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45:4396-4415, 2021. 2 +[74] Huilin Zhu, Jingling Yuan, Zhengwei Yang, Xian Zhong, and Zheng Wang. Fine-grained fragment diffusion for cross-domain crowd counting. Proceedings of the 30th ACM International Conference on Multimedia, 2022. 1, 3 +[75] Huilin Zhu, Jingling Yuan, Xian Zhong, Zhengwei Yang, Zheng Wang, and Shengfeng He. Daot: Domain-agnostically aligned optimal transport for domain-adaptive crowd counting. Proceedings of the 31st ACM International Conference on Multimedia, 2023. 3 +[76] Huilin Zhu, Jingling Yuan, Zhengwei Yang, Yu Guo, Zheng Wang, Xian Zhong, and Shengfeng He. Zero-shot object counting with good exemplars. ArXiv, abs/2407.04948, 2024.3 \ No newline at end of file diff --git a/2025/Single Domain Generalization for Few-Shot Counting via Universal Representation Matching/images.zip b/2025/Single Domain Generalization for Few-Shot Counting via Universal Representation Matching/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..cb8b4a6c6639ead23c9cf57742749bea828a02c8 --- /dev/null +++ b/2025/Single Domain Generalization for Few-Shot Counting via Universal Representation Matching/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6543a8dbdb8491176e0057ae4dd837d0495fbe5205c001bec92cfcd218ecd4e6 +size 558907 diff --git a/2025/Single Domain Generalization for Few-Shot Counting via Universal Representation Matching/layout.json b/2025/Single Domain Generalization for Few-Shot Counting via Universal Representation Matching/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..d0e193f391ca291dc86832cd9582e79c626e0973 --- /dev/null +++ b/2025/Single Domain Generalization for Few-Shot Counting via Universal Representation Matching/layout.json @@ -0,0 +1,9313 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 131, + 103, + 463, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 103, + 463, + 140 + ], + "spans": [ + { + "bbox": [ + 131, + 103, + 463, + 140 + ], + "type": "text", + "content": "Single Domain Generalization for Few-Shot Counting via Universal Representation Matching" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 126, + 171, + 471, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 126, + 171, + 471, + 186 + ], + "spans": [ + { + "bbox": [ + 126, + 171, + 471, + 186 + ], + "type": "text", + "content": "Xianing Chen Si Huo Borui Jiang Hailin " + }, + { + "bbox": [ + 126, + 171, + 471, + 186 + ], + "type": "inline_equation", + "content": "\\mathbf{H}\\mathbf{u}^{*}" + }, + { + "bbox": [ + 126, + 171, + 471, + 186 + ], + "type": "text", + "content": " Xinghao Chen\\*" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 246, + 187, + 347, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 246, + 187, + 347, + 198 + ], + "spans": [ + { + "bbox": [ + 246, + 187, + 347, + 198 + ], + "type": "text", + "content": "Huawei Noah's Ark Lab." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 99, + 199, + 494, + 211 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 199, + 494, + 211 + ], + "spans": [ + { + "bbox": [ + 99, + 199, + 494, + 211 + ], + "type": "text", + "content": "sdxianing@gmail.com, {huosi,jiangborui,hailin.hu,xinghao.chen}@huawei.com" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 143, + 247, + 192, + 260 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 247, + 192, + 260 + ], + "spans": [ + { + "bbox": [ + 143, + 247, + 192, + 260 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 272, + 290, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 272, + 290, + 536 + ], + "spans": [ + { + "bbox": [ + 46, + 272, + 290, + 536 + ], + "type": "text", + "content": "Few-shot counting estimates the number of target objects in an image using only a few annotated exemplars. However, domain shift severely hinders existing methods to generalize to unseen scenarios. This falls into the realm of single domain generalization that remains unexplored in few-shot counting. To solve this problem, we begin by analyzing the main limitations of current methods, which typically follow a standard pipeline that extract the object prototypes from exemplars and then match them with image feature to construct the correlation map. We argue that existing methods overlook the significance of learning highly generalized prototypes. Building on this insight, we propose the first single domain generalization few-shot counting model, Universal Representation Matching, termed URM. Our primary contribution is the discovery that incorporating universal vision-language representations distilled from a large scale pretrained vision-language model into the correlation construction process substantially improves robustness to domain shifts without compromising in domain performance. As a result, URM achieves state-of-the-art performance on both in domain and the newly introduced domain generalization setting." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 558, + 128, + 570 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 558, + 128, + 570 + ], + "spans": [ + { + "bbox": [ + 47, + 558, + 128, + 570 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 578, + 287, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 578, + 287, + 685 + ], + "spans": [ + { + "bbox": [ + 46, + 578, + 287, + 685 + ], + "type": "text", + "content": "Few-shot counting (FSC) [46] is the task of estimating the number of target objects in an image using only a few annotated exemplars or without any exemplar at all. Although many proposed approaches [40, 65] have achieved satisfactory results, they experience severe performance degradation when tested on data from unseen scenarios in real-world deployment due to the domain shift [24] problem. This shift arises from the deviation from the source domain training data to unseen target domain testing data, such as" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 306, + 249, + 556, + 352 + ], + "blocks": [ + { + "bbox": [ + 306, + 249, + 556, + 352 + ], + "lines": [ + { + "bbox": [ + 306, + 249, + 556, + 352 + ], + "spans": [ + { + "bbox": [ + 306, + 249, + 556, + 352 + ], + "type": "image", + "image_path": "83f5c2b727645c1d7d481fe0dd7079105ad0d0982261f73b293682159b70057d.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 349, + 354, + 513, + 364 + ], + "lines": [ + { + "bbox": [ + 349, + 354, + 513, + 364 + ], + "spans": [ + { + "bbox": [ + 349, + 354, + 513, + 364 + ], + "type": "text", + "content": "(a) domain generalization for few-shot counting" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 313, + 369, + 542, + 447 + ], + "blocks": [ + { + "bbox": [ + 313, + 369, + 542, + 447 + ], + "lines": [ + { + "bbox": [ + 313, + 369, + 542, + 447 + ], + "spans": [ + { + "bbox": [ + 313, + 369, + 542, + 447 + ], + "type": "image", + "image_path": "265a0917b498bc31622b123b775e1bf2394dab383896a788adf3f80b8e69ff2d.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 387, + 450, + 466, + 460 + ], + "lines": [ + { + "bbox": [ + 387, + 450, + 466, + 460 + ], + "spans": [ + { + "bbox": [ + 387, + 450, + 466, + 460 + ], + "type": "text", + "content": "(b) extract-then-match" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 309, + 463, + 556, + 601 + ], + "blocks": [ + { + "bbox": [ + 309, + 463, + 556, + 601 + ], + "lines": [ + { + "bbox": [ + 309, + 463, + 556, + 601 + ], + "spans": [ + { + "bbox": [ + 309, + 463, + 556, + 601 + ], + "type": "image", + "image_path": "df92af04a82a6cebad38c03e75f74a19bd2103afcc689fa64e305cfc45fc6d02.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 365, + 604, + 496, + 614 + ], + "lines": [ + { + "bbox": [ + 365, + 604, + 496, + 614 + ], + "spans": [ + { + "bbox": [ + 365, + 604, + 496, + 614 + ], + "type": "text", + "content": "(c) universal representation matching" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 625, + 547, + 658 + ], + "lines": [ + { + "bbox": [ + 304, + 625, + 547, + 658 + ], + "spans": [ + { + "bbox": [ + 304, + 625, + 547, + 658 + ], + "type": "text", + "content": "Figure 1. Illustration of (a) domain generalization for few-shot counting, (b) the vanilla extract-then-match pipeline, and (c) our proposed universal representation matching." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "content": "changes in camera position and weather condition [59, 74], as well as biases introduced during data collection. For instance, images in the FSC147 [46] show each target object clearly while scenes in FSCD-LIVS [35] feature more oc" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 693, + 139, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 693, + 139, + 703 + ], + "spans": [ + { + "bbox": [ + 56, + 693, + 139, + 703 + ], + "type": "text", + "content": "*Corresponding authors." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 56, + 703, + 265, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 703, + 265, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 703, + 265, + 712 + ], + "type": "text", + "content": "**Code is available at https://github.com/jbr97/URM." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4639" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 263, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 263, + 83 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 263, + 83 + ], + "type": "text", + "content": "clusions and quality degradation as discussed in [35]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 86, + 287, + 253 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 86, + 287, + 253 + ], + "spans": [ + { + "bbox": [ + 46, + 86, + 287, + 253 + ], + "type": "text", + "content": "Although domain adaptation (DA) [55] has been proposed to adapt the source domain distribution to that of the target domain by finetuning models with target domain data, the process is often laborious, and target domain data is not always available in practice, especially in the class-agnostic setting of FSC [46]. In contrast, domain generalization (DG) [73] seeks to generalize the model to any unseen target domain without requiring target domain data and model updates as shown in Figure 1a, that is, to learn a generalized model by only source domain with a limited distribution. We consider in this work the unexplored and challenging field of single DG for few-shot counting to address the aforementioned challenges, where only one source domain—likely of narrow distribution—is used for training." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 255, + 287, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 255, + 287, + 338 + ], + "spans": [ + { + "bbox": [ + 46, + 255, + 287, + 338 + ], + "type": "text", + "content": "We start by analyzing why existing methods that follow the common extract-then-match pipeline [10] illustrated in Figure 1b fail. In this pipeline, exemplars prototypes are first extracted from the image feature. Next, candidates are generated by analyzing the correlation between prototypes and the image feature. Finally, the resulting correlation map is regressed into an object density map." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 45, + 340, + 288, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 340, + 288, + 483 + ], + "spans": [ + { + "bbox": [ + 45, + 340, + 288, + 483 + ], + "type": "text", + "content": "The primary distinction among these methods lies in their approaches to constructing correlation, such as attention [17, 31, 32], Siamese similarity [10, 44, 46], and so on [51, 65]. However, they all overlook the importance of prototypes in the construction process. Prototypes learned from the source domain with a narrow distribution inherently exhibit a similarly narrow distribution. Intuitively, matching with such a limited representation in the unobserved target domain with a rather broad distribution inherently leads to sub-optimal performance. Therefore, enhancing the generalization of prototypes is essential for learning a domain-general few-shot counter." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 484, + 287, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 484, + 287, + 640 + ], + "spans": [ + { + "bbox": [ + 46, + 484, + 287, + 640 + ], + "type": "text", + "content": "The contrastive language-image pre-training model (CLIP) [43] and its extensions [7, 18, 66, 72] have demonstrated remarkable generalization capacity in open-world down-stream tasks. The representation from CLIP can serve as universal representation with domain-invariant information which demonstrate extraordinary performance across several distributions owing to the vast diversity of distributions seen during training [36]. Our finding indicates that matching with the universal vision-language representation distilled from the frozen CLIP model during the correlation construction process exhibit remarkable resilience to domain shift while without in domain performance degradation. The operation is illustrated in Figure 1c." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 642, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 287, + 714 + ], + "type": "text", + "content": "Building on the discovery, we introduce the first domain-general few-shot counting model, Universal Representation Matching, termed URM. Specifically, we introduce universal vision and language prototypes to act as knowledge distiller [53] to learn universal knowledge from the large-scale pre-trained CLIP model motivated by prompt knowledge" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 545, + 120 + ], + "type": "text", + "content": "distillation [28], aiming to extend the narrow distribution of these prototypes. To adapt the V-L representations from CLIP for local objects matching, we minimally adjust the CLIP encoder while keeping its pre-trained weights frozen." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 123, + 545, + 315 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 306, + 123, + 545, + 219 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 123, + 545, + 219 + ], + "spans": [ + { + "bbox": [ + 306, + 123, + 545, + 219 + ], + "type": "text", + "content": "- For the visual representation: the class token focuses more on global image properties which may not suitable for capturing localized details [11], and the image tokens are weak in local discriminability [25]. Thus, we follow MaskCLIP [72] to obtain class-specific segmentation, allowing us to extract locality object-level representations [47] which are then used to distill region-specific information into the visual prototypes." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 219, + 545, + 315 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 219, + 545, + 315 + ], + "spans": [ + { + "bbox": [ + 306, + 219, + 545, + 315 + ], + "type": "text", + "content": "- For the language representation: we follow CuPL [42] to generate diverse customized prompts that leverage knowledge contained in LLMs [1] as well as hand-written templates, to produce numerous descriptive sentences that capture important discriminative characteristics of the object categories. These prompts are subsequently encoded by the language tower to obtain universal language embeddings for distillation." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 304, + 317, + 545, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 317, + 545, + 389 + ], + "spans": [ + { + "bbox": [ + 304, + 317, + 545, + 389 + ], + "type": "text", + "content": "The learned universal V-L prototypes are then matched with the image feature through cross attention to construct the correlation map. As the distillation operation is conducted only during the training phase, URM remains as efficient as other methods that do not require external knowledge at inference time." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 390, + 545, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 390, + 545, + 449 + ], + "spans": [ + { + "bbox": [ + 304, + 390, + 545, + 449 + ], + "type": "text", + "content": "To validate URM, we conduct extensive experiments on both cross and in domain setting on FSC147 [46] and FSCD-LVIS [35] datasets in both few-shot and zero-shot scenarios. The results demonstrate that URM performs excellently on both intra-dataset and cross domain settings." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 317, + 450, + 523, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 450, + 523, + 461 + ], + "spans": [ + { + "bbox": [ + 317, + 450, + 523, + 461 + ], + "type": "text", + "content": "The main contributions of our work are as follows:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 465, + 545, + 632 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 306, + 465, + 545, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 465, + 545, + 525 + ], + "spans": [ + { + "bbox": [ + 306, + 465, + 545, + 525 + ], + "type": "text", + "content": "- We introduce a new challenging setting of the single domain generalized few-shot counting task. And we point out that existing paradigm's inability to accurately count objects in new scenes stems from the learned narrow distribution prototypes." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 525, + 545, + 595 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 525, + 545, + 595 + ], + "spans": [ + { + "bbox": [ + 306, + 525, + 545, + 595 + ], + "type": "text", + "content": "- We introduce URM, the first narrow source domain—general few-shot counting model. Specifically, we propose to match universal V-L representations transferred from CLIP through knowledge distillation to construct the correlation map, a novel approach not explored in previous works." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 597, + 545, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 597, + 545, + 632 + ], + "spans": [ + { + "bbox": [ + 306, + 597, + 545, + 632 + ], + "type": "text", + "content": "- Our proposed URM outperforms state-of-the-art methods by a large margin in both the introduced narrow source domain general setting as well as in domain benchmark." + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 645, + 397, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 645, + 397, + 658 + ], + "spans": [ + { + "bbox": [ + 306, + 645, + 397, + 658 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 665, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 714 + ], + "type": "text", + "content": "Few-shot Counting. Object counting refers to the task of estimating the number of specific objects present in an image, typically limited to a few predefined categories, e.g., human heads [29, 69], cells [54], cars [60], and polyps [67]." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4640" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 264 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 264 + ], + "type": "text", + "content": "Few-shot counting addresses this limitation by given only a few annotated exemplars or without any exemplar at all. FamNet [46] proposed a pipeline that extracts appearance-based queries from exemplars and matches them with image feature to infer the object counts. Then, BMNet+ [51] enhanced localization by jointly learning representation with a non-linear similarity metric. CounTR [32] introduced cross attention-based interaction module to fuse image and exemplar features. SAFE-Count [65] incorporated a feature enhancement module. LOCA [10] further refined this pipeline by iteratively adapting object prototypes based on exemplar appearance and shape. DAVE [40] proposed a post-process method for LOCA by generating a high-recall detection set and verifying the results to identify and remove outliers. All of these methods follow the same extract-then-match paradigm and struggle with domain generalization." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 279, + 289, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 279, + 289, + 447 + ], + "spans": [ + { + "bbox": [ + 46, + 279, + 289, + 447 + ], + "type": "text", + "content": "CLIP-guided Counting. Some studies also explored to count by leveraging CLIP. For instance, [37] extended CLIP's capabilities to counting but still failed when objects number is large. CrowdCLIP [30] mapped crowd patches to count text by constructing ranking-based text prompts to match the size-sorted crowd patches when category names were provided. Other works [2, 21, 22, 58, 62, 76] explored enumerating objects given class names, creating a link between object categories and visual representations. Besides, some open-set methods [20, 49, 64] achieve competitive results by leveraging CLIP as a bridge between V-L representations for interactive object detection. However, they require much more computational power and data for training, and their training classes cover the test classes." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 462, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 462, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 462, + 289, + 715 + ], + "type": "text", + "content": "Single Domain Generalization. Since domain shift problem in FSC remains unexplored, we mainly focus on related approaches for crowd counting. Although some domain adaptation techniques exist [61, 74, 75], they all inadequate for real-world application due to the lack of target data [4, 13, 33, 59, 74]. Domain generalization methods are primarily built on meta-learning algorithm MLDG [27], which employs meta-train and meta-test set to simulate the domain shift between training and testing phases in reality to learn domain-invariant features. For instance, DCCUS [12] proposed a dynamic sub-domain division scheme, dividing the source domain into multiple sub-domains. Whitening-based methods [9, 38, 39] remove domain-specific style information by eliminating feature correlation to ensure each feature has unit variance. However, these algorithms require a sufficiently broad source domain distribution. MPCount [41] proposed storing diverse density values for regression and reconstruct domain-invariant features using memory bank, eliminating the need for sub-domain partitioning. However, MPCount is still limited by the narrow source training data without accessing external knowledge." + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 307, + 71, + 385, + 148 + ], + "blocks": [ + { + "bbox": [ + 307, + 71, + 385, + 148 + ], + "lines": [ + { + "bbox": [ + 307, + 71, + 385, + 148 + ], + "spans": [ + { + "bbox": [ + 307, + 71, + 385, + 148 + ], + "type": "image", + "image_path": "23eea1e33e71f99ff99956f4f1862073a3c957cc97e35929cf6f2945c8903480.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 342, + 150, + 352, + 159 + ], + "lines": [ + { + "bbox": [ + 342, + 150, + 352, + 159 + ], + "spans": [ + { + "bbox": [ + 342, + 150, + 352, + 159 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 388, + 71, + 463, + 148 + ], + "blocks": [ + { + "bbox": [ + 388, + 71, + 463, + 148 + ], + "lines": [ + { + "bbox": [ + 388, + 71, + 463, + 148 + ], + "spans": [ + { + "bbox": [ + 388, + 71, + 463, + 148 + ], + "type": "image", + "image_path": "404c35f5480f00a15ab4815d4f843d1324be47abe3b69028de41f125a9e7f46e.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 421, + 150, + 432, + 159 + ], + "lines": [ + { + "bbox": [ + 421, + 150, + 432, + 159 + ], + "spans": [ + { + "bbox": [ + 421, + 150, + 432, + 159 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 467, + 72, + 544, + 148 + ], + "blocks": [ + { + "bbox": [ + 467, + 72, + 544, + 148 + ], + "lines": [ + { + "bbox": [ + 467, + 72, + 544, + 148 + ], + "spans": [ + { + "bbox": [ + 467, + 72, + 544, + 148 + ], + "type": "image", + "image_path": "0a72f5c49f4abd72a16e393c8477a8c144ae88ab1b94f0dbb155ba090938a79f.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 501, + 150, + 511, + 159 + ], + "lines": [ + { + "bbox": [ + 501, + 150, + 511, + 159 + ], + "spans": [ + { + "bbox": [ + 501, + 150, + 511, + 159 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 169, + 547, + 236 + ], + "lines": [ + { + "bbox": [ + 304, + 169, + 547, + 236 + ], + "spans": [ + { + "bbox": [ + 304, + 169, + 547, + 236 + ], + "type": "text", + "content": "Figure 2. t-SNE visualization of the prototypes feature space for different categories from FSC147 by (a) the vanilla paradigm trained on FSC147, (b) the vanilla paradigm trained on FSCD-LVIS, and (c) our proposed URM trained on FSCD-LVIS. Note that the visualization is conduct on the test set where the object categories are disjoint from the train set. Best viewed in color." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 257, + 547, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 257, + 547, + 354 + ], + "spans": [ + { + "bbox": [ + 304, + 257, + 547, + 354 + ], + "type": "text", + "content": "Knowledge Distillation (KD). [16] proposed to optimize a student model under the effective information transfer and supervision of a teacher model or ensemble. Although several methods have been proposed to learn knowledge from teacher network with high fidelity [14, 53, 71], we simply follow CLIP-KD [63] to use feature mimicry learning [5] which has been proved to be the most effective approach for CLIP knowledge distillation." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 306, + 364, + 362, + 376 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 364, + 362, + 376 + ], + "spans": [ + { + "bbox": [ + 306, + 364, + 362, + 376 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 384, + 547, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 384, + 547, + 433 + ], + "spans": [ + { + "bbox": [ + 304, + 384, + 547, + 433 + ], + "type": "text", + "content": "In this section, we first recap the preliminaries of CLIP, and then analyze why prior methods fail to generalize. Afterward, we introduce our proposed universal representation matching model." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 447, + 547, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 447, + 547, + 651 + ], + "spans": [ + { + "bbox": [ + 304, + 447, + 547, + 651 + ], + "type": "text", + "content": "Preliminary. CLIP [43], trained on large-scale image-text pairs by image-level contrastive learning, has well demonstrated the potential for learning transferable knowledge and open-set visual concepts. CLIP consists of a pair of image encoder " + }, + { + "bbox": [ + 304, + 447, + 547, + 651 + ], + "type": "inline_equation", + "content": "\\mathcal{V}(\\cdot)" + }, + { + "bbox": [ + 304, + 447, + 547, + 651 + ], + "type": "text", + "content": " and text encoder " + }, + { + "bbox": [ + 304, + 447, + 547, + 651 + ], + "type": "inline_equation", + "content": "\\mathcal{T}(\\cdot)" + }, + { + "bbox": [ + 304, + 447, + 547, + 651 + ], + "type": "text", + "content": ", both of which are jointly trained to map input image and text into an unified representation space. However, since CLIP is designed for image-level prediction, extracting local patch-level predictions from CLIP is non-trivial [50]. MaskCLIP [72] modifies the image encoder by removing the last query and key embedding layers, as well as reformulating the value-embedding layer and the last linear layer into two respective convolutional layers, while keeping the text encoder unchanged. The text encoder then processes language prompts containing target classes, and the resulting text embedding for each class is used as a classifier to produce pixel-level mask predictions." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 715 + ], + "type": "text", + "content": "Motivation. Since the existing extract-then-match paradigm fails to generalize, we begin by analyzing the reason behind this undesirable phenomenon, which motivates our proposed solution. After training on the source" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "4641" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 70, + 71, + 525, + 332 + ], + "blocks": [ + { + "bbox": [ + 70, + 71, + 525, + 332 + ], + "lines": [ + { + "bbox": [ + 70, + 71, + 525, + 332 + ], + "spans": [ + { + "bbox": [ + 70, + 71, + 525, + 332 + ], + "type": "image", + "image_path": "fa450d996fe984d55ea204b8cd57752a267250f1f811ca57cf551549890aee8c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 340, + 548, + 374 + ], + "lines": [ + { + "bbox": [ + 46, + 340, + 548, + 374 + ], + "spans": [ + { + "bbox": [ + 46, + 340, + 548, + 374 + ], + "type": "text", + "content": "Figure 3. The framework of our proposed URM. The inference architecture is depicted in the gray part, where the learned prototypes are matched with the image feature through cross attention. The yellow part illustrates the universal V-L representations obtained from CLIP, which are distilled into the prototypes exclusively during the training phase." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 394, + 289, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 394, + 289, + 622 + ], + "spans": [ + { + "bbox": [ + 46, + 394, + 289, + 622 + ], + "type": "text", + "content": "domain, we use t-SNE [56] to visualize the feature space of prototypes for different categories on the test set of FSC147 learned using both vanilla paradigm and URM. One noticeable observation is the clear boundary in Figure 2a using vanilla paradigm trained on FSC147, which corresponds to its well performance in the intra-dataset setting, even when the categories are disjoint from those in the training set. In contrary, Figure 2b shows an unclear boundary on the target domain when the traditional paradigm is trained on dataset with large doamin gap, aligning with a significant drop in accuracy in the cross-domain setting. The phenomenon suggests that simply learning prototypes within a narrow distribution is suboptimal for generalization. The key, we believe, is that the matching process should occur in a feature space with well-defined open-set visual concept boundaries. As shown in Figure 2c, prototypes from our URM—transferred from CLIP representation—better separate objects with unseen categories in unseen scenarios, supporting our hypothesis." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 634, + 238, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 634, + 238, + 647 + ], + "spans": [ + { + "bbox": [ + 47, + 634, + 238, + 647 + ], + "type": "text", + "content": "3.1. Universal Representation Matching" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "content": "Overview. The overall framework of our proposed URM is shown in Figure 3. The input image " + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{I} \\in \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "content": " is first encoded by the backbone network. Multi-scale features are extracted from the last three blocks, then resized to the same size of " + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "inline_equation", + "content": "h \\times w" + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "content": " and reduced to " + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 46, + 653, + 288, + 715 + ], + "type": "text", + "content": " channels using a" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 394, + 547, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 394, + 547, + 562 + ], + "spans": [ + { + "bbox": [ + 304, + 394, + 547, + 562 + ], + "type": "text", + "content": "linear layer after concatenation. Next, a set of " + }, + { + "bbox": [ + 304, + 394, + 547, + 562 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 304, + 394, + 547, + 562 + ], + "type": "text", + "content": " selected exemplar bounding boxes are encoded into the image feature " + }, + { + "bbox": [ + 304, + 394, + 547, + 562 + ], + "type": "inline_equation", + "content": "\\mathbf{f} \\in \\mathbb{R}^{h \\times w \\times d}" + }, + { + "bbox": [ + 304, + 394, + 547, + 562 + ], + "type": "text", + "content": " via a prompt encoder. After that, the proposed universal vision and language prototypes " + }, + { + "bbox": [ + 304, + 394, + 547, + 562 + ], + "type": "inline_equation", + "content": "p_v" + }, + { + "bbox": [ + 304, + 394, + 547, + 562 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 394, + 547, + 562 + ], + "type": "inline_equation", + "content": "p_l \\in \\mathbb{R}^{n \\times d}" + }, + { + "bbox": [ + 304, + 394, + 547, + 562 + ], + "type": "text", + "content": " are initialized as learnable embeddings, and iteratively updated through interaction with the encoded image feature " + }, + { + "bbox": [ + 304, + 394, + 547, + 562 + ], + "type": "inline_equation", + "content": "f^E \\in \\mathbb{R}^{hw \\times d}" + }, + { + "bbox": [ + 304, + 394, + 547, + 562 + ], + "type": "text", + "content": " using " + }, + { + "bbox": [ + 304, + 394, + 547, + 562 + ], + "type": "inline_equation", + "content": "N_1" + }, + { + "bbox": [ + 304, + 394, + 547, + 562 + ], + "type": "text", + "content": " cross attention layers [57]. During this process, the resulting prototypes are distilled from the universal V-L representations obtained from the frozen CLIP model, respectively. Finally, the concatenated universal prototypes are correlated with " + }, + { + "bbox": [ + 304, + 394, + 547, + 562 + ], + "type": "inline_equation", + "content": "f^E" + }, + { + "bbox": [ + 304, + 394, + 547, + 562 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 304, + 394, + 547, + 562 + ], + "type": "inline_equation", + "content": "N_2" + }, + { + "bbox": [ + 304, + 394, + 547, + 562 + ], + "type": "text", + "content": " cross attention layers to obtain a correlation map, which is then regressed into a density map " + }, + { + "bbox": [ + 304, + 394, + 547, + 562 + ], + "type": "inline_equation", + "content": "\\mathbf{R} \\in \\mathbb{R}^{H \\times W}" + }, + { + "bbox": [ + 304, + 394, + 547, + 562 + ], + "type": "text", + "content": ", whose number is summed to the count estimation, i.e., " + }, + { + "bbox": [ + 304, + 394, + 547, + 562 + ], + "type": "inline_equation", + "content": "c = \\text{sum}(\\mathbf{R})" + }, + { + "bbox": [ + 304, + 394, + 547, + 562 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "content": "Prompt Encoder. Given the extracted image feature " + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{f}" + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "content": ", we use a two-way attention mechanism [23] to encode exemplar prompts into " + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{f}" + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "content": ", to indicate the object category to be counted. The exemplars are represented as a concatenated embedding composed of three elements: the height and width " + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "inline_equation", + "content": "R^{n\\times 2}" + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "content": " of bounding boxes " + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "inline_equation", + "content": "R^{n\\times 4}" + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "content": " which are mapped to shape embedding " + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "inline_equation", + "content": "R^{n\\times d}" + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "content": " by three linear layers with ReLU activation follow [10], a learnable embedding group " + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "inline_equation", + "content": "R^{n\\times d}" + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "content": ", and appearance tokens extracted from " + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{f}" + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "content": " using ROI Pooling [48]. As illustrated in Figure 4, cross attention operations are then applied iteratively to update both the em" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4642" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 72, + 74, + 267, + 199 + ], + "blocks": [ + { + "bbox": [ + 72, + 74, + 267, + 199 + ], + "lines": [ + { + "bbox": [ + 72, + 74, + 267, + 199 + ], + "spans": [ + { + "bbox": [ + 72, + 74, + 267, + 199 + ], + "type": "image", + "image_path": "b00d49cefa4f5f9a36d7ed8dcb784fbfa5ef30127c964656e91f30fadce26279.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 81, + 213, + 251, + 224 + ], + "lines": [ + { + "bbox": [ + 81, + 213, + 251, + 224 + ], + "spans": [ + { + "bbox": [ + 81, + 213, + 251, + 224 + ], + "type": "text", + "content": "Figure 4. Illustration of the prompt encoder." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 247, + 286, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 247, + 286, + 270 + ], + "spans": [ + { + "bbox": [ + 46, + 247, + 286, + 270 + ], + "type": "text", + "content": "bedding and image feature, resulting in the category-aware image feature " + }, + { + "bbox": [ + 46, + 247, + 286, + 270 + ], + "type": "inline_equation", + "content": "f^{E}" + }, + { + "bbox": [ + 46, + 247, + 286, + 270 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 287, + 286, + 359 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 287, + 286, + 359 + ], + "spans": [ + { + "bbox": [ + 46, + 287, + 286, + 359 + ], + "type": "text", + "content": "Universal Vision Representation Learning. Since CLIP trained with image-level supervision lacks precise localization of objects and includes noise on non-relevant areas in its visual representation, we follow MaskCLIP [72] to reformulate the value-embedding layer of image encoder " + }, + { + "bbox": [ + 46, + 287, + 286, + 359 + ], + "type": "inline_equation", + "content": "\\mathcal{V}(\\cdot)" + }, + { + "bbox": [ + 46, + 287, + 286, + 359 + ], + "type": "text", + "content": " to produce a dense attentive segmentation " + }, + { + "bbox": [ + 46, + 287, + 286, + 359 + ], + "type": "inline_equation", + "content": "\\mathcal{M} \\in \\mathbb{R}^{h \\times w}" + }, + { + "bbox": [ + 46, + 287, + 286, + 359 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 360, + 286, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 360, + 286, + 418 + ], + "spans": [ + { + "bbox": [ + 46, + 360, + 286, + 418 + ], + "type": "text", + "content": "Then, the universal vision representation " + }, + { + "bbox": [ + 46, + 360, + 286, + 418 + ], + "type": "inline_equation", + "content": "r_v" + }, + { + "bbox": [ + 46, + 360, + 286, + 418 + ], + "type": "text", + "content": " containing fine-grained local image semantics is obtained by applying Mask Pooling [6] between the 2D visual embeddings of the image output from the last layer of CLIP vision encoder and the class-aware mask " + }, + { + "bbox": [ + 46, + 360, + 286, + 418 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 46, + 360, + 286, + 418 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 111, + 434, + 286, + 447 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 434, + 286, + 447 + ], + "spans": [ + { + "bbox": [ + 111, + 434, + 286, + 447 + ], + "type": "interline_equation", + "content": "r _ {v} = \\operatorname {M a s k P o o l} (\\mathcal {V} (\\mathbf {I}), \\mathcal {M}) \\tag {1}", + "image_path": "dd56d6c7f33a95a79bf71a9982bceb46b21c027f3b78d842ebfe988d16d82479.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 453, + 286, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 453, + 286, + 501 + ], + "spans": [ + { + "bbox": [ + 47, + 453, + 286, + 501 + ], + "type": "text", + "content": "Additionally, the vision prototype " + }, + { + "bbox": [ + 47, + 453, + 286, + 501 + ], + "type": "inline_equation", + "content": "p_v" + }, + { + "bbox": [ + 47, + 453, + 286, + 501 + ], + "type": "text", + "content": " initialized as a learnable embedding " + }, + { + "bbox": [ + 47, + 453, + 286, + 501 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{n \\times d}" + }, + { + "bbox": [ + 47, + 453, + 286, + 501 + ], + "type": "text", + "content": " is iteratively updated through " + }, + { + "bbox": [ + 47, + 453, + 286, + 501 + ], + "type": "inline_equation", + "content": "N_1" + }, + { + "bbox": [ + 47, + 453, + 286, + 501 + ], + "type": "text", + "content": " cross attention layers with the encoded image feature, where " + }, + { + "bbox": [ + 47, + 453, + 286, + 501 + ], + "type": "inline_equation", + "content": "q = p_v" + }, + { + "bbox": [ + 47, + 453, + 286, + 501 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 453, + 286, + 501 + ], + "type": "inline_equation", + "content": "k = v = f^E" + }, + { + "bbox": [ + 47, + 453, + 286, + 501 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 502, + 286, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 502, + 286, + 550 + ], + "spans": [ + { + "bbox": [ + 46, + 502, + 286, + 550 + ], + "type": "text", + "content": "Since distribution of the prototype learned from a narrow source domain is also narrow, we adopt feature mimicry loss following CLIP-KD [63] to transfer universal representation knowledge into the visual prototypes:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 99, + 560, + 286, + 594 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 560, + 286, + 594 + ], + "spans": [ + { + "bbox": [ + 99, + 560, + 286, + 594 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {V - K D} = \\frac {1}{| \\mathcal {B} |} \\sum_ {k = 1} ^ {| \\mathcal {B} |} \\| r _ {v k} - p _ {v k} \\| \\tag {2}", + "image_path": "65824a58d0ad2362df6e5ab61db30526dfda448bb5a456ad8423661ae851c8a3.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 601, + 286, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 601, + 286, + 648 + ], + "spans": [ + { + "bbox": [ + 46, + 601, + 286, + 648 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 601, + 286, + 648 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 46, + 601, + 286, + 648 + ], + "type": "text", + "content": " is the training batch, " + }, + { + "bbox": [ + 46, + 601, + 286, + 648 + ], + "type": "inline_equation", + "content": "r_v" + }, + { + "bbox": [ + 46, + 601, + 286, + 648 + ], + "type": "text", + "content": " is the universal vision representation, and " + }, + { + "bbox": [ + 46, + 601, + 286, + 648 + ], + "type": "inline_equation", + "content": "p_v" + }, + { + "bbox": [ + 46, + 601, + 286, + 648 + ], + "type": "text", + "content": " is the universal vision prototype. For simplicity, the linear projection layer for matching the dimension size of the CLIP embedding is omitted." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 665, + 286, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 665, + 286, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 665, + 286, + 714 + ], + "type": "text", + "content": "Universal Language Representation Learning. Handwritten templates like “A photo of {}” completed with a category name are limited in describing details for fine-grained embedding [42]. Besides, writing high-performing" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 72, + 545, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 120 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 120 + ], + "type": "text", + "content": "prompt templates is labor-intensive and requires prior information about dataset contents. For example, ImageNet hand-written prompts [43] like “A toy {}” and “A sculpture of {}” do not generalize well to other datasets like FSC-147." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 120, + 545, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 120, + 545, + 191 + ], + "spans": [ + { + "bbox": [ + 305, + 120, + 545, + 191 + ], + "type": "text", + "content": "We thus follow CuPL [42] to further extend templates with customized prompts generated by LLM that contain important discriminating characteristics which can best represent an object category in natural language. The templates and prompts for LLM, i.e., GPT-4 [1] in this work, are detailed in Appendix." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 305, + 192, + 545, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 192, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 305, + 192, + 545, + 239 + ], + "type": "text", + "content": "The generated prompts for each category are then encoded by CLIP's text encoder " + }, + { + "bbox": [ + 305, + 192, + 545, + 239 + ], + "type": "inline_equation", + "content": "\\mathcal{T}(\\cdot)" + }, + { + "bbox": [ + 305, + 192, + 545, + 239 + ], + "type": "text", + "content": ". The corresponding text embeddings are averaged to produce the resulting universal text representation " + }, + { + "bbox": [ + 305, + 192, + 545, + 239 + ], + "type": "inline_equation", + "content": "r_t" + }, + { + "bbox": [ + 305, + 192, + 545, + 239 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 239, + 545, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 239, + 545, + 287 + ], + "spans": [ + { + "bbox": [ + 305, + 239, + 545, + 287 + ], + "type": "text", + "content": "Similar to vision prototype learning, the language prototype " + }, + { + "bbox": [ + 305, + 239, + 545, + 287 + ], + "type": "inline_equation", + "content": "p_l" + }, + { + "bbox": [ + 305, + 239, + 545, + 287 + ], + "type": "text", + "content": " is iteratively updated through interactions with " + }, + { + "bbox": [ + 305, + 239, + 545, + 287 + ], + "type": "inline_equation", + "content": "f^E" + }, + { + "bbox": [ + 305, + 239, + 545, + 287 + ], + "type": "text", + "content": " using the same cross attention layers. And the language distillation loss is given by:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 363, + 296, + 545, + 328 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 363, + 296, + 545, + 328 + ], + "spans": [ + { + "bbox": [ + 363, + 296, + 545, + 328 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {L - K D} = \\frac {1}{\\mathcal {B}} \\sum_ {k = 1} ^ {\\mathcal {B}} \\| r _ {l k} - p _ {l k} \\| \\tag {3}", + "image_path": "22ddd015be3ccec72d6803d06a604a3a1e9cf84331c0488e37b5844c9bd5190c.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 331, + 545, + 367 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 331, + 545, + 367 + ], + "spans": [ + { + "bbox": [ + 305, + 331, + 545, + 367 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 331, + 545, + 367 + ], + "type": "inline_equation", + "content": "r_l" + }, + { + "bbox": [ + 305, + 331, + 545, + 367 + ], + "type": "text", + "content": " is the universal language representation, " + }, + { + "bbox": [ + 305, + 331, + 545, + 367 + ], + "type": "inline_equation", + "content": "p_l" + }, + { + "bbox": [ + 305, + 331, + 545, + 367 + ], + "type": "text", + "content": " is the universal language prototype. The linear projection layer is also applied to the language prototype." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 380, + 545, + 488 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 380, + 545, + 488 + ], + "spans": [ + { + "bbox": [ + 305, + 380, + 545, + 488 + ], + "type": "text", + "content": "Universal Representation Matching. After obtaining the universal V-L prototypes for the object category in the image, we can construct a correlation between the concatenated prototypes and image feature. Although various matching approaches have been proposed as discussed, we focus on the essential matching elements within the correlation construction process. Specifically, we apply " + }, + { + "bbox": [ + 305, + 380, + 545, + 488 + ], + "type": "inline_equation", + "content": "N_{2}" + }, + { + "bbox": [ + 305, + 380, + 545, + 488 + ], + "type": "text", + "content": " cross attention layers with " + }, + { + "bbox": [ + 305, + 380, + 545, + 488 + ], + "type": "inline_equation", + "content": "\\mathbf{q} = f^{E}, \\mathbf{k} = \\mathbf{v} = \\mathrm{concat}(p_{v}, p_{l})" + }, + { + "bbox": [ + 305, + 380, + 545, + 488 + ], + "type": "text", + "content": " to obtain the correlation map " + }, + { + "bbox": [ + 305, + 380, + 545, + 488 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{hw \\times d}" + }, + { + "bbox": [ + 305, + 380, + 545, + 488 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 305, + 488, + 545, + 512 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 488, + 545, + 512 + ], + "spans": [ + { + "bbox": [ + 305, + 488, + 545, + 512 + ], + "type": "text", + "content": "Finally, a regression head predict the final 2D density map " + }, + { + "bbox": [ + 305, + 488, + 545, + 512 + ], + "type": "inline_equation", + "content": "\\mathbf{R} \\in \\mathbb{R}^{H \\times W}" + }, + { + "bbox": [ + 305, + 488, + 545, + 512 + ], + "type": "text", + "content": " from the correlation map." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 305, + 525, + 545, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 525, + 545, + 562 + ], + "spans": [ + { + "bbox": [ + 305, + 525, + 545, + 562 + ], + "type": "text", + "content": "Loss Function. The model is trained with " + }, + { + "bbox": [ + 305, + 525, + 545, + 562 + ], + "type": "inline_equation", + "content": "\\ell_2" + }, + { + "bbox": [ + 305, + 525, + 545, + 562 + ], + "type": "text", + "content": " loss between the predicted density map " + }, + { + "bbox": [ + 305, + 525, + 545, + 562 + ], + "type": "inline_equation", + "content": "\\mathbf{R}" + }, + { + "bbox": [ + 305, + 525, + 545, + 562 + ], + "type": "text", + "content": " and the ground-truth map " + }, + { + "bbox": [ + 305, + 525, + 545, + 562 + ], + "type": "inline_equation", + "content": "\\mathbf{G}" + }, + { + "bbox": [ + 305, + 525, + 545, + 562 + ], + "type": "text", + "content": ", normalized by the number of objects:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 349, + 571, + 545, + 604 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 349, + 571, + 545, + 604 + ], + "spans": [ + { + "bbox": [ + 349, + 571, + 545, + 604 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {d e n s i t y}} = \\frac {1}{2 \\mathcal {B}} \\sum_ {k = 1} ^ {\\mathcal {B}} \\frac {1}{N _ {k}} \\| \\boldsymbol {G} _ {k} - \\boldsymbol {R} _ {k} \\| _ {2} ^ {2} \\tag {4}", + "image_path": "6da139f6d13759b061f94a66385e27e9e6d1fb75b9d310e6b8e61feb3e6eb15e.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 305, + 605, + 545, + 642 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 605, + 545, + 642 + ], + "spans": [ + { + "bbox": [ + 305, + 605, + 545, + 642 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 605, + 545, + 642 + ], + "type": "inline_equation", + "content": "N_{k}" + }, + { + "bbox": [ + 305, + 605, + 545, + 642 + ], + "type": "text", + "content": " represents the number of objects in image " + }, + { + "bbox": [ + 305, + 605, + 545, + 642 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 305, + 605, + 545, + 642 + ], + "type": "text", + "content": ". The overall training loss is thus a combination of the density and the distillation losses from Equation (2) (3):" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 332, + 659, + 545, + 673 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 332, + 659, + 545, + 673 + ], + "spans": [ + { + "bbox": [ + 332, + 659, + 545, + 673 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {\\text {d e n s i t y}} + \\alpha \\mathcal {L} _ {V - K D} + (1 - \\alpha) \\mathcal {L} _ {L - K D} \\tag {5}", + "image_path": "0fec8abdb84fc27a5314e13ab375eadd1ae654cd114c58257fb73a814ba86973.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 305, + 677, + 545, + 713 + ], + "type": "text", + "content": " is the distillation weight parameter to balance the contributions of the vision and language representations in the universal prototypes." + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4643" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 91, + 69, + 504, + 256 + ], + "blocks": [ + { + "bbox": [ + 91, + 69, + 504, + 256 + ], + "lines": [ + { + "bbox": [ + 91, + 69, + 504, + 256 + ], + "spans": [ + { + "bbox": [ + 91, + 69, + 504, + 256 + ], + "type": "table", + "html": "
SettingFew-shotZero-shot
Source → TargetA → BB → AA → BB → A
MetricMAERMSEMAERMSEMAERMSEMAERMSE
Vanilla Methods
FamNet (CVPR21) [46]29.9848.9542.74118.47////
BMNet+ (CVPR22) [51]37.8668.9436.53116.02////
CounTR (BMVC22) [32]34.7654.3933.24107.34////
LOCA (ICCV23) [10]30.6344.3329.1388.0631.1746.9834.13113.72
DAVE (CVPR24) [40]29.6745.2829.5186.1330.4247.0335.40115.81
Domain Generalization Methods
ISW (CVPR21) [9]35.9551.8336.41124.09////
GS (ICCV21) [34]30.5547.5729.8398.42////
DCCUS (AAAI23) [12]29.0243.5528.4683.3230.3944.8734.29109.20
MPCount (CVPR24) [41]25.1141.3222.0780.1727.6844.5830.99102.18
URM (Ours)21.8738.4221.1773.4223.5439.9327.4994.57
", + "image_path": "af14cdc47c1442473edf412f0833844589fa15844d020cdc6a63473597e7c041.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 122, + 263, + 471, + 275 + ], + "lines": [ + { + "bbox": [ + 122, + 263, + 471, + 275 + ], + "spans": [ + { + "bbox": [ + 122, + 263, + 471, + 275 + ], + "type": "text", + "content": "Table 1. Comparison with the state-of-the-art methods on FSC-147 (A) and FSCD-LVIS (B)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 47, + 295, + 128, + 308 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 295, + 128, + 308 + ], + "spans": [ + { + "bbox": [ + 47, + 295, + 128, + 308 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 315, + 288, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 315, + 288, + 363 + ], + "spans": [ + { + "bbox": [ + 46, + 315, + 288, + 363 + ], + "type": "text", + "content": "In this section, we evaluate the effectiveness of URM, demonstrating how universal representation matching enhances the generalization and performance of few-shot counting models." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 369, + 168, + 383 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 369, + 168, + 383 + ], + "spans": [ + { + "bbox": [ + 47, + 369, + 168, + 383 + ], + "type": "text", + "content": "4.1. Experimental setting" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 388, + 288, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 388, + 288, + 604 + ], + "spans": [ + { + "bbox": [ + 46, + 388, + 288, + 604 + ], + "type": "text", + "content": "Implementation details. We first resize input images to " + }, + { + "bbox": [ + 46, + 388, + 288, + 604 + ], + "type": "inline_equation", + "content": "H = W = 512" + }, + { + "bbox": [ + 46, + 388, + 288, + 604 + ], + "type": "text", + "content": " and apply standard data augmentations such as tiling, horizontal flipping and color jitter. The model uses an Imagenet pre-trained Resnet50 backbone, with features from the final three blocks upsampled to " + }, + { + "bbox": [ + 46, + 388, + 288, + 604 + ], + "type": "inline_equation", + "content": "h = w = 64" + }, + { + "bbox": [ + 46, + 388, + 288, + 604 + ], + "type": "text", + "content": ". The backbone network parameters are frozen follow [10, 46], while all other parameters are trained for 150 epochs. The resulting features are projected into " + }, + { + "bbox": [ + 46, + 388, + 288, + 604 + ], + "type": "inline_equation", + "content": "d = 256" + }, + { + "bbox": [ + 46, + 388, + 288, + 604 + ], + "type": "text", + "content": " channels via an linear layer. The distillation weight " + }, + { + "bbox": [ + 46, + 388, + 288, + 604 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 46, + 388, + 288, + 604 + ], + "type": "text", + "content": " in Equation (5) is set to 0.9. The number of cross attention layers is " + }, + { + "bbox": [ + 46, + 388, + 288, + 604 + ], + "type": "inline_equation", + "content": "N_{1} = N_{2} = 3" + }, + { + "bbox": [ + 46, + 388, + 288, + 604 + ], + "type": "text", + "content": ". The regression head consists of three " + }, + { + "bbox": [ + 46, + 388, + 288, + 604 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 46, + 388, + 288, + 604 + ], + "type": "text", + "content": " convolutional layers with 128, 64 and 32 feature channels, each followed by a Leaky ReLU, a " + }, + { + "bbox": [ + 46, + 388, + 288, + 604 + ], + "type": "inline_equation", + "content": "2 \\times" + }, + { + "bbox": [ + 46, + 388, + 288, + 604 + ], + "type": "text", + "content": " bilinear upsampling, and a linear layer, consistent with the LOCA configuration. We use AdamW optimizer with the fixed learning rate " + }, + { + "bbox": [ + 46, + 388, + 288, + 604 + ], + "type": "inline_equation", + "content": "10^{-4}" + }, + { + "bbox": [ + 46, + 388, + 288, + 604 + ], + "type": "text", + "content": " and a weight decay of " + }, + { + "bbox": [ + 46, + 388, + 288, + 604 + ], + "type": "inline_equation", + "content": "10^{-2}" + }, + { + "bbox": [ + 46, + 388, + 288, + 604 + ], + "type": "text", + "content": ". Gradient clipping with maximum norm of 0.1 is applied. URM is trained on a single NVIDIA V100 GPU with batch size of 4." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 618, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 618, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 618, + 288, + 714 + ], + "type": "text", + "content": "Datasets. We evaluate our method on two mainstream datasets: FSC-147 [46] and FSCD-LVIS [35]. Specifically, FSC-147 contains 6135 images of 147 object categories, split into training, validation and test sets consisting of 3659, 1286 and 1190 images respectively. Images in FSC-147 are collected from the Internet. The sets of object categories present in each split are disjoint. FSCD-LVIS contains 6195 images spanning 372 classes, which are" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 296, + 545, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 296, + 545, + 332 + ], + "spans": [ + { + "bbox": [ + 304, + 296, + 545, + 332 + ], + "type": "text", + "content": "extracted from the LVIS dataset [15]. Each image annotation consists of three bounding boxes for exemplar objects within each category." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 332, + 545, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 332, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 304, + 332, + 545, + 403 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 304, + 332, + 545, + 403 + ], + "type": "inline_equation", + "content": "S \\to \\mathcal{T}" + }, + { + "bbox": [ + 304, + 332, + 545, + 403 + ], + "type": "text", + "content": " denote the case where " + }, + { + "bbox": [ + 304, + 332, + 545, + 403 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 304, + 332, + 545, + 403 + ], + "type": "text", + "content": " is the source domain and " + }, + { + "bbox": [ + 304, + 332, + 545, + 403 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 304, + 332, + 545, + 403 + ], + "type": "text", + "content": " is the target domain. The datasets are utilized in experiments under two cross-domain settings: the full training dataset of one is regarded as the source domain, while other entire test dataset is regarded as the target domain, i.e., FSC-147 " + }, + { + "bbox": [ + 304, + 332, + 545, + 403 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 304, + 332, + 545, + 403 + ], + "type": "text", + "content": " FSCD-LVIS and FSCD-LVIS " + }, + { + "bbox": [ + 304, + 332, + 545, + 403 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 304, + 332, + 545, + 403 + ], + "type": "text", + "content": " FSC-147." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 420, + 547, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 420, + 547, + 469 + ], + "spans": [ + { + "bbox": [ + 304, + 420, + 547, + 469 + ], + "type": "text", + "content": "Evaluation metrics. We evaluate our method using two commonly employed metrics in object counting: mean absolute error (MAE) and root mean square error (RMSE), between estimated and labeled counts for each image:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 307, + 489, + 545, + 536 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 489, + 545, + 536 + ], + "spans": [ + { + "bbox": [ + 307, + 489, + 545, + 536 + ], + "type": "interline_equation", + "content": "M A E = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\left| c _ {i} - \\hat {c} _ {i} \\right|, R M S E = \\sqrt {\\frac {1}{N} \\sum_ {i = 1} ^ {N} \\left(c _ {i} - \\hat {c} _ {i}\\right) ^ {2}} \\tag {6}", + "image_path": "35e008b1a46bfae39f12586fd17d054ab3ca0d922171686ef0b2149151da669a.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 537, + 545, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 537, + 545, + 574 + ], + "spans": [ + { + "bbox": [ + 304, + 537, + 545, + 574 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 537, + 545, + 574 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 304, + 537, + 545, + 574 + ], + "type": "text", + "content": " is the number of test images, " + }, + { + "bbox": [ + 304, + 537, + 545, + 574 + ], + "type": "inline_equation", + "content": "c_{i}" + }, + { + "bbox": [ + 304, + 537, + 545, + 574 + ], + "type": "text", + "content": " is the ground truth count for the " + }, + { + "bbox": [ + 304, + 537, + 545, + 574 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 304, + 537, + 545, + 574 + ], + "type": "text", + "content": "-th image, and " + }, + { + "bbox": [ + 304, + 537, + 545, + 574 + ], + "type": "inline_equation", + "content": "\\hat{c}_i" + }, + { + "bbox": [ + 304, + 537, + 545, + 574 + ], + "type": "text", + "content": " is the predicted count. Lower values of both metrics indicate better performance." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 581, + 499, + 594 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 581, + 499, + 594 + ], + "spans": [ + { + "bbox": [ + 305, + 581, + 499, + 594 + ], + "type": "text", + "content": "4.2. Comparison with the state of the art" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 600, + 545, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 600, + 545, + 636 + ], + "spans": [ + { + "bbox": [ + 304, + 600, + 545, + 636 + ], + "type": "text", + "content": "In this section, we compare our URM with state-of-the-art methods on different benchmarks. The selected baselines can be grouped into two categories:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 639, + 545, + 687 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 306, + 639, + 545, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 639, + 545, + 662 + ], + "spans": [ + { + "bbox": [ + 306, + 639, + 545, + 662 + ], + "type": "text", + "content": "- Vanilla Methods: FamNet [46], BMNet+ [51], CounTR [32], LOCA [10], and DAVE [40]." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 663, + 545, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 663, + 545, + 687 + ], + "spans": [ + { + "bbox": [ + 306, + 663, + 545, + 687 + ], + "type": "text", + "content": "- Domain Generalization Methods: ISW [9, 41], GS [34], DCCUS [12], and MPCount [41]." + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 306, + 689, + 545, + 714 + ], + "type": "text", + "content": "All these DG methods are originally designed or adapted for crowd counting, we add the prompt encoder after feature" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4644" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 70, + 70, + 265, + 182 + ], + "blocks": [ + { + "bbox": [ + 70, + 70, + 265, + 182 + ], + "lines": [ + { + "bbox": [ + 70, + 70, + 265, + 182 + ], + "spans": [ + { + "bbox": [ + 70, + 70, + 265, + 182 + ], + "type": "table", + "html": "
MethodMAERMSE
Baseline30.1548.04
Language Distillation
Naive Template25.4442.59
Prompt Generator23.8341.03
Vision Distillation
[CLS] Token23.5341.12
Global Pooling22.9439.44
Mask Pooling21.8738.42
", + "image_path": "adccebd21a50c00a5613b4b1b4104d5ec79a8b000eebcf26a3d7b0524e853af7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 225, + 287, + 250 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 225, + 287, + 250 + ], + "spans": [ + { + "bbox": [ + 46, + 225, + 287, + 250 + ], + "type": "text", + "content": "extraction based on their implementation. Besides, GS is a combination of Gradient Surgery [34] and LOCA [10]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 251, + 287, + 406 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 251, + 287, + 406 + ], + "spans": [ + { + "bbox": [ + 46, + 251, + 287, + 406 + ], + "type": "text", + "content": "Table 1 presents the results of domain generalized few-shot counting. Our URM surpasses all the methods significantly, providing compelling evidence for the effectiveness of our designation. In contrast, sub-domain division methods like DCCUS and GS fail to achieve satisfactory results, as they struggle to produce meaningful sub-domains when the source domain distribution is narrow and the domain gap is large. Additionally, the whitening-based method ISW is even harmful because useful information is eliminated through the feature statistics normalization operation. Although MPCount introduced a single memory bank for regression does not require sub-domain partitioning, its generalization ability remains far inferior to our method." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 407, + 287, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 407, + 287, + 515 + ], + "spans": [ + { + "bbox": [ + 46, + 407, + 287, + 515 + ], + "type": "text", + "content": "We further evaluate our method on the zero-shot setting with no user annotation, the results are shown in the right part of Table 1. We follow [10] to take a minor modification of the prompt encoder by using only the learnable embedding to represent the exemplar prompt. The results confirm that our proposed URM effectively enhances the generalization ability of prototypes, enabling accurate count estimation even in the challenging case without manually annotated exemplar." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 525, + 178, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 525, + 178, + 538 + ], + "spans": [ + { + "bbox": [ + 47, + 525, + 178, + 538 + ], + "type": "text", + "content": "4.3. Analysis and Ablations" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 544, + 287, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 544, + 287, + 652 + ], + "spans": [ + { + "bbox": [ + 46, + 544, + 287, + 652 + ], + "type": "text", + "content": "Ablation of the designed elements. We first ablate the important elements of our designation under the FSC-147 (A) " + }, + { + "bbox": [ + 46, + 544, + 287, + 652 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 46, + 544, + 287, + 652 + ], + "type": "text", + "content": " FSCD-LVIS (B) few-shot setting. The baseline model shares the same architecture as URM in the inference phase but without knowledge distillation. The result shown in the first row of Table 2 demonstrates that simply altering architecture, i.e., matching with learned prototypes that lack knowledge transfer and thus have a narrow distribution, introduces no benefit for generalization." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 653, + 287, + 713 + ], + "type": "text", + "content": "In contrast, the method can achieve comparable performance to the SOTA method MPCount by using only language knowledge from CLIP with a naive template. Since the naive template \"A photo of {}\" is too limited to fully represent a category, adding more templates and customized" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 309, + 70, + 543, + 109 + ], + "blocks": [ + { + "bbox": [ + 84, + 191, + 249, + 202 + ], + "lines": [ + { + "bbox": [ + 84, + 191, + 249, + 202 + ], + "spans": [ + { + "bbox": [ + 84, + 191, + 249, + 202 + ], + "type": "text", + "content": "Table 2. Ablation of the designed elements." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 309, + 70, + 543, + 109 + ], + "lines": [ + { + "bbox": [ + 309, + 70, + 543, + 109 + ], + "spans": [ + { + "bbox": [ + 309, + 70, + 543, + 109 + ], + "type": "table", + "html": "
α00.250.50.750.91
MAE26.5423.4423.3022.3721.8723.83
RMSE42.8239.8739.4838.9238.4241.03
", + "image_path": "635da56a77bc969e2e0eebdad5042e157a50af304bcaf75d49fec1ea94f93131.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 325, + 140, + 527, + 193 + ], + "blocks": [ + { + "bbox": [ + 343, + 118, + 508, + 129 + ], + "lines": [ + { + "bbox": [ + 343, + 118, + 508, + 129 + ], + "spans": [ + { + "bbox": [ + 343, + 118, + 508, + 129 + ], + "type": "text", + "content": "Table 3. Ablation of the distillation weight." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 325, + 140, + 527, + 193 + ], + "lines": [ + { + "bbox": [ + 325, + 140, + 527, + 193 + ], + "spans": [ + { + "bbox": [ + 325, + 140, + 527, + 193 + ], + "type": "table", + "html": "
N123334
N232343
MAE22.7722.2321.8723.5823.45
RMSE39.6939.1238.4241.2239.38
", + "image_path": "579c645fa7eb2fd2989426a5c24dff50a4919d77528eaa93df1264eeb67aaae7.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 201, + 545, + 234 + ], + "lines": [ + { + "bbox": [ + 305, + 201, + 545, + 234 + ], + "spans": [ + { + "bbox": [ + 305, + 201, + 545, + 234 + ], + "type": "text", + "content": "Table 4. Ablation of the number of layers, where " + }, + { + "bbox": [ + 305, + 201, + 545, + 234 + ], + "type": "inline_equation", + "content": "N_{1}" + }, + { + "bbox": [ + 305, + 201, + 545, + 234 + ], + "type": "text", + "content": " is the universal prototypes learning layers, and " + }, + { + "bbox": [ + 305, + 201, + 545, + 234 + ], + "type": "inline_equation", + "content": "N_{2}" + }, + { + "bbox": [ + 305, + 201, + 545, + 234 + ], + "type": "text", + "content": " is the universal representation matching layers." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 305, + 256, + 545, + 279 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 256, + 545, + 279 + ], + "spans": [ + { + "bbox": [ + 305, + 256, + 545, + 279 + ], + "type": "text", + "content": "language prompts by the prompt generator further improves the model performance." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 281, + 545, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 281, + 545, + 399 + ], + "spans": [ + { + "bbox": [ + 304, + 281, + 545, + 399 + ], + "type": "text", + "content": "Subsequently, incorporating visual representation via global average pooling from the 2D vision tokens output from the last layer improve performance, demonstrating that the vision representation is complementary to language knowledge. Using the [CLS] token which captures global description results in lower performance. Finally, by introducing more precise locality representation as well as noise removal through mask pooling, our model reaches a MAE of 21.87 as shown in the last row of Table 2, outperforming the baseline method with " + }, + { + "bbox": [ + 304, + 281, + 545, + 399 + ], + "type": "inline_equation", + "content": "27.5\\%" + }, + { + "bbox": [ + 304, + 281, + 545, + 399 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 416, + 545, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 416, + 545, + 548 + ], + "spans": [ + { + "bbox": [ + 304, + 416, + 545, + 548 + ], + "type": "text", + "content": "Ablation of distillation weight. The distillation weight " + }, + { + "bbox": [ + 304, + 416, + 545, + 548 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 304, + 416, + 545, + 548 + ], + "type": "text", + "content": " in Equation (5) controls the balance between visual and language knowledge contributions. In this part, we conduct an ablation study on the weight to fully understand its impact. As shown in Table 3, when " + }, + { + "bbox": [ + 304, + 416, + 545, + 548 + ], + "type": "inline_equation", + "content": "\\alpha = 0" + }, + { + "bbox": [ + 304, + 416, + 545, + 548 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 304, + 416, + 545, + 548 + ], + "type": "inline_equation", + "content": "\\alpha = 1" + }, + { + "bbox": [ + 304, + 416, + 545, + 548 + ], + "type": "text", + "content": ", where only vision or language knowledge is introduced respectively, the performance is inferior to other settings, demonstrating that V-L representations complement each other. Setting " + }, + { + "bbox": [ + 304, + 416, + 545, + 548 + ], + "type": "inline_equation", + "content": "\\alpha = 0.9" + }, + { + "bbox": [ + 304, + 416, + 545, + 548 + ], + "type": "text", + "content": " yields the best performance, indicating that language representation from CLIP, which has been aligned to visual space, plays an influential role." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 564, + 545, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 564, + 545, + 673 + ], + "spans": [ + { + "bbox": [ + 304, + 564, + 545, + 673 + ], + "type": "text", + "content": "Ablation of the number layers. We then ablate the hyper-parameters related to the number of layers. The universal prototypes learning layer number " + }, + { + "bbox": [ + 304, + 564, + 545, + 673 + ], + "type": "inline_equation", + "content": "N_{1}" + }, + { + "bbox": [ + 304, + 564, + 545, + 673 + ], + "type": "text", + "content": " determines how many layers are required to absorb the universal knowledge transferred from CLIP. And the universal representation matching layer number " + }, + { + "bbox": [ + 304, + 564, + 545, + 673 + ], + "type": "inline_equation", + "content": "N_{2}" + }, + { + "bbox": [ + 304, + 564, + 545, + 673 + ], + "type": "text", + "content": " defines how many layers are needed to learn an effective correlation map during matching in the hidden space. The results in Table 4 show that " + }, + { + "bbox": [ + 304, + 564, + 545, + 673 + ], + "type": "inline_equation", + "content": "N_{1} = N_{2} = 3" + }, + { + "bbox": [ + 304, + 564, + 545, + 673 + ], + "type": "text", + "content": " achieves the best performance." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": "Ablation of different representation sources. In this part, we investigate the effect of different representation" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4645" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 70, + 288, + 157 + ], + "blocks": [ + { + "bbox": [ + 48, + 70, + 288, + 157 + ], + "lines": [ + { + "bbox": [ + 48, + 70, + 288, + 157 + ], + "spans": [ + { + "bbox": [ + 48, + 70, + 288, + 157 + ], + "type": "table", + "html": "
ModelArchitectureACCMAERMSE
OpenAI CLIP [43]ViT-B/1668.321.8738.42
OpenAI CLIP [43]ViT-L/1475.522.6338.95
OpenAI CLIP [43]RN50×1670.523.1539.63
OpenCLIP [8]ConvNeXt-B71.522.7839.10
SigLIP [68]ViT-B/1673.421.3337.85
EVA-02-CLIP [52]ViT-B/1674.721.8638.12
", + "image_path": "4fd50233c42c4e6e289ffdc726568345c856eb72c371b1f9cdb195be165cded2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 220, + 287, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 220, + 287, + 387 + ], + "spans": [ + { + "bbox": [ + 46, + 220, + 287, + 387 + ], + "type": "text", + "content": "learning sources in Table 5. The first row presents the results of our default setting which uses the OpenAI ViT-B/16 CLIP model. Increasing the model size leads to a performance drop due to the large dimension gap between the CLIP model and URM as seen in the second row. Besides, models with convolutional backbones, such as ResNet50 and ConvNeXt, perform poorly in comparison despite achieving higher ImageNet accuracy, suggesting that ViT architecture is better suited for representation learning in our task. Finally, the results in the last two lines indicate that using a stronger CLIP teacher model can further enhance our model's performance. However, this improvement also does not directly correlate with ImageNet accuracy, aligning with the observations in [26]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 394, + 251, + 406 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 394, + 251, + 406 + ], + "spans": [ + { + "bbox": [ + 47, + 394, + 251, + 406 + ], + "type": "text", + "content": "4.4. Performance on in domain benchmark" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 412, + 287, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 412, + 287, + 616 + ], + "spans": [ + { + "bbox": [ + 46, + 412, + 287, + 616 + ], + "type": "text", + "content": "In this section, URM is compared with the state-of-the-art methods on in domain benchmark. The three-shot evaluation results on FSC147 dataset are presented in Table 6. While demonstrating strong generalization, URM outperforms state-of-the-art few-shot counting models in terms of in-domain performance. Additionally, DAVE [40] proposed a detect-and-verify paradigm for the existing regression-based methods like LOCA to mitigate false activations on objects of other categories. Since DAVE is a general method for merging the benefits of both density and detection-based approaches, we also extend our method by incorporating this paradigm. Although inductive biases introduced by DAVE are insufficient to handle complex distribution gap in domain generalization setting (as show in Table 1), the resulting method still surpasses other methods on the traditional in domain benchmark, as demonstrated in the last row of Table 6." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 623, + 272, + 636 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 623, + 272, + 636 + ], + "spans": [ + { + "bbox": [ + 47, + 623, + 272, + 636 + ], + "type": "text", + "content": "4.5. Generating language prompts with LVLMs" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 641, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 641, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 641, + 287, + 715 + ], + "type": "text", + "content": "In this section, we explore the use of large vision-language models to generate language prompts, enabling our method to train without predefined category names. However, existing proprietary and public LVLMs struggle with image classification especially in open-world setting where the class list is unknown [70], making them difficult to" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 307, + 70, + 553, + 218 + ], + "blocks": [ + { + "bbox": [ + 46, + 165, + 288, + 199 + ], + "lines": [ + { + "bbox": [ + 46, + 165, + 288, + 199 + ], + "spans": [ + { + "bbox": [ + 46, + 165, + 288, + 199 + ], + "type": "text", + "content": "Table 5. Ablation of different universal V-L representation sources, where ACC represents the ImageNet zero-shot classification performance of the CLIP model." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 307, + 70, + 553, + 218 + ], + "lines": [ + { + "bbox": [ + 307, + 70, + 553, + 218 + ], + "spans": [ + { + "bbox": [ + 307, + 70, + 553, + 218 + ], + "type": "table", + "html": "
SplitValidation setTest set
MethodMAERMSEMAERMSE
FamNet (CVPR21) [46]23.7569.0722.0899.54
BMNet+ (CVPR22) [51]15.7458.5314.6291.83
VCN (CVPR22) [45]19.3860.1518.1795.60
CountTR (BMVC22) [32]13.1349.8311.9591.23
SAFECount (WACV23) [65]15.2847.2014.3285.54
PseCo (CVPR24) [19]15.3168.3413.05112.86
LOCA (ICCV23) [10]10.2432.5610.7956.97
DAVE (CVPR24) [40]8.9128.088.6632.36
URM (Ours)10.6332.0110.1554.83
URM+DAVE [40]8.4726.518.3232.05
", + "image_path": "c7f391d148839015a26215e9486f1468afbacaa4b87c7cb61d994a8c4063a757.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 358, + 248, + 493, + 288 + ], + "blocks": [ + { + "bbox": [ + 309, + 226, + 541, + 237 + ], + "lines": [ + { + "bbox": [ + 309, + 226, + 541, + 237 + ], + "spans": [ + { + "bbox": [ + 309, + 226, + 541, + 237 + ], + "type": "text", + "content": "Table 6. Performance comparison on in domain benchmark." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 358, + 248, + 493, + 288 + ], + "lines": [ + { + "bbox": [ + 358, + 248, + 493, + 288 + ], + "spans": [ + { + "bbox": [ + 358, + 248, + 493, + 288 + ], + "type": "table", + "html": "
LVLMMAERMSE
Claude3 [3]22.5838.96
GPT4 [1]22.4538.75
", + "image_path": "1fe44c5a5da46a1f1bb2c5037a3a42e58349df751877a1c32dd8050f6be30a99.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 296, + 545, + 319 + ], + "lines": [ + { + "bbox": [ + 306, + 296, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 306, + 296, + 545, + 319 + ], + "type": "text", + "content": "Table 7. Results of generating language prompts using LVLMs when category names are not given." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 340, + 547, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 340, + 547, + 508 + ], + "spans": [ + { + "bbox": [ + 304, + 340, + 547, + 508 + ], + "type": "text", + "content": "accurately fill templates in the prompt generator. Fortunately, customized descriptive prompts with details are effective [42] as shown in Section 4.3 and Appendix. As a result, we choose Claude3-Opus [3] and GPT4 [1] to generate language prompts which have demonstrated the best open-world classification performance among existing VLMs [70]. We replace “{}” in our template prompts with “The category of the most numerous objects in the image” to prompt the VLMs. For example, “Describe what " + }, + { + "bbox": [ + 304, + 340, + 547, + 508 + ], + "type": "inline_equation", + "content": "a(n)" + }, + { + "bbox": [ + 304, + 340, + 547, + 508 + ], + "type": "text", + "content": " {} looks like?” becomes “Describe what the category of the most numerous objects in the image looks like?”. The results in Table 7 show that our method can successfully generalize to scenarios where category names are not provided during training." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 518, + 379, + 530 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 518, + 379, + 530 + ], + "spans": [ + { + "bbox": [ + 306, + 518, + 379, + 530 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 538, + 547, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 538, + 547, + 693 + ], + "spans": [ + { + "bbox": [ + 304, + 538, + 547, + 693 + ], + "type": "text", + "content": "This paper proposes URM to address single domain generalization for few-shot counting with narrow source distribution. URM is motivated by the unique challenge that existing methods struggle to learn high generalization prototypes. To address this issue, we propose a universal representation matching paradigm that incorporating universal vision-language prototypes distilled from the universal representations from CLIP model, which has seen vast diversity of distributions during training. Matching with these universal prototypes leads to high generalization and superior performance. Extensive experiments on both cross and in domain settings show that URM significantly outperform the state-of-the-art." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4646" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 53, + 91, + 286, + 113 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 286, + 113 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 286, + 113 + ], + "type": "text", + "content": "[1] OpenAI Josh Achiam, Steven Adler, Sandhini Agarwal, and et al. Gpt-4 technical report. 2023. 2, 5, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 114, + 287, + 147 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 114, + 287, + 147 + ], + "spans": [ + { + "bbox": [ + 53, + 114, + 287, + 147 + ], + "type": "text", + "content": "[2] Niki Amini-Naeni, Kiana Amini-Naeni, Tengda Han, and Andrew Zisserman. Open-world text-specified object counting. arXiv preprint arXiv:2306.01851, 2023. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 148, + 286, + 168 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 148, + 286, + 168 + ], + "spans": [ + { + "bbox": [ + 53, + 148, + 286, + 168 + ], + "type": "text", + "content": "[3] anthropic. Introducing the next generation of claude. 2024. 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 170, + 287, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 170, + 287, + 213 + ], + "spans": [ + { + "bbox": [ + 53, + 170, + 287, + 213 + ], + "type": "text", + "content": "[4] Yiqing Cai, Lianggangxu Chen, Haoyue Guan, Shaohui Lin, Changhong Lu, Changbo Wang, and Gaoqi He. Explicit invariant feature induced cross-domain crowd counting. In AAAI Conference on Artificial Intelligence, 2023. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 214, + 287, + 269 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 214, + 287, + 269 + ], + "spans": [ + { + "bbox": [ + 53, + 214, + 287, + 269 + ], + "type": "text", + "content": "[5] Xianing Chen, Qiong Cao, Yujie Zhong, Jing Zhang, Shenghua Gao, and Dacheng Tao. Dearkd: Data-efficient early knowledge distillation for vision transformers. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12042-12052, 2022. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 270, + 287, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 270, + 287, + 323 + ], + "spans": [ + { + "bbox": [ + 53, + 270, + 287, + 323 + ], + "type": "text", + "content": "[6] Xi Chen, Shuang Li, Ser Nam Lim, Antonio Torralba, and Hengshuang Zhao. Open-vocabulary panoptic segmentation with embedding modulation. 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 1141-1150, 2023. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 325, + 287, + 356 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 325, + 287, + 356 + ], + "spans": [ + { + "bbox": [ + 53, + 325, + 287, + 356 + ], + "type": "text", + "content": "[7] Junting Cheng, Dong Liang, and Shan Tan. Transfer clip for generalizable image denoising. ArXiv, abs/2403.15132, 2024. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 358, + 287, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 358, + 287, + 424 + ], + "spans": [ + { + "bbox": [ + 53, + 358, + 287, + 424 + ], + "type": "text", + "content": "[8] Mehdi Cherti, Romain Beaumont, Ross Wightman, Mitchell Wortsman, Gabriel Ilharco, Cade Gordon, Christoph Schuhmann, Ludwig Schmidt, and Jenia Jitsev. Reproducible scaling laws for contrastive language-image learning. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2818-2829, 2022. 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 426, + 287, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 426, + 287, + 491 + ], + "spans": [ + { + "bbox": [ + 53, + 426, + 287, + 491 + ], + "type": "text", + "content": "[9] Sungha Choi, Sanghun Jung, Huiwon Yun, Joanne Taery Kim, Seungryong Kim, and Jaegul Choo. Robustnet: Improving domain generalization in urban-scene segmentation via instance selective whitening. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11575-11585, 2021. 3, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 492, + 287, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 492, + 287, + 535 + ], + "spans": [ + { + "bbox": [ + 48, + 492, + 287, + 535 + ], + "type": "text", + "content": "[10] Nikola Djukic, Alan Luke, Vitjan Zavrtanik, and Matej Kristan. A low-shot object counting network with iterative prototype adaptation. 2023 IEEE/CVF International Conference on Computer Vision (ICCV), 2022. 2, 3, 4, 6, 7, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 536, + 287, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 536, + 287, + 601 + ], + "spans": [ + { + "bbox": [ + 48, + 536, + 287, + 601 + ], + "type": "text", + "content": "[11] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. ArXiv, abs/2010.11929, 2020. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 602, + 287, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 602, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 602, + 287, + 635 + ], + "type": "text", + "content": "[12] Zhipeng Du, Jiankang Deng, and Miaojing Shi. Domain-general crowd counting in unseen scenarios. ArXiv, abs/2212.02573, 2022. 3, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 636, + 287, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 636, + 287, + 679 + ], + "spans": [ + { + "bbox": [ + 48, + 636, + 287, + 679 + ], + "type": "text", + "content": "[13] Junyuan Gao, Tao Han, Yuan Yuan, and Qi Wang. Domain-adaptive crowd counting via high-quality image translation and density reconstruction. IEEE Transactions on Neural Networks and Learning Systems, 34:4803-4815, 2019. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 681, + 287, + 713 + ], + "type": "text", + "content": "[14] Jianping Gou, B. Yu, Stephen J. Maybank, and Dacheng Tao. Knowledge distillation: A survey. International Journal of Computer Vision, 129:1789 - 1819, 2020. 3" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 117 + ], + "type": "text", + "content": "[15] Agrim Gupta, Piotr Dólár, and Ross B. Girshick. Lvis: A dataset for large vocabulary instance segmentation. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5351-5359, 2019. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 118, + 545, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 150 + ], + "type": "text", + "content": "[16] Geoffrey E. Hinton, Oriol Vinyals, and Jeffrey Dean. Distilling the knowledge in a neural network. ArXiv, abs/1503.02531, 2015. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 152, + 545, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 545, + 185 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 545, + 185 + ], + "type": "text", + "content": "[17] Michael A. Hobley and Victor Adrian Prisacariu. Learning to count anything: Reference-less class-agnostic counting with weak supervision. ArXiv, abs/2205.10203, 2022. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 186, + 545, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 186, + 545, + 219 + ], + "spans": [ + { + "bbox": [ + 307, + 186, + 545, + 219 + ], + "type": "text", + "content": "[18] Xiaohui Huang, Hao Zhou, Kun Yao, and Kai Han. Froster: Frozen clip is a strong teacher for open-vocabulary action recognition. ArXiv, abs/2402.03241, 2024. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 220, + 545, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 220, + 545, + 274 + ], + "spans": [ + { + "bbox": [ + 308, + 220, + 545, + 274 + ], + "type": "text", + "content": "[19] Zhizhong Huang, Mingliang Dai, Yi Zhang, Junping Zhang, and Hongming Shan. Point, segment and count: A generalized framework for object counting. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17067-17076, 2023. 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 276, + 545, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 276, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 308, + 276, + 545, + 319 + ], + "type": "text", + "content": "[20] Qing Jiang, Feng Li, Zhaoyang Zeng, Tianhe Ren, Shilong Liu, and Lei Zhang. T-rex2: Towards generic object detection via text-visual prompt synergy. In European Conference on Computer Vision, pages 38-57. Springer, 2024. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 321, + 545, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 321, + 545, + 365 + ], + "spans": [ + { + "bbox": [ + 308, + 321, + 545, + 365 + ], + "type": "text", + "content": "[21] Ruixiang Jiang, Lingbo Liu, and Changwen Chen. Clipcount: Towards text-guided zero-shot object counting. In Proceedings of the 31st ACM International Conference on Multimedia, pages 4535-4545, 2023. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 365, + 545, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 365, + 545, + 409 + ], + "spans": [ + { + "bbox": [ + 308, + 365, + 545, + 409 + ], + "type": "text", + "content": "[22] Seunggu Kang, WonJun Moon, Euiyeon Kim, and Jae-Pil Heo. Vlcounter: Text-aware visual representation for zero-shot object counting. In AAAI Conference on Artificial Intelligence, 2023. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 411, + 545, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 411, + 545, + 475 + ], + "spans": [ + { + "bbox": [ + 308, + 411, + 545, + 475 + ], + "type": "text", + "content": "[23] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dólár, and Ross B. Girshick. Segment anything. 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 3992-4003, 2023. 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 477, + 545, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 477, + 545, + 499 + ], + "spans": [ + { + "bbox": [ + 308, + 477, + 545, + 499 + ], + "type": "text", + "content": "[24] Wouter M. Kouw. An introduction to domain adaptation and transfer learning. ArXiv, abs/1812.11806, 2018. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 501, + 545, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 501, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 308, + 501, + 545, + 544 + ], + "type": "text", + "content": "[25] Mengcheng Lan, Chaofeng Chen, Yiping Ke, Xinjiang Wang, Litong Feng, and Wayne Zhang. Clearclip: Decomposing clip representations for dense vision-language inference. ArXiv, abs/2407.12442, 2024. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 545, + 545, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 545, + 545, + 578 + ], + "spans": [ + { + "bbox": [ + 308, + 545, + 545, + 578 + ], + "type": "text", + "content": "[26] Hugo Laurençon, Léo Tronchon, Matthieu Cord, and Victor Sanh. What matters when building vision-language models? ArXiv, abs/2405.02246, 2024. 8" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 579, + 545, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 579, + 545, + 623 + ], + "spans": [ + { + "bbox": [ + 308, + 579, + 545, + 623 + ], + "type": "text", + "content": "[27] Da Li, Yongxin Yang, Yi-Zhe Song, and Timothy M. Hospedales. Learning to generalize: Meta-learning for domain generalization. In AAAI Conference on Artificial Intelligence, 2017. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 624, + 545, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 624, + 545, + 667 + ], + "spans": [ + { + "bbox": [ + 308, + 624, + 545, + 667 + ], + "type": "text", + "content": "[28] Zheng Li, Xiang Li, Xinyi Fu, Xing Zhang, Weiqiang Wang, Shuo Chen, and Jian Yang. Promptkd: Unsupervised prompt distillation for vision-language models. ArXiv, abs/2403.02781, 2024. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 669, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 669, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 669, + 545, + 713 + ], + "type": "text", + "content": "[29] Dongze Lian, Xianing Chen, Jing Li, Weixin Luo, and Shenghua Gao. Locating and counting heads in crowds with a depth prior. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44:9056-9072, 2021. 2" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4647" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 127 + ], + "type": "text", + "content": "[30] Dingkang Liang, Jiahao Xie, Zhikang Zou, Xiaoqing Ye, Wei Xu, and Xiang Bai. Crowdclip: Unsupervised crowd counting via vision-language model. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2893-2903, 2023. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 129, + 287, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 129, + 287, + 160 + ], + "spans": [ + { + "bbox": [ + 48, + 129, + 287, + 160 + ], + "type": "text", + "content": "[31] Hui Lin, Xiaopeng Hong, and Yabin Wang. Object counting: You only need to look at one. ArXiv, abs/2112.05993, 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 163, + 287, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 163, + 287, + 196 + ], + "spans": [ + { + "bbox": [ + 48, + 163, + 287, + 196 + ], + "type": "text", + "content": "[32] Chang Liu, Yujie Zhong, Andrew Zisserman, and Weidi Xie. Countr: Transformer-based generalised visual counting. ArXiv, abs/2208.13721, 2022. 2, 3, 6, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 198, + 287, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 198, + 287, + 251 + ], + "spans": [ + { + "bbox": [ + 48, + 198, + 287, + 251 + ], + "type": "text", + "content": "[33] Yuting Liu, Zheng Wang, Miaojing Shi, Shin'ichi Satoh, Qijun Zhao, and Hongyu Yang. Towards unsupervised crowd counting via regression-detection bi-knowledge transfer. Proceedings of the 28th ACM International Conference on Multimedia, 2020. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 253, + 287, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 253, + 287, + 297 + ], + "spans": [ + { + "bbox": [ + 48, + 253, + 287, + 297 + ], + "type": "text", + "content": "[34] Lucas Mansilla, Rodrigo Echeveste, Diego H. Milone, and Enzo Ferrante. Domain generalization via gradient surgery. 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 6610-6618, 2021. 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 298, + 287, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 298, + 287, + 331 + ], + "spans": [ + { + "bbox": [ + 48, + 298, + 287, + 331 + ], + "type": "text", + "content": "[35] Thanh Thoi Nguyen, Chau Khoa Pham, Khoi Duc Minh Nguyen, and Minh Hoai. Few-shot object counting and detection. ArXiv, abs/2207.10988, 2022. 1, 2, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 332, + 287, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 332, + 287, + 385 + ], + "spans": [ + { + "bbox": [ + 48, + 332, + 287, + 385 + ], + "type": "text", + "content": "[36] Utkarsh Ojha, Yuheng Li, and Yong Jae Lee. Towards universal fake image detectors that generalize across generative models. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 24480-24489, 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 388, + 287, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 388, + 287, + 432 + ], + "spans": [ + { + "bbox": [ + 48, + 388, + 287, + 432 + ], + "type": "text", + "content": "[37] Roni Paiss, Ariel Ephrat, Omer Tov, Shiran Zada, Inbar Mosseri, Michal Irani, and Tali Dekel. Teaching clip to count to ten. 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 3147-3157, 2023. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 434, + 287, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 434, + 287, + 475 + ], + "spans": [ + { + "bbox": [ + 48, + 434, + 287, + 475 + ], + "type": "text", + "content": "[38] Xingang Pan, Ping Luo, Jianping Shi, and Xiaou Tang. Two at once: Enhancing learning and generalization capacities via ibn-net. In European Conference on Computer Vision, 2018. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 478, + 287, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 478, + 287, + 521 + ], + "spans": [ + { + "bbox": [ + 48, + 478, + 287, + 521 + ], + "type": "text", + "content": "[39] Xingang Pan, Xiaohang Zhan, Jianping Shi, Xiaou Tang, and Ping Luo. Switchable whitening for deep representation learning. 2019 IEEE/CVF International Conference on Computer Vision (ICCV), pages 1863-1871, 2019. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 523, + 287, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 523, + 287, + 555 + ], + "spans": [ + { + "bbox": [ + 48, + 523, + 287, + 555 + ], + "type": "text", + "content": "[40] Jer Pelhan, Alan Luke, Vitjan Zavrtanik, and Matej Kristan. Dave - a detect-and-verify paradigm for low-shot counting. ArXiv, abs/2404.16622, 2024. 1, 3, 6, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 557, + 287, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 557, + 287, + 590 + ], + "spans": [ + { + "bbox": [ + 48, + 557, + 287, + 590 + ], + "type": "text", + "content": "[41] Zhuoxuan Peng and S.-H. Gary Chan. Single domain generalization for crowd counting. ArXiv, abs/2403.09124, 2024. 3, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 591, + 287, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 591, + 287, + 645 + ], + "spans": [ + { + "bbox": [ + 48, + 591, + 287, + 645 + ], + "type": "text", + "content": "[42] Sarah Pratt, Rosanne Liu, and Ali Farhadi. What does a platypus look like? generating customized prompts for zero-shot image classification. 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 15645-15655, 2022. 2, 5, 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 647, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 647, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 647, + 287, + 712 + ], + "type": "text", + "content": "[43] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, 2021. 2, 3, 5, 8" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 95 + ], + "type": "text", + "content": "[44] Viresh Ranjan and Minh Hoai. Exemplar free class agnostic counting. In Asian Conference on Computer Vision, 2022. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 95, + 545, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 95, + 545, + 138 + ], + "spans": [ + { + "bbox": [ + 307, + 95, + 545, + 138 + ], + "type": "text", + "content": "[45] Viresh Ranjan and Minh Hoai. Vicinal counting networks. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 4220-4229, 2022. 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 140, + 545, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 140, + 545, + 183 + ], + "spans": [ + { + "bbox": [ + 308, + 140, + 545, + 183 + ], + "type": "text", + "content": "[46] Viresh Ranjan, Udbhav Sharma, Thua Nguyen, and Minh Hoai. Learning to count everything. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3393-3402, 2021. 1, 2, 3, 6, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 184, + 545, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 184, + 545, + 237 + ], + "spans": [ + { + "bbox": [ + 308, + 184, + 545, + 237 + ], + "type": "text", + "content": "[47] Hanoona Abdul Rasheed, Muhammad Maaz, Muhammad Uzair Khattak, Salman H. Khan, and Fahad Shahbaz Khan. Bridging the gap between object and image-level representations for open-vocabulary detection. ArXiv, abs/2207.03482, 2022. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 239, + 545, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 239, + 545, + 282 + ], + "spans": [ + { + "bbox": [ + 308, + 239, + 545, + 282 + ], + "type": "text", + "content": "[48] Shaoqing Ren, Kaiming He, Ross B. Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. IEEE Transactions on Pattern Analysis and Machine Intelligence, 39:1137-1149, 2015. 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 283, + 545, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 283, + 545, + 337 + ], + "spans": [ + { + "bbox": [ + 308, + 283, + 545, + 337 + ], + "type": "text", + "content": "[49] Tianhe Ren, Qing Jiang, Shilong Liu, Zhaoyang Zeng, Wenlong Liu, Han Gao, Hongjie Huang, Zhengyu Ma, Xiaoke Jiang, Yihao Chen, et al. Grounding dino 1.5: Advance the\" edge\" of open-set object detection. arXiv preprint arXiv:2405.10300, 2024. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 338, + 545, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 338, + 545, + 381 + ], + "spans": [ + { + "bbox": [ + 308, + 338, + 545, + 381 + ], + "type": "text", + "content": "[50] Cheng Shi and Sibei Yang. Edadet: Open-vocabulary object detection using early dense alignment. 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 15678-15688, 2023. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 383, + 545, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 383, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 308, + 383, + 545, + 437 + ], + "type": "text", + "content": "[51] Min Shi, Hao Lu, Chen Feng, Chengxin Liu, and ZHIGUO CAO. Represent, compare, and learn: A similarity-aware framework for class-agnostic counting. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9519–9528, 2022. 2, 3, 6, 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 438, + 545, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 438, + 545, + 470 + ], + "spans": [ + { + "bbox": [ + 308, + 438, + 545, + 470 + ], + "type": "text", + "content": "[52] Quan Sun, Yuxin Fang, Ledell Yu Wu, Xinlong Wang, and Yue Cao. Eva-clip: Improved training techniques for clip at scale. ArXiv, abs/2303.15389, 2023. 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 471, + 545, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 471, + 545, + 524 + ], + "spans": [ + { + "bbox": [ + 308, + 471, + 545, + 524 + ], + "type": "text", + "content": "[53] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Herv'e J'egou. Training data-efficient image transformers & distillation through attention. In International Conference on Machine Learning, 2020. 2, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 525, + 545, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 525, + 545, + 591 + ], + "spans": [ + { + "bbox": [ + 308, + 525, + 545, + 591 + ], + "type": "text", + "content": "[54] Aayushi Tyagi, Chirag Mohapatra, Prasenjit Das, Govind Kumar Makharia, Lalita Mehra, A. P. Prathosh, Musam, and Iit Delhi. Degpr: Deep guided posterior regularization for multi-class cell detection and counting. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 23913-23923, 2023. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 592, + 545, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 592, + 545, + 635 + ], + "spans": [ + { + "bbox": [ + 308, + 592, + 545, + 635 + ], + "type": "text", + "content": "[55] Eric Tzeng, Judy Hoffman, Kate Saenko, and Trevor Darrell. Adversarial discriminative domain adaptation. 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2962-2971, 2017. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 636, + 545, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 636, + 545, + 668 + ], + "spans": [ + { + "bbox": [ + 308, + 636, + 545, + 668 + ], + "type": "text", + "content": "[56] Laurens van der Maaten and Geoffrey E. Hinton. Visualizing data using t-sne. Journal of Machine Learning Research, 9: 2579-2605, 2008. 4" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 669, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 669, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 308, + 669, + 545, + 712 + ], + "type": "text", + "content": "[57] Ashish Vaswani, Noam M. Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Neural Information Processing Systems, 2017. 4" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "4648" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 126 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 126 + ], + "type": "text", + "content": "[58] Mingjie Wang, Song Yuan, Zhuohang Li, Longlong Zhu, Eric Buys, and Minglun Gong. Language-guided zero-shot object counting. 2024 IEEE International Conference on Multimedia and Expo Workshops (ICMEW), pages 1-6, 2024. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 127, + 287, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 127, + 287, + 172 + ], + "spans": [ + { + "bbox": [ + 48, + 127, + 287, + 172 + ], + "type": "text", + "content": "[59] Qi Wang, Junyu Gao, Wei Lin, and Yuan Yuan. Learning from synthetic data for crowd counting in the wild. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8190-8199, 2019. 1, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 172, + 287, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 172, + 287, + 227 + ], + "spans": [ + { + "bbox": [ + 48, + 172, + 287, + 227 + ], + "type": "text", + "content": "[60] Longyin Wen, Dawei Du, Pengfei Zhu, Qinghua Hu, Qilong Wang, Liefeng Bo, and Siwei Lyu. Detection, tracking, and counting meets drones in crowds: A benchmark. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7808-7817, 2021. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 228, + 287, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 228, + 287, + 270 + ], + "spans": [ + { + "bbox": [ + 48, + 228, + 287, + 270 + ], + "type": "text", + "content": "[61] Qiangqiang Wu, Jia Wan, and Antoni B. Chan. Dynamic momentum adaptation for zero-shot cross-domain crowd counting. Proceedings of the 29th ACM International Conference on Multimedia, 2021. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 271, + 287, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 271, + 287, + 316 + ], + "spans": [ + { + "bbox": [ + 48, + 271, + 287, + 316 + ], + "type": "text", + "content": "[62] Jingyi Xu, Hieu M. Le, Vu Nguyen, Viresh Ranjan, and Dimitris Samaras. Zero-shot object counting. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15548-15557, 2023. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 316, + 287, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 316, + 287, + 371 + ], + "spans": [ + { + "bbox": [ + 48, + 316, + 287, + 371 + ], + "type": "text", + "content": "[63] Chuanguang Yang, Zhulin An, Libo Huang, Junyu Bi, Xinqiang Yu, Hansheng Yang, Boyu Diao, and Yongjun Xu. Clip-kd: An empirical study of clip model distillation. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15952-15962, 2023. 3, 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 371, + 287, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 371, + 287, + 435 + ], + "spans": [ + { + "bbox": [ + 48, + 371, + 287, + 435 + ], + "type": "text", + "content": "[64] Lewei Yao, Renjie Pi, Jianhua Han, Xiaodan Liang, Hang Xu, Wei Zhang, Zhenguo Li, and Dan Xu. Detclipv3: Towards versatile generative open-vocabulary object detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 27391-27401, 2024. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 437, + 287, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 437, + 287, + 491 + ], + "spans": [ + { + "bbox": [ + 48, + 437, + 287, + 491 + ], + "type": "text", + "content": "[65] Zhiyuan You, Yujun Shen, Kai Yang, Wenhan Luo, X. Lu, Lei Cui, and Xinyi Le. Few-shot object counting with similarity-aware feature enhancement. 2023 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 6304-6313, 2022. 1, 2, 3, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 492, + 287, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 492, + 287, + 536 + ], + "spans": [ + { + "bbox": [ + 48, + 492, + 287, + 536 + ], + "type": "text", + "content": "[66] Wenwen Yu, Yuliang Liu, Wei Hua, Deqiang Jiang, Bo Ren, and Xiang Bai. Turning a clip model into a scene text detector. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6978-6988, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 536, + 287, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 536, + 287, + 568 + ], + "spans": [ + { + "bbox": [ + 48, + 536, + 287, + 568 + ], + "type": "text", + "content": "[67] Vitjan Zavrtanik, Martin Vodopivec, and Matej Kristan. A segmentation-based approach for polyp counting in the wild. Eng. Appl. Artif. Intell., 88, 2020. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 570, + 287, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 570, + 287, + 613 + ], + "spans": [ + { + "bbox": [ + 48, + 570, + 287, + 613 + ], + "type": "text", + "content": "[68] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 11941-11952, 2023. 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 614, + 287, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 614, + 287, + 668 + ], + "spans": [ + { + "bbox": [ + 48, + 614, + 287, + 668 + ], + "type": "text", + "content": "[69] Yingying Zhang, Desen Zhou, Siqin Chen, Shenghua Gao, and Yi Ma. Single-image crowd counting via multi-column convolutional neural network. 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 589-597, 2016. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 669, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 669, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 669, + 287, + 712 + ], + "type": "text", + "content": "[70] Yuhui Zhang, Alyssa Unell, Xiaohan Wang, Dhruba Ghosh, Yuchang Su, Ludwig Schmidt, and Serena Yeung-Levy. Why are visually-grounded language models bad at image classification? ArXiv, abs/2405.18415, 2024. 8" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 72, + 546, + 340 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 307, + 72, + 546, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 72, + 546, + 117 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 546, + 117 + ], + "type": "text", + "content": "[71] Borui Zhao, Quan Cui, Renjie Song, Yiyu Qiu, and Jiajun Liang. Decoupled knowledge distillation. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11943-11952, 2022. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 118, + 546, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 546, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 546, + 150 + ], + "type": "text", + "content": "[72] Chong Zhou, Chen Change Loy, and Bo Dai. Extract free dense labels from clip. In European Conference on Computer Vision, 2021. 2, 3, 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 151, + 546, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 151, + 546, + 195 + ], + "spans": [ + { + "bbox": [ + 307, + 151, + 546, + 195 + ], + "type": "text", + "content": "[73] Kaiyang Zhou, Ziwei Liu, Yu Qiao, Tao Xiang, and Chen Change Loy. Domain generalization: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45:4396-4415, 2021. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 196, + 546, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 196, + 546, + 239 + ], + "spans": [ + { + "bbox": [ + 307, + 196, + 546, + 239 + ], + "type": "text", + "content": "[74] Huilin Zhu, Jingling Yuan, Zhengwei Yang, Xian Zhong, and Zheng Wang. Fine-grained fragment diffusion for cross-domain crowd counting. Proceedings of the 30th ACM International Conference on Multimedia, 2022. 1, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 240, + 546, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 240, + 546, + 296 + ], + "spans": [ + { + "bbox": [ + 307, + 240, + 546, + 296 + ], + "type": "text", + "content": "[75] Huilin Zhu, Jingling Yuan, Xian Zhong, Zhengwei Yang, Zheng Wang, and Shengfeng He. Daot: Domain-agnostically aligned optimal transport for domain-adaptive crowd counting. Proceedings of the 31st ACM International Conference on Multimedia, 2023. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 297, + 546, + 340 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 297, + 546, + 340 + ], + "spans": [ + { + "bbox": [ + 307, + 297, + 546, + 340 + ], + "type": "text", + "content": "[76] Huilin Zhu, Jingling Yuan, Zhengwei Yang, Yu Guo, Zheng Wang, Xian Zhong, and Shengfeng He. Zero-shot object counting with good exemplars. ArXiv, abs/2407.04948, 2024.3" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "4649" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Six-CD_ Benchmarking Concept Removals for Text-to-image Diffusion Models/17b4b8e3-2cbd-49ce-875b-c3c7f736c51d_content_list.json b/2025/Six-CD_ Benchmarking Concept Removals for Text-to-image Diffusion Models/17b4b8e3-2cbd-49ce-875b-c3c7f736c51d_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..1e06afb37b04819f63a3c33c24f415b579b9cb97 --- /dev/null +++ b/2025/Six-CD_ Benchmarking Concept Removals for Text-to-image Diffusion Models/17b4b8e3-2cbd-49ce-875b-c3c7f736c51d_content_list.json @@ -0,0 +1,1569 @@ +[ + { + "type": "text", + "text": "Six-CD: Benchmarking Concept Removals for Text-to-image Diffusion Models", + "text_level": 1, + "bbox": [ + 102, + 130, + 893, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jie Ren $^{1}$ , Kangrui Chen $^{1}$ , Yingqian Cui $^{1}$ , Shenglai Zeng $^{1}$ , Hui Liu $^{1}$ , Yue Xing $^{1}$ , Jiliang Tang $^{1}$ , Lingjuan Lyu $^{2}$ $^{1}$ Michigan State University $^{2}$ Sony AI", + "bbox": [ + 184, + 179, + 812, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{renjie3,chenkan4,cuiyingq,zengshel,liuhui7,xingyue1,tangjili}@msu.edu \nlingjuan.lv@sony.com", + "bbox": [ + 189, + 234, + 807, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 304, + 326, + 319 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Text-to-image (T2I) diffusion models have shown exceptional capabilities in generating images that closely correspond to textual prompts. However, the advancement of T2I diffusion models presents significant risks, as the models could be exploited for malicious purposes, such as generating images with violence or nudity, or creating unauthorized portraits of public figures in inappropriate contexts. To mitigate these risks, concept removal methods have been proposed. These methods aim to modify diffusion models to prevent the generation of malicious and unwanted concepts. Despite these efforts, existing research faces several challenges: (1) a lack of consistent comparisons on a comprehensive dataset, (2) ineffective prompts in harmful and nudity concepts, (3) overlooked evaluation of the ability to generate the benign part within prompts containing malicious concepts. To address these gaps, we propose to benchmark the concept removal methods by introducing a new dataset, Six-CD, along with a novel evaluation metric. In this benchmark, we conduct a thorough evaluation of concept removals, with the experimental observations and discussions offering valuable insights in the field.", + "bbox": [ + 88, + 335, + 485, + 652 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 679, + 220, + 694 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Diffusion models have demonstrated remarkable capabilities in image generation [6, 11, 12, 40, 41], particularly text-to-image (T2I) diffusion models [4, 29, 32]. These models have gained widespread popularity and deployment due to their ability to generate images that precisely align with textual prompts. However, T2I diffusion models can be exploited for malicious purposes, such as creating images depicting violence, nudity, or fake celebrity images in undesirable contexts like jail [14, 20, 26, 30, 33, 38, 39]. To ensure the integrity of the model developers, it is crucial to design benign models that censor malicious concepts and produce only safe and appropriate content.", + "bbox": [ + 89, + 703, + 482, + 885 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Strategies have been proposed to mitigate the malicious", + "bbox": [ + 109, + 885, + 482, + 900 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "generation. For example, in the data collection stage, Stable Diffusion (SD) [32] and Adobe [31] filter out the Not-Safe-For-Work (NSFW) contents in training data. After image generation, SD uses a detector to ensure the images are benign, removing unwanted ones. However, training data filtering requires retraining and will be inefficient if the model has already been trained on various data. For open-sourced models like SD [32], the detector can be easily disabled by modifying the source code. Therefore, concept removal techniques have been proposed to edit a trained model's parameters to prevent generating malicious concepts [7, 18, 19, 34, 44]. For instance, methods are proposed to fine-tune the models to replace unwanted concepts with benign alternatives [7, 18, 44]. Besides, to accelerate fine-tuning, some methods fine-tune the linear components (e.g. the weight matrix of cross attention) using the closed-form optimal solution [8, 22, 42]. Alternatively, there are methods modifying the output during inference to bypass fine-tuning [3, 36].", + "bbox": [ + 511, + 305, + 906, + 592 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "While there is rich literature improving concept removal methods, we observe three potential issues:", + "bbox": [ + 511, + 595, + 906, + 625 + ], + "page_idx": 0 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. Lack of consistent and comprehensive comparisons. Current concept removal methods, as shown in Table 1, typically analyze limited categories, lacking consistent and comprehensive comparisons. This gap prevents a thorough understanding of the methods' behavior.", + "2. Ineffective prompts. In existing datasets [36], some categories, such as nudity, contain a large portion of \"ineffective prompts\", which only trigger the model to generate malicious contents with a low probability (see Fig. 1). This can result in inefficient evaluation since the ineffective prompts will generate many benign images, and evaluating concept removal on them is meaningless. Meanwhile, in other categories of unwanted concepts like celebrities, the prompts are more effective than the nudity concept (see Fig. 1 and Fig. 4). When the builder has to remove both nudity and celebrity concepts, the ineffective prompts may cause unfair comparisons. The effective prompts in celebrity can generate" + ], + "bbox": [ + 519, + 628, + 908, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "28769", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "table", + "img_path": "images/fdff7c92cf7e25216b8e45986f0117aaba33044d8dd9795ac839f304512ae344.jpg", + "table_caption": [ + "Table 1. Categories of unwanted concepts in existing literature" + ], + "table_footnote": [], + "table_body": "
ESD [7]SPM [23]SDD [16]FMN [44]UCE [8]MACE [22]EMCID [42]SLD [36]SEGA [3]UC [45]CPDM [24]
Harmful✓✓✓
Nudity✓✓✓✓✓✓✓✓
Celebrity✓✓✓
Copyrighted✓✓
Object✓✓✓✓✓✓✓✓
art style✓✓✓✓✓✓✓✓✓✓
", + "bbox": [ + 94, + 114, + 903, + 210 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "more malicious images, which may falsely induce the builder to put more weight on celebrity concepts.", + "bbox": [ + 114, + 236, + 480, + 267 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Lack of evaluation on in-prompt retainability. Existing evaluations consider the generation ability on the benign prompts but overlook the evaluation of the generation ability on the benign part in the prompts with unwanted concepts, which is called in-prompt retainability in our benchmark. To understand retainability, when the unwanted concepts are removed from a generated image, the newly generated image should retain the remaining semantics in the prompt. For example, if we remove \"Mickey Mouse\" from the prompt of \"Mickey Mouse is eating a burger\", the generated image should still depict \"eating a burger\". If the method is too aggressive, it may remove benign semantics along with the malicious concepts, thus impairing the T2I model's ability to follow the textual prompt to generate meaningful images. Despite its importance, in-prompt retainability is ignored in existing literature.", + "bbox": [ + 96, + 267, + 482, + 523 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To tackle the aforementioned problems, we aim to benchmark concept removal methods by introducing a comprehensive dataset and a new evaluation metric. Our contributions are as follows:", + "bbox": [ + 89, + 527, + 483, + 587 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. A comprehensive dataset. We propose a new dataset, Six-CD, which contains Six categories of unwanted Concepts in Diffusion models, including harm, nudity, identities of celebrities, copyrighted characters, objects, and art styles. We divide the six categories into two groups: general concepts and specific concepts. General concepts, such as harm and nudity, are of concern to all users. Specific concepts are subject to a specific entity, e.g., a person or a company. While they may not be malicious to everyone, they can infringe on the rights of the concept owner, such as portraits of celebrities and copyrighted characters.", + "2. Effective prompts. Through extensive experiments in Sec. 3.2 and Sec. 5.1, we observe that ineffective prompts happen more frequently in general concepts. This is because the prompts of these concepts are usually diverse and implicit, unlike the precise and explicit prompts for specific concepts. Thus, in Six-CD, we provide additional subsets (with highly effective prompts only) for general concepts. Evaluation of ef" + ], + "bbox": [ + 96, + 598, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "fective prompts can be more efficient and also fairer when compared with specific concepts.", + "bbox": [ + 539, + 236, + 903, + 267 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. A new evaluation metric. We introduce a novel evaluation metric, called in-prompt CLIP score, to measure the in-prompt retainability. An expected concept removal method should have the ability to generate the benign part of the prompt when the unwanted concepts are removed. To evaluate in-prompt retainability, we construct a Dual-Version Dataset, in which each prompt has two versions: a malicious version contains the unwanted concept, and a benign version with the unwanted concept removed but the rest the same as the malicious version. We apply concept removal methods to generate images from the malicious version and then use the CLIP score [10] to measure the similarity between these images and the benign prompts. Ideally, a successful concept removal method should preserve the benign part of the malicious prompt, resulting in a high in-prompt CLIP score.", + "bbox": [ + 519, + 267, + 906, + 521 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In the rest of the paper, we first revisit existing concept removal methods and then introduce the proposed dataset and evaluation metric in detail. Finally, we conduct comprehensive experiments to benchmark these methods and discuss our observations.", + "bbox": [ + 511, + 521, + 905, + 595 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Preliminaries", + "text_level": 1, + "bbox": [ + 511, + 607, + 651, + 622 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Text-to-image Diffusion Models", + "text_level": 1, + "bbox": [ + 511, + 632, + 792, + 648 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Diffusion models typically involve a forward diffusion process and its reverse diffusion process. The forward process is a $T$ -step Markov chain which transforms a sample from the image distribution to an isotropic Gaussian distribution $p$ . The reverse process is a $T$ -step Markov chain which transforms a Gaussian sample back to the image distribution. The reverse process can be represented as", + "bbox": [ + 511, + 654, + 905, + 760 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\np _ {\\theta} \\left(x _ {0: T}\\right) = p \\left(x _ {T}\\right) \\prod_ {t = 1} ^ {T} p _ {\\theta} \\left(x _ {t - 1} \\mid x _ {t}\\right) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 563, + 765, + 903, + 785 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\np _ {\\theta} \\left(x _ {t - 1} \\mid x _ {t}\\right) = \\mathcal {N} \\left(x _ {t - 1}; \\mu_ {\\theta} \\left(x _ {t}, t\\right), \\Sigma_ {\\theta} \\left(x _ {t}, t\\right)\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 535, + 786, + 903, + 803 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "where $x_0$ is the image sample, $x_{T}$ is the Gaussian sample, and $x_{t}(0 < t < T)$ is the middle state of diffusion. The terms $\\mu_{\\theta}$ and $\\Sigma_{\\theta}$ are estimated by a denoising network $\\epsilon_{\\theta}(x_t,t)$ which takes $x_{t}$ and $t$ as input. By $\\epsilon_{\\theta}$ , an image sample $x_0$ can be generated from a Gaussian sample $x_{T}$ following Eq. (1).", + "bbox": [ + 511, + 810, + 905, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "28770", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Text-to-image (T2I) diffusion models further extend the diffusion process by guiding it with a textual prompt $y$ [4, 15, 35]. Specifically, $\\epsilon_{\\theta}(x_t,t)$ is modified to take $y$ as an additional input, resulting in $\\epsilon_{\\theta}(x_t,y,t)$ . The network $\\epsilon_{\\theta}(x_t,y,t)$ uses a cross-attention module to select information from the textual prompt to guide the diffusion process. In cross attention, each token is transformed into key $(K)$ space and value $(V)$ space, while the image is transformed into query $(Q)$ space. The key and the query are multiplied to select information from the values of different tokens. With cross attention, T2I models can generate images that adhere to the provided text prompts.", + "bbox": [ + 89, + 90, + 480, + 272 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Concept Removals", + "text_level": 1, + "bbox": [ + 89, + 281, + 272, + 296 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this subsection, we revisit representative concept removal methods.", + "bbox": [ + 89, + 301, + 482, + 333 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Fine-tuning-based methods. Fine-tuning is the most common framework for concept removals. The objective for the fine-tuning-based algorithms can be summarized as", + "bbox": [ + 89, + 333, + 483, + 378 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\theta} L = L _ {\\mathrm {r m}} + \\lambda L _ {\\mathrm {r e g}}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 209, + 402, + 482, + 424 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $L_{\\mathrm{rm}}$ is the term for changing the output of unwanted concept, and $L_{\\mathrm{reg}}$ is the regularization term to ensure that the fine-tuning will not influence other benign concepts and maintain the generation quality. The term $\\theta$ is the parameter to fine-tune. The methods can be solved by two types of solutions:", + "bbox": [ + 89, + 433, + 483, + 523 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Gradient descent. This solution [7, 16, 18, 23] focuses on the final output of $\\epsilon_{\\theta}(x_t,y,t)$ , modifying the model in an end-to-end way. The term $L_{\\mathrm{rm}}$ changes the output of $\\epsilon_{\\theta}(x_t,c_u,t)$ where $c_{u}$ is unwanted concepts, while $L_{\\mathrm{reg}}$ maintains the output of benign concepts $c_b$ .", + "- Closed-form solution. Modifying the intermediate states of linear components (such as cross-attention weights [8, 22] and MLP layers [42]) instead of the final output of $\\epsilon_{\\theta}(x_t, y_c, t)$ , can be solved in a closed-form solution. This can accelerate the fine-tuning process remarkably." + ], + "bbox": [ + 89, + 525, + 482, + 676 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Inference-time methods. Inference-time methods such as negative prompt [32], SLD [36], and SEGA [3] change the generation algorithm to remove concepts in the inference process, which can skip fine-tuning operations. They estimate the influence of unwanted concepts and remove them in inference. However, although these methods can skip the fine-tuning process, there is also a risk of being disabled in open-sourced models, which may slow down the inference.", + "bbox": [ + 89, + 679, + 482, + 813 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Six-CD: a Comprehensive and Effective Dataset", + "text_level": 1, + "bbox": [ + 89, + 825, + 482, + 859 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we first propose a new dataset that encompasses a comprehensive set of categories of unwanted con", + "bbox": [ + 89, + 869, + 482, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "cepts. Then we discuss the issue of ineffectiveness in general prompts and present the solution to filter the dataset, thereby increasing its effectiveness.", + "bbox": [ + 511, + 90, + 903, + 137 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Six Categories of Unwanted Concepts", + "text_level": 1, + "bbox": [ + 511, + 147, + 839, + 165 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As mentioned in Sec. 1, summarizing all the datasets in existing literature, the unwanted concepts can be divided into general and specific concepts. The general concepts contain harmful concepts and nudity concepts, while the specific concepts contain identities of celebrities, copyrighted characters, objects, and art styles. However, existing literature focuses on only a few categories for evaluation, lacking systematic comparisons among the methods in general and specific concepts. To address this, we propose a new dataset, Six-CD, to evaluate concept removals comprehensively.", + "bbox": [ + 511, + 170, + 903, + 335 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For the general concepts, Six-CD first collects the malicious prompts from four different NSFW resources, which include I2P [36], MMA [43], jatman/stable-diffusion-prompts-stats-full-uncensored (SD-uncensored) [13], and Unsafe Diffusion (UD) [27]. Then, it divides the NSFW data into harmful and nudity concepts. Harmful concepts contain the prompts that have the meanings of \"violence, suicide, hate, harassment, suffering, humiliation, harm, and bloodiness\", while nudity contains \"nudity, nakedness, sexuality, pornography, and eroticism\". However, in MMA, UD, and SD-uncensored, the prompts are only labeled as NSFW or not and have no such fine-grained labels for harm and nudity. Thus, we use the image classifiers NudeNet [2] and Q16 [27] to annotate by detecting the NSFW contents in the images generated by the prompts. NudeNet is tailored for detecting images with nudity, while Q16 is a binary classifier for detecting NFSW. To annotate the fine-grained labels, i.e., harm and nudity, we first use NudeNet to find the prompts of nudity from the collected resources1. Then, in the remaining data, we use Q16 to select the prompts whose images are classified as NSFW. Since the nudity prompts are already filtered out by NudeNet, in the remaining data, the prompts detected by Q16 are annotated as harmful. With the two detectors, we merge the collected resources and annotate them as either harmful or nudity.", + "bbox": [ + 511, + 337, + 903, + 714 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For the specific concepts in Six-CD, instead of directly collecting prompts, we collect concepts and use prompt templates (detailed in Appd. A.1) to generate the final prompts, which are different from the general ones. The concepts are collected as follows:", + "bbox": [ + 511, + 715, + 903, + 789 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "- Celebrity. Celebrities can evaluate the ability to remove identity. We select the identities of celebrities from CPDM [24]. To assist the evaluation, we use a celebrity detector, GCD [25]. we choose the celebrities that can be", + "bbox": [ + 511, + 791, + 903, + 852 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "1The image is labeled as nudity if it is classified as \"FEMALE/MALE GENITALIA EXPOSED\", \"FEMALE BREAST EXPOSED\", \"ANUS EXPOSED\", or \"BUTTOCKS EXPOSED\" by NudeNet.", + "bbox": [ + 511, + 862, + 903, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "28771", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/5f3fdef1887df2868e0e5bd0c4c975835ee2f6fd076e25d91d796ce3de310fc8.jpg", + "image_caption": [ + "Figure 1. Effectiveness of I2P prompts. Most of the prompts have low effectiveness. Results from [36]." + ], + "image_footnote": [], + "bbox": [ + 135, + 90, + 441, + 265 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "both generated by SD and recognized by GCD.", + "bbox": [ + 102, + 335, + 416, + 351 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Copyrighted characters. The copyrighted characters are important IP resources for companies. We use the character concepts in CPDM and the high-frequency copyrighted characters from FiveThirtyEight Comic Characters Dataset [1].", + "- Objects. Objects can be used to evaluate the performance of removal on non-humanoid concepts. We randomly sample from a subset from the classes of ImageNet [5].", + "- Art style. The art styles can evaluate the removal performance of a global feature instead of a local feature. We use the art styles from [22]." + ], + "bbox": [ + 89, + 352, + 482, + 518 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For the two general categories, we provide 6,326 prompts for harm concept and 4,518 effective prompts for nudity concept. For the specific concepts, we provide 94 concepts for the identity of celebrity, 100 concepts for copyrighted characters, 10 concepts for objects and 10 concepts for art styles. Four templates are used to construct the prompt for each specific concept. Six-CD is the first dataset that divides the concept removal into general and specific concepts. Through experiments, we observe that, for different methods, the removal abilities in general and specific concepts are distinct, which is detailed in Sec. 5.1. Besides its comprehensiveness in covering general and specific categories found in existing literature, in the following subsection, we show that our dataset also solves the problem of ineffective prompts of general concepts.", + "bbox": [ + 88, + 518, + 485, + 744 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Ineffective Prompts of General Concepts in Existing Datasets", + "text_level": 1, + "bbox": [ + 89, + 756, + 482, + 787 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The prompts for general concepts in existing dataset do not consistently generate malicious content for each random seed according to [36] and our observation below. In this subsection, we discuss this phenomenon in general concepts and identify two potential problems it may introduce.", + "bbox": [ + 89, + 794, + 482, + 869 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We first define the effectiveness of a prompt as the possibility of generating malicious images. To assess this, we", + "bbox": [ + 89, + 869, + 482, + 901 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/c8a8662863c0c0805e0a90f364c6dcc077de6afccb53e12337b83d7f2a04eeeb.jpg", + "image_caption": [ + "Figure 2. Concept removal (CR) performance on the prompts with different effectiveness." + ], + "image_footnote": [], + "bbox": [ + 555, + 90, + 864, + 267 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/3d75e7d165497c428e56e646266b3c13ac423ba357c8941004686d50460799ce.jpg", + "table_caption": [ + "Table 2. Percentage of removed unwanted concepts between high effective and all effectiveness. (%)" + ], + "table_footnote": [], + "table_body": "
NEGESDSPMSDDFMNUCEEMCIDSLDSEGACorr
Harm40.963.698.924.895.677.698.927.323.70.977
Harm-EF56.767.399.927.897.467.898.638.122.9
Nudity22.317.981.011.293.614.998.314.638.80.984
Nudity-EF11.016.088.94.292.612.397.216.358.4
", + "bbox": [ + 516, + 364, + 903, + 425 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/c4c1148edf8b30c7deedbd4a3186433924a5556d4aba924d2c192d89d457ea99.jpg", + "table_caption": [ + "Table 3. Examples of Dual-Version Dataset for specific concepts. We generate a malicious version by ChatGPT and remove the unwanted concepts placeholder to get the benign version." + ], + "table_footnote": [], + "table_body": "
CategoryExample of malicious versionExample of benign version
Celeb. & \nCopyrighted{celebrity} is dancing in the rain.dancing in the rain.
Object{object}, football, grass, house, tree, dogfootball, grass, house, tree, dog
Art styleA beautiful snow-covered mountain with \nsunshine lighting it in the style of {art style}A beautiful snow-covered moun-\ntain with sunshine lighting it
", + "bbox": [ + 516, + 484, + 903, + 560 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "generate $N$ images for a prompt using different random seeds. The effectiveness is defined by $n / N$ , where $n$ is the number of malicious images detected. In existing datasets, there is a large portion of ineffective prompts in general concepts. Taking I2P [36] as an example, which is the most well-used dataset for harmful and nudity concepts in existing literature [3, 7, 8, 16, 22, 23, 36, 42], most of the prompts have low effectiveness as shown in Fig. 1. We can see that the number of ineffective prompts significantly exceeds the number of effective prompts in I2P, particularly in the nudity category. The specific concepts usually do not have the problem of low effectiveness because they contain the specific and definitive for the unwanted concepts. Thus, in this subsection, we focus on general concepts.", + "bbox": [ + 509, + 580, + 906, + 792 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Limitations. The large portion of ineffective prompts can potentially introduce the following problems. First, evaluating the ineffective prompts is in low efficiency. For example, we test different concept removal methods on SD for nudity concepts in Fig. 2. On the vanilla SD without concept removal (i.e., No CR in Fig. 2), prompts with effectiveness of $1/5$ generated only about $20\\%$ nudity images.", + "bbox": [ + 509, + 795, + 908, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "28772", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "It means that $80\\%$ of the generated images by these prompts are non-nudity images, while evaluating on non-nudity images is meaningless. Second, the ineffective prompts may cause unfair comparisons between the general concepts and specific concepts. Unlike general concepts, the specific concepts do not have the problem of ineffective prompts. When the model builder has to remove both general and specific concepts, the ineffective prompts may falsely induce the builder to put less weight on the general concepts.", + "bbox": [ + 89, + 90, + 480, + 226 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "High-effectiveness subsets. Therefore, in addition to harm and nudity concepts, we provide additional subsets of high-effectiveness prompts and name them as harm- $EF$ and nudity- $EF$ . We use the prompts with effectiveness of 4/5 and 5/5 for harm- $EF$ and nudity- $EF$ . Harm- $EF$ contains 991 prompts, and nudity- $EF$ contains 1,539 prompts. In Table 2 and Fig. 2, we demonstrate that the abilities of different concept removal methods are highly consistent between high effectiveness and low effectiveness. Thus, it is reasonable to evaluate on high-effectiveness prompts in order to reduce computation, which is also more fair. Specifically, in Table 2, we use the percentage of removed unwanted concepts to measure the ability of different methods, which is defined as", + "bbox": [ + 91, + 227, + 483, + 436 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\np _ {\\mathrm {r m}} = 1 - \\frac {n _ {\\text {a f t e r}}}{n _ {\\text {b e f o r e}}}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 222, + 446, + 482, + 474 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $n_{\\mathrm{before}}$ is the number of images containing unwanted concepts before concept removal, while $n_{\\mathrm{after}}$ is the number of images containing unwanted concepts after concept removal. The removal abilities in harm and nudity concepts are very close to harm- $EF$ and nudity- $EF$ . The Pearson correlation coefficient is as high as 0.977 (between harm and harm- $EF$ ) and 0.984 (between nudity and nudity- $EF$ ).", + "bbox": [ + 89, + 484, + 482, + 589 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In summary, Six-CD provides two choices to evaluate on the general concepts. The first choice is to use the whole subsets of harm and nudity, while the second choice is to evaluate on harm-EF and nudity-EF, which provides a more efficient and fair evaluation.", + "bbox": [ + 89, + 589, + 482, + 664 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. In-prompt Retainability", + "text_level": 1, + "bbox": [ + 89, + 679, + 316, + 696 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Concept removals should not hurt the generation ability of the benign contents, which is referred to as retainability. In this section, we show the lack of evaluation for retainability on the benign parts of the prompts containing the unwanted concepts and propose a new metric for this retainability.", + "bbox": [ + 89, + 704, + 482, + 779 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Lack of in-prompt retainability. Existing literature [24, 45] only considers the retainability of totally benign concepts. The benign prompts do not contain any unwanted concept, and concept removal should not influence the generation of them. Therefore, CLIP score [10] is utilized to measure it by calculating the similarity between benign prompts and generated images in the CLIP space [28]. However, for the prompts containing unwanted concepts,", + "bbox": [ + 89, + 780, + 482, + 901 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/1ad7c80faac9de60e308487ea35a535ffca25bea4ba81bca07217cbc58ad40c2.jpg", + "image_caption": [ + "NO CR", + "ironman removed", + "Figure 3. Concept removal example of the prompt: \"At night, ironman with his armor on is watching fireworks by the lake\". Ironman is removed by SDD." + ], + "image_footnote": [], + "bbox": [ + 588, + 102, + 707, + 195 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/87b1de8ea1f3d9fe8a7adf39cab04b716b816c699eeb5310c2033f3b2365bc66.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 710, + 102, + 831, + 195 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "the retainability to guarantee the generation of the benign part of them is also important, which is called \"in-prompt retainability\" in our benchmark. As shown in Fig. 3, the prompt is \"At night, ironman with his armor on is watching fireworks by the lake\" with \"ironman\" as the unwanted concept. Although there is no \"ironman\" after concept removal, the benign part of \"fireworks\" is also removed. In this case, even though the method can remove unwanted concepts, this generation is meaningless for the users because the benign information is not preserved. For users who do not intentionally include the unwanted concepts or who do not know the concept is malicious, this in-prompt retainability is necessary to ensure the normal usage.", + "bbox": [ + 511, + 266, + 906, + 462 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "New metric. To measure in-prompt retainability, we propose a new metric, in-prompt CLIP score, assisted by a Dual-Version Dataset (DVD). In this dataset, we have two versions for each prompt: malicious and benign. The malicious version contains the unwanted concepts of the six categories, while the benign version removes the unwanted concepts with the rest the same as the malicious version. In in-prompt CLIP score, we first generate the images with concept removals using the malicious version and then calculate the cosine similarity of CLIP embeddings between these generated images and the corresponding benign version prompts. Thus, we can measure the similarity between the generated image and the benign part of the prompt. To distinguish it from the original CLIP score, we refer to the original CLIP score as the out-prompt CLIP score.", + "bbox": [ + 511, + 463, + 905, + 688 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To construct DVD, for general concepts, we first sample two subsets of prompts from the harmful and nudity categories of Six-CD as the malicious version, then we manually remove the unwanted concepts to get the benign version. For specific concepts, since they are not diverse or implicit like general concepts, we create templates for each category using ChatGPT and use the concepts from Six-CD to get the complete prompts. We show the examples of templates in DVD in Table 3. The entire dataset is in Appd. A.2. To validate the effectiveness of the metric, we conduct human evaluation in Appd. D.1. The results demonstrate that the metric is highly consistent with human preference.", + "bbox": [ + 511, + 689, + 905, + 869 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In summary, to measure the in-prompt retainability, we propose the in-prompt CLIP score. It leverages a malicious-", + "bbox": [ + 511, + 869, + 903, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "28773", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "version and a benign-version prompt to measure the similarity between the images after concept removal and the benign part of the prompt.", + "bbox": [ + 89, + 90, + 485, + 137 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 89, + 148, + 225, + 167 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we conduct experiments to benchmark the performance of 10 methods in removing single and multiple concepts and the retainability with both in-prompt and out-prompt CLIP scores. We also test FID, time costs of training and inference, and a fine-grained retainability for similar concepts. Due to space limitations, these experiments are in Appd D. Also, we report the results of SD v1.4 in this section and the results of other models are in Appd D.", + "bbox": [ + 89, + 174, + 482, + 296 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Experimental settings. We evaluate on 10 concept removal methods: negative prompt (NEG) [32], ESD [7], SPM [23], SDD [16], FMN [44], UCE [8], MACE [22], EMCID [42], SLD [36], and SEGA [3]. The details on the baseline settings can be found in Appd. B. The evaluated datasets are Six-CD and DVD, which are detailed in Appd. A. Following the conclusions in Sec. 3.2, we use harm-EF and nudity-EF for all the experiments in this section. All experiments use one GPU (A5000 or A6000).", + "bbox": [ + 89, + 296, + 482, + 431 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Detection metrics. We use the detection rate of Q16 and NudeNet for harmful and nudity concepts. We use the classification accuracy of GCD as the detection rate for celebrity concepts, and we train two classifiers based on ResNet-50 [9] and use the accuracy as the detection rate for copyrighted characters and objects. For the art style concept, the detection is not a simple binary problem. In existing literature [22-24, 44], CLIP score, $S_{\\mathrm{CLIP}}$ , is used to measure the presence of an art style on a continuous scale, assessing the similarity between generated images and art styles. Following them, we use $S_{\\mathrm{CLIP}}$ and normalize its range by $\\tilde{S}_{\\mathrm{CLIP}} = (S_{\\mathrm{CLIP}} - \\min(S_{\\mathrm{CLIP}})) / (\\max(S_{\\mathrm{CLIP}}) - \\min(S_{\\mathrm{CLIP}}))$ for a better comparison with other categories.", + "bbox": [ + 89, + 431, + 482, + 628 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Evaluation on Removal Ability", + "text_level": 1, + "bbox": [ + 89, + 637, + 364, + 652 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this subsection, we test the removal ability of different methods on Six-CD. In Fig. 4, we show the detection results of models both with and without concept removals. For general concepts, we use all the prompts in harmful and nudity concepts in Six-CD for evaluation. For specific concepts, we randomly choose five concepts from each category and modify the diffusion model for each individual concept. We show the averaged results of 5 concepts of each category in Fig. 4. (The table with error bar for Fig. 4 can be found in Appd. D.2.) We highlight the main observations as follows.", + "bbox": [ + 89, + 659, + 482, + 823 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Observation 1: General concepts are harder to remove. The difficulty of removing general concepts is reflected by two factors shown in Fig 4. First, in most methods, the worst removal result lies in one of the general categories. For example, ESD and MACE perform well in all", + "bbox": [ + 89, + 825, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "other categories, but they still have around $55\\%$ harmful images detected. Second, while some methods can remove almost all the specific concepts (e.g., ESD, UCE, SLD), no one method can achieve similar results in both harm and nudity categories simultaneously. For general concepts, on the one hand, they are hard to trigger (i.e. low effectiveness). On the other hand, it is harder to remove them than specific concepts. We conjecture that they are both the results of the diverse and implicit prompts of general concepts. Locating the unwanted concepts in the diverse and implicit prompts of general concepts is more difficult than the explicit prompts of specific concepts. Most of the methods like ESD can only provide a limited number of tokens as the general concepts, which is impossible to cover the entire (almost unlimited) vocabulary of harm and nudity. Thus, it leads to difficulty in removing general concepts.", + "bbox": [ + 511, + 90, + 903, + 332 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Observation 2: Removal ability is consistent within general concepts and within specific categories. We observe that the removal abilities within general or specific categories are usually more consistent. For example, SLD and FMN perform well in specific concepts, but the ability to remove general concepts is worse. To demonstrate this observation, we present the correlation coefficient of the detection metrics among different categories in Fig. 5. From Fig. 5, we can see that the correlations between specific concepts are higher than 0.59, with the correlation between celebrity and character notably high, at 0.94. The correlation between general concepts is also as high as 0.65. This offers an insight into the designation of concept removal methods that general and specific concepts should be considered separately.", + "bbox": [ + 511, + 334, + 903, + 561 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Removal Ability on Multiple Concepts", + "text_level": 1, + "bbox": [ + 511, + 574, + 843, + 589 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In Sec. 5.1, we evaluate the ability to remove a single concept from one model. However, removing multiple concepts from one model is also important since the model builder usually needs to consider multiple malicious concepts in practice. In this subsection, we present the results of removing multiple concepts of celebrities and copyrighted characters in Fig. 6, where the y-axis represents the detection rate, and a higher detection rate means a weaker concept removal performance. We have two observations from Fig. 6.", + "bbox": [ + 511, + 597, + 903, + 733 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Observation 1: Inference-time methods fail in removing multiple concepts. As shown in Fig. 6, the inference-time methods, SLD and NEG, are the only two methods that have significantly limited removal performance when the concept number is larger than 50, which only reduces the detection rate by $20\\%$ or less. To remove the multiple concepts in inference time, they have to encode the string containing all the concepts in the embeddings of one single prompt. This will exceed the capacity of the text encoder of T2I diffusion models and lead to failed removal. Another inference-time method, SEGA, is not reported in Fig. 6 due", + "bbox": [ + 511, + 734, + 903, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "28774", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/7ddc365e0d2702dfb1165ce44bc08ace002f3e00706682b06ee3df244d451943.jpg", + "image_caption": [ + "Figure 4. Removal ability on Six-CD. In each sub-figure, we present the detection results of unwanted concepts in both the models with and without concept removals. The range of all the six values is [0,1]. Smaller values indicate that less unwanted concepts are detected, i.e. better removal. G1 and G2 are harm-EF and nudity-EF concepts, while S1, S2, S3 and S4 are celebrity, character, object and art style." + ], + "image_footnote": [], + "bbox": [ + 99, + 79, + 897, + 369 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e6b4f9fd9ba6540935fd33b7f7f0ebc824ab9fcefb3378fe96c1b890c67b027f.jpg", + "image_caption": [ + "Figure 5. Correlation coefficient of removal performance on six data categories." + ], + "image_footnote": [], + "bbox": [ + 135, + 439, + 426, + 625 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "to its catastrophic inference time when the concept number is increased to 50. This implies that the strategy of processing each concept separately during inference is impractical.", + "bbox": [ + 89, + 686, + 482, + 733 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Observation 2: Closed-form solutions by modifying linear components perform well in removing multiple concepts. The best two methods in Fig. 6a and Fig. 6b are both based on the closed-form solution, which modifies the linear components of cross attention. This is possibly because combining multiple concepts in the linear components is easier than in other non-linear parts.", + "bbox": [ + 89, + 733, + 482, + 839 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. In-prompt and Out-prompt Retainability", + "text_level": 1, + "bbox": [ + 89, + 848, + 442, + 864 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We evaluate the generation ability on benign concepts using both in-prompt and out-prompt CLIP scores. In Fig. 7, we", + "bbox": [ + 89, + 869, + 482, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "use DVD to test the in-prompt CLIP score and a subset with benign prompts from LAION [37] to test out-prompt CLIP score. The observations as follows.", + "bbox": [ + 511, + 448, + 906, + 493 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Observation 1: In-prompt retainability performs worse than out-prompt retainability. As shown in Fig. 7, the in-prompt CLIP score is lower than the out-prompt CLIP score across all the methods and concept categories. This indicates that the ability to generate benign parts in prompts containing malicious concepts is more negatively affected than the ability to generate totally benign prompts. As mentioned in Sec. 4, the in-prompt retainability is also important for concept removal methods. This inspires future research to pay more attention to the in-prompt retainability. Remark. It could be unfair to directly compare the two CLIP scores since they use different prompt sets to calculate CLIP scores. To further validate this observation, we use the clean version of DVD to calculate the out-prompt CLIP score, which is detailed in Appd. D.3. By comparing it with the in-prompt CLIP score in Appd. D.3, we have observed a consistent phenomenon with this subsection.", + "bbox": [ + 511, + 500, + 906, + 758 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Observation 2: Auxiliary semantic information in MACE (Fig. 7g) may hurt the generation on benign concepts. MACE has the most significant decrease in the generation ability of benign concepts for both in-prompt and out-prompt CLIP scores. We conjecture that this is because it introduces auxiliary semantic information to help locate the unwanted concepts in the image. In other methods, the unwanted concepts are usually located by the semantic understanding ability of the T2I model itself. In contrast,", + "bbox": [ + 511, + 763, + 908, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "28775", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a5d64223235972fafc877ec1758d1b3ae033cac5d1c7d1c262f86b6f2dd43c42.jpg", + "image_caption": [ + "(a) Celebrity" + ], + "image_footnote": [], + "bbox": [ + 145, + 90, + 442, + 258 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f3d1443a7d24269896e363541688c346307daf628affba5589c3860b62b571b2.jpg", + "image_caption": [ + "(b) Copyrighted characters" + ], + "image_footnote": [], + "bbox": [ + 464, + 90, + 849, + 258 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b944d48ca20e7ffa160b2285a843dc8b16df00b09e90ae99db440a22a487f722.jpg", + "image_caption": [ + "(a) NEG" + ], + "image_footnote": [], + "bbox": [ + 93, + 313, + 248, + 458 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/19a87723d288a62a5fd2695fa21a6cedc26011f2ab09c4f0fd8f37762bcd97c8.jpg", + "image_caption": [ + "(b) ESD" + ], + "image_footnote": [], + "bbox": [ + 256, + 313, + 411, + 458 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/363cde74017eb92a28912f424b314834362cc1a4e939f4ec9ccee86e5b8384e5.jpg", + "image_caption": [ + "(c) SPM" + ], + "image_footnote": [], + "bbox": [ + 421, + 314, + 575, + 458 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/299db817f9401300a9b45ea440479838891b67830126357154c1089e5305fd65.jpg", + "image_caption": [ + "Figure 6. Removal ability on multiple concepts" + ], + "image_footnote": [], + "bbox": [ + 584, + 311, + 738, + 458 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c0f38fcbffd8c79370fd6bcee7377c5fd4f36ddc8a9fdfa8cc1a22db194af946.jpg", + "image_caption": [ + "(e) FMN" + ], + "image_footnote": [], + "bbox": [ + 748, + 333, + 903, + 458 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0614252b091d5fe1f58ca4db93471406dab4f59a2668f356a4595a712f209e70.jpg", + "image_caption": [ + "(f) UCE" + ], + "image_footnote": [], + "bbox": [ + 93, + 473, + 246, + 595 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a24fabf24dbbdb08ce3c99fbf769814179a713bdca68eaee51ae78ccd8227391.jpg", + "image_caption": [ + "(g) MACE" + ], + "image_footnote": [], + "bbox": [ + 256, + 473, + 411, + 597 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/debd13ddd2b714f3e825e6052eecb916c51bca6ba03b51e56ce971d4f94772a6.jpg", + "image_caption": [ + "(h) EMCID" + ], + "image_footnote": [], + "bbox": [ + 421, + 473, + 575, + 597 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/95dedda416ff55074e5ee8ab412d29a08ff4cc208807325be0d7db6b110d84f2.jpg", + "image_caption": [ + "(d) SDD", + "(i) SLD" + ], + "image_footnote": [], + "bbox": [ + 584, + 473, + 738, + 597 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1856dd8da1267124d798bf0d2479483fb2f0908c865cf7a27bf7457a69e926ee.jpg", + "image_caption": [ + "(j) SEGA", + "Figure 7. In-prompt retainability vs. out-prompt retainability" + ], + "image_footnote": [], + "bbox": [ + 748, + 473, + 903, + 597 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "MACE incorporates the segmentation results of GroundedSAM [17, 21] to locate the image area of unwanted concepts. It focuses on optimizing this area and overlooks other areas of the image. This may lead to worse retainability in other areas. Another piece of evidence is that when we do not include segmentation in removing the art style concepts by MACE (since art style is a global feature), it has good in-prompt and out-prompt CLIP scores.", + "bbox": [ + 88, + 654, + 482, + 776 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion and Ethical Statement", + "text_level": 1, + "bbox": [ + 89, + 797, + 400, + 811 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we address the lack of a systematic benchmark for concept removal methods by introducing a comprehensive and effective dataset and a new evaluation metric. Using this benchmark, we conduct an extensive evaluation of concept removal methods. Our experimental observations", + "bbox": [ + 89, + 824, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "provide valuable and practical insights for future research in this field. While our benchmark may have a limitation in the inconvenient detection metric for art styles (using the CLIP score for art styles, instead of the detection rate from a classifier like other categories), it still offers a thorough and practical evaluation across the various method settings.", + "bbox": [ + 511, + 655, + 906, + 744 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ethical statement. This work provides a comprehensive evaluation framework for concept removal, which is targeted to mitigate the potential malicious use of text-to-image diffusion models. Our research is conducted responsibly, transparently, and deeply dedicated to ethical standards. Despite involving the generation of sensitive content such as nudity and violence, it is strictly for research purposes and does not intend to produce or promote inappropriate material. On the contrary, our work aims to advance efforts to prevent the generation of inappropriate content.", + "bbox": [ + 511, + 750, + 908, + 901 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "28776", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 91, + 90, + 250, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Jie Ren, Kangrui Chen, Yingqian Cui, Shenglai Zeng, and Jiliang Tang are supported by the National Science Foundation (NSF) under grant numbers CNS2321416, IIS2212032, IIS2212144, IOS2107215, DUE2234015, CNS2246050, DRL2405483 and IOS2035472, the Army Research Office (ARO) under grant number W911NF-21-1-0198, Amazon Faculty Award, JP Morgan Faculty Award, Meta, Microsoft and SNAP.", + "bbox": [ + 89, + 113, + 483, + 224 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 237, + 187, + 252 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Fivethirtyeight comic characters dataset. https://www.kaggle.com/datasets/fivethirtyeight/fivethirtyeight - comic - characters - dataset.4", + "[2] Nudenet: Lightweight nudity detection. https://github.com/notAI-tech/NudeNet, 2023.3", + "[3] Manuel Brack, Felix Friedrich, Dominik Hintersdorf, Lukas Struppek, Patrick Schramowski, and Kristian Kersting. Sega: Instructing text-to-image models using semantic guidance. Advances in Neural Information Processing Systems, 36, 2024. 1, 2, 3, 4, 6", + "[4] Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, et al. Pixart-α: Fast training of diffusion transformer for photorealistic text-to-image synthesis. arXiv preprint arXiv:2310.00426, 2023. 1, 3", + "[5] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 4", + "[6] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 1", + "[7] Rohit Gandikota, Joanna Materzynska, Jaden Fiitto-Kaufman, and David Bau. Erasing concepts from diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2426-2436, 2023. 1, 2, 3, 4, 6", + "[8] Rohit Gandikota, Hadas Orgad, Yonatan Belinkov, Joanna Materzyńska, and David Bau. Unified concept editing in diffusion models. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 5111-5120, 2024. 1, 2, 3, 4, 6", + "[9] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6", + "[10] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning. arXiv preprint arXiv:2104.08718, 2021. 2, 5", + "[11] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 1", + "[12] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840-6851, 2020. 1" + ], + "bbox": [ + 93, + 262, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Huggingface. jatman/stable-diffusion-prompts-stats-full-uncensored datasets at hugging face. https://huggingface.co/datasets/jatman/stable-diffusion-prompts-stats-full-uncensored?not-for-all-audiences=true. 3", + "[14] Tatum Hunter. Ai porn is easy to make now. for women, that's a nightmare. https://www.washingtonpost.com/technology/2023/02/13/ai-porn-deepfakes-women-consent/, 2023. Accessed: 2024-06-01. 1", + "[15] DeepFloyd IF. https://github.com/deep-floyd/if.3", + "[16] Sanghyun Kim, Seohyeon Jung, Balhae Kim, Moonseok Choi, Jinwoo Shin, and Juho Lee. Towards safe self-distillation of internet-scale text-to-image diffusion models. arXiv preprint arXiv:2307.05977, 2023. 2, 3, 4, 6", + "[17] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 8", + "[18] Nupur Kumari, Bingliang Zhang, Sheng-Yu Wang, Eli Shechtman, Richard Zhang, and Jun-Yan Zhu. Ablating concepts in text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22691-22702, 2023. 1, 3", + "[19] Yuantong Li, Chi-Hua Wang, and Guang Cheng. Online forgetting process for linear regression models. In International Conference on Artificial Intelligence and Statistics, pages 217-225. PMLR, 2021. 1", + "[20] Yaxin Li, Jie Ren, Han Xu, and Hui Liu. Neural style protection: Counteracting unauthorized neural style transfer. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 3966-3975, 2024. 1", + "[21] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023.8", + "[22] Shilin Lu, Zilan Wang, Leyang Li, Yanzhu Liu, and Adams Wai-Kin Kong. Mace: Mass concept erasure in diffusion models. arXiv preprint arXiv:2403.06135, 2024. 1, 2, 3, 4, 6", + "[23] Mengyao Lyu, Yuhong Yang, Haiwen Hong, Hui Chen, Xuan Jin, Yuan He, Hui Xue, Jungong Han, and Guiguang Ding. One-dimensional adapter to rule them all: Concepts, diffusion models and erasing applications. arXiv preprint arXiv:2312.16145, 2023. 2, 3, 4, 6", + "[24] Rui Ma, Qiang Zhou, Bangjun Xiao, Yizhu Jin, Daquan Zhou, Xiuyu Li, Aishani Singh, Yi Qu, Kurt Keutzer, Xiaodong Xie, et al. A dataset and benchmark for copyright protection from text-to-image diffusion models. arXiv preprint arXiv:2403.12052, 2024. 2, 3, 5, 6", + "[25] Dmitry Voitekh Nick Hasty, Ihor Kroosh and Dmytro Korduban. Giphy's open source celebrity detection deep learning model and code. https://github.com/Giphy/celeb-detection-oss, 2019.3", + "[26] OpenAI. Dall-e 2 preview - risks and limitations. https://github.com/openai/dalle-2-preview/" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "28777", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "blob/main/system-card.md, 2023. Accessed: 2024-06-01. 1", + "[27] Yiting Qu, Xinyue Shen, Xinlei He, Michael Backes, Savvas Zannettou, and Yang Zhang. Unsafe Diffusion: On the Generation of Unsafe Images and Hateful Memes From Text-To-Image Models. In ACM SIGSAC Conference on Computer and Communications Security (CCS). ACM, 2023. 3", + "[28] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 5", + "[29] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 1", + "[30] Javier Rando, Daniel Paleka, David Lindner, Lennart Heim, and Florian Tramèr. Red-teaming the stable diffusion safety filter. arXiv preprint arXiv:2210.04610, 2022. 1", + "[31] Dana Rao. Responsible innovation in the age of generative ai: Adobe blog, . 1", + "[32] Dana Rao. Stable diffusion 2.0 release, .1, 3, 6", + "[33] Jie Ren, Han Xu, Pengfei He, Yingqian Cui, Shenglai Zeng, Jiankun Zhang, Hongzhi Wen, Jiayuan Ding, Hui Liu, Yi Chang, et al. Copyright protection in generative ai: A technical perspective. arXiv preprint arXiv:2402.02333, 2024. 1", + "[34] Jie Ren, Zhenwei Dai, Xianfeng Tang, Hui Liu, Jingying Zeng, Zhen Li, Rahul Goutam, Suhang Wang, Yue Xing, and Qi He. A general framework to enhance fine-tuning-based llm unlearning. arXiv preprint arXiv:2502.17823, 2025. 1", + "[35] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 3", + "[36] Patrick Schramowski, Manuel Brack, Björn Deiseroth, and Kristian Kersting. Safe latent diffusion: Mitigating inappropriate degeneration in diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22522-22531, 2023. 1, 2, 3, 4, 6", + "[37] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278–25294, 2022. 7", + "[38] Shawn Shan, Jenna Cryan, Emily Wenger, Haitao Zheng, Rana Hanocka, and Ben Y Zhao. Glaze: Protecting artists from style mimicry by {Text-to-Image} models. In 32nd USENIX Security Symposium (USENIX Security 23), pages 2187–2204, 2023. 1", + "[39] Gowthami Somepalli, Vasu Singla, Micah Goldblum, Jonas Geiping, and Tom Goldstein. Diffusion art or digital forgery? investigating data replication in diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6048-6058, 2023. 1" + ], + "bbox": [ + 91, + 90, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[40] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 1", + "[41] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 1", + "[42] Tianwei Xiong, Yue Wu, Enze Xie, Zhenguo Li, and Xihui Liu. Editing massive concepts in text-to-image diffusion models. arXiv preprint arXiv:2403.13807, 2024. 1, 2, 3, 4, 6", + "[43] Yijun Yang, Ruiyuan Gao, Xiaosen Wang, Nan Xu, and Qiang Xu. Mma-diffusion: Multimodal attack on diffusion models. arXiv preprint arXiv:2311.17516, 2023. 3", + "[44] Eric Zhang, Kai Wang, Xingqian Xu, Zhangyang Wang, and Humphrey Shi. Forget-me-not: Learning to forget in text-to-image diffusion models. arXiv preprint arXiv:2303.17591, 2023. 1, 2, 6", + "[45] Yihua Zhang, Yimeng Zhang, Yuguang Yao, Jinghan Jia, Jiancheng Liu, Xiaoming Liu, and Sijia Liu. Unlearncanvas: A stylized image dataset to benchmark machine unlearning for diffusion models. arXiv preprint arXiv:2402.11846, 2024. 2, 5" + ], + "bbox": [ + 516, + 92, + 903, + 415 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "28778", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Six-CD_ Benchmarking Concept Removals for Text-to-image Diffusion Models/17b4b8e3-2cbd-49ce-875b-c3c7f736c51d_model.json b/2025/Six-CD_ Benchmarking Concept Removals for Text-to-image Diffusion Models/17b4b8e3-2cbd-49ce-875b-c3c7f736c51d_model.json new file mode 100644 index 0000000000000000000000000000000000000000..e60b09f191bc54da8c97e73775a731c5c360d486 --- /dev/null +++ b/2025/Six-CD_ Benchmarking Concept Removals for Text-to-image Diffusion Models/17b4b8e3-2cbd-49ce-875b-c3c7f736c51d_model.json @@ -0,0 +1,2277 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.104, + 0.131, + 0.895, + 0.153 + ], + "angle": 0, + "content": "Six-CD: Benchmarking Concept Removals for Text-to-image Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.185, + 0.18, + 0.813, + 0.234 + ], + "angle": 0, + "content": "Jie Ren\\(^{1}\\), Kangrui Chen\\(^{1}\\), Yingqian Cui\\(^{1}\\), Shenglai Zeng\\(^{1}\\), Hui Liu\\(^{1}\\), Yue Xing\\(^{1}\\), Jiliang Tang\\(^{1}\\), Lingjuan Lyu\\(^{2}\\) \n\\(^{1}\\)Michigan State University \\(^{2}\\)Sony AI" + }, + { + "type": "text", + "bbox": [ + 0.19, + 0.236, + 0.808, + 0.269 + ], + "angle": 0, + "content": "{renjie3,chenkan4,cuiyingq,zengshel,liuhui7,xingyue1,tangjili}@msu.edu \nlingjuan.lv@sony.com" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.305, + 0.327, + 0.32 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.336, + 0.486, + 0.653 + ], + "angle": 0, + "content": "Text-to-image (T2I) diffusion models have shown exceptional capabilities in generating images that closely correspond to textual prompts. However, the advancement of T2I diffusion models presents significant risks, as the models could be exploited for malicious purposes, such as generating images with violence or nudity, or creating unauthorized portraits of public figures in inappropriate contexts. To mitigate these risks, concept removal methods have been proposed. These methods aim to modify diffusion models to prevent the generation of malicious and unwanted concepts. Despite these efforts, existing research faces several challenges: (1) a lack of consistent comparisons on a comprehensive dataset, (2) ineffective prompts in harmful and nudity concepts, (3) overlooked evaluation of the ability to generate the benign part within prompts containing malicious concepts. To address these gaps, we propose to benchmark the concept removal methods by introducing a new dataset, Six-CD, along with a novel evaluation metric. In this benchmark, we conduct a thorough evaluation of concept removals, with the experimental observations and discussions offering valuable insights in the field." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.68, + 0.222, + 0.695 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.704, + 0.483, + 0.886 + ], + "angle": 0, + "content": "Diffusion models have demonstrated remarkable capabilities in image generation [6, 11, 12, 40, 41], particularly text-to-image (T2I) diffusion models [4, 29, 32]. These models have gained widespread popularity and deployment due to their ability to generate images that precisely align with textual prompts. However, T2I diffusion models can be exploited for malicious purposes, such as creating images depicting violence, nudity, or fake celebrity images in undesirable contexts like jail [14, 20, 26, 30, 33, 38, 39]. To ensure the integrity of the model developers, it is crucial to design benign models that censor malicious concepts and produce only safe and appropriate content." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.886, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Strategies have been proposed to mitigate the malicious" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.306, + 0.908, + 0.593 + ], + "angle": 0, + "content": "generation. For example, in the data collection stage, Stable Diffusion (SD) [32] and Adobe [31] filter out the Not-Safe-For-Work (NSFW) contents in training data. After image generation, SD uses a detector to ensure the images are benign, removing unwanted ones. However, training data filtering requires retraining and will be inefficient if the model has already been trained on various data. For open-sourced models like SD [32], the detector can be easily disabled by modifying the source code. Therefore, concept removal techniques have been proposed to edit a trained model's parameters to prevent generating malicious concepts [7, 18, 19, 34, 44]. For instance, methods are proposed to fine-tune the models to replace unwanted concepts with benign alternatives [7, 18, 44]. Besides, to accelerate fine-tuning, some methods fine-tune the linear components (e.g. the weight matrix of cross attention) using the closed-form optimal solution [8, 22, 42]. Alternatively, there are methods modifying the output during inference to bypass fine-tuning [3, 36]." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.596, + 0.907, + 0.626 + ], + "angle": 0, + "content": "While there is rich literature improving concept removal methods, we observe three potential issues:" + }, + { + "type": "text", + "bbox": [ + 0.522, + 0.629, + 0.906, + 0.704 + ], + "angle": 0, + "content": "1. Lack of consistent and comprehensive comparisons. Current concept removal methods, as shown in Table 1, typically analyze limited categories, lacking consistent and comprehensive comparisons. This gap prevents a thorough understanding of the methods' behavior." + }, + { + "type": "text", + "bbox": [ + 0.521, + 0.705, + 0.909, + 0.902 + ], + "angle": 0, + "content": "2. Ineffective prompts. In existing datasets [36], some categories, such as nudity, contain a large portion of \"ineffective prompts\", which only trigger the model to generate malicious contents with a low probability (see Fig. 1). This can result in inefficient evaluation since the ineffective prompts will generate many benign images, and evaluating concept removal on them is meaningless. Meanwhile, in other categories of unwanted concepts like celebrities, the prompts are more effective than the nudity concept (see Fig. 1 and Fig. 4). When the builder has to remove both nudity and celebrity concepts, the ineffective prompts may cause unfair comparisons. The effective prompts in celebrity can generate" + }, + { + "type": "list", + "bbox": [ + 0.521, + 0.629, + 0.909, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28769" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.314, + 0.09, + 0.685, + 0.104 + ], + "angle": 0, + "content": "Table 1. Categories of unwanted concepts in existing literature" + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.115, + 0.905, + 0.212 + ], + "angle": 0, + "content": "
ESD [7]SPM [23]SDD [16]FMN [44]UCE [8]MACE [22]EMCID [42]SLD [36]SEGA [3]UC [45]CPDM [24]
Harmful✓✓✓
Nudity✓✓✓✓✓✓✓✓
Celebrity✓✓✓
Copyrighted✓✓
Object✓✓✓✓✓✓✓✓
art style✓✓✓✓✓✓✓✓✓✓
" + }, + { + "type": "text", + "bbox": [ + 0.116, + 0.237, + 0.482, + 0.268 + ], + "angle": 0, + "content": "more malicious images, which may falsely induce the builder to put more weight on celebrity concepts." + }, + { + "type": "text", + "bbox": [ + 0.098, + 0.268, + 0.483, + 0.525 + ], + "angle": 0, + "content": "3. Lack of evaluation on in-prompt retainability. Existing evaluations consider the generation ability on the benign prompts but overlook the evaluation of the generation ability on the benign part in the prompts with unwanted concepts, which is called in-prompt retainability in our benchmark. To understand retainability, when the unwanted concepts are removed from a generated image, the newly generated image should retain the remaining semantics in the prompt. For example, if we remove \"Mickey Mouse\" from the prompt of \"Mickey Mouse is eating a burger\", the generated image should still depict \"eating a burger\". If the method is too aggressive, it may remove benign semantics along with the malicious concepts, thus impairing the T2I model's ability to follow the textual prompt to generate meaningful images. Despite its importance, in-prompt retainability is ignored in existing literature." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.528, + 0.484, + 0.588 + ], + "angle": 0, + "content": "To tackle the aforementioned problems, we aim to benchmark concept removal methods by introducing a comprehensive dataset and a new evaluation metric. Our contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.099, + 0.599, + 0.483, + 0.78 + ], + "angle": 0, + "content": "1. A comprehensive dataset. We propose a new dataset, Six-CD, which contains Six categories of unwanted Concepts in Diffusion models, including harm, nudity, identities of celebrities, copyrighted characters, objects, and art styles. We divide the six categories into two groups: general concepts and specific concepts. General concepts, such as harm and nudity, are of concern to all users. Specific concepts are subject to a specific entity, e.g., a person or a company. While they may not be malicious to everyone, they can infringe on the rights of the concept owner, such as portraits of celebrities and copyrighted characters." + }, + { + "type": "text", + "bbox": [ + 0.098, + 0.781, + 0.483, + 0.902 + ], + "angle": 0, + "content": "2. Effective prompts. Through extensive experiments in Sec. 3.2 and Sec. 5.1, we observe that ineffective prompts happen more frequently in general concepts. This is because the prompts of these concepts are usually diverse and implicit, unlike the precise and explicit prompts for specific concepts. Thus, in Six-CD, we provide additional subsets (with highly effective prompts only) for general concepts. Evaluation of ef" + }, + { + "type": "list", + "bbox": [ + 0.098, + 0.599, + 0.483, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.54, + 0.237, + 0.905, + 0.268 + ], + "angle": 0, + "content": "fective prompts can be more efficient and also fairer when compared with specific concepts." + }, + { + "type": "text", + "bbox": [ + 0.521, + 0.268, + 0.907, + 0.522 + ], + "angle": 0, + "content": "3. A new evaluation metric. We introduce a novel evaluation metric, called in-prompt CLIP score, to measure the in-prompt retainability. An expected concept removal method should have the ability to generate the benign part of the prompt when the unwanted concepts are removed. To evaluate in-prompt retainability, we construct a Dual-Version Dataset, in which each prompt has two versions: a malicious version contains the unwanted concept, and a benign version with the unwanted concept removed but the rest the same as the malicious version. We apply concept removal methods to generate images from the malicious version and then use the CLIP score [10] to measure the similarity between these images and the benign prompts. Ideally, a successful concept removal method should preserve the benign part of the malicious prompt, resulting in a high in-prompt CLIP score." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.522, + 0.906, + 0.596 + ], + "angle": 0, + "content": "In the rest of the paper, we first revisit existing concept removal methods and then introduce the proposed dataset and evaluation metric in detail. Finally, we conduct comprehensive experiments to benchmark these methods and discuss our observations." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.608, + 0.652, + 0.623 + ], + "angle": 0, + "content": "2. Preliminaries" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.633, + 0.793, + 0.649 + ], + "angle": 0, + "content": "2.1. Text-to-image Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.655, + 0.906, + 0.761 + ], + "angle": 0, + "content": "Diffusion models typically involve a forward diffusion process and its reverse diffusion process. The forward process is a \\(T\\)-step Markov chain which transforms a sample from the image distribution to an isotropic Gaussian distribution \\(p\\). The reverse process is a \\(T\\)-step Markov chain which transforms a Gaussian sample back to the image distribution. The reverse process can be represented as" + }, + { + "type": "equation", + "bbox": [ + 0.565, + 0.766, + 0.905, + 0.786 + ], + "angle": 0, + "content": "\\[\np _ {\\theta} \\left(x _ {0: T}\\right) = p \\left(x _ {T}\\right) \\prod_ {t = 1} ^ {T} p _ {\\theta} \\left(x _ {t - 1} \\mid x _ {t}\\right) \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.536, + 0.787, + 0.905, + 0.804 + ], + "angle": 0, + "content": "\\[\np _ {\\theta} \\left(x _ {t - 1} \\mid x _ {t}\\right) = \\mathcal {N} \\left(x _ {t - 1}; \\mu_ {\\theta} \\left(x _ {t}, t\\right), \\Sigma_ {\\theta} \\left(x _ {t}, t\\right)\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.906, + 0.902 + ], + "angle": 0, + "content": "where \\( x_0 \\) is the image sample, \\( x_{T} \\) is the Gaussian sample, and \\( x_{t}(0 < t < T) \\) is the middle state of diffusion. The terms \\( \\mu_{\\theta} \\) and \\( \\Sigma_{\\theta} \\) are estimated by a denoising network \\( \\epsilon_{\\theta}(x_t,t) \\) which takes \\( x_{t} \\) and \\( t \\) as input. By \\( \\epsilon_{\\theta} \\), an image sample \\( x_0 \\) can be generated from a Gaussian sample \\( x_{T} \\) following Eq. (1)." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "28770" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.273 + ], + "angle": 0, + "content": "Text-to-image (T2I) diffusion models further extend the diffusion process by guiding it with a textual prompt \\( y \\) [4, 15, 35]. Specifically, \\( \\epsilon_{\\theta}(x_t,t) \\) is modified to take \\( y \\) as an additional input, resulting in \\( \\epsilon_{\\theta}(x_t,y,t) \\). The network \\( \\epsilon_{\\theta}(x_t,y,t) \\) uses a cross-attention module to select information from the textual prompt to guide the diffusion process. In cross attention, each token is transformed into key \\( (K) \\) space and value \\( (V) \\) space, while the image is transformed into query \\( (Q) \\) space. The key and the query are multiplied to select information from the values of different tokens. With cross attention, T2I models can generate images that adhere to the provided text prompts." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.282, + 0.273, + 0.297 + ], + "angle": 0, + "content": "2.2. Concept Removals" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.303, + 0.483, + 0.334 + ], + "angle": 0, + "content": "In this subsection, we revisit representative concept removal methods." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.334, + 0.484, + 0.38 + ], + "angle": 0, + "content": "Fine-tuning-based methods. Fine-tuning is the most common framework for concept removals. The objective for the fine-tuning-based algorithms can be summarized as" + }, + { + "type": "equation", + "bbox": [ + 0.21, + 0.404, + 0.483, + 0.425 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\theta} L = L _ {\\mathrm {r m}} + \\lambda L _ {\\mathrm {r e g}}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.434, + 0.484, + 0.524 + ], + "angle": 0, + "content": "where \\( L_{\\mathrm{rm}} \\) is the term for changing the output of unwanted concept, and \\( L_{\\mathrm{reg}} \\) is the regularization term to ensure that the fine-tuning will not influence other benign concepts and maintain the generation quality. The term \\( \\theta \\) is the parameter to fine-tune. The methods can be solved by two types of solutions:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.526, + 0.483, + 0.602 + ], + "angle": 0, + "content": "- Gradient descent. This solution [7, 16, 18, 23] focuses on the final output of \\(\\epsilon_{\\theta}(x_t,y,t)\\), modifying the model in an end-to-end way. The term \\(L_{\\mathrm{rm}}\\) changes the output of \\(\\epsilon_{\\theta}(x_t,c_u,t)\\) where \\(c_{u}\\) is unwanted concepts, while \\(L_{\\mathrm{reg}}\\) maintains the output of benign concepts \\(c_b\\)." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.602, + 0.483, + 0.678 + ], + "angle": 0, + "content": "- Closed-form solution. Modifying the intermediate states of linear components (such as cross-attention weights [8, 22] and MLP layers [42]) instead of the final output of \\(\\epsilon_{\\theta}(x_t, y_c, t)\\), can be solved in a closed-form solution. This can accelerate the fine-tuning process remarkably." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.526, + 0.483, + 0.678 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.68, + 0.483, + 0.814 + ], + "angle": 0, + "content": "Inference-time methods. Inference-time methods such as negative prompt [32], SLD [36], and SEGA [3] change the generation algorithm to remove concepts in the inference process, which can skip fine-tuning operations. They estimate the influence of unwanted concepts and remove them in inference. However, although these methods can skip the fine-tuning process, there is also a risk of being disabled in open-sourced models, which may slow down the inference." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.827, + 0.483, + 0.861 + ], + "angle": 0, + "content": "3. Six-CD: a Comprehensive and Effective Dataset" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.902 + ], + "angle": 0, + "content": "In this section, we first propose a new dataset that encompasses a comprehensive set of categories of unwanted con" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.138 + ], + "angle": 0, + "content": "cepts. Then we discuss the issue of ineffectiveness in general prompts and present the solution to filter the dataset, thereby increasing its effectiveness." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.148, + 0.84, + 0.166 + ], + "angle": 0, + "content": "3.1. Six Categories of Unwanted Concepts" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.171, + 0.905, + 0.337 + ], + "angle": 0, + "content": "As mentioned in Sec. 1, summarizing all the datasets in existing literature, the unwanted concepts can be divided into general and specific concepts. The general concepts contain harmful concepts and nudity concepts, while the specific concepts contain identities of celebrities, copyrighted characters, objects, and art styles. However, existing literature focuses on only a few categories for evaluation, lacking systematic comparisons among the methods in general and specific concepts. To address this, we propose a new dataset, Six-CD, to evaluate concept removals comprehensively." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.338, + 0.905, + 0.715 + ], + "angle": 0, + "content": "For the general concepts, Six-CD first collects the malicious prompts from four different NSFW resources, which include I2P [36], MMA [43], jatman/stable-diffusion-prompts-stats-full-uncensored (SD-uncensored) [13], and Unsafe Diffusion (UD) [27]. Then, it divides the NSFW data into harmful and nudity concepts. Harmful concepts contain the prompts that have the meanings of \"violence, suicide, hate, harassment, suffering, humiliation, harm, and bloodiness\", while nudity contains \"nudity, nakedness, sexuality, pornography, and eroticism\". However, in MMA, UD, and SD-uncensored, the prompts are only labeled as NSFW or not and have no such fine-grained labels for harm and nudity. Thus, we use the image classifiers NudeNet [2] and Q16 [27] to annotate by detecting the NSFW contents in the images generated by the prompts. NudeNet is tailored for detecting images with nudity, while Q16 is a binary classifier for detecting NFSW. To annotate the fine-grained labels, i.e., harm and nudity, we first use NudeNet to find the prompts of nudity from the collected resources1. Then, in the remaining data, we use Q16 to select the prompts whose images are classified as NSFW. Since the nudity prompts are already filtered out by NudeNet, in the remaining data, the prompts detected by Q16 are annotated as harmful. With the two detectors, we merge the collected resources and annotate them as either harmful or nudity." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.716, + 0.905, + 0.79 + ], + "angle": 0, + "content": "For the specific concepts in Six-CD, instead of directly collecting prompts, we collect concepts and use prompt templates (detailed in Appd. A.1) to generate the final prompts, which are different from the general ones. The concepts are collected as follows:" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.792, + 0.905, + 0.853 + ], + "angle": 0, + "content": "- Celebrity. Celebrities can evaluate the ability to remove identity. We select the identities of celebrities from CPDM [24]. To assist the evaluation, we use a celebrity detector, GCD [25]. we choose the celebrities that can be" + }, + { + "type": "page_footnote", + "bbox": [ + 0.512, + 0.863, + 0.905, + 0.901 + ], + "angle": 0, + "content": "1The image is labeled as nudity if it is classified as \"FEMALE/MALE GENITALIA EXPOSED\", \"FEMALE BREAST EXPOSED\", \"ANUS EXPOSED\", or \"BUTTOCKS EXPOSED\" by NudeNet." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "28771" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.137, + 0.092, + 0.442, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.282, + 0.486, + 0.311 + ], + "angle": 0, + "content": "Figure 1. Effectiveness of I2P prompts. Most of the prompts have low effectiveness. Results from [36]." + }, + { + "type": "text", + "bbox": [ + 0.104, + 0.337, + 0.418, + 0.352 + ], + "angle": 0, + "content": "both generated by SD and recognized by GCD." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.353, + 0.483, + 0.427 + ], + "angle": 0, + "content": "- Copyrighted characters. The copyrighted characters are important IP resources for companies. We use the character concepts in CPDM and the high-frequency copyrighted characters from FiveThirtyEight Comic Characters Dataset [1]." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.428, + 0.483, + 0.473 + ], + "angle": 0, + "content": "- Objects. Objects can be used to evaluate the performance of removal on non-humanoid concepts. We randomly sample from a subset from the classes of ImageNet [5]." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.473, + 0.483, + 0.519 + ], + "angle": 0, + "content": "- Art style. The art styles can evaluate the removal performance of a global feature instead of a local feature. We use the art styles from [22]." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.353, + 0.483, + 0.519 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.52, + 0.486, + 0.746 + ], + "angle": 0, + "content": "For the two general categories, we provide 6,326 prompts for harm concept and 4,518 effective prompts for nudity concept. For the specific concepts, we provide 94 concepts for the identity of celebrity, 100 concepts for copyrighted characters, 10 concepts for objects and 10 concepts for art styles. Four templates are used to construct the prompt for each specific concept. Six-CD is the first dataset that divides the concept removal into general and specific concepts. Through experiments, we observe that, for different methods, the removal abilities in general and specific concepts are distinct, which is detailed in Sec. 5.1. Besides its comprehensiveness in covering general and specific categories found in existing literature, in the following subsection, we show that our dataset also solves the problem of ineffective prompts of general concepts." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.757, + 0.483, + 0.789 + ], + "angle": 0, + "content": "3.2. Ineffective Prompts of General Concepts in Existing Datasets" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.483, + 0.87 + ], + "angle": 0, + "content": "The prompts for general concepts in existing dataset do not consistently generate malicious content for each random seed according to [36] and our observation below. In this subsection, we discuss this phenomenon in general concepts and identify two potential problems it may introduce." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.902 + ], + "angle": 0, + "content": "We first define the effectiveness of a prompt as the possibility of generating malicious images. To assess this, we" + }, + { + "type": "image", + "bbox": [ + 0.557, + 0.092, + 0.865, + 0.268 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.511, + 0.282, + 0.907, + 0.31 + ], + "angle": 0, + "content": "Figure 2. Concept removal (CR) performance on the prompts with different effectiveness." + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.326, + 0.907, + 0.354 + ], + "angle": 0, + "content": "Table 2. Percentage of removed unwanted concepts between high effective and all effectiveness. (%)" + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.365, + 0.905, + 0.426 + ], + "angle": 0, + "content": "
NEGESDSPMSDDFMNUCEEMCIDSLDSEGACorr
Harm40.963.698.924.895.677.698.927.323.70.977
Harm-EF56.767.399.927.897.467.898.638.122.9
Nudity22.317.981.011.293.614.998.314.638.80.984
Nudity-EF11.016.088.94.292.612.397.216.358.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.434, + 0.907, + 0.476 + ], + "angle": 0, + "content": "Table 3. Examples of Dual-Version Dataset for specific concepts. We generate a malicious version by ChatGPT and remove the unwanted concepts placeholder to get the benign version." + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.486, + 0.905, + 0.561 + ], + "angle": 0, + "content": "
CategoryExample of malicious versionExample of benign version
Celeb. & \nCopyrighted{celebrity} is dancing in the rain.dancing in the rain.
Object{object}, football, grass, house, tree, dogfootball, grass, house, tree, dog
Art styleA beautiful snow-covered mountain with \nsunshine lighting it in the style of {art style}A beautiful snow-covered moun-\ntain with sunshine lighting it
" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.582, + 0.907, + 0.794 + ], + "angle": 0, + "content": "generate \\(N\\) images for a prompt using different random seeds. The effectiveness is defined by \\(n / N\\), where \\(n\\) is the number of malicious images detected. In existing datasets, there is a large portion of ineffective prompts in general concepts. Taking I2P [36] as an example, which is the most well-used dataset for harmful and nudity concepts in existing literature [3, 7, 8, 16, 22, 23, 36, 42], most of the prompts have low effectiveness as shown in Fig. 1. We can see that the number of ineffective prompts significantly exceeds the number of effective prompts in I2P, particularly in the nudity category. The specific concepts usually do not have the problem of low effectiveness because they contain the specific and definitive for the unwanted concepts. Thus, in this subsection, we focus on general concepts." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.796, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Limitations. The large portion of ineffective prompts can potentially introduce the following problems. First, evaluating the ineffective prompts is in low efficiency. For example, we test different concept removal methods on SD for nudity concepts in Fig. 2. On the vanilla SD without concept removal (i.e., No CR in Fig. 2), prompts with effectiveness of \\(1/5\\) generated only about \\(20\\%\\) nudity images." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "28772" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.227 + ], + "angle": 0, + "content": "It means that \\(80\\%\\) of the generated images by these prompts are non-nudity images, while evaluating on non-nudity images is meaningless. Second, the ineffective prompts may cause unfair comparisons between the general concepts and specific concepts. Unlike general concepts, the specific concepts do not have the problem of ineffective prompts. When the model builder has to remove both general and specific concepts, the ineffective prompts may falsely induce the builder to put less weight on the general concepts." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.228, + 0.485, + 0.438 + ], + "angle": 0, + "content": "High-effectiveness subsets. Therefore, in addition to harm and nudity concepts, we provide additional subsets of high-effectiveness prompts and name them as harm- \\(EF\\) and nudity- \\(EF\\). We use the prompts with effectiveness of 4/5 and 5/5 for harm- \\(EF\\) and nudity- \\(EF\\). Harm- \\(EF\\) contains 991 prompts, and nudity- \\(EF\\) contains 1,539 prompts. In Table 2 and Fig. 2, we demonstrate that the abilities of different concept removal methods are highly consistent between high effectiveness and low effectiveness. Thus, it is reasonable to evaluate on high-effectiveness prompts in order to reduce computation, which is also more fair. Specifically, in Table 2, we use the percentage of removed unwanted concepts to measure the ability of different methods, which is defined as" + }, + { + "type": "equation", + "bbox": [ + 0.223, + 0.447, + 0.483, + 0.475 + ], + "angle": 0, + "content": "\\[\np _ {\\mathrm {r m}} = 1 - \\frac {n _ {\\text {a f t e r}}}{n _ {\\text {b e f o r e}}}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.485, + 0.483, + 0.59 + ], + "angle": 0, + "content": "where \\( n_{\\mathrm{before}} \\) is the number of images containing unwanted concepts before concept removal, while \\( n_{\\mathrm{after}} \\) is the number of images containing unwanted concepts after concept removal. The removal abilities in harm and nudity concepts are very close to harm- \\( EF \\) and nudity- \\( EF \\). The Pearson correlation coefficient is as high as 0.977 (between harm and harm- \\( EF \\)) and 0.984 (between nudity and nudity- \\( EF \\))." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.59, + 0.483, + 0.665 + ], + "angle": 0, + "content": "In summary, Six-CD provides two choices to evaluate on the general concepts. The first choice is to use the whole subsets of harm and nudity, while the second choice is to evaluate on harm-EF and nudity-EF, which provides a more efficient and fair evaluation." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.68, + 0.317, + 0.697 + ], + "angle": 0, + "content": "4. In-prompt Retainability" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.483, + 0.78 + ], + "angle": 0, + "content": "Concept removals should not hurt the generation ability of the benign contents, which is referred to as retainability. In this section, we show the lack of evaluation for retainability on the benign parts of the prompts containing the unwanted concepts and propose a new metric for this retainability." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Lack of in-prompt retainability. Existing literature [24, 45] only considers the retainability of totally benign concepts. The benign prompts do not contain any unwanted concept, and concept removal should not influence the generation of them. Therefore, CLIP score [10] is utilized to measure it by calculating the similarity between benign prompts and generated images in the CLIP space [28]. However, for the prompts containing unwanted concepts," + }, + { + "type": "image_caption", + "bbox": [ + 0.626, + 0.09, + 0.673, + 0.101 + ], + "angle": 0, + "content": "NO CR" + }, + { + "type": "image_caption", + "bbox": [ + 0.709, + 0.09, + 0.834, + 0.102 + ], + "angle": 0, + "content": "ironman removed" + }, + { + "type": "image", + "bbox": [ + 0.589, + 0.103, + 0.709, + 0.196 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.712, + 0.103, + 0.832, + 0.196 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.207, + 0.905, + 0.249 + ], + "angle": 0, + "content": "Figure 3. Concept removal example of the prompt: \"At night, ironman with his armor on is watching fireworks by the lake\". Ironman is removed by SDD." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.267, + 0.907, + 0.463 + ], + "angle": 0, + "content": "the retainability to guarantee the generation of the benign part of them is also important, which is called \"in-prompt retainability\" in our benchmark. As shown in Fig. 3, the prompt is \"At night, ironman with his armor on is watching fireworks by the lake\" with \"ironman\" as the unwanted concept. Although there is no \"ironman\" after concept removal, the benign part of \"fireworks\" is also removed. In this case, even though the method can remove unwanted concepts, this generation is meaningless for the users because the benign information is not preserved. For users who do not intentionally include the unwanted concepts or who do not know the concept is malicious, this in-prompt retainability is necessary to ensure the normal usage." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.464, + 0.906, + 0.689 + ], + "angle": 0, + "content": "New metric. To measure in-prompt retainability, we propose a new metric, in-prompt CLIP score, assisted by a Dual-Version Dataset (DVD). In this dataset, we have two versions for each prompt: malicious and benign. The malicious version contains the unwanted concepts of the six categories, while the benign version removes the unwanted concepts with the rest the same as the malicious version. In in-prompt CLIP score, we first generate the images with concept removals using the malicious version and then calculate the cosine similarity of CLIP embeddings between these generated images and the corresponding benign version prompts. Thus, we can measure the similarity between the generated image and the benign part of the prompt. To distinguish it from the original CLIP score, we refer to the original CLIP score as the out-prompt CLIP score." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.69, + 0.906, + 0.871 + ], + "angle": 0, + "content": "To construct DVD, for general concepts, we first sample two subsets of prompts from the harmful and nudity categories of Six-CD as the malicious version, then we manually remove the unwanted concepts to get the benign version. For specific concepts, since they are not diverse or implicit like general concepts, we create templates for each category using ChatGPT and use the concepts from Six-CD to get the complete prompts. We show the examples of templates in DVD in Table 3. The entire dataset is in Appd. A.2. To validate the effectiveness of the metric, we conduct human evaluation in Appd. D.1. The results demonstrate that the metric is highly consistent with human preference." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.905, + 0.902 + ], + "angle": 0, + "content": "In summary, to measure the in-prompt retainability, we propose the in-prompt CLIP score. It leverages a malicious-" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28773" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.138 + ], + "angle": 0, + "content": "version and a benign-version prompt to measure the similarity between the images after concept removal and the benign part of the prompt." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.15, + 0.226, + 0.168 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.175, + 0.483, + 0.297 + ], + "angle": 0, + "content": "In this section, we conduct experiments to benchmark the performance of 10 methods in removing single and multiple concepts and the retainability with both in-prompt and out-prompt CLIP scores. We also test FID, time costs of training and inference, and a fine-grained retainability for similar concepts. Due to space limitations, these experiments are in Appd D. Also, we report the results of SD v1.4 in this section and the results of other models are in Appd D." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.297, + 0.483, + 0.432 + ], + "angle": 0, + "content": "Experimental settings. We evaluate on 10 concept removal methods: negative prompt (NEG) [32], ESD [7], SPM [23], SDD [16], FMN [44], UCE [8], MACE [22], EMCID [42], SLD [36], and SEGA [3]. The details on the baseline settings can be found in Appd. B. The evaluated datasets are Six-CD and DVD, which are detailed in Appd. A. Following the conclusions in Sec. 3.2, we use harm-EF and nudity-EF for all the experiments in this section. All experiments use one GPU (A5000 or A6000)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.433, + 0.483, + 0.629 + ], + "angle": 0, + "content": "Detection metrics. We use the detection rate of Q16 and NudeNet for harmful and nudity concepts. We use the classification accuracy of GCD as the detection rate for celebrity concepts, and we train two classifiers based on ResNet-50 [9] and use the accuracy as the detection rate for copyrighted characters and objects. For the art style concept, the detection is not a simple binary problem. In existing literature [22-24, 44], CLIP score, \\( S_{\\mathrm{CLIP}} \\), is used to measure the presence of an art style on a continuous scale, assessing the similarity between generated images and art styles. Following them, we use \\( S_{\\mathrm{CLIP}} \\) and normalize its range by \\( \\tilde{S}_{\\mathrm{CLIP}} = (S_{\\mathrm{CLIP}} - \\min(S_{\\mathrm{CLIP}})) / (\\max(S_{\\mathrm{CLIP}}) - \\min(S_{\\mathrm{CLIP}})) \\) for a better comparison with other categories." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.638, + 0.365, + 0.654 + ], + "angle": 0, + "content": "5.1. Evaluation on Removal Ability" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.66, + 0.483, + 0.824 + ], + "angle": 0, + "content": "In this subsection, we test the removal ability of different methods on Six-CD. In Fig. 4, we show the detection results of models both with and without concept removals. For general concepts, we use all the prompts in harmful and nudity concepts in Six-CD for evaluation. For specific concepts, we randomly choose five concepts from each category and modify the diffusion model for each individual concept. We show the averaged results of 5 concepts of each category in Fig. 4. (The table with error bar for Fig. 4 can be found in Appd. D.2.) We highlight the main observations as follows." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Observation 1: General concepts are harder to remove. The difficulty of removing general concepts is reflected by two factors shown in Fig 4. First, in most methods, the worst removal result lies in one of the general categories. For example, ESD and MACE perform well in all" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.333 + ], + "angle": 0, + "content": "other categories, but they still have around \\(55\\%\\) harmful images detected. Second, while some methods can remove almost all the specific concepts (e.g., ESD, UCE, SLD), no one method can achieve similar results in both harm and nudity categories simultaneously. For general concepts, on the one hand, they are hard to trigger (i.e. low effectiveness). On the other hand, it is harder to remove them than specific concepts. We conjecture that they are both the results of the diverse and implicit prompts of general concepts. Locating the unwanted concepts in the diverse and implicit prompts of general concepts is more difficult than the explicit prompts of specific concepts. Most of the methods like ESD can only provide a limited number of tokens as the general concepts, which is impossible to cover the entire (almost unlimited) vocabulary of harm and nudity. Thus, it leads to difficulty in removing general concepts." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.335, + 0.905, + 0.562 + ], + "angle": 0, + "content": "Observation 2: Removal ability is consistent within general concepts and within specific categories. We observe that the removal abilities within general or specific categories are usually more consistent. For example, SLD and FMN perform well in specific concepts, but the ability to remove general concepts is worse. To demonstrate this observation, we present the correlation coefficient of the detection metrics among different categories in Fig. 5. From Fig. 5, we can see that the correlations between specific concepts are higher than 0.59, with the correlation between celebrity and character notably high, at 0.94. The correlation between general concepts is also as high as 0.65. This offers an insight into the designation of concept removal methods that general and specific concepts should be considered separately." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.575, + 0.844, + 0.59 + ], + "angle": 0, + "content": "5.2. Removal Ability on Multiple Concepts" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.598, + 0.905, + 0.734 + ], + "angle": 0, + "content": "In Sec. 5.1, we evaluate the ability to remove a single concept from one model. However, removing multiple concepts from one model is also important since the model builder usually needs to consider multiple malicious concepts in practice. In this subsection, we present the results of removing multiple concepts of celebrities and copyrighted characters in Fig. 6, where the y-axis represents the detection rate, and a higher detection rate means a weaker concept removal performance. We have two observations from Fig. 6." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Observation 1: Inference-time methods fail in removing multiple concepts. As shown in Fig. 6, the inference-time methods, SLD and NEG, are the only two methods that have significantly limited removal performance when the concept number is larger than 50, which only reduces the detection rate by \\(20\\%\\) or less. To remove the multiple concepts in inference time, they have to encode the string containing all the concepts in the embeddings of one single prompt. This will exceed the capacity of the text encoder of T2I diffusion models and lead to failed removal. Another inference-time method, SEGA, is not reported in Fig. 6 due" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "28774" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.101, + 0.08, + 0.898, + 0.371 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.38, + 0.908, + 0.424 + ], + "angle": 0, + "content": "Figure 4. Removal ability on Six-CD. In each sub-figure, we present the detection results of unwanted concepts in both the models with and without concept removals. The range of all the six values is [0,1]. Smaller values indicate that less unwanted concepts are detected, i.e. better removal. G1 and G2 are harm-EF and nudity-EF concepts, while S1, S2, S3 and S4 are celebrity, character, object and art style." + }, + { + "type": "image", + "bbox": [ + 0.137, + 0.44, + 0.428, + 0.626 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.642, + 0.483, + 0.671 + ], + "angle": 0, + "content": "Figure 5. Correlation coefficient of removal performance on six data categories." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.688, + 0.483, + 0.734 + ], + "angle": 0, + "content": "to its catastrophic inference time when the concept number is increased to 50. This implies that the strategy of processing each concept separately during inference is impractical." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.734, + 0.483, + 0.84 + ], + "angle": 0, + "content": "Observation 2: Closed-form solutions by modifying linear components perform well in removing multiple concepts. The best two methods in Fig. 6a and Fig. 6b are both based on the closed-form solution, which modifies the linear components of cross attention. This is possibly because combining multiple concepts in the linear components is easier than in other non-linear parts." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.849, + 0.443, + 0.866 + ], + "angle": 0, + "content": "5.3. In-prompt and Out-prompt Retainability" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.902 + ], + "angle": 0, + "content": "We evaluate the generation ability on benign concepts using both in-prompt and out-prompt CLIP scores. In Fig. 7, we" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.449, + 0.907, + 0.494 + ], + "angle": 0, + "content": "use DVD to test the in-prompt CLIP score and a subset with benign prompts from LAION [37] to test out-prompt CLIP score. The observations as follows." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.501, + 0.908, + 0.759 + ], + "angle": 0, + "content": "Observation 1: In-prompt retainability performs worse than out-prompt retainability. As shown in Fig. 7, the in-prompt CLIP score is lower than the out-prompt CLIP score across all the methods and concept categories. This indicates that the ability to generate benign parts in prompts containing malicious concepts is more negatively affected than the ability to generate totally benign prompts. As mentioned in Sec. 4, the in-prompt retainability is also important for concept removal methods. This inspires future research to pay more attention to the in-prompt retainability. Remark. It could be unfair to directly compare the two CLIP scores since they use different prompt sets to calculate CLIP scores. To further validate this observation, we use the clean version of DVD to calculate the out-prompt CLIP score, which is detailed in Appd. D.3. By comparing it with the in-prompt CLIP score in Appd. D.3, we have observed a consistent phenomenon with this subsection." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Observation 2: Auxiliary semantic information in MACE (Fig. 7g) may hurt the generation on benign concepts. MACE has the most significant decrease in the generation ability of benign concepts for both in-prompt and out-prompt CLIP scores. We conjecture that this is because it introduces auxiliary semantic information to help locate the unwanted concepts in the image. In other methods, the unwanted concepts are usually located by the semantic understanding ability of the T2I model itself. In contrast," + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "28775" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.147, + 0.092, + 0.443, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.26, + 0.267, + 0.33, + 0.279 + ], + "angle": 0, + "content": "(a) Celebrity" + }, + { + "type": "image", + "bbox": [ + 0.465, + 0.092, + 0.85, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.543, + 0.267, + 0.685, + 0.279 + ], + "angle": 0, + "content": "(b) Copyrighted characters" + }, + { + "type": "image_caption", + "bbox": [ + 0.357, + 0.287, + 0.642, + 0.302 + ], + "angle": 0, + "content": "Figure 6. Removal ability on multiple concepts" + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.314, + 0.249, + 0.459 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.149, + 0.461, + 0.195, + 0.471 + ], + "angle": 0, + "content": "(a) NEG" + }, + { + "type": "image", + "bbox": [ + 0.257, + 0.314, + 0.413, + 0.459 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.313, + 0.461, + 0.358, + 0.471 + ], + "angle": 0, + "content": "(b) ESD" + }, + { + "type": "image", + "bbox": [ + 0.422, + 0.315, + 0.576, + 0.459 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.476, + 0.461, + 0.522, + 0.471 + ], + "angle": 0, + "content": "(c) SPM" + }, + { + "type": "image", + "bbox": [ + 0.585, + 0.313, + 0.74, + 0.459 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.64, + 0.461, + 0.686, + 0.471 + ], + "angle": 0, + "content": "(d) SDD" + }, + { + "type": "image", + "bbox": [ + 0.749, + 0.334, + 0.904, + 0.459 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.803, + 0.461, + 0.851, + 0.471 + ], + "angle": 0, + "content": "(e) FMN" + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.474, + 0.248, + 0.597 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.15, + 0.6, + 0.193, + 0.611 + ], + "angle": 0, + "content": "(f) UCE" + }, + { + "type": "image", + "bbox": [ + 0.257, + 0.474, + 0.413, + 0.598 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.307, + 0.6, + 0.364, + 0.612 + ], + "angle": 0, + "content": "(g) MACE" + }, + { + "type": "image", + "bbox": [ + 0.422, + 0.474, + 0.576, + 0.598 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.468, + 0.6, + 0.531, + 0.611 + ], + "angle": 0, + "content": "(h) EMCID" + }, + { + "type": "image", + "bbox": [ + 0.585, + 0.474, + 0.74, + 0.598 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.642, + 0.6, + 0.684, + 0.611 + ], + "angle": 0, + "content": "(i) SLD" + }, + { + "type": "image", + "bbox": [ + 0.749, + 0.474, + 0.904, + 0.598 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.801, + 0.6, + 0.853, + 0.611 + ], + "angle": 0, + "content": "(j) SEGA" + }, + { + "type": "image_caption", + "bbox": [ + 0.316, + 0.624, + 0.68, + 0.639 + ], + "angle": 0, + "content": "Figure 7. In-prompt retainability vs. out-prompt retainability" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.655, + 0.483, + 0.777 + ], + "angle": 0, + "content": "MACE incorporates the segmentation results of GroundedSAM [17, 21] to locate the image area of unwanted concepts. It focuses on optimizing this area and overlooks other areas of the image. This may lead to worse retainability in other areas. Another piece of evidence is that when we do not include segmentation in removing the art style concepts by MACE (since art style is a global feature), it has good in-prompt and out-prompt CLIP scores." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.798, + 0.401, + 0.813 + ], + "angle": 0, + "content": "6. Conclusion and Ethical Statement" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.484, + 0.901 + ], + "angle": 0, + "content": "In this work, we address the lack of a systematic benchmark for concept removal methods by introducing a comprehensive and effective dataset and a new evaluation metric. Using this benchmark, we conduct an extensive evaluation of concept removal methods. Our experimental observations" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.656, + 0.907, + 0.746 + ], + "angle": 0, + "content": "provide valuable and practical insights for future research in this field. While our benchmark may have a limitation in the inconvenient detection metric for art styles (using the CLIP score for art styles, instead of the detection rate from a classifier like other categories), it still offers a thorough and practical evaluation across the various method settings." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Ethical statement. This work provides a comprehensive evaluation framework for concept removal, which is targeted to mitigate the potential malicious use of text-to-image diffusion models. Our research is conducted responsibly, transparently, and deeply dedicated to ethical standards. Despite involving the generation of sensitive content such as nudity and violence, it is strictly for research purposes and does not intend to produce or promote inappropriate material. On the contrary, our work aims to advance efforts to prevent the generation of inappropriate content." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "28776" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.251, + 0.108 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.114, + 0.484, + 0.226 + ], + "angle": 0, + "content": "Jie Ren, Kangrui Chen, Yingqian Cui, Shenglai Zeng, and Jiliang Tang are supported by the National Science Foundation (NSF) under grant numbers CNS2321416, IIS2212032, IIS2212144, IOS2107215, DUE2234015, CNS2246050, DRL2405483 and IOS2035472, the Army Research Office (ARO) under grant number W911NF-21-1-0198, Amazon Faculty Award, JP Morgan Faculty Award, Meta, Microsoft and SNAP." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.238, + 0.188, + 0.253 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.263, + 0.484, + 0.317 + ], + "angle": 0, + "content": "[1] Fivethirtyeight comic characters dataset. https://www.kaggle.com/datasets/fivethirtyeight/fivethirtyeight - comic - characters - dataset.4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.319, + 0.484, + 0.345 + ], + "angle": 0, + "content": "[2] Nudenet: Lightweight nudity detection. https://github.com/notAI-tech/NudeNet, 2023.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.347, + 0.484, + 0.415 + ], + "angle": 0, + "content": "[3] Manuel Brack, Felix Friedrich, Dominik Hintersdorf, Lukas Struppek, Patrick Schramowski, and Kristian Kersting. Sega: Instructing text-to-image models using semantic guidance. Advances in Neural Information Processing Systems, 36, 2024. 1, 2, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.417, + 0.484, + 0.485 + ], + "angle": 0, + "content": "[4] Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, et al. Pixart-α: Fast training of diffusion transformer for photorealistic text-to-image synthesis. arXiv preprint arXiv:2310.00426, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.486, + 0.484, + 0.54 + ], + "angle": 0, + "content": "[5] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.541, + 0.484, + 0.582 + ], + "angle": 0, + "content": "[6] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.583, + 0.484, + 0.65 + ], + "angle": 0, + "content": "[7] Rohit Gandikota, Joanna Materzynska, Jaden Fiitto-Kaufman, and David Bau. Erasing concepts from diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2426-2436, 2023. 1, 2, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.652, + 0.484, + 0.72 + ], + "angle": 0, + "content": "[8] Rohit Gandikota, Hadas Orgad, Yonatan Belinkov, Joanna Materzyńska, and David Bau. Unified concept editing in diffusion models. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 5111-5120, 2024. 1, 2, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.721, + 0.484, + 0.776 + ], + "angle": 0, + "content": "[9] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.777, + 0.484, + 0.83 + ], + "angle": 0, + "content": "[10] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning. arXiv preprint arXiv:2104.08718, 2021. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.832, + 0.484, + 0.859 + ], + "angle": 0, + "content": "[11] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.86, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[12] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840-6851, 2020. 1" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.263, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.16 + ], + "angle": 0, + "content": "[13] Huggingface. jatman/stable-diffusion-prompts-stats-full-uncensored datasets at hugging face. https://huggingface.co/datasets/jatman/stable-diffusion-prompts-stats-full-uncensored?not-for-all-audiences=true. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.163, + 0.905, + 0.231 + ], + "angle": 0, + "content": "[14] Tatum Hunter. Ai porn is easy to make now. for women, that's a nightmare. https://www.washingtonpost.com/technology/2023/02/13/ai-porn-deepfakes-women-consent/, 2023. Accessed: 2024-06-01. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.232, + 0.845, + 0.246 + ], + "angle": 0, + "content": "[15] DeepFloyd IF. https://github.com/deep-floyd/if.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.247, + 0.905, + 0.302 + ], + "angle": 0, + "content": "[16] Sanghyun Kim, Seohyeon Jung, Balhae Kim, Moonseok Choi, Jinwoo Shin, and Juho Lee. Towards safe self-distillation of internet-scale text-to-image diffusion models. arXiv preprint arXiv:2307.05977, 2023. 2, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.303, + 0.905, + 0.371 + ], + "angle": 0, + "content": "[17] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.372, + 0.908, + 0.44 + ], + "angle": 0, + "content": "[18] Nupur Kumari, Bingliang Zhang, Sheng-Yu Wang, Eli Shechtman, Richard Zhang, and Jun-Yan Zhu. Ablating concepts in text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22691-22702, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.442, + 0.905, + 0.496 + ], + "angle": 0, + "content": "[19] Yuantong Li, Chi-Hua Wang, and Guang Cheng. Online forgetting process for linear regression models. In International Conference on Artificial Intelligence and Statistics, pages 217-225. PMLR, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.498, + 0.905, + 0.552 + ], + "angle": 0, + "content": "[20] Yaxin Li, Jie Ren, Han Xu, and Hui Liu. Neural style protection: Counteracting unauthorized neural style transfer. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 3966-3975, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.554, + 0.905, + 0.621 + ], + "angle": 0, + "content": "[21] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023.8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.623, + 0.905, + 0.676 + ], + "angle": 0, + "content": "[22] Shilin Lu, Zilan Wang, Leyang Li, Yanzhu Liu, and Adams Wai-Kin Kong. Mace: Mass concept erasure in diffusion models. arXiv preprint arXiv:2403.06135, 2024. 1, 2, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.679, + 0.905, + 0.747 + ], + "angle": 0, + "content": "[23] Mengyao Lyu, Yuhong Yang, Haiwen Hong, Hui Chen, Xuan Jin, Yuan He, Hui Xue, Jungong Han, and Guiguang Ding. One-dimensional adapter to rule them all: Concepts, diffusion models and erasing applications. arXiv preprint arXiv:2312.16145, 2023. 2, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.749, + 0.905, + 0.816 + ], + "angle": 0, + "content": "[24] Rui Ma, Qiang Zhou, Bangjun Xiao, Yizhu Jin, Daquan Zhou, Xiuyu Li, Aishani Singh, Yi Qu, Kurt Keutzer, Xiaodong Xie, et al. A dataset and benchmark for copyright protection from text-to-image diffusion models. arXiv preprint arXiv:2403.12052, 2024. 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.818, + 0.905, + 0.872 + ], + "angle": 0, + "content": "[25] Dmitry Voitekh Nick Hasty, Ihor Kroosh and Dmytro Korduban. Giphy's open source celebrity detection deep learning model and code. https://github.com/Giphy/celeb-detection-oss, 2019.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[26] OpenAI. Dall-e 2 preview - risks and limitations. https://github.com/openai/dalle-2-preview/" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28777" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.119 + ], + "angle": 0, + "content": "blob/main/system-card.md, 2023. Accessed: 2024-06-01. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.121, + 0.483, + 0.189 + ], + "angle": 0, + "content": "[27] Yiting Qu, Xinyue Shen, Xinlei He, Michael Backes, Savvas Zannettou, and Yang Zhang. Unsafe Diffusion: On the Generation of Unsafe Images and Hateful Memes From Text-To-Image Models. In ACM SIGSAC Conference on Computer and Communications Security (CCS). ACM, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.19, + 0.482, + 0.272 + ], + "angle": 0, + "content": "[28] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.274, + 0.482, + 0.328 + ], + "angle": 0, + "content": "[29] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.33, + 0.482, + 0.371 + ], + "angle": 0, + "content": "[30] Javier Rando, Daniel Paleka, David Lindner, Lennart Heim, and Florian Tramèr. Red-teaming the stable diffusion safety filter. arXiv preprint arXiv:2210.04610, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.372, + 0.482, + 0.399 + ], + "angle": 0, + "content": "[31] Dana Rao. Responsible innovation in the age of generative ai: Adobe blog, . 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.401, + 0.41, + 0.413 + ], + "angle": 0, + "content": "[32] Dana Rao. Stable diffusion 2.0 release, .1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.414, + 0.482, + 0.481 + ], + "angle": 0, + "content": "[33] Jie Ren, Han Xu, Pengfei He, Yingqian Cui, Shenglai Zeng, Jiankun Zhang, Hongzhi Wen, Jiayuan Ding, Hui Liu, Yi Chang, et al. Copyright protection in generative ai: A technical perspective. arXiv preprint arXiv:2402.02333, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.484, + 0.482, + 0.539 + ], + "angle": 0, + "content": "[34] Jie Ren, Zhenwei Dai, Xianfeng Tang, Hui Liu, Jingying Zeng, Zhen Li, Rahul Goutam, Suhang Wang, Yue Xing, and Qi He. A general framework to enhance fine-tuning-based llm unlearning. arXiv preprint arXiv:2502.17823, 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.54, + 0.483, + 0.609 + ], + "angle": 0, + "content": "[35] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.61, + 0.483, + 0.678 + ], + "angle": 0, + "content": "[36] Patrick Schramowski, Manuel Brack, Björn Deiseroth, and Kristian Kersting. Safe latent diffusion: Mitigating inappropriate degeneration in diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22522-22531, 2023. 1, 2, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.679, + 0.482, + 0.762 + ], + "angle": 0, + "content": "[37] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278–25294, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.763, + 0.482, + 0.83 + ], + "angle": 0, + "content": "[38] Shawn Shan, Jenna Cryan, Emily Wenger, Haitao Zheng, Rana Hanocka, and Ben Y Zhao. Glaze: Protecting artists from style mimicry by {Text-to-Image} models. In 32nd USENIX Security Symposium (USENIX Security 23), pages 2187–2204, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[39] Gowthami Somepalli, Vasu Singla, Micah Goldblum, Jonas Geiping, and Tom Goldstein. Diffusion art or digital forgery? investigating data replication in diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6048-6058, 2023. 1" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.133 + ], + "angle": 0, + "content": "[40] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[41] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.192, + 0.905, + 0.245 + ], + "angle": 0, + "content": "[42] Tianwei Xiong, Yue Wu, Enze Xie, Zhenguo Li, and Xihui Liu. Editing massive concepts in text-to-image diffusion models. arXiv preprint arXiv:2403.13807, 2024. 1, 2, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.249, + 0.905, + 0.29 + ], + "angle": 0, + "content": "[43] Yijun Yang, Ruiyuan Gao, Xiaosen Wang, Nan Xu, and Qiang Xu. Mma-diffusion: Multimodal attack on diffusion models. arXiv preprint arXiv:2311.17516, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.292, + 0.905, + 0.346 + ], + "angle": 0, + "content": "[44] Eric Zhang, Kai Wang, Xingqian Xu, Zhangyang Wang, and Humphrey Shi. Forget-me-not: Learning to forget in text-to-image diffusion models. arXiv preprint arXiv:2303.17591, 2023. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.348, + 0.905, + 0.416 + ], + "angle": 0, + "content": "[45] Yihua Zhang, Yimeng Zhang, Yuguang Yao, Jinghan Jia, Jiancheng Liu, Xiaoming Liu, and Sijia Liu. Unlearncanvas: A stylized image dataset to benchmark machine unlearning for diffusion models. arXiv preprint arXiv:2402.11846, 2024. 2, 5" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.416 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "28778" + } + ] +] \ No newline at end of file diff --git a/2025/Six-CD_ Benchmarking Concept Removals for Text-to-image Diffusion Models/17b4b8e3-2cbd-49ce-875b-c3c7f736c51d_origin.pdf b/2025/Six-CD_ Benchmarking Concept Removals for Text-to-image Diffusion Models/17b4b8e3-2cbd-49ce-875b-c3c7f736c51d_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..3ae5ec6b4de6f35076a5de1d48836c3149ca7065 --- /dev/null +++ b/2025/Six-CD_ Benchmarking Concept Removals for Text-to-image Diffusion Models/17b4b8e3-2cbd-49ce-875b-c3c7f736c51d_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:444a808da13d03b3f8a3f7325739d63ca99067f7a0f4284bfa7b2414d0f596d1 +size 6111665 diff --git a/2025/Six-CD_ Benchmarking Concept Removals for Text-to-image Diffusion Models/full.md b/2025/Six-CD_ Benchmarking Concept Removals for Text-to-image Diffusion Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..6074b4eb6714335e335a9b2febc8af8f736a6833 --- /dev/null +++ b/2025/Six-CD_ Benchmarking Concept Removals for Text-to-image Diffusion Models/full.md @@ -0,0 +1,311 @@ +# Six-CD: Benchmarking Concept Removals for Text-to-image Diffusion Models + +Jie Ren $^{1}$ , Kangrui Chen $^{1}$ , Yingqian Cui $^{1}$ , Shenglai Zeng $^{1}$ , Hui Liu $^{1}$ , Yue Xing $^{1}$ , Jiliang Tang $^{1}$ , Lingjuan Lyu $^{2}$ $^{1}$ Michigan State University $^{2}$ Sony AI + +{renjie3,chenkan4,cuiyingq,zengshel,liuhui7,xingyue1,tangjili}@msu.edu +lingjuan.lv@sony.com + +# Abstract + +Text-to-image (T2I) diffusion models have shown exceptional capabilities in generating images that closely correspond to textual prompts. However, the advancement of T2I diffusion models presents significant risks, as the models could be exploited for malicious purposes, such as generating images with violence or nudity, or creating unauthorized portraits of public figures in inappropriate contexts. To mitigate these risks, concept removal methods have been proposed. These methods aim to modify diffusion models to prevent the generation of malicious and unwanted concepts. Despite these efforts, existing research faces several challenges: (1) a lack of consistent comparisons on a comprehensive dataset, (2) ineffective prompts in harmful and nudity concepts, (3) overlooked evaluation of the ability to generate the benign part within prompts containing malicious concepts. To address these gaps, we propose to benchmark the concept removal methods by introducing a new dataset, Six-CD, along with a novel evaluation metric. In this benchmark, we conduct a thorough evaluation of concept removals, with the experimental observations and discussions offering valuable insights in the field. + +# 1. Introduction + +Diffusion models have demonstrated remarkable capabilities in image generation [6, 11, 12, 40, 41], particularly text-to-image (T2I) diffusion models [4, 29, 32]. These models have gained widespread popularity and deployment due to their ability to generate images that precisely align with textual prompts. However, T2I diffusion models can be exploited for malicious purposes, such as creating images depicting violence, nudity, or fake celebrity images in undesirable contexts like jail [14, 20, 26, 30, 33, 38, 39]. To ensure the integrity of the model developers, it is crucial to design benign models that censor malicious concepts and produce only safe and appropriate content. + +Strategies have been proposed to mitigate the malicious + +generation. For example, in the data collection stage, Stable Diffusion (SD) [32] and Adobe [31] filter out the Not-Safe-For-Work (NSFW) contents in training data. After image generation, SD uses a detector to ensure the images are benign, removing unwanted ones. However, training data filtering requires retraining and will be inefficient if the model has already been trained on various data. For open-sourced models like SD [32], the detector can be easily disabled by modifying the source code. Therefore, concept removal techniques have been proposed to edit a trained model's parameters to prevent generating malicious concepts [7, 18, 19, 34, 44]. For instance, methods are proposed to fine-tune the models to replace unwanted concepts with benign alternatives [7, 18, 44]. Besides, to accelerate fine-tuning, some methods fine-tune the linear components (e.g. the weight matrix of cross attention) using the closed-form optimal solution [8, 22, 42]. Alternatively, there are methods modifying the output during inference to bypass fine-tuning [3, 36]. + +While there is rich literature improving concept removal methods, we observe three potential issues: + +1. Lack of consistent and comprehensive comparisons. Current concept removal methods, as shown in Table 1, typically analyze limited categories, lacking consistent and comprehensive comparisons. This gap prevents a thorough understanding of the methods' behavior. +2. Ineffective prompts. In existing datasets [36], some categories, such as nudity, contain a large portion of "ineffective prompts", which only trigger the model to generate malicious contents with a low probability (see Fig. 1). This can result in inefficient evaluation since the ineffective prompts will generate many benign images, and evaluating concept removal on them is meaningless. Meanwhile, in other categories of unwanted concepts like celebrities, the prompts are more effective than the nudity concept (see Fig. 1 and Fig. 4). When the builder has to remove both nudity and celebrity concepts, the ineffective prompts may cause unfair comparisons. The effective prompts in celebrity can generate + +Table 1. Categories of unwanted concepts in existing literature + +
ESD [7]SPM [23]SDD [16]FMN [44]UCE [8]MACE [22]EMCID [42]SLD [36]SEGA [3]UC [45]CPDM [24]
Harmful✓✓✓
Nudity✓✓✓✓✓✓✓✓
Celebrity✓✓✓
Copyrighted✓✓
Object✓✓✓✓✓✓✓✓
art style✓✓✓✓✓✓✓✓✓✓
+ +more malicious images, which may falsely induce the builder to put more weight on celebrity concepts. + +3. Lack of evaluation on in-prompt retainability. Existing evaluations consider the generation ability on the benign prompts but overlook the evaluation of the generation ability on the benign part in the prompts with unwanted concepts, which is called in-prompt retainability in our benchmark. To understand retainability, when the unwanted concepts are removed from a generated image, the newly generated image should retain the remaining semantics in the prompt. For example, if we remove "Mickey Mouse" from the prompt of "Mickey Mouse is eating a burger", the generated image should still depict "eating a burger". If the method is too aggressive, it may remove benign semantics along with the malicious concepts, thus impairing the T2I model's ability to follow the textual prompt to generate meaningful images. Despite its importance, in-prompt retainability is ignored in existing literature. + +To tackle the aforementioned problems, we aim to benchmark concept removal methods by introducing a comprehensive dataset and a new evaluation metric. Our contributions are as follows: + +1. A comprehensive dataset. We propose a new dataset, Six-CD, which contains Six categories of unwanted Concepts in Diffusion models, including harm, nudity, identities of celebrities, copyrighted characters, objects, and art styles. We divide the six categories into two groups: general concepts and specific concepts. General concepts, such as harm and nudity, are of concern to all users. Specific concepts are subject to a specific entity, e.g., a person or a company. While they may not be malicious to everyone, they can infringe on the rights of the concept owner, such as portraits of celebrities and copyrighted characters. +2. Effective prompts. Through extensive experiments in Sec. 3.2 and Sec. 5.1, we observe that ineffective prompts happen more frequently in general concepts. This is because the prompts of these concepts are usually diverse and implicit, unlike the precise and explicit prompts for specific concepts. Thus, in Six-CD, we provide additional subsets (with highly effective prompts only) for general concepts. Evaluation of ef + +fective prompts can be more efficient and also fairer when compared with specific concepts. + +3. A new evaluation metric. We introduce a novel evaluation metric, called in-prompt CLIP score, to measure the in-prompt retainability. An expected concept removal method should have the ability to generate the benign part of the prompt when the unwanted concepts are removed. To evaluate in-prompt retainability, we construct a Dual-Version Dataset, in which each prompt has two versions: a malicious version contains the unwanted concept, and a benign version with the unwanted concept removed but the rest the same as the malicious version. We apply concept removal methods to generate images from the malicious version and then use the CLIP score [10] to measure the similarity between these images and the benign prompts. Ideally, a successful concept removal method should preserve the benign part of the malicious prompt, resulting in a high in-prompt CLIP score. + +In the rest of the paper, we first revisit existing concept removal methods and then introduce the proposed dataset and evaluation metric in detail. Finally, we conduct comprehensive experiments to benchmark these methods and discuss our observations. + +# 2. Preliminaries + +# 2.1. Text-to-image Diffusion Models + +Diffusion models typically involve a forward diffusion process and its reverse diffusion process. The forward process is a $T$ -step Markov chain which transforms a sample from the image distribution to an isotropic Gaussian distribution $p$ . The reverse process is a $T$ -step Markov chain which transforms a Gaussian sample back to the image distribution. The reverse process can be represented as + +$$ +p _ {\theta} \left(x _ {0: T}\right) = p \left(x _ {T}\right) \prod_ {t = 1} ^ {T} p _ {\theta} \left(x _ {t - 1} \mid x _ {t}\right) \tag {1} +$$ + +$$ +p _ {\theta} \left(x _ {t - 1} \mid x _ {t}\right) = \mathcal {N} \left(x _ {t - 1}; \mu_ {\theta} \left(x _ {t}, t\right), \Sigma_ {\theta} \left(x _ {t}, t\right)\right), \tag {2} +$$ + +where $x_0$ is the image sample, $x_{T}$ is the Gaussian sample, and $x_{t}(0 < t < T)$ is the middle state of diffusion. The terms $\mu_{\theta}$ and $\Sigma_{\theta}$ are estimated by a denoising network $\epsilon_{\theta}(x_t,t)$ which takes $x_{t}$ and $t$ as input. By $\epsilon_{\theta}$ , an image sample $x_0$ can be generated from a Gaussian sample $x_{T}$ following Eq. (1). + +Text-to-image (T2I) diffusion models further extend the diffusion process by guiding it with a textual prompt $y$ [4, 15, 35]. Specifically, $\epsilon_{\theta}(x_t,t)$ is modified to take $y$ as an additional input, resulting in $\epsilon_{\theta}(x_t,y,t)$ . The network $\epsilon_{\theta}(x_t,y,t)$ uses a cross-attention module to select information from the textual prompt to guide the diffusion process. In cross attention, each token is transformed into key $(K)$ space and value $(V)$ space, while the image is transformed into query $(Q)$ space. The key and the query are multiplied to select information from the values of different tokens. With cross attention, T2I models can generate images that adhere to the provided text prompts. + +# 2.2. Concept Removals + +In this subsection, we revisit representative concept removal methods. + +Fine-tuning-based methods. Fine-tuning is the most common framework for concept removals. The objective for the fine-tuning-based algorithms can be summarized as + +$$ +\min _ {\theta} L = L _ {\mathrm {r m}} + \lambda L _ {\mathrm {r e g}}, \tag {3} +$$ + +where $L_{\mathrm{rm}}$ is the term for changing the output of unwanted concept, and $L_{\mathrm{reg}}$ is the regularization term to ensure that the fine-tuning will not influence other benign concepts and maintain the generation quality. The term $\theta$ is the parameter to fine-tune. The methods can be solved by two types of solutions: + +- Gradient descent. This solution [7, 16, 18, 23] focuses on the final output of $\epsilon_{\theta}(x_t,y,t)$ , modifying the model in an end-to-end way. The term $L_{\mathrm{rm}}$ changes the output of $\epsilon_{\theta}(x_t,c_u,t)$ where $c_{u}$ is unwanted concepts, while $L_{\mathrm{reg}}$ maintains the output of benign concepts $c_b$ . +- Closed-form solution. Modifying the intermediate states of linear components (such as cross-attention weights [8, 22] and MLP layers [42]) instead of the final output of $\epsilon_{\theta}(x_t, y_c, t)$ , can be solved in a closed-form solution. This can accelerate the fine-tuning process remarkably. + +Inference-time methods. Inference-time methods such as negative prompt [32], SLD [36], and SEGA [3] change the generation algorithm to remove concepts in the inference process, which can skip fine-tuning operations. They estimate the influence of unwanted concepts and remove them in inference. However, although these methods can skip the fine-tuning process, there is also a risk of being disabled in open-sourced models, which may slow down the inference. + +# 3. Six-CD: a Comprehensive and Effective Dataset + +In this section, we first propose a new dataset that encompasses a comprehensive set of categories of unwanted con + +cepts. Then we discuss the issue of ineffectiveness in general prompts and present the solution to filter the dataset, thereby increasing its effectiveness. + +# 3.1. Six Categories of Unwanted Concepts + +As mentioned in Sec. 1, summarizing all the datasets in existing literature, the unwanted concepts can be divided into general and specific concepts. The general concepts contain harmful concepts and nudity concepts, while the specific concepts contain identities of celebrities, copyrighted characters, objects, and art styles. However, existing literature focuses on only a few categories for evaluation, lacking systematic comparisons among the methods in general and specific concepts. To address this, we propose a new dataset, Six-CD, to evaluate concept removals comprehensively. + +For the general concepts, Six-CD first collects the malicious prompts from four different NSFW resources, which include I2P [36], MMA [43], jatman/stable-diffusion-prompts-stats-full-uncensored (SD-uncensored) [13], and Unsafe Diffusion (UD) [27]. Then, it divides the NSFW data into harmful and nudity concepts. Harmful concepts contain the prompts that have the meanings of "violence, suicide, hate, harassment, suffering, humiliation, harm, and bloodiness", while nudity contains "nudity, nakedness, sexuality, pornography, and eroticism". However, in MMA, UD, and SD-uncensored, the prompts are only labeled as NSFW or not and have no such fine-grained labels for harm and nudity. Thus, we use the image classifiers NudeNet [2] and Q16 [27] to annotate by detecting the NSFW contents in the images generated by the prompts. NudeNet is tailored for detecting images with nudity, while Q16 is a binary classifier for detecting NFSW. To annotate the fine-grained labels, i.e., harm and nudity, we first use NudeNet to find the prompts of nudity from the collected resources1. Then, in the remaining data, we use Q16 to select the prompts whose images are classified as NSFW. Since the nudity prompts are already filtered out by NudeNet, in the remaining data, the prompts detected by Q16 are annotated as harmful. With the two detectors, we merge the collected resources and annotate them as either harmful or nudity. + +For the specific concepts in Six-CD, instead of directly collecting prompts, we collect concepts and use prompt templates (detailed in Appd. A.1) to generate the final prompts, which are different from the general ones. The concepts are collected as follows: + +- Celebrity. Celebrities can evaluate the ability to remove identity. We select the identities of celebrities from CPDM [24]. To assist the evaluation, we use a celebrity detector, GCD [25]. we choose the celebrities that can be + +![](images/5f3fdef1887df2868e0e5bd0c4c975835ee2f6fd076e25d91d796ce3de310fc8.jpg) +Figure 1. Effectiveness of I2P prompts. Most of the prompts have low effectiveness. Results from [36]. + +both generated by SD and recognized by GCD. + +- Copyrighted characters. The copyrighted characters are important IP resources for companies. We use the character concepts in CPDM and the high-frequency copyrighted characters from FiveThirtyEight Comic Characters Dataset [1]. +- Objects. Objects can be used to evaluate the performance of removal on non-humanoid concepts. We randomly sample from a subset from the classes of ImageNet [5]. +- Art style. The art styles can evaluate the removal performance of a global feature instead of a local feature. We use the art styles from [22]. + +For the two general categories, we provide 6,326 prompts for harm concept and 4,518 effective prompts for nudity concept. For the specific concepts, we provide 94 concepts for the identity of celebrity, 100 concepts for copyrighted characters, 10 concepts for objects and 10 concepts for art styles. Four templates are used to construct the prompt for each specific concept. Six-CD is the first dataset that divides the concept removal into general and specific concepts. Through experiments, we observe that, for different methods, the removal abilities in general and specific concepts are distinct, which is detailed in Sec. 5.1. Besides its comprehensiveness in covering general and specific categories found in existing literature, in the following subsection, we show that our dataset also solves the problem of ineffective prompts of general concepts. + +# 3.2. Ineffective Prompts of General Concepts in Existing Datasets + +The prompts for general concepts in existing dataset do not consistently generate malicious content for each random seed according to [36] and our observation below. In this subsection, we discuss this phenomenon in general concepts and identify two potential problems it may introduce. + +We first define the effectiveness of a prompt as the possibility of generating malicious images. To assess this, we + +![](images/c8a8662863c0c0805e0a90f364c6dcc077de6afccb53e12337b83d7f2a04eeeb.jpg) +Figure 2. Concept removal (CR) performance on the prompts with different effectiveness. + +Table 2. Percentage of removed unwanted concepts between high effective and all effectiveness. (%) + +
NEGESDSPMSDDFMNUCEEMCIDSLDSEGACorr
Harm40.963.698.924.895.677.698.927.323.70.977
Harm-EF56.767.399.927.897.467.898.638.122.9
Nudity22.317.981.011.293.614.998.314.638.80.984
Nudity-EF11.016.088.94.292.612.397.216.358.4
+ +Table 3. Examples of Dual-Version Dataset for specific concepts. We generate a malicious version by ChatGPT and remove the unwanted concepts placeholder to get the benign version. + +
CategoryExample of malicious versionExample of benign version
Celeb. & +Copyrighted{celebrity} is dancing in the rain.dancing in the rain.
Object{object}, football, grass, house, tree, dogfootball, grass, house, tree, dog
Art styleA beautiful snow-covered mountain with +sunshine lighting it in the style of {art style}A beautiful snow-covered moun- +tain with sunshine lighting it
+ +generate $N$ images for a prompt using different random seeds. The effectiveness is defined by $n / N$ , where $n$ is the number of malicious images detected. In existing datasets, there is a large portion of ineffective prompts in general concepts. Taking I2P [36] as an example, which is the most well-used dataset for harmful and nudity concepts in existing literature [3, 7, 8, 16, 22, 23, 36, 42], most of the prompts have low effectiveness as shown in Fig. 1. We can see that the number of ineffective prompts significantly exceeds the number of effective prompts in I2P, particularly in the nudity category. The specific concepts usually do not have the problem of low effectiveness because they contain the specific and definitive for the unwanted concepts. Thus, in this subsection, we focus on general concepts. + +Limitations. The large portion of ineffective prompts can potentially introduce the following problems. First, evaluating the ineffective prompts is in low efficiency. For example, we test different concept removal methods on SD for nudity concepts in Fig. 2. On the vanilla SD without concept removal (i.e., No CR in Fig. 2), prompts with effectiveness of $1/5$ generated only about $20\%$ nudity images. + +It means that $80\%$ of the generated images by these prompts are non-nudity images, while evaluating on non-nudity images is meaningless. Second, the ineffective prompts may cause unfair comparisons between the general concepts and specific concepts. Unlike general concepts, the specific concepts do not have the problem of ineffective prompts. When the model builder has to remove both general and specific concepts, the ineffective prompts may falsely induce the builder to put less weight on the general concepts. + +High-effectiveness subsets. Therefore, in addition to harm and nudity concepts, we provide additional subsets of high-effectiveness prompts and name them as harm- $EF$ and nudity- $EF$ . We use the prompts with effectiveness of 4/5 and 5/5 for harm- $EF$ and nudity- $EF$ . Harm- $EF$ contains 991 prompts, and nudity- $EF$ contains 1,539 prompts. In Table 2 and Fig. 2, we demonstrate that the abilities of different concept removal methods are highly consistent between high effectiveness and low effectiveness. Thus, it is reasonable to evaluate on high-effectiveness prompts in order to reduce computation, which is also more fair. Specifically, in Table 2, we use the percentage of removed unwanted concepts to measure the ability of different methods, which is defined as + +$$ +p _ {\mathrm {r m}} = 1 - \frac {n _ {\text {a f t e r}}}{n _ {\text {b e f o r e}}}, \tag {4} +$$ + +where $n_{\mathrm{before}}$ is the number of images containing unwanted concepts before concept removal, while $n_{\mathrm{after}}$ is the number of images containing unwanted concepts after concept removal. The removal abilities in harm and nudity concepts are very close to harm- $EF$ and nudity- $EF$ . The Pearson correlation coefficient is as high as 0.977 (between harm and harm- $EF$ ) and 0.984 (between nudity and nudity- $EF$ ). + +In summary, Six-CD provides two choices to evaluate on the general concepts. The first choice is to use the whole subsets of harm and nudity, while the second choice is to evaluate on harm-EF and nudity-EF, which provides a more efficient and fair evaluation. + +# 4. In-prompt Retainability + +Concept removals should not hurt the generation ability of the benign contents, which is referred to as retainability. In this section, we show the lack of evaluation for retainability on the benign parts of the prompts containing the unwanted concepts and propose a new metric for this retainability. + +Lack of in-prompt retainability. Existing literature [24, 45] only considers the retainability of totally benign concepts. The benign prompts do not contain any unwanted concept, and concept removal should not influence the generation of them. Therefore, CLIP score [10] is utilized to measure it by calculating the similarity between benign prompts and generated images in the CLIP space [28]. However, for the prompts containing unwanted concepts, + +![](images/1ad7c80faac9de60e308487ea35a535ffca25bea4ba81bca07217cbc58ad40c2.jpg) +NO CR +ironman removed +Figure 3. Concept removal example of the prompt: "At night, ironman with his armor on is watching fireworks by the lake". Ironman is removed by SDD. + +![](images/87b1de8ea1f3d9fe8a7adf39cab04b716b816c699eeb5310c2033f3b2365bc66.jpg) + +the retainability to guarantee the generation of the benign part of them is also important, which is called "in-prompt retainability" in our benchmark. As shown in Fig. 3, the prompt is "At night, ironman with his armor on is watching fireworks by the lake" with "ironman" as the unwanted concept. Although there is no "ironman" after concept removal, the benign part of "fireworks" is also removed. In this case, even though the method can remove unwanted concepts, this generation is meaningless for the users because the benign information is not preserved. For users who do not intentionally include the unwanted concepts or who do not know the concept is malicious, this in-prompt retainability is necessary to ensure the normal usage. + +New metric. To measure in-prompt retainability, we propose a new metric, in-prompt CLIP score, assisted by a Dual-Version Dataset (DVD). In this dataset, we have two versions for each prompt: malicious and benign. The malicious version contains the unwanted concepts of the six categories, while the benign version removes the unwanted concepts with the rest the same as the malicious version. In in-prompt CLIP score, we first generate the images with concept removals using the malicious version and then calculate the cosine similarity of CLIP embeddings between these generated images and the corresponding benign version prompts. Thus, we can measure the similarity between the generated image and the benign part of the prompt. To distinguish it from the original CLIP score, we refer to the original CLIP score as the out-prompt CLIP score. + +To construct DVD, for general concepts, we first sample two subsets of prompts from the harmful and nudity categories of Six-CD as the malicious version, then we manually remove the unwanted concepts to get the benign version. For specific concepts, since they are not diverse or implicit like general concepts, we create templates for each category using ChatGPT and use the concepts from Six-CD to get the complete prompts. We show the examples of templates in DVD in Table 3. The entire dataset is in Appd. A.2. To validate the effectiveness of the metric, we conduct human evaluation in Appd. D.1. The results demonstrate that the metric is highly consistent with human preference. + +In summary, to measure the in-prompt retainability, we propose the in-prompt CLIP score. It leverages a malicious- + +version and a benign-version prompt to measure the similarity between the images after concept removal and the benign part of the prompt. + +# 5. Experiments + +In this section, we conduct experiments to benchmark the performance of 10 methods in removing single and multiple concepts and the retainability with both in-prompt and out-prompt CLIP scores. We also test FID, time costs of training and inference, and a fine-grained retainability for similar concepts. Due to space limitations, these experiments are in Appd D. Also, we report the results of SD v1.4 in this section and the results of other models are in Appd D. + +Experimental settings. We evaluate on 10 concept removal methods: negative prompt (NEG) [32], ESD [7], SPM [23], SDD [16], FMN [44], UCE [8], MACE [22], EMCID [42], SLD [36], and SEGA [3]. The details on the baseline settings can be found in Appd. B. The evaluated datasets are Six-CD and DVD, which are detailed in Appd. A. Following the conclusions in Sec. 3.2, we use harm-EF and nudity-EF for all the experiments in this section. All experiments use one GPU (A5000 or A6000). + +Detection metrics. We use the detection rate of Q16 and NudeNet for harmful and nudity concepts. We use the classification accuracy of GCD as the detection rate for celebrity concepts, and we train two classifiers based on ResNet-50 [9] and use the accuracy as the detection rate for copyrighted characters and objects. For the art style concept, the detection is not a simple binary problem. In existing literature [22-24, 44], CLIP score, $S_{\mathrm{CLIP}}$ , is used to measure the presence of an art style on a continuous scale, assessing the similarity between generated images and art styles. Following them, we use $S_{\mathrm{CLIP}}$ and normalize its range by $\tilde{S}_{\mathrm{CLIP}} = (S_{\mathrm{CLIP}} - \min(S_{\mathrm{CLIP}})) / (\max(S_{\mathrm{CLIP}}) - \min(S_{\mathrm{CLIP}}))$ for a better comparison with other categories. + +# 5.1. Evaluation on Removal Ability + +In this subsection, we test the removal ability of different methods on Six-CD. In Fig. 4, we show the detection results of models both with and without concept removals. For general concepts, we use all the prompts in harmful and nudity concepts in Six-CD for evaluation. For specific concepts, we randomly choose five concepts from each category and modify the diffusion model for each individual concept. We show the averaged results of 5 concepts of each category in Fig. 4. (The table with error bar for Fig. 4 can be found in Appd. D.2.) We highlight the main observations as follows. + +Observation 1: General concepts are harder to remove. The difficulty of removing general concepts is reflected by two factors shown in Fig 4. First, in most methods, the worst removal result lies in one of the general categories. For example, ESD and MACE perform well in all + +other categories, but they still have around $55\%$ harmful images detected. Second, while some methods can remove almost all the specific concepts (e.g., ESD, UCE, SLD), no one method can achieve similar results in both harm and nudity categories simultaneously. For general concepts, on the one hand, they are hard to trigger (i.e. low effectiveness). On the other hand, it is harder to remove them than specific concepts. We conjecture that they are both the results of the diverse and implicit prompts of general concepts. Locating the unwanted concepts in the diverse and implicit prompts of general concepts is more difficult than the explicit prompts of specific concepts. Most of the methods like ESD can only provide a limited number of tokens as the general concepts, which is impossible to cover the entire (almost unlimited) vocabulary of harm and nudity. Thus, it leads to difficulty in removing general concepts. + +Observation 2: Removal ability is consistent within general concepts and within specific categories. We observe that the removal abilities within general or specific categories are usually more consistent. For example, SLD and FMN perform well in specific concepts, but the ability to remove general concepts is worse. To demonstrate this observation, we present the correlation coefficient of the detection metrics among different categories in Fig. 5. From Fig. 5, we can see that the correlations between specific concepts are higher than 0.59, with the correlation between celebrity and character notably high, at 0.94. The correlation between general concepts is also as high as 0.65. This offers an insight into the designation of concept removal methods that general and specific concepts should be considered separately. + +# 5.2. Removal Ability on Multiple Concepts + +In Sec. 5.1, we evaluate the ability to remove a single concept from one model. However, removing multiple concepts from one model is also important since the model builder usually needs to consider multiple malicious concepts in practice. In this subsection, we present the results of removing multiple concepts of celebrities and copyrighted characters in Fig. 6, where the y-axis represents the detection rate, and a higher detection rate means a weaker concept removal performance. We have two observations from Fig. 6. + +Observation 1: Inference-time methods fail in removing multiple concepts. As shown in Fig. 6, the inference-time methods, SLD and NEG, are the only two methods that have significantly limited removal performance when the concept number is larger than 50, which only reduces the detection rate by $20\%$ or less. To remove the multiple concepts in inference time, they have to encode the string containing all the concepts in the embeddings of one single prompt. This will exceed the capacity of the text encoder of T2I diffusion models and lead to failed removal. Another inference-time method, SEGA, is not reported in Fig. 6 due + +![](images/7ddc365e0d2702dfb1165ce44bc08ace002f3e00706682b06ee3df244d451943.jpg) +Figure 4. Removal ability on Six-CD. In each sub-figure, we present the detection results of unwanted concepts in both the models with and without concept removals. The range of all the six values is [0,1]. Smaller values indicate that less unwanted concepts are detected, i.e. better removal. G1 and G2 are harm-EF and nudity-EF concepts, while S1, S2, S3 and S4 are celebrity, character, object and art style. + +![](images/e6b4f9fd9ba6540935fd33b7f7f0ebc824ab9fcefb3378fe96c1b890c67b027f.jpg) +Figure 5. Correlation coefficient of removal performance on six data categories. + +to its catastrophic inference time when the concept number is increased to 50. This implies that the strategy of processing each concept separately during inference is impractical. + +Observation 2: Closed-form solutions by modifying linear components perform well in removing multiple concepts. The best two methods in Fig. 6a and Fig. 6b are both based on the closed-form solution, which modifies the linear components of cross attention. This is possibly because combining multiple concepts in the linear components is easier than in other non-linear parts. + +# 5.3. In-prompt and Out-prompt Retainability + +We evaluate the generation ability on benign concepts using both in-prompt and out-prompt CLIP scores. In Fig. 7, we + +use DVD to test the in-prompt CLIP score and a subset with benign prompts from LAION [37] to test out-prompt CLIP score. The observations as follows. + +Observation 1: In-prompt retainability performs worse than out-prompt retainability. As shown in Fig. 7, the in-prompt CLIP score is lower than the out-prompt CLIP score across all the methods and concept categories. This indicates that the ability to generate benign parts in prompts containing malicious concepts is more negatively affected than the ability to generate totally benign prompts. As mentioned in Sec. 4, the in-prompt retainability is also important for concept removal methods. This inspires future research to pay more attention to the in-prompt retainability. Remark. It could be unfair to directly compare the two CLIP scores since they use different prompt sets to calculate CLIP scores. To further validate this observation, we use the clean version of DVD to calculate the out-prompt CLIP score, which is detailed in Appd. D.3. By comparing it with the in-prompt CLIP score in Appd. D.3, we have observed a consistent phenomenon with this subsection. + +Observation 2: Auxiliary semantic information in MACE (Fig. 7g) may hurt the generation on benign concepts. MACE has the most significant decrease in the generation ability of benign concepts for both in-prompt and out-prompt CLIP scores. We conjecture that this is because it introduces auxiliary semantic information to help locate the unwanted concepts in the image. In other methods, the unwanted concepts are usually located by the semantic understanding ability of the T2I model itself. In contrast, + +![](images/a5d64223235972fafc877ec1758d1b3ae033cac5d1c7d1c262f86b6f2dd43c42.jpg) +(a) Celebrity + +![](images/f3d1443a7d24269896e363541688c346307daf628affba5589c3860b62b571b2.jpg) +(b) Copyrighted characters + +![](images/b944d48ca20e7ffa160b2285a843dc8b16df00b09e90ae99db440a22a487f722.jpg) +(a) NEG + +![](images/19a87723d288a62a5fd2695fa21a6cedc26011f2ab09c4f0fd8f37762bcd97c8.jpg) +(b) ESD + +![](images/363cde74017eb92a28912f424b314834362cc1a4e939f4ec9ccee86e5b8384e5.jpg) +(c) SPM + +![](images/299db817f9401300a9b45ea440479838891b67830126357154c1089e5305fd65.jpg) +Figure 6. Removal ability on multiple concepts + +![](images/c0f38fcbffd8c79370fd6bcee7377c5fd4f36ddc8a9fdfa8cc1a22db194af946.jpg) +(e) FMN + +![](images/0614252b091d5fe1f58ca4db93471406dab4f59a2668f356a4595a712f209e70.jpg) +(f) UCE + +![](images/a24fabf24dbbdb08ce3c99fbf769814179a713bdca68eaee51ae78ccd8227391.jpg) +(g) MACE + +![](images/debd13ddd2b714f3e825e6052eecb916c51bca6ba03b51e56ce971d4f94772a6.jpg) +(h) EMCID + +![](images/95dedda416ff55074e5ee8ab412d29a08ff4cc208807325be0d7db6b110d84f2.jpg) +(d) SDD +(i) SLD + +![](images/1856dd8da1267124d798bf0d2479483fb2f0908c865cf7a27bf7457a69e926ee.jpg) +(j) SEGA +Figure 7. In-prompt retainability vs. out-prompt retainability + +MACE incorporates the segmentation results of GroundedSAM [17, 21] to locate the image area of unwanted concepts. It focuses on optimizing this area and overlooks other areas of the image. This may lead to worse retainability in other areas. Another piece of evidence is that when we do not include segmentation in removing the art style concepts by MACE (since art style is a global feature), it has good in-prompt and out-prompt CLIP scores. + +# 6. Conclusion and Ethical Statement + +In this work, we address the lack of a systematic benchmark for concept removal methods by introducing a comprehensive and effective dataset and a new evaluation metric. Using this benchmark, we conduct an extensive evaluation of concept removal methods. Our experimental observations + +provide valuable and practical insights for future research in this field. While our benchmark may have a limitation in the inconvenient detection metric for art styles (using the CLIP score for art styles, instead of the detection rate from a classifier like other categories), it still offers a thorough and practical evaluation across the various method settings. + +Ethical statement. This work provides a comprehensive evaluation framework for concept removal, which is targeted to mitigate the potential malicious use of text-to-image diffusion models. Our research is conducted responsibly, transparently, and deeply dedicated to ethical standards. Despite involving the generation of sensitive content such as nudity and violence, it is strictly for research purposes and does not intend to produce or promote inappropriate material. On the contrary, our work aims to advance efforts to prevent the generation of inappropriate content. + +# Acknowledgement + +Jie Ren, Kangrui Chen, Yingqian Cui, Shenglai Zeng, and Jiliang Tang are supported by the National Science Foundation (NSF) under grant numbers CNS2321416, IIS2212032, IIS2212144, IOS2107215, DUE2234015, CNS2246050, DRL2405483 and IOS2035472, the Army Research Office (ARO) under grant number W911NF-21-1-0198, Amazon Faculty Award, JP Morgan Faculty Award, Meta, Microsoft and SNAP. + +# References + +[1] Fivethirtyeight comic characters dataset. https://www.kaggle.com/datasets/fivethirtyeight/fivethirtyeight - comic - characters - dataset.4 +[2] Nudenet: Lightweight nudity detection. https://github.com/notAI-tech/NudeNet, 2023.3 +[3] Manuel Brack, Felix Friedrich, Dominik Hintersdorf, Lukas Struppek, Patrick Schramowski, and Kristian Kersting. Sega: Instructing text-to-image models using semantic guidance. Advances in Neural Information Processing Systems, 36, 2024. 1, 2, 3, 4, 6 +[4] Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, et al. Pixart-α: Fast training of diffusion transformer for photorealistic text-to-image synthesis. arXiv preprint arXiv:2310.00426, 2023. 1, 3 +[5] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 4 +[6] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 1 +[7] Rohit Gandikota, Joanna Materzynska, Jaden Fiitto-Kaufman, and David Bau. Erasing concepts from diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2426-2436, 2023. 1, 2, 3, 4, 6 +[8] Rohit Gandikota, Hadas Orgad, Yonatan Belinkov, Joanna Materzyńska, and David Bau. Unified concept editing in diffusion models. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 5111-5120, 2024. 1, 2, 3, 4, 6 +[9] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6 +[10] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning. arXiv preprint arXiv:2104.08718, 2021. 2, 5 +[11] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 1 +[12] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840-6851, 2020. 1 + +[13] Huggingface. jatman/stable-diffusion-prompts-stats-full-uncensored datasets at hugging face. https://huggingface.co/datasets/jatman/stable-diffusion-prompts-stats-full-uncensored?not-for-all-audiences=true. 3 +[14] Tatum Hunter. Ai porn is easy to make now. for women, that's a nightmare. https://www.washingtonpost.com/technology/2023/02/13/ai-porn-deepfakes-women-consent/, 2023. Accessed: 2024-06-01. 1 +[15] DeepFloyd IF. https://github.com/deep-floyd/if.3 +[16] Sanghyun Kim, Seohyeon Jung, Balhae Kim, Moonseok Choi, Jinwoo Shin, and Juho Lee. Towards safe self-distillation of internet-scale text-to-image diffusion models. arXiv preprint arXiv:2307.05977, 2023. 2, 3, 4, 6 +[17] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 8 +[18] Nupur Kumari, Bingliang Zhang, Sheng-Yu Wang, Eli Shechtman, Richard Zhang, and Jun-Yan Zhu. Ablating concepts in text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22691-22702, 2023. 1, 3 +[19] Yuantong Li, Chi-Hua Wang, and Guang Cheng. Online forgetting process for linear regression models. In International Conference on Artificial Intelligence and Statistics, pages 217-225. PMLR, 2021. 1 +[20] Yaxin Li, Jie Ren, Han Xu, and Hui Liu. Neural style protection: Counteracting unauthorized neural style transfer. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 3966-3975, 2024. 1 +[21] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023.8 +[22] Shilin Lu, Zilan Wang, Leyang Li, Yanzhu Liu, and Adams Wai-Kin Kong. Mace: Mass concept erasure in diffusion models. arXiv preprint arXiv:2403.06135, 2024. 1, 2, 3, 4, 6 +[23] Mengyao Lyu, Yuhong Yang, Haiwen Hong, Hui Chen, Xuan Jin, Yuan He, Hui Xue, Jungong Han, and Guiguang Ding. One-dimensional adapter to rule them all: Concepts, diffusion models and erasing applications. arXiv preprint arXiv:2312.16145, 2023. 2, 3, 4, 6 +[24] Rui Ma, Qiang Zhou, Bangjun Xiao, Yizhu Jin, Daquan Zhou, Xiuyu Li, Aishani Singh, Yi Qu, Kurt Keutzer, Xiaodong Xie, et al. A dataset and benchmark for copyright protection from text-to-image diffusion models. arXiv preprint arXiv:2403.12052, 2024. 2, 3, 5, 6 +[25] Dmitry Voitekh Nick Hasty, Ihor Kroosh and Dmytro Korduban. Giphy's open source celebrity detection deep learning model and code. https://github.com/Giphy/celeb-detection-oss, 2019.3 +[26] OpenAI. Dall-e 2 preview - risks and limitations. https://github.com/openai/dalle-2-preview/ + +blob/main/system-card.md, 2023. Accessed: 2024-06-01. 1 +[27] Yiting Qu, Xinyue Shen, Xinlei He, Michael Backes, Savvas Zannettou, and Yang Zhang. Unsafe Diffusion: On the Generation of Unsafe Images and Hateful Memes From Text-To-Image Models. In ACM SIGSAC Conference on Computer and Communications Security (CCS). ACM, 2023. 3 +[28] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 5 +[29] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 1 +[30] Javier Rando, Daniel Paleka, David Lindner, Lennart Heim, and Florian Tramèr. Red-teaming the stable diffusion safety filter. arXiv preprint arXiv:2210.04610, 2022. 1 +[31] Dana Rao. Responsible innovation in the age of generative ai: Adobe blog, . 1 +[32] Dana Rao. Stable diffusion 2.0 release, .1, 3, 6 +[33] Jie Ren, Han Xu, Pengfei He, Yingqian Cui, Shenglai Zeng, Jiankun Zhang, Hongzhi Wen, Jiayuan Ding, Hui Liu, Yi Chang, et al. Copyright protection in generative ai: A technical perspective. arXiv preprint arXiv:2402.02333, 2024. 1 +[34] Jie Ren, Zhenwei Dai, Xianfeng Tang, Hui Liu, Jingying Zeng, Zhen Li, Rahul Goutam, Suhang Wang, Yue Xing, and Qi He. A general framework to enhance fine-tuning-based llm unlearning. arXiv preprint arXiv:2502.17823, 2025. 1 +[35] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 3 +[36] Patrick Schramowski, Manuel Brack, Björn Deiseroth, and Kristian Kersting. Safe latent diffusion: Mitigating inappropriate degeneration in diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22522-22531, 2023. 1, 2, 3, 4, 6 +[37] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278–25294, 2022. 7 +[38] Shawn Shan, Jenna Cryan, Emily Wenger, Haitao Zheng, Rana Hanocka, and Ben Y Zhao. Glaze: Protecting artists from style mimicry by {Text-to-Image} models. In 32nd USENIX Security Symposium (USENIX Security 23), pages 2187–2204, 2023. 1 +[39] Gowthami Somepalli, Vasu Singla, Micah Goldblum, Jonas Geiping, and Tom Goldstein. Diffusion art or digital forgery? investigating data replication in diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6048-6058, 2023. 1 + +[40] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 1 +[41] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 1 +[42] Tianwei Xiong, Yue Wu, Enze Xie, Zhenguo Li, and Xihui Liu. Editing massive concepts in text-to-image diffusion models. arXiv preprint arXiv:2403.13807, 2024. 1, 2, 3, 4, 6 +[43] Yijun Yang, Ruiyuan Gao, Xiaosen Wang, Nan Xu, and Qiang Xu. Mma-diffusion: Multimodal attack on diffusion models. arXiv preprint arXiv:2311.17516, 2023. 3 +[44] Eric Zhang, Kai Wang, Xingqian Xu, Zhangyang Wang, and Humphrey Shi. Forget-me-not: Learning to forget in text-to-image diffusion models. arXiv preprint arXiv:2303.17591, 2023. 1, 2, 6 +[45] Yihua Zhang, Yimeng Zhang, Yuguang Yao, Jinghan Jia, Jiancheng Liu, Xiaoming Liu, and Sijia Liu. Unlearncanvas: A stylized image dataset to benchmark machine unlearning for diffusion models. arXiv preprint arXiv:2402.11846, 2024. 2, 5 \ No newline at end of file diff --git a/2025/Six-CD_ Benchmarking Concept Removals for Text-to-image Diffusion Models/images.zip b/2025/Six-CD_ Benchmarking Concept Removals for Text-to-image Diffusion Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..f003a49bd504e4b018e1e4fb230ed7cd8c40261b --- /dev/null +++ b/2025/Six-CD_ Benchmarking Concept Removals for Text-to-image Diffusion Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:14dad6939de95cce6823deebf6981fd8eba4553bea88ee693f60446a34586023 +size 413356 diff --git a/2025/Six-CD_ Benchmarking Concept Removals for Text-to-image Diffusion Models/layout.json b/2025/Six-CD_ Benchmarking Concept Removals for Text-to-image Diffusion Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..59738c2732d3f12490efa0d3a7081df11cb679b9 --- /dev/null +++ b/2025/Six-CD_ Benchmarking Concept Removals for Text-to-image Diffusion Models/layout.json @@ -0,0 +1,8278 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 63, + 103, + 547, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 103, + 547, + 121 + ], + "spans": [ + { + "bbox": [ + 63, + 103, + 547, + 121 + ], + "type": "text", + "content": "Six-CD: Benchmarking Concept Removals for Text-to-image Diffusion Models" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 113, + 142, + 497, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 142, + 497, + 185 + ], + "spans": [ + { + "bbox": [ + 113, + 142, + 497, + 185 + ], + "type": "text", + "content": "Jie Ren" + }, + { + "bbox": [ + 113, + 142, + 497, + 185 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 113, + 142, + 497, + 185 + ], + "type": "text", + "content": ", Kangrui Chen" + }, + { + "bbox": [ + 113, + 142, + 497, + 185 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 113, + 142, + 497, + 185 + ], + "type": "text", + "content": ", Yingqian Cui" + }, + { + "bbox": [ + 113, + 142, + 497, + 185 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 113, + 142, + 497, + 185 + ], + "type": "text", + "content": ", Shenglai Zeng" + }, + { + "bbox": [ + 113, + 142, + 497, + 185 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 113, + 142, + 497, + 185 + ], + "type": "text", + "content": ", Hui Liu" + }, + { + "bbox": [ + 113, + 142, + 497, + 185 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 113, + 142, + 497, + 185 + ], + "type": "text", + "content": ", Yue Xing" + }, + { + "bbox": [ + 113, + 142, + 497, + 185 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 113, + 142, + 497, + 185 + ], + "type": "text", + "content": ", Jiliang Tang" + }, + { + "bbox": [ + 113, + 142, + 497, + 185 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 113, + 142, + 497, + 185 + ], + "type": "text", + "content": ", Lingjuan Lyu" + }, + { + "bbox": [ + 113, + 142, + 497, + 185 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 113, + 142, + 497, + 185 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 113, + 142, + 497, + 185 + ], + "type": "text", + "content": "Michigan State University " + }, + { + "bbox": [ + 113, + 142, + 497, + 185 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 113, + 142, + 497, + 185 + ], + "type": "text", + "content": "Sony AI" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 116, + 186, + 494, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 186, + 494, + 213 + ], + "spans": [ + { + "bbox": [ + 116, + 186, + 494, + 213 + ], + "type": "text", + "content": "{renjie3,chenkan4,cuiyingq,zengshel,liuhui7,xingyue1,tangjili}@msu.edu \nlingjuan.lv@sony.com" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 241, + 200, + 253 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 241, + 200, + 253 + ], + "spans": [ + { + "bbox": [ + 151, + 241, + 200, + 253 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 266, + 297, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 266, + 297, + 517 + ], + "spans": [ + { + "bbox": [ + 54, + 266, + 297, + 517 + ], + "type": "text", + "content": "Text-to-image (T2I) diffusion models have shown exceptional capabilities in generating images that closely correspond to textual prompts. However, the advancement of T2I diffusion models presents significant risks, as the models could be exploited for malicious purposes, such as generating images with violence or nudity, or creating unauthorized portraits of public figures in inappropriate contexts. To mitigate these risks, concept removal methods have been proposed. These methods aim to modify diffusion models to prevent the generation of malicious and unwanted concepts. Despite these efforts, existing research faces several challenges: (1) a lack of consistent comparisons on a comprehensive dataset, (2) ineffective prompts in harmful and nudity concepts, (3) overlooked evaluation of the ability to generate the benign part within prompts containing malicious concepts. To address these gaps, we propose to benchmark the concept removal methods by introducing a new dataset, Six-CD, along with a novel evaluation metric. In this benchmark, we conduct a thorough evaluation of concept removals, with the experimental observations and discussions offering valuable insights in the field." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 538, + 135, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 538, + 135, + 550 + ], + "spans": [ + { + "bbox": [ + 55, + 538, + 135, + 550 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 557, + 295, + 701 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 557, + 295, + 701 + ], + "spans": [ + { + "bbox": [ + 55, + 557, + 295, + 701 + ], + "type": "text", + "content": "Diffusion models have demonstrated remarkable capabilities in image generation [6, 11, 12, 40, 41], particularly text-to-image (T2I) diffusion models [4, 29, 32]. These models have gained widespread popularity and deployment due to their ability to generate images that precisely align with textual prompts. However, T2I diffusion models can be exploited for malicious purposes, such as creating images depicting violence, nudity, or fake celebrity images in undesirable contexts like jail [14, 20, 26, 30, 33, 38, 39]. To ensure the integrity of the model developers, it is crucial to design benign models that censor malicious concepts and produce only safe and appropriate content." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "type": "text", + "content": "Strategies have been proposed to mitigate the malicious" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 242, + 555, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 242, + 555, + 469 + ], + "spans": [ + { + "bbox": [ + 313, + 242, + 555, + 469 + ], + "type": "text", + "content": "generation. For example, in the data collection stage, Stable Diffusion (SD) [32] and Adobe [31] filter out the Not-Safe-For-Work (NSFW) contents in training data. After image generation, SD uses a detector to ensure the images are benign, removing unwanted ones. However, training data filtering requires retraining and will be inefficient if the model has already been trained on various data. For open-sourced models like SD [32], the detector can be easily disabled by modifying the source code. Therefore, concept removal techniques have been proposed to edit a trained model's parameters to prevent generating malicious concepts [7, 18, 19, 34, 44]. For instance, methods are proposed to fine-tune the models to replace unwanted concepts with benign alternatives [7, 18, 44]. Besides, to accelerate fine-tuning, some methods fine-tune the linear components (e.g. the weight matrix of cross attention) using the closed-form optimal solution [8, 22, 42]. Alternatively, there are methods modifying the output during inference to bypass fine-tuning [3, 36]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 472, + 555, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 472, + 555, + 495 + ], + "spans": [ + { + "bbox": [ + 313, + 472, + 555, + 495 + ], + "type": "text", + "content": "While there is rich literature improving concept removal methods, we observe three potential issues:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 318, + 498, + 556, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 319, + 498, + 554, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 498, + 554, + 557 + ], + "spans": [ + { + "bbox": [ + 319, + 498, + 554, + 557 + ], + "type": "text", + "content": "1. Lack of consistent and comprehensive comparisons. Current concept removal methods, as shown in Table 1, typically analyze limited categories, lacking consistent and comprehensive comparisons. This gap prevents a thorough understanding of the methods' behavior." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 318, + 558, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 558, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 318, + 558, + 556, + 714 + ], + "type": "text", + "content": "2. Ineffective prompts. In existing datasets [36], some categories, such as nudity, contain a large portion of \"ineffective prompts\", which only trigger the model to generate malicious contents with a low probability (see Fig. 1). This can result in inefficient evaluation since the ineffective prompts will generate many benign images, and evaluating concept removal on them is meaningless. Meanwhile, in other categories of unwanted concepts like celebrities, the prompts are more effective than the nudity concept (see Fig. 1 and Fig. 4). When the builder has to remove both nudity and celebrity concepts, the ineffective prompts may cause unfair comparisons. The effective prompts in celebrity can generate" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28769" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 91, + 553, + 167 + ], + "blocks": [ + { + "bbox": [ + 192, + 71, + 419, + 82 + ], + "lines": [ + { + "bbox": [ + 192, + 71, + 419, + 82 + ], + "spans": [ + { + "bbox": [ + 192, + 71, + 419, + 82 + ], + "type": "text", + "content": "Table 1. Categories of unwanted concepts in existing literature" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 91, + 553, + 167 + ], + "lines": [ + { + "bbox": [ + 58, + 91, + 553, + 167 + ], + "spans": [ + { + "bbox": [ + 58, + 91, + 553, + 167 + ], + "type": "table", + "html": "
ESD [7]SPM [23]SDD [16]FMN [44]UCE [8]MACE [22]EMCID [42]SLD [36]SEGA [3]UC [45]CPDM [24]
Harmful✓✓✓
Nudity✓✓✓✓✓✓✓✓
Celebrity✓✓✓
Copyrighted✓✓
Object✓✓✓✓✓✓✓✓
art style✓✓✓✓✓✓✓✓✓✓
", + "image_path": "fdff7c92cf7e25216b8e45986f0117aaba33044d8dd9795ac839f304512ae344.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 70, + 187, + 294, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 187, + 294, + 212 + ], + "spans": [ + { + "bbox": [ + 70, + 187, + 294, + 212 + ], + "type": "text", + "content": "more malicious images, which may falsely induce the builder to put more weight on celebrity concepts." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 59, + 212, + 295, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 212, + 295, + 415 + ], + "spans": [ + { + "bbox": [ + 59, + 212, + 295, + 415 + ], + "type": "text", + "content": "3. Lack of evaluation on in-prompt retainability. Existing evaluations consider the generation ability on the benign prompts but overlook the evaluation of the generation ability on the benign part in the prompts with unwanted concepts, which is called in-prompt retainability in our benchmark. To understand retainability, when the unwanted concepts are removed from a generated image, the newly generated image should retain the remaining semantics in the prompt. For example, if we remove \"Mickey Mouse\" from the prompt of \"Mickey Mouse is eating a burger\", the generated image should still depict \"eating a burger\". If the method is too aggressive, it may remove benign semantics along with the malicious concepts, thus impairing the T2I model's ability to follow the textual prompt to generate meaningful images. Despite its importance, in-prompt retainability is ignored in existing literature." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 418, + 296, + 465 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 418, + 296, + 465 + ], + "spans": [ + { + "bbox": [ + 55, + 418, + 296, + 465 + ], + "type": "text", + "content": "To tackle the aforementioned problems, we aim to benchmark concept removal methods by introducing a comprehensive dataset and a new evaluation metric. Our contributions are as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 59, + 474, + 295, + 714 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 60, + 474, + 295, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 474, + 295, + 617 + ], + "spans": [ + { + "bbox": [ + 60, + 474, + 295, + 617 + ], + "type": "text", + "content": "1. A comprehensive dataset. We propose a new dataset, Six-CD, which contains Six categories of unwanted Concepts in Diffusion models, including harm, nudity, identities of celebrities, copyrighted characters, objects, and art styles. We divide the six categories into two groups: general concepts and specific concepts. General concepts, such as harm and nudity, are of concern to all users. Specific concepts are subject to a specific entity, e.g., a person or a company. While they may not be malicious to everyone, they can infringe on the rights of the concept owner, such as portraits of celebrities and copyrighted characters." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 59, + 618, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 618, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 59, + 618, + 295, + 714 + ], + "type": "text", + "content": "2. Effective prompts. Through extensive experiments in Sec. 3.2 and Sec. 5.1, we observe that ineffective prompts happen more frequently in general concepts. This is because the prompts of these concepts are usually diverse and implicit, unlike the precise and explicit prompts for specific concepts. Thus, in Six-CD, we provide additional subsets (with highly effective prompts only) for general concepts. Evaluation of ef" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 330, + 187, + 553, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 330, + 187, + 553, + 212 + ], + "spans": [ + { + "bbox": [ + 330, + 187, + 553, + 212 + ], + "type": "text", + "content": "fective prompts can be more efficient and also fairer when compared with specific concepts." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 318, + 212, + 555, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 212, + 555, + 413 + ], + "spans": [ + { + "bbox": [ + 318, + 212, + 555, + 413 + ], + "type": "text", + "content": "3. A new evaluation metric. We introduce a novel evaluation metric, called in-prompt CLIP score, to measure the in-prompt retainability. An expected concept removal method should have the ability to generate the benign part of the prompt when the unwanted concepts are removed. To evaluate in-prompt retainability, we construct a Dual-Version Dataset, in which each prompt has two versions: a malicious version contains the unwanted concept, and a benign version with the unwanted concept removed but the rest the same as the malicious version. We apply concept removal methods to generate images from the malicious version and then use the CLIP score [10] to measure the similarity between these images and the benign prompts. Ideally, a successful concept removal method should preserve the benign part of the malicious prompt, resulting in a high in-prompt CLIP score." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 413, + 554, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 413, + 554, + 472 + ], + "spans": [ + { + "bbox": [ + 313, + 413, + 554, + 472 + ], + "type": "text", + "content": "In the rest of the paper, we first revisit existing concept removal methods and then introduce the proposed dataset and evaluation metric in detail. Finally, we conduct comprehensive experiments to benchmark these methods and discuss our observations." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 481, + 399, + 493 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 481, + 399, + 493 + ], + "spans": [ + { + "bbox": [ + 313, + 481, + 399, + 493 + ], + "type": "text", + "content": "2. Preliminaries" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 501, + 485, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 501, + 485, + 514 + ], + "spans": [ + { + "bbox": [ + 313, + 501, + 485, + 514 + ], + "type": "text", + "content": "2.1. Text-to-image Diffusion Models" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 518, + 554, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 518, + 554, + 602 + ], + "spans": [ + { + "bbox": [ + 313, + 518, + 554, + 602 + ], + "type": "text", + "content": "Diffusion models typically involve a forward diffusion process and its reverse diffusion process. The forward process is a " + }, + { + "bbox": [ + 313, + 518, + 554, + 602 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 518, + 554, + 602 + ], + "type": "text", + "content": "-step Markov chain which transforms a sample from the image distribution to an isotropic Gaussian distribution " + }, + { + "bbox": [ + 313, + 518, + 554, + 602 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 313, + 518, + 554, + 602 + ], + "type": "text", + "content": ". The reverse process is a " + }, + { + "bbox": [ + 313, + 518, + 554, + 602 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 518, + 554, + 602 + ], + "type": "text", + "content": "-step Markov chain which transforms a Gaussian sample back to the image distribution. The reverse process can be represented as" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 345, + 606, + 553, + 622 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 606, + 553, + 622 + ], + "spans": [ + { + "bbox": [ + 345, + 606, + 553, + 622 + ], + "type": "interline_equation", + "content": "p _ {\\theta} \\left(x _ {0: T}\\right) = p \\left(x _ {T}\\right) \\prod_ {t = 1} ^ {T} p _ {\\theta} \\left(x _ {t - 1} \\mid x _ {t}\\right) \\tag {1}", + "image_path": "7252e59cf1a76d05106334f782b46bc499ccf747bed9db765a48c3cd7b3c2273.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 328, + 623, + 553, + 636 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 623, + 553, + 636 + ], + "spans": [ + { + "bbox": [ + 328, + 623, + 553, + 636 + ], + "type": "interline_equation", + "content": "p _ {\\theta} \\left(x _ {t - 1} \\mid x _ {t}\\right) = \\mathcal {N} \\left(x _ {t - 1}; \\mu_ {\\theta} \\left(x _ {t}, t\\right), \\Sigma_ {\\theta} \\left(x _ {t}, t\\right)\\right), \\tag {2}", + "image_path": "694d9e7cb78d526e9caa1c27346d2aa45001e605dd30d7de0c45ae1df34d75e0.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 642, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 554, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 642, + 554, + 714 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 313, + 642, + 554, + 714 + ], + "type": "text", + "content": " is the image sample, " + }, + { + "bbox": [ + 313, + 642, + 554, + 714 + ], + "type": "inline_equation", + "content": "x_{T}" + }, + { + "bbox": [ + 313, + 642, + 554, + 714 + ], + "type": "text", + "content": " is the Gaussian sample, and " + }, + { + "bbox": [ + 313, + 642, + 554, + 714 + ], + "type": "inline_equation", + "content": "x_{t}(0 < t < T)" + }, + { + "bbox": [ + 313, + 642, + 554, + 714 + ], + "type": "text", + "content": " is the middle state of diffusion. The terms " + }, + { + "bbox": [ + 313, + 642, + 554, + 714 + ], + "type": "inline_equation", + "content": "\\mu_{\\theta}" + }, + { + "bbox": [ + 313, + 642, + 554, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 642, + 554, + 714 + ], + "type": "inline_equation", + "content": "\\Sigma_{\\theta}" + }, + { + "bbox": [ + 313, + 642, + 554, + 714 + ], + "type": "text", + "content": " are estimated by a denoising network " + }, + { + "bbox": [ + 313, + 642, + 554, + 714 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(x_t,t)" + }, + { + "bbox": [ + 313, + 642, + 554, + 714 + ], + "type": "text", + "content": " which takes " + }, + { + "bbox": [ + 313, + 642, + 554, + 714 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 313, + 642, + 554, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 642, + 554, + 714 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 642, + 554, + 714 + ], + "type": "text", + "content": " as input. By " + }, + { + "bbox": [ + 313, + 642, + 554, + 714 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 313, + 642, + 554, + 714 + ], + "type": "text", + "content": ", an image sample " + }, + { + "bbox": [ + 313, + 642, + 554, + 714 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 313, + 642, + 554, + 714 + ], + "type": "text", + "content": " can be generated from a Gaussian sample " + }, + { + "bbox": [ + 313, + 642, + 554, + 714 + ], + "type": "inline_equation", + "content": "x_{T}" + }, + { + "bbox": [ + 313, + 642, + 554, + 714 + ], + "type": "text", + "content": " following Eq. (1)." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "28770" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 216 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 216 + ], + "type": "text", + "content": "Text-to-image (T2I) diffusion models further extend the diffusion process by guiding it with a textual prompt " + }, + { + "bbox": [ + 55, + 72, + 294, + 216 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 55, + 72, + 294, + 216 + ], + "type": "text", + "content": " [4, 15, 35]. Specifically, " + }, + { + "bbox": [ + 55, + 72, + 294, + 216 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(x_t,t)" + }, + { + "bbox": [ + 55, + 72, + 294, + 216 + ], + "type": "text", + "content": " is modified to take " + }, + { + "bbox": [ + 55, + 72, + 294, + 216 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 55, + 72, + 294, + 216 + ], + "type": "text", + "content": " as an additional input, resulting in " + }, + { + "bbox": [ + 55, + 72, + 294, + 216 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(x_t,y,t)" + }, + { + "bbox": [ + 55, + 72, + 294, + 216 + ], + "type": "text", + "content": ". The network " + }, + { + "bbox": [ + 55, + 72, + 294, + 216 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(x_t,y,t)" + }, + { + "bbox": [ + 55, + 72, + 294, + 216 + ], + "type": "text", + "content": " uses a cross-attention module to select information from the textual prompt to guide the diffusion process. In cross attention, each token is transformed into key " + }, + { + "bbox": [ + 55, + 72, + 294, + 216 + ], + "type": "inline_equation", + "content": "(K)" + }, + { + "bbox": [ + 55, + 72, + 294, + 216 + ], + "type": "text", + "content": " space and value " + }, + { + "bbox": [ + 55, + 72, + 294, + 216 + ], + "type": "inline_equation", + "content": "(V)" + }, + { + "bbox": [ + 55, + 72, + 294, + 216 + ], + "type": "text", + "content": " space, while the image is transformed into query " + }, + { + "bbox": [ + 55, + 72, + 294, + 216 + ], + "type": "inline_equation", + "content": "(Q)" + }, + { + "bbox": [ + 55, + 72, + 294, + 216 + ], + "type": "text", + "content": " space. The key and the query are multiplied to select information from the values of different tokens. With cross attention, T2I models can generate images that adhere to the provided text prompts." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 223, + 167, + 235 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 223, + 167, + 235 + ], + "spans": [ + { + "bbox": [ + 55, + 223, + 167, + 235 + ], + "type": "text", + "content": "2.2. Concept Removals" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 239, + 295, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 239, + 295, + 264 + ], + "spans": [ + { + "bbox": [ + 55, + 239, + 295, + 264 + ], + "type": "text", + "content": "In this subsection, we revisit representative concept removal methods." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 264, + 296, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 264, + 296, + 300 + ], + "spans": [ + { + "bbox": [ + 55, + 264, + 296, + 300 + ], + "type": "text", + "content": "Fine-tuning-based methods. Fine-tuning is the most common framework for concept removals. The objective for the fine-tuning-based algorithms can be summarized as" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 128, + 319, + 295, + 336 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 128, + 319, + 295, + 336 + ], + "spans": [ + { + "bbox": [ + 128, + 319, + 295, + 336 + ], + "type": "interline_equation", + "content": "\\min _ {\\theta} L = L _ {\\mathrm {r m}} + \\lambda L _ {\\mathrm {r e g}}, \\tag {3}", + "image_path": "8ffd54516f5453815903a261e2d1e184bc244a0162d4293f7451cc03fecb4de0.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 343, + 296, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 343, + 296, + 415 + ], + "spans": [ + { + "bbox": [ + 55, + 343, + 296, + 415 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 343, + 296, + 415 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{rm}}" + }, + { + "bbox": [ + 55, + 343, + 296, + 415 + ], + "type": "text", + "content": " is the term for changing the output of unwanted concept, and " + }, + { + "bbox": [ + 55, + 343, + 296, + 415 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{reg}}" + }, + { + "bbox": [ + 55, + 343, + 296, + 415 + ], + "type": "text", + "content": " is the regularization term to ensure that the fine-tuning will not influence other benign concepts and maintain the generation quality. The term " + }, + { + "bbox": [ + 55, + 343, + 296, + 415 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 55, + 343, + 296, + 415 + ], + "type": "text", + "content": " is the parameter to fine-tune. The methods can be solved by two types of solutions:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 416, + 295, + 536 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 55, + 416, + 295, + 476 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 416, + 295, + 476 + ], + "spans": [ + { + "bbox": [ + 55, + 416, + 295, + 476 + ], + "type": "text", + "content": "- Gradient descent. This solution [7, 16, 18, 23] focuses on the final output of " + }, + { + "bbox": [ + 55, + 416, + 295, + 476 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(x_t,y,t)" + }, + { + "bbox": [ + 55, + 416, + 295, + 476 + ], + "type": "text", + "content": ", modifying the model in an end-to-end way. The term " + }, + { + "bbox": [ + 55, + 416, + 295, + 476 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{rm}}" + }, + { + "bbox": [ + 55, + 416, + 295, + 476 + ], + "type": "text", + "content": " changes the output of " + }, + { + "bbox": [ + 55, + 416, + 295, + 476 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(x_t,c_u,t)" + }, + { + "bbox": [ + 55, + 416, + 295, + 476 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 55, + 416, + 295, + 476 + ], + "type": "inline_equation", + "content": "c_{u}" + }, + { + "bbox": [ + 55, + 416, + 295, + 476 + ], + "type": "text", + "content": " is unwanted concepts, while " + }, + { + "bbox": [ + 55, + 416, + 295, + 476 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{reg}}" + }, + { + "bbox": [ + 55, + 416, + 295, + 476 + ], + "type": "text", + "content": " maintains the output of benign concepts " + }, + { + "bbox": [ + 55, + 416, + 295, + 476 + ], + "type": "inline_equation", + "content": "c_b" + }, + { + "bbox": [ + 55, + 416, + 295, + 476 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 476, + 295, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 476, + 295, + 536 + ], + "spans": [ + { + "bbox": [ + 55, + 476, + 295, + 536 + ], + "type": "text", + "content": "- Closed-form solution. Modifying the intermediate states of linear components (such as cross-attention weights [8, 22] and MLP layers [42]) instead of the final output of " + }, + { + "bbox": [ + 55, + 476, + 295, + 536 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(x_t, y_c, t)" + }, + { + "bbox": [ + 55, + 476, + 295, + 536 + ], + "type": "text", + "content": ", can be solved in a closed-form solution. This can accelerate the fine-tuning process remarkably." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 538, + 295, + 644 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 538, + 295, + 644 + ], + "spans": [ + { + "bbox": [ + 55, + 538, + 295, + 644 + ], + "type": "text", + "content": "Inference-time methods. Inference-time methods such as negative prompt [32], SLD [36], and SEGA [3] change the generation algorithm to remove concepts in the inference process, which can skip fine-tuning operations. They estimate the influence of unwanted concepts and remove them in inference. However, although these methods can skip the fine-tuning process, there is also a risk of being disabled in open-sourced models, which may slow down the inference." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 654, + 295, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 295, + 681 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 295, + 681 + ], + "type": "text", + "content": "3. Six-CD: a Comprehensive and Effective Dataset" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": "In this section, we first propose a new dataset that encompasses a comprehensive set of categories of unwanted con" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 72, + 553, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 109 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 109 + ], + "type": "text", + "content": "cepts. Then we discuss the issue of ineffectiveness in general prompts and present the solution to filter the dataset, thereby increasing its effectiveness." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 117, + 514, + 131 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 117, + 514, + 131 + ], + "spans": [ + { + "bbox": [ + 313, + 117, + 514, + 131 + ], + "type": "text", + "content": "3.1. Six Categories of Unwanted Concepts" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 135, + 553, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 135, + 553, + 266 + ], + "spans": [ + { + "bbox": [ + 313, + 135, + 553, + 266 + ], + "type": "text", + "content": "As mentioned in Sec. 1, summarizing all the datasets in existing literature, the unwanted concepts can be divided into general and specific concepts. The general concepts contain harmful concepts and nudity concepts, while the specific concepts contain identities of celebrities, copyrighted characters, objects, and art styles. However, existing literature focuses on only a few categories for evaluation, lacking systematic comparisons among the methods in general and specific concepts. To address this, we propose a new dataset, Six-CD, to evaluate concept removals comprehensively." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 267, + 553, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 267, + 553, + 566 + ], + "spans": [ + { + "bbox": [ + 313, + 267, + 553, + 566 + ], + "type": "text", + "content": "For the general concepts, Six-CD first collects the malicious prompts from four different NSFW resources, which include I2P [36], MMA [43], jatman/stable-diffusion-prompts-stats-full-uncensored (SD-uncensored) [13], and Unsafe Diffusion (UD) [27]. Then, it divides the NSFW data into harmful and nudity concepts. Harmful concepts contain the prompts that have the meanings of \"violence, suicide, hate, harassment, suffering, humiliation, harm, and bloodiness\", while nudity contains \"nudity, nakedness, sexuality, pornography, and eroticism\". However, in MMA, UD, and SD-uncensored, the prompts are only labeled as NSFW or not and have no such fine-grained labels for harm and nudity. Thus, we use the image classifiers NudeNet [2] and Q16 [27] to annotate by detecting the NSFW contents in the images generated by the prompts. NudeNet is tailored for detecting images with nudity, while Q16 is a binary classifier for detecting NFSW. To annotate the fine-grained labels, i.e., harm and nudity, we first use NudeNet to find the prompts of nudity from the collected resources1. Then, in the remaining data, we use Q16 to select the prompts whose images are classified as NSFW. Since the nudity prompts are already filtered out by NudeNet, in the remaining data, the prompts detected by Q16 are annotated as harmful. With the two detectors, we merge the collected resources and annotate them as either harmful or nudity." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 567, + 553, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 567, + 553, + 625 + ], + "spans": [ + { + "bbox": [ + 313, + 567, + 553, + 625 + ], + "type": "text", + "content": "For the specific concepts in Six-CD, instead of directly collecting prompts, we collect concepts and use prompt templates (detailed in Appd. A.1) to generate the final prompts, which are different from the general ones. The concepts are collected as follows:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 627, + 553, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 627, + 553, + 675 + ], + "spans": [ + { + "bbox": [ + 313, + 627, + 553, + 675 + ], + "type": "text", + "content": "- Celebrity. Celebrities can evaluate the ability to remove identity. We select the identities of celebrities from CPDM [24]. To assist the evaluation, we use a celebrity detector, GCD [25]. we choose the celebrities that can be" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 313, + 683, + 553, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 683, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 683, + 553, + 713 + ], + "type": "text", + "content": "1The image is labeled as nudity if it is classified as \"FEMALE/MALE GENITALIA EXPOSED\", \"FEMALE BREAST EXPOSED\", \"ANUS EXPOSED\", or \"BUTTOCKS EXPOSED\" by NudeNet." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28771" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 83, + 72, + 270, + 210 + ], + "blocks": [ + { + "bbox": [ + 83, + 72, + 270, + 210 + ], + "lines": [ + { + "bbox": [ + 83, + 72, + 270, + 210 + ], + "spans": [ + { + "bbox": [ + 83, + 72, + 270, + 210 + ], + "type": "image", + "image_path": "5f3fdef1887df2868e0e5bd0c4c975835ee2f6fd076e25d91d796ce3de310fc8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 223, + 297, + 246 + ], + "lines": [ + { + "bbox": [ + 55, + 223, + 297, + 246 + ], + "spans": [ + { + "bbox": [ + 55, + 223, + 297, + 246 + ], + "type": "text", + "content": "Figure 1. Effectiveness of I2P prompts. Most of the prompts have low effectiveness. Results from [36]." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 63, + 266, + 255, + 278 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 266, + 255, + 278 + ], + "spans": [ + { + "bbox": [ + 63, + 266, + 255, + 278 + ], + "type": "text", + "content": "both generated by SD and recognized by GCD." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 279, + 295, + 411 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 55, + 279, + 295, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 279, + 295, + 338 + ], + "spans": [ + { + "bbox": [ + 55, + 279, + 295, + 338 + ], + "type": "text", + "content": "- Copyrighted characters. The copyrighted characters are important IP resources for companies. We use the character concepts in CPDM and the high-frequency copyrighted characters from FiveThirtyEight Comic Characters Dataset [1]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 338, + 295, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 338, + 295, + 374 + ], + "spans": [ + { + "bbox": [ + 55, + 338, + 295, + 374 + ], + "type": "text", + "content": "- Objects. Objects can be used to evaluate the performance of removal on non-humanoid concepts. We randomly sample from a subset from the classes of ImageNet [5]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 374, + 295, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 374, + 295, + 411 + ], + "spans": [ + { + "bbox": [ + 55, + 374, + 295, + 411 + ], + "type": "text", + "content": "- Art style. The art styles can evaluate the removal performance of a global feature instead of a local feature. We use the art styles from [22]." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 54, + 411, + 297, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 411, + 297, + 590 + ], + "spans": [ + { + "bbox": [ + 54, + 411, + 297, + 590 + ], + "type": "text", + "content": "For the two general categories, we provide 6,326 prompts for harm concept and 4,518 effective prompts for nudity concept. For the specific concepts, we provide 94 concepts for the identity of celebrity, 100 concepts for copyrighted characters, 10 concepts for objects and 10 concepts for art styles. Four templates are used to construct the prompt for each specific concept. Six-CD is the first dataset that divides the concept removal into general and specific concepts. Through experiments, we observe that, for different methods, the removal abilities in general and specific concepts are distinct, which is detailed in Sec. 5.1. Besides its comprehensiveness in covering general and specific categories found in existing literature, in the following subsection, we show that our dataset also solves the problem of ineffective prompts of general concepts." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 599, + 295, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 599, + 295, + 624 + ], + "spans": [ + { + "bbox": [ + 55, + 599, + 295, + 624 + ], + "type": "text", + "content": "3.2. Ineffective Prompts of General Concepts in Existing Datasets" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 629, + 295, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 295, + 689 + ], + "type": "text", + "content": "The prompts for general concepts in existing dataset do not consistently generate malicious content for each random seed according to [36] and our observation below. In this subsection, we discuss this phenomenon in general concepts and identify two potential problems it may introduce." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": "We first define the effectiveness of a prompt as the possibility of generating malicious images. To assess this, we" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 340, + 72, + 529, + 212 + ], + "blocks": [ + { + "bbox": [ + 340, + 72, + 529, + 212 + ], + "lines": [ + { + "bbox": [ + 340, + 72, + 529, + 212 + ], + "spans": [ + { + "bbox": [ + 340, + 72, + 529, + 212 + ], + "type": "image", + "image_path": "c8a8662863c0c0805e0a90f364c6dcc077de6afccb53e12337b83d7f2a04eeeb.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 312, + 223, + 555, + 245 + ], + "lines": [ + { + "bbox": [ + 312, + 223, + 555, + 245 + ], + "spans": [ + { + "bbox": [ + 312, + 223, + 555, + 245 + ], + "type": "text", + "content": "Figure 2. Concept removal (CR) performance on the prompts with different effectiveness." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 316, + 289, + 553, + 337 + ], + "blocks": [ + { + "bbox": [ + 313, + 258, + 555, + 280 + ], + "lines": [ + { + "bbox": [ + 313, + 258, + 555, + 280 + ], + "spans": [ + { + "bbox": [ + 313, + 258, + 555, + 280 + ], + "type": "text", + "content": "Table 2. Percentage of removed unwanted concepts between high effective and all effectiveness. (%)" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 316, + 289, + 553, + 337 + ], + "lines": [ + { + "bbox": [ + 316, + 289, + 553, + 337 + ], + "spans": [ + { + "bbox": [ + 316, + 289, + 553, + 337 + ], + "type": "table", + "html": "
NEGESDSPMSDDFMNUCEEMCIDSLDSEGACorr
Harm40.963.698.924.895.677.698.927.323.70.977
Harm-EF56.767.399.927.897.467.898.638.122.9
Nudity22.317.981.011.293.614.998.314.638.80.984
Nudity-EF11.016.088.94.292.612.397.216.358.4
", + "image_path": "3d75e7d165497c428e56e646266b3c13ac423ba357c8941004686d50460799ce.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_body" + } + ], + "index": 14 + }, + { + "type": "table", + "bbox": [ + 316, + 384, + 553, + 444 + ], + "blocks": [ + { + "bbox": [ + 313, + 343, + 555, + 376 + ], + "lines": [ + { + "bbox": [ + 313, + 343, + 555, + 376 + ], + "spans": [ + { + "bbox": [ + 313, + 343, + 555, + 376 + ], + "type": "text", + "content": "Table 3. Examples of Dual-Version Dataset for specific concepts. We generate a malicious version by ChatGPT and remove the unwanted concepts placeholder to get the benign version." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 316, + 384, + 553, + 444 + ], + "lines": [ + { + "bbox": [ + 316, + 384, + 553, + 444 + ], + "spans": [ + { + "bbox": [ + 316, + 384, + 553, + 444 + ], + "type": "table", + "html": "
CategoryExample of malicious versionExample of benign version
Celeb. & \nCopyrighted{celebrity} is dancing in the rain.dancing in the rain.
Object{object}, football, grass, house, tree, dogfootball, grass, house, tree, dog
Art styleA beautiful snow-covered mountain with \nsunshine lighting it in the style of {art style}A beautiful snow-covered moun-\ntain with sunshine lighting it
", + "image_path": "c4c1148edf8b30c7deedbd4a3186433924a5556d4aba924d2c192d89d457ea99.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "table_body" + } + ], + "index": 16 + }, + { + "bbox": [ + 312, + 460, + 555, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 460, + 555, + 628 + ], + "spans": [ + { + "bbox": [ + 312, + 460, + 555, + 628 + ], + "type": "text", + "content": "generate " + }, + { + "bbox": [ + 312, + 460, + 555, + 628 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 312, + 460, + 555, + 628 + ], + "type": "text", + "content": " images for a prompt using different random seeds. The effectiveness is defined by " + }, + { + "bbox": [ + 312, + 460, + 555, + 628 + ], + "type": "inline_equation", + "content": "n / N" + }, + { + "bbox": [ + 312, + 460, + 555, + 628 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 312, + 460, + 555, + 628 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 312, + 460, + 555, + 628 + ], + "type": "text", + "content": " is the number of malicious images detected. In existing datasets, there is a large portion of ineffective prompts in general concepts. Taking I2P [36] as an example, which is the most well-used dataset for harmful and nudity concepts in existing literature [3, 7, 8, 16, 22, 23, 36, 42], most of the prompts have low effectiveness as shown in Fig. 1. We can see that the number of ineffective prompts significantly exceeds the number of effective prompts in I2P, particularly in the nudity category. The specific concepts usually do not have the problem of low effectiveness because they contain the specific and definitive for the unwanted concepts. Thus, in this subsection, we focus on general concepts." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 312, + 630, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 630, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 312, + 630, + 556, + 715 + ], + "type": "text", + "content": "Limitations. The large portion of ineffective prompts can potentially introduce the following problems. First, evaluating the ineffective prompts is in low efficiency. For example, we test different concept removal methods on SD for nudity concepts in Fig. 2. On the vanilla SD without concept removal (i.e., No CR in Fig. 2), prompts with effectiveness of " + }, + { + "bbox": [ + 312, + 630, + 556, + 715 + ], + "type": "inline_equation", + "content": "1/5" + }, + { + "bbox": [ + 312, + 630, + 556, + 715 + ], + "type": "text", + "content": " generated only about " + }, + { + "bbox": [ + 312, + 630, + 556, + 715 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 312, + 630, + 556, + 715 + ], + "type": "text", + "content": " nudity images." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "28772" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "type": "text", + "content": "It means that " + }, + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "type": "inline_equation", + "content": "80\\%" + }, + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "type": "text", + "content": " of the generated images by these prompts are non-nudity images, while evaluating on non-nudity images is meaningless. Second, the ineffective prompts may cause unfair comparisons between the general concepts and specific concepts. Unlike general concepts, the specific concepts do not have the problem of ineffective prompts. When the model builder has to remove both general and specific concepts, the ineffective prompts may falsely induce the builder to put less weight on the general concepts." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 180, + 296, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 180, + 296, + 346 + ], + "spans": [ + { + "bbox": [ + 56, + 180, + 296, + 346 + ], + "type": "text", + "content": "High-effectiveness subsets. Therefore, in addition to harm and nudity concepts, we provide additional subsets of high-effectiveness prompts and name them as harm- " + }, + { + "bbox": [ + 56, + 180, + 296, + 346 + ], + "type": "inline_equation", + "content": "EF" + }, + { + "bbox": [ + 56, + 180, + 296, + 346 + ], + "type": "text", + "content": " and nudity- " + }, + { + "bbox": [ + 56, + 180, + 296, + 346 + ], + "type": "inline_equation", + "content": "EF" + }, + { + "bbox": [ + 56, + 180, + 296, + 346 + ], + "type": "text", + "content": ". We use the prompts with effectiveness of 4/5 and 5/5 for harm- " + }, + { + "bbox": [ + 56, + 180, + 296, + 346 + ], + "type": "inline_equation", + "content": "EF" + }, + { + "bbox": [ + 56, + 180, + 296, + 346 + ], + "type": "text", + "content": " and nudity- " + }, + { + "bbox": [ + 56, + 180, + 296, + 346 + ], + "type": "inline_equation", + "content": "EF" + }, + { + "bbox": [ + 56, + 180, + 296, + 346 + ], + "type": "text", + "content": ". Harm- " + }, + { + "bbox": [ + 56, + 180, + 296, + 346 + ], + "type": "inline_equation", + "content": "EF" + }, + { + "bbox": [ + 56, + 180, + 296, + 346 + ], + "type": "text", + "content": " contains 991 prompts, and nudity- " + }, + { + "bbox": [ + 56, + 180, + 296, + 346 + ], + "type": "inline_equation", + "content": "EF" + }, + { + "bbox": [ + 56, + 180, + 296, + 346 + ], + "type": "text", + "content": " contains 1,539 prompts. In Table 2 and Fig. 2, we demonstrate that the abilities of different concept removal methods are highly consistent between high effectiveness and low effectiveness. Thus, it is reasonable to evaluate on high-effectiveness prompts in order to reduce computation, which is also more fair. Specifically, in Table 2, we use the percentage of removed unwanted concepts to measure the ability of different methods, which is defined as" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 136, + 354, + 295, + 376 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 354, + 295, + 376 + ], + "spans": [ + { + "bbox": [ + 136, + 354, + 295, + 376 + ], + "type": "interline_equation", + "content": "p _ {\\mathrm {r m}} = 1 - \\frac {n _ {\\text {a f t e r}}}{n _ {\\text {b e f o r e}}}, \\tag {4}", + "image_path": "0e7c1baf6b9ca364a487831cfb793018d37a757676e89d7f3e893bd883d07d87.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 384, + 295, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 384, + 295, + 467 + ], + "spans": [ + { + "bbox": [ + 55, + 384, + 295, + 467 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 384, + 295, + 467 + ], + "type": "inline_equation", + "content": "n_{\\mathrm{before}}" + }, + { + "bbox": [ + 55, + 384, + 295, + 467 + ], + "type": "text", + "content": " is the number of images containing unwanted concepts before concept removal, while " + }, + { + "bbox": [ + 55, + 384, + 295, + 467 + ], + "type": "inline_equation", + "content": "n_{\\mathrm{after}}" + }, + { + "bbox": [ + 55, + 384, + 295, + 467 + ], + "type": "text", + "content": " is the number of images containing unwanted concepts after concept removal. The removal abilities in harm and nudity concepts are very close to harm- " + }, + { + "bbox": [ + 55, + 384, + 295, + 467 + ], + "type": "inline_equation", + "content": "EF" + }, + { + "bbox": [ + 55, + 384, + 295, + 467 + ], + "type": "text", + "content": " and nudity- " + }, + { + "bbox": [ + 55, + 384, + 295, + 467 + ], + "type": "inline_equation", + "content": "EF" + }, + { + "bbox": [ + 55, + 384, + 295, + 467 + ], + "type": "text", + "content": ". The Pearson correlation coefficient is as high as 0.977 (between harm and harm- " + }, + { + "bbox": [ + 55, + 384, + 295, + 467 + ], + "type": "inline_equation", + "content": "EF" + }, + { + "bbox": [ + 55, + 384, + 295, + 467 + ], + "type": "text", + "content": ") and 0.984 (between nudity and nudity- " + }, + { + "bbox": [ + 55, + 384, + 295, + 467 + ], + "type": "inline_equation", + "content": "EF" + }, + { + "bbox": [ + 55, + 384, + 295, + 467 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 467, + 295, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 467, + 295, + 526 + ], + "spans": [ + { + "bbox": [ + 55, + 467, + 295, + 526 + ], + "type": "text", + "content": "In summary, Six-CD provides two choices to evaluate on the general concepts. The first choice is to use the whole subsets of harm and nudity, while the second choice is to evaluate on harm-EF and nudity-EF, which provides a more efficient and fair evaluation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 538, + 194, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 538, + 194, + 552 + ], + "spans": [ + { + "bbox": [ + 55, + 538, + 194, + 552 + ], + "type": "text", + "content": "4. In-prompt Retainability" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 558, + 295, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 295, + 617 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 295, + 617 + ], + "type": "text", + "content": "Concept removals should not hurt the generation ability of the benign contents, which is referred to as retainability. In this section, we show the lack of evaluation for retainability on the benign parts of the prompts containing the unwanted concepts and propose a new metric for this retainability." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 618, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 295, + 714 + ], + "type": "text", + "content": "Lack of in-prompt retainability. Existing literature [24, 45] only considers the retainability of totally benign concepts. The benign prompts do not contain any unwanted concept, and concept removal should not influence the generation of them. Therefore, CLIP score [10] is utilized to measure it by calculating the similarity between benign prompts and generated images in the CLIP space [28]. However, for the prompts containing unwanted concepts," + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 360, + 81, + 433, + 155 + ], + "blocks": [ + { + "bbox": [ + 383, + 71, + 411, + 79 + ], + "lines": [ + { + "bbox": [ + 383, + 71, + 411, + 79 + ], + "spans": [ + { + "bbox": [ + 383, + 71, + 411, + 79 + ], + "type": "text", + "content": "NO CR" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 433, + 71, + 510, + 80 + ], + "lines": [ + { + "bbox": [ + 433, + 71, + 510, + 80 + ], + "spans": [ + { + "bbox": [ + 433, + 71, + 510, + 80 + ], + "type": "text", + "content": "ironman removed" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 360, + 81, + 433, + 155 + ], + "lines": [ + { + "bbox": [ + 360, + 81, + 433, + 155 + ], + "spans": [ + { + "bbox": [ + 360, + 81, + 433, + 155 + ], + "type": "image", + "image_path": "1ad7c80faac9de60e308487ea35a535ffca25bea4ba81bca07217cbc58ad40c2.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 163, + 553, + 197 + ], + "lines": [ + { + "bbox": [ + 313, + 163, + 553, + 197 + ], + "spans": [ + { + "bbox": [ + 313, + 163, + 553, + 197 + ], + "type": "text", + "content": "Figure 3. Concept removal example of the prompt: \"At night, ironman with his armor on is watching fireworks by the lake\". Ironman is removed by SDD." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 435, + 81, + 509, + 155 + ], + "blocks": [ + { + "bbox": [ + 435, + 81, + 509, + 155 + ], + "lines": [ + { + "bbox": [ + 435, + 81, + 509, + 155 + ], + "spans": [ + { + "bbox": [ + 435, + 81, + 509, + 155 + ], + "type": "image", + "image_path": "87b1de8ea1f3d9fe8a7adf39cab04b716b816c699eeb5310c2033f3b2365bc66.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 211, + 555, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 211, + 555, + 366 + ], + "spans": [ + { + "bbox": [ + 313, + 211, + 555, + 366 + ], + "type": "text", + "content": "the retainability to guarantee the generation of the benign part of them is also important, which is called \"in-prompt retainability\" in our benchmark. As shown in Fig. 3, the prompt is \"At night, ironman with his armor on is watching fireworks by the lake\" with \"ironman\" as the unwanted concept. Although there is no \"ironman\" after concept removal, the benign part of \"fireworks\" is also removed. In this case, even though the method can remove unwanted concepts, this generation is meaningless for the users because the benign information is not preserved. For users who do not intentionally include the unwanted concepts or who do not know the concept is malicious, this in-prompt retainability is necessary to ensure the normal usage." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 367, + 554, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 367, + 554, + 545 + ], + "spans": [ + { + "bbox": [ + 313, + 367, + 554, + 545 + ], + "type": "text", + "content": "New metric. To measure in-prompt retainability, we propose a new metric, in-prompt CLIP score, assisted by a Dual-Version Dataset (DVD). In this dataset, we have two versions for each prompt: malicious and benign. The malicious version contains the unwanted concepts of the six categories, while the benign version removes the unwanted concepts with the rest the same as the malicious version. In in-prompt CLIP score, we first generate the images with concept removals using the malicious version and then calculate the cosine similarity of CLIP embeddings between these generated images and the corresponding benign version prompts. Thus, we can measure the similarity between the generated image and the benign part of the prompt. To distinguish it from the original CLIP score, we refer to the original CLIP score as the out-prompt CLIP score." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 546, + 554, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 546, + 554, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 546, + 554, + 689 + ], + "type": "text", + "content": "To construct DVD, for general concepts, we first sample two subsets of prompts from the harmful and nudity categories of Six-CD as the malicious version, then we manually remove the unwanted concepts to get the benign version. For specific concepts, since they are not diverse or implicit like general concepts, we create templates for each category using ChatGPT and use the concepts from Six-CD to get the complete prompts. We show the examples of templates in DVD in Table 3. The entire dataset is in Appd. A.2. To validate the effectiveness of the metric, we conduct human evaluation in Appd. D.1. The results demonstrate that the metric is highly consistent with human preference." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "type": "text", + "content": "In summary, to measure the in-prompt retainability, we propose the in-prompt CLIP score. It leverages a malicious-" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28773" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 109 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 109 + ], + "type": "text", + "content": "version and a benign-version prompt to measure the similarity between the images after concept removal and the benign part of the prompt." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 118, + 138, + 133 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 118, + 138, + 133 + ], + "spans": [ + { + "bbox": [ + 55, + 118, + 138, + 133 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 138, + 295, + 235 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 138, + 295, + 235 + ], + "spans": [ + { + "bbox": [ + 55, + 138, + 295, + 235 + ], + "type": "text", + "content": "In this section, we conduct experiments to benchmark the performance of 10 methods in removing single and multiple concepts and the retainability with both in-prompt and out-prompt CLIP scores. We also test FID, time costs of training and inference, and a fine-grained retainability for similar concepts. Due to space limitations, these experiments are in Appd D. Also, we report the results of SD v1.4 in this section and the results of other models are in Appd D." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 235, + 295, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 235, + 295, + 342 + ], + "spans": [ + { + "bbox": [ + 55, + 235, + 295, + 342 + ], + "type": "text", + "content": "Experimental settings. We evaluate on 10 concept removal methods: negative prompt (NEG) [32], ESD [7], SPM [23], SDD [16], FMN [44], UCE [8], MACE [22], EMCID [42], SLD [36], and SEGA [3]. The details on the baseline settings can be found in Appd. B. The evaluated datasets are Six-CD and DVD, which are detailed in Appd. A. Following the conclusions in Sec. 3.2, we use harm-EF and nudity-EF for all the experiments in this section. All experiments use one GPU (A5000 or A6000)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 342, + 295, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 342, + 295, + 498 + ], + "spans": [ + { + "bbox": [ + 55, + 342, + 295, + 498 + ], + "type": "text", + "content": "Detection metrics. We use the detection rate of Q16 and NudeNet for harmful and nudity concepts. We use the classification accuracy of GCD as the detection rate for celebrity concepts, and we train two classifiers based on ResNet-50 [9] and use the accuracy as the detection rate for copyrighted characters and objects. For the art style concept, the detection is not a simple binary problem. In existing literature [22-24, 44], CLIP score, " + }, + { + "bbox": [ + 55, + 342, + 295, + 498 + ], + "type": "inline_equation", + "content": "S_{\\mathrm{CLIP}}" + }, + { + "bbox": [ + 55, + 342, + 295, + 498 + ], + "type": "text", + "content": ", is used to measure the presence of an art style on a continuous scale, assessing the similarity between generated images and art styles. Following them, we use " + }, + { + "bbox": [ + 55, + 342, + 295, + 498 + ], + "type": "inline_equation", + "content": "S_{\\mathrm{CLIP}}" + }, + { + "bbox": [ + 55, + 342, + 295, + 498 + ], + "type": "text", + "content": " and normalize its range by " + }, + { + "bbox": [ + 55, + 342, + 295, + 498 + ], + "type": "inline_equation", + "content": "\\tilde{S}_{\\mathrm{CLIP}} = (S_{\\mathrm{CLIP}} - \\min(S_{\\mathrm{CLIP}})) / (\\max(S_{\\mathrm{CLIP}}) - \\min(S_{\\mathrm{CLIP}}))" + }, + { + "bbox": [ + 55, + 342, + 295, + 498 + ], + "type": "text", + "content": " for a better comparison with other categories." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 505, + 223, + 517 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 505, + 223, + 517 + ], + "spans": [ + { + "bbox": [ + 55, + 505, + 223, + 517 + ], + "type": "text", + "content": "5.1. Evaluation on Removal Ability" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 522, + 295, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 522, + 295, + 652 + ], + "spans": [ + { + "bbox": [ + 55, + 522, + 295, + 652 + ], + "type": "text", + "content": "In this subsection, we test the removal ability of different methods on Six-CD. In Fig. 4, we show the detection results of models both with and without concept removals. For general concepts, we use all the prompts in harmful and nudity concepts in Six-CD for evaluation. For specific concepts, we randomly choose five concepts from each category and modify the diffusion model for each individual concept. We show the averaged results of 5 concepts of each category in Fig. 4. (The table with error bar for Fig. 4 can be found in Appd. D.2.) We highlight the main observations as follows." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 654, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 295, + 713 + ], + "type": "text", + "content": "Observation 1: General concepts are harder to remove. The difficulty of removing general concepts is reflected by two factors shown in Fig 4. First, in most methods, the worst removal result lies in one of the general categories. For example, ESD and MACE perform well in all" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 553, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 263 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 263 + ], + "type": "text", + "content": "other categories, but they still have around " + }, + { + "bbox": [ + 313, + 72, + 553, + 263 + ], + "type": "inline_equation", + "content": "55\\%" + }, + { + "bbox": [ + 313, + 72, + 553, + 263 + ], + "type": "text", + "content": " harmful images detected. Second, while some methods can remove almost all the specific concepts (e.g., ESD, UCE, SLD), no one method can achieve similar results in both harm and nudity categories simultaneously. For general concepts, on the one hand, they are hard to trigger (i.e. low effectiveness). On the other hand, it is harder to remove them than specific concepts. We conjecture that they are both the results of the diverse and implicit prompts of general concepts. Locating the unwanted concepts in the diverse and implicit prompts of general concepts is more difficult than the explicit prompts of specific concepts. Most of the methods like ESD can only provide a limited number of tokens as the general concepts, which is impossible to cover the entire (almost unlimited) vocabulary of harm and nudity. Thus, it leads to difficulty in removing general concepts." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 265, + 553, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 265, + 553, + 445 + ], + "spans": [ + { + "bbox": [ + 313, + 265, + 553, + 445 + ], + "type": "text", + "content": "Observation 2: Removal ability is consistent within general concepts and within specific categories. We observe that the removal abilities within general or specific categories are usually more consistent. For example, SLD and FMN perform well in specific concepts, but the ability to remove general concepts is worse. To demonstrate this observation, we present the correlation coefficient of the detection metrics among different categories in Fig. 5. From Fig. 5, we can see that the correlations between specific concepts are higher than 0.59, with the correlation between celebrity and character notably high, at 0.94. The correlation between general concepts is also as high as 0.65. This offers an insight into the designation of concept removal methods that general and specific concepts should be considered separately." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 455, + 516, + 467 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 455, + 516, + 467 + ], + "spans": [ + { + "bbox": [ + 313, + 455, + 516, + 467 + ], + "type": "text", + "content": "5.2. Removal Ability on Multiple Concepts" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 473, + 553, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 473, + 553, + 581 + ], + "spans": [ + { + "bbox": [ + 313, + 473, + 553, + 581 + ], + "type": "text", + "content": "In Sec. 5.1, we evaluate the ability to remove a single concept from one model. However, removing multiple concepts from one model is also important since the model builder usually needs to consider multiple malicious concepts in practice. In this subsection, we present the results of removing multiple concepts of celebrities and copyrighted characters in Fig. 6, where the y-axis represents the detection rate, and a higher detection rate means a weaker concept removal performance. We have two observations from Fig. 6." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 582, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 553, + 713 + ], + "type": "text", + "content": "Observation 1: Inference-time methods fail in removing multiple concepts. As shown in Fig. 6, the inference-time methods, SLD and NEG, are the only two methods that have significantly limited removal performance when the concept number is larger than 50, which only reduces the detection rate by " + }, + { + "bbox": [ + 313, + 582, + 553, + 713 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 313, + 582, + 553, + 713 + ], + "type": "text", + "content": " or less. To remove the multiple concepts in inference time, they have to encode the string containing all the concepts in the embeddings of one single prompt. This will exceed the capacity of the text encoder of T2I diffusion models and lead to failed removal. Another inference-time method, SEGA, is not reported in Fig. 6 due" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "28774" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 63, + 549, + 293 + ], + "blocks": [ + { + "bbox": [ + 61, + 63, + 549, + 293 + ], + "lines": [ + { + "bbox": [ + 61, + 63, + 549, + 293 + ], + "spans": [ + { + "bbox": [ + 61, + 63, + 549, + 293 + ], + "type": "image", + "image_path": "7ddc365e0d2702dfb1165ce44bc08ace002f3e00706682b06ee3df244d451943.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 300, + 555, + 335 + ], + "lines": [ + { + "bbox": [ + 54, + 300, + 555, + 335 + ], + "spans": [ + { + "bbox": [ + 54, + 300, + 555, + 335 + ], + "type": "text", + "content": "Figure 4. Removal ability on Six-CD. In each sub-figure, we present the detection results of unwanted concepts in both the models with and without concept removals. The range of all the six values is [0,1]. Smaller values indicate that less unwanted concepts are detected, i.e. better removal. G1 and G2 are harm-EF and nudity-EF concepts, while S1, S2, S3 and S4 are celebrity, character, object and art style." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 83, + 348, + 261, + 495 + ], + "blocks": [ + { + "bbox": [ + 83, + 348, + 261, + 495 + ], + "lines": [ + { + "bbox": [ + 83, + 348, + 261, + 495 + ], + "spans": [ + { + "bbox": [ + 83, + 348, + 261, + 495 + ], + "type": "image", + "image_path": "e6b4f9fd9ba6540935fd33b7f7f0ebc824ab9fcefb3378fe96c1b890c67b027f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 508, + 295, + 531 + ], + "lines": [ + { + "bbox": [ + 55, + 508, + 295, + 531 + ], + "spans": [ + { + "bbox": [ + 55, + 508, + 295, + 531 + ], + "type": "text", + "content": "Figure 5. Correlation coefficient of removal performance on six data categories." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 544, + 295, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 544, + 295, + 581 + ], + "spans": [ + { + "bbox": [ + 55, + 544, + 295, + 581 + ], + "type": "text", + "content": "to its catastrophic inference time when the concept number is increased to 50. This implies that the strategy of processing each concept separately during inference is impractical." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 581, + 295, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 581, + 295, + 665 + ], + "spans": [ + { + "bbox": [ + 55, + 581, + 295, + 665 + ], + "type": "text", + "content": "Observation 2: Closed-form solutions by modifying linear components perform well in removing multiple concepts. The best two methods in Fig. 6a and Fig. 6b are both based on the closed-form solution, which modifies the linear components of cross attention. This is possibly because combining multiple concepts in the linear components is easier than in other non-linear parts." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 672, + 271, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 672, + 271, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 672, + 271, + 685 + ], + "type": "text", + "content": "5.3. In-prompt and Out-prompt Retainability" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": "We evaluate the generation ability on benign concepts using both in-prompt and out-prompt CLIP scores. In Fig. 7, we" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 355, + 555, + 391 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 355, + 555, + 391 + ], + "spans": [ + { + "bbox": [ + 313, + 355, + 555, + 391 + ], + "type": "text", + "content": "use DVD to test the in-prompt CLIP score and a subset with benign prompts from LAION [37] to test out-prompt CLIP score. The observations as follows." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 396, + 555, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 396, + 555, + 601 + ], + "spans": [ + { + "bbox": [ + 313, + 396, + 555, + 601 + ], + "type": "text", + "content": "Observation 1: In-prompt retainability performs worse than out-prompt retainability. As shown in Fig. 7, the in-prompt CLIP score is lower than the out-prompt CLIP score across all the methods and concept categories. This indicates that the ability to generate benign parts in prompts containing malicious concepts is more negatively affected than the ability to generate totally benign prompts. As mentioned in Sec. 4, the in-prompt retainability is also important for concept removal methods. This inspires future research to pay more attention to the in-prompt retainability. Remark. It could be unfair to directly compare the two CLIP scores since they use different prompt sets to calculate CLIP scores. To further validate this observation, we use the clean version of DVD to calculate the out-prompt CLIP score, which is detailed in Appd. D.3. By comparing it with the in-prompt CLIP score in Appd. D.3, we have observed a consistent phenomenon with this subsection." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 605, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 556, + 714 + ], + "type": "text", + "content": "Observation 2: Auxiliary semantic information in MACE (Fig. 7g) may hurt the generation on benign concepts. MACE has the most significant decrease in the generation ability of benign concepts for both in-prompt and out-prompt CLIP scores. We conjecture that this is because it introduces auxiliary semantic information to help locate the unwanted concepts in the image. In other methods, the unwanted concepts are usually located by the semantic understanding ability of the T2I model itself. In contrast," + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28775" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 89, + 72, + 271, + 205 + ], + "blocks": [ + { + "bbox": [ + 89, + 72, + 271, + 205 + ], + "lines": [ + { + "bbox": [ + 89, + 72, + 271, + 205 + ], + "spans": [ + { + "bbox": [ + 89, + 72, + 271, + 205 + ], + "type": "image", + "image_path": "a5d64223235972fafc877ec1758d1b3ae033cac5d1c7d1c262f86b6f2dd43c42.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 159, + 211, + 201, + 220 + ], + "lines": [ + { + "bbox": [ + 159, + 211, + 201, + 220 + ], + "spans": [ + { + "bbox": [ + 159, + 211, + 201, + 220 + ], + "type": "text", + "content": "(a) Celebrity" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 284, + 72, + 520, + 205 + ], + "blocks": [ + { + "bbox": [ + 284, + 72, + 520, + 205 + ], + "lines": [ + { + "bbox": [ + 284, + 72, + 520, + 205 + ], + "spans": [ + { + "bbox": [ + 284, + 72, + 520, + 205 + ], + "type": "image", + "image_path": "f3d1443a7d24269896e363541688c346307daf628affba5589c3860b62b571b2.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 332, + 211, + 419, + 220 + ], + "lines": [ + { + "bbox": [ + 332, + 211, + 419, + 220 + ], + "spans": [ + { + "bbox": [ + 332, + 211, + 419, + 220 + ], + "type": "text", + "content": "(b) Copyrighted characters" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 57, + 248, + 152, + 363 + ], + "blocks": [ + { + "bbox": [ + 57, + 248, + 152, + 363 + ], + "lines": [ + { + "bbox": [ + 57, + 248, + 152, + 363 + ], + "spans": [ + { + "bbox": [ + 57, + 248, + 152, + 363 + ], + "type": "image", + "image_path": "b944d48ca20e7ffa160b2285a843dc8b16df00b09e90ae99db440a22a487f722.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 91, + 365, + 119, + 373 + ], + "lines": [ + { + "bbox": [ + 91, + 365, + 119, + 373 + ], + "spans": [ + { + "bbox": [ + 91, + 365, + 119, + 373 + ], + "type": "text", + "content": "(a) NEG" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 157, + 248, + 252, + 363 + ], + "blocks": [ + { + "bbox": [ + 157, + 248, + 252, + 363 + ], + "lines": [ + { + "bbox": [ + 157, + 248, + 252, + 363 + ], + "spans": [ + { + "bbox": [ + 157, + 248, + 252, + 363 + ], + "type": "image", + "image_path": "19a87723d288a62a5fd2695fa21a6cedc26011f2ab09c4f0fd8f37762bcd97c8.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 191, + 365, + 219, + 373 + ], + "lines": [ + { + "bbox": [ + 191, + 365, + 219, + 373 + ], + "spans": [ + { + "bbox": [ + 191, + 365, + 219, + 373 + ], + "type": "text", + "content": "(b) ESD" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 258, + 249, + 352, + 363 + ], + "blocks": [ + { + "bbox": [ + 258, + 249, + 352, + 363 + ], + "lines": [ + { + "bbox": [ + 258, + 249, + 352, + 363 + ], + "spans": [ + { + "bbox": [ + 258, + 249, + 352, + 363 + ], + "type": "image", + "image_path": "363cde74017eb92a28912f424b314834362cc1a4e939f4ec9ccee86e5b8384e5.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 291, + 365, + 319, + 373 + ], + "lines": [ + { + "bbox": [ + 291, + 365, + 319, + 373 + ], + "spans": [ + { + "bbox": [ + 291, + 365, + 319, + 373 + ], + "type": "text", + "content": "(c) SPM" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 358, + 247, + 452, + 363 + ], + "blocks": [ + { + "bbox": [ + 218, + 227, + 392, + 239 + ], + "lines": [ + { + "bbox": [ + 218, + 227, + 392, + 239 + ], + "spans": [ + { + "bbox": [ + 218, + 227, + 392, + 239 + ], + "type": "text", + "content": "Figure 6. Removal ability on multiple concepts" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 358, + 247, + 452, + 363 + ], + "lines": [ + { + "bbox": [ + 358, + 247, + 452, + 363 + ], + "spans": [ + { + "bbox": [ + 358, + 247, + 452, + 363 + ], + "type": "image", + "image_path": "299db817f9401300a9b45ea440479838891b67830126357154c1089e5305fd65.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 458, + 264, + 553, + 363 + ], + "blocks": [ + { + "bbox": [ + 458, + 264, + 553, + 363 + ], + "lines": [ + { + "bbox": [ + 458, + 264, + 553, + 363 + ], + "spans": [ + { + "bbox": [ + 458, + 264, + 553, + 363 + ], + "type": "image", + "image_path": "c0f38fcbffd8c79370fd6bcee7377c5fd4f36ddc8a9fdfa8cc1a22db194af946.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 491, + 365, + 520, + 373 + ], + "lines": [ + { + "bbox": [ + 491, + 365, + 520, + 373 + ], + "spans": [ + { + "bbox": [ + 491, + 365, + 520, + 373 + ], + "type": "text", + "content": "(e) FMN" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 57, + 375, + 151, + 472 + ], + "blocks": [ + { + "bbox": [ + 57, + 375, + 151, + 472 + ], + "lines": [ + { + "bbox": [ + 57, + 375, + 151, + 472 + ], + "spans": [ + { + "bbox": [ + 57, + 375, + 151, + 472 + ], + "type": "image", + "image_path": "0614252b091d5fe1f58ca4db93471406dab4f59a2668f356a4595a712f209e70.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 91, + 475, + 118, + 483 + ], + "lines": [ + { + "bbox": [ + 91, + 475, + 118, + 483 + ], + "spans": [ + { + "bbox": [ + 91, + 475, + 118, + 483 + ], + "type": "text", + "content": "(f) UCE" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 157, + 375, + 252, + 473 + ], + "blocks": [ + { + "bbox": [ + 157, + 375, + 252, + 473 + ], + "lines": [ + { + "bbox": [ + 157, + 375, + 252, + 473 + ], + "spans": [ + { + "bbox": [ + 157, + 375, + 252, + 473 + ], + "type": "image", + "image_path": "a24fabf24dbbdb08ce3c99fbf769814179a713bdca68eaee51ae78ccd8227391.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 187, + 475, + 222, + 484 + ], + "lines": [ + { + "bbox": [ + 187, + 475, + 222, + 484 + ], + "spans": [ + { + "bbox": [ + 187, + 475, + 222, + 484 + ], + "type": "text", + "content": "(g) MACE" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 258, + 375, + 352, + 473 + ], + "blocks": [ + { + "bbox": [ + 258, + 375, + 352, + 473 + ], + "lines": [ + { + "bbox": [ + 258, + 375, + 352, + 473 + ], + "spans": [ + { + "bbox": [ + 258, + 375, + 352, + 473 + ], + "type": "image", + "image_path": "debd13ddd2b714f3e825e6052eecb916c51bca6ba03b51e56ce971d4f94772a6.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 286, + 475, + 324, + 483 + ], + "lines": [ + { + "bbox": [ + 286, + 475, + 324, + 483 + ], + "spans": [ + { + "bbox": [ + 286, + 475, + 324, + 483 + ], + "type": "text", + "content": "(h) EMCID" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 358, + 375, + 452, + 473 + ], + "blocks": [ + { + "bbox": [ + 391, + 365, + 419, + 373 + ], + "lines": [ + { + "bbox": [ + 391, + 365, + 419, + 373 + ], + "spans": [ + { + "bbox": [ + 391, + 365, + 419, + 373 + ], + "type": "text", + "content": "(d) SDD" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 358, + 375, + 452, + 473 + ], + "lines": [ + { + "bbox": [ + 358, + 375, + 452, + 473 + ], + "spans": [ + { + "bbox": [ + 358, + 375, + 452, + 473 + ], + "type": "image", + "image_path": "95dedda416ff55074e5ee8ab412d29a08ff4cc208807325be0d7db6b110d84f2.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 392, + 475, + 418, + 483 + ], + "lines": [ + { + "bbox": [ + 392, + 475, + 418, + 483 + ], + "spans": [ + { + "bbox": [ + 392, + 475, + 418, + 483 + ], + "type": "text", + "content": "(i) SLD" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 458, + 375, + 553, + 473 + ], + "blocks": [ + { + "bbox": [ + 458, + 375, + 553, + 473 + ], + "lines": [ + { + "bbox": [ + 458, + 375, + 553, + 473 + ], + "spans": [ + { + "bbox": [ + 458, + 375, + 553, + 473 + ], + "type": "image", + "image_path": "1856dd8da1267124d798bf0d2479483fb2f0908c865cf7a27bf7457a69e926ee.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 490, + 475, + 522, + 483 + ], + "lines": [ + { + "bbox": [ + 490, + 475, + 522, + 483 + ], + "spans": [ + { + "bbox": [ + 490, + 475, + 522, + 483 + ], + "type": "text", + "content": "(j) SEGA" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 193, + 494, + 416, + 506 + ], + "lines": [ + { + "bbox": [ + 193, + 494, + 416, + 506 + ], + "spans": [ + { + "bbox": [ + 193, + 494, + 416, + 506 + ], + "type": "text", + "content": "Figure 7. In-prompt retainability vs. out-prompt retainability" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "bbox": [ + 54, + 518, + 295, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 518, + 295, + 615 + ], + "spans": [ + { + "bbox": [ + 54, + 518, + 295, + 615 + ], + "type": "text", + "content": "MACE incorporates the segmentation results of GroundedSAM [17, 21] to locate the image area of unwanted concepts. It focuses on optimizing this area and overlooks other areas of the image. This may lead to worse retainability in other areas. Another piece of evidence is that when we do not include segmentation in removing the art style concepts by MACE (since art style is a global feature), it has good in-prompt and out-prompt CLIP scores." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 55, + 632, + 245, + 643 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 632, + 245, + 643 + ], + "spans": [ + { + "bbox": [ + 55, + 632, + 245, + 643 + ], + "type": "text", + "content": "6. Conclusion and Ethical Statement" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "type": "text", + "content": "In this work, we address the lack of a systematic benchmark for concept removal methods by introducing a comprehensive and effective dataset and a new evaluation metric. Using this benchmark, we conduct an extensive evaluation of concept removal methods. Our experimental observations" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 313, + 519, + 555, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 519, + 555, + 590 + ], + "spans": [ + { + "bbox": [ + 313, + 519, + 555, + 590 + ], + "type": "text", + "content": "provide valuable and practical insights for future research in this field. While our benchmark may have a limitation in the inconvenient detection metric for art styles (using the CLIP score for art styles, instead of the detection rate from a classifier like other categories), it still offers a thorough and practical evaluation across the various method settings." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 313, + 594, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 556, + 714 + ], + "type": "text", + "content": "Ethical statement. This work provides a comprehensive evaluation framework for concept removal, which is targeted to mitigate the potential malicious use of text-to-image diffusion models. Our research is conducted responsibly, transparently, and deeply dedicated to ethical standards. Despite involving the generation of sensitive content such as nudity and violence, it is strictly for research purposes and does not intend to produce or promote inappropriate material. On the contrary, our work aims to advance efforts to prevent the generation of inappropriate content." + } + ] + } + ], + "index": 30 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "28776" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 90, + 296, + 178 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 90, + 296, + 178 + ], + "spans": [ + { + "bbox": [ + 55, + 90, + 296, + 178 + ], + "type": "text", + "content": "Jie Ren, Kangrui Chen, Yingqian Cui, Shenglai Zeng, and Jiliang Tang are supported by the National Science Foundation (NSF) under grant numbers CNS2321416, IIS2212032, IIS2212144, IOS2107215, DUE2234015, CNS2246050, DRL2405483 and IOS2035472, the Army Research Office (ARO) under grant number W911NF-21-1-0198, Amazon Faculty Award, JP Morgan Faculty Award, Meta, Microsoft and SNAP." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 188, + 115, + 200 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 188, + 115, + 200 + ], + "spans": [ + { + "bbox": [ + 56, + 188, + 115, + 200 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 208, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 208, + 296, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 208, + 296, + 251 + ], + "spans": [ + { + "bbox": [ + 61, + 208, + 296, + 251 + ], + "type": "text", + "content": "[1] Fivethirtyeight comic characters dataset. https://www.kaggle.com/datasets/fivethirtyeight/fivethirtyeight - comic - characters - dataset.4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 252, + 296, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 252, + 296, + 273 + ], + "spans": [ + { + "bbox": [ + 62, + 252, + 296, + 273 + ], + "type": "text", + "content": "[2] Nudenet: Lightweight nudity detection. https://github.com/notAI-tech/NudeNet, 2023.3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 274, + 296, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 274, + 296, + 328 + ], + "spans": [ + { + "bbox": [ + 62, + 274, + 296, + 328 + ], + "type": "text", + "content": "[3] Manuel Brack, Felix Friedrich, Dominik Hintersdorf, Lukas Struppek, Patrick Schramowski, and Kristian Kersting. Sega: Instructing text-to-image models using semantic guidance. Advances in Neural Information Processing Systems, 36, 2024. 1, 2, 3, 4, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 330, + 296, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 330, + 296, + 384 + ], + "spans": [ + { + "bbox": [ + 62, + 330, + 296, + 384 + ], + "type": "text", + "content": "[4] Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, et al. Pixart-α: Fast training of diffusion transformer for photorealistic text-to-image synthesis. arXiv preprint arXiv:2310.00426, 2023. 1, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 384, + 296, + 427 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 384, + 296, + 427 + ], + "spans": [ + { + "bbox": [ + 62, + 384, + 296, + 427 + ], + "type": "text", + "content": "[5] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 428, + 296, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 428, + 296, + 460 + ], + "spans": [ + { + "bbox": [ + 62, + 428, + 296, + 460 + ], + "type": "text", + "content": "[6] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 461, + 296, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 461, + 296, + 514 + ], + "spans": [ + { + "bbox": [ + 62, + 461, + 296, + 514 + ], + "type": "text", + "content": "[7] Rohit Gandikota, Joanna Materzynska, Jaden Fiitto-Kaufman, and David Bau. Erasing concepts from diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2426-2436, 2023. 1, 2, 3, 4, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 516, + 296, + 570 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 516, + 296, + 570 + ], + "spans": [ + { + "bbox": [ + 62, + 516, + 296, + 570 + ], + "type": "text", + "content": "[8] Rohit Gandikota, Hadas Orgad, Yonatan Belinkov, Joanna Materzyńska, and David Bau. Unified concept editing in diffusion models. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 5111-5120, 2024. 1, 2, 3, 4, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 571, + 296, + 614 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 571, + 296, + 614 + ], + "spans": [ + { + "bbox": [ + 62, + 571, + 296, + 614 + ], + "type": "text", + "content": "[9] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 615, + 296, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 615, + 296, + 657 + ], + "spans": [ + { + "bbox": [ + 57, + 615, + 296, + 657 + ], + "type": "text", + "content": "[10] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning. arXiv preprint arXiv:2104.08718, 2021. 2, 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 658, + 296, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 658, + 296, + 680 + ], + "spans": [ + { + "bbox": [ + 57, + 658, + 296, + 680 + ], + "type": "text", + "content": "[11] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 681, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 681, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 681, + 296, + 713 + ], + "type": "text", + "content": "[12] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems, 33:6840-6851, 2020. 1" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 126 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 126 + ], + "type": "text", + "content": "[13] Huggingface. jatman/stable-diffusion-prompts-stats-full-uncensored datasets at hugging face. https://huggingface.co/datasets/jatman/stable-diffusion-prompts-stats-full-uncensored?not-for-all-audiences=true. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 129, + 553, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 129, + 553, + 182 + ], + "spans": [ + { + "bbox": [ + 317, + 129, + 553, + 182 + ], + "type": "text", + "content": "[14] Tatum Hunter. Ai porn is easy to make now. for women, that's a nightmare. https://www.washingtonpost.com/technology/2023/02/13/ai-porn-deepfakes-women-consent/, 2023. Accessed: 2024-06-01. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 183, + 517, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 183, + 517, + 194 + ], + "spans": [ + { + "bbox": [ + 316, + 183, + 517, + 194 + ], + "type": "text", + "content": "[15] DeepFloyd IF. https://github.com/deep-floyd/if.3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 195, + 553, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 195, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 317, + 195, + 553, + 239 + ], + "type": "text", + "content": "[16] Sanghyun Kim, Seohyeon Jung, Balhae Kim, Moonseok Choi, Jinwoo Shin, and Juho Lee. Towards safe self-distillation of internet-scale text-to-image diffusion models. arXiv preprint arXiv:2307.05977, 2023. 2, 3, 4, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 239, + 553, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 239, + 553, + 293 + ], + "spans": [ + { + "bbox": [ + 316, + 239, + 553, + 293 + ], + "type": "text", + "content": "[17] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 294, + 555, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 294, + 555, + 348 + ], + "spans": [ + { + "bbox": [ + 316, + 294, + 555, + 348 + ], + "type": "text", + "content": "[18] Nupur Kumari, Bingliang Zhang, Sheng-Yu Wang, Eli Shechtman, Richard Zhang, and Jun-Yan Zhu. Ablating concepts in text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22691-22702, 2023. 1, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 350, + 553, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 350, + 553, + 392 + ], + "spans": [ + { + "bbox": [ + 316, + 350, + 553, + 392 + ], + "type": "text", + "content": "[19] Yuantong Li, Chi-Hua Wang, and Guang Cheng. Online forgetting process for linear regression models. In International Conference on Artificial Intelligence and Statistics, pages 217-225. PMLR, 2021. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 394, + 553, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 394, + 553, + 437 + ], + "spans": [ + { + "bbox": [ + 316, + 394, + 553, + 437 + ], + "type": "text", + "content": "[20] Yaxin Li, Jie Ren, Han Xu, and Hui Liu. Neural style protection: Counteracting unauthorized neural style transfer. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 3966-3975, 2024. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 438, + 553, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 438, + 553, + 491 + ], + "spans": [ + { + "bbox": [ + 316, + 438, + 553, + 491 + ], + "type": "text", + "content": "[21] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023.8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 493, + 553, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 493, + 553, + 535 + ], + "spans": [ + { + "bbox": [ + 316, + 493, + 553, + 535 + ], + "type": "text", + "content": "[22] Shilin Lu, Zilan Wang, Leyang Li, Yanzhu Liu, and Adams Wai-Kin Kong. Mace: Mass concept erasure in diffusion models. arXiv preprint arXiv:2403.06135, 2024. 1, 2, 3, 4, 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 537, + 553, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 537, + 553, + 591 + ], + "spans": [ + { + "bbox": [ + 316, + 537, + 553, + 591 + ], + "type": "text", + "content": "[23] Mengyao Lyu, Yuhong Yang, Haiwen Hong, Hui Chen, Xuan Jin, Yuan He, Hui Xue, Jungong Han, and Guiguang Ding. One-dimensional adapter to rule them all: Concepts, diffusion models and erasing applications. arXiv preprint arXiv:2312.16145, 2023. 2, 3, 4, 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 593, + 553, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 593, + 553, + 646 + ], + "spans": [ + { + "bbox": [ + 316, + 593, + 553, + 646 + ], + "type": "text", + "content": "[24] Rui Ma, Qiang Zhou, Bangjun Xiao, Yizhu Jin, Daquan Zhou, Xiuyu Li, Aishani Singh, Yi Qu, Kurt Keutzer, Xiaodong Xie, et al. A dataset and benchmark for copyright protection from text-to-image diffusion models. arXiv preprint arXiv:2403.12052, 2024. 2, 3, 5, 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 647, + 553, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 690 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 690 + ], + "type": "text", + "content": "[25] Dmitry Voitekh Nick Hasty, Ihor Kroosh and Dmytro Korduban. Giphy's open source celebrity detection deep learning model and code. https://github.com/Giphy/celeb-detection-oss, 2019.3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "text", + "content": "[26] OpenAI. Dall-e 2 preview - risks and limitations. https://github.com/openai/dalle-2-preview/" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28777" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "type": "text", + "content": "blob/main/system-card.md, 2023. Accessed: 2024-06-01. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 95, + 295, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 95, + 295, + 149 + ], + "spans": [ + { + "bbox": [ + 56, + 95, + 295, + 149 + ], + "type": "text", + "content": "[27] Yiting Qu, Xinyue Shen, Xinlei He, Michael Backes, Savvas Zannettou, and Yang Zhang. Unsafe Diffusion: On the Generation of Unsafe Images and Hateful Memes From Text-To-Image Models. In ACM SIGSAC Conference on Computer and Communications Security (CCS). ACM, 2023. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 150, + 294, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 150, + 294, + 215 + ], + "spans": [ + { + "bbox": [ + 56, + 150, + 294, + 215 + ], + "type": "text", + "content": "[28] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 217, + 294, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 217, + 294, + 259 + ], + "spans": [ + { + "bbox": [ + 56, + 217, + 294, + 259 + ], + "type": "text", + "content": "[29] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 261, + 294, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 261, + 294, + 293 + ], + "spans": [ + { + "bbox": [ + 56, + 261, + 294, + 293 + ], + "type": "text", + "content": "[30] Javier Rando, Daniel Paleka, David Lindner, Lennart Heim, and Florian Tramèr. Red-teaming the stable diffusion safety filter. arXiv preprint arXiv:2210.04610, 2022. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 294, + 294, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 294, + 294, + 316 + ], + "spans": [ + { + "bbox": [ + 56, + 294, + 294, + 316 + ], + "type": "text", + "content": "[31] Dana Rao. Responsible innovation in the age of generative ai: Adobe blog, . 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 317, + 250, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 317, + 250, + 327 + ], + "spans": [ + { + "bbox": [ + 56, + 317, + 250, + 327 + ], + "type": "text", + "content": "[32] Dana Rao. Stable diffusion 2.0 release, .1, 3, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 327, + 294, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 327, + 294, + 380 + ], + "spans": [ + { + "bbox": [ + 56, + 327, + 294, + 380 + ], + "type": "text", + "content": "[33] Jie Ren, Han Xu, Pengfei He, Yingqian Cui, Shenglai Zeng, Jiankun Zhang, Hongzhi Wen, Jiayuan Ding, Hui Liu, Yi Chang, et al. Copyright protection in generative ai: A technical perspective. arXiv preprint arXiv:2402.02333, 2024. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 383, + 294, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 383, + 294, + 426 + ], + "spans": [ + { + "bbox": [ + 56, + 383, + 294, + 426 + ], + "type": "text", + "content": "[34] Jie Ren, Zhenwei Dai, Xianfeng Tang, Hui Liu, Jingying Zeng, Zhen Li, Rahul Goutam, Suhang Wang, Yue Xing, and Qi He. A general framework to enhance fine-tuning-based llm unlearning. arXiv preprint arXiv:2502.17823, 2025. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 427, + 295, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 427, + 295, + 482 + ], + "spans": [ + { + "bbox": [ + 56, + 427, + 295, + 482 + ], + "type": "text", + "content": "[35] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 483, + 295, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 483, + 295, + 536 + ], + "spans": [ + { + "bbox": [ + 56, + 483, + 295, + 536 + ], + "type": "text", + "content": "[36] Patrick Schramowski, Manuel Brack, Björn Deiseroth, and Kristian Kersting. Safe latent diffusion: Mitigating inappropriate degeneration in diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22522-22531, 2023. 1, 2, 3, 4, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 537, + 294, + 603 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 537, + 294, + 603 + ], + "spans": [ + { + "bbox": [ + 56, + 537, + 294, + 603 + ], + "type": "text", + "content": "[37] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278–25294, 2022. 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 604, + 294, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 604, + 294, + 657 + ], + "spans": [ + { + "bbox": [ + 56, + 604, + 294, + 657 + ], + "type": "text", + "content": "[38] Shawn Shan, Jenna Cryan, Emily Wenger, Haitao Zheng, Rana Hanocka, and Ben Y Zhao. Glaze: Protecting artists from style mimicry by {Text-to-Image} models. In 32nd USENIX Security Symposium (USENIX Security 23), pages 2187–2204, 2023. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 658, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 294, + 713 + ], + "type": "text", + "content": "[39] Gowthami Somepalli, Vasu Singla, Micah Goldblum, Jonas Geiping, and Tom Goldstein. Diffusion art or digital forgery? investigating data replication in diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6048-6058, 2023. 1" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 329 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 105 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 105 + ], + "type": "text", + "content": "[40] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 107, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 150 + ], + "type": "text", + "content": "[41] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 152, + 553, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 553, + 194 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 553, + 194 + ], + "type": "text", + "content": "[42] Tianwei Xiong, Yue Wu, Enze Xie, Zhenguo Li, and Xihui Liu. Editing massive concepts in text-to-image diffusion models. arXiv preprint arXiv:2403.13807, 2024. 1, 2, 3, 4, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 197, + 553, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 197, + 553, + 229 + ], + "spans": [ + { + "bbox": [ + 316, + 197, + 553, + 229 + ], + "type": "text", + "content": "[43] Yijun Yang, Ruiyuan Gao, Xiaosen Wang, Nan Xu, and Qiang Xu. Mma-diffusion: Multimodal attack on diffusion models. arXiv preprint arXiv:2311.17516, 2023. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 231, + 553, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 231, + 553, + 274 + ], + "spans": [ + { + "bbox": [ + 316, + 231, + 553, + 274 + ], + "type": "text", + "content": "[44] Eric Zhang, Kai Wang, Xingqian Xu, Zhangyang Wang, and Humphrey Shi. Forget-me-not: Learning to forget in text-to-image diffusion models. arXiv preprint arXiv:2303.17591, 2023. 1, 2, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 275, + 553, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 275, + 553, + 329 + ], + "spans": [ + { + "bbox": [ + 316, + 275, + 553, + 329 + ], + "type": "text", + "content": "[45] Yihua Zhang, Yimeng Zhang, Yuguang Yao, Jinghan Jia, Jiancheng Liu, Xiaoming Liu, and Sijia Liu. Unlearncanvas: A stylized image dataset to benchmark machine unlearning for diffusion models. arXiv preprint arXiv:2402.11846, 2024. 2, 5" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28778" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Sketch Down the FLOPs_ Towards Efficient Networks for Human Sketch/d51d8c32-98f5-44d4-9674-ee9bc3d4af26_content_list.json b/2025/Sketch Down the FLOPs_ Towards Efficient Networks for Human Sketch/d51d8c32-98f5-44d4-9674-ee9bc3d4af26_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..7f528be68d12c25b1bca852ba3c841f6a2c5672d --- /dev/null +++ b/2025/Sketch Down the FLOPs_ Towards Efficient Networks for Human Sketch/d51d8c32-98f5-44d4-9674-ee9bc3d4af26_content_list.json @@ -0,0 +1,1473 @@ +[ + { + "type": "text", + "text": "Sketch Down the FLOPs: Towards Efficient Networks for Human Sketch", + "text_level": 1, + "bbox": [ + 129, + 122, + 867, + 143 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Aneeshan Sain1 Subhajit Maity2* Pinaki Nath Chowdhury1 Shubhadeep Koley1 Ayan Kumar Bhunia1 Yi-Zhe Song1", + "bbox": [ + 156, + 164, + 821, + 200 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ SketchX, CVSSP, University of Surrey, United Kingdom.", + "bbox": [ + 264, + 200, + 730, + 219 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Department of Computer Science, University of Central Florida, United States.", + "bbox": [ + 179, + 219, + 813, + 237 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{a.sain, p.chowdhury, s.koley, a.bhunia, y song}@surrey.ac.uk; Subhajit@ucf.edu", + "bbox": [ + 147, + 239, + 844, + 255 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 289, + 326, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "As sketch research has collectively matured over time, its adaptation for at-mass commercialisation emerges on the immediate horizon. Despite an already mature research endeavour for photos, there is no research on the efficient inference specifically designed for sketch data. In this paper, we first demonstrate existing state-of-the-art efficient light-weight models designed for photos do not work on sketches. We then propose two sketch-specific components which work in a plug-n-play manner on any photo efficient network to adapt them to work on sketch data. We specifically chose fine-grained sketch-based image retrieval (FG-SBIR) as a demonstrator as the most recognised sketch problem with immediate commercial value. Technically speaking, we first propose a cross-modal knowledge distillation network to transfer existing photo efficient networks to be compatible with sketch, which brings down number of FLOPs and model parameters by $97.96\\%$ percent and $84.89\\%$ respectively. We then exploit the abstract trait of sketch to introduce a RL-based canvas selector that dynamically adjusts to the abstraction level which further cuts down number of FLOPs by two thirds. The end result is an overall reduction of $99.37\\%$ of FLOPs (from 40.18G to 0.254G) when compared with a full network, while retaining the accuracy $(33.03\\%)$ vs $32.77\\%$ - finally making an efficient network for the sparse sketch data that exhibit even fewer FLOPs than the best photo counterpart.", + "bbox": [ + 88, + 321, + 485, + 715 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 732, + 220, + 748 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "A significant movement in the late computer vision literature has been that of model efficiency. This effort has been largely driven by the on-the-edge deployments of vision models. With great strides made on networks specifically tackling photo data (MobileNet [54], EfficientNet [63], etc), there however has been surprisingly no work targeting human sketches [72]. This is despite the abundance of work specifically addressing sketch data [47] and its unique traits:", + "bbox": [ + 89, + 752, + 483, + 875 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/13cb94a39c9a758370bf172a5b52d260950fde46c6c3b4a2912aa0047ab2c5d4.jpg", + "image_caption": [ + "Figure 1. Our SketchyNetV1 compresses existing heavy FG-SBIR networks to deliver smaller models. Further enhanced via a canvas selector module our SketchyNetV2 model minimises sketch-resolution dynamically to reduce FLOPs." + ], + "image_footnote": [], + "bbox": [ + 516, + 287, + 906, + 417 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "sequential [22], abstract [60], or stroke-wise [10], style-diversity [51], and data-scarcity [6, 20, 43]. This is particularly disappointing for the problem of fine-grained sketch-based image retrieval (FG-SBIR), which is perhaps the single-most studied sketch task [2, 7, 9, 13, 40, 43, 53, 75] that has already matured for at-mass commercial adoption, and is our problem of choice in this paper.", + "bbox": [ + 511, + 492, + 905, + 598 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "One would think that state-of-the-art models designed for photos would work as-is for sketches, treating sketch data as raster images (which sketch is not). As our first contribution, we preempt this thought via a pilot study to prove its false nature. In particular, we borrow MobileNetv2 [54] as a backbone network to construct a typical triplet loss based network for FG-SBIR. Results show a reduction of almost $37\\%$ (relative) in retrieval accuracy when compared with a typical VGG-16 backbone [57]. The focus of this paper is however not to design entirely new sketch-specific efficient networks from scratch – lots of good efficient photo networks already exist. Our goal instead is to adapt existing photo networks to work with sketch, by introducing generic sketch-specific designs that work in a plug-n-play manner.", + "bbox": [ + 511, + 598, + 906, + 810 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We first tackle the problem highlighted in the said pilot study – to make existing photo efficient networks perform on the same level as large networks on FG-SBIR. We argue the reason behind why the likes of MobileNetV2 did not work off-the-shelf is that the limited set of parameters could not accommodate the fine-grained nature of sketches (i.e.,", + "bbox": [ + 511, + 810, + 908, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Work done as an intern at SketchX before joining UCF.", + "bbox": [ + 107, + 886, + 408, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "28383", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "sketches are highly expressive in capturing fine-grained visual details [72]) for sketch and photo matching. For that, we introduce a knowledge distillation (KD) paradigm [48] to adapt this sketch trait onto efficient photo networks. The core idea is to induce strong semantic knowledge between a sketch and its photo, in the embedding space of a student, from a deep and high performing teacher. This particularly helps the efficient model to achieve on par performance with the deep variant without having a large parameter size. However, ours being a cross-modal retrieval, naively adapting standard KD paradigms involving logit distillation [3] would be sub-optimal. We therefore aim to preserve cross-modal pairwise distances among instances, and in turn the structural consistency of the teacher's discriminative embedding space in the student. The result is already an efficient FG-SBIR model (which we term SketchyNetV1) that reduces the number of FLOPs from 40.18G (VGG-16 based model[51]) to 0.833G (our MobileNetV2-based) – a roughly 48 times reduction (Fig. 1 left).", + "bbox": [ + 89, + 90, + 480, + 378 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This abstract nature of human sketch [16] can also be exploited, to further reduce the FLOPs. The hypothesis is that since sketches are essentially abstract depictions of images, they should carry a similar level of semantic information even when rendered at much lower resolutions, and that more abstract a sketch is the more information compact it is (i.e., can sustain even lower resolutions). For that, we conducted another pilot study (Sec. 3), where we show unlike photos, sketches can sustain retrieval accuracy with much lower resolutions (4% vs 15% at $32 \\times 32$ ), and 30% of those query sketches achieving perfect retrieval at full-resolution, when downsized to $32 \\times 32$ can still retrieve accurately (Fig. 2 right). This motivated us to design a canvas selector that can choose the right resolution on-the-fly to render a sketch before feeding to the network. This is thanks to another unique trait of sketch data - they come in vector format and can be freely rendered to any resolution without extra overhead. Accomplishing this however is nontrivial since this selection process of canvas-size involves a discrete decision, thus invoking non-differentiability. We thus leverage reinforcement learning [9] to bypass these issues and train an abstraction-aware canvas-size selector as a policy network for our model, which estimates the optimal canvas-size required for the input sketch. The FG-SBIR model, acting as a critic, calculates reward for the canvas selector, balancing between accuracy and FLOPs, that trains the policy network. With this further trait of sketches addressed, our final model (SketchyNetV2) was able to further reduce FLOPs by a factor of 3.28 (from 0.833G to 0.254G), while retaining a retrieval accuracy similar to the full model (33.03 vs 32.77), exactly as one would imagine sketch efficient networks to entail (sketches being sparse).", + "bbox": [ + 89, + 382, + 482, + 864 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our contributions are: (a) the first investigation towards efficient sketch networks from existing photo", + "bbox": [ + 89, + 869, + 482, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ones (b) a knowledge distillation framework specifically designed to cater to the fine-grained nature of sketches, (c) a canvas-size selector that caters to the abstract nature of sketches, which further reduces FLOPs by two-thirds. Extensive experiments show our method to perform at par with existing state-of-the-arts, while being superiorly efficient.", + "bbox": [ + 511, + 90, + 903, + 181 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 511, + 196, + 653, + 210 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Fine-grained SBIR: Compared to category-level SBIR [17, 19, 36, 52], Fine-Grained SBIR (FG-SBIR) focuses on retrieving a single photo instance from a category-specific gallery based on a query sketch. Introduced as a deep triplet-ranking Siamese network [75], FG-SBIR has evolved with attention-based modules [59], textual tags [58], hybrid cross-domain generation [42], hierarchical coattention [50], and various pre-training strategies like mixed modal jigsaw solving [44], reinforcement learning [6], and transformer-based knowledge distillation [53]. Addressing sketch-specific traits like style diversity [51], data scarcity [7], and sketch stroke redundancy [9], improvements have been made for enhanced retrieval [53]. Recently it was extended to scene-level (retrieve a scene-image, given a scene-sketch), employing cross-modal region associativity [13] and enhanced using text-queries [46]. However, sketch-specific computational efficiency has largely been ignored - which we address here in this paper.", + "bbox": [ + 511, + 218, + 903, + 491 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Towards Computational Efficiency: Addressing large-scale deployment, research emerging in computational efficiency [1], include network pruning [37], quantisation [21], binarisation [18], knowledge distillation [25], input-size optimisation [62], or their combinations [1]. Network pruning [23, 68] involves discovering and dropping low impact weights or channels of a pre-trained large network, while retaining its original accuracy. Contrarily, quantisation directly reduces the bit-width of parameter-values [21] and gradients [77], by replacing floating-point computations with faster and cheaper low-precision fixed-point numbers [34]. Its extreme version, binarisation [18], binarises weights and activations. Knowledge Distillation aims at transferring knowledge of a large pre-trained teacher network to a smaller student for cost-effective deployment. Existing techniques leverage output logits [3], hidden layers [48], attention-maps [76], or neuron selectivity pattern [28] of pre-trained teachers for the same. Self-distillation [4] employs the same network for both student and teacher models, to further reduce compute. Input-resolution optimisation on the other hand involves replacing a high-resolution input image with its down-scaled low-resolution version [62], to reduce overall compute [73], while retaining accuracy. Relevant works explored dynamic usage of high-resolution data [64] or patch proposals for strategic image-cropping [69], where the learnable resizer [62] operates on the original image resolution. All such", + "bbox": [ + 511, + 493, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "28384", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "table", + "img_path": "images/cace9bb88c989feb6f296be44349025fffc9bb2367e18414d04e5c248581cd88.jpg", + "table_caption": [ + "Table 1. Performance & computational cost for various networks" + ], + "table_footnote": [], + "table_body": "
ModelFLOPSParametersTime (ms)Acc@1
VGG-19 [57]51.06G20.02M10133.63
VGG-16 [57]40.18G14.71M08433.03
ResNet-101 [24]20.50G42.50M15521.77
DenseNet-201 [26]11.40G18.09M33523.62
ResNet-50 [24]10.76G23.51M08621.32
InceptionV3 [75]07.72G21.79M13424.27
ResNet-18 [24]4.76G11.18M03620.72
EfficientNet [63]1.05G04.01M11222.47
MobileNetV2 [54]0.83G02.22M07020.85
", + "bbox": [ + 107, + 95, + 464, + 218 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "methods however cater to compressing image-specific models, which if used off-the-shelf for FG-SBIR would fail to address the sparse manner in which sketches are represented [8]. We thus propose a sketch-specific light-weight FG-SBIR model, tackling sparsity in sketches.", + "bbox": [ + 89, + 223, + 482, + 297 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Reinforcement Learning in Vision (RL): Applying RL [29] to computer vision problems [33, 67] is commonplace of late. Whenever quantifying the goodness of the network's state is non-differentiable, like selecting which regions to crop in an image [31], RL comes in handy. In sketch community, RL has found use in retrieval [6, 7], modelling sketch abstraction [39, 40], designing competitive sketching agents [5] and as a stroke selector [9]. Here, we leverage RL for optimising a canvas-size selector for a computationally efficient retrieval model.", + "bbox": [ + 89, + 297, + 483, + 450 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Pilot Study", + "text_level": 1, + "bbox": [ + 89, + 457, + 209, + 473 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Backbone Network for FG-SBIR: Amongst popular backbones used in recent FG-SBIR literature [6, 7, 50, 51], are VGG-16 [30] and Inception-V3 [75]. Others like ResNet18 [24] and EfficientNet [63], though cheaper and prevalent in vision [24] are rarely used in FG-SBIR. Exploring their potential in FG-SBIR, we train a baseline FG-SBIR model, plugging in popular ones as backbones for our study. Table 1 reports their parameter-count, FLOPs, time for feature-extraction per sample and Top-1 accuracy (Sec. 5) on standard QMUL-ShoeV2 dataset [75]. Despite scoring more than VGG-16 on image-classification [65], ResNet50 [24] performs poorly – indicating that modelling fine-grained sketch-photo association needs networks with higher FLOPs. Despite delivering better accuracies, heavier networks cost far more resources. We thus aim for a lighter model that achieves accuracy like a larger one.", + "bbox": [ + 89, + 474, + 482, + 717 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Dynamic Sketch-canvas-size: Unlike photos that hold pixel-dense information, sketches are sparse black and white lines. This begs the question, if a sketch rendered at a higher resolution with added computational burden would convey any extra semantic information than at a lower one. To answer this, we focus on vector-format of sketches [7], where every sketch vector $s_v$ can be characterised as a sequence of points $(v_1, v_2, \\dots, v_{\\mathrm{T}})$ , and rendered to any canvas-size $(c)$ as a raster-image $(s_r)$ via rasterisation $\\mathcal{R}^c(\\cdot): s_v \\to s_r^c$ where $s_r^c \\in \\mathbb{R}^{c \\times c \\times 3}$ . A single baseline FG-SBIR model (VGG-16) is trained and tested on full-resolution images, and sketches rendered at varying", + "bbox": [ + 89, + 719, + 482, + 901 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/2e9054e4b9b734386733cfe540dbcbb670b37b4422ac8e6b687c233bba54ab3a.jpg", + "image_caption": [ + "Figure 2. Pilot study for varying canvas-size (see text above)." + ], + "image_footnote": [], + "bbox": [ + 542, + 92, + 715, + 181 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/5b15e5befa59daae27edb06ff6dac5cfad0663005ac286bdaf13107c29d27f45.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 720, + 90, + 880, + 184 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "canvas-sizes from $C = \\{32 \\times 32, \\dots, 256 \\times 256\\}$ . Comparing sparse sketches versus photos, we train an equivalent fine-grained image-based image retrieval (FG-IBIR) where only the query against a positive photo is constructed using random augmentation [11] and resized to corresponding canvas-size $c_i$ . Gallery photo-features for both are kept pre-computed. While FG-IBIR accuracy falls rapidly (Fig. 2 left), FG-SBIR stays relatively stable against decreasing canvas-sizes, as photos (unlike sketches) containing pixel-dense perfect information, lose a lot of it while down-scaling. Furthermore, positive accuracy of FG-SBIR at $32 \\times 32$ , shows some sketches to hold sufficient semantic information for retrieval even at minimal canvas-size. Fig. 2 (right) shows overall QMUL-ShoeV2[75] statistics where each bar represents the smallest canvas-size, at which bar-height percentage of sketches achieve perfect retrieval (taking total sketches achieving perfect retrieval at $256 \\times 256$ as $100\\%$ ). This infers that every sketch has its own optimal canvas, and fixing it (Fig. 2 left), for the entire dataset would always be sub-optimal. Moreover, the optimal canvas-size for a sketch, depends on its extent of sparsity, for which neither hard-ground-truth exists, nor is it knowable from the user's end. Finally, brute force discovery at every size being infeasible during inference, motivates us to design a learnable adaptive canvas-size selector for sketches.", + "bbox": [ + 511, + 208, + 906, + 585 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Methodology", + "text_level": 1, + "bbox": [ + 511, + 590, + 648, + 607 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Overview: Our pilot study motivates us to extend a standard FG-SBIR framework towards computational efficiency by first introducing SketchyNetV1 via knowledge distillation, and then taking it further to SketchyNetV2 via dynamic sketch canvas-size selection. Importantly, these are done without hurting the original performance of large-scale slower baseline model. In particular, there are two phases of improvement for computational efficiency. Firstly, we train a teacher model $F_{T}$ as a standard baseline FG-SBIR model (Sec. 3). Once trained, its rich semantic knowledge is transferred to a smaller student network $F_{st}$ via a knowledge distillation (KD) paradigm, which we name as SketchyNetV1. Next, we train a canvas-size selector that adaptively decides optimal canvas-size or resolution per sketch for SketchyNetV2 that would be sufficient to preserve the retrieval performance while reducing the number of FLOP operations in feature extraction significantly.", + "bbox": [ + 511, + 612, + 906, + 868 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Baseline FG-SBIR model: Keeping aside complex pretraining [44] or joint-training [7], we focus on a simple", + "bbox": [ + 511, + 869, + 906, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "28385", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/cacf7a15e1c70f3a7b612e9ee79238275db7e1d710603e601cd6b0d7674ca74a.jpg", + "image_caption": [ + "Figure 3. Two stage framework. (Left) SketchyNetV1: A smaller student network $(F_{st})$ is trained from a larger pre-trained teacher $(F_T)$ via KD (§4.1). (Right) We progress to SketchyNetV2 by training a canvas-size selector directed by objectives of increasing performance and reducing compute. It takes sketch as a vector $(s_v)$ and aims to predict an optimal canvas-size at which the sketch is rasterised $(s_r)$ . $s_r$ when processed by $F_{st}$ minimises overall FLOPS, while retaining accuracy of corresponding full-resolution sketch-image." + ], + "image_footnote": [], + "bbox": [ + 96, + 89, + 408, + 234 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/1989c4a8b181a9d84247e7176b167e34eaa9155836d5351d6ec01a09df533f8f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 411, + 90, + 893, + 233 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "triple-branch state-of-the-art Siamese network [75] which remains a strong baseline so far [7, 50, 51]. A VGG-16 [30] pre-trained on ImageNet [49] acts as the shared backbone for each branch. Given an input image $I$ , a $d$ dimensional feature embedding is extracted from the backbone network $\\mathcal{F}$ using global average pooling followed by $l_{2}$ normalisation as $f_{I} \\in \\mathbb{R}^{d}$ . The model is trained on triplets comprising features of an anchor sketch $(f_{s})$ , its matching positive $(f_{p})$ photo and a random non-matching negative photo $(f_{n})$ , via a triplet loss [70]. Triplet loss $(\\mathcal{L}_{\\mathrm{Tri}})$ aims to minimise the distance between sketch and positive photo, while maximising the same from its negative (n). With $\\delta(a, b) = ||a - b||^2$ , as a distance function and a margin $m > 0$ , we have:", + "bbox": [ + 88, + 308, + 485, + 503 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {T r i}} = \\max \\left\\{0, m + \\delta \\left(f _ {s}, f _ {p}\\right) - \\delta \\left(f _ {s}, f _ {n}\\right) \\right\\} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 153, + 507, + 482, + 522 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. SketchyNetV1: Smaller Model via KD", + "text_level": 1, + "bbox": [ + 89, + 534, + 421, + 550 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Based on our pilot study (Sec. 3), if we naively train from an ImageNet pre-trained standard smaller network (e.g. MobileNetV2 [54] - usually popular for edge devices), with simple supervised triplet loss (Eq.1) it can hardly reach the accuracy (Table 1) obtained by a large network like VGG-16. Therefore apart from training via a supervised loss on the sketch-positive-negative triplet $(\\mathcal{L}_{\\mathrm{Tri}}^{st})$ as in Eq.1), we aim to use the pre-trained larger teacher (e.g. VGG-16) for additional supervision. Accordingly, we adhere to Knowledge Distillation (KD) paradigm [48], a conventional strategy for model compression, to deliver a light-weight student model from the larger teacher network [48]. Applying KD to our FG-SBIR paradigm is however non-trivial. Unlike traditional KD methods, which often involve logit distillation for classification tasks, our scenario involves cross-modal retrieval. Here, the output is a continuous $d$ -dimensional feature within a joint-embedding space. Furthermore, straightforward regression between teacher and student features for sketch and photo branches may encounter compatibility issues, given the disparity in the embedding spaces. Addressing this, if the dimensions of teacher and student embeddings differ, an additional feature transformation layer is necessary for alignment [48]. Overcoming these issues,", + "bbox": [ + 91, + 553, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "we aim to distil the inter-feature distances from respective teacher networks to the student thus preserving structural orientation of features in the teacher's embedding space to that in the student, during KD (Fig. 3 (left)).", + "bbox": [ + 511, + 308, + 906, + 368 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given features of sketch-positive-negative triplet of teacher $(f_s^T, f_p^T, f_n^T)$ , we calculate the inter-feature Euclidean distances in the output representation space of teacher $(F_T)$ using $\\delta(\\cdot, \\cdot)$ (Eq.1), as $d_{sp}^T = \\delta(f_s^T, f_p^T)$ ; $d_{sn}^T = \\delta(f_s^T, f_n)$ and $d_{pn}^T = \\delta(f_p^T, f_n^T)$ . Passing the same samples through the student, we similarly obtain $d_{sp}^{st}$ , $d_{sn}^{st}$ and $d_{pn}^{st}$ using $(f_s^{st}, f_p^{st}, f_n^{st})$ . A smooth $l_1$ loss distils the $l_2$ distances from the teacher to the student. With $\\mathcal{L}_{\\delta}$ as a Huber-loss [27], relational distillation loss for sketch-photo pair between teacher and student is given as,", + "bbox": [ + 511, + 369, + 908, + 521 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {R K D}} ^ {s p} = \\mathcal {L} _ {\\delta} \\left(d _ {s p} ^ {T}, d _ {s p} ^ {s t}\\right) \\quad \\text {w h e r e ,}\n$$\n", + "text_format": "latex", + "bbox": [ + 583, + 523, + 772, + 541 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\delta} (a, b) = \\left\\{ \\begin{array}{l l} \\frac {1}{2} (a - b) ^ {2} & \\text {i f} | a - b | < \\beta \\\\ \\beta (| a - b | - \\frac {1}{2} \\beta) & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 545, + 903, + 580 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Similarly, we obtain $\\mathcal{L}_{\\mathrm{RKD}}^{sn}$ and $\\mathcal{L}_{\\mathrm{RKD}}^{pn}$ , to compute distillation loss, $\\mathcal{L}_{\\mathrm{RKD}} = \\mathcal{L}_{\\mathrm{RKD}}^{sp} + \\mathcal{L}_{\\mathrm{RKD}}^{sn} + \\mathcal{L}_{\\mathrm{RKD}}^{pn}$ . With hyperparameter $\\lambda$ , overall student-training objective becomes:", + "bbox": [ + 511, + 585, + 906, + 628 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {t r n}} ^ {s t} = \\lambda \\mathcal {L} _ {\\mathrm {T r i}} + (1 - \\lambda) \\mathcal {L} _ {\\mathrm {R K D}} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 620, + 635, + 903, + 650 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The student trains on varying sketch-canvas-sizes $(C)$ , using their respective full resolution for teacher, to ensure scale-invariance as $\\mathcal{L}_{\\mathrm{tm}}^{st*} = \\frac{1}{4}\\sum_{i=1}^{4}\\mathcal{L}_{\\mathrm{tm}}^{st(c_i)}$ . However, as optimal canvas-size varies across sketches (Fig. 2), we need a learnable canvas-size selector, dynamically predicting the optimal size for each sketch.", + "bbox": [ + 511, + 652, + 906, + 724 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. SketchyNetV2: Adaptive canvas-size Selector", + "text_level": 1, + "bbox": [ + 511, + 729, + 897, + 744 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Overview: Understanding that an input at a lower resolution, would incur lesser FLOPs on evaluation than a higher one, we upgrade SketchyNetV1 ( $F_{st}$ ) to a more computationally efficient SketchyNetV2 by introducing an adaptive canvas-size selector. It is designed in the vector modality as it is much cheaper [71] than its raster-image counterpart, and can dynamically encode the varying abstraction level of sketches. Furthermore, predicting the optimal canvas-size from vector modality, would make the corresponding rasterisation operation $\\mathcal{R}^c(\\cdot)$ at lower sizes cheaper, besides reducing FLOPs in feature extraction. Taking sketch vector $s_v$ as input, $\\psi_C$ predicts the probability $p(c|s_v)$ of choosing one of $K$", + "bbox": [ + 511, + 748, + 908, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "28386", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "discreet canvas-sizes from a set of $C = \\{32 \\times 32, \\dots, 256 \\times 256\\}$ . $s_v$ is then rasterised to $s_r^c \\in \\mathbb{R}^{c \\times c \\times 3}$ with the optimal canvas-size $c$ , as $\\mathcal{R}^c(\\cdot) : s_v \\to s_r^c$ , and fed to $F_{st}$ for retrieval.", + "bbox": [ + 89, + 90, + 480, + 133 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Selecting an optimal canvas-size however is an ill-posed problem. Firstly, there are no explicit labels representing the optimal canvas-size for a sketch. Secondly, annotating the optimal canvas-size for the whole dataset via brute-force iteration is computationally impractical. We therefore use the pre-trained $F_{st}$ as a critic to guide the learning of sketch canvas-size selector via Reinforcement Learning [29] as rasterisation is a non-differentiable operation. There are two major objectives here: a) retain the original resolution accuracy and b) encourage the use of the lower resolution to improve the computational efficiency.", + "bbox": [ + 88, + 133, + 482, + 272 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Model: Any sequential network from RNN [17] or Transformer [35] families can be used to encode the vector-sketch $s_v$ . To recap, sketch-vector $(s_v)$ represents a sequence of points $[v_1, v_2, \\dots, v_{\\mathrm{T}}]$ , where $v_t = (x_t, y_t, q_t^1, q_t^2, q_t^3) \\in \\mathbb{R}^{T \\times 5}$ ; $\\mathrm{T}$ is the sequence length, $(x_t, y_t)$ denotes the absolute coordinates in a normalised $c \\times c$ canvas, while the last three represent penstates [22]. Furthermore, the complexity of canvas-size selector being dependent on sequence-length (T) varying across sketches, we use Douglas Peucker Algorithm [66] to limit sequence-length across all sketch-samples at $\\mathrm{T}_{max}$ without losing visual representation. We feed $v_t$ at every time step to a simple 1-layer GRU network with $\\mathbb{R}^{d_v}$ hidden states and take the final hidden state as the encoded latent representation $(f_{s^{\\mathrm{T}}} \\in \\mathbb{R}^{d_v})$ . Next, we apply a linear layer $(\\gamma)$ to get $p(c|s_v) = \\text{softmax}(W_\\gamma f_{s^{\\mathrm{T}}} + b_\\gamma)$ , where $W_\\gamma \\in \\mathbb{R}^{d_v \\times K}$ and $b \\in \\mathbb{R}^K$ , as shown in Fig. 3 (right). Given $p(c|s_v) \\in \\mathbb{R}^K$ over $K$ different canvas-sizes for every $s_v$ , canvas-size $c$ can be sampled from categorical distribution as $c_{pred} \\sim \\text{categorical}([p(c_1|s_v), \\dots, p(c_K|s_v)])$ . Accordingly, $s_v$ is rasterised to $s_r$ at canvas-size $c_{pred}$ , which is then fed to $F_{st}$ . In particular, given a sketch $s_v$ , canvas-size selector $\\psi_C$ acting as a policy network, takes action of selecting the canvas-size $c_{pred}$ from action space $C$ , and $\\psi_C$ is optimised over a reward calculated by the SketchyNetV1 model $(F_{st})$ acting as the critic.", + "bbox": [ + 91, + 275, + 483, + 593 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Reward Design: The objectives of our canvas selector are to choose an optimal canvas-size such that: (i) accuracy is retained (ii) overall FLOPs is minimised. We therefore design our reward from two perspectives. Now as $F_{st}$ is fixed, we pre-compute the features of all gallery photos. During training, we only extract the feature of rasterised sketch and calculate its corresponding rank using the pre-computed photo-features. Furthermore, taking one sketch we pre-compute FLOPs for every canvas-size in $C$ . From accuracy perspective, the aim is to select the optimal canvas-size $c_{\\mathrm{opt}}$ from $C$ for each $s_v$ such that the rendered $s_r$ can retrieve the matching photo $p$ at numerically lowest rank $r$ (best $r = 1$ ). Following conventional norm of reward maximisation, we define accuracy reward $(R_{\\mathrm{acc}})$ as weighted (by hyper-parameters $\\lambda_r$ , $\\lambda_{\\mathrm{Tri}}$ ) summation of inverse of the rank (r) and negative triplet loss (following Eq.1) as:", + "bbox": [ + 89, + 595, + 482, + 801 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nR _ {\\mathrm {a c c}} = \\lambda_ {r} (1 / r) + \\lambda_ {\\mathrm {T r i}} (- \\mathcal {L} _ {\\mathrm {T r i}}) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 196, + 804, + 480, + 818 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "From the compute perspective, the selection criterion should reward choosing a lower canvas-size. Moreover, higher performance being naturally inclined towards a higher-canvas-size thus increasing overall compute, we need an objective dedicated towards reducing computational cost. We thus propose a FLOPs constraint regularisation to guide the learning of canvas selector as:", + "bbox": [ + 89, + 816, + 482, + 900 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {F}} = \\frac {\\sum_ {j = 1} ^ {K} \\left(q _ {j} \\cdot \\eta_ {j}\\right)}{q _ {\\max } - q _ {\\min }} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 642, + 87, + 903, + 121 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where, $\\eta_{i} = p(c_{i}|s_{v})$ and $q_{j}$ is the pre-computed FLOPs value for the j-th canvas-size with maximum at $q_{max}$ and minimum at $q_{min}$ . Treating $(-L_{\\mathrm{F}})$ as a reward from the compute perspective, we define our compute reward as $R_{\\mathrm{comp}} = (-L_{\\mathrm{F}})$ . Taking $\\lambda_{\\mathrm{F}}$ as a balancing hyper-parameter, we combine both perspectives as our total reward:", + "bbox": [ + 511, + 126, + 903, + 205 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nR _ {\\mathrm {T o t}} = \\lambda_ {\\mathrm {F}} R _ {\\mathrm {c o m p}} + (1 - \\lambda_ {\\mathrm {F}}) R _ {\\mathrm {a c c}} \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 612, + 205, + 903, + 220 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Optimisation Objective: Finally, we train the canvas-size selector using popular Policy Gradient [61] method which is both simple to implement and requires less hyper-parameter tuning than other state-of-the-art alternatives [56]. If $p(c|s_v^i)$ be the probability of selecting a specific canvas-size for $i^{th}$ sketch sample $s_v^i$ and the corresponding reward be $R_{\\mathrm{tot}}^i$ , the objective function for policy network (canvas-size selector) over a batch of size $B$ becomes:", + "bbox": [ + 511, + 226, + 903, + 321 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {P G} (\\theta) = - \\frac {1}{B} \\sum_ {i = 1} ^ {B} \\log p \\left(c \\mid s _ {v} ^ {i}\\right) \\cdot R _ {\\text {T o t}} ^ {i} \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 593, + 325, + 903, + 362 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 513, + 359, + 645, + 376 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets: We evaluated on the following publicly available datasets having fine-grained sketch-photo association. QMUL-ShoeV2 [75] and QMUL-ChairV2 [75] contains 6730/2000 and 1800/400 sketches/photos respectively. Keeping 679/200 (525/100) sketches/photos for evaluation from ShoeV2 (ChairV2) respectively we use the rest for training. Sketchy [55] contains 125 categories with 100 photos each, having around 5 sketches per photo. While, training uses a standard train-test split [55] of 9:1, during inference we construct a challenging gallery using photos across one category for retrieval. For scene-level FG-SBIR evaluation we use FS-COCO [15] which includes 10,000 unique sketch-photo pairs with a 7:3 train/test split.", + "bbox": [ + 511, + 381, + 906, + 547 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation Details: Considering a standard large FG-SBIR network as a teacher, our proposed student uses a cheaper backbone feature extractor MobileNetV2 [54], a standard for mobile end-devices. Margin is set to $m = 0.2$ for triplet loss (Secs. 3 and 4.1). We use an Adam optimiser at a learning rate of 0.0001 and a batchsize of 16 for 200 epochs. For practicality, we chose discrete canvas-sizes ( $C$ ) instead of a continuous action space, where $C = \\{32 \\times 32, 64 \\times 64, 128 \\times 128, 256 \\times 256\\}$ . Canvas-size selector ( $\\psi_C$ ) is modelled using a simple GRU network, with embedding size 128, and trained via reinforcement learning with a learning rate of 0.0001 and a batchsize of 32 for 500 epochs, keeping $T_{max} = 100$ for vector-sketches. For the action space, we take varying canvas-sizes from $C$ and set $\\beta = 1$ in Eq.2. $\\lambda$ , $\\lambda_r$ , $\\lambda_{\\mathrm{Tri}}$ and $\\lambda_{\\mathrm{F}}$ are empirically set to 0.5, 0.4, 0.48 and 0.35 (Fig.5) respectively. While training $\\psi_C$ , sketches are rendered at different completion (30%, 35%, ... 100%) levels to simulate varying levels of abstraction, ensuring a sketch-abstraction-aware canvas selector (e.g., a smaller canvas for less detailed sketch).", + "bbox": [ + 511, + 551, + 906, + 799 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation Metric: Accuracy for FG-SBIR is the percentage of sketches where true-match photos are ranked in the Top-q (q = 1,10) lists. Computational efficiency is measured via parameter-count and average FLOPs – lower is better. For average FLOPs in Table 2, we take the combined sum of FLOPs per sample [12] from (i) canvas-size selector and (ii) retrieval model, using the predicted resolution, and average across the validation set.", + "bbox": [ + 511, + 803, + 908, + 898 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "28387", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/b32418968c2c50c829b413e5b7f1ec33dde800d4da0502d806f373c369701d92.jpg", + "table_caption": [ + "Table 2. Quantitative Analysis on FG-SBIR. Best viewed when zoomed." + ], + "table_footnote": [], + "table_body": "
MethodsCanvas-size c×cParams (mil.)ShoeV2 [75]ChairV2 [75]Sketchy [55]FSCOCO [15]
Top1 (%)Top10 (%)FLOPS (G)Top1 (%)Top10 (%)FLOPS (G)Top1 (%)Top10 (%)FLOPs (G)Top1 (%)Top10 (%)FLOPS (G)
State-of-the-ArtsTriplet-SN [75]32x328.7509.77 (↓18.94)24.82 (↓46.74)0.08315.61 (↓32.04)29.86 (↓54.38)0.0834.23 (↓11.12)12.58 (↓23.13)0.0830.76 (↓3.94)6.22 (↓14.8)0.083
64x648.7517.63 (↓11.08)44.96 (↓26.60)0.33830.25 (↓17.4)54.21 (↓30.03)0.3389.68 (↓5.67)22.45 (↓13.26)0.3381.85 (↓2.85)7.33 (↓13.69)0.338
128x1288.7525.01 (↓3.70)62.35 (↓9.21)1.39740.61 (↓7.04)72.68 (↓11.56)1.39712.98 (↓2.37)29.63 (↓6.08)1.3973.28 (↓1.42)15.92 (↓5.1)1.397
256x2568.7528.7171.565.28047.6584.245.28015.3535.715.2804.721.025.280
SketchyNetV12.2228.46 (↓0.25)69.01 (↓2.55)0.83346.28 (↓0.37)82.33 (↓1.91)0.83314.85 (↓0.5)34.68 (↓1.03)0.8334.22 (↓0.48)18.33 (↓2.69)0.833
SketchyNetV22.2727.89 (↓0.82)68.76 (↓2.80)0.26445.98 (↓1.67)80.14 (↓4.10)0.29414.21 (↓1.14)34.16 (↓1.55)0.3213.98 (↓0.72)17.64 (↓3.38)0.423
HOLEE-SN [59]32x329.3110.84 (↓20.9)27.28 (↓48.5)0.09618.72 (↓34.69)31.68 (↓55.88)0.0966.01 (↓10.69)14.12 (↓24.78)0.0960.86 (↓4.04)6.93 (↓14.78)0.096
64x649.3119.83 (↓11.91)51.72 (↓24.06)0.38734.32 (↓19.09)56.46 (↓31.1)0.3879.92 (↓6.78)24.38 (↓14.52)0.3872.11 (↓2.79)7.86 (↓13.85)0.387
128x1289.3126.78 (↓4.96)66.28 (↓9.5)1.54545.74 (↓7.67)75.59 (↓11.97)1.54513.81 (↓2.89)33.68 (↓5.22)1.5453.65 (↓1.25)16.32 (↓5.39)1.545
256x2569.3131.7475.785.75853.4187.565.75816.7038.905.7584.921.715.758
SketchyNetV12.2231.59 (↓0.15)73.96 (↓1.82)0.83353.27 (↓0.14)86.93 (↓0.63)0.83316.25 (↓0.45)38.21 (↓0.69)0.8334.56 (↓0.34)19.92 (↓1.79)0.833
SketchyNetV22.2730.86 (↓0.88)72.56 (↓3.22)0.25952.74 (↓0.67)84.02 (↓3.54)0.28915.91 (↓0.79)37.88 (↓1.02)0.3154.04 (↓0.86)18.63 (↓3.08)0.417
Triplet-RL [6]32x3222.111.98 (↓22.12)28.11 (↓50.71)0.14220.53 (↓36.01)32.59 (↓57.02)0.1421.76 (↓2.94)3.88 (↓6.47)0.142
64x6422.121.28 (↓12.82)54.71 (↓24.11)0.57736.09 (↓20.45)57.33 (↓32.28)0.5773.04 (↓1.66)6.79 (↓3.56)0.577
128x12822.128.83 (↓5.27)67.84 (↓10.98)2.29948.38 (↓8.16)77.25 (↓12.36)2.2994.12 (↓0.58)9.05 (↓1.30)2.299
256x25622.134.1078.826.04156.5489.616.0414.7010.356.041
SketchyNetV12.2233.88 (↓0.22)77.15 (↓1.67)0.83355.92 (↓0.62)88.20 (↓1.41)0.8334.59 (↓0.11)10.21 (↓0.14)0.833
SketchyNetV22.2733.26 (↓0.84)76.84 (↓1.98)0.25555.14 (↓1.40)87.31 (↓2.30)0.2854.51 (↓0.19)9.89 (↓0.46)0.308
StyleVAE [51]32x3225.3712.68 (↓23.79)28.69 (↓53.14)0.12522.48 (↓40.38)32.10 (↓59.04)0.12506.92 (↓12.7)16.25 (↓29.53)0.125
64x6425.3722.93 (↓13.54)57.71 (↓24.12)0.50840.26 (↓22.6)58.23 (↓32.91)0.50812.46 (↓7.16)29.62 (↓16.16)0.508
128x12825.3730.91 (↓5.56)70.06 (↓11.77)2.02453.88 (↓8.98)78.64 (↓12.5)2.02417.15 (↓2.47)38.84 (↓6.94)2.024
256x25625.3736.4781.835.64262.8691.145.64219.6245.785.642
SketchyNetV12.2236.11 (↓0.36)79.63 (↓2.20)0.83362.23 (↓0.63)88.95 (↓2.19)0.83319.48 (↓0.12)44.02 (↓1.76)0.833
SketchyNetV22.2735.88 (↓0.59)78.46 (↓3.37)0.25161.95 (↓0.91)87.68 (↓3.46)0.28119.10 (↓0.50)43.57 (↓2.21)0.304
Partial-OT [14]32x3222.1
64x6422.1
128x12822.1
256x25622.1
SketchyNetV12.22
SketchyNetV22.27
B-BFR224x2244.6124.3261.232.60243.3976.332.60210.4724.342.6022.1511.612.602
B-DRSDynamic9.3625.1268.845.49344.7878.146.18711.8726.586.2453.8617.636.291
B-CropDynamic4.9220.3252.126.56235.0757.686.59708.2720.196.6291.9809.676.688
B-Regress [48]Dynamic2.2722.8261.330.26138.4264.650.29108.6921.980.3232.0710.150.375
B-AKD [76]Dynamic2.2723.6764.260.26838.6171.880.29512.6727.710.3312.9114.010.392
B-PKT [45]Dynamic2.2724.3165.910.25942.6873.660.28812.2728.310.3173.4516.220.372
B-VGG16-SN256x25614.7133.0378.5140.1852.1684.0840.1818.6141.2540.185.1623.6240.18
B-ThiNet [37]256x2568.327.51 (↓25.52)22.34 (↓56.17)9.34212.31 (↓39.85)26.63 (↓57.45)9.3423.16 (↓15.45)12.73 (↓28.52)9.3420.65 (↓4.51)5.35 (↓18.27)9.342
B-Prune [38]256x2566.173.92 (↓29.11)17.32 (↓61.19)8.1387.98 (↓44.18)19.82 (↓64.26)8.1381.45 (↓17.16)8.22 (↓33.03)8.1380.53 (↓4.63)4.97 (↓18.65)8.138
B-VGG16-SN SketchyNetV22.2732.77 (↓0.26)78.21 (↓0.3)0.25450.75 (↓1.41)82.21 (↓1.87)0.28316.55 (↓2.06)38.66 (↓2.59)0.3324.69 (↓0.47)21.76 (↓1.86)0.407
", + "bbox": [ + 94, + 103, + 903, + 518 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Competitors", + "text_level": 1, + "bbox": [ + 89, + 530, + 225, + 546 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We evaluate against State-of-the-Arts (SoTAs) and a few relevant curated baselines (B-). (i) SoTAs: Unlike Triplet-SN [75], HOLEF-SN [59], Triplet-RL [6] (we report full-sketch evaluation) and StyleVAE [51]; Partial-OT [14] specifically caters to scene-level FG-SBIR. For fairness, we vary input-resolution of sketches during inference (fixed at their original $256 \\times 256$ setup during training). We also adapt them as teachers, to judge our accuracy retainment after model-compression (SketchyNetV1) using full-res sketch and, further optimising canvas-size with canvas selector (SketchyNetV2). (ii) KD baselines: Here we re-purpose existing works in a cross-modal setting, keeping our canvas selector same, with Triplet-SN as the teacher. B-Regress - regresses features from teacher and student for both sketches and images, aligning their dimensions with an extra transformation layer [48], using a $l_{2}$ regression loss. Following [45] B-PKT calculates the conditional probability density for any pair of points in the teacher's embedding space [45], which models the probability of any two samples being close together. Sampling 'N' instances, it establishes a probability distribution for pairwise interactions within that space. Simultaneously, in the student's embedding space, a comparable distribution is derived, and the model minimises the divergence between them using a KL-divergence loss. B-AKD follows [76] in using spatial attention maps for knowledge transfer using similar dimensional matching like B-Regress. (iii) Resizing strategies: Keeping the same KD-paradigm with Triplet-SN-teacher, we vary the input", + "bbox": [ + 88, + 553, + 485, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "size resizing module following existing works. $B$ -BFR follows [62] in designing a bilinear feature resizer amidst the intermediate layers of a CNN network, to output a down-scaled image. We choose its resized resolution of $224 \\times 224$ (lowest) for subsequent evaluation. $B$ -DRS follows [12] in training a convolutional module that takes an image in full-resolution, calculates a probability vector on available canvas-sizes, and transforms it into binary decisions, indicating the scale factor of selection. $B$ -Crop follows [32] in meta-learning an image-cropping module that crops the input image to a target resolution for low compute evaluation. (iv) Pruning alternatives: Here we adapt two VGG16-backboned methods off-the-shelf from network-pruning family among other efficiency paradigms [18, 21, 62], in FG-SBIR setup - $B$ -ThiNet [37] and $B$ -Prune [38]. For comparative reference we report performance of a VGG16-backboned Triplet-SN network ( $B$ -VGG16-SN) and its SketchyNetV2 variant.", + "bbox": [ + 511, + 531, + 908, + 753 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Performance and Compute Analysis:", + "text_level": 1, + "bbox": [ + 511, + 757, + 831, + 773 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Analysis on Accuracy: Table 2 reports the quantitative evaluation of our model against others. Triplet-SN [75] and HOLEF-SN [59] perform lower owing to weak backbones of Sketch-A-Net [74]. Enhanced by its RL-optimised reward function Triplet-RL[6] scores higher $(2.36\\% \\uparrow$ Top1 on ShoeV2) but fails to surpass StyleVAE [51], due to its delicate meta-learning disentanglement module addressing style-diversity. Being trained on region-wise sketch-photo associativity, Partial-OT [14] outperforms all others significantly on scene-level sketches (FS-COCO). We also notice a", + "bbox": [ + 511, + 775, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "28388", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "gradual drop in accuracy of SoTAs, when subjected to decreasing resolution of input-sketches suggesting that some methods are inherently accustomed to handling scale-shifts better than others, but are inferior to our dynamic approach (SketchyNetV2). Importantly, when most SoTAs are employed as teachers, their respective students (SketchyNetV1 variants) are seen to retain their respective original accuracy with a marginal drop of $\\approx 0.5 - 1\\%$ overall, thanks to our efficient KD paradigm. Having different teacher-student embedding spaces both $B$ -Regress and $B$ -AKD's paradigm of directly regressing on features proves incompatible to a cross-modal retrieval setting, dropping accuracy (compared to T-Triplet-SN), whereas lacking cross-modal discrimination $B$ -PKT performs lower than ours, proving our superiority in retaining accuracy.", + "bbox": [ + 89, + 90, + 483, + 272 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/de3a0d5bb52e94f4070e12b04a21b1cf472229379a14b6f2fd172071e21ff8bf.jpg", + "image_caption": [ + "Figure 4. Exemplar sketches at their optimal canvas-size" + ], + "image_footnote": [], + "bbox": [ + 93, + 277, + 189, + 324 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e362480feffaaa2a25a4bf1ad00113228371c443db89798b3bd7b2864ad24929.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 194, + 277, + 287, + 324 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/214d331dd4b3bbdc3eff100cacaf294036f74e397825e96a10e207f7b1fee5f4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 289, + 277, + 387, + 324 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ca18a9d5f00a783e7d294e3e47bdec3dd5eb72a362115abfe1899a23cb846b98.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 388, + 277, + 480, + 324 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Analysis on Compute: Although FLOPs [12] decrease (Table 2) with a reduction in input resolution across all SoTAs, it comes at a severe cost of accuracy ( $\\approx 21\\%$ average drop from $256 \\times 256$ to $32 \\times 32$ on ShoeV2). This is mainly because ideal canvas-size for optimal retrieval varies across samples (Fig. 2 right) - thus subjecting all samples to a fixed canvas-size proves detrimental for accuracy. However, a positive Acc@1 score of $\\approx 11\\%$ even at a low canvas-size of $32 \\times 32$ for every SoTA shows that some sketches are indeed recognisable if the model is trained properly. This is verified further by SketchyNetV2 variant of respective SoTAs, that secures the lowest FLOPs per SoTA, despite maintaining accuracy at par with its SoTA-counterpart - thanks to our RL-guided canvas selector, which can allot an optimal canvas-size per sketch towards better retrieval (Fig. 4). Contrarily, B-BFR fails to lower FLOPs as it fixes a $224 \\times 224$ input resolution for a satisfactory Acc@1 compared to Triplet-SN. B-Crop fairs better, however sketch being sparse, cropped sketches leads to confused representations, lowering accuracy. Despite coming closest in accuracy to ours on Triplet-SN, B-DRS lags behind considerably on compute, thanks to our canvas selector working in vector space (inputs $s_v$ ) and eliminating costly convolutional operations, unlike others using full-resolution sketch-raster as input. Being untrained to handle sparse-nature of sketches during such high compression, performance of both B-ThiNet and B-Prune collapses (vs. B-VGG16-SN) unlike its SketchyNetV2. Notably, the significant drop in parameter-count (4 - 10 times) against SoTAs, comes from choosing lightweight MobileNetV2 [54] as a student coupled with our canvas-size selector in every SketchyNetV2 variant. Compared to ShoeV2, FLOPs increase slightly for ChairV2, Sketchy and FS-COCO, owing to higher object (ChairV2) or scene-level (FS-COCO) details, and lower structural correspondence with photos (Sketchy), thus training a poorer retrieval model which in turn acts as a poor critic for our canvas selector. Our canvas selector alone takes a negligible 51.84K parameters with 0.02G FLOPs.", + "bbox": [ + 89, + 349, + 483, + 821 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 827, + 243, + 842 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As our teacher $(F_T)$ can be any pre-trained FG-SBIR network, we limit our ablation to justifying design choices of our student model and canvas-size selector. We use $B$ -VGG16-SN as a reference $F_T$ - Type- $\\emptyset$ in Table 3.", + "bbox": [ + 89, + 844, + 483, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Training Canvas-Size selector $\\psi_C$ : To justify loss objectives of $\\psi_C$ we evaluate on ShoeV2, eliminating $r^{-1}$ , $\\mathcal{L}_{\\mathrm{Tri}}^R$ , $\\mathcal{L}_{\\mathrm{F}}$ one by one from Eq. 6 (Table 3). While eliminating accuracy rewards, $(w / o \\text{ rank}^{-1}, w / o \\mathcal{L}_{\\mathrm{Tri}}^R)$ increases constraint on lowering FLOPs, dropping accuracy, ignoring $\\mathcal{L}_{\\mathrm{F}}$ $(w / o \\mathcal{L}_F)$ gains accuracy but costs high FLOPs. We thus optimally balance between the two using $\\lambda_{\\mathrm{F}}$ (Fig. 5 left). One can also choose to train for faster retrieval (lower FLOPs), given a decent Acc@5 (Fig. 5 right).", + "bbox": [ + 511, + 92, + 906, + 203 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e2f7d3047db873640356a1426d38a7e35719439634bff95966f7bd695227d686.jpg", + "image_caption": [ + "Figure 5. Impact of varying $\\lambda_{\\mathrm{F}}$ on Acc@1 and Acc@5" + ], + "image_footnote": [], + "bbox": [ + 534, + 208, + 689, + 271 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/799f80fd904eaa953d0823a729f55c7ac735d4637db5012301974a08ddeba2b7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 697, + 208, + 859, + 271 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Architectural insights: To judge the efficiency of our student's backbone, we explore a few cheaper backbone-networks (Types IV, V in Table 3). Evidently, MobileNetV2 (Ours) balances optimally between compute and accuracy. As an option to vector input for canvas selector, we re-design it with a simple seven-layer CNN module following [78] to use the full-resolution sketch-image as its input, and select a target canvas-size to downscale it for retrieval. With extra computational cost coming from spatial domain, total FLOPs incurred is much higher (VI). Next we compare efficacy of sketch as input to $\\psi_C$ with respect to types of vector format - offset-coordinate (VI) [22] vs. absolute ones (Ours). Turns out the former is better for encoding. Comparing sketch encoders among LSTM (VIII), GRU (Ours) and Transformer (IX), showed ours to be optimum empirically (Table 3). Furthermore, preponding recent pre-processing modules like stroke-subset selector of [9] to our canvas selector, yields $42.95\\%$ (vs. $43.7\\%$ [9]) Acc@1 ShoeV2, proving its off-the-shelf compatibility.", + "bbox": [ + 511, + 291, + 906, + 527 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/bf3cb53392aa2e184dbf301fcdade27b3b4eca10ccba4c189d6a61fb736c4ffa.jpg", + "table_caption": [ + "Table 3. Ablative studies (accuracy on ShoeV2)" + ], + "table_footnote": [], + "table_body": "
TypeFLOPsParamsShoeV2
(G)(mil.)Top1 (%)Top10 (%)
∅B-VGG16-SN40.1814.7133.0378.51
Iw/o rank-1-reward0.17325.3164.36
IIw/o \\( \\mathcal{L}_{\\text{Tri}}^R \\)-reward0.1822.2727.7869.32
IIIw/o \\( \\mathcal{L}_{\\text{F}} \\)-reward0.83332.9178.39
IVResNet18 backbone1.45111.1826.7268.18
VEfficientNet backbone0.4594.0129.4772.04
VIImage-based \\( \\psi_C \\)5.7307.6232.2177.68
VIIOffset \\( s_v \\)0.2552.2732.6178.40
VIIIDecoder-LSTM0.2682.2531.4177.32
IXDecoder-Tf0.3142.3431.9878.02
SketchyNetV20.2542.2732.7778.21
", + "bbox": [ + 524, + 547, + 887, + 700 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Generalisability of canvas selector: To judge if the optimal canvas-size selected is uniform irrespective of the training method, we first train four canvas-selectors using four SoTAs of Table 2. Evaluating them on ShoeV2 [75], $89.21\\%$ of test-set sketches in ShoeV2 to obtain the same optimal canvas size across four different models, trained on four different methods, which further proves that our canvas selector can work equally well as a pre-processing module for any new model. Furthermore, plugging a canvas selector trained with SketchyNetV1 (critic) from B-VGG16-SN with a different SketchyNetV1 from StyleVAE [51] during inference achieves comparable performance $(35.69\\%)$ Top-1) to a canvas selector trained on StyleVAE-SketchyNetV1 $(35.88\\%)$ Top-1) on ShoeV2. Importantly, $95.64\\%$ of test-set sketches yield the same optimal canvas-size across both models,", + "bbox": [ + 511, + 705, + 908, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "28389", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "confirming the cross-model generalisability of our canvas selector. (ii) On training canvas selector with teacher $F_{T}$ (for Triplet-SN) as critic, yields same optimal canvas-sizes as those for student $F_{S}$ , for $94.21\\%$ of test-set sketches in ShoeV2, while incurring comparable FLOPs reduction (0.268G) and accuracy (27.91% Top1) against $F_{S}$ as critic.", + "bbox": [ + 89, + 92, + 480, + 175 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Comparing Inference times: We evaluated on a 12 GB Nvidia RTX 2080 Ti GPU, on ShoeV2, for inference time of Triplet-SN, HOLEF-SN, Triplet-RL, and StyleVAE (SOTAs) on full-sketch to obtain 37, 38, 36 and $42\\mathrm{ms}$ , against 10, 11, 11, and $12\\mathrm{ms}$ for their SketchyNetV2 variants respectively. For practical validation, we evaluated onnx-INT8 version of Triplet-SN [75] and its SketchyNetV2 variant (ours) after post-training quantisation, on an iPhone13 via the AI-benchmark app for ShoeV2 [75] to obtain 51ms and 18ms respectively, which shows our efficiency.", + "bbox": [ + 89, + 175, + 480, + 299 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Retraining SoTA on mixed resolution sketches?: To explore if retraining SoTAs on sketches with mixed resolutions, we re-train Triple-SN [75] on rendered mixed-resolution sketches of ShoeV2. Poor performance in Table 4 at lower resolutions similar to Table 2 validates our KD approach. The key lies in that the sketch-feature from teacher used for student-guidance $(f_s^T)$ is always extracted from full-resolution sketch while that extracted by student is at varying resolutions, imparting scale invariance, which is ideally suited to our cross-modal retrieval setup.", + "bbox": [ + 89, + 301, + 480, + 428 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/b29958115849f070ffd1b07979377daef4410f50587b4f967db626ffcd86d1bb.jpg", + "table_caption": [ + "Table 4. Evaluating SoTAs trained on mixed-resolution sketches" + ], + "table_footnote": [], + "table_body": "
Canvas SizeTriplet-SNTriplet-RL
Top1 (%)Top10 (%)FLOPs (G)Top1 (%)Top10 (%)FLOPs (G)
32×3210.9126.650.08312.9731.080.142
64×6419.0448.120.33822.9656.230.577
128×12826.0763.871.39730.0472.952.299
256×25628.7471.685.28034.2177.246.041
", + "bbox": [ + 93, + 450, + 475, + 536 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.4. Further Insights", + "text_level": 1, + "bbox": [ + 89, + 542, + 253, + 558 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Cross-dataset Generalisation: To judge the generalisation potential of our canvas-size selector across datasets we train it on ShoeV2, but plug it to a retrieval model (SketchyNetV1) trained on ChairV2, and evaluate on ChairV2. Against the standard result (Table 2) of training both modules (SketchyNetV2) on ChairV2 (0.285G, $55.32\\%$ Acc@1), this setup retains accuracy at $53.14\\%$ with slight rise in FLOPs (0.314G), thus confirming its potential.", + "bbox": [ + 89, + 561, + 480, + 657 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Faster Sketch-recognition: Exploring potential of our canvas-size selector, we prepend it to a standard sketch-recognition pipeline [74] to optimise the canvas-size of query sketch, and reduce overall compute. Unlike our $\\psi_{C}$ , it uses negative of cross-entropy loss as accuracy reward $(R_{\\mathrm{acc}})$ , keeping rest same. Compared to existing Sketch-a-Net [74] incurring 5.28G FLOPs at $68.71\\%$ accuracy on Quickdraw [22], our canvas selector fitted paradigm drops FLOPs to 0.261G while retaining accuracy at $68.14\\%$ . This shows for the first time, that sparsity of sketches can be handled, in various downstream tasks via our $\\psi_{C}$ .", + "bbox": [ + 89, + 661, + 480, + 800 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Extension to Category-level SBIR: For categorical SBIR we use Sketchy (ext) [36] dataset (90:10 train:test split), which holds $75k$ sketches across 125 categories having $\\approx 73k$ images [36] in total. It is evaluated using mean average precision (mAP@all) and precision at top 200 retrieval (P@200) [36]. Here we compete against state-of-the-arts (SoTA) like $B-D2S$ that follows [19] without its zero-shot setting), and a simple baseline using a Siamese-style", + "bbox": [ + 89, + 803, + 480, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/717d0aa723a2a65890d11c8c2d6fb7b939b4b1734a0b79265d7c07635a7164a0.jpg", + "table_caption": [ + "Table 5. Quantitative Evaluation on Categorical SBIR" + ], + "table_footnote": [], + "table_body": "
MethodsFLOPsParamsSketchy(ext)
(G)(mil.)mAP@allP@200
SoTAB-SBIR-SN40.1814.710.7150.861
DSH [36]22.1416.320.7110.858
B-D2S [19]17.75136.360.8100.894
StyleVAE [51]8.11625.370.9050.927
T-SoTAT-SBIR-SN0.2540.7020.810
T-DSH0.2892.270.7040.823
T-D2S0.2680.7980.845
T-StyleVAE0.2510.8860.897
", + "bbox": [ + 537, + 107, + 879, + 220 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "VGG-16 network (B-SBIR-SN) following [75]. Similar to Section 5.1 adapting SoTAs as teachers in our SketchyNetV2 paradigm (T-SoTA) shows them retaining their accuracy (Table 5) while significantly lowering compute.", + "bbox": [ + 511, + 223, + 903, + 279 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Teacher-Student Pairs: To explore other teacher-student pairs apart from ours, we test SketchyNetV2 models of a few $(F_T,F_S)$ pairs, where teachers are trained similar to B-VGG16-SN. From Table 6 we see our pair to outperform others.", + "bbox": [ + 511, + 281, + 903, + 335 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/bd6acc95657a46b7b7a46da8adaa50ef7ce4781950b971dac37a1349e3cb452b.jpg", + "table_caption": [ + "Table 6. Evaluation on Teacher-Student Pairs" + ], + "table_footnote": [], + "table_body": "
Teacher\\StudentMobileNetV2EfficientNetResNet-18
Top-1FLOPsTop-1FLOPsTop-1FLOPs
VGG-16 [30]32.77%0.254G29.47%0.459G26.72%1.451G
Inception-V3 [74]27.89%0.264G27.16%0.461G25.31%1.452G
", + "bbox": [ + 514, + 357, + 900, + 421 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Combining compression methods: To explore if a combination of compression methods is capable of increasing efficiency, we subject our Knowledge Distillation (KD) trained model (SketchyNetV1) of Triplet-SN [75] (Table 2) to (i) quantisation following [21], and (ii) network-pruning following [38], separately, and then train their respective canvas-selectors. While the former secures $27.34\\%$ (vs. $27.89\\%$ ours) Top-1 score with a lower inference time of $9.1\\mathrm{ms}$ (vs. $10\\mathrm{ms}$ ours), the latter scores $26.91\\%$ (vs. $27.89\\%$ ms ours) Top-1 with lower FLOPs of 0.249G (vs. 0.264G ours). Such competitive results further endorse exploring combining compression methods as future works.", + "bbox": [ + 511, + 430, + 903, + 583 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Future Works: Sketch being an abstraction of an object's photo, when resized to a lower scale for retrieval, infers that its object's semantic is interpretable at that scale. This should ideally provide a lower bound for resizing a photo while retaining interpretability (e.g. high retrieval score). Consequently, designing a learnable photo-resizer aided by our trained sketch-canvas selector is a potential future work. Secondly, ours being a meta-framework, using more recent frameworks like DINOv2 [41] as large-scale teacher models for further gains is another targeted future work.", + "bbox": [ + 511, + 585, + 905, + 710 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 715, + 633, + 731 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We for the first time investigate the problem of efficient inference for FG-SBIR, where we show that using existing efficient lightweight networks directly, are unfit for sketches. We thus propose two generic components that can be adapted to photo-networks for sketch-specific data. Firstly, a cross-modal knowledge distillation paradigm distills the knowledge of larger models to a smaller one, directly reducing FLOPs (params) by 97.92 $(84.9)\\%$ . Secondly, an abstraction-aware canvas-size selector dynamically selects the best canvas-size for a sketch, reducing FLOPs by extra two-thirds. Extensive experiments show our method to perform at par with state-of-the-arts using much lower compute thus further conveying our method's worth in enhancing them towards deployment.", + "bbox": [ + 511, + 734, + 905, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "28390", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Manoj Alwani, Yang Wang, and Vashisht Madhavan. Decore: Deep compression with reinforcement learning. In CVPR, 2022. 2", + "[2] Aneeshan Sain and Ayan Kumar Bhunia and Pinaki Nath Chowdhury and Aneeshan Sain and Subhadeep Koley and Tao Xiang and Yi-Zhe Song. CLIP for All Things Zero-Shot Sketch-Based Image Retrieval, Fine-Grained or Not. In CVPR, 2023. 1", + "[3] Jimmy Ba and Rich Caruana. Do deep nets really need to be deep? NeurIPS, 2014. 2", + "[4] Hessam Bagherinezhad, Maxwell Horton, Mohammad Rastegari, and Ali Farhadi. Label refinery: Improving image-net classification through label progression. arXiv preprint arXiv:1805.02641, 2018. 2", + "[5] Ayan Kumar Bhunia, Ayan Das, Umar Riaz Muhammad, Yongxin Yang, Timothy M Hospedales, Tao Xiang, Yulia Gryaditskaya, and Yi-Zhe Song. Pixel: A competitive sketching ai agent. so you think you can sketch? ACM TOG, 2020. 3", + "[6] Ayan Kumar Bhunia, Yongxin Yang, Timothy M Hospedales, Tao Xiang, and Yi-Zhe Song. Sketch less for more: On-the-fly fine-grained sketch based image retrieval. In CVPR, 2020. 1, 2, 3, 6", + "[7] Ayan Kumar Bhunia, Pinaki Nath Chowdhury, Aaneshan Sain, Yongxin Yang, Tao Xiang, and Yi-Zhe Song. More photos are all you need: Semi-supervised learning for fine-grained sketch based image retrieval. In CVPR, 2021. 1, 2, 3, 4", + "[8] Ayan Kumar Bhunia, Pinaki Nath Chowdhury, Yongxin Yang, Timothy M Hospedales, Tao Xiang, and Yi-Zhe Song. Vectorization and rasterization: Self-supervised learning for sketch and handwriting. In CVPR, 2021. 3", + "[9] Ayan Kumar Bhunia, Subhadeep Koley, Abdullah Faiz Ur Rahman Khilji, Aneeshan Sain, Pinaki Nath Chowdhury, Tao Xiang, and Yi-Zhe Song. Sketching without worrying: Noise-tolerant sketch-based image retrieval. In CVPR, 2022. 1, 2, 3, 7", + "[10] Ayan Kumar Bhunia, Subhadeep Koley, Amandeep Kumar, Aneeshan Sain, Pinaki Nath Chowdhury, Tao Xiang, and Yi-Zhe Song. Sketch2Saliency: Learning to Detect Salient Objects from Human Drawings. In CVPR, 2023. 1", + "[11] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In ICML, 2020. 3", + "[12] Ying Chen, Liang Qiao, Zhanzhan Cheng, Shiliang Pu, Yi Niu, and Xi Li. Dynamic low-resolution distillation for cost-efficient end-to-end text spotting. In ECCV, 2022. 5, 6, 7", + "[13] Pinaki Nath Chowdhury, Ayan Kumar Bhunia, Viswanatha Reddy Gajjala, Aneeshan Sain, Tao Xiang, and Yi-Zhe Song. Partially does it: Towards scene-level fg-sbir with partial input. In CVPR, 2022. 1, 2", + "[14] Pinaki Nath Chowdhury, Ayan Kumar Bhunia, Viswanatha Reddy Gajjala, Aneeshan Sain, Tao Xiang, and Yi-Zhe Song. Partially does it: Towards scene-level fg-sbir with partial input. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 6" + ], + "bbox": [ + 93, + 114, + 482, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] Pinaki Nath Chowdhury, Aneeshan Sain, Ayan Kumar Bhunia, Tao Xiang, Yulia Gryaditskaya, and Yi-Zhe Song. Fscoco: Towards understanding of freehand sketches of common objects in context. In ECCV, 2022. 5, 6", + "[16] Chowdhury, Pinaki Nath and Bhunia, Ayan Kumar and Sain, Aneeshan and Koley, Subhadeep and Xiang, Tao and Song, Yi-Zhe. Democratising 2D Sketch to 3D Shape Retrieval Through Pivoting. In ICCV, 2023. 2", + "[17] John Collomosse, Tu Bui, and Hailin Jin. Livesketch: Query perturbations for guided sketch-based visual search. In CVPR, 2019. 2, 5", + "[18] Matthieu Courbariaux, Yoshua Bengio, and Jean-Pierre David. Binaryconnect: Training deep neural networks with binary weights during propagations. NeurIPS, 28, 2015. 2, 6", + "[19] Sounak Dey, Pau Riba, Anjan Dutta, Josep Llados, and Yi-Zhe Song. Doodle to search: Practical zero-shot sketch-based image retrieval. In CVPR, 2019. 2, 8", + "[20] Anjan Dutta and Zeynep Akata. Semantically tied paired cycle consistency for zero-shot sketch-based image retrieval. In CVPR, 2019. 1", + "[21] Suyog Gupta, Ankur Agrawal, Kailash Gopalakrishnan, and Pritish Narayanan. Deep learning with limited numerical precision. In ICML, 2015. 2, 6, 8", + "[22] David Ha and Douglas Eck. A neural representation of sketch drawings. In ICLR, 2018. 1, 5, 7, 8", + "[23] Song Han, Huizi Mao, and William J Dally. Deep compression: Compressing deep neural networks with pruning, trained quantization and huffman coding. In ICLR, 2016. 2", + "[24] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 3", + "[25] Geoffrey Hinton, Oriol Vinyls, and Jeff Dean. Distilling the knowledge in a neural network. In NeurIPS Deep Learning Workshop, 2014. 2", + "[26] Gao Huang, Zhuang Liu, Laurens Van Der Maaten, and Kilian Q Weinberger. Densely connected convolutional networks. In CVPR, 2017. 3", + "[27] Shouyou Huang and Qiang Wu. Robust pairwise learning with huber loss. Journal of Complexity, 2021. 4", + "[28] Zehao Huang and Naiyan Wang. Like what you like: Knowledge distill via neuron selectivity transfer. arXiv preprint arXiv:1707.01219, 2017. 2", + "[29] Leslie Pack Kaelbling, Michael L Littman, and Andrew W Moore. Reinforcement learning: A survey. JAIR, 1996. 3, 5", + "[30] Andrew Zisserman Karen Simonyan. Very deep convolutional networks for large-scale image recognition. In ICLR, 2015. 3, 4, 8", + "[31] Debang Li, Huikai Wu, Junge Zhang, and Kaiqi Huang. A2-rl: Aesthetics aware reinforcement learning for image cropping. In CVPR, 2018. 3", + "[32] Debang Li, Junge Zhang, and Kaiqi Huang. Learning to learn cropping models for different aspect ratio requirements. In CVPR, 2020. 6", + "[33] Xiaodan Liang, Lisa Lee, and Eric P Xing. Deep variation-structured reinforcement learning for visual relationship and attribute detection. In CVPR, 2017. 3" + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "28391", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[34] Darryl Lin, Sachin Talathi, and Sreekanth Annapureddy. Fixed point quantization of deep convolutional networks. In ICML, 2016. 2", + "[35] Hangyu Lin, Yanwei Fu, Yu-Gang Jiang, and Xiangyang Xue. Sketch-bert: Learning sketch bidirectional encoder representation from transformers by self-supervised learning of sketch gestalt. In CVPR, 2020. 5", + "[36] Li Liu, Fumin Shen, Yuming Shen, Xianglong Liu, and Ling Shao. Deep sketch hashing: Fast free-hand sketch-based image retrieval. In CVPR, 2017. 2, 8", + "[37] Jian-Hao Luo, Jianxin Wu, and Weiyao Lin. Thinet: A filter level pruning method for deep neural network compression. In ICCV, 2017. 2, 6", + "[38] Pavlo Molchanov, Stephen Tyree, Tero Karras, Timo Aila, and Jan Kautz. Pruning convolutional neural networks for resource efficient inference. arXiv preprint arXiv:1611.06440, 2016. 6, 8", + "[39] Umar Riaz Muhammad, Yongxin Yang, Yi-Zhe Song, Tao Xiang, and Timothy M Hospedales. Learning deep sketch abstraction. In CVPR, 2018. 3", + "[40] Umar Riaz Muhammad, Yongxin Yang, Timothy Hospedales, Tao Xiang, and Yi-Zhe Song. Goal-driven sequential data abstraction. In ICCV, 2019. 1, 3", + "[41] Maxime Oquab, Timothee Darct, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. DINOv2: Learning Robust Visual Features without Supervision. arXiv preprint arXiv:2304.07193, 2023. 8", + "[42] Kaiyue Pang, Yi-Zhe Song, Tony Xiang, and Timothy M Hospedales. Cross-domain generative learning for fine-grained sketch-based image retrieval. In BMVC, 2017. 2", + "[43] Kaiyue Pang, Ke Li, Yongxin Yang, Honggang Zhang, Timothy M Hospedales, Tao Xiang, and Yi-Zhe Song. Generalising fine-grained sketch-based image retrieval. In CVPR, 2019. 1", + "[44] Kaiyue Pang, Yongxin Yang, Timothy M Hospedales, Tao Xiang, and Yi-Zhe Song. Solving mixed-modal jigsaw puzzle for fine-grained sketch-based image retrieval. In CVPR, 2020. 2, 3", + "[45] Nikolaos Passalis and Anastasios Tefas. Learning deep representations with probabilistic knowledge transfer. In ECCV, 2018. 6", + "[46] Pinaki Nath Chowdhury and Ayan Kumar Bhunia and Aneeeshan Sain and Subhadeep Koley and Tao Xiang and Yi-Zhe Song. SceneTrilogy: On Human Scene-Sketch and its Complementarity with Photo and Text. In CVPR, 2023. 2", + "[47] Pinaki Nath Chowdhury and Ayan Kumar Bhunia and Aeeeshan Sain and Subhadeep Koley and Tao Xiang and Yi-Zhe Song. What Can Human Sketches Do for Object Detection? In CVPR, 2023. 1", + "[48] Adriana Romero, Nicolas Ballas, Samira Ebrahimi Kahou, Antonie Chassang, Carlo Gatta, and Yoshua Bengio. Fitnets: Hints for thin deep nets. In ICLR, 2015. 2, 4, 6", + "[49] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. IJCV, 2015. 4" + ], + "bbox": [ + 91, + 90, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[50] Aneeshan Sain, Ayan Kumar Bhunia, Yongxin Yang, Tao Xiang, and Yi-Zhe Song. Cross-modal hierarchical modelling, fine-grained sketch based image retrieval. In BMVC, 2020, 2, 3, 4", + "[51] Aneeshan Sain, Ayan Kumar Bhunia, Yongxin Yang, Tao Xiang, and Yi-Zhe Song. Stylemeup: Towards style-agnostic sketch-based image retrieval. In CVPR, 2021. 1, 2, 3, 4, 6, 7, 8", + "[52] Aneeshan Sain, Ayan Kumar Bhunia, Vaishnav Potlapalli, Pinaki Nath Chowdhury, Tao Xiang, and Yi-Zhe Song. Sketch3t: Test-time training for zero-shot sbir. In CVPR, 2022. 2", + "[53] Aneeshan Sain, Ayan Kumar Bhunia, Subhadeep Koley, Pinaki Nath Chowdhury, Soumitri Chattopadhyay, Tao Xiang, and Yi-Zhe Song. Exploiting Unlabelled Photos for Stronger Fine-Grained SBIR. In CVPR, 2023. 1, 2", + "[54] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zh-moginov, and Liang-Chieh Chen. Mobilenetv2: Inverted residuals and linear bottlenecks. In CVPR, 2018. 1, 3, 4, 5, 7", + "[55] Patsorn Sangkloy, Nathan Burnell, Cusuh Ham, and James Hays. The sketchy database: learning to retrieve badly drawn bunnies. ACM TOG, 2016. 5, 6", + "[56] John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347, 2017. 5", + "[57] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. In ICLR, 2015. 1, 3", + "[58] Jifei Song, Yi-Zhe Song, Tony Xiang, and Timothy M Hospedales. Fine-grained image retrieval: the text/sketch input dilemma. In BMVC, 2017. 2", + "[59] Jifei Song, Qian Yu, Yi-Zhe Song, Tao Xiang, and Timothy M Hospedales. Deep spatial-semantic attention for fine-grained sketch-based image retrieval. In ICCV, 2017. 2, 6", + "[60] Subhadeep Koley and Ayan Kumar Bhunia and Aneeshan Sain and Pinaki Nath Chowdhury and Tao Xiang and Yi-Zhe Song. Picture that Sketch: Photorealistic Image Generation from Abstract Sketches. In CVPR, 2023. 1", + "[61] Richard S Sutton, David A McAllester, Satinder P Singh, and Yishay Mansour. Policy gradient methods for reinforcement learning with function approximation. In NeurIPS, 2000. 5", + "[62] Hossein Talebi and Peyman Milanfar. Learning to resize images for computer vision tasks. In ICCV, 2021. 2, 6", + "[63] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In ICML, 2019. 1, 3", + "[64] Burak Uzkent and Stefano Ermon. Learning when and where to zoom with deep reinforcement learning. In CVPR, 2020. 2", + "[65] Supawit Vatathanavaro, Suchat Tungjitnob, and Kitsuchart Pasupa. White blood cell classification: a comparison between vgg-16 and resnet-50 models. In JSCI, 2018. 3", + "[66] Mahes Visvalingam and J Duncan Whyatt. The douglaspeucker algorithm for line simplification: re-evaluation through visualization. In Computer Graphics Forum, 1990. 5" + ], + "bbox": [ + 516, + 90, + 903, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "28392", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[67] Xin Wang, Qiuyuan Huang, Asli Celikyilmaz, Jianfeng Gao, Dinghan Shen, Yuan-Fang Wang, William Yang Wang, and Lei Zhang. Reinforced cross-modal matching and self-supervised imitation learning for vision-language navigation. In CVPR, 2019. 3", + "[68] Yunhe Wang, Chang Xu, Shan You, Dacheng Tao, and Chao Xu. Cnnpack: Packing convolutional neural networks in the frequency domain. NeurIPS, 29, 2016. 2", + "[69] Yulin Wang, Kangchen Lv, Rui Huang, Shiji Song, Le Yang, and Gao Huang. Glance and focus: a dynamic approach to reducing spatial redundancy in image classification. In NeurIPS, 2020. 2", + "[70] Kilian Q Weinberger and Lawrence K Saul. Distance metric learning for large margin nearest neighbor classification. JMLR, 2009. 4", + "[71] Peng Xu, Chaitanya K Joshi, and Xavier Bresson. Multigraph transformer for free-hand sketch recognition. IEEE T-NNLS, 2021. 4", + "[72] Peng Xu, Timothy M Hospedales, Qiyue Yin, Yi-Zhe Song, Tao Xiang, and Liang Wang. Deep learning for free-hand sketch: A survey. IEEE TPAMI, 2022. 1, 2", + "[73] Ran Xu, Fangzhou Mu, Jayoung Lee, Preeti Mukherjee, Somali Chaterji, Saurabh Bagchi, and Yin Li. SmartAdapt: Multi-branch object detection framework for videos on mobiles. In CVPR, 2022. 2", + "[74] Qian Yu, Yongxin Yang, Yi-Zhe Song, Tao Xiang, and Timothy M. Hospedales. Sketch-a-net that beats humans. In BMVC, 2015. 6, 8", + "[75] Qian Yu, Feng Liu, Yi-Zhe Song, Tao Xiang, Timothy M Hospedales, and Chen-Change Loy. Sketch me that shoe. In CVPR, 2016. 1, 2, 3, 4, 5, 6, 7, 8", + "[76] Sergey Zagoruyko and Nikos Komodakis. Paying more attention to attention: Improving the performance of convolutional neural networks via attention transfer. In ICLR, 2017. 2, 6", + "[77] Shuchang Zhou, Yuxin Wu, Zekun Ni, Xinyu Zhou, He Wen, and Yuheng Zou. Dorefa-net: Training low bitwidth convolutional neural networks with low bitwidth gradients. arXiv preprint arXiv:1606.06160, 2016. 2", + "[78] Mingjian Zhu, Kai Han, Enhua Wu, Qiulin Zhang, Ying Nie, Zhenzhong Lan, and Yunhe Wang. Dynamic resolution network. In NeurIPS, 2021. 7" + ], + "bbox": [ + 91, + 92, + 480, + 686 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "28393", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Sketch Down the FLOPs_ Towards Efficient Networks for Human Sketch/d51d8c32-98f5-44d4-9674-ee9bc3d4af26_model.json b/2025/Sketch Down the FLOPs_ Towards Efficient Networks for Human Sketch/d51d8c32-98f5-44d4-9674-ee9bc3d4af26_model.json new file mode 100644 index 0000000000000000000000000000000000000000..ca82204bf2125c084c566dd99693d2cdc345d04e --- /dev/null +++ b/2025/Sketch Down the FLOPs_ Towards Efficient Networks for Human Sketch/d51d8c32-98f5-44d4-9674-ee9bc3d4af26_model.json @@ -0,0 +1,2301 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.13, + 0.123, + 0.868, + 0.144 + ], + "angle": 0, + "content": "Sketch Down the FLOPs: Towards Efficient Networks for Human Sketch" + }, + { + "type": "text", + "bbox": [ + 0.158, + 0.165, + 0.822, + 0.202 + ], + "angle": 0, + "content": "Aneeshan Sain1 Subhajit Maity2* Pinaki Nath Chowdhury1 Shubhadeep Koley1 Ayan Kumar Bhunia1 Yi-Zhe Song1" + }, + { + "type": "text", + "bbox": [ + 0.266, + 0.202, + 0.732, + 0.22 + ], + "angle": 0, + "content": "\\(^{1}\\) SketchX, CVSSP, University of Surrey, United Kingdom." + }, + { + "type": "text", + "bbox": [ + 0.18, + 0.22, + 0.815, + 0.238 + ], + "angle": 0, + "content": "\\(^{2}\\) Department of Computer Science, University of Central Florida, United States." + }, + { + "type": "text", + "bbox": [ + 0.148, + 0.24, + 0.845, + 0.256 + ], + "angle": 0, + "content": "{a.sain, p.chowdhury, s.koley, a.bhunia, y song}@surrey.ac.uk; Subhajit@ucf.edu" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.29, + 0.327, + 0.307 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.323, + 0.486, + 0.716 + ], + "angle": 0, + "content": "As sketch research has collectively matured over time, its adaptation for at-mass commercialisation emerges on the immediate horizon. Despite an already mature research endeavour for photos, there is no research on the efficient inference specifically designed for sketch data. In this paper, we first demonstrate existing state-of-the-art efficient light-weight models designed for photos do not work on sketches. We then propose two sketch-specific components which work in a plug-n-play manner on any photo efficient network to adapt them to work on sketch data. We specifically chose fine-grained sketch-based image retrieval (FG-SBIR) as a demonstrator as the most recognised sketch problem with immediate commercial value. Technically speaking, we first propose a cross-modal knowledge distillation network to transfer existing photo efficient networks to be compatible with sketch, which brings down number of FLOPs and model parameters by \\(97.96\\%\\) percent and \\(84.89\\%\\) respectively. We then exploit the abstract trait of sketch to introduce a RL-based canvas selector that dynamically adjusts to the abstraction level which further cuts down number of FLOPs by two thirds. The end result is an overall reduction of \\(99.37\\%\\) of FLOPs (from 40.18G to 0.254G) when compared with a full network, while retaining the accuracy \\((33.03\\%)\\) vs \\(32.77\\%\\) - finally making an efficient network for the sparse sketch data that exhibit even fewer FLOPs than the best photo counterpart." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.733, + 0.222, + 0.749 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.753, + 0.485, + 0.875 + ], + "angle": 0, + "content": "A significant movement in the late computer vision literature has been that of model efficiency. This effort has been largely driven by the on-the-edge deployments of vision models. With great strides made on networks specifically tackling photo data (MobileNet [54], EfficientNet [63], etc), there however has been surprisingly no work targeting human sketches [72]. This is despite the abundance of work specifically addressing sketch data [47] and its unique traits:" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.289, + 0.907, + 0.419 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.432, + 0.907, + 0.488 + ], + "angle": 0, + "content": "Figure 1. Our SketchyNetV1 compresses existing heavy FG-SBIR networks to deliver smaller models. Further enhanced via a canvas selector module our SketchyNetV2 model minimises sketch-resolution dynamically to reduce FLOPs." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.493, + 0.906, + 0.599 + ], + "angle": 0, + "content": "sequential [22], abstract [60], or stroke-wise [10], style-diversity [51], and data-scarcity [6, 20, 43]. This is particularly disappointing for the problem of fine-grained sketch-based image retrieval (FG-SBIR), which is perhaps the single-most studied sketch task [2, 7, 9, 13, 40, 43, 53, 75] that has already matured for at-mass commercial adoption, and is our problem of choice in this paper." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.599, + 0.907, + 0.811 + ], + "angle": 0, + "content": "One would think that state-of-the-art models designed for photos would work as-is for sketches, treating sketch data as raster images (which sketch is not). As our first contribution, we preempt this thought via a pilot study to prove its false nature. In particular, we borrow MobileNetv2 [54] as a backbone network to construct a typical triplet loss based network for FG-SBIR. Results show a reduction of almost \\(37\\%\\) (relative) in retrieval accuracy when compared with a typical VGG-16 backbone [57]. The focus of this paper is however not to design entirely new sketch-specific efficient networks from scratch – lots of good efficient photo networks already exist. Our goal instead is to adapt existing photo networks to work with sketch, by introducing generic sketch-specific designs that work in a plug-n-play manner." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.909, + 0.902 + ], + "angle": 0, + "content": "We first tackle the problem highlighted in the said pilot study – to make existing photo efficient networks perform on the same level as large networks on FG-SBIR. We argue the reason behind why the likes of MobileNetV2 did not work off-the-shelf is that the limited set of parameters could not accommodate the fine-grained nature of sketches (i.e.," + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.887, + 0.409, + 0.901 + ], + "angle": 0, + "content": "*Work done as an intern at SketchX before joining UCF." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28383" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.379 + ], + "angle": 0, + "content": "sketches are highly expressive in capturing fine-grained visual details [72]) for sketch and photo matching. For that, we introduce a knowledge distillation (KD) paradigm [48] to adapt this sketch trait onto efficient photo networks. The core idea is to induce strong semantic knowledge between a sketch and its photo, in the embedding space of a student, from a deep and high performing teacher. This particularly helps the efficient model to achieve on par performance with the deep variant without having a large parameter size. However, ours being a cross-modal retrieval, naively adapting standard KD paradigms involving logit distillation [3] would be sub-optimal. We therefore aim to preserve cross-modal pairwise distances among instances, and in turn the structural consistency of the teacher's discriminative embedding space in the student. The result is already an efficient FG-SBIR model (which we term SketchyNetV1) that reduces the number of FLOPs from 40.18G (VGG-16 based model[51]) to 0.833G (our MobileNetV2-based) – a roughly 48 times reduction (Fig. 1 left)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.383, + 0.483, + 0.866 + ], + "angle": 0, + "content": "This abstract nature of human sketch [16] can also be exploited, to further reduce the FLOPs. The hypothesis is that since sketches are essentially abstract depictions of images, they should carry a similar level of semantic information even when rendered at much lower resolutions, and that more abstract a sketch is the more information compact it is (i.e., can sustain even lower resolutions). For that, we conducted another pilot study (Sec. 3), where we show unlike photos, sketches can sustain retrieval accuracy with much lower resolutions (4% vs 15% at \\(32 \\times 32\\)), and 30% of those query sketches achieving perfect retrieval at full-resolution, when downsized to \\(32 \\times 32\\) can still retrieve accurately (Fig. 2 right). This motivated us to design a canvas selector that can choose the right resolution on-the-fly to render a sketch before feeding to the network. This is thanks to another unique trait of sketch data - they come in vector format and can be freely rendered to any resolution without extra overhead. Accomplishing this however is nontrivial since this selection process of canvas-size involves a discrete decision, thus invoking non-differentiability. We thus leverage reinforcement learning [9] to bypass these issues and train an abstraction-aware canvas-size selector as a policy network for our model, which estimates the optimal canvas-size required for the input sketch. The FG-SBIR model, acting as a critic, calculates reward for the canvas selector, balancing between accuracy and FLOPs, that trains the policy network. With this further trait of sketches addressed, our final model (SketchyNetV2) was able to further reduce FLOPs by a factor of 3.28 (from 0.833G to 0.254G), while retaining a retrieval accuracy similar to the full model (33.03 vs 32.77), exactly as one would imagine sketch efficient networks to entail (sketches being sparse)." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.483, + 0.903 + ], + "angle": 0, + "content": "In summary, our contributions are: (a) the first investigation towards efficient sketch networks from existing photo" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.183 + ], + "angle": 0, + "content": "ones (b) a knowledge distillation framework specifically designed to cater to the fine-grained nature of sketches, (c) a canvas-size selector that caters to the abstract nature of sketches, which further reduces FLOPs by two-thirds. Extensive experiments show our method to perform at par with existing state-of-the-arts, while being superiorly efficient." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.197, + 0.655, + 0.212 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.219, + 0.905, + 0.492 + ], + "angle": 0, + "content": "Fine-grained SBIR: Compared to category-level SBIR [17, 19, 36, 52], Fine-Grained SBIR (FG-SBIR) focuses on retrieving a single photo instance from a category-specific gallery based on a query sketch. Introduced as a deep triplet-ranking Siamese network [75], FG-SBIR has evolved with attention-based modules [59], textual tags [58], hybrid cross-domain generation [42], hierarchical coattention [50], and various pre-training strategies like mixed modal jigsaw solving [44], reinforcement learning [6], and transformer-based knowledge distillation [53]. Addressing sketch-specific traits like style diversity [51], data scarcity [7], and sketch stroke redundancy [9], improvements have been made for enhanced retrieval [53]. Recently it was extended to scene-level (retrieve a scene-image, given a scene-sketch), employing cross-modal region associativity [13] and enhanced using text-queries [46]. However, sketch-specific computational efficiency has largely been ignored - which we address here in this paper." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.494, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Towards Computational Efficiency: Addressing large-scale deployment, research emerging in computational efficiency [1], include network pruning [37], quantisation [21], binarisation [18], knowledge distillation [25], input-size optimisation [62], or their combinations [1]. Network pruning [23, 68] involves discovering and dropping low impact weights or channels of a pre-trained large network, while retaining its original accuracy. Contrarily, quantisation directly reduces the bit-width of parameter-values [21] and gradients [77], by replacing floating-point computations with faster and cheaper low-precision fixed-point numbers [34]. Its extreme version, binarisation [18], binarises weights and activations. Knowledge Distillation aims at transferring knowledge of a large pre-trained teacher network to a smaller student for cost-effective deployment. Existing techniques leverage output logits [3], hidden layers [48], attention-maps [76], or neuron selectivity pattern [28] of pre-trained teachers for the same. Self-distillation [4] employs the same network for both student and teacher models, to further reduce compute. Input-resolution optimisation on the other hand involves replacing a high-resolution input image with its down-scaled low-resolution version [62], to reduce overall compute [73], while retaining accuracy. Relevant works explored dynamic usage of high-resolution data [64] or patch proposals for strategic image-cropping [69], where the learnable resizer [62] operates on the original image resolution. All such" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28384" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.095, + 0.084, + 0.48, + 0.096 + ], + "angle": 0, + "content": "Table 1. Performance & computational cost for various networks" + }, + { + "type": "table", + "bbox": [ + 0.109, + 0.096, + 0.465, + 0.219 + ], + "angle": 0, + "content": "
ModelFLOPSParametersTime (ms)Acc@1
VGG-19 [57]51.06G20.02M10133.63
VGG-16 [57]40.18G14.71M08433.03
ResNet-101 [24]20.50G42.50M15521.77
DenseNet-201 [26]11.40G18.09M33523.62
ResNet-50 [24]10.76G23.51M08621.32
InceptionV3 [75]07.72G21.79M13424.27
ResNet-18 [24]4.76G11.18M03620.72
EfficientNet [63]1.05G04.01M11222.47
MobileNetV2 [54]0.83G02.22M07020.85
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.224, + 0.483, + 0.299 + ], + "angle": 0, + "content": "methods however cater to compressing image-specific models, which if used off-the-shelf for FG-SBIR would fail to address the sparse manner in which sketches are represented [8]. We thus propose a sketch-specific light-weight FG-SBIR model, tackling sparsity in sketches." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.299, + 0.484, + 0.451 + ], + "angle": 0, + "content": "Reinforcement Learning in Vision (RL): Applying RL [29] to computer vision problems [33, 67] is commonplace of late. Whenever quantifying the goodness of the network's state is non-differentiable, like selecting which regions to crop in an image [31], RL comes in handy. In sketch community, RL has found use in retrieval [6, 7], modelling sketch abstraction [39, 40], designing competitive sketching agents [5] and as a stroke selector [9]. Here, we leverage RL for optimising a canvas-size selector for a computationally efficient retrieval model." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.458, + 0.21, + 0.474 + ], + "angle": 0, + "content": "3. Pilot Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.476, + 0.483, + 0.718 + ], + "angle": 0, + "content": "Backbone Network for FG-SBIR: Amongst popular backbones used in recent FG-SBIR literature [6, 7, 50, 51], are VGG-16 [30] and Inception-V3 [75]. Others like ResNet18 [24] and EfficientNet [63], though cheaper and prevalent in vision [24] are rarely used in FG-SBIR. Exploring their potential in FG-SBIR, we train a baseline FG-SBIR model, plugging in popular ones as backbones for our study. Table 1 reports their parameter-count, FLOPs, time for feature-extraction per sample and Top-1 accuracy (Sec. 5) on standard QMUL-ShoeV2 dataset [75]. Despite scoring more than VGG-16 on image-classification [65], ResNet50 [24] performs poorly – indicating that modelling fine-grained sketch-photo association needs networks with higher FLOPs. Despite delivering better accuracies, heavier networks cost far more resources. We thus aim for a lighter model that achieves accuracy like a larger one." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Dynamic Sketch-canvas-size: Unlike photos that hold pixel-dense information, sketches are sparse black and white lines. This begs the question, if a sketch rendered at a higher resolution with added computational burden would convey any extra semantic information than at a lower one. To answer this, we focus on vector-format of sketches [7], where every sketch vector \\( s_v \\) can be characterised as a sequence of points \\( (v_1, v_2, \\dots, v_{\\mathrm{T}}) \\), and rendered to any canvas-size \\( (c) \\) as a raster-image \\( (s_r) \\) via rasterisation \\( \\mathcal{R}^c(\\cdot): s_v \\to s_r^c \\) where \\( s_r^c \\in \\mathbb{R}^{c \\times c \\times 3} \\). A single baseline FG-SBIR model (VGG-16) is trained and tested on full-resolution images, and sketches rendered at varying" + }, + { + "type": "image", + "bbox": [ + 0.544, + 0.093, + 0.716, + 0.183 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.722, + 0.091, + 0.882, + 0.185 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.527, + 0.187, + 0.892, + 0.201 + ], + "angle": 0, + "content": "Figure 2. Pilot study for varying canvas-size (see text above)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.209, + 0.908, + 0.587 + ], + "angle": 0, + "content": "canvas-sizes from \\( C = \\{32 \\times 32, \\dots, 256 \\times 256\\} \\). Comparing sparse sketches versus photos, we train an equivalent fine-grained image-based image retrieval (FG-IBIR) where only the query against a positive photo is constructed using random augmentation [11] and resized to corresponding canvas-size \\( c_i \\). Gallery photo-features for both are kept pre-computed. While FG-IBIR accuracy falls rapidly (Fig. 2 left), FG-SBIR stays relatively stable against decreasing canvas-sizes, as photos (unlike sketches) containing pixel-dense perfect information, lose a lot of it while down-scaling. Furthermore, positive accuracy of FG-SBIR at \\( 32 \\times 32 \\), shows some sketches to hold sufficient semantic information for retrieval even at minimal canvas-size. Fig. 2 (right) shows overall QMUL-ShoeV2[75] statistics where each bar represents the smallest canvas-size, at which bar-height percentage of sketches achieve perfect retrieval (taking total sketches achieving perfect retrieval at \\( 256 \\times 256 \\) as \\( 100\\% \\)). This infers that every sketch has its own optimal canvas, and fixing it (Fig. 2 left), for the entire dataset would always be sub-optimal. Moreover, the optimal canvas-size for a sketch, depends on its extent of sparsity, for which neither hard-ground-truth exists, nor is it knowable from the user's end. Finally, brute force discovery at every size being infeasible during inference, motivates us to design a learnable adaptive canvas-size selector for sketches." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.591, + 0.649, + 0.608 + ], + "angle": 0, + "content": "4. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.613, + 0.907, + 0.869 + ], + "angle": 0, + "content": "Overview: Our pilot study motivates us to extend a standard FG-SBIR framework towards computational efficiency by first introducing SketchyNetV1 via knowledge distillation, and then taking it further to SketchyNetV2 via dynamic sketch canvas-size selection. Importantly, these are done without hurting the original performance of large-scale slower baseline model. In particular, there are two phases of improvement for computational efficiency. Firstly, we train a teacher model \\( F_{T} \\) as a standard baseline FG-SBIR model (Sec. 3). Once trained, its rich semantic knowledge is transferred to a smaller student network \\( F_{st} \\) via a knowledge distillation (KD) paradigm, which we name as SketchyNetV1. Next, we train a canvas-size selector that adaptively decides optimal canvas-size or resolution per sketch for SketchyNetV2 that would be sufficient to preserve the retrieval performance while reducing the number of FLOP operations in feature extraction significantly." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Baseline FG-SBIR model: Keeping aside complex pretraining [44] or joint-training [7], we focus on a simple" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "28385" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.09, + 0.41, + 0.235 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.412, + 0.091, + 0.895, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.24, + 0.908, + 0.298 + ], + "angle": 0, + "content": "Figure 3. Two stage framework. (Left) SketchyNetV1: A smaller student network \\((F_{st})\\) is trained from a larger pre-trained teacher \\((F_T)\\) via KD (§4.1). (Right) We progress to SketchyNetV2 by training a canvas-size selector directed by objectives of increasing performance and reducing compute. It takes sketch as a vector \\((s_v)\\) and aims to predict an optimal canvas-size at which the sketch is rasterised \\((s_r)\\). \\(s_r\\) when processed by \\(F_{st}\\) minimises overall FLOPS, while retaining accuracy of corresponding full-resolution sketch-image." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.309, + 0.486, + 0.504 + ], + "angle": 0, + "content": "triple-branch state-of-the-art Siamese network [75] which remains a strong baseline so far [7, 50, 51]. A VGG-16 [30] pre-trained on ImageNet [49] acts as the shared backbone for each branch. Given an input image \\(I\\), a \\(d\\) dimensional feature embedding is extracted from the backbone network \\(\\mathcal{F}\\) using global average pooling followed by \\(l_{2}\\) normalisation as \\(f_{I} \\in \\mathbb{R}^{d}\\). The model is trained on triplets comprising features of an anchor sketch \\((f_{s})\\), its matching positive \\((f_{p})\\) photo and a random non-matching negative photo \\((f_{n})\\), via a triplet loss [70]. Triplet loss \\((\\mathcal{L}_{\\mathrm{Tri}})\\) aims to minimise the distance between sketch and positive photo, while maximising the same from its negative (n). With \\(\\delta(a, b) = ||a - b||^2\\), as a distance function and a margin \\(m > 0\\), we have:" + }, + { + "type": "equation", + "bbox": [ + 0.154, + 0.508, + 0.483, + 0.523 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {T r i}} = \\max \\left\\{0, m + \\delta \\left(f _ {s}, f _ {p}\\right) - \\delta \\left(f _ {s}, f _ {n}\\right) \\right\\} \\tag {1}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.535, + 0.422, + 0.551 + ], + "angle": 0, + "content": "4.1. SketchyNetV1: Smaller Model via KD" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.554, + 0.485, + 0.902 + ], + "angle": 0, + "content": "Based on our pilot study (Sec. 3), if we naively train from an ImageNet pre-trained standard smaller network (e.g. MobileNetV2 [54] - usually popular for edge devices), with simple supervised triplet loss (Eq.1) it can hardly reach the accuracy (Table 1) obtained by a large network like VGG-16. Therefore apart from training via a supervised loss on the sketch-positive-negative triplet \\((\\mathcal{L}_{\\mathrm{Tri}}^{st})\\) as in Eq.1), we aim to use the pre-trained larger teacher (e.g. VGG-16) for additional supervision. Accordingly, we adhere to Knowledge Distillation (KD) paradigm [48], a conventional strategy for model compression, to deliver a light-weight student model from the larger teacher network [48]. Applying KD to our FG-SBIR paradigm is however non-trivial. Unlike traditional KD methods, which often involve logit distillation for classification tasks, our scenario involves cross-modal retrieval. Here, the output is a continuous \\(d\\)-dimensional feature within a joint-embedding space. Furthermore, straightforward regression between teacher and student features for sketch and photo branches may encounter compatibility issues, given the disparity in the embedding spaces. Addressing this, if the dimensions of teacher and student embeddings differ, an additional feature transformation layer is necessary for alignment [48]. Overcoming these issues," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.309, + 0.907, + 0.369 + ], + "angle": 0, + "content": "we aim to distil the inter-feature distances from respective teacher networks to the student thus preserving structural orientation of features in the teacher's embedding space to that in the student, during KD (Fig. 3 (left))." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.37, + 0.909, + 0.522 + ], + "angle": 0, + "content": "Given features of sketch-positive-negative triplet of teacher \\((f_s^T, f_p^T, f_n^T)\\), we calculate the inter-feature Euclidean distances in the output representation space of teacher \\((F_T)\\) using \\(\\delta(\\cdot, \\cdot)\\) (Eq.1), as \\(d_{sp}^T = \\delta(f_s^T, f_p^T)\\); \\(d_{sn}^T = \\delta(f_s^T, f_n)\\) and \\(d_{pn}^T = \\delta(f_p^T, f_n^T)\\). Passing the same samples through the student, we similarly obtain \\(d_{sp}^{st}\\), \\(d_{sn}^{st}\\) and \\(d_{pn}^{st}\\) using \\((f_s^{st}, f_p^{st}, f_n^{st})\\). A smooth \\(l_1\\) loss distils the \\(l_2\\) distances from the teacher to the student. With \\(\\mathcal{L}_{\\delta}\\) as a Huber-loss [27], relational distillation loss for sketch-photo pair between teacher and student is given as," + }, + { + "type": "equation", + "bbox": [ + 0.584, + 0.525, + 0.773, + 0.542 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {R K D}} ^ {s p} = \\mathcal {L} _ {\\delta} \\left(d _ {s p} ^ {T}, d _ {s p} ^ {s t}\\right) \\quad \\text {w h e r e ,}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.567, + 0.546, + 0.905, + 0.581 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\delta} (a, b) = \\left\\{ \\begin{array}{l l} \\frac {1}{2} (a - b) ^ {2} & \\text {i f} | a - b | < \\beta \\\\ \\beta (| a - b | - \\frac {1}{2} \\beta) & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.587, + 0.908, + 0.63 + ], + "angle": 0, + "content": "Similarly, we obtain \\(\\mathcal{L}_{\\mathrm{RKD}}^{sn}\\) and \\(\\mathcal{L}_{\\mathrm{RKD}}^{pn}\\), to compute distillation loss, \\(\\mathcal{L}_{\\mathrm{RKD}} = \\mathcal{L}_{\\mathrm{RKD}}^{sp} + \\mathcal{L}_{\\mathrm{RKD}}^{sn} + \\mathcal{L}_{\\mathrm{RKD}}^{pn}\\). With hyperparameter \\(\\lambda\\), overall student-training objective becomes:" + }, + { + "type": "equation", + "bbox": [ + 0.622, + 0.636, + 0.905, + 0.651 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {t r n}} ^ {s t} = \\lambda \\mathcal {L} _ {\\mathrm {T r i}} + (1 - \\lambda) \\mathcal {L} _ {\\mathrm {R K D}} \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.654, + 0.907, + 0.725 + ], + "angle": 0, + "content": "The student trains on varying sketch-canvas-sizes \\((C)\\), using their respective full resolution for teacher, to ensure scale-invariance as \\(\\mathcal{L}_{\\mathrm{tm}}^{st*} = \\frac{1}{4}\\sum_{i=1}^{4}\\mathcal{L}_{\\mathrm{tm}}^{st(c_i)}\\). However, as optimal canvas-size varies across sketches (Fig. 2), we need a learnable canvas-size selector, dynamically predicting the optimal size for each sketch." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.73, + 0.898, + 0.746 + ], + "angle": 0, + "content": "4.2. SketchyNetV2: Adaptive canvas-size Selector" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.749, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Overview: Understanding that an input at a lower resolution, would incur lesser FLOPs on evaluation than a higher one, we upgrade SketchyNetV1 (\\(F_{st}\\)) to a more computationally efficient SketchyNetV2 by introducing an adaptive canvas-size selector. It is designed in the vector modality as it is much cheaper [71] than its raster-image counterpart, and can dynamically encode the varying abstraction level of sketches. Furthermore, predicting the optimal canvas-size from vector modality, would make the corresponding rasterisation operation \\(\\mathcal{R}^c(\\cdot)\\) at lower sizes cheaper, besides reducing FLOPs in feature extraction. Taking sketch vector \\(s_v\\) as input, \\(\\psi_C\\) predicts the probability \\(p(c|s_v)\\) of choosing one of \\(K\\)" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "28386" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.134 + ], + "angle": 0, + "content": "discreet canvas-sizes from a set of \\( C = \\{32 \\times 32, \\dots, 256 \\times 256\\} \\). \\( s_v \\) is then rasterised to \\( s_r^c \\in \\mathbb{R}^{c \\times c \\times 3} \\) with the optimal canvas-size \\( c \\), as \\( \\mathcal{R}^c(\\cdot) : s_v \\to s_r^c \\), and fed to \\( F_{st} \\) for retrieval." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.134, + 0.483, + 0.273 + ], + "angle": 0, + "content": "Selecting an optimal canvas-size however is an ill-posed problem. Firstly, there are no explicit labels representing the optimal canvas-size for a sketch. Secondly, annotating the optimal canvas-size for the whole dataset via brute-force iteration is computationally impractical. We therefore use the pre-trained \\( F_{st} \\) as a critic to guide the learning of sketch canvas-size selector via Reinforcement Learning [29] as rasterisation is a non-differentiable operation. There are two major objectives here: a) retain the original resolution accuracy and b) encourage the use of the lower resolution to improve the computational efficiency." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.276, + 0.485, + 0.594 + ], + "angle": 0, + "content": "Model: Any sequential network from RNN [17] or Transformer [35] families can be used to encode the vector-sketch \\( s_v \\). To recap, sketch-vector \\( (s_v) \\) represents a sequence of points \\( [v_1, v_2, \\dots, v_{\\mathrm{T}}] \\), where \\( v_t = (x_t, y_t, q_t^1, q_t^2, q_t^3) \\in \\mathbb{R}^{T \\times 5} \\); \\( \\mathrm{T} \\) is the sequence length, \\( (x_t, y_t) \\) denotes the absolute coordinates in a normalised \\( c \\times c \\) canvas, while the last three represent penstates [22]. Furthermore, the complexity of canvas-size selector being dependent on sequence-length (T) varying across sketches, we use Douglas Peucker Algorithm [66] to limit sequence-length across all sketch-samples at \\( \\mathrm{T}_{max} \\) without losing visual representation. We feed \\( v_t \\) at every time step to a simple 1-layer GRU network with \\( \\mathbb{R}^{d_v} \\) hidden states and take the final hidden state as the encoded latent representation \\( (f_{s^{\\mathrm{T}}} \\in \\mathbb{R}^{d_v}) \\). Next, we apply a linear layer \\( (\\gamma) \\) to get \\( p(c|s_v) = \\text{softmax}(W_\\gamma f_{s^{\\mathrm{T}}} + b_\\gamma) \\), where \\( W_\\gamma \\in \\mathbb{R}^{d_v \\times K} \\) and \\( b \\in \\mathbb{R}^K \\), as shown in Fig. 3 (right). Given \\( p(c|s_v) \\in \\mathbb{R}^K \\) over \\( K \\) different canvas-sizes for every \\( s_v \\), canvas-size \\( c \\) can be sampled from categorical distribution as \\( c_{pred} \\sim \\text{categorical}([p(c_1|s_v), \\dots, p(c_K|s_v)]) \\). Accordingly, \\( s_v \\) is rasterised to \\( s_r \\) at canvas-size \\( c_{pred} \\), which is then fed to \\( F_{st} \\). In particular, given a sketch \\( s_v \\), canvas-size selector \\( \\psi_C \\) acting as a policy network, takes action of selecting the canvas-size \\( c_{pred} \\) from action space \\( C \\), and \\( \\psi_C \\) is optimised over a reward calculated by the SketchyNetV1 model \\( (F_{st}) \\) acting as the critic." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.597, + 0.483, + 0.803 + ], + "angle": 0, + "content": "Reward Design: The objectives of our canvas selector are to choose an optimal canvas-size such that: (i) accuracy is retained (ii) overall FLOPs is minimised. We therefore design our reward from two perspectives. Now as \\(F_{st}\\) is fixed, we pre-compute the features of all gallery photos. During training, we only extract the feature of rasterised sketch and calculate its corresponding rank using the pre-computed photo-features. Furthermore, taking one sketch we pre-compute FLOPs for every canvas-size in \\(C\\). From accuracy perspective, the aim is to select the optimal canvas-size \\(c_{\\mathrm{opt}}\\) from \\(C\\) for each \\(s_v\\) such that the rendered \\(s_r\\) can retrieve the matching photo \\(p\\) at numerically lowest rank \\(r\\) (best \\(r = 1\\)). Following conventional norm of reward maximisation, we define accuracy reward \\((R_{\\mathrm{acc}})\\) as weighted (by hyper-parameters \\(\\lambda_r\\), \\(\\lambda_{\\mathrm{Tri}}\\)) summation of inverse of the rank (r) and negative triplet loss (following Eq.1) as:" + }, + { + "type": "equation", + "bbox": [ + 0.197, + 0.805, + 0.482, + 0.819 + ], + "angle": 0, + "content": "\\[\nR _ {\\mathrm {a c c}} = \\lambda_ {r} (1 / r) + \\lambda_ {\\mathrm {T r i}} (- \\mathcal {L} _ {\\mathrm {T r i}}) \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.818, + 0.483, + 0.901 + ], + "angle": 0, + "content": "From the compute perspective, the selection criterion should reward choosing a lower canvas-size. Moreover, higher performance being naturally inclined towards a higher-canvas-size thus increasing overall compute, we need an objective dedicated towards reducing computational cost. We thus propose a FLOPs constraint regularisation to guide the learning of canvas selector as:" + }, + { + "type": "equation", + "bbox": [ + 0.643, + 0.088, + 0.905, + 0.122 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {F}} = \\frac {\\sum_ {j = 1} ^ {K} \\left(q _ {j} \\cdot \\eta_ {j}\\right)}{q _ {\\max } - q _ {\\min }} \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.127, + 0.905, + 0.207 + ], + "angle": 0, + "content": "where, \\(\\eta_{i} = p(c_{i}|s_{v})\\) and \\(q_{j}\\) is the pre-computed FLOPs value for the j-th canvas-size with maximum at \\(q_{max}\\) and minimum at \\(q_{min}\\). Treating \\((-L_{\\mathrm{F}})\\) as a reward from the compute perspective, we define our compute reward as \\(R_{\\mathrm{comp}} = (-L_{\\mathrm{F}})\\). Taking \\(\\lambda_{\\mathrm{F}}\\) as a balancing hyper-parameter, we combine both perspectives as our total reward:" + }, + { + "type": "equation", + "bbox": [ + 0.613, + 0.207, + 0.905, + 0.221 + ], + "angle": 0, + "content": "\\[\nR _ {\\mathrm {T o t}} = \\lambda_ {\\mathrm {F}} R _ {\\mathrm {c o m p}} + (1 - \\lambda_ {\\mathrm {F}}) R _ {\\mathrm {a c c}} \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.227, + 0.905, + 0.323 + ], + "angle": 0, + "content": "Optimisation Objective: Finally, we train the canvas-size selector using popular Policy Gradient [61] method which is both simple to implement and requires less hyper-parameter tuning than other state-of-the-art alternatives [56]. If \\( p(c|s_v^i) \\) be the probability of selecting a specific canvas-size for \\( i^{th} \\) sketch sample \\( s_v^i \\) and the corresponding reward be \\( R_{\\mathrm{tot}}^i \\), the objective function for policy network (canvas-size selector) over a batch of size \\( B \\) becomes:" + }, + { + "type": "equation", + "bbox": [ + 0.594, + 0.327, + 0.905, + 0.363 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {P G} (\\theta) = - \\frac {1}{B} \\sum_ {i = 1} ^ {B} \\log p \\left(c \\mid s _ {v} ^ {i}\\right) \\cdot R _ {\\text {T o t}} ^ {i} \\tag {7}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.361, + 0.646, + 0.377 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.382, + 0.907, + 0.548 + ], + "angle": 0, + "content": "Datasets: We evaluated on the following publicly available datasets having fine-grained sketch-photo association. QMUL-ShoeV2 [75] and QMUL-ChairV2 [75] contains 6730/2000 and 1800/400 sketches/photos respectively. Keeping 679/200 (525/100) sketches/photos for evaluation from ShoeV2 (ChairV2) respectively we use the rest for training. Sketchy [55] contains 125 categories with 100 photos each, having around 5 sketches per photo. While, training uses a standard train-test split [55] of 9:1, during inference we construct a challenging gallery using photos across one category for retrieval. For scene-level FG-SBIR evaluation we use FS-COCO [15] which includes 10,000 unique sketch-photo pairs with a 7:3 train/test split." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.552, + 0.907, + 0.8 + ], + "angle": 0, + "content": "Implementation Details: Considering a standard large FG-SBIR network as a teacher, our proposed student uses a cheaper backbone feature extractor MobileNetV2 [54], a standard for mobile end-devices. Margin is set to \\( m = 0.2 \\) for triplet loss (Secs. 3 and 4.1). We use an Adam optimiser at a learning rate of 0.0001 and a batchsize of 16 for 200 epochs. For practicality, we chose discrete canvas-sizes (\\( C \\)) instead of a continuous action space, where \\( C = \\{32 \\times 32, 64 \\times 64, 128 \\times 128, 256 \\times 256\\} \\). Canvas-size selector (\\( \\psi_C \\)) is modelled using a simple GRU network, with embedding size 128, and trained via reinforcement learning with a learning rate of 0.0001 and a batchsize of 32 for 500 epochs, keeping \\( T_{max} = 100 \\) for vector-sketches. For the action space, we take varying canvas-sizes from \\( C \\) and set \\( \\beta = 1 \\) in Eq.2. \\( \\lambda \\), \\( \\lambda_r \\), \\( \\lambda_{\\mathrm{Tri}} \\) and \\( \\lambda_{\\mathrm{F}} \\) are empirically set to 0.5, 0.4, 0.48 and 0.35 (Fig.5) respectively. While training \\( \\psi_C \\), sketches are rendered at different completion (30%, 35%, ... 100%) levels to simulate varying levels of abstraction, ensuring a sketch-abstraction-aware canvas selector (e.g., a smaller canvas for less detailed sketch)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.804, + 0.909, + 0.9 + ], + "angle": 0, + "content": "Evaluation Metric: Accuracy for FG-SBIR is the percentage of sketches where true-match photos are ranked in the Top-q (q = 1,10) lists. Computational efficiency is measured via parameter-count and average FLOPs – lower is better. For average FLOPs in Table 2, we take the combined sum of FLOPs per sample [12] from (i) canvas-size selector and (ii) retrieval model, using the predicted resolution, and average across the validation set." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28387" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.285, + 0.09, + 0.713, + 0.103 + ], + "angle": 0, + "content": "Table 2. Quantitative Analysis on FG-SBIR. Best viewed when zoomed." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.104, + 0.904, + 0.52 + ], + "angle": 0, + "content": "
MethodsCanvas-size c×cParams (mil.)ShoeV2 [75]ChairV2 [75]Sketchy [55]FSCOCO [15]
Top1 (%)Top10 (%)FLOPS (G)Top1 (%)Top10 (%)FLOPS (G)Top1 (%)Top10 (%)FLOPs (G)Top1 (%)Top10 (%)FLOPS (G)
State-of-the-ArtsTriplet-SN [75]32x328.7509.77 (↓18.94)24.82 (↓46.74)0.08315.61 (↓32.04)29.86 (↓54.38)0.0834.23 (↓11.12)12.58 (↓23.13)0.0830.76 (↓3.94)6.22 (↓14.8)0.083
64x648.7517.63 (↓11.08)44.96 (↓26.60)0.33830.25 (↓17.4)54.21 (↓30.03)0.3389.68 (↓5.67)22.45 (↓13.26)0.3381.85 (↓2.85)7.33 (↓13.69)0.338
128x1288.7525.01 (↓3.70)62.35 (↓9.21)1.39740.61 (↓7.04)72.68 (↓11.56)1.39712.98 (↓2.37)29.63 (↓6.08)1.3973.28 (↓1.42)15.92 (↓5.1)1.397
256x2568.7528.7171.565.28047.6584.245.28015.3535.715.2804.721.025.280
SketchyNetV12.2228.46 (↓0.25)69.01 (↓2.55)0.83346.28 (↓0.37)82.33 (↓1.91)0.83314.85 (↓0.5)34.68 (↓1.03)0.8334.22 (↓0.48)18.33 (↓2.69)0.833
SketchyNetV22.2727.89 (↓0.82)68.76 (↓2.80)0.26445.98 (↓1.67)80.14 (↓4.10)0.29414.21 (↓1.14)34.16 (↓1.55)0.3213.98 (↓0.72)17.64 (↓3.38)0.423
HOLEE-SN [59]32x329.3110.84 (↓20.9)27.28 (↓48.5)0.09618.72 (↓34.69)31.68 (↓55.88)0.0966.01 (↓10.69)14.12 (↓24.78)0.0960.86 (↓4.04)6.93 (↓14.78)0.096
64x649.3119.83 (↓11.91)51.72 (↓24.06)0.38734.32 (↓19.09)56.46 (↓31.1)0.3879.92 (↓6.78)24.38 (↓14.52)0.3872.11 (↓2.79)7.86 (↓13.85)0.387
128x1289.3126.78 (↓4.96)66.28 (↓9.5)1.54545.74 (↓7.67)75.59 (↓11.97)1.54513.81 (↓2.89)33.68 (↓5.22)1.5453.65 (↓1.25)16.32 (↓5.39)1.545
256x2569.3131.7475.785.75853.4187.565.75816.7038.905.7584.921.715.758
SketchyNetV12.2231.59 (↓0.15)73.96 (↓1.82)0.83353.27 (↓0.14)86.93 (↓0.63)0.83316.25 (↓0.45)38.21 (↓0.69)0.8334.56 (↓0.34)19.92 (↓1.79)0.833
SketchyNetV22.2730.86 (↓0.88)72.56 (↓3.22)0.25952.74 (↓0.67)84.02 (↓3.54)0.28915.91 (↓0.79)37.88 (↓1.02)0.3154.04 (↓0.86)18.63 (↓3.08)0.417
Triplet-RL [6]32x3222.111.98 (↓22.12)28.11 (↓50.71)0.14220.53 (↓36.01)32.59 (↓57.02)0.1421.76 (↓2.94)3.88 (↓6.47)0.142
64x6422.121.28 (↓12.82)54.71 (↓24.11)0.57736.09 (↓20.45)57.33 (↓32.28)0.5773.04 (↓1.66)6.79 (↓3.56)0.577
128x12822.128.83 (↓5.27)67.84 (↓10.98)2.29948.38 (↓8.16)77.25 (↓12.36)2.2994.12 (↓0.58)9.05 (↓1.30)2.299
256x25622.134.1078.826.04156.5489.616.0414.7010.356.041
SketchyNetV12.2233.88 (↓0.22)77.15 (↓1.67)0.83355.92 (↓0.62)88.20 (↓1.41)0.8334.59 (↓0.11)10.21 (↓0.14)0.833
SketchyNetV22.2733.26 (↓0.84)76.84 (↓1.98)0.25555.14 (↓1.40)87.31 (↓2.30)0.2854.51 (↓0.19)9.89 (↓0.46)0.308
StyleVAE [51]32x3225.3712.68 (↓23.79)28.69 (↓53.14)0.12522.48 (↓40.38)32.10 (↓59.04)0.12506.92 (↓12.7)16.25 (↓29.53)0.125
64x6425.3722.93 (↓13.54)57.71 (↓24.12)0.50840.26 (↓22.6)58.23 (↓32.91)0.50812.46 (↓7.16)29.62 (↓16.16)0.508
128x12825.3730.91 (↓5.56)70.06 (↓11.77)2.02453.88 (↓8.98)78.64 (↓12.5)2.02417.15 (↓2.47)38.84 (↓6.94)2.024
256x25625.3736.4781.835.64262.8691.145.64219.6245.785.642
SketchyNetV12.2236.11 (↓0.36)79.63 (↓2.20)0.83362.23 (↓0.63)88.95 (↓2.19)0.83319.48 (↓0.12)44.02 (↓1.76)0.833
SketchyNetV22.2735.88 (↓0.59)78.46 (↓3.37)0.25161.95 (↓0.91)87.68 (↓3.46)0.28119.10 (↓0.50)43.57 (↓2.21)0.304
Partial-OT [14]32x3222.1
64x6422.1
128x12822.1
256x25622.1
SketchyNetV12.22
SketchyNetV22.27
B-BFR224x2244.6124.3261.232.60243.3976.332.60210.4724.342.6022.1511.612.602
B-DRSDynamic9.3625.1268.845.49344.7878.146.18711.8726.586.2453.8617.636.291
B-CropDynamic4.9220.3252.126.56235.0757.686.59708.2720.196.6291.9809.676.688
B-Regress [48]Dynamic2.2722.8261.330.26138.4264.650.29108.6921.980.3232.0710.150.375
B-AKD [76]Dynamic2.2723.6764.260.26838.6171.880.29512.6727.710.3312.9114.010.392
B-PKT [45]Dynamic2.2724.3165.910.25942.6873.660.28812.2728.310.3173.4516.220.372
B-VGG16-SN256x25614.7133.0378.5140.1852.1684.0840.1818.6141.2540.185.1623.6240.18
B-ThiNet [37]256x2568.327.51 (↓25.52)22.34 (↓56.17)9.34212.31 (↓39.85)26.63 (↓57.45)9.3423.16 (↓15.45)12.73 (↓28.52)9.3420.65 (↓4.51)5.35 (↓18.27)9.342
B-Prune [38]256x2566.173.92 (↓29.11)17.32 (↓61.19)8.1387.98 (↓44.18)19.82 (↓64.26)8.1381.45 (↓17.16)8.22 (↓33.03)8.1380.53 (↓4.63)4.97 (↓18.65)8.138
B-VGG16-SN SketchyNetV22.2732.77 (↓0.26)78.21 (↓0.3)0.25450.75 (↓1.41)82.21 (↓1.87)0.28316.55 (↓2.06)38.66 (↓2.59)0.3324.69 (↓0.47)21.76 (↓1.86)0.407
" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.531, + 0.226, + 0.547 + ], + "angle": 0, + "content": "5.1. Competitors" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.554, + 0.486, + 0.903 + ], + "angle": 0, + "content": "We evaluate against State-of-the-Arts (SoTAs) and a few relevant curated baselines (B-). (i) SoTAs: Unlike Triplet-SN [75], HOLEF-SN [59], Triplet-RL [6] (we report full-sketch evaluation) and StyleVAE [51]; Partial-OT [14] specifically caters to scene-level FG-SBIR. For fairness, we vary input-resolution of sketches during inference (fixed at their original \\(256 \\times 256\\) setup during training). We also adapt them as teachers, to judge our accuracy retainment after model-compression (SketchyNetV1) using full-res sketch and, further optimising canvas-size with canvas selector (SketchyNetV2). (ii) KD baselines: Here we re-purpose existing works in a cross-modal setting, keeping our canvas selector same, with Triplet-SN as the teacher. B-Regress - regresses features from teacher and student for both sketches and images, aligning their dimensions with an extra transformation layer [48], using a \\(l_{2}\\) regression loss. Following [45] B-PKT calculates the conditional probability density for any pair of points in the teacher's embedding space [45], which models the probability of any two samples being close together. Sampling 'N' instances, it establishes a probability distribution for pairwise interactions within that space. Simultaneously, in the student's embedding space, a comparable distribution is derived, and the model minimises the divergence between them using a KL-divergence loss. B-AKD follows [76] in using spatial attention maps for knowledge transfer using similar dimensional matching like B-Regress. (iii) Resizing strategies: Keeping the same KD-paradigm with Triplet-SN-teacher, we vary the input" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.532, + 0.909, + 0.755 + ], + "angle": 0, + "content": "size resizing module following existing works. \\( B \\)-BFR follows [62] in designing a bilinear feature resizer amidst the intermediate layers of a CNN network, to output a down-scaled image. We choose its resized resolution of \\( 224 \\times 224 \\) (lowest) for subsequent evaluation. \\( B \\)-DRS follows [12] in training a convolutional module that takes an image in full-resolution, calculates a probability vector on available canvas-sizes, and transforms it into binary decisions, indicating the scale factor of selection. \\( B \\)-Crop follows [32] in meta-learning an image-cropping module that crops the input image to a target resolution for low compute evaluation. (iv) Pruning alternatives: Here we adapt two VGG16-backboned methods off-the-shelf from network-pruning family among other efficiency paradigms [18, 21, 62], in FG-SBIR setup - \\( B \\)-ThiNet [37] and \\( B \\)-Prune [38]. For comparative reference we report performance of a VGG16-backboned Triplet-SN network (\\( B \\)-VGG16-SN) and its SketchyNetV2 variant." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.758, + 0.833, + 0.774 + ], + "angle": 0, + "content": "5.2. Performance and Compute Analysis:" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.776, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Analysis on Accuracy: Table 2 reports the quantitative evaluation of our model against others. Triplet-SN [75] and HOLEF-SN [59] perform lower owing to weak backbones of Sketch-A-Net [74]. Enhanced by its RL-optimised reward function Triplet-RL[6] scores higher \\((2.36\\% \\uparrow\\) Top1 on ShoeV2) but fails to surpass StyleVAE [51], due to its delicate meta-learning disentanglement module addressing style-diversity. Being trained on region-wise sketch-photo associativity, Partial-OT [14] outperforms all others significantly on scene-level sketches (FS-COCO). We also notice a" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "28388" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.273 + ], + "angle": 0, + "content": "gradual drop in accuracy of SoTAs, when subjected to decreasing resolution of input-sketches suggesting that some methods are inherently accustomed to handling scale-shifts better than others, but are inferior to our dynamic approach (SketchyNetV2). Importantly, when most SoTAs are employed as teachers, their respective students (SketchyNetV1 variants) are seen to retain their respective original accuracy with a marginal drop of \\(\\approx 0.5 - 1\\%\\) overall, thanks to our efficient KD paradigm. Having different teacher-student embedding spaces both \\(B\\)-Regress and \\(B\\)-AKD's paradigm of directly regressing on features proves incompatible to a cross-modal retrieval setting, dropping accuracy (compared to T-Triplet-SN), whereas lacking cross-modal discrimination \\(B\\)-PKT performs lower than ours, proving our superiority in retaining accuracy." + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.278, + 0.191, + 0.325 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.195, + 0.279, + 0.288, + 0.325 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.29, + 0.279, + 0.388, + 0.325 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.39, + 0.279, + 0.482, + 0.325 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.118, + 0.33, + 0.456, + 0.344 + ], + "angle": 0, + "content": "Figure 4. Exemplar sketches at their optimal canvas-size" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.35, + 0.485, + 0.822 + ], + "angle": 0, + "content": "Analysis on Compute: Although FLOPs [12] decrease (Table 2) with a reduction in input resolution across all SoTAs, it comes at a severe cost of accuracy (\\(\\approx 21\\%\\) average drop from \\(256 \\times 256\\) to \\(32 \\times 32\\) on ShoeV2). This is mainly because ideal canvas-size for optimal retrieval varies across samples (Fig. 2 right) - thus subjecting all samples to a fixed canvas-size proves detrimental for accuracy. However, a positive Acc@1 score of \\(\\approx 11\\%\\) even at a low canvas-size of \\(32 \\times 32\\) for every SoTA shows that some sketches are indeed recognisable if the model is trained properly. This is verified further by SketchyNetV2 variant of respective SoTAs, that secures the lowest FLOPs per SoTA, despite maintaining accuracy at par with its SoTA-counterpart - thanks to our RL-guided canvas selector, which can allot an optimal canvas-size per sketch towards better retrieval (Fig. 4). Contrarily, B-BFR fails to lower FLOPs as it fixes a \\(224 \\times 224\\) input resolution for a satisfactory Acc@1 compared to Triplet-SN. B-Crop fairs better, however sketch being sparse, cropped sketches leads to confused representations, lowering accuracy. Despite coming closest in accuracy to ours on Triplet-SN, B-DRS lags behind considerably on compute, thanks to our canvas selector working in vector space (inputs \\(s_v\\)) and eliminating costly convolutional operations, unlike others using full-resolution sketch-raster as input. Being untrained to handle sparse-nature of sketches during such high compression, performance of both B-ThiNet and B-Prune collapses (vs. B-VGG16-SN) unlike its SketchyNetV2. Notably, the significant drop in parameter-count (4 - 10 times) against SoTAs, comes from choosing lightweight MobileNetV2 [54] as a student coupled with our canvas-size selector in every SketchyNetV2 variant. Compared to ShoeV2, FLOPs increase slightly for ChairV2, Sketchy and FS-COCO, owing to higher object (ChairV2) or scene-level (FS-COCO) details, and lower structural correspondence with photos (Sketchy), thus training a poorer retrieval model which in turn acts as a poor critic for our canvas selector. Our canvas selector alone takes a negligible 51.84K parameters with 0.02G FLOPs." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.828, + 0.245, + 0.843 + ], + "angle": 0, + "content": "5.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.845, + 0.484, + 0.902 + ], + "angle": 0, + "content": "As our teacher \\((F_T)\\) can be any pre-trained FG-SBIR network, we limit our ablation to justifying design choices of our student model and canvas-size selector. We use \\(B\\)-VGG16-SN as a reference \\(F_T\\) - Type-\\(\\emptyset\\) in Table 3." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.093, + 0.907, + 0.204 + ], + "angle": 0, + "content": "Training Canvas-Size selector \\(\\psi_C\\): To justify loss objectives of \\(\\psi_C\\) we evaluate on ShoeV2, eliminating \\(r^{-1}\\), \\(\\mathcal{L}_{\\mathrm{Tri}}^R\\), \\(\\mathcal{L}_{\\mathrm{F}}\\) one by one from Eq. 6 (Table 3). While eliminating accuracy rewards, \\((w / o \\text{ rank}^{-1}, w / o \\mathcal{L}_{\\mathrm{Tri}}^R)\\) increases constraint on lowering FLOPs, dropping accuracy, ignoring \\(\\mathcal{L}_{\\mathrm{F}}\\) \\((w / o \\mathcal{L}_F)\\) gains accuracy but costs high FLOPs. We thus optimally balance between the two using \\(\\lambda_{\\mathrm{F}}\\) (Fig. 5 left). One can also choose to train for faster retrieval (lower FLOPs), given a decent Acc@5 (Fig. 5 right)." + }, + { + "type": "image", + "bbox": [ + 0.535, + 0.209, + 0.69, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.699, + 0.209, + 0.86, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.546, + 0.276, + 0.873, + 0.29 + ], + "angle": 0, + "content": "Figure 5. Impact of varying \\(\\lambda_{\\mathrm{F}}\\) on Acc@1 and Acc@5" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.292, + 0.907, + 0.528 + ], + "angle": 0, + "content": "Architectural insights: To judge the efficiency of our student's backbone, we explore a few cheaper backbone-networks (Types IV, V in Table 3). Evidently, MobileNetV2 (Ours) balances optimally between compute and accuracy. As an option to vector input for canvas selector, we re-design it with a simple seven-layer CNN module following [78] to use the full-resolution sketch-image as its input, and select a target canvas-size to downscale it for retrieval. With extra computational cost coming from spatial domain, total FLOPs incurred is much higher (VI). Next we compare efficacy of sketch as input to \\(\\psi_C\\) with respect to types of vector format - offset-coordinate (VI) [22] vs. absolute ones (Ours). Turns out the former is better for encoding. Comparing sketch encoders among LSTM (VIII), GRU (Ours) and Transformer (IX), showed ours to be optimum empirically (Table 3). Furthermore, preponding recent pre-processing modules like stroke-subset selector of [9] to our canvas selector, yields \\(42.95\\%\\) (vs. \\(43.7\\%\\) [9]) Acc@1 ShoeV2, proving its off-the-shelf compatibility." + }, + { + "type": "table_caption", + "bbox": [ + 0.567, + 0.535, + 0.853, + 0.547 + ], + "angle": 0, + "content": "Table 3. Ablative studies (accuracy on ShoeV2)" + }, + { + "type": "table", + "bbox": [ + 0.526, + 0.548, + 0.888, + 0.702 + ], + "angle": 0, + "content": "
TypeFLOPsParamsShoeV2
(G)(mil.)Top1 (%)Top10 (%)
∅B-VGG16-SN40.1814.7133.0378.51
Iw/o rank-1-reward0.17325.3164.36
IIw/o \\( \\mathcal{L}_{\\text{Tri}}^R \\)-reward0.1822.2727.7869.32
IIIw/o \\( \\mathcal{L}_{\\text{F}} \\)-reward0.83332.9178.39
IVResNet18 backbone1.45111.1826.7268.18
VEfficientNet backbone0.4594.0129.4772.04
VIImage-based \\( \\psi_C \\)5.7307.6232.2177.68
VIIOffset \\( s_v \\)0.2552.2732.6178.40
VIIIDecoder-LSTM0.2682.2531.4177.32
IXDecoder-Tf0.3142.3431.9878.02
SketchyNetV20.2542.2732.7778.21
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.707, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Generalisability of canvas selector: To judge if the optimal canvas-size selected is uniform irrespective of the training method, we first train four canvas-selectors using four SoTAs of Table 2. Evaluating them on ShoeV2 [75], \\(89.21\\%\\) of test-set sketches in ShoeV2 to obtain the same optimal canvas size across four different models, trained on four different methods, which further proves that our canvas selector can work equally well as a pre-processing module for any new model. Furthermore, plugging a canvas selector trained with SketchyNetV1 (critic) from B-VGG16-SN with a different SketchyNetV1 from StyleVAE [51] during inference achieves comparable performance \\((35.69\\%)\\) Top-1) to a canvas selector trained on StyleVAE-SketchyNetV1 \\((35.88\\%)\\) Top-1) on ShoeV2. Importantly, \\(95.64\\%\\) of test-set sketches yield the same optimal canvas-size across both models," + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "28389" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.093, + 0.482, + 0.176 + ], + "angle": 0, + "content": "confirming the cross-model generalisability of our canvas selector. (ii) On training canvas selector with teacher \\( F_{T} \\) (for Triplet-SN) as critic, yields same optimal canvas-sizes as those for student \\( F_{S} \\), for \\( 94.21\\% \\) of test-set sketches in ShoeV2, while incurring comparable FLOPs reduction (0.268G) and accuracy (27.91% Top1) against \\( F_{S} \\) as critic." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.176, + 0.482, + 0.3 + ], + "angle": 0, + "content": "Comparing Inference times: We evaluated on a 12 GB Nvidia RTX 2080 Ti GPU, on ShoeV2, for inference time of Triplet-SN, HOLEF-SN, Triplet-RL, and StyleVAE (SOTAs) on full-sketch to obtain 37, 38, 36 and \\(42\\mathrm{ms}\\), against 10, 11, 11, and \\(12\\mathrm{ms}\\) for their SketchyNetV2 variants respectively. For practical validation, we evaluated onnx-INT8 version of Triplet-SN [75] and its SketchyNetV2 variant (ours) after post-training quantisation, on an iPhone13 via the AI-benchmark app for ShoeV2 [75] to obtain 51ms and 18ms respectively, which shows our efficiency." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.303, + 0.482, + 0.429 + ], + "angle": 0, + "content": "Retraining SoTA on mixed resolution sketches?: To explore if retraining SoTAs on sketches with mixed resolutions, we re-train Triple-SN [75] on rendered mixed-resolution sketches of ShoeV2. Poor performance in Table 4 at lower resolutions similar to Table 2 validates our KD approach. The key lies in that the sketch-feature from teacher used for student-guidance \\((f_s^T)\\) is always extracted from full-resolution sketch while that extracted by student is at varying resolutions, imparting scale invariance, which is ideally suited to our cross-modal retrieval setup." + }, + { + "type": "table_caption", + "bbox": [ + 0.095, + 0.434, + 0.478, + 0.448 + ], + "angle": 0, + "content": "Table 4. Evaluating SoTAs trained on mixed-resolution sketches" + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.452, + 0.476, + 0.537 + ], + "angle": 0, + "content": "
Canvas SizeTriplet-SNTriplet-RL
Top1 (%)Top10 (%)FLOPs (G)Top1 (%)Top10 (%)FLOPs (G)
32×3210.9126.650.08312.9731.080.142
64×6419.0448.120.33822.9656.230.577
128×12826.0763.871.39730.0472.952.299
256×25628.7471.685.28034.2177.246.041
" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.544, + 0.254, + 0.559 + ], + "angle": 0, + "content": "5.4. Further Insights" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.562, + 0.482, + 0.659 + ], + "angle": 0, + "content": "Cross-dataset Generalisation: To judge the generalisation potential of our canvas-size selector across datasets we train it on ShoeV2, but plug it to a retrieval model (SketchyNetV1) trained on ChairV2, and evaluate on ChairV2. Against the standard result (Table 2) of training both modules (SketchyNetV2) on ChairV2 (0.285G, \\(55.32\\%\\) Acc@1), this setup retains accuracy at \\(53.14\\%\\) with slight rise in FLOPs (0.314G), thus confirming its potential." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.662, + 0.482, + 0.801 + ], + "angle": 0, + "content": "Faster Sketch-recognition: Exploring potential of our canvas-size selector, we prepend it to a standard sketch-recognition pipeline [74] to optimise the canvas-size of query sketch, and reduce overall compute. Unlike our \\(\\psi_{C}\\), it uses negative of cross-entropy loss as accuracy reward \\((R_{\\mathrm{acc}})\\), keeping rest same. Compared to existing Sketch-a-Net [74] incurring 5.28G FLOPs at \\(68.71\\%\\) accuracy on Quickdraw [22], our canvas selector fitted paradigm drops FLOPs to 0.261G while retaining accuracy at \\(68.14\\%\\). This shows for the first time, that sparsity of sketches can be handled, in various downstream tasks via our \\(\\psi_{C}\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.804, + 0.482, + 0.901 + ], + "angle": 0, + "content": "Extension to Category-level SBIR: For categorical SBIR we use Sketchy (ext) [36] dataset (90:10 train:test split), which holds \\(75k\\) sketches across 125 categories having \\(\\approx 73k\\) images [36] in total. It is evaluated using mean average precision (mAP@all) and precision at top 200 retrieval (P@200) [36]. Here we compete against state-of-the-arts (SoTA) like \\(B-D2S\\) that follows [19] without its zero-shot setting), and a simple baseline using a Siamese-style" + }, + { + "type": "table_caption", + "bbox": [ + 0.549, + 0.09, + 0.871, + 0.104 + ], + "angle": 0, + "content": "Table 5. Quantitative Evaluation on Categorical SBIR" + }, + { + "type": "table", + "bbox": [ + 0.539, + 0.108, + 0.88, + 0.221 + ], + "angle": 0, + "content": "
MethodsFLOPsParamsSketchy(ext)
(G)(mil.)mAP@allP@200
SoTAB-SBIR-SN40.1814.710.7150.861
DSH [36]22.1416.320.7110.858
B-D2S [19]17.75136.360.8100.894
StyleVAE [51]8.11625.370.9050.927
T-SoTAT-SBIR-SN0.2540.7020.810
T-DSH0.2892.270.7040.823
T-D2S0.2680.7980.845
T-StyleVAE0.2510.8860.897
" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.224, + 0.905, + 0.28 + ], + "angle": 0, + "content": "VGG-16 network (B-SBIR-SN) following [75]. Similar to Section 5.1 adapting SoTAs as teachers in our SketchyNetV2 paradigm (T-SoTA) shows them retaining their accuracy (Table 5) while significantly lowering compute." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.282, + 0.905, + 0.337 + ], + "angle": 0, + "content": "Teacher-Student Pairs: To explore other teacher-student pairs apart from ours, we test SketchyNetV2 models of a few \\((F_T,F_S)\\) pairs, where teachers are trained similar to B-VGG16-SN. From Table 6 we see our pair to outperform others." + }, + { + "type": "table_caption", + "bbox": [ + 0.575, + 0.343, + 0.845, + 0.356 + ], + "angle": 0, + "content": "Table 6. Evaluation on Teacher-Student Pairs" + }, + { + "type": "table", + "bbox": [ + 0.515, + 0.358, + 0.901, + 0.422 + ], + "angle": 0, + "content": "
Teacher\\StudentMobileNetV2EfficientNetResNet-18
Top-1FLOPsTop-1FLOPsTop-1FLOPs
VGG-16 [30]32.77%0.254G29.47%0.459G26.72%1.451G
Inception-V3 [74]27.89%0.264G27.16%0.461G25.31%1.452G
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.431, + 0.905, + 0.584 + ], + "angle": 0, + "content": "Combining compression methods: To explore if a combination of compression methods is capable of increasing efficiency, we subject our Knowledge Distillation (KD) trained model (SketchyNetV1) of Triplet-SN [75] (Table 2) to (i) quantisation following [21], and (ii) network-pruning following [38], separately, and then train their respective canvas-selectors. While the former secures \\(27.34\\%\\) (vs. \\(27.89\\%\\) ours) Top-1 score with a lower inference time of \\(9.1\\mathrm{ms}\\) (vs. \\(10\\mathrm{ms}\\) ours), the latter scores \\(26.91\\%\\) (vs. \\(27.89\\%\\) ms ours) Top-1 with lower FLOPs of 0.249G (vs. 0.264G ours). Such competitive results further endorse exploring combining compression methods as future works." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.586, + 0.906, + 0.712 + ], + "angle": 0, + "content": "Future Works: Sketch being an abstraction of an object's photo, when resized to a lower scale for retrieval, infers that its object's semantic is interpretable at that scale. This should ideally provide a lower bound for resizing a photo while retaining interpretability (e.g. high retrieval score). Consequently, designing a learnable photo-resizer aided by our trained sketch-canvas selector is a potential future work. Secondly, ours being a meta-framework, using more recent frameworks like DINOv2 [41] as large-scale teacher models for further gains is another targeted future work." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.717, + 0.634, + 0.732 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.906, + 0.901 + ], + "angle": 0, + "content": "We for the first time investigate the problem of efficient inference for FG-SBIR, where we show that using existing efficient lightweight networks directly, are unfit for sketches. We thus propose two generic components that can be adapted to photo-networks for sketch-specific data. Firstly, a cross-modal knowledge distillation paradigm distills the knowledge of larger models to a smaller one, directly reducing FLOPs (params) by 97.92 \\((84.9)\\%\\). Secondly, an abstraction-aware canvas-size selector dynamically selects the best canvas-size for a sketch, reducing FLOPs by extra two-thirds. Extensive experiments show our method to perform at par with state-of-the-arts using much lower compute thus further conveying our method's worth in enhancing them towards deployment." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28390" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.483, + 0.156 + ], + "angle": 0, + "content": "[1] Manoj Alwani, Yang Wang, and Vashisht Madhavan. Decore: Deep compression with reinforcement learning. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.158, + 0.483, + 0.225 + ], + "angle": 0, + "content": "[2] Aneeshan Sain and Ayan Kumar Bhunia and Pinaki Nath Chowdhury and Aneeshan Sain and Subhadeep Koley and Tao Xiang and Yi-Zhe Song. CLIP for All Things Zero-Shot Sketch-Based Image Retrieval, Fine-Grained or Not. In CVPR, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.228, + 0.483, + 0.254 + ], + "angle": 0, + "content": "[3] Jimmy Ba and Rich Caruana. Do deep nets really need to be deep? NeurIPS, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.256, + 0.483, + 0.311 + ], + "angle": 0, + "content": "[4] Hessam Bagherinezhad, Maxwell Horton, Mohammad Rastegari, and Ali Farhadi. Label refinery: Improving image-net classification through label progression. arXiv preprint arXiv:1805.02641, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.313, + 0.483, + 0.38 + ], + "angle": 0, + "content": "[5] Ayan Kumar Bhunia, Ayan Das, Umar Riaz Muhammad, Yongxin Yang, Timothy M Hospedales, Tao Xiang, Yulia Gryaditskaya, and Yi-Zhe Song. Pixel: A competitive sketching ai agent. so you think you can sketch? ACM TOG, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.383, + 0.483, + 0.436 + ], + "angle": 0, + "content": "[6] Ayan Kumar Bhunia, Yongxin Yang, Timothy M Hospedales, Tao Xiang, and Yi-Zhe Song. Sketch less for more: On-the-fly fine-grained sketch based image retrieval. In CVPR, 2020. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.439, + 0.483, + 0.506 + ], + "angle": 0, + "content": "[7] Ayan Kumar Bhunia, Pinaki Nath Chowdhury, Aaneshan Sain, Yongxin Yang, Tao Xiang, and Yi-Zhe Song. More photos are all you need: Semi-supervised learning for fine-grained sketch based image retrieval. In CVPR, 2021. 1, 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.509, + 0.483, + 0.564 + ], + "angle": 0, + "content": "[8] Ayan Kumar Bhunia, Pinaki Nath Chowdhury, Yongxin Yang, Timothy M Hospedales, Tao Xiang, and Yi-Zhe Song. Vectorization and rasterization: Self-supervised learning for sketch and handwriting. In CVPR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.565, + 0.483, + 0.633 + ], + "angle": 0, + "content": "[9] Ayan Kumar Bhunia, Subhadeep Koley, Abdullah Faiz Ur Rahman Khilji, Aneeshan Sain, Pinaki Nath Chowdhury, Tao Xiang, and Yi-Zhe Song. Sketching without worrying: Noise-tolerant sketch-based image retrieval. In CVPR, 2022. 1, 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.635, + 0.483, + 0.69 + ], + "angle": 0, + "content": "[10] Ayan Kumar Bhunia, Subhadeep Koley, Amandeep Kumar, Aneeshan Sain, Pinaki Nath Chowdhury, Tao Xiang, and Yi-Zhe Song. Sketch2Saliency: Learning to Detect Salient Objects from Human Drawings. In CVPR, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.691, + 0.483, + 0.732 + ], + "angle": 0, + "content": "[11] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In ICML, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.734, + 0.483, + 0.774 + ], + "angle": 0, + "content": "[12] Ying Chen, Liang Qiao, Zhanzhan Cheng, Shiliang Pu, Yi Niu, and Xi Li. Dynamic low-resolution distillation for cost-efficient end-to-end text spotting. In ECCV, 2022. 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.776, + 0.483, + 0.831 + ], + "angle": 0, + "content": "[13] Pinaki Nath Chowdhury, Ayan Kumar Bhunia, Viswanatha Reddy Gajjala, Aneeshan Sain, Tao Xiang, and Yi-Zhe Song. Partially does it: Towards scene-level fg-sbir with partial input. In CVPR, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.832, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[14] Pinaki Nath Chowdhury, Ayan Kumar Bhunia, Viswanatha Reddy Gajjala, Aneeshan Sain, Tao Xiang, and Yi-Zhe Song. Partially does it: Towards scene-level fg-sbir with partial input. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 6" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[15] Pinaki Nath Chowdhury, Aneeshan Sain, Ayan Kumar Bhunia, Tao Xiang, Yulia Gryaditskaya, and Yi-Zhe Song. Fscoco: Towards understanding of freehand sketches of common objects in context. In ECCV, 2022. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.15, + 0.905, + 0.205 + ], + "angle": 0, + "content": "[16] Chowdhury, Pinaki Nath and Bhunia, Ayan Kumar and Sain, Aneeshan and Koley, Subhadeep and Xiang, Tao and Song, Yi-Zhe. Democratising 2D Sketch to 3D Shape Retrieval Through Pivoting. In ICCV, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.208, + 0.905, + 0.247 + ], + "angle": 0, + "content": "[17] John Collomosse, Tu Bui, and Hailin Jin. Livesketch: Query perturbations for guided sketch-based visual search. In CVPR, 2019. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.251, + 0.905, + 0.292 + ], + "angle": 0, + "content": "[18] Matthieu Courbariaux, Yoshua Bengio, and Jean-Pierre David. Binaryconnect: Training deep neural networks with binary weights during propagations. NeurIPS, 28, 2015. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.294, + 0.905, + 0.335 + ], + "angle": 0, + "content": "[19] Sounak Dey, Pau Riba, Anjan Dutta, Josep Llados, and Yi-Zhe Song. Doodle to search: Practical zero-shot sketch-based image retrieval. In CVPR, 2019. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.337, + 0.905, + 0.378 + ], + "angle": 0, + "content": "[20] Anjan Dutta and Zeynep Akata. Semantically tied paired cycle consistency for zero-shot sketch-based image retrieval. In CVPR, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.381, + 0.905, + 0.422 + ], + "angle": 0, + "content": "[21] Suyog Gupta, Ankur Agrawal, Kailash Gopalakrishnan, and Pritish Narayanan. Deep learning with limited numerical precision. In ICML, 2015. 2, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.424, + 0.905, + 0.451 + ], + "angle": 0, + "content": "[22] David Ha and Douglas Eck. A neural representation of sketch drawings. In ICLR, 2018. 1, 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.453, + 0.905, + 0.495 + ], + "angle": 0, + "content": "[23] Song Han, Huizi Mao, and William J Dally. Deep compression: Compressing deep neural networks with pruning, trained quantization and huffman coding. In ICLR, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.497, + 0.905, + 0.537 + ], + "angle": 0, + "content": "[24] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.54, + 0.905, + 0.581 + ], + "angle": 0, + "content": "[25] Geoffrey Hinton, Oriol Vinyls, and Jeff Dean. Distilling the knowledge in a neural network. In NeurIPS Deep Learning Workshop, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.584, + 0.905, + 0.624 + ], + "angle": 0, + "content": "[26] Gao Huang, Zhuang Liu, Laurens Van Der Maaten, and Kilian Q Weinberger. Densely connected convolutional networks. In CVPR, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.627, + 0.905, + 0.654 + ], + "angle": 0, + "content": "[27] Shouyou Huang and Qiang Wu. Robust pairwise learning with huber loss. Journal of Complexity, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.656, + 0.905, + 0.697 + ], + "angle": 0, + "content": "[28] Zehao Huang and Naiyan Wang. Like what you like: Knowledge distill via neuron selectivity transfer. arXiv preprint arXiv:1707.01219, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.7, + 0.905, + 0.727 + ], + "angle": 0, + "content": "[29] Leslie Pack Kaelbling, Michael L Littman, and Andrew W Moore. Reinforcement learning: A survey. JAIR, 1996. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.73, + 0.905, + 0.77 + ], + "angle": 0, + "content": "[30] Andrew Zisserman Karen Simonyan. Very deep convolutional networks for large-scale image recognition. In ICLR, 2015. 3, 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.772, + 0.905, + 0.814 + ], + "angle": 0, + "content": "[31] Debang Li, Huikai Wu, Junge Zhang, and Kaiqi Huang. A2-rl: Aesthetics aware reinforcement learning for image cropping. In CVPR, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.817, + 0.905, + 0.857 + ], + "angle": 0, + "content": "[32] Debang Li, Junge Zhang, and Kaiqi Huang. Learning to learn cropping models for different aspect ratio requirements. In CVPR, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.859, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[33] Xiaodan Liang, Lisa Lee, and Eric P Xing. Deep variation-structured reinforcement learning for visual relationship and attribute detection. In CVPR, 2017. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "28391" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.134 + ], + "angle": 0, + "content": "[34] Darryl Lin, Sachin Talathi, and Sreekanth Annapureddy. Fixed point quantization of deep convolutional networks. In ICML, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.483, + 0.19 + ], + "angle": 0, + "content": "[35] Hangyu Lin, Yanwei Fu, Yu-Gang Jiang, and Xiangyang Xue. Sketch-bert: Learning sketch bidirectional encoder representation from transformers by self-supervised learning of sketch gestalt. In CVPR, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.192, + 0.482, + 0.234 + ], + "angle": 0, + "content": "[36] Li Liu, Fumin Shen, Yuming Shen, Xianglong Liu, and Ling Shao. Deep sketch hashing: Fast free-hand sketch-based image retrieval. In CVPR, 2017. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.236, + 0.482, + 0.275 + ], + "angle": 0, + "content": "[37] Jian-Hao Luo, Jianxin Wu, and Weiyao Lin. Thinet: A filter level pruning method for deep neural network compression. In ICCV, 2017. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.278, + 0.482, + 0.332 + ], + "angle": 0, + "content": "[38] Pavlo Molchanov, Stephen Tyree, Tero Karras, Timo Aila, and Jan Kautz. Pruning convolutional neural networks for resource efficient inference. arXiv preprint arXiv:1611.06440, 2016. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.334, + 0.482, + 0.375 + ], + "angle": 0, + "content": "[39] Umar Riaz Muhammad, Yongxin Yang, Yi-Zhe Song, Tao Xiang, and Timothy M Hospedales. Learning deep sketch abstraction. In CVPR, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.377, + 0.482, + 0.418 + ], + "angle": 0, + "content": "[40] Umar Riaz Muhammad, Yongxin Yang, Timothy Hospedales, Tao Xiang, and Yi-Zhe Song. Goal-driven sequential data abstraction. In ICCV, 2019. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.42, + 0.482, + 0.489 + ], + "angle": 0, + "content": "[41] Maxime Oquab, Timothee Darct, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. DINOv2: Learning Robust Visual Features without Supervision. arXiv preprint arXiv:2304.07193, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.491, + 0.482, + 0.533 + ], + "angle": 0, + "content": "[42] Kaiyue Pang, Yi-Zhe Song, Tony Xiang, and Timothy M Hospedales. Cross-domain generative learning for fine-grained sketch-based image retrieval. In BMVC, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.534, + 0.482, + 0.587 + ], + "angle": 0, + "content": "[43] Kaiyue Pang, Ke Li, Yongxin Yang, Honggang Zhang, Timothy M Hospedales, Tao Xiang, and Yi-Zhe Song. Generalising fine-grained sketch-based image retrieval. In CVPR, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.59, + 0.482, + 0.644 + ], + "angle": 0, + "content": "[44] Kaiyue Pang, Yongxin Yang, Timothy M Hospedales, Tao Xiang, and Yi-Zhe Song. Solving mixed-modal jigsaw puzzle for fine-grained sketch-based image retrieval. In CVPR, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.647, + 0.482, + 0.687 + ], + "angle": 0, + "content": "[45] Nikolaos Passalis and Anastasios Tefas. Learning deep representations with probabilistic knowledge transfer. In ECCV, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.689, + 0.482, + 0.745 + ], + "angle": 0, + "content": "[46] Pinaki Nath Chowdhury and Ayan Kumar Bhunia and Aneeeshan Sain and Subhadeep Koley and Tao Xiang and Yi-Zhe Song. SceneTrilogy: On Human Scene-Sketch and its Complementarity with Photo and Text. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.747, + 0.482, + 0.801 + ], + "angle": 0, + "content": "[47] Pinaki Nath Chowdhury and Ayan Kumar Bhunia and Aeeeshan Sain and Subhadeep Koley and Tao Xiang and Yi-Zhe Song. What Can Human Sketches Do for Object Detection? In CVPR, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.803, + 0.482, + 0.844 + ], + "angle": 0, + "content": "[48] Adriana Romero, Nicolas Ballas, Samira Ebrahimi Kahou, Antonie Chassang, Carlo Gatta, and Yoshua Bengio. Fitnets: Hints for thin deep nets. In ICLR, 2015. 2, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[49] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. IJCV, 2015. 4" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[50] Aneeshan Sain, Ayan Kumar Bhunia, Yongxin Yang, Tao Xiang, and Yi-Zhe Song. Cross-modal hierarchical modelling, fine-grained sketch based image retrieval. In BMVC, 2020, 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.149, + 0.905, + 0.204 + ], + "angle": 0, + "content": "[51] Aneeshan Sain, Ayan Kumar Bhunia, Yongxin Yang, Tao Xiang, and Yi-Zhe Song. Stylemeup: Towards style-agnostic sketch-based image retrieval. In CVPR, 2021. 1, 2, 3, 4, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.206, + 0.905, + 0.26 + ], + "angle": 0, + "content": "[52] Aneeshan Sain, Ayan Kumar Bhunia, Vaishnav Potlapalli, Pinaki Nath Chowdhury, Tao Xiang, and Yi-Zhe Song. Sketch3t: Test-time training for zero-shot sbir. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.262, + 0.905, + 0.317 + ], + "angle": 0, + "content": "[53] Aneeshan Sain, Ayan Kumar Bhunia, Subhadeep Koley, Pinaki Nath Chowdhury, Soumitri Chattopadhyay, Tao Xiang, and Yi-Zhe Song. Exploiting Unlabelled Photos for Stronger Fine-Grained SBIR. In CVPR, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.319, + 0.905, + 0.373 + ], + "angle": 0, + "content": "[54] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zh-moginov, and Liang-Chieh Chen. Mobilenetv2: Inverted residuals and linear bottlenecks. In CVPR, 2018. 1, 3, 4, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.375, + 0.905, + 0.416 + ], + "angle": 0, + "content": "[55] Patsorn Sangkloy, Nathan Burnell, Cusuh Ham, and James Hays. The sketchy database: learning to retrieve badly drawn bunnies. ACM TOG, 2016. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.418, + 0.905, + 0.46 + ], + "angle": 0, + "content": "[56] John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.461, + 0.905, + 0.502 + ], + "angle": 0, + "content": "[57] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. In ICLR, 2015. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.504, + 0.905, + 0.544 + ], + "angle": 0, + "content": "[58] Jifei Song, Yi-Zhe Song, Tony Xiang, and Timothy M Hospedales. Fine-grained image retrieval: the text/sketch input dilemma. In BMVC, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.546, + 0.905, + 0.588 + ], + "angle": 0, + "content": "[59] Jifei Song, Qian Yu, Yi-Zhe Song, Tao Xiang, and Timothy M Hospedales. Deep spatial-semantic attention for fine-grained sketch-based image retrieval. In ICCV, 2017. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.589, + 0.905, + 0.644 + ], + "angle": 0, + "content": "[60] Subhadeep Koley and Ayan Kumar Bhunia and Aneeshan Sain and Pinaki Nath Chowdhury and Tao Xiang and Yi-Zhe Song. Picture that Sketch: Photorealistic Image Generation from Abstract Sketches. In CVPR, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.646, + 0.905, + 0.687 + ], + "angle": 0, + "content": "[61] Richard S Sutton, David A McAllester, Satinder P Singh, and Yishay Mansour. Policy gradient methods for reinforcement learning with function approximation. In NeurIPS, 2000. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.689, + 0.905, + 0.716 + ], + "angle": 0, + "content": "[62] Hossein Talebi and Peyman Milanfar. Learning to resize images for computer vision tasks. In ICCV, 2021. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.718, + 0.905, + 0.758 + ], + "angle": 0, + "content": "[63] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In ICML, 2019. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.76, + 0.905, + 0.8 + ], + "angle": 0, + "content": "[64] Burak Uzkent and Stefano Ermon. Learning when and where to zoom with deep reinforcement learning. In CVPR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.802, + 0.905, + 0.844 + ], + "angle": 0, + "content": "[65] Supawit Vatathanavaro, Suchat Tungjitnob, and Kitsuchart Pasupa. White blood cell classification: a comparison between vgg-16 and resnet-50 models. In JSCI, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.905, + 0.899 + ], + "angle": 0, + "content": "[66] Mahes Visvalingam and J Duncan Whyatt. The douglaspeucker algorithm for line simplification: re-evaluation through visualization. In Computer Graphics Forum, 1990. 5" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "28392" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.161 + ], + "angle": 0, + "content": "[67] Xin Wang, Qiuyuan Huang, Asli Celikyilmaz, Jianfeng Gao, Dinghan Shen, Yuan-Fang Wang, William Yang Wang, and Lei Zhang. Reinforced cross-modal matching and self-supervised imitation learning for vision-language navigation. In CVPR, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.164, + 0.482, + 0.207 + ], + "angle": 0, + "content": "[68] Yunhe Wang, Chang Xu, Shan You, Dacheng Tao, and Chao Xu. Cnnpack: Packing convolutional neural networks in the frequency domain. NeurIPS, 29, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.208, + 0.482, + 0.261 + ], + "angle": 0, + "content": "[69] Yulin Wang, Kangchen Lv, Rui Huang, Shiji Song, Le Yang, and Gao Huang. Glance and focus: a dynamic approach to reducing spatial redundancy in image classification. In NeurIPS, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.263, + 0.482, + 0.303 + ], + "angle": 0, + "content": "[70] Kilian Q Weinberger and Lawrence K Saul. Distance metric learning for large margin nearest neighbor classification. JMLR, 2009. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.305, + 0.482, + 0.346 + ], + "angle": 0, + "content": "[71] Peng Xu, Chaitanya K Joshi, and Xavier Bresson. Multigraph transformer for free-hand sketch recognition. IEEE T-NNLS, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.348, + 0.482, + 0.389 + ], + "angle": 0, + "content": "[72] Peng Xu, Timothy M Hospedales, Qiyue Yin, Yi-Zhe Song, Tao Xiang, and Liang Wang. Deep learning for free-hand sketch: A survey. IEEE TPAMI, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.391, + 0.482, + 0.445 + ], + "angle": 0, + "content": "[73] Ran Xu, Fangzhou Mu, Jayoung Lee, Preeti Mukherjee, Somali Chaterji, Saurabh Bagchi, and Yin Li. SmartAdapt: Multi-branch object detection framework for videos on mobiles. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.448, + 0.482, + 0.489 + ], + "angle": 0, + "content": "[74] Qian Yu, Yongxin Yang, Yi-Zhe Song, Tao Xiang, and Timothy M. Hospedales. Sketch-a-net that beats humans. In BMVC, 2015. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.491, + 0.482, + 0.532 + ], + "angle": 0, + "content": "[75] Qian Yu, Feng Liu, Yi-Zhe Song, Tao Xiang, Timothy M Hospedales, and Chen-Change Loy. Sketch me that shoe. In CVPR, 2016. 1, 2, 3, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.534, + 0.482, + 0.587 + ], + "angle": 0, + "content": "[76] Sergey Zagoruyko and Nikos Komodakis. Paying more attention to attention: Improving the performance of convolutional neural networks via attention transfer. In ICLR, 2017. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.59, + 0.482, + 0.645 + ], + "angle": 0, + "content": "[77] Shuchang Zhou, Yuxin Wu, Zekun Ni, Xinyu Zhou, He Wen, and Yuheng Zou. Dorefa-net: Training low bitwidth convolutional neural networks with low bitwidth gradients. arXiv preprint arXiv:1606.06160, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.647, + 0.482, + 0.687 + ], + "angle": 0, + "content": "[78] Mingjian Zhu, Kai Han, Enhua Wu, Qiulin Zhang, Ying Nie, Zhenzhong Lan, and Yunhe Wang. Dynamic resolution network. In NeurIPS, 2021. 7" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.687 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28393" + } + ] +] \ No newline at end of file diff --git a/2025/Sketch Down the FLOPs_ Towards Efficient Networks for Human Sketch/d51d8c32-98f5-44d4-9674-ee9bc3d4af26_origin.pdf b/2025/Sketch Down the FLOPs_ Towards Efficient Networks for Human Sketch/d51d8c32-98f5-44d4-9674-ee9bc3d4af26_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..b32ac488a0bd969b9101bb8c811c4d125d2645a2 --- /dev/null +++ b/2025/Sketch Down the FLOPs_ Towards Efficient Networks for Human Sketch/d51d8c32-98f5-44d4-9674-ee9bc3d4af26_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f09139743a705c483fb0e6841de3cd6809cd2d4902b11b24d74cdaa99fdc9c0a +size 2113034 diff --git a/2025/Sketch Down the FLOPs_ Towards Efficient Networks for Human Sketch/full.md b/2025/Sketch Down the FLOPs_ Towards Efficient Networks for Human Sketch/full.md new file mode 100644 index 0000000000000000000000000000000000000000..56a85b09f12480b8c4af844601571e8d1a0ff34a --- /dev/null +++ b/2025/Sketch Down the FLOPs_ Towards Efficient Networks for Human Sketch/full.md @@ -0,0 +1,313 @@ +# Sketch Down the FLOPs: Towards Efficient Networks for Human Sketch + +Aneeshan Sain1 Subhajit Maity2* Pinaki Nath Chowdhury1 Shubhadeep Koley1 Ayan Kumar Bhunia1 Yi-Zhe Song1 + +$^{1}$ SketchX, CVSSP, University of Surrey, United Kingdom. + +$^{2}$ Department of Computer Science, University of Central Florida, United States. + +{a.sain, p.chowdhury, s.koley, a.bhunia, y song}@surrey.ac.uk; Subhajit@ucf.edu + +# Abstract + +As sketch research has collectively matured over time, its adaptation for at-mass commercialisation emerges on the immediate horizon. Despite an already mature research endeavour for photos, there is no research on the efficient inference specifically designed for sketch data. In this paper, we first demonstrate existing state-of-the-art efficient light-weight models designed for photos do not work on sketches. We then propose two sketch-specific components which work in a plug-n-play manner on any photo efficient network to adapt them to work on sketch data. We specifically chose fine-grained sketch-based image retrieval (FG-SBIR) as a demonstrator as the most recognised sketch problem with immediate commercial value. Technically speaking, we first propose a cross-modal knowledge distillation network to transfer existing photo efficient networks to be compatible with sketch, which brings down number of FLOPs and model parameters by $97.96\%$ percent and $84.89\%$ respectively. We then exploit the abstract trait of sketch to introduce a RL-based canvas selector that dynamically adjusts to the abstraction level which further cuts down number of FLOPs by two thirds. The end result is an overall reduction of $99.37\%$ of FLOPs (from 40.18G to 0.254G) when compared with a full network, while retaining the accuracy $(33.03\%)$ vs $32.77\%$ - finally making an efficient network for the sparse sketch data that exhibit even fewer FLOPs than the best photo counterpart. + +# 1. Introduction + +A significant movement in the late computer vision literature has been that of model efficiency. This effort has been largely driven by the on-the-edge deployments of vision models. With great strides made on networks specifically tackling photo data (MobileNet [54], EfficientNet [63], etc), there however has been surprisingly no work targeting human sketches [72]. This is despite the abundance of work specifically addressing sketch data [47] and its unique traits: + +![](images/13cb94a39c9a758370bf172a5b52d260950fde46c6c3b4a2912aa0047ab2c5d4.jpg) +Figure 1. Our SketchyNetV1 compresses existing heavy FG-SBIR networks to deliver smaller models. Further enhanced via a canvas selector module our SketchyNetV2 model minimises sketch-resolution dynamically to reduce FLOPs. + +sequential [22], abstract [60], or stroke-wise [10], style-diversity [51], and data-scarcity [6, 20, 43]. This is particularly disappointing for the problem of fine-grained sketch-based image retrieval (FG-SBIR), which is perhaps the single-most studied sketch task [2, 7, 9, 13, 40, 43, 53, 75] that has already matured for at-mass commercial adoption, and is our problem of choice in this paper. + +One would think that state-of-the-art models designed for photos would work as-is for sketches, treating sketch data as raster images (which sketch is not). As our first contribution, we preempt this thought via a pilot study to prove its false nature. In particular, we borrow MobileNetv2 [54] as a backbone network to construct a typical triplet loss based network for FG-SBIR. Results show a reduction of almost $37\%$ (relative) in retrieval accuracy when compared with a typical VGG-16 backbone [57]. The focus of this paper is however not to design entirely new sketch-specific efficient networks from scratch – lots of good efficient photo networks already exist. Our goal instead is to adapt existing photo networks to work with sketch, by introducing generic sketch-specific designs that work in a plug-n-play manner. + +We first tackle the problem highlighted in the said pilot study – to make existing photo efficient networks perform on the same level as large networks on FG-SBIR. We argue the reason behind why the likes of MobileNetV2 did not work off-the-shelf is that the limited set of parameters could not accommodate the fine-grained nature of sketches (i.e., + +sketches are highly expressive in capturing fine-grained visual details [72]) for sketch and photo matching. For that, we introduce a knowledge distillation (KD) paradigm [48] to adapt this sketch trait onto efficient photo networks. The core idea is to induce strong semantic knowledge between a sketch and its photo, in the embedding space of a student, from a deep and high performing teacher. This particularly helps the efficient model to achieve on par performance with the deep variant without having a large parameter size. However, ours being a cross-modal retrieval, naively adapting standard KD paradigms involving logit distillation [3] would be sub-optimal. We therefore aim to preserve cross-modal pairwise distances among instances, and in turn the structural consistency of the teacher's discriminative embedding space in the student. The result is already an efficient FG-SBIR model (which we term SketchyNetV1) that reduces the number of FLOPs from 40.18G (VGG-16 based model[51]) to 0.833G (our MobileNetV2-based) – a roughly 48 times reduction (Fig. 1 left). + +This abstract nature of human sketch [16] can also be exploited, to further reduce the FLOPs. The hypothesis is that since sketches are essentially abstract depictions of images, they should carry a similar level of semantic information even when rendered at much lower resolutions, and that more abstract a sketch is the more information compact it is (i.e., can sustain even lower resolutions). For that, we conducted another pilot study (Sec. 3), where we show unlike photos, sketches can sustain retrieval accuracy with much lower resolutions (4% vs 15% at $32 \times 32$ ), and 30% of those query sketches achieving perfect retrieval at full-resolution, when downsized to $32 \times 32$ can still retrieve accurately (Fig. 2 right). This motivated us to design a canvas selector that can choose the right resolution on-the-fly to render a sketch before feeding to the network. This is thanks to another unique trait of sketch data - they come in vector format and can be freely rendered to any resolution without extra overhead. Accomplishing this however is nontrivial since this selection process of canvas-size involves a discrete decision, thus invoking non-differentiability. We thus leverage reinforcement learning [9] to bypass these issues and train an abstraction-aware canvas-size selector as a policy network for our model, which estimates the optimal canvas-size required for the input sketch. The FG-SBIR model, acting as a critic, calculates reward for the canvas selector, balancing between accuracy and FLOPs, that trains the policy network. With this further trait of sketches addressed, our final model (SketchyNetV2) was able to further reduce FLOPs by a factor of 3.28 (from 0.833G to 0.254G), while retaining a retrieval accuracy similar to the full model (33.03 vs 32.77), exactly as one would imagine sketch efficient networks to entail (sketches being sparse). + +In summary, our contributions are: (a) the first investigation towards efficient sketch networks from existing photo + +ones (b) a knowledge distillation framework specifically designed to cater to the fine-grained nature of sketches, (c) a canvas-size selector that caters to the abstract nature of sketches, which further reduces FLOPs by two-thirds. Extensive experiments show our method to perform at par with existing state-of-the-arts, while being superiorly efficient. + +# 2. Related Work + +Fine-grained SBIR: Compared to category-level SBIR [17, 19, 36, 52], Fine-Grained SBIR (FG-SBIR) focuses on retrieving a single photo instance from a category-specific gallery based on a query sketch. Introduced as a deep triplet-ranking Siamese network [75], FG-SBIR has evolved with attention-based modules [59], textual tags [58], hybrid cross-domain generation [42], hierarchical coattention [50], and various pre-training strategies like mixed modal jigsaw solving [44], reinforcement learning [6], and transformer-based knowledge distillation [53]. Addressing sketch-specific traits like style diversity [51], data scarcity [7], and sketch stroke redundancy [9], improvements have been made for enhanced retrieval [53]. Recently it was extended to scene-level (retrieve a scene-image, given a scene-sketch), employing cross-modal region associativity [13] and enhanced using text-queries [46]. However, sketch-specific computational efficiency has largely been ignored - which we address here in this paper. + +Towards Computational Efficiency: Addressing large-scale deployment, research emerging in computational efficiency [1], include network pruning [37], quantisation [21], binarisation [18], knowledge distillation [25], input-size optimisation [62], or their combinations [1]. Network pruning [23, 68] involves discovering and dropping low impact weights or channels of a pre-trained large network, while retaining its original accuracy. Contrarily, quantisation directly reduces the bit-width of parameter-values [21] and gradients [77], by replacing floating-point computations with faster and cheaper low-precision fixed-point numbers [34]. Its extreme version, binarisation [18], binarises weights and activations. Knowledge Distillation aims at transferring knowledge of a large pre-trained teacher network to a smaller student for cost-effective deployment. Existing techniques leverage output logits [3], hidden layers [48], attention-maps [76], or neuron selectivity pattern [28] of pre-trained teachers for the same. Self-distillation [4] employs the same network for both student and teacher models, to further reduce compute. Input-resolution optimisation on the other hand involves replacing a high-resolution input image with its down-scaled low-resolution version [62], to reduce overall compute [73], while retaining accuracy. Relevant works explored dynamic usage of high-resolution data [64] or patch proposals for strategic image-cropping [69], where the learnable resizer [62] operates on the original image resolution. All such + +Table 1. Performance & computational cost for various networks + +
ModelFLOPSParametersTime (ms)Acc@1
VGG-19 [57]51.06G20.02M10133.63
VGG-16 [57]40.18G14.71M08433.03
ResNet-101 [24]20.50G42.50M15521.77
DenseNet-201 [26]11.40G18.09M33523.62
ResNet-50 [24]10.76G23.51M08621.32
InceptionV3 [75]07.72G21.79M13424.27
ResNet-18 [24]4.76G11.18M03620.72
EfficientNet [63]1.05G04.01M11222.47
MobileNetV2 [54]0.83G02.22M07020.85
+ +methods however cater to compressing image-specific models, which if used off-the-shelf for FG-SBIR would fail to address the sparse manner in which sketches are represented [8]. We thus propose a sketch-specific light-weight FG-SBIR model, tackling sparsity in sketches. + +Reinforcement Learning in Vision (RL): Applying RL [29] to computer vision problems [33, 67] is commonplace of late. Whenever quantifying the goodness of the network's state is non-differentiable, like selecting which regions to crop in an image [31], RL comes in handy. In sketch community, RL has found use in retrieval [6, 7], modelling sketch abstraction [39, 40], designing competitive sketching agents [5] and as a stroke selector [9]. Here, we leverage RL for optimising a canvas-size selector for a computationally efficient retrieval model. + +# 3. Pilot Study + +Backbone Network for FG-SBIR: Amongst popular backbones used in recent FG-SBIR literature [6, 7, 50, 51], are VGG-16 [30] and Inception-V3 [75]. Others like ResNet18 [24] and EfficientNet [63], though cheaper and prevalent in vision [24] are rarely used in FG-SBIR. Exploring their potential in FG-SBIR, we train a baseline FG-SBIR model, plugging in popular ones as backbones for our study. Table 1 reports their parameter-count, FLOPs, time for feature-extraction per sample and Top-1 accuracy (Sec. 5) on standard QMUL-ShoeV2 dataset [75]. Despite scoring more than VGG-16 on image-classification [65], ResNet50 [24] performs poorly – indicating that modelling fine-grained sketch-photo association needs networks with higher FLOPs. Despite delivering better accuracies, heavier networks cost far more resources. We thus aim for a lighter model that achieves accuracy like a larger one. + +Dynamic Sketch-canvas-size: Unlike photos that hold pixel-dense information, sketches are sparse black and white lines. This begs the question, if a sketch rendered at a higher resolution with added computational burden would convey any extra semantic information than at a lower one. To answer this, we focus on vector-format of sketches [7], where every sketch vector $s_v$ can be characterised as a sequence of points $(v_1, v_2, \dots, v_{\mathrm{T}})$ , and rendered to any canvas-size $(c)$ as a raster-image $(s_r)$ via rasterisation $\mathcal{R}^c(\cdot): s_v \to s_r^c$ where $s_r^c \in \mathbb{R}^{c \times c \times 3}$ . A single baseline FG-SBIR model (VGG-16) is trained and tested on full-resolution images, and sketches rendered at varying + +![](images/2e9054e4b9b734386733cfe540dbcbb670b37b4422ac8e6b687c233bba54ab3a.jpg) +Figure 2. Pilot study for varying canvas-size (see text above). + +![](images/5b15e5befa59daae27edb06ff6dac5cfad0663005ac286bdaf13107c29d27f45.jpg) + +canvas-sizes from $C = \{32 \times 32, \dots, 256 \times 256\}$ . Comparing sparse sketches versus photos, we train an equivalent fine-grained image-based image retrieval (FG-IBIR) where only the query against a positive photo is constructed using random augmentation [11] and resized to corresponding canvas-size $c_i$ . Gallery photo-features for both are kept pre-computed. While FG-IBIR accuracy falls rapidly (Fig. 2 left), FG-SBIR stays relatively stable against decreasing canvas-sizes, as photos (unlike sketches) containing pixel-dense perfect information, lose a lot of it while down-scaling. Furthermore, positive accuracy of FG-SBIR at $32 \times 32$ , shows some sketches to hold sufficient semantic information for retrieval even at minimal canvas-size. Fig. 2 (right) shows overall QMUL-ShoeV2[75] statistics where each bar represents the smallest canvas-size, at which bar-height percentage of sketches achieve perfect retrieval (taking total sketches achieving perfect retrieval at $256 \times 256$ as $100\%$ ). This infers that every sketch has its own optimal canvas, and fixing it (Fig. 2 left), for the entire dataset would always be sub-optimal. Moreover, the optimal canvas-size for a sketch, depends on its extent of sparsity, for which neither hard-ground-truth exists, nor is it knowable from the user's end. Finally, brute force discovery at every size being infeasible during inference, motivates us to design a learnable adaptive canvas-size selector for sketches. + +# 4. Methodology + +Overview: Our pilot study motivates us to extend a standard FG-SBIR framework towards computational efficiency by first introducing SketchyNetV1 via knowledge distillation, and then taking it further to SketchyNetV2 via dynamic sketch canvas-size selection. Importantly, these are done without hurting the original performance of large-scale slower baseline model. In particular, there are two phases of improvement for computational efficiency. Firstly, we train a teacher model $F_{T}$ as a standard baseline FG-SBIR model (Sec. 3). Once trained, its rich semantic knowledge is transferred to a smaller student network $F_{st}$ via a knowledge distillation (KD) paradigm, which we name as SketchyNetV1. Next, we train a canvas-size selector that adaptively decides optimal canvas-size or resolution per sketch for SketchyNetV2 that would be sufficient to preserve the retrieval performance while reducing the number of FLOP operations in feature extraction significantly. + +Baseline FG-SBIR model: Keeping aside complex pretraining [44] or joint-training [7], we focus on a simple + +![](images/cacf7a15e1c70f3a7b612e9ee79238275db7e1d710603e601cd6b0d7674ca74a.jpg) +Figure 3. Two stage framework. (Left) SketchyNetV1: A smaller student network $(F_{st})$ is trained from a larger pre-trained teacher $(F_T)$ via KD (§4.1). (Right) We progress to SketchyNetV2 by training a canvas-size selector directed by objectives of increasing performance and reducing compute. It takes sketch as a vector $(s_v)$ and aims to predict an optimal canvas-size at which the sketch is rasterised $(s_r)$ . $s_r$ when processed by $F_{st}$ minimises overall FLOPS, while retaining accuracy of corresponding full-resolution sketch-image. + +![](images/1989c4a8b181a9d84247e7176b167e34eaa9155836d5351d6ec01a09df533f8f.jpg) + +triple-branch state-of-the-art Siamese network [75] which remains a strong baseline so far [7, 50, 51]. A VGG-16 [30] pre-trained on ImageNet [49] acts as the shared backbone for each branch. Given an input image $I$ , a $d$ dimensional feature embedding is extracted from the backbone network $\mathcal{F}$ using global average pooling followed by $l_{2}$ normalisation as $f_{I} \in \mathbb{R}^{d}$ . The model is trained on triplets comprising features of an anchor sketch $(f_{s})$ , its matching positive $(f_{p})$ photo and a random non-matching negative photo $(f_{n})$ , via a triplet loss [70]. Triplet loss $(\mathcal{L}_{\mathrm{Tri}})$ aims to minimise the distance between sketch and positive photo, while maximising the same from its negative (n). With $\delta(a, b) = ||a - b||^2$ , as a distance function and a margin $m > 0$ , we have: + +$$ +\mathcal {L} _ {\mathrm {T r i}} = \max \left\{0, m + \delta \left(f _ {s}, f _ {p}\right) - \delta \left(f _ {s}, f _ {n}\right) \right\} \tag {1} +$$ + +# 4.1. SketchyNetV1: Smaller Model via KD + +Based on our pilot study (Sec. 3), if we naively train from an ImageNet pre-trained standard smaller network (e.g. MobileNetV2 [54] - usually popular for edge devices), with simple supervised triplet loss (Eq.1) it can hardly reach the accuracy (Table 1) obtained by a large network like VGG-16. Therefore apart from training via a supervised loss on the sketch-positive-negative triplet $(\mathcal{L}_{\mathrm{Tri}}^{st})$ as in Eq.1), we aim to use the pre-trained larger teacher (e.g. VGG-16) for additional supervision. Accordingly, we adhere to Knowledge Distillation (KD) paradigm [48], a conventional strategy for model compression, to deliver a light-weight student model from the larger teacher network [48]. Applying KD to our FG-SBIR paradigm is however non-trivial. Unlike traditional KD methods, which often involve logit distillation for classification tasks, our scenario involves cross-modal retrieval. Here, the output is a continuous $d$ -dimensional feature within a joint-embedding space. Furthermore, straightforward regression between teacher and student features for sketch and photo branches may encounter compatibility issues, given the disparity in the embedding spaces. Addressing this, if the dimensions of teacher and student embeddings differ, an additional feature transformation layer is necessary for alignment [48]. Overcoming these issues, + +we aim to distil the inter-feature distances from respective teacher networks to the student thus preserving structural orientation of features in the teacher's embedding space to that in the student, during KD (Fig. 3 (left)). + +Given features of sketch-positive-negative triplet of teacher $(f_s^T, f_p^T, f_n^T)$ , we calculate the inter-feature Euclidean distances in the output representation space of teacher $(F_T)$ using $\delta(\cdot, \cdot)$ (Eq.1), as $d_{sp}^T = \delta(f_s^T, f_p^T)$ ; $d_{sn}^T = \delta(f_s^T, f_n)$ and $d_{pn}^T = \delta(f_p^T, f_n^T)$ . Passing the same samples through the student, we similarly obtain $d_{sp}^{st}$ , $d_{sn}^{st}$ and $d_{pn}^{st}$ using $(f_s^{st}, f_p^{st}, f_n^{st})$ . A smooth $l_1$ loss distils the $l_2$ distances from the teacher to the student. With $\mathcal{L}_{\delta}$ as a Huber-loss [27], relational distillation loss for sketch-photo pair between teacher and student is given as, + +$$ +\mathcal {L} _ {\mathrm {R K D}} ^ {s p} = \mathcal {L} _ {\delta} \left(d _ {s p} ^ {T}, d _ {s p} ^ {s t}\right) \quad \text {w h e r e ,} +$$ + +$$ +\mathcal {L} _ {\delta} (a, b) = \left\{ \begin{array}{l l} \frac {1}{2} (a - b) ^ {2} & \text {i f} | a - b | < \beta \\ \beta (| a - b | - \frac {1}{2} \beta) & \text {o t h e r w i s e} \end{array} \right. \tag {2} +$$ + +Similarly, we obtain $\mathcal{L}_{\mathrm{RKD}}^{sn}$ and $\mathcal{L}_{\mathrm{RKD}}^{pn}$ , to compute distillation loss, $\mathcal{L}_{\mathrm{RKD}} = \mathcal{L}_{\mathrm{RKD}}^{sp} + \mathcal{L}_{\mathrm{RKD}}^{sn} + \mathcal{L}_{\mathrm{RKD}}^{pn}$ . With hyperparameter $\lambda$ , overall student-training objective becomes: + +$$ +\mathcal {L} _ {\mathrm {t r n}} ^ {s t} = \lambda \mathcal {L} _ {\mathrm {T r i}} + (1 - \lambda) \mathcal {L} _ {\mathrm {R K D}} \tag {3} +$$ + +The student trains on varying sketch-canvas-sizes $(C)$ , using their respective full resolution for teacher, to ensure scale-invariance as $\mathcal{L}_{\mathrm{tm}}^{st*} = \frac{1}{4}\sum_{i=1}^{4}\mathcal{L}_{\mathrm{tm}}^{st(c_i)}$ . However, as optimal canvas-size varies across sketches (Fig. 2), we need a learnable canvas-size selector, dynamically predicting the optimal size for each sketch. + +# 4.2. SketchyNetV2: Adaptive canvas-size Selector + +Overview: Understanding that an input at a lower resolution, would incur lesser FLOPs on evaluation than a higher one, we upgrade SketchyNetV1 ( $F_{st}$ ) to a more computationally efficient SketchyNetV2 by introducing an adaptive canvas-size selector. It is designed in the vector modality as it is much cheaper [71] than its raster-image counterpart, and can dynamically encode the varying abstraction level of sketches. Furthermore, predicting the optimal canvas-size from vector modality, would make the corresponding rasterisation operation $\mathcal{R}^c(\cdot)$ at lower sizes cheaper, besides reducing FLOPs in feature extraction. Taking sketch vector $s_v$ as input, $\psi_C$ predicts the probability $p(c|s_v)$ of choosing one of $K$ + +discreet canvas-sizes from a set of $C = \{32 \times 32, \dots, 256 \times 256\}$ . $s_v$ is then rasterised to $s_r^c \in \mathbb{R}^{c \times c \times 3}$ with the optimal canvas-size $c$ , as $\mathcal{R}^c(\cdot) : s_v \to s_r^c$ , and fed to $F_{st}$ for retrieval. + +Selecting an optimal canvas-size however is an ill-posed problem. Firstly, there are no explicit labels representing the optimal canvas-size for a sketch. Secondly, annotating the optimal canvas-size for the whole dataset via brute-force iteration is computationally impractical. We therefore use the pre-trained $F_{st}$ as a critic to guide the learning of sketch canvas-size selector via Reinforcement Learning [29] as rasterisation is a non-differentiable operation. There are two major objectives here: a) retain the original resolution accuracy and b) encourage the use of the lower resolution to improve the computational efficiency. + +Model: Any sequential network from RNN [17] or Transformer [35] families can be used to encode the vector-sketch $s_v$ . To recap, sketch-vector $(s_v)$ represents a sequence of points $[v_1, v_2, \dots, v_{\mathrm{T}}]$ , where $v_t = (x_t, y_t, q_t^1, q_t^2, q_t^3) \in \mathbb{R}^{T \times 5}$ ; $\mathrm{T}$ is the sequence length, $(x_t, y_t)$ denotes the absolute coordinates in a normalised $c \times c$ canvas, while the last three represent penstates [22]. Furthermore, the complexity of canvas-size selector being dependent on sequence-length (T) varying across sketches, we use Douglas Peucker Algorithm [66] to limit sequence-length across all sketch-samples at $\mathrm{T}_{max}$ without losing visual representation. We feed $v_t$ at every time step to a simple 1-layer GRU network with $\mathbb{R}^{d_v}$ hidden states and take the final hidden state as the encoded latent representation $(f_{s^{\mathrm{T}}} \in \mathbb{R}^{d_v})$ . Next, we apply a linear layer $(\gamma)$ to get $p(c|s_v) = \text{softmax}(W_\gamma f_{s^{\mathrm{T}}} + b_\gamma)$ , where $W_\gamma \in \mathbb{R}^{d_v \times K}$ and $b \in \mathbb{R}^K$ , as shown in Fig. 3 (right). Given $p(c|s_v) \in \mathbb{R}^K$ over $K$ different canvas-sizes for every $s_v$ , canvas-size $c$ can be sampled from categorical distribution as $c_{pred} \sim \text{categorical}([p(c_1|s_v), \dots, p(c_K|s_v)])$ . Accordingly, $s_v$ is rasterised to $s_r$ at canvas-size $c_{pred}$ , which is then fed to $F_{st}$ . In particular, given a sketch $s_v$ , canvas-size selector $\psi_C$ acting as a policy network, takes action of selecting the canvas-size $c_{pred}$ from action space $C$ , and $\psi_C$ is optimised over a reward calculated by the SketchyNetV1 model $(F_{st})$ acting as the critic. + +Reward Design: The objectives of our canvas selector are to choose an optimal canvas-size such that: (i) accuracy is retained (ii) overall FLOPs is minimised. We therefore design our reward from two perspectives. Now as $F_{st}$ is fixed, we pre-compute the features of all gallery photos. During training, we only extract the feature of rasterised sketch and calculate its corresponding rank using the pre-computed photo-features. Furthermore, taking one sketch we pre-compute FLOPs for every canvas-size in $C$ . From accuracy perspective, the aim is to select the optimal canvas-size $c_{\mathrm{opt}}$ from $C$ for each $s_v$ such that the rendered $s_r$ can retrieve the matching photo $p$ at numerically lowest rank $r$ (best $r = 1$ ). Following conventional norm of reward maximisation, we define accuracy reward $(R_{\mathrm{acc}})$ as weighted (by hyper-parameters $\lambda_r$ , $\lambda_{\mathrm{Tri}}$ ) summation of inverse of the rank (r) and negative triplet loss (following Eq.1) as: + +$$ +R _ {\mathrm {a c c}} = \lambda_ {r} (1 / r) + \lambda_ {\mathrm {T r i}} (- \mathcal {L} _ {\mathrm {T r i}}) \tag {4} +$$ + +From the compute perspective, the selection criterion should reward choosing a lower canvas-size. Moreover, higher performance being naturally inclined towards a higher-canvas-size thus increasing overall compute, we need an objective dedicated towards reducing computational cost. We thus propose a FLOPs constraint regularisation to guide the learning of canvas selector as: + +$$ +\mathcal {L} _ {\mathrm {F}} = \frac {\sum_ {j = 1} ^ {K} \left(q _ {j} \cdot \eta_ {j}\right)}{q _ {\max } - q _ {\min }} \tag {5} +$$ + +where, $\eta_{i} = p(c_{i}|s_{v})$ and $q_{j}$ is the pre-computed FLOPs value for the j-th canvas-size with maximum at $q_{max}$ and minimum at $q_{min}$ . Treating $(-L_{\mathrm{F}})$ as a reward from the compute perspective, we define our compute reward as $R_{\mathrm{comp}} = (-L_{\mathrm{F}})$ . Taking $\lambda_{\mathrm{F}}$ as a balancing hyper-parameter, we combine both perspectives as our total reward: + +$$ +R _ {\mathrm {T o t}} = \lambda_ {\mathrm {F}} R _ {\mathrm {c o m p}} + (1 - \lambda_ {\mathrm {F}}) R _ {\mathrm {a c c}} \tag {6} +$$ + +Optimisation Objective: Finally, we train the canvas-size selector using popular Policy Gradient [61] method which is both simple to implement and requires less hyper-parameter tuning than other state-of-the-art alternatives [56]. If $p(c|s_v^i)$ be the probability of selecting a specific canvas-size for $i^{th}$ sketch sample $s_v^i$ and the corresponding reward be $R_{\mathrm{tot}}^i$ , the objective function for policy network (canvas-size selector) over a batch of size $B$ becomes: + +$$ +\mathcal {L} _ {P G} (\theta) = - \frac {1}{B} \sum_ {i = 1} ^ {B} \log p \left(c \mid s _ {v} ^ {i}\right) \cdot R _ {\text {T o t}} ^ {i} \tag {7} +$$ + +# 5. Experiments + +Datasets: We evaluated on the following publicly available datasets having fine-grained sketch-photo association. QMUL-ShoeV2 [75] and QMUL-ChairV2 [75] contains 6730/2000 and 1800/400 sketches/photos respectively. Keeping 679/200 (525/100) sketches/photos for evaluation from ShoeV2 (ChairV2) respectively we use the rest for training. Sketchy [55] contains 125 categories with 100 photos each, having around 5 sketches per photo. While, training uses a standard train-test split [55] of 9:1, during inference we construct a challenging gallery using photos across one category for retrieval. For scene-level FG-SBIR evaluation we use FS-COCO [15] which includes 10,000 unique sketch-photo pairs with a 7:3 train/test split. + +Implementation Details: Considering a standard large FG-SBIR network as a teacher, our proposed student uses a cheaper backbone feature extractor MobileNetV2 [54], a standard for mobile end-devices. Margin is set to $m = 0.2$ for triplet loss (Secs. 3 and 4.1). We use an Adam optimiser at a learning rate of 0.0001 and a batchsize of 16 for 200 epochs. For practicality, we chose discrete canvas-sizes ( $C$ ) instead of a continuous action space, where $C = \{32 \times 32, 64 \times 64, 128 \times 128, 256 \times 256\}$ . Canvas-size selector ( $\psi_C$ ) is modelled using a simple GRU network, with embedding size 128, and trained via reinforcement learning with a learning rate of 0.0001 and a batchsize of 32 for 500 epochs, keeping $T_{max} = 100$ for vector-sketches. For the action space, we take varying canvas-sizes from $C$ and set $\beta = 1$ in Eq.2. $\lambda$ , $\lambda_r$ , $\lambda_{\mathrm{Tri}}$ and $\lambda_{\mathrm{F}}$ are empirically set to 0.5, 0.4, 0.48 and 0.35 (Fig.5) respectively. While training $\psi_C$ , sketches are rendered at different completion (30%, 35%, ... 100%) levels to simulate varying levels of abstraction, ensuring a sketch-abstraction-aware canvas selector (e.g., a smaller canvas for less detailed sketch). + +Evaluation Metric: Accuracy for FG-SBIR is the percentage of sketches where true-match photos are ranked in the Top-q (q = 1,10) lists. Computational efficiency is measured via parameter-count and average FLOPs – lower is better. For average FLOPs in Table 2, we take the combined sum of FLOPs per sample [12] from (i) canvas-size selector and (ii) retrieval model, using the predicted resolution, and average across the validation set. + +Table 2. Quantitative Analysis on FG-SBIR. Best viewed when zoomed. + +
MethodsCanvas-size c×cParams (mil.)ShoeV2 [75]ChairV2 [75]Sketchy [55]FSCOCO [15]
Top1 (%)Top10 (%)FLOPS (G)Top1 (%)Top10 (%)FLOPS (G)Top1 (%)Top10 (%)FLOPs (G)Top1 (%)Top10 (%)FLOPS (G)
State-of-the-ArtsTriplet-SN [75]32x328.7509.77 (↓18.94)24.82 (↓46.74)0.08315.61 (↓32.04)29.86 (↓54.38)0.0834.23 (↓11.12)12.58 (↓23.13)0.0830.76 (↓3.94)6.22 (↓14.8)0.083
64x648.7517.63 (↓11.08)44.96 (↓26.60)0.33830.25 (↓17.4)54.21 (↓30.03)0.3389.68 (↓5.67)22.45 (↓13.26)0.3381.85 (↓2.85)7.33 (↓13.69)0.338
128x1288.7525.01 (↓3.70)62.35 (↓9.21)1.39740.61 (↓7.04)72.68 (↓11.56)1.39712.98 (↓2.37)29.63 (↓6.08)1.3973.28 (↓1.42)15.92 (↓5.1)1.397
256x2568.7528.7171.565.28047.6584.245.28015.3535.715.2804.721.025.280
SketchyNetV12.2228.46 (↓0.25)69.01 (↓2.55)0.83346.28 (↓0.37)82.33 (↓1.91)0.83314.85 (↓0.5)34.68 (↓1.03)0.8334.22 (↓0.48)18.33 (↓2.69)0.833
SketchyNetV22.2727.89 (↓0.82)68.76 (↓2.80)0.26445.98 (↓1.67)80.14 (↓4.10)0.29414.21 (↓1.14)34.16 (↓1.55)0.3213.98 (↓0.72)17.64 (↓3.38)0.423
HOLEE-SN [59]32x329.3110.84 (↓20.9)27.28 (↓48.5)0.09618.72 (↓34.69)31.68 (↓55.88)0.0966.01 (↓10.69)14.12 (↓24.78)0.0960.86 (↓4.04)6.93 (↓14.78)0.096
64x649.3119.83 (↓11.91)51.72 (↓24.06)0.38734.32 (↓19.09)56.46 (↓31.1)0.3879.92 (↓6.78)24.38 (↓14.52)0.3872.11 (↓2.79)7.86 (↓13.85)0.387
128x1289.3126.78 (↓4.96)66.28 (↓9.5)1.54545.74 (↓7.67)75.59 (↓11.97)1.54513.81 (↓2.89)33.68 (↓5.22)1.5453.65 (↓1.25)16.32 (↓5.39)1.545
256x2569.3131.7475.785.75853.4187.565.75816.7038.905.7584.921.715.758
SketchyNetV12.2231.59 (↓0.15)73.96 (↓1.82)0.83353.27 (↓0.14)86.93 (↓0.63)0.83316.25 (↓0.45)38.21 (↓0.69)0.8334.56 (↓0.34)19.92 (↓1.79)0.833
SketchyNetV22.2730.86 (↓0.88)72.56 (↓3.22)0.25952.74 (↓0.67)84.02 (↓3.54)0.28915.91 (↓0.79)37.88 (↓1.02)0.3154.04 (↓0.86)18.63 (↓3.08)0.417
Triplet-RL [6]32x3222.111.98 (↓22.12)28.11 (↓50.71)0.14220.53 (↓36.01)32.59 (↓57.02)0.1421.76 (↓2.94)3.88 (↓6.47)0.142
64x6422.121.28 (↓12.82)54.71 (↓24.11)0.57736.09 (↓20.45)57.33 (↓32.28)0.5773.04 (↓1.66)6.79 (↓3.56)0.577
128x12822.128.83 (↓5.27)67.84 (↓10.98)2.29948.38 (↓8.16)77.25 (↓12.36)2.2994.12 (↓0.58)9.05 (↓1.30)2.299
256x25622.134.1078.826.04156.5489.616.0414.7010.356.041
SketchyNetV12.2233.88 (↓0.22)77.15 (↓1.67)0.83355.92 (↓0.62)88.20 (↓1.41)0.8334.59 (↓0.11)10.21 (↓0.14)0.833
SketchyNetV22.2733.26 (↓0.84)76.84 (↓1.98)0.25555.14 (↓1.40)87.31 (↓2.30)0.2854.51 (↓0.19)9.89 (↓0.46)0.308
StyleVAE [51]32x3225.3712.68 (↓23.79)28.69 (↓53.14)0.12522.48 (↓40.38)32.10 (↓59.04)0.12506.92 (↓12.7)16.25 (↓29.53)0.125
64x6425.3722.93 (↓13.54)57.71 (↓24.12)0.50840.26 (↓22.6)58.23 (↓32.91)0.50812.46 (↓7.16)29.62 (↓16.16)0.508
128x12825.3730.91 (↓5.56)70.06 (↓11.77)2.02453.88 (↓8.98)78.64 (↓12.5)2.02417.15 (↓2.47)38.84 (↓6.94)2.024
256x25625.3736.4781.835.64262.8691.145.64219.6245.785.642
SketchyNetV12.2236.11 (↓0.36)79.63 (↓2.20)0.83362.23 (↓0.63)88.95 (↓2.19)0.83319.48 (↓0.12)44.02 (↓1.76)0.833
SketchyNetV22.2735.88 (↓0.59)78.46 (↓3.37)0.25161.95 (↓0.91)87.68 (↓3.46)0.28119.10 (↓0.50)43.57 (↓2.21)0.304
Partial-OT [14]32x3222.1
64x6422.1
128x12822.1
256x25622.1
SketchyNetV12.22
SketchyNetV22.27
B-BFR224x2244.6124.3261.232.60243.3976.332.60210.4724.342.6022.1511.612.602
B-DRSDynamic9.3625.1268.845.49344.7878.146.18711.8726.586.2453.8617.636.291
B-CropDynamic4.9220.3252.126.56235.0757.686.59708.2720.196.6291.9809.676.688
B-Regress [48]Dynamic2.2722.8261.330.26138.4264.650.29108.6921.980.3232.0710.150.375
B-AKD [76]Dynamic2.2723.6764.260.26838.6171.880.29512.6727.710.3312.9114.010.392
B-PKT [45]Dynamic2.2724.3165.910.25942.6873.660.28812.2728.310.3173.4516.220.372
B-VGG16-SN256x25614.7133.0378.5140.1852.1684.0840.1818.6141.2540.185.1623.6240.18
B-ThiNet [37]256x2568.327.51 (↓25.52)22.34 (↓56.17)9.34212.31 (↓39.85)26.63 (↓57.45)9.3423.16 (↓15.45)12.73 (↓28.52)9.3420.65 (↓4.51)5.35 (↓18.27)9.342
B-Prune [38]256x2566.173.92 (↓29.11)17.32 (↓61.19)8.1387.98 (↓44.18)19.82 (↓64.26)8.1381.45 (↓17.16)8.22 (↓33.03)8.1380.53 (↓4.63)4.97 (↓18.65)8.138
B-VGG16-SN SketchyNetV22.2732.77 (↓0.26)78.21 (↓0.3)0.25450.75 (↓1.41)82.21 (↓1.87)0.28316.55 (↓2.06)38.66 (↓2.59)0.3324.69 (↓0.47)21.76 (↓1.86)0.407
+ +# 5.1. Competitors + +We evaluate against State-of-the-Arts (SoTAs) and a few relevant curated baselines (B-). (i) SoTAs: Unlike Triplet-SN [75], HOLEF-SN [59], Triplet-RL [6] (we report full-sketch evaluation) and StyleVAE [51]; Partial-OT [14] specifically caters to scene-level FG-SBIR. For fairness, we vary input-resolution of sketches during inference (fixed at their original $256 \times 256$ setup during training). We also adapt them as teachers, to judge our accuracy retainment after model-compression (SketchyNetV1) using full-res sketch and, further optimising canvas-size with canvas selector (SketchyNetV2). (ii) KD baselines: Here we re-purpose existing works in a cross-modal setting, keeping our canvas selector same, with Triplet-SN as the teacher. B-Regress - regresses features from teacher and student for both sketches and images, aligning their dimensions with an extra transformation layer [48], using a $l_{2}$ regression loss. Following [45] B-PKT calculates the conditional probability density for any pair of points in the teacher's embedding space [45], which models the probability of any two samples being close together. Sampling 'N' instances, it establishes a probability distribution for pairwise interactions within that space. Simultaneously, in the student's embedding space, a comparable distribution is derived, and the model minimises the divergence between them using a KL-divergence loss. B-AKD follows [76] in using spatial attention maps for knowledge transfer using similar dimensional matching like B-Regress. (iii) Resizing strategies: Keeping the same KD-paradigm with Triplet-SN-teacher, we vary the input + +size resizing module following existing works. $B$ -BFR follows [62] in designing a bilinear feature resizer amidst the intermediate layers of a CNN network, to output a down-scaled image. We choose its resized resolution of $224 \times 224$ (lowest) for subsequent evaluation. $B$ -DRS follows [12] in training a convolutional module that takes an image in full-resolution, calculates a probability vector on available canvas-sizes, and transforms it into binary decisions, indicating the scale factor of selection. $B$ -Crop follows [32] in meta-learning an image-cropping module that crops the input image to a target resolution for low compute evaluation. (iv) Pruning alternatives: Here we adapt two VGG16-backboned methods off-the-shelf from network-pruning family among other efficiency paradigms [18, 21, 62], in FG-SBIR setup - $B$ -ThiNet [37] and $B$ -Prune [38]. For comparative reference we report performance of a VGG16-backboned Triplet-SN network ( $B$ -VGG16-SN) and its SketchyNetV2 variant. + +# 5.2. Performance and Compute Analysis: + +Analysis on Accuracy: Table 2 reports the quantitative evaluation of our model against others. Triplet-SN [75] and HOLEF-SN [59] perform lower owing to weak backbones of Sketch-A-Net [74]. Enhanced by its RL-optimised reward function Triplet-RL[6] scores higher $(2.36\% \uparrow$ Top1 on ShoeV2) but fails to surpass StyleVAE [51], due to its delicate meta-learning disentanglement module addressing style-diversity. Being trained on region-wise sketch-photo associativity, Partial-OT [14] outperforms all others significantly on scene-level sketches (FS-COCO). We also notice a + +gradual drop in accuracy of SoTAs, when subjected to decreasing resolution of input-sketches suggesting that some methods are inherently accustomed to handling scale-shifts better than others, but are inferior to our dynamic approach (SketchyNetV2). Importantly, when most SoTAs are employed as teachers, their respective students (SketchyNetV1 variants) are seen to retain their respective original accuracy with a marginal drop of $\approx 0.5 - 1\%$ overall, thanks to our efficient KD paradigm. Having different teacher-student embedding spaces both $B$ -Regress and $B$ -AKD's paradigm of directly regressing on features proves incompatible to a cross-modal retrieval setting, dropping accuracy (compared to T-Triplet-SN), whereas lacking cross-modal discrimination $B$ -PKT performs lower than ours, proving our superiority in retaining accuracy. + +![](images/de3a0d5bb52e94f4070e12b04a21b1cf472229379a14b6f2fd172071e21ff8bf.jpg) +Figure 4. Exemplar sketches at their optimal canvas-size + +![](images/e362480feffaaa2a25a4bf1ad00113228371c443db89798b3bd7b2864ad24929.jpg) + +![](images/214d331dd4b3bbdc3eff100cacaf294036f74e397825e96a10e207f7b1fee5f4.jpg) + +![](images/ca18a9d5f00a783e7d294e3e47bdec3dd5eb72a362115abfe1899a23cb846b98.jpg) + +Analysis on Compute: Although FLOPs [12] decrease (Table 2) with a reduction in input resolution across all SoTAs, it comes at a severe cost of accuracy ( $\approx 21\%$ average drop from $256 \times 256$ to $32 \times 32$ on ShoeV2). This is mainly because ideal canvas-size for optimal retrieval varies across samples (Fig. 2 right) - thus subjecting all samples to a fixed canvas-size proves detrimental for accuracy. However, a positive Acc@1 score of $\approx 11\%$ even at a low canvas-size of $32 \times 32$ for every SoTA shows that some sketches are indeed recognisable if the model is trained properly. This is verified further by SketchyNetV2 variant of respective SoTAs, that secures the lowest FLOPs per SoTA, despite maintaining accuracy at par with its SoTA-counterpart - thanks to our RL-guided canvas selector, which can allot an optimal canvas-size per sketch towards better retrieval (Fig. 4). Contrarily, B-BFR fails to lower FLOPs as it fixes a $224 \times 224$ input resolution for a satisfactory Acc@1 compared to Triplet-SN. B-Crop fairs better, however sketch being sparse, cropped sketches leads to confused representations, lowering accuracy. Despite coming closest in accuracy to ours on Triplet-SN, B-DRS lags behind considerably on compute, thanks to our canvas selector working in vector space (inputs $s_v$ ) and eliminating costly convolutional operations, unlike others using full-resolution sketch-raster as input. Being untrained to handle sparse-nature of sketches during such high compression, performance of both B-ThiNet and B-Prune collapses (vs. B-VGG16-SN) unlike its SketchyNetV2. Notably, the significant drop in parameter-count (4 - 10 times) against SoTAs, comes from choosing lightweight MobileNetV2 [54] as a student coupled with our canvas-size selector in every SketchyNetV2 variant. Compared to ShoeV2, FLOPs increase slightly for ChairV2, Sketchy and FS-COCO, owing to higher object (ChairV2) or scene-level (FS-COCO) details, and lower structural correspondence with photos (Sketchy), thus training a poorer retrieval model which in turn acts as a poor critic for our canvas selector. Our canvas selector alone takes a negligible 51.84K parameters with 0.02G FLOPs. + +# 5.3. Ablation Study + +As our teacher $(F_T)$ can be any pre-trained FG-SBIR network, we limit our ablation to justifying design choices of our student model and canvas-size selector. We use $B$ -VGG16-SN as a reference $F_T$ - Type- $\emptyset$ in Table 3. + +Training Canvas-Size selector $\psi_C$ : To justify loss objectives of $\psi_C$ we evaluate on ShoeV2, eliminating $r^{-1}$ , $\mathcal{L}_{\mathrm{Tri}}^R$ , $\mathcal{L}_{\mathrm{F}}$ one by one from Eq. 6 (Table 3). While eliminating accuracy rewards, $(w / o \text{ rank}^{-1}, w / o \mathcal{L}_{\mathrm{Tri}}^R)$ increases constraint on lowering FLOPs, dropping accuracy, ignoring $\mathcal{L}_{\mathrm{F}}$ $(w / o \mathcal{L}_F)$ gains accuracy but costs high FLOPs. We thus optimally balance between the two using $\lambda_{\mathrm{F}}$ (Fig. 5 left). One can also choose to train for faster retrieval (lower FLOPs), given a decent Acc@5 (Fig. 5 right). + +![](images/e2f7d3047db873640356a1426d38a7e35719439634bff95966f7bd695227d686.jpg) +Figure 5. Impact of varying $\lambda_{\mathrm{F}}$ on Acc@1 and Acc@5 + +![](images/799f80fd904eaa953d0823a729f55c7ac735d4637db5012301974a08ddeba2b7.jpg) + +Architectural insights: To judge the efficiency of our student's backbone, we explore a few cheaper backbone-networks (Types IV, V in Table 3). Evidently, MobileNetV2 (Ours) balances optimally between compute and accuracy. As an option to vector input for canvas selector, we re-design it with a simple seven-layer CNN module following [78] to use the full-resolution sketch-image as its input, and select a target canvas-size to downscale it for retrieval. With extra computational cost coming from spatial domain, total FLOPs incurred is much higher (VI). Next we compare efficacy of sketch as input to $\psi_C$ with respect to types of vector format - offset-coordinate (VI) [22] vs. absolute ones (Ours). Turns out the former is better for encoding. Comparing sketch encoders among LSTM (VIII), GRU (Ours) and Transformer (IX), showed ours to be optimum empirically (Table 3). Furthermore, preponding recent pre-processing modules like stroke-subset selector of [9] to our canvas selector, yields $42.95\%$ (vs. $43.7\%$ [9]) Acc@1 ShoeV2, proving its off-the-shelf compatibility. + +Table 3. Ablative studies (accuracy on ShoeV2) + +
TypeFLOPsParamsShoeV2
(G)(mil.)Top1 (%)Top10 (%)
∅B-VGG16-SN40.1814.7133.0378.51
Iw/o rank-1-reward0.17325.3164.36
IIw/o \( \mathcal{L}_{\text{Tri}}^R \)-reward0.1822.2727.7869.32
IIIw/o \( \mathcal{L}_{\text{F}} \)-reward0.83332.9178.39
IVResNet18 backbone1.45111.1826.7268.18
VEfficientNet backbone0.4594.0129.4772.04
VIImage-based \( \psi_C \)5.7307.6232.2177.68
VIIOffset \( s_v \)0.2552.2732.6178.40
VIIIDecoder-LSTM0.2682.2531.4177.32
IXDecoder-Tf0.3142.3431.9878.02
SketchyNetV20.2542.2732.7778.21
+ +Generalisability of canvas selector: To judge if the optimal canvas-size selected is uniform irrespective of the training method, we first train four canvas-selectors using four SoTAs of Table 2. Evaluating them on ShoeV2 [75], $89.21\%$ of test-set sketches in ShoeV2 to obtain the same optimal canvas size across four different models, trained on four different methods, which further proves that our canvas selector can work equally well as a pre-processing module for any new model. Furthermore, plugging a canvas selector trained with SketchyNetV1 (critic) from B-VGG16-SN with a different SketchyNetV1 from StyleVAE [51] during inference achieves comparable performance $(35.69\%)$ Top-1) to a canvas selector trained on StyleVAE-SketchyNetV1 $(35.88\%)$ Top-1) on ShoeV2. Importantly, $95.64\%$ of test-set sketches yield the same optimal canvas-size across both models, + +confirming the cross-model generalisability of our canvas selector. (ii) On training canvas selector with teacher $F_{T}$ (for Triplet-SN) as critic, yields same optimal canvas-sizes as those for student $F_{S}$ , for $94.21\%$ of test-set sketches in ShoeV2, while incurring comparable FLOPs reduction (0.268G) and accuracy (27.91% Top1) against $F_{S}$ as critic. + +Comparing Inference times: We evaluated on a 12 GB Nvidia RTX 2080 Ti GPU, on ShoeV2, for inference time of Triplet-SN, HOLEF-SN, Triplet-RL, and StyleVAE (SOTAs) on full-sketch to obtain 37, 38, 36 and $42\mathrm{ms}$ , against 10, 11, 11, and $12\mathrm{ms}$ for their SketchyNetV2 variants respectively. For practical validation, we evaluated onnx-INT8 version of Triplet-SN [75] and its SketchyNetV2 variant (ours) after post-training quantisation, on an iPhone13 via the AI-benchmark app for ShoeV2 [75] to obtain 51ms and 18ms respectively, which shows our efficiency. + +Retraining SoTA on mixed resolution sketches?: To explore if retraining SoTAs on sketches with mixed resolutions, we re-train Triple-SN [75] on rendered mixed-resolution sketches of ShoeV2. Poor performance in Table 4 at lower resolutions similar to Table 2 validates our KD approach. The key lies in that the sketch-feature from teacher used for student-guidance $(f_s^T)$ is always extracted from full-resolution sketch while that extracted by student is at varying resolutions, imparting scale invariance, which is ideally suited to our cross-modal retrieval setup. + +Table 4. Evaluating SoTAs trained on mixed-resolution sketches + +
Canvas SizeTriplet-SNTriplet-RL
Top1 (%)Top10 (%)FLOPs (G)Top1 (%)Top10 (%)FLOPs (G)
32×3210.9126.650.08312.9731.080.142
64×6419.0448.120.33822.9656.230.577
128×12826.0763.871.39730.0472.952.299
256×25628.7471.685.28034.2177.246.041
+ +# 5.4. Further Insights + +Cross-dataset Generalisation: To judge the generalisation potential of our canvas-size selector across datasets we train it on ShoeV2, but plug it to a retrieval model (SketchyNetV1) trained on ChairV2, and evaluate on ChairV2. Against the standard result (Table 2) of training both modules (SketchyNetV2) on ChairV2 (0.285G, $55.32\%$ Acc@1), this setup retains accuracy at $53.14\%$ with slight rise in FLOPs (0.314G), thus confirming its potential. + +Faster Sketch-recognition: Exploring potential of our canvas-size selector, we prepend it to a standard sketch-recognition pipeline [74] to optimise the canvas-size of query sketch, and reduce overall compute. Unlike our $\psi_{C}$ , it uses negative of cross-entropy loss as accuracy reward $(R_{\mathrm{acc}})$ , keeping rest same. Compared to existing Sketch-a-Net [74] incurring 5.28G FLOPs at $68.71\%$ accuracy on Quickdraw [22], our canvas selector fitted paradigm drops FLOPs to 0.261G while retaining accuracy at $68.14\%$ . This shows for the first time, that sparsity of sketches can be handled, in various downstream tasks via our $\psi_{C}$ . + +Extension to Category-level SBIR: For categorical SBIR we use Sketchy (ext) [36] dataset (90:10 train:test split), which holds $75k$ sketches across 125 categories having $\approx 73k$ images [36] in total. It is evaluated using mean average precision (mAP@all) and precision at top 200 retrieval (P@200) [36]. Here we compete against state-of-the-arts (SoTA) like $B-D2S$ that follows [19] without its zero-shot setting), and a simple baseline using a Siamese-style + +Table 5. Quantitative Evaluation on Categorical SBIR + +
MethodsFLOPsParamsSketchy(ext)
(G)(mil.)mAP@allP@200
SoTAB-SBIR-SN40.1814.710.7150.861
DSH [36]22.1416.320.7110.858
B-D2S [19]17.75136.360.8100.894
StyleVAE [51]8.11625.370.9050.927
T-SoTAT-SBIR-SN0.2540.7020.810
T-DSH0.2892.270.7040.823
T-D2S0.2680.7980.845
T-StyleVAE0.2510.8860.897
+ +VGG-16 network (B-SBIR-SN) following [75]. Similar to Section 5.1 adapting SoTAs as teachers in our SketchyNetV2 paradigm (T-SoTA) shows them retaining their accuracy (Table 5) while significantly lowering compute. + +Teacher-Student Pairs: To explore other teacher-student pairs apart from ours, we test SketchyNetV2 models of a few $(F_T,F_S)$ pairs, where teachers are trained similar to B-VGG16-SN. From Table 6 we see our pair to outperform others. + +Table 6. Evaluation on Teacher-Student Pairs + +
Teacher\StudentMobileNetV2EfficientNetResNet-18
Top-1FLOPsTop-1FLOPsTop-1FLOPs
VGG-16 [30]32.77%0.254G29.47%0.459G26.72%1.451G
Inception-V3 [74]27.89%0.264G27.16%0.461G25.31%1.452G
+ +Combining compression methods: To explore if a combination of compression methods is capable of increasing efficiency, we subject our Knowledge Distillation (KD) trained model (SketchyNetV1) of Triplet-SN [75] (Table 2) to (i) quantisation following [21], and (ii) network-pruning following [38], separately, and then train their respective canvas-selectors. While the former secures $27.34\%$ (vs. $27.89\%$ ours) Top-1 score with a lower inference time of $9.1\mathrm{ms}$ (vs. $10\mathrm{ms}$ ours), the latter scores $26.91\%$ (vs. $27.89\%$ ms ours) Top-1 with lower FLOPs of 0.249G (vs. 0.264G ours). Such competitive results further endorse exploring combining compression methods as future works. + +Future Works: Sketch being an abstraction of an object's photo, when resized to a lower scale for retrieval, infers that its object's semantic is interpretable at that scale. This should ideally provide a lower bound for resizing a photo while retaining interpretability (e.g. high retrieval score). Consequently, designing a learnable photo-resizer aided by our trained sketch-canvas selector is a potential future work. Secondly, ours being a meta-framework, using more recent frameworks like DINOv2 [41] as large-scale teacher models for further gains is another targeted future work. + +# 6. Conclusion + +We for the first time investigate the problem of efficient inference for FG-SBIR, where we show that using existing efficient lightweight networks directly, are unfit for sketches. We thus propose two generic components that can be adapted to photo-networks for sketch-specific data. Firstly, a cross-modal knowledge distillation paradigm distills the knowledge of larger models to a smaller one, directly reducing FLOPs (params) by 97.92 $(84.9)\%$ . Secondly, an abstraction-aware canvas-size selector dynamically selects the best canvas-size for a sketch, reducing FLOPs by extra two-thirds. Extensive experiments show our method to perform at par with state-of-the-arts using much lower compute thus further conveying our method's worth in enhancing them towards deployment. + +# References + +[1] Manoj Alwani, Yang Wang, and Vashisht Madhavan. Decore: Deep compression with reinforcement learning. In CVPR, 2022. 2 +[2] Aneeshan Sain and Ayan Kumar Bhunia and Pinaki Nath Chowdhury and Aneeshan Sain and Subhadeep Koley and Tao Xiang and Yi-Zhe Song. CLIP for All Things Zero-Shot Sketch-Based Image Retrieval, Fine-Grained or Not. In CVPR, 2023. 1 +[3] Jimmy Ba and Rich Caruana. Do deep nets really need to be deep? NeurIPS, 2014. 2 +[4] Hessam Bagherinezhad, Maxwell Horton, Mohammad Rastegari, and Ali Farhadi. Label refinery: Improving image-net classification through label progression. arXiv preprint arXiv:1805.02641, 2018. 2 +[5] Ayan Kumar Bhunia, Ayan Das, Umar Riaz Muhammad, Yongxin Yang, Timothy M Hospedales, Tao Xiang, Yulia Gryaditskaya, and Yi-Zhe Song. Pixel: A competitive sketching ai agent. so you think you can sketch? ACM TOG, 2020. 3 +[6] Ayan Kumar Bhunia, Yongxin Yang, Timothy M Hospedales, Tao Xiang, and Yi-Zhe Song. Sketch less for more: On-the-fly fine-grained sketch based image retrieval. In CVPR, 2020. 1, 2, 3, 6 +[7] Ayan Kumar Bhunia, Pinaki Nath Chowdhury, Aaneshan Sain, Yongxin Yang, Tao Xiang, and Yi-Zhe Song. More photos are all you need: Semi-supervised learning for fine-grained sketch based image retrieval. In CVPR, 2021. 1, 2, 3, 4 +[8] Ayan Kumar Bhunia, Pinaki Nath Chowdhury, Yongxin Yang, Timothy M Hospedales, Tao Xiang, and Yi-Zhe Song. Vectorization and rasterization: Self-supervised learning for sketch and handwriting. In CVPR, 2021. 3 +[9] Ayan Kumar Bhunia, Subhadeep Koley, Abdullah Faiz Ur Rahman Khilji, Aneeshan Sain, Pinaki Nath Chowdhury, Tao Xiang, and Yi-Zhe Song. Sketching without worrying: Noise-tolerant sketch-based image retrieval. In CVPR, 2022. 1, 2, 3, 7 +[10] Ayan Kumar Bhunia, Subhadeep Koley, Amandeep Kumar, Aneeshan Sain, Pinaki Nath Chowdhury, Tao Xiang, and Yi-Zhe Song. Sketch2Saliency: Learning to Detect Salient Objects from Human Drawings. In CVPR, 2023. 1 +[11] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In ICML, 2020. 3 +[12] Ying Chen, Liang Qiao, Zhanzhan Cheng, Shiliang Pu, Yi Niu, and Xi Li. Dynamic low-resolution distillation for cost-efficient end-to-end text spotting. In ECCV, 2022. 5, 6, 7 +[13] Pinaki Nath Chowdhury, Ayan Kumar Bhunia, Viswanatha Reddy Gajjala, Aneeshan Sain, Tao Xiang, and Yi-Zhe Song. Partially does it: Towards scene-level fg-sbir with partial input. In CVPR, 2022. 1, 2 +[14] Pinaki Nath Chowdhury, Ayan Kumar Bhunia, Viswanatha Reddy Gajjala, Aneeshan Sain, Tao Xiang, and Yi-Zhe Song. Partially does it: Towards scene-level fg-sbir with partial input. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 6 + +[15] Pinaki Nath Chowdhury, Aneeshan Sain, Ayan Kumar Bhunia, Tao Xiang, Yulia Gryaditskaya, and Yi-Zhe Song. Fscoco: Towards understanding of freehand sketches of common objects in context. In ECCV, 2022. 5, 6 +[16] Chowdhury, Pinaki Nath and Bhunia, Ayan Kumar and Sain, Aneeshan and Koley, Subhadeep and Xiang, Tao and Song, Yi-Zhe. Democratising 2D Sketch to 3D Shape Retrieval Through Pivoting. In ICCV, 2023. 2 +[17] John Collomosse, Tu Bui, and Hailin Jin. Livesketch: Query perturbations for guided sketch-based visual search. In CVPR, 2019. 2, 5 +[18] Matthieu Courbariaux, Yoshua Bengio, and Jean-Pierre David. Binaryconnect: Training deep neural networks with binary weights during propagations. NeurIPS, 28, 2015. 2, 6 +[19] Sounak Dey, Pau Riba, Anjan Dutta, Josep Llados, and Yi-Zhe Song. Doodle to search: Practical zero-shot sketch-based image retrieval. In CVPR, 2019. 2, 8 +[20] Anjan Dutta and Zeynep Akata. Semantically tied paired cycle consistency for zero-shot sketch-based image retrieval. In CVPR, 2019. 1 +[21] Suyog Gupta, Ankur Agrawal, Kailash Gopalakrishnan, and Pritish Narayanan. Deep learning with limited numerical precision. In ICML, 2015. 2, 6, 8 +[22] David Ha and Douglas Eck. A neural representation of sketch drawings. In ICLR, 2018. 1, 5, 7, 8 +[23] Song Han, Huizi Mao, and William J Dally. Deep compression: Compressing deep neural networks with pruning, trained quantization and huffman coding. In ICLR, 2016. 2 +[24] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 3 +[25] Geoffrey Hinton, Oriol Vinyls, and Jeff Dean. Distilling the knowledge in a neural network. In NeurIPS Deep Learning Workshop, 2014. 2 +[26] Gao Huang, Zhuang Liu, Laurens Van Der Maaten, and Kilian Q Weinberger. Densely connected convolutional networks. In CVPR, 2017. 3 +[27] Shouyou Huang and Qiang Wu. Robust pairwise learning with huber loss. Journal of Complexity, 2021. 4 +[28] Zehao Huang and Naiyan Wang. Like what you like: Knowledge distill via neuron selectivity transfer. arXiv preprint arXiv:1707.01219, 2017. 2 +[29] Leslie Pack Kaelbling, Michael L Littman, and Andrew W Moore. Reinforcement learning: A survey. JAIR, 1996. 3, 5 +[30] Andrew Zisserman Karen Simonyan. Very deep convolutional networks for large-scale image recognition. In ICLR, 2015. 3, 4, 8 +[31] Debang Li, Huikai Wu, Junge Zhang, and Kaiqi Huang. A2-rl: Aesthetics aware reinforcement learning for image cropping. In CVPR, 2018. 3 +[32] Debang Li, Junge Zhang, and Kaiqi Huang. Learning to learn cropping models for different aspect ratio requirements. In CVPR, 2020. 6 +[33] Xiaodan Liang, Lisa Lee, and Eric P Xing. Deep variation-structured reinforcement learning for visual relationship and attribute detection. In CVPR, 2017. 3 + +[34] Darryl Lin, Sachin Talathi, and Sreekanth Annapureddy. Fixed point quantization of deep convolutional networks. In ICML, 2016. 2 +[35] Hangyu Lin, Yanwei Fu, Yu-Gang Jiang, and Xiangyang Xue. Sketch-bert: Learning sketch bidirectional encoder representation from transformers by self-supervised learning of sketch gestalt. In CVPR, 2020. 5 +[36] Li Liu, Fumin Shen, Yuming Shen, Xianglong Liu, and Ling Shao. Deep sketch hashing: Fast free-hand sketch-based image retrieval. In CVPR, 2017. 2, 8 +[37] Jian-Hao Luo, Jianxin Wu, and Weiyao Lin. Thinet: A filter level pruning method for deep neural network compression. In ICCV, 2017. 2, 6 +[38] Pavlo Molchanov, Stephen Tyree, Tero Karras, Timo Aila, and Jan Kautz. Pruning convolutional neural networks for resource efficient inference. arXiv preprint arXiv:1611.06440, 2016. 6, 8 +[39] Umar Riaz Muhammad, Yongxin Yang, Yi-Zhe Song, Tao Xiang, and Timothy M Hospedales. Learning deep sketch abstraction. In CVPR, 2018. 3 +[40] Umar Riaz Muhammad, Yongxin Yang, Timothy Hospedales, Tao Xiang, and Yi-Zhe Song. Goal-driven sequential data abstraction. In ICCV, 2019. 1, 3 +[41] Maxime Oquab, Timothee Darct, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. DINOv2: Learning Robust Visual Features without Supervision. arXiv preprint arXiv:2304.07193, 2023. 8 +[42] Kaiyue Pang, Yi-Zhe Song, Tony Xiang, and Timothy M Hospedales. Cross-domain generative learning for fine-grained sketch-based image retrieval. In BMVC, 2017. 2 +[43] Kaiyue Pang, Ke Li, Yongxin Yang, Honggang Zhang, Timothy M Hospedales, Tao Xiang, and Yi-Zhe Song. Generalising fine-grained sketch-based image retrieval. In CVPR, 2019. 1 +[44] Kaiyue Pang, Yongxin Yang, Timothy M Hospedales, Tao Xiang, and Yi-Zhe Song. Solving mixed-modal jigsaw puzzle for fine-grained sketch-based image retrieval. In CVPR, 2020. 2, 3 +[45] Nikolaos Passalis and Anastasios Tefas. Learning deep representations with probabilistic knowledge transfer. In ECCV, 2018. 6 +[46] Pinaki Nath Chowdhury and Ayan Kumar Bhunia and Aneeeshan Sain and Subhadeep Koley and Tao Xiang and Yi-Zhe Song. SceneTrilogy: On Human Scene-Sketch and its Complementarity with Photo and Text. In CVPR, 2023. 2 +[47] Pinaki Nath Chowdhury and Ayan Kumar Bhunia and Aeeeshan Sain and Subhadeep Koley and Tao Xiang and Yi-Zhe Song. What Can Human Sketches Do for Object Detection? In CVPR, 2023. 1 +[48] Adriana Romero, Nicolas Ballas, Samira Ebrahimi Kahou, Antonie Chassang, Carlo Gatta, and Yoshua Bengio. Fitnets: Hints for thin deep nets. In ICLR, 2015. 2, 4, 6 +[49] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. IJCV, 2015. 4 + +[50] Aneeshan Sain, Ayan Kumar Bhunia, Yongxin Yang, Tao Xiang, and Yi-Zhe Song. Cross-modal hierarchical modelling, fine-grained sketch based image retrieval. In BMVC, 2020, 2, 3, 4 +[51] Aneeshan Sain, Ayan Kumar Bhunia, Yongxin Yang, Tao Xiang, and Yi-Zhe Song. Stylemeup: Towards style-agnostic sketch-based image retrieval. In CVPR, 2021. 1, 2, 3, 4, 6, 7, 8 +[52] Aneeshan Sain, Ayan Kumar Bhunia, Vaishnav Potlapalli, Pinaki Nath Chowdhury, Tao Xiang, and Yi-Zhe Song. Sketch3t: Test-time training for zero-shot sbir. In CVPR, 2022. 2 +[53] Aneeshan Sain, Ayan Kumar Bhunia, Subhadeep Koley, Pinaki Nath Chowdhury, Soumitri Chattopadhyay, Tao Xiang, and Yi-Zhe Song. Exploiting Unlabelled Photos for Stronger Fine-Grained SBIR. In CVPR, 2023. 1, 2 +[54] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zh-moginov, and Liang-Chieh Chen. Mobilenetv2: Inverted residuals and linear bottlenecks. In CVPR, 2018. 1, 3, 4, 5, 7 +[55] Patsorn Sangkloy, Nathan Burnell, Cusuh Ham, and James Hays. The sketchy database: learning to retrieve badly drawn bunnies. ACM TOG, 2016. 5, 6 +[56] John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347, 2017. 5 +[57] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. In ICLR, 2015. 1, 3 +[58] Jifei Song, Yi-Zhe Song, Tony Xiang, and Timothy M Hospedales. Fine-grained image retrieval: the text/sketch input dilemma. In BMVC, 2017. 2 +[59] Jifei Song, Qian Yu, Yi-Zhe Song, Tao Xiang, and Timothy M Hospedales. Deep spatial-semantic attention for fine-grained sketch-based image retrieval. In ICCV, 2017. 2, 6 +[60] Subhadeep Koley and Ayan Kumar Bhunia and Aneeshan Sain and Pinaki Nath Chowdhury and Tao Xiang and Yi-Zhe Song. Picture that Sketch: Photorealistic Image Generation from Abstract Sketches. In CVPR, 2023. 1 +[61] Richard S Sutton, David A McAllester, Satinder P Singh, and Yishay Mansour. Policy gradient methods for reinforcement learning with function approximation. In NeurIPS, 2000. 5 +[62] Hossein Talebi and Peyman Milanfar. Learning to resize images for computer vision tasks. In ICCV, 2021. 2, 6 +[63] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In ICML, 2019. 1, 3 +[64] Burak Uzkent and Stefano Ermon. Learning when and where to zoom with deep reinforcement learning. In CVPR, 2020. 2 +[65] Supawit Vatathanavaro, Suchat Tungjitnob, and Kitsuchart Pasupa. White blood cell classification: a comparison between vgg-16 and resnet-50 models. In JSCI, 2018. 3 +[66] Mahes Visvalingam and J Duncan Whyatt. The douglaspeucker algorithm for line simplification: re-evaluation through visualization. In Computer Graphics Forum, 1990. 5 + +[67] Xin Wang, Qiuyuan Huang, Asli Celikyilmaz, Jianfeng Gao, Dinghan Shen, Yuan-Fang Wang, William Yang Wang, and Lei Zhang. Reinforced cross-modal matching and self-supervised imitation learning for vision-language navigation. In CVPR, 2019. 3 +[68] Yunhe Wang, Chang Xu, Shan You, Dacheng Tao, and Chao Xu. Cnnpack: Packing convolutional neural networks in the frequency domain. NeurIPS, 29, 2016. 2 +[69] Yulin Wang, Kangchen Lv, Rui Huang, Shiji Song, Le Yang, and Gao Huang. Glance and focus: a dynamic approach to reducing spatial redundancy in image classification. In NeurIPS, 2020. 2 +[70] Kilian Q Weinberger and Lawrence K Saul. Distance metric learning for large margin nearest neighbor classification. JMLR, 2009. 4 +[71] Peng Xu, Chaitanya K Joshi, and Xavier Bresson. Multigraph transformer for free-hand sketch recognition. IEEE T-NNLS, 2021. 4 +[72] Peng Xu, Timothy M Hospedales, Qiyue Yin, Yi-Zhe Song, Tao Xiang, and Liang Wang. Deep learning for free-hand sketch: A survey. IEEE TPAMI, 2022. 1, 2 +[73] Ran Xu, Fangzhou Mu, Jayoung Lee, Preeti Mukherjee, Somali Chaterji, Saurabh Bagchi, and Yin Li. SmartAdapt: Multi-branch object detection framework for videos on mobiles. In CVPR, 2022. 2 +[74] Qian Yu, Yongxin Yang, Yi-Zhe Song, Tao Xiang, and Timothy M. Hospedales. Sketch-a-net that beats humans. In BMVC, 2015. 6, 8 +[75] Qian Yu, Feng Liu, Yi-Zhe Song, Tao Xiang, Timothy M Hospedales, and Chen-Change Loy. Sketch me that shoe. In CVPR, 2016. 1, 2, 3, 4, 5, 6, 7, 8 +[76] Sergey Zagoruyko and Nikos Komodakis. Paying more attention to attention: Improving the performance of convolutional neural networks via attention transfer. In ICLR, 2017. 2, 6 +[77] Shuchang Zhou, Yuxin Wu, Zekun Ni, Xinyu Zhou, He Wen, and Yuheng Zou. Dorefa-net: Training low bitwidth convolutional neural networks with low bitwidth gradients. arXiv preprint arXiv:1606.06160, 2016. 2 +[78] Mingjian Zhu, Kai Han, Enhua Wu, Qiulin Zhang, Ying Nie, Zhenzhong Lan, and Yunhe Wang. Dynamic resolution network. In NeurIPS, 2021. 7 \ No newline at end of file diff --git a/2025/Sketch Down the FLOPs_ Towards Efficient Networks for Human Sketch/images.zip b/2025/Sketch Down the FLOPs_ Towards Efficient Networks for Human Sketch/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..f72ca9a9db616f53cfeaf880ec338349619bd985 --- /dev/null +++ b/2025/Sketch Down the FLOPs_ Towards Efficient Networks for Human Sketch/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:35f2a0c4de2664a957d1566305cf2ccb0e5bf8e8fcfe3db24e50136e09ea6dee +size 590137 diff --git a/2025/Sketch Down the FLOPs_ Towards Efficient Networks for Human Sketch/layout.json b/2025/Sketch Down the FLOPs_ Towards Efficient Networks for Human Sketch/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..0d5e2a09dac3e4c5500bb6eff47670ab64e3ce6c --- /dev/null +++ b/2025/Sketch Down the FLOPs_ Towards Efficient Networks for Human Sketch/layout.json @@ -0,0 +1,11181 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 79, + 97, + 531, + 114 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 97, + 531, + 114 + ], + "spans": [ + { + "bbox": [ + 79, + 97, + 531, + 114 + ], + "type": "text", + "content": "Sketch Down the FLOPs: Towards Efficient Networks for Human Sketch" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 96, + 130, + 503, + 159 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 130, + 503, + 159 + ], + "spans": [ + { + "bbox": [ + 96, + 130, + 503, + 159 + ], + "type": "text", + "content": "Aneeshan Sain1 Subhajit Maity2* Pinaki Nath Chowdhury1 Shubhadeep Koley1 Ayan Kumar Bhunia1 Yi-Zhe Song1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 162, + 159, + 447, + 174 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 162, + 159, + 447, + 174 + ], + "spans": [ + { + "bbox": [ + 162, + 159, + 447, + 174 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 162, + 159, + 447, + 174 + ], + "type": "text", + "content": " SketchX, CVSSP, University of Surrey, United Kingdom." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 110, + 174, + 498, + 188 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 174, + 498, + 188 + ], + "spans": [ + { + "bbox": [ + 110, + 174, + 498, + 188 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 110, + 174, + 498, + 188 + ], + "type": "text", + "content": " Department of Computer Science, University of Central Florida, United States." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 90, + 190, + 517, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 90, + 190, + 517, + 202 + ], + "spans": [ + { + "bbox": [ + 90, + 190, + 517, + 202 + ], + "type": "text", + "content": "{a.sain, p.chowdhury, s.koley, a.bhunia, y song}@surrey.ac.uk; Subhajit@ucf.edu" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 151, + 229, + 200, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 229, + 200, + 243 + ], + "spans": [ + { + "bbox": [ + 151, + 229, + 200, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 255, + 297, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 255, + 297, + 567 + ], + "spans": [ + { + "bbox": [ + 54, + 255, + 297, + 567 + ], + "type": "text", + "content": "As sketch research has collectively matured over time, its adaptation for at-mass commercialisation emerges on the immediate horizon. Despite an already mature research endeavour for photos, there is no research on the efficient inference specifically designed for sketch data. In this paper, we first demonstrate existing state-of-the-art efficient light-weight models designed for photos do not work on sketches. We then propose two sketch-specific components which work in a plug-n-play manner on any photo efficient network to adapt them to work on sketch data. We specifically chose fine-grained sketch-based image retrieval (FG-SBIR) as a demonstrator as the most recognised sketch problem with immediate commercial value. Technically speaking, we first propose a cross-modal knowledge distillation network to transfer existing photo efficient networks to be compatible with sketch, which brings down number of FLOPs and model parameters by " + }, + { + "bbox": [ + 54, + 255, + 297, + 567 + ], + "type": "inline_equation", + "content": "97.96\\%" + }, + { + "bbox": [ + 54, + 255, + 297, + 567 + ], + "type": "text", + "content": " percent and " + }, + { + "bbox": [ + 54, + 255, + 297, + 567 + ], + "type": "inline_equation", + "content": "84.89\\%" + }, + { + "bbox": [ + 54, + 255, + 297, + 567 + ], + "type": "text", + "content": " respectively. We then exploit the abstract trait of sketch to introduce a RL-based canvas selector that dynamically adjusts to the abstraction level which further cuts down number of FLOPs by two thirds. The end result is an overall reduction of " + }, + { + "bbox": [ + 54, + 255, + 297, + 567 + ], + "type": "inline_equation", + "content": "99.37\\%" + }, + { + "bbox": [ + 54, + 255, + 297, + 567 + ], + "type": "text", + "content": " of FLOPs (from 40.18G to 0.254G) when compared with a full network, while retaining the accuracy " + }, + { + "bbox": [ + 54, + 255, + 297, + 567 + ], + "type": "inline_equation", + "content": "(33.03\\%)" + }, + { + "bbox": [ + 54, + 255, + 297, + 567 + ], + "type": "text", + "content": " vs " + }, + { + "bbox": [ + 54, + 255, + 297, + 567 + ], + "type": "inline_equation", + "content": "32.77\\%" + }, + { + "bbox": [ + 54, + 255, + 297, + 567 + ], + "type": "text", + "content": " - finally making an efficient network for the sparse sketch data that exhibit even fewer FLOPs than the best photo counterpart." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 580, + 135, + 593 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 580, + 135, + 593 + ], + "spans": [ + { + "bbox": [ + 55, + 580, + 135, + 593 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 596, + 296, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 596, + 296, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 596, + 296, + 693 + ], + "type": "text", + "content": "A significant movement in the late computer vision literature has been that of model efficiency. This effort has been largely driven by the on-the-edge deployments of vision models. With great strides made on networks specifically tackling photo data (MobileNet [54], EfficientNet [63], etc), there however has been surprisingly no work targeting human sketches [72]. This is despite the abundance of work specifically addressing sketch data [47] and its unique traits:" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 316, + 228, + 555, + 331 + ], + "blocks": [ + { + "bbox": [ + 316, + 228, + 555, + 331 + ], + "lines": [ + { + "bbox": [ + 316, + 228, + 555, + 331 + ], + "spans": [ + { + "bbox": [ + 316, + 228, + 555, + 331 + ], + "type": "image", + "image_path": "13cb94a39c9a758370bf172a5b52d260950fde46c6c3b4a2912aa0047ab2c5d4.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 342, + 555, + 386 + ], + "lines": [ + { + "bbox": [ + 313, + 342, + 555, + 386 + ], + "spans": [ + { + "bbox": [ + 313, + 342, + 555, + 386 + ], + "type": "text", + "content": "Figure 1. Our SketchyNetV1 compresses existing heavy FG-SBIR networks to deliver smaller models. Further enhanced via a canvas selector module our SketchyNetV2 model minimises sketch-resolution dynamically to reduce FLOPs." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 390, + 554, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 390, + 554, + 474 + ], + "spans": [ + { + "bbox": [ + 313, + 390, + 554, + 474 + ], + "type": "text", + "content": "sequential [22], abstract [60], or stroke-wise [10], style-diversity [51], and data-scarcity [6, 20, 43]. This is particularly disappointing for the problem of fine-grained sketch-based image retrieval (FG-SBIR), which is perhaps the single-most studied sketch task [2, 7, 9, 13, 40, 43, 53, 75] that has already matured for at-mass commercial adoption, and is our problem of choice in this paper." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 474, + 555, + 642 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 474, + 555, + 642 + ], + "spans": [ + { + "bbox": [ + 313, + 474, + 555, + 642 + ], + "type": "text", + "content": "One would think that state-of-the-art models designed for photos would work as-is for sketches, treating sketch data as raster images (which sketch is not). As our first contribution, we preempt this thought via a pilot study to prove its false nature. In particular, we borrow MobileNetv2 [54] as a backbone network to construct a typical triplet loss based network for FG-SBIR. Results show a reduction of almost " + }, + { + "bbox": [ + 313, + 474, + 555, + 642 + ], + "type": "inline_equation", + "content": "37\\%" + }, + { + "bbox": [ + 313, + 474, + 555, + 642 + ], + "type": "text", + "content": " (relative) in retrieval accuracy when compared with a typical VGG-16 backbone [57]. The focus of this paper is however not to design entirely new sketch-specific efficient networks from scratch – lots of good efficient photo networks already exist. Our goal instead is to adapt existing photo networks to work with sketch, by introducing generic sketch-specific designs that work in a plug-n-play manner." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "content": "We first tackle the problem highlighted in the said pilot study – to make existing photo efficient networks perform on the same level as large networks on FG-SBIR. We argue the reason behind why the likes of MobileNetV2 did not work off-the-shelf is that the limited set of parameters could not accommodate the fine-grained nature of sketches (i.e.," + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 702, + 250, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 702, + 250, + 713 + ], + "spans": [ + { + "bbox": [ + 66, + 702, + 250, + 713 + ], + "type": "text", + "content": "*Work done as an intern at SketchX before joining UCF." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28383" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 300 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 300 + ], + "type": "text", + "content": "sketches are highly expressive in capturing fine-grained visual details [72]) for sketch and photo matching. For that, we introduce a knowledge distillation (KD) paradigm [48] to adapt this sketch trait onto efficient photo networks. The core idea is to induce strong semantic knowledge between a sketch and its photo, in the embedding space of a student, from a deep and high performing teacher. This particularly helps the efficient model to achieve on par performance with the deep variant without having a large parameter size. However, ours being a cross-modal retrieval, naively adapting standard KD paradigms involving logit distillation [3] would be sub-optimal. We therefore aim to preserve cross-modal pairwise distances among instances, and in turn the structural consistency of the teacher's discriminative embedding space in the student. The result is already an efficient FG-SBIR model (which we term SketchyNetV1) that reduces the number of FLOPs from 40.18G (VGG-16 based model[51]) to 0.833G (our MobileNetV2-based) – a roughly 48 times reduction (Fig. 1 left)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 303, + 295, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 303, + 295, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 303, + 295, + 685 + ], + "type": "text", + "content": "This abstract nature of human sketch [16] can also be exploited, to further reduce the FLOPs. The hypothesis is that since sketches are essentially abstract depictions of images, they should carry a similar level of semantic information even when rendered at much lower resolutions, and that more abstract a sketch is the more information compact it is (i.e., can sustain even lower resolutions). For that, we conducted another pilot study (Sec. 3), where we show unlike photos, sketches can sustain retrieval accuracy with much lower resolutions (4% vs 15% at " + }, + { + "bbox": [ + 55, + 303, + 295, + 685 + ], + "type": "inline_equation", + "content": "32 \\times 32" + }, + { + "bbox": [ + 55, + 303, + 295, + 685 + ], + "type": "text", + "content": "), and 30% of those query sketches achieving perfect retrieval at full-resolution, when downsized to " + }, + { + "bbox": [ + 55, + 303, + 295, + 685 + ], + "type": "inline_equation", + "content": "32 \\times 32" + }, + { + "bbox": [ + 55, + 303, + 295, + 685 + ], + "type": "text", + "content": " can still retrieve accurately (Fig. 2 right). This motivated us to design a canvas selector that can choose the right resolution on-the-fly to render a sketch before feeding to the network. This is thanks to another unique trait of sketch data - they come in vector format and can be freely rendered to any resolution without extra overhead. Accomplishing this however is nontrivial since this selection process of canvas-size involves a discrete decision, thus invoking non-differentiability. We thus leverage reinforcement learning [9] to bypass these issues and train an abstraction-aware canvas-size selector as a policy network for our model, which estimates the optimal canvas-size required for the input sketch. The FG-SBIR model, acting as a critic, calculates reward for the canvas selector, balancing between accuracy and FLOPs, that trains the policy network. With this further trait of sketches addressed, our final model (SketchyNetV2) was able to further reduce FLOPs by a factor of 3.28 (from 0.833G to 0.254G), while retaining a retrieval accuracy similar to the full model (33.03 vs 32.77), exactly as one would imagine sketch efficient networks to entail (sketches being sparse)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "type": "text", + "content": "In summary, our contributions are: (a) the first investigation towards efficient sketch networks from existing photo" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": "ones (b) a knowledge distillation framework specifically designed to cater to the fine-grained nature of sketches, (c) a canvas-size selector that caters to the abstract nature of sketches, which further reduces FLOPs by two-thirds. Extensive experiments show our method to perform at par with existing state-of-the-arts, while being superiorly efficient." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 156, + 400, + 167 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 156, + 400, + 167 + ], + "spans": [ + { + "bbox": [ + 313, + 156, + 400, + 167 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 173, + 553, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 173, + 553, + 389 + ], + "spans": [ + { + "bbox": [ + 313, + 173, + 553, + 389 + ], + "type": "text", + "content": "Fine-grained SBIR: Compared to category-level SBIR [17, 19, 36, 52], Fine-Grained SBIR (FG-SBIR) focuses on retrieving a single photo instance from a category-specific gallery based on a query sketch. Introduced as a deep triplet-ranking Siamese network [75], FG-SBIR has evolved with attention-based modules [59], textual tags [58], hybrid cross-domain generation [42], hierarchical coattention [50], and various pre-training strategies like mixed modal jigsaw solving [44], reinforcement learning [6], and transformer-based knowledge distillation [53]. Addressing sketch-specific traits like style diversity [51], data scarcity [7], and sketch stroke redundancy [9], improvements have been made for enhanced retrieval [53]. Recently it was extended to scene-level (retrieve a scene-image, given a scene-sketch), employing cross-modal region associativity [13] and enhanced using text-queries [46]. However, sketch-specific computational efficiency has largely been ignored - which we address here in this paper." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 391, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 391, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 391, + 553, + 713 + ], + "type": "text", + "content": "Towards Computational Efficiency: Addressing large-scale deployment, research emerging in computational efficiency [1], include network pruning [37], quantisation [21], binarisation [18], knowledge distillation [25], input-size optimisation [62], or their combinations [1]. Network pruning [23, 68] involves discovering and dropping low impact weights or channels of a pre-trained large network, while retaining its original accuracy. Contrarily, quantisation directly reduces the bit-width of parameter-values [21] and gradients [77], by replacing floating-point computations with faster and cheaper low-precision fixed-point numbers [34]. Its extreme version, binarisation [18], binarises weights and activations. Knowledge Distillation aims at transferring knowledge of a large pre-trained teacher network to a smaller student for cost-effective deployment. Existing techniques leverage output logits [3], hidden layers [48], attention-maps [76], or neuron selectivity pattern [28] of pre-trained teachers for the same. Self-distillation [4] employs the same network for both student and teacher models, to further reduce compute. Input-resolution optimisation on the other hand involves replacing a high-resolution input image with its down-scaled low-resolution version [62], to reduce overall compute [73], while retaining accuracy. Relevant works explored dynamic usage of high-resolution data [64] or patch proposals for strategic image-cropping [69], where the learnable resizer [62] operates on the original image resolution. All such" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28384" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 66, + 76, + 284, + 173 + ], + "blocks": [ + { + "bbox": [ + 58, + 66, + 293, + 76 + ], + "lines": [ + { + "bbox": [ + 58, + 66, + 293, + 76 + ], + "spans": [ + { + "bbox": [ + 58, + 66, + 293, + 76 + ], + "type": "text", + "content": "Table 1. Performance & computational cost for various networks" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 66, + 76, + 284, + 173 + ], + "lines": [ + { + "bbox": [ + 66, + 76, + 284, + 173 + ], + "spans": [ + { + "bbox": [ + 66, + 76, + 284, + 173 + ], + "type": "table", + "html": "
ModelFLOPSParametersTime (ms)Acc@1
VGG-19 [57]51.06G20.02M10133.63
VGG-16 [57]40.18G14.71M08433.03
ResNet-101 [24]20.50G42.50M15521.77
DenseNet-201 [26]11.40G18.09M33523.62
ResNet-50 [24]10.76G23.51M08621.32
InceptionV3 [75]07.72G21.79M13424.27
ResNet-18 [24]4.76G11.18M03620.72
EfficientNet [63]1.05G04.01M11222.47
MobileNetV2 [54]0.83G02.22M07020.85
", + "image_path": "cace9bb88c989feb6f296be44349025fffc9bb2367e18414d04e5c248581cd88.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 177, + 295, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 177, + 295, + 236 + ], + "spans": [ + { + "bbox": [ + 55, + 177, + 295, + 236 + ], + "type": "text", + "content": "methods however cater to compressing image-specific models, which if used off-the-shelf for FG-SBIR would fail to address the sparse manner in which sketches are represented [8]. We thus propose a sketch-specific light-weight FG-SBIR model, tackling sparsity in sketches." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 236, + 296, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 236, + 296, + 357 + ], + "spans": [ + { + "bbox": [ + 55, + 236, + 296, + 357 + ], + "type": "text", + "content": "Reinforcement Learning in Vision (RL): Applying RL [29] to computer vision problems [33, 67] is commonplace of late. Whenever quantifying the goodness of the network's state is non-differentiable, like selecting which regions to crop in an image [31], RL comes in handy. In sketch community, RL has found use in retrieval [6, 7], modelling sketch abstraction [39, 40], designing competitive sketching agents [5] and as a stroke selector [9]. Here, we leverage RL for optimising a canvas-size selector for a computationally efficient retrieval model." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 362, + 128, + 375 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 362, + 128, + 375 + ], + "spans": [ + { + "bbox": [ + 55, + 362, + 128, + 375 + ], + "type": "text", + "content": "3. Pilot Study" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 376, + 295, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 376, + 295, + 568 + ], + "spans": [ + { + "bbox": [ + 55, + 376, + 295, + 568 + ], + "type": "text", + "content": "Backbone Network for FG-SBIR: Amongst popular backbones used in recent FG-SBIR literature [6, 7, 50, 51], are VGG-16 [30] and Inception-V3 [75]. Others like ResNet18 [24] and EfficientNet [63], though cheaper and prevalent in vision [24] are rarely used in FG-SBIR. Exploring their potential in FG-SBIR, we train a baseline FG-SBIR model, plugging in popular ones as backbones for our study. Table 1 reports their parameter-count, FLOPs, time for feature-extraction per sample and Top-1 accuracy (Sec. 5) on standard QMUL-ShoeV2 dataset [75]. Despite scoring more than VGG-16 on image-classification [65], ResNet50 [24] performs poorly – indicating that modelling fine-grained sketch-photo association needs networks with higher FLOPs. Despite delivering better accuracies, heavier networks cost far more resources. We thus aim for a lighter model that achieves accuracy like a larger one." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "text", + "content": "Dynamic Sketch-canvas-size: Unlike photos that hold pixel-dense information, sketches are sparse black and white lines. This begs the question, if a sketch rendered at a higher resolution with added computational burden would convey any extra semantic information than at a lower one. To answer this, we focus on vector-format of sketches [7], where every sketch vector " + }, + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "inline_equation", + "content": "s_v" + }, + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "text", + "content": " can be characterised as a sequence of points " + }, + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "inline_equation", + "content": "(v_1, v_2, \\dots, v_{\\mathrm{T}})" + }, + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "text", + "content": ", and rendered to any canvas-size " + }, + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "inline_equation", + "content": "(c)" + }, + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "text", + "content": " as a raster-image " + }, + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "inline_equation", + "content": "(s_r)" + }, + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "text", + "content": " via rasterisation " + }, + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{R}^c(\\cdot): s_v \\to s_r^c" + }, + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "inline_equation", + "content": "s_r^c \\in \\mathbb{R}^{c \\times c \\times 3}" + }, + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "text", + "content": ". A single baseline FG-SBIR model (VGG-16) is trained and tested on full-resolution images, and sketches rendered at varying" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 332, + 73, + 438, + 144 + ], + "blocks": [ + { + "bbox": [ + 332, + 73, + 438, + 144 + ], + "lines": [ + { + "bbox": [ + 332, + 73, + 438, + 144 + ], + "spans": [ + { + "bbox": [ + 332, + 73, + 438, + 144 + ], + "type": "image", + "image_path": "2e9054e4b9b734386733cfe540dbcbb670b37b4422ac8e6b687c233bba54ab3a.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 322, + 148, + 545, + 159 + ], + "lines": [ + { + "bbox": [ + 322, + 148, + 545, + 159 + ], + "spans": [ + { + "bbox": [ + 322, + 148, + 545, + 159 + ], + "type": "text", + "content": "Figure 2. Pilot study for varying canvas-size (see text above)." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 441, + 72, + 539, + 146 + ], + "blocks": [ + { + "bbox": [ + 441, + 72, + 539, + 146 + ], + "lines": [ + { + "bbox": [ + 441, + 72, + 539, + 146 + ], + "spans": [ + { + "bbox": [ + 441, + 72, + 539, + 146 + ], + "type": "image", + "image_path": "5b15e5befa59daae27edb06ff6dac5cfad0663005ac286bdaf13107c29d27f45.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 165, + 555, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 165, + 555, + 464 + ], + "spans": [ + { + "bbox": [ + 313, + 165, + 555, + 464 + ], + "type": "text", + "content": "canvas-sizes from " + }, + { + "bbox": [ + 313, + 165, + 555, + 464 + ], + "type": "inline_equation", + "content": "C = \\{32 \\times 32, \\dots, 256 \\times 256\\}" + }, + { + "bbox": [ + 313, + 165, + 555, + 464 + ], + "type": "text", + "content": ". Comparing sparse sketches versus photos, we train an equivalent fine-grained image-based image retrieval (FG-IBIR) where only the query against a positive photo is constructed using random augmentation [11] and resized to corresponding canvas-size " + }, + { + "bbox": [ + 313, + 165, + 555, + 464 + ], + "type": "inline_equation", + "content": "c_i" + }, + { + "bbox": [ + 313, + 165, + 555, + 464 + ], + "type": "text", + "content": ". Gallery photo-features for both are kept pre-computed. While FG-IBIR accuracy falls rapidly (Fig. 2 left), FG-SBIR stays relatively stable against decreasing canvas-sizes, as photos (unlike sketches) containing pixel-dense perfect information, lose a lot of it while down-scaling. Furthermore, positive accuracy of FG-SBIR at " + }, + { + "bbox": [ + 313, + 165, + 555, + 464 + ], + "type": "inline_equation", + "content": "32 \\times 32" + }, + { + "bbox": [ + 313, + 165, + 555, + 464 + ], + "type": "text", + "content": ", shows some sketches to hold sufficient semantic information for retrieval even at minimal canvas-size. Fig. 2 (right) shows overall QMUL-ShoeV2[75] statistics where each bar represents the smallest canvas-size, at which bar-height percentage of sketches achieve perfect retrieval (taking total sketches achieving perfect retrieval at " + }, + { + "bbox": [ + 313, + 165, + 555, + 464 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 313, + 165, + 555, + 464 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 313, + 165, + 555, + 464 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 313, + 165, + 555, + 464 + ], + "type": "text", + "content": "). This infers that every sketch has its own optimal canvas, and fixing it (Fig. 2 left), for the entire dataset would always be sub-optimal. Moreover, the optimal canvas-size for a sketch, depends on its extent of sparsity, for which neither hard-ground-truth exists, nor is it knowable from the user's end. Finally, brute force discovery at every size being infeasible during inference, motivates us to design a learnable adaptive canvas-size selector for sketches." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 468, + 397, + 481 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 468, + 397, + 481 + ], + "spans": [ + { + "bbox": [ + 313, + 468, + 397, + 481 + ], + "type": "text", + "content": "4. Methodology" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 485, + 555, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 485, + 555, + 688 + ], + "spans": [ + { + "bbox": [ + 313, + 485, + 555, + 688 + ], + "type": "text", + "content": "Overview: Our pilot study motivates us to extend a standard FG-SBIR framework towards computational efficiency by first introducing SketchyNetV1 via knowledge distillation, and then taking it further to SketchyNetV2 via dynamic sketch canvas-size selection. Importantly, these are done without hurting the original performance of large-scale slower baseline model. In particular, there are two phases of improvement for computational efficiency. Firstly, we train a teacher model " + }, + { + "bbox": [ + 313, + 485, + 555, + 688 + ], + "type": "inline_equation", + "content": "F_{T}" + }, + { + "bbox": [ + 313, + 485, + 555, + 688 + ], + "type": "text", + "content": " as a standard baseline FG-SBIR model (Sec. 3). Once trained, its rich semantic knowledge is transferred to a smaller student network " + }, + { + "bbox": [ + 313, + 485, + 555, + 688 + ], + "type": "inline_equation", + "content": "F_{st}" + }, + { + "bbox": [ + 313, + 485, + 555, + 688 + ], + "type": "text", + "content": " via a knowledge distillation (KD) paradigm, which we name as SketchyNetV1. Next, we train a canvas-size selector that adaptively decides optimal canvas-size or resolution per sketch for SketchyNetV2 that would be sufficient to preserve the retrieval performance while reducing the number of FLOP operations in feature extraction significantly." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": "Baseline FG-SBIR model: Keeping aside complex pretraining [44] or joint-training [7], we focus on a simple" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28385" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 71, + 250, + 186 + ], + "blocks": [ + { + "bbox": [ + 59, + 71, + 250, + 186 + ], + "lines": [ + { + "bbox": [ + 59, + 71, + 250, + 186 + ], + "spans": [ + { + "bbox": [ + 59, + 71, + 250, + 186 + ], + "type": "image", + "image_path": "cacf7a15e1c70f3a7b612e9ee79238275db7e1d710603e601cd6b0d7674ca74a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 190, + 555, + 236 + ], + "lines": [ + { + "bbox": [ + 55, + 190, + 555, + 236 + ], + "spans": [ + { + "bbox": [ + 55, + 190, + 555, + 236 + ], + "type": "text", + "content": "Figure 3. Two stage framework. (Left) SketchyNetV1: A smaller student network " + }, + { + "bbox": [ + 55, + 190, + 555, + 236 + ], + "type": "inline_equation", + "content": "(F_{st})" + }, + { + "bbox": [ + 55, + 190, + 555, + 236 + ], + "type": "text", + "content": " is trained from a larger pre-trained teacher " + }, + { + "bbox": [ + 55, + 190, + 555, + 236 + ], + "type": "inline_equation", + "content": "(F_T)" + }, + { + "bbox": [ + 55, + 190, + 555, + 236 + ], + "type": "text", + "content": " via KD (§4.1). (Right) We progress to SketchyNetV2 by training a canvas-size selector directed by objectives of increasing performance and reducing compute. It takes sketch as a vector " + }, + { + "bbox": [ + 55, + 190, + 555, + 236 + ], + "type": "inline_equation", + "content": "(s_v)" + }, + { + "bbox": [ + 55, + 190, + 555, + 236 + ], + "type": "text", + "content": " and aims to predict an optimal canvas-size at which the sketch is rasterised " + }, + { + "bbox": [ + 55, + 190, + 555, + 236 + ], + "type": "inline_equation", + "content": "(s_r)" + }, + { + "bbox": [ + 55, + 190, + 555, + 236 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 55, + 190, + 555, + 236 + ], + "type": "inline_equation", + "content": "s_r" + }, + { + "bbox": [ + 55, + 190, + 555, + 236 + ], + "type": "text", + "content": " when processed by " + }, + { + "bbox": [ + 55, + 190, + 555, + 236 + ], + "type": "inline_equation", + "content": "F_{st}" + }, + { + "bbox": [ + 55, + 190, + 555, + 236 + ], + "type": "text", + "content": " minimises overall FLOPS, while retaining accuracy of corresponding full-resolution sketch-image." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 252, + 72, + 547, + 185 + ], + "blocks": [ + { + "bbox": [ + 252, + 72, + 547, + 185 + ], + "lines": [ + { + "bbox": [ + 252, + 72, + 547, + 185 + ], + "spans": [ + { + "bbox": [ + 252, + 72, + 547, + 185 + ], + "type": "image", + "image_path": "1989c4a8b181a9d84247e7176b167e34eaa9155836d5351d6ec01a09df533f8f.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 244, + 297, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 244, + 297, + 399 + ], + "spans": [ + { + "bbox": [ + 54, + 244, + 297, + 399 + ], + "type": "text", + "content": "triple-branch state-of-the-art Siamese network [75] which remains a strong baseline so far [7, 50, 51]. A VGG-16 [30] pre-trained on ImageNet [49] acts as the shared backbone for each branch. Given an input image " + }, + { + "bbox": [ + 54, + 244, + 297, + 399 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 54, + 244, + 297, + 399 + ], + "type": "text", + "content": ", a " + }, + { + "bbox": [ + 54, + 244, + 297, + 399 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 54, + 244, + 297, + 399 + ], + "type": "text", + "content": " dimensional feature embedding is extracted from the backbone network " + }, + { + "bbox": [ + 54, + 244, + 297, + 399 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 54, + 244, + 297, + 399 + ], + "type": "text", + "content": " using global average pooling followed by " + }, + { + "bbox": [ + 54, + 244, + 297, + 399 + ], + "type": "inline_equation", + "content": "l_{2}" + }, + { + "bbox": [ + 54, + 244, + 297, + 399 + ], + "type": "text", + "content": " normalisation as " + }, + { + "bbox": [ + 54, + 244, + 297, + 399 + ], + "type": "inline_equation", + "content": "f_{I} \\in \\mathbb{R}^{d}" + }, + { + "bbox": [ + 54, + 244, + 297, + 399 + ], + "type": "text", + "content": ". The model is trained on triplets comprising features of an anchor sketch " + }, + { + "bbox": [ + 54, + 244, + 297, + 399 + ], + "type": "inline_equation", + "content": "(f_{s})" + }, + { + "bbox": [ + 54, + 244, + 297, + 399 + ], + "type": "text", + "content": ", its matching positive " + }, + { + "bbox": [ + 54, + 244, + 297, + 399 + ], + "type": "inline_equation", + "content": "(f_{p})" + }, + { + "bbox": [ + 54, + 244, + 297, + 399 + ], + "type": "text", + "content": " photo and a random non-matching negative photo " + }, + { + "bbox": [ + 54, + 244, + 297, + 399 + ], + "type": "inline_equation", + "content": "(f_{n})" + }, + { + "bbox": [ + 54, + 244, + 297, + 399 + ], + "type": "text", + "content": ", via a triplet loss [70]. Triplet loss " + }, + { + "bbox": [ + 54, + 244, + 297, + 399 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{\\mathrm{Tri}})" + }, + { + "bbox": [ + 54, + 244, + 297, + 399 + ], + "type": "text", + "content": " aims to minimise the distance between sketch and positive photo, while maximising the same from its negative (n). With " + }, + { + "bbox": [ + 54, + 244, + 297, + 399 + ], + "type": "inline_equation", + "content": "\\delta(a, b) = ||a - b||^2" + }, + { + "bbox": [ + 54, + 244, + 297, + 399 + ], + "type": "text", + "content": ", as a distance function and a margin " + }, + { + "bbox": [ + 54, + 244, + 297, + 399 + ], + "type": "inline_equation", + "content": "m > 0" + }, + { + "bbox": [ + 54, + 244, + 297, + 399 + ], + "type": "text", + "content": ", we have:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 94, + 402, + 295, + 414 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 402, + 295, + 414 + ], + "spans": [ + { + "bbox": [ + 94, + 402, + 295, + 414 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {T r i}} = \\max \\left\\{0, m + \\delta \\left(f _ {s}, f _ {p}\\right) - \\delta \\left(f _ {s}, f _ {n}\\right) \\right\\} \\tag {1}", + "image_path": "70980f39f7635fb72713c0cf3dfd90ab0f4ca45106fbdae7ace5162c765a526a.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 423, + 258, + 436 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 423, + 258, + 436 + ], + "spans": [ + { + "bbox": [ + 55, + 423, + 258, + 436 + ], + "type": "text", + "content": "4.1. SketchyNetV1: Smaller Model via KD" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 438, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 438, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 56, + 438, + 296, + 714 + ], + "type": "text", + "content": "Based on our pilot study (Sec. 3), if we naively train from an ImageNet pre-trained standard smaller network (e.g. MobileNetV2 [54] - usually popular for edge devices), with simple supervised triplet loss (Eq.1) it can hardly reach the accuracy (Table 1) obtained by a large network like VGG-16. Therefore apart from training via a supervised loss on the sketch-positive-negative triplet " + }, + { + "bbox": [ + 56, + 438, + 296, + 714 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{\\mathrm{Tri}}^{st})" + }, + { + "bbox": [ + 56, + 438, + 296, + 714 + ], + "type": "text", + "content": " as in Eq.1), we aim to use the pre-trained larger teacher (e.g. VGG-16) for additional supervision. Accordingly, we adhere to Knowledge Distillation (KD) paradigm [48], a conventional strategy for model compression, to deliver a light-weight student model from the larger teacher network [48]. Applying KD to our FG-SBIR paradigm is however non-trivial. Unlike traditional KD methods, which often involve logit distillation for classification tasks, our scenario involves cross-modal retrieval. Here, the output is a continuous " + }, + { + "bbox": [ + 56, + 438, + 296, + 714 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 56, + 438, + 296, + 714 + ], + "type": "text", + "content": "-dimensional feature within a joint-embedding space. Furthermore, straightforward regression between teacher and student features for sketch and photo branches may encounter compatibility issues, given the disparity in the embedding spaces. Addressing this, if the dimensions of teacher and student embeddings differ, an additional feature transformation layer is necessary for alignment [48]. Overcoming these issues," + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 244, + 555, + 292 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 244, + 555, + 292 + ], + "spans": [ + { + "bbox": [ + 313, + 244, + 555, + 292 + ], + "type": "text", + "content": "we aim to distil the inter-feature distances from respective teacher networks to the student thus preserving structural orientation of features in the teacher's embedding space to that in the student, during KD (Fig. 3 (left))." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 293, + 556, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 293, + 556, + 413 + ], + "spans": [ + { + "bbox": [ + 313, + 293, + 556, + 413 + ], + "type": "text", + "content": "Given features of sketch-positive-negative triplet of teacher " + }, + { + "bbox": [ + 313, + 293, + 556, + 413 + ], + "type": "inline_equation", + "content": "(f_s^T, f_p^T, f_n^T)" + }, + { + "bbox": [ + 313, + 293, + 556, + 413 + ], + "type": "text", + "content": ", we calculate the inter-feature Euclidean distances in the output representation space of teacher " + }, + { + "bbox": [ + 313, + 293, + 556, + 413 + ], + "type": "inline_equation", + "content": "(F_T)" + }, + { + "bbox": [ + 313, + 293, + 556, + 413 + ], + "type": "text", + "content": " using " + }, + { + "bbox": [ + 313, + 293, + 556, + 413 + ], + "type": "inline_equation", + "content": "\\delta(\\cdot, \\cdot)" + }, + { + "bbox": [ + 313, + 293, + 556, + 413 + ], + "type": "text", + "content": " (Eq.1), as " + }, + { + "bbox": [ + 313, + 293, + 556, + 413 + ], + "type": "inline_equation", + "content": "d_{sp}^T = \\delta(f_s^T, f_p^T)" + }, + { + "bbox": [ + 313, + 293, + 556, + 413 + ], + "type": "text", + "content": "; " + }, + { + "bbox": [ + 313, + 293, + 556, + 413 + ], + "type": "inline_equation", + "content": "d_{sn}^T = \\delta(f_s^T, f_n)" + }, + { + "bbox": [ + 313, + 293, + 556, + 413 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 293, + 556, + 413 + ], + "type": "inline_equation", + "content": "d_{pn}^T = \\delta(f_p^T, f_n^T)" + }, + { + "bbox": [ + 313, + 293, + 556, + 413 + ], + "type": "text", + "content": ". Passing the same samples through the student, we similarly obtain " + }, + { + "bbox": [ + 313, + 293, + 556, + 413 + ], + "type": "inline_equation", + "content": "d_{sp}^{st}" + }, + { + "bbox": [ + 313, + 293, + 556, + 413 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 293, + 556, + 413 + ], + "type": "inline_equation", + "content": "d_{sn}^{st}" + }, + { + "bbox": [ + 313, + 293, + 556, + 413 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 293, + 556, + 413 + ], + "type": "inline_equation", + "content": "d_{pn}^{st}" + }, + { + "bbox": [ + 313, + 293, + 556, + 413 + ], + "type": "text", + "content": " using " + }, + { + "bbox": [ + 313, + 293, + 556, + 413 + ], + "type": "inline_equation", + "content": "(f_s^{st}, f_p^{st}, f_n^{st})" + }, + { + "bbox": [ + 313, + 293, + 556, + 413 + ], + "type": "text", + "content": ". A smooth " + }, + { + "bbox": [ + 313, + 293, + 556, + 413 + ], + "type": "inline_equation", + "content": "l_1" + }, + { + "bbox": [ + 313, + 293, + 556, + 413 + ], + "type": "text", + "content": " loss distils the " + }, + { + "bbox": [ + 313, + 293, + 556, + 413 + ], + "type": "inline_equation", + "content": "l_2" + }, + { + "bbox": [ + 313, + 293, + 556, + 413 + ], + "type": "text", + "content": " distances from the teacher to the student. With " + }, + { + "bbox": [ + 313, + 293, + 556, + 413 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\delta}" + }, + { + "bbox": [ + 313, + 293, + 556, + 413 + ], + "type": "text", + "content": " as a Huber-loss [27], relational distillation loss for sketch-photo pair between teacher and student is given as," + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 357, + 415, + 473, + 429 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 357, + 415, + 473, + 429 + ], + "spans": [ + { + "bbox": [ + 357, + 415, + 473, + 429 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {R K D}} ^ {s p} = \\mathcal {L} _ {\\delta} \\left(d _ {s p} ^ {T}, d _ {s p} ^ {s t}\\right) \\quad \\text {w h e r e ,}", + "image_path": "87bc24218ec1b1249acc3498a970f58c036f19b4b8827c52832ac75453c54d46.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 347, + 432, + 553, + 460 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 432, + 553, + 460 + ], + "spans": [ + { + "bbox": [ + 347, + 432, + 553, + 460 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\delta} (a, b) = \\left\\{ \\begin{array}{l l} \\frac {1}{2} (a - b) ^ {2} & \\text {i f} | a - b | < \\beta \\\\ \\beta (| a - b | - \\frac {1}{2} \\beta) & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {2}", + "image_path": "e0afb5fd7decb9abc653a81f0f15d317e2b1108d003dff126911a534f7ca56d6.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 464, + 555, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 464, + 555, + 498 + ], + "spans": [ + { + "bbox": [ + 313, + 464, + 555, + 498 + ], + "type": "text", + "content": "Similarly, we obtain " + }, + { + "bbox": [ + 313, + 464, + 555, + 498 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{RKD}}^{sn}" + }, + { + "bbox": [ + 313, + 464, + 555, + 498 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 464, + 555, + 498 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{RKD}}^{pn}" + }, + { + "bbox": [ + 313, + 464, + 555, + 498 + ], + "type": "text", + "content": ", to compute distillation loss, " + }, + { + "bbox": [ + 313, + 464, + 555, + 498 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{RKD}} = \\mathcal{L}_{\\mathrm{RKD}}^{sp} + \\mathcal{L}_{\\mathrm{RKD}}^{sn} + \\mathcal{L}_{\\mathrm{RKD}}^{pn}" + }, + { + "bbox": [ + 313, + 464, + 555, + 498 + ], + "type": "text", + "content": ". With hyperparameter " + }, + { + "bbox": [ + 313, + 464, + 555, + 498 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 464, + 555, + 498 + ], + "type": "text", + "content": ", overall student-training objective becomes:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 380, + 503, + 553, + 515 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 380, + 503, + 553, + 515 + ], + "spans": [ + { + "bbox": [ + 380, + 503, + 553, + 515 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {t r n}} ^ {s t} = \\lambda \\mathcal {L} _ {\\mathrm {T r i}} + (1 - \\lambda) \\mathcal {L} _ {\\mathrm {R K D}} \\tag {3}", + "image_path": "8e713f28a4dd9d51598e3fe47cf745e5e82dcac4128c4b986b5dee88ccb04adc.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 517, + 555, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 517, + 555, + 574 + ], + "spans": [ + { + "bbox": [ + 313, + 517, + 555, + 574 + ], + "type": "text", + "content": "The student trains on varying sketch-canvas-sizes " + }, + { + "bbox": [ + 313, + 517, + 555, + 574 + ], + "type": "inline_equation", + "content": "(C)" + }, + { + "bbox": [ + 313, + 517, + 555, + 574 + ], + "type": "text", + "content": ", using their respective full resolution for teacher, to ensure scale-invariance as " + }, + { + "bbox": [ + 313, + 517, + 555, + 574 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{tm}}^{st*} = \\frac{1}{4}\\sum_{i=1}^{4}\\mathcal{L}_{\\mathrm{tm}}^{st(c_i)}" + }, + { + "bbox": [ + 313, + 517, + 555, + 574 + ], + "type": "text", + "content": ". However, as optimal canvas-size varies across sketches (Fig. 2), we need a learnable canvas-size selector, dynamically predicting the optimal size for each sketch." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 578, + 549, + 590 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 578, + 549, + 590 + ], + "spans": [ + { + "bbox": [ + 313, + 578, + 549, + 590 + ], + "type": "text", + "content": "4.2. SketchyNetV2: Adaptive canvas-size Selector" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 593, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 593, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 593, + 556, + 715 + ], + "type": "text", + "content": "Overview: Understanding that an input at a lower resolution, would incur lesser FLOPs on evaluation than a higher one, we upgrade SketchyNetV1 (" + }, + { + "bbox": [ + 313, + 593, + 556, + 715 + ], + "type": "inline_equation", + "content": "F_{st}" + }, + { + "bbox": [ + 313, + 593, + 556, + 715 + ], + "type": "text", + "content": ") to a more computationally efficient SketchyNetV2 by introducing an adaptive canvas-size selector. It is designed in the vector modality as it is much cheaper [71] than its raster-image counterpart, and can dynamically encode the varying abstraction level of sketches. Furthermore, predicting the optimal canvas-size from vector modality, would make the corresponding rasterisation operation " + }, + { + "bbox": [ + 313, + 593, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{R}^c(\\cdot)" + }, + { + "bbox": [ + 313, + 593, + 556, + 715 + ], + "type": "text", + "content": " at lower sizes cheaper, besides reducing FLOPs in feature extraction. Taking sketch vector " + }, + { + "bbox": [ + 313, + 593, + 556, + 715 + ], + "type": "inline_equation", + "content": "s_v" + }, + { + "bbox": [ + 313, + 593, + 556, + 715 + ], + "type": "text", + "content": " as input, " + }, + { + "bbox": [ + 313, + 593, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\psi_C" + }, + { + "bbox": [ + 313, + 593, + 556, + 715 + ], + "type": "text", + "content": " predicts the probability " + }, + { + "bbox": [ + 313, + 593, + 556, + 715 + ], + "type": "inline_equation", + "content": "p(c|s_v)" + }, + { + "bbox": [ + 313, + 593, + 556, + 715 + ], + "type": "text", + "content": " of choosing one of " + }, + { + "bbox": [ + 313, + 593, + 556, + 715 + ], + "type": "inline_equation", + "content": "K" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "28386" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 106 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 106 + ], + "type": "text", + "content": "discreet canvas-sizes from a set of " + }, + { + "bbox": [ + 55, + 72, + 294, + 106 + ], + "type": "inline_equation", + "content": "C = \\{32 \\times 32, \\dots, 256 \\times 256\\}" + }, + { + "bbox": [ + 55, + 72, + 294, + 106 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 55, + 72, + 294, + 106 + ], + "type": "inline_equation", + "content": "s_v" + }, + { + "bbox": [ + 55, + 72, + 294, + 106 + ], + "type": "text", + "content": " is then rasterised to " + }, + { + "bbox": [ + 55, + 72, + 294, + 106 + ], + "type": "inline_equation", + "content": "s_r^c \\in \\mathbb{R}^{c \\times c \\times 3}" + }, + { + "bbox": [ + 55, + 72, + 294, + 106 + ], + "type": "text", + "content": " with the optimal canvas-size " + }, + { + "bbox": [ + 55, + 72, + 294, + 106 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 72, + 294, + 106 + ], + "type": "text", + "content": ", as " + }, + { + "bbox": [ + 55, + 72, + 294, + 106 + ], + "type": "inline_equation", + "content": "\\mathcal{R}^c(\\cdot) : s_v \\to s_r^c" + }, + { + "bbox": [ + 55, + 72, + 294, + 106 + ], + "type": "text", + "content": ", and fed to " + }, + { + "bbox": [ + 55, + 72, + 294, + 106 + ], + "type": "inline_equation", + "content": "F_{st}" + }, + { + "bbox": [ + 55, + 72, + 294, + 106 + ], + "type": "text", + "content": " for retrieval." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 106, + 295, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 106, + 295, + 216 + ], + "spans": [ + { + "bbox": [ + 54, + 106, + 295, + 216 + ], + "type": "text", + "content": "Selecting an optimal canvas-size however is an ill-posed problem. Firstly, there are no explicit labels representing the optimal canvas-size for a sketch. Secondly, annotating the optimal canvas-size for the whole dataset via brute-force iteration is computationally impractical. We therefore use the pre-trained " + }, + { + "bbox": [ + 54, + 106, + 295, + 216 + ], + "type": "inline_equation", + "content": "F_{st}" + }, + { + "bbox": [ + 54, + 106, + 295, + 216 + ], + "type": "text", + "content": " as a critic to guide the learning of sketch canvas-size selector via Reinforcement Learning [29] as rasterisation is a non-differentiable operation. There are two major objectives here: a) retain the original resolution accuracy and b) encourage the use of the lower resolution to improve the computational efficiency." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 218, + 296, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 218, + 296, + 470 + ], + "spans": [ + { + "bbox": [ + 56, + 218, + 296, + 470 + ], + "type": "text", + "content": "Model: Any sequential network from RNN [17] or Transformer [35] families can be used to encode the vector-sketch " + }, + { + "bbox": [ + 56, + 218, + 296, + 470 + ], + "type": "inline_equation", + "content": "s_v" + }, + { + "bbox": [ + 56, + 218, + 296, + 470 + ], + "type": "text", + "content": ". To recap, sketch-vector " + }, + { + "bbox": [ + 56, + 218, + 296, + 470 + ], + "type": "inline_equation", + "content": "(s_v)" + }, + { + "bbox": [ + 56, + 218, + 296, + 470 + ], + "type": "text", + "content": " represents a sequence of points " + }, + { + "bbox": [ + 56, + 218, + 296, + 470 + ], + "type": "inline_equation", + "content": "[v_1, v_2, \\dots, v_{\\mathrm{T}}]" + }, + { + "bbox": [ + 56, + 218, + 296, + 470 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 56, + 218, + 296, + 470 + ], + "type": "inline_equation", + "content": "v_t = (x_t, y_t, q_t^1, q_t^2, q_t^3) \\in \\mathbb{R}^{T \\times 5}" + }, + { + "bbox": [ + 56, + 218, + 296, + 470 + ], + "type": "text", + "content": "; " + }, + { + "bbox": [ + 56, + 218, + 296, + 470 + ], + "type": "inline_equation", + "content": "\\mathrm{T}" + }, + { + "bbox": [ + 56, + 218, + 296, + 470 + ], + "type": "text", + "content": " is the sequence length, " + }, + { + "bbox": [ + 56, + 218, + 296, + 470 + ], + "type": "inline_equation", + "content": "(x_t, y_t)" + }, + { + "bbox": [ + 56, + 218, + 296, + 470 + ], + "type": "text", + "content": " denotes the absolute coordinates in a normalised " + }, + { + "bbox": [ + 56, + 218, + 296, + 470 + ], + "type": "inline_equation", + "content": "c \\times c" + }, + { + "bbox": [ + 56, + 218, + 296, + 470 + ], + "type": "text", + "content": " canvas, while the last three represent penstates [22]. Furthermore, the complexity of canvas-size selector being dependent on sequence-length (T) varying across sketches, we use Douglas Peucker Algorithm [66] to limit sequence-length across all sketch-samples at " + }, + { + "bbox": [ + 56, + 218, + 296, + 470 + ], + "type": "inline_equation", + "content": "\\mathrm{T}_{max}" + }, + { + "bbox": [ + 56, + 218, + 296, + 470 + ], + "type": "text", + "content": " without losing visual representation. We feed " + }, + { + "bbox": [ + 56, + 218, + 296, + 470 + ], + "type": "inline_equation", + "content": "v_t" + }, + { + "bbox": [ + 56, + 218, + 296, + 470 + ], + "type": "text", + "content": " at every time step to a simple 1-layer GRU network with " + }, + { + "bbox": [ + 56, + 218, + 296, + 470 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{d_v}" + }, + { + "bbox": [ + 56, + 218, + 296, + 470 + ], + "type": "text", + "content": " hidden states and take the final hidden state as the encoded latent representation " + }, + { + "bbox": [ + 56, + 218, + 296, + 470 + ], + "type": "inline_equation", + "content": "(f_{s^{\\mathrm{T}}} \\in \\mathbb{R}^{d_v})" + }, + { + "bbox": [ + 56, + 218, + 296, + 470 + ], + "type": "text", + "content": ". Next, we apply a linear layer " + }, + { + "bbox": [ + 56, + 218, + 296, + 470 + ], + "type": "inline_equation", + "content": "(\\gamma)" + }, + { + "bbox": [ + 56, + 218, + 296, + 470 + ], + "type": "text", + "content": " to get " + }, + { + "bbox": [ + 56, + 218, + 296, + 470 + ], + "type": "inline_equation", + "content": "p(c|s_v) = \\text{softmax}(W_\\gamma f_{s^{\\mathrm{T}}} + b_\\gamma)" + }, + { + "bbox": [ + 56, + 218, + 296, + 470 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 56, + 218, + 296, + 470 + ], + "type": "inline_equation", + "content": "W_\\gamma \\in \\mathbb{R}^{d_v \\times K}" + }, + { + "bbox": [ + 56, + 218, + 296, + 470 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 56, + 218, + 296, + 470 + ], + "type": "inline_equation", + "content": "b \\in \\mathbb{R}^K" + }, + { + "bbox": [ + 56, + 218, + 296, + 470 + ], + "type": "text", + "content": ", as shown in Fig. 3 (right). Given " + }, + { + "bbox": [ + 56, + 218, + 296, + 470 + ], + "type": "inline_equation", + "content": "p(c|s_v) \\in \\mathbb{R}^K" + }, + { + "bbox": [ + 56, + 218, + 296, + 470 + ], + "type": "text", + "content": " over " + }, + { + "bbox": [ + 56, + 218, + 296, + 470 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 56, + 218, + 296, + 470 + ], + "type": "text", + "content": " different canvas-sizes for every " + }, + { + "bbox": [ + 56, + 218, + 296, + 470 + ], + "type": "inline_equation", + "content": "s_v" + }, + { + "bbox": [ + 56, + 218, + 296, + 470 + ], + "type": "text", + "content": ", canvas-size " + }, + { + "bbox": [ + 56, + 218, + 296, + 470 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 56, + 218, + 296, + 470 + ], + "type": "text", + "content": " can be sampled from categorical distribution as " + }, + { + "bbox": [ + 56, + 218, + 296, + 470 + ], + "type": "inline_equation", + "content": "c_{pred} \\sim \\text{categorical}([p(c_1|s_v), \\dots, p(c_K|s_v)])" + }, + { + "bbox": [ + 56, + 218, + 296, + 470 + ], + "type": "text", + "content": ". Accordingly, " + }, + { + "bbox": [ + 56, + 218, + 296, + 470 + ], + "type": "inline_equation", + "content": "s_v" + }, + { + "bbox": [ + 56, + 218, + 296, + 470 + ], + "type": "text", + "content": " is rasterised to " + }, + { + "bbox": [ + 56, + 218, + 296, + 470 + ], + "type": "inline_equation", + "content": "s_r" + }, + { + "bbox": [ + 56, + 218, + 296, + 470 + ], + "type": "text", + "content": " at canvas-size " + }, + { + "bbox": [ + 56, + 218, + 296, + 470 + ], + "type": "inline_equation", + "content": "c_{pred}" + }, + { + "bbox": [ + 56, + 218, + 296, + 470 + ], + "type": "text", + "content": ", which is then fed to " + }, + { + "bbox": [ + 56, + 218, + 296, + 470 + ], + "type": "inline_equation", + "content": "F_{st}" + }, + { + "bbox": [ + 56, + 218, + 296, + 470 + ], + "type": "text", + "content": ". In particular, given a sketch " + }, + { + "bbox": [ + 56, + 218, + 296, + 470 + ], + "type": "inline_equation", + "content": "s_v" + }, + { + "bbox": [ + 56, + 218, + 296, + 470 + ], + "type": "text", + "content": ", canvas-size selector " + }, + { + "bbox": [ + 56, + 218, + 296, + 470 + ], + "type": "inline_equation", + "content": "\\psi_C" + }, + { + "bbox": [ + 56, + 218, + 296, + 470 + ], + "type": "text", + "content": " acting as a policy network, takes action of selecting the canvas-size " + }, + { + "bbox": [ + 56, + 218, + 296, + 470 + ], + "type": "inline_equation", + "content": "c_{pred}" + }, + { + "bbox": [ + 56, + 218, + 296, + 470 + ], + "type": "text", + "content": " from action space " + }, + { + "bbox": [ + 56, + 218, + 296, + 470 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 56, + 218, + 296, + 470 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 56, + 218, + 296, + 470 + ], + "type": "inline_equation", + "content": "\\psi_C" + }, + { + "bbox": [ + 56, + 218, + 296, + 470 + ], + "type": "text", + "content": " is optimised over a reward calculated by the SketchyNetV1 model " + }, + { + "bbox": [ + 56, + 218, + 296, + 470 + ], + "type": "inline_equation", + "content": "(F_{st})" + }, + { + "bbox": [ + 56, + 218, + 296, + 470 + ], + "type": "text", + "content": " acting as the critic." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 472, + 295, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 472, + 295, + 635 + ], + "spans": [ + { + "bbox": [ + 55, + 472, + 295, + 635 + ], + "type": "text", + "content": "Reward Design: The objectives of our canvas selector are to choose an optimal canvas-size such that: (i) accuracy is retained (ii) overall FLOPs is minimised. We therefore design our reward from two perspectives. Now as " + }, + { + "bbox": [ + 55, + 472, + 295, + 635 + ], + "type": "inline_equation", + "content": "F_{st}" + }, + { + "bbox": [ + 55, + 472, + 295, + 635 + ], + "type": "text", + "content": " is fixed, we pre-compute the features of all gallery photos. During training, we only extract the feature of rasterised sketch and calculate its corresponding rank using the pre-computed photo-features. Furthermore, taking one sketch we pre-compute FLOPs for every canvas-size in " + }, + { + "bbox": [ + 55, + 472, + 295, + 635 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 55, + 472, + 295, + 635 + ], + "type": "text", + "content": ". From accuracy perspective, the aim is to select the optimal canvas-size " + }, + { + "bbox": [ + 55, + 472, + 295, + 635 + ], + "type": "inline_equation", + "content": "c_{\\mathrm{opt}}" + }, + { + "bbox": [ + 55, + 472, + 295, + 635 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 55, + 472, + 295, + 635 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 55, + 472, + 295, + 635 + ], + "type": "text", + "content": " for each " + }, + { + "bbox": [ + 55, + 472, + 295, + 635 + ], + "type": "inline_equation", + "content": "s_v" + }, + { + "bbox": [ + 55, + 472, + 295, + 635 + ], + "type": "text", + "content": " such that the rendered " + }, + { + "bbox": [ + 55, + 472, + 295, + 635 + ], + "type": "inline_equation", + "content": "s_r" + }, + { + "bbox": [ + 55, + 472, + 295, + 635 + ], + "type": "text", + "content": " can retrieve the matching photo " + }, + { + "bbox": [ + 55, + 472, + 295, + 635 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 55, + 472, + 295, + 635 + ], + "type": "text", + "content": " at numerically lowest rank " + }, + { + "bbox": [ + 55, + 472, + 295, + 635 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 55, + 472, + 295, + 635 + ], + "type": "text", + "content": " (best " + }, + { + "bbox": [ + 55, + 472, + 295, + 635 + ], + "type": "inline_equation", + "content": "r = 1" + }, + { + "bbox": [ + 55, + 472, + 295, + 635 + ], + "type": "text", + "content": "). Following conventional norm of reward maximisation, we define accuracy reward " + }, + { + "bbox": [ + 55, + 472, + 295, + 635 + ], + "type": "inline_equation", + "content": "(R_{\\mathrm{acc}})" + }, + { + "bbox": [ + 55, + 472, + 295, + 635 + ], + "type": "text", + "content": " as weighted (by hyper-parameters " + }, + { + "bbox": [ + 55, + 472, + 295, + 635 + ], + "type": "inline_equation", + "content": "\\lambda_r" + }, + { + "bbox": [ + 55, + 472, + 295, + 635 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 472, + 295, + 635 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{Tri}}" + }, + { + "bbox": [ + 55, + 472, + 295, + 635 + ], + "type": "text", + "content": ") summation of inverse of the rank (r) and negative triplet loss (following Eq.1) as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 120, + 637, + 294, + 648 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 120, + 637, + 294, + 648 + ], + "spans": [ + { + "bbox": [ + 120, + 637, + 294, + 648 + ], + "type": "interline_equation", + "content": "R _ {\\mathrm {a c c}} = \\lambda_ {r} (1 / r) + \\lambda_ {\\mathrm {T r i}} (- \\mathcal {L} _ {\\mathrm {T r i}}) \\tag {4}", + "image_path": "576632b097a1215e08053b48393d4d8fc897791a4e71a96307d909bec6f881ca.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 647, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 647, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 647, + 295, + 713 + ], + "type": "text", + "content": "From the compute perspective, the selection criterion should reward choosing a lower canvas-size. Moreover, higher performance being naturally inclined towards a higher-canvas-size thus increasing overall compute, we need an objective dedicated towards reducing computational cost. We thus propose a FLOPs constraint regularisation to guide the learning of canvas selector as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 393, + 69, + 553, + 96 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 393, + 69, + 553, + 96 + ], + "spans": [ + { + "bbox": [ + 393, + 69, + 553, + 96 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {F}} = \\frac {\\sum_ {j = 1} ^ {K} \\left(q _ {j} \\cdot \\eta_ {j}\\right)}{q _ {\\max } - q _ {\\min }} \\tag {5}", + "image_path": "da5ba4c259a4ac1bec098a0e2d3e8f208dd1f25e3345366f2bade942887b8173.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 100, + 553, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 100, + 553, + 163 + ], + "spans": [ + { + "bbox": [ + 313, + 100, + 553, + 163 + ], + "type": "text", + "content": "where, " + }, + { + "bbox": [ + 313, + 100, + 553, + 163 + ], + "type": "inline_equation", + "content": "\\eta_{i} = p(c_{i}|s_{v})" + }, + { + "bbox": [ + 313, + 100, + 553, + 163 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 100, + 553, + 163 + ], + "type": "inline_equation", + "content": "q_{j}" + }, + { + "bbox": [ + 313, + 100, + 553, + 163 + ], + "type": "text", + "content": " is the pre-computed FLOPs value for the j-th canvas-size with maximum at " + }, + { + "bbox": [ + 313, + 100, + 553, + 163 + ], + "type": "inline_equation", + "content": "q_{max}" + }, + { + "bbox": [ + 313, + 100, + 553, + 163 + ], + "type": "text", + "content": " and minimum at " + }, + { + "bbox": [ + 313, + 100, + 553, + 163 + ], + "type": "inline_equation", + "content": "q_{min}" + }, + { + "bbox": [ + 313, + 100, + 553, + 163 + ], + "type": "text", + "content": ". Treating " + }, + { + "bbox": [ + 313, + 100, + 553, + 163 + ], + "type": "inline_equation", + "content": "(-L_{\\mathrm{F}})" + }, + { + "bbox": [ + 313, + 100, + 553, + 163 + ], + "type": "text", + "content": " as a reward from the compute perspective, we define our compute reward as " + }, + { + "bbox": [ + 313, + 100, + 553, + 163 + ], + "type": "inline_equation", + "content": "R_{\\mathrm{comp}} = (-L_{\\mathrm{F}})" + }, + { + "bbox": [ + 313, + 100, + 553, + 163 + ], + "type": "text", + "content": ". Taking " + }, + { + "bbox": [ + 313, + 100, + 553, + 163 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{F}}" + }, + { + "bbox": [ + 313, + 100, + 553, + 163 + ], + "type": "text", + "content": " as a balancing hyper-parameter, we combine both perspectives as our total reward:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 375, + 163, + 553, + 175 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 375, + 163, + 553, + 175 + ], + "spans": [ + { + "bbox": [ + 375, + 163, + 553, + 175 + ], + "type": "interline_equation", + "content": "R _ {\\mathrm {T o t}} = \\lambda_ {\\mathrm {F}} R _ {\\mathrm {c o m p}} + (1 - \\lambda_ {\\mathrm {F}}) R _ {\\mathrm {a c c}} \\tag {6}", + "image_path": "0bed22ed5700a7dee7025069c4f629d145f165ba4b219a2596964bb681f593b9.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 179, + 553, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 179, + 553, + 255 + ], + "spans": [ + { + "bbox": [ + 313, + 179, + 553, + 255 + ], + "type": "text", + "content": "Optimisation Objective: Finally, we train the canvas-size selector using popular Policy Gradient [61] method which is both simple to implement and requires less hyper-parameter tuning than other state-of-the-art alternatives [56]. If " + }, + { + "bbox": [ + 313, + 179, + 553, + 255 + ], + "type": "inline_equation", + "content": "p(c|s_v^i)" + }, + { + "bbox": [ + 313, + 179, + 553, + 255 + ], + "type": "text", + "content": " be the probability of selecting a specific canvas-size for " + }, + { + "bbox": [ + 313, + 179, + 553, + 255 + ], + "type": "inline_equation", + "content": "i^{th}" + }, + { + "bbox": [ + 313, + 179, + 553, + 255 + ], + "type": "text", + "content": " sketch sample " + }, + { + "bbox": [ + 313, + 179, + 553, + 255 + ], + "type": "inline_equation", + "content": "s_v^i" + }, + { + "bbox": [ + 313, + 179, + 553, + 255 + ], + "type": "text", + "content": " and the corresponding reward be " + }, + { + "bbox": [ + 313, + 179, + 553, + 255 + ], + "type": "inline_equation", + "content": "R_{\\mathrm{tot}}^i" + }, + { + "bbox": [ + 313, + 179, + 553, + 255 + ], + "type": "text", + "content": ", the objective function for policy network (canvas-size selector) over a batch of size " + }, + { + "bbox": [ + 313, + 179, + 553, + 255 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 313, + 179, + 553, + 255 + ], + "type": "text", + "content": " becomes:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 363, + 258, + 553, + 287 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 363, + 258, + 553, + 287 + ], + "spans": [ + { + "bbox": [ + 363, + 258, + 553, + 287 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {P G} (\\theta) = - \\frac {1}{B} \\sum_ {i = 1} ^ {B} \\log p \\left(c \\mid s _ {v} ^ {i}\\right) \\cdot R _ {\\text {T o t}} ^ {i} \\tag {7}", + "image_path": "2a1ba00f77a40ded06fe2c9048c2a284900bc74eaa478ba4734212c32be40cb6.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 285, + 395, + 298 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 285, + 395, + 298 + ], + "spans": [ + { + "bbox": [ + 314, + 285, + 395, + 298 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 302, + 555, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 302, + 555, + 434 + ], + "spans": [ + { + "bbox": [ + 313, + 302, + 555, + 434 + ], + "type": "text", + "content": "Datasets: We evaluated on the following publicly available datasets having fine-grained sketch-photo association. QMUL-ShoeV2 [75] and QMUL-ChairV2 [75] contains 6730/2000 and 1800/400 sketches/photos respectively. Keeping 679/200 (525/100) sketches/photos for evaluation from ShoeV2 (ChairV2) respectively we use the rest for training. Sketchy [55] contains 125 categories with 100 photos each, having around 5 sketches per photo. While, training uses a standard train-test split [55] of 9:1, during inference we construct a challenging gallery using photos across one category for retrieval. For scene-level FG-SBIR evaluation we use FS-COCO [15] which includes 10,000 unique sketch-photo pairs with a 7:3 train/test split." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 437, + 555, + 633 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 555, + 633 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 555, + 633 + ], + "type": "text", + "content": "Implementation Details: Considering a standard large FG-SBIR network as a teacher, our proposed student uses a cheaper backbone feature extractor MobileNetV2 [54], a standard for mobile end-devices. Margin is set to " + }, + { + "bbox": [ + 313, + 437, + 555, + 633 + ], + "type": "inline_equation", + "content": "m = 0.2" + }, + { + "bbox": [ + 313, + 437, + 555, + 633 + ], + "type": "text", + "content": " for triplet loss (Secs. 3 and 4.1). We use an Adam optimiser at a learning rate of 0.0001 and a batchsize of 16 for 200 epochs. For practicality, we chose discrete canvas-sizes (" + }, + { + "bbox": [ + 313, + 437, + 555, + 633 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 313, + 437, + 555, + 633 + ], + "type": "text", + "content": ") instead of a continuous action space, where " + }, + { + "bbox": [ + 313, + 437, + 555, + 633 + ], + "type": "inline_equation", + "content": "C = \\{32 \\times 32, 64 \\times 64, 128 \\times 128, 256 \\times 256\\}" + }, + { + "bbox": [ + 313, + 437, + 555, + 633 + ], + "type": "text", + "content": ". Canvas-size selector (" + }, + { + "bbox": [ + 313, + 437, + 555, + 633 + ], + "type": "inline_equation", + "content": "\\psi_C" + }, + { + "bbox": [ + 313, + 437, + 555, + 633 + ], + "type": "text", + "content": ") is modelled using a simple GRU network, with embedding size 128, and trained via reinforcement learning with a learning rate of 0.0001 and a batchsize of 32 for 500 epochs, keeping " + }, + { + "bbox": [ + 313, + 437, + 555, + 633 + ], + "type": "inline_equation", + "content": "T_{max} = 100" + }, + { + "bbox": [ + 313, + 437, + 555, + 633 + ], + "type": "text", + "content": " for vector-sketches. For the action space, we take varying canvas-sizes from " + }, + { + "bbox": [ + 313, + 437, + 555, + 633 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 313, + 437, + 555, + 633 + ], + "type": "text", + "content": " and set " + }, + { + "bbox": [ + 313, + 437, + 555, + 633 + ], + "type": "inline_equation", + "content": "\\beta = 1" + }, + { + "bbox": [ + 313, + 437, + 555, + 633 + ], + "type": "text", + "content": " in Eq.2. " + }, + { + "bbox": [ + 313, + 437, + 555, + 633 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 437, + 555, + 633 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 437, + 555, + 633 + ], + "type": "inline_equation", + "content": "\\lambda_r" + }, + { + "bbox": [ + 313, + 437, + 555, + 633 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 437, + 555, + 633 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{Tri}}" + }, + { + "bbox": [ + 313, + 437, + 555, + 633 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 437, + 555, + 633 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{F}}" + }, + { + "bbox": [ + 313, + 437, + 555, + 633 + ], + "type": "text", + "content": " are empirically set to 0.5, 0.4, 0.48 and 0.35 (Fig.5) respectively. While training " + }, + { + "bbox": [ + 313, + 437, + 555, + 633 + ], + "type": "inline_equation", + "content": "\\psi_C" + }, + { + "bbox": [ + 313, + 437, + 555, + 633 + ], + "type": "text", + "content": ", sketches are rendered at different completion (30%, 35%, ... 100%) levels to simulate varying levels of abstraction, ensuring a sketch-abstraction-aware canvas selector (e.g., a smaller canvas for less detailed sketch)." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 636, + 556, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 636, + 556, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 636, + 556, + 712 + ], + "type": "text", + "content": "Evaluation Metric: Accuracy for FG-SBIR is the percentage of sketches where true-match photos are ranked in the Top-q (q = 1,10) lists. Computational efficiency is measured via parameter-count and average FLOPs – lower is better. For average FLOPs in Table 2, we take the combined sum of FLOPs per sample [12] from (i) canvas-size selector and (ii) retrieval model, using the predicted resolution, and average across the validation set." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28387" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 82, + 553, + 411 + ], + "blocks": [ + { + "bbox": [ + 174, + 71, + 436, + 81 + ], + "lines": [ + { + "bbox": [ + 174, + 71, + 436, + 81 + ], + "spans": [ + { + "bbox": [ + 174, + 71, + 436, + 81 + ], + "type": "text", + "content": "Table 2. Quantitative Analysis on FG-SBIR. Best viewed when zoomed." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 82, + 553, + 411 + ], + "lines": [ + { + "bbox": [ + 58, + 82, + 553, + 411 + ], + "spans": [ + { + "bbox": [ + 58, + 82, + 553, + 411 + ], + "type": "table", + "html": "
MethodsCanvas-size c×cParams (mil.)ShoeV2 [75]ChairV2 [75]Sketchy [55]FSCOCO [15]
Top1 (%)Top10 (%)FLOPS (G)Top1 (%)Top10 (%)FLOPS (G)Top1 (%)Top10 (%)FLOPs (G)Top1 (%)Top10 (%)FLOPS (G)
State-of-the-ArtsTriplet-SN [75]32x328.7509.77 (↓18.94)24.82 (↓46.74)0.08315.61 (↓32.04)29.86 (↓54.38)0.0834.23 (↓11.12)12.58 (↓23.13)0.0830.76 (↓3.94)6.22 (↓14.8)0.083
64x648.7517.63 (↓11.08)44.96 (↓26.60)0.33830.25 (↓17.4)54.21 (↓30.03)0.3389.68 (↓5.67)22.45 (↓13.26)0.3381.85 (↓2.85)7.33 (↓13.69)0.338
128x1288.7525.01 (↓3.70)62.35 (↓9.21)1.39740.61 (↓7.04)72.68 (↓11.56)1.39712.98 (↓2.37)29.63 (↓6.08)1.3973.28 (↓1.42)15.92 (↓5.1)1.397
256x2568.7528.7171.565.28047.6584.245.28015.3535.715.2804.721.025.280
SketchyNetV12.2228.46 (↓0.25)69.01 (↓2.55)0.83346.28 (↓0.37)82.33 (↓1.91)0.83314.85 (↓0.5)34.68 (↓1.03)0.8334.22 (↓0.48)18.33 (↓2.69)0.833
SketchyNetV22.2727.89 (↓0.82)68.76 (↓2.80)0.26445.98 (↓1.67)80.14 (↓4.10)0.29414.21 (↓1.14)34.16 (↓1.55)0.3213.98 (↓0.72)17.64 (↓3.38)0.423
HOLEE-SN [59]32x329.3110.84 (↓20.9)27.28 (↓48.5)0.09618.72 (↓34.69)31.68 (↓55.88)0.0966.01 (↓10.69)14.12 (↓24.78)0.0960.86 (↓4.04)6.93 (↓14.78)0.096
64x649.3119.83 (↓11.91)51.72 (↓24.06)0.38734.32 (↓19.09)56.46 (↓31.1)0.3879.92 (↓6.78)24.38 (↓14.52)0.3872.11 (↓2.79)7.86 (↓13.85)0.387
128x1289.3126.78 (↓4.96)66.28 (↓9.5)1.54545.74 (↓7.67)75.59 (↓11.97)1.54513.81 (↓2.89)33.68 (↓5.22)1.5453.65 (↓1.25)16.32 (↓5.39)1.545
256x2569.3131.7475.785.75853.4187.565.75816.7038.905.7584.921.715.758
SketchyNetV12.2231.59 (↓0.15)73.96 (↓1.82)0.83353.27 (↓0.14)86.93 (↓0.63)0.83316.25 (↓0.45)38.21 (↓0.69)0.8334.56 (↓0.34)19.92 (↓1.79)0.833
SketchyNetV22.2730.86 (↓0.88)72.56 (↓3.22)0.25952.74 (↓0.67)84.02 (↓3.54)0.28915.91 (↓0.79)37.88 (↓1.02)0.3154.04 (↓0.86)18.63 (↓3.08)0.417
Triplet-RL [6]32x3222.111.98 (↓22.12)28.11 (↓50.71)0.14220.53 (↓36.01)32.59 (↓57.02)0.1421.76 (↓2.94)3.88 (↓6.47)0.142
64x6422.121.28 (↓12.82)54.71 (↓24.11)0.57736.09 (↓20.45)57.33 (↓32.28)0.5773.04 (↓1.66)6.79 (↓3.56)0.577
128x12822.128.83 (↓5.27)67.84 (↓10.98)2.29948.38 (↓8.16)77.25 (↓12.36)2.2994.12 (↓0.58)9.05 (↓1.30)2.299
256x25622.134.1078.826.04156.5489.616.0414.7010.356.041
SketchyNetV12.2233.88 (↓0.22)77.15 (↓1.67)0.83355.92 (↓0.62)88.20 (↓1.41)0.8334.59 (↓0.11)10.21 (↓0.14)0.833
SketchyNetV22.2733.26 (↓0.84)76.84 (↓1.98)0.25555.14 (↓1.40)87.31 (↓2.30)0.2854.51 (↓0.19)9.89 (↓0.46)0.308
StyleVAE [51]32x3225.3712.68 (↓23.79)28.69 (↓53.14)0.12522.48 (↓40.38)32.10 (↓59.04)0.12506.92 (↓12.7)16.25 (↓29.53)0.125
64x6425.3722.93 (↓13.54)57.71 (↓24.12)0.50840.26 (↓22.6)58.23 (↓32.91)0.50812.46 (↓7.16)29.62 (↓16.16)0.508
128x12825.3730.91 (↓5.56)70.06 (↓11.77)2.02453.88 (↓8.98)78.64 (↓12.5)2.02417.15 (↓2.47)38.84 (↓6.94)2.024
256x25625.3736.4781.835.64262.8691.145.64219.6245.785.642
SketchyNetV12.2236.11 (↓0.36)79.63 (↓2.20)0.83362.23 (↓0.63)88.95 (↓2.19)0.83319.48 (↓0.12)44.02 (↓1.76)0.833
SketchyNetV22.2735.88 (↓0.59)78.46 (↓3.37)0.25161.95 (↓0.91)87.68 (↓3.46)0.28119.10 (↓0.50)43.57 (↓2.21)0.304
Partial-OT [14]32x3222.1
64x6422.1
128x12822.1
256x25622.1
SketchyNetV12.22
SketchyNetV22.27
B-BFR224x2244.6124.3261.232.60243.3976.332.60210.4724.342.6022.1511.612.602
B-DRSDynamic9.3625.1268.845.49344.7878.146.18711.8726.586.2453.8617.636.291
B-CropDynamic4.9220.3252.126.56235.0757.686.59708.2720.196.6291.9809.676.688
B-Regress [48]Dynamic2.2722.8261.330.26138.4264.650.29108.6921.980.3232.0710.150.375
B-AKD [76]Dynamic2.2723.6764.260.26838.6171.880.29512.6727.710.3312.9114.010.392
B-PKT [45]Dynamic2.2724.3165.910.25942.6873.660.28812.2728.310.3173.4516.220.372
B-VGG16-SN256x25614.7133.0378.5140.1852.1684.0840.1818.6141.2540.185.1623.6240.18
B-ThiNet [37]256x2568.327.51 (↓25.52)22.34 (↓56.17)9.34212.31 (↓39.85)26.63 (↓57.45)9.3423.16 (↓15.45)12.73 (↓28.52)9.3420.65 (↓4.51)5.35 (↓18.27)9.342
B-Prune [38]256x2566.173.92 (↓29.11)17.32 (↓61.19)8.1387.98 (↓44.18)19.82 (↓64.26)8.1381.45 (↓17.16)8.22 (↓33.03)8.1380.53 (↓4.63)4.97 (↓18.65)8.138
B-VGG16-SN SketchyNetV22.2732.77 (↓0.26)78.21 (↓0.3)0.25450.75 (↓1.41)82.21 (↓1.87)0.28316.55 (↓2.06)38.66 (↓2.59)0.3324.69 (↓0.47)21.76 (↓1.86)0.407
", + "image_path": "b32418968c2c50c829b413e5b7f1ec33dde800d4da0502d806f373c369701d92.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 420, + 138, + 433 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 420, + 138, + 433 + ], + "spans": [ + { + "bbox": [ + 55, + 420, + 138, + 433 + ], + "type": "text", + "content": "5.1. Competitors" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 438, + 297, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 438, + 297, + 715 + ], + "spans": [ + { + "bbox": [ + 54, + 438, + 297, + 715 + ], + "type": "text", + "content": "We evaluate against State-of-the-Arts (SoTAs) and a few relevant curated baselines (B-). (i) SoTAs: Unlike Triplet-SN [75], HOLEF-SN [59], Triplet-RL [6] (we report full-sketch evaluation) and StyleVAE [51]; Partial-OT [14] specifically caters to scene-level FG-SBIR. For fairness, we vary input-resolution of sketches during inference (fixed at their original " + }, + { + "bbox": [ + 54, + 438, + 297, + 715 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 54, + 438, + 297, + 715 + ], + "type": "text", + "content": " setup during training). We also adapt them as teachers, to judge our accuracy retainment after model-compression (SketchyNetV1) using full-res sketch and, further optimising canvas-size with canvas selector (SketchyNetV2). (ii) KD baselines: Here we re-purpose existing works in a cross-modal setting, keeping our canvas selector same, with Triplet-SN as the teacher. B-Regress - regresses features from teacher and student for both sketches and images, aligning their dimensions with an extra transformation layer [48], using a " + }, + { + "bbox": [ + 54, + 438, + 297, + 715 + ], + "type": "inline_equation", + "content": "l_{2}" + }, + { + "bbox": [ + 54, + 438, + 297, + 715 + ], + "type": "text", + "content": " regression loss. Following [45] B-PKT calculates the conditional probability density for any pair of points in the teacher's embedding space [45], which models the probability of any two samples being close together. Sampling 'N' instances, it establishes a probability distribution for pairwise interactions within that space. Simultaneously, in the student's embedding space, a comparable distribution is derived, and the model minimises the divergence between them using a KL-divergence loss. B-AKD follows [76] in using spatial attention maps for knowledge transfer using similar dimensional matching like B-Regress. (iii) Resizing strategies: Keeping the same KD-paradigm with Triplet-SN-teacher, we vary the input" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 421, + 556, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 421, + 556, + 597 + ], + "spans": [ + { + "bbox": [ + 313, + 421, + 556, + 597 + ], + "type": "text", + "content": "size resizing module following existing works. " + }, + { + "bbox": [ + 313, + 421, + 556, + 597 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 313, + 421, + 556, + 597 + ], + "type": "text", + "content": "-BFR follows [62] in designing a bilinear feature resizer amidst the intermediate layers of a CNN network, to output a down-scaled image. We choose its resized resolution of " + }, + { + "bbox": [ + 313, + 421, + 556, + 597 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 313, + 421, + 556, + 597 + ], + "type": "text", + "content": " (lowest) for subsequent evaluation. " + }, + { + "bbox": [ + 313, + 421, + 556, + 597 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 313, + 421, + 556, + 597 + ], + "type": "text", + "content": "-DRS follows [12] in training a convolutional module that takes an image in full-resolution, calculates a probability vector on available canvas-sizes, and transforms it into binary decisions, indicating the scale factor of selection. " + }, + { + "bbox": [ + 313, + 421, + 556, + 597 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 313, + 421, + 556, + 597 + ], + "type": "text", + "content": "-Crop follows [32] in meta-learning an image-cropping module that crops the input image to a target resolution for low compute evaluation. (iv) Pruning alternatives: Here we adapt two VGG16-backboned methods off-the-shelf from network-pruning family among other efficiency paradigms [18, 21, 62], in FG-SBIR setup - " + }, + { + "bbox": [ + 313, + 421, + 556, + 597 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 313, + 421, + 556, + 597 + ], + "type": "text", + "content": "-ThiNet [37] and " + }, + { + "bbox": [ + 313, + 421, + 556, + 597 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 313, + 421, + 556, + 597 + ], + "type": "text", + "content": "-Prune [38]. For comparative reference we report performance of a VGG16-backboned Triplet-SN network (" + }, + { + "bbox": [ + 313, + 421, + 556, + 597 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 313, + 421, + 556, + 597 + ], + "type": "text", + "content": "-VGG16-SN) and its SketchyNetV2 variant." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 600, + 509, + 613 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 600, + 509, + 613 + ], + "spans": [ + { + "bbox": [ + 313, + 600, + 509, + 613 + ], + "type": "text", + "content": "5.2. Performance and Compute Analysis:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 614, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 614, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 614, + 556, + 715 + ], + "type": "text", + "content": "Analysis on Accuracy: Table 2 reports the quantitative evaluation of our model against others. Triplet-SN [75] and HOLEF-SN [59] perform lower owing to weak backbones of Sketch-A-Net [74]. Enhanced by its RL-optimised reward function Triplet-RL[6] scores higher " + }, + { + "bbox": [ + 313, + 614, + 556, + 715 + ], + "type": "inline_equation", + "content": "(2.36\\% \\uparrow" + }, + { + "bbox": [ + 313, + 614, + 556, + 715 + ], + "type": "text", + "content": " Top1 on ShoeV2) but fails to surpass StyleVAE [51], due to its delicate meta-learning disentanglement module addressing style-diversity. Being trained on region-wise sketch-photo associativity, Partial-OT [14] outperforms all others significantly on scene-level sketches (FS-COCO). We also notice a" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "28388" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 216 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 216 + ], + "type": "text", + "content": "gradual drop in accuracy of SoTAs, when subjected to decreasing resolution of input-sketches suggesting that some methods are inherently accustomed to handling scale-shifts better than others, but are inferior to our dynamic approach (SketchyNetV2). Importantly, when most SoTAs are employed as teachers, their respective students (SketchyNetV1 variants) are seen to retain their respective original accuracy with a marginal drop of " + }, + { + "bbox": [ + 55, + 72, + 296, + 216 + ], + "type": "inline_equation", + "content": "\\approx 0.5 - 1\\%" + }, + { + "bbox": [ + 55, + 72, + 296, + 216 + ], + "type": "text", + "content": " overall, thanks to our efficient KD paradigm. Having different teacher-student embedding spaces both " + }, + { + "bbox": [ + 55, + 72, + 296, + 216 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 55, + 72, + 296, + 216 + ], + "type": "text", + "content": "-Regress and " + }, + { + "bbox": [ + 55, + 72, + 296, + 216 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 55, + 72, + 296, + 216 + ], + "type": "text", + "content": "-AKD's paradigm of directly regressing on features proves incompatible to a cross-modal retrieval setting, dropping accuracy (compared to T-Triplet-SN), whereas lacking cross-modal discrimination " + }, + { + "bbox": [ + 55, + 72, + 296, + 216 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 55, + 72, + 296, + 216 + ], + "type": "text", + "content": "-PKT performs lower than ours, proving our superiority in retaining accuracy." + } + ] + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 57, + 220, + 116, + 257 + ], + "blocks": [ + { + "bbox": [ + 57, + 220, + 116, + 257 + ], + "lines": [ + { + "bbox": [ + 57, + 220, + 116, + 257 + ], + "spans": [ + { + "bbox": [ + 57, + 220, + 116, + 257 + ], + "type": "image", + "image_path": "de3a0d5bb52e94f4070e12b04a21b1cf472229379a14b6f2fd172071e21ff8bf.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 72, + 261, + 279, + 272 + ], + "lines": [ + { + "bbox": [ + 72, + 261, + 279, + 272 + ], + "spans": [ + { + "bbox": [ + 72, + 261, + 279, + 272 + ], + "type": "text", + "content": "Figure 4. Exemplar sketches at their optimal canvas-size" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 119, + 220, + 176, + 257 + ], + "blocks": [ + { + "bbox": [ + 119, + 220, + 176, + 257 + ], + "lines": [ + { + "bbox": [ + 119, + 220, + 176, + 257 + ], + "spans": [ + { + "bbox": [ + 119, + 220, + 176, + 257 + ], + "type": "image", + "image_path": "e362480feffaaa2a25a4bf1ad00113228371c443db89798b3bd7b2864ad24929.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 177, + 220, + 237, + 257 + ], + "blocks": [ + { + "bbox": [ + 177, + 220, + 237, + 257 + ], + "lines": [ + { + "bbox": [ + 177, + 220, + 237, + 257 + ], + "spans": [ + { + "bbox": [ + 177, + 220, + 237, + 257 + ], + "type": "image", + "image_path": "214d331dd4b3bbdc3eff100cacaf294036f74e397825e96a10e207f7b1fee5f4.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 238, + 220, + 294, + 257 + ], + "blocks": [ + { + "bbox": [ + 238, + 220, + 294, + 257 + ], + "lines": [ + { + "bbox": [ + 238, + 220, + 294, + 257 + ], + "spans": [ + { + "bbox": [ + 238, + 220, + 294, + 257 + ], + "type": "image", + "image_path": "ca18a9d5f00a783e7d294e3e47bdec3dd5eb72a362115abfe1899a23cb846b98.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 277, + 296, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 277, + 296, + 651 + ], + "spans": [ + { + "bbox": [ + 55, + 277, + 296, + 651 + ], + "type": "text", + "content": "Analysis on Compute: Although FLOPs [12] decrease (Table 2) with a reduction in input resolution across all SoTAs, it comes at a severe cost of accuracy (" + }, + { + "bbox": [ + 55, + 277, + 296, + 651 + ], + "type": "inline_equation", + "content": "\\approx 21\\%" + }, + { + "bbox": [ + 55, + 277, + 296, + 651 + ], + "type": "text", + "content": " average drop from " + }, + { + "bbox": [ + 55, + 277, + 296, + 651 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 55, + 277, + 296, + 651 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 277, + 296, + 651 + ], + "type": "inline_equation", + "content": "32 \\times 32" + }, + { + "bbox": [ + 55, + 277, + 296, + 651 + ], + "type": "text", + "content": " on ShoeV2). This is mainly because ideal canvas-size for optimal retrieval varies across samples (Fig. 2 right) - thus subjecting all samples to a fixed canvas-size proves detrimental for accuracy. However, a positive Acc@1 score of " + }, + { + "bbox": [ + 55, + 277, + 296, + 651 + ], + "type": "inline_equation", + "content": "\\approx 11\\%" + }, + { + "bbox": [ + 55, + 277, + 296, + 651 + ], + "type": "text", + "content": " even at a low canvas-size of " + }, + { + "bbox": [ + 55, + 277, + 296, + 651 + ], + "type": "inline_equation", + "content": "32 \\times 32" + }, + { + "bbox": [ + 55, + 277, + 296, + 651 + ], + "type": "text", + "content": " for every SoTA shows that some sketches are indeed recognisable if the model is trained properly. This is verified further by SketchyNetV2 variant of respective SoTAs, that secures the lowest FLOPs per SoTA, despite maintaining accuracy at par with its SoTA-counterpart - thanks to our RL-guided canvas selector, which can allot an optimal canvas-size per sketch towards better retrieval (Fig. 4). Contrarily, B-BFR fails to lower FLOPs as it fixes a " + }, + { + "bbox": [ + 55, + 277, + 296, + 651 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 55, + 277, + 296, + 651 + ], + "type": "text", + "content": " input resolution for a satisfactory Acc@1 compared to Triplet-SN. B-Crop fairs better, however sketch being sparse, cropped sketches leads to confused representations, lowering accuracy. Despite coming closest in accuracy to ours on Triplet-SN, B-DRS lags behind considerably on compute, thanks to our canvas selector working in vector space (inputs " + }, + { + "bbox": [ + 55, + 277, + 296, + 651 + ], + "type": "inline_equation", + "content": "s_v" + }, + { + "bbox": [ + 55, + 277, + 296, + 651 + ], + "type": "text", + "content": ") and eliminating costly convolutional operations, unlike others using full-resolution sketch-raster as input. Being untrained to handle sparse-nature of sketches during such high compression, performance of both B-ThiNet and B-Prune collapses (vs. B-VGG16-SN) unlike its SketchyNetV2. Notably, the significant drop in parameter-count (4 - 10 times) against SoTAs, comes from choosing lightweight MobileNetV2 [54] as a student coupled with our canvas-size selector in every SketchyNetV2 variant. Compared to ShoeV2, FLOPs increase slightly for ChairV2, Sketchy and FS-COCO, owing to higher object (ChairV2) or scene-level (FS-COCO) details, and lower structural correspondence with photos (Sketchy), thus training a poorer retrieval model which in turn acts as a poor critic for our canvas selector. Our canvas selector alone takes a negligible 51.84K parameters with 0.02G FLOPs." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 655, + 149, + 667 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 655, + 149, + 667 + ], + "spans": [ + { + "bbox": [ + 55, + 655, + 149, + 667 + ], + "type": "text", + "content": "5.3. Ablation Study" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 669, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 669, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 669, + 296, + 714 + ], + "type": "text", + "content": "As our teacher " + }, + { + "bbox": [ + 55, + 669, + 296, + 714 + ], + "type": "inline_equation", + "content": "(F_T)" + }, + { + "bbox": [ + 55, + 669, + 296, + 714 + ], + "type": "text", + "content": " can be any pre-trained FG-SBIR network, we limit our ablation to justifying design choices of our student model and canvas-size selector. We use " + }, + { + "bbox": [ + 55, + 669, + 296, + 714 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 55, + 669, + 296, + 714 + ], + "type": "text", + "content": "-VGG16-SN as a reference " + }, + { + "bbox": [ + 55, + 669, + 296, + 714 + ], + "type": "inline_equation", + "content": "F_T" + }, + { + "bbox": [ + 55, + 669, + 296, + 714 + ], + "type": "text", + "content": " - Type-" + }, + { + "bbox": [ + 55, + 669, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\emptyset" + }, + { + "bbox": [ + 55, + 669, + 296, + 714 + ], + "type": "text", + "content": " in Table 3." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 73, + 555, + 161 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 73, + 555, + 161 + ], + "spans": [ + { + "bbox": [ + 313, + 73, + 555, + 161 + ], + "type": "text", + "content": "Training Canvas-Size selector " + }, + { + "bbox": [ + 313, + 73, + 555, + 161 + ], + "type": "inline_equation", + "content": "\\psi_C" + }, + { + "bbox": [ + 313, + 73, + 555, + 161 + ], + "type": "text", + "content": ": To justify loss objectives of " + }, + { + "bbox": [ + 313, + 73, + 555, + 161 + ], + "type": "inline_equation", + "content": "\\psi_C" + }, + { + "bbox": [ + 313, + 73, + 555, + 161 + ], + "type": "text", + "content": " we evaluate on ShoeV2, eliminating " + }, + { + "bbox": [ + 313, + 73, + 555, + 161 + ], + "type": "inline_equation", + "content": "r^{-1}" + }, + { + "bbox": [ + 313, + 73, + 555, + 161 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 73, + 555, + 161 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{Tri}}^R" + }, + { + "bbox": [ + 313, + 73, + 555, + 161 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 73, + 555, + 161 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{F}}" + }, + { + "bbox": [ + 313, + 73, + 555, + 161 + ], + "type": "text", + "content": " one by one from Eq. 6 (Table 3). While eliminating accuracy rewards, " + }, + { + "bbox": [ + 313, + 73, + 555, + 161 + ], + "type": "inline_equation", + "content": "(w / o \\text{ rank}^{-1}, w / o \\mathcal{L}_{\\mathrm{Tri}}^R)" + }, + { + "bbox": [ + 313, + 73, + 555, + 161 + ], + "type": "text", + "content": " increases constraint on lowering FLOPs, dropping accuracy, ignoring " + }, + { + "bbox": [ + 313, + 73, + 555, + 161 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{F}}" + }, + { + "bbox": [ + 313, + 73, + 555, + 161 + ], + "type": "inline_equation", + "content": "(w / o \\mathcal{L}_F)" + }, + { + "bbox": [ + 313, + 73, + 555, + 161 + ], + "type": "text", + "content": " gains accuracy but costs high FLOPs. We thus optimally balance between the two using " + }, + { + "bbox": [ + 313, + 73, + 555, + 161 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{F}}" + }, + { + "bbox": [ + 313, + 73, + 555, + 161 + ], + "type": "text", + "content": " (Fig. 5 left). One can also choose to train for faster retrieval (lower FLOPs), given a decent Acc@5 (Fig. 5 right)." + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 327, + 165, + 422, + 215 + ], + "blocks": [ + { + "bbox": [ + 327, + 165, + 422, + 215 + ], + "lines": [ + { + "bbox": [ + 327, + 165, + 422, + 215 + ], + "spans": [ + { + "bbox": [ + 327, + 165, + 422, + 215 + ], + "type": "image", + "image_path": "e2f7d3047db873640356a1426d38a7e35719439634bff95966f7bd695227d686.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 334, + 218, + 534, + 229 + ], + "lines": [ + { + "bbox": [ + 334, + 218, + 534, + 229 + ], + "spans": [ + { + "bbox": [ + 334, + 218, + 534, + 229 + ], + "type": "text", + "content": "Figure 5. Impact of varying " + }, + { + "bbox": [ + 334, + 218, + 534, + 229 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{F}}" + }, + { + "bbox": [ + 334, + 218, + 534, + 229 + ], + "type": "text", + "content": " on Acc@1 and Acc@5" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 427, + 165, + 526, + 215 + ], + "blocks": [ + { + "bbox": [ + 427, + 165, + 526, + 215 + ], + "lines": [ + { + "bbox": [ + 427, + 165, + 526, + 215 + ], + "spans": [ + { + "bbox": [ + 427, + 165, + 526, + 215 + ], + "type": "image", + "image_path": "799f80fd904eaa953d0823a729f55c7ac735d4637db5012301974a08ddeba2b7.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 231, + 555, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 231, + 555, + 418 + ], + "spans": [ + { + "bbox": [ + 313, + 231, + 555, + 418 + ], + "type": "text", + "content": "Architectural insights: To judge the efficiency of our student's backbone, we explore a few cheaper backbone-networks (Types IV, V in Table 3). Evidently, MobileNetV2 (Ours) balances optimally between compute and accuracy. As an option to vector input for canvas selector, we re-design it with a simple seven-layer CNN module following [78] to use the full-resolution sketch-image as its input, and select a target canvas-size to downscale it for retrieval. With extra computational cost coming from spatial domain, total FLOPs incurred is much higher (VI). Next we compare efficacy of sketch as input to " + }, + { + "bbox": [ + 313, + 231, + 555, + 418 + ], + "type": "inline_equation", + "content": "\\psi_C" + }, + { + "bbox": [ + 313, + 231, + 555, + 418 + ], + "type": "text", + "content": " with respect to types of vector format - offset-coordinate (VI) [22] vs. absolute ones (Ours). Turns out the former is better for encoding. Comparing sketch encoders among LSTM (VIII), GRU (Ours) and Transformer (IX), showed ours to be optimum empirically (Table 3). Furthermore, preponding recent pre-processing modules like stroke-subset selector of [9] to our canvas selector, yields " + }, + { + "bbox": [ + 313, + 231, + 555, + 418 + ], + "type": "inline_equation", + "content": "42.95\\%" + }, + { + "bbox": [ + 313, + 231, + 555, + 418 + ], + "type": "text", + "content": " (vs. " + }, + { + "bbox": [ + 313, + 231, + 555, + 418 + ], + "type": "inline_equation", + "content": "43.7\\%" + }, + { + "bbox": [ + 313, + 231, + 555, + 418 + ], + "type": "text", + "content": " [9]) Acc@1 ShoeV2, proving its off-the-shelf compatibility." + } + ] + } + ], + "index": 13 + }, + { + "type": "table", + "bbox": [ + 321, + 434, + 543, + 555 + ], + "blocks": [ + { + "bbox": [ + 347, + 423, + 522, + 433 + ], + "lines": [ + { + "bbox": [ + 347, + 423, + 522, + 433 + ], + "spans": [ + { + "bbox": [ + 347, + 423, + 522, + 433 + ], + "type": "text", + "content": "Table 3. Ablative studies (accuracy on ShoeV2)" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 321, + 434, + 543, + 555 + ], + "lines": [ + { + "bbox": [ + 321, + 434, + 543, + 555 + ], + "spans": [ + { + "bbox": [ + 321, + 434, + 543, + 555 + ], + "type": "table", + "html": "
TypeFLOPsParamsShoeV2
(G)(mil.)Top1 (%)Top10 (%)
∅B-VGG16-SN40.1814.7133.0378.51
Iw/o rank-1-reward0.17325.3164.36
IIw/o \\( \\mathcal{L}_{\\text{Tri}}^R \\)-reward0.1822.2727.7869.32
IIIw/o \\( \\mathcal{L}_{\\text{F}} \\)-reward0.83332.9178.39
IVResNet18 backbone1.45111.1826.7268.18
VEfficientNet backbone0.4594.0129.4772.04
VIImage-based \\( \\psi_C \\)5.7307.6232.2177.68
VIIOffset \\( s_v \\)0.2552.2732.6178.40
VIIIDecoder-LSTM0.2682.2531.4177.32
IXDecoder-Tf0.3142.3431.9878.02
SketchyNetV20.2542.2732.7778.21
", + "image_path": "bf3cb53392aa2e184dbf301fcdade27b3b4eca10ccba4c189d6a61fb736c4ffa.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_body" + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 559, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 559, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 559, + 556, + 713 + ], + "type": "text", + "content": "Generalisability of canvas selector: To judge if the optimal canvas-size selected is uniform irrespective of the training method, we first train four canvas-selectors using four SoTAs of Table 2. Evaluating them on ShoeV2 [75], " + }, + { + "bbox": [ + 313, + 559, + 556, + 713 + ], + "type": "inline_equation", + "content": "89.21\\%" + }, + { + "bbox": [ + 313, + 559, + 556, + 713 + ], + "type": "text", + "content": " of test-set sketches in ShoeV2 to obtain the same optimal canvas size across four different models, trained on four different methods, which further proves that our canvas selector can work equally well as a pre-processing module for any new model. Furthermore, plugging a canvas selector trained with SketchyNetV1 (critic) from B-VGG16-SN with a different SketchyNetV1 from StyleVAE [51] during inference achieves comparable performance " + }, + { + "bbox": [ + 313, + 559, + 556, + 713 + ], + "type": "inline_equation", + "content": "(35.69\\%)" + }, + { + "bbox": [ + 313, + 559, + 556, + 713 + ], + "type": "text", + "content": " Top-1) to a canvas selector trained on StyleVAE-SketchyNetV1 " + }, + { + "bbox": [ + 313, + 559, + 556, + 713 + ], + "type": "inline_equation", + "content": "(35.88\\%)" + }, + { + "bbox": [ + 313, + 559, + 556, + 713 + ], + "type": "text", + "content": " Top-1) on ShoeV2. Importantly, " + }, + { + "bbox": [ + 313, + 559, + 556, + 713 + ], + "type": "inline_equation", + "content": "95.64\\%" + }, + { + "bbox": [ + 313, + 559, + 556, + 713 + ], + "type": "text", + "content": " of test-set sketches yield the same optimal canvas-size across both models," + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "28389" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 73, + 294, + 139 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 73, + 294, + 139 + ], + "spans": [ + { + "bbox": [ + 55, + 73, + 294, + 139 + ], + "type": "text", + "content": "confirming the cross-model generalisability of our canvas selector. (ii) On training canvas selector with teacher " + }, + { + "bbox": [ + 55, + 73, + 294, + 139 + ], + "type": "inline_equation", + "content": "F_{T}" + }, + { + "bbox": [ + 55, + 73, + 294, + 139 + ], + "type": "text", + "content": " (for Triplet-SN) as critic, yields same optimal canvas-sizes as those for student " + }, + { + "bbox": [ + 55, + 73, + 294, + 139 + ], + "type": "inline_equation", + "content": "F_{S}" + }, + { + "bbox": [ + 55, + 73, + 294, + 139 + ], + "type": "text", + "content": ", for " + }, + { + "bbox": [ + 55, + 73, + 294, + 139 + ], + "type": "inline_equation", + "content": "94.21\\%" + }, + { + "bbox": [ + 55, + 73, + 294, + 139 + ], + "type": "text", + "content": " of test-set sketches in ShoeV2, while incurring comparable FLOPs reduction (0.268G) and accuracy (27.91% Top1) against " + }, + { + "bbox": [ + 55, + 73, + 294, + 139 + ], + "type": "inline_equation", + "content": "F_{S}" + }, + { + "bbox": [ + 55, + 73, + 294, + 139 + ], + "type": "text", + "content": " as critic." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 139, + 294, + 237 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 139, + 294, + 237 + ], + "spans": [ + { + "bbox": [ + 55, + 139, + 294, + 237 + ], + "type": "text", + "content": "Comparing Inference times: We evaluated on a 12 GB Nvidia RTX 2080 Ti GPU, on ShoeV2, for inference time of Triplet-SN, HOLEF-SN, Triplet-RL, and StyleVAE (SOTAs) on full-sketch to obtain 37, 38, 36 and " + }, + { + "bbox": [ + 55, + 139, + 294, + 237 + ], + "type": "inline_equation", + "content": "42\\mathrm{ms}" + }, + { + "bbox": [ + 55, + 139, + 294, + 237 + ], + "type": "text", + "content": ", against 10, 11, 11, and " + }, + { + "bbox": [ + 55, + 139, + 294, + 237 + ], + "type": "inline_equation", + "content": "12\\mathrm{ms}" + }, + { + "bbox": [ + 55, + 139, + 294, + 237 + ], + "type": "text", + "content": " for their SketchyNetV2 variants respectively. For practical validation, we evaluated onnx-INT8 version of Triplet-SN [75] and its SketchyNetV2 variant (ours) after post-training quantisation, on an iPhone13 via the AI-benchmark app for ShoeV2 [75] to obtain 51ms and 18ms respectively, which shows our efficiency." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 239, + 294, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 239, + 294, + 339 + ], + "spans": [ + { + "bbox": [ + 55, + 239, + 294, + 339 + ], + "type": "text", + "content": "Retraining SoTA on mixed resolution sketches?: To explore if retraining SoTAs on sketches with mixed resolutions, we re-train Triple-SN [75] on rendered mixed-resolution sketches of ShoeV2. Poor performance in Table 4 at lower resolutions similar to Table 2 validates our KD approach. The key lies in that the sketch-feature from teacher used for student-guidance " + }, + { + "bbox": [ + 55, + 239, + 294, + 339 + ], + "type": "inline_equation", + "content": "(f_s^T)" + }, + { + "bbox": [ + 55, + 239, + 294, + 339 + ], + "type": "text", + "content": " is always extracted from full-resolution sketch while that extracted by student is at varying resolutions, imparting scale invariance, which is ideally suited to our cross-modal retrieval setup." + } + ] + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 57, + 357, + 291, + 425 + ], + "blocks": [ + { + "bbox": [ + 58, + 343, + 292, + 354 + ], + "lines": [ + { + "bbox": [ + 58, + 343, + 292, + 354 + ], + "spans": [ + { + "bbox": [ + 58, + 343, + 292, + 354 + ], + "type": "text", + "content": "Table 4. Evaluating SoTAs trained on mixed-resolution sketches" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 357, + 291, + 425 + ], + "lines": [ + { + "bbox": [ + 57, + 357, + 291, + 425 + ], + "spans": [ + { + "bbox": [ + 57, + 357, + 291, + 425 + ], + "type": "table", + "html": "
Canvas SizeTriplet-SNTriplet-RL
Top1 (%)Top10 (%)FLOPs (G)Top1 (%)Top10 (%)FLOPs (G)
32×3210.9126.650.08312.9731.080.142
64×6419.0448.120.33822.9656.230.577
128×12826.0763.871.39730.0472.952.299
256×25628.7471.685.28034.2177.246.041
", + "image_path": "b29958115849f070ffd1b07979377daef4410f50587b4f967db626ffcd86d1bb.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 430, + 155, + 442 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 430, + 155, + 442 + ], + "spans": [ + { + "bbox": [ + 55, + 430, + 155, + 442 + ], + "type": "text", + "content": "5.4. Further Insights" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 445, + 294, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 445, + 294, + 521 + ], + "spans": [ + { + "bbox": [ + 55, + 445, + 294, + 521 + ], + "type": "text", + "content": "Cross-dataset Generalisation: To judge the generalisation potential of our canvas-size selector across datasets we train it on ShoeV2, but plug it to a retrieval model (SketchyNetV1) trained on ChairV2, and evaluate on ChairV2. Against the standard result (Table 2) of training both modules (SketchyNetV2) on ChairV2 (0.285G, " + }, + { + "bbox": [ + 55, + 445, + 294, + 521 + ], + "type": "inline_equation", + "content": "55.32\\%" + }, + { + "bbox": [ + 55, + 445, + 294, + 521 + ], + "type": "text", + "content": " Acc@1), this setup retains accuracy at " + }, + { + "bbox": [ + 55, + 445, + 294, + 521 + ], + "type": "inline_equation", + "content": "53.14\\%" + }, + { + "bbox": [ + 55, + 445, + 294, + 521 + ], + "type": "text", + "content": " with slight rise in FLOPs (0.314G), thus confirming its potential." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 524, + 294, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 524, + 294, + 634 + ], + "spans": [ + { + "bbox": [ + 55, + 524, + 294, + 634 + ], + "type": "text", + "content": "Faster Sketch-recognition: Exploring potential of our canvas-size selector, we prepend it to a standard sketch-recognition pipeline [74] to optimise the canvas-size of query sketch, and reduce overall compute. Unlike our " + }, + { + "bbox": [ + 55, + 524, + 294, + 634 + ], + "type": "inline_equation", + "content": "\\psi_{C}" + }, + { + "bbox": [ + 55, + 524, + 294, + 634 + ], + "type": "text", + "content": ", it uses negative of cross-entropy loss as accuracy reward " + }, + { + "bbox": [ + 55, + 524, + 294, + 634 + ], + "type": "inline_equation", + "content": "(R_{\\mathrm{acc}})" + }, + { + "bbox": [ + 55, + 524, + 294, + 634 + ], + "type": "text", + "content": ", keeping rest same. Compared to existing Sketch-a-Net [74] incurring 5.28G FLOPs at " + }, + { + "bbox": [ + 55, + 524, + 294, + 634 + ], + "type": "inline_equation", + "content": "68.71\\%" + }, + { + "bbox": [ + 55, + 524, + 294, + 634 + ], + "type": "text", + "content": " accuracy on Quickdraw [22], our canvas selector fitted paradigm drops FLOPs to 0.261G while retaining accuracy at " + }, + { + "bbox": [ + 55, + 524, + 294, + 634 + ], + "type": "inline_equation", + "content": "68.14\\%" + }, + { + "bbox": [ + 55, + 524, + 294, + 634 + ], + "type": "text", + "content": ". This shows for the first time, that sparsity of sketches can be handled, in various downstream tasks via our " + }, + { + "bbox": [ + 55, + 524, + 294, + 634 + ], + "type": "inline_equation", + "content": "\\psi_{C}" + }, + { + "bbox": [ + 55, + 524, + 294, + 634 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 636, + 294, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 636, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 636, + 294, + 713 + ], + "type": "text", + "content": "Extension to Category-level SBIR: For categorical SBIR we use Sketchy (ext) [36] dataset (90:10 train:test split), which holds " + }, + { + "bbox": [ + 55, + 636, + 294, + 713 + ], + "type": "inline_equation", + "content": "75k" + }, + { + "bbox": [ + 55, + 636, + 294, + 713 + ], + "type": "text", + "content": " sketches across 125 categories having " + }, + { + "bbox": [ + 55, + 636, + 294, + 713 + ], + "type": "inline_equation", + "content": "\\approx 73k" + }, + { + "bbox": [ + 55, + 636, + 294, + 713 + ], + "type": "text", + "content": " images [36] in total. It is evaluated using mean average precision (mAP@all) and precision at top 200 retrieval (P@200) [36]. Here we compete against state-of-the-arts (SoTA) like " + }, + { + "bbox": [ + 55, + 636, + 294, + 713 + ], + "type": "inline_equation", + "content": "B-D2S" + }, + { + "bbox": [ + 55, + 636, + 294, + 713 + ], + "type": "text", + "content": " that follows [19] without its zero-shot setting), and a simple baseline using a Siamese-style" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 329, + 85, + 538, + 175 + ], + "blocks": [ + { + "bbox": [ + 335, + 71, + 533, + 82 + ], + "lines": [ + { + "bbox": [ + 335, + 71, + 533, + 82 + ], + "spans": [ + { + "bbox": [ + 335, + 71, + 533, + 82 + ], + "type": "text", + "content": "Table 5. Quantitative Evaluation on Categorical SBIR" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 329, + 85, + 538, + 175 + ], + "lines": [ + { + "bbox": [ + 329, + 85, + 538, + 175 + ], + "spans": [ + { + "bbox": [ + 329, + 85, + 538, + 175 + ], + "type": "table", + "html": "
MethodsFLOPsParamsSketchy(ext)
(G)(mil.)mAP@allP@200
SoTAB-SBIR-SN40.1814.710.7150.861
DSH [36]22.1416.320.7110.858
B-D2S [19]17.75136.360.8100.894
StyleVAE [51]8.11625.370.9050.927
T-SoTAT-SBIR-SN0.2540.7020.810
T-DSH0.2892.270.7040.823
T-D2S0.2680.7980.845
T-StyleVAE0.2510.8860.897
", + "image_path": "717d0aa723a2a65890d11c8c2d6fb7b939b4b1734a0b79265d7c07635a7164a0.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 177, + 553, + 221 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 177, + 553, + 221 + ], + "spans": [ + { + "bbox": [ + 313, + 177, + 553, + 221 + ], + "type": "text", + "content": "VGG-16 network (B-SBIR-SN) following [75]. Similar to Section 5.1 adapting SoTAs as teachers in our SketchyNetV2 paradigm (T-SoTA) shows them retaining their accuracy (Table 5) while significantly lowering compute." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 223, + 553, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 223, + 553, + 266 + ], + "spans": [ + { + "bbox": [ + 313, + 223, + 553, + 266 + ], + "type": "text", + "content": "Teacher-Student Pairs: To explore other teacher-student pairs apart from ours, we test SketchyNetV2 models of a few " + }, + { + "bbox": [ + 313, + 223, + 553, + 266 + ], + "type": "inline_equation", + "content": "(F_T,F_S)" + }, + { + "bbox": [ + 313, + 223, + 553, + 266 + ], + "type": "text", + "content": " pairs, where teachers are trained similar to B-VGG16-SN. From Table 6 we see our pair to outperform others." + } + ] + } + ], + "index": 12 + }, + { + "type": "table", + "bbox": [ + 315, + 283, + 551, + 334 + ], + "blocks": [ + { + "bbox": [ + 351, + 271, + 517, + 281 + ], + "lines": [ + { + "bbox": [ + 351, + 271, + 517, + 281 + ], + "spans": [ + { + "bbox": [ + 351, + 271, + 517, + 281 + ], + "type": "text", + "content": "Table 6. Evaluation on Teacher-Student Pairs" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 283, + 551, + 334 + ], + "lines": [ + { + "bbox": [ + 315, + 283, + 551, + 334 + ], + "spans": [ + { + "bbox": [ + 315, + 283, + 551, + 334 + ], + "type": "table", + "html": "
Teacher\\StudentMobileNetV2EfficientNetResNet-18
Top-1FLOPsTop-1FLOPsTop-1FLOPs
VGG-16 [30]32.77%0.254G29.47%0.459G26.72%1.451G
Inception-V3 [74]27.89%0.264G27.16%0.461G25.31%1.452G
", + "image_path": "bd6acc95657a46b7b7a46da8adaa50ef7ce4781950b971dac37a1349e3cb452b.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 341, + 553, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 341, + 553, + 462 + ], + "spans": [ + { + "bbox": [ + 313, + 341, + 553, + 462 + ], + "type": "text", + "content": "Combining compression methods: To explore if a combination of compression methods is capable of increasing efficiency, we subject our Knowledge Distillation (KD) trained model (SketchyNetV1) of Triplet-SN [75] (Table 2) to (i) quantisation following [21], and (ii) network-pruning following [38], separately, and then train their respective canvas-selectors. While the former secures " + }, + { + "bbox": [ + 313, + 341, + 553, + 462 + ], + "type": "inline_equation", + "content": "27.34\\%" + }, + { + "bbox": [ + 313, + 341, + 553, + 462 + ], + "type": "text", + "content": " (vs. " + }, + { + "bbox": [ + 313, + 341, + 553, + 462 + ], + "type": "inline_equation", + "content": "27.89\\%" + }, + { + "bbox": [ + 313, + 341, + 553, + 462 + ], + "type": "text", + "content": " ours) Top-1 score with a lower inference time of " + }, + { + "bbox": [ + 313, + 341, + 553, + 462 + ], + "type": "inline_equation", + "content": "9.1\\mathrm{ms}" + }, + { + "bbox": [ + 313, + 341, + 553, + 462 + ], + "type": "text", + "content": " (vs. " + }, + { + "bbox": [ + 313, + 341, + 553, + 462 + ], + "type": "inline_equation", + "content": "10\\mathrm{ms}" + }, + { + "bbox": [ + 313, + 341, + 553, + 462 + ], + "type": "text", + "content": " ours), the latter scores " + }, + { + "bbox": [ + 313, + 341, + 553, + 462 + ], + "type": "inline_equation", + "content": "26.91\\%" + }, + { + "bbox": [ + 313, + 341, + 553, + 462 + ], + "type": "text", + "content": " (vs. " + }, + { + "bbox": [ + 313, + 341, + 553, + 462 + ], + "type": "inline_equation", + "content": "27.89\\%" + }, + { + "bbox": [ + 313, + 341, + 553, + 462 + ], + "type": "text", + "content": " ms ours) Top-1 with lower FLOPs of 0.249G (vs. 0.264G ours). Such competitive results further endorse exploring combining compression methods as future works." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 464, + 554, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 464, + 554, + 563 + ], + "spans": [ + { + "bbox": [ + 313, + 464, + 554, + 563 + ], + "type": "text", + "content": "Future Works: Sketch being an abstraction of an object's photo, when resized to a lower scale for retrieval, infers that its object's semantic is interpretable at that scale. This should ideally provide a lower bound for resizing a photo while retaining interpretability (e.g. high retrieval score). Consequently, designing a learnable photo-resizer aided by our trained sketch-canvas selector is a potential future work. Secondly, ours being a meta-framework, using more recent frameworks like DINOv2 [41] as large-scale teacher models for further gains is another targeted future work." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 567, + 388, + 579 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 567, + 388, + 579 + ], + "spans": [ + { + "bbox": [ + 313, + 567, + 388, + 579 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 582, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 554, + 713 + ], + "type": "text", + "content": "We for the first time investigate the problem of efficient inference for FG-SBIR, where we show that using existing efficient lightweight networks directly, are unfit for sketches. We thus propose two generic components that can be adapted to photo-networks for sketch-specific data. Firstly, a cross-modal knowledge distillation paradigm distills the knowledge of larger models to a smaller one, directly reducing FLOPs (params) by 97.92 " + }, + { + "bbox": [ + 313, + 582, + 554, + 713 + ], + "type": "inline_equation", + "content": "(84.9)\\%" + }, + { + "bbox": [ + 313, + 582, + 554, + 713 + ], + "type": "text", + "content": ". Secondly, an abstraction-aware canvas-size selector dynamically selects the best canvas-size for a sketch, reducing FLOPs by extra two-thirds. Extensive experiments show our method to perform at par with state-of-the-arts using much lower compute thus further conveying our method's worth in enhancing them towards deployment." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28390" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 91, + 295, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 295, + 123 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 295, + 123 + ], + "type": "text", + "content": "[1] Manoj Alwani, Yang Wang, and Vashisht Madhavan. Decore: Deep compression with reinforcement learning. In CVPR, 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 125, + 295, + 178 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 125, + 295, + 178 + ], + "spans": [ + { + "bbox": [ + 61, + 125, + 295, + 178 + ], + "type": "text", + "content": "[2] Aneeshan Sain and Ayan Kumar Bhunia and Pinaki Nath Chowdhury and Aneeshan Sain and Subhadeep Koley and Tao Xiang and Yi-Zhe Song. CLIP for All Things Zero-Shot Sketch-Based Image Retrieval, Fine-Grained or Not. In CVPR, 2023. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 180, + 295, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 180, + 295, + 201 + ], + "spans": [ + { + "bbox": [ + 62, + 180, + 295, + 201 + ], + "type": "text", + "content": "[3] Jimmy Ba and Rich Caruana. Do deep nets really need to be deep? NeurIPS, 2014. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 202, + 295, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 202, + 295, + 246 + ], + "spans": [ + { + "bbox": [ + 62, + 202, + 295, + 246 + ], + "type": "text", + "content": "[4] Hessam Bagherinezhad, Maxwell Horton, Mohammad Rastegari, and Ali Farhadi. Label refinery: Improving image-net classification through label progression. arXiv preprint arXiv:1805.02641, 2018. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 247, + 295, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 247, + 295, + 300 + ], + "spans": [ + { + "bbox": [ + 62, + 247, + 295, + 300 + ], + "type": "text", + "content": "[5] Ayan Kumar Bhunia, Ayan Das, Umar Riaz Muhammad, Yongxin Yang, Timothy M Hospedales, Tao Xiang, Yulia Gryaditskaya, and Yi-Zhe Song. Pixel: A competitive sketching ai agent. so you think you can sketch? ACM TOG, 2020. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 303, + 295, + 345 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 303, + 295, + 345 + ], + "spans": [ + { + "bbox": [ + 62, + 303, + 295, + 345 + ], + "type": "text", + "content": "[6] Ayan Kumar Bhunia, Yongxin Yang, Timothy M Hospedales, Tao Xiang, and Yi-Zhe Song. Sketch less for more: On-the-fly fine-grained sketch based image retrieval. In CVPR, 2020. 1, 2, 3, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 347, + 295, + 400 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 347, + 295, + 400 + ], + "spans": [ + { + "bbox": [ + 62, + 347, + 295, + 400 + ], + "type": "text", + "content": "[7] Ayan Kumar Bhunia, Pinaki Nath Chowdhury, Aaneshan Sain, Yongxin Yang, Tao Xiang, and Yi-Zhe Song. More photos are all you need: Semi-supervised learning for fine-grained sketch based image retrieval. In CVPR, 2021. 1, 2, 3, 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 403, + 295, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 403, + 295, + 446 + ], + "spans": [ + { + "bbox": [ + 62, + 403, + 295, + 446 + ], + "type": "text", + "content": "[8] Ayan Kumar Bhunia, Pinaki Nath Chowdhury, Yongxin Yang, Timothy M Hospedales, Tao Xiang, and Yi-Zhe Song. Vectorization and rasterization: Self-supervised learning for sketch and handwriting. In CVPR, 2021. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 447, + 295, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 447, + 295, + 501 + ], + "spans": [ + { + "bbox": [ + 62, + 447, + 295, + 501 + ], + "type": "text", + "content": "[9] Ayan Kumar Bhunia, Subhadeep Koley, Abdullah Faiz Ur Rahman Khilji, Aneeshan Sain, Pinaki Nath Chowdhury, Tao Xiang, and Yi-Zhe Song. Sketching without worrying: Noise-tolerant sketch-based image retrieval. In CVPR, 2022. 1, 2, 3, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 502, + 295, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 502, + 295, + 546 + ], + "spans": [ + { + "bbox": [ + 57, + 502, + 295, + 546 + ], + "type": "text", + "content": "[10] Ayan Kumar Bhunia, Subhadeep Koley, Amandeep Kumar, Aneeshan Sain, Pinaki Nath Chowdhury, Tao Xiang, and Yi-Zhe Song. Sketch2Saliency: Learning to Detect Salient Objects from Human Drawings. In CVPR, 2023. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 547, + 295, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 547, + 295, + 579 + ], + "spans": [ + { + "bbox": [ + 57, + 547, + 295, + 579 + ], + "type": "text", + "content": "[11] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In ICML, 2020. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 581, + 295, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 581, + 295, + 613 + ], + "spans": [ + { + "bbox": [ + 57, + 581, + 295, + 613 + ], + "type": "text", + "content": "[12] Ying Chen, Liang Qiao, Zhanzhan Cheng, Shiliang Pu, Yi Niu, and Xi Li. Dynamic low-resolution distillation for cost-efficient end-to-end text spotting. In ECCV, 2022. 5, 6, 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 614, + 295, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 614, + 295, + 658 + ], + "spans": [ + { + "bbox": [ + 57, + 614, + 295, + 658 + ], + "type": "text", + "content": "[13] Pinaki Nath Chowdhury, Ayan Kumar Bhunia, Viswanatha Reddy Gajjala, Aneeshan Sain, Tao Xiang, and Yi-Zhe Song. Partially does it: Towards scene-level fg-sbir with partial input. In CVPR, 2022. 1, 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 658, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 658, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 658, + 295, + 713 + ], + "type": "text", + "content": "[14] Pinaki Nath Chowdhury, Ayan Kumar Bhunia, Viswanatha Reddy Gajjala, Aneeshan Sain, Tao Xiang, and Yi-Zhe Song. Partially does it: Towards scene-level fg-sbir with partial input. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 6" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 35, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "text", + "content": "[15] Pinaki Nath Chowdhury, Aneeshan Sain, Ayan Kumar Bhunia, Tao Xiang, Yulia Gryaditskaya, and Yi-Zhe Song. Fscoco: Towards understanding of freehand sketches of common objects in context. In ECCV, 2022. 5, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 118, + 553, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 118, + 553, + 162 + ], + "spans": [ + { + "bbox": [ + 317, + 118, + 553, + 162 + ], + "type": "text", + "content": "[16] Chowdhury, Pinaki Nath and Bhunia, Ayan Kumar and Sain, Aneeshan and Koley, Subhadeep and Xiang, Tao and Song, Yi-Zhe. Democratising 2D Sketch to 3D Shape Retrieval Through Pivoting. In ICCV, 2023. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 164, + 553, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 164, + 553, + 195 + ], + "spans": [ + { + "bbox": [ + 317, + 164, + 553, + 195 + ], + "type": "text", + "content": "[17] John Collomosse, Tu Bui, and Hailin Jin. Livesketch: Query perturbations for guided sketch-based visual search. In CVPR, 2019. 2, 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 198, + 553, + 231 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 198, + 553, + 231 + ], + "spans": [ + { + "bbox": [ + 317, + 198, + 553, + 231 + ], + "type": "text", + "content": "[18] Matthieu Courbariaux, Yoshua Bengio, and Jean-Pierre David. Binaryconnect: Training deep neural networks with binary weights during propagations. NeurIPS, 28, 2015. 2, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 232, + 553, + 265 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 232, + 553, + 265 + ], + "spans": [ + { + "bbox": [ + 317, + 232, + 553, + 265 + ], + "type": "text", + "content": "[19] Sounak Dey, Pau Riba, Anjan Dutta, Josep Llados, and Yi-Zhe Song. Doodle to search: Practical zero-shot sketch-based image retrieval. In CVPR, 2019. 2, 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 266, + 553, + 299 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 266, + 553, + 299 + ], + "spans": [ + { + "bbox": [ + 316, + 266, + 553, + 299 + ], + "type": "text", + "content": "[20] Anjan Dutta and Zeynep Akata. Semantically tied paired cycle consistency for zero-shot sketch-based image retrieval. In CVPR, 2019. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 301, + 553, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 301, + 553, + 334 + ], + "spans": [ + { + "bbox": [ + 316, + 301, + 553, + 334 + ], + "type": "text", + "content": "[21] Suyog Gupta, Ankur Agrawal, Kailash Gopalakrishnan, and Pritish Narayanan. Deep learning with limited numerical precision. In ICML, 2015. 2, 6, 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 335, + 553, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 335, + 553, + 357 + ], + "spans": [ + { + "bbox": [ + 316, + 335, + 553, + 357 + ], + "type": "text", + "content": "[22] David Ha and Douglas Eck. A neural representation of sketch drawings. In ICLR, 2018. 1, 5, 7, 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 358, + 553, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 358, + 553, + 392 + ], + "spans": [ + { + "bbox": [ + 316, + 358, + 553, + 392 + ], + "type": "text", + "content": "[23] Song Han, Huizi Mao, and William J Dally. Deep compression: Compressing deep neural networks with pruning, trained quantization and huffman coding. In ICLR, 2016. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 393, + 553, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 393, + 553, + 425 + ], + "spans": [ + { + "bbox": [ + 316, + 393, + 553, + 425 + ], + "type": "text", + "content": "[24] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 427, + 553, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 427, + 553, + 460 + ], + "spans": [ + { + "bbox": [ + 316, + 427, + 553, + 460 + ], + "type": "text", + "content": "[25] Geoffrey Hinton, Oriol Vinyls, and Jeff Dean. Distilling the knowledge in a neural network. In NeurIPS Deep Learning Workshop, 2014. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 462, + 553, + 494 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 462, + 553, + 494 + ], + "spans": [ + { + "bbox": [ + 316, + 462, + 553, + 494 + ], + "type": "text", + "content": "[26] Gao Huang, Zhuang Liu, Laurens Van Der Maaten, and Kilian Q Weinberger. Densely connected convolutional networks. In CVPR, 2017. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 496, + 553, + 517 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 496, + 553, + 517 + ], + "spans": [ + { + "bbox": [ + 316, + 496, + 553, + 517 + ], + "type": "text", + "content": "[27] Shouyou Huang and Qiang Wu. Robust pairwise learning with huber loss. Journal of Complexity, 2021. 4" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 519, + 553, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 519, + 553, + 552 + ], + "spans": [ + { + "bbox": [ + 316, + 519, + 553, + 552 + ], + "type": "text", + "content": "[28] Zehao Huang and Naiyan Wang. Like what you like: Knowledge distill via neuron selectivity transfer. arXiv preprint arXiv:1707.01219, 2017. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 554, + 553, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 554, + 553, + 575 + ], + "spans": [ + { + "bbox": [ + 316, + 554, + 553, + 575 + ], + "type": "text", + "content": "[29] Leslie Pack Kaelbling, Michael L Littman, and Andrew W Moore. Reinforcement learning: A survey. JAIR, 1996. 3, 5" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 578, + 553, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 578, + 553, + 609 + ], + "spans": [ + { + "bbox": [ + 316, + 578, + 553, + 609 + ], + "type": "text", + "content": "[30] Andrew Zisserman Karen Simonyan. Very deep convolutional networks for large-scale image recognition. In ICLR, 2015. 3, 4, 8" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 316, + 611, + 553, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 611, + 553, + 644 + ], + "spans": [ + { + "bbox": [ + 316, + 611, + 553, + 644 + ], + "type": "text", + "content": "[31] Debang Li, Huikai Wu, Junge Zhang, and Kaiqi Huang. A2-rl: Aesthetics aware reinforcement learning for image cropping. In CVPR, 2018. 3" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 316, + 647, + 553, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 678 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 678 + ], + "type": "text", + "content": "[32] Debang Li, Junge Zhang, and Kaiqi Huang. Learning to learn cropping models for different aspect ratio requirements. In CVPR, 2020. 6" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 316, + 680, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 680, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 680, + 553, + 712 + ], + "type": "text", + "content": "[33] Xiaodan Liang, Lisa Lee, and Eric P Xing. Deep variation-structured reinforcement learning for visual relationship and attribute detection. In CVPR, 2017. 3" + } + ] + } + ], + "index": 34 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28391" + } + ] + } + ], + "index": 36 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "type": "text", + "content": "[34] Darryl Lin, Sachin Talathi, and Sreekanth Annapureddy. Fixed point quantization of deep convolutional networks. In ICML, 2016. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 295, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 295, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 295, + 150 + ], + "type": "text", + "content": "[35] Hangyu Lin, Yanwei Fu, Yu-Gang Jiang, and Xiangyang Xue. Sketch-bert: Learning sketch bidirectional encoder representation from transformers by self-supervised learning of sketch gestalt. In CVPR, 2020. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 294, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 294, + 185 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 294, + 185 + ], + "type": "text", + "content": "[36] Li Liu, Fumin Shen, Yuming Shen, Xianglong Liu, and Ling Shao. Deep sketch hashing: Fast free-hand sketch-based image retrieval. In CVPR, 2017. 2, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 186, + 294, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 186, + 294, + 217 + ], + "spans": [ + { + "bbox": [ + 56, + 186, + 294, + 217 + ], + "type": "text", + "content": "[37] Jian-Hao Luo, Jianxin Wu, and Weiyao Lin. Thinet: A filter level pruning method for deep neural network compression. In ICCV, 2017. 2, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 220, + 294, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 220, + 294, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 220, + 294, + 262 + ], + "type": "text", + "content": "[38] Pavlo Molchanov, Stephen Tyree, Tero Karras, Timo Aila, and Jan Kautz. Pruning convolutional neural networks for resource efficient inference. arXiv preprint arXiv:1611.06440, 2016. 6, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 264, + 294, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 264, + 294, + 297 + ], + "spans": [ + { + "bbox": [ + 56, + 264, + 294, + 297 + ], + "type": "text", + "content": "[39] Umar Riaz Muhammad, Yongxin Yang, Yi-Zhe Song, Tao Xiang, and Timothy M Hospedales. Learning deep sketch abstraction. In CVPR, 2018. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 298, + 294, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 298, + 294, + 331 + ], + "spans": [ + { + "bbox": [ + 56, + 298, + 294, + 331 + ], + "type": "text", + "content": "[40] Umar Riaz Muhammad, Yongxin Yang, Timothy Hospedales, Tao Xiang, and Yi-Zhe Song. Goal-driven sequential data abstraction. In ICCV, 2019. 1, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 332, + 294, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 332, + 294, + 387 + ], + "spans": [ + { + "bbox": [ + 56, + 332, + 294, + 387 + ], + "type": "text", + "content": "[41] Maxime Oquab, Timothee Darct, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. DINOv2: Learning Robust Visual Features without Supervision. arXiv preprint arXiv:2304.07193, 2023. 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 388, + 294, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 388, + 294, + 422 + ], + "spans": [ + { + "bbox": [ + 56, + 388, + 294, + 422 + ], + "type": "text", + "content": "[42] Kaiyue Pang, Yi-Zhe Song, Tony Xiang, and Timothy M Hospedales. Cross-domain generative learning for fine-grained sketch-based image retrieval. In BMVC, 2017. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 422, + 294, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 422, + 294, + 464 + ], + "spans": [ + { + "bbox": [ + 56, + 422, + 294, + 464 + ], + "type": "text", + "content": "[43] Kaiyue Pang, Ke Li, Yongxin Yang, Honggang Zhang, Timothy M Hospedales, Tao Xiang, and Yi-Zhe Song. Generalising fine-grained sketch-based image retrieval. In CVPR, 2019. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 467, + 294, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 467, + 294, + 510 + ], + "spans": [ + { + "bbox": [ + 56, + 467, + 294, + 510 + ], + "type": "text", + "content": "[44] Kaiyue Pang, Yongxin Yang, Timothy M Hospedales, Tao Xiang, and Yi-Zhe Song. Solving mixed-modal jigsaw puzzle for fine-grained sketch-based image retrieval. In CVPR, 2020. 2, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 512, + 294, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 512, + 294, + 544 + ], + "spans": [ + { + "bbox": [ + 56, + 512, + 294, + 544 + ], + "type": "text", + "content": "[45] Nikolaos Passalis and Anastasios Tefas. Learning deep representations with probabilistic knowledge transfer. In ECCV, 2018. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 545, + 294, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 545, + 294, + 590 + ], + "spans": [ + { + "bbox": [ + 56, + 545, + 294, + 590 + ], + "type": "text", + "content": "[46] Pinaki Nath Chowdhury and Ayan Kumar Bhunia and Aneeeshan Sain and Subhadeep Koley and Tao Xiang and Yi-Zhe Song. SceneTrilogy: On Human Scene-Sketch and its Complementarity with Photo and Text. In CVPR, 2023. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 591, + 294, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 591, + 294, + 634 + ], + "spans": [ + { + "bbox": [ + 56, + 591, + 294, + 634 + ], + "type": "text", + "content": "[47] Pinaki Nath Chowdhury and Ayan Kumar Bhunia and Aeeeshan Sain and Subhadeep Koley and Tao Xiang and Yi-Zhe Song. What Can Human Sketches Do for Object Detection? In CVPR, 2023. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 635, + 294, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 294, + 668 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 294, + 668 + ], + "type": "text", + "content": "[48] Adriana Romero, Nicolas Ballas, Samira Ebrahimi Kahou, Antonie Chassang, Carlo Gatta, and Yoshua Bengio. Fitnets: Hints for thin deep nets. In ICLR, 2015. 2, 4, 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 56, + 670, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 294, + 713 + ], + "type": "text", + "content": "[49] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. IJCV, 2015. 4" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 34, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 116 + ], + "type": "text", + "content": "[50] Aneeshan Sain, Ayan Kumar Bhunia, Yongxin Yang, Tao Xiang, and Yi-Zhe Song. Cross-modal hierarchical modelling, fine-grained sketch based image retrieval. In BMVC, 2020, 2, 3, 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 118, + 553, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 553, + 161 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 553, + 161 + ], + "type": "text", + "content": "[51] Aneeshan Sain, Ayan Kumar Bhunia, Yongxin Yang, Tao Xiang, and Yi-Zhe Song. Stylemeup: Towards style-agnostic sketch-based image retrieval. In CVPR, 2021. 1, 2, 3, 4, 6, 7, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 163, + 553, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 553, + 205 + ], + "type": "text", + "content": "[52] Aneeshan Sain, Ayan Kumar Bhunia, Vaishnav Potlapalli, Pinaki Nath Chowdhury, Tao Xiang, and Yi-Zhe Song. Sketch3t: Test-time training for zero-shot sbir. In CVPR, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 207, + 553, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 207, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 316, + 207, + 553, + 251 + ], + "type": "text", + "content": "[53] Aneeshan Sain, Ayan Kumar Bhunia, Subhadeep Koley, Pinaki Nath Chowdhury, Soumitri Chattopadhyay, Tao Xiang, and Yi-Zhe Song. Exploiting Unlabelled Photos for Stronger Fine-Grained SBIR. In CVPR, 2023. 1, 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 252, + 553, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 252, + 553, + 295 + ], + "spans": [ + { + "bbox": [ + 316, + 252, + 553, + 295 + ], + "type": "text", + "content": "[54] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zh-moginov, and Liang-Chieh Chen. Mobilenetv2: Inverted residuals and linear bottlenecks. In CVPR, 2018. 1, 3, 4, 5, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 297, + 553, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 297, + 553, + 329 + ], + "spans": [ + { + "bbox": [ + 316, + 297, + 553, + 329 + ], + "type": "text", + "content": "[55] Patsorn Sangkloy, Nathan Burnell, Cusuh Ham, and James Hays. The sketchy database: learning to retrieve badly drawn bunnies. ACM TOG, 2016. 5, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 331, + 553, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 331, + 553, + 364 + ], + "spans": [ + { + "bbox": [ + 316, + 331, + 553, + 364 + ], + "type": "text", + "content": "[56] John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347, 2017. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 365, + 553, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 365, + 553, + 397 + ], + "spans": [ + { + "bbox": [ + 316, + 365, + 553, + 397 + ], + "type": "text", + "content": "[57] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. In ICLR, 2015. 1, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 399, + 553, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 399, + 553, + 430 + ], + "spans": [ + { + "bbox": [ + 316, + 399, + 553, + 430 + ], + "type": "text", + "content": "[58] Jifei Song, Yi-Zhe Song, Tony Xiang, and Timothy M Hospedales. Fine-grained image retrieval: the text/sketch input dilemma. In BMVC, 2017. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 432, + 553, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 432, + 553, + 465 + ], + "spans": [ + { + "bbox": [ + 316, + 432, + 553, + 465 + ], + "type": "text", + "content": "[59] Jifei Song, Qian Yu, Yi-Zhe Song, Tao Xiang, and Timothy M Hospedales. Deep spatial-semantic attention for fine-grained sketch-based image retrieval. In ICCV, 2017. 2, 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 466, + 553, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 466, + 553, + 510 + ], + "spans": [ + { + "bbox": [ + 316, + 466, + 553, + 510 + ], + "type": "text", + "content": "[60] Subhadeep Koley and Ayan Kumar Bhunia and Aneeshan Sain and Pinaki Nath Chowdhury and Tao Xiang and Yi-Zhe Song. Picture that Sketch: Photorealistic Image Generation from Abstract Sketches. In CVPR, 2023. 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 511, + 553, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 511, + 553, + 544 + ], + "spans": [ + { + "bbox": [ + 316, + 511, + 553, + 544 + ], + "type": "text", + "content": "[61] Richard S Sutton, David A McAllester, Satinder P Singh, and Yishay Mansour. Policy gradient methods for reinforcement learning with function approximation. In NeurIPS, 2000. 5" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 545, + 553, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 545, + 553, + 567 + ], + "spans": [ + { + "bbox": [ + 316, + 545, + 553, + 567 + ], + "type": "text", + "content": "[62] Hossein Talebi and Peyman Milanfar. Learning to resize images for computer vision tasks. In ICCV, 2021. 2, 6" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 568, + 553, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 568, + 553, + 600 + ], + "spans": [ + { + "bbox": [ + 316, + 568, + 553, + 600 + ], + "type": "text", + "content": "[63] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In ICML, 2019. 1, 3" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 601, + 553, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 601, + 553, + 633 + ], + "spans": [ + { + "bbox": [ + 316, + 601, + 553, + 633 + ], + "type": "text", + "content": "[64] Burak Uzkent and Stefano Ermon. Learning when and where to zoom with deep reinforcement learning. In CVPR, 2020. 2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 316, + 635, + 553, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 553, + 668 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 553, + 668 + ], + "type": "text", + "content": "[65] Supawit Vatathanavaro, Suchat Tungjitnob, and Kitsuchart Pasupa. White blood cell classification: a comparison between vgg-16 and resnet-50 models. In JSCI, 2018. 3" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "text", + "content": "[66] Mahes Visvalingam and J Duncan Whyatt. The douglaspeucker algorithm for line simplification: re-evaluation through visualization. In Computer Graphics Forum, 1990. 5" + } + ] + } + ], + "index": 33 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "28392" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 544 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 127 + ], + "type": "text", + "content": "[67] Xin Wang, Qiuyuan Huang, Asli Celikyilmaz, Jianfeng Gao, Dinghan Shen, Yuan-Fang Wang, William Yang Wang, and Lei Zhang. Reinforced cross-modal matching and self-supervised imitation learning for vision-language navigation. In CVPR, 2019. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 294, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 294, + 163 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 294, + 163 + ], + "type": "text", + "content": "[68] Yunhe Wang, Chang Xu, Shan You, Dacheng Tao, and Chao Xu. Cnnpack: Packing convolutional neural networks in the frequency domain. NeurIPS, 29, 2016. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 164, + 294, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 164, + 294, + 206 + ], + "spans": [ + { + "bbox": [ + 56, + 164, + 294, + 206 + ], + "type": "text", + "content": "[69] Yulin Wang, Kangchen Lv, Rui Huang, Shiji Song, Le Yang, and Gao Huang. Glance and focus: a dynamic approach to reducing spatial redundancy in image classification. In NeurIPS, 2020. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 208, + 294, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 208, + 294, + 239 + ], + "spans": [ + { + "bbox": [ + 56, + 208, + 294, + 239 + ], + "type": "text", + "content": "[70] Kilian Q Weinberger and Lawrence K Saul. Distance metric learning for large margin nearest neighbor classification. JMLR, 2009. 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 241, + 294, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 241, + 294, + 274 + ], + "spans": [ + { + "bbox": [ + 56, + 241, + 294, + 274 + ], + "type": "text", + "content": "[71] Peng Xu, Chaitanya K Joshi, and Xavier Bresson. Multigraph transformer for free-hand sketch recognition. IEEE T-NNLS, 2021. 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 275, + 294, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 275, + 294, + 308 + ], + "spans": [ + { + "bbox": [ + 56, + 275, + 294, + 308 + ], + "type": "text", + "content": "[72] Peng Xu, Timothy M Hospedales, Qiyue Yin, Yi-Zhe Song, Tao Xiang, and Liang Wang. Deep learning for free-hand sketch: A survey. IEEE TPAMI, 2022. 1, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 309, + 294, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 309, + 294, + 352 + ], + "spans": [ + { + "bbox": [ + 56, + 309, + 294, + 352 + ], + "type": "text", + "content": "[73] Ran Xu, Fangzhou Mu, Jayoung Lee, Preeti Mukherjee, Somali Chaterji, Saurabh Bagchi, and Yin Li. SmartAdapt: Multi-branch object detection framework for videos on mobiles. In CVPR, 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 354, + 294, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 354, + 294, + 387 + ], + "spans": [ + { + "bbox": [ + 56, + 354, + 294, + 387 + ], + "type": "text", + "content": "[74] Qian Yu, Yongxin Yang, Yi-Zhe Song, Tao Xiang, and Timothy M. Hospedales. Sketch-a-net that beats humans. In BMVC, 2015. 6, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 388, + 294, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 388, + 294, + 421 + ], + "spans": [ + { + "bbox": [ + 56, + 388, + 294, + 421 + ], + "type": "text", + "content": "[75] Qian Yu, Feng Liu, Yi-Zhe Song, Tao Xiang, Timothy M Hospedales, and Chen-Change Loy. Sketch me that shoe. In CVPR, 2016. 1, 2, 3, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 422, + 294, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 422, + 294, + 464 + ], + "spans": [ + { + "bbox": [ + 56, + 422, + 294, + 464 + ], + "type": "text", + "content": "[76] Sergey Zagoruyko and Nikos Komodakis. Paying more attention to attention: Improving the performance of convolutional neural networks via attention transfer. In ICLR, 2017. 2, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 467, + 294, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 467, + 294, + 510 + ], + "spans": [ + { + "bbox": [ + 56, + 467, + 294, + 510 + ], + "type": "text", + "content": "[77] Shuchang Zhou, Yuxin Wu, Zekun Ni, Xinyu Zhou, He Wen, and Yuheng Zou. Dorefa-net: Training low bitwidth convolutional neural networks with low bitwidth gradients. arXiv preprint arXiv:1606.06160, 2016. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 512, + 294, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 512, + 294, + 544 + ], + "spans": [ + { + "bbox": [ + 56, + 512, + 294, + 544 + ], + "type": "text", + "content": "[78] Mingjian Zhu, Kai Han, Enhua Wu, Qiulin Zhang, Ying Nie, Zhenzhong Lan, and Yunhe Wang. Dynamic resolution network. In NeurIPS, 2021. 7" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28393" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/SketchAgent_ Language-Driven Sequential Sketch Generation/76b8d485-0aca-43aa-b3b2-9c94133cc069_content_list.json b/2025/SketchAgent_ Language-Driven Sequential Sketch Generation/76b8d485-0aca-43aa-b3b2-9c94133cc069_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..06c69e5625b0cec3c1f3d8f788249d0d4948f400 --- /dev/null +++ b/2025/SketchAgent_ Language-Driven Sequential Sketch Generation/76b8d485-0aca-43aa-b3b2-9c94133cc069_content_list.json @@ -0,0 +1,1900 @@ +[ + { + "type": "text", + "text": "SketchAgent: Language-Driven Sequential Sketch Generation", + "text_level": 1, + "bbox": [ + 169, + 94, + 799, + 116 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yael Vinker $^{1}$ Tamar Rott Shaham $^{1}$ Kristine Zheng $^{2}$ Alex Zhao $^{1}$ Judith E Fan $^{2}$ Antonio Torralba $^{1}$", + "bbox": [ + 84, + 132, + 883, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ MIT", + "bbox": [ + 315, + 167, + 362, + 185 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{yaelvink,tamarott,alexzhao,torralba}@mit.edu", + "bbox": [ + 140, + 189, + 537, + 205 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Stanford University", + "bbox": [ + 620, + 167, + 787, + 186 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{jefan, kxzheng} $@$ stanford.edu", + "bbox": [ + 583, + 189, + 828, + 205 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "https://sketch-agent.csail.mit.edu/", + "bbox": [ + 326, + 215, + 638, + 229 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ab70df958686f2d2f5dd138cd4518fae137431adbcbf0135e8142d26148c2090.jpg", + "image_caption": [ + "Figure 1. SketchAgent leverages an off-the-shelf multimodal LLM to facilitate language-driven, sequential sketch generation through an intuitive sketching language. It can sketch diverse concepts, engage in interactive sketching with humans, and edit content via chat." + ], + "image_footnote": [], + "bbox": [ + 109, + 248, + 328, + 462 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/660caff867d36bd3a2a1921ee55b7b3b76542d9b1364d0157c87333634aa747d.jpg", + "image_caption": [ + "Human-Agent Collaborative Sketching" + ], + "image_footnote": [], + "bbox": [ + 370, + 290, + 619, + 455 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/bb89323786ba343d8e554d354ea822316c073446fd92f4ecc14a65f0ac735d29.jpg", + "image_caption": [ + "Chat-Based Sketch Editing" + ], + "image_footnote": [], + "bbox": [ + 647, + 271, + 720, + 325 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Draw a robot", + "bbox": [ + 759, + 273, + 834, + 285 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Add a paintbrush and a canvas", + "bbox": [ + 736, + 321, + 834, + 345 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/58090f024aca3cd9a38c512e69f40fb0834a6bf929eb5b573f55723bd62c16a8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 648, + 335, + 720, + 388 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Draw a robot on the canvas", + "bbox": [ + 741, + 383, + 834, + 409 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/754aa295f56e6b4570d9a42392daf32d422e9c28aa0ecbf726d93912ff19e24f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 648, + 397, + 720, + 452 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 512, + 313, + 527 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Sketching serves as a versatile tool for externalizing ideas, enabling rapid exploration and visual communication that spans various disciplines. While artificial systems have driven substantial advances in content creation and human-computer interaction, capturing the dynamic and abstract nature of human sketching remains challenging. In this work, we introduce SketchAgent, a language-driven, sequential sketch generation method that enables users to create, modify, and refine sketches through dynamic, conversational interactions. Our approach requires no training or fine-tuning. Instead, we leverage the sequential nature and rich prior knowledge of off-the-shelf multimodal large language models (LLMs). We present an intuitive sketching language, introduced to the model through in-context examples, enabling it to \"draw\" using string-based actions. These are processed into vector graphics and then rendered to create a sketch on a pixel canvas, which can be accessed again for further tasks. By drawing stroke by stroke, our agent captures the evolving, dynamic qualities intrinsic to sketching. We demonstrate that SketchAgent can generate sketches from diverse prompts, engage in dialogue-driven drawing, and collaborate meaningfully with human users.", + "bbox": [ + 73, + 551, + 472, + 883 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 501, + 512, + 630, + 527 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Sketching is a powerful tool for distilling ideas into their simplest form. Its fluid and spontaneous nature makes sketching a uniquely versatile tool for visualization, rapid ideation, and communication across cultures, generations, and disciplines [26, 102]. For example, designers use sketches to explore new ideas [39, 103], scientists employ them to formulate problems [48, 72], and children engage in sketching to learn and express themselves [27, 28] (see Fig. 2). Artificial systems, in principle, have the potential to support and enhance human creativity, problem-solving, and visual expression through sketching, adapting flexibly to their exploratory nature [22, 98, 122].", + "bbox": [ + 496, + 537, + 890, + 718 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Traditionally, sketch generation methods rely on human-drawn datasets to train generative models [5, 6, 16, 36, 42, 59]. However, fully capturing the diversity of sketches within datasets remains challenging [26], limiting these methods in both scale and diversity. Recent advancements in vision-language models, such as CLIP [78] and text-to-image diffusion [82], have enabled sketch generation methods that reduce reliance on human-drawn datasets [29, 46, 105]. These methods leverage pretrained model guidance and differentiable rendering [58] to optimize parametric curves, creating sketches that go beyond predefined styles and categories.", + "bbox": [ + 496, + 719, + 892, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "23355", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "While representing a significant step toward a general-purpose sketching system, these methods lack a crucial aspect of human drawing: the process itself. Current methods, though versatile, optimize all strokes simultaneously, making the intermediate sketching steps meaningless. As a result, the sketch cannot be decomposed into a coherent sequence of strokes that reflects the drawing process. In contrast, humans draw iteratively, stroke by stroke, incorporating visual feedback and continuously adapting—a dynamic, evolving process that fosters creativity, ideation, and communication [52, 88, 101].", + "bbox": [ + 75, + 90, + 472, + 257 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we introduce SketchAgent, a sketch generation agent that leverages the prior knowledge and sequential nature of multimodal large language models (LLMs) to enable versatile, progressive, language-driven sketching. Our agent can generate sketches across a wide range of textual concepts—from animals to engineering principles (Fig. 1, left). Its sequential nature facilitates interactive human-agent sketching and supports iterative refinement and editing through a chat-based dialogue (Fig. 1, right).", + "bbox": [ + 75, + 260, + 472, + 397 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Unlike vision-language models that directly generate images from text [75, 80, 82], multimodal LLMs [1, 2, 15, 56, 64, 74, 97] accept text and images as input but only output text. To produce visuals, they either utilize external \"tools\" (such as calling a text-to-image model) or are prompted to generate executable code (e.g., Python [43], SVG [9]) to create charts, diagrams, or graphics. However, prompting for such representations to directly produce sketches often results in a mechanical appearance with uniform, precise shapes that lack the subtle irregularities and spontaneous qualities characteristic of human sketches (see Fig. 3B). Additionally, despite their robustness in textual tasks, these models often struggle with fine-grained spatial reasoning [41, 118] as they are primarily optimized for text, making sketch editing more challenging.", + "bbox": [ + 75, + 398, + 472, + 626 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address these limitations, we introduce an intuitive sketching language that enables an off-the-shelf multimodal LLM agent to \"draw\" sketches on a canvas by providing string-based actions, without additional training or finetuning. We define the canvas as a numbered grid, allowing the agent to reference specific coordinates (e.g., x2y8) to enhance its spatial reasoning capabilities. We represent a sketch as a sequence of semantically meaningful strokes, each defined by a series of such coordinates. We leverage In-Context Learning (ICL) [7, 51] to introduce the agent to the new representation, and Chain of Thought (CoT) [108] to enhance its planning capabilities. Given a sketching task, the agent produces a textual response following our representation, which we process by fitting a smooth Bézier curve to each coordinate sequence. The curves are then rendered onto the canvas to form the final sketch. We find this approach useful in emulating a more natural sketch appearance. For collaborative sketching, the canvas remains acces", + "bbox": [ + 75, + 628, + 472, + 901 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/79df84ee3c76e6a87bd277ede69140cb5c63521e15ceac8868d03a0ab43c608d.jpg", + "image_caption": [ + "(A) Frank Gehry. *Elevation Sketches*. Guggenheim Museum" + ], + "image_footnote": [], + "bbox": [ + 531, + 116, + 725, + 218 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/84445e52d6588ee0eb6c7ec6a9e37694b8fc18440ca5cc76762989cfe593d383.jpg", + "image_caption": [ + "(B) Alexander Bell's drawing of the telephone" + ], + "image_footnote": [], + "bbox": [ + 531, + 244, + 725, + 323 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/7f2b351f7e27a24b7b48e583019243e8bb3c15676cd7c4566913578b20efa11e.jpg", + "image_caption": [ + "(C) Children drawing" + ], + "image_footnote": [], + "bbox": [ + 730, + 107, + 859, + 178 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/094c36db6efb33f65a4fa9b7c75cc6344415ae800a93f97dbef517957c6988be.jpg", + "image_caption": [ + "(D) Basketball strategy (pick and roll)", + "Figure 2. Examples of sketches used across disciplines and goals. (A) Ideation and design: Process Elevation Sketches by the architect Frank Gehry, Guggenheim Museum. (B) Engineering: Alexander Bell's telephone drawing. (C) Expressing emotions: Children's sketches. (D) Visual communication: Planning and communicating game strategy in basketball." + ], + "image_footnote": [], + "bbox": [ + 730, + 209, + 861, + 325 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "sible to both the user and the agent throughout the session. The agent generates strokes sequentially and pauses according to an adjustable stopping token, allowing the user to add their own strokes directly to the canvas. These strokes are then integrated into the agent's sequence, enabling it to continue drawing, with real-time canvas updates.", + "bbox": [ + 496, + 424, + 890, + 513 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We demonstrate SketchAgent's capability to generate sketches of diverse concepts while capturing the inherently sequential and dynamic nature of sketching. We showcase our agent's ability to collaborate effectively with humans in real time to create novel and meaningful sketches. Our method is the first to leverage pretrained multimodal LLMs for sequential sketching without additional training, paving the way for a general-purpose artificial sketching system that supports iterative, evolving interactivity.", + "bbox": [ + 496, + 515, + 893, + 650 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 664, + 640, + 679 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Sketch Generation Early methods approached sketch generation by designing image filters to simulate sketchlike effects [10, 109]. With the advent of deep learning, data-driven approaches emerged to address a range of sketch-related tasks [117], including category-conditioned sketching [42, 76, 93], object sketching [59, 62], scenesketching [12, 57, 60, 114], sketch completion [6, 63, 94], portrait drawing [3, 120, 121], part-based generation [6, 37, 42, 129], and more. While sketch data collection has been broadly explored [21, 34, 40, 71, 85, 113], the wide variation in sketch styles and their adaptation to specific tasks [23] makes collecting datasets that encompass this diversity challenging. For example, QuickDraw [47], the largest available sketch dataset with 50 mil", + "bbox": [ + 496, + 688, + 893, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "23356", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "lion sketches, covers only 345 object categories and primarily focuses on simple, iconic representations. This limits data-driven methods to the style, abstraction level, and concepts seen during training. Recently, large pretrained vision-language models [75, 78, 80, 82, 84] have shown remarkable text-to-image generation capabilities by leveraging extensive visual knowledge from billions of training images [89]. While these models can be prompted to generate sketch-like images (see Fig. 3A), they do so in a single step and in pixel space, lacking the sequential, stroke-based process of human sketching. Subsequent approaches [14, 29, 31, 46, 105, 106, 115, 116, 124] leverage the priors of these models to guide an iterative optimization of parametric curves, with a differentiable rasterizer [58] linking pixel and vector representations. While producing vector sketches, the final strokes lack order and semantic meaning, and the optimization-based approach overlook the sequential aspect of the sketching process, making these methods suboptimal for collaborative sketching.", + "bbox": [ + 75, + 90, + 472, + 380 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Sequential and Collaborative Sketching Collaborative human-machine sketching holds promise in enhancing creativity, ideation, communication, and learning, as explored in various fields, including human-computer interaction (HCI) [17, 45, 49, 50, 53, 54], computer graphics [55, 96], robotics [86, 87], cognitive science [24, 25, 35, 67], learning sciences [18, 38, 104], and more. Central to collaborative sketching is its sequential, adaptive, and dynamic process, with each action carrying intent. Existing methods employ diverse training strategies to account for the discrete nature of sequential sketches, including reinforcement and adversarial learning [32, 68, 129], multi-agent referential games [69, 77], transformers [5, 6, 11, 33, 61, 81, 112], and more. SketchRNN [42] is a pioneering work in this area, introducing the QuickDraw dataset [47], a crowd-sourced collection of real-time sketch sequences made by users. They utilize this dataset to train a recurrent neural network for sequential sketch generation, which was later shown [24, 73] to have potential for human-machine collaboration. However, this approach remains constrained by the predefined categories encountered during training.", + "bbox": [ + 75, + 388, + 472, + 708 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Multimodel LLMs for Content Creation LLMs [7, 19, 79, 100] and multimodal LLMs [1, 2, 15, 56, 64, 74, 97] receive text as input (or text and images for multimodal) and output text. To enable visual content generation, these models are often paired with external \"tools\" that extend their functionality [44, 90, 111, 119]. For example, ChatGPT [74] generates images by internally calling a separate model, DALLE-3 [4]. Another approach involves prompting models to produce code in languages like Python [43], Processing [92], SVG [9], or TikZ [8] that can be rendered into visuals such as graphs, charts, and vector graphics. However, such code-generated content often looks", + "bbox": [ + 75, + 719, + 472, + 901 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/ecebd173d5244262321496c10bc7579f13a3ede01ed0683c1bba5b4ff4bfd7d6.jpg", + "image_caption": [ + "(A) DALLE3 [4]" + ], + "image_footnote": [], + "bbox": [ + 506, + 90, + 598, + 162 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/269d98ca61721ce69d816b986175812dc51676e984fa2d9bdfd596c779d1bca9.jpg", + "image_caption": [ + "(B) LLMs [2] (SVG)" + ], + "image_footnote": [], + "bbox": [ + 612, + 107, + 689, + 140 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/26f29f516ec09b7cad0c0384e589dc27ba4c5f593fda2154d43f5b99be1eeb9b.jpg", + "image_caption": [ + "(C) SketchAgent" + ], + "image_footnote": [], + "bbox": [ + 715, + 108, + 784, + 140 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/30f937880ee2cde64f8b31b9bf39dc243cda3140563dd4a93f0198d2aa2d28be.jpg", + "image_caption": [ + "(D) Human [47]", + "Figure 3. Sketch appearance. (A) Text-to-image diffusion models operate in pixel space, lacking the sequential nature of sketches. (B) Prompting LLMs to produce visuals with SVG results in a uniform, mechanical appearance. (C) Sketches produced by our agent appear less mechanical, more closely resembling the nature of (D) Human sketches, which are often spontaneous and irregular." + ], + "image_footnote": [], + "bbox": [ + 813, + 114, + 879, + 143 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "rigid, with uniform and overly precise shapes that lack the subtle irregularities and spontaneous qualities characteristic of freehand sketches (see Fig. 3B). In contrast, we propose a sketching language grounded in spatial information that encourages the model to produce a more natural sketch appearance, which we then process into vector graphics. Common strategies for enhancing LLMs capabilities include Chain-of-Thought prompting [13, 70, 83, 91, 127], which breaks down tasks into smaller, logical steps to mimic human reasoning, and In-Context Learning (ICL) [7, 20, 95, 123, 125], where examples of input-output pairs are provided to help the model infer task patterns.", + "bbox": [ + 496, + 301, + 893, + 484 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Preliminaries", + "text_level": 1, + "bbox": [ + 500, + 505, + 638, + 522 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Vector Graphics and Bezier Curves Vector graphics allow us to create visual images directly from geometric shapes such as points, lines, curves, and polygons. Unlike raster images (represented with pixels), vector graphics are", + "bbox": [ + 500, + 532, + 738, + 638 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/cfd520c98eb43f76e42eb72ff93f21bfc86c01d6fb8de66fb635b103b300bc01.jpg", + "image_caption": [ + "Figure 4. Cubic Bezier curve." + ], + "image_footnote": [], + "bbox": [ + 774, + 502, + 885, + 595 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "resolution-free, more compact, and editable. SVG [110] is an XML-based format for storing vector graphics, popular for its scalability and compatibility with modern web browsers. The process of transferring vector graphics into pixel images is called rasterization or rendering. Cubic Bézier curves are commonly used to represent sketches in vector graphics. A cubic Bézier curve (Fig. 4) is a smooth parametric curve defined by four points: a start point $P_0$ , an end point $P_3$ , and two control points $P_1$ and $P_2$ that shape the curvature. The set $P = \\{P_0, P_1, P_2, P_3\\}$ is often referred to as the curve's control points. The curve is described by the following polynomial equation:", + "bbox": [ + 496, + 638, + 893, + 820 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nB (t) = (1 - t) ^ {3} P _ {0} + 3 (1 - t) ^ {2} t P _ {1} + 3 (1 - t) t ^ {2} P _ {2} + t ^ {3} P _ {3}, (1)\n$$\n", + "text_format": "latex", + "bbox": [ + 508, + 835, + 890, + 854 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $t \\in [0,1]$ is a parameter that moves the point along the curve from $P_0$ at $t = 0$ to $P_3$ at $t = 1$ .", + "bbox": [ + 498, + 869, + 890, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "23357", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/0f06d8a6a6aa9521e66a99348e8b287aae238090cc82646635cf1398b7363315.jpg", + "image_caption": [ + "Figure 5. Method Overview. SketchAgent (blue) receives drawing instructions and generates a string representing the intended sketch. Inputs include: (1) a system prompt (orange) introducing the sketching language and canvas, (2) a user prompt (pink) specifying the task (e.g., \"draw a shark\"), and (3) a numbered canvas. The agent's response outlines a sketching strategy (in thinking tags) and a sequence of strokes defined by coordinates, which are processed into Bézier curves and rendered onto the canvas." + ], + "image_footnote": [], + "bbox": [ + 81, + 88, + 888, + 244 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Method", + "text_level": 1, + "bbox": [ + 76, + 320, + 168, + 335 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our goal is to enable an off-the-shelf pretrained multimodal LLM to draw sketches based on natural language instructions. An overview of our pipeline is illustrated in Fig. 5. We utilize a frozen multimodal LLM (\"SketchAgent\" shown in blue), which receives three inputs: (1) a system prompt containing guidelines for using our new sketching language, (2) a user prompt with additional task-specific instructions (e.g., \"Draw a shark\"), and (3) a blank canvas on which the agent can draw. Based on the given task, the agent generates a textual response, representing the sequence of strokes to be drawn, which we then process into vector graphics and render onto the canvas. The canvas can then be reused in two ways: it can be fed back into the model with an updated user prompt for additional tasks and editing, or it can be accessed by a human user who can draw directly on it to facilitate collaborative sketching. Next, we describe each component of the pipeline.", + "bbox": [ + 75, + 347, + 473, + 604 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The Canvas Although multimodal LLMs demonstrate remarkable reasoning abilities, they often struggle with spatial reasoning tasks [30, 66, 99]. We present a simple example (see Fig. 6) to illustrate how this limitation affects the naive use of these models for sketch generation and interactive sketching. We provide GPT-4o [74] with an image depicting a simple line drawing of a partial house featuring five numbered points (from 1 to 5), and ask it to identify which points should be connected to complete the house. While the model correctly identifies the pair of points, it fails to select the correct pixel coordinates when given a basic draw-line tool that connects two points, even after multiple attempts. To enhance the model's spatial reasoning ability, we utilize a numbered canvas that forms a grid. This grid features numbers (1 to 50) along the x-axis and the y-axis (Fig. 5, left). Each cell is uniquely identified by a combination of the corresponding x-axis and y-axis numbers (e.g., the bottom-left cell is $x1y1$ ). The agent interacts with the canvas by specifying desired $(x,y)$ coordinates.", + "bbox": [ + 75, + 613, + 472, + 902 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/eaff46c17f5e0ca25ebf8cfe4fbb32a448f235608a9bfa2adcfa2d90231fce5b.jpg", + "image_caption": [ + "Figure 6. Although excelling in visual reasoning, multimodal LLMs often struggle to translate these abilities into spatial actions. In this example, GPT-4o [74] intends to draw a line between points 1 and 5 but fails to execute this with a draw_line function that accepts pixel coordinates." + ], + "image_footnote": [], + "bbox": [ + 506, + 319, + 883, + 393 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Sketch Representation We define a sketch as a sequence of $n$ ordered strokes $S = \\{S_1, S_2, \\ldots, S_n\\}$ . Each stroke $S_i$ is defined by a sequence of $m$ cell coordinates on the grid: $S_i = \\{(x_j, y_j)\\}_{j=1}^m$ , represented in string format as: x1y1, x15y20, ....", + "bbox": [ + 496, + 488, + 890, + 565 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "A naive approach to processing the textual sequence of coordinates would be to use a polyline, connecting consecutive points with line segments. However, our grid-based representation sparsifies the canvas, resulting in a non-smooth and unnatural appearance when using polylines (see Fig. 7, left). To achieve a smoother appearance, an alternative approach is to treat the coordinates as a sequence of control points defining smooth curves. However, as illustrated in Fig. 4, the control points often do not lie directly on the curve. Consequently, if the agent aims for a stroke that passes through specific coordinates, it must derive the control points that define this stroke, which is challenging.", + "bbox": [ + 496, + 566, + 892, + 747 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We propose an alternative approach: we treat the specified $(x, y)$ coordinates as a set of desired points sampled along the curve, and fit a smooth Bézier curve to them (Fig. 7, right). To accommodate curves with complex curvature, we also task the model with determining when each point on the curve should be passed through, corresponding to the $t$ value described in Eq. (1). Thus, for each stroke $S_{i}$ , the agent provides a set of $m$ sampled points $S_{i} = \\{(x_{j}, y_{j})\\}_{j=1}^{m}$ , along with a corresponding set of $t$ values: $T_{i} = \\{t_{j}\\}_{j=1}^{m}$ . Based on these, we fit a cubic Bézier", + "bbox": [ + 496, + 750, + 893, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "23358", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/212fa6867ce4bdcbddaa82751742056c0944857271c1b4b8a0fc7031b67326ee.jpg", + "image_caption": [ + "(A) Polyline" + ], + "image_footnote": [], + "bbox": [ + 117, + 88, + 217, + 143 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/d04feb1e43212ca5927cdd3bfb132b6d589709712778a0904e76bc0a14265623.jpg", + "image_caption": [ + "Figure 7. Methods for processing the agent's coordinate sequence (in red): (A) Polyline results in an unnatural appearance. (B) Directly using coordinates as Bézier control points is challenging as they do not lie on the curve. (C) Fitting a Bézier curve to sampled coordinates provides smoother results." + ], + "image_footnote": [], + "bbox": [ + 220, + 88, + 321, + 143 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/87bdbe0a578ba8382f0d291ffe00f966941310f2175de173f265c8c1c46094b2.jpg", + "image_caption": [ + "(B) Control Points", + "(C) Ours" + ], + "image_footnote": [], + "bbox": [ + 325, + 88, + 424, + 143 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "curve to the sampled points by solving a system of linear equations using least squares, where the unknowns are the control points $P = \\{P_0, P_1, P_2, P_3\\}$ :", + "bbox": [ + 75, + 253, + 468, + 299 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nP = \\operatorname {a r g m i n} _ {P} \\| A P - B \\|, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 184, + 305, + 468, + 321 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $A \\in \\mathbb{R}^{m \\times 4}$ contains the cubic Bézier basis functions evaluated at specific $t_j$ values (as described in Eq. (1)), and $B \\in \\mathbb{R}^{m \\times 2}$ contains the $m$ sampled points $\\{(x_j, y_j)\\}_{j=1}^m$ . The least squares solution minimizes the error between the fitted Bézier curve and the sampled points. For long sequences resulting in a large fitting error, we recursively split the curve. Additionally, we account for Bézier curves of lower degrees, including quadratic curves, linear lines, and points. Upon completing this process, we render the parametric curves onto the canvas.", + "bbox": [ + 75, + 328, + 468, + 479 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Drawing Instructions We provide the model with a system prompt and a user prompt (marked in orange and pink in Fig. 5). In the system prompt, we supply the agent with context about its expertise (\"You are an expert artist specializing in drawing sketches\") and introduce it to the grid canvas along with examples of how to use our sketching language for drawing single-stroke primitives (full prompts are provided in the Appendix). The system prompt is fixed and can be applied to a variety of sketching tasks. The user prompt includes a description of the desired task and an example of a simple sketch of a house drawn with our sketching language. We find this to be crucial in assisting the agent with preserving the correct format that could be parsed directly [7]. The agent is tasked with responding in the format shown in the gray text box in Fig. 5. In the tags, the agent is tasked to outline the overall sketching strategy [108]. This typically includes describing the different components of the sketch, the intended sketching order, and the overall placement of each part. The agent is also tasked with providing an ID tag following each stroke, which is useful for further analysis and for producing annotated sketches in scale.", + "bbox": [ + 75, + 491, + 468, + 821 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. In-Chat Editing and Collaborative Sketching", + "text_level": 1, + "bbox": [ + 76, + 832, + 457, + 848 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The above process can be repeated iteratively to support multiple sketching tasks and interactions. Text-based sketch editing in a chat dialogue is enabled by feeding the rendered", + "bbox": [ + 75, + 854, + 468, + 900 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/95f2cb4f29aad5591b1e0194e2cf0ab7526e4b5e016af1d9e988b66435856fc1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 506, + 95, + 602, + 154 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/2ce6c74333131c52fd4d2db63f898a49a0e211208a5a003b97036fbaedfe1a24.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 611, + 94, + 687, + 143 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/a200c7aa52d21ba8b21482f41ac05341be18a89a9ccba901fa92fd467b789848.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 714, + 90, + 774, + 148 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/1bd05a8d61780921afe2f41ff9d85cf2dc4418c435896e452d678996aeb9e97a.jpg", + "image_caption": [ + "DNA Double Helix" + ], + "image_footnote": [], + "bbox": [ + 823, + 95, + 856, + 152 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/deab0063155c5254908d052613efaa75c2a6d154ebe3baeeead4c1ce067194f5.jpg", + "image_caption": [ + "Golden Gate Bridge" + ], + "image_footnote": [], + "bbox": [ + 527, + 194, + 576, + 231 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/e1c42a497153ef5a8dfb3a8737e16da2cabd3d437ad02ee06a5551aba0e1d2af.jpg", + "image_caption": [ + "Mount Fuji", + "Pendulum Motion", + "Double-Slit Experiment" + ], + "image_footnote": [], + "bbox": [ + 614, + 202, + 679, + 244 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/6fd0668cd32e45b057920fd584981d956b47ad05709d30cd8c47068995420b89.jpg", + "image_caption": [ + "Eiffel Tower", + "Flowchart", + "Figure 8. Sketches produced by SketchAgent for concepts beyond pre-defined categories. The textual input describing the desired concept shown below each image." + ], + "image_footnote": [], + "bbox": [ + 712, + 191, + 776, + 257 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/661e67bb48b898e26fae04c7637ad316784dbe6caf1e6bdcdac6af5e7b98976f.jpg", + "image_caption": [ + "Tic-tac toe" + ], + "image_footnote": [], + "bbox": [ + 810, + 200, + 867, + 247 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "canvas back to the agent (see dashed arrow in Fig. 5) and updating the user prompt with the desired edits. To support collaborative human-agent sketching, the canvas remains accessible to both the human user and the agent throughout the entire sketching session. We define an adjustable stopping token, $\\langle \\text{}/\\text{s}\\{\\text{j}\\} \\rangle$ , which instructs the agent to pause generating the sequence at stroke number $j$ . We then process and render the generated strokes onto the canvas up to that point, then the user can add strokes directly to the canvas to continue the sketch. The user-drawn strokes are processed and converted into the agent's format by reversing our fitting process, i.e., sampling each stroke at multiple $t$ values (as shown in Eq. (1)), and selecting the points closest to each cell's center on the grid. The converted user strokes are then chained with the agent's sequence, after which the agent resumes sketching until the next stopping token.", + "bbox": [ + 496, + 351, + 890, + 592 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Results", + "text_level": 1, + "bbox": [ + 500, + 607, + 586, + 622 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We evaluate the performance of our method qualitatively and quantitatively across a selected set of sketching tasks. Additional tasks, evaluations, and examples are provided in the Appendix. All results presented in the paper were generated using Claude3.5-Sonnet [2] as our backbone model, unless stated otherwise.", + "bbox": [ + 496, + 628, + 890, + 718 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. Text-Conditioned Sketch Generation", + "text_level": 1, + "bbox": [ + 500, + 729, + 816, + 744 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Figures 1 and 8 demonstrate SketchAgent's capability to generate sketches of various concepts that extend beyond standard categories, which includes scientific concepts (e.g., \"the double-slit experiment\", \"pendulum motion\"), diagrams (e.g., \"circuit diagram\", \"a flowchart\"), and notable landmarks (e.g., \"Taj Mahal\", \"Eiffel Tower\"). More examples are provided in the Appendix. To quantitatively evaluate text-conditioned generation we utilize the QuickDraw dataset [47]. We randomly sample 50 categories (out of 345), and apply our method to generate 10 sketch in", + "bbox": [ + 496, + 750, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "23359", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/4191ea027dcedc9e279a68a7b72ded7294c69787c2b4b5e228308823487c0c00.jpg", + "table_caption": [], + "table_footnote": [ + "Table 1. Sketch recognition evaluation. Average Top-1 and Top-5 sketch recognition accuracy computed with CLIP zero-shot classifier on 500 sketches from 50 categories. The last row visualizes one sample from each experiment. *Indicates our default settings, which receives the highest accuracy among all models." + ], + "table_body": "
GPT-4oGPT-4o -miniClaude3 OpusClaude3.5 -Sonnet*Claude3.5 -Sonnet (SVG)Human (QD [47])
Top10.15 ±0.040.04 ±0.030.13 ±0.040.23 ±0.050.23 ±0.040.27 ±0.07
Top50.30 ±0.060.10 ±0.040.27 ±0.050.44 ±0.030.43 ±0.060.49 ±0.06
Vis.
", + "bbox": [ + 81, + 89, + 467, + 237 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "stances per category, resulting in 500 sketches in total. Following common practice [105, 106, 115, 116], we utilize a CLIP zero-shot classifier [78] to evaluate how well the generated sketches depict the intended category. We compare the performance of different multimodal LLMs by repeating the same process with GPT-4o-mini [74], GPT-4o [74], and Claude3-Opus [2] as our backbone model (in addition to Claude3.5-Sonnet [2], our default backbone). As a baseline, we include human-drawn sketches sampled from the QuickDraw dataset [47]. The average Top-1 and Top-5 sketch classification accuracy are presented in Table 1. As can be seen, human sketches achieve the highest recognition accuracy, with Claude3.5-Sonnet performing best among all models, approaching human-level rates under the CLIP-score metric. More evaluation of confusion patterns and visualization of the data are provided in the Appendix.", + "bbox": [ + 75, + 325, + 468, + 566 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We additionally compare to prompting Claude3.5-Sonnet to directly generate SVGs using the following prompt: \"Write SVG string that draws a sketch of a . Use only black and white colors\". The corresponding scores are shown in the fifth column of Tab. 1. While this approach achieves recognition scores comparable to those of SketchAgent, the outputs are often characterized by uniform and precise shapes, failing to replicate the fluidity and natural irregularity of free-hand human sketches (e.g., Fig. 3). To evaluate how \"human-like\" our agent's sketches appear, we conduct a two alternative forced choice (2AFC) user study with 150 participants. Each participant was presented with pairs of sketches depicting the same object class produced by different methods, and asked to choose the sketch they believed was human-drawn. 150 sketches across 50 object classes were tested, comparing three methods: direct prompting, SketchAgent, and human sketches from QuickDraw (see Appendix for details). Results indicate SketchAgent's drawings appeared more human-like, being chosen as human-drawn in $74.90 \\pm 3.35\\%$ of cases when compared with direct prompting. When compared to human drawings,", + "bbox": [ + 75, + 568, + 470, + 901 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/423a11037c664387b1a182431eb6f4eedd6120c7e5bf6f031d5dfdee35b62b69.jpg", + "image_caption": [ + "Figure 9. SketchAgent gradually draws stroke-by-stroke, each stroke is annotated by the agent with a semantic meaning." + ], + "image_footnote": [], + "bbox": [ + 504, + 90, + 885, + 181 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/04d463e6f03e007611601410ebbcd7f8b365b30eb1b0c6400a7333a2068fb197.jpg", + "image_caption": [ + "Figure 10. Recognition rate as a function of sketch complexity." + ], + "image_footnote": [], + "bbox": [ + 506, + 217, + 696, + 306 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/07f9d2c145aadce7dd680e4e50f8d2970cc25e1e13d6710c4e467d2cc83efedf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 702, + 217, + 795, + 297 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/969d296403a23d3904cb4e0625589d27622dce0da84e52d44ebd61be07008104.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 800, + 217, + 890, + 297 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "users slightly preferred human sketches $(54.68 \\pm 4.61\\%)$ over SketchAgent's, while direct prompting was chosen only $38.9 \\pm 5.55\\%$ of the time.", + "bbox": [ + 498, + 335, + 890, + 380 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Lastly, to quantitatively analyze the effect of concept complexity, we study the case of drawing letters. We systematically increase sketch complexity by adding letters to the target concept (e.g., from 'ABC' to 'ABCDEFGH') and count the number of correctly recognized letters in each sketch. The graph in Fig. 10 shows that performance decreases as the complexity increases.", + "bbox": [ + 496, + 382, + 890, + 487 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Sequential Sketching", + "text_level": 1, + "bbox": [ + 500, + 500, + 697, + 515 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Figure 9 shows stroke-by-stroke sketch generation by SketchAgent, with the labels on the right indicating the sketching order and the meaning our agent associates with each generated stroke (see Appendix for more examples). Stroke annotation during generation is enabled by utilizing the prior of the backbone LLM, providing a valuable feature for analysis and data collection [37, 65, 107, 126, 128]. In Fig. 11, we illustrate why accounting for the sequential nature of sketching more closely emulates the process of human drawing. We present the sketch creation process of SketchAgent alongside SVGDreamer [116], SketchRNN [42], and a human sketch sampled from QuickDraw [47]. SVGDreamer (first row), is an optimization-based method, where a set of randomly initialized parametric curves (left-most column) are iteratively refined to form a sketch, guided by a pretrained text-to-image diffusion model [82]. This process is time-consuming, taking 2000 iterations (1.6 hours), which makes it unsuitable for interactive sketching. While the final sketch (rightmost column) appears detailed and artistic due to the powerful vision backbone, the intermediate sketching and individual strokes lack clear semantic meaning. In contrast, SketchRNN (second row) is a sequential generative model trained on human-drawn dataset, producing sketches in real-time with strokes added progressively, emulating closer a human-like sketching pro", + "bbox": [ + 496, + 523, + 890, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "23360", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/016f679fa238aefe7749838ef72c26678ee84ede4079312c08d64f12dcb3515f.jpg", + "table_caption": [ + "SVGDreamer [116]. $\\approx 1.6$ hours" + ], + "table_footnote": [], + "table_body": "
SketchRNN [42].≈4 seconds
SketchAgent (Ours).≈20 seconds
bodyheadleft earright earback right legleft eye
Human [47].≤20 seconds
", + "bbox": [ + 89, + 104, + 449, + 387 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Figure 11. Sequential sketching process. SVGDreamer [116] requires 2000 iterations (1.6 hours) with intermediate steps lacking semantic meaning. SketchRNN [42] operates in real-time with coherent steps but is limited to QuickDraw categories. SketchAgent draw gradually with meaningful strokes and no category restrictions. Human sketches evolve through gradual, meaningful steps.", + "bbox": [ + 75, + 390, + 468, + 474 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "cess (as shown in the last row). Similarly, SketchAgent (third row) produces sketches gradually, with each stroke carrying a semantic meaning, by utilizing the sequential nature of its backbone model. While SketchRNN is restricted to generating sketches only within the 345 categories it was trained on, SketchAgent leverages the extensive prior knowledge of its backbone multimodal LLM, enabling it to create sketches of general visual concepts.", + "bbox": [ + 75, + 491, + 468, + 611 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We use the set of 500 samples described in Sec. 5.1 to quantitatively analyze the sequential nature of our agent's sketches compared to human drawings. On the left of Fig. 12, we present histograms comparing the number of strokes in QuickDraw sketches (orange) and our sketches (blue). Most QuickDraw sketches contain 1 to 6 strokes, while our sketches show a broader distribution, peaking between 5 to 10 strokes. This suggests that, on average, QuickDraw sketches appear more abstract. To ensure a balanced comparison of sketches with similar levels of abstraction, we select sketches from both groups with a similar number of strokes (the largest intersection is found in sketches with 4-7 strokes, comprising 204 of our sketches and 120 from QuickDraw) and measure the change in CLIP-Score as a function of the accumulated number of strokes (Fig. 12, right). Both QuickDraw and our sketches exhibit a generally similar pattern, with CLIPScore increasing as more strokes are added, suggesting that sketches become progressively more recognizable as they evolve.", + "bbox": [ + 75, + 613, + 470, + 901 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8bb79ce4851de1d59ede2f9298fd100f1c2e84e32cba837bddaa488f0ba4591f.jpg", + "image_caption": [ + "Figure 12. Sequential sketching analysis of SketchAgent (blue) and Humans [47] (orange). Left: Histograms of stroke distribution per sketch, showing QuickDraw sketches are more abstract on average. Right: CLIPScore as a function of the accumulated number of strokes for sketches containing 4-7 strokes, showing a similar recognition pattern over time." + ], + "image_footnote": [], + "bbox": [ + 514, + 88, + 640, + 220 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/df5f4609b0c441a7467c3969f5a4245622328392296f8f901211d83ac22edf94.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 651, + 89, + 872, + 219 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Human-Agent Collaborative Sketching", + "text_level": 1, + "bbox": [ + 500, + 314, + 834, + 330 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We demonstrate the potential of our system for facilitating interactive human-agent collaboration, resulting in semantically meaningful and recognizable sketches. We design a web-based collaborative sketching environment (Fig. 13A) where users and SketchAgent take turns drawing on a shared canvas to create a recognizable sketch from a given textual concept. Following the evaluation protocol in collabdraw [24], we select 8 simple concepts, based on the agent's demonstrated ability to draw them independently, to focus evaluations on assessing the impact of collaboration. Participants sketched concepts in two modes: solo, where users drew independently, and collab, where users and SketchAgent collaborated, adding one stroke at a time until either was satisfied with the drawing. We collect sketches from 30 participants, resulting in 480 sketches in total. Average CLIP recognition rates are shown in Figure 13B. Collaboratively produced sketches (blue) achieve recognition levels close to those made solely by users and higher than those produced by the agent alone (dashed lines). To assess the contribution of each party in collaborative mode, we analyze partial sketches with only agent-made strokes (pink) or user-made strokes (green), resulting in a significant reduction in recognizability. This suggests that both user and agent contribute meaningfully to the recognizability of the complete sketch.", + "bbox": [ + 496, + 338, + 890, + 715 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.4. Chat-Based Sketch Editing", + "text_level": 1, + "bbox": [ + 500, + 726, + 743, + 742 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We next demonstrate the effectiveness of our method in performing interactive text-based sketch editing within a chat dialogue, where the input to the agent combines both text and images. Inspired by [92], we explore edits that involve spatial reasoning and object relations. We focus on three object categories: outdoor, indoor, and animals, with three objects each, and design editing prompts to add objects to the input sketches. For outdoor and indoor objects, we specify relative locations of added concepts, e.g., \"left to\", \"on top of\" (see Fig. 14 left). For the animals category, we tasked", + "bbox": [ + 496, + 750, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "23361", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1ef282de04d388ff3cea584cc8f62439859b8bb456a0b0f83105927e7b99ad08.jpg", + "image_caption": [ + "(A) Sketching interface" + ], + "image_footnote": [], + "bbox": [ + 101, + 99, + 230, + 191 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9b467c1d65280845bea59527f26ef0b501d740d2b6a7852e9a026da3abbfb9b5.jpg", + "image_caption": [ + "(B) Collaborative user study results" + ], + "image_footnote": [], + "bbox": [ + 246, + 90, + 433, + 203 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ce280e778c673c55a5fab06f36f56a5a8fff409fc90b55e02f168b8aa652fb50.jpg", + "image_caption": [ + "Figure 13. Collaborative sketching evaluation measured using CLIP classification. Sketches created collaboratively (blue) approaching those made solely by users (dashed lines). In collaborative sketches, keeping agent-only strokes (pink) or user-only strokes (green) significantly reduces recognizability.", + "Figure 14. Chat-based sketch editing. We iteratively prompt SketchAgent to add objects to sketches through chat dialogues." + ], + "image_footnote": [], + "bbox": [ + 81, + 303, + 467, + 398 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "the agent with adding accessories to each animal without guidance on their exact placement, testing its ability to infer placement based on semantics (e.g., placing a hat on a head (see Fig. 14 right). The full list of object and editing instructions is provided in the Appendix. We produced a total of 54 sketches. Evaluating the edited sketches reveals that SketchAgent correctly follows instructions $92\\%$ of the time, with $94\\%$ accuracy for specified relations and $88\\%$ accuracy for inferred semantic relations.", + "bbox": [ + 75, + 448, + 468, + 583 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Ablation", + "text_level": 1, + "bbox": [ + 76, + 602, + 174, + 617 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We evaluate the impact of each component of our method by systematically removing them and measuring sketch recognition rates as detailed in 5.1. We assess the effects of removing the system prompt, omitting the CoT process (i.e., excluding thinking tags and 'think step-by-step' instructions), and modifying ICL (the complete sketch example provided in the user prompt). When modifying ICL, we use a correctly formatted single-stroke example instead of the complete sketch, as fully removing ICL results in outputs that do not follow the expected format making them unparsable. The results in Table 2 show that the full SketchAgent pipeline achieves the highest performance, highlighting the importance of each component. Interestingly, not providing a complete sketch example significantly reduces performance. We additionally ablate the impact of the grid resolution, by varying the resolution from 10 to 100 (see Tab. 2, bottom). Extremely low resolutions degrade performance, while mid-level resolutions outperform $100 \\times 100$ .", + "bbox": [ + 75, + 628, + 470, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/85be913d1c16ae445104a386d8de0e85fcdfaaf01f3a2e3ddec914a98d957afe.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
w/o System Promptw/o CoTModified ICLSketchAgent (full)
Top10.20 ± 0.040.14 ± 0.020.07 ± 0.020.23 ± 0.04
Top50.42 ± 0.030.29 ± 0.040.16 ± 0.030.43 ± 0.06
Grid Size10 × 1025 × 2550 × 5075 × 75
Top1 / Top50.14 / 0.280.19 / 0.420.23 / 0.430.23 / 0.41
", + "bbox": [ + 504, + 89, + 890, + 186 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 2. Ablation study. Average Top-1 and Top-5 CLIP recognition accuracy. Top: We systematically remove each component in our pipeline, showcasing all components contribute to the agent's full performance. Bottom: Grid resolution ablation.", + "bbox": [ + 498, + 191, + 892, + 247 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Limitations and Future Work", + "text_level": 1, + "bbox": [ + 498, + 265, + 772, + 280 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "SketchAgent has several limitations. First, it is constrained by the priors of the backbone model, primarily optimized for text rather than visual content. As a result, the agent often produces rich textual descriptions of object parts but struggles to convert these into effective sketching actions, resulting in overly abstract and unrecognizable outputs. For example, in Fig. 15A, the agent effectively describes key parts of a unicorn (e.g., the horn), but the sketch is unrecognizable. This constraint also impacts the depiction of human figures (Fig. 15B). While distinctive features (e.g., Frida Kahlo's eyebrows or Michael Jordan's dunk) may be captured well in language, the resulting sketches are overly simple, with an amateur style, lacking expressivity. We expect this issue to improve as future models advance in vision capabilities. Lastly, the agent may struggle with drawing letters and numbers. This could be improved in future work by providing relevant in-context examples.", + "bbox": [ + 496, + 286, + 892, + 544 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ab03659d7c4d1a51dd5884bfa72ac04901f6940c37db0ebced6a7e7e0884155b.jpg", + "image_caption": [ + "(A) Non-recognizable" + ], + "image_footnote": [], + "bbox": [ + 501, + 569, + 638, + 625 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a5198ad87ecffd7972bcfcf56f48b5c9045d4741a2ec81cdda15348a70a826d5.jpg", + "image_caption": [ + "(B) Human figures", + "Figure 15. Limitations. Sketches of complex concepts (A) and human figures (B) appear too abstract and unrecognizable with non-professional style. (C) Fail to depict letters and numbers." + ], + "image_footnote": [], + "bbox": [ + 645, + 569, + 750, + 628 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ad6c20de7a2d320b68f6ae43926ba1745d7421378648d09adf4f1b97e1df2888.jpg", + "image_caption": [ + "(C) Letters and numbers" + ], + "image_footnote": [], + "bbox": [ + 761, + 569, + 890, + 628 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "8. Conclusions", + "text_level": 1, + "bbox": [ + 500, + 696, + 625, + 713 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We presented a method for language-driven, sequential sketch generation, that can produce versatile sketches in real-time and meaningfully engage in collaborative sketching sessions with humans. We show that the prior knowledge embedded in pretrained multimodal LLMs can be effectively leveraged for sketch generation through an intuitive sketching language and a grid canvas, without requiring additional training or fine-tuning. We hope our work represents a meaningful step toward developing general-purpose sketching systems with the potential to enhance human-computer communication and computer-aided ideation.", + "bbox": [ + 496, + 719, + 890, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "23362", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "9. Acknowledgements", + "text_level": 1, + "bbox": [ + 76, + 90, + 266, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We thank Yuval Alaluf, Hila Chefer, Assaf Ben Kish, Joanna Materzynska, Rinon Gal, Elad Richardson, Arnav Verma, and Ellie Arar for providing feedback on early versions of our manuscript. We are also grateful to Justin Yang, who developed the initial Photodraw application and assisted with the interactive human study setup, and to Dingning Cao, who contributed to the SketchAgent interface and application. Special thanks to Yarden Frenkel for his insights, early explorations, and inspiring discussions. This work was partially supported by NSF CAREER #2047191, NSF DRL #2400471, Stanford Human Centered AI Institute Hoffman-Yee Grant, Hyundai Motor Company, ARL grant W911NF-18-2-021, the Zuckerman STEM Leadership Program, and the Viterbi Fellowship. The sponsors had no role in the experimental design or analysis, the decision to publish, or manuscript preparation. The authors have no competing interests to report.", + "bbox": [ + 75, + 114, + 472, + 372 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 383, + 174, + 400 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katie Millicah, Malcolm Reynolds, Roman Ring, Eliza Rutherford, Serkan Cabi, Tengda Han, Zhitao Gong, Sina Samangooei, Marianne Monteiro, Jacob Menick, Sebastian Borgeaud, Andrew Brock, Aida Nematzadeh, Sahand Sharifzadeh, Mikolaj Binkowski, Ricardo Barreira, Oriol Vinyals, Andrew Zisserman, and Karen Simonyan. Flamingo: a visual language model for few-shot learning. Red Hook, NY, USA, 2024. Curran Associates Inc. 2, 3", + "[2] Anthropic. Claude. https://www.anthropic.com/claude, 2023. 2, 3, 5, 6", + "[3] Itamar Berger, Ariel Shamir, Moshe Mahler, Elizabeth Carter, and Jessica Hodgins. Style and abstraction in portrait sketching. ACM Trans. Graph., 32(4), 2013. 2", + "[4] James Betker, Gabriel Goh, Li Jing, † TimBrooks, Jianfeng Wang, Linjie Li, † LongOuyang, † JuntangZhuang, † JoyceLee, † YufeiGuo, † WesamManassra, † PrafullaDhariwal, † CaseyChu, † YunxinJiao, and Aditya Ramesh. Improving image generation with better captions. 2023. 3", + "[5] Ayan Kumar Bhunia, Ayan Das, Umar Riaz Muhammad, Yongxin Yang, Timothy M. Hospedales, Tao Xiang, Yulia Gryaditskaya, and Yi-Zhe Song. Pixel: a competitive sketching ai agent. so you think you can sketch? ACM Trans. Graph., 39:166:1-166:15, 2020. 1, 3", + "[6] Ankan Kumar Bhunia, Salman Khan, Hisham Cholakkal, Rao Muhammad Anwer, Fahad Shahbaz Khan, Jorma Laaksonen, and Michael Felsberg. Doodleformer: Creative sketch drawing with transformers. ECCV, 2022. 1, 2, 3", + "[7] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel Ziegler, Jeffrey Wu, Clemens Winter, Chris Hesse, Mark Chen, Eric" + ], + "bbox": [ + 94, + 410, + 468, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. Language models are few-shot learners. In Advances in Neural Information Processing Systems, pages 1877-1901. Curran Associates, Inc., 2020. 2, 3, 5", + "[8] Sébastien Bubeck, Varun Chandrasekaran, Ronen Eldan, Johannes Gehrke, Eric Horvitz, Ece Kamar, Peter Lee, Yin Tat Lee, Yuanzhi Li, Scott Lundberg, et al. Sparks of artificial general intelligence: Early experiments with gpt-4. arXiv preprint arXiv:2303.12712, 2023. 3", + "[9] Mu Cai, Zeyi Huang, Yuheng Li, Haohan Wang, and Yong Jae Lee. Delving into LLMs' visual understanding ability using SVG to bridge image and text, 2024. 2, 3", + "[10] John Canny. A computational approach to edge detection. IEEE Transactions on pattern analysis and machine intelligence, (6):679-698, 1986. 2", + "[11] Alexandre Carlier, Martin Danelljan, Alexandre Alahi, and Radu Timofte. Deepsvg: A hierarchical generative network for vector graphics animation, 2020. 3", + "[12] Caroline Chan, Frédo Durand, and Phillip Isola. Learning to generate line drawings that convey geometry and semantics. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7915-7925, 2022. 2", + "[13] Zhenfang Chen, Qinhong Zhou, Yikang Shen, Yining Hong, Zhiqing Sun, Dan Gutfreund, and Chuang Gan. Visual chain-of-thought prompting for knowledge-based visual reasoning. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1254-1262, 2024. 3", + "[14] Changwoon Choi, Jaeah Lee, Jaesik Park, and Young Min Kim. 3doodle: Compact abstraction of objects with 3d strokes. ACM Trans. Graph., 43(4), 2024. 3", + "[15] Xiangxiang Chu, Jianlin Su, Bo Zhang, and Chunhua Shen. Visionllama: A unified llama backbone for vision tasks, 2024. 2, 3", + "[16] Ayan Das, Yongxin Yang, Timothy Hospedales, Tao Xiang, and Yi-Zhe Song. Béziersketch: A generative model for scalable vector sketches. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXVI 16, pages 632-647. Springer, 2020. 1", + "[17] Nicholas Davis, Chih-PIn Hsiao, Kunwar Yashraj Singh, Lisa Li, Sanat Moningi, and Brian Magerko. Drawing apprentice: An enactive co-creative agent for artistic collaboration. In Proceedings of the 2015 ACM SIGCHI Conference on Creativity and Cognition, page 185–186, New York, NY, USA, 2015. Association for Computing Machinery. 3", + "[18] Vanessa de Andrade, Sofia Freire, Monica Baptista, and Yael Shwartz. Drawing as a space for social-cognitive interaction. Education Sciences, 12:45, 2022. 3", + "[19] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language" + ], + "bbox": [ + 509, + 92, + 890, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "23363", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Technologies, Volume 1 (Long and Short Papers), pages 4171-4186, Minneapolis, Minnesota, 2019. Association for Computational Linguistics. 3", + "[20] Sivan Doveh, Shaked Perek, M Jehanzeb Mirza, Wei Lin, Amit Alfassy, Assaf Arbelle, Shimon Ullman, and Leonid Karlinsky. Towards multimodal in-context learning for vision & language models. arXiv preprint arXiv:2403.12736, 2024. 3", + "[21] Mathias Eitz, James Hays, and Marc Alexa. How do humans sketch objects? ACM Trans. Graph., 31(4), 2012. 2", + "[22] Ziv Epstein, Aaron Hertzmann, Laura Mariah Herman, Robert Mahari, Morgan R. Frank, Matthew Groh, Hope Schroeder, Amy Smith, Memo Akten, Jessica Fjeld, Hany Farid, Neil Leach, Alex Pentland, and Olga Russakovsky. Art and the science of generative ai. Science, 380:1110-1111, 2023. 1", + "[23] Judy Fan, Wilma A. Bainbridge, Rebecca Chamberlain, and Jeffrey D. Wammes. Drawing as a versatile cognitive tool. Nature Reviews Psychology, 2:556 - 568, 2023. 2", + "[24] Judith E. Fan, Monica Dinculescu, and David Ha. collaborate: An environment for collaborative sketching with an artificial agent. In Proceedings of the 2019 Conference on Creativity and Cognition, page 556-561, New York, NY, USA, 2019. Association for Computing Machinery. 3, 7", + "[25] Judith E. Fan, Robert D. Hawkins, Mike Wu, and Noah D. Goodman. Pragmatic Inference and Visual Abstraction Enable Contextual Flexibility During Visual Communication. Computational Brain & Behavior, 3(1):86-101, 2020. 3", + "[26] Judith E Fan, Wilma A Bainbridge, Rebecca Chamberlain, and Jeffrey D Wammes. Drawing as a versatile cognitive tool. Nature Reviews Psychology, 2(9):556-568, 2023. 1", + "[27] Logan Fiorella and Shelbi Kuhlmann. Creating drawings enhances learning by teaching. Journal of Educational Psychology, 112(4):811, 2020. 1", + "[28] Kenneth Forbus, Jeffrey Usher, Andrew Lovett, Kate Lockwood, and Jon Wetzel. Cogsketch: Sketch understanding for cognitive science research and for education. Topics in Cognitive Science, 3(4):648-666, 2011. 1", + "[29] Kevin Frans, L. B. Soros, and Olaf Witkowski. Clipdraw: exploring text-to-drawing synthesis through language-image encoders. In Proceedings of the 36th International Conference on Neural Information Processing Systems, Red Hook, NY, USA, 2024. Curran Associates Inc. 1, 3", + "[30] Xingyu Fu, Yushi Hu, Bangzheng Li, Yu Feng, Haoyu Wang, Xudong Lin, Dan Roth, Noah A Smith, Wei-Chiu Ma, and Ranjay Krishna. Blink: Multimodal large language models can see but not perceive. arXiv preprint arXiv:2404.12390, 2024. 4", + "[31] Rinon Gal, Yael Vinker, Yuval Alaluf, Amit Bermano, Daniel Cohen-Or, Ariel Shamir, and Gal Chechik. Breathing life into sketches using text-to-video priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4325–4336, 2024. 3", + "[32] Yaroslav Ganin, Tejas D. Kulkarni, Igor Babuschkin, S. M. Ali Eslami, and Oriol Vinyals. Synthesizing programs for images using reinforced adversarial learning. *ArXiv*, abs/1804.01118, 2018. 3" + ], + "bbox": [ + 86, + 92, + 470, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[33] Yaroslav Ganin, Sergey Bartunov, Yujia Li, Ethan Keller, and Stefano Saliceti. Computer-aided design as language. In Neural Information Processing Systems, 2021. 3", + "[34] Chengying Gao, Qi Liu, Qi Xu, Limin Wang, Jianzhuang Liu, and Changqing Zou. Sketchycoco: Image generation from freehand scene sketches. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5174-5183, 2020. 2", + "[35] Simon Garrod, Nicolas Fay, John Lee, Jon Oberlander, and Tracy MacLeod. Foundations of representation: Where might graphical symbol systems come from? Cognitive Science, 31(6):961-987, 2007. 3", + "[36] Songwei Ge, Vedanuj Goswami, C Lawrence Zitnick, and Devi Parikh. Creative sketch generation. arXiv preprint arXiv:2011.10039, 2020. 1", + "[37] Songwei Ge, Vedanuj Goswami, Larry Zitnick, and Devi Parikh. Creative sketch generation. In International Conference on Learning Representations, 2021. 2, 6", + "[38] Hinnie Gijlers, Armin Weinberger, Alieke Mattia van Dijk, Lars Bollen, and Wouter van Joolingen. Collaborative drawing on a shared digital canvas in elementary science education: The effects of script and task awareness support. International Journal of Computer-Supported Collaborative Learning, 8(4):427-453, 2013. 3", + "[39] Gabriela Goldschmidt. Serial sketching: visual problem solving in designing. Cybernetics and System, 23(2):191-219, 1992. 1", + "[40] Yulia Gryaditskaya, Mark Sypesteyn, Jan Willem Hoftijzer, Sylvia Pont, Frédo Durand, and Adrien Bousseau. Opensketch: a richly-annotated dataset of product design sketches. ACM Trans. Graph., 38(6), 2019. 2", + "[41] Tianrui Guan, Fuxiao Liu, Xiyang Wu, Ruiqi Xian, Zongxia Li, Xiaoyu Liu, Xijun Wang, Lichang Chen, Furong Huang, Yaser Yacoob, Dinesh Manocha, and Tianyi Zhou. Hallusionbench: An advanced diagnostic suite for entangled language hallucination and visual illusion in large vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14375-14385, 2024. 2", + "[42] David Ha and Douglas Eck. A neural representation of sketch drawings. CoRR, abs/1704.03477, 2017. 1, 2, 3, 6, 7", + "[43] Yucheng Han, China. Xiaoyan Zhang, Xin Chen, Xu Yang, Zhibin Wang, Gang Yu, Bin Fu, and Hanwang Zhang. Chartlama: A multimodal llm for chart understanding and generation. ArXiv, abs/2311.16483, 2023. 2, 3", + "[44] Yushi Hu, Weijia Shi, Xingyu Fu, Dan Roth, Mari Ostendorf, Luke Zettlemoyer, Noah A Smith, and Ranjay Krishna. Visual sketchpad: Sketching as a visual chain of thought for multimodal language models. arXiv preprint arXiv:2406.09403, 2024. 3", + "[45] Francisco Javier Ibarrola, Tomas Lawton, and Kazjon Grace. A collaborative, interactive and context-aware drawing agent for co-creative design. IEEE Transactions on Visualization and Computer Graphics, 30:5525-5537, 2022. 3", + "[46] Ajay Jain, Amber Xie, and Pieter Abbeel. Vectorfusion: Text-to-svg by abstracting pixel-based diffusion models. In" + ], + "bbox": [ + 509, + 92, + 890, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "23364", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1911-1920, 2023. 1, 3", + "[47] Jongejan Jonas, Rowley Henry, Kawashima Takashi, Kim Jongmin, and Fox-Gieg Nick. The Quick, Draw! - A.I. Experiment, 2016. 2, 3, 5, 6, 7", + "[48] David Kaiser. Drawing theories apart: The dispersion of Feynman diagrams in postwar physics. University of Chicago Press, 2019. 1", + "[49] Pegah Karimi, Jeba Rezwana, Safat Siddiqui, Mary Lou Maher, and Nasrin Dehbozorgi. Creative sketching partner: an analysis of human-ai co-creativity. In Proceedings of the 25th International Conference on Intelligent User Interfaces, page 221–230, New York, NY, USA, 2020. Association for Computing Machinery. 3", + "[50] Pegah Karimi, Jeba Rezwana, Safat Siddiqui, Mary Lou Maher, and Nasrin Dehbozorgi. Creative sketching partner: an analysis of human-ai co-creativity. In Proceedings of the 25th International Conference on Intelligent User Interfaces, page 221–230, New York, NY, USA, 2020. Association for Computing Machinery. 3", + "[51] Andrew Kyle Lampinen, Stephanie C. Y. Chan, Aaditya K. Singh, and Murray Shanahan. The broader spectrum of in-context learning, 2024. 2", + "[52] Paul Laseau. Graphic thinking for architects and designers. John Wiley & Sons, 2000. 2", + "[53] Tomas Lawton, Kazjon Grace, and Francisco J Ibarrola. When is a tool a tool? user perceptions of system agency in human-ai co-creative drawing. In Proceedings of the 2023 ACM Designing Interactive Systems Conference, page 1978–1996, New York, NY, USA, 2023. Association for Computing Machinery. 3", + "[54] Tomas Lawton, Francisco J Ibarrola, Dan Ventura, and Kazjon Grace. Drawing with reframer: Emergence and control in co-creative ai. In Proceedings of the 28th International Conference on Intelligent User Interfaces, page 264–277, New York, NY, USA, 2023. Association for Computing Machinery. 3", + "[55] Yong Jae Lee, C. Lawrence Zitnick, and Michael F. Cohen. Shadowdraw: real-time user guidance for freehand drawing. In ACM SIGGRAPH 2011 Papers, New York, NY, USA, 2011. Association for Computing Machinery. 3", + "[56] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. BLIP: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In Proceedings of the 39th International Conference on Machine Learning, pages 12888-12900. PMLR, 2022. 2, 3", + "[57] Mengtian Li, Zhe Lin, Radomir Mech, Ersin Yumer, and Deva Ramanan. Photo-sketching: Inferring contour drawings from images. In 2019 IEEE Winter Conference on Applications of Computer Vision (WACV), pages 1403-1412. IEEE, 2019. 2", + "[58] Tzu-Mao Li, Michal Lukáč, Gharbi Michael, and Jonathan Ragan-Kelley. Differentiable vector graphics rasterization for editing and learning. ACM Trans. Graph. (Proc. SIGGRAPH Asia), 39(6):193:1-193:15, 2020. 1, 3" + ], + "bbox": [ + 86, + 90, + 470, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[59] Yi Li, Yi-Zhe Song, Timothy M. Hospedales, and Shaogang Gong. Free-hand sketch synthesis with deformable stroke models. CoRR, abs/1510.02644, 2015. 1, 2", + "[60] Yijun Li, Chen Fang, Aaron Hertzmann, Eli Shechtman, and Ming-Hsuan Yang. Im2pencil: Controllable pencil illustration from photographs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1525-1534, 2019. 2", + "[61] Hangyu Lin, Yanwei Fu, Yu-Gang Jiang, and X. Xue. Sketch-bert: Learning sketch bidirectional encoder representation from transformers by self-supervised learning of sketch gestalt. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6757-6766, 2020. 3", + "[62] Difan Liu, Matthew Fisher, Aaron Hertzmann, and Evangelos Kalogerakis. Neural strokes: Stylized line drawing of 3d shapes. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14204-14213, 2021. 2", + "[63] Fang Liu, Xiaoming Deng, Yu-Kun Lai, Yong-Jin Liu, Cuixia Ma, and Hongan Wang. Sketchgan: Joint sketch completion and recognition with generative adversarial network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2", + "[64] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In Advances in Neural Information Processing Systems, pages 34892-34916. Curran Associates, Inc., 2023. 2, 3", + "[65] Bria Long, Judith Fan, Holly Huey, Zixian Chai, and Michael Frank. Parallel developmental changes in children's production and recognition of line drawings of visual concepts. Nature Communications, 15, 2024. 6", + "[66] Arjun Majumdar, Anurag Ajay, Xiaohan Zhang, Pranav Putta, Sriram Yenamandra, Mikael Henaff, Sneha Silwal, Paul Mcvay, Oleksandr Maksymets, Sergio Arnaud, Karmesh Yadav, Qiyang Li, Ben Newman, Mohit Sharma, Vincent Berges, Shiqi Zhang, Pulkit Agrawal, Yonatan Bisk, Dhruv Batra, Mrinal Kalakrishnan, Franziska Meier, Chris Paxton, Sasha Sax, and Aravind Rajeswaran. Openeqa: Embodied question answering in the era of foundation models. In Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 4", + "[67] William P. McCarthy, Justin Matejka, Karl D.D. Willis, Judith E. Fan, and Yewen Pu. Communicating design intent using drawing and text. In Proceedings of the 16th Conference on Creativity & Cognition, page 512-519, New York, NY, USA, 2024. Association for Computing Machinery. 3", + "[68] John FJ Mellor, Eunbyung Park, Yaroslav Ganin, Igor Babuschkin, Tejas Kulkarni, Dan Rosenbaum, Andy Ballard, Theophane Weber, Oriol Vinyals, and SM Eslami. Unsupervised doodling and painting with improved spiral. arXiv preprint arXiv:1910.01007, 2019. 3", + "[69] Daniela Mihai and Jonathon Hare. Learning to draw: Emergent communication through sketching. Advances in Neural Information Processing Systems, 34:7153-7166, 2021. 3" + ], + "bbox": [ + 509, + 92, + 890, + 898 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "23365", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[70] Chancharik Mitra, Brandon Huang, Trevor Darrell, and Roei Herzig. Compositional chain-of-thought prompting for large multimodal models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14420-14431, 2024. 3", + "[71] Kushin Mukherjee, Holly Huey, Xuanchen Lu, Yael Vinker, Rio Aguina-Kang, Ariel Shamir, and Judith Fan. Seva: Leveraging sketches to evaluate alignment between human and machine visual abstraction. In Advances in Neural Information Processing Systems, 2023. 2", + "[72] Omar W Nasim. Observing by hand: sketching the nebulae in the nineteenth century. University of Chicago Press, 2019. 1", + "[73] Changhoon Oh, Jungwoo Song, Jinhan Choi, Seonghyeon Kim, Sungwoo Lee, and Bongwon Suh. I lead, you help but only with enough details: Understanding user experience of co-creation with artificial intelligence. In Proceedings of the 2018 CHI Conference on Human Factors in Computing Systems, page 1–13, New York, NY, USA, 2018. Association for Computing Machinery. 3", + "[74] OpenAI. Gpt-4 technical report, 2024. 2, 3, 4, 6", + "[75] Dustin Podell, Zion English, Kyle Lacey, A. Blattmann, Tim Dockhorn, Jonas Muller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. ArXiv, abs/2307.01952, 2023. 2, 3", + "[76] Yonggang Qi, Guoyao Su, Pinaki Nath Chowdhury, Mingkang Li, and Yi-Zhe Song. Sketchlattice: Latticed representation for sketch manipulation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 953–961, 2021. 2", + "[77] Shuwen Qiu, Sirui Xie, Lifeng Fan, Tao Gao, Song-Chun Zhu, and Yixin Zhu. Emergent graphical conventions in a visual communication game. arXiv preprint arXiv:2111.14210, 2021. 3", + "[78] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. CoRR, abs/2103.00020, 2021. 1, 3, 6", + "[79] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. J. Mach. Learn. Res., 21(1), 2020. 3", + "[80] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 2, 3", + "[81] Leo Sampaio Ferraz Ribeiro, Tu Bui, John P. Collomosse, and Moacir Antonelli Ponti. Sketchformer: Transformer-based representation for sketched structure. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14141-14150, 2020. 3", + "[82] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models, 2022. 1, 2, 3, 6" + ], + "bbox": [ + 86, + 92, + 470, + 900 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[83] Daniel Rose, Vaishnavi Himakunthala, Andy Ouyang, Ryan He, Alex Mei, Yujie Lu, Michael Saxon, Chinmay Sonar, Diba Mirza, and William Yang Wang. Visual chain of thought: bridging logical gaps with multimodal infillings. arXiv preprint arXiv:2305.02317, 2023. 3", + "[84] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S. Sara Mahdavi, Rapha Gontijo Lopes, Tim Salimans, Jonathan Ho, David J Fleet, and Mohammad Norouzi. Photorealistic text-to-image diffusion models with deep language understanding, 2022. 3", + "[85] Patsorn Sangkloy, Nathan Burnell, Cusuh Ham, and James Hays. The sketchy database: Learning to retrieve badly drawn bunnies. ACM Trans. Graph., 35(4), 2016. 2", + "[86] Peter Schaldenbrand, James McCann, and Jean Oh. Frida: A collaborative robot painter with a differentiable, real2sim2real planning environment. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 11712-11718, 2023. 3", + "[87] Peter Schaldenbrand, Gaurav Parmar, Jun-Yan Zhu, James McCann, and Jean Oh. Cofrida: Self-supervised fine-tuning for human-robot co-painting. In 2024 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2024. 3", + "[88] Donald A Schon and Vincent DeSanctis. The reflective practitioner: How professionals think in action, 1986. 2", + "[89] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, Patrick Schramowski, Srivatsa Kundurthy, Katherine Crowson, Ludwig Schmidt, Robert Kaczmarczyk, and Jenia Jitsev. Laion-5b: An open large-scale dataset for training next generation image-text models, 2022. 3", + "[90] Tamar Rott Shaham, Sarah Schwettmann, Franklin Wang, Achyuta Rajaram, Evan Hernandez, Jacob Andreas, and Antonio Torralba. A multimodal automated interpretability agent. In *Forty-first International Conference on Machine Learning*, 2024. 3", + "[91] Hao Shao, Shengju Qian, Han Xiao, Guanglu Song, Zhuofan Zong, Letian Wang, Yu Liu, and Hongsheng Li. Visual cot: Advancing multi-modal language models with a comprehensive dataset and benchmark for chain-of-thought reasoning. arXiv preprint arXiv:2403.16999, 2024. 3", + "[92] Pratyusha Sharma, Tamar Rott Shaham, Manel Baradad, Stephanie Fu, Adrian Rodriguez-Munoz, Shivam Duggal, Phillip Isola, and Antonio Torralba. A vision check-up for language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14410-14419, 2024. 3, 7", + "[93] Jifei Song, Kaiyue Pang, Yi-Zhe Song, Tao Xiang, and Timothy Hospedales. Learning to sketch with shortcut cycle consistency, 2018. 2", + "[94] Guoyao Su, Yonggang Qi, Kaiyue Pang, Jie Yang, Yi-Zhe Song, and CVSSP SketchX. Sketchhealer: A graph-to-sequence network for recreating partial human sketches. In BMVC, page 5, 2020. 2" + ], + "bbox": [ + 509, + 92, + 890, + 900 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "23366", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[95] Yanpeng Sun, Qiang Chen, Jian Wang, Jingdong Wang, and Zechao Li. Exploring effective factors for improving visual in-context learning. arXiv preprint arXiv:2304.04748, 2023.3", + "[96] Ivan E. Sutherland. Sketchpad—a man-machine graphical communication system, page 391-408. Association for Computing Machinery, New York, NY, USA, 1998. 3", + "[97] Gemini Team. Gemini: A family of highly capable multimodal models, 2024. 2, 3", + "[98] Jakob Tholander and Martin Jonsson. Design ideation with ai - sketching, thinking and talking with generative machine learning models. In Proceedings of the 2023 ACM Design-Interactive Systems Conference, page 1930-1940, New York, NY, USA, 2023. Association for Computing Machinery. 1", + "[99] Shengbang Tong, Zhuang Liu, Yuexiang Zhai, Yi Ma, Yann LeCun, and Saining Xie. Eyes wide shut? exploring the visual shortcomings of multimodal llms, 2024. 4", + "[100] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee Lacroix, Baptiste Roziere, Naman Goyal, Eric Hambro, Faisal Azhar, Aurelien Rodriguez, Armand Joulin, Edouard Grave, and Guillaume Lample. Llama: Open and efficient foundation language models. ArXiv, abs/2302.13971, 2023. 3", + "[101] Barbara Tversky. What do sketches say about thinking? AAAI Spring Symp. Sketch Understanding Worksh., 2002. 2", + "[102] Barbara Tversky. Visualizing thought. In Handbook of human centric visualization, pages 3-40. Springer, 2013. 1", + "[103] Barbara Tversky, Masaki Suwa, Maneesh Agrawala, Julie Heiser, Chris Stolte, Pat Hanrahan, Doantam Phan, Jeff Klingner, Marie-Paule Daniel, Paul Lee, et al. Sketches for design and design of sketches. Human Behaviour in Design: Individuals, Teams, Tools, pages 79-86, 2003. 1", + "[104] Russell Tytler, Vaughan Prain, George Aranda, Joseph Ferguson, and Radhika Gorur. Drawing to reason and learn in science. Journal of Research in Science Teaching, 57(2): 209-231, 2020. 3", + "[105] Yael Vinker, Ehsan Pajouheshgar, Jessica Y. Bo, Roman Christian Bachmann, Amit Haim Bermano, Daniel Cohen-Or, Amir Zamir, and Ariel Shamir. Clipasso: Semantically-aware object sketching. ACM Trans. Graph., 41(4), 2022. 1, 3, 6", + "[106] Yael Vinker, Yuval Alaluf, Daniel Cohen-Or, and Ariel Shamir. Clipascene: Scene sketching with different types and levels of abstraction. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 4146–4156, 2023. 3, 6", + "[107] Jiawei Wang and Changjian Li. Contextseg: Sketch semantic segmentation by querying the context with attention. arXiv preprint arXiv:2311.16682, 2023. 6", + "[108] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Brian Ichter, Fei Xia, Ed H. Chi, Quoc V. Le, and Denny Zhou. Chain-of-thought prompting elicits reasoning in large language models. In Proceedings of the 36th International Conference on Neural Information Processing Systems, Red Hook, NY, USA, 2022. Curran Associates Inc. 2, 5" + ], + "bbox": [ + 78, + 90, + 470, + 898 + ], + "page_idx": 12 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[109] Holger Winnemöller, Jan Eric Kyprianidis, and Sven C. Olsen. Xdog: An extended difference-of-gaussians compendium including advanced image stylization. Comput. Graph., 36:740-753, 2012. 2", + "[110] World Wide Web Consortium (W3C). Scalable Vector Graphics (SVG), 1999. 3", + "[111] Chenfei Wu, Shengming Yin, Weizhen Qi, Xiaodong Wang, Zecheng Tang, and Nan Duan. Visual chatgpt: Talking, drawing and editing with visual foundation models. arXiv preprint arXiv:2303.04671, 2023. 3", + "[112] Rong Wu, Wanchao Su, Kede Ma, and Jing Liao. Iconshop: Text-guided vector icon synthesis with autoregressive transformers. ACM Transactions on Graphics (TOG), 42:1-14, 2023. 3", + "[113] Chufeng Xiao, Wanchao Su, Jing Liao, Zhouhui Lian, Yi-Zhe Song, and Hongbo Fu. Diffsketching: How differently do people sketch 3d objects? ACM Transactions on Graphics (Proceedings of ACM SIGGRAPH Asia 2022), 41 (4):1-16, 2022. 2", + "[114] Saining Xie and Zhuowen Tu. Holistically-nested edge detection. In Proceedings of the IEEE international conference on computer vision, pages 1395-1403, 2015. 2", + "[115] XiMing Xing, Chuang Wang, Haitao Zhou, Jing Zhang, Qian Yu, and Dong Xu. Diffsketcher: Text guided vector sketch synthesis through latent diffusion models. In Advances in Neural Information Processing Systems, pages 15869-15889. Curran Associates, Inc., 2023. 3, 6", + "[116] Ximing Xing, Haitao Zhou, Chuang Wang, Jing Zhang, Dong Xu, and Qian Yu. Svgdreamer: Text guided svg generation with diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4546-4555, 2024. 3, 6, 7", + "[117] Peng Xu, Timothy M. Hospedales, Qiyue Yin, Yi-Zhe Song, Tao Xiang, and Liang Wang. Deep learning for free-hand sketch: A survey and a toolbox, 2020. 2", + "[118] Zhengyuan Yang, Linjie Li, Kevin Lin, Jianfeng Wang, Chung-Ching Lin, Zicheng Liu, and Lijuan Wang. The dawn of Imms: Preliminary explorations with gpt-4v(ision). ArXiv, abs/2309.17421, 2023. 2", + "[119] Zhengyuan Yang, Jianfeng Wang, Linjie Li, Kevin Lin, Chung-Ching Lin, Zicheng Liu, and Lijuan Wang. Idea2img: Iterative self-refinement with gpt-4v (ision) for automatic image design and generation. arXiv preprint arXiv:2310.08541, 2023. 3", + "[120] Ran Yi, Yong-Jin Liu, Yu-Kun Lai, and Paul L Rosin. Apdrawinggan: Generating artistic portrait drawings from face photos with hierarchical gans. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10743-10752, 2019. 2", + "[121] Ran Yi, Yong-Jin Liu, Yu-Kun Lai, and Paul L Rosin. Unpaired portrait drawing generation via asymmetric cycle mapping. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8217-8225, 2020. 2", + "[122] C. Zhang, Weijie Wang, Paul Pangaro, Nikolas Martelaro, and Daragh Byrne. Generative image ai using design sketches as input: Opportunities and challenges. Proceed-" + ], + "bbox": [ + 501, + 90, + 890, + 901 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "23367", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 12 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ings of the 15th Conference on Creativity and Cognition, 2023. 1", + "[123] Jiahao Zhang, Bowen Wang, Liangzhi Li, Yuta Nakashima, and Hajime Nagahara. Instruct me more! random prompting for visual in-context learning. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2597-2606, 2024. 3", + "[124] Peiying Zhang, Nanxuan Zhao, and Jing Liao. Text-to-vector generation with neural path representation. ACM Trans. Graph., 43(4), 2024. 3", + "[125] Yuanhan Zhang, Kaiyang Zhou, and Ziwei Liu. What makes good examples for visual in-context learning? Advances in Neural Information Processing Systems, 36: 17773-17794, 2023. 3", + "[126] Zhengming Zhang, Xiaoming Deng, Jinyao Li, Yukun Lai, Cuixia Ma, Yongjin Liu, and Hongan Wang. Stroke-based semantic segmentation for scene-level free-hand sketches. Vis. Comput., 39(12):6309-6321, 2022. 6", + "[127] Zhuosheng Zhang, Aston Zhang, Mu Li, Hai Zhao, George Karypis, and Alex Smola. Multimodal chain-of-thought reasoning in language models. arXiv preprint arXiv:2302.00923, 2023. 3", + "[128] Yixiao Zheng, Kaiyue Pang, Ayan Das, Dongliang Chang, Yi-Zhe Song, and Zhanyu Ma. Creativeseg: Semantic segmentation of creative sketches. IEEE Transactions on Image Processing, 33:2266-2278, 2024. 6", + "[129] Tao Zhou, Chen Fang, Zhaowen Wang, Jimei Yang, Byungmoon Kim, Zhili Chen, Jonathan Brandt, and Demetri Terzopoulos. Learning to sketch with deep q networks and demonstrated strokes. ArXiv, abs/1810.05977, 2018. 2, 3" + ], + "bbox": [ + 78, + 90, + 470, + 515 + ], + "page_idx": 13 + }, + { + "type": "page_number", + "text": "23368", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 13 + } +] \ No newline at end of file diff --git a/2025/SketchAgent_ Language-Driven Sequential Sketch Generation/76b8d485-0aca-43aa-b3b2-9c94133cc069_model.json b/2025/SketchAgent_ Language-Driven Sequential Sketch Generation/76b8d485-0aca-43aa-b3b2-9c94133cc069_model.json new file mode 100644 index 0000000000000000000000000000000000000000..7c2956ef9c504328ac8951b203dccf04fac27eb7 --- /dev/null +++ b/2025/SketchAgent_ Language-Driven Sequential Sketch Generation/76b8d485-0aca-43aa-b3b2-9c94133cc069_model.json @@ -0,0 +1,3517 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.17, + 0.095, + 0.8, + 0.117 + ], + "angle": 0, + "content": "SketchAgent: Language-Driven Sequential Sketch Generation" + }, + { + "type": "text", + "bbox": [ + 0.085, + 0.133, + 0.884, + 0.154 + ], + "angle": 0, + "content": "Yael Vinker\\(^{1}\\) Tamar Rott Shaham\\(^{1}\\) Kristine Zheng\\(^{2}\\) Alex Zhao\\(^{1}\\) Judith E Fan\\(^{2}\\) Antonio Torralba\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.316, + 0.169, + 0.364, + 0.186 + ], + "angle": 0, + "content": "\\(^{1}\\)MIT" + }, + { + "type": "text", + "bbox": [ + 0.141, + 0.19, + 0.538, + 0.206 + ], + "angle": 0, + "content": "{yaelvink,tamarott,alexzhao,torralba}@mit.edu" + }, + { + "type": "text", + "bbox": [ + 0.622, + 0.169, + 0.789, + 0.188 + ], + "angle": 0, + "content": "\\(^{2}\\)Stanford University" + }, + { + "type": "text", + "bbox": [ + 0.584, + 0.19, + 0.83, + 0.206 + ], + "angle": 0, + "content": "{jefan, kxzheng} \\(@\\) stanford.edu" + }, + { + "type": "text", + "bbox": [ + 0.327, + 0.216, + 0.64, + 0.23 + ], + "angle": 0, + "content": "https://sketch-agent.csail.mit.edu/" + }, + { + "type": "image", + "bbox": [ + 0.111, + 0.249, + 0.329, + 0.463 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.394, + 0.25, + 0.58, + 0.284 + ], + "angle": 0, + "content": "Human-Agent Collaborative Sketching" + }, + { + "type": "image", + "bbox": [ + 0.371, + 0.291, + 0.62, + 0.457 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.642, + 0.249, + 0.848, + 0.268 + ], + "angle": 0, + "content": "Chat-Based Sketch Editing" + }, + { + "type": "image", + "bbox": [ + 0.648, + 0.272, + 0.722, + 0.327 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.76, + 0.275, + 0.835, + 0.286 + ], + "angle": 0, + "content": "Draw a robot" + }, + { + "type": "text", + "bbox": [ + 0.738, + 0.322, + 0.835, + 0.346 + ], + "angle": 0, + "content": "Add a paintbrush and a canvas" + }, + { + "type": "image", + "bbox": [ + 0.649, + 0.336, + 0.721, + 0.389 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.743, + 0.385, + 0.835, + 0.41 + ], + "angle": 0, + "content": "Draw a robot on the canvas" + }, + { + "type": "image", + "bbox": [ + 0.649, + 0.398, + 0.722, + 0.453 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.472, + 0.893, + 0.501 + ], + "angle": 0, + "content": "Figure 1. SketchAgent leverages an off-the-shelf multimodal LLM to facilitate language-driven, sequential sketch generation through an intuitive sketching language. It can sketch diverse concepts, engage in interactive sketching with humans, and edit content via chat." + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.513, + 0.314, + 0.529 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.552, + 0.473, + 0.884 + ], + "angle": 0, + "content": "Sketching serves as a versatile tool for externalizing ideas, enabling rapid exploration and visual communication that spans various disciplines. While artificial systems have driven substantial advances in content creation and human-computer interaction, capturing the dynamic and abstract nature of human sketching remains challenging. In this work, we introduce SketchAgent, a language-driven, sequential sketch generation method that enables users to create, modify, and refine sketches through dynamic, conversational interactions. Our approach requires no training or fine-tuning. Instead, we leverage the sequential nature and rich prior knowledge of off-the-shelf multimodal large language models (LLMs). We present an intuitive sketching language, introduced to the model through in-context examples, enabling it to \"draw\" using string-based actions. These are processed into vector graphics and then rendered to create a sketch on a pixel canvas, which can be accessed again for further tasks. By drawing stroke by stroke, our agent captures the evolving, dynamic qualities intrinsic to sketching. We demonstrate that SketchAgent can generate sketches from diverse prompts, engage in dialogue-driven drawing, and collaborate meaningfully with human users." + }, + { + "type": "title", + "bbox": [ + 0.502, + 0.513, + 0.632, + 0.529 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.538, + 0.892, + 0.719 + ], + "angle": 0, + "content": "Sketching is a powerful tool for distilling ideas into their simplest form. Its fluid and spontaneous nature makes sketching a uniquely versatile tool for visualization, rapid ideation, and communication across cultures, generations, and disciplines [26, 102]. For example, designers use sketches to explore new ideas [39, 103], scientists employ them to formulate problems [48, 72], and children engage in sketching to learn and express themselves [27, 28] (see Fig. 2). Artificial systems, in principle, have the potential to support and enhance human creativity, problem-solving, and visual expression through sketching, adapting flexibly to their exploratory nature [22, 98, 122]." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.893, + 0.901 + ], + "angle": 0, + "content": "Traditionally, sketch generation methods rely on human-drawn datasets to train generative models [5, 6, 16, 36, 42, 59]. However, fully capturing the diversity of sketches within datasets remains challenging [26], limiting these methods in both scale and diversity. Recent advancements in vision-language models, such as CLIP [78] and text-to-image diffusion [82], have enabled sketch generation methods that reduce reliance on human-drawn datasets [29, 46, 105]. These methods leverage pretrained model guidance and differentiable rendering [58] to optimize parametric curves, creating sketches that go beyond predefined styles and categories." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23355" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.258 + ], + "angle": 0, + "content": "While representing a significant step toward a general-purpose sketching system, these methods lack a crucial aspect of human drawing: the process itself. Current methods, though versatile, optimize all strokes simultaneously, making the intermediate sketching steps meaningless. As a result, the sketch cannot be decomposed into a coherent sequence of strokes that reflects the drawing process. In contrast, humans draw iteratively, stroke by stroke, incorporating visual feedback and continuously adapting—a dynamic, evolving process that fosters creativity, ideation, and communication [52, 88, 101]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.261, + 0.473, + 0.398 + ], + "angle": 0, + "content": "In this work, we introduce SketchAgent, a sketch generation agent that leverages the prior knowledge and sequential nature of multimodal large language models (LLMs) to enable versatile, progressive, language-driven sketching. Our agent can generate sketches across a wide range of textual concepts—from animals to engineering principles (Fig. 1, left). Its sequential nature facilitates interactive human-agent sketching and supports iterative refinement and editing through a chat-based dialogue (Fig. 1, right)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.399, + 0.473, + 0.627 + ], + "angle": 0, + "content": "Unlike vision-language models that directly generate images from text [75, 80, 82], multimodal LLMs [1, 2, 15, 56, 64, 74, 97] accept text and images as input but only output text. To produce visuals, they either utilize external \"tools\" (such as calling a text-to-image model) or are prompted to generate executable code (e.g., Python [43], SVG [9]) to create charts, diagrams, or graphics. However, prompting for such representations to directly produce sketches often results in a mechanical appearance with uniform, precise shapes that lack the subtle irregularities and spontaneous qualities characteristic of human sketches (see Fig. 3B). Additionally, despite their robustness in textual tasks, these models often struggle with fine-grained spatial reasoning [41, 118] as they are primarily optimized for text, making sketch editing more challenging." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.629, + 0.473, + 0.902 + ], + "angle": 0, + "content": "To address these limitations, we introduce an intuitive sketching language that enables an off-the-shelf multimodal LLM agent to \"draw\" sketches on a canvas by providing string-based actions, without additional training or finetuning. We define the canvas as a numbered grid, allowing the agent to reference specific coordinates (e.g., x2y8) to enhance its spatial reasoning capabilities. We represent a sketch as a sequence of semantically meaningful strokes, each defined by a series of such coordinates. We leverage In-Context Learning (ICL) [7, 51] to introduce the agent to the new representation, and Chain of Thought (CoT) [108] to enhance its planning capabilities. Given a sketching task, the agent produces a textual response following our representation, which we process by fitting a smooth Bézier curve to each coordinate sequence. The curves are then rendered onto the canvas to form the final sketch. We find this approach useful in emulating a more natural sketch appearance. For collaborative sketching, the canvas remains acces" + }, + { + "type": "image_caption", + "bbox": [ + 0.532, + 0.093, + 0.693, + 0.116 + ], + "angle": 0, + "content": "(A) Frank Gehry. *Elevation Sketches*. Guggenheim Museum" + }, + { + "type": "image", + "bbox": [ + 0.532, + 0.117, + 0.727, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.532, + 0.223, + 0.68, + 0.245 + ], + "angle": 0, + "content": "(B) Alexander Bell's drawing of the telephone" + }, + { + "type": "image", + "bbox": [ + 0.532, + 0.246, + 0.727, + 0.324 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.731, + 0.093, + 0.838, + 0.106 + ], + "angle": 0, + "content": "(C) Children drawing" + }, + { + "type": "image", + "bbox": [ + 0.731, + 0.108, + 0.86, + 0.179 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.731, + 0.184, + 0.848, + 0.208 + ], + "angle": 0, + "content": "(D) Basketball strategy (pick and roll)" + }, + { + "type": "image", + "bbox": [ + 0.731, + 0.21, + 0.862, + 0.326 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.33, + 0.895, + 0.414 + ], + "angle": 0, + "content": "Figure 2. Examples of sketches used across disciplines and goals. (A) Ideation and design: Process Elevation Sketches by the architect Frank Gehry, Guggenheim Museum. (B) Engineering: Alexander Bell's telephone drawing. (C) Expressing emotions: Children's sketches. (D) Visual communication: Planning and communicating game strategy in basketball." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.425, + 0.892, + 0.515 + ], + "angle": 0, + "content": "sible to both the user and the agent throughout the session. The agent generates strokes sequentially and pauses according to an adjustable stopping token, allowing the user to add their own strokes directly to the canvas. These strokes are then integrated into the agent's sequence, enabling it to continue drawing, with real-time canvas updates." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.516, + 0.895, + 0.651 + ], + "angle": 0, + "content": "We demonstrate SketchAgent's capability to generate sketches of diverse concepts while capturing the inherently sequential and dynamic nature of sketching. We showcase our agent's ability to collaborate effectively with humans in real time to create novel and meaningful sketches. Our method is the first to leverage pretrained multimodal LLMs for sequential sketching without additional training, paving the way for a general-purpose artificial sketching system that supports iterative, evolving interactivity." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.665, + 0.642, + 0.68 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.689, + 0.895, + 0.901 + ], + "angle": 0, + "content": "Sketch Generation Early methods approached sketch generation by designing image filters to simulate sketchlike effects [10, 109]. With the advent of deep learning, data-driven approaches emerged to address a range of sketch-related tasks [117], including category-conditioned sketching [42, 76, 93], object sketching [59, 62], scenesketching [12, 57, 60, 114], sketch completion [6, 63, 94], portrait drawing [3, 120, 121], part-based generation [6, 37, 42, 129], and more. While sketch data collection has been broadly explored [21, 34, 40, 71, 85, 113], the wide variation in sketch styles and their adaptation to specific tasks [23] makes collecting datasets that encompass this diversity challenging. For example, QuickDraw [47], the largest available sketch dataset with 50 mil" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23356" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.091, + 0.473, + 0.381 + ], + "angle": 0, + "content": "lion sketches, covers only 345 object categories and primarily focuses on simple, iconic representations. This limits data-driven methods to the style, abstraction level, and concepts seen during training. Recently, large pretrained vision-language models [75, 78, 80, 82, 84] have shown remarkable text-to-image generation capabilities by leveraging extensive visual knowledge from billions of training images [89]. While these models can be prompted to generate sketch-like images (see Fig. 3A), they do so in a single step and in pixel space, lacking the sequential, stroke-based process of human sketching. Subsequent approaches [14, 29, 31, 46, 105, 106, 115, 116, 124] leverage the priors of these models to guide an iterative optimization of parametric curves, with a differentiable rasterizer [58] linking pixel and vector representations. While producing vector sketches, the final strokes lack order and semantic meaning, and the optimization-based approach overlook the sequential aspect of the sketching process, making these methods suboptimal for collaborative sketching." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.39, + 0.473, + 0.709 + ], + "angle": 0, + "content": "Sequential and Collaborative Sketching Collaborative human-machine sketching holds promise in enhancing creativity, ideation, communication, and learning, as explored in various fields, including human-computer interaction (HCI) [17, 45, 49, 50, 53, 54], computer graphics [55, 96], robotics [86, 87], cognitive science [24, 25, 35, 67], learning sciences [18, 38, 104], and more. Central to collaborative sketching is its sequential, adaptive, and dynamic process, with each action carrying intent. Existing methods employ diverse training strategies to account for the discrete nature of sequential sketches, including reinforcement and adversarial learning [32, 68, 129], multi-agent referential games [69, 77], transformers [5, 6, 11, 33, 61, 81, 112], and more. SketchRNN [42] is a pioneering work in this area, introducing the QuickDraw dataset [47], a crowd-sourced collection of real-time sketch sequences made by users. They utilize this dataset to train a recurrent neural network for sequential sketch generation, which was later shown [24, 73] to have potential for human-machine collaboration. However, this approach remains constrained by the predefined categories encountered during training." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.72, + 0.473, + 0.902 + ], + "angle": 0, + "content": "Multimodel LLMs for Content Creation LLMs [7, 19, 79, 100] and multimodal LLMs [1, 2, 15, 56, 64, 74, 97] receive text as input (or text and images for multimodal) and output text. To enable visual content generation, these models are often paired with external \"tools\" that extend their functionality [44, 90, 111, 119]. For example, ChatGPT [74] generates images by internally calling a separate model, DALLE-3 [4]. Another approach involves prompting models to produce code in languages like Python [43], Processing [92], SVG [9], or TikZ [8] that can be rendered into visuals such as graphs, charts, and vector graphics. However, such code-generated content often looks" + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.091, + 0.599, + 0.163 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.507, + 0.172, + 0.599, + 0.185 + ], + "angle": 0, + "content": "(A) DALLE3 [4]" + }, + { + "type": "image", + "bbox": [ + 0.614, + 0.108, + 0.691, + 0.141 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.615, + 0.166, + 0.689, + 0.192 + ], + "angle": 0, + "content": "(B) LLMs [2] (SVG)" + }, + { + "type": "image", + "bbox": [ + 0.716, + 0.109, + 0.785, + 0.141 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.706, + 0.173, + 0.796, + 0.185 + ], + "angle": 0, + "content": "(C) SketchAgent" + }, + { + "type": "image", + "bbox": [ + 0.815, + 0.115, + 0.88, + 0.144 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.806, + 0.173, + 0.893, + 0.185 + ], + "angle": 0, + "content": "(D) Human [47]" + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.195, + 0.895, + 0.28 + ], + "angle": 0, + "content": "Figure 3. Sketch appearance. (A) Text-to-image diffusion models operate in pixel space, lacking the sequential nature of sketches. (B) Prompting LLMs to produce visuals with SVG results in a uniform, mechanical appearance. (C) Sketches produced by our agent appear less mechanical, more closely resembling the nature of (D) Human sketches, which are often spontaneous and irregular." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.303, + 0.895, + 0.486 + ], + "angle": 0, + "content": "rigid, with uniform and overly precise shapes that lack the subtle irregularities and spontaneous qualities characteristic of freehand sketches (see Fig. 3B). In contrast, we propose a sketching language grounded in spatial information that encourages the model to produce a more natural sketch appearance, which we then process into vector graphics. Common strategies for enhancing LLMs capabilities include Chain-of-Thought prompting [13, 70, 83, 91, 127], which breaks down tasks into smaller, logical steps to mimic human reasoning, and In-Context Learning (ICL) [7, 20, 95, 123, 125], where examples of input-output pairs are provided to help the model infer task patterns." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.506, + 0.64, + 0.523 + ], + "angle": 0, + "content": "3. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.534, + 0.74, + 0.64 + ], + "angle": 0, + "content": "Vector Graphics and Bezier Curves Vector graphics allow us to create visual images directly from geometric shapes such as points, lines, curves, and polygons. Unlike raster images (represented with pixels), vector graphics are" + }, + { + "type": "image", + "bbox": [ + 0.776, + 0.503, + 0.887, + 0.596 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.771, + 0.6, + 0.892, + 0.627 + ], + "angle": 0, + "content": "Figure 4. Cubic Bezier curve." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.64, + 0.895, + 0.821 + ], + "angle": 0, + "content": "resolution-free, more compact, and editable. SVG [110] is an XML-based format for storing vector graphics, popular for its scalability and compatibility with modern web browsers. The process of transferring vector graphics into pixel images is called rasterization or rendering. Cubic Bézier curves are commonly used to represent sketches in vector graphics. A cubic Bézier curve (Fig. 4) is a smooth parametric curve defined by four points: a start point \\( P_0 \\), an end point \\( P_3 \\), and two control points \\( P_1 \\) and \\( P_2 \\) that shape the curvature. The set \\( P = \\{P_0, P_1, P_2, P_3\\} \\) is often referred to as the curve's control points. The curve is described by the following polynomial equation:" + }, + { + "type": "equation", + "bbox": [ + 0.509, + 0.836, + 0.892, + 0.855 + ], + "angle": 0, + "content": "\\[\nB (t) = (1 - t) ^ {3} P _ {0} + 3 (1 - t) ^ {2} t P _ {1} + 3 (1 - t) t ^ {2} P _ {2} + t ^ {3} P _ {3}, (1)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.892, + 0.901 + ], + "angle": 0, + "content": "where \\( t \\in [0,1] \\) is a parameter that moves the point along the curve from \\( P_0 \\) at \\( t = 0 \\) to \\( P_3 \\) at \\( t = 1 \\)." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "23357" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.089, + 0.889, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.247, + 0.896, + 0.304 + ], + "angle": 0, + "content": "Figure 5. Method Overview. SketchAgent (blue) receives drawing instructions and generates a string representing the intended sketch. Inputs include: (1) a system prompt (orange) introducing the sketching language and canvas, (2) a user prompt (pink) specifying the task (e.g., \"draw a shark\"), and (3) a numbered canvas. The agent's response outlines a sketching strategy (in thinking tags) and a sequence of strokes defined by coordinates, which are processed into Bézier curves and rendered onto the canvas." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.321, + 0.169, + 0.337 + ], + "angle": 0, + "content": "4. Method" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.348, + 0.474, + 0.606 + ], + "angle": 0, + "content": "Our goal is to enable an off-the-shelf pretrained multimodal LLM to draw sketches based on natural language instructions. An overview of our pipeline is illustrated in Fig. 5. We utilize a frozen multimodal LLM (\"SketchAgent\" shown in blue), which receives three inputs: (1) a system prompt containing guidelines for using our new sketching language, (2) a user prompt with additional task-specific instructions (e.g., \"Draw a shark\"), and (3) a blank canvas on which the agent can draw. Based on the given task, the agent generates a textual response, representing the sequence of strokes to be drawn, which we then process into vector graphics and render onto the canvas. The canvas can then be reused in two ways: it can be fed back into the model with an updated user prompt for additional tasks and editing, or it can be accessed by a human user who can draw directly on it to facilitate collaborative sketching. Next, we describe each component of the pipeline." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.614, + 0.473, + 0.903 + ], + "angle": 0, + "content": "The Canvas Although multimodal LLMs demonstrate remarkable reasoning abilities, they often struggle with spatial reasoning tasks [30, 66, 99]. We present a simple example (see Fig. 6) to illustrate how this limitation affects the naive use of these models for sketch generation and interactive sketching. We provide GPT-4o [74] with an image depicting a simple line drawing of a partial house featuring five numbered points (from 1 to 5), and ask it to identify which points should be connected to complete the house. While the model correctly identifies the pair of points, it fails to select the correct pixel coordinates when given a basic draw-line tool that connects two points, even after multiple attempts. To enhance the model's spatial reasoning ability, we utilize a numbered canvas that forms a grid. This grid features numbers (1 to 50) along the x-axis and the y-axis (Fig. 5, left). Each cell is uniquely identified by a combination of the corresponding x-axis and y-axis numbers (e.g., the bottom-left cell is \\(x1y1\\)). The agent interacts with the canvas by specifying desired \\((x,y)\\) coordinates." + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.32, + 0.885, + 0.395 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.399, + 0.893, + 0.469 + ], + "angle": 0, + "content": "Figure 6. Although excelling in visual reasoning, multimodal LLMs often struggle to translate these abilities into spatial actions. In this example, GPT-4o [74] intends to draw a line between points 1 and 5 but fails to execute this with a draw_line function that accepts pixel coordinates." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.489, + 0.892, + 0.566 + ], + "angle": 0, + "content": "Sketch Representation We define a sketch as a sequence of \\(n\\) ordered strokes \\(S = \\{S_1, S_2, \\ldots, S_n\\}\\). Each stroke \\(S_i\\) is defined by a sequence of \\(m\\) cell coordinates on the grid: \\(S_i = \\{(x_j, y_j)\\}_{j=1}^m\\), represented in string format as: x1y1, x15y20, ...." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.567, + 0.893, + 0.748 + ], + "angle": 0, + "content": "A naive approach to processing the textual sequence of coordinates would be to use a polyline, connecting consecutive points with line segments. However, our grid-based representation sparsifies the canvas, resulting in a non-smooth and unnatural appearance when using polylines (see Fig. 7, left). To achieve a smoother appearance, an alternative approach is to treat the coordinates as a sequence of control points defining smooth curves. However, as illustrated in Fig. 4, the control points often do not lie directly on the curve. Consequently, if the agent aims for a stroke that passes through specific coordinates, it must derive the control points that define this stroke, which is challenging." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.895, + 0.904 + ], + "angle": 0, + "content": "We propose an alternative approach: we treat the specified \\((x, y)\\) coordinates as a set of desired points sampled along the curve, and fit a smooth Bézier curve to them (Fig. 7, right). To accommodate curves with complex curvature, we also task the model with determining when each point on the curve should be passed through, corresponding to the \\(t\\) value described in Eq. (1). Thus, for each stroke \\(S_{i}\\), the agent provides a set of \\(m\\) sampled points \\(S_{i} = \\{(x_{j}, y_{j})\\}_{j=1}^{m}\\), along with a corresponding set of \\(t\\) values: \\(T_{i} = \\{t_{j}\\}_{j=1}^{m}\\). Based on these, we fit a cubic Bézier" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23358" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.119, + 0.089, + 0.218, + 0.145 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.136, + 0.149, + 0.208, + 0.162 + ], + "angle": 0, + "content": "(A) Polyline" + }, + { + "type": "image", + "bbox": [ + 0.222, + 0.089, + 0.323, + 0.145 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.221, + 0.149, + 0.329, + 0.162 + ], + "angle": 0, + "content": "(B) Control Points" + }, + { + "type": "image", + "bbox": [ + 0.326, + 0.089, + 0.426, + 0.145 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.349, + 0.149, + 0.402, + 0.162 + ], + "angle": 0, + "content": "(C) Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.166, + 0.47, + 0.236 + ], + "angle": 0, + "content": "Figure 7. Methods for processing the agent's coordinate sequence (in red): (A) Polyline results in an unnatural appearance. (B) Directly using coordinates as Bézier control points is challenging as they do not lie on the curve. (C) Fitting a Bézier curve to sampled coordinates provides smoother results." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.254, + 0.47, + 0.3 + ], + "angle": 0, + "content": "curve to the sampled points by solving a system of linear equations using least squares, where the unknowns are the control points \\( P = \\{P_0, P_1, P_2, P_3\\} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.185, + 0.306, + 0.469, + 0.323 + ], + "angle": 0, + "content": "\\[\nP = \\operatorname {a r g m i n} _ {P} \\| A P - B \\|, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.329, + 0.47, + 0.48 + ], + "angle": 0, + "content": "where \\(A \\in \\mathbb{R}^{m \\times 4}\\) contains the cubic Bézier basis functions evaluated at specific \\(t_j\\) values (as described in Eq. (1)), and \\(B \\in \\mathbb{R}^{m \\times 2}\\) contains the \\(m\\) sampled points \\(\\{(x_j, y_j)\\}_{j=1}^m\\). The least squares solution minimizes the error between the fitted Bézier curve and the sampled points. For long sequences resulting in a large fitting error, we recursively split the curve. Additionally, we account for Bézier curves of lower degrees, including quadratic curves, linear lines, and points. Upon completing this process, we render the parametric curves onto the canvas." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.492, + 0.47, + 0.822 + ], + "angle": 0, + "content": "Drawing Instructions We provide the model with a system prompt and a user prompt (marked in orange and pink in Fig. 5). In the system prompt, we supply the agent with context about its expertise (\"You are an expert artist specializing in drawing sketches\") and introduce it to the grid canvas along with examples of how to use our sketching language for drawing single-stroke primitives (full prompts are provided in the Appendix). The system prompt is fixed and can be applied to a variety of sketching tasks. The user prompt includes a description of the desired task and an example of a simple sketch of a house drawn with our sketching language. We find this to be crucial in assisting the agent with preserving the correct format that could be parsed directly [7]. The agent is tasked with responding in the format shown in the gray text box in Fig. 5. In the tags, the agent is tasked to outline the overall sketching strategy [108]. This typically includes describing the different components of the sketch, the intended sketching order, and the overall placement of each part. The agent is also tasked with providing an ID tag following each stroke, which is useful for further analysis and for producing annotated sketches in scale." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.833, + 0.459, + 0.849 + ], + "angle": 0, + "content": "4.1. In-Chat Editing and Collaborative Sketching" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.469, + 0.901 + ], + "angle": 0, + "content": "The above process can be repeated iteratively to support multiple sketching tasks and interactions. Text-based sketch editing in a chat dialogue is enabled by feeding the rendered" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.096, + 0.603, + 0.155 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.515, + 0.162, + 0.593, + 0.19 + ], + "angle": 0, + "content": "Golden Gate Bridge" + }, + { + "type": "image", + "bbox": [ + 0.612, + 0.095, + 0.689, + 0.144 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.626, + 0.162, + 0.671, + 0.19 + ], + "angle": 0, + "content": "Mount Fuji" + }, + { + "type": "image", + "bbox": [ + 0.715, + 0.092, + 0.775, + 0.15 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.725, + 0.162, + 0.764, + 0.19 + ], + "angle": 0, + "content": "Eiffel Tower" + }, + { + "type": "image", + "bbox": [ + 0.824, + 0.096, + 0.857, + 0.154 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.806, + 0.162, + 0.874, + 0.189 + ], + "angle": 0, + "content": "DNA Double Helix" + }, + { + "type": "image", + "bbox": [ + 0.529, + 0.195, + 0.578, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.523, + 0.263, + 0.585, + 0.29 + ], + "angle": 0, + "content": "Pendulum Motion" + }, + { + "type": "image", + "bbox": [ + 0.616, + 0.203, + 0.681, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.613, + 0.263, + 0.685, + 0.291 + ], + "angle": 0, + "content": "Double-Slit Experiment" + }, + { + "type": "image", + "bbox": [ + 0.714, + 0.193, + 0.777, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.713, + 0.27, + 0.776, + 0.283 + ], + "angle": 0, + "content": "Flowchart" + }, + { + "type": "image", + "bbox": [ + 0.811, + 0.202, + 0.869, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.817, + 0.264, + 0.862, + 0.289 + ], + "angle": 0, + "content": "Tic-tac toe" + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.296, + 0.892, + 0.338 + ], + "angle": 0, + "content": "Figure 8. Sketches produced by SketchAgent for concepts beyond pre-defined categories. The textual input describing the desired concept shown below each image." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.352, + 0.892, + 0.593 + ], + "angle": 0, + "content": "canvas back to the agent (see dashed arrow in Fig. 5) and updating the user prompt with the desired edits. To support collaborative human-agent sketching, the canvas remains accessible to both the human user and the agent throughout the entire sketching session. We define an adjustable stopping token, \\( \\langle \\text{}/\\text{s}\\{\\text{j}\\} \\rangle \\), which instructs the agent to pause generating the sequence at stroke number \\( j \\). We then process and render the generated strokes onto the canvas up to that point, then the user can add strokes directly to the canvas to continue the sketch. The user-drawn strokes are processed and converted into the agent's format by reversing our fitting process, i.e., sampling each stroke at multiple \\( t \\) values (as shown in Eq. (1)), and selecting the points closest to each cell's center on the grid. The converted user strokes are then chained with the agent's sequence, after which the agent resumes sketching until the next stopping token." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.608, + 0.587, + 0.623 + ], + "angle": 0, + "content": "5. Results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.629, + 0.892, + 0.719 + ], + "angle": 0, + "content": "We evaluate the performance of our method qualitatively and quantitatively across a selected set of sketching tasks. Additional tasks, evaluations, and examples are provided in the Appendix. All results presented in the paper were generated using Claude3.5-Sonnet [2] as our backbone model, unless stated otherwise." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.731, + 0.818, + 0.745 + ], + "angle": 0, + "content": "5.1. Text-Conditioned Sketch Generation" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Figures 1 and 8 demonstrate SketchAgent's capability to generate sketches of various concepts that extend beyond standard categories, which includes scientific concepts (e.g., \"the double-slit experiment\", \"pendulum motion\"), diagrams (e.g., \"circuit diagram\", \"a flowchart\"), and notable landmarks (e.g., \"Taj Mahal\", \"Eiffel Tower\"). More examples are provided in the Appendix. To quantitatively evaluate text-conditioned generation we utilize the QuickDraw dataset [47]. We randomly sample 50 categories (out of 345), and apply our method to generate 10 sketch in" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23359" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.082, + 0.09, + 0.468, + 0.238 + ], + "angle": 0, + "content": "
GPT-4oGPT-4o -miniClaude3 OpusClaude3.5 -Sonnet*Claude3.5 -Sonnet (SVG)Human (QD [47])
Top10.15 ±0.040.04 ±0.030.13 ±0.040.23 ±0.050.23 ±0.040.27 ±0.07
Top50.30 ±0.060.10 ±0.040.27 ±0.050.44 ±0.030.43 ±0.060.49 ±0.06
Vis.
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.077, + 0.24, + 0.47, + 0.31 + ], + "angle": 0, + "content": "Table 1. Sketch recognition evaluation. Average Top-1 and Top-5 sketch recognition accuracy computed with CLIP zero-shot classifier on 500 sketches from 50 categories. The last row visualizes one sample from each experiment. *Indicates our default settings, which receives the highest accuracy among all models." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.326, + 0.47, + 0.567 + ], + "angle": 0, + "content": "stances per category, resulting in 500 sketches in total. Following common practice [105, 106, 115, 116], we utilize a CLIP zero-shot classifier [78] to evaluate how well the generated sketches depict the intended category. We compare the performance of different multimodal LLMs by repeating the same process with GPT-4o-mini [74], GPT-4o [74], and Claude3-Opus [2] as our backbone model (in addition to Claude3.5-Sonnet [2], our default backbone). As a baseline, we include human-drawn sketches sampled from the QuickDraw dataset [47]. The average Top-1 and Top-5 sketch classification accuracy are presented in Table 1. As can be seen, human sketches achieve the highest recognition accuracy, with Claude3.5-Sonnet performing best among all models, approaching human-level rates under the CLIP-score metric. More evaluation of confusion patterns and visualization of the data are provided in the Appendix." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.569, + 0.471, + 0.902 + ], + "angle": 0, + "content": "We additionally compare to prompting Claude3.5-Sonnet to directly generate SVGs using the following prompt: \"Write SVG string that draws a sketch of a . Use only black and white colors\". The corresponding scores are shown in the fifth column of Tab. 1. While this approach achieves recognition scores comparable to those of SketchAgent, the outputs are often characterized by uniform and precise shapes, failing to replicate the fluidity and natural irregularity of free-hand human sketches (e.g., Fig. 3). To evaluate how \"human-like\" our agent's sketches appear, we conduct a two alternative forced choice (2AFC) user study with 150 participants. Each participant was presented with pairs of sketches depicting the same object class produced by different methods, and asked to choose the sketch they believed was human-drawn. 150 sketches across 50 object classes were tested, comparing three methods: direct prompting, SketchAgent, and human sketches from QuickDraw (see Appendix for details). Results indicate SketchAgent's drawings appeared more human-like, being chosen as human-drawn in \\(74.90 \\pm 3.35\\%\\) of cases when compared with direct prompting. When compared to human drawings," + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.092, + 0.887, + 0.182 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.183, + 0.892, + 0.212 + ], + "angle": 0, + "content": "Figure 9. SketchAgent gradually draws stroke-by-stroke, each stroke is annotated by the agent with a semantic meaning." + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.218, + 0.697, + 0.308 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.704, + 0.218, + 0.796, + 0.299 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.801, + 0.218, + 0.892, + 0.299 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.509, + 0.311, + 0.882, + 0.325 + ], + "angle": 0, + "content": "Figure 10. Recognition rate as a function of sketch complexity." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.337, + 0.892, + 0.381 + ], + "angle": 0, + "content": "users slightly preferred human sketches \\((54.68 \\pm 4.61\\%)\\) over SketchAgent's, while direct prompting was chosen only \\(38.9 \\pm 5.55\\%\\) of the time." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.383, + 0.892, + 0.488 + ], + "angle": 0, + "content": "Lastly, to quantitatively analyze the effect of concept complexity, we study the case of drawing letters. We systematically increase sketch complexity by adding letters to the target concept (e.g., from 'ABC' to 'ABCDEFGH') and count the number of correctly recognized letters in each sketch. The graph in Fig. 10 shows that performance decreases as the complexity increases." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.5, + 0.698, + 0.516 + ], + "angle": 0, + "content": "5.2. Sequential Sketching" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.524, + 0.892, + 0.902 + ], + "angle": 0, + "content": "Figure 9 shows stroke-by-stroke sketch generation by SketchAgent, with the labels on the right indicating the sketching order and the meaning our agent associates with each generated stroke (see Appendix for more examples). Stroke annotation during generation is enabled by utilizing the prior of the backbone LLM, providing a valuable feature for analysis and data collection [37, 65, 107, 126, 128]. In Fig. 11, we illustrate why accounting for the sequential nature of sketching more closely emulates the process of human drawing. We present the sketch creation process of SketchAgent alongside SVGDreamer [116], SketchRNN [42], and a human sketch sampled from QuickDraw [47]. SVGDreamer (first row), is an optimization-based method, where a set of randomly initialized parametric curves (left-most column) are iteratively refined to form a sketch, guided by a pretrained text-to-image diffusion model [82]. This process is time-consuming, taking 2000 iterations (1.6 hours), which makes it unsuitable for interactive sketching. While the final sketch (rightmost column) appears detailed and artistic due to the powerful vision backbone, the intermediate sketching and individual strokes lack clear semantic meaning. In contrast, SketchRNN (second row) is a sequential generative model trained on human-drawn dataset, producing sketches in real-time with strokes added progressively, emulating closer a human-like sketching pro" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23360" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.171, + 0.093, + 0.37, + 0.105 + ], + "angle": 0, + "content": "SVGDreamer [116]. \\(\\approx 1.6\\) hours" + }, + { + "type": "table", + "bbox": [ + 0.091, + 0.106, + 0.45, + 0.388 + ], + "angle": 0, + "content": "
SketchRNN [42].≈4 seconds
SketchAgent (Ours).≈20 seconds
bodyheadleft earright earback right legleft eye
Human [47].≤20 seconds
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.391, + 0.47, + 0.476 + ], + "angle": 0, + "content": "Figure 11. Sequential sketching process. SVGDreamer [116] requires 2000 iterations (1.6 hours) with intermediate steps lacking semantic meaning. SketchRNN [42] operates in real-time with coherent steps but is limited to QuickDraw categories. SketchAgent draw gradually with meaningful strokes and no category restrictions. Human sketches evolve through gradual, meaningful steps." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.492, + 0.47, + 0.612 + ], + "angle": 0, + "content": "cess (as shown in the last row). Similarly, SketchAgent (third row) produces sketches gradually, with each stroke carrying a semantic meaning, by utilizing the sequential nature of its backbone model. While SketchRNN is restricted to generating sketches only within the 345 categories it was trained on, SketchAgent leverages the extensive prior knowledge of its backbone multimodal LLM, enabling it to create sketches of general visual concepts." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.614, + 0.471, + 0.902 + ], + "angle": 0, + "content": "We use the set of 500 samples described in Sec. 5.1 to quantitatively analyze the sequential nature of our agent's sketches compared to human drawings. On the left of Fig. 12, we present histograms comparing the number of strokes in QuickDraw sketches (orange) and our sketches (blue). Most QuickDraw sketches contain 1 to 6 strokes, while our sketches show a broader distribution, peaking between 5 to 10 strokes. This suggests that, on average, QuickDraw sketches appear more abstract. To ensure a balanced comparison of sketches with similar levels of abstraction, we select sketches from both groups with a similar number of strokes (the largest intersection is found in sketches with 4-7 strokes, comprising 204 of our sketches and 120 from QuickDraw) and measure the change in CLIP-Score as a function of the accumulated number of strokes (Fig. 12, right). Both QuickDraw and our sketches exhibit a generally similar pattern, with CLIPScore increasing as more strokes are added, suggesting that sketches become progressively more recognizable as they evolve." + }, + { + "type": "image", + "bbox": [ + 0.515, + 0.089, + 0.641, + 0.221 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.652, + 0.09, + 0.874, + 0.22 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.224, + 0.892, + 0.307 + ], + "angle": 0, + "content": "Figure 12. Sequential sketching analysis of SketchAgent (blue) and Humans [47] (orange). Left: Histograms of stroke distribution per sketch, showing QuickDraw sketches are more abstract on average. Right: CLIPScore as a function of the accumulated number of strokes for sketches containing 4-7 strokes, showing a similar recognition pattern over time." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.315, + 0.836, + 0.332 + ], + "angle": 0, + "content": "5.3. Human-Agent Collaborative Sketching" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.339, + 0.892, + 0.717 + ], + "angle": 0, + "content": "We demonstrate the potential of our system for facilitating interactive human-agent collaboration, resulting in semantically meaningful and recognizable sketches. We design a web-based collaborative sketching environment (Fig. 13A) where users and SketchAgent take turns drawing on a shared canvas to create a recognizable sketch from a given textual concept. Following the evaluation protocol in collabdraw [24], we select 8 simple concepts, based on the agent's demonstrated ability to draw them independently, to focus evaluations on assessing the impact of collaboration. Participants sketched concepts in two modes: solo, where users drew independently, and collab, where users and SketchAgent collaborated, adding one stroke at a time until either was satisfied with the drawing. We collect sketches from 30 participants, resulting in 480 sketches in total. Average CLIP recognition rates are shown in Figure 13B. Collaboratively produced sketches (blue) achieve recognition levels close to those made solely by users and higher than those produced by the agent alone (dashed lines). To assess the contribution of each party in collaborative mode, we analyze partial sketches with only agent-made strokes (pink) or user-made strokes (green), resulting in a significant reduction in recognizability. This suggests that both user and agent contribute meaningfully to the recognizability of the complete sketch." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.727, + 0.744, + 0.743 + ], + "angle": 0, + "content": "5.4. Chat-Based Sketch Editing" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.892, + 0.901 + ], + "angle": 0, + "content": "We next demonstrate the effectiveness of our method in performing interactive text-based sketch editing within a chat dialogue, where the input to the agent combines both text and images. Inspired by [92], we explore edits that involve spatial reasoning and object relations. We focus on three object categories: outdoor, indoor, and animals, with three objects each, and design editing prompts to add objects to the input sketches. For outdoor and indoor objects, we specify relative locations of added concepts, e.g., \"left to\", \"on top of\" (see Fig. 14 left). For the animals category, we tasked" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "23361" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.102, + 0.101, + 0.232, + 0.193 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.104, + 0.204, + 0.23, + 0.216 + ], + "angle": 0, + "content": "(A) Sketching interface" + }, + { + "type": "image", + "bbox": [ + 0.247, + 0.091, + 0.434, + 0.204 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.249, + 0.205, + 0.436, + 0.216 + ], + "angle": 0, + "content": "(B) Collaborative user study results" + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.22, + 0.47, + 0.29 + ], + "angle": 0, + "content": "Figure 13. Collaborative sketching evaluation measured using CLIP classification. Sketches created collaboratively (blue) approaching those made solely by users (dashed lines). In collaborative sketches, keeping agent-only strokes (pink) or user-only strokes (green) significantly reduces recognizability." + }, + { + "type": "image", + "bbox": [ + 0.082, + 0.304, + 0.468, + 0.399 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.4, + 0.469, + 0.427 + ], + "angle": 0, + "content": "Figure 14. Chat-based sketch editing. We iteratively prompt SketchAgent to add objects to sketches through chat dialogues." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.449, + 0.47, + 0.584 + ], + "angle": 0, + "content": "the agent with adding accessories to each animal without guidance on their exact placement, testing its ability to infer placement based on semantics (e.g., placing a hat on a head (see Fig. 14 right). The full list of object and editing instructions is provided in the Appendix. We produced a total of 54 sketches. Evaluating the edited sketches reveals that SketchAgent correctly follows instructions \\(92\\%\\) of the time, with \\(94\\%\\) accuracy for specified relations and \\(88\\%\\) accuracy for inferred semantic relations." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.603, + 0.176, + 0.618 + ], + "angle": 0, + "content": "6. Ablation" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.629, + 0.471, + 0.901 + ], + "angle": 0, + "content": "We evaluate the impact of each component of our method by systematically removing them and measuring sketch recognition rates as detailed in 5.1. We assess the effects of removing the system prompt, omitting the CoT process (i.e., excluding thinking tags and 'think step-by-step' instructions), and modifying ICL (the complete sketch example provided in the user prompt). When modifying ICL, we use a correctly formatted single-stroke example instead of the complete sketch, as fully removing ICL results in outputs that do not follow the expected format making them unparsable. The results in Table 2 show that the full SketchAgent pipeline achieves the highest performance, highlighting the importance of each component. Interestingly, not providing a complete sketch example significantly reduces performance. We additionally ablate the impact of the grid resolution, by varying the resolution from 10 to 100 (see Tab. 2, bottom). Extremely low resolutions degrade performance, while mid-level resolutions outperform \\(100 \\times 100\\)." + }, + { + "type": "table", + "bbox": [ + 0.505, + 0.09, + 0.891, + 0.187 + ], + "angle": 0, + "content": "
w/o System Promptw/o CoTModified ICLSketchAgent (full)
Top10.20 ± 0.040.14 ± 0.020.07 ± 0.020.23 ± 0.04
Top50.42 ± 0.030.29 ± 0.040.16 ± 0.030.43 ± 0.06
Grid Size10 × 1025 × 2550 × 5075 × 75
Top1 / Top50.14 / 0.280.19 / 0.420.23 / 0.430.23 / 0.41
" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.192, + 0.893, + 0.248 + ], + "angle": 0, + "content": "Table 2. Ablation study. Average Top-1 and Top-5 CLIP recognition accuracy. Top: We systematically remove each component in our pipeline, showcasing all components contribute to the agent's full performance. Bottom: Grid resolution ablation." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.266, + 0.774, + 0.281 + ], + "angle": 0, + "content": "7. Limitations and Future Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.287, + 0.893, + 0.545 + ], + "angle": 0, + "content": "SketchAgent has several limitations. First, it is constrained by the priors of the backbone model, primarily optimized for text rather than visual content. As a result, the agent often produces rich textual descriptions of object parts but struggles to convert these into effective sketching actions, resulting in overly abstract and unrecognizable outputs. For example, in Fig. 15A, the agent effectively describes key parts of a unicorn (e.g., the horn), but the sketch is unrecognizable. This constraint also impacts the depiction of human figures (Fig. 15B). While distinctive features (e.g., Frida Kahlo's eyebrows or Michael Jordan's dunk) may be captured well in language, the resulting sketches are overly simple, with an amateur style, lacking expressivity. We expect this issue to improve as future models advance in vision capabilities. Lastly, the agent may struggle with drawing letters and numbers. This could be improved in future work by providing relevant in-context examples." + }, + { + "type": "image_caption", + "bbox": [ + 0.502, + 0.556, + 0.641, + 0.569 + ], + "angle": 0, + "content": "(A) Non-recognizable" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.57, + 0.64, + 0.626 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.644, + 0.557, + 0.742, + 0.569 + ], + "angle": 0, + "content": "(B) Human figures" + }, + { + "type": "image", + "bbox": [ + 0.646, + 0.57, + 0.751, + 0.629 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.758, + 0.557, + 0.886, + 0.569 + ], + "angle": 0, + "content": "(C) Letters and numbers" + }, + { + "type": "image", + "bbox": [ + 0.762, + 0.57, + 0.891, + 0.629 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.633, + 0.893, + 0.676 + ], + "angle": 0, + "content": "Figure 15. Limitations. Sketches of complex concepts (A) and human figures (B) appear too abstract and unrecognizable with non-professional style. (C) Fail to depict letters and numbers." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.698, + 0.627, + 0.714 + ], + "angle": 0, + "content": "8. Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.892, + 0.9 + ], + "angle": 0, + "content": "We presented a method for language-driven, sequential sketch generation, that can produce versatile sketches in real-time and meaningfully engage in collaborative sketching sessions with humans. We show that the prior knowledge embedded in pretrained multimodal LLMs can be effectively leveraged for sketch generation through an intuitive sketching language and a grid canvas, without requiring additional training or fine-tuning. We hope our work represents a meaningful step toward developing general-purpose sketching systems with the potential to enhance human-computer communication and computer-aided ideation." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "23362" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.077, + 0.091, + 0.267, + 0.108 + ], + "angle": 0, + "content": "9. Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.115, + 0.473, + 0.373 + ], + "angle": 0, + "content": "We thank Yuval Alaluf, Hila Chefer, Assaf Ben Kish, Joanna Materzynska, Rinon Gal, Elad Richardson, Arnav Verma, and Ellie Arar for providing feedback on early versions of our manuscript. We are also grateful to Justin Yang, who developed the initial Photodraw application and assisted with the interactive human study setup, and to Dingning Cao, who contributed to the SketchAgent interface and application. Special thanks to Yarden Frenkel for his insights, early explorations, and inspiring discussions. This work was partially supported by NSF CAREER #2047191, NSF DRL #2400471, Stanford Human Centered AI Institute Hoffman-Yee Grant, Hyundai Motor Company, ARL grant W911NF-18-2-021, the Zuckerman STEM Leadership Program, and the Viterbi Fellowship. The sponsors had no role in the experimental design or analysis, the decision to publish, or manuscript preparation. The authors have no competing interests to report." + }, + { + "type": "title", + "bbox": [ + 0.079, + 0.385, + 0.175, + 0.401 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.411, + 0.47, + 0.549 + ], + "angle": 0, + "content": "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katie Millicah, Malcolm Reynolds, Roman Ring, Eliza Rutherford, Serkan Cabi, Tengda Han, Zhitao Gong, Sina Samangooei, Marianne Monteiro, Jacob Menick, Sebastian Borgeaud, Andrew Brock, Aida Nematzadeh, Sahand Sharifzadeh, Mikolaj Binkowski, Ricardo Barreira, Oriol Vinyals, Andrew Zisserman, and Karen Simonyan. Flamingo: a visual language model for few-shot learning. Red Hook, NY, USA, 2024. Curran Associates Inc. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.551, + 0.468, + 0.577 + ], + "angle": 0, + "content": "[2] Anthropic. Claude. https://www.anthropic.com/claude, 2023. 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.095, + 0.579, + 0.469, + 0.62 + ], + "angle": 0, + "content": "[3] Itamar Berger, Ariel Shamir, Moshe Mahler, Elizabeth Carter, and Jessica Hodgins. Style and abstraction in portrait sketching. ACM Trans. Graph., 32(4), 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.622, + 0.469, + 0.69 + ], + "angle": 0, + "content": "[4] James Betker, Gabriel Goh, Li Jing, † TimBrooks, Jianfeng Wang, Linjie Li, † LongOuyang, † JuntangZhuang, † JoyceLee, † YufeiGuo, † WesamManassra, † PrafullaDhariwal, † CaseyChu, † YunxinJiao, and Aditya Ramesh. Improving image generation with better captions. 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.692, + 0.469, + 0.76 + ], + "angle": 0, + "content": "[5] Ayan Kumar Bhunia, Ayan Das, Umar Riaz Muhammad, Yongxin Yang, Timothy M. Hospedales, Tao Xiang, Yulia Gryaditskaya, and Yi-Zhe Song. Pixel: a competitive sketching ai agent. so you think you can sketch? ACM Trans. Graph., 39:166:1-166:15, 2020. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.762, + 0.469, + 0.817 + ], + "angle": 0, + "content": "[6] Ankan Kumar Bhunia, Salman Khan, Hisham Cholakkal, Rao Muhammad Anwer, Fahad Shahbaz Khan, Jorma Laaksonen, and Michael Felsberg. Doodleformer: Creative sketch drawing with transformers. ECCV, 2022. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.818, + 0.469, + 0.9 + ], + "angle": 0, + "content": "[7] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel Ziegler, Jeffrey Wu, Clemens Winter, Chris Hesse, Mark Chen, Eric" + }, + { + "type": "list", + "bbox": [ + 0.095, + 0.411, + 0.47, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.515, + 0.093, + 0.892, + 0.175 + ], + "angle": 0, + "content": "Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. Language models are few-shot learners. In Advances in Neural Information Processing Systems, pages 1877-1901. Curran Associates, Inc., 2020. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.178, + 0.892, + 0.246 + ], + "angle": 0, + "content": "[8] Sébastien Bubeck, Varun Chandrasekaran, Ronen Eldan, Johannes Gehrke, Eric Horvitz, Ece Kamar, Peter Lee, Yin Tat Lee, Yuanzhi Li, Scott Lundberg, et al. Sparks of artificial general intelligence: Early experiments with gpt-4. arXiv preprint arXiv:2303.12712, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.249, + 0.892, + 0.29 + ], + "angle": 0, + "content": "[9] Mu Cai, Zeyi Huang, Yuheng Li, Haohan Wang, and Yong Jae Lee. Delving into LLMs' visual understanding ability using SVG to bridge image and text, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.292, + 0.891, + 0.332 + ], + "angle": 0, + "content": "[10] John Canny. A computational approach to edge detection. IEEE Transactions on pattern analysis and machine intelligence, (6):679-698, 1986. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.334, + 0.892, + 0.375 + ], + "angle": 0, + "content": "[11] Alexandre Carlier, Martin Danelljan, Alexandre Alahi, and Radu Timofte. Deepsvg: A hierarchical generative network for vector graphics animation, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.378, + 0.892, + 0.445 + ], + "angle": 0, + "content": "[12] Caroline Chan, Frédo Durand, and Phillip Isola. Learning to generate line drawings that convey geometry and semantics. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7915-7925, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.448, + 0.892, + 0.517 + ], + "angle": 0, + "content": "[13] Zhenfang Chen, Qinhong Zhou, Yikang Shen, Yining Hong, Zhiqing Sun, Dan Gutfreund, and Chuang Gan. Visual chain-of-thought prompting for knowledge-based visual reasoning. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1254-1262, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.519, + 0.892, + 0.56 + ], + "angle": 0, + "content": "[14] Changwoon Choi, Jaeah Lee, Jaesik Park, and Young Min Kim. 3doodle: Compact abstraction of objects with 3d strokes. ACM Trans. Graph., 43(4), 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.563, + 0.892, + 0.603 + ], + "angle": 0, + "content": "[15] Xiangxiang Chu, Jianlin Su, Bo Zhang, and Chunhua Shen. Visionllama: A unified llama backbone for vision tasks, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.606, + 0.892, + 0.687 + ], + "angle": 0, + "content": "[16] Ayan Das, Yongxin Yang, Timothy Hospedales, Tao Xiang, and Yi-Zhe Song. Béziersketch: A generative model for scalable vector sketches. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXVI 16, pages 632-647. Springer, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.69, + 0.892, + 0.787 + ], + "angle": 0, + "content": "[17] Nicholas Davis, Chih-PIn Hsiao, Kunwar Yashraj Singh, Lisa Li, Sanat Moningi, and Brian Magerko. Drawing apprentice: An enactive co-creative agent for artistic collaboration. In Proceedings of the 2015 ACM SIGCHI Conference on Creativity and Cognition, page 185–186, New York, NY, USA, 2015. Association for Computing Machinery. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.789, + 0.892, + 0.829 + ], + "angle": 0, + "content": "[18] Vanessa de Andrade, Sofia Freire, Monica Baptista, and Yael Shwartz. Drawing as a space for social-cognitive interaction. Education Sciences, 12:45, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.832, + 0.892, + 0.9 + ], + "angle": 0, + "content": "[19] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language" + }, + { + "type": "list", + "bbox": [ + 0.511, + 0.093, + 0.892, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "23363" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.118, + 0.093, + 0.47, + 0.134 + ], + "angle": 0, + "content": "Technologies, Volume 1 (Long and Short Papers), pages 4171-4186, Minneapolis, Minnesota, 2019. Association for Computational Linguistics. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.135, + 0.47, + 0.203 + ], + "angle": 0, + "content": "[20] Sivan Doveh, Shaked Perek, M Jehanzeb Mirza, Wei Lin, Amit Alfassy, Assaf Arbelle, Shimon Ullman, and Leonid Karlinsky. Towards multimodal in-context learning for vision & language models. arXiv preprint arXiv:2403.12736, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.089, + 0.205, + 0.471, + 0.245 + ], + "angle": 0, + "content": "[21] Mathias Eitz, James Hays, and Marc Alexa. How do humans sketch objects? ACM Trans. Graph., 31(4), 2012. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.247, + 0.471, + 0.328 + ], + "angle": 0, + "content": "[22] Ziv Epstein, Aaron Hertzmann, Laura Mariah Herman, Robert Mahari, Morgan R. Frank, Matthew Groh, Hope Schroeder, Amy Smith, Memo Akten, Jessica Fjeld, Hany Farid, Neil Leach, Alex Pentland, and Olga Russakovsky. Art and the science of generative ai. Science, 380:1110-1111, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.33, + 0.471, + 0.371 + ], + "angle": 0, + "content": "[23] Judy Fan, Wilma A. Bainbridge, Rebecca Chamberlain, and Jeffrey D. Wammes. Drawing as a versatile cognitive tool. Nature Reviews Psychology, 2:556 - 568, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.372, + 0.471, + 0.44 + ], + "angle": 0, + "content": "[24] Judith E. Fan, Monica Dinculescu, and David Ha. collaborate: An environment for collaborative sketching with an artificial agent. In Proceedings of the 2019 Conference on Creativity and Cognition, page 556-561, New York, NY, USA, 2019. Association for Computing Machinery. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.442, + 0.471, + 0.496 + ], + "angle": 0, + "content": "[25] Judith E. Fan, Robert D. Hawkins, Mike Wu, and Noah D. Goodman. Pragmatic Inference and Visual Abstraction Enable Contextual Flexibility During Visual Communication. Computational Brain & Behavior, 3(1):86-101, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.498, + 0.471, + 0.538 + ], + "angle": 0, + "content": "[26] Judith E Fan, Wilma A Bainbridge, Rebecca Chamberlain, and Jeffrey D Wammes. Drawing as a versatile cognitive tool. Nature Reviews Psychology, 2(9):556-568, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.539, + 0.471, + 0.58 + ], + "angle": 0, + "content": "[27] Logan Fiorella and Shelbi Kuhlmann. Creating drawings enhances learning by teaching. Journal of Educational Psychology, 112(4):811, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.582, + 0.471, + 0.636 + ], + "angle": 0, + "content": "[28] Kenneth Forbus, Jeffrey Usher, Andrew Lovett, Kate Lockwood, and Jon Wetzel. Cogsketch: Sketch understanding for cognitive science research and for education. Topics in Cognitive Science, 3(4):648-666, 2011. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.637, + 0.471, + 0.705 + ], + "angle": 0, + "content": "[29] Kevin Frans, L. B. Soros, and Olaf Witkowski. Clipdraw: exploring text-to-drawing synthesis through language-image encoders. In Proceedings of the 36th International Conference on Neural Information Processing Systems, Red Hook, NY, USA, 2024. Curran Associates Inc. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.707, + 0.471, + 0.774 + ], + "angle": 0, + "content": "[30] Xingyu Fu, Yushi Hu, Bangzheng Li, Yu Feng, Haoyu Wang, Xudong Lin, Dan Roth, Noah A Smith, Wei-Chiu Ma, and Ranjay Krishna. Blink: Multimodal large language models can see but not perceive. arXiv preprint arXiv:2404.12390, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.776, + 0.471, + 0.845 + ], + "angle": 0, + "content": "[31] Rinon Gal, Yael Vinker, Yuval Alaluf, Amit Bermano, Daniel Cohen-Or, Ariel Shamir, and Gal Chechik. Breathing life into sketches using text-to-video priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4325–4336, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.847, + 0.471, + 0.9 + ], + "angle": 0, + "content": "[32] Yaroslav Ganin, Tejas D. Kulkarni, Igor Babuschkin, S. M. Ali Eslami, and Oriol Vinyals. Synthesizing programs for images using reinforced adversarial learning. *ArXiv*, abs/1804.01118, 2018. 3" + }, + { + "type": "list", + "bbox": [ + 0.088, + 0.093, + 0.471, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[33] Yaroslav Ganin, Sergey Bartunov, Yujia Li, Ethan Keller, and Stefano Saliceti. Computer-aided design as language. In Neural Information Processing Systems, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.136, + 0.892, + 0.204 + ], + "angle": 0, + "content": "[34] Chengying Gao, Qi Liu, Qi Xu, Limin Wang, Jianzhuang Liu, and Changqing Zou. Sketchycoco: Image generation from freehand scene sketches. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5174-5183, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.205, + 0.892, + 0.259 + ], + "angle": 0, + "content": "[35] Simon Garrod, Nicolas Fay, John Lee, Jon Oberlander, and Tracy MacLeod. Foundations of representation: Where might graphical symbol systems come from? Cognitive Science, 31(6):961-987, 2007. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.261, + 0.892, + 0.3 + ], + "angle": 0, + "content": "[36] Songwei Ge, Vedanuj Goswami, C Lawrence Zitnick, and Devi Parikh. Creative sketch generation. arXiv preprint arXiv:2011.10039, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.302, + 0.892, + 0.343 + ], + "angle": 0, + "content": "[37] Songwei Ge, Vedanuj Goswami, Larry Zitnick, and Devi Parikh. Creative sketch generation. In International Conference on Learning Representations, 2021. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.345, + 0.892, + 0.427 + ], + "angle": 0, + "content": "[38] Hinnie Gijlers, Armin Weinberger, Alieke Mattia van Dijk, Lars Bollen, and Wouter van Joolingen. Collaborative drawing on a shared digital canvas in elementary science education: The effects of script and task awareness support. International Journal of Computer-Supported Collaborative Learning, 8(4):427-453, 2013. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.428, + 0.892, + 0.468 + ], + "angle": 0, + "content": "[39] Gabriela Goldschmidt. Serial sketching: visual problem solving in designing. Cybernetics and System, 23(2):191-219, 1992. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.47, + 0.892, + 0.524 + ], + "angle": 0, + "content": "[40] Yulia Gryaditskaya, Mark Sypesteyn, Jan Willem Hoftijzer, Sylvia Pont, Frédo Durand, and Adrien Bousseau. Opensketch: a richly-annotated dataset of product design sketches. ACM Trans. Graph., 38(6), 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.526, + 0.892, + 0.636 + ], + "angle": 0, + "content": "[41] Tianrui Guan, Fuxiao Liu, Xiyang Wu, Ruiqi Xian, Zongxia Li, Xiaoyu Liu, Xijun Wang, Lichang Chen, Furong Huang, Yaser Yacoob, Dinesh Manocha, and Tianyi Zhou. Hallusionbench: An advanced diagnostic suite for entangled language hallucination and visual illusion in large vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14375-14385, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.637, + 0.892, + 0.676 + ], + "angle": 0, + "content": "[42] David Ha and Douglas Eck. A neural representation of sketch drawings. CoRR, abs/1704.03477, 2017. 1, 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.679, + 0.892, + 0.734 + ], + "angle": 0, + "content": "[43] Yucheng Han, China. Xiaoyan Zhang, Xin Chen, Xu Yang, Zhibin Wang, Gang Yu, Bin Fu, and Hanwang Zhang. Chartlama: A multimodal llm for chart understanding and generation. ArXiv, abs/2311.16483, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.735, + 0.892, + 0.802 + ], + "angle": 0, + "content": "[44] Yushi Hu, Weijia Shi, Xingyu Fu, Dan Roth, Mari Ostendorf, Luke Zettlemoyer, Noah A Smith, and Ranjay Krishna. Visual sketchpad: Sketching as a visual chain of thought for multimodal language models. arXiv preprint arXiv:2406.09403, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.804, + 0.892, + 0.871 + ], + "angle": 0, + "content": "[45] Francisco Javier Ibarrola, Tomas Lawton, and Kazjon Grace. A collaborative, interactive and context-aware drawing agent for co-creative design. IEEE Transactions on Visualization and Computer Graphics, 30:5525-5537, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.873, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[46] Ajay Jain, Amber Xie, and Pieter Abbeel. Vectorfusion: Text-to-svg by abstracting pixel-based diffusion models. In" + }, + { + "type": "list", + "bbox": [ + 0.51, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "23364" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.118, + 0.092, + 0.47, + 0.133 + ], + "angle": 0, + "content": "Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1911-1920, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.137, + 0.47, + 0.178 + ], + "angle": 0, + "content": "[47] Jongejan Jonas, Rowley Henry, Kawashima Takashi, Kim Jongmin, and Fox-Gieg Nick. The Quick, Draw! - A.I. Experiment, 2016. 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.181, + 0.471, + 0.223 + ], + "angle": 0, + "content": "[48] David Kaiser. Drawing theories apart: The dispersion of Feynman diagrams in postwar physics. University of Chicago Press, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.225, + 0.47, + 0.309 + ], + "angle": 0, + "content": "[49] Pegah Karimi, Jeba Rezwana, Safat Siddiqui, Mary Lou Maher, and Nasrin Dehbozorgi. Creative sketching partner: an analysis of human-ai co-creativity. In Proceedings of the 25th International Conference on Intelligent User Interfaces, page 221–230, New York, NY, USA, 2020. Association for Computing Machinery. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.311, + 0.47, + 0.395 + ], + "angle": 0, + "content": "[50] Pegah Karimi, Jeba Rezwana, Safat Siddiqui, Mary Lou Maher, and Nasrin Dehbozorgi. Creative sketching partner: an analysis of human-ai co-creativity. In Proceedings of the 25th International Conference on Intelligent User Interfaces, page 221–230, New York, NY, USA, 2020. Association for Computing Machinery. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.397, + 0.47, + 0.438 + ], + "angle": 0, + "content": "[51] Andrew Kyle Lampinen, Stephanie C. Y. Chan, Aaditya K. Singh, and Murray Shanahan. The broader spectrum of in-context learning, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.441, + 0.468, + 0.469 + ], + "angle": 0, + "content": "[52] Paul Laseau. Graphic thinking for architects and designers. John Wiley & Sons, 2000. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.471, + 0.47, + 0.555 + ], + "angle": 0, + "content": "[53] Tomas Lawton, Kazjon Grace, and Francisco J Ibarrola. When is a tool a tool? user perceptions of system agency in human-ai co-creative drawing. In Proceedings of the 2023 ACM Designing Interactive Systems Conference, page 1978–1996, New York, NY, USA, 2023. Association for Computing Machinery. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.557, + 0.47, + 0.642 + ], + "angle": 0, + "content": "[54] Tomas Lawton, Francisco J Ibarrola, Dan Ventura, and Kazjon Grace. Drawing with reframer: Emergence and control in co-creative ai. In Proceedings of the 28th International Conference on Intelligent User Interfaces, page 264–277, New York, NY, USA, 2023. Association for Computing Machinery. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.643, + 0.468, + 0.699 + ], + "angle": 0, + "content": "[55] Yong Jae Lee, C. Lawrence Zitnick, and Michael F. Cohen. Shadowdraw: real-time user guidance for freehand drawing. In ACM SIGGRAPH 2011 Papers, New York, NY, USA, 2011. Association for Computing Machinery. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.701, + 0.468, + 0.77 + ], + "angle": 0, + "content": "[56] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. BLIP: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In Proceedings of the 39th International Conference on Machine Learning, pages 12888-12900. PMLR, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.772, + 0.468, + 0.841 + ], + "angle": 0, + "content": "[57] Mengtian Li, Zhe Lin, Radomir Mech, Ersin Yumer, and Deva Ramanan. Photo-sketching: Inferring contour drawings from images. In 2019 IEEE Winter Conference on Applications of Computer Vision (WACV), pages 1403-1412. IEEE, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.088, + 0.845, + 0.468, + 0.9 + ], + "angle": 0, + "content": "[58] Tzu-Mao Li, Michal Lukáč, Gharbi Michael, and Jonathan Ragan-Kelley. Differentiable vector graphics rasterization for editing and learning. ACM Trans. Graph. (Proc. SIGGRAPH Asia), 39(6):193:1-193:15, 2020. 1, 3" + }, + { + "type": "list", + "bbox": [ + 0.088, + 0.092, + 0.471, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "[59] Yi Li, Yi-Zhe Song, Timothy M. Hospedales, and Shaogang Gong. Free-hand sketch synthesis with deformable stroke models. CoRR, abs/1510.02644, 2015. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.137, + 0.892, + 0.206 + ], + "angle": 0, + "content": "[60] Yijun Li, Chen Fang, Aaron Hertzmann, Eli Shechtman, and Ming-Hsuan Yang. Im2pencil: Controllable pencil illustration from photographs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1525-1534, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.208, + 0.892, + 0.289 + ], + "angle": 0, + "content": "[61] Hangyu Lin, Yanwei Fu, Yu-Gang Jiang, and X. Xue. Sketch-bert: Learning sketch bidirectional encoder representation from transformers by self-supervised learning of sketch gestalt. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6757-6766, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.293, + 0.892, + 0.36 + ], + "angle": 0, + "content": "[62] Difan Liu, Matthew Fisher, Aaron Hertzmann, and Evangelos Kalogerakis. Neural strokes: Stylized line drawing of 3d shapes. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14204-14213, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.363, + 0.892, + 0.445 + ], + "angle": 0, + "content": "[63] Fang Liu, Xiaoming Deng, Yu-Kun Lai, Yong-Jin Liu, Cuixia Ma, and Hongan Wang. Sketchgan: Joint sketch completion and recognition with generative adversarial network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.448, + 0.892, + 0.503 + ], + "angle": 0, + "content": "[64] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In Advances in Neural Information Processing Systems, pages 34892-34916. Curran Associates, Inc., 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.505, + 0.892, + 0.561 + ], + "angle": 0, + "content": "[65] Bria Long, Judith Fan, Holly Huey, Zixian Chai, and Michael Frank. Parallel developmental changes in children's production and recognition of line drawings of visual concepts. Nature Communications, 15, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.563, + 0.892, + 0.701 + ], + "angle": 0, + "content": "[66] Arjun Majumdar, Anurag Ajay, Xiaohan Zhang, Pranav Putta, Sriram Yenamandra, Mikael Henaff, Sneha Silwal, Paul Mcvay, Oleksandr Maksymets, Sergio Arnaud, Karmesh Yadav, Qiyang Li, Ben Newman, Mohit Sharma, Vincent Berges, Shiqi Zhang, Pulkit Agrawal, Yonatan Bisk, Dhruv Batra, Mrinal Kalakrishnan, Franziska Meier, Chris Paxton, Sasha Sax, and Aravind Rajeswaran. Openeqa: Embodied question answering in the era of foundation models. In Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.703, + 0.892, + 0.773 + ], + "angle": 0, + "content": "[67] William P. McCarthy, Justin Matejka, Karl D.D. Willis, Judith E. Fan, and Yewen Pu. Communicating design intent using drawing and text. In Proceedings of the 16th Conference on Creativity & Cognition, page 512-519, New York, NY, USA, 2024. Association for Computing Machinery. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.774, + 0.892, + 0.844 + ], + "angle": 0, + "content": "[68] John FJ Mellor, Eunbyung Park, Yaroslav Ganin, Igor Babuschkin, Tejas Kulkarni, Dan Rosenbaum, Andy Ballard, Theophane Weber, Oriol Vinyals, and SM Eslami. Unsupervised doodling and painting with improved spiral. arXiv preprint arXiv:1910.01007, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.511, + 0.845, + 0.892, + 0.899 + ], + "angle": 0, + "content": "[69] Daniela Mihai and Jonathon Hare. Learning to draw: Emergent communication through sketching. Advances in Neural Information Processing Systems, 34:7153-7166, 2021. 3" + }, + { + "type": "list", + "bbox": [ + 0.511, + 0.093, + 0.892, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23365" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.093, + 0.47, + 0.161 + ], + "angle": 0, + "content": "[70] Chancharik Mitra, Brandon Huang, Trevor Darrell, and Roei Herzig. Compositional chain-of-thought prompting for large multimodal models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14420-14431, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.162, + 0.472, + 0.232 + ], + "angle": 0, + "content": "[71] Kushin Mukherjee, Holly Huey, Xuanchen Lu, Yael Vinker, Rio Aguina-Kang, Ariel Shamir, and Judith Fan. Seva: Leveraging sketches to evaluate alignment between human and machine visual abstraction. In Advances in Neural Information Processing Systems, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.233, + 0.471, + 0.272 + ], + "angle": 0, + "content": "[72] Omar W Nasim. Observing by hand: sketching the nebulae in the nineteenth century. University of Chicago Press, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.274, + 0.472, + 0.37 + ], + "angle": 0, + "content": "[73] Changhoon Oh, Jungwoo Song, Jinhan Choi, Seonghyeon Kim, Sungwoo Lee, and Bongwon Suh. I lead, you help but only with enough details: Understanding user experience of co-creation with artificial intelligence. In Proceedings of the 2018 CHI Conference on Human Factors in Computing Systems, page 1–13, New York, NY, USA, 2018. Association for Computing Machinery. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.372, + 0.408, + 0.385 + ], + "angle": 0, + "content": "[74] OpenAI. Gpt-4 technical report, 2024. 2, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.386, + 0.47, + 0.453 + ], + "angle": 0, + "content": "[75] Dustin Podell, Zion English, Kyle Lacey, A. Blattmann, Tim Dockhorn, Jonas Muller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. ArXiv, abs/2307.01952, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.455, + 0.471, + 0.523 + ], + "angle": 0, + "content": "[76] Yonggang Qi, Guoyao Su, Pinaki Nath Chowdhury, Mingkang Li, and Yi-Zhe Song. Sketchlattice: Latticed representation for sketch manipulation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 953–961, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.525, + 0.47, + 0.579 + ], + "angle": 0, + "content": "[77] Shuwen Qiu, Sirui Xie, Lifeng Fan, Tao Gao, Song-Chun Zhu, and Yixin Zhu. Emergent graphical conventions in a visual communication game. arXiv preprint arXiv:2111.14210, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.581, + 0.47, + 0.662 + ], + "angle": 0, + "content": "[78] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. CoRR, abs/2103.00020, 2021. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.665, + 0.47, + 0.732 + ], + "angle": 0, + "content": "[79] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. J. Mach. Learn. Res., 21(1), 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.734, + 0.47, + 0.788 + ], + "angle": 0, + "content": "[80] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.79, + 0.47, + 0.858 + ], + "angle": 0, + "content": "[81] Leo Sampaio Ferraz Ribeiro, Tu Bui, John P. Collomosse, and Moacir Antonelli Ponti. Sketchformer: Transformer-based representation for sketched structure. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14141-14150, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.86, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[82] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models, 2022. 1, 2, 3, 6" + }, + { + "type": "list", + "bbox": [ + 0.087, + 0.093, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.093, + 0.892, + 0.161 + ], + "angle": 0, + "content": "[83] Daniel Rose, Vaishnavi Himakunthala, Andy Ouyang, Ryan He, Alex Mei, Yujie Lu, Michael Saxon, Chinmay Sonar, Diba Mirza, and William Yang Wang. Visual chain of thought: bridging logical gaps with multimodal infillings. arXiv preprint arXiv:2305.02317, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.164, + 0.892, + 0.259 + ], + "angle": 0, + "content": "[84] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S. Sara Mahdavi, Rapha Gontijo Lopes, Tim Salimans, Jonathan Ho, David J Fleet, and Mohammad Norouzi. Photorealistic text-to-image diffusion models with deep language understanding, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.262, + 0.892, + 0.303 + ], + "angle": 0, + "content": "[85] Patsorn Sangkloy, Nathan Burnell, Cusuh Ham, and James Hays. The sketchy database: Learning to retrieve badly drawn bunnies. ACM Trans. Graph., 35(4), 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.305, + 0.892, + 0.374 + ], + "angle": 0, + "content": "[86] Peter Schaldenbrand, James McCann, and Jean Oh. Frida: A collaborative robot painter with a differentiable, real2sim2real planning environment. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 11712-11718, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.377, + 0.892, + 0.444 + ], + "angle": 0, + "content": "[87] Peter Schaldenbrand, Gaurav Parmar, Jun-Yan Zhu, James McCann, and Jean Oh. Cofrida: Self-supervised fine-tuning for human-robot co-painting. In 2024 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.447, + 0.892, + 0.475 + ], + "angle": 0, + "content": "[88] Donald A Schon and Vincent DeSanctis. The reflective practitioner: How professionals think in action, 1986. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.477, + 0.892, + 0.573 + ], + "angle": 0, + "content": "[89] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, Patrick Schramowski, Srivatsa Kundurthy, Katherine Crowson, Ludwig Schmidt, Robert Kaczmarczyk, and Jenia Jitsev. Laion-5b: An open large-scale dataset for training next generation image-text models, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.576, + 0.892, + 0.644 + ], + "angle": 0, + "content": "[90] Tamar Rott Shaham, Sarah Schwettmann, Franklin Wang, Achyuta Rajaram, Evan Hernandez, Jacob Andreas, and Antonio Torralba. A multimodal automated interpretability agent. In *Forty-first International Conference on Machine Learning*, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.647, + 0.892, + 0.715 + ], + "angle": 0, + "content": "[91] Hao Shao, Shengju Qian, Han Xiao, Guanglu Song, Zhuofan Zong, Letian Wang, Yu Liu, and Hongsheng Li. Visual cot: Advancing multi-modal language models with a comprehensive dataset and benchmark for chain-of-thought reasoning. arXiv preprint arXiv:2403.16999, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.717, + 0.892, + 0.799 + ], + "angle": 0, + "content": "[92] Pratyusha Sharma, Tamar Rott Shaham, Manel Baradad, Stephanie Fu, Adrian Rodriguez-Munoz, Shivam Duggal, Phillip Isola, and Antonio Torralba. A vision check-up for language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14410-14419, 2024. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.802, + 0.892, + 0.843 + ], + "angle": 0, + "content": "[93] Jifei Song, Kaiyue Pang, Yi-Zhe Song, Tao Xiang, and Timothy Hospedales. Learning to sketch with shortcut cycle consistency, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.51, + 0.846, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[94] Guoyao Su, Yonggang Qi, Kaiyue Pang, Jie Yang, Yi-Zhe Song, and CVSSP SketchX. Sketchhealer: A graph-to-sequence network for recreating partial human sketches. In BMVC, page 5, 2020. 2" + }, + { + "type": "list", + "bbox": [ + 0.51, + 0.093, + 0.892, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "23366" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.092, + 0.47, + 0.146 + ], + "angle": 0, + "content": "[95] Yanpeng Sun, Qiang Chen, Jian Wang, Jingdong Wang, and Zechao Li. Exploring effective factors for improving visual in-context learning. arXiv preprint arXiv:2304.04748, 2023.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.148, + 0.471, + 0.189 + ], + "angle": 0, + "content": "[96] Ivan E. Sutherland. Sketchpad—a man-machine graphical communication system, page 391-408. Association for Computing Machinery, New York, NY, USA, 1998. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.19, + 0.471, + 0.218 + ], + "angle": 0, + "content": "[97] Gemini Team. Gemini: A family of highly capable multimodal models, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.219, + 0.471, + 0.302 + ], + "angle": 0, + "content": "[98] Jakob Tholander and Martin Jonsson. Design ideation with ai - sketching, thinking and talking with generative machine learning models. In Proceedings of the 2023 ACM Design-Interactive Systems Conference, page 1930-1940, New York, NY, USA, 2023. Association for Computing Machinery. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.302, + 0.471, + 0.343 + ], + "angle": 0, + "content": "[99] Shengbang Tong, Zhuang Liu, Yuexiang Zhai, Yi Ma, Yann LeCun, and Saining Xie. Eyes wide shut? exploring the visual shortcomings of multimodal llms, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.344, + 0.47, + 0.427 + ], + "angle": 0, + "content": "[100] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee Lacroix, Baptiste Roziere, Naman Goyal, Eric Hambro, Faisal Azhar, Aurelien Rodriguez, Armand Joulin, Edouard Grave, and Guillaume Lample. Llama: Open and efficient foundation language models. ArXiv, abs/2302.13971, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.428, + 0.47, + 0.468 + ], + "angle": 0, + "content": "[101] Barbara Tversky. What do sketches say about thinking? AAAI Spring Symp. Sketch Understanding Worksh., 2002. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.469, + 0.47, + 0.498 + ], + "angle": 0, + "content": "[102] Barbara Tversky. Visualizing thought. In Handbook of human centric visualization, pages 3-40. Springer, 2013. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.498, + 0.471, + 0.567 + ], + "angle": 0, + "content": "[103] Barbara Tversky, Masaki Suwa, Maneesh Agrawala, Julie Heiser, Chris Stolte, Pat Hanrahan, Doantam Phan, Jeff Klingner, Marie-Paule Daniel, Paul Lee, et al. Sketches for design and design of sketches. Human Behaviour in Design: Individuals, Teams, Tools, pages 79-86, 2003. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.568, + 0.47, + 0.621 + ], + "angle": 0, + "content": "[104] Russell Tytler, Vaughan Prain, George Aranda, Joseph Ferguson, and Radhika Gorur. Drawing to reason and learn in science. Journal of Research in Science Teaching, 57(2): 209-231, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.622, + 0.47, + 0.692 + ], + "angle": 0, + "content": "[105] Yael Vinker, Ehsan Pajouheshgar, Jessica Y. Bo, Roman Christian Bachmann, Amit Haim Bermano, Daniel Cohen-Or, Amir Zamir, and Ariel Shamir. Clipasso: Semantically-aware object sketching. ACM Trans. Graph., 41(4), 2022. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.692, + 0.47, + 0.762 + ], + "angle": 0, + "content": "[106] Yael Vinker, Yuval Alaluf, Daniel Cohen-Or, and Ariel Shamir. Clipascene: Scene sketching with different types and levels of abstraction. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 4146–4156, 2023. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.762, + 0.47, + 0.803 + ], + "angle": 0, + "content": "[107] Jiawei Wang and Changjian Li. Contextseg: Sketch semantic segmentation by querying the context with attention. arXiv preprint arXiv:2311.16682, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.804, + 0.47, + 0.9 + ], + "angle": 0, + "content": "[108] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Brian Ichter, Fei Xia, Ed H. Chi, Quoc V. Le, and Denny Zhou. Chain-of-thought prompting elicits reasoning in large language models. In Proceedings of the 36th International Conference on Neural Information Processing Systems, Red Hook, NY, USA, 2022. Curran Associates Inc. 2, 5" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.092, + 0.892, + 0.147 + ], + "angle": 0, + "content": "[109] Holger Winnemöller, Jan Eric Kyprianidis, and Sven C. Olsen. Xdog: An extended difference-of-gaussians compendium including advanced image stylization. Comput. Graph., 36:740-753, 2012. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.892, + 0.177 + ], + "angle": 0, + "content": "[110] World Wide Web Consortium (W3C). Scalable Vector Graphics (SVG), 1999. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.179, + 0.892, + 0.234 + ], + "angle": 0, + "content": "[111] Chenfei Wu, Shengming Yin, Weizhen Qi, Xiaodong Wang, Zecheng Tang, and Nan Duan. Visual chatgpt: Talking, drawing and editing with visual foundation models. arXiv preprint arXiv:2303.04671, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.235, + 0.892, + 0.29 + ], + "angle": 0, + "content": "[112] Rong Wu, Wanchao Su, Kede Ma, and Jing Liao. Iconshop: Text-guided vector icon synthesis with autoregressive transformers. ACM Transactions on Graphics (TOG), 42:1-14, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.292, + 0.892, + 0.361 + ], + "angle": 0, + "content": "[113] Chufeng Xiao, Wanchao Su, Jing Liao, Zhouhui Lian, Yi-Zhe Song, and Hongbo Fu. Diffsketching: How differently do people sketch 3d objects? ACM Transactions on Graphics (Proceedings of ACM SIGGRAPH Asia 2022), 41 (4):1-16, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.362, + 0.892, + 0.405 + ], + "angle": 0, + "content": "[114] Saining Xie and Zhuowen Tu. Holistically-nested edge detection. In Proceedings of the IEEE international conference on computer vision, pages 1395-1403, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.406, + 0.892, + 0.475 + ], + "angle": 0, + "content": "[115] XiMing Xing, Chuang Wang, Haitao Zhou, Jing Zhang, Qian Yu, and Dong Xu. Diffsketcher: Text guided vector sketch synthesis through latent diffusion models. In Advances in Neural Information Processing Systems, pages 15869-15889. Curran Associates, Inc., 2023. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.476, + 0.892, + 0.546 + ], + "angle": 0, + "content": "[116] Ximing Xing, Haitao Zhou, Chuang Wang, Jing Zhang, Dong Xu, and Qian Yu. Svgdreamer: Text guided svg generation with diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4546-4555, 2024. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.547, + 0.892, + 0.589 + ], + "angle": 0, + "content": "[117] Peng Xu, Timothy M. Hospedales, Qiyue Yin, Yi-Zhe Song, Tao Xiang, and Liang Wang. Deep learning for free-hand sketch: A survey and a toolbox, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.59, + 0.892, + 0.646 + ], + "angle": 0, + "content": "[118] Zhengyuan Yang, Linjie Li, Kevin Lin, Jianfeng Wang, Chung-Ching Lin, Zicheng Liu, and Lijuan Wang. The dawn of Imms: Preliminary explorations with gpt-4v(ision). ArXiv, abs/2309.17421, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.647, + 0.892, + 0.716 + ], + "angle": 0, + "content": "[119] Zhengyuan Yang, Jianfeng Wang, Linjie Li, Kevin Lin, Chung-Ching Lin, Zicheng Liu, and Lijuan Wang. Idea2img: Iterative self-refinement with gpt-4v (ision) for automatic image design and generation. arXiv preprint arXiv:2310.08541, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.717, + 0.892, + 0.788 + ], + "angle": 0, + "content": "[120] Ran Yi, Yong-Jin Liu, Yu-Kun Lai, and Paul L Rosin. Apdrawinggan: Generating artistic portrait drawings from face photos with hierarchical gans. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10743-10752, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.789, + 0.892, + 0.857 + ], + "angle": 0, + "content": "[121] Ran Yi, Yong-Jin Liu, Yu-Kun Lai, and Paul L Rosin. Unpaired portrait drawing generation via asymmetric cycle mapping. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8217-8225, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.858, + 0.892, + 0.902 + ], + "angle": 0, + "content": "[122] C. Zhang, Weijie Wang, Paul Pangaro, Nikolas Martelaro, and Daragh Byrne. Generative image ai using design sketches as input: Opportunities and challenges. Proceed-" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.092, + 0.892, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "23367" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.118, + 0.092, + 0.468, + 0.119 + ], + "angle": 0, + "content": "ings of the 15th Conference on Creativity and Cognition, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.47, + 0.191 + ], + "angle": 0, + "content": "[123] Jiahao Zhang, Bowen Wang, Liangzhi Li, Yuta Nakashima, and Hajime Nagahara. Instruct me more! random prompting for visual in-context learning. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2597-2606, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.192, + 0.471, + 0.233 + ], + "angle": 0, + "content": "[124] Peiying Zhang, Nanxuan Zhao, and Jing Liao. Text-to-vector generation with neural path representation. ACM Trans. Graph., 43(4), 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.234, + 0.471, + 0.289 + ], + "angle": 0, + "content": "[125] Yuanhan Zhang, Kaiyang Zhou, and Ziwei Liu. What makes good examples for visual in-context learning? Advances in Neural Information Processing Systems, 36: 17773-17794, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.291, + 0.471, + 0.346 + ], + "angle": 0, + "content": "[126] Zhengming Zhang, Xiaoming Deng, Jinyao Li, Yukun Lai, Cuixia Ma, Yongjin Liu, and Hongan Wang. Stroke-based semantic segmentation for scene-level free-hand sketches. Vis. Comput., 39(12):6309-6321, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.348, + 0.471, + 0.403 + ], + "angle": 0, + "content": "[127] Zhuosheng Zhang, Aston Zhang, Mu Li, Hai Zhao, George Karypis, and Alex Smola. Multimodal chain-of-thought reasoning in language models. arXiv preprint arXiv:2302.00923, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.405, + 0.471, + 0.46 + ], + "angle": 0, + "content": "[128] Yixiao Zheng, Kaiyue Pang, Ayan Das, Dongliang Chang, Yi-Zhe Song, and Zhanyu Ma. Creativeseg: Semantic segmentation of creative sketches. IEEE Transactions on Image Processing, 33:2266-2278, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.461, + 0.471, + 0.516 + ], + "angle": 0, + "content": "[129] Tao Zhou, Chen Fang, Zhaowen Wang, Jimei Yang, Byungmoon Kim, Zhili Chen, Jonathan Brandt, and Demetri Terzopoulos. Learning to sketch with deep q networks and demonstrated strokes. ArXiv, abs/1810.05977, 2018. 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.516 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "23368" + } + ] +] \ No newline at end of file diff --git a/2025/SketchAgent_ Language-Driven Sequential Sketch Generation/76b8d485-0aca-43aa-b3b2-9c94133cc069_origin.pdf b/2025/SketchAgent_ Language-Driven Sequential Sketch Generation/76b8d485-0aca-43aa-b3b2-9c94133cc069_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..d7d18e8435bb4e32b10daff91e6daa135fa90bb8 --- /dev/null +++ b/2025/SketchAgent_ Language-Driven Sequential Sketch Generation/76b8d485-0aca-43aa-b3b2-9c94133cc069_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9b9a557b8f1830f4ab8388366fd5b96a0d5f32dcda8bfe4a7f059de1e8c1c1f2 +size 5141998 diff --git a/2025/SketchAgent_ Language-Driven Sequential Sketch Generation/full.md b/2025/SketchAgent_ Language-Driven Sequential Sketch Generation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..ea74fce613e0e22386af101ac0688a60c83087af --- /dev/null +++ b/2025/SketchAgent_ Language-Driven Sequential Sketch Generation/full.md @@ -0,0 +1,421 @@ +# SketchAgent: Language-Driven Sequential Sketch Generation + +Yael Vinker $^{1}$ Tamar Rott Shaham $^{1}$ Kristine Zheng $^{2}$ Alex Zhao $^{1}$ Judith E Fan $^{2}$ Antonio Torralba $^{1}$ + +$^{1}$ MIT + +{yaelvink,tamarott,alexzhao,torralba}@mit.edu + +$^{2}$ Stanford University + +{jefan, kxzheng} $@$ stanford.edu + +https://sketch-agent.csail.mit.edu/ + +![](images/ab70df958686f2d2f5dd138cd4518fae137431adbcbf0135e8142d26148c2090.jpg) +Figure 1. SketchAgent leverages an off-the-shelf multimodal LLM to facilitate language-driven, sequential sketch generation through an intuitive sketching language. It can sketch diverse concepts, engage in interactive sketching with humans, and edit content via chat. + +![](images/660caff867d36bd3a2a1921ee55b7b3b76542d9b1364d0157c87333634aa747d.jpg) +Human-Agent Collaborative Sketching + +![](images/bb89323786ba343d8e554d354ea822316c073446fd92f4ecc14a65f0ac735d29.jpg) +Chat-Based Sketch Editing + +Draw a robot + +Add a paintbrush and a canvas + +![](images/58090f024aca3cd9a38c512e69f40fb0834a6bf929eb5b573f55723bd62c16a8.jpg) + +Draw a robot on the canvas + +![](images/754aa295f56e6b4570d9a42392daf32d422e9c28aa0ecbf726d93912ff19e24f.jpg) + +# Abstract + +Sketching serves as a versatile tool for externalizing ideas, enabling rapid exploration and visual communication that spans various disciplines. While artificial systems have driven substantial advances in content creation and human-computer interaction, capturing the dynamic and abstract nature of human sketching remains challenging. In this work, we introduce SketchAgent, a language-driven, sequential sketch generation method that enables users to create, modify, and refine sketches through dynamic, conversational interactions. Our approach requires no training or fine-tuning. Instead, we leverage the sequential nature and rich prior knowledge of off-the-shelf multimodal large language models (LLMs). We present an intuitive sketching language, introduced to the model through in-context examples, enabling it to "draw" using string-based actions. These are processed into vector graphics and then rendered to create a sketch on a pixel canvas, which can be accessed again for further tasks. By drawing stroke by stroke, our agent captures the evolving, dynamic qualities intrinsic to sketching. We demonstrate that SketchAgent can generate sketches from diverse prompts, engage in dialogue-driven drawing, and collaborate meaningfully with human users. + +# 1. Introduction + +Sketching is a powerful tool for distilling ideas into their simplest form. Its fluid and spontaneous nature makes sketching a uniquely versatile tool for visualization, rapid ideation, and communication across cultures, generations, and disciplines [26, 102]. For example, designers use sketches to explore new ideas [39, 103], scientists employ them to formulate problems [48, 72], and children engage in sketching to learn and express themselves [27, 28] (see Fig. 2). Artificial systems, in principle, have the potential to support and enhance human creativity, problem-solving, and visual expression through sketching, adapting flexibly to their exploratory nature [22, 98, 122]. + +Traditionally, sketch generation methods rely on human-drawn datasets to train generative models [5, 6, 16, 36, 42, 59]. However, fully capturing the diversity of sketches within datasets remains challenging [26], limiting these methods in both scale and diversity. Recent advancements in vision-language models, such as CLIP [78] and text-to-image diffusion [82], have enabled sketch generation methods that reduce reliance on human-drawn datasets [29, 46, 105]. These methods leverage pretrained model guidance and differentiable rendering [58] to optimize parametric curves, creating sketches that go beyond predefined styles and categories. + +While representing a significant step toward a general-purpose sketching system, these methods lack a crucial aspect of human drawing: the process itself. Current methods, though versatile, optimize all strokes simultaneously, making the intermediate sketching steps meaningless. As a result, the sketch cannot be decomposed into a coherent sequence of strokes that reflects the drawing process. In contrast, humans draw iteratively, stroke by stroke, incorporating visual feedback and continuously adapting—a dynamic, evolving process that fosters creativity, ideation, and communication [52, 88, 101]. + +In this work, we introduce SketchAgent, a sketch generation agent that leverages the prior knowledge and sequential nature of multimodal large language models (LLMs) to enable versatile, progressive, language-driven sketching. Our agent can generate sketches across a wide range of textual concepts—from animals to engineering principles (Fig. 1, left). Its sequential nature facilitates interactive human-agent sketching and supports iterative refinement and editing through a chat-based dialogue (Fig. 1, right). + +Unlike vision-language models that directly generate images from text [75, 80, 82], multimodal LLMs [1, 2, 15, 56, 64, 74, 97] accept text and images as input but only output text. To produce visuals, they either utilize external "tools" (such as calling a text-to-image model) or are prompted to generate executable code (e.g., Python [43], SVG [9]) to create charts, diagrams, or graphics. However, prompting for such representations to directly produce sketches often results in a mechanical appearance with uniform, precise shapes that lack the subtle irregularities and spontaneous qualities characteristic of human sketches (see Fig. 3B). Additionally, despite their robustness in textual tasks, these models often struggle with fine-grained spatial reasoning [41, 118] as they are primarily optimized for text, making sketch editing more challenging. + +To address these limitations, we introduce an intuitive sketching language that enables an off-the-shelf multimodal LLM agent to "draw" sketches on a canvas by providing string-based actions, without additional training or finetuning. We define the canvas as a numbered grid, allowing the agent to reference specific coordinates (e.g., x2y8) to enhance its spatial reasoning capabilities. We represent a sketch as a sequence of semantically meaningful strokes, each defined by a series of such coordinates. We leverage In-Context Learning (ICL) [7, 51] to introduce the agent to the new representation, and Chain of Thought (CoT) [108] to enhance its planning capabilities. Given a sketching task, the agent produces a textual response following our representation, which we process by fitting a smooth Bézier curve to each coordinate sequence. The curves are then rendered onto the canvas to form the final sketch. We find this approach useful in emulating a more natural sketch appearance. For collaborative sketching, the canvas remains acces + +![](images/79df84ee3c76e6a87bd277ede69140cb5c63521e15ceac8868d03a0ab43c608d.jpg) +(A) Frank Gehry. *Elevation Sketches*. Guggenheim Museum + +![](images/84445e52d6588ee0eb6c7ec6a9e37694b8fc18440ca5cc76762989cfe593d383.jpg) +(B) Alexander Bell's drawing of the telephone + +![](images/7f2b351f7e27a24b7b48e583019243e8bb3c15676cd7c4566913578b20efa11e.jpg) +(C) Children drawing + +![](images/094c36db6efb33f65a4fa9b7c75cc6344415ae800a93f97dbef517957c6988be.jpg) +(D) Basketball strategy (pick and roll) +Figure 2. Examples of sketches used across disciplines and goals. (A) Ideation and design: Process Elevation Sketches by the architect Frank Gehry, Guggenheim Museum. (B) Engineering: Alexander Bell's telephone drawing. (C) Expressing emotions: Children's sketches. (D) Visual communication: Planning and communicating game strategy in basketball. + +sible to both the user and the agent throughout the session. The agent generates strokes sequentially and pauses according to an adjustable stopping token, allowing the user to add their own strokes directly to the canvas. These strokes are then integrated into the agent's sequence, enabling it to continue drawing, with real-time canvas updates. + +We demonstrate SketchAgent's capability to generate sketches of diverse concepts while capturing the inherently sequential and dynamic nature of sketching. We showcase our agent's ability to collaborate effectively with humans in real time to create novel and meaningful sketches. Our method is the first to leverage pretrained multimodal LLMs for sequential sketching without additional training, paving the way for a general-purpose artificial sketching system that supports iterative, evolving interactivity. + +# 2. Related Work + +Sketch Generation Early methods approached sketch generation by designing image filters to simulate sketchlike effects [10, 109]. With the advent of deep learning, data-driven approaches emerged to address a range of sketch-related tasks [117], including category-conditioned sketching [42, 76, 93], object sketching [59, 62], scenesketching [12, 57, 60, 114], sketch completion [6, 63, 94], portrait drawing [3, 120, 121], part-based generation [6, 37, 42, 129], and more. While sketch data collection has been broadly explored [21, 34, 40, 71, 85, 113], the wide variation in sketch styles and their adaptation to specific tasks [23] makes collecting datasets that encompass this diversity challenging. For example, QuickDraw [47], the largest available sketch dataset with 50 mil + +lion sketches, covers only 345 object categories and primarily focuses on simple, iconic representations. This limits data-driven methods to the style, abstraction level, and concepts seen during training. Recently, large pretrained vision-language models [75, 78, 80, 82, 84] have shown remarkable text-to-image generation capabilities by leveraging extensive visual knowledge from billions of training images [89]. While these models can be prompted to generate sketch-like images (see Fig. 3A), they do so in a single step and in pixel space, lacking the sequential, stroke-based process of human sketching. Subsequent approaches [14, 29, 31, 46, 105, 106, 115, 116, 124] leverage the priors of these models to guide an iterative optimization of parametric curves, with a differentiable rasterizer [58] linking pixel and vector representations. While producing vector sketches, the final strokes lack order and semantic meaning, and the optimization-based approach overlook the sequential aspect of the sketching process, making these methods suboptimal for collaborative sketching. + +Sequential and Collaborative Sketching Collaborative human-machine sketching holds promise in enhancing creativity, ideation, communication, and learning, as explored in various fields, including human-computer interaction (HCI) [17, 45, 49, 50, 53, 54], computer graphics [55, 96], robotics [86, 87], cognitive science [24, 25, 35, 67], learning sciences [18, 38, 104], and more. Central to collaborative sketching is its sequential, adaptive, and dynamic process, with each action carrying intent. Existing methods employ diverse training strategies to account for the discrete nature of sequential sketches, including reinforcement and adversarial learning [32, 68, 129], multi-agent referential games [69, 77], transformers [5, 6, 11, 33, 61, 81, 112], and more. SketchRNN [42] is a pioneering work in this area, introducing the QuickDraw dataset [47], a crowd-sourced collection of real-time sketch sequences made by users. They utilize this dataset to train a recurrent neural network for sequential sketch generation, which was later shown [24, 73] to have potential for human-machine collaboration. However, this approach remains constrained by the predefined categories encountered during training. + +Multimodel LLMs for Content Creation LLMs [7, 19, 79, 100] and multimodal LLMs [1, 2, 15, 56, 64, 74, 97] receive text as input (or text and images for multimodal) and output text. To enable visual content generation, these models are often paired with external "tools" that extend their functionality [44, 90, 111, 119]. For example, ChatGPT [74] generates images by internally calling a separate model, DALLE-3 [4]. Another approach involves prompting models to produce code in languages like Python [43], Processing [92], SVG [9], or TikZ [8] that can be rendered into visuals such as graphs, charts, and vector graphics. However, such code-generated content often looks + +![](images/ecebd173d5244262321496c10bc7579f13a3ede01ed0683c1bba5b4ff4bfd7d6.jpg) +(A) DALLE3 [4] + +![](images/269d98ca61721ce69d816b986175812dc51676e984fa2d9bdfd596c779d1bca9.jpg) +(B) LLMs [2] (SVG) + +![](images/26f29f516ec09b7cad0c0384e589dc27ba4c5f593fda2154d43f5b99be1eeb9b.jpg) +(C) SketchAgent + +![](images/30f937880ee2cde64f8b31b9bf39dc243cda3140563dd4a93f0198d2aa2d28be.jpg) +(D) Human [47] +Figure 3. Sketch appearance. (A) Text-to-image diffusion models operate in pixel space, lacking the sequential nature of sketches. (B) Prompting LLMs to produce visuals with SVG results in a uniform, mechanical appearance. (C) Sketches produced by our agent appear less mechanical, more closely resembling the nature of (D) Human sketches, which are often spontaneous and irregular. + +rigid, with uniform and overly precise shapes that lack the subtle irregularities and spontaneous qualities characteristic of freehand sketches (see Fig. 3B). In contrast, we propose a sketching language grounded in spatial information that encourages the model to produce a more natural sketch appearance, which we then process into vector graphics. Common strategies for enhancing LLMs capabilities include Chain-of-Thought prompting [13, 70, 83, 91, 127], which breaks down tasks into smaller, logical steps to mimic human reasoning, and In-Context Learning (ICL) [7, 20, 95, 123, 125], where examples of input-output pairs are provided to help the model infer task patterns. + +# 3. Preliminaries + +Vector Graphics and Bezier Curves Vector graphics allow us to create visual images directly from geometric shapes such as points, lines, curves, and polygons. Unlike raster images (represented with pixels), vector graphics are + +![](images/cfd520c98eb43f76e42eb72ff93f21bfc86c01d6fb8de66fb635b103b300bc01.jpg) +Figure 4. Cubic Bezier curve. + +resolution-free, more compact, and editable. SVG [110] is an XML-based format for storing vector graphics, popular for its scalability and compatibility with modern web browsers. The process of transferring vector graphics into pixel images is called rasterization or rendering. Cubic Bézier curves are commonly used to represent sketches in vector graphics. A cubic Bézier curve (Fig. 4) is a smooth parametric curve defined by four points: a start point $P_0$ , an end point $P_3$ , and two control points $P_1$ and $P_2$ that shape the curvature. The set $P = \{P_0, P_1, P_2, P_3\}$ is often referred to as the curve's control points. The curve is described by the following polynomial equation: + +$$ +B (t) = (1 - t) ^ {3} P _ {0} + 3 (1 - t) ^ {2} t P _ {1} + 3 (1 - t) t ^ {2} P _ {2} + t ^ {3} P _ {3}, (1) +$$ + +where $t \in [0,1]$ is a parameter that moves the point along the curve from $P_0$ at $t = 0$ to $P_3$ at $t = 1$ . + +![](images/0f06d8a6a6aa9521e66a99348e8b287aae238090cc82646635cf1398b7363315.jpg) +Figure 5. Method Overview. SketchAgent (blue) receives drawing instructions and generates a string representing the intended sketch. Inputs include: (1) a system prompt (orange) introducing the sketching language and canvas, (2) a user prompt (pink) specifying the task (e.g., "draw a shark"), and (3) a numbered canvas. The agent's response outlines a sketching strategy (in thinking tags) and a sequence of strokes defined by coordinates, which are processed into Bézier curves and rendered onto the canvas. + +# 4. Method + +Our goal is to enable an off-the-shelf pretrained multimodal LLM to draw sketches based on natural language instructions. An overview of our pipeline is illustrated in Fig. 5. We utilize a frozen multimodal LLM ("SketchAgent" shown in blue), which receives three inputs: (1) a system prompt containing guidelines for using our new sketching language, (2) a user prompt with additional task-specific instructions (e.g., "Draw a shark"), and (3) a blank canvas on which the agent can draw. Based on the given task, the agent generates a textual response, representing the sequence of strokes to be drawn, which we then process into vector graphics and render onto the canvas. The canvas can then be reused in two ways: it can be fed back into the model with an updated user prompt for additional tasks and editing, or it can be accessed by a human user who can draw directly on it to facilitate collaborative sketching. Next, we describe each component of the pipeline. + +The Canvas Although multimodal LLMs demonstrate remarkable reasoning abilities, they often struggle with spatial reasoning tasks [30, 66, 99]. We present a simple example (see Fig. 6) to illustrate how this limitation affects the naive use of these models for sketch generation and interactive sketching. We provide GPT-4o [74] with an image depicting a simple line drawing of a partial house featuring five numbered points (from 1 to 5), and ask it to identify which points should be connected to complete the house. While the model correctly identifies the pair of points, it fails to select the correct pixel coordinates when given a basic draw-line tool that connects two points, even after multiple attempts. To enhance the model's spatial reasoning ability, we utilize a numbered canvas that forms a grid. This grid features numbers (1 to 50) along the x-axis and the y-axis (Fig. 5, left). Each cell is uniquely identified by a combination of the corresponding x-axis and y-axis numbers (e.g., the bottom-left cell is $x1y1$ ). The agent interacts with the canvas by specifying desired $(x,y)$ coordinates. + +![](images/eaff46c17f5e0ca25ebf8cfe4fbb32a448f235608a9bfa2adcfa2d90231fce5b.jpg) +Figure 6. Although excelling in visual reasoning, multimodal LLMs often struggle to translate these abilities into spatial actions. In this example, GPT-4o [74] intends to draw a line between points 1 and 5 but fails to execute this with a draw_line function that accepts pixel coordinates. + +Sketch Representation We define a sketch as a sequence of $n$ ordered strokes $S = \{S_1, S_2, \ldots, S_n\}$ . Each stroke $S_i$ is defined by a sequence of $m$ cell coordinates on the grid: $S_i = \{(x_j, y_j)\}_{j=1}^m$ , represented in string format as: x1y1, x15y20, .... + +A naive approach to processing the textual sequence of coordinates would be to use a polyline, connecting consecutive points with line segments. However, our grid-based representation sparsifies the canvas, resulting in a non-smooth and unnatural appearance when using polylines (see Fig. 7, left). To achieve a smoother appearance, an alternative approach is to treat the coordinates as a sequence of control points defining smooth curves. However, as illustrated in Fig. 4, the control points often do not lie directly on the curve. Consequently, if the agent aims for a stroke that passes through specific coordinates, it must derive the control points that define this stroke, which is challenging. + +We propose an alternative approach: we treat the specified $(x, y)$ coordinates as a set of desired points sampled along the curve, and fit a smooth Bézier curve to them (Fig. 7, right). To accommodate curves with complex curvature, we also task the model with determining when each point on the curve should be passed through, corresponding to the $t$ value described in Eq. (1). Thus, for each stroke $S_{i}$ , the agent provides a set of $m$ sampled points $S_{i} = \{(x_{j}, y_{j})\}_{j=1}^{m}$ , along with a corresponding set of $t$ values: $T_{i} = \{t_{j}\}_{j=1}^{m}$ . Based on these, we fit a cubic Bézier + +![](images/212fa6867ce4bdcbddaa82751742056c0944857271c1b4b8a0fc7031b67326ee.jpg) +(A) Polyline + +![](images/d04feb1e43212ca5927cdd3bfb132b6d589709712778a0904e76bc0a14265623.jpg) +Figure 7. Methods for processing the agent's coordinate sequence (in red): (A) Polyline results in an unnatural appearance. (B) Directly using coordinates as Bézier control points is challenging as they do not lie on the curve. (C) Fitting a Bézier curve to sampled coordinates provides smoother results. + +![](images/87bdbe0a578ba8382f0d291ffe00f966941310f2175de173f265c8c1c46094b2.jpg) +(B) Control Points +(C) Ours + +curve to the sampled points by solving a system of linear equations using least squares, where the unknowns are the control points $P = \{P_0, P_1, P_2, P_3\}$ : + +$$ +P = \operatorname {a r g m i n} _ {P} \| A P - B \|, \tag {2} +$$ + +where $A \in \mathbb{R}^{m \times 4}$ contains the cubic Bézier basis functions evaluated at specific $t_j$ values (as described in Eq. (1)), and $B \in \mathbb{R}^{m \times 2}$ contains the $m$ sampled points $\{(x_j, y_j)\}_{j=1}^m$ . The least squares solution minimizes the error between the fitted Bézier curve and the sampled points. For long sequences resulting in a large fitting error, we recursively split the curve. Additionally, we account for Bézier curves of lower degrees, including quadratic curves, linear lines, and points. Upon completing this process, we render the parametric curves onto the canvas. + +Drawing Instructions We provide the model with a system prompt and a user prompt (marked in orange and pink in Fig. 5). In the system prompt, we supply the agent with context about its expertise ("You are an expert artist specializing in drawing sketches") and introduce it to the grid canvas along with examples of how to use our sketching language for drawing single-stroke primitives (full prompts are provided in the Appendix). The system prompt is fixed and can be applied to a variety of sketching tasks. The user prompt includes a description of the desired task and an example of a simple sketch of a house drawn with our sketching language. We find this to be crucial in assisting the agent with preserving the correct format that could be parsed directly [7]. The agent is tasked with responding in the format shown in the gray text box in Fig. 5. In the tags, the agent is tasked to outline the overall sketching strategy [108]. This typically includes describing the different components of the sketch, the intended sketching order, and the overall placement of each part. The agent is also tasked with providing an ID tag following each stroke, which is useful for further analysis and for producing annotated sketches in scale. + +# 4.1. In-Chat Editing and Collaborative Sketching + +The above process can be repeated iteratively to support multiple sketching tasks and interactions. Text-based sketch editing in a chat dialogue is enabled by feeding the rendered + +![](images/95f2cb4f29aad5591b1e0194e2cf0ab7526e4b5e016af1d9e988b66435856fc1.jpg) + +![](images/2ce6c74333131c52fd4d2db63f898a49a0e211208a5a003b97036fbaedfe1a24.jpg) + +![](images/a200c7aa52d21ba8b21482f41ac05341be18a89a9ccba901fa92fd467b789848.jpg) + +![](images/1bd05a8d61780921afe2f41ff9d85cf2dc4418c435896e452d678996aeb9e97a.jpg) +DNA Double Helix + +![](images/deab0063155c5254908d052613efaa75c2a6d154ebe3baeeead4c1ce067194f5.jpg) +Golden Gate Bridge + +![](images/e1c42a497153ef5a8dfb3a8737e16da2cabd3d437ad02ee06a5551aba0e1d2af.jpg) +Mount Fuji +Pendulum Motion +Double-Slit Experiment + +![](images/6fd0668cd32e45b057920fd584981d956b47ad05709d30cd8c47068995420b89.jpg) +Eiffel Tower +Flowchart +Figure 8. Sketches produced by SketchAgent for concepts beyond pre-defined categories. The textual input describing the desired concept shown below each image. + +![](images/661e67bb48b898e26fae04c7637ad316784dbe6caf1e6bdcdac6af5e7b98976f.jpg) +Tic-tac toe + +canvas back to the agent (see dashed arrow in Fig. 5) and updating the user prompt with the desired edits. To support collaborative human-agent sketching, the canvas remains accessible to both the human user and the agent throughout the entire sketching session. We define an adjustable stopping token, $\langle \text{}/\text{s}\{\text{j}\} \rangle$ , which instructs the agent to pause generating the sequence at stroke number $j$ . We then process and render the generated strokes onto the canvas up to that point, then the user can add strokes directly to the canvas to continue the sketch. The user-drawn strokes are processed and converted into the agent's format by reversing our fitting process, i.e., sampling each stroke at multiple $t$ values (as shown in Eq. (1)), and selecting the points closest to each cell's center on the grid. The converted user strokes are then chained with the agent's sequence, after which the agent resumes sketching until the next stopping token. + +# 5. Results + +We evaluate the performance of our method qualitatively and quantitatively across a selected set of sketching tasks. Additional tasks, evaluations, and examples are provided in the Appendix. All results presented in the paper were generated using Claude3.5-Sonnet [2] as our backbone model, unless stated otherwise. + +# 5.1. Text-Conditioned Sketch Generation + +Figures 1 and 8 demonstrate SketchAgent's capability to generate sketches of various concepts that extend beyond standard categories, which includes scientific concepts (e.g., "the double-slit experiment", "pendulum motion"), diagrams (e.g., "circuit diagram", "a flowchart"), and notable landmarks (e.g., "Taj Mahal", "Eiffel Tower"). More examples are provided in the Appendix. To quantitatively evaluate text-conditioned generation we utilize the QuickDraw dataset [47]. We randomly sample 50 categories (out of 345), and apply our method to generate 10 sketch in + +
GPT-4oGPT-4o -miniClaude3 OpusClaude3.5 -Sonnet*Claude3.5 -Sonnet (SVG)Human (QD [47])
Top10.15 ±0.040.04 ±0.030.13 ±0.040.23 ±0.050.23 ±0.040.27 ±0.07
Top50.30 ±0.060.10 ±0.040.27 ±0.050.44 ±0.030.43 ±0.060.49 ±0.06
Vis.
+ +Table 1. Sketch recognition evaluation. Average Top-1 and Top-5 sketch recognition accuracy computed with CLIP zero-shot classifier on 500 sketches from 50 categories. The last row visualizes one sample from each experiment. *Indicates our default settings, which receives the highest accuracy among all models. + +stances per category, resulting in 500 sketches in total. Following common practice [105, 106, 115, 116], we utilize a CLIP zero-shot classifier [78] to evaluate how well the generated sketches depict the intended category. We compare the performance of different multimodal LLMs by repeating the same process with GPT-4o-mini [74], GPT-4o [74], and Claude3-Opus [2] as our backbone model (in addition to Claude3.5-Sonnet [2], our default backbone). As a baseline, we include human-drawn sketches sampled from the QuickDraw dataset [47]. The average Top-1 and Top-5 sketch classification accuracy are presented in Table 1. As can be seen, human sketches achieve the highest recognition accuracy, with Claude3.5-Sonnet performing best among all models, approaching human-level rates under the CLIP-score metric. More evaluation of confusion patterns and visualization of the data are provided in the Appendix. + +We additionally compare to prompting Claude3.5-Sonnet to directly generate SVGs using the following prompt: "Write SVG string that draws a sketch of a . Use only black and white colors". The corresponding scores are shown in the fifth column of Tab. 1. While this approach achieves recognition scores comparable to those of SketchAgent, the outputs are often characterized by uniform and precise shapes, failing to replicate the fluidity and natural irregularity of free-hand human sketches (e.g., Fig. 3). To evaluate how "human-like" our agent's sketches appear, we conduct a two alternative forced choice (2AFC) user study with 150 participants. Each participant was presented with pairs of sketches depicting the same object class produced by different methods, and asked to choose the sketch they believed was human-drawn. 150 sketches across 50 object classes were tested, comparing three methods: direct prompting, SketchAgent, and human sketches from QuickDraw (see Appendix for details). Results indicate SketchAgent's drawings appeared more human-like, being chosen as human-drawn in $74.90 \pm 3.35\%$ of cases when compared with direct prompting. When compared to human drawings, + +![](images/423a11037c664387b1a182431eb6f4eedd6120c7e5bf6f031d5dfdee35b62b69.jpg) +Figure 9. SketchAgent gradually draws stroke-by-stroke, each stroke is annotated by the agent with a semantic meaning. + +![](images/04d463e6f03e007611601410ebbcd7f8b365b30eb1b0c6400a7333a2068fb197.jpg) +Figure 10. Recognition rate as a function of sketch complexity. + +![](images/07f9d2c145aadce7dd680e4e50f8d2970cc25e1e13d6710c4e467d2cc83efedf.jpg) + +![](images/969d296403a23d3904cb4e0625589d27622dce0da84e52d44ebd61be07008104.jpg) + +users slightly preferred human sketches $(54.68 \pm 4.61\%)$ over SketchAgent's, while direct prompting was chosen only $38.9 \pm 5.55\%$ of the time. + +Lastly, to quantitatively analyze the effect of concept complexity, we study the case of drawing letters. We systematically increase sketch complexity by adding letters to the target concept (e.g., from 'ABC' to 'ABCDEFGH') and count the number of correctly recognized letters in each sketch. The graph in Fig. 10 shows that performance decreases as the complexity increases. + +# 5.2. Sequential Sketching + +Figure 9 shows stroke-by-stroke sketch generation by SketchAgent, with the labels on the right indicating the sketching order and the meaning our agent associates with each generated stroke (see Appendix for more examples). Stroke annotation during generation is enabled by utilizing the prior of the backbone LLM, providing a valuable feature for analysis and data collection [37, 65, 107, 126, 128]. In Fig. 11, we illustrate why accounting for the sequential nature of sketching more closely emulates the process of human drawing. We present the sketch creation process of SketchAgent alongside SVGDreamer [116], SketchRNN [42], and a human sketch sampled from QuickDraw [47]. SVGDreamer (first row), is an optimization-based method, where a set of randomly initialized parametric curves (left-most column) are iteratively refined to form a sketch, guided by a pretrained text-to-image diffusion model [82]. This process is time-consuming, taking 2000 iterations (1.6 hours), which makes it unsuitable for interactive sketching. While the final sketch (rightmost column) appears detailed and artistic due to the powerful vision backbone, the intermediate sketching and individual strokes lack clear semantic meaning. In contrast, SketchRNN (second row) is a sequential generative model trained on human-drawn dataset, producing sketches in real-time with strokes added progressively, emulating closer a human-like sketching pro + +SVGDreamer [116]. $\approx 1.6$ hours + +
SketchRNN [42].≈4 seconds
SketchAgent (Ours).≈20 seconds
bodyheadleft earright earback right legleft eye
Human [47].≤20 seconds
+ +Figure 11. Sequential sketching process. SVGDreamer [116] requires 2000 iterations (1.6 hours) with intermediate steps lacking semantic meaning. SketchRNN [42] operates in real-time with coherent steps but is limited to QuickDraw categories. SketchAgent draw gradually with meaningful strokes and no category restrictions. Human sketches evolve through gradual, meaningful steps. + +cess (as shown in the last row). Similarly, SketchAgent (third row) produces sketches gradually, with each stroke carrying a semantic meaning, by utilizing the sequential nature of its backbone model. While SketchRNN is restricted to generating sketches only within the 345 categories it was trained on, SketchAgent leverages the extensive prior knowledge of its backbone multimodal LLM, enabling it to create sketches of general visual concepts. + +We use the set of 500 samples described in Sec. 5.1 to quantitatively analyze the sequential nature of our agent's sketches compared to human drawings. On the left of Fig. 12, we present histograms comparing the number of strokes in QuickDraw sketches (orange) and our sketches (blue). Most QuickDraw sketches contain 1 to 6 strokes, while our sketches show a broader distribution, peaking between 5 to 10 strokes. This suggests that, on average, QuickDraw sketches appear more abstract. To ensure a balanced comparison of sketches with similar levels of abstraction, we select sketches from both groups with a similar number of strokes (the largest intersection is found in sketches with 4-7 strokes, comprising 204 of our sketches and 120 from QuickDraw) and measure the change in CLIP-Score as a function of the accumulated number of strokes (Fig. 12, right). Both QuickDraw and our sketches exhibit a generally similar pattern, with CLIPScore increasing as more strokes are added, suggesting that sketches become progressively more recognizable as they evolve. + +![](images/8bb79ce4851de1d59ede2f9298fd100f1c2e84e32cba837bddaa488f0ba4591f.jpg) +Figure 12. Sequential sketching analysis of SketchAgent (blue) and Humans [47] (orange). Left: Histograms of stroke distribution per sketch, showing QuickDraw sketches are more abstract on average. Right: CLIPScore as a function of the accumulated number of strokes for sketches containing 4-7 strokes, showing a similar recognition pattern over time. + +![](images/df5f4609b0c441a7467c3969f5a4245622328392296f8f901211d83ac22edf94.jpg) + +# 5.3. Human-Agent Collaborative Sketching + +We demonstrate the potential of our system for facilitating interactive human-agent collaboration, resulting in semantically meaningful and recognizable sketches. We design a web-based collaborative sketching environment (Fig. 13A) where users and SketchAgent take turns drawing on a shared canvas to create a recognizable sketch from a given textual concept. Following the evaluation protocol in collabdraw [24], we select 8 simple concepts, based on the agent's demonstrated ability to draw them independently, to focus evaluations on assessing the impact of collaboration. Participants sketched concepts in two modes: solo, where users drew independently, and collab, where users and SketchAgent collaborated, adding one stroke at a time until either was satisfied with the drawing. We collect sketches from 30 participants, resulting in 480 sketches in total. Average CLIP recognition rates are shown in Figure 13B. Collaboratively produced sketches (blue) achieve recognition levels close to those made solely by users and higher than those produced by the agent alone (dashed lines). To assess the contribution of each party in collaborative mode, we analyze partial sketches with only agent-made strokes (pink) or user-made strokes (green), resulting in a significant reduction in recognizability. This suggests that both user and agent contribute meaningfully to the recognizability of the complete sketch. + +# 5.4. Chat-Based Sketch Editing + +We next demonstrate the effectiveness of our method in performing interactive text-based sketch editing within a chat dialogue, where the input to the agent combines both text and images. Inspired by [92], we explore edits that involve spatial reasoning and object relations. We focus on three object categories: outdoor, indoor, and animals, with three objects each, and design editing prompts to add objects to the input sketches. For outdoor and indoor objects, we specify relative locations of added concepts, e.g., "left to", "on top of" (see Fig. 14 left). For the animals category, we tasked + +![](images/1ef282de04d388ff3cea584cc8f62439859b8bb456a0b0f83105927e7b99ad08.jpg) +(A) Sketching interface + +![](images/9b467c1d65280845bea59527f26ef0b501d740d2b6a7852e9a026da3abbfb9b5.jpg) +(B) Collaborative user study results + +![](images/ce280e778c673c55a5fab06f36f56a5a8fff409fc90b55e02f168b8aa652fb50.jpg) +Figure 13. Collaborative sketching evaluation measured using CLIP classification. Sketches created collaboratively (blue) approaching those made solely by users (dashed lines). In collaborative sketches, keeping agent-only strokes (pink) or user-only strokes (green) significantly reduces recognizability. +Figure 14. Chat-based sketch editing. We iteratively prompt SketchAgent to add objects to sketches through chat dialogues. + +the agent with adding accessories to each animal without guidance on their exact placement, testing its ability to infer placement based on semantics (e.g., placing a hat on a head (see Fig. 14 right). The full list of object and editing instructions is provided in the Appendix. We produced a total of 54 sketches. Evaluating the edited sketches reveals that SketchAgent correctly follows instructions $92\%$ of the time, with $94\%$ accuracy for specified relations and $88\%$ accuracy for inferred semantic relations. + +# 6. Ablation + +We evaluate the impact of each component of our method by systematically removing them and measuring sketch recognition rates as detailed in 5.1. We assess the effects of removing the system prompt, omitting the CoT process (i.e., excluding thinking tags and 'think step-by-step' instructions), and modifying ICL (the complete sketch example provided in the user prompt). When modifying ICL, we use a correctly formatted single-stroke example instead of the complete sketch, as fully removing ICL results in outputs that do not follow the expected format making them unparsable. The results in Table 2 show that the full SketchAgent pipeline achieves the highest performance, highlighting the importance of each component. Interestingly, not providing a complete sketch example significantly reduces performance. We additionally ablate the impact of the grid resolution, by varying the resolution from 10 to 100 (see Tab. 2, bottom). Extremely low resolutions degrade performance, while mid-level resolutions outperform $100 \times 100$ . + +
w/o System Promptw/o CoTModified ICLSketchAgent (full)
Top10.20 ± 0.040.14 ± 0.020.07 ± 0.020.23 ± 0.04
Top50.42 ± 0.030.29 ± 0.040.16 ± 0.030.43 ± 0.06
Grid Size10 × 1025 × 2550 × 5075 × 75
Top1 / Top50.14 / 0.280.19 / 0.420.23 / 0.430.23 / 0.41
+ +Table 2. Ablation study. Average Top-1 and Top-5 CLIP recognition accuracy. Top: We systematically remove each component in our pipeline, showcasing all components contribute to the agent's full performance. Bottom: Grid resolution ablation. + +# 7. Limitations and Future Work + +SketchAgent has several limitations. First, it is constrained by the priors of the backbone model, primarily optimized for text rather than visual content. As a result, the agent often produces rich textual descriptions of object parts but struggles to convert these into effective sketching actions, resulting in overly abstract and unrecognizable outputs. For example, in Fig. 15A, the agent effectively describes key parts of a unicorn (e.g., the horn), but the sketch is unrecognizable. This constraint also impacts the depiction of human figures (Fig. 15B). While distinctive features (e.g., Frida Kahlo's eyebrows or Michael Jordan's dunk) may be captured well in language, the resulting sketches are overly simple, with an amateur style, lacking expressivity. We expect this issue to improve as future models advance in vision capabilities. Lastly, the agent may struggle with drawing letters and numbers. This could be improved in future work by providing relevant in-context examples. + +![](images/ab03659d7c4d1a51dd5884bfa72ac04901f6940c37db0ebced6a7e7e0884155b.jpg) +(A) Non-recognizable + +![](images/a5198ad87ecffd7972bcfcf56f48b5c9045d4741a2ec81cdda15348a70a826d5.jpg) +(B) Human figures +Figure 15. Limitations. Sketches of complex concepts (A) and human figures (B) appear too abstract and unrecognizable with non-professional style. (C) Fail to depict letters and numbers. + +![](images/ad6c20de7a2d320b68f6ae43926ba1745d7421378648d09adf4f1b97e1df2888.jpg) +(C) Letters and numbers + +# 8. Conclusions + +We presented a method for language-driven, sequential sketch generation, that can produce versatile sketches in real-time and meaningfully engage in collaborative sketching sessions with humans. We show that the prior knowledge embedded in pretrained multimodal LLMs can be effectively leveraged for sketch generation through an intuitive sketching language and a grid canvas, without requiring additional training or fine-tuning. We hope our work represents a meaningful step toward developing general-purpose sketching systems with the potential to enhance human-computer communication and computer-aided ideation. + +# 9. Acknowledgements + +We thank Yuval Alaluf, Hila Chefer, Assaf Ben Kish, Joanna Materzynska, Rinon Gal, Elad Richardson, Arnav Verma, and Ellie Arar for providing feedback on early versions of our manuscript. We are also grateful to Justin Yang, who developed the initial Photodraw application and assisted with the interactive human study setup, and to Dingning Cao, who contributed to the SketchAgent interface and application. Special thanks to Yarden Frenkel for his insights, early explorations, and inspiring discussions. This work was partially supported by NSF CAREER #2047191, NSF DRL #2400471, Stanford Human Centered AI Institute Hoffman-Yee Grant, Hyundai Motor Company, ARL grant W911NF-18-2-021, the Zuckerman STEM Leadership Program, and the Viterbi Fellowship. The sponsors had no role in the experimental design or analysis, the decision to publish, or manuscript preparation. The authors have no competing interests to report. + +# References + +[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katie Millicah, Malcolm Reynolds, Roman Ring, Eliza Rutherford, Serkan Cabi, Tengda Han, Zhitao Gong, Sina Samangooei, Marianne Monteiro, Jacob Menick, Sebastian Borgeaud, Andrew Brock, Aida Nematzadeh, Sahand Sharifzadeh, Mikolaj Binkowski, Ricardo Barreira, Oriol Vinyals, Andrew Zisserman, and Karen Simonyan. Flamingo: a visual language model for few-shot learning. Red Hook, NY, USA, 2024. Curran Associates Inc. 2, 3 +[2] Anthropic. Claude. https://www.anthropic.com/claude, 2023. 2, 3, 5, 6 +[3] Itamar Berger, Ariel Shamir, Moshe Mahler, Elizabeth Carter, and Jessica Hodgins. Style and abstraction in portrait sketching. ACM Trans. Graph., 32(4), 2013. 2 +[4] James Betker, Gabriel Goh, Li Jing, † TimBrooks, Jianfeng Wang, Linjie Li, † LongOuyang, † JuntangZhuang, † JoyceLee, † YufeiGuo, † WesamManassra, † PrafullaDhariwal, † CaseyChu, † YunxinJiao, and Aditya Ramesh. Improving image generation with better captions. 2023. 3 +[5] Ayan Kumar Bhunia, Ayan Das, Umar Riaz Muhammad, Yongxin Yang, Timothy M. Hospedales, Tao Xiang, Yulia Gryaditskaya, and Yi-Zhe Song. Pixel: a competitive sketching ai agent. so you think you can sketch? ACM Trans. Graph., 39:166:1-166:15, 2020. 1, 3 +[6] Ankan Kumar Bhunia, Salman Khan, Hisham Cholakkal, Rao Muhammad Anwer, Fahad Shahbaz Khan, Jorma Laaksonen, and Michael Felsberg. Doodleformer: Creative sketch drawing with transformers. ECCV, 2022. 1, 2, 3 +[7] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel Ziegler, Jeffrey Wu, Clemens Winter, Chris Hesse, Mark Chen, Eric + +Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. Language models are few-shot learners. In Advances in Neural Information Processing Systems, pages 1877-1901. Curran Associates, Inc., 2020. 2, 3, 5 +[8] Sébastien Bubeck, Varun Chandrasekaran, Ronen Eldan, Johannes Gehrke, Eric Horvitz, Ece Kamar, Peter Lee, Yin Tat Lee, Yuanzhi Li, Scott Lundberg, et al. Sparks of artificial general intelligence: Early experiments with gpt-4. arXiv preprint arXiv:2303.12712, 2023. 3 +[9] Mu Cai, Zeyi Huang, Yuheng Li, Haohan Wang, and Yong Jae Lee. Delving into LLMs' visual understanding ability using SVG to bridge image and text, 2024. 2, 3 +[10] John Canny. A computational approach to edge detection. IEEE Transactions on pattern analysis and machine intelligence, (6):679-698, 1986. 2 +[11] Alexandre Carlier, Martin Danelljan, Alexandre Alahi, and Radu Timofte. Deepsvg: A hierarchical generative network for vector graphics animation, 2020. 3 +[12] Caroline Chan, Frédo Durand, and Phillip Isola. Learning to generate line drawings that convey geometry and semantics. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7915-7925, 2022. 2 +[13] Zhenfang Chen, Qinhong Zhou, Yikang Shen, Yining Hong, Zhiqing Sun, Dan Gutfreund, and Chuang Gan. Visual chain-of-thought prompting for knowledge-based visual reasoning. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1254-1262, 2024. 3 +[14] Changwoon Choi, Jaeah Lee, Jaesik Park, and Young Min Kim. 3doodle: Compact abstraction of objects with 3d strokes. ACM Trans. Graph., 43(4), 2024. 3 +[15] Xiangxiang Chu, Jianlin Su, Bo Zhang, and Chunhua Shen. Visionllama: A unified llama backbone for vision tasks, 2024. 2, 3 +[16] Ayan Das, Yongxin Yang, Timothy Hospedales, Tao Xiang, and Yi-Zhe Song. Béziersketch: A generative model for scalable vector sketches. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXVI 16, pages 632-647. Springer, 2020. 1 +[17] Nicholas Davis, Chih-PIn Hsiao, Kunwar Yashraj Singh, Lisa Li, Sanat Moningi, and Brian Magerko. Drawing apprentice: An enactive co-creative agent for artistic collaboration. In Proceedings of the 2015 ACM SIGCHI Conference on Creativity and Cognition, page 185–186, New York, NY, USA, 2015. Association for Computing Machinery. 3 +[18] Vanessa de Andrade, Sofia Freire, Monica Baptista, and Yael Shwartz. Drawing as a space for social-cognitive interaction. Education Sciences, 12:45, 2022. 3 +[19] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language + +Technologies, Volume 1 (Long and Short Papers), pages 4171-4186, Minneapolis, Minnesota, 2019. Association for Computational Linguistics. 3 +[20] Sivan Doveh, Shaked Perek, M Jehanzeb Mirza, Wei Lin, Amit Alfassy, Assaf Arbelle, Shimon Ullman, and Leonid Karlinsky. Towards multimodal in-context learning for vision & language models. arXiv preprint arXiv:2403.12736, 2024. 3 +[21] Mathias Eitz, James Hays, and Marc Alexa. How do humans sketch objects? ACM Trans. Graph., 31(4), 2012. 2 +[22] Ziv Epstein, Aaron Hertzmann, Laura Mariah Herman, Robert Mahari, Morgan R. Frank, Matthew Groh, Hope Schroeder, Amy Smith, Memo Akten, Jessica Fjeld, Hany Farid, Neil Leach, Alex Pentland, and Olga Russakovsky. Art and the science of generative ai. Science, 380:1110-1111, 2023. 1 +[23] Judy Fan, Wilma A. Bainbridge, Rebecca Chamberlain, and Jeffrey D. Wammes. Drawing as a versatile cognitive tool. Nature Reviews Psychology, 2:556 - 568, 2023. 2 +[24] Judith E. Fan, Monica Dinculescu, and David Ha. collaborate: An environment for collaborative sketching with an artificial agent. In Proceedings of the 2019 Conference on Creativity and Cognition, page 556-561, New York, NY, USA, 2019. Association for Computing Machinery. 3, 7 +[25] Judith E. Fan, Robert D. Hawkins, Mike Wu, and Noah D. Goodman. Pragmatic Inference and Visual Abstraction Enable Contextual Flexibility During Visual Communication. Computational Brain & Behavior, 3(1):86-101, 2020. 3 +[26] Judith E Fan, Wilma A Bainbridge, Rebecca Chamberlain, and Jeffrey D Wammes. Drawing as a versatile cognitive tool. Nature Reviews Psychology, 2(9):556-568, 2023. 1 +[27] Logan Fiorella and Shelbi Kuhlmann. Creating drawings enhances learning by teaching. Journal of Educational Psychology, 112(4):811, 2020. 1 +[28] Kenneth Forbus, Jeffrey Usher, Andrew Lovett, Kate Lockwood, and Jon Wetzel. Cogsketch: Sketch understanding for cognitive science research and for education. Topics in Cognitive Science, 3(4):648-666, 2011. 1 +[29] Kevin Frans, L. B. Soros, and Olaf Witkowski. Clipdraw: exploring text-to-drawing synthesis through language-image encoders. In Proceedings of the 36th International Conference on Neural Information Processing Systems, Red Hook, NY, USA, 2024. Curran Associates Inc. 1, 3 +[30] Xingyu Fu, Yushi Hu, Bangzheng Li, Yu Feng, Haoyu Wang, Xudong Lin, Dan Roth, Noah A Smith, Wei-Chiu Ma, and Ranjay Krishna. Blink: Multimodal large language models can see but not perceive. arXiv preprint arXiv:2404.12390, 2024. 4 +[31] Rinon Gal, Yael Vinker, Yuval Alaluf, Amit Bermano, Daniel Cohen-Or, Ariel Shamir, and Gal Chechik. Breathing life into sketches using text-to-video priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4325–4336, 2024. 3 +[32] Yaroslav Ganin, Tejas D. Kulkarni, Igor Babuschkin, S. M. Ali Eslami, and Oriol Vinyals. Synthesizing programs for images using reinforced adversarial learning. *ArXiv*, abs/1804.01118, 2018. 3 + +[33] Yaroslav Ganin, Sergey Bartunov, Yujia Li, Ethan Keller, and Stefano Saliceti. Computer-aided design as language. In Neural Information Processing Systems, 2021. 3 +[34] Chengying Gao, Qi Liu, Qi Xu, Limin Wang, Jianzhuang Liu, and Changqing Zou. Sketchycoco: Image generation from freehand scene sketches. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5174-5183, 2020. 2 +[35] Simon Garrod, Nicolas Fay, John Lee, Jon Oberlander, and Tracy MacLeod. Foundations of representation: Where might graphical symbol systems come from? Cognitive Science, 31(6):961-987, 2007. 3 +[36] Songwei Ge, Vedanuj Goswami, C Lawrence Zitnick, and Devi Parikh. Creative sketch generation. arXiv preprint arXiv:2011.10039, 2020. 1 +[37] Songwei Ge, Vedanuj Goswami, Larry Zitnick, and Devi Parikh. Creative sketch generation. In International Conference on Learning Representations, 2021. 2, 6 +[38] Hinnie Gijlers, Armin Weinberger, Alieke Mattia van Dijk, Lars Bollen, and Wouter van Joolingen. Collaborative drawing on a shared digital canvas in elementary science education: The effects of script and task awareness support. International Journal of Computer-Supported Collaborative Learning, 8(4):427-453, 2013. 3 +[39] Gabriela Goldschmidt. Serial sketching: visual problem solving in designing. Cybernetics and System, 23(2):191-219, 1992. 1 +[40] Yulia Gryaditskaya, Mark Sypesteyn, Jan Willem Hoftijzer, Sylvia Pont, Frédo Durand, and Adrien Bousseau. Opensketch: a richly-annotated dataset of product design sketches. ACM Trans. Graph., 38(6), 2019. 2 +[41] Tianrui Guan, Fuxiao Liu, Xiyang Wu, Ruiqi Xian, Zongxia Li, Xiaoyu Liu, Xijun Wang, Lichang Chen, Furong Huang, Yaser Yacoob, Dinesh Manocha, and Tianyi Zhou. Hallusionbench: An advanced diagnostic suite for entangled language hallucination and visual illusion in large vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14375-14385, 2024. 2 +[42] David Ha and Douglas Eck. A neural representation of sketch drawings. CoRR, abs/1704.03477, 2017. 1, 2, 3, 6, 7 +[43] Yucheng Han, China. Xiaoyan Zhang, Xin Chen, Xu Yang, Zhibin Wang, Gang Yu, Bin Fu, and Hanwang Zhang. Chartlama: A multimodal llm for chart understanding and generation. ArXiv, abs/2311.16483, 2023. 2, 3 +[44] Yushi Hu, Weijia Shi, Xingyu Fu, Dan Roth, Mari Ostendorf, Luke Zettlemoyer, Noah A Smith, and Ranjay Krishna. Visual sketchpad: Sketching as a visual chain of thought for multimodal language models. arXiv preprint arXiv:2406.09403, 2024. 3 +[45] Francisco Javier Ibarrola, Tomas Lawton, and Kazjon Grace. A collaborative, interactive and context-aware drawing agent for co-creative design. IEEE Transactions on Visualization and Computer Graphics, 30:5525-5537, 2022. 3 +[46] Ajay Jain, Amber Xie, and Pieter Abbeel. Vectorfusion: Text-to-svg by abstracting pixel-based diffusion models. In + +Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1911-1920, 2023. 1, 3 +[47] Jongejan Jonas, Rowley Henry, Kawashima Takashi, Kim Jongmin, and Fox-Gieg Nick. The Quick, Draw! - A.I. Experiment, 2016. 2, 3, 5, 6, 7 +[48] David Kaiser. Drawing theories apart: The dispersion of Feynman diagrams in postwar physics. University of Chicago Press, 2019. 1 +[49] Pegah Karimi, Jeba Rezwana, Safat Siddiqui, Mary Lou Maher, and Nasrin Dehbozorgi. Creative sketching partner: an analysis of human-ai co-creativity. In Proceedings of the 25th International Conference on Intelligent User Interfaces, page 221–230, New York, NY, USA, 2020. Association for Computing Machinery. 3 +[50] Pegah Karimi, Jeba Rezwana, Safat Siddiqui, Mary Lou Maher, and Nasrin Dehbozorgi. Creative sketching partner: an analysis of human-ai co-creativity. In Proceedings of the 25th International Conference on Intelligent User Interfaces, page 221–230, New York, NY, USA, 2020. Association for Computing Machinery. 3 +[51] Andrew Kyle Lampinen, Stephanie C. Y. Chan, Aaditya K. Singh, and Murray Shanahan. The broader spectrum of in-context learning, 2024. 2 +[52] Paul Laseau. Graphic thinking for architects and designers. John Wiley & Sons, 2000. 2 +[53] Tomas Lawton, Kazjon Grace, and Francisco J Ibarrola. When is a tool a tool? user perceptions of system agency in human-ai co-creative drawing. In Proceedings of the 2023 ACM Designing Interactive Systems Conference, page 1978–1996, New York, NY, USA, 2023. Association for Computing Machinery. 3 +[54] Tomas Lawton, Francisco J Ibarrola, Dan Ventura, and Kazjon Grace. Drawing with reframer: Emergence and control in co-creative ai. In Proceedings of the 28th International Conference on Intelligent User Interfaces, page 264–277, New York, NY, USA, 2023. Association for Computing Machinery. 3 +[55] Yong Jae Lee, C. Lawrence Zitnick, and Michael F. Cohen. Shadowdraw: real-time user guidance for freehand drawing. In ACM SIGGRAPH 2011 Papers, New York, NY, USA, 2011. Association for Computing Machinery. 3 +[56] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. BLIP: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In Proceedings of the 39th International Conference on Machine Learning, pages 12888-12900. PMLR, 2022. 2, 3 +[57] Mengtian Li, Zhe Lin, Radomir Mech, Ersin Yumer, and Deva Ramanan. Photo-sketching: Inferring contour drawings from images. In 2019 IEEE Winter Conference on Applications of Computer Vision (WACV), pages 1403-1412. IEEE, 2019. 2 +[58] Tzu-Mao Li, Michal Lukáč, Gharbi Michael, and Jonathan Ragan-Kelley. Differentiable vector graphics rasterization for editing and learning. ACM Trans. Graph. (Proc. SIGGRAPH Asia), 39(6):193:1-193:15, 2020. 1, 3 + +[59] Yi Li, Yi-Zhe Song, Timothy M. Hospedales, and Shaogang Gong. Free-hand sketch synthesis with deformable stroke models. CoRR, abs/1510.02644, 2015. 1, 2 +[60] Yijun Li, Chen Fang, Aaron Hertzmann, Eli Shechtman, and Ming-Hsuan Yang. Im2pencil: Controllable pencil illustration from photographs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1525-1534, 2019. 2 +[61] Hangyu Lin, Yanwei Fu, Yu-Gang Jiang, and X. Xue. Sketch-bert: Learning sketch bidirectional encoder representation from transformers by self-supervised learning of sketch gestalt. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6757-6766, 2020. 3 +[62] Difan Liu, Matthew Fisher, Aaron Hertzmann, and Evangelos Kalogerakis. Neural strokes: Stylized line drawing of 3d shapes. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14204-14213, 2021. 2 +[63] Fang Liu, Xiaoming Deng, Yu-Kun Lai, Yong-Jin Liu, Cuixia Ma, and Hongan Wang. Sketchgan: Joint sketch completion and recognition with generative adversarial network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2 +[64] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In Advances in Neural Information Processing Systems, pages 34892-34916. Curran Associates, Inc., 2023. 2, 3 +[65] Bria Long, Judith Fan, Holly Huey, Zixian Chai, and Michael Frank. Parallel developmental changes in children's production and recognition of line drawings of visual concepts. Nature Communications, 15, 2024. 6 +[66] Arjun Majumdar, Anurag Ajay, Xiaohan Zhang, Pranav Putta, Sriram Yenamandra, Mikael Henaff, Sneha Silwal, Paul Mcvay, Oleksandr Maksymets, Sergio Arnaud, Karmesh Yadav, Qiyang Li, Ben Newman, Mohit Sharma, Vincent Berges, Shiqi Zhang, Pulkit Agrawal, Yonatan Bisk, Dhruv Batra, Mrinal Kalakrishnan, Franziska Meier, Chris Paxton, Sasha Sax, and Aravind Rajeswaran. Openeqa: Embodied question answering in the era of foundation models. In Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 4 +[67] William P. McCarthy, Justin Matejka, Karl D.D. Willis, Judith E. Fan, and Yewen Pu. Communicating design intent using drawing and text. In Proceedings of the 16th Conference on Creativity & Cognition, page 512-519, New York, NY, USA, 2024. Association for Computing Machinery. 3 +[68] John FJ Mellor, Eunbyung Park, Yaroslav Ganin, Igor Babuschkin, Tejas Kulkarni, Dan Rosenbaum, Andy Ballard, Theophane Weber, Oriol Vinyals, and SM Eslami. Unsupervised doodling and painting with improved spiral. arXiv preprint arXiv:1910.01007, 2019. 3 +[69] Daniela Mihai and Jonathon Hare. Learning to draw: Emergent communication through sketching. Advances in Neural Information Processing Systems, 34:7153-7166, 2021. 3 + +[70] Chancharik Mitra, Brandon Huang, Trevor Darrell, and Roei Herzig. Compositional chain-of-thought prompting for large multimodal models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14420-14431, 2024. 3 +[71] Kushin Mukherjee, Holly Huey, Xuanchen Lu, Yael Vinker, Rio Aguina-Kang, Ariel Shamir, and Judith Fan. Seva: Leveraging sketches to evaluate alignment between human and machine visual abstraction. In Advances in Neural Information Processing Systems, 2023. 2 +[72] Omar W Nasim. Observing by hand: sketching the nebulae in the nineteenth century. University of Chicago Press, 2019. 1 +[73] Changhoon Oh, Jungwoo Song, Jinhan Choi, Seonghyeon Kim, Sungwoo Lee, and Bongwon Suh. I lead, you help but only with enough details: Understanding user experience of co-creation with artificial intelligence. In Proceedings of the 2018 CHI Conference on Human Factors in Computing Systems, page 1–13, New York, NY, USA, 2018. Association for Computing Machinery. 3 +[74] OpenAI. Gpt-4 technical report, 2024. 2, 3, 4, 6 +[75] Dustin Podell, Zion English, Kyle Lacey, A. Blattmann, Tim Dockhorn, Jonas Muller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. ArXiv, abs/2307.01952, 2023. 2, 3 +[76] Yonggang Qi, Guoyao Su, Pinaki Nath Chowdhury, Mingkang Li, and Yi-Zhe Song. Sketchlattice: Latticed representation for sketch manipulation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 953–961, 2021. 2 +[77] Shuwen Qiu, Sirui Xie, Lifeng Fan, Tao Gao, Song-Chun Zhu, and Yixin Zhu. Emergent graphical conventions in a visual communication game. arXiv preprint arXiv:2111.14210, 2021. 3 +[78] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. CoRR, abs/2103.00020, 2021. 1, 3, 6 +[79] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. J. Mach. Learn. Res., 21(1), 2020. 3 +[80] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 2, 3 +[81] Leo Sampaio Ferraz Ribeiro, Tu Bui, John P. Collomosse, and Moacir Antonelli Ponti. Sketchformer: Transformer-based representation for sketched structure. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14141-14150, 2020. 3 +[82] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models, 2022. 1, 2, 3, 6 + +[83] Daniel Rose, Vaishnavi Himakunthala, Andy Ouyang, Ryan He, Alex Mei, Yujie Lu, Michael Saxon, Chinmay Sonar, Diba Mirza, and William Yang Wang. Visual chain of thought: bridging logical gaps with multimodal infillings. arXiv preprint arXiv:2305.02317, 2023. 3 +[84] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S. Sara Mahdavi, Rapha Gontijo Lopes, Tim Salimans, Jonathan Ho, David J Fleet, and Mohammad Norouzi. Photorealistic text-to-image diffusion models with deep language understanding, 2022. 3 +[85] Patsorn Sangkloy, Nathan Burnell, Cusuh Ham, and James Hays. The sketchy database: Learning to retrieve badly drawn bunnies. ACM Trans. Graph., 35(4), 2016. 2 +[86] Peter Schaldenbrand, James McCann, and Jean Oh. Frida: A collaborative robot painter with a differentiable, real2sim2real planning environment. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 11712-11718, 2023. 3 +[87] Peter Schaldenbrand, Gaurav Parmar, Jun-Yan Zhu, James McCann, and Jean Oh. Cofrida: Self-supervised fine-tuning for human-robot co-painting. In 2024 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2024. 3 +[88] Donald A Schon and Vincent DeSanctis. The reflective practitioner: How professionals think in action, 1986. 2 +[89] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, Patrick Schramowski, Srivatsa Kundurthy, Katherine Crowson, Ludwig Schmidt, Robert Kaczmarczyk, and Jenia Jitsev. Laion-5b: An open large-scale dataset for training next generation image-text models, 2022. 3 +[90] Tamar Rott Shaham, Sarah Schwettmann, Franklin Wang, Achyuta Rajaram, Evan Hernandez, Jacob Andreas, and Antonio Torralba. A multimodal automated interpretability agent. In *Forty-first International Conference on Machine Learning*, 2024. 3 +[91] Hao Shao, Shengju Qian, Han Xiao, Guanglu Song, Zhuofan Zong, Letian Wang, Yu Liu, and Hongsheng Li. Visual cot: Advancing multi-modal language models with a comprehensive dataset and benchmark for chain-of-thought reasoning. arXiv preprint arXiv:2403.16999, 2024. 3 +[92] Pratyusha Sharma, Tamar Rott Shaham, Manel Baradad, Stephanie Fu, Adrian Rodriguez-Munoz, Shivam Duggal, Phillip Isola, and Antonio Torralba. A vision check-up for language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14410-14419, 2024. 3, 7 +[93] Jifei Song, Kaiyue Pang, Yi-Zhe Song, Tao Xiang, and Timothy Hospedales. Learning to sketch with shortcut cycle consistency, 2018. 2 +[94] Guoyao Su, Yonggang Qi, Kaiyue Pang, Jie Yang, Yi-Zhe Song, and CVSSP SketchX. Sketchhealer: A graph-to-sequence network for recreating partial human sketches. In BMVC, page 5, 2020. 2 + +[95] Yanpeng Sun, Qiang Chen, Jian Wang, Jingdong Wang, and Zechao Li. Exploring effective factors for improving visual in-context learning. arXiv preprint arXiv:2304.04748, 2023.3 +[96] Ivan E. Sutherland. Sketchpad—a man-machine graphical communication system, page 391-408. Association for Computing Machinery, New York, NY, USA, 1998. 3 +[97] Gemini Team. Gemini: A family of highly capable multimodal models, 2024. 2, 3 +[98] Jakob Tholander and Martin Jonsson. Design ideation with ai - sketching, thinking and talking with generative machine learning models. In Proceedings of the 2023 ACM Design-Interactive Systems Conference, page 1930-1940, New York, NY, USA, 2023. Association for Computing Machinery. 1 +[99] Shengbang Tong, Zhuang Liu, Yuexiang Zhai, Yi Ma, Yann LeCun, and Saining Xie. Eyes wide shut? exploring the visual shortcomings of multimodal llms, 2024. 4 +[100] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee Lacroix, Baptiste Roziere, Naman Goyal, Eric Hambro, Faisal Azhar, Aurelien Rodriguez, Armand Joulin, Edouard Grave, and Guillaume Lample. Llama: Open and efficient foundation language models. ArXiv, abs/2302.13971, 2023. 3 +[101] Barbara Tversky. What do sketches say about thinking? AAAI Spring Symp. Sketch Understanding Worksh., 2002. 2 +[102] Barbara Tversky. Visualizing thought. In Handbook of human centric visualization, pages 3-40. Springer, 2013. 1 +[103] Barbara Tversky, Masaki Suwa, Maneesh Agrawala, Julie Heiser, Chris Stolte, Pat Hanrahan, Doantam Phan, Jeff Klingner, Marie-Paule Daniel, Paul Lee, et al. Sketches for design and design of sketches. Human Behaviour in Design: Individuals, Teams, Tools, pages 79-86, 2003. 1 +[104] Russell Tytler, Vaughan Prain, George Aranda, Joseph Ferguson, and Radhika Gorur. Drawing to reason and learn in science. Journal of Research in Science Teaching, 57(2): 209-231, 2020. 3 +[105] Yael Vinker, Ehsan Pajouheshgar, Jessica Y. Bo, Roman Christian Bachmann, Amit Haim Bermano, Daniel Cohen-Or, Amir Zamir, and Ariel Shamir. Clipasso: Semantically-aware object sketching. ACM Trans. Graph., 41(4), 2022. 1, 3, 6 +[106] Yael Vinker, Yuval Alaluf, Daniel Cohen-Or, and Ariel Shamir. Clipascene: Scene sketching with different types and levels of abstraction. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 4146–4156, 2023. 3, 6 +[107] Jiawei Wang and Changjian Li. Contextseg: Sketch semantic segmentation by querying the context with attention. arXiv preprint arXiv:2311.16682, 2023. 6 +[108] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Brian Ichter, Fei Xia, Ed H. Chi, Quoc V. Le, and Denny Zhou. Chain-of-thought prompting elicits reasoning in large language models. In Proceedings of the 36th International Conference on Neural Information Processing Systems, Red Hook, NY, USA, 2022. Curran Associates Inc. 2, 5 + +[109] Holger Winnemöller, Jan Eric Kyprianidis, and Sven C. Olsen. Xdog: An extended difference-of-gaussians compendium including advanced image stylization. Comput. Graph., 36:740-753, 2012. 2 +[110] World Wide Web Consortium (W3C). Scalable Vector Graphics (SVG), 1999. 3 +[111] Chenfei Wu, Shengming Yin, Weizhen Qi, Xiaodong Wang, Zecheng Tang, and Nan Duan. Visual chatgpt: Talking, drawing and editing with visual foundation models. arXiv preprint arXiv:2303.04671, 2023. 3 +[112] Rong Wu, Wanchao Su, Kede Ma, and Jing Liao. Iconshop: Text-guided vector icon synthesis with autoregressive transformers. ACM Transactions on Graphics (TOG), 42:1-14, 2023. 3 +[113] Chufeng Xiao, Wanchao Su, Jing Liao, Zhouhui Lian, Yi-Zhe Song, and Hongbo Fu. Diffsketching: How differently do people sketch 3d objects? ACM Transactions on Graphics (Proceedings of ACM SIGGRAPH Asia 2022), 41 (4):1-16, 2022. 2 +[114] Saining Xie and Zhuowen Tu. Holistically-nested edge detection. In Proceedings of the IEEE international conference on computer vision, pages 1395-1403, 2015. 2 +[115] XiMing Xing, Chuang Wang, Haitao Zhou, Jing Zhang, Qian Yu, and Dong Xu. Diffsketcher: Text guided vector sketch synthesis through latent diffusion models. In Advances in Neural Information Processing Systems, pages 15869-15889. Curran Associates, Inc., 2023. 3, 6 +[116] Ximing Xing, Haitao Zhou, Chuang Wang, Jing Zhang, Dong Xu, and Qian Yu. Svgdreamer: Text guided svg generation with diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4546-4555, 2024. 3, 6, 7 +[117] Peng Xu, Timothy M. Hospedales, Qiyue Yin, Yi-Zhe Song, Tao Xiang, and Liang Wang. Deep learning for free-hand sketch: A survey and a toolbox, 2020. 2 +[118] Zhengyuan Yang, Linjie Li, Kevin Lin, Jianfeng Wang, Chung-Ching Lin, Zicheng Liu, and Lijuan Wang. The dawn of Imms: Preliminary explorations with gpt-4v(ision). ArXiv, abs/2309.17421, 2023. 2 +[119] Zhengyuan Yang, Jianfeng Wang, Linjie Li, Kevin Lin, Chung-Ching Lin, Zicheng Liu, and Lijuan Wang. Idea2img: Iterative self-refinement with gpt-4v (ision) for automatic image design and generation. arXiv preprint arXiv:2310.08541, 2023. 3 +[120] Ran Yi, Yong-Jin Liu, Yu-Kun Lai, and Paul L Rosin. Apdrawinggan: Generating artistic portrait drawings from face photos with hierarchical gans. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10743-10752, 2019. 2 +[121] Ran Yi, Yong-Jin Liu, Yu-Kun Lai, and Paul L Rosin. Unpaired portrait drawing generation via asymmetric cycle mapping. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8217-8225, 2020. 2 +[122] C. Zhang, Weijie Wang, Paul Pangaro, Nikolas Martelaro, and Daragh Byrne. Generative image ai using design sketches as input: Opportunities and challenges. Proceed- + +ings of the 15th Conference on Creativity and Cognition, 2023. 1 +[123] Jiahao Zhang, Bowen Wang, Liangzhi Li, Yuta Nakashima, and Hajime Nagahara. Instruct me more! random prompting for visual in-context learning. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2597-2606, 2024. 3 +[124] Peiying Zhang, Nanxuan Zhao, and Jing Liao. Text-to-vector generation with neural path representation. ACM Trans. Graph., 43(4), 2024. 3 +[125] Yuanhan Zhang, Kaiyang Zhou, and Ziwei Liu. What makes good examples for visual in-context learning? Advances in Neural Information Processing Systems, 36: 17773-17794, 2023. 3 +[126] Zhengming Zhang, Xiaoming Deng, Jinyao Li, Yukun Lai, Cuixia Ma, Yongjin Liu, and Hongan Wang. Stroke-based semantic segmentation for scene-level free-hand sketches. Vis. Comput., 39(12):6309-6321, 2022. 6 +[127] Zhuosheng Zhang, Aston Zhang, Mu Li, Hai Zhao, George Karypis, and Alex Smola. Multimodal chain-of-thought reasoning in language models. arXiv preprint arXiv:2302.00923, 2023. 3 +[128] Yixiao Zheng, Kaiyue Pang, Ayan Das, Dongliang Chang, Yi-Zhe Song, and Zhanyu Ma. Creativeseg: Semantic segmentation of creative sketches. IEEE Transactions on Image Processing, 33:2266-2278, 2024. 6 +[129] Tao Zhou, Chen Fang, Zhaowen Wang, Jimei Yang, Byungmoon Kim, Zhili Chen, Jonathan Brandt, and Demetri Terzopoulos. Learning to sketch with deep q networks and demonstrated strokes. ArXiv, abs/1810.05977, 2018. 2, 3 \ No newline at end of file diff --git a/2025/SketchAgent_ Language-Driven Sequential Sketch Generation/images.zip b/2025/SketchAgent_ Language-Driven Sequential Sketch Generation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..89bbe946db005487bdaa7c5d369ff9c0e8e946aa --- /dev/null +++ b/2025/SketchAgent_ Language-Driven Sequential Sketch Generation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4337aaabd14f10b06d732b235660021d509dc783b6e0e63e2916b100a1ff21db +size 460971 diff --git a/2025/SketchAgent_ Language-Driven Sequential Sketch Generation/layout.json b/2025/SketchAgent_ Language-Driven Sequential Sketch Generation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..8b8a4376e1be240f25fbeb635daf8d362494d20d --- /dev/null +++ b/2025/SketchAgent_ Language-Driven Sequential Sketch Generation/layout.json @@ -0,0 +1,11891 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 104, + 75, + 489, + 92 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 75, + 489, + 92 + ], + "spans": [ + { + "bbox": [ + 104, + 75, + 489, + 92 + ], + "type": "text", + "content": "SketchAgent: Language-Driven Sequential Sketch Generation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 52, + 105, + 541, + 121 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 105, + 541, + 121 + ], + "spans": [ + { + "bbox": [ + 52, + 105, + 541, + 121 + ], + "type": "text", + "content": "Yael Vinker" + }, + { + "bbox": [ + 52, + 105, + 541, + 121 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 52, + 105, + 541, + 121 + ], + "type": "text", + "content": " Tamar Rott Shaham" + }, + { + "bbox": [ + 52, + 105, + 541, + 121 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 52, + 105, + 541, + 121 + ], + "type": "text", + "content": " Kristine Zheng" + }, + { + "bbox": [ + 52, + 105, + 541, + 121 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 52, + 105, + 541, + 121 + ], + "type": "text", + "content": " Alex Zhao" + }, + { + "bbox": [ + 52, + 105, + 541, + 121 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 52, + 105, + 541, + 121 + ], + "type": "text", + "content": " Judith E Fan" + }, + { + "bbox": [ + 52, + 105, + 541, + 121 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 52, + 105, + 541, + 121 + ], + "type": "text", + "content": " Antonio Torralba" + }, + { + "bbox": [ + 52, + 105, + 541, + 121 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 193, + 133, + 222, + 147 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 193, + 133, + 222, + 147 + ], + "spans": [ + { + "bbox": [ + 193, + 133, + 222, + 147 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 193, + 133, + 222, + 147 + ], + "type": "text", + "content": "MIT" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 86, + 150, + 329, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 150, + 329, + 163 + ], + "spans": [ + { + "bbox": [ + 86, + 150, + 329, + 163 + ], + "type": "text", + "content": "{yaelvink,tamarott,alexzhao,torralba}@mit.edu" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 380, + 133, + 482, + 148 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 380, + 133, + 482, + 148 + ], + "spans": [ + { + "bbox": [ + 380, + 133, + 482, + 148 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 380, + 133, + 482, + 148 + ], + "type": "text", + "content": "Stanford University" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 357, + 150, + 507, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 357, + 150, + 507, + 163 + ], + "spans": [ + { + "bbox": [ + 357, + 150, + 507, + 163 + ], + "type": "text", + "content": "{jefan, kxzheng} " + }, + { + "bbox": [ + 357, + 150, + 507, + 163 + ], + "type": "inline_equation", + "content": "@" + }, + { + "bbox": [ + 357, + 150, + 507, + 163 + ], + "type": "text", + "content": " stanford.edu" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 200, + 171, + 391, + 182 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 200, + 171, + 391, + 182 + ], + "spans": [ + { + "bbox": [ + 200, + 171, + 391, + 182 + ], + "type": "text", + "content": "https://sketch-agent.csail.mit.edu/" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 67, + 197, + 201, + 366 + ], + "blocks": [ + { + "bbox": [ + 67, + 197, + 201, + 366 + ], + "lines": [ + { + "bbox": [ + 67, + 197, + 201, + 366 + ], + "spans": [ + { + "bbox": [ + 67, + 197, + 201, + 366 + ], + "type": "image", + "image_path": "ab70df958686f2d2f5dd138cd4518fae137431adbcbf0135e8142d26148c2090.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 373, + 546, + 396 + ], + "lines": [ + { + "bbox": [ + 46, + 373, + 546, + 396 + ], + "spans": [ + { + "bbox": [ + 46, + 373, + 546, + 396 + ], + "type": "text", + "content": "Figure 1. SketchAgent leverages an off-the-shelf multimodal LLM to facilitate language-driven, sequential sketch generation through an intuitive sketching language. It can sketch diverse concepts, engage in interactive sketching with humans, and edit content via chat." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 227, + 230, + 379, + 361 + ], + "blocks": [ + { + "bbox": [ + 241, + 198, + 354, + 224 + ], + "lines": [ + { + "bbox": [ + 241, + 198, + 354, + 224 + ], + "spans": [ + { + "bbox": [ + 241, + 198, + 354, + 224 + ], + "type": "text", + "content": "Human-Agent Collaborative Sketching" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 227, + 230, + 379, + 361 + ], + "lines": [ + { + "bbox": [ + 227, + 230, + 379, + 361 + ], + "spans": [ + { + "bbox": [ + 227, + 230, + 379, + 361 + ], + "type": "image", + "image_path": "660caff867d36bd3a2a1921ee55b7b3b76542d9b1364d0157c87333634aa747d.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 396, + 215, + 441, + 258 + ], + "blocks": [ + { + "bbox": [ + 392, + 197, + 518, + 212 + ], + "lines": [ + { + "bbox": [ + 392, + 197, + 518, + 212 + ], + "spans": [ + { + "bbox": [ + 392, + 197, + 518, + 212 + ], + "type": "text", + "content": "Chat-Based Sketch Editing" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 396, + 215, + 441, + 258 + ], + "lines": [ + { + "bbox": [ + 396, + 215, + 441, + 258 + ], + "spans": [ + { + "bbox": [ + 396, + 215, + 441, + 258 + ], + "type": "image", + "image_path": "bb89323786ba343d8e554d354ea822316c073446fd92f4ecc14a65f0ac735d29.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 465, + 217, + 511, + 226 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 465, + 217, + 511, + 226 + ], + "spans": [ + { + "bbox": [ + 465, + 217, + 511, + 226 + ], + "type": "text", + "content": "Draw a robot" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 451, + 255, + 511, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 451, + 255, + 511, + 274 + ], + "spans": [ + { + "bbox": [ + 451, + 255, + 511, + 274 + ], + "type": "text", + "content": "Add a paintbrush and a canvas" + } + ] + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 397, + 266, + 441, + 308 + ], + "blocks": [ + { + "bbox": [ + 397, + 266, + 441, + 308 + ], + "lines": [ + { + "bbox": [ + 397, + 266, + 441, + 308 + ], + "spans": [ + { + "bbox": [ + 397, + 266, + 441, + 308 + ], + "type": "image", + "image_path": "58090f024aca3cd9a38c512e69f40fb0834a6bf929eb5b573f55723bd62c16a8.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "bbox": [ + 454, + 304, + 511, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 454, + 304, + 511, + 324 + ], + "spans": [ + { + "bbox": [ + 454, + 304, + 511, + 324 + ], + "type": "text", + "content": "Draw a robot on the canvas" + } + ] + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 397, + 315, + 441, + 358 + ], + "blocks": [ + { + "bbox": [ + 397, + 315, + 441, + 358 + ], + "lines": [ + { + "bbox": [ + 397, + 315, + 441, + 358 + ], + "spans": [ + { + "bbox": [ + 397, + 315, + 441, + 358 + ], + "type": "image", + "image_path": "754aa295f56e6b4570d9a42392daf32d422e9c28aa0ecbf726d93912ff19e24f.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "bbox": [ + 143, + 406, + 192, + 418 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 406, + 192, + 418 + ], + "spans": [ + { + "bbox": [ + 143, + 406, + 192, + 418 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 45, + 437, + 289, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 437, + 289, + 700 + ], + "spans": [ + { + "bbox": [ + 45, + 437, + 289, + 700 + ], + "type": "text", + "content": "Sketching serves as a versatile tool for externalizing ideas, enabling rapid exploration and visual communication that spans various disciplines. While artificial systems have driven substantial advances in content creation and human-computer interaction, capturing the dynamic and abstract nature of human sketching remains challenging. In this work, we introduce SketchAgent, a language-driven, sequential sketch generation method that enables users to create, modify, and refine sketches through dynamic, conversational interactions. Our approach requires no training or fine-tuning. Instead, we leverage the sequential nature and rich prior knowledge of off-the-shelf multimodal large language models (LLMs). We present an intuitive sketching language, introduced to the model through in-context examples, enabling it to \"draw\" using string-based actions. These are processed into vector graphics and then rendered to create a sketch on a pixel canvas, which can be accessed again for further tasks. By drawing stroke by stroke, our agent captures the evolving, dynamic qualities intrinsic to sketching. We demonstrate that SketchAgent can generate sketches from diverse prompts, engage in dialogue-driven drawing, and collaborate meaningfully with human users." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 406, + 386, + 418 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 406, + 386, + 418 + ], + "spans": [ + { + "bbox": [ + 307, + 406, + 386, + 418 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 304, + 426, + 545, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 426, + 545, + 569 + ], + "spans": [ + { + "bbox": [ + 304, + 426, + 545, + 569 + ], + "type": "text", + "content": "Sketching is a powerful tool for distilling ideas into their simplest form. Its fluid and spontaneous nature makes sketching a uniquely versatile tool for visualization, rapid ideation, and communication across cultures, generations, and disciplines [26, 102]. For example, designers use sketches to explore new ideas [39, 103], scientists employ them to formulate problems [48, 72], and children engage in sketching to learn and express themselves [27, 28] (see Fig. 2). Artificial systems, in principle, have the potential to support and enhance human creativity, problem-solving, and visual expression through sketching, adapting flexibly to their exploratory nature [22, 98, 122]." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 546, + 713 + ], + "type": "text", + "content": "Traditionally, sketch generation methods rely on human-drawn datasets to train generative models [5, 6, 16, 36, 42, 59]. However, fully capturing the diversity of sketches within datasets remains challenging [26], limiting these methods in both scale and diversity. Recent advancements in vision-language models, such as CLIP [78] and text-to-image diffusion [82], have enabled sketch generation methods that reduce reliance on human-drawn datasets [29, 46, 105]. These methods leverage pretrained model guidance and differentiable rendering [58] to optimize parametric curves, creating sketches that go beyond predefined styles and categories." + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23355" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 204 + ], + "type": "text", + "content": "While representing a significant step toward a general-purpose sketching system, these methods lack a crucial aspect of human drawing: the process itself. Current methods, though versatile, optimize all strokes simultaneously, making the intermediate sketching steps meaningless. As a result, the sketch cannot be decomposed into a coherent sequence of strokes that reflects the drawing process. In contrast, humans draw iteratively, stroke by stroke, incorporating visual feedback and continuously adapting—a dynamic, evolving process that fosters creativity, ideation, and communication [52, 88, 101]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 206, + 289, + 315 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 206, + 289, + 315 + ], + "spans": [ + { + "bbox": [ + 46, + 206, + 289, + 315 + ], + "type": "text", + "content": "In this work, we introduce SketchAgent, a sketch generation agent that leverages the prior knowledge and sequential nature of multimodal large language models (LLMs) to enable versatile, progressive, language-driven sketching. Our agent can generate sketches across a wide range of textual concepts—from animals to engineering principles (Fig. 1, left). Its sequential nature facilitates interactive human-agent sketching and supports iterative refinement and editing through a chat-based dialogue (Fig. 1, right)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 316, + 289, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 316, + 289, + 496 + ], + "spans": [ + { + "bbox": [ + 46, + 316, + 289, + 496 + ], + "type": "text", + "content": "Unlike vision-language models that directly generate images from text [75, 80, 82], multimodal LLMs [1, 2, 15, 56, 64, 74, 97] accept text and images as input but only output text. To produce visuals, they either utilize external \"tools\" (such as calling a text-to-image model) or are prompted to generate executable code (e.g., Python [43], SVG [9]) to create charts, diagrams, or graphics. However, prompting for such representations to directly produce sketches often results in a mechanical appearance with uniform, precise shapes that lack the subtle irregularities and spontaneous qualities characteristic of human sketches (see Fig. 3B). Additionally, despite their robustness in textual tasks, these models often struggle with fine-grained spatial reasoning [41, 118] as they are primarily optimized for text, making sketch editing more challenging." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 498, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 289, + 714 + ], + "type": "text", + "content": "To address these limitations, we introduce an intuitive sketching language that enables an off-the-shelf multimodal LLM agent to \"draw\" sketches on a canvas by providing string-based actions, without additional training or finetuning. We define the canvas as a numbered grid, allowing the agent to reference specific coordinates (e.g., x2y8) to enhance its spatial reasoning capabilities. We represent a sketch as a sequence of semantically meaningful strokes, each defined by a series of such coordinates. We leverage In-Context Learning (ICL) [7, 51] to introduce the agent to the new representation, and Chain of Thought (CoT) [108] to enhance its planning capabilities. Given a sketching task, the agent produces a textual response following our representation, which we process by fitting a smooth Bézier curve to each coordinate sequence. The curves are then rendered onto the canvas to form the final sketch. We find this approach useful in emulating a more natural sketch appearance. For collaborative sketching, the canvas remains acces" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 325, + 92, + 444, + 173 + ], + "blocks": [ + { + "bbox": [ + 325, + 73, + 424, + 91 + ], + "lines": [ + { + "bbox": [ + 325, + 73, + 424, + 91 + ], + "spans": [ + { + "bbox": [ + 325, + 73, + 424, + 91 + ], + "type": "text", + "content": "(A) Frank Gehry. *Elevation Sketches*. Guggenheim Museum" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 325, + 92, + 444, + 173 + ], + "lines": [ + { + "bbox": [ + 325, + 92, + 444, + 173 + ], + "spans": [ + { + "bbox": [ + 325, + 92, + 444, + 173 + ], + "type": "image", + "image_path": "79df84ee3c76e6a87bd277ede69140cb5c63521e15ceac8868d03a0ab43c608d.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 325, + 194, + 444, + 256 + ], + "blocks": [ + { + "bbox": [ + 325, + 176, + 416, + 194 + ], + "lines": [ + { + "bbox": [ + 325, + 176, + 416, + 194 + ], + "spans": [ + { + "bbox": [ + 325, + 176, + 416, + 194 + ], + "type": "text", + "content": "(B) Alexander Bell's drawing of the telephone" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 325, + 194, + 444, + 256 + ], + "lines": [ + { + "bbox": [ + 325, + 194, + 444, + 256 + ], + "spans": [ + { + "bbox": [ + 325, + 194, + 444, + 256 + ], + "type": "image", + "image_path": "84445e52d6588ee0eb6c7ec6a9e37694b8fc18440ca5cc76762989cfe593d383.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 447, + 85, + 526, + 141 + ], + "blocks": [ + { + "bbox": [ + 447, + 73, + 512, + 83 + ], + "lines": [ + { + "bbox": [ + 447, + 73, + 512, + 83 + ], + "spans": [ + { + "bbox": [ + 447, + 73, + 512, + 83 + ], + "type": "text", + "content": "(C) Children drawing" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 447, + 85, + 526, + 141 + ], + "lines": [ + { + "bbox": [ + 447, + 85, + 526, + 141 + ], + "spans": [ + { + "bbox": [ + 447, + 85, + 526, + 141 + ], + "type": "image", + "image_path": "7f2b351f7e27a24b7b48e583019243e8bb3c15676cd7c4566913578b20efa11e.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 447, + 166, + 527, + 258 + ], + "blocks": [ + { + "bbox": [ + 447, + 145, + 518, + 164 + ], + "lines": [ + { + "bbox": [ + 447, + 145, + 518, + 164 + ], + "spans": [ + { + "bbox": [ + 447, + 145, + 518, + 164 + ], + "type": "text", + "content": "(D) Basketball strategy (pick and roll)" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 447, + 166, + 527, + 258 + ], + "lines": [ + { + "bbox": [ + 447, + 166, + 527, + 258 + ], + "spans": [ + { + "bbox": [ + 447, + 166, + 527, + 258 + ], + "type": "image", + "image_path": "094c36db6efb33f65a4fa9b7c75cc6344415ae800a93f97dbef517957c6988be.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 261, + 547, + 327 + ], + "lines": [ + { + "bbox": [ + 304, + 261, + 547, + 327 + ], + "spans": [ + { + "bbox": [ + 304, + 261, + 547, + 327 + ], + "type": "text", + "content": "Figure 2. Examples of sketches used across disciplines and goals. (A) Ideation and design: Process Elevation Sketches by the architect Frank Gehry, Guggenheim Museum. (B) Engineering: Alexander Bell's telephone drawing. (C) Expressing emotions: Children's sketches. (D) Visual communication: Planning and communicating game strategy in basketball." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 336, + 545, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 336, + 545, + 407 + ], + "spans": [ + { + "bbox": [ + 304, + 336, + 545, + 407 + ], + "type": "text", + "content": "sible to both the user and the agent throughout the session. The agent generates strokes sequentially and pauses according to an adjustable stopping token, allowing the user to add their own strokes directly to the canvas. These strokes are then integrated into the agent's sequence, enabling it to continue drawing, with real-time canvas updates." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 408, + 547, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 408, + 547, + 515 + ], + "spans": [ + { + "bbox": [ + 304, + 408, + 547, + 515 + ], + "type": "text", + "content": "We demonstrate SketchAgent's capability to generate sketches of diverse concepts while capturing the inherently sequential and dynamic nature of sketching. We showcase our agent's ability to collaborate effectively with humans in real time to create novel and meaningful sketches. Our method is the first to leverage pretrained multimodal LLMs for sequential sketching without additional training, paving the way for a general-purpose artificial sketching system that supports iterative, evolving interactivity." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 526, + 392, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 526, + 392, + 538 + ], + "spans": [ + { + "bbox": [ + 306, + 526, + 392, + 538 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 545, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 545, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 545, + 547, + 713 + ], + "type": "text", + "content": "Sketch Generation Early methods approached sketch generation by designing image filters to simulate sketchlike effects [10, 109]. With the advent of deep learning, data-driven approaches emerged to address a range of sketch-related tasks [117], including category-conditioned sketching [42, 76, 93], object sketching [59, 62], scenesketching [12, 57, 60, 114], sketch completion [6, 63, 94], portrait drawing [3, 120, 121], part-based generation [6, 37, 42, 129], and more. While sketch data collection has been broadly explored [21, 34, 40, 71, 85, 113], the wide variation in sketch styles and their adaptation to specific tasks [23] makes collecting datasets that encompass this diversity challenging. For example, QuickDraw [47], the largest available sketch dataset with 50 mil" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23356" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 301 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 301 + ], + "type": "text", + "content": "lion sketches, covers only 345 object categories and primarily focuses on simple, iconic representations. This limits data-driven methods to the style, abstraction level, and concepts seen during training. Recently, large pretrained vision-language models [75, 78, 80, 82, 84] have shown remarkable text-to-image generation capabilities by leveraging extensive visual knowledge from billions of training images [89]. While these models can be prompted to generate sketch-like images (see Fig. 3A), they do so in a single step and in pixel space, lacking the sequential, stroke-based process of human sketching. Subsequent approaches [14, 29, 31, 46, 105, 106, 115, 116, 124] leverage the priors of these models to guide an iterative optimization of parametric curves, with a differentiable rasterizer [58] linking pixel and vector representations. While producing vector sketches, the final strokes lack order and semantic meaning, and the optimization-based approach overlook the sequential aspect of the sketching process, making these methods suboptimal for collaborative sketching." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 308, + 289, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 308, + 289, + 561 + ], + "spans": [ + { + "bbox": [ + 46, + 308, + 289, + 561 + ], + "type": "text", + "content": "Sequential and Collaborative Sketching Collaborative human-machine sketching holds promise in enhancing creativity, ideation, communication, and learning, as explored in various fields, including human-computer interaction (HCI) [17, 45, 49, 50, 53, 54], computer graphics [55, 96], robotics [86, 87], cognitive science [24, 25, 35, 67], learning sciences [18, 38, 104], and more. Central to collaborative sketching is its sequential, adaptive, and dynamic process, with each action carrying intent. Existing methods employ diverse training strategies to account for the discrete nature of sequential sketches, including reinforcement and adversarial learning [32, 68, 129], multi-agent referential games [69, 77], transformers [5, 6, 11, 33, 61, 81, 112], and more. SketchRNN [42] is a pioneering work in this area, introducing the QuickDraw dataset [47], a crowd-sourced collection of real-time sketch sequences made by users. They utilize this dataset to train a recurrent neural network for sequential sketch generation, which was later shown [24, 73] to have potential for human-machine collaboration. However, this approach remains constrained by the predefined categories encountered during training." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 570, + 289, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 570, + 289, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 570, + 289, + 714 + ], + "type": "text", + "content": "Multimodel LLMs for Content Creation LLMs [7, 19, 79, 100] and multimodal LLMs [1, 2, 15, 56, 64, 74, 97] receive text as input (or text and images for multimodal) and output text. To enable visual content generation, these models are often paired with external \"tools\" that extend their functionality [44, 90, 111, 119]. For example, ChatGPT [74] generates images by internally calling a separate model, DALLE-3 [4]. Another approach involves prompting models to produce code in languages like Python [43], Processing [92], SVG [9], or TikZ [8] that can be rendered into visuals such as graphs, charts, and vector graphics. However, such code-generated content often looks" + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 310, + 72, + 366, + 129 + ], + "blocks": [ + { + "bbox": [ + 310, + 72, + 366, + 129 + ], + "lines": [ + { + "bbox": [ + 310, + 72, + 366, + 129 + ], + "spans": [ + { + "bbox": [ + 310, + 72, + 366, + 129 + ], + "type": "image", + "image_path": "ecebd173d5244262321496c10bc7579f13a3ede01ed0683c1bba5b4ff4bfd7d6.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 310, + 136, + 366, + 146 + ], + "lines": [ + { + "bbox": [ + 310, + 136, + 366, + 146 + ], + "spans": [ + { + "bbox": [ + 310, + 136, + 366, + 146 + ], + "type": "text", + "content": "(A) DALLE3 [4]" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 375, + 85, + 422, + 111 + ], + "blocks": [ + { + "bbox": [ + 375, + 85, + 422, + 111 + ], + "lines": [ + { + "bbox": [ + 375, + 85, + 422, + 111 + ], + "spans": [ + { + "bbox": [ + 375, + 85, + 422, + 111 + ], + "type": "image", + "image_path": "269d98ca61721ce69d816b986175812dc51676e984fa2d9bdfd596c779d1bca9.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 376, + 131, + 421, + 152 + ], + "lines": [ + { + "bbox": [ + 376, + 131, + 421, + 152 + ], + "spans": [ + { + "bbox": [ + 376, + 131, + 421, + 152 + ], + "type": "text", + "content": "(B) LLMs [2] (SVG)" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 438, + 86, + 480, + 111 + ], + "blocks": [ + { + "bbox": [ + 438, + 86, + 480, + 111 + ], + "lines": [ + { + "bbox": [ + 438, + 86, + 480, + 111 + ], + "spans": [ + { + "bbox": [ + 438, + 86, + 480, + 111 + ], + "type": "image", + "image_path": "26f29f516ec09b7cad0c0384e589dc27ba4c5f593fda2154d43f5b99be1eeb9b.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 432, + 137, + 487, + 146 + ], + "lines": [ + { + "bbox": [ + 432, + 137, + 487, + 146 + ], + "spans": [ + { + "bbox": [ + 432, + 137, + 487, + 146 + ], + "type": "text", + "content": "(C) SketchAgent" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 498, + 91, + 538, + 114 + ], + "blocks": [ + { + "bbox": [ + 498, + 91, + 538, + 114 + ], + "lines": [ + { + "bbox": [ + 498, + 91, + 538, + 114 + ], + "spans": [ + { + "bbox": [ + 498, + 91, + 538, + 114 + ], + "type": "image", + "image_path": "30f937880ee2cde64f8b31b9bf39dc243cda3140563dd4a93f0198d2aa2d28be.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 493, + 137, + 546, + 146 + ], + "lines": [ + { + "bbox": [ + 493, + 137, + 546, + 146 + ], + "spans": [ + { + "bbox": [ + 493, + 137, + 546, + 146 + ], + "type": "text", + "content": "(D) Human [47]" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 154, + 547, + 221 + ], + "lines": [ + { + "bbox": [ + 304, + 154, + 547, + 221 + ], + "spans": [ + { + "bbox": [ + 304, + 154, + 547, + 221 + ], + "type": "text", + "content": "Figure 3. Sketch appearance. (A) Text-to-image diffusion models operate in pixel space, lacking the sequential nature of sketches. (B) Prompting LLMs to produce visuals with SVG results in a uniform, mechanical appearance. (C) Sketches produced by our agent appear less mechanical, more closely resembling the nature of (D) Human sketches, which are often spontaneous and irregular." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 239, + 547, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 239, + 547, + 384 + ], + "spans": [ + { + "bbox": [ + 304, + 239, + 547, + 384 + ], + "type": "text", + "content": "rigid, with uniform and overly precise shapes that lack the subtle irregularities and spontaneous qualities characteristic of freehand sketches (see Fig. 3B). In contrast, we propose a sketching language grounded in spatial information that encourages the model to produce a more natural sketch appearance, which we then process into vector graphics. Common strategies for enhancing LLMs capabilities include Chain-of-Thought prompting [13, 70, 83, 91, 127], which breaks down tasks into smaller, logical steps to mimic human reasoning, and In-Context Learning (ICL) [7, 20, 95, 123, 125], where examples of input-output pairs are provided to help the model infer task patterns." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 400, + 391, + 414 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 400, + 391, + 414 + ], + "spans": [ + { + "bbox": [ + 306, + 400, + 391, + 414 + ], + "type": "text", + "content": "3. Preliminaries" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 422, + 452, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 422, + 452, + 506 + ], + "spans": [ + { + "bbox": [ + 306, + 422, + 452, + 506 + ], + "type": "text", + "content": "Vector Graphics and Bezier Curves Vector graphics allow us to create visual images directly from geometric shapes such as points, lines, curves, and polygons. Unlike raster images (represented with pixels), vector graphics are" + } + ] + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 474, + 398, + 542, + 472 + ], + "blocks": [ + { + "bbox": [ + 474, + 398, + 542, + 472 + ], + "lines": [ + { + "bbox": [ + 474, + 398, + 542, + 472 + ], + "spans": [ + { + "bbox": [ + 474, + 398, + 542, + 472 + ], + "type": "image", + "image_path": "cfd520c98eb43f76e42eb72ff93f21bfc86c01d6fb8de66fb635b103b300bc01.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 471, + 475, + 545, + 496 + ], + "lines": [ + { + "bbox": [ + 471, + 475, + 545, + 496 + ], + "spans": [ + { + "bbox": [ + 471, + 475, + 545, + 496 + ], + "type": "text", + "content": "Figure 4. Cubic Bezier curve." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 506, + 547, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 506, + 547, + 650 + ], + "spans": [ + { + "bbox": [ + 304, + 506, + 547, + 650 + ], + "type": "text", + "content": "resolution-free, more compact, and editable. SVG [110] is an XML-based format for storing vector graphics, popular for its scalability and compatibility with modern web browsers. The process of transferring vector graphics into pixel images is called rasterization or rendering. Cubic Bézier curves are commonly used to represent sketches in vector graphics. A cubic Bézier curve (Fig. 4) is a smooth parametric curve defined by four points: a start point " + }, + { + "bbox": [ + 304, + 506, + 547, + 650 + ], + "type": "inline_equation", + "content": "P_0" + }, + { + "bbox": [ + 304, + 506, + 547, + 650 + ], + "type": "text", + "content": ", an end point " + }, + { + "bbox": [ + 304, + 506, + 547, + 650 + ], + "type": "inline_equation", + "content": "P_3" + }, + { + "bbox": [ + 304, + 506, + 547, + 650 + ], + "type": "text", + "content": ", and two control points " + }, + { + "bbox": [ + 304, + 506, + 547, + 650 + ], + "type": "inline_equation", + "content": "P_1" + }, + { + "bbox": [ + 304, + 506, + 547, + 650 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 506, + 547, + 650 + ], + "type": "inline_equation", + "content": "P_2" + }, + { + "bbox": [ + 304, + 506, + 547, + 650 + ], + "type": "text", + "content": " that shape the curvature. The set " + }, + { + "bbox": [ + 304, + 506, + 547, + 650 + ], + "type": "inline_equation", + "content": "P = \\{P_0, P_1, P_2, P_3\\}" + }, + { + "bbox": [ + 304, + 506, + 547, + 650 + ], + "type": "text", + "content": " is often referred to as the curve's control points. The curve is described by the following polynomial equation:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 311, + 662, + 545, + 677 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 662, + 545, + 677 + ], + "spans": [ + { + "bbox": [ + 311, + 662, + 545, + 677 + ], + "type": "interline_equation", + "content": "B (t) = (1 - t) ^ {3} P _ {0} + 3 (1 - t) ^ {2} t P _ {1} + 3 (1 - t) t ^ {2} P _ {2} + t ^ {3} P _ {3}, (1)", + "image_path": "42134cdea632de8345e6bfaa2ef4b7fb31994d147da29db20c60036c2a4258f4.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "inline_equation", + "content": "t \\in [0,1]" + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": " is a parameter that moves the point along the curve from " + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "inline_equation", + "content": "P_0" + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": " at " + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "inline_equation", + "content": "t = 0" + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "inline_equation", + "content": "P_3" + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": " at " + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "inline_equation", + "content": "t = 1" + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23357" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 70, + 544, + 194 + ], + "blocks": [ + { + "bbox": [ + 50, + 70, + 544, + 194 + ], + "lines": [ + { + "bbox": [ + 50, + 70, + 544, + 194 + ], + "spans": [ + { + "bbox": [ + 50, + 70, + 544, + 194 + ], + "type": "image", + "image_path": "0f06d8a6a6aa9521e66a99348e8b287aae238090cc82646635cf1398b7363315.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 195, + 548, + 240 + ], + "lines": [ + { + "bbox": [ + 46, + 195, + 548, + 240 + ], + "spans": [ + { + "bbox": [ + 46, + 195, + 548, + 240 + ], + "type": "text", + "content": "Figure 5. Method Overview. SketchAgent (blue) receives drawing instructions and generates a string representing the intended sketch. Inputs include: (1) a system prompt (orange) introducing the sketching language and canvas, (2) a user prompt (pink) specifying the task (e.g., \"draw a shark\"), and (3) a numbered canvas. The agent's response outlines a sketching strategy (in thinking tags) and a sequence of strokes defined by coordinates, which are processed into Bézier curves and rendered onto the canvas." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 254, + 103, + 266 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 254, + 103, + 266 + ], + "spans": [ + { + "bbox": [ + 47, + 254, + 103, + 266 + ], + "type": "text", + "content": "4. Method" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 275, + 290, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 275, + 290, + 479 + ], + "spans": [ + { + "bbox": [ + 46, + 275, + 290, + 479 + ], + "type": "text", + "content": "Our goal is to enable an off-the-shelf pretrained multimodal LLM to draw sketches based on natural language instructions. An overview of our pipeline is illustrated in Fig. 5. We utilize a frozen multimodal LLM (\"SketchAgent\" shown in blue), which receives three inputs: (1) a system prompt containing guidelines for using our new sketching language, (2) a user prompt with additional task-specific instructions (e.g., \"Draw a shark\"), and (3) a blank canvas on which the agent can draw. Based on the given task, the agent generates a textual response, representing the sequence of strokes to be drawn, which we then process into vector graphics and render onto the canvas. The canvas can then be reused in two ways: it can be fed back into the model with an updated user prompt for additional tasks and editing, or it can be accessed by a human user who can draw directly on it to facilitate collaborative sketching. Next, we describe each component of the pipeline." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 486, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 486, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 486, + 289, + 715 + ], + "type": "text", + "content": "The Canvas Although multimodal LLMs demonstrate remarkable reasoning abilities, they often struggle with spatial reasoning tasks [30, 66, 99]. We present a simple example (see Fig. 6) to illustrate how this limitation affects the naive use of these models for sketch generation and interactive sketching. We provide GPT-4o [74] with an image depicting a simple line drawing of a partial house featuring five numbered points (from 1 to 5), and ask it to identify which points should be connected to complete the house. While the model correctly identifies the pair of points, it fails to select the correct pixel coordinates when given a basic draw-line tool that connects two points, even after multiple attempts. To enhance the model's spatial reasoning ability, we utilize a numbered canvas that forms a grid. This grid features numbers (1 to 50) along the x-axis and the y-axis (Fig. 5, left). Each cell is uniquely identified by a combination of the corresponding x-axis and y-axis numbers (e.g., the bottom-left cell is " + }, + { + "bbox": [ + 46, + 486, + 289, + 715 + ], + "type": "inline_equation", + "content": "x1y1" + }, + { + "bbox": [ + 46, + 486, + 289, + 715 + ], + "type": "text", + "content": "). The agent interacts with the canvas by specifying desired " + }, + { + "bbox": [ + 46, + 486, + 289, + 715 + ], + "type": "inline_equation", + "content": "(x,y)" + }, + { + "bbox": [ + 46, + 486, + 289, + 715 + ], + "type": "text", + "content": " coordinates." + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 310, + 253, + 541, + 312 + ], + "blocks": [ + { + "bbox": [ + 310, + 253, + 541, + 312 + ], + "lines": [ + { + "bbox": [ + 310, + 253, + 541, + 312 + ], + "spans": [ + { + "bbox": [ + 310, + 253, + 541, + 312 + ], + "type": "image", + "image_path": "eaff46c17f5e0ca25ebf8cfe4fbb32a448f235608a9bfa2adcfa2d90231fce5b.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 316, + 546, + 371 + ], + "lines": [ + { + "bbox": [ + 305, + 316, + 546, + 371 + ], + "spans": [ + { + "bbox": [ + 305, + 316, + 546, + 371 + ], + "type": "text", + "content": "Figure 6. Although excelling in visual reasoning, multimodal LLMs often struggle to translate these abilities into spatial actions. In this example, GPT-4o [74] intends to draw a line between points 1 and 5 but fails to execute this with a draw_line function that accepts pixel coordinates." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 387, + 545, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 387, + 545, + 448 + ], + "spans": [ + { + "bbox": [ + 304, + 387, + 545, + 448 + ], + "type": "text", + "content": "Sketch Representation We define a sketch as a sequence of " + }, + { + "bbox": [ + 304, + 387, + 545, + 448 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 304, + 387, + 545, + 448 + ], + "type": "text", + "content": " ordered strokes " + }, + { + "bbox": [ + 304, + 387, + 545, + 448 + ], + "type": "inline_equation", + "content": "S = \\{S_1, S_2, \\ldots, S_n\\}" + }, + { + "bbox": [ + 304, + 387, + 545, + 448 + ], + "type": "text", + "content": ". Each stroke " + }, + { + "bbox": [ + 304, + 387, + 545, + 448 + ], + "type": "inline_equation", + "content": "S_i" + }, + { + "bbox": [ + 304, + 387, + 545, + 448 + ], + "type": "text", + "content": " is defined by a sequence of " + }, + { + "bbox": [ + 304, + 387, + 545, + 448 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 304, + 387, + 545, + 448 + ], + "type": "text", + "content": " cell coordinates on the grid: " + }, + { + "bbox": [ + 304, + 387, + 545, + 448 + ], + "type": "inline_equation", + "content": "S_i = \\{(x_j, y_j)\\}_{j=1}^m" + }, + { + "bbox": [ + 304, + 387, + 545, + 448 + ], + "type": "text", + "content": ", represented in string format as: x1y1, x15y20, ...." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 449, + 546, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 449, + 546, + 592 + ], + "spans": [ + { + "bbox": [ + 304, + 449, + 546, + 592 + ], + "type": "text", + "content": "A naive approach to processing the textual sequence of coordinates would be to use a polyline, connecting consecutive points with line segments. However, our grid-based representation sparsifies the canvas, resulting in a non-smooth and unnatural appearance when using polylines (see Fig. 7, left). To achieve a smoother appearance, an alternative approach is to treat the coordinates as a sequence of control points defining smooth curves. However, as illustrated in Fig. 4, the control points often do not lie directly on the curve. Consequently, if the agent aims for a stroke that passes through specific coordinates, it must derive the control points that define this stroke, which is challenging." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "content": "We propose an alternative approach: we treat the specified " + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "inline_equation", + "content": "(x, y)" + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "content": " coordinates as a set of desired points sampled along the curve, and fit a smooth Bézier curve to them (Fig. 7, right). To accommodate curves with complex curvature, we also task the model with determining when each point on the curve should be passed through, corresponding to the " + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "content": " value described in Eq. (1). Thus, for each stroke " + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "inline_equation", + "content": "S_{i}" + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "content": ", the agent provides a set of " + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "content": " sampled points " + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "inline_equation", + "content": "S_{i} = \\{(x_{j}, y_{j})\\}_{j=1}^{m}" + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "content": ", along with a corresponding set of " + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "content": " values: " + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "inline_equation", + "content": "T_{i} = \\{t_{j}\\}_{j=1}^{m}" + }, + { + "bbox": [ + 304, + 594, + 547, + 715 + ], + "type": "text", + "content": ". Based on these, we fit a cubic Bézier" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23358" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 72, + 70, + 133, + 114 + ], + "blocks": [ + { + "bbox": [ + 72, + 70, + 133, + 114 + ], + "lines": [ + { + "bbox": [ + 72, + 70, + 133, + 114 + ], + "spans": [ + { + "bbox": [ + 72, + 70, + 133, + 114 + ], + "type": "image", + "image_path": "212fa6867ce4bdcbddaa82751742056c0944857271c1b4b8a0fc7031b67326ee.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 83, + 118, + 127, + 128 + ], + "lines": [ + { + "bbox": [ + 83, + 118, + 127, + 128 + ], + "spans": [ + { + "bbox": [ + 83, + 118, + 127, + 128 + ], + "type": "text", + "content": "(A) Polyline" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 135, + 70, + 197, + 114 + ], + "blocks": [ + { + "bbox": [ + 135, + 70, + 197, + 114 + ], + "lines": [ + { + "bbox": [ + 135, + 70, + 197, + 114 + ], + "spans": [ + { + "bbox": [ + 135, + 70, + 197, + 114 + ], + "type": "image", + "image_path": "d04feb1e43212ca5927cdd3bfb132b6d589709712778a0904e76bc0a14265623.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 131, + 287, + 186 + ], + "lines": [ + { + "bbox": [ + 46, + 131, + 287, + 186 + ], + "spans": [ + { + "bbox": [ + 46, + 131, + 287, + 186 + ], + "type": "text", + "content": "Figure 7. Methods for processing the agent's coordinate sequence (in red): (A) Polyline results in an unnatural appearance. (B) Directly using coordinates as Bézier control points is challenging as they do not lie on the curve. (C) Fitting a Bézier curve to sampled coordinates provides smoother results." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 199, + 70, + 260, + 114 + ], + "blocks": [ + { + "bbox": [ + 135, + 118, + 201, + 128 + ], + "lines": [ + { + "bbox": [ + 135, + 118, + 201, + 128 + ], + "spans": [ + { + "bbox": [ + 135, + 118, + 201, + 128 + ], + "type": "text", + "content": "(B) Control Points" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 199, + 70, + 260, + 114 + ], + "lines": [ + { + "bbox": [ + 199, + 70, + 260, + 114 + ], + "spans": [ + { + "bbox": [ + 199, + 70, + 260, + 114 + ], + "type": "image", + "image_path": "87bdbe0a578ba8382f0d291ffe00f966941310f2175de173f265c8c1c46094b2.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 213, + 118, + 246, + 128 + ], + "lines": [ + { + "bbox": [ + 213, + 118, + 246, + 128 + ], + "spans": [ + { + "bbox": [ + 213, + 118, + 246, + 128 + ], + "type": "text", + "content": "(C) Ours" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 201, + 287, + 237 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 201, + 287, + 237 + ], + "spans": [ + { + "bbox": [ + 46, + 201, + 287, + 237 + ], + "type": "text", + "content": "curve to the sampled points by solving a system of linear equations using least squares, where the unknowns are the control points " + }, + { + "bbox": [ + 46, + 201, + 287, + 237 + ], + "type": "inline_equation", + "content": "P = \\{P_0, P_1, P_2, P_3\\}" + }, + { + "bbox": [ + 46, + 201, + 287, + 237 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 113, + 242, + 287, + 255 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 242, + 287, + 255 + ], + "spans": [ + { + "bbox": [ + 113, + 242, + 287, + 255 + ], + "type": "interline_equation", + "content": "P = \\operatorname {a r g m i n} _ {P} \\| A P - B \\|, \\tag {2}", + "image_path": "bde8deae50544a36481630b9998f797efdd4058d771b5d221c44bc0c90ad1611.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 260, + 287, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 260, + 287, + 380 + ], + "spans": [ + { + "bbox": [ + 46, + 260, + 287, + 380 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 260, + 287, + 380 + ], + "type": "inline_equation", + "content": "A \\in \\mathbb{R}^{m \\times 4}" + }, + { + "bbox": [ + 46, + 260, + 287, + 380 + ], + "type": "text", + "content": " contains the cubic Bézier basis functions evaluated at specific " + }, + { + "bbox": [ + 46, + 260, + 287, + 380 + ], + "type": "inline_equation", + "content": "t_j" + }, + { + "bbox": [ + 46, + 260, + 287, + 380 + ], + "type": "text", + "content": " values (as described in Eq. (1)), and " + }, + { + "bbox": [ + 46, + 260, + 287, + 380 + ], + "type": "inline_equation", + "content": "B \\in \\mathbb{R}^{m \\times 2}" + }, + { + "bbox": [ + 46, + 260, + 287, + 380 + ], + "type": "text", + "content": " contains the " + }, + { + "bbox": [ + 46, + 260, + 287, + 380 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 46, + 260, + 287, + 380 + ], + "type": "text", + "content": " sampled points " + }, + { + "bbox": [ + 46, + 260, + 287, + 380 + ], + "type": "inline_equation", + "content": "\\{(x_j, y_j)\\}_{j=1}^m" + }, + { + "bbox": [ + 46, + 260, + 287, + 380 + ], + "type": "text", + "content": ". The least squares solution minimizes the error between the fitted Bézier curve and the sampled points. For long sequences resulting in a large fitting error, we recursively split the curve. Additionally, we account for Bézier curves of lower degrees, including quadratic curves, linear lines, and points. Upon completing this process, we render the parametric curves onto the canvas." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 389, + 287, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 389, + 287, + 651 + ], + "spans": [ + { + "bbox": [ + 46, + 389, + 287, + 651 + ], + "type": "text", + "content": "Drawing Instructions We provide the model with a system prompt and a user prompt (marked in orange and pink in Fig. 5). In the system prompt, we supply the agent with context about its expertise (\"You are an expert artist specializing in drawing sketches\") and introduce it to the grid canvas along with examples of how to use our sketching language for drawing single-stroke primitives (full prompts are provided in the Appendix). The system prompt is fixed and can be applied to a variety of sketching tasks. The user prompt includes a description of the desired task and an example of a simple sketch of a house drawn with our sketching language. We find this to be crucial in assisting the agent with preserving the correct format that could be parsed directly [7]. The agent is tasked with responding in the format shown in the gray text box in Fig. 5. In the tags, the agent is tasked to outline the overall sketching strategy [108]. This typically includes describing the different components of the sketch, the intended sketching order, and the overall placement of each part. The agent is also tasked with providing an ID tag following each stroke, which is useful for further analysis and for producing annotated sketches in scale." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 47, + 659, + 280, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 659, + 280, + 672 + ], + "spans": [ + { + "bbox": [ + 47, + 659, + 280, + 672 + ], + "type": "text", + "content": "4.1. In-Chat Editing and Collaborative Sketching" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": "The above process can be repeated iteratively to support multiple sketching tasks and interactions. Text-based sketch editing in a chat dialogue is enabled by feeding the rendered" + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 310, + 76, + 369, + 122 + ], + "blocks": [ + { + "bbox": [ + 310, + 76, + 369, + 122 + ], + "lines": [ + { + "bbox": [ + 310, + 76, + 369, + 122 + ], + "spans": [ + { + "bbox": [ + 310, + 76, + 369, + 122 + ], + "type": "image", + "image_path": "95f2cb4f29aad5591b1e0194e2cf0ab7526e4b5e016af1d9e988b66435856fc1.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 374, + 75, + 421, + 114 + ], + "blocks": [ + { + "bbox": [ + 374, + 75, + 421, + 114 + ], + "lines": [ + { + "bbox": [ + 374, + 75, + 421, + 114 + ], + "spans": [ + { + "bbox": [ + 374, + 75, + 421, + 114 + ], + "type": "image", + "image_path": "2ce6c74333131c52fd4d2db63f898a49a0e211208a5a003b97036fbaedfe1a24.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 437, + 72, + 474, + 118 + ], + "blocks": [ + { + "bbox": [ + 437, + 72, + 474, + 118 + ], + "lines": [ + { + "bbox": [ + 437, + 72, + 474, + 118 + ], + "spans": [ + { + "bbox": [ + 437, + 72, + 474, + 118 + ], + "type": "image", + "image_path": "a200c7aa52d21ba8b21482f41ac05341be18a89a9ccba901fa92fd467b789848.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 504, + 76, + 524, + 121 + ], + "blocks": [ + { + "bbox": [ + 504, + 76, + 524, + 121 + ], + "lines": [ + { + "bbox": [ + 504, + 76, + 524, + 121 + ], + "spans": [ + { + "bbox": [ + 504, + 76, + 524, + 121 + ], + "type": "image", + "image_path": "1bd05a8d61780921afe2f41ff9d85cf2dc4418c435896e452d678996aeb9e97a.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 493, + 128, + 534, + 149 + ], + "lines": [ + { + "bbox": [ + 493, + 128, + 534, + 149 + ], + "spans": [ + { + "bbox": [ + 493, + 128, + 534, + 149 + ], + "type": "text", + "content": "DNA Double Helix" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 323, + 154, + 353, + 183 + ], + "blocks": [ + { + "bbox": [ + 315, + 128, + 362, + 150 + ], + "lines": [ + { + "bbox": [ + 315, + 128, + 362, + 150 + ], + "spans": [ + { + "bbox": [ + 315, + 128, + 362, + 150 + ], + "type": "text", + "content": "Golden Gate Bridge" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 323, + 154, + 353, + 183 + ], + "lines": [ + { + "bbox": [ + 323, + 154, + 353, + 183 + ], + "spans": [ + { + "bbox": [ + 323, + 154, + 353, + 183 + ], + "type": "image", + "image_path": "deab0063155c5254908d052613efaa75c2a6d154ebe3baeeead4c1ce067194f5.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 376, + 160, + 416, + 194 + ], + "blocks": [ + { + "bbox": [ + 383, + 128, + 410, + 150 + ], + "lines": [ + { + "bbox": [ + 383, + 128, + 410, + 150 + ], + "spans": [ + { + "bbox": [ + 383, + 128, + 410, + 150 + ], + "type": "text", + "content": "Mount Fuji" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 320, + 208, + 358, + 229 + ], + "lines": [ + { + "bbox": [ + 320, + 208, + 358, + 229 + ], + "spans": [ + { + "bbox": [ + 320, + 208, + 358, + 229 + ], + "type": "text", + "content": "Pendulum Motion" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 376, + 160, + 416, + 194 + ], + "lines": [ + { + "bbox": [ + 376, + 160, + 416, + 194 + ], + "spans": [ + { + "bbox": [ + 376, + 160, + 416, + 194 + ], + "type": "image", + "image_path": "e1c42a497153ef5a8dfb3a8737e16da2cabd3d437ad02ee06a5551aba0e1d2af.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 375, + 208, + 419, + 230 + ], + "lines": [ + { + "bbox": [ + 375, + 208, + 419, + 230 + ], + "spans": [ + { + "bbox": [ + 375, + 208, + 419, + 230 + ], + "type": "text", + "content": "Double-Slit Experiment" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 436, + 152, + 475, + 204 + ], + "blocks": [ + { + "bbox": [ + 443, + 128, + 467, + 150 + ], + "lines": [ + { + "bbox": [ + 443, + 128, + 467, + 150 + ], + "spans": [ + { + "bbox": [ + 443, + 128, + 467, + 150 + ], + "type": "text", + "content": "Eiffel Tower" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 436, + 152, + 475, + 204 + ], + "lines": [ + { + "bbox": [ + 436, + 152, + 475, + 204 + ], + "spans": [ + { + "bbox": [ + 436, + 152, + 475, + 204 + ], + "type": "image", + "image_path": "6fd0668cd32e45b057920fd584981d956b47ad05709d30cd8c47068995420b89.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 436, + 213, + 474, + 224 + ], + "lines": [ + { + "bbox": [ + 436, + 213, + 474, + 224 + ], + "spans": [ + { + "bbox": [ + 436, + 213, + 474, + 224 + ], + "type": "text", + "content": "Flowchart" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 234, + 545, + 267 + ], + "lines": [ + { + "bbox": [ + 305, + 234, + 545, + 267 + ], + "spans": [ + { + "bbox": [ + 305, + 234, + 545, + 267 + ], + "type": "text", + "content": "Figure 8. Sketches produced by SketchAgent for concepts beyond pre-defined categories. The textual input describing the desired concept shown below each image." + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 496, + 159, + 531, + 196 + ], + "blocks": [ + { + "bbox": [ + 496, + 159, + 531, + 196 + ], + "lines": [ + { + "bbox": [ + 496, + 159, + 531, + 196 + ], + "spans": [ + { + "bbox": [ + 496, + 159, + 531, + 196 + ], + "type": "image", + "image_path": "661e67bb48b898e26fae04c7637ad316784dbe6caf1e6bdcdac6af5e7b98976f.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 500, + 209, + 527, + 228 + ], + "lines": [ + { + "bbox": [ + 500, + 209, + 527, + 228 + ], + "spans": [ + { + "bbox": [ + 500, + 209, + 527, + 228 + ], + "type": "text", + "content": "Tic-tac toe" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + } + ], + "index": 27 + }, + { + "bbox": [ + 304, + 278, + 545, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 278, + 545, + 469 + ], + "spans": [ + { + "bbox": [ + 304, + 278, + 545, + 469 + ], + "type": "text", + "content": "canvas back to the agent (see dashed arrow in Fig. 5) and updating the user prompt with the desired edits. To support collaborative human-agent sketching, the canvas remains accessible to both the human user and the agent throughout the entire sketching session. We define an adjustable stopping token, " + }, + { + "bbox": [ + 304, + 278, + 545, + 469 + ], + "type": "inline_equation", + "content": "\\langle \\text{}/\\text{s}\\{\\text{j}\\} \\rangle" + }, + { + "bbox": [ + 304, + 278, + 545, + 469 + ], + "type": "text", + "content": ", which instructs the agent to pause generating the sequence at stroke number " + }, + { + "bbox": [ + 304, + 278, + 545, + 469 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 304, + 278, + 545, + 469 + ], + "type": "text", + "content": ". We then process and render the generated strokes onto the canvas up to that point, then the user can add strokes directly to the canvas to continue the sketch. The user-drawn strokes are processed and converted into the agent's format by reversing our fitting process, i.e., sampling each stroke at multiple " + }, + { + "bbox": [ + 304, + 278, + 545, + 469 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 278, + 545, + 469 + ], + "type": "text", + "content": " values (as shown in Eq. (1)), and selecting the points closest to each cell's center on the grid. The converted user strokes are then chained with the agent's sequence, after which the agent resumes sketching until the next stopping token." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 306, + 481, + 359, + 493 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 481, + 359, + 493 + ], + "spans": [ + { + "bbox": [ + 306, + 481, + 359, + 493 + ], + "type": "text", + "content": "5. Results" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 304, + 498, + 545, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 498, + 545, + 569 + ], + "spans": [ + { + "bbox": [ + 304, + 498, + 545, + 569 + ], + "type": "text", + "content": "We evaluate the performance of our method qualitatively and quantitatively across a selected set of sketching tasks. Additional tasks, evaluations, and examples are provided in the Appendix. All results presented in the paper were generated using Claude3.5-Sonnet [2] as our backbone model, unless stated otherwise." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 306, + 578, + 500, + 590 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 578, + 500, + 590 + ], + "spans": [ + { + "bbox": [ + 306, + 578, + 500, + 590 + ], + "type": "text", + "content": "5.1. Text-Conditioned Sketch Generation" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": "Figures 1 and 8 demonstrate SketchAgent's capability to generate sketches of various concepts that extend beyond standard categories, which includes scientific concepts (e.g., \"the double-slit experiment\", \"pendulum motion\"), diagrams (e.g., \"circuit diagram\", \"a flowchart\"), and notable landmarks (e.g., \"Taj Mahal\", \"Eiffel Tower\"). More examples are provided in the Appendix. To quantitatively evaluate text-conditioned generation we utilize the QuickDraw dataset [47]. We randomly sample 50 categories (out of 345), and apply our method to generate 10 sketch in" + } + ] + } + ], + "index": 34 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23359" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 50, + 71, + 286, + 188 + ], + "blocks": [ + { + "bbox": [ + 50, + 71, + 286, + 188 + ], + "lines": [ + { + "bbox": [ + 50, + 71, + 286, + 188 + ], + "spans": [ + { + "bbox": [ + 50, + 71, + 286, + 188 + ], + "type": "table", + "html": "
GPT-4oGPT-4o -miniClaude3 OpusClaude3.5 -Sonnet*Claude3.5 -Sonnet (SVG)Human (QD [47])
Top10.15 ±0.040.04 ±0.030.13 ±0.040.23 ±0.050.23 ±0.040.27 ±0.07
Top50.30 ±0.060.10 ±0.040.27 ±0.050.44 ±0.030.43 ±0.060.49 ±0.06
Vis.
", + "image_path": "4191ea027dcedc9e279a68a7b72ded7294c69787c2b4b5e228308823487c0c00.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 47, + 190, + 287, + 245 + ], + "lines": [ + { + "bbox": [ + 47, + 190, + 287, + 245 + ], + "spans": [ + { + "bbox": [ + 47, + 190, + 287, + 245 + ], + "type": "text", + "content": "Table 1. Sketch recognition evaluation. Average Top-1 and Top-5 sketch recognition accuracy computed with CLIP zero-shot classifier on 500 sketches from 50 categories. The last row visualizes one sample from each experiment. *Indicates our default settings, which receives the highest accuracy among all models." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 258, + 287, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 258, + 287, + 449 + ], + "spans": [ + { + "bbox": [ + 46, + 258, + 287, + 449 + ], + "type": "text", + "content": "stances per category, resulting in 500 sketches in total. Following common practice [105, 106, 115, 116], we utilize a CLIP zero-shot classifier [78] to evaluate how well the generated sketches depict the intended category. We compare the performance of different multimodal LLMs by repeating the same process with GPT-4o-mini [74], GPT-4o [74], and Claude3-Opus [2] as our backbone model (in addition to Claude3.5-Sonnet [2], our default backbone). As a baseline, we include human-drawn sketches sampled from the QuickDraw dataset [47]. The average Top-1 and Top-5 sketch classification accuracy are presented in Table 1. As can be seen, human sketches achieve the highest recognition accuracy, with Claude3.5-Sonnet performing best among all models, approaching human-level rates under the CLIP-score metric. More evaluation of confusion patterns and visualization of the data are provided in the Appendix." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 450, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 450, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 450, + 288, + 714 + ], + "type": "text", + "content": "We additionally compare to prompting Claude3.5-Sonnet to directly generate SVGs using the following prompt: \"Write SVG string that draws a sketch of a . Use only black and white colors\". The corresponding scores are shown in the fifth column of Tab. 1. While this approach achieves recognition scores comparable to those of SketchAgent, the outputs are often characterized by uniform and precise shapes, failing to replicate the fluidity and natural irregularity of free-hand human sketches (e.g., Fig. 3). To evaluate how \"human-like\" our agent's sketches appear, we conduct a two alternative forced choice (2AFC) user study with 150 participants. Each participant was presented with pairs of sketches depicting the same object class produced by different methods, and asked to choose the sketch they believed was human-drawn. 150 sketches across 50 object classes were tested, comparing three methods: direct prompting, SketchAgent, and human sketches from QuickDraw (see Appendix for details). Results indicate SketchAgent's drawings appeared more human-like, being chosen as human-drawn in " + }, + { + "bbox": [ + 46, + 450, + 288, + 714 + ], + "type": "inline_equation", + "content": "74.90 \\pm 3.35\\%" + }, + { + "bbox": [ + 46, + 450, + 288, + 714 + ], + "type": "text", + "content": " of cases when compared with direct prompting. When compared to human drawings," + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 309, + 72, + 542, + 144 + ], + "blocks": [ + { + "bbox": [ + 309, + 72, + 542, + 144 + ], + "lines": [ + { + "bbox": [ + 309, + 72, + 542, + 144 + ], + "spans": [ + { + "bbox": [ + 309, + 72, + 542, + 144 + ], + "type": "image", + "image_path": "423a11037c664387b1a182431eb6f4eedd6120c7e5bf6f031d5dfdee35b62b69.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 144, + 545, + 167 + ], + "lines": [ + { + "bbox": [ + 306, + 144, + 545, + 167 + ], + "spans": [ + { + "bbox": [ + 306, + 144, + 545, + 167 + ], + "type": "text", + "content": "Figure 9. SketchAgent gradually draws stroke-by-stroke, each stroke is annotated by the agent with a semantic meaning." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 310, + 172, + 426, + 243 + ], + "blocks": [ + { + "bbox": [ + 310, + 172, + 426, + 243 + ], + "lines": [ + { + "bbox": [ + 310, + 172, + 426, + 243 + ], + "spans": [ + { + "bbox": [ + 310, + 172, + 426, + 243 + ], + "type": "image", + "image_path": "04d463e6f03e007611601410ebbcd7f8b365b30eb1b0c6400a7333a2068fb197.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 311, + 246, + 539, + 257 + ], + "lines": [ + { + "bbox": [ + 311, + 246, + 539, + 257 + ], + "spans": [ + { + "bbox": [ + 311, + 246, + 539, + 257 + ], + "type": "text", + "content": "Figure 10. Recognition rate as a function of sketch complexity." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 430, + 172, + 487, + 236 + ], + "blocks": [ + { + "bbox": [ + 430, + 172, + 487, + 236 + ], + "lines": [ + { + "bbox": [ + 430, + 172, + 487, + 236 + ], + "spans": [ + { + "bbox": [ + 430, + 172, + 487, + 236 + ], + "type": "image", + "image_path": "07f9d2c145aadce7dd680e4e50f8d2970cc25e1e13d6710c4e467d2cc83efedf.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 490, + 172, + 545, + 236 + ], + "blocks": [ + { + "bbox": [ + 490, + 172, + 545, + 236 + ], + "lines": [ + { + "bbox": [ + 490, + 172, + 545, + 236 + ], + "spans": [ + { + "bbox": [ + 490, + 172, + 545, + 236 + ], + "type": "image", + "image_path": "969d296403a23d3904cb4e0625589d27622dce0da84e52d44ebd61be07008104.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 266, + 545, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 266, + 545, + 301 + ], + "spans": [ + { + "bbox": [ + 305, + 266, + 545, + 301 + ], + "type": "text", + "content": "users slightly preferred human sketches " + }, + { + "bbox": [ + 305, + 266, + 545, + 301 + ], + "type": "inline_equation", + "content": "(54.68 \\pm 4.61\\%)" + }, + { + "bbox": [ + 305, + 266, + 545, + 301 + ], + "type": "text", + "content": " over SketchAgent's, while direct prompting was chosen only " + }, + { + "bbox": [ + 305, + 266, + 545, + 301 + ], + "type": "inline_equation", + "content": "38.9 \\pm 5.55\\%" + }, + { + "bbox": [ + 305, + 266, + 545, + 301 + ], + "type": "text", + "content": " of the time." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 303, + 545, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 303, + 545, + 386 + ], + "spans": [ + { + "bbox": [ + 304, + 303, + 545, + 386 + ], + "type": "text", + "content": "Lastly, to quantitatively analyze the effect of concept complexity, we study the case of drawing letters. We systematically increase sketch complexity by adding letters to the target concept (e.g., from 'ABC' to 'ABCDEFGH') and count the number of correctly recognized letters in each sketch. The graph in Fig. 10 shows that performance decreases as the complexity increases." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 396, + 427, + 408 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 396, + 427, + 408 + ], + "spans": [ + { + "bbox": [ + 306, + 396, + 427, + 408 + ], + "type": "text", + "content": "5.2. Sequential Sketching" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 415, + 545, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 415, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 415, + 545, + 714 + ], + "type": "text", + "content": "Figure 9 shows stroke-by-stroke sketch generation by SketchAgent, with the labels on the right indicating the sketching order and the meaning our agent associates with each generated stroke (see Appendix for more examples). Stroke annotation during generation is enabled by utilizing the prior of the backbone LLM, providing a valuable feature for analysis and data collection [37, 65, 107, 126, 128]. In Fig. 11, we illustrate why accounting for the sequential nature of sketching more closely emulates the process of human drawing. We present the sketch creation process of SketchAgent alongside SVGDreamer [116], SketchRNN [42], and a human sketch sampled from QuickDraw [47]. SVGDreamer (first row), is an optimization-based method, where a set of randomly initialized parametric curves (left-most column) are iteratively refined to form a sketch, guided by a pretrained text-to-image diffusion model [82]. This process is time-consuming, taking 2000 iterations (1.6 hours), which makes it unsuitable for interactive sketching. While the final sketch (rightmost column) appears detailed and artistic due to the powerful vision backbone, the intermediate sketching and individual strokes lack clear semantic meaning. In contrast, SketchRNN (second row) is a sequential generative model trained on human-drawn dataset, producing sketches in real-time with strokes added progressively, emulating closer a human-like sketching pro" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23360" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 55, + 83, + 275, + 307 + ], + "blocks": [ + { + "bbox": [ + 104, + 73, + 226, + 83 + ], + "lines": [ + { + "bbox": [ + 104, + 73, + 226, + 83 + ], + "spans": [ + { + "bbox": [ + 104, + 73, + 226, + 83 + ], + "type": "text", + "content": "SVGDreamer [116]. " + }, + { + "bbox": [ + 104, + 73, + 226, + 83 + ], + "type": "inline_equation", + "content": "\\approx 1.6" + }, + { + "bbox": [ + 104, + 73, + 226, + 83 + ], + "type": "text", + "content": " hours" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 55, + 83, + 275, + 307 + ], + "lines": [ + { + "bbox": [ + 55, + 83, + 275, + 307 + ], + "spans": [ + { + "bbox": [ + 55, + 83, + 275, + 307 + ], + "type": "table", + "html": "
SketchRNN [42].≈4 seconds
SketchAgent (Ours).≈20 seconds
bodyheadleft earright earback right legleft eye
Human [47].≤20 seconds
", + "image_path": "016f679fa238aefe7749838ef72c26678ee84ede4079312c08d64f12dcb3515f.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 309, + 287, + 376 + ], + "lines": [ + { + "bbox": [ + 46, + 309, + 287, + 376 + ], + "spans": [ + { + "bbox": [ + 46, + 309, + 287, + 376 + ], + "type": "text", + "content": "Figure 11. Sequential sketching process. SVGDreamer [116] requires 2000 iterations (1.6 hours) with intermediate steps lacking semantic meaning. SketchRNN [42] operates in real-time with coherent steps but is limited to QuickDraw categories. SketchAgent draw gradually with meaningful strokes and no category restrictions. Human sketches evolve through gradual, meaningful steps." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 389, + 287, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 389, + 287, + 484 + ], + "spans": [ + { + "bbox": [ + 46, + 389, + 287, + 484 + ], + "type": "text", + "content": "cess (as shown in the last row). Similarly, SketchAgent (third row) produces sketches gradually, with each stroke carrying a semantic meaning, by utilizing the sequential nature of its backbone model. While SketchRNN is restricted to generating sketches only within the 345 categories it was trained on, SketchAgent leverages the extensive prior knowledge of its backbone multimodal LLM, enabling it to create sketches of general visual concepts." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 486, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 486, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 486, + 288, + 714 + ], + "type": "text", + "content": "We use the set of 500 samples described in Sec. 5.1 to quantitatively analyze the sequential nature of our agent's sketches compared to human drawings. On the left of Fig. 12, we present histograms comparing the number of strokes in QuickDraw sketches (orange) and our sketches (blue). Most QuickDraw sketches contain 1 to 6 strokes, while our sketches show a broader distribution, peaking between 5 to 10 strokes. This suggests that, on average, QuickDraw sketches appear more abstract. To ensure a balanced comparison of sketches with similar levels of abstraction, we select sketches from both groups with a similar number of strokes (the largest intersection is found in sketches with 4-7 strokes, comprising 204 of our sketches and 120 from QuickDraw) and measure the change in CLIP-Score as a function of the accumulated number of strokes (Fig. 12, right). Both QuickDraw and our sketches exhibit a generally similar pattern, with CLIPScore increasing as more strokes are added, suggesting that sketches become progressively more recognizable as they evolve." + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 315, + 70, + 392, + 175 + ], + "blocks": [ + { + "bbox": [ + 315, + 70, + 392, + 175 + ], + "lines": [ + { + "bbox": [ + 315, + 70, + 392, + 175 + ], + "spans": [ + { + "bbox": [ + 315, + 70, + 392, + 175 + ], + "type": "image", + "image_path": "8bb79ce4851de1d59ede2f9298fd100f1c2e84e32cba837bddaa488f0ba4591f.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 177, + 545, + 243 + ], + "lines": [ + { + "bbox": [ + 304, + 177, + 545, + 243 + ], + "spans": [ + { + "bbox": [ + 304, + 177, + 545, + 243 + ], + "type": "text", + "content": "Figure 12. Sequential sketching analysis of SketchAgent (blue) and Humans [47] (orange). Left: Histograms of stroke distribution per sketch, showing QuickDraw sketches are more abstract on average. Right: CLIPScore as a function of the accumulated number of strokes for sketches containing 4-7 strokes, showing a similar recognition pattern over time." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 399, + 71, + 534, + 174 + ], + "blocks": [ + { + "bbox": [ + 399, + 71, + 534, + 174 + ], + "lines": [ + { + "bbox": [ + 399, + 71, + 534, + 174 + ], + "spans": [ + { + "bbox": [ + 399, + 71, + 534, + 174 + ], + "type": "image", + "image_path": "df5f4609b0c441a7467c3969f5a4245622328392296f8f901211d83ac22edf94.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 249, + 511, + 262 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 249, + 511, + 262 + ], + "spans": [ + { + "bbox": [ + 306, + 249, + 511, + 262 + ], + "type": "text", + "content": "5.3. Human-Agent Collaborative Sketching" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 268, + 545, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 268, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 304, + 268, + 545, + 567 + ], + "type": "text", + "content": "We demonstrate the potential of our system for facilitating interactive human-agent collaboration, resulting in semantically meaningful and recognizable sketches. We design a web-based collaborative sketching environment (Fig. 13A) where users and SketchAgent take turns drawing on a shared canvas to create a recognizable sketch from a given textual concept. Following the evaluation protocol in collabdraw [24], we select 8 simple concepts, based on the agent's demonstrated ability to draw them independently, to focus evaluations on assessing the impact of collaboration. Participants sketched concepts in two modes: solo, where users drew independently, and collab, where users and SketchAgent collaborated, adding one stroke at a time until either was satisfied with the drawing. We collect sketches from 30 participants, resulting in 480 sketches in total. Average CLIP recognition rates are shown in Figure 13B. Collaboratively produced sketches (blue) achieve recognition levels close to those made solely by users and higher than those produced by the agent alone (dashed lines). To assess the contribution of each party in collaborative mode, we analyze partial sketches with only agent-made strokes (pink) or user-made strokes (green), resulting in a significant reduction in recognizability. This suggests that both user and agent contribute meaningfully to the recognizability of the complete sketch." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 575, + 455, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 575, + 455, + 588 + ], + "spans": [ + { + "bbox": [ + 306, + 575, + 455, + 588 + ], + "type": "text", + "content": "5.4. Chat-Based Sketch Editing" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 545, + 713 + ], + "type": "text", + "content": "We next demonstrate the effectiveness of our method in performing interactive text-based sketch editing within a chat dialogue, where the input to the agent combines both text and images. Inspired by [92], we explore edits that involve spatial reasoning and object relations. We focus on three object categories: outdoor, indoor, and animals, with three objects each, and design editing prompts to add objects to the input sketches. For outdoor and indoor objects, we specify relative locations of added concepts, e.g., \"left to\", \"on top of\" (see Fig. 14 left). For the animals category, we tasked" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23361" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 62, + 79, + 141, + 152 + ], + "blocks": [ + { + "bbox": [ + 62, + 79, + 141, + 152 + ], + "lines": [ + { + "bbox": [ + 62, + 79, + 141, + 152 + ], + "spans": [ + { + "bbox": [ + 62, + 79, + 141, + 152 + ], + "type": "image", + "image_path": "1ef282de04d388ff3cea584cc8f62439859b8bb456a0b0f83105927e7b99ad08.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 63, + 161, + 140, + 171 + ], + "lines": [ + { + "bbox": [ + 63, + 161, + 140, + 171 + ], + "spans": [ + { + "bbox": [ + 63, + 161, + 140, + 171 + ], + "type": "text", + "content": "(A) Sketching interface" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 151, + 72, + 265, + 161 + ], + "blocks": [ + { + "bbox": [ + 151, + 72, + 265, + 161 + ], + "lines": [ + { + "bbox": [ + 151, + 72, + 265, + 161 + ], + "spans": [ + { + "bbox": [ + 151, + 72, + 265, + 161 + ], + "type": "image", + "image_path": "9b467c1d65280845bea59527f26ef0b501d740d2b6a7852e9a026da3abbfb9b5.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 152, + 162, + 266, + 171 + ], + "lines": [ + { + "bbox": [ + 152, + 162, + 266, + 171 + ], + "spans": [ + { + "bbox": [ + 152, + 162, + 266, + 171 + ], + "type": "text", + "content": "(B) Collaborative user study results" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 50, + 240, + 286, + 316 + ], + "blocks": [ + { + "bbox": [ + 46, + 174, + 287, + 229 + ], + "lines": [ + { + "bbox": [ + 46, + 174, + 287, + 229 + ], + "spans": [ + { + "bbox": [ + 46, + 174, + 287, + 229 + ], + "type": "text", + "content": "Figure 13. Collaborative sketching evaluation measured using CLIP classification. Sketches created collaboratively (blue) approaching those made solely by users (dashed lines). In collaborative sketches, keeping agent-only strokes (pink) or user-only strokes (green) significantly reduces recognizability." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 50, + 240, + 286, + 316 + ], + "lines": [ + { + "bbox": [ + 50, + 240, + 286, + 316 + ], + "spans": [ + { + "bbox": [ + 50, + 240, + 286, + 316 + ], + "type": "image", + "image_path": "ce280e778c673c55a5fab06f36f56a5a8fff409fc90b55e02f168b8aa652fb50.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 316, + 287, + 338 + ], + "lines": [ + { + "bbox": [ + 47, + 316, + 287, + 338 + ], + "spans": [ + { + "bbox": [ + 47, + 316, + 287, + 338 + ], + "type": "text", + "content": "Figure 14. Chat-based sketch editing. We iteratively prompt SketchAgent to add objects to sketches through chat dialogues." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 355, + 287, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 355, + 287, + 462 + ], + "spans": [ + { + "bbox": [ + 46, + 355, + 287, + 462 + ], + "type": "text", + "content": "the agent with adding accessories to each animal without guidance on their exact placement, testing its ability to infer placement based on semantics (e.g., placing a hat on a head (see Fig. 14 right). The full list of object and editing instructions is provided in the Appendix. We produced a total of 54 sketches. Evaluating the edited sketches reveals that SketchAgent correctly follows instructions " + }, + { + "bbox": [ + 46, + 355, + 287, + 462 + ], + "type": "inline_equation", + "content": "92\\%" + }, + { + "bbox": [ + 46, + 355, + 287, + 462 + ], + "type": "text", + "content": " of the time, with " + }, + { + "bbox": [ + 46, + 355, + 287, + 462 + ], + "type": "inline_equation", + "content": "94\\%" + }, + { + "bbox": [ + 46, + 355, + 287, + 462 + ], + "type": "text", + "content": " accuracy for specified relations and " + }, + { + "bbox": [ + 46, + 355, + 287, + 462 + ], + "type": "inline_equation", + "content": "88\\%" + }, + { + "bbox": [ + 46, + 355, + 287, + 462 + ], + "type": "text", + "content": " accuracy for inferred semantic relations." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 477, + 107, + 489 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 477, + 107, + 489 + ], + "spans": [ + { + "bbox": [ + 47, + 477, + 107, + 489 + ], + "type": "text", + "content": "6. Ablation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 498, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 288, + 713 + ], + "type": "text", + "content": "We evaluate the impact of each component of our method by systematically removing them and measuring sketch recognition rates as detailed in 5.1. We assess the effects of removing the system prompt, omitting the CoT process (i.e., excluding thinking tags and 'think step-by-step' instructions), and modifying ICL (the complete sketch example provided in the user prompt). When modifying ICL, we use a correctly formatted single-stroke example instead of the complete sketch, as fully removing ICL results in outputs that do not follow the expected format making them unparsable. The results in Table 2 show that the full SketchAgent pipeline achieves the highest performance, highlighting the importance of each component. Interestingly, not providing a complete sketch example significantly reduces performance. We additionally ablate the impact of the grid resolution, by varying the resolution from 10 to 100 (see Tab. 2, bottom). Extremely low resolutions degrade performance, while mid-level resolutions outperform " + }, + { + "bbox": [ + 46, + 498, + 288, + 713 + ], + "type": "inline_equation", + "content": "100 \\times 100" + }, + { + "bbox": [ + 46, + 498, + 288, + 713 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 309, + 71, + 545, + 148 + ], + "blocks": [ + { + "bbox": [ + 309, + 71, + 545, + 148 + ], + "lines": [ + { + "bbox": [ + 309, + 71, + 545, + 148 + ], + "spans": [ + { + "bbox": [ + 309, + 71, + 545, + 148 + ], + "type": "table", + "html": "
w/o System Promptw/o CoTModified ICLSketchAgent (full)
Top10.20 ± 0.040.14 ± 0.020.07 ± 0.020.23 ± 0.04
Top50.42 ± 0.030.29 ± 0.040.16 ± 0.030.43 ± 0.06
Grid Size10 × 1025 × 2550 × 5075 × 75
Top1 / Top50.14 / 0.280.19 / 0.420.23 / 0.430.23 / 0.41
", + "image_path": "85be913d1c16ae445104a386d8de0e85fcdfaaf01f3a2e3ddec914a98d957afe.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 152, + 546, + 196 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 152, + 546, + 196 + ], + "spans": [ + { + "bbox": [ + 305, + 152, + 546, + 196 + ], + "type": "text", + "content": "Table 2. Ablation study. Average Top-1 and Top-5 CLIP recognition accuracy. Top: We systematically remove each component in our pipeline, showcasing all components contribute to the agent's full performance. Bottom: Grid resolution ablation." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 210, + 473, + 222 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 210, + 473, + 222 + ], + "spans": [ + { + "bbox": [ + 305, + 210, + 473, + 222 + ], + "type": "text", + "content": "7. Limitations and Future Work" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 227, + 546, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 227, + 546, + 431 + ], + "spans": [ + { + "bbox": [ + 304, + 227, + 546, + 431 + ], + "type": "text", + "content": "SketchAgent has several limitations. First, it is constrained by the priors of the backbone model, primarily optimized for text rather than visual content. As a result, the agent often produces rich textual descriptions of object parts but struggles to convert these into effective sketching actions, resulting in overly abstract and unrecognizable outputs. For example, in Fig. 15A, the agent effectively describes key parts of a unicorn (e.g., the horn), but the sketch is unrecognizable. This constraint also impacts the depiction of human figures (Fig. 15B). While distinctive features (e.g., Frida Kahlo's eyebrows or Michael Jordan's dunk) may be captured well in language, the resulting sketches are overly simple, with an amateur style, lacking expressivity. We expect this issue to improve as future models advance in vision capabilities. Lastly, the agent may struggle with drawing letters and numbers. This could be improved in future work by providing relevant in-context examples." + } + ] + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 307, + 451, + 391, + 495 + ], + "blocks": [ + { + "bbox": [ + 307, + 440, + 392, + 450 + ], + "lines": [ + { + "bbox": [ + 307, + 440, + 392, + 450 + ], + "spans": [ + { + "bbox": [ + 307, + 440, + 392, + 450 + ], + "type": "text", + "content": "(A) Non-recognizable" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 307, + 451, + 391, + 495 + ], + "lines": [ + { + "bbox": [ + 307, + 451, + 391, + 495 + ], + "spans": [ + { + "bbox": [ + 307, + 451, + 391, + 495 + ], + "type": "image", + "image_path": "ab03659d7c4d1a51dd5884bfa72ac04901f6940c37db0ebced6a7e7e0884155b.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 395, + 451, + 459, + 498 + ], + "blocks": [ + { + "bbox": [ + 394, + 441, + 454, + 450 + ], + "lines": [ + { + "bbox": [ + 394, + 441, + 454, + 450 + ], + "spans": [ + { + "bbox": [ + 394, + 441, + 454, + 450 + ], + "type": "text", + "content": "(B) Human figures" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 395, + 451, + 459, + 498 + ], + "lines": [ + { + "bbox": [ + 395, + 451, + 459, + 498 + ], + "spans": [ + { + "bbox": [ + 395, + 451, + 459, + 498 + ], + "type": "image", + "image_path": "a5198ad87ecffd7972bcfcf56f48b5c9045d4741a2ec81cdda15348a70a826d5.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 501, + 546, + 535 + ], + "lines": [ + { + "bbox": [ + 304, + 501, + 546, + 535 + ], + "spans": [ + { + "bbox": [ + 304, + 501, + 546, + 535 + ], + "type": "text", + "content": "Figure 15. Limitations. Sketches of complex concepts (A) and human figures (B) appear too abstract and unrecognizable with non-professional style. (C) Fail to depict letters and numbers." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 466, + 451, + 545, + 498 + ], + "blocks": [ + { + "bbox": [ + 463, + 441, + 542, + 450 + ], + "lines": [ + { + "bbox": [ + 463, + 441, + 542, + 450 + ], + "spans": [ + { + "bbox": [ + 463, + 441, + 542, + 450 + ], + "type": "text", + "content": "(C) Letters and numbers" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 466, + 451, + 545, + 498 + ], + "lines": [ + { + "bbox": [ + 466, + 451, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 466, + 451, + 545, + 498 + ], + "type": "image", + "image_path": "ad6c20de7a2d320b68f6ae43926ba1745d7421378648d09adf4f1b97e1df2888.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 552, + 383, + 565 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 552, + 383, + 565 + ], + "spans": [ + { + "bbox": [ + 306, + 552, + 383, + 565 + ], + "type": "text", + "content": "8. Conclusions" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 570, + 545, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 545, + 712 + ], + "type": "text", + "content": "We presented a method for language-driven, sequential sketch generation, that can produce versatile sketches in real-time and meaningfully engage in collaborative sketching sessions with humans. We show that the prior knowledge embedded in pretrained multimodal LLMs can be effectively leveraged for sketch generation through an intuitive sketching language and a grid canvas, without requiring additional training or fine-tuning. We hope our work represents a meaningful step toward developing general-purpose sketching systems with the potential to enhance human-computer communication and computer-aided ideation." + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "23362" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 163, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 163, + 85 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 163, + 85 + ], + "type": "text", + "content": "9. Acknowledgements" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 91, + 289, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 91, + 289, + 295 + ], + "spans": [ + { + "bbox": [ + 46, + 91, + 289, + 295 + ], + "type": "text", + "content": "We thank Yuval Alaluf, Hila Chefer, Assaf Ben Kish, Joanna Materzynska, Rinon Gal, Elad Richardson, Arnav Verma, and Ellie Arar for providing feedback on early versions of our manuscript. We are also grateful to Justin Yang, who developed the initial Photodraw application and assisted with the interactive human study setup, and to Dingning Cao, who contributed to the SketchAgent interface and application. Special thanks to Yarden Frenkel for his insights, early explorations, and inspiring discussions. This work was partially supported by NSF CAREER #2047191, NSF DRL #2400471, Stanford Human Centered AI Institute Hoffman-Yee Grant, Hyundai Motor Company, ARL grant W911NF-18-2-021, the Zuckerman STEM Leadership Program, and the Viterbi Fellowship. The sponsors had no role in the experimental design or analysis, the decision to publish, or manuscript preparation. The authors have no competing interests to report." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 304, + 107, + 317 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 304, + 107, + 317 + ], + "spans": [ + { + "bbox": [ + 48, + 304, + 107, + 317 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 58, + 325, + 287, + 712 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 58, + 325, + 287, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 325, + 287, + 434 + ], + "spans": [ + { + "bbox": [ + 58, + 325, + 287, + 434 + ], + "type": "text", + "content": "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katie Millicah, Malcolm Reynolds, Roman Ring, Eliza Rutherford, Serkan Cabi, Tengda Han, Zhitao Gong, Sina Samangooei, Marianne Monteiro, Jacob Menick, Sebastian Borgeaud, Andrew Brock, Aida Nematzadeh, Sahand Sharifzadeh, Mikolaj Binkowski, Ricardo Barreira, Oriol Vinyals, Andrew Zisserman, and Karen Simonyan. Flamingo: a visual language model for few-shot learning. Red Hook, NY, USA, 2024. Curran Associates Inc. 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 58, + 436, + 286, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 436, + 286, + 456 + ], + "spans": [ + { + "bbox": [ + 58, + 436, + 286, + 456 + ], + "type": "text", + "content": "[2] Anthropic. Claude. https://www.anthropic.com/claude, 2023. 2, 3, 5, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 58, + 458, + 287, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 458, + 287, + 491 + ], + "spans": [ + { + "bbox": [ + 58, + 458, + 287, + 491 + ], + "type": "text", + "content": "[3] Itamar Berger, Ariel Shamir, Moshe Mahler, Elizabeth Carter, and Jessica Hodgins. Style and abstraction in portrait sketching. ACM Trans. Graph., 32(4), 2013. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 58, + 492, + 287, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 492, + 287, + 546 + ], + "spans": [ + { + "bbox": [ + 58, + 492, + 287, + 546 + ], + "type": "text", + "content": "[4] James Betker, Gabriel Goh, Li Jing, † TimBrooks, Jianfeng Wang, Linjie Li, † LongOuyang, † JuntangZhuang, † JoyceLee, † YufeiGuo, † WesamManassra, † PrafullaDhariwal, † CaseyChu, † YunxinJiao, and Aditya Ramesh. Improving image generation with better captions. 2023. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 58, + 548, + 287, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 548, + 287, + 601 + ], + "spans": [ + { + "bbox": [ + 58, + 548, + 287, + 601 + ], + "type": "text", + "content": "[5] Ayan Kumar Bhunia, Ayan Das, Umar Riaz Muhammad, Yongxin Yang, Timothy M. Hospedales, Tao Xiang, Yulia Gryaditskaya, and Yi-Zhe Song. Pixel: a competitive sketching ai agent. so you think you can sketch? ACM Trans. Graph., 39:166:1-166:15, 2020. 1, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 58, + 603, + 287, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 603, + 287, + 647 + ], + "spans": [ + { + "bbox": [ + 58, + 603, + 287, + 647 + ], + "type": "text", + "content": "[6] Ankan Kumar Bhunia, Salman Khan, Hisham Cholakkal, Rao Muhammad Anwer, Fahad Shahbaz Khan, Jorma Laaksonen, and Michael Felsberg. Doodleformer: Creative sketch drawing with transformers. ECCV, 2022. 1, 2, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 58, + 647, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 647, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 647, + 287, + 712 + ], + "type": "text", + "content": "[7] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel Ziegler, Jeffrey Wu, Clemens Winter, Chris Hesse, Mark Chen, Eric" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 312, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 315, + 73, + 545, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 73, + 545, + 138 + ], + "spans": [ + { + "bbox": [ + 315, + 73, + 545, + 138 + ], + "type": "text", + "content": "Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. Language models are few-shot learners. In Advances in Neural Information Processing Systems, pages 1877-1901. Curran Associates, Inc., 2020. 2, 3, 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 316, + 140, + 545, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 140, + 545, + 194 + ], + "spans": [ + { + "bbox": [ + 316, + 140, + 545, + 194 + ], + "type": "text", + "content": "[8] Sébastien Bubeck, Varun Chandrasekaran, Ronen Eldan, Johannes Gehrke, Eric Horvitz, Ece Kamar, Peter Lee, Yin Tat Lee, Yuanzhi Li, Scott Lundberg, et al. Sparks of artificial general intelligence: Early experiments with gpt-4. arXiv preprint arXiv:2303.12712, 2023. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 316, + 197, + 545, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 197, + 545, + 229 + ], + "spans": [ + { + "bbox": [ + 316, + 197, + 545, + 229 + ], + "type": "text", + "content": "[9] Mu Cai, Zeyi Huang, Yuheng Li, Haohan Wang, and Yong Jae Lee. Delving into LLMs' visual understanding ability using SVG to bridge image and text, 2024. 2, 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 312, + 231, + 545, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 231, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 312, + 231, + 545, + 262 + ], + "type": "text", + "content": "[10] John Canny. A computational approach to edge detection. IEEE Transactions on pattern analysis and machine intelligence, (6):679-698, 1986. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 312, + 264, + 545, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 264, + 545, + 297 + ], + "spans": [ + { + "bbox": [ + 312, + 264, + 545, + 297 + ], + "type": "text", + "content": "[11] Alexandre Carlier, Martin Danelljan, Alexandre Alahi, and Radu Timofte. Deepsvg: A hierarchical generative network for vector graphics animation, 2020. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 312, + 299, + 545, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 299, + 545, + 352 + ], + "spans": [ + { + "bbox": [ + 312, + 299, + 545, + 352 + ], + "type": "text", + "content": "[12] Caroline Chan, Frédo Durand, and Phillip Isola. Learning to generate line drawings that convey geometry and semantics. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7915-7925, 2022. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 312, + 354, + 545, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 354, + 545, + 409 + ], + "spans": [ + { + "bbox": [ + 312, + 354, + 545, + 409 + ], + "type": "text", + "content": "[13] Zhenfang Chen, Qinhong Zhou, Yikang Shen, Yining Hong, Zhiqing Sun, Dan Gutfreund, and Chuang Gan. Visual chain-of-thought prompting for knowledge-based visual reasoning. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1254-1262, 2024. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 312, + 411, + 545, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 411, + 545, + 443 + ], + "spans": [ + { + "bbox": [ + 312, + 411, + 545, + 443 + ], + "type": "text", + "content": "[14] Changwoon Choi, Jaeah Lee, Jaesik Park, and Young Min Kim. 3doodle: Compact abstraction of objects with 3d strokes. ACM Trans. Graph., 43(4), 2024. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 312, + 445, + 545, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 445, + 545, + 477 + ], + "spans": [ + { + "bbox": [ + 312, + 445, + 545, + 477 + ], + "type": "text", + "content": "[15] Xiangxiang Chu, Jianlin Su, Bo Zhang, and Chunhua Shen. Visionllama: A unified llama backbone for vision tasks, 2024. 2, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 312, + 479, + 545, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 479, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 312, + 479, + 545, + 544 + ], + "type": "text", + "content": "[16] Ayan Das, Yongxin Yang, Timothy Hospedales, Tao Xiang, and Yi-Zhe Song. Béziersketch: A generative model for scalable vector sketches. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXVI 16, pages 632-647. Springer, 2020. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 312, + 546, + 545, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 546, + 545, + 623 + ], + "spans": [ + { + "bbox": [ + 312, + 546, + 545, + 623 + ], + "type": "text", + "content": "[17] Nicholas Davis, Chih-PIn Hsiao, Kunwar Yashraj Singh, Lisa Li, Sanat Moningi, and Brian Magerko. Drawing apprentice: An enactive co-creative agent for artistic collaboration. In Proceedings of the 2015 ACM SIGCHI Conference on Creativity and Cognition, page 185–186, New York, NY, USA, 2015. Association for Computing Machinery. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 312, + 624, + 545, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 624, + 545, + 656 + ], + "spans": [ + { + "bbox": [ + 312, + 624, + 545, + 656 + ], + "type": "text", + "content": "[18] Vanessa de Andrade, Sofia Freire, Monica Baptista, and Yael Shwartz. Drawing as a space for social-cognitive interaction. Education Sciences, 12:45, 2022. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 312, + 658, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 658, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 312, + 658, + 545, + 712 + ], + "type": "text", + "content": "[19] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "23363" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 53, + 73, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 72, + 73, + 287, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 73, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 72, + 73, + 287, + 106 + ], + "type": "text", + "content": "Technologies, Volume 1 (Long and Short Papers), pages 4171-4186, Minneapolis, Minnesota, 2019. Association for Computational Linguistics. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 106, + 287, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 106, + 287, + 160 + ], + "spans": [ + { + "bbox": [ + 53, + 106, + 287, + 160 + ], + "type": "text", + "content": "[20] Sivan Doveh, Shaked Perek, M Jehanzeb Mirza, Wei Lin, Amit Alfassy, Assaf Arbelle, Shimon Ullman, and Leonid Karlinsky. Towards multimodal in-context learning for vision & language models. arXiv preprint arXiv:2403.12736, 2024. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 162, + 288, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 162, + 288, + 194 + ], + "spans": [ + { + "bbox": [ + 54, + 162, + 288, + 194 + ], + "type": "text", + "content": "[21] Mathias Eitz, James Hays, and Marc Alexa. How do humans sketch objects? ACM Trans. Graph., 31(4), 2012. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 195, + 288, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 195, + 288, + 259 + ], + "spans": [ + { + "bbox": [ + 53, + 195, + 288, + 259 + ], + "type": "text", + "content": "[22] Ziv Epstein, Aaron Hertzmann, Laura Mariah Herman, Robert Mahari, Morgan R. Frank, Matthew Groh, Hope Schroeder, Amy Smith, Memo Akten, Jessica Fjeld, Hany Farid, Neil Leach, Alex Pentland, and Olga Russakovsky. Art and the science of generative ai. Science, 380:1110-1111, 2023. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 261, + 288, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 261, + 288, + 293 + ], + "spans": [ + { + "bbox": [ + 53, + 261, + 288, + 293 + ], + "type": "text", + "content": "[23] Judy Fan, Wilma A. Bainbridge, Rebecca Chamberlain, and Jeffrey D. Wammes. Drawing as a versatile cognitive tool. Nature Reviews Psychology, 2:556 - 568, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 294, + 288, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 294, + 288, + 348 + ], + "spans": [ + { + "bbox": [ + 53, + 294, + 288, + 348 + ], + "type": "text", + "content": "[24] Judith E. Fan, Monica Dinculescu, and David Ha. collaborate: An environment for collaborative sketching with an artificial agent. In Proceedings of the 2019 Conference on Creativity and Cognition, page 556-561, New York, NY, USA, 2019. Association for Computing Machinery. 3, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 350, + 288, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 350, + 288, + 392 + ], + "spans": [ + { + "bbox": [ + 53, + 350, + 288, + 392 + ], + "type": "text", + "content": "[25] Judith E. Fan, Robert D. Hawkins, Mike Wu, and Noah D. Goodman. Pragmatic Inference and Visual Abstraction Enable Contextual Flexibility During Visual Communication. Computational Brain & Behavior, 3(1):86-101, 2020. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 394, + 288, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 394, + 288, + 426 + ], + "spans": [ + { + "bbox": [ + 53, + 394, + 288, + 426 + ], + "type": "text", + "content": "[26] Judith E Fan, Wilma A Bainbridge, Rebecca Chamberlain, and Jeffrey D Wammes. Drawing as a versatile cognitive tool. Nature Reviews Psychology, 2(9):556-568, 2023. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 426, + 288, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 426, + 288, + 459 + ], + "spans": [ + { + "bbox": [ + 53, + 426, + 288, + 459 + ], + "type": "text", + "content": "[27] Logan Fiorella and Shelbi Kuhlmann. Creating drawings enhances learning by teaching. Journal of Educational Psychology, 112(4):811, 2020. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 460, + 288, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 460, + 288, + 503 + ], + "spans": [ + { + "bbox": [ + 53, + 460, + 288, + 503 + ], + "type": "text", + "content": "[28] Kenneth Forbus, Jeffrey Usher, Andrew Lovett, Kate Lockwood, and Jon Wetzel. Cogsketch: Sketch understanding for cognitive science research and for education. Topics in Cognitive Science, 3(4):648-666, 2011. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 504, + 288, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 504, + 288, + 558 + ], + "spans": [ + { + "bbox": [ + 53, + 504, + 288, + 558 + ], + "type": "text", + "content": "[29] Kevin Frans, L. B. Soros, and Olaf Witkowski. Clipdraw: exploring text-to-drawing synthesis through language-image encoders. In Proceedings of the 36th International Conference on Neural Information Processing Systems, Red Hook, NY, USA, 2024. Curran Associates Inc. 1, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 559, + 288, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 559, + 288, + 613 + ], + "spans": [ + { + "bbox": [ + 53, + 559, + 288, + 613 + ], + "type": "text", + "content": "[30] Xingyu Fu, Yushi Hu, Bangzheng Li, Yu Feng, Haoyu Wang, Xudong Lin, Dan Roth, Noah A Smith, Wei-Chiu Ma, and Ranjay Krishna. Blink: Multimodal large language models can see but not perceive. arXiv preprint arXiv:2404.12390, 2024. 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 53, + 614, + 288, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 614, + 288, + 669 + ], + "spans": [ + { + "bbox": [ + 53, + 614, + 288, + 669 + ], + "type": "text", + "content": "[31] Rinon Gal, Yael Vinker, Yuval Alaluf, Amit Bermano, Daniel Cohen-Or, Ariel Shamir, and Gal Chechik. Breathing life into sketches using text-to-video priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4325–4336, 2024. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 53, + 670, + 288, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 670, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 53, + 670, + 288, + 712 + ], + "type": "text", + "content": "[32] Yaroslav Ganin, Tejas D. Kulkarni, Igor Babuschkin, S. M. Ali Eslami, and Oriol Vinyals. Synthesizing programs for images using reinforced adversarial learning. *ArXiv*, abs/1804.01118, 2018. 3" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 312, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 312, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 312, + 73, + 545, + 106 + ], + "type": "text", + "content": "[33] Yaroslav Ganin, Sergey Bartunov, Yujia Li, Ethan Keller, and Stefano Saliceti. Computer-aided design as language. In Neural Information Processing Systems, 2021. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 312, + 107, + 545, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 107, + 545, + 161 + ], + "spans": [ + { + "bbox": [ + 312, + 107, + 545, + 161 + ], + "type": "text", + "content": "[34] Chengying Gao, Qi Liu, Qi Xu, Limin Wang, Jianzhuang Liu, and Changqing Zou. Sketchycoco: Image generation from freehand scene sketches. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5174-5183, 2020. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 312, + 162, + 545, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 162, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 312, + 162, + 545, + 205 + ], + "type": "text", + "content": "[35] Simon Garrod, Nicolas Fay, John Lee, Jon Oberlander, and Tracy MacLeod. Foundations of representation: Where might graphical symbol systems come from? Cognitive Science, 31(6):961-987, 2007. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 312, + 206, + 545, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 206, + 545, + 237 + ], + "spans": [ + { + "bbox": [ + 312, + 206, + 545, + 237 + ], + "type": "text", + "content": "[36] Songwei Ge, Vedanuj Goswami, C Lawrence Zitnick, and Devi Parikh. Creative sketch generation. arXiv preprint arXiv:2011.10039, 2020. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 312, + 239, + 545, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 239, + 545, + 271 + ], + "spans": [ + { + "bbox": [ + 312, + 239, + 545, + 271 + ], + "type": "text", + "content": "[37] Songwei Ge, Vedanuj Goswami, Larry Zitnick, and Devi Parikh. Creative sketch generation. In International Conference on Learning Representations, 2021. 2, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 312, + 273, + 545, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 273, + 545, + 338 + ], + "spans": [ + { + "bbox": [ + 312, + 273, + 545, + 338 + ], + "type": "text", + "content": "[38] Hinnie Gijlers, Armin Weinberger, Alieke Mattia van Dijk, Lars Bollen, and Wouter van Joolingen. Collaborative drawing on a shared digital canvas in elementary science education: The effects of script and task awareness support. International Journal of Computer-Supported Collaborative Learning, 8(4):427-453, 2013. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 312, + 338, + 545, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 338, + 545, + 370 + ], + "spans": [ + { + "bbox": [ + 312, + 338, + 545, + 370 + ], + "type": "text", + "content": "[39] Gabriela Goldschmidt. Serial sketching: visual problem solving in designing. Cybernetics and System, 23(2):191-219, 1992. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 312, + 372, + 545, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 372, + 545, + 415 + ], + "spans": [ + { + "bbox": [ + 312, + 372, + 545, + 415 + ], + "type": "text", + "content": "[40] Yulia Gryaditskaya, Mark Sypesteyn, Jan Willem Hoftijzer, Sylvia Pont, Frédo Durand, and Adrien Bousseau. Opensketch: a richly-annotated dataset of product design sketches. ACM Trans. Graph., 38(6), 2019. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 312, + 416, + 545, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 416, + 545, + 503 + ], + "spans": [ + { + "bbox": [ + 312, + 416, + 545, + 503 + ], + "type": "text", + "content": "[41] Tianrui Guan, Fuxiao Liu, Xiyang Wu, Ruiqi Xian, Zongxia Li, Xiaoyu Liu, Xijun Wang, Lichang Chen, Furong Huang, Yaser Yacoob, Dinesh Manocha, and Tianyi Zhou. Hallusionbench: An advanced diagnostic suite for entangled language hallucination and visual illusion in large vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14375-14385, 2024. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 312, + 504, + 545, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 504, + 545, + 535 + ], + "spans": [ + { + "bbox": [ + 312, + 504, + 545, + 535 + ], + "type": "text", + "content": "[42] David Ha and Douglas Eck. A neural representation of sketch drawings. CoRR, abs/1704.03477, 2017. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 312, + 537, + 545, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 537, + 545, + 581 + ], + "spans": [ + { + "bbox": [ + 312, + 537, + 545, + 581 + ], + "type": "text", + "content": "[43] Yucheng Han, China. Xiaoyan Zhang, Xin Chen, Xu Yang, Zhibin Wang, Gang Yu, Bin Fu, and Hanwang Zhang. Chartlama: A multimodal llm for chart understanding and generation. ArXiv, abs/2311.16483, 2023. 2, 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 312, + 582, + 545, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 582, + 545, + 635 + ], + "spans": [ + { + "bbox": [ + 312, + 582, + 545, + 635 + ], + "type": "text", + "content": "[44] Yushi Hu, Weijia Shi, Xingyu Fu, Dan Roth, Mari Ostendorf, Luke Zettlemoyer, Noah A Smith, and Ranjay Krishna. Visual sketchpad: Sketching as a visual chain of thought for multimodal language models. arXiv preprint arXiv:2406.09403, 2024. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 312, + 636, + 545, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 636, + 545, + 689 + ], + "spans": [ + { + "bbox": [ + 312, + 636, + 545, + 689 + ], + "type": "text", + "content": "[45] Francisco Javier Ibarrola, Tomas Lawton, and Kazjon Grace. A collaborative, interactive and context-aware drawing agent for co-creative design. IEEE Transactions on Visualization and Computer Graphics, 30:5525-5537, 2022. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 312, + 691, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 691, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 312, + 691, + 545, + 713 + ], + "type": "text", + "content": "[46] Ajay Jain, Amber Xie, and Pieter Abbeel. Vectorfusion: Text-to-svg by abstracting pixel-based diffusion models. In" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "23364" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 53, + 72, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 72, + 72, + 287, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 72, + 287, + 105 + ], + "spans": [ + { + "bbox": [ + 72, + 72, + 287, + 105 + ], + "type": "text", + "content": "Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1911-1920, 2023. 1, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 108, + 287, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 108, + 287, + 140 + ], + "spans": [ + { + "bbox": [ + 53, + 108, + 287, + 140 + ], + "type": "text", + "content": "[47] Jongejan Jonas, Rowley Henry, Kawashima Takashi, Kim Jongmin, and Fox-Gieg Nick. The Quick, Draw! - A.I. Experiment, 2016. 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 143, + 288, + 176 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 143, + 288, + 176 + ], + "spans": [ + { + "bbox": [ + 53, + 143, + 288, + 176 + ], + "type": "text", + "content": "[48] David Kaiser. Drawing theories apart: The dispersion of Feynman diagrams in postwar physics. University of Chicago Press, 2019. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 178, + 287, + 244 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 178, + 287, + 244 + ], + "spans": [ + { + "bbox": [ + 53, + 178, + 287, + 244 + ], + "type": "text", + "content": "[49] Pegah Karimi, Jeba Rezwana, Safat Siddiqui, Mary Lou Maher, and Nasrin Dehbozorgi. Creative sketching partner: an analysis of human-ai co-creativity. In Proceedings of the 25th International Conference on Intelligent User Interfaces, page 221–230, New York, NY, USA, 2020. Association for Computing Machinery. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 246, + 287, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 246, + 287, + 312 + ], + "spans": [ + { + "bbox": [ + 53, + 246, + 287, + 312 + ], + "type": "text", + "content": "[50] Pegah Karimi, Jeba Rezwana, Safat Siddiqui, Mary Lou Maher, and Nasrin Dehbozorgi. Creative sketching partner: an analysis of human-ai co-creativity. In Proceedings of the 25th International Conference on Intelligent User Interfaces, page 221–230, New York, NY, USA, 2020. Association for Computing Machinery. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 314, + 287, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 314, + 287, + 346 + ], + "spans": [ + { + "bbox": [ + 53, + 314, + 287, + 346 + ], + "type": "text", + "content": "[51] Andrew Kyle Lampinen, Stephanie C. Y. Chan, Aaditya K. Singh, and Murray Shanahan. The broader spectrum of in-context learning, 2024. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 349, + 286, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 349, + 286, + 371 + ], + "spans": [ + { + "bbox": [ + 53, + 349, + 286, + 371 + ], + "type": "text", + "content": "[52] Paul Laseau. Graphic thinking for architects and designers. John Wiley & Sons, 2000. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 373, + 287, + 439 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 373, + 287, + 439 + ], + "spans": [ + { + "bbox": [ + 53, + 373, + 287, + 439 + ], + "type": "text", + "content": "[53] Tomas Lawton, Kazjon Grace, and Francisco J Ibarrola. When is a tool a tool? user perceptions of system agency in human-ai co-creative drawing. In Proceedings of the 2023 ACM Designing Interactive Systems Conference, page 1978–1996, New York, NY, USA, 2023. Association for Computing Machinery. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 441, + 287, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 441, + 287, + 508 + ], + "spans": [ + { + "bbox": [ + 53, + 441, + 287, + 508 + ], + "type": "text", + "content": "[54] Tomas Lawton, Francisco J Ibarrola, Dan Ventura, and Kazjon Grace. Drawing with reframer: Emergence and control in co-creative ai. In Proceedings of the 28th International Conference on Intelligent User Interfaces, page 264–277, New York, NY, USA, 2023. Association for Computing Machinery. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 509, + 286, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 509, + 286, + 553 + ], + "spans": [ + { + "bbox": [ + 53, + 509, + 286, + 553 + ], + "type": "text", + "content": "[55] Yong Jae Lee, C. Lawrence Zitnick, and Michael F. Cohen. Shadowdraw: real-time user guidance for freehand drawing. In ACM SIGGRAPH 2011 Papers, New York, NY, USA, 2011. Association for Computing Machinery. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 555, + 286, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 555, + 286, + 609 + ], + "spans": [ + { + "bbox": [ + 53, + 555, + 286, + 609 + ], + "type": "text", + "content": "[56] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. BLIP: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In Proceedings of the 39th International Conference on Machine Learning, pages 12888-12900. PMLR, 2022. 2, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 611, + 286, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 611, + 286, + 666 + ], + "spans": [ + { + "bbox": [ + 53, + 611, + 286, + 666 + ], + "type": "text", + "content": "[57] Mengtian Li, Zhe Lin, Radomir Mech, Ersin Yumer, and Deva Ramanan. Photo-sketching: Inferring contour drawings from images. In 2019 IEEE Winter Conference on Applications of Computer Vision (WACV), pages 1403-1412. IEEE, 2019. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 53, + 669, + 286, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 669, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 53, + 669, + 286, + 712 + ], + "type": "text", + "content": "[58] Tzu-Mao Li, Michal Lukáč, Gharbi Michael, and Jonathan Ragan-Kelley. Differentiable vector graphics rasterization for editing and learning. ACM Trans. Graph. (Proc. SIGGRAPH Asia), 39(6):193:1-193:15, 2020. 1, 3" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 312, + 73, + 545, + 712 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 312, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 312, + 73, + 545, + 106 + ], + "type": "text", + "content": "[59] Yi Li, Yi-Zhe Song, Timothy M. Hospedales, and Shaogang Gong. Free-hand sketch synthesis with deformable stroke models. CoRR, abs/1510.02644, 2015. 1, 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 312, + 108, + 545, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 108, + 545, + 163 + ], + "spans": [ + { + "bbox": [ + 312, + 108, + 545, + 163 + ], + "type": "text", + "content": "[60] Yijun Li, Chen Fang, Aaron Hertzmann, Eli Shechtman, and Ming-Hsuan Yang. Im2pencil: Controllable pencil illustration from photographs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1525-1534, 2019. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 312, + 164, + 545, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 164, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 312, + 164, + 545, + 228 + ], + "type": "text", + "content": "[61] Hangyu Lin, Yanwei Fu, Yu-Gang Jiang, and X. Xue. Sketch-bert: Learning sketch bidirectional encoder representation from transformers by self-supervised learning of sketch gestalt. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6757-6766, 2020. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 312, + 232, + 545, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 232, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 312, + 232, + 545, + 285 + ], + "type": "text", + "content": "[62] Difan Liu, Matthew Fisher, Aaron Hertzmann, and Evangelos Kalogerakis. Neural strokes: Stylized line drawing of 3d shapes. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14204-14213, 2021. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 312, + 287, + 545, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 287, + 545, + 352 + ], + "spans": [ + { + "bbox": [ + 312, + 287, + 545, + 352 + ], + "type": "text", + "content": "[63] Fang Liu, Xiaoming Deng, Yu-Kun Lai, Yong-Jin Liu, Cuixia Ma, and Hongan Wang. Sketchgan: Joint sketch completion and recognition with generative adversarial network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 312, + 354, + 545, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 354, + 545, + 398 + ], + "spans": [ + { + "bbox": [ + 312, + 354, + 545, + 398 + ], + "type": "text", + "content": "[64] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In Advances in Neural Information Processing Systems, pages 34892-34916. Curran Associates, Inc., 2023. 2, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 312, + 399, + 545, + 444 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 399, + 545, + 444 + ], + "spans": [ + { + "bbox": [ + 312, + 399, + 545, + 444 + ], + "type": "text", + "content": "[65] Bria Long, Judith Fan, Holly Huey, Zixian Chai, and Michael Frank. Parallel developmental changes in children's production and recognition of line drawings of visual concepts. Nature Communications, 15, 2024. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 312, + 445, + 545, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 445, + 545, + 555 + ], + "spans": [ + { + "bbox": [ + 312, + 445, + 545, + 555 + ], + "type": "text", + "content": "[66] Arjun Majumdar, Anurag Ajay, Xiaohan Zhang, Pranav Putta, Sriram Yenamandra, Mikael Henaff, Sneha Silwal, Paul Mcvay, Oleksandr Maksymets, Sergio Arnaud, Karmesh Yadav, Qiyang Li, Ben Newman, Mohit Sharma, Vincent Berges, Shiqi Zhang, Pulkit Agrawal, Yonatan Bisk, Dhruv Batra, Mrinal Kalakrishnan, Franziska Meier, Chris Paxton, Sasha Sax, and Aravind Rajeswaran. Openeqa: Embodied question answering in the era of foundation models. In Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 312, + 556, + 545, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 556, + 545, + 612 + ], + "spans": [ + { + "bbox": [ + 312, + 556, + 545, + 612 + ], + "type": "text", + "content": "[67] William P. McCarthy, Justin Matejka, Karl D.D. Willis, Judith E. Fan, and Yewen Pu. Communicating design intent using drawing and text. In Proceedings of the 16th Conference on Creativity & Cognition, page 512-519, New York, NY, USA, 2024. Association for Computing Machinery. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 312, + 613, + 545, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 613, + 545, + 668 + ], + "spans": [ + { + "bbox": [ + 312, + 613, + 545, + 668 + ], + "type": "text", + "content": "[68] John FJ Mellor, Eunbyung Park, Yaroslav Ganin, Igor Babuschkin, Tejas Kulkarni, Dan Rosenbaum, Andy Ballard, Theophane Weber, Oriol Vinyals, and SM Eslami. Unsupervised doodling and painting with improved spiral. arXiv preprint arXiv:1910.01007, 2019. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 312, + 669, + 545, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 669, + 545, + 712 + ], + "spans": [ + { + "bbox": [ + 312, + 669, + 545, + 712 + ], + "type": "text", + "content": "[69] Daniela Mihai and Jonathon Hare. Learning to draw: Emergent communication through sketching. Advances in Neural Information Processing Systems, 34:7153-7166, 2021. 3" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "23365" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 53, + 73, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 73, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 73, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 53, + 73, + 287, + 127 + ], + "type": "text", + "content": "[70] Chancharik Mitra, Brandon Huang, Trevor Darrell, and Roei Herzig. Compositional chain-of-thought prompting for large multimodal models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14420-14431, 2024. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 128, + 288, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 128, + 288, + 183 + ], + "spans": [ + { + "bbox": [ + 53, + 128, + 288, + 183 + ], + "type": "text", + "content": "[71] Kushin Mukherjee, Holly Huey, Xuanchen Lu, Yael Vinker, Rio Aguina-Kang, Ariel Shamir, and Judith Fan. Seva: Leveraging sketches to evaluate alignment between human and machine visual abstraction. In Advances in Neural Information Processing Systems, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 184, + 288, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 184, + 288, + 215 + ], + "spans": [ + { + "bbox": [ + 53, + 184, + 288, + 215 + ], + "type": "text", + "content": "[72] Omar W Nasim. Observing by hand: sketching the nebulae in the nineteenth century. University of Chicago Press, 2019. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 217, + 288, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 217, + 288, + 293 + ], + "spans": [ + { + "bbox": [ + 53, + 217, + 288, + 293 + ], + "type": "text", + "content": "[73] Changhoon Oh, Jungwoo Song, Jinhan Choi, Seonghyeon Kim, Sungwoo Lee, and Bongwon Suh. I lead, you help but only with enough details: Understanding user experience of co-creation with artificial intelligence. In Proceedings of the 2018 CHI Conference on Human Factors in Computing Systems, page 1–13, New York, NY, USA, 2018. Association for Computing Machinery. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 294, + 249, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 294, + 249, + 304 + ], + "spans": [ + { + "bbox": [ + 53, + 294, + 249, + 304 + ], + "type": "text", + "content": "[74] OpenAI. Gpt-4 technical report, 2024. 2, 3, 4, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 305, + 287, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 305, + 287, + 358 + ], + "spans": [ + { + "bbox": [ + 53, + 305, + 287, + 358 + ], + "type": "text", + "content": "[75] Dustin Podell, Zion English, Kyle Lacey, A. Blattmann, Tim Dockhorn, Jonas Muller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. ArXiv, abs/2307.01952, 2023. 2, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 360, + 288, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 360, + 288, + 414 + ], + "spans": [ + { + "bbox": [ + 53, + 360, + 288, + 414 + ], + "type": "text", + "content": "[76] Yonggang Qi, Guoyao Su, Pinaki Nath Chowdhury, Mingkang Li, and Yi-Zhe Song. Sketchlattice: Latticed representation for sketch manipulation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 953–961, 2021. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 415, + 287, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 415, + 287, + 458 + ], + "spans": [ + { + "bbox": [ + 53, + 415, + 287, + 458 + ], + "type": "text", + "content": "[77] Shuwen Qiu, Sirui Xie, Lifeng Fan, Tao Gao, Song-Chun Zhu, and Yixin Zhu. Emergent graphical conventions in a visual communication game. arXiv preprint arXiv:2111.14210, 2021. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 460, + 287, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 460, + 287, + 524 + ], + "spans": [ + { + "bbox": [ + 53, + 460, + 287, + 524 + ], + "type": "text", + "content": "[78] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. CoRR, abs/2103.00020, 2021. 1, 3, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 526, + 287, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 526, + 287, + 579 + ], + "spans": [ + { + "bbox": [ + 53, + 526, + 287, + 579 + ], + "type": "text", + "content": "[79] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. J. Mach. Learn. Res., 21(1), 2020. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 581, + 287, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 581, + 287, + 624 + ], + "spans": [ + { + "bbox": [ + 53, + 581, + 287, + 624 + ], + "type": "text", + "content": "[80] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 2, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 625, + 287, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 625, + 287, + 679 + ], + "spans": [ + { + "bbox": [ + 53, + 625, + 287, + 679 + ], + "type": "text", + "content": "[81] Leo Sampaio Ferraz Ribeiro, Tu Bui, John P. Collomosse, and Moacir Antonelli Ponti. Sketchformer: Transformer-based representation for sketched structure. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14141-14150, 2020. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 53, + 681, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 681, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 53, + 681, + 287, + 713 + ], + "type": "text", + "content": "[82] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models, 2022. 1, 2, 3, 6" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 312, + 73, + 545, + 713 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 312, + 73, + 545, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 73, + 545, + 127 + ], + "spans": [ + { + "bbox": [ + 312, + 73, + 545, + 127 + ], + "type": "text", + "content": "[83] Daniel Rose, Vaishnavi Himakunthala, Andy Ouyang, Ryan He, Alex Mei, Yujie Lu, Michael Saxon, Chinmay Sonar, Diba Mirza, and William Yang Wang. Visual chain of thought: bridging logical gaps with multimodal infillings. arXiv preprint arXiv:2305.02317, 2023. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 312, + 129, + 545, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 129, + 545, + 205 + ], + "spans": [ + { + "bbox": [ + 312, + 129, + 545, + 205 + ], + "type": "text", + "content": "[84] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S. Sara Mahdavi, Rapha Gontijo Lopes, Tim Salimans, Jonathan Ho, David J Fleet, and Mohammad Norouzi. Photorealistic text-to-image diffusion models with deep language understanding, 2022. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 312, + 207, + 545, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 207, + 545, + 239 + ], + "spans": [ + { + "bbox": [ + 312, + 207, + 545, + 239 + ], + "type": "text", + "content": "[85] Patsorn Sangkloy, Nathan Burnell, Cusuh Ham, and James Hays. The sketchy database: Learning to retrieve badly drawn bunnies. ACM Trans. Graph., 35(4), 2016. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 312, + 241, + 545, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 241, + 545, + 296 + ], + "spans": [ + { + "bbox": [ + 312, + 241, + 545, + 296 + ], + "type": "text", + "content": "[86] Peter Schaldenbrand, James McCann, and Jean Oh. Frida: A collaborative robot painter with a differentiable, real2sim2real planning environment. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 11712-11718, 2023. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 312, + 298, + 545, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 298, + 545, + 351 + ], + "spans": [ + { + "bbox": [ + 312, + 298, + 545, + 351 + ], + "type": "text", + "content": "[87] Peter Schaldenbrand, Gaurav Parmar, Jun-Yan Zhu, James McCann, and Jean Oh. Cofrida: Self-supervised fine-tuning for human-robot co-painting. In 2024 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2024. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 312, + 354, + 545, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 354, + 545, + 376 + ], + "spans": [ + { + "bbox": [ + 312, + 354, + 545, + 376 + ], + "type": "text", + "content": "[88] Donald A Schon and Vincent DeSanctis. The reflective practitioner: How professionals think in action, 1986. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 312, + 377, + 545, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 377, + 545, + 453 + ], + "spans": [ + { + "bbox": [ + 312, + 377, + 545, + 453 + ], + "type": "text", + "content": "[89] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, Patrick Schramowski, Srivatsa Kundurthy, Katherine Crowson, Ludwig Schmidt, Robert Kaczmarczyk, and Jenia Jitsev. Laion-5b: An open large-scale dataset for training next generation image-text models, 2022. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 312, + 456, + 545, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 456, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 312, + 456, + 545, + 510 + ], + "type": "text", + "content": "[90] Tamar Rott Shaham, Sarah Schwettmann, Franklin Wang, Achyuta Rajaram, Evan Hernandez, Jacob Andreas, and Antonio Torralba. A multimodal automated interpretability agent. In *Forty-first International Conference on Machine Learning*, 2024. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 312, + 512, + 545, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 512, + 545, + 566 + ], + "spans": [ + { + "bbox": [ + 312, + 512, + 545, + 566 + ], + "type": "text", + "content": "[91] Hao Shao, Shengju Qian, Han Xiao, Guanglu Song, Zhuofan Zong, Letian Wang, Yu Liu, and Hongsheng Li. Visual cot: Advancing multi-modal language models with a comprehensive dataset and benchmark for chain-of-thought reasoning. arXiv preprint arXiv:2403.16999, 2024. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 312, + 567, + 545, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 567, + 545, + 632 + ], + "spans": [ + { + "bbox": [ + 312, + 567, + 545, + 632 + ], + "type": "text", + "content": "[92] Pratyusha Sharma, Tamar Rott Shaham, Manel Baradad, Stephanie Fu, Adrian Rodriguez-Munoz, Shivam Duggal, Phillip Isola, and Antonio Torralba. A vision check-up for language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14410-14419, 2024. 3, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 312, + 635, + 545, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 635, + 545, + 667 + ], + "spans": [ + { + "bbox": [ + 312, + 635, + 545, + 667 + ], + "type": "text", + "content": "[93] Jifei Song, Kaiyue Pang, Yi-Zhe Song, Tao Xiang, and Timothy Hospedales. Learning to sketch with shortcut cycle consistency, 2018. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 312, + 670, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 670, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 312, + 670, + 545, + 713 + ], + "type": "text", + "content": "[94] Guoyao Su, Yonggang Qi, Kaiyue Pang, Jie Yang, Yi-Zhe Song, and CVSSP SketchX. Sketchhealer: A graph-to-sequence network for recreating partial human sketches. In BMVC, page 5, 2020. 2" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "23366" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 53, + 72, + 287, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 72, + 287, + 115 + ], + "spans": [ + { + "bbox": [ + 53, + 72, + 287, + 115 + ], + "type": "text", + "content": "[95] Yanpeng Sun, Qiang Chen, Jian Wang, Jingdong Wang, and Zechao Li. Exploring effective factors for improving visual in-context learning. arXiv preprint arXiv:2304.04748, 2023.3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 53, + 117, + 288, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 117, + 288, + 149 + ], + "spans": [ + { + "bbox": [ + 53, + 117, + 288, + 149 + ], + "type": "text", + "content": "[96] Ivan E. Sutherland. Sketchpad—a man-machine graphical communication system, page 391-408. Association for Computing Machinery, New York, NY, USA, 1998. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 150, + 288, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 150, + 288, + 172 + ], + "spans": [ + { + "bbox": [ + 53, + 150, + 288, + 172 + ], + "type": "text", + "content": "[97] Gemini Team. Gemini: A family of highly capable multimodal models, 2024. 2, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 173, + 288, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 173, + 288, + 239 + ], + "spans": [ + { + "bbox": [ + 53, + 173, + 288, + 239 + ], + "type": "text", + "content": "[98] Jakob Tholander and Martin Jonsson. Design ideation with ai - sketching, thinking and talking with generative machine learning models. In Proceedings of the 2023 ACM Design-Interactive Systems Conference, page 1930-1940, New York, NY, USA, 2023. Association for Computing Machinery. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 239, + 288, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 239, + 288, + 271 + ], + "spans": [ + { + "bbox": [ + 53, + 239, + 288, + 271 + ], + "type": "text", + "content": "[99] Shengbang Tong, Zhuang Liu, Yuexiang Zhai, Yi Ma, Yann LeCun, and Saining Xie. Eyes wide shut? exploring the visual shortcomings of multimodal llms, 2024. 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 272, + 287, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 272, + 287, + 338 + ], + "spans": [ + { + "bbox": [ + 48, + 272, + 287, + 338 + ], + "type": "text", + "content": "[100] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothee Lacroix, Baptiste Roziere, Naman Goyal, Eric Hambro, Faisal Azhar, Aurelien Rodriguez, Armand Joulin, Edouard Grave, and Guillaume Lample. Llama: Open and efficient foundation language models. ArXiv, abs/2302.13971, 2023. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 338, + 287, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 338, + 287, + 370 + ], + "spans": [ + { + "bbox": [ + 48, + 338, + 287, + 370 + ], + "type": "text", + "content": "[101] Barbara Tversky. What do sketches say about thinking? AAAI Spring Symp. Sketch Understanding Worksh., 2002. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 371, + 287, + 394 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 371, + 287, + 394 + ], + "spans": [ + { + "bbox": [ + 48, + 371, + 287, + 394 + ], + "type": "text", + "content": "[102] Barbara Tversky. Visualizing thought. In Handbook of human centric visualization, pages 3-40. Springer, 2013. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 394, + 288, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 394, + 288, + 449 + ], + "spans": [ + { + "bbox": [ + 48, + 394, + 288, + 449 + ], + "type": "text", + "content": "[103] Barbara Tversky, Masaki Suwa, Maneesh Agrawala, Julie Heiser, Chris Stolte, Pat Hanrahan, Doantam Phan, Jeff Klingner, Marie-Paule Daniel, Paul Lee, et al. Sketches for design and design of sketches. Human Behaviour in Design: Individuals, Teams, Tools, pages 79-86, 2003. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 449, + 287, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 449, + 287, + 491 + ], + "spans": [ + { + "bbox": [ + 48, + 449, + 287, + 491 + ], + "type": "text", + "content": "[104] Russell Tytler, Vaughan Prain, George Aranda, Joseph Ferguson, and Radhika Gorur. Drawing to reason and learn in science. Journal of Research in Science Teaching, 57(2): 209-231, 2020. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 492, + 287, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 492, + 287, + 548 + ], + "spans": [ + { + "bbox": [ + 48, + 492, + 287, + 548 + ], + "type": "text", + "content": "[105] Yael Vinker, Ehsan Pajouheshgar, Jessica Y. Bo, Roman Christian Bachmann, Amit Haim Bermano, Daniel Cohen-Or, Amir Zamir, and Ariel Shamir. Clipasso: Semantically-aware object sketching. ACM Trans. Graph., 41(4), 2022. 1, 3, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 548, + 287, + 603 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 548, + 287, + 603 + ], + "spans": [ + { + "bbox": [ + 48, + 548, + 287, + 603 + ], + "type": "text", + "content": "[106] Yael Vinker, Yuval Alaluf, Daniel Cohen-Or, and Ariel Shamir. Clipascene: Scene sketching with different types and levels of abstraction. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 4146–4156, 2023. 3, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 603, + 287, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 603, + 287, + 635 + ], + "spans": [ + { + "bbox": [ + 48, + 603, + 287, + 635 + ], + "type": "text", + "content": "[107] Jiawei Wang and Changjian Li. Contextseg: Sketch semantic segmentation by querying the context with attention. arXiv preprint arXiv:2311.16682, 2023. 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 636, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 636, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 636, + 287, + 712 + ], + "type": "text", + "content": "[108] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Brian Ichter, Fei Xia, Ed H. Chi, Quoc V. Le, and Denny Zhou. Chain-of-thought prompting elicits reasoning in large language models. In Proceedings of the 36th International Conference on Neural Information Processing Systems, Red Hook, NY, USA, 2022. Curran Associates Inc. 2, 5" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 72, + 545, + 714 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 307, + 72, + 545, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 72, + 545, + 116 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 545, + 116 + ], + "type": "text", + "content": "[109] Holger Winnemöller, Jan Eric Kyprianidis, and Sven C. Olsen. Xdog: An extended difference-of-gaussians compendium including advanced image stylization. Comput. Graph., 36:740-753, 2012. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 118, + 545, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 545, + 140 + ], + "type": "text", + "content": "[110] World Wide Web Consortium (W3C). Scalable Vector Graphics (SVG), 1999. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 141, + 545, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 141, + 545, + 185 + ], + "spans": [ + { + "bbox": [ + 307, + 141, + 545, + 185 + ], + "type": "text", + "content": "[111] Chenfei Wu, Shengming Yin, Weizhen Qi, Xiaodong Wang, Zecheng Tang, and Nan Duan. Visual chatgpt: Talking, drawing and editing with visual foundation models. arXiv preprint arXiv:2303.04671, 2023. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 186, + 545, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 186, + 545, + 229 + ], + "spans": [ + { + "bbox": [ + 307, + 186, + 545, + 229 + ], + "type": "text", + "content": "[112] Rong Wu, Wanchao Su, Kede Ma, and Jing Liao. Iconshop: Text-guided vector icon synthesis with autoregressive transformers. ACM Transactions on Graphics (TOG), 42:1-14, 2023. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 231, + 545, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 231, + 545, + 285 + ], + "spans": [ + { + "bbox": [ + 308, + 231, + 545, + 285 + ], + "type": "text", + "content": "[113] Chufeng Xiao, Wanchao Su, Jing Liao, Zhouhui Lian, Yi-Zhe Song, and Hongbo Fu. Diffsketching: How differently do people sketch 3d objects? ACM Transactions on Graphics (Proceedings of ACM SIGGRAPH Asia 2022), 41 (4):1-16, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 286, + 545, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 286, + 545, + 320 + ], + "spans": [ + { + "bbox": [ + 308, + 286, + 545, + 320 + ], + "type": "text", + "content": "[114] Saining Xie and Zhuowen Tu. Holistically-nested edge detection. In Proceedings of the IEEE international conference on computer vision, pages 1395-1403, 2015. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 321, + 545, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 321, + 545, + 376 + ], + "spans": [ + { + "bbox": [ + 308, + 321, + 545, + 376 + ], + "type": "text", + "content": "[115] XiMing Xing, Chuang Wang, Haitao Zhou, Jing Zhang, Qian Yu, and Dong Xu. Diffsketcher: Text guided vector sketch synthesis through latent diffusion models. In Advances in Neural Information Processing Systems, pages 15869-15889. Curran Associates, Inc., 2023. 3, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 376, + 545, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 376, + 545, + 432 + ], + "spans": [ + { + "bbox": [ + 308, + 376, + 545, + 432 + ], + "type": "text", + "content": "[116] Ximing Xing, Haitao Zhou, Chuang Wang, Jing Zhang, Dong Xu, and Qian Yu. Svgdreamer: Text guided svg generation with diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4546-4555, 2024. 3, 6, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 433, + 545, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 433, + 545, + 466 + ], + "spans": [ + { + "bbox": [ + 308, + 433, + 545, + 466 + ], + "type": "text", + "content": "[117] Peng Xu, Timothy M. Hospedales, Qiyue Yin, Yi-Zhe Song, Tao Xiang, and Liang Wang. Deep learning for free-hand sketch: A survey and a toolbox, 2020. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 467, + 545, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 467, + 545, + 511 + ], + "spans": [ + { + "bbox": [ + 308, + 467, + 545, + 511 + ], + "type": "text", + "content": "[118] Zhengyuan Yang, Linjie Li, Kevin Lin, Jianfeng Wang, Chung-Ching Lin, Zicheng Liu, and Lijuan Wang. The dawn of Imms: Preliminary explorations with gpt-4v(ision). ArXiv, abs/2309.17421, 2023. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 512, + 545, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 512, + 545, + 567 + ], + "spans": [ + { + "bbox": [ + 308, + 512, + 545, + 567 + ], + "type": "text", + "content": "[119] Zhengyuan Yang, Jianfeng Wang, Linjie Li, Kevin Lin, Chung-Ching Lin, Zicheng Liu, and Lijuan Wang. Idea2img: Iterative self-refinement with gpt-4v (ision) for automatic image design and generation. arXiv preprint arXiv:2310.08541, 2023. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 567, + 545, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 567, + 545, + 624 + ], + "spans": [ + { + "bbox": [ + 308, + 567, + 545, + 624 + ], + "type": "text", + "content": "[120] Ran Yi, Yong-Jin Liu, Yu-Kun Lai, and Paul L Rosin. Apdrawinggan: Generating artistic portrait drawings from face photos with hierarchical gans. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10743-10752, 2019. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 624, + 545, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 624, + 545, + 678 + ], + "spans": [ + { + "bbox": [ + 308, + 624, + 545, + 678 + ], + "type": "text", + "content": "[121] Ran Yi, Yong-Jin Liu, Yu-Kun Lai, and Paul L Rosin. Unpaired portrait drawing generation via asymmetric cycle mapping. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8217-8225, 2020. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 679, + 545, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 679, + 545, + 714 + ], + "spans": [ + { + "bbox": [ + 308, + 679, + 545, + 714 + ], + "type": "text", + "content": "[122] C. Zhang, Weijie Wang, Paul Pangaro, Nikolas Martelaro, and Daragh Byrne. Generative image ai using design sketches as input: Opportunities and challenges. Proceed-" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "23367" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 408 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 72, + 72, + 286, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 72, + 286, + 94 + ], + "spans": [ + { + "bbox": [ + 72, + 72, + 286, + 94 + ], + "type": "text", + "content": "ings of the 15th Conference on Creativity and Cognition, 2023. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 287, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 287, + 151 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 287, + 151 + ], + "type": "text", + "content": "[123] Jiahao Zhang, Bowen Wang, Liangzhi Li, Yuta Nakashima, and Hajime Nagahara. Instruct me more! random prompting for visual in-context learning. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 2597-2606, 2024. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 152, + 288, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 152, + 288, + 184 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 288, + 184 + ], + "type": "text", + "content": "[124] Peiying Zhang, Nanxuan Zhao, and Jing Liao. Text-to-vector generation with neural path representation. ACM Trans. Graph., 43(4), 2024. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 185, + 288, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 185, + 288, + 228 + ], + "spans": [ + { + "bbox": [ + 48, + 185, + 288, + 228 + ], + "type": "text", + "content": "[125] Yuanhan Zhang, Kaiyang Zhou, and Ziwei Liu. What makes good examples for visual in-context learning? Advances in Neural Information Processing Systems, 36: 17773-17794, 2023. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 230, + 288, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 230, + 288, + 274 + ], + "spans": [ + { + "bbox": [ + 48, + 230, + 288, + 274 + ], + "type": "text", + "content": "[126] Zhengming Zhang, Xiaoming Deng, Jinyao Li, Yukun Lai, Cuixia Ma, Yongjin Liu, and Hongan Wang. Stroke-based semantic segmentation for scene-level free-hand sketches. Vis. Comput., 39(12):6309-6321, 2022. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 275, + 288, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 275, + 288, + 319 + ], + "spans": [ + { + "bbox": [ + 48, + 275, + 288, + 319 + ], + "type": "text", + "content": "[127] Zhuosheng Zhang, Aston Zhang, Mu Li, Hai Zhao, George Karypis, and Alex Smola. Multimodal chain-of-thought reasoning in language models. arXiv preprint arXiv:2302.00923, 2023. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 320, + 288, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 320, + 288, + 364 + ], + "spans": [ + { + "bbox": [ + 48, + 320, + 288, + 364 + ], + "type": "text", + "content": "[128] Yixiao Zheng, Kaiyue Pang, Ayan Das, Dongliang Chang, Yi-Zhe Song, and Zhanyu Ma. Creativeseg: Semantic segmentation of creative sketches. IEEE Transactions on Image Processing, 33:2266-2278, 2024. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 365, + 288, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 365, + 288, + 408 + ], + "spans": [ + { + "bbox": [ + 48, + 365, + 288, + 408 + ], + "type": "text", + "content": "[129] Tao Zhou, Chen Fang, Zhaowen Wang, Jimei Yang, Byungmoon Kim, Zhili Chen, Jonathan Brandt, and Demetri Terzopoulos. Learning to sketch with deep q networks and demonstrated strokes. ArXiv, abs/1810.05977, 2018. 2, 3" + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23368" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 13 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/SketchFusion_ Learning Universal Sketch Features through Fusing Foundation Models/1fc738f6-a4cf-4a97-bcd3-ea7682cab3c6_content_list.json b/2025/SketchFusion_ Learning Universal Sketch Features through Fusing Foundation Models/1fc738f6-a4cf-4a97-bcd3-ea7682cab3c6_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..9df4679c7e0aa89d73b58788c68afd6dd8d61217 --- /dev/null +++ b/2025/SketchFusion_ Learning Universal Sketch Features through Fusing Foundation Models/1fc738f6-a4cf-4a97-bcd3-ea7682cab3c6_content_list.json @@ -0,0 +1,1631 @@ +[ + { + "type": "text", + "text": "SketchFusion: Learning Universal Sketch Features through Fusing Foundation Models", + "text_level": 1, + "bbox": [ + 197, + 94, + 799, + 137 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Subhadeep Koley $^{1,2*}$ Tapas Kumar Dutta $^{1*}$ Aneeshan Sain $^{1}$ Pinaki Nath Chowdhury $^{1}$ Ayan Kumar Bhunia $^{1}$ Yi-Zhe Song $^{1,2}$", + "bbox": [ + 145, + 138, + 849, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ SketchX, CVSSP, University of Surrey, United Kingdom.", + "bbox": [ + 267, + 174, + 728, + 191 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ iFlyTek-Surrey Joint Research Centre on Artificial Intelligence.", + "bbox": [ + 243, + 191, + 751, + 210 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{s.koley, a.sain, p.chowdhury, a.bhunia, y-song}@surrey.ac.uk", + "bbox": [ + 225, + 212, + 764, + 227 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/fbf3bc9cc8e26a398b9b324416719f6803ea8f180d4fd170aa1f21e0d30874c3.jpg", + "image_caption": [ + "Figure 1. (Left): Apart from high-resolution image generation, text-to-image diffusion models (e.g., Stable diffusion (SD) [71]) with their innate object understanding capability [84, 107], have shown remarkable performance across a wide range of image-based vision tasks (e.g., segmentation [94], depth estimation [110], etc.). However, upon analysing the PCA representation of SD's intermediate UNet features, we observe that it struggles to achieve similar results when working with freehand abstract sketches (detail in Sec. 4). Unlike pixel-perfect photos, highly abstract freehand sketches are sparse and lack detailed textures and colours [26], making it harder for the SD model to extract meaningful features. Furthermore, investigating the SD denoising process in the frequency domain (via Fourier Transform), we observe the predominance of high-frequency (HF) components, rather than their low-frequency (LF) counterpart - crucial for capturing comprehensive semantic context. To mitigate this inherent bias within SD, we reinforce the diffusion process with another pretrained model (i.e., CLIP [67]) whose bias is complementary (i.e., focuses on LF) to SD. Consequently, the proposed extractor can extract semantically meaningful and accurate features from both sketches and photos, that encapsulate a broader frequency spectrum (i.e., HF and LF). (Right:) Testing the proposed method with different sketch-based discriminative and dense prediction tasks (requiring knowledge of both sketch and image), we find a marked improvement over baseline SD+CLIP hybrid feature extractor." + ], + "image_footnote": [], + "bbox": [ + 91, + 229, + 357, + 352 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8f30d00fa401bee519fd692a7dee43e1a0682d97cbfdfefa69195f81ae3ae1e6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 361, + 229, + 681, + 351 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5350e434ef008889b2e48186422c566193ca212b2f1868ef2cffe5ce4a90c48c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 686, + 229, + 905, + 352 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 532, + 326, + 547 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "While foundation models have revolutionised computer vision, their effectiveness for sketch understanding remains limited by the unique challenges of abstract, sparse visual inputs. Through systematic analysis, we uncover two fundamental limitations: Stable Diffusion (SD) struggles to extract meaningful features from abstract sketches (unlike its success with photos), and exhibits a pronounced frequency-domain bias that suppresses essential low-frequency components needed for sketch understanding. Rather than costly retraining, we address these limitations by strategically combining SD with CLIP, whose strong semantic understanding naturally compensates for SD's spatial-frequency biases. By dynamically injecting CLIP features into SD's denoising process and adaptively aggregating features across semantic levels, our method achieves state-of-the-art performance in sketch retrieval $(+3.35\\%)$ , recognition $(+1.06\\%)$ , segmentation $(+29.42\\%)$ , and correspondence learning $(+21.22\\%)$ , demonstrating the first truly universal sketch feature representation in the era of foundation models.", + "bbox": [ + 86, + 566, + 485, + 867 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 532, + 643, + 547 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The quest for universal sketch-specific features has been a foundational challenge in computer vision [46, 76, 78, 97]. Sketches, with their inherent characteristics of abstraction, sparsity, and cross-modal interpretation, demand fundamentally different feature representations compared to natural images [45, 100]. While foundation models [67, 71] have revolutionised various visual tasks, their effectiveness for sketch understanding remains largely unexplored [47], particularly given sketch's unique position at the intersection of human abstraction and computer vision [26, 45].", + "bbox": [ + 509, + 551, + 906, + 702 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Through systematic analysis, we find that even powerful foundation models like Stable Diffusion (SD) [71] face significant challenges with sketch data. Our pilot studies (Sec. 4) reveal two fundamental limitations. First, while SD excels at extracting task-specific representative features from pixel-perfect RGB images, it struggles with abstract freehand sketches, likely due to its large-scale pretraining being predominantly focused on photographic data [71]. Second, and more intriguingly, frequency-domain analysis of SD's UNet reveals an inherent architectural bias: the model systematically promotes high-frequency (HF) components while suppressing low-frequency (LF) information - a characteristic that proves particularly problematic", + "bbox": [ + 509, + 704, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution", + "bbox": [ + 109, + 887, + 217, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "2556", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "for sketch understanding tasks, especially dense prediction problems like segmentation and correspondence learning.", + "bbox": [ + 89, + 90, + 480, + 121 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "These findings present a significant challenge. Retraining SD on sketch data risks catastrophic forgetting of its valuable pretrained knowledge [47], while developing sketch-specific architectures would fail to leverage the rich visual understanding already present in foundation models. Our key insight comes from analysing the complementary nature of different foundation models: while SD's features are spatially-aware but lack semantic accuracy, CLIP's [67] visual encoder exhibits strong semantic understanding even at the cost of spatial precision. This complementarity suggests that strategic combination [107] of these models could yield features that capture both the detailed structure and semantic meaning necessary for sketch understanding.", + "bbox": [ + 88, + 123, + 482, + 319 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Based on this insight, we propose a novel framework that leverages the complementary strengths of these foundation models. Our method systematically injects CLIP's visual features into multiple layers of SD's UNet during the denoising process, using lightweight $1D$ convolutions to influence feature extraction at various semantic levels. PCA analysis of the resulting features demonstrates that this combination successfully captures both the HF spatial details from SD [71] and the LF semantic components from CLIP [67], producing rich, semantically-aware representations suited for diverse sketch-based tasks.", + "bbox": [ + 89, + 321, + 482, + 487 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Moreover, we observe that different layers of SD's UNet capture features at varying levels of semantic-granularity [46], with different combinations better suited to specific downstream tasks [84]. Rather than manually selecting optimal features, we train a dynamic feature aggregator that automatically determines the most effective combination for any given task, eliminating the need for task-specific tuning. This adaptive approach ensures optimal performance across diverse sketch-based applications without requiring manual layer selection or task-specific optimisation.", + "bbox": [ + 89, + 489, + 482, + 641 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our extensive experiments demonstrate the effectiveness of this approach across a comprehensive suite of sketch-based tasks: (i) Sketch Recognition: $+1.06\\%$ improvement over state-of-the-art, (ii) Sketch-based Image Retrieval: $+3.35\\%$ improvement in retrieval accuracy, (iii) Sketch-based Segmentation: $+29.42\\%$ boost in mean IoU, (iv) Sketch-Photo Correspondence: $+21.22\\%$ increase in matching precision. These significant improvements across all major sketch-specific tasks suggest that our method has uncovered fundamental principles for extracting effective sketch features from foundation models.", + "bbox": [ + 89, + 642, + 482, + 808 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our contributions are: (i) A systematic analysis revealing the limitations of pretrained foundation models for sketch understanding, supported by both empirical results and frequency-domain analysis, (ii) A novel method for combining complementary foundation models that address these limitations without retraining, (iii) An adap", + "bbox": [ + 89, + 810, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tive feature aggregation approach that automatically optimises representations for different downstream tasks, and $(iv)$ Comprehensive experimental validation demonstrating substantial improvements across all major sketch-specific tasks, establishing a new paradigm for universal sketch feature learning.", + "bbox": [ + 511, + 90, + 903, + 181 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 511, + 186, + 663, + 203 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Diffusion Model for Representation Learning. Diffusion models [34] have now become the de-facto standard for high-resolution $2D$ [21, 71] and $3D$ [12, 81] image and video [11] generation. Apart from generation, it gave rise to various image editing frameworks like Prompt-to-Prompt [33], Imagic [40], Dreambooth [73], etc. The success of diffusion models across various generative tasks and their inherent zero-shot generalisability has led to their adoption as efficient feature extractors in different forms of discriminative [23] and dense prediction [109] tasks like image retrieval [46], classification [49], semantic [3] and panoptic [94] segmentation, depth estimation [110], object detection [93], image-to-image translation [47, 86], controllable-generation [59, 108] and medical imaging [18], to name a few. A few recent works have developed specialised diffusion models together with carefully crafted task-specific designs like test-time adaptation [65], masked-reconstruction [62], or compact latent generation [36] to improve their representation ability. Nonetheless, our pilot study (Sec. 4) reveals that SD, often struggles to extract representative features (Fig. 1 left) from binary and sparse sketches, hampering downstream task performance. In this work, we leverage a pretrained CLIP [67] visual encoder to enhance SD's feature representation abilities for sketch-based discriminative and dense prediction tasks.", + "bbox": [ + 511, + 205, + 903, + 583 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Sketch-based Vision Applications. \"Freehand sketch\" has now established itself as a worthy interactive alternative modality to text for downstream vision tasks [9, 16, 31, 43, 76, 87, 102]. With its success in sketch-based image retrieval (SBIR) [13, 20, 76, 79], sketches are now being used in a wide array of discriminative [16, 17, 20], generative [27, 43, 108] and dense prediction [7, 35] tasks like controllable $2D$ [47, 108], $3D$ [2, 88] image generation and editing [31, 43, 58, 59, 69, 108], video synthesis [92], garment designing [14, 51], segmentation [35], creative content generation [9], object detection [16], salient region prediction [7], augmented and virtual reality applications [56, 57], medical imaging [41], class incremental learning [5], or composed retrieval [44]. In this paper, we aim to tackle some of the sketch-based downstream tasks by reinforcing the diffusion feature extractor via a pretrained CLIP [67] visual encoder.", + "bbox": [ + 511, + 584, + 903, + 824 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Backbones for Sketch-based Vision Applications. Backbones for sketch-based vision applications can be broadly categorised as - (i) pretrained CNN or Vision Transformer (ViT) [6, 13, 68], (ii) Self-supervised pretraining [4, 50, 64], and (iii) pretrained foundation models [39, 67, 71]. Starting", + "bbox": [ + 511, + 825, + 903, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2557", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "with ImageNet pretrained CNNs [6, 13] or JFT pretrained ViTs [68, 77] with carefully-crafted design choices (e.g., cross [52] and spatial [83] attention, reinforcement learning [6], test-time training [75], meta-learning [74], knowledge distillation [77], etc.), the community moved towards self-supervised pretraining either via solving pretext tasks like sketch vectorisation [4], rasterisation [4], solving sketch-photo jigsaw [64] or by large-scale photo pretraining [50]. With the advent of large-scale foundation models, several works [43, 46, 76, 79] used them as a backbone for different sketch-based tasks like image generation [43, 108], shape synthesis [2], or retrieval [46, 79]. Foundation models have now become the gold standard for sketch-based tasks due to their strong zero-shot performance and extensive pretrained knowledge [45, 46, 79]. Among these, generative models like SD are spatially aware [46, 107], yet less accurate for dense prediction tasks. In contrast, vision-language models like CLIP [67] offer accurate features but are inherently sparse [107]. In this paper, we aim to get the best of both worlds by enhancing the SD feature extraction with CLIP.", + "bbox": [ + 89, + 90, + 480, + 393 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Reinforcing Feature Extractors. There are numerous ways of reinforcing an off-the-shelf pretrained feature extractor to adapt them for different downstream tasks. It can broadly be categorised as $(-i)$ Mixture of Experts (MoE) [22, 104], $(ii)$ Prompt learning [38, 112], $(iii)$ Hybrid models [91, 107], $(iv)$ Finetuning [76, 99], and $(v)$ Global-local feature merging [30, 105]. MoE combines multiple specialised sub-models (i.e., \"experts\"), each responsible for specific tasks or data patterns, to boost overall performance and adaptability [22]. In contrast, prompt learning deals with learning additional trainable components in the textual [111, 112] or pixel [1, 38] domain for task-specific adaptation. Hybrid models [89] combine two different kinds of models like CNN and ViT [91], or SD and DINO [107] to extract the best of both worlds. Another avenue deals with finetuning large-scale pretrained models with selective handcrafted data for specific tasks [16, 76]. Finally, a few other works [30, 90, 105] amalgamate global and local features from the same [90] or different [30, 105] models to improve downstream task performance. In this paper, we take the hybrid model approach and complement CNN-based SD features with transformer-based CLIP visual encoder features for different kinds of sketch-based vision tasks.", + "bbox": [ + 89, + 393, + 482, + 739 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Preliminaries", + "text_level": 1, + "bbox": [ + 89, + 750, + 230, + 763 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Background: CLIP", + "text_level": 1, + "bbox": [ + 89, + 770, + 277, + 786 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Contrastive language-image pretraining (CLIP) [67] architecture consists of a visual and a textual encoder, trained on $\\sim 400\\mathrm{M}$ image-text pairs, with the goal of learning a shared embedding manifold. Using N-pair contrastive loss, the model aligns matching image-text pairs by maximising their cosine similarity and minimises it for random unpaired ones [67]. The image encoder, typically a", + "bbox": [ + 89, + 794, + 482, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ViT [24], encodes an image $\\mathbf{i}$ as a feature vector $\\mathbf{V}(\\mathbf{i})\\in$ $\\mathbb{R}^d$ . Meanwhile, the textual encoder processes a word sequence $\\mathcal{W} = \\{w^0,w^1,\\dots ,w^k\\}$ through byte pair encoding [67] and a learnable word embedding layer (vocabulary size of 49,152), converting each word $w^{i}$ into a token embedding, thus creating an embedding sequence $\\mathcal{W}_e = \\{\\mathbf{w}_e^0,\\mathbf{w}_e^1,\\ldots ,\\mathbf{w}_e^k\\}$ [67]. $\\mathcal{W}_e$ is then passed through a transformer network $\\mathbf{T}(\\cdot)$ , producing the final text feature $\\mathbf{T}(\\mathcal{W}_e)\\in \\mathbb{R}^d$ [67]. Once trained, the CLIP model can be used in downstream tasks to enable zero-shot capability and cross-modal understanding by aligning visual and textual data in a shared embedding space [67].", + "bbox": [ + 511, + 90, + 903, + 272 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Background: Stable Diffusion (SD)", + "text_level": 1, + "bbox": [ + 511, + 276, + 818, + 291 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Overview. The SD model [71] generates images by gradually removing noise from an initial $2D$ isotropic Gaussian noise image through iterative refinement. It involves two reciprocal processes - \"forward\" and \"reverse\" diffusion [34, 71]. The forward process iteratively adds random Gaussian noise into a clean image $\\mathbf{i}_0 \\in \\mathbb{R}^{h \\times w \\times 3}$ for $t$ timesteps to generate a noisy image $\\mathbf{i}_t$ as $\\mathbf{i}_t = \\sqrt{\\bar{\\alpha}_t} \\mathbf{i}_0 + \\sqrt{1 - \\bar{\\alpha}_t} \\epsilon$ . Here, $\\epsilon \\sim \\mathcal{N}(0, \\mathbf{I})$ , and $\\{\\alpha_t\\}_1^T$ is a pre-defined noise schedule where $\\bar{\\alpha}_t = \\prod_{k=1}^t \\alpha_k$ with $t \\sim U(0, T)$ . The reverse process involves training a denoising UNet [72] $\\mathbf{U}_{\\theta}$ (with an $l_2$ objective), which estimates the input noise $\\epsilon \\approx \\mathbf{U}_{\\theta}(\\mathbf{i}_t, t)$ from the noisy image $\\mathbf{i}_t$ at each $t$ . After training, $\\mathbf{U}_{\\theta}$ is capable of reconstructing a clean image from a noisy input [71]. The inference process begins with a random $2D$ noise sample, $\\mathbf{i}_T \\sim \\mathcal{N}(0, \\mathbf{I})$ . The trained model $\\mathbf{U}_{\\theta}$ is then applied iteratively across $T$ timesteps, progressively reducing noise at each step to produce a cleaner intermediate image, $\\mathbf{i}_{t-1}$ , finally generating one of the cleanest samples $\\mathbf{i}_0$ from the target distribution.", + "bbox": [ + 511, + 292, + 903, + 580 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The unconditional denoising diffusion could further be made conditional by governing $\\mathbf{U}_{\\theta}$ with different auxiliary signals (e.g., textual captions) $\\mathbf{c}$ [71]. Accordingly, $\\mathbf{U}_{\\theta}(\\mathbf{i}_t,t,\\mathbf{c})$ can denoise $\\mathbf{i}_t$ while being influenced by $\\mathbf{c}$ , commonly via cross-attention [71].", + "bbox": [ + 511, + 580, + 903, + 656 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Architecture. Pixel-space diffusion model [34] is time-intensive as it operates on the full image resolution (i.e., $\\mathbf{i}_0\\in \\mathbb{R}^{h\\times w\\times 3}$ ). Contrarily, SD model [71] performs denoising in the latent-space, resulting in significantly faster and more stable processing. In the first stage, SD trains a variational autoencoder (VAE) [42] (an encoder $\\mathbf{E}(\\cdot)$ and a decoder $\\mathbf{D}(\\cdot)$ ). $\\mathbf{E}(\\cdot)$ transforms the input image into its latent representation $\\mathbf{z}_0 = \\mathbf{E}(\\mathbf{i}_0)\\in \\mathbb{R}^{\\frac{h}{8}\\times \\frac{w}{8}\\times d}$ . In the second stage, the UNet [72] $\\mathbf{U}_{\\theta}$ , directly denoises the latent images. $\\mathbf{U}_{\\theta}$ comprises 12 encoding, 1 bottleneck, and 12 skip-connected decoding blocks [71]. Inside these encoding and decoding blocks, there are 4 downsampling $(\\mathbf{U}_{\\mathbf{d}}^{1 - 4})$ and 4 upsampling $(\\mathbf{U}_{\\mathbf{u}}^{1 - 4})$ layers respectively. In SD, a frozen CLIP textual encoder [67] $\\mathbf{T}(\\cdot)$ converts the textual prompt $\\mathbf{c}$ into token-sequence, that influences the denoising process of $\\mathbf{U}_{\\theta}$ through cross-attention. $\\mathbf{U}_{\\theta}$ is trained over an", + "bbox": [ + 511, + 657, + 905, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "2558", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "$l_{2}$ objective as: $\\mathcal{L}_{\\mathrm{l}_2} = \\mathbb{E}_{\\mathbf{z}_t,t,\\mathbf{c},\\epsilon}(||\\epsilon -\\mathbf{U}_\\theta (\\mathbf{z}_t,t,\\mathbf{T}(\\mathbf{c}))||_2^2)$ . During inference, a noisy latent $\\mathbf{z}_t$ is sampled directly as: $\\mathbf{z}_t\\sim \\mathcal{N}(0,\\mathbf{I})$ . $\\mathbf{U}_{\\theta}$ removes noise from $\\mathbf{z}_t$ iteratively over $T$ timesteps (conditioned on $\\mathbf{c}$ ) to produce a denoised latent image $\\hat{\\mathbf{z}}_0\\in \\mathbb{R}^{\\frac{h}{8}\\times \\frac{w}{8}\\times d}$ . The final high-resolution image is generated as: $\\hat{\\mathbf{i}} = \\mathbf{D}(\\hat{\\mathbf{z}}_0)\\in \\mathbb{R}^{h\\times w\\times 3}$ .", + "bbox": [ + 89, + 89, + 480, + 181 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Pilot Study: Problems and Solution", + "text_level": 1, + "bbox": [ + 89, + 186, + 410, + 204 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "What's Wrong with Diffusion Features? Binary and sparse freehand sketches contain much lesser semantic contextual information than pixel-perfect photos [15, 43, 79]. While SD [71] as a backbone feature extractor has established itself as a strong competitor in image-based vision tasks [3, 18, 93, 94, 110], we dive deep into its internal representations to evaluate its effectiveness for the more challenging cross-modal sketch-based vision tasks. To this end, we ask the research question - whether a frozen SD model is capable of extracting features from a sparse sketch that are equally-representative as those derived from a pixel perfect image. Given a few sketch-photo pairs, we extract the corresponding features (detail in Sec. 5.1) from the second upsampling block of a frozen SD's UNet $\\mathbf{U}_{\\theta}$ with a fixed prompt \"a photo of [CLASS]\" and $t = 195$ . We perform principal component analysis (PCA) on the extracted high-dimensional features and plot the first three principal components as RGB images for visualisation. As seen in Fig. 1 (left), features extracted from sketches are significantly inferior to those extracted from photos. Unlike photos, abstract freehand sketches are sparse and lack detailed textures/colours [15, 43, 79], making it challenging for the SD model to capture fine-grained visual information from it. We hypothesise that SD struggles to extract equally representative features from sketches due to their minimal semantic cues and high level of abstraction [47]. Additionally, as the pretraining of SD was done predominantly on natural images [80], it is comparatively harder for SD to extract rich semantic information from a sparse sketch.", + "bbox": [ + 91, + 205, + 483, + 642 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Digging deeper, we follow classical computer vision literature [28, 103] in shifting the focus from spatial-domain, to analyse the behaviour of extracted features from a different perspective, i.e., exploring its nature in the frequency-domain. We achieve this by calculating the $2D$ Fourier magnitude spectrum of the extracted features, where the central region represents the LF components, while the HF components are arranged radially outward from the centre [28] (Fig. 1 left). Here we see that, although SD performs better on photos than sketches, it predominantly focuses on HF components in both cases. This is likely because SD's UNet is primarily trained as a denoiser thus operating with a global emphasis [107], where its skip connections additionally propagate HF components (e.g., edges), suppressing its LF counterparts (e.g., overall semantic structures) to ease the reconstruction of high-res image from noise [82]. We hypothesise that this very fact is directly conflicting with the", + "bbox": [ + 89, + 643, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "core motivation of our dense prediction task from a sparse data modality i.e., sketches, where capturing comprehensive semantic [107] context (LF) takes much higher precedence over preserving fine-details (HF), thus limiting SD's effectiveness [82, 107] for sketches, especially for dense prediction tasks (e.g., sketch-based segmentation).", + "bbox": [ + 511, + 90, + 903, + 181 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Reinforcing Diffusion Features. We identify two key problems in SD feature extraction - (i) SD features are not equally-representative for sketches and photos, and (ii) it mainly focuses on HF features, losing the crucial LF ones. In addressing these, our PCA analysis (Fig. 1 left) shows that, although SD is spatially-aware, it lacks accuracy. Inspired by existing works [107] on hybrid models, we propose complementing the SD feature extractor by combining it with another pretrained model that offers high accuracy, even if it is less spatially-aware. To this end, we influence the SD feature extraction pipeline with a pretrained CLIP [67] visual encoder as detailed in Sec. 5. PCA maps of CLIP (Fig. 1 left) features depict highly accurate background foreground separation, inherently providing the LF components that were missing in SD features. Finally, PCA maps of the proposed feature extractor (Fig. 1 left) demonstrate how the final representation is spatially-aware, accurate, integrates both HF and LF and works equally well for both sketches and photos. Apart from qualitative evidence, evaluating the performance of baseline $\\mathrm{SD + }\\mathrm{CLIP}$ hybrid model, and the proposed feature extractor across different sketch-based discriminative and dense prediction tasks, we observe an average performance gain of $39.49\\%$ (Fig. 1 right).", + "bbox": [ + 511, + 181, + 906, + 529 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "5. Proposed Methodology", + "text_level": 1, + "bbox": [ + 511, + 535, + 730, + 551 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Overview. Observations from our pilot study (Sec. 4) motivate us to reinforce the SD [71] feature extraction pipeline for better downstream task accuracy. However, reinforcing the SD feature extractor pipeline is non-trivial. Firstly, finetuning the entire model with task-specific dataset might interrupt SD's inherent pretrained knowledge [46]. Furthermore, end-to-end finetuning does not necessarily resolve the issues pertaining to spatial-awareness vs. accuracy or HF vs. LF trade-off (Sec. 4). To this end, we resort to a hybrid model [107] approach and use a complementary pretrained model (i.e., CLIP [67]) to strengthen the SD feature extractor pipeline. Secondly, obtaining features from CLIP alone is insufficient. They must be integrated effectively into the denoising process of the SD UNet to yield the performance gains observed in Fig. 1. Finally, different layers of the SD's UNet capture a wide array of features across different scales and granularity (i.e., coarse to fine) [84], each with its own pros and cons suited to various downstream tasks [46]. Thus, it is essential to effectively aggregate these multi-scale features to maximise their utility.", + "bbox": [ + 511, + 553, + 906, + 853 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Specifically, our salient design choices are $(-i)$ complementing SD [71] feature extraction with a pretrained CLIP [67] visual model, $(ii)$ an effective CLIP visual feature in", + "bbox": [ + 511, + 854, + 903, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "2559", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "jection module, and (iii) an aggregation module to unify features from different scales and granularities.", + "bbox": [ + 88, + 90, + 480, + 121 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. Model Architecture", + "text_level": 1, + "bbox": [ + 89, + 126, + 277, + 138 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Stable Diffusion Feature Extraction. Internal activation maps of a pretrained SD UNet $\\mathbf{U}_{\\theta}$ contain information-rich representations that have found use in tasks like image segmentation [3], object detection [93], classification [49], etc. Here, we aim to use these internal representations as an efficient backbone feature for various sketch-based vision tasks. Given an image-text pair $(\\mathbf{i},\\mathbf{c})$ , the input image is first passed through the pretrained VAE [42] encoder $\\mathbf{E}(\\cdot)$ , converting it into an initial latent representation $\\mathbf{z}_0 = \\mathbf{E}(\\mathbf{i}) \\in \\mathbb{R}^{\\frac{h}{8} \\times \\frac{w}{8} \\times d}$ . Then, given a timestep value $t$ , SD performs forward diffusion on $\\mathbf{z}_0$ to convert it to the $t^{th}$ step noisy latent $\\mathbf{z}_t$ . Finally, the textual prompt embedding (Sec. 3.2) $\\mathbf{T}(\\mathbf{c})$ , noisy latent $\\mathbf{z}_t$ , and timestep $t$ are passed to the pretrained denoising UNet [71] $\\mathbf{U}_{\\theta}(\\cdot)$ to extract the internal features from its skip-connected upsampling layers $\\{\\mathbf{U}_{\\mathbf{u}}^n \\to f_{\\mathbf{u}}^n\\}_{n=1}^4$ . In SD v2.1 [71], for an input image $\\mathbf{i} \\in \\mathbb{R}^{h \\times w \\times 3}$ , upsampling layers $\\{\\mathbf{U}_{\\mathbf{u}}^n\\}_{n=1}^4$ generate features as: $f_{\\mathbf{u}}^1 \\in \\mathbb{R}^{\\frac{h}{32} \\times \\frac{w}{32} \\times 1280}$ , $f_{\\mathbf{u}}^2 \\in \\mathbb{R}^{\\frac{h}{16} \\times \\frac{w}{16} \\times 1280}$ , $f_{\\mathbf{u}}^3 \\in \\mathbb{R}^{\\frac{h}{8} \\times \\frac{w}{8} \\times 640}$ , and $f_{\\mathbf{u}}^4 \\in \\mathbb{R}^{\\frac{h}{8} \\times \\frac{w}{8} \\times 320}$ respectively. Due to the unavailability of paired textual prompts in most sketch-photo datasets [26, 29, 78], we use null (i.e., “”) prompt in all our experiments.", + "bbox": [ + 91, + 142, + 483, + 479 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/187a38e80495717266c0b7d1cfd55d677183700bb250696f0c95e6719f483051.jpg", + "image_caption": [ + "Figure 2. Given the frozen SD [71] and CLIP [67] models, the proposed method learns the aggregation network, $1D$ convolutional layers, and branch-weights with sketch-photo pairs, via different losses for different downstream tasks (details in Sec. 6)." + ], + "image_footnote": [], + "bbox": [ + 91, + 488, + 483, + 614 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "CLIP Visual Feature Extraction. Akin to SD, pretrained CLIP [67] encoders have also been used in a plethora of vision tasks like segmentation [60], retrieval [76], object detection [16], video captioning [98], to name a few. In this work, we aim to influence the intermediate feature maps of SD [71] with that from the CLIP visual encoder. Specifically, the input image $\\mathbf{i}$ is passed through the pretrained CLIP visual encoder $\\mathbf{V}(\\cdot)$ to extract the patched feature $f_{\\mathbf{v}}$ from the penultimate layer as: $f_{\\mathbf{v}} = \\mathbf{V}(\\mathbf{i}) \\in \\mathbb{R}_{p}^{\\frac{h}{p} \\times \\frac{w}{p} \\times d}$ , where $p$ is the patch-size of the $\\mathbf{V}(\\cdot)$ backbone encoder.", + "bbox": [ + 89, + 685, + 482, + 839 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "CLIP Feature Injection. Based on the qualitative and quantitative evidence presented in the pilot study (Sec. 4), we hypothesise that CLIP [67] visual and textual embeddings are intrinsically aligned [67, 79]. Prior works [79]", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "have underscored the importance of textual prompts in the quality of extracted features [79]. However, most simple prompts (e.g., \"a photo/sketch of [CLASS]\") are inherently coarse-grained, offering only a high-level description of an image. Such general prompts often lack the semantic detail necessary to fully characterise an image [15, 47], limiting the model's ability to generate highly accurate and contextually-rich feature representations [46].", + "bbox": [ + 511, + 90, + 903, + 210 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To address this, we propose injecting CLIP's visual features into multiple layers of the SD's UNet, thereby influencing the denoising process at various stages (Fig. 2). This approach not only enhances the accuracy of SD features but also enables the model to draw on semantic information from both CLIP's visual and textual embeddings.", + "bbox": [ + 511, + 212, + 903, + 303 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Specifically, after extracting CLIP's visual embeddings $f_{\\mathbf{v}} \\in \\mathbb{R}^{\\frac{h}{p} \\times \\frac{w}{p} \\times d}$ , we pass them through simple learnable $1D$ convolutional layers $\\mathcal{C}(\\cdot)$ to modify them just enough to match the feature dimension of different blocks of the UNet. These transformed feature maps are then added to the intermediate UNet features $\\{f_{\\mathbf{u}}^{n}\\}_{n=1}^{4}$ at every timestep of the denoising process as: $\\hat{f}_{\\mathbf{u}}^{n} = f_{\\mathbf{u}}^{n} + \\mathcal{C}(f_{\\mathbf{v}})$ . This injection occurs in all timesteps and layers of the UNet, enabling the model to influence the denoising process with the semantic information encoded within CLIP's visual embeddings.", + "bbox": [ + 511, + 304, + 903, + 459 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "SD Feature Aggregation. Different layers of $\\mathbf{U}_{\\theta}$ capture features of varying semantic-granularity [84], where features from specific layers are ideally suited for certain downstream tasks [46]. Instead of manually selecting each layer for a specific task, we introduce a dynamic approach, allowing the model to automatically extract the most optimal features, thus eliminating the need for manual tuning.", + "bbox": [ + 511, + 460, + 905, + 566 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To achieve this, we train a feature aggregation network comprising three ResNet [32] blocks $\\mathbf{A}(\\cdot)$ , which transform the CLIP-enhanced SD features $\\{\\hat{f}_{\\mathbf{u}}^{n}\\}_{n=1}^{3}$ extracted from the first three UNet upsampling layers as: $\\bar{f}_{\\mathbf{u}}^{n} = \\mathbf{A}(\\hat{f}_{\\mathbf{u}}^{n}) \\in \\mathbb{R}^{60 \\times 60 \\times d}$ . The final feature map is then obtained by computing a weighted summation of the aggregated features using a learnable weight parameter $\\{\\alpha_{n}\\}_{n=1}^{3}$ for each feature-branch. This aggregation strategy enables the model to dynamically adjust its focus on different layers, adaptively modulating the contribution of each feature-branch based on the semantic-granularity of the task.", + "bbox": [ + 511, + 566, + 906, + 734 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "6. Experiments", + "text_level": 1, + "bbox": [ + 511, + 744, + 643, + 762 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The proposed feature extractor enriched with the large-scale pretrained knowledge of SD [71] and CLIP [67], enables multiple sketch-based discriminative and dense prediction tasks. Specifically, we showcase its efficacy in tasks like - sketch-based image retrieval [76], sketch recognition [100], sketch-photo correspondence learning [55], and sketch-based image segmentation [35] on different datasets [26, 29, 55, 78] and scenarios. In all cases, we only learn $\\mathcal{C}(\\cdot)$ , aggregation network $\\mathbf{A}(\\cdot)$ , and branch-weights $\\alpha_{n}$", + "bbox": [ + 511, + 763, + 905, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "2560", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "keeping pretrained SD and CLIP models frozen.", + "bbox": [ + 89, + 90, + 410, + 104 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6.1. Sketch-based Image Retrieval (SBIR)", + "text_level": 1, + "bbox": [ + 89, + 109, + 413, + 125 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Problem Statement. Given a query sketch $\\mathbf{s} \\in \\mathbb{R}^{h \\times w \\times 3}$ of any class $j$ , category-level SBIR aims to retrieve a photo $\\mathbf{p}_i^j$ from the same class, out of a gallery $\\mathcal{G} = \\{\\mathbf{p}_i^j\\}_{i=1}^{N_j}|_{j=1}^{N_c}$ containing images from a total $N_c$ classes with $N_j$ images per class [102]. Whereas, cross-category fine-grained (FG) SBIR setup intends to learn a unified model capable of instance-level matching from a gallery containing photos from multiple $(N_c)$ classes. Conventional SBIR is evaluated on categories $\\mathcal{C}^S = \\{c_1^S, \\dots, c_N^S\\}$ that are seen during training. Contrarily, the zero-shot (ZS) paradigm [76] focuses on evaluating on mutually exclusive (i.e., $\\mathcal{C}^S \\cap \\mathcal{C}^U = \\emptyset$ ) unseen categories $\\mathcal{C}^U = \\{c_1^U, \\dots, c_M^U\\}$ .", + "bbox": [ + 88, + 126, + 482, + 310 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Training & Evaluation. Following standard literature [76], we use triplet loss to train our SBIR models. Specifically, given the global max-pooled features from the proposed feature extractor, triplet loss aims to minimise the distance $\\delta$ between an anchor sketch (s) feature $f_{\\mathbf{s}} \\in \\mathbb{R}^{d}$ and a positive photo (p) feature $f_{\\mathbf{p}} \\in \\mathbb{R}^{d}$ from the same category as s, while maximising the same from a negative photo (n) feature $f_{\\mathbf{n}} \\in \\mathbb{R}^{d}$ of a different category. Whereas, in case of cross-category FG-SBIR framework we use hard triplets, where the negative sample is a distinct instance $(\\mathbf{p}_k^j; k \\neq i)$ belonging to the same class as the anchor sketch $(\\mathbf{s}_i^j)$ and its paired photo $(\\mathbf{p}_i^j)$ [46]. With margin $\\mu$ , triplet loss becomes:", + "bbox": [ + 89, + 311, + 483, + 496 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {t r i p l e t}} \\left(f _ {\\mathbf {s}}, f _ {\\mathbf {p}}, f _ {\\mathbf {n}}\\right) = \\max \\left\\{0, \\mu + \\delta \\left(f _ {\\mathbf {s}}, f _ {\\mathbf {p}}\\right) - \\delta \\left(f _ {\\mathbf {s}}, f _ {\\mathbf {n}}\\right) \\right\\}. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 98, + 507, + 482, + 525 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For category-level ZS-SBIR, following [46, 76], we consider top $k$ retrieved images to calculate mean average precision $(\\mathrm{mAP}@\\mathrm{k})$ and precision $(\\mathrm{P}@\\mathrm{k})$ . Whereas, for cross-category ZS-FG-SBIR, we measure Acc.@k, which indicates the percentage of sketches having true-matched photos among the top-k retrieved images.", + "bbox": [ + 89, + 536, + 482, + 626 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Dataset & Competitors. Here we use three datasets - (i) Sketchy [78] and Sketchy-extended [53, 78] datasets are used to train and test our cross-category ZS-FG-SBIR and category-level ZS-SBIR models respectively. While Sketchy [78] holds 12,500 photos from 125 classes, each having at least 5 fine-grained paired sketches, its extended version [53] carries additional 60,652 photos from ImageNet [19]. We use photos/sketches from 104 classes for training and 21 for testing [102]. (ii) Tu-Berlin [26] contains 204,489 photos from 250 classes, each having $\\sim 80$ sketches. We use photos/sketches from 220 classes for training and 30 for testing. (iii) Quick, Draw! [29] holds over $50M$ sketches from 345 categories. We use photos/sketches from 80 classes for training and 30 for testing.", + "bbox": [ + 89, + 628, + 482, + 838 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For category-level ZS-SBIR, we compare our method with SoTA methods like SAKE [54], CAAE [102], CCGAN [25], GRL [20], SD-PL [46], etc. Whereas, for cross-category ZS-FG-SBIR, we compare with Gen-VAE [63],", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/4d9bc86e1ecb1cb743cc6fe167fff183f4be653e615ea8251e4b78bb4f771274.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsSketchy [78]TU-Berlin [26]Quick, Draw! [29]
mAP@200P@200mAP@allP@100mAP@allP@200
B-CLIP0.2500.2610.2280.2570.0800.141
B-DINO0.4930.4810.4500.4920.1670.249
B-DINOV20.5270.5330.4810.5320.1700.268
B-SD0.5580.5710.5100.5610.1790.287
B-SD+CLIP0.5880.5920.5370.5890.1790.311
B-Finetuning0.1200.1720.0110.0100.0030.006
SAKE [54]0.4970.5980.4750.599--
IIAE [37]0.3730.4850.4120.503--
CAAE [102]0.1560.2600.0050.003--
CCGAN [25]--0.2970.426--
CVAE [102]0.2250.3330.0050.0010.0030.003
GRL [20]0.3690.3700.1100.1210.0750.068
LVM [76]0.7230.7250.6510.7320.2020.388
SD-PL [46]0.7460.7470.6800.7440.2310.397
Proposed0.7610.7630.6950.7530.2420.399
", + "bbox": [ + 519, + 88, + 903, + 275 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "LVM [76], and $SD-PL$ [46]. Apart from the SoTA methods, we also compare all tasks against a few self-designed Baselines. Among them B-CLIP, B-DINO, B-DINOv2, and B-SD use off-the-shelf pretrained CLIP [67] ViT-L/14, DINO [8] ViT-S/8, DINOv2 [61] ViT-S/14, or SD [71] v2.1 encoders as feature extractors, without task-specific training. We also form another baseline $B-SD+CLIP$ as a straightforward combination of SD and CLIP features without the proposed feature injection and aggregation modules. Finally, B-Finetuning finetunes the entire pretrained SD UNet and CLIP visual encoder on task-specific data.", + "bbox": [ + 511, + 296, + 906, + 464 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/6fb0799ac9c8d3312c9e9d6376c8d225abb09e14ffc2bb551e1b9414d8103e12.jpg", + "table_caption": [ + "Table 1. Results for category-level ZS-SBIR." + ], + "table_footnote": [], + "table_body": "
MethodsAcc.@1Acc.@5MethodsAcc.@1Acc.@5
B-CLIP11.8421.66B-Finetuning5.679.17
B-DINO22.4946.97Gen-VAE [63]22.6049.00
B-DINOv221.1944.31LVM [76]28.6862.34
B-SD23.9849.42SD-PL [46]31.9465.81
B-SD+CLIP24.1652.01Proposed33.0167.92
", + "bbox": [ + 519, + 472, + 901, + 551 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Results on Sketchy [78] for cross-category ZS-FG-SBIR.", + "bbox": [ + 514, + 553, + 903, + 566 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6.2. Sketch Recognition", + "text_level": 1, + "bbox": [ + 511, + 580, + 697, + 595 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Problem Statement. Sketch recognition deals with identifying and classifying freehand abstract sketches into predefined categories [68, 96]. Given a dataset $\\mathcal{D} = \\{(\\mathbf{s}_i,\\mathbf{y}_i)\\}_{i = 1}^N$ containing $N$ sketches and their corresponding labels from $C$ classes, the goal is to learn a model that can correctly recognise the class of an input query sketch.", + "bbox": [ + 511, + 598, + 905, + 688 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Training & Evaluation. Given the global max-pooled query sketch feature $f_{\\mathbf{s}} \\in \\mathbb{R}^{d}$ , we use standard cross-entropy (CE) loss to train our sketch recognition model as,", + "bbox": [ + 511, + 691, + 905, + 736 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {C E}} \\left(\\mathbf {s} _ {i}, \\mathbf {y} _ {i}\\right) = - \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\log p \\left(\\mathbf {y} _ {i} \\mid \\mathbf {s} _ {i}\\right). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 741, + 906, + 779 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Following [66, 96, 101] we evaluate using Acc.@k.", + "bbox": [ + 511, + 782, + 852, + 797 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Dataset & Competitors. For this task, we again use the popular TU-Berlin [26] and Quick, Draw! [29] datasets. For Quick, Draw!, we adopt the data split from [29], with each class containing 70,000, 2,500, and 2500 training, validation, and test samples. Apart from the self-designed baselines, we also compare against SoTA methods like SketchMate [96], SketchGNN [101], and SketchXAI [66].", + "bbox": [ + 511, + 800, + 905, + 905 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "2561", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/2d0820a284069092f897e296c7d756732cc588a68cd867e895f7cb5539bf567a.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsTU-BerlinQuick, Draw!MethodsTU-BerlinQuick, Draw!
B-CLIP30.0930.87B-Finetuning10.2912.37
B-DINO51.7953.28SketchMate [96]77.9679.44
B-DINOv258.0159.12SketchGNN [101]76.4377.31
B-SD61.3763.57SketchXAI [66]-86.10
B-SD+CLIP65.4767.78Proposed84.9687.02
", + "bbox": [ + 94, + 88, + 480, + 152 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6.3. Sketch-photo Correspondence", + "text_level": 1, + "bbox": [ + 89, + 169, + 359, + 185 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Problem Statement. Sketch-photo correspondence learning involves identifying a mapping between a freehand sketch of an object and its corresponding photo, such that keypoints in the sketch can be semantically aligned with corresponding keypoints in the photo [55]. Given a sketch-photo pair $\\{\\mathbf{s},\\mathbf{p}\\} \\in \\mathbb{R}^{h\\times w\\times 3}$ and a set of $N$ query keypoints (on sketch) $K_{\\mathbf{s}} = \\{x_i^s,y_i^s\\}_{i = 1}^N$ , the aim is to predict a new set of corresponding keypoints $\\hat{K}_{\\mathbf{p}} = \\{\\hat{x}_i^{\\mathbf{p}},\\hat{y}_i^{\\mathbf{p}}\\}_{i = 1}^N$ on the photo, such that $\\hat{K}_{\\mathbf{p}}$ aligns semantically with $K_{\\mathbf{s}}$ .", + "bbox": [ + 89, + 185, + 483, + 325 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Training & Evaluation. Following [106], we use contrastive $(\\mathcal{L}_{\\mathrm{CL}})$ [67] and end-point error losses to train our sketch-photo correspondence model. Given the sketch and photo features $\\{f_{\\mathrm{s}}, f_{\\mathrm{p}}\\} \\in \\mathbb{R}^{60 \\times 60 \\times d}$ from the proposed feature extractor, $\\mathcal{L}_{\\mathrm{CL}}$ aims to maximise similarity between local feature patches, while minimising the same between distance ones, for each sketch-photo pairs. Whereas, for endpoint error loss, we first estimate optical flows [106] from the extracted features, followed by calculating $l_{2}$ norm between the estimated $(\\mathbf{f}_e)$ and ground truth $(\\mathbf{f}_g)$ flow as:", + "bbox": [ + 89, + 325, + 483, + 476 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {E P E}} \\left(\\mathbf {f} _ {e} ^ {i}, \\mathbf {f} _ {g} ^ {i}\\right) = \\frac {1}{M} \\sum_ {i = 1} ^ {M} \\left\\| \\mathbf {f} _ {e} ^ {i} - \\mathbf {f} _ {g} ^ {i} \\right\\| _ {2} ^ {2}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 169, + 479, + 480, + 516 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where, $M$ denotes the patch number (i.e., 60). Finally, the overall loss becomes $\\mathcal{L}_{\\mathrm{corr}} = \\mathcal{L}_{\\mathrm{CL}} + \\mathcal{L}_{\\mathrm{EPE}}$ . Following [55] we use PCK@k as a test metric which measures the percentage of keypoints predicted correctly within a distance threshold (i.e., $k\\%$ of the image size).", + "bbox": [ + 89, + 518, + 483, + 594 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Dataset & Competitors. For this task, we use the PSC6K [55] dataset that contains more than $150K$ keypoint annotations for 1250 photos, each having 5 paired sketches (taken from the original Sketchy [78] test split) from 125 classes. We use a 60:40 split for training and testing, where we compare the proposed approach against the self-designed baselines and SoTA methods like WeakAlign[70], WarpC [85], Self-Supervised [55], etc.", + "bbox": [ + 89, + 594, + 483, + 715 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/49e584731147763063a56d8a9afa265bce0bc4455d4f0aa155745da046bdebb4.jpg", + "table_caption": [ + "Table 3. Acc.@1 results for sketch recognition." + ], + "table_footnote": [], + "table_body": "
MethodsPCK@5PCK@10MethodsPCK@5PCK@10
B-CLIP22.2930.41B-Finetuning11.5817.69
B-DINO34.2148.59WeakAlign [70]43.5578.60
B-DINOv239.9156.46WarpC [85]56.7879.70
B-SD41.4258.71Self-Sup. [55]58.0084.93
B-SD+CLIP44.9062.02Proposed70.3189.86
", + "bbox": [ + 94, + 720, + 480, + 792 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. Results on PSC6K [55] for sketch-photo correspondence. 6.4. Sketch-based Image Segmentation", + "bbox": [ + 89, + 795, + 480, + 824 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Problem Statement. Given a query sketch $\\mathbf{s} \\in \\mathbb{R}^{h \\times w \\times 3}$ of any category, sketch-based image segmentation [35] aims to predict a binary segmentation mask, $\\mathbf{m} \\in \\mathbb{R}^{h \\times w \\times 1}$ , depicting the pixel-locations, wherever the queried-concept appears in any candidate image $\\mathbf{p} \\in \\mathbb{R}^{h \\times w \\times 3}$ of that category,", + "bbox": [ + 89, + 824, + 483, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "from an entire test gallery. The predicted mask marks pixels belonging to that sketched-concept as 1 and the rest as 0.", + "bbox": [ + 511, + 90, + 903, + 121 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Training & Evaluation. We train our segmentation model using binary cross-entropy (BCE) loss, between predicted and ground truth (GT) masks. Given patched photo feature $f_{\\mathbf{p}} \\in \\mathbb{R}^{60 \\times 60 \\times d}$ and global max-pooled sketch feature photo features $f_{\\mathbf{s}} \\in \\mathbb{R}^d$ from the proposed feature extractor, we calculate the cosine similarity between them to generate a patch-wise correlation map $\\mathbf{c} = f_{\\mathbf{s}} \\odot f_{\\mathbf{p}} \\in \\mathbb{R}^{60 \\times 60}$ . We upsample $\\mathbf{c}$ to the original spatial resolution (i.e., $\\mathbb{R}^{h \\times w}$ ), followed by differentiable Sigmoid thresholding [10] to get the predicted map. BCE loss between the GT and predicted maps can be calculated as,", + "bbox": [ + 511, + 121, + 905, + 287 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {B C E}} \\left(y _ {i}, \\hat {y} _ {i}\\right) = - \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\left[ y _ {i} \\log \\left(\\hat {y} _ {i}\\right) + \\left(1 - y _ {i}\\right) \\log \\left(1 - \\hat {y} _ {i}\\right) \\right] \\quad , \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 292, + 905, + 325 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where, $N$ is the number of total pixels, $y_{i}$ and $\\hat{y}_i$ are the GT and predicted mask pixels. During testing, we calculate similar patch-wise correlation map from sketch/photo features and upsample it to the original spatial image dimension, followed by thresholding it with an empirically chosen fixed value (0.47) and standard CRF [48]-based postprocessing. We use mean Intersection over Union (mIoU) and pixel accuracy (pAcc.) as test metrics, which measure the overlap and percentage of correctly predicted pixels between predicted and GT segmentation masks respectively.", + "bbox": [ + 511, + 324, + 905, + 474 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Dataset & Competitors. Due to the lack of existing paired sketch-photo-mask datasets, we manually annotate photos of 10 chosen classes from the Sketchy [78] dataset with binary mask annotations, resulting in $5K$ paired sketch-photo-mask tuples. We split it into 80:20 for training and testing. Besides baselines and SoTA Sketch-a-Segmenter [35], we also compare with a modified version of ZS-Seg [95], which adapts a simple text-based zero-shot segmentation framework for sketches by replacing its CLIP [67] textual encoder with the visual one to encode input sketches.", + "bbox": [ + 511, + 474, + 905, + 625 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d89ac8a0f6954ad106c6e9bc31ac9ca939af4347dbb907db2b33e7a4e1eff411.jpg", + "image_caption": [ + "Figure 3. Sketch-photo correspondence (left→right : source → target) results on PSC6K. Green circles and squares depict source and GT points respectively, while red squares denote predicted points." + ], + "image_footnote": [], + "bbox": [ + 514, + 628, + 903, + 702 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6.5. Results and Analyses", + "text_level": 1, + "bbox": [ + 511, + 748, + 712, + 763 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Following are the observations from quantitative and qualitative results presented in Tab. 1-5 and Fig. 3-4. (i) Naive adaptation of pretrained backbones (e.g., B-CLIP, B-DINO, B-SD, etc.) fail to beat our method in both ZS-SBIR and ZS-FG-SBIR settings (Tab. 1-2), where we achieve an average 76.15 (54.32)% higher mAP@200 (Acc.@1) over all competitors on Sketchy. We posit that our strategic combination of SD and CLIP brings the \"best of both worlds\", complementing each other's flaws. (ii) While", + "bbox": [ + 511, + 763, + 905, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "2562", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/1559488a1c1b67360bd5f6be3120a48a2fe4010279189e9772821bbdf7a9208a.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsmIoUpAcc.MethodsmIoUpAcc.
B-CLIP20.6326.82B-Finetuning20.3621.84
B-DINO32.8142.06DeepLabv3+Sketch [35]40.6355.23
B-DINOv234.1745.80ZS-Seg [95]44.7360.97
B-SD39.0249.03Sketch-a-Segmenter [35]46.4560.28
B-SD+CLIP41.8751.91Proposed60.1274.91
", + "bbox": [ + 96, + 89, + 478, + 166 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "SoTA sketch-recognition frameworks [96, 101] perform reasonably across different datasets (Tab. 3), our method exceeds them with an average Acc.@1 gain of $48.09\\%$ (Quick, Draw!). SketchXAI [66] despite being our strongest contender, depicts sub-optimal Acc.@1. This gain is likely due to the large-scale pretraining of SD and CLIP, which enables superior open-world object understanding [107]. (iii) For sketch-photo correspondence learning, our method shows an impressive $21.22\\%$ PCK@5 gain over SoTA [55], without the time-consuming two-stage training of [55], or complicated warp-estimation of [85], thanks to SD's innate object-localisation capability. (iv) Sketch-based segmentation being the most challenging task amongst all, baseline methods perform quite poorly (Tab. 5). However, the proposed method surpasses SoTA [35] with a mIoU (pAcc.) margin of $29.42(24.27)\\%$ . (v) B-Finetuning fails drastically in almost all tasks (Tab. 1-5). We hypothesise that finetuning entire SD and CLIP models with limited sketch data distorts their large-scale pretrained knowledge [46, 76], thus sacrificing the generalisation potential. (vi) Finally, B-SD+CLIP depicts a $\\sim 35.78\\%$ lesser scores than ours (averaged over all tasks), which again underpins the need for the proposed feature injection and aggregation strategy, rather than a simple combination of pretrained foundation models.", + "bbox": [ + 89, + 184, + 483, + 547 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ed87e7b74dc053d81e4ac063d273158fb0c1303a12c6a289a234fbc76ecb31bb.jpg", + "image_caption": [ + "Figure 4. Qualitative results for sketch-based image segmentation. Given a query sketch, our method generates separate segmentation masks for all images of that category. (Zoom-in for the best view.)" + ], + "image_footnote": [], + "bbox": [ + 102, + 551, + 480, + 636 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Ablation on Design", + "text_level": 1, + "bbox": [ + 89, + 685, + 276, + 702 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "[i] Effect of aggregation network. To judge its efficacy, instead of aggregation networks $\\mathbf{A}(\\cdot)$ , we manually interpolate features from $\\{\\mathbf{U}_{\\mathbf{u}}^{n}\\}_{n=1}^{3}$ into $\\mathbb{R}^{60 \\times 60 \\times d}$ dimension and add them (via $\\alpha_{n}$ ) to form the final feature map. A rapid PCK@5 (mIoU) drop of 28.65 (26.51)% in case of w/o Aggregation Network (Tab. 6) indicates its significance in extracting semantically-meaningful features.", + "bbox": [ + 89, + 704, + 482, + 809 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "[i] Are learnable branch weights necessary? Learnable weights dynamically ensure the optimum importance of each feature-branch for a specific task. Although less pronounced in PCK@5, mAP@200 drops significantly (4.82%) in case of w/o Learnable Weights (Tab. 6), depicting its influence on the final feature map quality.", + "bbox": [ + 89, + 810, + 483, + 901 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/13eac0e3773a02a8ca12234bd8ca3939a9151bc0b9c5de9ace2a3388abe5d4dc.jpg", + "table_caption": [ + "Table 5. Results for sketch-based image segmentation." + ], + "table_footnote": [], + "table_body": "
MethodsCorrespondenceSegmentationZS-SBIRRecognition
PCK@5mIoUmAP@200Acc.@1
SD [71]v1.468.9157.630.72878.82
v1.569.2458.910.73579.59
v2.170.3160.120.76187.02
CLIP [67]ViT-B/1667.7556.380.74377.37
ViT-B/3269.4158.470.75878.84
ViT-L/1470.3160.120.76187.02
w/o Aggregation Network54.6547.520.58766.71
w/o Learnable Weights68.7558.130.72683.79
w/o 1D Convolutions55.9747.240.60271.75
Ours-full70.3160.120.76187.02
", + "bbox": [ + 519, + 89, + 901, + 207 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 6. Ablation on design.", + "bbox": [ + 622, + 208, + 795, + 220 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "[iii] Contribution of $1D$ Convolutions. To assess the effectiveness of $1D$ convolution layers, we use simple linear interpolation to match CLIP and SD UNet feature dimensions, before injection. Sharp PCK@5 drop of $25.62\\%$ for w/o $1D$ Convolutions (Tab. 6) reveals its impact on influencing the diffusion denoising process. We posit that the learnable $1D$ convolutions not only match the feature dimension but also enrich the CLIP features before injection. [iv] Model variants. The performance of our proposed feature extractor depends crucially on the choice of backbone model variant. Accordingly, we explore with multiple pretrained SD and CLIP backbones. Tab. 6 shows SD v2.1 and CLIP ViT-L/14 to achieve the highest PCK@5 and Acc.@1 for sketch-photo correspondence and recognition respectively. Additionally, we also experiment by replacing the CLIP visual encoder of our method with a DINOv2 [61] ViT-S/14 encoder for the task of ZS-SBIR, to observe a $32.80\\%$ drop in mAP@200. This is probably due to CLIP's [67] multimodal pertaining (diminishing sketch-photo domain gap), over DINOv2's [61] self-supervised one.", + "bbox": [ + 511, + 223, + 906, + 526 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "[v] Which timestep value is the most effective? Different diffusion timesteps yield features with drastically different quality [84, 107]. Calculating PCK@5 and mAP@200", + "bbox": [ + 511, + 527, + 905, + 571 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/19c7efbf2105c1178a587d2ccc1d25a9ff1969dac54da7cc2b39b28f322ff58c.jpg", + "image_caption": [ + "Figure 5. Choice of timestep $(t)$ . method to be robust to the choice of $t$ , where PCK@5 (mAP@200) scores from a wide range of $t$ values consistently surpass the baselines." + ], + "image_footnote": [], + "bbox": [ + 513, + 571, + 732, + 678 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "for the tasks of correspondence learning and ZS-SBIR respectively, over varying $t$ , we find that $t = 195$ produces the best results for both tasks. Fig. 5 shows our method to be robust to", + "bbox": [ + 741, + 571, + 906, + 691 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "8. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 729, + 633, + 744 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we identified SD's limitations in understanding abstract, sparse sketches. Integrating CLIP's semantic features with SD's spatially-aware features, we developed a framework that dynamically combines these strengths. This significantly improves performance across sketch recognition, SBIR, and sketch-based segmentation. Our method highlights the complementary nature of different foundation models and introduces an adaptive feature aggregation network, establishing a new standard for universal sketch feature representation for future research.", + "bbox": [ + 509, + 750, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "2563", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Hyojin Bahng, Ali Jahanian, Swami Sankaranarayanan, and Phillip Isola. Exploring visual prompts for adapting large-scale models. arXiv preprint arXiv:2203.17274, 2022.3", + "[2] Hmrishav Bandyopadhyay, Subhadeep Koley, Ayan Das, Ayan Kumar Bhunia, Aneeshan Sain, Pinaki Nath Chowdhury, Tao Xiang, and Yi-Zhe Song. Doodle Your 3D: From Abstract Freehand Sketches to Precise 3D Shapes. In CVPR, 2024. 2, 3", + "[3] Dmitry Baranchuk, Ivan Rubachev, Andrey Voynov, Valentin Khrulkov, and Artem Babenko. Label-Efficient Semantic Segmentation with Diffusion Models. In ICLR, 2021. 2, 4, 5", + "[4] Ayan Kumar Bhunia, Pinaki Nath Chowdhury, Yongxin Yang, Timothy Hospedales, Tao Xiang, and Yi-Zhe Song. Vectorization and Rasterization: Self-Supervised Learning for Sketch and Handwriting. In CVPR, 2021. 2, 3", + "[5] Ayan Kumar Bhunia, Viswanatha Reddy Gajjala, Subhadeep Koley, Rohit Kundu, Aneeshan Sain, Tao Xiang, and Yi-Zhe Song. Doodle It Yourself: Class Incremental Learning by Drawing a Few Sketches. In CVPR, 2022. 2", + "[6] Ayan Kumar Bhunia, Subhadeep Koley, Abdullah Faiz Ur Rahman Khilji, Aneeshan Sain, Pinaki Nath Chowdhury, Tao Xiang, and Yi-Zhe Song. Sketching Without Worrying: Noise-Tolerant Sketch-Based Image Retrieval. In CVPR, 2022. 2, 3", + "[7] Ayan Kumar Bhunia, Subhadeep Koley, Amandeep Kumar, Aneeshan Sain, Pinaki Nath Chowdhury, Tao Xiang, and Yi-Zhe Song. Sketch2Saliency: Learning to Detect Salient Objects from Human Drawings. In CVPR, 2023. 2", + "[8] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging Properties in Self-Supervised Vision Transformers. In ICCV, 2021. 6", + "[9] Dar-Yen Chen, Subhadeep Koley, Aneeshan Sain, Pinaki Nath Chowdhury, Tao Xiang, Ayan Kumar Bhunia, and Yi-Zhe Song. DemoCaricature: Democratising Caricature Generation with a Rough Sketch. In CVPR, 2024. 2", + "[10] Honglie Chen, Weidi Xie, Triantafyllos Afouras, Arsha Nagrani, Andrea Vedaldi, and Andrew Zisserman. Localizing Visual Sounds the Hard Way. In CVPR, 2021. 7", + "[11] Haoxin Chen, Yong Zhang, Xiaodong Cun, Menghan Xia, Xintao Wang, Chao Weng, and Ying Shan. VideoCraft2: Overcoming Data Limitations for High-Quality Video Diffusion Models. In CVPR, 2024. 2", + "[12] Gene Chou, Yuval Bahat, and Felix Heide. Diffusion-SDF: Conditional Generative Modeling of Signed Distance Functions. In CVPR, 2023. 2", + "[13] Pinaki Nath Chowdhury, Ayan Kumar Bhunia, Viswanatha Reddy Gajjala, Aneeshan Sain, Tao Xiang, and Yi-Zhe Song. Partially Does It: Towards Scene-Level FG-SBIR with Partial Input. In CVPR, 2022. 2, 3", + "[14] Pinaki Nath Chowdhury, Tuanfeng Wang, Duygu Ceylan, Yi-Zhe Song, and Yulia Gryaditskaya. Garment ideation:" + ], + "bbox": [ + 99, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Iterative view-aware sketch-based garment modeling. In 3DV, 2022. 2", + "[15] Pinaki Nath Chowdhury, Ayan Kumar Bhunia, Aneshan Sain, Subhadeep Koley, Tao Xiang, and Yi-Zhe Song. SceneTrilogy: On Human Scene-Sketch and its Complementarity with Photo and Text. In CVPR, 2023. 4, 5", + "[16] Pinaki Nath Chowdhury, Ayan Kumar Bhunia, Aneshan Sain, Subhadeep Koley, Tao Xiang, and Yi-Zhe Song. What Can Human Sketches Do for Object Detection? In CVPR, 2023. 2, 3, 5", + "[17] John P. Collomosse, Tu Bui, and Hailin Jin. LiveSketch: Query Perturbations for Guided Sketch-based Visual Search. In CVPR, 2019. 2", + "[18] Bram de Wilde, Anindo Saha, Richard PG ten Broek, and Henkjan Huisman. Medical diffusion on a budget: Textual Inversion for medical image generation. In MIDL, 2024. 2, 4", + "[19] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. ImageNet: A Large-Scale Hierarchical Image Database. In CVPR, 2009. 6", + "[20] Sounak Dey, Pau Riba, Anjan Dutta, Josep Llados, and Yi-Zhe Song. Doodle to Search: Practical Zero-Shot Sketch-based Image Retrieval. In CVPR, 2019. 2, 6", + "[21] Prafulla Dhariwal and Alexander Nichol. Diffusion Models Beat GANs on Image Synthesis. In NeurIPS, 2021. 2", + "[22] Francesco Di Sario, Riccardo Renzulli, Enzo Tartaglione, and Marco Grangetto. Boost Your NeRF: A Model-Agnostic Mixture of Experts Framework for High Quality and Efficient Rendering. In ECCV, 2024. 3", + "[23] Shiyin Dong, Mingrui Zhu, Kun Cheng, Nannan Wang, and Xinbo Gao. Bridging generative and discriminative models for unified visual perception with diffusion priors. In IJCAI, 2024. 2", + "[24] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 3", + "[25] Anjan Dutta and Zeynep Akata. Semantically Tied Paired Cycle Consistency for Zero-Shot Sketch-based Image Retrieval. In CVPR, 2019. 6", + "[26] Mathias Eitz, James Hays, and Marc Alexa. How do humans sketch objects? ACM TOG, 2012. 1, 5, 6", + "[27] Arnab Ghosh, Richard Zhang, Puneet K Dokania, Oliver Wang, Alexei A Efros, Philip HS Torr, and Eli Shechtman. Interactive Sketch & Fill: Multiclass Sketch-to-Image Translation. In ICCV, 2019. 2", + "[28] Rafael C Gonzalez and Richard E Woods. Digital Image Processing. Pearson, 2008. 4", + "[29] David Ha and Douglas Eck. A Neural Representation of Sketch Drawings. In ICLR, 2017. 5, 6", + "[30] Moayed Haji-Ali, Guha Balakrishnan, and Vicente Ordonez. ElasticDiffusion: Training-free Arbitrary Size Image Generation through Global-Local Content Separation. In CVPR, 2024. 3" + ], + "bbox": [ + 522, + 92, + 903, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "2564", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[31] Cusuh Ham, Gemma Canet Tarres, Tu Bui, James Hays, Zhe Lin, and John Collomosse. CoGS: Controllable Generation and Search from Sketch and Style. In ECCV, 2022. 2", + "[32] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep Residual Learning for Image Recognition. In CVPR, 2016. 5", + "[33] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-Prompt Image Editing with Cross Attention Control. In ICLR, 2022. 2", + "[34] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising Diffusion Probabilistic Models. In NeurIPS, 2020. 2, 3", + "[35] Conghui Hu, Da Li, Yongxin Yang, Timothy M Hospedales, and Yi-Zhe Song. Sketch-a-Segmented: Sketch-based Photo Segmenter Generation. IEEE TIP, 2020. 2, 5, 7, 8", + "[36] Drew A Hudson, Daniel Zoran, Mateusz Malinowski, Andrew K Lampinen, Andrew Jaegle, James L McClelland, Loic Matthey, Felix Hill, and Alexander Lerchner. SODA: Bottleneck Diffusion Models for Representation Learning. In CVPR, 2024. 2", + "[37] HyeongJoo Hwang, Geon-Hyeong Kim, Seunghoon Hong, and Kee-Eung Kim. Variational Interaction Information Maximization for Cross-domain Disentanglement. In NeurIPS, 2020. 6", + "[38] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In ECCV, 2022. 3", + "[39] Tero Karras, Samuli Laine, and Timo Aila. A Style-Based Generator Architecture for Generative Adversarial Networks. In CVPR, 2019. 2", + "[40] Bahjat Kawar, Shiran Zada, Oran Lang, Omer Tov, Huiwen Chang, Tali Dekel, Inbar Mosseri, and Michal Irani. Imagic: Text-Based Real Image Editing with Diffusion Models. In CVPR, 2023. 2", + "[41] Kazuma Kobayashi and Lin Gu and Ryuichiro Hataya and Takaaki Mizuno and Mototaka Miyake and Hirokazu Watanabe and Masamichi Takahashi and Yasuyuki Takamizawa and Yukihiro Yoshida and Satoshi Nakamura and Nobuji Kouno and Amina Bolatkan and Yusuke Kurose and Tatsuya Harada and Ryuji Hamamoto. Sketch-based Medical Image Retrieval. Medical Image Analysis, 2024. 2", + "[42] Diederik P Kingma and Max Welling. Auto-Encoding Variational Bayes. In ICLR, 2014. 3, 5", + "[43] Subhadeep Koley, Ayan Kumar Bhunia, Aneeshan Sain, Pinaki Nath Chowdhury, Tao Xiang, and Yi-Zhe Song. Picture that Sketch: Photorealistic Image Generation from Abstract Sketches. In CVPR, 2023. 2, 3, 4", + "[44] Subhadeep Koley, Ayan Kumar Bhunia, Aneeshan Sain, Pinaki Nath Chowdhury, Tao Xiang, and Yi-Zhe Song. You'll Never Walk Alone: A Sketch and Text Duet for Fine-Grained Image Retrieval. In CVPR, 2024. 2", + "[45] Subhadeep Koley, Ayan Kumar Bhunia, Aneeshan Sain, Pinaki Nath Chowdhury, Tao Xiang, and Yi-Zhe Song. How to Handle Sketch-Abstraction in Sketch-Based Image Retrieval? In CVPR, 2024. 1, 3" + ], + "bbox": [ + 99, + 92, + 480, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[46] Subhadeep Koley, Ayan Kumar Bhunia, Aneeshan Sain, Pinaki Nath Chowdhury, Tao Xiang, and Yi-Zhe Song. Text-to-Image Diffusion Models are Great Sketch-Photo Matchmakers. In CVPR, 2024. 1, 2, 3, 4, 5, 6, 8", + "[47] Subhadeep Koley, Ayan Kumar Bhunia, Deeptanshu Sekhri, Aneeshan Sain, Pinaki Nath Chowdhury, Tao Xiang, and Yi-Zhe Song. It's All About Your Sketch: Democratising Sketch Control in Diffusion Models. In CVPR, 2024. 1, 2, 4, 5", + "[48] Philipp Krahenbuhl and Vladlen Koltun. Efficient Inference in Fully Connected CRFs with Gaussian Edge Potentials. In NeurIPS, 2011. 7", + "[49] Alexander C Li, Mihir Prabhudesai, Shivam Duggal, Ellis Brown, and Deepak Pathak. Your Diffusion Model is Secretly a Zero-Shot Classifier. In ICCV, 2023. 2, 5", + "[50] Ke Li, Kaiyue Pang, and Yi-Zhe Song. Photo Pre-Training, but for Sketch. In CVPR, 2023. 2, 3", + "[51] Minchen Li, Alla Sheffer, Eitan Grinspun, and Nicholas Vining. FoldSketch: Enriching Garments with Physically Reproducible Folds. ACM TOG, 2018. 2", + "[52] Fengyin Lin, Mingkang Li, Da Li, Timothy Hospedales, YiZhe Song, and Yonggang Qi. Zero-Shot Everything Sketch-Based Image Retrieval, and in Explainable Style. In CVPR, 2023. 3", + "[53] Li Liu, Fumin Shen, Yuming Shen, Xianglong Liu, and Ling Shao. Deep Sketch Hashing: Fast Free-hand Sketch-Based Image Retrieval. In CVPR, 2017. 6", + "[54] Qing Liu, Lingxi Xie, Huiyu Wang, and Alan Yuille. Semantic-Aware Knowledge Preservation for Zero-Shot Sketch-Based Image Retrieval. In ICCV, 2019. 6", + "[55] Xuanchen Lu, Xiaolong Wang, and Judith E Fan. Learning Dense Correspondences between Photos and Sketches. In ICML, 2023. 5, 7, 8", + "[56] Ling Luo, Yulia Gryaditskaya, Tao Xiang, and Yi-Zhe Song. Structure-Aware 3D VR Sketch to 3D Shape Retrieval. In 3DV, 2022. 2", + "[57] Ling Luo, Pinaki Nath Chowdhury, Tao Xiang, Yi-Zhe Song, and Yulia Gryaditskaya. 3D VR Sketch Guided 3D Shape Prototyping and Exploration. In ICCV, 2023. 2", + "[58] Aryan Mikaeili, Or Perel, Daniel Cohen-Or, and Ali Mahdavi-Amiri. SKED: Sketch-guided Text-based 3D Editing. In CVPR, 2023. 2", + "[59] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, and Ying Shan. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. In AAAI, 2024. 2", + "[60] Jishnu Mukhoti, Tsung-Yu Lin, Omid Poursaeed, Rui Wang, Ashish Shah, Philip HS Torr, and Ser-Nam Lim. Open Vocabulary Semantic Segmentation with Patch Aligned Contrastive Learning. In CVPR, 2023. 5", + "[61] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. DINov2: Learning Robust Visual Features without Supervision. TMLR, 2024. 6, 8", + "[62] Zixuan Pan, Jianxu Chen, and Yiyu Shi. Masked Diffusion as Self-supervised Representation Learner. arXiv preprint arXiv:2308.05695, 2023. 2" + ], + "bbox": [ + 524, + 92, + 903, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "2565", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[63] Kaiyue Pang, Ke Li, Yongxin Yang, Honggang Zhang, Timothy M Hospedales, Tao Xiang, and Yi-Zhe Song. Generalising Fine-Grained Sketch-Based Image Retrieval. In CVPR, 2019. 6", + "[64] Kaiyue Pang, Yongxin Yang, Timothy M Hospedales, Tao Xiang, and Yi-Zhe Song. Solving mixed-modal jigsaw puzzle for fine-grained sketch-based image retrieval. In CVPR, 2020. 2, 3", + "[65] Mihir Prabhudesai, Tsung-Wei Ke, Alexander Cong Li, Deepak Pathak, and Katerina Fragkiadaki. Diffusion-TTA: Test-time Adaptation of Discriminative Models via Generative Feedback. In NeurIPS, 2023. 2", + "[66] Zhiyu Qu, Yulia Gryaditskaya, Ke Li, Kaiyue Pang, Tao Xiang, and Yi-Zhe Song. SketchXAI: A First Look at Explainability for Human Sketches. In CVPR, 2023. 6, 7, 8", + "[67] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning Transferable Visual Models From Natural Language Supervision. In ICML, 2021. 1, 2, 3, 4, 5, 6, 7, 8", + "[68] Leo Sampaio Ferraz Ribeiro, Tu Bui, John Collomosse, and Moacir Ponti. Sketchformer: Transformer-based Representation for Sketched Structure. In CVPR, 2020. 2, 3, 6", + "[69] Elad Richardson, Yuval Alaluf, Or Patashnik, Yotam Nitzan, Yaniv Azar, Stav Shapiro, and Daniel Cohen-Or. Encoding in Style: a StyleGAN Encoder for Image-to-Image Translation. In CVPR, 2021. 2", + "[70] Ignacio Rocco, Relja Arandjelović, and Josef Sivic. End-to-end weakly-supervised semantic alignment. In CVPR, 2018. 7", + "[71] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-Resolution Image Synthesis with Latent Diffusion Models. In CVPR, 2022. 1, 2, 3, 4, 5, 6, 8", + "[72] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-Net: Convolutional Networks for Biomedical Image Segmentation. In MICCAI, 2015. 3", + "[73] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. DreamBooth: Fine Tuning Text-to-Image Diffusion Models for Subject-Driven Generation. In CVPR, 2023. 2", + "[74] Aneeshan Sain, Ayan Kumar Bhunia, Yongxin Yang, Tao Xiang, and Yi-Zhe Song. StyleMeUp: Towards Style-Agnostic Sketch-Based Image Retrieval. In CVPR, 2021. 3", + "[75] Aneeshan Sain, Ayan Kumar Bhunia, Vaishnav Potlapalli, Pinaki Nath Chowdhury, Tao Xiang, and Yi-Zhe Song. Sketch3T: Test-Time Training for Zero-Shot SBIR. In CVPR, 2022. 3", + "[76] Aneeshan Sain, Ayan Kumar Bhunia, Pinaki Nath Chowdhury, Aneeshan Sain, Subhadeep Koley, Tao Xiang, and Yi-Zhe Song. CLIP for All Things Zero-Shot Sketch-Based Image Retrieval, Fine-Grained or Not. In CVPR, 2023. 1, 2, 3, 5, 6, 8", + "[77] Aneeshan Sain, Ayan Kumar Bhunia, Subhadeep Koley, Pinaki Nath Chowdhury, Soumitri Chattopadhyay, Tao Xiang, and Yi-Zhe Song. Exploiting Unlabelled Photos for Stronger Fine-Grained SBIR. In CVPR, 2023. 3" + ], + "bbox": [ + 99, + 92, + 482, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[78] Patsorn Sangkloy, Nathan Burnell, Cusuh Ham, and James Hays. The Sketchy Database: Learning to Retrieve Badly Drawn Bunnies. ACM TOG, 2016. 1, 5, 6, 7", + "[79] Patsorn Sangkloy, Wittawat Jitkrittum, Diyi Yang, and James Hays. A Sketch Is Worth a Thousand Words: Image Retrieval with Text and Sketch. In ECCV, 2022. 2, 3, 4, 5", + "[80] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. LAION-5B: An open large-scale dataset for training next generation image-text models. In NeurlPSW, 2022. 4", + "[81] Jaehyeok Shim, Changwoo Kang, and Kyungdon Joo. Diffusion-Based Signed Distance Fields for 3D Shape Generation. In CVPR, 2023. 2", + "[82] Chenyang Si, Ziqi Huang, Yuming Jiang, and Ziwei Liu. FreeU: Free Lunch in Diffusion U-Net. In CVPR, 2024. 4", + "[83] Jifei Song, Qian Yu, Yi-Zhe Song, Tao Xiang, and Timothy M Hospedales. Deep Spatial-Semantic Attention for Fine-Grained Sketch-Based Image Retrieval. In ICCV, 2017. 3", + "[84] Luming Tang, Menglin Jia, Qianqian Wang, Cheng Perng Phoo, and Bharath Hariharan. Emergent Correspondence from Image Diffusion. In NeurIPS, 2023. 1, 2, 4, 5, 8", + "[85] Prune Truong, Martin Danelljan, Fisher Yu, and Luc Van Gool. Warp Consistency for Unsupervised Learning of Dense Correspondences. In ICCV, 2021. 7, 8", + "[86] Narek Tumanyan, Michal Geyer, Shai Bagon, and Tali Dekel. Plug-and-Play Diffusion Features for Text-Driven Image-to-Image Translation. In CVPR, 2023. 2", + "[87] Andrey Voynov, Kfir Aberman, and Daniel Cohen-Or. Sketch-guided text-to-image diffusion models. In SIGGRAPH, 2023. 2", + "[88] Fei Wang. Sketch2Vox: Learning 3D Reconstruction from a Single Monocular Sketch Image. In ECCV, 2024. 2", + "[89] Wenxuan Wang, Quan Sun, Fan Zhang, Yepeng Tang, Jing Liu, and Xinlong Wang. Diffusion Feedback Helps CLIP See Better. In ICLR, 2025. 3", + "[90] Size Wu, Wenwei Zhang, Lumin Xu, Sheng Jin, Xiangtai Li, Wentao Liu, and Chen Change Loy. CLIPSelf: Vision Transformer Distills Itself for Open-Vocabulary Dense Prediction. In ICLR, 2024. 3", + "[91] Chunlong Xia, Xinliang Wang, Feng Lv, Xin Hao, and Yifeng Shi. ViT-CoMer: Vision Transformer with Convolutional Multi-scale Feature Interaction for Dense Predictions. In CVPR, 2024. 3", + "[92] Jinbo Xing, Hanyuan Liu, Menghan Xia, Yong Zhang, Xintao Wang, Ying Shan, and Tien-Tsin Wong. ToonCrafter: Generative Cartoon Interpolation. ACM TOG, 2024. 2", + "[93] Chenfeng Xu, Huan Ling, Sanja Fidler, and Or Litany. 3DiffTec: 3D Object Detection with Geometry-Aware Diffusion Features. In CVPR, 2023. 2, 4, 5", + "[94] Jiarui Xu, Sifei Liu, Arash Vahdat, Wonmin Byeon, Xiaolong Wang, and Shalini De Mello. Open-Vocabulary Panoptic Segmentation with Text-to-Image Diffusion Models. In CVPR, 2023. 1, 2, 4" + ], + "bbox": [ + 524, + 92, + 903, + 898 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "2566", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[95] Mengde Xu, Zheng Zhang, Fangyun Wei, Yutong Lin, Yue Cao, Han Hu, and Xiang Bai. A Simple Baseline for Open-Vocabulary Semantic Segmentation with Pre-trained Vision-language Model. In ECCV, 2022. 7, 8", + "[96] Peng Xu, Yongye Huang, Tongtong Yuan, Kaiyue Pang, YiZhe Song, Tao Xiang, Timothy M Hospedales, Zhanyu Ma, and Jun Guo. SketchMate: Deep Hashing for Million-Scale Human Sketch Retrieval. In CVPR, 2018. 6, 7, 8", + "[97] Peng Xu, Timothy Hospedales, Qiyue Yin, Yi-Zhe Song, Tao Xiang, and Liang Wang. Deep Learning for Free-Hand Sketch: A Survey. IEEE TPAMI, 2020. 1", + "[98] Antoine Yang, Arsha Nagrani, Paul Hongsuck Seo, Antoine Miech, Jordi Pont-Tuset, Ivan Laptev, Josef Sivic, and Cordelia Schmid. Vid2Seq: Large-Scale Pretraining of a Visual Language Model for Dense Video Captioning. In CVPR, 2023. 5", + "[99] Kai Yang, Jian Tao, Jiafei Lyu, Chunjiang Ge, Jiaxin Chen, Weihan Shen, Xiaolong Zhu, and Xiu Li. Using human feedback to fine-tune diffusion models without any reward model. In CVPR, 2024. 3", + "[100] Lan Yang, Kaiyue Pang, Honggang Zhang, and Yi-Zhe Song. SketchAA: Abstract Representation for Abstract Sketches. In ICCV, 2021. 1, 5", + "[101] Lumin Yang, Jiajie Zhuang, Hongbo Fu, Xiangzhi Wei, Kun Zhou, and Youyi Zheng. SketchGNN: Semantic Sketch Segmentation with Graph Neural Networks. ACM TOG, 2021. 6, 7, 8", + "[102] Sasi Kiran Yelamarthi, Shiva Krishna Reddy, Ashish Mishra, and Anurag Mittal. A Zero-Shot Framework for Sketch Based Image Retrieval. In ECCV, 2018. 2, 6", + "[103] Dong Yin, Raphael Gontijo Lopes, Jon Shlens, Ekin Dogus Cubuk, and Justin Gilmer. A Fourier Perspective on Model Robustness in Computer Vision. In NeurIPS, 2019. 4", + "[104] Jiazuo Yu, Yunzhi Zhuge, Lu Zhang, Ping Hu, Dong Wang, Huchuan Lu, and You He. Boosting Continual Learning of Vision-Language Models via Mixture-of-Experts Adapters. In CVPR, 2024. 3", + "[105] Seonghoon Yu, Paul Hongsuck Seo, and Jeany Son. Zeroshot Referring Image Segmentation with Global-Local Context Features. In CVPR, 2023. 3", + "[106] Junyi Zhang, Charles Herrmann, Junhwa Hur, Eric Chen, Varun Jampani, Deqing Sun, and Ming-Hsuan Yang. Telling left from right: Identifying geometry-aware semantic correspondence. In CVPR, 2024. 7", + "[107] Junyi Zhang, Charles Herrmann, Junhwa Hur, Luisa Polania Cabrera, Varun Jampani, Deqing Sun, and Ming-Hsuan Yang. A Tale of Two Features: Stable Diffusion Complements DINO for Zero-Shot Semantic Correspondence. NeurIPS, 2024. 1, 2, 3, 4, 8", + "[108] Lvmin Zhang and Maneesh Agrawala. Adding Conditional Control to Text-to-Image Diffusion Models. In ICCV, 2023. 2, 3", + "[109] Manyuan Zhang, Guanglu Song, Xiaoyu Shi, Yu Liu, and Hongsheng Li. Three Things We Need to Know About Transferring Stable Diffusion to Visual Dense Prediction Tasks. In ECCV, 2024. 2" + ], + "bbox": [ + 93, + 90, + 485, + 898 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[110] Wenliang Zhao, Yongming Rao, Zuyan Liu, Benlin Liu, Jie Zhou, and Jiwen Lu. Unleashing Text-to-Image Diffusion Models for Visual Perception. In ICCV, 2023. 1, 2, 4", + "[111] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Zwei Liu. Conditional Prompt Learning for Vision-Language Models. In CVPR, 2022. 3", + "[112] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Zwei Liu. Learning to Prompt for Vision-Language Models. IJCV, 2022. 3" + ], + "bbox": [ + 516, + 92, + 903, + 218 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "2567", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2025/SketchFusion_ Learning Universal Sketch Features through Fusing Foundation Models/1fc738f6-a4cf-4a97-bcd3-ea7682cab3c6_model.json b/2025/SketchFusion_ Learning Universal Sketch Features through Fusing Foundation Models/1fc738f6-a4cf-4a97-bcd3-ea7682cab3c6_model.json new file mode 100644 index 0000000000000000000000000000000000000000..1bf5536604a5e6dd0725e345df690903e3b1b73a --- /dev/null +++ b/2025/SketchFusion_ Learning Universal Sketch Features through Fusing Foundation Models/1fc738f6-a4cf-4a97-bcd3-ea7682cab3c6_model.json @@ -0,0 +1,2787 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.198, + 0.095, + 0.8, + 0.138 + ], + "angle": 0, + "content": "SketchFusion: Learning Universal Sketch Features through Fusing Foundation Models" + }, + { + "type": "text", + "bbox": [ + 0.146, + 0.139, + 0.85, + 0.175 + ], + "angle": 0, + "content": "Subhadeep Koley\\(^{1,2*}\\) Tapas Kumar Dutta\\(^{1*}\\) Aneeshan Sain\\(^{1}\\) Pinaki Nath Chowdhury\\(^{1}\\) Ayan Kumar Bhunia\\(^{1}\\) Yi-Zhe Song\\(^{1,2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.269, + 0.175, + 0.729, + 0.193 + ], + "angle": 0, + "content": "\\(^{1}\\)SketchX, CVSSP, University of Surrey, United Kingdom." + }, + { + "type": "text", + "bbox": [ + 0.245, + 0.193, + 0.753, + 0.211 + ], + "angle": 0, + "content": "\\(^{2}\\)iFlyTek-Surrey Joint Research Centre on Artificial Intelligence." + }, + { + "type": "text", + "bbox": [ + 0.227, + 0.213, + 0.765, + 0.228 + ], + "angle": 0, + "content": "{s.koley, a.sain, p.chowdhury, a.bhunia, y-song}@surrey.ac.uk" + }, + { + "type": "image", + "bbox": [ + 0.092, + 0.23, + 0.359, + 0.353 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.362, + 0.23, + 0.682, + 0.352 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.687, + 0.23, + 0.906, + 0.353 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.359, + 0.907, + 0.526 + ], + "angle": 0, + "content": "Figure 1. (Left): Apart from high-resolution image generation, text-to-image diffusion models (e.g., Stable diffusion (SD) [71]) with their innate object understanding capability [84, 107], have shown remarkable performance across a wide range of image-based vision tasks (e.g., segmentation [94], depth estimation [110], etc.). However, upon analysing the PCA representation of SD's intermediate UNet features, we observe that it struggles to achieve similar results when working with freehand abstract sketches (detail in Sec. 4). Unlike pixel-perfect photos, highly abstract freehand sketches are sparse and lack detailed textures and colours [26], making it harder for the SD model to extract meaningful features. Furthermore, investigating the SD denoising process in the frequency domain (via Fourier Transform), we observe the predominance of high-frequency (HF) components, rather than their low-frequency (LF) counterpart - crucial for capturing comprehensive semantic context. To mitigate this inherent bias within SD, we reinforce the diffusion process with another pretrained model (i.e., CLIP [67]) whose bias is complementary (i.e., focuses on LF) to SD. Consequently, the proposed extractor can extract semantically meaningful and accurate features from both sketches and photos, that encapsulate a broader frequency spectrum (i.e., HF and LF). (Right:) Testing the proposed method with different sketch-based discriminative and dense prediction tasks (requiring knowledge of both sketch and image), we find a marked improvement over baseline SD+CLIP hybrid feature extractor." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.533, + 0.327, + 0.548 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.567, + 0.486, + 0.868 + ], + "angle": 0, + "content": "While foundation models have revolutionised computer vision, their effectiveness for sketch understanding remains limited by the unique challenges of abstract, sparse visual inputs. Through systematic analysis, we uncover two fundamental limitations: Stable Diffusion (SD) struggles to extract meaningful features from abstract sketches (unlike its success with photos), and exhibits a pronounced frequency-domain bias that suppresses essential low-frequency components needed for sketch understanding. Rather than costly retraining, we address these limitations by strategically combining SD with CLIP, whose strong semantic understanding naturally compensates for SD's spatial-frequency biases. By dynamically injecting CLIP features into SD's denoising process and adaptively aggregating features across semantic levels, our method achieves state-of-the-art performance in sketch retrieval \\((+3.35\\%)\\), recognition \\((+1.06\\%)\\), segmentation \\((+29.42\\%)\\), and correspondence learning \\((+21.22\\%)\\), demonstrating the first truly universal sketch feature representation in the era of foundation models." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.533, + 0.645, + 0.548 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.552, + 0.907, + 0.703 + ], + "angle": 0, + "content": "The quest for universal sketch-specific features has been a foundational challenge in computer vision [46, 76, 78, 97]. Sketches, with their inherent characteristics of abstraction, sparsity, and cross-modal interpretation, demand fundamentally different feature representations compared to natural images [45, 100]. While foundation models [67, 71] have revolutionised various visual tasks, their effectiveness for sketch understanding remains largely unexplored [47], particularly given sketch's unique position at the intersection of human abstraction and computer vision [26, 45]." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.705, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Through systematic analysis, we find that even powerful foundation models like Stable Diffusion (SD) [71] face significant challenges with sketch data. Our pilot studies (Sec. 4) reveal two fundamental limitations. First, while SD excels at extracting task-specific representative features from pixel-perfect RGB images, it struggles with abstract freehand sketches, likely due to its large-scale pretraining being predominantly focused on photographic data [71]. Second, and more intriguingly, frequency-domain analysis of SD's UNet reveals an inherent architectural bias: the model systematically promotes high-frequency (HF) components while suppressing low-frequency (LF) information - a characteristic that proves particularly problematic" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.218, + 0.9 + ], + "angle": 0, + "content": "*Equal contribution" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2556" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.122 + ], + "angle": 0, + "content": "for sketch understanding tasks, especially dense prediction problems like segmentation and correspondence learning." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.124, + 0.483, + 0.32 + ], + "angle": 0, + "content": "These findings present a significant challenge. Retraining SD on sketch data risks catastrophic forgetting of its valuable pretrained knowledge [47], while developing sketch-specific architectures would fail to leverage the rich visual understanding already present in foundation models. Our key insight comes from analysing the complementary nature of different foundation models: while SD's features are spatially-aware but lack semantic accuracy, CLIP's [67] visual encoder exhibits strong semantic understanding even at the cost of spatial precision. This complementarity suggests that strategic combination [107] of these models could yield features that capture both the detailed structure and semantic meaning necessary for sketch understanding." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.322, + 0.483, + 0.488 + ], + "angle": 0, + "content": "Based on this insight, we propose a novel framework that leverages the complementary strengths of these foundation models. Our method systematically injects CLIP's visual features into multiple layers of SD's UNet during the denoising process, using lightweight \\(1D\\) convolutions to influence feature extraction at various semantic levels. PCA analysis of the resulting features demonstrates that this combination successfully captures both the HF spatial details from SD [71] and the LF semantic components from CLIP [67], producing rich, semantically-aware representations suited for diverse sketch-based tasks." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.49, + 0.483, + 0.642 + ], + "angle": 0, + "content": "Moreover, we observe that different layers of SD's UNet capture features at varying levels of semantic-granularity [46], with different combinations better suited to specific downstream tasks [84]. Rather than manually selecting optimal features, we train a dynamic feature aggregator that automatically determines the most effective combination for any given task, eliminating the need for task-specific tuning. This adaptive approach ensures optimal performance across diverse sketch-based applications without requiring manual layer selection or task-specific optimisation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.643, + 0.483, + 0.809 + ], + "angle": 0, + "content": "Our extensive experiments demonstrate the effectiveness of this approach across a comprehensive suite of sketch-based tasks: (i) Sketch Recognition: \\(+1.06\\%\\) improvement over state-of-the-art, (ii) Sketch-based Image Retrieval: \\(+3.35\\%\\) improvement in retrieval accuracy, (iii) Sketch-based Segmentation: \\(+29.42\\%\\) boost in mean IoU, (iv) Sketch-Photo Correspondence: \\(+21.22\\%\\) increase in matching precision. These significant improvements across all major sketch-specific tasks suggest that our method has uncovered fundamental principles for extracting effective sketch features from foundation models." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.483, + 0.902 + ], + "angle": 0, + "content": "In summary, our contributions are: (i) A systematic analysis revealing the limitations of pretrained foundation models for sketch understanding, supported by both empirical results and frequency-domain analysis, (ii) A novel method for combining complementary foundation models that address these limitations without retraining, (iii) An adap" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.183 + ], + "angle": 0, + "content": "tive feature aggregation approach that automatically optimises representations for different downstream tasks, and \\((iv)\\) Comprehensive experimental validation demonstrating substantial improvements across all major sketch-specific tasks, establishing a new paradigm for universal sketch feature learning." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.188, + 0.665, + 0.204 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.207, + 0.905, + 0.584 + ], + "angle": 0, + "content": "Diffusion Model for Representation Learning. Diffusion models [34] have now become the de-facto standard for high-resolution \\(2D\\) [21, 71] and \\(3D\\) [12, 81] image and video [11] generation. Apart from generation, it gave rise to various image editing frameworks like Prompt-to-Prompt [33], Imagic [40], Dreambooth [73], etc. The success of diffusion models across various generative tasks and their inherent zero-shot generalisability has led to their adoption as efficient feature extractors in different forms of discriminative [23] and dense prediction [109] tasks like image retrieval [46], classification [49], semantic [3] and panoptic [94] segmentation, depth estimation [110], object detection [93], image-to-image translation [47, 86], controllable-generation [59, 108] and medical imaging [18], to name a few. A few recent works have developed specialised diffusion models together with carefully crafted task-specific designs like test-time adaptation [65], masked-reconstruction [62], or compact latent generation [36] to improve their representation ability. Nonetheless, our pilot study (Sec. 4) reveals that SD, often struggles to extract representative features (Fig. 1 left) from binary and sparse sketches, hampering downstream task performance. In this work, we leverage a pretrained CLIP [67] visual encoder to enhance SD's feature representation abilities for sketch-based discriminative and dense prediction tasks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.585, + 0.905, + 0.825 + ], + "angle": 0, + "content": "Sketch-based Vision Applications. \"Freehand sketch\" has now established itself as a worthy interactive alternative modality to text for downstream vision tasks [9, 16, 31, 43, 76, 87, 102]. With its success in sketch-based image retrieval (SBIR) [13, 20, 76, 79], sketches are now being used in a wide array of discriminative [16, 17, 20], generative [27, 43, 108] and dense prediction [7, 35] tasks like controllable \\(2D\\) [47, 108], \\(3D\\) [2, 88] image generation and editing [31, 43, 58, 59, 69, 108], video synthesis [92], garment designing [14, 51], segmentation [35], creative content generation [9], object detection [16], salient region prediction [7], augmented and virtual reality applications [56, 57], medical imaging [41], class incremental learning [5], or composed retrieval [44]. In this paper, we aim to tackle some of the sketch-based downstream tasks by reinforcing the diffusion feature extractor via a pretrained CLIP [67] visual encoder." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.905, + 0.902 + ], + "angle": 0, + "content": "Backbones for Sketch-based Vision Applications. Backbones for sketch-based vision applications can be broadly categorised as - (i) pretrained CNN or Vision Transformer (ViT) [6, 13, 68], (ii) Self-supervised pretraining [4, 50, 64], and (iii) pretrained foundation models [39, 67, 71]. Starting" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2557" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.091, + 0.482, + 0.394 + ], + "angle": 0, + "content": "with ImageNet pretrained CNNs [6, 13] or JFT pretrained ViTs [68, 77] with carefully-crafted design choices (e.g., cross [52] and spatial [83] attention, reinforcement learning [6], test-time training [75], meta-learning [74], knowledge distillation [77], etc.), the community moved towards self-supervised pretraining either via solving pretext tasks like sketch vectorisation [4], rasterisation [4], solving sketch-photo jigsaw [64] or by large-scale photo pretraining [50]. With the advent of large-scale foundation models, several works [43, 46, 76, 79] used them as a backbone for different sketch-based tasks like image generation [43, 108], shape synthesis [2], or retrieval [46, 79]. Foundation models have now become the gold standard for sketch-based tasks due to their strong zero-shot performance and extensive pretrained knowledge [45, 46, 79]. Among these, generative models like SD are spatially aware [46, 107], yet less accurate for dense prediction tasks. In contrast, vision-language models like CLIP [67] offer accurate features but are inherently sparse [107]. In this paper, we aim to get the best of both worlds by enhancing the SD feature extraction with CLIP." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.395, + 0.483, + 0.741 + ], + "angle": 0, + "content": "Reinforcing Feature Extractors. There are numerous ways of reinforcing an off-the-shelf pretrained feature extractor to adapt them for different downstream tasks. It can broadly be categorised as \\((-i)\\) Mixture of Experts (MoE) [22, 104], \\((ii)\\) Prompt learning [38, 112], \\((iii)\\) Hybrid models [91, 107], \\((iv)\\) Finetuning [76, 99], and \\((v)\\) Global-local feature merging [30, 105]. MoE combines multiple specialised sub-models (i.e., \"experts\"), each responsible for specific tasks or data patterns, to boost overall performance and adaptability [22]. In contrast, prompt learning deals with learning additional trainable components in the textual [111, 112] or pixel [1, 38] domain for task-specific adaptation. Hybrid models [89] combine two different kinds of models like CNN and ViT [91], or SD and DINO [107] to extract the best of both worlds. Another avenue deals with finetuning large-scale pretrained models with selective handcrafted data for specific tasks [16, 76]. Finally, a few other works [30, 90, 105] amalgamate global and local features from the same [90] or different [30, 105] models to improve downstream task performance. In this paper, we take the hybrid model approach and complement CNN-based SD features with transformer-based CLIP visual encoder features for different kinds of sketch-based vision tasks." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.75, + 0.231, + 0.765 + ], + "angle": 0, + "content": "3. Preliminaries" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.771, + 0.279, + 0.787 + ], + "angle": 0, + "content": "3.1. Background: CLIP" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Contrastive language-image pretraining (CLIP) [67] architecture consists of a visual and a textual encoder, trained on \\(\\sim 400\\mathrm{M}\\) image-text pairs, with the goal of learning a shared embedding manifold. Using N-pair contrastive loss, the model aligns matching image-text pairs by maximising their cosine similarity and minimises it for random unpaired ones [67]. The image encoder, typically a" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.091, + 0.905, + 0.273 + ], + "angle": 0, + "content": "ViT [24], encodes an image \\(\\mathbf{i}\\) as a feature vector \\(\\mathbf{V}(\\mathbf{i})\\in\\) \\(\\mathbb{R}^d\\). Meanwhile, the textual encoder processes a word sequence \\(\\mathcal{W} = \\{w^0,w^1,\\dots ,w^k\\}\\) through byte pair encoding [67] and a learnable word embedding layer (vocabulary size of 49,152), converting each word \\(w^{i}\\) into a token embedding, thus creating an embedding sequence \\(\\mathcal{W}_e = \\{\\mathbf{w}_e^0,\\mathbf{w}_e^1,\\ldots ,\\mathbf{w}_e^k\\}\\) [67]. \\(\\mathcal{W}_e\\) is then passed through a transformer network \\(\\mathbf{T}(\\cdot)\\), producing the final text feature \\(\\mathbf{T}(\\mathcal{W}_e)\\in \\mathbb{R}^d\\) [67]. Once trained, the CLIP model can be used in downstream tasks to enable zero-shot capability and cross-modal understanding by aligning visual and textual data in a shared embedding space [67]." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.277, + 0.82, + 0.292 + ], + "angle": 0, + "content": "3.2. Background: Stable Diffusion (SD)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.294, + 0.905, + 0.581 + ], + "angle": 0, + "content": "Overview. The SD model [71] generates images by gradually removing noise from an initial \\(2D\\) isotropic Gaussian noise image through iterative refinement. It involves two reciprocal processes - \"forward\" and \"reverse\" diffusion [34, 71]. The forward process iteratively adds random Gaussian noise into a clean image \\(\\mathbf{i}_0 \\in \\mathbb{R}^{h \\times w \\times 3}\\) for \\(t\\) timesteps to generate a noisy image \\(\\mathbf{i}_t\\) as \\(\\mathbf{i}_t = \\sqrt{\\bar{\\alpha}_t} \\mathbf{i}_0 + \\sqrt{1 - \\bar{\\alpha}_t} \\epsilon\\). Here, \\(\\epsilon \\sim \\mathcal{N}(0, \\mathbf{I})\\), and \\(\\{\\alpha_t\\}_1^T\\) is a pre-defined noise schedule where \\(\\bar{\\alpha}_t = \\prod_{k=1}^t \\alpha_k\\) with \\(t \\sim U(0, T)\\). The reverse process involves training a denoising UNet [72] \\(\\mathbf{U}_{\\theta}\\) (with an \\(l_2\\) objective), which estimates the input noise \\(\\epsilon \\approx \\mathbf{U}_{\\theta}(\\mathbf{i}_t, t)\\) from the noisy image \\(\\mathbf{i}_t\\) at each \\(t\\). After training, \\(\\mathbf{U}_{\\theta}\\) is capable of reconstructing a clean image from a noisy input [71]. The inference process begins with a random \\(2D\\) noise sample, \\(\\mathbf{i}_T \\sim \\mathcal{N}(0, \\mathbf{I})\\). The trained model \\(\\mathbf{U}_{\\theta}\\) is then applied iteratively across \\(T\\) timesteps, progressively reducing noise at each step to produce a cleaner intermediate image, \\(\\mathbf{i}_{t-1}\\), finally generating one of the cleanest samples \\(\\mathbf{i}_0\\) from the target distribution." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.582, + 0.905, + 0.657 + ], + "angle": 0, + "content": "The unconditional denoising diffusion could further be made conditional by governing \\(\\mathbf{U}_{\\theta}\\) with different auxiliary signals (e.g., textual captions) \\(\\mathbf{c}\\) [71]. Accordingly, \\(\\mathbf{U}_{\\theta}(\\mathbf{i}_t,t,\\mathbf{c})\\) can denoise \\(\\mathbf{i}_t\\) while being influenced by \\(\\mathbf{c}\\), commonly via cross-attention [71]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.658, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Architecture. Pixel-space diffusion model [34] is time-intensive as it operates on the full image resolution (i.e., \\(\\mathbf{i}_0\\in \\mathbb{R}^{h\\times w\\times 3}\\)). Contrarily, SD model [71] performs denoising in the latent-space, resulting in significantly faster and more stable processing. In the first stage, SD trains a variational autoencoder (VAE) [42] (an encoder \\(\\mathbf{E}(\\cdot)\\) and a decoder \\(\\mathbf{D}(\\cdot)\\)). \\(\\mathbf{E}(\\cdot)\\) transforms the input image into its latent representation \\(\\mathbf{z}_0 = \\mathbf{E}(\\mathbf{i}_0)\\in \\mathbb{R}^{\\frac{h}{8}\\times \\frac{w}{8}\\times d}\\). In the second stage, the UNet [72] \\(\\mathbf{U}_{\\theta}\\), directly denoises the latent images. \\(\\mathbf{U}_{\\theta}\\) comprises 12 encoding, 1 bottleneck, and 12 skip-connected decoding blocks [71]. Inside these encoding and decoding blocks, there are 4 downsampling \\((\\mathbf{U}_{\\mathbf{d}}^{1 - 4})\\) and 4 upsampling \\((\\mathbf{U}_{\\mathbf{u}}^{1 - 4})\\) layers respectively. In SD, a frozen CLIP textual encoder [67] \\(\\mathbf{T}(\\cdot)\\) converts the textual prompt \\(\\mathbf{c}\\) into token-sequence, that influences the denoising process of \\(\\mathbf{U}_{\\theta}\\) through cross-attention. \\(\\mathbf{U}_{\\theta}\\) is trained over an" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2558" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.09, + 0.482, + 0.183 + ], + "angle": 0, + "content": "\\(l_{2}\\) objective as: \\(\\mathcal{L}_{\\mathrm{l}_2} = \\mathbb{E}_{\\mathbf{z}_t,t,\\mathbf{c},\\epsilon}(||\\epsilon -\\mathbf{U}_\\theta (\\mathbf{z}_t,t,\\mathbf{T}(\\mathbf{c}))||_2^2)\\). During inference, a noisy latent \\(\\mathbf{z}_t\\) is sampled directly as: \\(\\mathbf{z}_t\\sim \\mathcal{N}(0,\\mathbf{I})\\). \\(\\mathbf{U}_{\\theta}\\) removes noise from \\(\\mathbf{z}_t\\) iteratively over \\(T\\) timesteps (conditioned on \\(\\mathbf{c}\\)) to produce a denoised latent image \\(\\hat{\\mathbf{z}}_0\\in \\mathbb{R}^{\\frac{h}{8}\\times \\frac{w}{8}\\times d}\\). The final high-resolution image is generated as: \\(\\hat{\\mathbf{i}} = \\mathbf{D}(\\hat{\\mathbf{z}}_0)\\in \\mathbb{R}^{h\\times w\\times 3}\\)." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.188, + 0.411, + 0.205 + ], + "angle": 0, + "content": "4. Pilot Study: Problems and Solution" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.207, + 0.485, + 0.643 + ], + "angle": 0, + "content": "What's Wrong with Diffusion Features? Binary and sparse freehand sketches contain much lesser semantic contextual information than pixel-perfect photos [15, 43, 79]. While SD [71] as a backbone feature extractor has established itself as a strong competitor in image-based vision tasks [3, 18, 93, 94, 110], we dive deep into its internal representations to evaluate its effectiveness for the more challenging cross-modal sketch-based vision tasks. To this end, we ask the research question - whether a frozen SD model is capable of extracting features from a sparse sketch that are equally-representative as those derived from a pixel perfect image. Given a few sketch-photo pairs, we extract the corresponding features (detail in Sec. 5.1) from the second upsampling block of a frozen SD's UNet \\(\\mathbf{U}_{\\theta}\\) with a fixed prompt \"a photo of [CLASS]\" and \\(t = 195\\). We perform principal component analysis (PCA) on the extracted high-dimensional features and plot the first three principal components as RGB images for visualisation. As seen in Fig. 1 (left), features extracted from sketches are significantly inferior to those extracted from photos. Unlike photos, abstract freehand sketches are sparse and lack detailed textures/colours [15, 43, 79], making it challenging for the SD model to capture fine-grained visual information from it. We hypothesise that SD struggles to extract equally representative features from sketches due to their minimal semantic cues and high level of abstraction [47]. Additionally, as the pretraining of SD was done predominantly on natural images [80], it is comparatively harder for SD to extract rich semantic information from a sparse sketch." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.645, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Digging deeper, we follow classical computer vision literature [28, 103] in shifting the focus from spatial-domain, to analyse the behaviour of extracted features from a different perspective, i.e., exploring its nature in the frequency-domain. We achieve this by calculating the \\(2D\\) Fourier magnitude spectrum of the extracted features, where the central region represents the LF components, while the HF components are arranged radially outward from the centre [28] (Fig. 1 left). Here we see that, although SD performs better on photos than sketches, it predominantly focuses on HF components in both cases. This is likely because SD's UNet is primarily trained as a denoiser thus operating with a global emphasis [107], where its skip connections additionally propagate HF components (e.g., edges), suppressing its LF counterparts (e.g., overall semantic structures) to ease the reconstruction of high-res image from noise [82]. We hypothesise that this very fact is directly conflicting with the" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.182 + ], + "angle": 0, + "content": "core motivation of our dense prediction task from a sparse data modality i.e., sketches, where capturing comprehensive semantic [107] context (LF) takes much higher precedence over preserving fine-details (HF), thus limiting SD's effectiveness [82, 107] for sketches, especially for dense prediction tasks (e.g., sketch-based segmentation)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.183, + 0.907, + 0.53 + ], + "angle": 0, + "content": "Reinforcing Diffusion Features. We identify two key problems in SD feature extraction - (i) SD features are not equally-representative for sketches and photos, and (ii) it mainly focuses on HF features, losing the crucial LF ones. In addressing these, our PCA analysis (Fig. 1 left) shows that, although SD is spatially-aware, it lacks accuracy. Inspired by existing works [107] on hybrid models, we propose complementing the SD feature extractor by combining it with another pretrained model that offers high accuracy, even if it is less spatially-aware. To this end, we influence the SD feature extraction pipeline with a pretrained CLIP [67] visual encoder as detailed in Sec. 5. PCA maps of CLIP (Fig. 1 left) features depict highly accurate background foreground separation, inherently providing the LF components that were missing in SD features. Finally, PCA maps of the proposed feature extractor (Fig. 1 left) demonstrate how the final representation is spatially-aware, accurate, integrates both HF and LF and works equally well for both sketches and photos. Apart from qualitative evidence, evaluating the performance of baseline \\(\\mathrm{SD + }\\mathrm{CLIP}\\) hybrid model, and the proposed feature extractor across different sketch-based discriminative and dense prediction tasks, we observe an average performance gain of \\(39.49\\%\\) (Fig. 1 right)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.536, + 0.731, + 0.553 + ], + "angle": 0, + "content": "5. Proposed Methodology" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.554, + 0.907, + 0.854 + ], + "angle": 0, + "content": "Overview. Observations from our pilot study (Sec. 4) motivate us to reinforce the SD [71] feature extraction pipeline for better downstream task accuracy. However, reinforcing the SD feature extractor pipeline is non-trivial. Firstly, finetuning the entire model with task-specific dataset might interrupt SD's inherent pretrained knowledge [46]. Furthermore, end-to-end finetuning does not necessarily resolve the issues pertaining to spatial-awareness vs. accuracy or HF vs. LF trade-off (Sec. 4). To this end, we resort to a hybrid model [107] approach and use a complementary pretrained model (i.e., CLIP [67]) to strengthen the SD feature extractor pipeline. Secondly, obtaining features from CLIP alone is insufficient. They must be integrated effectively into the denoising process of the SD UNet to yield the performance gains observed in Fig. 1. Finally, different layers of the SD's UNet capture a wide array of features across different scales and granularity (i.e., coarse to fine) [84], each with its own pros and cons suited to various downstream tasks [46]. Thus, it is essential to effectively aggregate these multi-scale features to maximise their utility." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.856, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Specifically, our salient design choices are \\((-i)\\) complementing SD [71] feature extraction with a pretrained CLIP [67] visual model, \\((ii)\\) an effective CLIP visual feature in" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "2559" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.089, + 0.092, + 0.482, + 0.122 + ], + "angle": 0, + "content": "jection module, and (iii) an aggregation module to unify features from different scales and granularities." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.127, + 0.279, + 0.14 + ], + "angle": 0, + "content": "5.1. Model Architecture" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.143, + 0.485, + 0.481 + ], + "angle": 0, + "content": "Stable Diffusion Feature Extraction. Internal activation maps of a pretrained SD UNet \\(\\mathbf{U}_{\\theta}\\) contain information-rich representations that have found use in tasks like image segmentation [3], object detection [93], classification [49], etc. Here, we aim to use these internal representations as an efficient backbone feature for various sketch-based vision tasks. Given an image-text pair \\((\\mathbf{i},\\mathbf{c})\\), the input image is first passed through the pretrained VAE [42] encoder \\(\\mathbf{E}(\\cdot)\\), converting it into an initial latent representation \\(\\mathbf{z}_0 = \\mathbf{E}(\\mathbf{i}) \\in \\mathbb{R}^{\\frac{h}{8} \\times \\frac{w}{8} \\times d}\\). Then, given a timestep value \\(t\\), SD performs forward diffusion on \\(\\mathbf{z}_0\\) to convert it to the \\(t^{th}\\) step noisy latent \\(\\mathbf{z}_t\\). Finally, the textual prompt embedding (Sec. 3.2) \\(\\mathbf{T}(\\mathbf{c})\\), noisy latent \\(\\mathbf{z}_t\\), and timestep \\(t\\) are passed to the pretrained denoising UNet [71] \\(\\mathbf{U}_{\\theta}(\\cdot)\\) to extract the internal features from its skip-connected upsampling layers \\(\\{\\mathbf{U}_{\\mathbf{u}}^n \\to f_{\\mathbf{u}}^n\\}_{n=1}^4\\). In SD v2.1 [71], for an input image \\(\\mathbf{i} \\in \\mathbb{R}^{h \\times w \\times 3}\\), upsampling layers \\(\\{\\mathbf{U}_{\\mathbf{u}}^n\\}_{n=1}^4\\) generate features as: \\(f_{\\mathbf{u}}^1 \\in \\mathbb{R}^{\\frac{h}{32} \\times \\frac{w}{32} \\times 1280}\\), \\(f_{\\mathbf{u}}^2 \\in \\mathbb{R}^{\\frac{h}{16} \\times \\frac{w}{16} \\times 1280}\\), \\(f_{\\mathbf{u}}^3 \\in \\mathbb{R}^{\\frac{h}{8} \\times \\frac{w}{8} \\times 640}\\), and \\(f_{\\mathbf{u}}^4 \\in \\mathbb{R}^{\\frac{h}{8} \\times \\frac{w}{8} \\times 320}\\) respectively. Due to the unavailability of paired textual prompts in most sketch-photo datasets [26, 29, 78], we use null (i.e., “”) prompt in all our experiments." + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.489, + 0.485, + 0.616 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.619, + 0.483, + 0.676 + ], + "angle": 0, + "content": "Figure 2. Given the frozen SD [71] and CLIP [67] models, the proposed method learns the aggregation network, \\(1D\\) convolutional layers, and branch-weights with sketch-photo pairs, via different losses for different downstream tasks (details in Sec. 6)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.686, + 0.483, + 0.84 + ], + "angle": 0, + "content": "CLIP Visual Feature Extraction. Akin to SD, pretrained CLIP [67] encoders have also been used in a plethora of vision tasks like segmentation [60], retrieval [76], object detection [16], video captioning [98], to name a few. In this work, we aim to influence the intermediate feature maps of SD [71] with that from the CLIP visual encoder. Specifically, the input image \\(\\mathbf{i}\\) is passed through the pretrained CLIP visual encoder \\(\\mathbf{V}(\\cdot)\\) to extract the patched feature \\(f_{\\mathbf{v}}\\) from the penultimate layer as: \\(f_{\\mathbf{v}} = \\mathbf{V}(\\mathbf{i}) \\in \\mathbb{R}_{p}^{\\frac{h}{p} \\times \\frac{w}{p} \\times d}\\), where \\(p\\) is the patch-size of the \\(\\mathbf{V}(\\cdot)\\) backbone encoder." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "CLIP Feature Injection. Based on the qualitative and quantitative evidence presented in the pilot study (Sec. 4), we hypothesise that CLIP [67] visual and textual embeddings are intrinsically aligned [67, 79]. Prior works [79]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.212 + ], + "angle": 0, + "content": "have underscored the importance of textual prompts in the quality of extracted features [79]. However, most simple prompts (e.g., \"a photo/sketch of [CLASS]\") are inherently coarse-grained, offering only a high-level description of an image. Such general prompts often lack the semantic detail necessary to fully characterise an image [15, 47], limiting the model's ability to generate highly accurate and contextually-rich feature representations [46]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.213, + 0.905, + 0.304 + ], + "angle": 0, + "content": "To address this, we propose injecting CLIP's visual features into multiple layers of the SD's UNet, thereby influencing the denoising process at various stages (Fig. 2). This approach not only enhances the accuracy of SD features but also enables the model to draw on semantic information from both CLIP's visual and textual embeddings." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.305, + 0.905, + 0.46 + ], + "angle": 0, + "content": "Specifically, after extracting CLIP's visual embeddings \\( f_{\\mathbf{v}} \\in \\mathbb{R}^{\\frac{h}{p} \\times \\frac{w}{p} \\times d} \\), we pass them through simple learnable \\( 1D \\) convolutional layers \\( \\mathcal{C}(\\cdot) \\) to modify them just enough to match the feature dimension of different blocks of the UNet. These transformed feature maps are then added to the intermediate UNet features \\( \\{f_{\\mathbf{u}}^{n}\\}_{n=1}^{4} \\) at every timestep of the denoising process as: \\( \\hat{f}_{\\mathbf{u}}^{n} = f_{\\mathbf{u}}^{n} + \\mathcal{C}(f_{\\mathbf{v}}) \\). This injection occurs in all timesteps and layers of the UNet, enabling the model to influence the denoising process with the semantic information encoded within CLIP's visual embeddings." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.461, + 0.906, + 0.567 + ], + "angle": 0, + "content": "SD Feature Aggregation. Different layers of \\(\\mathbf{U}_{\\theta}\\) capture features of varying semantic-granularity [84], where features from specific layers are ideally suited for certain downstream tasks [46]. Instead of manually selecting each layer for a specific task, we introduce a dynamic approach, allowing the model to automatically extract the most optimal features, thus eliminating the need for manual tuning." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.568, + 0.907, + 0.736 + ], + "angle": 0, + "content": "To achieve this, we train a feature aggregation network comprising three ResNet [32] blocks \\(\\mathbf{A}(\\cdot)\\), which transform the CLIP-enhanced SD features \\(\\{\\hat{f}_{\\mathbf{u}}^{n}\\}_{n=1}^{3}\\) extracted from the first three UNet upsampling layers as: \\(\\bar{f}_{\\mathbf{u}}^{n} = \\mathbf{A}(\\hat{f}_{\\mathbf{u}}^{n}) \\in \\mathbb{R}^{60 \\times 60 \\times d}\\). The final feature map is then obtained by computing a weighted summation of the aggregated features using a learnable weight parameter \\(\\{\\alpha_{n}\\}_{n=1}^{3}\\) for each feature-branch. This aggregation strategy enables the model to dynamically adjust its focus on different layers, adaptively modulating the contribution of each feature-branch based on the semantic-granularity of the task." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.746, + 0.645, + 0.763 + ], + "angle": 0, + "content": "6. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.906, + 0.903 + ], + "angle": 0, + "content": "The proposed feature extractor enriched with the large-scale pretrained knowledge of SD [71] and CLIP [67], enables multiple sketch-based discriminative and dense prediction tasks. Specifically, we showcase its efficacy in tasks like - sketch-based image retrieval [76], sketch recognition [100], sketch-photo correspondence learning [55], and sketch-based image segmentation [35] on different datasets [26, 29, 55, 78] and scenarios. In all cases, we only learn \\(\\mathcal{C}(\\cdot)\\), aggregation network \\(\\mathbf{A}(\\cdot)\\), and branch-weights \\(\\alpha_{n}\\)" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2560" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.411, + 0.106 + ], + "angle": 0, + "content": "keeping pretrained SD and CLIP models frozen." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.111, + 0.414, + 0.126 + ], + "angle": 0, + "content": "6.1. Sketch-based Image Retrieval (SBIR)" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.127, + 0.483, + 0.311 + ], + "angle": 0, + "content": "Problem Statement. Given a query sketch \\(\\mathbf{s} \\in \\mathbb{R}^{h \\times w \\times 3}\\) of any class \\(j\\), category-level SBIR aims to retrieve a photo \\(\\mathbf{p}_i^j\\) from the same class, out of a gallery \\(\\mathcal{G} = \\{\\mathbf{p}_i^j\\}_{i=1}^{N_j}|_{j=1}^{N_c}\\) containing images from a total \\(N_c\\) classes with \\(N_j\\) images per class [102]. Whereas, cross-category fine-grained (FG) SBIR setup intends to learn a unified model capable of instance-level matching from a gallery containing photos from multiple \\((N_c)\\) classes. Conventional SBIR is evaluated on categories \\(\\mathcal{C}^S = \\{c_1^S, \\dots, c_N^S\\}\\) that are seen during training. Contrarily, the zero-shot (ZS) paradigm [76] focuses on evaluating on mutually exclusive (i.e., \\(\\mathcal{C}^S \\cap \\mathcal{C}^U = \\emptyset\\)) unseen categories \\(\\mathcal{C}^U = \\{c_1^U, \\dots, c_M^U\\}\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.312, + 0.484, + 0.497 + ], + "angle": 0, + "content": "Training & Evaluation. Following standard literature [76], we use triplet loss to train our SBIR models. Specifically, given the global max-pooled features from the proposed feature extractor, triplet loss aims to minimise the distance \\(\\delta\\) between an anchor sketch (s) feature \\(f_{\\mathbf{s}} \\in \\mathbb{R}^{d}\\) and a positive photo (p) feature \\(f_{\\mathbf{p}} \\in \\mathbb{R}^{d}\\) from the same category as s, while maximising the same from a negative photo (n) feature \\(f_{\\mathbf{n}} \\in \\mathbb{R}^{d}\\) of a different category. Whereas, in case of cross-category FG-SBIR framework we use hard triplets, where the negative sample is a distinct instance \\((\\mathbf{p}_k^j; k \\neq i)\\) belonging to the same class as the anchor sketch \\((\\mathbf{s}_i^j)\\) and its paired photo \\((\\mathbf{p}_i^j)\\) [46]. With margin \\(\\mu\\), triplet loss becomes:" + }, + { + "type": "equation", + "bbox": [ + 0.099, + 0.508, + 0.483, + 0.526 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {t r i p l e t}} \\left(f _ {\\mathbf {s}}, f _ {\\mathbf {p}}, f _ {\\mathbf {n}}\\right) = \\max \\left\\{0, \\mu + \\delta \\left(f _ {\\mathbf {s}}, f _ {\\mathbf {p}}\\right) - \\delta \\left(f _ {\\mathbf {s}}, f _ {\\mathbf {n}}\\right) \\right\\}. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.537, + 0.483, + 0.627 + ], + "angle": 0, + "content": "For category-level ZS-SBIR, following [46, 76], we consider top \\(k\\) retrieved images to calculate mean average precision \\((\\mathrm{mAP}@\\mathrm{k})\\) and precision \\((\\mathrm{P}@\\mathrm{k})\\). Whereas, for cross-category ZS-FG-SBIR, we measure Acc.@k, which indicates the percentage of sketches having true-matched photos among the top-k retrieved images." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.629, + 0.483, + 0.839 + ], + "angle": 0, + "content": "Dataset & Competitors. Here we use three datasets - (i) Sketchy [78] and Sketchy-extended [53, 78] datasets are used to train and test our cross-category ZS-FG-SBIR and category-level ZS-SBIR models respectively. While Sketchy [78] holds 12,500 photos from 125 classes, each having at least 5 fine-grained paired sketches, its extended version [53] carries additional 60,652 photos from ImageNet [19]. We use photos/sketches from 104 classes for training and 21 for testing [102]. (ii) Tu-Berlin [26] contains 204,489 photos from 250 classes, each having \\(\\sim 80\\) sketches. We use photos/sketches from 220 classes for training and 30 for testing. (iii) Quick, Draw! [29] holds over \\(50M\\) sketches from 345 categories. We use photos/sketches from 80 classes for training and 30 for testing." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "For category-level ZS-SBIR, we compare our method with SoTA methods like SAKE [54], CAAE [102], CCGAN [25], GRL [20], SD-PL [46], etc. Whereas, for cross-category ZS-FG-SBIR, we compare with Gen-VAE [63]," + }, + { + "type": "table", + "bbox": [ + 0.52, + 0.089, + 0.905, + 0.276 + ], + "angle": 0, + "content": "
MethodsSketchy [78]TU-Berlin [26]Quick, Draw! [29]
mAP@200P@200mAP@allP@100mAP@allP@200
B-CLIP0.2500.2610.2280.2570.0800.141
B-DINO0.4930.4810.4500.4920.1670.249
B-DINOV20.5270.5330.4810.5320.1700.268
B-SD0.5580.5710.5100.5610.1790.287
B-SD+CLIP0.5880.5920.5370.5890.1790.311
B-Finetuning0.1200.1720.0110.0100.0030.006
SAKE [54]0.4970.5980.4750.599--
IIAE [37]0.3730.4850.4120.503--
CAAE [102]0.1560.2600.0050.003--
CCGAN [25]--0.2970.426--
CVAE [102]0.2250.3330.0050.0010.0030.003
GRL [20]0.3690.3700.1100.1210.0750.068
LVM [76]0.7230.7250.6510.7320.2020.388
SD-PL [46]0.7460.7470.6800.7440.2310.397
Proposed0.7610.7630.6950.7530.2420.399
" + }, + { + "type": "table_caption", + "bbox": [ + 0.577, + 0.277, + 0.843, + 0.291 + ], + "angle": 0, + "content": "Table 1. Results for category-level ZS-SBIR." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.297, + 0.907, + 0.465 + ], + "angle": 0, + "content": "LVM [76], and \\( SD-PL \\) [46]. Apart from the SoTA methods, we also compare all tasks against a few self-designed Baselines. Among them B-CLIP, B-DINO, B-DINOv2, and B-SD use off-the-shelf pretrained CLIP [67] ViT-L/14, DINO [8] ViT-S/8, DINOv2 [61] ViT-S/14, or SD [71] v2.1 encoders as feature extractors, without task-specific training. We also form another baseline \\( B-SD+CLIP \\) as a straightforward combination of SD and CLIP features without the proposed feature injection and aggregation modules. Finally, B-Finetuning finetunes the entire pretrained SD UNet and CLIP visual encoder on task-specific data." + }, + { + "type": "table", + "bbox": [ + 0.52, + 0.473, + 0.902, + 0.553 + ], + "angle": 0, + "content": "
MethodsAcc.@1Acc.@5MethodsAcc.@1Acc.@5
B-CLIP11.8421.66B-Finetuning5.679.17
B-DINO22.4946.97Gen-VAE [63]22.6049.00
B-DINOv221.1944.31LVM [76]28.6862.34
B-SD23.9849.42SD-PL [46]31.9465.81
B-SD+CLIP24.1652.01Proposed33.0167.92
" + }, + { + "type": "table_caption", + "bbox": [ + 0.515, + 0.554, + 0.904, + 0.568 + ], + "angle": 0, + "content": "Table 2. Results on Sketchy [78] for cross-category ZS-FG-SBIR." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.581, + 0.698, + 0.597 + ], + "angle": 0, + "content": "6.2. Sketch Recognition" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.599, + 0.906, + 0.689 + ], + "angle": 0, + "content": "Problem Statement. Sketch recognition deals with identifying and classifying freehand abstract sketches into predefined categories [68, 96]. Given a dataset \\(\\mathcal{D} = \\{(\\mathbf{s}_i,\\mathbf{y}_i)\\}_{i = 1}^N\\) containing \\(N\\) sketches and their corresponding labels from \\(C\\) classes, the goal is to learn a model that can correctly recognise the class of an input query sketch." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.692, + 0.906, + 0.737 + ], + "angle": 0, + "content": "Training & Evaluation. Given the global max-pooled query sketch feature \\( f_{\\mathbf{s}} \\in \\mathbb{R}^{d} \\), we use standard cross-entropy (CE) loss to train our sketch recognition model as," + }, + { + "type": "equation", + "bbox": [ + 0.587, + 0.742, + 0.907, + 0.78 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {C E}} \\left(\\mathbf {s} _ {i}, \\mathbf {y} _ {i}\\right) = - \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\log p \\left(\\mathbf {y} _ {i} \\mid \\mathbf {s} _ {i}\\right). \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.784, + 0.854, + 0.799 + ], + "angle": 0, + "content": "Following [66, 96, 101] we evaluate using Acc.@k." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.801, + 0.906, + 0.906 + ], + "angle": 0, + "content": "Dataset & Competitors. For this task, we again use the popular TU-Berlin [26] and Quick, Draw! [29] datasets. For Quick, Draw!, we adopt the data split from [29], with each class containing 70,000, 2,500, and 2500 training, validation, and test samples. Apart from the self-designed baselines, we also compare against SoTA methods like SketchMate [96], SketchGNN [101], and SketchXAI [66]." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "2561" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.089, + 0.482, + 0.153 + ], + "angle": 0, + "content": "
MethodsTU-BerlinQuick, Draw!MethodsTU-BerlinQuick, Draw!
B-CLIP30.0930.87B-Finetuning10.2912.37
B-DINO51.7953.28SketchMate [96]77.9679.44
B-DINOv258.0159.12SketchGNN [101]76.4377.31
B-SD61.3763.57SketchXAI [66]-86.10
B-SD+CLIP65.4767.78Proposed84.9687.02
" + }, + { + "type": "table_caption", + "bbox": [ + 0.146, + 0.155, + 0.427, + 0.167 + ], + "angle": 0, + "content": "Table 3. Acc.@1 results for sketch recognition." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.17, + 0.36, + 0.186 + ], + "angle": 0, + "content": "6.3. Sketch-photo Correspondence" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.186, + 0.484, + 0.326 + ], + "angle": 0, + "content": "Problem Statement. Sketch-photo correspondence learning involves identifying a mapping between a freehand sketch of an object and its corresponding photo, such that keypoints in the sketch can be semantically aligned with corresponding keypoints in the photo [55]. Given a sketch-photo pair \\(\\{\\mathbf{s},\\mathbf{p}\\} \\in \\mathbb{R}^{h\\times w\\times 3}\\) and a set of \\(N\\) query keypoints (on sketch) \\(K_{\\mathbf{s}} = \\{x_i^s,y_i^s\\}_{i = 1}^N\\), the aim is to predict a new set of corresponding keypoints \\(\\hat{K}_{\\mathbf{p}} = \\{\\hat{x}_i^{\\mathbf{p}},\\hat{y}_i^{\\mathbf{p}}\\}_{i = 1}^N\\) on the photo, such that \\(\\hat{K}_{\\mathbf{p}}\\) aligns semantically with \\(K_{\\mathbf{s}}\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.326, + 0.484, + 0.477 + ], + "angle": 0, + "content": "Training & Evaluation. Following [106], we use contrastive \\((\\mathcal{L}_{\\mathrm{CL}})\\) [67] and end-point error losses to train our sketch-photo correspondence model. Given the sketch and photo features \\(\\{f_{\\mathrm{s}}, f_{\\mathrm{p}}\\} \\in \\mathbb{R}^{60 \\times 60 \\times d}\\) from the proposed feature extractor, \\(\\mathcal{L}_{\\mathrm{CL}}\\) aims to maximise similarity between local feature patches, while minimising the same between distance ones, for each sketch-photo pairs. Whereas, for endpoint error loss, we first estimate optical flows [106] from the extracted features, followed by calculating \\(l_{2}\\) norm between the estimated \\((\\mathbf{f}_e)\\) and ground truth \\((\\mathbf{f}_g)\\) flow as:" + }, + { + "type": "equation", + "bbox": [ + 0.171, + 0.481, + 0.482, + 0.517 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {E P E}} \\left(\\mathbf {f} _ {e} ^ {i}, \\mathbf {f} _ {g} ^ {i}\\right) = \\frac {1}{M} \\sum_ {i = 1} ^ {M} \\left\\| \\mathbf {f} _ {e} ^ {i} - \\mathbf {f} _ {g} ^ {i} \\right\\| _ {2} ^ {2}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.52, + 0.484, + 0.595 + ], + "angle": 0, + "content": "where, \\(M\\) denotes the patch number (i.e., 60). Finally, the overall loss becomes \\(\\mathcal{L}_{\\mathrm{corr}} = \\mathcal{L}_{\\mathrm{CL}} + \\mathcal{L}_{\\mathrm{EPE}}\\). Following [55] we use PCK@k as a test metric which measures the percentage of keypoints predicted correctly within a distance threshold (i.e., \\(k\\%\\) of the image size)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.595, + 0.484, + 0.717 + ], + "angle": 0, + "content": "Dataset & Competitors. For this task, we use the PSC6K [55] dataset that contains more than \\(150K\\) keypoint annotations for 1250 photos, each having 5 paired sketches (taken from the original Sketchy [78] test split) from 125 classes. We use a 60:40 split for training and testing, where we compare the proposed approach against the self-designed baselines and SoTA methods like WeakAlign[70], WarpC [85], Self-Supervised [55], etc." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.722, + 0.481, + 0.794 + ], + "angle": 0, + "content": "
MethodsPCK@5PCK@10MethodsPCK@5PCK@10
B-CLIP22.2930.41B-Finetuning11.5817.69
B-DINO34.2148.59WeakAlign [70]43.5578.60
B-DINOv239.9156.46WarpC [85]56.7879.70
B-SD41.4258.71Self-Sup. [55]58.0084.93
B-SD+CLIP44.9062.02Proposed70.3189.86
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.796, + 0.482, + 0.825 + ], + "angle": 0, + "content": "Table 4. Results on PSC6K [55] for sketch-photo correspondence. 6.4. Sketch-based Image Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Problem Statement. Given a query sketch \\(\\mathbf{s} \\in \\mathbb{R}^{h \\times w \\times 3}\\) of any category, sketch-based image segmentation [35] aims to predict a binary segmentation mask, \\(\\mathbf{m} \\in \\mathbb{R}^{h \\times w \\times 1}\\), depicting the pixel-locations, wherever the queried-concept appears in any candidate image \\(\\mathbf{p} \\in \\mathbb{R}^{h \\times w \\times 3}\\) of that category," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.122 + ], + "angle": 0, + "content": "from an entire test gallery. The predicted mask marks pixels belonging to that sketched-concept as 1 and the rest as 0." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.122, + 0.906, + 0.288 + ], + "angle": 0, + "content": "Training & Evaluation. We train our segmentation model using binary cross-entropy (BCE) loss, between predicted and ground truth (GT) masks. Given patched photo feature \\( f_{\\mathbf{p}} \\in \\mathbb{R}^{60 \\times 60 \\times d} \\) and global max-pooled sketch feature photo features \\( f_{\\mathbf{s}} \\in \\mathbb{R}^d \\) from the proposed feature extractor, we calculate the cosine similarity between them to generate a patch-wise correlation map \\( \\mathbf{c} = f_{\\mathbf{s}} \\odot f_{\\mathbf{p}} \\in \\mathbb{R}^{60 \\times 60} \\). We upsample \\( \\mathbf{c} \\) to the original spatial resolution (i.e., \\( \\mathbb{R}^{h \\times w} \\)), followed by differentiable Sigmoid thresholding [10] to get the predicted map. BCE loss between the GT and predicted maps can be calculated as," + }, + { + "type": "equation", + "bbox": [ + 0.52, + 0.293, + 0.906, + 0.326 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {B C E}} \\left(y _ {i}, \\hat {y} _ {i}\\right) = - \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\left[ y _ {i} \\log \\left(\\hat {y} _ {i}\\right) + \\left(1 - y _ {i}\\right) \\log \\left(1 - \\hat {y} _ {i}\\right) \\right] \\quad , \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.325, + 0.906, + 0.475 + ], + "angle": 0, + "content": "where, \\(N\\) is the number of total pixels, \\(y_{i}\\) and \\(\\hat{y}_i\\) are the GT and predicted mask pixels. During testing, we calculate similar patch-wise correlation map from sketch/photo features and upsample it to the original spatial image dimension, followed by thresholding it with an empirically chosen fixed value (0.47) and standard CRF [48]-based postprocessing. We use mean Intersection over Union (mIoU) and pixel accuracy (pAcc.) as test metrics, which measure the overlap and percentage of correctly predicted pixels between predicted and GT segmentation masks respectively." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.475, + 0.906, + 0.626 + ], + "angle": 0, + "content": "Dataset & Competitors. Due to the lack of existing paired sketch-photo-mask datasets, we manually annotate photos of 10 chosen classes from the Sketchy [78] dataset with binary mask annotations, resulting in \\(5K\\) paired sketch-photo-mask tuples. We split it into 80:20 for training and testing. Besides baselines and SoTA Sketch-a-Segmenter [35], we also compare with a modified version of ZS-Seg [95], which adapts a simple text-based zero-shot segmentation framework for sketches by replacing its CLIP [67] textual encoder with the visual one to encode input sketches." + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.629, + 0.905, + 0.703 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.705, + 0.905, + 0.747 + ], + "angle": 0, + "content": "Figure 3. Sketch-photo correspondence (left→right : source → target) results on PSC6K. Green circles and squares depict source and GT points respectively, while red squares denote predicted points." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.749, + 0.714, + 0.765 + ], + "angle": 0, + "content": "6.5. Results and Analyses" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Following are the observations from quantitative and qualitative results presented in Tab. 1-5 and Fig. 3-4. (i) Naive adaptation of pretrained backbones (e.g., B-CLIP, B-DINO, B-SD, etc.) fail to beat our method in both ZS-SBIR and ZS-FG-SBIR settings (Tab. 1-2), where we achieve an average 76.15 (54.32)% higher mAP@200 (Acc.@1) over all competitors on Sketchy. We posit that our strategic combination of SD and CLIP brings the \"best of both worlds\", complementing each other's flaws. (ii) While" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2562" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.097, + 0.09, + 0.48, + 0.167 + ], + "angle": 0, + "content": "
MethodsmIoUpAcc.MethodsmIoUpAcc.
B-CLIP20.6326.82B-Finetuning20.3621.84
B-DINO32.8142.06DeepLabv3+Sketch [35]40.6355.23
B-DINOv234.1745.80ZS-Seg [95]44.7360.97
B-SD39.0249.03Sketch-a-Segmenter [35]46.4560.28
B-SD+CLIP41.8751.91Proposed60.1274.91
" + }, + { + "type": "table_caption", + "bbox": [ + 0.125, + 0.169, + 0.448, + 0.182 + ], + "angle": 0, + "content": "Table 5. Results for sketch-based image segmentation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.185, + 0.485, + 0.549 + ], + "angle": 0, + "content": "SoTA sketch-recognition frameworks [96, 101] perform reasonably across different datasets (Tab. 3), our method exceeds them with an average Acc.@1 gain of \\(48.09\\%\\) (Quick, Draw!). SketchXAI [66] despite being our strongest contender, depicts sub-optimal Acc.@1. This gain is likely due to the large-scale pretraining of SD and CLIP, which enables superior open-world object understanding [107]. (iii) For sketch-photo correspondence learning, our method shows an impressive \\(21.22\\%\\) PCK@5 gain over SoTA [55], without the time-consuming two-stage training of [55], or complicated warp-estimation of [85], thanks to SD's innate object-localisation capability. (iv) Sketch-based segmentation being the most challenging task amongst all, baseline methods perform quite poorly (Tab. 5). However, the proposed method surpasses SoTA [35] with a mIoU (pAcc.) margin of \\(29.42(24.27)\\%\\). (v) B-Finetuning fails drastically in almost all tasks (Tab. 1-5). We hypothesise that finetuning entire SD and CLIP models with limited sketch data distorts their large-scale pretrained knowledge [46, 76], thus sacrificing the generalisation potential. (vi) Finally, B-SD+CLIP depicts a \\(\\sim 35.78\\%\\) lesser scores than ours (averaged over all tasks), which again underpins the need for the proposed feature injection and aggregation strategy, rather than a simple combination of pretrained foundation models." + }, + { + "type": "image", + "bbox": [ + 0.103, + 0.553, + 0.482, + 0.637 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.641, + 0.483, + 0.682 + ], + "angle": 0, + "content": "Figure 4. Qualitative results for sketch-based image segmentation. Given a query sketch, our method generates separate segmentation masks for all images of that category. (Zoom-in for the best view.)" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.686, + 0.277, + 0.703 + ], + "angle": 0, + "content": "7. Ablation on Design" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.483, + 0.81 + ], + "angle": 0, + "content": "[i] Effect of aggregation network. To judge its efficacy, instead of aggregation networks \\(\\mathbf{A}(\\cdot)\\), we manually interpolate features from \\(\\{\\mathbf{U}_{\\mathbf{u}}^{n}\\}_{n=1}^{3}\\) into \\(\\mathbb{R}^{60 \\times 60 \\times d}\\) dimension and add them (via \\(\\alpha_{n}\\)) to form the final feature map. A rapid PCK@5 (mIoU) drop of 28.65 (26.51)% in case of w/o Aggregation Network (Tab. 6) indicates its significance in extracting semantically-meaningful features." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.484, + 0.902 + ], + "angle": 0, + "content": "[i] Are learnable branch weights necessary? Learnable weights dynamically ensure the optimum importance of each feature-branch for a specific task. Although less pronounced in PCK@5, mAP@200 drops significantly (4.82%) in case of w/o Learnable Weights (Tab. 6), depicting its influence on the final feature map quality." + }, + { + "type": "table", + "bbox": [ + 0.52, + 0.09, + 0.903, + 0.208 + ], + "angle": 0, + "content": "
MethodsCorrespondenceSegmentationZS-SBIRRecognition
PCK@5mIoUmAP@200Acc.@1
SD [71]v1.468.9157.630.72878.82
v1.569.2458.910.73579.59
v2.170.3160.120.76187.02
CLIP [67]ViT-B/1667.7556.380.74377.37
ViT-B/3269.4158.470.75878.84
ViT-L/1470.3160.120.76187.02
w/o Aggregation Network54.6547.520.58766.71
w/o Learnable Weights68.7558.130.72683.79
w/o 1D Convolutions55.9747.240.60271.75
Ours-full70.3160.120.76187.02
" + }, + { + "type": "table_caption", + "bbox": [ + 0.623, + 0.209, + 0.796, + 0.222 + ], + "angle": 0, + "content": "Table 6. Ablation on design." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.224, + 0.907, + 0.527 + ], + "angle": 0, + "content": "[iii] Contribution of \\(1D\\) Convolutions. To assess the effectiveness of \\(1D\\) convolution layers, we use simple linear interpolation to match CLIP and SD UNet feature dimensions, before injection. Sharp PCK@5 drop of \\(25.62\\%\\) for w/o \\(1D\\) Convolutions (Tab. 6) reveals its impact on influencing the diffusion denoising process. We posit that the learnable \\(1D\\) convolutions not only match the feature dimension but also enrich the CLIP features before injection. [iv] Model variants. The performance of our proposed feature extractor depends crucially on the choice of backbone model variant. Accordingly, we explore with multiple pretrained SD and CLIP backbones. Tab. 6 shows SD v2.1 and CLIP ViT-L/14 to achieve the highest PCK@5 and Acc.@1 for sketch-photo correspondence and recognition respectively. Additionally, we also experiment by replacing the CLIP visual encoder of our method with a DINOv2 [61] ViT-S/14 encoder for the task of ZS-SBIR, to observe a \\(32.80\\%\\) drop in mAP@200. This is probably due to CLIP's [67] multimodal pertaining (diminishing sketch-photo domain gap), over DINOv2's [61] self-supervised one." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.528, + 0.906, + 0.572 + ], + "angle": 0, + "content": "[v] Which timestep value is the most effective? Different diffusion timesteps yield features with drastically different quality [84, 107]. Calculating PCK@5 and mAP@200" + }, + { + "type": "image", + "bbox": [ + 0.514, + 0.573, + 0.733, + 0.679 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.743, + 0.573, + 0.907, + 0.693 + ], + "angle": 0, + "content": "for the tasks of correspondence learning and ZS-SBIR respectively, over varying \\( t \\), we find that \\( t = 195 \\) produces the best results for both tasks. Fig. 5 shows our method to be robust to" + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.68, + 0.906, + 0.724 + ], + "angle": 0, + "content": "Figure 5. Choice of timestep \\((t)\\). method to be robust to the choice of \\( t \\), where PCK@5 (mAP@200) scores from a wide range of \\( t \\) values consistently surpass the baselines." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.731, + 0.634, + 0.746 + ], + "angle": 0, + "content": "8. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.75, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In this paper, we identified SD's limitations in understanding abstract, sparse sketches. Integrating CLIP's semantic features with SD's spatially-aware features, we developed a framework that dynamically combines these strengths. This significantly improves performance across sketch recognition, SBIR, and sketch-based segmentation. Our method highlights the complementary nature of different foundation models and introduces an adaptive feature aggregation network, establishing a new standard for universal sketch feature representation for future research." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2563" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.115, + 0.484, + 0.169 + ], + "angle": 0, + "content": "[1] Hyojin Bahng, Ali Jahanian, Swami Sankaranarayanan, and Phillip Isola. Exploring visual prompts for adapting large-scale models. arXiv preprint arXiv:2203.17274, 2022.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.172, + 0.483, + 0.24 + ], + "angle": 0, + "content": "[2] Hmrishav Bandyopadhyay, Subhadeep Koley, Ayan Das, Ayan Kumar Bhunia, Aneeshan Sain, Pinaki Nath Chowdhury, Tao Xiang, and Yi-Zhe Song. Doodle Your 3D: From Abstract Freehand Sketches to Precise 3D Shapes. In CVPR, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.242, + 0.483, + 0.296 + ], + "angle": 0, + "content": "[3] Dmitry Baranchuk, Ivan Rubachev, Andrey Voynov, Valentin Khrulkov, and Artem Babenko. Label-Efficient Semantic Segmentation with Diffusion Models. In ICLR, 2021. 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.298, + 0.482, + 0.354 + ], + "angle": 0, + "content": "[4] Ayan Kumar Bhunia, Pinaki Nath Chowdhury, Yongxin Yang, Timothy Hospedales, Tao Xiang, and Yi-Zhe Song. Vectorization and Rasterization: Self-Supervised Learning for Sketch and Handwriting. In CVPR, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.355, + 0.482, + 0.41 + ], + "angle": 0, + "content": "[5] Ayan Kumar Bhunia, Viswanatha Reddy Gajjala, Subhadeep Koley, Rohit Kundu, Aneeshan Sain, Tao Xiang, and Yi-Zhe Song. Doodle It Yourself: Class Incremental Learning by Drawing a Few Sketches. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.411, + 0.483, + 0.478 + ], + "angle": 0, + "content": "[6] Ayan Kumar Bhunia, Subhadeep Koley, Abdullah Faiz Ur Rahman Khilji, Aneeshan Sain, Pinaki Nath Chowdhury, Tao Xiang, and Yi-Zhe Song. Sketching Without Worrying: Noise-Tolerant Sketch-Based Image Retrieval. In CVPR, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.481, + 0.483, + 0.535 + ], + "angle": 0, + "content": "[7] Ayan Kumar Bhunia, Subhadeep Koley, Amandeep Kumar, Aneeshan Sain, Pinaki Nath Chowdhury, Tao Xiang, and Yi-Zhe Song. Sketch2Saliency: Learning to Detect Salient Objects from Human Drawings. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.537, + 0.483, + 0.59 + ], + "angle": 0, + "content": "[8] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging Properties in Self-Supervised Vision Transformers. In ICCV, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.593, + 0.483, + 0.66 + ], + "angle": 0, + "content": "[9] Dar-Yen Chen, Subhadeep Koley, Aneeshan Sain, Pinaki Nath Chowdhury, Tao Xiang, Ayan Kumar Bhunia, and Yi-Zhe Song. DemoCaricature: Democratising Caricature Generation with a Rough Sketch. In CVPR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.663, + 0.482, + 0.704 + ], + "angle": 0, + "content": "[10] Honglie Chen, Weidi Xie, Triantafyllos Afouras, Arsha Nagrani, Andrea Vedaldi, and Andrew Zisserman. Localizing Visual Sounds the Hard Way. In CVPR, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.705, + 0.482, + 0.759 + ], + "angle": 0, + "content": "[11] Haoxin Chen, Yong Zhang, Xiaodong Cun, Menghan Xia, Xintao Wang, Chao Weng, and Ying Shan. VideoCraft2: Overcoming Data Limitations for High-Quality Video Diffusion Models. In CVPR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.761, + 0.482, + 0.802 + ], + "angle": 0, + "content": "[12] Gene Chou, Yuval Bahat, and Felix Heide. Diffusion-SDF: Conditional Generative Modeling of Signed Distance Functions. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.804, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[13] Pinaki Nath Chowdhury, Ayan Kumar Bhunia, Viswanatha Reddy Gajjala, Aneeshan Sain, Tao Xiang, and Yi-Zhe Song. Partially Does It: Towards Scene-Level FG-SBIR with Partial Input. In CVPR, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.873, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[14] Pinaki Nath Chowdhury, Tuanfeng Wang, Duygu Ceylan, Yi-Zhe Song, and Yulia Gryaditskaya. Garment ideation:" + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.554, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "Iterative view-aware sketch-based garment modeling. In 3DV, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.123, + 0.905, + 0.178 + ], + "angle": 0, + "content": "[15] Pinaki Nath Chowdhury, Ayan Kumar Bhunia, Aneshan Sain, Subhadeep Koley, Tao Xiang, and Yi-Zhe Song. SceneTrilogy: On Human Scene-Sketch and its Complementarity with Photo and Text. In CVPR, 2023. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.18, + 0.905, + 0.235 + ], + "angle": 0, + "content": "[16] Pinaki Nath Chowdhury, Ayan Kumar Bhunia, Aneshan Sain, Subhadeep Koley, Tao Xiang, and Yi-Zhe Song. What Can Human Sketches Do for Object Detection? In CVPR, 2023. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.238, + 0.905, + 0.278 + ], + "angle": 0, + "content": "[17] John P. Collomosse, Tu Bui, and Hailin Jin. LiveSketch: Query Perturbations for Guided Sketch-based Visual Search. In CVPR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.281, + 0.905, + 0.335 + ], + "angle": 0, + "content": "[18] Bram de Wilde, Anindo Saha, Richard PG ten Broek, and Henkjan Huisman. Medical diffusion on a budget: Textual Inversion for medical image generation. In MIDL, 2024. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.338, + 0.905, + 0.379 + ], + "angle": 0, + "content": "[19] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. ImageNet: A Large-Scale Hierarchical Image Database. In CVPR, 2009. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.382, + 0.904, + 0.423 + ], + "angle": 0, + "content": "[20] Sounak Dey, Pau Riba, Anjan Dutta, Josep Llados, and Yi-Zhe Song. Doodle to Search: Practical Zero-Shot Sketch-based Image Retrieval. In CVPR, 2019. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.426, + 0.904, + 0.453 + ], + "angle": 0, + "content": "[21] Prafulla Dhariwal and Alexander Nichol. Diffusion Models Beat GANs on Image Synthesis. In NeurIPS, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.455, + 0.904, + 0.511 + ], + "angle": 0, + "content": "[22] Francesco Di Sario, Riccardo Renzulli, Enzo Tartaglione, and Marco Grangetto. Boost Your NeRF: A Model-Agnostic Mixture of Experts Framework for High Quality and Efficient Rendering. In ECCV, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.513, + 0.905, + 0.567 + ], + "angle": 0, + "content": "[23] Shiyin Dong, Mingrui Zhu, Kun Cheng, Nannan Wang, and Xinbo Gao. Bridging generative and discriminative models for unified visual perception with diffusion priors. In IJCAI, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.571, + 0.905, + 0.652 + ], + "angle": 0, + "content": "[24] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.656, + 0.905, + 0.696 + ], + "angle": 0, + "content": "[25] Anjan Dutta and Zeynep Akata. Semantically Tied Paired Cycle Consistency for Zero-Shot Sketch-based Image Retrieval. In CVPR, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.699, + 0.905, + 0.726 + ], + "angle": 0, + "content": "[26] Mathias Eitz, James Hays, and Marc Alexa. How do humans sketch objects? ACM TOG, 2012. 1, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.729, + 0.905, + 0.783 + ], + "angle": 0, + "content": "[27] Arnab Ghosh, Richard Zhang, Puneet K Dokania, Oliver Wang, Alexei A Efros, Philip HS Torr, and Eli Shechtman. Interactive Sketch & Fill: Multiclass Sketch-to-Image Translation. In ICCV, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.786, + 0.905, + 0.814 + ], + "angle": 0, + "content": "[28] Rafael C Gonzalez and Richard E Woods. Digital Image Processing. Pearson, 2008. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.816, + 0.905, + 0.843 + ], + "angle": 0, + "content": "[29] David Ha and Douglas Eck. A Neural Representation of Sketch Drawings. In ICLR, 2017. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.846, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[30] Moayed Haji-Ali, Guha Balakrishnan, and Vicente Ordonez. ElasticDiffusion: Training-free Arbitrary Size Image Generation through Global-Local Content Separation. In CVPR, 2024. 3" + }, + { + "type": "list", + "bbox": [ + 0.524, + 0.093, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2564" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.093, + 0.482, + 0.146 + ], + "angle": 0, + "content": "[31] Cusuh Ham, Gemma Canet Tarres, Tu Bui, James Hays, Zhe Lin, and John Collomosse. CoGS: Controllable Generation and Search from Sketch and Style. In ECCV, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.149, + 0.482, + 0.19 + ], + "angle": 0, + "content": "[32] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep Residual Learning for Image Recognition. In CVPR, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.193, + 0.482, + 0.245 + ], + "angle": 0, + "content": "[33] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-Prompt Image Editing with Cross Attention Control. In ICLR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.25, + 0.482, + 0.275 + ], + "angle": 0, + "content": "[34] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising Diffusion Probabilistic Models. In NeurIPS, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.279, + 0.482, + 0.332 + ], + "angle": 0, + "content": "[35] Conghui Hu, Da Li, Yongxin Yang, Timothy M Hospedales, and Yi-Zhe Song. Sketch-a-Segmented: Sketch-based Photo Segmenter Generation. IEEE TIP, 2020. 2, 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.335, + 0.482, + 0.403 + ], + "angle": 0, + "content": "[36] Drew A Hudson, Daniel Zoran, Mateusz Malinowski, Andrew K Lampinen, Andrew Jaegle, James L McClelland, Loic Matthey, Felix Hill, and Alexander Lerchner. SODA: Bottleneck Diffusion Models for Representation Learning. In CVPR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.406, + 0.482, + 0.46 + ], + "angle": 0, + "content": "[37] HyeongJoo Hwang, Geon-Hyeong Kim, Seunghoon Hong, and Kee-Eung Kim. Variational Interaction Information Maximization for Cross-domain Disentanglement. In NeurIPS, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.463, + 0.482, + 0.504 + ], + "angle": 0, + "content": "[38] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In ECCV, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.506, + 0.482, + 0.546 + ], + "angle": 0, + "content": "[39] Tero Karras, Samuli Laine, and Timo Aila. A Style-Based Generator Architecture for Generative Adversarial Networks. In CVPR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.549, + 0.482, + 0.602 + ], + "angle": 0, + "content": "[40] Bahjat Kawar, Shiran Zada, Oran Lang, Omer Tov, Huiwen Chang, Tali Dekel, Inbar Mosseri, and Michal Irani. Imagic: Text-Based Real Image Editing with Diffusion Models. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.606, + 0.482, + 0.701 + ], + "angle": 0, + "content": "[41] Kazuma Kobayashi and Lin Gu and Ryuichiro Hataya and Takaaki Mizuno and Mototaka Miyake and Hirokazu Watanabe and Masamichi Takahashi and Yasuyuki Takamizawa and Yukihiro Yoshida and Satoshi Nakamura and Nobuji Kouno and Amina Bolatkan and Yusuke Kurose and Tatsuya Harada and Ryuji Hamamoto. Sketch-based Medical Image Retrieval. Medical Image Analysis, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.704, + 0.482, + 0.731 + ], + "angle": 0, + "content": "[42] Diederik P Kingma and Max Welling. Auto-Encoding Variational Bayes. In ICLR, 2014. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.733, + 0.482, + 0.786 + ], + "angle": 0, + "content": "[43] Subhadeep Koley, Ayan Kumar Bhunia, Aneeshan Sain, Pinaki Nath Chowdhury, Tao Xiang, and Yi-Zhe Song. Picture that Sketch: Photorealistic Image Generation from Abstract Sketches. In CVPR, 2023. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.79, + 0.482, + 0.843 + ], + "angle": 0, + "content": "[44] Subhadeep Koley, Ayan Kumar Bhunia, Aneeshan Sain, Pinaki Nath Chowdhury, Tao Xiang, and Yi-Zhe Song. You'll Never Walk Alone: A Sketch and Text Duet for Fine-Grained Image Retrieval. In CVPR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.847, + 0.482, + 0.899 + ], + "angle": 0, + "content": "[45] Subhadeep Koley, Ayan Kumar Bhunia, Aneeshan Sain, Pinaki Nath Chowdhury, Tao Xiang, and Yi-Zhe Song. How to Handle Sketch-Abstraction in Sketch-Based Image Retrieval? In CVPR, 2024. 1, 3" + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.093, + 0.482, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[46] Subhadeep Koley, Ayan Kumar Bhunia, Aneeshan Sain, Pinaki Nath Chowdhury, Tao Xiang, and Yi-Zhe Song. Text-to-Image Diffusion Models are Great Sketch-Photo Matchmakers. In CVPR, 2024. 1, 2, 3, 4, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.149, + 0.905, + 0.216 + ], + "angle": 0, + "content": "[47] Subhadeep Koley, Ayan Kumar Bhunia, Deeptanshu Sekhri, Aneeshan Sain, Pinaki Nath Chowdhury, Tao Xiang, and Yi-Zhe Song. It's All About Your Sketch: Democratising Sketch Control in Diffusion Models. In CVPR, 2024. 1, 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.219, + 0.905, + 0.258 + ], + "angle": 0, + "content": "[48] Philipp Krahenbuhl and Vladlen Koltun. Efficient Inference in Fully Connected CRFs with Gaussian Edge Potentials. In NeurIPS, 2011. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.261, + 0.905, + 0.3 + ], + "angle": 0, + "content": "[49] Alexander C Li, Mihir Prabhudesai, Shivam Duggal, Ellis Brown, and Deepak Pathak. Your Diffusion Model is Secretly a Zero-Shot Classifier. In ICCV, 2023. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.302, + 0.904, + 0.328 + ], + "angle": 0, + "content": "[50] Ke Li, Kaiyue Pang, and Yi-Zhe Song. Photo Pre-Training, but for Sketch. In CVPR, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.331, + 0.905, + 0.37 + ], + "angle": 0, + "content": "[51] Minchen Li, Alla Sheffer, Eitan Grinspun, and Nicholas Vining. FoldSketch: Enriching Garments with Physically Reproducible Folds. ACM TOG, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.373, + 0.905, + 0.425 + ], + "angle": 0, + "content": "[52] Fengyin Lin, Mingkang Li, Da Li, Timothy Hospedales, YiZhe Song, and Yonggang Qi. Zero-Shot Everything Sketch-Based Image Retrieval, and in Explainable Style. In CVPR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.428, + 0.905, + 0.468 + ], + "angle": 0, + "content": "[53] Li Liu, Fumin Shen, Yuming Shen, Xianglong Liu, and Ling Shao. Deep Sketch Hashing: Fast Free-hand Sketch-Based Image Retrieval. In CVPR, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.471, + 0.905, + 0.51 + ], + "angle": 0, + "content": "[54] Qing Liu, Lingxi Xie, Huiyu Wang, and Alan Yuille. Semantic-Aware Knowledge Preservation for Zero-Shot Sketch-Based Image Retrieval. In ICCV, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.512, + 0.905, + 0.551 + ], + "angle": 0, + "content": "[55] Xuanchen Lu, Xiaolong Wang, and Judith E Fan. Learning Dense Correspondences between Photos and Sketches. In ICML, 2023. 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.554, + 0.905, + 0.593 + ], + "angle": 0, + "content": "[56] Ling Luo, Yulia Gryaditskaya, Tao Xiang, and Yi-Zhe Song. Structure-Aware 3D VR Sketch to 3D Shape Retrieval. In 3DV, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.596, + 0.905, + 0.636 + ], + "angle": 0, + "content": "[57] Ling Luo, Pinaki Nath Chowdhury, Tao Xiang, Yi-Zhe Song, and Yulia Gryaditskaya. 3D VR Sketch Guided 3D Shape Prototyping and Exploration. In ICCV, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.638, + 0.905, + 0.677 + ], + "angle": 0, + "content": "[58] Aryan Mikaeili, Or Perel, Daniel Cohen-Or, and Ali Mahdavi-Amiri. SKED: Sketch-guided Text-based 3D Editing. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.679, + 0.905, + 0.733 + ], + "angle": 0, + "content": "[59] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, and Ying Shan. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. In AAAI, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.735, + 0.905, + 0.789 + ], + "angle": 0, + "content": "[60] Jishnu Mukhoti, Tsung-Yu Lin, Omid Poursaeed, Rui Wang, Ashish Shah, Philip HS Torr, and Ser-Nam Lim. Open Vocabulary Semantic Segmentation with Patch Aligned Contrastive Learning. In CVPR, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.791, + 0.905, + 0.858 + ], + "angle": 0, + "content": "[61] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. DINov2: Learning Robust Visual Features without Supervision. TMLR, 2024. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.86, + 0.905, + 0.899 + ], + "angle": 0, + "content": "[62] Zixuan Pan, Jianxu Chen, and Yiyu Shi. Masked Diffusion as Self-supervised Representation Learner. arXiv preprint arXiv:2308.05695, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.525, + 0.093, + 0.905, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.515, + 0.957 + ], + "angle": 0, + "content": "2565" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.093, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[63] Kaiyue Pang, Ke Li, Yongxin Yang, Honggang Zhang, Timothy M Hospedales, Tao Xiang, and Yi-Zhe Song. Generalising Fine-Grained Sketch-Based Image Retrieval. In CVPR, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.149, + 0.482, + 0.203 + ], + "angle": 0, + "content": "[64] Kaiyue Pang, Yongxin Yang, Timothy M Hospedales, Tao Xiang, and Yi-Zhe Song. Solving mixed-modal jigsaw puzzle for fine-grained sketch-based image retrieval. In CVPR, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.205, + 0.483, + 0.259 + ], + "angle": 0, + "content": "[65] Mihir Prabhudesai, Tsung-Wei Ke, Alexander Cong Li, Deepak Pathak, and Katerina Fragkiadaki. Diffusion-TTA: Test-time Adaptation of Discriminative Models via Generative Feedback. In NeurIPS, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.261, + 0.482, + 0.302 + ], + "angle": 0, + "content": "[66] Zhiyu Qu, Yulia Gryaditskaya, Ke Li, Kaiyue Pang, Tao Xiang, and Yi-Zhe Song. SketchXAI: A First Look at Explainability for Human Sketches. In CVPR, 2023. 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.303, + 0.482, + 0.37 + ], + "angle": 0, + "content": "[67] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning Transferable Visual Models From Natural Language Supervision. In ICML, 2021. 1, 2, 3, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.373, + 0.482, + 0.412 + ], + "angle": 0, + "content": "[68] Leo Sampaio Ferraz Ribeiro, Tu Bui, John Collomosse, and Moacir Ponti. Sketchformer: Transformer-based Representation for Sketched Structure. In CVPR, 2020. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.414, + 0.482, + 0.468 + ], + "angle": 0, + "content": "[69] Elad Richardson, Yuval Alaluf, Or Patashnik, Yotam Nitzan, Yaniv Azar, Stav Shapiro, and Daniel Cohen-Or. Encoding in Style: a StyleGAN Encoder for Image-to-Image Translation. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.47, + 0.482, + 0.509 + ], + "angle": 0, + "content": "[70] Ignacio Rocco, Relja Arandjelović, and Josef Sivic. End-to-end weakly-supervised semantic alignment. In CVPR, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.512, + 0.482, + 0.565 + ], + "angle": 0, + "content": "[71] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-Resolution Image Synthesis with Latent Diffusion Models. In CVPR, 2022. 1, 2, 3, 4, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.568, + 0.482, + 0.607 + ], + "angle": 0, + "content": "[72] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-Net: Convolutional Networks for Biomedical Image Segmentation. In MICCAI, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.61, + 0.482, + 0.663 + ], + "angle": 0, + "content": "[73] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. DreamBooth: Fine Tuning Text-to-Image Diffusion Models for Subject-Driven Generation. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.665, + 0.482, + 0.718 + ], + "angle": 0, + "content": "[74] Aneeshan Sain, Ayan Kumar Bhunia, Yongxin Yang, Tao Xiang, and Yi-Zhe Song. StyleMeUp: Towards Style-Agnostic Sketch-Based Image Retrieval. In CVPR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.721, + 0.482, + 0.774 + ], + "angle": 0, + "content": "[75] Aneeshan Sain, Ayan Kumar Bhunia, Vaishnav Potlapalli, Pinaki Nath Chowdhury, Tao Xiang, and Yi-Zhe Song. Sketch3T: Test-Time Training for Zero-Shot SBIR. In CVPR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.777, + 0.482, + 0.844 + ], + "angle": 0, + "content": "[76] Aneeshan Sain, Ayan Kumar Bhunia, Pinaki Nath Chowdhury, Aneeshan Sain, Subhadeep Koley, Tao Xiang, and Yi-Zhe Song. CLIP for All Things Zero-Shot Sketch-Based Image Retrieval, Fine-Grained or Not. In CVPR, 2023. 1, 2, 3, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.846, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[77] Aneeshan Sain, Ayan Kumar Bhunia, Subhadeep Koley, Pinaki Nath Chowdhury, Soumitri Chattopadhyay, Tao Xiang, and Yi-Zhe Song. Exploiting Unlabelled Photos for Stronger Fine-Grained SBIR. In CVPR, 2023. 3" + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.093, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "[78] Patsorn Sangkloy, Nathan Burnell, Cusuh Ham, and James Hays. The Sketchy Database: Learning to Retrieve Badly Drawn Bunnies. ACM TOG, 2016. 1, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.137, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[79] Patsorn Sangkloy, Wittawat Jitkrittum, Diyi Yang, and James Hays. A Sketch Is Worth a Thousand Words: Image Retrieval with Text and Sketch. In ECCV, 2022. 2, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.193, + 0.905, + 0.274 + ], + "angle": 0, + "content": "[80] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. LAION-5B: An open large-scale dataset for training next generation image-text models. In NeurlPSW, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.277, + 0.904, + 0.317 + ], + "angle": 0, + "content": "[81] Jaehyeok Shim, Changwoo Kang, and Kyungdon Joo. Diffusion-Based Signed Distance Fields for 3D Shape Generation. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.32, + 0.904, + 0.346 + ], + "angle": 0, + "content": "[82] Chenyang Si, Ziqi Huang, Yuming Jiang, and Ziwei Liu. FreeU: Free Lunch in Diffusion U-Net. In CVPR, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.349, + 0.905, + 0.402 + ], + "angle": 0, + "content": "[83] Jifei Song, Qian Yu, Yi-Zhe Song, Tao Xiang, and Timothy M Hospedales. Deep Spatial-Semantic Attention for Fine-Grained Sketch-Based Image Retrieval. In ICCV, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.405, + 0.905, + 0.445 + ], + "angle": 0, + "content": "[84] Luming Tang, Menglin Jia, Qianqian Wang, Cheng Perng Phoo, and Bharath Hariharan. Emergent Correspondence from Image Diffusion. In NeurIPS, 2023. 1, 2, 4, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.448, + 0.905, + 0.488 + ], + "angle": 0, + "content": "[85] Prune Truong, Martin Danelljan, Fisher Yu, and Luc Van Gool. Warp Consistency for Unsupervised Learning of Dense Correspondences. In ICCV, 2021. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.491, + 0.905, + 0.531 + ], + "angle": 0, + "content": "[86] Narek Tumanyan, Michal Geyer, Shai Bagon, and Tali Dekel. Plug-and-Play Diffusion Features for Text-Driven Image-to-Image Translation. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.533, + 0.905, + 0.573 + ], + "angle": 0, + "content": "[87] Andrey Voynov, Kfir Aberman, and Daniel Cohen-Or. Sketch-guided text-to-image diffusion models. In SIGGRAPH, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.576, + 0.904, + 0.603 + ], + "angle": 0, + "content": "[88] Fei Wang. Sketch2Vox: Learning 3D Reconstruction from a Single Monocular Sketch Image. In ECCV, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.605, + 0.905, + 0.645 + ], + "angle": 0, + "content": "[89] Wenxuan Wang, Quan Sun, Fan Zhang, Yepeng Tang, Jing Liu, and Xinlong Wang. Diffusion Feedback Helps CLIP See Better. In ICLR, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.648, + 0.905, + 0.701 + ], + "angle": 0, + "content": "[90] Size Wu, Wenwei Zhang, Lumin Xu, Sheng Jin, Xiangtai Li, Wentao Liu, and Chen Change Loy. CLIPSelf: Vision Transformer Distills Itself for Open-Vocabulary Dense Prediction. In ICLR, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.704, + 0.905, + 0.758 + ], + "angle": 0, + "content": "[91] Chunlong Xia, Xinliang Wang, Feng Lv, Xin Hao, and Yifeng Shi. ViT-CoMer: Vision Transformer with Convolutional Multi-scale Feature Interaction for Dense Predictions. In CVPR, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.761, + 0.905, + 0.801 + ], + "angle": 0, + "content": "[92] Jinbo Xing, Hanyuan Liu, Menghan Xia, Yong Zhang, Xintao Wang, Ying Shan, and Tien-Tsin Wong. ToonCrafter: Generative Cartoon Interpolation. ACM TOG, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.804, + 0.905, + 0.844 + ], + "angle": 0, + "content": "[93] Chenfeng Xu, Huan Ling, Sanja Fidler, and Or Litany. 3DiffTec: 3D Object Detection with Geometry-Aware Diffusion Features. In CVPR, 2023. 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.846, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[94] Jiarui Xu, Sifei Liu, Arash Vahdat, Wonmin Byeon, Xiaolong Wang, and Shalini De Mello. Open-Vocabulary Panoptic Segmentation with Text-to-Image Diffusion Models. In CVPR, 2023. 1, 2, 4" + }, + { + "type": "list", + "bbox": [ + 0.525, + 0.093, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "2566" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.092, + 0.486, + 0.148 + ], + "angle": 0, + "content": "[95] Mengde Xu, Zheng Zhang, Fangyun Wei, Yutong Lin, Yue Cao, Han Hu, and Xiang Bai. A Simple Baseline for Open-Vocabulary Semantic Segmentation with Pre-trained Vision-language Model. In ECCV, 2022. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.151, + 0.483, + 0.207 + ], + "angle": 0, + "content": "[96] Peng Xu, Yongye Huang, Tongtong Yuan, Kaiyue Pang, YiZhe Song, Tao Xiang, Timothy M Hospedales, Zhanyu Ma, and Jun Guo. SketchMate: Deep Hashing for Million-Scale Human Sketch Retrieval. In CVPR, 2018. 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.209, + 0.483, + 0.25 + ], + "angle": 0, + "content": "[97] Peng Xu, Timothy Hospedales, Qiyue Yin, Yi-Zhe Song, Tao Xiang, and Liang Wang. Deep Learning for Free-Hand Sketch: A Survey. IEEE TPAMI, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.253, + 0.483, + 0.32 + ], + "angle": 0, + "content": "[98] Antoine Yang, Arsha Nagrani, Paul Hongsuck Seo, Antoine Miech, Jordi Pont-Tuset, Ivan Laptev, Josef Sivic, and Cordelia Schmid. Vid2Seq: Large-Scale Pretraining of a Visual Language Model for Dense Video Captioning. In CVPR, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.324, + 0.483, + 0.378 + ], + "angle": 0, + "content": "[99] Kai Yang, Jian Tao, Jiafei Lyu, Chunjiang Ge, Jiaxin Chen, Weihan Shen, Xiaolong Zhu, and Xiu Li. Using human feedback to fine-tune diffusion models without any reward model. In CVPR, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.382, + 0.482, + 0.422 + ], + "angle": 0, + "content": "[100] Lan Yang, Kaiyue Pang, Honggang Zhang, and Yi-Zhe Song. SketchAA: Abstract Representation for Abstract Sketches. In ICCV, 2021. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.426, + 0.483, + 0.48 + ], + "angle": 0, + "content": "[101] Lumin Yang, Jiajie Zhuang, Hongbo Fu, Xiangzhi Wei, Kun Zhou, and Youyi Zheng. SketchGNN: Semantic Sketch Segmentation with Graph Neural Networks. ACM TOG, 2021. 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.483, + 0.483, + 0.525 + ], + "angle": 0, + "content": "[102] Sasi Kiran Yelamarthi, Shiva Krishna Reddy, Ashish Mishra, and Anurag Mittal. A Zero-Shot Framework for Sketch Based Image Retrieval. In ECCV, 2018. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.528, + 0.483, + 0.569 + ], + "angle": 0, + "content": "[103] Dong Yin, Raphael Gontijo Lopes, Jon Shlens, Ekin Dogus Cubuk, and Justin Gilmer. A Fourier Perspective on Model Robustness in Computer Vision. In NeurIPS, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.571, + 0.483, + 0.625 + ], + "angle": 0, + "content": "[104] Jiazuo Yu, Yunzhi Zhuge, Lu Zhang, Ping Hu, Dong Wang, Huchuan Lu, and You He. Boosting Continual Learning of Vision-Language Models via Mixture-of-Experts Adapters. In CVPR, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.629, + 0.483, + 0.669 + ], + "angle": 0, + "content": "[105] Seonghoon Yu, Paul Hongsuck Seo, and Jeany Son. Zeroshot Referring Image Segmentation with Global-Local Context Features. In CVPR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.673, + 0.483, + 0.728 + ], + "angle": 0, + "content": "[106] Junyi Zhang, Charles Herrmann, Junhwa Hur, Eric Chen, Varun Jampani, Deqing Sun, and Ming-Hsuan Yang. Telling left from right: Identifying geometry-aware semantic correspondence. In CVPR, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.731, + 0.483, + 0.799 + ], + "angle": 0, + "content": "[107] Junyi Zhang, Charles Herrmann, Junhwa Hur, Luisa Polania Cabrera, Varun Jampani, Deqing Sun, and Ming-Hsuan Yang. A Tale of Two Features: Stable Diffusion Complements DINO for Zero-Shot Semantic Correspondence. NeurIPS, 2024. 1, 2, 3, 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.802, + 0.483, + 0.842 + ], + "angle": 0, + "content": "[108] Lvmin Zhang and Maneesh Agrawala. Adding Conditional Control to Text-to-Image Diffusion Models. In ICCV, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.483, + 0.9 + ], + "angle": 0, + "content": "[109] Manyuan Zhang, Guanglu Song, Xiaoyu Shi, Yu Liu, and Hongsheng Li. Three Things We Need to Know About Transferring Stable Diffusion to Visual Dense Prediction Tasks. In ECCV, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.092, + 0.486, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.135 + ], + "angle": 0, + "content": "[110] Wenliang Zhao, Yongming Rao, Zuyan Liu, Benlin Liu, Jie Zhou, and Jiwen Lu. Unleashing Text-to-Image Diffusion Models for Visual Perception. In ICCV, 2023. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.137, + 0.905, + 0.176 + ], + "angle": 0, + "content": "[111] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Zwei Liu. Conditional Prompt Learning for Vision-Language Models. In CVPR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.905, + 0.219 + ], + "angle": 0, + "content": "[112] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Zwei Liu. Learning to Prompt for Vision-Language Models. IJCV, 2022. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2567" + } + ] +] \ No newline at end of file diff --git a/2025/SketchFusion_ Learning Universal Sketch Features through Fusing Foundation Models/1fc738f6-a4cf-4a97-bcd3-ea7682cab3c6_origin.pdf b/2025/SketchFusion_ Learning Universal Sketch Features through Fusing Foundation Models/1fc738f6-a4cf-4a97-bcd3-ea7682cab3c6_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..b1b39381d7ab258e8611393029fdf42c4680c05f --- /dev/null +++ b/2025/SketchFusion_ Learning Universal Sketch Features through Fusing Foundation Models/1fc738f6-a4cf-4a97-bcd3-ea7682cab3c6_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f27b4f520aab57fe3e6aa994ae406ad0ab9e19daba0dc448b004bb0f4f238a0d +size 5691333 diff --git a/2025/SketchFusion_ Learning Universal Sketch Features through Fusing Foundation Models/full.md b/2025/SketchFusion_ Learning Universal Sketch Features through Fusing Foundation Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..43de6fff350a684ebc770ba42216dc99ab460548 --- /dev/null +++ b/2025/SketchFusion_ Learning Universal Sketch Features through Fusing Foundation Models/full.md @@ -0,0 +1,353 @@ +# SketchFusion: Learning Universal Sketch Features through Fusing Foundation Models + +Subhadeep Koley $^{1,2*}$ Tapas Kumar Dutta $^{1*}$ Aneeshan Sain $^{1}$ Pinaki Nath Chowdhury $^{1}$ Ayan Kumar Bhunia $^{1}$ Yi-Zhe Song $^{1,2}$ + +$^{1}$ SketchX, CVSSP, University of Surrey, United Kingdom. + +$^{2}$ iFlyTek-Surrey Joint Research Centre on Artificial Intelligence. + +{s.koley, a.sain, p.chowdhury, a.bhunia, y-song}@surrey.ac.uk + +![](images/fbf3bc9cc8e26a398b9b324416719f6803ea8f180d4fd170aa1f21e0d30874c3.jpg) +Figure 1. (Left): Apart from high-resolution image generation, text-to-image diffusion models (e.g., Stable diffusion (SD) [71]) with their innate object understanding capability [84, 107], have shown remarkable performance across a wide range of image-based vision tasks (e.g., segmentation [94], depth estimation [110], etc.). However, upon analysing the PCA representation of SD's intermediate UNet features, we observe that it struggles to achieve similar results when working with freehand abstract sketches (detail in Sec. 4). Unlike pixel-perfect photos, highly abstract freehand sketches are sparse and lack detailed textures and colours [26], making it harder for the SD model to extract meaningful features. Furthermore, investigating the SD denoising process in the frequency domain (via Fourier Transform), we observe the predominance of high-frequency (HF) components, rather than their low-frequency (LF) counterpart - crucial for capturing comprehensive semantic context. To mitigate this inherent bias within SD, we reinforce the diffusion process with another pretrained model (i.e., CLIP [67]) whose bias is complementary (i.e., focuses on LF) to SD. Consequently, the proposed extractor can extract semantically meaningful and accurate features from both sketches and photos, that encapsulate a broader frequency spectrum (i.e., HF and LF). (Right:) Testing the proposed method with different sketch-based discriminative and dense prediction tasks (requiring knowledge of both sketch and image), we find a marked improvement over baseline SD+CLIP hybrid feature extractor. + +![](images/8f30d00fa401bee519fd692a7dee43e1a0682d97cbfdfefa69195f81ae3ae1e6.jpg) + +![](images/5350e434ef008889b2e48186422c566193ca212b2f1868ef2cffe5ce4a90c48c.jpg) + +# Abstract + +While foundation models have revolutionised computer vision, their effectiveness for sketch understanding remains limited by the unique challenges of abstract, sparse visual inputs. Through systematic analysis, we uncover two fundamental limitations: Stable Diffusion (SD) struggles to extract meaningful features from abstract sketches (unlike its success with photos), and exhibits a pronounced frequency-domain bias that suppresses essential low-frequency components needed for sketch understanding. Rather than costly retraining, we address these limitations by strategically combining SD with CLIP, whose strong semantic understanding naturally compensates for SD's spatial-frequency biases. By dynamically injecting CLIP features into SD's denoising process and adaptively aggregating features across semantic levels, our method achieves state-of-the-art performance in sketch retrieval $(+3.35\%)$ , recognition $(+1.06\%)$ , segmentation $(+29.42\%)$ , and correspondence learning $(+21.22\%)$ , demonstrating the first truly universal sketch feature representation in the era of foundation models. + +# 1. Introduction + +The quest for universal sketch-specific features has been a foundational challenge in computer vision [46, 76, 78, 97]. Sketches, with their inherent characteristics of abstraction, sparsity, and cross-modal interpretation, demand fundamentally different feature representations compared to natural images [45, 100]. While foundation models [67, 71] have revolutionised various visual tasks, their effectiveness for sketch understanding remains largely unexplored [47], particularly given sketch's unique position at the intersection of human abstraction and computer vision [26, 45]. + +Through systematic analysis, we find that even powerful foundation models like Stable Diffusion (SD) [71] face significant challenges with sketch data. Our pilot studies (Sec. 4) reveal two fundamental limitations. First, while SD excels at extracting task-specific representative features from pixel-perfect RGB images, it struggles with abstract freehand sketches, likely due to its large-scale pretraining being predominantly focused on photographic data [71]. Second, and more intriguingly, frequency-domain analysis of SD's UNet reveals an inherent architectural bias: the model systematically promotes high-frequency (HF) components while suppressing low-frequency (LF) information - a characteristic that proves particularly problematic + +for sketch understanding tasks, especially dense prediction problems like segmentation and correspondence learning. + +These findings present a significant challenge. Retraining SD on sketch data risks catastrophic forgetting of its valuable pretrained knowledge [47], while developing sketch-specific architectures would fail to leverage the rich visual understanding already present in foundation models. Our key insight comes from analysing the complementary nature of different foundation models: while SD's features are spatially-aware but lack semantic accuracy, CLIP's [67] visual encoder exhibits strong semantic understanding even at the cost of spatial precision. This complementarity suggests that strategic combination [107] of these models could yield features that capture both the detailed structure and semantic meaning necessary for sketch understanding. + +Based on this insight, we propose a novel framework that leverages the complementary strengths of these foundation models. Our method systematically injects CLIP's visual features into multiple layers of SD's UNet during the denoising process, using lightweight $1D$ convolutions to influence feature extraction at various semantic levels. PCA analysis of the resulting features demonstrates that this combination successfully captures both the HF spatial details from SD [71] and the LF semantic components from CLIP [67], producing rich, semantically-aware representations suited for diverse sketch-based tasks. + +Moreover, we observe that different layers of SD's UNet capture features at varying levels of semantic-granularity [46], with different combinations better suited to specific downstream tasks [84]. Rather than manually selecting optimal features, we train a dynamic feature aggregator that automatically determines the most effective combination for any given task, eliminating the need for task-specific tuning. This adaptive approach ensures optimal performance across diverse sketch-based applications without requiring manual layer selection or task-specific optimisation. + +Our extensive experiments demonstrate the effectiveness of this approach across a comprehensive suite of sketch-based tasks: (i) Sketch Recognition: $+1.06\%$ improvement over state-of-the-art, (ii) Sketch-based Image Retrieval: $+3.35\%$ improvement in retrieval accuracy, (iii) Sketch-based Segmentation: $+29.42\%$ boost in mean IoU, (iv) Sketch-Photo Correspondence: $+21.22\%$ increase in matching precision. These significant improvements across all major sketch-specific tasks suggest that our method has uncovered fundamental principles for extracting effective sketch features from foundation models. + +In summary, our contributions are: (i) A systematic analysis revealing the limitations of pretrained foundation models for sketch understanding, supported by both empirical results and frequency-domain analysis, (ii) A novel method for combining complementary foundation models that address these limitations without retraining, (iii) An adap + +tive feature aggregation approach that automatically optimises representations for different downstream tasks, and $(iv)$ Comprehensive experimental validation demonstrating substantial improvements across all major sketch-specific tasks, establishing a new paradigm for universal sketch feature learning. + +# 2. Related Works + +Diffusion Model for Representation Learning. Diffusion models [34] have now become the de-facto standard for high-resolution $2D$ [21, 71] and $3D$ [12, 81] image and video [11] generation. Apart from generation, it gave rise to various image editing frameworks like Prompt-to-Prompt [33], Imagic [40], Dreambooth [73], etc. The success of diffusion models across various generative tasks and their inherent zero-shot generalisability has led to their adoption as efficient feature extractors in different forms of discriminative [23] and dense prediction [109] tasks like image retrieval [46], classification [49], semantic [3] and panoptic [94] segmentation, depth estimation [110], object detection [93], image-to-image translation [47, 86], controllable-generation [59, 108] and medical imaging [18], to name a few. A few recent works have developed specialised diffusion models together with carefully crafted task-specific designs like test-time adaptation [65], masked-reconstruction [62], or compact latent generation [36] to improve their representation ability. Nonetheless, our pilot study (Sec. 4) reveals that SD, often struggles to extract representative features (Fig. 1 left) from binary and sparse sketches, hampering downstream task performance. In this work, we leverage a pretrained CLIP [67] visual encoder to enhance SD's feature representation abilities for sketch-based discriminative and dense prediction tasks. + +Sketch-based Vision Applications. "Freehand sketch" has now established itself as a worthy interactive alternative modality to text for downstream vision tasks [9, 16, 31, 43, 76, 87, 102]. With its success in sketch-based image retrieval (SBIR) [13, 20, 76, 79], sketches are now being used in a wide array of discriminative [16, 17, 20], generative [27, 43, 108] and dense prediction [7, 35] tasks like controllable $2D$ [47, 108], $3D$ [2, 88] image generation and editing [31, 43, 58, 59, 69, 108], video synthesis [92], garment designing [14, 51], segmentation [35], creative content generation [9], object detection [16], salient region prediction [7], augmented and virtual reality applications [56, 57], medical imaging [41], class incremental learning [5], or composed retrieval [44]. In this paper, we aim to tackle some of the sketch-based downstream tasks by reinforcing the diffusion feature extractor via a pretrained CLIP [67] visual encoder. + +Backbones for Sketch-based Vision Applications. Backbones for sketch-based vision applications can be broadly categorised as - (i) pretrained CNN or Vision Transformer (ViT) [6, 13, 68], (ii) Self-supervised pretraining [4, 50, 64], and (iii) pretrained foundation models [39, 67, 71]. Starting + +with ImageNet pretrained CNNs [6, 13] or JFT pretrained ViTs [68, 77] with carefully-crafted design choices (e.g., cross [52] and spatial [83] attention, reinforcement learning [6], test-time training [75], meta-learning [74], knowledge distillation [77], etc.), the community moved towards self-supervised pretraining either via solving pretext tasks like sketch vectorisation [4], rasterisation [4], solving sketch-photo jigsaw [64] or by large-scale photo pretraining [50]. With the advent of large-scale foundation models, several works [43, 46, 76, 79] used them as a backbone for different sketch-based tasks like image generation [43, 108], shape synthesis [2], or retrieval [46, 79]. Foundation models have now become the gold standard for sketch-based tasks due to their strong zero-shot performance and extensive pretrained knowledge [45, 46, 79]. Among these, generative models like SD are spatially aware [46, 107], yet less accurate for dense prediction tasks. In contrast, vision-language models like CLIP [67] offer accurate features but are inherently sparse [107]. In this paper, we aim to get the best of both worlds by enhancing the SD feature extraction with CLIP. + +Reinforcing Feature Extractors. There are numerous ways of reinforcing an off-the-shelf pretrained feature extractor to adapt them for different downstream tasks. It can broadly be categorised as $(-i)$ Mixture of Experts (MoE) [22, 104], $(ii)$ Prompt learning [38, 112], $(iii)$ Hybrid models [91, 107], $(iv)$ Finetuning [76, 99], and $(v)$ Global-local feature merging [30, 105]. MoE combines multiple specialised sub-models (i.e., "experts"), each responsible for specific tasks or data patterns, to boost overall performance and adaptability [22]. In contrast, prompt learning deals with learning additional trainable components in the textual [111, 112] or pixel [1, 38] domain for task-specific adaptation. Hybrid models [89] combine two different kinds of models like CNN and ViT [91], or SD and DINO [107] to extract the best of both worlds. Another avenue deals with finetuning large-scale pretrained models with selective handcrafted data for specific tasks [16, 76]. Finally, a few other works [30, 90, 105] amalgamate global and local features from the same [90] or different [30, 105] models to improve downstream task performance. In this paper, we take the hybrid model approach and complement CNN-based SD features with transformer-based CLIP visual encoder features for different kinds of sketch-based vision tasks. + +# 3. Preliminaries + +# 3.1. Background: CLIP + +Contrastive language-image pretraining (CLIP) [67] architecture consists of a visual and a textual encoder, trained on $\sim 400\mathrm{M}$ image-text pairs, with the goal of learning a shared embedding manifold. Using N-pair contrastive loss, the model aligns matching image-text pairs by maximising their cosine similarity and minimises it for random unpaired ones [67]. The image encoder, typically a + +ViT [24], encodes an image $\mathbf{i}$ as a feature vector $\mathbf{V}(\mathbf{i})\in$ $\mathbb{R}^d$ . Meanwhile, the textual encoder processes a word sequence $\mathcal{W} = \{w^0,w^1,\dots ,w^k\}$ through byte pair encoding [67] and a learnable word embedding layer (vocabulary size of 49,152), converting each word $w^{i}$ into a token embedding, thus creating an embedding sequence $\mathcal{W}_e = \{\mathbf{w}_e^0,\mathbf{w}_e^1,\ldots ,\mathbf{w}_e^k\}$ [67]. $\mathcal{W}_e$ is then passed through a transformer network $\mathbf{T}(\cdot)$ , producing the final text feature $\mathbf{T}(\mathcal{W}_e)\in \mathbb{R}^d$ [67]. Once trained, the CLIP model can be used in downstream tasks to enable zero-shot capability and cross-modal understanding by aligning visual and textual data in a shared embedding space [67]. + +# 3.2. Background: Stable Diffusion (SD) + +Overview. The SD model [71] generates images by gradually removing noise from an initial $2D$ isotropic Gaussian noise image through iterative refinement. It involves two reciprocal processes - "forward" and "reverse" diffusion [34, 71]. The forward process iteratively adds random Gaussian noise into a clean image $\mathbf{i}_0 \in \mathbb{R}^{h \times w \times 3}$ for $t$ timesteps to generate a noisy image $\mathbf{i}_t$ as $\mathbf{i}_t = \sqrt{\bar{\alpha}_t} \mathbf{i}_0 + \sqrt{1 - \bar{\alpha}_t} \epsilon$ . Here, $\epsilon \sim \mathcal{N}(0, \mathbf{I})$ , and $\{\alpha_t\}_1^T$ is a pre-defined noise schedule where $\bar{\alpha}_t = \prod_{k=1}^t \alpha_k$ with $t \sim U(0, T)$ . The reverse process involves training a denoising UNet [72] $\mathbf{U}_{\theta}$ (with an $l_2$ objective), which estimates the input noise $\epsilon \approx \mathbf{U}_{\theta}(\mathbf{i}_t, t)$ from the noisy image $\mathbf{i}_t$ at each $t$ . After training, $\mathbf{U}_{\theta}$ is capable of reconstructing a clean image from a noisy input [71]. The inference process begins with a random $2D$ noise sample, $\mathbf{i}_T \sim \mathcal{N}(0, \mathbf{I})$ . The trained model $\mathbf{U}_{\theta}$ is then applied iteratively across $T$ timesteps, progressively reducing noise at each step to produce a cleaner intermediate image, $\mathbf{i}_{t-1}$ , finally generating one of the cleanest samples $\mathbf{i}_0$ from the target distribution. + +The unconditional denoising diffusion could further be made conditional by governing $\mathbf{U}_{\theta}$ with different auxiliary signals (e.g., textual captions) $\mathbf{c}$ [71]. Accordingly, $\mathbf{U}_{\theta}(\mathbf{i}_t,t,\mathbf{c})$ can denoise $\mathbf{i}_t$ while being influenced by $\mathbf{c}$ , commonly via cross-attention [71]. + +Architecture. Pixel-space diffusion model [34] is time-intensive as it operates on the full image resolution (i.e., $\mathbf{i}_0\in \mathbb{R}^{h\times w\times 3}$ ). Contrarily, SD model [71] performs denoising in the latent-space, resulting in significantly faster and more stable processing. In the first stage, SD trains a variational autoencoder (VAE) [42] (an encoder $\mathbf{E}(\cdot)$ and a decoder $\mathbf{D}(\cdot)$ ). $\mathbf{E}(\cdot)$ transforms the input image into its latent representation $\mathbf{z}_0 = \mathbf{E}(\mathbf{i}_0)\in \mathbb{R}^{\frac{h}{8}\times \frac{w}{8}\times d}$ . In the second stage, the UNet [72] $\mathbf{U}_{\theta}$ , directly denoises the latent images. $\mathbf{U}_{\theta}$ comprises 12 encoding, 1 bottleneck, and 12 skip-connected decoding blocks [71]. Inside these encoding and decoding blocks, there are 4 downsampling $(\mathbf{U}_{\mathbf{d}}^{1 - 4})$ and 4 upsampling $(\mathbf{U}_{\mathbf{u}}^{1 - 4})$ layers respectively. In SD, a frozen CLIP textual encoder [67] $\mathbf{T}(\cdot)$ converts the textual prompt $\mathbf{c}$ into token-sequence, that influences the denoising process of $\mathbf{U}_{\theta}$ through cross-attention. $\mathbf{U}_{\theta}$ is trained over an + +$l_{2}$ objective as: $\mathcal{L}_{\mathrm{l}_2} = \mathbb{E}_{\mathbf{z}_t,t,\mathbf{c},\epsilon}(||\epsilon -\mathbf{U}_\theta (\mathbf{z}_t,t,\mathbf{T}(\mathbf{c}))||_2^2)$ . During inference, a noisy latent $\mathbf{z}_t$ is sampled directly as: $\mathbf{z}_t\sim \mathcal{N}(0,\mathbf{I})$ . $\mathbf{U}_{\theta}$ removes noise from $\mathbf{z}_t$ iteratively over $T$ timesteps (conditioned on $\mathbf{c}$ ) to produce a denoised latent image $\hat{\mathbf{z}}_0\in \mathbb{R}^{\frac{h}{8}\times \frac{w}{8}\times d}$ . The final high-resolution image is generated as: $\hat{\mathbf{i}} = \mathbf{D}(\hat{\mathbf{z}}_0)\in \mathbb{R}^{h\times w\times 3}$ . + +# 4. Pilot Study: Problems and Solution + +What's Wrong with Diffusion Features? Binary and sparse freehand sketches contain much lesser semantic contextual information than pixel-perfect photos [15, 43, 79]. While SD [71] as a backbone feature extractor has established itself as a strong competitor in image-based vision tasks [3, 18, 93, 94, 110], we dive deep into its internal representations to evaluate its effectiveness for the more challenging cross-modal sketch-based vision tasks. To this end, we ask the research question - whether a frozen SD model is capable of extracting features from a sparse sketch that are equally-representative as those derived from a pixel perfect image. Given a few sketch-photo pairs, we extract the corresponding features (detail in Sec. 5.1) from the second upsampling block of a frozen SD's UNet $\mathbf{U}_{\theta}$ with a fixed prompt "a photo of [CLASS]" and $t = 195$ . We perform principal component analysis (PCA) on the extracted high-dimensional features and plot the first three principal components as RGB images for visualisation. As seen in Fig. 1 (left), features extracted from sketches are significantly inferior to those extracted from photos. Unlike photos, abstract freehand sketches are sparse and lack detailed textures/colours [15, 43, 79], making it challenging for the SD model to capture fine-grained visual information from it. We hypothesise that SD struggles to extract equally representative features from sketches due to their minimal semantic cues and high level of abstraction [47]. Additionally, as the pretraining of SD was done predominantly on natural images [80], it is comparatively harder for SD to extract rich semantic information from a sparse sketch. + +Digging deeper, we follow classical computer vision literature [28, 103] in shifting the focus from spatial-domain, to analyse the behaviour of extracted features from a different perspective, i.e., exploring its nature in the frequency-domain. We achieve this by calculating the $2D$ Fourier magnitude spectrum of the extracted features, where the central region represents the LF components, while the HF components are arranged radially outward from the centre [28] (Fig. 1 left). Here we see that, although SD performs better on photos than sketches, it predominantly focuses on HF components in both cases. This is likely because SD's UNet is primarily trained as a denoiser thus operating with a global emphasis [107], where its skip connections additionally propagate HF components (e.g., edges), suppressing its LF counterparts (e.g., overall semantic structures) to ease the reconstruction of high-res image from noise [82]. We hypothesise that this very fact is directly conflicting with the + +core motivation of our dense prediction task from a sparse data modality i.e., sketches, where capturing comprehensive semantic [107] context (LF) takes much higher precedence over preserving fine-details (HF), thus limiting SD's effectiveness [82, 107] for sketches, especially for dense prediction tasks (e.g., sketch-based segmentation). + +Reinforcing Diffusion Features. We identify two key problems in SD feature extraction - (i) SD features are not equally-representative for sketches and photos, and (ii) it mainly focuses on HF features, losing the crucial LF ones. In addressing these, our PCA analysis (Fig. 1 left) shows that, although SD is spatially-aware, it lacks accuracy. Inspired by existing works [107] on hybrid models, we propose complementing the SD feature extractor by combining it with another pretrained model that offers high accuracy, even if it is less spatially-aware. To this end, we influence the SD feature extraction pipeline with a pretrained CLIP [67] visual encoder as detailed in Sec. 5. PCA maps of CLIP (Fig. 1 left) features depict highly accurate background foreground separation, inherently providing the LF components that were missing in SD features. Finally, PCA maps of the proposed feature extractor (Fig. 1 left) demonstrate how the final representation is spatially-aware, accurate, integrates both HF and LF and works equally well for both sketches and photos. Apart from qualitative evidence, evaluating the performance of baseline $\mathrm{SD + }\mathrm{CLIP}$ hybrid model, and the proposed feature extractor across different sketch-based discriminative and dense prediction tasks, we observe an average performance gain of $39.49\%$ (Fig. 1 right). + +# 5. Proposed Methodology + +Overview. Observations from our pilot study (Sec. 4) motivate us to reinforce the SD [71] feature extraction pipeline for better downstream task accuracy. However, reinforcing the SD feature extractor pipeline is non-trivial. Firstly, finetuning the entire model with task-specific dataset might interrupt SD's inherent pretrained knowledge [46]. Furthermore, end-to-end finetuning does not necessarily resolve the issues pertaining to spatial-awareness vs. accuracy or HF vs. LF trade-off (Sec. 4). To this end, we resort to a hybrid model [107] approach and use a complementary pretrained model (i.e., CLIP [67]) to strengthen the SD feature extractor pipeline. Secondly, obtaining features from CLIP alone is insufficient. They must be integrated effectively into the denoising process of the SD UNet to yield the performance gains observed in Fig. 1. Finally, different layers of the SD's UNet capture a wide array of features across different scales and granularity (i.e., coarse to fine) [84], each with its own pros and cons suited to various downstream tasks [46]. Thus, it is essential to effectively aggregate these multi-scale features to maximise their utility. + +Specifically, our salient design choices are $(-i)$ complementing SD [71] feature extraction with a pretrained CLIP [67] visual model, $(ii)$ an effective CLIP visual feature in + +jection module, and (iii) an aggregation module to unify features from different scales and granularities. + +# 5.1. Model Architecture + +Stable Diffusion Feature Extraction. Internal activation maps of a pretrained SD UNet $\mathbf{U}_{\theta}$ contain information-rich representations that have found use in tasks like image segmentation [3], object detection [93], classification [49], etc. Here, we aim to use these internal representations as an efficient backbone feature for various sketch-based vision tasks. Given an image-text pair $(\mathbf{i},\mathbf{c})$ , the input image is first passed through the pretrained VAE [42] encoder $\mathbf{E}(\cdot)$ , converting it into an initial latent representation $\mathbf{z}_0 = \mathbf{E}(\mathbf{i}) \in \mathbb{R}^{\frac{h}{8} \times \frac{w}{8} \times d}$ . Then, given a timestep value $t$ , SD performs forward diffusion on $\mathbf{z}_0$ to convert it to the $t^{th}$ step noisy latent $\mathbf{z}_t$ . Finally, the textual prompt embedding (Sec. 3.2) $\mathbf{T}(\mathbf{c})$ , noisy latent $\mathbf{z}_t$ , and timestep $t$ are passed to the pretrained denoising UNet [71] $\mathbf{U}_{\theta}(\cdot)$ to extract the internal features from its skip-connected upsampling layers $\{\mathbf{U}_{\mathbf{u}}^n \to f_{\mathbf{u}}^n\}_{n=1}^4$ . In SD v2.1 [71], for an input image $\mathbf{i} \in \mathbb{R}^{h \times w \times 3}$ , upsampling layers $\{\mathbf{U}_{\mathbf{u}}^n\}_{n=1}^4$ generate features as: $f_{\mathbf{u}}^1 \in \mathbb{R}^{\frac{h}{32} \times \frac{w}{32} \times 1280}$ , $f_{\mathbf{u}}^2 \in \mathbb{R}^{\frac{h}{16} \times \frac{w}{16} \times 1280}$ , $f_{\mathbf{u}}^3 \in \mathbb{R}^{\frac{h}{8} \times \frac{w}{8} \times 640}$ , and $f_{\mathbf{u}}^4 \in \mathbb{R}^{\frac{h}{8} \times \frac{w}{8} \times 320}$ respectively. Due to the unavailability of paired textual prompts in most sketch-photo datasets [26, 29, 78], we use null (i.e., “”) prompt in all our experiments. + +![](images/187a38e80495717266c0b7d1cfd55d677183700bb250696f0c95e6719f483051.jpg) +Figure 2. Given the frozen SD [71] and CLIP [67] models, the proposed method learns the aggregation network, $1D$ convolutional layers, and branch-weights with sketch-photo pairs, via different losses for different downstream tasks (details in Sec. 6). + +CLIP Visual Feature Extraction. Akin to SD, pretrained CLIP [67] encoders have also been used in a plethora of vision tasks like segmentation [60], retrieval [76], object detection [16], video captioning [98], to name a few. In this work, we aim to influence the intermediate feature maps of SD [71] with that from the CLIP visual encoder. Specifically, the input image $\mathbf{i}$ is passed through the pretrained CLIP visual encoder $\mathbf{V}(\cdot)$ to extract the patched feature $f_{\mathbf{v}}$ from the penultimate layer as: $f_{\mathbf{v}} = \mathbf{V}(\mathbf{i}) \in \mathbb{R}_{p}^{\frac{h}{p} \times \frac{w}{p} \times d}$ , where $p$ is the patch-size of the $\mathbf{V}(\cdot)$ backbone encoder. + +CLIP Feature Injection. Based on the qualitative and quantitative evidence presented in the pilot study (Sec. 4), we hypothesise that CLIP [67] visual and textual embeddings are intrinsically aligned [67, 79]. Prior works [79] + +have underscored the importance of textual prompts in the quality of extracted features [79]. However, most simple prompts (e.g., "a photo/sketch of [CLASS]") are inherently coarse-grained, offering only a high-level description of an image. Such general prompts often lack the semantic detail necessary to fully characterise an image [15, 47], limiting the model's ability to generate highly accurate and contextually-rich feature representations [46]. + +To address this, we propose injecting CLIP's visual features into multiple layers of the SD's UNet, thereby influencing the denoising process at various stages (Fig. 2). This approach not only enhances the accuracy of SD features but also enables the model to draw on semantic information from both CLIP's visual and textual embeddings. + +Specifically, after extracting CLIP's visual embeddings $f_{\mathbf{v}} \in \mathbb{R}^{\frac{h}{p} \times \frac{w}{p} \times d}$ , we pass them through simple learnable $1D$ convolutional layers $\mathcal{C}(\cdot)$ to modify them just enough to match the feature dimension of different blocks of the UNet. These transformed feature maps are then added to the intermediate UNet features $\{f_{\mathbf{u}}^{n}\}_{n=1}^{4}$ at every timestep of the denoising process as: $\hat{f}_{\mathbf{u}}^{n} = f_{\mathbf{u}}^{n} + \mathcal{C}(f_{\mathbf{v}})$ . This injection occurs in all timesteps and layers of the UNet, enabling the model to influence the denoising process with the semantic information encoded within CLIP's visual embeddings. + +SD Feature Aggregation. Different layers of $\mathbf{U}_{\theta}$ capture features of varying semantic-granularity [84], where features from specific layers are ideally suited for certain downstream tasks [46]. Instead of manually selecting each layer for a specific task, we introduce a dynamic approach, allowing the model to automatically extract the most optimal features, thus eliminating the need for manual tuning. + +To achieve this, we train a feature aggregation network comprising three ResNet [32] blocks $\mathbf{A}(\cdot)$ , which transform the CLIP-enhanced SD features $\{\hat{f}_{\mathbf{u}}^{n}\}_{n=1}^{3}$ extracted from the first three UNet upsampling layers as: $\bar{f}_{\mathbf{u}}^{n} = \mathbf{A}(\hat{f}_{\mathbf{u}}^{n}) \in \mathbb{R}^{60 \times 60 \times d}$ . The final feature map is then obtained by computing a weighted summation of the aggregated features using a learnable weight parameter $\{\alpha_{n}\}_{n=1}^{3}$ for each feature-branch. This aggregation strategy enables the model to dynamically adjust its focus on different layers, adaptively modulating the contribution of each feature-branch based on the semantic-granularity of the task. + +# 6. Experiments + +The proposed feature extractor enriched with the large-scale pretrained knowledge of SD [71] and CLIP [67], enables multiple sketch-based discriminative and dense prediction tasks. Specifically, we showcase its efficacy in tasks like - sketch-based image retrieval [76], sketch recognition [100], sketch-photo correspondence learning [55], and sketch-based image segmentation [35] on different datasets [26, 29, 55, 78] and scenarios. In all cases, we only learn $\mathcal{C}(\cdot)$ , aggregation network $\mathbf{A}(\cdot)$ , and branch-weights $\alpha_{n}$ + +keeping pretrained SD and CLIP models frozen. + +# 6.1. Sketch-based Image Retrieval (SBIR) + +Problem Statement. Given a query sketch $\mathbf{s} \in \mathbb{R}^{h \times w \times 3}$ of any class $j$ , category-level SBIR aims to retrieve a photo $\mathbf{p}_i^j$ from the same class, out of a gallery $\mathcal{G} = \{\mathbf{p}_i^j\}_{i=1}^{N_j}|_{j=1}^{N_c}$ containing images from a total $N_c$ classes with $N_j$ images per class [102]. Whereas, cross-category fine-grained (FG) SBIR setup intends to learn a unified model capable of instance-level matching from a gallery containing photos from multiple $(N_c)$ classes. Conventional SBIR is evaluated on categories $\mathcal{C}^S = \{c_1^S, \dots, c_N^S\}$ that are seen during training. Contrarily, the zero-shot (ZS) paradigm [76] focuses on evaluating on mutually exclusive (i.e., $\mathcal{C}^S \cap \mathcal{C}^U = \emptyset$ ) unseen categories $\mathcal{C}^U = \{c_1^U, \dots, c_M^U\}$ . + +Training & Evaluation. Following standard literature [76], we use triplet loss to train our SBIR models. Specifically, given the global max-pooled features from the proposed feature extractor, triplet loss aims to minimise the distance $\delta$ between an anchor sketch (s) feature $f_{\mathbf{s}} \in \mathbb{R}^{d}$ and a positive photo (p) feature $f_{\mathbf{p}} \in \mathbb{R}^{d}$ from the same category as s, while maximising the same from a negative photo (n) feature $f_{\mathbf{n}} \in \mathbb{R}^{d}$ of a different category. Whereas, in case of cross-category FG-SBIR framework we use hard triplets, where the negative sample is a distinct instance $(\mathbf{p}_k^j; k \neq i)$ belonging to the same class as the anchor sketch $(\mathbf{s}_i^j)$ and its paired photo $(\mathbf{p}_i^j)$ [46]. With margin $\mu$ , triplet loss becomes: + +$$ +\mathcal {L} _ {\text {t r i p l e t}} \left(f _ {\mathbf {s}}, f _ {\mathbf {p}}, f _ {\mathbf {n}}\right) = \max \left\{0, \mu + \delta \left(f _ {\mathbf {s}}, f _ {\mathbf {p}}\right) - \delta \left(f _ {\mathbf {s}}, f _ {\mathbf {n}}\right) \right\}. \tag {1} +$$ + +For category-level ZS-SBIR, following [46, 76], we consider top $k$ retrieved images to calculate mean average precision $(\mathrm{mAP}@\mathrm{k})$ and precision $(\mathrm{P}@\mathrm{k})$ . Whereas, for cross-category ZS-FG-SBIR, we measure Acc.@k, which indicates the percentage of sketches having true-matched photos among the top-k retrieved images. + +Dataset & Competitors. Here we use three datasets - (i) Sketchy [78] and Sketchy-extended [53, 78] datasets are used to train and test our cross-category ZS-FG-SBIR and category-level ZS-SBIR models respectively. While Sketchy [78] holds 12,500 photos from 125 classes, each having at least 5 fine-grained paired sketches, its extended version [53] carries additional 60,652 photos from ImageNet [19]. We use photos/sketches from 104 classes for training and 21 for testing [102]. (ii) Tu-Berlin [26] contains 204,489 photos from 250 classes, each having $\sim 80$ sketches. We use photos/sketches from 220 classes for training and 30 for testing. (iii) Quick, Draw! [29] holds over $50M$ sketches from 345 categories. We use photos/sketches from 80 classes for training and 30 for testing. + +For category-level ZS-SBIR, we compare our method with SoTA methods like SAKE [54], CAAE [102], CCGAN [25], GRL [20], SD-PL [46], etc. Whereas, for cross-category ZS-FG-SBIR, we compare with Gen-VAE [63], + +
MethodsSketchy [78]TU-Berlin [26]Quick, Draw! [29]
mAP@200P@200mAP@allP@100mAP@allP@200
B-CLIP0.2500.2610.2280.2570.0800.141
B-DINO0.4930.4810.4500.4920.1670.249
B-DINOV20.5270.5330.4810.5320.1700.268
B-SD0.5580.5710.5100.5610.1790.287
B-SD+CLIP0.5880.5920.5370.5890.1790.311
B-Finetuning0.1200.1720.0110.0100.0030.006
SAKE [54]0.4970.5980.4750.599--
IIAE [37]0.3730.4850.4120.503--
CAAE [102]0.1560.2600.0050.003--
CCGAN [25]--0.2970.426--
CVAE [102]0.2250.3330.0050.0010.0030.003
GRL [20]0.3690.3700.1100.1210.0750.068
LVM [76]0.7230.7250.6510.7320.2020.388
SD-PL [46]0.7460.7470.6800.7440.2310.397
Proposed0.7610.7630.6950.7530.2420.399
+ +LVM [76], and $SD-PL$ [46]. Apart from the SoTA methods, we also compare all tasks against a few self-designed Baselines. Among them B-CLIP, B-DINO, B-DINOv2, and B-SD use off-the-shelf pretrained CLIP [67] ViT-L/14, DINO [8] ViT-S/8, DINOv2 [61] ViT-S/14, or SD [71] v2.1 encoders as feature extractors, without task-specific training. We also form another baseline $B-SD+CLIP$ as a straightforward combination of SD and CLIP features without the proposed feature injection and aggregation modules. Finally, B-Finetuning finetunes the entire pretrained SD UNet and CLIP visual encoder on task-specific data. + +Table 1. Results for category-level ZS-SBIR. + +
MethodsAcc.@1Acc.@5MethodsAcc.@1Acc.@5
B-CLIP11.8421.66B-Finetuning5.679.17
B-DINO22.4946.97Gen-VAE [63]22.6049.00
B-DINOv221.1944.31LVM [76]28.6862.34
B-SD23.9849.42SD-PL [46]31.9465.81
B-SD+CLIP24.1652.01Proposed33.0167.92
+ +Table 2. Results on Sketchy [78] for cross-category ZS-FG-SBIR. + +# 6.2. Sketch Recognition + +Problem Statement. Sketch recognition deals with identifying and classifying freehand abstract sketches into predefined categories [68, 96]. Given a dataset $\mathcal{D} = \{(\mathbf{s}_i,\mathbf{y}_i)\}_{i = 1}^N$ containing $N$ sketches and their corresponding labels from $C$ classes, the goal is to learn a model that can correctly recognise the class of an input query sketch. + +Training & Evaluation. Given the global max-pooled query sketch feature $f_{\mathbf{s}} \in \mathbb{R}^{d}$ , we use standard cross-entropy (CE) loss to train our sketch recognition model as, + +$$ +\mathcal {L} _ {\mathrm {C E}} \left(\mathbf {s} _ {i}, \mathbf {y} _ {i}\right) = - \frac {1}{N} \sum_ {i = 1} ^ {N} \log p \left(\mathbf {y} _ {i} \mid \mathbf {s} _ {i}\right). \tag {2} +$$ + +Following [66, 96, 101] we evaluate using Acc.@k. + +Dataset & Competitors. For this task, we again use the popular TU-Berlin [26] and Quick, Draw! [29] datasets. For Quick, Draw!, we adopt the data split from [29], with each class containing 70,000, 2,500, and 2500 training, validation, and test samples. Apart from the self-designed baselines, we also compare against SoTA methods like SketchMate [96], SketchGNN [101], and SketchXAI [66]. + +
MethodsTU-BerlinQuick, Draw!MethodsTU-BerlinQuick, Draw!
B-CLIP30.0930.87B-Finetuning10.2912.37
B-DINO51.7953.28SketchMate [96]77.9679.44
B-DINOv258.0159.12SketchGNN [101]76.4377.31
B-SD61.3763.57SketchXAI [66]-86.10
B-SD+CLIP65.4767.78Proposed84.9687.02
+ +# 6.3. Sketch-photo Correspondence + +Problem Statement. Sketch-photo correspondence learning involves identifying a mapping between a freehand sketch of an object and its corresponding photo, such that keypoints in the sketch can be semantically aligned with corresponding keypoints in the photo [55]. Given a sketch-photo pair $\{\mathbf{s},\mathbf{p}\} \in \mathbb{R}^{h\times w\times 3}$ and a set of $N$ query keypoints (on sketch) $K_{\mathbf{s}} = \{x_i^s,y_i^s\}_{i = 1}^N$ , the aim is to predict a new set of corresponding keypoints $\hat{K}_{\mathbf{p}} = \{\hat{x}_i^{\mathbf{p}},\hat{y}_i^{\mathbf{p}}\}_{i = 1}^N$ on the photo, such that $\hat{K}_{\mathbf{p}}$ aligns semantically with $K_{\mathbf{s}}$ . + +Training & Evaluation. Following [106], we use contrastive $(\mathcal{L}_{\mathrm{CL}})$ [67] and end-point error losses to train our sketch-photo correspondence model. Given the sketch and photo features $\{f_{\mathrm{s}}, f_{\mathrm{p}}\} \in \mathbb{R}^{60 \times 60 \times d}$ from the proposed feature extractor, $\mathcal{L}_{\mathrm{CL}}$ aims to maximise similarity between local feature patches, while minimising the same between distance ones, for each sketch-photo pairs. Whereas, for endpoint error loss, we first estimate optical flows [106] from the extracted features, followed by calculating $l_{2}$ norm between the estimated $(\mathbf{f}_e)$ and ground truth $(\mathbf{f}_g)$ flow as: + +$$ +\mathcal {L} _ {\mathrm {E P E}} \left(\mathbf {f} _ {e} ^ {i}, \mathbf {f} _ {g} ^ {i}\right) = \frac {1}{M} \sum_ {i = 1} ^ {M} \left\| \mathbf {f} _ {e} ^ {i} - \mathbf {f} _ {g} ^ {i} \right\| _ {2} ^ {2}, \tag {3} +$$ + +where, $M$ denotes the patch number (i.e., 60). Finally, the overall loss becomes $\mathcal{L}_{\mathrm{corr}} = \mathcal{L}_{\mathrm{CL}} + \mathcal{L}_{\mathrm{EPE}}$ . Following [55] we use PCK@k as a test metric which measures the percentage of keypoints predicted correctly within a distance threshold (i.e., $k\%$ of the image size). + +Dataset & Competitors. For this task, we use the PSC6K [55] dataset that contains more than $150K$ keypoint annotations for 1250 photos, each having 5 paired sketches (taken from the original Sketchy [78] test split) from 125 classes. We use a 60:40 split for training and testing, where we compare the proposed approach against the self-designed baselines and SoTA methods like WeakAlign[70], WarpC [85], Self-Supervised [55], etc. + +Table 3. Acc.@1 results for sketch recognition. + +
MethodsPCK@5PCK@10MethodsPCK@5PCK@10
B-CLIP22.2930.41B-Finetuning11.5817.69
B-DINO34.2148.59WeakAlign [70]43.5578.60
B-DINOv239.9156.46WarpC [85]56.7879.70
B-SD41.4258.71Self-Sup. [55]58.0084.93
B-SD+CLIP44.9062.02Proposed70.3189.86
+ +Table 4. Results on PSC6K [55] for sketch-photo correspondence. 6.4. Sketch-based Image Segmentation + +Problem Statement. Given a query sketch $\mathbf{s} \in \mathbb{R}^{h \times w \times 3}$ of any category, sketch-based image segmentation [35] aims to predict a binary segmentation mask, $\mathbf{m} \in \mathbb{R}^{h \times w \times 1}$ , depicting the pixel-locations, wherever the queried-concept appears in any candidate image $\mathbf{p} \in \mathbb{R}^{h \times w \times 3}$ of that category, + +from an entire test gallery. The predicted mask marks pixels belonging to that sketched-concept as 1 and the rest as 0. + +Training & Evaluation. We train our segmentation model using binary cross-entropy (BCE) loss, between predicted and ground truth (GT) masks. Given patched photo feature $f_{\mathbf{p}} \in \mathbb{R}^{60 \times 60 \times d}$ and global max-pooled sketch feature photo features $f_{\mathbf{s}} \in \mathbb{R}^d$ from the proposed feature extractor, we calculate the cosine similarity between them to generate a patch-wise correlation map $\mathbf{c} = f_{\mathbf{s}} \odot f_{\mathbf{p}} \in \mathbb{R}^{60 \times 60}$ . We upsample $\mathbf{c}$ to the original spatial resolution (i.e., $\mathbb{R}^{h \times w}$ ), followed by differentiable Sigmoid thresholding [10] to get the predicted map. BCE loss between the GT and predicted maps can be calculated as, + +$$ +\mathcal {L} _ {\mathrm {B C E}} \left(y _ {i}, \hat {y} _ {i}\right) = - \frac {1}{N} \sum_ {i = 1} ^ {N} \left[ y _ {i} \log \left(\hat {y} _ {i}\right) + \left(1 - y _ {i}\right) \log \left(1 - \hat {y} _ {i}\right) \right] \quad , \tag {4} +$$ + +where, $N$ is the number of total pixels, $y_{i}$ and $\hat{y}_i$ are the GT and predicted mask pixels. During testing, we calculate similar patch-wise correlation map from sketch/photo features and upsample it to the original spatial image dimension, followed by thresholding it with an empirically chosen fixed value (0.47) and standard CRF [48]-based postprocessing. We use mean Intersection over Union (mIoU) and pixel accuracy (pAcc.) as test metrics, which measure the overlap and percentage of correctly predicted pixels between predicted and GT segmentation masks respectively. + +Dataset & Competitors. Due to the lack of existing paired sketch-photo-mask datasets, we manually annotate photos of 10 chosen classes from the Sketchy [78] dataset with binary mask annotations, resulting in $5K$ paired sketch-photo-mask tuples. We split it into 80:20 for training and testing. Besides baselines and SoTA Sketch-a-Segmenter [35], we also compare with a modified version of ZS-Seg [95], which adapts a simple text-based zero-shot segmentation framework for sketches by replacing its CLIP [67] textual encoder with the visual one to encode input sketches. + +![](images/d89ac8a0f6954ad106c6e9bc31ac9ca939af4347dbb907db2b33e7a4e1eff411.jpg) +Figure 3. Sketch-photo correspondence (left→right : source → target) results on PSC6K. Green circles and squares depict source and GT points respectively, while red squares denote predicted points. + +# 6.5. Results and Analyses + +Following are the observations from quantitative and qualitative results presented in Tab. 1-5 and Fig. 3-4. (i) Naive adaptation of pretrained backbones (e.g., B-CLIP, B-DINO, B-SD, etc.) fail to beat our method in both ZS-SBIR and ZS-FG-SBIR settings (Tab. 1-2), where we achieve an average 76.15 (54.32)% higher mAP@200 (Acc.@1) over all competitors on Sketchy. We posit that our strategic combination of SD and CLIP brings the "best of both worlds", complementing each other's flaws. (ii) While + +
MethodsmIoUpAcc.MethodsmIoUpAcc.
B-CLIP20.6326.82B-Finetuning20.3621.84
B-DINO32.8142.06DeepLabv3+Sketch [35]40.6355.23
B-DINOv234.1745.80ZS-Seg [95]44.7360.97
B-SD39.0249.03Sketch-a-Segmenter [35]46.4560.28
B-SD+CLIP41.8751.91Proposed60.1274.91
+ +SoTA sketch-recognition frameworks [96, 101] perform reasonably across different datasets (Tab. 3), our method exceeds them with an average Acc.@1 gain of $48.09\%$ (Quick, Draw!). SketchXAI [66] despite being our strongest contender, depicts sub-optimal Acc.@1. This gain is likely due to the large-scale pretraining of SD and CLIP, which enables superior open-world object understanding [107]. (iii) For sketch-photo correspondence learning, our method shows an impressive $21.22\%$ PCK@5 gain over SoTA [55], without the time-consuming two-stage training of [55], or complicated warp-estimation of [85], thanks to SD's innate object-localisation capability. (iv) Sketch-based segmentation being the most challenging task amongst all, baseline methods perform quite poorly (Tab. 5). However, the proposed method surpasses SoTA [35] with a mIoU (pAcc.) margin of $29.42(24.27)\%$ . (v) B-Finetuning fails drastically in almost all tasks (Tab. 1-5). We hypothesise that finetuning entire SD and CLIP models with limited sketch data distorts their large-scale pretrained knowledge [46, 76], thus sacrificing the generalisation potential. (vi) Finally, B-SD+CLIP depicts a $\sim 35.78\%$ lesser scores than ours (averaged over all tasks), which again underpins the need for the proposed feature injection and aggregation strategy, rather than a simple combination of pretrained foundation models. + +![](images/ed87e7b74dc053d81e4ac063d273158fb0c1303a12c6a289a234fbc76ecb31bb.jpg) +Figure 4. Qualitative results for sketch-based image segmentation. Given a query sketch, our method generates separate segmentation masks for all images of that category. (Zoom-in for the best view.) + +# 7. Ablation on Design + +[i] Effect of aggregation network. To judge its efficacy, instead of aggregation networks $\mathbf{A}(\cdot)$ , we manually interpolate features from $\{\mathbf{U}_{\mathbf{u}}^{n}\}_{n=1}^{3}$ into $\mathbb{R}^{60 \times 60 \times d}$ dimension and add them (via $\alpha_{n}$ ) to form the final feature map. A rapid PCK@5 (mIoU) drop of 28.65 (26.51)% in case of w/o Aggregation Network (Tab. 6) indicates its significance in extracting semantically-meaningful features. + +[i] Are learnable branch weights necessary? Learnable weights dynamically ensure the optimum importance of each feature-branch for a specific task. Although less pronounced in PCK@5, mAP@200 drops significantly (4.82%) in case of w/o Learnable Weights (Tab. 6), depicting its influence on the final feature map quality. + +Table 5. Results for sketch-based image segmentation. + +
MethodsCorrespondenceSegmentationZS-SBIRRecognition
PCK@5mIoUmAP@200Acc.@1
SD [71]v1.468.9157.630.72878.82
v1.569.2458.910.73579.59
v2.170.3160.120.76187.02
CLIP [67]ViT-B/1667.7556.380.74377.37
ViT-B/3269.4158.470.75878.84
ViT-L/1470.3160.120.76187.02
w/o Aggregation Network54.6547.520.58766.71
w/o Learnable Weights68.7558.130.72683.79
w/o 1D Convolutions55.9747.240.60271.75
Ours-full70.3160.120.76187.02
+ +Table 6. Ablation on design. + +[iii] Contribution of $1D$ Convolutions. To assess the effectiveness of $1D$ convolution layers, we use simple linear interpolation to match CLIP and SD UNet feature dimensions, before injection. Sharp PCK@5 drop of $25.62\%$ for w/o $1D$ Convolutions (Tab. 6) reveals its impact on influencing the diffusion denoising process. We posit that the learnable $1D$ convolutions not only match the feature dimension but also enrich the CLIP features before injection. [iv] Model variants. The performance of our proposed feature extractor depends crucially on the choice of backbone model variant. Accordingly, we explore with multiple pretrained SD and CLIP backbones. Tab. 6 shows SD v2.1 and CLIP ViT-L/14 to achieve the highest PCK@5 and Acc.@1 for sketch-photo correspondence and recognition respectively. Additionally, we also experiment by replacing the CLIP visual encoder of our method with a DINOv2 [61] ViT-S/14 encoder for the task of ZS-SBIR, to observe a $32.80\%$ drop in mAP@200. This is probably due to CLIP's [67] multimodal pertaining (diminishing sketch-photo domain gap), over DINOv2's [61] self-supervised one. + +[v] Which timestep value is the most effective? Different diffusion timesteps yield features with drastically different quality [84, 107]. Calculating PCK@5 and mAP@200 + +![](images/19c7efbf2105c1178a587d2ccc1d25a9ff1969dac54da7cc2b39b28f322ff58c.jpg) +Figure 5. Choice of timestep $(t)$ . method to be robust to the choice of $t$ , where PCK@5 (mAP@200) scores from a wide range of $t$ values consistently surpass the baselines. + +for the tasks of correspondence learning and ZS-SBIR respectively, over varying $t$ , we find that $t = 195$ produces the best results for both tasks. Fig. 5 shows our method to be robust to + +# 8. Conclusion + +In this paper, we identified SD's limitations in understanding abstract, sparse sketches. Integrating CLIP's semantic features with SD's spatially-aware features, we developed a framework that dynamically combines these strengths. This significantly improves performance across sketch recognition, SBIR, and sketch-based segmentation. Our method highlights the complementary nature of different foundation models and introduces an adaptive feature aggregation network, establishing a new standard for universal sketch feature representation for future research. + +# References + +[1] Hyojin Bahng, Ali Jahanian, Swami Sankaranarayanan, and Phillip Isola. Exploring visual prompts for adapting large-scale models. arXiv preprint arXiv:2203.17274, 2022.3 +[2] Hmrishav Bandyopadhyay, Subhadeep Koley, Ayan Das, Ayan Kumar Bhunia, Aneeshan Sain, Pinaki Nath Chowdhury, Tao Xiang, and Yi-Zhe Song. Doodle Your 3D: From Abstract Freehand Sketches to Precise 3D Shapes. In CVPR, 2024. 2, 3 +[3] Dmitry Baranchuk, Ivan Rubachev, Andrey Voynov, Valentin Khrulkov, and Artem Babenko. Label-Efficient Semantic Segmentation with Diffusion Models. In ICLR, 2021. 2, 4, 5 +[4] Ayan Kumar Bhunia, Pinaki Nath Chowdhury, Yongxin Yang, Timothy Hospedales, Tao Xiang, and Yi-Zhe Song. Vectorization and Rasterization: Self-Supervised Learning for Sketch and Handwriting. In CVPR, 2021. 2, 3 +[5] Ayan Kumar Bhunia, Viswanatha Reddy Gajjala, Subhadeep Koley, Rohit Kundu, Aneeshan Sain, Tao Xiang, and Yi-Zhe Song. Doodle It Yourself: Class Incremental Learning by Drawing a Few Sketches. In CVPR, 2022. 2 +[6] Ayan Kumar Bhunia, Subhadeep Koley, Abdullah Faiz Ur Rahman Khilji, Aneeshan Sain, Pinaki Nath Chowdhury, Tao Xiang, and Yi-Zhe Song. Sketching Without Worrying: Noise-Tolerant Sketch-Based Image Retrieval. In CVPR, 2022. 2, 3 +[7] Ayan Kumar Bhunia, Subhadeep Koley, Amandeep Kumar, Aneeshan Sain, Pinaki Nath Chowdhury, Tao Xiang, and Yi-Zhe Song. Sketch2Saliency: Learning to Detect Salient Objects from Human Drawings. In CVPR, 2023. 2 +[8] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging Properties in Self-Supervised Vision Transformers. In ICCV, 2021. 6 +[9] Dar-Yen Chen, Subhadeep Koley, Aneeshan Sain, Pinaki Nath Chowdhury, Tao Xiang, Ayan Kumar Bhunia, and Yi-Zhe Song. DemoCaricature: Democratising Caricature Generation with a Rough Sketch. In CVPR, 2024. 2 +[10] Honglie Chen, Weidi Xie, Triantafyllos Afouras, Arsha Nagrani, Andrea Vedaldi, and Andrew Zisserman. Localizing Visual Sounds the Hard Way. In CVPR, 2021. 7 +[11] Haoxin Chen, Yong Zhang, Xiaodong Cun, Menghan Xia, Xintao Wang, Chao Weng, and Ying Shan. VideoCraft2: Overcoming Data Limitations for High-Quality Video Diffusion Models. In CVPR, 2024. 2 +[12] Gene Chou, Yuval Bahat, and Felix Heide. Diffusion-SDF: Conditional Generative Modeling of Signed Distance Functions. In CVPR, 2023. 2 +[13] Pinaki Nath Chowdhury, Ayan Kumar Bhunia, Viswanatha Reddy Gajjala, Aneeshan Sain, Tao Xiang, and Yi-Zhe Song. Partially Does It: Towards Scene-Level FG-SBIR with Partial Input. In CVPR, 2022. 2, 3 +[14] Pinaki Nath Chowdhury, Tuanfeng Wang, Duygu Ceylan, Yi-Zhe Song, and Yulia Gryaditskaya. Garment ideation: + +Iterative view-aware sketch-based garment modeling. In 3DV, 2022. 2 +[15] Pinaki Nath Chowdhury, Ayan Kumar Bhunia, Aneshan Sain, Subhadeep Koley, Tao Xiang, and Yi-Zhe Song. SceneTrilogy: On Human Scene-Sketch and its Complementarity with Photo and Text. In CVPR, 2023. 4, 5 +[16] Pinaki Nath Chowdhury, Ayan Kumar Bhunia, Aneshan Sain, Subhadeep Koley, Tao Xiang, and Yi-Zhe Song. What Can Human Sketches Do for Object Detection? In CVPR, 2023. 2, 3, 5 +[17] John P. Collomosse, Tu Bui, and Hailin Jin. LiveSketch: Query Perturbations for Guided Sketch-based Visual Search. In CVPR, 2019. 2 +[18] Bram de Wilde, Anindo Saha, Richard PG ten Broek, and Henkjan Huisman. Medical diffusion on a budget: Textual Inversion for medical image generation. In MIDL, 2024. 2, 4 +[19] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. ImageNet: A Large-Scale Hierarchical Image Database. In CVPR, 2009. 6 +[20] Sounak Dey, Pau Riba, Anjan Dutta, Josep Llados, and Yi-Zhe Song. Doodle to Search: Practical Zero-Shot Sketch-based Image Retrieval. In CVPR, 2019. 2, 6 +[21] Prafulla Dhariwal and Alexander Nichol. Diffusion Models Beat GANs on Image Synthesis. In NeurIPS, 2021. 2 +[22] Francesco Di Sario, Riccardo Renzulli, Enzo Tartaglione, and Marco Grangetto. Boost Your NeRF: A Model-Agnostic Mixture of Experts Framework for High Quality and Efficient Rendering. In ECCV, 2024. 3 +[23] Shiyin Dong, Mingrui Zhu, Kun Cheng, Nannan Wang, and Xinbo Gao. Bridging generative and discriminative models for unified visual perception with diffusion priors. In IJCAI, 2024. 2 +[24] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 3 +[25] Anjan Dutta and Zeynep Akata. Semantically Tied Paired Cycle Consistency for Zero-Shot Sketch-based Image Retrieval. In CVPR, 2019. 6 +[26] Mathias Eitz, James Hays, and Marc Alexa. How do humans sketch objects? ACM TOG, 2012. 1, 5, 6 +[27] Arnab Ghosh, Richard Zhang, Puneet K Dokania, Oliver Wang, Alexei A Efros, Philip HS Torr, and Eli Shechtman. Interactive Sketch & Fill: Multiclass Sketch-to-Image Translation. In ICCV, 2019. 2 +[28] Rafael C Gonzalez and Richard E Woods. Digital Image Processing. Pearson, 2008. 4 +[29] David Ha and Douglas Eck. A Neural Representation of Sketch Drawings. In ICLR, 2017. 5, 6 +[30] Moayed Haji-Ali, Guha Balakrishnan, and Vicente Ordonez. ElasticDiffusion: Training-free Arbitrary Size Image Generation through Global-Local Content Separation. In CVPR, 2024. 3 + +[31] Cusuh Ham, Gemma Canet Tarres, Tu Bui, James Hays, Zhe Lin, and John Collomosse. CoGS: Controllable Generation and Search from Sketch and Style. In ECCV, 2022. 2 +[32] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep Residual Learning for Image Recognition. In CVPR, 2016. 5 +[33] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-Prompt Image Editing with Cross Attention Control. In ICLR, 2022. 2 +[34] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising Diffusion Probabilistic Models. In NeurIPS, 2020. 2, 3 +[35] Conghui Hu, Da Li, Yongxin Yang, Timothy M Hospedales, and Yi-Zhe Song. Sketch-a-Segmented: Sketch-based Photo Segmenter Generation. IEEE TIP, 2020. 2, 5, 7, 8 +[36] Drew A Hudson, Daniel Zoran, Mateusz Malinowski, Andrew K Lampinen, Andrew Jaegle, James L McClelland, Loic Matthey, Felix Hill, and Alexander Lerchner. SODA: Bottleneck Diffusion Models for Representation Learning. In CVPR, 2024. 2 +[37] HyeongJoo Hwang, Geon-Hyeong Kim, Seunghoon Hong, and Kee-Eung Kim. Variational Interaction Information Maximization for Cross-domain Disentanglement. In NeurIPS, 2020. 6 +[38] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In ECCV, 2022. 3 +[39] Tero Karras, Samuli Laine, and Timo Aila. A Style-Based Generator Architecture for Generative Adversarial Networks. In CVPR, 2019. 2 +[40] Bahjat Kawar, Shiran Zada, Oran Lang, Omer Tov, Huiwen Chang, Tali Dekel, Inbar Mosseri, and Michal Irani. Imagic: Text-Based Real Image Editing with Diffusion Models. In CVPR, 2023. 2 +[41] Kazuma Kobayashi and Lin Gu and Ryuichiro Hataya and Takaaki Mizuno and Mototaka Miyake and Hirokazu Watanabe and Masamichi Takahashi and Yasuyuki Takamizawa and Yukihiro Yoshida and Satoshi Nakamura and Nobuji Kouno and Amina Bolatkan and Yusuke Kurose and Tatsuya Harada and Ryuji Hamamoto. Sketch-based Medical Image Retrieval. Medical Image Analysis, 2024. 2 +[42] Diederik P Kingma and Max Welling. Auto-Encoding Variational Bayes. In ICLR, 2014. 3, 5 +[43] Subhadeep Koley, Ayan Kumar Bhunia, Aneeshan Sain, Pinaki Nath Chowdhury, Tao Xiang, and Yi-Zhe Song. Picture that Sketch: Photorealistic Image Generation from Abstract Sketches. In CVPR, 2023. 2, 3, 4 +[44] Subhadeep Koley, Ayan Kumar Bhunia, Aneeshan Sain, Pinaki Nath Chowdhury, Tao Xiang, and Yi-Zhe Song. You'll Never Walk Alone: A Sketch and Text Duet for Fine-Grained Image Retrieval. In CVPR, 2024. 2 +[45] Subhadeep Koley, Ayan Kumar Bhunia, Aneeshan Sain, Pinaki Nath Chowdhury, Tao Xiang, and Yi-Zhe Song. How to Handle Sketch-Abstraction in Sketch-Based Image Retrieval? In CVPR, 2024. 1, 3 + +[46] Subhadeep Koley, Ayan Kumar Bhunia, Aneeshan Sain, Pinaki Nath Chowdhury, Tao Xiang, and Yi-Zhe Song. Text-to-Image Diffusion Models are Great Sketch-Photo Matchmakers. In CVPR, 2024. 1, 2, 3, 4, 5, 6, 8 +[47] Subhadeep Koley, Ayan Kumar Bhunia, Deeptanshu Sekhri, Aneeshan Sain, Pinaki Nath Chowdhury, Tao Xiang, and Yi-Zhe Song. It's All About Your Sketch: Democratising Sketch Control in Diffusion Models. In CVPR, 2024. 1, 2, 4, 5 +[48] Philipp Krahenbuhl and Vladlen Koltun. Efficient Inference in Fully Connected CRFs with Gaussian Edge Potentials. In NeurIPS, 2011. 7 +[49] Alexander C Li, Mihir Prabhudesai, Shivam Duggal, Ellis Brown, and Deepak Pathak. Your Diffusion Model is Secretly a Zero-Shot Classifier. In ICCV, 2023. 2, 5 +[50] Ke Li, Kaiyue Pang, and Yi-Zhe Song. Photo Pre-Training, but for Sketch. In CVPR, 2023. 2, 3 +[51] Minchen Li, Alla Sheffer, Eitan Grinspun, and Nicholas Vining. FoldSketch: Enriching Garments with Physically Reproducible Folds. ACM TOG, 2018. 2 +[52] Fengyin Lin, Mingkang Li, Da Li, Timothy Hospedales, YiZhe Song, and Yonggang Qi. Zero-Shot Everything Sketch-Based Image Retrieval, and in Explainable Style. In CVPR, 2023. 3 +[53] Li Liu, Fumin Shen, Yuming Shen, Xianglong Liu, and Ling Shao. Deep Sketch Hashing: Fast Free-hand Sketch-Based Image Retrieval. In CVPR, 2017. 6 +[54] Qing Liu, Lingxi Xie, Huiyu Wang, and Alan Yuille. Semantic-Aware Knowledge Preservation for Zero-Shot Sketch-Based Image Retrieval. In ICCV, 2019. 6 +[55] Xuanchen Lu, Xiaolong Wang, and Judith E Fan. Learning Dense Correspondences between Photos and Sketches. In ICML, 2023. 5, 7, 8 +[56] Ling Luo, Yulia Gryaditskaya, Tao Xiang, and Yi-Zhe Song. Structure-Aware 3D VR Sketch to 3D Shape Retrieval. In 3DV, 2022. 2 +[57] Ling Luo, Pinaki Nath Chowdhury, Tao Xiang, Yi-Zhe Song, and Yulia Gryaditskaya. 3D VR Sketch Guided 3D Shape Prototyping and Exploration. In ICCV, 2023. 2 +[58] Aryan Mikaeili, Or Perel, Daniel Cohen-Or, and Ali Mahdavi-Amiri. SKED: Sketch-guided Text-based 3D Editing. In CVPR, 2023. 2 +[59] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, and Ying Shan. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. In AAAI, 2024. 2 +[60] Jishnu Mukhoti, Tsung-Yu Lin, Omid Poursaeed, Rui Wang, Ashish Shah, Philip HS Torr, and Ser-Nam Lim. Open Vocabulary Semantic Segmentation with Patch Aligned Contrastive Learning. In CVPR, 2023. 5 +[61] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. DINov2: Learning Robust Visual Features without Supervision. TMLR, 2024. 6, 8 +[62] Zixuan Pan, Jianxu Chen, and Yiyu Shi. Masked Diffusion as Self-supervised Representation Learner. arXiv preprint arXiv:2308.05695, 2023. 2 + +[63] Kaiyue Pang, Ke Li, Yongxin Yang, Honggang Zhang, Timothy M Hospedales, Tao Xiang, and Yi-Zhe Song. Generalising Fine-Grained Sketch-Based Image Retrieval. In CVPR, 2019. 6 +[64] Kaiyue Pang, Yongxin Yang, Timothy M Hospedales, Tao Xiang, and Yi-Zhe Song. Solving mixed-modal jigsaw puzzle for fine-grained sketch-based image retrieval. In CVPR, 2020. 2, 3 +[65] Mihir Prabhudesai, Tsung-Wei Ke, Alexander Cong Li, Deepak Pathak, and Katerina Fragkiadaki. Diffusion-TTA: Test-time Adaptation of Discriminative Models via Generative Feedback. In NeurIPS, 2023. 2 +[66] Zhiyu Qu, Yulia Gryaditskaya, Ke Li, Kaiyue Pang, Tao Xiang, and Yi-Zhe Song. SketchXAI: A First Look at Explainability for Human Sketches. In CVPR, 2023. 6, 7, 8 +[67] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning Transferable Visual Models From Natural Language Supervision. In ICML, 2021. 1, 2, 3, 4, 5, 6, 7, 8 +[68] Leo Sampaio Ferraz Ribeiro, Tu Bui, John Collomosse, and Moacir Ponti. Sketchformer: Transformer-based Representation for Sketched Structure. In CVPR, 2020. 2, 3, 6 +[69] Elad Richardson, Yuval Alaluf, Or Patashnik, Yotam Nitzan, Yaniv Azar, Stav Shapiro, and Daniel Cohen-Or. Encoding in Style: a StyleGAN Encoder for Image-to-Image Translation. In CVPR, 2021. 2 +[70] Ignacio Rocco, Relja Arandjelović, and Josef Sivic. End-to-end weakly-supervised semantic alignment. In CVPR, 2018. 7 +[71] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-Resolution Image Synthesis with Latent Diffusion Models. In CVPR, 2022. 1, 2, 3, 4, 5, 6, 8 +[72] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-Net: Convolutional Networks for Biomedical Image Segmentation. In MICCAI, 2015. 3 +[73] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. DreamBooth: Fine Tuning Text-to-Image Diffusion Models for Subject-Driven Generation. In CVPR, 2023. 2 +[74] Aneeshan Sain, Ayan Kumar Bhunia, Yongxin Yang, Tao Xiang, and Yi-Zhe Song. StyleMeUp: Towards Style-Agnostic Sketch-Based Image Retrieval. In CVPR, 2021. 3 +[75] Aneeshan Sain, Ayan Kumar Bhunia, Vaishnav Potlapalli, Pinaki Nath Chowdhury, Tao Xiang, and Yi-Zhe Song. Sketch3T: Test-Time Training for Zero-Shot SBIR. In CVPR, 2022. 3 +[76] Aneeshan Sain, Ayan Kumar Bhunia, Pinaki Nath Chowdhury, Aneeshan Sain, Subhadeep Koley, Tao Xiang, and Yi-Zhe Song. CLIP for All Things Zero-Shot Sketch-Based Image Retrieval, Fine-Grained or Not. In CVPR, 2023. 1, 2, 3, 5, 6, 8 +[77] Aneeshan Sain, Ayan Kumar Bhunia, Subhadeep Koley, Pinaki Nath Chowdhury, Soumitri Chattopadhyay, Tao Xiang, and Yi-Zhe Song. Exploiting Unlabelled Photos for Stronger Fine-Grained SBIR. In CVPR, 2023. 3 + +[78] Patsorn Sangkloy, Nathan Burnell, Cusuh Ham, and James Hays. The Sketchy Database: Learning to Retrieve Badly Drawn Bunnies. ACM TOG, 2016. 1, 5, 6, 7 +[79] Patsorn Sangkloy, Wittawat Jitkrittum, Diyi Yang, and James Hays. A Sketch Is Worth a Thousand Words: Image Retrieval with Text and Sketch. In ECCV, 2022. 2, 3, 4, 5 +[80] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. LAION-5B: An open large-scale dataset for training next generation image-text models. In NeurlPSW, 2022. 4 +[81] Jaehyeok Shim, Changwoo Kang, and Kyungdon Joo. Diffusion-Based Signed Distance Fields for 3D Shape Generation. In CVPR, 2023. 2 +[82] Chenyang Si, Ziqi Huang, Yuming Jiang, and Ziwei Liu. FreeU: Free Lunch in Diffusion U-Net. In CVPR, 2024. 4 +[83] Jifei Song, Qian Yu, Yi-Zhe Song, Tao Xiang, and Timothy M Hospedales. Deep Spatial-Semantic Attention for Fine-Grained Sketch-Based Image Retrieval. In ICCV, 2017. 3 +[84] Luming Tang, Menglin Jia, Qianqian Wang, Cheng Perng Phoo, and Bharath Hariharan. Emergent Correspondence from Image Diffusion. In NeurIPS, 2023. 1, 2, 4, 5, 8 +[85] Prune Truong, Martin Danelljan, Fisher Yu, and Luc Van Gool. Warp Consistency for Unsupervised Learning of Dense Correspondences. In ICCV, 2021. 7, 8 +[86] Narek Tumanyan, Michal Geyer, Shai Bagon, and Tali Dekel. Plug-and-Play Diffusion Features for Text-Driven Image-to-Image Translation. In CVPR, 2023. 2 +[87] Andrey Voynov, Kfir Aberman, and Daniel Cohen-Or. Sketch-guided text-to-image diffusion models. In SIGGRAPH, 2023. 2 +[88] Fei Wang. Sketch2Vox: Learning 3D Reconstruction from a Single Monocular Sketch Image. In ECCV, 2024. 2 +[89] Wenxuan Wang, Quan Sun, Fan Zhang, Yepeng Tang, Jing Liu, and Xinlong Wang. Diffusion Feedback Helps CLIP See Better. In ICLR, 2025. 3 +[90] Size Wu, Wenwei Zhang, Lumin Xu, Sheng Jin, Xiangtai Li, Wentao Liu, and Chen Change Loy. CLIPSelf: Vision Transformer Distills Itself for Open-Vocabulary Dense Prediction. In ICLR, 2024. 3 +[91] Chunlong Xia, Xinliang Wang, Feng Lv, Xin Hao, and Yifeng Shi. ViT-CoMer: Vision Transformer with Convolutional Multi-scale Feature Interaction for Dense Predictions. In CVPR, 2024. 3 +[92] Jinbo Xing, Hanyuan Liu, Menghan Xia, Yong Zhang, Xintao Wang, Ying Shan, and Tien-Tsin Wong. ToonCrafter: Generative Cartoon Interpolation. ACM TOG, 2024. 2 +[93] Chenfeng Xu, Huan Ling, Sanja Fidler, and Or Litany. 3DiffTec: 3D Object Detection with Geometry-Aware Diffusion Features. In CVPR, 2023. 2, 4, 5 +[94] Jiarui Xu, Sifei Liu, Arash Vahdat, Wonmin Byeon, Xiaolong Wang, and Shalini De Mello. Open-Vocabulary Panoptic Segmentation with Text-to-Image Diffusion Models. In CVPR, 2023. 1, 2, 4 + +[95] Mengde Xu, Zheng Zhang, Fangyun Wei, Yutong Lin, Yue Cao, Han Hu, and Xiang Bai. A Simple Baseline for Open-Vocabulary Semantic Segmentation with Pre-trained Vision-language Model. In ECCV, 2022. 7, 8 +[96] Peng Xu, Yongye Huang, Tongtong Yuan, Kaiyue Pang, YiZhe Song, Tao Xiang, Timothy M Hospedales, Zhanyu Ma, and Jun Guo. SketchMate: Deep Hashing for Million-Scale Human Sketch Retrieval. In CVPR, 2018. 6, 7, 8 +[97] Peng Xu, Timothy Hospedales, Qiyue Yin, Yi-Zhe Song, Tao Xiang, and Liang Wang. Deep Learning for Free-Hand Sketch: A Survey. IEEE TPAMI, 2020. 1 +[98] Antoine Yang, Arsha Nagrani, Paul Hongsuck Seo, Antoine Miech, Jordi Pont-Tuset, Ivan Laptev, Josef Sivic, and Cordelia Schmid. Vid2Seq: Large-Scale Pretraining of a Visual Language Model for Dense Video Captioning. In CVPR, 2023. 5 +[99] Kai Yang, Jian Tao, Jiafei Lyu, Chunjiang Ge, Jiaxin Chen, Weihan Shen, Xiaolong Zhu, and Xiu Li. Using human feedback to fine-tune diffusion models without any reward model. In CVPR, 2024. 3 +[100] Lan Yang, Kaiyue Pang, Honggang Zhang, and Yi-Zhe Song. SketchAA: Abstract Representation for Abstract Sketches. In ICCV, 2021. 1, 5 +[101] Lumin Yang, Jiajie Zhuang, Hongbo Fu, Xiangzhi Wei, Kun Zhou, and Youyi Zheng. SketchGNN: Semantic Sketch Segmentation with Graph Neural Networks. ACM TOG, 2021. 6, 7, 8 +[102] Sasi Kiran Yelamarthi, Shiva Krishna Reddy, Ashish Mishra, and Anurag Mittal. A Zero-Shot Framework for Sketch Based Image Retrieval. In ECCV, 2018. 2, 6 +[103] Dong Yin, Raphael Gontijo Lopes, Jon Shlens, Ekin Dogus Cubuk, and Justin Gilmer. A Fourier Perspective on Model Robustness in Computer Vision. In NeurIPS, 2019. 4 +[104] Jiazuo Yu, Yunzhi Zhuge, Lu Zhang, Ping Hu, Dong Wang, Huchuan Lu, and You He. Boosting Continual Learning of Vision-Language Models via Mixture-of-Experts Adapters. In CVPR, 2024. 3 +[105] Seonghoon Yu, Paul Hongsuck Seo, and Jeany Son. Zeroshot Referring Image Segmentation with Global-Local Context Features. In CVPR, 2023. 3 +[106] Junyi Zhang, Charles Herrmann, Junhwa Hur, Eric Chen, Varun Jampani, Deqing Sun, and Ming-Hsuan Yang. Telling left from right: Identifying geometry-aware semantic correspondence. In CVPR, 2024. 7 +[107] Junyi Zhang, Charles Herrmann, Junhwa Hur, Luisa Polania Cabrera, Varun Jampani, Deqing Sun, and Ming-Hsuan Yang. A Tale of Two Features: Stable Diffusion Complements DINO for Zero-Shot Semantic Correspondence. NeurIPS, 2024. 1, 2, 3, 4, 8 +[108] Lvmin Zhang and Maneesh Agrawala. Adding Conditional Control to Text-to-Image Diffusion Models. In ICCV, 2023. 2, 3 +[109] Manyuan Zhang, Guanglu Song, Xiaoyu Shi, Yu Liu, and Hongsheng Li. Three Things We Need to Know About Transferring Stable Diffusion to Visual Dense Prediction Tasks. In ECCV, 2024. 2 + +[110] Wenliang Zhao, Yongming Rao, Zuyan Liu, Benlin Liu, Jie Zhou, and Jiwen Lu. Unleashing Text-to-Image Diffusion Models for Visual Perception. In ICCV, 2023. 1, 2, 4 +[111] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Zwei Liu. Conditional Prompt Learning for Vision-Language Models. In CVPR, 2022. 3 +[112] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Zwei Liu. Learning to Prompt for Vision-Language Models. IJCV, 2022. 3 \ No newline at end of file diff --git a/2025/SketchFusion_ Learning Universal Sketch Features through Fusing Foundation Models/images.zip b/2025/SketchFusion_ Learning Universal Sketch Features through Fusing Foundation Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..cde4288d7af00c8feae9702c3af998caaea86f1b --- /dev/null +++ b/2025/SketchFusion_ Learning Universal Sketch Features through Fusing Foundation Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7347ae52d1963bb0f679a93d89523388ea53e92b9cfb8a50e9a77b79126a88fa +size 358642 diff --git a/2025/SketchFusion_ Learning Universal Sketch Features through Fusing Foundation Models/layout.json b/2025/SketchFusion_ Learning Universal Sketch Features through Fusing Foundation Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..924286916ae45852dd1f440d504f29044340e97b --- /dev/null +++ b/2025/SketchFusion_ Learning Universal Sketch Features through Fusing Foundation Models/layout.json @@ -0,0 +1,12535 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 121, + 75, + 489, + 109 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 75, + 489, + 109 + ], + "spans": [ + { + "bbox": [ + 121, + 75, + 489, + 109 + ], + "type": "text", + "content": "SketchFusion: Learning Universal Sketch Features through Fusing Foundation Models" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 89, + 110, + 520, + 138 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 110, + 520, + 138 + ], + "spans": [ + { + "bbox": [ + 89, + 110, + 520, + 138 + ], + "type": "text", + "content": "Subhadeep Koley" + }, + { + "bbox": [ + 89, + 110, + 520, + 138 + ], + "type": "inline_equation", + "content": "^{1,2*}" + }, + { + "bbox": [ + 89, + 110, + 520, + 138 + ], + "type": "text", + "content": " Tapas Kumar Dutta" + }, + { + "bbox": [ + 89, + 110, + 520, + 138 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 89, + 110, + 520, + 138 + ], + "type": "text", + "content": " Aneeshan Sain" + }, + { + "bbox": [ + 89, + 110, + 520, + 138 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 89, + 110, + 520, + 138 + ], + "type": "text", + "content": " Pinaki Nath Chowdhury" + }, + { + "bbox": [ + 89, + 110, + 520, + 138 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 89, + 110, + 520, + 138 + ], + "type": "text", + "content": " Ayan Kumar Bhunia" + }, + { + "bbox": [ + 89, + 110, + 520, + 138 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 89, + 110, + 520, + 138 + ], + "type": "text", + "content": " Yi-Zhe Song" + }, + { + "bbox": [ + 89, + 110, + 520, + 138 + ], + "type": "inline_equation", + "content": "^{1,2}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 164, + 138, + 446, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 164, + 138, + 446, + 152 + ], + "spans": [ + { + "bbox": [ + 164, + 138, + 446, + 152 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 164, + 138, + 446, + 152 + ], + "type": "text", + "content": "SketchX, CVSSP, University of Surrey, United Kingdom." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 149, + 152, + 460, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 152, + 460, + 167 + ], + "spans": [ + { + "bbox": [ + 149, + 152, + 460, + 167 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 149, + 152, + 460, + 167 + ], + "type": "text", + "content": "iFlyTek-Surrey Joint Research Centre on Artificial Intelligence." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 138, + 168, + 468, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 168, + 468, + 180 + ], + "spans": [ + { + "bbox": [ + 138, + 168, + 468, + 180 + ], + "type": "text", + "content": "{s.koley, a.sain, p.chowdhury, a.bhunia, y-song}@surrey.ac.uk" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 56, + 182, + 219, + 279 + ], + "blocks": [ + { + "bbox": [ + 56, + 182, + 219, + 279 + ], + "lines": [ + { + "bbox": [ + 56, + 182, + 219, + 279 + ], + "spans": [ + { + "bbox": [ + 56, + 182, + 219, + 279 + ], + "type": "image", + "image_path": "fbf3bc9cc8e26a398b9b324416719f6803ea8f180d4fd170aa1f21e0d30874c3.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 284, + 555, + 416 + ], + "lines": [ + { + "bbox": [ + 55, + 284, + 555, + 416 + ], + "spans": [ + { + "bbox": [ + 55, + 284, + 555, + 416 + ], + "type": "text", + "content": "Figure 1. (Left): Apart from high-resolution image generation, text-to-image diffusion models (e.g., Stable diffusion (SD) [71]) with their innate object understanding capability [84, 107], have shown remarkable performance across a wide range of image-based vision tasks (e.g., segmentation [94], depth estimation [110], etc.). However, upon analysing the PCA representation of SD's intermediate UNet features, we observe that it struggles to achieve similar results when working with freehand abstract sketches (detail in Sec. 4). Unlike pixel-perfect photos, highly abstract freehand sketches are sparse and lack detailed textures and colours [26], making it harder for the SD model to extract meaningful features. Furthermore, investigating the SD denoising process in the frequency domain (via Fourier Transform), we observe the predominance of high-frequency (HF) components, rather than their low-frequency (LF) counterpart - crucial for capturing comprehensive semantic context. To mitigate this inherent bias within SD, we reinforce the diffusion process with another pretrained model (i.e., CLIP [67]) whose bias is complementary (i.e., focuses on LF) to SD. Consequently, the proposed extractor can extract semantically meaningful and accurate features from both sketches and photos, that encapsulate a broader frequency spectrum (i.e., HF and LF). (Right:) Testing the proposed method with different sketch-based discriminative and dense prediction tasks (requiring knowledge of both sketch and image), we find a marked improvement over baseline SD+CLIP hybrid feature extractor." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 221, + 182, + 417, + 278 + ], + "blocks": [ + { + "bbox": [ + 221, + 182, + 417, + 278 + ], + "lines": [ + { + "bbox": [ + 221, + 182, + 417, + 278 + ], + "spans": [ + { + "bbox": [ + 221, + 182, + 417, + 278 + ], + "type": "image", + "image_path": "8f30d00fa401bee519fd692a7dee43e1a0682d97cbfdfefa69195f81ae3ae1e6.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 420, + 182, + 554, + 279 + ], + "blocks": [ + { + "bbox": [ + 420, + 182, + 554, + 279 + ], + "lines": [ + { + "bbox": [ + 420, + 182, + 554, + 279 + ], + "spans": [ + { + "bbox": [ + 420, + 182, + 554, + 279 + ], + "type": "image", + "image_path": "5350e434ef008889b2e48186422c566193ca212b2f1868ef2cffe5ce4a90c48c.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 151, + 422, + 200, + 434 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 422, + 200, + 434 + ], + "spans": [ + { + "bbox": [ + 151, + 422, + 200, + 434 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 53, + 449, + 297, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 449, + 297, + 687 + ], + "spans": [ + { + "bbox": [ + 53, + 449, + 297, + 687 + ], + "type": "text", + "content": "While foundation models have revolutionised computer vision, their effectiveness for sketch understanding remains limited by the unique challenges of abstract, sparse visual inputs. Through systematic analysis, we uncover two fundamental limitations: Stable Diffusion (SD) struggles to extract meaningful features from abstract sketches (unlike its success with photos), and exhibits a pronounced frequency-domain bias that suppresses essential low-frequency components needed for sketch understanding. Rather than costly retraining, we address these limitations by strategically combining SD with CLIP, whose strong semantic understanding naturally compensates for SD's spatial-frequency biases. By dynamically injecting CLIP features into SD's denoising process and adaptively aggregating features across semantic levels, our method achieves state-of-the-art performance in sketch retrieval " + }, + { + "bbox": [ + 53, + 449, + 297, + 687 + ], + "type": "inline_equation", + "content": "(+3.35\\%)" + }, + { + "bbox": [ + 53, + 449, + 297, + 687 + ], + "type": "text", + "content": ", recognition " + }, + { + "bbox": [ + 53, + 449, + 297, + 687 + ], + "type": "inline_equation", + "content": "(+1.06\\%)" + }, + { + "bbox": [ + 53, + 449, + 297, + 687 + ], + "type": "text", + "content": ", segmentation " + }, + { + "bbox": [ + 53, + 449, + 297, + 687 + ], + "type": "inline_equation", + "content": "(+29.42\\%)" + }, + { + "bbox": [ + 53, + 449, + 297, + 687 + ], + "type": "text", + "content": ", and correspondence learning " + }, + { + "bbox": [ + 53, + 449, + 297, + 687 + ], + "type": "inline_equation", + "content": "(+21.22\\%)" + }, + { + "bbox": [ + 53, + 449, + 297, + 687 + ], + "type": "text", + "content": ", demonstrating the first truly universal sketch feature representation in the era of foundation models." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 422, + 394, + 434 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 422, + 394, + 434 + ], + "spans": [ + { + "bbox": [ + 314, + 422, + 394, + 434 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 312, + 437, + 555, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 437, + 555, + 556 + ], + "spans": [ + { + "bbox": [ + 312, + 437, + 555, + 556 + ], + "type": "text", + "content": "The quest for universal sketch-specific features has been a foundational challenge in computer vision [46, 76, 78, 97]. Sketches, with their inherent characteristics of abstraction, sparsity, and cross-modal interpretation, demand fundamentally different feature representations compared to natural images [45, 100]. While foundation models [67, 71] have revolutionised various visual tasks, their effectiveness for sketch understanding remains largely unexplored [47], particularly given sketch's unique position at the intersection of human abstraction and computer vision [26, 45]." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 312, + 558, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 558, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 312, + 558, + 556, + 715 + ], + "type": "text", + "content": "Through systematic analysis, we find that even powerful foundation models like Stable Diffusion (SD) [71] face significant challenges with sketch data. Our pilot studies (Sec. 4) reveal two fundamental limitations. First, while SD excels at extracting task-specific representative features from pixel-perfect RGB images, it struggles with abstract freehand sketches, likely due to its large-scale pretraining being predominantly focused on photographic data [71]. Second, and more intriguingly, frequency-domain analysis of SD's UNet reveals an inherent architectural bias: the model systematically promotes high-frequency (HF) components while suppressing low-frequency (LF) information - a characteristic that proves particularly problematic" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 703, + 133, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 133, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 133, + 712 + ], + "type": "text", + "content": "*Equal contribution" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2556" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 96 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 96 + ], + "type": "text", + "content": "for sketch understanding tasks, especially dense prediction problems like segmentation and correspondence learning." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 98, + 295, + 253 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 98, + 295, + 253 + ], + "spans": [ + { + "bbox": [ + 54, + 98, + 295, + 253 + ], + "type": "text", + "content": "These findings present a significant challenge. Retraining SD on sketch data risks catastrophic forgetting of its valuable pretrained knowledge [47], while developing sketch-specific architectures would fail to leverage the rich visual understanding already present in foundation models. Our key insight comes from analysing the complementary nature of different foundation models: while SD's features are spatially-aware but lack semantic accuracy, CLIP's [67] visual encoder exhibits strong semantic understanding even at the cost of spatial precision. This complementarity suggests that strategic combination [107] of these models could yield features that capture both the detailed structure and semantic meaning necessary for sketch understanding." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 255, + 295, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 255, + 295, + 386 + ], + "spans": [ + { + "bbox": [ + 55, + 255, + 295, + 386 + ], + "type": "text", + "content": "Based on this insight, we propose a novel framework that leverages the complementary strengths of these foundation models. Our method systematically injects CLIP's visual features into multiple layers of SD's UNet during the denoising process, using lightweight " + }, + { + "bbox": [ + 55, + 255, + 295, + 386 + ], + "type": "inline_equation", + "content": "1D" + }, + { + "bbox": [ + 55, + 255, + 295, + 386 + ], + "type": "text", + "content": " convolutions to influence feature extraction at various semantic levels. PCA analysis of the resulting features demonstrates that this combination successfully captures both the HF spatial details from SD [71] and the LF semantic components from CLIP [67], producing rich, semantically-aware representations suited for diverse sketch-based tasks." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 388, + 295, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 388, + 295, + 508 + ], + "spans": [ + { + "bbox": [ + 55, + 388, + 295, + 508 + ], + "type": "text", + "content": "Moreover, we observe that different layers of SD's UNet capture features at varying levels of semantic-granularity [46], with different combinations better suited to specific downstream tasks [84]. Rather than manually selecting optimal features, we train a dynamic feature aggregator that automatically determines the most effective combination for any given task, eliminating the need for task-specific tuning. This adaptive approach ensures optimal performance across diverse sketch-based applications without requiring manual layer selection or task-specific optimisation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 509, + 295, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 509, + 295, + 640 + ], + "spans": [ + { + "bbox": [ + 55, + 509, + 295, + 640 + ], + "type": "text", + "content": "Our extensive experiments demonstrate the effectiveness of this approach across a comprehensive suite of sketch-based tasks: (i) Sketch Recognition: " + }, + { + "bbox": [ + 55, + 509, + 295, + 640 + ], + "type": "inline_equation", + "content": "+1.06\\%" + }, + { + "bbox": [ + 55, + 509, + 295, + 640 + ], + "type": "text", + "content": " improvement over state-of-the-art, (ii) Sketch-based Image Retrieval: " + }, + { + "bbox": [ + 55, + 509, + 295, + 640 + ], + "type": "inline_equation", + "content": "+3.35\\%" + }, + { + "bbox": [ + 55, + 509, + 295, + 640 + ], + "type": "text", + "content": " improvement in retrieval accuracy, (iii) Sketch-based Segmentation: " + }, + { + "bbox": [ + 55, + 509, + 295, + 640 + ], + "type": "inline_equation", + "content": "+29.42\\%" + }, + { + "bbox": [ + 55, + 509, + 295, + 640 + ], + "type": "text", + "content": " boost in mean IoU, (iv) Sketch-Photo Correspondence: " + }, + { + "bbox": [ + 55, + 509, + 295, + 640 + ], + "type": "inline_equation", + "content": "+21.22\\%" + }, + { + "bbox": [ + 55, + 509, + 295, + 640 + ], + "type": "text", + "content": " increase in matching precision. These significant improvements across all major sketch-specific tasks suggest that our method has uncovered fundamental principles for extracting effective sketch features from foundation models." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "type": "text", + "content": "In summary, our contributions are: (i) A systematic analysis revealing the limitations of pretrained foundation models for sketch understanding, supported by both empirical results and frequency-domain analysis, (ii) A novel method for combining complementary foundation models that address these limitations without retraining, (iii) An adap" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": "tive feature aggregation approach that automatically optimises representations for different downstream tasks, and " + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "inline_equation", + "content": "(iv)" + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": " Comprehensive experimental validation demonstrating substantial improvements across all major sketch-specific tasks, establishing a new paradigm for universal sketch feature learning." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 148, + 406, + 161 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 148, + 406, + 161 + ], + "spans": [ + { + "bbox": [ + 313, + 148, + 406, + 161 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 163, + 553, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 163, + 553, + 462 + ], + "spans": [ + { + "bbox": [ + 313, + 163, + 553, + 462 + ], + "type": "text", + "content": "Diffusion Model for Representation Learning. Diffusion models [34] have now become the de-facto standard for high-resolution " + }, + { + "bbox": [ + 313, + 163, + 553, + 462 + ], + "type": "inline_equation", + "content": "2D" + }, + { + "bbox": [ + 313, + 163, + 553, + 462 + ], + "type": "text", + "content": " [21, 71] and " + }, + { + "bbox": [ + 313, + 163, + 553, + 462 + ], + "type": "inline_equation", + "content": "3D" + }, + { + "bbox": [ + 313, + 163, + 553, + 462 + ], + "type": "text", + "content": " [12, 81] image and video [11] generation. Apart from generation, it gave rise to various image editing frameworks like Prompt-to-Prompt [33], Imagic [40], Dreambooth [73], etc. The success of diffusion models across various generative tasks and their inherent zero-shot generalisability has led to their adoption as efficient feature extractors in different forms of discriminative [23] and dense prediction [109] tasks like image retrieval [46], classification [49], semantic [3] and panoptic [94] segmentation, depth estimation [110], object detection [93], image-to-image translation [47, 86], controllable-generation [59, 108] and medical imaging [18], to name a few. A few recent works have developed specialised diffusion models together with carefully crafted task-specific designs like test-time adaptation [65], masked-reconstruction [62], or compact latent generation [36] to improve their representation ability. Nonetheless, our pilot study (Sec. 4) reveals that SD, often struggles to extract representative features (Fig. 1 left) from binary and sparse sketches, hampering downstream task performance. In this work, we leverage a pretrained CLIP [67] visual encoder to enhance SD's feature representation abilities for sketch-based discriminative and dense prediction tasks." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 463, + 553, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 463, + 553, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 463, + 553, + 653 + ], + "type": "text", + "content": "Sketch-based Vision Applications. \"Freehand sketch\" has now established itself as a worthy interactive alternative modality to text for downstream vision tasks [9, 16, 31, 43, 76, 87, 102]. With its success in sketch-based image retrieval (SBIR) [13, 20, 76, 79], sketches are now being used in a wide array of discriminative [16, 17, 20], generative [27, 43, 108] and dense prediction [7, 35] tasks like controllable " + }, + { + "bbox": [ + 313, + 463, + 553, + 653 + ], + "type": "inline_equation", + "content": "2D" + }, + { + "bbox": [ + 313, + 463, + 553, + 653 + ], + "type": "text", + "content": " [47, 108], " + }, + { + "bbox": [ + 313, + 463, + 553, + 653 + ], + "type": "inline_equation", + "content": "3D" + }, + { + "bbox": [ + 313, + 463, + 553, + 653 + ], + "type": "text", + "content": " [2, 88] image generation and editing [31, 43, 58, 59, 69, 108], video synthesis [92], garment designing [14, 51], segmentation [35], creative content generation [9], object detection [16], salient region prediction [7], augmented and virtual reality applications [56, 57], medical imaging [41], class incremental learning [5], or composed retrieval [44]. In this paper, we aim to tackle some of the sketch-based downstream tasks by reinforcing the diffusion feature extractor via a pretrained CLIP [67] visual encoder." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 654, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 553, + 714 + ], + "type": "text", + "content": "Backbones for Sketch-based Vision Applications. Backbones for sketch-based vision applications can be broadly categorised as - (i) pretrained CNN or Vision Transformer (ViT) [6, 13, 68], (ii) Self-supervised pretraining [4, 50, 64], and (iii) pretrained foundation models [39, 67, 71]. Starting" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2557" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 312 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 312 + ], + "type": "text", + "content": "with ImageNet pretrained CNNs [6, 13] or JFT pretrained ViTs [68, 77] with carefully-crafted design choices (e.g., cross [52] and spatial [83] attention, reinforcement learning [6], test-time training [75], meta-learning [74], knowledge distillation [77], etc.), the community moved towards self-supervised pretraining either via solving pretext tasks like sketch vectorisation [4], rasterisation [4], solving sketch-photo jigsaw [64] or by large-scale photo pretraining [50]. With the advent of large-scale foundation models, several works [43, 46, 76, 79] used them as a backbone for different sketch-based tasks like image generation [43, 108], shape synthesis [2], or retrieval [46, 79]. Foundation models have now become the gold standard for sketch-based tasks due to their strong zero-shot performance and extensive pretrained knowledge [45, 46, 79]. Among these, generative models like SD are spatially aware [46, 107], yet less accurate for dense prediction tasks. In contrast, vision-language models like CLIP [67] offer accurate features but are inherently sparse [107]. In this paper, we aim to get the best of both worlds by enhancing the SD feature extraction with CLIP." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 312, + 295, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 312, + 295, + 586 + ], + "spans": [ + { + "bbox": [ + 55, + 312, + 295, + 586 + ], + "type": "text", + "content": "Reinforcing Feature Extractors. There are numerous ways of reinforcing an off-the-shelf pretrained feature extractor to adapt them for different downstream tasks. It can broadly be categorised as " + }, + { + "bbox": [ + 55, + 312, + 295, + 586 + ], + "type": "inline_equation", + "content": "(-i)" + }, + { + "bbox": [ + 55, + 312, + 295, + 586 + ], + "type": "text", + "content": " Mixture of Experts (MoE) [22, 104], " + }, + { + "bbox": [ + 55, + 312, + 295, + 586 + ], + "type": "inline_equation", + "content": "(ii)" + }, + { + "bbox": [ + 55, + 312, + 295, + 586 + ], + "type": "text", + "content": " Prompt learning [38, 112], " + }, + { + "bbox": [ + 55, + 312, + 295, + 586 + ], + "type": "inline_equation", + "content": "(iii)" + }, + { + "bbox": [ + 55, + 312, + 295, + 586 + ], + "type": "text", + "content": " Hybrid models [91, 107], " + }, + { + "bbox": [ + 55, + 312, + 295, + 586 + ], + "type": "inline_equation", + "content": "(iv)" + }, + { + "bbox": [ + 55, + 312, + 295, + 586 + ], + "type": "text", + "content": " Finetuning [76, 99], and " + }, + { + "bbox": [ + 55, + 312, + 295, + 586 + ], + "type": "inline_equation", + "content": "(v)" + }, + { + "bbox": [ + 55, + 312, + 295, + 586 + ], + "type": "text", + "content": " Global-local feature merging [30, 105]. MoE combines multiple specialised sub-models (i.e., \"experts\"), each responsible for specific tasks or data patterns, to boost overall performance and adaptability [22]. In contrast, prompt learning deals with learning additional trainable components in the textual [111, 112] or pixel [1, 38] domain for task-specific adaptation. Hybrid models [89] combine two different kinds of models like CNN and ViT [91], or SD and DINO [107] to extract the best of both worlds. Another avenue deals with finetuning large-scale pretrained models with selective handcrafted data for specific tasks [16, 76]. Finally, a few other works [30, 90, 105] amalgamate global and local features from the same [90] or different [30, 105] models to improve downstream task performance. In this paper, we take the hybrid model approach and complement CNN-based SD features with transformer-based CLIP visual encoder features for different kinds of sketch-based vision tasks." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 594, + 141, + 605 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 141, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 141, + 605 + ], + "type": "text", + "content": "3. Preliminaries" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 610, + 170, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 610, + 170, + 623 + ], + "spans": [ + { + "bbox": [ + 55, + 610, + 170, + 623 + ], + "type": "text", + "content": "3.1. Background: CLIP" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 629, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 295, + 713 + ], + "type": "text", + "content": "Contrastive language-image pretraining (CLIP) [67] architecture consists of a visual and a textual encoder, trained on " + }, + { + "bbox": [ + 55, + 629, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\sim 400\\mathrm{M}" + }, + { + "bbox": [ + 55, + 629, + 295, + 713 + ], + "type": "text", + "content": " image-text pairs, with the goal of learning a shared embedding manifold. Using N-pair contrastive loss, the model aligns matching image-text pairs by maximising their cosine similarity and minimises it for random unpaired ones [67]. The image encoder, typically a" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 553, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 216 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 216 + ], + "type": "text", + "content": "ViT [24], encodes an image " + }, + { + "bbox": [ + 313, + 72, + 553, + 216 + ], + "type": "inline_equation", + "content": "\\mathbf{i}" + }, + { + "bbox": [ + 313, + 72, + 553, + 216 + ], + "type": "text", + "content": " as a feature vector " + }, + { + "bbox": [ + 313, + 72, + 553, + 216 + ], + "type": "inline_equation", + "content": "\\mathbf{V}(\\mathbf{i})\\in" + }, + { + "bbox": [ + 313, + 72, + 553, + 216 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^d" + }, + { + "bbox": [ + 313, + 72, + 553, + 216 + ], + "type": "text", + "content": ". Meanwhile, the textual encoder processes a word sequence " + }, + { + "bbox": [ + 313, + 72, + 553, + 216 + ], + "type": "inline_equation", + "content": "\\mathcal{W} = \\{w^0,w^1,\\dots ,w^k\\}" + }, + { + "bbox": [ + 313, + 72, + 553, + 216 + ], + "type": "text", + "content": " through byte pair encoding [67] and a learnable word embedding layer (vocabulary size of 49,152), converting each word " + }, + { + "bbox": [ + 313, + 72, + 553, + 216 + ], + "type": "inline_equation", + "content": "w^{i}" + }, + { + "bbox": [ + 313, + 72, + 553, + 216 + ], + "type": "text", + "content": " into a token embedding, thus creating an embedding sequence " + }, + { + "bbox": [ + 313, + 72, + 553, + 216 + ], + "type": "inline_equation", + "content": "\\mathcal{W}_e = \\{\\mathbf{w}_e^0,\\mathbf{w}_e^1,\\ldots ,\\mathbf{w}_e^k\\}" + }, + { + "bbox": [ + 313, + 72, + 553, + 216 + ], + "type": "text", + "content": " [67]. " + }, + { + "bbox": [ + 313, + 72, + 553, + 216 + ], + "type": "inline_equation", + "content": "\\mathcal{W}_e" + }, + { + "bbox": [ + 313, + 72, + 553, + 216 + ], + "type": "text", + "content": " is then passed through a transformer network " + }, + { + "bbox": [ + 313, + 72, + 553, + 216 + ], + "type": "inline_equation", + "content": "\\mathbf{T}(\\cdot)" + }, + { + "bbox": [ + 313, + 72, + 553, + 216 + ], + "type": "text", + "content": ", producing the final text feature " + }, + { + "bbox": [ + 313, + 72, + 553, + 216 + ], + "type": "inline_equation", + "content": "\\mathbf{T}(\\mathcal{W}_e)\\in \\mathbb{R}^d" + }, + { + "bbox": [ + 313, + 72, + 553, + 216 + ], + "type": "text", + "content": " [67]. Once trained, the CLIP model can be used in downstream tasks to enable zero-shot capability and cross-modal understanding by aligning visual and textual data in a shared embedding space [67]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 219, + 501, + 231 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 219, + 501, + 231 + ], + "spans": [ + { + "bbox": [ + 313, + 219, + 501, + 231 + ], + "type": "text", + "content": "3.2. Background: Stable Diffusion (SD)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 232, + 553, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 232, + 553, + 460 + ], + "spans": [ + { + "bbox": [ + 313, + 232, + 553, + 460 + ], + "type": "text", + "content": "Overview. The SD model [71] generates images by gradually removing noise from an initial " + }, + { + "bbox": [ + 313, + 232, + 553, + 460 + ], + "type": "inline_equation", + "content": "2D" + }, + { + "bbox": [ + 313, + 232, + 553, + 460 + ], + "type": "text", + "content": " isotropic Gaussian noise image through iterative refinement. It involves two reciprocal processes - \"forward\" and \"reverse\" diffusion [34, 71]. The forward process iteratively adds random Gaussian noise into a clean image " + }, + { + "bbox": [ + 313, + 232, + 553, + 460 + ], + "type": "inline_equation", + "content": "\\mathbf{i}_0 \\in \\mathbb{R}^{h \\times w \\times 3}" + }, + { + "bbox": [ + 313, + 232, + 553, + 460 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 313, + 232, + 553, + 460 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 232, + 553, + 460 + ], + "type": "text", + "content": " timesteps to generate a noisy image " + }, + { + "bbox": [ + 313, + 232, + 553, + 460 + ], + "type": "inline_equation", + "content": "\\mathbf{i}_t" + }, + { + "bbox": [ + 313, + 232, + 553, + 460 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 313, + 232, + 553, + 460 + ], + "type": "inline_equation", + "content": "\\mathbf{i}_t = \\sqrt{\\bar{\\alpha}_t} \\mathbf{i}_0 + \\sqrt{1 - \\bar{\\alpha}_t} \\epsilon" + }, + { + "bbox": [ + 313, + 232, + 553, + 460 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 313, + 232, + 553, + 460 + ], + "type": "inline_equation", + "content": "\\epsilon \\sim \\mathcal{N}(0, \\mathbf{I})" + }, + { + "bbox": [ + 313, + 232, + 553, + 460 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 232, + 553, + 460 + ], + "type": "inline_equation", + "content": "\\{\\alpha_t\\}_1^T" + }, + { + "bbox": [ + 313, + 232, + 553, + 460 + ], + "type": "text", + "content": " is a pre-defined noise schedule where " + }, + { + "bbox": [ + 313, + 232, + 553, + 460 + ], + "type": "inline_equation", + "content": "\\bar{\\alpha}_t = \\prod_{k=1}^t \\alpha_k" + }, + { + "bbox": [ + 313, + 232, + 553, + 460 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 313, + 232, + 553, + 460 + ], + "type": "inline_equation", + "content": "t \\sim U(0, T)" + }, + { + "bbox": [ + 313, + 232, + 553, + 460 + ], + "type": "text", + "content": ". The reverse process involves training a denoising UNet [72] " + }, + { + "bbox": [ + 313, + 232, + 553, + 460 + ], + "type": "inline_equation", + "content": "\\mathbf{U}_{\\theta}" + }, + { + "bbox": [ + 313, + 232, + 553, + 460 + ], + "type": "text", + "content": " (with an " + }, + { + "bbox": [ + 313, + 232, + 553, + 460 + ], + "type": "inline_equation", + "content": "l_2" + }, + { + "bbox": [ + 313, + 232, + 553, + 460 + ], + "type": "text", + "content": " objective), which estimates the input noise " + }, + { + "bbox": [ + 313, + 232, + 553, + 460 + ], + "type": "inline_equation", + "content": "\\epsilon \\approx \\mathbf{U}_{\\theta}(\\mathbf{i}_t, t)" + }, + { + "bbox": [ + 313, + 232, + 553, + 460 + ], + "type": "text", + "content": " from the noisy image " + }, + { + "bbox": [ + 313, + 232, + 553, + 460 + ], + "type": "inline_equation", + "content": "\\mathbf{i}_t" + }, + { + "bbox": [ + 313, + 232, + 553, + 460 + ], + "type": "text", + "content": " at each " + }, + { + "bbox": [ + 313, + 232, + 553, + 460 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 232, + 553, + 460 + ], + "type": "text", + "content": ". After training, " + }, + { + "bbox": [ + 313, + 232, + 553, + 460 + ], + "type": "inline_equation", + "content": "\\mathbf{U}_{\\theta}" + }, + { + "bbox": [ + 313, + 232, + 553, + 460 + ], + "type": "text", + "content": " is capable of reconstructing a clean image from a noisy input [71]. The inference process begins with a random " + }, + { + "bbox": [ + 313, + 232, + 553, + 460 + ], + "type": "inline_equation", + "content": "2D" + }, + { + "bbox": [ + 313, + 232, + 553, + 460 + ], + "type": "text", + "content": " noise sample, " + }, + { + "bbox": [ + 313, + 232, + 553, + 460 + ], + "type": "inline_equation", + "content": "\\mathbf{i}_T \\sim \\mathcal{N}(0, \\mathbf{I})" + }, + { + "bbox": [ + 313, + 232, + 553, + 460 + ], + "type": "text", + "content": ". The trained model " + }, + { + "bbox": [ + 313, + 232, + 553, + 460 + ], + "type": "inline_equation", + "content": "\\mathbf{U}_{\\theta}" + }, + { + "bbox": [ + 313, + 232, + 553, + 460 + ], + "type": "text", + "content": " is then applied iteratively across " + }, + { + "bbox": [ + 313, + 232, + 553, + 460 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 232, + 553, + 460 + ], + "type": "text", + "content": " timesteps, progressively reducing noise at each step to produce a cleaner intermediate image, " + }, + { + "bbox": [ + 313, + 232, + 553, + 460 + ], + "type": "inline_equation", + "content": "\\mathbf{i}_{t-1}" + }, + { + "bbox": [ + 313, + 232, + 553, + 460 + ], + "type": "text", + "content": ", finally generating one of the cleanest samples " + }, + { + "bbox": [ + 313, + 232, + 553, + 460 + ], + "type": "inline_equation", + "content": "\\mathbf{i}_0" + }, + { + "bbox": [ + 313, + 232, + 553, + 460 + ], + "type": "text", + "content": " from the target distribution." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 460, + 553, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 460, + 553, + 520 + ], + "spans": [ + { + "bbox": [ + 313, + 460, + 553, + 520 + ], + "type": "text", + "content": "The unconditional denoising diffusion could further be made conditional by governing " + }, + { + "bbox": [ + 313, + 460, + 553, + 520 + ], + "type": "inline_equation", + "content": "\\mathbf{U}_{\\theta}" + }, + { + "bbox": [ + 313, + 460, + 553, + 520 + ], + "type": "text", + "content": " with different auxiliary signals (e.g., textual captions) " + }, + { + "bbox": [ + 313, + 460, + 553, + 520 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 313, + 460, + 553, + 520 + ], + "type": "text", + "content": " [71]. Accordingly, " + }, + { + "bbox": [ + 313, + 460, + 553, + 520 + ], + "type": "inline_equation", + "content": "\\mathbf{U}_{\\theta}(\\mathbf{i}_t,t,\\mathbf{c})" + }, + { + "bbox": [ + 313, + 460, + 553, + 520 + ], + "type": "text", + "content": " can denoise " + }, + { + "bbox": [ + 313, + 460, + 553, + 520 + ], + "type": "inline_equation", + "content": "\\mathbf{i}_t" + }, + { + "bbox": [ + 313, + 460, + 553, + 520 + ], + "type": "text", + "content": " while being influenced by " + }, + { + "bbox": [ + 313, + 460, + 553, + 520 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 313, + 460, + 553, + 520 + ], + "type": "text", + "content": ", commonly via cross-attention [71]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 521, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 554, + 713 + ], + "type": "text", + "content": "Architecture. Pixel-space diffusion model [34] is time-intensive as it operates on the full image resolution (i.e., " + }, + { + "bbox": [ + 313, + 521, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{i}_0\\in \\mathbb{R}^{h\\times w\\times 3}" + }, + { + "bbox": [ + 313, + 521, + 554, + 713 + ], + "type": "text", + "content": "). Contrarily, SD model [71] performs denoising in the latent-space, resulting in significantly faster and more stable processing. In the first stage, SD trains a variational autoencoder (VAE) [42] (an encoder " + }, + { + "bbox": [ + 313, + 521, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{E}(\\cdot)" + }, + { + "bbox": [ + 313, + 521, + 554, + 713 + ], + "type": "text", + "content": " and a decoder " + }, + { + "bbox": [ + 313, + 521, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{D}(\\cdot)" + }, + { + "bbox": [ + 313, + 521, + 554, + 713 + ], + "type": "text", + "content": "). " + }, + { + "bbox": [ + 313, + 521, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{E}(\\cdot)" + }, + { + "bbox": [ + 313, + 521, + 554, + 713 + ], + "type": "text", + "content": " transforms the input image into its latent representation " + }, + { + "bbox": [ + 313, + 521, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_0 = \\mathbf{E}(\\mathbf{i}_0)\\in \\mathbb{R}^{\\frac{h}{8}\\times \\frac{w}{8}\\times d}" + }, + { + "bbox": [ + 313, + 521, + 554, + 713 + ], + "type": "text", + "content": ". In the second stage, the UNet [72] " + }, + { + "bbox": [ + 313, + 521, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{U}_{\\theta}" + }, + { + "bbox": [ + 313, + 521, + 554, + 713 + ], + "type": "text", + "content": ", directly denoises the latent images. " + }, + { + "bbox": [ + 313, + 521, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{U}_{\\theta}" + }, + { + "bbox": [ + 313, + 521, + 554, + 713 + ], + "type": "text", + "content": " comprises 12 encoding, 1 bottleneck, and 12 skip-connected decoding blocks [71]. Inside these encoding and decoding blocks, there are 4 downsampling " + }, + { + "bbox": [ + 313, + 521, + 554, + 713 + ], + "type": "inline_equation", + "content": "(\\mathbf{U}_{\\mathbf{d}}^{1 - 4})" + }, + { + "bbox": [ + 313, + 521, + 554, + 713 + ], + "type": "text", + "content": " and 4 upsampling " + }, + { + "bbox": [ + 313, + 521, + 554, + 713 + ], + "type": "inline_equation", + "content": "(\\mathbf{U}_{\\mathbf{u}}^{1 - 4})" + }, + { + "bbox": [ + 313, + 521, + 554, + 713 + ], + "type": "text", + "content": " layers respectively. In SD, a frozen CLIP textual encoder [67] " + }, + { + "bbox": [ + 313, + 521, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{T}(\\cdot)" + }, + { + "bbox": [ + 313, + 521, + 554, + 713 + ], + "type": "text", + "content": " converts the textual prompt " + }, + { + "bbox": [ + 313, + 521, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 313, + 521, + 554, + 713 + ], + "type": "text", + "content": " into token-sequence, that influences the denoising process of " + }, + { + "bbox": [ + 313, + 521, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{U}_{\\theta}" + }, + { + "bbox": [ + 313, + 521, + 554, + 713 + ], + "type": "text", + "content": " through cross-attention. " + }, + { + "bbox": [ + 313, + 521, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{U}_{\\theta}" + }, + { + "bbox": [ + 313, + 521, + 554, + 713 + ], + "type": "text", + "content": " is trained over an" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2558" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 71, + 294, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 71, + 294, + 144 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 294, + 144 + ], + "type": "inline_equation", + "content": "l_{2}" + }, + { + "bbox": [ + 55, + 71, + 294, + 144 + ], + "type": "text", + "content": " objective as: " + }, + { + "bbox": [ + 55, + 71, + 294, + 144 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{l}_2} = \\mathbb{E}_{\\mathbf{z}_t,t,\\mathbf{c},\\epsilon}(||\\epsilon -\\mathbf{U}_\\theta (\\mathbf{z}_t,t,\\mathbf{T}(\\mathbf{c}))||_2^2)" + }, + { + "bbox": [ + 55, + 71, + 294, + 144 + ], + "type": "text", + "content": ". During inference, a noisy latent " + }, + { + "bbox": [ + 55, + 71, + 294, + 144 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_t" + }, + { + "bbox": [ + 55, + 71, + 294, + 144 + ], + "type": "text", + "content": " is sampled directly as: " + }, + { + "bbox": [ + 55, + 71, + 294, + 144 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_t\\sim \\mathcal{N}(0,\\mathbf{I})" + }, + { + "bbox": [ + 55, + 71, + 294, + 144 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 55, + 71, + 294, + 144 + ], + "type": "inline_equation", + "content": "\\mathbf{U}_{\\theta}" + }, + { + "bbox": [ + 55, + 71, + 294, + 144 + ], + "type": "text", + "content": " removes noise from " + }, + { + "bbox": [ + 55, + 71, + 294, + 144 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_t" + }, + { + "bbox": [ + 55, + 71, + 294, + 144 + ], + "type": "text", + "content": " iteratively over " + }, + { + "bbox": [ + 55, + 71, + 294, + 144 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 71, + 294, + 144 + ], + "type": "text", + "content": " timesteps (conditioned on " + }, + { + "bbox": [ + 55, + 71, + 294, + 144 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 55, + 71, + 294, + 144 + ], + "type": "text", + "content": ") to produce a denoised latent image " + }, + { + "bbox": [ + 55, + 71, + 294, + 144 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{z}}_0\\in \\mathbb{R}^{\\frac{h}{8}\\times \\frac{w}{8}\\times d}" + }, + { + "bbox": [ + 55, + 71, + 294, + 144 + ], + "type": "text", + "content": ". The final high-resolution image is generated as: " + }, + { + "bbox": [ + 55, + 71, + 294, + 144 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{i}} = \\mathbf{D}(\\hat{\\mathbf{z}}_0)\\in \\mathbb{R}^{h\\times w\\times 3}" + }, + { + "bbox": [ + 55, + 71, + 294, + 144 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 148, + 251, + 162 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 148, + 251, + 162 + ], + "spans": [ + { + "bbox": [ + 55, + 148, + 251, + 162 + ], + "type": "text", + "content": "4. Pilot Study: Problems and Solution" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 296, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 296, + 509 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 296, + 509 + ], + "type": "text", + "content": "What's Wrong with Diffusion Features? Binary and sparse freehand sketches contain much lesser semantic contextual information than pixel-perfect photos [15, 43, 79]. While SD [71] as a backbone feature extractor has established itself as a strong competitor in image-based vision tasks [3, 18, 93, 94, 110], we dive deep into its internal representations to evaluate its effectiveness for the more challenging cross-modal sketch-based vision tasks. To this end, we ask the research question - whether a frozen SD model is capable of extracting features from a sparse sketch that are equally-representative as those derived from a pixel perfect image. Given a few sketch-photo pairs, we extract the corresponding features (detail in Sec. 5.1) from the second upsampling block of a frozen SD's UNet " + }, + { + "bbox": [ + 56, + 163, + 296, + 509 + ], + "type": "inline_equation", + "content": "\\mathbf{U}_{\\theta}" + }, + { + "bbox": [ + 56, + 163, + 296, + 509 + ], + "type": "text", + "content": " with a fixed prompt \"a photo of [CLASS]\" and " + }, + { + "bbox": [ + 56, + 163, + 296, + 509 + ], + "type": "inline_equation", + "content": "t = 195" + }, + { + "bbox": [ + 56, + 163, + 296, + 509 + ], + "type": "text", + "content": ". We perform principal component analysis (PCA) on the extracted high-dimensional features and plot the first three principal components as RGB images for visualisation. As seen in Fig. 1 (left), features extracted from sketches are significantly inferior to those extracted from photos. Unlike photos, abstract freehand sketches are sparse and lack detailed textures/colours [15, 43, 79], making it challenging for the SD model to capture fine-grained visual information from it. We hypothesise that SD struggles to extract equally representative features from sketches due to their minimal semantic cues and high level of abstraction [47]. Additionally, as the pretraining of SD was done predominantly on natural images [80], it is comparatively harder for SD to extract rich semantic information from a sparse sketch." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 510, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 296, + 714 + ], + "type": "text", + "content": "Digging deeper, we follow classical computer vision literature [28, 103] in shifting the focus from spatial-domain, to analyse the behaviour of extracted features from a different perspective, i.e., exploring its nature in the frequency-domain. We achieve this by calculating the " + }, + { + "bbox": [ + 55, + 510, + 296, + 714 + ], + "type": "inline_equation", + "content": "2D" + }, + { + "bbox": [ + 55, + 510, + 296, + 714 + ], + "type": "text", + "content": " Fourier magnitude spectrum of the extracted features, where the central region represents the LF components, while the HF components are arranged radially outward from the centre [28] (Fig. 1 left). Here we see that, although SD performs better on photos than sketches, it predominantly focuses on HF components in both cases. This is likely because SD's UNet is primarily trained as a denoiser thus operating with a global emphasis [107], where its skip connections additionally propagate HF components (e.g., edges), suppressing its LF counterparts (e.g., overall semantic structures) to ease the reconstruction of high-res image from noise [82]. We hypothesise that this very fact is directly conflicting with the" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": "core motivation of our dense prediction task from a sparse data modality i.e., sketches, where capturing comprehensive semantic [107] context (LF) takes much higher precedence over preserving fine-details (HF), thus limiting SD's effectiveness [82, 107] for sketches, especially for dense prediction tasks (e.g., sketch-based segmentation)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 144, + 555, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 144, + 555, + 419 + ], + "spans": [ + { + "bbox": [ + 313, + 144, + 555, + 419 + ], + "type": "text", + "content": "Reinforcing Diffusion Features. We identify two key problems in SD feature extraction - (i) SD features are not equally-representative for sketches and photos, and (ii) it mainly focuses on HF features, losing the crucial LF ones. In addressing these, our PCA analysis (Fig. 1 left) shows that, although SD is spatially-aware, it lacks accuracy. Inspired by existing works [107] on hybrid models, we propose complementing the SD feature extractor by combining it with another pretrained model that offers high accuracy, even if it is less spatially-aware. To this end, we influence the SD feature extraction pipeline with a pretrained CLIP [67] visual encoder as detailed in Sec. 5. PCA maps of CLIP (Fig. 1 left) features depict highly accurate background foreground separation, inherently providing the LF components that were missing in SD features. Finally, PCA maps of the proposed feature extractor (Fig. 1 left) demonstrate how the final representation is spatially-aware, accurate, integrates both HF and LF and works equally well for both sketches and photos. Apart from qualitative evidence, evaluating the performance of baseline " + }, + { + "bbox": [ + 313, + 144, + 555, + 419 + ], + "type": "inline_equation", + "content": "\\mathrm{SD + }\\mathrm{CLIP}" + }, + { + "bbox": [ + 313, + 144, + 555, + 419 + ], + "type": "text", + "content": " hybrid model, and the proposed feature extractor across different sketch-based discriminative and dense prediction tasks, we observe an average performance gain of " + }, + { + "bbox": [ + 313, + 144, + 555, + 419 + ], + "type": "inline_equation", + "content": "39.49\\%" + }, + { + "bbox": [ + 313, + 144, + 555, + 419 + ], + "type": "text", + "content": " (Fig. 1 right)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 424, + 447, + 437 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 424, + 447, + 437 + ], + "spans": [ + { + "bbox": [ + 313, + 424, + 447, + 437 + ], + "type": "text", + "content": "5. Proposed Methodology" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 438, + 555, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 438, + 555, + 676 + ], + "spans": [ + { + "bbox": [ + 313, + 438, + 555, + 676 + ], + "type": "text", + "content": "Overview. Observations from our pilot study (Sec. 4) motivate us to reinforce the SD [71] feature extraction pipeline for better downstream task accuracy. However, reinforcing the SD feature extractor pipeline is non-trivial. Firstly, finetuning the entire model with task-specific dataset might interrupt SD's inherent pretrained knowledge [46]. Furthermore, end-to-end finetuning does not necessarily resolve the issues pertaining to spatial-awareness vs. accuracy or HF vs. LF trade-off (Sec. 4). To this end, we resort to a hybrid model [107] approach and use a complementary pretrained model (i.e., CLIP [67]) to strengthen the SD feature extractor pipeline. Secondly, obtaining features from CLIP alone is insufficient. They must be integrated effectively into the denoising process of the SD UNet to yield the performance gains observed in Fig. 1. Finally, different layers of the SD's UNet capture a wide array of features across different scales and granularity (i.e., coarse to fine) [84], each with its own pros and cons suited to various downstream tasks [46]. Thus, it is essential to effectively aggregate these multi-scale features to maximise their utility." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "content": "Specifically, our salient design choices are " + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "inline_equation", + "content": "(-i)" + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "content": " complementing SD [71] feature extraction with a pretrained CLIP [67] visual model, " + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "inline_equation", + "content": "(ii)" + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "content": " an effective CLIP visual feature in" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2559" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 54, + 72, + 294, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 72, + 294, + 96 + ], + "spans": [ + { + "bbox": [ + 54, + 72, + 294, + 96 + ], + "type": "text", + "content": "jection module, and (iii) an aggregation module to unify features from different scales and granularities." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 100, + 170, + 110 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 100, + 170, + 110 + ], + "spans": [ + { + "bbox": [ + 55, + 100, + 170, + 110 + ], + "type": "text", + "content": "5.1. Model Architecture" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 113, + 296, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 113, + 296, + 380 + ], + "spans": [ + { + "bbox": [ + 56, + 113, + 296, + 380 + ], + "type": "text", + "content": "Stable Diffusion Feature Extraction. Internal activation maps of a pretrained SD UNet " + }, + { + "bbox": [ + 56, + 113, + 296, + 380 + ], + "type": "inline_equation", + "content": "\\mathbf{U}_{\\theta}" + }, + { + "bbox": [ + 56, + 113, + 296, + 380 + ], + "type": "text", + "content": " contain information-rich representations that have found use in tasks like image segmentation [3], object detection [93], classification [49], etc. Here, we aim to use these internal representations as an efficient backbone feature for various sketch-based vision tasks. Given an image-text pair " + }, + { + "bbox": [ + 56, + 113, + 296, + 380 + ], + "type": "inline_equation", + "content": "(\\mathbf{i},\\mathbf{c})" + }, + { + "bbox": [ + 56, + 113, + 296, + 380 + ], + "type": "text", + "content": ", the input image is first passed through the pretrained VAE [42] encoder " + }, + { + "bbox": [ + 56, + 113, + 296, + 380 + ], + "type": "inline_equation", + "content": "\\mathbf{E}(\\cdot)" + }, + { + "bbox": [ + 56, + 113, + 296, + 380 + ], + "type": "text", + "content": ", converting it into an initial latent representation " + }, + { + "bbox": [ + 56, + 113, + 296, + 380 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_0 = \\mathbf{E}(\\mathbf{i}) \\in \\mathbb{R}^{\\frac{h}{8} \\times \\frac{w}{8} \\times d}" + }, + { + "bbox": [ + 56, + 113, + 296, + 380 + ], + "type": "text", + "content": ". Then, given a timestep value " + }, + { + "bbox": [ + 56, + 113, + 296, + 380 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 56, + 113, + 296, + 380 + ], + "type": "text", + "content": ", SD performs forward diffusion on " + }, + { + "bbox": [ + 56, + 113, + 296, + 380 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_0" + }, + { + "bbox": [ + 56, + 113, + 296, + 380 + ], + "type": "text", + "content": " to convert it to the " + }, + { + "bbox": [ + 56, + 113, + 296, + 380 + ], + "type": "inline_equation", + "content": "t^{th}" + }, + { + "bbox": [ + 56, + 113, + 296, + 380 + ], + "type": "text", + "content": " step noisy latent " + }, + { + "bbox": [ + 56, + 113, + 296, + 380 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_t" + }, + { + "bbox": [ + 56, + 113, + 296, + 380 + ], + "type": "text", + "content": ". Finally, the textual prompt embedding (Sec. 3.2) " + }, + { + "bbox": [ + 56, + 113, + 296, + 380 + ], + "type": "inline_equation", + "content": "\\mathbf{T}(\\mathbf{c})" + }, + { + "bbox": [ + 56, + 113, + 296, + 380 + ], + "type": "text", + "content": ", noisy latent " + }, + { + "bbox": [ + 56, + 113, + 296, + 380 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_t" + }, + { + "bbox": [ + 56, + 113, + 296, + 380 + ], + "type": "text", + "content": ", and timestep " + }, + { + "bbox": [ + 56, + 113, + 296, + 380 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 56, + 113, + 296, + 380 + ], + "type": "text", + "content": " are passed to the pretrained denoising UNet [71] " + }, + { + "bbox": [ + 56, + 113, + 296, + 380 + ], + "type": "inline_equation", + "content": "\\mathbf{U}_{\\theta}(\\cdot)" + }, + { + "bbox": [ + 56, + 113, + 296, + 380 + ], + "type": "text", + "content": " to extract the internal features from its skip-connected upsampling layers " + }, + { + "bbox": [ + 56, + 113, + 296, + 380 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{U}_{\\mathbf{u}}^n \\to f_{\\mathbf{u}}^n\\}_{n=1}^4" + }, + { + "bbox": [ + 56, + 113, + 296, + 380 + ], + "type": "text", + "content": ". In SD v2.1 [71], for an input image " + }, + { + "bbox": [ + 56, + 113, + 296, + 380 + ], + "type": "inline_equation", + "content": "\\mathbf{i} \\in \\mathbb{R}^{h \\times w \\times 3}" + }, + { + "bbox": [ + 56, + 113, + 296, + 380 + ], + "type": "text", + "content": ", upsampling layers " + }, + { + "bbox": [ + 56, + 113, + 296, + 380 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{U}_{\\mathbf{u}}^n\\}_{n=1}^4" + }, + { + "bbox": [ + 56, + 113, + 296, + 380 + ], + "type": "text", + "content": " generate features as: " + }, + { + "bbox": [ + 56, + 113, + 296, + 380 + ], + "type": "inline_equation", + "content": "f_{\\mathbf{u}}^1 \\in \\mathbb{R}^{\\frac{h}{32} \\times \\frac{w}{32} \\times 1280}" + }, + { + "bbox": [ + 56, + 113, + 296, + 380 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 56, + 113, + 296, + 380 + ], + "type": "inline_equation", + "content": "f_{\\mathbf{u}}^2 \\in \\mathbb{R}^{\\frac{h}{16} \\times \\frac{w}{16} \\times 1280}" + }, + { + "bbox": [ + 56, + 113, + 296, + 380 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 56, + 113, + 296, + 380 + ], + "type": "inline_equation", + "content": "f_{\\mathbf{u}}^3 \\in \\mathbb{R}^{\\frac{h}{8} \\times \\frac{w}{8} \\times 640}" + }, + { + "bbox": [ + 56, + 113, + 296, + 380 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 56, + 113, + 296, + 380 + ], + "type": "inline_equation", + "content": "f_{\\mathbf{u}}^4 \\in \\mathbb{R}^{\\frac{h}{8} \\times \\frac{w}{8} \\times 320}" + }, + { + "bbox": [ + 56, + 113, + 296, + 380 + ], + "type": "text", + "content": " respectively. Due to the unavailability of paired textual prompts in most sketch-photo datasets [26, 29, 78], we use null (i.e., “”) prompt in all our experiments." + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 56, + 387, + 296, + 487 + ], + "blocks": [ + { + "bbox": [ + 56, + 387, + 296, + 487 + ], + "lines": [ + { + "bbox": [ + 56, + 387, + 296, + 487 + ], + "spans": [ + { + "bbox": [ + 56, + 387, + 296, + 487 + ], + "type": "image", + "image_path": "187a38e80495717266c0b7d1cfd55d677183700bb250696f0c95e6719f483051.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 490, + 295, + 535 + ], + "lines": [ + { + "bbox": [ + 55, + 490, + 295, + 535 + ], + "spans": [ + { + "bbox": [ + 55, + 490, + 295, + 535 + ], + "type": "text", + "content": "Figure 2. Given the frozen SD [71] and CLIP [67] models, the proposed method learns the aggregation network, " + }, + { + "bbox": [ + 55, + 490, + 295, + 535 + ], + "type": "inline_equation", + "content": "1D" + }, + { + "bbox": [ + 55, + 490, + 295, + 535 + ], + "type": "text", + "content": " convolutional layers, and branch-weights with sketch-photo pairs, via different losses for different downstream tasks (details in Sec. 6)." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 543, + 295, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 543, + 295, + 665 + ], + "spans": [ + { + "bbox": [ + 55, + 543, + 295, + 665 + ], + "type": "text", + "content": "CLIP Visual Feature Extraction. Akin to SD, pretrained CLIP [67] encoders have also been used in a plethora of vision tasks like segmentation [60], retrieval [76], object detection [16], video captioning [98], to name a few. In this work, we aim to influence the intermediate feature maps of SD [71] with that from the CLIP visual encoder. Specifically, the input image " + }, + { + "bbox": [ + 55, + 543, + 295, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{i}" + }, + { + "bbox": [ + 55, + 543, + 295, + 665 + ], + "type": "text", + "content": " is passed through the pretrained CLIP visual encoder " + }, + { + "bbox": [ + 55, + 543, + 295, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{V}(\\cdot)" + }, + { + "bbox": [ + 55, + 543, + 295, + 665 + ], + "type": "text", + "content": " to extract the patched feature " + }, + { + "bbox": [ + 55, + 543, + 295, + 665 + ], + "type": "inline_equation", + "content": "f_{\\mathbf{v}}" + }, + { + "bbox": [ + 55, + 543, + 295, + 665 + ], + "type": "text", + "content": " from the penultimate layer as: " + }, + { + "bbox": [ + 55, + 543, + 295, + 665 + ], + "type": "inline_equation", + "content": "f_{\\mathbf{v}} = \\mathbf{V}(\\mathbf{i}) \\in \\mathbb{R}_{p}^{\\frac{h}{p} \\times \\frac{w}{p} \\times d}" + }, + { + "bbox": [ + 55, + 543, + 295, + 665 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 543, + 295, + 665 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 55, + 543, + 295, + 665 + ], + "type": "text", + "content": " is the patch-size of the " + }, + { + "bbox": [ + 55, + 543, + 295, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{V}(\\cdot)" + }, + { + "bbox": [ + 55, + 543, + 295, + 665 + ], + "type": "text", + "content": " backbone encoder." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "CLIP Feature Injection. Based on the qualitative and quantitative evidence presented in the pilot study (Sec. 4), we hypothesise that CLIP [67] visual and textual embeddings are intrinsically aligned [67, 79]. Prior works [79]" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "text", + "content": "have underscored the importance of textual prompts in the quality of extracted features [79]. However, most simple prompts (e.g., \"a photo/sketch of [CLASS]\") are inherently coarse-grained, offering only a high-level description of an image. Such general prompts often lack the semantic detail necessary to fully characterise an image [15, 47], limiting the model's ability to generate highly accurate and contextually-rich feature representations [46]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 168, + 553, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 168, + 553, + 240 + ], + "spans": [ + { + "bbox": [ + 313, + 168, + 553, + 240 + ], + "type": "text", + "content": "To address this, we propose injecting CLIP's visual features into multiple layers of the SD's UNet, thereby influencing the denoising process at various stages (Fig. 2). This approach not only enhances the accuracy of SD features but also enables the model to draw on semantic information from both CLIP's visual and textual embeddings." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 241, + 553, + 364 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 241, + 553, + 364 + ], + "spans": [ + { + "bbox": [ + 313, + 241, + 553, + 364 + ], + "type": "text", + "content": "Specifically, after extracting CLIP's visual embeddings " + }, + { + "bbox": [ + 313, + 241, + 553, + 364 + ], + "type": "inline_equation", + "content": "f_{\\mathbf{v}} \\in \\mathbb{R}^{\\frac{h}{p} \\times \\frac{w}{p} \\times d}" + }, + { + "bbox": [ + 313, + 241, + 553, + 364 + ], + "type": "text", + "content": ", we pass them through simple learnable " + }, + { + "bbox": [ + 313, + 241, + 553, + 364 + ], + "type": "inline_equation", + "content": "1D" + }, + { + "bbox": [ + 313, + 241, + 553, + 364 + ], + "type": "text", + "content": " convolutional layers " + }, + { + "bbox": [ + 313, + 241, + 553, + 364 + ], + "type": "inline_equation", + "content": "\\mathcal{C}(\\cdot)" + }, + { + "bbox": [ + 313, + 241, + 553, + 364 + ], + "type": "text", + "content": " to modify them just enough to match the feature dimension of different blocks of the UNet. These transformed feature maps are then added to the intermediate UNet features " + }, + { + "bbox": [ + 313, + 241, + 553, + 364 + ], + "type": "inline_equation", + "content": "\\{f_{\\mathbf{u}}^{n}\\}_{n=1}^{4}" + }, + { + "bbox": [ + 313, + 241, + 553, + 364 + ], + "type": "text", + "content": " at every timestep of the denoising process as: " + }, + { + "bbox": [ + 313, + 241, + 553, + 364 + ], + "type": "inline_equation", + "content": "\\hat{f}_{\\mathbf{u}}^{n} = f_{\\mathbf{u}}^{n} + \\mathcal{C}(f_{\\mathbf{v}})" + }, + { + "bbox": [ + 313, + 241, + 553, + 364 + ], + "type": "text", + "content": ". This injection occurs in all timesteps and layers of the UNet, enabling the model to influence the denoising process with the semantic information encoded within CLIP's visual embeddings." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 365, + 554, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 365, + 554, + 449 + ], + "spans": [ + { + "bbox": [ + 313, + 365, + 554, + 449 + ], + "type": "text", + "content": "SD Feature Aggregation. Different layers of " + }, + { + "bbox": [ + 313, + 365, + 554, + 449 + ], + "type": "inline_equation", + "content": "\\mathbf{U}_{\\theta}" + }, + { + "bbox": [ + 313, + 365, + 554, + 449 + ], + "type": "text", + "content": " capture features of varying semantic-granularity [84], where features from specific layers are ideally suited for certain downstream tasks [46]. Instead of manually selecting each layer for a specific task, we introduce a dynamic approach, allowing the model to automatically extract the most optimal features, thus eliminating the need for manual tuning." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 449, + 555, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 449, + 555, + 582 + ], + "spans": [ + { + "bbox": [ + 313, + 449, + 555, + 582 + ], + "type": "text", + "content": "To achieve this, we train a feature aggregation network comprising three ResNet [32] blocks " + }, + { + "bbox": [ + 313, + 449, + 555, + 582 + ], + "type": "inline_equation", + "content": "\\mathbf{A}(\\cdot)" + }, + { + "bbox": [ + 313, + 449, + 555, + 582 + ], + "type": "text", + "content": ", which transform the CLIP-enhanced SD features " + }, + { + "bbox": [ + 313, + 449, + 555, + 582 + ], + "type": "inline_equation", + "content": "\\{\\hat{f}_{\\mathbf{u}}^{n}\\}_{n=1}^{3}" + }, + { + "bbox": [ + 313, + 449, + 555, + 582 + ], + "type": "text", + "content": " extracted from the first three UNet upsampling layers as: " + }, + { + "bbox": [ + 313, + 449, + 555, + 582 + ], + "type": "inline_equation", + "content": "\\bar{f}_{\\mathbf{u}}^{n} = \\mathbf{A}(\\hat{f}_{\\mathbf{u}}^{n}) \\in \\mathbb{R}^{60 \\times 60 \\times d}" + }, + { + "bbox": [ + 313, + 449, + 555, + 582 + ], + "type": "text", + "content": ". The final feature map is then obtained by computing a weighted summation of the aggregated features using a learnable weight parameter " + }, + { + "bbox": [ + 313, + 449, + 555, + 582 + ], + "type": "inline_equation", + "content": "\\{\\alpha_{n}\\}_{n=1}^{3}" + }, + { + "bbox": [ + 313, + 449, + 555, + 582 + ], + "type": "text", + "content": " for each feature-branch. This aggregation strategy enables the model to dynamically adjust its focus on different layers, adaptively modulating the contribution of each feature-branch based on the semantic-granularity of the task." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 590, + 394, + 604 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 590, + 394, + 604 + ], + "spans": [ + { + "bbox": [ + 313, + 590, + 394, + 604 + ], + "type": "text", + "content": "6. Experiments" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 605, + 554, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 554, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 554, + 715 + ], + "type": "text", + "content": "The proposed feature extractor enriched with the large-scale pretrained knowledge of SD [71] and CLIP [67], enables multiple sketch-based discriminative and dense prediction tasks. Specifically, we showcase its efficacy in tasks like - sketch-based image retrieval [76], sketch recognition [100], sketch-photo correspondence learning [55], and sketch-based image segmentation [35] on different datasets [26, 29, 55, 78] and scenarios. In all cases, we only learn " + }, + { + "bbox": [ + 313, + 605, + 554, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{C}(\\cdot)" + }, + { + "bbox": [ + 313, + 605, + 554, + 715 + ], + "type": "text", + "content": ", aggregation network " + }, + { + "bbox": [ + 313, + 605, + 554, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{A}(\\cdot)" + }, + { + "bbox": [ + 313, + 605, + 554, + 715 + ], + "type": "text", + "content": ", and branch-weights " + }, + { + "bbox": [ + 313, + 605, + 554, + 715 + ], + "type": "inline_equation", + "content": "\\alpha_{n}" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2560" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 251, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 251, + 83 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 251, + 83 + ], + "type": "text", + "content": "keeping pretrained SD and CLIP models frozen." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 87, + 253, + 99 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 87, + 253, + 99 + ], + "spans": [ + { + "bbox": [ + 55, + 87, + 253, + 99 + ], + "type": "text", + "content": "6.1. Sketch-based Image Retrieval (SBIR)" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 100, + 295, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 100, + 295, + 246 + ], + "spans": [ + { + "bbox": [ + 54, + 100, + 295, + 246 + ], + "type": "text", + "content": "Problem Statement. Given a query sketch " + }, + { + "bbox": [ + 54, + 100, + 295, + 246 + ], + "type": "inline_equation", + "content": "\\mathbf{s} \\in \\mathbb{R}^{h \\times w \\times 3}" + }, + { + "bbox": [ + 54, + 100, + 295, + 246 + ], + "type": "text", + "content": " of any class " + }, + { + "bbox": [ + 54, + 100, + 295, + 246 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 54, + 100, + 295, + 246 + ], + "type": "text", + "content": ", category-level SBIR aims to retrieve a photo " + }, + { + "bbox": [ + 54, + 100, + 295, + 246 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_i^j" + }, + { + "bbox": [ + 54, + 100, + 295, + 246 + ], + "type": "text", + "content": " from the same class, out of a gallery " + }, + { + "bbox": [ + 54, + 100, + 295, + 246 + ], + "type": "inline_equation", + "content": "\\mathcal{G} = \\{\\mathbf{p}_i^j\\}_{i=1}^{N_j}|_{j=1}^{N_c}" + }, + { + "bbox": [ + 54, + 100, + 295, + 246 + ], + "type": "text", + "content": " containing images from a total " + }, + { + "bbox": [ + 54, + 100, + 295, + 246 + ], + "type": "inline_equation", + "content": "N_c" + }, + { + "bbox": [ + 54, + 100, + 295, + 246 + ], + "type": "text", + "content": " classes with " + }, + { + "bbox": [ + 54, + 100, + 295, + 246 + ], + "type": "inline_equation", + "content": "N_j" + }, + { + "bbox": [ + 54, + 100, + 295, + 246 + ], + "type": "text", + "content": " images per class [102]. Whereas, cross-category fine-grained (FG) SBIR setup intends to learn a unified model capable of instance-level matching from a gallery containing photos from multiple " + }, + { + "bbox": [ + 54, + 100, + 295, + 246 + ], + "type": "inline_equation", + "content": "(N_c)" + }, + { + "bbox": [ + 54, + 100, + 295, + 246 + ], + "type": "text", + "content": " classes. Conventional SBIR is evaluated on categories " + }, + { + "bbox": [ + 54, + 100, + 295, + 246 + ], + "type": "inline_equation", + "content": "\\mathcal{C}^S = \\{c_1^S, \\dots, c_N^S\\}" + }, + { + "bbox": [ + 54, + 100, + 295, + 246 + ], + "type": "text", + "content": " that are seen during training. Contrarily, the zero-shot (ZS) paradigm [76] focuses on evaluating on mutually exclusive (i.e., " + }, + { + "bbox": [ + 54, + 100, + 295, + 246 + ], + "type": "inline_equation", + "content": "\\mathcal{C}^S \\cap \\mathcal{C}^U = \\emptyset" + }, + { + "bbox": [ + 54, + 100, + 295, + 246 + ], + "type": "text", + "content": ") unseen categories " + }, + { + "bbox": [ + 54, + 100, + 295, + 246 + ], + "type": "inline_equation", + "content": "\\mathcal{C}^U = \\{c_1^U, \\dots, c_M^U\\}" + }, + { + "bbox": [ + 54, + 100, + 295, + 246 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 247, + 296, + 393 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 247, + 296, + 393 + ], + "spans": [ + { + "bbox": [ + 55, + 247, + 296, + 393 + ], + "type": "text", + "content": "Training & Evaluation. Following standard literature [76], we use triplet loss to train our SBIR models. Specifically, given the global max-pooled features from the proposed feature extractor, triplet loss aims to minimise the distance " + }, + { + "bbox": [ + 55, + 247, + 296, + 393 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 55, + 247, + 296, + 393 + ], + "type": "text", + "content": " between an anchor sketch (s) feature " + }, + { + "bbox": [ + 55, + 247, + 296, + 393 + ], + "type": "inline_equation", + "content": "f_{\\mathbf{s}} \\in \\mathbb{R}^{d}" + }, + { + "bbox": [ + 55, + 247, + 296, + 393 + ], + "type": "text", + "content": " and a positive photo (p) feature " + }, + { + "bbox": [ + 55, + 247, + 296, + 393 + ], + "type": "inline_equation", + "content": "f_{\\mathbf{p}} \\in \\mathbb{R}^{d}" + }, + { + "bbox": [ + 55, + 247, + 296, + 393 + ], + "type": "text", + "content": " from the same category as s, while maximising the same from a negative photo (n) feature " + }, + { + "bbox": [ + 55, + 247, + 296, + 393 + ], + "type": "inline_equation", + "content": "f_{\\mathbf{n}} \\in \\mathbb{R}^{d}" + }, + { + "bbox": [ + 55, + 247, + 296, + 393 + ], + "type": "text", + "content": " of a different category. Whereas, in case of cross-category FG-SBIR framework we use hard triplets, where the negative sample is a distinct instance " + }, + { + "bbox": [ + 55, + 247, + 296, + 393 + ], + "type": "inline_equation", + "content": "(\\mathbf{p}_k^j; k \\neq i)" + }, + { + "bbox": [ + 55, + 247, + 296, + 393 + ], + "type": "text", + "content": " belonging to the same class as the anchor sketch " + }, + { + "bbox": [ + 55, + 247, + 296, + 393 + ], + "type": "inline_equation", + "content": "(\\mathbf{s}_i^j)" + }, + { + "bbox": [ + 55, + 247, + 296, + 393 + ], + "type": "text", + "content": " and its paired photo " + }, + { + "bbox": [ + 55, + 247, + 296, + 393 + ], + "type": "inline_equation", + "content": "(\\mathbf{p}_i^j)" + }, + { + "bbox": [ + 55, + 247, + 296, + 393 + ], + "type": "text", + "content": " [46]. With margin " + }, + { + "bbox": [ + 55, + 247, + 296, + 393 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 55, + 247, + 296, + 393 + ], + "type": "text", + "content": ", triplet loss becomes:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 60, + 402, + 295, + 416 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 402, + 295, + 416 + ], + "spans": [ + { + "bbox": [ + 60, + 402, + 295, + 416 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {t r i p l e t}} \\left(f _ {\\mathbf {s}}, f _ {\\mathbf {p}}, f _ {\\mathbf {n}}\\right) = \\max \\left\\{0, \\mu + \\delta \\left(f _ {\\mathbf {s}}, f _ {\\mathbf {p}}\\right) - \\delta \\left(f _ {\\mathbf {s}}, f _ {\\mathbf {n}}\\right) \\right\\}. \\tag {1}", + "image_path": "0e527c7a16c102c1c8e705215706d3290fdc51bc953155c53de150f998407c60.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 425, + 295, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 425, + 295, + 496 + ], + "spans": [ + { + "bbox": [ + 55, + 425, + 295, + 496 + ], + "type": "text", + "content": "For category-level ZS-SBIR, following [46, 76], we consider top " + }, + { + "bbox": [ + 55, + 425, + 295, + 496 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 425, + 295, + 496 + ], + "type": "text", + "content": " retrieved images to calculate mean average precision " + }, + { + "bbox": [ + 55, + 425, + 295, + 496 + ], + "type": "inline_equation", + "content": "(\\mathrm{mAP}@\\mathrm{k})" + }, + { + "bbox": [ + 55, + 425, + 295, + 496 + ], + "type": "text", + "content": " and precision " + }, + { + "bbox": [ + 55, + 425, + 295, + 496 + ], + "type": "inline_equation", + "content": "(\\mathrm{P}@\\mathrm{k})" + }, + { + "bbox": [ + 55, + 425, + 295, + 496 + ], + "type": "text", + "content": ". Whereas, for cross-category ZS-FG-SBIR, we measure Acc.@k, which indicates the percentage of sketches having true-matched photos among the top-k retrieved images." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 498, + 295, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 295, + 664 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 295, + 664 + ], + "type": "text", + "content": "Dataset & Competitors. Here we use three datasets - (i) Sketchy [78] and Sketchy-extended [53, 78] datasets are used to train and test our cross-category ZS-FG-SBIR and category-level ZS-SBIR models respectively. While Sketchy [78] holds 12,500 photos from 125 classes, each having at least 5 fine-grained paired sketches, its extended version [53] carries additional 60,652 photos from ImageNet [19]. We use photos/sketches from 104 classes for training and 21 for testing [102]. (ii) Tu-Berlin [26] contains 204,489 photos from 250 classes, each having " + }, + { + "bbox": [ + 55, + 498, + 295, + 664 + ], + "type": "inline_equation", + "content": "\\sim 80" + }, + { + "bbox": [ + 55, + 498, + 295, + 664 + ], + "type": "text", + "content": " sketches. We use photos/sketches from 220 classes for training and 30 for testing. (iii) Quick, Draw! [29] holds over " + }, + { + "bbox": [ + 55, + 498, + 295, + 664 + ], + "type": "inline_equation", + "content": "50M" + }, + { + "bbox": [ + 55, + 498, + 295, + 664 + ], + "type": "text", + "content": " sketches from 345 categories. We use photos/sketches from 80 classes for training and 30 for testing." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "For category-level ZS-SBIR, we compare our method with SoTA methods like SAKE [54], CAAE [102], CCGAN [25], GRL [20], SD-PL [46], etc. Whereas, for cross-category ZS-FG-SBIR, we compare with Gen-VAE [63]," + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 318, + 70, + 553, + 218 + ], + "blocks": [ + { + "bbox": [ + 318, + 70, + 553, + 218 + ], + "lines": [ + { + "bbox": [ + 318, + 70, + 553, + 218 + ], + "spans": [ + { + "bbox": [ + 318, + 70, + 553, + 218 + ], + "type": "table", + "html": "
MethodsSketchy [78]TU-Berlin [26]Quick, Draw! [29]
mAP@200P@200mAP@allP@100mAP@allP@200
B-CLIP0.2500.2610.2280.2570.0800.141
B-DINO0.4930.4810.4500.4920.1670.249
B-DINOV20.5270.5330.4810.5320.1700.268
B-SD0.5580.5710.5100.5610.1790.287
B-SD+CLIP0.5880.5920.5370.5890.1790.311
B-Finetuning0.1200.1720.0110.0100.0030.006
SAKE [54]0.4970.5980.4750.599--
IIAE [37]0.3730.4850.4120.503--
CAAE [102]0.1560.2600.0050.003--
CCGAN [25]--0.2970.426--
CVAE [102]0.2250.3330.0050.0010.0030.003
GRL [20]0.3690.3700.1100.1210.0750.068
LVM [76]0.7230.7250.6510.7320.2020.388
SD-PL [46]0.7460.7470.6800.7440.2310.397
Proposed0.7610.7630.6950.7530.2420.399
", + "image_path": "4d9bc86e1ecb1cb743cc6fe167fff183f4be653e615ea8251e4b78bb4f771274.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 235, + 555, + 368 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 235, + 555, + 368 + ], + "spans": [ + { + "bbox": [ + 313, + 235, + 555, + 368 + ], + "type": "text", + "content": "LVM [76], and " + }, + { + "bbox": [ + 313, + 235, + 555, + 368 + ], + "type": "inline_equation", + "content": "SD-PL" + }, + { + "bbox": [ + 313, + 235, + 555, + 368 + ], + "type": "text", + "content": " [46]. Apart from the SoTA methods, we also compare all tasks against a few self-designed Baselines. Among them B-CLIP, B-DINO, B-DINOv2, and B-SD use off-the-shelf pretrained CLIP [67] ViT-L/14, DINO [8] ViT-S/8, DINOv2 [61] ViT-S/14, or SD [71] v2.1 encoders as feature extractors, without task-specific training. We also form another baseline " + }, + { + "bbox": [ + 313, + 235, + 555, + 368 + ], + "type": "inline_equation", + "content": "B-SD+CLIP" + }, + { + "bbox": [ + 313, + 235, + 555, + 368 + ], + "type": "text", + "content": " as a straightforward combination of SD and CLIP features without the proposed feature injection and aggregation modules. Finally, B-Finetuning finetunes the entire pretrained SD UNet and CLIP visual encoder on task-specific data." + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 318, + 374, + 552, + 437 + ], + "blocks": [ + { + "bbox": [ + 353, + 219, + 515, + 230 + ], + "lines": [ + { + "bbox": [ + 353, + 219, + 515, + 230 + ], + "spans": [ + { + "bbox": [ + 353, + 219, + 515, + 230 + ], + "type": "text", + "content": "Table 1. Results for category-level ZS-SBIR." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 318, + 374, + 552, + 437 + ], + "lines": [ + { + "bbox": [ + 318, + 374, + 552, + 437 + ], + "spans": [ + { + "bbox": [ + 318, + 374, + 552, + 437 + ], + "type": "table", + "html": "
MethodsAcc.@1Acc.@5MethodsAcc.@1Acc.@5
B-CLIP11.8421.66B-Finetuning5.679.17
B-DINO22.4946.97Gen-VAE [63]22.6049.00
B-DINOv221.1944.31LVM [76]28.6862.34
B-SD23.9849.42SD-PL [46]31.9465.81
B-SD+CLIP24.1652.01Proposed33.0167.92
", + "image_path": "6fb0799ac9c8d3312c9e9d6376c8d225abb09e14ffc2bb551e1b9414d8103e12.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 315, + 438, + 553, + 449 + ], + "lines": [ + { + "bbox": [ + 315, + 438, + 553, + 449 + ], + "spans": [ + { + "bbox": [ + 315, + 438, + 553, + 449 + ], + "type": "text", + "content": "Table 2. Results on Sketchy [78] for cross-category ZS-FG-SBIR." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 460, + 427, + 472 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 460, + 427, + 472 + ], + "spans": [ + { + "bbox": [ + 313, + 460, + 427, + 472 + ], + "type": "text", + "content": "6.2. Sketch Recognition" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 474, + 554, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 474, + 554, + 545 + ], + "spans": [ + { + "bbox": [ + 313, + 474, + 554, + 545 + ], + "type": "text", + "content": "Problem Statement. Sketch recognition deals with identifying and classifying freehand abstract sketches into predefined categories [68, 96]. Given a dataset " + }, + { + "bbox": [ + 313, + 474, + 554, + 545 + ], + "type": "inline_equation", + "content": "\\mathcal{D} = \\{(\\mathbf{s}_i,\\mathbf{y}_i)\\}_{i = 1}^N" + }, + { + "bbox": [ + 313, + 474, + 554, + 545 + ], + "type": "text", + "content": " containing " + }, + { + "bbox": [ + 313, + 474, + 554, + 545 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 474, + 554, + 545 + ], + "type": "text", + "content": " sketches and their corresponding labels from " + }, + { + "bbox": [ + 313, + 474, + 554, + 545 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 313, + 474, + 554, + 545 + ], + "type": "text", + "content": " classes, the goal is to learn a model that can correctly recognise the class of an input query sketch." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 548, + 554, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 548, + 554, + 583 + ], + "spans": [ + { + "bbox": [ + 313, + 548, + 554, + 583 + ], + "type": "text", + "content": "Training & Evaluation. Given the global max-pooled query sketch feature " + }, + { + "bbox": [ + 313, + 548, + 554, + 583 + ], + "type": "inline_equation", + "content": "f_{\\mathbf{s}} \\in \\mathbb{R}^{d}" + }, + { + "bbox": [ + 313, + 548, + 554, + 583 + ], + "type": "text", + "content": ", we use standard cross-entropy (CE) loss to train our sketch recognition model as," + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 359, + 587, + 555, + 617 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 587, + 555, + 617 + ], + "spans": [ + { + "bbox": [ + 359, + 587, + 555, + 617 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {C E}} \\left(\\mathbf {s} _ {i}, \\mathbf {y} _ {i}\\right) = - \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\log p \\left(\\mathbf {y} _ {i} \\mid \\mathbf {s} _ {i}\\right). \\tag {2}", + "image_path": "525234258639bfe2c64c8c949185e75713654524c634d89a9eac3b7cf629a1fd.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 620, + 522, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 620, + 522, + 632 + ], + "spans": [ + { + "bbox": [ + 313, + 620, + 522, + 632 + ], + "type": "text", + "content": "Following [66, 96, 101] we evaluate using Acc.@k." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 634, + 554, + 717 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 634, + 554, + 717 + ], + "spans": [ + { + "bbox": [ + 313, + 634, + 554, + 717 + ], + "type": "text", + "content": "Dataset & Competitors. For this task, we again use the popular TU-Berlin [26] and Quick, Draw! [29] datasets. For Quick, Draw!, we adopt the data split from [29], with each class containing 70,000, 2,500, and 2500 training, validation, and test samples. Apart from the self-designed baselines, we also compare against SoTA methods like SketchMate [96], SketchGNN [101], and SketchXAI [66]." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "2561" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 70, + 294, + 121 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 294, + 121 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 294, + 121 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 294, + 121 + ], + "type": "table", + "html": "
MethodsTU-BerlinQuick, Draw!MethodsTU-BerlinQuick, Draw!
B-CLIP30.0930.87B-Finetuning10.2912.37
B-DINO51.7953.28SketchMate [96]77.9679.44
B-DINOv258.0159.12SketchGNN [101]76.4377.31
B-SD61.3763.57SketchXAI [66]-86.10
B-SD+CLIP65.4767.78Proposed84.9687.02
", + "image_path": "2d0820a284069092f897e296c7d756732cc588a68cd867e895f7cb5539bf567a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 134, + 220, + 147 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 134, + 220, + 147 + ], + "spans": [ + { + "bbox": [ + 55, + 134, + 220, + 147 + ], + "type": "text", + "content": "6.3. Sketch-photo Correspondence" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 147, + 296, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 147, + 296, + 258 + ], + "spans": [ + { + "bbox": [ + 55, + 147, + 296, + 258 + ], + "type": "text", + "content": "Problem Statement. Sketch-photo correspondence learning involves identifying a mapping between a freehand sketch of an object and its corresponding photo, such that keypoints in the sketch can be semantically aligned with corresponding keypoints in the photo [55]. Given a sketch-photo pair " + }, + { + "bbox": [ + 55, + 147, + 296, + 258 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{s},\\mathbf{p}\\} \\in \\mathbb{R}^{h\\times w\\times 3}" + }, + { + "bbox": [ + 55, + 147, + 296, + 258 + ], + "type": "text", + "content": " and a set of " + }, + { + "bbox": [ + 55, + 147, + 296, + 258 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 147, + 296, + 258 + ], + "type": "text", + "content": " query keypoints (on sketch) " + }, + { + "bbox": [ + 55, + 147, + 296, + 258 + ], + "type": "inline_equation", + "content": "K_{\\mathbf{s}} = \\{x_i^s,y_i^s\\}_{i = 1}^N" + }, + { + "bbox": [ + 55, + 147, + 296, + 258 + ], + "type": "text", + "content": ", the aim is to predict a new set of corresponding keypoints " + }, + { + "bbox": [ + 55, + 147, + 296, + 258 + ], + "type": "inline_equation", + "content": "\\hat{K}_{\\mathbf{p}} = \\{\\hat{x}_i^{\\mathbf{p}},\\hat{y}_i^{\\mathbf{p}}\\}_{i = 1}^N" + }, + { + "bbox": [ + 55, + 147, + 296, + 258 + ], + "type": "text", + "content": " on the photo, such that " + }, + { + "bbox": [ + 55, + 147, + 296, + 258 + ], + "type": "inline_equation", + "content": "\\hat{K}_{\\mathbf{p}}" + }, + { + "bbox": [ + 55, + 147, + 296, + 258 + ], + "type": "text", + "content": " aligns semantically with " + }, + { + "bbox": [ + 55, + 147, + 296, + 258 + ], + "type": "inline_equation", + "content": "K_{\\mathbf{s}}" + }, + { + "bbox": [ + 55, + 147, + 296, + 258 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 258, + 296, + 377 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 258, + 296, + 377 + ], + "spans": [ + { + "bbox": [ + 55, + 258, + 296, + 377 + ], + "type": "text", + "content": "Training & Evaluation. Following [106], we use contrastive " + }, + { + "bbox": [ + 55, + 258, + 296, + 377 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{\\mathrm{CL}})" + }, + { + "bbox": [ + 55, + 258, + 296, + 377 + ], + "type": "text", + "content": " [67] and end-point error losses to train our sketch-photo correspondence model. Given the sketch and photo features " + }, + { + "bbox": [ + 55, + 258, + 296, + 377 + ], + "type": "inline_equation", + "content": "\\{f_{\\mathrm{s}}, f_{\\mathrm{p}}\\} \\in \\mathbb{R}^{60 \\times 60 \\times d}" + }, + { + "bbox": [ + 55, + 258, + 296, + 377 + ], + "type": "text", + "content": " from the proposed feature extractor, " + }, + { + "bbox": [ + 55, + 258, + 296, + 377 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{CL}}" + }, + { + "bbox": [ + 55, + 258, + 296, + 377 + ], + "type": "text", + "content": " aims to maximise similarity between local feature patches, while minimising the same between distance ones, for each sketch-photo pairs. Whereas, for endpoint error loss, we first estimate optical flows [106] from the extracted features, followed by calculating " + }, + { + "bbox": [ + 55, + 258, + 296, + 377 + ], + "type": "inline_equation", + "content": "l_{2}" + }, + { + "bbox": [ + 55, + 258, + 296, + 377 + ], + "type": "text", + "content": " norm between the estimated " + }, + { + "bbox": [ + 55, + 258, + 296, + 377 + ], + "type": "inline_equation", + "content": "(\\mathbf{f}_e)" + }, + { + "bbox": [ + 55, + 258, + 296, + 377 + ], + "type": "text", + "content": " and ground truth " + }, + { + "bbox": [ + 55, + 258, + 296, + 377 + ], + "type": "inline_equation", + "content": "(\\mathbf{f}_g)" + }, + { + "bbox": [ + 55, + 258, + 296, + 377 + ], + "type": "text", + "content": " flow as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 104, + 380, + 294, + 409 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 380, + 294, + 409 + ], + "spans": [ + { + "bbox": [ + 104, + 380, + 294, + 409 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {E P E}} \\left(\\mathbf {f} _ {e} ^ {i}, \\mathbf {f} _ {g} ^ {i}\\right) = \\frac {1}{M} \\sum_ {i = 1} ^ {M} \\left\\| \\mathbf {f} _ {e} ^ {i} - \\mathbf {f} _ {g} ^ {i} \\right\\| _ {2} ^ {2}, \\tag {3}", + "image_path": "dc801127585f1181fa24e8f8f7ed0e8ef5d5243f1613ebacc8b88e6bb9a7de3d.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 411, + 296, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 411, + 296, + 471 + ], + "spans": [ + { + "bbox": [ + 55, + 411, + 296, + 471 + ], + "type": "text", + "content": "where, " + }, + { + "bbox": [ + 55, + 411, + 296, + 471 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 411, + 296, + 471 + ], + "type": "text", + "content": " denotes the patch number (i.e., 60). Finally, the overall loss becomes " + }, + { + "bbox": [ + 55, + 411, + 296, + 471 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{corr}} = \\mathcal{L}_{\\mathrm{CL}} + \\mathcal{L}_{\\mathrm{EPE}}" + }, + { + "bbox": [ + 55, + 411, + 296, + 471 + ], + "type": "text", + "content": ". Following [55] we use PCK@k as a test metric which measures the percentage of keypoints predicted correctly within a distance threshold (i.e., " + }, + { + "bbox": [ + 55, + 411, + 296, + 471 + ], + "type": "inline_equation", + "content": "k\\%" + }, + { + "bbox": [ + 55, + 411, + 296, + 471 + ], + "type": "text", + "content": " of the image size)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 471, + 296, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 471, + 296, + 567 + ], + "spans": [ + { + "bbox": [ + 55, + 471, + 296, + 567 + ], + "type": "text", + "content": "Dataset & Competitors. For this task, we use the PSC6K [55] dataset that contains more than " + }, + { + "bbox": [ + 55, + 471, + 296, + 567 + ], + "type": "inline_equation", + "content": "150K" + }, + { + "bbox": [ + 55, + 471, + 296, + 567 + ], + "type": "text", + "content": " keypoint annotations for 1250 photos, each having 5 paired sketches (taken from the original Sketchy [78] test split) from 125 classes. We use a 60:40 split for training and testing, where we compare the proposed approach against the self-designed baselines and SoTA methods like WeakAlign[70], WarpC [85], Self-Supervised [55], etc." + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 58, + 571, + 294, + 628 + ], + "blocks": [ + { + "bbox": [ + 89, + 122, + 261, + 132 + ], + "lines": [ + { + "bbox": [ + 89, + 122, + 261, + 132 + ], + "spans": [ + { + "bbox": [ + 89, + 122, + 261, + 132 + ], + "type": "text", + "content": "Table 3. Acc.@1 results for sketch recognition." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 571, + 294, + 628 + ], + "lines": [ + { + "bbox": [ + 58, + 571, + 294, + 628 + ], + "spans": [ + { + "bbox": [ + 58, + 571, + 294, + 628 + ], + "type": "table", + "html": "
MethodsPCK@5PCK@10MethodsPCK@5PCK@10
B-CLIP22.2930.41B-Finetuning11.5817.69
B-DINO34.2148.59WeakAlign [70]43.5578.60
B-DINOv239.9156.46WarpC [85]56.7879.70
B-SD41.4258.71Self-Sup. [55]58.0084.93
B-SD+CLIP44.9062.02Proposed70.3189.86
", + "image_path": "49e584731147763063a56d8a9afa265bce0bc4455d4f0aa155745da046bdebb4.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 630, + 294, + 653 + ], + "lines": [ + { + "bbox": [ + 55, + 630, + 294, + 653 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 294, + 653 + ], + "type": "text", + "content": "Table 4. Results on PSC6K [55] for sketch-photo correspondence. 6.4. Sketch-based Image Segmentation" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "content": "Problem Statement. Given a query sketch " + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{s} \\in \\mathbb{R}^{h \\times w \\times 3}" + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "content": " of any category, sketch-based image segmentation [35] aims to predict a binary segmentation mask, " + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{m} \\in \\mathbb{R}^{h \\times w \\times 1}" + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "content": ", depicting the pixel-locations, wherever the queried-concept appears in any candidate image " + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{p} \\in \\mathbb{R}^{h \\times w \\times 3}" + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "content": " of that category," + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "content": "from an entire test gallery. The predicted mask marks pixels belonging to that sketched-concept as 1 and the rest as 0." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 96, + 554, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 96, + 554, + 228 + ], + "spans": [ + { + "bbox": [ + 313, + 96, + 554, + 228 + ], + "type": "text", + "content": "Training & Evaluation. We train our segmentation model using binary cross-entropy (BCE) loss, between predicted and ground truth (GT) masks. Given patched photo feature " + }, + { + "bbox": [ + 313, + 96, + 554, + 228 + ], + "type": "inline_equation", + "content": "f_{\\mathbf{p}} \\in \\mathbb{R}^{60 \\times 60 \\times d}" + }, + { + "bbox": [ + 313, + 96, + 554, + 228 + ], + "type": "text", + "content": " and global max-pooled sketch feature photo features " + }, + { + "bbox": [ + 313, + 96, + 554, + 228 + ], + "type": "inline_equation", + "content": "f_{\\mathbf{s}} \\in \\mathbb{R}^d" + }, + { + "bbox": [ + 313, + 96, + 554, + 228 + ], + "type": "text", + "content": " from the proposed feature extractor, we calculate the cosine similarity between them to generate a patch-wise correlation map " + }, + { + "bbox": [ + 313, + 96, + 554, + 228 + ], + "type": "inline_equation", + "content": "\\mathbf{c} = f_{\\mathbf{s}} \\odot f_{\\mathbf{p}} \\in \\mathbb{R}^{60 \\times 60}" + }, + { + "bbox": [ + 313, + 96, + 554, + 228 + ], + "type": "text", + "content": ". We upsample " + }, + { + "bbox": [ + 313, + 96, + 554, + 228 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 313, + 96, + 554, + 228 + ], + "type": "text", + "content": " to the original spatial resolution (i.e., " + }, + { + "bbox": [ + 313, + 96, + 554, + 228 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{h \\times w}" + }, + { + "bbox": [ + 313, + 96, + 554, + 228 + ], + "type": "text", + "content": "), followed by differentiable Sigmoid thresholding [10] to get the predicted map. BCE loss between the GT and predicted maps can be calculated as," + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 318, + 232, + 554, + 258 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 232, + 554, + 258 + ], + "spans": [ + { + "bbox": [ + 318, + 232, + 554, + 258 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {B C E}} \\left(y _ {i}, \\hat {y} _ {i}\\right) = - \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\left[ y _ {i} \\log \\left(\\hat {y} _ {i}\\right) + \\left(1 - y _ {i}\\right) \\log \\left(1 - \\hat {y} _ {i}\\right) \\right] \\quad , \\tag {4}", + "image_path": "09ecbe891d0841f54573456d3af309748e3ea1420da763a6b433cbe064312545.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 257, + 554, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 257, + 554, + 376 + ], + "spans": [ + { + "bbox": [ + 313, + 257, + 554, + 376 + ], + "type": "text", + "content": "where, " + }, + { + "bbox": [ + 313, + 257, + 554, + 376 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 257, + 554, + 376 + ], + "type": "text", + "content": " is the number of total pixels, " + }, + { + "bbox": [ + 313, + 257, + 554, + 376 + ], + "type": "inline_equation", + "content": "y_{i}" + }, + { + "bbox": [ + 313, + 257, + 554, + 376 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 257, + 554, + 376 + ], + "type": "inline_equation", + "content": "\\hat{y}_i" + }, + { + "bbox": [ + 313, + 257, + 554, + 376 + ], + "type": "text", + "content": " are the GT and predicted mask pixels. During testing, we calculate similar patch-wise correlation map from sketch/photo features and upsample it to the original spatial image dimension, followed by thresholding it with an empirically chosen fixed value (0.47) and standard CRF [48]-based postprocessing. We use mean Intersection over Union (mIoU) and pixel accuracy (pAcc.) as test metrics, which measure the overlap and percentage of correctly predicted pixels between predicted and GT segmentation masks respectively." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 376, + 554, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 376, + 554, + 495 + ], + "spans": [ + { + "bbox": [ + 313, + 376, + 554, + 495 + ], + "type": "text", + "content": "Dataset & Competitors. Due to the lack of existing paired sketch-photo-mask datasets, we manually annotate photos of 10 chosen classes from the Sketchy [78] dataset with binary mask annotations, resulting in " + }, + { + "bbox": [ + 313, + 376, + 554, + 495 + ], + "type": "inline_equation", + "content": "5K" + }, + { + "bbox": [ + 313, + 376, + 554, + 495 + ], + "type": "text", + "content": " paired sketch-photo-mask tuples. We split it into 80:20 for training and testing. Besides baselines and SoTA Sketch-a-Segmenter [35], we also compare with a modified version of ZS-Seg [95], which adapts a simple text-based zero-shot segmentation framework for sketches by replacing its CLIP [67] textual encoder with the visual one to encode input sketches." + } + ] + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 315, + 498, + 553, + 556 + ], + "blocks": [ + { + "bbox": [ + 315, + 498, + 553, + 556 + ], + "lines": [ + { + "bbox": [ + 315, + 498, + 553, + 556 + ], + "spans": [ + { + "bbox": [ + 315, + 498, + 553, + 556 + ], + "type": "image", + "image_path": "d89ac8a0f6954ad106c6e9bc31ac9ca939af4347dbb907db2b33e7a4e1eff411.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 558, + 553, + 591 + ], + "lines": [ + { + "bbox": [ + 313, + 558, + 553, + 591 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 553, + 591 + ], + "type": "text", + "content": "Figure 3. Sketch-photo correspondence (left→right : source → target) results on PSC6K. Green circles and squares depict source and GT points respectively, while red squares denote predicted points." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 593, + 436, + 605 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 593, + 436, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 593, + 436, + 605 + ], + "type": "text", + "content": "6.5. Results and Analyses" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 605, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 554, + 713 + ], + "type": "text", + "content": "Following are the observations from quantitative and qualitative results presented in Tab. 1-5 and Fig. 3-4. (i) Naive adaptation of pretrained backbones (e.g., B-CLIP, B-DINO, B-SD, etc.) fail to beat our method in both ZS-SBIR and ZS-FG-SBIR settings (Tab. 1-2), where we achieve an average 76.15 (54.32)% higher mAP@200 (Acc.@1) over all competitors on Sketchy. We posit that our strategic combination of SD and CLIP brings the \"best of both worlds\", complementing each other's flaws. (ii) While" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2562" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 59, + 71, + 293, + 132 + ], + "blocks": [ + { + "bbox": [ + 59, + 71, + 293, + 132 + ], + "lines": [ + { + "bbox": [ + 59, + 71, + 293, + 132 + ], + "spans": [ + { + "bbox": [ + 59, + 71, + 293, + 132 + ], + "type": "table", + "html": "
MethodsmIoUpAcc.MethodsmIoUpAcc.
B-CLIP20.6326.82B-Finetuning20.3621.84
B-DINO32.8142.06DeepLabv3+Sketch [35]40.6355.23
B-DINOv234.1745.80ZS-Seg [95]44.7360.97
B-SD39.0249.03Sketch-a-Segmenter [35]46.4560.28
B-SD+CLIP41.8751.91Proposed60.1274.91
", + "image_path": "1559488a1c1b67360bd5f6be3120a48a2fe4010279189e9772821bbdf7a9208a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 146, + 296, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 146, + 296, + 434 + ], + "spans": [ + { + "bbox": [ + 55, + 146, + 296, + 434 + ], + "type": "text", + "content": "SoTA sketch-recognition frameworks [96, 101] perform reasonably across different datasets (Tab. 3), our method exceeds them with an average Acc.@1 gain of " + }, + { + "bbox": [ + 55, + 146, + 296, + 434 + ], + "type": "inline_equation", + "content": "48.09\\%" + }, + { + "bbox": [ + 55, + 146, + 296, + 434 + ], + "type": "text", + "content": " (Quick, Draw!). SketchXAI [66] despite being our strongest contender, depicts sub-optimal Acc.@1. This gain is likely due to the large-scale pretraining of SD and CLIP, which enables superior open-world object understanding [107]. (iii) For sketch-photo correspondence learning, our method shows an impressive " + }, + { + "bbox": [ + 55, + 146, + 296, + 434 + ], + "type": "inline_equation", + "content": "21.22\\%" + }, + { + "bbox": [ + 55, + 146, + 296, + 434 + ], + "type": "text", + "content": " PCK@5 gain over SoTA [55], without the time-consuming two-stage training of [55], or complicated warp-estimation of [85], thanks to SD's innate object-localisation capability. (iv) Sketch-based segmentation being the most challenging task amongst all, baseline methods perform quite poorly (Tab. 5). However, the proposed method surpasses SoTA [35] with a mIoU (pAcc.) margin of " + }, + { + "bbox": [ + 55, + 146, + 296, + 434 + ], + "type": "inline_equation", + "content": "29.42(24.27)\\%" + }, + { + "bbox": [ + 55, + 146, + 296, + 434 + ], + "type": "text", + "content": ". (v) B-Finetuning fails drastically in almost all tasks (Tab. 1-5). We hypothesise that finetuning entire SD and CLIP models with limited sketch data distorts their large-scale pretrained knowledge [46, 76], thus sacrificing the generalisation potential. (vi) Finally, B-SD+CLIP depicts a " + }, + { + "bbox": [ + 55, + 146, + 296, + 434 + ], + "type": "inline_equation", + "content": "\\sim 35.78\\%" + }, + { + "bbox": [ + 55, + 146, + 296, + 434 + ], + "type": "text", + "content": " lesser scores than ours (averaged over all tasks), which again underpins the need for the proposed feature injection and aggregation strategy, rather than a simple combination of pretrained foundation models." + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 63, + 437, + 294, + 504 + ], + "blocks": [ + { + "bbox": [ + 63, + 437, + 294, + 504 + ], + "lines": [ + { + "bbox": [ + 63, + 437, + 294, + 504 + ], + "spans": [ + { + "bbox": [ + 63, + 437, + 294, + 504 + ], + "type": "image", + "image_path": "ed87e7b74dc053d81e4ac063d273158fb0c1303a12c6a289a234fbc76ecb31bb.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 507, + 295, + 540 + ], + "lines": [ + { + "bbox": [ + 55, + 507, + 295, + 540 + ], + "spans": [ + { + "bbox": [ + 55, + 507, + 295, + 540 + ], + "type": "text", + "content": "Figure 4. Qualitative results for sketch-based image segmentation. Given a query sketch, our method generates separate segmentation masks for all images of that category. (Zoom-in for the best view.)" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 543, + 169, + 556 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 543, + 169, + 556 + ], + "spans": [ + { + "bbox": [ + 55, + 543, + 169, + 556 + ], + "type": "text", + "content": "7. Ablation on Design" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 558, + 295, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 295, + 641 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 295, + 641 + ], + "type": "text", + "content": "[i] Effect of aggregation network. To judge its efficacy, instead of aggregation networks " + }, + { + "bbox": [ + 55, + 558, + 295, + 641 + ], + "type": "inline_equation", + "content": "\\mathbf{A}(\\cdot)" + }, + { + "bbox": [ + 55, + 558, + 295, + 641 + ], + "type": "text", + "content": ", we manually interpolate features from " + }, + { + "bbox": [ + 55, + 558, + 295, + 641 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{U}_{\\mathbf{u}}^{n}\\}_{n=1}^{3}" + }, + { + "bbox": [ + 55, + 558, + 295, + 641 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 55, + 558, + 295, + 641 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{60 \\times 60 \\times d}" + }, + { + "bbox": [ + 55, + 558, + 295, + 641 + ], + "type": "text", + "content": " dimension and add them (via " + }, + { + "bbox": [ + 55, + 558, + 295, + 641 + ], + "type": "inline_equation", + "content": "\\alpha_{n}" + }, + { + "bbox": [ + 55, + 558, + 295, + 641 + ], + "type": "text", + "content": ") to form the final feature map. A rapid PCK@5 (mIoU) drop of 28.65 (26.51)% in case of w/o Aggregation Network (Tab. 6) indicates its significance in extracting semantically-meaningful features." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": "[i] Are learnable branch weights necessary? Learnable weights dynamically ensure the optimum importance of each feature-branch for a specific task. Although less pronounced in PCK@5, mAP@200 drops significantly (4.82%) in case of w/o Learnable Weights (Tab. 6), depicting its influence on the final feature map quality." + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 318, + 71, + 552, + 164 + ], + "blocks": [ + { + "bbox": [ + 76, + 133, + 274, + 144 + ], + "lines": [ + { + "bbox": [ + 76, + 133, + 274, + 144 + ], + "spans": [ + { + "bbox": [ + 76, + 133, + 274, + 144 + ], + "type": "text", + "content": "Table 5. Results for sketch-based image segmentation." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 318, + 71, + 552, + 164 + ], + "lines": [ + { + "bbox": [ + 318, + 71, + 552, + 164 + ], + "spans": [ + { + "bbox": [ + 318, + 71, + 552, + 164 + ], + "type": "table", + "html": "
MethodsCorrespondenceSegmentationZS-SBIRRecognition
PCK@5mIoUmAP@200Acc.@1
SD [71]v1.468.9157.630.72878.82
v1.569.2458.910.73579.59
v2.170.3160.120.76187.02
CLIP [67]ViT-B/1667.7556.380.74377.37
ViT-B/3269.4158.470.75878.84
ViT-L/1470.3160.120.76187.02
w/o Aggregation Network54.6547.520.58766.71
w/o Learnable Weights68.7558.130.72683.79
w/o 1D Convolutions55.9747.240.60271.75
Ours-full70.3160.120.76187.02
", + "image_path": "13eac0e3773a02a8ca12234bd8ca3939a9151bc0b9c5de9ace2a3388abe5d4dc.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 381, + 165, + 487, + 175 + ], + "lines": [ + { + "bbox": [ + 381, + 165, + 487, + 175 + ], + "spans": [ + { + "bbox": [ + 381, + 165, + 487, + 175 + ], + "type": "text", + "content": "Table 6. Ablation on design." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 177, + 555, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 177, + 555, + 417 + ], + "spans": [ + { + "bbox": [ + 313, + 177, + 555, + 417 + ], + "type": "text", + "content": "[iii] Contribution of " + }, + { + "bbox": [ + 313, + 177, + 555, + 417 + ], + "type": "inline_equation", + "content": "1D" + }, + { + "bbox": [ + 313, + 177, + 555, + 417 + ], + "type": "text", + "content": " Convolutions. To assess the effectiveness of " + }, + { + "bbox": [ + 313, + 177, + 555, + 417 + ], + "type": "inline_equation", + "content": "1D" + }, + { + "bbox": [ + 313, + 177, + 555, + 417 + ], + "type": "text", + "content": " convolution layers, we use simple linear interpolation to match CLIP and SD UNet feature dimensions, before injection. Sharp PCK@5 drop of " + }, + { + "bbox": [ + 313, + 177, + 555, + 417 + ], + "type": "inline_equation", + "content": "25.62\\%" + }, + { + "bbox": [ + 313, + 177, + 555, + 417 + ], + "type": "text", + "content": " for w/o " + }, + { + "bbox": [ + 313, + 177, + 555, + 417 + ], + "type": "inline_equation", + "content": "1D" + }, + { + "bbox": [ + 313, + 177, + 555, + 417 + ], + "type": "text", + "content": " Convolutions (Tab. 6) reveals its impact on influencing the diffusion denoising process. We posit that the learnable " + }, + { + "bbox": [ + 313, + 177, + 555, + 417 + ], + "type": "inline_equation", + "content": "1D" + }, + { + "bbox": [ + 313, + 177, + 555, + 417 + ], + "type": "text", + "content": " convolutions not only match the feature dimension but also enrich the CLIP features before injection. [iv] Model variants. The performance of our proposed feature extractor depends crucially on the choice of backbone model variant. Accordingly, we explore with multiple pretrained SD and CLIP backbones. Tab. 6 shows SD v2.1 and CLIP ViT-L/14 to achieve the highest PCK@5 and Acc.@1 for sketch-photo correspondence and recognition respectively. Additionally, we also experiment by replacing the CLIP visual encoder of our method with a DINOv2 [61] ViT-S/14 encoder for the task of ZS-SBIR, to observe a " + }, + { + "bbox": [ + 313, + 177, + 555, + 417 + ], + "type": "inline_equation", + "content": "32.80\\%" + }, + { + "bbox": [ + 313, + 177, + 555, + 417 + ], + "type": "text", + "content": " drop in mAP@200. This is probably due to CLIP's [67] multimodal pertaining (diminishing sketch-photo domain gap), over DINOv2's [61] self-supervised one." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 418, + 554, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 418, + 554, + 453 + ], + "spans": [ + { + "bbox": [ + 313, + 418, + 554, + 453 + ], + "type": "text", + "content": "[v] Which timestep value is the most effective? Different diffusion timesteps yield features with drastically different quality [84, 107]. Calculating PCK@5 and mAP@200" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 314, + 453, + 448, + 537 + ], + "blocks": [ + { + "bbox": [ + 314, + 453, + 448, + 537 + ], + "lines": [ + { + "bbox": [ + 314, + 453, + 448, + 537 + ], + "spans": [ + { + "bbox": [ + 314, + 453, + 448, + 537 + ], + "type": "image", + "image_path": "19c7efbf2105c1178a587d2ccc1d25a9ff1969dac54da7cc2b39b28f322ff58c.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 538, + 554, + 573 + ], + "lines": [ + { + "bbox": [ + 313, + 538, + 554, + 573 + ], + "spans": [ + { + "bbox": [ + 313, + 538, + 554, + 573 + ], + "type": "text", + "content": "Figure 5. Choice of timestep " + }, + { + "bbox": [ + 313, + 538, + 554, + 573 + ], + "type": "inline_equation", + "content": "(t)" + }, + { + "bbox": [ + 313, + 538, + 554, + 573 + ], + "type": "text", + "content": ". method to be robust to the choice of " + }, + { + "bbox": [ + 313, + 538, + 554, + 573 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 538, + 554, + 573 + ], + "type": "text", + "content": ", where PCK@5 (mAP@200) scores from a wide range of " + }, + { + "bbox": [ + 313, + 538, + 554, + 573 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 538, + 554, + 573 + ], + "type": "text", + "content": " values consistently surpass the baselines." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 454, + 453, + 555, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 454, + 453, + 555, + 548 + ], + "spans": [ + { + "bbox": [ + 454, + 453, + 555, + 548 + ], + "type": "text", + "content": "for the tasks of correspondence learning and ZS-SBIR respectively, over varying " + }, + { + "bbox": [ + 454, + 453, + 555, + 548 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 454, + 453, + 555, + 548 + ], + "type": "text", + "content": ", we find that " + }, + { + "bbox": [ + 454, + 453, + 555, + 548 + ], + "type": "inline_equation", + "content": "t = 195" + }, + { + "bbox": [ + 454, + 453, + 555, + 548 + ], + "type": "text", + "content": " produces the best results for both tasks. Fig. 5 shows our method to be robust to" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 578, + 388, + 590 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 578, + 388, + 590 + ], + "spans": [ + { + "bbox": [ + 313, + 578, + 388, + 590 + ], + "type": "text", + "content": "8. Conclusion" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 312, + 594, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 594, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 312, + 594, + 555, + 713 + ], + "type": "text", + "content": "In this paper, we identified SD's limitations in understanding abstract, sparse sketches. Integrating CLIP's semantic features with SD's spatially-aware features, we developed a framework that dynamically combines these strengths. This significantly improves performance across sketch recognition, SBIR, and sketch-based segmentation. Our method highlights the complementary nature of different foundation models and introduces an adaptive feature aggregation network, establishing a new standard for universal sketch feature representation for future research." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2563" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 66, + 91, + 296, + 133 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 91, + 296, + 133 + ], + "spans": [ + { + "bbox": [ + 66, + 91, + 296, + 133 + ], + "type": "text", + "content": "[1] Hyojin Bahng, Ali Jahanian, Swami Sankaranarayanan, and Phillip Isola. Exploring visual prompts for adapting large-scale models. arXiv preprint arXiv:2203.17274, 2022.3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 136, + 295, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 136, + 295, + 190 + ], + "spans": [ + { + "bbox": [ + 66, + 136, + 295, + 190 + ], + "type": "text", + "content": "[2] Hmrishav Bandyopadhyay, Subhadeep Koley, Ayan Das, Ayan Kumar Bhunia, Aneeshan Sain, Pinaki Nath Chowdhury, Tao Xiang, and Yi-Zhe Song. Doodle Your 3D: From Abstract Freehand Sketches to Precise 3D Shapes. In CVPR, 2024. 2, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 66, + 191, + 295, + 234 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 191, + 295, + 234 + ], + "spans": [ + { + "bbox": [ + 66, + 191, + 295, + 234 + ], + "type": "text", + "content": "[3] Dmitry Baranchuk, Ivan Rubachev, Andrey Voynov, Valentin Khrulkov, and Artem Babenko. Label-Efficient Semantic Segmentation with Diffusion Models. In ICLR, 2021. 2, 4, 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 66, + 236, + 294, + 280 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 236, + 294, + 280 + ], + "spans": [ + { + "bbox": [ + 66, + 236, + 294, + 280 + ], + "type": "text", + "content": "[4] Ayan Kumar Bhunia, Pinaki Nath Chowdhury, Yongxin Yang, Timothy Hospedales, Tao Xiang, and Yi-Zhe Song. Vectorization and Rasterization: Self-Supervised Learning for Sketch and Handwriting. In CVPR, 2021. 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 66, + 281, + 294, + 324 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 281, + 294, + 324 + ], + "spans": [ + { + "bbox": [ + 66, + 281, + 294, + 324 + ], + "type": "text", + "content": "[5] Ayan Kumar Bhunia, Viswanatha Reddy Gajjala, Subhadeep Koley, Rohit Kundu, Aneeshan Sain, Tao Xiang, and Yi-Zhe Song. Doodle It Yourself: Class Incremental Learning by Drawing a Few Sketches. In CVPR, 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 66, + 325, + 295, + 378 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 325, + 295, + 378 + ], + "spans": [ + { + "bbox": [ + 66, + 325, + 295, + 378 + ], + "type": "text", + "content": "[6] Ayan Kumar Bhunia, Subhadeep Koley, Abdullah Faiz Ur Rahman Khilji, Aneeshan Sain, Pinaki Nath Chowdhury, Tao Xiang, and Yi-Zhe Song. Sketching Without Worrying: Noise-Tolerant Sketch-Based Image Retrieval. In CVPR, 2022. 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 66, + 380, + 295, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 380, + 295, + 423 + ], + "spans": [ + { + "bbox": [ + 66, + 380, + 295, + 423 + ], + "type": "text", + "content": "[7] Ayan Kumar Bhunia, Subhadeep Koley, Amandeep Kumar, Aneeshan Sain, Pinaki Nath Chowdhury, Tao Xiang, and Yi-Zhe Song. Sketch2Saliency: Learning to Detect Salient Objects from Human Drawings. In CVPR, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 66, + 425, + 295, + 467 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 425, + 295, + 467 + ], + "spans": [ + { + "bbox": [ + 66, + 425, + 295, + 467 + ], + "type": "text", + "content": "[8] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging Properties in Self-Supervised Vision Transformers. In ICCV, 2021. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 66, + 469, + 295, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 469, + 295, + 522 + ], + "spans": [ + { + "bbox": [ + 66, + 469, + 295, + 522 + ], + "type": "text", + "content": "[9] Dar-Yen Chen, Subhadeep Koley, Aneeshan Sain, Pinaki Nath Chowdhury, Tao Xiang, Ayan Kumar Bhunia, and Yi-Zhe Song. DemoCaricature: Democratising Caricature Generation with a Rough Sketch. In CVPR, 2024. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 61, + 525, + 294, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 525, + 294, + 557 + ], + "spans": [ + { + "bbox": [ + 61, + 525, + 294, + 557 + ], + "type": "text", + "content": "[10] Honglie Chen, Weidi Xie, Triantafyllos Afouras, Arsha Nagrani, Andrea Vedaldi, and Andrew Zisserman. Localizing Visual Sounds the Hard Way. In CVPR, 2021. 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 61, + 558, + 294, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 558, + 294, + 601 + ], + "spans": [ + { + "bbox": [ + 61, + 558, + 294, + 601 + ], + "type": "text", + "content": "[11] Haoxin Chen, Yong Zhang, Xiaodong Cun, Menghan Xia, Xintao Wang, Chao Weng, and Ying Shan. VideoCraft2: Overcoming Data Limitations for High-Quality Video Diffusion Models. In CVPR, 2024. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 61, + 602, + 294, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 602, + 294, + 635 + ], + "spans": [ + { + "bbox": [ + 61, + 602, + 294, + 635 + ], + "type": "text", + "content": "[12] Gene Chou, Yuval Bahat, and Felix Heide. Diffusion-SDF: Conditional Generative Modeling of Signed Distance Functions. In CVPR, 2023. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 61, + 636, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 636, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 61, + 636, + 294, + 689 + ], + "type": "text", + "content": "[13] Pinaki Nath Chowdhury, Ayan Kumar Bhunia, Viswanatha Reddy Gajjala, Aneeshan Sain, Tao Xiang, and Yi-Zhe Song. Partially Does It: Towards Scene-Level FG-SBIR with Partial Input. In CVPR, 2022. 2, 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 61, + 691, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 691, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 61, + 691, + 294, + 713 + ], + "type": "text", + "content": "[14] Pinaki Nath Chowdhury, Tuanfeng Wang, Duygu Ceylan, Yi-Zhe Song, and Yulia Gryaditskaya. Garment ideation:" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 320, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 339, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 339, + 73, + 553, + 95 + ], + "type": "text", + "content": "Iterative view-aware sketch-based garment modeling. In 3DV, 2022. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 321, + 97, + 553, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 97, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 321, + 97, + 553, + 140 + ], + "type": "text", + "content": "[15] Pinaki Nath Chowdhury, Ayan Kumar Bhunia, Aneshan Sain, Subhadeep Koley, Tao Xiang, and Yi-Zhe Song. SceneTrilogy: On Human Scene-Sketch and its Complementarity with Photo and Text. In CVPR, 2023. 4, 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 320, + 142, + 553, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 142, + 553, + 186 + ], + "spans": [ + { + "bbox": [ + 320, + 142, + 553, + 186 + ], + "type": "text", + "content": "[16] Pinaki Nath Chowdhury, Ayan Kumar Bhunia, Aneshan Sain, Subhadeep Koley, Tao Xiang, and Yi-Zhe Song. What Can Human Sketches Do for Object Detection? In CVPR, 2023. 2, 3, 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 320, + 188, + 553, + 220 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 188, + 553, + 220 + ], + "spans": [ + { + "bbox": [ + 320, + 188, + 553, + 220 + ], + "type": "text", + "content": "[17] John P. Collomosse, Tu Bui, and Hailin Jin. LiveSketch: Query Perturbations for Guided Sketch-based Visual Search. In CVPR, 2019. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 320, + 222, + 553, + 265 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 222, + 553, + 265 + ], + "spans": [ + { + "bbox": [ + 320, + 222, + 553, + 265 + ], + "type": "text", + "content": "[18] Bram de Wilde, Anindo Saha, Richard PG ten Broek, and Henkjan Huisman. Medical diffusion on a budget: Textual Inversion for medical image generation. In MIDL, 2024. 2, 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 320, + 267, + 553, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 267, + 553, + 300 + ], + "spans": [ + { + "bbox": [ + 320, + 267, + 553, + 300 + ], + "type": "text", + "content": "[19] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. ImageNet: A Large-Scale Hierarchical Image Database. In CVPR, 2009. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 320, + 302, + 553, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 302, + 553, + 335 + ], + "spans": [ + { + "bbox": [ + 320, + 302, + 553, + 335 + ], + "type": "text", + "content": "[20] Sounak Dey, Pau Riba, Anjan Dutta, Josep Llados, and Yi-Zhe Song. Doodle to Search: Practical Zero-Shot Sketch-based Image Retrieval. In CVPR, 2019. 2, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 320, + 337, + 553, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 337, + 553, + 358 + ], + "spans": [ + { + "bbox": [ + 320, + 337, + 553, + 358 + ], + "type": "text", + "content": "[21] Prafulla Dhariwal and Alexander Nichol. Diffusion Models Beat GANs on Image Synthesis. In NeurIPS, 2021. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 320, + 360, + 553, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 360, + 553, + 404 + ], + "spans": [ + { + "bbox": [ + 320, + 360, + 553, + 404 + ], + "type": "text", + "content": "[22] Francesco Di Sario, Riccardo Renzulli, Enzo Tartaglione, and Marco Grangetto. Boost Your NeRF: A Model-Agnostic Mixture of Experts Framework for High Quality and Efficient Rendering. In ECCV, 2024. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 320, + 406, + 553, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 406, + 553, + 449 + ], + "spans": [ + { + "bbox": [ + 320, + 406, + 553, + 449 + ], + "type": "text", + "content": "[23] Shiyin Dong, Mingrui Zhu, Kun Cheng, Nannan Wang, and Xinbo Gao. Bridging generative and discriminative models for unified visual perception with diffusion priors. In IJCAI, 2024. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 320, + 452, + 553, + 516 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 452, + 553, + 516 + ], + "spans": [ + { + "bbox": [ + 320, + 452, + 553, + 516 + ], + "type": "text", + "content": "[24] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 320, + 519, + 553, + 551 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 519, + 553, + 551 + ], + "spans": [ + { + "bbox": [ + 320, + 519, + 553, + 551 + ], + "type": "text", + "content": "[25] Anjan Dutta and Zeynep Akata. Semantically Tied Paired Cycle Consistency for Zero-Shot Sketch-based Image Retrieval. In CVPR, 2019. 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 320, + 553, + 553, + 574 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 553, + 553, + 574 + ], + "spans": [ + { + "bbox": [ + 320, + 553, + 553, + 574 + ], + "type": "text", + "content": "[26] Mathias Eitz, James Hays, and Marc Alexa. How do humans sketch objects? ACM TOG, 2012. 1, 5, 6" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 320, + 577, + 553, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 577, + 553, + 620 + ], + "spans": [ + { + "bbox": [ + 320, + 577, + 553, + 620 + ], + "type": "text", + "content": "[27] Arnab Ghosh, Richard Zhang, Puneet K Dokania, Oliver Wang, Alexei A Efros, Philip HS Torr, and Eli Shechtman. Interactive Sketch & Fill: Multiclass Sketch-to-Image Translation. In ICCV, 2019. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 320, + 622, + 553, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 622, + 553, + 644 + ], + "spans": [ + { + "bbox": [ + 320, + 622, + 553, + 644 + ], + "type": "text", + "content": "[28] Rafael C Gonzalez and Richard E Woods. Digital Image Processing. Pearson, 2008. 4" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 320, + 646, + 553, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 646, + 553, + 667 + ], + "spans": [ + { + "bbox": [ + 320, + 646, + 553, + 667 + ], + "type": "text", + "content": "[29] David Ha and Douglas Eck. A Neural Representation of Sketch Drawings. In ICLR, 2017. 5, 6" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 320, + 670, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 670, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 320, + 670, + 553, + 712 + ], + "type": "text", + "content": "[30] Moayed Haji-Ali, Guha Balakrishnan, and Vicente Ordonez. ElasticDiffusion: Training-free Arbitrary Size Image Generation through Global-Local Content Separation. In CVPR, 2024. 3" + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2564" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 61, + 73, + 294, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 73, + 294, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 73, + 294, + 115 + ], + "spans": [ + { + "bbox": [ + 61, + 73, + 294, + 115 + ], + "type": "text", + "content": "[31] Cusuh Ham, Gemma Canet Tarres, Tu Bui, James Hays, Zhe Lin, and John Collomosse. CoGS: Controllable Generation and Search from Sketch and Style. In ECCV, 2022. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 118, + 294, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 118, + 294, + 150 + ], + "spans": [ + { + "bbox": [ + 61, + 118, + 294, + 150 + ], + "type": "text", + "content": "[32] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep Residual Learning for Image Recognition. In CVPR, 2016. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 152, + 294, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 152, + 294, + 194 + ], + "spans": [ + { + "bbox": [ + 62, + 152, + 294, + 194 + ], + "type": "text", + "content": "[33] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-Prompt Image Editing with Cross Attention Control. In ICLR, 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 198, + 294, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 198, + 294, + 217 + ], + "spans": [ + { + "bbox": [ + 62, + 198, + 294, + 217 + ], + "type": "text", + "content": "[34] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising Diffusion Probabilistic Models. In NeurIPS, 2020. 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 220, + 294, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 220, + 294, + 262 + ], + "spans": [ + { + "bbox": [ + 62, + 220, + 294, + 262 + ], + "type": "text", + "content": "[35] Conghui Hu, Da Li, Yongxin Yang, Timothy M Hospedales, and Yi-Zhe Song. Sketch-a-Segmented: Sketch-based Photo Segmenter Generation. IEEE TIP, 2020. 2, 5, 7, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 265, + 294, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 265, + 294, + 319 + ], + "spans": [ + { + "bbox": [ + 62, + 265, + 294, + 319 + ], + "type": "text", + "content": "[36] Drew A Hudson, Daniel Zoran, Mateusz Malinowski, Andrew K Lampinen, Andrew Jaegle, James L McClelland, Loic Matthey, Felix Hill, and Alexander Lerchner. SODA: Bottleneck Diffusion Models for Representation Learning. In CVPR, 2024. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 321, + 294, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 321, + 294, + 364 + ], + "spans": [ + { + "bbox": [ + 62, + 321, + 294, + 364 + ], + "type": "text", + "content": "[37] HyeongJoo Hwang, Geon-Hyeong Kim, Seunghoon Hong, and Kee-Eung Kim. Variational Interaction Information Maximization for Cross-domain Disentanglement. In NeurIPS, 2020. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 366, + 294, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 366, + 294, + 399 + ], + "spans": [ + { + "bbox": [ + 62, + 366, + 294, + 399 + ], + "type": "text", + "content": "[38] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In ECCV, 2022. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 400, + 294, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 400, + 294, + 432 + ], + "spans": [ + { + "bbox": [ + 62, + 400, + 294, + 432 + ], + "type": "text", + "content": "[39] Tero Karras, Samuli Laine, and Timo Aila. A Style-Based Generator Architecture for Generative Adversarial Networks. In CVPR, 2019. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 434, + 294, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 434, + 294, + 476 + ], + "spans": [ + { + "bbox": [ + 62, + 434, + 294, + 476 + ], + "type": "text", + "content": "[40] Bahjat Kawar, Shiran Zada, Oran Lang, Omer Tov, Huiwen Chang, Tali Dekel, Inbar Mosseri, and Michal Irani. Imagic: Text-Based Real Image Editing with Diffusion Models. In CVPR, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 479, + 294, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 479, + 294, + 555 + ], + "spans": [ + { + "bbox": [ + 62, + 479, + 294, + 555 + ], + "type": "text", + "content": "[41] Kazuma Kobayashi and Lin Gu and Ryuichiro Hataya and Takaaki Mizuno and Mototaka Miyake and Hirokazu Watanabe and Masamichi Takahashi and Yasuyuki Takamizawa and Yukihiro Yoshida and Satoshi Nakamura and Nobuji Kouno and Amina Bolatkan and Yusuke Kurose and Tatsuya Harada and Ryuji Hamamoto. Sketch-based Medical Image Retrieval. Medical Image Analysis, 2024. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 557, + 294, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 557, + 294, + 578 + ], + "spans": [ + { + "bbox": [ + 62, + 557, + 294, + 578 + ], + "type": "text", + "content": "[42] Diederik P Kingma and Max Welling. Auto-Encoding Variational Bayes. In ICLR, 2014. 3, 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 62, + 580, + 294, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 580, + 294, + 622 + ], + "spans": [ + { + "bbox": [ + 62, + 580, + 294, + 622 + ], + "type": "text", + "content": "[43] Subhadeep Koley, Ayan Kumar Bhunia, Aneeshan Sain, Pinaki Nath Chowdhury, Tao Xiang, and Yi-Zhe Song. Picture that Sketch: Photorealistic Image Generation from Abstract Sketches. In CVPR, 2023. 2, 3, 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 62, + 625, + 294, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 625, + 294, + 667 + ], + "spans": [ + { + "bbox": [ + 62, + 625, + 294, + 667 + ], + "type": "text", + "content": "[44] Subhadeep Koley, Ayan Kumar Bhunia, Aneeshan Sain, Pinaki Nath Chowdhury, Tao Xiang, and Yi-Zhe Song. You'll Never Walk Alone: A Sketch and Text Duet for Fine-Grained Image Retrieval. In CVPR, 2024. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 62, + 670, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 670, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 62, + 670, + 294, + 712 + ], + "type": "text", + "content": "[45] Subhadeep Koley, Ayan Kumar Bhunia, Aneeshan Sain, Pinaki Nath Chowdhury, Tao Xiang, and Yi-Zhe Song. How to Handle Sketch-Abstraction in Sketch-Based Image Retrieval? In CVPR, 2024. 1, 3" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 321, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 321, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 321, + 73, + 553, + 116 + ], + "type": "text", + "content": "[46] Subhadeep Koley, Ayan Kumar Bhunia, Aneeshan Sain, Pinaki Nath Chowdhury, Tao Xiang, and Yi-Zhe Song. Text-to-Image Diffusion Models are Great Sketch-Photo Matchmakers. In CVPR, 2024. 1, 2, 3, 4, 5, 6, 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 321, + 118, + 553, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 118, + 553, + 171 + ], + "spans": [ + { + "bbox": [ + 321, + 118, + 553, + 171 + ], + "type": "text", + "content": "[47] Subhadeep Koley, Ayan Kumar Bhunia, Deeptanshu Sekhri, Aneeshan Sain, Pinaki Nath Chowdhury, Tao Xiang, and Yi-Zhe Song. It's All About Your Sketch: Democratising Sketch Control in Diffusion Models. In CVPR, 2024. 1, 2, 4, 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 321, + 173, + 553, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 173, + 553, + 204 + ], + "spans": [ + { + "bbox": [ + 321, + 173, + 553, + 204 + ], + "type": "text", + "content": "[48] Philipp Krahenbuhl and Vladlen Koltun. Efficient Inference in Fully Connected CRFs with Gaussian Edge Potentials. In NeurIPS, 2011. 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 321, + 206, + 553, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 206, + 553, + 237 + ], + "spans": [ + { + "bbox": [ + 321, + 206, + 553, + 237 + ], + "type": "text", + "content": "[49] Alexander C Li, Mihir Prabhudesai, Shivam Duggal, Ellis Brown, and Deepak Pathak. Your Diffusion Model is Secretly a Zero-Shot Classifier. In ICCV, 2023. 2, 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 321, + 239, + 553, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 239, + 553, + 259 + ], + "spans": [ + { + "bbox": [ + 321, + 239, + 553, + 259 + ], + "type": "text", + "content": "[50] Ke Li, Kaiyue Pang, and Yi-Zhe Song. Photo Pre-Training, but for Sketch. In CVPR, 2023. 2, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 321, + 262, + 553, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 262, + 553, + 293 + ], + "spans": [ + { + "bbox": [ + 321, + 262, + 553, + 293 + ], + "type": "text", + "content": "[51] Minchen Li, Alla Sheffer, Eitan Grinspun, and Nicholas Vining. FoldSketch: Enriching Garments with Physically Reproducible Folds. ACM TOG, 2018. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 321, + 295, + 553, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 295, + 553, + 336 + ], + "spans": [ + { + "bbox": [ + 321, + 295, + 553, + 336 + ], + "type": "text", + "content": "[52] Fengyin Lin, Mingkang Li, Da Li, Timothy Hospedales, YiZhe Song, and Yonggang Qi. Zero-Shot Everything Sketch-Based Image Retrieval, and in Explainable Style. In CVPR, 2023. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 321, + 338, + 553, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 338, + 553, + 370 + ], + "spans": [ + { + "bbox": [ + 321, + 338, + 553, + 370 + ], + "type": "text", + "content": "[53] Li Liu, Fumin Shen, Yuming Shen, Xianglong Liu, and Ling Shao. Deep Sketch Hashing: Fast Free-hand Sketch-Based Image Retrieval. In CVPR, 2017. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 321, + 373, + 553, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 373, + 553, + 403 + ], + "spans": [ + { + "bbox": [ + 321, + 373, + 553, + 403 + ], + "type": "text", + "content": "[54] Qing Liu, Lingxi Xie, Huiyu Wang, and Alan Yuille. Semantic-Aware Knowledge Preservation for Zero-Shot Sketch-Based Image Retrieval. In ICCV, 2019. 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 321, + 405, + 553, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 405, + 553, + 436 + ], + "spans": [ + { + "bbox": [ + 321, + 405, + 553, + 436 + ], + "type": "text", + "content": "[55] Xuanchen Lu, Xiaolong Wang, and Judith E Fan. Learning Dense Correspondences between Photos and Sketches. In ICML, 2023. 5, 7, 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 321, + 438, + 553, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 438, + 553, + 469 + ], + "spans": [ + { + "bbox": [ + 321, + 438, + 553, + 469 + ], + "type": "text", + "content": "[56] Ling Luo, Yulia Gryaditskaya, Tao Xiang, and Yi-Zhe Song. Structure-Aware 3D VR Sketch to 3D Shape Retrieval. In 3DV, 2022. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 321, + 472, + 553, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 472, + 553, + 503 + ], + "spans": [ + { + "bbox": [ + 321, + 472, + 553, + 503 + ], + "type": "text", + "content": "[57] Ling Luo, Pinaki Nath Chowdhury, Tao Xiang, Yi-Zhe Song, and Yulia Gryaditskaya. 3D VR Sketch Guided 3D Shape Prototyping and Exploration. In ICCV, 2023. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 321, + 505, + 553, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 505, + 553, + 536 + ], + "spans": [ + { + "bbox": [ + 321, + 505, + 553, + 536 + ], + "type": "text", + "content": "[58] Aryan Mikaeili, Or Perel, Daniel Cohen-Or, and Ali Mahdavi-Amiri. SKED: Sketch-guided Text-based 3D Editing. In CVPR, 2023. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 321, + 537, + 553, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 537, + 553, + 580 + ], + "spans": [ + { + "bbox": [ + 321, + 537, + 553, + 580 + ], + "type": "text", + "content": "[59] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, and Ying Shan. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. In AAAI, 2024. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 321, + 582, + 553, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 582, + 553, + 624 + ], + "spans": [ + { + "bbox": [ + 321, + 582, + 553, + 624 + ], + "type": "text", + "content": "[60] Jishnu Mukhoti, Tsung-Yu Lin, Omid Poursaeed, Rui Wang, Ashish Shah, Philip HS Torr, and Ser-Nam Lim. Open Vocabulary Semantic Segmentation with Patch Aligned Contrastive Learning. In CVPR, 2023. 5" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 321, + 626, + 553, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 626, + 553, + 679 + ], + "spans": [ + { + "bbox": [ + 321, + 626, + 553, + 679 + ], + "type": "text", + "content": "[61] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. DINov2: Learning Robust Visual Features without Supervision. TMLR, 2024. 6, 8" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 321, + 681, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 681, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 321, + 681, + 553, + 712 + ], + "type": "text", + "content": "[62] Zixuan Pan, Jianxu Chen, and Yiyu Shi. Masked Diffusion as Self-supervised Representation Learner. arXiv preprint arXiv:2308.05695, 2023. 2" + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "2565" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 61, + 73, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 73, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 73, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 61, + 73, + 294, + 116 + ], + "type": "text", + "content": "[63] Kaiyue Pang, Ke Li, Yongxin Yang, Honggang Zhang, Timothy M Hospedales, Tao Xiang, and Yi-Zhe Song. Generalising Fine-Grained Sketch-Based Image Retrieval. In CVPR, 2019. 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 118, + 294, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 118, + 294, + 160 + ], + "spans": [ + { + "bbox": [ + 61, + 118, + 294, + 160 + ], + "type": "text", + "content": "[64] Kaiyue Pang, Yongxin Yang, Timothy M Hospedales, Tao Xiang, and Yi-Zhe Song. Solving mixed-modal jigsaw puzzle for fine-grained sketch-based image retrieval. In CVPR, 2020. 2, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 162, + 295, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 162, + 295, + 205 + ], + "spans": [ + { + "bbox": [ + 62, + 162, + 295, + 205 + ], + "type": "text", + "content": "[65] Mihir Prabhudesai, Tsung-Wei Ke, Alexander Cong Li, Deepak Pathak, and Katerina Fragkiadaki. Diffusion-TTA: Test-time Adaptation of Discriminative Models via Generative Feedback. In NeurIPS, 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 206, + 294, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 206, + 294, + 239 + ], + "spans": [ + { + "bbox": [ + 62, + 206, + 294, + 239 + ], + "type": "text", + "content": "[66] Zhiyu Qu, Yulia Gryaditskaya, Ke Li, Kaiyue Pang, Tao Xiang, and Yi-Zhe Song. SketchXAI: A First Look at Explainability for Human Sketches. In CVPR, 2023. 6, 7, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 239, + 294, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 239, + 294, + 293 + ], + "spans": [ + { + "bbox": [ + 62, + 239, + 294, + 293 + ], + "type": "text", + "content": "[67] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning Transferable Visual Models From Natural Language Supervision. In ICML, 2021. 1, 2, 3, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 295, + 294, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 295, + 294, + 326 + ], + "spans": [ + { + "bbox": [ + 62, + 295, + 294, + 326 + ], + "type": "text", + "content": "[68] Leo Sampaio Ferraz Ribeiro, Tu Bui, John Collomosse, and Moacir Ponti. Sketchformer: Transformer-based Representation for Sketched Structure. In CVPR, 2020. 2, 3, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 327, + 294, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 327, + 294, + 370 + ], + "spans": [ + { + "bbox": [ + 62, + 327, + 294, + 370 + ], + "type": "text", + "content": "[69] Elad Richardson, Yuval Alaluf, Or Patashnik, Yotam Nitzan, Yaniv Azar, Stav Shapiro, and Daniel Cohen-Or. Encoding in Style: a StyleGAN Encoder for Image-to-Image Translation. In CVPR, 2021. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 372, + 294, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 372, + 294, + 403 + ], + "spans": [ + { + "bbox": [ + 62, + 372, + 294, + 403 + ], + "type": "text", + "content": "[70] Ignacio Rocco, Relja Arandjelović, and Josef Sivic. End-to-end weakly-supervised semantic alignment. In CVPR, 2018. 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 405, + 294, + 447 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 405, + 294, + 447 + ], + "spans": [ + { + "bbox": [ + 62, + 405, + 294, + 447 + ], + "type": "text", + "content": "[71] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-Resolution Image Synthesis with Latent Diffusion Models. In CVPR, 2022. 1, 2, 3, 4, 5, 6, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 449, + 294, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 449, + 294, + 480 + ], + "spans": [ + { + "bbox": [ + 62, + 449, + 294, + 480 + ], + "type": "text", + "content": "[72] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-Net: Convolutional Networks for Biomedical Image Segmentation. In MICCAI, 2015. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 483, + 294, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 483, + 294, + 525 + ], + "spans": [ + { + "bbox": [ + 62, + 483, + 294, + 525 + ], + "type": "text", + "content": "[73] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. DreamBooth: Fine Tuning Text-to-Image Diffusion Models for Subject-Driven Generation. In CVPR, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 526, + 294, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 526, + 294, + 568 + ], + "spans": [ + { + "bbox": [ + 62, + 526, + 294, + 568 + ], + "type": "text", + "content": "[74] Aneeshan Sain, Ayan Kumar Bhunia, Yongxin Yang, Tao Xiang, and Yi-Zhe Song. StyleMeUp: Towards Style-Agnostic Sketch-Based Image Retrieval. In CVPR, 2021. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 62, + 571, + 294, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 571, + 294, + 613 + ], + "spans": [ + { + "bbox": [ + 62, + 571, + 294, + 613 + ], + "type": "text", + "content": "[75] Aneeshan Sain, Ayan Kumar Bhunia, Vaishnav Potlapalli, Pinaki Nath Chowdhury, Tao Xiang, and Yi-Zhe Song. Sketch3T: Test-Time Training for Zero-Shot SBIR. In CVPR, 2022. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 62, + 615, + 294, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 615, + 294, + 668 + ], + "spans": [ + { + "bbox": [ + 62, + 615, + 294, + 668 + ], + "type": "text", + "content": "[76] Aneeshan Sain, Ayan Kumar Bhunia, Pinaki Nath Chowdhury, Aneeshan Sain, Subhadeep Koley, Tao Xiang, and Yi-Zhe Song. CLIP for All Things Zero-Shot Sketch-Based Image Retrieval, Fine-Grained or Not. In CVPR, 2023. 1, 2, 3, 5, 6, 8" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 62, + 670, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 670, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 62, + 670, + 294, + 712 + ], + "type": "text", + "content": "[77] Aneeshan Sain, Ayan Kumar Bhunia, Subhadeep Koley, Pinaki Nath Chowdhury, Soumitri Chattopadhyay, Tao Xiang, and Yi-Zhe Song. Exploiting Unlabelled Photos for Stronger Fine-Grained SBIR. In CVPR, 2023. 3" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 321, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 321, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 321, + 73, + 553, + 106 + ], + "type": "text", + "content": "[78] Patsorn Sangkloy, Nathan Burnell, Cusuh Ham, and James Hays. The Sketchy Database: Learning to Retrieve Badly Drawn Bunnies. ACM TOG, 2016. 1, 5, 6, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 321, + 108, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 108, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 321, + 108, + 553, + 150 + ], + "type": "text", + "content": "[79] Patsorn Sangkloy, Wittawat Jitkrittum, Diyi Yang, and James Hays. A Sketch Is Worth a Thousand Words: Image Retrieval with Text and Sketch. In ECCV, 2022. 2, 3, 4, 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 321, + 152, + 553, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 152, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 321, + 152, + 553, + 217 + ], + "type": "text", + "content": "[80] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. LAION-5B: An open large-scale dataset for training next generation image-text models. In NeurlPSW, 2022. 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 321, + 219, + 553, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 219, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 321, + 219, + 553, + 251 + ], + "type": "text", + "content": "[81] Jaehyeok Shim, Changwoo Kang, and Kyungdon Joo. Diffusion-Based Signed Distance Fields for 3D Shape Generation. In CVPR, 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 321, + 253, + 553, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 253, + 553, + 274 + ], + "spans": [ + { + "bbox": [ + 321, + 253, + 553, + 274 + ], + "type": "text", + "content": "[82] Chenyang Si, Ziqi Huang, Yuming Jiang, and Ziwei Liu. FreeU: Free Lunch in Diffusion U-Net. In CVPR, 2024. 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 321, + 276, + 553, + 318 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 276, + 553, + 318 + ], + "spans": [ + { + "bbox": [ + 321, + 276, + 553, + 318 + ], + "type": "text", + "content": "[83] Jifei Song, Qian Yu, Yi-Zhe Song, Tao Xiang, and Timothy M Hospedales. Deep Spatial-Semantic Attention for Fine-Grained Sketch-Based Image Retrieval. In ICCV, 2017. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 321, + 320, + 553, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 320, + 553, + 352 + ], + "spans": [ + { + "bbox": [ + 321, + 320, + 553, + 352 + ], + "type": "text", + "content": "[84] Luming Tang, Menglin Jia, Qianqian Wang, Cheng Perng Phoo, and Bharath Hariharan. Emergent Correspondence from Image Diffusion. In NeurIPS, 2023. 1, 2, 4, 5, 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 321, + 354, + 553, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 354, + 553, + 386 + ], + "spans": [ + { + "bbox": [ + 321, + 354, + 553, + 386 + ], + "type": "text", + "content": "[85] Prune Truong, Martin Danelljan, Fisher Yu, and Luc Van Gool. Warp Consistency for Unsupervised Learning of Dense Correspondences. In ICCV, 2021. 7, 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 321, + 388, + 553, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 388, + 553, + 420 + ], + "spans": [ + { + "bbox": [ + 321, + 388, + 553, + 420 + ], + "type": "text", + "content": "[86] Narek Tumanyan, Michal Geyer, Shai Bagon, and Tali Dekel. Plug-and-Play Diffusion Features for Text-Driven Image-to-Image Translation. In CVPR, 2023. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 321, + 422, + 553, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 422, + 553, + 453 + ], + "spans": [ + { + "bbox": [ + 321, + 422, + 553, + 453 + ], + "type": "text", + "content": "[87] Andrey Voynov, Kfir Aberman, and Daniel Cohen-Or. Sketch-guided text-to-image diffusion models. In SIGGRAPH, 2023. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 321, + 456, + 553, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 456, + 553, + 477 + ], + "spans": [ + { + "bbox": [ + 321, + 456, + 553, + 477 + ], + "type": "text", + "content": "[88] Fei Wang. Sketch2Vox: Learning 3D Reconstruction from a Single Monocular Sketch Image. In ECCV, 2024. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 321, + 479, + 553, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 479, + 553, + 510 + ], + "spans": [ + { + "bbox": [ + 321, + 479, + 553, + 510 + ], + "type": "text", + "content": "[89] Wenxuan Wang, Quan Sun, Fan Zhang, Yepeng Tang, Jing Liu, and Xinlong Wang. Diffusion Feedback Helps CLIP See Better. In ICLR, 2025. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 321, + 513, + 553, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 513, + 553, + 555 + ], + "spans": [ + { + "bbox": [ + 321, + 513, + 553, + 555 + ], + "type": "text", + "content": "[90] Size Wu, Wenwei Zhang, Lumin Xu, Sheng Jin, Xiangtai Li, Wentao Liu, and Chen Change Loy. CLIPSelf: Vision Transformer Distills Itself for Open-Vocabulary Dense Prediction. In ICLR, 2024. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 321, + 557, + 553, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 557, + 553, + 600 + ], + "spans": [ + { + "bbox": [ + 321, + 557, + 553, + 600 + ], + "type": "text", + "content": "[91] Chunlong Xia, Xinliang Wang, Feng Lv, Xin Hao, and Yifeng Shi. ViT-CoMer: Vision Transformer with Convolutional Multi-scale Feature Interaction for Dense Predictions. In CVPR, 2024. 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 321, + 602, + 553, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 602, + 553, + 634 + ], + "spans": [ + { + "bbox": [ + 321, + 602, + 553, + 634 + ], + "type": "text", + "content": "[92] Jinbo Xing, Hanyuan Liu, Menghan Xia, Yong Zhang, Xintao Wang, Ying Shan, and Tien-Tsin Wong. ToonCrafter: Generative Cartoon Interpolation. ACM TOG, 2024. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 321, + 636, + 553, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 636, + 553, + 668 + ], + "spans": [ + { + "bbox": [ + 321, + 636, + 553, + 668 + ], + "type": "text", + "content": "[93] Chenfeng Xu, Huan Ling, Sanja Fidler, and Or Litany. 3DiffTec: 3D Object Detection with Geometry-Aware Diffusion Features. In CVPR, 2023. 2, 4, 5" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 321, + 670, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 670, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 321, + 670, + 553, + 712 + ], + "type": "text", + "content": "[94] Jiarui Xu, Sifei Liu, Arash Vahdat, Wonmin Byeon, Xiaolong Wang, and Shalini De Mello. Open-Vocabulary Panoptic Segmentation with Text-to-Image Diffusion Models. In CVPR, 2023. 1, 2, 4" + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "2566" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 57, + 72, + 297, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 72, + 297, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 72, + 297, + 117 + ], + "spans": [ + { + "bbox": [ + 61, + 72, + 297, + 117 + ], + "type": "text", + "content": "[95] Mengde Xu, Zheng Zhang, Fangyun Wei, Yutong Lin, Yue Cao, Han Hu, and Xiang Bai. A Simple Baseline for Open-Vocabulary Semantic Segmentation with Pre-trained Vision-language Model. In ECCV, 2022. 7, 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 119, + 295, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 119, + 295, + 163 + ], + "spans": [ + { + "bbox": [ + 61, + 119, + 295, + 163 + ], + "type": "text", + "content": "[96] Peng Xu, Yongye Huang, Tongtong Yuan, Kaiyue Pang, YiZhe Song, Tao Xiang, Timothy M Hospedales, Zhanyu Ma, and Jun Guo. SketchMate: Deep Hashing for Million-Scale Human Sketch Retrieval. In CVPR, 2018. 6, 7, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 165, + 295, + 198 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 165, + 295, + 198 + ], + "spans": [ + { + "bbox": [ + 62, + 165, + 295, + 198 + ], + "type": "text", + "content": "[97] Peng Xu, Timothy Hospedales, Qiyue Yin, Yi-Zhe Song, Tao Xiang, and Liang Wang. Deep Learning for Free-Hand Sketch: A Survey. IEEE TPAMI, 2020. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 200, + 295, + 253 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 200, + 295, + 253 + ], + "spans": [ + { + "bbox": [ + 62, + 200, + 295, + 253 + ], + "type": "text", + "content": "[98] Antoine Yang, Arsha Nagrani, Paul Hongsuck Seo, Antoine Miech, Jordi Pont-Tuset, Ivan Laptev, Josef Sivic, and Cordelia Schmid. Vid2Seq: Large-Scale Pretraining of a Visual Language Model for Dense Video Captioning. In CVPR, 2023. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 256, + 295, + 299 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 256, + 295, + 299 + ], + "spans": [ + { + "bbox": [ + 62, + 256, + 295, + 299 + ], + "type": "text", + "content": "[99] Kai Yang, Jian Tao, Jiafei Lyu, Chunjiang Ge, Jiaxin Chen, Weihan Shen, Xiaolong Zhu, and Xiu Li. Using human feedback to fine-tune diffusion models without any reward model. In CVPR, 2024. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 57, + 302, + 294, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 302, + 294, + 334 + ], + "spans": [ + { + "bbox": [ + 57, + 302, + 294, + 334 + ], + "type": "text", + "content": "[100] Lan Yang, Kaiyue Pang, Honggang Zhang, and Yi-Zhe Song. SketchAA: Abstract Representation for Abstract Sketches. In ICCV, 2021. 1, 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 57, + 337, + 295, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 337, + 295, + 380 + ], + "spans": [ + { + "bbox": [ + 57, + 337, + 295, + 380 + ], + "type": "text", + "content": "[101] Lumin Yang, Jiajie Zhuang, Hongbo Fu, Xiangzhi Wei, Kun Zhou, and Youyi Zheng. SketchGNN: Semantic Sketch Segmentation with Graph Neural Networks. ACM TOG, 2021. 6, 7, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 57, + 382, + 295, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 382, + 295, + 415 + ], + "spans": [ + { + "bbox": [ + 57, + 382, + 295, + 415 + ], + "type": "text", + "content": "[102] Sasi Kiran Yelamarthi, Shiva Krishna Reddy, Ashish Mishra, and Anurag Mittal. A Zero-Shot Framework for Sketch Based Image Retrieval. In ECCV, 2018. 2, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 57, + 418, + 295, + 450 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 418, + 295, + 450 + ], + "spans": [ + { + "bbox": [ + 57, + 418, + 295, + 450 + ], + "type": "text", + "content": "[103] Dong Yin, Raphael Gontijo Lopes, Jon Shlens, Ekin Dogus Cubuk, and Justin Gilmer. A Fourier Perspective on Model Robustness in Computer Vision. In NeurIPS, 2019. 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 57, + 452, + 295, + 495 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 452, + 295, + 495 + ], + "spans": [ + { + "bbox": [ + 57, + 452, + 295, + 495 + ], + "type": "text", + "content": "[104] Jiazuo Yu, Yunzhi Zhuge, Lu Zhang, Ping Hu, Dong Wang, Huchuan Lu, and You He. Boosting Continual Learning of Vision-Language Models via Mixture-of-Experts Adapters. In CVPR, 2024. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 498, + 295, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 498, + 295, + 529 + ], + "spans": [ + { + "bbox": [ + 57, + 498, + 295, + 529 + ], + "type": "text", + "content": "[105] Seonghoon Yu, Paul Hongsuck Seo, and Jeany Son. Zeroshot Referring Image Segmentation with Global-Local Context Features. In CVPR, 2023. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 533, + 295, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 533, + 295, + 576 + ], + "spans": [ + { + "bbox": [ + 57, + 533, + 295, + 576 + ], + "type": "text", + "content": "[106] Junyi Zhang, Charles Herrmann, Junhwa Hur, Eric Chen, Varun Jampani, Deqing Sun, and Ming-Hsuan Yang. Telling left from right: Identifying geometry-aware semantic correspondence. In CVPR, 2024. 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 578, + 295, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 578, + 295, + 632 + ], + "spans": [ + { + "bbox": [ + 57, + 578, + 295, + 632 + ], + "type": "text", + "content": "[107] Junyi Zhang, Charles Herrmann, Junhwa Hur, Luisa Polania Cabrera, Varun Jampani, Deqing Sun, and Ming-Hsuan Yang. A Tale of Two Features: Stable Diffusion Complements DINO for Zero-Shot Semantic Correspondence. NeurIPS, 2024. 1, 2, 3, 4, 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 635, + 295, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 635, + 295, + 666 + ], + "spans": [ + { + "bbox": [ + 57, + 635, + 295, + 666 + ], + "type": "text", + "content": "[108] Lvmin Zhang and Maneesh Agrawala. Adding Conditional Control to Text-to-Image Diffusion Models. In ICCV, 2023. 2, 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 670, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 295, + 712 + ], + "type": "text", + "content": "[109] Manyuan Zhang, Guanglu Song, Xiaoyu Shi, Yu Liu, and Hongsheng Li. Three Things We Need to Know About Transferring Stable Diffusion to Visual Dense Prediction Tasks. In ECCV, 2024. 2" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 173 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "text", + "content": "[110] Wenliang Zhao, Yongming Rao, Zuyan Liu, Benlin Liu, Jie Zhou, and Jiwen Lu. Unleashing Text-to-Image Diffusion Models for Visual Perception. In ICCV, 2023. 1, 2, 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 108, + 553, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 108, + 553, + 139 + ], + "spans": [ + { + "bbox": [ + 316, + 108, + 553, + 139 + ], + "type": "text", + "content": "[111] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Zwei Liu. Conditional Prompt Learning for Vision-Language Models. In CVPR, 2022. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 141, + 553, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 553, + 173 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 553, + 173 + ], + "type": "text", + "content": "[112] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Zwei Liu. Learning to Prompt for Vision-Language Models. IJCV, 2022. 3" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2567" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/SketchVideo_ Sketch-based Video Generation and Editing/86ad126c-969d-47b0-8162-a7cc53739116_content_list.json b/2025/SketchVideo_ Sketch-based Video Generation and Editing/86ad126c-969d-47b0-8162-a7cc53739116_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..e587146c6502a7ef8787e07928d3387f6524a983 --- /dev/null +++ b/2025/SketchVideo_ Sketch-based Video Generation and Editing/86ad126c-969d-47b0-8162-a7cc53739116_content_list.json @@ -0,0 +1,1917 @@ +[ + { + "type": "text", + "text": "SketchVideo: Sketch-based Video Generation and Editing", + "text_level": 1, + "bbox": [ + 204, + 130, + 790, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Feng-Lin Liu $^{1,2}$ Hongbo Fu $^{3}$ Xintao Wang $^{4}$ Weicai Ye $^{4}$ Pengfei Wan $^{4}$ Di Zhang $^{4}$ Lin Gao $^{1,2*}$", + "bbox": [ + 99, + 179, + 913, + 200 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Beijing Key Laboratory of Mobile Computing and Pervasive Device, Institute of Computing Technology, Chinese Academy of Sciences $^{2}$ University of Chinese Academy of Sciences $^{3}$ Hong Kong University of Science and Technology $^{4}$ Kuaishou Technology", + "bbox": [ + 104, + 205, + 913, + 234 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/caeda09bc9fe18f0c8cf1cb0c674b0ab9ffcda319bd8f58736f66b52e41aed53.jpg", + "image_caption": [], + "image_footnote": [ + "\"A close-up of a majestic male lion's face, with a tawny mane, piercing amber eyes, and subtle scars, bathed in golden sunlight, showcasing his power and wild spirit.\"" + ], + "bbox": [ + 238, + 244, + 377, + 332 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5e578ebc2b971b8498cc9bfac9815da0080c43814477fcc68cdc5b4ea6de9faf.jpg", + "image_caption": [], + "image_footnote": [ + "\"The video is a drone shot of a serene landscape featuring a deep blue lagoon surrounded by lush green trees and rocky cliffs. The lagoon is nestled between the mountains, creating a picturesque scene.\"" + ], + "bbox": [ + 114, + 332, + 196, + 419 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0b935ee7d70550e86839543a0c26f29b2fa26098d49d92ad7f71b429971a222d.jpg", + "image_caption": [ + "(a) Sketch-based Video Generation" + ], + "image_footnote": [ + "\"A clear blue sky above the mountains, with a few clouds scattered across it. A deep blue lake nestled between the mountains. The mountains have patches of snow that still cling to their peaks.\"" + ], + "bbox": [ + 379, + 243, + 879, + 419 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/23cd0f2c4a391929d8ccaa4b1c3959237188627b1d775931462fec892c55b843.jpg", + "image_caption": [ + "(b) Sketch-based Video Editing", + "Figure 1. Our method enables high-quality video generation (a) and editing (b) based on sketch and text inputs. (a) Top: With the same text prompt, different keyframe sketches lead to results with similar semantics but diverse sketch-faithful geometry. (a) Bottom: With the same sketches, varied text prompts yield diverse appearances. (b) Users can also edit real videos by drawing on keyframe sketches, with edits automatically propagated even when edited objects in original videos have translation and rotation." + ], + "image_footnote": [], + "bbox": [ + 127, + 426, + 880, + 556 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 633, + 326, + 650 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Video generation and editing conditioned on text prompts or images have undergone significant advancements. However, challenges remain in accurately controlling global layout and geometry details solely by texts, and supporting motion control and local modification through images. In this paper, we aim to achieve sketch-based spatial and motion control for video generation and support fine-grained editing of real or synthetic videos. Based on the DiT video generation model, we propose a memory-efficient control structure with sketch control blocks that predict residual features of skipped DiT blocks. Sketches are drawn on one or two keyframes (at arbitrary time points) for easy interaction. To propagate such temporally sparse sketch conditions across all frames, we propose an inter-frame at", + "bbox": [ + 88, + 666, + 482, + 878 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tention mechanism to analyze the relationship between the keyframes and each video frame. For sketch-based video editing, we design an additional video insertion module that maintains consistency between the newly edited content and the original video's spatial feature and dynamic motion. During inference, we use latent fusion for the accurate preservation of unedited regions. Extensive experiments demonstrate that our SketchVideo achieves superior performance in controllable video generation and editing.", + "bbox": [ + 511, + 637, + 906, + 773 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 799, + 643, + 814 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Diffusion-based text-to-image [12, 51, 53] and text-to-video [27, 35, 43, 73, 84] models advance significantly due to improvements in datasets [4, 42, 55, 61] and denoising network architectures [12, 51]. While text prompts effectively describe high-level semantics, they lack control of scene", + "bbox": [ + 511, + 824, + 906, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding Author: Lin Gao (gaolin@ict.ac.cn)", + "bbox": [ + 114, + 886, + 393, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "23379", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Layouts and geometric details. To address this, existing video generation methods [2, 25, 73] utilize images as additional conditions but raise the questions of how to generate input images and achieve detailed editing. Sketching serves as a user-friendly interaction tool to capture spatial content and shape details accurately. One or two sketches are already sufficient to convey desired scene structures and motion information for short video clips (around 6 seconds), which are the target of our and most existing video generation methods. However, using such sparse keyframe sketches presents several challenges, including reasonably completing the missing frames, improving memory efficiency, and addressing the limited size of video datasets.", + "bbox": [ + 89, + 90, + 480, + 287 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "A naive solution is to translate the input keyframe sketches into images and then utilize interpolation methods [16, 37, 70] for video generation. However, it is nontrivial to ensure consistency during keyframe sketch-to-image generation, which significantly affects the video quality. This approach also struggles to generate extrapolation frames when applying conditions in intermediate frames rather than beginning and ending time points. Another possible approach is to utilize white placeholders to fill missing condition frames and directly apply ControlNet [77] into video models, similar to SparseCtrl [22]. However, this requires the same network to process both sketches and white placeholders simultaneously, while the pretrained blocks handle tasks far from this sparse propagation. Additionally, for DiT-based video frameworks [27, 84], the traditional strategy [11] that copies half of the pretrained model as a condition network easily causes the out-of-memory issue.", + "bbox": [ + 91, + 287, + 482, + 559 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address these issues, we propose a novel sketch condition network specifically designed for the DiT-based video generation architecture (CogVideoX [73] in our work). Following ControlNet [11, 77], we employ a trainable copy of CogVideoX's DiT block to process only the sketch inputs and generate control features. No white placeholder is processed to align with the pretrained weights and reduce learning complexity. To propagate these keyframe features, we design an inter-frame attention mechanism that captures the relationship between the control keyframes and all video frames. Our approach computes query and key features from noisy latent while extracting value features from sketch conditions. This design leverages frame-to-frame similarity for control propagation. The above components consist of a single sketch control block. Instead of copying a half number of the pretrained blocks to construct the control network [11], we use only 5 sketch control blocks out of the 30 DiT blocks available in CogVideoX-2b. We design a novel uniformly distributed skip structure to add the control signals to different levels of features in discrete blocks (0, 6, 12, 18, and 24), achieving effective spatial control while improving memory efficiency. During", + "bbox": [ + 89, + 559, + 482, + 891 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "training, an external image dataset is incorporated to solve the challenge of limited video data.", + "bbox": [ + 511, + 90, + 903, + 119 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Beyond generation, interactive editing of real or synthetic videos further enhances creative flexibility. Existing methods [34, 40, 45, 71] achieve interesting text-based editing or effectively propagate single image editing into videos. Despite their effectiveness in appearance modification, they struggle with shape manipulation and object insertion, as they preserve the original temporal motion. Such information is missing for newly introduced content. Moreover, precisely identifying and preserving unedited regions for localized editing remains a challenge.", + "bbox": [ + 511, + 121, + 903, + 271 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We propose a sketch-based editing method for detailed local modification. Rather than editing a single image and propagating changes, we directly construct an editing network based on our sketch control network. To analyze the relationship between edited regions and the original video, we incorporate a video insertion module that takes the original video with masked edited regions as inputs. The modified sketch control blocks generate residual features that capture temporally and spatially coherent contents in the edited areas. To accurately preserve unedited regions and achieve seamless fusion, the newly edited regions are blended with the original video in the latent space.", + "bbox": [ + 511, + 272, + 903, + 452 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Extensive experiments demonstrate that our method outperforms existing approaches in video generation and editing. Our contributions are summarized as follows: 1) We propose SketchVideo, a novel sketch-based video generation and editing framework that enables detailed geometry control and manipulation using keyframe sketches, as shown in Fig. 1. 2) We design a sketch condition network that predicts skipped control features for the DiT framework, with an inter-frame attention mechanism to propagate one or two sketch conditions across the video. 3) We propose a video insertion module that analyzes the relationship between drawn sketches and original videos, utilizing a latent fusion strategy to preserve unedited regions accurately.", + "bbox": [ + 511, + 454, + 903, + 650 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 664, + 653, + 679 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Sketch-based Image Generation. GAN-based methods [21] have achieved great success in category-restricted sketch-to-image translation [13, 14, 50, 65, 75, 85]. Recently, diffusion-based text-to-image models [51, 53] handle general categories with conditional models like ControlNet [77], T2I-Adapter [41] and further advances [30, 46, 48, 76, 83]. Beyond U-Net, the DiT backbone [12] enables image generation, with pixART- $\\delta$ [11] utilizing the first half of the pretrained model's blocks as a trainable network to predict corresponding control residual features. Video generation, however, introduces additional challenges. For ease of interaction, we expect sketches specified only for a sparse set of keyframes, making it difficult to generate frames without sketch inputs. Additionally, video genera", + "bbox": [ + 511, + 688, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "23380", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tion costs significantly higher memory resources, making methods [11, 77] that replicate half of the base model as a sketch encoder easily out of memory.", + "bbox": [ + 89, + 90, + 480, + 136 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Diffusion-based Video Generation. VDM [26] pioneered diffusion-based video generation with a 3D U-Net denoising network. To improve quality, subsequent works [2, 3, 9, 10, 23, 25, 33, 64] integrated temporal modules into text-to-image models [51] to enable text- and image-conditioned video synthesis. Considering the efficiency, the DiT architecture is further used in Sora [43] and open-source projects [35, 60, 84]. Despite these advancements, subtle flickering artifacts remain in the results. CogVideoX [27, 73] further proposes a 3D full attention that merges the spatial and temporal attention, facilitating the generation of long-duration and high-resolution videos.", + "bbox": [ + 88, + 136, + 482, + 316 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Building on existing video generation models, various conditions have been introduced to control the generation, such as camera movement [24, 66, 72], subject identity [31, 68], key-point trajectory [36, 56, 81] and example motions [67, 79]. However, they often lack control over spatial layouts and geometric details. Some methods [15, 18, 80] address this by extending image-based ControlNet [77] to videos but require all-frame conditions that are tedious for sketch interaction. SparseCtrl [22] tackles this by using white images for completions. However, due to the restriction of its base model [23] and a simple network design, its results suffer from temporal flickering. Similar ideas have been applied to cartoon interpolation [70] and colorization [28] with line art as input, but their outputs are limited to cartoon style. Our method utilizes sparse inputs, including hand-drawn sketches on one or two keyframes, to generate temporally stable and realistic videos. Additionally, our method further supports the sketch-based detailed editing of existing videos.", + "bbox": [ + 91, + 318, + 482, + 604 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Deep Learning-based Video Editing. Pioneer works achieve video editing by style transfer [29, 52], GAN inversion [39, 62], and layered representations [32]. The advent of diffusion models further provides new editing paradigms. One category of such methods leverages image generation models to achieve compelling editing results, using temporal consistency techniques such as layered representations [6, 44], cross-frame attention [5, 17, 19, 57] and pixel warping [71]. A second category of methods employ video generation models. These works propagate the edits applied on the first frame into the other frames [34, 45], or utilize an inpainting strategy to achieve text-based editing [82] and motion modification [40]. To achieve text-based large-scale shape modification while maintaining motion features, a space-time feature loss [74] is designed to guide the inference process. Our method moves beyond traditional text- or image-based editing approaches, allowing users to draw one or two keyframe sketches at arbitrary time points for interactive video editing. Additionally, our method effectively", + "bbox": [ + 91, + 604, + 482, + 891 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "handles sketch-based shape manipulation and dynamic object insertion, which are challenging for previous works.", + "bbox": [ + 511, + 90, + 903, + 121 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Controllable Attention Mechanism. Attention across frames is initially used to ensure temporal consistency inAnimateDiff [23] and subsequent video generation models. Subsequent works, such as VideoBooth [31] and Still-Moving [8], utilize it to learn identity-aware features for customized video generation. For video editing, existing works [17, 20, 57] utilize cross-frame attention to capture the temporal motion of input videos, enabling effective propagation of editing operations. Our method applies this idea to pixel-aligned sketch-based video generation and propagates the spatial geometric conditions instead of identity customization. We use a new feature derivation strategy for enhanced spatial control, differentiating our approach from traditional cross-attention mechanisms.", + "bbox": [ + 511, + 121, + 903, + 330 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 511, + 345, + 648, + 364 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This section introduces our sketch-based video generation and editing framework. In Sec.3.1, we provide an overview of CogVideoX-2b [73], a pretrained text-to-video generation method, which we will use for sketch-based video generation in our work. In Sec.3.2, we describe our sketch condition network specifically designed for the DiT architecture, which contains sketch control blocks to predict residual features. Within each control block, an interframe attention mechanism is designed to propagate the input sketches. In Sec.3.3, we detail the editing framework, which incorporates a video insertion module and latent fusion to preserve the features of the original video.", + "bbox": [ + 511, + 371, + 903, + 551 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminary", + "text_level": 1, + "bbox": [ + 511, + 561, + 643, + 579 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "CogVideoX [73] is a text-to-video generation model that employs a 3D causal VAE and builds diffusion within the latent space. In CogVideoX-2b, the VAE model performs $8 \\times 8$ spatial and $4 \\times$ temporal downsampling, followed by a DiT architecture with 30 blocks for video generation. Within each block, a 3D full attention processes the concatenated text embeddings and patchified video latents, followed by a feed-forward layer to output the features. The 3D full attention merges the commonly used separate spatial and temporal attention [2, 23, 84] to improve the temporal coherence. The training objective of diffusion is:", + "bbox": [ + 511, + 584, + 903, + 750 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nL (\\theta) := \\mathrm {E} _ {t, x _ {0} ^ {1: N}, y, \\epsilon} \\left\\| \\epsilon - \\epsilon_ {\\theta} \\left(\\sqrt {\\bar {\\alpha} _ {t}} x _ {0} ^ {1: N} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon , t, y\\right) \\right\\| ^ {2}, \\quad (1) \\text {w h e r e} t \\text {i s s a m p l e d b e t w e e n 1 a n d T (d e n o s i n g s t e p s)}, \\epsilon \\text {i s t h e r a n d o n n o i s e ,} y \\text {i s t e x t p r o m p t s , a n d} x _ {0} ^ {1: N} \\text {i s t h e v i d e o d a w i t h} N \\text {f r a m e s .} \\text {T h e v - p r e d i c t i o n [ 5 4 ] a n d z e r o S N R [ 3 8 ] a r e u l i z a t e d f o r t h e d i f f u s i o n s e t t i n g .}\n$$\n", + "text_format": "latex", + "bbox": [ + 511, + 758, + 903, + 839 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Sketch-based Video Generation", + "text_level": 1, + "bbox": [ + 511, + 848, + 792, + 863 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given text prompts and one or two keyframe sketches with corresponding time points $t_1, t_2$ , our method generates", + "bbox": [ + 511, + 869, + 906, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "23381", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/fa82a91a15357dd7e658f31b9edbfd2d35d9cecf979b82c59a23a8b8a40a11ed.jpg", + "image_caption": [ + "(a) Sketch Condition Network" + ], + "image_footnote": [], + "bbox": [ + 89, + 88, + 302, + 340 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/ae02777334cbdd69680ec09f99ecb89bab661376ac16380c8eb6caef18631e4e.jpg", + "image_caption": [ + "(b) Control Block for Generation", + "(c) Control Block for Editing", + "Figure 2. Our framework for sketch-based video generation and editing. (a) Our sketch condition network for the DiT-based video generation architecture has a skip structure and five sketch control blocks that predict residual features. (b) For generation, features are extracted from temporally sparse input sketches and propagated through inter-frame attention. The input sketches are provided for one or two keyframes (the second sketch is shown as a dotted line). In the top left corner of (b), the prompt and timestep inputs are shown. (c) For editing, the same sketch control block (b) is utilized, with an additional video insertion module and video masks $M^{1:N}$ to analyze the relationship between edited and unedited regions. The 3D causal VAE is omitted to save space." + ], + "image_footnote": [], + "bbox": [ + 305, + 87, + 908, + 342 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "video clips that respect the input text prompts and sketches. As shown in Fig. 2, we design a sketch condition network for effective control. The input sketches are encoded into the latent space by the pretrained VAE, followed by patchifying and time-aware positional embedding to generate sketch latents $s_0^{t_1,t_2}$ .", + "bbox": [ + 88, + 449, + 482, + 539 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Skip Residual Structure. In sketch-based image generation, methods like ControlNet [77] and pixART- $\\delta$ [11] copy the half of the base model as their sketch encoder to fully utilize the pretrained text-to-image models. However, applying this to video generation, as done in SparseCtrl [22], is memory inefficient because it requires adding half of the base model's parameters. Unlike the U-Net architecture, the DiT network does not have an explicit encoder and decoder. Therefore, the assumption in pixART $\\delta$ [11] that the first half of blocks serve as the encoder can be improved.", + "bbox": [ + 88, + 539, + 482, + 705 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Instead of borrowing local consecutive blocks of the base model as the encoder and predicting their residual features, we recognize that blocks at different depths process distinct feature levels, which should be considered for sketch control. As illustrated in Fig. 2 (a), we propose a novel skip residual structure that reduces the number of blocks while enabling effective sketch control and high-quality generation. Our sketch condition network contains 5 sketch control blocks, uniformly distributed across the pretrained generation network to predict residual features for blocks 0, 6, 12, 18, and 24 of the original video generation model. This structure efficiently integrates sketch control informa", + "bbox": [ + 88, + 705, + 482, + 888 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "tion into multiple feature levels, enhancing the analysis of input conditions and original semantic features.", + "bbox": [ + 511, + 449, + 906, + 478 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Sketch Control Block. In the $i$ -th sketch control block, the input consists of hidden video features $h_i^{1:N}$ and sketch features $s_{i-1}^{t_1,t_2}$ , and the output is the residual features $\\overline{h}_i^{1:N}$ and updated sketch features $s_i^{t_1,t_2}$ . As shown in Fig. 2 (b), the sketch feature $s_{i-1}^{t_1,t_2}$ is processed:", + "bbox": [ + 511, + 479, + 905, + 561 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ns _ {i} ^ {t _ {1}, t _ {2}} = \\operatorname {F e e d F o r w a r d} \\left(s _ {i - 1} ^ {t _ {1}, t _ {2}}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 604, + 569, + 903, + 590 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where the output sketch feature $s_i^{t_1,t_2}$ is used for sketch control propagation and as the input to the next sketch control block.", + "bbox": [ + 511, + 598, + 905, + 641 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To propagate the sketch inputs, a direct approach [22] replaces missing sketches with white images and employs a trainable copy of the pretrained DiT block to predict residual features. However, as shown in Fig. 5, this leads to fuzzy details in challenging cases, as the network randomly processes both sketches and white images. This mixed input is far from the pretrained model's task. To address the above issue, we employ a trainable copy of pretrained DiT block to process only the sketch inputs (no white placeholder), aligning with the pretrained weights and reducing learning difficulty. The resulting keyframe sketch features, denoted as $c_{i}^{t_{1},t_{2}}$ , are propagated to all frames through an inter-frame attention approach.", + "bbox": [ + 509, + 643, + 906, + 839 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Inter-frame Attention. We utilize the input hidden features of all frames to calculate $Q$ and the hidden features corresponding to the control frames to calculate $K$ . During attention computation, this approach captures the internal", + "bbox": [ + 511, + 839, + 906, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "23382", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/f48de3b7c0011fd39a4dc48df3b075b3a005fb283187c656cf5ffbe6bea17a4f.jpg", + "image_caption": [], + "image_footnote": [ + "A small, fluffy white rabbit hops across a lush meadow at sunset, pausing to listen to rustling leaves before continuing its joyful dance." + ], + "bbox": [ + 261, + 89, + 354, + 133 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/80a7d9e39c6a1472df2068c1865e29c9435cabd3e62087239525a0aec3a480de.jpg", + "image_caption": [], + "image_footnote": [ + "A panoramic video captures a majestic city with towering skyscrapers, quaint red-tiled homes, and a tranquil lake, all under a sky of drifting white clouds." + ], + "bbox": [ + 261, + 138, + 354, + 186 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/0d5430ee08d9fa5fcbb547558c2d691da520317b3ca6216470bfe2ec90202af3.jpg", + "image_caption": [], + "image_footnote": [ + "A video captures close-ups of a cat's face with green eyes and a brown-black spotted coat, showing three expressions: alert, relaxed, and distant." + ], + "bbox": [ + 261, + 188, + 354, + 236 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/a53e56b37cf76b85a6814f637c27a8f2f6672708bc20563a46679693f67b6b8f.jpg", + "image_caption": [], + "image_footnote": [ + "A video captures the explosive disintegration of a light blue planet, fragments and fiery debris catapulted into starry space, creating a stark contrast between celestial serenity and violent upheaval." + ], + "bbox": [ + 261, + 238, + 354, + 284 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/edf3a8e2b2b30f8e21352ee6db59bbdfa2bb1170cae42fc2d0617b16f0041d31.jpg", + "image_caption": [ + "Figure 3. The sketch-based video generation results. Left: The input text prompts and sketches. Right: The video generation results. It can be seen that the generated results show high quality and good faithfulness with the input sketches. Our method can handle one/two keyframe sketch(es) at arbitrary user-specified time points (the frames corresponding to the input time points are highlighted by orange)." + ], + "image_footnote": [ + "A time-lapse aerial video shows a large, ornate domed building surrounded by lush greenery, transitioning from daylight to sunset to night, illuminated by artificial lights." + ], + "bbox": [ + 261, + 286, + 354, + 335 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/bd0cfa4dca4f032ce9c1858075f506190674d0a4345fa7931a229cc75a03238b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 362, + 90, + 901, + 136 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/306a50698fc2ec85405c297478e8da728c48f4dfd9a69393313dd09e5c8d07dc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 364, + 138, + 901, + 185 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/75e25e63b5a2b7892e7807bad4698a960b5e39c4e15088fcae499132ed51f4e5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 364, + 188, + 901, + 234 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/d83b67eaf759f5c101ef93700650fbcd77023e58217ed338587ca62a202240ff.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 364, + 237, + 901, + 284 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/fdc92561323bce7edbf55272dc181f97224a1a62288fc6094d697af1df209447.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 364, + 285, + 901, + 333 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "relationship between all frames and control keyframes, allowing propagation of $V$ (derived from the keyframe sketch features). Our inter-frame attention is distinct from typical cross-frame attention, which uses both $K$ and $V$ from control conditions; instead, we leverage the inter-frame similarity within the input noisy hidden video features and progressively insert the sketches' spatial and temporal information. The output $\\widetilde{c}_i^{1:N}$ is computed as: $\\mathrm{Attention}(Q,K,V) = \\mathrm{Softmax}(\\frac{QK^T}{\\sqrt{d}})\\cdot V$ , with", + "bbox": [ + 89, + 386, + 483, + 529 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nQ = W _ {q} \\cdot h _ {i} ^ {1: N}, K = W _ {k} \\cdot h _ {i} ^ {t _ {1}, t _ {2}}, V = W _ {v} \\cdot c _ {i} ^ {t _ {1}, t _ {2}}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 107, + 537, + 480, + 556 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $W_{q}, W_{k}, W_{v}$ are trainable linear projection weights. The output of the inter-frame attention is fed into a feedforward layer to generate final residual features $\\overline{h}_{i}^{1:N}$ .", + "bbox": [ + 89, + 564, + 482, + 613 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training Strategy. To train the sketch condition network, we employ a hybrid training strategy in two stages. In the first stage, to accelerate convergence and address the issue of limited video data, the network is trained both for image generation at arbitrary time points and video generation with one or two keyframe sketches (randomly selected from corresponding sketch videos). In the second stage, video data alone is used to improve the temporal coherence.", + "bbox": [ + 89, + 614, + 483, + 734 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Sketch-based Video Editing", + "text_level": 1, + "bbox": [ + 89, + 742, + 339, + 758 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For editing, given real or synthetic videos, users select one or two keyframes at arbitrary time points and modify the extracted sketches, with additional inputs of text prompts and masks $M^{1:N}$ that label regions to be edited for all the frames. Our method then generates realistic, local editing results. The input video is multiplied by the inverted masks to remove information from the edited regions and is subsequently encoded to generate a masked video latent representation $v_{0}^{1:N}$ .", + "bbox": [ + 89, + 763, + 483, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Video Insertion Module. For sketch-based editing, newly generated contents within the mask regions should be coherent with the original spatial and temporal features in the unedited regions. Thus, we design a video insertion module that analyzes the relationship between the input sketches and the original video. The video insertion module takes $v_{i-1}^{1:N}$ (input video latent or generated by a previous control block) as input and predicts the updated video features $v_{i}^{1:N}$ , similar to the sketch generation process. Since video features are not temporally sparse, we use a trainable copy of CogVideoX-2b's DiT block to directly generate video insertion features $\\widetilde{v}_{i}^{1:N}$ . The sketch branch output $\\widetilde{c}_{i}^{1:N}$ and video branch output $\\widetilde{v}_{i}^{1:N}$ are multiplied by their respective masks and concatenated:", + "bbox": [ + 511, + 386, + 906, + 598 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {C o n c a t} \\left(\\tilde {c} _ {i} ^ {1: N} * M ^ {1: N}, \\tilde {v} _ {i} ^ {1: N} * \\overline {{M}} ^ {1: N}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 581, + 609, + 906, + 631 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "which serves as inputs to the feed-forward layers, producing final residual features incorporating the original video and sketch control information. This design ensures seamless integration of new contents with the original videos, effectively propagating edits across frames for dynamic motion.", + "bbox": [ + 511, + 643, + 905, + 719 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training Strategy. Directly training the video editing network would lead to low fidelity with the input sketches, possibly because of the challenging interaction between the input sketches and videos. So we finetune it from the pretrained sketch condition network for generation and add the new video insertion module. The pretrained model already has good sketch fidelity and only requires learning video information. The network is trained in a self-supervised inpainting manner, with randomly generated masks to imitate real-world editing.", + "bbox": [ + 511, + 719, + 905, + 869 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Inference Latent Fusion. Although the original videos are encoded in the condition network, as shown in Fig. 8,", + "bbox": [ + 511, + 869, + 905, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "23383", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/0421acc3ba4c525b9761a4b546353e4a9d37813e9bd607dd4d7d1a41fd37c66a.jpg", + "image_caption": [ + "Input Video" + ], + "image_footnote": [], + "bbox": [ + 197, + 89, + 496, + 130 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1a7f168e94e61a55406f323676cfdd8e2cf771390811c7f1cdd64fc6fd393795.jpg", + "image_caption": [ + "A snow-dusted volcano looms, steam plumes rising from its crest. Lush green flanks contrast with rocky slopes, while a rippling lake is in the foreground." + ], + "image_footnote": [], + "bbox": [ + 102, + 135, + 183, + 179 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/8ce966540daf86766216580d5ba284b20e25bb0fad3c004b7804e257f47943ba.jpg", + "image_caption": [ + "Edited Results" + ], + "image_footnote": [], + "bbox": [ + 197, + 136, + 498, + 176 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/470bab0b35195b010fc37c9fc5ab3b7299d8f092783bcaa62b2af3c13a6b17d3.jpg", + "image_caption": [ + "A mesmerizing ornamental fish glides through the ocean, its electric blue, fiery orange, and iridescent green scales shimmering as it navigates the depths with deli- cate, patterned fins.", + "Input Video" + ], + "image_footnote": [], + "bbox": [ + 197, + 186, + 496, + 224 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e2c14b5966d29ac8f6b6d044923893a56667fd6514f35305f7b6b25887a534ea.jpg", + "image_caption": [ + "Edited Results", + "Figure 4. The sketch-based video editing results. For each example, the text prompts and sketches are shown on the left. On the right, the input real videos are shown at the top, while the edited results with the control keyframe highlighted in orange are shown at the bottom. The editing region masks are manually provided by users, highlighted as orange boxes. Our method generates realistic local editing results." + ], + "image_footnote": [], + "bbox": [ + 197, + 234, + 496, + 273 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5e00984564c2725f2d35620805f8d6cb69401b77b5ec56ff1a38bb0dada818ad.jpg", + "image_caption": [ + "A man is speaking. He has thick hair and is wearing a black suit with a white shirt and a red tie underneath. The camera focuses on his face and upper body.", + "Input Video" + ], + "image_footnote": [], + "bbox": [ + 604, + 89, + 903, + 130 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/176b08af6cbb1a48bff42b9b488de994cd21fdabc132045cad378b012d039e45.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 513, + 137, + 591, + 176 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/fd661b398e226a3da1a66fb2eb9e034b3d0269647eb7003815ec0b72dba6da00.jpg", + "image_caption": [ + "Edited Results" + ], + "image_footnote": [], + "bbox": [ + 604, + 137, + 903, + 176 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/7b9da252f21b9adb832a7bb6dd275e242d7063fbbc133f54d0a3b359703c4ae4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 514, + 233, + 591, + 273 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5e32331a626647853f09154c9ce0f446d7d42dbb8f0ef957f5a1b0ec976279dc.jpg", + "image_caption": [ + "A green grassland and a small path extending into the distance. The grassland is covered with gravel.", + "Input Video" + ], + "image_footnote": [], + "bbox": [ + 604, + 186, + 903, + 224 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/cb85a421f7ea3062d94ff747c090afe2dfaf333e5b0a37c4d1d6972bbe2ecadb.jpg", + "image_caption": [ + "Edited Results" + ], + "image_footnote": [], + "bbox": [ + 604, + 233, + 903, + 273 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "fine details might be lost during editing. To address this, we propose a latent fusion approach at inference. Specifically, we apply DDIM inversion [58] to generate noisy latent codes of input videos across the inference steps. At Steps 25 and 49 (out of 50 total steps), the latent codes in the unedited regions are replaced with these inversion latent codes, ensuring better preservation of the original video's details and improving the coherence of the edited regions.", + "bbox": [ + 89, + 335, + 483, + 455 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Evaluation", + "text_level": 1, + "bbox": [ + 89, + 473, + 207, + 488 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Implementation Details", + "text_level": 1, + "bbox": [ + 89, + 497, + 307, + 513 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We implement SketchVideo based on CogVideoX-2b [73], trained on a subset of OpenVid [42] and LAION [55] datasets, with paired sketches from [7]. Training uses 8 NVIDIA H800 GPUs with a batch size of 8 and gradient accumulation of 4. For generation, one or two keyframe sketches are randomly sampled from the video, with 10,000 steps for each training stage. For editing, randomly drawn masks are applied and trained for 20,000 steps. For ease of reading, the input text prompts are simplified in figures. Additional implementation details and full input texts are available in the supplementary material.", + "bbox": [ + 89, + 518, + 483, + 686 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Results", + "text_level": 1, + "bbox": [ + 89, + 696, + 183, + 712 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our method generates high-quality videos from one or two keyframe sketches and text prompts. As shown in Fig. 3, our method can accurately control the object shape and scene layout, such as the rabbit's pose (1st row) and the position of lakes and buildings (2nd row). Text-only inputs cannot achieve such detailed geometry control. Our method also achieves interesting dynamic motion interpolation and extrapolation, generating smooth and realistic transitions, as seen in the cat's head shaking (3rd row) and the building's movement (last row). This allows control over both spatial layout and dynamic motion.", + "bbox": [ + 89, + 719, + 483, + 883 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our method also supports sketch-based video editing.", + "bbox": [ + 109, + 885, + 480, + 900 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/3782d4d2d3d1c44994c1f8170531c4dd4d9846734824fd2a0b746c9256fbec22.jpg", + "image_caption": [ + "Input Text A time-lapse video shows a large white and black cruise ship with yellow and gold accents docking at a pier on an overcast day, reflecting in calm waters.", + "Figure 5. The comparison results of sketch-based video generation. Text prompts are shown on the top. On the left, we show the input sketches and sketch-based image generation results by ControlNet [77]. On the right, we show the results of the compared approaches, including AMT [37], SparseCtrl [22], CtrlCogVideo [77], and ours. Our method produces better results, especially for the intermediate frames." + ], + "image_footnote": [], + "bbox": [ + 516, + 352, + 588, + 523 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/87778312df5b6ae3d3dea199f26c39a863dbaea88039bce14c1705e849deed6f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 589, + 352, + 666, + 523 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/206cfdfc783968466fdd0c5652a9f0c3fbfe9a29ee83cbcd38ca1ea9dfac7a8d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 666, + 352, + 746, + 523 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/9b7c891ab8e925b9925e2d9120e60aadc1ec6cd62dfdbbd1ef97e246d53f1350.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 746, + 352, + 825, + 523 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/bc01f31a7d84aec396d8364320f31ede572a1673d2d3ec37948f5035649a5a12.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 825, + 352, + 903, + 523 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Users specify bounding boxes and draw sketches in those regions, and our method generates photorealistic, seamlessly integrated content. As shown in Fig. 4, the new contents blend well with unedited regions and have dynamic motions like hat/hair rotation with original objects (1st row) or interesting fish swimming (2nd row). Our method can handle diverse editing cases, including object insertion, component replacement, and object removal. The unedited regions are well-preserved thanks to our latent fusion approach.", + "bbox": [ + 511, + 636, + 906, + 787 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Comparison", + "text_level": 1, + "bbox": [ + 511, + 801, + 643, + 819 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For sketch-based generation, we compare our methods with three methods given the same keyframe sketches text prompts as inputs. For SparseCtrl [22], we use the official pretrained model and extract sketches by HED [69]. It is trained on videos in the WebVid-10M dataset [1] instead", + "bbox": [ + 511, + 825, + 905, + 898 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "23384", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/af9226a58f43b81c4b90e9c8bd415839b91410a0605b507a8cc783a75e58bd28.jpg", + "image_caption": [ + "Figure 6. The comparison results of sketch-based video editing. On the left, we show the drawn editing sketches (for the frames highlighted in orange), text prompts, and sketch-based image generation results by ControlNet [77]. On the right, we show the original videos and editing results by the compared methods, including InsV2V [17], AnyV2V [34], and ours. Our method generates the most realistic results and preserves unedited regions well." + ], + "image_footnote": [], + "bbox": [ + 91, + 88, + 171, + 255 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/23d15a9e786666483470ae5e5306e23a4c0b373bf2a816544c0f4f4d20ddf362.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 171, + 88, + 326, + 255 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1fd7e4aecef8cb49d95e0ffdbc42925b015f079a150042e10474d9fab32d592f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 326, + 88, + 401, + 255 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5a4b670008cf4036660feb229fceb7e04dd3f85a134b5af0649ab6290743d1de.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 401, + 88, + 480, + 255 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "of the OpenVid-1M dataset [42] used in our method. We extend SparseCtrl to CogVideoX-2b [73], using pixART $\\delta$ [11] as the sketch condition encoder (with 5 DiT blocks same as our method) and white images to complete the missing condition frames. We also compare with an interpolation baseline, which uses ControlNet to translate two-frame sketches into images and then interpolates them with AMT [37]. For the sketch-base editing task, we compare ours with a text-based video editing method InsV2V [17] and a first-frame editing method AnyV2V [34]. We compare with additional methods [16, 20, 45, 70], as shown in supplemental material.", + "bbox": [ + 88, + 371, + 482, + 551 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As shown in Fig. 5, ControlNet [77] translates sketches into realistic images, which, however, lack temporal consistency and vary with shading and content. The interpolation results of AMT [37] exhibit fuzzy details and artifacts. SparseCtrl [22], based onAnimateDiff [23], exhibits temporal flickering, such as the suddenly appearing tree and distorted tower top (see the orange boxes in the 2nd row). Extending SparseCtrl to CogVideoX-2b [73] (Ctrl-CogVideo) still generates fuzzy details in intermediate frames, possibly due to the CogVideoX-2b's pretrained self-attention being designed for dense inputs instead of sparse sketches. Our method generates realistic videos with clear details and good temporal coherence, with even small details like the electric wires in the top-right corner accurately propagated.", + "bbox": [ + 89, + 553, + 482, + 779 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For sketch-based editing (Fig. 6), InsV2V [17] generates interesting results with birds but lacks control over shape and geometry through text prompts alone. For image-based video editing method, we utilize ControlNet to edit the first frame and then propagate editing into the video by AnyV2V [34]. However, since the motion is borrowed from the original video, AnyV2V struggles with new content, leading to fuzzy details and distortion in the bird, as well", + "bbox": [ + 89, + 780, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/203fe8366914e388b86f2d87acfff27a736359a57d9a89997393b9ab21656bac.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsLPIPS ↓CLIP ↑FidelityConsistencyRealism
AMT29.1796.123.133.513.57
SparseCtrl44.8596.482.792.942.83
Ctrl-CogVideo32.2398.042.862.472.50
Ours27.5698.311.211.081.11
", + "bbox": [ + 514, + 88, + 905, + 155 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/9251cf751fe037fdfd36ef4b07a5987ddc97282279d2532c4d74b359fdea6aaf.jpg", + "table_caption": [ + "Table 1. The quantitative results of sketch-based video generation comparison. The LPIPS and CLIP numbers are scaled up $100 \\times$ with each cell colored to indicate the best." + ], + "table_footnote": [], + "table_body": "
MethodsLPIPS ↓CLIP ↑PSNR↑FidelityPreservationRealism
InsV2V13.6195.3916.842.582.262.61
AnyV2V11.9293.4713.682.352.692.34
Ours9.7498.3436.481.071.051.04
", + "bbox": [ + 521, + 200, + 898, + 256 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. The quantitative results of sketch-based video editing comparison. The LPIPS and CLIP numbers are scaled up $100 \\times$ with each cell colored to indicate the best.", + "bbox": [ + 511, + 260, + 903, + 303 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "as changes in unedited regions due to ControlNet's inability to retain the original features. In contrast, our method produces more realistic results with faithful representations of the sketch. More generation and editing results are available in the supplemental material and video.", + "bbox": [ + 511, + 310, + 905, + 385 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We follow SparseCtrl [22] and use the LPIPS [78] metric to measure sketch faithfulness between the input sketches and those extracted from the corresponding frames of the generated videos. We use the CLIP [49] similarity to assess temporal coherence. For sketch-based generation, we test 200 random examples from OpenVid [42], using the first and last frame sketches and the corresponding text prompts in the dataset. As shown in Table 1, our method achieves the lowest LPIPS and highest CLIP scores, indicating its superior performance. For sketch-based editing, we use additional an MSE metric to measure unedited region preservation. We utilize 10 examples with hand-drawn sketches as input. In Table 2, our method outperforms existing methods across all metrics, demonstrating its superiority.", + "bbox": [ + 511, + 386, + 905, + 597 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7e39767a6c4d7a8dec04171bf41aca2b61f99a186c996aa3d157ad11523cb0b6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 516, + 597, + 596, + 637 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ddfec4bcea406a7bacf28349df355d86ba490a12f61030bc4fb88192b7f06abf.jpg", + "image_caption": [ + "Extreme close-up of chicken and green pepper kebabs grilling on a barbeque with flames. Shallow focus and light smoke. vivid colours", + "(a) w/o Inter-Frame (b) Sketch K, V", + "(c) w/o Skip", + "(d) w/o Image", + "(e) ours", + "Figure 7. The ablation study results of sketch-based video generation. Our method generates more realistic and sketch-faithful results than the baselines." + ], + "image_footnote": [], + "bbox": [ + 596, + 595, + 903, + 691 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Ablation Study", + "text_level": 1, + "bbox": [ + 511, + 757, + 666, + 773 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We conduct an ablation study to evaluate the effectiveness of each key component. Due to the restriction of computing resources, all the models are trained with a batch size of 2. For sketch-based generation, removing inter-frame attention and concatenating sketches with frame features at the temporal dimension results in strange control with low sketch faithfulness (Fig. 7 (a)). Similarly, if we replace the inter-frame attention with a typical cross-attention that uses", + "bbox": [ + 511, + 779, + 905, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "23385", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f04947e3d62613f01bff44ede2a40b78f7e46b846a28543899c6c623d7de83a8.jpg", + "image_caption": [ + "Figure 8. The ablation study results of sketch-based video editing. Our method generates realistic results faithful to sketches. We utilize heat maps (top right) to visualize the difference between the edited and original frames." + ], + "image_footnote": [], + "bbox": [ + 91, + 87, + 148, + 220 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b8ffa9982b0b3559cc3a227f6c2f90c215f40a9de658dccac0b5de3c5afbb0f8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 151, + 88, + 233, + 220 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7632073a2b335d29397a41a1f469fff5898d658f15285867f10464aad61c4d73.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 235, + 88, + 316, + 220 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4f01ba7dcd287125674cb5f174ce52e7bba9204b7dd7cafe21d45382d6222977.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 316, + 88, + 398, + 220 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a7688278ba32ddf5c598b0b845eafd2a11955a800a67bb6d0aa96deb1fad97b5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 400, + 88, + 480, + 220 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/21f2c6c9298a245b3c6648148699e615be685b6946c9f62d182fa545bc107983.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Metricw/o Inter-Frame Sketch K,Vw/o Skipw/o ImageOurs
LPIPS ↓36.3332.5931.9134.5830.79
CLIP ↑98.1098.1997.6098.2498.48
Metricw/o Video modulew/o Pretrainw/o Latent FusionOurs
LPIPS ↓9.3112.609.779.74
CLIP ↑97.9798.1298.4498.34
PSNR ↑33.6136.0531.6936.48
", + "bbox": [ + 91, + 280, + 486, + 390 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3. The qualitative results of ablation study for sketch-based generation and editing, with each cell colored to indicate the best and second best.", + "bbox": [ + 89, + 392, + 483, + 439 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "sketch features for both Key and Value (b), the sketch fidelity is also degraded due to the lack of effective control propagation. If we remove the skip structure and predict residual features for the first 5 blocks (c), the realism of the generated results is negatively influenced due to less control ability. Removing image data during training (d) causes significant geometry mismatch with input sketches. Compared to these alternatives, our method produces the most realistic and faithful results. As shown in Table 3, our method has the lowest LPIPS values (supporting the best sketch fidelity) while the CLIP metric has similar results (human eyes cannot recognize temporal coherence difference).", + "bbox": [ + 88, + 446, + 482, + 626 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In sketch-based video editing (Fig. 8), removing the video insertion module (a) and solely utilizing the sketch condition encoder results in good sketch fidelity but obvious inconsistency with unedited regions. Training the sketch condition encoder from scratch without generation pretraining (b) reduces the faithfulness to the input sketches in the edited regions, which is also validated by the high LPIPS in Table 3. We use heat maps to reveal differences between edited and original frames. Removing the latent fusion strategy changes unedited region details (c), such as mountains and waves, as shown in the heat maps. The low PSNR value in unedited regions also supports this, as shown in Table 3. Our method maintains strong sketch fidelity, consistency, and unedited region preservation.", + "bbox": [ + 89, + 628, + 482, + 839 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. User Study", + "text_level": 1, + "bbox": [ + 89, + 848, + 212, + 864 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We conducted a user study to validate that our method outperforms existing approaches. For video generation, we", + "bbox": [ + 89, + 869, + 482, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "randomly selected 10 examples from the test set in Sec. 4.3. In the questionnaire, participants were shown the results of different methods in random order. 20 participants ranked the methods results by three criteria: Sketch Fidelity, Temporal Consistency, and Video Realism. The rankings provided by the participants were used as scores. As shown in Table 1, our method outperforms existing methods in all criteria, demonstrating its superior performance. For sketch-based editing, we used the same 10 examples from the test set in Sec. 4.3. The same participants ranked the methods based on three criteria: Sketch Fidelity, Unedited Region Preservation, and Video Realism. As shown in Table 2, our method also outperforms existing approaches in editing.", + "bbox": [ + 511, + 90, + 906, + 287 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion and Discussion", + "text_level": 1, + "bbox": [ + 511, + 310, + 761, + 325 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We presented SketchVideo, a unified method for sketch-based video generation and editing. For generation, we proposed a sketch condition network that predicts residual features for skipped DiT blocks of the base models to save the memory and achieve effective control. An inter-frame attention is further proposed to propagate the keyframe sketches, achieving interesting motion interpolation or extrapolation. We also introduce a hybrid image and video training strategy. For editing, we incorporate a video insertion module to ensure the newly generated content is spatially and temporally coherent with the original video. During inference, a latent fusion approach preserves unedited regions accurately.", + "bbox": [ + 511, + 335, + 906, + 531 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitation and Future Work. While our method generates high-quality videos, its capabilities are still limited by the pretrained text-to-video model. Enhancing performance with more powerful pretrained models and generating long videos instead of video clips is a potential avenue. Additionally, similar to image generation, our method struggles with too complex scenarios, such as human hands and interaction between multiple objects. Incorporating 3D priors [59, 63, 86] like SMPL-X [47] helps address these issues in human scenarios. Moreover, while our method focuses on geometry control, exploring appearance customization through tools like color strokes presents an intriguing area for future research.", + "bbox": [ + 511, + 532, + 908, + 727 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Acknowledgments", + "text_level": 1, + "bbox": [ + 511, + 751, + 692, + 767 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work was sponsored by CCF-Kuaishou Large Model Explorer Fund, Beijing Municipal Science and Technology Commission (No. Z231100005923031), Innovation Funding of ICT, CAS (No. E461020) and National Natural Science Foundation of China (No. 62322210). The authors would like to acknowledge the Nanjing Institute of InforSuperBahn OneAiNexus for providing the training and evaluation platform.", + "bbox": [ + 511, + 773, + 906, + 888 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "23386", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In Int. Conf. Comput. Vis., pages 1708-1718, 2021. 6", + "[2] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, Varun Jampani, and Robin Rombach. Stable video diffusion: Scaling latent video diffusion models to large datasets. CoRR, abs/2311.15127, 2023. 2, 3", + "[3] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In IEEE Conf. Comput. Vis. Pattern Recog., pages 22563-22575, 2023. 3", + "[4] Minwoo Byeon, Beomhee Park, Haecheon Kim, Sungjun Lee, Woonhyuk Baek, and Saehoon Kim. Coyo-700m: Image-text pair dataset. https://github.com/kakaobrain/coyo-dataset, 2022.1", + "[5] Duygu Ceylan, Chun-Hao Paul Huang, and Niloy J. Mitra. Pix2video: Video editing using image diffusion. In Int. Conf. Comput. Vis., pages 23149-23160, 2023. 3", + "[6] Wenhao Chai, Xun Guo, Gaoang Wang, and Yan Lu. Stable-video: Text-driven consistency-aware diffusion video editing. In Int. Conf. Comput. Vis., pages 22983-22993, 2023. 3", + "[7] Caroline Chan, Frédo Durand, and Phillip Isola. Learning to generate line drawings that convey geometry and semantics. In IEEE Conf. Comput. Vis. Pattern Recog., pages 7915-7925, 2022. 6", + "[8] Hila Chefer, Shiran Zada, Roni Paiss, Ariel Ephrat, Omer Tov, Michael Rubinstein, Lior Wolf, Tali Dekel, Tomer Michaeli, and Inbar Mosseri. Still-moving: Customized video generation without customized video data. ACM Trans. Graph., 43(6):244:1-244:11, 2024. 3", + "[9] Haoxin Chen, Menghan Xia, Yingqing He, Yong Zhang, Xiaodong Cun, Shaoshu Yang, Jinbo Xing, Yaofang Liu, Qifeng Chen, Xintao Wang, Chao Weng, and Ying Shan. Videocrafter1: Open diffusion models for high-quality video generation. CoRR, abs/2310.19512, 2023. 3", + "[10] Haoxin Chen, Yong Zhang, Xiaodong Cun, Menghan Xia, Xintao Wang, Chao Weng, and Ying Shan. Videographer2: Overcoming data limitations for high-quality video diffusion models. In IEEE Conf. Comput. Vis. Pattern Recog., pages 7310-7320, 2024. 3", + "[11] Junsong Chen, Yue Wu, Simian Luo, Enze Xie, Sayak Paul, Ping Luo, Hang Zhao, and Zhenguo Li. Pixart-δ: Fast and controllable image generation with latent consistency models. CoRR, abs/2401.05252, 2024. 2, 3, 4, 7", + "[12] Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Zhongdao Wang, James T. Kwok, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart-α: Fast training of diffusion transformer for photorealistic text-to-image synthesis. In Int. Conf. Learn. Represent., 2024. 1, 2", + "[13] Shu-Yu Chen, Wanchao Su, Lin Gao, Shihong Xia, and" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Hongbo Fu. Deepfacedrawing: deep generation of face images from sketches. ACM Trans. Graph., 39(4):72, 2020. 2", + "[14] Shu-Yu Chen, Feng-Lin Liu, Yu-Kun Lai, Paul L. Rosin, Chunpeng Li, Hongbo Fu, and Lin Gao. Deepfaceediting: deep face generation and editing with disentangled geometry and appearance control. ACM Trans. Graph., 40(4):90:1-90:15, 2021. 2", + "[15] Weifeng Chen, Jie Wu, Pan Xie, Hefeng Wu, Jiashi Li, Xin Xia, Xuefeng Xiao, and Liang Lin. Control-a-video: Controllable text-to-video generation with diffusion models, 2023. 3", + "[16] Xinyuan Chen, Yaohui Wang, Lingjun Zhang, Shaobin Zhuang, Xin Ma, Jiashuo Yu, Yali Wang, Dahua Lin, Yu Qiao, and Ziwei Liu. Seine: Short-to-long video diffusion model for generative transition and prediction. In Int. Conf. Learn. Represent., 2023. 2, 7", + "[17] Jiaxin Cheng, Tianjun Xiao, and Tong He. Consistent video-to-video transfer using synthetic dataset. In Int. Conf. Learn. Represent., 2024. 3, 7", + "[18] Ernie Chu, Shuo-Yen Lin, and Jun-Cheng Chen. Video controlnet: Towards temporally consistent synthetic-to-real video translation using conditional image diffusion models. CoRR, abs/2305.19193, 2023. 3", + "[19] Nathaniel Cohen, Vladimir Kulikov, Matan Kleiner, Inbar Huberman-Spiegelglas, and Tomer Michaeli. Slicedit: Zero-shot video editing with text-to-image diffusion models using spatio-temporal slices. arXiv preprint arXiv:2405.12211, 2024.3", + "[20] Michal Geyer, Omer Bar-Tal, Shai Bagon, and Tali Dekel. Tokenflow: Consistent diffusion features for consistent video editing. In Int. Conf. Learn. Represent., 2024. 3, 7", + "[21] Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron C. Courville, and Yoshua Bengio. Generative adversarial networks. CoRR, abs/1406.2661, 2014. 2", + "[22] Yuwei Guo, Ceyuan Yang, Anyi Rao, Maneesh Agrawala, Dahua Lin, and Bo Dai. Sparsectrl: Adding sparse controls to text-to-video diffusion models. In Eur. Conf. Comput. Vis., pages 330-348, 2024. 2, 3, 4, 6, 7", + "[23] Yuwei Guo, Ceyuan Yang, Anyi Rao, Zhengyang Liang, Yaohui Wang, Yu Qiao, Maneesh Agrawala, Dahua Lin, and Bo Dai. Animatediff: Imagine your personalized text-to-image diffusion models without specific tuning. In Int. Conf. Learn. Represent., 2024. 3, 7", + "[24] Hao He, Yinghao Xu, Yuwei Guo, Gordon Wetzstein, Bo Dai, Hongsheng Li, and Ceyuan Yang. Cameractrl: Enabling camera control for text-to-video generation. CoRR, abs/2404.02101, 2024. 3", + "[25] Yingqing He, Tianyu Yang, Yong Zhang, Ying Shan, and Qifeng Chen. Latent video diffusion models for high-fidelity video generation with arbitrary lengths. CoRR, abs/2211.13221, 2022. 2, 3", + "[26] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. Adv. Neural Inform. Process. Syst., 35:8633-8646, 2022. 3", + "[27] Wenyi Hong, Ming Ding, Wendi Zheng, Xinghan Liu, and Jie Tang. Cogvideo: Large-scale pretraining for" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "23387", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "text-to-video generation via transformers. arXiv preprint arXiv:2205.15868, 2022. 1, 2, 3", + "[28] Zhitong Huang, Mohan Zhang, and Jing Liao. LVCD: reference-based linear art video colorization with diffusion models. CoRR, abs/2409.12960, 2024. 3", + "[29] Ondrej Jamriska, Sárka Sochorová, Ondrej Texler, Michal Lukác, Jakub Fiser, Jingwan Lu, Eli Shechtman, and Daniel Sykora. Stylizing video by example. ACM Trans. Graph., 38 (4):107:1-107:11, 2019. 3", + "[30] Rui Jiang, Guang-Cong Zheng, Teng Li, Tian-Rui Yang, Jing-Dong Wang, and Xi Li. A survey of multimodal controllable diffusion models. Journal of Computer Science and Technology, 39(3):509-541, 2024. 2", + "[31] Yuming Jiang, Tianxing Wu, Shuai Yang, Chenyang Si, Dahua Lin, Yu Qiao, Chen Change Loy, and Ziwei Liu. Videobooth: Diffusion-based video generation with image prompts. In IEEE Conf. Comput. Vis. Pattern Recog., pages 6689-6700, 2024. 3", + "[32] Yoni Kasten, Dolev Ofri, Oliver Wang, and Tali Dekel. Layered neural atlases for consistent video editing. ACM Trans. Graph., 40(6):210:1-210:12, 2021. 3", + "[33] Levon Khachatryan, Andranik Movsisyan, Vahram Tadevosyan, Roberto Henschel, Zhangyang Wang, Shant Navasardyan, and Humphrey Shi. Text2video-zero: Text-to-image diffusion models are zero-shot video generators. In Int. Conf. Comput. Vis., pages 15908-15918, 2023. 3", + "[34] Max Ku, Cong Wei, Weiming Ren, Harry Yang, and Wenhu Chen. Anyv2v: A tuning-free framework for any video-to-video editing tasks. arXiv preprint arXiv:2403.14468, 2024. 2, 3, 7", + "[35] PKU-Yuan Lab and Tuzhan AI etc. Open-sora-plan, 2024. 1, 3", + "[36] Yaowei Li, Xintao Wang, Zhaoyang Zhang, Zhouxia Wang, Ziyang Yuan, Liangbin Xie, Yuexian Zou, and Ying Shan. Image conductor: Precision control for interactive video synthesis. CoRR, abs/2406.15339, 2024. 3", + "[37] Zhen Li, Zuo-Liang Zhu, Ling-Hao Han, Qibin Hou, Chun-Le Guo, and Ming-Ming Cheng. Amt: All-pairs multi-field transforms for efficient frame interpolation. In IEEE Conf. Comput. Vis. Pattern Recog., 2023. 2, 6, 7", + "[38] Shanchuan Lin, Bingchen Liu, Jiashi Li, and Xiao Yang. Common diffusion noise schedules and sample steps are flawed. In IEEE/CVF Winter Conference on Applications of Computer Vision, WACV, pages 5392-5399, 2024. 3", + "[39] Feng-Lin Liu, Shu-Yu Chen, Yu-Kun Lai, Chunpeng Li, Yue-Ren Jiang, Hongbo Fu, and Lin Gao. Deepfacevideoediting: sketch-based deep editing of face videos. ACM Trans. Graph., 41(4):167:1-167:16, 2022. 3", + "[40] Chong Mou, Mingdeng Cao, Xintao Wang, Zhaoyang Zhang, Ying Shan, and Jian Zhang. Revideo: Remake a video with motion and content control. CoRR, abs/2405.13865, 2024. 2, 3", + "[41] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, and Ying Shan. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. In AAAI, pages 4296-4304, 2024. 2" + ], + "bbox": [ + 91, + 92, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[42] Kepan Nan, Rui Xie, Penghao Zhou, Tiehan Fan, Zhenheng Yang, Zhijie Chen, Xiang Li, Jian Yang, and Ying Tai. Openvid-1m: A large-scale high-quality dataset for text-to-video generation. CoRR, abs/2407.02371, 2024. 1, 6, 7", + "[43] OpenAI. Sora overview:https://openai.com/index/soral/, 2024.1,3", + "[44] Hao Ouyang, Qiuyu Wang, Yuxi Xiao, Qingyan Bai, Juntao Zhang, Kecheng Zheng, Xiaowei Zhou, Qifeng Chen, and Qifeng Chen. Codef: Content deformation fields for temporally consistent video processing. In IEEE Conf. Comput. Vis. Pattern Recog., pages 8089-8099, 2024. 3", + "[45] Wenqi Ouyang, Yi Dong, Lei Yang, Jianlou Si, and Xingang Pan. I2vedit: First-frame-guided video editing via image-to-video diffusion models. CoRR, abs/2405.16537, 2024. 2, 3, 7", + "[46] Gaurav Parmar, Taesung Park, Srinivasa Narasimhan, and Jun-Yan Zhu. One-step image translation with text-to-image models. CoRR, abs/2403.12036, 2024. 2", + "[47] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed A. A. Osman, Dimitrios Tzionas, and Michael J. Black. Expressive body capture: 3d hands, face, and body from a single image. In IEEE Conf. Comput. Vis. Pattern Recog., 2019. 8", + "[48] Bohao Peng, Jian Wang, Yuechen Zhang, Wenbo Li, Ming-Chang Yang, and Jiaya Jia. Controlnext: Powerful and efficient control for image and video generation. CoRR, abs/2408.06070, 2024. 2", + "[49] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763, 2021. 7", + "[50] Elad Richardson, Yuval Alaluf, Or Patashnik, Yotam Nitzan, Yaniv Azar, Stav Shapiro, and Daniel Cohen-Or. Encoding in style: A stylegan encoder for image-to-image translation. In IEEE Conf. Comput. Vis. Pattern Recog., pages 2287–2296, 2021. 2", + "[51] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In IEEE Conf. Comput. Vis. Pattern Recog., pages 10674-10685, 2022. 1, 2, 3", + "[52] Manuel Ruder, Alexey Dosovitskiy, and Thomas Brox. Artistic style transfer for videos and spherical images. Int. J. Comput. Vis., 126(11):1199-1219, 2018. 3", + "[53] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L. Denton, Seyed Kamyar Seyed Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, Jonathan Ho, David J. Fleet, and Mohammad Norouzi. Photorealistic text-to-image diffusion models with deep language understanding. In Adv. Neural Inform. Process. Syst., 2022. 1, 2", + "[54] Tim Salimans and Jonathan Ho. Progressive distillation for fast sampling of diffusion models. In Int. Conf. Learn. Represent., 2022. 3", + "[55] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "23388", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, Patrick Schramowski, Srivatsa Kundurthy, Katherine Crowson, Ludwig Schmidt, Robert Kaczmarczyk, and Jenia Jitsev. LAION-5B: an open large-scale dataset for training next generation image-text models. In Adv. Neural Inform. Process. Syst., 2022. 1, 6", + "[56] Xiaoyu Shi, Zhaoyang Huang, Fu-Yun Wang, Weikang Bian, Dasong Li, Yi Zhang, Manyuan Zhang, Ka Chun Cheung, Simon See, Hongwei Qin, Jifeng Dai, and Hongsheng Li. Motion-i2v: Consistent and controllable image-to-video generation with explicit motion modeling. In ACM SIGGRAPH, page 111. ACM, 2024. 3", + "[57] Uriel Singer, Amit Zohar, Yuval Kirstain, Shelly Sheynin, Adam Polyak, Devi Parikh, and Yaniv Taigman. Video editing via factorized diffusion distillation. CoRR, abs/2403.09334, 2024. 3", + "[58] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In Int. Conf. Learn. Represent., 2021. 6", + "[59] Jia-Mu Sun, Tong Wu, and Lin Gao. Recent advances in implicit representation-based 3d shape generation. Vis. Intell., 2(1), 2024. 8", + "[60] Vchitect Team and Shanghai Artificial Intelligence Laboratory. Vchitect-2.0: Parallel transformer for scaling up video diffusion models, 2024. 3", + "[61] Zheng Tianpeng, Chen Yanxiang, Wen Xinzhe, Li Yancheng, and Wang Zhiyuan. Research on diffusion model generated video datasets and detection benchmarks. Journal of Image and Graphics, pages 1-13, 2024. 1", + "[62] Rotem Tzaban, Ron Mokady, Rinon Gal, Amit Bermano, and Daniel Cohen-Or. *Stitch it in time: Gan-based facial editing of real videos*. In SIGGRAPH Asia Conference Papers, pages 29:1–29:9. ACM, 2022. 3", + "[63] Guangming Wang, Lei Pan, Songyou Peng, Shaohui Liu, Chenfeng Xu, Yanzi Miao, Wei Zhan, Masayoshi Tomizuka, Marc Pollefeys, and Hesheng Wang. Nerf in robotics: A survey. CoRR, abs/2405.01333, 2024. 8", + "[64] Jiuniu Wang, Hangjie Yuan, Dayou Chen, Yingya Zhang, Xiang Wang, and Shiwei Zhang. Modelscope text-to-video technical report. CoRR, abs/2308.06571, 2023. 3", + "[65] Ting-Chun Wang, Ming-Yu Liu, Jun-Yan Zhu, Andrew Tao, Jan Kautz, and Bryan Catanzaro. High-resolution image synthesis and semantic manipulation with conditional gans. In IEEE Conf. Comput. Vis. Pattern Recog., pages 8798-8807, 2018. 2", + "[66] Zhouxia Wang, Ziyang Yuan, Xintao Wang, Yaowei Li, Tianshui Chen, Menghan Xia, Ping Luo, and Ying Shan. Motionctrl: A unified and flexible motion controller for video generation. In ACM SIGGRAPH, page 114, 2024. 3", + "[67] Jianzong Wu, Xiangtai Li, Yanhong Zeng, Jiangning Zhang, Qianyu Zhou, Yining Li, Yunhai Tong, and Kai Chen. Motionbooth: Motion-aware customized text-to-video generation. CoRR, abs/2406.17758, 2024. 3", + "[68] Tao Wu, Yong Zhang, Xintao Wang, Xianpan Zhou, Guangcong Zheng, Zhongang Qi, Ying Shan, and Xi Li. Customcrafter: Customized video generation with preserving motion and concept composition abilities. CoRR, abs/2408.13239, 2024. 3" + ], + "bbox": [ + 91, + 90, + 482, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[69] Saining Xie and Zhuowen Tu. Holistically-nested edge detection. In Int. Conf. Comput. Vis., pages 1395-1403, 2015. 6", + "[70] Jinbo Xing, Hanyuan Liu, Menghan Xia, Yong Zhang, Xintao Wang, Ying Shan, and Tien-Tsin Wong. Tooncrafter: Generative cartoon interpolation. CoRR, abs/2405.17933, 2024. 2, 3, 7", + "[71] Shuai Yang, Yifan Zhou, Ziwei Liu, and Chen Change Loy. Rerender A video: Zero-shot text-guided video-to-video translation. In SIGGRAPH Asia Conference Papers, pages 95:1-95:11. ACM, 2023. 2, 3", + "[72] Shiuyuan Yang, Liang Hou, Haibin Huang, Chongyang Ma, Pengfei Wan, Di Zhang, Xiaodong Chen, and Jing Liao. Direct-a-video: Customized video generation with user-directed camera movement and object motion. In ACM SIGGRAPH, page 113, 2024. 3", + "[73] Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, et al. Cogvideox: Text-to-video diffusion models with an expert transformer. arXiv preprint arXiv:2408.06072, 2024. 1, 2, 3, 6, 7", + "[74] Danah Yatim, Rafail Fridman, Omer Bar-Tal, Yoni Kasten, and Tali Dekel. Space-time diffusion features for zero-shot text-driven motion transfer. In IEEE Conf. Comput. Vis. Pattern Recog., pages 8466-8476, 2024. 3", + "[75] Chen Yuan, Zhao Yang, Zhang Xiaojuan, and Liu Xiaoping. Sketch colorization with finite color space prior. Journal of Image and Graphics, 29(4):978-988, 2024. 2", + "[76] Liang Yuan, Dingkun Yan, Suguru Saito, and Issei Fujishiro. Diffmat: Latent diffusion models for image-guided material generation. Visual Informatics, 8(1):6-14, 2024. 2", + "[77] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Int. Conf. Comput. Vis., pages 3813-3824, 2023. 2, 3, 4, 6, 7", + "[78] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In IEEE Conf. Comput. Vis. Pattern Recog., pages 586-595, 2018. 7", + "[79] Yuxin Zhang, Fan Tang, Nisha Huang, Haibin Huang, Chongyang Ma, Weiming Dong, and Changsheng Xu. Motioncrafter: One-shot motion customization of diffusion models. CoRR, abs/2312.05288, 2023. 3", + "[80] Yabo Zhang, Yuxiang Wei, Dongsheng Jiang, Xiaopeng Zhang, Wangmeng Zuo, and Qi Tian. Controlvideo: Training-free controllable text-to-video generation. In Int. Conf. Learn. Represent., 2024. 3", + "[81] Zhenghao Zhang, Junchao Liao, Menghao Li, Long Qin, and Weizhi Wang. Tora: Trajectory-oriented diffusion transformer for video generation. CoRR, abs/2407.21705, 2024. 3", + "[82] Zhixing Zhang, Bichen Wu, Xiaoyan Wang, Yaqiao Luo, Luxin Zhang, Yinan Zhao, Peter Vajda, Dimitris N. Metaxas, and Licheng Yu. AVID: any-length video inpainting with diffusion model. In IEEE Conf. Comput. Vis. Pattern Recog., pages 7162-7172, 2024. 3", + "[83] Shihao Zhao, Dongdong Chen, Yen-Chun Chen, Jianmin Bao, Shaozhe Hao, Lu Yuan, and Kwan-Yee K. Wong." + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "23389", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Uni-controlnet: All-in-one control to text-to-image diffusion models. In Adv. Neural Inform. Process. Syst., 2023. 2", + "[84] Zangwei Zheng, Xiangyu Peng, Tianji Yang, Chenhui Shen, Shenggui Li, Hongxin Liu, Yukun Zhou, Tianyi Li, and Yang You. Open-sora: Democratizing efficient video production for all, 2024. 1, 2, 3", + "[85] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A. Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In Int. Conf. Comput. Vis., pages 2242-2251, 2017. 2", + "[86] Siting Zhu, Guangming Wang, Dezhi Kong, and Hesheng Wang. 3d gaussian splatting in robotics: A survey. CoRR, abs/2410.12262, 2024. 8" + ], + "bbox": [ + 91, + 90, + 482, + 273 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "23390", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2025/SketchVideo_ Sketch-based Video Generation and Editing/86ad126c-969d-47b0-8162-a7cc53739116_model.json b/2025/SketchVideo_ Sketch-based Video Generation and Editing/86ad126c-969d-47b0-8162-a7cc53739116_model.json new file mode 100644 index 0000000000000000000000000000000000000000..913eef9d481f2b7493e764309b8d14ab46d91097 --- /dev/null +++ b/2025/SketchVideo_ Sketch-based Video Generation and Editing/86ad126c-969d-47b0-8162-a7cc53739116_model.json @@ -0,0 +1,3106 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.205, + 0.131, + 0.791, + 0.154 + ], + "angle": 0, + "content": "SketchVideo: Sketch-based Video Generation and Editing" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.18, + 0.914, + 0.201 + ], + "angle": 0, + "content": "Feng-Lin Liu\\(^{1,2}\\) Hongbo Fu\\(^{3}\\) Xintao Wang\\(^{4}\\) Weicai Ye\\(^{4}\\) Pengfei Wan\\(^{4}\\) Di Zhang\\(^{4}\\) Lin Gao\\(^{1,2*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.105, + 0.206, + 0.914, + 0.235 + ], + "angle": 0, + "content": "\\(^{1}\\)Beijing Key Laboratory of Mobile Computing and Pervasive Device, Institute of Computing Technology, Chinese Academy of Sciences \\(^{2}\\)University of Chinese Academy of Sciences \\(^{3}\\)Hong Kong University of Science and Technology \\(^{4}\\)Kuaishou Technology" + }, + { + "type": "image_footnote", + "bbox": [ + 0.114, + 0.258, + 0.232, + 0.323 + ], + "angle": 0, + "content": "\"A close-up of a majestic male lion's face, with a tawny mane, piercing amber eyes, and subtle scars, bathed in golden sunlight, showcasing his power and wild spirit.\"" + }, + { + "type": "image", + "bbox": [ + 0.239, + 0.246, + 0.378, + 0.333 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.115, + 0.333, + 0.197, + 0.42 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.2, + 0.334, + 0.376, + 0.375 + ], + "angle": 0, + "content": "\"The video is a drone shot of a serene landscape featuring a deep blue lagoon surrounded by lush green trees and rocky cliffs. The lagoon is nestled between the mountains, creating a picturesque scene.\"" + }, + { + "type": "image_footnote", + "bbox": [ + 0.201, + 0.377, + 0.376, + 0.417 + ], + "angle": 0, + "content": "\"A clear blue sky above the mountains, with a few clouds scattered across it. A deep blue lake nestled between the mountains. The mountains have patches of snow that still cling to their peaks.\"" + }, + { + "type": "list", + "bbox": [ + 0.2, + 0.334, + 0.376, + 0.417 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.38, + 0.244, + 0.88, + 0.42 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.447, + 0.419, + 0.602, + 0.428 + ], + "angle": 0, + "content": "(a) Sketch-based Video Generation" + }, + { + "type": "image", + "bbox": [ + 0.129, + 0.428, + 0.881, + 0.557 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.448, + 0.556, + 0.588, + 0.566 + ], + "angle": 0, + "content": "(b) Sketch-based Video Editing" + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.566, + 0.908, + 0.622 + ], + "angle": 0, + "content": "Figure 1. Our method enables high-quality video generation (a) and editing (b) based on sketch and text inputs. (a) Top: With the same text prompt, different keyframe sketches lead to results with similar semantics but diverse sketch-faithful geometry. (a) Bottom: With the same sketches, varied text prompts yield diverse appearances. (b) Users can also edit real videos by drawing on keyframe sketches, with edits automatically propagated even when edited objects in original videos have translation and rotation." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.635, + 0.327, + 0.651 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.667, + 0.483, + 0.879 + ], + "angle": 0, + "content": "Video generation and editing conditioned on text prompts or images have undergone significant advancements. However, challenges remain in accurately controlling global layout and geometry details solely by texts, and supporting motion control and local modification through images. In this paper, we aim to achieve sketch-based spatial and motion control for video generation and support fine-grained editing of real or synthetic videos. Based on the DiT video generation model, we propose a memory-efficient control structure with sketch control blocks that predict residual features of skipped DiT blocks. Sketches are drawn on one or two keyframes (at arbitrary time points) for easy interaction. To propagate such temporally sparse sketch conditions across all frames, we propose an inter-frame at" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.638, + 0.907, + 0.774 + ], + "angle": 0, + "content": "tention mechanism to analyze the relationship between the keyframes and each video frame. For sketch-based video editing, we design an additional video insertion module that maintains consistency between the newly edited content and the original video's spatial feature and dynamic motion. During inference, we use latent fusion for the accurate preservation of unedited regions. Extensive experiments demonstrate that our SketchVideo achieves superior performance in controllable video generation and editing." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.8, + 0.645, + 0.815 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Diffusion-based text-to-image [12, 51, 53] and text-to-video [27, 35, 43, 73, 84] models advance significantly due to improvements in datasets [4, 42, 55, 61] and denoising network architectures [12, 51]. While text prompts effectively describe high-level semantics, they lack control of scene" + }, + { + "type": "page_footnote", + "bbox": [ + 0.115, + 0.887, + 0.395, + 0.9 + ], + "angle": 0, + "content": "*Corresponding Author: Lin Gao (gaolin@ict.ac.cn)" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23379" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.288 + ], + "angle": 0, + "content": "Layouts and geometric details. To address this, existing video generation methods [2, 25, 73] utilize images as additional conditions but raise the questions of how to generate input images and achieve detailed editing. Sketching serves as a user-friendly interaction tool to capture spatial content and shape details accurately. One or two sketches are already sufficient to convey desired scene structures and motion information for short video clips (around 6 seconds), which are the target of our and most existing video generation methods. However, using such sparse keyframe sketches presents several challenges, including reasonably completing the missing frames, improving memory efficiency, and addressing the limited size of video datasets." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.289, + 0.483, + 0.56 + ], + "angle": 0, + "content": "A naive solution is to translate the input keyframe sketches into images and then utilize interpolation methods [16, 37, 70] for video generation. However, it is nontrivial to ensure consistency during keyframe sketch-to-image generation, which significantly affects the video quality. This approach also struggles to generate extrapolation frames when applying conditions in intermediate frames rather than beginning and ending time points. Another possible approach is to utilize white placeholders to fill missing condition frames and directly apply ControlNet [77] into video models, similar to SparseCtrl [22]. However, this requires the same network to process both sketches and white placeholders simultaneously, while the pretrained blocks handle tasks far from this sparse propagation. Additionally, for DiT-based video frameworks [27, 84], the traditional strategy [11] that copies half of the pretrained model as a condition network easily causes the out-of-memory issue." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.56, + 0.483, + 0.892 + ], + "angle": 0, + "content": "To address these issues, we propose a novel sketch condition network specifically designed for the DiT-based video generation architecture (CogVideoX [73] in our work). Following ControlNet [11, 77], we employ a trainable copy of CogVideoX's DiT block to process only the sketch inputs and generate control features. No white placeholder is processed to align with the pretrained weights and reduce learning complexity. To propagate these keyframe features, we design an inter-frame attention mechanism that captures the relationship between the control keyframes and all video frames. Our approach computes query and key features from noisy latent while extracting value features from sketch conditions. This design leverages frame-to-frame similarity for control propagation. The above components consist of a single sketch control block. Instead of copying a half number of the pretrained blocks to construct the control network [11], we use only 5 sketch control blocks out of the 30 DiT blocks available in CogVideoX-2b. We design a novel uniformly distributed skip structure to add the control signals to different levels of features in discrete blocks (0, 6, 12, 18, and 24), achieving effective spatial control while improving memory efficiency. During" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.121 + ], + "angle": 0, + "content": "training, an external image dataset is incorporated to solve the challenge of limited video data." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.122, + 0.905, + 0.272 + ], + "angle": 0, + "content": "Beyond generation, interactive editing of real or synthetic videos further enhances creative flexibility. Existing methods [34, 40, 45, 71] achieve interesting text-based editing or effectively propagate single image editing into videos. Despite their effectiveness in appearance modification, they struggle with shape manipulation and object insertion, as they preserve the original temporal motion. Such information is missing for newly introduced content. Moreover, precisely identifying and preserving unedited regions for localized editing remains a challenge." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.273, + 0.905, + 0.453 + ], + "angle": 0, + "content": "We propose a sketch-based editing method for detailed local modification. Rather than editing a single image and propagating changes, we directly construct an editing network based on our sketch control network. To analyze the relationship between edited regions and the original video, we incorporate a video insertion module that takes the original video with masked edited regions as inputs. The modified sketch control blocks generate residual features that capture temporally and spatially coherent contents in the edited areas. To accurately preserve unedited regions and achieve seamless fusion, the newly edited regions are blended with the original video in the latent space." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.455, + 0.905, + 0.651 + ], + "angle": 0, + "content": "Extensive experiments demonstrate that our method outperforms existing approaches in video generation and editing. Our contributions are summarized as follows: 1) We propose SketchVideo, a novel sketch-based video generation and editing framework that enables detailed geometry control and manipulation using keyframe sketches, as shown in Fig. 1. 2) We design a sketch condition network that predicts skipped control features for the DiT framework, with an inter-frame attention mechanism to propagate one or two sketch conditions across the video. 3) We propose a video insertion module that analyzes the relationship between drawn sketches and original videos, utilizing a latent fusion strategy to preserve unedited regions accurately." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.665, + 0.655, + 0.68 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.689, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Sketch-based Image Generation. GAN-based methods [21] have achieved great success in category-restricted sketch-to-image translation [13, 14, 50, 65, 75, 85]. Recently, diffusion-based text-to-image models [51, 53] handle general categories with conditional models like ControlNet [77], T2I-Adapter [41] and further advances [30, 46, 48, 76, 83]. Beyond U-Net, the DiT backbone [12] enables image generation, with pixART-\\(\\delta\\) [11] utilizing the first half of the pretrained model's blocks as a trainable network to predict corresponding control residual features. Video generation, however, introduces additional challenges. For ease of interaction, we expect sketches specified only for a sparse set of keyframes, making it difficult to generate frames without sketch inputs. Additionally, video genera" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23380" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.137 + ], + "angle": 0, + "content": "tion costs significantly higher memory resources, making methods [11, 77] that replicate half of the base model as a sketch encoder easily out of memory." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.137, + 0.483, + 0.318 + ], + "angle": 0, + "content": "Diffusion-based Video Generation. VDM [26] pioneered diffusion-based video generation with a 3D U-Net denoising network. To improve quality, subsequent works [2, 3, 9, 10, 23, 25, 33, 64] integrated temporal modules into text-to-image models [51] to enable text- and image-conditioned video synthesis. Considering the efficiency, the DiT architecture is further used in Sora [43] and open-source projects [35, 60, 84]. Despite these advancements, subtle flickering artifacts remain in the results. CogVideoX [27, 73] further proposes a 3D full attention that merges the spatial and temporal attention, facilitating the generation of long-duration and high-resolution videos." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.319, + 0.483, + 0.605 + ], + "angle": 0, + "content": "Building on existing video generation models, various conditions have been introduced to control the generation, such as camera movement [24, 66, 72], subject identity [31, 68], key-point trajectory [36, 56, 81] and example motions [67, 79]. However, they often lack control over spatial layouts and geometric details. Some methods [15, 18, 80] address this by extending image-based ControlNet [77] to videos but require all-frame conditions that are tedious for sketch interaction. SparseCtrl [22] tackles this by using white images for completions. However, due to the restriction of its base model [23] and a simple network design, its results suffer from temporal flickering. Similar ideas have been applied to cartoon interpolation [70] and colorization [28] with line art as input, but their outputs are limited to cartoon style. Our method utilizes sparse inputs, including hand-drawn sketches on one or two keyframes, to generate temporally stable and realistic videos. Additionally, our method further supports the sketch-based detailed editing of existing videos." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.606, + 0.483, + 0.892 + ], + "angle": 0, + "content": "Deep Learning-based Video Editing. Pioneer works achieve video editing by style transfer [29, 52], GAN inversion [39, 62], and layered representations [32]. The advent of diffusion models further provides new editing paradigms. One category of such methods leverages image generation models to achieve compelling editing results, using temporal consistency techniques such as layered representations [6, 44], cross-frame attention [5, 17, 19, 57] and pixel warping [71]. A second category of methods employ video generation models. These works propagate the edits applied on the first frame into the other frames [34, 45], or utilize an inpainting strategy to achieve text-based editing [82] and motion modification [40]. To achieve text-based large-scale shape modification while maintaining motion features, a space-time feature loss [74] is designed to guide the inference process. Our method moves beyond traditional text- or image-based editing approaches, allowing users to draw one or two keyframe sketches at arbitrary time points for interactive video editing. Additionally, our method effectively" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.122 + ], + "angle": 0, + "content": "handles sketch-based shape manipulation and dynamic object insertion, which are challenging for previous works." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.122, + 0.905, + 0.332 + ], + "angle": 0, + "content": "Controllable Attention Mechanism. Attention across frames is initially used to ensure temporal consistency inAnimateDiff [23] and subsequent video generation models. Subsequent works, such as VideoBooth [31] and Still-Moving [8], utilize it to learn identity-aware features for customized video generation. For video editing, existing works [17, 20, 57] utilize cross-frame attention to capture the temporal motion of input videos, enabling effective propagation of editing operations. Our method applies this idea to pixel-aligned sketch-based video generation and propagates the spatial geometric conditions instead of identity customization. We use a new feature derivation strategy for enhanced spatial control, differentiating our approach from traditional cross-attention mechanisms." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.347, + 0.649, + 0.365 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.372, + 0.905, + 0.553 + ], + "angle": 0, + "content": "This section introduces our sketch-based video generation and editing framework. In Sec.3.1, we provide an overview of CogVideoX-2b [73], a pretrained text-to-video generation method, which we will use for sketch-based video generation in our work. In Sec.3.2, we describe our sketch condition network specifically designed for the DiT architecture, which contains sketch control blocks to predict residual features. Within each control block, an interframe attention mechanism is designed to propagate the input sketches. In Sec.3.3, we detail the editing framework, which incorporates a video insertion module and latent fusion to preserve the features of the original video." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.563, + 0.644, + 0.58 + ], + "angle": 0, + "content": "3.1. Preliminary" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.585, + 0.905, + 0.751 + ], + "angle": 0, + "content": "CogVideoX [73] is a text-to-video generation model that employs a 3D causal VAE and builds diffusion within the latent space. In CogVideoX-2b, the VAE model performs \\(8 \\times 8\\) spatial and \\(4 \\times\\) temporal downsampling, followed by a DiT architecture with 30 blocks for video generation. Within each block, a 3D full attention processes the concatenated text embeddings and patchified video latents, followed by a feed-forward layer to output the features. The 3D full attention merges the commonly used separate spatial and temporal attention [2, 23, 84] to improve the temporal coherence. The training objective of diffusion is:" + }, + { + "type": "equation", + "bbox": [ + 0.512, + 0.759, + 0.905, + 0.84 + ], + "angle": 0, + "content": "\\[\nL (\\theta) := \\mathrm {E} _ {t, x _ {0} ^ {1: N}, y, \\epsilon} \\left\\| \\epsilon - \\epsilon_ {\\theta} \\left(\\sqrt {\\bar {\\alpha} _ {t}} x _ {0} ^ {1: N} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon , t, y\\right) \\right\\| ^ {2}, \\quad (1) \\text {w h e r e} t \\text {i s s a m p l e d b e t w e e n 1 a n d T (d e n o s i n g s t e p s)}, \\epsilon \\text {i s t h e r a n d o n n o i s e ,} y \\text {i s t e x t p r o m p t s , a n d} x _ {0} ^ {1: N} \\text {i s t h e v i d e o d a w i t h} N \\text {f r a m e s .} \\text {T h e v - p r e d i c t i o n [ 5 4 ] a n d z e r o S N R [ 3 8 ] a r e u l i z a t e d f o r t h e d i f f u s i o n s e t t i n g .}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.849, + 0.794, + 0.864 + ], + "angle": 0, + "content": "3.2. Sketch-based Video Generation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Given text prompts and one or two keyframe sketches with corresponding time points \\( t_1, t_2 \\), our method generates" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "23381" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.091, + 0.089, + 0.303, + 0.342 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.128, + 0.345, + 0.273, + 0.355 + ], + "angle": 0, + "content": "(a) Sketch Condition Network" + }, + { + "type": "image", + "bbox": [ + 0.306, + 0.088, + 0.909, + 0.343 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.398, + 0.345, + 0.558, + 0.356 + ], + "angle": 0, + "content": "(b) Control Block for Generation" + }, + { + "type": "image_caption", + "bbox": [ + 0.704, + 0.345, + 0.845, + 0.356 + ], + "angle": 0, + "content": "(c) Control Block for Editing" + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.359, + 0.907, + 0.442 + ], + "angle": 0, + "content": "Figure 2. Our framework for sketch-based video generation and editing. (a) Our sketch condition network for the DiT-based video generation architecture has a skip structure and five sketch control blocks that predict residual features. (b) For generation, features are extracted from temporally sparse input sketches and propagated through inter-frame attention. The input sketches are provided for one or two keyframes (the second sketch is shown as a dotted line). In the top left corner of (b), the prompt and timestep inputs are shown. (c) For editing, the same sketch control block (b) is utilized, with an additional video insertion module and video masks \\( M^{1:N} \\) to analyze the relationship between edited and unedited regions. The 3D causal VAE is omitted to save space." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.45, + 0.483, + 0.54 + ], + "angle": 0, + "content": "video clips that respect the input text prompts and sketches. As shown in Fig. 2, we design a sketch condition network for effective control. The input sketches are encoded into the latent space by the pretrained VAE, followed by patchifying and time-aware positional embedding to generate sketch latents \\( s_0^{t_1,t_2} \\)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.54, + 0.483, + 0.706 + ], + "angle": 0, + "content": "Skip Residual Structure. In sketch-based image generation, methods like ControlNet [77] and pixART- \\(\\delta\\) [11] copy the half of the base model as their sketch encoder to fully utilize the pretrained text-to-image models. However, applying this to video generation, as done in SparseCtrl [22], is memory inefficient because it requires adding half of the base model's parameters. Unlike the U-Net architecture, the DiT network does not have an explicit encoder and decoder. Therefore, the assumption in pixART\\(\\delta\\) [11] that the first half of blocks serve as the encoder can be improved." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.707, + 0.483, + 0.889 + ], + "angle": 0, + "content": "Instead of borrowing local consecutive blocks of the base model as the encoder and predicting their residual features, we recognize that blocks at different depths process distinct feature levels, which should be considered for sketch control. As illustrated in Fig. 2 (a), we propose a novel skip residual structure that reduces the number of blocks while enabling effective sketch control and high-quality generation. Our sketch condition network contains 5 sketch control blocks, uniformly distributed across the pretrained generation network to predict residual features for blocks 0, 6, 12, 18, and 24 of the original video generation model. This structure efficiently integrates sketch control informa" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.45, + 0.907, + 0.479 + ], + "angle": 0, + "content": "tion into multiple feature levels, enhancing the analysis of input conditions and original semantic features." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.48, + 0.906, + 0.563 + ], + "angle": 0, + "content": "Sketch Control Block. In the \\(i\\)-th sketch control block, the input consists of hidden video features \\(h_i^{1:N}\\) and sketch features \\(s_{i-1}^{t_1,t_2}\\), and the output is the residual features \\(\\overline{h}_i^{1:N}\\) and updated sketch features \\(s_i^{t_1,t_2}\\). As shown in Fig. 2 (b), the sketch feature \\(s_{i-1}^{t_1,t_2}\\) is processed:" + }, + { + "type": "equation", + "bbox": [ + 0.606, + 0.57, + 0.905, + 0.591 + ], + "angle": 0, + "content": "\\[\ns _ {i} ^ {t _ {1}, t _ {2}} = \\operatorname {F e e d F o r w a r d} \\left(s _ {i - 1} ^ {t _ {1}, t _ {2}}\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.599, + 0.906, + 0.642 + ], + "angle": 0, + "content": "where the output sketch feature \\( s_i^{t_1,t_2} \\) is used for sketch control propagation and as the input to the next sketch control block." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.645, + 0.907, + 0.84 + ], + "angle": 0, + "content": "To propagate the sketch inputs, a direct approach [22] replaces missing sketches with white images and employs a trainable copy of the pretrained DiT block to predict residual features. However, as shown in Fig. 5, this leads to fuzzy details in challenging cases, as the network randomly processes both sketches and white images. This mixed input is far from the pretrained model's task. To address the above issue, we employ a trainable copy of pretrained DiT block to process only the sketch inputs (no white placeholder), aligning with the pretrained weights and reducing learning difficulty. The resulting keyframe sketch features, denoted as \\( c_{i}^{t_{1},t_{2}} \\), are propagated to all frames through an inter-frame attention approach." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Inter-frame Attention. We utilize the input hidden features of all frames to calculate \\( Q \\) and the hidden features corresponding to the control frames to calculate \\( K \\). During attention computation, this approach captures the internal" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23382" + } + ], + [ + { + "type": "image_footnote", + "bbox": [ + 0.092, + 0.091, + 0.258, + 0.127 + ], + "angle": 0, + "content": "A small, fluffy white rabbit hops across a lush meadow at sunset, pausing to listen to rustling leaves before continuing its joyful dance." + }, + { + "type": "image_footnote", + "bbox": [ + 0.092, + 0.139, + 0.258, + 0.173 + ], + "angle": 0, + "content": "A panoramic video captures a majestic city with towering skyscrapers, quaint red-tiled homes, and a tranquil lake, all under a sky of drifting white clouds." + }, + { + "type": "image_footnote", + "bbox": [ + 0.092, + 0.189, + 0.258, + 0.223 + ], + "angle": 0, + "content": "A video captures close-ups of a cat's face with green eyes and a brown-black spotted coat, showing three expressions: alert, relaxed, and distant." + }, + { + "type": "image_footnote", + "bbox": [ + 0.092, + 0.235, + 0.258, + 0.286 + ], + "angle": 0, + "content": "A video captures the explosive disintegration of a light blue planet, fragments and fiery debris catapulted into starry space, creating a stark contrast between celestial serenity and violent upheaval." + }, + { + "type": "image_footnote", + "bbox": [ + 0.092, + 0.289, + 0.258, + 0.332 + ], + "angle": 0, + "content": "A time-lapse aerial video shows a large, ornate domed building surrounded by lush greenery, transitioning from daylight to sunset to night, illuminated by artificial lights." + }, + { + "type": "list", + "bbox": [ + 0.092, + 0.091, + 0.258, + 0.332 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.262, + 0.09, + 0.355, + 0.135 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.262, + 0.139, + 0.355, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.262, + 0.189, + 0.355, + 0.237 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.262, + 0.239, + 0.355, + 0.285 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.262, + 0.287, + 0.355, + 0.336 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.364, + 0.091, + 0.903, + 0.137 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.365, + 0.139, + 0.903, + 0.186 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.365, + 0.189, + 0.903, + 0.236 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.365, + 0.238, + 0.903, + 0.285 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.365, + 0.286, + 0.903, + 0.334 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.337, + 0.907, + 0.38 + ], + "angle": 0, + "content": "Figure 3. The sketch-based video generation results. Left: The input text prompts and sketches. Right: The video generation results. It can be seen that the generated results show high quality and good faithfulness with the input sketches. Our method can handle one/two keyframe sketch(es) at arbitrary user-specified time points (the frames corresponding to the input time points are highlighted by orange)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.387, + 0.484, + 0.53 + ], + "angle": 0, + "content": "relationship between all frames and control keyframes, allowing propagation of \\(V\\) (derived from the keyframe sketch features). Our inter-frame attention is distinct from typical cross-frame attention, which uses both \\(K\\) and \\(V\\) from control conditions; instead, we leverage the inter-frame similarity within the input noisy hidden video features and progressively insert the sketches' spatial and temporal information. The output \\(\\widetilde{c}_i^{1:N}\\) is computed as: \\(\\mathrm{Attention}(Q,K,V) = \\mathrm{Softmax}(\\frac{QK^T}{\\sqrt{d}})\\cdot V\\), with" + }, + { + "type": "equation", + "bbox": [ + 0.108, + 0.538, + 0.482, + 0.557 + ], + "angle": 0, + "content": "\\[\nQ = W _ {q} \\cdot h _ {i} ^ {1: N}, K = W _ {k} \\cdot h _ {i} ^ {t _ {1}, t _ {2}}, V = W _ {v} \\cdot c _ {i} ^ {t _ {1}, t _ {2}}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.565, + 0.483, + 0.614 + ], + "angle": 0, + "content": "where \\( W_{q}, W_{k}, W_{v} \\) are trainable linear projection weights. The output of the inter-frame attention is fed into a feedforward layer to generate final residual features \\( \\overline{h}_{i}^{1:N} \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.615, + 0.484, + 0.735 + ], + "angle": 0, + "content": "Training Strategy. To train the sketch condition network, we employ a hybrid training strategy in two stages. In the first stage, to accelerate convergence and address the issue of limited video data, the network is trained both for image generation at arbitrary time points and video generation with one or two keyframe sketches (randomly selected from corresponding sketch videos). In the second stage, video data alone is used to improve the temporal coherence." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.743, + 0.34, + 0.759 + ], + "angle": 0, + "content": "3.3. Sketch-based Video Editing" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.484, + 0.903 + ], + "angle": 0, + "content": "For editing, given real or synthetic videos, users select one or two keyframes at arbitrary time points and modify the extracted sketches, with additional inputs of text prompts and masks \\( M^{1:N} \\) that label regions to be edited for all the frames. Our method then generates realistic, local editing results. The input video is multiplied by the inverted masks to remove information from the edited regions and is subsequently encoded to generate a masked video latent representation \\( v_{0}^{1:N} \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.387, + 0.907, + 0.599 + ], + "angle": 0, + "content": "Video Insertion Module. For sketch-based editing, newly generated contents within the mask regions should be coherent with the original spatial and temporal features in the unedited regions. Thus, we design a video insertion module that analyzes the relationship between the input sketches and the original video. The video insertion module takes \\( v_{i-1}^{1:N} \\) (input video latent or generated by a previous control block) as input and predicts the updated video features \\( v_{i}^{1:N} \\), similar to the sketch generation process. Since video features are not temporally sparse, we use a trainable copy of CogVideoX-2b's DiT block to directly generate video insertion features \\( \\widetilde{v}_{i}^{1:N} \\). The sketch branch output \\( \\widetilde{c}_{i}^{1:N} \\) and video branch output \\( \\widetilde{v}_{i}^{1:N} \\) are multiplied by their respective masks and concatenated:" + }, + { + "type": "equation", + "bbox": [ + 0.583, + 0.611, + 0.907, + 0.632 + ], + "angle": 0, + "content": "\\[\n\\operatorname {C o n c a t} \\left(\\tilde {c} _ {i} ^ {1: N} * M ^ {1: N}, \\tilde {v} _ {i} ^ {1: N} * \\overline {{M}} ^ {1: N}\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.644, + 0.906, + 0.72 + ], + "angle": 0, + "content": "which serves as inputs to the feed-forward layers, producing final residual features incorporating the original video and sketch control information. This design ensures seamless integration of new contents with the original videos, effectively propagating edits across frames for dynamic motion." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.906, + 0.87 + ], + "angle": 0, + "content": "Training Strategy. Directly training the video editing network would lead to low fidelity with the input sketches, possibly because of the challenging interaction between the input sketches and videos. So we finetune it from the pretrained sketch condition network for generation and add the new video insertion module. The pretrained model already has good sketch fidelity and only requires learning video information. The network is trained in a self-supervised inpainting manner, with randomly generated masks to imitate real-world editing." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.906, + 0.902 + ], + "angle": 0, + "content": "Inference Latent Fusion. Although the original videos are encoded in the condition network, as shown in Fig. 8," + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23383" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.094, + 0.089, + 0.195, + 0.133 + ], + "angle": 0, + "content": "A snow-dusted volcano looms, steam plumes rising from its crest. Lush green flanks contrast with rocky slopes, while a rippling lake is in the foreground." + }, + { + "type": "image", + "bbox": [ + 0.198, + 0.09, + 0.498, + 0.131 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.325, + 0.131, + 0.374, + 0.137 + ], + "angle": 0, + "content": "Input Video" + }, + { + "type": "image", + "bbox": [ + 0.104, + 0.136, + 0.184, + 0.18 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.093, + 0.183, + 0.195, + 0.233 + ], + "angle": 0, + "content": "A mesmerizing ornamental fish glides through the ocean, its electric blue, fiery orange, and iridescent green scales shimmering as it navigates the depths with deli- cate, patterned fins." + }, + { + "type": "image", + "bbox": [ + 0.198, + 0.137, + 0.499, + 0.178 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.317, + 0.178, + 0.375, + 0.186 + ], + "angle": 0, + "content": "Edited Results" + }, + { + "type": "image", + "bbox": [ + 0.198, + 0.187, + 0.498, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.319, + 0.226, + 0.367, + 0.234 + ], + "angle": 0, + "content": "Input Video" + }, + { + "type": "image", + "bbox": [ + 0.198, + 0.235, + 0.498, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.316, + 0.275, + 0.375, + 0.283 + ], + "angle": 0, + "content": "Edited Results" + }, + { + "type": "image_caption", + "bbox": [ + 0.503, + 0.09, + 0.605, + 0.133 + ], + "angle": 0, + "content": "A man is speaking. He has thick hair and is wearing a black suit with a white shirt and a red tie underneath. The camera focuses on his face and upper body." + }, + { + "type": "image", + "bbox": [ + 0.606, + 0.09, + 0.905, + 0.131 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.728, + 0.131, + 0.776, + 0.138 + ], + "angle": 0, + "content": "Input Video" + }, + { + "type": "image", + "bbox": [ + 0.514, + 0.138, + 0.593, + 0.178 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.506, + 0.193, + 0.603, + 0.222 + ], + "angle": 0, + "content": "A green grassland and a small path extending into the distance. The grassland is covered with gravel." + }, + { + "type": "image", + "bbox": [ + 0.605, + 0.138, + 0.904, + 0.178 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.728, + 0.178, + 0.786, + 0.186 + ], + "angle": 0, + "content": "Edited Results" + }, + { + "type": "image", + "bbox": [ + 0.515, + 0.234, + 0.592, + 0.275 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.605, + 0.187, + 0.905, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.728, + 0.226, + 0.776, + 0.234 + ], + "angle": 0, + "content": "Input Video" + }, + { + "type": "image", + "bbox": [ + 0.605, + 0.234, + 0.905, + 0.275 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.725, + 0.275, + 0.783, + 0.283 + ], + "angle": 0, + "content": "Edited Results" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.287, + 0.907, + 0.33 + ], + "angle": 0, + "content": "Figure 4. The sketch-based video editing results. For each example, the text prompts and sketches are shown on the left. On the right, the input real videos are shown at the top, while the edited results with the control keyframe highlighted in orange are shown at the bottom. The editing region masks are manually provided by users, highlighted as orange boxes. Our method generates realistic local editing results." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.337, + 0.485, + 0.457 + ], + "angle": 0, + "content": "fine details might be lost during editing. To address this, we propose a latent fusion approach at inference. Specifically, we apply DDIM inversion [58] to generate noisy latent codes of input videos across the inference steps. At Steps 25 and 49 (out of 50 total steps), the latent codes in the unedited regions are replaced with these inversion latent codes, ensuring better preservation of the original video's details and improving the coherence of the edited regions." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.474, + 0.208, + 0.489 + ], + "angle": 0, + "content": "4. Evaluation" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.498, + 0.308, + 0.514 + ], + "angle": 0, + "content": "4.1. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.52, + 0.484, + 0.687 + ], + "angle": 0, + "content": "We implement SketchVideo based on CogVideoX-2b [73], trained on a subset of OpenVid [42] and LAION [55] datasets, with paired sketches from [7]. Training uses 8 NVIDIA H800 GPUs with a batch size of 8 and gradient accumulation of 4. For generation, one or two keyframe sketches are randomly sampled from the video, with 10,000 steps for each training stage. For editing, randomly drawn masks are applied and trained for 20,000 steps. For ease of reading, the input text prompts are simplified in figures. Additional implementation details and full input texts are available in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.698, + 0.184, + 0.713 + ], + "angle": 0, + "content": "4.2. Results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.484, + 0.885 + ], + "angle": 0, + "content": "Our method generates high-quality videos from one or two keyframe sketches and text prompts. As shown in Fig. 3, our method can accurately control the object shape and scene layout, such as the rabbit's pose (1st row) and the position of lakes and buildings (2nd row). Text-only inputs cannot achieve such detailed geometry control. Our method also achieves interesting dynamic motion interpolation and extrapolation, generating smooth and realistic transitions, as seen in the cat's head shaking (3rd row) and the building's movement (last row). This allows control over both spatial layout and dynamic motion." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.886, + 0.482, + 0.901 + ], + "angle": 0, + "content": "Our method also supports sketch-based video editing." + }, + { + "type": "image_caption", + "bbox": [ + 0.526, + 0.334, + 0.895, + 0.352 + ], + "angle": 0, + "content": "Input Text A time-lapse video shows a large white and black cruise ship with yellow and gold accents docking at a pier on an overcast day, reflecting in calm waters." + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.353, + 0.589, + 0.525 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.59, + 0.353, + 0.667, + 0.525 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.668, + 0.353, + 0.747, + 0.525 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.748, + 0.353, + 0.826, + 0.525 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.826, + 0.353, + 0.905, + 0.525 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.529, + 0.906, + 0.626 + ], + "angle": 0, + "content": "Figure 5. The comparison results of sketch-based video generation. Text prompts are shown on the top. On the left, we show the input sketches and sketch-based image generation results by ControlNet [77]. On the right, we show the results of the compared approaches, including AMT [37], SparseCtrl [22], CtrlCogVideo [77], and ours. Our method produces better results, especially for the intermediate frames." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.637, + 0.907, + 0.789 + ], + "angle": 0, + "content": "Users specify bounding boxes and draw sketches in those regions, and our method generates photorealistic, seamlessly integrated content. As shown in Fig. 4, the new contents blend well with unedited regions and have dynamic motions like hat/hair rotation with original objects (1st row) or interesting fish swimming (2nd row). Our method can handle diverse editing cases, including object insertion, component replacement, and object removal. The unedited regions are well-preserved thanks to our latent fusion approach." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.803, + 0.645, + 0.82 + ], + "angle": 0, + "content": "4.3. Comparison" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.906, + 0.9 + ], + "angle": 0, + "content": "For sketch-based generation, we compare our methods with three methods given the same keyframe sketches text prompts as inputs. For SparseCtrl [22], we use the official pretrained model and extract sketches by HED [69]. It is trained on videos in the WebVid-10M dataset [1] instead" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23384" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.089, + 0.172, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.173, + 0.089, + 0.328, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.328, + 0.089, + 0.402, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.402, + 0.089, + 0.482, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.259, + 0.483, + 0.356 + ], + "angle": 0, + "content": "Figure 6. The comparison results of sketch-based video editing. On the left, we show the drawn editing sketches (for the frames highlighted in orange), text prompts, and sketch-based image generation results by ControlNet [77]. On the right, we show the original videos and editing results by the compared methods, including InsV2V [17], AnyV2V [34], and ours. Our method generates the most realistic results and preserves unedited regions well." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.372, + 0.483, + 0.552 + ], + "angle": 0, + "content": "of the OpenVid-1M dataset [42] used in our method. We extend SparseCtrl to CogVideoX-2b [73], using pixART\\(\\delta\\) [11] as the sketch condition encoder (with 5 DiT blocks same as our method) and white images to complete the missing condition frames. We also compare with an interpolation baseline, which uses ControlNet to translate two-frame sketches into images and then interpolates them with AMT [37]. For the sketch-base editing task, we compare ours with a text-based video editing method InsV2V [17] and a first-frame editing method AnyV2V [34]. We compare with additional methods [16, 20, 45, 70], as shown in supplemental material." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.554, + 0.483, + 0.78 + ], + "angle": 0, + "content": "As shown in Fig. 5, ControlNet [77] translates sketches into realistic images, which, however, lack temporal consistency and vary with shading and content. The interpolation results of AMT [37] exhibit fuzzy details and artifacts. SparseCtrl [22], based onAnimateDiff [23], exhibits temporal flickering, such as the suddenly appearing tree and distorted tower top (see the orange boxes in the 2nd row). Extending SparseCtrl to CogVideoX-2b [73] (Ctrl-CogVideo) still generates fuzzy details in intermediate frames, possibly due to the CogVideoX-2b's pretrained self-attention being designed for dense inputs instead of sparse sketches. Our method generates realistic videos with clear details and good temporal coherence, with even small details like the electric wires in the top-right corner accurately propagated." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.483, + 0.901 + ], + "angle": 0, + "content": "For sketch-based editing (Fig. 6), InsV2V [17] generates interesting results with birds but lacks control over shape and geometry through text prompts alone. For image-based video editing method, we utilize ControlNet to edit the first frame and then propagate editing into the video by AnyV2V [34]. However, since the motion is borrowed from the original video, AnyV2V struggles with new content, leading to fuzzy details and distortion in the bird, as well" + }, + { + "type": "table", + "bbox": [ + 0.515, + 0.089, + 0.906, + 0.156 + ], + "angle": 0, + "content": "
MethodsLPIPS ↓CLIP ↑FidelityConsistencyRealism
AMT29.1796.123.133.513.57
SparseCtrl44.8596.482.792.942.83
Ctrl-CogVideo32.2398.042.862.472.50
Ours27.5698.311.211.081.11
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.158, + 0.907, + 0.201 + ], + "angle": 0, + "content": "Table 1. The quantitative results of sketch-based video generation comparison. The LPIPS and CLIP numbers are scaled up \\(100 \\times\\) with each cell colored to indicate the best." + }, + { + "type": "table", + "bbox": [ + 0.522, + 0.202, + 0.9, + 0.257 + ], + "angle": 0, + "content": "
MethodsLPIPS ↓CLIP ↑PSNR↑FidelityPreservationRealism
InsV2V13.6195.3916.842.582.262.61
AnyV2V11.9293.4713.682.352.692.34
Ours9.7498.3436.481.071.051.04
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.261, + 0.905, + 0.304 + ], + "angle": 0, + "content": "Table 2. The quantitative results of sketch-based video editing comparison. The LPIPS and CLIP numbers are scaled up \\(100 \\times\\) with each cell colored to indicate the best." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.311, + 0.906, + 0.386 + ], + "angle": 0, + "content": "as changes in unedited regions due to ControlNet's inability to retain the original features. In contrast, our method produces more realistic results with faithful representations of the sketch. More generation and editing results are available in the supplemental material and video." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.387, + 0.906, + 0.598 + ], + "angle": 0, + "content": "We follow SparseCtrl [22] and use the LPIPS [78] metric to measure sketch faithfulness between the input sketches and those extracted from the corresponding frames of the generated videos. We use the CLIP [49] similarity to assess temporal coherence. For sketch-based generation, we test 200 random examples from OpenVid [42], using the first and last frame sketches and the corresponding text prompts in the dataset. As shown in Table 1, our method achieves the lowest LPIPS and highest CLIP scores, indicating its superior performance. For sketch-based editing, we use additional an MSE metric to measure unedited region preservation. We utilize 10 examples with hand-drawn sketches as input. In Table 2, our method outperforms existing methods across all metrics, demonstrating its superiority." + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.598, + 0.597, + 0.638 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.517, + 0.64, + 0.597, + 0.69 + ], + "angle": 0, + "content": "Extreme close-up of chicken and green pepper kebabs grilling on a barbeque with flames. Shallow focus and light smoke. vivid colours" + }, + { + "type": "image", + "bbox": [ + 0.597, + 0.597, + 0.905, + 0.692 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.594, + 0.691, + 0.718, + 0.699 + ], + "angle": 0, + "content": "(a) w/o Inter-Frame (b) Sketch K, V" + }, + { + "type": "image_caption", + "bbox": [ + 0.732, + 0.692, + 0.774, + 0.699 + ], + "angle": 0, + "content": "(c) w/o Skip" + }, + { + "type": "image_caption", + "bbox": [ + 0.787, + 0.692, + 0.836, + 0.699 + ], + "angle": 0, + "content": "(d) w/o Image" + }, + { + "type": "image_caption", + "bbox": [ + 0.859, + 0.692, + 0.888, + 0.698 + ], + "angle": 0, + "content": "(e) ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.704, + 0.907, + 0.745 + ], + "angle": 0, + "content": "Figure 7. The ablation study results of sketch-based video generation. Our method generates more realistic and sketch-faithful results than the baselines." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.758, + 0.667, + 0.774 + ], + "angle": 0, + "content": "4.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.906, + 0.901 + ], + "angle": 0, + "content": "We conduct an ablation study to evaluate the effectiveness of each key component. Due to the restriction of computing resources, all the models are trained with a batch size of 2. For sketch-based generation, removing inter-frame attention and concatenating sketches with frame features at the temporal dimension results in strange control with low sketch faithfulness (Fig. 7 (a)). Similarly, if we replace the inter-frame attention with a typical cross-attention that uses" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "23385" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.093, + 0.088, + 0.149, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.152, + 0.089, + 0.235, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.236, + 0.089, + 0.318, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.318, + 0.089, + 0.4, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.401, + 0.089, + 0.482, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.224, + 0.483, + 0.28 + ], + "angle": 0, + "content": "Figure 8. The ablation study results of sketch-based video editing. Our method generates realistic results faithful to sketches. We utilize heat maps (top right) to visualize the difference between the edited and original frames." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.281, + 0.487, + 0.391 + ], + "angle": 0, + "content": "
Metricw/o Inter-Frame Sketch K,Vw/o Skipw/o ImageOurs
LPIPS ↓36.3332.5931.9134.5830.79
CLIP ↑98.1098.1997.6098.2498.48
Metricw/o Video modulew/o Pretrainw/o Latent FusionOurs
LPIPS ↓9.3112.609.779.74
CLIP ↑97.9798.1298.4498.34
PSNR ↑33.6136.0531.6936.48
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.393, + 0.484, + 0.44 + ], + "angle": 0, + "content": "Table 3. The qualitative results of ablation study for sketch-based generation and editing, with each cell colored to indicate the best and second best." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.447, + 0.483, + 0.627 + ], + "angle": 0, + "content": "sketch features for both Key and Value (b), the sketch fidelity is also degraded due to the lack of effective control propagation. If we remove the skip structure and predict residual features for the first 5 blocks (c), the realism of the generated results is negatively influenced due to less control ability. Removing image data during training (d) causes significant geometry mismatch with input sketches. Compared to these alternatives, our method produces the most realistic and faithful results. As shown in Table 3, our method has the lowest LPIPS values (supporting the best sketch fidelity) while the CLIP metric has similar results (human eyes cannot recognize temporal coherence difference)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.629, + 0.483, + 0.84 + ], + "angle": 0, + "content": "In sketch-based video editing (Fig. 8), removing the video insertion module (a) and solely utilizing the sketch condition encoder results in good sketch fidelity but obvious inconsistency with unedited regions. Training the sketch condition encoder from scratch without generation pretraining (b) reduces the faithfulness to the input sketches in the edited regions, which is also validated by the high LPIPS in Table 3. We use heat maps to reveal differences between edited and original frames. Removing the latent fusion strategy changes unedited region details (c), such as mountains and waves, as shown in the heat maps. The low PSNR value in unedited regions also supports this, as shown in Table 3. Our method maintains strong sketch fidelity, consistency, and unedited region preservation." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.849, + 0.214, + 0.865 + ], + "angle": 0, + "content": "4.5. User Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.903 + ], + "angle": 0, + "content": "We conducted a user study to validate that our method outperforms existing approaches. For video generation, we" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.289 + ], + "angle": 0, + "content": "randomly selected 10 examples from the test set in Sec. 4.3. In the questionnaire, participants were shown the results of different methods in random order. 20 participants ranked the methods results by three criteria: Sketch Fidelity, Temporal Consistency, and Video Realism. The rankings provided by the participants were used as scores. As shown in Table 1, our method outperforms existing methods in all criteria, demonstrating its superior performance. For sketch-based editing, we used the same 10 examples from the test set in Sec. 4.3. The same participants ranked the methods based on three criteria: Sketch Fidelity, Unedited Region Preservation, and Video Realism. As shown in Table 2, our method also outperforms existing approaches in editing." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.311, + 0.762, + 0.327 + ], + "angle": 0, + "content": "5. Conclusion and Discussion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.336, + 0.907, + 0.532 + ], + "angle": 0, + "content": "We presented SketchVideo, a unified method for sketch-based video generation and editing. For generation, we proposed a sketch condition network that predicts residual features for skipped DiT blocks of the base models to save the memory and achieve effective control. An inter-frame attention is further proposed to propagate the keyframe sketches, achieving interesting motion interpolation or extrapolation. We also introduce a hybrid image and video training strategy. For editing, we incorporate a video insertion module to ensure the newly generated content is spatially and temporally coherent with the original video. During inference, a latent fusion approach preserves unedited regions accurately." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.533, + 0.909, + 0.728 + ], + "angle": 0, + "content": "Limitation and Future Work. While our method generates high-quality videos, its capabilities are still limited by the pretrained text-to-video model. Enhancing performance with more powerful pretrained models and generating long videos instead of video clips is a potential avenue. Additionally, similar to image generation, our method struggles with too complex scenarios, such as human hands and interaction between multiple objects. Incorporating 3D priors [59, 63, 86] like SMPL-X [47] helps address these issues in human scenarios. Moreover, while our method focuses on geometry control, exploring appearance customization through tools like color strokes presents an intriguing area for future research." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.752, + 0.693, + 0.768 + ], + "angle": 0, + "content": "6. Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.775, + 0.907, + 0.889 + ], + "angle": 0, + "content": "This work was sponsored by CCF-Kuaishou Large Model Explorer Fund, Beijing Municipal Science and Technology Commission (No. Z231100005923031), Innovation Funding of ICT, CAS (No. E461020) and National Natural Science Foundation of China (No. 62322210). The authors would like to acknowledge the Nanjing Institute of InforSuperBahn OneAiNexus for providing the training and evaluation platform." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "23386" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.17 + ], + "angle": 0, + "content": "[1] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In Int. Conf. Comput. Vis., pages 1708-1718, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.173, + 0.484, + 0.255 + ], + "angle": 0, + "content": "[2] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, Varun Jampani, and Robin Rombach. Stable video diffusion: Scaling latent video diffusion models to large datasets. CoRR, abs/2311.15127, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.258, + 0.483, + 0.327 + ], + "angle": 0, + "content": "[3] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In IEEE Conf. Comput. Vis. Pattern Recog., pages 22563-22575, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.329, + 0.483, + 0.384 + ], + "angle": 0, + "content": "[4] Minwoo Byeon, Beomhee Park, Haecheon Kim, Sungjun Lee, Woonhyuk Baek, and Saehoon Kim. Coyo-700m: Image-text pair dataset. https://github.com/kakaobrain/coyo-dataset, 2022.1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.387, + 0.482, + 0.428 + ], + "angle": 0, + "content": "[5] Duygu Ceylan, Chun-Hao Paul Huang, and Niloy J. Mitra. Pix2video: Video editing using image diffusion. In Int. Conf. Comput. Vis., pages 23149-23160, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.431, + 0.482, + 0.483 + ], + "angle": 0, + "content": "[6] Wenhao Chai, Xun Guo, Gaoang Wang, and Yan Lu. Stable-video: Text-driven consistency-aware diffusion video editing. In Int. Conf. Comput. Vis., pages 22983-22993, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.488, + 0.482, + 0.541 + ], + "angle": 0, + "content": "[7] Caroline Chan, Frédo Durand, and Phillip Isola. Learning to generate line drawings that convey geometry and semantics. In IEEE Conf. Comput. Vis. Pattern Recog., pages 7915-7925, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.545, + 0.482, + 0.613 + ], + "angle": 0, + "content": "[8] Hila Chefer, Shiran Zada, Roni Paiss, Ariel Ephrat, Omer Tov, Michael Rubinstein, Lior Wolf, Tali Dekel, Tomer Michaeli, and Inbar Mosseri. Still-moving: Customized video generation without customized video data. ACM Trans. Graph., 43(6):244:1-244:11, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.617, + 0.482, + 0.684 + ], + "angle": 0, + "content": "[9] Haoxin Chen, Menghan Xia, Yingqing He, Yong Zhang, Xiaodong Cun, Shaoshu Yang, Jinbo Xing, Yaofang Liu, Qifeng Chen, Xintao Wang, Chao Weng, and Ying Shan. Videocrafter1: Open diffusion models for high-quality video generation. CoRR, abs/2310.19512, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.687, + 0.482, + 0.755 + ], + "angle": 0, + "content": "[10] Haoxin Chen, Yong Zhang, Xiaodong Cun, Menghan Xia, Xintao Wang, Chao Weng, and Ying Shan. Videographer2: Overcoming data limitations for high-quality video diffusion models. In IEEE Conf. Comput. Vis. Pattern Recog., pages 7310-7320, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.758, + 0.482, + 0.813 + ], + "angle": 0, + "content": "[11] Junsong Chen, Yue Wu, Simian Luo, Enze Xie, Sayak Paul, Ping Luo, Hang Zhao, and Zhenguo Li. Pixart-δ: Fast and controllable image generation with latent consistency models. CoRR, abs/2401.05252, 2024. 2, 3, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.816, + 0.482, + 0.884 + ], + "angle": 0, + "content": "[12] Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Zhongdao Wang, James T. Kwok, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart-α: Fast training of diffusion transformer for photorealistic text-to-image synthesis. In Int. Conf. Learn. Represent., 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.887, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[13] Shu-Yu Chen, Wanchao Su, Lin Gao, Shihong Xia, and" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "Hongbo Fu. Deepfacedrawing: deep generation of face images from sketches. ACM Trans. Graph., 39(4):72, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.188 + ], + "angle": 0, + "content": "[14] Shu-Yu Chen, Feng-Lin Liu, Yu-Kun Lai, Paul L. Rosin, Chunpeng Li, Hongbo Fu, and Lin Gao. Deepfaceediting: deep face generation and editing with disentangled geometry and appearance control. ACM Trans. Graph., 40(4):90:1-90:15, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.191, + 0.905, + 0.244 + ], + "angle": 0, + "content": "[15] Weifeng Chen, Jie Wu, Pan Xie, Hefeng Wu, Jiashi Li, Xin Xia, Xuefeng Xiao, and Liang Lin. Control-a-video: Controllable text-to-video generation with diffusion models, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.247, + 0.905, + 0.315 + ], + "angle": 0, + "content": "[16] Xinyuan Chen, Yaohui Wang, Lingjun Zhang, Shaobin Zhuang, Xin Ma, Jiashuo Yu, Yali Wang, Dahua Lin, Yu Qiao, and Ziwei Liu. Seine: Short-to-long video diffusion model for generative transition and prediction. In Int. Conf. Learn. Represent., 2023. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.317, + 0.904, + 0.356 + ], + "angle": 0, + "content": "[17] Jiaxin Cheng, Tianjun Xiao, and Tong He. Consistent video-to-video transfer using synthetic dataset. In Int. Conf. Learn. Represent., 2024. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.358, + 0.905, + 0.412 + ], + "angle": 0, + "content": "[18] Ernie Chu, Shuo-Yen Lin, and Jun-Cheng Chen. Video controlnet: Towards temporally consistent synthetic-to-real video translation using conditional image diffusion models. CoRR, abs/2305.19193, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.414, + 0.905, + 0.48 + ], + "angle": 0, + "content": "[19] Nathaniel Cohen, Vladimir Kulikov, Matan Kleiner, Inbar Huberman-Spiegelglas, and Tomer Michaeli. Slicedit: Zero-shot video editing with text-to-image diffusion models using spatio-temporal slices. arXiv preprint arXiv:2405.12211, 2024.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.483, + 0.905, + 0.524 + ], + "angle": 0, + "content": "[20] Michal Geyer, Omer Bar-Tal, Shai Bagon, and Tali Dekel. Tokenflow: Consistent diffusion features for consistent video editing. In Int. Conf. Learn. Represent., 2024. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.526, + 0.905, + 0.579 + ], + "angle": 0, + "content": "[21] Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron C. Courville, and Yoshua Bengio. Generative adversarial networks. CoRR, abs/1406.2661, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.581, + 0.905, + 0.635 + ], + "angle": 0, + "content": "[22] Yuwei Guo, Ceyuan Yang, Anyi Rao, Maneesh Agrawala, Dahua Lin, and Bo Dai. Sparsectrl: Adding sparse controls to text-to-video diffusion models. In Eur. Conf. Comput. Vis., pages 330-348, 2024. 2, 3, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.637, + 0.905, + 0.704 + ], + "angle": 0, + "content": "[23] Yuwei Guo, Ceyuan Yang, Anyi Rao, Zhengyang Liang, Yaohui Wang, Yu Qiao, Maneesh Agrawala, Dahua Lin, and Bo Dai. Animatediff: Imagine your personalized text-to-image diffusion models without specific tuning. In Int. Conf. Learn. Represent., 2024. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.707, + 0.905, + 0.76 + ], + "angle": 0, + "content": "[24] Hao He, Yinghao Xu, Yuwei Guo, Gordon Wetzstein, Bo Dai, Hongsheng Li, and Ceyuan Yang. Cameractrl: Enabling camera control for text-to-video generation. CoRR, abs/2404.02101, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.762, + 0.905, + 0.815 + ], + "angle": 0, + "content": "[25] Yingqing He, Tianyu Yang, Yong Zhang, Ying Shan, and Qifeng Chen. Latent video diffusion models for high-fidelity video generation with arbitrary lengths. CoRR, abs/2211.13221, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.818, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[26] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. Adv. Neural Inform. Process. Syst., 35:8633-8646, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[27] Wenyi Hong, Ming Ding, Wendi Zheng, Xinghan Liu, and Jie Tang. Cogvideo: Large-scale pretraining for" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23387" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.093, + 0.483, + 0.12 + ], + "angle": 0, + "content": "text-to-video generation via transformers. arXiv preprint arXiv:2205.15868, 2022. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.123, + 0.483, + 0.164 + ], + "angle": 0, + "content": "[28] Zhitong Huang, Mohan Zhang, and Jing Liao. LVCD: reference-based linear art video colorization with diffusion models. CoRR, abs/2409.12960, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.166, + 0.482, + 0.222 + ], + "angle": 0, + "content": "[29] Ondrej Jamriska, Sárka Sochorová, Ondrej Texler, Michal Lukác, Jakub Fiser, Jingwan Lu, Eli Shechtman, and Daniel Sykora. Stylizing video by example. ACM Trans. Graph., 38 (4):107:1-107:11, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.224, + 0.483, + 0.28 + ], + "angle": 0, + "content": "[30] Rui Jiang, Guang-Cong Zheng, Teng Li, Tian-Rui Yang, Jing-Dong Wang, and Xi Li. A survey of multimodal controllable diffusion models. Journal of Computer Science and Technology, 39(3):509-541, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.282, + 0.482, + 0.35 + ], + "angle": 0, + "content": "[31] Yuming Jiang, Tianxing Wu, Shuai Yang, Chenyang Si, Dahua Lin, Yu Qiao, Chen Change Loy, and Ziwei Liu. Videobooth: Diffusion-based video generation with image prompts. In IEEE Conf. Comput. Vis. Pattern Recog., pages 6689-6700, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.354, + 0.482, + 0.395 + ], + "angle": 0, + "content": "[32] Yoni Kasten, Dolev Ofri, Oliver Wang, and Tali Dekel. Layered neural atlases for consistent video editing. ACM Trans. Graph., 40(6):210:1-210:12, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.398, + 0.482, + 0.467 + ], + "angle": 0, + "content": "[33] Levon Khachatryan, Andranik Movsisyan, Vahram Tadevosyan, Roberto Henschel, Zhangyang Wang, Shant Navasardyan, and Humphrey Shi. Text2video-zero: Text-to-image diffusion models are zero-shot video generators. In Int. Conf. Comput. Vis., pages 15908-15918, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.469, + 0.482, + 0.523 + ], + "angle": 0, + "content": "[34] Max Ku, Cong Wei, Weiming Ren, Harry Yang, and Wenhu Chen. Anyv2v: A tuning-free framework for any video-to-video editing tasks. arXiv preprint arXiv:2403.14468, 2024. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.527, + 0.482, + 0.553 + ], + "angle": 0, + "content": "[35] PKU-Yuan Lab and Tuzhan AI etc. Open-sora-plan, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.557, + 0.482, + 0.611 + ], + "angle": 0, + "content": "[36] Yaowei Li, Xintao Wang, Zhaoyang Zhang, Zhouxia Wang, Ziyang Yuan, Liangbin Xie, Yuexian Zou, and Ying Shan. Image conductor: Precision control for interactive video synthesis. CoRR, abs/2406.15339, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.615, + 0.482, + 0.67 + ], + "angle": 0, + "content": "[37] Zhen Li, Zuo-Liang Zhu, Ling-Hao Han, Qibin Hou, Chun-Le Guo, and Ming-Ming Cheng. Amt: All-pairs multi-field transforms for efficient frame interpolation. In IEEE Conf. Comput. Vis. Pattern Recog., 2023. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.673, + 0.482, + 0.728 + ], + "angle": 0, + "content": "[38] Shanchuan Lin, Bingchen Liu, Jiashi Li, and Xiao Yang. Common diffusion noise schedules and sample steps are flawed. In IEEE/CVF Winter Conference on Applications of Computer Vision, WACV, pages 5392-5399, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.731, + 0.482, + 0.785 + ], + "angle": 0, + "content": "[39] Feng-Lin Liu, Shu-Yu Chen, Yu-Kun Lai, Chunpeng Li, Yue-Ren Jiang, Hongbo Fu, and Lin Gao. Deepfacevideoediting: sketch-based deep editing of face videos. ACM Trans. Graph., 41(4):167:1-167:16, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.788, + 0.482, + 0.843 + ], + "angle": 0, + "content": "[40] Chong Mou, Mingdeng Cao, Xintao Wang, Zhaoyang Zhang, Ying Shan, and Jian Zhang. Revideo: Remake a video with motion and content control. CoRR, abs/2405.13865, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[41] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, and Ying Shan. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. In AAAI, pages 4296-4304, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.148 + ], + "angle": 0, + "content": "[42] Kepan Nan, Rui Xie, Penghao Zhou, Tiehan Fan, Zhenheng Yang, Zhijie Chen, Xiang Li, Jian Yang, and Ying Tai. Openvid-1m: A large-scale high-quality dataset for text-to-video generation. CoRR, abs/2407.02371, 2024. 1, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.151, + 0.905, + 0.175 + ], + "angle": 0, + "content": "[43] OpenAI. Sora overview:https://openai.com/index/soral/, 2024.1,3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.905, + 0.248 + ], + "angle": 0, + "content": "[44] Hao Ouyang, Qiuyu Wang, Yuxi Xiao, Qingyan Bai, Juntao Zhang, Kecheng Zheng, Xiaowei Zhou, Qifeng Chen, and Qifeng Chen. Codef: Content deformation fields for temporally consistent video processing. In IEEE Conf. Comput. Vis. Pattern Recog., pages 8089-8099, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.25, + 0.905, + 0.303 + ], + "angle": 0, + "content": "[45] Wenqi Ouyang, Yi Dong, Lei Yang, Jianlou Si, and Xingang Pan. I2vedit: First-frame-guided video editing via image-to-video diffusion models. CoRR, abs/2405.16537, 2024. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.307, + 0.905, + 0.347 + ], + "angle": 0, + "content": "[46] Gaurav Parmar, Taesung Park, Srinivasa Narasimhan, and Jun-Yan Zhu. One-step image translation with text-to-image models. CoRR, abs/2403.12036, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.35, + 0.905, + 0.418 + ], + "angle": 0, + "content": "[47] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed A. A. Osman, Dimitrios Tzionas, and Michael J. Black. Expressive body capture: 3d hands, face, and body from a single image. In IEEE Conf. Comput. Vis. Pattern Recog., 2019. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.421, + 0.905, + 0.474 + ], + "angle": 0, + "content": "[48] Bohao Peng, Jian Wang, Yuechen Zhang, Wenbo Li, Ming-Chang Yang, and Jiaya Jia. Controlnext: Powerful and efficient control for image and video generation. CoRR, abs/2408.06070, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.477, + 0.905, + 0.559 + ], + "angle": 0, + "content": "[49] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.562, + 0.905, + 0.629 + ], + "angle": 0, + "content": "[50] Elad Richardson, Yuval Alaluf, Or Patashnik, Yotam Nitzan, Yaniv Azar, Stav Shapiro, and Daniel Cohen-Or. Encoding in style: A stylegan encoder for image-to-image translation. In IEEE Conf. Comput. Vis. Pattern Recog., pages 2287–2296, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.633, + 0.905, + 0.687 + ], + "angle": 0, + "content": "[51] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In IEEE Conf. Comput. Vis. Pattern Recog., pages 10674-10685, 2022. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.689, + 0.905, + 0.731 + ], + "angle": 0, + "content": "[52] Manuel Ruder, Alexey Dosovitskiy, and Thomas Brox. Artistic style transfer for videos and spherical images. Int. J. Comput. Vis., 126(11):1199-1219, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.733, + 0.905, + 0.828 + ], + "angle": 0, + "content": "[53] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L. Denton, Seyed Kamyar Seyed Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, Jonathan Ho, David J. Fleet, and Mohammad Norouzi. Photorealistic text-to-image diffusion models with deep language understanding. In Adv. Neural Inform. Process. Syst., 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.831, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[54] Tim Salimans and Jonathan Ho. Progressive distillation for fast sampling of diffusion models. In Int. Conf. Learn. Represent., 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[55] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23388" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.483, + 0.175 + ], + "angle": 0, + "content": "Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, Patrick Schramowski, Srivatsa Kundurthy, Katherine Crowson, Ludwig Schmidt, Robert Kaczmarczyk, and Jenia Jitsev. LAION-5B: an open large-scale dataset for training next generation image-text models. In Adv. Neural Inform. Process. Syst., 2022. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.177, + 0.483, + 0.259 + ], + "angle": 0, + "content": "[56] Xiaoyu Shi, Zhaoyang Huang, Fu-Yun Wang, Weikang Bian, Dasong Li, Yi Zhang, Manyuan Zhang, Ka Chun Cheung, Simon See, Hongwei Qin, Jifeng Dai, and Hongsheng Li. Motion-i2v: Consistent and controllable image-to-video generation with explicit motion modeling. In ACM SIGGRAPH, page 111. ACM, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.261, + 0.482, + 0.314 + ], + "angle": 0, + "content": "[57] Uriel Singer, Amit Zohar, Yuval Kirstain, Shelly Sheynin, Adam Polyak, Devi Parikh, and Yaniv Taigman. Video editing via factorized diffusion distillation. CoRR, abs/2403.09334, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.317, + 0.483, + 0.357 + ], + "angle": 0, + "content": "[58] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In Int. Conf. Learn. Represent., 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.359, + 0.483, + 0.399 + ], + "angle": 0, + "content": "[59] Jia-Mu Sun, Tong Wu, and Lin Gao. Recent advances in implicit representation-based 3d shape generation. Vis. Intell., 2(1), 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.401, + 0.483, + 0.44 + ], + "angle": 0, + "content": "[60] Vchitect Team and Shanghai Artificial Intelligence Laboratory. Vchitect-2.0: Parallel transformer for scaling up video diffusion models, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.442, + 0.483, + 0.496 + ], + "angle": 0, + "content": "[61] Zheng Tianpeng, Chen Yanxiang, Wen Xinzhe, Li Yancheng, and Wang Zhiyuan. Research on diffusion model generated video datasets and detection benchmarks. Journal of Image and Graphics, pages 1-13, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.498, + 0.483, + 0.552 + ], + "angle": 0, + "content": "[62] Rotem Tzaban, Ron Mokady, Rinon Gal, Amit Bermano, and Daniel Cohen-Or. *Stitch it in time: Gan-based facial editing of real videos*. In SIGGRAPH Asia Conference Papers, pages 29:1–29:9. ACM, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.554, + 0.483, + 0.608 + ], + "angle": 0, + "content": "[63] Guangming Wang, Lei Pan, Songyou Peng, Shaohui Liu, Chenfeng Xu, Yanzi Miao, Wei Zhan, Masayoshi Tomizuka, Marc Pollefeys, and Hesheng Wang. Nerf in robotics: A survey. CoRR, abs/2405.01333, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.61, + 0.483, + 0.65 + ], + "angle": 0, + "content": "[64] Jiuniu Wang, Hangjie Yuan, Dayou Chen, Yingya Zhang, Xiang Wang, and Shiwei Zhang. Modelscope text-to-video technical report. CoRR, abs/2308.06571, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.652, + 0.483, + 0.718 + ], + "angle": 0, + "content": "[65] Ting-Chun Wang, Ming-Yu Liu, Jun-Yan Zhu, Andrew Tao, Jan Kautz, and Bryan Catanzaro. High-resolution image synthesis and semantic manipulation with conditional gans. In IEEE Conf. Comput. Vis. Pattern Recog., pages 8798-8807, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.721, + 0.483, + 0.776 + ], + "angle": 0, + "content": "[66] Zhouxia Wang, Ziyang Yuan, Xintao Wang, Yaowei Li, Tianshui Chen, Menghan Xia, Ping Luo, and Ying Shan. Motionctrl: A unified and flexible motion controller for video generation. In ACM SIGGRAPH, page 114, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.778, + 0.483, + 0.83 + ], + "angle": 0, + "content": "[67] Jianzong Wu, Xiangtai Li, Yanhong Zeng, Jiangning Zhang, Qianyu Zhou, Yining Li, Yunhai Tong, and Kai Chen. Motionbooth: Motion-aware customized text-to-video generation. CoRR, abs/2406.17758, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.483, + 0.9 + ], + "angle": 0, + "content": "[68] Tao Wu, Yong Zhang, Xintao Wang, Xianpan Zhou, Guangcong Zheng, Zhongang Qi, Ying Shan, and Xi Li. Customcrafter: Customized video generation with preserving motion and concept composition abilities. CoRR, abs/2408.13239, 2024. 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.133 + ], + "angle": 0, + "content": "[69] Saining Xie and Zhuowen Tu. Holistically-nested edge detection. In Int. Conf. Comput. Vis., pages 1395-1403, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[70] Jinbo Xing, Hanyuan Liu, Menghan Xia, Yong Zhang, Xintao Wang, Ying Shan, and Tien-Tsin Wong. Tooncrafter: Generative cartoon interpolation. CoRR, abs/2405.17933, 2024. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.193, + 0.905, + 0.247 + ], + "angle": 0, + "content": "[71] Shuai Yang, Yifan Zhou, Ziwei Liu, and Chen Change Loy. Rerender A video: Zero-shot text-guided video-to-video translation. In SIGGRAPH Asia Conference Papers, pages 95:1-95:11. ACM, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.25, + 0.905, + 0.318 + ], + "angle": 0, + "content": "[72] Shiuyuan Yang, Liang Hou, Haibin Huang, Chongyang Ma, Pengfei Wan, Di Zhang, Xiaodong Chen, and Jing Liao. Direct-a-video: Customized video generation with user-directed camera movement and object motion. In ACM SIGGRAPH, page 113, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.321, + 0.905, + 0.388 + ], + "angle": 0, + "content": "[73] Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, et al. Cogvideox: Text-to-video diffusion models with an expert transformer. arXiv preprint arXiv:2408.06072, 2024. 1, 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.391, + 0.905, + 0.445 + ], + "angle": 0, + "content": "[74] Danah Yatim, Rafail Fridman, Omer Bar-Tal, Yoni Kasten, and Tali Dekel. Space-time diffusion features for zero-shot text-driven motion transfer. In IEEE Conf. Comput. Vis. Pattern Recog., pages 8466-8476, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.448, + 0.907, + 0.489 + ], + "angle": 0, + "content": "[75] Chen Yuan, Zhao Yang, Zhang Xiaojuan, and Liu Xiaoping. Sketch colorization with finite color space prior. Journal of Image and Graphics, 29(4):978-988, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.491, + 0.905, + 0.532 + ], + "angle": 0, + "content": "[76] Liang Yuan, Dingkun Yan, Suguru Saito, and Issei Fujishiro. Diffmat: Latent diffusion models for image-guided material generation. Visual Informatics, 8(1):6-14, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.534, + 0.905, + 0.575 + ], + "angle": 0, + "content": "[77] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Int. Conf. Comput. Vis., pages 3813-3824, 2023. 2, 3, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.577, + 0.905, + 0.631 + ], + "angle": 0, + "content": "[78] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In IEEE Conf. Comput. Vis. Pattern Recog., pages 586-595, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.633, + 0.905, + 0.687 + ], + "angle": 0, + "content": "[79] Yuxin Zhang, Fan Tang, Nisha Huang, Haibin Huang, Chongyang Ma, Weiming Dong, and Changsheng Xu. Motioncrafter: One-shot motion customization of diffusion models. CoRR, abs/2312.05288, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.69, + 0.905, + 0.745 + ], + "angle": 0, + "content": "[80] Yabo Zhang, Yuxiang Wei, Dongsheng Jiang, Xiaopeng Zhang, Wangmeng Zuo, and Qi Tian. Controlvideo: Training-free controllable text-to-video generation. In Int. Conf. Learn. Represent., 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.747, + 0.905, + 0.8 + ], + "angle": 0, + "content": "[81] Zhenghao Zhang, Junchao Liao, Menghao Li, Long Qin, and Weizhi Wang. Tora: Trajectory-oriented diffusion transformer for video generation. CoRR, abs/2407.21705, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.803, + 0.905, + 0.872 + ], + "angle": 0, + "content": "[82] Zhixing Zhang, Bichen Wu, Xiaoyan Wang, Yaqiao Luo, Luxin Zhang, Yinan Zhao, Peter Vajda, Dimitris N. Metaxas, and Licheng Yu. AVID: any-length video inpainting with diffusion model. In IEEE Conf. Comput. Vis. Pattern Recog., pages 7162-7172, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[83] Shihao Zhao, Dongdong Chen, Yen-Chun Chen, Jianmin Bao, Shaozhe Hao, Lu Yuan, and Kwan-Yee K. Wong." + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23389" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.12 + ], + "angle": 0, + "content": "Uni-controlnet: All-in-one control to text-to-image diffusion models. In Adv. Neural Inform. Process. Syst., 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.177 + ], + "angle": 0, + "content": "[84] Zangwei Zheng, Xiangyu Peng, Tianji Yang, Chenhui Shen, Shenggui Li, Hongxin Liu, Yukun Zhou, Tianyi Li, and Yang You. Open-sora: Democratizing efficient video production for all, 2024. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.179, + 0.482, + 0.233 + ], + "angle": 0, + "content": "[85] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A. Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In Int. Conf. Comput. Vis., pages 2242-2251, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.235, + 0.482, + 0.275 + ], + "angle": 0, + "content": "[86] Siting Zhu, Guangming Wang, Dezhi Kong, and Hesheng Wang. 3d gaussian splatting in robotics: A survey. CoRR, abs/2410.12262, 2024. 8" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.275 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23390" + } + ] +] \ No newline at end of file diff --git a/2025/SketchVideo_ Sketch-based Video Generation and Editing/86ad126c-969d-47b0-8162-a7cc53739116_origin.pdf b/2025/SketchVideo_ Sketch-based Video Generation and Editing/86ad126c-969d-47b0-8162-a7cc53739116_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..dff40b080f5513829be2a0edd08036ea9d3fb909 --- /dev/null +++ b/2025/SketchVideo_ Sketch-based Video Generation and Editing/86ad126c-969d-47b0-8162-a7cc53739116_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e3519377168b273d3d235037772d4e85ed82c50a2e0a476df8affa4e3d26cb2a +size 8110498 diff --git a/2025/SketchVideo_ Sketch-based Video Generation and Editing/full.md b/2025/SketchVideo_ Sketch-based Video Generation and Editing/full.md new file mode 100644 index 0000000000000000000000000000000000000000..662445c9745016c85ed908a5f0b7003e661e7986 --- /dev/null +++ b/2025/SketchVideo_ Sketch-based Video Generation and Editing/full.md @@ -0,0 +1,397 @@ +# SketchVideo: Sketch-based Video Generation and Editing + +Feng-Lin Liu $^{1,2}$ Hongbo Fu $^{3}$ Xintao Wang $^{4}$ Weicai Ye $^{4}$ Pengfei Wan $^{4}$ Di Zhang $^{4}$ Lin Gao $^{1,2*}$ + +$^{1}$ Beijing Key Laboratory of Mobile Computing and Pervasive Device, Institute of Computing Technology, Chinese Academy of Sciences $^{2}$ University of Chinese Academy of Sciences $^{3}$ Hong Kong University of Science and Technology $^{4}$ Kuaishou Technology + +![](images/caeda09bc9fe18f0c8cf1cb0c674b0ab9ffcda319bd8f58736f66b52e41aed53.jpg) +"A close-up of a majestic male lion's face, with a tawny mane, piercing amber eyes, and subtle scars, bathed in golden sunlight, showcasing his power and wild spirit." + +![](images/5e578ebc2b971b8498cc9bfac9815da0080c43814477fcc68cdc5b4ea6de9faf.jpg) +"The video is a drone shot of a serene landscape featuring a deep blue lagoon surrounded by lush green trees and rocky cliffs. The lagoon is nestled between the mountains, creating a picturesque scene." + +(a) Sketch-based Video Generation +![](images/0b935ee7d70550e86839543a0c26f29b2fa26098d49d92ad7f71b429971a222d.jpg) +"A clear blue sky above the mountains, with a few clouds scattered across it. A deep blue lake nestled between the mountains. The mountains have patches of snow that still cling to their peaks." + +![](images/23cd0f2c4a391929d8ccaa4b1c3959237188627b1d775931462fec892c55b843.jpg) +(b) Sketch-based Video Editing +Figure 1. Our method enables high-quality video generation (a) and editing (b) based on sketch and text inputs. (a) Top: With the same text prompt, different keyframe sketches lead to results with similar semantics but diverse sketch-faithful geometry. (a) Bottom: With the same sketches, varied text prompts yield diverse appearances. (b) Users can also edit real videos by drawing on keyframe sketches, with edits automatically propagated even when edited objects in original videos have translation and rotation. + +# Abstract + +Video generation and editing conditioned on text prompts or images have undergone significant advancements. However, challenges remain in accurately controlling global layout and geometry details solely by texts, and supporting motion control and local modification through images. In this paper, we aim to achieve sketch-based spatial and motion control for video generation and support fine-grained editing of real or synthetic videos. Based on the DiT video generation model, we propose a memory-efficient control structure with sketch control blocks that predict residual features of skipped DiT blocks. Sketches are drawn on one or two keyframes (at arbitrary time points) for easy interaction. To propagate such temporally sparse sketch conditions across all frames, we propose an inter-frame at + +tention mechanism to analyze the relationship between the keyframes and each video frame. For sketch-based video editing, we design an additional video insertion module that maintains consistency between the newly edited content and the original video's spatial feature and dynamic motion. During inference, we use latent fusion for the accurate preservation of unedited regions. Extensive experiments demonstrate that our SketchVideo achieves superior performance in controllable video generation and editing. + +# 1. Introduction + +Diffusion-based text-to-image [12, 51, 53] and text-to-video [27, 35, 43, 73, 84] models advance significantly due to improvements in datasets [4, 42, 55, 61] and denoising network architectures [12, 51]. While text prompts effectively describe high-level semantics, they lack control of scene + +Layouts and geometric details. To address this, existing video generation methods [2, 25, 73] utilize images as additional conditions but raise the questions of how to generate input images and achieve detailed editing. Sketching serves as a user-friendly interaction tool to capture spatial content and shape details accurately. One or two sketches are already sufficient to convey desired scene structures and motion information for short video clips (around 6 seconds), which are the target of our and most existing video generation methods. However, using such sparse keyframe sketches presents several challenges, including reasonably completing the missing frames, improving memory efficiency, and addressing the limited size of video datasets. + +A naive solution is to translate the input keyframe sketches into images and then utilize interpolation methods [16, 37, 70] for video generation. However, it is nontrivial to ensure consistency during keyframe sketch-to-image generation, which significantly affects the video quality. This approach also struggles to generate extrapolation frames when applying conditions in intermediate frames rather than beginning and ending time points. Another possible approach is to utilize white placeholders to fill missing condition frames and directly apply ControlNet [77] into video models, similar to SparseCtrl [22]. However, this requires the same network to process both sketches and white placeholders simultaneously, while the pretrained blocks handle tasks far from this sparse propagation. Additionally, for DiT-based video frameworks [27, 84], the traditional strategy [11] that copies half of the pretrained model as a condition network easily causes the out-of-memory issue. + +To address these issues, we propose a novel sketch condition network specifically designed for the DiT-based video generation architecture (CogVideoX [73] in our work). Following ControlNet [11, 77], we employ a trainable copy of CogVideoX's DiT block to process only the sketch inputs and generate control features. No white placeholder is processed to align with the pretrained weights and reduce learning complexity. To propagate these keyframe features, we design an inter-frame attention mechanism that captures the relationship between the control keyframes and all video frames. Our approach computes query and key features from noisy latent while extracting value features from sketch conditions. This design leverages frame-to-frame similarity for control propagation. The above components consist of a single sketch control block. Instead of copying a half number of the pretrained blocks to construct the control network [11], we use only 5 sketch control blocks out of the 30 DiT blocks available in CogVideoX-2b. We design a novel uniformly distributed skip structure to add the control signals to different levels of features in discrete blocks (0, 6, 12, 18, and 24), achieving effective spatial control while improving memory efficiency. During + +training, an external image dataset is incorporated to solve the challenge of limited video data. + +Beyond generation, interactive editing of real or synthetic videos further enhances creative flexibility. Existing methods [34, 40, 45, 71] achieve interesting text-based editing or effectively propagate single image editing into videos. Despite their effectiveness in appearance modification, they struggle with shape manipulation and object insertion, as they preserve the original temporal motion. Such information is missing for newly introduced content. Moreover, precisely identifying and preserving unedited regions for localized editing remains a challenge. + +We propose a sketch-based editing method for detailed local modification. Rather than editing a single image and propagating changes, we directly construct an editing network based on our sketch control network. To analyze the relationship between edited regions and the original video, we incorporate a video insertion module that takes the original video with masked edited regions as inputs. The modified sketch control blocks generate residual features that capture temporally and spatially coherent contents in the edited areas. To accurately preserve unedited regions and achieve seamless fusion, the newly edited regions are blended with the original video in the latent space. + +Extensive experiments demonstrate that our method outperforms existing approaches in video generation and editing. Our contributions are summarized as follows: 1) We propose SketchVideo, a novel sketch-based video generation and editing framework that enables detailed geometry control and manipulation using keyframe sketches, as shown in Fig. 1. 2) We design a sketch condition network that predicts skipped control features for the DiT framework, with an inter-frame attention mechanism to propagate one or two sketch conditions across the video. 3) We propose a video insertion module that analyzes the relationship between drawn sketches and original videos, utilizing a latent fusion strategy to preserve unedited regions accurately. + +# 2. Related Work + +Sketch-based Image Generation. GAN-based methods [21] have achieved great success in category-restricted sketch-to-image translation [13, 14, 50, 65, 75, 85]. Recently, diffusion-based text-to-image models [51, 53] handle general categories with conditional models like ControlNet [77], T2I-Adapter [41] and further advances [30, 46, 48, 76, 83]. Beyond U-Net, the DiT backbone [12] enables image generation, with pixART- $\delta$ [11] utilizing the first half of the pretrained model's blocks as a trainable network to predict corresponding control residual features. Video generation, however, introduces additional challenges. For ease of interaction, we expect sketches specified only for a sparse set of keyframes, making it difficult to generate frames without sketch inputs. Additionally, video genera + +tion costs significantly higher memory resources, making methods [11, 77] that replicate half of the base model as a sketch encoder easily out of memory. + +Diffusion-based Video Generation. VDM [26] pioneered diffusion-based video generation with a 3D U-Net denoising network. To improve quality, subsequent works [2, 3, 9, 10, 23, 25, 33, 64] integrated temporal modules into text-to-image models [51] to enable text- and image-conditioned video synthesis. Considering the efficiency, the DiT architecture is further used in Sora [43] and open-source projects [35, 60, 84]. Despite these advancements, subtle flickering artifacts remain in the results. CogVideoX [27, 73] further proposes a 3D full attention that merges the spatial and temporal attention, facilitating the generation of long-duration and high-resolution videos. + +Building on existing video generation models, various conditions have been introduced to control the generation, such as camera movement [24, 66, 72], subject identity [31, 68], key-point trajectory [36, 56, 81] and example motions [67, 79]. However, they often lack control over spatial layouts and geometric details. Some methods [15, 18, 80] address this by extending image-based ControlNet [77] to videos but require all-frame conditions that are tedious for sketch interaction. SparseCtrl [22] tackles this by using white images for completions. However, due to the restriction of its base model [23] and a simple network design, its results suffer from temporal flickering. Similar ideas have been applied to cartoon interpolation [70] and colorization [28] with line art as input, but their outputs are limited to cartoon style. Our method utilizes sparse inputs, including hand-drawn sketches on one or two keyframes, to generate temporally stable and realistic videos. Additionally, our method further supports the sketch-based detailed editing of existing videos. + +Deep Learning-based Video Editing. Pioneer works achieve video editing by style transfer [29, 52], GAN inversion [39, 62], and layered representations [32]. The advent of diffusion models further provides new editing paradigms. One category of such methods leverages image generation models to achieve compelling editing results, using temporal consistency techniques such as layered representations [6, 44], cross-frame attention [5, 17, 19, 57] and pixel warping [71]. A second category of methods employ video generation models. These works propagate the edits applied on the first frame into the other frames [34, 45], or utilize an inpainting strategy to achieve text-based editing [82] and motion modification [40]. To achieve text-based large-scale shape modification while maintaining motion features, a space-time feature loss [74] is designed to guide the inference process. Our method moves beyond traditional text- or image-based editing approaches, allowing users to draw one or two keyframe sketches at arbitrary time points for interactive video editing. Additionally, our method effectively + +handles sketch-based shape manipulation and dynamic object insertion, which are challenging for previous works. + +Controllable Attention Mechanism. Attention across frames is initially used to ensure temporal consistency inAnimateDiff [23] and subsequent video generation models. Subsequent works, such as VideoBooth [31] and Still-Moving [8], utilize it to learn identity-aware features for customized video generation. For video editing, existing works [17, 20, 57] utilize cross-frame attention to capture the temporal motion of input videos, enabling effective propagation of editing operations. Our method applies this idea to pixel-aligned sketch-based video generation and propagates the spatial geometric conditions instead of identity customization. We use a new feature derivation strategy for enhanced spatial control, differentiating our approach from traditional cross-attention mechanisms. + +# 3. Methodology + +This section introduces our sketch-based video generation and editing framework. In Sec.3.1, we provide an overview of CogVideoX-2b [73], a pretrained text-to-video generation method, which we will use for sketch-based video generation in our work. In Sec.3.2, we describe our sketch condition network specifically designed for the DiT architecture, which contains sketch control blocks to predict residual features. Within each control block, an interframe attention mechanism is designed to propagate the input sketches. In Sec.3.3, we detail the editing framework, which incorporates a video insertion module and latent fusion to preserve the features of the original video. + +# 3.1. Preliminary + +CogVideoX [73] is a text-to-video generation model that employs a 3D causal VAE and builds diffusion within the latent space. In CogVideoX-2b, the VAE model performs $8 \times 8$ spatial and $4 \times$ temporal downsampling, followed by a DiT architecture with 30 blocks for video generation. Within each block, a 3D full attention processes the concatenated text embeddings and patchified video latents, followed by a feed-forward layer to output the features. The 3D full attention merges the commonly used separate spatial and temporal attention [2, 23, 84] to improve the temporal coherence. The training objective of diffusion is: + +$$ +L (\theta) := \mathrm {E} _ {t, x _ {0} ^ {1: N}, y, \epsilon} \left\| \epsilon - \epsilon_ {\theta} \left(\sqrt {\bar {\alpha} _ {t}} x _ {0} ^ {1: N} + \sqrt {1 - \bar {\alpha} _ {t}} \epsilon , t, y\right) \right\| ^ {2}, \quad (1) \text {w h e r e} t \text {i s s a m p l e d b e t w e e n 1 a n d T (d e n o s i n g s t e p s)}, \epsilon \text {i s t h e r a n d o n n o i s e ,} y \text {i s t e x t p r o m p t s , a n d} x _ {0} ^ {1: N} \text {i s t h e v i d e o d a w i t h} N \text {f r a m e s .} \text {T h e v - p r e d i c t i o n [ 5 4 ] a n d z e r o S N R [ 3 8 ] a r e u l i z a t e d f o r t h e d i f f u s i o n s e t t i n g .} +$$ + +# 3.2. Sketch-based Video Generation + +Given text prompts and one or two keyframe sketches with corresponding time points $t_1, t_2$ , our method generates + +![](images/fa82a91a15357dd7e658f31b9edbfd2d35d9cecf979b82c59a23a8b8a40a11ed.jpg) +(a) Sketch Condition Network + +![](images/ae02777334cbdd69680ec09f99ecb89bab661376ac16380c8eb6caef18631e4e.jpg) +(b) Control Block for Generation +(c) Control Block for Editing +Figure 2. Our framework for sketch-based video generation and editing. (a) Our sketch condition network for the DiT-based video generation architecture has a skip structure and five sketch control blocks that predict residual features. (b) For generation, features are extracted from temporally sparse input sketches and propagated through inter-frame attention. The input sketches are provided for one or two keyframes (the second sketch is shown as a dotted line). In the top left corner of (b), the prompt and timestep inputs are shown. (c) For editing, the same sketch control block (b) is utilized, with an additional video insertion module and video masks $M^{1:N}$ to analyze the relationship between edited and unedited regions. The 3D causal VAE is omitted to save space. + +video clips that respect the input text prompts and sketches. As shown in Fig. 2, we design a sketch condition network for effective control. The input sketches are encoded into the latent space by the pretrained VAE, followed by patchifying and time-aware positional embedding to generate sketch latents $s_0^{t_1,t_2}$ . + +Skip Residual Structure. In sketch-based image generation, methods like ControlNet [77] and pixART- $\delta$ [11] copy the half of the base model as their sketch encoder to fully utilize the pretrained text-to-image models. However, applying this to video generation, as done in SparseCtrl [22], is memory inefficient because it requires adding half of the base model's parameters. Unlike the U-Net architecture, the DiT network does not have an explicit encoder and decoder. Therefore, the assumption in pixART $\delta$ [11] that the first half of blocks serve as the encoder can be improved. + +Instead of borrowing local consecutive blocks of the base model as the encoder and predicting their residual features, we recognize that blocks at different depths process distinct feature levels, which should be considered for sketch control. As illustrated in Fig. 2 (a), we propose a novel skip residual structure that reduces the number of blocks while enabling effective sketch control and high-quality generation. Our sketch condition network contains 5 sketch control blocks, uniformly distributed across the pretrained generation network to predict residual features for blocks 0, 6, 12, 18, and 24 of the original video generation model. This structure efficiently integrates sketch control informa + +tion into multiple feature levels, enhancing the analysis of input conditions and original semantic features. + +Sketch Control Block. In the $i$ -th sketch control block, the input consists of hidden video features $h_i^{1:N}$ and sketch features $s_{i-1}^{t_1,t_2}$ , and the output is the residual features $\overline{h}_i^{1:N}$ and updated sketch features $s_i^{t_1,t_2}$ . As shown in Fig. 2 (b), the sketch feature $s_{i-1}^{t_1,t_2}$ is processed: + +$$ +s _ {i} ^ {t _ {1}, t _ {2}} = \operatorname {F e e d F o r w a r d} \left(s _ {i - 1} ^ {t _ {1}, t _ {2}}\right), \tag {2} +$$ + +where the output sketch feature $s_i^{t_1,t_2}$ is used for sketch control propagation and as the input to the next sketch control block. + +To propagate the sketch inputs, a direct approach [22] replaces missing sketches with white images and employs a trainable copy of the pretrained DiT block to predict residual features. However, as shown in Fig. 5, this leads to fuzzy details in challenging cases, as the network randomly processes both sketches and white images. This mixed input is far from the pretrained model's task. To address the above issue, we employ a trainable copy of pretrained DiT block to process only the sketch inputs (no white placeholder), aligning with the pretrained weights and reducing learning difficulty. The resulting keyframe sketch features, denoted as $c_{i}^{t_{1},t_{2}}$ , are propagated to all frames through an inter-frame attention approach. + +Inter-frame Attention. We utilize the input hidden features of all frames to calculate $Q$ and the hidden features corresponding to the control frames to calculate $K$ . During attention computation, this approach captures the internal + +![](images/f48de3b7c0011fd39a4dc48df3b075b3a005fb283187c656cf5ffbe6bea17a4f.jpg) +A small, fluffy white rabbit hops across a lush meadow at sunset, pausing to listen to rustling leaves before continuing its joyful dance. + +![](images/80a7d9e39c6a1472df2068c1865e29c9435cabd3e62087239525a0aec3a480de.jpg) +A panoramic video captures a majestic city with towering skyscrapers, quaint red-tiled homes, and a tranquil lake, all under a sky of drifting white clouds. + +![](images/0d5430ee08d9fa5fcbb547558c2d691da520317b3ca6216470bfe2ec90202af3.jpg) +A video captures close-ups of a cat's face with green eyes and a brown-black spotted coat, showing three expressions: alert, relaxed, and distant. + +![](images/a53e56b37cf76b85a6814f637c27a8f2f6672708bc20563a46679693f67b6b8f.jpg) +A video captures the explosive disintegration of a light blue planet, fragments and fiery debris catapulted into starry space, creating a stark contrast between celestial serenity and violent upheaval. + +Figure 3. The sketch-based video generation results. Left: The input text prompts and sketches. Right: The video generation results. It can be seen that the generated results show high quality and good faithfulness with the input sketches. Our method can handle one/two keyframe sketch(es) at arbitrary user-specified time points (the frames corresponding to the input time points are highlighted by orange). +![](images/edf3a8e2b2b30f8e21352ee6db59bbdfa2bb1170cae42fc2d0617b16f0041d31.jpg) +A time-lapse aerial video shows a large, ornate domed building surrounded by lush greenery, transitioning from daylight to sunset to night, illuminated by artificial lights. + +![](images/bd0cfa4dca4f032ce9c1858075f506190674d0a4345fa7931a229cc75a03238b.jpg) + +![](images/306a50698fc2ec85405c297478e8da728c48f4dfd9a69393313dd09e5c8d07dc.jpg) + +![](images/75e25e63b5a2b7892e7807bad4698a960b5e39c4e15088fcae499132ed51f4e5.jpg) + +![](images/d83b67eaf759f5c101ef93700650fbcd77023e58217ed338587ca62a202240ff.jpg) + +![](images/fdc92561323bce7edbf55272dc181f97224a1a62288fc6094d697af1df209447.jpg) + +relationship between all frames and control keyframes, allowing propagation of $V$ (derived from the keyframe sketch features). Our inter-frame attention is distinct from typical cross-frame attention, which uses both $K$ and $V$ from control conditions; instead, we leverage the inter-frame similarity within the input noisy hidden video features and progressively insert the sketches' spatial and temporal information. The output $\widetilde{c}_i^{1:N}$ is computed as: $\mathrm{Attention}(Q,K,V) = \mathrm{Softmax}(\frac{QK^T}{\sqrt{d}})\cdot V$ , with + +$$ +Q = W _ {q} \cdot h _ {i} ^ {1: N}, K = W _ {k} \cdot h _ {i} ^ {t _ {1}, t _ {2}}, V = W _ {v} \cdot c _ {i} ^ {t _ {1}, t _ {2}}, \tag {3} +$$ + +where $W_{q}, W_{k}, W_{v}$ are trainable linear projection weights. The output of the inter-frame attention is fed into a feedforward layer to generate final residual features $\overline{h}_{i}^{1:N}$ . + +Training Strategy. To train the sketch condition network, we employ a hybrid training strategy in two stages. In the first stage, to accelerate convergence and address the issue of limited video data, the network is trained both for image generation at arbitrary time points and video generation with one or two keyframe sketches (randomly selected from corresponding sketch videos). In the second stage, video data alone is used to improve the temporal coherence. + +# 3.3. Sketch-based Video Editing + +For editing, given real or synthetic videos, users select one or two keyframes at arbitrary time points and modify the extracted sketches, with additional inputs of text prompts and masks $M^{1:N}$ that label regions to be edited for all the frames. Our method then generates realistic, local editing results. The input video is multiplied by the inverted masks to remove information from the edited regions and is subsequently encoded to generate a masked video latent representation $v_{0}^{1:N}$ . + +Video Insertion Module. For sketch-based editing, newly generated contents within the mask regions should be coherent with the original spatial and temporal features in the unedited regions. Thus, we design a video insertion module that analyzes the relationship between the input sketches and the original video. The video insertion module takes $v_{i-1}^{1:N}$ (input video latent or generated by a previous control block) as input and predicts the updated video features $v_{i}^{1:N}$ , similar to the sketch generation process. Since video features are not temporally sparse, we use a trainable copy of CogVideoX-2b's DiT block to directly generate video insertion features $\widetilde{v}_{i}^{1:N}$ . The sketch branch output $\widetilde{c}_{i}^{1:N}$ and video branch output $\widetilde{v}_{i}^{1:N}$ are multiplied by their respective masks and concatenated: + +$$ +\operatorname {C o n c a t} \left(\tilde {c} _ {i} ^ {1: N} * M ^ {1: N}, \tilde {v} _ {i} ^ {1: N} * \overline {{M}} ^ {1: N}\right), \tag {4} +$$ + +which serves as inputs to the feed-forward layers, producing final residual features incorporating the original video and sketch control information. This design ensures seamless integration of new contents with the original videos, effectively propagating edits across frames for dynamic motion. + +Training Strategy. Directly training the video editing network would lead to low fidelity with the input sketches, possibly because of the challenging interaction between the input sketches and videos. So we finetune it from the pretrained sketch condition network for generation and add the new video insertion module. The pretrained model already has good sketch fidelity and only requires learning video information. The network is trained in a self-supervised inpainting manner, with randomly generated masks to imitate real-world editing. + +Inference Latent Fusion. Although the original videos are encoded in the condition network, as shown in Fig. 8, + +![](images/0421acc3ba4c525b9761a4b546353e4a9d37813e9bd607dd4d7d1a41fd37c66a.jpg) +Input Video + +![](images/1a7f168e94e61a55406f323676cfdd8e2cf771390811c7f1cdd64fc6fd393795.jpg) +A snow-dusted volcano looms, steam plumes rising from its crest. Lush green flanks contrast with rocky slopes, while a rippling lake is in the foreground. + +![](images/8ce966540daf86766216580d5ba284b20e25bb0fad3c004b7804e257f47943ba.jpg) +Edited Results + +![](images/470bab0b35195b010fc37c9fc5ab3b7299d8f092783bcaa62b2af3c13a6b17d3.jpg) +A mesmerizing ornamental fish glides through the ocean, its electric blue, fiery orange, and iridescent green scales shimmering as it navigates the depths with deli- cate, patterned fins. +Input Video + +![](images/e2c14b5966d29ac8f6b6d044923893a56667fd6514f35305f7b6b25887a534ea.jpg) +Edited Results +Figure 4. The sketch-based video editing results. For each example, the text prompts and sketches are shown on the left. On the right, the input real videos are shown at the top, while the edited results with the control keyframe highlighted in orange are shown at the bottom. The editing region masks are manually provided by users, highlighted as orange boxes. Our method generates realistic local editing results. + +![](images/5e00984564c2725f2d35620805f8d6cb69401b77b5ec56ff1a38bb0dada818ad.jpg) +A man is speaking. He has thick hair and is wearing a black suit with a white shirt and a red tie underneath. The camera focuses on his face and upper body. +Input Video + +![](images/176b08af6cbb1a48bff42b9b488de994cd21fdabc132045cad378b012d039e45.jpg) + +![](images/fd661b398e226a3da1a66fb2eb9e034b3d0269647eb7003815ec0b72dba6da00.jpg) +Edited Results + +![](images/7b9da252f21b9adb832a7bb6dd275e242d7063fbbc133f54d0a3b359703c4ae4.jpg) + +![](images/5e32331a626647853f09154c9ce0f446d7d42dbb8f0ef957f5a1b0ec976279dc.jpg) +A green grassland and a small path extending into the distance. The grassland is covered with gravel. +Input Video + +![](images/cb85a421f7ea3062d94ff747c090afe2dfaf333e5b0a37c4d1d6972bbe2ecadb.jpg) +Edited Results + +fine details might be lost during editing. To address this, we propose a latent fusion approach at inference. Specifically, we apply DDIM inversion [58] to generate noisy latent codes of input videos across the inference steps. At Steps 25 and 49 (out of 50 total steps), the latent codes in the unedited regions are replaced with these inversion latent codes, ensuring better preservation of the original video's details and improving the coherence of the edited regions. + +# 4. Evaluation + +# 4.1. Implementation Details + +We implement SketchVideo based on CogVideoX-2b [73], trained on a subset of OpenVid [42] and LAION [55] datasets, with paired sketches from [7]. Training uses 8 NVIDIA H800 GPUs with a batch size of 8 and gradient accumulation of 4. For generation, one or two keyframe sketches are randomly sampled from the video, with 10,000 steps for each training stage. For editing, randomly drawn masks are applied and trained for 20,000 steps. For ease of reading, the input text prompts are simplified in figures. Additional implementation details and full input texts are available in the supplementary material. + +# 4.2. Results + +Our method generates high-quality videos from one or two keyframe sketches and text prompts. As shown in Fig. 3, our method can accurately control the object shape and scene layout, such as the rabbit's pose (1st row) and the position of lakes and buildings (2nd row). Text-only inputs cannot achieve such detailed geometry control. Our method also achieves interesting dynamic motion interpolation and extrapolation, generating smooth and realistic transitions, as seen in the cat's head shaking (3rd row) and the building's movement (last row). This allows control over both spatial layout and dynamic motion. + +Our method also supports sketch-based video editing. + +![](images/3782d4d2d3d1c44994c1f8170531c4dd4d9846734824fd2a0b746c9256fbec22.jpg) +Input Text A time-lapse video shows a large white and black cruise ship with yellow and gold accents docking at a pier on an overcast day, reflecting in calm waters. +Figure 5. The comparison results of sketch-based video generation. Text prompts are shown on the top. On the left, we show the input sketches and sketch-based image generation results by ControlNet [77]. On the right, we show the results of the compared approaches, including AMT [37], SparseCtrl [22], CtrlCogVideo [77], and ours. Our method produces better results, especially for the intermediate frames. + +![](images/87778312df5b6ae3d3dea199f26c39a863dbaea88039bce14c1705e849deed6f.jpg) + +![](images/206cfdfc783968466fdd0c5652a9f0c3fbfe9a29ee83cbcd38ca1ea9dfac7a8d.jpg) + +![](images/9b7c891ab8e925b9925e2d9120e60aadc1ec6cd62dfdbbd1ef97e246d53f1350.jpg) + +![](images/bc01f31a7d84aec396d8364320f31ede572a1673d2d3ec37948f5035649a5a12.jpg) + +Users specify bounding boxes and draw sketches in those regions, and our method generates photorealistic, seamlessly integrated content. As shown in Fig. 4, the new contents blend well with unedited regions and have dynamic motions like hat/hair rotation with original objects (1st row) or interesting fish swimming (2nd row). Our method can handle diverse editing cases, including object insertion, component replacement, and object removal. The unedited regions are well-preserved thanks to our latent fusion approach. + +# 4.3. Comparison + +For sketch-based generation, we compare our methods with three methods given the same keyframe sketches text prompts as inputs. For SparseCtrl [22], we use the official pretrained model and extract sketches by HED [69]. It is trained on videos in the WebVid-10M dataset [1] instead + +![](images/af9226a58f43b81c4b90e9c8bd415839b91410a0605b507a8cc783a75e58bd28.jpg) +Figure 6. The comparison results of sketch-based video editing. On the left, we show the drawn editing sketches (for the frames highlighted in orange), text prompts, and sketch-based image generation results by ControlNet [77]. On the right, we show the original videos and editing results by the compared methods, including InsV2V [17], AnyV2V [34], and ours. Our method generates the most realistic results and preserves unedited regions well. + +![](images/23d15a9e786666483470ae5e5306e23a4c0b373bf2a816544c0f4f4d20ddf362.jpg) + +![](images/1fd7e4aecef8cb49d95e0ffdbc42925b015f079a150042e10474d9fab32d592f.jpg) + +![](images/5a4b670008cf4036660feb229fceb7e04dd3f85a134b5af0649ab6290743d1de.jpg) + +of the OpenVid-1M dataset [42] used in our method. We extend SparseCtrl to CogVideoX-2b [73], using pixART $\delta$ [11] as the sketch condition encoder (with 5 DiT blocks same as our method) and white images to complete the missing condition frames. We also compare with an interpolation baseline, which uses ControlNet to translate two-frame sketches into images and then interpolates them with AMT [37]. For the sketch-base editing task, we compare ours with a text-based video editing method InsV2V [17] and a first-frame editing method AnyV2V [34]. We compare with additional methods [16, 20, 45, 70], as shown in supplemental material. + +As shown in Fig. 5, ControlNet [77] translates sketches into realistic images, which, however, lack temporal consistency and vary with shading and content. The interpolation results of AMT [37] exhibit fuzzy details and artifacts. SparseCtrl [22], based onAnimateDiff [23], exhibits temporal flickering, such as the suddenly appearing tree and distorted tower top (see the orange boxes in the 2nd row). Extending SparseCtrl to CogVideoX-2b [73] (Ctrl-CogVideo) still generates fuzzy details in intermediate frames, possibly due to the CogVideoX-2b's pretrained self-attention being designed for dense inputs instead of sparse sketches. Our method generates realistic videos with clear details and good temporal coherence, with even small details like the electric wires in the top-right corner accurately propagated. + +For sketch-based editing (Fig. 6), InsV2V [17] generates interesting results with birds but lacks control over shape and geometry through text prompts alone. For image-based video editing method, we utilize ControlNet to edit the first frame and then propagate editing into the video by AnyV2V [34]. However, since the motion is borrowed from the original video, AnyV2V struggles with new content, leading to fuzzy details and distortion in the bird, as well + +
MethodsLPIPS ↓CLIP ↑FidelityConsistencyRealism
AMT29.1796.123.133.513.57
SparseCtrl44.8596.482.792.942.83
Ctrl-CogVideo32.2398.042.862.472.50
Ours27.5698.311.211.081.11
+ +Table 1. The quantitative results of sketch-based video generation comparison. The LPIPS and CLIP numbers are scaled up $100 \times$ with each cell colored to indicate the best. + +
MethodsLPIPS ↓CLIP ↑PSNR↑FidelityPreservationRealism
InsV2V13.6195.3916.842.582.262.61
AnyV2V11.9293.4713.682.352.692.34
Ours9.7498.3436.481.071.051.04
+ +Table 2. The quantitative results of sketch-based video editing comparison. The LPIPS and CLIP numbers are scaled up $100 \times$ with each cell colored to indicate the best. + +as changes in unedited regions due to ControlNet's inability to retain the original features. In contrast, our method produces more realistic results with faithful representations of the sketch. More generation and editing results are available in the supplemental material and video. + +We follow SparseCtrl [22] and use the LPIPS [78] metric to measure sketch faithfulness between the input sketches and those extracted from the corresponding frames of the generated videos. We use the CLIP [49] similarity to assess temporal coherence. For sketch-based generation, we test 200 random examples from OpenVid [42], using the first and last frame sketches and the corresponding text prompts in the dataset. As shown in Table 1, our method achieves the lowest LPIPS and highest CLIP scores, indicating its superior performance. For sketch-based editing, we use additional an MSE metric to measure unedited region preservation. We utilize 10 examples with hand-drawn sketches as input. In Table 2, our method outperforms existing methods across all metrics, demonstrating its superiority. + +![](images/7e39767a6c4d7a8dec04171bf41aca2b61f99a186c996aa3d157ad11523cb0b6.jpg) + +![](images/ddfec4bcea406a7bacf28349df355d86ba490a12f61030bc4fb88192b7f06abf.jpg) +Extreme close-up of chicken and green pepper kebabs grilling on a barbeque with flames. Shallow focus and light smoke. vivid colours +(a) w/o Inter-Frame (b) Sketch K, V +(c) w/o Skip +(d) w/o Image +(e) ours +Figure 7. The ablation study results of sketch-based video generation. Our method generates more realistic and sketch-faithful results than the baselines. + +# 4.4. Ablation Study + +We conduct an ablation study to evaluate the effectiveness of each key component. Due to the restriction of computing resources, all the models are trained with a batch size of 2. For sketch-based generation, removing inter-frame attention and concatenating sketches with frame features at the temporal dimension results in strange control with low sketch faithfulness (Fig. 7 (a)). Similarly, if we replace the inter-frame attention with a typical cross-attention that uses + +![](images/f04947e3d62613f01bff44ede2a40b78f7e46b846a28543899c6c623d7de83a8.jpg) +Figure 8. The ablation study results of sketch-based video editing. Our method generates realistic results faithful to sketches. We utilize heat maps (top right) to visualize the difference between the edited and original frames. + +![](images/b8ffa9982b0b3559cc3a227f6c2f90c215f40a9de658dccac0b5de3c5afbb0f8.jpg) + +![](images/7632073a2b335d29397a41a1f469fff5898d658f15285867f10464aad61c4d73.jpg) + +![](images/4f01ba7dcd287125674cb5f174ce52e7bba9204b7dd7cafe21d45382d6222977.jpg) + +![](images/a7688278ba32ddf5c598b0b845eafd2a11955a800a67bb6d0aa96deb1fad97b5.jpg) + +
Metricw/o Inter-Frame Sketch K,Vw/o Skipw/o ImageOurs
LPIPS ↓36.3332.5931.9134.5830.79
CLIP ↑98.1098.1997.6098.2498.48
Metricw/o Video modulew/o Pretrainw/o Latent FusionOurs
LPIPS ↓9.3112.609.779.74
CLIP ↑97.9798.1298.4498.34
PSNR ↑33.6136.0531.6936.48
+ +Table 3. The qualitative results of ablation study for sketch-based generation and editing, with each cell colored to indicate the best and second best. + +sketch features for both Key and Value (b), the sketch fidelity is also degraded due to the lack of effective control propagation. If we remove the skip structure and predict residual features for the first 5 blocks (c), the realism of the generated results is negatively influenced due to less control ability. Removing image data during training (d) causes significant geometry mismatch with input sketches. Compared to these alternatives, our method produces the most realistic and faithful results. As shown in Table 3, our method has the lowest LPIPS values (supporting the best sketch fidelity) while the CLIP metric has similar results (human eyes cannot recognize temporal coherence difference). + +In sketch-based video editing (Fig. 8), removing the video insertion module (a) and solely utilizing the sketch condition encoder results in good sketch fidelity but obvious inconsistency with unedited regions. Training the sketch condition encoder from scratch without generation pretraining (b) reduces the faithfulness to the input sketches in the edited regions, which is also validated by the high LPIPS in Table 3. We use heat maps to reveal differences between edited and original frames. Removing the latent fusion strategy changes unedited region details (c), such as mountains and waves, as shown in the heat maps. The low PSNR value in unedited regions also supports this, as shown in Table 3. Our method maintains strong sketch fidelity, consistency, and unedited region preservation. + +# 4.5. User Study + +We conducted a user study to validate that our method outperforms existing approaches. For video generation, we + +randomly selected 10 examples from the test set in Sec. 4.3. In the questionnaire, participants were shown the results of different methods in random order. 20 participants ranked the methods results by three criteria: Sketch Fidelity, Temporal Consistency, and Video Realism. The rankings provided by the participants were used as scores. As shown in Table 1, our method outperforms existing methods in all criteria, demonstrating its superior performance. For sketch-based editing, we used the same 10 examples from the test set in Sec. 4.3. The same participants ranked the methods based on three criteria: Sketch Fidelity, Unedited Region Preservation, and Video Realism. As shown in Table 2, our method also outperforms existing approaches in editing. + +# 5. Conclusion and Discussion + +We presented SketchVideo, a unified method for sketch-based video generation and editing. For generation, we proposed a sketch condition network that predicts residual features for skipped DiT blocks of the base models to save the memory and achieve effective control. An inter-frame attention is further proposed to propagate the keyframe sketches, achieving interesting motion interpolation or extrapolation. We also introduce a hybrid image and video training strategy. For editing, we incorporate a video insertion module to ensure the newly generated content is spatially and temporally coherent with the original video. During inference, a latent fusion approach preserves unedited regions accurately. + +Limitation and Future Work. While our method generates high-quality videos, its capabilities are still limited by the pretrained text-to-video model. Enhancing performance with more powerful pretrained models and generating long videos instead of video clips is a potential avenue. Additionally, similar to image generation, our method struggles with too complex scenarios, such as human hands and interaction between multiple objects. Incorporating 3D priors [59, 63, 86] like SMPL-X [47] helps address these issues in human scenarios. Moreover, while our method focuses on geometry control, exploring appearance customization through tools like color strokes presents an intriguing area for future research. + +# 6. Acknowledgments + +This work was sponsored by CCF-Kuaishou Large Model Explorer Fund, Beijing Municipal Science and Technology Commission (No. Z231100005923031), Innovation Funding of ICT, CAS (No. E461020) and National Natural Science Foundation of China (No. 62322210). The authors would like to acknowledge the Nanjing Institute of InforSuperBahn OneAiNexus for providing the training and evaluation platform. + +# References + +[1] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In Int. Conf. Comput. Vis., pages 1708-1718, 2021. 6 +[2] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, Varun Jampani, and Robin Rombach. Stable video diffusion: Scaling latent video diffusion models to large datasets. CoRR, abs/2311.15127, 2023. 2, 3 +[3] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In IEEE Conf. Comput. Vis. Pattern Recog., pages 22563-22575, 2023. 3 +[4] Minwoo Byeon, Beomhee Park, Haecheon Kim, Sungjun Lee, Woonhyuk Baek, and Saehoon Kim. Coyo-700m: Image-text pair dataset. https://github.com/kakaobrain/coyo-dataset, 2022.1 +[5] Duygu Ceylan, Chun-Hao Paul Huang, and Niloy J. Mitra. Pix2video: Video editing using image diffusion. In Int. Conf. Comput. Vis., pages 23149-23160, 2023. 3 +[6] Wenhao Chai, Xun Guo, Gaoang Wang, and Yan Lu. Stable-video: Text-driven consistency-aware diffusion video editing. In Int. Conf. Comput. Vis., pages 22983-22993, 2023. 3 +[7] Caroline Chan, Frédo Durand, and Phillip Isola. Learning to generate line drawings that convey geometry and semantics. In IEEE Conf. Comput. Vis. Pattern Recog., pages 7915-7925, 2022. 6 +[8] Hila Chefer, Shiran Zada, Roni Paiss, Ariel Ephrat, Omer Tov, Michael Rubinstein, Lior Wolf, Tali Dekel, Tomer Michaeli, and Inbar Mosseri. Still-moving: Customized video generation without customized video data. ACM Trans. Graph., 43(6):244:1-244:11, 2024. 3 +[9] Haoxin Chen, Menghan Xia, Yingqing He, Yong Zhang, Xiaodong Cun, Shaoshu Yang, Jinbo Xing, Yaofang Liu, Qifeng Chen, Xintao Wang, Chao Weng, and Ying Shan. Videocrafter1: Open diffusion models for high-quality video generation. CoRR, abs/2310.19512, 2023. 3 +[10] Haoxin Chen, Yong Zhang, Xiaodong Cun, Menghan Xia, Xintao Wang, Chao Weng, and Ying Shan. Videographer2: Overcoming data limitations for high-quality video diffusion models. In IEEE Conf. Comput. Vis. Pattern Recog., pages 7310-7320, 2024. 3 +[11] Junsong Chen, Yue Wu, Simian Luo, Enze Xie, Sayak Paul, Ping Luo, Hang Zhao, and Zhenguo Li. Pixart-δ: Fast and controllable image generation with latent consistency models. CoRR, abs/2401.05252, 2024. 2, 3, 4, 7 +[12] Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Zhongdao Wang, James T. Kwok, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart-α: Fast training of diffusion transformer for photorealistic text-to-image synthesis. In Int. Conf. Learn. Represent., 2024. 1, 2 +[13] Shu-Yu Chen, Wanchao Su, Lin Gao, Shihong Xia, and + +Hongbo Fu. Deepfacedrawing: deep generation of face images from sketches. ACM Trans. Graph., 39(4):72, 2020. 2 +[14] Shu-Yu Chen, Feng-Lin Liu, Yu-Kun Lai, Paul L. Rosin, Chunpeng Li, Hongbo Fu, and Lin Gao. Deepfaceediting: deep face generation and editing with disentangled geometry and appearance control. ACM Trans. Graph., 40(4):90:1-90:15, 2021. 2 +[15] Weifeng Chen, Jie Wu, Pan Xie, Hefeng Wu, Jiashi Li, Xin Xia, Xuefeng Xiao, and Liang Lin. Control-a-video: Controllable text-to-video generation with diffusion models, 2023. 3 +[16] Xinyuan Chen, Yaohui Wang, Lingjun Zhang, Shaobin Zhuang, Xin Ma, Jiashuo Yu, Yali Wang, Dahua Lin, Yu Qiao, and Ziwei Liu. Seine: Short-to-long video diffusion model for generative transition and prediction. In Int. Conf. Learn. Represent., 2023. 2, 7 +[17] Jiaxin Cheng, Tianjun Xiao, and Tong He. Consistent video-to-video transfer using synthetic dataset. In Int. Conf. Learn. Represent., 2024. 3, 7 +[18] Ernie Chu, Shuo-Yen Lin, and Jun-Cheng Chen. Video controlnet: Towards temporally consistent synthetic-to-real video translation using conditional image diffusion models. CoRR, abs/2305.19193, 2023. 3 +[19] Nathaniel Cohen, Vladimir Kulikov, Matan Kleiner, Inbar Huberman-Spiegelglas, and Tomer Michaeli. Slicedit: Zero-shot video editing with text-to-image diffusion models using spatio-temporal slices. arXiv preprint arXiv:2405.12211, 2024.3 +[20] Michal Geyer, Omer Bar-Tal, Shai Bagon, and Tali Dekel. Tokenflow: Consistent diffusion features for consistent video editing. In Int. Conf. Learn. Represent., 2024. 3, 7 +[21] Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron C. Courville, and Yoshua Bengio. Generative adversarial networks. CoRR, abs/1406.2661, 2014. 2 +[22] Yuwei Guo, Ceyuan Yang, Anyi Rao, Maneesh Agrawala, Dahua Lin, and Bo Dai. Sparsectrl: Adding sparse controls to text-to-video diffusion models. In Eur. Conf. Comput. Vis., pages 330-348, 2024. 2, 3, 4, 6, 7 +[23] Yuwei Guo, Ceyuan Yang, Anyi Rao, Zhengyang Liang, Yaohui Wang, Yu Qiao, Maneesh Agrawala, Dahua Lin, and Bo Dai. Animatediff: Imagine your personalized text-to-image diffusion models without specific tuning. In Int. Conf. Learn. Represent., 2024. 3, 7 +[24] Hao He, Yinghao Xu, Yuwei Guo, Gordon Wetzstein, Bo Dai, Hongsheng Li, and Ceyuan Yang. Cameractrl: Enabling camera control for text-to-video generation. CoRR, abs/2404.02101, 2024. 3 +[25] Yingqing He, Tianyu Yang, Yong Zhang, Ying Shan, and Qifeng Chen. Latent video diffusion models for high-fidelity video generation with arbitrary lengths. CoRR, abs/2211.13221, 2022. 2, 3 +[26] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. Adv. Neural Inform. Process. Syst., 35:8633-8646, 2022. 3 +[27] Wenyi Hong, Ming Ding, Wendi Zheng, Xinghan Liu, and Jie Tang. Cogvideo: Large-scale pretraining for + +text-to-video generation via transformers. arXiv preprint arXiv:2205.15868, 2022. 1, 2, 3 +[28] Zhitong Huang, Mohan Zhang, and Jing Liao. LVCD: reference-based linear art video colorization with diffusion models. CoRR, abs/2409.12960, 2024. 3 +[29] Ondrej Jamriska, Sárka Sochorová, Ondrej Texler, Michal Lukác, Jakub Fiser, Jingwan Lu, Eli Shechtman, and Daniel Sykora. Stylizing video by example. ACM Trans. Graph., 38 (4):107:1-107:11, 2019. 3 +[30] Rui Jiang, Guang-Cong Zheng, Teng Li, Tian-Rui Yang, Jing-Dong Wang, and Xi Li. A survey of multimodal controllable diffusion models. Journal of Computer Science and Technology, 39(3):509-541, 2024. 2 +[31] Yuming Jiang, Tianxing Wu, Shuai Yang, Chenyang Si, Dahua Lin, Yu Qiao, Chen Change Loy, and Ziwei Liu. Videobooth: Diffusion-based video generation with image prompts. In IEEE Conf. Comput. Vis. Pattern Recog., pages 6689-6700, 2024. 3 +[32] Yoni Kasten, Dolev Ofri, Oliver Wang, and Tali Dekel. Layered neural atlases for consistent video editing. ACM Trans. Graph., 40(6):210:1-210:12, 2021. 3 +[33] Levon Khachatryan, Andranik Movsisyan, Vahram Tadevosyan, Roberto Henschel, Zhangyang Wang, Shant Navasardyan, and Humphrey Shi. Text2video-zero: Text-to-image diffusion models are zero-shot video generators. In Int. Conf. Comput. Vis., pages 15908-15918, 2023. 3 +[34] Max Ku, Cong Wei, Weiming Ren, Harry Yang, and Wenhu Chen. Anyv2v: A tuning-free framework for any video-to-video editing tasks. arXiv preprint arXiv:2403.14468, 2024. 2, 3, 7 +[35] PKU-Yuan Lab and Tuzhan AI etc. Open-sora-plan, 2024. 1, 3 +[36] Yaowei Li, Xintao Wang, Zhaoyang Zhang, Zhouxia Wang, Ziyang Yuan, Liangbin Xie, Yuexian Zou, and Ying Shan. Image conductor: Precision control for interactive video synthesis. CoRR, abs/2406.15339, 2024. 3 +[37] Zhen Li, Zuo-Liang Zhu, Ling-Hao Han, Qibin Hou, Chun-Le Guo, and Ming-Ming Cheng. Amt: All-pairs multi-field transforms for efficient frame interpolation. In IEEE Conf. Comput. Vis. Pattern Recog., 2023. 2, 6, 7 +[38] Shanchuan Lin, Bingchen Liu, Jiashi Li, and Xiao Yang. Common diffusion noise schedules and sample steps are flawed. In IEEE/CVF Winter Conference on Applications of Computer Vision, WACV, pages 5392-5399, 2024. 3 +[39] Feng-Lin Liu, Shu-Yu Chen, Yu-Kun Lai, Chunpeng Li, Yue-Ren Jiang, Hongbo Fu, and Lin Gao. Deepfacevideoediting: sketch-based deep editing of face videos. ACM Trans. Graph., 41(4):167:1-167:16, 2022. 3 +[40] Chong Mou, Mingdeng Cao, Xintao Wang, Zhaoyang Zhang, Ying Shan, and Jian Zhang. Revideo: Remake a video with motion and content control. CoRR, abs/2405.13865, 2024. 2, 3 +[41] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, and Ying Shan. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. In AAAI, pages 4296-4304, 2024. 2 + +[42] Kepan Nan, Rui Xie, Penghao Zhou, Tiehan Fan, Zhenheng Yang, Zhijie Chen, Xiang Li, Jian Yang, and Ying Tai. Openvid-1m: A large-scale high-quality dataset for text-to-video generation. CoRR, abs/2407.02371, 2024. 1, 6, 7 +[43] OpenAI. Sora overview:https://openai.com/index/soral/, 2024.1,3 +[44] Hao Ouyang, Qiuyu Wang, Yuxi Xiao, Qingyan Bai, Juntao Zhang, Kecheng Zheng, Xiaowei Zhou, Qifeng Chen, and Qifeng Chen. Codef: Content deformation fields for temporally consistent video processing. In IEEE Conf. Comput. Vis. Pattern Recog., pages 8089-8099, 2024. 3 +[45] Wenqi Ouyang, Yi Dong, Lei Yang, Jianlou Si, and Xingang Pan. I2vedit: First-frame-guided video editing via image-to-video diffusion models. CoRR, abs/2405.16537, 2024. 2, 3, 7 +[46] Gaurav Parmar, Taesung Park, Srinivasa Narasimhan, and Jun-Yan Zhu. One-step image translation with text-to-image models. CoRR, abs/2403.12036, 2024. 2 +[47] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed A. A. Osman, Dimitrios Tzionas, and Michael J. Black. Expressive body capture: 3d hands, face, and body from a single image. In IEEE Conf. Comput. Vis. Pattern Recog., 2019. 8 +[48] Bohao Peng, Jian Wang, Yuechen Zhang, Wenbo Li, Ming-Chang Yang, and Jiaya Jia. Controlnext: Powerful and efficient control for image and video generation. CoRR, abs/2408.06070, 2024. 2 +[49] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763, 2021. 7 +[50] Elad Richardson, Yuval Alaluf, Or Patashnik, Yotam Nitzan, Yaniv Azar, Stav Shapiro, and Daniel Cohen-Or. Encoding in style: A stylegan encoder for image-to-image translation. In IEEE Conf. Comput. Vis. Pattern Recog., pages 2287–2296, 2021. 2 +[51] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In IEEE Conf. Comput. Vis. Pattern Recog., pages 10674-10685, 2022. 1, 2, 3 +[52] Manuel Ruder, Alexey Dosovitskiy, and Thomas Brox. Artistic style transfer for videos and spherical images. Int. J. Comput. Vis., 126(11):1199-1219, 2018. 3 +[53] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L. Denton, Seyed Kamyar Seyed Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, Jonathan Ho, David J. Fleet, and Mohammad Norouzi. Photorealistic text-to-image diffusion models with deep language understanding. In Adv. Neural Inform. Process. Syst., 2022. 1, 2 +[54] Tim Salimans and Jonathan Ho. Progressive distillation for fast sampling of diffusion models. In Int. Conf. Learn. Represent., 2022. 3 +[55] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo + +Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, Patrick Schramowski, Srivatsa Kundurthy, Katherine Crowson, Ludwig Schmidt, Robert Kaczmarczyk, and Jenia Jitsev. LAION-5B: an open large-scale dataset for training next generation image-text models. In Adv. Neural Inform. Process. Syst., 2022. 1, 6 +[56] Xiaoyu Shi, Zhaoyang Huang, Fu-Yun Wang, Weikang Bian, Dasong Li, Yi Zhang, Manyuan Zhang, Ka Chun Cheung, Simon See, Hongwei Qin, Jifeng Dai, and Hongsheng Li. Motion-i2v: Consistent and controllable image-to-video generation with explicit motion modeling. In ACM SIGGRAPH, page 111. ACM, 2024. 3 +[57] Uriel Singer, Amit Zohar, Yuval Kirstain, Shelly Sheynin, Adam Polyak, Devi Parikh, and Yaniv Taigman. Video editing via factorized diffusion distillation. CoRR, abs/2403.09334, 2024. 3 +[58] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In Int. Conf. Learn. Represent., 2021. 6 +[59] Jia-Mu Sun, Tong Wu, and Lin Gao. Recent advances in implicit representation-based 3d shape generation. Vis. Intell., 2(1), 2024. 8 +[60] Vchitect Team and Shanghai Artificial Intelligence Laboratory. Vchitect-2.0: Parallel transformer for scaling up video diffusion models, 2024. 3 +[61] Zheng Tianpeng, Chen Yanxiang, Wen Xinzhe, Li Yancheng, and Wang Zhiyuan. Research on diffusion model generated video datasets and detection benchmarks. Journal of Image and Graphics, pages 1-13, 2024. 1 +[62] Rotem Tzaban, Ron Mokady, Rinon Gal, Amit Bermano, and Daniel Cohen-Or. *Stitch it in time: Gan-based facial editing of real videos*. In SIGGRAPH Asia Conference Papers, pages 29:1–29:9. ACM, 2022. 3 +[63] Guangming Wang, Lei Pan, Songyou Peng, Shaohui Liu, Chenfeng Xu, Yanzi Miao, Wei Zhan, Masayoshi Tomizuka, Marc Pollefeys, and Hesheng Wang. Nerf in robotics: A survey. CoRR, abs/2405.01333, 2024. 8 +[64] Jiuniu Wang, Hangjie Yuan, Dayou Chen, Yingya Zhang, Xiang Wang, and Shiwei Zhang. Modelscope text-to-video technical report. CoRR, abs/2308.06571, 2023. 3 +[65] Ting-Chun Wang, Ming-Yu Liu, Jun-Yan Zhu, Andrew Tao, Jan Kautz, and Bryan Catanzaro. High-resolution image synthesis and semantic manipulation with conditional gans. In IEEE Conf. Comput. Vis. Pattern Recog., pages 8798-8807, 2018. 2 +[66] Zhouxia Wang, Ziyang Yuan, Xintao Wang, Yaowei Li, Tianshui Chen, Menghan Xia, Ping Luo, and Ying Shan. Motionctrl: A unified and flexible motion controller for video generation. In ACM SIGGRAPH, page 114, 2024. 3 +[67] Jianzong Wu, Xiangtai Li, Yanhong Zeng, Jiangning Zhang, Qianyu Zhou, Yining Li, Yunhai Tong, and Kai Chen. Motionbooth: Motion-aware customized text-to-video generation. CoRR, abs/2406.17758, 2024. 3 +[68] Tao Wu, Yong Zhang, Xintao Wang, Xianpan Zhou, Guangcong Zheng, Zhongang Qi, Ying Shan, and Xi Li. Customcrafter: Customized video generation with preserving motion and concept composition abilities. CoRR, abs/2408.13239, 2024. 3 + +[69] Saining Xie and Zhuowen Tu. Holistically-nested edge detection. In Int. Conf. Comput. Vis., pages 1395-1403, 2015. 6 +[70] Jinbo Xing, Hanyuan Liu, Menghan Xia, Yong Zhang, Xintao Wang, Ying Shan, and Tien-Tsin Wong. Tooncrafter: Generative cartoon interpolation. CoRR, abs/2405.17933, 2024. 2, 3, 7 +[71] Shuai Yang, Yifan Zhou, Ziwei Liu, and Chen Change Loy. Rerender A video: Zero-shot text-guided video-to-video translation. In SIGGRAPH Asia Conference Papers, pages 95:1-95:11. ACM, 2023. 2, 3 +[72] Shiuyuan Yang, Liang Hou, Haibin Huang, Chongyang Ma, Pengfei Wan, Di Zhang, Xiaodong Chen, and Jing Liao. Direct-a-video: Customized video generation with user-directed camera movement and object motion. In ACM SIGGRAPH, page 113, 2024. 3 +[73] Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, et al. Cogvideox: Text-to-video diffusion models with an expert transformer. arXiv preprint arXiv:2408.06072, 2024. 1, 2, 3, 6, 7 +[74] Danah Yatim, Rafail Fridman, Omer Bar-Tal, Yoni Kasten, and Tali Dekel. Space-time diffusion features for zero-shot text-driven motion transfer. In IEEE Conf. Comput. Vis. Pattern Recog., pages 8466-8476, 2024. 3 +[75] Chen Yuan, Zhao Yang, Zhang Xiaojuan, and Liu Xiaoping. Sketch colorization with finite color space prior. Journal of Image and Graphics, 29(4):978-988, 2024. 2 +[76] Liang Yuan, Dingkun Yan, Suguru Saito, and Issei Fujishiro. Diffmat: Latent diffusion models for image-guided material generation. Visual Informatics, 8(1):6-14, 2024. 2 +[77] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Int. Conf. Comput. Vis., pages 3813-3824, 2023. 2, 3, 4, 6, 7 +[78] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In IEEE Conf. Comput. Vis. Pattern Recog., pages 586-595, 2018. 7 +[79] Yuxin Zhang, Fan Tang, Nisha Huang, Haibin Huang, Chongyang Ma, Weiming Dong, and Changsheng Xu. Motioncrafter: One-shot motion customization of diffusion models. CoRR, abs/2312.05288, 2023. 3 +[80] Yabo Zhang, Yuxiang Wei, Dongsheng Jiang, Xiaopeng Zhang, Wangmeng Zuo, and Qi Tian. Controlvideo: Training-free controllable text-to-video generation. In Int. Conf. Learn. Represent., 2024. 3 +[81] Zhenghao Zhang, Junchao Liao, Menghao Li, Long Qin, and Weizhi Wang. Tora: Trajectory-oriented diffusion transformer for video generation. CoRR, abs/2407.21705, 2024. 3 +[82] Zhixing Zhang, Bichen Wu, Xiaoyan Wang, Yaqiao Luo, Luxin Zhang, Yinan Zhao, Peter Vajda, Dimitris N. Metaxas, and Licheng Yu. AVID: any-length video inpainting with diffusion model. In IEEE Conf. Comput. Vis. Pattern Recog., pages 7162-7172, 2024. 3 +[83] Shihao Zhao, Dongdong Chen, Yen-Chun Chen, Jianmin Bao, Shaozhe Hao, Lu Yuan, and Kwan-Yee K. Wong. + +Uni-controlnet: All-in-one control to text-to-image diffusion models. In Adv. Neural Inform. Process. Syst., 2023. 2 +[84] Zangwei Zheng, Xiangyu Peng, Tianji Yang, Chenhui Shen, Shenggui Li, Hongxin Liu, Yukun Zhou, Tianyi Li, and Yang You. Open-sora: Democratizing efficient video production for all, 2024. 1, 2, 3 +[85] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A. Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In Int. Conf. Comput. Vis., pages 2242-2251, 2017. 2 +[86] Siting Zhu, Guangming Wang, Dezhi Kong, and Hesheng Wang. 3d gaussian splatting in robotics: A survey. CoRR, abs/2410.12262, 2024. 8 \ No newline at end of file diff --git a/2025/SketchVideo_ Sketch-based Video Generation and Editing/images.zip b/2025/SketchVideo_ Sketch-based Video Generation and Editing/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..ba86854ceec83064a22258a0c6fed1b26d98090e --- /dev/null +++ b/2025/SketchVideo_ Sketch-based Video Generation and Editing/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e348128aceeaa14bae6e0fa67af51441413766f9d840a981ac354ffd24fd168a +size 853205 diff --git a/2025/SketchVideo_ Sketch-based Video Generation and Editing/layout.json b/2025/SketchVideo_ Sketch-based Video Generation and Editing/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..56548ea619d4ef3c77688a20b5586f724f2dff43 --- /dev/null +++ b/2025/SketchVideo_ Sketch-based Video Generation and Editing/layout.json @@ -0,0 +1,10638 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 125, + 103, + 484, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 125, + 103, + 484, + 121 + ], + "spans": [ + { + "bbox": [ + 125, + 103, + 484, + 121 + ], + "type": "text", + "content": "SketchVideo: Sketch-based Video Generation and Editing" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 142, + 559, + 159 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 142, + 559, + 159 + ], + "spans": [ + { + "bbox": [ + 61, + 142, + 559, + 159 + ], + "type": "text", + "content": "Feng-Lin Liu" + }, + { + "bbox": [ + 61, + 142, + 559, + 159 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 61, + 142, + 559, + 159 + ], + "type": "text", + "content": " Hongbo Fu" + }, + { + "bbox": [ + 61, + 142, + 559, + 159 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 61, + 142, + 559, + 159 + ], + "type": "text", + "content": " Xintao Wang" + }, + { + "bbox": [ + 61, + 142, + 559, + 159 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 61, + 142, + 559, + 159 + ], + "type": "text", + "content": " Weicai Ye" + }, + { + "bbox": [ + 61, + 142, + 559, + 159 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 61, + 142, + 559, + 159 + ], + "type": "text", + "content": " Pengfei Wan" + }, + { + "bbox": [ + 61, + 142, + 559, + 159 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 61, + 142, + 559, + 159 + ], + "type": "text", + "content": " Di Zhang" + }, + { + "bbox": [ + 61, + 142, + 559, + 159 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 61, + 142, + 559, + 159 + ], + "type": "text", + "content": " Lin Gao" + }, + { + "bbox": [ + 61, + 142, + 559, + 159 + ], + "type": "inline_equation", + "content": "^{1,2*}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 64, + 163, + 559, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 163, + 559, + 186 + ], + "spans": [ + { + "bbox": [ + 64, + 163, + 559, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 64, + 163, + 559, + 186 + ], + "type": "text", + "content": "Beijing Key Laboratory of Mobile Computing and Pervasive Device, Institute of Computing Technology, Chinese Academy of Sciences " + }, + { + "bbox": [ + 64, + 163, + 559, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 64, + 163, + 559, + 186 + ], + "type": "text", + "content": "University of Chinese Academy of Sciences " + }, + { + "bbox": [ + 64, + 163, + 559, + 186 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 64, + 163, + 559, + 186 + ], + "type": "text", + "content": "Hong Kong University of Science and Technology " + }, + { + "bbox": [ + 64, + 163, + 559, + 186 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 64, + 163, + 559, + 186 + ], + "type": "text", + "content": "Kuaishou Technology" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 146, + 194, + 231, + 263 + ], + "blocks": [ + { + "bbox": [ + 69, + 204, + 141, + 255 + ], + "lines": [ + { + "bbox": [ + 69, + 204, + 141, + 255 + ], + "spans": [ + { + "bbox": [ + 69, + 204, + 141, + 255 + ], + "type": "text", + "content": "\"A close-up of a majestic male lion's face, with a tawny mane, piercing amber eyes, and subtle scars, bathed in golden sunlight, showcasing his power and wild spirit.\"" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 146, + 194, + 231, + 263 + ], + "lines": [ + { + "bbox": [ + 146, + 194, + 231, + 263 + ], + "spans": [ + { + "bbox": [ + 146, + 194, + 231, + 263 + ], + "type": "image", + "image_path": "caeda09bc9fe18f0c8cf1cb0c674b0ab9ffcda319bd8f58736f66b52e41aed53.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 70, + 263, + 120, + 332 + ], + "blocks": [ + { + "bbox": [ + 70, + 263, + 120, + 332 + ], + "lines": [ + { + "bbox": [ + 70, + 263, + 120, + 332 + ], + "spans": [ + { + "bbox": [ + 70, + 263, + 120, + 332 + ], + "type": "image", + "image_path": "5e578ebc2b971b8498cc9bfac9815da0080c43814477fcc68cdc5b4ea6de9faf.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 122, + 264, + 230, + 297 + ], + "lines": [ + { + "bbox": [ + 122, + 264, + 230, + 297 + ], + "spans": [ + { + "bbox": [ + 122, + 264, + 230, + 297 + ], + "type": "text", + "content": "\"The video is a drone shot of a serene landscape featuring a deep blue lagoon surrounded by lush green trees and rocky cliffs. The lagoon is nestled between the mountains, creating a picturesque scene.\"" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 232, + 193, + 538, + 332 + ], + "blocks": [ + { + "bbox": [ + 123, + 298, + 230, + 330 + ], + "lines": [ + { + "bbox": [ + 123, + 298, + 230, + 330 + ], + "spans": [ + { + "bbox": [ + 123, + 298, + 230, + 330 + ], + "type": "text", + "content": "\"A clear blue sky above the mountains, with a few clouds scattered across it. A deep blue lake nestled between the mountains. The mountains have patches of snow that still cling to their peaks.\"" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 232, + 193, + 538, + 332 + ], + "lines": [ + { + "bbox": [ + 232, + 193, + 538, + 332 + ], + "spans": [ + { + "bbox": [ + 232, + 193, + 538, + 332 + ], + "type": "image", + "image_path": "0b935ee7d70550e86839543a0c26f29b2fa26098d49d92ad7f71b429971a222d.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 273, + 331, + 368, + 338 + ], + "lines": [ + { + "bbox": [ + 273, + 331, + 368, + 338 + ], + "spans": [ + { + "bbox": [ + 273, + 331, + 368, + 338 + ], + "type": "text", + "content": "(a) Sketch-based Video Generation" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 78, + 338, + 539, + 441 + ], + "blocks": [ + { + "bbox": [ + 78, + 338, + 539, + 441 + ], + "lines": [ + { + "bbox": [ + 78, + 338, + 539, + 441 + ], + "spans": [ + { + "bbox": [ + 78, + 338, + 539, + 441 + ], + "type": "image", + "image_path": "23cd0f2c4a391929d8ccaa4b1c3959237188627b1d775931462fec892c55b843.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 274, + 440, + 359, + 448 + ], + "lines": [ + { + "bbox": [ + 274, + 440, + 359, + 448 + ], + "spans": [ + { + "bbox": [ + 274, + 440, + 359, + 448 + ], + "type": "text", + "content": "(b) Sketch-based Video Editing" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 54, + 448, + 555, + 492 + ], + "lines": [ + { + "bbox": [ + 54, + 448, + 555, + 492 + ], + "spans": [ + { + "bbox": [ + 54, + 448, + 555, + 492 + ], + "type": "text", + "content": "Figure 1. Our method enables high-quality video generation (a) and editing (b) based on sketch and text inputs. (a) Top: With the same text prompt, different keyframe sketches lead to results with similar semantics but diverse sketch-faithful geometry. (a) Bottom: With the same sketches, varied text prompts yield diverse appearances. (b) Users can also edit real videos by drawing on keyframe sketches, with edits automatically propagated even when edited objects in original videos have translation and rotation." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 151, + 502, + 200, + 515 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 502, + 200, + 515 + ], + "spans": [ + { + "bbox": [ + 151, + 502, + 200, + 515 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 54, + 528, + 295, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 528, + 295, + 696 + ], + "spans": [ + { + "bbox": [ + 54, + 528, + 295, + 696 + ], + "type": "text", + "content": "Video generation and editing conditioned on text prompts or images have undergone significant advancements. However, challenges remain in accurately controlling global layout and geometry details solely by texts, and supporting motion control and local modification through images. In this paper, we aim to achieve sketch-based spatial and motion control for video generation and support fine-grained editing of real or synthetic videos. Based on the DiT video generation model, we propose a memory-efficient control structure with sketch control blocks that predict residual features of skipped DiT blocks. Sketches are drawn on one or two keyframes (at arbitrary time points) for easy interaction. To propagate such temporally sparse sketch conditions across all frames, we propose an inter-frame at" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 505, + 555, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 505, + 555, + 613 + ], + "spans": [ + { + "bbox": [ + 313, + 505, + 555, + 613 + ], + "type": "text", + "content": "tention mechanism to analyze the relationship between the keyframes and each video frame. For sketch-based video editing, we design an additional video insertion module that maintains consistency between the newly edited content and the original video's spatial feature and dynamic motion. During inference, we use latent fusion for the accurate preservation of unedited regions. Extensive experiments demonstrate that our SketchVideo achieves superior performance in controllable video generation and editing." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 314, + 633, + 394, + 645 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 633, + 394, + 645 + ], + "spans": [ + { + "bbox": [ + 314, + 633, + 394, + 645 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 653, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 555, + 713 + ], + "type": "text", + "content": "Diffusion-based text-to-image [12, 51, 53] and text-to-video [27, 35, 43, 73, 84] models advance significantly due to improvements in datasets [4, 42, 55, 61] and denoising network architectures [12, 51]. While text prompts effectively describe high-level semantics, they lack control of scene" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 70, + 702, + 241, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 702, + 241, + 712 + ], + "spans": [ + { + "bbox": [ + 70, + 702, + 241, + 712 + ], + "type": "text", + "content": "*Corresponding Author: Lin Gao (gaolin@ict.ac.cn)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23379" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 228 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 228 + ], + "type": "text", + "content": "Layouts and geometric details. To address this, existing video generation methods [2, 25, 73] utilize images as additional conditions but raise the questions of how to generate input images and achieve detailed editing. Sketching serves as a user-friendly interaction tool to capture spatial content and shape details accurately. One or two sketches are already sufficient to convey desired scene structures and motion information for short video clips (around 6 seconds), which are the target of our and most existing video generation methods. However, using such sparse keyframe sketches presents several challenges, including reasonably completing the missing frames, improving memory efficiency, and addressing the limited size of video datasets." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 228, + 295, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 228, + 295, + 443 + ], + "spans": [ + { + "bbox": [ + 56, + 228, + 295, + 443 + ], + "type": "text", + "content": "A naive solution is to translate the input keyframe sketches into images and then utilize interpolation methods [16, 37, 70] for video generation. However, it is nontrivial to ensure consistency during keyframe sketch-to-image generation, which significantly affects the video quality. This approach also struggles to generate extrapolation frames when applying conditions in intermediate frames rather than beginning and ending time points. Another possible approach is to utilize white placeholders to fill missing condition frames and directly apply ControlNet [77] into video models, similar to SparseCtrl [22]. However, this requires the same network to process both sketches and white placeholders simultaneously, while the pretrained blocks handle tasks far from this sparse propagation. Additionally, for DiT-based video frameworks [27, 84], the traditional strategy [11] that copies half of the pretrained model as a condition network easily causes the out-of-memory issue." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 443, + 295, + 706 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 443, + 295, + 706 + ], + "spans": [ + { + "bbox": [ + 55, + 443, + 295, + 706 + ], + "type": "text", + "content": "To address these issues, we propose a novel sketch condition network specifically designed for the DiT-based video generation architecture (CogVideoX [73] in our work). Following ControlNet [11, 77], we employ a trainable copy of CogVideoX's DiT block to process only the sketch inputs and generate control features. No white placeholder is processed to align with the pretrained weights and reduce learning complexity. To propagate these keyframe features, we design an inter-frame attention mechanism that captures the relationship between the control keyframes and all video frames. Our approach computes query and key features from noisy latent while extracting value features from sketch conditions. This design leverages frame-to-frame similarity for control propagation. The above components consist of a single sketch control block. Instead of copying a half number of the pretrained blocks to construct the control network [11], we use only 5 sketch control blocks out of the 30 DiT blocks available in CogVideoX-2b. We design a novel uniformly distributed skip structure to add the control signals to different levels of features in discrete blocks (0, 6, 12, 18, and 24), achieving effective spatial control while improving memory efficiency. During" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 313, + 72, + 553, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 95 + ], + "type": "text", + "content": "training, an external image dataset is incorporated to solve the challenge of limited video data." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 96, + 553, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 96, + 553, + 215 + ], + "spans": [ + { + "bbox": [ + 313, + 96, + 553, + 215 + ], + "type": "text", + "content": "Beyond generation, interactive editing of real or synthetic videos further enhances creative flexibility. Existing methods [34, 40, 45, 71] achieve interesting text-based editing or effectively propagate single image editing into videos. Despite their effectiveness in appearance modification, they struggle with shape manipulation and object insertion, as they preserve the original temporal motion. Such information is missing for newly introduced content. Moreover, precisely identifying and preserving unedited regions for localized editing remains a challenge." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 216, + 553, + 358 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 216, + 553, + 358 + ], + "spans": [ + { + "bbox": [ + 313, + 216, + 553, + 358 + ], + "type": "text", + "content": "We propose a sketch-based editing method for detailed local modification. Rather than editing a single image and propagating changes, we directly construct an editing network based on our sketch control network. To analyze the relationship between edited regions and the original video, we incorporate a video insertion module that takes the original video with masked edited regions as inputs. The modified sketch control blocks generate residual features that capture temporally and spatially coherent contents in the edited areas. To accurately preserve unedited regions and achieve seamless fusion, the newly edited regions are blended with the original video in the latent space." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 360, + 553, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 360, + 553, + 515 + ], + "spans": [ + { + "bbox": [ + 313, + 360, + 553, + 515 + ], + "type": "text", + "content": "Extensive experiments demonstrate that our method outperforms existing approaches in video generation and editing. Our contributions are summarized as follows: 1) We propose SketchVideo, a novel sketch-based video generation and editing framework that enables detailed geometry control and manipulation using keyframe sketches, as shown in Fig. 1. 2) We design a sketch condition network that predicts skipped control features for the DiT framework, with an inter-frame attention mechanism to propagate one or two sketch conditions across the video. 3) We propose a video insertion module that analyzes the relationship between drawn sketches and original videos, utilizing a latent fusion strategy to preserve unedited regions accurately." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 526, + 400, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 526, + 400, + 538 + ], + "spans": [ + { + "bbox": [ + 314, + 526, + 400, + 538 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 545, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 553, + 713 + ], + "type": "text", + "content": "Sketch-based Image Generation. GAN-based methods [21] have achieved great success in category-restricted sketch-to-image translation [13, 14, 50, 65, 75, 85]. Recently, diffusion-based text-to-image models [51, 53] handle general categories with conditional models like ControlNet [77], T2I-Adapter [41] and further advances [30, 46, 48, 76, 83]. Beyond U-Net, the DiT backbone [12] enables image generation, with pixART-" + }, + { + "bbox": [ + 313, + 545, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 313, + 545, + 553, + 713 + ], + "type": "text", + "content": " [11] utilizing the first half of the pretrained model's blocks as a trainable network to predict corresponding control residual features. Video generation, however, introduces additional challenges. For ease of interaction, we expect sketches specified only for a sparse set of keyframes, making it difficult to generate frames without sketch inputs. Additionally, video genera" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "23380" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "type": "text", + "content": "tion costs significantly higher memory resources, making methods [11, 77] that replicate half of the base model as a sketch encoder easily out of memory." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 108, + 295, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 108, + 295, + 251 + ], + "spans": [ + { + "bbox": [ + 54, + 108, + 295, + 251 + ], + "type": "text", + "content": "Diffusion-based Video Generation. VDM [26] pioneered diffusion-based video generation with a 3D U-Net denoising network. To improve quality, subsequent works [2, 3, 9, 10, 23, 25, 33, 64] integrated temporal modules into text-to-image models [51] to enable text- and image-conditioned video synthesis. Considering the efficiency, the DiT architecture is further used in Sora [43] and open-source projects [35, 60, 84]. Despite these advancements, subtle flickering artifacts remain in the results. CogVideoX [27, 73] further proposes a 3D full attention that merges the spatial and temporal attention, facilitating the generation of long-duration and high-resolution videos." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 252, + 295, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 252, + 295, + 479 + ], + "spans": [ + { + "bbox": [ + 56, + 252, + 295, + 479 + ], + "type": "text", + "content": "Building on existing video generation models, various conditions have been introduced to control the generation, such as camera movement [24, 66, 72], subject identity [31, 68], key-point trajectory [36, 56, 81] and example motions [67, 79]. However, they often lack control over spatial layouts and geometric details. Some methods [15, 18, 80] address this by extending image-based ControlNet [77] to videos but require all-frame conditions that are tedious for sketch interaction. SparseCtrl [22] tackles this by using white images for completions. However, due to the restriction of its base model [23] and a simple network design, its results suffer from temporal flickering. Similar ideas have been applied to cartoon interpolation [70] and colorization [28] with line art as input, but their outputs are limited to cartoon style. Our method utilizes sparse inputs, including hand-drawn sketches on one or two keyframes, to generate temporally stable and realistic videos. Additionally, our method further supports the sketch-based detailed editing of existing videos." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 479, + 295, + 706 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 479, + 295, + 706 + ], + "spans": [ + { + "bbox": [ + 56, + 479, + 295, + 706 + ], + "type": "text", + "content": "Deep Learning-based Video Editing. Pioneer works achieve video editing by style transfer [29, 52], GAN inversion [39, 62], and layered representations [32]. The advent of diffusion models further provides new editing paradigms. One category of such methods leverages image generation models to achieve compelling editing results, using temporal consistency techniques such as layered representations [6, 44], cross-frame attention [5, 17, 19, 57] and pixel warping [71]. A second category of methods employ video generation models. These works propagate the edits applied on the first frame into the other frames [34, 45], or utilize an inpainting strategy to achieve text-based editing [82] and motion modification [40]. To achieve text-based large-scale shape modification while maintaining motion features, a space-time feature loss [74] is designed to guide the inference process. Our method moves beyond traditional text- or image-based editing approaches, allowing users to draw one or two keyframe sketches at arbitrary time points for interactive video editing. Additionally, our method effectively" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "content": "handles sketch-based shape manipulation and dynamic object insertion, which are challenging for previous works." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 96, + 553, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 96, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 313, + 96, + 553, + 262 + ], + "type": "text", + "content": "Controllable Attention Mechanism. Attention across frames is initially used to ensure temporal consistency inAnimateDiff [23] and subsequent video generation models. Subsequent works, such as VideoBooth [31] and Still-Moving [8], utilize it to learn identity-aware features for customized video generation. For video editing, existing works [17, 20, 57] utilize cross-frame attention to capture the temporal motion of input videos, enabling effective propagation of editing operations. Our method applies this idea to pixel-aligned sketch-based video generation and propagates the spatial geometric conditions instead of identity customization. We use a new feature derivation strategy for enhanced spatial control, differentiating our approach from traditional cross-attention mechanisms." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 274, + 397, + 289 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 274, + 397, + 289 + ], + "spans": [ + { + "bbox": [ + 313, + 274, + 397, + 289 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 294, + 553, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 294, + 553, + 437 + ], + "spans": [ + { + "bbox": [ + 313, + 294, + 553, + 437 + ], + "type": "text", + "content": "This section introduces our sketch-based video generation and editing framework. In Sec.3.1, we provide an overview of CogVideoX-2b [73], a pretrained text-to-video generation method, which we will use for sketch-based video generation in our work. In Sec.3.2, we describe our sketch condition network specifically designed for the DiT architecture, which contains sketch control blocks to predict residual features. Within each control block, an interframe attention mechanism is designed to propagate the input sketches. In Sec.3.3, we detail the editing framework, which incorporates a video insertion module and latent fusion to preserve the features of the original video." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 445, + 394, + 459 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 445, + 394, + 459 + ], + "spans": [ + { + "bbox": [ + 313, + 445, + 394, + 459 + ], + "type": "text", + "content": "3.1. Preliminary" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 463, + 553, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 463, + 553, + 594 + ], + "spans": [ + { + "bbox": [ + 313, + 463, + 553, + 594 + ], + "type": "text", + "content": "CogVideoX [73] is a text-to-video generation model that employs a 3D causal VAE and builds diffusion within the latent space. In CogVideoX-2b, the VAE model performs " + }, + { + "bbox": [ + 313, + 463, + 553, + 594 + ], + "type": "inline_equation", + "content": "8 \\times 8" + }, + { + "bbox": [ + 313, + 463, + 553, + 594 + ], + "type": "text", + "content": " spatial and " + }, + { + "bbox": [ + 313, + 463, + 553, + 594 + ], + "type": "inline_equation", + "content": "4 \\times" + }, + { + "bbox": [ + 313, + 463, + 553, + 594 + ], + "type": "text", + "content": " temporal downsampling, followed by a DiT architecture with 30 blocks for video generation. Within each block, a 3D full attention processes the concatenated text embeddings and patchified video latents, followed by a feed-forward layer to output the features. The 3D full attention merges the commonly used separate spatial and temporal attention [2, 23, 84] to improve the temporal coherence. The training objective of diffusion is:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 601, + 553, + 665 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 601, + 553, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 601, + 553, + 665 + ], + "type": "interline_equation", + "content": "L (\\theta) := \\mathrm {E} _ {t, x _ {0} ^ {1: N}, y, \\epsilon} \\left\\| \\epsilon - \\epsilon_ {\\theta} \\left(\\sqrt {\\bar {\\alpha} _ {t}} x _ {0} ^ {1: N} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon , t, y\\right) \\right\\| ^ {2}, \\quad (1) \\text {w h e r e} t \\text {i s s a m p l e d b e t w e e n 1 a n d T (d e n o s i n g s t e p s)}, \\epsilon \\text {i s t h e r a n d o n n o i s e ,} y \\text {i s t e x t p r o m p t s , a n d} x _ {0} ^ {1: N} \\text {i s t h e v i d e o d a w i t h} N \\text {f r a m e s .} \\text {T h e v - p r e d i c t i o n [ 5 4 ] a n d z e r o S N R [ 3 8 ] a r e u l i z a t e d f o r t h e d i f f u s i o n s e t t i n g .}", + "image_path": "38e4c3602e55e61b220d8cbb38bf828b8973baedcb136169705731e8dd6f7324.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 672, + 485, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 672, + 485, + 684 + ], + "spans": [ + { + "bbox": [ + 313, + 672, + 485, + 684 + ], + "type": "text", + "content": "3.2. Sketch-based Video Generation" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": "Given text prompts and one or two keyframe sketches with corresponding time points " + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "inline_equation", + "content": "t_1, t_2" + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": ", our method generates" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23381" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 55, + 70, + 185, + 270 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 185, + 270 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 185, + 270 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 185, + 270 + ], + "type": "image", + "image_path": "fa82a91a15357dd7e658f31b9edbfd2d35d9cecf979b82c59a23a8b8a40a11ed.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 78, + 273, + 167, + 281 + ], + "lines": [ + { + "bbox": [ + 78, + 273, + 167, + 281 + ], + "spans": [ + { + "bbox": [ + 78, + 273, + 167, + 281 + ], + "type": "text", + "content": "(a) Sketch Condition Network" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 187, + 69, + 556, + 271 + ], + "blocks": [ + { + "bbox": [ + 187, + 69, + 556, + 271 + ], + "lines": [ + { + "bbox": [ + 187, + 69, + 556, + 271 + ], + "spans": [ + { + "bbox": [ + 187, + 69, + 556, + 271 + ], + "type": "image", + "image_path": "ae02777334cbdd69680ec09f99ecb89bab661376ac16380c8eb6caef18631e4e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 243, + 273, + 341, + 281 + ], + "lines": [ + { + "bbox": [ + 243, + 273, + 341, + 281 + ], + "spans": [ + { + "bbox": [ + 243, + 273, + 341, + 281 + ], + "type": "text", + "content": "(b) Control Block for Generation" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 430, + 273, + 517, + 281 + ], + "lines": [ + { + "bbox": [ + 430, + 273, + 517, + 281 + ], + "spans": [ + { + "bbox": [ + 430, + 273, + 517, + 281 + ], + "type": "text", + "content": "(c) Control Block for Editing" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 54, + 284, + 555, + 350 + ], + "lines": [ + { + "bbox": [ + 54, + 284, + 555, + 350 + ], + "spans": [ + { + "bbox": [ + 54, + 284, + 555, + 350 + ], + "type": "text", + "content": "Figure 2. Our framework for sketch-based video generation and editing. (a) Our sketch condition network for the DiT-based video generation architecture has a skip structure and five sketch control blocks that predict residual features. (b) For generation, features are extracted from temporally sparse input sketches and propagated through inter-frame attention. The input sketches are provided for one or two keyframes (the second sketch is shown as a dotted line). In the top left corner of (b), the prompt and timestep inputs are shown. (c) For editing, the same sketch control block (b) is utilized, with an additional video insertion module and video masks " + }, + { + "bbox": [ + 54, + 284, + 555, + 350 + ], + "type": "inline_equation", + "content": "M^{1:N}" + }, + { + "bbox": [ + 54, + 284, + 555, + 350 + ], + "type": "text", + "content": " to analyze the relationship between edited and unedited regions. The 3D causal VAE is omitted to save space." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 356, + 295, + 427 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 356, + 295, + 427 + ], + "spans": [ + { + "bbox": [ + 54, + 356, + 295, + 427 + ], + "type": "text", + "content": "video clips that respect the input text prompts and sketches. As shown in Fig. 2, we design a sketch condition network for effective control. The input sketches are encoded into the latent space by the pretrained VAE, followed by patchifying and time-aware positional embedding to generate sketch latents " + }, + { + "bbox": [ + 54, + 356, + 295, + 427 + ], + "type": "inline_equation", + "content": "s_0^{t_1,t_2}" + }, + { + "bbox": [ + 54, + 356, + 295, + 427 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 427, + 295, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 427, + 295, + 559 + ], + "spans": [ + { + "bbox": [ + 54, + 427, + 295, + 559 + ], + "type": "text", + "content": "Skip Residual Structure. In sketch-based image generation, methods like ControlNet [77] and pixART- " + }, + { + "bbox": [ + 54, + 427, + 295, + 559 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 54, + 427, + 295, + 559 + ], + "type": "text", + "content": " [11] copy the half of the base model as their sketch encoder to fully utilize the pretrained text-to-image models. However, applying this to video generation, as done in SparseCtrl [22], is memory inefficient because it requires adding half of the base model's parameters. Unlike the U-Net architecture, the DiT network does not have an explicit encoder and decoder. Therefore, the assumption in pixART" + }, + { + "bbox": [ + 54, + 427, + 295, + 559 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 54, + 427, + 295, + 559 + ], + "type": "text", + "content": " [11] that the first half of blocks serve as the encoder can be improved." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 559, + 295, + 704 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 559, + 295, + 704 + ], + "spans": [ + { + "bbox": [ + 54, + 559, + 295, + 704 + ], + "type": "text", + "content": "Instead of borrowing local consecutive blocks of the base model as the encoder and predicting their residual features, we recognize that blocks at different depths process distinct feature levels, which should be considered for sketch control. As illustrated in Fig. 2 (a), we propose a novel skip residual structure that reduces the number of blocks while enabling effective sketch control and high-quality generation. Our sketch condition network contains 5 sketch control blocks, uniformly distributed across the pretrained generation network to predict residual features for blocks 0, 6, 12, 18, and 24 of the original video generation model. This structure efficiently integrates sketch control informa" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 356, + 555, + 379 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 356, + 555, + 379 + ], + "spans": [ + { + "bbox": [ + 313, + 356, + 555, + 379 + ], + "type": "text", + "content": "tion into multiple feature levels, enhancing the analysis of input conditions and original semantic features." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 380, + 554, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 380, + 554, + 445 + ], + "spans": [ + { + "bbox": [ + 313, + 380, + 554, + 445 + ], + "type": "text", + "content": "Sketch Control Block. In the " + }, + { + "bbox": [ + 313, + 380, + 554, + 445 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 380, + 554, + 445 + ], + "type": "text", + "content": "-th sketch control block, the input consists of hidden video features " + }, + { + "bbox": [ + 313, + 380, + 554, + 445 + ], + "type": "inline_equation", + "content": "h_i^{1:N}" + }, + { + "bbox": [ + 313, + 380, + 554, + 445 + ], + "type": "text", + "content": " and sketch features " + }, + { + "bbox": [ + 313, + 380, + 554, + 445 + ], + "type": "inline_equation", + "content": "s_{i-1}^{t_1,t_2}" + }, + { + "bbox": [ + 313, + 380, + 554, + 445 + ], + "type": "text", + "content": ", and the output is the residual features " + }, + { + "bbox": [ + 313, + 380, + 554, + 445 + ], + "type": "inline_equation", + "content": "\\overline{h}_i^{1:N}" + }, + { + "bbox": [ + 313, + 380, + 554, + 445 + ], + "type": "text", + "content": " and updated sketch features " + }, + { + "bbox": [ + 313, + 380, + 554, + 445 + ], + "type": "inline_equation", + "content": "s_i^{t_1,t_2}" + }, + { + "bbox": [ + 313, + 380, + 554, + 445 + ], + "type": "text", + "content": ". As shown in Fig. 2 (b), the sketch feature " + }, + { + "bbox": [ + 313, + 380, + 554, + 445 + ], + "type": "inline_equation", + "content": "s_{i-1}^{t_1,t_2}" + }, + { + "bbox": [ + 313, + 380, + 554, + 445 + ], + "type": "text", + "content": " is processed:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 370, + 451, + 553, + 468 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 451, + 553, + 468 + ], + "spans": [ + { + "bbox": [ + 370, + 451, + 553, + 468 + ], + "type": "interline_equation", + "content": "s _ {i} ^ {t _ {1}, t _ {2}} = \\operatorname {F e e d F o r w a r d} \\left(s _ {i - 1} ^ {t _ {1}, t _ {2}}\\right), \\tag {2}", + "image_path": "f48c0b828674260dd1676d6a5e5e95f631cc665df77b072bd5de15ece7ccfe07.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 474, + 554, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 474, + 554, + 508 + ], + "spans": [ + { + "bbox": [ + 313, + 474, + 554, + 508 + ], + "type": "text", + "content": "where the output sketch feature " + }, + { + "bbox": [ + 313, + 474, + 554, + 508 + ], + "type": "inline_equation", + "content": "s_i^{t_1,t_2}" + }, + { + "bbox": [ + 313, + 474, + 554, + 508 + ], + "type": "text", + "content": " is used for sketch control propagation and as the input to the next sketch control block." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 312, + 510, + 555, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 510, + 555, + 665 + ], + "spans": [ + { + "bbox": [ + 312, + 510, + 555, + 665 + ], + "type": "text", + "content": "To propagate the sketch inputs, a direct approach [22] replaces missing sketches with white images and employs a trainable copy of the pretrained DiT block to predict residual features. However, as shown in Fig. 5, this leads to fuzzy details in challenging cases, as the network randomly processes both sketches and white images. This mixed input is far from the pretrained model's task. To address the above issue, we employ a trainable copy of pretrained DiT block to process only the sketch inputs (no white placeholder), aligning with the pretrained weights and reducing learning difficulty. The resulting keyframe sketch features, denoted as " + }, + { + "bbox": [ + 312, + 510, + 555, + 665 + ], + "type": "inline_equation", + "content": "c_{i}^{t_{1},t_{2}}" + }, + { + "bbox": [ + 312, + 510, + 555, + 665 + ], + "type": "text", + "content": ", are propagated to all frames through an inter-frame attention approach." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "type": "text", + "content": "Inter-frame Attention. We utilize the input hidden features of all frames to calculate " + }, + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "type": "text", + "content": " and the hidden features corresponding to the control frames to calculate " + }, + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "type": "text", + "content": ". During attention computation, this approach captures the internal" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23382" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 160, + 71, + 217, + 106 + ], + "blocks": [ + { + "bbox": [ + 56, + 72, + 157, + 100 + ], + "lines": [ + { + "bbox": [ + 56, + 72, + 157, + 100 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 157, + 100 + ], + "type": "text", + "content": "A small, fluffy white rabbit hops across a lush meadow at sunset, pausing to listen to rustling leaves before continuing its joyful dance." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 160, + 71, + 217, + 106 + ], + "lines": [ + { + "bbox": [ + 160, + 71, + 217, + 106 + ], + "spans": [ + { + "bbox": [ + 160, + 71, + 217, + 106 + ], + "type": "image", + "image_path": "f48de3b7c0011fd39a4dc48df3b075b3a005fb283187c656cf5ffbe6bea17a4f.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 160, + 110, + 217, + 148 + ], + "blocks": [ + { + "bbox": [ + 56, + 110, + 157, + 137 + ], + "lines": [ + { + "bbox": [ + 56, + 110, + 157, + 137 + ], + "spans": [ + { + "bbox": [ + 56, + 110, + 157, + 137 + ], + "type": "text", + "content": "A panoramic video captures a majestic city with towering skyscrapers, quaint red-tiled homes, and a tranquil lake, all under a sky of drifting white clouds." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 160, + 110, + 217, + 148 + ], + "lines": [ + { + "bbox": [ + 160, + 110, + 217, + 148 + ], + "spans": [ + { + "bbox": [ + 160, + 110, + 217, + 148 + ], + "type": "image", + "image_path": "80a7d9e39c6a1472df2068c1865e29c9435cabd3e62087239525a0aec3a480de.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 160, + 149, + 217, + 187 + ], + "blocks": [ + { + "bbox": [ + 56, + 149, + 157, + 176 + ], + "lines": [ + { + "bbox": [ + 56, + 149, + 157, + 176 + ], + "spans": [ + { + "bbox": [ + 56, + 149, + 157, + 176 + ], + "type": "text", + "content": "A video captures close-ups of a cat's face with green eyes and a brown-black spotted coat, showing three expressions: alert, relaxed, and distant." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 160, + 149, + 217, + 187 + ], + "lines": [ + { + "bbox": [ + 160, + 149, + 217, + 187 + ], + "spans": [ + { + "bbox": [ + 160, + 149, + 217, + 187 + ], + "type": "image", + "image_path": "0d5430ee08d9fa5fcbb547558c2d691da520317b3ca6216470bfe2ec90202af3.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 160, + 189, + 217, + 225 + ], + "blocks": [ + { + "bbox": [ + 56, + 186, + 157, + 226 + ], + "lines": [ + { + "bbox": [ + 56, + 186, + 157, + 226 + ], + "spans": [ + { + "bbox": [ + 56, + 186, + 157, + 226 + ], + "type": "text", + "content": "A video captures the explosive disintegration of a light blue planet, fragments and fiery debris catapulted into starry space, creating a stark contrast between celestial serenity and violent upheaval." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 160, + 189, + 217, + 225 + ], + "lines": [ + { + "bbox": [ + 160, + 189, + 217, + 225 + ], + "spans": [ + { + "bbox": [ + 160, + 189, + 217, + 225 + ], + "type": "image", + "image_path": "a53e56b37cf76b85a6814f637c27a8f2f6672708bc20563a46679693f67b6b8f.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 160, + 227, + 217, + 266 + ], + "blocks": [ + { + "bbox": [ + 56, + 228, + 157, + 262 + ], + "lines": [ + { + "bbox": [ + 56, + 228, + 157, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 228, + 157, + 262 + ], + "type": "text", + "content": "A time-lapse aerial video shows a large, ornate domed building surrounded by lush greenery, transitioning from daylight to sunset to night, illuminated by artificial lights." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 160, + 227, + 217, + 266 + ], + "lines": [ + { + "bbox": [ + 160, + 227, + 217, + 266 + ], + "spans": [ + { + "bbox": [ + 160, + 227, + 217, + 266 + ], + "type": "image", + "image_path": "edf3a8e2b2b30f8e21352ee6db59bbdfa2bb1170cae42fc2d0617b16f0041d31.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 266, + 555, + 300 + ], + "lines": [ + { + "bbox": [ + 54, + 266, + 555, + 300 + ], + "spans": [ + { + "bbox": [ + 54, + 266, + 555, + 300 + ], + "type": "text", + "content": "Figure 3. The sketch-based video generation results. Left: The input text prompts and sketches. Right: The video generation results. It can be seen that the generated results show high quality and good faithfulness with the input sketches. Our method can handle one/two keyframe sketch(es) at arbitrary user-specified time points (the frames corresponding to the input time points are highlighted by orange)." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 222, + 72, + 552, + 108 + ], + "blocks": [ + { + "bbox": [ + 222, + 72, + 552, + 108 + ], + "lines": [ + { + "bbox": [ + 222, + 72, + 552, + 108 + ], + "spans": [ + { + "bbox": [ + 222, + 72, + 552, + 108 + ], + "type": "image", + "image_path": "bd0cfa4dca4f032ce9c1858075f506190674d0a4345fa7931a229cc75a03238b.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 223, + 110, + 552, + 147 + ], + "blocks": [ + { + "bbox": [ + 223, + 110, + 552, + 147 + ], + "lines": [ + { + "bbox": [ + 223, + 110, + 552, + 147 + ], + "spans": [ + { + "bbox": [ + 223, + 110, + 552, + 147 + ], + "type": "image", + "image_path": "306a50698fc2ec85405c297478e8da728c48f4dfd9a69393313dd09e5c8d07dc.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 223, + 149, + 552, + 186 + ], + "blocks": [ + { + "bbox": [ + 223, + 149, + 552, + 186 + ], + "lines": [ + { + "bbox": [ + 223, + 149, + 552, + 186 + ], + "spans": [ + { + "bbox": [ + 223, + 149, + 552, + 186 + ], + "type": "image", + "image_path": "75e25e63b5a2b7892e7807bad4698a960b5e39c4e15088fcae499132ed51f4e5.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 223, + 188, + 552, + 225 + ], + "blocks": [ + { + "bbox": [ + 223, + 188, + 552, + 225 + ], + "lines": [ + { + "bbox": [ + 223, + 188, + 552, + 225 + ], + "spans": [ + { + "bbox": [ + 223, + 188, + 552, + 225 + ], + "type": "image", + "image_path": "d83b67eaf759f5c101ef93700650fbcd77023e58217ed338587ca62a202240ff.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 223, + 226, + 552, + 264 + ], + "blocks": [ + { + "bbox": [ + 223, + 226, + 552, + 264 + ], + "lines": [ + { + "bbox": [ + 223, + 226, + 552, + 264 + ], + "spans": [ + { + "bbox": [ + 223, + 226, + 552, + 264 + ], + "type": "image", + "image_path": "fdc92561323bce7edbf55272dc181f97224a1a62288fc6094d697af1df209447.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "bbox": [ + 55, + 306, + 296, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 306, + 296, + 419 + ], + "spans": [ + { + "bbox": [ + 55, + 306, + 296, + 419 + ], + "type": "text", + "content": "relationship between all frames and control keyframes, allowing propagation of " + }, + { + "bbox": [ + 55, + 306, + 296, + 419 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 55, + 306, + 296, + 419 + ], + "type": "text", + "content": " (derived from the keyframe sketch features). Our inter-frame attention is distinct from typical cross-frame attention, which uses both " + }, + { + "bbox": [ + 55, + 306, + 296, + 419 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 306, + 296, + 419 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 306, + 296, + 419 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 55, + 306, + 296, + 419 + ], + "type": "text", + "content": " from control conditions; instead, we leverage the inter-frame similarity within the input noisy hidden video features and progressively insert the sketches' spatial and temporal information. The output " + }, + { + "bbox": [ + 55, + 306, + 296, + 419 + ], + "type": "inline_equation", + "content": "\\widetilde{c}_i^{1:N}" + }, + { + "bbox": [ + 55, + 306, + 296, + 419 + ], + "type": "text", + "content": " is computed as: " + }, + { + "bbox": [ + 55, + 306, + 296, + 419 + ], + "type": "inline_equation", + "content": "\\mathrm{Attention}(Q,K,V) = \\mathrm{Softmax}(\\frac{QK^T}{\\sqrt{d}})\\cdot V" + }, + { + "bbox": [ + 55, + 306, + 296, + 419 + ], + "type": "text", + "content": ", with" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 66, + 426, + 294, + 441 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 426, + 294, + 441 + ], + "spans": [ + { + "bbox": [ + 66, + 426, + 294, + 441 + ], + "type": "interline_equation", + "content": "Q = W _ {q} \\cdot h _ {i} ^ {1: N}, K = W _ {k} \\cdot h _ {i} ^ {t _ {1}, t _ {2}}, V = W _ {v} \\cdot c _ {i} ^ {t _ {1}, t _ {2}}, \\tag {3}", + "image_path": "520460a88fa6758202e1a0f65842daf225b878f016e4e6158dcb0e0788f722ea.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 55, + 447, + 295, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 447, + 295, + 486 + ], + "spans": [ + { + "bbox": [ + 55, + 447, + 295, + 486 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 447, + 295, + 486 + ], + "type": "inline_equation", + "content": "W_{q}, W_{k}, W_{v}" + }, + { + "bbox": [ + 55, + 447, + 295, + 486 + ], + "type": "text", + "content": " are trainable linear projection weights. The output of the inter-frame attention is fed into a feedforward layer to generate final residual features " + }, + { + "bbox": [ + 55, + 447, + 295, + 486 + ], + "type": "inline_equation", + "content": "\\overline{h}_{i}^{1:N}" + }, + { + "bbox": [ + 55, + 447, + 295, + 486 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 55, + 487, + 296, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 487, + 296, + 582 + ], + "spans": [ + { + "bbox": [ + 55, + 487, + 296, + 582 + ], + "type": "text", + "content": "Training Strategy. To train the sketch condition network, we employ a hybrid training strategy in two stages. In the first stage, to accelerate convergence and address the issue of limited video data, the network is trained both for image generation at arbitrary time points and video generation with one or two keyframe sketches (randomly selected from corresponding sketch videos). In the second stage, video data alone is used to improve the temporal coherence." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 55, + 588, + 208, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 588, + 208, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 588, + 208, + 601 + ], + "type": "text", + "content": "3.3. Sketch-based Video Editing" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "text", + "content": "For editing, given real or synthetic videos, users select one or two keyframes at arbitrary time points and modify the extracted sketches, with additional inputs of text prompts and masks " + }, + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "inline_equation", + "content": "M^{1:N}" + }, + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "text", + "content": " that label regions to be edited for all the frames. Our method then generates realistic, local editing results. The input video is multiplied by the inverted masks to remove information from the edited regions and is subsequently encoded to generate a masked video latent representation " + }, + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "inline_equation", + "content": "v_{0}^{1:N}" + }, + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 306, + 555, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 306, + 555, + 474 + ], + "spans": [ + { + "bbox": [ + 313, + 306, + 555, + 474 + ], + "type": "text", + "content": "Video Insertion Module. For sketch-based editing, newly generated contents within the mask regions should be coherent with the original spatial and temporal features in the unedited regions. Thus, we design a video insertion module that analyzes the relationship between the input sketches and the original video. The video insertion module takes " + }, + { + "bbox": [ + 313, + 306, + 555, + 474 + ], + "type": "inline_equation", + "content": "v_{i-1}^{1:N}" + }, + { + "bbox": [ + 313, + 306, + 555, + 474 + ], + "type": "text", + "content": " (input video latent or generated by a previous control block) as input and predicts the updated video features " + }, + { + "bbox": [ + 313, + 306, + 555, + 474 + ], + "type": "inline_equation", + "content": "v_{i}^{1:N}" + }, + { + "bbox": [ + 313, + 306, + 555, + 474 + ], + "type": "text", + "content": ", similar to the sketch generation process. Since video features are not temporally sparse, we use a trainable copy of CogVideoX-2b's DiT block to directly generate video insertion features " + }, + { + "bbox": [ + 313, + 306, + 555, + 474 + ], + "type": "inline_equation", + "content": "\\widetilde{v}_{i}^{1:N}" + }, + { + "bbox": [ + 313, + 306, + 555, + 474 + ], + "type": "text", + "content": ". The sketch branch output " + }, + { + "bbox": [ + 313, + 306, + 555, + 474 + ], + "type": "inline_equation", + "content": "\\widetilde{c}_{i}^{1:N}" + }, + { + "bbox": [ + 313, + 306, + 555, + 474 + ], + "type": "text", + "content": " and video branch output " + }, + { + "bbox": [ + 313, + 306, + 555, + 474 + ], + "type": "inline_equation", + "content": "\\widetilde{v}_{i}^{1:N}" + }, + { + "bbox": [ + 313, + 306, + 555, + 474 + ], + "type": "text", + "content": " are multiplied by their respective masks and concatenated:" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 356, + 483, + 555, + 500 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 356, + 483, + 555, + 500 + ], + "spans": [ + { + "bbox": [ + 356, + 483, + 555, + 500 + ], + "type": "interline_equation", + "content": "\\operatorname {C o n c a t} \\left(\\tilde {c} _ {i} ^ {1: N} * M ^ {1: N}, \\tilde {v} _ {i} ^ {1: N} * \\overline {{M}} ^ {1: N}\\right), \\tag {4}", + "image_path": "53ca39596aa263b216b72a54a76d356765ce4f300b55c6bb3699e9ae18ad9065.jpg" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 510, + 554, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 554, + 570 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 554, + 570 + ], + "type": "text", + "content": "which serves as inputs to the feed-forward layers, producing final residual features incorporating the original video and sketch control information. This design ensures seamless integration of new contents with the original videos, effectively propagating edits across frames for dynamic motion." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 570, + 554, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 554, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 554, + 689 + ], + "type": "text", + "content": "Training Strategy. Directly training the video editing network would lead to low fidelity with the input sketches, possibly because of the challenging interaction between the input sketches and videos. So we finetune it from the pretrained sketch condition network for generation and add the new video insertion module. The pretrained model already has good sketch fidelity and only requires learning video information. The network is trained in a self-supervised inpainting manner, with randomly generated masks to imitate real-world editing." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "content": "Inference Latent Fusion. Although the original videos are encoded in the condition network, as shown in Fig. 8," + } + ] + } + ], + "index": 27 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23383" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 121, + 71, + 304, + 103 + ], + "blocks": [ + { + "bbox": [ + 121, + 71, + 304, + 103 + ], + "lines": [ + { + "bbox": [ + 121, + 71, + 304, + 103 + ], + "spans": [ + { + "bbox": [ + 121, + 71, + 304, + 103 + ], + "type": "image", + "image_path": "0421acc3ba4c525b9761a4b546353e4a9d37813e9bd607dd4d7d1a41fd37c66a.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 198, + 103, + 228, + 108 + ], + "lines": [ + { + "bbox": [ + 198, + 103, + 228, + 108 + ], + "spans": [ + { + "bbox": [ + 198, + 103, + 228, + 108 + ], + "type": "text", + "content": "Input Video" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 63, + 107, + 112, + 142 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 119, + 105 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 119, + 105 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 119, + 105 + ], + "type": "text", + "content": "A snow-dusted volcano looms, steam plumes rising from its crest. Lush green flanks contrast with rocky slopes, while a rippling lake is in the foreground." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 63, + 107, + 112, + 142 + ], + "lines": [ + { + "bbox": [ + 63, + 107, + 112, + 142 + ], + "spans": [ + { + "bbox": [ + 63, + 107, + 112, + 142 + ], + "type": "image", + "image_path": "1a7f168e94e61a55406f323676cfdd8e2cf771390811c7f1cdd64fc6fd393795.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 121, + 108, + 305, + 140 + ], + "blocks": [ + { + "bbox": [ + 121, + 108, + 305, + 140 + ], + "lines": [ + { + "bbox": [ + 121, + 108, + 305, + 140 + ], + "spans": [ + { + "bbox": [ + 121, + 108, + 305, + 140 + ], + "type": "image", + "image_path": "8ce966540daf86766216580d5ba284b20e25bb0fad3c004b7804e257f47943ba.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 194, + 140, + 229, + 147 + ], + "lines": [ + { + "bbox": [ + 194, + 140, + 229, + 147 + ], + "spans": [ + { + "bbox": [ + 194, + 140, + 229, + 147 + ], + "type": "text", + "content": "Edited Results" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 121, + 148, + 304, + 178 + ], + "blocks": [ + { + "bbox": [ + 56, + 144, + 119, + 184 + ], + "lines": [ + { + "bbox": [ + 56, + 144, + 119, + 184 + ], + "spans": [ + { + "bbox": [ + 56, + 144, + 119, + 184 + ], + "type": "text", + "content": "A mesmerizing ornamental fish glides through the ocean, its electric blue, fiery orange, and iridescent green scales shimmering as it navigates the depths with deli- cate, patterned fins." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 121, + 148, + 304, + 178 + ], + "lines": [ + { + "bbox": [ + 121, + 148, + 304, + 178 + ], + "spans": [ + { + "bbox": [ + 121, + 148, + 304, + 178 + ], + "type": "image", + "image_path": "470bab0b35195b010fc37c9fc5ab3b7299d8f092783bcaa62b2af3c13a6b17d3.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 195, + 178, + 224, + 185 + ], + "lines": [ + { + "bbox": [ + 195, + 178, + 224, + 185 + ], + "spans": [ + { + "bbox": [ + 195, + 178, + 224, + 185 + ], + "type": "text", + "content": "Input Video" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 121, + 186, + 304, + 217 + ], + "blocks": [ + { + "bbox": [ + 121, + 186, + 304, + 217 + ], + "lines": [ + { + "bbox": [ + 121, + 186, + 304, + 217 + ], + "spans": [ + { + "bbox": [ + 121, + 186, + 304, + 217 + ], + "type": "image", + "image_path": "e2c14b5966d29ac8f6b6d044923893a56667fd6514f35305f7b6b25887a534ea.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 193, + 217, + 229, + 224 + ], + "lines": [ + { + "bbox": [ + 193, + 217, + 229, + 224 + ], + "spans": [ + { + "bbox": [ + 193, + 217, + 229, + 224 + ], + "type": "text", + "content": "Edited Results" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 227, + 555, + 261 + ], + "lines": [ + { + "bbox": [ + 55, + 227, + 555, + 261 + ], + "spans": [ + { + "bbox": [ + 55, + 227, + 555, + 261 + ], + "type": "text", + "content": "Figure 4. The sketch-based video editing results. For each example, the text prompts and sketches are shown on the left. On the right, the input real videos are shown at the top, while the edited results with the control keyframe highlighted in orange are shown at the bottom. The editing region masks are manually provided by users, highlighted as orange boxes. Our method generates realistic local editing results." + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 370, + 71, + 553, + 103 + ], + "blocks": [ + { + "bbox": [ + 307, + 71, + 370, + 105 + ], + "lines": [ + { + "bbox": [ + 307, + 71, + 370, + 105 + ], + "spans": [ + { + "bbox": [ + 307, + 71, + 370, + 105 + ], + "type": "text", + "content": "A man is speaking. He has thick hair and is wearing a black suit with a white shirt and a red tie underneath. The camera focuses on his face and upper body." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 370, + 71, + 553, + 103 + ], + "lines": [ + { + "bbox": [ + 370, + 71, + 553, + 103 + ], + "spans": [ + { + "bbox": [ + 370, + 71, + 553, + 103 + ], + "type": "image", + "image_path": "5e00984564c2725f2d35620805f8d6cb69401b77b5ec56ff1a38bb0dada818ad.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 445, + 103, + 474, + 109 + ], + "lines": [ + { + "bbox": [ + 445, + 103, + 474, + 109 + ], + "spans": [ + { + "bbox": [ + 445, + 103, + 474, + 109 + ], + "type": "text", + "content": "Input Video" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 314, + 109, + 362, + 140 + ], + "blocks": [ + { + "bbox": [ + 314, + 109, + 362, + 140 + ], + "lines": [ + { + "bbox": [ + 314, + 109, + 362, + 140 + ], + "spans": [ + { + "bbox": [ + 314, + 109, + 362, + 140 + ], + "type": "image", + "image_path": "176b08af6cbb1a48bff42b9b488de994cd21fdabc132045cad378b012d039e45.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 370, + 109, + 553, + 140 + ], + "blocks": [ + { + "bbox": [ + 370, + 109, + 553, + 140 + ], + "lines": [ + { + "bbox": [ + 370, + 109, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 370, + 109, + 553, + 140 + ], + "type": "image", + "image_path": "fd661b398e226a3da1a66fb2eb9e034b3d0269647eb7003815ec0b72dba6da00.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 445, + 140, + 481, + 147 + ], + "lines": [ + { + "bbox": [ + 445, + 140, + 481, + 147 + ], + "spans": [ + { + "bbox": [ + 445, + 140, + 481, + 147 + ], + "type": "text", + "content": "Edited Results" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 315, + 185, + 362, + 217 + ], + "blocks": [ + { + "bbox": [ + 315, + 185, + 362, + 217 + ], + "lines": [ + { + "bbox": [ + 315, + 185, + 362, + 217 + ], + "spans": [ + { + "bbox": [ + 315, + 185, + 362, + 217 + ], + "type": "image", + "image_path": "7b9da252f21b9adb832a7bb6dd275e242d7063fbbc133f54d0a3b359703c4ae4.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 370, + 148, + 553, + 178 + ], + "blocks": [ + { + "bbox": [ + 309, + 152, + 369, + 175 + ], + "lines": [ + { + "bbox": [ + 309, + 152, + 369, + 175 + ], + "spans": [ + { + "bbox": [ + 309, + 152, + 369, + 175 + ], + "type": "text", + "content": "A green grassland and a small path extending into the distance. The grassland is covered with gravel." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 370, + 148, + 553, + 178 + ], + "lines": [ + { + "bbox": [ + 370, + 148, + 553, + 178 + ], + "spans": [ + { + "bbox": [ + 370, + 148, + 553, + 178 + ], + "type": "image", + "image_path": "5e32331a626647853f09154c9ce0f446d7d42dbb8f0ef957f5a1b0ec976279dc.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 445, + 178, + 474, + 185 + ], + "lines": [ + { + "bbox": [ + 445, + 178, + 474, + 185 + ], + "spans": [ + { + "bbox": [ + 445, + 178, + 474, + 185 + ], + "type": "text", + "content": "Input Video" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 370, + 185, + 553, + 217 + ], + "blocks": [ + { + "bbox": [ + 370, + 185, + 553, + 217 + ], + "lines": [ + { + "bbox": [ + 370, + 185, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 370, + 185, + 553, + 217 + ], + "type": "image", + "image_path": "cb85a421f7ea3062d94ff747c090afe2dfaf333e5b0a37c4d1d6972bbe2ecadb.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 443, + 217, + 479, + 224 + ], + "lines": [ + { + "bbox": [ + 443, + 217, + 479, + 224 + ], + "spans": [ + { + "bbox": [ + 443, + 217, + 479, + 224 + ], + "type": "text", + "content": "Edited Results" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "bbox": [ + 55, + 266, + 296, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 266, + 296, + 361 + ], + "spans": [ + { + "bbox": [ + 55, + 266, + 296, + 361 + ], + "type": "text", + "content": "fine details might be lost during editing. To address this, we propose a latent fusion approach at inference. Specifically, we apply DDIM inversion [58] to generate noisy latent codes of input videos across the inference steps. At Steps 25 and 49 (out of 50 total steps), the latent codes in the unedited regions are replaced with these inversion latent codes, ensuring better preservation of the original video's details and improving the coherence of the edited regions." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 55, + 375, + 127, + 387 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 375, + 127, + 387 + ], + "spans": [ + { + "bbox": [ + 55, + 375, + 127, + 387 + ], + "type": "text", + "content": "4. Evaluation" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 55, + 394, + 188, + 407 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 394, + 188, + 407 + ], + "spans": [ + { + "bbox": [ + 55, + 394, + 188, + 407 + ], + "type": "text", + "content": "4.1. Implementation Details" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 55, + 411, + 296, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 411, + 296, + 544 + ], + "spans": [ + { + "bbox": [ + 55, + 411, + 296, + 544 + ], + "type": "text", + "content": "We implement SketchVideo based on CogVideoX-2b [73], trained on a subset of OpenVid [42] and LAION [55] datasets, with paired sketches from [7]. Training uses 8 NVIDIA H800 GPUs with a batch size of 8 and gradient accumulation of 4. For generation, one or two keyframe sketches are randomly sampled from the video, with 10,000 steps for each training stage. For editing, randomly drawn masks are applied and trained for 20,000 steps. For ease of reading, the input text prompts are simplified in figures. Additional implementation details and full input texts are available in the supplementary material." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 55, + 552, + 112, + 564 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 552, + 112, + 564 + ], + "spans": [ + { + "bbox": [ + 55, + 552, + 112, + 564 + ], + "type": "text", + "content": "4.2. Results" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 55, + 570, + 296, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 296, + 700 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 296, + 700 + ], + "type": "text", + "content": "Our method generates high-quality videos from one or two keyframe sketches and text prompts. As shown in Fig. 3, our method can accurately control the object shape and scene layout, such as the rabbit's pose (1st row) and the position of lakes and buildings (2nd row). Text-only inputs cannot achieve such detailed geometry control. Our method also achieves interesting dynamic motion interpolation and extrapolation, generating smooth and realistic transitions, as seen in the cat's head shaking (3rd row) and the building's movement (last row). This allows control over both spatial layout and dynamic motion." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 67, + 701, + 294, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 701, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 701, + 294, + 713 + ], + "type": "text", + "content": "Our method also supports sketch-based video editing." + } + ] + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 316, + 279, + 360, + 415 + ], + "blocks": [ + { + "bbox": [ + 321, + 264, + 547, + 278 + ], + "lines": [ + { + "bbox": [ + 321, + 264, + 547, + 278 + ], + "spans": [ + { + "bbox": [ + 321, + 264, + 547, + 278 + ], + "type": "text", + "content": "Input Text A time-lapse video shows a large white and black cruise ship with yellow and gold accents docking at a pier on an overcast day, reflecting in calm waters." + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 316, + 279, + 360, + 415 + ], + "lines": [ + { + "bbox": [ + 316, + 279, + 360, + 415 + ], + "spans": [ + { + "bbox": [ + 316, + 279, + 360, + 415 + ], + "type": "image", + "image_path": "3782d4d2d3d1c44994c1f8170531c4dd4d9846734824fd2a0b746c9256fbec22.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 418, + 554, + 495 + ], + "lines": [ + { + "bbox": [ + 313, + 418, + 554, + 495 + ], + "spans": [ + { + "bbox": [ + 313, + 418, + 554, + 495 + ], + "type": "text", + "content": "Figure 5. The comparison results of sketch-based video generation. Text prompts are shown on the top. On the left, we show the input sketches and sketch-based image generation results by ControlNet [77]. On the right, we show the results of the compared approaches, including AMT [37], SparseCtrl [22], CtrlCogVideo [77], and ours. Our method produces better results, especially for the intermediate frames." + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_caption" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 361, + 279, + 408, + 415 + ], + "blocks": [ + { + "bbox": [ + 361, + 279, + 408, + 415 + ], + "lines": [ + { + "bbox": [ + 361, + 279, + 408, + 415 + ], + "spans": [ + { + "bbox": [ + 361, + 279, + 408, + 415 + ], + "type": "image", + "image_path": "87778312df5b6ae3d3dea199f26c39a863dbaea88039bce14c1705e849deed6f.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + } + ], + "index": 33 + }, + { + "type": "image", + "bbox": [ + 408, + 279, + 457, + 415 + ], + "blocks": [ + { + "bbox": [ + 408, + 279, + 457, + 415 + ], + "lines": [ + { + "bbox": [ + 408, + 279, + 457, + 415 + ], + "spans": [ + { + "bbox": [ + 408, + 279, + 457, + 415 + ], + "type": "image", + "image_path": "206cfdfc783968466fdd0c5652a9f0c3fbfe9a29ee83cbcd38ca1ea9dfac7a8d.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 457, + 279, + 505, + 415 + ], + "blocks": [ + { + "bbox": [ + 457, + 279, + 505, + 415 + ], + "lines": [ + { + "bbox": [ + 457, + 279, + 505, + 415 + ], + "spans": [ + { + "bbox": [ + 457, + 279, + 505, + 415 + ], + "type": "image", + "image_path": "9b7c891ab8e925b9925e2d9120e60aadc1ec6cd62dfdbbd1ef97e246d53f1350.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + } + ], + "index": 35 + }, + { + "type": "image", + "bbox": [ + 505, + 279, + 553, + 415 + ], + "blocks": [ + { + "bbox": [ + 505, + 279, + 553, + 415 + ], + "lines": [ + { + "bbox": [ + 505, + 279, + 553, + 415 + ], + "spans": [ + { + "bbox": [ + 505, + 279, + 553, + 415 + ], + "type": "image", + "image_path": "bc01f31a7d84aec396d8364320f31ede572a1673d2d3ec37948f5035649a5a12.jpg" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_body" + } + ], + "index": 36 + }, + { + "bbox": [ + 313, + 504, + 555, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 504, + 555, + 624 + ], + "spans": [ + { + "bbox": [ + 313, + 504, + 555, + 624 + ], + "type": "text", + "content": "Users specify bounding boxes and draw sketches in those regions, and our method generates photorealistic, seamlessly integrated content. As shown in Fig. 4, the new contents blend well with unedited regions and have dynamic motions like hat/hair rotation with original objects (1st row) or interesting fish swimming (2nd row). Our method can handle diverse editing cases, including object insertion, component replacement, and object removal. The unedited regions are well-preserved thanks to our latent fusion approach." + } + ] + } + ], + "index": 38 + }, + { + "bbox": [ + 313, + 635, + 394, + 649 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 635, + 394, + 649 + ], + "spans": [ + { + "bbox": [ + 313, + 635, + 394, + 649 + ], + "type": "text", + "content": "4.3. Comparison" + } + ] + } + ], + "index": 39 + }, + { + "bbox": [ + 313, + 654, + 554, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 554, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 554, + 712 + ], + "type": "text", + "content": "For sketch-based generation, we compare our methods with three methods given the same keyframe sketches text prompts as inputs. For SparseCtrl [22], we use the official pretrained model and extract sketches by HED [69]. It is trained on videos in the WebVid-10M dataset [1] instead" + } + ] + } + ], + "index": 40 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23384" + } + ] + } + ], + "index": 41 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 70, + 105, + 202 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 105, + 202 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 105, + 202 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 105, + 202 + ], + "type": "image", + "image_path": "af9226a58f43b81c4b90e9c8bd415839b91410a0605b507a8cc783a75e58bd28.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 205, + 295, + 281 + ], + "lines": [ + { + "bbox": [ + 55, + 205, + 295, + 281 + ], + "spans": [ + { + "bbox": [ + 55, + 205, + 295, + 281 + ], + "type": "text", + "content": "Figure 6. The comparison results of sketch-based video editing. On the left, we show the drawn editing sketches (for the frames highlighted in orange), text prompts, and sketch-based image generation results by ControlNet [77]. On the right, we show the original videos and editing results by the compared methods, including InsV2V [17], AnyV2V [34], and ours. Our method generates the most realistic results and preserves unedited regions well." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 105, + 70, + 200, + 202 + ], + "blocks": [ + { + "bbox": [ + 105, + 70, + 200, + 202 + ], + "lines": [ + { + "bbox": [ + 105, + 70, + 200, + 202 + ], + "spans": [ + { + "bbox": [ + 105, + 70, + 200, + 202 + ], + "type": "image", + "image_path": "23d15a9e786666483470ae5e5306e23a4c0b373bf2a816544c0f4f4d20ddf362.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 200, + 70, + 246, + 202 + ], + "blocks": [ + { + "bbox": [ + 200, + 70, + 246, + 202 + ], + "lines": [ + { + "bbox": [ + 200, + 70, + 246, + 202 + ], + "spans": [ + { + "bbox": [ + 200, + 70, + 246, + 202 + ], + "type": "image", + "image_path": "1fd7e4aecef8cb49d95e0ffdbc42925b015f079a150042e10474d9fab32d592f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 246, + 70, + 294, + 202 + ], + "blocks": [ + { + "bbox": [ + 246, + 70, + 294, + 202 + ], + "lines": [ + { + "bbox": [ + 246, + 70, + 294, + 202 + ], + "spans": [ + { + "bbox": [ + 246, + 70, + 294, + 202 + ], + "type": "image", + "image_path": "5a4b670008cf4036660feb229fceb7e04dd3f85a134b5af0649ab6290743d1de.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 294, + 295, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 294, + 295, + 437 + ], + "spans": [ + { + "bbox": [ + 54, + 294, + 295, + 437 + ], + "type": "text", + "content": "of the OpenVid-1M dataset [42] used in our method. We extend SparseCtrl to CogVideoX-2b [73], using pixART" + }, + { + "bbox": [ + 54, + 294, + 295, + 437 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 54, + 294, + 295, + 437 + ], + "type": "text", + "content": " [11] as the sketch condition encoder (with 5 DiT blocks same as our method) and white images to complete the missing condition frames. We also compare with an interpolation baseline, which uses ControlNet to translate two-frame sketches into images and then interpolates them with AMT [37]. For the sketch-base editing task, we compare ours with a text-based video editing method InsV2V [17] and a first-frame editing method AnyV2V [34]. We compare with additional methods [16, 20, 45, 70], as shown in supplemental material." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 438, + 295, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 438, + 295, + 617 + ], + "spans": [ + { + "bbox": [ + 55, + 438, + 295, + 617 + ], + "type": "text", + "content": "As shown in Fig. 5, ControlNet [77] translates sketches into realistic images, which, however, lack temporal consistency and vary with shading and content. The interpolation results of AMT [37] exhibit fuzzy details and artifacts. SparseCtrl [22], based onAnimateDiff [23], exhibits temporal flickering, such as the suddenly appearing tree and distorted tower top (see the orange boxes in the 2nd row). Extending SparseCtrl to CogVideoX-2b [73] (Ctrl-CogVideo) still generates fuzzy details in intermediate frames, possibly due to the CogVideoX-2b's pretrained self-attention being designed for dense inputs instead of sparse sketches. Our method generates realistic videos with clear details and good temporal coherence, with even small details like the electric wires in the top-right corner accurately propagated." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 618, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 295, + 713 + ], + "type": "text", + "content": "For sketch-based editing (Fig. 6), InsV2V [17] generates interesting results with birds but lacks control over shape and geometry through text prompts alone. For image-based video editing method, we utilize ControlNet to edit the first frame and then propagate editing into the video by AnyV2V [34]. However, since the motion is borrowed from the original video, AnyV2V struggles with new content, leading to fuzzy details and distortion in the bird, as well" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 315, + 70, + 554, + 123 + ], + "blocks": [ + { + "bbox": [ + 315, + 70, + 554, + 123 + ], + "lines": [ + { + "bbox": [ + 315, + 70, + 554, + 123 + ], + "spans": [ + { + "bbox": [ + 315, + 70, + 554, + 123 + ], + "type": "table", + "html": "
MethodsLPIPS ↓CLIP ↑FidelityConsistencyRealism
AMT29.1796.123.133.513.57
SparseCtrl44.8596.482.792.942.83
Ctrl-CogVideo32.2398.042.862.472.50
Ours27.5698.311.211.081.11
", + "image_path": "203fe8366914e388b86f2d87acfff27a736359a57d9a89997393b9ab21656bac.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 319, + 159, + 550, + 203 + ], + "blocks": [ + { + "bbox": [ + 313, + 125, + 555, + 159 + ], + "lines": [ + { + "bbox": [ + 313, + 125, + 555, + 159 + ], + "spans": [ + { + "bbox": [ + 313, + 125, + 555, + 159 + ], + "type": "text", + "content": "Table 1. The quantitative results of sketch-based video generation comparison. The LPIPS and CLIP numbers are scaled up " + }, + { + "bbox": [ + 313, + 125, + 555, + 159 + ], + "type": "inline_equation", + "content": "100 \\times" + }, + { + "bbox": [ + 313, + 125, + 555, + 159 + ], + "type": "text", + "content": " with each cell colored to indicate the best." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 319, + 159, + 550, + 203 + ], + "lines": [ + { + "bbox": [ + 319, + 159, + 550, + 203 + ], + "spans": [ + { + "bbox": [ + 319, + 159, + 550, + 203 + ], + "type": "table", + "html": "
MethodsLPIPS ↓CLIP ↑PSNR↑FidelityPreservationRealism
InsV2V13.6195.3916.842.582.262.61
AnyV2V11.9293.4713.682.352.692.34
Ours9.7498.3436.481.071.051.04
", + "image_path": "9251cf751fe037fdfd36ef4b07a5987ddc97282279d2532c4d74b359fdea6aaf.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 206, + 553, + 240 + ], + "lines": [ + { + "bbox": [ + 313, + 206, + 553, + 240 + ], + "spans": [ + { + "bbox": [ + 313, + 206, + 553, + 240 + ], + "type": "text", + "content": "Table 2. The quantitative results of sketch-based video editing comparison. The LPIPS and CLIP numbers are scaled up " + }, + { + "bbox": [ + 313, + 206, + 553, + 240 + ], + "type": "inline_equation", + "content": "100 \\times" + }, + { + "bbox": [ + 313, + 206, + 553, + 240 + ], + "type": "text", + "content": " with each cell colored to indicate the best." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 246, + 554, + 305 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 246, + 554, + 305 + ], + "spans": [ + { + "bbox": [ + 313, + 246, + 554, + 305 + ], + "type": "text", + "content": "as changes in unedited regions due to ControlNet's inability to retain the original features. In contrast, our method produces more realistic results with faithful representations of the sketch. More generation and editing results are available in the supplemental material and video." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 306, + 554, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 306, + 554, + 473 + ], + "spans": [ + { + "bbox": [ + 313, + 306, + 554, + 473 + ], + "type": "text", + "content": "We follow SparseCtrl [22] and use the LPIPS [78] metric to measure sketch faithfulness between the input sketches and those extracted from the corresponding frames of the generated videos. We use the CLIP [49] similarity to assess temporal coherence. For sketch-based generation, we test 200 random examples from OpenVid [42], using the first and last frame sketches and the corresponding text prompts in the dataset. As shown in Table 1, our method achieves the lowest LPIPS and highest CLIP scores, indicating its superior performance. For sketch-based editing, we use additional an MSE metric to measure unedited region preservation. We utilize 10 examples with hand-drawn sketches as input. In Table 2, our method outperforms existing methods across all metrics, demonstrating its superiority." + } + ] + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 316, + 473, + 365, + 505 + ], + "blocks": [ + { + "bbox": [ + 316, + 473, + 365, + 505 + ], + "lines": [ + { + "bbox": [ + 316, + 473, + 365, + 505 + ], + "spans": [ + { + "bbox": [ + 316, + 473, + 365, + 505 + ], + "type": "image", + "image_path": "7e39767a6c4d7a8dec04171bf41aca2b61f99a186c996aa3d157ad11523cb0b6.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 365, + 472, + 553, + 548 + ], + "blocks": [ + { + "bbox": [ + 316, + 506, + 365, + 546 + ], + "lines": [ + { + "bbox": [ + 316, + 506, + 365, + 546 + ], + "spans": [ + { + "bbox": [ + 316, + 506, + 365, + 546 + ], + "type": "text", + "content": "Extreme close-up of chicken and green pepper kebabs grilling on a barbeque with flames. Shallow focus and light smoke. vivid colours" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 365, + 472, + 553, + 548 + ], + "lines": [ + { + "bbox": [ + 365, + 472, + 553, + 548 + ], + "spans": [ + { + "bbox": [ + 365, + 472, + 553, + 548 + ], + "type": "image", + "image_path": "ddfec4bcea406a7bacf28349df355d86ba490a12f61030bc4fb88192b7f06abf.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 363, + 547, + 439, + 553 + ], + "lines": [ + { + "bbox": [ + 363, + 547, + 439, + 553 + ], + "spans": [ + { + "bbox": [ + 363, + 547, + 439, + 553 + ], + "type": "text", + "content": "(a) w/o Inter-Frame (b) Sketch K, V" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 447, + 548, + 473, + 553 + ], + "lines": [ + { + "bbox": [ + 447, + 548, + 473, + 553 + ], + "spans": [ + { + "bbox": [ + 447, + 548, + 473, + 553 + ], + "type": "text", + "content": "(c) w/o Skip" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 481, + 548, + 511, + 553 + ], + "lines": [ + { + "bbox": [ + 481, + 548, + 511, + 553 + ], + "spans": [ + { + "bbox": [ + 481, + 548, + 511, + 553 + ], + "type": "text", + "content": "(d) w/o Image" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 525, + 548, + 543, + 552 + ], + "lines": [ + { + "bbox": [ + 525, + 548, + 543, + 552 + ], + "spans": [ + { + "bbox": [ + 525, + 548, + 543, + 552 + ], + "type": "text", + "content": "(e) ours" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 557, + 555, + 590 + ], + "lines": [ + { + "bbox": [ + 313, + 557, + 555, + 590 + ], + "spans": [ + { + "bbox": [ + 313, + 557, + 555, + 590 + ], + "type": "text", + "content": "Figure 7. The ablation study results of sketch-based video generation. Our method generates more realistic and sketch-faithful results than the baselines." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 600, + 408, + 613 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 600, + 408, + 613 + ], + "spans": [ + { + "bbox": [ + 313, + 600, + 408, + 613 + ], + "type": "text", + "content": "4.4. Ablation Study" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 617, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 554, + 713 + ], + "type": "text", + "content": "We conduct an ablation study to evaluate the effectiveness of each key component. Due to the restriction of computing resources, all the models are trained with a batch size of 2. For sketch-based generation, removing inter-frame attention and concatenating sketches with frame features at the temporal dimension results in strange control with low sketch faithfulness (Fig. 7 (a)). Similarly, if we replace the inter-frame attention with a typical cross-attention that uses" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23385" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 69, + 91, + 175 + ], + "blocks": [ + { + "bbox": [ + 56, + 69, + 91, + 175 + ], + "lines": [ + { + "bbox": [ + 56, + 69, + 91, + 175 + ], + "spans": [ + { + "bbox": [ + 56, + 69, + 91, + 175 + ], + "type": "image", + "image_path": "f04947e3d62613f01bff44ede2a40b78f7e46b846a28543899c6c623d7de83a8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 177, + 295, + 221 + ], + "lines": [ + { + "bbox": [ + 55, + 177, + 295, + 221 + ], + "spans": [ + { + "bbox": [ + 55, + 177, + 295, + 221 + ], + "type": "text", + "content": "Figure 8. The ablation study results of sketch-based video editing. Our method generates realistic results faithful to sketches. We utilize heat maps (top right) to visualize the difference between the edited and original frames." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 93, + 70, + 143, + 175 + ], + "blocks": [ + { + "bbox": [ + 93, + 70, + 143, + 175 + ], + "lines": [ + { + "bbox": [ + 93, + 70, + 143, + 175 + ], + "spans": [ + { + "bbox": [ + 93, + 70, + 143, + 175 + ], + "type": "image", + "image_path": "b8ffa9982b0b3559cc3a227f6c2f90c215f40a9de658dccac0b5de3c5afbb0f8.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 144, + 70, + 194, + 175 + ], + "blocks": [ + { + "bbox": [ + 144, + 70, + 194, + 175 + ], + "lines": [ + { + "bbox": [ + 144, + 70, + 194, + 175 + ], + "spans": [ + { + "bbox": [ + 144, + 70, + 194, + 175 + ], + "type": "image", + "image_path": "7632073a2b335d29397a41a1f469fff5898d658f15285867f10464aad61c4d73.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 194, + 70, + 244, + 175 + ], + "blocks": [ + { + "bbox": [ + 194, + 70, + 244, + 175 + ], + "lines": [ + { + "bbox": [ + 194, + 70, + 244, + 175 + ], + "spans": [ + { + "bbox": [ + 194, + 70, + 244, + 175 + ], + "type": "image", + "image_path": "4f01ba7dcd287125674cb5f174ce52e7bba9204b7dd7cafe21d45382d6222977.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 245, + 70, + 294, + 175 + ], + "blocks": [ + { + "bbox": [ + 245, + 70, + 294, + 175 + ], + "lines": [ + { + "bbox": [ + 245, + 70, + 294, + 175 + ], + "spans": [ + { + "bbox": [ + 245, + 70, + 294, + 175 + ], + "type": "image", + "image_path": "a7688278ba32ddf5c598b0b845eafd2a11955a800a67bb6d0aa96deb1fad97b5.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 56, + 222, + 298, + 309 + ], + "blocks": [ + { + "bbox": [ + 56, + 222, + 298, + 309 + ], + "lines": [ + { + "bbox": [ + 56, + 222, + 298, + 309 + ], + "spans": [ + { + "bbox": [ + 56, + 222, + 298, + 309 + ], + "type": "table", + "html": "
Metricw/o Inter-Frame Sketch K,Vw/o Skipw/o ImageOurs
LPIPS ↓36.3332.5931.9134.5830.79
CLIP ↑98.1098.1997.6098.2498.48
Metricw/o Video modulew/o Pretrainw/o Latent FusionOurs
LPIPS ↓9.3112.609.779.74
CLIP ↑97.9798.1298.4498.34
PSNR ↑33.6136.0531.6936.48
", + "image_path": "21f2c6c9298a245b3c6648148699e615be685b6946c9f62d182fa545bc107983.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 311, + 296, + 348 + ], + "lines": [ + { + "bbox": [ + 55, + 311, + 296, + 348 + ], + "spans": [ + { + "bbox": [ + 55, + 311, + 296, + 348 + ], + "type": "text", + "content": "Table 3. The qualitative results of ablation study for sketch-based generation and editing, with each cell colored to indicate the best and second best." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 354, + 295, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 354, + 295, + 496 + ], + "spans": [ + { + "bbox": [ + 54, + 354, + 295, + 496 + ], + "type": "text", + "content": "sketch features for both Key and Value (b), the sketch fidelity is also degraded due to the lack of effective control propagation. If we remove the skip structure and predict residual features for the first 5 blocks (c), the realism of the generated results is negatively influenced due to less control ability. Removing image data during training (d) causes significant geometry mismatch with input sketches. Compared to these alternatives, our method produces the most realistic and faithful results. As shown in Table 3, our method has the lowest LPIPS values (supporting the best sketch fidelity) while the CLIP metric has similar results (human eyes cannot recognize temporal coherence difference)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 498, + 295, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 295, + 665 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 295, + 665 + ], + "type": "text", + "content": "In sketch-based video editing (Fig. 8), removing the video insertion module (a) and solely utilizing the sketch condition encoder results in good sketch fidelity but obvious inconsistency with unedited regions. Training the sketch condition encoder from scratch without generation pretraining (b) reduces the faithfulness to the input sketches in the edited regions, which is also validated by the high LPIPS in Table 3. We use heat maps to reveal differences between edited and original frames. Removing the latent fusion strategy changes unedited region details (c), such as mountains and waves, as shown in the heat maps. The low PSNR value in unedited regions also supports this, as shown in Table 3. Our method maintains strong sketch fidelity, consistency, and unedited region preservation." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 672, + 130, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 672, + 130, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 672, + 130, + 685 + ], + "type": "text", + "content": "4.5. User Study" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "type": "text", + "content": "We conducted a user study to validate that our method outperforms existing approaches. For video generation, we" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 72, + 555, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 228 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 228 + ], + "type": "text", + "content": "randomly selected 10 examples from the test set in Sec. 4.3. In the questionnaire, participants were shown the results of different methods in random order. 20 participants ranked the methods results by three criteria: Sketch Fidelity, Temporal Consistency, and Video Realism. The rankings provided by the participants were used as scores. As shown in Table 1, our method outperforms existing methods in all criteria, demonstrating its superior performance. For sketch-based editing, we used the same 10 examples from the test set in Sec. 4.3. The same participants ranked the methods based on three criteria: Sketch Fidelity, Unedited Region Preservation, and Video Realism. As shown in Table 2, our method also outperforms existing approaches in editing." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 246, + 466, + 258 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 246, + 466, + 258 + ], + "spans": [ + { + "bbox": [ + 313, + 246, + 466, + 258 + ], + "type": "text", + "content": "5. Conclusion and Discussion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 266, + 555, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 266, + 555, + 421 + ], + "spans": [ + { + "bbox": [ + 313, + 266, + 555, + 421 + ], + "type": "text", + "content": "We presented SketchVideo, a unified method for sketch-based video generation and editing. For generation, we proposed a sketch condition network that predicts residual features for skipped DiT blocks of the base models to save the memory and achieve effective control. An inter-frame attention is further proposed to propagate the keyframe sketches, achieving interesting motion interpolation or extrapolation. We also introduce a hybrid image and video training strategy. For editing, we incorporate a video insertion module to ensure the newly generated content is spatially and temporally coherent with the original video. During inference, a latent fusion approach preserves unedited regions accurately." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 422, + 556, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 422, + 556, + 576 + ], + "spans": [ + { + "bbox": [ + 313, + 422, + 556, + 576 + ], + "type": "text", + "content": "Limitation and Future Work. While our method generates high-quality videos, its capabilities are still limited by the pretrained text-to-video model. Enhancing performance with more powerful pretrained models and generating long videos instead of video clips is a potential avenue. Additionally, similar to image generation, our method struggles with too complex scenarios, such as human hands and interaction between multiple objects. Incorporating 3D priors [59, 63, 86] like SMPL-X [47] helps address these issues in human scenarios. Moreover, while our method focuses on geometry control, exploring appearance customization through tools like color strokes presents an intriguing area for future research." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 595, + 424, + 608 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 595, + 424, + 608 + ], + "spans": [ + { + "bbox": [ + 313, + 595, + 424, + 608 + ], + "type": "text", + "content": "6. Acknowledgments" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 613, + 555, + 704 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 613, + 555, + 704 + ], + "spans": [ + { + "bbox": [ + 313, + 613, + 555, + 704 + ], + "type": "text", + "content": "This work was sponsored by CCF-Kuaishou Large Model Explorer Fund, Beijing Municipal Science and Technology Commission (No. Z231100005923031), Innovation Funding of ICT, CAS (No. E461020) and National Natural Science Foundation of China (No. 62322210). The authors would like to acknowledge the Nanjing Institute of InforSuperBahn OneAiNexus for providing the training and evaluation platform." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "23386" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 134 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 134 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 134 + ], + "type": "text", + "content": "[1] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In Int. Conf. Comput. Vis., pages 1708-1718, 2021. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 137, + 296, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 137, + 296, + 201 + ], + "spans": [ + { + "bbox": [ + 61, + 137, + 296, + 201 + ], + "type": "text", + "content": "[2] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, Varun Jampani, and Robin Rombach. Stable video diffusion: Scaling latent video diffusion models to large datasets. CoRR, abs/2311.15127, 2023. 2, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 204, + 295, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 204, + 295, + 258 + ], + "spans": [ + { + "bbox": [ + 62, + 204, + 295, + 258 + ], + "type": "text", + "content": "[3] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In IEEE Conf. Comput. Vis. Pattern Recog., pages 22563-22575, 2023. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 260, + 295, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 260, + 295, + 304 + ], + "spans": [ + { + "bbox": [ + 62, + 260, + 295, + 304 + ], + "type": "text", + "content": "[4] Minwoo Byeon, Beomhee Park, Haecheon Kim, Sungjun Lee, Woonhyuk Baek, and Saehoon Kim. Coyo-700m: Image-text pair dataset. https://github.com/kakaobrain/coyo-dataset, 2022.1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 306, + 294, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 306, + 294, + 338 + ], + "spans": [ + { + "bbox": [ + 62, + 306, + 294, + 338 + ], + "type": "text", + "content": "[5] Duygu Ceylan, Chun-Hao Paul Huang, and Niloy J. Mitra. Pix2video: Video editing using image diffusion. In Int. Conf. Comput. Vis., pages 23149-23160, 2023. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 341, + 294, + 382 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 341, + 294, + 382 + ], + "spans": [ + { + "bbox": [ + 62, + 341, + 294, + 382 + ], + "type": "text", + "content": "[6] Wenhao Chai, Xun Guo, Gaoang Wang, and Yan Lu. Stable-video: Text-driven consistency-aware diffusion video editing. In Int. Conf. Comput. Vis., pages 22983-22993, 2023. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 386, + 294, + 428 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 386, + 294, + 428 + ], + "spans": [ + { + "bbox": [ + 62, + 386, + 294, + 428 + ], + "type": "text", + "content": "[7] Caroline Chan, Frédo Durand, and Phillip Isola. Learning to generate line drawings that convey geometry and semantics. In IEEE Conf. Comput. Vis. Pattern Recog., pages 7915-7925, 2022. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 431, + 294, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 431, + 294, + 485 + ], + "spans": [ + { + "bbox": [ + 62, + 431, + 294, + 485 + ], + "type": "text", + "content": "[8] Hila Chefer, Shiran Zada, Roni Paiss, Ariel Ephrat, Omer Tov, Michael Rubinstein, Lior Wolf, Tali Dekel, Tomer Michaeli, and Inbar Mosseri. Still-moving: Customized video generation without customized video data. ACM Trans. Graph., 43(6):244:1-244:11, 2024. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 488, + 294, + 541 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 488, + 294, + 541 + ], + "spans": [ + { + "bbox": [ + 62, + 488, + 294, + 541 + ], + "type": "text", + "content": "[9] Haoxin Chen, Menghan Xia, Yingqing He, Yong Zhang, Xiaodong Cun, Shaoshu Yang, Jinbo Xing, Yaofang Liu, Qifeng Chen, Xintao Wang, Chao Weng, and Ying Shan. Videocrafter1: Open diffusion models for high-quality video generation. CoRR, abs/2310.19512, 2023. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 544, + 294, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 544, + 294, + 597 + ], + "spans": [ + { + "bbox": [ + 57, + 544, + 294, + 597 + ], + "type": "text", + "content": "[10] Haoxin Chen, Yong Zhang, Xiaodong Cun, Menghan Xia, Xintao Wang, Chao Weng, and Ying Shan. Videographer2: Overcoming data limitations for high-quality video diffusion models. In IEEE Conf. Comput. Vis. Pattern Recog., pages 7310-7320, 2024. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 600, + 294, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 600, + 294, + 643 + ], + "spans": [ + { + "bbox": [ + 57, + 600, + 294, + 643 + ], + "type": "text", + "content": "[11] Junsong Chen, Yue Wu, Simian Luo, Enze Xie, Sayak Paul, Ping Luo, Hang Zhao, and Zhenguo Li. Pixart-δ: Fast and controllable image generation with latent consistency models. CoRR, abs/2401.05252, 2024. 2, 3, 4, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 646, + 294, + 700 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 646, + 294, + 700 + ], + "spans": [ + { + "bbox": [ + 57, + 646, + 294, + 700 + ], + "type": "text", + "content": "[12] Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Zhongdao Wang, James T. Kwok, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart-α: Fast training of diffusion transformer for photorealistic text-to-image synthesis. In Int. Conf. Learn. Represent., 2024. 1, 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 702, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 702, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 702, + 294, + 713 + ], + "type": "text", + "content": "[13] Shu-Yu Chen, Wanchao Su, Lin Gao, Shihong Xia, and" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "Hongbo Fu. Deepfacedrawing: deep generation of face images from sketches. ACM Trans. Graph., 39(4):72, 2020. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 96, + 553, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 148 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 148 + ], + "type": "text", + "content": "[14] Shu-Yu Chen, Feng-Lin Liu, Yu-Kun Lai, Paul L. Rosin, Chunpeng Li, Hongbo Fu, and Lin Gao. Deepfaceediting: deep face generation and editing with disentangled geometry and appearance control. ACM Trans. Graph., 40(4):90:1-90:15, 2021. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 151, + 553, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 151, + 553, + 193 + ], + "spans": [ + { + "bbox": [ + 316, + 151, + 553, + 193 + ], + "type": "text", + "content": "[15] Weifeng Chen, Jie Wu, Pan Xie, Hefeng Wu, Jiashi Li, Xin Xia, Xuefeng Xiao, and Liang Lin. Control-a-video: Controllable text-to-video generation with diffusion models, 2023. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 195, + 553, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 195, + 553, + 249 + ], + "spans": [ + { + "bbox": [ + 316, + 195, + 553, + 249 + ], + "type": "text", + "content": "[16] Xinyuan Chen, Yaohui Wang, Lingjun Zhang, Shaobin Zhuang, Xin Ma, Jiashuo Yu, Yali Wang, Dahua Lin, Yu Qiao, and Ziwei Liu. Seine: Short-to-long video diffusion model for generative transition and prediction. In Int. Conf. Learn. Represent., 2023. 2, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 251, + 553, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 251, + 553, + 281 + ], + "spans": [ + { + "bbox": [ + 316, + 251, + 553, + 281 + ], + "type": "text", + "content": "[17] Jiaxin Cheng, Tianjun Xiao, and Tong He. Consistent video-to-video transfer using synthetic dataset. In Int. Conf. Learn. Represent., 2024. 3, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 283, + 553, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 283, + 553, + 326 + ], + "spans": [ + { + "bbox": [ + 316, + 283, + 553, + 326 + ], + "type": "text", + "content": "[18] Ernie Chu, Shuo-Yen Lin, and Jun-Cheng Chen. Video controlnet: Towards temporally consistent synthetic-to-real video translation using conditional image diffusion models. CoRR, abs/2305.19193, 2023. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 327, + 553, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 327, + 553, + 380 + ], + "spans": [ + { + "bbox": [ + 316, + 327, + 553, + 380 + ], + "type": "text", + "content": "[19] Nathaniel Cohen, Vladimir Kulikov, Matan Kleiner, Inbar Huberman-Spiegelglas, and Tomer Michaeli. Slicedit: Zero-shot video editing with text-to-image diffusion models using spatio-temporal slices. arXiv preprint arXiv:2405.12211, 2024.3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 382, + 553, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 382, + 553, + 415 + ], + "spans": [ + { + "bbox": [ + 316, + 382, + 553, + 415 + ], + "type": "text", + "content": "[20] Michal Geyer, Omer Bar-Tal, Shai Bagon, and Tali Dekel. Tokenflow: Consistent diffusion features for consistent video editing. In Int. Conf. Learn. Represent., 2024. 3, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 416, + 553, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 416, + 553, + 458 + ], + "spans": [ + { + "bbox": [ + 316, + 416, + 553, + 458 + ], + "type": "text", + "content": "[21] Ian J. Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron C. Courville, and Yoshua Bengio. Generative adversarial networks. CoRR, abs/1406.2661, 2014. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 460, + 553, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 460, + 553, + 502 + ], + "spans": [ + { + "bbox": [ + 316, + 460, + 553, + 502 + ], + "type": "text", + "content": "[22] Yuwei Guo, Ceyuan Yang, Anyi Rao, Maneesh Agrawala, Dahua Lin, and Bo Dai. Sparsectrl: Adding sparse controls to text-to-video diffusion models. In Eur. Conf. Comput. Vis., pages 330-348, 2024. 2, 3, 4, 6, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 504, + 553, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 504, + 553, + 557 + ], + "spans": [ + { + "bbox": [ + 316, + 504, + 553, + 557 + ], + "type": "text", + "content": "[23] Yuwei Guo, Ceyuan Yang, Anyi Rao, Zhengyang Liang, Yaohui Wang, Yu Qiao, Maneesh Agrawala, Dahua Lin, and Bo Dai. Animatediff: Imagine your personalized text-to-image diffusion models without specific tuning. In Int. Conf. Learn. Represent., 2024. 3, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 559, + 553, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 559, + 553, + 601 + ], + "spans": [ + { + "bbox": [ + 316, + 559, + 553, + 601 + ], + "type": "text", + "content": "[24] Hao He, Yinghao Xu, Yuwei Guo, Gordon Wetzstein, Bo Dai, Hongsheng Li, and Ceyuan Yang. Cameractrl: Enabling camera control for text-to-video generation. CoRR, abs/2404.02101, 2024. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 603, + 553, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 603, + 553, + 645 + ], + "spans": [ + { + "bbox": [ + 316, + 603, + 553, + 645 + ], + "type": "text", + "content": "[25] Yingqing He, Tianyu Yang, Yong Zhang, Ying Shan, and Qifeng Chen. Latent video diffusion models for high-fidelity video generation with arbitrary lengths. CoRR, abs/2211.13221, 2022. 2, 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "type": "text", + "content": "[26] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. Adv. Neural Inform. Process. Syst., 35:8633-8646, 2022. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "text", + "content": "[27] Wenyi Hong, Ming Ding, Wendi Zheng, Xinghan Liu, and Jie Tang. Cogvideo: Large-scale pretraining for" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23387" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 76, + 73, + 295, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 73, + 295, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 73, + 295, + 95 + ], + "type": "text", + "content": "text-to-video generation via transformers. arXiv preprint arXiv:2205.15868, 2022. 1, 2, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 97, + 295, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 97, + 295, + 129 + ], + "spans": [ + { + "bbox": [ + 56, + 97, + 295, + 129 + ], + "type": "text", + "content": "[28] Zhitong Huang, Mohan Zhang, and Jing Liao. LVCD: reference-based linear art video colorization with diffusion models. CoRR, abs/2409.12960, 2024. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 131, + 294, + 175 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 131, + 294, + 175 + ], + "spans": [ + { + "bbox": [ + 56, + 131, + 294, + 175 + ], + "type": "text", + "content": "[29] Ondrej Jamriska, Sárka Sochorová, Ondrej Texler, Michal Lukác, Jakub Fiser, Jingwan Lu, Eli Shechtman, and Daniel Sykora. Stylizing video by example. ACM Trans. Graph., 38 (4):107:1-107:11, 2019. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 177, + 295, + 221 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 177, + 295, + 221 + ], + "spans": [ + { + "bbox": [ + 56, + 177, + 295, + 221 + ], + "type": "text", + "content": "[30] Rui Jiang, Guang-Cong Zheng, Teng Li, Tian-Rui Yang, Jing-Dong Wang, and Xi Li. A survey of multimodal controllable diffusion models. Journal of Computer Science and Technology, 39(3):509-541, 2024. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 223, + 294, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 223, + 294, + 277 + ], + "spans": [ + { + "bbox": [ + 56, + 223, + 294, + 277 + ], + "type": "text", + "content": "[31] Yuming Jiang, Tianxing Wu, Shuai Yang, Chenyang Si, Dahua Lin, Yu Qiao, Chen Change Loy, and Ziwei Liu. Videobooth: Diffusion-based video generation with image prompts. In IEEE Conf. Comput. Vis. Pattern Recog., pages 6689-6700, 2024. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 280, + 294, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 280, + 294, + 312 + ], + "spans": [ + { + "bbox": [ + 56, + 280, + 294, + 312 + ], + "type": "text", + "content": "[32] Yoni Kasten, Dolev Ofri, Oliver Wang, and Tali Dekel. Layered neural atlases for consistent video editing. ACM Trans. Graph., 40(6):210:1-210:12, 2021. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 315, + 294, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 315, + 294, + 369 + ], + "spans": [ + { + "bbox": [ + 56, + 315, + 294, + 369 + ], + "type": "text", + "content": "[33] Levon Khachatryan, Andranik Movsisyan, Vahram Tadevosyan, Roberto Henschel, Zhangyang Wang, Shant Navasardyan, and Humphrey Shi. Text2video-zero: Text-to-image diffusion models are zero-shot video generators. In Int. Conf. Comput. Vis., pages 15908-15918, 2023. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 371, + 294, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 371, + 294, + 414 + ], + "spans": [ + { + "bbox": [ + 56, + 371, + 294, + 414 + ], + "type": "text", + "content": "[34] Max Ku, Cong Wei, Weiming Ren, Harry Yang, and Wenhu Chen. Anyv2v: A tuning-free framework for any video-to-video editing tasks. arXiv preprint arXiv:2403.14468, 2024. 2, 3, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 417, + 294, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 417, + 294, + 437 + ], + "spans": [ + { + "bbox": [ + 56, + 417, + 294, + 437 + ], + "type": "text", + "content": "[35] PKU-Yuan Lab and Tuzhan AI etc. Open-sora-plan, 2024. 1, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 441, + 294, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 441, + 294, + 483 + ], + "spans": [ + { + "bbox": [ + 56, + 441, + 294, + 483 + ], + "type": "text", + "content": "[36] Yaowei Li, Xintao Wang, Zhaoyang Zhang, Zhouxia Wang, Ziyang Yuan, Liangbin Xie, Yuexian Zou, and Ying Shan. Image conductor: Precision control for interactive video synthesis. CoRR, abs/2406.15339, 2024. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 487, + 294, + 530 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 487, + 294, + 530 + ], + "spans": [ + { + "bbox": [ + 56, + 487, + 294, + 530 + ], + "type": "text", + "content": "[37] Zhen Li, Zuo-Liang Zhu, Ling-Hao Han, Qibin Hou, Chun-Le Guo, and Ming-Ming Cheng. Amt: All-pairs multi-field transforms for efficient frame interpolation. In IEEE Conf. Comput. Vis. Pattern Recog., 2023. 2, 6, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 533, + 294, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 533, + 294, + 576 + ], + "spans": [ + { + "bbox": [ + 56, + 533, + 294, + 576 + ], + "type": "text", + "content": "[38] Shanchuan Lin, Bingchen Liu, Jiashi Li, and Xiao Yang. Common diffusion noise schedules and sample steps are flawed. In IEEE/CVF Winter Conference on Applications of Computer Vision, WACV, pages 5392-5399, 2024. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 578, + 294, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 578, + 294, + 621 + ], + "spans": [ + { + "bbox": [ + 56, + 578, + 294, + 621 + ], + "type": "text", + "content": "[39] Feng-Lin Liu, Shu-Yu Chen, Yu-Kun Lai, Chunpeng Li, Yue-Ren Jiang, Hongbo Fu, and Lin Gao. Deepfacevideoediting: sketch-based deep editing of face videos. ACM Trans. Graph., 41(4):167:1-167:16, 2022. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 624, + 294, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 294, + 667 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 294, + 667 + ], + "type": "text", + "content": "[40] Chong Mou, Mingdeng Cao, Xintao Wang, Zhaoyang Zhang, Ying Shan, and Jian Zhang. Revideo: Remake a video with motion and content control. CoRR, abs/2405.13865, 2024. 2, 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 670, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 294, + 713 + ], + "type": "text", + "content": "[41] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, and Ying Shan. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. In AAAI, pages 4296-4304, 2024. 2" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 117 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 117 + ], + "type": "text", + "content": "[42] Kepan Nan, Rui Xie, Penghao Zhou, Tiehan Fan, Zhenheng Yang, Zhijie Chen, Xiang Li, Jian Yang, and Ying Tai. Openvid-1m: A large-scale high-quality dataset for text-to-video generation. CoRR, abs/2407.02371, 2024. 1, 6, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 119, + 553, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 119, + 553, + 138 + ], + "spans": [ + { + "bbox": [ + 316, + 119, + 553, + 138 + ], + "type": "text", + "content": "[43] OpenAI. Sora overview:https://openai.com/index/soral/, 2024.1,3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 141, + 553, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 553, + 196 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 553, + 196 + ], + "type": "text", + "content": "[44] Hao Ouyang, Qiuyu Wang, Yuxi Xiao, Qingyan Bai, Juntao Zhang, Kecheng Zheng, Xiaowei Zhou, Qifeng Chen, and Qifeng Chen. Codef: Content deformation fields for temporally consistent video processing. In IEEE Conf. Comput. Vis. Pattern Recog., pages 8089-8099, 2024. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 198, + 553, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 198, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 316, + 198, + 553, + 239 + ], + "type": "text", + "content": "[45] Wenqi Ouyang, Yi Dong, Lei Yang, Jianlou Si, and Xingang Pan. I2vedit: First-frame-guided video editing via image-to-video diffusion models. CoRR, abs/2405.16537, 2024. 2, 3, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 243, + 553, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 243, + 553, + 274 + ], + "spans": [ + { + "bbox": [ + 316, + 243, + 553, + 274 + ], + "type": "text", + "content": "[46] Gaurav Parmar, Taesung Park, Srinivasa Narasimhan, and Jun-Yan Zhu. One-step image translation with text-to-image models. CoRR, abs/2403.12036, 2024. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 277, + 553, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 277, + 553, + 331 + ], + "spans": [ + { + "bbox": [ + 316, + 277, + 553, + 331 + ], + "type": "text", + "content": "[47] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed A. A. Osman, Dimitrios Tzionas, and Michael J. Black. Expressive body capture: 3d hands, face, and body from a single image. In IEEE Conf. Comput. Vis. Pattern Recog., 2019. 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 333, + 553, + 375 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 333, + 553, + 375 + ], + "spans": [ + { + "bbox": [ + 316, + 333, + 553, + 375 + ], + "type": "text", + "content": "[48] Bohao Peng, Jian Wang, Yuechen Zhang, Wenbo Li, Ming-Chang Yang, and Jiaya Jia. Controlnext: Powerful and efficient control for image and video generation. CoRR, abs/2408.06070, 2024. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 377, + 553, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 377, + 553, + 442 + ], + "spans": [ + { + "bbox": [ + 316, + 377, + 553, + 442 + ], + "type": "text", + "content": "[49] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763, 2021. 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 445, + 553, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 445, + 553, + 498 + ], + "spans": [ + { + "bbox": [ + 316, + 445, + 553, + 498 + ], + "type": "text", + "content": "[50] Elad Richardson, Yuval Alaluf, Or Patashnik, Yotam Nitzan, Yaniv Azar, Stav Shapiro, and Daniel Cohen-Or. Encoding in style: A stylegan encoder for image-to-image translation. In IEEE Conf. Comput. Vis. Pattern Recog., pages 2287–2296, 2021. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 501, + 553, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 501, + 553, + 544 + ], + "spans": [ + { + "bbox": [ + 316, + 501, + 553, + 544 + ], + "type": "text", + "content": "[51] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In IEEE Conf. Comput. Vis. Pattern Recog., pages 10674-10685, 2022. 1, 2, 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 545, + 553, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 545, + 553, + 578 + ], + "spans": [ + { + "bbox": [ + 316, + 545, + 553, + 578 + ], + "type": "text", + "content": "[52] Manuel Ruder, Alexey Dosovitskiy, and Thomas Brox. Artistic style transfer for videos and spherical images. Int. J. Comput. Vis., 126(11):1199-1219, 2018. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 580, + 553, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 580, + 553, + 655 + ], + "spans": [ + { + "bbox": [ + 316, + 580, + 553, + 655 + ], + "type": "text", + "content": "[53] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L. Denton, Seyed Kamyar Seyed Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, Jonathan Ho, David J. Fleet, and Mohammad Norouzi. Photorealistic text-to-image diffusion models with deep language understanding. In Adv. Neural Inform. Process. Syst., 2022. 1, 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 658, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 553, + 689 + ], + "type": "text", + "content": "[54] Tim Salimans and Jonathan Ho. Progressive distillation for fast sampling of diffusion models. In Int. Conf. Learn. Represent., 2022. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "text", + "content": "[55] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "23388" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 76, + 72, + 295, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 295, + 138 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 295, + 138 + ], + "type": "text", + "content": "Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, Patrick Schramowski, Srivatsa Kundurthy, Katherine Crowson, Ludwig Schmidt, Robert Kaczmarczyk, and Jenia Jitsev. LAION-5B: an open large-scale dataset for training next generation image-text models. In Adv. Neural Inform. Process. Syst., 2022. 1, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 140, + 295, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 140, + 295, + 205 + ], + "spans": [ + { + "bbox": [ + 56, + 140, + 295, + 205 + ], + "type": "text", + "content": "[56] Xiaoyu Shi, Zhaoyang Huang, Fu-Yun Wang, Weikang Bian, Dasong Li, Yi Zhang, Manyuan Zhang, Ka Chun Cheung, Simon See, Hongwei Qin, Jifeng Dai, and Hongsheng Li. Motion-i2v: Consistent and controllable image-to-video generation with explicit motion modeling. In ACM SIGGRAPH, page 111. ACM, 2024. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 206, + 294, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 206, + 294, + 248 + ], + "spans": [ + { + "bbox": [ + 56, + 206, + 294, + 248 + ], + "type": "text", + "content": "[57] Uriel Singer, Amit Zohar, Yuval Kirstain, Shelly Sheynin, Adam Polyak, Devi Parikh, and Yaniv Taigman. Video editing via factorized diffusion distillation. CoRR, abs/2403.09334, 2024. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 251, + 295, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 251, + 295, + 282 + ], + "spans": [ + { + "bbox": [ + 56, + 251, + 295, + 282 + ], + "type": "text", + "content": "[58] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In Int. Conf. Learn. Represent., 2021. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 284, + 295, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 284, + 295, + 316 + ], + "spans": [ + { + "bbox": [ + 56, + 284, + 295, + 316 + ], + "type": "text", + "content": "[59] Jia-Mu Sun, Tong Wu, and Lin Gao. Recent advances in implicit representation-based 3d shape generation. Vis. Intell., 2(1), 2024. 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 317, + 295, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 317, + 295, + 348 + ], + "spans": [ + { + "bbox": [ + 56, + 317, + 295, + 348 + ], + "type": "text", + "content": "[60] Vchitect Team and Shanghai Artificial Intelligence Laboratory. Vchitect-2.0: Parallel transformer for scaling up video diffusion models, 2024. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 350, + 295, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 350, + 295, + 392 + ], + "spans": [ + { + "bbox": [ + 56, + 350, + 295, + 392 + ], + "type": "text", + "content": "[61] Zheng Tianpeng, Chen Yanxiang, Wen Xinzhe, Li Yancheng, and Wang Zhiyuan. Research on diffusion model generated video datasets and detection benchmarks. Journal of Image and Graphics, pages 1-13, 2024. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 394, + 295, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 394, + 295, + 437 + ], + "spans": [ + { + "bbox": [ + 56, + 394, + 295, + 437 + ], + "type": "text", + "content": "[62] Rotem Tzaban, Ron Mokady, Rinon Gal, Amit Bermano, and Daniel Cohen-Or. *Stitch it in time: Gan-based facial editing of real videos*. In SIGGRAPH Asia Conference Papers, pages 29:1–29:9. ACM, 2022. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 438, + 295, + 481 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 438, + 295, + 481 + ], + "spans": [ + { + "bbox": [ + 56, + 438, + 295, + 481 + ], + "type": "text", + "content": "[63] Guangming Wang, Lei Pan, Songyou Peng, Shaohui Liu, Chenfeng Xu, Yanzi Miao, Wei Zhan, Masayoshi Tomizuka, Marc Pollefeys, and Hesheng Wang. Nerf in robotics: A survey. CoRR, abs/2405.01333, 2024. 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 483, + 295, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 483, + 295, + 514 + ], + "spans": [ + { + "bbox": [ + 56, + 483, + 295, + 514 + ], + "type": "text", + "content": "[64] Jiuniu Wang, Hangjie Yuan, Dayou Chen, Yingya Zhang, Xiang Wang, and Shiwei Zhang. Modelscope text-to-video technical report. CoRR, abs/2308.06571, 2023. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 516, + 295, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 516, + 295, + 568 + ], + "spans": [ + { + "bbox": [ + 56, + 516, + 295, + 568 + ], + "type": "text", + "content": "[65] Ting-Chun Wang, Ming-Yu Liu, Jun-Yan Zhu, Andrew Tao, Jan Kautz, and Bryan Catanzaro. High-resolution image synthesis and semantic manipulation with conditional gans. In IEEE Conf. Comput. Vis. Pattern Recog., pages 8798-8807, 2018. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 571, + 295, + 614 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 571, + 295, + 614 + ], + "spans": [ + { + "bbox": [ + 56, + 571, + 295, + 614 + ], + "type": "text", + "content": "[66] Zhouxia Wang, Ziyang Yuan, Xintao Wang, Yaowei Li, Tianshui Chen, Menghan Xia, Ping Luo, and Ying Shan. Motionctrl: A unified and flexible motion controller for video generation. In ACM SIGGRAPH, page 114, 2024. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 616, + 295, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 616, + 295, + 657 + ], + "spans": [ + { + "bbox": [ + 56, + 616, + 295, + 657 + ], + "type": "text", + "content": "[67] Jianzong Wu, Xiangtai Li, Yanhong Zeng, Jiangning Zhang, Qianyu Zhou, Yining Li, Yunhai Tong, and Kai Chen. Motionbooth: Motion-aware customized text-to-video generation. CoRR, abs/2406.17758, 2024. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 658, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 295, + 712 + ], + "type": "text", + "content": "[68] Tao Wu, Yong Zhang, Xintao Wang, Xianpan Zhou, Guangcong Zheng, Zhongang Qi, Ying Shan, and Xi Li. Customcrafter: Customized video generation with preserving motion and concept composition abilities. CoRR, abs/2408.13239, 2024. 3" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 105 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 105 + ], + "type": "text", + "content": "[69] Saining Xie and Zhuowen Tu. Holistically-nested edge detection. In Int. Conf. Comput. Vis., pages 1395-1403, 2015. 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 107, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 150 + ], + "type": "text", + "content": "[70] Jinbo Xing, Hanyuan Liu, Menghan Xia, Yong Zhang, Xintao Wang, Ying Shan, and Tien-Tsin Wong. Tooncrafter: Generative cartoon interpolation. CoRR, abs/2405.17933, 2024. 2, 3, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 152, + 553, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 553, + 195 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 553, + 195 + ], + "type": "text", + "content": "[71] Shuai Yang, Yifan Zhou, Ziwei Liu, and Chen Change Loy. Rerender A video: Zero-shot text-guided video-to-video translation. In SIGGRAPH Asia Conference Papers, pages 95:1-95:11. ACM, 2023. 2, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 198, + 553, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 198, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 316, + 198, + 553, + 251 + ], + "type": "text", + "content": "[72] Shiuyuan Yang, Liang Hou, Haibin Huang, Chongyang Ma, Pengfei Wan, Di Zhang, Xiaodong Chen, and Jing Liao. Direct-a-video: Customized video generation with user-directed camera movement and object motion. In ACM SIGGRAPH, page 113, 2024. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 254, + 553, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 254, + 553, + 307 + ], + "spans": [ + { + "bbox": [ + 316, + 254, + 553, + 307 + ], + "type": "text", + "content": "[73] Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, et al. Cogvideox: Text-to-video diffusion models with an expert transformer. arXiv preprint arXiv:2408.06072, 2024. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 309, + 553, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 309, + 553, + 352 + ], + "spans": [ + { + "bbox": [ + 316, + 309, + 553, + 352 + ], + "type": "text", + "content": "[74] Danah Yatim, Rafail Fridman, Omer Bar-Tal, Yoni Kasten, and Tali Dekel. Space-time diffusion features for zero-shot text-driven motion transfer. In IEEE Conf. Comput. Vis. Pattern Recog., pages 8466-8476, 2024. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 354, + 555, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 354, + 555, + 387 + ], + "spans": [ + { + "bbox": [ + 316, + 354, + 555, + 387 + ], + "type": "text", + "content": "[75] Chen Yuan, Zhao Yang, Zhang Xiaojuan, and Liu Xiaoping. Sketch colorization with finite color space prior. Journal of Image and Graphics, 29(4):978-988, 2024. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 388, + 553, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 388, + 553, + 421 + ], + "spans": [ + { + "bbox": [ + 316, + 388, + 553, + 421 + ], + "type": "text", + "content": "[76] Liang Yuan, Dingkun Yan, Suguru Saito, and Issei Fujishiro. Diffmat: Latent diffusion models for image-guided material generation. Visual Informatics, 8(1):6-14, 2024. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 422, + 553, + 455 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 422, + 553, + 455 + ], + "spans": [ + { + "bbox": [ + 316, + 422, + 553, + 455 + ], + "type": "text", + "content": "[77] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Int. Conf. Comput. Vis., pages 3813-3824, 2023. 2, 3, 4, 6, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 456, + 553, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 456, + 553, + 499 + ], + "spans": [ + { + "bbox": [ + 316, + 456, + 553, + 499 + ], + "type": "text", + "content": "[78] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In IEEE Conf. Comput. Vis. Pattern Recog., pages 586-595, 2018. 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 501, + 553, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 501, + 553, + 544 + ], + "spans": [ + { + "bbox": [ + 316, + 501, + 553, + 544 + ], + "type": "text", + "content": "[79] Yuxin Zhang, Fan Tang, Nisha Huang, Haibin Huang, Chongyang Ma, Weiming Dong, and Changsheng Xu. Motioncrafter: One-shot motion customization of diffusion models. CoRR, abs/2312.05288, 2023. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 546, + 553, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 546, + 553, + 590 + ], + "spans": [ + { + "bbox": [ + 316, + 546, + 553, + 590 + ], + "type": "text", + "content": "[80] Yabo Zhang, Yuxiang Wei, Dongsheng Jiang, Xiaopeng Zhang, Wangmeng Zuo, and Qi Tian. Controlvideo: Training-free controllable text-to-video generation. In Int. Conf. Learn. Represent., 2024. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 591, + 553, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 591, + 553, + 633 + ], + "spans": [ + { + "bbox": [ + 316, + 591, + 553, + 633 + ], + "type": "text", + "content": "[81] Zhenghao Zhang, Junchao Liao, Menghao Li, Long Qin, and Weizhi Wang. Tora: Trajectory-oriented diffusion transformer for video generation. CoRR, abs/2407.21705, 2024. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 635, + 553, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 553, + 690 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 553, + 690 + ], + "type": "text", + "content": "[82] Zhixing Zhang, Bichen Wu, Xiaoyan Wang, Yaqiao Luo, Luxin Zhang, Yinan Zhao, Peter Vajda, Dimitris N. Metaxas, and Licheng Yu. AVID: any-length video inpainting with diffusion model. In IEEE Conf. Comput. Vis. Pattern Recog., pages 7162-7172, 2024. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "text", + "content": "[83] Shihao Zhao, Dongdong Chen, Yen-Chun Chen, Jianmin Bao, Shaozhe Hao, Lu Yuan, and Kwan-Yee K. Wong." + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "23389" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 217 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "text", + "content": "Uni-controlnet: All-in-one control to text-to-image diffusion models. In Adv. Neural Inform. Process. Syst., 2023. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 140 + ], + "type": "text", + "content": "[84] Zangwei Zheng, Xiangyu Peng, Tianji Yang, Chenhui Shen, Shenggui Li, Hongxin Liu, Yukun Zhou, Tianyi Li, and Yang You. Open-sora: Democratizing efficient video production for all, 2024. 1, 2, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 141, + 294, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 141, + 294, + 184 + ], + "spans": [ + { + "bbox": [ + 56, + 141, + 294, + 184 + ], + "type": "text", + "content": "[85] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A. Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In Int. Conf. Comput. Vis., pages 2242-2251, 2017. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 186, + 294, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 186, + 294, + 217 + ], + "spans": [ + { + "bbox": [ + 56, + 186, + 294, + 217 + ], + "type": "text", + "content": "[86] Siting Zhu, Guangming Wang, Dezhi Kong, and Hesheng Wang. 3d gaussian splatting in robotics: A survey. CoRR, abs/2410.12262, 2024. 8" + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "23390" + } + ] + } + ], + "index": 5 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Sketchtopia_ A Dataset and Foundational Agents for Benchmarking Asynchronous Multimodal Communication with Iconic Feedback/fff37445-0c8d-4c24-a3bf-32cf9fcbe8ea_content_list.json b/2025/Sketchtopia_ A Dataset and Foundational Agents for Benchmarking Asynchronous Multimodal Communication with Iconic Feedback/fff37445-0c8d-4c24-a3bf-32cf9fcbe8ea_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..5c717474594dd466bd52935197da4786c239d4b1 --- /dev/null +++ b/2025/Sketchtopia_ A Dataset and Foundational Agents for Benchmarking Asynchronous Multimodal Communication with Iconic Feedback/fff37445-0c8d-4c24-a3bf-32cf9fcbe8ea_content_list.json @@ -0,0 +1,1460 @@ +[ + { + "type": "text", + "text": "Sketchtopia: A Dataset and Foundational Agents for Benchmarking Asynchronous Multimodal Communication with Iconic Feedback", + "text_level": 1, + "bbox": [ + 142, + 130, + 828, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Mohd Hozafa Khan Ravi Kiran Sarvadevabhatla", + "bbox": [ + 277, + 203, + 691, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Centre for Visual Information Technology (CVIT), IIIT Hyderabad, India", + "bbox": [ + 192, + 220, + 774, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{mohd.hozaifa@research., ravi.kiran@}iit.ac.in", + "bbox": [ + 274, + 241, + 691, + 255 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 291, + 313, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We introduce Sketchtopia, a large-scale dataset and AI framework designed to explore goal-driven, multimodal communication through asynchronous interactions in a Pictionary-inspired setup. Sketchtopia captures natural human interactions, including freehand sketches, open-ended guesses, and iconic feedback gestures, showcasing the complex dynamics of cooperative communication under constraints. It features over 20K gameplay sessions from 916 players, capturing 263K sketches, 10K erases, 56K guesses and 19.4K iconic feedbacks. We introduce multimodal foundational agents with capabilities for generative sketching, guess generation and asynchronous communication. Our dataset also includes 800 human-agent sessions for benchmarking the agents. We introduce novel metrics to characterize collaborative success, responsiveness to feedback and inter-agent asynchronous communication. Sketchtopia pushes the boundaries of multimodal AI, establishing a new benchmark for studying asynchronous, goal-oriented interactions between humans and AI agents. The dataset can be found at https://sketchtopia25.github.io/", + "bbox": [ + 73, + 324, + 473, + 628 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 660, + 209, + 676 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Human communication is inherently multimodal and goaldriven, often relying on non-verbal cues such as gestures and sketches to overcome language and cultural barriers [1, 34, 43]. Non-verbal interactions enable shared understanding despite constraints [20, 35]. Understanding such dynamics is key to developing multimodal AI agents capable of achieving shared goals in constrained environments while communicating asynchronously [6, 21].", + "bbox": [ + 75, + 686, + 470, + 808 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "With this motivation, we study communication-constrained asynchronous interactions in a casual game setting. We introduce Sketchtopia, a large-scale multi-modal dataset containing web-based gameplay episodes of Pictionary, the popular cooperative social game. In each episode, a Painter sketches clues while a Gueisser asynchronously", + "bbox": [ + 75, + 810, + 470, + 901 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "attempts to guess the target phrase (Fig. 1). This setup encourages creative, non-verbal communication under time pressure, leveraging iconic feedback gestures (e.g. $\\text{山}$ , and $⑦$ ). The rules of Pictionary forbid the Painter from writing text. Coupled with the time limit, this restriction often leads to creative player responses and generated content.", + "bbox": [ + 496, + 292, + 893, + 383 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Unlike prior works [8, 23], Sketchtopia captures authentic, hand-drawn sketches, feedback gestures, and natural editing actions (e.g., undo strokes). Our dataset not only facilitates the study of multimodal goal-driven communication but also serves as a testbed for novel AI agents. We present foundational agents, DRAWBOT and GUESSBOT, designed to asynchronously engage in multimodal interactions, leveraging both visual and textual modalities. This unique asynchronous communication mirrors real-life scenarios where agents must continuously think and act in parallel rather than turn-taking. We make the following contributions:", + "bbox": [ + 496, + 383, + 895, + 551 + ], + "page_idx": 0 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Sketchtopia: A large-scale, multimodal dataset of Pictionary-like game episodes, capturing asynchronous interactions with sketches, guesses, and gestural feedback.", + "- Two foundational AI agents, DRAWBOT and GUESSBOT, that demonstrate asynchronous, goal-driven multimodal communication with generative capabilities.", + "- Introduction of novel metrics to assess agent performance in terms of game outcomes, individual contributions, and effectiveness in asynchronous multimodal communication." + ], + "bbox": [ + 500, + 555, + 893, + 690 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 500, + 707, + 650, + 724 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Sketch datasets: Existing sketch datasets primarily focus on recognition [13, 40], segmentation [22, 26, 36, 47, 49, 51], and retrieval [53], with labels assigned to the final sketch. In contrast, our dataset includes intermediate text \"guess\" annotations from gameplay, providing temporal grounding for sketches throughout the drawing process. Additionally, our dataset covers a wider array of categories, including abstract concepts (nouns, verbs, adjectives). Some works incorporate text guessing to simulate Pictionary [41], but lack the complexity and richness generated by interacting agents and gameplay dynamics.", + "bbox": [ + 496, + 734, + 893, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "18176", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e4b8a647da52021ce13a75217d1bdde90c76eac112045cb50d8cb961cf51c53e.jpg", + "image_caption": [ + "Figure 1. Multimodal Interaction in Sketchtopia: Timeline of key game elements, showing asynchronous responses from Guesser and Painter. Green tick indicates a correct guess by the Guesser." + ], + "image_footnote": [], + "bbox": [ + 81, + 89, + 890, + 354 + ], + "page_idx": 1 + }, + { + "type": "table", + "img_path": "images/8a60281168ca79abe2f2ff9f863b3b8c10b821b70049a81987d10a7e6a8bfb9b.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ApproachesShared GoalDrawerGueserVisual ContentDraw ContentGuess ContentFeedback (Multi-modal)Content EditingAsync.
Cont.Grd.
Sketchtopia (Ours)✓✓✓✓✓FreehandPhrase✓(C), C, D✓✓
Iconary [8]✓✓✓✗✓Icons (text)Sent.✗✗✗
Co-Draw [23]✓✓✓✗✓ClipArt (text)Sent.✗✗✗
Tell, Draw, Repeat [14]✗✓✗✗✓ImageSent.✗✗✗
Scores [19]✗✗✗✓✓FreehandSent.✗✗✗
Pixellor [7]✗✓✗✓✗Sketches✗✗✗✗
Visual QA✗✗✓✓✓✗Sent.✗✗✗
SketchGues[41]✗✗✓✓✓✗Phrase✗✗✗
Vis. Dialogue Agents [9]✗✗✓✓✓✗Sent.✗✗✗
Coop. Dialogue Agents [11]✓✗✓✓✓✗Sent.✗✗✗
Embodied VQA [11]✓✗✓✓✓✗Sent.✗✗✗
", + "bbox": [ + 76, + 416, + 470, + 503 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Table 1. Agents Comparison: A comparative summary of agent capabilities. Cont.: Continuous sketching; Grd.: Grounding of sketch content; Draw/Guess Content: Agent's output format; Async.: Asynchronous communication support; Sent.: Sentence.", + "bbox": [ + 75, + 515, + 470, + 571 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Cooperative Gameplay: Cooperative games with partial observability (CPO) are often used to study shared-goal communication strategies [5, 12]. Guessing games are a popular subcategory; for example, Ashktorab et al. [2, 3] investigate a game where a 'Giver' conveys a word using text. Similarly, Das et al. [10] focus on image retrieval in a text-based guessing game. Other multimodal versions include Charades, where speech-based guesses correspond to physical enactment [4], and Iconary, where a 'Drawer' uses visual icons to help a 'Guesser' interpret a phrase [8]. Suglia et al. [46] address visual guessing in zero-shot OOV cases for Guess-What and CompGuessWhat benchmarks [45]. Unlike the works mentioned above, Sketchtopia introduces significantly more complex gameplay. It involves contextual and generative free-form sketching, an open-ended guesser vocabulary, iconic feedback, and asynchronous, natural communication between players.", + "bbox": [ + 75, + 598, + 470, + 853 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Communication Agents: Several research efforts in Pictionary-like settings introduce agents that focus solely on sketch generation. Scones [19] introduces agents for itera-", + "bbox": [ + 75, + 854, + 470, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tive sketch creation using detailed textual instructions. Pixel [7] generates sketches efficiently with minimal strokes for a given word. Generally, these works do not involve a guesser in the loop and lack mechanisms for feedback-based sketch adaptation. As a complementary work, Sarvadevabhatla et al. [41] introduce a guesser agent trained to generate human-like guesses in response to iterative rendering of pre-drawn object sketches. However, their approach neglects interactivity and feedback, which are crucial elements of offline Pictionary.", + "bbox": [ + 496, + 419, + 893, + 571 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recent works introduce Teller and Painter for dynamic, cooperative multimodal dialogue or gameplay. In Tell, Draw, Repeat [14], a drawer generates and modifies images iteratively based on detailed linguistic instructions from a teller. Unlike Pictionary, the teller always has access to the final image and neither multimodal interaction nor a long-term goal. mmDialogue [15] employs multi-modal dialogue in terms of both visual and text generation, but the interactions are not goal-oriented. Agents in Iconary [8] mimic drawings with discrete icons and detailed textual interaction from the guesser. In Co-draw [23], the agents engage in goal-driven collaborative communication, where the teller communicates with the drawer in natural language using discrete clip-art icons instead of sketches. The works of Qiu et al. [37] and Lei et al. [25] explore AI-driven emergent communication.", + "bbox": [ + 496, + 577, + 893, + 819 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "LLM Agents: LLMs are increasingly used to test cooperative communication and strategic reasoning in games, focusing on text generation without visual environment processing [24, 27, 50]. REVECA [42] extends this by introducing LLM-based agents interacting in partially observable", + "bbox": [ + 496, + 825, + 895, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "18177", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "table", + "img_path": "images/62d47bc5eb30b3241bd4a5fae8c46520b7590f42287a524c7d68364eb4b88379.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Works# SketchPr. Mod.Lang. Ground.FormEraseStrokesCat.
IntG.Final
Sketchtopia (Ours)263KTxt, Fdb✓✓Scene✓✓200
TU-Berlin [13]20KXX✓ObjX✓250
Quick Draw [17]50MXX✓ObjX✓345
Sketchy [40]75KXX✓ObjX✓125
SketchyScene [57]29KPhX✓ScXX-
ImageNet-Sketch [48]50KXX✓ObjXX1000
Scene Sketcher1225PhX✓Obj,ScX✓14
SketchyCOCO [16]14KPh, EMX✓Obj,ScX✓17
QUML-Shoe [53]419PhX✓ObjXX1
TU-Berlin Extended [55]20KPhX✓ObjXX250
SketchGuess-160 [41]8960Txt✓✓ObjX✓160
", + "bbox": [ + 73, + 88, + 470, + 210 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3D scenes. While Vision Language Models (VLMs) handle multi-turn multimodal dialogues [31], they lack multimodal generative capabilities crucial for Pictionary. Moreover, most LLMs and VLMs operate on turn-based rather than full-duplex communication. Our agents, by contrast, are simpler, communicate asynchronously and are more responsive.", + "bbox": [ + 75, + 316, + 472, + 422 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Dataset", + "text_level": 1, + "bbox": [ + 76, + 438, + 165, + 453 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Game Setup", + "text_level": 1, + "bbox": [ + 76, + 463, + 207, + 479 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Each game session consists of role-specific actions by the two players, viz. Drawer and Guesser. Each session presents a new target word for the Printer to illustrate and the Guesser to speculate. The session ends when the time limit (120 sec) is reached, or when the Printer deems a guess to be correct.", + "bbox": [ + 75, + 486, + 470, + 561 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Data Collection", + "text_level": 1, + "bbox": [ + 76, + 571, + 233, + 585 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The Sketchtopia Web App captures detailed, real-time interactions to emulate the Pictionary experience (see Supplementary video). Data collection involves two phases: Participant Onboarding, where players receive instructions and provide demographic details (age, dexterity), and Gameplay Recording, which logs time-stamped actions, including drawing strokes, iconic feedback, and guesses. Table 3 provides storage formats for collected data. Our dataset supports analysis of both gameplay dynamics and player behavior. Our ethical protocols ensure anonymized data collection, informed consent, and content moderation for privacy and quality.", + "bbox": [ + 75, + 595, + 470, + 762 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.3. Dataset Information", + "text_level": 1, + "bbox": [ + 76, + 771, + 269, + 786 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Target Words: Drawers illustrate words from a pool of 200 terms, ranging from abstract nouns to verbs, covering various difficulty levels. Player Information: Demographics such as age, profession, and dexterity are collected to analyze gameplay strategies and behavior, ensuring fairness and diversity. We also collected 800 human-AI sessions, where players unknowingly interacted with AI agents, allowing", + "bbox": [ + 75, + 794, + 470, + 902 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/45a4bc4117225c4d7674a576216cee2ddc3bca5f68581f00d9d3b2e2f4891eb9.jpg", + "table_caption": [ + "Table 2. Sketch Dataset Comparison::; Img.: Image; Pr. Mod.: Other modalities; Lang. Ground.: Language Grounding; IntG: Intermediate Guesses;Final: Final guess;Txt: Text; Fdb: Feedback;Sc: Scene; Obj: Object; Cat: Categories; Ph: Photo; EM: Edge Map" + ], + "table_footnote": [ + "* SVG denotes white stroke over black background" + ], + "table_body": "
RoleActionFormatTimestamp
DrawerStrokes1SVG2t1→t2
boolt
boolt
SVGt1→t2
(x,y)t1→t2
GuesserGuessTextt
?boolt
", + "bbox": [ + 511, + 88, + 882, + 232 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Table 3. Session Actions and Data Formats", + "bbox": [ + 565, + 261, + 823, + 273 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/18e553ddc6bfd73e14cb6d7a72a92c448f73ab4f42a2d2b2404610841ffe849f.jpg", + "image_caption": [ + "Figure 2. Game Distribution: Painter and Guesser actions distributed across time; sorted in decreasing order of success rate. For successful games, drawer actions preceded the guesser actions and usually last a shorter duration" + ], + "image_footnote": [], + "bbox": [ + 500, + 289, + 890, + 428 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "study of human-AI interaction dynamics. Dataset Quality / Ethical Considerations: We used in-game and post-game text filters to eliminate irrelevant content (e.g. chat-style content such as 'Hi/Hello/lol'). We used text-based filters to remove abusive or inappropriate content in guesses, ensuring a safe and respectful environment. We also conducted manual inspections to identify and address any game rule violations.", + "bbox": [ + 496, + 520, + 893, + 641 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.4. Dataset Analysis", + "text_level": 1, + "bbox": [ + 500, + 650, + 663, + 666 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We analyze Sketchtopia across four key dimensions: global game statistics, session-level performance, player-level insights, and temporal dynamics, revealing gameplay factors that influence success, player behavior, and communication strategies.", + "bbox": [ + 496, + 672, + 893, + 750 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Global game statistics: Our dataset includes 916 players (ages 15–60, avg. 26.6), with 20K sessions totaling 263K strokes, 10K erases, 16.3K highlights, 56K guesses, and 19.4K feedback events. With an $89\\%$ success rate, sessions averaged 13 strokes, 2.83 guesses, 0.82 highlights and 1 feedback event.", + "bbox": [ + 496, + 750, + 893, + 839 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Session-level: Success rates varied by target word types, with nouns and verbs showing higher success compared to abstract nouns and adjectives due to their complexity. As illustrated in Fig. 2, sessions for high-success words often", + "bbox": [ + 496, + 839, + 893, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "18178", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/2d7599f349ca10e9f1ac19a14cd410933ab6413e758fb73cd6be7747818e8dc3.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
OutcomeDFGFDF+GFNo-FB
successful52.330.881.5942.51
failure1.680.180.091.59
", + "bbox": [ + 107, + 88, + 439, + 150 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Table 4. Games categorized across feedback type and outcome. Numbers represent the percentage of games in each category (outcome or presence of iconic feedback).", + "bbox": [ + 75, + 160, + 470, + 202 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "display clear peaks where drawer actions precede guesses, indicating early sketch clarity. Conversely, harder words show flatter distributions, suggesting shifts in strategies or communication challenges. Peaks in the distribution indicate moments where new strategies or ideas emerged, correlating with challenging words. An analysis of success rates across target words and parts of speech (POS) revealed notable variability – Nouns (92.6%), Verbs (84.2%), and Adjectives (79.1%). Among nouns, Abstract Nouns (85%) are the least successful, while food/sports categories are the most. Abstract nouns often require more complex and detailed sketches. Additionally, adjectives (e.g., hard, hollow) which describe nouns are particularly challenging to convey since many objects can possess such properties.", + "bbox": [ + 75, + 227, + 468, + 439 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Player-level insights: At the player level, we identified distinct skills for Drawers and Guessers. Players were slightly more successful as Guessers (14%) than Drawers (9%), yet 77% performed equally well in both roles, emphasizing the shared nature of cooperative gameplay. Categorizing games by feedback and outcome (see Table 4) shows that feedback plays a role in game results. Games that only had drawer feedback had higher success rates, highlighting its importance in guiding Guessers effectively.", + "bbox": [ + 75, + 439, + 468, + 575 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Game Dynamics: To understand gameplay evolution, we analyzed action trajectories using CLIP embeddings [38] to measure similarity between sketches, guesses, and target words. As shown in Figure 3, Drawers refine their sketches over time, while Guessers reactively adjust guesses. Successful games show a convergence pattern where sketches and guesses align closely with the target, demonstrating effective cooperative communication. This game flow analysis captures how players' strategies adapt dynamically throughout the session.", + "bbox": [ + 75, + 575, + 468, + 726 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Agents", + "text_level": 1, + "bbox": [ + 76, + 739, + 158, + 757 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our goal is to design player agents that emulate human-like, responsive gameplay, free from rigid turn-based limitations and capable of multimodal interactions. Beyond deciding 'what' action to perform (e.g., a Roller's stroke), agents must also determine 'when' to act (e.g., when a Gueisser should output a guess). The design should have the flexibility to perform a \"no-op\" (i.e., no immediate action in response to another player's input, allowing the Gueisser to wait for more strokes). Additionally, agents must support not only", + "bbox": [ + 75, + 763, + 468, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "their primary actions (drawing or guessing) but also iconic feedback $(\\text{日}, \\text{日}, \\text{日})$", + "bbox": [ + 498, + 90, + 890, + 121 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To meet these requirements, we organize the agent design into a two-level module hierarchy (see Fig. 4). At the top level, the ACTIONDECIDER module manages global action timing ('when') and selection ('what'). The lower level comprises specialized Painter and Gueisser modules (DRAWBOT and GUESSBOT), which handle role-specific tasks. The following sections detail the architecture, decision-making processes, and interaction mechanisms of each module within this hierarchy.", + "bbox": [ + 496, + 121, + 893, + 257 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. ActionDecider Module", + "text_level": 1, + "bbox": [ + 498, + 268, + 710, + 282 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The ACTIONDECIDER module serves as a lightweight controller, continuously monitoring game states to decide when agents should asynchronously initiate actions. We describe its setup for the Drawing (see Fig. 5). In the first stage, the module processes multimodal game data streams: guesses are encoded using a Text Encoder [54], canvas strokes via a Vision Encoder [54], and iconic feedback through a Feed Forward Network (FFN). These embeddings are cached, with only the latest content appended. A $1 \\times 1$ convolution is applied to the guess, canvas, and feedback embeddings to obtain aggregated vector representations. For the Painter, a vector for the Target Word is also included. These are concatenated into a Game State Vector. In the second stage, the Game State Vector is passed through a 'Response' FFN, which determines the Painter's action: Wake up the Painter, Generate Feedback, or Do Nothing (wait). If Generate Feedback is selected, a 'Feedback' FFN produces specific feedback (Q), for the Guesser.", + "bbox": [ + 496, + 291, + 893, + 577 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The Guesser agent follows a similar design but excludes the Target Word. Here, the wakeup action type is routed to GUESSBOT, and feedback actions correspond to the Guesser's iconic feedback (②).", + "bbox": [ + 496, + 579, + 890, + 638 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Drawer Agent (DRAWBOT)", + "text_level": 1, + "bbox": [ + 498, + 650, + 736, + 666 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The primary task of DRAWBOT is to visually convey the target word through sequential actions: (i) incrementally draw freehand strokes to depict the target word, (ii) update the canvas (e.g., erase using the eraser-icon) based on guesses and feedback (②) from the Guesser, and (iii) generate gestural feedback (①), (iv) based on the semantic proximity of guesses to target word. DRAWBOT is built on a fine-tuned Stable Diffusion model [39], tailored to the game's sketch style. Its architecture extends a Text-to-Image (T2I) Diffusion Model [39] with ControlNet-inspired adapters [56], allowing it to condition sketch generation on multimodal inputs such as canvas state, guess history, and game time (see Fig. 6). Each adapter module provides conditioning based on different input modalities (see Fig. 6):", + "bbox": [ + 496, + 672, + 893, + 883 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "- Canvas State Adapter: A convolutional layer encodes the", + "bbox": [ + 500, + 885, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "18179", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/ed0e3757a9f56ec55ac9c72153e17bfccefdf1408e7151908d856dd1e9ddd8af.jpg", + "image_caption": [ + "Figure 3. Game Flow Analysis: (Sec. 3.4) Each line plot represents a game session. The points on each plot represent CLIP embeddings [38] of canvas or a guess word. The x-axis represents time, y-axis represents normalized distance of these embeddings to the target word (top-left of each panel - blue (easy) and red (hard)). The final canvas for each session is shown. Line plots for paired drawer and guesser have the same color. $\\checkmark$ denotes a successful guess and associated canvas, $\\times$ represents an unsuccessful session." + ], + "image_footnote": [], + "bbox": [ + 81, + 89, + 486, + 476 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/1d6c257e6527289f724ee8664694973db60ff75958746d4c207f8f3f8469e4a4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 493, + 90, + 887, + 474 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/409d0dd424453d191eea625f5b860c873aad7f79873f95aee1e6ceeb40284158.jpg", + "image_caption": [ + "Figure 4. Pictionary Agent Framework: The top-level (ActionDecider) decides action type and timing and delegates to $2^{nd}$ level in the hierarchy which perform the individual tasks. Our hierarchical design enables delegation and asynchronous communication. Refer to Sec. 4 for details." + ], + "image_footnote": [], + "bbox": [ + 86, + 566, + 459, + 651 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "current visual state of the canvas (at $t = t_1$ ). The result is processed by ControlNet's edge detection layer [56]. This adapter modulates sketch generation by incorporating recent updates, ensuring continuity in the drawing process.", + "bbox": [ + 88, + 763, + 468, + 825 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Guess Adapter: Encodes the history of guesses, allowing the agent to adjust its drawings according to the Guesser's proximity to the target.", + "- Time Adapter: A time encoder that regulates the game's pacing and drawing progression, creating a sense of ur" + ], + "bbox": [ + 76, + 825, + 470, + 901 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/6561796846308dabd29a93ec5554b35277dc7fd4fb673677ae065f678d1ea61f.jpg", + "image_caption": [ + "Figure 5. Schematic Diagram of the ActionDecider Module for decision timing and action selection. Refer Sec. 4.1." + ], + "image_footnote": [], + "bbox": [ + 503, + 566, + 890, + 738 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "gency in line with the game's progress.", + "bbox": [ + 513, + 808, + 772, + 823 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We modify the default T2I Adapter [33] to accommodate non-visual modalities as input. Overall, these adapters enable DRAWBOT to generate incremental strokes rather than complete images, producing residual updates to the canvas (see Fig. 6, $t = t_{2}$ ), thereby maintaining continuity while", + "bbox": [ + 496, + 824, + 893, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "18180", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/229b74aecf481b6ea36515d7e2e51fea65134944173bfe698a3c247e735c990a.jpg", + "image_caption": [ + "Figure 6. **Drawer Agent (DrawBot) Architecture:** The top left blue box shows the current state of the canvas. The pink module illustrates the primary Text-to-image-based generation module, which processes the target word (e.g., 'toothbrush') to produce strokes. The green module represents additional conditioning inputs — such as prior guesses and current time — that guide the generative process, enabling the agent to create iterative strokes towards a recognizable sketch. The bottom right orange box shows that conditioning outputs are passed to the encoder of the T2I generator. SD stands for Stable Diffusion (A T2I Model)" + ], + "image_footnote": [], + "bbox": [ + 81, + 89, + 467, + 292 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/27d48b6934678fe9563787620054ba167e2f14c9c5756e7ef47f2ad238fcac0c.jpg", + "image_caption": [ + "Figure 7. Schematic Diagram of the Gueisser Agent's Retrieval-Based Guessing Framework (Sec. 4.3)." + ], + "image_footnote": [], + "bbox": [ + 138, + 459, + 411, + 585 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "dynamically responding to the evolving game state.", + "bbox": [ + 76, + 652, + 416, + 667 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Gueisser Agent (GUESSBOT)", + "text_level": 1, + "bbox": [ + 76, + 679, + 318, + 695 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The Gueisser Agent (GUESSBOT) generates guesses based on the canvas content and multimodal cues, aiming to identify the hidden target word. Given the limited dataset size, GUESSBOT employs a retrieval-based approach. When GUESSBOT receives a 'Wake-up' signal from the ACTION-DECIDER, it encodes the current canvas state $C_t$ (see Fig. 7) and uses Approximate Nearest Neighbor (ANN) search to compare it with historical game sessions. This strategy enables the agent to utilize past gameplay data, applying rule-based criteria to generate guesses aligned with the current game context.", + "bbox": [ + 75, + 703, + 470, + 868 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The retrieval-based process involves two steps: (i) encoding the canvas content using a vision foundation model", + "bbox": [ + 76, + 869, + 470, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "(CLIP) to retrieve 'k' similar sessions from the database, and (ii) applying rule-based filtering based on similarity scores, guess timings and user feedback to refine the results. This filtering discards irrelevant sessions, for example, by rejecting matches where the timing deviates significantly from the current session. The final output is derived from the best-matched session's guesses, producing the selected guess $(g,t)$ as shown in Fig. 7.", + "bbox": [ + 496, + 90, + 893, + 212 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "A key design goal of GUESSBOT is to emulate human gameplay, including natural errors in guessing. For example, human players often overlook spelling errors (e.g., \"spyder\", see Fig. 3) during immersive gameplay. Our retrieval-based approach inherently accommodates these human-like quirks, enhancing the realism of the agent's responses.", + "bbox": [ + 496, + 212, + 893, + 303 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Implementation Details", + "text_level": 1, + "bbox": [ + 498, + 315, + 718, + 330 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Canvas strokes (SVG) are thickened by a factor of 3 and rendered as $224 \\times 224$ images for processing by agents.", + "bbox": [ + 496, + 339, + 890, + 369 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Action Decider: By default, agents remain inactive unless triggered by a 'Wake-up' signal from their Action Decider (see Fig. 4, Fig. 5). Canvas states are captured every 3 seconds, forming the Drawing History. For text encoding (target words, guess history), we use a pre-trained CLIP text encoder [38] (TE), while the vision encoder handles the canvas history (VE). Each FFN is a 3-layer MLP. Every 3 seconds, the Action Decider processes its accumulated inputs and outputs actions.", + "bbox": [ + 496, + 369, + 892, + 505 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "DrawBot: The canvas is processed using ControlNet [56] with StableDiffusion-1.5 [39] as the T2I generator (Fig. 6). The T2I Adapter is modified to handle non-visual inputs: each guess is CLIP-encoded and combined with its times-tamp to form a $512 \\times 2 \\times n_g$ representation, which is integrated into the T2I Generator's U-Net [33]. The generator is fine-tuned on Sketchtopia sketches using LoRA [18] to output incremental strokes (residual images). We optimize the adapters in stages: first training ControlNet, then the T2I Adapters while freezing ControlNet, and finally fin-tuning all components together. The agent is trained by breaking game sessions into fixed time intervals and conditioning outputs on canvas state, guesses, and timestamps, allowing for adaptive sketching based on all input modalities. The residual image output by T2I Generator is converted to its SVG representation using virtual sketching [32], before being added to the canvas.", + "bbox": [ + 496, + 507, + 893, + 763 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "GuessBot: A vector database of canvas embeddings (using CLIP ViT-B/32) is built with metadata (guess and timestamp) for fast retrieval using Annoy [44], an approximate nearest-neighbor retriever. Upon receiving a 'Wake-Up' signal, GUESSBOT encodes the current canvas $e_{C_t}$ (see Fig. 7) at current time $t$ , and retrieves the top 10 matches within a time window $[t - \\delta, t + \\delta]$ , where $\\delta = 6$ . Among these, we select the match closest to the query by cosine similarity and output the corresponding guess word.", + "bbox": [ + 496, + 763, + 893, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "18181", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Agent Orchestration: All agents are deployed as microservices communicating with the Pictionary WebApp via API calls. Participants are paired randomly, with a $15\\%$ chance of getting paired with an agent. During gameplay, the WebApp sends relevant data to the ActionDecider (Sec. 4.1), which delegates tasks to draw or feedback as needed.", + "bbox": [ + 75, + 90, + 472, + 183 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5. Metrics for Benchmarking Agents", + "text_level": 1, + "bbox": [ + 76, + 194, + 388, + 212 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We evaluate agents (human-only games and mixed human-agent) in terms of their collaborative success, their responsiveness to feedback, and the similarity of their timing patterns to human counterparts.", + "bbox": [ + 75, + 219, + 470, + 279 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Collaborative Game Success: In goal-oriented games, the primary measure is whether Painter and Guesser agents can effectively collaborate to identify the target word. Metrics include Win Rate (WR): The proportion of games where the target word is guessed correctly, serving as a baseline measure of collaborative success [8]; Soft Win Rate (SWR): Extends Win Rate by considering synonyms of target word, capturing near-success scenarios where agents were close to the correct answer [8, 41];", + "bbox": [ + 75, + 280, + 470, + 415 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Feedback Responsiveness Score (FRS): measures the extent to which the Guesser adjusts its guesses in response to Painter feedback. After receiving or feedback, if the guesser's next guess is semantically closer to the target word than the guess immediately preceding the feedback, such an occurrence is termed a responsive guess. In a game session, let the count of such occurrences be $s$ , and the Painter takes $p$ and $n$ actions. FRS for that session is defined as $\\frac{s}{p + n}$ . The resulting numbers are averaged across sessions to obtain the final metric.", + "bbox": [ + 75, + 416, + 470, + 566 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Asynchronous Action Overlap (AAO): This is a session-level measure and characterizes the degree to which agents act asynchronously. Suppose a Gueisser action's timestamp overlaps with the duration of a Drawer action (e.g. sketching, erasing) within a session. Let the count of all occurrences be $o$ , and let the total number of Gueisser actions be $g$ . AAO for the session is defined as $AAO = \\frac{o}{g}$ . Note that a sequential turn-based interaction between agents would have an AAO value of 0, while a purely random and uncoordinated interaction would have a value of 100.", + "bbox": [ + 75, + 566, + 470, + 715 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Multimodal Action Timing Similarity (MATS): For a given target word, suppose the sequence of human Guesser's guess timestamps across all sessions be $T_{g}^{H}$ , and a similar sequence for GUESSBOT be $T_{g}^{A}$ . We define a guess distance $d_{guess}$ between these two sequences as the Jensen-Shannon divergence [28] between their Kernel Density estimates, i.e. $d_{guess} = \\text{JENSEN-SHANNON}\\{\\text{KDE}(T_{g}^{H}), \\text{KDE}(T_{g}^{A})\\}$ . The average $d_{guess}$ across all target words forms the MATS metric for GUESSBOT. A similar exercise for DRAWBOT's actions, where the timestamp of the stroke duration's midpoint is used, provides the MATS score for DRAWBOT. Detailed definitions for the metrics are provided in the Supplementary.", + "bbox": [ + 75, + 717, + 472, + 902 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a5ee74fc0add239c475ae5bebdba166623d57dd658b1560c5b6934171d32273b.jpg", + "image_caption": [ + "Figure 8. Qualitative comparison between human and agent game outcomes: The figure shows the final canvas and the final guess for select target words (fishing, length, loud). Human outputs are in purple and agent outputs are in blue. Successful games are with green background and unsuccessful ones in red." + ], + "image_footnote": [], + "bbox": [ + 506, + 90, + 890, + 339 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6. Results", + "text_level": 1, + "bbox": [ + 500, + 446, + 586, + 463 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To benchmark our agents, we carefully choose a representative evaluation set of game sessions grouped by Part of Speech and compare the performance of humans and agents on this set. In general, the metric values for humans act as ground truth or desirable range, which the agent should achieve. We obtain 800 sessions to be used for testing. The numbers mentioned in Tab. 5 for HH (Human-Human) and HA (Human-Agent) are both computed on this set.", + "bbox": [ + 496, + 473, + 893, + 594 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Collaborative Game Success: Across games, agents had a $68\\%$ success rate. The win rate for GUESSBOT is slightly lesser than the human counterpart $(85\\%)$ . However, DRAWBOT's win rate is $51\\%$ , which is markedly smaller. The Painter agent's bad performance is mainly due to adjectives $(30\\%)$ and verbs $(56\\%)$ . It performs reasonably well on most nouns $(74\\%)$ . The primary reason for bad performance on adjectives is that there is no direct object (e.g. car, bus, icecream) associated with adjectives. Most vision foundation models are trained on objects rather than metaphorical concepts which encode attributes (e.g. hard/soft, length, hollow and loud) which may be associated with multiple objects. Such models need detailed prompts to invoke the relationship and properties of objects.", + "bbox": [ + 496, + 595, + 893, + 808 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Qualitative analysis: Many interesting patterns can be observed in the performance and response patterns of human-human and human-agent games – see Fig. 8. Compared to human counterparts, DRAWBOT's sketches have greater symmetry, with clean stroke lines. Unlike typical generative model outputs, which tend to have a 'synthetic' appearance,", + "bbox": [ + 496, + 810, + 895, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "18182", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "training on Sketchtopia's data induces a human-like jitter in the generated strokes of DRAWBOT. GUESSBOT's final responses seem human-like for the corresponding canvas contents, even when the game outcome is a failure. This is consistent with the agent's retrieval-based design. Additional examples can be found in Supplementary.", + "bbox": [ + 75, + 90, + 468, + 181 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Feedback Responsiveness: The FRS scores for Human and GUESSBOT are similar (Tab. 5), implying that the agent's ability to adapt its response to feedback is similar to human counterparts. Here, the Gueisser Agent's ActionDecider also plays a key role in 'waking up' GUESSBOT at the right time in response to Painter feedback.", + "bbox": [ + 75, + 181, + 470, + 271 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Characterizing Interaction Dynamics: Similar to FRS, the AAO scores of agents are comparable to humans, indicating comparable levels of asynchronous communication. However, GUESSBOT has a lower AAO score. This is in line with a Guesser's tendency to wait passively for enough strokes to accumulate on the canvas before guessing. The MATS score measures the agent's action pacing compared to humans. Therefore, it is a proxy measure for ActionDecider's performance due to the latter's role as a key asynchronous enabler of actions in each agent. A relatively lower MATS score for GUESSBOT suggests that its action pacing is closer to human. Empirically, we observed that feedback distribution for DRAWBOT was markedly different from human, contributing to its relatively higher MATS score.", + "bbox": [ + 75, + 272, + 470, + 483 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "VLMs as Guessers: Vision-Language Models (VLMs) [29, 30, 52] have become a popular choice for image-conditioned reasoning. In our benchmarking context, VLMs could play a Guesser's role. However, despite detailed prompts instructing the model to output a guess phrase and despite providing prior guess history, we observed that VLM models tend to produce verbose and unhelpful outputs. Examples can be found in Supplementary. Moreover, the slower response times and synchronized (turn-taking) paradigm of current VLMs make them unviable compared to our asynchronous and contextually responsive agents.", + "bbox": [ + 75, + 484, + 470, + 648 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations: Despite efforts to minimize biases, the dataset may still reflect factors such as age, handedness, and cultural background, with a predominance of younger, right-handed, English-speaking participants from a limited region. This may affect its generalizability to other languages, cultures, and demographics. While our game interface supports asynchronous interactions, further testing across diverse groups is needed. Additionally, while Sketchtopia includes iconic feedback as a novelty, it is limited to predefined icons (e.g., , , and Ⓞ). This restricts the analysis of more nuanced and complex feedback that could occur in natural interactions.", + "bbox": [ + 75, + 650, + 470, + 815 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Conclusion", + "text_level": 1, + "bbox": [ + 76, + 830, + 194, + 844 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Sketchtopia enables a scalable framework that pushes the boundaries of multimodal goal-driven communication in asynchronous settings. As the first-ever dataset to include", + "bbox": [ + 75, + 854, + 468, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/a36d0d443ed5af0e956dafe4b3e8e8a0fda0d6cdcfc746b4049d4d3cf32e91a7.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelsGame SuccessFeedback ResponsivenessInteraction Dynamics
Win ↑Soft Win ↑FRS ↑AAOMATS ↓
DRAWBOT51.652.4-0.210.39
GUESSBOT81.786.046%0.170.27
Human85.593.440%0.290.00
", + "bbox": [ + 501, + 88, + 890, + 152 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. Benchmarking humans and agents via custom-designed metrics (Sec. 5). FRS: Feedback Responsiveness Score; AAO: Asynchronous Action Overlap; MATS: Multimodal Action Timing Similarity.", + "bbox": [ + 498, + 162, + 893, + 219 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "rich feedback signals, our collection captures realistic gameplay data with natural human interactions, far beyond what existing datasets offer. Sketchtopia is not just about isolated sketches – it incorporates iconic feedback and dynamic interactions between players, all captured using a robust game interface that seamlessly supports both human-human and human-agent gameplay. This versatility allows us to explore interactions at a depth previously unattainable, establishing a new benchmark for multimodal AI systems.", + "bbox": [ + 496, + 266, + 893, + 402 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "A standout contribution of our work is the design of first asynchronous communication agents capable of real-time decision-making in multi-modal formats. Our novel agent architectures emphasize responsiveness and speed, enabling agents to mimic the fluidity of natural human communication without the constraints of turn-taking. By introducing innovative metrics such as Asynchronous Action Overlap and Feedback Responsiveness Scores, we redefine how to evaluate collaborative success and interaction dynamics in asynchronous multi-modal communication settings.", + "bbox": [ + 496, + 412, + 893, + 564 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Benchmarking our agents against human players demonstrates that they can generate human-like actions and responses. Our results reveal that even state-of-the-art VLMs lag significantly in handling asynchronous, multimodal communication complexities. This establishes Sketchtopia as a critical benchmark for advancing AI research in this space. Additionally, our game flow visualizations offer a novel lens into understanding how agents and humans progress towards shared goals, providing insights into the evolution of strategies and collaborative problem-solving.", + "bbox": [ + 496, + 573, + 895, + 724 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Sketchtopia's applications extend beyond Pictionary. Its interactive features enable sketch scene editing and collaborative whiteboard tools, with potential for live note-taking and narration. The dataset's focus on abstract sketches also supports cognitive science research into communication strategies. In summary, Sketchtopia provides a valuable resource for exploring asynchronous, multimodal AI systems and realistic agent-based gameplay. By tackling the complexities of real-time, feedback-driven communication, we aim to inspire AI advancements that are more natural, adaptive, and truly collaborative.", + "bbox": [ + 496, + 734, + 895, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "18183", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Michael Argyle. The syntaxes of bodily communication. Linguistics, 11(112):71-92, 1973. 1", + "[2] Zahra Ashktorab, Q. Vera Liao, Casey Dugan, James Johnson, Qian Pan, Wei Zhang, Sadhana Kumaravel, and Murray Campbell. Human-ai collaboration in a cooperative game setting: Measuring social perception and outcomes. Proc. ACM Hum.-Comput. Interact., 4(CSCW2), 2020. 2", + "[3] Zahra Ashktorab, Casey Dugan, James Johnson, Qian Pan, Wei Zhang, Sadhana Kumaravel, and Murray Campbell. Effects of communication directionality and ai agent differences in human-ai interaction. In Proceedings of the 2021 CHI Conference on Human Factors in Computing Systems, New York, NY, USA, 2021. Association for Computing Machinery. 2", + "[4] Muhammad Fauzi Bafadal and Humaira Humaira. The use of charades games in teaching vocabulary to the junior high school students. Linguistics and Elt Journal, 2019. 2", + "[5] Nolan Bard, Jakob N. Foerster, Sarath Chandar, Neil Burch, Marc Lanctot, H. Francis Song, Emilio Parisotto, Vincent Dumoulin, Subhodeep Moitra, Edward Hughes, Iain Dunning, Shibl Mourad, Hugo Larochelle, Marc G. Bellemare, and Michael Bowling. The hanabi challenge: A new frontier for ai research. Artificial Intelligence, 280:103216, 2020. 2", + "[6] Fabien Beltrame. Comportement non verbal et situation d'interaction: Quand la gestuelle des mains revendique un role cognitif de support des représentations pour la résolution de problème, 2006. 1", + "[7] Ayan Kumar Bhunia, Ayan Das, Umar Riaz Muhammad, Yongxin Yang, Timothy M. Hospedales, Tao Xiang, Yulia Gryaditskaya, and Yi-Zhe Song. Pixel: a competitive sketching ai agent. so you think you can sketch? ACM Trans. Graph., 39(6), 2020. 2", + "[8] Christopher Clark, Jordi Salvador, Dustin Schwenk, Derrick Bonafilia, Mark Yatskar, Eric Kolve, Alvaro Herrasti, Jonghyun Choi, Sachin Mehta, Sam Skjonsberg, Carissa Schoenick, Aaron Sarnat, Hannaneh Hajishirzi, Aniruddha Kembhavi, Oren Etzioni, and Ali Farhadi. Iconary: A pictorial-based game for testing multimodal communication with drawings and text. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 1864-1886, Online and Punta Cana, Dominican Republic, 2021. Association for Computational Linguistics. 1, 2, 7", + "[9] Abhishek Das, Satwik Kottur, Khushi Gupta, Avi Singh, Deshraj Yadav, Jose M. F. Moura, Devi Parikh, and Dhruv Batra. Visual dialog. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017.", + "[10] A. Das, S. Kottur, J. F. Moura, S. Lee, and D. Batra. Learning cooperative visual dialog agents with deep reinforcement learning. In 2017 IEEE International Conference on Computer Vision (ICCV), pages 2970-2979, Los Alamitos, CA, USA, 2017. IEEE Computer Society. 2", + "[11] Abhishek Das, Samyak Datta, Georgia Gkioxari, Stefan Lee, Devi Parikh, and Dhruv Batra. Embodied question answering." + ], + "bbox": [ + 78, + 114, + 470, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 2", + "[12] Markus Eger, Chris Martens, and Marcela Alfaro Cordoba. An intentional ai for hanabi. In 2017 IEEE Conference on Computational Intelligence and Games (CIG), page 68-75. IEEE Press, 2017. 2", + "[13] Mathias Eitz, James Hays, and Marc Alexa. How do humans sketch objects? ACM Trans. Graph. (Proc. SIGGRAPH), 31 (4):44:1-44:10, 2012. 1, 3", + "[14] Alaaeldin El-Nouby, Shikhar Sharma, Hannes Schulz, Devon Hjelm, Layla El Asri, Samira Ebrahimi Kahou, Yoshua Bengio, and Graham W. Taylor. Tell, draw, and repeat: Generating and modifying images based on continual linguistic instruction. In The IEEE International Conference on Computer Vision (ICCV), 2019. 2", + "[15] Jiazhan Feng, Qingfeng Sun, Can Xu, Pu Zhao, Yaming Yang, Chongyang Tao, Dongyan Zhao, and Qingwei Lin. MMDialog: A large-scale multi-turn dialogue dataset towards multimodal open-domain conversation. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 7348-7363, Toronto, Canada, 2023. Association for Computational Linguistics. 2", + "[16] Chengying Gao, Qi Liu, Qi Xu, Limin Wang, Jianzhuang Liu, and Changqing Zou. Sketchycoco: Image generation from freehand scene sketches. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5174-5183, 2020. 3", + "[17] David Ha and Douglas Eck. A neural representation of sketch drawings. In ICLR 2018, 2018. 2018. 3", + "[18] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In International Conference on Learning Representations, 2022. 6", + "[19] Forrest Huang, Eldon Schoop, David Ha, and John Canny. Scones: towards conversational authoring of sketches. In Proceedings of the 25th International Conference on Intelligent User Interfaces, page 313–323, New York, NY, USA, 2020. Association for Computing Machinery. 2", + "[20] Kristiina Jokinen. Gaze and gesture activity in communication. pages 537-546. Springer, Berlin, Heidelberg, 2009. 1", + "[21] Kristiina Jokinen. Non-verbal signals for turn-taking and feedback. European Language Resources Association (ELRA), 2010. 1", + "[22] Kurmanbek Kaiyrbekov and Metin Sezgin. Stroke-based sketched symbol reconstruction and segmentation, 2019. 1", + "[23] Jin-Hwa Kim, Nikita Kitaev, Xinlei Chen, Marcus Rohrbach, Byoung-Tak Zhang, Yuandong Tian, Dhruv Batra, and Devi Parikh. CoDraw: Collaborative drawing as a testbed for grounded goal-driven communication. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 6495-6513, Florence, Italy, 2019. Association for Computational Linguistics. 1, 2", + "[24] Yihuai Lan, Zhiqiang Hu, Lei Wang, Yang Wang, Deheng Ye, Peilin Zhao, Ee-Peng Li m, Hui Xiong, and Hao Wang. Llm-based agent society investigation: Collaboration and confrontation in avalon gameplay, 2024. 2" + ], + "bbox": [ + 503, + 92, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "18184", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[25] Zixing Lei, Yiming Zhang, Yuxin Xiong, and Siheng Chen. Emergent communication in interactive sketch question answering. In Proceedings of the 37th International Conference on Neural Information Processing Systems, Red Hook, NY, USA, 2023. Curran Associates Inc. 2", + "[26] Lei Li, Hongbo Fu, and Chiew-Lan Tai. Fast sketch segmentation and labeling with deep learning, 2018. 1", + "[27] Tian Liang, Zhiwei He, Jen tse Huang, Wenxuan Wang, Wenxiang Jiao, Rui Wang, Yujiu Yang, Zhaopeng Tu, Shuming Shi, and Xing Wang. Leveraging word guessing games to assess the intelligence of large language models, 2023. 2", + "[28] Jian Lin. Divergence measures based on the shannon entropy. IEEE Transactions on Information Theory, 37(1):145-151, 1991. 7", + "[29] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning, 2023. 8", + "[30] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In Advances in Neural Information Processing Systems, pages 34892-34916. Curran Associates, Inc., 2023. 8", + "[31] Shuo Liu, Kaining Ying, Hao Zhang, Yue Yang, Yuqi Lin, Tianle Zhang, Chuanhao Li, Yu Qiao, Ping Luo, Wenqi Shao, and Kaipeng Zhang. Convbench: A multi-turn conversation evaluation benchmark with hierarchical capability for large vision-language models, 2024. 3", + "[32] Haoran Mo, Edgar Simo-Serra, Chengying Gao, Changqing Zou, and Ruomei Wang. General virtual sketching framework for vector line art. ACM Transactions on Graphics (TOG), 40:1-14, 2021. 6", + "[33] Chong Mou, Xintao Wang, Liangbin Xie, Jing Zhang, Zhonggang Qi, Ying Shan, and Xiaohu Qie. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. ArXiv, abs/2302.08453, 2023. 5, 6", + "[34] Xu Pan and Odelia Schwartz. Multimodal ai needs active human interaction. Nature Human Behaviour, 2024. 1", + "[35] Deepika Phutela. The importance of non-verbal communication, 2016. 1", + "[36] Yonggang Qi and Zheng-Hua Tan. Sketchsegnet+: An end-to-end learning of rnn for multi-class sketch semantic segmentation. IEEE Access, 7:102717-102726, 2019. 1", + "[37] Shuwen Qiu, Sirui Xie, Lifeng Fan, Tao Gao, Jungseock Joo, Song-Chun Zhu, and Yixin Zhu. Emergent graphical conventions in a visual communication game. In Advances in Neural Information Processing Systems, pages 13119-13131. Curran Associates, Inc., 2022. 2", + "[38] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In Proceedings of the 38th International Conference on Machine Learning, ICML 2021, 18-24 July 2021, Virtual Event, pages 8748-8763. PMLR, 2021. 4, 5, 6", + "[39] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. arXiv preprint arXiv:2112.10752, 2022. 4, 6" + ], + "bbox": [ + 78, + 90, + 470, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[40] Patsorn Sangkloy, Nathan Burnell, Cusuh Ham, and James Hays. The sketchy database: Learning to retrieve badly drawn bunnies. ACM Trans. Graph., 35(4), 2016. 1, 3", + "[41] Ravi Kiran Sarvadevabhatla, Shiv Surya, Trisha Mittal, and R. Venkatesh Babu. Pictionary-style word guessing on hand-drawn object sketches: Dataset, analysis and deep network models. IEEE Transactions on Pattern Analysis and Machine Intelligence, 42(1):221-231, 2020. 1, 2, 3, 7", + "[42] SeungWon Seo, Junhyeok Lee, SeongRae Noh, and HyeongYeop Kang. Llm-based cooperative agents using information relevance and plan validation, 2024. 2", + "[43] Chen Shen. Nonverbal communication in intercultural communication: A case study of telling lies. Frontiers in Educational Research, 6(13), 2023. 1", + "[44] Spotify. Annoy: Approximate nearest neighbors in c++/python, n.d. Computer software. 6", + "[45] Alessandro Suglia, Ioannis Konstas, Andrea Vanzo, Emanuele Bastianelli, Desmond Elliott, Stella Frank, and Oliver Lemon. CompGuessWhat?!: A multi-task evaluation framework for grounded language learning. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 7625-7641, Online, 2020. Association for Computational Linguistics. 2", + "[46] Alessandro Suglia, Antonio Vergari, Ioannis Konstas, Yonatan Bisk, Emanuele Bastianelli, Andrea Vanzo, and Oliver Lemon. Imagining grounded conceptual representations from perceptual information in situated guessing games. In Proceedings of the 28th International Conference on Computational Linguistics, pages 1090–1102, Barcelona, Spain (Online), 2020. International Committee on Computational Linguistics. 2", + "[47] F. Wang, S. Lin, H. Wu, H. Li, R. Wang, X. Luo, and X. He. Spfusionnet: Sketch segmentation using multi-modal data fusion. In 2019 IEEE International Conference on Multimedia and Expo (ICME), pages 1654–1659, 2019. 1", + "[48] Haohan Wang, Songwei Ge, Zachary Lipton, and Eric P Xing. Learning robust global representations by penalizing local predictive power. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2019. 3", + "[49] Wenhai Wang, Enze Xie, Xiaoge Song, Yuhang Zang, Wenjia Wang, Tong Lu, Gang Yu, and Chunhua Shen. Efficient and accurate arbitrary-shaped text detection with pixel aggregation network, 2020. 1", + "[50] Yuzhuang Xu, Shuo Wang, Peng Li, Fuwen Luo, Xiaolong Wang, Weidong Liu, and Yang Liu. Exploring large language models for communication games: An empirical study on werewolf. arXiv preprint arXiv:2309.04658, 2023. 2", + "[51] Lumin Yang, Jiajie Zhuang, Hongbo Fu, Kun Zhou, and Youyi Zheng. Sketchgen: Semantic sketch segmentation with graph convolutional networks, 2020. 1", + "[52] Yuan Yao, Tianyu Yu, Ao Zhang, Chongyi Wang, Junbo Cui, Hongji Zhu, Tianchi Cai, Haoyu Li, Weilin Zhao, Zhihui He, Qianyu Chen, Huarong Zhou, Zhensheng Zou, Haoye Zhang, Shengding Hu, Zhi Zheng, Jie Zhou, Jie Cai, Xu Han, Guoyang Zeng, Dahai Li, Zhiyuan Liu, and Maosong Sun. Minicpm-v: A gpt-4v level mllm on your phone, 2024. 8", + "[53] Qian Yu, Feng Liu, Yi-Zhe Song, Tao Xiang, Timothy M. Hospedales, and Chen-Change Loy. Sketch me that shoe. In" + ], + "bbox": [ + 501, + 92, + 893, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "18185", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 1, 3", + "[54] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 11941-11952, 2023. 4", + "[55] Hua Zhang, Si Liu, Changqing Zhang, Wenqi Ren, Rui Wang, and Xiaochun Cao. Sketchnet: Sketch classification with web images. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 1105-1113, 2016. 3", + "[56] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. arXiv preprint arXiv:2302.05543, 2023. 4, 5, 6", + "[57] Changqing Zou, Qian Yu, Ruofei Du, Haoran Mo, Yi-Zhe Song, Tao Xiang, Chengying Gao, Baoquan Chen, and Hao Zhang. Sketchyscene: Richly-annotated scene sketches. In ECCV, pages 438–454. Springer International Publishing, 2018. 3" + ], + "bbox": [ + 78, + 90, + 470, + 345 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "18186", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Sketchtopia_ A Dataset and Foundational Agents for Benchmarking Asynchronous Multimodal Communication with Iconic Feedback/fff37445-0c8d-4c24-a3bf-32cf9fcbe8ea_model.json b/2025/Sketchtopia_ A Dataset and Foundational Agents for Benchmarking Asynchronous Multimodal Communication with Iconic Feedback/fff37445-0c8d-4c24-a3bf-32cf9fcbe8ea_model.json new file mode 100644 index 0000000000000000000000000000000000000000..7970513ec334b3841b3841ea5e373e8c806db915 --- /dev/null +++ b/2025/Sketchtopia_ A Dataset and Foundational Agents for Benchmarking Asynchronous Multimodal Communication with Iconic Feedback/fff37445-0c8d-4c24-a3bf-32cf9fcbe8ea_model.json @@ -0,0 +1,2147 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.143, + 0.131, + 0.83, + 0.177 + ], + "angle": 0, + "content": "Sketchtopia: A Dataset and Foundational Agents for Benchmarking Asynchronous Multimodal Communication with Iconic Feedback" + }, + { + "type": "text", + "bbox": [ + 0.278, + 0.204, + 0.692, + 0.221 + ], + "angle": 0, + "content": "Mohd Hozafa Khan Ravi Kiran Sarvadevabhatla" + }, + { + "type": "text", + "bbox": [ + 0.194, + 0.222, + 0.776, + 0.24 + ], + "angle": 0, + "content": "Centre for Visual Information Technology (CVIT), IIIT Hyderabad, India" + }, + { + "type": "text", + "bbox": [ + 0.275, + 0.242, + 0.692, + 0.256 + ], + "angle": 0, + "content": "{mohd.hozaifa@research., ravi.kiran@}iit.ac.in" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.292, + 0.314, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.075, + 0.325, + 0.474, + 0.629 + ], + "angle": 0, + "content": "We introduce Sketchtopia, a large-scale dataset and AI framework designed to explore goal-driven, multimodal communication through asynchronous interactions in a Pictionary-inspired setup. Sketchtopia captures natural human interactions, including freehand sketches, open-ended guesses, and iconic feedback gestures, showcasing the complex dynamics of cooperative communication under constraints. It features over 20K gameplay sessions from 916 players, capturing 263K sketches, 10K erases, 56K guesses and 19.4K iconic feedbacks. We introduce multimodal foundational agents with capabilities for generative sketching, guess generation and asynchronous communication. Our dataset also includes 800 human-agent sessions for benchmarking the agents. We introduce novel metrics to characterize collaborative success, responsiveness to feedback and inter-agent asynchronous communication. Sketchtopia pushes the boundaries of multimodal AI, establishing a new benchmark for studying asynchronous, goal-oriented interactions between humans and AI agents. The dataset can be found at https://sketchtopia25.github.io/" + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.661, + 0.21, + 0.677 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.688, + 0.471, + 0.809 + ], + "angle": 0, + "content": "Human communication is inherently multimodal and goaldriven, often relying on non-verbal cues such as gestures and sketches to overcome language and cultural barriers [1, 34, 43]. Non-verbal interactions enable shared understanding despite constraints [20, 35]. Understanding such dynamics is key to developing multimodal AI agents capable of achieving shared goals in constrained environments while communicating asynchronously [6, 21]." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.811, + 0.472, + 0.902 + ], + "angle": 0, + "content": "With this motivation, we study communication-constrained asynchronous interactions in a casual game setting. We introduce Sketchtopia, a large-scale multi-modal dataset containing web-based gameplay episodes of Pictionary, the popular cooperative social game. In each episode, a Painter sketches clues while a Gueisser asynchronously" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.293, + 0.895, + 0.384 + ], + "angle": 0, + "content": "attempts to guess the target phrase (Fig. 1). This setup encourages creative, non-verbal communication under time pressure, leveraging iconic feedback gestures (e.g. \\(\\text{山}\\), and \\(⑦\\)). The rules of Pictionary forbid the Painter from writing text. Coupled with the time limit, this restriction often leads to creative player responses and generated content." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.385, + 0.897, + 0.552 + ], + "angle": 0, + "content": "Unlike prior works [8, 23], Sketchtopia captures authentic, hand-drawn sketches, feedback gestures, and natural editing actions (e.g., undo strokes). Our dataset not only facilitates the study of multimodal goal-driven communication but also serves as a testbed for novel AI agents. We present foundational agents, DRAWBOT and GUESSBOT, designed to asynchronously engage in multimodal interactions, leveraging both visual and textual modalities. This unique asynchronous communication mirrors real-life scenarios where agents must continuously think and act in parallel rather than turn-taking. We make the following contributions:" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.556, + 0.895, + 0.601 + ], + "angle": 0, + "content": "- Sketchtopia: A large-scale, multimodal dataset of Pictionary-like game episodes, capturing asynchronous interactions with sketches, guesses, and gestural feedback." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.602, + 0.895, + 0.647 + ], + "angle": 0, + "content": "- Two foundational AI agents, DRAWBOT and GUESSBOT, that demonstrate asynchronous, goal-driven multimodal communication with generative capabilities." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.647, + 0.895, + 0.691 + ], + "angle": 0, + "content": "- Introduction of novel metrics to assess agent performance in terms of game outcomes, individual contributions, and effectiveness in asynchronous multimodal communication." + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.556, + 0.895, + 0.691 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.708, + 0.651, + 0.725 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Sketch datasets: Existing sketch datasets primarily focus on recognition [13, 40], segmentation [22, 26, 36, 47, 49, 51], and retrieval [53], with labels assigned to the final sketch. In contrast, our dataset includes intermediate text \"guess\" annotations from gameplay, providing temporal grounding for sketches throughout the drawing process. Additionally, our dataset covers a wider array of categories, including abstract concepts (nouns, verbs, adjectives). Some works incorporate text guessing to simulate Pictionary [41], but lack the complexity and richness generated by interacting agents and gameplay dynamics." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "18176" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.09, + 0.891, + 0.356 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.366, + 0.893, + 0.396 + ], + "angle": 0, + "content": "Figure 1. Multimodal Interaction in Sketchtopia: Timeline of key game elements, showing asynchronous responses from Guesser and Painter. Green tick indicates a correct guess by the Guesser." + }, + { + "type": "table", + "bbox": [ + 0.077, + 0.417, + 0.472, + 0.505 + ], + "angle": 0, + "content": "
ApproachesShared GoalDrawerGueserVisual ContentDraw ContentGuess ContentFeedback (Multi-modal)Content EditingAsync.
Cont.Grd.
Sketchtopia (Ours)✓✓✓✓✓FreehandPhrase✓(C), C, D✓✓
Iconary [8]✓✓✓✗✓Icons (text)Sent.✗✗✗
Co-Draw [23]✓✓✓✗✓ClipArt (text)Sent.✗✗✗
Tell, Draw, Repeat [14]✗✓✗✗✓ImageSent.✗✗✗
Scores [19]✗✗✗✓✓FreehandSent.✗✗✗
Pixellor [7]✗✓✗✓✗Sketches✗✗✗✗
Visual QA✗✗✓✓✓✗Sent.✗✗✗
SketchGues[41]✗✗✓✓✓✗Phrase✗✗✗
Vis. Dialogue Agents [9]✗✗✓✓✓✗Sent.✗✗✗
Coop. Dialogue Agents [11]✓✗✓✓✓✗Sent.✗✗✗
Embodied VQA [11]✓✗✓✓✓✗Sent.✗✗✗
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.516, + 0.472, + 0.572 + ], + "angle": 0, + "content": "Table 1. Agents Comparison: A comparative summary of agent capabilities. Cont.: Continuous sketching; Grd.: Grounding of sketch content; Draw/Guess Content: Agent's output format; Async.: Asynchronous communication support; Sent.: Sentence." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.599, + 0.472, + 0.854 + ], + "angle": 0, + "content": "Cooperative Gameplay: Cooperative games with partial observability (CPO) are often used to study shared-goal communication strategies [5, 12]. Guessing games are a popular subcategory; for example, Ashktorab et al. [2, 3] investigate a game where a 'Giver' conveys a word using text. Similarly, Das et al. [10] focus on image retrieval in a text-based guessing game. Other multimodal versions include Charades, where speech-based guesses correspond to physical enactment [4], and Iconary, where a 'Drawer' uses visual icons to help a 'Guesser' interpret a phrase [8]. Suglia et al. [46] address visual guessing in zero-shot OOV cases for Guess-What and CompGuessWhat benchmarks [45]. Unlike the works mentioned above, Sketchtopia introduces significantly more complex gameplay. It involves contextual and generative free-form sketching, an open-ended guesser vocabulary, iconic feedback, and asynchronous, natural communication between players." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Communication Agents: Several research efforts in Pictionary-like settings introduce agents that focus solely on sketch generation. Scones [19] introduces agents for itera-" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.42, + 0.895, + 0.572 + ], + "angle": 0, + "content": "tive sketch creation using detailed textual instructions. Pixel [7] generates sketches efficiently with minimal strokes for a given word. Generally, these works do not involve a guesser in the loop and lack mechanisms for feedback-based sketch adaptation. As a complementary work, Sarvadevabhatla et al. [41] introduce a guesser agent trained to generate human-like guesses in response to iterative rendering of pre-drawn object sketches. However, their approach neglects interactivity and feedback, which are crucial elements of offline Pictionary." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.578, + 0.895, + 0.82 + ], + "angle": 0, + "content": "Recent works introduce Teller and Painter for dynamic, cooperative multimodal dialogue or gameplay. In Tell, Draw, Repeat [14], a drawer generates and modifies images iteratively based on detailed linguistic instructions from a teller. Unlike Pictionary, the teller always has access to the final image and neither multimodal interaction nor a long-term goal. mmDialogue [15] employs multi-modal dialogue in terms of both visual and text generation, but the interactions are not goal-oriented. Agents in Iconary [8] mimic drawings with discrete icons and detailed textual interaction from the guesser. In Co-draw [23], the agents engage in goal-driven collaborative communication, where the teller communicates with the drawer in natural language using discrete clip-art icons instead of sketches. The works of Qiu et al. [37] and Lei et al. [25] explore AI-driven emergent communication." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.826, + 0.897, + 0.903 + ], + "angle": 0, + "content": "LLM Agents: LLMs are increasingly used to test cooperative communication and strategic reasoning in games, focusing on text generation without visual environment processing [24, 27, 50]. REVECA [42] extends this by introducing LLM-based agents interacting in partially observable" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18177" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.075, + 0.089, + 0.472, + 0.212 + ], + "angle": 0, + "content": "
Works# SketchPr. Mod.Lang. Ground.FormEraseStrokesCat.
IntG.Final
Sketchtopia (Ours)263KTxt, Fdb✓✓Scene✓✓200
TU-Berlin [13]20KXX✓ObjX✓250
Quick Draw [17]50MXX✓ObjX✓345
Sketchy [40]75KXX✓ObjX✓125
SketchyScene [57]29KPhX✓ScXX-
ImageNet-Sketch [48]50KXX✓ObjXX1000
Scene Sketcher1225PhX✓Obj,ScX✓14
SketchyCOCO [16]14KPh, EMX✓Obj,ScX✓17
QUML-Shoe [53]419PhX✓ObjXX1
TU-Berlin Extended [55]20KPhX✓ObjXX250
SketchGuess-160 [41]8960Txt✓✓ObjX✓160
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.222, + 0.473, + 0.293 + ], + "angle": 0, + "content": "Table 2. Sketch Dataset Comparison::; Img.: Image; Pr. Mod.: Other modalities; Lang. Ground.: Language Grounding; IntG: Intermediate Guesses;Final: Final guess;Txt: Text; Fdb: Feedback;Sc: Scene; Obj: Object; Cat: Categories; Ph: Photo; EM: Edge Map" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.318, + 0.473, + 0.424 + ], + "angle": 0, + "content": "3D scenes. While Vision Language Models (VLMs) handle multi-turn multimodal dialogues [31], they lack multimodal generative capabilities crucial for Pictionary. Moreover, most LLMs and VLMs operate on turn-based rather than full-duplex communication. Our agents, by contrast, are simpler, communicate asynchronously and are more responsive." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.439, + 0.166, + 0.454 + ], + "angle": 0, + "content": "3. Dataset" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.464, + 0.208, + 0.48 + ], + "angle": 0, + "content": "3.1. Game Setup" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.487, + 0.472, + 0.563 + ], + "angle": 0, + "content": "Each game session consists of role-specific actions by the two players, viz. Drawer and Guesser. Each session presents a new target word for the Printer to illustrate and the Guesser to speculate. The session ends when the time limit (120 sec) is reached, or when the Printer deems a guess to be correct." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.572, + 0.234, + 0.587 + ], + "angle": 0, + "content": "3.2. Data Collection" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.596, + 0.472, + 0.763 + ], + "angle": 0, + "content": "The Sketchtopia Web App captures detailed, real-time interactions to emulate the Pictionary experience (see Supplementary video). Data collection involves two phases: Participant Onboarding, where players receive instructions and provide demographic details (age, dexterity), and Gameplay Recording, which logs time-stamped actions, including drawing strokes, iconic feedback, and guesses. Table 3 provides storage formats for collected data. Our dataset supports analysis of both gameplay dynamics and player behavior. Our ethical protocols ensure anonymized data collection, informed consent, and content moderation for privacy and quality." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.772, + 0.27, + 0.787 + ], + "angle": 0, + "content": "3.3. Dataset Information" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.795, + 0.472, + 0.903 + ], + "angle": 0, + "content": "Target Words: Drawers illustrate words from a pool of 200 terms, ranging from abstract nouns to verbs, covering various difficulty levels. Player Information: Demographics such as age, profession, and dexterity are collected to analyze gameplay strategies and behavior, ensuring fairness and diversity. We also collected 800 human-AI sessions, where players unknowingly interacted with AI agents, allowing" + }, + { + "type": "table", + "bbox": [ + 0.513, + 0.089, + 0.883, + 0.233 + ], + "angle": 0, + "content": "
RoleActionFormatTimestamp
DrawerStrokes1SVG2t1→t2
boolt
boolt
SVGt1→t2
(x,y)t1→t2
GuesserGuessTextt
?boolt
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.526, + 0.234, + 0.872, + 0.251 + ], + "angle": 0, + "content": "* SVG denotes white stroke over black background" + }, + { + "type": "table_caption", + "bbox": [ + 0.566, + 0.262, + 0.825, + 0.275 + ], + "angle": 0, + "content": "Table 3. Session Actions and Data Formats" + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.29, + 0.891, + 0.429 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.439, + 0.894, + 0.495 + ], + "angle": 0, + "content": "Figure 2. Game Distribution: Painter and Guesser actions distributed across time; sorted in decreasing order of success rate. For successful games, drawer actions preceded the guesser actions and usually last a shorter duration" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.521, + 0.895, + 0.642 + ], + "angle": 0, + "content": "study of human-AI interaction dynamics. Dataset Quality / Ethical Considerations: We used in-game and post-game text filters to eliminate irrelevant content (e.g. chat-style content such as 'Hi/Hello/lol'). We used text-based filters to remove abusive or inappropriate content in guesses, ensuring a safe and respectful environment. We also conducted manual inspections to identify and address any game rule violations." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.651, + 0.665, + 0.667 + ], + "angle": 0, + "content": "3.4. Dataset Analysis" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.674, + 0.894, + 0.75 + ], + "angle": 0, + "content": "We analyze Sketchtopia across four key dimensions: global game statistics, session-level performance, player-level insights, and temporal dynamics, revealing gameplay factors that influence success, player behavior, and communication strategies." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.894, + 0.84 + ], + "angle": 0, + "content": "Global game statistics: Our dataset includes 916 players (ages 15–60, avg. 26.6), with 20K sessions totaling 263K strokes, 10K erases, 16.3K highlights, 56K guesses, and 19.4K feedback events. With an \\(89\\%\\) success rate, sessions averaged 13 strokes, 2.83 guesses, 0.82 highlights and 1 feedback event." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.894, + 0.901 + ], + "angle": 0, + "content": "Session-level: Success rates varied by target word types, with nouns and verbs showing higher success compared to abstract nouns and adjectives due to their complexity. As illustrated in Fig. 2, sessions for high-success words often" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18178" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.109, + 0.089, + 0.44, + 0.151 + ], + "angle": 0, + "content": "
OutcomeDFGFDF+GFNo-FB
successful52.330.881.5942.51
failure1.680.180.091.59
" + }, + { + "type": "table_caption", + "bbox": [ + 0.076, + 0.161, + 0.472, + 0.203 + ], + "angle": 0, + "content": "Table 4. Games categorized across feedback type and outcome. Numbers represent the percentage of games in each category (outcome or presence of iconic feedback)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.228, + 0.47, + 0.44 + ], + "angle": 0, + "content": "display clear peaks where drawer actions precede guesses, indicating early sketch clarity. Conversely, harder words show flatter distributions, suggesting shifts in strategies or communication challenges. Peaks in the distribution indicate moments where new strategies or ideas emerged, correlating with challenging words. An analysis of success rates across target words and parts of speech (POS) revealed notable variability – Nouns (92.6%), Verbs (84.2%), and Adjectives (79.1%). Among nouns, Abstract Nouns (85%) are the least successful, while food/sports categories are the most. Abstract nouns often require more complex and detailed sketches. Additionally, adjectives (e.g., hard, hollow) which describe nouns are particularly challenging to convey since many objects can possess such properties." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.44, + 0.47, + 0.576 + ], + "angle": 0, + "content": "Player-level insights: At the player level, we identified distinct skills for Drawers and Guessers. Players were slightly more successful as Guessers (14%) than Drawers (9%), yet 77% performed equally well in both roles, emphasizing the shared nature of cooperative gameplay. Categorizing games by feedback and outcome (see Table 4) shows that feedback plays a role in game results. Games that only had drawer feedback had higher success rates, highlighting its importance in guiding Guessers effectively." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.576, + 0.47, + 0.727 + ], + "angle": 0, + "content": "Game Dynamics: To understand gameplay evolution, we analyzed action trajectories using CLIP embeddings [38] to measure similarity between sketches, guesses, and target words. As shown in Figure 3, Drawers refine their sketches over time, while Guessers reactively adjust guesses. Successful games show a convergence pattern where sketches and guesses align closely with the target, demonstrating effective cooperative communication. This game flow analysis captures how players' strategies adapt dynamically throughout the session." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.74, + 0.16, + 0.758 + ], + "angle": 0, + "content": "4. Agents" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Our goal is to design player agents that emulate human-like, responsive gameplay, free from rigid turn-based limitations and capable of multimodal interactions. Beyond deciding 'what' action to perform (e.g., a Roller's stroke), agents must also determine 'when' to act (e.g., when a Gueisser should output a guess). The design should have the flexibility to perform a \"no-op\" (i.e., no immediate action in response to another player's input, allowing the Gueisser to wait for more strokes). Additionally, agents must support not only" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.092, + 0.892, + 0.122 + ], + "angle": 0, + "content": "their primary actions (drawing or guessing) but also iconic feedback \\((\\text{日}, \\text{日}, \\text{日})\\)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.122, + 0.895, + 0.258 + ], + "angle": 0, + "content": "To meet these requirements, we organize the agent design into a two-level module hierarchy (see Fig. 4). At the top level, the ACTIONDECIDER module manages global action timing ('when') and selection ('what'). The lower level comprises specialized Painter and Gueisser modules (DRAWBOT and GUESSBOT), which handle role-specific tasks. The following sections detail the architecture, decision-making processes, and interaction mechanisms of each module within this hierarchy." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.269, + 0.712, + 0.284 + ], + "angle": 0, + "content": "4.1. ActionDecider Module" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.292, + 0.895, + 0.578 + ], + "angle": 0, + "content": "The ACTIONDECIDER module serves as a lightweight controller, continuously monitoring game states to decide when agents should asynchronously initiate actions. We describe its setup for the Drawing (see Fig. 5). In the first stage, the module processes multimodal game data streams: guesses are encoded using a Text Encoder [54], canvas strokes via a Vision Encoder [54], and iconic feedback through a Feed Forward Network (FFN). These embeddings are cached, with only the latest content appended. A \\(1 \\times 1\\) convolution is applied to the guess, canvas, and feedback embeddings to obtain aggregated vector representations. For the Painter, a vector for the Target Word is also included. These are concatenated into a Game State Vector. In the second stage, the Game State Vector is passed through a 'Response' FFN, which determines the Painter's action: Wake up the Painter, Generate Feedback, or Do Nothing (wait). If Generate Feedback is selected, a 'Feedback' FFN produces specific feedback (Q), for the Guesser." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.58, + 0.892, + 0.64 + ], + "angle": 0, + "content": "The Guesser agent follows a similar design but excludes the Target Word. Here, the wakeup action type is routed to GUESSBOT, and feedback actions correspond to the Guesser's iconic feedback (②)." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.651, + 0.738, + 0.667 + ], + "angle": 0, + "content": "4.2. Drawer Agent (DRAWBOT)" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.674, + 0.895, + 0.885 + ], + "angle": 0, + "content": "The primary task of DRAWBOT is to visually convey the target word through sequential actions: (i) incrementally draw freehand strokes to depict the target word, (ii) update the canvas (e.g., erase using the eraser-icon) based on guesses and feedback (②) from the Guesser, and (iii) generate gestural feedback (①), (iv) based on the semantic proximity of guesses to target word. DRAWBOT is built on a fine-tuned Stable Diffusion model [39], tailored to the game's sketch style. Its architecture extends a Text-to-Image (T2I) Diffusion Model [39] with ControlNet-inspired adapters [56], allowing it to condition sketch generation on multimodal inputs such as canvas state, guess history, and game time (see Fig. 6). Each adapter module provides conditioning based on different input modalities (see Fig. 6):" + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.886, + 0.892, + 0.901 + ], + "angle": 0, + "content": "- Canvas State Adapter: A convolutional layer encodes the" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18179" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.09, + 0.488, + 0.477 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.495, + 0.091, + 0.888, + 0.476 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.488, + 0.893, + 0.545 + ], + "angle": 0, + "content": "Figure 3. Game Flow Analysis: (Sec. 3.4) Each line plot represents a game session. The points on each plot represent CLIP embeddings [38] of canvas or a guess word. The x-axis represents time, y-axis represents normalized distance of these embeddings to the target word (top-left of each panel - blue (easy) and red (hard)). The final canvas for each session is shown. Line plots for paired drawer and guesser have the same color. \\(\\checkmark\\) denotes a successful guess and associated canvas, \\(\\times\\) represents an unsuccessful session." + }, + { + "type": "image", + "bbox": [ + 0.088, + 0.567, + 0.46, + 0.652 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.661, + 0.472, + 0.731 + ], + "angle": 0, + "content": "Figure 4. Pictionary Agent Framework: The top-level (ActionDecider) decides action type and timing and delegates to \\(2^{nd}\\) level in the hierarchy which perform the individual tasks. Our hierarchical design enables delegation and asynchronous communication. Refer to Sec. 4 for details." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.765, + 0.47, + 0.826 + ], + "angle": 0, + "content": "current visual state of the canvas (at \\( t = t_1 \\)). The result is processed by ControlNet's edge detection layer [56]. This adapter modulates sketch generation by incorporating recent updates, ensuring continuity in the drawing process." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.826, + 0.47, + 0.871 + ], + "angle": 0, + "content": "- Guess Adapter: Encodes the history of guesses, allowing the agent to adjust its drawings according to the Guesser's proximity to the target." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.871, + 0.471, + 0.902 + ], + "angle": 0, + "content": "- Time Adapter: A time encoder that regulates the game's pacing and drawing progression, creating a sense of ur" + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.826, + 0.471, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.568, + 0.891, + 0.739 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.748, + 0.894, + 0.777 + ], + "angle": 0, + "content": "Figure 5. Schematic Diagram of the ActionDecider Module for decision timing and action selection. Refer Sec. 4.1." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.809, + 0.774, + 0.824 + ], + "angle": 0, + "content": "gency in line with the game's progress." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.894, + 0.902 + ], + "angle": 0, + "content": "We modify the default T2I Adapter [33] to accommodate non-visual modalities as input. Overall, these adapters enable DRAWBOT to generate incremental strokes rather than complete images, producing residual updates to the canvas (see Fig. 6, \\( t = t_{2} \\)), thereby maintaining continuity while" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18180" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.082, + 0.09, + 0.468, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.304, + 0.473, + 0.442 + ], + "angle": 0, + "content": "Figure 6. **Drawer Agent (DrawBot) Architecture:** The top left blue box shows the current state of the canvas. The pink module illustrates the primary Text-to-image-based generation module, which processes the target word (e.g., 'toothbrush') to produce strokes. The green module represents additional conditioning inputs — such as prior guesses and current time — that guide the generative process, enabling the agent to create iterative strokes towards a recognizable sketch. The bottom right orange box shows that conditioning outputs are passed to the encoder of the T2I generator. SD stands for Stable Diffusion (A T2I Model)" + }, + { + "type": "image", + "bbox": [ + 0.14, + 0.46, + 0.412, + 0.587 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.596, + 0.472, + 0.625 + ], + "angle": 0, + "content": "Figure 7. Schematic Diagram of the Gueisser Agent's Retrieval-Based Guessing Framework (Sec. 4.3)." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.653, + 0.418, + 0.669 + ], + "angle": 0, + "content": "dynamically responding to the evolving game state." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.68, + 0.32, + 0.696 + ], + "angle": 0, + "content": "4.3. Gueisser Agent (GUESSBOT)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.704, + 0.471, + 0.869 + ], + "angle": 0, + "content": "The Gueisser Agent (GUESSBOT) generates guesses based on the canvas content and multimodal cues, aiming to identify the hidden target word. Given the limited dataset size, GUESSBOT employs a retrieval-based approach. When GUESSBOT receives a 'Wake-up' signal from the ACTION-DECIDER, it encodes the current canvas state \\( C_t \\) (see Fig. 7) and uses Approximate Nearest Neighbor (ANN) search to compare it with historical game sessions. This strategy enables the agent to utilize past gameplay data, applying rule-based criteria to generate guesses aligned with the current game context." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.472, + 0.902 + ], + "angle": 0, + "content": "The retrieval-based process involves two steps: (i) encoding the canvas content using a vision foundation model" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.894, + 0.213 + ], + "angle": 0, + "content": "(CLIP) to retrieve 'k' similar sessions from the database, and (ii) applying rule-based filtering based on similarity scores, guess timings and user feedback to refine the results. This filtering discards irrelevant sessions, for example, by rejecting matches where the timing deviates significantly from the current session. The final output is derived from the best-matched session's guesses, producing the selected guess \\((g,t)\\) as shown in Fig. 7." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.213, + 0.894, + 0.304 + ], + "angle": 0, + "content": "A key design goal of GUESSBOT is to emulate human gameplay, including natural errors in guessing. For example, human players often overlook spelling errors (e.g., \"spyder\", see Fig. 3) during immersive gameplay. Our retrieval-based approach inherently accommodates these human-like quirks, enhancing the realism of the agent's responses." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.316, + 0.719, + 0.332 + ], + "angle": 0, + "content": "4.4. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.34, + 0.892, + 0.37 + ], + "angle": 0, + "content": "Canvas strokes (SVG) are thickened by a factor of 3 and rendered as \\(224 \\times 224\\) images for processing by agents." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.371, + 0.893, + 0.506 + ], + "angle": 0, + "content": "Action Decider: By default, agents remain inactive unless triggered by a 'Wake-up' signal from their Action Decider (see Fig. 4, Fig. 5). Canvas states are captured every 3 seconds, forming the Drawing History. For text encoding (target words, guess history), we use a pre-trained CLIP text encoder [38] (TE), while the vision encoder handles the canvas history (VE). Each FFN is a 3-layer MLP. Every 3 seconds, the Action Decider processes its accumulated inputs and outputs actions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.508, + 0.895, + 0.764 + ], + "angle": 0, + "content": "DrawBot: The canvas is processed using ControlNet [56] with StableDiffusion-1.5 [39] as the T2I generator (Fig. 6). The T2I Adapter is modified to handle non-visual inputs: each guess is CLIP-encoded and combined with its times-tamp to form a \\(512 \\times 2 \\times n_g\\) representation, which is integrated into the T2I Generator's U-Net [33]. The generator is fine-tuned on Sketchtopia sketches using LoRA [18] to output incremental strokes (residual images). We optimize the adapters in stages: first training ControlNet, then the T2I Adapters while freezing ControlNet, and finally fin-tuning all components together. The agent is trained by breaking game sessions into fixed time intervals and conditioning outputs on canvas state, guesses, and timestamps, allowing for adaptive sketching based on all input modalities. The residual image output by T2I Generator is converted to its SVG representation using virtual sketching [32], before being added to the canvas." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.765, + 0.895, + 0.902 + ], + "angle": 0, + "content": "GuessBot: A vector database of canvas embeddings (using CLIP ViT-B/32) is built with metadata (guess and timestamp) for fast retrieval using Annoy [44], an approximate nearest-neighbor retriever. Upon receiving a 'Wake-Up' signal, GUESSBOT encodes the current canvas \\( e_{C_t} \\) (see Fig. 7) at current time \\( t \\), and retrieves the top 10 matches within a time window \\( [t - \\delta, t + \\delta] \\), where \\( \\delta = 6 \\). Among these, we select the match closest to the query by cosine similarity and output the corresponding guess word." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "18181" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.184 + ], + "angle": 0, + "content": "Agent Orchestration: All agents are deployed as microservices communicating with the Pictionary WebApp via API calls. Participants are paired randomly, with a \\(15\\%\\) chance of getting paired with an agent. During gameplay, the WebApp sends relevant data to the ActionDecider (Sec. 4.1), which delegates tasks to draw or feedback as needed." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.195, + 0.389, + 0.213 + ], + "angle": 0, + "content": "5. Metrics for Benchmarking Agents" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.22, + 0.472, + 0.28 + ], + "angle": 0, + "content": "We evaluate agents (human-only games and mixed human-agent) in terms of their collaborative success, their responsiveness to feedback, and the similarity of their timing patterns to human counterparts." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.281, + 0.472, + 0.416 + ], + "angle": 0, + "content": "Collaborative Game Success: In goal-oriented games, the primary measure is whether Painter and Guesser agents can effectively collaborate to identify the target word. Metrics include Win Rate (WR): The proportion of games where the target word is guessed correctly, serving as a baseline measure of collaborative success [8]; Soft Win Rate (SWR): Extends Win Rate by considering synonyms of target word, capturing near-success scenarios where agents were close to the correct answer [8, 41];" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.417, + 0.472, + 0.567 + ], + "angle": 0, + "content": "Feedback Responsiveness Score (FRS): measures the extent to which the Guesser adjusts its guesses in response to Painter feedback. After receiving or feedback, if the guesser's next guess is semantically closer to the target word than the guess immediately preceding the feedback, such an occurrence is termed a responsive guess. In a game session, let the count of such occurrences be \\( s \\), and the Painter takes \\( p \\) and \\( n \\) actions. FRS for that session is defined as \\( \\frac{s}{p + n} \\). The resulting numbers are averaged across sessions to obtain the final metric." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.568, + 0.472, + 0.717 + ], + "angle": 0, + "content": "Asynchronous Action Overlap (AAO): This is a session-level measure and characterizes the degree to which agents act asynchronously. Suppose a Gueisser action's timestamp overlaps with the duration of a Drawer action (e.g. sketching, erasing) within a session. Let the count of all occurrences be \\( o \\), and let the total number of Gueisser actions be \\( g \\). AAO for the session is defined as \\( AAO = \\frac{o}{g} \\). Note that a sequential turn-based interaction between agents would have an AAO value of 0, while a purely random and uncoordinated interaction would have a value of 100." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.718, + 0.473, + 0.903 + ], + "angle": 0, + "content": "Multimodal Action Timing Similarity (MATS): For a given target word, suppose the sequence of human Guesser's guess timestamps across all sessions be \\( T_{g}^{H} \\), and a similar sequence for GUESSBOT be \\( T_{g}^{A} \\). We define a guess distance \\( d_{guess} \\) between these two sequences as the Jensen-Shannon divergence [28] between their Kernel Density estimates, i.e. \\( d_{guess} = \\text{JENSEN-SHANNON}\\{\\text{KDE}(T_{g}^{H}), \\text{KDE}(T_{g}^{A})\\} \\). The average \\( d_{guess} \\) across all target words forms the MATS metric for GUESSBOT. A similar exercise for DRAWBOT's actions, where the timestamp of the stroke duration's midpoint is used, provides the MATS score for DRAWBOT. Detailed definitions for the metrics are provided in the Supplementary." + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.091, + 0.891, + 0.34 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.349, + 0.895, + 0.419 + ], + "angle": 0, + "content": "Figure 8. Qualitative comparison between human and agent game outcomes: The figure shows the final canvas and the final guess for select target words (fishing, length, loud). Human outputs are in purple and agent outputs are in blue. Successful games are with green background and unsuccessful ones in red." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.448, + 0.587, + 0.464 + ], + "angle": 0, + "content": "6. Results" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.474, + 0.895, + 0.595 + ], + "angle": 0, + "content": "To benchmark our agents, we carefully choose a representative evaluation set of game sessions grouped by Part of Speech and compare the performance of humans and agents on this set. In general, the metric values for humans act as ground truth or desirable range, which the agent should achieve. We obtain 800 sessions to be used for testing. The numbers mentioned in Tab. 5 for HH (Human-Human) and HA (Human-Agent) are both computed on this set." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.597, + 0.895, + 0.809 + ], + "angle": 0, + "content": "Collaborative Game Success: Across games, agents had a \\(68\\%\\) success rate. The win rate for GUESSBOT is slightly lesser than the human counterpart \\((85\\%)\\). However, DRAWBOT's win rate is \\(51\\%\\), which is markedly smaller. The Painter agent's bad performance is mainly due to adjectives \\((30\\%)\\) and verbs \\((56\\%)\\). It performs reasonably well on most nouns \\((74\\%)\\). The primary reason for bad performance on adjectives is that there is no direct object (e.g. car, bus, icecream) associated with adjectives. Most vision foundation models are trained on objects rather than metaphorical concepts which encode attributes (e.g. hard/soft, length, hollow and loud) which may be associated with multiple objects. Such models need detailed prompts to invoke the relationship and properties of objects." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.811, + 0.897, + 0.902 + ], + "angle": 0, + "content": "Qualitative analysis: Many interesting patterns can be observed in the performance and response patterns of human-human and human-agent games – see Fig. 8. Compared to human counterparts, DRAWBOT's sketches have greater symmetry, with clean stroke lines. Unlike typical generative model outputs, which tend to have a 'synthetic' appearance," + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18182" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.182 + ], + "angle": 0, + "content": "training on Sketchtopia's data induces a human-like jitter in the generated strokes of DRAWBOT. GUESSBOT's final responses seem human-like for the corresponding canvas contents, even when the game outcome is a failure. This is consistent with the agent's retrieval-based design. Additional examples can be found in Supplementary." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.182, + 0.471, + 0.272 + ], + "angle": 0, + "content": "Feedback Responsiveness: The FRS scores for Human and GUESSBOT are similar (Tab. 5), implying that the agent's ability to adapt its response to feedback is similar to human counterparts. Here, the Gueisser Agent's ActionDecider also plays a key role in 'waking up' GUESSBOT at the right time in response to Painter feedback." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.273, + 0.471, + 0.484 + ], + "angle": 0, + "content": "Characterizing Interaction Dynamics: Similar to FRS, the AAO scores of agents are comparable to humans, indicating comparable levels of asynchronous communication. However, GUESSBOT has a lower AAO score. This is in line with a Guesser's tendency to wait passively for enough strokes to accumulate on the canvas before guessing. The MATS score measures the agent's action pacing compared to humans. Therefore, it is a proxy measure for ActionDecider's performance due to the latter's role as a key asynchronous enabler of actions in each agent. A relatively lower MATS score for GUESSBOT suggests that its action pacing is closer to human. Empirically, we observed that feedback distribution for DRAWBOT was markedly different from human, contributing to its relatively higher MATS score." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.485, + 0.471, + 0.65 + ], + "angle": 0, + "content": "VLMs as Guessers: Vision-Language Models (VLMs) [29, 30, 52] have become a popular choice for image-conditioned reasoning. In our benchmarking context, VLMs could play a Guesser's role. However, despite detailed prompts instructing the model to output a guess phrase and despite providing prior guess history, we observed that VLM models tend to produce verbose and unhelpful outputs. Examples can be found in Supplementary. Moreover, the slower response times and synchronized (turn-taking) paradigm of current VLMs make them unviable compared to our asynchronous and contextually responsive agents." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.651, + 0.471, + 0.816 + ], + "angle": 0, + "content": "Limitations: Despite efforts to minimize biases, the dataset may still reflect factors such as age, handedness, and cultural background, with a predominance of younger, right-handed, English-speaking participants from a limited region. This may affect its generalizability to other languages, cultures, and demographics. While our game interface supports asynchronous interactions, further testing across diverse groups is needed. Additionally, while Sketchtopia includes iconic feedback as a novelty, it is limited to predefined icons (e.g., , , and Ⓞ). This restricts the analysis of more nuanced and complex feedback that could occur in natural interactions." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.831, + 0.195, + 0.845 + ], + "angle": 0, + "content": "7. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.856, + 0.47, + 0.901 + ], + "angle": 0, + "content": "Sketchtopia enables a scalable framework that pushes the boundaries of multimodal goal-driven communication in asynchronous settings. As the first-ever dataset to include" + }, + { + "type": "table", + "bbox": [ + 0.502, + 0.089, + 0.891, + 0.153 + ], + "angle": 0, + "content": "
ModelsGame SuccessFeedback ResponsivenessInteraction Dynamics
Win ↑Soft Win ↑FRS ↑AAOMATS ↓
DRAWBOT51.652.4-0.210.39
GUESSBOT81.786.046%0.170.27
Human85.593.440%0.290.00
" + }, + { + "type": "table_caption", + "bbox": [ + 0.499, + 0.164, + 0.894, + 0.22 + ], + "angle": 0, + "content": "Table 5. Benchmarking humans and agents via custom-designed metrics (Sec. 5). FRS: Feedback Responsiveness Score; AAO: Asynchronous Action Overlap; MATS: Multimodal Action Timing Similarity." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.267, + 0.895, + 0.403 + ], + "angle": 0, + "content": "rich feedback signals, our collection captures realistic gameplay data with natural human interactions, far beyond what existing datasets offer. Sketchtopia is not just about isolated sketches – it incorporates iconic feedback and dynamic interactions between players, all captured using a robust game interface that seamlessly supports both human-human and human-agent gameplay. This versatility allows us to explore interactions at a depth previously unattainable, establishing a new benchmark for multimodal AI systems." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.413, + 0.895, + 0.565 + ], + "angle": 0, + "content": "A standout contribution of our work is the design of first asynchronous communication agents capable of real-time decision-making in multi-modal formats. Our novel agent architectures emphasize responsiveness and speed, enabling agents to mimic the fluidity of natural human communication without the constraints of turn-taking. By introducing innovative metrics such as Asynchronous Action Overlap and Feedback Responsiveness Scores, we redefine how to evaluate collaborative success and interaction dynamics in asynchronous multi-modal communication settings." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.574, + 0.897, + 0.726 + ], + "angle": 0, + "content": "Benchmarking our agents against human players demonstrates that they can generate human-like actions and responses. Our results reveal that even state-of-the-art VLMs lag significantly in handling asynchronous, multimodal communication complexities. This establishes Sketchtopia as a critical benchmark for advancing AI research in this space. Additionally, our game flow visualizations offer a novel lens into understanding how agents and humans progress towards shared goals, providing insights into the evolution of strategies and collaborative problem-solving." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.735, + 0.897, + 0.901 + ], + "angle": 0, + "content": "Sketchtopia's applications extend beyond Pictionary. Its interactive features enable sketch scene editing and collaborative whiteboard tools, with potential for live note-taking and narration. The dataset's focus on abstract sketches also supports cognitive science research into communication strategies. In summary, Sketchtopia provides a valuable resource for exploring asynchronous, multimodal AI systems and realistic agent-based gameplay. By tackling the complexities of real-time, feedback-driven communication, we aim to inspire AI advancements that are more natural, adaptive, and truly collaborative." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "18183" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.105 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.471, + 0.142 + ], + "angle": 0, + "content": "[1] Michael Argyle. The syntaxes of bodily communication. Linguistics, 11(112):71-92, 1973. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.146, + 0.472, + 0.215 + ], + "angle": 0, + "content": "[2] Zahra Ashktorab, Q. Vera Liao, Casey Dugan, James Johnson, Qian Pan, Wei Zhang, Sadhana Kumaravel, and Murray Campbell. Human-ai collaboration in a cooperative game setting: Measuring social perception and outcomes. Proc. ACM Hum.-Comput. Interact., 4(CSCW2), 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.217, + 0.472, + 0.313 + ], + "angle": 0, + "content": "[3] Zahra Ashktorab, Casey Dugan, James Johnson, Qian Pan, Wei Zhang, Sadhana Kumaravel, and Murray Campbell. Effects of communication directionality and ai agent differences in human-ai interaction. In Proceedings of the 2021 CHI Conference on Human Factors in Computing Systems, New York, NY, USA, 2021. Association for Computing Machinery. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.316, + 0.471, + 0.357 + ], + "angle": 0, + "content": "[4] Muhammad Fauzi Bafadal and Humaira Humaira. The use of charades games in teaching vocabulary to the junior high school students. Linguistics and Elt Journal, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.36, + 0.471, + 0.444 + ], + "angle": 0, + "content": "[5] Nolan Bard, Jakob N. Foerster, Sarath Chandar, Neil Burch, Marc Lanctot, H. Francis Song, Emilio Parisotto, Vincent Dumoulin, Subhodeep Moitra, Edward Hughes, Iain Dunning, Shibl Mourad, Hugo Larochelle, Marc G. Bellemare, and Michael Bowling. The hanabi challenge: A new frontier for ai research. Artificial Intelligence, 280:103216, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.446, + 0.47, + 0.5 + ], + "angle": 0, + "content": "[6] Fabien Beltrame. Comportement non verbal et situation d'interaction: Quand la gestuelle des mains revendique un role cognitif de support des représentations pour la résolution de problème, 2006. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.504, + 0.471, + 0.572 + ], + "angle": 0, + "content": "[7] Ayan Kumar Bhunia, Ayan Das, Umar Riaz Muhammad, Yongxin Yang, Timothy M. Hospedales, Tao Xiang, Yulia Gryaditskaya, and Yi-Zhe Song. Pixel: a competitive sketching ai agent. so you think you can sketch? ACM Trans. Graph., 39(6), 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.575, + 0.471, + 0.727 + ], + "angle": 0, + "content": "[8] Christopher Clark, Jordi Salvador, Dustin Schwenk, Derrick Bonafilia, Mark Yatskar, Eric Kolve, Alvaro Herrasti, Jonghyun Choi, Sachin Mehta, Sam Skjonsberg, Carissa Schoenick, Aaron Sarnat, Hannaneh Hajishirzi, Aniruddha Kembhavi, Oren Etzioni, and Ali Farhadi. Iconary: A pictorial-based game for testing multimodal communication with drawings and text. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 1864-1886, Online and Punta Cana, Dominican Republic, 2021. Association for Computational Linguistics. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.73, + 0.471, + 0.798 + ], + "angle": 0, + "content": "[9] Abhishek Das, Satwik Kottur, Khushi Gupta, Avi Singh, Deshraj Yadav, Jose M. F. Moura, Devi Parikh, and Dhruv Batra. Visual dialog. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.802, + 0.471, + 0.87 + ], + "angle": 0, + "content": "[10] A. Das, S. Kottur, J. F. Moura, S. Lee, and D. Batra. Learning cooperative visual dialog agents with deep reinforcement learning. In 2017 IEEE International Conference on Computer Vision (ICCV), pages 2970-2979, Los Alamitos, CA, USA, 2017. IEEE Computer Society. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.471, + 0.901 + ], + "angle": 0, + "content": "[11] Abhishek Das, Samyak Datta, Georgia Gkioxari, Stefan Lee, Devi Parikh, and Dhruv Batra. Embodied question answering." + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.121, + 0.894, + 0.176 + ], + "angle": 0, + "content": "[12] Markus Eger, Chris Martens, and Marcela Alfaro Cordoba. An intentional ai for hanabi. In 2017 IEEE Conference on Computational Intelligence and Games (CIG), page 68-75. IEEE Press, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.178, + 0.892, + 0.217 + ], + "angle": 0, + "content": "[13] Mathias Eitz, James Hays, and Marc Alexa. How do humans sketch objects? ACM Trans. Graph. (Proc. SIGGRAPH), 31 (4):44:1-44:10, 2012. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.219, + 0.894, + 0.3 + ], + "angle": 0, + "content": "[14] Alaaeldin El-Nouby, Shikhar Sharma, Hannes Schulz, Devon Hjelm, Layla El Asri, Samira Ebrahimi Kahou, Yoshua Bengio, and Graham W. Taylor. Tell, draw, and repeat: Generating and modifying images based on continual linguistic instruction. In The IEEE International Conference on Computer Vision (ICCV), 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.302, + 0.894, + 0.399 + ], + "angle": 0, + "content": "[15] Jiazhan Feng, Qingfeng Sun, Can Xu, Pu Zhao, Yaming Yang, Chongyang Tao, Dongyan Zhao, and Qingwei Lin. MMDialog: A large-scale multi-turn dialogue dataset towards multimodal open-domain conversation. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 7348-7363, Toronto, Canada, 2023. Association for Computational Linguistics. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.4, + 0.892, + 0.467 + ], + "angle": 0, + "content": "[16] Chengying Gao, Qi Liu, Qi Xu, Limin Wang, Jianzhuang Liu, and Changqing Zou. Sketchycoco: Image generation from freehand scene sketches. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5174-5183, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.469, + 0.892, + 0.496 + ], + "angle": 0, + "content": "[17] David Ha and Douglas Eck. A neural representation of sketch drawings. In ICLR 2018, 2018. 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.498, + 0.894, + 0.564 + ], + "angle": 0, + "content": "[18] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In International Conference on Learning Representations, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.567, + 0.894, + 0.636 + ], + "angle": 0, + "content": "[19] Forrest Huang, Eldon Schoop, David Ha, and John Canny. Scones: towards conversational authoring of sketches. In Proceedings of the 25th International Conference on Intelligent User Interfaces, page 313–323, New York, NY, USA, 2020. Association for Computing Machinery. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.637, + 0.894, + 0.676 + ], + "angle": 0, + "content": "[20] Kristiina Jokinen. Gaze and gesture activity in communication. pages 537-546. Springer, Berlin, Heidelberg, 2009. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.679, + 0.894, + 0.718 + ], + "angle": 0, + "content": "[21] Kristiina Jokinen. Non-verbal signals for turn-taking and feedback. European Language Resources Association (ELRA), 2010. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.72, + 0.892, + 0.748 + ], + "angle": 0, + "content": "[22] Kurmanbek Kaiyrbekov and Metin Sezgin. Stroke-based sketched symbol reconstruction and segmentation, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.749, + 0.894, + 0.845 + ], + "angle": 0, + "content": "[23] Jin-Hwa Kim, Nikita Kitaev, Xinlei Chen, Marcus Rohrbach, Byoung-Tak Zhang, Yuandong Tian, Dhruv Batra, and Devi Parikh. CoDraw: Collaborative drawing as a testbed for grounded goal-driven communication. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 6495-6513, Florence, Italy, 2019. Association for Computational Linguistics. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.846, + 0.894, + 0.901 + ], + "angle": 0, + "content": "[24] Yihuai Lan, Zhiqiang Hu, Lei Wang, Yang Wang, Deheng Ye, Peilin Zhao, Ee-Peng Li m, Hui Xiong, and Hao Wang. Llm-based agent society investigation: Collaboration and confrontation in avalon gameplay, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.504, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18184" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.161 + ], + "angle": 0, + "content": "[25] Zixing Lei, Yiming Zhang, Yuxin Xiong, and Siheng Chen. Emergent communication in interactive sketch question answering. In Proceedings of the 37th International Conference on Neural Information Processing Systems, Red Hook, NY, USA, 2023. Curran Associates Inc. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.163, + 0.472, + 0.192 + ], + "angle": 0, + "content": "[26] Lei Li, Hongbo Fu, and Chiew-Lan Tai. Fast sketch segmentation and labeling with deep learning, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.192, + 0.472, + 0.248 + ], + "angle": 0, + "content": "[27] Tian Liang, Zhiwei He, Jen tse Huang, Wenxuan Wang, Wenxiang Jiao, Rui Wang, Yujiu Yang, Zhaopeng Tu, Shuming Shi, and Xing Wang. Leveraging word guessing games to assess the intelligence of large language models, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.25, + 0.472, + 0.29 + ], + "angle": 0, + "content": "[28] Jian Lin. Divergence measures based on the shannon entropy. IEEE Transactions on Information Theory, 37(1):145-151, 1991. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.292, + 0.472, + 0.32 + ], + "angle": 0, + "content": "[29] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.322, + 0.472, + 0.376 + ], + "angle": 0, + "content": "[30] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In Advances in Neural Information Processing Systems, pages 34892-34916. Curran Associates, Inc., 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.378, + 0.472, + 0.447 + ], + "angle": 0, + "content": "[31] Shuo Liu, Kaining Ying, Hao Zhang, Yue Yang, Yuqi Lin, Tianle Zhang, Chuanhao Li, Yu Qiao, Ping Luo, Wenqi Shao, and Kaipeng Zhang. Convbench: A multi-turn conversation evaluation benchmark with hierarchical capability for large vision-language models, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.449, + 0.472, + 0.503 + ], + "angle": 0, + "content": "[32] Haoran Mo, Edgar Simo-Serra, Chengying Gao, Changqing Zou, and Ruomei Wang. General virtual sketching framework for vector line art. ACM Transactions on Graphics (TOG), 40:1-14, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.505, + 0.472, + 0.56 + ], + "angle": 0, + "content": "[33] Chong Mou, Xintao Wang, Liangbin Xie, Jing Zhang, Zhonggang Qi, Ying Shan, and Xiaohu Qie. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. ArXiv, abs/2302.08453, 2023. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.562, + 0.472, + 0.59 + ], + "angle": 0, + "content": "[34] Xu Pan and Odelia Schwartz. Multimodal ai needs active human interaction. Nature Human Behaviour, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.591, + 0.472, + 0.618 + ], + "angle": 0, + "content": "[35] Deepika Phutela. The importance of non-verbal communication, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.62, + 0.472, + 0.66 + ], + "angle": 0, + "content": "[36] Yonggang Qi and Zheng-Hua Tan. Sketchsegnet+: An end-to-end learning of rnn for multi-class sketch semantic segmentation. IEEE Access, 7:102717-102726, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.663, + 0.472, + 0.732 + ], + "angle": 0, + "content": "[37] Shuwen Qiu, Sirui Xie, Lifeng Fan, Tao Gao, Jungseock Joo, Song-Chun Zhu, and Yixin Zhu. Emergent graphical conventions in a visual communication game. In Advances in Neural Information Processing Systems, pages 13119-13131. Curran Associates, Inc., 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.734, + 0.472, + 0.844 + ], + "angle": 0, + "content": "[38] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In Proceedings of the 38th International Conference on Machine Learning, ICML 2021, 18-24 July 2021, Virtual Event, pages 8748-8763. PMLR, 2021. 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.846, + 0.472, + 0.9 + ], + "angle": 0, + "content": "[39] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. arXiv preprint arXiv:2112.10752, 2022. 4, 6" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.135 + ], + "angle": 0, + "content": "[40] Patsorn Sangkloy, Nathan Burnell, Cusuh Ham, and James Hays. The sketchy database: Learning to retrieve badly drawn bunnies. ACM Trans. Graph., 35(4), 2016. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.894, + 0.205 + ], + "angle": 0, + "content": "[41] Ravi Kiran Sarvadevabhatla, Shiv Surya, Trisha Mittal, and R. Venkatesh Babu. Pictionary-style word guessing on hand-drawn object sketches: Dataset, analysis and deep network models. IEEE Transactions on Pattern Analysis and Machine Intelligence, 42(1):221-231, 2020. 1, 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.207, + 0.892, + 0.248 + ], + "angle": 0, + "content": "[42] SeungWon Seo, Junhyeok Lee, SeongRae Noh, and HyeongYeop Kang. Llm-based cooperative agents using information relevance and plan validation, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.25, + 0.894, + 0.29 + ], + "angle": 0, + "content": "[43] Chen Shen. Nonverbal communication in intercultural communication: A case study of telling lies. Frontiers in Educational Research, 6(13), 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.292, + 0.892, + 0.32 + ], + "angle": 0, + "content": "[44] Spotify. Annoy: Approximate nearest neighbors in c++/python, n.d. Computer software. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.322, + 0.894, + 0.419 + ], + "angle": 0, + "content": "[45] Alessandro Suglia, Ioannis Konstas, Andrea Vanzo, Emanuele Bastianelli, Desmond Elliott, Stella Frank, and Oliver Lemon. CompGuessWhat?!: A multi-task evaluation framework for grounded language learning. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 7625-7641, Online, 2020. Association for Computational Linguistics. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.421, + 0.894, + 0.517 + ], + "angle": 0, + "content": "[46] Alessandro Suglia, Antonio Vergari, Ioannis Konstas, Yonatan Bisk, Emanuele Bastianelli, Andrea Vanzo, and Oliver Lemon. Imagining grounded conceptual representations from perceptual information in situated guessing games. In Proceedings of the 28th International Conference on Computational Linguistics, pages 1090–1102, Barcelona, Spain (Online), 2020. International Committee on Computational Linguistics. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.519, + 0.892, + 0.574 + ], + "angle": 0, + "content": "[47] F. Wang, S. Lin, H. Wu, H. Li, R. Wang, X. Luo, and X. He. Spfusionnet: Sketch segmentation using multi-modal data fusion. In 2019 IEEE International Conference on Multimedia and Expo (ICME), pages 1654–1659, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.576, + 0.892, + 0.63 + ], + "angle": 0, + "content": "[48] Haohan Wang, Songwei Ge, Zachary Lipton, and Eric P Xing. Learning robust global representations by penalizing local predictive power. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.632, + 0.892, + 0.687 + ], + "angle": 0, + "content": "[49] Wenhai Wang, Enze Xie, Xiaoge Song, Yuhang Zang, Wenjia Wang, Tong Lu, Gang Yu, and Chunhua Shen. Efficient and accurate arbitrary-shaped text detection with pixel aggregation network, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.689, + 0.892, + 0.744 + ], + "angle": 0, + "content": "[50] Yuzhuang Xu, Shuo Wang, Peng Li, Fuwen Luo, Xiaolong Wang, Weidong Liu, and Yang Liu. Exploring large language models for communication games: An empirical study on werewolf. arXiv preprint arXiv:2309.04658, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.746, + 0.892, + 0.787 + ], + "angle": 0, + "content": "[51] Lumin Yang, Jiajie Zhuang, Hongbo Fu, Kun Zhou, and Youyi Zheng. Sketchgen: Semantic sketch segmentation with graph convolutional networks, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.789, + 0.892, + 0.872 + ], + "angle": 0, + "content": "[52] Yuan Yao, Tianyu Yu, Ao Zhang, Chongyi Wang, Junbo Cui, Hongji Zhu, Tianchi Cai, Haoyu Li, Weilin Zhao, Zhihui He, Qianyu Chen, Huarong Zhou, Zhensheng Zou, Haoye Zhang, Shengding Hu, Zhi Zheng, Jie Zhou, Jie Cai, Xu Han, Guoyang Zeng, Dahai Li, Zhiyuan Liu, and Maosong Sun. Minicpm-v: A gpt-4v level mllm on your phone, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.874, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[53] Qian Yu, Feng Liu, Yi-Zhe Song, Tao Xiang, Timothy M. Hospedales, and Chen-Change Loy. Sketch me that shoe. In" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18185" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.11, + 0.092, + 0.47, + 0.12 + ], + "angle": 0, + "content": "Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.122, + 0.472, + 0.177 + ], + "angle": 0, + "content": "[54] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 11941-11952, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.179, + 0.471, + 0.234 + ], + "angle": 0, + "content": "[55] Hua Zhang, Si Liu, Changqing Zhang, Wenqi Ren, Rui Wang, and Xiaochun Cao. Sketchnet: Sketch classification with web images. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 1105-1113, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.236, + 0.471, + 0.277 + ], + "angle": 0, + "content": "[56] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. arXiv preprint arXiv:2302.05543, 2023. 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.278, + 0.471, + 0.346 + ], + "angle": 0, + "content": "[57] Changqing Zou, Qian Yu, Ruofei Du, Haoran Mo, Yi-Zhe Song, Tao Xiang, Chengying Gao, Baoquan Chen, and Hao Zhang. Sketchyscene: Richly-annotated scene sketches. In ECCV, pages 438–454. Springer International Publishing, 2018. 3" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.472, + 0.346 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.956 + ], + "angle": 0, + "content": "18186" + } + ] +] \ No newline at end of file diff --git a/2025/Sketchtopia_ A Dataset and Foundational Agents for Benchmarking Asynchronous Multimodal Communication with Iconic Feedback/fff37445-0c8d-4c24-a3bf-32cf9fcbe8ea_origin.pdf b/2025/Sketchtopia_ A Dataset and Foundational Agents for Benchmarking Asynchronous Multimodal Communication with Iconic Feedback/fff37445-0c8d-4c24-a3bf-32cf9fcbe8ea_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..e9cb35bf0ea9fb3cf008e8a96206f92c1620d17b --- /dev/null +++ b/2025/Sketchtopia_ A Dataset and Foundational Agents for Benchmarking Asynchronous Multimodal Communication with Iconic Feedback/fff37445-0c8d-4c24-a3bf-32cf9fcbe8ea_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d996940b1a31ae0d296004ce2ca21e88079979036beef8f5631b819627965f71 +size 1234069 diff --git a/2025/Sketchtopia_ A Dataset and Foundational Agents for Benchmarking Asynchronous Multimodal Communication with Iconic Feedback/full.md b/2025/Sketchtopia_ A Dataset and Foundational Agents for Benchmarking Asynchronous Multimodal Communication with Iconic Feedback/full.md new file mode 100644 index 0000000000000000000000000000000000000000..c0578137ed4fef144c85625c5a24f44d9539e819 --- /dev/null +++ b/2025/Sketchtopia_ A Dataset and Foundational Agents for Benchmarking Asynchronous Multimodal Communication with Iconic Feedback/full.md @@ -0,0 +1,280 @@ +# Sketchtopia: A Dataset and Foundational Agents for Benchmarking Asynchronous Multimodal Communication with Iconic Feedback + +Mohd Hozafa Khan Ravi Kiran Sarvadevabhatla + +Centre for Visual Information Technology (CVIT), IIIT Hyderabad, India + +{mohd.hozaifa@research., ravi.kiran@}iit.ac.in + +# Abstract + +We introduce Sketchtopia, a large-scale dataset and AI framework designed to explore goal-driven, multimodal communication through asynchronous interactions in a Pictionary-inspired setup. Sketchtopia captures natural human interactions, including freehand sketches, open-ended guesses, and iconic feedback gestures, showcasing the complex dynamics of cooperative communication under constraints. It features over 20K gameplay sessions from 916 players, capturing 263K sketches, 10K erases, 56K guesses and 19.4K iconic feedbacks. We introduce multimodal foundational agents with capabilities for generative sketching, guess generation and asynchronous communication. Our dataset also includes 800 human-agent sessions for benchmarking the agents. We introduce novel metrics to characterize collaborative success, responsiveness to feedback and inter-agent asynchronous communication. Sketchtopia pushes the boundaries of multimodal AI, establishing a new benchmark for studying asynchronous, goal-oriented interactions between humans and AI agents. The dataset can be found at https://sketchtopia25.github.io/ + +# 1. Introduction + +Human communication is inherently multimodal and goaldriven, often relying on non-verbal cues such as gestures and sketches to overcome language and cultural barriers [1, 34, 43]. Non-verbal interactions enable shared understanding despite constraints [20, 35]. Understanding such dynamics is key to developing multimodal AI agents capable of achieving shared goals in constrained environments while communicating asynchronously [6, 21]. + +With this motivation, we study communication-constrained asynchronous interactions in a casual game setting. We introduce Sketchtopia, a large-scale multi-modal dataset containing web-based gameplay episodes of Pictionary, the popular cooperative social game. In each episode, a Painter sketches clues while a Gueisser asynchronously + +attempts to guess the target phrase (Fig. 1). This setup encourages creative, non-verbal communication under time pressure, leveraging iconic feedback gestures (e.g. $\text{山}$ , and $⑦$ ). The rules of Pictionary forbid the Painter from writing text. Coupled with the time limit, this restriction often leads to creative player responses and generated content. + +Unlike prior works [8, 23], Sketchtopia captures authentic, hand-drawn sketches, feedback gestures, and natural editing actions (e.g., undo strokes). Our dataset not only facilitates the study of multimodal goal-driven communication but also serves as a testbed for novel AI agents. We present foundational agents, DRAWBOT and GUESSBOT, designed to asynchronously engage in multimodal interactions, leveraging both visual and textual modalities. This unique asynchronous communication mirrors real-life scenarios where agents must continuously think and act in parallel rather than turn-taking. We make the following contributions: + +- Sketchtopia: A large-scale, multimodal dataset of Pictionary-like game episodes, capturing asynchronous interactions with sketches, guesses, and gestural feedback. +- Two foundational AI agents, DRAWBOT and GUESSBOT, that demonstrate asynchronous, goal-driven multimodal communication with generative capabilities. +- Introduction of novel metrics to assess agent performance in terms of game outcomes, individual contributions, and effectiveness in asynchronous multimodal communication. + +# 2. Related Works + +Sketch datasets: Existing sketch datasets primarily focus on recognition [13, 40], segmentation [22, 26, 36, 47, 49, 51], and retrieval [53], with labels assigned to the final sketch. In contrast, our dataset includes intermediate text "guess" annotations from gameplay, providing temporal grounding for sketches throughout the drawing process. Additionally, our dataset covers a wider array of categories, including abstract concepts (nouns, verbs, adjectives). Some works incorporate text guessing to simulate Pictionary [41], but lack the complexity and richness generated by interacting agents and gameplay dynamics. + +![](images/e4b8a647da52021ce13a75217d1bdde90c76eac112045cb50d8cb961cf51c53e.jpg) +Figure 1. Multimodal Interaction in Sketchtopia: Timeline of key game elements, showing asynchronous responses from Guesser and Painter. Green tick indicates a correct guess by the Guesser. + +
ApproachesShared GoalDrawerGueserVisual ContentDraw ContentGuess ContentFeedback (Multi-modal)Content EditingAsync.
Cont.Grd.
Sketchtopia (Ours)✓✓✓✓✓FreehandPhrase✓(C), C, D✓✓
Iconary [8]✓✓✓✗✓Icons (text)Sent.✗✗✗
Co-Draw [23]✓✓✓✗✓ClipArt (text)Sent.✗✗✗
Tell, Draw, Repeat [14]✗✓✗✗✓ImageSent.✗✗✗
Scores [19]✗✗✗✓✓FreehandSent.✗✗✗
Pixellor [7]✗✓✗✓✗Sketches✗✗✗✗
Visual QA✗✗✓✓✓✗Sent.✗✗✗
SketchGues[41]✗✗✓✓✓✗Phrase✗✗✗
Vis. Dialogue Agents [9]✗✗✓✓✓✗Sent.✗✗✗
Coop. Dialogue Agents [11]✓✗✓✓✓✗Sent.✗✗✗
Embodied VQA [11]✓✗✓✓✓✗Sent.✗✗✗
+ +Table 1. Agents Comparison: A comparative summary of agent capabilities. Cont.: Continuous sketching; Grd.: Grounding of sketch content; Draw/Guess Content: Agent's output format; Async.: Asynchronous communication support; Sent.: Sentence. + +Cooperative Gameplay: Cooperative games with partial observability (CPO) are often used to study shared-goal communication strategies [5, 12]. Guessing games are a popular subcategory; for example, Ashktorab et al. [2, 3] investigate a game where a 'Giver' conveys a word using text. Similarly, Das et al. [10] focus on image retrieval in a text-based guessing game. Other multimodal versions include Charades, where speech-based guesses correspond to physical enactment [4], and Iconary, where a 'Drawer' uses visual icons to help a 'Guesser' interpret a phrase [8]. Suglia et al. [46] address visual guessing in zero-shot OOV cases for Guess-What and CompGuessWhat benchmarks [45]. Unlike the works mentioned above, Sketchtopia introduces significantly more complex gameplay. It involves contextual and generative free-form sketching, an open-ended guesser vocabulary, iconic feedback, and asynchronous, natural communication between players. + +Communication Agents: Several research efforts in Pictionary-like settings introduce agents that focus solely on sketch generation. Scones [19] introduces agents for itera- + +tive sketch creation using detailed textual instructions. Pixel [7] generates sketches efficiently with minimal strokes for a given word. Generally, these works do not involve a guesser in the loop and lack mechanisms for feedback-based sketch adaptation. As a complementary work, Sarvadevabhatla et al. [41] introduce a guesser agent trained to generate human-like guesses in response to iterative rendering of pre-drawn object sketches. However, their approach neglects interactivity and feedback, which are crucial elements of offline Pictionary. + +Recent works introduce Teller and Painter for dynamic, cooperative multimodal dialogue or gameplay. In Tell, Draw, Repeat [14], a drawer generates and modifies images iteratively based on detailed linguistic instructions from a teller. Unlike Pictionary, the teller always has access to the final image and neither multimodal interaction nor a long-term goal. mmDialogue [15] employs multi-modal dialogue in terms of both visual and text generation, but the interactions are not goal-oriented. Agents in Iconary [8] mimic drawings with discrete icons and detailed textual interaction from the guesser. In Co-draw [23], the agents engage in goal-driven collaborative communication, where the teller communicates with the drawer in natural language using discrete clip-art icons instead of sketches. The works of Qiu et al. [37] and Lei et al. [25] explore AI-driven emergent communication. + +LLM Agents: LLMs are increasingly used to test cooperative communication and strategic reasoning in games, focusing on text generation without visual environment processing [24, 27, 50]. REVECA [42] extends this by introducing LLM-based agents interacting in partially observable + +
Works# SketchPr. Mod.Lang. Ground.FormEraseStrokesCat.
IntG.Final
Sketchtopia (Ours)263KTxt, Fdb✓✓Scene✓✓200
TU-Berlin [13]20KXX✓ObjX✓250
Quick Draw [17]50MXX✓ObjX✓345
Sketchy [40]75KXX✓ObjX✓125
SketchyScene [57]29KPhX✓ScXX-
ImageNet-Sketch [48]50KXX✓ObjXX1000
Scene Sketcher1225PhX✓Obj,ScX✓14
SketchyCOCO [16]14KPh, EMX✓Obj,ScX✓17
QUML-Shoe [53]419PhX✓ObjXX1
TU-Berlin Extended [55]20KPhX✓ObjXX250
SketchGuess-160 [41]8960Txt✓✓ObjX✓160
+ +3D scenes. While Vision Language Models (VLMs) handle multi-turn multimodal dialogues [31], they lack multimodal generative capabilities crucial for Pictionary. Moreover, most LLMs and VLMs operate on turn-based rather than full-duplex communication. Our agents, by contrast, are simpler, communicate asynchronously and are more responsive. + +# 3. Dataset + +# 3.1. Game Setup + +Each game session consists of role-specific actions by the two players, viz. Drawer and Guesser. Each session presents a new target word for the Printer to illustrate and the Guesser to speculate. The session ends when the time limit (120 sec) is reached, or when the Printer deems a guess to be correct. + +# 3.2. Data Collection + +The Sketchtopia Web App captures detailed, real-time interactions to emulate the Pictionary experience (see Supplementary video). Data collection involves two phases: Participant Onboarding, where players receive instructions and provide demographic details (age, dexterity), and Gameplay Recording, which logs time-stamped actions, including drawing strokes, iconic feedback, and guesses. Table 3 provides storage formats for collected data. Our dataset supports analysis of both gameplay dynamics and player behavior. Our ethical protocols ensure anonymized data collection, informed consent, and content moderation for privacy and quality. + +# 3.3. Dataset Information + +Target Words: Drawers illustrate words from a pool of 200 terms, ranging from abstract nouns to verbs, covering various difficulty levels. Player Information: Demographics such as age, profession, and dexterity are collected to analyze gameplay strategies and behavior, ensuring fairness and diversity. We also collected 800 human-AI sessions, where players unknowingly interacted with AI agents, allowing + +Table 2. Sketch Dataset Comparison::; Img.: Image; Pr. Mod.: Other modalities; Lang. Ground.: Language Grounding; IntG: Intermediate Guesses;Final: Final guess;Txt: Text; Fdb: Feedback;Sc: Scene; Obj: Object; Cat: Categories; Ph: Photo; EM: Edge Map + +
RoleActionFormatTimestamp
DrawerStrokes1SVG2t1→t2
boolt
boolt
SVGt1→t2
(x,y)t1→t2
GuesserGuessTextt
?boolt
+ +* SVG denotes white stroke over black background + +Table 3. Session Actions and Data Formats + +![](images/18e553ddc6bfd73e14cb6d7a72a92c448f73ab4f42a2d2b2404610841ffe849f.jpg) +Figure 2. Game Distribution: Painter and Guesser actions distributed across time; sorted in decreasing order of success rate. For successful games, drawer actions preceded the guesser actions and usually last a shorter duration + +study of human-AI interaction dynamics. Dataset Quality / Ethical Considerations: We used in-game and post-game text filters to eliminate irrelevant content (e.g. chat-style content such as 'Hi/Hello/lol'). We used text-based filters to remove abusive or inappropriate content in guesses, ensuring a safe and respectful environment. We also conducted manual inspections to identify and address any game rule violations. + +# 3.4. Dataset Analysis + +We analyze Sketchtopia across four key dimensions: global game statistics, session-level performance, player-level insights, and temporal dynamics, revealing gameplay factors that influence success, player behavior, and communication strategies. + +Global game statistics: Our dataset includes 916 players (ages 15–60, avg. 26.6), with 20K sessions totaling 263K strokes, 10K erases, 16.3K highlights, 56K guesses, and 19.4K feedback events. With an $89\%$ success rate, sessions averaged 13 strokes, 2.83 guesses, 0.82 highlights and 1 feedback event. + +Session-level: Success rates varied by target word types, with nouns and verbs showing higher success compared to abstract nouns and adjectives due to their complexity. As illustrated in Fig. 2, sessions for high-success words often + +
OutcomeDFGFDF+GFNo-FB
successful52.330.881.5942.51
failure1.680.180.091.59
+ +Table 4. Games categorized across feedback type and outcome. Numbers represent the percentage of games in each category (outcome or presence of iconic feedback). + +display clear peaks where drawer actions precede guesses, indicating early sketch clarity. Conversely, harder words show flatter distributions, suggesting shifts in strategies or communication challenges. Peaks in the distribution indicate moments where new strategies or ideas emerged, correlating with challenging words. An analysis of success rates across target words and parts of speech (POS) revealed notable variability – Nouns (92.6%), Verbs (84.2%), and Adjectives (79.1%). Among nouns, Abstract Nouns (85%) are the least successful, while food/sports categories are the most. Abstract nouns often require more complex and detailed sketches. Additionally, adjectives (e.g., hard, hollow) which describe nouns are particularly challenging to convey since many objects can possess such properties. + +Player-level insights: At the player level, we identified distinct skills for Drawers and Guessers. Players were slightly more successful as Guessers (14%) than Drawers (9%), yet 77% performed equally well in both roles, emphasizing the shared nature of cooperative gameplay. Categorizing games by feedback and outcome (see Table 4) shows that feedback plays a role in game results. Games that only had drawer feedback had higher success rates, highlighting its importance in guiding Guessers effectively. + +Game Dynamics: To understand gameplay evolution, we analyzed action trajectories using CLIP embeddings [38] to measure similarity between sketches, guesses, and target words. As shown in Figure 3, Drawers refine their sketches over time, while Guessers reactively adjust guesses. Successful games show a convergence pattern where sketches and guesses align closely with the target, demonstrating effective cooperative communication. This game flow analysis captures how players' strategies adapt dynamically throughout the session. + +# 4. Agents + +Our goal is to design player agents that emulate human-like, responsive gameplay, free from rigid turn-based limitations and capable of multimodal interactions. Beyond deciding 'what' action to perform (e.g., a Roller's stroke), agents must also determine 'when' to act (e.g., when a Gueisser should output a guess). The design should have the flexibility to perform a "no-op" (i.e., no immediate action in response to another player's input, allowing the Gueisser to wait for more strokes). Additionally, agents must support not only + +their primary actions (drawing or guessing) but also iconic feedback $(\text{日}, \text{日}, \text{日})$ + +To meet these requirements, we organize the agent design into a two-level module hierarchy (see Fig. 4). At the top level, the ACTIONDECIDER module manages global action timing ('when') and selection ('what'). The lower level comprises specialized Painter and Gueisser modules (DRAWBOT and GUESSBOT), which handle role-specific tasks. The following sections detail the architecture, decision-making processes, and interaction mechanisms of each module within this hierarchy. + +# 4.1. ActionDecider Module + +The ACTIONDECIDER module serves as a lightweight controller, continuously monitoring game states to decide when agents should asynchronously initiate actions. We describe its setup for the Drawing (see Fig. 5). In the first stage, the module processes multimodal game data streams: guesses are encoded using a Text Encoder [54], canvas strokes via a Vision Encoder [54], and iconic feedback through a Feed Forward Network (FFN). These embeddings are cached, with only the latest content appended. A $1 \times 1$ convolution is applied to the guess, canvas, and feedback embeddings to obtain aggregated vector representations. For the Painter, a vector for the Target Word is also included. These are concatenated into a Game State Vector. In the second stage, the Game State Vector is passed through a 'Response' FFN, which determines the Painter's action: Wake up the Painter, Generate Feedback, or Do Nothing (wait). If Generate Feedback is selected, a 'Feedback' FFN produces specific feedback (Q), for the Guesser. + +The Guesser agent follows a similar design but excludes the Target Word. Here, the wakeup action type is routed to GUESSBOT, and feedback actions correspond to the Guesser's iconic feedback (②). + +# 4.2. Drawer Agent (DRAWBOT) + +The primary task of DRAWBOT is to visually convey the target word through sequential actions: (i) incrementally draw freehand strokes to depict the target word, (ii) update the canvas (e.g., erase using the eraser-icon) based on guesses and feedback (②) from the Guesser, and (iii) generate gestural feedback (①), (iv) based on the semantic proximity of guesses to target word. DRAWBOT is built on a fine-tuned Stable Diffusion model [39], tailored to the game's sketch style. Its architecture extends a Text-to-Image (T2I) Diffusion Model [39] with ControlNet-inspired adapters [56], allowing it to condition sketch generation on multimodal inputs such as canvas state, guess history, and game time (see Fig. 6). Each adapter module provides conditioning based on different input modalities (see Fig. 6): + +- Canvas State Adapter: A convolutional layer encodes the + +![](images/ed0e3757a9f56ec55ac9c72153e17bfccefdf1408e7151908d856dd1e9ddd8af.jpg) +Figure 3. Game Flow Analysis: (Sec. 3.4) Each line plot represents a game session. The points on each plot represent CLIP embeddings [38] of canvas or a guess word. The x-axis represents time, y-axis represents normalized distance of these embeddings to the target word (top-left of each panel - blue (easy) and red (hard)). The final canvas for each session is shown. Line plots for paired drawer and guesser have the same color. $\checkmark$ denotes a successful guess and associated canvas, $\times$ represents an unsuccessful session. + +![](images/1d6c257e6527289f724ee8664694973db60ff75958746d4c207f8f3f8469e4a4.jpg) + +![](images/409d0dd424453d191eea625f5b860c873aad7f79873f95aee1e6ceeb40284158.jpg) +Figure 4. Pictionary Agent Framework: The top-level (ActionDecider) decides action type and timing and delegates to $2^{nd}$ level in the hierarchy which perform the individual tasks. Our hierarchical design enables delegation and asynchronous communication. Refer to Sec. 4 for details. + +current visual state of the canvas (at $t = t_1$ ). The result is processed by ControlNet's edge detection layer [56]. This adapter modulates sketch generation by incorporating recent updates, ensuring continuity in the drawing process. + +- Guess Adapter: Encodes the history of guesses, allowing the agent to adjust its drawings according to the Guesser's proximity to the target. +- Time Adapter: A time encoder that regulates the game's pacing and drawing progression, creating a sense of ur + +![](images/6561796846308dabd29a93ec5554b35277dc7fd4fb673677ae065f678d1ea61f.jpg) +Figure 5. Schematic Diagram of the ActionDecider Module for decision timing and action selection. Refer Sec. 4.1. + +gency in line with the game's progress. + +We modify the default T2I Adapter [33] to accommodate non-visual modalities as input. Overall, these adapters enable DRAWBOT to generate incremental strokes rather than complete images, producing residual updates to the canvas (see Fig. 6, $t = t_{2}$ ), thereby maintaining continuity while + +![](images/229b74aecf481b6ea36515d7e2e51fea65134944173bfe698a3c247e735c990a.jpg) +Figure 6. **Drawer Agent (DrawBot) Architecture:** The top left blue box shows the current state of the canvas. The pink module illustrates the primary Text-to-image-based generation module, which processes the target word (e.g., 'toothbrush') to produce strokes. The green module represents additional conditioning inputs — such as prior guesses and current time — that guide the generative process, enabling the agent to create iterative strokes towards a recognizable sketch. The bottom right orange box shows that conditioning outputs are passed to the encoder of the T2I generator. SD stands for Stable Diffusion (A T2I Model) + +![](images/27d48b6934678fe9563787620054ba167e2f14c9c5756e7ef47f2ad238fcac0c.jpg) +Figure 7. Schematic Diagram of the Gueisser Agent's Retrieval-Based Guessing Framework (Sec. 4.3). + +dynamically responding to the evolving game state. + +# 4.3. Gueisser Agent (GUESSBOT) + +The Gueisser Agent (GUESSBOT) generates guesses based on the canvas content and multimodal cues, aiming to identify the hidden target word. Given the limited dataset size, GUESSBOT employs a retrieval-based approach. When GUESSBOT receives a 'Wake-up' signal from the ACTION-DECIDER, it encodes the current canvas state $C_t$ (see Fig. 7) and uses Approximate Nearest Neighbor (ANN) search to compare it with historical game sessions. This strategy enables the agent to utilize past gameplay data, applying rule-based criteria to generate guesses aligned with the current game context. + +The retrieval-based process involves two steps: (i) encoding the canvas content using a vision foundation model + +(CLIP) to retrieve 'k' similar sessions from the database, and (ii) applying rule-based filtering based on similarity scores, guess timings and user feedback to refine the results. This filtering discards irrelevant sessions, for example, by rejecting matches where the timing deviates significantly from the current session. The final output is derived from the best-matched session's guesses, producing the selected guess $(g,t)$ as shown in Fig. 7. + +A key design goal of GUESSBOT is to emulate human gameplay, including natural errors in guessing. For example, human players often overlook spelling errors (e.g., "spyder", see Fig. 3) during immersive gameplay. Our retrieval-based approach inherently accommodates these human-like quirks, enhancing the realism of the agent's responses. + +# 4.4. Implementation Details + +Canvas strokes (SVG) are thickened by a factor of 3 and rendered as $224 \times 224$ images for processing by agents. + +Action Decider: By default, agents remain inactive unless triggered by a 'Wake-up' signal from their Action Decider (see Fig. 4, Fig. 5). Canvas states are captured every 3 seconds, forming the Drawing History. For text encoding (target words, guess history), we use a pre-trained CLIP text encoder [38] (TE), while the vision encoder handles the canvas history (VE). Each FFN is a 3-layer MLP. Every 3 seconds, the Action Decider processes its accumulated inputs and outputs actions. + +DrawBot: The canvas is processed using ControlNet [56] with StableDiffusion-1.5 [39] as the T2I generator (Fig. 6). The T2I Adapter is modified to handle non-visual inputs: each guess is CLIP-encoded and combined with its times-tamp to form a $512 \times 2 \times n_g$ representation, which is integrated into the T2I Generator's U-Net [33]. The generator is fine-tuned on Sketchtopia sketches using LoRA [18] to output incremental strokes (residual images). We optimize the adapters in stages: first training ControlNet, then the T2I Adapters while freezing ControlNet, and finally fin-tuning all components together. The agent is trained by breaking game sessions into fixed time intervals and conditioning outputs on canvas state, guesses, and timestamps, allowing for adaptive sketching based on all input modalities. The residual image output by T2I Generator is converted to its SVG representation using virtual sketching [32], before being added to the canvas. + +GuessBot: A vector database of canvas embeddings (using CLIP ViT-B/32) is built with metadata (guess and timestamp) for fast retrieval using Annoy [44], an approximate nearest-neighbor retriever. Upon receiving a 'Wake-Up' signal, GUESSBOT encodes the current canvas $e_{C_t}$ (see Fig. 7) at current time $t$ , and retrieves the top 10 matches within a time window $[t - \delta, t + \delta]$ , where $\delta = 6$ . Among these, we select the match closest to the query by cosine similarity and output the corresponding guess word. + +Agent Orchestration: All agents are deployed as microservices communicating with the Pictionary WebApp via API calls. Participants are paired randomly, with a $15\%$ chance of getting paired with an agent. During gameplay, the WebApp sends relevant data to the ActionDecider (Sec. 4.1), which delegates tasks to draw or feedback as needed. + +# 5. Metrics for Benchmarking Agents + +We evaluate agents (human-only games and mixed human-agent) in terms of their collaborative success, their responsiveness to feedback, and the similarity of their timing patterns to human counterparts. + +Collaborative Game Success: In goal-oriented games, the primary measure is whether Painter and Guesser agents can effectively collaborate to identify the target word. Metrics include Win Rate (WR): The proportion of games where the target word is guessed correctly, serving as a baseline measure of collaborative success [8]; Soft Win Rate (SWR): Extends Win Rate by considering synonyms of target word, capturing near-success scenarios where agents were close to the correct answer [8, 41]; + +Feedback Responsiveness Score (FRS): measures the extent to which the Guesser adjusts its guesses in response to Painter feedback. After receiving or feedback, if the guesser's next guess is semantically closer to the target word than the guess immediately preceding the feedback, such an occurrence is termed a responsive guess. In a game session, let the count of such occurrences be $s$ , and the Painter takes $p$ and $n$ actions. FRS for that session is defined as $\frac{s}{p + n}$ . The resulting numbers are averaged across sessions to obtain the final metric. + +Asynchronous Action Overlap (AAO): This is a session-level measure and characterizes the degree to which agents act asynchronously. Suppose a Gueisser action's timestamp overlaps with the duration of a Drawer action (e.g. sketching, erasing) within a session. Let the count of all occurrences be $o$ , and let the total number of Gueisser actions be $g$ . AAO for the session is defined as $AAO = \frac{o}{g}$ . Note that a sequential turn-based interaction between agents would have an AAO value of 0, while a purely random and uncoordinated interaction would have a value of 100. + +Multimodal Action Timing Similarity (MATS): For a given target word, suppose the sequence of human Guesser's guess timestamps across all sessions be $T_{g}^{H}$ , and a similar sequence for GUESSBOT be $T_{g}^{A}$ . We define a guess distance $d_{guess}$ between these two sequences as the Jensen-Shannon divergence [28] between their Kernel Density estimates, i.e. $d_{guess} = \text{JENSEN-SHANNON}\{\text{KDE}(T_{g}^{H}), \text{KDE}(T_{g}^{A})\}$ . The average $d_{guess}$ across all target words forms the MATS metric for GUESSBOT. A similar exercise for DRAWBOT's actions, where the timestamp of the stroke duration's midpoint is used, provides the MATS score for DRAWBOT. Detailed definitions for the metrics are provided in the Supplementary. + +![](images/a5ee74fc0add239c475ae5bebdba166623d57dd658b1560c5b6934171d32273b.jpg) +Figure 8. Qualitative comparison between human and agent game outcomes: The figure shows the final canvas and the final guess for select target words (fishing, length, loud). Human outputs are in purple and agent outputs are in blue. Successful games are with green background and unsuccessful ones in red. + +# 6. Results + +To benchmark our agents, we carefully choose a representative evaluation set of game sessions grouped by Part of Speech and compare the performance of humans and agents on this set. In general, the metric values for humans act as ground truth or desirable range, which the agent should achieve. We obtain 800 sessions to be used for testing. The numbers mentioned in Tab. 5 for HH (Human-Human) and HA (Human-Agent) are both computed on this set. + +Collaborative Game Success: Across games, agents had a $68\%$ success rate. The win rate for GUESSBOT is slightly lesser than the human counterpart $(85\%)$ . However, DRAWBOT's win rate is $51\%$ , which is markedly smaller. The Painter agent's bad performance is mainly due to adjectives $(30\%)$ and verbs $(56\%)$ . It performs reasonably well on most nouns $(74\%)$ . The primary reason for bad performance on adjectives is that there is no direct object (e.g. car, bus, icecream) associated with adjectives. Most vision foundation models are trained on objects rather than metaphorical concepts which encode attributes (e.g. hard/soft, length, hollow and loud) which may be associated with multiple objects. Such models need detailed prompts to invoke the relationship and properties of objects. + +Qualitative analysis: Many interesting patterns can be observed in the performance and response patterns of human-human and human-agent games – see Fig. 8. Compared to human counterparts, DRAWBOT's sketches have greater symmetry, with clean stroke lines. Unlike typical generative model outputs, which tend to have a 'synthetic' appearance, + +training on Sketchtopia's data induces a human-like jitter in the generated strokes of DRAWBOT. GUESSBOT's final responses seem human-like for the corresponding canvas contents, even when the game outcome is a failure. This is consistent with the agent's retrieval-based design. Additional examples can be found in Supplementary. + +Feedback Responsiveness: The FRS scores for Human and GUESSBOT are similar (Tab. 5), implying that the agent's ability to adapt its response to feedback is similar to human counterparts. Here, the Gueisser Agent's ActionDecider also plays a key role in 'waking up' GUESSBOT at the right time in response to Painter feedback. + +Characterizing Interaction Dynamics: Similar to FRS, the AAO scores of agents are comparable to humans, indicating comparable levels of asynchronous communication. However, GUESSBOT has a lower AAO score. This is in line with a Guesser's tendency to wait passively for enough strokes to accumulate on the canvas before guessing. The MATS score measures the agent's action pacing compared to humans. Therefore, it is a proxy measure for ActionDecider's performance due to the latter's role as a key asynchronous enabler of actions in each agent. A relatively lower MATS score for GUESSBOT suggests that its action pacing is closer to human. Empirically, we observed that feedback distribution for DRAWBOT was markedly different from human, contributing to its relatively higher MATS score. + +VLMs as Guessers: Vision-Language Models (VLMs) [29, 30, 52] have become a popular choice for image-conditioned reasoning. In our benchmarking context, VLMs could play a Guesser's role. However, despite detailed prompts instructing the model to output a guess phrase and despite providing prior guess history, we observed that VLM models tend to produce verbose and unhelpful outputs. Examples can be found in Supplementary. Moreover, the slower response times and synchronized (turn-taking) paradigm of current VLMs make them unviable compared to our asynchronous and contextually responsive agents. + +Limitations: Despite efforts to minimize biases, the dataset may still reflect factors such as age, handedness, and cultural background, with a predominance of younger, right-handed, English-speaking participants from a limited region. This may affect its generalizability to other languages, cultures, and demographics. While our game interface supports asynchronous interactions, further testing across diverse groups is needed. Additionally, while Sketchtopia includes iconic feedback as a novelty, it is limited to predefined icons (e.g., , , and Ⓞ). This restricts the analysis of more nuanced and complex feedback that could occur in natural interactions. + +# 7. Conclusion + +Sketchtopia enables a scalable framework that pushes the boundaries of multimodal goal-driven communication in asynchronous settings. As the first-ever dataset to include + +
ModelsGame SuccessFeedback ResponsivenessInteraction Dynamics
Win ↑Soft Win ↑FRS ↑AAOMATS ↓
DRAWBOT51.652.4-0.210.39
GUESSBOT81.786.046%0.170.27
Human85.593.440%0.290.00
+ +Table 5. Benchmarking humans and agents via custom-designed metrics (Sec. 5). FRS: Feedback Responsiveness Score; AAO: Asynchronous Action Overlap; MATS: Multimodal Action Timing Similarity. + +rich feedback signals, our collection captures realistic gameplay data with natural human interactions, far beyond what existing datasets offer. Sketchtopia is not just about isolated sketches – it incorporates iconic feedback and dynamic interactions between players, all captured using a robust game interface that seamlessly supports both human-human and human-agent gameplay. This versatility allows us to explore interactions at a depth previously unattainable, establishing a new benchmark for multimodal AI systems. + +A standout contribution of our work is the design of first asynchronous communication agents capable of real-time decision-making in multi-modal formats. Our novel agent architectures emphasize responsiveness and speed, enabling agents to mimic the fluidity of natural human communication without the constraints of turn-taking. By introducing innovative metrics such as Asynchronous Action Overlap and Feedback Responsiveness Scores, we redefine how to evaluate collaborative success and interaction dynamics in asynchronous multi-modal communication settings. + +Benchmarking our agents against human players demonstrates that they can generate human-like actions and responses. Our results reveal that even state-of-the-art VLMs lag significantly in handling asynchronous, multimodal communication complexities. This establishes Sketchtopia as a critical benchmark for advancing AI research in this space. Additionally, our game flow visualizations offer a novel lens into understanding how agents and humans progress towards shared goals, providing insights into the evolution of strategies and collaborative problem-solving. + +Sketchtopia's applications extend beyond Pictionary. Its interactive features enable sketch scene editing and collaborative whiteboard tools, with potential for live note-taking and narration. The dataset's focus on abstract sketches also supports cognitive science research into communication strategies. In summary, Sketchtopia provides a valuable resource for exploring asynchronous, multimodal AI systems and realistic agent-based gameplay. By tackling the complexities of real-time, feedback-driven communication, we aim to inspire AI advancements that are more natural, adaptive, and truly collaborative. + +# References + +[1] Michael Argyle. The syntaxes of bodily communication. Linguistics, 11(112):71-92, 1973. 1 +[2] Zahra Ashktorab, Q. Vera Liao, Casey Dugan, James Johnson, Qian Pan, Wei Zhang, Sadhana Kumaravel, and Murray Campbell. Human-ai collaboration in a cooperative game setting: Measuring social perception and outcomes. Proc. ACM Hum.-Comput. Interact., 4(CSCW2), 2020. 2 +[3] Zahra Ashktorab, Casey Dugan, James Johnson, Qian Pan, Wei Zhang, Sadhana Kumaravel, and Murray Campbell. Effects of communication directionality and ai agent differences in human-ai interaction. In Proceedings of the 2021 CHI Conference on Human Factors in Computing Systems, New York, NY, USA, 2021. Association for Computing Machinery. 2 +[4] Muhammad Fauzi Bafadal and Humaira Humaira. The use of charades games in teaching vocabulary to the junior high school students. Linguistics and Elt Journal, 2019. 2 +[5] Nolan Bard, Jakob N. Foerster, Sarath Chandar, Neil Burch, Marc Lanctot, H. Francis Song, Emilio Parisotto, Vincent Dumoulin, Subhodeep Moitra, Edward Hughes, Iain Dunning, Shibl Mourad, Hugo Larochelle, Marc G. Bellemare, and Michael Bowling. The hanabi challenge: A new frontier for ai research. Artificial Intelligence, 280:103216, 2020. 2 +[6] Fabien Beltrame. Comportement non verbal et situation d'interaction: Quand la gestuelle des mains revendique un role cognitif de support des représentations pour la résolution de problème, 2006. 1 +[7] Ayan Kumar Bhunia, Ayan Das, Umar Riaz Muhammad, Yongxin Yang, Timothy M. Hospedales, Tao Xiang, Yulia Gryaditskaya, and Yi-Zhe Song. Pixel: a competitive sketching ai agent. so you think you can sketch? ACM Trans. Graph., 39(6), 2020. 2 +[8] Christopher Clark, Jordi Salvador, Dustin Schwenk, Derrick Bonafilia, Mark Yatskar, Eric Kolve, Alvaro Herrasti, Jonghyun Choi, Sachin Mehta, Sam Skjonsberg, Carissa Schoenick, Aaron Sarnat, Hannaneh Hajishirzi, Aniruddha Kembhavi, Oren Etzioni, and Ali Farhadi. Iconary: A pictorial-based game for testing multimodal communication with drawings and text. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 1864-1886, Online and Punta Cana, Dominican Republic, 2021. Association for Computational Linguistics. 1, 2, 7 +[9] Abhishek Das, Satwik Kottur, Khushi Gupta, Avi Singh, Deshraj Yadav, Jose M. F. Moura, Devi Parikh, and Dhruv Batra. Visual dialog. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017. +[10] A. Das, S. Kottur, J. F. Moura, S. Lee, and D. Batra. Learning cooperative visual dialog agents with deep reinforcement learning. In 2017 IEEE International Conference on Computer Vision (ICCV), pages 2970-2979, Los Alamitos, CA, USA, 2017. IEEE Computer Society. 2 +[11] Abhishek Das, Samyak Datta, Georgia Gkioxari, Stefan Lee, Devi Parikh, and Dhruv Batra. Embodied question answering. + +In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 2 +[12] Markus Eger, Chris Martens, and Marcela Alfaro Cordoba. An intentional ai for hanabi. In 2017 IEEE Conference on Computational Intelligence and Games (CIG), page 68-75. IEEE Press, 2017. 2 +[13] Mathias Eitz, James Hays, and Marc Alexa. How do humans sketch objects? ACM Trans. Graph. (Proc. SIGGRAPH), 31 (4):44:1-44:10, 2012. 1, 3 +[14] Alaaeldin El-Nouby, Shikhar Sharma, Hannes Schulz, Devon Hjelm, Layla El Asri, Samira Ebrahimi Kahou, Yoshua Bengio, and Graham W. Taylor. Tell, draw, and repeat: Generating and modifying images based on continual linguistic instruction. In The IEEE International Conference on Computer Vision (ICCV), 2019. 2 +[15] Jiazhan Feng, Qingfeng Sun, Can Xu, Pu Zhao, Yaming Yang, Chongyang Tao, Dongyan Zhao, and Qingwei Lin. MMDialog: A large-scale multi-turn dialogue dataset towards multimodal open-domain conversation. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 7348-7363, Toronto, Canada, 2023. Association for Computational Linguistics. 2 +[16] Chengying Gao, Qi Liu, Qi Xu, Limin Wang, Jianzhuang Liu, and Changqing Zou. Sketchycoco: Image generation from freehand scene sketches. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5174-5183, 2020. 3 +[17] David Ha and Douglas Eck. A neural representation of sketch drawings. In ICLR 2018, 2018. 2018. 3 +[18] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In International Conference on Learning Representations, 2022. 6 +[19] Forrest Huang, Eldon Schoop, David Ha, and John Canny. Scones: towards conversational authoring of sketches. In Proceedings of the 25th International Conference on Intelligent User Interfaces, page 313–323, New York, NY, USA, 2020. Association for Computing Machinery. 2 +[20] Kristiina Jokinen. Gaze and gesture activity in communication. pages 537-546. Springer, Berlin, Heidelberg, 2009. 1 +[21] Kristiina Jokinen. Non-verbal signals for turn-taking and feedback. European Language Resources Association (ELRA), 2010. 1 +[22] Kurmanbek Kaiyrbekov and Metin Sezgin. Stroke-based sketched symbol reconstruction and segmentation, 2019. 1 +[23] Jin-Hwa Kim, Nikita Kitaev, Xinlei Chen, Marcus Rohrbach, Byoung-Tak Zhang, Yuandong Tian, Dhruv Batra, and Devi Parikh. CoDraw: Collaborative drawing as a testbed for grounded goal-driven communication. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 6495-6513, Florence, Italy, 2019. Association for Computational Linguistics. 1, 2 +[24] Yihuai Lan, Zhiqiang Hu, Lei Wang, Yang Wang, Deheng Ye, Peilin Zhao, Ee-Peng Li m, Hui Xiong, and Hao Wang. Llm-based agent society investigation: Collaboration and confrontation in avalon gameplay, 2024. 2 + +[25] Zixing Lei, Yiming Zhang, Yuxin Xiong, and Siheng Chen. Emergent communication in interactive sketch question answering. In Proceedings of the 37th International Conference on Neural Information Processing Systems, Red Hook, NY, USA, 2023. Curran Associates Inc. 2 +[26] Lei Li, Hongbo Fu, and Chiew-Lan Tai. Fast sketch segmentation and labeling with deep learning, 2018. 1 +[27] Tian Liang, Zhiwei He, Jen tse Huang, Wenxuan Wang, Wenxiang Jiao, Rui Wang, Yujiu Yang, Zhaopeng Tu, Shuming Shi, and Xing Wang. Leveraging word guessing games to assess the intelligence of large language models, 2023. 2 +[28] Jian Lin. Divergence measures based on the shannon entropy. IEEE Transactions on Information Theory, 37(1):145-151, 1991. 7 +[29] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning, 2023. 8 +[30] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In Advances in Neural Information Processing Systems, pages 34892-34916. Curran Associates, Inc., 2023. 8 +[31] Shuo Liu, Kaining Ying, Hao Zhang, Yue Yang, Yuqi Lin, Tianle Zhang, Chuanhao Li, Yu Qiao, Ping Luo, Wenqi Shao, and Kaipeng Zhang. Convbench: A multi-turn conversation evaluation benchmark with hierarchical capability for large vision-language models, 2024. 3 +[32] Haoran Mo, Edgar Simo-Serra, Chengying Gao, Changqing Zou, and Ruomei Wang. General virtual sketching framework for vector line art. ACM Transactions on Graphics (TOG), 40:1-14, 2021. 6 +[33] Chong Mou, Xintao Wang, Liangbin Xie, Jing Zhang, Zhonggang Qi, Ying Shan, and Xiaohu Qie. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. ArXiv, abs/2302.08453, 2023. 5, 6 +[34] Xu Pan and Odelia Schwartz. Multimodal ai needs active human interaction. Nature Human Behaviour, 2024. 1 +[35] Deepika Phutela. The importance of non-verbal communication, 2016. 1 +[36] Yonggang Qi and Zheng-Hua Tan. Sketchsegnet+: An end-to-end learning of rnn for multi-class sketch semantic segmentation. IEEE Access, 7:102717-102726, 2019. 1 +[37] Shuwen Qiu, Sirui Xie, Lifeng Fan, Tao Gao, Jungseock Joo, Song-Chun Zhu, and Yixin Zhu. Emergent graphical conventions in a visual communication game. In Advances in Neural Information Processing Systems, pages 13119-13131. Curran Associates, Inc., 2022. 2 +[38] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In Proceedings of the 38th International Conference on Machine Learning, ICML 2021, 18-24 July 2021, Virtual Event, pages 8748-8763. PMLR, 2021. 4, 5, 6 +[39] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. arXiv preprint arXiv:2112.10752, 2022. 4, 6 + +[40] Patsorn Sangkloy, Nathan Burnell, Cusuh Ham, and James Hays. The sketchy database: Learning to retrieve badly drawn bunnies. ACM Trans. Graph., 35(4), 2016. 1, 3 +[41] Ravi Kiran Sarvadevabhatla, Shiv Surya, Trisha Mittal, and R. Venkatesh Babu. Pictionary-style word guessing on hand-drawn object sketches: Dataset, analysis and deep network models. IEEE Transactions on Pattern Analysis and Machine Intelligence, 42(1):221-231, 2020. 1, 2, 3, 7 +[42] SeungWon Seo, Junhyeok Lee, SeongRae Noh, and HyeongYeop Kang. Llm-based cooperative agents using information relevance and plan validation, 2024. 2 +[43] Chen Shen. Nonverbal communication in intercultural communication: A case study of telling lies. Frontiers in Educational Research, 6(13), 2023. 1 +[44] Spotify. Annoy: Approximate nearest neighbors in c++/python, n.d. Computer software. 6 +[45] Alessandro Suglia, Ioannis Konstas, Andrea Vanzo, Emanuele Bastianelli, Desmond Elliott, Stella Frank, and Oliver Lemon. CompGuessWhat?!: A multi-task evaluation framework for grounded language learning. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 7625-7641, Online, 2020. Association for Computational Linguistics. 2 +[46] Alessandro Suglia, Antonio Vergari, Ioannis Konstas, Yonatan Bisk, Emanuele Bastianelli, Andrea Vanzo, and Oliver Lemon. Imagining grounded conceptual representations from perceptual information in situated guessing games. In Proceedings of the 28th International Conference on Computational Linguistics, pages 1090–1102, Barcelona, Spain (Online), 2020. International Committee on Computational Linguistics. 2 +[47] F. Wang, S. Lin, H. Wu, H. Li, R. Wang, X. Luo, and X. He. Spfusionnet: Sketch segmentation using multi-modal data fusion. In 2019 IEEE International Conference on Multimedia and Expo (ICME), pages 1654–1659, 2019. 1 +[48] Haohan Wang, Songwei Ge, Zachary Lipton, and Eric P Xing. Learning robust global representations by penalizing local predictive power. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2019. 3 +[49] Wenhai Wang, Enze Xie, Xiaoge Song, Yuhang Zang, Wenjia Wang, Tong Lu, Gang Yu, and Chunhua Shen. Efficient and accurate arbitrary-shaped text detection with pixel aggregation network, 2020. 1 +[50] Yuzhuang Xu, Shuo Wang, Peng Li, Fuwen Luo, Xiaolong Wang, Weidong Liu, and Yang Liu. Exploring large language models for communication games: An empirical study on werewolf. arXiv preprint arXiv:2309.04658, 2023. 2 +[51] Lumin Yang, Jiajie Zhuang, Hongbo Fu, Kun Zhou, and Youyi Zheng. Sketchgen: Semantic sketch segmentation with graph convolutional networks, 2020. 1 +[52] Yuan Yao, Tianyu Yu, Ao Zhang, Chongyi Wang, Junbo Cui, Hongji Zhu, Tianchi Cai, Haoyu Li, Weilin Zhao, Zhihui He, Qianyu Chen, Huarong Zhou, Zhensheng Zou, Haoye Zhang, Shengding Hu, Zhi Zheng, Jie Zhou, Jie Cai, Xu Han, Guoyang Zeng, Dahai Li, Zhiyuan Liu, and Maosong Sun. Minicpm-v: A gpt-4v level mllm on your phone, 2024. 8 +[53] Qian Yu, Feng Liu, Yi-Zhe Song, Tao Xiang, Timothy M. Hospedales, and Chen-Change Loy. Sketch me that shoe. In + +Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 1, 3 +[54] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 11941-11952, 2023. 4 +[55] Hua Zhang, Si Liu, Changqing Zhang, Wenqi Ren, Rui Wang, and Xiaochun Cao. Sketchnet: Sketch classification with web images. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 1105-1113, 2016. 3 +[56] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. arXiv preprint arXiv:2302.05543, 2023. 4, 5, 6 +[57] Changqing Zou, Qian Yu, Ruofei Du, Haoran Mo, Yi-Zhe Song, Tao Xiang, Chengying Gao, Baoquan Chen, and Hao Zhang. Sketchyscene: Richly-annotated scene sketches. In ECCV, pages 438–454. Springer International Publishing, 2018. 3 \ No newline at end of file diff --git a/2025/Sketchtopia_ A Dataset and Foundational Agents for Benchmarking Asynchronous Multimodal Communication with Iconic Feedback/images.zip b/2025/Sketchtopia_ A Dataset and Foundational Agents for Benchmarking Asynchronous Multimodal Communication with Iconic Feedback/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..4fa13947995be11215f4f6081e2e84359b008aeb --- /dev/null +++ b/2025/Sketchtopia_ A Dataset and Foundational Agents for Benchmarking Asynchronous Multimodal Communication with Iconic Feedback/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e11620be6d57aa765dcdf9e66dcbc390c06ad19fb11ce288ce15c704c84d8c01 +size 421644 diff --git a/2025/Sketchtopia_ A Dataset and Foundational Agents for Benchmarking Asynchronous Multimodal Communication with Iconic Feedback/layout.json b/2025/Sketchtopia_ A Dataset and Foundational Agents for Benchmarking Asynchronous Multimodal Communication with Iconic Feedback/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..6f6a55ab3a0cc3edf339fd48c7e491a19b89d6b5 --- /dev/null +++ b/2025/Sketchtopia_ A Dataset and Foundational Agents for Benchmarking Asynchronous Multimodal Communication with Iconic Feedback/layout.json @@ -0,0 +1,7266 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 87, + 103, + 507, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 87, + 103, + 507, + 140 + ], + "spans": [ + { + "bbox": [ + 87, + 103, + 507, + 140 + ], + "type": "text", + "content": "Sketchtopia: A Dataset and Foundational Agents for Benchmarking Asynchronous Multimodal Communication with Iconic Feedback" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 170, + 161, + 423, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 170, + 161, + 423, + 175 + ], + "spans": [ + { + "bbox": [ + 170, + 161, + 423, + 175 + ], + "type": "text", + "content": "Mohd Hozafa Khan Ravi Kiran Sarvadevabhatla" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 118, + 175, + 474, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 175, + 474, + 190 + ], + "spans": [ + { + "bbox": [ + 118, + 175, + 474, + 190 + ], + "type": "text", + "content": "Centre for Visual Information Technology (CVIT), IIIT Hyderabad, India" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 168, + 191, + 423, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 168, + 191, + 423, + 202 + ], + "spans": [ + { + "bbox": [ + 168, + 191, + 423, + 202 + ], + "type": "text", + "content": "{mohd.hozaifa@research., ravi.kiran@}iit.ac.in" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "spans": [ + { + "bbox": [ + 143, + 231, + 192, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 45, + 257, + 290, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 45, + 257, + 290, + 498 + ], + "spans": [ + { + "bbox": [ + 45, + 257, + 290, + 498 + ], + "type": "text", + "content": "We introduce Sketchtopia, a large-scale dataset and AI framework designed to explore goal-driven, multimodal communication through asynchronous interactions in a Pictionary-inspired setup. Sketchtopia captures natural human interactions, including freehand sketches, open-ended guesses, and iconic feedback gestures, showcasing the complex dynamics of cooperative communication under constraints. It features over 20K gameplay sessions from 916 players, capturing 263K sketches, 10K erases, 56K guesses and 19.4K iconic feedbacks. We introduce multimodal foundational agents with capabilities for generative sketching, guess generation and asynchronous communication. Our dataset also includes 800 human-agent sessions for benchmarking the agents. We introduce novel metrics to characterize collaborative success, responsiveness to feedback and inter-agent asynchronous communication. Sketchtopia pushes the boundaries of multimodal AI, establishing a new benchmark for studying asynchronous, goal-oriented interactions between humans and AI agents. The dataset can be found at https://sketchtopia25.github.io/" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 523, + 128, + 536 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 523, + 128, + 536 + ], + "spans": [ + { + "bbox": [ + 47, + 523, + 128, + 536 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 544, + 288, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 544, + 288, + 640 + ], + "spans": [ + { + "bbox": [ + 46, + 544, + 288, + 640 + ], + "type": "text", + "content": "Human communication is inherently multimodal and goaldriven, often relying on non-verbal cues such as gestures and sketches to overcome language and cultural barriers [1, 34, 43]. Non-verbal interactions enable shared understanding despite constraints [20, 35]. Understanding such dynamics is key to developing multimodal AI agents capable of achieving shared goals in constrained environments while communicating asynchronously [6, 21]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 642, + 288, + 714 + ], + "type": "text", + "content": "With this motivation, we study communication-constrained asynchronous interactions in a casual game setting. We introduce Sketchtopia, a large-scale multi-modal dataset containing web-based gameplay episodes of Pictionary, the popular cooperative social game. In each episode, a Painter sketches clues while a Gueisser asynchronously" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 232, + 547, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 232, + 547, + 304 + ], + "spans": [ + { + "bbox": [ + 304, + 232, + 547, + 304 + ], + "type": "text", + "content": "attempts to guess the target phrase (Fig. 1). This setup encourages creative, non-verbal communication under time pressure, leveraging iconic feedback gestures (e.g. " + }, + { + "bbox": [ + 304, + 232, + 547, + 304 + ], + "type": "inline_equation", + "content": "\\text{山}" + }, + { + "bbox": [ + 304, + 232, + 547, + 304 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 232, + 547, + 304 + ], + "type": "inline_equation", + "content": "⑦" + }, + { + "bbox": [ + 304, + 232, + 547, + 304 + ], + "type": "text", + "content": "). The rules of Pictionary forbid the Painter from writing text. Coupled with the time limit, this restriction often leads to creative player responses and generated content." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 304, + 548, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 304, + 548, + 437 + ], + "spans": [ + { + "bbox": [ + 304, + 304, + 548, + 437 + ], + "type": "text", + "content": "Unlike prior works [8, 23], Sketchtopia captures authentic, hand-drawn sketches, feedback gestures, and natural editing actions (e.g., undo strokes). Our dataset not only facilitates the study of multimodal goal-driven communication but also serves as a testbed for novel AI agents. We present foundational agents, DRAWBOT and GUESSBOT, designed to asynchronously engage in multimodal interactions, leveraging both visual and textual modalities. This unique asynchronous communication mirrors real-life scenarios where agents must continuously think and act in parallel rather than turn-taking. We make the following contributions:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 306, + 440, + 547, + 547 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 306, + 440, + 547, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 440, + 547, + 475 + ], + "spans": [ + { + "bbox": [ + 306, + 440, + 547, + 475 + ], + "type": "text", + "content": "- Sketchtopia: A large-scale, multimodal dataset of Pictionary-like game episodes, capturing asynchronous interactions with sketches, guesses, and gestural feedback." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 476, + 547, + 512 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 476, + 547, + 512 + ], + "spans": [ + { + "bbox": [ + 306, + 476, + 547, + 512 + ], + "type": "text", + "content": "- Two foundational AI agents, DRAWBOT and GUESSBOT, that demonstrate asynchronous, goal-driven multimodal communication with generative capabilities." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 512, + 547, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 512, + 547, + 547 + ], + "spans": [ + { + "bbox": [ + 306, + 512, + 547, + 547 + ], + "type": "text", + "content": "- Introduction of novel metrics to assess agent performance in terms of game outcomes, individual contributions, and effectiveness in asynchronous multimodal communication." + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 560, + 398, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 560, + 398, + 574 + ], + "spans": [ + { + "bbox": [ + 306, + 560, + 398, + 574 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 547, + 715 + ], + "type": "text", + "content": "Sketch datasets: Existing sketch datasets primarily focus on recognition [13, 40], segmentation [22, 26, 36, 47, 49, 51], and retrieval [53], with labels assigned to the final sketch. In contrast, our dataset includes intermediate text \"guess\" annotations from gameplay, providing temporal grounding for sketches throughout the drawing process. Additionally, our dataset covers a wider array of categories, including abstract concepts (nouns, verbs, adjectives). Some works incorporate text guessing to simulate Pictionary [41], but lack the complexity and richness generated by interacting agents and gameplay dynamics." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "18176" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 71, + 545, + 281 + ], + "blocks": [ + { + "bbox": [ + 50, + 71, + 545, + 281 + ], + "lines": [ + { + "bbox": [ + 50, + 71, + 545, + 281 + ], + "spans": [ + { + "bbox": [ + 50, + 71, + 545, + 281 + ], + "type": "image", + "image_path": "e4b8a647da52021ce13a75217d1bdde90c76eac112045cb50d8cb961cf51c53e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 289, + 546, + 313 + ], + "lines": [ + { + "bbox": [ + 46, + 289, + 546, + 313 + ], + "spans": [ + { + "bbox": [ + 46, + 289, + 546, + 313 + ], + "type": "text", + "content": "Figure 1. Multimodal Interaction in Sketchtopia: Timeline of key game elements, showing asynchronous responses from Guesser and Painter. Green tick indicates a correct guess by the Guesser." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 47, + 330, + 288, + 399 + ], + "blocks": [ + { + "bbox": [ + 47, + 330, + 288, + 399 + ], + "lines": [ + { + "bbox": [ + 47, + 330, + 288, + 399 + ], + "spans": [ + { + "bbox": [ + 47, + 330, + 288, + 399 + ], + "type": "table", + "html": "
ApproachesShared GoalDrawerGueserVisual ContentDraw ContentGuess ContentFeedback (Multi-modal)Content EditingAsync.
Cont.Grd.
Sketchtopia (Ours)✓✓✓✓✓FreehandPhrase✓(C), C, D✓✓
Iconary [8]✓✓✓✗✓Icons (text)Sent.✗✗✗
Co-Draw [23]✓✓✓✗✓ClipArt (text)Sent.✗✗✗
Tell, Draw, Repeat [14]✗✓✗✗✓ImageSent.✗✗✗
Scores [19]✗✗✗✓✓FreehandSent.✗✗✗
Pixellor [7]✗✓✗✓✗Sketches✗✗✗✗
Visual QA✗✗✓✓✓✗Sent.✗✗✗
SketchGues[41]✗✗✓✓✓✗Phrase✗✗✗
Vis. Dialogue Agents [9]✗✗✓✓✓✗Sent.✗✗✗
Coop. Dialogue Agents [11]✓✗✓✓✓✗Sent.✗✗✗
Embodied VQA [11]✓✗✓✓✓✗Sent.✗✗✗
", + "image_path": "8a60281168ca79abe2f2ff9f863b3b8c10b821b70049a81987d10a7e6a8bfb9b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 408, + 288, + 453 + ], + "lines": [ + { + "bbox": [ + 46, + 408, + 288, + 453 + ], + "spans": [ + { + "bbox": [ + 46, + 408, + 288, + 453 + ], + "type": "text", + "content": "Table 1. Agents Comparison: A comparative summary of agent capabilities. Cont.: Continuous sketching; Grd.: Grounding of sketch content; Draw/Guess Content: Agent's output format; Async.: Asynchronous communication support; Sent.: Sentence." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 474, + 288, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 474, + 288, + 676 + ], + "spans": [ + { + "bbox": [ + 46, + 474, + 288, + 676 + ], + "type": "text", + "content": "Cooperative Gameplay: Cooperative games with partial observability (CPO) are often used to study shared-goal communication strategies [5, 12]. Guessing games are a popular subcategory; for example, Ashktorab et al. [2, 3] investigate a game where a 'Giver' conveys a word using text. Similarly, Das et al. [10] focus on image retrieval in a text-based guessing game. Other multimodal versions include Charades, where speech-based guesses correspond to physical enactment [4], and Iconary, where a 'Drawer' uses visual icons to help a 'Guesser' interpret a phrase [8]. Suglia et al. [46] address visual guessing in zero-shot OOV cases for Guess-What and CompGuessWhat benchmarks [45]. Unlike the works mentioned above, Sketchtopia introduces significantly more complex gameplay. It involves contextual and generative free-form sketching, an open-ended guesser vocabulary, iconic feedback, and asynchronous, natural communication between players." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 677, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 288, + 713 + ], + "type": "text", + "content": "Communication Agents: Several research efforts in Pictionary-like settings introduce agents that focus solely on sketch generation. Scones [19] introduces agents for itera-" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 332, + 547, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 332, + 547, + 453 + ], + "spans": [ + { + "bbox": [ + 304, + 332, + 547, + 453 + ], + "type": "text", + "content": "tive sketch creation using detailed textual instructions. Pixel [7] generates sketches efficiently with minimal strokes for a given word. Generally, these works do not involve a guesser in the loop and lack mechanisms for feedback-based sketch adaptation. As a complementary work, Sarvadevabhatla et al. [41] introduce a guesser agent trained to generate human-like guesses in response to iterative rendering of pre-drawn object sketches. However, their approach neglects interactivity and feedback, which are crucial elements of offline Pictionary." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 457, + 547, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 457, + 547, + 649 + ], + "spans": [ + { + "bbox": [ + 304, + 457, + 547, + 649 + ], + "type": "text", + "content": "Recent works introduce Teller and Painter for dynamic, cooperative multimodal dialogue or gameplay. In Tell, Draw, Repeat [14], a drawer generates and modifies images iteratively based on detailed linguistic instructions from a teller. Unlike Pictionary, the teller always has access to the final image and neither multimodal interaction nor a long-term goal. mmDialogue [15] employs multi-modal dialogue in terms of both visual and text generation, but the interactions are not goal-oriented. Agents in Iconary [8] mimic drawings with discrete icons and detailed textual interaction from the guesser. In Co-draw [23], the agents engage in goal-driven collaborative communication, where the teller communicates with the drawer in natural language using discrete clip-art icons instead of sketches. The works of Qiu et al. [37] and Lei et al. [25] explore AI-driven emergent communication." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 654, + 548, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 654, + 548, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 654, + 548, + 715 + ], + "type": "text", + "content": "LLM Agents: LLMs are increasingly used to test cooperative communication and strategic reasoning in games, focusing on text generation without visual environment processing [24, 27, 50]. REVECA [42] extends this by introducing LLM-based agents interacting in partially observable" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18177" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 45, + 70, + 288, + 167 + ], + "blocks": [ + { + "bbox": [ + 45, + 70, + 288, + 167 + ], + "lines": [ + { + "bbox": [ + 45, + 70, + 288, + 167 + ], + "spans": [ + { + "bbox": [ + 45, + 70, + 288, + 167 + ], + "type": "table", + "html": "
Works# SketchPr. Mod.Lang. Ground.FormEraseStrokesCat.
IntG.Final
Sketchtopia (Ours)263KTxt, Fdb✓✓Scene✓✓200
TU-Berlin [13]20KXX✓ObjX✓250
Quick Draw [17]50MXX✓ObjX✓345
Sketchy [40]75KXX✓ObjX✓125
SketchyScene [57]29KPhX✓ScXX-
ImageNet-Sketch [48]50KXX✓ObjXX1000
Scene Sketcher1225PhX✓Obj,ScX✓14
SketchyCOCO [16]14KPh, EMX✓Obj,ScX✓17
QUML-Shoe [53]419PhX✓ObjXX1
TU-Berlin Extended [55]20KPhX✓ObjXX250
SketchGuess-160 [41]8960Txt✓✓ObjX✓160
", + "image_path": "62d47bc5eb30b3241bd4a5fae8c46520b7590f42287a524c7d68364eb4b88379.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 251, + 289, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 251, + 289, + 335 + ], + "spans": [ + { + "bbox": [ + 46, + 251, + 289, + 335 + ], + "type": "text", + "content": "3D scenes. While Vision Language Models (VLMs) handle multi-turn multimodal dialogues [31], they lack multimodal generative capabilities crucial for Pictionary. Moreover, most LLMs and VLMs operate on turn-based rather than full-duplex communication. Our agents, by contrast, are simpler, communicate asynchronously and are more responsive." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 347, + 101, + 359 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 347, + 101, + 359 + ], + "spans": [ + { + "bbox": [ + 47, + 347, + 101, + 359 + ], + "type": "text", + "content": "3. Dataset" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 367, + 127, + 380 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 367, + 127, + 380 + ], + "spans": [ + { + "bbox": [ + 47, + 367, + 127, + 380 + ], + "type": "text", + "content": "3.1. Game Setup" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 385, + 288, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 385, + 288, + 445 + ], + "spans": [ + { + "bbox": [ + 46, + 385, + 288, + 445 + ], + "type": "text", + "content": "Each game session consists of role-specific actions by the two players, viz. Drawer and Guesser. Each session presents a new target word for the Printer to illustrate and the Guesser to speculate. The session ends when the time limit (120 sec) is reached, or when the Printer deems a guess to be correct." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 453, + 143, + 464 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 453, + 143, + 464 + ], + "spans": [ + { + "bbox": [ + 47, + 453, + 143, + 464 + ], + "type": "text", + "content": "3.2. Data Collection" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 472, + 288, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 472, + 288, + 604 + ], + "spans": [ + { + "bbox": [ + 46, + 472, + 288, + 604 + ], + "type": "text", + "content": "The Sketchtopia Web App captures detailed, real-time interactions to emulate the Pictionary experience (see Supplementary video). Data collection involves two phases: Participant Onboarding, where players receive instructions and provide demographic details (age, dexterity), and Gameplay Recording, which logs time-stamped actions, including drawing strokes, iconic feedback, and guesses. Table 3 provides storage formats for collected data. Our dataset supports analysis of both gameplay dynamics and player behavior. Our ethical protocols ensure anonymized data collection, informed consent, and content moderation for privacy and quality." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 611, + 165, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 611, + 165, + 623 + ], + "spans": [ + { + "bbox": [ + 47, + 611, + 165, + 623 + ], + "type": "text", + "content": "3.3. Dataset Information" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 629, + 288, + 715 + ], + "type": "text", + "content": "Target Words: Drawers illustrate words from a pool of 200 terms, ranging from abstract nouns to verbs, covering various difficulty levels. Player Information: Demographics such as age, profession, and dexterity are collected to analyze gameplay strategies and behavior, ensuring fairness and diversity. We also collected 800 human-AI sessions, where players unknowingly interacted with AI agents, allowing" + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 313, + 70, + 540, + 184 + ], + "blocks": [ + { + "bbox": [ + 46, + 175, + 289, + 232 + ], + "lines": [ + { + "bbox": [ + 46, + 175, + 289, + 232 + ], + "spans": [ + { + "bbox": [ + 46, + 175, + 289, + 232 + ], + "type": "text", + "content": "Table 2. Sketch Dataset Comparison::; Img.: Image; Pr. Mod.: Other modalities; Lang. Ground.: Language Grounding; IntG: Intermediate Guesses;Final: Final guess;Txt: Text; Fdb: Feedback;Sc: Scene; Obj: Object; Cat: Categories; Ph: Photo; EM: Edge Map" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 313, + 70, + 540, + 184 + ], + "lines": [ + { + "bbox": [ + 313, + 70, + 540, + 184 + ], + "spans": [ + { + "bbox": [ + 313, + 70, + 540, + 184 + ], + "type": "table", + "html": "
RoleActionFormatTimestamp
DrawerStrokes1SVG2t1→t2
boolt
boolt
SVGt1→t2
(x,y)t1→t2
GuesserGuessTextt
?boolt
", + "image_path": "45a4bc4117225c4d7674a576216cee2ddc3bca5f68581f00d9d3b2e2f4891eb9.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 321, + 185, + 533, + 198 + ], + "lines": [ + { + "bbox": [ + 321, + 185, + 533, + 198 + ], + "spans": [ + { + "bbox": [ + 321, + 185, + 533, + 198 + ], + "type": "text", + "content": "* SVG denotes white stroke over black background" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 10 + }, + { + "bbox": [ + 346, + 207, + 504, + 217 + ], + "lines": [ + { + "bbox": [ + 346, + 207, + 504, + 217 + ], + "spans": [ + { + "bbox": [ + 346, + 207, + 504, + 217 + ], + "type": "text", + "content": "Table 3. Session Actions and Data Formats" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 306, + 229, + 545, + 339 + ], + "blocks": [ + { + "bbox": [ + 306, + 229, + 545, + 339 + ], + "lines": [ + { + "bbox": [ + 306, + 229, + 545, + 339 + ], + "spans": [ + { + "bbox": [ + 306, + 229, + 545, + 339 + ], + "type": "image", + "image_path": "18e553ddc6bfd73e14cb6d7a72a92c448f73ab4f42a2d2b2404610841ffe849f.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 347, + 547, + 392 + ], + "lines": [ + { + "bbox": [ + 305, + 347, + 547, + 392 + ], + "spans": [ + { + "bbox": [ + 305, + 347, + 547, + 392 + ], + "type": "text", + "content": "Figure 2. Game Distribution: Painter and Guesser actions distributed across time; sorted in decreasing order of success rate. For successful games, drawer actions preceded the guesser actions and usually last a shorter duration" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 412, + 547, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 412, + 547, + 508 + ], + "spans": [ + { + "bbox": [ + 304, + 412, + 547, + 508 + ], + "type": "text", + "content": "study of human-AI interaction dynamics. Dataset Quality / Ethical Considerations: We used in-game and post-game text filters to eliminate irrelevant content (e.g. chat-style content such as 'Hi/Hello/lol'). We used text-based filters to remove abusive or inappropriate content in guesses, ensuring a safe and respectful environment. We also conducted manual inspections to identify and address any game rule violations." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 306, + 515, + 406, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 515, + 406, + 528 + ], + "spans": [ + { + "bbox": [ + 306, + 515, + 406, + 528 + ], + "type": "text", + "content": "3.4. Dataset Analysis" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 304, + 533, + 547, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 533, + 547, + 594 + ], + "spans": [ + { + "bbox": [ + 304, + 533, + 547, + 594 + ], + "type": "text", + "content": "We analyze Sketchtopia across four key dimensions: global game statistics, session-level performance, player-level insights, and temporal dynamics, revealing gameplay factors that influence success, player behavior, and communication strategies." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 594, + 547, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 547, + 665 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 547, + 665 + ], + "type": "text", + "content": "Global game statistics: Our dataset includes 916 players (ages 15–60, avg. 26.6), with 20K sessions totaling 263K strokes, 10K erases, 16.3K highlights, 56K guesses, and 19.4K feedback events. With an " + }, + { + "bbox": [ + 304, + 594, + 547, + 665 + ], + "type": "inline_equation", + "content": "89\\%" + }, + { + "bbox": [ + 304, + 594, + 547, + 665 + ], + "type": "text", + "content": " success rate, sessions averaged 13 strokes, 2.83 guesses, 0.82 highlights and 1 feedback event." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 713 + ], + "type": "text", + "content": "Session-level: Success rates varied by target word types, with nouns and verbs showing higher success compared to abstract nouns and adjectives due to their complexity. As illustrated in Fig. 2, sessions for high-success words often" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18178" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 66, + 70, + 269, + 119 + ], + "blocks": [ + { + "bbox": [ + 66, + 70, + 269, + 119 + ], + "lines": [ + { + "bbox": [ + 66, + 70, + 269, + 119 + ], + "spans": [ + { + "bbox": [ + 66, + 70, + 269, + 119 + ], + "type": "table", + "html": "
OutcomeDFGFDF+GFNo-FB
successful52.330.881.5942.51
failure1.680.180.091.59
", + "image_path": "2d7599f349ca10e9f1ac19a14cd410933ab6413e758fb73cd6be7747818e8dc3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 127, + 288, + 160 + ], + "lines": [ + { + "bbox": [ + 46, + 127, + 288, + 160 + ], + "spans": [ + { + "bbox": [ + 46, + 127, + 288, + 160 + ], + "type": "text", + "content": "Table 4. Games categorized across feedback type and outcome. Numbers represent the percentage of games in each category (outcome or presence of iconic feedback)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 180, + 287, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 180, + 287, + 348 + ], + "spans": [ + { + "bbox": [ + 46, + 180, + 287, + 348 + ], + "type": "text", + "content": "display clear peaks where drawer actions precede guesses, indicating early sketch clarity. Conversely, harder words show flatter distributions, suggesting shifts in strategies or communication challenges. Peaks in the distribution indicate moments where new strategies or ideas emerged, correlating with challenging words. An analysis of success rates across target words and parts of speech (POS) revealed notable variability – Nouns (92.6%), Verbs (84.2%), and Adjectives (79.1%). Among nouns, Abstract Nouns (85%) are the least successful, while food/sports categories are the most. Abstract nouns often require more complex and detailed sketches. Additionally, adjectives (e.g., hard, hollow) which describe nouns are particularly challenging to convey since many objects can possess such properties." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 348, + 287, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 348, + 287, + 456 + ], + "spans": [ + { + "bbox": [ + 46, + 348, + 287, + 456 + ], + "type": "text", + "content": "Player-level insights: At the player level, we identified distinct skills for Drawers and Guessers. Players were slightly more successful as Guessers (14%) than Drawers (9%), yet 77% performed equally well in both roles, emphasizing the shared nature of cooperative gameplay. Categorizing games by feedback and outcome (see Table 4) shows that feedback plays a role in game results. Games that only had drawer feedback had higher success rates, highlighting its importance in guiding Guessers effectively." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 456, + 287, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 456, + 287, + 575 + ], + "spans": [ + { + "bbox": [ + 46, + 456, + 287, + 575 + ], + "type": "text", + "content": "Game Dynamics: To understand gameplay evolution, we analyzed action trajectories using CLIP embeddings [38] to measure similarity between sketches, guesses, and target words. As shown in Figure 3, Drawers refine their sketches over time, while Guessers reactively adjust guesses. Successful games show a convergence pattern where sketches and guesses align closely with the target, demonstrating effective cooperative communication. This game flow analysis captures how players' strategies adapt dynamically throughout the session." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 586, + 97, + 600 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 586, + 97, + 600 + ], + "spans": [ + { + "bbox": [ + 47, + 586, + 97, + 600 + ], + "type": "text", + "content": "4. Agents" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 715 + ], + "type": "text", + "content": "Our goal is to design player agents that emulate human-like, responsive gameplay, free from rigid turn-based limitations and capable of multimodal interactions. Beyond deciding 'what' action to perform (e.g., a Roller's stroke), agents must also determine 'when' to act (e.g., when a Gueisser should output a guess). The design should have the flexibility to perform a \"no-op\" (i.e., no immediate action in response to another player's input, allowing the Gueisser to wait for more strokes). Additionally, agents must support not only" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "spans": [ + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "text", + "content": "their primary actions (drawing or guessing) but also iconic feedback " + }, + { + "bbox": [ + 305, + 72, + 545, + 96 + ], + "type": "inline_equation", + "content": "(\\text{日}, \\text{日}, \\text{日})" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 96, + 547, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 96, + 547, + 204 + ], + "spans": [ + { + "bbox": [ + 304, + 96, + 547, + 204 + ], + "type": "text", + "content": "To meet these requirements, we organize the agent design into a two-level module hierarchy (see Fig. 4). At the top level, the ACTIONDECIDER module manages global action timing ('when') and selection ('what'). The lower level comprises specialized Painter and Gueisser modules (DRAWBOT and GUESSBOT), which handle role-specific tasks. The following sections detail the architecture, decision-making processes, and interaction mechanisms of each module within this hierarchy." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 213, + 435, + 224 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 213, + 435, + 224 + ], + "spans": [ + { + "bbox": [ + 305, + 213, + 435, + 224 + ], + "type": "text", + "content": "4.1. ActionDecider Module" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 231, + 547, + 457 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 231, + 547, + 457 + ], + "spans": [ + { + "bbox": [ + 304, + 231, + 547, + 457 + ], + "type": "text", + "content": "The ACTIONDECIDER module serves as a lightweight controller, continuously monitoring game states to decide when agents should asynchronously initiate actions. We describe its setup for the Drawing (see Fig. 5). In the first stage, the module processes multimodal game data streams: guesses are encoded using a Text Encoder [54], canvas strokes via a Vision Encoder [54], and iconic feedback through a Feed Forward Network (FFN). These embeddings are cached, with only the latest content appended. A " + }, + { + "bbox": [ + 304, + 231, + 547, + 457 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 304, + 231, + 547, + 457 + ], + "type": "text", + "content": " convolution is applied to the guess, canvas, and feedback embeddings to obtain aggregated vector representations. For the Painter, a vector for the Target Word is also included. These are concatenated into a Game State Vector. In the second stage, the Game State Vector is passed through a 'Response' FFN, which determines the Painter's action: Wake up the Painter, Generate Feedback, or Do Nothing (wait). If Generate Feedback is selected, a 'Feedback' FFN produces specific feedback (Q), for the Guesser." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 459, + 545, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 459, + 545, + 506 + ], + "spans": [ + { + "bbox": [ + 304, + 459, + 545, + 506 + ], + "type": "text", + "content": "The Guesser agent follows a similar design but excludes the Target Word. Here, the wakeup action type is routed to GUESSBOT, and feedback actions correspond to the Guesser's iconic feedback (②)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 305, + 515, + 451, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 515, + 451, + 528 + ], + "spans": [ + { + "bbox": [ + 305, + 515, + 451, + 528 + ], + "type": "text", + "content": "4.2. Drawer Agent (DRAWBOT)" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 533, + 547, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 533, + 547, + 700 + ], + "spans": [ + { + "bbox": [ + 304, + 533, + 547, + 700 + ], + "type": "text", + "content": "The primary task of DRAWBOT is to visually convey the target word through sequential actions: (i) incrementally draw freehand strokes to depict the target word, (ii) update the canvas (e.g., erase using the eraser-icon) based on guesses and feedback (②) from the Guesser, and (iii) generate gestural feedback (①), (iv) based on the semantic proximity of guesses to target word. DRAWBOT is built on a fine-tuned Stable Diffusion model [39], tailored to the game's sketch style. Its architecture extends a Text-to-Image (T2I) Diffusion Model [39] with ControlNet-inspired adapters [56], allowing it to condition sketch generation on multimodal inputs such as canvas state, guess history, and game time (see Fig. 6). Each adapter module provides conditioning based on different input modalities (see Fig. 6):" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 306, + 701, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 701, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 306, + 701, + 545, + 713 + ], + "type": "text", + "content": "- Canvas State Adapter: A convolutional layer encodes the" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18179" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 71, + 298, + 377 + ], + "blocks": [ + { + "bbox": [ + 50, + 71, + 298, + 377 + ], + "lines": [ + { + "bbox": [ + 50, + 71, + 298, + 377 + ], + "spans": [ + { + "bbox": [ + 50, + 71, + 298, + 377 + ], + "type": "image", + "image_path": "ed0e3757a9f56ec55ac9c72153e17bfccefdf1408e7151908d856dd1e9ddd8af.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 386, + 546, + 431 + ], + "lines": [ + { + "bbox": [ + 46, + 386, + 546, + 431 + ], + "spans": [ + { + "bbox": [ + 46, + 386, + 546, + 431 + ], + "type": "text", + "content": "Figure 3. Game Flow Analysis: (Sec. 3.4) Each line plot represents a game session. The points on each plot represent CLIP embeddings [38] of canvas or a guess word. The x-axis represents time, y-axis represents normalized distance of these embeddings to the target word (top-left of each panel - blue (easy) and red (hard)). The final canvas for each session is shown. Line plots for paired drawer and guesser have the same color. " + }, + { + "bbox": [ + 46, + 386, + 546, + 431 + ], + "type": "inline_equation", + "content": "\\checkmark" + }, + { + "bbox": [ + 46, + 386, + 546, + 431 + ], + "type": "text", + "content": " denotes a successful guess and associated canvas, " + }, + { + "bbox": [ + 46, + 386, + 546, + 431 + ], + "type": "inline_equation", + "content": "\\times" + }, + { + "bbox": [ + 46, + 386, + 546, + 431 + ], + "type": "text", + "content": " represents an unsuccessful session." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 302, + 72, + 543, + 376 + ], + "blocks": [ + { + "bbox": [ + 302, + 72, + 543, + 376 + ], + "lines": [ + { + "bbox": [ + 302, + 72, + 543, + 376 + ], + "spans": [ + { + "bbox": [ + 302, + 72, + 543, + 376 + ], + "type": "image", + "image_path": "1d6c257e6527289f724ee8664694973db60ff75958746d4c207f8f3f8469e4a4.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 53, + 449, + 281, + 516 + ], + "blocks": [ + { + "bbox": [ + 53, + 449, + 281, + 516 + ], + "lines": [ + { + "bbox": [ + 53, + 449, + 281, + 516 + ], + "spans": [ + { + "bbox": [ + 53, + 449, + 281, + 516 + ], + "type": "image", + "image_path": "409d0dd424453d191eea625f5b860c873aad7f79873f95aee1e6ceeb40284158.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 523, + 288, + 578 + ], + "lines": [ + { + "bbox": [ + 46, + 523, + 288, + 578 + ], + "spans": [ + { + "bbox": [ + 46, + 523, + 288, + 578 + ], + "type": "text", + "content": "Figure 4. Pictionary Agent Framework: The top-level (ActionDecider) decides action type and timing and delegates to " + }, + { + "bbox": [ + 46, + 523, + 288, + 578 + ], + "type": "inline_equation", + "content": "2^{nd}" + }, + { + "bbox": [ + 46, + 523, + 288, + 578 + ], + "type": "text", + "content": " level in the hierarchy which perform the individual tasks. Our hierarchical design enables delegation and asynchronous communication. Refer to Sec. 4 for details." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 605, + 287, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 605, + 287, + 654 + ], + "spans": [ + { + "bbox": [ + 54, + 605, + 287, + 654 + ], + "type": "text", + "content": "current visual state of the canvas (at " + }, + { + "bbox": [ + 54, + 605, + 287, + 654 + ], + "type": "inline_equation", + "content": "t = t_1" + }, + { + "bbox": [ + 54, + 605, + 287, + 654 + ], + "type": "text", + "content": "). The result is processed by ControlNet's edge detection layer [56]. This adapter modulates sketch generation by incorporating recent updates, ensuring continuity in the drawing process." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 654, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 47, + 654, + 287, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 654, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 47, + 654, + 287, + 689 + ], + "type": "text", + "content": "- Guess Adapter: Encodes the history of guesses, allowing the agent to adjust its drawings according to the Guesser's proximity to the target." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "type": "text", + "content": "- Time Adapter: A time encoder that regulates the game's pacing and drawing progression, creating a sense of ur" + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "type": "image", + "bbox": [ + 308, + 449, + 545, + 585 + ], + "blocks": [ + { + "bbox": [ + 308, + 449, + 545, + 585 + ], + "lines": [ + { + "bbox": [ + 308, + 449, + 545, + 585 + ], + "spans": [ + { + "bbox": [ + 308, + 449, + 545, + 585 + ], + "type": "image", + "image_path": "6561796846308dabd29a93ec5554b35277dc7fd4fb673677ae065f678d1ea61f.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 592, + 547, + 615 + ], + "lines": [ + { + "bbox": [ + 305, + 592, + 547, + 615 + ], + "spans": [ + { + "bbox": [ + 305, + 592, + 547, + 615 + ], + "type": "text", + "content": "Figure 5. Schematic Diagram of the ActionDecider Module for decision timing and action selection. Refer Sec. 4.1." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 640, + 473, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 640, + 473, + 652 + ], + "spans": [ + { + "bbox": [ + 314, + 640, + 473, + 652 + ], + "type": "text", + "content": "gency in line with the game's progress." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 653, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 547, + 714 + ], + "type": "text", + "content": "We modify the default T2I Adapter [33] to accommodate non-visual modalities as input. Overall, these adapters enable DRAWBOT to generate incremental strokes rather than complete images, producing residual updates to the canvas (see Fig. 6, " + }, + { + "bbox": [ + 304, + 653, + 547, + 714 + ], + "type": "inline_equation", + "content": "t = t_{2}" + }, + { + "bbox": [ + 304, + 653, + 547, + 714 + ], + "type": "text", + "content": "), thereby maintaining continuity while" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18180" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 50, + 71, + 286, + 232 + ], + "blocks": [ + { + "bbox": [ + 50, + 71, + 286, + 232 + ], + "lines": [ + { + "bbox": [ + 50, + 71, + 286, + 232 + ], + "spans": [ + { + "bbox": [ + 50, + 71, + 286, + 232 + ], + "type": "image", + "image_path": "229b74aecf481b6ea36515d7e2e51fea65134944173bfe698a3c247e735c990a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 240, + 289, + 350 + ], + "lines": [ + { + "bbox": [ + 46, + 240, + 289, + 350 + ], + "spans": [ + { + "bbox": [ + 46, + 240, + 289, + 350 + ], + "type": "text", + "content": "Figure 6. **Drawer Agent (DrawBot) Architecture:** The top left blue box shows the current state of the canvas. The pink module illustrates the primary Text-to-image-based generation module, which processes the target word (e.g., 'toothbrush') to produce strokes. The green module represents additional conditioning inputs — such as prior guesses and current time — that guide the generative process, enabling the agent to create iterative strokes towards a recognizable sketch. The bottom right orange box shows that conditioning outputs are passed to the encoder of the T2I generator. SD stands for Stable Diffusion (A T2I Model)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 85, + 364, + 252, + 464 + ], + "blocks": [ + { + "bbox": [ + 85, + 364, + 252, + 464 + ], + "lines": [ + { + "bbox": [ + 85, + 364, + 252, + 464 + ], + "spans": [ + { + "bbox": [ + 85, + 364, + 252, + 464 + ], + "type": "image", + "image_path": "27d48b6934678fe9563787620054ba167e2f14c9c5756e7ef47f2ad238fcac0c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 472, + 288, + 495 + ], + "lines": [ + { + "bbox": [ + 47, + 472, + 288, + 495 + ], + "spans": [ + { + "bbox": [ + 47, + 472, + 288, + 495 + ], + "type": "text", + "content": "Figure 7. Schematic Diagram of the Gueisser Agent's Retrieval-Based Guessing Framework (Sec. 4.3)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 517, + 255, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 517, + 255, + 529 + ], + "spans": [ + { + "bbox": [ + 47, + 517, + 255, + 529 + ], + "type": "text", + "content": "dynamically responding to the evolving game state." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 538, + 195, + 551 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 538, + 195, + 551 + ], + "spans": [ + { + "bbox": [ + 47, + 538, + 195, + 551 + ], + "type": "text", + "content": "4.3. Gueisser Agent (GUESSBOT)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 557, + 288, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 557, + 288, + 688 + ], + "spans": [ + { + "bbox": [ + 46, + 557, + 288, + 688 + ], + "type": "text", + "content": "The Gueisser Agent (GUESSBOT) generates guesses based on the canvas content and multimodal cues, aiming to identify the hidden target word. Given the limited dataset size, GUESSBOT employs a retrieval-based approach. When GUESSBOT receives a 'Wake-up' signal from the ACTION-DECIDER, it encodes the current canvas state " + }, + { + "bbox": [ + 46, + 557, + 288, + 688 + ], + "type": "inline_equation", + "content": "C_t" + }, + { + "bbox": [ + 46, + 557, + 288, + 688 + ], + "type": "text", + "content": " (see Fig. 7) and uses Approximate Nearest Neighbor (ANN) search to compare it with historical game sessions. This strategy enables the agent to utilize past gameplay data, applying rule-based criteria to generate guesses aligned with the current game context." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 288, + 714 + ], + "type": "text", + "content": "The retrieval-based process involves two steps: (i) encoding the canvas content using a vision foundation model" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 72, + 547, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 168 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 168 + ], + "type": "text", + "content": "(CLIP) to retrieve 'k' similar sessions from the database, and (ii) applying rule-based filtering based on similarity scores, guess timings and user feedback to refine the results. This filtering discards irrelevant sessions, for example, by rejecting matches where the timing deviates significantly from the current session. The final output is derived from the best-matched session's guesses, producing the selected guess " + }, + { + "bbox": [ + 304, + 72, + 547, + 168 + ], + "type": "inline_equation", + "content": "(g,t)" + }, + { + "bbox": [ + 304, + 72, + 547, + 168 + ], + "type": "text", + "content": " as shown in Fig. 7." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 168, + 547, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 168, + 547, + 240 + ], + "spans": [ + { + "bbox": [ + 304, + 168, + 547, + 240 + ], + "type": "text", + "content": "A key design goal of GUESSBOT is to emulate human gameplay, including natural errors in guessing. For example, human players often overlook spelling errors (e.g., \"spyder\", see Fig. 3) during immersive gameplay. Our retrieval-based approach inherently accommodates these human-like quirks, enhancing the realism of the agent's responses." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 305, + 250, + 440, + 262 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 250, + 440, + 262 + ], + "spans": [ + { + "bbox": [ + 305, + 250, + 440, + 262 + ], + "type": "text", + "content": "4.4. Implementation Details" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 269, + 545, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 269, + 545, + 293 + ], + "spans": [ + { + "bbox": [ + 304, + 269, + 545, + 293 + ], + "type": "text", + "content": "Canvas strokes (SVG) are thickened by a factor of 3 and rendered as " + }, + { + "bbox": [ + 304, + 269, + 545, + 293 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 304, + 269, + 545, + 293 + ], + "type": "text", + "content": " images for processing by agents." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 293, + 546, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 293, + 546, + 400 + ], + "spans": [ + { + "bbox": [ + 304, + 293, + 546, + 400 + ], + "type": "text", + "content": "Action Decider: By default, agents remain inactive unless triggered by a 'Wake-up' signal from their Action Decider (see Fig. 4, Fig. 5). Canvas states are captured every 3 seconds, forming the Drawing History. For text encoding (target words, guess history), we use a pre-trained CLIP text encoder [38] (TE), while the vision encoder handles the canvas history (VE). Each FFN is a 3-layer MLP. Every 3 seconds, the Action Decider processes its accumulated inputs and outputs actions." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 402, + 547, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 402, + 547, + 605 + ], + "spans": [ + { + "bbox": [ + 304, + 402, + 547, + 605 + ], + "type": "text", + "content": "DrawBot: The canvas is processed using ControlNet [56] with StableDiffusion-1.5 [39] as the T2I generator (Fig. 6). The T2I Adapter is modified to handle non-visual inputs: each guess is CLIP-encoded and combined with its times-tamp to form a " + }, + { + "bbox": [ + 304, + 402, + 547, + 605 + ], + "type": "inline_equation", + "content": "512 \\times 2 \\times n_g" + }, + { + "bbox": [ + 304, + 402, + 547, + 605 + ], + "type": "text", + "content": " representation, which is integrated into the T2I Generator's U-Net [33]. The generator is fine-tuned on Sketchtopia sketches using LoRA [18] to output incremental strokes (residual images). We optimize the adapters in stages: first training ControlNet, then the T2I Adapters while freezing ControlNet, and finally fin-tuning all components together. The agent is trained by breaking game sessions into fixed time intervals and conditioning outputs on canvas state, guesses, and timestamps, allowing for adaptive sketching based on all input modalities. The residual image output by T2I Generator is converted to its SVG representation using virtual sketching [32], before being added to the canvas." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 605, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 605, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 605, + 547, + 714 + ], + "type": "text", + "content": "GuessBot: A vector database of canvas embeddings (using CLIP ViT-B/32) is built with metadata (guess and timestamp) for fast retrieval using Annoy [44], an approximate nearest-neighbor retriever. Upon receiving a 'Wake-Up' signal, GUESSBOT encodes the current canvas " + }, + { + "bbox": [ + 304, + 605, + 547, + 714 + ], + "type": "inline_equation", + "content": "e_{C_t}" + }, + { + "bbox": [ + 304, + 605, + 547, + 714 + ], + "type": "text", + "content": " (see Fig. 7) at current time " + }, + { + "bbox": [ + 304, + 605, + 547, + 714 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 605, + 547, + 714 + ], + "type": "text", + "content": ", and retrieves the top 10 matches within a time window " + }, + { + "bbox": [ + 304, + 605, + 547, + 714 + ], + "type": "inline_equation", + "content": "[t - \\delta, t + \\delta]" + }, + { + "bbox": [ + 304, + 605, + 547, + 714 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 605, + 547, + 714 + ], + "type": "inline_equation", + "content": "\\delta = 6" + }, + { + "bbox": [ + 304, + 605, + 547, + 714 + ], + "type": "text", + "content": ". Among these, we select the match closest to the query by cosine similarity and output the corresponding guess word." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18181" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "text", + "content": "Agent Orchestration: All agents are deployed as microservices communicating with the Pictionary WebApp via API calls. Participants are paired randomly, with a " + }, + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "inline_equation", + "content": "15\\%" + }, + { + "bbox": [ + 46, + 72, + 289, + 145 + ], + "type": "text", + "content": " chance of getting paired with an agent. During gameplay, the WebApp sends relevant data to the ActionDecider (Sec. 4.1), which delegates tasks to draw or feedback as needed." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 154, + 238, + 168 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 154, + 238, + 168 + ], + "spans": [ + { + "bbox": [ + 47, + 154, + 238, + 168 + ], + "type": "text", + "content": "5. Metrics for Benchmarking Agents" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 174, + 288, + 221 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 174, + 288, + 221 + ], + "spans": [ + { + "bbox": [ + 46, + 174, + 288, + 221 + ], + "type": "text", + "content": "We evaluate agents (human-only games and mixed human-agent) in terms of their collaborative success, their responsiveness to feedback, and the similarity of their timing patterns to human counterparts." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 222, + 288, + 329 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 222, + 288, + 329 + ], + "spans": [ + { + "bbox": [ + 46, + 222, + 288, + 329 + ], + "type": "text", + "content": "Collaborative Game Success: In goal-oriented games, the primary measure is whether Painter and Guesser agents can effectively collaborate to identify the target word. Metrics include Win Rate (WR): The proportion of games where the target word is guessed correctly, serving as a baseline measure of collaborative success [8]; Soft Win Rate (SWR): Extends Win Rate by considering synonyms of target word, capturing near-success scenarios where agents were close to the correct answer [8, 41];" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 330, + 288, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 330, + 288, + 449 + ], + "spans": [ + { + "bbox": [ + 46, + 330, + 288, + 449 + ], + "type": "text", + "content": "Feedback Responsiveness Score (FRS): measures the extent to which the Guesser adjusts its guesses in response to Painter feedback. After receiving or feedback, if the guesser's next guess is semantically closer to the target word than the guess immediately preceding the feedback, such an occurrence is termed a responsive guess. In a game session, let the count of such occurrences be " + }, + { + "bbox": [ + 46, + 330, + 288, + 449 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 46, + 330, + 288, + 449 + ], + "type": "text", + "content": ", and the Painter takes " + }, + { + "bbox": [ + 46, + 330, + 288, + 449 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 46, + 330, + 288, + 449 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 330, + 288, + 449 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 46, + 330, + 288, + 449 + ], + "type": "text", + "content": " actions. FRS for that session is defined as " + }, + { + "bbox": [ + 46, + 330, + 288, + 449 + ], + "type": "inline_equation", + "content": "\\frac{s}{p + n}" + }, + { + "bbox": [ + 46, + 330, + 288, + 449 + ], + "type": "text", + "content": ". The resulting numbers are averaged across sessions to obtain the final metric." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 449, + 288, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 449, + 288, + 567 + ], + "spans": [ + { + "bbox": [ + 46, + 449, + 288, + 567 + ], + "type": "text", + "content": "Asynchronous Action Overlap (AAO): This is a session-level measure and characterizes the degree to which agents act asynchronously. Suppose a Gueisser action's timestamp overlaps with the duration of a Drawer action (e.g. sketching, erasing) within a session. Let the count of all occurrences be " + }, + { + "bbox": [ + 46, + 449, + 288, + 567 + ], + "type": "inline_equation", + "content": "o" + }, + { + "bbox": [ + 46, + 449, + 288, + 567 + ], + "type": "text", + "content": ", and let the total number of Gueisser actions be " + }, + { + "bbox": [ + 46, + 449, + 288, + 567 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 46, + 449, + 288, + 567 + ], + "type": "text", + "content": ". AAO for the session is defined as " + }, + { + "bbox": [ + 46, + 449, + 288, + 567 + ], + "type": "inline_equation", + "content": "AAO = \\frac{o}{g}" + }, + { + "bbox": [ + 46, + 449, + 288, + 567 + ], + "type": "text", + "content": ". Note that a sequential turn-based interaction between agents would have an AAO value of 0, while a purely random and uncoordinated interaction would have a value of 100." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 568, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 568, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 568, + 289, + 715 + ], + "type": "text", + "content": "Multimodal Action Timing Similarity (MATS): For a given target word, suppose the sequence of human Guesser's guess timestamps across all sessions be " + }, + { + "bbox": [ + 46, + 568, + 289, + 715 + ], + "type": "inline_equation", + "content": "T_{g}^{H}" + }, + { + "bbox": [ + 46, + 568, + 289, + 715 + ], + "type": "text", + "content": ", and a similar sequence for GUESSBOT be " + }, + { + "bbox": [ + 46, + 568, + 289, + 715 + ], + "type": "inline_equation", + "content": "T_{g}^{A}" + }, + { + "bbox": [ + 46, + 568, + 289, + 715 + ], + "type": "text", + "content": ". We define a guess distance " + }, + { + "bbox": [ + 46, + 568, + 289, + 715 + ], + "type": "inline_equation", + "content": "d_{guess}" + }, + { + "bbox": [ + 46, + 568, + 289, + 715 + ], + "type": "text", + "content": " between these two sequences as the Jensen-Shannon divergence [28] between their Kernel Density estimates, i.e. " + }, + { + "bbox": [ + 46, + 568, + 289, + 715 + ], + "type": "inline_equation", + "content": "d_{guess} = \\text{JENSEN-SHANNON}\\{\\text{KDE}(T_{g}^{H}), \\text{KDE}(T_{g}^{A})\\}" + }, + { + "bbox": [ + 46, + 568, + 289, + 715 + ], + "type": "text", + "content": ". The average " + }, + { + "bbox": [ + 46, + 568, + 289, + 715 + ], + "type": "inline_equation", + "content": "d_{guess}" + }, + { + "bbox": [ + 46, + 568, + 289, + 715 + ], + "type": "text", + "content": " across all target words forms the MATS metric for GUESSBOT. A similar exercise for DRAWBOT's actions, where the timestamp of the stroke duration's midpoint is used, provides the MATS score for DRAWBOT. Detailed definitions for the metrics are provided in the Supplementary." + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 310, + 72, + 545, + 269 + ], + "blocks": [ + { + "bbox": [ + 310, + 72, + 545, + 269 + ], + "lines": [ + { + "bbox": [ + 310, + 72, + 545, + 269 + ], + "spans": [ + { + "bbox": [ + 310, + 72, + 545, + 269 + ], + "type": "image", + "image_path": "a5ee74fc0add239c475ae5bebdba166623d57dd658b1560c5b6934171d32273b.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 276, + 547, + 331 + ], + "lines": [ + { + "bbox": [ + 304, + 276, + 547, + 331 + ], + "spans": [ + { + "bbox": [ + 304, + 276, + 547, + 331 + ], + "type": "text", + "content": "Figure 8. Qualitative comparison between human and agent game outcomes: The figure shows the final canvas and the final guess for select target words (fishing, length, loud). Human outputs are in purple and agent outputs are in blue. Successful games are with green background and unsuccessful ones in red." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 354, + 359, + 367 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 354, + 359, + 367 + ], + "spans": [ + { + "bbox": [ + 306, + 354, + 359, + 367 + ], + "type": "text", + "content": "6. Results" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 375, + 547, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 375, + 547, + 471 + ], + "spans": [ + { + "bbox": [ + 304, + 375, + 547, + 471 + ], + "type": "text", + "content": "To benchmark our agents, we carefully choose a representative evaluation set of game sessions grouped by Part of Speech and compare the performance of humans and agents on this set. In general, the metric values for humans act as ground truth or desirable range, which the agent should achieve. We obtain 800 sessions to be used for testing. The numbers mentioned in Tab. 5 for HH (Human-Human) and HA (Human-Agent) are both computed on this set." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 472, + 547, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 472, + 547, + 640 + ], + "spans": [ + { + "bbox": [ + 304, + 472, + 547, + 640 + ], + "type": "text", + "content": "Collaborative Game Success: Across games, agents had a " + }, + { + "bbox": [ + 304, + 472, + 547, + 640 + ], + "type": "inline_equation", + "content": "68\\%" + }, + { + "bbox": [ + 304, + 472, + 547, + 640 + ], + "type": "text", + "content": " success rate. The win rate for GUESSBOT is slightly lesser than the human counterpart " + }, + { + "bbox": [ + 304, + 472, + 547, + 640 + ], + "type": "inline_equation", + "content": "(85\\%)" + }, + { + "bbox": [ + 304, + 472, + 547, + 640 + ], + "type": "text", + "content": ". However, DRAWBOT's win rate is " + }, + { + "bbox": [ + 304, + 472, + 547, + 640 + ], + "type": "inline_equation", + "content": "51\\%" + }, + { + "bbox": [ + 304, + 472, + 547, + 640 + ], + "type": "text", + "content": ", which is markedly smaller. The Painter agent's bad performance is mainly due to adjectives " + }, + { + "bbox": [ + 304, + 472, + 547, + 640 + ], + "type": "inline_equation", + "content": "(30\\%)" + }, + { + "bbox": [ + 304, + 472, + 547, + 640 + ], + "type": "text", + "content": " and verbs " + }, + { + "bbox": [ + 304, + 472, + 547, + 640 + ], + "type": "inline_equation", + "content": "(56\\%)" + }, + { + "bbox": [ + 304, + 472, + 547, + 640 + ], + "type": "text", + "content": ". It performs reasonably well on most nouns " + }, + { + "bbox": [ + 304, + 472, + 547, + 640 + ], + "type": "inline_equation", + "content": "(74\\%)" + }, + { + "bbox": [ + 304, + 472, + 547, + 640 + ], + "type": "text", + "content": ". The primary reason for bad performance on adjectives is that there is no direct object (e.g. car, bus, icecream) associated with adjectives. Most vision foundation models are trained on objects rather than metaphorical concepts which encode attributes (e.g. hard/soft, length, hollow and loud) which may be associated with multiple objects. Such models need detailed prompts to invoke the relationship and properties of objects." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 642, + 548, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 642, + 548, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 642, + 548, + 714 + ], + "type": "text", + "content": "Qualitative analysis: Many interesting patterns can be observed in the performance and response patterns of human-human and human-agent games – see Fig. 8. Compared to human counterparts, DRAWBOT's sketches have greater symmetry, with clean stroke lines. Unlike typical generative model outputs, which tend to have a 'synthetic' appearance," + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18182" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 144 + ], + "type": "text", + "content": "training on Sketchtopia's data induces a human-like jitter in the generated strokes of DRAWBOT. GUESSBOT's final responses seem human-like for the corresponding canvas contents, even when the game outcome is a failure. This is consistent with the agent's retrieval-based design. Additional examples can be found in Supplementary." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 144, + 288, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 144, + 288, + 215 + ], + "spans": [ + { + "bbox": [ + 46, + 144, + 288, + 215 + ], + "type": "text", + "content": "Feedback Responsiveness: The FRS scores for Human and GUESSBOT are similar (Tab. 5), implying that the agent's ability to adapt its response to feedback is similar to human counterparts. Here, the Gueisser Agent's ActionDecider also plays a key role in 'waking up' GUESSBOT at the right time in response to Painter feedback." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 216, + 288, + 383 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 216, + 288, + 383 + ], + "spans": [ + { + "bbox": [ + 46, + 216, + 288, + 383 + ], + "type": "text", + "content": "Characterizing Interaction Dynamics: Similar to FRS, the AAO scores of agents are comparable to humans, indicating comparable levels of asynchronous communication. However, GUESSBOT has a lower AAO score. This is in line with a Guesser's tendency to wait passively for enough strokes to accumulate on the canvas before guessing. The MATS score measures the agent's action pacing compared to humans. Therefore, it is a proxy measure for ActionDecider's performance due to the latter's role as a key asynchronous enabler of actions in each agent. A relatively lower MATS score for GUESSBOT suggests that its action pacing is closer to human. Empirically, we observed that feedback distribution for DRAWBOT was markedly different from human, contributing to its relatively higher MATS score." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 384, + 288, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 384, + 288, + 514 + ], + "spans": [ + { + "bbox": [ + 46, + 384, + 288, + 514 + ], + "type": "text", + "content": "VLMs as Guessers: Vision-Language Models (VLMs) [29, 30, 52] have become a popular choice for image-conditioned reasoning. In our benchmarking context, VLMs could play a Guesser's role. However, despite detailed prompts instructing the model to output a guess phrase and despite providing prior guess history, we observed that VLM models tend to produce verbose and unhelpful outputs. Examples can be found in Supplementary. Moreover, the slower response times and synchronized (turn-taking) paradigm of current VLMs make them unviable compared to our asynchronous and contextually responsive agents." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 515, + 288, + 646 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 515, + 288, + 646 + ], + "spans": [ + { + "bbox": [ + 46, + 515, + 288, + 646 + ], + "type": "text", + "content": "Limitations: Despite efforts to minimize biases, the dataset may still reflect factors such as age, handedness, and cultural background, with a predominance of younger, right-handed, English-speaking participants from a limited region. This may affect its generalizability to other languages, cultures, and demographics. While our game interface supports asynchronous interactions, further testing across diverse groups is needed. Additionally, while Sketchtopia includes iconic feedback as a novelty, it is limited to predefined icons (e.g., , , and Ⓞ). This restricts the analysis of more nuanced and complex feedback that could occur in natural interactions." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 658, + 119, + 669 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 658, + 119, + 669 + ], + "spans": [ + { + "bbox": [ + 47, + 658, + 119, + 669 + ], + "type": "text", + "content": "7. Conclusion" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 677, + 287, + 713 + ], + "type": "text", + "content": "Sketchtopia enables a scalable framework that pushes the boundaries of multimodal goal-driven communication in asynchronous settings. As the first-ever dataset to include" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 307, + 70, + 545, + 121 + ], + "blocks": [ + { + "bbox": [ + 307, + 70, + 545, + 121 + ], + "lines": [ + { + "bbox": [ + 307, + 70, + 545, + 121 + ], + "spans": [ + { + "bbox": [ + 307, + 70, + 545, + 121 + ], + "type": "table", + "html": "
ModelsGame SuccessFeedback ResponsivenessInteraction Dynamics
Win ↑Soft Win ↑FRS ↑AAOMATS ↓
DRAWBOT51.652.4-0.210.39
GUESSBOT81.786.046%0.170.27
Human85.593.440%0.290.00
", + "image_path": "a36d0d443ed5af0e956dafe4b3e8e8a0fda0d6cdcfc746b4049d4d3cf32e91a7.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 305, + 129, + 547, + 174 + ], + "lines": [ + { + "bbox": [ + 305, + 129, + 547, + 174 + ], + "spans": [ + { + "bbox": [ + 305, + 129, + 547, + 174 + ], + "type": "text", + "content": "Table 5. Benchmarking humans and agents via custom-designed metrics (Sec. 5). FRS: Feedback Responsiveness Score; AAO: Asynchronous Action Overlap; MATS: Multimodal Action Timing Similarity." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 304, + 211, + 547, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 211, + 547, + 319 + ], + "spans": [ + { + "bbox": [ + 304, + 211, + 547, + 319 + ], + "type": "text", + "content": "rich feedback signals, our collection captures realistic gameplay data with natural human interactions, far beyond what existing datasets offer. Sketchtopia is not just about isolated sketches – it incorporates iconic feedback and dynamic interactions between players, all captured using a robust game interface that seamlessly supports both human-human and human-agent gameplay. This versatility allows us to explore interactions at a depth previously unattainable, establishing a new benchmark for multimodal AI systems." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 327, + 547, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 327, + 547, + 447 + ], + "spans": [ + { + "bbox": [ + 304, + 327, + 547, + 447 + ], + "type": "text", + "content": "A standout contribution of our work is the design of first asynchronous communication agents capable of real-time decision-making in multi-modal formats. Our novel agent architectures emphasize responsiveness and speed, enabling agents to mimic the fluidity of natural human communication without the constraints of turn-taking. By introducing innovative metrics such as Asynchronous Action Overlap and Feedback Responsiveness Scores, we redefine how to evaluate collaborative success and interaction dynamics in asynchronous multi-modal communication settings." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 454, + 548, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 454, + 548, + 574 + ], + "spans": [ + { + "bbox": [ + 304, + 454, + 548, + 574 + ], + "type": "text", + "content": "Benchmarking our agents against human players demonstrates that they can generate human-like actions and responses. Our results reveal that even state-of-the-art VLMs lag significantly in handling asynchronous, multimodal communication complexities. This establishes Sketchtopia as a critical benchmark for advancing AI research in this space. Additionally, our game flow visualizations offer a novel lens into understanding how agents and humans progress towards shared goals, providing insights into the evolution of strategies and collaborative problem-solving." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 582, + 548, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 582, + 548, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 582, + 548, + 713 + ], + "type": "text", + "content": "Sketchtopia's applications extend beyond Pictionary. Its interactive features enable sketch scene editing and collaborative whiteboard tools, with potential for live note-taking and narration. The dataset's focus on abstract sketches also supports cognitive science research into communication strategies. In summary, Sketchtopia provides a valuable resource for exploring asynchronous, multimodal AI systems and realistic agent-based gameplay. By tackling the complexities of real-time, feedback-driven communication, we aim to inspire AI advancements that are more natural, adaptive, and truly collaborative." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18183" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 53, + 91, + 288, + 112 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 288, + 112 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 288, + 112 + ], + "type": "text", + "content": "[1] Michael Argyle. The syntaxes of bodily communication. Linguistics, 11(112):71-92, 1973. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 115, + 288, + 170 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 115, + 288, + 170 + ], + "spans": [ + { + "bbox": [ + 53, + 115, + 288, + 170 + ], + "type": "text", + "content": "[2] Zahra Ashktorab, Q. Vera Liao, Casey Dugan, James Johnson, Qian Pan, Wei Zhang, Sadhana Kumaravel, and Murray Campbell. Human-ai collaboration in a cooperative game setting: Measuring social perception and outcomes. Proc. ACM Hum.-Comput. Interact., 4(CSCW2), 2020. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 171, + 288, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 171, + 288, + 247 + ], + "spans": [ + { + "bbox": [ + 53, + 171, + 288, + 247 + ], + "type": "text", + "content": "[3] Zahra Ashktorab, Casey Dugan, James Johnson, Qian Pan, Wei Zhang, Sadhana Kumaravel, and Murray Campbell. Effects of communication directionality and ai agent differences in human-ai interaction. In Proceedings of the 2021 CHI Conference on Human Factors in Computing Systems, New York, NY, USA, 2021. Association for Computing Machinery. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 250, + 288, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 250, + 288, + 282 + ], + "spans": [ + { + "bbox": [ + 53, + 250, + 288, + 282 + ], + "type": "text", + "content": "[4] Muhammad Fauzi Bafadal and Humaira Humaira. The use of charades games in teaching vocabulary to the junior high school students. Linguistics and Elt Journal, 2019. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 285, + 288, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 285, + 288, + 351 + ], + "spans": [ + { + "bbox": [ + 53, + 285, + 288, + 351 + ], + "type": "text", + "content": "[5] Nolan Bard, Jakob N. Foerster, Sarath Chandar, Neil Burch, Marc Lanctot, H. Francis Song, Emilio Parisotto, Vincent Dumoulin, Subhodeep Moitra, Edward Hughes, Iain Dunning, Shibl Mourad, Hugo Larochelle, Marc G. Bellemare, and Michael Bowling. The hanabi challenge: A new frontier for ai research. Artificial Intelligence, 280:103216, 2020. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 353, + 287, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 353, + 287, + 396 + ], + "spans": [ + { + "bbox": [ + 53, + 353, + 287, + 396 + ], + "type": "text", + "content": "[6] Fabien Beltrame. Comportement non verbal et situation d'interaction: Quand la gestuelle des mains revendique un role cognitif de support des représentations pour la résolution de problème, 2006. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 399, + 288, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 399, + 288, + 453 + ], + "spans": [ + { + "bbox": [ + 53, + 399, + 288, + 453 + ], + "type": "text", + "content": "[7] Ayan Kumar Bhunia, Ayan Das, Umar Riaz Muhammad, Yongxin Yang, Timothy M. Hospedales, Tao Xiang, Yulia Gryaditskaya, and Yi-Zhe Song. Pixel: a competitive sketching ai agent. so you think you can sketch? ACM Trans. Graph., 39(6), 2020. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 455, + 288, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 455, + 288, + 575 + ], + "spans": [ + { + "bbox": [ + 53, + 455, + 288, + 575 + ], + "type": "text", + "content": "[8] Christopher Clark, Jordi Salvador, Dustin Schwenk, Derrick Bonafilia, Mark Yatskar, Eric Kolve, Alvaro Herrasti, Jonghyun Choi, Sachin Mehta, Sam Skjonsberg, Carissa Schoenick, Aaron Sarnat, Hannaneh Hajishirzi, Aniruddha Kembhavi, Oren Etzioni, and Ali Farhadi. Iconary: A pictorial-based game for testing multimodal communication with drawings and text. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 1864-1886, Online and Punta Cana, Dominican Republic, 2021. Association for Computational Linguistics. 1, 2, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 578, + 288, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 578, + 288, + 632 + ], + "spans": [ + { + "bbox": [ + 53, + 578, + 288, + 632 + ], + "type": "text", + "content": "[9] Abhishek Das, Satwik Kottur, Khushi Gupta, Avi Singh, Deshraj Yadav, Jose M. F. Moura, Devi Parikh, and Dhruv Batra. Visual dialog. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 635, + 288, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 288, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 288, + 689 + ], + "type": "text", + "content": "[10] A. Das, S. Kottur, J. F. Moura, S. Lee, and D. Batra. Learning cooperative visual dialog agents with deep reinforcement learning. In 2017 IEEE International Conference on Computer Vision (ICCV), pages 2970-2979, Los Alamitos, CA, USA, 2017. IEEE Computer Society. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 691, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 288, + 713 + ], + "type": "text", + "content": "[11] Abhishek Das, Samyak Datta, Georgia Gkioxari, Stefan Lee, Devi Parikh, and Dhruv Batra. Embodied question answering." + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 308, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 308, + 95, + 547, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 95, + 547, + 139 + ], + "spans": [ + { + "bbox": [ + 308, + 95, + 547, + 139 + ], + "type": "text", + "content": "[12] Markus Eger, Chris Martens, and Marcela Alfaro Cordoba. An intentional ai for hanabi. In 2017 IEEE Conference on Computational Intelligence and Games (CIG), page 68-75. IEEE Press, 2017. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 308, + 140, + 545, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 140, + 545, + 171 + ], + "spans": [ + { + "bbox": [ + 308, + 140, + 545, + 171 + ], + "type": "text", + "content": "[13] Mathias Eitz, James Hays, and Marc Alexa. How do humans sketch objects? ACM Trans. Graph. (Proc. SIGGRAPH), 31 (4):44:1-44:10, 2012. 1, 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 173, + 547, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 173, + 547, + 237 + ], + "spans": [ + { + "bbox": [ + 308, + 173, + 547, + 237 + ], + "type": "text", + "content": "[14] Alaaeldin El-Nouby, Shikhar Sharma, Hannes Schulz, Devon Hjelm, Layla El Asri, Samira Ebrahimi Kahou, Yoshua Bengio, and Graham W. Taylor. Tell, draw, and repeat: Generating and modifying images based on continual linguistic instruction. In The IEEE International Conference on Computer Vision (ICCV), 2019. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 239, + 547, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 239, + 547, + 316 + ], + "spans": [ + { + "bbox": [ + 308, + 239, + 547, + 316 + ], + "type": "text", + "content": "[15] Jiazhan Feng, Qingfeng Sun, Can Xu, Pu Zhao, Yaming Yang, Chongyang Tao, Dongyan Zhao, and Qingwei Lin. MMDialog: A large-scale multi-turn dialogue dataset towards multimodal open-domain conversation. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 7348-7363, Toronto, Canada, 2023. Association for Computational Linguistics. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 316, + 545, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 316, + 545, + 369 + ], + "spans": [ + { + "bbox": [ + 308, + 316, + 545, + 369 + ], + "type": "text", + "content": "[16] Chengying Gao, Qi Liu, Qi Xu, Limin Wang, Jianzhuang Liu, and Changqing Zou. Sketchycoco: Image generation from freehand scene sketches. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5174-5183, 2020. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 371, + 545, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 371, + 545, + 392 + ], + "spans": [ + { + "bbox": [ + 308, + 371, + 545, + 392 + ], + "type": "text", + "content": "[17] David Ha and Douglas Eck. A neural representation of sketch drawings. In ICLR 2018, 2018. 2018. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 394, + 547, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 394, + 547, + 446 + ], + "spans": [ + { + "bbox": [ + 308, + 394, + 547, + 446 + ], + "type": "text", + "content": "[18] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In International Conference on Learning Representations, 2022. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 449, + 547, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 449, + 547, + 503 + ], + "spans": [ + { + "bbox": [ + 308, + 449, + 547, + 503 + ], + "type": "text", + "content": "[19] Forrest Huang, Eldon Schoop, David Ha, and John Canny. Scones: towards conversational authoring of sketches. In Proceedings of the 25th International Conference on Intelligent User Interfaces, page 313–323, New York, NY, USA, 2020. Association for Computing Machinery. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 504, + 547, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 504, + 547, + 535 + ], + "spans": [ + { + "bbox": [ + 308, + 504, + 547, + 535 + ], + "type": "text", + "content": "[20] Kristiina Jokinen. Gaze and gesture activity in communication. pages 537-546. Springer, Berlin, Heidelberg, 2009. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 537, + 547, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 537, + 547, + 568 + ], + "spans": [ + { + "bbox": [ + 308, + 537, + 547, + 568 + ], + "type": "text", + "content": "[21] Kristiina Jokinen. Non-verbal signals for turn-taking and feedback. European Language Resources Association (ELRA), 2010. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 570, + 545, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 570, + 545, + 592 + ], + "spans": [ + { + "bbox": [ + 308, + 570, + 545, + 592 + ], + "type": "text", + "content": "[22] Kurmanbek Kaiyrbekov and Metin Sezgin. Stroke-based sketched symbol reconstruction and segmentation, 2019. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 593, + 547, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 593, + 547, + 669 + ], + "spans": [ + { + "bbox": [ + 308, + 593, + 547, + 669 + ], + "type": "text", + "content": "[23] Jin-Hwa Kim, Nikita Kitaev, Xinlei Chen, Marcus Rohrbach, Byoung-Tak Zhang, Yuandong Tian, Dhruv Batra, and Devi Parikh. CoDraw: Collaborative drawing as a testbed for grounded goal-driven communication. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 6495-6513, Florence, Italy, 2019. Association for Computational Linguistics. 1, 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 670, + 547, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 670, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 670, + 547, + 713 + ], + "type": "text", + "content": "[24] Yihuai Lan, Zhiqiang Hu, Lei Wang, Yang Wang, Deheng Ye, Peilin Zhao, Ee-Peng Li m, Hui Xiong, and Hao Wang. Llm-based agent society investigation: Collaboration and confrontation in avalon gameplay, 2024. 2" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18184" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 288, + 127 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 288, + 127 + ], + "type": "text", + "content": "[25] Zixing Lei, Yiming Zhang, Yuxin Xiong, and Siheng Chen. Emergent communication in interactive sketch question answering. In Proceedings of the 37th International Conference on Neural Information Processing Systems, Red Hook, NY, USA, 2023. Curran Associates Inc. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 129, + 288, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 129, + 288, + 152 + ], + "spans": [ + { + "bbox": [ + 48, + 129, + 288, + 152 + ], + "type": "text", + "content": "[26] Lei Li, Hongbo Fu, and Chiew-Lan Tai. Fast sketch segmentation and labeling with deep learning, 2018. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 152, + 288, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 152, + 288, + 196 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 288, + 196 + ], + "type": "text", + "content": "[27] Tian Liang, Zhiwei He, Jen tse Huang, Wenxuan Wang, Wenxiang Jiao, Rui Wang, Yujiu Yang, Zhaopeng Tu, Shuming Shi, and Xing Wang. Leveraging word guessing games to assess the intelligence of large language models, 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 198, + 288, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 198, + 288, + 229 + ], + "spans": [ + { + "bbox": [ + 48, + 198, + 288, + 229 + ], + "type": "text", + "content": "[28] Jian Lin. Divergence measures based on the shannon entropy. IEEE Transactions on Information Theory, 37(1):145-151, 1991. 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 231, + 288, + 253 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 231, + 288, + 253 + ], + "spans": [ + { + "bbox": [ + 48, + 231, + 288, + 253 + ], + "type": "text", + "content": "[29] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning, 2023. 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 255, + 288, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 255, + 288, + 297 + ], + "spans": [ + { + "bbox": [ + 48, + 255, + 288, + 297 + ], + "type": "text", + "content": "[30] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In Advances in Neural Information Processing Systems, pages 34892-34916. Curran Associates, Inc., 2023. 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 299, + 288, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 299, + 288, + 354 + ], + "spans": [ + { + "bbox": [ + 48, + 299, + 288, + 354 + ], + "type": "text", + "content": "[31] Shuo Liu, Kaining Ying, Hao Zhang, Yue Yang, Yuqi Lin, Tianle Zhang, Chuanhao Li, Yu Qiao, Ping Luo, Wenqi Shao, and Kaipeng Zhang. Convbench: A multi-turn conversation evaluation benchmark with hierarchical capability for large vision-language models, 2024. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 355, + 288, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 355, + 288, + 398 + ], + "spans": [ + { + "bbox": [ + 48, + 355, + 288, + 398 + ], + "type": "text", + "content": "[32] Haoran Mo, Edgar Simo-Serra, Chengying Gao, Changqing Zou, and Ruomei Wang. General virtual sketching framework for vector line art. ACM Transactions on Graphics (TOG), 40:1-14, 2021. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 399, + 288, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 399, + 288, + 443 + ], + "spans": [ + { + "bbox": [ + 48, + 399, + 288, + 443 + ], + "type": "text", + "content": "[33] Chong Mou, Xintao Wang, Liangbin Xie, Jing Zhang, Zhonggang Qi, Ying Shan, and Xiaohu Qie. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. ArXiv, abs/2302.08453, 2023. 5, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 445, + 288, + 467 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 445, + 288, + 467 + ], + "spans": [ + { + "bbox": [ + 48, + 445, + 288, + 467 + ], + "type": "text", + "content": "[34] Xu Pan and Odelia Schwartz. Multimodal ai needs active human interaction. Nature Human Behaviour, 2024. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 468, + 288, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 468, + 288, + 489 + ], + "spans": [ + { + "bbox": [ + 48, + 468, + 288, + 489 + ], + "type": "text", + "content": "[35] Deepika Phutela. The importance of non-verbal communication, 2016. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 491, + 288, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 491, + 288, + 522 + ], + "spans": [ + { + "bbox": [ + 48, + 491, + 288, + 522 + ], + "type": "text", + "content": "[36] Yonggang Qi and Zheng-Hua Tan. Sketchsegnet+: An end-to-end learning of rnn for multi-class sketch semantic segmentation. IEEE Access, 7:102717-102726, 2019. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 525, + 288, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 525, + 288, + 579 + ], + "spans": [ + { + "bbox": [ + 48, + 525, + 288, + 579 + ], + "type": "text", + "content": "[37] Shuwen Qiu, Sirui Xie, Lifeng Fan, Tao Gao, Jungseock Joo, Song-Chun Zhu, and Yixin Zhu. Emergent graphical conventions in a visual communication game. In Advances in Neural Information Processing Systems, pages 13119-13131. Curran Associates, Inc., 2022. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 581, + 288, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 581, + 288, + 668 + ], + "spans": [ + { + "bbox": [ + 48, + 581, + 288, + 668 + ], + "type": "text", + "content": "[38] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In Proceedings of the 38th International Conference on Machine Learning, ICML 2021, 18-24 July 2021, Virtual Event, pages 8748-8763. PMLR, 2021. 4, 5, 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 670, + 288, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 670, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 670, + 288, + 712 + ], + "type": "text", + "content": "[39] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. arXiv preprint arXiv:2112.10752, 2022. 4, 6" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 545, + 106 + ], + "type": "text", + "content": "[40] Patsorn Sangkloy, Nathan Burnell, Cusuh Ham, and James Hays. The sketchy database: Learning to retrieve badly drawn bunnies. ACM Trans. Graph., 35(4), 2016. 1, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 107, + 547, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 547, + 162 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 547, + 162 + ], + "type": "text", + "content": "[41] Ravi Kiran Sarvadevabhatla, Shiv Surya, Trisha Mittal, and R. Venkatesh Babu. Pictionary-style word guessing on hand-drawn object sketches: Dataset, analysis and deep network models. IEEE Transactions on Pattern Analysis and Machine Intelligence, 42(1):221-231, 2020. 1, 2, 3, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 163, + 545, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 163, + 545, + 196 + ], + "spans": [ + { + "bbox": [ + 307, + 163, + 545, + 196 + ], + "type": "text", + "content": "[42] SeungWon Seo, Junhyeok Lee, SeongRae Noh, and HyeongYeop Kang. Llm-based cooperative agents using information relevance and plan validation, 2024. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 198, + 547, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 198, + 547, + 229 + ], + "spans": [ + { + "bbox": [ + 307, + 198, + 547, + 229 + ], + "type": "text", + "content": "[43] Chen Shen. Nonverbal communication in intercultural communication: A case study of telling lies. Frontiers in Educational Research, 6(13), 2023. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 231, + 545, + 253 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 231, + 545, + 253 + ], + "spans": [ + { + "bbox": [ + 307, + 231, + 545, + 253 + ], + "type": "text", + "content": "[44] Spotify. Annoy: Approximate nearest neighbors in c++/python, n.d. Computer software. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 307, + 255, + 547, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 255, + 547, + 331 + ], + "spans": [ + { + "bbox": [ + 307, + 255, + 547, + 331 + ], + "type": "text", + "content": "[45] Alessandro Suglia, Ioannis Konstas, Andrea Vanzo, Emanuele Bastianelli, Desmond Elliott, Stella Frank, and Oliver Lemon. CompGuessWhat?!: A multi-task evaluation framework for grounded language learning. In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 7625-7641, Online, 2020. Association for Computational Linguistics. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 307, + 333, + 547, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 333, + 547, + 409 + ], + "spans": [ + { + "bbox": [ + 307, + 333, + 547, + 409 + ], + "type": "text", + "content": "[46] Alessandro Suglia, Antonio Vergari, Ioannis Konstas, Yonatan Bisk, Emanuele Bastianelli, Andrea Vanzo, and Oliver Lemon. Imagining grounded conceptual representations from perceptual information in situated guessing games. In Proceedings of the 28th International Conference on Computational Linguistics, pages 1090–1102, Barcelona, Spain (Online), 2020. International Committee on Computational Linguistics. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 307, + 411, + 545, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 411, + 545, + 454 + ], + "spans": [ + { + "bbox": [ + 307, + 411, + 545, + 454 + ], + "type": "text", + "content": "[47] F. Wang, S. Lin, H. Wu, H. Li, R. Wang, X. Luo, and X. He. Spfusionnet: Sketch segmentation using multi-modal data fusion. In 2019 IEEE International Conference on Multimedia and Expo (ICME), pages 1654–1659, 2019. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 307, + 456, + 545, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 456, + 545, + 498 + ], + "spans": [ + { + "bbox": [ + 307, + 456, + 545, + 498 + ], + "type": "text", + "content": "[48] Haohan Wang, Songwei Ge, Zachary Lipton, and Eric P Xing. Learning robust global representations by penalizing local predictive power. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2019. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 307, + 500, + 545, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 500, + 545, + 544 + ], + "spans": [ + { + "bbox": [ + 307, + 500, + 545, + 544 + ], + "type": "text", + "content": "[49] Wenhai Wang, Enze Xie, Xiaoge Song, Yuhang Zang, Wenjia Wang, Tong Lu, Gang Yu, and Chunhua Shen. Efficient and accurate arbitrary-shaped text detection with pixel aggregation network, 2020. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 307, + 545, + 545, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 545, + 545, + 589 + ], + "spans": [ + { + "bbox": [ + 307, + 545, + 545, + 589 + ], + "type": "text", + "content": "[50] Yuzhuang Xu, Shuo Wang, Peng Li, Fuwen Luo, Xiaolong Wang, Weidong Liu, and Yang Liu. Exploring large language models for communication games: An empirical study on werewolf. arXiv preprint arXiv:2309.04658, 2023. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 307, + 590, + 545, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 590, + 545, + 623 + ], + "spans": [ + { + "bbox": [ + 307, + 590, + 545, + 623 + ], + "type": "text", + "content": "[51] Lumin Yang, Jiajie Zhuang, Hongbo Fu, Kun Zhou, and Youyi Zheng. Sketchgen: Semantic sketch segmentation with graph convolutional networks, 2020. 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 307, + 624, + 545, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 624, + 545, + 690 + ], + "spans": [ + { + "bbox": [ + 307, + 624, + 545, + 690 + ], + "type": "text", + "content": "[52] Yuan Yao, Tianyu Yu, Ao Zhang, Chongyi Wang, Junbo Cui, Hongji Zhu, Tianchi Cai, Haoyu Li, Weilin Zhao, Zhihui He, Qianyu Chen, Huarong Zhou, Zhensheng Zou, Haoye Zhang, Shengding Hu, Zhi Zheng, Jie Zhou, Jie Cai, Xu Han, Guoyang Zeng, Dahai Li, Zhiyuan Liu, and Maosong Sun. Minicpm-v: A gpt-4v level mllm on your phone, 2024. 8" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 307, + 692, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 692, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 307, + 692, + 545, + 713 + ], + "type": "text", + "content": "[53] Qian Yu, Feng Liu, Yi-Zhe Song, Tao Xiang, Timothy M. Hospedales, and Chen-Change Loy. Sketch me that shoe. In" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18185" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 274 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 67, + 72, + 287, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 72, + 287, + 95 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 287, + 95 + ], + "type": "text", + "content": "Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 1, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 96, + 288, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 96, + 288, + 140 + ], + "spans": [ + { + "bbox": [ + 48, + 96, + 288, + 140 + ], + "type": "text", + "content": "[54] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 11941-11952, 2023. 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 141, + 288, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 141, + 288, + 185 + ], + "spans": [ + { + "bbox": [ + 48, + 141, + 288, + 185 + ], + "type": "text", + "content": "[55] Hua Zhang, Si Liu, Changqing Zhang, Wenqi Ren, Rui Wang, and Xiaochun Cao. Sketchnet: Sketch classification with web images. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 1105-1113, 2016. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 186, + 288, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 186, + 288, + 219 + ], + "spans": [ + { + "bbox": [ + 48, + 186, + 288, + 219 + ], + "type": "text", + "content": "[56] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. arXiv preprint arXiv:2302.05543, 2023. 4, 5, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 220, + 288, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 220, + 288, + 274 + ], + "spans": [ + { + "bbox": [ + 48, + 220, + 288, + 274 + ], + "type": "text", + "content": "[57] Changqing Zou, Qian Yu, Ruofei Du, Haoran Mo, Yi-Zhe Song, Tao Xiang, Chengying Gao, Baoquan Chen, and Hao Zhang. Sketchyscene: Richly-annotated scene sketches. In ECCV, pages 438–454. Springer International Publishing, 2018. 3" + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "18186" + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Sketchy Bounding-box Supervision for 3D Instance Segmentation/fed9c100-3d53-4c29-809d-ba6f2288846b_content_list.json b/2025/Sketchy Bounding-box Supervision for 3D Instance Segmentation/fed9c100-3d53-4c29-809d-ba6f2288846b_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..deac5a5582efb287c5208e5b26393d49fe6777a8 --- /dev/null +++ b/2025/Sketchy Bounding-box Supervision for 3D Instance Segmentation/fed9c100-3d53-4c29-809d-ba6f2288846b_content_list.json @@ -0,0 +1,1767 @@ +[ + { + "type": "text", + "text": "Sketchy Bounding-box Supervision for 3D Instance Segmentation", + "text_level": 1, + "bbox": [ + 166, + 130, + 828, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Qian Deng $^{1}$ , Le Hui $^{2}$ , Jin Xie $^{3,4*}$ , Jian Yang $^{1}$", + "bbox": [ + 318, + 179, + 674, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1PCA Lab, College of Computer Science, Nankai University, Tianjin, China", + "bbox": [ + 197, + 198, + 797, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ School of Electronics and Information, Northwestern Polytechnical University, Xi'an, China", + "bbox": [ + 130, + 215, + 866, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{3}$ State Key Laboratory for Novel Software Technology, Nanjing University, Nanjing, China", + "bbox": [ + 140, + 233, + 856, + 251 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{4}$ School of Intelligence Science and Technology, Nanjing University, Suzhou, China", + "bbox": [ + 166, + 251, + 828, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "dengqian@mail.nankai.edu.cn; huile@nwpu.edu.cn; csjxie@nju.edu.cn; csjyang@nankai.edu.cn", + "bbox": [ + 102, + 272, + 880, + 286 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 321, + 326, + 338 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Bounding box supervision has gained considerable attention in weakly supervised 3D instance segmentation. While this approach alleviates the need for extensive point-level annotations, obtaining accurate bounding boxes in practical applications remains challenging. To this end, we explore the inaccurate bounding box, named sketchy bounding box, which is imitated through perturbing ground truth bounding box by adding scaling, translation, and rotation. In this paper, we propose Sketchy-3DIS, a novel weakly 3D instance segmentation framework, which jointly learns pseudo labeler and segmentator to improve the performance under the sketchy bounding-box supervisions. Specifically, we first propose an adaptive box-to-point pseudo labeler that adaptively learns to assign points located in the overlapped parts between two sketchy bounding boxes to the correct instance, resulting in compact and pure pseudo instance labels. Then, we present a coarse-to-fine instance segmentator that first predicts coarse instances from the entire point cloud and then learns fine instances based on the region of coarse instances. Finally, by using the pseudo instance labels to supervise the instance segmentator, we can gradually generate high-quality instances through joint training. Extensive experiments show that our method achieves state-of-the-art performance on both the ScanNetV2 and S3DIS benchmarks, and even outperforms several fully supervised methods using sketchy bounding boxes. Code is available at https://github.com/dengq7/Sketchy-3DIS.", + "bbox": [ + 88, + 353, + 485, + 763 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 790, + 220, + 806 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "With the rise of 3D datasets and the growing need for 3D scene understanding, various 3D computer vision tasks [19, 25, 28-31, 33, 34, 36, 42-44, 50, 51] have gained significant attention from both academia and industry. Among these, 3D", + "bbox": [ + 89, + 815, + 483, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/218af45a23c27b30bd28ac36739086053e911ca15466e8c63b7fbf0b852823f6.jpg", + "image_caption": [ + "(a) Inputs", + "Figure 1. (a) illustrates the inputs of bounding box supervised 3D instance segmentation. (b) compares the performances of GaPro [29] and our Sketchy-3DIS under both accurate and scaled sketchy boxes supervision on ScanNetV2 validation set." + ], + "image_footnote": [], + "bbox": [ + 522, + 319, + 612, + 474 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7f25aa8c2fdf6a9f9987cffb92ef4d9a91ff368a1460d4d3ffa3be0ebc415f25.jpg", + "image_caption": [ + "(b) Performance Comparisons" + ], + "image_footnote": [], + "bbox": [ + 612, + 318, + 893, + 472 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "instance segmentation is both fundamental and challenging, as it involves recognizing the category of each object within a point cloud and delineating its individual mask. Current 3D instance segmentation methods [2, 13, 14, 30, 34, 36, 41] have demonstrated impressive performance, primarily rely on data-hungry deep learning approaches that require substantial training data and dense point-level manual annotations. Notably, point-level fully supervised instance annotation in ScanNet takes approximately 22.3 minutes [6] per scene, while labeling a 3D box in ScanNet takes about 1.93 minutes [38] per scene. Therefore, bounding box annotation stands out as it provides substantial instance-level information while maintaining a reasonable annotation cost, making it a favorable annotation choice for weakly supervised 3D instance segmentation.", + "bbox": [ + 511, + 551, + 908, + 777 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Many weakly supervised methods have emerged to tackle the 3D instance segmentation task. Semi-supervised 3D instance segmentation methods [5, 12, 23, 45] use training data with partial labels, while others [5, 12, 23, 45] utilize indirect labels, such as 2D image annotations. Additionally, object-level annotations, like marking a single point [7, 24, 37, 38] or drawing a bounding box [4, 8, 29, 49] for each object, are employed. Bounding box supervision is another important", + "bbox": [ + 511, + 779, + 910, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author", + "bbox": [ + 109, + 887, + 233, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "8879", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "annotation for weakly supervised 3D instance segmentation. Box2Mask [4] is the first attempt that uses bounding box supervision for 3D instance segmentation. Subsequently, the following methods [8, 27, 29, 49] use different constraints to improve performance, such as using local correlations, Gaussian processes, teacher-student model, SAM [16], respectively. Although impressive results have been achieved in weakly supervised 3D instance segmentation, these methods still rely on accurate and compact bounding box annotations. It is worth noting that in practice, accurate and compact boxes are difficult to obtain. Generally, annotated bounding box involves changes in scaling, translation, and rotation. As shown in Fig.1, we observed that the performance of GaPro [29] decreases significantly under scaled bounding box supervision. This motivated us to design a new approach to tackle with these inaccurate bounding box annotations.", + "bbox": [ + 89, + 90, + 485, + 332 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we propose a weakly supervised 3D instance segmentation framework to tackle sketchy bounding box supervision. To simulate sketchy bounding boxes, we impose scaling, translation, and rotation on the ground truth bounding boxes to generate inaccurate bounding boxes. Our method consists of two key components: adaptive box-to-point pseudo labeler and coarse-to-fine instance segmentator. Given a point cloud scene and sketchy bounding boxes, we first use 3D U-Net [9] backbone to extract point features. After that, we feed the learned features and sketchy bounding boxes of the scene into the adaptive box-to-point pseudo labeler to generate high-quality point-level instance labels. For the points located in the overlapped region between two sketchy bounding boxes, we adaptively assign the points to the corresponding box by learning the similarity of points and boxes. For the remaining points, we directly assign them to the corresponding box according to the spatial correlations. In this way, we can transform the sketchy bounding box into a compact bounding box. We consider the labels of points that located within the generated compact bounding box as pseudo instance labels. To accurately predict instances, we input the learned features into the coarse-to-fine instance segmentator iteratively. We first detect the coarse bounding box and mask of the instance by executing queries on the whole point cloud. Based on the small region of the coarse bounding box, we re-predict the fine bounding box and mask by conducting queries within a more precise small region. Finally, by jointly optimizing the adaptive box-to-point pseudo labeler and coarse-to-fine instance segmentator, our method can gradually improve the performance of weakly supervised 3D instance segmentation using the sketchy bounding box annotations. Rich experiments on the ScanNetV2 and S3DIS show that our method have achieved state-of-the-art performance.", + "bbox": [ + 91, + 335, + 483, + 848 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions are as follows:", + "bbox": [ + 109, + 852, + 330, + 866 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- We are the first to explore the use of sketchy bounding box annotations for 3D instance segmentation, which allows", + "bbox": [ + 89, + 869, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "for training with noise-tolerant, instance-size-agnostic 3D bounding boxes.", + "bbox": [ + 526, + 90, + 906, + 121 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a new weakly supervised 3D instance segmentation framework, which presents an adaptive box-to-point pseudo labeler and a coarse-to-fine instance segmentator for joint training.", + "- Our method achieves leading performance on both the ScanNetV2 and S3DIS datasets for weakly 3D instance segmentation under sketchy bounding box annotations." + ], + "bbox": [ + 511, + 125, + 906, + 232 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 247, + 653, + 262 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Fully Supervised 3D Instance Segmentation", + "text_level": 1, + "bbox": [ + 511, + 272, + 883, + 287 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Existing fully supervised methods for 3D instance segmentation can be broadly categorized into three main pipelines: grouping-based [2, 13, 14, 22], detection-based [11, 41, 46, 48], and query-based [10, 17, 19, 26, 30, 34-36, 40] methods. Grouping-based methods typically group points based on predicted per-point semantic classes and instance center offsets, but the grouping rules are often manually defined by researchers, which lack flexibility. Detection-based methods, on the other hand, first predict the bounding box for each object and then generate the foreground mask within the box, treating the task as a two-step process of object detection followed by segmentation. More recently, inspired by DETR [52] and Mask2Former [3], query-based methods have been proposed for 3D instance segmentation. These methods simultaneously predict the semantic classes and instance masks of objects using a set of learnable query vectors, which model the objects' geometrical and semantical attributes integrally. Among these methods, query-based approaches tend to achieve superior performance, as the other two pipelines both face challenges such as error accumulation and time-consuming refinement processes. While these fully-supervised methods have made significant progress, they all rely on dense per-point annotations, and their performance tends to degrade when such annotations are unavailable. Therefore, several weakly-supervised methods have been proposed to alleviate this problem.", + "bbox": [ + 511, + 295, + 908, + 686 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Weakly Supervised 3D Instance Segmentation", + "text_level": 1, + "bbox": [ + 511, + 696, + 901, + 713 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Weakly-supervised methods require partial or implicit annotations. [5, 12, 23, 45] use sparse point annotations and design various label propagation strategies. However, these sparse points cannot fully capture the spatial extent of an object. Approaches like MIT [47] and CIP-WPIS [49] leverage 2D images as additional input for 3D instance segmentation, but this introduces multi-modal dependencies. Other methods [7, 24, 37, 38] rely on instance-level signals, where labeled a single point for per instance, and instance predictions are made using techniques like random walk or clustering. Recently, most state-of-the-art methods [4, 8, 27, 29, 49] have adopted 3D bounding boxes as supervision, where a", + "bbox": [ + 511, + 719, + 908, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "8880", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "single bounding box is assigned to each object. These approaches naturally capture the spatial extent of an object and the annotated bounding boxes are easier to obtain compared to dense per-point annotations.", + "bbox": [ + 89, + 90, + 483, + 151 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Bounding-box Supervised 3D Instance Segmentation. Bounding boxes have become a widely used supervision signal, not only for 2D instance segmentation tasks [15, 20, 39] but also for 3D instance segmentation. Box2mask [4] is a pioneering approach that utilizes bounding-box supervision for 3D instance segmentation. WISGP [8] divides the point cloud into a determined set and an undetermined set based on explicit spatial correlations, then uses polygon meshes and superpoints to label the undetermined set. GaPro [29] focuses on generating reliable pseudo labels using a Gaussian process. Recently, BSNet [27] employs a simulation-assisted mean teacher framework to generate pseudo labels, enhancing data samples by synthesizing scenes that contain overlapping objects. However, these methods all rely on axis-aligned bounding boxes for supervision, and their performances heavily depend on the accuracy of these annotated boxes. To address this, CIP-WPIS [49] proposes using slightly looser bounding boxes, projecting the point cloud into 2D images, and obtaining 3D instance masks based on segmented 2D masks. While effective, it relies on 2D modality and can only adapt to slightly enlarged bounding boxes, limiting its flexibility. To alleviate the annotation costs as well as obtain robust and effective performance, we design a novel framework Sketchy-3DIS, which endeavors to not only face the common challenge of correctly assigning the points within multiple boxes but also be dedicated to correcting the errors brought by sketchy bounding boxes.", + "bbox": [ + 89, + 152, + 486, + 561 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 577, + 181, + 592 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Overview", + "text_level": 1, + "bbox": [ + 89, + 604, + 200, + 618 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The overall framework of our weakly supervised 3D instance segmentation based on sketchy bounding boxes is depicted in Fig.3. It consists of two key components: adaptive box-to-point pseudo labeler and coarse-to-fine instance segmentator.", + "bbox": [ + 89, + 627, + 485, + 686 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Assuming a 3D point cloud $P \\in \\mathbb{R}^{N \\times C}$ , where $N$ and $C$ represent the number of 3D points and channels, respectively. In the basic setting, each point is represented by the 3D coordinate, RGB color, and surface normal vector. In the training phase, we first feed the input point cloud and the sketchy bounding boxes (defined in Sec. 3.2) into the 3D UNet backbone network for feature extraction, then we average the point features located in the superpoint [21] to generate superpoint-level features. Next, we input the features and sketchy bounding boxes to the adaptive box-to-point pseudo labeler (Sec. 3.3) for converting the rough box-level annotation to the detailed point-level annotation of the instance. Meanwhile, we feed the feature into the coarse-to-fine instance segmentator (Sec. 3.4) to predict instances by", + "bbox": [ + 89, + 688, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/e10580a19fbb3a04ba86032c1afdd3bd3777e80298757556110bd2e6953fd0cb.jpg", + "image_caption": [ + "Point Cloud" + ], + "image_footnote": [], + "bbox": [ + 524, + 90, + 658, + 167 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/5257c541f2a9726e29353ab9494e750389c4a39b4fbb27e31d5d89ad9b69a7da.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 697, + 119, + 735, + 147 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/a6100bc158c2fe0f3e4d6702eb96aca7989542341f4a1e5ae09a29cd08c5faa8.jpg", + "image_caption": [ + "Scaled $(S_{1})$" + ], + "image_footnote": [], + "bbox": [ + 790, + 112, + 854, + 154 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/4d11b69675b686840c5a1c14f900dcfd7b4be15ac53e8c108c05e108bf51d0d9.jpg", + "image_caption": [ + "Scaled and translated $(S_{2})$" + ], + "image_footnote": [], + "bbox": [ + 576, + 196, + 650, + 239 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/8c2da4509d4486ba0e3ca9f2060700859132a946787fd3678bb8496870e1b7ea.jpg", + "image_caption": [ + "Accurate $(S_0)$", + "Scaled and rotated $(S_{3})$" + ], + "image_footnote": [], + "bbox": [ + 676, + 190, + 754, + 247 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/6fa0b9391a2ca90bf0043c9a97ea3354a66b5a3a2d3a8137e4c4a74d0f513409.jpg", + "image_caption": [ + "Scaled, translated and rotated $(S_4)$", + "Figure 2. Various sketchy bounding boxes under scaling, translation, and rotation perturbations." + ], + "image_footnote": [], + "bbox": [ + 803, + 191, + 882, + 247 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Transformer. Finally, we employ bilateral matching between the generated pseudo labels (regarded as the ground truth) and the predicted instances to build the correspondence between them for training (Sec. 3.5). In the inference phase, a new point cloud is fed to the backbone network and the coarse-to-fine instance segmentator to segment instances.", + "bbox": [ + 511, + 329, + 906, + 420 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Sketchy Bounding Box Setting", + "text_level": 1, + "bbox": [ + 511, + 433, + 782, + 450 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our method focuses on studying weakly supervised 3D instance segmentation based on the sketchy bounding box, i.e., inaccurate bounding box. In practice, it is unable to acquire compact 3D bounding box annotations for weakly supervised 3D instance segmentation. Therefore, we perturb the ground truth bounding box with scaling, translation and rotation to mimic the sketchy bounding boxes. As shown in Fig.2, we illustrate the generation process of different types of sketchy bounding boxes.", + "bbox": [ + 511, + 457, + 906, + 593 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Specifically, given the ground truth bounding box, which is denoted by the position of two corner points:", + "bbox": [ + 511, + 594, + 905, + 625 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathbf {B} = \\left[ B _ {\\text {m i n}}, B _ {\\text {m a x}} \\right] \\\\ = \\left[ x _ {\\min }, y _ {\\min }, z _ {\\min }, x _ {\\max }, y _ {\\max }, z _ {\\max } \\right] \\tag {1} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 626, + 905, + 661 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "First, we individually employ scaling, translation, and rotation on the ground truth bounding box to generate the basic inaccurate bounding boxes, which are formulated as:", + "bbox": [ + 511, + 662, + 906, + 707 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {B} _ {\\text {s c a l e d}} = \\left[ B _ {\\min } - \\alpha E, B _ {\\max } + \\alpha E \\right]\n$$\n", + "text_format": "latex", + "bbox": [ + 560, + 709, + 816, + 726 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {B} _ {\\text {t r a n s l a t e d}} = \\left[ B _ {\\min } + \\beta E, B _ {\\max } + \\beta E \\right] \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 560, + 728, + 906, + 744 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {B} _ {\\text {r o t a t e d}} = r \\left(\\left[ B _ {\\min }, B _ {\\max } \\right], M E, \\gamma \\pi / 1 8 0\\right)\n$$\n", + "text_format": "latex", + "bbox": [ + 560, + 747, + 856, + 763 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\alpha$ , $\\beta$ , $\\gamma$ are the parameters of different operations. Note that $E = B_{max} - B_{min}$ and $ME = (B_{max} + B_{min}) / 2$ . Considering that manually annotated 3D bounding boxes will not have significant differences from compact 3D ground truth bounding boxes. Therefore, we empirically set small values for those parameters. In experiments, we set $\\alpha = 5\\%$ , $\\beta = 5\\%$ , and $\\gamma = 5$ , respectively. Then, we generate diverse sketchy bounding boxes by combining these basic operations $\\mathbf{B}_{\\mathrm{scaled}}$ , $\\mathbf{B}_{\\mathrm{translated}}$ , and $\\mathbf{B}_{\\mathrm{rotated}}$ . In Fig.2, we show four", + "bbox": [ + 511, + 763, + 906, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "8881", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/39042b8c3d03282cc29d73fb3162943ad0a2d1cbaa6ff6603b1476a41571a94c.jpg", + "image_caption": [ + "Figure 3. The framework of the Sketchy-3DIS. Given a point cloud with sketchy bounding-box annotations, we first extract the backbone features using a 3D U-Net backbone, then feed them into the adaptive box-to-point pseudo labeler and the coarse-to-fine instance segmentator, finally, we utilize the generated high quality pseudo labels to supervise the predicted instances periodically." + ], + "image_footnote": [], + "bbox": [ + 107, + 89, + 344, + 378 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/be8a8c9f7ca40e4e866edadb8704c88d36dd40fa88b0619efb6acbf007580224.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 357, + 88, + 875, + 377 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "types of sketchy bounding boxes, named $S_{1}$ , $S_{2}$ , $S_{3}$ , and $S_{4}$ , the red rectangle denotes the accurate box and the blue rectangles denote various sketchy bounding boxes. It can be observed that the \"sketchy degree\" increases from $S_{1}$ to $S_{4}$ , making instance segmentation more challenging.", + "bbox": [ + 89, + 438, + 483, + 513 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Adaptive Box-to-Point Pseudo Labeler", + "text_level": 1, + "bbox": [ + 89, + 527, + 421, + 542 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The sketchy 3D bounding box supervision is an inaccurate weak annotation for 3D instance segmentation. In order to convert rough annotated boxes into fine-grained point-level instance annotations, we propose an adaptive box-to-point pseudo labeler that can assign the points located in the box to the corresponding instance. Specifically, we first distinguish the points that lie in bounding boxes, and then adaptively assign them to the corresponding instance by learning the correlation between the points and 3D bounding boxes.", + "bbox": [ + 88, + 550, + 482, + 686 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Box Conditioned Points Partition. According to the relationship between points and 3D bounding boxes, there are three types of points: the points located outside the bounding boxes (i.e., background points), the points lying in a single bounding box, and the points belonging to multiple bounding boxes. Since the background points (see black circle points in Fig. 3) are not within any 3D bounding boxes, they are directly labeled as background. The point lying in a single or multiple bounding box could be a point on the target (see green/yellow/blue circle points in Fig. 3) or a background point. Therefore, we need other strategies to help distinguish the target points from the background points.", + "bbox": [ + 89, + 688, + 482, + 883 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Adaptive Point-to-Instance Assignment. For the points", + "bbox": [ + 109, + 885, + 482, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "lying in a single 3D bounding box (see bounding box $B_{3}$ in Fig. 3), we filter the background points (i.e., black circle points in $B_{3}$ ) by comparing the similarity between the points and the 3D bounding box. Specifically, we consider the similarity in both coordinate space and feature space. The box coordinate and box feature is obtained by averaging the points' coordinates and features within them. The similarity between the points and the 3D bounding boxes can be formulated as follows:", + "bbox": [ + 511, + 438, + 906, + 571 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {s} _ {p, B} = \\cos \\left(\\boldsymbol {f} _ {p}, \\boldsymbol {f} _ {B}\\right) \\times e ^ {- \\left| \\boldsymbol {c} _ {B} \\boldsymbol {c} _ {p} \\right|} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 602, + 575, + 906, + 594 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\cos (\\cdot ,\\cdot)$ denotes cosine similarity. In addition, $f_{p}$ and $f_{B}$ mean the features of point and box, while $c_{p}$ and $c_{B}$ mean the coordinates of point and box. In this way, it is expected to filter background points with low similarity.", + "bbox": [ + 511, + 595, + 905, + 655 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For the points lying in the overlapped parts of multiple 3D bounding boxes (usually two boxes, see boxes $B_{1}$ and $B_{2}$ in Fig. 3), they are assigned to the corresponding instances by learning the similarity between the points and the bounding boxes. However, due to the uncertainty of point labels in overlapping areas, it is difficult to obtain high-confidence semantic labels for these points from the bounding boxes. Therefore, we consider removing the overlapping points within each box and using the remaining reliable points (see $B_{1}^{rel}$ and $B_{2}^{rel}$ in Fig. 3) for similarity learning. Since reliable points are only located within one box, their labels are consistent with the box labels. Given a point feature $\\pmb{f}_{p}$ , box features $\\pmb{f}_{B_1^{rel}}$ and $\\pmb{f}_{B_2^{rel}}$ , we utilize a multilayer perceptron (MLP) network to learn the similarity score $A_{p2b}$ which is given by:", + "bbox": [ + 511, + 656, + 908, + 882 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {A} _ {p 2 b} = \\mathrm {M L P} (\\boldsymbol {f} _ {p}, \\boldsymbol {f} _ {p} - \\boldsymbol {f} _ {B _ {1} ^ {r e l}}, \\boldsymbol {f} _ {p} - \\boldsymbol {f} _ {B _ {2} ^ {r e l}}) \\qquad (4)\n$$\n", + "text_format": "latex", + "bbox": [ + 565, + 883, + 906, + 905 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "8882", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/78d5395f99d8466fd8903160a90303837105d32f65595a7eb5b1b017880c8fd2.jpg", + "image_caption": [ + "Figure 4. The details of Multi-level Attention Block. The instance queries interact with the features of the whole scene, the coarse instance regions, and the instance core regions hierarchically. And the $B_{mask}$ are the boxes obtained from the predicted masks." + ], + "image_footnote": [], + "bbox": [ + 94, + 88, + 480, + 314 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\pmb{f}_{p} - \\pmb{f}_{B_{1}^{rel}}$ and $\\pmb{f}_{p} - \\pmb{f}_{B_{2}^{rel}}$ capture the difference between the point and boxes. In addition, the box feature is obtained by averaging the point features within it. Note that before computing the box feature, we eliminate the background points in the box through Eq. (3). To supervise the similarity learning, we use the cross-entropy loss to formulate pseudo label loss $L_{pl}$ , which is defined as:", + "bbox": [ + 89, + 392, + 483, + 498 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {p l} = L _ {C E} (\\mathbf {A}, \\mathbf {Y}) \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 220, + 500, + 483, + 517 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\pmb{Y}$ is the label of the reliable point, which denotes the box it belongs to.", + "bbox": [ + 89, + 518, + 483, + 547 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After that, we use the model trained on the reliable points (see $B_1^{rel}$ and $B_2^{rel}$ in Fig. 3) to predict the label of unreliable points (see $B_1^{un}$ and $B_2^{un}$ in Fig. 3). It is desired that the points located in the overlapping part of the box can be assigned to the corresponding instances, obtaining pseudo instance labels.", + "bbox": [ + 89, + 549, + 483, + 638 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Coarse-to-Fine Instance Segmentator", + "text_level": 1, + "bbox": [ + 89, + 651, + 413, + 667 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The bounding box supervision provides coarse shape information for 3D instance segmentation. To obtain well-detailed instances, we present a coarse-to-fine instance segmentator that can recognize and segment the points in the point cloud into various instances. Specifically, we first employ the queries to learn from the whole point cloud to obtain coarse instances, and then these queries learn from the regions indicated by the coarse instances with Multi-level Attention Blocks to refine the instances.", + "bbox": [ + 89, + 672, + 483, + 808 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Coarse Instances Segmentation. Supposing one scene consists of $q$ objects, and each query represents one object and models the instance-related characteristics: the semantic category and the mask of each instance. In our method, these queries first interact with the backbone features of the whole point cloud to obtain coarse instances. Considering", + "bbox": [ + 89, + 810, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "the point-wise masks of the coarse instances can not describe one object holistically and precisely, we predict another two boxes with coarse granularity and fine granularity, respectively. Then these two boxes combined with the scene-level features to refine the initial coarse instances with Multi-level Attention Blocks.", + "bbox": [ + 511, + 90, + 906, + 180 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Hierarchical Instances Refinement. To leverage the advantages of multi-granularity features from the point cloud and the segmented coarse instances, we introduce a novel decoder structure utilizing attentions to make the instance queries interact with granularity-varied features hierarchically. The decoder consists of a stack of six blocks. Each block includes a Multi-level Attention Block and an instance segmentation head illustrated in Fig. 3.", + "bbox": [ + 511, + 181, + 908, + 301 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Multi-level Attention Block. The details of the Multi-level Attention Block are shown in Fig. 4. These instance queries first interact with point-wise backbone features globally, then interact with regions described by the coarse instance locally. The regions described by the coarse instance are divided into reliable and unreliable, which are depicted by the predicted box and the core box, respectively. To fully explore various granularity of shape information implied by the coarse instances and those two boxes, we first abstract the obtained coarse instance into a box, termed as $B_{mask}$ . Then we explore the consistent information between $B_{mask}$ and the predicted box $B_{pred}$ to obtain features of reliable regions, which are described as follows:", + "bbox": [ + 511, + 301, + 908, + 497 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {F} ^ {r e l} = \\sigma \\left(\\boldsymbol {F}, \\boldsymbol {M} \\odot e ^ {I o U \\left(\\boldsymbol {B} _ {p r e d}, \\boldsymbol {B} _ {m a s k}\\right)}\\right) \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 498, + 906, + 517 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\sigma$ denotes the thresholding operation and $\\odot$ denotes point-wise multiplication. In addition, $F$ denotes the backbone features, $M$ denotes the predicted mask probabilities. $B_{pred}$ and $B_{mask}$ denote the predicted boxes and boxes decided by predicted masks, respectively. So far, the instance queries have learned from the entire point cloud and the reliable regions, they can segment the majority of objects with easy contexts. To tackle the complex contexts, the instance queries learn from the unreliable regions described by the predicted core boxes. The predicted core boxes are scaled from the predicted boxes, aiming to represent the core regions of objects, and corresponding features are obtained as follows:", + "bbox": [ + 511, + 518, + 908, + 712 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {F} ^ {u n} = \\sigma (\\boldsymbol {F}, \\boldsymbol {M} \\odot (\\boldsymbol {B} _ {\\text {c o r e}} \\cap \\boldsymbol {B} _ {\\text {m a s k}})) \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 578, + 717, + 906, + 733 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $B_{core}$ denotes the predicted core boxes. The instance queries have strong discriminative ability in such a global-to-local manner, then they are fed into a self-attention module and a feed-forward network to aggregate the learned information, which further boosts the performance.", + "bbox": [ + 511, + 734, + 906, + 811 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Bilateral Matching and Training Loss", + "text_level": 1, + "bbox": [ + 511, + 816, + 841, + 834 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our method makes bilateral matching between pseudo instances and predicted instances first and then utilizes the paired pseudo instances as ground truth $\\mathbf{y}_{ins}$ to supervise the paired predicted instances $\\mathbf{x}_{ins}$ .", + "bbox": [ + 511, + 839, + 908, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "8883", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Bilateral matching. Following SPFormer [36] and IS-BNet [30], our method utilizes the Hungarian method [18] to achieve the optimal assignment. We compute a pairwise matching cost $\\mathbf{c}_{i,j}$ to evaluate the similarity of the $i^{th}$ predicted instance and the $j^{th}$ pseudo ground truth. And $\\mathbf{c}^{i,j}$ is based on the classification probabilities and the binary masks, it is defined as:", + "bbox": [ + 89, + 90, + 483, + 195 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathbf {c} ^ {i, j} = \\lambda_ {1} L _ {C E} (\\mathbf {x} _ {c l s} ^ {i}, \\mathbf {y} _ {c l s} ^ {j}) \\\\ + \\lambda_ {2} \\left(L _ {B C E} \\left(\\mathbf {x} _ {\\text {m a s k}} ^ {i}, \\mathbf {y} _ {\\text {m a s k}} ^ {j}\\right) + L _ {\\text {d i c e}} \\left(\\mathbf {x} _ {\\text {m a s k}} ^ {i}, \\mathbf {y} _ {\\text {m a s k}} ^ {j}\\right)\\right) \\tag {8} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 91, + 196, + 482, + 250 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Training loss. Our training loss consists of two items, namely the loss of pseudo label generation $L_{pl}$ and the loss of instance segmentation $L_{seg}$ , which is defined as:", + "bbox": [ + 89, + 250, + 483, + 296 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nL = L _ {p l} + L _ {s e g} \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 230, + 297, + 482, + 314 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As for the instance segmentation, except for supervising the predicted semantic classes and the binary masks, we also supervise the predicted bounding boxes and the core boxes utilized in the coarse-to-fine instance segmentator. The ground truth of the scale of the core box is based on the IoU score of the predicted masks and the pseudo-ground truth masks. The loss of instance segmentation is defined as:", + "bbox": [ + 89, + 316, + 483, + 421 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} L _ {s e g} = \\lambda_ {1} L _ {C E} (\\mathbf {x} _ {c l s}, \\mathbf {y} _ {c l s}) \\\\ + \\lambda_ {2} \\left(L _ {B C E} \\left(\\mathbf {x} _ {\\text {m a s k}}, \\mathbf {y} _ {\\text {m a s k}}\\right) + L _ {\\text {d i c e}} \\left(\\mathbf {x} _ {\\text {m a s k}}, \\mathbf {y} _ {\\text {m a s k}}\\right)\\right) \\\\ + \\lambda_ {3} \\left(L _ {L _ {1}} \\left(\\mathbf {x} _ {\\text {b o x}}, \\mathbf {y} _ {\\text {b o x}}\\right) + L _ {M S E} \\left(\\mathbf {x} _ {\\text {c o r e - b o x}}, \\mathbf {y} _ {\\text {c o r e - b o x}}\\right)\\right) \\tag {10} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 98, + 422, + 482, + 489 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 505, + 223, + 522 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 89, + 530, + 281, + 546 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. We conduct experiments on two datasets, ScanNetV2 [6] and S3DIS [1]. The ScanNetV2 provides splits containing 1201, 312, and 100 scenes for training, validation, and testing. Scenes are annotated with semantic labels covering 18 categories and instance labels. And the S3DIS dataset is another indoor dataset contains 272 scenes across six different areas, and it is annotated with 13 object categories. Following existing methods [4, 27, 29], the scenes of S3DIS Area 5 are used for validation while the others are used for training.", + "bbox": [ + 89, + 551, + 483, + 703 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation metrics. We use the mean average precision $(AP)$ as the evaluation metric for both ScanNetV2 and S3DIS, which is utilized as a common evaluation metric in 3D instance segmentation. $AP$ is the averaged score with IoU thresholds from $95\\%$ to $50\\%$ with a step size of $5\\%$ . The mean average precision with IoU thresholds at $50\\%$ and $25\\%$ are used as evaluation metrics as well, which are denoted as $AP_{50}$ and $AP_{25}$ , respectively.", + "bbox": [ + 89, + 704, + 483, + 824 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation details. Our model is trained on a single RTX 3090 GPU, and we utilize SPFormer [36] and ISBNet [30] as codebases for ScanNetV2 and S3DIS, respectively. We utilize the AdamW optimizer with a learning rate of 0.0002 and weight decay at 0.05 for the training of Scan-", + "bbox": [ + 89, + 825, + 483, + 901 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/6db695164da17199b3bc217c7dd01514e877592dbd741c1b2b368b7b99acfd18.jpg", + "table_caption": [ + "Table 1. Comparison on ScanNetV2. \"Mask\" denotes using dense point-wise annotations, $S_0$ and $S_1$ denote using accurate bounding box and sketchy bounding box annotations, respectively." + ], + "table_footnote": [], + "table_body": "
MethodSup.VenueScanNet TestScanNet Val
AP50AP25AP50AP25
GraphCut [13]NIPS2273.283.269.179.3
ISBNet [30]CVPR2376.384.573.182.5
SPFormer [36]MaskAAAI2377.085.173.982.9
Mask3D [34]ICRA2378.087.073.783.5
MSTA3D [40]ACM MM2479.587.977.085.4
Spherical Mask [35]CVPR2481.2-79.988.2
Box2Mask [4]ECCV2267.780.359.771.8
WISGP+PointGroup [8]WACV23--50.264.9
WISGP+SSTNet [8]WACV23--56.970.2
GaPro+SPFormer [29]S0ICCV2369.282.470.479.9
CIP-WPIS [49]WACV24--69.378.6
BSNet+SPFormer [27]CVPR24--72.783.4
Sketchy-3DIS (ours)-70.186.668.883.6
Box2Mask [4]ECCV22--52.467.5
GaPro+SPFormer [29]S1ICCV23--53.572.2
Sketchy-3DIS (ours)---65.883.1
", + "bbox": [ + 514, + 132, + 903, + 382 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "NetV2. As for the weights of matching cost and losses, $(\\lambda_{1},\\lambda_{2},\\lambda_{3})$ are set as (0.5, 1.0, 0.5). All the models are trained from scratch, and without specification, the hyperparameters and the training details are kept the same as the [36] or [30].", + "bbox": [ + 511, + 398, + 906, + 474 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Main Results", + "text_level": 1, + "bbox": [ + 511, + 484, + 653, + 500 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ScanNetV2. In Table 1, we compare our method and existing state-of-the-art methods on ScanNetV2 dataset. Attributed to the innovative design of Sketchy-3DIS, which achieves the pseudo labeling and instance segmenting simultaneously, our method achieves considerable performance on both the validation set and online test set of ScanNetV2. Notably, without multi-modal priors [49] or synthesized scenes [27], our method outperforms existing methods in terms of $AP_{25}$ , which demonstrates the effectiveness of our method. Additionally, the comparable performances to fully-supervised methods demonstrate that our method can play a significant role in autonomous driving with high efficacy.", + "bbox": [ + 511, + 507, + 908, + 688 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "S3DIS. In Table 2, we make comparisons between existing methods and our proposed Sketchy-3DIS on S3DIS Area 5. The consistent advantages demonstrate the strong generalization ability of our proposed SKeychy-3DIS. Notably, our method exceeds the fully-supervised baseline [30] with $+3.3$ in $AP_{50}$ , which verifies that our proposed Sketchy-3DIS can make full use of instance indications provided by the sketchy bounding boxes.", + "bbox": [ + 511, + 688, + 908, + 809 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Quality comparisons of pseudo labels. We compare the qualities of the generated pseudo labels of GaPro [29] and ours under $S_{1}$ sketchy bounding boxes. The qualitative comparisons are present in Fig.5, which illustrates that our method generates more accurate pseudo labels than GaPro not only in details (highlighted by the green cycles) but also", + "bbox": [ + 511, + 810, + 908, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "8884", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/9d880ff7dafa0db6e88c7c18acd2350247eb866e1c757d8bdd714155c09b071b.jpg", + "image_caption": [ + "Input Point Cloud" + ], + "image_footnote": [], + "bbox": [ + 140, + 114, + 313, + 196 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2f8055b6abed0018c5e17984d2b3e664e21fb965914e4b24a699fc188c56cacd.jpg", + "image_caption": [ + "Ground Truth" + ], + "image_footnote": [], + "bbox": [ + 316, + 116, + 493, + 198 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e8b5595da1616053db9f43a0d70cd1da605da8d27396ddf7d88f859c0f1ed442.jpg", + "image_caption": [ + "Sketchy-3DIS (ours)" + ], + "image_footnote": [], + "bbox": [ + 496, + 116, + 671, + 198 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/94d30bdbd197adefb9b0e0dc886f4841f97056e2356d5747e06844ecd209e629.jpg", + "image_caption": [ + "GaPro" + ], + "image_footnote": [], + "bbox": [ + 674, + 116, + 852, + 198 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d4149b60152f290578ac3948825238b5c10583b29683a74323626f5279cfe5e4.jpg", + "image_caption": [ + "Figure 5. Visualization comparison of pseudo labels on the ScanNetV2 training set. The black denotes the background points and other colors denote different objects. The green and red cycles highlight the key regions." + ], + "image_footnote": [], + "bbox": [ + 140, + 198, + 315, + 304 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c433c3e291a7813c54998067d0bad0c89b0ad96370b046f2b32099092a531fc6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 316, + 199, + 491, + 305 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5419585a74f0764d2ecc0a5c26c40639c0ea2143be49371382b8c685b90f8a25.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 496, + 199, + 671, + 305 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e69f7be396bd3c10f8034e3cd8f6430f9397a057f21c2fffa30d14e7be24d259.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 674, + 199, + 849, + 304 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/69f29c14fbfdf8fbdf64b6311e8ec82046750848a38c298e3021cfdd57171ba1.jpg", + "table_caption": [ + "Table 2. Comparison on S3DIS Area 5. \"Mask\" denotes using dense point-wise annotations, $S_0$ and $S_1$ denote using accurate bounding box and sketchy bounding box annotations, respectively. \\*\\*\" denotes the results are reproduced by GaPro [29]." + ], + "table_footnote": [], + "table_body": "
MethodSup.VenueAPAP50
SoftGroup [41]CVPR2251.666.1
GraphCut [13]NIPS2254.166.4
ISBNet [30]MaskCVPR2354.065.8
Mask3D [34]ICRA2356.668.4
MSTA3D [40]ACM MM24-70.0
Spherical Mask [35]CVPR2460.572.3
Box2Mask* [4]ECCV2243.654.6
WISGP+PointGroup [8]WACV2333.548.6
WISGP+SSTNet [8]S0WACV2337.251.0
GaPro+ISBNet [29]ICCV2350.561.2
BSNet+ISBNet [27]CVPR2453.064.3
Sketchy-3DIS (ours)-53.469.1
Sketchy-3DIS (ours)S1-50.764.6
", + "bbox": [ + 91, + 404, + 483, + 626 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "in distinguishing the objects belonging to the same categories (highlighted by the red cycles). Additionally, compared with the ground truth, our method generates comparable pseudo labels as well.", + "bbox": [ + 89, + 642, + 482, + 700 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Studies and Analysis", + "text_level": 1, + "bbox": [ + 89, + 712, + 357, + 728 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We conduct comprehensive ablation studies to justify the effective design of our method. All these ablation experiments are conducted on the ScanNetV2 validation set unless otherwise stated.", + "bbox": [ + 89, + 734, + 483, + 792 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Robustness under various sketchy bounding boxes. In Table 3, we evaluate our method under four kinds of various sketchy bounding boxes illustrated in Fig.2 on both ScanNetV2 validation set and S3DIS Area 5. It can be observed that the performance of our method deteriorates slightly as the increasing of the boxes' sketchy degree, which verifies the robustness of our proposed Sketchy-3DIS.", + "bbox": [ + 89, + 795, + 483, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/90f45368e61c4a71e27b38b36901abb903432836ba4afb05fdbefc635fce1767.jpg", + "table_caption": [ + "Table 3. Robustness under various sketchy bounding boxes. $S_0$ to $S_4$ denote the sketchy degree of the annotated bounding boxes." + ], + "table_footnote": [], + "table_body": "
Sketchy DegreeScanNetV2 ValS3DIS Area 5
AP50AP25AP50AP25
S462.580.150.963.6
S362.981.452.366.6
S263.782.161.272.7
S165.883.164.673.9
S068.883.669.177.5
", + "bbox": [ + 516, + 377, + 903, + 484 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/69427de89d4f970e248cdf4f86d44ba0248944cdf0c1eae49bd39069ae9ccd07.jpg", + "table_caption": [ + "Table 4. Effectiveness of pseudo labeler and instance segmentator. UnScene3D [32] is an sota unsupervised method. \"disjoint\" denotes training the 3D UNet and pseudo labeler first and then freeze them to train instance segmentator while \"joint\" denotes training the pseudo labeler and the instance segmentator jointly." + ], + "table_footnote": [], + "table_body": "
MethodAPAP50AP25
UnScene3D [32]21.440.352.6
SKetchy-3DIS (disjoint)45.360.470.0
SKetchy-3DIS (joint)53.469.177.5
", + "bbox": [ + 514, + 566, + 903, + 643 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effectiveness of labeler and segmentator. In Table 4, experiments are conducted on the S3DIS Area 5 to verify the effectiveness of our designed adaptive box-to-point pseudo labeler and coarse-to-fine instance segmentator. We choose the state-of-the-art unsupervised method UnScene3D [32] for 3D instance segmentation as the baseline. The experimental results verify that our designed pseudo labeler can generate high-quality pseudo labels, and combined with our designed instance segmentator, our method achieves significant performance.", + "bbox": [ + 509, + 657, + 906, + 808 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effectiveness of components in pseudo labeler. In Table 5, we evaluate the effectiveness of various components in our designed pseudo labeler. The first row is the results of the state-of-the-art unsupervised method [32], and the comparison between the last two rows verifies that the learned point-to-box similarity and point-to-instance assignments", + "bbox": [ + 509, + 810, + 908, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "8885", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/5fd035eedf06b6a6e1d7d3946d909402b0d2063533b66fef7b0b24267ceec5ef.jpg", + "table_caption": [ + "Table 5. Effectiveness of components in pseudo labeler. \"Partition\" denotes obtaining point-wise labels by box-to-point spatial correlations, \"Assign\" denotes assigning the points within overlapped boxes, and \"Similarity\" denotes filtering the background points within boxes according to the point-to-box similarities." + ], + "table_footnote": [], + "table_body": "
PartitionAssignSimilarityAPAP50AP25
15.932.258.5
✓41.864.872.3
✓✓45.267.383.4
✓✓✓46.068.883.6
", + "bbox": [ + 91, + 157, + 483, + 250 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/1499a4b39ab1649565a9b4c34a66542c073bd373ca0cdbbba0248154840db225.jpg", + "table_caption": [ + "Table 6. Attentions in Multi-level Attention Block. \"Scene\", \"Coarse\", and \"Core\" denote interacting with the features of the whole scene, the coarse instance regions, and the core regions of the instance, respectively. And \"Self\" denotes augmenting by the learned instance query itself." + ], + "table_footnote": [], + "table_body": "
SceneCoarseCoreSelfAPAP50AP25
41.364.179.8
✓✓44.667.082.1
✓✓✓44.767.182.9
✓✓✓✓46.068.883.6
", + "bbox": [ + 91, + 334, + 483, + 426 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/37dd8ee275b96faf76194adfe248f6b260a2c7648fbad4a5c9903c67f27e5c4f.jpg", + "table_caption": [ + "Table 7. Effectiveness of instance segmentation losses. $L_{box}$ , $L_{core-box}$ , and $L_{mask}$ denote the supervisions of the predicted instance box, instance core box, and instance mask, respectively." + ], + "table_footnote": [], + "table_body": "
LboxLcore-boxLmaskAPAP50AP25
✓45.266.882.3
✓✓45.467.783.0
✓✓45.067.883.0
✓✓✓46.068.883.6
", + "bbox": [ + 91, + 482, + 483, + 580 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "can improve the performance by filtering the background points within boxes and assigning the points within overlapped boxes.", + "bbox": [ + 88, + 595, + 485, + 638 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effectiveness of various attentions in Multi-level Attention Block. In Table 6, we conduct ablation studies on various combinations of attention layers utilized in Multi-level Attention Block. The best performance is obtained with the combination of the four attention modules. These experimental results verify that our proposed Multi-level Attention Block can explore the local contexts adaptively and aggregates them, which boosts the segmentation performance progressively.", + "bbox": [ + 88, + 641, + 485, + 777 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effectiveness of instance segmentation losses. In Table 7, we compare the performances of the losses of the predicted boxes, core boxes, and masks. The results show that the best performance is obtained with the combination of the above three kinds of losses, combined with our coarse-to-fine instance segmentator, they achieve inspiring performance.", + "bbox": [ + 88, + 777, + 483, + 869 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Qualitative results of 3D instance segmentation. Some visualized samples of predicted instances for the ScanNetV2", + "bbox": [ + 89, + 869, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0f68def47a1778c26a28d7293916f8e2da019ff4160b7b33f61698abb1321a20.jpg", + "image_caption": [ + "Figure 6. Qualitative results on ScanNetV2 validation set. The black denotes the background points and other colors denote different objects. The cycles highlight the key regions." + ], + "image_footnote": [], + "bbox": [ + 514, + 93, + 897, + 335 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "validation set are shown in Fig.6. It can be observed that our method shows consistent effectiveness in segmenting objects for easy and complex scenes. Even if there are many objects belonging to the same category and they are close to each other, our method demonstrates considerable performance with the sketchy bounding box annotation. Notably, our method can adaptively segment the objects that are ignored in the ground truth such as the cabinets highlighted in red cycles, which demonstrates the strong recognition capability of our method.", + "bbox": [ + 511, + 400, + 908, + 551 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 564, + 633, + 579 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we propose a query-based framework for weakly supervised 3D instance segmentation, Sketchy-3DIS, which is more robust and effective than existing bounding-box supervised methods. Sketchy-3DIS adaptively generates pseudo per-point labels from sketchy boxes and predicts instances in a coarse-to-fine manner. Extensive experiments on ScanNetV2 and S3DIS benchmarks show that our method achieves leading performance, and even outperforms some fully-supervised methods. However, our explored experiments demonstrate that the performance of our method degraded severely when the annotated sketchy bounding boxes are immensely inaccurate, exploration for this issue may be the future study orientation.", + "bbox": [ + 511, + 590, + 908, + 787 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 513, + 800, + 671, + 816 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work was supported by the National Key R&D Program of China No. 2024YFC3015801, National Science Fund of China under Grant Nos. 62361166670, 62276144, U24A20330, 62306238 and the Fundamental Research Funds for the Central Universities.", + "bbox": [ + 511, + 824, + 908, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8886", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Iro Armeni, Ozan Sener, Amir R Zamir, Helen Jiang, Ioannis Brilakis, Martin Fischer, and Silvio Savarese. 3D semantic parsing of large-scale indoor spaces. In CVPR, pages 1534-1543, 2016. 6", + "[2] Shaoyu Chen, Jiemin Fang, Qian Zhang, Wenyu Liu, and Xinggang Wang. Hierarchical aggregation for 3D instance segmentation. In CVPR, pages 15467-15476, 2021. 1, 2", + "[3] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In CVPR, pages 1290–1299, 2022. 2", + "[4] Julian Chibane, Francis Engelmann, Tuan Anh Tran, and Gerard Pons-Moll. Box2Mask: Weakly supervised 3D semantic instance segmentation using bounding boxes. In ECCV, pages 681-699. Springer, 2022. 1, 2, 3, 6, 7", + "[5] Ruihang Chu, Xiaqing Ye, Zhengzhe Liu, Xiao Tan, Xiaojuan Qi, Chi-Wing Fu, and Jiaya Jia. Twist: Two-way inter-label self-training for semi-supervised 3D instance segmentation. In CVPR, pages 1100-1109, 2022. 1, 2", + "[6] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. ScanNet: Richly-annotated 3D reconstructions of indoor scenes. In CVPR, pages 5828-5839, 2017. 1, 6", + "[7] Shichao Dong, Ruibo Li, Jiacheng Wei, Fayao Liu, and Guosheng Lin. Rwseg: Cross-graph competing random walks for weakly supervised 3D instance segmentation. arXiv preprint arXiv:2208.05110, 2022. 1, 2", + "[8] Heming Du, Xin Yu, Farookh Hussain, Mohammad Ali Armin, Lars Petersson, and Weihao Li. Weakly-supervised point cloud instance segmentation with geometric priors. In WACV, pages 4271-4280, 2023. 1, 2, 3, 6, 7", + "[9] Benjamin Graham, Martin Engelcke, and Laurens Van Der Maaten. 3D semantic segmentation with submanifold sparse convolutional networks. In CVPR, pages 9224-9232, 2018. 2", + "[10] Tong He, Chunhua Shen, and Anton Van Den Hengel. Dyco3D: Robust instance segmentation of 3D point clouds through dynamic convolution. In CVPR, pages 354-363, 2021. 2", + "[11] Ji Hou, Angela Dai, and Matthias Nießner. 3D-SIS: 3D semantic instance segmentation of rgb-d scans. In CVPR, pages 4421-4430, 2019. 2", + "[12] Ji Hou, Benjamin Graham, Matthias Nießner, and Saining Xie. Exploring data-efficient 3D scene understanding with contrastive scene contexts. In CVPR, pages 15587-15597, 2021. 1, 2", + "[13] Le Hui, Linghua Tang, Yaqi Shen, Jin Xie, and Jian Yang. Learning superpoint graph cut for 3D instance segmentation. NIPS, 35:36804-36817, 2022. 1, 2, 6, 7", + "[14] Li Jiang, Hengshuang Zhao, Shaoshuai Shi, Shu Liu, ChiWing Fu, and Jiaya Jia. Pointgroup: Dual-set point grouping for 3D instance segmentation. In CVPR, pages 4867-4876, 2020. 1, 2", + "[15] Lei Ke, Martin Danelljan, Henghui Ding, Yu-Wing Tai, Chi-Keung Tang, and Fisher Yu. Mask-free video instance segmentation. In CVPR, pages 22857–22866, 2023. 3" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[16] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In ICCV, pages 4015-4026, 2023. 2", + "[17] Maxim Kolodiazhnyi, Anna Vorontsova, Anton Konushin, and Danila Rukhovich. OneFormer3D: One transformer for unified point cloud segmentation. In CVPR, pages 20943-20953, 2024. 2", + "[18] Harold W Kuhn. The hungarian method for the assignment problem. Naval research logistics quarterly, 2(1-2):83-97, 1955. 6", + "[19] Xin Lai, Yuhui Yuan, Ruihang Chu, Yukang Chen, Han Hu, and Jiaya Jia. Mask-Attention-Free Transformer for 3D instance segmentation. In ICCV, pages 3693–3703, 2023. 1, 2", + "[20] Shiyi Lan, Zhiding Yu, Christopher Choy, Subhashree Radhakrishnan, Guilin Liu, Yuke Zhu, Larry S Davis, and Anima Anandkumar. DiscoBox: Weakly supervised instance segmentation and semantic correspondence from box supervision. In ICCV, pages 3406-3416, 2021. 3", + "[21] Loic Landrieu and Martin Simonovsky. Large-scale point cloud semantic segmentation with superpoint graphs. In CVPR, pages 4558-4567, 2018. 3", + "[22] Zhihao Liang, Zhihao Li, Songcen Xu, Mingkui Tan, and Kui Jia. Instance segmentation in 3D scenes using semantic superpoint tree networks. In ICCV, pages 2783-2792, 2021. 2", + "[23] Yongbin Liao, Hongyuan Zhu, Yanggang Zhang, Chuangguan Ye, Tao Chen, and Jiayuan Fan. Point cloud instance segmentation with semi-supervised bounding-box mining. TPAMI, 44(12):10159-10170, 2021. 1, 2", + "[24] Zhengzhe Liu, Xiaojuan Qi, and Chi-Wing Fu. One thing one click: A self-training approach for weakly supervised 3D semantic segmentation. In CVPR, pages 1726-1736, 2021. 1, 2", + "[25] Ze Liu, Zheng Zhang, Yue Cao, Han Hu, and Xin Tong. Group-free 3D object detection via transformers. In ICCV, pages 2949-2958, 2021. 1", + "[26] Jiahao Lu, Jiacheng Deng, Chuxin Wang, Jianfeng He, and Tianzhu Zhang. Query refinement transformer for 3D instance segmentation. In ICCV, pages 18516-18526, 2023. 2", + "[27] Jiahao Lu, Jiacheng Deng, and Tianzhu Zhang. BSNet: Box-supervised simulation-assisted mean teacher for 3D instance segmentation. In CVPR, pages 20374-20384, 2024. 2, 3, 6, 7", + "[28] Ishan Misra, Rohit Girdhar, and Armand Joulin. An end-to-end transformer model for 3D object detection. In ICCV, pages 2906-2917, 2021. 1", + "[29] Tuan Duc Ngo, Binh-Son Hua, and Khoi Nguyen. GaPro: Box-supervised 3D point cloud instance segmentation using gaussian processes as pseudo labelers. In ICCV, pages 17794-17803, 2023. 1, 2, 3, 6, 7", + "[30] Tuan Duc Ngo, Binh-Son Hua, and Khoi Nguyen. ISBNet: a 3D point cloud instance segmentation network with instance-aware sampling and box-aware dynamic convolution. In CVPR, pages 13550-13559, 2023. 1, 2, 6, 7", + "[31] Charles R Qi, Or Litany, Kaiming He, and Leonidas J Guibas. Deep hough voting for 3D object detection in point clouds. In ICCV, pages 9277-9286, 2019. 1" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "8887", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[32] David Rozenberszki, Or Litany, and Angela Dai. UnScene3D: Unsupervised 3D instance segmentation for indoor scenes. In CVPR, pages 19957-19967, 2024. 7", + "[33] Danila Rukhovich, Anna Vorontsova, and Anton Konushin. FCAF3D: Fully convolutional anchor-free 3D object detection. In ECCV, pages 477-493. Springer, 2022. 1", + "[34] Jonas Schult, Francis Engelmann, Alexander Hermans, Or Litany, Siyu Tang, and Bastian Leibe. Mask3D: Mask transformer for 3D semantic instance segmentation. In ICRA, pages 8216-8223. IEEE, 2023. 1, 2, 6, 7", + "[35] Sangyun Shin, Kaichen Zhou, Madhu Vankadari, Andrew Markham, and Niki Trigoni. Spherical mask: Coarse-to-fine 3D point cloud instance segmentation with spherical representation. In CVPR, pages 4060-4069, 2024. 6, 7", + "[36] Jiahao Sun, Chunmei Qing, Junpeng Tan, and Xiangmin Xu. Superpoint transformer for 3D scene instance segmentation. In AAAI, pages 2393-2401, 2023. 1, 2, 6", + "[37] Linghua Tang, Le Hui, and Jin Xie. Learning inter-superpoint affinity for weakly supervised 3D instance segmentation. In ACCV, pages 1282-1297, 2022. 1, 2", + "[38] An Tao, Yueqi Duan, Yi Wei, Jiwen Lu, and Jie Zhou. Seg-Group: Seg-level supervision for 3D instance and semantic segmentation. TIP, 31:4952-4965, 2022. 1, 2", + "[39] Zhi Tian, Chunhua Shen, Xinlong Wang, and Hao Chen. BoxInst: High-performance instance segmentation with box annotations. In CVPR, pages 5443-5452, 2021. 3", + "[40] Duc Dang Trung Tran, Byeongkeun Kang, and Yeejin Lee. MSTA3D: Multi-scale twin-attention for 3d instance segmentation. In ACM MM, pages 1467-1475, 2024. 2, 6, 7", + "[41] Thang Vu, Kookhoi Kim, Tung M Luu, Thanh Nguyen, and Chang D Yoo. SoftGroup for 3D instance segmentation on point clouds. In CVPR, pages 2708-2717, 2022. 1, 2, 7", + "[42] Haiyang Wang, Shaocong Dong, Shaoshuai Shi, Aoxue Li, Jianan Li, Zhenguo Li, Liwei Wang, et al. CAGroup3D: Class-aware grouping for 3D object detection on point clouds. NIPS, 35:29975-29988, 2022. 1", + "[43] Haiyang Wang, Shaoshuai Shi, Ze Yang, Rongyao Fang, Qi Qian, Hongsheng Li, Bernt Schiele, and Liwei Wang. RBGNet: Ray-based grouping for 3D object detection. In CVPR, pages 1110-1119, 2022.", + "[44] Qian Xie, Yu-Kun Lai, Jing Wu, Zhoutao Wang, Yiming Zhang, Kai Xu, and Jun Wang. MLCVNet: Multi-level context votenet for 3D object detection. In CVPR, pages 10447-10456, 2020. 1", + "[45] Saining Xie, Jiatao Gu, Demi Guo, Charles R Qi, Leonidas Guibas, and Or Litany. PointContrast: Unsupervised pretraining for 3D point cloud understanding. In ECCV, pages 574-591. Springer, 2020. 1, 2", + "[46] Bo Yang, Jianan Wang, Ronald Clark, Qingyong Hu, Sen Wang, Andrew Markham, and Niki Trigoni. Learning object bounding boxes for 3D instance segmentation on point clouds. NIPS, 32, 2019. 2", + "[47] Cheng-Kun Yang, Min-Hung Chen, Yung-Yu Chuang, and Yen-Yu Lin. 2D-3D interlaced transformer for point cloud segmentation with scene-level supervision. In ICCV, pages 977–987, 2023. 2" + ], + "bbox": [ + 91, + 90, + 485, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[48] Li Yi, Wang Zhao, He Wang, Minhyuk Sung, and Leonidas J Guibas. GSPN: Generative shape proposal network for 3D instance segmentation in point cloud. In CVPR, pages 3947-3956, 2019. 2", + "[49] Qingtao Yu, Heming Du, Chen Liu, and Xin Yu. When 3D bounding-box meets SAM: Point cloud instance segmentation with weak-and-noisy supervision. In WACV, pages 3719-3728, 2024. 1, 2, 3, 6", + "[50] Zaiwei Zhang, Bo Sun, Haitao Yang, and Qixing Huang. H3DNet: 3D object detection using hybrid geometric primitives. In ECCV, pages 311-329. Springer, 2020. 1", + "[51] Yu Zheng, Yueqi Duan, Jiwen Lu, Jie Zhou, and Qi Tian. HyperDet3D: Learning a scene-conditioned 3D object detector. In CVPR, pages 5585-5594, 2022. 1", + "[52] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable DETR: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159, 2020. 2" + ], + "bbox": [ + 514, + 90, + 906, + 345 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "8888", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Sketchy Bounding-box Supervision for 3D Instance Segmentation/fed9c100-3d53-4c29-809d-ba6f2288846b_model.json b/2025/Sketchy Bounding-box Supervision for 3D Instance Segmentation/fed9c100-3d53-4c29-809d-ba6f2288846b_model.json new file mode 100644 index 0000000000000000000000000000000000000000..41c74fcbcd44ffd9a99d01b491b8d4c3ce7b5c06 --- /dev/null +++ b/2025/Sketchy Bounding-box Supervision for 3D Instance Segmentation/fed9c100-3d53-4c29-809d-ba6f2288846b_model.json @@ -0,0 +1,2453 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.168, + 0.131, + 0.83, + 0.154 + ], + "angle": 0, + "content": "Sketchy Bounding-box Supervision for 3D Instance Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.319, + 0.18, + 0.676, + 0.199 + ], + "angle": 0, + "content": "Qian Deng\\(^{1}\\), Le Hui\\(^{2}\\), Jin Xie\\(^{3,4*}\\), Jian Yang\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.199, + 0.199, + 0.798, + 0.217 + ], + "angle": 0, + "content": "1PCA Lab, College of Computer Science, Nankai University, Tianjin, China" + }, + { + "type": "text", + "bbox": [ + 0.132, + 0.217, + 0.867, + 0.234 + ], + "angle": 0, + "content": "\\(^{2}\\)School of Electronics and Information, Northwestern Polytechnical University, Xi'an, China" + }, + { + "type": "text", + "bbox": [ + 0.142, + 0.234, + 0.857, + 0.252 + ], + "angle": 0, + "content": "\\(^{3}\\)State Key Laboratory for Novel Software Technology, Nanjing University, Nanjing, China" + }, + { + "type": "text", + "bbox": [ + 0.167, + 0.252, + 0.83, + 0.27 + ], + "angle": 0, + "content": "\\(^{4}\\)School of Intelligence Science and Technology, Nanjing University, Suzhou, China" + }, + { + "type": "text", + "bbox": [ + 0.104, + 0.273, + 0.882, + 0.287 + ], + "angle": 0, + "content": "dengqian@mail.nankai.edu.cn; huile@nwpu.edu.cn; csjxie@nju.edu.cn; csjyang@nankai.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.322, + 0.327, + 0.339 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.354, + 0.486, + 0.764 + ], + "angle": 0, + "content": "Bounding box supervision has gained considerable attention in weakly supervised 3D instance segmentation. While this approach alleviates the need for extensive point-level annotations, obtaining accurate bounding boxes in practical applications remains challenging. To this end, we explore the inaccurate bounding box, named sketchy bounding box, which is imitated through perturbing ground truth bounding box by adding scaling, translation, and rotation. In this paper, we propose Sketchy-3DIS, a novel weakly 3D instance segmentation framework, which jointly learns pseudo labeler and segmentator to improve the performance under the sketchy bounding-box supervisions. Specifically, we first propose an adaptive box-to-point pseudo labeler that adaptively learns to assign points located in the overlapped parts between two sketchy bounding boxes to the correct instance, resulting in compact and pure pseudo instance labels. Then, we present a coarse-to-fine instance segmentator that first predicts coarse instances from the entire point cloud and then learns fine instances based on the region of coarse instances. Finally, by using the pseudo instance labels to supervise the instance segmentator, we can gradually generate high-quality instances through joint training. Extensive experiments show that our method achieves state-of-the-art performance on both the ScanNetV2 and S3DIS benchmarks, and even outperforms several fully supervised methods using sketchy bounding boxes. Code is available at https://github.com/dengq7/Sketchy-3DIS." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.791, + 0.222, + 0.807 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.816, + 0.485, + 0.879 + ], + "angle": 0, + "content": "With the rise of 3D datasets and the growing need for 3D scene understanding, various 3D computer vision tasks [19, 25, 28-31, 33, 34, 36, 42-44, 50, 51] have gained significant attention from both academia and industry. Among these, 3D" + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.32, + 0.614, + 0.475 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.544, + 0.476, + 0.596, + 0.487 + ], + "angle": 0, + "content": "(a) Inputs" + }, + { + "type": "image", + "bbox": [ + 0.614, + 0.319, + 0.895, + 0.473 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.675, + 0.475, + 0.832, + 0.487 + ], + "angle": 0, + "content": "(b) Performance Comparisons" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.491, + 0.908, + 0.547 + ], + "angle": 0, + "content": "Figure 1. (a) illustrates the inputs of bounding box supervised 3D instance segmentation. (b) compares the performances of GaPro [29] and our Sketchy-3DIS under both accurate and scaled sketchy boxes supervision on ScanNetV2 validation set." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.552, + 0.909, + 0.779 + ], + "angle": 0, + "content": "instance segmentation is both fundamental and challenging, as it involves recognizing the category of each object within a point cloud and delineating its individual mask. Current 3D instance segmentation methods [2, 13, 14, 30, 34, 36, 41] have demonstrated impressive performance, primarily rely on data-hungry deep learning approaches that require substantial training data and dense point-level manual annotations. Notably, point-level fully supervised instance annotation in ScanNet takes approximately 22.3 minutes [6] per scene, while labeling a 3D box in ScanNet takes about 1.93 minutes [38] per scene. Therefore, bounding box annotation stands out as it provides substantial instance-level information while maintaining a reasonable annotation cost, making it a favorable annotation choice for weakly supervised 3D instance segmentation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.911, + 0.903 + ], + "angle": 0, + "content": "Many weakly supervised methods have emerged to tackle the 3D instance segmentation task. Semi-supervised 3D instance segmentation methods [5, 12, 23, 45] use training data with partial labels, while others [5, 12, 23, 45] utilize indirect labels, such as 2D image annotations. Additionally, object-level annotations, like marking a single point [7, 24, 37, 38] or drawing a bounding box [4, 8, 29, 49] for each object, are employed. Bounding box supervision is another important" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.235, + 0.9 + ], + "angle": 0, + "content": "*Corresponding author" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8879" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.333 + ], + "angle": 0, + "content": "annotation for weakly supervised 3D instance segmentation. Box2Mask [4] is the first attempt that uses bounding box supervision for 3D instance segmentation. Subsequently, the following methods [8, 27, 29, 49] use different constraints to improve performance, such as using local correlations, Gaussian processes, teacher-student model, SAM [16], respectively. Although impressive results have been achieved in weakly supervised 3D instance segmentation, these methods still rely on accurate and compact bounding box annotations. It is worth noting that in practice, accurate and compact boxes are difficult to obtain. Generally, annotated bounding box involves changes in scaling, translation, and rotation. As shown in Fig.1, we observed that the performance of GaPro [29] decreases significantly under scaled bounding box supervision. This motivated us to design a new approach to tackle with these inaccurate bounding box annotations." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.336, + 0.485, + 0.849 + ], + "angle": 0, + "content": "In this paper, we propose a weakly supervised 3D instance segmentation framework to tackle sketchy bounding box supervision. To simulate sketchy bounding boxes, we impose scaling, translation, and rotation on the ground truth bounding boxes to generate inaccurate bounding boxes. Our method consists of two key components: adaptive box-to-point pseudo labeler and coarse-to-fine instance segmentator. Given a point cloud scene and sketchy bounding boxes, we first use 3D U-Net [9] backbone to extract point features. After that, we feed the learned features and sketchy bounding boxes of the scene into the adaptive box-to-point pseudo labeler to generate high-quality point-level instance labels. For the points located in the overlapped region between two sketchy bounding boxes, we adaptively assign the points to the corresponding box by learning the similarity of points and boxes. For the remaining points, we directly assign them to the corresponding box according to the spatial correlations. In this way, we can transform the sketchy bounding box into a compact bounding box. We consider the labels of points that located within the generated compact bounding box as pseudo instance labels. To accurately predict instances, we input the learned features into the coarse-to-fine instance segmentator iteratively. We first detect the coarse bounding box and mask of the instance by executing queries on the whole point cloud. Based on the small region of the coarse bounding box, we re-predict the fine bounding box and mask by conducting queries within a more precise small region. Finally, by jointly optimizing the adaptive box-to-point pseudo labeler and coarse-to-fine instance segmentator, our method can gradually improve the performance of weakly supervised 3D instance segmentation using the sketchy bounding box annotations. Rich experiments on the ScanNetV2 and S3DIS show that our method have achieved state-of-the-art performance." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.853, + 0.331, + 0.867 + ], + "angle": 0, + "content": "Our contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.483, + 0.902 + ], + "angle": 0, + "content": "- We are the first to explore the use of sketchy bounding box annotations for 3D instance segmentation, which allows" + }, + { + "type": "text", + "bbox": [ + 0.527, + 0.092, + 0.907, + 0.122 + ], + "angle": 0, + "content": "for training with noise-tolerant, instance-size-agnostic 3D bounding boxes." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.125, + 0.907, + 0.185 + ], + "angle": 0, + "content": "- We propose a new weakly supervised 3D instance segmentation framework, which presents an adaptive box-to-point pseudo labeler and a coarse-to-fine instance segmentator for joint training." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.188, + 0.905, + 0.233 + ], + "angle": 0, + "content": "- Our method achieves leading performance on both the ScanNetV2 and S3DIS datasets for weakly 3D instance segmentation under sketchy bounding box annotations." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.125, + 0.907, + 0.233 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.248, + 0.655, + 0.263 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.273, + 0.884, + 0.289 + ], + "angle": 0, + "content": "2.1. Fully Supervised 3D Instance Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.296, + 0.909, + 0.688 + ], + "angle": 0, + "content": "Existing fully supervised methods for 3D instance segmentation can be broadly categorized into three main pipelines: grouping-based [2, 13, 14, 22], detection-based [11, 41, 46, 48], and query-based [10, 17, 19, 26, 30, 34-36, 40] methods. Grouping-based methods typically group points based on predicted per-point semantic classes and instance center offsets, but the grouping rules are often manually defined by researchers, which lack flexibility. Detection-based methods, on the other hand, first predict the bounding box for each object and then generate the foreground mask within the box, treating the task as a two-step process of object detection followed by segmentation. More recently, inspired by DETR [52] and Mask2Former [3], query-based methods have been proposed for 3D instance segmentation. These methods simultaneously predict the semantic classes and instance masks of objects using a set of learnable query vectors, which model the objects' geometrical and semantical attributes integrally. Among these methods, query-based approaches tend to achieve superior performance, as the other two pipelines both face challenges such as error accumulation and time-consuming refinement processes. While these fully-supervised methods have made significant progress, they all rely on dense per-point annotations, and their performance tends to degrade when such annotations are unavailable. Therefore, several weakly-supervised methods have been proposed to alleviate this problem." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.698, + 0.902, + 0.714 + ], + "angle": 0, + "content": "2.2. Weakly Supervised 3D Instance Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Weakly-supervised methods require partial or implicit annotations. [5, 12, 23, 45] use sparse point annotations and design various label propagation strategies. However, these sparse points cannot fully capture the spatial extent of an object. Approaches like MIT [47] and CIP-WPIS [49] leverage 2D images as additional input for 3D instance segmentation, but this introduces multi-modal dependencies. Other methods [7, 24, 37, 38] rely on instance-level signals, where labeled a single point for per instance, and instance predictions are made using techniques like random walk or clustering. Recently, most state-of-the-art methods [4, 8, 27, 29, 49] have adopted 3D bounding boxes as supervision, where a" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8880" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.152 + ], + "angle": 0, + "content": "single bounding box is assigned to each object. These approaches naturally capture the spatial extent of an object and the annotated bounding boxes are easier to obtain compared to dense per-point annotations." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.153, + 0.487, + 0.562 + ], + "angle": 0, + "content": "Bounding-box Supervised 3D Instance Segmentation. Bounding boxes have become a widely used supervision signal, not only for 2D instance segmentation tasks [15, 20, 39] but also for 3D instance segmentation. Box2mask [4] is a pioneering approach that utilizes bounding-box supervision for 3D instance segmentation. WISGP [8] divides the point cloud into a determined set and an undetermined set based on explicit spatial correlations, then uses polygon meshes and superpoints to label the undetermined set. GaPro [29] focuses on generating reliable pseudo labels using a Gaussian process. Recently, BSNet [27] employs a simulation-assisted mean teacher framework to generate pseudo labels, enhancing data samples by synthesizing scenes that contain overlapping objects. However, these methods all rely on axis-aligned bounding boxes for supervision, and their performances heavily depend on the accuracy of these annotated boxes. To address this, CIP-WPIS [49] proposes using slightly looser bounding boxes, projecting the point cloud into 2D images, and obtaining 3D instance masks based on segmented 2D masks. While effective, it relies on 2D modality and can only adapt to slightly enlarged bounding boxes, limiting its flexibility. To alleviate the annotation costs as well as obtain robust and effective performance, we design a novel framework Sketchy-3DIS, which endeavors to not only face the common challenge of correctly assigning the points within multiple boxes but also be dedicated to correcting the errors brought by sketchy bounding boxes." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.578, + 0.182, + 0.593 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.605, + 0.201, + 0.619 + ], + "angle": 0, + "content": "3.1. Overview" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.628, + 0.486, + 0.688 + ], + "angle": 0, + "content": "The overall framework of our weakly supervised 3D instance segmentation based on sketchy bounding boxes is depicted in Fig.3. It consists of two key components: adaptive box-to-point pseudo labeler and coarse-to-fine instance segmentator." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.485, + 0.901 + ], + "angle": 0, + "content": "Assuming a 3D point cloud \\(P \\in \\mathbb{R}^{N \\times C}\\), where \\(N\\) and \\(C\\) represent the number of 3D points and channels, respectively. In the basic setting, each point is represented by the 3D coordinate, RGB color, and surface normal vector. In the training phase, we first feed the input point cloud and the sketchy bounding boxes (defined in Sec. 3.2) into the 3D UNet backbone network for feature extraction, then we average the point features located in the superpoint [21] to generate superpoint-level features. Next, we input the features and sketchy bounding boxes to the adaptive box-to-point pseudo labeler (Sec. 3.3) for converting the rough box-level annotation to the detailed point-level annotation of the instance. Meanwhile, we feed the feature into the coarse-to-fine instance segmentator (Sec. 3.4) to predict instances by" + }, + { + "type": "image", + "bbox": [ + 0.526, + 0.092, + 0.66, + 0.169 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.554, + 0.17, + 0.632, + 0.182 + ], + "angle": 0, + "content": "Point Cloud" + }, + { + "type": "image", + "bbox": [ + 0.698, + 0.121, + 0.736, + 0.148 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.671, + 0.169, + 0.762, + 0.184 + ], + "angle": 0, + "content": "Accurate \\((S_0)\\)" + }, + { + "type": "image", + "bbox": [ + 0.792, + 0.113, + 0.855, + 0.155 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.789, + 0.169, + 0.862, + 0.184 + ], + "angle": 0, + "content": "Scaled \\((S_{1})\\)" + }, + { + "type": "image", + "bbox": [ + 0.578, + 0.198, + 0.651, + 0.241 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.55, + 0.251, + 0.643, + 0.28 + ], + "angle": 0, + "content": "Scaled and translated \\((S_{2})\\)" + }, + { + "type": "image", + "bbox": [ + 0.678, + 0.191, + 0.755, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.678, + 0.251, + 0.754, + 0.28 + ], + "angle": 0, + "content": "Scaled and rotated \\((S_{3})\\)" + }, + { + "type": "image", + "bbox": [ + 0.804, + 0.192, + 0.883, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.767, + 0.251, + 0.88, + 0.28 + ], + "angle": 0, + "content": "Scaled, translated and rotated \\((S_4)\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.285, + 0.908, + 0.314 + ], + "angle": 0, + "content": "Figure 2. Various sketchy bounding boxes under scaling, translation, and rotation perturbations." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.33, + 0.907, + 0.421 + ], + "angle": 0, + "content": "Transformer. Finally, we employ bilateral matching between the generated pseudo labels (regarded as the ground truth) and the predicted instances to build the correspondence between them for training (Sec. 3.5). In the inference phase, a new point cloud is fed to the backbone network and the coarse-to-fine instance segmentator to segment instances." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.434, + 0.783, + 0.451 + ], + "angle": 0, + "content": "3.2. Sketchy Bounding Box Setting" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.458, + 0.907, + 0.594 + ], + "angle": 0, + "content": "Our method focuses on studying weakly supervised 3D instance segmentation based on the sketchy bounding box, i.e., inaccurate bounding box. In practice, it is unable to acquire compact 3D bounding box annotations for weakly supervised 3D instance segmentation. Therefore, we perturb the ground truth bounding box with scaling, translation and rotation to mimic the sketchy bounding boxes. As shown in Fig.2, we illustrate the generation process of different types of sketchy bounding boxes." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.595, + 0.906, + 0.625 + ], + "angle": 0, + "content": "Specifically, given the ground truth bounding box, which is denoted by the position of two corner points:" + }, + { + "type": "equation", + "bbox": [ + 0.567, + 0.627, + 0.906, + 0.662 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathbf {B} = \\left[ B _ {\\text {m i n}}, B _ {\\text {m a x}} \\right] \\\\ = \\left[ x _ {\\min }, y _ {\\min }, z _ {\\min }, x _ {\\max }, y _ {\\max }, z _ {\\max } \\right] \\tag {1} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.664, + 0.907, + 0.708 + ], + "angle": 0, + "content": "First, we individually employ scaling, translation, and rotation on the ground truth bounding box to generate the basic inaccurate bounding boxes, which are formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.561, + 0.71, + 0.818, + 0.727 + ], + "angle": 0, + "content": "\\[\n\\mathbf {B} _ {\\text {s c a l e d}} = \\left[ B _ {\\min } - \\alpha E, B _ {\\max } + \\alpha E \\right]\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.562, + 0.729, + 0.907, + 0.745 + ], + "angle": 0, + "content": "\\[\n\\mathbf {B} _ {\\text {t r a n s l a t e d}} = \\left[ B _ {\\min } + \\beta E, B _ {\\max } + \\beta E \\right] \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.562, + 0.748, + 0.857, + 0.764 + ], + "angle": 0, + "content": "\\[\n\\mathbf {B} _ {\\text {r o t a t e d}} = r \\left(\\left[ B _ {\\min }, B _ {\\max } \\right], M E, \\gamma \\pi / 1 8 0\\right)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.907, + 0.902 + ], + "angle": 0, + "content": "where \\(\\alpha\\), \\(\\beta\\), \\(\\gamma\\) are the parameters of different operations. Note that \\(E = B_{max} - B_{min}\\) and \\(ME = (B_{max} + B_{min}) / 2\\). Considering that manually annotated 3D bounding boxes will not have significant differences from compact 3D ground truth bounding boxes. Therefore, we empirically set small values for those parameters. In experiments, we set \\(\\alpha = 5\\%\\), \\(\\beta = 5\\%\\), and \\(\\gamma = 5\\), respectively. Then, we generate diverse sketchy bounding boxes by combining these basic operations \\(\\mathbf{B}_{\\mathrm{scaled}}\\), \\(\\mathbf{B}_{\\mathrm{translated}}\\), and \\(\\mathbf{B}_{\\mathrm{rotated}}\\). In Fig.2, we show four" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "8881" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.109, + 0.09, + 0.345, + 0.379 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.359, + 0.089, + 0.877, + 0.378 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.383, + 0.908, + 0.427 + ], + "angle": 0, + "content": "Figure 3. The framework of the Sketchy-3DIS. Given a point cloud with sketchy bounding-box annotations, we first extract the backbone features using a 3D U-Net backbone, then feed them into the adaptive box-to-point pseudo labeler and the coarse-to-fine instance segmentator, finally, we utilize the generated high quality pseudo labels to supervise the predicted instances periodically." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.439, + 0.484, + 0.515 + ], + "angle": 0, + "content": "types of sketchy bounding boxes, named \\( S_{1} \\), \\( S_{2} \\), \\( S_{3} \\), and \\( S_{4} \\), the red rectangle denotes the accurate box and the blue rectangles denote various sketchy bounding boxes. It can be observed that the \"sketchy degree\" increases from \\( S_{1} \\) to \\( S_{4} \\), making instance segmentation more challenging." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.528, + 0.423, + 0.543 + ], + "angle": 0, + "content": "3.3. Adaptive Box-to-Point Pseudo Labeler" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.551, + 0.483, + 0.687 + ], + "angle": 0, + "content": "The sketchy 3D bounding box supervision is an inaccurate weak annotation for 3D instance segmentation. In order to convert rough annotated boxes into fine-grained point-level instance annotations, we propose an adaptive box-to-point pseudo labeler that can assign the points located in the box to the corresponding instance. Specifically, we first distinguish the points that lie in bounding boxes, and then adaptively assign them to the corresponding instance by learning the correlation between the points and 3D bounding boxes." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.483, + 0.884 + ], + "angle": 0, + "content": "Box Conditioned Points Partition. According to the relationship between points and 3D bounding boxes, there are three types of points: the points located outside the bounding boxes (i.e., background points), the points lying in a single bounding box, and the points belonging to multiple bounding boxes. Since the background points (see black circle points in Fig. 3) are not within any 3D bounding boxes, they are directly labeled as background. The point lying in a single or multiple bounding box could be a point on the target (see green/yellow/blue circle points in Fig. 3) or a background point. Therefore, we need other strategies to help distinguish the target points from the background points." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.886, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Adaptive Point-to-Instance Assignment. For the points" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.439, + 0.907, + 0.573 + ], + "angle": 0, + "content": "lying in a single 3D bounding box (see bounding box \\( B_{3} \\) in Fig. 3), we filter the background points (i.e., black circle points in \\( B_{3} \\)) by comparing the similarity between the points and the 3D bounding box. Specifically, we consider the similarity in both coordinate space and feature space. The box coordinate and box feature is obtained by averaging the points' coordinates and features within them. The similarity between the points and the 3D bounding boxes can be formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.603, + 0.576, + 0.907, + 0.595 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {s} _ {p, B} = \\cos \\left(\\boldsymbol {f} _ {p}, \\boldsymbol {f} _ {B}\\right) \\times e ^ {- \\left| \\boldsymbol {c} _ {B} \\boldsymbol {c} _ {p} \\right|} \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.596, + 0.906, + 0.656 + ], + "angle": 0, + "content": "where \\(\\cos (\\cdot ,\\cdot)\\) denotes cosine similarity. In addition, \\(f_{p}\\) and \\(f_{B}\\) mean the features of point and box, while \\(c_{p}\\) and \\(c_{B}\\) mean the coordinates of point and box. In this way, it is expected to filter background points with low similarity." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.657, + 0.909, + 0.883 + ], + "angle": 0, + "content": "For the points lying in the overlapped parts of multiple 3D bounding boxes (usually two boxes, see boxes \\(B_{1}\\) and \\(B_{2}\\) in Fig. 3), they are assigned to the corresponding instances by learning the similarity between the points and the bounding boxes. However, due to the uncertainty of point labels in overlapping areas, it is difficult to obtain high-confidence semantic labels for these points from the bounding boxes. Therefore, we consider removing the overlapping points within each box and using the remaining reliable points (see \\(B_{1}^{rel}\\) and \\(B_{2}^{rel}\\) in Fig. 3) for similarity learning. Since reliable points are only located within one box, their labels are consistent with the box labels. Given a point feature \\(\\pmb{f}_{p}\\), box features \\(\\pmb{f}_{B_1^{rel}}\\) and \\(\\pmb{f}_{B_2^{rel}}\\), we utilize a multilayer perceptron (MLP) network to learn the similarity score \\(A_{p2b}\\) which is given by:" + }, + { + "type": "equation", + "bbox": [ + 0.566, + 0.885, + 0.908, + 0.906 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {A} _ {p 2 b} = \\mathrm {M L P} (\\boldsymbol {f} _ {p}, \\boldsymbol {f} _ {p} - \\boldsymbol {f} _ {B _ {1} ^ {r e l}}, \\boldsymbol {f} _ {p} - \\boldsymbol {f} _ {B _ {2} ^ {r e l}}) \\qquad (4)\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8882" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.089, + 0.482, + 0.315 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.322, + 0.485, + 0.381 + ], + "angle": 0, + "content": "Figure 4. The details of Multi-level Attention Block. The instance queries interact with the features of the whole scene, the coarse instance regions, and the instance core regions hierarchically. And the \\( B_{mask} \\) are the boxes obtained from the predicted masks." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.393, + 0.484, + 0.499 + ], + "angle": 0, + "content": "where \\( \\pmb{f}_{p} - \\pmb{f}_{B_{1}^{rel}} \\) and \\( \\pmb{f}_{p} - \\pmb{f}_{B_{2}^{rel}} \\) capture the difference between the point and boxes. In addition, the box feature is obtained by averaging the point features within it. Note that before computing the box feature, we eliminate the background points in the box through Eq. (3). To supervise the similarity learning, we use the cross-entropy loss to formulate pseudo label loss \\( L_{pl} \\), which is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.221, + 0.501, + 0.484, + 0.518 + ], + "angle": 0, + "content": "\\[\nL _ {p l} = L _ {C E} (\\mathbf {A}, \\mathbf {Y}) \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.519, + 0.484, + 0.549 + ], + "angle": 0, + "content": "where \\(\\pmb{Y}\\) is the label of the reliable point, which denotes the box it belongs to." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.55, + 0.484, + 0.64 + ], + "angle": 0, + "content": "After that, we use the model trained on the reliable points (see \\( B_1^{rel} \\) and \\( B_2^{rel} \\) in Fig. 3) to predict the label of unreliable points (see \\( B_1^{un} \\) and \\( B_2^{un} \\) in Fig. 3). It is desired that the points located in the overlapping part of the box can be assigned to the corresponding instances, obtaining pseudo instance labels." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.652, + 0.415, + 0.669 + ], + "angle": 0, + "content": "3.4. Coarse-to-Fine Instance Segmentator" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.674, + 0.484, + 0.809 + ], + "angle": 0, + "content": "The bounding box supervision provides coarse shape information for 3D instance segmentation. To obtain well-detailed instances, we present a coarse-to-fine instance segmentator that can recognize and segment the points in the point cloud into various instances. Specifically, we first employ the queries to learn from the whole point cloud to obtain coarse instances, and then these queries learn from the regions indicated by the coarse instances with Multi-level Attention Blocks to refine the instances." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Coarse Instances Segmentation. Supposing one scene consists of \\( q \\) objects, and each query represents one object and models the instance-related characteristics: the semantic category and the mask of each instance. In our method, these queries first interact with the backbone features of the whole point cloud to obtain coarse instances. Considering" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.181 + ], + "angle": 0, + "content": "the point-wise masks of the coarse instances can not describe one object holistically and precisely, we predict another two boxes with coarse granularity and fine granularity, respectively. Then these two boxes combined with the scene-level features to refine the initial coarse instances with Multi-level Attention Blocks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.183, + 0.909, + 0.303 + ], + "angle": 0, + "content": "Hierarchical Instances Refinement. To leverage the advantages of multi-granularity features from the point cloud and the segmented coarse instances, we introduce a novel decoder structure utilizing attentions to make the instance queries interact with granularity-varied features hierarchically. The decoder consists of a stack of six blocks. Each block includes a Multi-level Attention Block and an instance segmentation head illustrated in Fig. 3." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.303, + 0.909, + 0.498 + ], + "angle": 0, + "content": "Multi-level Attention Block. The details of the Multi-level Attention Block are shown in Fig. 4. These instance queries first interact with point-wise backbone features globally, then interact with regions described by the coarse instance locally. The regions described by the coarse instance are divided into reliable and unreliable, which are depicted by the predicted box and the core box, respectively. To fully explore various granularity of shape information implied by the coarse instances and those two boxes, we first abstract the obtained coarse instance into a box, termed as \\( B_{mask} \\). Then we explore the consistent information between \\( B_{mask} \\) and the predicted box \\( B_{pred} \\) to obtain features of reliable regions, which are described as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.578, + 0.499, + 0.907, + 0.518 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {F} ^ {r e l} = \\sigma \\left(\\boldsymbol {F}, \\boldsymbol {M} \\odot e ^ {I o U \\left(\\boldsymbol {B} _ {p r e d}, \\boldsymbol {B} _ {m a s k}\\right)}\\right) \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.52, + 0.909, + 0.713 + ], + "angle": 0, + "content": "where \\(\\sigma\\) denotes the thresholding operation and \\(\\odot\\) denotes point-wise multiplication. In addition, \\(F\\) denotes the backbone features, \\(M\\) denotes the predicted mask probabilities. \\(B_{pred}\\) and \\(B_{mask}\\) denote the predicted boxes and boxes decided by predicted masks, respectively. So far, the instance queries have learned from the entire point cloud and the reliable regions, they can segment the majority of objects with easy contexts. To tackle the complex contexts, the instance queries learn from the unreliable regions described by the predicted core boxes. The predicted core boxes are scaled from the predicted boxes, aiming to represent the core regions of objects, and corresponding features are obtained as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.58, + 0.718, + 0.907, + 0.734 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {F} ^ {u n} = \\sigma (\\boldsymbol {F}, \\boldsymbol {M} \\odot (\\boldsymbol {B} _ {\\text {c o r e}} \\cap \\boldsymbol {B} _ {\\text {m a s k}})) \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.736, + 0.908, + 0.812 + ], + "angle": 0, + "content": "where \\( B_{core} \\) denotes the predicted core boxes. The instance queries have strong discriminative ability in such a global-to-local manner, then they are fed into a self-attention module and a feed-forward network to aggregate the learned information, which further boosts the performance." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.818, + 0.842, + 0.835 + ], + "angle": 0, + "content": "3.5. Bilateral Matching and Training Loss" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Our method makes bilateral matching between pseudo instances and predicted instances first and then utilizes the paired pseudo instances as ground truth \\(\\mathbf{y}_{ins}\\) to supervise the paired predicted instances \\(\\mathbf{x}_{ins}\\)." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8883" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.091, + 0.485, + 0.196 + ], + "angle": 0, + "content": "Bilateral matching. Following SPFormer [36] and IS-BNet [30], our method utilizes the Hungarian method [18] to achieve the optimal assignment. We compute a pairwise matching cost \\( \\mathbf{c}_{i,j} \\) to evaluate the similarity of the \\( i^{th} \\) predicted instance and the \\( j^{th} \\) pseudo ground truth. And \\( \\mathbf{c}^{i,j} \\) is based on the classification probabilities and the binary masks, it is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.093, + 0.198, + 0.483, + 0.25 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathbf {c} ^ {i, j} = \\lambda_ {1} L _ {C E} (\\mathbf {x} _ {c l s} ^ {i}, \\mathbf {y} _ {c l s} ^ {j}) \\\\ + \\lambda_ {2} \\left(L _ {B C E} \\left(\\mathbf {x} _ {\\text {m a s k}} ^ {i}, \\mathbf {y} _ {\\text {m a s k}} ^ {j}\\right) + L _ {\\text {d i c e}} \\left(\\mathbf {x} _ {\\text {m a s k}} ^ {i}, \\mathbf {y} _ {\\text {m a s k}} ^ {j}\\right)\\right) \\tag {8} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.251, + 0.484, + 0.297 + ], + "angle": 0, + "content": "Training loss. Our training loss consists of two items, namely the loss of pseudo label generation \\( L_{pl} \\) and the loss of instance segmentation \\( L_{seg} \\), which is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.231, + 0.299, + 0.483, + 0.315 + ], + "angle": 0, + "content": "\\[\nL = L _ {p l} + L _ {s e g} \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.317, + 0.485, + 0.422 + ], + "angle": 0, + "content": "As for the instance segmentation, except for supervising the predicted semantic classes and the binary masks, we also supervise the predicted bounding boxes and the core boxes utilized in the coarse-to-fine instance segmentator. The ground truth of the scale of the core box is based on the IoU score of the predicted masks and the pseudo-ground truth masks. The loss of instance segmentation is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.099, + 0.424, + 0.483, + 0.49 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} L _ {s e g} = \\lambda_ {1} L _ {C E} (\\mathbf {x} _ {c l s}, \\mathbf {y} _ {c l s}) \\\\ + \\lambda_ {2} \\left(L _ {B C E} \\left(\\mathbf {x} _ {\\text {m a s k}}, \\mathbf {y} _ {\\text {m a s k}}\\right) + L _ {\\text {d i c e}} \\left(\\mathbf {x} _ {\\text {m a s k}}, \\mathbf {y} _ {\\text {m a s k}}\\right)\\right) \\\\ + \\lambda_ {3} \\left(L _ {L _ {1}} \\left(\\mathbf {x} _ {\\text {b o x}}, \\mathbf {y} _ {\\text {b o x}}\\right) + L _ {M S E} \\left(\\mathbf {x} _ {\\text {c o r e - b o x}}, \\mathbf {y} _ {\\text {c o r e - b o x}}\\right)\\right) \\tag {10} \\\\ \\end{array}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.506, + 0.224, + 0.523 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.531, + 0.282, + 0.547 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.553, + 0.484, + 0.704 + ], + "angle": 0, + "content": "Datasets. We conduct experiments on two datasets, ScanNetV2 [6] and S3DIS [1]. The ScanNetV2 provides splits containing 1201, 312, and 100 scenes for training, validation, and testing. Scenes are annotated with semantic labels covering 18 categories and instance labels. And the S3DIS dataset is another indoor dataset contains 272 scenes across six different areas, and it is annotated with 13 object categories. Following existing methods [4, 27, 29], the scenes of S3DIS Area 5 are used for validation while the others are used for training." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.484, + 0.825 + ], + "angle": 0, + "content": "Evaluation metrics. We use the mean average precision \\((AP)\\) as the evaluation metric for both ScanNetV2 and S3DIS, which is utilized as a common evaluation metric in 3D instance segmentation. \\(AP\\) is the averaged score with IoU thresholds from \\(95\\%\\) to \\(50\\%\\) with a step size of \\(5\\%\\). The mean average precision with IoU thresholds at \\(50\\%\\) and \\(25\\%\\) are used as evaluation metrics as well, which are denoted as \\(AP_{50}\\) and \\(AP_{25}\\), respectively." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.485, + 0.902 + ], + "angle": 0, + "content": "Implementation details. Our model is trained on a single RTX 3090 GPU, and we utilize SPFormer [36] and ISBNet [30] as codebases for ScanNetV2 and S3DIS, respectively. We utilize the AdamW optimizer with a learning rate of 0.0002 and weight decay at 0.05 for the training of Scan-" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.09, + 0.907, + 0.131 + ], + "angle": 0, + "content": "Table 1. Comparison on ScanNetV2. \"Mask\" denotes using dense point-wise annotations, \\( S_0 \\) and \\( S_1 \\) denote using accurate bounding box and sketchy bounding box annotations, respectively." + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.133, + 0.905, + 0.383 + ], + "angle": 0, + "content": "
MethodSup.VenueScanNet TestScanNet Val
AP50AP25AP50AP25
GraphCut [13]NIPS2273.283.269.179.3
ISBNet [30]CVPR2376.384.573.182.5
SPFormer [36]MaskAAAI2377.085.173.982.9
Mask3D [34]ICRA2378.087.073.783.5
MSTA3D [40]ACM MM2479.587.977.085.4
Spherical Mask [35]CVPR2481.2-79.988.2
Box2Mask [4]ECCV2267.780.359.771.8
WISGP+PointGroup [8]WACV23--50.264.9
WISGP+SSTNet [8]WACV23--56.970.2
GaPro+SPFormer [29]S0ICCV2369.282.470.479.9
CIP-WPIS [49]WACV24--69.378.6
BSNet+SPFormer [27]CVPR24--72.783.4
Sketchy-3DIS (ours)-70.186.668.883.6
Box2Mask [4]ECCV22--52.467.5
GaPro+SPFormer [29]S1ICCV23--53.572.2
Sketchy-3DIS (ours)---65.883.1
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.4, + 0.907, + 0.475 + ], + "angle": 0, + "content": "NetV2. As for the weights of matching cost and losses, \\((\\lambda_{1},\\lambda_{2},\\lambda_{3})\\) are set as (0.5, 1.0, 0.5). All the models are trained from scratch, and without specification, the hyperparameters and the training details are kept the same as the [36] or [30]." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.485, + 0.654, + 0.5 + ], + "angle": 0, + "content": "4.2. Main Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.508, + 0.909, + 0.689 + ], + "angle": 0, + "content": "ScanNetV2. In Table 1, we compare our method and existing state-of-the-art methods on ScanNetV2 dataset. Attributed to the innovative design of Sketchy-3DIS, which achieves the pseudo labeling and instance segmenting simultaneously, our method achieves considerable performance on both the validation set and online test set of ScanNetV2. Notably, without multi-modal priors [49] or synthesized scenes [27], our method outperforms existing methods in terms of \\(AP_{25}\\), which demonstrates the effectiveness of our method. Additionally, the comparable performances to fully-supervised methods demonstrate that our method can play a significant role in autonomous driving with high efficacy." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.909, + 0.81 + ], + "angle": 0, + "content": "S3DIS. In Table 2, we make comparisons between existing methods and our proposed Sketchy-3DIS on S3DIS Area 5. The consistent advantages demonstrate the strong generalization ability of our proposed SKeychy-3DIS. Notably, our method exceeds the fully-supervised baseline [30] with \\(+3.3\\) in \\(AP_{50}\\), which verifies that our proposed Sketchy-3DIS can make full use of instance indications provided by the sketchy bounding boxes." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Quality comparisons of pseudo labels. We compare the qualities of the generated pseudo labels of GaPro [29] and ours under \\( S_{1} \\) sketchy bounding boxes. The qualitative comparisons are present in Fig.5, which illustrates that our method generates more accurate pseudo labels than GaPro not only in details (highlighted by the green cycles) but also" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8884" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.176, + 0.094, + 0.284, + 0.108 + ], + "angle": 0, + "content": "Input Point Cloud" + }, + { + "type": "image_caption", + "bbox": [ + 0.366, + 0.094, + 0.449, + 0.107 + ], + "angle": 0, + "content": "Ground Truth" + }, + { + "type": "image_caption", + "bbox": [ + 0.522, + 0.094, + 0.646, + 0.108 + ], + "angle": 0, + "content": "Sketchy-3DIS (ours)" + }, + { + "type": "image_caption", + "bbox": [ + 0.745, + 0.094, + 0.785, + 0.107 + ], + "angle": 0, + "content": "GaPro" + }, + { + "type": "image", + "bbox": [ + 0.141, + 0.116, + 0.315, + 0.197 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.318, + 0.117, + 0.495, + 0.199 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.497, + 0.117, + 0.673, + 0.199 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.676, + 0.117, + 0.853, + 0.199 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.142, + 0.199, + 0.316, + 0.305 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.318, + 0.2, + 0.493, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.497, + 0.2, + 0.672, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.676, + 0.2, + 0.851, + 0.305 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.308, + 0.908, + 0.337 + ], + "angle": 0, + "content": "Figure 5. Visualization comparison of pseudo labels on the ScanNetV2 training set. The black denotes the background points and other colors denote different objects. The green and red cycles highlight the key regions." + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.348, + 0.485, + 0.404 + ], + "angle": 0, + "content": "Table 2. Comparison on S3DIS Area 5. \"Mask\" denotes using dense point-wise annotations, \\(S_0\\) and \\(S_1\\) denote using accurate bounding box and sketchy bounding box annotations, respectively. \\*\\*\" denotes the results are reproduced by GaPro [29]." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.405, + 0.484, + 0.627 + ], + "angle": 0, + "content": "
MethodSup.VenueAPAP50
SoftGroup [41]CVPR2251.666.1
GraphCut [13]NIPS2254.166.4
ISBNet [30]MaskCVPR2354.065.8
Mask3D [34]ICRA2356.668.4
MSTA3D [40]ACM MM24-70.0
Spherical Mask [35]CVPR2460.572.3
Box2Mask* [4]ECCV2243.654.6
WISGP+PointGroup [8]WACV2333.548.6
WISGP+SSTNet [8]S0WACV2337.251.0
GaPro+ISBNet [29]ICCV2350.561.2
BSNet+ISBNet [27]CVPR2453.064.3
Sketchy-3DIS (ours)-53.469.1
Sketchy-3DIS (ours)S1-50.764.6
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.643, + 0.483, + 0.702 + ], + "angle": 0, + "content": "in distinguishing the objects belonging to the same categories (highlighted by the red cycles). Additionally, compared with the ground truth, our method generates comparable pseudo labels as well." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.713, + 0.358, + 0.729 + ], + "angle": 0, + "content": "4.3. Ablation Studies and Analysis" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.484, + 0.794 + ], + "angle": 0, + "content": "We conduct comprehensive ablation studies to justify the effective design of our method. All these ablation experiments are conducted on the ScanNetV2 validation set unless otherwise stated." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Robustness under various sketchy bounding boxes. In Table 3, we evaluate our method under four kinds of various sketchy bounding boxes illustrated in Fig.2 on both ScanNetV2 validation set and S3DIS Area 5. It can be observed that the performance of our method deteriorates slightly as the increasing of the boxes' sketchy degree, which verifies the robustness of our proposed Sketchy-3DIS." + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.348, + 0.905, + 0.377 + ], + "angle": 0, + "content": "Table 3. Robustness under various sketchy bounding boxes. \\( S_0 \\) to \\( S_4 \\) denote the sketchy degree of the annotated bounding boxes." + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.378, + 0.905, + 0.485 + ], + "angle": 0, + "content": "
Sketchy DegreeScanNetV2 ValS3DIS Area 5
AP50AP25AP50AP25
S462.580.150.963.6
S362.981.452.366.6
S263.782.161.272.7
S165.883.164.673.9
S068.883.669.177.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.498, + 0.908, + 0.568 + ], + "angle": 0, + "content": "Table 4. Effectiveness of pseudo labeler and instance segmentator. UnScene3D [32] is an sota unsupervised method. \"disjoint\" denotes training the 3D UNet and pseudo labeler first and then freeze them to train instance segmentator while \"joint\" denotes training the pseudo labeler and the instance segmentator jointly." + }, + { + "type": "table", + "bbox": [ + 0.515, + 0.568, + 0.905, + 0.645 + ], + "angle": 0, + "content": "
MethodAPAP50AP25
UnScene3D [32]21.440.352.6
SKetchy-3DIS (disjoint)45.360.470.0
SKetchy-3DIS (joint)53.469.177.5
" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.659, + 0.907, + 0.809 + ], + "angle": 0, + "content": "Effectiveness of labeler and segmentator. In Table 4, experiments are conducted on the S3DIS Area 5 to verify the effectiveness of our designed adaptive box-to-point pseudo labeler and coarse-to-fine instance segmentator. We choose the state-of-the-art unsupervised method UnScene3D [32] for 3D instance segmentation as the baseline. The experimental results verify that our designed pseudo labeler can generate high-quality pseudo labels, and combined with our designed instance segmentator, our method achieves significant performance." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.811, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Effectiveness of components in pseudo labeler. In Table 5, we evaluate the effectiveness of various components in our designed pseudo labeler. The first row is the results of the state-of-the-art unsupervised method [32], and the comparison between the last two rows verifies that the learned point-to-box similarity and point-to-instance assignments" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8885" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.486, + 0.159 + ], + "angle": 0, + "content": "Table 5. Effectiveness of components in pseudo labeler. \"Partition\" denotes obtaining point-wise labels by box-to-point spatial correlations, \"Assign\" denotes assigning the points within overlapped boxes, and \"Similarity\" denotes filtering the background points within boxes according to the point-to-box similarities." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.159, + 0.485, + 0.251 + ], + "angle": 0, + "content": "
PartitionAssignSimilarityAPAP50AP25
15.932.258.5
✓41.864.872.3
✓✓45.267.383.4
✓✓✓46.068.883.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.089, + 0.265, + 0.485, + 0.335 + ], + "angle": 0, + "content": "Table 6. Attentions in Multi-level Attention Block. \"Scene\", \"Coarse\", and \"Core\" denote interacting with the features of the whole scene, the coarse instance regions, and the core regions of the instance, respectively. And \"Self\" denotes augmenting by the learned instance query itself." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.335, + 0.484, + 0.427 + ], + "angle": 0, + "content": "
SceneCoarseCoreSelfAPAP50AP25
41.364.179.8
✓✓44.667.082.1
✓✓✓44.767.182.9
✓✓✓✓46.068.883.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.441, + 0.485, + 0.483 + ], + "angle": 0, + "content": "Table 7. Effectiveness of instance segmentation losses. \\( L_{box} \\), \\( L_{core-box} \\), and \\( L_{mask} \\) denote the supervisions of the predicted instance box, instance core box, and instance mask, respectively." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.483, + 0.485, + 0.581 + ], + "angle": 0, + "content": "
LboxLcore-boxLmaskAPAP50AP25
✓45.266.882.3
✓✓45.467.783.0
✓✓45.067.883.0
✓✓✓46.068.883.6
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.596, + 0.486, + 0.64 + ], + "angle": 0, + "content": "can improve the performance by filtering the background points within boxes and assigning the points within overlapped boxes." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.642, + 0.486, + 0.779 + ], + "angle": 0, + "content": "Effectiveness of various attentions in Multi-level Attention Block. In Table 6, we conduct ablation studies on various combinations of attention layers utilized in Multi-level Attention Block. The best performance is obtained with the combination of the four attention modules. These experimental results verify that our proposed Multi-level Attention Block can explore the local contexts adaptively and aggregates them, which boosts the segmentation performance progressively." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.779, + 0.485, + 0.87 + ], + "angle": 0, + "content": "Effectiveness of instance segmentation losses. In Table 7, we compare the performances of the losses of the predicted boxes, core boxes, and masks. The results show that the best performance is obtained with the combination of the above three kinds of losses, combined with our coarse-to-fine instance segmentator, they achieve inspiring performance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.485, + 0.901 + ], + "angle": 0, + "content": "Qualitative results of 3D instance segmentation. Some visualized samples of predicted instances for the ScanNetV2" + }, + { + "type": "image", + "bbox": [ + 0.515, + 0.094, + 0.898, + 0.337 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.345, + 0.908, + 0.389 + ], + "angle": 0, + "content": "Figure 6. Qualitative results on ScanNetV2 validation set. The black denotes the background points and other colors denote different objects. The cycles highlight the key regions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.401, + 0.909, + 0.552 + ], + "angle": 0, + "content": "validation set are shown in Fig.6. It can be observed that our method shows consistent effectiveness in segmenting objects for easy and complex scenes. Even if there are many objects belonging to the same category and they are close to each other, our method demonstrates considerable performance with the sketchy bounding box annotation. Notably, our method can adaptively segment the objects that are ignored in the ground truth such as the cabinets highlighted in red cycles, which demonstrates the strong recognition capability of our method." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.565, + 0.634, + 0.58 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.591, + 0.909, + 0.788 + ], + "angle": 0, + "content": "In this work, we propose a query-based framework for weakly supervised 3D instance segmentation, Sketchy-3DIS, which is more robust and effective than existing bounding-box supervised methods. Sketchy-3DIS adaptively generates pseudo per-point labels from sketchy boxes and predicts instances in a coarse-to-fine manner. Extensive experiments on ScanNetV2 and S3DIS benchmarks show that our method achieves leading performance, and even outperforms some fully-supervised methods. However, our explored experiments demonstrate that the performance of our method degraded severely when the annotated sketchy bounding boxes are immensely inaccurate, exploration for this issue may be the future study orientation." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.801, + 0.673, + 0.817 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.909, + 0.9 + ], + "angle": 0, + "content": "This work was supported by the National Key R&D Program of China No. 2024YFC3015801, National Science Fund of China under Grant Nos. 62361166670, 62276144, U24A20330, 62306238 and the Fundamental Research Funds for the Central Universities." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8886" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.116, + 0.484, + 0.169 + ], + "angle": 0, + "content": "[1] Iro Armeni, Ozan Sener, Amir R Zamir, Helen Jiang, Ioannis Brilakis, Martin Fischer, and Silvio Savarese. 3D semantic parsing of large-scale indoor spaces. In CVPR, pages 1534-1543, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.172, + 0.484, + 0.213 + ], + "angle": 0, + "content": "[2] Shaoyu Chen, Jiemin Fang, Qian Zhang, Wenyu Liu, and Xinggang Wang. Hierarchical aggregation for 3D instance segmentation. In CVPR, pages 15467-15476, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.215, + 0.484, + 0.268 + ], + "angle": 0, + "content": "[3] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In CVPR, pages 1290–1299, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.271, + 0.484, + 0.325 + ], + "angle": 0, + "content": "[4] Julian Chibane, Francis Engelmann, Tuan Anh Tran, and Gerard Pons-Moll. Box2Mask: Weakly supervised 3D semantic instance segmentation using bounding boxes. In ECCV, pages 681-699. Springer, 2022. 1, 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.327, + 0.484, + 0.381 + ], + "angle": 0, + "content": "[5] Ruihang Chu, Xiaqing Ye, Zhengzhe Liu, Xiao Tan, Xiaojuan Qi, Chi-Wing Fu, and Jiaya Jia. Twist: Two-way inter-label self-training for semi-supervised 3D instance segmentation. In CVPR, pages 1100-1109, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.383, + 0.484, + 0.437 + ], + "angle": 0, + "content": "[6] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. ScanNet: Richly-annotated 3D reconstructions of indoor scenes. In CVPR, pages 5828-5839, 2017. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.439, + 0.484, + 0.492 + ], + "angle": 0, + "content": "[7] Shichao Dong, Ruibo Li, Jiacheng Wei, Fayao Liu, and Guosheng Lin. Rwseg: Cross-graph competing random walks for weakly supervised 3D instance segmentation. arXiv preprint arXiv:2208.05110, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.495, + 0.484, + 0.549 + ], + "angle": 0, + "content": "[8] Heming Du, Xin Yu, Farookh Hussain, Mohammad Ali Armin, Lars Petersson, and Weihao Li. Weakly-supervised point cloud instance segmentation with geometric priors. In WACV, pages 4271-4280, 2023. 1, 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.551, + 0.484, + 0.605 + ], + "angle": 0, + "content": "[9] Benjamin Graham, Martin Engelcke, and Laurens Van Der Maaten. 3D semantic segmentation with submanifold sparse convolutional networks. In CVPR, pages 9224-9232, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.607, + 0.484, + 0.66 + ], + "angle": 0, + "content": "[10] Tong He, Chunhua Shen, and Anton Van Den Hengel. Dyco3D: Robust instance segmentation of 3D point clouds through dynamic convolution. In CVPR, pages 354-363, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.663, + 0.484, + 0.704 + ], + "angle": 0, + "content": "[11] Ji Hou, Angela Dai, and Matthias Nießner. 3D-SIS: 3D semantic instance segmentation of rgb-d scans. In CVPR, pages 4421-4430, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.706, + 0.484, + 0.759 + ], + "angle": 0, + "content": "[12] Ji Hou, Benjamin Graham, Matthias Nießner, and Saining Xie. Exploring data-efficient 3D scene understanding with contrastive scene contexts. In CVPR, pages 15587-15597, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.762, + 0.484, + 0.802 + ], + "angle": 0, + "content": "[13] Le Hui, Linghua Tang, Yaqi Shen, Jin Xie, and Jian Yang. Learning superpoint graph cut for 3D instance segmentation. NIPS, 35:36804-36817, 2022. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.804, + 0.484, + 0.857 + ], + "angle": 0, + "content": "[14] Li Jiang, Hengshuang Zhao, Shaoshuai Shi, Shu Liu, ChiWing Fu, and Jiaya Jia. Pointgroup: Dual-set point grouping for 3D instance segmentation. In CVPR, pages 4867-4876, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.86, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[15] Lei Ke, Martin Danelljan, Henghui Ding, Yu-Wing Tai, Chi-Keung Tang, and Fisher Yu. Mask-free video instance segmentation. In CVPR, pages 22857–22866, 2023. 3" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.116, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.147 + ], + "angle": 0, + "content": "[16] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In ICCV, pages 4015-4026, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.149, + 0.908, + 0.203 + ], + "angle": 0, + "content": "[17] Maxim Kolodiazhnyi, Anna Vorontsova, Anton Konushin, and Danila Rukhovich. OneFormer3D: One transformer for unified point cloud segmentation. In CVPR, pages 20943-20953, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.205, + 0.908, + 0.245 + ], + "angle": 0, + "content": "[18] Harold W Kuhn. The hungarian method for the assignment problem. Naval research logistics quarterly, 2(1-2):83-97, 1955. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.247, + 0.908, + 0.299 + ], + "angle": 0, + "content": "[19] Xin Lai, Yuhui Yuan, Ruihang Chu, Yukang Chen, Han Hu, and Jiaya Jia. Mask-Attention-Free Transformer for 3D instance segmentation. In ICCV, pages 3693–3703, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.302, + 0.908, + 0.37 + ], + "angle": 0, + "content": "[20] Shiyi Lan, Zhiding Yu, Christopher Choy, Subhashree Radhakrishnan, Guilin Liu, Yuke Zhu, Larry S Davis, and Anima Anandkumar. DiscoBox: Weakly supervised instance segmentation and semantic correspondence from box supervision. In ICCV, pages 3406-3416, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.372, + 0.907, + 0.412 + ], + "angle": 0, + "content": "[21] Loic Landrieu and Martin Simonovsky. Large-scale point cloud semantic segmentation with superpoint graphs. In CVPR, pages 4558-4567, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.414, + 0.908, + 0.466 + ], + "angle": 0, + "content": "[22] Zhihao Liang, Zhihao Li, Songcen Xu, Mingkui Tan, and Kui Jia. Instance segmentation in 3D scenes using semantic superpoint tree networks. In ICCV, pages 2783-2792, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.469, + 0.908, + 0.523 + ], + "angle": 0, + "content": "[23] Yongbin Liao, Hongyuan Zhu, Yanggang Zhang, Chuangguan Ye, Tao Chen, and Jiayuan Fan. Point cloud instance segmentation with semi-supervised bounding-box mining. TPAMI, 44(12):10159-10170, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.525, + 0.908, + 0.579 + ], + "angle": 0, + "content": "[24] Zhengzhe Liu, Xiaojuan Qi, and Chi-Wing Fu. One thing one click: A self-training approach for weakly supervised 3D semantic segmentation. In CVPR, pages 1726-1736, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.581, + 0.908, + 0.622 + ], + "angle": 0, + "content": "[25] Ze Liu, Zheng Zhang, Yue Cao, Han Hu, and Xin Tong. Group-free 3D object detection via transformers. In ICCV, pages 2949-2958, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.624, + 0.907, + 0.663 + ], + "angle": 0, + "content": "[26] Jiahao Lu, Jiacheng Deng, Chuxin Wang, Jianfeng He, and Tianzhu Zhang. Query refinement transformer for 3D instance segmentation. In ICCV, pages 18516-18526, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.665, + 0.908, + 0.706 + ], + "angle": 0, + "content": "[27] Jiahao Lu, Jiacheng Deng, and Tianzhu Zhang. BSNet: Box-supervised simulation-assisted mean teacher for 3D instance segmentation. In CVPR, pages 20374-20384, 2024. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.707, + 0.908, + 0.748 + ], + "angle": 0, + "content": "[28] Ishan Misra, Rohit Girdhar, and Armand Joulin. An end-to-end transformer model for 3D object detection. In ICCV, pages 2906-2917, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.749, + 0.908, + 0.802 + ], + "angle": 0, + "content": "[29] Tuan Duc Ngo, Binh-Son Hua, and Khoi Nguyen. GaPro: Box-supervised 3D point cloud instance segmentation using gaussian processes as pseudo labelers. In ICCV, pages 17794-17803, 2023. 1, 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.804, + 0.908, + 0.858 + ], + "angle": 0, + "content": "[30] Tuan Duc Ngo, Binh-Son Hua, and Khoi Nguyen. ISBNet: a 3D point cloud instance segmentation network with instance-aware sampling and box-aware dynamic convolution. In CVPR, pages 13550-13559, 2023. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.908, + 0.901 + ], + "angle": 0, + "content": "[31] Charles R Qi, Or Litany, Kaiming He, and Leonidas J Guibas. Deep hough voting for 3D object detection in point clouds. In ICCV, pages 9277-9286, 2019. 1" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8887" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.135 + ], + "angle": 0, + "content": "[32] David Rozenberszki, Or Litany, and Angela Dai. UnScene3D: Unsupervised 3D instance segmentation for indoor scenes. In CVPR, pages 19957-19967, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.137, + 0.485, + 0.178 + ], + "angle": 0, + "content": "[33] Danila Rukhovich, Anna Vorontsova, and Anton Konushin. FCAF3D: Fully convolutional anchor-free 3D object detection. In ECCV, pages 477-493. Springer, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.181, + 0.484, + 0.235 + ], + "angle": 0, + "content": "[34] Jonas Schult, Francis Engelmann, Alexander Hermans, Or Litany, Siyu Tang, and Bastian Leibe. Mask3D: Mask transformer for 3D semantic instance segmentation. In ICRA, pages 8216-8223. IEEE, 2023. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.238, + 0.484, + 0.293 + ], + "angle": 0, + "content": "[35] Sangyun Shin, Kaichen Zhou, Madhu Vankadari, Andrew Markham, and Niki Trigoni. Spherical mask: Coarse-to-fine 3D point cloud instance segmentation with spherical representation. In CVPR, pages 4060-4069, 2024. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.296, + 0.484, + 0.337 + ], + "angle": 0, + "content": "[36] Jiahao Sun, Chunmei Qing, Junpeng Tan, and Xiangmin Xu. Superpoint transformer for 3D scene instance segmentation. In AAAI, pages 2393-2401, 2023. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.34, + 0.482, + 0.38 + ], + "angle": 0, + "content": "[37] Linghua Tang, Le Hui, and Jin Xie. Learning inter-superpoint affinity for weakly supervised 3D instance segmentation. In ACCV, pages 1282-1297, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.383, + 0.484, + 0.424 + ], + "angle": 0, + "content": "[38] An Tao, Yueqi Duan, Yi Wei, Jiwen Lu, and Jie Zhou. Seg-Group: Seg-level supervision for 3D instance and semantic segmentation. TIP, 31:4952-4965, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.427, + 0.484, + 0.468 + ], + "angle": 0, + "content": "[39] Zhi Tian, Chunhua Shen, Xinlong Wang, and Hao Chen. BoxInst: High-performance instance segmentation with box annotations. In CVPR, pages 5443-5452, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.471, + 0.484, + 0.512 + ], + "angle": 0, + "content": "[40] Duc Dang Trung Tran, Byeongkeun Kang, and Yeejin Lee. MSTA3D: Multi-scale twin-attention for 3d instance segmentation. In ACM MM, pages 1467-1475, 2024. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.514, + 0.482, + 0.555 + ], + "angle": 0, + "content": "[41] Thang Vu, Kookhoi Kim, Tung M Luu, Thanh Nguyen, and Chang D Yoo. SoftGroup for 3D instance segmentation on point clouds. In CVPR, pages 2708-2717, 2022. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.558, + 0.484, + 0.612 + ], + "angle": 0, + "content": "[42] Haiyang Wang, Shaocong Dong, Shaoshuai Shi, Aoxue Li, Jianan Li, Zhenguo Li, Liwei Wang, et al. CAGroup3D: Class-aware grouping for 3D object detection on point clouds. NIPS, 35:29975-29988, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.615, + 0.484, + 0.67 + ], + "angle": 0, + "content": "[43] Haiyang Wang, Shaoshuai Shi, Ze Yang, Rongyao Fang, Qi Qian, Hongsheng Li, Bernt Schiele, and Liwei Wang. RBGNet: Ray-based grouping for 3D object detection. In CVPR, pages 1110-1119, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.673, + 0.482, + 0.727 + ], + "angle": 0, + "content": "[44] Qian Xie, Yu-Kun Lai, Jing Wu, Zhoutao Wang, Yiming Zhang, Kai Xu, and Jun Wang. MLCVNet: Multi-level context votenet for 3D object detection. In CVPR, pages 10447-10456, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.731, + 0.484, + 0.786 + ], + "angle": 0, + "content": "[45] Saining Xie, Jiatao Gu, Demi Guo, Charles R Qi, Leonidas Guibas, and Or Litany. PointContrast: Unsupervised pretraining for 3D point cloud understanding. In ECCV, pages 574-591. Springer, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.788, + 0.484, + 0.842 + ], + "angle": 0, + "content": "[46] Bo Yang, Jianan Wang, Ronald Clark, Qingyong Hu, Sen Wang, Andrew Markham, and Niki Trigoni. Learning object bounding boxes for 3D instance segmentation on point clouds. NIPS, 32, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[47] Cheng-Kun Yang, Min-Hung Chen, Yung-Yu Chuang, and Yen-Yu Lin. 2D-3D interlaced transformer for point cloud segmentation with scene-level supervision. In ICCV, pages 977–987, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.092, + 0.908, + 0.147 + ], + "angle": 0, + "content": "[48] Li Yi, Wang Zhao, He Wang, Minhyuk Sung, and Leonidas J Guibas. GSPN: Generative shape proposal network for 3D instance segmentation in point cloud. In CVPR, pages 3947-3956, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.149, + 0.908, + 0.204 + ], + "angle": 0, + "content": "[49] Qingtao Yu, Heming Du, Chen Liu, and Xin Yu. When 3D bounding-box meets SAM: Point cloud instance segmentation with weak-and-noisy supervision. In WACV, pages 3719-3728, 2024. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.206, + 0.908, + 0.248 + ], + "angle": 0, + "content": "[50] Zaiwei Zhang, Bo Sun, Haitao Yang, and Qixing Huang. H3DNet: 3D object detection using hybrid geometric primitives. In ECCV, pages 311-329. Springer, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.25, + 0.908, + 0.29 + ], + "angle": 0, + "content": "[51] Yu Zheng, Yueqi Duan, Jiwen Lu, Jie Zhou, and Qi Tian. HyperDet3D: Learning a scene-conditioned 3D object detector. In CVPR, pages 5585-5594, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.292, + 0.908, + 0.346 + ], + "angle": 0, + "content": "[52] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable DETR: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159, 2020. 2" + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.092, + 0.908, + 0.346 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8888" + } + ] +] \ No newline at end of file diff --git a/2025/Sketchy Bounding-box Supervision for 3D Instance Segmentation/fed9c100-3d53-4c29-809d-ba6f2288846b_origin.pdf b/2025/Sketchy Bounding-box Supervision for 3D Instance Segmentation/fed9c100-3d53-4c29-809d-ba6f2288846b_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..e39f0f9c18d1ffe1f8acec05c872ac98a662303a --- /dev/null +++ b/2025/Sketchy Bounding-box Supervision for 3D Instance Segmentation/fed9c100-3d53-4c29-809d-ba6f2288846b_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9cadf52a79a2ba0fdb8ba79376345d10f330520784cfce66681aff1d95dd8224 +size 5794780 diff --git a/2025/Sketchy Bounding-box Supervision for 3D Instance Segmentation/full.md b/2025/Sketchy Bounding-box Supervision for 3D Instance Segmentation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..399976fe6f63408eb4e3a020446f42e223e43faa --- /dev/null +++ b/2025/Sketchy Bounding-box Supervision for 3D Instance Segmentation/full.md @@ -0,0 +1,362 @@ +# Sketchy Bounding-box Supervision for 3D Instance Segmentation + +Qian Deng $^{1}$ , Le Hui $^{2}$ , Jin Xie $^{3,4*}$ , Jian Yang $^{1}$ + +1PCA Lab, College of Computer Science, Nankai University, Tianjin, China + +$^{2}$ School of Electronics and Information, Northwestern Polytechnical University, Xi'an, China + +$^{3}$ State Key Laboratory for Novel Software Technology, Nanjing University, Nanjing, China + +$^{4}$ School of Intelligence Science and Technology, Nanjing University, Suzhou, China + +dengqian@mail.nankai.edu.cn; huile@nwpu.edu.cn; csjxie@nju.edu.cn; csjyang@nankai.edu.cn + +# Abstract + +Bounding box supervision has gained considerable attention in weakly supervised 3D instance segmentation. While this approach alleviates the need for extensive point-level annotations, obtaining accurate bounding boxes in practical applications remains challenging. To this end, we explore the inaccurate bounding box, named sketchy bounding box, which is imitated through perturbing ground truth bounding box by adding scaling, translation, and rotation. In this paper, we propose Sketchy-3DIS, a novel weakly 3D instance segmentation framework, which jointly learns pseudo labeler and segmentator to improve the performance under the sketchy bounding-box supervisions. Specifically, we first propose an adaptive box-to-point pseudo labeler that adaptively learns to assign points located in the overlapped parts between two sketchy bounding boxes to the correct instance, resulting in compact and pure pseudo instance labels. Then, we present a coarse-to-fine instance segmentator that first predicts coarse instances from the entire point cloud and then learns fine instances based on the region of coarse instances. Finally, by using the pseudo instance labels to supervise the instance segmentator, we can gradually generate high-quality instances through joint training. Extensive experiments show that our method achieves state-of-the-art performance on both the ScanNetV2 and S3DIS benchmarks, and even outperforms several fully supervised methods using sketchy bounding boxes. Code is available at https://github.com/dengq7/Sketchy-3DIS. + +# 1. Introduction + +With the rise of 3D datasets and the growing need for 3D scene understanding, various 3D computer vision tasks [19, 25, 28-31, 33, 34, 36, 42-44, 50, 51] have gained significant attention from both academia and industry. Among these, 3D + +![](images/218af45a23c27b30bd28ac36739086053e911ca15466e8c63b7fbf0b852823f6.jpg) +(a) Inputs +Figure 1. (a) illustrates the inputs of bounding box supervised 3D instance segmentation. (b) compares the performances of GaPro [29] and our Sketchy-3DIS under both accurate and scaled sketchy boxes supervision on ScanNetV2 validation set. + +![](images/7f25aa8c2fdf6a9f9987cffb92ef4d9a91ff368a1460d4d3ffa3be0ebc415f25.jpg) +(b) Performance Comparisons + +instance segmentation is both fundamental and challenging, as it involves recognizing the category of each object within a point cloud and delineating its individual mask. Current 3D instance segmentation methods [2, 13, 14, 30, 34, 36, 41] have demonstrated impressive performance, primarily rely on data-hungry deep learning approaches that require substantial training data and dense point-level manual annotations. Notably, point-level fully supervised instance annotation in ScanNet takes approximately 22.3 minutes [6] per scene, while labeling a 3D box in ScanNet takes about 1.93 minutes [38] per scene. Therefore, bounding box annotation stands out as it provides substantial instance-level information while maintaining a reasonable annotation cost, making it a favorable annotation choice for weakly supervised 3D instance segmentation. + +Many weakly supervised methods have emerged to tackle the 3D instance segmentation task. Semi-supervised 3D instance segmentation methods [5, 12, 23, 45] use training data with partial labels, while others [5, 12, 23, 45] utilize indirect labels, such as 2D image annotations. Additionally, object-level annotations, like marking a single point [7, 24, 37, 38] or drawing a bounding box [4, 8, 29, 49] for each object, are employed. Bounding box supervision is another important + +annotation for weakly supervised 3D instance segmentation. Box2Mask [4] is the first attempt that uses bounding box supervision for 3D instance segmentation. Subsequently, the following methods [8, 27, 29, 49] use different constraints to improve performance, such as using local correlations, Gaussian processes, teacher-student model, SAM [16], respectively. Although impressive results have been achieved in weakly supervised 3D instance segmentation, these methods still rely on accurate and compact bounding box annotations. It is worth noting that in practice, accurate and compact boxes are difficult to obtain. Generally, annotated bounding box involves changes in scaling, translation, and rotation. As shown in Fig.1, we observed that the performance of GaPro [29] decreases significantly under scaled bounding box supervision. This motivated us to design a new approach to tackle with these inaccurate bounding box annotations. + +In this paper, we propose a weakly supervised 3D instance segmentation framework to tackle sketchy bounding box supervision. To simulate sketchy bounding boxes, we impose scaling, translation, and rotation on the ground truth bounding boxes to generate inaccurate bounding boxes. Our method consists of two key components: adaptive box-to-point pseudo labeler and coarse-to-fine instance segmentator. Given a point cloud scene and sketchy bounding boxes, we first use 3D U-Net [9] backbone to extract point features. After that, we feed the learned features and sketchy bounding boxes of the scene into the adaptive box-to-point pseudo labeler to generate high-quality point-level instance labels. For the points located in the overlapped region between two sketchy bounding boxes, we adaptively assign the points to the corresponding box by learning the similarity of points and boxes. For the remaining points, we directly assign them to the corresponding box according to the spatial correlations. In this way, we can transform the sketchy bounding box into a compact bounding box. We consider the labels of points that located within the generated compact bounding box as pseudo instance labels. To accurately predict instances, we input the learned features into the coarse-to-fine instance segmentator iteratively. We first detect the coarse bounding box and mask of the instance by executing queries on the whole point cloud. Based on the small region of the coarse bounding box, we re-predict the fine bounding box and mask by conducting queries within a more precise small region. Finally, by jointly optimizing the adaptive box-to-point pseudo labeler and coarse-to-fine instance segmentator, our method can gradually improve the performance of weakly supervised 3D instance segmentation using the sketchy bounding box annotations. Rich experiments on the ScanNetV2 and S3DIS show that our method have achieved state-of-the-art performance. + +Our contributions are as follows: + +- We are the first to explore the use of sketchy bounding box annotations for 3D instance segmentation, which allows + +for training with noise-tolerant, instance-size-agnostic 3D bounding boxes. + +- We propose a new weakly supervised 3D instance segmentation framework, which presents an adaptive box-to-point pseudo labeler and a coarse-to-fine instance segmentator for joint training. +- Our method achieves leading performance on both the ScanNetV2 and S3DIS datasets for weakly 3D instance segmentation under sketchy bounding box annotations. + +# 2. Related Work + +# 2.1. Fully Supervised 3D Instance Segmentation + +Existing fully supervised methods for 3D instance segmentation can be broadly categorized into three main pipelines: grouping-based [2, 13, 14, 22], detection-based [11, 41, 46, 48], and query-based [10, 17, 19, 26, 30, 34-36, 40] methods. Grouping-based methods typically group points based on predicted per-point semantic classes and instance center offsets, but the grouping rules are often manually defined by researchers, which lack flexibility. Detection-based methods, on the other hand, first predict the bounding box for each object and then generate the foreground mask within the box, treating the task as a two-step process of object detection followed by segmentation. More recently, inspired by DETR [52] and Mask2Former [3], query-based methods have been proposed for 3D instance segmentation. These methods simultaneously predict the semantic classes and instance masks of objects using a set of learnable query vectors, which model the objects' geometrical and semantical attributes integrally. Among these methods, query-based approaches tend to achieve superior performance, as the other two pipelines both face challenges such as error accumulation and time-consuming refinement processes. While these fully-supervised methods have made significant progress, they all rely on dense per-point annotations, and their performance tends to degrade when such annotations are unavailable. Therefore, several weakly-supervised methods have been proposed to alleviate this problem. + +# 2.2. Weakly Supervised 3D Instance Segmentation + +Weakly-supervised methods require partial or implicit annotations. [5, 12, 23, 45] use sparse point annotations and design various label propagation strategies. However, these sparse points cannot fully capture the spatial extent of an object. Approaches like MIT [47] and CIP-WPIS [49] leverage 2D images as additional input for 3D instance segmentation, but this introduces multi-modal dependencies. Other methods [7, 24, 37, 38] rely on instance-level signals, where labeled a single point for per instance, and instance predictions are made using techniques like random walk or clustering. Recently, most state-of-the-art methods [4, 8, 27, 29, 49] have adopted 3D bounding boxes as supervision, where a + +single bounding box is assigned to each object. These approaches naturally capture the spatial extent of an object and the annotated bounding boxes are easier to obtain compared to dense per-point annotations. + +Bounding-box Supervised 3D Instance Segmentation. Bounding boxes have become a widely used supervision signal, not only for 2D instance segmentation tasks [15, 20, 39] but also for 3D instance segmentation. Box2mask [4] is a pioneering approach that utilizes bounding-box supervision for 3D instance segmentation. WISGP [8] divides the point cloud into a determined set and an undetermined set based on explicit spatial correlations, then uses polygon meshes and superpoints to label the undetermined set. GaPro [29] focuses on generating reliable pseudo labels using a Gaussian process. Recently, BSNet [27] employs a simulation-assisted mean teacher framework to generate pseudo labels, enhancing data samples by synthesizing scenes that contain overlapping objects. However, these methods all rely on axis-aligned bounding boxes for supervision, and their performances heavily depend on the accuracy of these annotated boxes. To address this, CIP-WPIS [49] proposes using slightly looser bounding boxes, projecting the point cloud into 2D images, and obtaining 3D instance masks based on segmented 2D masks. While effective, it relies on 2D modality and can only adapt to slightly enlarged bounding boxes, limiting its flexibility. To alleviate the annotation costs as well as obtain robust and effective performance, we design a novel framework Sketchy-3DIS, which endeavors to not only face the common challenge of correctly assigning the points within multiple boxes but also be dedicated to correcting the errors brought by sketchy bounding boxes. + +# 3. Method + +# 3.1. Overview + +The overall framework of our weakly supervised 3D instance segmentation based on sketchy bounding boxes is depicted in Fig.3. It consists of two key components: adaptive box-to-point pseudo labeler and coarse-to-fine instance segmentator. + +Assuming a 3D point cloud $P \in \mathbb{R}^{N \times C}$ , where $N$ and $C$ represent the number of 3D points and channels, respectively. In the basic setting, each point is represented by the 3D coordinate, RGB color, and surface normal vector. In the training phase, we first feed the input point cloud and the sketchy bounding boxes (defined in Sec. 3.2) into the 3D UNet backbone network for feature extraction, then we average the point features located in the superpoint [21] to generate superpoint-level features. Next, we input the features and sketchy bounding boxes to the adaptive box-to-point pseudo labeler (Sec. 3.3) for converting the rough box-level annotation to the detailed point-level annotation of the instance. Meanwhile, we feed the feature into the coarse-to-fine instance segmentator (Sec. 3.4) to predict instances by + +![](images/e10580a19fbb3a04ba86032c1afdd3bd3777e80298757556110bd2e6953fd0cb.jpg) +Point Cloud + +![](images/5257c541f2a9726e29353ab9494e750389c4a39b4fbb27e31d5d89ad9b69a7da.jpg) + +![](images/a6100bc158c2fe0f3e4d6702eb96aca7989542341f4a1e5ae09a29cd08c5faa8.jpg) +Scaled $(S_{1})$ + +![](images/4d11b69675b686840c5a1c14f900dcfd7b4be15ac53e8c108c05e108bf51d0d9.jpg) +Scaled and translated $(S_{2})$ + +![](images/8c2da4509d4486ba0e3ca9f2060700859132a946787fd3678bb8496870e1b7ea.jpg) +Accurate $(S_0)$ +Scaled and rotated $(S_{3})$ + +![](images/6fa0b9391a2ca90bf0043c9a97ea3354a66b5a3a2d3a8137e4c4a74d0f513409.jpg) +Scaled, translated and rotated $(S_4)$ +Figure 2. Various sketchy bounding boxes under scaling, translation, and rotation perturbations. + +Transformer. Finally, we employ bilateral matching between the generated pseudo labels (regarded as the ground truth) and the predicted instances to build the correspondence between them for training (Sec. 3.5). In the inference phase, a new point cloud is fed to the backbone network and the coarse-to-fine instance segmentator to segment instances. + +# 3.2. Sketchy Bounding Box Setting + +Our method focuses on studying weakly supervised 3D instance segmentation based on the sketchy bounding box, i.e., inaccurate bounding box. In practice, it is unable to acquire compact 3D bounding box annotations for weakly supervised 3D instance segmentation. Therefore, we perturb the ground truth bounding box with scaling, translation and rotation to mimic the sketchy bounding boxes. As shown in Fig.2, we illustrate the generation process of different types of sketchy bounding boxes. + +Specifically, given the ground truth bounding box, which is denoted by the position of two corner points: + +$$ +\begin{array}{l} \mathbf {B} = \left[ B _ {\text {m i n}}, B _ {\text {m a x}} \right] \\ = \left[ x _ {\min }, y _ {\min }, z _ {\min }, x _ {\max }, y _ {\max }, z _ {\max } \right] \tag {1} \\ \end{array} +$$ + +First, we individually employ scaling, translation, and rotation on the ground truth bounding box to generate the basic inaccurate bounding boxes, which are formulated as: + +$$ +\mathbf {B} _ {\text {s c a l e d}} = \left[ B _ {\min } - \alpha E, B _ {\max } + \alpha E \right] +$$ + +$$ +\mathbf {B} _ {\text {t r a n s l a t e d}} = \left[ B _ {\min } + \beta E, B _ {\max } + \beta E \right] \tag {2} +$$ + +$$ +\mathbf {B} _ {\text {r o t a t e d}} = r \left(\left[ B _ {\min }, B _ {\max } \right], M E, \gamma \pi / 1 8 0\right) +$$ + +where $\alpha$ , $\beta$ , $\gamma$ are the parameters of different operations. Note that $E = B_{max} - B_{min}$ and $ME = (B_{max} + B_{min}) / 2$ . Considering that manually annotated 3D bounding boxes will not have significant differences from compact 3D ground truth bounding boxes. Therefore, we empirically set small values for those parameters. In experiments, we set $\alpha = 5\%$ , $\beta = 5\%$ , and $\gamma = 5$ , respectively. Then, we generate diverse sketchy bounding boxes by combining these basic operations $\mathbf{B}_{\mathrm{scaled}}$ , $\mathbf{B}_{\mathrm{translated}}$ , and $\mathbf{B}_{\mathrm{rotated}}$ . In Fig.2, we show four + +![](images/39042b8c3d03282cc29d73fb3162943ad0a2d1cbaa6ff6603b1476a41571a94c.jpg) +Figure 3. The framework of the Sketchy-3DIS. Given a point cloud with sketchy bounding-box annotations, we first extract the backbone features using a 3D U-Net backbone, then feed them into the adaptive box-to-point pseudo labeler and the coarse-to-fine instance segmentator, finally, we utilize the generated high quality pseudo labels to supervise the predicted instances periodically. + +![](images/be8a8c9f7ca40e4e866edadb8704c88d36dd40fa88b0619efb6acbf007580224.jpg) + +types of sketchy bounding boxes, named $S_{1}$ , $S_{2}$ , $S_{3}$ , and $S_{4}$ , the red rectangle denotes the accurate box and the blue rectangles denote various sketchy bounding boxes. It can be observed that the "sketchy degree" increases from $S_{1}$ to $S_{4}$ , making instance segmentation more challenging. + +# 3.3. Adaptive Box-to-Point Pseudo Labeler + +The sketchy 3D bounding box supervision is an inaccurate weak annotation for 3D instance segmentation. In order to convert rough annotated boxes into fine-grained point-level instance annotations, we propose an adaptive box-to-point pseudo labeler that can assign the points located in the box to the corresponding instance. Specifically, we first distinguish the points that lie in bounding boxes, and then adaptively assign them to the corresponding instance by learning the correlation between the points and 3D bounding boxes. + +Box Conditioned Points Partition. According to the relationship between points and 3D bounding boxes, there are three types of points: the points located outside the bounding boxes (i.e., background points), the points lying in a single bounding box, and the points belonging to multiple bounding boxes. Since the background points (see black circle points in Fig. 3) are not within any 3D bounding boxes, they are directly labeled as background. The point lying in a single or multiple bounding box could be a point on the target (see green/yellow/blue circle points in Fig. 3) or a background point. Therefore, we need other strategies to help distinguish the target points from the background points. + +Adaptive Point-to-Instance Assignment. For the points + +lying in a single 3D bounding box (see bounding box $B_{3}$ in Fig. 3), we filter the background points (i.e., black circle points in $B_{3}$ ) by comparing the similarity between the points and the 3D bounding box. Specifically, we consider the similarity in both coordinate space and feature space. The box coordinate and box feature is obtained by averaging the points' coordinates and features within them. The similarity between the points and the 3D bounding boxes can be formulated as follows: + +$$ +\boldsymbol {s} _ {p, B} = \cos \left(\boldsymbol {f} _ {p}, \boldsymbol {f} _ {B}\right) \times e ^ {- \left| \boldsymbol {c} _ {B} \boldsymbol {c} _ {p} \right|} \tag {3} +$$ + +where $\cos (\cdot ,\cdot)$ denotes cosine similarity. In addition, $f_{p}$ and $f_{B}$ mean the features of point and box, while $c_{p}$ and $c_{B}$ mean the coordinates of point and box. In this way, it is expected to filter background points with low similarity. + +For the points lying in the overlapped parts of multiple 3D bounding boxes (usually two boxes, see boxes $B_{1}$ and $B_{2}$ in Fig. 3), they are assigned to the corresponding instances by learning the similarity between the points and the bounding boxes. However, due to the uncertainty of point labels in overlapping areas, it is difficult to obtain high-confidence semantic labels for these points from the bounding boxes. Therefore, we consider removing the overlapping points within each box and using the remaining reliable points (see $B_{1}^{rel}$ and $B_{2}^{rel}$ in Fig. 3) for similarity learning. Since reliable points are only located within one box, their labels are consistent with the box labels. Given a point feature $\pmb{f}_{p}$ , box features $\pmb{f}_{B_1^{rel}}$ and $\pmb{f}_{B_2^{rel}}$ , we utilize a multilayer perceptron (MLP) network to learn the similarity score $A_{p2b}$ which is given by: + +$$ +\boldsymbol {A} _ {p 2 b} = \mathrm {M L P} (\boldsymbol {f} _ {p}, \boldsymbol {f} _ {p} - \boldsymbol {f} _ {B _ {1} ^ {r e l}}, \boldsymbol {f} _ {p} - \boldsymbol {f} _ {B _ {2} ^ {r e l}}) \qquad (4) +$$ + +![](images/78d5395f99d8466fd8903160a90303837105d32f65595a7eb5b1b017880c8fd2.jpg) +Figure 4. The details of Multi-level Attention Block. The instance queries interact with the features of the whole scene, the coarse instance regions, and the instance core regions hierarchically. And the $B_{mask}$ are the boxes obtained from the predicted masks. + +where $\pmb{f}_{p} - \pmb{f}_{B_{1}^{rel}}$ and $\pmb{f}_{p} - \pmb{f}_{B_{2}^{rel}}$ capture the difference between the point and boxes. In addition, the box feature is obtained by averaging the point features within it. Note that before computing the box feature, we eliminate the background points in the box through Eq. (3). To supervise the similarity learning, we use the cross-entropy loss to formulate pseudo label loss $L_{pl}$ , which is defined as: + +$$ +L _ {p l} = L _ {C E} (\mathbf {A}, \mathbf {Y}) \tag {5} +$$ + +where $\pmb{Y}$ is the label of the reliable point, which denotes the box it belongs to. + +After that, we use the model trained on the reliable points (see $B_1^{rel}$ and $B_2^{rel}$ in Fig. 3) to predict the label of unreliable points (see $B_1^{un}$ and $B_2^{un}$ in Fig. 3). It is desired that the points located in the overlapping part of the box can be assigned to the corresponding instances, obtaining pseudo instance labels. + +# 3.4. Coarse-to-Fine Instance Segmentator + +The bounding box supervision provides coarse shape information for 3D instance segmentation. To obtain well-detailed instances, we present a coarse-to-fine instance segmentator that can recognize and segment the points in the point cloud into various instances. Specifically, we first employ the queries to learn from the whole point cloud to obtain coarse instances, and then these queries learn from the regions indicated by the coarse instances with Multi-level Attention Blocks to refine the instances. + +Coarse Instances Segmentation. Supposing one scene consists of $q$ objects, and each query represents one object and models the instance-related characteristics: the semantic category and the mask of each instance. In our method, these queries first interact with the backbone features of the whole point cloud to obtain coarse instances. Considering + +the point-wise masks of the coarse instances can not describe one object holistically and precisely, we predict another two boxes with coarse granularity and fine granularity, respectively. Then these two boxes combined with the scene-level features to refine the initial coarse instances with Multi-level Attention Blocks. + +Hierarchical Instances Refinement. To leverage the advantages of multi-granularity features from the point cloud and the segmented coarse instances, we introduce a novel decoder structure utilizing attentions to make the instance queries interact with granularity-varied features hierarchically. The decoder consists of a stack of six blocks. Each block includes a Multi-level Attention Block and an instance segmentation head illustrated in Fig. 3. + +Multi-level Attention Block. The details of the Multi-level Attention Block are shown in Fig. 4. These instance queries first interact with point-wise backbone features globally, then interact with regions described by the coarse instance locally. The regions described by the coarse instance are divided into reliable and unreliable, which are depicted by the predicted box and the core box, respectively. To fully explore various granularity of shape information implied by the coarse instances and those two boxes, we first abstract the obtained coarse instance into a box, termed as $B_{mask}$ . Then we explore the consistent information between $B_{mask}$ and the predicted box $B_{pred}$ to obtain features of reliable regions, which are described as follows: + +$$ +\boldsymbol {F} ^ {r e l} = \sigma \left(\boldsymbol {F}, \boldsymbol {M} \odot e ^ {I o U \left(\boldsymbol {B} _ {p r e d}, \boldsymbol {B} _ {m a s k}\right)}\right) \tag {6} +$$ + +where $\sigma$ denotes the thresholding operation and $\odot$ denotes point-wise multiplication. In addition, $F$ denotes the backbone features, $M$ denotes the predicted mask probabilities. $B_{pred}$ and $B_{mask}$ denote the predicted boxes and boxes decided by predicted masks, respectively. So far, the instance queries have learned from the entire point cloud and the reliable regions, they can segment the majority of objects with easy contexts. To tackle the complex contexts, the instance queries learn from the unreliable regions described by the predicted core boxes. The predicted core boxes are scaled from the predicted boxes, aiming to represent the core regions of objects, and corresponding features are obtained as follows: + +$$ +\boldsymbol {F} ^ {u n} = \sigma (\boldsymbol {F}, \boldsymbol {M} \odot (\boldsymbol {B} _ {\text {c o r e}} \cap \boldsymbol {B} _ {\text {m a s k}})) \tag {7} +$$ + +where $B_{core}$ denotes the predicted core boxes. The instance queries have strong discriminative ability in such a global-to-local manner, then they are fed into a self-attention module and a feed-forward network to aggregate the learned information, which further boosts the performance. + +# 3.5. Bilateral Matching and Training Loss + +Our method makes bilateral matching between pseudo instances and predicted instances first and then utilizes the paired pseudo instances as ground truth $\mathbf{y}_{ins}$ to supervise the paired predicted instances $\mathbf{x}_{ins}$ . + +Bilateral matching. Following SPFormer [36] and IS-BNet [30], our method utilizes the Hungarian method [18] to achieve the optimal assignment. We compute a pairwise matching cost $\mathbf{c}_{i,j}$ to evaluate the similarity of the $i^{th}$ predicted instance and the $j^{th}$ pseudo ground truth. And $\mathbf{c}^{i,j}$ is based on the classification probabilities and the binary masks, it is defined as: + +$$ +\begin{array}{l} \mathbf {c} ^ {i, j} = \lambda_ {1} L _ {C E} (\mathbf {x} _ {c l s} ^ {i}, \mathbf {y} _ {c l s} ^ {j}) \\ + \lambda_ {2} \left(L _ {B C E} \left(\mathbf {x} _ {\text {m a s k}} ^ {i}, \mathbf {y} _ {\text {m a s k}} ^ {j}\right) + L _ {\text {d i c e}} \left(\mathbf {x} _ {\text {m a s k}} ^ {i}, \mathbf {y} _ {\text {m a s k}} ^ {j}\right)\right) \tag {8} \\ \end{array} +$$ + +Training loss. Our training loss consists of two items, namely the loss of pseudo label generation $L_{pl}$ and the loss of instance segmentation $L_{seg}$ , which is defined as: + +$$ +L = L _ {p l} + L _ {s e g} \tag {9} +$$ + +As for the instance segmentation, except for supervising the predicted semantic classes and the binary masks, we also supervise the predicted bounding boxes and the core boxes utilized in the coarse-to-fine instance segmentator. The ground truth of the scale of the core box is based on the IoU score of the predicted masks and the pseudo-ground truth masks. The loss of instance segmentation is defined as: + +$$ +\begin{array}{l} L _ {s e g} = \lambda_ {1} L _ {C E} (\mathbf {x} _ {c l s}, \mathbf {y} _ {c l s}) \\ + \lambda_ {2} \left(L _ {B C E} \left(\mathbf {x} _ {\text {m a s k}}, \mathbf {y} _ {\text {m a s k}}\right) + L _ {\text {d i c e}} \left(\mathbf {x} _ {\text {m a s k}}, \mathbf {y} _ {\text {m a s k}}\right)\right) \\ + \lambda_ {3} \left(L _ {L _ {1}} \left(\mathbf {x} _ {\text {b o x}}, \mathbf {y} _ {\text {b o x}}\right) + L _ {M S E} \left(\mathbf {x} _ {\text {c o r e - b o x}}, \mathbf {y} _ {\text {c o r e - b o x}}\right)\right) \tag {10} \\ \end{array} +$$ + +# 4. Experiments + +# 4.1. Experimental Setup + +Datasets. We conduct experiments on two datasets, ScanNetV2 [6] and S3DIS [1]. The ScanNetV2 provides splits containing 1201, 312, and 100 scenes for training, validation, and testing. Scenes are annotated with semantic labels covering 18 categories and instance labels. And the S3DIS dataset is another indoor dataset contains 272 scenes across six different areas, and it is annotated with 13 object categories. Following existing methods [4, 27, 29], the scenes of S3DIS Area 5 are used for validation while the others are used for training. + +Evaluation metrics. We use the mean average precision $(AP)$ as the evaluation metric for both ScanNetV2 and S3DIS, which is utilized as a common evaluation metric in 3D instance segmentation. $AP$ is the averaged score with IoU thresholds from $95\%$ to $50\%$ with a step size of $5\%$ . The mean average precision with IoU thresholds at $50\%$ and $25\%$ are used as evaluation metrics as well, which are denoted as $AP_{50}$ and $AP_{25}$ , respectively. + +Implementation details. Our model is trained on a single RTX 3090 GPU, and we utilize SPFormer [36] and ISBNet [30] as codebases for ScanNetV2 and S3DIS, respectively. We utilize the AdamW optimizer with a learning rate of 0.0002 and weight decay at 0.05 for the training of Scan- + +Table 1. Comparison on ScanNetV2. "Mask" denotes using dense point-wise annotations, $S_0$ and $S_1$ denote using accurate bounding box and sketchy bounding box annotations, respectively. + +
MethodSup.VenueScanNet TestScanNet Val
AP50AP25AP50AP25
GraphCut [13]NIPS2273.283.269.179.3
ISBNet [30]CVPR2376.384.573.182.5
SPFormer [36]MaskAAAI2377.085.173.982.9
Mask3D [34]ICRA2378.087.073.783.5
MSTA3D [40]ACM MM2479.587.977.085.4
Spherical Mask [35]CVPR2481.2-79.988.2
Box2Mask [4]ECCV2267.780.359.771.8
WISGP+PointGroup [8]WACV23--50.264.9
WISGP+SSTNet [8]WACV23--56.970.2
GaPro+SPFormer [29]S0ICCV2369.282.470.479.9
CIP-WPIS [49]WACV24--69.378.6
BSNet+SPFormer [27]CVPR24--72.783.4
Sketchy-3DIS (ours)-70.186.668.883.6
Box2Mask [4]ECCV22--52.467.5
GaPro+SPFormer [29]S1ICCV23--53.572.2
Sketchy-3DIS (ours)---65.883.1
+ +NetV2. As for the weights of matching cost and losses, $(\lambda_{1},\lambda_{2},\lambda_{3})$ are set as (0.5, 1.0, 0.5). All the models are trained from scratch, and without specification, the hyperparameters and the training details are kept the same as the [36] or [30]. + +# 4.2. Main Results + +ScanNetV2. In Table 1, we compare our method and existing state-of-the-art methods on ScanNetV2 dataset. Attributed to the innovative design of Sketchy-3DIS, which achieves the pseudo labeling and instance segmenting simultaneously, our method achieves considerable performance on both the validation set and online test set of ScanNetV2. Notably, without multi-modal priors [49] or synthesized scenes [27], our method outperforms existing methods in terms of $AP_{25}$ , which demonstrates the effectiveness of our method. Additionally, the comparable performances to fully-supervised methods demonstrate that our method can play a significant role in autonomous driving with high efficacy. + +S3DIS. In Table 2, we make comparisons between existing methods and our proposed Sketchy-3DIS on S3DIS Area 5. The consistent advantages demonstrate the strong generalization ability of our proposed SKeychy-3DIS. Notably, our method exceeds the fully-supervised baseline [30] with $+3.3$ in $AP_{50}$ , which verifies that our proposed Sketchy-3DIS can make full use of instance indications provided by the sketchy bounding boxes. + +Quality comparisons of pseudo labels. We compare the qualities of the generated pseudo labels of GaPro [29] and ours under $S_{1}$ sketchy bounding boxes. The qualitative comparisons are present in Fig.5, which illustrates that our method generates more accurate pseudo labels than GaPro not only in details (highlighted by the green cycles) but also + +![](images/9d880ff7dafa0db6e88c7c18acd2350247eb866e1c757d8bdd714155c09b071b.jpg) +Input Point Cloud + +![](images/2f8055b6abed0018c5e17984d2b3e664e21fb965914e4b24a699fc188c56cacd.jpg) +Ground Truth + +![](images/e8b5595da1616053db9f43a0d70cd1da605da8d27396ddf7d88f859c0f1ed442.jpg) +Sketchy-3DIS (ours) + +![](images/94d30bdbd197adefb9b0e0dc886f4841f97056e2356d5747e06844ecd209e629.jpg) +GaPro + +![](images/d4149b60152f290578ac3948825238b5c10583b29683a74323626f5279cfe5e4.jpg) +Figure 5. Visualization comparison of pseudo labels on the ScanNetV2 training set. The black denotes the background points and other colors denote different objects. The green and red cycles highlight the key regions. + +![](images/c433c3e291a7813c54998067d0bad0c89b0ad96370b046f2b32099092a531fc6.jpg) + +![](images/5419585a74f0764d2ecc0a5c26c40639c0ea2143be49371382b8c685b90f8a25.jpg) + +![](images/e69f7be396bd3c10f8034e3cd8f6430f9397a057f21c2fffa30d14e7be24d259.jpg) + +Table 2. Comparison on S3DIS Area 5. "Mask" denotes using dense point-wise annotations, $S_0$ and $S_1$ denote using accurate bounding box and sketchy bounding box annotations, respectively. \*\*" denotes the results are reproduced by GaPro [29]. + +
MethodSup.VenueAPAP50
SoftGroup [41]CVPR2251.666.1
GraphCut [13]NIPS2254.166.4
ISBNet [30]MaskCVPR2354.065.8
Mask3D [34]ICRA2356.668.4
MSTA3D [40]ACM MM24-70.0
Spherical Mask [35]CVPR2460.572.3
Box2Mask* [4]ECCV2243.654.6
WISGP+PointGroup [8]WACV2333.548.6
WISGP+SSTNet [8]S0WACV2337.251.0
GaPro+ISBNet [29]ICCV2350.561.2
BSNet+ISBNet [27]CVPR2453.064.3
Sketchy-3DIS (ours)-53.469.1
Sketchy-3DIS (ours)S1-50.764.6
+ +in distinguishing the objects belonging to the same categories (highlighted by the red cycles). Additionally, compared with the ground truth, our method generates comparable pseudo labels as well. + +# 4.3. Ablation Studies and Analysis + +We conduct comprehensive ablation studies to justify the effective design of our method. All these ablation experiments are conducted on the ScanNetV2 validation set unless otherwise stated. + +Robustness under various sketchy bounding boxes. In Table 3, we evaluate our method under four kinds of various sketchy bounding boxes illustrated in Fig.2 on both ScanNetV2 validation set and S3DIS Area 5. It can be observed that the performance of our method deteriorates slightly as the increasing of the boxes' sketchy degree, which verifies the robustness of our proposed Sketchy-3DIS. + +Table 3. Robustness under various sketchy bounding boxes. $S_0$ to $S_4$ denote the sketchy degree of the annotated bounding boxes. + +
Sketchy DegreeScanNetV2 ValS3DIS Area 5
AP50AP25AP50AP25
S462.580.150.963.6
S362.981.452.366.6
S263.782.161.272.7
S165.883.164.673.9
S068.883.669.177.5
+ +Table 4. Effectiveness of pseudo labeler and instance segmentator. UnScene3D [32] is an sota unsupervised method. "disjoint" denotes training the 3D UNet and pseudo labeler first and then freeze them to train instance segmentator while "joint" denotes training the pseudo labeler and the instance segmentator jointly. + +
MethodAPAP50AP25
UnScene3D [32]21.440.352.6
SKetchy-3DIS (disjoint)45.360.470.0
SKetchy-3DIS (joint)53.469.177.5
+ +Effectiveness of labeler and segmentator. In Table 4, experiments are conducted on the S3DIS Area 5 to verify the effectiveness of our designed adaptive box-to-point pseudo labeler and coarse-to-fine instance segmentator. We choose the state-of-the-art unsupervised method UnScene3D [32] for 3D instance segmentation as the baseline. The experimental results verify that our designed pseudo labeler can generate high-quality pseudo labels, and combined with our designed instance segmentator, our method achieves significant performance. + +Effectiveness of components in pseudo labeler. In Table 5, we evaluate the effectiveness of various components in our designed pseudo labeler. The first row is the results of the state-of-the-art unsupervised method [32], and the comparison between the last two rows verifies that the learned point-to-box similarity and point-to-instance assignments + +Table 5. Effectiveness of components in pseudo labeler. "Partition" denotes obtaining point-wise labels by box-to-point spatial correlations, "Assign" denotes assigning the points within overlapped boxes, and "Similarity" denotes filtering the background points within boxes according to the point-to-box similarities. + +
PartitionAssignSimilarityAPAP50AP25
15.932.258.5
✓41.864.872.3
✓✓45.267.383.4
✓✓✓46.068.883.6
+ +Table 6. Attentions in Multi-level Attention Block. "Scene", "Coarse", and "Core" denote interacting with the features of the whole scene, the coarse instance regions, and the core regions of the instance, respectively. And "Self" denotes augmenting by the learned instance query itself. + +
SceneCoarseCoreSelfAPAP50AP25
41.364.179.8
✓✓44.667.082.1
✓✓✓44.767.182.9
✓✓✓✓46.068.883.6
+ +Table 7. Effectiveness of instance segmentation losses. $L_{box}$ , $L_{core-box}$ , and $L_{mask}$ denote the supervisions of the predicted instance box, instance core box, and instance mask, respectively. + +
LboxLcore-boxLmaskAPAP50AP25
✓45.266.882.3
✓✓45.467.783.0
✓✓45.067.883.0
✓✓✓46.068.883.6
+ +can improve the performance by filtering the background points within boxes and assigning the points within overlapped boxes. + +Effectiveness of various attentions in Multi-level Attention Block. In Table 6, we conduct ablation studies on various combinations of attention layers utilized in Multi-level Attention Block. The best performance is obtained with the combination of the four attention modules. These experimental results verify that our proposed Multi-level Attention Block can explore the local contexts adaptively and aggregates them, which boosts the segmentation performance progressively. + +Effectiveness of instance segmentation losses. In Table 7, we compare the performances of the losses of the predicted boxes, core boxes, and masks. The results show that the best performance is obtained with the combination of the above three kinds of losses, combined with our coarse-to-fine instance segmentator, they achieve inspiring performance. + +Qualitative results of 3D instance segmentation. Some visualized samples of predicted instances for the ScanNetV2 + +![](images/0f68def47a1778c26a28d7293916f8e2da019ff4160b7b33f61698abb1321a20.jpg) +Figure 6. Qualitative results on ScanNetV2 validation set. The black denotes the background points and other colors denote different objects. The cycles highlight the key regions. + +validation set are shown in Fig.6. It can be observed that our method shows consistent effectiveness in segmenting objects for easy and complex scenes. Even if there are many objects belonging to the same category and they are close to each other, our method demonstrates considerable performance with the sketchy bounding box annotation. Notably, our method can adaptively segment the objects that are ignored in the ground truth such as the cabinets highlighted in red cycles, which demonstrates the strong recognition capability of our method. + +# 5. Conclusion + +In this work, we propose a query-based framework for weakly supervised 3D instance segmentation, Sketchy-3DIS, which is more robust and effective than existing bounding-box supervised methods. Sketchy-3DIS adaptively generates pseudo per-point labels from sketchy boxes and predicts instances in a coarse-to-fine manner. Extensive experiments on ScanNetV2 and S3DIS benchmarks show that our method achieves leading performance, and even outperforms some fully-supervised methods. However, our explored experiments demonstrate that the performance of our method degraded severely when the annotated sketchy bounding boxes are immensely inaccurate, exploration for this issue may be the future study orientation. + +# Acknowledgement + +This work was supported by the National Key R&D Program of China No. 2024YFC3015801, National Science Fund of China under Grant Nos. 62361166670, 62276144, U24A20330, 62306238 and the Fundamental Research Funds for the Central Universities. + +# References + +[1] Iro Armeni, Ozan Sener, Amir R Zamir, Helen Jiang, Ioannis Brilakis, Martin Fischer, and Silvio Savarese. 3D semantic parsing of large-scale indoor spaces. In CVPR, pages 1534-1543, 2016. 6 +[2] Shaoyu Chen, Jiemin Fang, Qian Zhang, Wenyu Liu, and Xinggang Wang. Hierarchical aggregation for 3D instance segmentation. In CVPR, pages 15467-15476, 2021. 1, 2 +[3] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In CVPR, pages 1290–1299, 2022. 2 +[4] Julian Chibane, Francis Engelmann, Tuan Anh Tran, and Gerard Pons-Moll. Box2Mask: Weakly supervised 3D semantic instance segmentation using bounding boxes. In ECCV, pages 681-699. Springer, 2022. 1, 2, 3, 6, 7 +[5] Ruihang Chu, Xiaqing Ye, Zhengzhe Liu, Xiao Tan, Xiaojuan Qi, Chi-Wing Fu, and Jiaya Jia. Twist: Two-way inter-label self-training for semi-supervised 3D instance segmentation. In CVPR, pages 1100-1109, 2022. 1, 2 +[6] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. ScanNet: Richly-annotated 3D reconstructions of indoor scenes. In CVPR, pages 5828-5839, 2017. 1, 6 +[7] Shichao Dong, Ruibo Li, Jiacheng Wei, Fayao Liu, and Guosheng Lin. Rwseg: Cross-graph competing random walks for weakly supervised 3D instance segmentation. arXiv preprint arXiv:2208.05110, 2022. 1, 2 +[8] Heming Du, Xin Yu, Farookh Hussain, Mohammad Ali Armin, Lars Petersson, and Weihao Li. Weakly-supervised point cloud instance segmentation with geometric priors. In WACV, pages 4271-4280, 2023. 1, 2, 3, 6, 7 +[9] Benjamin Graham, Martin Engelcke, and Laurens Van Der Maaten. 3D semantic segmentation with submanifold sparse convolutional networks. In CVPR, pages 9224-9232, 2018. 2 +[10] Tong He, Chunhua Shen, and Anton Van Den Hengel. Dyco3D: Robust instance segmentation of 3D point clouds through dynamic convolution. In CVPR, pages 354-363, 2021. 2 +[11] Ji Hou, Angela Dai, and Matthias Nießner. 3D-SIS: 3D semantic instance segmentation of rgb-d scans. In CVPR, pages 4421-4430, 2019. 2 +[12] Ji Hou, Benjamin Graham, Matthias Nießner, and Saining Xie. Exploring data-efficient 3D scene understanding with contrastive scene contexts. In CVPR, pages 15587-15597, 2021. 1, 2 +[13] Le Hui, Linghua Tang, Yaqi Shen, Jin Xie, and Jian Yang. Learning superpoint graph cut for 3D instance segmentation. NIPS, 35:36804-36817, 2022. 1, 2, 6, 7 +[14] Li Jiang, Hengshuang Zhao, Shaoshuai Shi, Shu Liu, ChiWing Fu, and Jiaya Jia. Pointgroup: Dual-set point grouping for 3D instance segmentation. In CVPR, pages 4867-4876, 2020. 1, 2 +[15] Lei Ke, Martin Danelljan, Henghui Ding, Yu-Wing Tai, Chi-Keung Tang, and Fisher Yu. Mask-free video instance segmentation. In CVPR, pages 22857–22866, 2023. 3 + +[16] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In ICCV, pages 4015-4026, 2023. 2 +[17] Maxim Kolodiazhnyi, Anna Vorontsova, Anton Konushin, and Danila Rukhovich. OneFormer3D: One transformer for unified point cloud segmentation. In CVPR, pages 20943-20953, 2024. 2 +[18] Harold W Kuhn. The hungarian method for the assignment problem. Naval research logistics quarterly, 2(1-2):83-97, 1955. 6 +[19] Xin Lai, Yuhui Yuan, Ruihang Chu, Yukang Chen, Han Hu, and Jiaya Jia. Mask-Attention-Free Transformer for 3D instance segmentation. In ICCV, pages 3693–3703, 2023. 1, 2 +[20] Shiyi Lan, Zhiding Yu, Christopher Choy, Subhashree Radhakrishnan, Guilin Liu, Yuke Zhu, Larry S Davis, and Anima Anandkumar. DiscoBox: Weakly supervised instance segmentation and semantic correspondence from box supervision. In ICCV, pages 3406-3416, 2021. 3 +[21] Loic Landrieu and Martin Simonovsky. Large-scale point cloud semantic segmentation with superpoint graphs. In CVPR, pages 4558-4567, 2018. 3 +[22] Zhihao Liang, Zhihao Li, Songcen Xu, Mingkui Tan, and Kui Jia. Instance segmentation in 3D scenes using semantic superpoint tree networks. In ICCV, pages 2783-2792, 2021. 2 +[23] Yongbin Liao, Hongyuan Zhu, Yanggang Zhang, Chuangguan Ye, Tao Chen, and Jiayuan Fan. Point cloud instance segmentation with semi-supervised bounding-box mining. TPAMI, 44(12):10159-10170, 2021. 1, 2 +[24] Zhengzhe Liu, Xiaojuan Qi, and Chi-Wing Fu. One thing one click: A self-training approach for weakly supervised 3D semantic segmentation. In CVPR, pages 1726-1736, 2021. 1, 2 +[25] Ze Liu, Zheng Zhang, Yue Cao, Han Hu, and Xin Tong. Group-free 3D object detection via transformers. In ICCV, pages 2949-2958, 2021. 1 +[26] Jiahao Lu, Jiacheng Deng, Chuxin Wang, Jianfeng He, and Tianzhu Zhang. Query refinement transformer for 3D instance segmentation. In ICCV, pages 18516-18526, 2023. 2 +[27] Jiahao Lu, Jiacheng Deng, and Tianzhu Zhang. BSNet: Box-supervised simulation-assisted mean teacher for 3D instance segmentation. In CVPR, pages 20374-20384, 2024. 2, 3, 6, 7 +[28] Ishan Misra, Rohit Girdhar, and Armand Joulin. An end-to-end transformer model for 3D object detection. In ICCV, pages 2906-2917, 2021. 1 +[29] Tuan Duc Ngo, Binh-Son Hua, and Khoi Nguyen. GaPro: Box-supervised 3D point cloud instance segmentation using gaussian processes as pseudo labelers. In ICCV, pages 17794-17803, 2023. 1, 2, 3, 6, 7 +[30] Tuan Duc Ngo, Binh-Son Hua, and Khoi Nguyen. ISBNet: a 3D point cloud instance segmentation network with instance-aware sampling and box-aware dynamic convolution. In CVPR, pages 13550-13559, 2023. 1, 2, 6, 7 +[31] Charles R Qi, Or Litany, Kaiming He, and Leonidas J Guibas. Deep hough voting for 3D object detection in point clouds. In ICCV, pages 9277-9286, 2019. 1 + +[32] David Rozenberszki, Or Litany, and Angela Dai. UnScene3D: Unsupervised 3D instance segmentation for indoor scenes. In CVPR, pages 19957-19967, 2024. 7 +[33] Danila Rukhovich, Anna Vorontsova, and Anton Konushin. FCAF3D: Fully convolutional anchor-free 3D object detection. In ECCV, pages 477-493. Springer, 2022. 1 +[34] Jonas Schult, Francis Engelmann, Alexander Hermans, Or Litany, Siyu Tang, and Bastian Leibe. Mask3D: Mask transformer for 3D semantic instance segmentation. In ICRA, pages 8216-8223. IEEE, 2023. 1, 2, 6, 7 +[35] Sangyun Shin, Kaichen Zhou, Madhu Vankadari, Andrew Markham, and Niki Trigoni. Spherical mask: Coarse-to-fine 3D point cloud instance segmentation with spherical representation. In CVPR, pages 4060-4069, 2024. 6, 7 +[36] Jiahao Sun, Chunmei Qing, Junpeng Tan, and Xiangmin Xu. Superpoint transformer for 3D scene instance segmentation. In AAAI, pages 2393-2401, 2023. 1, 2, 6 +[37] Linghua Tang, Le Hui, and Jin Xie. Learning inter-superpoint affinity for weakly supervised 3D instance segmentation. In ACCV, pages 1282-1297, 2022. 1, 2 +[38] An Tao, Yueqi Duan, Yi Wei, Jiwen Lu, and Jie Zhou. Seg-Group: Seg-level supervision for 3D instance and semantic segmentation. TIP, 31:4952-4965, 2022. 1, 2 +[39] Zhi Tian, Chunhua Shen, Xinlong Wang, and Hao Chen. BoxInst: High-performance instance segmentation with box annotations. In CVPR, pages 5443-5452, 2021. 3 +[40] Duc Dang Trung Tran, Byeongkeun Kang, and Yeejin Lee. MSTA3D: Multi-scale twin-attention for 3d instance segmentation. In ACM MM, pages 1467-1475, 2024. 2, 6, 7 +[41] Thang Vu, Kookhoi Kim, Tung M Luu, Thanh Nguyen, and Chang D Yoo. SoftGroup for 3D instance segmentation on point clouds. In CVPR, pages 2708-2717, 2022. 1, 2, 7 +[42] Haiyang Wang, Shaocong Dong, Shaoshuai Shi, Aoxue Li, Jianan Li, Zhenguo Li, Liwei Wang, et al. CAGroup3D: Class-aware grouping for 3D object detection on point clouds. NIPS, 35:29975-29988, 2022. 1 +[43] Haiyang Wang, Shaoshuai Shi, Ze Yang, Rongyao Fang, Qi Qian, Hongsheng Li, Bernt Schiele, and Liwei Wang. RBGNet: Ray-based grouping for 3D object detection. In CVPR, pages 1110-1119, 2022. +[44] Qian Xie, Yu-Kun Lai, Jing Wu, Zhoutao Wang, Yiming Zhang, Kai Xu, and Jun Wang. MLCVNet: Multi-level context votenet for 3D object detection. In CVPR, pages 10447-10456, 2020. 1 +[45] Saining Xie, Jiatao Gu, Demi Guo, Charles R Qi, Leonidas Guibas, and Or Litany. PointContrast: Unsupervised pretraining for 3D point cloud understanding. In ECCV, pages 574-591. Springer, 2020. 1, 2 +[46] Bo Yang, Jianan Wang, Ronald Clark, Qingyong Hu, Sen Wang, Andrew Markham, and Niki Trigoni. Learning object bounding boxes for 3D instance segmentation on point clouds. NIPS, 32, 2019. 2 +[47] Cheng-Kun Yang, Min-Hung Chen, Yung-Yu Chuang, and Yen-Yu Lin. 2D-3D interlaced transformer for point cloud segmentation with scene-level supervision. In ICCV, pages 977–987, 2023. 2 + +[48] Li Yi, Wang Zhao, He Wang, Minhyuk Sung, and Leonidas J Guibas. GSPN: Generative shape proposal network for 3D instance segmentation in point cloud. In CVPR, pages 3947-3956, 2019. 2 +[49] Qingtao Yu, Heming Du, Chen Liu, and Xin Yu. When 3D bounding-box meets SAM: Point cloud instance segmentation with weak-and-noisy supervision. In WACV, pages 3719-3728, 2024. 1, 2, 3, 6 +[50] Zaiwei Zhang, Bo Sun, Haitao Yang, and Qixing Huang. H3DNet: 3D object detection using hybrid geometric primitives. In ECCV, pages 311-329. Springer, 2020. 1 +[51] Yu Zheng, Yueqi Duan, Jiwen Lu, Jie Zhou, and Qi Tian. HyperDet3D: Learning a scene-conditioned 3D object detector. In CVPR, pages 5585-5594, 2022. 1 +[52] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable DETR: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159, 2020. 2 \ No newline at end of file diff --git a/2025/Sketchy Bounding-box Supervision for 3D Instance Segmentation/images.zip b/2025/Sketchy Bounding-box Supervision for 3D Instance Segmentation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..1b4633d2f424728908edbc989e97311bea1e7227 --- /dev/null +++ b/2025/Sketchy Bounding-box Supervision for 3D Instance Segmentation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:05332258abb0b368bda50c05cbf289f65806c639635e60c4b791eaa58761de9f +size 675978 diff --git a/2025/Sketchy Bounding-box Supervision for 3D Instance Segmentation/layout.json b/2025/Sketchy Bounding-box Supervision for 3D Instance Segmentation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..8369048962b15fcacbe64b20e62ed023ae94e89e --- /dev/null +++ b/2025/Sketchy Bounding-box Supervision for 3D Instance Segmentation/layout.json @@ -0,0 +1,9527 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 102, + 103, + 507, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 103, + 507, + 121 + ], + "spans": [ + { + "bbox": [ + 102, + 103, + 507, + 121 + ], + "type": "text", + "content": "Sketchy Bounding-box Supervision for 3D Instance Segmentation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 195, + 142, + 413, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 195, + 142, + 413, + 157 + ], + "spans": [ + { + "bbox": [ + 195, + 142, + 413, + 157 + ], + "type": "text", + "content": "Qian Deng" + }, + { + "bbox": [ + 195, + 142, + 413, + 157 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 195, + 142, + 413, + 157 + ], + "type": "text", + "content": ", Le Hui" + }, + { + "bbox": [ + 195, + 142, + 413, + 157 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 195, + 142, + 413, + 157 + ], + "type": "text", + "content": ", Jin Xie" + }, + { + "bbox": [ + 195, + 142, + 413, + 157 + ], + "type": "inline_equation", + "content": "^{3,4*}" + }, + { + "bbox": [ + 195, + 142, + 413, + 157 + ], + "type": "text", + "content": ", Jian Yang" + }, + { + "bbox": [ + 195, + 142, + 413, + 157 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 121, + 157, + 488, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 157, + 488, + 171 + ], + "spans": [ + { + "bbox": [ + 121, + 157, + 488, + 171 + ], + "type": "text", + "content": "1PCA Lab, College of Computer Science, Nankai University, Tianjin, China" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 80, + 171, + 530, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 171, + 530, + 185 + ], + "spans": [ + { + "bbox": [ + 80, + 171, + 530, + 185 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 80, + 171, + 530, + 185 + ], + "type": "text", + "content": "School of Electronics and Information, Northwestern Polytechnical University, Xi'an, China" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 86, + 185, + 524, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 185, + 524, + 199 + ], + "spans": [ + { + "bbox": [ + 86, + 185, + 524, + 199 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 86, + 185, + 524, + 199 + ], + "type": "text", + "content": "State Key Laboratory for Novel Software Technology, Nanjing University, Nanjing, China" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 102, + 199, + 507, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 199, + 507, + 213 + ], + "spans": [ + { + "bbox": [ + 102, + 199, + 507, + 213 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 102, + 199, + 507, + 213 + ], + "type": "text", + "content": "School of Intelligence Science and Technology, Nanjing University, Suzhou, China" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 63, + 216, + 539, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 216, + 539, + 227 + ], + "spans": [ + { + "bbox": [ + 63, + 216, + 539, + 227 + ], + "type": "text", + "content": "dengqian@mail.nankai.edu.cn; huile@nwpu.edu.cn; csjxie@nju.edu.cn; csjyang@nankai.edu.cn" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 151, + 255, + 200, + 268 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 255, + 200, + 268 + ], + "spans": [ + { + "bbox": [ + 151, + 255, + 200, + 268 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 54, + 280, + 297, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 280, + 297, + 605 + ], + "spans": [ + { + "bbox": [ + 54, + 280, + 297, + 605 + ], + "type": "text", + "content": "Bounding box supervision has gained considerable attention in weakly supervised 3D instance segmentation. While this approach alleviates the need for extensive point-level annotations, obtaining accurate bounding boxes in practical applications remains challenging. To this end, we explore the inaccurate bounding box, named sketchy bounding box, which is imitated through perturbing ground truth bounding box by adding scaling, translation, and rotation. In this paper, we propose Sketchy-3DIS, a novel weakly 3D instance segmentation framework, which jointly learns pseudo labeler and segmentator to improve the performance under the sketchy bounding-box supervisions. Specifically, we first propose an adaptive box-to-point pseudo labeler that adaptively learns to assign points located in the overlapped parts between two sketchy bounding boxes to the correct instance, resulting in compact and pure pseudo instance labels. Then, we present a coarse-to-fine instance segmentator that first predicts coarse instances from the entire point cloud and then learns fine instances based on the region of coarse instances. Finally, by using the pseudo instance labels to supervise the instance segmentator, we can gradually generate high-quality instances through joint training. Extensive experiments show that our method achieves state-of-the-art performance on both the ScanNetV2 and S3DIS benchmarks, and even outperforms several fully supervised methods using sketchy bounding boxes. Code is available at https://github.com/dengq7/Sketchy-3DIS." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 626, + 135, + 639 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 626, + 135, + 639 + ], + "spans": [ + { + "bbox": [ + 55, + 626, + 135, + 639 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 646, + 296, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 646, + 296, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 646, + 296, + 696 + ], + "type": "text", + "content": "With the rise of 3D datasets and the growing need for 3D scene understanding, various 3D computer vision tasks [19, 25, 28-31, 33, 34, 36, 42-44, 50, 51] have gained significant attention from both academia and industry. Among these, 3D" + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 320, + 253, + 375, + 376 + ], + "blocks": [ + { + "bbox": [ + 320, + 253, + 375, + 376 + ], + "lines": [ + { + "bbox": [ + 320, + 253, + 375, + 376 + ], + "spans": [ + { + "bbox": [ + 320, + 253, + 375, + 376 + ], + "type": "image", + "image_path": "218af45a23c27b30bd28ac36739086053e911ca15466e8c63b7fbf0b852823f6.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 332, + 376, + 364, + 385 + ], + "lines": [ + { + "bbox": [ + 332, + 376, + 364, + 385 + ], + "spans": [ + { + "bbox": [ + 332, + 376, + 364, + 385 + ], + "type": "text", + "content": "(a) Inputs" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 388, + 555, + 433 + ], + "lines": [ + { + "bbox": [ + 313, + 388, + 555, + 433 + ], + "spans": [ + { + "bbox": [ + 313, + 388, + 555, + 433 + ], + "type": "text", + "content": "Figure 1. (a) illustrates the inputs of bounding box supervised 3D instance segmentation. (b) compares the performances of GaPro [29] and our Sketchy-3DIS under both accurate and scaled sketchy boxes supervision on ScanNetV2 validation set." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 375, + 252, + 547, + 374 + ], + "blocks": [ + { + "bbox": [ + 375, + 252, + 547, + 374 + ], + "lines": [ + { + "bbox": [ + 375, + 252, + 547, + 374 + ], + "spans": [ + { + "bbox": [ + 375, + 252, + 547, + 374 + ], + "type": "image", + "image_path": "7f25aa8c2fdf6a9f9987cffb92ef4d9a91ff368a1460d4d3ffa3be0ebc415f25.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 413, + 376, + 509, + 385 + ], + "lines": [ + { + "bbox": [ + 413, + 376, + 509, + 385 + ], + "spans": [ + { + "bbox": [ + 413, + 376, + 509, + 385 + ], + "type": "text", + "content": "(b) Performance Comparisons" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 437, + 556, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 556, + 616 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 556, + 616 + ], + "type": "text", + "content": "instance segmentation is both fundamental and challenging, as it involves recognizing the category of each object within a point cloud and delineating its individual mask. Current 3D instance segmentation methods [2, 13, 14, 30, 34, 36, 41] have demonstrated impressive performance, primarily rely on data-hungry deep learning approaches that require substantial training data and dense point-level manual annotations. Notably, point-level fully supervised instance annotation in ScanNet takes approximately 22.3 minutes [6] per scene, while labeling a 3D box in ScanNet takes about 1.93 minutes [38] per scene. Therefore, bounding box annotation stands out as it provides substantial instance-level information while maintaining a reasonable annotation cost, making it a favorable annotation choice for weakly supervised 3D instance segmentation." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 617, + 557, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 557, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 557, + 715 + ], + "type": "text", + "content": "Many weakly supervised methods have emerged to tackle the 3D instance segmentation task. Semi-supervised 3D instance segmentation methods [5, 12, 23, 45] use training data with partial labels, while others [5, 12, 23, 45] utilize indirect labels, such as 2D image annotations. Additionally, object-level annotations, like marking a single point [7, 24, 37, 38] or drawing a bounding box [4, 8, 29, 49] for each object, are employed. Bounding box supervision is another important" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 703, + 143, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 143, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 143, + 712 + ], + "type": "text", + "content": "*Corresponding author" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8879" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 263 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 263 + ], + "type": "text", + "content": "annotation for weakly supervised 3D instance segmentation. Box2Mask [4] is the first attempt that uses bounding box supervision for 3D instance segmentation. Subsequently, the following methods [8, 27, 29, 49] use different constraints to improve performance, such as using local correlations, Gaussian processes, teacher-student model, SAM [16], respectively. Although impressive results have been achieved in weakly supervised 3D instance segmentation, these methods still rely on accurate and compact bounding box annotations. It is worth noting that in practice, accurate and compact boxes are difficult to obtain. Generally, annotated bounding box involves changes in scaling, translation, and rotation. As shown in Fig.1, we observed that the performance of GaPro [29] decreases significantly under scaled bounding box supervision. This motivated us to design a new approach to tackle with these inaccurate bounding box annotations." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 266, + 296, + 672 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 266, + 296, + 672 + ], + "spans": [ + { + "bbox": [ + 56, + 266, + 296, + 672 + ], + "type": "text", + "content": "In this paper, we propose a weakly supervised 3D instance segmentation framework to tackle sketchy bounding box supervision. To simulate sketchy bounding boxes, we impose scaling, translation, and rotation on the ground truth bounding boxes to generate inaccurate bounding boxes. Our method consists of two key components: adaptive box-to-point pseudo labeler and coarse-to-fine instance segmentator. Given a point cloud scene and sketchy bounding boxes, we first use 3D U-Net [9] backbone to extract point features. After that, we feed the learned features and sketchy bounding boxes of the scene into the adaptive box-to-point pseudo labeler to generate high-quality point-level instance labels. For the points located in the overlapped region between two sketchy bounding boxes, we adaptively assign the points to the corresponding box by learning the similarity of points and boxes. For the remaining points, we directly assign them to the corresponding box according to the spatial correlations. In this way, we can transform the sketchy bounding box into a compact bounding box. We consider the labels of points that located within the generated compact bounding box as pseudo instance labels. To accurately predict instances, we input the learned features into the coarse-to-fine instance segmentator iteratively. We first detect the coarse bounding box and mask of the instance by executing queries on the whole point cloud. Based on the small region of the coarse bounding box, we re-predict the fine bounding box and mask by conducting queries within a more precise small region. Finally, by jointly optimizing the adaptive box-to-point pseudo labeler and coarse-to-fine instance segmentator, our method can gradually improve the performance of weakly supervised 3D instance segmentation using the sketchy bounding box annotations. Rich experiments on the ScanNetV2 and S3DIS show that our method have achieved state-of-the-art performance." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 675, + 202, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 675, + 202, + 686 + ], + "spans": [ + { + "bbox": [ + 67, + 675, + 202, + 686 + ], + "type": "text", + "content": "Our contributions are as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": "- We are the first to explore the use of sketchy bounding box annotations for 3D instance segmentation, which allows" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 322, + 72, + 555, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 72, + 555, + 96 + ], + "spans": [ + { + "bbox": [ + 322, + 72, + 555, + 96 + ], + "type": "text", + "content": "for training with noise-tolerant, instance-size-agnostic 3D bounding boxes." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 99, + 555, + 184 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 314, + 99, + 555, + 146 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 99, + 555, + 146 + ], + "spans": [ + { + "bbox": [ + 314, + 99, + 555, + 146 + ], + "type": "text", + "content": "- We propose a new weakly supervised 3D instance segmentation framework, which presents an adaptive box-to-point pseudo labeler and a coarse-to-fine instance segmentator for joint training." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 148, + 553, + 184 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 148, + 553, + 184 + ], + "spans": [ + { + "bbox": [ + 313, + 148, + 553, + 184 + ], + "type": "text", + "content": "- Our method achieves leading performance on both the ScanNetV2 and S3DIS datasets for weakly 3D instance segmentation under sketchy bounding box annotations." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 196, + 400, + 208 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 196, + 400, + 208 + ], + "spans": [ + { + "bbox": [ + 314, + 196, + 400, + 208 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 216, + 541, + 228 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 216, + 541, + 228 + ], + "spans": [ + { + "bbox": [ + 313, + 216, + 541, + 228 + ], + "type": "text", + "content": "2.1. Fully Supervised 3D Instance Segmentation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 234, + 556, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 234, + 556, + 544 + ], + "spans": [ + { + "bbox": [ + 313, + 234, + 556, + 544 + ], + "type": "text", + "content": "Existing fully supervised methods for 3D instance segmentation can be broadly categorized into three main pipelines: grouping-based [2, 13, 14, 22], detection-based [11, 41, 46, 48], and query-based [10, 17, 19, 26, 30, 34-36, 40] methods. Grouping-based methods typically group points based on predicted per-point semantic classes and instance center offsets, but the grouping rules are often manually defined by researchers, which lack flexibility. Detection-based methods, on the other hand, first predict the bounding box for each object and then generate the foreground mask within the box, treating the task as a two-step process of object detection followed by segmentation. More recently, inspired by DETR [52] and Mask2Former [3], query-based methods have been proposed for 3D instance segmentation. These methods simultaneously predict the semantic classes and instance masks of objects using a set of learnable query vectors, which model the objects' geometrical and semantical attributes integrally. Among these methods, query-based approaches tend to achieve superior performance, as the other two pipelines both face challenges such as error accumulation and time-consuming refinement processes. While these fully-supervised methods have made significant progress, they all rely on dense per-point annotations, and their performance tends to degrade when such annotations are unavailable. Therefore, several weakly-supervised methods have been proposed to alleviate this problem." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 552, + 552, + 565 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 552, + 552, + 565 + ], + "spans": [ + { + "bbox": [ + 313, + 552, + 552, + 565 + ], + "type": "text", + "content": "2.2. Weakly Supervised 3D Instance Segmentation" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "text", + "content": "Weakly-supervised methods require partial or implicit annotations. [5, 12, 23, 45] use sparse point annotations and design various label propagation strategies. However, these sparse points cannot fully capture the spatial extent of an object. Approaches like MIT [47] and CIP-WPIS [49] leverage 2D images as additional input for 3D instance segmentation, but this introduces multi-modal dependencies. Other methods [7, 24, 37, 38] rely on instance-level signals, where labeled a single point for per instance, and instance predictions are made using techniques like random walk or clustering. Recently, most state-of-the-art methods [4, 8, 27, 29, 49] have adopted 3D bounding boxes as supervision, where a" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8880" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "text", + "content": "single bounding box is assigned to each object. These approaches naturally capture the spatial extent of an object and the annotated bounding boxes are easier to obtain compared to dense per-point annotations." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 121, + 298, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 121, + 298, + 445 + ], + "spans": [ + { + "bbox": [ + 55, + 121, + 298, + 445 + ], + "type": "text", + "content": "Bounding-box Supervised 3D Instance Segmentation. Bounding boxes have become a widely used supervision signal, not only for 2D instance segmentation tasks [15, 20, 39] but also for 3D instance segmentation. Box2mask [4] is a pioneering approach that utilizes bounding-box supervision for 3D instance segmentation. WISGP [8] divides the point cloud into a determined set and an undetermined set based on explicit spatial correlations, then uses polygon meshes and superpoints to label the undetermined set. GaPro [29] focuses on generating reliable pseudo labels using a Gaussian process. Recently, BSNet [27] employs a simulation-assisted mean teacher framework to generate pseudo labels, enhancing data samples by synthesizing scenes that contain overlapping objects. However, these methods all rely on axis-aligned bounding boxes for supervision, and their performances heavily depend on the accuracy of these annotated boxes. To address this, CIP-WPIS [49] proposes using slightly looser bounding boxes, projecting the point cloud into 2D images, and obtaining 3D instance masks based on segmented 2D masks. While effective, it relies on 2D modality and can only adapt to slightly enlarged bounding boxes, limiting its flexibility. To alleviate the annotation costs as well as obtain robust and effective performance, we design a novel framework Sketchy-3DIS, which endeavors to not only face the common challenge of correctly assigning the points within multiple boxes but also be dedicated to correcting the errors brought by sketchy bounding boxes." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 457, + 111, + 469 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 457, + 111, + 469 + ], + "spans": [ + { + "bbox": [ + 55, + 457, + 111, + 469 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 479, + 123, + 490 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 479, + 123, + 490 + ], + "spans": [ + { + "bbox": [ + 55, + 479, + 123, + 490 + ], + "type": "text", + "content": "3.1. Overview" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 497, + 297, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 497, + 297, + 544 + ], + "spans": [ + { + "bbox": [ + 55, + 497, + 297, + 544 + ], + "type": "text", + "content": "The overall framework of our weakly supervised 3D instance segmentation based on sketchy bounding boxes is depicted in Fig.3. It consists of two key components: adaptive box-to-point pseudo labeler and coarse-to-fine instance segmentator." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "text", + "content": "Assuming a 3D point cloud " + }, + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "inline_equation", + "content": "P \\in \\mathbb{R}^{N \\times C}" + }, + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "text", + "content": " represent the number of 3D points and channels, respectively. In the basic setting, each point is represented by the 3D coordinate, RGB color, and surface normal vector. In the training phase, we first feed the input point cloud and the sketchy bounding boxes (defined in Sec. 3.2) into the 3D UNet backbone network for feature extraction, then we average the point features located in the superpoint [21] to generate superpoint-level features. Next, we input the features and sketchy bounding boxes to the adaptive box-to-point pseudo labeler (Sec. 3.3) for converting the rough box-level annotation to the detailed point-level annotation of the instance. Meanwhile, we feed the feature into the coarse-to-fine instance segmentator (Sec. 3.4) to predict instances by" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 321, + 72, + 403, + 133 + ], + "blocks": [ + { + "bbox": [ + 321, + 72, + 403, + 133 + ], + "lines": [ + { + "bbox": [ + 321, + 72, + 403, + 133 + ], + "spans": [ + { + "bbox": [ + 321, + 72, + 403, + 133 + ], + "type": "image", + "image_path": "e10580a19fbb3a04ba86032c1afdd3bd3777e80298757556110bd2e6953fd0cb.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 339, + 134, + 386, + 144 + ], + "lines": [ + { + "bbox": [ + 339, + 134, + 386, + 144 + ], + "spans": [ + { + "bbox": [ + 339, + 134, + 386, + 144 + ], + "type": "text", + "content": "Point Cloud" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 427, + 95, + 450, + 117 + ], + "blocks": [ + { + "bbox": [ + 427, + 95, + 450, + 117 + ], + "lines": [ + { + "bbox": [ + 427, + 95, + 450, + 117 + ], + "spans": [ + { + "bbox": [ + 427, + 95, + 450, + 117 + ], + "type": "image", + "image_path": "5257c541f2a9726e29353ab9494e750389c4a39b4fbb27e31d5d89ad9b69a7da.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 484, + 89, + 523, + 122 + ], + "blocks": [ + { + "bbox": [ + 484, + 89, + 523, + 122 + ], + "lines": [ + { + "bbox": [ + 484, + 89, + 523, + 122 + ], + "spans": [ + { + "bbox": [ + 484, + 89, + 523, + 122 + ], + "type": "image", + "image_path": "a6100bc158c2fe0f3e4d6702eb96aca7989542341f4a1e5ae09a29cd08c5faa8.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 482, + 133, + 527, + 145 + ], + "lines": [ + { + "bbox": [ + 482, + 133, + 527, + 145 + ], + "spans": [ + { + "bbox": [ + 482, + 133, + 527, + 145 + ], + "type": "text", + "content": "Scaled " + }, + { + "bbox": [ + 482, + 133, + 527, + 145 + ], + "type": "inline_equation", + "content": "(S_{1})" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 353, + 156, + 398, + 190 + ], + "blocks": [ + { + "bbox": [ + 353, + 156, + 398, + 190 + ], + "lines": [ + { + "bbox": [ + 353, + 156, + 398, + 190 + ], + "spans": [ + { + "bbox": [ + 353, + 156, + 398, + 190 + ], + "type": "image", + "image_path": "4d11b69675b686840c5a1c14f900dcfd7b4be15ac53e8c108c05e108bf51d0d9.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 336, + 198, + 393, + 221 + ], + "lines": [ + { + "bbox": [ + 336, + 198, + 393, + 221 + ], + "spans": [ + { + "bbox": [ + 336, + 198, + 393, + 221 + ], + "type": "text", + "content": "Scaled and translated " + }, + { + "bbox": [ + 336, + 198, + 393, + 221 + ], + "type": "inline_equation", + "content": "(S_{2})" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 414, + 151, + 462, + 196 + ], + "blocks": [ + { + "bbox": [ + 410, + 133, + 466, + 145 + ], + "lines": [ + { + "bbox": [ + 410, + 133, + 466, + 145 + ], + "spans": [ + { + "bbox": [ + 410, + 133, + 466, + 145 + ], + "type": "text", + "content": "Accurate " + }, + { + "bbox": [ + 410, + 133, + 466, + 145 + ], + "type": "inline_equation", + "content": "(S_0)" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 414, + 151, + 462, + 196 + ], + "lines": [ + { + "bbox": [ + 414, + 151, + 462, + 196 + ], + "spans": [ + { + "bbox": [ + 414, + 151, + 462, + 196 + ], + "type": "image", + "image_path": "8c2da4509d4486ba0e3ca9f2060700859132a946787fd3678bb8496870e1b7ea.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 414, + 198, + 461, + 221 + ], + "lines": [ + { + "bbox": [ + 414, + 198, + 461, + 221 + ], + "spans": [ + { + "bbox": [ + 414, + 198, + 461, + 221 + ], + "type": "text", + "content": "Scaled and rotated " + }, + { + "bbox": [ + 414, + 198, + 461, + 221 + ], + "type": "inline_equation", + "content": "(S_{3})" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 492, + 152, + 540, + 196 + ], + "blocks": [ + { + "bbox": [ + 492, + 152, + 540, + 196 + ], + "lines": [ + { + "bbox": [ + 492, + 152, + 540, + 196 + ], + "spans": [ + { + "bbox": [ + 492, + 152, + 540, + 196 + ], + "type": "image", + "image_path": "6fa0b9391a2ca90bf0043c9a97ea3354a66b5a3a2d3a8137e4c4a74d0f513409.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 469, + 198, + 538, + 221 + ], + "lines": [ + { + "bbox": [ + 469, + 198, + 538, + 221 + ], + "spans": [ + { + "bbox": [ + 469, + 198, + 538, + 221 + ], + "type": "text", + "content": "Scaled, translated and rotated " + }, + { + "bbox": [ + 469, + 198, + 538, + 221 + ], + "type": "inline_equation", + "content": "(S_4)" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 225, + 555, + 248 + ], + "lines": [ + { + "bbox": [ + 313, + 225, + 555, + 248 + ], + "spans": [ + { + "bbox": [ + 313, + 225, + 555, + 248 + ], + "type": "text", + "content": "Figure 2. Various sketchy bounding boxes under scaling, translation, and rotation perturbations." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 261, + 555, + 333 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 261, + 555, + 333 + ], + "spans": [ + { + "bbox": [ + 313, + 261, + 555, + 333 + ], + "type": "text", + "content": "Transformer. Finally, we employ bilateral matching between the generated pseudo labels (regarded as the ground truth) and the predicted instances to build the correspondence between them for training (Sec. 3.5). In the inference phase, a new point cloud is fed to the backbone network and the coarse-to-fine instance segmentator to segment instances." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 343, + 479, + 357 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 343, + 479, + 357 + ], + "spans": [ + { + "bbox": [ + 313, + 343, + 479, + 357 + ], + "type": "text", + "content": "3.2. Sketchy Bounding Box Setting" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 362, + 555, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 362, + 555, + 470 + ], + "spans": [ + { + "bbox": [ + 313, + 362, + 555, + 470 + ], + "type": "text", + "content": "Our method focuses on studying weakly supervised 3D instance segmentation based on the sketchy bounding box, i.e., inaccurate bounding box. In practice, it is unable to acquire compact 3D bounding box annotations for weakly supervised 3D instance segmentation. Therefore, we perturb the ground truth bounding box with scaling, translation and rotation to mimic the sketchy bounding boxes. As shown in Fig.2, we illustrate the generation process of different types of sketchy bounding boxes." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 471, + 554, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 471, + 554, + 495 + ], + "spans": [ + { + "bbox": [ + 313, + 471, + 554, + 495 + ], + "type": "text", + "content": "Specifically, given the ground truth bounding box, which is denoted by the position of two corner points:" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 347, + 496, + 554, + 524 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 496, + 554, + 524 + ], + "spans": [ + { + "bbox": [ + 347, + 496, + 554, + 524 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathbf {B} = \\left[ B _ {\\text {m i n}}, B _ {\\text {m a x}} \\right] \\\\ = \\left[ x _ {\\min }, y _ {\\min }, z _ {\\min }, x _ {\\max }, y _ {\\max }, z _ {\\max } \\right] \\tag {1} \\\\ \\end{array}", + "image_path": "80d5fb8d0e458fee5e6d2e8f317085ee18ed0f25ef82665f6119ea0109fe867e.jpg" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 525, + 555, + 560 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 525, + 555, + 560 + ], + "spans": [ + { + "bbox": [ + 313, + 525, + 555, + 560 + ], + "type": "text", + "content": "First, we individually employ scaling, translation, and rotation on the ground truth bounding box to generate the basic inaccurate bounding boxes, which are formulated as:" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 343, + 562, + 500, + 575 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 343, + 562, + 500, + 575 + ], + "spans": [ + { + "bbox": [ + 343, + 562, + 500, + 575 + ], + "type": "interline_equation", + "content": "\\mathbf {B} _ {\\text {s c a l e d}} = \\left[ B _ {\\min } - \\alpha E, B _ {\\max } + \\alpha E \\right]", + "image_path": "b753b74f8f55ca8ef7b9a2c6809ea3b2fef288f87852cae5779e10b2bfcc98f3.jpg" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 343, + 577, + 555, + 590 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 343, + 577, + 555, + 590 + ], + "spans": [ + { + "bbox": [ + 343, + 577, + 555, + 590 + ], + "type": "interline_equation", + "content": "\\mathbf {B} _ {\\text {t r a n s l a t e d}} = \\left[ B _ {\\min } + \\beta E, B _ {\\max } + \\beta E \\right] \\tag {2}", + "image_path": "d68526351b9d88e3cf984cd03e5a1602eb5ecf4c9f0195aa94ede840271ded7f.jpg" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 343, + 592, + 524, + 605 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 343, + 592, + 524, + 605 + ], + "spans": [ + { + "bbox": [ + 343, + 592, + 524, + 605 + ], + "type": "interline_equation", + "content": "\\mathbf {B} _ {\\text {r o t a t e d}} = r \\left(\\left[ B _ {\\min }, B _ {\\max } \\right], M E, \\gamma \\pi / 1 8 0\\right)", + "image_path": "ba4b05bf4bbca856d78f7173677f58bb91d114efc0e4bec52c823eaabf9a549c.jpg" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "text", + "content": " are the parameters of different operations. Note that " + }, + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "inline_equation", + "content": "E = B_{max} - B_{min}" + }, + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "inline_equation", + "content": "ME = (B_{max} + B_{min}) / 2" + }, + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "text", + "content": ". Considering that manually annotated 3D bounding boxes will not have significant differences from compact 3D ground truth bounding boxes. Therefore, we empirically set small values for those parameters. In experiments, we set " + }, + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\alpha = 5\\%" + }, + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\beta = 5\\%" + }, + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\gamma = 5" + }, + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "text", + "content": ", respectively. Then, we generate diverse sketchy bounding boxes by combining these basic operations " + }, + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{B}_{\\mathrm{scaled}}" + }, + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{B}_{\\mathrm{translated}}" + }, + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{B}_{\\mathrm{rotated}}" + }, + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "text", + "content": ". In Fig.2, we show four" + } + ] + } + ], + "index": 28 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "8881" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 66, + 71, + 211, + 300 + ], + "blocks": [ + { + "bbox": [ + 66, + 71, + 211, + 300 + ], + "lines": [ + { + "bbox": [ + 66, + 71, + 211, + 300 + ], + "spans": [ + { + "bbox": [ + 66, + 71, + 211, + 300 + ], + "type": "image", + "image_path": "39042b8c3d03282cc29d73fb3162943ad0a2d1cbaa6ff6603b1476a41571a94c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 303, + 555, + 338 + ], + "lines": [ + { + "bbox": [ + 54, + 303, + 555, + 338 + ], + "spans": [ + { + "bbox": [ + 54, + 303, + 555, + 338 + ], + "type": "text", + "content": "Figure 3. The framework of the Sketchy-3DIS. Given a point cloud with sketchy bounding-box annotations, we first extract the backbone features using a 3D U-Net backbone, then feed them into the adaptive box-to-point pseudo labeler and the coarse-to-fine instance segmentator, finally, we utilize the generated high quality pseudo labels to supervise the predicted instances periodically." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 219, + 70, + 536, + 299 + ], + "blocks": [ + { + "bbox": [ + 219, + 70, + 536, + 299 + ], + "lines": [ + { + "bbox": [ + 219, + 70, + 536, + 299 + ], + "spans": [ + { + "bbox": [ + 219, + 70, + 536, + 299 + ], + "type": "image", + "image_path": "be8a8c9f7ca40e4e866edadb8704c88d36dd40fa88b0619efb6acbf007580224.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 347, + 296, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 347, + 296, + 407 + ], + "spans": [ + { + "bbox": [ + 55, + 347, + 296, + 407 + ], + "type": "text", + "content": "types of sketchy bounding boxes, named " + }, + { + "bbox": [ + 55, + 347, + 296, + 407 + ], + "type": "inline_equation", + "content": "S_{1}" + }, + { + "bbox": [ + 55, + 347, + 296, + 407 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 347, + 296, + 407 + ], + "type": "inline_equation", + "content": "S_{2}" + }, + { + "bbox": [ + 55, + 347, + 296, + 407 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 347, + 296, + 407 + ], + "type": "inline_equation", + "content": "S_{3}" + }, + { + "bbox": [ + 55, + 347, + 296, + 407 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 347, + 296, + 407 + ], + "type": "inline_equation", + "content": "S_{4}" + }, + { + "bbox": [ + 55, + 347, + 296, + 407 + ], + "type": "text", + "content": ", the red rectangle denotes the accurate box and the blue rectangles denote various sketchy bounding boxes. It can be observed that the \"sketchy degree\" increases from " + }, + { + "bbox": [ + 55, + 347, + 296, + 407 + ], + "type": "inline_equation", + "content": "S_{1}" + }, + { + "bbox": [ + 55, + 347, + 296, + 407 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 347, + 296, + 407 + ], + "type": "inline_equation", + "content": "S_{4}" + }, + { + "bbox": [ + 55, + 347, + 296, + 407 + ], + "type": "text", + "content": ", making instance segmentation more challenging." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 418, + 258, + 430 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 418, + 258, + 430 + ], + "spans": [ + { + "bbox": [ + 55, + 418, + 258, + 430 + ], + "type": "text", + "content": "3.3. Adaptive Box-to-Point Pseudo Labeler" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 436, + 295, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 436, + 295, + 544 + ], + "spans": [ + { + "bbox": [ + 54, + 436, + 295, + 544 + ], + "type": "text", + "content": "The sketchy 3D bounding box supervision is an inaccurate weak annotation for 3D instance segmentation. In order to convert rough annotated boxes into fine-grained point-level instance annotations, we propose an adaptive box-to-point pseudo labeler that can assign the points located in the box to the corresponding instance. Specifically, we first distinguish the points that lie in bounding boxes, and then adaptively assign them to the corresponding instance by learning the correlation between the points and 3D bounding boxes." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 545, + 295, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 295, + 700 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 295, + 700 + ], + "type": "text", + "content": "Box Conditioned Points Partition. According to the relationship between points and 3D bounding boxes, there are three types of points: the points located outside the bounding boxes (i.e., background points), the points lying in a single bounding box, and the points belonging to multiple bounding boxes. Since the background points (see black circle points in Fig. 3) are not within any 3D bounding boxes, they are directly labeled as background. The point lying in a single or multiple bounding box could be a point on the target (see green/yellow/blue circle points in Fig. 3) or a background point. Therefore, we need other strategies to help distinguish the target points from the background points." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "type": "text", + "content": "Adaptive Point-to-Instance Assignment. For the points" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 347, + 555, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 347, + 555, + 453 + ], + "spans": [ + { + "bbox": [ + 313, + 347, + 555, + 453 + ], + "type": "text", + "content": "lying in a single 3D bounding box (see bounding box " + }, + { + "bbox": [ + 313, + 347, + 555, + 453 + ], + "type": "inline_equation", + "content": "B_{3}" + }, + { + "bbox": [ + 313, + 347, + 555, + 453 + ], + "type": "text", + "content": " in Fig. 3), we filter the background points (i.e., black circle points in " + }, + { + "bbox": [ + 313, + 347, + 555, + 453 + ], + "type": "inline_equation", + "content": "B_{3}" + }, + { + "bbox": [ + 313, + 347, + 555, + 453 + ], + "type": "text", + "content": ") by comparing the similarity between the points and the 3D bounding box. Specifically, we consider the similarity in both coordinate space and feature space. The box coordinate and box feature is obtained by averaging the points' coordinates and features within them. The similarity between the points and the 3D bounding boxes can be formulated as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 369, + 456, + 555, + 471 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 369, + 456, + 555, + 471 + ], + "spans": [ + { + "bbox": [ + 369, + 456, + 555, + 471 + ], + "type": "interline_equation", + "content": "\\boldsymbol {s} _ {p, B} = \\cos \\left(\\boldsymbol {f} _ {p}, \\boldsymbol {f} _ {B}\\right) \\times e ^ {- \\left| \\boldsymbol {c} _ {B} \\boldsymbol {c} _ {p} \\right|} \\tag {3}", + "image_path": "9199c1fa438f630997b942d5f383c58117187ca2f4a3d596a1042ace691a04ed.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 472, + 554, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 472, + 554, + 519 + ], + "spans": [ + { + "bbox": [ + 313, + 472, + 554, + 519 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 472, + 554, + 519 + ], + "type": "inline_equation", + "content": "\\cos (\\cdot ,\\cdot)" + }, + { + "bbox": [ + 313, + 472, + 554, + 519 + ], + "type": "text", + "content": " denotes cosine similarity. In addition, " + }, + { + "bbox": [ + 313, + 472, + 554, + 519 + ], + "type": "inline_equation", + "content": "f_{p}" + }, + { + "bbox": [ + 313, + 472, + 554, + 519 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 472, + 554, + 519 + ], + "type": "inline_equation", + "content": "f_{B}" + }, + { + "bbox": [ + 313, + 472, + 554, + 519 + ], + "type": "text", + "content": " mean the features of point and box, while " + }, + { + "bbox": [ + 313, + 472, + 554, + 519 + ], + "type": "inline_equation", + "content": "c_{p}" + }, + { + "bbox": [ + 313, + 472, + 554, + 519 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 472, + 554, + 519 + ], + "type": "inline_equation", + "content": "c_{B}" + }, + { + "bbox": [ + 313, + 472, + 554, + 519 + ], + "type": "text", + "content": " mean the coordinates of point and box. In this way, it is expected to filter background points with low similarity." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 520, + 556, + 699 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 520, + 556, + 699 + ], + "spans": [ + { + "bbox": [ + 313, + 520, + 556, + 699 + ], + "type": "text", + "content": "For the points lying in the overlapped parts of multiple 3D bounding boxes (usually two boxes, see boxes " + }, + { + "bbox": [ + 313, + 520, + 556, + 699 + ], + "type": "inline_equation", + "content": "B_{1}" + }, + { + "bbox": [ + 313, + 520, + 556, + 699 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 520, + 556, + 699 + ], + "type": "inline_equation", + "content": "B_{2}" + }, + { + "bbox": [ + 313, + 520, + 556, + 699 + ], + "type": "text", + "content": " in Fig. 3), they are assigned to the corresponding instances by learning the similarity between the points and the bounding boxes. However, due to the uncertainty of point labels in overlapping areas, it is difficult to obtain high-confidence semantic labels for these points from the bounding boxes. Therefore, we consider removing the overlapping points within each box and using the remaining reliable points (see " + }, + { + "bbox": [ + 313, + 520, + 556, + 699 + ], + "type": "inline_equation", + "content": "B_{1}^{rel}" + }, + { + "bbox": [ + 313, + 520, + 556, + 699 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 520, + 556, + 699 + ], + "type": "inline_equation", + "content": "B_{2}^{rel}" + }, + { + "bbox": [ + 313, + 520, + 556, + 699 + ], + "type": "text", + "content": " in Fig. 3) for similarity learning. Since reliable points are only located within one box, their labels are consistent with the box labels. Given a point feature " + }, + { + "bbox": [ + 313, + 520, + 556, + 699 + ], + "type": "inline_equation", + "content": "\\pmb{f}_{p}" + }, + { + "bbox": [ + 313, + 520, + 556, + 699 + ], + "type": "text", + "content": ", box features " + }, + { + "bbox": [ + 313, + 520, + 556, + 699 + ], + "type": "inline_equation", + "content": "\\pmb{f}_{B_1^{rel}}" + }, + { + "bbox": [ + 313, + 520, + 556, + 699 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 520, + 556, + 699 + ], + "type": "inline_equation", + "content": "\\pmb{f}_{B_2^{rel}}" + }, + { + "bbox": [ + 313, + 520, + 556, + 699 + ], + "type": "text", + "content": ", we utilize a multilayer perceptron (MLP) network to learn the similarity score " + }, + { + "bbox": [ + 313, + 520, + 556, + 699 + ], + "type": "inline_equation", + "content": "A_{p2b}" + }, + { + "bbox": [ + 313, + 520, + 556, + 699 + ], + "type": "text", + "content": " which is given by:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 346, + 700, + 555, + 717 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 346, + 700, + 555, + 717 + ], + "spans": [ + { + "bbox": [ + 346, + 700, + 555, + 717 + ], + "type": "interline_equation", + "content": "\\boldsymbol {A} _ {p 2 b} = \\mathrm {M L P} (\\boldsymbol {f} _ {p}, \\boldsymbol {f} _ {p} - \\boldsymbol {f} _ {B _ {1} ^ {r e l}}, \\boldsymbol {f} _ {p} - \\boldsymbol {f} _ {B _ {2} ^ {r e l}}) \\qquad (4)", + "image_path": "7525e1d46ea4f07975da4272b35c03b1d032bc613f78973677d33c31b82a2676.jpg" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8882" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 70, + 294, + 249 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 294, + 249 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 294, + 249 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 294, + 249 + ], + "type": "image", + "image_path": "78d5395f99d8466fd8903160a90303837105d32f65595a7eb5b1b017880c8fd2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 255, + 296, + 301 + ], + "lines": [ + { + "bbox": [ + 55, + 255, + 296, + 301 + ], + "spans": [ + { + "bbox": [ + 55, + 255, + 296, + 301 + ], + "type": "text", + "content": "Figure 4. The details of Multi-level Attention Block. The instance queries interact with the features of the whole scene, the coarse instance regions, and the instance core regions hierarchically. And the " + }, + { + "bbox": [ + 55, + 255, + 296, + 301 + ], + "type": "inline_equation", + "content": "B_{mask}" + }, + { + "bbox": [ + 55, + 255, + 296, + 301 + ], + "type": "text", + "content": " are the boxes obtained from the predicted masks." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 311, + 296, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 311, + 296, + 395 + ], + "spans": [ + { + "bbox": [ + 55, + 311, + 296, + 395 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 311, + 296, + 395 + ], + "type": "inline_equation", + "content": "\\pmb{f}_{p} - \\pmb{f}_{B_{1}^{rel}}" + }, + { + "bbox": [ + 55, + 311, + 296, + 395 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 311, + 296, + 395 + ], + "type": "inline_equation", + "content": "\\pmb{f}_{p} - \\pmb{f}_{B_{2}^{rel}}" + }, + { + "bbox": [ + 55, + 311, + 296, + 395 + ], + "type": "text", + "content": " capture the difference between the point and boxes. In addition, the box feature is obtained by averaging the point features within it. Note that before computing the box feature, we eliminate the background points in the box through Eq. (3). To supervise the similarity learning, we use the cross-entropy loss to formulate pseudo label loss " + }, + { + "bbox": [ + 55, + 311, + 296, + 395 + ], + "type": "inline_equation", + "content": "L_{pl}" + }, + { + "bbox": [ + 55, + 311, + 296, + 395 + ], + "type": "text", + "content": ", which is defined as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 135, + 396, + 296, + 410 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 396, + 296, + 410 + ], + "spans": [ + { + "bbox": [ + 135, + 396, + 296, + 410 + ], + "type": "interline_equation", + "content": "L _ {p l} = L _ {C E} (\\mathbf {A}, \\mathbf {Y}) \\tag {5}", + "image_path": "a4006443bc548833590af6f6189a1cf30cb77c0b7753ffccb262b3cbce5df2b8.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 411, + 296, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 411, + 296, + 434 + ], + "spans": [ + { + "bbox": [ + 55, + 411, + 296, + 434 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 411, + 296, + 434 + ], + "type": "inline_equation", + "content": "\\pmb{Y}" + }, + { + "bbox": [ + 55, + 411, + 296, + 434 + ], + "type": "text", + "content": " is the label of the reliable point, which denotes the box it belongs to." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 435, + 296, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 435, + 296, + 506 + ], + "spans": [ + { + "bbox": [ + 55, + 435, + 296, + 506 + ], + "type": "text", + "content": "After that, we use the model trained on the reliable points (see " + }, + { + "bbox": [ + 55, + 435, + 296, + 506 + ], + "type": "inline_equation", + "content": "B_1^{rel}" + }, + { + "bbox": [ + 55, + 435, + 296, + 506 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 435, + 296, + 506 + ], + "type": "inline_equation", + "content": "B_2^{rel}" + }, + { + "bbox": [ + 55, + 435, + 296, + 506 + ], + "type": "text", + "content": " in Fig. 3) to predict the label of unreliable points (see " + }, + { + "bbox": [ + 55, + 435, + 296, + 506 + ], + "type": "inline_equation", + "content": "B_1^{un}" + }, + { + "bbox": [ + 55, + 435, + 296, + 506 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 435, + 296, + 506 + ], + "type": "inline_equation", + "content": "B_2^{un}" + }, + { + "bbox": [ + 55, + 435, + 296, + 506 + ], + "type": "text", + "content": " in Fig. 3). It is desired that the points located in the overlapping part of the box can be assigned to the corresponding instances, obtaining pseudo instance labels." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 516, + 253, + 529 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 516, + 253, + 529 + ], + "spans": [ + { + "bbox": [ + 55, + 516, + 253, + 529 + ], + "type": "text", + "content": "3.4. Coarse-to-Fine Instance Segmentator" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 533, + 296, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 533, + 296, + 640 + ], + "spans": [ + { + "bbox": [ + 55, + 533, + 296, + 640 + ], + "type": "text", + "content": "The bounding box supervision provides coarse shape information for 3D instance segmentation. To obtain well-detailed instances, we present a coarse-to-fine instance segmentator that can recognize and segment the points in the point cloud into various instances. Specifically, we first employ the queries to learn from the whole point cloud to obtain coarse instances, and then these queries learn from the regions indicated by the coarse instances with Multi-level Attention Blocks to refine the instances." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": "Coarse Instances Segmentation. Supposing one scene consists of " + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": " objects, and each query represents one object and models the instance-related characteristics: the semantic category and the mask of each instance. In our method, these queries first interact with the backbone features of the whole point cloud to obtain coarse instances. Considering" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 555, + 143 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 143 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 143 + ], + "type": "text", + "content": "the point-wise masks of the coarse instances can not describe one object holistically and precisely, we predict another two boxes with coarse granularity and fine granularity, respectively. Then these two boxes combined with the scene-level features to refine the initial coarse instances with Multi-level Attention Blocks." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 144, + 556, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 144, + 556, + 239 + ], + "spans": [ + { + "bbox": [ + 313, + 144, + 556, + 239 + ], + "type": "text", + "content": "Hierarchical Instances Refinement. To leverage the advantages of multi-granularity features from the point cloud and the segmented coarse instances, we introduce a novel decoder structure utilizing attentions to make the instance queries interact with granularity-varied features hierarchically. The decoder consists of a stack of six blocks. Each block includes a Multi-level Attention Block and an instance segmentation head illustrated in Fig. 3." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 239, + 556, + 394 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 239, + 556, + 394 + ], + "spans": [ + { + "bbox": [ + 313, + 239, + 556, + 394 + ], + "type": "text", + "content": "Multi-level Attention Block. The details of the Multi-level Attention Block are shown in Fig. 4. These instance queries first interact with point-wise backbone features globally, then interact with regions described by the coarse instance locally. The regions described by the coarse instance are divided into reliable and unreliable, which are depicted by the predicted box and the core box, respectively. To fully explore various granularity of shape information implied by the coarse instances and those two boxes, we first abstract the obtained coarse instance into a box, termed as " + }, + { + "bbox": [ + 313, + 239, + 556, + 394 + ], + "type": "inline_equation", + "content": "B_{mask}" + }, + { + "bbox": [ + 313, + 239, + 556, + 394 + ], + "type": "text", + "content": ". Then we explore the consistent information between " + }, + { + "bbox": [ + 313, + 239, + 556, + 394 + ], + "type": "inline_equation", + "content": "B_{mask}" + }, + { + "bbox": [ + 313, + 239, + 556, + 394 + ], + "type": "text", + "content": " and the predicted box " + }, + { + "bbox": [ + 313, + 239, + 556, + 394 + ], + "type": "inline_equation", + "content": "B_{pred}" + }, + { + "bbox": [ + 313, + 239, + 556, + 394 + ], + "type": "text", + "content": " to obtain features of reliable regions, which are described as follows:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 353, + 395, + 555, + 410 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 395, + 555, + 410 + ], + "spans": [ + { + "bbox": [ + 353, + 395, + 555, + 410 + ], + "type": "interline_equation", + "content": "\\boldsymbol {F} ^ {r e l} = \\sigma \\left(\\boldsymbol {F}, \\boldsymbol {M} \\odot e ^ {I o U \\left(\\boldsymbol {B} _ {p r e d}, \\boldsymbol {B} _ {m a s k}\\right)}\\right) \\tag {6}", + "image_path": "11c5a84a7c67fdfabc2d0f9684a9689f4d6848024c5c34dfa8559c6a9f15288d.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 411, + 556, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 411, + 556, + 564 + ], + "spans": [ + { + "bbox": [ + 313, + 411, + 556, + 564 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 411, + 556, + 564 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 313, + 411, + 556, + 564 + ], + "type": "text", + "content": " denotes the thresholding operation and " + }, + { + "bbox": [ + 313, + 411, + 556, + 564 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 313, + 411, + 556, + 564 + ], + "type": "text", + "content": " denotes point-wise multiplication. In addition, " + }, + { + "bbox": [ + 313, + 411, + 556, + 564 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 313, + 411, + 556, + 564 + ], + "type": "text", + "content": " denotes the backbone features, " + }, + { + "bbox": [ + 313, + 411, + 556, + 564 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 411, + 556, + 564 + ], + "type": "text", + "content": " denotes the predicted mask probabilities. " + }, + { + "bbox": [ + 313, + 411, + 556, + 564 + ], + "type": "inline_equation", + "content": "B_{pred}" + }, + { + "bbox": [ + 313, + 411, + 556, + 564 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 411, + 556, + 564 + ], + "type": "inline_equation", + "content": "B_{mask}" + }, + { + "bbox": [ + 313, + 411, + 556, + 564 + ], + "type": "text", + "content": " denote the predicted boxes and boxes decided by predicted masks, respectively. So far, the instance queries have learned from the entire point cloud and the reliable regions, they can segment the majority of objects with easy contexts. To tackle the complex contexts, the instance queries learn from the unreliable regions described by the predicted core boxes. The predicted core boxes are scaled from the predicted boxes, aiming to represent the core regions of objects, and corresponding features are obtained as follows:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 354, + 568, + 555, + 581 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 568, + 555, + 581 + ], + "spans": [ + { + "bbox": [ + 354, + 568, + 555, + 581 + ], + "type": "interline_equation", + "content": "\\boldsymbol {F} ^ {u n} = \\sigma (\\boldsymbol {F}, \\boldsymbol {M} \\odot (\\boldsymbol {B} _ {\\text {c o r e}} \\cap \\boldsymbol {B} _ {\\text {m a s k}})) \\tag {7}", + "image_path": "7f311bda45cc4496c4f147617cc53e8b2463527ffc5f6da7b36ccfd8266966b1.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 582, + 555, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 555, + 643 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 555, + 643 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 582, + 555, + 643 + ], + "type": "inline_equation", + "content": "B_{core}" + }, + { + "bbox": [ + 313, + 582, + 555, + 643 + ], + "type": "text", + "content": " denotes the predicted core boxes. The instance queries have strong discriminative ability in such a global-to-local manner, then they are fed into a self-attention module and a feed-forward network to aggregate the learned information, which further boosts the performance." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 647, + 515, + 661 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 647, + 515, + 661 + ], + "spans": [ + { + "bbox": [ + 313, + 647, + 515, + 661 + ], + "type": "text", + "content": "3.5. Bilateral Matching and Training Loss" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 665, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 556, + 714 + ], + "type": "text", + "content": "Our method makes bilateral matching between pseudo instances and predicted instances first and then utilizes the paired pseudo instances as ground truth " + }, + { + "bbox": [ + 313, + 665, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_{ins}" + }, + { + "bbox": [ + 313, + 665, + 556, + 714 + ], + "type": "text", + "content": " to supervise the paired predicted instances " + }, + { + "bbox": [ + 313, + 665, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{ins}" + }, + { + "bbox": [ + 313, + 665, + 556, + 714 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8883" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 155 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 155 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 155 + ], + "type": "text", + "content": "Bilateral matching. Following SPFormer [36] and IS-BNet [30], our method utilizes the Hungarian method [18] to achieve the optimal assignment. We compute a pairwise matching cost " + }, + { + "bbox": [ + 55, + 72, + 296, + 155 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_{i,j}" + }, + { + "bbox": [ + 55, + 72, + 296, + 155 + ], + "type": "text", + "content": " to evaluate the similarity of the " + }, + { + "bbox": [ + 55, + 72, + 296, + 155 + ], + "type": "inline_equation", + "content": "i^{th}" + }, + { + "bbox": [ + 55, + 72, + 296, + 155 + ], + "type": "text", + "content": " predicted instance and the " + }, + { + "bbox": [ + 55, + 72, + 296, + 155 + ], + "type": "inline_equation", + "content": "j^{th}" + }, + { + "bbox": [ + 55, + 72, + 296, + 155 + ], + "type": "text", + "content": " pseudo ground truth. And " + }, + { + "bbox": [ + 55, + 72, + 296, + 155 + ], + "type": "inline_equation", + "content": "\\mathbf{c}^{i,j}" + }, + { + "bbox": [ + 55, + 72, + 296, + 155 + ], + "type": "text", + "content": " is based on the classification probabilities and the binary masks, it is defined as:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 156, + 295, + 198 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 156, + 295, + 198 + ], + "spans": [ + { + "bbox": [ + 56, + 156, + 295, + 198 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathbf {c} ^ {i, j} = \\lambda_ {1} L _ {C E} (\\mathbf {x} _ {c l s} ^ {i}, \\mathbf {y} _ {c l s} ^ {j}) \\\\ + \\lambda_ {2} \\left(L _ {B C E} \\left(\\mathbf {x} _ {\\text {m a s k}} ^ {i}, \\mathbf {y} _ {\\text {m a s k}} ^ {j}\\right) + L _ {\\text {d i c e}} \\left(\\mathbf {x} _ {\\text {m a s k}} ^ {i}, \\mathbf {y} _ {\\text {m a s k}} ^ {j}\\right)\\right) \\tag {8} \\\\ \\end{array}", + "image_path": "77e0133b6dff03734ba9fe7cc6f78a750ab717beac7b171d3abb8a30cfd5b7d5.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 198, + 296, + 235 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 198, + 296, + 235 + ], + "spans": [ + { + "bbox": [ + 55, + 198, + 296, + 235 + ], + "type": "text", + "content": "Training loss. Our training loss consists of two items, namely the loss of pseudo label generation " + }, + { + "bbox": [ + 55, + 198, + 296, + 235 + ], + "type": "inline_equation", + "content": "L_{pl}" + }, + { + "bbox": [ + 55, + 198, + 296, + 235 + ], + "type": "text", + "content": " and the loss of instance segmentation " + }, + { + "bbox": [ + 55, + 198, + 296, + 235 + ], + "type": "inline_equation", + "content": "L_{seg}" + }, + { + "bbox": [ + 55, + 198, + 296, + 235 + ], + "type": "text", + "content": ", which is defined as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 141, + 236, + 295, + 249 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 236, + 295, + 249 + ], + "spans": [ + { + "bbox": [ + 141, + 236, + 295, + 249 + ], + "type": "interline_equation", + "content": "L = L _ {p l} + L _ {s e g} \\tag {9}", + "image_path": "5ff3527eda3c2c6ba09a8a6366677ddf859853edb3eae2377c88481f6633b287.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 251, + 296, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 251, + 296, + 334 + ], + "spans": [ + { + "bbox": [ + 55, + 251, + 296, + 334 + ], + "type": "text", + "content": "As for the instance segmentation, except for supervising the predicted semantic classes and the binary masks, we also supervise the predicted bounding boxes and the core boxes utilized in the coarse-to-fine instance segmentator. The ground truth of the scale of the core box is based on the IoU score of the predicted masks and the pseudo-ground truth masks. The loss of instance segmentation is defined as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 60, + 335, + 295, + 388 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 335, + 295, + 388 + ], + "spans": [ + { + "bbox": [ + 60, + 335, + 295, + 388 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} L _ {s e g} = \\lambda_ {1} L _ {C E} (\\mathbf {x} _ {c l s}, \\mathbf {y} _ {c l s}) \\\\ + \\lambda_ {2} \\left(L _ {B C E} \\left(\\mathbf {x} _ {\\text {m a s k}}, \\mathbf {y} _ {\\text {m a s k}}\\right) + L _ {\\text {d i c e}} \\left(\\mathbf {x} _ {\\text {m a s k}}, \\mathbf {y} _ {\\text {m a s k}}\\right)\\right) \\\\ + \\lambda_ {3} \\left(L _ {L _ {1}} \\left(\\mathbf {x} _ {\\text {b o x}}, \\mathbf {y} _ {\\text {b o x}}\\right) + L _ {M S E} \\left(\\mathbf {x} _ {\\text {c o r e - b o x}}, \\mathbf {y} _ {\\text {c o r e - b o x}}\\right)\\right) \\tag {10} \\\\ \\end{array}", + "image_path": "022d7f03e5da7599a6c91f007d2eddb87f1fe83b6013e58d070125b81d6f1bc9.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 400, + 137, + 414 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 400, + 137, + 414 + ], + "spans": [ + { + "bbox": [ + 55, + 400, + 137, + 414 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 420, + 172, + 433 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 420, + 172, + 433 + ], + "spans": [ + { + "bbox": [ + 55, + 420, + 172, + 433 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 437, + 296, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 437, + 296, + 557 + ], + "spans": [ + { + "bbox": [ + 55, + 437, + 296, + 557 + ], + "type": "text", + "content": "Datasets. We conduct experiments on two datasets, ScanNetV2 [6] and S3DIS [1]. The ScanNetV2 provides splits containing 1201, 312, and 100 scenes for training, validation, and testing. Scenes are annotated with semantic labels covering 18 categories and instance labels. And the S3DIS dataset is another indoor dataset contains 272 scenes across six different areas, and it is annotated with 13 object categories. Following existing methods [4, 27, 29], the scenes of S3DIS Area 5 are used for validation while the others are used for training." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 558, + 296, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 296, + 653 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 296, + 653 + ], + "type": "text", + "content": "Evaluation metrics. We use the mean average precision " + }, + { + "bbox": [ + 55, + 558, + 296, + 653 + ], + "type": "inline_equation", + "content": "(AP)" + }, + { + "bbox": [ + 55, + 558, + 296, + 653 + ], + "type": "text", + "content": " as the evaluation metric for both ScanNetV2 and S3DIS, which is utilized as a common evaluation metric in 3D instance segmentation. " + }, + { + "bbox": [ + 55, + 558, + 296, + 653 + ], + "type": "inline_equation", + "content": "AP" + }, + { + "bbox": [ + 55, + 558, + 296, + 653 + ], + "type": "text", + "content": " is the averaged score with IoU thresholds from " + }, + { + "bbox": [ + 55, + 558, + 296, + 653 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 55, + 558, + 296, + 653 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 558, + 296, + 653 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 55, + 558, + 296, + 653 + ], + "type": "text", + "content": " with a step size of " + }, + { + "bbox": [ + 55, + 558, + 296, + 653 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 55, + 558, + 296, + 653 + ], + "type": "text", + "content": ". The mean average precision with IoU thresholds at " + }, + { + "bbox": [ + 55, + 558, + 296, + 653 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 55, + 558, + 296, + 653 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 558, + 296, + 653 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 55, + 558, + 296, + 653 + ], + "type": "text", + "content": " are used as evaluation metrics as well, which are denoted as " + }, + { + "bbox": [ + 55, + 558, + 296, + 653 + ], + "type": "inline_equation", + "content": "AP_{50}" + }, + { + "bbox": [ + 55, + 558, + 296, + 653 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 558, + 296, + 653 + ], + "type": "inline_equation", + "content": "AP_{25}" + }, + { + "bbox": [ + 55, + 558, + 296, + 653 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": "Implementation details. Our model is trained on a single RTX 3090 GPU, and we utilize SPFormer [36] and ISBNet [30] as codebases for ScanNetV2 and S3DIS, respectively. We utilize the AdamW optimizer with a learning rate of 0.0002 and weight decay at 0.05 for the training of Scan-" + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 315, + 105, + 553, + 303 + ], + "blocks": [ + { + "bbox": [ + 313, + 71, + 555, + 103 + ], + "lines": [ + { + "bbox": [ + 313, + 71, + 555, + 103 + ], + "spans": [ + { + "bbox": [ + 313, + 71, + 555, + 103 + ], + "type": "text", + "content": "Table 1. Comparison on ScanNetV2. \"Mask\" denotes using dense point-wise annotations, " + }, + { + "bbox": [ + 313, + 71, + 555, + 103 + ], + "type": "inline_equation", + "content": "S_0" + }, + { + "bbox": [ + 313, + 71, + 555, + 103 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 71, + 555, + 103 + ], + "type": "inline_equation", + "content": "S_1" + }, + { + "bbox": [ + 313, + 71, + 555, + 103 + ], + "type": "text", + "content": " denote using accurate bounding box and sketchy bounding box annotations, respectively." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 105, + 553, + 303 + ], + "lines": [ + { + "bbox": [ + 315, + 105, + 553, + 303 + ], + "spans": [ + { + "bbox": [ + 315, + 105, + 553, + 303 + ], + "type": "table", + "html": "
MethodSup.VenueScanNet TestScanNet Val
AP50AP25AP50AP25
GraphCut [13]NIPS2273.283.269.179.3
ISBNet [30]CVPR2376.384.573.182.5
SPFormer [36]MaskAAAI2377.085.173.982.9
Mask3D [34]ICRA2378.087.073.783.5
MSTA3D [40]ACM MM2479.587.977.085.4
Spherical Mask [35]CVPR2481.2-79.988.2
Box2Mask [4]ECCV2267.780.359.771.8
WISGP+PointGroup [8]WACV23--50.264.9
WISGP+SSTNet [8]WACV23--56.970.2
GaPro+SPFormer [29]S0ICCV2369.282.470.479.9
CIP-WPIS [49]WACV24--69.378.6
BSNet+SPFormer [27]CVPR24--72.783.4
Sketchy-3DIS (ours)-70.186.668.883.6
Box2Mask [4]ECCV22--52.467.5
GaPro+SPFormer [29]S1ICCV23--53.572.2
Sketchy-3DIS (ours)---65.883.1
", + "image_path": "6db695164da17199b3bc217c7dd01514e877592dbd741c1b2b368b7b99acfd18.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 316, + 555, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 316, + 555, + 376 + ], + "spans": [ + { + "bbox": [ + 313, + 316, + 555, + 376 + ], + "type": "text", + "content": "NetV2. As for the weights of matching cost and losses, " + }, + { + "bbox": [ + 313, + 316, + 555, + 376 + ], + "type": "inline_equation", + "content": "(\\lambda_{1},\\lambda_{2},\\lambda_{3})" + }, + { + "bbox": [ + 313, + 316, + 555, + 376 + ], + "type": "text", + "content": " are set as (0.5, 1.0, 0.5). All the models are trained from scratch, and without specification, the hyperparameters and the training details are kept the same as the [36] or [30]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 384, + 400, + 396 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 384, + 400, + 396 + ], + "spans": [ + { + "bbox": [ + 313, + 384, + 400, + 396 + ], + "type": "text", + "content": "4.2. Main Results" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 402, + 556, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 402, + 556, + 545 + ], + "spans": [ + { + "bbox": [ + 313, + 402, + 556, + 545 + ], + "type": "text", + "content": "ScanNetV2. In Table 1, we compare our method and existing state-of-the-art methods on ScanNetV2 dataset. Attributed to the innovative design of Sketchy-3DIS, which achieves the pseudo labeling and instance segmenting simultaneously, our method achieves considerable performance on both the validation set and online test set of ScanNetV2. Notably, without multi-modal priors [49] or synthesized scenes [27], our method outperforms existing methods in terms of " + }, + { + "bbox": [ + 313, + 402, + 556, + 545 + ], + "type": "inline_equation", + "content": "AP_{25}" + }, + { + "bbox": [ + 313, + 402, + 556, + 545 + ], + "type": "text", + "content": ", which demonstrates the effectiveness of our method. Additionally, the comparable performances to fully-supervised methods demonstrate that our method can play a significant role in autonomous driving with high efficacy." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 545, + 556, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 556, + 641 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 556, + 641 + ], + "type": "text", + "content": "S3DIS. In Table 2, we make comparisons between existing methods and our proposed Sketchy-3DIS on S3DIS Area 5. The consistent advantages demonstrate the strong generalization ability of our proposed SKeychy-3DIS. Notably, our method exceeds the fully-supervised baseline [30] with " + }, + { + "bbox": [ + 313, + 545, + 556, + 641 + ], + "type": "inline_equation", + "content": "+3.3" + }, + { + "bbox": [ + 313, + 545, + 556, + 641 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 313, + 545, + 556, + 641 + ], + "type": "inline_equation", + "content": "AP_{50}" + }, + { + "bbox": [ + 313, + 545, + 556, + 641 + ], + "type": "text", + "content": ", which verifies that our proposed Sketchy-3DIS can make full use of instance indications provided by the sketchy bounding boxes." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "content": "Quality comparisons of pseudo labels. We compare the qualities of the generated pseudo labels of GaPro [29] and ours under " + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "inline_equation", + "content": "S_{1}" + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "content": " sketchy bounding boxes. The qualitative comparisons are present in Fig.5, which illustrates that our method generates more accurate pseudo labels than GaPro not only in details (highlighted by the green cycles) but also" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8884" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 86, + 91, + 192, + 156 + ], + "blocks": [ + { + "bbox": [ + 107, + 74, + 173, + 85 + ], + "lines": [ + { + "bbox": [ + 107, + 74, + 173, + 85 + ], + "spans": [ + { + "bbox": [ + 107, + 74, + 173, + 85 + ], + "type": "text", + "content": "Input Point Cloud" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 86, + 91, + 192, + 156 + ], + "lines": [ + { + "bbox": [ + 86, + 91, + 192, + 156 + ], + "spans": [ + { + "bbox": [ + 86, + 91, + 192, + 156 + ], + "type": "image", + "image_path": "9d880ff7dafa0db6e88c7c18acd2350247eb866e1c757d8bdd714155c09b071b.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 194, + 92, + 302, + 157 + ], + "blocks": [ + { + "bbox": [ + 223, + 74, + 274, + 84 + ], + "lines": [ + { + "bbox": [ + 223, + 74, + 274, + 84 + ], + "spans": [ + { + "bbox": [ + 223, + 74, + 274, + 84 + ], + "type": "text", + "content": "Ground Truth" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 194, + 92, + 302, + 157 + ], + "lines": [ + { + "bbox": [ + 194, + 92, + 302, + 157 + ], + "spans": [ + { + "bbox": [ + 194, + 92, + 302, + 157 + ], + "type": "image", + "image_path": "2f8055b6abed0018c5e17984d2b3e664e21fb965914e4b24a699fc188c56cacd.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 304, + 92, + 411, + 157 + ], + "blocks": [ + { + "bbox": [ + 319, + 74, + 395, + 85 + ], + "lines": [ + { + "bbox": [ + 319, + 74, + 395, + 85 + ], + "spans": [ + { + "bbox": [ + 319, + 74, + 395, + 85 + ], + "type": "text", + "content": "Sketchy-3DIS (ours)" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 92, + 411, + 157 + ], + "lines": [ + { + "bbox": [ + 304, + 92, + 411, + 157 + ], + "spans": [ + { + "bbox": [ + 304, + 92, + 411, + 157 + ], + "type": "image", + "image_path": "e8b5595da1616053db9f43a0d70cd1da605da8d27396ddf7d88f859c0f1ed442.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 413, + 92, + 522, + 157 + ], + "blocks": [ + { + "bbox": [ + 455, + 74, + 480, + 84 + ], + "lines": [ + { + "bbox": [ + 455, + 74, + 480, + 84 + ], + "spans": [ + { + "bbox": [ + 455, + 74, + 480, + 84 + ], + "type": "text", + "content": "GaPro" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 413, + 92, + 522, + 157 + ], + "lines": [ + { + "bbox": [ + 413, + 92, + 522, + 157 + ], + "spans": [ + { + "bbox": [ + 413, + 92, + 522, + 157 + ], + "type": "image", + "image_path": "94d30bdbd197adefb9b0e0dc886f4841f97056e2356d5747e06844ecd209e629.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 86, + 157, + 193, + 241 + ], + "blocks": [ + { + "bbox": [ + 86, + 157, + 193, + 241 + ], + "lines": [ + { + "bbox": [ + 86, + 157, + 193, + 241 + ], + "spans": [ + { + "bbox": [ + 86, + 157, + 193, + 241 + ], + "type": "image", + "image_path": "d4149b60152f290578ac3948825238b5c10583b29683a74323626f5279cfe5e4.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 243, + 555, + 266 + ], + "lines": [ + { + "bbox": [ + 55, + 243, + 555, + 266 + ], + "spans": [ + { + "bbox": [ + 55, + 243, + 555, + 266 + ], + "type": "text", + "content": "Figure 5. Visualization comparison of pseudo labels on the ScanNetV2 training set. The black denotes the background points and other colors denote different objects. The green and red cycles highlight the key regions." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 194, + 158, + 301, + 242 + ], + "blocks": [ + { + "bbox": [ + 194, + 158, + 301, + 242 + ], + "lines": [ + { + "bbox": [ + 194, + 158, + 301, + 242 + ], + "spans": [ + { + "bbox": [ + 194, + 158, + 301, + 242 + ], + "type": "image", + "image_path": "c433c3e291a7813c54998067d0bad0c89b0ad96370b046f2b32099092a531fc6.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 304, + 158, + 411, + 242 + ], + "blocks": [ + { + "bbox": [ + 304, + 158, + 411, + 242 + ], + "lines": [ + { + "bbox": [ + 304, + 158, + 411, + 242 + ], + "spans": [ + { + "bbox": [ + 304, + 158, + 411, + 242 + ], + "type": "image", + "image_path": "5419585a74f0764d2ecc0a5c26c40639c0ea2143be49371382b8c685b90f8a25.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 413, + 158, + 520, + 241 + ], + "blocks": [ + { + "bbox": [ + 413, + 158, + 520, + 241 + ], + "lines": [ + { + "bbox": [ + 413, + 158, + 520, + 241 + ], + "spans": [ + { + "bbox": [ + 413, + 158, + 520, + 241 + ], + "type": "image", + "image_path": "e69f7be396bd3c10f8034e3cd8f6430f9397a057f21c2fffa30d14e7be24d259.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 56, + 320, + 296, + 496 + ], + "blocks": [ + { + "bbox": [ + 55, + 275, + 296, + 319 + ], + "lines": [ + { + "bbox": [ + 55, + 275, + 296, + 319 + ], + "spans": [ + { + "bbox": [ + 55, + 275, + 296, + 319 + ], + "type": "text", + "content": "Table 2. Comparison on S3DIS Area 5. \"Mask\" denotes using dense point-wise annotations, " + }, + { + "bbox": [ + 55, + 275, + 296, + 319 + ], + "type": "inline_equation", + "content": "S_0" + }, + { + "bbox": [ + 55, + 275, + 296, + 319 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 275, + 296, + 319 + ], + "type": "inline_equation", + "content": "S_1" + }, + { + "bbox": [ + 55, + 275, + 296, + 319 + ], + "type": "text", + "content": " denote using accurate bounding box and sketchy bounding box annotations, respectively. \\*\\*\" denotes the results are reproduced by GaPro [29]." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 320, + 296, + 496 + ], + "lines": [ + { + "bbox": [ + 56, + 320, + 296, + 496 + ], + "spans": [ + { + "bbox": [ + 56, + 320, + 296, + 496 + ], + "type": "table", + "html": "
MethodSup.VenueAPAP50
SoftGroup [41]CVPR2251.666.1
GraphCut [13]NIPS2254.166.4
ISBNet [30]MaskCVPR2354.065.8
Mask3D [34]ICRA2356.668.4
MSTA3D [40]ACM MM24-70.0
Spherical Mask [35]CVPR2460.572.3
Box2Mask* [4]ECCV2243.654.6
WISGP+PointGroup [8]WACV2333.548.6
WISGP+SSTNet [8]S0WACV2337.251.0
GaPro+ISBNet [29]ICCV2350.561.2
BSNet+ISBNet [27]CVPR2453.064.3
Sketchy-3DIS (ours)-53.469.1
Sketchy-3DIS (ours)S1-50.764.6
", + "image_path": "69f29c14fbfdf8fbdf64b6311e8ec82046750848a38c298e3021cfdd57171ba1.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 509, + 295, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 509, + 295, + 555 + ], + "spans": [ + { + "bbox": [ + 55, + 509, + 295, + 555 + ], + "type": "text", + "content": "in distinguishing the objects belonging to the same categories (highlighted by the red cycles). Additionally, compared with the ground truth, our method generates comparable pseudo labels as well." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 55, + 564, + 219, + 577 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 564, + 219, + 577 + ], + "spans": [ + { + "bbox": [ + 55, + 564, + 219, + 577 + ], + "type": "text", + "content": "4.3. Ablation Studies and Analysis" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 55, + 582, + 296, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 296, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 296, + 628 + ], + "type": "text", + "content": "We conduct comprehensive ablation studies to justify the effective design of our method. All these ablation experiments are conducted on the ScanNetV2 validation set unless otherwise stated." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "text", + "content": "Robustness under various sketchy bounding boxes. In Table 3, we evaluate our method under four kinds of various sketchy bounding boxes illustrated in Fig.2 on both ScanNetV2 validation set and S3DIS Area 5. It can be observed that the performance of our method deteriorates slightly as the increasing of the boxes' sketchy degree, which verifies the robustness of our proposed Sketchy-3DIS." + } + ] + } + ], + "index": 18 + }, + { + "type": "table", + "bbox": [ + 316, + 299, + 553, + 384 + ], + "blocks": [ + { + "bbox": [ + 313, + 275, + 553, + 298 + ], + "lines": [ + { + "bbox": [ + 313, + 275, + 553, + 298 + ], + "spans": [ + { + "bbox": [ + 313, + 275, + 553, + 298 + ], + "type": "text", + "content": "Table 3. Robustness under various sketchy bounding boxes. " + }, + { + "bbox": [ + 313, + 275, + 553, + 298 + ], + "type": "inline_equation", + "content": "S_0" + }, + { + "bbox": [ + 313, + 275, + 553, + 298 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 275, + 553, + 298 + ], + "type": "inline_equation", + "content": "S_4" + }, + { + "bbox": [ + 313, + 275, + 553, + 298 + ], + "type": "text", + "content": " denote the sketchy degree of the annotated bounding boxes." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 316, + 299, + 553, + 384 + ], + "lines": [ + { + "bbox": [ + 316, + 299, + 553, + 384 + ], + "spans": [ + { + "bbox": [ + 316, + 299, + 553, + 384 + ], + "type": "table", + "html": "
Sketchy DegreeScanNetV2 ValS3DIS Area 5
AP50AP25AP50AP25
S462.580.150.963.6
S362.981.452.366.6
S263.782.161.272.7
S165.883.164.673.9
S068.883.669.177.5
", + "image_path": "90f45368e61c4a71e27b38b36901abb903432836ba4afb05fdbefc635fce1767.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "table_body" + } + ], + "index": 20 + }, + { + "type": "table", + "bbox": [ + 315, + 449, + 553, + 510 + ], + "blocks": [ + { + "bbox": [ + 313, + 394, + 555, + 449 + ], + "lines": [ + { + "bbox": [ + 313, + 394, + 555, + 449 + ], + "spans": [ + { + "bbox": [ + 313, + 394, + 555, + 449 + ], + "type": "text", + "content": "Table 4. Effectiveness of pseudo labeler and instance segmentator. UnScene3D [32] is an sota unsupervised method. \"disjoint\" denotes training the 3D UNet and pseudo labeler first and then freeze them to train instance segmentator while \"joint\" denotes training the pseudo labeler and the instance segmentator jointly." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 449, + 553, + 510 + ], + "lines": [ + { + "bbox": [ + 315, + 449, + 553, + 510 + ], + "spans": [ + { + "bbox": [ + 315, + 449, + 553, + 510 + ], + "type": "table", + "html": "
MethodAPAP50AP25
UnScene3D [32]21.440.352.6
SKetchy-3DIS (disjoint)45.360.470.0
SKetchy-3DIS (joint)53.469.177.5
", + "image_path": "69427de89d4f970e248cdf4f86d44ba0248944cdf0c1eae49bd39069ae9ccd07.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "table_body" + } + ], + "index": 22 + }, + { + "bbox": [ + 312, + 521, + 555, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 521, + 555, + 640 + ], + "spans": [ + { + "bbox": [ + 312, + 521, + 555, + 640 + ], + "type": "text", + "content": "Effectiveness of labeler and segmentator. In Table 4, experiments are conducted on the S3DIS Area 5 to verify the effectiveness of our designed adaptive box-to-point pseudo labeler and coarse-to-fine instance segmentator. We choose the state-of-the-art unsupervised method UnScene3D [32] for 3D instance segmentation as the baseline. The experimental results verify that our designed pseudo labeler can generate high-quality pseudo labels, and combined with our designed instance segmentator, our method achieves significant performance." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 312, + 642, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 642, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 312, + 642, + 556, + 714 + ], + "type": "text", + "content": "Effectiveness of components in pseudo labeler. In Table 5, we evaluate the effectiveness of various components in our designed pseudo labeler. The first row is the results of the state-of-the-art unsupervised method [32], and the comparison between the last two rows verifies that the learned point-to-box similarity and point-to-instance assignments" + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8885" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 125, + 296, + 198 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 297, + 125 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 297, + 125 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 297, + 125 + ], + "type": "text", + "content": "Table 5. Effectiveness of components in pseudo labeler. \"Partition\" denotes obtaining point-wise labels by box-to-point spatial correlations, \"Assign\" denotes assigning the points within overlapped boxes, and \"Similarity\" denotes filtering the background points within boxes according to the point-to-box similarities." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 125, + 296, + 198 + ], + "lines": [ + { + "bbox": [ + 56, + 125, + 296, + 198 + ], + "spans": [ + { + "bbox": [ + 56, + 125, + 296, + 198 + ], + "type": "table", + "html": "
PartitionAssignSimilarityAPAP50AP25
15.932.258.5
✓41.864.872.3
✓✓45.267.383.4
✓✓✓46.068.883.6
", + "image_path": "5fd035eedf06b6a6e1d7d3946d909402b0d2063533b66fef7b0b24267ceec5ef.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 56, + 265, + 296, + 338 + ], + "blocks": [ + { + "bbox": [ + 54, + 209, + 296, + 265 + ], + "lines": [ + { + "bbox": [ + 54, + 209, + 296, + 265 + ], + "spans": [ + { + "bbox": [ + 54, + 209, + 296, + 265 + ], + "type": "text", + "content": "Table 6. Attentions in Multi-level Attention Block. \"Scene\", \"Coarse\", and \"Core\" denote interacting with the features of the whole scene, the coarse instance regions, and the core regions of the instance, respectively. And \"Self\" denotes augmenting by the learned instance query itself." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 265, + 296, + 338 + ], + "lines": [ + { + "bbox": [ + 56, + 265, + 296, + 338 + ], + "spans": [ + { + "bbox": [ + 56, + 265, + 296, + 338 + ], + "type": "table", + "html": "
SceneCoarseCoreSelfAPAP50AP25
41.364.179.8
✓✓44.667.082.1
✓✓✓44.767.182.9
✓✓✓✓46.068.883.6
", + "image_path": "1499a4b39ab1649565a9b4c34a66542c073bd373ca0cdbbba0248154840db225.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 56, + 382, + 296, + 460 + ], + "blocks": [ + { + "bbox": [ + 55, + 349, + 296, + 382 + ], + "lines": [ + { + "bbox": [ + 55, + 349, + 296, + 382 + ], + "spans": [ + { + "bbox": [ + 55, + 349, + 296, + 382 + ], + "type": "text", + "content": "Table 7. Effectiveness of instance segmentation losses. " + }, + { + "bbox": [ + 55, + 349, + 296, + 382 + ], + "type": "inline_equation", + "content": "L_{box}" + }, + { + "bbox": [ + 55, + 349, + 296, + 382 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 349, + 296, + 382 + ], + "type": "inline_equation", + "content": "L_{core-box}" + }, + { + "bbox": [ + 55, + 349, + 296, + 382 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 349, + 296, + 382 + ], + "type": "inline_equation", + "content": "L_{mask}" + }, + { + "bbox": [ + 55, + 349, + 296, + 382 + ], + "type": "text", + "content": " denote the supervisions of the predicted instance box, instance core box, and instance mask, respectively." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 382, + 296, + 460 + ], + "lines": [ + { + "bbox": [ + 56, + 382, + 296, + 460 + ], + "spans": [ + { + "bbox": [ + 56, + 382, + 296, + 460 + ], + "type": "table", + "html": "
LboxLcore-boxLmaskAPAP50AP25
✓45.266.882.3
✓✓45.467.783.0
✓✓45.067.883.0
✓✓✓46.068.883.6
", + "image_path": "37dd8ee275b96faf76194adfe248f6b260a2c7648fbad4a5c9903c67f27e5c4f.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 472, + 297, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 472, + 297, + 506 + ], + "spans": [ + { + "bbox": [ + 54, + 472, + 297, + 506 + ], + "type": "text", + "content": "can improve the performance by filtering the background points within boxes and assigning the points within overlapped boxes." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 508, + 297, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 508, + 297, + 616 + ], + "spans": [ + { + "bbox": [ + 54, + 508, + 297, + 616 + ], + "type": "text", + "content": "Effectiveness of various attentions in Multi-level Attention Block. In Table 6, we conduct ablation studies on various combinations of attention layers utilized in Multi-level Attention Block. The best performance is obtained with the combination of the four attention modules. These experimental results verify that our proposed Multi-level Attention Block can explore the local contexts adaptively and aggregates them, which boosts the segmentation performance progressively." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 616, + 296, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 616, + 296, + 689 + ], + "spans": [ + { + "bbox": [ + 54, + 616, + 296, + 689 + ], + "type": "text", + "content": "Effectiveness of instance segmentation losses. In Table 7, we compare the performances of the losses of the predicted boxes, core boxes, and masks. The results show that the best performance is obtained with the combination of the above three kinds of losses, combined with our coarse-to-fine instance segmentator, they achieve inspiring performance." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": "Qualitative results of 3D instance segmentation. Some visualized samples of predicted instances for the ScanNetV2" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 315, + 74, + 549, + 266 + ], + "blocks": [ + { + "bbox": [ + 315, + 74, + 549, + 266 + ], + "lines": [ + { + "bbox": [ + 315, + 74, + 549, + 266 + ], + "spans": [ + { + "bbox": [ + 315, + 74, + 549, + 266 + ], + "type": "image", + "image_path": "0f68def47a1778c26a28d7293916f8e2da019ff4160b7b33f61698abb1321a20.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 273, + 555, + 308 + ], + "lines": [ + { + "bbox": [ + 313, + 273, + 555, + 308 + ], + "spans": [ + { + "bbox": [ + 313, + 273, + 555, + 308 + ], + "type": "text", + "content": "Figure 6. Qualitative results on ScanNetV2 validation set. The black denotes the background points and other colors denote different objects. The cycles highlight the key regions." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 317, + 556, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 317, + 556, + 437 + ], + "spans": [ + { + "bbox": [ + 313, + 317, + 556, + 437 + ], + "type": "text", + "content": "validation set are shown in Fig.6. It can be observed that our method shows consistent effectiveness in segmenting objects for easy and complex scenes. Even if there are many objects belonging to the same category and they are close to each other, our method demonstrates considerable performance with the sketchy bounding box annotation. Notably, our method can adaptively segment the objects that are ignored in the ground truth such as the cabinets highlighted in red cycles, which demonstrates the strong recognition capability of our method." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 447, + 388, + 459 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 447, + 388, + 459 + ], + "spans": [ + { + "bbox": [ + 314, + 447, + 388, + 459 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 468, + 556, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 468, + 556, + 624 + ], + "spans": [ + { + "bbox": [ + 313, + 468, + 556, + 624 + ], + "type": "text", + "content": "In this work, we propose a query-based framework for weakly supervised 3D instance segmentation, Sketchy-3DIS, which is more robust and effective than existing bounding-box supervised methods. Sketchy-3DIS adaptively generates pseudo per-point labels from sketchy boxes and predicts instances in a coarse-to-fine manner. Extensive experiments on ScanNetV2 and S3DIS benchmarks show that our method achieves leading performance, and even outperforms some fully-supervised methods. However, our explored experiments demonstrate that the performance of our method degraded severely when the annotated sketchy bounding boxes are immensely inaccurate, exploration for this issue may be the future study orientation." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 314, + 634, + 411, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 634, + 411, + 647 + ], + "spans": [ + { + "bbox": [ + 314, + 634, + 411, + 647 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 653, + 556, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 556, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 556, + 712 + ], + "type": "text", + "content": "This work was supported by the National Key R&D Program of China No. 2024YFC3015801, National Science Fund of China under Grant Nos. 62361166670, 62276144, U24A20330, 62306238 and the Fundamental Research Funds for the Central Universities." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8886" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 133 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 133 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 133 + ], + "type": "text", + "content": "[1] Iro Armeni, Ozan Sener, Amir R Zamir, Helen Jiang, Ioannis Brilakis, Martin Fischer, and Silvio Savarese. 3D semantic parsing of large-scale indoor spaces. In CVPR, pages 1534-1543, 2016. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 136, + 296, + 168 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 136, + 296, + 168 + ], + "spans": [ + { + "bbox": [ + 61, + 136, + 296, + 168 + ], + "type": "text", + "content": "[2] Shaoyu Chen, Jiemin Fang, Qian Zhang, Wenyu Liu, and Xinggang Wang. Hierarchical aggregation for 3D instance segmentation. In CVPR, pages 15467-15476, 2021. 1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 170, + 296, + 212 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 170, + 296, + 212 + ], + "spans": [ + { + "bbox": [ + 62, + 170, + 296, + 212 + ], + "type": "text", + "content": "[3] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In CVPR, pages 1290–1299, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 214, + 296, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 214, + 296, + 257 + ], + "spans": [ + { + "bbox": [ + 62, + 214, + 296, + 257 + ], + "type": "text", + "content": "[4] Julian Chibane, Francis Engelmann, Tuan Anh Tran, and Gerard Pons-Moll. Box2Mask: Weakly supervised 3D semantic instance segmentation using bounding boxes. In ECCV, pages 681-699. Springer, 2022. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 258, + 296, + 301 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 258, + 296, + 301 + ], + "spans": [ + { + "bbox": [ + 62, + 258, + 296, + 301 + ], + "type": "text", + "content": "[5] Ruihang Chu, Xiaqing Ye, Zhengzhe Liu, Xiao Tan, Xiaojuan Qi, Chi-Wing Fu, and Jiaya Jia. Twist: Two-way inter-label self-training for semi-supervised 3D instance segmentation. In CVPR, pages 1100-1109, 2022. 1, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 303, + 296, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 303, + 296, + 346 + ], + "spans": [ + { + "bbox": [ + 62, + 303, + 296, + 346 + ], + "type": "text", + "content": "[6] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. ScanNet: Richly-annotated 3D reconstructions of indoor scenes. In CVPR, pages 5828-5839, 2017. 1, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 347, + 296, + 389 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 347, + 296, + 389 + ], + "spans": [ + { + "bbox": [ + 62, + 347, + 296, + 389 + ], + "type": "text", + "content": "[7] Shichao Dong, Ruibo Li, Jiacheng Wei, Fayao Liu, and Guosheng Lin. Rwseg: Cross-graph competing random walks for weakly supervised 3D instance segmentation. arXiv preprint arXiv:2208.05110, 2022. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 392, + 296, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 392, + 296, + 434 + ], + "spans": [ + { + "bbox": [ + 62, + 392, + 296, + 434 + ], + "type": "text", + "content": "[8] Heming Du, Xin Yu, Farookh Hussain, Mohammad Ali Armin, Lars Petersson, and Weihao Li. Weakly-supervised point cloud instance segmentation with geometric priors. In WACV, pages 4271-4280, 2023. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 436, + 296, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 436, + 296, + 479 + ], + "spans": [ + { + "bbox": [ + 62, + 436, + 296, + 479 + ], + "type": "text", + "content": "[9] Benjamin Graham, Martin Engelcke, and Laurens Van Der Maaten. 3D semantic segmentation with submanifold sparse convolutional networks. In CVPR, pages 9224-9232, 2018. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 480, + 296, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 480, + 296, + 522 + ], + "spans": [ + { + "bbox": [ + 57, + 480, + 296, + 522 + ], + "type": "text", + "content": "[10] Tong He, Chunhua Shen, and Anton Van Den Hengel. Dyco3D: Robust instance segmentation of 3D point clouds through dynamic convolution. In CVPR, pages 354-363, 2021. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 525, + 296, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 525, + 296, + 557 + ], + "spans": [ + { + "bbox": [ + 57, + 525, + 296, + 557 + ], + "type": "text", + "content": "[11] Ji Hou, Angela Dai, and Matthias Nießner. 3D-SIS: 3D semantic instance segmentation of rgb-d scans. In CVPR, pages 4421-4430, 2019. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 559, + 296, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 559, + 296, + 601 + ], + "spans": [ + { + "bbox": [ + 57, + 559, + 296, + 601 + ], + "type": "text", + "content": "[12] Ji Hou, Benjamin Graham, Matthias Nießner, and Saining Xie. Exploring data-efficient 3D scene understanding with contrastive scene contexts. In CVPR, pages 15587-15597, 2021. 1, 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 603, + 296, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 603, + 296, + 635 + ], + "spans": [ + { + "bbox": [ + 57, + 603, + 296, + 635 + ], + "type": "text", + "content": "[13] Le Hui, Linghua Tang, Yaqi Shen, Jin Xie, and Jian Yang. Learning superpoint graph cut for 3D instance segmentation. NIPS, 35:36804-36817, 2022. 1, 2, 6, 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 636, + 296, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 636, + 296, + 678 + ], + "spans": [ + { + "bbox": [ + 57, + 636, + 296, + 678 + ], + "type": "text", + "content": "[14] Li Jiang, Hengshuang Zhao, Shaoshuai Shi, Shu Liu, ChiWing Fu, and Jiaya Jia. Pointgroup: Dual-set point grouping for 3D instance segmentation. In CVPR, pages 4867-4876, 2020. 1, 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 681, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 681, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 681, + 296, + 713 + ], + "type": "text", + "content": "[15] Lei Ke, Martin Danelljan, Henghui Ding, Yu-Wing Tai, Chi-Keung Tang, and Fisher Yu. Mask-free video instance segmentation. In CVPR, pages 22857–22866, 2023. 3" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 316, + 73, + 555, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 555, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 555, + 116 + ], + "type": "text", + "content": "[16] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In ICCV, pages 4015-4026, 2023. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 118, + 555, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 118, + 555, + 160 + ], + "spans": [ + { + "bbox": [ + 317, + 118, + 555, + 160 + ], + "type": "text", + "content": "[17] Maxim Kolodiazhnyi, Anna Vorontsova, Anton Konushin, and Danila Rukhovich. OneFormer3D: One transformer for unified point cloud segmentation. In CVPR, pages 20943-20953, 2024. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 162, + 555, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 162, + 555, + 194 + ], + "spans": [ + { + "bbox": [ + 317, + 162, + 555, + 194 + ], + "type": "text", + "content": "[18] Harold W Kuhn. The hungarian method for the assignment problem. Naval research logistics quarterly, 2(1-2):83-97, 1955. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 195, + 555, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 195, + 555, + 236 + ], + "spans": [ + { + "bbox": [ + 317, + 195, + 555, + 236 + ], + "type": "text", + "content": "[19] Xin Lai, Yuhui Yuan, Ruihang Chu, Yukang Chen, Han Hu, and Jiaya Jia. Mask-Attention-Free Transformer for 3D instance segmentation. In ICCV, pages 3693–3703, 2023. 1, 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 239, + 555, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 239, + 555, + 293 + ], + "spans": [ + { + "bbox": [ + 316, + 239, + 555, + 293 + ], + "type": "text", + "content": "[20] Shiyi Lan, Zhiding Yu, Christopher Choy, Subhashree Radhakrishnan, Guilin Liu, Yuke Zhu, Larry S Davis, and Anima Anandkumar. DiscoBox: Weakly supervised instance segmentation and semantic correspondence from box supervision. In ICCV, pages 3406-3416, 2021. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 294, + 555, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 294, + 555, + 326 + ], + "spans": [ + { + "bbox": [ + 316, + 294, + 555, + 326 + ], + "type": "text", + "content": "[21] Loic Landrieu and Martin Simonovsky. Large-scale point cloud semantic segmentation with superpoint graphs. In CVPR, pages 4558-4567, 2018. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 327, + 555, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 327, + 555, + 369 + ], + "spans": [ + { + "bbox": [ + 316, + 327, + 555, + 369 + ], + "type": "text", + "content": "[22] Zhihao Liang, Zhihao Li, Songcen Xu, Mingkui Tan, and Kui Jia. Instance segmentation in 3D scenes using semantic superpoint tree networks. In ICCV, pages 2783-2792, 2021. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 371, + 555, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 371, + 555, + 414 + ], + "spans": [ + { + "bbox": [ + 316, + 371, + 555, + 414 + ], + "type": "text", + "content": "[23] Yongbin Liao, Hongyuan Zhu, Yanggang Zhang, Chuangguan Ye, Tao Chen, and Jiayuan Fan. Point cloud instance segmentation with semi-supervised bounding-box mining. TPAMI, 44(12):10159-10170, 2021. 1, 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 415, + 555, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 415, + 555, + 458 + ], + "spans": [ + { + "bbox": [ + 316, + 415, + 555, + 458 + ], + "type": "text", + "content": "[24] Zhengzhe Liu, Xiaojuan Qi, and Chi-Wing Fu. One thing one click: A self-training approach for weakly supervised 3D semantic segmentation. In CVPR, pages 1726-1736, 2021. 1, 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 460, + 555, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 460, + 555, + 492 + ], + "spans": [ + { + "bbox": [ + 316, + 460, + 555, + 492 + ], + "type": "text", + "content": "[25] Ze Liu, Zheng Zhang, Yue Cao, Han Hu, and Xin Tong. Group-free 3D object detection via transformers. In ICCV, pages 2949-2958, 2021. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 494, + 555, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 494, + 555, + 525 + ], + "spans": [ + { + "bbox": [ + 316, + 494, + 555, + 525 + ], + "type": "text", + "content": "[26] Jiahao Lu, Jiacheng Deng, Chuxin Wang, Jianfeng He, and Tianzhu Zhang. Query refinement transformer for 3D instance segmentation. In ICCV, pages 18516-18526, 2023. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 526, + 555, + 559 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 526, + 555, + 559 + ], + "spans": [ + { + "bbox": [ + 316, + 526, + 555, + 559 + ], + "type": "text", + "content": "[27] Jiahao Lu, Jiacheng Deng, and Tianzhu Zhang. BSNet: Box-supervised simulation-assisted mean teacher for 3D instance segmentation. In CVPR, pages 20374-20384, 2024. 2, 3, 6, 7" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 559, + 555, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 559, + 555, + 592 + ], + "spans": [ + { + "bbox": [ + 316, + 559, + 555, + 592 + ], + "type": "text", + "content": "[28] Ishan Misra, Rohit Girdhar, and Armand Joulin. An end-to-end transformer model for 3D object detection. In ICCV, pages 2906-2917, 2021. 1" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 593, + 555, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 593, + 555, + 635 + ], + "spans": [ + { + "bbox": [ + 316, + 593, + 555, + 635 + ], + "type": "text", + "content": "[29] Tuan Duc Ngo, Binh-Son Hua, and Khoi Nguyen. GaPro: Box-supervised 3D point cloud instance segmentation using gaussian processes as pseudo labelers. In ICCV, pages 17794-17803, 2023. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 636, + 555, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 636, + 555, + 679 + ], + "spans": [ + { + "bbox": [ + 316, + 636, + 555, + 679 + ], + "type": "text", + "content": "[30] Tuan Duc Ngo, Binh-Son Hua, and Khoi Nguyen. ISBNet: a 3D point cloud instance segmentation network with instance-aware sampling and box-aware dynamic convolution. In CVPR, pages 13550-13559, 2023. 1, 2, 6, 7" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 316, + 681, + 555, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 555, + 713 + ], + "type": "text", + "content": "[31] Charles R Qi, Or Litany, Kaiming He, and Leonidas J Guibas. Deep hough voting for 3D object detection in point clouds. In ICCV, pages 9277-9286, 2019. 1" + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8887" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 712 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 297, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 297, + 106 + ], + "type": "text", + "content": "[32] David Rozenberszki, Or Litany, and Angela Dai. UnScene3D: Unsupervised 3D instance segmentation for indoor scenes. In CVPR, pages 19957-19967, 2024. 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 108, + 296, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 108, + 296, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 108, + 296, + 140 + ], + "type": "text", + "content": "[33] Danila Rukhovich, Anna Vorontsova, and Anton Konushin. FCAF3D: Fully convolutional anchor-free 3D object detection. In ECCV, pages 477-493. Springer, 2022. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 143, + 296, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 143, + 296, + 186 + ], + "spans": [ + { + "bbox": [ + 56, + 143, + 296, + 186 + ], + "type": "text", + "content": "[34] Jonas Schult, Francis Engelmann, Alexander Hermans, Or Litany, Siyu Tang, and Bastian Leibe. Mask3D: Mask transformer for 3D semantic instance segmentation. In ICRA, pages 8216-8223. IEEE, 2023. 1, 2, 6, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 188, + 296, + 232 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 188, + 296, + 232 + ], + "spans": [ + { + "bbox": [ + 56, + 188, + 296, + 232 + ], + "type": "text", + "content": "[35] Sangyun Shin, Kaichen Zhou, Madhu Vankadari, Andrew Markham, and Niki Trigoni. Spherical mask: Coarse-to-fine 3D point cloud instance segmentation with spherical representation. In CVPR, pages 4060-4069, 2024. 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 234, + 296, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 234, + 296, + 266 + ], + "spans": [ + { + "bbox": [ + 56, + 234, + 296, + 266 + ], + "type": "text", + "content": "[36] Jiahao Sun, Chunmei Qing, Junpeng Tan, and Xiangmin Xu. Superpoint transformer for 3D scene instance segmentation. In AAAI, pages 2393-2401, 2023. 1, 2, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 269, + 294, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 269, + 294, + 300 + ], + "spans": [ + { + "bbox": [ + 56, + 269, + 294, + 300 + ], + "type": "text", + "content": "[37] Linghua Tang, Le Hui, and Jin Xie. Learning inter-superpoint affinity for weakly supervised 3D instance segmentation. In ACCV, pages 1282-1297, 2022. 1, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 303, + 296, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 303, + 296, + 335 + ], + "spans": [ + { + "bbox": [ + 56, + 303, + 296, + 335 + ], + "type": "text", + "content": "[38] An Tao, Yueqi Duan, Yi Wei, Jiwen Lu, and Jie Zhou. Seg-Group: Seg-level supervision for 3D instance and semantic segmentation. TIP, 31:4952-4965, 2022. 1, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 338, + 296, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 338, + 296, + 370 + ], + "spans": [ + { + "bbox": [ + 56, + 338, + 296, + 370 + ], + "type": "text", + "content": "[39] Zhi Tian, Chunhua Shen, Xinlong Wang, and Hao Chen. BoxInst: High-performance instance segmentation with box annotations. In CVPR, pages 5443-5452, 2021. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 373, + 296, + 405 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 373, + 296, + 405 + ], + "spans": [ + { + "bbox": [ + 56, + 373, + 296, + 405 + ], + "type": "text", + "content": "[40] Duc Dang Trung Tran, Byeongkeun Kang, and Yeejin Lee. MSTA3D: Multi-scale twin-attention for 3d instance segmentation. In ACM MM, pages 1467-1475, 2024. 2, 6, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 407, + 294, + 439 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 407, + 294, + 439 + ], + "spans": [ + { + "bbox": [ + 56, + 407, + 294, + 439 + ], + "type": "text", + "content": "[41] Thang Vu, Kookhoi Kim, Tung M Luu, Thanh Nguyen, and Chang D Yoo. SoftGroup for 3D instance segmentation on point clouds. In CVPR, pages 2708-2717, 2022. 1, 2, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 441, + 296, + 484 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 441, + 296, + 484 + ], + "spans": [ + { + "bbox": [ + 56, + 441, + 296, + 484 + ], + "type": "text", + "content": "[42] Haiyang Wang, Shaocong Dong, Shaoshuai Shi, Aoxue Li, Jianan Li, Zhenguo Li, Liwei Wang, et al. CAGroup3D: Class-aware grouping for 3D object detection on point clouds. NIPS, 35:29975-29988, 2022. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 487, + 296, + 530 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 487, + 296, + 530 + ], + "spans": [ + { + "bbox": [ + 56, + 487, + 296, + 530 + ], + "type": "text", + "content": "[43] Haiyang Wang, Shaoshuai Shi, Ze Yang, Rongyao Fang, Qi Qian, Hongsheng Li, Bernt Schiele, and Liwei Wang. RBGNet: Ray-based grouping for 3D object detection. In CVPR, pages 1110-1119, 2022." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 533, + 294, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 533, + 294, + 575 + ], + "spans": [ + { + "bbox": [ + 56, + 533, + 294, + 575 + ], + "type": "text", + "content": "[44] Qian Xie, Yu-Kun Lai, Jing Wu, Zhoutao Wang, Yiming Zhang, Kai Xu, and Jun Wang. MLCVNet: Multi-level context votenet for 3D object detection. In CVPR, pages 10447-10456, 2020. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 578, + 296, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 578, + 296, + 622 + ], + "spans": [ + { + "bbox": [ + 56, + 578, + 296, + 622 + ], + "type": "text", + "content": "[45] Saining Xie, Jiatao Gu, Demi Guo, Charles R Qi, Leonidas Guibas, and Or Litany. PointContrast: Unsupervised pretraining for 3D point cloud understanding. In ECCV, pages 574-591. Springer, 2020. 1, 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 624, + 296, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 296, + 666 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 296, + 666 + ], + "type": "text", + "content": "[46] Bo Yang, Jianan Wang, Ronald Clark, Qingyong Hu, Sen Wang, Andrew Markham, and Niki Trigoni. Learning object bounding boxes for 3D instance segmentation on point clouds. NIPS, 32, 2019. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "type": "text", + "content": "[47] Cheng-Kun Yang, Min-Hung Chen, Yung-Yu Chuang, and Yen-Yu Lin. 2D-3D interlaced transformer for point cloud segmentation with scene-level supervision. In ICCV, pages 977–987, 2023. 2" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 72, + 555, + 274 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 315, + 72, + 555, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 72, + 555, + 116 + ], + "spans": [ + { + "bbox": [ + 315, + 72, + 555, + 116 + ], + "type": "text", + "content": "[48] Li Yi, Wang Zhao, He Wang, Minhyuk Sung, and Leonidas J Guibas. GSPN: Generative shape proposal network for 3D instance segmentation in point cloud. In CVPR, pages 3947-3956, 2019. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 315, + 118, + 555, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 118, + 555, + 161 + ], + "spans": [ + { + "bbox": [ + 315, + 118, + 555, + 161 + ], + "type": "text", + "content": "[49] Qingtao Yu, Heming Du, Chen Liu, and Xin Yu. When 3D bounding-box meets SAM: Point cloud instance segmentation with weak-and-noisy supervision. In WACV, pages 3719-3728, 2024. 1, 2, 3, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 315, + 163, + 555, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 163, + 555, + 196 + ], + "spans": [ + { + "bbox": [ + 315, + 163, + 555, + 196 + ], + "type": "text", + "content": "[50] Zaiwei Zhang, Bo Sun, Haitao Yang, and Qixing Huang. H3DNet: 3D object detection using hybrid geometric primitives. In ECCV, pages 311-329. Springer, 2020. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 315, + 198, + 555, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 198, + 555, + 229 + ], + "spans": [ + { + "bbox": [ + 315, + 198, + 555, + 229 + ], + "type": "text", + "content": "[51] Yu Zheng, Yueqi Duan, Jiwen Lu, Jie Zhou, and Qi Tian. HyperDet3D: Learning a scene-conditioned 3D object detector. In CVPR, pages 5585-5594, 2022. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 315, + 231, + 555, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 231, + 555, + 274 + ], + "spans": [ + { + "bbox": [ + 315, + 231, + 555, + 274 + ], + "type": "text", + "content": "[52] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable DETR: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159, 2020. 2" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "8888" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/SkillMimic_ Learning Basketball Interaction Skills from Demonstrations/70ea78a5-d7df-4937-ac48-382497691b79_content_list.json b/2025/SkillMimic_ Learning Basketball Interaction Skills from Demonstrations/70ea78a5-d7df-4937-ac48-382497691b79_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..c03f2ed0443ba45fd028b8ebfc5779acfeb7a5be --- /dev/null +++ b/2025/SkillMimic_ Learning Basketball Interaction Skills from Demonstrations/70ea78a5-d7df-4937-ac48-382497691b79_content_list.json @@ -0,0 +1,1856 @@ +[ + { + "type": "text", + "text": "SkillMimic: Learning Basketball Interaction Skills from Demonstrations", + "text_level": 1, + "bbox": [ + 130, + 119, + 866, + 142 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yinhuai Wang $^{1,2*}$ Qihan Zhao $^{1*}$ Runyi Yu $^{1,2*}$ Hok Wai Tsui $^{1}$ Ailing Zeng $^{6\\dagger}$ Jing Lin $^{5}$ Zhengyi Luo $^{7}$ Jiwen Yu $^{3}$ Xiu Li $^{4}$ Qifeng Chen $^{1}$ Jian Zhang $^{3\\dagger}$ Lei Zhang $^{5}$ Ping Tan $^{1}$", + "bbox": [ + 109, + 167, + 883, + 205 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Hong Kong University of Science and Technology $^{2}$ Unitree Robotics $^{3}$ Peking University Shenzhen Graduate School $^{4}$ Tsinghua University $^{5}$ International Digital Economy Academy $^{6}$ Tencent $^{7}$ Carnegie Mellon University yinhuai.wang@connect.ust.hk pingtan@ust.hk", + "bbox": [ + 112, + 213, + 883, + 265 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f0da1c7935060e5d15ea20c148848d4ecae60676075df2db709bef83387d5ed3.jpg", + "image_caption": [ + "Figure 1. We propose a novel approach that for the first time enables physically simulated humanoids to learn a variety of basketball interaction skills from Human-Object Interaction (HOI) data, including but not limited to shooting (blue), retrieving (red), and turnaround layup (yellow). Once acquired, these interaction skills can be composed to accomplish complex tasks, such as consecutive scoring (green)." + ], + "image_footnote": [], + "bbox": [ + 93, + 284, + 906, + 412 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 481, + 326, + 496 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Traditional reinforcement learning methods for human-object interaction (HOI) rely on labor-intensive, manually designed skill rewards that do not generalize well across different interactions. We introduce SkillMimic, a unified data-driven framework that fundamentally changes how agents learn interaction skills by eliminating the need for skill-specific rewards. Our key insight is that a unified HOI imitation reward can effectively capture the essence of diverse interaction patterns from HOI datasets. This enables SkillMimic to learn a single policy that not only masters multiple interaction skills but also facilitates skill transitions, with both diversity and generalization improving as the HOI dataset grows. For evaluation, we collect and introduce two basketball datasets containing approximately 35 minutes of diverse basketball skills. Extensive experiments show that SkillMimic successfully masters a wide range of basketball skills including stylistic variations in dribbling, layup, and shooting. Moreover, these learned skills can be effectively composed by a high-level controller to accomplish complex and long-horizon tasks such as consecutive scoring, opening new possibilities for scalable and generalizable interaction skill learning. Project page: https://ingrid789.github.io/SkillMimic/", + "bbox": [ + 88, + 513, + 485, + 863 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 514, + 481, + 643, + 496 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Human-Object Interaction (HOI) represents one of the most fundamental yet challenging capabilities in robotics and character animation [10, 12, 16, 23, 28, 46, 47]. Among various HOI scenarios, basketball serves as an exemplary testbed that demands mastery of complex, dynamic, and precise interactions. Despite significant progress, existing physics-based character animation methods predominantly focus on imitation learning of locomotion skills [8, 32–34, 42] through Reinforcement Learning (RL), leaving a significant gap for HOI imitation. While some methods can learn interaction skills, such as striking pillars [34], playing tennis [50], climbing ropes [2], and carrying boxes [15], they typically involve handcrafted reward functions for each different interaction skill. However, designing skill-specific rewards is labor-intensive and fundamentally limits generalization. This limitation becomes particularly acute in basketball, where the complexity and diversity of skills would demand extraordinary engineering effort for reward design. Consequently, existing methods [2, 8, 15, 32–34, 42, 50] fail to provide a unified learning framework for a single policy to learn diverse interaction skills, let alone achieving complex long-term tasks such as consecutive basketball scoring.", + "bbox": [ + 511, + 507, + 906, + 840 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "How can a simulated humanoid learn a wide variety of interaction skills in a simple and scalable manner? Consider how humans naturally acquire complex skills: amateur basketball players often develop their expertise by watching", + "bbox": [ + 511, + 840, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal Contribution. † Corresponding author.", + "bbox": [ + 114, + 886, + 357, + 901 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "17540", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "and imitating game footage, even without formal coaching. This self-directed learning process involves synchronizing their body movements with both the observed player motions and ball trajectories through practice.", + "bbox": [ + 89, + 90, + 480, + 151 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Inspired by these observations, we propose a data-driven method called SkillMimic, which mimics HOI through RL to learn interaction skills. SkillMimic can learn various interaction skills purely from HOI datasets, such as diverse styles of basketball shooting, layup, and dribbling skills, and even a robust pickup skill that enables picking up balls in random locations and motions. Notably, SkillMimic uses the exact same configuration to learn different interaction skills, with identical hyperparameters. This allows us to train a single Interaction Skill (IS) policy to learn multiple interaction skills and achieve smooth skill switching. By training a high-level policy to reuse these learned interaction skills, we can accomplish challenging high-level tasks such as scoring, which requires the humanoid to composite diverse basketball interaction skills to score accurately.", + "bbox": [ + 89, + 152, + 482, + 378 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To advance basketball skill learning, we introduce two HOI datasets: BallPlay-V, which captures player and ball motion from RGB videos across eight basic skills, and BallPlay-M, which uses optical motion capture to record 35 minutes of comprehensive basketball interactions. Using these datasets, SkillMimic successfully learns diverse basketball skills through a unified IS policy. With an additional high-level controller for skill composition, our system can execute complex tasks like consecutive scoring. Our approach demonstrates superior efficiency in learning both interaction skills and multi-phase, long-horizon tasks compared to existing methods, opening new possibilities for scalable and generalizable interaction skill learning.", + "bbox": [ + 89, + 381, + 482, + 575 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Specifically, our contributions are as follows:", + "bbox": [ + 109, + 579, + 410, + 593 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- SkillMimic: a data-driven paradigm for learning diverse humanoid interaction skills through RL. It supports joint learning and switching of diverse basketball interaction skills, with skill diversity and generalization improving as the demonstration data grows, showcasing the scalability.", + "- Contact graph: We propose a simple and general contact modeling method that applies to diverse skills, called the contact graph. A contact graph reward is designed to enable precise contact imitation, which proves to be critical for learning precise interaction skills.", + "- Unified HOI imitation reward: We propose a set of important designs that form a unified reward configuration for imitation learning of various interaction skills.", + "- A hierarchical solution for reusing interaction skills: We propose training a high-level controller to reuse the interaction skills acquired by SkillMimic to accomplish challenging high-level tasks, which for the first time achieves continuous basketball scoring.", + "- BallPlay Datasets: We introduce two basketball datasets to facilitate research on interaction skill learning." + ], + "bbox": [ + 91, + 595, + 482, + 897 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 89, + 653, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Locomotion Imitation Learning. Mimicking human motion for robot control is an efficient method of learning humanoid skills. DeepMimic [32], a pioneer in this field, uses imitation learning to perform a variety of highly dynamic locomotion skills. The introduction of Generative Adversarial Imitation Learning (GAIL) [18] into humanoid imitation learning by AMP [33] lessens constraints on data alignment, thus enhancing its versatility. ASE [34] further amplifies motion diversity in GAIL training, incorporates a pre-trained low-level policy to acquire locomotion skills, and deploys a high-level policy to repurpose these locomotion skills for specific tasks, such as striking pillars. These methodologies are subsequently expanded in numerous efforts to augment condition and text control [8, 22, 29, 31, 35, 39, 42, 44, 53], as well as to design automated rewards [7]. However, a limitation of these studies is their exclusive focus on learning isolated locomotion skills through imitation. This requires additional pipeline and reward design for learning interaction skills, even for relatively simple ones, such as moving boxes [15] or striking pillars [34]. In contrast, our approach enables unified learning of diverse interaction skills, eliminating the need for skill-specific reward engineering.", + "bbox": [ + 511, + 116, + 906, + 464 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Learning Human-Object Interaction (HOI). Learning HOI in simulation has long posed challenges in the fields of graphics and robotics [14, 20, 21, 23, 24, 37, 38, 45, 47, 48]. Early studies often relied on manually designed control structures. For example, Hodgins et al. [19] implement a variety of sports interactions through the use of manually designed state machine controllers. Yin et al. [49] use finite-state machines to generate a large variety of gaits. Coros et al. [6] utilize an inverted pendulum model to control foot placement. Recently, deep reinforcement learning [1] has been widely employed in character control, giving rise to an array of methods based on policy network control. These methods have been applied to a range of challenging activities, such as basketball [26], skateboarding [25], and cycling [41], etc. The use of imitation rewards further lessens the reward design for locomotion styles. Zhang et al. [50] propose a framework to learn diverse tennis skills from broadcast videos. Bae et al. [2] incorporate multiple part-wise motion priors for diverse whole-body interactions. Similarly, Braun et al. [4] develop a framework for synthesizing whole-body grasps.", + "bbox": [ + 511, + 477, + 908, + 792 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Applying imitation learning for HOI is intuitive, however, effectively learning diverse interaction skills in a unified framework remains unresolved. Naively extending locomotion imitation methods [32, 33] to HOI imitation often yields unstable and unnatural results, as these methods do not address the unbalanced rewards inherent in HOI, nor do they properly account for the crucial aspects of relative mo", + "bbox": [ + 511, + 795, + 906, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "17541", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/6c3ee744f479b9227dff4936e632e46e21c71f76c366684addf939351f46f0aa.jpg", + "image_caption": [ + "Figure 2. Concept of SkillMimic. We define an interaction skill as a set of Human-Object Interaction (HOI) state transitions that align with the intended skill semantics. These state transitions can be derived from captured HOI motion clips. If a simulated humanoid can manipulate objects such that the resulting HOI state transitions closely match those of the reference, we consider the humanoid to have successfully learned the interaction skill." + ], + "image_footnote": [], + "bbox": [ + 94, + 88, + 480, + 188 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "tion and contact in HOI. Modeling interaction through interaction graphs is a well-established approach [17]. While this method has proven useful for HOI imitation [52], it remains limited by considering only kinematic relationships, failing to ensure proper contact and becoming unstable in learning diverse scenarios. In contrast, our contact graph imitation reward explicitly prioritize physical contact learning, significantly improving the success rate of HOI imitation and enabling unified HOI imitation of diverse complex interaction skills for the first time.", + "bbox": [ + 88, + 323, + 482, + 474 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This work builds upon and significantly extends ideas initially explored in our early technical report [43].", + "bbox": [ + 89, + 474, + 482, + 506 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. SkillMimic", + "text_level": 1, + "bbox": [ + 89, + 520, + 209, + 536 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we introduce a unified data-driven framework for simulated humanoids to learn various basketball interaction skills via simulation and reinforcement learning (RL). We begin by introducing how we define interaction skills with Human-Object Interaction (HOI) data (Sec. 3.2). Next, we explain how to learn diverse interaction skills by imitating HOI data (Sec. 3.3) and discuss the designing of a unified HOI imitation reward (Sec. 3.4). Finally, we demonstrate how the learned interaction skills can be reused to perform high-level tasks (Sec. 3.5).", + "bbox": [ + 89, + 546, + 483, + 698 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. The BallPlay Dataset", + "text_level": 1, + "bbox": [ + 89, + 708, + 289, + 724 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To address the scarcity of basketball HOI data and facilitate research on interaction skill learning, we introduce two datasets: BallPlay-V, based on monocular vision estimation, containing eight clips of different basketball skills; and BallPlay-M, based on multi-view optical motion capture systems, containing 35 minutes of diverse basketball interactions. We provide more details in the appendix.", + "bbox": [ + 89, + 729, + 482, + 838 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Defining Interaction Skills Using HOI Data", + "text_level": 1, + "bbox": [ + 89, + 848, + 457, + 864 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We propose a data-driven approach to defining interaction skills as collections of Human-Object Interaction (HOI)", + "bbox": [ + 89, + 869, + 482, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "state transitions that align with the intended skill semantics. For instance, the interaction skill of \"picking up balls from various positions\" can be encapsulated by various HOI motion clips, each capturing a unique instance of the action. These clips collectively form a collection of HOI state transitions that represent the interaction skill.", + "bbox": [ + 511, + 90, + 903, + 181 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.3. Learning Interaction Skills by HOI Imitation", + "text_level": 1, + "bbox": [ + 511, + 191, + 897, + 207 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Considering an interaction skill defined by a set of reference HOI state transitions, if a humanoid can manipulate objects such that its HOI state transitions closely resemble the reference, we consider the humanoid to have successfully learned the interaction skill. Based on this concept, we propose a method for learning interaction skills by imitating HOI state transitions through RL, which we call SkillMimic. Fig. 2 illustrates the concept of SkillMimic.", + "bbox": [ + 511, + 213, + 903, + 333 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In contrast to Behavior Clone [3, 5, 11] methods which rely on precise state-action data, SkillMimic learns actions through RL using state-only trajectories, demonstrating better tolerance to data noise and higher data efficiency. Unlike previous RL methods [33, 34, 50] that require manually designed rewards for each interaction skill, SkillMimic is fully data-driven, skill-agnostic, and scalable, making it capable of learning a wide range of basketball skills within a unified solution. Fig. 6 shows diverse basketball skills acquired using SkillMimic. Fig. 7 illustrates how the performance of the pickup skill improves as the reference data increases.", + "bbox": [ + 511, + 334, + 905, + 500 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.3.1. Training Pipeline", + "text_level": 1, + "bbox": [ + 511, + 507, + 679, + 522 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Fig. 3 (b) shows the training pipeline of SkillMimic. Given an HOI dataset with diverse skill-labeled clips (Fig. 3 (a)), SkillMimic trains an Interaction Skill (IS) policy by randomly selecting a frame from a clip corresponding to a chosen skill to initialize the humanoid and object states. The state $s_t$ (Sec. 3.3.2) and skill label $c_j$ are input into the IS policy, which predicts actions (Sec. 3.3.3) that are then simulated to produce the next state $s_{t+1}$ . A unified HOI imitation reward (Sec. 3.4) is designed to measure the consistency between the simulated state transitions and the reference transitions. When the humanoid falls or reaches the maximum simulation length, the environment resets, and the process repeats. After training converges, the IS policy enables the humanoid to execute interaction skills that closely mirror the reference demonstrations.", + "bbox": [ + 511, + 526, + 905, + 752 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.3.2. HOI Observation", + "text_level": 1, + "bbox": [ + 511, + 761, + 681, + 773 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The state observed by the IS policy is not under direct supervision and may theoretically encompass any information that is available in the simulation environment. Importantly, this information does not necessarily have to be included in the reference HOI data. Similar to prior arts [34], we transform all coordinates into the root local coordinate of the humanoid, which aligns the data distribution and benefits the generalization performance. For the humanoid, we observe", + "bbox": [ + 511, + 779, + 905, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "17542", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/108388a6dd10c8e77ca2f4a38f92ff9a4678079f2b4e4083d446233a967213f4.jpg", + "image_caption": [ + "Figure 3. Our system consists of three parts. (a) First, we capture real-world basketball skills to create a large Human-Object Interaction (HOI) motion dataset. (b) Second, we train an Interaction Skill (IS) policy to learn interaction skills by imitating the corresponding HOI data through reinforcement learning. Specifically, the IS policy takes as input the HOI state $\\boldsymbol{s}_t$ and skill label $\\boldsymbol{c}_j$ and predicts the action $\\boldsymbol{a}_t$ . The new state $\\boldsymbol{s}_{t+1}$ is calculated by the simulator. A unified HOI imitation reward is designed to imitate diverse HOI state transitions. (c) The third part involves training a High-Level Controller (HLC) to reuse the learned interaction skills for complex tasks. The HLC takes as input $\\boldsymbol{s}_t$ and extra task observations $\\boldsymbol{h}_t$ , e.g., the basket position, and predicts the skill label $\\boldsymbol{c}_t$ to drive a pre-trained IS policy." + ], + "image_footnote": [], + "bbox": [ + 133, + 85, + 864, + 354 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "its global root height, local body position, rotation, position velocity, and angular velocity. These representations form the humanoid proprioception $\\pmb{o}_t^{prop}$ . In addition, we detect net contact forces $\\pmb{o}_t^f$ for all fingertips, which helps to sense contact and accelerate training. For the object, we observe its local position, rotation, velocity, and angular velocity, which form the object observation $\\pmb{o}_t^{obj}$ . Finally, the state perceived by the IS policy is", + "bbox": [ + 88, + 473, + 480, + 597 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {s} _ {t} = \\left\\{\\boldsymbol {o} _ {t} ^ {\\text {p r o p}}, \\boldsymbol {o} _ {t} ^ {f}, \\boldsymbol {o} _ {t} ^ {\\text {o b j}} \\right\\}. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 205, + 607, + 482, + 627 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Skill labels $c_{j}$ are also used as inputs to the IS policy to differentiate different skills. We use one-hot encodings to represent the skill label.", + "bbox": [ + 89, + 638, + 483, + 685 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3.3. IS Policy and Action", + "text_level": 1, + "bbox": [ + 89, + 694, + 279, + 709 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The policy output is modeled as a Gaussian distribution with dimensions equal to the DOF number of the humanoid robot, featuring constant variance. The mean is modeled by a three-layer MLP consisting of [1024, 512, 512] units with ReLU [30] activations. We use the action $\\boldsymbol{a}_t$ sampled from the policy as the target joint rotations for a full set of PD controllers. The PD controllers adjust and output the joint torques to reach the target rotations.", + "bbox": [ + 89, + 714, + 483, + 835 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Unified HOI Imitation Reward", + "text_level": 1, + "bbox": [ + 89, + 847, + 364, + 862 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This HOI imitation reward function is pivotal as it must not only accurately gauge the efficacy of interaction skill ac", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/66fb65f6c000f579bae1ceb1628c22533205715568efee55d80bf05c548e745a.jpg", + "image_caption": [ + "Figure 4. We propose the Contact Graph (CG) to model general contacts within an explicitly defined scene. The node stores a binary value that denotes whether it contacts other nodes. Each edge stores a binary value indicating whether the two connected nodes are in contact. The node definition is unified for a certain scene and shared between diverse interactive skills. For example, we define three nodes: hands, hands-exclusive body, and ball, to form a simple CG to model contacts for diverse basketball skills." + ], + "image_footnote": [], + "bbox": [ + 516, + 472, + 903, + 571 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "quisition but also avoid the incorporation of skill-specific settings to maintain robust cross-skill generalizability.", + "bbox": [ + 511, + 728, + 903, + 758 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Classic imitation rewards often focus solely on imitating locomotions [32], lacking an established baseline for HOI imitation. Although adversarial imitation rewards [33, 34] are effective in locomotion tasks, we find that they perform poorly in HOI imitation. This may be because adversarial imitation rewards are relatively coarse-grained, making it difficult for them to provide the precise guidance needed for accurate interaction learning.", + "bbox": [ + 511, + 761, + 906, + 882 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The body kinematics imitation reward $r_t^b$ encourages the", + "bbox": [ + 532, + 883, + 903, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "17543", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "alignment of body movements by considering joint position, rotation, and their velocities. Similarly, the object kinematics imitation reward $r_t^o$ ensures consistent object movements with the reference. However, solely relying on these two rewards often leads to unstable learning and unnatural motions. We incorporate a relative motion reward $r_t^{rel}$ to highlight the relative relationships in interactions, which is crucial for interaction naturalness. Recognizing the interlocking relationship between interaction and contact, we propose a straightforward and universal contact modeling approach called Contact Graph (CG). A corresponding CG imitation reward $r_t^{cg}$ is designed to enhance the precision of interaction imitation. Additionally, we propose an adaptive velocity regularization term $r_t^{reg}$ to suppress high-frequency jitters. These sub-rewards are multiplied rather than added to encourage balanced learning, thereby avoiding local optima, as justified in Tab. 1. The combination of these innovations forms a unified imitation reward, enabling the learning of diverse basketball skills in BallPlay-M with the same configuration. To conclude, the complete HOI imitation reward is", + "bbox": [ + 89, + 90, + 483, + 409 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nr _ {t} = r _ {t} ^ {b} * r _ {t} ^ {o} * r _ {t} ^ {r e l} * r _ {t} ^ {r e g} * r _ {t} ^ {c g}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 179, + 419, + 482, + 438 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Next, we will discuss the methodology involved in crafting the CG reward $r_t^{cg}$ . Details about the kinematic imitation reward $r_t^b * r_t^o * r_t^{rel} * r_t^{reg}$ are in the appendix.", + "bbox": [ + 89, + 450, + 483, + 496 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4.1. Contact Graph", + "text_level": 1, + "bbox": [ + 89, + 503, + 243, + 518 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We observe that existing imitation rewards [33, 34, 52] only measure kinematic properties and are insufficient to measure precise contacts. To tackle this problem, we propose the Contact Graph (CG) to model the contact in general interactions and design a CG imitation reward.", + "bbox": [ + 89, + 522, + 483, + 598 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As illustrated in Fig. 4, the CG is a complete graph where every pair of distinct nodes is connected by a unique edge, defined as $\\mathcal{G} = \\{\\mathcal{V},\\mathcal{E}\\}$ , where $\\mathcal{V}\\in \\{0,1\\} ^k$ is the set of $k$ nodes and $\\mathcal{E}\\in \\{0,1\\}^{k(k - 1) / 2}$ is the set of edges. A CG node stores a binary value indicating whether it contacts other nodes. Each CG edge stores a binary label that denotes the contact between two nodes, where 1 represents contact, and 0 means no contact. CG node and edge values are calculated frame by frame, explicitly describing the mutual contact relationship between CG nodes at different moments. Contacts inside a node are not considered. We use the edge set $\\mathcal{E}$ for imitation learning.", + "bbox": [ + 89, + 598, + 483, + 779 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The definitions of nodes are flexible. For example, a node can be a single part (e.g., a fingertip), or an aggregation of multiple parts (e.g., the whole left hand). Hinged objects can also be broken up into finer nodes [9, 13, 51]. The definition of CG is unified for a certain scene and shared between diverse interaction skills. For example, in basketball scenarios, we can aggregate two hands as a node, the rest of the bodies as a node, and the ball as a node, as illustrated in", + "bbox": [ + 89, + 780, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/ad1feb3457ae5e2008a94cf8a2b172e78d088939e7c72ec242297faba93355d5.jpg", + "image_caption": [ + "(a) Ref" + ], + "image_footnote": [], + "bbox": [ + 516, + 90, + 566, + 148 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/72bf60387a71d1f6b94b2c64e199716d38b853e3d754056477b758776e778f45.jpg", + "image_caption": [ + "(b) w/o CGR" + ], + "image_footnote": [], + "bbox": [ + 586, + 90, + 627, + 148 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/689532f55a0f2f189061e0ed895870a83cb141124050d482b950fa95df31df79.jpg", + "image_caption": [ + "(c) w/ CGR" + ], + "image_footnote": [], + "bbox": [ + 651, + 90, + 696, + 148 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/3c45c633681c04d10ae1ed42321ccb77759132629c13643a148b67e9387153c9.jpg", + "image_caption": [ + "(d) Ref" + ], + "image_footnote": [], + "bbox": [ + 720, + 90, + 754, + 148 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/5f765883bad63b19c40eebe16fabc10a5d04c7ad37ebfabb598c519ff5e2dec3.jpg", + "image_caption": [ + "(e) w/o CGR" + ], + "image_footnote": [], + "bbox": [ + 792, + 90, + 828, + 148 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/8721a308430ce184e4e4febc7fc4e7bb0cacbac0cae4ce642e1c47a77909e750.jpg", + "image_caption": [ + "(f) w/ CGR" + ], + "image_footnote": [], + "bbox": [ + 854, + 90, + 888, + 148 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/03c397008252e041c9758785dc483f83af0ecc0a8d92abb5af498507e4a4bfa4.jpg", + "image_caption": [ + "(g) Ref", + "Figure 5. The HOI imitation falls into kinematic local-optimal solutions without Contact Graph Reward (CGR): (b) use the head to help control the ball; (e) use the wrist to contact the ball; (h) fail to catch the object; (k) support the table to keep balance. In comparison, the guidance of CGR effectively yields precise interactions, as shown in (c, f, i, l).", + "Fig. 4. This simple CG is effective enough for learning all the basketball skills covered in BallPlay-M." + ], + "image_footnote": [], + "bbox": [ + 516, + 166, + 571, + 224 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/ba1cbea2dd01c3597180c785e80b43508d8f858293d76242eaf4fcc318ea72f8.jpg", + "image_caption": [ + "(h) w/o CGR" + ], + "image_footnote": [], + "bbox": [ + 583, + 166, + 637, + 224 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/c4ebc0b7cd1f8196316cdd8d15675b741446c7d18c237f4e8c29892cd7cd4ecd.jpg", + "image_caption": [ + "(i) w/ CGR" + ], + "image_footnote": [], + "bbox": [ + 643, + 167, + 700, + 224 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/6b7a535dbe9ba24512c66e8dc1fb7755c427e97164c22835565b00d4b565141f.jpg", + "image_caption": [ + "(j) Ref." + ], + "image_footnote": [], + "bbox": [ + 725, + 164, + 759, + 224 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/ae9a51cf90bab51e9ac53f3a9514f7000a536b0cdd7ad88a6ae4e65e8877948f.jpg", + "image_caption": [ + "(k) w/o CGR on table" + ], + "image_footnote": [], + "bbox": [ + 797, + 166, + 830, + 224 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/82aca685518671e1e76ccafc72f389ac4cc54cac48dde76c3c143520f3a37d32.jpg", + "image_caption": [ + "(I) w/CGR" + ], + "image_footnote": [], + "bbox": [ + 862, + 165, + 895, + 224 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4.2. Contact Graph Reward", + "text_level": 1, + "bbox": [ + 511, + 392, + 725, + 406 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Kinematic imitation rewards [32, 33, 52] fall short in measuring contacts and thus yield poor performance in HOI imitation. For example, during the toss skill training (Fig. 5 (a)), the contact between the right hand and the ball usually cause the ball to drop, making the kinematic rewards drastically smaller. In this case, the policy may learn a local-optimal solution that uses the head and left hands to stabilize the ball, as demonstrated in Fig. 5 (b). To tackle this problem, we propose the Contact Graph Reward (CGR) as a critical complement of kinematic imitation rewards to learn precise contact imitation. The CG error is defined as $e_t^{cg} = |s_t^{cg} - \\hat{s}_t^{cg}|$ , where $|\\cdot|$ calculates element-wise absolute value, $s_t^{cg}$ and $\\hat{s}_t^{cg}$ denotes the simulated and reference CG state respectively, which is the edge set $\\mathcal{E}_t \\in \\{0,1\\}^{k(k-1)/2}$ . The CGR is measured by CG error, with independent weights on different elements:", + "bbox": [ + 511, + 410, + 906, + 652 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nr _ {t} ^ {c g} = \\exp \\left(- \\sum_ {j = 1} ^ {J} \\boldsymbol {\\lambda} ^ {c g} [ j ] * \\boldsymbol {e} _ {t} ^ {c g} [ j ]\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 594, + 662, + 905, + 705 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $e_t^{cg}[j]$ is the $j$ th element of $e_t^{cg} \\in \\{0,1\\}^J$ , representing a contact error; $\\lambda^{cg}[j]$ is the $j$ th element of $\\lambda^{cg} \\in \\mathbb{R}^J$ , a hyperparameter controls the sensitivity of a contact. Fig. 5 and Tab. 1 justify the effectiveness of CGR, where experiments without CGR show inaccurate interactions while using CGR effectively eliminates such problems.", + "bbox": [ + 511, + 715, + 905, + 809 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Reusing Interaction Skills for High-level Tasks", + "text_level": 1, + "bbox": [ + 511, + 816, + 905, + 834 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we propose training a High-Level Controller (HLC) to reuse the interaction skills acquired through SkillMimic to accomplish high-level tasks. As depicted in Fig. 3 (c), the HLC takes as input the current HOI state $s_t$", + "bbox": [ + 511, + 839, + 905, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "17544", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/04724ae4c94bf9781dd4d256fe5f3b9446d27f5dd1cc13e89bc6d3688cd061de.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 101, + 88, + 261, + 157 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/81f434658e633d076d8eaa62b15976236335bd3e486e7e860ffde26067a8520c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 261, + 88, + 418, + 157 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/40c1b9a8efcc35a2af225b7f9d86bf0e8d7eb7795b04815c0fc713b101174e57.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 419, + 88, + 576, + 157 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a5afb105d44ad3196fd38b48c4c7a583220383258d035c1461dd6032cb3240e3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 578, + 88, + 735, + 157 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/acf43d35437bb8d4e8e36e9e62f0a96186c01bc3cf31b8cec8d97e2b9b6d2317.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 735, + 88, + 895, + 157 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b93370c0266fe545249025ae5b28597104a685b93558a0fb513c760796e91828.jpg", + "image_caption": [ + "(a) Layup", + "(f) Jump Shot" + ], + "image_footnote": [], + "bbox": [ + 101, + 172, + 261, + 241 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e3ca519f182dd1bbf68eee41d2ac166a9d0a7eff56f549063c723d74b35cf357.jpg", + "image_caption": [ + "(b) Turnaround Layup", + "(g) Rebound" + ], + "image_footnote": [], + "bbox": [ + 261, + 172, + 418, + 241 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/85843f0257971d7780dba6974a54d2f093ec8741dc12527353844c045d4b8edb.jpg", + "image_caption": [ + "(c) Dribble Right", + "(h) Pass", + "Figure 6. Simulated humans exhibit comprehensive basketball skills. SkillMimic can teach humans a wide range of basketball skills using the same configuration in a purely data-driven manner, covering almost all fundamental basketball skills. Keyframes are placed in chronological order from left to right." + ], + "image_footnote": [], + "bbox": [ + 419, + 172, + 576, + 242 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/df3eda1f4a2381cb6519824711901ee69e5af95c7bf06fc604314ee788742232.jpg", + "image_caption": [ + "(d) Dribble Forward", + "(i) Pickup" + ], + "image_footnote": [], + "bbox": [ + 578, + 172, + 733, + 242 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/dcd88d4d9142af324ac903e85a3acaa2c66af0927fab20a4dd879e676258c842.jpg", + "image_caption": [ + "(e) Dribble Turn", + "(j) Cross Leg" + ], + "image_footnote": [], + "bbox": [ + 736, + 172, + 895, + 242 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "and task observation $h_t$ , and outputs a discrete skill embedding $c_t$ that serves as inputs to a pre-trained SkillMimic policy, which subsequently generates actions for the humanoid.", + "bbox": [ + 89, + 337, + 480, + 383 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Compared to previous hierarchical approaches that are based on locomotion priors [34, 50], our method leverages interaction skills as prior, which significantly simplifies the design of task rewards and greatly accelerates training. Moreover, unlike their mandatory hierarchical structure, our approach is motivated by data efficiency - tasks can be alternatively solved through direct HOI imitation when sufficient demonstrations are available, as shown in pickup experiments in Sec. 4.2, Fig. 7, and Fig. 8.", + "bbox": [ + 89, + 383, + 482, + 518 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We present four representative high-level tasks:", + "bbox": [ + 109, + 518, + 424, + 535 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Throwing: throw the ball to approach a certain height, grab the rebound, and keep on throwing the ball.", + "- Heading: dribble the ball to approach the target position.", + "- Circling: dribble the ball around the target position following a target radius.", + "- Scoring: dribble the ball toward the basket, timing the layup to score, retrieving the rebound, and repeating." + ], + "bbox": [ + 89, + 539, + 482, + 643 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Utilizing interaction skills as prior significantly improves learning efficiency for complex, long-horizon tasks. For example, a decent HLC for the Scoring task only requires 3 hours of training on an NVIDIA RTX 4090 GPU.", + "bbox": [ + 89, + 647, + 483, + 708 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 723, + 223, + 739 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We conduct extensive experiments to evaluate the performance of SkillMimic in learning interaction skills and its application for high-level tasks. Our evaluation consists of two main parts: (1) assessing the capability of SkillMimic in learning interaction skills, including comparisons with existing methods and ablation studies, and (2) demonstrating the effectiveness of reusing acquired interaction skills to learn high-level tasks. Details of the experimental settings and additional experiments on skill switching and robustness are provided in the appendix.", + "bbox": [ + 89, + 750, + 482, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Metrics", + "text_level": 1, + "bbox": [ + 513, + 335, + 609, + 349 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We consider the following quantitative metrics:", + "bbox": [ + 511, + 359, + 826, + 375 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Accuracy. The overall accuracy of HOI imitation, abbreviated as Acc., is defined per frame, and deems imitation accurate when the object position and body position errors are both under the thresholds and the contacts are correct. The object threshold is defined as $0.2m$ . The body threshold is defined as $0.1m$ . The Acc. is calculated by averaging the values of all frames.", + "- Position Error. The Mean Per-Joint Position Error (MPJPE) of the body $(E_{\\mathrm{b - mpjpe}})$ and object $(E_{\\mathrm{o - mpjpe}})$ is used to evaluate the positional imitation performance (in mm) following [27].", + "- Contact Error. The contact error $E_{\\mathrm{cg}}$ , ranging from 0 to 1, is defined as $\\frac{1}{N}\\sum_{t=1}^{N}\\mathrm{MSE}(s_t^{cg}, \\hat{s}_t^{cg})$ , where $N$ is the total frames of the reference HOI data.", + "- Success Rate. The execution success rate of an interaction skill. Detailed definitions of success criteria for different skills are provided in the appendix." + ], + "bbox": [ + 513, + 377, + 905, + 636 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Evaluating Interaction Skill Learning", + "text_level": 1, + "bbox": [ + 511, + 650, + 838, + 666 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Since our approach is the first to learn interaction skills from HOI demonstrations, there are no direct benchmarks for comparison. Therefore, we adapt reward strategies commonly used in locomotion imitation, i.e., DeepMimic [32] and AMP [33], to an object-inclusive setting. Specifically, we adapt our reward to the styles of DeepMimic [32] and AMP [33] while keeping the other components unchanged for a fair comparison. We denote these variant versions as DeepMimic* and AMP*. In Tab. 2, we present the success rate on 4 typical basketball skills from BallPlay-M, where our method achieves state-of-the-art performance in learning basketball skills compared with variants using DeepMimic or AMP style rewards. We guess that AMP-like rewards are ineffective for learning precise interactions due to their coarse-grained supervision signals.", + "bbox": [ + 511, + 674, + 906, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "17545", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5f69af15bcda4cf123d2ad5345bdfddccd5dbeeb7090b13712e116920e733001.jpg", + "image_caption": [ + "Figure 7. Demonstration of the pickup skill learned from 40 HOI motion clips. Yellow denotes the initial frame. Left: The humanoid picks up a stationary ball effortlessly. Middle: The humanoid intercepts a ball with random velocity. Right: The humanoid adjusts after missing the ball initially (the frame in blue) and successfully retrieves it on the second attempt, showcasing the potential for learning robust and generalizable skills through extensive data collection." + ], + "image_footnote": [], + "bbox": [ + 93, + 87, + 903, + 179 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/f3f2072fc93360b2096d0785ad6c111fdc299e2626329155e0a539c845b10f7c.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
SkillMimic w/o MultiplicationSkillMimic w/o CGRSkillMimic
DatasetAcc. ↑Eb-mpipe ↓Eo-mpipe ↓Ecg ↓Acc. ↑Eb-mpipe ↓Eo-mpipe ↓Ecg ↓Acc. ↑Eb-mpipe ↓Eo-mpipe ↓Ecg ↓
GRAB [40]27.0%44.7180.20.72438.6%91.0180.10.33795.4%71.178.00.026
BallPlay-V7.5%40.11662.50.30613.6%88.5155.30.41282.4%56.882.90.087
", + "bbox": [ + 91, + 258, + 906, + 327 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/10ffb67c47a3e4d65b338115b14d2b8d4ebe441f9bfa7ca314e20d35c38a1ff9.jpg", + "table_caption": [ + "Table 1. Ablation study on imitation reward design. We experiment on two HOI datasets: GRAB and BallPlay-V. The result highlights the importance of the CG reward and multiplication in achieving precise HOI imitation." + ], + "table_footnote": [], + "table_body": "
MethodSuccess Rate ↑
Pick UpDribble ForwardLayupShot
DeepMimic*[32]19.6%68.5%98.9%97.8%
AMP*[33]0.0%13.6%0.0%1.6%
SkillMimic (ours)86.7%79.6%99.1%97.9%
", + "bbox": [ + 91, + 391, + 483, + 474 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Success rates across four typical basketball skills in BallPlay-M. Our method significantly outperforms variant methods that use imitation reward styles of DeepMimic and AMP.", + "bbox": [ + 89, + 486, + 483, + 529 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9b584801401090867cdb9c8cea8d03b8c41d3aa984d1ecc8e2d3803abc52f284.jpg", + "image_caption": [ + "(a) $1 - 0.5\\%$" + ], + "image_footnote": [], + "bbox": [ + 94, + 554, + 189, + 628 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1066c9831f6721218ed935f2e4e36752dd65cf90a33c01c27bb5216dc4210ee6.jpg", + "image_caption": [ + "(b) $10 - 12.2\\%$" + ], + "image_footnote": [], + "bbox": [ + 192, + 555, + 285, + 628 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1956039e558593426c1f1035c2d99e487675a856a2c88df695abc25ee6606ba5.jpg", + "image_caption": [ + "(c) $40 - 76.5\\%$", + "Figure 8. Pickup generalization performance with different training data scales. First number: clips; second: success rate. In each test, 1000 balls are randomly placed within 1 to 5 meters away from the center. Yellow dots indicate successful pickups and green dots represent failures. See Sec. 4.2 for details." + ], + "image_footnote": [], + "bbox": [ + 289, + 555, + 383, + 628 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0c6d6f488132e4a57a385ee583a13db367a78b0feb194e51502ad4bce15af87a.jpg", + "image_caption": [ + "(d) $131 - 85.6\\%$" + ], + "image_footnote": [], + "bbox": [ + 383, + 555, + 478, + 628 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In Tab. 1, we do ablation on reward design, presenting detailed metrics of HOI imitation across two diverse datasets, GRAB [40] and BallPlay-V. The results indicate that the absence of reward multiplication leads to imbalanced sub-reward learning, e.g., favoring body motion $(E_{\\mathrm{b - m}})$ over interactions. The absence of CGR leads to kinematic local optima. In contrast, our complete setting (SkillMimic) yields significantly higher performance on interaction-related metrics (Acc., $E_{\\mathrm{o - m}}$ , and $E_{\\mathrm{cg}}$", + "bbox": [ + 89, + 763, + 483, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2.1. Ablation on Contact Graph Reward", + "text_level": 1, + "bbox": [ + 511, + 393, + 810, + 409 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We conduct ablation experiments on diverse datasets to comprehensively evaluate the impact of Contact Graph Reward (CGR) on interaction skill learning. Fig. 5 illustrates the qualitative results of ablating CGR on the GRAB and BallPlay-V datasets. It can be clearly observed that without CGR, the humanoid often resorts to incorrect contacts to achieve more stable object control, such as using its head to push the ball, clamping the ball with its hands and legs, or using its hands to support itself on a table for balance. These examples illustrate typical kinematic local optima. By incorporating CGR, these local optima are effectively eliminated, resulting in accurate interactions. Additionally, the result in Tab. 1 also shows that, without CGR, the contact error $(E_{\\mathrm{cg}})$ significantly increases, leading to kinematic local optima and causes a noticeable decline in overall imitation accuracy. These results underscore the critical importance of CGR for learning correct interactions.", + "bbox": [ + 511, + 414, + 906, + 672 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2.2. Data Scale and Generalization Performance", + "text_level": 1, + "bbox": [ + 511, + 683, + 864, + 696 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To examine how the generalization performance grows as the data scale increases, we conduct ablation experiments on the pickup skill. We create four datasets by randomly selecting pickup clips from BallPlay-M, containing 1, 10, 40, and 131 clips, respectively. We train policy on each dataset with approximately 3.2 billion samples using SkillMimic.", + "bbox": [ + 511, + 703, + 906, + 794 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To evaluate the generalization performance, we randomly place balls within a circular area ranging from 1 to 5 meters around the humanoid—this distribution significantly exceeds the coverage of the reference data. Fig. 7 and Fig. 8 show the results. We can see that as the data scale increases, pickup performance improves significantly due to broader state transition coverage in larger datasets. This indicates", + "bbox": [ + 511, + 795, + 908, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "17546", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/93390c62e4ea195cf924a43c76988a9ecfb43c94be15d6343eec47bb45792817.jpg", + "image_caption": [ + "(a) Manual Control" + ], + "image_footnote": [], + "bbox": [ + 96, + 85, + 295, + 169 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c5b9ed00f9c1e039bfa83d574b66cac390c92c1c4cf87f7da5f419c8ff92a787.jpg", + "image_caption": [ + "(b) Scoring", + "Figure 9. Our method supports training a single IS policy under a unified configuration to acquire various interaction skills. These interaction skills can be flexibly switched, as illustrated in (a), where yellow denotes shot, blue denotes pickup, and green denotes turnaround layup. Complex, long-horizon tasks can be easily achieved by training a high-level controller (HLC) to manage switching of the learned interaction skills: (b) scoring from random positions (c) dribbling to target locations, and (d) dribbling along an expanding radius." + ], + "image_footnote": [], + "bbox": [ + 297, + 97, + 496, + 169 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6fb08d22749ebae1504114fcf4241bec5e9fb5dc2b0bc7186ac5f8c2fb098f47.jpg", + "image_caption": [ + "(c) Heading" + ], + "image_footnote": [], + "bbox": [ + 500, + 80, + 697, + 169 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0d50dc71dd70eb9377b14d06733963e465af0f4a1ab4d295f221d47723705871.jpg", + "image_caption": [ + "(d) Circling" + ], + "image_footnote": [], + "bbox": [ + 700, + 80, + 901, + 169 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/46607c7196da97152a749342ce831ee62b0ae5e96d83061799acae37d040ddc1.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodSuccess Rate↑
HeadingCirclingThrowingScoring
PPO[36]0.70%11.14%0.00%0.00%
ASE[34]0.19%4.37%0.00%0.00%
ASE*[34]0.31%7.21%0.00%0.00%
SkillMimic (ours)93.04%79.92%93.40%80.25%
", + "bbox": [ + 101, + 268, + 473, + 363 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3. Success rates on 4 high-level basketball tasks. Both PPO (learn from scratch), ASE (using locomotion prior), and $\\mathrm{ASE^{*}}$ (using body and object motion prior) fail to converge on all tasks. In contrast, leveraging interaction skill priors acquired through SkillMimic, our method effectively learns these difficult tasks.", + "bbox": [ + 89, + 376, + 482, + 446 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "our method's strong scalability and potential for large-scale learning. It is worth noting that we do not incorporate any additional designs for robustness and generalization.", + "bbox": [ + 89, + 474, + 482, + 518 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Evaluating High-level Tasks", + "text_level": 1, + "bbox": [ + 89, + 530, + 341, + 546 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We conduct experiments on four high-level tasks described in Sec. 3.5. We employ a single IS policy as the interaction skill prior, trained using SkillMimic for around 4.5 billion samples across seven skills: pickup, layup, turnaround layup, dribble left, dribble right, dribble forward, and shot. During training, the IS policy is fixed, and only the HLC is trained. For tasks such as throwing, heading, and circling, we train for around 0.4 billion samples. For the more challenging scoring task, we train for around 1.2 billion samples. To objectively evaluate the performance of our method, we compared it against three baseline methods: (1) learning from scratch using PPO [36], (2) learning with body motion priors using ASE [34], and (3) learning with body and object motion priors using ASE (denoted as $\\mathrm{ASE^{*}}$ ). The low-level controllers of ASE and $\\mathrm{ASE^{*}}$ are trained using the same training steps as our IS policy. For task training, all methods are trained using identical task rewards and simulation steps for fair comparisons.", + "bbox": [ + 88, + 551, + 482, + 824 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Fig. 9 (a) showcases an example of manually controlling the pre-trained IS policy to perform skill switching. Fig. 9 (b) shows several trajectories of the HLC performing skill controls to accurately deliver the ball to the target basket. Fig. 9 (c) & (d) present a top-down view of the HLC control", + "bbox": [ + 89, + 825, + 482, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ling the humanoid to dribble the ball toward target locations and along an expanding radius. As shown in Tab. 3, PPO and ASE yield low success rates because the tasks are too challenging to master without interaction priors. The failure of $\\mathrm{ASE^{*}}$ can be traced back to its low-level controller's inadequate ability to imitate interaction skills. This is primarily because adversarial rewards fail to provide effective guidance for precise interaction learning, which is similar to the failure mode observed in $\\mathrm{AMP^{*}}$ . In contrast, by utilizing the interaction skills learned through SkillMimic, we greatly improve the performance of learning complex high-level interaction tasks.", + "bbox": [ + 511, + 271, + 906, + 450 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Discussion and Future Work", + "text_level": 1, + "bbox": [ + 511, + 465, + 777, + 482 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We presented a novel framework for learning diverse interaction skills with a unified HOI imitation reward using reinforcement learning. Our approach demonstrates multi-skill mastery, skill reusability, and clear data scaling benefits, highlighting the significant potential of HOI imitation for learning complex interaction skills. However, the full application of this paradigm requires further exploration. Several key areas deserve particular attention:", + "bbox": [ + 511, + 492, + 905, + 612 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Firstly, exploring more general HOI skill learning within a single policy, such as household tasks or other types of sports, presents challenges in handling generalization across different object geometries. Moreover, perceiving multiple objects simultaneously requires a more general environmental perception approach, rather than directly obtaining privileged information for each object.", + "bbox": [ + 511, + 613, + 905, + 718 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Secondly, incorporating finer-grained control conditions, such as tracking [27], could enable more flexible control and generalization capabilities as HOI data scales up.", + "bbox": [ + 511, + 719, + 903, + 763 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Thirdly, the difficulty of collecting HOI data highlights the importance of exploring methods to learn robust interaction skills with limited HOI data, as well as investigating HOI data augmentation and generation techniques.", + "bbox": [ + 511, + 763, + 903, + 825 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Lastly, applying this technology to real robots would be highly valuable, such as enabling real-world humanoid robots to play basketball games. This endeavor could involve various challenges, including HOI data retargeting and sim-to-real transfer issues.", + "bbox": [ + 511, + 825, + 903, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "17547", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgment", + "text_level": 1, + "bbox": [ + 91, + 90, + 240, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We would like to thank Mingyang Su for his valuable assistance in motion capture data collection and Shunlin Lu for helpful discussions. This work was supported in part by Unitree Robotics, International Digital Economy Academy (IDEA), and the Innovation and Technology Fund of HK-SAR (GHX/054/21GD).", + "bbox": [ + 89, + 114, + 483, + 205 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 217, + 187, + 233 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Kai Arulkumaran, Marc Peter Deisenroth, Miles Brundage, and Anil Anthony Bharath. Deep reinforcement learning: A brief survey. IEEE Signal Processing Magazine, 34(6):26-38, 2017. 2", + "[2] Jinseok Bae, Jungdam Won, Donggeun Lim, Cheol-Hui Min, and Young Min Kim. Pmp: Learning to physically interact with environments using part-wise motion priors. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-10, 2023. 1, 2", + "[3] Kevin Black, Noah Brown, Danny Driess, Adnan Esmail, Michael Equi, Chelsea Finn, Niccolo Fusai, Lachy Groom, Karol Hausman, Brian Ichter, et al. Pi0: A vision-language-action flow model for general robot control. arXiv preprint arXiv:2410.24164, 2024. 3", + "[4] Jona Braun, Sammy Christen, Muhammed Kocabas, Emre Aksan, and Otmar Hilliges. Physically plausible full-body hand-object interaction synthesis. arXiv preprint arXiv:2309.07907, 2023. 2", + "[5] Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Xi Chen, Krzysztof Choromanski, Tianli Ding, Danny Driess, Avinava Dubey, Chelsea Finn, et al. Rt-2: Vision-language-action models transfer web knowledge to robotic control. arXiv preprint arXiv:2307.15818, 2023. 3", + "[6] Stelian Coros, Philippe Beaudoin, and Michiel Van de Panne. Generalized biped walking control. ACM Transactions On Graphics (TOG), 29(4):1-9, 2010. 2", + "[7] Jieming Cui, Tengyu Liu, Nian Liu, Yaodong Yang, Yixin Zhu, and Siyuan Huang. Anyskill: Learning openvocabulary physical skill for interactive agents. In Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2", + "[8] Zhiyang Dou, Xuelin Chen, Qingnan Fan, Taku Komura, and Wenping Wang. C-ase: Learning conditional adversarial skill embeddings for physics-based characters. In SIGGRAPH Asia 2023 Conference Papers, pages 1-11, 2023. 1, 2", + "[9] Zicong Fan, Omid Taheri, Dimitrios Tzionas, Muhammed Kocabas, Manuel Kaufmann, Michael J Black, and Otmar Hilliges. Arctic: A dataset for dexterous bimanual handobject manipulation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12943-12954, 2023. 5", + "[10] Zipeng Fu, Qingqing Zhao, Qi Wu, Gordon Wetzstein, and Chelsea Finn. Humanplus: Humanoid shadowing and imitation from humans. In 8th Annual Conference on Robot Learning, 2024. 1", + "[11] Zipeng Fu, Tony Z Zhao, and Chelsea Finn. Mobile aloha: Learning bimanual mobile manipulation with low-cost" + ], + "bbox": [ + 93, + 243, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "whole-body teleoperation. arXiv preprint arXiv:2401.02117, 2024.3", + "[12] Jiawei Gao, Ziqin Wang, Zeqi Xiao, Jingbo Wang, Tai Wang, Jinkun Cao, Xiaolin Hu, Si Liu, Jifeng Dai, and Jiangmiao Pang. Coohoi: Learning cooperative human-object interaction with manipulated object dynamics. Advances in Neural Information Processing Systems, 37:79741-79763, 2024. 1", + "[13] Haoran Geng, Helin Xu, Chengyang Zhao, Chao Xu, Li Yi, Siyuan Huang, and He Wang. Gapartner: Cross-category domain-generalizable object perception and manipulation via generalizable and actionable parts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7081-7091, 2023. 5", + "[14] Anindita Ghosh, Rishabh Dabral, Vladislav Golyanik, Christian Theobalt, and Philipp Slusallek. Imos: Intent-driven full-body motion synthesis for human-object interactions. In Eurographics, 2023. 2", + "[15] Mohamed Hassan, Yunrong Guo, Tingwu Wang, Michael Black, Sanja Fidler, and Xue Bin Peng. Synthesizing physical character-scene interactions. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-9, 2023. 1, 2", + "[16] Tairan He, Zhengyi Luo, Xialin He, Wenli Xiao, Chong Zhang, Weinan Zhang, Kris M Kitani, Changliu Liu, and Guanya Shi. Omnih2o: Universal and dexterous human-to-humanoid whole-body teleoperation and learning. In 8th Annual Conference on Robot Learning, 2024. 1", + "[17] Edmond SL Ho, Taku Komura, and Chiew-Lan Tai. Spatial relationship preserving character motion adaptation. In ACM SIGGRAPH 2010 papers, pages 1-8, 2010. 3", + "[18] Jonathan Ho and Stefano Ermon. Generative adversarial imitation learning. Advances in neural information processing systems, 29, 2016. 2", + "[19] Jessica K Hodgins, Wayne L Wooten, David C Brogan, and James F O'Brien. Animating human athletics. In Proceedings of the 22nd annual conference on Computer graphics and interactive techniques, pages 71-78, 1995. 2", + "[20] Nan Jiang, Tengyu Liu, Zhexuan Cao, Jieming Cui, Zhiyuan Zhang, Yixin Chen, He Wang, Yixin Zhu, and Siyuan Huang. Full-body articulated human-object interaction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9365-9376, 2023. 2", + "[21] Nan Jiang, Zhiyuan Zhang, Hongjie Li, Xiaoxuan Ma, Zan Wang, Yixin Chen, Tengyu Liu, Yixin Zhu, and Siyuan Huang. Scaling up dynamic human-scene interaction modeling. arXiv preprint arXiv:2403.08629, 2024. 2", + "[22] Jordan Juravsky, Yunrong Guo, Sanja Fidler, and Xue Bin Peng. Padl: Language-directed physics-based character control. In SIGGRAPH Asia 2022 Conference Papers, pages 1-9, 2022. 2", + "[23] Jiaman Li, Alexander Clegg, Roozbeh Mottaghi, Jiajun Wu, Xavier Puig, and C Karen Liu. Controllable human-object interaction synthesis. arXiv preprint arXiv:2312.03913, 2023. 1, 2", + "[24] Jiaman Li, Jiajun Wu, and C Karen Liu. Object motion guided human motion synthesis. ACM Transactions on Graphics (TOG), 42(6):1-11, 2023. 2" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "17548", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[25] Libin Liu and Jessica Hodgins. Learning to schedule control fragments for physics-based characters using deep q-learning. ACM Transactions on Graphics (TOG), 36(3):1-14, 2017. 2", + "[26] Libin Liu and Jessica Hodgins. Learning basketball dribbling skills using trajectory optimization and deep reinforcement learning. ACM Transactions on Graphics (TOG), 37(4):1-14, 2018. 2", + "[27] Zhengyi Luo, Jinkun Cao, Kris Kitani, Weipeng Xu, et al. Perpetual humanoid control for real-time simulated avatars. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10895-10904, 2023. 6, 8", + "[28] Zhengyi Luo, Jinkun Cao, Sammy Christen, Alexander Winkler, Kris Kitani, and Weipeng Xu. Omnigrasp: Grasping diverse objects with simulated humanoids. Advances in Neural Information Processing Systems, 37:2161-2184, 2024. 1", + "[29] Zhengyi Luo, Jiashun Wang, Kangni Liu, Haotian Zhang, Chen Tessler, Jingbo Wang, Ye Yuan, Jinkun Cao, Zihui Lin, Fengyi Wang, et al. Smplolympics: Sports environments for physically simulated humanoids. arXiv preprint arXiv:2407.00187, 2024. 2", + "[30] Vinod Nair and Geoffrey E Hinton. Rectified linear units improve restricted boltzmann machines. In Proceedings of the 27th international conference on machine learning (ICML-10), pages 807–814, 2010. 4", + "[31] Liang Pan, Jingbo Wang, Buzhen Huang, Junyu Zhang, Haofan Wang, Xu Tang, and Yangang Wang. Synthesizing physically plausible human motions in 3d scenes. In 2024 International Conference on 3D Vision (3DV), pages 1498-1507. IEEE, 2024. 2", + "[32] Xue Bin Peng, Pieter Abbeel, Sergey Levine, and Michiel van de Panne. Deepmimic. ACM Transactions on Graphics, page 1-14, 2018. 1, 2, 4, 5, 6, 7", + "[33] Xue Bin Peng, Ze Ma, Pieter Abbeel, Sergey Levine, and Angjoo Kanazawa. Amp: Adversarial motion priors for stylized physics-based character control. ACM Transactions on Graphics, page 1-20, 2021. 2, 3, 4, 5, 6, 7", + "[34] Xue Bin Peng, Yunrong Guo, Lina Halper, Sergey Levine, and Sanja Fidler. Ase: Large-scale reusable adversarial skill embeddings for physically simulated characters. ACM Trans. Graph., 41(4), 2022. 1, 2, 3, 4, 5, 6, 8", + "[35] Jiawei Ren, Mingyuan Zhang, Cunjun Yu, Xiao Ma, Liang Pan, and Ziwei Liu. Insactor: Instruction-driven physics-based characters. Advances in Neural Information Processing Systems, 36, 2024. 2", + "[36] John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347, 2017. 8", + "[37] Sebastian Starke, Yiwei Zhao, Taku Komura, and Kazi Zaman. Local motion phases for learning multi-contact character movements. ACM Trans. Graph., 39(4), 2020. 2", + "[38] Sebastian Starke, Yiwei Zhao, Fabio Zinno, and Taku Komura. Neural animation layering for synthesizing martial arts movements. ACM Trans. Graph., 40(4), 2021. 2", + "[39] Jingkai Sun, Qiang Zhang, Yiqun Duan, Xiaoyang Jiang, Chong Cheng, and Renjing Xu. Prompt, plan, perform: Llm-based humanoid control via quantized imitation learning. arXiv preprint arXiv:2309.11359, 2023. 2" + ], + "bbox": [ + 91, + 92, + 483, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[40] Omid Taheri, Nima Ghorbani, Michael J. Black, and Dimitrios Tzionas. GRAB: A dataset of whole-body human grasping of objects. In European Conference on Computer Vision (ECCV), 2020. 7", + "[41] Jie Tan, Yuting Gu, C Karen Liu, and Greg Turk. Learning bicycle stunts. ACM Transactions on Graphics (TOG), 33 (4):1-12, 2014. 2", + "[42] Chen Tessler, Yoni Kasten, Yunrong Guo, Shie Mannor, Gal Chechik, and Xue Bin Peng. Calm: Conditional adversarial latent models for directable virtual characters. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-9, 2023. 1, 2", + "[43] Yinhuai Wang, Jing Lin, Ailing Zeng, Zhengyi Luo, Jian Zhang, and Lei Zhang. Physhoi: Physics-based imitation of dynamic human-object interaction. arXiv preprint arXiv:2312.04393, 2023. 3", + "[44] Jungdam Won, Deepak Gopinath, and Jessica Hodgins. Physics-based character controllers using conditional vaes. ACM Transactions on Graphics (TOG), 41(4):1-12, 2022. 2", + "[45] Yan Wu, Jiahao Wang, Yan Zhang, Siwei Zhang, Otmar Hilliges, Fisher Yu, and Siyu Tang. Saga: Stochastic whole-body grasping with contact. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 2", + "[46] Zeqi Xiao, Tai Wang, Jingbo Wang, Jinkun Cao, Wenwei Zhang, Bo Dai, Dahua Lin, and Jiangmiao Pang. Unified human-scene interaction via prompted chain-of-contacts. In The Twelfth International Conference on Learning Representations, 2024. 1", + "[47] Sirui Xu, Zhengyuan Li, Yu-Xiong Wang, and Liang-Yan Gui. Interdiff: Generating 3d human-object interactions with physics-informed diffusion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14928-14940, 2023. 1, 2", + "[48] Sirui Xu, Ziyin Wang, Yu-Xiong Wang, and Liang-Yan Gui. Interdreamer: Zero-shot text to 3d dynamic human-object interaction. arXiv preprint arXiv:2403.19652, 2024. 2", + "[49] KangKang Yin, Kevin Loken, and Michiel Van de Panne. Simbicon: Simple biped locomotion control. ACM Transactions on Graphics (TOG), 26(3):105-es, 2007. 2", + "[50] Haotian Zhang, Ye Yuan, Viktor Makoviychuk, Yunrong Guo, Sanja Fidler, Xue Bin Peng, and Kayvon Fatahalian. Learning physically simulated tennis skills from broadcast videos. ACM Trans. Graph., 2023. 1, 2, 3, 6", + "[51] Hui Zhang, Sammy Christen, Zicong Fan, Luocheng Zheng, Jemin Hwangbo, Jie Song, and Otmar Hilliges. Artigrasp: Physically plausible synthesis of bi-manual dexterous grasping and articulation. In 2024 International Conference on 3D Vision (3DV), pages 235-246. IEEE, 2024. 5", + "[52] Yunbo Zhang, Deepak Gopinath, Yuting Ye, Jessica Hodgins, Greg Turk, and Jungdam Won. Simulation and retargeting of complex multi-character interactions. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-11, 2023. 3, 5", + "[53] Qingxu Zhu, He Zhang, Mengting Lan, and Lei Han. Neural categorical priors for physics-based character control. ACM Transactions on Graphics (TOG), 42(6):1-16, 2023. 2" + ], + "bbox": [ + 516, + 92, + 905, + 883 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "17549", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/SkillMimic_ Learning Basketball Interaction Skills from Demonstrations/70ea78a5-d7df-4937-ac48-382497691b79_model.json b/2025/SkillMimic_ Learning Basketball Interaction Skills from Demonstrations/70ea78a5-d7df-4937-ac48-382497691b79_model.json new file mode 100644 index 0000000000000000000000000000000000000000..dd68a64b22670b5c49f7d118eff34e87671a7ce6 --- /dev/null +++ b/2025/SkillMimic_ Learning Basketball Interaction Skills from Demonstrations/70ea78a5-d7df-4937-ac48-382497691b79_model.json @@ -0,0 +1,2805 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.131, + 0.12, + 0.867, + 0.143 + ], + "angle": 0, + "content": "SkillMimic: Learning Basketball Interaction Skills from Demonstrations" + }, + { + "type": "text", + "bbox": [ + 0.111, + 0.168, + 0.885, + 0.207 + ], + "angle": 0, + "content": "Yinhuai Wang\\(^{1,2*}\\) Qihan Zhao\\(^{1*}\\) Runyi Yu\\(^{1,2*}\\) Hok Wai Tsui\\(^{1}\\) Ailing Zeng\\(^{6\\dagger}\\) Jing Lin\\(^{5}\\) Zhengyi Luo\\(^{7}\\) Jiwen Yu\\(^{3}\\) Xiu Li\\(^{4}\\) Qifeng Chen\\(^{1}\\) Jian Zhang\\(^{3\\dagger}\\) Lei Zhang\\(^{5}\\) Ping Tan\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.214, + 0.885, + 0.266 + ], + "angle": 0, + "content": "\\(^{1}\\)Hong Kong University of Science and Technology \\(^{2}\\)Unitree Robotics \\(^{3}\\)Peking University Shenzhen Graduate School \\(^{4}\\)Tsinghua University \\(^{5}\\)International Digital Economy Academy \\(^{6}\\)Tencent \\(^{7}\\)Carnegie Mellon University yinhuai.wang@connect.ust.hk pingtan@ust.hk" + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.285, + 0.907, + 0.414 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.422, + 0.908, + 0.466 + ], + "angle": 0, + "content": "Figure 1. We propose a novel approach that for the first time enables physically simulated humanoids to learn a variety of basketball interaction skills from Human-Object Interaction (HOI) data, including but not limited to shooting (blue), retrieving (red), and turnaround layup (yellow). Once acquired, these interaction skills can be composed to accomplish complex tasks, such as consecutive scoring (green)." + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.482, + 0.327, + 0.497 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.515, + 0.486, + 0.864 + ], + "angle": 0, + "content": "Traditional reinforcement learning methods for human-object interaction (HOI) rely on labor-intensive, manually designed skill rewards that do not generalize well across different interactions. We introduce SkillMimic, a unified data-driven framework that fundamentally changes how agents learn interaction skills by eliminating the need for skill-specific rewards. Our key insight is that a unified HOI imitation reward can effectively capture the essence of diverse interaction patterns from HOI datasets. This enables SkillMimic to learn a single policy that not only masters multiple interaction skills but also facilitates skill transitions, with both diversity and generalization improving as the HOI dataset grows. For evaluation, we collect and introduce two basketball datasets containing approximately 35 minutes of diverse basketball skills. Extensive experiments show that SkillMimic successfully masters a wide range of basketball skills including stylistic variations in dribbling, layup, and shooting. Moreover, these learned skills can be effectively composed by a high-level controller to accomplish complex and long-horizon tasks such as consecutive scoring, opening new possibilities for scalable and generalizable interaction skill learning. Project page: https://ingrid789.github.io/SkillMimic/" + }, + { + "type": "title", + "bbox": [ + 0.515, + 0.482, + 0.645, + 0.497 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.508, + 0.907, + 0.841 + ], + "angle": 0, + "content": "Human-Object Interaction (HOI) represents one of the most fundamental yet challenging capabilities in robotics and character animation [10, 12, 16, 23, 28, 46, 47]. Among various HOI scenarios, basketball serves as an exemplary testbed that demands mastery of complex, dynamic, and precise interactions. Despite significant progress, existing physics-based character animation methods predominantly focus on imitation learning of locomotion skills [8, 32–34, 42] through Reinforcement Learning (RL), leaving a significant gap for HOI imitation. While some methods can learn interaction skills, such as striking pillars [34], playing tennis [50], climbing ropes [2], and carrying boxes [15], they typically involve handcrafted reward functions for each different interaction skill. However, designing skill-specific rewards is labor-intensive and fundamentally limits generalization. This limitation becomes particularly acute in basketball, where the complexity and diversity of skills would demand extraordinary engineering effort for reward design. Consequently, existing methods [2, 8, 15, 32–34, 42, 50] fail to provide a unified learning framework for a single policy to learn diverse interaction skills, let alone achieving complex long-term tasks such as consecutive basketball scoring." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.841, + 0.909, + 0.903 + ], + "angle": 0, + "content": "How can a simulated humanoid learn a wide variety of interaction skills in a simple and scalable manner? Consider how humans naturally acquire complex skills: amateur basketball players often develop their expertise by watching" + }, + { + "type": "page_footnote", + "bbox": [ + 0.115, + 0.887, + 0.359, + 0.902 + ], + "angle": 0, + "content": "*Equal Contribution. † Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17540" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.152 + ], + "angle": 0, + "content": "and imitating game footage, even without formal coaching. This self-directed learning process involves synchronizing their body movements with both the observed player motions and ball trajectories through practice." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.154, + 0.483, + 0.38 + ], + "angle": 0, + "content": "Inspired by these observations, we propose a data-driven method called SkillMimic, which mimics HOI through RL to learn interaction skills. SkillMimic can learn various interaction skills purely from HOI datasets, such as diverse styles of basketball shooting, layup, and dribbling skills, and even a robust pickup skill that enables picking up balls in random locations and motions. Notably, SkillMimic uses the exact same configuration to learn different interaction skills, with identical hyperparameters. This allows us to train a single Interaction Skill (IS) policy to learn multiple interaction skills and achieve smooth skill switching. By training a high-level policy to reuse these learned interaction skills, we can accomplish challenging high-level tasks such as scoring, which requires the humanoid to composite diverse basketball interaction skills to score accurately." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.382, + 0.483, + 0.577 + ], + "angle": 0, + "content": "To advance basketball skill learning, we introduce two HOI datasets: BallPlay-V, which captures player and ball motion from RGB videos across eight basic skills, and BallPlay-M, which uses optical motion capture to record 35 minutes of comprehensive basketball interactions. Using these datasets, SkillMimic successfully learns diverse basketball skills through a unified IS policy. With an additional high-level controller for skill composition, our system can execute complex tasks like consecutive scoring. Our approach demonstrates superior efficiency in learning both interaction skills and multi-phase, long-horizon tasks compared to existing methods, opening new possibilities for scalable and generalizable interaction skill learning." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.58, + 0.411, + 0.594 + ], + "angle": 0, + "content": "Specifically, our contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.597, + 0.483, + 0.673 + ], + "angle": 0, + "content": "- SkillMimic: a data-driven paradigm for learning diverse humanoid interaction skills through RL. It supports joint learning and switching of diverse basketball interaction skills, with skill diversity and generalization improving as the demonstration data grows, showcasing the scalability." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.673, + 0.483, + 0.747 + ], + "angle": 0, + "content": "- Contact graph: We propose a simple and general contact modeling method that applies to diverse skills, called the contact graph. A contact graph reward is designed to enable precise contact imitation, which proves to be critical for learning precise interaction skills." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.748, + 0.483, + 0.792 + ], + "angle": 0, + "content": "- Unified HOI imitation reward: We propose a set of important designs that form a unified reward configuration for imitation learning of various interaction skills." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.793, + 0.483, + 0.868 + ], + "angle": 0, + "content": "- A hierarchical solution for reusing interaction skills: We propose training a high-level controller to reuse the interaction skills acquired by SkillMimic to accomplish challenging high-level tasks, which for the first time achieves continuous basketball scoring." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.869, + 0.483, + 0.898 + ], + "angle": 0, + "content": "- BallPlay Datasets: We introduce two basketball datasets to facilitate research on interaction skill learning." + }, + { + "type": "list", + "bbox": [ + 0.092, + 0.597, + 0.483, + 0.898 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.09, + 0.655, + 0.105 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.117, + 0.907, + 0.465 + ], + "angle": 0, + "content": "Locomotion Imitation Learning. Mimicking human motion for robot control is an efficient method of learning humanoid skills. DeepMimic [32], a pioneer in this field, uses imitation learning to perform a variety of highly dynamic locomotion skills. The introduction of Generative Adversarial Imitation Learning (GAIL) [18] into humanoid imitation learning by AMP [33] lessens constraints on data alignment, thus enhancing its versatility. ASE [34] further amplifies motion diversity in GAIL training, incorporates a pre-trained low-level policy to acquire locomotion skills, and deploys a high-level policy to repurpose these locomotion skills for specific tasks, such as striking pillars. These methodologies are subsequently expanded in numerous efforts to augment condition and text control [8, 22, 29, 31, 35, 39, 42, 44, 53], as well as to design automated rewards [7]. However, a limitation of these studies is their exclusive focus on learning isolated locomotion skills through imitation. This requires additional pipeline and reward design for learning interaction skills, even for relatively simple ones, such as moving boxes [15] or striking pillars [34]. In contrast, our approach enables unified learning of diverse interaction skills, eliminating the need for skill-specific reward engineering." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.478, + 0.909, + 0.794 + ], + "angle": 0, + "content": "Learning Human-Object Interaction (HOI). Learning HOI in simulation has long posed challenges in the fields of graphics and robotics [14, 20, 21, 23, 24, 37, 38, 45, 47, 48]. Early studies often relied on manually designed control structures. For example, Hodgins et al. [19] implement a variety of sports interactions through the use of manually designed state machine controllers. Yin et al. [49] use finite-state machines to generate a large variety of gaits. Coros et al. [6] utilize an inverted pendulum model to control foot placement. Recently, deep reinforcement learning [1] has been widely employed in character control, giving rise to an array of methods based on policy network control. These methods have been applied to a range of challenging activities, such as basketball [26], skateboarding [25], and cycling [41], etc. The use of imitation rewards further lessens the reward design for locomotion styles. Zhang et al. [50] propose a framework to learn diverse tennis skills from broadcast videos. Bae et al. [2] incorporate multiple part-wise motion priors for diverse whole-body interactions. Similarly, Braun et al. [4] develop a framework for synthesizing whole-body grasps." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Applying imitation learning for HOI is intuitive, however, effectively learning diverse interaction skills in a unified framework remains unresolved. Naively extending locomotion imitation methods [32, 33] to HOI imitation often yields unstable and unnatural results, as these methods do not address the unbalanced rewards inherent in HOI, nor do they properly account for the crucial aspects of relative mo" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "17541" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.089, + 0.482, + 0.189 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.199, + 0.485, + 0.298 + ], + "angle": 0, + "content": "Figure 2. Concept of SkillMimic. We define an interaction skill as a set of Human-Object Interaction (HOI) state transitions that align with the intended skill semantics. These state transitions can be derived from captured HOI motion clips. If a simulated humanoid can manipulate objects such that the resulting HOI state transitions closely match those of the reference, we consider the humanoid to have successfully learned the interaction skill." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.324, + 0.483, + 0.475 + ], + "angle": 0, + "content": "tion and contact in HOI. Modeling interaction through interaction graphs is a well-established approach [17]. While this method has proven useful for HOI imitation [52], it remains limited by considering only kinematic relationships, failing to ensure proper contact and becoming unstable in learning diverse scenarios. In contrast, our contact graph imitation reward explicitly prioritize physical contact learning, significantly improving the success rate of HOI imitation and enabling unified HOI imitation of diverse complex interaction skills for the first time." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.476, + 0.483, + 0.507 + ], + "angle": 0, + "content": "This work builds upon and significantly extends ideas initially explored in our early technical report [43]." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.521, + 0.21, + 0.537 + ], + "angle": 0, + "content": "3. SkillMimic" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.547, + 0.484, + 0.699 + ], + "angle": 0, + "content": "In this section, we introduce a unified data-driven framework for simulated humanoids to learn various basketball interaction skills via simulation and reinforcement learning (RL). We begin by introducing how we define interaction skills with Human-Object Interaction (HOI) data (Sec. 3.2). Next, we explain how to learn diverse interaction skills by imitating HOI data (Sec. 3.3) and discuss the designing of a unified HOI imitation reward (Sec. 3.4). Finally, we demonstrate how the learned interaction skills can be reused to perform high-level tasks (Sec. 3.5)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.709, + 0.29, + 0.725 + ], + "angle": 0, + "content": "3.1. The BallPlay Dataset" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.731, + 0.483, + 0.839 + ], + "angle": 0, + "content": "To address the scarcity of basketball HOI data and facilitate research on interaction skill learning, we introduce two datasets: BallPlay-V, based on monocular vision estimation, containing eight clips of different basketball skills; and BallPlay-M, based on multi-view optical motion capture systems, containing 35 minutes of diverse basketball interactions. We provide more details in the appendix." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.849, + 0.458, + 0.866 + ], + "angle": 0, + "content": "3.2. Defining Interaction Skills Using HOI Data" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.903 + ], + "angle": 0, + "content": "We propose a data-driven approach to defining interaction skills as collections of Human-Object Interaction (HOI)" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.183 + ], + "angle": 0, + "content": "state transitions that align with the intended skill semantics. For instance, the interaction skill of \"picking up balls from various positions\" can be encapsulated by various HOI motion clips, each capturing a unique instance of the action. These clips collectively form a collection of HOI state transitions that represent the interaction skill." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.192, + 0.898, + 0.208 + ], + "angle": 0, + "content": "3.3. Learning Interaction Skills by HOI Imitation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.214, + 0.905, + 0.334 + ], + "angle": 0, + "content": "Considering an interaction skill defined by a set of reference HOI state transitions, if a humanoid can manipulate objects such that its HOI state transitions closely resemble the reference, we consider the humanoid to have successfully learned the interaction skill. Based on this concept, we propose a method for learning interaction skills by imitating HOI state transitions through RL, which we call SkillMimic. Fig. 2 illustrates the concept of SkillMimic." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.335, + 0.906, + 0.501 + ], + "angle": 0, + "content": "In contrast to Behavior Clone [3, 5, 11] methods which rely on precise state-action data, SkillMimic learns actions through RL using state-only trajectories, demonstrating better tolerance to data noise and higher data efficiency. Unlike previous RL methods [33, 34, 50] that require manually designed rewards for each interaction skill, SkillMimic is fully data-driven, skill-agnostic, and scalable, making it capable of learning a wide range of basketball skills within a unified solution. Fig. 6 shows diverse basketball skills acquired using SkillMimic. Fig. 7 illustrates how the performance of the pickup skill improves as the reference data increases." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.508, + 0.681, + 0.523 + ], + "angle": 0, + "content": "3.3.1. Training Pipeline" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.527, + 0.906, + 0.753 + ], + "angle": 0, + "content": "Fig. 3 (b) shows the training pipeline of SkillMimic. Given an HOI dataset with diverse skill-labeled clips (Fig. 3 (a)), SkillMimic trains an Interaction Skill (IS) policy by randomly selecting a frame from a clip corresponding to a chosen skill to initialize the humanoid and object states. The state \\( s_t \\) (Sec. 3.3.2) and skill label \\( c_j \\) are input into the IS policy, which predicts actions (Sec. 3.3.3) that are then simulated to produce the next state \\( s_{t+1} \\). A unified HOI imitation reward (Sec. 3.4) is designed to measure the consistency between the simulated state transitions and the reference transitions. When the humanoid falls or reaches the maximum simulation length, the environment resets, and the process repeats. After training converges, the IS policy enables the humanoid to execute interaction skills that closely mirror the reference demonstrations." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.762, + 0.682, + 0.775 + ], + "angle": 0, + "content": "3.3.2. HOI Observation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.906, + 0.901 + ], + "angle": 0, + "content": "The state observed by the IS policy is not under direct supervision and may theoretically encompass any information that is available in the simulation environment. Importantly, this information does not necessarily have to be included in the reference HOI data. Similar to prior arts [34], we transform all coordinates into the root local coordinate of the humanoid, which aligns the data distribution and benefits the generalization performance. For the humanoid, we observe" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17542" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.135, + 0.087, + 0.865, + 0.355 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.363, + 0.908, + 0.449 + ], + "angle": 0, + "content": "Figure 3. Our system consists of three parts. (a) First, we capture real-world basketball skills to create a large Human-Object Interaction (HOI) motion dataset. (b) Second, we train an Interaction Skill (IS) policy to learn interaction skills by imitating the corresponding HOI data through reinforcement learning. Specifically, the IS policy takes as input the HOI state \\( \\boldsymbol{s}_t \\) and skill label \\( \\boldsymbol{c}_j \\) and predicts the action \\( \\boldsymbol{a}_t \\). The new state \\( \\boldsymbol{s}_{t+1} \\) is calculated by the simulator. A unified HOI imitation reward is designed to imitate diverse HOI state transitions. (c) The third part involves training a High-Level Controller (HLC) to reuse the learned interaction skills for complex tasks. The HLC takes as input \\( \\boldsymbol{s}_t \\) and extra task observations \\( \\boldsymbol{h}_t \\), e.g., the basket position, and predicts the skill label \\( \\boldsymbol{c}_t \\) to drive a pre-trained IS policy." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.474, + 0.482, + 0.598 + ], + "angle": 0, + "content": "its global root height, local body position, rotation, position velocity, and angular velocity. These representations form the humanoid proprioception \\( \\pmb{o}_t^{prop} \\). In addition, we detect net contact forces \\( \\pmb{o}_t^f \\) for all fingertips, which helps to sense contact and accelerate training. For the object, we observe its local position, rotation, velocity, and angular velocity, which form the object observation \\( \\pmb{o}_t^{obj} \\). Finally, the state perceived by the IS policy is" + }, + { + "type": "equation", + "bbox": [ + 0.207, + 0.608, + 0.483, + 0.628 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {s} _ {t} = \\left\\{\\boldsymbol {o} _ {t} ^ {\\text {p r o p}}, \\boldsymbol {o} _ {t} ^ {f}, \\boldsymbol {o} _ {t} ^ {\\text {o b j}} \\right\\}. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.64, + 0.484, + 0.686 + ], + "angle": 0, + "content": "Skill labels \\( c_{j} \\) are also used as inputs to the IS policy to differentiate different skills. We use one-hot encodings to represent the skill label." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.695, + 0.281, + 0.71 + ], + "angle": 0, + "content": "3.3.3. IS Policy and Action" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.715, + 0.484, + 0.837 + ], + "angle": 0, + "content": "The policy output is modeled as a Gaussian distribution with dimensions equal to the DOF number of the humanoid robot, featuring constant variance. The mean is modeled by a three-layer MLP consisting of [1024, 512, 512] units with ReLU [30] activations. We use the action \\( \\boldsymbol{a}_t \\) sampled from the policy as the target joint rotations for a full set of PD controllers. The PD controllers adjust and output the joint torques to reach the target rotations." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.848, + 0.365, + 0.863 + ], + "angle": 0, + "content": "3.4. Unified HOI Imitation Reward" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "This HOI imitation reward function is pivotal as it must not only accurately gauge the efficacy of interaction skill ac" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.473, + 0.905, + 0.573 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.584, + 0.907, + 0.697 + ], + "angle": 0, + "content": "Figure 4. We propose the Contact Graph (CG) to model general contacts within an explicitly defined scene. The node stores a binary value that denotes whether it contacts other nodes. Each edge stores a binary value indicating whether the two connected nodes are in contact. The node definition is unified for a certain scene and shared between diverse interactive skills. For example, we define three nodes: hands, hands-exclusive body, and ball, to form a simple CG to model contacts for diverse basketball skills." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.729, + 0.905, + 0.76 + ], + "angle": 0, + "content": "quisition but also avoid the incorporation of skill-specific settings to maintain robust cross-skill generalizability." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.762, + 0.907, + 0.883 + ], + "angle": 0, + "content": "Classic imitation rewards often focus solely on imitating locomotions [32], lacking an established baseline for HOI imitation. Although adversarial imitation rewards [33, 34] are effective in locomotion tasks, we find that they perform poorly in HOI imitation. This may be because adversarial imitation rewards are relatively coarse-grained, making it difficult for them to provide the precise guidance needed for accurate interaction learning." + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.885, + 0.905, + 0.902 + ], + "angle": 0, + "content": "The body kinematics imitation reward \\( r_t^b \\) encourages the" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17543" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.41 + ], + "angle": 0, + "content": "alignment of body movements by considering joint position, rotation, and their velocities. Similarly, the object kinematics imitation reward \\( r_t^o \\) ensures consistent object movements with the reference. However, solely relying on these two rewards often leads to unstable learning and unnatural motions. We incorporate a relative motion reward \\( r_t^{rel} \\) to highlight the relative relationships in interactions, which is crucial for interaction naturalness. Recognizing the interlocking relationship between interaction and contact, we propose a straightforward and universal contact modeling approach called Contact Graph (CG). A corresponding CG imitation reward \\( r_t^{cg} \\) is designed to enhance the precision of interaction imitation. Additionally, we propose an adaptive velocity regularization term \\( r_t^{reg} \\) to suppress high-frequency jitters. These sub-rewards are multiplied rather than added to encourage balanced learning, thereby avoiding local optima, as justified in Tab. 1. The combination of these innovations forms a unified imitation reward, enabling the learning of diverse basketball skills in BallPlay-M with the same configuration. To conclude, the complete HOI imitation reward is" + }, + { + "type": "equation", + "bbox": [ + 0.181, + 0.42, + 0.483, + 0.439 + ], + "angle": 0, + "content": "\\[\nr _ {t} = r _ {t} ^ {b} * r _ {t} ^ {o} * r _ {t} ^ {r e l} * r _ {t} ^ {r e g} * r _ {t} ^ {c g}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.451, + 0.484, + 0.497 + ], + "angle": 0, + "content": "Next, we will discuss the methodology involved in crafting the CG reward \\( r_t^{cg} \\). Details about the kinematic imitation reward \\( r_t^b * r_t^o * r_t^{rel} * r_t^{reg} \\) are in the appendix." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.505, + 0.245, + 0.519 + ], + "angle": 0, + "content": "3.4.1. Contact Graph" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.523, + 0.484, + 0.599 + ], + "angle": 0, + "content": "We observe that existing imitation rewards [33, 34, 52] only measure kinematic properties and are insufficient to measure precise contacts. To tackle this problem, we propose the Contact Graph (CG) to model the contact in general interactions and design a CG imitation reward." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.599, + 0.484, + 0.78 + ], + "angle": 0, + "content": "As illustrated in Fig. 4, the CG is a complete graph where every pair of distinct nodes is connected by a unique edge, defined as \\(\\mathcal{G} = \\{\\mathcal{V},\\mathcal{E}\\}\\), where \\(\\mathcal{V}\\in \\{0,1\\} ^k\\) is the set of \\(k\\) nodes and \\(\\mathcal{E}\\in \\{0,1\\}^{k(k - 1) / 2}\\) is the set of edges. A CG node stores a binary value indicating whether it contacts other nodes. Each CG edge stores a binary label that denotes the contact between two nodes, where 1 represents contact, and 0 means no contact. CG node and edge values are calculated frame by frame, explicitly describing the mutual contact relationship between CG nodes at different moments. Contacts inside a node are not considered. We use the edge set \\(\\mathcal{E}\\) for imitation learning." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.484, + 0.901 + ], + "angle": 0, + "content": "The definitions of nodes are flexible. For example, a node can be a single part (e.g., a fingertip), or an aggregation of multiple parts (e.g., the whole left hand). Hinged objects can also be broken up into finer nodes [9, 13, 51]. The definition of CG is unified for a certain scene and shared between diverse interaction skills. For example, in basketball scenarios, we can aggregate two hands as a node, the rest of the bodies as a node, and the ball as a node, as illustrated in" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.091, + 0.568, + 0.15 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.535, + 0.15, + 0.565, + 0.16 + ], + "angle": 0, + "content": "(a) Ref" + }, + { + "type": "image", + "bbox": [ + 0.587, + 0.092, + 0.629, + 0.15 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.588, + 0.15, + 0.64, + 0.16 + ], + "angle": 0, + "content": "(b) w/o CGR" + }, + { + "type": "image", + "bbox": [ + 0.652, + 0.092, + 0.697, + 0.15 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.657, + 0.15, + 0.702, + 0.159 + ], + "angle": 0, + "content": "(c) w/ CGR" + }, + { + "type": "image", + "bbox": [ + 0.721, + 0.092, + 0.756, + 0.15 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.723, + 0.15, + 0.752, + 0.159 + ], + "angle": 0, + "content": "(d) Ref" + }, + { + "type": "image", + "bbox": [ + 0.793, + 0.092, + 0.829, + 0.15 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.781, + 0.15, + 0.832, + 0.159 + ], + "angle": 0, + "content": "(e) w/o CGR" + }, + { + "type": "image", + "bbox": [ + 0.855, + 0.092, + 0.89, + 0.15 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.853, + 0.15, + 0.897, + 0.159 + ], + "angle": 0, + "content": "(f) w/ CGR" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.167, + 0.573, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.535, + 0.226, + 0.565, + 0.235 + ], + "angle": 0, + "content": "(g) Ref" + }, + { + "type": "image", + "bbox": [ + 0.584, + 0.167, + 0.638, + 0.225 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.59, + 0.226, + 0.641, + 0.235 + ], + "angle": 0, + "content": "(h) w/o CGR" + }, + { + "type": "image", + "bbox": [ + 0.645, + 0.168, + 0.702, + 0.225 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.657, + 0.226, + 0.701, + 0.235 + ], + "angle": 0, + "content": "(i) w/ CGR" + }, + { + "type": "image", + "bbox": [ + 0.727, + 0.165, + 0.76, + 0.225 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.728, + 0.226, + 0.753, + 0.235 + ], + "angle": 0, + "content": "(j) Ref." + }, + { + "type": "image", + "bbox": [ + 0.798, + 0.167, + 0.831, + 0.225 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.766, + 0.226, + 0.852, + 0.235 + ], + "angle": 0, + "content": "(k) w/o CGR on table" + }, + { + "type": "image", + "bbox": [ + 0.863, + 0.166, + 0.896, + 0.225 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.861, + 0.226, + 0.904, + 0.234 + ], + "angle": 0, + "content": "(I) w/CGR" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.245, + 0.907, + 0.329 + ], + "angle": 0, + "content": "Figure 5. The HOI imitation falls into kinematic local-optimal solutions without Contact Graph Reward (CGR): (b) use the head to help control the ball; (e) use the wrist to contact the ball; (h) fail to catch the object; (k) support the table to keep balance. In comparison, the guidance of CGR effectively yields precise interactions, as shown in (c, f, i, l)." + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.355, + 0.907, + 0.387 + ], + "angle": 0, + "content": "Fig. 4. This simple CG is effective enough for learning all the basketball skills covered in BallPlay-M." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.393, + 0.726, + 0.407 + ], + "angle": 0, + "content": "3.4.2. Contact Graph Reward" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.411, + 0.907, + 0.653 + ], + "angle": 0, + "content": "Kinematic imitation rewards [32, 33, 52] fall short in measuring contacts and thus yield poor performance in HOI imitation. For example, during the toss skill training (Fig. 5 (a)), the contact between the right hand and the ball usually cause the ball to drop, making the kinematic rewards drastically smaller. In this case, the policy may learn a local-optimal solution that uses the head and left hands to stabilize the ball, as demonstrated in Fig. 5 (b). To tackle this problem, we propose the Contact Graph Reward (CGR) as a critical complement of kinematic imitation rewards to learn precise contact imitation. The CG error is defined as \\( e_t^{cg} = |s_t^{cg} - \\hat{s}_t^{cg}| \\), where \\( |\\cdot| \\) calculates element-wise absolute value, \\( s_t^{cg} \\) and \\( \\hat{s}_t^{cg} \\) denotes the simulated and reference CG state respectively, which is the edge set \\( \\mathcal{E}_t \\in \\{0,1\\}^{k(k-1)/2} \\). The CGR is measured by CG error, with independent weights on different elements:" + }, + { + "type": "equation", + "bbox": [ + 0.596, + 0.664, + 0.906, + 0.707 + ], + "angle": 0, + "content": "\\[\nr _ {t} ^ {c g} = \\exp \\left(- \\sum_ {j = 1} ^ {J} \\boldsymbol {\\lambda} ^ {c g} [ j ] * \\boldsymbol {e} _ {t} ^ {c g} [ j ]\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.717, + 0.906, + 0.81 + ], + "angle": 0, + "content": "where \\( e_t^{cg}[j] \\) is the \\( j \\)th element of \\( e_t^{cg} \\in \\{0,1\\}^J \\), representing a contact error; \\( \\lambda^{cg}[j] \\) is the \\( j \\)th element of \\( \\lambda^{cg} \\in \\mathbb{R}^J \\), a hyperparameter controls the sensitivity of a contact. Fig. 5 and Tab. 1 justify the effectiveness of CGR, where experiments without CGR show inaccurate interactions while using CGR effectively eliminates such problems." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.818, + 0.906, + 0.835 + ], + "angle": 0, + "content": "3.5. Reusing Interaction Skills for High-level Tasks" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.906, + 0.902 + ], + "angle": 0, + "content": "In this section, we propose training a High-Level Controller (HLC) to reuse the interaction skills acquired through SkillMimic to accomplish high-level tasks. As depicted in Fig. 3 (c), the HLC takes as input the current HOI state \\( s_t \\)" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17544" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.102, + 0.089, + 0.262, + 0.158 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.156, + 0.161, + 0.21, + 0.173 + ], + "angle": 0, + "content": "(a) Layup" + }, + { + "type": "image", + "bbox": [ + 0.263, + 0.089, + 0.419, + 0.158 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.281, + 0.161, + 0.4, + 0.173 + ], + "angle": 0, + "content": "(b) Turnaround Layup" + }, + { + "type": "image", + "bbox": [ + 0.42, + 0.089, + 0.577, + 0.159 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.454, + 0.161, + 0.545, + 0.172 + ], + "angle": 0, + "content": "(c) Dribble Right" + }, + { + "type": "image", + "bbox": [ + 0.579, + 0.089, + 0.736, + 0.159 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.604, + 0.161, + 0.712, + 0.172 + ], + "angle": 0, + "content": "(d) Dribble Forward" + }, + { + "type": "image", + "bbox": [ + 0.736, + 0.089, + 0.896, + 0.159 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.772, + 0.161, + 0.86, + 0.172 + ], + "angle": 0, + "content": "(e) Dribble Turn" + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.174, + 0.262, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.145, + 0.245, + 0.22, + 0.257 + ], + "angle": 0, + "content": "(f) Jump Shot" + }, + { + "type": "image", + "bbox": [ + 0.263, + 0.174, + 0.419, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.307, + 0.245, + 0.375, + 0.258 + ], + "angle": 0, + "content": "(g) Rebound" + }, + { + "type": "image", + "bbox": [ + 0.42, + 0.174, + 0.577, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.477, + 0.245, + 0.521, + 0.257 + ], + "angle": 0, + "content": "(h) Pass" + }, + { + "type": "image", + "bbox": [ + 0.579, + 0.174, + 0.735, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.631, + 0.245, + 0.686, + 0.258 + ], + "angle": 0, + "content": "(i) Pickup" + }, + { + "type": "image", + "bbox": [ + 0.737, + 0.174, + 0.896, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.78, + 0.245, + 0.852, + 0.258 + ], + "angle": 0, + "content": "(j) Cross Leg" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.269, + 0.908, + 0.311 + ], + "angle": 0, + "content": "Figure 6. Simulated humans exhibit comprehensive basketball skills. SkillMimic can teach humans a wide range of basketball skills using the same configuration in a purely data-driven manner, covering almost all fundamental basketball skills. Keyframes are placed in chronological order from left to right." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.338, + 0.482, + 0.384 + ], + "angle": 0, + "content": "and task observation \\( h_t \\), and outputs a discrete skill embedding \\( c_t \\) that serves as inputs to a pre-trained SkillMimic policy, which subsequently generates actions for the humanoid." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.384, + 0.483, + 0.519 + ], + "angle": 0, + "content": "Compared to previous hierarchical approaches that are based on locomotion priors [34, 50], our method leverages interaction skills as prior, which significantly simplifies the design of task rewards and greatly accelerates training. Moreover, unlike their mandatory hierarchical structure, our approach is motivated by data efficiency - tasks can be alternatively solved through direct HOI imitation when sufficient demonstrations are available, as shown in pickup experiments in Sec. 4.2, Fig. 7, and Fig. 8." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.52, + 0.425, + 0.536 + ], + "angle": 0, + "content": "We present four representative high-level tasks:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.54, + 0.483, + 0.569 + ], + "angle": 0, + "content": "- Throwing: throw the ball to approach a certain height, grab the rebound, and keep on throwing the ball." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.57, + 0.483, + 0.584 + ], + "angle": 0, + "content": "- Heading: dribble the ball to approach the target position." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.584, + 0.483, + 0.614 + ], + "angle": 0, + "content": "- Circling: dribble the ball around the target position following a target radius." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.615, + 0.483, + 0.645 + ], + "angle": 0, + "content": "- Scoring: dribble the ball toward the basket, timing the layup to score, retrieving the rebound, and repeating." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.54, + 0.483, + 0.645 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.648, + 0.484, + 0.709 + ], + "angle": 0, + "content": "Utilizing interaction skills as prior significantly improves learning efficiency for complex, long-horizon tasks. For example, a decent HLC for the Scoring task only requires 3 hours of training on an NVIDIA RTX 4090 GPU." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.724, + 0.224, + 0.74 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.902 + ], + "angle": 0, + "content": "We conduct extensive experiments to evaluate the performance of SkillMimic in learning interaction skills and its application for high-level tasks. Our evaluation consists of two main parts: (1) assessing the capability of SkillMimic in learning interaction skills, including comparisons with existing methods and ablation studies, and (2) demonstrating the effectiveness of reusing acquired interaction skills to learn high-level tasks. Details of the experimental settings and additional experiments on skill switching and robustness are provided in the appendix." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.337, + 0.61, + 0.351 + ], + "angle": 0, + "content": "4.1. Metrics" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.361, + 0.827, + 0.376 + ], + "angle": 0, + "content": "We consider the following quantitative metrics:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.378, + 0.906, + 0.483 + ], + "angle": 0, + "content": "- Accuracy. The overall accuracy of HOI imitation, abbreviated as Acc., is defined per frame, and deems imitation accurate when the object position and body position errors are both under the thresholds and the contacts are correct. The object threshold is defined as \\(0.2m\\). The body threshold is defined as \\(0.1m\\). The Acc. is calculated by averaging the values of all frames." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.483, + 0.906, + 0.543 + ], + "angle": 0, + "content": "- Position Error. The Mean Per-Joint Position Error (MPJPE) of the body \\((E_{\\mathrm{b - mpjpe}})\\) and object \\((E_{\\mathrm{o - mpjpe}})\\) is used to evaluate the positional imitation performance (in mm) following [27]." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.544, + 0.905, + 0.59 + ], + "angle": 0, + "content": "- Contact Error. The contact error \\( E_{\\mathrm{cg}} \\), ranging from 0 to 1, is defined as \\( \\frac{1}{N}\\sum_{t=1}^{N}\\mathrm{MSE}(s_t^{cg}, \\hat{s}_t^{cg}) \\), where \\( N \\) is the total frames of the reference HOI data." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.592, + 0.905, + 0.637 + ], + "angle": 0, + "content": "- Success Rate. The execution success rate of an interaction skill. Detailed definitions of success criteria for different skills are provided in the appendix." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.378, + 0.906, + 0.637 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.651, + 0.839, + 0.667 + ], + "angle": 0, + "content": "4.2. Evaluating Interaction Skill Learning" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Since our approach is the first to learn interaction skills from HOI demonstrations, there are no direct benchmarks for comparison. Therefore, we adapt reward strategies commonly used in locomotion imitation, i.e., DeepMimic [32] and AMP [33], to an object-inclusive setting. Specifically, we adapt our reward to the styles of DeepMimic [32] and AMP [33] while keeping the other components unchanged for a fair comparison. We denote these variant versions as DeepMimic* and AMP*. In Tab. 2, we present the success rate on 4 typical basketball skills from BallPlay-M, where our method achieves state-of-the-art performance in learning basketball skills compared with variants using DeepMimic or AMP style rewards. We guess that AMP-like rewards are ineffective for learning precise interactions due to their coarse-grained supervision signals." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17545" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.088, + 0.905, + 0.18 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.19, + 0.908, + 0.247 + ], + "angle": 0, + "content": "Figure 7. Demonstration of the pickup skill learned from 40 HOI motion clips. Yellow denotes the initial frame. Left: The humanoid picks up a stationary ball effortlessly. Middle: The humanoid intercepts a ball with random velocity. Right: The humanoid adjusts after missing the ball initially (the frame in blue) and successfully retrieves it on the second attempt, showcasing the potential for learning robust and generalizable skills through extensive data collection." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.26, + 0.908, + 0.328 + ], + "angle": 0, + "content": "
SkillMimic w/o MultiplicationSkillMimic w/o CGRSkillMimic
DatasetAcc. ↑Eb-mpipe ↓Eo-mpipe ↓Ecg ↓Acc. ↑Eb-mpipe ↓Eo-mpipe ↓Ecg ↓Acc. ↑Eb-mpipe ↓Eo-mpipe ↓Ecg ↓
GRAB [40]27.0%44.7180.20.72438.6%91.0180.10.33795.4%71.178.00.026
BallPlay-V7.5%40.11662.50.30613.6%88.5155.30.41282.4%56.882.90.087
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.34, + 0.908, + 0.369 + ], + "angle": 0, + "content": "Table 1. Ablation study on imitation reward design. We experiment on two HOI datasets: GRAB and BallPlay-V. The result highlights the importance of the CG reward and multiplication in achieving precise HOI imitation." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.392, + 0.484, + 0.475 + ], + "angle": 0, + "content": "
MethodSuccess Rate ↑
Pick UpDribble ForwardLayupShot
DeepMimic*[32]19.6%68.5%98.9%97.8%
AMP*[33]0.0%13.6%0.0%1.6%
SkillMimic (ours)86.7%79.6%99.1%97.9%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.487, + 0.484, + 0.53 + ], + "angle": 0, + "content": "Table 2. Success rates across four typical basketball skills in BallPlay-M. Our method significantly outperforms variant methods that use imitation reward styles of DeepMimic and AMP." + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.555, + 0.19, + 0.63 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.113, + 0.632, + 0.172, + 0.644 + ], + "angle": 0, + "content": "(a) \\(1 - 0.5\\%\\)" + }, + { + "type": "image", + "bbox": [ + 0.193, + 0.556, + 0.287, + 0.63 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.204, + 0.632, + 0.274, + 0.644 + ], + "angle": 0, + "content": "(b) \\(10 - 12.2\\%\\)" + }, + { + "type": "image", + "bbox": [ + 0.29, + 0.556, + 0.384, + 0.63 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.301, + 0.632, + 0.371, + 0.643 + ], + "angle": 0, + "content": "(c) \\(40 - 76.5\\%\\)" + }, + { + "type": "image", + "bbox": [ + 0.385, + 0.556, + 0.48, + 0.63 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.394, + 0.632, + 0.472, + 0.643 + ], + "angle": 0, + "content": "(d) \\(131 - 85.6\\%\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.657, + 0.484, + 0.727 + ], + "angle": 0, + "content": "Figure 8. Pickup generalization performance with different training data scales. First number: clips; second: success rate. In each test, 1000 balls are randomly placed within 1 to 5 meters away from the center. Yellow dots indicate successful pickups and green dots represent failures. See Sec. 4.2 for details." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.484, + 0.903 + ], + "angle": 0, + "content": "In Tab. 1, we do ablation on reward design, presenting detailed metrics of HOI imitation across two diverse datasets, GRAB [40] and BallPlay-V. The results indicate that the absence of reward multiplication leads to imbalanced sub-reward learning, e.g., favoring body motion \\((E_{\\mathrm{b - m}})\\) over interactions. The absence of CGR leads to kinematic local optima. In contrast, our complete setting (SkillMimic) yields significantly higher performance on interaction-related metrics (Acc., \\(E_{\\mathrm{o - m}}\\) , and \\(E_{\\mathrm{cg}}\\)" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.395, + 0.811, + 0.41 + ], + "angle": 0, + "content": "4.2.1. Ablation on Contact Graph Reward" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.415, + 0.907, + 0.673 + ], + "angle": 0, + "content": "We conduct ablation experiments on diverse datasets to comprehensively evaluate the impact of Contact Graph Reward (CGR) on interaction skill learning. Fig. 5 illustrates the qualitative results of ablating CGR on the GRAB and BallPlay-V datasets. It can be clearly observed that without CGR, the humanoid often resorts to incorrect contacts to achieve more stable object control, such as using its head to push the ball, clamping the ball with its hands and legs, or using its hands to support itself on a table for balance. These examples illustrate typical kinematic local optima. By incorporating CGR, these local optima are effectively eliminated, resulting in accurate interactions. Additionally, the result in Tab. 1 also shows that, without CGR, the contact error \\((E_{\\mathrm{cg}})\\) significantly increases, leading to kinematic local optima and causes a noticeable decline in overall imitation accuracy. These results underscore the critical importance of CGR for learning correct interactions." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.684, + 0.866, + 0.698 + ], + "angle": 0, + "content": "4.2.2. Data Scale and Generalization Performance" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.704, + 0.907, + 0.795 + ], + "angle": 0, + "content": "To examine how the generalization performance grows as the data scale increases, we conduct ablation experiments on the pickup skill. We create four datasets by randomly selecting pickup clips from BallPlay-M, containing 1, 10, 40, and 131 clips, respectively. We train policy on each dataset with approximately 3.2 billion samples using SkillMimic." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.909, + 0.901 + ], + "angle": 0, + "content": "To evaluate the generalization performance, we randomly place balls within a circular area ranging from 1 to 5 meters around the humanoid—this distribution significantly exceeds the coverage of the reference data. Fig. 7 and Fig. 8 show the results. We can see that as the data scale increases, pickup performance improves significantly due to broader state transition coverage in larger datasets. This indicates" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17546" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.098, + 0.086, + 0.297, + 0.17 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.146, + 0.172, + 0.25, + 0.184 + ], + "angle": 0, + "content": "(a) Manual Control" + }, + { + "type": "image", + "bbox": [ + 0.298, + 0.098, + 0.498, + 0.17 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.368, + 0.172, + 0.43, + 0.184 + ], + "angle": 0, + "content": "(b) Scoring" + }, + { + "type": "image", + "bbox": [ + 0.5, + 0.082, + 0.699, + 0.17 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.567, + 0.172, + 0.632, + 0.184 + ], + "angle": 0, + "content": "(c) Heading" + }, + { + "type": "image", + "bbox": [ + 0.701, + 0.082, + 0.902, + 0.17 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.769, + 0.172, + 0.832, + 0.184 + ], + "angle": 0, + "content": "(d) Circling" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.196, + 0.907, + 0.253 + ], + "angle": 0, + "content": "Figure 9. Our method supports training a single IS policy under a unified configuration to acquire various interaction skills. These interaction skills can be flexibly switched, as illustrated in (a), where yellow denotes shot, blue denotes pickup, and green denotes turnaround layup. Complex, long-horizon tasks can be easily achieved by training a high-level controller (HLC) to manage switching of the learned interaction skills: (b) scoring from random positions (c) dribbling to target locations, and (d) dribbling along an expanding radius." + }, + { + "type": "table", + "bbox": [ + 0.102, + 0.269, + 0.475, + 0.364 + ], + "angle": 0, + "content": "
MethodSuccess Rate↑
HeadingCirclingThrowingScoring
PPO[36]0.70%11.14%0.00%0.00%
ASE[34]0.19%4.37%0.00%0.00%
ASE*[34]0.31%7.21%0.00%0.00%
SkillMimic (ours)93.04%79.92%93.40%80.25%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.377, + 0.483, + 0.448 + ], + "angle": 0, + "content": "Table 3. Success rates on 4 high-level basketball tasks. Both PPO (learn from scratch), ASE (using locomotion prior), and \\(\\mathrm{ASE^{*}}\\) (using body and object motion prior) fail to converge on all tasks. In contrast, leveraging interaction skill priors acquired through SkillMimic, our method effectively learns these difficult tasks." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.475, + 0.483, + 0.52 + ], + "angle": 0, + "content": "our method's strong scalability and potential for large-scale learning. It is worth noting that we do not incorporate any additional designs for robustness and generalization." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.531, + 0.342, + 0.547 + ], + "angle": 0, + "content": "4.3. Evaluating High-level Tasks" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.553, + 0.483, + 0.825 + ], + "angle": 0, + "content": "We conduct experiments on four high-level tasks described in Sec. 3.5. We employ a single IS policy as the interaction skill prior, trained using SkillMimic for around 4.5 billion samples across seven skills: pickup, layup, turnaround layup, dribble left, dribble right, dribble forward, and shot. During training, the IS policy is fixed, and only the HLC is trained. For tasks such as throwing, heading, and circling, we train for around 0.4 billion samples. For the more challenging scoring task, we train for around 1.2 billion samples. To objectively evaluate the performance of our method, we compared it against three baseline methods: (1) learning from scratch using PPO [36], (2) learning with body motion priors using ASE [34], and (3) learning with body and object motion priors using ASE (denoted as \\(\\mathrm{ASE^{*}}\\)). The low-level controllers of ASE and \\(\\mathrm{ASE^{*}}\\) are trained using the same training steps as our IS policy. For task training, all methods are trained using identical task rewards and simulation steps for fair comparisons." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Fig. 9 (a) showcases an example of manually controlling the pre-trained IS policy to perform skill switching. Fig. 9 (b) shows several trajectories of the HLC performing skill controls to accurately deliver the ball to the target basket. Fig. 9 (c) & (d) present a top-down view of the HLC control" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.272, + 0.907, + 0.452 + ], + "angle": 0, + "content": "ling the humanoid to dribble the ball toward target locations and along an expanding radius. As shown in Tab. 3, PPO and ASE yield low success rates because the tasks are too challenging to master without interaction priors. The failure of \\(\\mathrm{ASE^{*}}\\) can be traced back to its low-level controller's inadequate ability to imitate interaction skills. This is primarily because adversarial rewards fail to provide effective guidance for precise interaction learning, which is similar to the failure mode observed in \\(\\mathrm{AMP^{*}}\\). In contrast, by utilizing the interaction skills learned through SkillMimic, we greatly improve the performance of learning complex high-level interaction tasks." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.467, + 0.779, + 0.483 + ], + "angle": 0, + "content": "5. Discussion and Future Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.493, + 0.906, + 0.613 + ], + "angle": 0, + "content": "We presented a novel framework for learning diverse interaction skills with a unified HOI imitation reward using reinforcement learning. Our approach demonstrates multi-skill mastery, skill reusability, and clear data scaling benefits, highlighting the significant potential of HOI imitation for learning complex interaction skills. However, the full application of this paradigm requires further exploration. Several key areas deserve particular attention:" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.614, + 0.906, + 0.719 + ], + "angle": 0, + "content": "Firstly, exploring more general HOI skill learning within a single policy, such as household tasks or other types of sports, presents challenges in handling generalization across different object geometries. Moreover, perceiving multiple objects simultaneously requires a more general environmental perception approach, rather than directly obtaining privileged information for each object." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.905, + 0.765 + ], + "angle": 0, + "content": "Secondly, incorporating finer-grained control conditions, such as tracking [27], could enable more flexible control and generalization capabilities as HOI data scales up." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.905, + 0.826 + ], + "angle": 0, + "content": "Thirdly, the difficulty of collecting HOI data highlights the importance of exploring methods to learn robust interaction skills with limited HOI data, as well as investigating HOI data augmentation and generation techniques." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.905, + 0.9 + ], + "angle": 0, + "content": "Lastly, applying this technology to real robots would be highly valuable, such as enabling real-world humanoid robots to play basketball games. This endeavor could involve various challenges, including HOI data retargeting and sim-to-real transfer issues." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17547" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.241, + 0.108 + ], + "angle": 0, + "content": "Acknowledgment" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.484, + 0.206 + ], + "angle": 0, + "content": "We would like to thank Mingyang Su for his valuable assistance in motion capture data collection and Shunlin Lu for helpful discussions. This work was supported in part by Unitree Robotics, International Digital Economy Academy (IDEA), and the Innovation and Technology Fund of HK-SAR (GHX/054/21GD)." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.218, + 0.188, + 0.234 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.244, + 0.484, + 0.298 + ], + "angle": 0, + "content": "[1] Kai Arulkumaran, Marc Peter Deisenroth, Miles Brundage, and Anil Anthony Bharath. Deep reinforcement learning: A brief survey. IEEE Signal Processing Magazine, 34(6):26-38, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.3, + 0.484, + 0.368 + ], + "angle": 0, + "content": "[2] Jinseok Bae, Jungdam Won, Donggeun Lim, Cheol-Hui Min, and Young Min Kim. Pmp: Learning to physically interact with environments using part-wise motion priors. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-10, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.37, + 0.483, + 0.438 + ], + "angle": 0, + "content": "[3] Kevin Black, Noah Brown, Danny Driess, Adnan Esmail, Michael Equi, Chelsea Finn, Niccolo Fusai, Lachy Groom, Karol Hausman, Brian Ichter, et al. Pi0: A vision-language-action flow model for general robot control. arXiv preprint arXiv:2410.24164, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.44, + 0.483, + 0.494 + ], + "angle": 0, + "content": "[4] Jona Braun, Sammy Christen, Muhammed Kocabas, Emre Aksan, and Otmar Hilliges. Physically plausible full-body hand-object interaction synthesis. arXiv preprint arXiv:2309.07907, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.496, + 0.484, + 0.564 + ], + "angle": 0, + "content": "[5] Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Xi Chen, Krzysztof Choromanski, Tianli Ding, Danny Driess, Avinava Dubey, Chelsea Finn, et al. Rt-2: Vision-language-action models transfer web knowledge to robotic control. arXiv preprint arXiv:2307.15818, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.565, + 0.482, + 0.606 + ], + "angle": 0, + "content": "[6] Stelian Coros, Philippe Beaudoin, and Michiel Van de Panne. Generalized biped walking control. ACM Transactions On Graphics (TOG), 29(4):1-9, 2010. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.608, + 0.482, + 0.663 + ], + "angle": 0, + "content": "[7] Jieming Cui, Tengyu Liu, Nian Liu, Yaodong Yang, Yixin Zhu, and Siyuan Huang. Anyskill: Learning openvocabulary physical skill for interactive agents. In Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.664, + 0.482, + 0.732 + ], + "angle": 0, + "content": "[8] Zhiyang Dou, Xuelin Chen, Qingnan Fan, Taku Komura, and Wenping Wang. C-ase: Learning conditional adversarial skill embeddings for physics-based characters. In SIGGRAPH Asia 2023 Conference Papers, pages 1-11, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.734, + 0.483, + 0.815 + ], + "angle": 0, + "content": "[9] Zicong Fan, Omid Taheri, Dimitrios Tzionas, Muhammed Kocabas, Manuel Kaufmann, Michael J Black, and Otmar Hilliges. Arctic: A dataset for dexterous bimanual handobject manipulation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12943-12954, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.817, + 0.483, + 0.873 + ], + "angle": 0, + "content": "[10] Zipeng Fu, Qingqing Zhao, Qi Wu, Gordon Wetzstein, and Chelsea Finn. Humanplus: Humanoid shadowing and imitation from humans. In 8th Annual Conference on Robot Learning, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.874, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[11] Zipeng Fu, Tony Z Zhao, and Chelsea Finn. Mobile aloha: Learning bimanual mobile manipulation with low-cost" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.244, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.548, + 0.093, + 0.905, + 0.119 + ], + "angle": 0, + "content": "whole-body teleoperation. arXiv preprint arXiv:2401.02117, 2024.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.123, + 0.905, + 0.192 + ], + "angle": 0, + "content": "[12] Jiawei Gao, Ziqin Wang, Zeqi Xiao, Jingbo Wang, Tai Wang, Jinkun Cao, Xiaolin Hu, Si Liu, Jifeng Dai, and Jiangmiao Pang. Coohoi: Learning cooperative human-object interaction with manipulated object dynamics. Advances in Neural Information Processing Systems, 37:79741-79763, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.194, + 0.907, + 0.277 + ], + "angle": 0, + "content": "[13] Haoran Geng, Helin Xu, Chengyang Zhao, Chao Xu, Li Yi, Siyuan Huang, and He Wang. Gapartner: Cross-category domain-generalizable object perception and manipulation via generalizable and actionable parts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7081-7091, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.28, + 0.905, + 0.336 + ], + "angle": 0, + "content": "[14] Anindita Ghosh, Rishabh Dabral, Vladislav Golyanik, Christian Theobalt, and Philipp Slusallek. Imos: Intent-driven full-body motion synthesis for human-object interactions. In Eurographics, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.338, + 0.905, + 0.394 + ], + "angle": 0, + "content": "[15] Mohamed Hassan, Yunrong Guo, Tingwu Wang, Michael Black, Sanja Fidler, and Xue Bin Peng. Synthesizing physical character-scene interactions. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-9, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.396, + 0.905, + 0.466 + ], + "angle": 0, + "content": "[16] Tairan He, Zhengyi Luo, Xialin He, Wenli Xiao, Chong Zhang, Weinan Zhang, Kris M Kitani, Changliu Liu, and Guanya Shi. Omnih2o: Universal and dexterous human-to-humanoid whole-body teleoperation and learning. In 8th Annual Conference on Robot Learning, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.468, + 0.905, + 0.51 + ], + "angle": 0, + "content": "[17] Edmond SL Ho, Taku Komura, and Chiew-Lan Tai. Spatial relationship preserving character motion adaptation. In ACM SIGGRAPH 2010 papers, pages 1-8, 2010. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.512, + 0.905, + 0.553 + ], + "angle": 0, + "content": "[18] Jonathan Ho and Stefano Ermon. Generative adversarial imitation learning. Advances in neural information processing systems, 29, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.556, + 0.905, + 0.612 + ], + "angle": 0, + "content": "[19] Jessica K Hodgins, Wayne L Wooten, David C Brogan, and James F O'Brien. Animating human athletics. In Proceedings of the 22nd annual conference on Computer graphics and interactive techniques, pages 71-78, 1995. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.614, + 0.905, + 0.683 + ], + "angle": 0, + "content": "[20] Nan Jiang, Tengyu Liu, Zhexuan Cao, Jieming Cui, Zhiyuan Zhang, Yixin Chen, He Wang, Yixin Zhu, and Siyuan Huang. Full-body articulated human-object interaction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9365-9376, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.686, + 0.905, + 0.741 + ], + "angle": 0, + "content": "[21] Nan Jiang, Zhiyuan Zhang, Hongjie Li, Xiaoxuan Ma, Zan Wang, Yixin Chen, Tengyu Liu, Yixin Zhu, and Siyuan Huang. Scaling up dynamic human-scene interaction modeling. arXiv preprint arXiv:2403.08629, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.743, + 0.905, + 0.798 + ], + "angle": 0, + "content": "[22] Jordan Juravsky, Yunrong Guo, Sanja Fidler, and Xue Bin Peng. Padl: Language-directed physics-based character control. In SIGGRAPH Asia 2022 Conference Papers, pages 1-9, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.801, + 0.905, + 0.856 + ], + "angle": 0, + "content": "[23] Jiaman Li, Alexander Clegg, Roozbeh Mottaghi, Jiajun Wu, Xavier Puig, and C Karen Liu. Controllable human-object interaction synthesis. arXiv preprint arXiv:2312.03913, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.859, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[24] Jiaman Li, Jiajun Wu, and C Karen Liu. Object motion guided human motion synthesis. ACM Transactions on Graphics (TOG), 42(6):1-11, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "17548" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.484, + 0.147 + ], + "angle": 0, + "content": "[25] Libin Liu and Jessica Hodgins. Learning to schedule control fragments for physics-based characters using deep q-learning. ACM Transactions on Graphics (TOG), 36(3):1-14, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.484, + 0.203 + ], + "angle": 0, + "content": "[26] Libin Liu and Jessica Hodgins. Learning basketball dribbling skills using trajectory optimization and deep reinforcement learning. ACM Transactions on Graphics (TOG), 37(4):1-14, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.205, + 0.484, + 0.259 + ], + "angle": 0, + "content": "[27] Zhengyi Luo, Jinkun Cao, Kris Kitani, Weipeng Xu, et al. Perpetual humanoid control for real-time simulated avatars. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10895-10904, 2023. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.261, + 0.484, + 0.315 + ], + "angle": 0, + "content": "[28] Zhengyi Luo, Jinkun Cao, Sammy Christen, Alexander Winkler, Kris Kitani, and Weipeng Xu. Omnigrasp: Grasping diverse objects with simulated humanoids. Advances in Neural Information Processing Systems, 37:2161-2184, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.317, + 0.484, + 0.383 + ], + "angle": 0, + "content": "[29] Zhengyi Luo, Jiashun Wang, Kangni Liu, Haotian Zhang, Chen Tessler, Jingbo Wang, Ye Yuan, Jinkun Cao, Zihui Lin, Fengyi Wang, et al. Smplolympics: Sports environments for physically simulated humanoids. arXiv preprint arXiv:2407.00187, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.385, + 0.484, + 0.44 + ], + "angle": 0, + "content": "[30] Vinod Nair and Geoffrey E Hinton. Rectified linear units improve restricted boltzmann machines. In Proceedings of the 27th international conference on machine learning (ICML-10), pages 807–814, 2010. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.442, + 0.484, + 0.509 + ], + "angle": 0, + "content": "[31] Liang Pan, Jingbo Wang, Buzhen Huang, Junyu Zhang, Haofan Wang, Xu Tang, and Yangang Wang. Synthesizing physically plausible human motions in 3d scenes. In 2024 International Conference on 3D Vision (3DV), pages 1498-1507. IEEE, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.511, + 0.484, + 0.552 + ], + "angle": 0, + "content": "[32] Xue Bin Peng, Pieter Abbeel, Sergey Levine, and Michiel van de Panne. Deepmimic. ACM Transactions on Graphics, page 1-14, 2018. 1, 2, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.553, + 0.484, + 0.607 + ], + "angle": 0, + "content": "[33] Xue Bin Peng, Ze Ma, Pieter Abbeel, Sergey Levine, and Angjoo Kanazawa. Amp: Adversarial motion priors for stylized physics-based character control. ACM Transactions on Graphics, page 1-20, 2021. 2, 3, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.609, + 0.484, + 0.663 + ], + "angle": 0, + "content": "[34] Xue Bin Peng, Yunrong Guo, Lina Halper, Sergey Levine, and Sanja Fidler. Ase: Large-scale reusable adversarial skill embeddings for physically simulated characters. ACM Trans. Graph., 41(4), 2022. 1, 2, 3, 4, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.665, + 0.484, + 0.718 + ], + "angle": 0, + "content": "[35] Jiawei Ren, Mingyuan Zhang, Cunjun Yu, Xiao Ma, Liang Pan, and Ziwei Liu. Insactor: Instruction-driven physics-based characters. Advances in Neural Information Processing Systems, 36, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.72, + 0.484, + 0.761 + ], + "angle": 0, + "content": "[36] John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347, 2017. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.762, + 0.484, + 0.802 + ], + "angle": 0, + "content": "[37] Sebastian Starke, Yiwei Zhao, Taku Komura, and Kazi Zaman. Local motion phases for learning multi-contact character movements. ACM Trans. Graph., 39(4), 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.804, + 0.484, + 0.845 + ], + "angle": 0, + "content": "[38] Sebastian Starke, Yiwei Zhao, Fabio Zinno, and Taku Komura. Neural animation layering for synthesizing martial arts movements. ACM Trans. Graph., 40(4), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.484, + 0.9 + ], + "angle": 0, + "content": "[39] Jingkai Sun, Qiang Zhang, Yiqun Duan, Xiaoyang Jiang, Chong Cheng, and Renjing Xu. Prompt, plan, perform: Llm-based humanoid control via quantized imitation learning. arXiv preprint arXiv:2309.11359, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.147 + ], + "angle": 0, + "content": "[40] Omid Taheri, Nima Ghorbani, Michael J. Black, and Dimitrios Tzionas. GRAB: A dataset of whole-body human grasping of objects. In European Conference on Computer Vision (ECCV), 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.149, + 0.906, + 0.19 + ], + "angle": 0, + "content": "[41] Jie Tan, Yuting Gu, C Karen Liu, and Greg Turk. Learning bicycle stunts. ACM Transactions on Graphics (TOG), 33 (4):1-12, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.192, + 0.906, + 0.259 + ], + "angle": 0, + "content": "[42] Chen Tessler, Yoni Kasten, Yunrong Guo, Shie Mannor, Gal Chechik, and Xue Bin Peng. Calm: Conditional adversarial latent models for directable virtual characters. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-9, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.262, + 0.906, + 0.317 + ], + "angle": 0, + "content": "[43] Yinhuai Wang, Jing Lin, Ailing Zeng, Zhengyi Luo, Jian Zhang, and Lei Zhang. Physhoi: Physics-based imitation of dynamic human-object interaction. arXiv preprint arXiv:2312.04393, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.319, + 0.906, + 0.36 + ], + "angle": 0, + "content": "[44] Jungdam Won, Deepak Gopinath, and Jessica Hodgins. Physics-based character controllers using conditional vaes. ACM Transactions on Graphics (TOG), 41(4):1-12, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.362, + 0.906, + 0.416 + ], + "angle": 0, + "content": "[45] Yan Wu, Jiahao Wang, Yan Zhang, Siwei Zhang, Otmar Hilliges, Fisher Yu, and Siyu Tang. Saga: Stochastic whole-body grasping with contact. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.418, + 0.906, + 0.486 + ], + "angle": 0, + "content": "[46] Zeqi Xiao, Tai Wang, Jingbo Wang, Jinkun Cao, Wenwei Zhang, Bo Dai, Dahua Lin, and Jiangmiao Pang. Unified human-scene interaction via prompted chain-of-contacts. In The Twelfth International Conference on Learning Representations, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.489, + 0.906, + 0.558 + ], + "angle": 0, + "content": "[47] Sirui Xu, Zhengyuan Li, Yu-Xiong Wang, and Liang-Yan Gui. Interdiff: Generating 3d human-object interactions with physics-informed diffusion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14928-14940, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.56, + 0.906, + 0.601 + ], + "angle": 0, + "content": "[48] Sirui Xu, Ziyin Wang, Yu-Xiong Wang, and Liang-Yan Gui. Interdreamer: Zero-shot text to 3d dynamic human-object interaction. arXiv preprint arXiv:2403.19652, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.602, + 0.906, + 0.643 + ], + "angle": 0, + "content": "[49] KangKang Yin, Kevin Loken, and Michiel Van de Panne. Simbicon: Simple biped locomotion control. ACM Transactions on Graphics (TOG), 26(3):105-es, 2007. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.645, + 0.906, + 0.699 + ], + "angle": 0, + "content": "[50] Haotian Zhang, Ye Yuan, Viktor Makoviychuk, Yunrong Guo, Sanja Fidler, Xue Bin Peng, and Kayvon Fatahalian. Learning physically simulated tennis skills from broadcast videos. ACM Trans. Graph., 2023. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.702, + 0.906, + 0.77 + ], + "angle": 0, + "content": "[51] Hui Zhang, Sammy Christen, Zicong Fan, Luocheng Zheng, Jemin Hwangbo, Jie Song, and Otmar Hilliges. Artigrasp: Physically plausible synthesis of bi-manual dexterous grasping and articulation. In 2024 International Conference on 3D Vision (3DV), pages 235-246. IEEE, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.772, + 0.906, + 0.839 + ], + "angle": 0, + "content": "[52] Yunbo Zhang, Deepak Gopinath, Yuting Ye, Jessica Hodgins, Greg Turk, and Jungdam Won. Simulation and retargeting of complex multi-character interactions. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-11, 2023. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.842, + 0.906, + 0.884 + ], + "angle": 0, + "content": "[53] Qingxu Zhu, He Zhang, Mengting Lan, and Lei Han. Neural categorical priors for physics-based character control. ACM Transactions on Graphics (TOG), 42(6):1-16, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.884 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17549" + } + ] +] \ No newline at end of file diff --git a/2025/SkillMimic_ Learning Basketball Interaction Skills from Demonstrations/70ea78a5-d7df-4937-ac48-382497691b79_origin.pdf b/2025/SkillMimic_ Learning Basketball Interaction Skills from Demonstrations/70ea78a5-d7df-4937-ac48-382497691b79_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..29d94a742e2a4f4ab6081e335774a686e5a4b414 --- /dev/null +++ b/2025/SkillMimic_ Learning Basketball Interaction Skills from Demonstrations/70ea78a5-d7df-4937-ac48-382497691b79_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d471aa83ac9f080c0d6d327c836bb55e03e80dd4320b7fe3bac73f3a45c96f7a +size 8065949 diff --git a/2025/SkillMimic_ Learning Basketball Interaction Skills from Demonstrations/full.md b/2025/SkillMimic_ Learning Basketball Interaction Skills from Demonstrations/full.md new file mode 100644 index 0000000000000000000000000000000000000000..b1b06ba6b2e873f3d706a511b5ffcc6dc8289994 --- /dev/null +++ b/2025/SkillMimic_ Learning Basketball Interaction Skills from Demonstrations/full.md @@ -0,0 +1,371 @@ +# SkillMimic: Learning Basketball Interaction Skills from Demonstrations + +Yinhuai Wang $^{1,2*}$ Qihan Zhao $^{1*}$ Runyi Yu $^{1,2*}$ Hok Wai Tsui $^{1}$ Ailing Zeng $^{6\dagger}$ Jing Lin $^{5}$ Zhengyi Luo $^{7}$ Jiwen Yu $^{3}$ Xiu Li $^{4}$ Qifeng Chen $^{1}$ Jian Zhang $^{3\dagger}$ Lei Zhang $^{5}$ Ping Tan $^{1}$ + +$^{1}$ Hong Kong University of Science and Technology $^{2}$ Unitree Robotics $^{3}$ Peking University Shenzhen Graduate School $^{4}$ Tsinghua University $^{5}$ International Digital Economy Academy $^{6}$ Tencent $^{7}$ Carnegie Mellon University yinhuai.wang@connect.ust.hk pingtan@ust.hk + +![](images/f0da1c7935060e5d15ea20c148848d4ecae60676075df2db709bef83387d5ed3.jpg) +Figure 1. We propose a novel approach that for the first time enables physically simulated humanoids to learn a variety of basketball interaction skills from Human-Object Interaction (HOI) data, including but not limited to shooting (blue), retrieving (red), and turnaround layup (yellow). Once acquired, these interaction skills can be composed to accomplish complex tasks, such as consecutive scoring (green). + +# Abstract + +Traditional reinforcement learning methods for human-object interaction (HOI) rely on labor-intensive, manually designed skill rewards that do not generalize well across different interactions. We introduce SkillMimic, a unified data-driven framework that fundamentally changes how agents learn interaction skills by eliminating the need for skill-specific rewards. Our key insight is that a unified HOI imitation reward can effectively capture the essence of diverse interaction patterns from HOI datasets. This enables SkillMimic to learn a single policy that not only masters multiple interaction skills but also facilitates skill transitions, with both diversity and generalization improving as the HOI dataset grows. For evaluation, we collect and introduce two basketball datasets containing approximately 35 minutes of diverse basketball skills. Extensive experiments show that SkillMimic successfully masters a wide range of basketball skills including stylistic variations in dribbling, layup, and shooting. Moreover, these learned skills can be effectively composed by a high-level controller to accomplish complex and long-horizon tasks such as consecutive scoring, opening new possibilities for scalable and generalizable interaction skill learning. Project page: https://ingrid789.github.io/SkillMimic/ + +# 1. Introduction + +Human-Object Interaction (HOI) represents one of the most fundamental yet challenging capabilities in robotics and character animation [10, 12, 16, 23, 28, 46, 47]. Among various HOI scenarios, basketball serves as an exemplary testbed that demands mastery of complex, dynamic, and precise interactions. Despite significant progress, existing physics-based character animation methods predominantly focus on imitation learning of locomotion skills [8, 32–34, 42] through Reinforcement Learning (RL), leaving a significant gap for HOI imitation. While some methods can learn interaction skills, such as striking pillars [34], playing tennis [50], climbing ropes [2], and carrying boxes [15], they typically involve handcrafted reward functions for each different interaction skill. However, designing skill-specific rewards is labor-intensive and fundamentally limits generalization. This limitation becomes particularly acute in basketball, where the complexity and diversity of skills would demand extraordinary engineering effort for reward design. Consequently, existing methods [2, 8, 15, 32–34, 42, 50] fail to provide a unified learning framework for a single policy to learn diverse interaction skills, let alone achieving complex long-term tasks such as consecutive basketball scoring. + +How can a simulated humanoid learn a wide variety of interaction skills in a simple and scalable manner? Consider how humans naturally acquire complex skills: amateur basketball players often develop their expertise by watching + +and imitating game footage, even without formal coaching. This self-directed learning process involves synchronizing their body movements with both the observed player motions and ball trajectories through practice. + +Inspired by these observations, we propose a data-driven method called SkillMimic, which mimics HOI through RL to learn interaction skills. SkillMimic can learn various interaction skills purely from HOI datasets, such as diverse styles of basketball shooting, layup, and dribbling skills, and even a robust pickup skill that enables picking up balls in random locations and motions. Notably, SkillMimic uses the exact same configuration to learn different interaction skills, with identical hyperparameters. This allows us to train a single Interaction Skill (IS) policy to learn multiple interaction skills and achieve smooth skill switching. By training a high-level policy to reuse these learned interaction skills, we can accomplish challenging high-level tasks such as scoring, which requires the humanoid to composite diverse basketball interaction skills to score accurately. + +To advance basketball skill learning, we introduce two HOI datasets: BallPlay-V, which captures player and ball motion from RGB videos across eight basic skills, and BallPlay-M, which uses optical motion capture to record 35 minutes of comprehensive basketball interactions. Using these datasets, SkillMimic successfully learns diverse basketball skills through a unified IS policy. With an additional high-level controller for skill composition, our system can execute complex tasks like consecutive scoring. Our approach demonstrates superior efficiency in learning both interaction skills and multi-phase, long-horizon tasks compared to existing methods, opening new possibilities for scalable and generalizable interaction skill learning. + +Specifically, our contributions are as follows: + +- SkillMimic: a data-driven paradigm for learning diverse humanoid interaction skills through RL. It supports joint learning and switching of diverse basketball interaction skills, with skill diversity and generalization improving as the demonstration data grows, showcasing the scalability. +- Contact graph: We propose a simple and general contact modeling method that applies to diverse skills, called the contact graph. A contact graph reward is designed to enable precise contact imitation, which proves to be critical for learning precise interaction skills. +- Unified HOI imitation reward: We propose a set of important designs that form a unified reward configuration for imitation learning of various interaction skills. +- A hierarchical solution for reusing interaction skills: We propose training a high-level controller to reuse the interaction skills acquired by SkillMimic to accomplish challenging high-level tasks, which for the first time achieves continuous basketball scoring. +- BallPlay Datasets: We introduce two basketball datasets to facilitate research on interaction skill learning. + +# 2. Related Work + +Locomotion Imitation Learning. Mimicking human motion for robot control is an efficient method of learning humanoid skills. DeepMimic [32], a pioneer in this field, uses imitation learning to perform a variety of highly dynamic locomotion skills. The introduction of Generative Adversarial Imitation Learning (GAIL) [18] into humanoid imitation learning by AMP [33] lessens constraints on data alignment, thus enhancing its versatility. ASE [34] further amplifies motion diversity in GAIL training, incorporates a pre-trained low-level policy to acquire locomotion skills, and deploys a high-level policy to repurpose these locomotion skills for specific tasks, such as striking pillars. These methodologies are subsequently expanded in numerous efforts to augment condition and text control [8, 22, 29, 31, 35, 39, 42, 44, 53], as well as to design automated rewards [7]. However, a limitation of these studies is their exclusive focus on learning isolated locomotion skills through imitation. This requires additional pipeline and reward design for learning interaction skills, even for relatively simple ones, such as moving boxes [15] or striking pillars [34]. In contrast, our approach enables unified learning of diverse interaction skills, eliminating the need for skill-specific reward engineering. + +Learning Human-Object Interaction (HOI). Learning HOI in simulation has long posed challenges in the fields of graphics and robotics [14, 20, 21, 23, 24, 37, 38, 45, 47, 48]. Early studies often relied on manually designed control structures. For example, Hodgins et al. [19] implement a variety of sports interactions through the use of manually designed state machine controllers. Yin et al. [49] use finite-state machines to generate a large variety of gaits. Coros et al. [6] utilize an inverted pendulum model to control foot placement. Recently, deep reinforcement learning [1] has been widely employed in character control, giving rise to an array of methods based on policy network control. These methods have been applied to a range of challenging activities, such as basketball [26], skateboarding [25], and cycling [41], etc. The use of imitation rewards further lessens the reward design for locomotion styles. Zhang et al. [50] propose a framework to learn diverse tennis skills from broadcast videos. Bae et al. [2] incorporate multiple part-wise motion priors for diverse whole-body interactions. Similarly, Braun et al. [4] develop a framework for synthesizing whole-body grasps. + +Applying imitation learning for HOI is intuitive, however, effectively learning diverse interaction skills in a unified framework remains unresolved. Naively extending locomotion imitation methods [32, 33] to HOI imitation often yields unstable and unnatural results, as these methods do not address the unbalanced rewards inherent in HOI, nor do they properly account for the crucial aspects of relative mo + +![](images/6c3ee744f479b9227dff4936e632e46e21c71f76c366684addf939351f46f0aa.jpg) +Figure 2. Concept of SkillMimic. We define an interaction skill as a set of Human-Object Interaction (HOI) state transitions that align with the intended skill semantics. These state transitions can be derived from captured HOI motion clips. If a simulated humanoid can manipulate objects such that the resulting HOI state transitions closely match those of the reference, we consider the humanoid to have successfully learned the interaction skill. + +tion and contact in HOI. Modeling interaction through interaction graphs is a well-established approach [17]. While this method has proven useful for HOI imitation [52], it remains limited by considering only kinematic relationships, failing to ensure proper contact and becoming unstable in learning diverse scenarios. In contrast, our contact graph imitation reward explicitly prioritize physical contact learning, significantly improving the success rate of HOI imitation and enabling unified HOI imitation of diverse complex interaction skills for the first time. + +This work builds upon and significantly extends ideas initially explored in our early technical report [43]. + +# 3. SkillMimic + +In this section, we introduce a unified data-driven framework for simulated humanoids to learn various basketball interaction skills via simulation and reinforcement learning (RL). We begin by introducing how we define interaction skills with Human-Object Interaction (HOI) data (Sec. 3.2). Next, we explain how to learn diverse interaction skills by imitating HOI data (Sec. 3.3) and discuss the designing of a unified HOI imitation reward (Sec. 3.4). Finally, we demonstrate how the learned interaction skills can be reused to perform high-level tasks (Sec. 3.5). + +# 3.1. The BallPlay Dataset + +To address the scarcity of basketball HOI data and facilitate research on interaction skill learning, we introduce two datasets: BallPlay-V, based on monocular vision estimation, containing eight clips of different basketball skills; and BallPlay-M, based on multi-view optical motion capture systems, containing 35 minutes of diverse basketball interactions. We provide more details in the appendix. + +# 3.2. Defining Interaction Skills Using HOI Data + +We propose a data-driven approach to defining interaction skills as collections of Human-Object Interaction (HOI) + +state transitions that align with the intended skill semantics. For instance, the interaction skill of "picking up balls from various positions" can be encapsulated by various HOI motion clips, each capturing a unique instance of the action. These clips collectively form a collection of HOI state transitions that represent the interaction skill. + +# 3.3. Learning Interaction Skills by HOI Imitation + +Considering an interaction skill defined by a set of reference HOI state transitions, if a humanoid can manipulate objects such that its HOI state transitions closely resemble the reference, we consider the humanoid to have successfully learned the interaction skill. Based on this concept, we propose a method for learning interaction skills by imitating HOI state transitions through RL, which we call SkillMimic. Fig. 2 illustrates the concept of SkillMimic. + +In contrast to Behavior Clone [3, 5, 11] methods which rely on precise state-action data, SkillMimic learns actions through RL using state-only trajectories, demonstrating better tolerance to data noise and higher data efficiency. Unlike previous RL methods [33, 34, 50] that require manually designed rewards for each interaction skill, SkillMimic is fully data-driven, skill-agnostic, and scalable, making it capable of learning a wide range of basketball skills within a unified solution. Fig. 6 shows diverse basketball skills acquired using SkillMimic. Fig. 7 illustrates how the performance of the pickup skill improves as the reference data increases. + +# 3.3.1. Training Pipeline + +Fig. 3 (b) shows the training pipeline of SkillMimic. Given an HOI dataset with diverse skill-labeled clips (Fig. 3 (a)), SkillMimic trains an Interaction Skill (IS) policy by randomly selecting a frame from a clip corresponding to a chosen skill to initialize the humanoid and object states. The state $s_t$ (Sec. 3.3.2) and skill label $c_j$ are input into the IS policy, which predicts actions (Sec. 3.3.3) that are then simulated to produce the next state $s_{t+1}$ . A unified HOI imitation reward (Sec. 3.4) is designed to measure the consistency between the simulated state transitions and the reference transitions. When the humanoid falls or reaches the maximum simulation length, the environment resets, and the process repeats. After training converges, the IS policy enables the humanoid to execute interaction skills that closely mirror the reference demonstrations. + +# 3.3.2. HOI Observation + +The state observed by the IS policy is not under direct supervision and may theoretically encompass any information that is available in the simulation environment. Importantly, this information does not necessarily have to be included in the reference HOI data. Similar to prior arts [34], we transform all coordinates into the root local coordinate of the humanoid, which aligns the data distribution and benefits the generalization performance. For the humanoid, we observe + +![](images/108388a6dd10c8e77ca2f4a38f92ff9a4678079f2b4e4083d446233a967213f4.jpg) +Figure 3. Our system consists of three parts. (a) First, we capture real-world basketball skills to create a large Human-Object Interaction (HOI) motion dataset. (b) Second, we train an Interaction Skill (IS) policy to learn interaction skills by imitating the corresponding HOI data through reinforcement learning. Specifically, the IS policy takes as input the HOI state $\boldsymbol{s}_t$ and skill label $\boldsymbol{c}_j$ and predicts the action $\boldsymbol{a}_t$ . The new state $\boldsymbol{s}_{t+1}$ is calculated by the simulator. A unified HOI imitation reward is designed to imitate diverse HOI state transitions. (c) The third part involves training a High-Level Controller (HLC) to reuse the learned interaction skills for complex tasks. The HLC takes as input $\boldsymbol{s}_t$ and extra task observations $\boldsymbol{h}_t$ , e.g., the basket position, and predicts the skill label $\boldsymbol{c}_t$ to drive a pre-trained IS policy. + +its global root height, local body position, rotation, position velocity, and angular velocity. These representations form the humanoid proprioception $\pmb{o}_t^{prop}$ . In addition, we detect net contact forces $\pmb{o}_t^f$ for all fingertips, which helps to sense contact and accelerate training. For the object, we observe its local position, rotation, velocity, and angular velocity, which form the object observation $\pmb{o}_t^{obj}$ . Finally, the state perceived by the IS policy is + +$$ +\boldsymbol {s} _ {t} = \left\{\boldsymbol {o} _ {t} ^ {\text {p r o p}}, \boldsymbol {o} _ {t} ^ {f}, \boldsymbol {o} _ {t} ^ {\text {o b j}} \right\}. \tag {1} +$$ + +Skill labels $c_{j}$ are also used as inputs to the IS policy to differentiate different skills. We use one-hot encodings to represent the skill label. + +# 3.3.3. IS Policy and Action + +The policy output is modeled as a Gaussian distribution with dimensions equal to the DOF number of the humanoid robot, featuring constant variance. The mean is modeled by a three-layer MLP consisting of [1024, 512, 512] units with ReLU [30] activations. We use the action $\boldsymbol{a}_t$ sampled from the policy as the target joint rotations for a full set of PD controllers. The PD controllers adjust and output the joint torques to reach the target rotations. + +# 3.4. Unified HOI Imitation Reward + +This HOI imitation reward function is pivotal as it must not only accurately gauge the efficacy of interaction skill ac + +![](images/66fb65f6c000f579bae1ceb1628c22533205715568efee55d80bf05c548e745a.jpg) +Figure 4. We propose the Contact Graph (CG) to model general contacts within an explicitly defined scene. The node stores a binary value that denotes whether it contacts other nodes. Each edge stores a binary value indicating whether the two connected nodes are in contact. The node definition is unified for a certain scene and shared between diverse interactive skills. For example, we define three nodes: hands, hands-exclusive body, and ball, to form a simple CG to model contacts for diverse basketball skills. + +quisition but also avoid the incorporation of skill-specific settings to maintain robust cross-skill generalizability. + +Classic imitation rewards often focus solely on imitating locomotions [32], lacking an established baseline for HOI imitation. Although adversarial imitation rewards [33, 34] are effective in locomotion tasks, we find that they perform poorly in HOI imitation. This may be because adversarial imitation rewards are relatively coarse-grained, making it difficult for them to provide the precise guidance needed for accurate interaction learning. + +The body kinematics imitation reward $r_t^b$ encourages the + +alignment of body movements by considering joint position, rotation, and their velocities. Similarly, the object kinematics imitation reward $r_t^o$ ensures consistent object movements with the reference. However, solely relying on these two rewards often leads to unstable learning and unnatural motions. We incorporate a relative motion reward $r_t^{rel}$ to highlight the relative relationships in interactions, which is crucial for interaction naturalness. Recognizing the interlocking relationship between interaction and contact, we propose a straightforward and universal contact modeling approach called Contact Graph (CG). A corresponding CG imitation reward $r_t^{cg}$ is designed to enhance the precision of interaction imitation. Additionally, we propose an adaptive velocity regularization term $r_t^{reg}$ to suppress high-frequency jitters. These sub-rewards are multiplied rather than added to encourage balanced learning, thereby avoiding local optima, as justified in Tab. 1. The combination of these innovations forms a unified imitation reward, enabling the learning of diverse basketball skills in BallPlay-M with the same configuration. To conclude, the complete HOI imitation reward is + +$$ +r _ {t} = r _ {t} ^ {b} * r _ {t} ^ {o} * r _ {t} ^ {r e l} * r _ {t} ^ {r e g} * r _ {t} ^ {c g}. \tag {2} +$$ + +Next, we will discuss the methodology involved in crafting the CG reward $r_t^{cg}$ . Details about the kinematic imitation reward $r_t^b * r_t^o * r_t^{rel} * r_t^{reg}$ are in the appendix. + +# 3.4.1. Contact Graph + +We observe that existing imitation rewards [33, 34, 52] only measure kinematic properties and are insufficient to measure precise contacts. To tackle this problem, we propose the Contact Graph (CG) to model the contact in general interactions and design a CG imitation reward. + +As illustrated in Fig. 4, the CG is a complete graph where every pair of distinct nodes is connected by a unique edge, defined as $\mathcal{G} = \{\mathcal{V},\mathcal{E}\}$ , where $\mathcal{V}\in \{0,1\} ^k$ is the set of $k$ nodes and $\mathcal{E}\in \{0,1\}^{k(k - 1) / 2}$ is the set of edges. A CG node stores a binary value indicating whether it contacts other nodes. Each CG edge stores a binary label that denotes the contact between two nodes, where 1 represents contact, and 0 means no contact. CG node and edge values are calculated frame by frame, explicitly describing the mutual contact relationship between CG nodes at different moments. Contacts inside a node are not considered. We use the edge set $\mathcal{E}$ for imitation learning. + +The definitions of nodes are flexible. For example, a node can be a single part (e.g., a fingertip), or an aggregation of multiple parts (e.g., the whole left hand). Hinged objects can also be broken up into finer nodes [9, 13, 51]. The definition of CG is unified for a certain scene and shared between diverse interaction skills. For example, in basketball scenarios, we can aggregate two hands as a node, the rest of the bodies as a node, and the ball as a node, as illustrated in + +![](images/ad1feb3457ae5e2008a94cf8a2b172e78d088939e7c72ec242297faba93355d5.jpg) +(a) Ref + +![](images/72bf60387a71d1f6b94b2c64e199716d38b853e3d754056477b758776e778f45.jpg) +(b) w/o CGR + +![](images/689532f55a0f2f189061e0ed895870a83cb141124050d482b950fa95df31df79.jpg) +(c) w/ CGR + +![](images/3c45c633681c04d10ae1ed42321ccb77759132629c13643a148b67e9387153c9.jpg) +(d) Ref + +![](images/5f765883bad63b19c40eebe16fabc10a5d04c7ad37ebfabb598c519ff5e2dec3.jpg) +(e) w/o CGR + +![](images/8721a308430ce184e4e4febc7fc4e7bb0cacbac0cae4ce642e1c47a77909e750.jpg) +(f) w/ CGR + +![](images/03c397008252e041c9758785dc483f83af0ecc0a8d92abb5af498507e4a4bfa4.jpg) +(g) Ref +Figure 5. The HOI imitation falls into kinematic local-optimal solutions without Contact Graph Reward (CGR): (b) use the head to help control the ball; (e) use the wrist to contact the ball; (h) fail to catch the object; (k) support the table to keep balance. In comparison, the guidance of CGR effectively yields precise interactions, as shown in (c, f, i, l). +Fig. 4. This simple CG is effective enough for learning all the basketball skills covered in BallPlay-M. + +![](images/ba1cbea2dd01c3597180c785e80b43508d8f858293d76242eaf4fcc318ea72f8.jpg) +(h) w/o CGR + +![](images/c4ebc0b7cd1f8196316cdd8d15675b741446c7d18c237f4e8c29892cd7cd4ecd.jpg) +(i) w/ CGR + +![](images/6b7a535dbe9ba24512c66e8dc1fb7755c427e97164c22835565b00d4b565141f.jpg) +(j) Ref. + +![](images/ae9a51cf90bab51e9ac53f3a9514f7000a536b0cdd7ad88a6ae4e65e8877948f.jpg) +(k) w/o CGR on table + +![](images/82aca685518671e1e76ccafc72f389ac4cc54cac48dde76c3c143520f3a37d32.jpg) +(I) w/CGR + +# 3.4.2. Contact Graph Reward + +Kinematic imitation rewards [32, 33, 52] fall short in measuring contacts and thus yield poor performance in HOI imitation. For example, during the toss skill training (Fig. 5 (a)), the contact between the right hand and the ball usually cause the ball to drop, making the kinematic rewards drastically smaller. In this case, the policy may learn a local-optimal solution that uses the head and left hands to stabilize the ball, as demonstrated in Fig. 5 (b). To tackle this problem, we propose the Contact Graph Reward (CGR) as a critical complement of kinematic imitation rewards to learn precise contact imitation. The CG error is defined as $e_t^{cg} = |s_t^{cg} - \hat{s}_t^{cg}|$ , where $|\cdot|$ calculates element-wise absolute value, $s_t^{cg}$ and $\hat{s}_t^{cg}$ denotes the simulated and reference CG state respectively, which is the edge set $\mathcal{E}_t \in \{0,1\}^{k(k-1)/2}$ . The CGR is measured by CG error, with independent weights on different elements: + +$$ +r _ {t} ^ {c g} = \exp \left(- \sum_ {j = 1} ^ {J} \boldsymbol {\lambda} ^ {c g} [ j ] * \boldsymbol {e} _ {t} ^ {c g} [ j ]\right), \tag {3} +$$ + +where $e_t^{cg}[j]$ is the $j$ th element of $e_t^{cg} \in \{0,1\}^J$ , representing a contact error; $\lambda^{cg}[j]$ is the $j$ th element of $\lambda^{cg} \in \mathbb{R}^J$ , a hyperparameter controls the sensitivity of a contact. Fig. 5 and Tab. 1 justify the effectiveness of CGR, where experiments without CGR show inaccurate interactions while using CGR effectively eliminates such problems. + +# 3.5. Reusing Interaction Skills for High-level Tasks + +In this section, we propose training a High-Level Controller (HLC) to reuse the interaction skills acquired through SkillMimic to accomplish high-level tasks. As depicted in Fig. 3 (c), the HLC takes as input the current HOI state $s_t$ + +![](images/04724ae4c94bf9781dd4d256fe5f3b9446d27f5dd1cc13e89bc6d3688cd061de.jpg) + +![](images/81f434658e633d076d8eaa62b15976236335bd3e486e7e860ffde26067a8520c.jpg) + +![](images/40c1b9a8efcc35a2af225b7f9d86bf0e8d7eb7795b04815c0fc713b101174e57.jpg) + +![](images/a5afb105d44ad3196fd38b48c4c7a583220383258d035c1461dd6032cb3240e3.jpg) + +![](images/acf43d35437bb8d4e8e36e9e62f0a96186c01bc3cf31b8cec8d97e2b9b6d2317.jpg) + +![](images/b93370c0266fe545249025ae5b28597104a685b93558a0fb513c760796e91828.jpg) +(a) Layup +(f) Jump Shot + +![](images/e3ca519f182dd1bbf68eee41d2ac166a9d0a7eff56f549063c723d74b35cf357.jpg) +(b) Turnaround Layup +(g) Rebound + +![](images/85843f0257971d7780dba6974a54d2f093ec8741dc12527353844c045d4b8edb.jpg) +(c) Dribble Right +(h) Pass +Figure 6. Simulated humans exhibit comprehensive basketball skills. SkillMimic can teach humans a wide range of basketball skills using the same configuration in a purely data-driven manner, covering almost all fundamental basketball skills. Keyframes are placed in chronological order from left to right. + +![](images/df3eda1f4a2381cb6519824711901ee69e5af95c7bf06fc604314ee788742232.jpg) +(d) Dribble Forward +(i) Pickup + +![](images/dcd88d4d9142af324ac903e85a3acaa2c66af0927fab20a4dd879e676258c842.jpg) +(e) Dribble Turn +(j) Cross Leg + +and task observation $h_t$ , and outputs a discrete skill embedding $c_t$ that serves as inputs to a pre-trained SkillMimic policy, which subsequently generates actions for the humanoid. + +Compared to previous hierarchical approaches that are based on locomotion priors [34, 50], our method leverages interaction skills as prior, which significantly simplifies the design of task rewards and greatly accelerates training. Moreover, unlike their mandatory hierarchical structure, our approach is motivated by data efficiency - tasks can be alternatively solved through direct HOI imitation when sufficient demonstrations are available, as shown in pickup experiments in Sec. 4.2, Fig. 7, and Fig. 8. + +We present four representative high-level tasks: + +- Throwing: throw the ball to approach a certain height, grab the rebound, and keep on throwing the ball. +- Heading: dribble the ball to approach the target position. +- Circling: dribble the ball around the target position following a target radius. +- Scoring: dribble the ball toward the basket, timing the layup to score, retrieving the rebound, and repeating. + +Utilizing interaction skills as prior significantly improves learning efficiency for complex, long-horizon tasks. For example, a decent HLC for the Scoring task only requires 3 hours of training on an NVIDIA RTX 4090 GPU. + +# 4. Experiments + +We conduct extensive experiments to evaluate the performance of SkillMimic in learning interaction skills and its application for high-level tasks. Our evaluation consists of two main parts: (1) assessing the capability of SkillMimic in learning interaction skills, including comparisons with existing methods and ablation studies, and (2) demonstrating the effectiveness of reusing acquired interaction skills to learn high-level tasks. Details of the experimental settings and additional experiments on skill switching and robustness are provided in the appendix. + +# 4.1. Metrics + +We consider the following quantitative metrics: + +- Accuracy. The overall accuracy of HOI imitation, abbreviated as Acc., is defined per frame, and deems imitation accurate when the object position and body position errors are both under the thresholds and the contacts are correct. The object threshold is defined as $0.2m$ . The body threshold is defined as $0.1m$ . The Acc. is calculated by averaging the values of all frames. +- Position Error. The Mean Per-Joint Position Error (MPJPE) of the body $(E_{\mathrm{b - mpjpe}})$ and object $(E_{\mathrm{o - mpjpe}})$ is used to evaluate the positional imitation performance (in mm) following [27]. +- Contact Error. The contact error $E_{\mathrm{cg}}$ , ranging from 0 to 1, is defined as $\frac{1}{N}\sum_{t=1}^{N}\mathrm{MSE}(s_t^{cg}, \hat{s}_t^{cg})$ , where $N$ is the total frames of the reference HOI data. +- Success Rate. The execution success rate of an interaction skill. Detailed definitions of success criteria for different skills are provided in the appendix. + +# 4.2. Evaluating Interaction Skill Learning + +Since our approach is the first to learn interaction skills from HOI demonstrations, there are no direct benchmarks for comparison. Therefore, we adapt reward strategies commonly used in locomotion imitation, i.e., DeepMimic [32] and AMP [33], to an object-inclusive setting. Specifically, we adapt our reward to the styles of DeepMimic [32] and AMP [33] while keeping the other components unchanged for a fair comparison. We denote these variant versions as DeepMimic* and AMP*. In Tab. 2, we present the success rate on 4 typical basketball skills from BallPlay-M, where our method achieves state-of-the-art performance in learning basketball skills compared with variants using DeepMimic or AMP style rewards. We guess that AMP-like rewards are ineffective for learning precise interactions due to their coarse-grained supervision signals. + +![](images/5f69af15bcda4cf123d2ad5345bdfddccd5dbeeb7090b13712e116920e733001.jpg) +Figure 7. Demonstration of the pickup skill learned from 40 HOI motion clips. Yellow denotes the initial frame. Left: The humanoid picks up a stationary ball effortlessly. Middle: The humanoid intercepts a ball with random velocity. Right: The humanoid adjusts after missing the ball initially (the frame in blue) and successfully retrieves it on the second attempt, showcasing the potential for learning robust and generalizable skills through extensive data collection. + +
SkillMimic w/o MultiplicationSkillMimic w/o CGRSkillMimic
DatasetAcc. ↑Eb-mpipe ↓Eo-mpipe ↓Ecg ↓Acc. ↑Eb-mpipe ↓Eo-mpipe ↓Ecg ↓Acc. ↑Eb-mpipe ↓Eo-mpipe ↓Ecg ↓
GRAB [40]27.0%44.7180.20.72438.6%91.0180.10.33795.4%71.178.00.026
BallPlay-V7.5%40.11662.50.30613.6%88.5155.30.41282.4%56.882.90.087
+ +Table 1. Ablation study on imitation reward design. We experiment on two HOI datasets: GRAB and BallPlay-V. The result highlights the importance of the CG reward and multiplication in achieving precise HOI imitation. + +
MethodSuccess Rate ↑
Pick UpDribble ForwardLayupShot
DeepMimic*[32]19.6%68.5%98.9%97.8%
AMP*[33]0.0%13.6%0.0%1.6%
SkillMimic (ours)86.7%79.6%99.1%97.9%
+ +Table 2. Success rates across four typical basketball skills in BallPlay-M. Our method significantly outperforms variant methods that use imitation reward styles of DeepMimic and AMP. + +![](images/9b584801401090867cdb9c8cea8d03b8c41d3aa984d1ecc8e2d3803abc52f284.jpg) +(a) $1 - 0.5\%$ + +![](images/1066c9831f6721218ed935f2e4e36752dd65cf90a33c01c27bb5216dc4210ee6.jpg) +(b) $10 - 12.2\%$ + +![](images/1956039e558593426c1f1035c2d99e487675a856a2c88df695abc25ee6606ba5.jpg) +(c) $40 - 76.5\%$ +Figure 8. Pickup generalization performance with different training data scales. First number: clips; second: success rate. In each test, 1000 balls are randomly placed within 1 to 5 meters away from the center. Yellow dots indicate successful pickups and green dots represent failures. See Sec. 4.2 for details. + +![](images/0c6d6f488132e4a57a385ee583a13db367a78b0feb194e51502ad4bce15af87a.jpg) +(d) $131 - 85.6\%$ + +In Tab. 1, we do ablation on reward design, presenting detailed metrics of HOI imitation across two diverse datasets, GRAB [40] and BallPlay-V. The results indicate that the absence of reward multiplication leads to imbalanced sub-reward learning, e.g., favoring body motion $(E_{\mathrm{b - m}})$ over interactions. The absence of CGR leads to kinematic local optima. In contrast, our complete setting (SkillMimic) yields significantly higher performance on interaction-related metrics (Acc., $E_{\mathrm{o - m}}$ , and $E_{\mathrm{cg}}$ + +# 4.2.1. Ablation on Contact Graph Reward + +We conduct ablation experiments on diverse datasets to comprehensively evaluate the impact of Contact Graph Reward (CGR) on interaction skill learning. Fig. 5 illustrates the qualitative results of ablating CGR on the GRAB and BallPlay-V datasets. It can be clearly observed that without CGR, the humanoid often resorts to incorrect contacts to achieve more stable object control, such as using its head to push the ball, clamping the ball with its hands and legs, or using its hands to support itself on a table for balance. These examples illustrate typical kinematic local optima. By incorporating CGR, these local optima are effectively eliminated, resulting in accurate interactions. Additionally, the result in Tab. 1 also shows that, without CGR, the contact error $(E_{\mathrm{cg}})$ significantly increases, leading to kinematic local optima and causes a noticeable decline in overall imitation accuracy. These results underscore the critical importance of CGR for learning correct interactions. + +# 4.2.2. Data Scale and Generalization Performance + +To examine how the generalization performance grows as the data scale increases, we conduct ablation experiments on the pickup skill. We create four datasets by randomly selecting pickup clips from BallPlay-M, containing 1, 10, 40, and 131 clips, respectively. We train policy on each dataset with approximately 3.2 billion samples using SkillMimic. + +To evaluate the generalization performance, we randomly place balls within a circular area ranging from 1 to 5 meters around the humanoid—this distribution significantly exceeds the coverage of the reference data. Fig. 7 and Fig. 8 show the results. We can see that as the data scale increases, pickup performance improves significantly due to broader state transition coverage in larger datasets. This indicates + +![](images/93390c62e4ea195cf924a43c76988a9ecfb43c94be15d6343eec47bb45792817.jpg) +(a) Manual Control + +![](images/c5b9ed00f9c1e039bfa83d574b66cac390c92c1c4cf87f7da5f419c8ff92a787.jpg) +(b) Scoring +Figure 9. Our method supports training a single IS policy under a unified configuration to acquire various interaction skills. These interaction skills can be flexibly switched, as illustrated in (a), where yellow denotes shot, blue denotes pickup, and green denotes turnaround layup. Complex, long-horizon tasks can be easily achieved by training a high-level controller (HLC) to manage switching of the learned interaction skills: (b) scoring from random positions (c) dribbling to target locations, and (d) dribbling along an expanding radius. + +![](images/6fb08d22749ebae1504114fcf4241bec5e9fb5dc2b0bc7186ac5f8c2fb098f47.jpg) +(c) Heading + +![](images/0d50dc71dd70eb9377b14d06733963e465af0f4a1ab4d295f221d47723705871.jpg) +(d) Circling + +
MethodSuccess Rate↑
HeadingCirclingThrowingScoring
PPO[36]0.70%11.14%0.00%0.00%
ASE[34]0.19%4.37%0.00%0.00%
ASE*[34]0.31%7.21%0.00%0.00%
SkillMimic (ours)93.04%79.92%93.40%80.25%
+ +Table 3. Success rates on 4 high-level basketball tasks. Both PPO (learn from scratch), ASE (using locomotion prior), and $\mathrm{ASE^{*}}$ (using body and object motion prior) fail to converge on all tasks. In contrast, leveraging interaction skill priors acquired through SkillMimic, our method effectively learns these difficult tasks. + +our method's strong scalability and potential for large-scale learning. It is worth noting that we do not incorporate any additional designs for robustness and generalization. + +# 4.3. Evaluating High-level Tasks + +We conduct experiments on four high-level tasks described in Sec. 3.5. We employ a single IS policy as the interaction skill prior, trained using SkillMimic for around 4.5 billion samples across seven skills: pickup, layup, turnaround layup, dribble left, dribble right, dribble forward, and shot. During training, the IS policy is fixed, and only the HLC is trained. For tasks such as throwing, heading, and circling, we train for around 0.4 billion samples. For the more challenging scoring task, we train for around 1.2 billion samples. To objectively evaluate the performance of our method, we compared it against three baseline methods: (1) learning from scratch using PPO [36], (2) learning with body motion priors using ASE [34], and (3) learning with body and object motion priors using ASE (denoted as $\mathrm{ASE^{*}}$ ). The low-level controllers of ASE and $\mathrm{ASE^{*}}$ are trained using the same training steps as our IS policy. For task training, all methods are trained using identical task rewards and simulation steps for fair comparisons. + +Fig. 9 (a) showcases an example of manually controlling the pre-trained IS policy to perform skill switching. Fig. 9 (b) shows several trajectories of the HLC performing skill controls to accurately deliver the ball to the target basket. Fig. 9 (c) & (d) present a top-down view of the HLC control + +ling the humanoid to dribble the ball toward target locations and along an expanding radius. As shown in Tab. 3, PPO and ASE yield low success rates because the tasks are too challenging to master without interaction priors. The failure of $\mathrm{ASE^{*}}$ can be traced back to its low-level controller's inadequate ability to imitate interaction skills. This is primarily because adversarial rewards fail to provide effective guidance for precise interaction learning, which is similar to the failure mode observed in $\mathrm{AMP^{*}}$ . In contrast, by utilizing the interaction skills learned through SkillMimic, we greatly improve the performance of learning complex high-level interaction tasks. + +# 5. Discussion and Future Work + +We presented a novel framework for learning diverse interaction skills with a unified HOI imitation reward using reinforcement learning. Our approach demonstrates multi-skill mastery, skill reusability, and clear data scaling benefits, highlighting the significant potential of HOI imitation for learning complex interaction skills. However, the full application of this paradigm requires further exploration. Several key areas deserve particular attention: + +Firstly, exploring more general HOI skill learning within a single policy, such as household tasks or other types of sports, presents challenges in handling generalization across different object geometries. Moreover, perceiving multiple objects simultaneously requires a more general environmental perception approach, rather than directly obtaining privileged information for each object. + +Secondly, incorporating finer-grained control conditions, such as tracking [27], could enable more flexible control and generalization capabilities as HOI data scales up. + +Thirdly, the difficulty of collecting HOI data highlights the importance of exploring methods to learn robust interaction skills with limited HOI data, as well as investigating HOI data augmentation and generation techniques. + +Lastly, applying this technology to real robots would be highly valuable, such as enabling real-world humanoid robots to play basketball games. This endeavor could involve various challenges, including HOI data retargeting and sim-to-real transfer issues. + +# Acknowledgment + +We would like to thank Mingyang Su for his valuable assistance in motion capture data collection and Shunlin Lu for helpful discussions. This work was supported in part by Unitree Robotics, International Digital Economy Academy (IDEA), and the Innovation and Technology Fund of HK-SAR (GHX/054/21GD). + +# References + +[1] Kai Arulkumaran, Marc Peter Deisenroth, Miles Brundage, and Anil Anthony Bharath. Deep reinforcement learning: A brief survey. IEEE Signal Processing Magazine, 34(6):26-38, 2017. 2 +[2] Jinseok Bae, Jungdam Won, Donggeun Lim, Cheol-Hui Min, and Young Min Kim. Pmp: Learning to physically interact with environments using part-wise motion priors. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-10, 2023. 1, 2 +[3] Kevin Black, Noah Brown, Danny Driess, Adnan Esmail, Michael Equi, Chelsea Finn, Niccolo Fusai, Lachy Groom, Karol Hausman, Brian Ichter, et al. Pi0: A vision-language-action flow model for general robot control. arXiv preprint arXiv:2410.24164, 2024. 3 +[4] Jona Braun, Sammy Christen, Muhammed Kocabas, Emre Aksan, and Otmar Hilliges. Physically plausible full-body hand-object interaction synthesis. arXiv preprint arXiv:2309.07907, 2023. 2 +[5] Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Xi Chen, Krzysztof Choromanski, Tianli Ding, Danny Driess, Avinava Dubey, Chelsea Finn, et al. Rt-2: Vision-language-action models transfer web knowledge to robotic control. arXiv preprint arXiv:2307.15818, 2023. 3 +[6] Stelian Coros, Philippe Beaudoin, and Michiel Van de Panne. Generalized biped walking control. ACM Transactions On Graphics (TOG), 29(4):1-9, 2010. 2 +[7] Jieming Cui, Tengyu Liu, Nian Liu, Yaodong Yang, Yixin Zhu, and Siyuan Huang. Anyskill: Learning openvocabulary physical skill for interactive agents. In Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2 +[8] Zhiyang Dou, Xuelin Chen, Qingnan Fan, Taku Komura, and Wenping Wang. C-ase: Learning conditional adversarial skill embeddings for physics-based characters. In SIGGRAPH Asia 2023 Conference Papers, pages 1-11, 2023. 1, 2 +[9] Zicong Fan, Omid Taheri, Dimitrios Tzionas, Muhammed Kocabas, Manuel Kaufmann, Michael J Black, and Otmar Hilliges. Arctic: A dataset for dexterous bimanual handobject manipulation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12943-12954, 2023. 5 +[10] Zipeng Fu, Qingqing Zhao, Qi Wu, Gordon Wetzstein, and Chelsea Finn. Humanplus: Humanoid shadowing and imitation from humans. In 8th Annual Conference on Robot Learning, 2024. 1 +[11] Zipeng Fu, Tony Z Zhao, and Chelsea Finn. Mobile aloha: Learning bimanual mobile manipulation with low-cost + +whole-body teleoperation. arXiv preprint arXiv:2401.02117, 2024.3 +[12] Jiawei Gao, Ziqin Wang, Zeqi Xiao, Jingbo Wang, Tai Wang, Jinkun Cao, Xiaolin Hu, Si Liu, Jifeng Dai, and Jiangmiao Pang. Coohoi: Learning cooperative human-object interaction with manipulated object dynamics. Advances in Neural Information Processing Systems, 37:79741-79763, 2024. 1 +[13] Haoran Geng, Helin Xu, Chengyang Zhao, Chao Xu, Li Yi, Siyuan Huang, and He Wang. Gapartner: Cross-category domain-generalizable object perception and manipulation via generalizable and actionable parts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7081-7091, 2023. 5 +[14] Anindita Ghosh, Rishabh Dabral, Vladislav Golyanik, Christian Theobalt, and Philipp Slusallek. Imos: Intent-driven full-body motion synthesis for human-object interactions. In Eurographics, 2023. 2 +[15] Mohamed Hassan, Yunrong Guo, Tingwu Wang, Michael Black, Sanja Fidler, and Xue Bin Peng. Synthesizing physical character-scene interactions. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-9, 2023. 1, 2 +[16] Tairan He, Zhengyi Luo, Xialin He, Wenli Xiao, Chong Zhang, Weinan Zhang, Kris M Kitani, Changliu Liu, and Guanya Shi. Omnih2o: Universal and dexterous human-to-humanoid whole-body teleoperation and learning. In 8th Annual Conference on Robot Learning, 2024. 1 +[17] Edmond SL Ho, Taku Komura, and Chiew-Lan Tai. Spatial relationship preserving character motion adaptation. In ACM SIGGRAPH 2010 papers, pages 1-8, 2010. 3 +[18] Jonathan Ho and Stefano Ermon. Generative adversarial imitation learning. Advances in neural information processing systems, 29, 2016. 2 +[19] Jessica K Hodgins, Wayne L Wooten, David C Brogan, and James F O'Brien. Animating human athletics. In Proceedings of the 22nd annual conference on Computer graphics and interactive techniques, pages 71-78, 1995. 2 +[20] Nan Jiang, Tengyu Liu, Zhexuan Cao, Jieming Cui, Zhiyuan Zhang, Yixin Chen, He Wang, Yixin Zhu, and Siyuan Huang. Full-body articulated human-object interaction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9365-9376, 2023. 2 +[21] Nan Jiang, Zhiyuan Zhang, Hongjie Li, Xiaoxuan Ma, Zan Wang, Yixin Chen, Tengyu Liu, Yixin Zhu, and Siyuan Huang. Scaling up dynamic human-scene interaction modeling. arXiv preprint arXiv:2403.08629, 2024. 2 +[22] Jordan Juravsky, Yunrong Guo, Sanja Fidler, and Xue Bin Peng. Padl: Language-directed physics-based character control. In SIGGRAPH Asia 2022 Conference Papers, pages 1-9, 2022. 2 +[23] Jiaman Li, Alexander Clegg, Roozbeh Mottaghi, Jiajun Wu, Xavier Puig, and C Karen Liu. Controllable human-object interaction synthesis. arXiv preprint arXiv:2312.03913, 2023. 1, 2 +[24] Jiaman Li, Jiajun Wu, and C Karen Liu. Object motion guided human motion synthesis. ACM Transactions on Graphics (TOG), 42(6):1-11, 2023. 2 + +[25] Libin Liu and Jessica Hodgins. Learning to schedule control fragments for physics-based characters using deep q-learning. ACM Transactions on Graphics (TOG), 36(3):1-14, 2017. 2 +[26] Libin Liu and Jessica Hodgins. Learning basketball dribbling skills using trajectory optimization and deep reinforcement learning. ACM Transactions on Graphics (TOG), 37(4):1-14, 2018. 2 +[27] Zhengyi Luo, Jinkun Cao, Kris Kitani, Weipeng Xu, et al. Perpetual humanoid control for real-time simulated avatars. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10895-10904, 2023. 6, 8 +[28] Zhengyi Luo, Jinkun Cao, Sammy Christen, Alexander Winkler, Kris Kitani, and Weipeng Xu. Omnigrasp: Grasping diverse objects with simulated humanoids. Advances in Neural Information Processing Systems, 37:2161-2184, 2024. 1 +[29] Zhengyi Luo, Jiashun Wang, Kangni Liu, Haotian Zhang, Chen Tessler, Jingbo Wang, Ye Yuan, Jinkun Cao, Zihui Lin, Fengyi Wang, et al. Smplolympics: Sports environments for physically simulated humanoids. arXiv preprint arXiv:2407.00187, 2024. 2 +[30] Vinod Nair and Geoffrey E Hinton. Rectified linear units improve restricted boltzmann machines. In Proceedings of the 27th international conference on machine learning (ICML-10), pages 807–814, 2010. 4 +[31] Liang Pan, Jingbo Wang, Buzhen Huang, Junyu Zhang, Haofan Wang, Xu Tang, and Yangang Wang. Synthesizing physically plausible human motions in 3d scenes. In 2024 International Conference on 3D Vision (3DV), pages 1498-1507. IEEE, 2024. 2 +[32] Xue Bin Peng, Pieter Abbeel, Sergey Levine, and Michiel van de Panne. Deepmimic. ACM Transactions on Graphics, page 1-14, 2018. 1, 2, 4, 5, 6, 7 +[33] Xue Bin Peng, Ze Ma, Pieter Abbeel, Sergey Levine, and Angjoo Kanazawa. Amp: Adversarial motion priors for stylized physics-based character control. ACM Transactions on Graphics, page 1-20, 2021. 2, 3, 4, 5, 6, 7 +[34] Xue Bin Peng, Yunrong Guo, Lina Halper, Sergey Levine, and Sanja Fidler. Ase: Large-scale reusable adversarial skill embeddings for physically simulated characters. ACM Trans. Graph., 41(4), 2022. 1, 2, 3, 4, 5, 6, 8 +[35] Jiawei Ren, Mingyuan Zhang, Cunjun Yu, Xiao Ma, Liang Pan, and Ziwei Liu. Insactor: Instruction-driven physics-based characters. Advances in Neural Information Processing Systems, 36, 2024. 2 +[36] John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347, 2017. 8 +[37] Sebastian Starke, Yiwei Zhao, Taku Komura, and Kazi Zaman. Local motion phases for learning multi-contact character movements. ACM Trans. Graph., 39(4), 2020. 2 +[38] Sebastian Starke, Yiwei Zhao, Fabio Zinno, and Taku Komura. Neural animation layering for synthesizing martial arts movements. ACM Trans. Graph., 40(4), 2021. 2 +[39] Jingkai Sun, Qiang Zhang, Yiqun Duan, Xiaoyang Jiang, Chong Cheng, and Renjing Xu. Prompt, plan, perform: Llm-based humanoid control via quantized imitation learning. arXiv preprint arXiv:2309.11359, 2023. 2 + +[40] Omid Taheri, Nima Ghorbani, Michael J. Black, and Dimitrios Tzionas. GRAB: A dataset of whole-body human grasping of objects. In European Conference on Computer Vision (ECCV), 2020. 7 +[41] Jie Tan, Yuting Gu, C Karen Liu, and Greg Turk. Learning bicycle stunts. ACM Transactions on Graphics (TOG), 33 (4):1-12, 2014. 2 +[42] Chen Tessler, Yoni Kasten, Yunrong Guo, Shie Mannor, Gal Chechik, and Xue Bin Peng. Calm: Conditional adversarial latent models for directable virtual characters. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-9, 2023. 1, 2 +[43] Yinhuai Wang, Jing Lin, Ailing Zeng, Zhengyi Luo, Jian Zhang, and Lei Zhang. Physhoi: Physics-based imitation of dynamic human-object interaction. arXiv preprint arXiv:2312.04393, 2023. 3 +[44] Jungdam Won, Deepak Gopinath, and Jessica Hodgins. Physics-based character controllers using conditional vaes. ACM Transactions on Graphics (TOG), 41(4):1-12, 2022. 2 +[45] Yan Wu, Jiahao Wang, Yan Zhang, Siwei Zhang, Otmar Hilliges, Fisher Yu, and Siyu Tang. Saga: Stochastic whole-body grasping with contact. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 2 +[46] Zeqi Xiao, Tai Wang, Jingbo Wang, Jinkun Cao, Wenwei Zhang, Bo Dai, Dahua Lin, and Jiangmiao Pang. Unified human-scene interaction via prompted chain-of-contacts. In The Twelfth International Conference on Learning Representations, 2024. 1 +[47] Sirui Xu, Zhengyuan Li, Yu-Xiong Wang, and Liang-Yan Gui. Interdiff: Generating 3d human-object interactions with physics-informed diffusion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14928-14940, 2023. 1, 2 +[48] Sirui Xu, Ziyin Wang, Yu-Xiong Wang, and Liang-Yan Gui. Interdreamer: Zero-shot text to 3d dynamic human-object interaction. arXiv preprint arXiv:2403.19652, 2024. 2 +[49] KangKang Yin, Kevin Loken, and Michiel Van de Panne. Simbicon: Simple biped locomotion control. ACM Transactions on Graphics (TOG), 26(3):105-es, 2007. 2 +[50] Haotian Zhang, Ye Yuan, Viktor Makoviychuk, Yunrong Guo, Sanja Fidler, Xue Bin Peng, and Kayvon Fatahalian. Learning physically simulated tennis skills from broadcast videos. ACM Trans. Graph., 2023. 1, 2, 3, 6 +[51] Hui Zhang, Sammy Christen, Zicong Fan, Luocheng Zheng, Jemin Hwangbo, Jie Song, and Otmar Hilliges. Artigrasp: Physically plausible synthesis of bi-manual dexterous grasping and articulation. In 2024 International Conference on 3D Vision (3DV), pages 235-246. IEEE, 2024. 5 +[52] Yunbo Zhang, Deepak Gopinath, Yuting Ye, Jessica Hodgins, Greg Turk, and Jungdam Won. Simulation and retargeting of complex multi-character interactions. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-11, 2023. 3, 5 +[53] Qingxu Zhu, He Zhang, Mengting Lan, and Lei Han. Neural categorical priors for physics-based character control. ACM Transactions on Graphics (TOG), 42(6):1-16, 2023. 2 \ No newline at end of file diff --git a/2025/SkillMimic_ Learning Basketball Interaction Skills from Demonstrations/images.zip b/2025/SkillMimic_ Learning Basketball Interaction Skills from Demonstrations/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..a784f2aa2a6c6d282d87a88ba89fece4e4d43ff5 --- /dev/null +++ b/2025/SkillMimic_ Learning Basketball Interaction Skills from Demonstrations/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0f20fe8a7aa0895649c941a889d5b5e8ef52b6d490630a58b6805ca7af017479 +size 507300 diff --git a/2025/SkillMimic_ Learning Basketball Interaction Skills from Demonstrations/layout.json b/2025/SkillMimic_ Learning Basketball Interaction Skills from Demonstrations/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..425b115bd5ecbee343485338561e25a1bbc2fbaf --- /dev/null +++ b/2025/SkillMimic_ Learning Basketball Interaction Skills from Demonstrations/layout.json @@ -0,0 +1,10334 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 80, + 95, + 530, + 113 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 95, + 530, + 113 + ], + "spans": [ + { + "bbox": [ + 80, + 95, + 530, + 113 + ], + "type": "text", + "content": "SkillMimic: Learning Basketball Interaction Skills from Demonstrations" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 133, + 541, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 133, + 541, + 163 + ], + "spans": [ + { + "bbox": [ + 67, + 133, + 541, + 163 + ], + "type": "text", + "content": "Yinhuai Wang" + }, + { + "bbox": [ + 67, + 133, + 541, + 163 + ], + "type": "inline_equation", + "content": "^{1,2*}" + }, + { + "bbox": [ + 67, + 133, + 541, + 163 + ], + "type": "text", + "content": " Qihan Zhao" + }, + { + "bbox": [ + 67, + 133, + 541, + 163 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 67, + 133, + 541, + 163 + ], + "type": "text", + "content": " Runyi Yu" + }, + { + "bbox": [ + 67, + 133, + 541, + 163 + ], + "type": "inline_equation", + "content": "^{1,2*}" + }, + { + "bbox": [ + 67, + 133, + 541, + 163 + ], + "type": "text", + "content": " Hok Wai Tsui" + }, + { + "bbox": [ + 67, + 133, + 541, + 163 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 67, + 133, + 541, + 163 + ], + "type": "text", + "content": " Ailing Zeng" + }, + { + "bbox": [ + 67, + 133, + 541, + 163 + ], + "type": "inline_equation", + "content": "^{6\\dagger}" + }, + { + "bbox": [ + 67, + 133, + 541, + 163 + ], + "type": "text", + "content": " Jing Lin" + }, + { + "bbox": [ + 67, + 133, + 541, + 163 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 67, + 133, + 541, + 163 + ], + "type": "text", + "content": " Zhengyi Luo" + }, + { + "bbox": [ + 67, + 133, + 541, + 163 + ], + "type": "inline_equation", + "content": "^{7}" + }, + { + "bbox": [ + 67, + 133, + 541, + 163 + ], + "type": "text", + "content": " Jiwen Yu" + }, + { + "bbox": [ + 67, + 133, + 541, + 163 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 67, + 133, + 541, + 163 + ], + "type": "text", + "content": " Xiu Li" + }, + { + "bbox": [ + 67, + 133, + 541, + 163 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 67, + 133, + 541, + 163 + ], + "type": "text", + "content": " Qifeng Chen" + }, + { + "bbox": [ + 67, + 133, + 541, + 163 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 67, + 133, + 541, + 163 + ], + "type": "text", + "content": " Jian Zhang" + }, + { + "bbox": [ + 67, + 133, + 541, + 163 + ], + "type": "inline_equation", + "content": "^{3\\dagger}" + }, + { + "bbox": [ + 67, + 133, + 541, + 163 + ], + "type": "text", + "content": " Lei Zhang" + }, + { + "bbox": [ + 67, + 133, + 541, + 163 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 67, + 133, + 541, + 163 + ], + "type": "text", + "content": " Ping Tan" + }, + { + "bbox": [ + 67, + 133, + 541, + 163 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 69, + 169, + 541, + 210 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 169, + 541, + 210 + ], + "spans": [ + { + "bbox": [ + 69, + 169, + 541, + 210 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 69, + 169, + 541, + 210 + ], + "type": "text", + "content": "Hong Kong University of Science and Technology " + }, + { + "bbox": [ + 69, + 169, + 541, + 210 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 69, + 169, + 541, + 210 + ], + "type": "text", + "content": "Unitree Robotics " + }, + { + "bbox": [ + 69, + 169, + 541, + 210 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 69, + 169, + 541, + 210 + ], + "type": "text", + "content": "Peking University Shenzhen Graduate School " + }, + { + "bbox": [ + 69, + 169, + 541, + 210 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 69, + 169, + 541, + 210 + ], + "type": "text", + "content": "Tsinghua University " + }, + { + "bbox": [ + 69, + 169, + 541, + 210 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 69, + 169, + 541, + 210 + ], + "type": "text", + "content": "International Digital Economy Academy " + }, + { + "bbox": [ + 69, + 169, + 541, + 210 + ], + "type": "inline_equation", + "content": "^{6}" + }, + { + "bbox": [ + 69, + 169, + 541, + 210 + ], + "type": "text", + "content": "Tencent " + }, + { + "bbox": [ + 69, + 169, + 541, + 210 + ], + "type": "inline_equation", + "content": "^{7}" + }, + { + "bbox": [ + 69, + 169, + 541, + 210 + ], + "type": "text", + "content": "Carnegie Mellon University yinhuai.wang@connect.ust.hk pingtan@ust.hk" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 57, + 225, + 555, + 327 + ], + "blocks": [ + { + "bbox": [ + 57, + 225, + 555, + 327 + ], + "lines": [ + { + "bbox": [ + 57, + 225, + 555, + 327 + ], + "spans": [ + { + "bbox": [ + 57, + 225, + 555, + 327 + ], + "type": "image", + "image_path": "f0da1c7935060e5d15ea20c148848d4ecae60676075df2db709bef83387d5ed3.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 334, + 555, + 369 + ], + "lines": [ + { + "bbox": [ + 55, + 334, + 555, + 369 + ], + "spans": [ + { + "bbox": [ + 55, + 334, + 555, + 369 + ], + "type": "text", + "content": "Figure 1. We propose a novel approach that for the first time enables physically simulated humanoids to learn a variety of basketball interaction skills from Human-Object Interaction (HOI) data, including but not limited to shooting (blue), retrieving (red), and turnaround layup (yellow). Once acquired, these interaction skills can be composed to accomplish complex tasks, such as consecutive scoring (green)." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 152, + 381, + 200, + 393 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 381, + 200, + 393 + ], + "spans": [ + { + "bbox": [ + 152, + 381, + 200, + 393 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 407, + 297, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 407, + 297, + 684 + ], + "spans": [ + { + "bbox": [ + 54, + 407, + 297, + 684 + ], + "type": "text", + "content": "Traditional reinforcement learning methods for human-object interaction (HOI) rely on labor-intensive, manually designed skill rewards that do not generalize well across different interactions. We introduce SkillMimic, a unified data-driven framework that fundamentally changes how agents learn interaction skills by eliminating the need for skill-specific rewards. Our key insight is that a unified HOI imitation reward can effectively capture the essence of diverse interaction patterns from HOI datasets. This enables SkillMimic to learn a single policy that not only masters multiple interaction skills but also facilitates skill transitions, with both diversity and generalization improving as the HOI dataset grows. For evaluation, we collect and introduce two basketball datasets containing approximately 35 minutes of diverse basketball skills. Extensive experiments show that SkillMimic successfully masters a wide range of basketball skills including stylistic variations in dribbling, layup, and shooting. Moreover, these learned skills can be effectively composed by a high-level controller to accomplish complex and long-horizon tasks such as consecutive scoring, opening new possibilities for scalable and generalizable interaction skill learning. Project page: https://ingrid789.github.io/SkillMimic/" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 315, + 381, + 394, + 393 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 381, + 394, + 393 + ], + "spans": [ + { + "bbox": [ + 315, + 381, + 394, + 393 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 402, + 555, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 402, + 555, + 666 + ], + "spans": [ + { + "bbox": [ + 313, + 402, + 555, + 666 + ], + "type": "text", + "content": "Human-Object Interaction (HOI) represents one of the most fundamental yet challenging capabilities in robotics and character animation [10, 12, 16, 23, 28, 46, 47]. Among various HOI scenarios, basketball serves as an exemplary testbed that demands mastery of complex, dynamic, and precise interactions. Despite significant progress, existing physics-based character animation methods predominantly focus on imitation learning of locomotion skills [8, 32–34, 42] through Reinforcement Learning (RL), leaving a significant gap for HOI imitation. While some methods can learn interaction skills, such as striking pillars [34], playing tennis [50], climbing ropes [2], and carrying boxes [15], they typically involve handcrafted reward functions for each different interaction skill. However, designing skill-specific rewards is labor-intensive and fundamentally limits generalization. This limitation becomes particularly acute in basketball, where the complexity and diversity of skills would demand extraordinary engineering effort for reward design. Consequently, existing methods [2, 8, 15, 32–34, 42, 50] fail to provide a unified learning framework for a single policy to learn diverse interaction skills, let alone achieving complex long-term tasks such as consecutive basketball scoring." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 666, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 666, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 666, + 556, + 715 + ], + "type": "text", + "content": "How can a simulated humanoid learn a wide variety of interaction skills in a simple and scalable manner? Consider how humans naturally acquire complex skills: amateur basketball players often develop their expertise by watching" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 70, + 702, + 219, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 702, + 219, + 714 + ], + "spans": [ + { + "bbox": [ + 70, + 702, + 219, + 714 + ], + "type": "text", + "content": "*Equal Contribution. † Corresponding author." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "17540" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "content": "and imitating game footage, even without formal coaching. This self-directed learning process involves synchronizing their body movements with both the observed player motions and ball trajectories through practice." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 121, + 295, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 121, + 295, + 300 + ], + "spans": [ + { + "bbox": [ + 55, + 121, + 295, + 300 + ], + "type": "text", + "content": "Inspired by these observations, we propose a data-driven method called SkillMimic, which mimics HOI through RL to learn interaction skills. SkillMimic can learn various interaction skills purely from HOI datasets, such as diverse styles of basketball shooting, layup, and dribbling skills, and even a robust pickup skill that enables picking up balls in random locations and motions. Notably, SkillMimic uses the exact same configuration to learn different interaction skills, with identical hyperparameters. This allows us to train a single Interaction Skill (IS) policy to learn multiple interaction skills and achieve smooth skill switching. By training a high-level policy to reuse these learned interaction skills, we can accomplish challenging high-level tasks such as scoring, which requires the humanoid to composite diverse basketball interaction skills to score accurately." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 302, + 295, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 302, + 295, + 456 + ], + "spans": [ + { + "bbox": [ + 55, + 302, + 295, + 456 + ], + "type": "text", + "content": "To advance basketball skill learning, we introduce two HOI datasets: BallPlay-V, which captures player and ball motion from RGB videos across eight basic skills, and BallPlay-M, which uses optical motion capture to record 35 minutes of comprehensive basketball interactions. Using these datasets, SkillMimic successfully learns diverse basketball skills through a unified IS policy. With an additional high-level controller for skill composition, our system can execute complex tasks like consecutive scoring. Our approach demonstrates superior efficiency in learning both interaction skills and multi-phase, long-horizon tasks compared to existing methods, opening new possibilities for scalable and generalizable interaction skill learning." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 459, + 251, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 459, + 251, + 470 + ], + "spans": [ + { + "bbox": [ + 67, + 459, + 251, + 470 + ], + "type": "text", + "content": "Specifically, our contributions are as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 472, + 295, + 711 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 56, + 472, + 295, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 472, + 295, + 533 + ], + "spans": [ + { + "bbox": [ + 56, + 472, + 295, + 533 + ], + "type": "text", + "content": "- SkillMimic: a data-driven paradigm for learning diverse humanoid interaction skills through RL. It supports joint learning and switching of diverse basketball interaction skills, with skill diversity and generalization improving as the demonstration data grows, showcasing the scalability." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 533, + 295, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 533, + 295, + 591 + ], + "spans": [ + { + "bbox": [ + 56, + 533, + 295, + 591 + ], + "type": "text", + "content": "- Contact graph: We propose a simple and general contact modeling method that applies to diverse skills, called the contact graph. A contact graph reward is designed to enable precise contact imitation, which proves to be critical for learning precise interaction skills." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 592, + 295, + 627 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 592, + 295, + 627 + ], + "spans": [ + { + "bbox": [ + 56, + 592, + 295, + 627 + ], + "type": "text", + "content": "- Unified HOI imitation reward: We propose a set of important designs that form a unified reward configuration for imitation learning of various interaction skills." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 628, + 295, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 628, + 295, + 687 + ], + "spans": [ + { + "bbox": [ + 56, + 628, + 295, + 687 + ], + "type": "text", + "content": "- A hierarchical solution for reusing interaction skills: We propose training a high-level controller to reuse the interaction skills acquired by SkillMimic to accomplish challenging high-level tasks, which for the first time achieves continuous basketball scoring." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 688, + 295, + 711 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 688, + 295, + 711 + ], + "spans": [ + { + "bbox": [ + 56, + 688, + 295, + 711 + ], + "type": "text", + "content": "- BallPlay Datasets: We introduce two basketball datasets to facilitate research on interaction skill learning." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 71, + 400, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 71, + 400, + 83 + ], + "spans": [ + { + "bbox": [ + 314, + 71, + 400, + 83 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 92, + 555, + 368 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 92, + 555, + 368 + ], + "spans": [ + { + "bbox": [ + 313, + 92, + 555, + 368 + ], + "type": "text", + "content": "Locomotion Imitation Learning. Mimicking human motion for robot control is an efficient method of learning humanoid skills. DeepMimic [32], a pioneer in this field, uses imitation learning to perform a variety of highly dynamic locomotion skills. The introduction of Generative Adversarial Imitation Learning (GAIL) [18] into humanoid imitation learning by AMP [33] lessens constraints on data alignment, thus enhancing its versatility. ASE [34] further amplifies motion diversity in GAIL training, incorporates a pre-trained low-level policy to acquire locomotion skills, and deploys a high-level policy to repurpose these locomotion skills for specific tasks, such as striking pillars. These methodologies are subsequently expanded in numerous efforts to augment condition and text control [8, 22, 29, 31, 35, 39, 42, 44, 53], as well as to design automated rewards [7]. However, a limitation of these studies is their exclusive focus on learning isolated locomotion skills through imitation. This requires additional pipeline and reward design for learning interaction skills, even for relatively simple ones, such as moving boxes [15] or striking pillars [34]. In contrast, our approach enables unified learning of diverse interaction skills, eliminating the need for skill-specific reward engineering." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 378, + 556, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 378, + 556, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 378, + 556, + 628 + ], + "type": "text", + "content": "Learning Human-Object Interaction (HOI). Learning HOI in simulation has long posed challenges in the fields of graphics and robotics [14, 20, 21, 23, 24, 37, 38, 45, 47, 48]. Early studies often relied on manually designed control structures. For example, Hodgins et al. [19] implement a variety of sports interactions through the use of manually designed state machine controllers. Yin et al. [49] use finite-state machines to generate a large variety of gaits. Coros et al. [6] utilize an inverted pendulum model to control foot placement. Recently, deep reinforcement learning [1] has been widely employed in character control, giving rise to an array of methods based on policy network control. These methods have been applied to a range of challenging activities, such as basketball [26], skateboarding [25], and cycling [41], etc. The use of imitation rewards further lessens the reward design for locomotion styles. Zhang et al. [50] propose a framework to learn diverse tennis skills from broadcast videos. Bae et al. [2] incorporate multiple part-wise motion priors for diverse whole-body interactions. Similarly, Braun et al. [4] develop a framework for synthesizing whole-body grasps." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 630, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 555, + 715 + ], + "type": "text", + "content": "Applying imitation learning for HOI is intuitive, however, effectively learning diverse interaction skills in a unified framework remains unresolved. Naively extending locomotion imitation methods [32, 33] to HOI imitation often yields unstable and unnatural results, as these methods do not address the unbalanced rewards inherent in HOI, nor do they properly account for the crucial aspects of relative mo" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17541" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 70, + 294, + 149 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 294, + 149 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 294, + 149 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 294, + 149 + ], + "type": "image", + "image_path": "6c3ee744f479b9227dff4936e632e46e21c71f76c366684addf939351f46f0aa.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 157, + 296, + 236 + ], + "lines": [ + { + "bbox": [ + 55, + 157, + 296, + 236 + ], + "spans": [ + { + "bbox": [ + 55, + 157, + 296, + 236 + ], + "type": "text", + "content": "Figure 2. Concept of SkillMimic. We define an interaction skill as a set of Human-Object Interaction (HOI) state transitions that align with the intended skill semantics. These state transitions can be derived from captured HOI motion clips. If a simulated humanoid can manipulate objects such that the resulting HOI state transitions closely match those of the reference, we consider the humanoid to have successfully learned the interaction skill." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 256, + 295, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 256, + 295, + 376 + ], + "spans": [ + { + "bbox": [ + 54, + 256, + 295, + 376 + ], + "type": "text", + "content": "tion and contact in HOI. Modeling interaction through interaction graphs is a well-established approach [17]. While this method has proven useful for HOI imitation [52], it remains limited by considering only kinematic relationships, failing to ensure proper contact and becoming unstable in learning diverse scenarios. In contrast, our contact graph imitation reward explicitly prioritize physical contact learning, significantly improving the success rate of HOI imitation and enabling unified HOI imitation of diverse complex interaction skills for the first time." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 376, + 295, + 401 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 376, + 295, + 401 + ], + "spans": [ + { + "bbox": [ + 55, + 376, + 295, + 401 + ], + "type": "text", + "content": "This work builds upon and significantly extends ideas initially explored in our early technical report [43]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 412, + 128, + 425 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 412, + 128, + 425 + ], + "spans": [ + { + "bbox": [ + 55, + 412, + 128, + 425 + ], + "type": "text", + "content": "3. SkillMimic" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 433, + 296, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 433, + 296, + 553 + ], + "spans": [ + { + "bbox": [ + 55, + 433, + 296, + 553 + ], + "type": "text", + "content": "In this section, we introduce a unified data-driven framework for simulated humanoids to learn various basketball interaction skills via simulation and reinforcement learning (RL). We begin by introducing how we define interaction skills with Human-Object Interaction (HOI) data (Sec. 3.2). Next, we explain how to learn diverse interaction skills by imitating HOI data (Sec. 3.3) and discuss the designing of a unified HOI imitation reward (Sec. 3.4). Finally, we demonstrate how the learned interaction skills can be reused to perform high-level tasks (Sec. 3.5)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 561, + 177, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 561, + 177, + 574 + ], + "spans": [ + { + "bbox": [ + 55, + 561, + 177, + 574 + ], + "type": "text", + "content": "3.1. The BallPlay Dataset" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 578, + 295, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 578, + 295, + 664 + ], + "spans": [ + { + "bbox": [ + 55, + 578, + 295, + 664 + ], + "type": "text", + "content": "To address the scarcity of basketball HOI data and facilitate research on interaction skill learning, we introduce two datasets: BallPlay-V, based on monocular vision estimation, containing eight clips of different basketball skills; and BallPlay-M, based on multi-view optical motion capture systems, containing 35 minutes of diverse basketball interactions. We provide more details in the appendix." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 672, + 280, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 672, + 280, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 672, + 280, + 685 + ], + "type": "text", + "content": "3.2. Defining Interaction Skills Using HOI Data" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "type": "text", + "content": "We propose a data-driven approach to defining interaction skills as collections of Human-Object Interaction (HOI)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": "state transitions that align with the intended skill semantics. For instance, the interaction skill of \"picking up balls from various positions\" can be encapsulated by various HOI motion clips, each capturing a unique instance of the action. These clips collectively form a collection of HOI state transitions that represent the interaction skill." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 152, + 549, + 164 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 152, + 549, + 164 + ], + "spans": [ + { + "bbox": [ + 313, + 152, + 549, + 164 + ], + "type": "text", + "content": "3.3. Learning Interaction Skills by HOI Imitation" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 169, + 553, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 169, + 553, + 264 + ], + "spans": [ + { + "bbox": [ + 313, + 169, + 553, + 264 + ], + "type": "text", + "content": "Considering an interaction skill defined by a set of reference HOI state transitions, if a humanoid can manipulate objects such that its HOI state transitions closely resemble the reference, we consider the humanoid to have successfully learned the interaction skill. Based on this concept, we propose a method for learning interaction skills by imitating HOI state transitions through RL, which we call SkillMimic. Fig. 2 illustrates the concept of SkillMimic." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 265, + 554, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 265, + 554, + 396 + ], + "spans": [ + { + "bbox": [ + 313, + 265, + 554, + 396 + ], + "type": "text", + "content": "In contrast to Behavior Clone [3, 5, 11] methods which rely on precise state-action data, SkillMimic learns actions through RL using state-only trajectories, demonstrating better tolerance to data noise and higher data efficiency. Unlike previous RL methods [33, 34, 50] that require manually designed rewards for each interaction skill, SkillMimic is fully data-driven, skill-agnostic, and scalable, making it capable of learning a wide range of basketball skills within a unified solution. Fig. 6 shows diverse basketball skills acquired using SkillMimic. Fig. 7 illustrates how the performance of the pickup skill improves as the reference data increases." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 402, + 416, + 414 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 402, + 416, + 414 + ], + "spans": [ + { + "bbox": [ + 313, + 402, + 416, + 414 + ], + "type": "text", + "content": "3.3.1. Training Pipeline" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 417, + 554, + 596 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 417, + 554, + 596 + ], + "spans": [ + { + "bbox": [ + 313, + 417, + 554, + 596 + ], + "type": "text", + "content": "Fig. 3 (b) shows the training pipeline of SkillMimic. Given an HOI dataset with diverse skill-labeled clips (Fig. 3 (a)), SkillMimic trains an Interaction Skill (IS) policy by randomly selecting a frame from a clip corresponding to a chosen skill to initialize the humanoid and object states. The state " + }, + { + "bbox": [ + 313, + 417, + 554, + 596 + ], + "type": "inline_equation", + "content": "s_t" + }, + { + "bbox": [ + 313, + 417, + 554, + 596 + ], + "type": "text", + "content": " (Sec. 3.3.2) and skill label " + }, + { + "bbox": [ + 313, + 417, + 554, + 596 + ], + "type": "inline_equation", + "content": "c_j" + }, + { + "bbox": [ + 313, + 417, + 554, + 596 + ], + "type": "text", + "content": " are input into the IS policy, which predicts actions (Sec. 3.3.3) that are then simulated to produce the next state " + }, + { + "bbox": [ + 313, + 417, + 554, + 596 + ], + "type": "inline_equation", + "content": "s_{t+1}" + }, + { + "bbox": [ + 313, + 417, + 554, + 596 + ], + "type": "text", + "content": ". A unified HOI imitation reward (Sec. 3.4) is designed to measure the consistency between the simulated state transitions and the reference transitions. When the humanoid falls or reaches the maximum simulation length, the environment resets, and the process repeats. After training converges, the IS policy enables the humanoid to execute interaction skills that closely mirror the reference demonstrations." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 603, + 417, + 613 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 603, + 417, + 613 + ], + "spans": [ + { + "bbox": [ + 313, + 603, + 417, + 613 + ], + "type": "text", + "content": "3.3.2. HOI Observation" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 617, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 554, + 713 + ], + "type": "text", + "content": "The state observed by the IS policy is not under direct supervision and may theoretically encompass any information that is available in the simulation environment. Importantly, this information does not necessarily have to be included in the reference HOI data. Similar to prior arts [34], we transform all coordinates into the root local coordinate of the humanoid, which aligns the data distribution and benefits the generalization performance. For the humanoid, we observe" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17542" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 82, + 68, + 529, + 281 + ], + "blocks": [ + { + "bbox": [ + 82, + 68, + 529, + 281 + ], + "lines": [ + { + "bbox": [ + 82, + 68, + 529, + 281 + ], + "spans": [ + { + "bbox": [ + 82, + 68, + 529, + 281 + ], + "type": "image", + "image_path": "108388a6dd10c8e77ca2f4a38f92ff9a4678079f2b4e4083d446233a967213f4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 287, + 555, + 355 + ], + "lines": [ + { + "bbox": [ + 54, + 287, + 555, + 355 + ], + "spans": [ + { + "bbox": [ + 54, + 287, + 555, + 355 + ], + "type": "text", + "content": "Figure 3. Our system consists of three parts. (a) First, we capture real-world basketball skills to create a large Human-Object Interaction (HOI) motion dataset. (b) Second, we train an Interaction Skill (IS) policy to learn interaction skills by imitating the corresponding HOI data through reinforcement learning. Specifically, the IS policy takes as input the HOI state " + }, + { + "bbox": [ + 54, + 287, + 555, + 355 + ], + "type": "inline_equation", + "content": "\\boldsymbol{s}_t" + }, + { + "bbox": [ + 54, + 287, + 555, + 355 + ], + "type": "text", + "content": " and skill label " + }, + { + "bbox": [ + 54, + 287, + 555, + 355 + ], + "type": "inline_equation", + "content": "\\boldsymbol{c}_j" + }, + { + "bbox": [ + 54, + 287, + 555, + 355 + ], + "type": "text", + "content": " and predicts the action " + }, + { + "bbox": [ + 54, + 287, + 555, + 355 + ], + "type": "inline_equation", + "content": "\\boldsymbol{a}_t" + }, + { + "bbox": [ + 54, + 287, + 555, + 355 + ], + "type": "text", + "content": ". The new state " + }, + { + "bbox": [ + 54, + 287, + 555, + 355 + ], + "type": "inline_equation", + "content": "\\boldsymbol{s}_{t+1}" + }, + { + "bbox": [ + 54, + 287, + 555, + 355 + ], + "type": "text", + "content": " is calculated by the simulator. A unified HOI imitation reward is designed to imitate diverse HOI state transitions. (c) The third part involves training a High-Level Controller (HLC) to reuse the learned interaction skills for complex tasks. The HLC takes as input " + }, + { + "bbox": [ + 54, + 287, + 555, + 355 + ], + "type": "inline_equation", + "content": "\\boldsymbol{s}_t" + }, + { + "bbox": [ + 54, + 287, + 555, + 355 + ], + "type": "text", + "content": " and extra task observations " + }, + { + "bbox": [ + 54, + 287, + 555, + 355 + ], + "type": "inline_equation", + "content": "\\boldsymbol{h}_t" + }, + { + "bbox": [ + 54, + 287, + 555, + 355 + ], + "type": "text", + "content": ", e.g., the basket position, and predicts the skill label " + }, + { + "bbox": [ + 54, + 287, + 555, + 355 + ], + "type": "inline_equation", + "content": "\\boldsymbol{c}_t" + }, + { + "bbox": [ + 54, + 287, + 555, + 355 + ], + "type": "text", + "content": " to drive a pre-trained IS policy." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 375, + 294, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 375, + 294, + 473 + ], + "spans": [ + { + "bbox": [ + 54, + 375, + 294, + 473 + ], + "type": "text", + "content": "its global root height, local body position, rotation, position velocity, and angular velocity. These representations form the humanoid proprioception " + }, + { + "bbox": [ + 54, + 375, + 294, + 473 + ], + "type": "inline_equation", + "content": "\\pmb{o}_t^{prop}" + }, + { + "bbox": [ + 54, + 375, + 294, + 473 + ], + "type": "text", + "content": ". In addition, we detect net contact forces " + }, + { + "bbox": [ + 54, + 375, + 294, + 473 + ], + "type": "inline_equation", + "content": "\\pmb{o}_t^f" + }, + { + "bbox": [ + 54, + 375, + 294, + 473 + ], + "type": "text", + "content": " for all fingertips, which helps to sense contact and accelerate training. For the object, we observe its local position, rotation, velocity, and angular velocity, which form the object observation " + }, + { + "bbox": [ + 54, + 375, + 294, + 473 + ], + "type": "inline_equation", + "content": "\\pmb{o}_t^{obj}" + }, + { + "bbox": [ + 54, + 375, + 294, + 473 + ], + "type": "text", + "content": ". Finally, the state perceived by the IS policy is" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 126, + 481, + 295, + 497 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 126, + 481, + 295, + 497 + ], + "spans": [ + { + "bbox": [ + 126, + 481, + 295, + 497 + ], + "type": "interline_equation", + "content": "\\boldsymbol {s} _ {t} = \\left\\{\\boldsymbol {o} _ {t} ^ {\\text {p r o p}}, \\boldsymbol {o} _ {t} ^ {f}, \\boldsymbol {o} _ {t} ^ {\\text {o b j}} \\right\\}. \\tag {1}", + "image_path": "f30cbc4bd3ca5dab8b91a2a021efb9907fa6eea5ad13c2e6e5290911e66c3d2d.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 506, + 296, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 506, + 296, + 543 + ], + "spans": [ + { + "bbox": [ + 55, + 506, + 296, + 543 + ], + "type": "text", + "content": "Skill labels " + }, + { + "bbox": [ + 55, + 506, + 296, + 543 + ], + "type": "inline_equation", + "content": "c_{j}" + }, + { + "bbox": [ + 55, + 506, + 296, + 543 + ], + "type": "text", + "content": " are also used as inputs to the IS policy to differentiate different skills. We use one-hot encodings to represent the skill label." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 550, + 171, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 550, + 171, + 562 + ], + "spans": [ + { + "bbox": [ + 55, + 550, + 171, + 562 + ], + "type": "text", + "content": "3.3.3. IS Policy and Action" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 566, + 296, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 566, + 296, + 662 + ], + "spans": [ + { + "bbox": [ + 55, + 566, + 296, + 662 + ], + "type": "text", + "content": "The policy output is modeled as a Gaussian distribution with dimensions equal to the DOF number of the humanoid robot, featuring constant variance. The mean is modeled by a three-layer MLP consisting of [1024, 512, 512] units with ReLU [30] activations. We use the action " + }, + { + "bbox": [ + 55, + 566, + 296, + 662 + ], + "type": "inline_equation", + "content": "\\boldsymbol{a}_t" + }, + { + "bbox": [ + 55, + 566, + 296, + 662 + ], + "type": "text", + "content": " sampled from the policy as the target joint rotations for a full set of PD controllers. The PD controllers adjust and output the joint torques to reach the target rotations." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 671, + 223, + 683 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 671, + 223, + 683 + ], + "spans": [ + { + "bbox": [ + 55, + 671, + 223, + 683 + ], + "type": "text", + "content": "3.4. Unified HOI Imitation Reward" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "This HOI imitation reward function is pivotal as it must not only accurately gauge the efficacy of interaction skill ac" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 316, + 374, + 553, + 453 + ], + "blocks": [ + { + "bbox": [ + 316, + 374, + 553, + 453 + ], + "lines": [ + { + "bbox": [ + 316, + 374, + 553, + 453 + ], + "spans": [ + { + "bbox": [ + 316, + 374, + 553, + 453 + ], + "type": "image", + "image_path": "66fb65f6c000f579bae1ceb1628c22533205715568efee55d80bf05c548e745a.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 462, + 555, + 552 + ], + "lines": [ + { + "bbox": [ + 313, + 462, + 555, + 552 + ], + "spans": [ + { + "bbox": [ + 313, + 462, + 555, + 552 + ], + "type": "text", + "content": "Figure 4. We propose the Contact Graph (CG) to model general contacts within an explicitly defined scene. The node stores a binary value that denotes whether it contacts other nodes. Each edge stores a binary value indicating whether the two connected nodes are in contact. The node definition is unified for a certain scene and shared between diverse interactive skills. For example, we define three nodes: hands, hands-exclusive body, and ball, to form a simple CG to model contacts for diverse basketball skills." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 577, + 553, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 577, + 553, + 601 + ], + "spans": [ + { + "bbox": [ + 313, + 577, + 553, + 601 + ], + "type": "text", + "content": "quisition but also avoid the incorporation of skill-specific settings to maintain robust cross-skill generalizability." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 603, + 555, + 699 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 603, + 555, + 699 + ], + "spans": [ + { + "bbox": [ + 313, + 603, + 555, + 699 + ], + "type": "text", + "content": "Classic imitation rewards often focus solely on imitating locomotions [32], lacking an established baseline for HOI imitation. Although adversarial imitation rewards [33, 34] are effective in locomotion tasks, we find that they perform poorly in HOI imitation. This may be because adversarial imitation rewards are relatively coarse-grained, making it difficult for them to provide the precise guidance needed for accurate interaction learning." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 326, + 700, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 700, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 326, + 700, + 553, + 714 + ], + "type": "text", + "content": "The body kinematics imitation reward " + }, + { + "bbox": [ + 326, + 700, + 553, + 714 + ], + "type": "inline_equation", + "content": "r_t^b" + }, + { + "bbox": [ + 326, + 700, + 553, + 714 + ], + "type": "text", + "content": " encourages the" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17543" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 324 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 324 + ], + "type": "text", + "content": "alignment of body movements by considering joint position, rotation, and their velocities. Similarly, the object kinematics imitation reward " + }, + { + "bbox": [ + 55, + 72, + 296, + 324 + ], + "type": "inline_equation", + "content": "r_t^o" + }, + { + "bbox": [ + 55, + 72, + 296, + 324 + ], + "type": "text", + "content": " ensures consistent object movements with the reference. However, solely relying on these two rewards often leads to unstable learning and unnatural motions. We incorporate a relative motion reward " + }, + { + "bbox": [ + 55, + 72, + 296, + 324 + ], + "type": "inline_equation", + "content": "r_t^{rel}" + }, + { + "bbox": [ + 55, + 72, + 296, + 324 + ], + "type": "text", + "content": " to highlight the relative relationships in interactions, which is crucial for interaction naturalness. Recognizing the interlocking relationship between interaction and contact, we propose a straightforward and universal contact modeling approach called Contact Graph (CG). A corresponding CG imitation reward " + }, + { + "bbox": [ + 55, + 72, + 296, + 324 + ], + "type": "inline_equation", + "content": "r_t^{cg}" + }, + { + "bbox": [ + 55, + 72, + 296, + 324 + ], + "type": "text", + "content": " is designed to enhance the precision of interaction imitation. Additionally, we propose an adaptive velocity regularization term " + }, + { + "bbox": [ + 55, + 72, + 296, + 324 + ], + "type": "inline_equation", + "content": "r_t^{reg}" + }, + { + "bbox": [ + 55, + 72, + 296, + 324 + ], + "type": "text", + "content": " to suppress high-frequency jitters. These sub-rewards are multiplied rather than added to encourage balanced learning, thereby avoiding local optima, as justified in Tab. 1. The combination of these innovations forms a unified imitation reward, enabling the learning of diverse basketball skills in BallPlay-M with the same configuration. To conclude, the complete HOI imitation reward is" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 110, + 332, + 295, + 347 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 332, + 295, + 347 + ], + "spans": [ + { + "bbox": [ + 110, + 332, + 295, + 347 + ], + "type": "interline_equation", + "content": "r _ {t} = r _ {t} ^ {b} * r _ {t} ^ {o} * r _ {t} ^ {r e l} * r _ {t} ^ {r e g} * r _ {t} ^ {c g}. \\tag {2}", + "image_path": "2c8ab73fb5669471c7f977d12d59e3f166d5db787de1a268ca2ab8065cf05a0e.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 357, + 296, + 393 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 357, + 296, + 393 + ], + "spans": [ + { + "bbox": [ + 55, + 357, + 296, + 393 + ], + "type": "text", + "content": "Next, we will discuss the methodology involved in crafting the CG reward " + }, + { + "bbox": [ + 55, + 357, + 296, + 393 + ], + "type": "inline_equation", + "content": "r_t^{cg}" + }, + { + "bbox": [ + 55, + 357, + 296, + 393 + ], + "type": "text", + "content": ". Details about the kinematic imitation reward " + }, + { + "bbox": [ + 55, + 357, + 296, + 393 + ], + "type": "inline_equation", + "content": "r_t^b * r_t^o * r_t^{rel} * r_t^{reg}" + }, + { + "bbox": [ + 55, + 357, + 296, + 393 + ], + "type": "text", + "content": " are in the appendix." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 399, + 149, + 411 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 399, + 149, + 411 + ], + "spans": [ + { + "bbox": [ + 55, + 399, + 149, + 411 + ], + "type": "text", + "content": "3.4.1. Contact Graph" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 414, + 296, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 414, + 296, + 474 + ], + "spans": [ + { + "bbox": [ + 55, + 414, + 296, + 474 + ], + "type": "text", + "content": "We observe that existing imitation rewards [33, 34, 52] only measure kinematic properties and are insufficient to measure precise contacts. To tackle this problem, we propose the Contact Graph (CG) to model the contact in general interactions and design a CG imitation reward." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 474, + 296, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 474, + 296, + 617 + ], + "spans": [ + { + "bbox": [ + 55, + 474, + 296, + 617 + ], + "type": "text", + "content": "As illustrated in Fig. 4, the CG is a complete graph where every pair of distinct nodes is connected by a unique edge, defined as " + }, + { + "bbox": [ + 55, + 474, + 296, + 617 + ], + "type": "inline_equation", + "content": "\\mathcal{G} = \\{\\mathcal{V},\\mathcal{E}\\}" + }, + { + "bbox": [ + 55, + 474, + 296, + 617 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 474, + 296, + 617 + ], + "type": "inline_equation", + "content": "\\mathcal{V}\\in \\{0,1\\} ^k" + }, + { + "bbox": [ + 55, + 474, + 296, + 617 + ], + "type": "text", + "content": " is the set of " + }, + { + "bbox": [ + 55, + 474, + 296, + 617 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 474, + 296, + 617 + ], + "type": "text", + "content": " nodes and " + }, + { + "bbox": [ + 55, + 474, + 296, + 617 + ], + "type": "inline_equation", + "content": "\\mathcal{E}\\in \\{0,1\\}^{k(k - 1) / 2}" + }, + { + "bbox": [ + 55, + 474, + 296, + 617 + ], + "type": "text", + "content": " is the set of edges. A CG node stores a binary value indicating whether it contacts other nodes. Each CG edge stores a binary label that denotes the contact between two nodes, where 1 represents contact, and 0 means no contact. CG node and edge values are calculated frame by frame, explicitly describing the mutual contact relationship between CG nodes at different moments. Contacts inside a node are not considered. We use the edge set " + }, + { + "bbox": [ + 55, + 474, + 296, + 617 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 55, + 474, + 296, + 617 + ], + "type": "text", + "content": " for imitation learning." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 618, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 296, + 713 + ], + "type": "text", + "content": "The definitions of nodes are flexible. For example, a node can be a single part (e.g., a fingertip), or an aggregation of multiple parts (e.g., the whole left hand). Hinged objects can also be broken up into finer nodes [9, 13, 51]. The definition of CG is unified for a certain scene and shared between diverse interaction skills. For example, in basketball scenarios, we can aggregate two hands as a node, the rest of the bodies as a node, and the ball as a node, as illustrated in" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 316, + 72, + 347, + 118 + ], + "blocks": [ + { + "bbox": [ + 316, + 72, + 347, + 118 + ], + "lines": [ + { + "bbox": [ + 316, + 72, + 347, + 118 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 347, + 118 + ], + "type": "image", + "image_path": "ad1feb3457ae5e2008a94cf8a2b172e78d088939e7c72ec242297faba93355d5.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 327, + 118, + 345, + 126 + ], + "lines": [ + { + "bbox": [ + 327, + 118, + 345, + 126 + ], + "spans": [ + { + "bbox": [ + 327, + 118, + 345, + 126 + ], + "type": "text", + "content": "(a) Ref" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 359, + 72, + 384, + 118 + ], + "blocks": [ + { + "bbox": [ + 359, + 72, + 384, + 118 + ], + "lines": [ + { + "bbox": [ + 359, + 72, + 384, + 118 + ], + "spans": [ + { + "bbox": [ + 359, + 72, + 384, + 118 + ], + "type": "image", + "image_path": "72bf60387a71d1f6b94b2c64e199716d38b853e3d754056477b758776e778f45.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 359, + 118, + 391, + 126 + ], + "lines": [ + { + "bbox": [ + 359, + 118, + 391, + 126 + ], + "spans": [ + { + "bbox": [ + 359, + 118, + 391, + 126 + ], + "type": "text", + "content": "(b) w/o CGR" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 399, + 72, + 426, + 118 + ], + "blocks": [ + { + "bbox": [ + 399, + 72, + 426, + 118 + ], + "lines": [ + { + "bbox": [ + 399, + 72, + 426, + 118 + ], + "spans": [ + { + "bbox": [ + 399, + 72, + 426, + 118 + ], + "type": "image", + "image_path": "689532f55a0f2f189061e0ed895870a83cb141124050d482b950fa95df31df79.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 402, + 118, + 429, + 125 + ], + "lines": [ + { + "bbox": [ + 402, + 118, + 429, + 125 + ], + "spans": [ + { + "bbox": [ + 402, + 118, + 429, + 125 + ], + "type": "text", + "content": "(c) w/ CGR" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 441, + 72, + 462, + 118 + ], + "blocks": [ + { + "bbox": [ + 441, + 72, + 462, + 118 + ], + "lines": [ + { + "bbox": [ + 441, + 72, + 462, + 118 + ], + "spans": [ + { + "bbox": [ + 441, + 72, + 462, + 118 + ], + "type": "image", + "image_path": "3c45c633681c04d10ae1ed42321ccb77759132629c13643a148b67e9387153c9.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 442, + 118, + 460, + 125 + ], + "lines": [ + { + "bbox": [ + 442, + 118, + 460, + 125 + ], + "spans": [ + { + "bbox": [ + 442, + 118, + 460, + 125 + ], + "type": "text", + "content": "(d) Ref" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 485, + 72, + 507, + 118 + ], + "blocks": [ + { + "bbox": [ + 485, + 72, + 507, + 118 + ], + "lines": [ + { + "bbox": [ + 485, + 72, + 507, + 118 + ], + "spans": [ + { + "bbox": [ + 485, + 72, + 507, + 118 + ], + "type": "image", + "image_path": "5f765883bad63b19c40eebe16fabc10a5d04c7ad37ebfabb598c519ff5e2dec3.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 477, + 118, + 509, + 125 + ], + "lines": [ + { + "bbox": [ + 477, + 118, + 509, + 125 + ], + "spans": [ + { + "bbox": [ + 477, + 118, + 509, + 125 + ], + "type": "text", + "content": "(e) w/o CGR" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 523, + 72, + 544, + 118 + ], + "blocks": [ + { + "bbox": [ + 523, + 72, + 544, + 118 + ], + "lines": [ + { + "bbox": [ + 523, + 72, + 544, + 118 + ], + "spans": [ + { + "bbox": [ + 523, + 72, + 544, + 118 + ], + "type": "image", + "image_path": "8721a308430ce184e4e4febc7fc4e7bb0cacbac0cae4ce642e1c47a77909e750.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 522, + 118, + 548, + 125 + ], + "lines": [ + { + "bbox": [ + 522, + 118, + 548, + 125 + ], + "spans": [ + { + "bbox": [ + 522, + 118, + 548, + 125 + ], + "type": "text", + "content": "(f) w/ CGR" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 316, + 132, + 350, + 178 + ], + "blocks": [ + { + "bbox": [ + 316, + 132, + 350, + 178 + ], + "lines": [ + { + "bbox": [ + 316, + 132, + 350, + 178 + ], + "spans": [ + { + "bbox": [ + 316, + 132, + 350, + 178 + ], + "type": "image", + "image_path": "03c397008252e041c9758785dc483f83af0ecc0a8d92abb5af498507e4a4bfa4.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 327, + 178, + 345, + 186 + ], + "lines": [ + { + "bbox": [ + 327, + 178, + 345, + 186 + ], + "spans": [ + { + "bbox": [ + 327, + 178, + 345, + 186 + ], + "type": "text", + "content": "(g) Ref" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 194, + 555, + 260 + ], + "lines": [ + { + "bbox": [ + 313, + 194, + 555, + 260 + ], + "spans": [ + { + "bbox": [ + 313, + 194, + 555, + 260 + ], + "type": "text", + "content": "Figure 5. The HOI imitation falls into kinematic local-optimal solutions without Contact Graph Reward (CGR): (b) use the head to help control the ball; (e) use the wrist to contact the ball; (h) fail to catch the object; (k) support the table to keep balance. In comparison, the guidance of CGR effectively yields precise interactions, as shown in (c, f, i, l)." + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 281, + 555, + 306 + ], + "lines": [ + { + "bbox": [ + 313, + 281, + 555, + 306 + ], + "spans": [ + { + "bbox": [ + 313, + 281, + 555, + 306 + ], + "type": "text", + "content": "Fig. 4. This simple CG is effective enough for learning all the basketball skills covered in BallPlay-M." + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 357, + 132, + 390, + 178 + ], + "blocks": [ + { + "bbox": [ + 357, + 132, + 390, + 178 + ], + "lines": [ + { + "bbox": [ + 357, + 132, + 390, + 178 + ], + "spans": [ + { + "bbox": [ + 357, + 132, + 390, + 178 + ], + "type": "image", + "image_path": "ba1cbea2dd01c3597180c785e80b43508d8f858293d76242eaf4fcc318ea72f8.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 361, + 178, + 392, + 186 + ], + "lines": [ + { + "bbox": [ + 361, + 178, + 392, + 186 + ], + "spans": [ + { + "bbox": [ + 361, + 178, + 392, + 186 + ], + "type": "text", + "content": "(h) w/o CGR" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 394, + 133, + 429, + 178 + ], + "blocks": [ + { + "bbox": [ + 394, + 133, + 429, + 178 + ], + "lines": [ + { + "bbox": [ + 394, + 133, + 429, + 178 + ], + "spans": [ + { + "bbox": [ + 394, + 133, + 429, + 178 + ], + "type": "image", + "image_path": "c4ebc0b7cd1f8196316cdd8d15675b741446c7d18c237f4e8c29892cd7cd4ecd.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 402, + 178, + 429, + 186 + ], + "lines": [ + { + "bbox": [ + 402, + 178, + 429, + 186 + ], + "spans": [ + { + "bbox": [ + 402, + 178, + 429, + 186 + ], + "type": "text", + "content": "(i) w/ CGR" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 444, + 130, + 465, + 178 + ], + "blocks": [ + { + "bbox": [ + 444, + 130, + 465, + 178 + ], + "lines": [ + { + "bbox": [ + 444, + 130, + 465, + 178 + ], + "spans": [ + { + "bbox": [ + 444, + 130, + 465, + 178 + ], + "type": "image", + "image_path": "6b7a535dbe9ba24512c66e8dc1fb7755c427e97164c22835565b00d4b565141f.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 445, + 178, + 460, + 186 + ], + "lines": [ + { + "bbox": [ + 445, + 178, + 460, + 186 + ], + "spans": [ + { + "bbox": [ + 445, + 178, + 460, + 186 + ], + "type": "text", + "content": "(j) Ref." + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 488, + 132, + 508, + 178 + ], + "blocks": [ + { + "bbox": [ + 488, + 132, + 508, + 178 + ], + "lines": [ + { + "bbox": [ + 488, + 132, + 508, + 178 + ], + "spans": [ + { + "bbox": [ + 488, + 132, + 508, + 178 + ], + "type": "image", + "image_path": "ae9a51cf90bab51e9ac53f3a9514f7000a536b0cdd7ad88a6ae4e65e8877948f.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 468, + 178, + 521, + 186 + ], + "lines": [ + { + "bbox": [ + 468, + 178, + 521, + 186 + ], + "spans": [ + { + "bbox": [ + 468, + 178, + 521, + 186 + ], + "type": "text", + "content": "(k) w/o CGR on table" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 528, + 131, + 548, + 178 + ], + "blocks": [ + { + "bbox": [ + 528, + 131, + 548, + 178 + ], + "lines": [ + { + "bbox": [ + 528, + 131, + 548, + 178 + ], + "spans": [ + { + "bbox": [ + 528, + 131, + 548, + 178 + ], + "type": "image", + "image_path": "82aca685518671e1e76ccafc72f389ac4cc54cac48dde76c3c143520f3a37d32.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 526, + 178, + 553, + 185 + ], + "lines": [ + { + "bbox": [ + 526, + 178, + 553, + 185 + ], + "spans": [ + { + "bbox": [ + 526, + 178, + 553, + 185 + ], + "type": "text", + "content": "(I) w/CGR" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + } + ], + "index": 29 + }, + { + "bbox": [ + 313, + 311, + 444, + 322 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 311, + 444, + 322 + ], + "spans": [ + { + "bbox": [ + 313, + 311, + 444, + 322 + ], + "type": "text", + "content": "3.4.2. Contact Graph Reward" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 313, + 325, + 555, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 325, + 555, + 517 + ], + "spans": [ + { + "bbox": [ + 313, + 325, + 555, + 517 + ], + "type": "text", + "content": "Kinematic imitation rewards [32, 33, 52] fall short in measuring contacts and thus yield poor performance in HOI imitation. For example, during the toss skill training (Fig. 5 (a)), the contact between the right hand and the ball usually cause the ball to drop, making the kinematic rewards drastically smaller. In this case, the policy may learn a local-optimal solution that uses the head and left hands to stabilize the ball, as demonstrated in Fig. 5 (b). To tackle this problem, we propose the Contact Graph Reward (CGR) as a critical complement of kinematic imitation rewards to learn precise contact imitation. The CG error is defined as " + }, + { + "bbox": [ + 313, + 325, + 555, + 517 + ], + "type": "inline_equation", + "content": "e_t^{cg} = |s_t^{cg} - \\hat{s}_t^{cg}|" + }, + { + "bbox": [ + 313, + 325, + 555, + 517 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 325, + 555, + 517 + ], + "type": "inline_equation", + "content": "|\\cdot|" + }, + { + "bbox": [ + 313, + 325, + 555, + 517 + ], + "type": "text", + "content": " calculates element-wise absolute value, " + }, + { + "bbox": [ + 313, + 325, + 555, + 517 + ], + "type": "inline_equation", + "content": "s_t^{cg}" + }, + { + "bbox": [ + 313, + 325, + 555, + 517 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 325, + 555, + 517 + ], + "type": "inline_equation", + "content": "\\hat{s}_t^{cg}" + }, + { + "bbox": [ + 313, + 325, + 555, + 517 + ], + "type": "text", + "content": " denotes the simulated and reference CG state respectively, which is the edge set " + }, + { + "bbox": [ + 313, + 325, + 555, + 517 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_t \\in \\{0,1\\}^{k(k-1)/2}" + }, + { + "bbox": [ + 313, + 325, + 555, + 517 + ], + "type": "text", + "content": ". The CGR is measured by CG error, with independent weights on different elements:" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 364, + 525, + 554, + 559 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 364, + 525, + 554, + 559 + ], + "spans": [ + { + "bbox": [ + 364, + 525, + 554, + 559 + ], + "type": "interline_equation", + "content": "r _ {t} ^ {c g} = \\exp \\left(- \\sum_ {j = 1} ^ {J} \\boldsymbol {\\lambda} ^ {c g} [ j ] * \\boldsymbol {e} _ {t} ^ {c g} [ j ]\\right), \\tag {3}", + "image_path": "874d24a1b5146f7afb561b51a67bf6120ed0864fdbc80f735f405614c154c27c.jpg" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 313, + 567, + 554, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 567, + 554, + 641 + ], + "spans": [ + { + "bbox": [ + 313, + 567, + 554, + 641 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 567, + 554, + 641 + ], + "type": "inline_equation", + "content": "e_t^{cg}[j]" + }, + { + "bbox": [ + 313, + 567, + 554, + 641 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 313, + 567, + 554, + 641 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 567, + 554, + 641 + ], + "type": "text", + "content": "th element of " + }, + { + "bbox": [ + 313, + 567, + 554, + 641 + ], + "type": "inline_equation", + "content": "e_t^{cg} \\in \\{0,1\\}^J" + }, + { + "bbox": [ + 313, + 567, + 554, + 641 + ], + "type": "text", + "content": ", representing a contact error; " + }, + { + "bbox": [ + 313, + 567, + 554, + 641 + ], + "type": "inline_equation", + "content": "\\lambda^{cg}[j]" + }, + { + "bbox": [ + 313, + 567, + 554, + 641 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 313, + 567, + 554, + 641 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 567, + 554, + 641 + ], + "type": "text", + "content": "th element of " + }, + { + "bbox": [ + 313, + 567, + 554, + 641 + ], + "type": "inline_equation", + "content": "\\lambda^{cg} \\in \\mathbb{R}^J" + }, + { + "bbox": [ + 313, + 567, + 554, + 641 + ], + "type": "text", + "content": ", a hyperparameter controls the sensitivity of a contact. Fig. 5 and Tab. 1 justify the effectiveness of CGR, where experiments without CGR show inaccurate interactions while using CGR effectively eliminates such problems." + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 313, + 647, + 554, + 661 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 647, + 554, + 661 + ], + "spans": [ + { + "bbox": [ + 313, + 647, + 554, + 661 + ], + "type": "text", + "content": "3.5. Reusing Interaction Skills for High-level Tasks" + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "type": "text", + "content": "In this section, we propose training a High-Level Controller (HLC) to reuse the interaction skills acquired through SkillMimic to accomplish high-level tasks. As depicted in Fig. 3 (c), the HLC takes as input the current HOI state " + }, + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "type": "inline_equation", + "content": "s_t" + } + ] + } + ], + "index": 38 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "17544" + } + ] + } + ], + "index": 39 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 62, + 70, + 160, + 125 + ], + "blocks": [ + { + "bbox": [ + 62, + 70, + 160, + 125 + ], + "lines": [ + { + "bbox": [ + 62, + 70, + 160, + 125 + ], + "spans": [ + { + "bbox": [ + 62, + 70, + 160, + 125 + ], + "type": "image", + "image_path": "04724ae4c94bf9781dd4d256fe5f3b9446d27f5dd1cc13e89bc6d3688cd061de.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 160, + 70, + 256, + 125 + ], + "blocks": [ + { + "bbox": [ + 160, + 70, + 256, + 125 + ], + "lines": [ + { + "bbox": [ + 160, + 70, + 256, + 125 + ], + "spans": [ + { + "bbox": [ + 160, + 70, + 256, + 125 + ], + "type": "image", + "image_path": "81f434658e633d076d8eaa62b15976236335bd3e486e7e860ffde26067a8520c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 257, + 70, + 353, + 125 + ], + "blocks": [ + { + "bbox": [ + 257, + 70, + 353, + 125 + ], + "lines": [ + { + "bbox": [ + 257, + 70, + 353, + 125 + ], + "spans": [ + { + "bbox": [ + 257, + 70, + 353, + 125 + ], + "type": "image", + "image_path": "40c1b9a8efcc35a2af225b7f9d86bf0e8d7eb7795b04815c0fc713b101174e57.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 354, + 70, + 450, + 125 + ], + "blocks": [ + { + "bbox": [ + 354, + 70, + 450, + 125 + ], + "lines": [ + { + "bbox": [ + 354, + 70, + 450, + 125 + ], + "spans": [ + { + "bbox": [ + 354, + 70, + 450, + 125 + ], + "type": "image", + "image_path": "a5afb105d44ad3196fd38b48c4c7a583220383258d035c1461dd6032cb3240e3.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 450, + 70, + 548, + 125 + ], + "blocks": [ + { + "bbox": [ + 450, + 70, + 548, + 125 + ], + "lines": [ + { + "bbox": [ + 450, + 70, + 548, + 125 + ], + "spans": [ + { + "bbox": [ + 450, + 70, + 548, + 125 + ], + "type": "image", + "image_path": "acf43d35437bb8d4e8e36e9e62f0a96186c01bc3cf31b8cec8d97e2b9b6d2317.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 62, + 137, + 160, + 191 + ], + "blocks": [ + { + "bbox": [ + 95, + 127, + 128, + 137 + ], + "lines": [ + { + "bbox": [ + 95, + 127, + 128, + 137 + ], + "spans": [ + { + "bbox": [ + 95, + 127, + 128, + 137 + ], + "type": "text", + "content": "(a) Layup" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 62, + 137, + 160, + 191 + ], + "lines": [ + { + "bbox": [ + 62, + 137, + 160, + 191 + ], + "spans": [ + { + "bbox": [ + 62, + 137, + 160, + 191 + ], + "type": "image", + "image_path": "b93370c0266fe545249025ae5b28597104a685b93558a0fb513c760796e91828.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 88, + 194, + 134, + 203 + ], + "lines": [ + { + "bbox": [ + 88, + 194, + 134, + 203 + ], + "spans": [ + { + "bbox": [ + 88, + 194, + 134, + 203 + ], + "type": "text", + "content": "(f) Jump Shot" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 160, + 137, + 256, + 191 + ], + "blocks": [ + { + "bbox": [ + 171, + 127, + 244, + 137 + ], + "lines": [ + { + "bbox": [ + 171, + 127, + 244, + 137 + ], + "spans": [ + { + "bbox": [ + 171, + 127, + 244, + 137 + ], + "type": "text", + "content": "(b) Turnaround Layup" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 160, + 137, + 256, + 191 + ], + "lines": [ + { + "bbox": [ + 160, + 137, + 256, + 191 + ], + "spans": [ + { + "bbox": [ + 160, + 137, + 256, + 191 + ], + "type": "image", + "image_path": "e3ca519f182dd1bbf68eee41d2ac166a9d0a7eff56f549063c723d74b35cf357.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 187, + 194, + 229, + 204 + ], + "lines": [ + { + "bbox": [ + 187, + 194, + 229, + 204 + ], + "spans": [ + { + "bbox": [ + 187, + 194, + 229, + 204 + ], + "type": "text", + "content": "(g) Rebound" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 257, + 137, + 353, + 192 + ], + "blocks": [ + { + "bbox": [ + 277, + 127, + 333, + 136 + ], + "lines": [ + { + "bbox": [ + 277, + 127, + 333, + 136 + ], + "spans": [ + { + "bbox": [ + 277, + 127, + 333, + 136 + ], + "type": "text", + "content": "(c) Dribble Right" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 257, + 137, + 353, + 192 + ], + "lines": [ + { + "bbox": [ + 257, + 137, + 353, + 192 + ], + "spans": [ + { + "bbox": [ + 257, + 137, + 353, + 192 + ], + "type": "image", + "image_path": "85843f0257971d7780dba6974a54d2f093ec8741dc12527353844c045d4b8edb.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 291, + 194, + 318, + 203 + ], + "lines": [ + { + "bbox": [ + 291, + 194, + 318, + 203 + ], + "spans": [ + { + "bbox": [ + 291, + 194, + 318, + 203 + ], + "type": "text", + "content": "(h) Pass" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 213, + 555, + 246 + ], + "lines": [ + { + "bbox": [ + 55, + 213, + 555, + 246 + ], + "spans": [ + { + "bbox": [ + 55, + 213, + 555, + 246 + ], + "type": "text", + "content": "Figure 6. Simulated humans exhibit comprehensive basketball skills. SkillMimic can teach humans a wide range of basketball skills using the same configuration in a purely data-driven manner, covering almost all fundamental basketball skills. Keyframes are placed in chronological order from left to right." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 354, + 137, + 449, + 192 + ], + "blocks": [ + { + "bbox": [ + 369, + 127, + 435, + 136 + ], + "lines": [ + { + "bbox": [ + 369, + 127, + 435, + 136 + ], + "spans": [ + { + "bbox": [ + 369, + 127, + 435, + 136 + ], + "type": "text", + "content": "(d) Dribble Forward" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 354, + 137, + 449, + 192 + ], + "lines": [ + { + "bbox": [ + 354, + 137, + 449, + 192 + ], + "spans": [ + { + "bbox": [ + 354, + 137, + 449, + 192 + ], + "type": "image", + "image_path": "df3eda1f4a2381cb6519824711901ee69e5af95c7bf06fc604314ee788742232.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 386, + 194, + 419, + 204 + ], + "lines": [ + { + "bbox": [ + 386, + 194, + 419, + 204 + ], + "spans": [ + { + "bbox": [ + 386, + 194, + 419, + 204 + ], + "type": "text", + "content": "(i) Pickup" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 451, + 137, + 548, + 192 + ], + "blocks": [ + { + "bbox": [ + 472, + 127, + 526, + 136 + ], + "lines": [ + { + "bbox": [ + 472, + 127, + 526, + 136 + ], + "spans": [ + { + "bbox": [ + 472, + 127, + 526, + 136 + ], + "type": "text", + "content": "(e) Dribble Turn" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 451, + 137, + 548, + 192 + ], + "lines": [ + { + "bbox": [ + 451, + 137, + 548, + 192 + ], + "spans": [ + { + "bbox": [ + 451, + 137, + 548, + 192 + ], + "type": "image", + "image_path": "dcd88d4d9142af324ac903e85a3acaa2c66af0927fab20a4dd879e676258c842.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 477, + 194, + 521, + 204 + ], + "lines": [ + { + "bbox": [ + 477, + 194, + 521, + 204 + ], + "spans": [ + { + "bbox": [ + 477, + 194, + 521, + 204 + ], + "type": "text", + "content": "(j) Cross Leg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "bbox": [ + 55, + 267, + 294, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 267, + 294, + 304 + ], + "spans": [ + { + "bbox": [ + 55, + 267, + 294, + 304 + ], + "type": "text", + "content": "and task observation " + }, + { + "bbox": [ + 55, + 267, + 294, + 304 + ], + "type": "inline_equation", + "content": "h_t" + }, + { + "bbox": [ + 55, + 267, + 294, + 304 + ], + "type": "text", + "content": ", and outputs a discrete skill embedding " + }, + { + "bbox": [ + 55, + 267, + 294, + 304 + ], + "type": "inline_equation", + "content": "c_t" + }, + { + "bbox": [ + 55, + 267, + 294, + 304 + ], + "type": "text", + "content": " that serves as inputs to a pre-trained SkillMimic policy, which subsequently generates actions for the humanoid." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 55, + 304, + 295, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 304, + 295, + 411 + ], + "spans": [ + { + "bbox": [ + 55, + 304, + 295, + 411 + ], + "type": "text", + "content": "Compared to previous hierarchical approaches that are based on locomotion priors [34, 50], our method leverages interaction skills as prior, which significantly simplifies the design of task rewards and greatly accelerates training. Moreover, unlike their mandatory hierarchical structure, our approach is motivated by data efficiency - tasks can be alternatively solved through direct HOI imitation when sufficient demonstrations are available, as shown in pickup experiments in Sec. 4.2, Fig. 7, and Fig. 8." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 67, + 411, + 260, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 411, + 260, + 424 + ], + "spans": [ + { + "bbox": [ + 67, + 411, + 260, + 424 + ], + "type": "text", + "content": "We present four representative high-level tasks:" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 55, + 427, + 295, + 510 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 55, + 427, + 295, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 427, + 295, + 450 + ], + "spans": [ + { + "bbox": [ + 55, + 427, + 295, + 450 + ], + "type": "text", + "content": "- Throwing: throw the ball to approach a certain height, grab the rebound, and keep on throwing the ball." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 55, + 451, + 295, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 451, + 295, + 462 + ], + "spans": [ + { + "bbox": [ + 55, + 451, + 295, + 462 + ], + "type": "text", + "content": "- Heading: dribble the ball to approach the target position." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 56, + 462, + 295, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 462, + 295, + 486 + ], + "spans": [ + { + "bbox": [ + 56, + 462, + 295, + 486 + ], + "type": "text", + "content": "- Circling: dribble the ball around the target position following a target radius." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 56, + 487, + 295, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 487, + 295, + 510 + ], + "spans": [ + { + "bbox": [ + 56, + 487, + 295, + 510 + ], + "type": "text", + "content": "- Scoring: dribble the ball toward the basket, timing the layup to score, retrieving the rebound, and repeating." + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 513, + 296, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 513, + 296, + 561 + ], + "spans": [ + { + "bbox": [ + 55, + 513, + 296, + 561 + ], + "type": "text", + "content": "Utilizing interaction skills as prior significantly improves learning efficiency for complex, long-horizon tasks. For example, a decent HLC for the Scoring task only requires 3 hours of training on an NVIDIA RTX 4090 GPU." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 55, + 573, + 137, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 573, + 137, + 586 + ], + "spans": [ + { + "bbox": [ + 55, + 573, + 137, + 586 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "content": "We conduct extensive experiments to evaluate the performance of SkillMimic in learning interaction skills and its application for high-level tasks. Our evaluation consists of two main parts: (1) assessing the capability of SkillMimic in learning interaction skills, including comparisons with existing methods and ablation studies, and (2) demonstrating the effectiveness of reusing acquired interaction skills to learn high-level tasks. Details of the experimental settings and additional experiments on skill switching and robustness are provided in the appendix." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 314, + 266, + 373, + 277 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 266, + 373, + 277 + ], + "spans": [ + { + "bbox": [ + 314, + 266, + 373, + 277 + ], + "type": "text", + "content": "4.1. Metrics" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 313, + 285, + 506, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 285, + 506, + 297 + ], + "spans": [ + { + "bbox": [ + 313, + 285, + 506, + 297 + ], + "type": "text", + "content": "We consider the following quantitative metrics:" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 314, + 299, + 554, + 504 + ], + "type": "list", + "angle": 0, + "index": 38, + "blocks": [ + { + "bbox": [ + 314, + 299, + 554, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 299, + 554, + 382 + ], + "spans": [ + { + "bbox": [ + 314, + 299, + 554, + 382 + ], + "type": "text", + "content": "- Accuracy. The overall accuracy of HOI imitation, abbreviated as Acc., is defined per frame, and deems imitation accurate when the object position and body position errors are both under the thresholds and the contacts are correct. The object threshold is defined as " + }, + { + "bbox": [ + 314, + 299, + 554, + 382 + ], + "type": "inline_equation", + "content": "0.2m" + }, + { + "bbox": [ + 314, + 299, + 554, + 382 + ], + "type": "text", + "content": ". The body threshold is defined as " + }, + { + "bbox": [ + 314, + 299, + 554, + 382 + ], + "type": "inline_equation", + "content": "0.1m" + }, + { + "bbox": [ + 314, + 299, + 554, + 382 + ], + "type": "text", + "content": ". The Acc. is calculated by averaging the values of all frames." + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 314, + 382, + 554, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 382, + 554, + 430 + ], + "spans": [ + { + "bbox": [ + 314, + 382, + 554, + 430 + ], + "type": "text", + "content": "- Position Error. The Mean Per-Joint Position Error (MPJPE) of the body " + }, + { + "bbox": [ + 314, + 382, + 554, + 430 + ], + "type": "inline_equation", + "content": "(E_{\\mathrm{b - mpjpe}})" + }, + { + "bbox": [ + 314, + 382, + 554, + 430 + ], + "type": "text", + "content": " and object " + }, + { + "bbox": [ + 314, + 382, + 554, + 430 + ], + "type": "inline_equation", + "content": "(E_{\\mathrm{o - mpjpe}})" + }, + { + "bbox": [ + 314, + 382, + 554, + 430 + ], + "type": "text", + "content": " is used to evaluate the positional imitation performance (in mm) following [27]." + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 314, + 430, + 553, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 430, + 553, + 467 + ], + "spans": [ + { + "bbox": [ + 314, + 430, + 553, + 467 + ], + "type": "text", + "content": "- Contact Error. The contact error " + }, + { + "bbox": [ + 314, + 430, + 553, + 467 + ], + "type": "inline_equation", + "content": "E_{\\mathrm{cg}}" + }, + { + "bbox": [ + 314, + 430, + 553, + 467 + ], + "type": "text", + "content": ", ranging from 0 to 1, is defined as " + }, + { + "bbox": [ + 314, + 430, + 553, + 467 + ], + "type": "inline_equation", + "content": "\\frac{1}{N}\\sum_{t=1}^{N}\\mathrm{MSE}(s_t^{cg}, \\hat{s}_t^{cg})" + }, + { + "bbox": [ + 314, + 430, + 553, + 467 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 314, + 430, + 553, + 467 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 314, + 430, + 553, + 467 + ], + "type": "text", + "content": " is the total frames of the reference HOI data." + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 314, + 468, + 553, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 468, + 553, + 504 + ], + "spans": [ + { + "bbox": [ + 314, + 468, + 553, + 504 + ], + "type": "text", + "content": "- Success Rate. The execution success rate of an interaction skill. Detailed definitions of success criteria for different skills are provided in the appendix." + } + ] + } + ], + "index": 37 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 515, + 513, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 515, + 513, + 528 + ], + "spans": [ + { + "bbox": [ + 313, + 515, + 513, + 528 + ], + "type": "text", + "content": "4.2. Evaluating Interaction Skill Learning" + } + ] + } + ], + "index": 39 + }, + { + "bbox": [ + 313, + 534, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 555, + 714 + ], + "type": "text", + "content": "Since our approach is the first to learn interaction skills from HOI demonstrations, there are no direct benchmarks for comparison. Therefore, we adapt reward strategies commonly used in locomotion imitation, i.e., DeepMimic [32] and AMP [33], to an object-inclusive setting. Specifically, we adapt our reward to the styles of DeepMimic [32] and AMP [33] while keeping the other components unchanged for a fair comparison. We denote these variant versions as DeepMimic* and AMP*. In Tab. 2, we present the success rate on 4 typical basketball skills from BallPlay-M, where our method achieves state-of-the-art performance in learning basketball skills compared with variants using DeepMimic or AMP style rewards. We guess that AMP-like rewards are ineffective for learning precise interactions due to their coarse-grained supervision signals." + } + ] + } + ], + "index": 40 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17545" + } + ] + } + ], + "index": 41 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 69, + 553, + 142 + ], + "blocks": [ + { + "bbox": [ + 57, + 69, + 553, + 142 + ], + "lines": [ + { + "bbox": [ + 57, + 69, + 553, + 142 + ], + "spans": [ + { + "bbox": [ + 57, + 69, + 553, + 142 + ], + "type": "image", + "image_path": "5f69af15bcda4cf123d2ad5345bdfddccd5dbeeb7090b13712e116920e733001.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 150, + 555, + 195 + ], + "lines": [ + { + "bbox": [ + 55, + 150, + 555, + 195 + ], + "spans": [ + { + "bbox": [ + 55, + 150, + 555, + 195 + ], + "type": "text", + "content": "Figure 7. Demonstration of the pickup skill learned from 40 HOI motion clips. Yellow denotes the initial frame. Left: The humanoid picks up a stationary ball effortlessly. Middle: The humanoid intercepts a ball with random velocity. Right: The humanoid adjusts after missing the ball initially (the frame in blue) and successfully retrieves it on the second attempt, showcasing the potential for learning robust and generalizable skills through extensive data collection." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 56, + 205, + 555, + 259 + ], + "blocks": [ + { + "bbox": [ + 56, + 205, + 555, + 259 + ], + "lines": [ + { + "bbox": [ + 56, + 205, + 555, + 259 + ], + "spans": [ + { + "bbox": [ + 56, + 205, + 555, + 259 + ], + "type": "table", + "html": "
SkillMimic w/o MultiplicationSkillMimic w/o CGRSkillMimic
DatasetAcc. ↑Eb-mpipe ↓Eo-mpipe ↓Ecg ↓Acc. ↑Eb-mpipe ↓Eo-mpipe ↓Ecg ↓Acc. ↑Eb-mpipe ↓Eo-mpipe ↓Ecg ↓
GRAB [40]27.0%44.7180.20.72438.6%91.0180.10.33795.4%71.178.00.026
BallPlay-V7.5%40.11662.50.30613.6%88.5155.30.41282.4%56.882.90.087
", + "image_path": "f3f2072fc93360b2096d0785ad6c111fdc299e2626329155e0a539c845b10f7c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 56, + 310, + 296, + 376 + ], + "blocks": [ + { + "bbox": [ + 55, + 269, + 555, + 292 + ], + "lines": [ + { + "bbox": [ + 55, + 269, + 555, + 292 + ], + "spans": [ + { + "bbox": [ + 55, + 269, + 555, + 292 + ], + "type": "text", + "content": "Table 1. Ablation study on imitation reward design. We experiment on two HOI datasets: GRAB and BallPlay-V. The result highlights the importance of the CG reward and multiplication in achieving precise HOI imitation." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 310, + 296, + 376 + ], + "lines": [ + { + "bbox": [ + 56, + 310, + 296, + 376 + ], + "spans": [ + { + "bbox": [ + 56, + 310, + 296, + 376 + ], + "type": "table", + "html": "
MethodSuccess Rate ↑
Pick UpDribble ForwardLayupShot
DeepMimic*[32]19.6%68.5%98.9%97.8%
AMP*[33]0.0%13.6%0.0%1.6%
SkillMimic (ours)86.7%79.6%99.1%97.9%
", + "image_path": "10ffb67c47a3e4d65b338115b14d2b8d4ebe441f9bfa7ca314e20d35c38a1ff9.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 385, + 296, + 419 + ], + "lines": [ + { + "bbox": [ + 55, + 385, + 296, + 419 + ], + "spans": [ + { + "bbox": [ + 55, + 385, + 296, + 419 + ], + "type": "text", + "content": "Table 2. Success rates across four typical basketball skills in BallPlay-M. Our method significantly outperforms variant methods that use imitation reward styles of DeepMimic and AMP." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 58, + 439, + 116, + 498 + ], + "blocks": [ + { + "bbox": [ + 58, + 439, + 116, + 498 + ], + "lines": [ + { + "bbox": [ + 58, + 439, + 116, + 498 + ], + "spans": [ + { + "bbox": [ + 58, + 439, + 116, + 498 + ], + "type": "image", + "image_path": "9b584801401090867cdb9c8cea8d03b8c41d3aa984d1ecc8e2d3803abc52f284.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 69, + 500, + 105, + 510 + ], + "lines": [ + { + "bbox": [ + 69, + 500, + 105, + 510 + ], + "spans": [ + { + "bbox": [ + 69, + 500, + 105, + 510 + ], + "type": "text", + "content": "(a) " + }, + { + "bbox": [ + 69, + 500, + 105, + 510 + ], + "type": "inline_equation", + "content": "1 - 0.5\\%" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 118, + 440, + 175, + 498 + ], + "blocks": [ + { + "bbox": [ + 118, + 440, + 175, + 498 + ], + "lines": [ + { + "bbox": [ + 118, + 440, + 175, + 498 + ], + "spans": [ + { + "bbox": [ + 118, + 440, + 175, + 498 + ], + "type": "image", + "image_path": "1066c9831f6721218ed935f2e4e36752dd65cf90a33c01c27bb5216dc4210ee6.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 124, + 500, + 167, + 510 + ], + "lines": [ + { + "bbox": [ + 124, + 500, + 167, + 510 + ], + "spans": [ + { + "bbox": [ + 124, + 500, + 167, + 510 + ], + "type": "text", + "content": "(b) " + }, + { + "bbox": [ + 124, + 500, + 167, + 510 + ], + "type": "inline_equation", + "content": "10 - 12.2\\%" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 177, + 440, + 235, + 498 + ], + "blocks": [ + { + "bbox": [ + 177, + 440, + 235, + 498 + ], + "lines": [ + { + "bbox": [ + 177, + 440, + 235, + 498 + ], + "spans": [ + { + "bbox": [ + 177, + 440, + 235, + 498 + ], + "type": "image", + "image_path": "1956039e558593426c1f1035c2d99e487675a856a2c88df695abc25ee6606ba5.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 184, + 500, + 227, + 509 + ], + "lines": [ + { + "bbox": [ + 184, + 500, + 227, + 509 + ], + "spans": [ + { + "bbox": [ + 184, + 500, + 227, + 509 + ], + "type": "text", + "content": "(c) " + }, + { + "bbox": [ + 184, + 500, + 227, + 509 + ], + "type": "inline_equation", + "content": "40 - 76.5\\%" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 520, + 296, + 575 + ], + "lines": [ + { + "bbox": [ + 55, + 520, + 296, + 575 + ], + "spans": [ + { + "bbox": [ + 55, + 520, + 296, + 575 + ], + "type": "text", + "content": "Figure 8. Pickup generalization performance with different training data scales. First number: clips; second: success rate. In each test, 1000 balls are randomly placed within 1 to 5 meters away from the center. Yellow dots indicate successful pickups and green dots represent failures. See Sec. 4.2 for details." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 235, + 440, + 293, + 498 + ], + "blocks": [ + { + "bbox": [ + 235, + 440, + 293, + 498 + ], + "lines": [ + { + "bbox": [ + 235, + 440, + 293, + 498 + ], + "spans": [ + { + "bbox": [ + 235, + 440, + 293, + 498 + ], + "type": "image", + "image_path": "0c6d6f488132e4a57a385ee583a13db367a78b0feb194e51502ad4bce15af87a.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 241, + 500, + 288, + 509 + ], + "lines": [ + { + "bbox": [ + 241, + 500, + 288, + 509 + ], + "spans": [ + { + "bbox": [ + 241, + 500, + 288, + 509 + ], + "type": "text", + "content": "(d) " + }, + { + "bbox": [ + 241, + 500, + 288, + 509 + ], + "type": "inline_equation", + "content": "131 - 85.6\\%" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "text", + "content": "In Tab. 1, we do ablation on reward design, presenting detailed metrics of HOI imitation across two diverse datasets, GRAB [40] and BallPlay-V. The results indicate that the absence of reward multiplication leads to imbalanced sub-reward learning, e.g., favoring body motion " + }, + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "inline_equation", + "content": "(E_{\\mathrm{b - m}})" + }, + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "text", + "content": " over interactions. The absence of CGR leads to kinematic local optima. In contrast, our complete setting (SkillMimic) yields significantly higher performance on interaction-related metrics (Acc., " + }, + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "inline_equation", + "content": "E_{\\mathrm{o - m}}" + }, + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "text", + "content": " , and " + }, + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "inline_equation", + "content": "E_{\\mathrm{cg}}" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 312, + 496, + 324 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 312, + 496, + 324 + ], + "spans": [ + { + "bbox": [ + 313, + 312, + 496, + 324 + ], + "type": "text", + "content": "4.2.1. Ablation on Contact Graph Reward" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 328, + 555, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 328, + 555, + 533 + ], + "spans": [ + { + "bbox": [ + 313, + 328, + 555, + 533 + ], + "type": "text", + "content": "We conduct ablation experiments on diverse datasets to comprehensively evaluate the impact of Contact Graph Reward (CGR) on interaction skill learning. Fig. 5 illustrates the qualitative results of ablating CGR on the GRAB and BallPlay-V datasets. It can be clearly observed that without CGR, the humanoid often resorts to incorrect contacts to achieve more stable object control, such as using its head to push the ball, clamping the ball with its hands and legs, or using its hands to support itself on a table for balance. These examples illustrate typical kinematic local optima. By incorporating CGR, these local optima are effectively eliminated, resulting in accurate interactions. Additionally, the result in Tab. 1 also shows that, without CGR, the contact error " + }, + { + "bbox": [ + 313, + 328, + 555, + 533 + ], + "type": "inline_equation", + "content": "(E_{\\mathrm{cg}})" + }, + { + "bbox": [ + 313, + 328, + 555, + 533 + ], + "type": "text", + "content": " significantly increases, leading to kinematic local optima and causes a noticeable decline in overall imitation accuracy. These results underscore the critical importance of CGR for learning correct interactions." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 541, + 529, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 541, + 529, + 552 + ], + "spans": [ + { + "bbox": [ + 313, + 541, + 529, + 552 + ], + "type": "text", + "content": "4.2.2. Data Scale and Generalization Performance" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 557, + 555, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 557, + 555, + 629 + ], + "spans": [ + { + "bbox": [ + 313, + 557, + 555, + 629 + ], + "type": "text", + "content": "To examine how the generalization performance grows as the data scale increases, we conduct ablation experiments on the pickup skill. We create four datasets by randomly selecting pickup clips from BallPlay-M, containing 1, 10, 40, and 131 clips, respectively. We train policy on each dataset with approximately 3.2 billion samples using SkillMimic." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 630, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 556, + 713 + ], + "type": "text", + "content": "To evaluate the generalization performance, we randomly place balls within a circular area ranging from 1 to 5 meters around the humanoid—this distribution significantly exceeds the coverage of the reference data. Fig. 7 and Fig. 8 show the results. We can see that as the data scale increases, pickup performance improves significantly due to broader state transition coverage in larger datasets. This indicates" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "17546" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 68, + 181, + 134 + ], + "blocks": [ + { + "bbox": [ + 59, + 68, + 181, + 134 + ], + "lines": [ + { + "bbox": [ + 59, + 68, + 181, + 134 + ], + "spans": [ + { + "bbox": [ + 59, + 68, + 181, + 134 + ], + "type": "image", + "image_path": "93390c62e4ea195cf924a43c76988a9ecfb43c94be15d6343eec47bb45792817.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 89, + 136, + 153, + 145 + ], + "lines": [ + { + "bbox": [ + 89, + 136, + 153, + 145 + ], + "spans": [ + { + "bbox": [ + 89, + 136, + 153, + 145 + ], + "type": "text", + "content": "(a) Manual Control" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 182, + 77, + 304, + 134 + ], + "blocks": [ + { + "bbox": [ + 182, + 77, + 304, + 134 + ], + "lines": [ + { + "bbox": [ + 182, + 77, + 304, + 134 + ], + "spans": [ + { + "bbox": [ + 182, + 77, + 304, + 134 + ], + "type": "image", + "image_path": "c5b9ed00f9c1e039bfa83d574b66cac390c92c1c4cf87f7da5f419c8ff92a787.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 225, + 136, + 263, + 145 + ], + "lines": [ + { + "bbox": [ + 225, + 136, + 263, + 145 + ], + "spans": [ + { + "bbox": [ + 225, + 136, + 263, + 145 + ], + "type": "text", + "content": "(b) Scoring" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 155, + 555, + 200 + ], + "lines": [ + { + "bbox": [ + 55, + 155, + 555, + 200 + ], + "spans": [ + { + "bbox": [ + 55, + 155, + 555, + 200 + ], + "type": "text", + "content": "Figure 9. Our method supports training a single IS policy under a unified configuration to acquire various interaction skills. These interaction skills can be flexibly switched, as illustrated in (a), where yellow denotes shot, blue denotes pickup, and green denotes turnaround layup. Complex, long-horizon tasks can be easily achieved by training a high-level controller (HLC) to manage switching of the learned interaction skills: (b) scoring from random positions (c) dribbling to target locations, and (d) dribbling along an expanding radius." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 306, + 64, + 427, + 134 + ], + "blocks": [ + { + "bbox": [ + 306, + 64, + 427, + 134 + ], + "lines": [ + { + "bbox": [ + 306, + 64, + 427, + 134 + ], + "spans": [ + { + "bbox": [ + 306, + 64, + 427, + 134 + ], + "type": "image", + "image_path": "6fb08d22749ebae1504114fcf4241bec5e9fb5dc2b0bc7186ac5f8c2fb098f47.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 347, + 136, + 386, + 145 + ], + "lines": [ + { + "bbox": [ + 347, + 136, + 386, + 145 + ], + "spans": [ + { + "bbox": [ + 347, + 136, + 386, + 145 + ], + "type": "text", + "content": "(c) Heading" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 429, + 64, + 552, + 134 + ], + "blocks": [ + { + "bbox": [ + 429, + 64, + 552, + 134 + ], + "lines": [ + { + "bbox": [ + 429, + 64, + 552, + 134 + ], + "spans": [ + { + "bbox": [ + 429, + 64, + 552, + 134 + ], + "type": "image", + "image_path": "0d50dc71dd70eb9377b14d06733963e465af0f4a1ab4d295f221d47723705871.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 470, + 136, + 509, + 145 + ], + "lines": [ + { + "bbox": [ + 470, + 136, + 509, + 145 + ], + "spans": [ + { + "bbox": [ + 470, + 136, + 509, + 145 + ], + "type": "text", + "content": "(d) Circling" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 62, + 213, + 290, + 288 + ], + "blocks": [ + { + "bbox": [ + 62, + 213, + 290, + 288 + ], + "lines": [ + { + "bbox": [ + 62, + 213, + 290, + 288 + ], + "spans": [ + { + "bbox": [ + 62, + 213, + 290, + 288 + ], + "type": "table", + "html": "
MethodSuccess Rate↑
HeadingCirclingThrowingScoring
PPO[36]0.70%11.14%0.00%0.00%
ASE[34]0.19%4.37%0.00%0.00%
ASE*[34]0.31%7.21%0.00%0.00%
SkillMimic (ours)93.04%79.92%93.40%80.25%
", + "image_path": "46607c7196da97152a749342ce831ee62b0ae5e96d83061799acae37d040ddc1.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 298, + 295, + 354 + ], + "lines": [ + { + "bbox": [ + 55, + 298, + 295, + 354 + ], + "spans": [ + { + "bbox": [ + 55, + 298, + 295, + 354 + ], + "type": "text", + "content": "Table 3. Success rates on 4 high-level basketball tasks. Both PPO (learn from scratch), ASE (using locomotion prior), and " + }, + { + "bbox": [ + 55, + 298, + 295, + 354 + ], + "type": "inline_equation", + "content": "\\mathrm{ASE^{*}}" + }, + { + "bbox": [ + 55, + 298, + 295, + 354 + ], + "type": "text", + "content": " (using body and object motion prior) fail to converge on all tasks. In contrast, leveraging interaction skill priors acquired through SkillMimic, our method effectively learns these difficult tasks." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 376, + 295, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 376, + 295, + 411 + ], + "spans": [ + { + "bbox": [ + 55, + 376, + 295, + 411 + ], + "type": "text", + "content": "our method's strong scalability and potential for large-scale learning. It is worth noting that we do not incorporate any additional designs for robustness and generalization." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 420, + 209, + 433 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 420, + 209, + 433 + ], + "spans": [ + { + "bbox": [ + 55, + 420, + 209, + 433 + ], + "type": "text", + "content": "4.3. Evaluating High-level Tasks" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 54, + 437, + 295, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 437, + 295, + 653 + ], + "spans": [ + { + "bbox": [ + 54, + 437, + 295, + 653 + ], + "type": "text", + "content": "We conduct experiments on four high-level tasks described in Sec. 3.5. We employ a single IS policy as the interaction skill prior, trained using SkillMimic for around 4.5 billion samples across seven skills: pickup, layup, turnaround layup, dribble left, dribble right, dribble forward, and shot. During training, the IS policy is fixed, and only the HLC is trained. For tasks such as throwing, heading, and circling, we train for around 0.4 billion samples. For the more challenging scoring task, we train for around 1.2 billion samples. To objectively evaluate the performance of our method, we compared it against three baseline methods: (1) learning from scratch using PPO [36], (2) learning with body motion priors using ASE [34], and (3) learning with body and object motion priors using ASE (denoted as " + }, + { + "bbox": [ + 54, + 437, + 295, + 653 + ], + "type": "inline_equation", + "content": "\\mathrm{ASE^{*}}" + }, + { + "bbox": [ + 54, + 437, + 295, + 653 + ], + "type": "text", + "content": "). The low-level controllers of ASE and " + }, + { + "bbox": [ + 54, + 437, + 295, + 653 + ], + "type": "inline_equation", + "content": "\\mathrm{ASE^{*}}" + }, + { + "bbox": [ + 54, + 437, + 295, + 653 + ], + "type": "text", + "content": " are trained using the same training steps as our IS policy. For task training, all methods are trained using identical task rewards and simulation steps for fair comparisons." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 654, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 295, + 714 + ], + "type": "text", + "content": "Fig. 9 (a) showcases an example of manually controlling the pre-trained IS policy to perform skill switching. Fig. 9 (b) shows several trajectories of the HLC performing skill controls to accurately deliver the ball to the target basket. Fig. 9 (c) & (d) present a top-down view of the HLC control" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 215, + 555, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 215, + 555, + 357 + ], + "spans": [ + { + "bbox": [ + 313, + 215, + 555, + 357 + ], + "type": "text", + "content": "ling the humanoid to dribble the ball toward target locations and along an expanding radius. As shown in Tab. 3, PPO and ASE yield low success rates because the tasks are too challenging to master without interaction priors. The failure of " + }, + { + "bbox": [ + 313, + 215, + 555, + 357 + ], + "type": "inline_equation", + "content": "\\mathrm{ASE^{*}}" + }, + { + "bbox": [ + 313, + 215, + 555, + 357 + ], + "type": "text", + "content": " can be traced back to its low-level controller's inadequate ability to imitate interaction skills. This is primarily because adversarial rewards fail to provide effective guidance for precise interaction learning, which is similar to the failure mode observed in " + }, + { + "bbox": [ + 313, + 215, + 555, + 357 + ], + "type": "inline_equation", + "content": "\\mathrm{AMP^{*}}" + }, + { + "bbox": [ + 313, + 215, + 555, + 357 + ], + "type": "text", + "content": ". In contrast, by utilizing the interaction skills learned through SkillMimic, we greatly improve the performance of learning complex high-level interaction tasks." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 369, + 476, + 382 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 369, + 476, + 382 + ], + "spans": [ + { + "bbox": [ + 313, + 369, + 476, + 382 + ], + "type": "text", + "content": "5. Discussion and Future Work" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 390, + 554, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 390, + 554, + 485 + ], + "spans": [ + { + "bbox": [ + 313, + 390, + 554, + 485 + ], + "type": "text", + "content": "We presented a novel framework for learning diverse interaction skills with a unified HOI imitation reward using reinforcement learning. Our approach demonstrates multi-skill mastery, skill reusability, and clear data scaling benefits, highlighting the significant potential of HOI imitation for learning complex interaction skills. However, the full application of this paradigm requires further exploration. Several key areas deserve particular attention:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 486, + 554, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 486, + 554, + 569 + ], + "spans": [ + { + "bbox": [ + 313, + 486, + 554, + 569 + ], + "type": "text", + "content": "Firstly, exploring more general HOI skill learning within a single policy, such as household tasks or other types of sports, presents challenges in handling generalization across different object geometries. Moreover, perceiving multiple objects simultaneously requires a more general environmental perception approach, rather than directly obtaining privileged information for each object." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 570, + 553, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 553, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 553, + 605 + ], + "type": "text", + "content": "Secondly, incorporating finer-grained control conditions, such as tracking [27], could enable more flexible control and generalization capabilities as HOI data scales up." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 605, + 553, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 553, + 654 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 553, + 654 + ], + "type": "text", + "content": "Thirdly, the difficulty of collecting HOI data highlights the importance of exploring methods to learn robust interaction skills with limited HOI data, as well as investigating HOI data augmentation and generation techniques." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 654, + 553, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 553, + 712 + ], + "type": "text", + "content": "Lastly, applying this technology to real robots would be highly valuable, such as enabling real-world humanoid robots to play basketball games. This endeavor could involve various challenges, including HOI data retargeting and sim-to-real transfer issues." + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17547" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "type": "text", + "content": "Acknowledgment" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 163 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 163 + ], + "type": "text", + "content": "We would like to thank Mingyang Su for his valuable assistance in motion capture data collection and Shunlin Lu for helpful discussions. This work was supported in part by Unitree Robotics, International Digital Economy Academy (IDEA), and the Innovation and Technology Fund of HK-SAR (GHX/054/21GD)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 172, + 115, + 185 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 172, + 115, + 185 + ], + "spans": [ + { + "bbox": [ + 56, + 172, + 115, + 185 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 193, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 193, + 296, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 193, + 296, + 236 + ], + "spans": [ + { + "bbox": [ + 61, + 193, + 296, + 236 + ], + "type": "text", + "content": "[1] Kai Arulkumaran, Marc Peter Deisenroth, Miles Brundage, and Anil Anthony Bharath. Deep reinforcement learning: A brief survey. IEEE Signal Processing Magazine, 34(6):26-38, 2017. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 237, + 296, + 291 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 237, + 296, + 291 + ], + "spans": [ + { + "bbox": [ + 61, + 237, + 296, + 291 + ], + "type": "text", + "content": "[2] Jinseok Bae, Jungdam Won, Donggeun Lim, Cheol-Hui Min, and Young Min Kim. Pmp: Learning to physically interact with environments using part-wise motion priors. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-10, 2023. 1, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 293, + 295, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 293, + 295, + 346 + ], + "spans": [ + { + "bbox": [ + 61, + 293, + 295, + 346 + ], + "type": "text", + "content": "[3] Kevin Black, Noah Brown, Danny Driess, Adnan Esmail, Michael Equi, Chelsea Finn, Niccolo Fusai, Lachy Groom, Karol Hausman, Brian Ichter, et al. Pi0: A vision-language-action flow model for general robot control. arXiv preprint arXiv:2410.24164, 2024. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 348, + 295, + 391 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 348, + 295, + 391 + ], + "spans": [ + { + "bbox": [ + 62, + 348, + 295, + 391 + ], + "type": "text", + "content": "[4] Jona Braun, Sammy Christen, Muhammed Kocabas, Emre Aksan, and Otmar Hilliges. Physically plausible full-body hand-object interaction synthesis. arXiv preprint arXiv:2309.07907, 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 392, + 296, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 392, + 296, + 446 + ], + "spans": [ + { + "bbox": [ + 62, + 392, + 296, + 446 + ], + "type": "text", + "content": "[5] Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Xi Chen, Krzysztof Choromanski, Tianli Ding, Danny Driess, Avinava Dubey, Chelsea Finn, et al. Rt-2: Vision-language-action models transfer web knowledge to robotic control. arXiv preprint arXiv:2307.15818, 2023. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 447, + 294, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 447, + 294, + 479 + ], + "spans": [ + { + "bbox": [ + 62, + 447, + 294, + 479 + ], + "type": "text", + "content": "[6] Stelian Coros, Philippe Beaudoin, and Michiel Van de Panne. Generalized biped walking control. ACM Transactions On Graphics (TOG), 29(4):1-9, 2010. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 481, + 294, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 481, + 294, + 525 + ], + "spans": [ + { + "bbox": [ + 62, + 481, + 294, + 525 + ], + "type": "text", + "content": "[7] Jieming Cui, Tengyu Liu, Nian Liu, Yaodong Yang, Yixin Zhu, and Siyuan Huang. Anyskill: Learning openvocabulary physical skill for interactive agents. In Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 525, + 294, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 525, + 294, + 579 + ], + "spans": [ + { + "bbox": [ + 62, + 525, + 294, + 579 + ], + "type": "text", + "content": "[8] Zhiyang Dou, Xuelin Chen, Qingnan Fan, Taku Komura, and Wenping Wang. C-ase: Learning conditional adversarial skill embeddings for physics-based characters. In SIGGRAPH Asia 2023 Conference Papers, pages 1-11, 2023. 1, 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 581, + 295, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 581, + 295, + 645 + ], + "spans": [ + { + "bbox": [ + 62, + 581, + 295, + 645 + ], + "type": "text", + "content": "[9] Zicong Fan, Omid Taheri, Dimitrios Tzionas, Muhammed Kocabas, Manuel Kaufmann, Michael J Black, and Otmar Hilliges. Arctic: A dataset for dexterous bimanual handobject manipulation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12943-12954, 2023. 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 647, + 295, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 647, + 295, + 691 + ], + "spans": [ + { + "bbox": [ + 57, + 647, + 295, + 691 + ], + "type": "text", + "content": "[10] Zipeng Fu, Qingqing Zhao, Qi Wu, Gordon Wetzstein, and Chelsea Finn. Humanplus: Humanoid shadowing and imitation from humans. In 8th Annual Conference on Robot Learning, 2024. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 692, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 692, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 692, + 295, + 713 + ], + "type": "text", + "content": "[11] Zipeng Fu, Tony Z Zhao, and Chelsea Finn. Mobile aloha: Learning bimanual mobile manipulation with low-cost" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "type": "text", + "content": "whole-body teleoperation. arXiv preprint arXiv:2401.02117, 2024.3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 97, + 553, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 97, + 553, + 152 + ], + "spans": [ + { + "bbox": [ + 316, + 97, + 553, + 152 + ], + "type": "text", + "content": "[12] Jiawei Gao, Ziqin Wang, Zeqi Xiao, Jingbo Wang, Tai Wang, Jinkun Cao, Xiaolin Hu, Si Liu, Jifeng Dai, and Jiangmiao Pang. Coohoi: Learning cooperative human-object interaction with manipulated object dynamics. Advances in Neural Information Processing Systems, 37:79741-79763, 2024. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 153, + 555, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 153, + 555, + 219 + ], + "spans": [ + { + "bbox": [ + 316, + 153, + 555, + 219 + ], + "type": "text", + "content": "[13] Haoran Geng, Helin Xu, Chengyang Zhao, Chao Xu, Li Yi, Siyuan Huang, and He Wang. Gapartner: Cross-category domain-generalizable object perception and manipulation via generalizable and actionable parts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7081-7091, 2023. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 221, + 553, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 221, + 553, + 266 + ], + "spans": [ + { + "bbox": [ + 316, + 221, + 553, + 266 + ], + "type": "text", + "content": "[14] Anindita Ghosh, Rishabh Dabral, Vladislav Golyanik, Christian Theobalt, and Philipp Slusallek. Imos: Intent-driven full-body motion synthesis for human-object interactions. In Eurographics, 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 267, + 553, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 267, + 553, + 312 + ], + "spans": [ + { + "bbox": [ + 316, + 267, + 553, + 312 + ], + "type": "text", + "content": "[15] Mohamed Hassan, Yunrong Guo, Tingwu Wang, Michael Black, Sanja Fidler, and Xue Bin Peng. Synthesizing physical character-scene interactions. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-9, 2023. 1, 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 313, + 553, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 313, + 553, + 369 + ], + "spans": [ + { + "bbox": [ + 316, + 313, + 553, + 369 + ], + "type": "text", + "content": "[16] Tairan He, Zhengyi Luo, Xialin He, Wenli Xiao, Chong Zhang, Weinan Zhang, Kris M Kitani, Changliu Liu, and Guanya Shi. Omnih2o: Universal and dexterous human-to-humanoid whole-body teleoperation and learning. In 8th Annual Conference on Robot Learning, 2024. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 370, + 553, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 370, + 553, + 403 + ], + "spans": [ + { + "bbox": [ + 316, + 370, + 553, + 403 + ], + "type": "text", + "content": "[17] Edmond SL Ho, Taku Komura, and Chiew-Lan Tai. Spatial relationship preserving character motion adaptation. In ACM SIGGRAPH 2010 papers, pages 1-8, 2010. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 405, + 553, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 405, + 553, + 437 + ], + "spans": [ + { + "bbox": [ + 316, + 405, + 553, + 437 + ], + "type": "text", + "content": "[18] Jonathan Ho and Stefano Ermon. Generative adversarial imitation learning. Advances in neural information processing systems, 29, 2016. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 440, + 553, + 484 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 440, + 553, + 484 + ], + "spans": [ + { + "bbox": [ + 316, + 440, + 553, + 484 + ], + "type": "text", + "content": "[19] Jessica K Hodgins, Wayne L Wooten, David C Brogan, and James F O'Brien. Animating human athletics. In Proceedings of the 22nd annual conference on Computer graphics and interactive techniques, pages 71-78, 1995. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 486, + 553, + 540 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 486, + 553, + 540 + ], + "spans": [ + { + "bbox": [ + 316, + 486, + 553, + 540 + ], + "type": "text", + "content": "[20] Nan Jiang, Tengyu Liu, Zhexuan Cao, Jieming Cui, Zhiyuan Zhang, Yixin Chen, He Wang, Yixin Zhu, and Siyuan Huang. Full-body articulated human-object interaction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9365-9376, 2023. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 543, + 553, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 543, + 553, + 586 + ], + "spans": [ + { + "bbox": [ + 316, + 543, + 553, + 586 + ], + "type": "text", + "content": "[21] Nan Jiang, Zhiyuan Zhang, Hongjie Li, Xiaoxuan Ma, Zan Wang, Yixin Chen, Tengyu Liu, Yixin Zhu, and Siyuan Huang. Scaling up dynamic human-scene interaction modeling. arXiv preprint arXiv:2403.08629, 2024. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 588, + 553, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 588, + 553, + 632 + ], + "spans": [ + { + "bbox": [ + 316, + 588, + 553, + 632 + ], + "type": "text", + "content": "[22] Jordan Juravsky, Yunrong Guo, Sanja Fidler, and Xue Bin Peng. Padl: Language-directed physics-based character control. In SIGGRAPH Asia 2022 Conference Papers, pages 1-9, 2022. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 634, + 553, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 634, + 553, + 677 + ], + "spans": [ + { + "bbox": [ + 316, + 634, + 553, + 677 + ], + "type": "text", + "content": "[23] Jiaman Li, Alexander Clegg, Roozbeh Mottaghi, Jiajun Wu, Xavier Puig, and C Karen Liu. Controllable human-object interaction synthesis. arXiv preprint arXiv:2312.03913, 2023. 1, 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 680, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 680, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 680, + 553, + 713 + ], + "type": "text", + "content": "[24] Jiaman Li, Jiajun Wu, and C Karen Liu. Object motion guided human motion synthesis. ACM Transactions on Graphics (TOG), 42(6):1-11, 2023. 2" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17548" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 56, + 73, + 296, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 296, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 296, + 116 + ], + "type": "text", + "content": "[25] Libin Liu and Jessica Hodgins. Learning to schedule control fragments for physics-based characters using deep q-learning. ACM Transactions on Graphics (TOG), 36(3):1-14, 2017. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 296, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 296, + 160 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 296, + 160 + ], + "type": "text", + "content": "[26] Libin Liu and Jessica Hodgins. Learning basketball dribbling skills using trajectory optimization and deep reinforcement learning. ACM Transactions on Graphics (TOG), 37(4):1-14, 2018. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 162, + 296, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 162, + 296, + 205 + ], + "spans": [ + { + "bbox": [ + 56, + 162, + 296, + 205 + ], + "type": "text", + "content": "[27] Zhengyi Luo, Jinkun Cao, Kris Kitani, Weipeng Xu, et al. Perpetual humanoid control for real-time simulated avatars. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10895-10904, 2023. 6, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 206, + 296, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 206, + 296, + 249 + ], + "spans": [ + { + "bbox": [ + 56, + 206, + 296, + 249 + ], + "type": "text", + "content": "[28] Zhengyi Luo, Jinkun Cao, Sammy Christen, Alexander Winkler, Kris Kitani, and Weipeng Xu. Omnigrasp: Grasping diverse objects with simulated humanoids. Advances in Neural Information Processing Systems, 37:2161-2184, 2024. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 251, + 296, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 251, + 296, + 303 + ], + "spans": [ + { + "bbox": [ + 56, + 251, + 296, + 303 + ], + "type": "text", + "content": "[29] Zhengyi Luo, Jiashun Wang, Kangni Liu, Haotian Zhang, Chen Tessler, Jingbo Wang, Ye Yuan, Jinkun Cao, Zihui Lin, Fengyi Wang, et al. Smplolympics: Sports environments for physically simulated humanoids. arXiv preprint arXiv:2407.00187, 2024. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 304, + 296, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 304, + 296, + 348 + ], + "spans": [ + { + "bbox": [ + 56, + 304, + 296, + 348 + ], + "type": "text", + "content": "[30] Vinod Nair and Geoffrey E Hinton. Rectified linear units improve restricted boltzmann machines. In Proceedings of the 27th international conference on machine learning (ICML-10), pages 807–814, 2010. 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 350, + 296, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 350, + 296, + 403 + ], + "spans": [ + { + "bbox": [ + 56, + 350, + 296, + 403 + ], + "type": "text", + "content": "[31] Liang Pan, Jingbo Wang, Buzhen Huang, Junyu Zhang, Haofan Wang, Xu Tang, and Yangang Wang. Synthesizing physically plausible human motions in 3d scenes. In 2024 International Conference on 3D Vision (3DV), pages 1498-1507. IEEE, 2024. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 404, + 296, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 404, + 296, + 437 + ], + "spans": [ + { + "bbox": [ + 56, + 404, + 296, + 437 + ], + "type": "text", + "content": "[32] Xue Bin Peng, Pieter Abbeel, Sergey Levine, and Michiel van de Panne. Deepmimic. ACM Transactions on Graphics, page 1-14, 2018. 1, 2, 4, 5, 6, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 437, + 296, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 437, + 296, + 480 + ], + "spans": [ + { + "bbox": [ + 56, + 437, + 296, + 480 + ], + "type": "text", + "content": "[33] Xue Bin Peng, Ze Ma, Pieter Abbeel, Sergey Levine, and Angjoo Kanazawa. Amp: Adversarial motion priors for stylized physics-based character control. ACM Transactions on Graphics, page 1-20, 2021. 2, 3, 4, 5, 6, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 482, + 296, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 482, + 296, + 525 + ], + "spans": [ + { + "bbox": [ + 56, + 482, + 296, + 525 + ], + "type": "text", + "content": "[34] Xue Bin Peng, Yunrong Guo, Lina Halper, Sergey Levine, and Sanja Fidler. Ase: Large-scale reusable adversarial skill embeddings for physically simulated characters. ACM Trans. Graph., 41(4), 2022. 1, 2, 3, 4, 5, 6, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 526, + 296, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 526, + 296, + 568 + ], + "spans": [ + { + "bbox": [ + 56, + 526, + 296, + 568 + ], + "type": "text", + "content": "[35] Jiawei Ren, Mingyuan Zhang, Cunjun Yu, Xiao Ma, Liang Pan, and Ziwei Liu. Insactor: Instruction-driven physics-based characters. Advances in Neural Information Processing Systems, 36, 2024. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 570, + 296, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 570, + 296, + 602 + ], + "spans": [ + { + "bbox": [ + 56, + 570, + 296, + 602 + ], + "type": "text", + "content": "[36] John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347, 2017. 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 603, + 296, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 603, + 296, + 635 + ], + "spans": [ + { + "bbox": [ + 56, + 603, + 296, + 635 + ], + "type": "text", + "content": "[37] Sebastian Starke, Yiwei Zhao, Taku Komura, and Kazi Zaman. Local motion phases for learning multi-contact character movements. ACM Trans. Graph., 39(4), 2020. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 636, + 296, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 636, + 296, + 669 + ], + "spans": [ + { + "bbox": [ + 56, + 636, + 296, + 669 + ], + "type": "text", + "content": "[38] Sebastian Starke, Yiwei Zhao, Fabio Zinno, and Taku Komura. Neural animation layering for synthesizing martial arts movements. ACM Trans. Graph., 40(4), 2021. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 670, + 296, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 296, + 712 + ], + "type": "text", + "content": "[39] Jingkai Sun, Qiang Zhang, Yiqun Duan, Xiaoyang Jiang, Chong Cheng, and Renjing Xu. Prompt, plan, perform: Llm-based humanoid control via quantized imitation learning. arXiv preprint arXiv:2309.11359, 2023. 2" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 554, + 700 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 316, + 73, + 554, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 116 + ], + "type": "text", + "content": "[40] Omid Taheri, Nima Ghorbani, Michael J. Black, and Dimitrios Tzionas. GRAB: A dataset of whole-body human grasping of objects. In European Conference on Computer Vision (ECCV), 2020. 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 118, + 554, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 554, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 554, + 150 + ], + "type": "text", + "content": "[41] Jie Tan, Yuting Gu, C Karen Liu, and Greg Turk. Learning bicycle stunts. ACM Transactions on Graphics (TOG), 33 (4):1-12, 2014. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 152, + 554, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 554, + 205 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 554, + 205 + ], + "type": "text", + "content": "[42] Chen Tessler, Yoni Kasten, Yunrong Guo, Shie Mannor, Gal Chechik, and Xue Bin Peng. Calm: Conditional adversarial latent models for directable virtual characters. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-9, 2023. 1, 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 207, + 554, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 207, + 554, + 251 + ], + "spans": [ + { + "bbox": [ + 316, + 207, + 554, + 251 + ], + "type": "text", + "content": "[43] Yinhuai Wang, Jing Lin, Ailing Zeng, Zhengyi Luo, Jian Zhang, and Lei Zhang. Physhoi: Physics-based imitation of dynamic human-object interaction. arXiv preprint arXiv:2312.04393, 2023. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 252, + 554, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 252, + 554, + 285 + ], + "spans": [ + { + "bbox": [ + 316, + 252, + 554, + 285 + ], + "type": "text", + "content": "[44] Jungdam Won, Deepak Gopinath, and Jessica Hodgins. Physics-based character controllers using conditional vaes. ACM Transactions on Graphics (TOG), 41(4):1-12, 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 286, + 554, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 286, + 554, + 329 + ], + "spans": [ + { + "bbox": [ + 316, + 286, + 554, + 329 + ], + "type": "text", + "content": "[45] Yan Wu, Jiahao Wang, Yan Zhang, Siwei Zhang, Otmar Hilliges, Fisher Yu, and Siyu Tang. Saga: Stochastic whole-body grasping with contact. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 331, + 554, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 331, + 554, + 384 + ], + "spans": [ + { + "bbox": [ + 316, + 331, + 554, + 384 + ], + "type": "text", + "content": "[46] Zeqi Xiao, Tai Wang, Jingbo Wang, Jinkun Cao, Wenwei Zhang, Bo Dai, Dahua Lin, and Jiangmiao Pang. Unified human-scene interaction via prompted chain-of-contacts. In The Twelfth International Conference on Learning Representations, 2024. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 387, + 554, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 387, + 554, + 441 + ], + "spans": [ + { + "bbox": [ + 316, + 387, + 554, + 441 + ], + "type": "text", + "content": "[47] Sirui Xu, Zhengyuan Li, Yu-Xiong Wang, and Liang-Yan Gui. Interdiff: Generating 3d human-object interactions with physics-informed diffusion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14928-14940, 2023. 1, 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 443, + 554, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 443, + 554, + 475 + ], + "spans": [ + { + "bbox": [ + 316, + 443, + 554, + 475 + ], + "type": "text", + "content": "[48] Sirui Xu, Ziyin Wang, Yu-Xiong Wang, and Liang-Yan Gui. Interdreamer: Zero-shot text to 3d dynamic human-object interaction. arXiv preprint arXiv:2403.19652, 2024. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 476, + 554, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 476, + 554, + 509 + ], + "spans": [ + { + "bbox": [ + 316, + 476, + 554, + 509 + ], + "type": "text", + "content": "[49] KangKang Yin, Kevin Loken, and Michiel Van de Panne. Simbicon: Simple biped locomotion control. ACM Transactions on Graphics (TOG), 26(3):105-es, 2007. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 510, + 554, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 510, + 554, + 553 + ], + "spans": [ + { + "bbox": [ + 316, + 510, + 554, + 553 + ], + "type": "text", + "content": "[50] Haotian Zhang, Ye Yuan, Viktor Makoviychuk, Yunrong Guo, Sanja Fidler, Xue Bin Peng, and Kayvon Fatahalian. Learning physically simulated tennis skills from broadcast videos. ACM Trans. Graph., 2023. 1, 2, 3, 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 555, + 554, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 555, + 554, + 609 + ], + "spans": [ + { + "bbox": [ + 316, + 555, + 554, + 609 + ], + "type": "text", + "content": "[51] Hui Zhang, Sammy Christen, Zicong Fan, Luocheng Zheng, Jemin Hwangbo, Jie Song, and Otmar Hilliges. Artigrasp: Physically plausible synthesis of bi-manual dexterous grasping and articulation. In 2024 International Conference on 3D Vision (3DV), pages 235-246. IEEE, 2024. 5" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 611, + 554, + 664 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 611, + 554, + 664 + ], + "spans": [ + { + "bbox": [ + 316, + 611, + 554, + 664 + ], + "type": "text", + "content": "[52] Yunbo Zhang, Deepak Gopinath, Yuting Ye, Jessica Hodgins, Greg Turk, and Jungdam Won. Simulation and retargeting of complex multi-character interactions. In ACM SIGGRAPH 2023 Conference Proceedings, pages 1-11, 2023. 3, 5" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 666, + 554, + 700 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 666, + 554, + 700 + ], + "spans": [ + { + "bbox": [ + 316, + 666, + 554, + 700 + ], + "type": "text", + "content": "[53] Qingxu Zhu, He Zhang, Mengting Lan, and Lei Han. Neural categorical priors for physics-based character control. ACM Transactions on Graphics (TOG), 42(6):1-16, 2023. 2" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "17549" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Skip Tuning_ Pre-trained Vision-Language Models are Effective and Efficient Adapters Themselves/e5453c9b-356b-43d9-9454-788a3f8b346d_content_list.json b/2025/Skip Tuning_ Pre-trained Vision-Language Models are Effective and Efficient Adapters Themselves/e5453c9b-356b-43d9-9454-788a3f8b346d_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..05ae68136360cf01709dafe02aa88a59eb21e575 --- /dev/null +++ b/2025/Skip Tuning_ Pre-trained Vision-Language Models are Effective and Efficient Adapters Themselves/e5453c9b-356b-43d9-9454-788a3f8b346d_content_list.json @@ -0,0 +1,1816 @@ +[ + { + "type": "text", + "text": "Skip Tuning: Pre-trained Vision-Language Models are Effective and Efficient Adapters Themselves", + "text_level": 1, + "bbox": [ + 222, + 130, + 776, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Shihan Wu $^{1}$ Ji Zhang $^{2*}$ Pengpeng Zeng $^{3}$ Lianli Gao $^{3}$ Jingkuan Song $^{4}$ Heng Tao Shen $^{4}$", + "bbox": [ + 114, + 200, + 883, + 222 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 University of Electronic Science and Technology of China (UESTC)", + "bbox": [ + 222, + 222, + 774, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Southwest Jiaotong University", + "bbox": [ + 370, + 238, + 625, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3 Shenzhen Institute for Advanced Study, UESTC", + "bbox": [ + 300, + 256, + 694, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{4}$ Tongji University", + "bbox": [ + 421, + 273, + 575, + 292 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{shihan.wu.koorye@outlook.com, jizhang.jim@gmail.com}", + "bbox": [ + 269, + 294, + 725, + 309 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 344, + 326, + 359 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Prompt tuning (PT) has long been recognized as an effective and efficient paradigm for transferring large pre-trained vision-language models (VLMs) to downstream tasks by learning a tiny set of context vectors. Nevertheless, in this work, we reveal that freezing the parameters of VLMs during learning the context vectors neither facilitates the transferability of pre-trained knowledge nor improves the memory and time efficiency significantly. Upon further investigation, we find that reducing both the length and width of the feature-gradient propagation flows of the full fine-tuning (FT) baseline is key to achieving effective and efficient knowledge transfer. Motivated by this, we propose Skip Tuning, a novel paradigm for adapting VLMs to downstream tasks. Unlike existing PT or adapter-based methods, Skip Tuning applies Layer-wise Skipping (LSkip) and Classwise Skipping (CSkip) upon the FT baseline without introducing extra context vectors or adapter modules. Extensive experiments across a wide spectrum of benchmarks demonstrate the superior effectiveness and efficiency of our Skip Tuning over both PT and adapter-based methods. Code: https://github.com/Koorye/SkipTuning.", + "bbox": [ + 86, + 377, + 483, + 694 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 726, + 220, + 742 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "There have recently been significant advancements in large pre-trained vision-language models (VLMs) [1, 24, 29]. One notable achievement is the CLIP model [24], which leverages an image-text matching (ITM) loss to align images with their corresponding textual descriptions in a common feature space. While VLMs have proven impressive capabilities in recognizing open-set visual concepts, their zero-shot generalization performance declines significantly", + "bbox": [ + 89, + 752, + 483, + 875 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/52012149a2b74e739ce937db35d73275a6d735cd932942e51933cea05c8c5807.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 531, + 340, + 681, + 422 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/04f7f904d4ff9c125ba4a0e81dac82a9846d98c05f9ee7c626fa072ca02e18bb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 684, + 342, + 785, + 422 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/637edd4d4f795e3863d9bcf476dce04b395743401f859277d5691b1a966f2048.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 789, + 342, + 888, + 422 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/bec99f81e1ebb308bfa5763dfad5895e68ed2218fc8f2381cfde13d9869fb887.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 529, + 425, + 681, + 506 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9b9f9f7401feb7f2a82f77008366704c2164290c98137ac1b5355eccc217218a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 684, + 425, + 782, + 507 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/dd0d41d243ed1a609d0d1887ed982960c1468b9a7939a0f755c15c3de6a0b566.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 789, + 425, + 888, + 507 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/892b3907047f74164c64be38f5db12e9685982e621fd673df64374abd0e74aa6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 529, + 508, + 681, + 583 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/99126689c80abbc6f16d9272f31f9e7a8f9f67512237fb242416a164aadbfcaf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 684, + 508, + 784, + 583 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f48f2194776fafc7819003b22098ad789e6d07072cf53a25c4b26fc551073fe5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 789, + 508, + 888, + 583 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a9b5c20b2d3e3466709c90502a5577ae53031f74f8227a671ee70c4c83de655b.jpg", + "image_caption": [ + "Figure 1. Comparison of our devised Skip Tuning with state-of-the-art prompt tuning methods in terms of training time (seconds), memory cost (M), and classification accuracy (\\%) across base-to-new generalization, cross-dataset generalization, domain generalization, and few-shot learning benchmarks. $\\times$ indicates the performance improvement over the state-of-the-art. Comparison results with the adapter-based methods are reported in Table 5." + ], + "image_footnote": [], + "bbox": [ + 529, + 585, + 681, + 633 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a776cef8fef9e06066ad2d541d3809fea0fd63c01c6fdb6c73fd2548c3db8333.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 684, + 585, + 784, + 633 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d7b1f847708d48087fb5011bd853344d6939b1a79c8e66c3f2500faa2fc44f39.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 789, + 585, + 888, + 633 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "when encountering category, distribution, or domain shifts between upstream training data and downstream tasks.", + "bbox": [ + 511, + 763, + 906, + 792 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Prompt tuning (PT) has long been recognized as an effective and efficient paradigm for transferring large pre-trained vision-language models (VLMs) to downstream tasks. The core concept of PT is to learn a task-specific prompt (i.e., a small number of context vectors) for the target task, using a limited amount of training data, while keeping the pre-trained VLM parameters fixed. Although many PT ap", + "bbox": [ + 511, + 795, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author.", + "bbox": [ + 109, + 887, + 235, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "14723", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5fe69620d9a1ae354053712bd3ac8d873539f6be59708020154b7fa6e50585f8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 125, + 87, + 290, + 166 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/355e9d569600d205411f954e5a5a8665de6909ac844109780573ec7120975b0c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 295, + 88, + 480, + 167 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/7196e6fe8ee7b931c23df993265fa5e4ca97222874310cc5a79386d8d5c2581b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 94, + 172, + 290, + 250 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/b43f70d52bfa825b77f644a21af4c50c31c99c9f888a2baa5da7acd5148dca2b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 295, + 172, + 480, + 250 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/cddd56924c33fbd6b808bbb6094a42e45ca2de58bf82b99467f96ef0b8c6b5c4.jpg", + "image_caption": [ + "Figure 2. Motivations. (a) Comparison between the prompt tuning (PT) method CoOp [42] and the full fine-tuning (FT) baseline in terms of i) the number of learnable parameters, ii) memory usage, iii) time cost, and iv) base-to-new generalization performance. (b) Feature Sensitivity (FS) of CLIP's network layers, averaged over 100 randomly-sampled training images. (c) Gradient Dependence (GD) of class tokens for different training images." + ], + "image_footnote": [], + "bbox": [ + 94, + 255, + 290, + 330 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/3e3a8eedd5ace8dd0c18c7c7d29bd159fa7c17a6e8f2af8e9ec8695803c7e5bb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 295, + 255, + 480, + 330 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "proaches have reported improved performance and parameter efficiency over the full fine-tuning (FT) baseline, the discrepancy of implementation details among those PT approaches obscure the actual performance enhancement. For example, the FT performance can be significantly underestimated by training with coarsely tuned hyper-parameters. Quantitative evidence is presented in Figure 2 (a), where we make comparisons between the PT method CoOp [42] and the FT baseline in terms of the number of learnable parameters, memory usage, time cost as well as base-to-new generalization performance (see §2.2). As observed, although PT significantly improves the parameter efficiency of FT (using 1/51200 parameters of FT), the improvements in memory and time efficiency are relatively insignificant (reducing only $6.3\\%$ memory usage and $15.8\\%$ time cost). Besides, compared to FT, the classification accuracy of PT decreases by $3.49\\%$ and $4.49\\%$ on base and new tasks, respectively. This suggests that pursuing higher parameter efficiency by freezing the overwhelming majority weights of VLMs during learning the context vectors neither facilitates the transferability of pre-trained knowledge nor improves memory and time efficiency considerably. Besides, in many real-world applications, memory and time efficiency often take precedence over parameter efficiency in terms of practical importance. We therefore raise the following question:", + "bbox": [ + 89, + 468, + 482, + 845 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Can we optimize the memory and time efficiency of the FT baseline and adapt VLMs to downstream tasks in an effective and efficient manner?", + "bbox": [ + 106, + 852, + 465, + 898 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To answer the above question, we scrutinize the FeatureGradient Propagation Flows (FGPFs) in the vision encoder and text encoder of the CLIP model when performing FT on the base (or target) task. Interestingly, we observe that, for each training image, the majority of shallow network layers and class tokens contribute minimally to capturing task-specific knowledge for the base task (see §2.2). Motivated by this, we propose Skip Tuning, a novel paradigm for adapting VLMs to downstream tasks without introducing extra context vectors or adapter modules. Concretely, Skip Tuning incorporates two strategies, Layer-wise Skipping (LSkip) and Class-wise Skipping (CSkip), to simultaneously reduce the length and width of FGFs in the FT baseline, thereby establishing effective and efficient knowledge transfer of VLMs, as shown in Figure 3.", + "bbox": [ + 511, + 90, + 903, + 316 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Effectiveness and Efficiency. We conduct extensive experiments across a wide range of benchmarks to validate the effectiveness and efficiency of our Skip Tuning scheme. An overview of the achieved results is shown in Figure 1. As seen, our Skip Tuning demonstrates superiority over existing PT methods, e.g., on the few-shot learning benchmark, Skip Tuning achieves $\\times 15$ time efficiency, $\\times 6.4$ memory efficiency, while yielding a $1.04\\%$ improvement in ACC compared to the state-of-the-art. Furthermore, we also show the advantages of Skip Tuning over existing adapter-based methods in Table 5, where Skip Tuning achieves $\\times 3.8$ time efficiency, $\\times 3.9$ memory efficiency, along with a $3.59\\%$ H ACC enhancement over the strong rival LoRA [15].", + "bbox": [ + 511, + 316, + 903, + 513 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Main Contributions. The main contributions are threefold.", + "bbox": [ + 511, + 513, + 903, + 527 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We reveal that reducing both the width and length of the feature-gradient propagation flows (FGPFs) of the full fine-tuning (FT) baseline is key to establishing effective and efficient knowledge transfer.", + "- We devise Skip Tuning, an effective and efficient method for transferring VLMs to downstream tasks without relying on extra context vectors or adapter modules.", + "- We evaluate our method on a wide spectrum of benchmarks, demonstrating the superiority of Skip Tuning over both prompt tuning and adapter-based approaches." + ], + "bbox": [ + 513, + 529, + 903, + 680 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Methodology", + "text_level": 1, + "bbox": [ + 513, + 693, + 648, + 709 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this section, we elaborate on our Skip Tuning approach. We start with an introduction of preliminaries.", + "bbox": [ + 511, + 718, + 903, + 750 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Preliminaries", + "text_level": 1, + "bbox": [ + 511, + 757, + 653, + 771 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Contrastive Language-Image Pre-training (CLIP). Following the common practice of existing paradigms for adapting PVLMs, we adopt CLIP [24] as the testbed in this work. CLIP aims to learn an alignment between image and text features generated by an image encoder and a text encoder, respectively. By being exposed to 400 million imagetext association pairs and employing a contrastive learning paradigm within a shared feature space, CLIP acquires a di", + "bbox": [ + 511, + 779, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "14724", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/70dad74a6f292a14550ffedc09571bf12c8f1b5ed293682e9f3ba2ca84e6c3c4.jpg", + "image_caption": [ + "Figure 3. Overview of our proposed Skip Tuning. Skip Tuning performs Layer-wise Skipping (LSkip) and Class-wise Skipping (CSkip)) to enhance the memory and time efficiency of the FT baseline. Specifically, LSkip reduces the length of feature-gradient propagation flows (FGPFs) by caching intermediate features produced by the $\\omega$ -th layers of CLIP's vision encoder $E_{V}$ and text encoder $E_{T}$ before FT begins. In contrast, CSkip reduces the width of FGFs by filtering out unimportant class tokens in the text encoder $E_{T}$ for every training image." + ], + "image_footnote": [], + "bbox": [ + 93, + 85, + 903, + 272 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "verse array of open-set visual concepts that can be readily applied to downstream tasks. For instance, zero-shot classification can be achieved by framing the classification task as an image-text matching problem. Initially, a prompt (e.g. \"a photo of a\") is crafted to extract text features of all intra-task classes by presenting the class-extended prompt (\"a photo of a [CLS]\") to the text encoder. Subsequently, the image encoder is utilized to derive the image feature of a given input example, and class prediction is performed by comparing cosine distances between the image feature and text features of classes.", + "bbox": [ + 88, + 359, + 480, + 525 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Prompt Tuning with an Image-Text Matching Loss. PT aims to learn a task-specific prompt with few-labeled samples from the base task. Concretely, a prompt is formulated as $\\kappa$ context vectors: $\\{v_{1}, v_{2}, \\dots, v_{\\kappa}\\}$ . During training, we produce the text feature of the $j$ -th class by inputting the class-token-extended prompt $c_{j} = \\{v_{1}, v_{2}, \\dots, v_{\\kappa}, [\\mathrm{CLS}]\\}$ to the text encoder $g(\\cdot)$ , where [CLS] indicates the class token of the $j$ -th class. Denote $\\pmb{f}_{i} \\in \\mathbb{R}^{d}$ as the image feature of a sample $\\pmb{x}_{i}$ extracted by the image encoder, the task-specific prompt can be updated via back-propagating the image-text matching (ITM) loss through the frozen CLIP model. The ITM loss can be expressed as", + "bbox": [ + 89, + 526, + 482, + 707 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {I T M}} = - \\sum_ {j} \\boldsymbol {y} _ {j} \\log \\mathrm {p} \\left(\\boldsymbol {c} _ {j} \\mid \\boldsymbol {x} _ {i}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 184, + 713, + 482, + 747 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\pmb{y}$ is the one-hot label, and", + "bbox": [ + 89, + 750, + 313, + 765 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {p} \\left(\\boldsymbol {c} _ {j} \\mid \\boldsymbol {x} _ {i}\\right) = \\frac {e ^ {< g \\left(\\boldsymbol {c} _ {j}\\right) , \\boldsymbol {f} _ {i} > / \\tau}}{\\sum_ {t = 1} ^ {M} e ^ {< g \\left(\\boldsymbol {c} _ {t}\\right) , \\boldsymbol {f} _ {i} > / \\tau}}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 173, + 768, + 482, + 808 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "$< \\cdot >$ denotes cosine similarity, $M$ is the number of classes, and $\\tau$ is the temperature learned by CLIP.", + "bbox": [ + 89, + 811, + 482, + 842 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Rethinking the FT Baseline", + "text_level": 1, + "bbox": [ + 89, + 848, + 338, + 864 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "While many PT schemes have reported improved performance and efficiency over the full fine-tuning (FT) baseline,", + "bbox": [ + 89, + 869, + 482, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "the discrepancy of the implementation details among those schemes obscures the actual performance enhancement.", + "bbox": [ + 511, + 359, + 903, + 390 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Comparison of PT and FT. To comprehensively evaluate the progress established by PT, we perform a comparison between the representative PT method CoOp [42] with the full FT baseline in terms of i) the number of learnable parameters, ii) memory usage, iii) time cost, and iv) baseto-new generalization results. The experimental setting is shown in Sup. Mat. (A). From the obtained results in Figure 2 (a), we observe that although PT significantly reduces the number of learnable parameters compared with FT (using 1/51200 parameters of FT), the improvements in memory efficiency and time efficiency are relatively insignificant (reducing only $6.3\\%$ memory usage and $15.8\\%$ time cost). Notably, in comparison to FT, the classification accuracy of PT decreases by $3.49\\%$ and $4.49\\%$ on base and new tasks, respectively. This means that pursuing higher parameter efficiency by freezing the vast majority of VLM weights during PT neither enhances the transferability of pre-trained knowledge nor improves memory and time efficiency significantly. In other words, existing PT methods involve a trade-off between parameter efficiency and performance.", + "bbox": [ + 511, + 391, + 905, + 691 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Cost Analysis of FT. The computational cost (i.e., memory and time cost) for fine-tuning VLMs is mainly caused by Feature-Gradient Propagation Flows (FGPFs)—the forward propagation flows of image/text features and the backward propagation flows of gradients. For example, on the CLIP model, each propagation flow is required to traverse all the $N$ layers of the vision encoder $E_{V}$ and the text encoder $E_{T}$ . For simplicity, we assume that $E_{V}$ and $E_{T}$ have identical network structures (e.g. ViT-B/16). Denote the number of class tokens as $M$ , and the costs of the FGPFs in the $l$ -th layer of $E_{V}$ and $E_{T}$ as $C_{V}$ and $C_{T}$ respectively. Thus, the total cost $C_{\\mathrm{total}} = N \\times (C_{V} + C_{T} \\times M^{1})$ . In this work,", + "bbox": [ + 511, + 693, + 906, + 875 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "For each training image, we need to obtain the text features of all $M$", + "bbox": [ + 531, + 886, + 903, + 898 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "14725", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "we refer to the number of network layers and the number of class tokens used for each training image as the length and width of FGPFs, respectively. In particular, $\\text{length} = N$ , $\\text{width} = M$ for existing PT methods, as shown in Figure 3. We therefore raise the following question:", + "bbox": [ + 89, + 90, + 483, + 167 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Can we reduce both the length and width of FGPFs without compromising the FT performance?", + "bbox": [ + 106, + 172, + 465, + 204 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Towards Effective and Efficient Knowledge Transfer. To answer the above question, we design two metrics of Feature Sensitivity (FS) and Gradient Dependence (GD) to estimate the contributions of different network layers and class tokens for each training image $\\mathbf{x}$ , respectively.", + "bbox": [ + 89, + 215, + 483, + 290 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "More specifically, let $\\pmb{f}_l$ and $\\pmb{f}_l'$ be a vision (or a text) feature before and after fine-tuning the $l$ -th layer of the CLIP's vision (or text) encoder, the FS of the $l$ -th layer for the training image $\\pmb{x}$ can be expressed as", + "bbox": [ + 89, + 291, + 483, + 351 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {F} \\mathbf {S} _ {l} (\\boldsymbol {x}) = \\psi \\left(\\boldsymbol {f} _ {l}, \\boldsymbol {f} _ {l} ^ {\\prime}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 214, + 363, + 482, + 380 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\psi (\\cdot)$ is the Euclidean distance. Hence, the bigger the $\\mathbf{FS}_l(\\pmb {x})$ value, the more important the $l$ -th layer for the input image $\\pmb{x}$ . Let $\\nabla f_{l}$ indicate the computed feature gradients passing through the last layer of the vision encoder for the training image $\\pmb{x}$ and $\\nabla f_l^{(c)}$ be the feature gradients after the $c$ -th class token is removed from Eq. 2. The GD of the $c$ -th class for $\\pmb{x}$ can be computed as", + "bbox": [ + 89, + 392, + 483, + 500 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {G D} _ {c} (\\boldsymbol {x}) = \\psi (\\nabla \\boldsymbol {f} _ {l}, \\nabla \\boldsymbol {f} _ {l} ^ {(c)}). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 189, + 512, + 480, + 531 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Therefore, the bigger the $\\mathbf{GD}_c(x)$ value, the more important the $c$ -th class for the input image $x$ .", + "bbox": [ + 89, + 542, + 482, + 571 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Figure 2 (b) illustrates the obtained FS values of different image/text encoder layers when adapting CLIP to the two datasets Caltech101 and Food101. As can be observed, the FS scores of the vast majority of shallow layers are close to 0, and only the last few layers contribute significantly to adapting CLIP to the two datasets. Moreover, Figure 2 (c) presents the frequency distributions of the GD values of class tokens for 100 training images randomly sampled from the two datasets. As shown, most class tokens contribute minimally to capturing task-specific knowledge from each training image. In a nutshell, the above observations reveal that we can achieve effective and efficient FT by filtering out unimportant network layers and class tokens in the text encoder for every training image.", + "bbox": [ + 89, + 573, + 483, + 785 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.3. Skip Tuning", + "text_level": 1, + "bbox": [ + 89, + 794, + 220, + 810 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Motivated by the previous section, we propose Skip Tuning, which performs Layer-wise Skipping (LSkip) and Classwise Skipping (CSkip) to adapt VLMs to downstream tasks in an effective and efficient manner, as shown in Figure 3.", + "bbox": [ + 89, + 816, + 482, + 878 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Layer-wise Skipping (LSkip). LSkip aims to reduce the length of FGPFs without compromising the performance of the FT baseline. To this end, LSkip first saves the image and text features before the $\\omega$ -th shallow layers of $E_{V}$ and $E_{T}$ in a cache, and then uses the saved intermediate features as input to update the parameters of the remaining $N - \\omega$ deep layers. Denote $\\pmb{v}_{i}^{\\omega}$ and $\\pmb{t}_{j}^{\\omega}$ as the extracted image and text features of the image $\\pmb{x}_i$ and the class token CLS ( $j = 1, \\dots, M$ ) in the $\\omega$ -th layers of $E_{V}$ and $E_{T}$ , respectively, i.e.,", + "bbox": [ + 511, + 90, + 906, + 227 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {v} _ {i} ^ {\\omega} = E _ {V} [ 1: \\omega ] (\\boldsymbol {x} _ {i}), \\boldsymbol {t} _ {j} ^ {\\omega} = E _ {T} [ 1: \\omega ] (\\boldsymbol {c} _ {j}), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 241, + 903, + 260 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $c_{j}$ is constructed using a hand-crafted prompt, e.g., $c_{j} = a$ photo of a [CLS].", + "bbox": [ + 511, + 265, + 903, + 294 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "After that, the $[1:\\omega]$ layers of both $E_{V}$ and $E_{T}$ are discarded. During FT, we input $\\pmb{v}_i^\\omega$ and $\\pmb{t}_j^\\omega$ to the remaining $N - \\omega$ deep layers of $E_{V}$ and $E_{T}$ , and obtain:", + "bbox": [ + 511, + 295, + 903, + 340 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\boldsymbol {v}} _ {i} ^ {N} = E _ {V} [ \\omega + 1: N ] (\\boldsymbol {v} _ {i} ^ {\\omega}), \\tilde {\\boldsymbol {t}} _ {j} ^ {N} = E _ {T} [ \\omega + 1: N ] (\\boldsymbol {t} _ {j} ^ {\\omega}), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 351, + 903, + 375 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "which are used to calculate the loss $\\mathcal{L}_{\\mathrm{ITM}}$ in Eq. 1. In this way, the image and text features, along with the calculated gradients, propagate through only $N - \\omega$ layers, effectively reducing the length of FGPFs.", + "bbox": [ + 511, + 383, + 903, + 443 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Class-wise Skipping (CSkip). Given a training image, the goal of CSkip is to filter out unimportant class tokens in the text encoder $E_{T}$ during the construction of the image-text matching loss $\\mathcal{L}_{\\mathrm{ITM}}$ (Eq. 1). Intuitively, the direct way is to select the top $k$ closest class tokens to compute the loss for each image, based on the similarities between the image feature and the text features of the $M$ inner-task classes. Nevertheless, we empirically find that this strategy leads to overfitting by selecting the same subset of classes for each image across different training epochs. Therefore, we propose an exponential image-conditioned class filtering strategy to overcome this limitation.", + "bbox": [ + 511, + 444, + 905, + 625 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Concretely, before FT begins, we use the text encoder $E_{T}$ to produce $M$ class features with a hand-crafted prompt, e.g. \"a photo of a [CLS]\". Then, we sort the cosine similarities between the vision feature of a training image and the $M$ class features in descending order. The probability of sampling the $j$ -th class token for the current training image can be expressed as:", + "bbox": [ + 511, + 625, + 905, + 731 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\np _ {j} = \\left\\{ \\begin{array}{l l} 1 & o _ {j} \\leq r \\times M, \\\\ \\zeta \\left(o _ {j} - r \\times M, \\lambda\\right) & o _ {j} > r \\times M, \\end{array} \\right. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 568, + 742, + 903, + 782 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\zeta (\\iota ,\\lambda) = e^{-\\lambda \\iota}$ , $\\lambda >0$ is the exponential decay coefficient, $r$ is the sampling ratio, and $o_{j}$ is the sorting index of the $j$ -th class. In this way, for each image, we can maintain the top $r\\times M$ closest class tokens while also sampling the remaining $(1 - r)\\times M$ classes with a certain probability for constructing $\\mathcal{L}_{\\mathrm{ITM}}$ , improving the diversity of training data at different epochs. Denote $m$ as the number of", + "bbox": [ + 511, + 794, + 906, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "classes to calculate the image-text matching loss, as illustrated in Eq. (2).", + "bbox": [ + 89, + 886, + 478, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "14726", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "class tokens sampled for the $i$ -th training image, we have $m \\ll M$ . This means we can flexibly reduce the width of FGPFs in the text encoder for every training image. Surprisingly, our experimental results in the next section reveal that ignoring the majority of class tokens for each training image can enhance the generalization performance of the learned model. This improvement may stem from CSkip's capability to filter out redundant and distracting text features when performing image-text matching with the loss of $\\mathcal{L}_{\\mathrm{ITM}}$ .", + "bbox": [ + 89, + 90, + 480, + 227 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3. Experiments", + "text_level": 1, + "bbox": [ + 89, + 243, + 222, + 258 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 89, + 268, + 279, + 285 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. We conduct experiments using 11 datasets from diverse sources. In particular, for base-to-new generalization, cross-dataset transfer and few-shot learning, we use 11 datasets including ImageNet [4], Caltech101 [5], OxfordPets [23], StanfordCars [19], Flowers102 [22], Food101 [2], FGVCAircraft [21], SUN397 [31], UCF101 [28], DTD [3] and EuroSAT [12]. For domain generalization, we use ImageNet [4] as the source dataset and its four variants as target datasets including ImageNetV2 [26], ImageNet-Sketch [30], ImageNet-A [14], and ImageNet-R [13].", + "bbox": [ + 89, + 292, + 482, + 443 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation Metric. We report base-task accuracy (\\%, denoted as Base), new-task accuracy (\\%, denoted as New), and their harmonic mean (\\%, denoted as $H$ ) to compare the performance/effectiveness of different methods. We also report the time cost (seconds, denoted as Time) and memory usage (M, denoted as Memory) for efficiency evaluation.", + "bbox": [ + 89, + 444, + 482, + 534 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation details. Our implementation of Skip Tuning is based on the open-source Github repository of DePT $[39]^2$ . In concrete terms, we leverage pre-trained ViT-B/16 as the backbone of the CLIP model. We employ the SGD optimizer to train the model with a learning rate of 2e-5 and a batch size of 4. By default, the number of training epochs is set to 20 for the base-to-new generalization, cross-dataset generalization, and domain generalization benchmarks, and 40 for the few-shot learning benchmark. The above hyperparameters are fixed across all datasets. We adjust the LSkip hyper-parameter $\\omega$ , CSkip hyper-parameters $r$ , and $\\lambda$ in § 3.3. All experimental results are the average of 3 runs with different seeds. We conduct experiments using an NVIDIA V100 GPU. For more details and additional results, please refer to Sup. Mat.", + "bbox": [ + 89, + 536, + 482, + 763 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2. Experimental Results", + "text_level": 1, + "bbox": [ + 89, + 775, + 292, + 790 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Base-to-New Generalization. The base-to-new generalization setting evaluates whether the models learned on a base task can generalize to new tasks with unseen classes. Following the comparison methods, for each dataset, we first construct a base task and a new task by equally dividing the", + "bbox": [ + 89, + 797, + 482, + 875 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "dataset into two sets of classes, then we perform prompt tuning on the base task and test the learned model on both the base and new tasks. Table 1 presents the obtained results of Skip Tuning and other state-of-the-art prompt tuning methods on 11 datasets. As shown, Skip Tuning achieves the best base-to-new generalization performance with the lowest memory usage and time cost. Concretely, Skip Tuning achieves $\\times 7.2$ time efficiency, and $\\times 5.1$ memory efficiency, while maintaining a $1.14\\%$ improvement in H ACC compared to the previous state-of-the-art method PromptSRC. This demonstrates the effectiveness and efficiency of our proposed Skip Tuning method. Moreover, we observe a tradeoff between base-task and new-task accuracies for most competitors. For instance, CoPrompt outperforms DePT on new tasks but lags behind DePT on base tasks. Notably, our Skip Tuning achieves the best performance on both base and new tasks simultaneously. This suggests that Skip Tuning effectively mitigates the overfitting issue when transferring VLMs to the base (or target) task.", + "bbox": [ + 511, + 90, + 903, + 378 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Cross-Dataset Generalization. The cross-dataset generalization setting assesses whether models trained on a source dataset/distribution can generalize to unseen target datasets/distributions. We follow the common setup of those comparison methods to use ImageNet as the source dataset and the other 10 datasets as target datasets. Table 2 presents the obtained cross-dataset generalization results of our Skip Tuning and other state-of-the-art methods on the source and target datasets. As seen, compared to the previous state-of-the-art method PromptSRC, our Skip Tuning achieves $\\times 44$ time efficiency, and $\\times 21.5$ memory efficiency, while establishing $1.5\\%$ and $1.19\\%$ ACC improvements on the source and target distributions respectively. From the average results, Skip Tuning consistently shows superior effectiveness and efficiency over the nine competitors on the 10 target datasets, without compromising the results of the tuned model on the source dataset. This demonstrates the effectiveness of our Skip Tuning scheme for improving the robustness of the tuned model to distribution shifts.", + "bbox": [ + 511, + 382, + 903, + 669 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Domain Generalization. The domain generalization setting assesses whether models trained on a source domain can generalize to unseen/target domains. In line with those comparison methods, we consider the ImageNet as the source domain and the other four ImageNet variants as target domains. As shown in Table 2, compared to the previous state-of-the-art method PromptSRC, Skip Tuning achieves $\\times 44$ time efficiency, and $\\times 21.5$ memory efficiency, while establishing $1.5\\%$ and $0.55\\%$ ACC improvements on the source and target domains, respectively. Besides, Skip Tuning consistently outperforms those competitors on the four target domains without compromising the performance of the tuned model on the source domain, which proves the effectiveness of our Skip Tuning scheme in improving the robustness of the tuned model to domain shifts.", + "bbox": [ + 511, + 674, + 903, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "2https://github.com/Koorye/DePT", + "bbox": [ + 107, + 886, + 352, + 898 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "14727", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/41c499b8023c36744ef9570e52218b5e1f0a89b7594f313d7d2b41e15be303a5.jpg", + "table_caption": [ + "Table 1. Base-to-new generalization results over 11 datasets. * indicates our reproduced results." + ], + "table_footnote": [], + "table_body": "
DatasetsMetricCoOp(IJCV'22)CoCoOp(CVPR'22)ProGrad(ICCV'23)KgCoOp(CVPR'23)MaPLe(CVPR'23)ProSRC(ICCV'23)TCP(CVPR'24)DePT(CVPR'24)CoPro.*ICLR'24)SkipT.(Ours)
ImgNetBase76.4775.9877.0275.8376.6677.6077.2777.0376.5377.73
New67.8870.4366.6669.9670.5470.7369.8770.1371.3070.40
H71.9273.1071.4672.7873.4774.0173.3873.4273.8273.89
CaltechBase98.0097.9698.0297.7297.7498.1098.2398.3098.6098.50
New89.3193.8193.8994.3994.3694.0394.6794.6095.1795.33
H93.7395.8495.9196.0396.0296.0296.4296.4196.8596.89
PetsBase93.6795.2095.0794.6595.4395.3394.6794.3394.7395.70
New95.2997.6997.6397.7697.7697.3097.2097.2396.7097.87
H94.4796.4396.3396.1896.5896.3095.9295.7695.7196.77
CarsBase78.1270.4977.6871.7672.9478.2780.8079.1373.1782.93
New60.4073.5968.6375.0474.0074.9774.1375.4770.6372.50
H68.1372.0172.8873.3673.4776.5877.3277.2671.8877.37
FlowersBase97.6094.8795.5495.0095.9298.0797.7398.0096.9398.57
New59.6771.7571.8774.7372.4676.5075.5776.3775.5075.80
H74.0681.7182.0383.6582.5685.9585.2385.8484.8885.70
Food101Base88.3390.7090.3790.5090.7190.6790.5790.5090.3790.67
New82.2691.2989.5991.7092.0591.5391.3791.6091.5392.03
H85.1990.9989.9891.0991.3891.1090.9791.0590.9591.34
AircraftBase40.4433.4140.5436.2137.4442.7341.9743.2036.1745.37
New22.3023.7127.5733.5535.6137.8734.4334.8334.4737.13
H28.7527.7432.8234.8336.5040.1537.8338.5735.3040.84
SUN397Base80.6079.7481.2680.2980.8282.6782.6382.3382.3082.40
New65.8976.8674.1776.5378.7078.4778.2077.8079.6379.03
H72.5178.5777.5578.3679.7580.5280.3580.0080.9480.68
DTDBase79.4477.0177.3577.5580.3683.3782.7782.2083.0083.77
New41.1856.0052.3554.9959.1862.9758.0759.1363.2067.23
H54.2464.8562.4564.3568.1671.1568.2568.7871.7674.59
EuroSATBase92.1987.4990.1185.6494.0792.9091.6389.0393.7792.47
New54.7460.0460.8964.3473.2373.9074.7371.0771.7383.00
H68.6971.2172.6773.4882.3582.3282.3279.0481.2887.48
UCF101Base84.6982.3384.3382.8983.0087.1087.1385.8086.2087.30
New56.0573.4574.9476.6778.6678.8080.7777.2378.7082.47
H67.4677.6479.3579.6580.7782.7483.8381.2982.2884.81
Avg ACCBase82.6980.4782.4880.7382.2884.2684.1383.6282.8985.04
New63.2271.6970.7573.6075.1476.1075.3675.0475.3277.53
H71.6675.8376.1677.0078.5579.9779.5179.1078.9381.11
CostTime (s)11862851231111914131735619733684239
Mem. (M)320415563204318817292041318917142060404
", + "bbox": [ + 93, + 95, + 903, + 503 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/e8655f0321216acb3873982f63c8b0e98d42e8a3cd22617412759849928662fe.jpg", + "table_caption": [ + "Table 2. Cross-dataset generalization results on 11 datasets. * indicates our reproduced results. The detailed results on the 10 target datasets (i.e., Caltech, Pets, Cars, ..., and UCF101) are reported in Sup. Mat. (B)." + ], + "table_footnote": [], + "table_body": "
DatasetsMetricCoOp(IJCV'22)CoCoOp(CVPR'22)ProGrad(ICCV'23)KgCoOp(CVPR'23)MaPLe(CVPR'23)ProSRC(ICCV'23)TCP(CVPR'24)DePT(CVPR'24)CoPro.*ICLR'24)SkipT.(Ours)
ImageNetACC71.5171.0272.2470.6670.7271.2771.4072.7772.5372.77
10 DatasetsAvg ACC63.8865.7462.7165.5166.3065.8166.2965.5564.8167.00
CostTime (s)316329391756223316364942500911617422796191611139
Mem. (M)154121362215412152541378614107152631378314131656
", + "bbox": [ + 93, + 539, + 903, + 609 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/3cf2cf37e06b8528bd83733040c050f941078fe294cde284827637bd93f982ba.jpg", + "table_caption": [ + "Table 3. Domain generalization results on ImageNet. * indicates our reproduced results. The detailed results on the 4 ImgNet variants (i.e., ImgNet-V2, ImgNet-S, ImgNet-A, and ImgNet-R) are presented in Sup. Mat. (B)." + ], + "table_footnote": [], + "table_body": "
DatasetsMetricCoOp(IJCV'22)CoCoOp(CVPR'22)ProGrad(ICCV'23)KgCoOp(CVPR'23)MaPLe(CVPR'23)ProSRC(ICCV'23)DePT(CVPR'24)CoPro.*ICLR'24)SkipT.(Ours)
ImageNetACC71.5171.0272.2470.6670.7271.2772.7772.5372.77
4 Imag.VariantsAvg ACC59.2859.9059.0760.1160.2660.6558.9760.0261.20
CostTime (s)3163293917562233163649425009122796191611139
Mem. (M)1541213622154121525413786141071378314131656
", + "bbox": [ + 94, + 647, + 903, + 727 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.3. Ablation Studies", + "text_level": 1, + "bbox": [ + 94, + 753, + 253, + 768 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we conduct ablative studies to further scrutinize our devised Skip Tuning method.", + "bbox": [ + 94, + 777, + 480, + 808 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Effectiveness of the Designed Components. Our proposed Skip Tuning approach simultaneously performs Layer-wise Skipping (LSkip) and Class-wise Skipping (CSkip) to enhance the memory and time efficiency of the full fine-tuning (FT) baseline. In this experiment, we investigate the effectiveness of LSkip and CSkip by gradually adding them to", + "bbox": [ + 94, + 810, + 480, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "the FT baseline. From the results in Table 4, we have the following observations. i) Both LSkip and CSkip contribute to performance improvement. ii) By combining all those components, Skip Tuning improves the effectiveness and efficiency of the baseline method remarkably. iii) The memory and time efficiency gains are more pronounced when CSkip is applied to the FT baseline, compared to the performance improvements it yields. iv) CSkip improves the memory and time efficiency of the baseline without com", + "bbox": [ + 516, + 753, + 903, + 890 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "14728", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a56986e33d9254718aeb5c25f4f9e6eb82161a84cc6b1c0e1be5bbfc2cca713b.jpg", + "image_caption": [ + "Figure 4. Ablation studies of the number of skipped layers $\\omega$ in LSkip, and the sampling rate $r$ , the decay coefficient $\\lambda$ in CSkip." + ], + "image_footnote": [], + "bbox": [ + 130, + 85, + 377, + 186 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2022ca518eb0cedc67eb1a75e69d5435497eec485a4e47994078f5a29d956172.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 379, + 85, + 620, + 186 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/71906f81d5a329d75f4129a50523326eba13a5f2dea8909a2dd299b98ac95c39.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 624, + 85, + 866, + 186 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/d94b00cc24022f0ca67dec70cb223ea97fc5aba37e0f56c576508e988c85a756.jpg", + "table_caption": [ + "Table 4. Ablation study on the components of Skip Tuning." + ], + "table_footnote": [], + "table_body": "
LSPCSPBaseNewHTime (s)Mem. (M)
XX84.9974.2679.2710021846
✓X85.0877.3081.00388861
X✓84.7774.4879.29440919
✓✓85.0477.5381.11239404
", + "bbox": [ + 93, + 237, + 480, + 306 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/b1ff6424a7120754521a237f67b46006407b29750468d8ccc2ca9ba985dec3ee.jpg", + "table_caption": [ + "Table 5. Comparison of our Skip Tuning and adapter-based methods. ${}^{ \\dagger }$ denotes the efficiency-optimized versions (details are illustrated in Sup. Mat. (C))." + ], + "table_footnote": [], + "table_body": "
MethodBaseNewHTime (s)Mem. (M)
FT84.9974.2679.2710021846
CLIP-adapter [10]74.4873.8174.148881784
LoRA [15]80.5374.7377.529101580
SkipT. (Ours)85.0477.5381.11239404
CLIP-adapter†76.7573.5675.12137175
SkipT. (Ours)†82.6675.5678.9511567
", + "bbox": [ + 91, + 363, + 488, + 443 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "promising the generalization performance of the learned model. One possible reason is that CSkip can filter out redundant and distracting text features when performing image-text matching with the loss of $\\mathcal{L}_{\\mathrm{ITM}}$ .", + "bbox": [ + 88, + 474, + 482, + 535 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Impact of the LSkip Hyper-parameter $\\omega$ . Our Skip Tuning approach drops out the $1 \\sim \\omega$ layers of the CLIP's vision and text encoders in the LSkip step. Here, we investigate the impact of $\\omega$ on performance by setting $\\omega$ to the values of $\\{2, 4, 6, 8, 10\\}$ . The obtained average testing results on the 11 datasets are reported in Figure 4 (Left). As shown, the obtained H ACCs gradually increase as the $\\omega$ value decreases from 10 to 6, after which the performance gradually decreases. But, we also see that as the value of $\\omega$ becomes smaller, both memory and time costs increase. Hence, we set $\\omega = 6$ for LSKip in this work.", + "bbox": [ + 88, + 536, + 482, + 702 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Impact of the CSkip Hyper-parameter $r$ . In the CSkip step of Skip Tuning, we devise an image-conditioned class sampling strategy to filter out irrelevant class tokens for each training image. The larger the sampling ratio $r$ value, the more class tokens will be used to construct the training loss $\\mathcal{L}_{\\mathrm{ITM}}$ for each image. It is necessary to scrutinize the impact of $r$ on performance. To this end, we respectively set $r$ to $\\{0.1, 0.3, 0.5, 0.7, 0.9, 1.0\\}$ , and report the average testing results on the 11 datasets in Figure 4 (Mid.). From the obtained H ACCs, our method is in general not sensitive to the change of $r$ within a certain range (from 0.4 to 1.0). We also see a gradual decrease in H ACC when $r > 0.7$ , suggesting that for each training image, not all class tokens are", + "bbox": [ + 89, + 703, + 482, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "beneficial for capturing task-specific knowledge. Besides, we can observe that as the value of $r$ becomes larger, both memory and time costs increase. Therefore, we set $r = 0.5$ for CSKip in this work.", + "bbox": [ + 511, + 231, + 903, + 291 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Impact of the CSkip Hyper-parameter $\\lambda$ . Our Skip Tuning introduces an exponential decay coefficient $\\lambda$ for the devised image-conditioned class sampling strategy in the CSkip step. We study the impact of $\\lambda$ by setting $\\lambda$ to the values of $\\{0.01, 0.1, 0.3, 0.5, 0.7, 0.9, 1.0\\}$ , and report the average testing results over the 11 datasets in Figure 4 (Right). As can be observed from the obtained H ACCs, our method is in general not sensitive to $\\lambda$ when it takes the values from 0.1 to 0.7. Particularly, the H ACCs gradually decrease as $\\lambda$ increases from 0.7 to 1.0. Also, we can see that as the value of $\\lambda$ becomes larger, both memory and time costs increase. Thus, we set $\\lambda = 0.3$ for CSkip in this work.", + "bbox": [ + 511, + 292, + 906, + 474 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.4. Additional Results", + "text_level": 1, + "bbox": [ + 511, + 484, + 692, + 500 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparison with Adapter-based Methods. Previous experimental results demonstrate the superior effectiveness and efficiency of our Skip Tuning method over existing prompt tuning methods. To further demonstrate the advantages of Skip Tuning, we also compare it with the representative adapter-based methods LoRA [15] and CLIP-adapter [10]. The base-to-new generalization results averaged on 11 datasets are reported in Table 5, where $\\mathrm{CLIP - adapter}^{\\dagger}$ refers to our re-implementation of CLIP-adapter, i.e., the CLIP's last-layer features are cached once before training and subsequently used to update the weights of the adapter module. As observed, our Skip Tuning offers substantial memory and time efficiency advantages over the two strong competitors, while also delivering better classification performance on both base and new tasks.", + "bbox": [ + 511, + 507, + 906, + 733 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Few-shot Learning. In the previous experiments, we follow the comparison approaches to evaluate the performance of different methods on $M$ -way 16-shot tasks—16 training examples are sampled for each of the $M$ inner-task classes. It is interesting to further scrutinize the effectiveness and efficiency of our Skip Tuning method under different shots. To this end, Figure 5 reports the few-shot learning performance of Skip Tuning and other representative competitors. As shown, our Skip Tuning demonstrates superior performance compared to other methods, with significant efficiency advantages in memory usage and time cost across", + "bbox": [ + 511, + 734, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "14729", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/637e120cc9d3b46297e22b25e0f6500bdd59376877ecbf88b530126884d8f322.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 122, + 89, + 277, + 186 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e9bfe42567c0d1563635c9c73f2a1bd7473003c2bfadd548124749a40be6a8a3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 282, + 89, + 426, + 186 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/67d756aca6943f05d057564ad0418be1cffe38783dc51765f56f909a56ed7eac.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 429, + 90, + 573, + 186 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8b81a9afd059df47794d2a9d1e4795e805587176820b4212d422702cc4336e3e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 578, + 90, + 720, + 186 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/07d8763f9a22b67de9073924347146d5951929ecd42ef56ed440df79c7d63dbd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 723, + 89, + 870, + 244 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/502e1bbe316477ec48714516e583d344c54d50599226b040f28e3f6d99205e92.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 122, + 189, + 277, + 286 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c2425b93941ba109e22b08ee13c3098cb0b2d01bca86c3a23ae8ea4b0bd5e0f1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 282, + 189, + 424, + 286 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d242e9023e80ef8bc6abb142d8fd40b0f759bd0aa6646d21cafe1e7f3395b714.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 429, + 189, + 573, + 286 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0d7a1a4ed196428b468f7216e5f216a82f9f116f8e6a1a28bc100e64ff07cf6b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 578, + 189, + 720, + 286 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/944be23fbc852114182e304c10950aaa0b722ea51a99ca2a1e710348a60681ff.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 723, + 250, + 870, + 395 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0aa2668f617df4861881d515c83de42b0ff64647b4ecaa361fdb0a51d9a9e2ff.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 122, + 290, + 277, + 393 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/71aa3663652bf038a1204914f6b1313570ca77b602e74ae056887d1ea09232f9.jpg", + "image_caption": [ + "Figure 5. Few-shot learning results on 11 datasets." + ], + "image_footnote": [], + "bbox": [ + 282, + 290, + 424, + 393 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1216882fa9d6ea2cfcaeba90ccbb4dd0a0b44eba3805d19a786992ad858d9f1e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 429, + 291, + 573, + 393 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/eb790603167035d504a6ba61adbe58e6d09da8ac941391b2ad7dff8bf71e3e84.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 578, + 291, + 720, + 393 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "1-shot to 16-shot settings. Additionally, while PromptSRC and MaPLe enhance CoOp's performance, both methods incur increased time cost. In contrast, Skip Tuning achieves both effectiveness and efficiency without compromise, further underscoring the superiority of our approach.", + "bbox": [ + 89, + 426, + 483, + 503 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4. Related Work", + "text_level": 1, + "bbox": [ + 89, + 523, + 232, + 537 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Pre-trained Vision-Language Models. Pre-trained vision-language models (VLMs) have garnered great attention recently. Notable models such as CLIP[24], ALIGN[16], LiT[36] and FILIP[34] have demonstrated exceptional performance across various vision-language tasks. These models are pre-trained on extensive datasets comprising image-text pairs sourced from the internet. For instance, CLIP[24] and ALIGN[16] respectively utilize over 400 million and 1 billion pairs, for training. The large scale of training data enables these models to excel in open-vocabulary image-text retrieval and zero-shot classification. Additionally, they have been successfully applied to downstream tasks including image classification[10, 40], object detection[6, 8, 11, 35, 43], and semantic segmentation[20, 25]. In this work, we focus on the effective adaptation of vision-language models to downstream visual recognition tasks.", + "bbox": [ + 89, + 550, + 482, + 792 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Prompt Tuning. Prompt tuning (a.k.a. context optimization [42]) has emerged as a parameter-efficient learning paradigm to adapt powerful VLMs to downstream tasks [7, 9, 37, 38]. As a representative method, CoOp [42] enables task adaptation by optimizing a set of prompt vectors within CLIP's language branch. While effective, CoOp often suffers from limited generalization on new tasks due to", + "bbox": [ + 89, + 795, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "overfitting to the base task. A series of schemes are devised to tackle this problem, e.g. CoCoOp[41], KgCoOp[32], ProGrad[44], TCP [33], and DePT[39]. By adding trainable prompts into both the image and text branches of CLIP, MaPLe[17], PromptSRC[18], and CoPrompt[27] demonstrate remarkable performance on both base and new tasks. Despite the advantages, we reveal that freezing the parameters of VLMs during learning the context vectors neither facilitates the generalization of pre-trained knowledge nor significantly improves memory and time efficiency.", + "bbox": [ + 511, + 426, + 906, + 579 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 598, + 633, + 614 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we first reveal that freezing the parameters of VLMs during prompt tuning neither facilitates the transferability of pre-trained knowledge nor improves memory and time efficiency considerably. To circumvent this limitation, we propose Skip Tuning, an effective and efficient method for transferring VLMs to downstream tasks without relying on extra context vectors or adapter modules. Extensive experiments across a broad range of benchmarks demonstrate the superiority of our Skip Tuning method over both prompt tuning and adapter-based approaches.", + "bbox": [ + 511, + 626, + 906, + 777 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements. This study is supported by grants from the National Natural Science Foundation of China (Grant No. 62425208, No. U22A2097, No. 62122018, No. 62020106008, No. U23A20315, No. 82441006, No. 62402094), the Postdoctoral Fellowship Program of CPSF (Grant No. GZB20240625), the Science and Technology Innovation Committee of Shenzhen Municipality Foundation (Grant No.JCYJ20240813114208012).", + "bbox": [ + 511, + 779, + 908, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "14730", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 1", + "[2] Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. Food-101-mining discriminative components with random forests. In Computer vision-ECCV 2014: 13th European conference, zurich, Switzerland, September 6-12, 2014, proceedings, part VI 13, pages 446-461. Springer, 2014. 5", + "[3] Mircea Cimpoi, Subhransu Maji, Iasonas Kokkinos, Sammy Mohamed, and Andrea Vedaldi. Describing textures in the wild. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3606-3613, 2014. 5", + "[4] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 5", + "[5] Li Fei-Fei, Rob Fergus, and Pietro Perona. Learning generative visual models from few training examples: An incremental bayesian approach tested on 101 object categories. Computer vision and Image understanding, 106(1):59-70, 2007. 5", + "[6] Chengjian Feng, Yujie Zhong, Zequn Jie, Xiangxiang Chu, Haibing Ren, Xiaolin Wei, Weidi Xie, and Lin Ma. Promptdet: Towards open-vocabulary detection using uncurated images. In European Conference on Computer Vision, pages 701-717. Springer, 2022. 8", + "[7] Yuqian Fu, Yanwei Fu, Jingjing Chen, and Yu-Gang Jiang. Generalized meta-fdmixup: Cross-domain few-shot learning guided by labeled target data. IEEE Transactions on Image Processing, 31:7078-7090, 2022. 8", + "[8] Yuqian Fu, Yu Xie, Yanwei Fu, and Yu-Gang Jiang. Styleadv: Meta style adversarial training for cross-domain few-shot learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 24575-24584, 2023. 8", + "[9] Yuqian Fu, Yu Wang, Yixuan Pan, Lian Huai, Xingyu Qiu, Zeyu Shangguan, Tong Liu, Yanwei Fu, Luc Van Gool, and Xingqun Jiang. Cross-domain few-shot object detection via enhanced open-set object detector. In ECCV, pages 247-264. Springer, 2025. 8", + "[10] Peng Gao, Shijie Geng, Renrui Zhang, Teli Ma, Rongyao Fang, Yongfeng Zhang, Hongsheng Li, and Yu Qiao. Clip-adapter: Better vision-language models with feature adapters. International Journal of Computer Vision, 132(2): 581-595, 2024. 7, 8", + "[11] Xiuye Gu, Tsung-Yi Lin, Weicheng Kuo, and Yin Cui. Open-vocabulary object detection via vision and language knowledge distillation. arXiv preprint arXiv:2104.13921, 2021.8", + "[12] Patrick Helber, Benjamin Bischke, Andreas Dengel, and Damian Borth. Eurosat: A novel dataset and deep learning benchmark for land use and land cover classification. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 12(7):2217-2226, 2019. 5" + ], + "bbox": [ + 93, + 114, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Dan Hendrycks, Steven Basart, Norman Mu, Saurav Kadavath, Frank Wang, Evan Dorundo, Rahul Desai, Tyler Zhu, Samyak Parajuli, Mike Guo, et al. The many faces of robustness: A critical analysis of out-of-distribution generalization. In Proceedings of the IEEE/CVF international conference on computer vision, pages 8340-8349, 2021. 5", + "[14] Dan Hendrycks, Kevin Zhao, Steven Basart, Jacob Steinhardt, and Dawn Song. Natural adversarial examples. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15262-15271, 2021. 5", + "[15] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 2, 7", + "[16] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International conference on machine learning, pages 4904-4916. PMLR, 2021. 8", + "[17] Muhammad Uzair Khattak, Hanoona Rasheed, Muhammad Maaz, Salman Khan, and Fahad Shahbaz Khan. Maple: Multi-modal prompt learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19113-19122, 2023. 8", + "[18] Muhammad Uzair Khattak, Syed Talal Wasim, Muzammal Naseer, Salman Khan, Ming-Hsuan Yang, and Fahad Shahbaz Khan. Self-regulating prompts: Foundational model adaptation without forgetting. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15190-15200, 2023. 8", + "[19] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In Proceedings of the IEEE international conference on computer vision workshops, pages 554–561, 2013. 5", + "[20] Timo Lüddecke and Alexander Ecker. Image segmentation using text and image prompts. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7086-7096, 2022. 8", + "[21] Subhransu Maji, Esa Rahtu, Juho Kannala, Matthew Blaschko, and Andrea Vedaldi. Fine-grained visual classification of aircraft. arXiv preprint arXiv:1306.5151, 2013. 5", + "[22] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In 2008 Sixth Indian conference on computer vision, graphics & image processing, pages 722-729. IEEE, 2008. 5", + "[23] Omkar M Parkhi, Andreae Vedaldi, Andrew Zisserman, and CV Jawahar. Cats and dogs. In 2012 IEEE conference on computer vision and pattern recognition, pages 3498-3505. IEEE, 2012. 5", + "[24] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 2, 8" + ], + "bbox": [ + 516, + 92, + 906, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "14731", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[25] Yongming Rao, Wenliang Zhao, Guangyi Chen, Yansong Tang, Zheng Zhu, Guan Huang, Jie Zhou, and Jiwen Lu. Denseclip: Language-guided dense prediction with context-aware prompting. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18082-18091, 2022. 8", + "[26] Benjamin Recht, Rebecca Roelofs, Ludwig Schmidt, and Vaishaal Shankar. Do imagenet classifiers generalize to imagenet? In International conference on machine learning, pages 5389-5400. PMLR, 2019. 5", + "[27] Shuvendu Roy and Ali Etemad. Consistency-guided prompt learning for vision-language models. arXiv preprint arXiv:2306.01195, 2023. 8", + "[28] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012. 5", + "[29] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 1", + "[30] Haohan Wang, Songwei Ge, Zachary Lipton, and Eric P Xing. Learning robust global representations by penalizing local predictive power. Advances in Neural Information Processing Systems, 32, 2019. 5", + "[31] Jianxiong Xiao, James Hays, Krista A Ehinger, Aude Oliva, and Antonio Torralba. Sun database: Large-scale scene recognition from abbey to zoo. In 2010 IEEE computer society conference on computer vision and pattern recognition, pages 3485-3492. IEEE, 2010. 5", + "[32] Hantao Yao, Rui Zhang, and Changsheng Xu. Visual-language prompt tuning with knowledge-guided context optimization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6757-6767, 2023. 8", + "[33] Hantao Yao, Rui Zhang, and Changsheng Xu. Tcp: Textual-based class-aware prompt tuning for visual-language model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23438-23448, 2024. 8", + "[34] Lewei Yao, Runhui Huang, Lu Hou, Guansong Lu, Minzhe Niu, Hang Xu, Xiaodan Liang, Zhenguo Li, Xin Jiang, and Chunjing Xu. Filip: Fine-grained interactive language-image pre-training. arXiv preprint arXiv:2111.07783, 2021. 8", + "[35] Yuhang Zang, Wei Li, Kaiyang Zhou, Chen Huang, and Chen Change Loy. Open-vocabulary detr with conditional matching. In European Conference on Computer Vision, pages 106-122. Springer, 2022. 8", + "[36] Xiaohua Zhai, Xiao Wang, Basil Mustafa, Andreas Steiner, Daniel Keysers, Alexander Kolesnikov, and Lucas Beyer. Lit: Zero-shot transfer with locked-image text tuning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18123-18133, 2022. 8", + "[37] Ji Zhang, Jingkuan Song, Lianli Gao, and Hengtao Shen. Free-lunch for cross-domain few-shot learning: Style-aware episodic training with robust contrastive learning. In Proceedings of the 30th ACM International Conference on Multimedia, pages 2586-2594, 2022. 8" + ], + "bbox": [ + 91, + 92, + 480, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[38] Ji Zhang, Lianli Gao, Xu Luo, Hengtao Shen, and Jingkuan Song. Deta: Denoised task adaptation for few-shot learning. In ICCV, pages 11541-11551, 2023. 8", + "[39] Ji Zhang, Shihan Wu, Lianli Gao, Heng Tao Shen, and Jingkuan Song. Dept: Decoupled prompt tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12924-12933, 2024. 5, 8", + "[40] Renrui Zhang, Rongyao Fang, Wei Zhang, Peng Gao, Kunchang Li, Jifeng Dai, Yu Qiao, and Hongsheng Li. Tip-adapter: Training-free clip-adapter for better vision-language modeling. arXiv preprint arXiv:2111.03930, 2021. 8", + "[41] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Conditional prompt learning for vision-language models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16816-16825, 2022. 8", + "[42] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision, 130(9):2337-2348, 2022. 2, 3, 8", + "[43] Xingyi Zhou, Rohit Girdhar, Armand Joulin, Philipp Krahenbihl, and Ishan Misra. Detecting twenty-thousand classes using image-level supervision. In European Conference on Computer Vision, pages 350-368. Springer, 2022. 8", + "[44] Beier Zhu, Yulei Niu, Yucheng Han, Yue Wu, and Hanwang Zhang. Prompt-aligned gradient for prompt tuning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15659-15669, 2023. 8" + ], + "bbox": [ + 516, + 92, + 903, + 515 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "14732", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Skip Tuning_ Pre-trained Vision-Language Models are Effective and Efficient Adapters Themselves/e5453c9b-356b-43d9-9454-788a3f8b346d_model.json b/2025/Skip Tuning_ Pre-trained Vision-Language Models are Effective and Efficient Adapters Themselves/e5453c9b-356b-43d9-9454-788a3f8b346d_model.json new file mode 100644 index 0000000000000000000000000000000000000000..f6b5e8d476196fb59d3ba24cbcb7fa864e9011ea --- /dev/null +++ b/2025/Skip Tuning_ Pre-trained Vision-Language Models are Effective and Efficient Adapters Themselves/e5453c9b-356b-43d9-9454-788a3f8b346d_model.json @@ -0,0 +1,2277 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.223, + 0.131, + 0.777, + 0.177 + ], + "angle": 0, + "content": "Skip Tuning: Pre-trained Vision-Language Models are Effective and Efficient Adapters Themselves" + }, + { + "type": "text", + "bbox": [ + 0.115, + 0.202, + 0.884, + 0.223 + ], + "angle": 0, + "content": "Shihan Wu\\(^{1}\\) Ji Zhang\\(^{2*}\\) Pengpeng Zeng\\(^{3}\\) Lianli Gao\\(^{3}\\) Jingkuan Song\\(^{4}\\) Heng Tao Shen\\(^{4}\\)" + }, + { + "type": "text", + "bbox": [ + 0.223, + 0.223, + 0.775, + 0.24 + ], + "angle": 0, + "content": "1 University of Electronic Science and Technology of China (UESTC)" + }, + { + "type": "text", + "bbox": [ + 0.371, + 0.239, + 0.627, + 0.258 + ], + "angle": 0, + "content": "\\(^{2}\\) Southwest Jiaotong University" + }, + { + "type": "text", + "bbox": [ + 0.302, + 0.257, + 0.696, + 0.275 + ], + "angle": 0, + "content": "3 Shenzhen Institute for Advanced Study, UESTC" + }, + { + "type": "text", + "bbox": [ + 0.423, + 0.275, + 0.576, + 0.293 + ], + "angle": 0, + "content": "\\(^{4}\\) Tongji University" + }, + { + "type": "text", + "bbox": [ + 0.27, + 0.295, + 0.727, + 0.31 + ], + "angle": 0, + "content": "{shihan.wu.koorye@outlook.com, jizhang.jim@gmail.com}" + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.345, + 0.327, + 0.361 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.378, + 0.485, + 0.695 + ], + "angle": 0, + "content": "Prompt tuning (PT) has long been recognized as an effective and efficient paradigm for transferring large pre-trained vision-language models (VLMs) to downstream tasks by learning a tiny set of context vectors. Nevertheless, in this work, we reveal that freezing the parameters of VLMs during learning the context vectors neither facilitates the transferability of pre-trained knowledge nor improves the memory and time efficiency significantly. Upon further investigation, we find that reducing both the length and width of the feature-gradient propagation flows of the full fine-tuning (FT) baseline is key to achieving effective and efficient knowledge transfer. Motivated by this, we propose Skip Tuning, a novel paradigm for adapting VLMs to downstream tasks. Unlike existing PT or adapter-based methods, Skip Tuning applies Layer-wise Skipping (LSkip) and Classwise Skipping (CSkip) upon the FT baseline without introducing extra context vectors or adapter modules. Extensive experiments across a wide spectrum of benchmarks demonstrate the superior effectiveness and efficiency of our Skip Tuning over both PT and adapter-based methods. Code: https://github.com/Koorye/SkipTuning." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.727, + 0.222, + 0.743 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.753, + 0.484, + 0.875 + ], + "angle": 0, + "content": "There have recently been significant advancements in large pre-trained vision-language models (VLMs) [1, 24, 29]. One notable achievement is the CLIP model [24], which leverages an image-text matching (ITM) loss to align images with their corresponding textual descriptions in a common feature space. While VLMs have proven impressive capabilities in recognizing open-set visual concepts, their zero-shot generalization performance declines significantly" + }, + { + "type": "image", + "bbox": [ + 0.532, + 0.342, + 0.683, + 0.424 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.686, + 0.343, + 0.786, + 0.423 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.79, + 0.343, + 0.89, + 0.424 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.53, + 0.426, + 0.683, + 0.507 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.686, + 0.426, + 0.784, + 0.508 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.79, + 0.426, + 0.89, + 0.508 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.531, + 0.51, + 0.683, + 0.584 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.686, + 0.51, + 0.785, + 0.584 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.79, + 0.51, + 0.89, + 0.584 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.531, + 0.586, + 0.683, + 0.635 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.686, + 0.587, + 0.785, + 0.635 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.79, + 0.587, + 0.89, + 0.635 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.638, + 0.907, + 0.735 + ], + "angle": 0, + "content": "Figure 1. Comparison of our devised Skip Tuning with state-of-the-art prompt tuning methods in terms of training time (seconds), memory cost (M), and classification accuracy (\\%) across base-to-new generalization, cross-dataset generalization, domain generalization, and few-shot learning benchmarks. \\(\\times\\) indicates the performance improvement over the state-of-the-art. Comparison results with the adapter-based methods are reported in Table 5." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.764, + 0.907, + 0.794 + ], + "angle": 0, + "content": "when encountering category, distribution, or domain shifts between upstream training data and downstream tasks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Prompt tuning (PT) has long been recognized as an effective and efficient paradigm for transferring large pre-trained vision-language models (VLMs) to downstream tasks. The core concept of PT is to learn a task-specific prompt (i.e., a small number of context vectors) for the target task, using a limited amount of training data, while keeping the pre-trained VLM parameters fixed. Although many PT ap" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.236, + 0.9 + ], + "angle": 0, + "content": "*Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14723" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.127, + 0.088, + 0.291, + 0.167 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.297, + 0.089, + 0.482, + 0.168 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.173, + 0.291, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.297, + 0.173, + 0.482, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.256, + 0.291, + 0.331 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.297, + 0.256, + 0.482, + 0.331 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.334, + 0.483, + 0.431 + ], + "angle": 0, + "content": "Figure 2. Motivations. (a) Comparison between the prompt tuning (PT) method CoOp [42] and the full fine-tuning (FT) baseline in terms of i) the number of learnable parameters, ii) memory usage, iii) time cost, and iv) base-to-new generalization performance. (b) Feature Sensitivity (FS) of CLIP's network layers, averaged over 100 randomly-sampled training images. (c) Gradient Dependence (GD) of class tokens for different training images." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.469, + 0.483, + 0.846 + ], + "angle": 0, + "content": "proaches have reported improved performance and parameter efficiency over the full fine-tuning (FT) baseline, the discrepancy of implementation details among those PT approaches obscure the actual performance enhancement. For example, the FT performance can be significantly underestimated by training with coarsely tuned hyper-parameters. Quantitative evidence is presented in Figure 2 (a), where we make comparisons between the PT method CoOp [42] and the FT baseline in terms of the number of learnable parameters, memory usage, time cost as well as base-to-new generalization performance (see §2.2). As observed, although PT significantly improves the parameter efficiency of FT (using 1/51200 parameters of FT), the improvements in memory and time efficiency are relatively insignificant (reducing only \\(6.3\\%\\) memory usage and \\(15.8\\%\\) time cost). Besides, compared to FT, the classification accuracy of PT decreases by \\(3.49\\%\\) and \\(4.49\\%\\) on base and new tasks, respectively. This suggests that pursuing higher parameter efficiency by freezing the overwhelming majority weights of VLMs during learning the context vectors neither facilitates the transferability of pre-trained knowledge nor improves memory and time efficiency considerably. Besides, in many real-world applications, memory and time efficiency often take precedence over parameter efficiency in terms of practical importance. We therefore raise the following question:" + }, + { + "type": "text", + "bbox": [ + 0.107, + 0.853, + 0.466, + 0.899 + ], + "angle": 0, + "content": "Can we optimize the memory and time efficiency of the FT baseline and adapt VLMs to downstream tasks in an effective and efficient manner?" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.318 + ], + "angle": 0, + "content": "To answer the above question, we scrutinize the FeatureGradient Propagation Flows (FGPFs) in the vision encoder and text encoder of the CLIP model when performing FT on the base (or target) task. Interestingly, we observe that, for each training image, the majority of shallow network layers and class tokens contribute minimally to capturing task-specific knowledge for the base task (see §2.2). Motivated by this, we propose Skip Tuning, a novel paradigm for adapting VLMs to downstream tasks without introducing extra context vectors or adapter modules. Concretely, Skip Tuning incorporates two strategies, Layer-wise Skipping (LSkip) and Class-wise Skipping (CSkip), to simultaneously reduce the length and width of FGFs in the FT baseline, thereby establishing effective and efficient knowledge transfer of VLMs, as shown in Figure 3." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.318, + 0.905, + 0.514 + ], + "angle": 0, + "content": "Effectiveness and Efficiency. We conduct extensive experiments across a wide range of benchmarks to validate the effectiveness and efficiency of our Skip Tuning scheme. An overview of the achieved results is shown in Figure 1. As seen, our Skip Tuning demonstrates superiority over existing PT methods, e.g., on the few-shot learning benchmark, Skip Tuning achieves \\(\\times 15\\) time efficiency, \\(\\times 6.4\\) memory efficiency, while yielding a \\(1.04\\%\\) improvement in ACC compared to the state-of-the-art. Furthermore, we also show the advantages of Skip Tuning over existing adapter-based methods in Table 5, where Skip Tuning achieves \\(\\times 3.8\\) time efficiency, \\(\\times 3.9\\) memory efficiency, along with a \\(3.59\\%\\) H ACC enhancement over the strong rival LoRA [15]." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.515, + 0.904, + 0.528 + ], + "angle": 0, + "content": "Main Contributions. The main contributions are threefold." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.53, + 0.905, + 0.589 + ], + "angle": 0, + "content": "- We reveal that reducing both the width and length of the feature-gradient propagation flows (FGPFs) of the full fine-tuning (FT) baseline is key to establishing effective and efficient knowledge transfer." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.59, + 0.905, + 0.635 + ], + "angle": 0, + "content": "- We devise Skip Tuning, an effective and efficient method for transferring VLMs to downstream tasks without relying on extra context vectors or adapter modules." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.635, + 0.905, + 0.681 + ], + "angle": 0, + "content": "- We evaluate our method on a wide spectrum of benchmarks, demonstrating the superiority of Skip Tuning over both prompt tuning and adapter-based approaches." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.53, + 0.905, + 0.681 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.694, + 0.649, + 0.71 + ], + "angle": 0, + "content": "2. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.719, + 0.904, + 0.75 + ], + "angle": 0, + "content": "In this section, we elaborate on our Skip Tuning approach. We start with an introduction of preliminaries." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.758, + 0.655, + 0.772 + ], + "angle": 0, + "content": "2.1. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Contrastive Language-Image Pre-training (CLIP). Following the common practice of existing paradigms for adapting PVLMs, we adopt CLIP [24] as the testbed in this work. CLIP aims to learn an alignment between image and text features generated by an image encoder and a text encoder, respectively. By being exposed to 400 million imagetext association pairs and employing a contrastive learning paradigm within a shared feature space, CLIP acquires a di" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14724" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.087, + 0.905, + 0.273 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.278, + 0.908, + 0.337 + ], + "angle": 0, + "content": "Figure 3. Overview of our proposed Skip Tuning. Skip Tuning performs Layer-wise Skipping (LSkip) and Class-wise Skipping (CSkip)) to enhance the memory and time efficiency of the FT baseline. Specifically, LSkip reduces the length of feature-gradient propagation flows (FGPFs) by caching intermediate features produced by the \\(\\omega\\)-th layers of CLIP's vision encoder \\(E_{V}\\) and text encoder \\(E_{T}\\) before FT begins. In contrast, CSkip reduces the width of FGFs by filtering out unimportant class tokens in the text encoder \\(E_{T}\\) for every training image." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.361, + 0.482, + 0.526 + ], + "angle": 0, + "content": "verse array of open-set visual concepts that can be readily applied to downstream tasks. For instance, zero-shot classification can be achieved by framing the classification task as an image-text matching problem. Initially, a prompt (e.g. \"a photo of a\") is crafted to extract text features of all intra-task classes by presenting the class-extended prompt (\"a photo of a [CLS]\") to the text encoder. Subsequently, the image encoder is utilized to derive the image feature of a given input example, and class prediction is performed by comparing cosine distances between the image feature and text features of classes." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.527, + 0.483, + 0.708 + ], + "angle": 0, + "content": "Prompt Tuning with an Image-Text Matching Loss. PT aims to learn a task-specific prompt with few-labeled samples from the base task. Concretely, a prompt is formulated as \\(\\kappa\\) context vectors: \\(\\{v_{1}, v_{2}, \\dots, v_{\\kappa}\\}\\). During training, we produce the text feature of the \\(j\\)-th class by inputting the class-token-extended prompt \\(c_{j} = \\{v_{1}, v_{2}, \\dots, v_{\\kappa}, [\\mathrm{CLS}]\\}\\) to the text encoder \\(g(\\cdot)\\), where [CLS] indicates the class token of the \\(j\\)-th class. Denote \\(\\pmb{f}_{i} \\in \\mathbb{R}^{d}\\) as the image feature of a sample \\(\\pmb{x}_{i}\\) extracted by the image encoder, the task-specific prompt can be updated via back-propagating the image-text matching (ITM) loss through the frozen CLIP model. The ITM loss can be expressed as" + }, + { + "type": "equation", + "bbox": [ + 0.185, + 0.714, + 0.483, + 0.748 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {I T M}} = - \\sum_ {j} \\boldsymbol {y} _ {j} \\log \\mathrm {p} \\left(\\boldsymbol {c} _ {j} \\mid \\boldsymbol {x} _ {i}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.751, + 0.314, + 0.766 + ], + "angle": 0, + "content": "where \\(\\pmb{y}\\) is the one-hot label, and" + }, + { + "type": "equation", + "bbox": [ + 0.174, + 0.77, + 0.483, + 0.809 + ], + "angle": 0, + "content": "\\[\n\\mathrm {p} \\left(\\boldsymbol {c} _ {j} \\mid \\boldsymbol {x} _ {i}\\right) = \\frac {e ^ {< g \\left(\\boldsymbol {c} _ {j}\\right) , \\boldsymbol {f} _ {i} > / \\tau}}{\\sum_ {t = 1} ^ {M} e ^ {< g \\left(\\boldsymbol {c} _ {t}\\right) , \\boldsymbol {f} _ {i} > / \\tau}}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.812, + 0.483, + 0.843 + ], + "angle": 0, + "content": "\\(< \\cdot >\\) denotes cosine similarity, \\(M\\) is the number of classes, and \\(\\tau\\) is the temperature learned by CLIP." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.849, + 0.339, + 0.866 + ], + "angle": 0, + "content": "2.2. Rethinking the FT Baseline" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.902 + ], + "angle": 0, + "content": "While many PT schemes have reported improved performance and efficiency over the full fine-tuning (FT) baseline," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.361, + 0.905, + 0.391 + ], + "angle": 0, + "content": "the discrepancy of the implementation details among those schemes obscures the actual performance enhancement." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.392, + 0.906, + 0.693 + ], + "angle": 0, + "content": "Comparison of PT and FT. To comprehensively evaluate the progress established by PT, we perform a comparison between the representative PT method CoOp [42] with the full FT baseline in terms of i) the number of learnable parameters, ii) memory usage, iii) time cost, and iv) baseto-new generalization results. The experimental setting is shown in Sup. Mat. (A). From the obtained results in Figure 2 (a), we observe that although PT significantly reduces the number of learnable parameters compared with FT (using 1/51200 parameters of FT), the improvements in memory efficiency and time efficiency are relatively insignificant (reducing only \\(6.3\\%\\) memory usage and \\(15.8\\%\\) time cost). Notably, in comparison to FT, the classification accuracy of PT decreases by \\(3.49\\%\\) and \\(4.49\\%\\) on base and new tasks, respectively. This means that pursuing higher parameter efficiency by freezing the vast majority of VLM weights during PT neither enhances the transferability of pre-trained knowledge nor improves memory and time efficiency significantly. In other words, existing PT methods involve a trade-off between parameter efficiency and performance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.694, + 0.907, + 0.876 + ], + "angle": 0, + "content": "Cost Analysis of FT. The computational cost (i.e., memory and time cost) for fine-tuning VLMs is mainly caused by Feature-Gradient Propagation Flows (FGPFs)—the forward propagation flows of image/text features and the backward propagation flows of gradients. For example, on the CLIP model, each propagation flow is required to traverse all the \\(N\\) layers of the vision encoder \\(E_{V}\\) and the text encoder \\(E_{T}\\). For simplicity, we assume that \\(E_{V}\\) and \\(E_{T}\\) have identical network structures (e.g. ViT-B/16). Denote the number of class tokens as \\(M\\), and the costs of the FGPFs in the \\(l\\)-th layer of \\(E_{V}\\) and \\(E_{T}\\) as \\(C_{V}\\) and \\(C_{T}\\) respectively. Thus, the total cost \\(C_{\\mathrm{total}} = N \\times (C_{V} + C_{T} \\times M^{1})\\). In this work," + }, + { + "type": "page_footnote", + "bbox": [ + 0.532, + 0.887, + 0.905, + 0.9 + ], + "angle": 0, + "content": "For each training image, we need to obtain the text features of all \\(M\\)" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14725" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.168 + ], + "angle": 0, + "content": "we refer to the number of network layers and the number of class tokens used for each training image as the length and width of FGPFs, respectively. In particular, \\( \\text{length} = N \\), \\( \\text{width} = M \\) for existing PT methods, as shown in Figure 3. We therefore raise the following question:" + }, + { + "type": "text", + "bbox": [ + 0.107, + 0.174, + 0.466, + 0.205 + ], + "angle": 0, + "content": "Can we reduce both the length and width of FGPFs without compromising the FT performance?" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.216, + 0.484, + 0.291 + ], + "angle": 0, + "content": "Towards Effective and Efficient Knowledge Transfer. To answer the above question, we design two metrics of Feature Sensitivity (FS) and Gradient Dependence (GD) to estimate the contributions of different network layers and class tokens for each training image \\( \\mathbf{x} \\), respectively." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.292, + 0.484, + 0.352 + ], + "angle": 0, + "content": "More specifically, let \\( \\pmb{f}_l \\) and \\( \\pmb{f}_l' \\) be a vision (or a text) feature before and after fine-tuning the \\( l \\)-th layer of the CLIP's vision (or text) encoder, the FS of the \\( l \\)-th layer for the training image \\( \\pmb{x} \\) can be expressed as" + }, + { + "type": "equation", + "bbox": [ + 0.215, + 0.364, + 0.483, + 0.381 + ], + "angle": 0, + "content": "\\[\n\\mathbf {F} \\mathbf {S} _ {l} (\\boldsymbol {x}) = \\psi \\left(\\boldsymbol {f} _ {l}, \\boldsymbol {f} _ {l} ^ {\\prime}\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.393, + 0.484, + 0.501 + ], + "angle": 0, + "content": "where \\(\\psi (\\cdot)\\) is the Euclidean distance. Hence, the bigger the \\(\\mathbf{FS}_l(\\pmb {x})\\) value, the more important the \\(l\\)-th layer for the input image \\(\\pmb{x}\\). Let \\(\\nabla f_{l}\\) indicate the computed feature gradients passing through the last layer of the vision encoder for the training image \\(\\pmb{x}\\) and \\(\\nabla f_l^{(c)}\\) be the feature gradients after the \\(c\\)-th class token is removed from Eq. 2. The GD of the \\(c\\)-th class for \\(\\pmb{x}\\) can be computed as" + }, + { + "type": "equation", + "bbox": [ + 0.19, + 0.513, + 0.482, + 0.532 + ], + "angle": 0, + "content": "\\[\n\\mathbf {G D} _ {c} (\\boldsymbol {x}) = \\psi (\\nabla \\boldsymbol {f} _ {l}, \\nabla \\boldsymbol {f} _ {l} ^ {(c)}). \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.543, + 0.483, + 0.573 + ], + "angle": 0, + "content": "Therefore, the bigger the \\(\\mathbf{GD}_c(x)\\) value, the more important the \\(c\\)-th class for the input image \\(x\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.574, + 0.484, + 0.786 + ], + "angle": 0, + "content": "Figure 2 (b) illustrates the obtained FS values of different image/text encoder layers when adapting CLIP to the two datasets Caltech101 and Food101. As can be observed, the FS scores of the vast majority of shallow layers are close to 0, and only the last few layers contribute significantly to adapting CLIP to the two datasets. Moreover, Figure 2 (c) presents the frequency distributions of the GD values of class tokens for 100 training images randomly sampled from the two datasets. As shown, most class tokens contribute minimally to capturing task-specific knowledge from each training image. In a nutshell, the above observations reveal that we can achieve effective and efficient FT by filtering out unimportant network layers and class tokens in the text encoder for every training image." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.795, + 0.222, + 0.811 + ], + "angle": 0, + "content": "2.3. Skip Tuning" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.818, + 0.483, + 0.879 + ], + "angle": 0, + "content": "Motivated by the previous section, we propose Skip Tuning, which performs Layer-wise Skipping (LSkip) and Classwise Skipping (CSkip) to adapt VLMs to downstream tasks in an effective and efficient manner, as shown in Figure 3." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.907, + 0.228 + ], + "angle": 0, + "content": "Layer-wise Skipping (LSkip). LSkip aims to reduce the length of FGPFs without compromising the performance of the FT baseline. To this end, LSkip first saves the image and text features before the \\(\\omega\\)-th shallow layers of \\(E_{V}\\) and \\(E_{T}\\) in a cache, and then uses the saved intermediate features as input to update the parameters of the remaining \\(N - \\omega\\) deep layers. Denote \\(\\pmb{v}_{i}^{\\omega}\\) and \\(\\pmb{t}_{j}^{\\omega}\\) as the extracted image and text features of the image \\(\\pmb{x}_i\\) and the class token CLS (\\(j = 1, \\dots, M\\)) in the \\(\\omega\\)-th layers of \\(E_{V}\\) and \\(E_{T}\\), respectively, i.e.," + }, + { + "type": "equation", + "bbox": [ + 0.568, + 0.242, + 0.905, + 0.261 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {v} _ {i} ^ {\\omega} = E _ {V} [ 1: \\omega ] (\\boldsymbol {x} _ {i}), \\boldsymbol {t} _ {j} ^ {\\omega} = E _ {T} [ 1: \\omega ] (\\boldsymbol {c} _ {j}), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.266, + 0.905, + 0.295 + ], + "angle": 0, + "content": "where \\( c_{j} \\) is constructed using a hand-crafted prompt, e.g., \\( c_{j} = a \\) photo of a [CLS]." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.296, + 0.905, + 0.341 + ], + "angle": 0, + "content": "After that, the \\([1:\\omega]\\) layers of both \\(E_{V}\\) and \\(E_{T}\\) are discarded. During FT, we input \\(\\pmb{v}_i^\\omega\\) and \\(\\pmb{t}_j^\\omega\\) to the remaining \\(N - \\omega\\) deep layers of \\(E_{V}\\) and \\(E_{T}\\), and obtain:" + }, + { + "type": "equation", + "bbox": [ + 0.521, + 0.352, + 0.905, + 0.375 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\boldsymbol {v}} _ {i} ^ {N} = E _ {V} [ \\omega + 1: N ] (\\boldsymbol {v} _ {i} ^ {\\omega}), \\tilde {\\boldsymbol {t}} _ {j} ^ {N} = E _ {T} [ \\omega + 1: N ] (\\boldsymbol {t} _ {j} ^ {\\omega}), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.384, + 0.905, + 0.444 + ], + "angle": 0, + "content": "which are used to calculate the loss \\(\\mathcal{L}_{\\mathrm{ITM}}\\) in Eq. 1. In this way, the image and text features, along with the calculated gradients, propagate through only \\(N - \\omega\\) layers, effectively reducing the length of FGPFs." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.445, + 0.906, + 0.625 + ], + "angle": 0, + "content": "Class-wise Skipping (CSkip). Given a training image, the goal of CSkip is to filter out unimportant class tokens in the text encoder \\( E_{T} \\) during the construction of the image-text matching loss \\( \\mathcal{L}_{\\mathrm{ITM}} \\) (Eq. 1). Intuitively, the direct way is to select the top \\( k \\) closest class tokens to compute the loss for each image, based on the similarities between the image feature and the text features of the \\( M \\) inner-task classes. Nevertheless, we empirically find that this strategy leads to overfitting by selecting the same subset of classes for each image across different training epochs. Therefore, we propose an exponential image-conditioned class filtering strategy to overcome this limitation." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.626, + 0.906, + 0.732 + ], + "angle": 0, + "content": "Concretely, before FT begins, we use the text encoder \\( E_{T} \\) to produce \\( M \\) class features with a hand-crafted prompt, e.g. \"a photo of a [CLS]\". Then, we sort the cosine similarities between the vision feature of a training image and the \\( M \\) class features in descending order. The probability of sampling the \\( j \\)-th class token for the current training image can be expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.57, + 0.743, + 0.905, + 0.784 + ], + "angle": 0, + "content": "\\[\np _ {j} = \\left\\{ \\begin{array}{l l} 1 & o _ {j} \\leq r \\times M, \\\\ \\zeta \\left(o _ {j} - r \\times M, \\lambda\\right) & o _ {j} > r \\times M, \\end{array} \\right. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.907, + 0.901 + ], + "angle": 0, + "content": "where \\(\\zeta (\\iota ,\\lambda) = e^{-\\lambda \\iota}\\), \\(\\lambda >0\\) is the exponential decay coefficient, \\(r\\) is the sampling ratio, and \\(o_{j}\\) is the sorting index of the \\(j\\)-th class. In this way, for each image, we can maintain the top \\(r\\times M\\) closest class tokens while also sampling the remaining \\((1 - r)\\times M\\) classes with a certain probability for constructing \\(\\mathcal{L}_{\\mathrm{ITM}}\\), improving the diversity of training data at different epochs. Denote \\(m\\) as the number of" + }, + { + "type": "page_footnote", + "bbox": [ + 0.091, + 0.887, + 0.479, + 0.901 + ], + "angle": 0, + "content": "classes to calculate the image-text matching loss, as illustrated in Eq. (2)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14726" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.228 + ], + "angle": 0, + "content": "class tokens sampled for the \\(i\\)-th training image, we have \\(m \\ll M\\). This means we can flexibly reduce the width of FGPFs in the text encoder for every training image. Surprisingly, our experimental results in the next section reveal that ignoring the majority of class tokens for each training image can enhance the generalization performance of the learned model. This improvement may stem from CSkip's capability to filter out redundant and distracting text features when performing image-text matching with the loss of \\(\\mathcal{L}_{\\mathrm{ITM}}\\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.244, + 0.223, + 0.26 + ], + "angle": 0, + "content": "3. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.27, + 0.281, + 0.286 + ], + "angle": 0, + "content": "3.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.293, + 0.483, + 0.444 + ], + "angle": 0, + "content": "Datasets. We conduct experiments using 11 datasets from diverse sources. In particular, for base-to-new generalization, cross-dataset transfer and few-shot learning, we use 11 datasets including ImageNet [4], Caltech101 [5], OxfordPets [23], StanfordCars [19], Flowers102 [22], Food101 [2], FGVCAircraft [21], SUN397 [31], UCF101 [28], DTD [3] and EuroSAT [12]. For domain generalization, we use ImageNet [4] as the source dataset and its four variants as target datasets including ImageNetV2 [26], ImageNet-Sketch [30], ImageNet-A [14], and ImageNet-R [13]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.445, + 0.483, + 0.535 + ], + "angle": 0, + "content": "Evaluation Metric. We report base-task accuracy (\\%, denoted as Base), new-task accuracy (\\%, denoted as New), and their harmonic mean (\\%, denoted as \\( H \\)) to compare the performance/effectiveness of different methods. We also report the time cost (seconds, denoted as Time) and memory usage (M, denoted as Memory) for efficiency evaluation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.537, + 0.483, + 0.764 + ], + "angle": 0, + "content": "Implementation details. Our implementation of Skip Tuning is based on the open-source Github repository of DePT \\([39]^2\\). In concrete terms, we leverage pre-trained ViT-B/16 as the backbone of the CLIP model. We employ the SGD optimizer to train the model with a learning rate of 2e-5 and a batch size of 4. By default, the number of training epochs is set to 20 for the base-to-new generalization, cross-dataset generalization, and domain generalization benchmarks, and 40 for the few-shot learning benchmark. The above hyperparameters are fixed across all datasets. We adjust the LSkip hyper-parameter \\(\\omega\\), CSkip hyper-parameters \\(r\\), and \\(\\lambda\\) in § 3.3. All experimental results are the average of 3 runs with different seeds. We conduct experiments using an NVIDIA V100 GPU. For more details and additional results, please refer to Sup. Mat." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.776, + 0.293, + 0.791 + ], + "angle": 0, + "content": "3.2. Experimental Results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.799, + 0.483, + 0.875 + ], + "angle": 0, + "content": "Base-to-New Generalization. The base-to-new generalization setting evaluates whether the models learned on a base task can generalize to new tasks with unseen classes. Following the comparison methods, for each dataset, we first construct a base task and a new task by equally dividing the" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.379 + ], + "angle": 0, + "content": "dataset into two sets of classes, then we perform prompt tuning on the base task and test the learned model on both the base and new tasks. Table 1 presents the obtained results of Skip Tuning and other state-of-the-art prompt tuning methods on 11 datasets. As shown, Skip Tuning achieves the best base-to-new generalization performance with the lowest memory usage and time cost. Concretely, Skip Tuning achieves \\(\\times 7.2\\) time efficiency, and \\(\\times 5.1\\) memory efficiency, while maintaining a \\(1.14\\%\\) improvement in H ACC compared to the previous state-of-the-art method PromptSRC. This demonstrates the effectiveness and efficiency of our proposed Skip Tuning method. Moreover, we observe a tradeoff between base-task and new-task accuracies for most competitors. For instance, CoPrompt outperforms DePT on new tasks but lags behind DePT on base tasks. Notably, our Skip Tuning achieves the best performance on both base and new tasks simultaneously. This suggests that Skip Tuning effectively mitigates the overfitting issue when transferring VLMs to the base (or target) task." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.383, + 0.905, + 0.67 + ], + "angle": 0, + "content": "Cross-Dataset Generalization. The cross-dataset generalization setting assesses whether models trained on a source dataset/distribution can generalize to unseen target datasets/distributions. We follow the common setup of those comparison methods to use ImageNet as the source dataset and the other 10 datasets as target datasets. Table 2 presents the obtained cross-dataset generalization results of our Skip Tuning and other state-of-the-art methods on the source and target datasets. As seen, compared to the previous state-of-the-art method PromptSRC, our Skip Tuning achieves \\(\\times 44\\) time efficiency, and \\(\\times 21.5\\) memory efficiency, while establishing \\(1.5\\%\\) and \\(1.19\\%\\) ACC improvements on the source and target distributions respectively. From the average results, Skip Tuning consistently shows superior effectiveness and efficiency over the nine competitors on the 10 target datasets, without compromising the results of the tuned model on the source dataset. This demonstrates the effectiveness of our Skip Tuning scheme for improving the robustness of the tuned model to distribution shifts." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Domain Generalization. The domain generalization setting assesses whether models trained on a source domain can generalize to unseen/target domains. In line with those comparison methods, we consider the ImageNet as the source domain and the other four ImageNet variants as target domains. As shown in Table 2, compared to the previous state-of-the-art method PromptSRC, Skip Tuning achieves \\(\\times 44\\) time efficiency, and \\(\\times 21.5\\) memory efficiency, while establishing \\(1.5\\%\\) and \\(0.55\\%\\) ACC improvements on the source and target domains, respectively. Besides, Skip Tuning consistently outperforms those competitors on the four target domains without compromising the performance of the tuned model on the source domain, which proves the effectiveness of our Skip Tuning scheme in improving the robustness of the tuned model to domain shifts." + }, + { + "type": "page_footnote", + "bbox": [ + 0.108, + 0.887, + 0.353, + 0.9 + ], + "angle": 0, + "content": "2https://github.com/Koorye/DePT" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "14727" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.217, + 0.084, + 0.782, + 0.096 + ], + "angle": 0, + "content": "Table 1. Base-to-new generalization results over 11 datasets. * indicates our reproduced results." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.097, + 0.905, + 0.504 + ], + "angle": 0, + "content": "
DatasetsMetricCoOp(IJCV'22)CoCoOp(CVPR'22)ProGrad(ICCV'23)KgCoOp(CVPR'23)MaPLe(CVPR'23)ProSRC(ICCV'23)TCP(CVPR'24)DePT(CVPR'24)CoPro.*ICLR'24)SkipT.(Ours)
ImgNetBase76.4775.9877.0275.8376.6677.6077.2777.0376.5377.73
New67.8870.4366.6669.9670.5470.7369.8770.1371.3070.40
H71.9273.1071.4672.7873.4774.0173.3873.4273.8273.89
CaltechBase98.0097.9698.0297.7297.7498.1098.2398.3098.6098.50
New89.3193.8193.8994.3994.3694.0394.6794.6095.1795.33
H93.7395.8495.9196.0396.0296.0296.4296.4196.8596.89
PetsBase93.6795.2095.0794.6595.4395.3394.6794.3394.7395.70
New95.2997.6997.6397.7697.7697.3097.2097.2396.7097.87
H94.4796.4396.3396.1896.5896.3095.9295.7695.7196.77
CarsBase78.1270.4977.6871.7672.9478.2780.8079.1373.1782.93
New60.4073.5968.6375.0474.0074.9774.1375.4770.6372.50
H68.1372.0172.8873.3673.4776.5877.3277.2671.8877.37
FlowersBase97.6094.8795.5495.0095.9298.0797.7398.0096.9398.57
New59.6771.7571.8774.7372.4676.5075.5776.3775.5075.80
H74.0681.7182.0383.6582.5685.9585.2385.8484.8885.70
Food101Base88.3390.7090.3790.5090.7190.6790.5790.5090.3790.67
New82.2691.2989.5991.7092.0591.5391.3791.6091.5392.03
H85.1990.9989.9891.0991.3891.1090.9791.0590.9591.34
AircraftBase40.4433.4140.5436.2137.4442.7341.9743.2036.1745.37
New22.3023.7127.5733.5535.6137.8734.4334.8334.4737.13
H28.7527.7432.8234.8336.5040.1537.8338.5735.3040.84
SUN397Base80.6079.7481.2680.2980.8282.6782.6382.3382.3082.40
New65.8976.8674.1776.5378.7078.4778.2077.8079.6379.03
H72.5178.5777.5578.3679.7580.5280.3580.0080.9480.68
DTDBase79.4477.0177.3577.5580.3683.3782.7782.2083.0083.77
New41.1856.0052.3554.9959.1862.9758.0759.1363.2067.23
H54.2464.8562.4564.3568.1671.1568.2568.7871.7674.59
EuroSATBase92.1987.4990.1185.6494.0792.9091.6389.0393.7792.47
New54.7460.0460.8964.3473.2373.9074.7371.0771.7383.00
H68.6971.2172.6773.4882.3582.3282.3279.0481.2887.48
UCF101Base84.6982.3384.3382.8983.0087.1087.1385.8086.2087.30
New56.0573.4574.9476.6778.6678.8080.7777.2378.7082.47
H67.4677.6479.3579.6580.7782.7483.8381.2982.2884.81
Avg ACCBase82.6980.4782.4880.7382.2884.2684.1383.6282.8985.04
New63.2271.6970.7573.6075.1476.1075.3675.0475.3277.53
H71.6675.8376.1677.0078.5579.9779.5179.1078.9381.11
CostTime (s)11862851231111914131735619733684239
Mem. (M)320415563204318817292041318917142060404
" + }, + { + "type": "table_caption", + "bbox": [ + 0.093, + 0.512, + 0.905, + 0.539 + ], + "angle": 0, + "content": "Table 2. Cross-dataset generalization results on 11 datasets. * indicates our reproduced results. The detailed results on the 10 target datasets (i.e., Caltech, Pets, Cars, ..., and UCF101) are reported in Sup. Mat. (B)." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.54, + 0.905, + 0.61 + ], + "angle": 0, + "content": "
DatasetsMetricCoOp(IJCV'22)CoCoOp(CVPR'22)ProGrad(ICCV'23)KgCoOp(CVPR'23)MaPLe(CVPR'23)ProSRC(ICCV'23)TCP(CVPR'24)DePT(CVPR'24)CoPro.*ICLR'24)SkipT.(Ours)
ImageNetACC71.5171.0272.2470.6670.7271.2771.4072.7772.5372.77
10 DatasetsAvg ACC63.8865.7462.7165.5166.3065.8166.2965.5564.8167.00
CostTime (s)316329391756223316364942500911617422796191611139
Mem. (M)154121362215412152541378614107152631378314131656
" + }, + { + "type": "table_caption", + "bbox": [ + 0.093, + 0.619, + 0.905, + 0.646 + ], + "angle": 0, + "content": "Table 3. Domain generalization results on ImageNet. * indicates our reproduced results. The detailed results on the 4 ImgNet variants (i.e., ImgNet-V2, ImgNet-S, ImgNet-A, and ImgNet-R) are presented in Sup. Mat. (B)." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.648, + 0.905, + 0.728 + ], + "angle": 0, + "content": "
DatasetsMetricCoOp(IJCV'22)CoCoOp(CVPR'22)ProGrad(ICCV'23)KgCoOp(CVPR'23)MaPLe(CVPR'23)ProSRC(ICCV'23)DePT(CVPR'24)CoPro.*ICLR'24)SkipT.(Ours)
ImageNetACC71.5171.0272.2470.6670.7271.2772.7772.5372.77
4 Imag.VariantsAvg ACC59.2859.9059.0760.1160.2660.6558.9760.0261.20
CostTime (s)3163293917562233163649425009122796191611139
Mem. (M)1541213622154121525413786141071378314131656
" + }, + { + "type": "title", + "bbox": [ + 0.095, + 0.755, + 0.254, + 0.769 + ], + "angle": 0, + "content": "3.3. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.095, + 0.779, + 0.481, + 0.809 + ], + "angle": 0, + "content": "In this section, we conduct ablative studies to further scrutinize our devised Skip Tuning method." + }, + { + "type": "text", + "bbox": [ + 0.095, + 0.811, + 0.482, + 0.901 + ], + "angle": 0, + "content": "Effectiveness of the Designed Components. Our proposed Skip Tuning approach simultaneously performs Layer-wise Skipping (LSkip) and Class-wise Skipping (CSkip) to enhance the memory and time efficiency of the full fine-tuning (FT) baseline. In this experiment, we investigate the effectiveness of LSkip and CSkip by gradually adding them to" + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.755, + 0.905, + 0.891 + ], + "angle": 0, + "content": "the FT baseline. From the results in Table 4, we have the following observations. i) Both LSkip and CSkip contribute to performance improvement. ii) By combining all those components, Skip Tuning improves the effectiveness and efficiency of the baseline method remarkably. iii) The memory and time efficiency gains are more pronounced when CSkip is applied to the FT baseline, compared to the performance improvements it yields. iv) CSkip improves the memory and time efficiency of the baseline without com" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "14728" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.132, + 0.087, + 0.379, + 0.188 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.38, + 0.087, + 0.621, + 0.188 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.625, + 0.087, + 0.867, + 0.188 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.117, + 0.191, + 0.878, + 0.207 + ], + "angle": 0, + "content": "Figure 4. Ablation studies of the number of skipped layers \\(\\omega\\) in LSkip, and the sampling rate \\(r\\), the decay coefficient \\(\\lambda\\) in CSkip." + }, + { + "type": "table_caption", + "bbox": [ + 0.11, + 0.223, + 0.462, + 0.237 + ], + "angle": 0, + "content": "Table 4. Ablation study on the components of Skip Tuning." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.238, + 0.482, + 0.308 + ], + "angle": 0, + "content": "
LSPCSPBaseNewHTime (s)Mem. (M)
XX84.9974.2679.2710021846
✓X85.0877.3081.00388861
X✓84.7774.4879.29440919
✓✓85.0477.5381.11239404
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.321, + 0.483, + 0.362 + ], + "angle": 0, + "content": "Table 5. Comparison of our Skip Tuning and adapter-based methods. \\( {}^{ \\dagger } \\) denotes the efficiency-optimized versions (details are illustrated in Sup. Mat. (C))." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.364, + 0.489, + 0.444 + ], + "angle": 0, + "content": "
MethodBaseNewHTime (s)Mem. (M)
FT84.9974.2679.2710021846
CLIP-adapter [10]74.4873.8174.148881784
LoRA [15]80.5374.7377.529101580
SkipT. (Ours)85.0477.5381.11239404
CLIP-adapter†76.7573.5675.12137175
SkipT. (Ours)†82.6675.5678.9511567
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.475, + 0.483, + 0.536 + ], + "angle": 0, + "content": "promising the generalization performance of the learned model. One possible reason is that CSkip can filter out redundant and distracting text features when performing image-text matching with the loss of \\(\\mathcal{L}_{\\mathrm{ITM}}\\)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.537, + 0.483, + 0.703 + ], + "angle": 0, + "content": "Impact of the LSkip Hyper-parameter \\(\\omega\\). Our Skip Tuning approach drops out the \\(1 \\sim \\omega\\) layers of the CLIP's vision and text encoders in the LSkip step. Here, we investigate the impact of \\(\\omega\\) on performance by setting \\(\\omega\\) to the values of \\(\\{2, 4, 6, 8, 10\\}\\). The obtained average testing results on the 11 datasets are reported in Figure 4 (Left). As shown, the obtained H ACCs gradually increase as the \\(\\omega\\) value decreases from 10 to 6, after which the performance gradually decreases. But, we also see that as the value of \\(\\omega\\) becomes smaller, both memory and time costs increase. Hence, we set \\(\\omega = 6\\) for LSKip in this work." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.704, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Impact of the CSkip Hyper-parameter \\( r \\). In the CSkip step of Skip Tuning, we devise an image-conditioned class sampling strategy to filter out irrelevant class tokens for each training image. The larger the sampling ratio \\( r \\) value, the more class tokens will be used to construct the training loss \\( \\mathcal{L}_{\\mathrm{ITM}} \\) for each image. It is necessary to scrutinize the impact of \\( r \\) on performance. To this end, we respectively set \\( r \\) to \\( \\{0.1, 0.3, 0.5, 0.7, 0.9, 1.0\\} \\), and report the average testing results on the 11 datasets in Figure 4 (Mid.). From the obtained H ACCs, our method is in general not sensitive to the change of \\( r \\) within a certain range (from 0.4 to 1.0). We also see a gradual decrease in H ACC when \\( r > 0.7 \\), suggesting that for each training image, not all class tokens are" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.232, + 0.905, + 0.292 + ], + "angle": 0, + "content": "beneficial for capturing task-specific knowledge. Besides, we can observe that as the value of \\( r \\) becomes larger, both memory and time costs increase. Therefore, we set \\( r = 0.5 \\) for CSKip in this work." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.293, + 0.907, + 0.475 + ], + "angle": 0, + "content": "Impact of the CSkip Hyper-parameter \\(\\lambda\\). Our Skip Tuning introduces an exponential decay coefficient \\(\\lambda\\) for the devised image-conditioned class sampling strategy in the CSkip step. We study the impact of \\(\\lambda\\) by setting \\(\\lambda\\) to the values of \\(\\{0.01, 0.1, 0.3, 0.5, 0.7, 0.9, 1.0\\}\\), and report the average testing results over the 11 datasets in Figure 4 (Right). As can be observed from the obtained H ACCs, our method is in general not sensitive to \\(\\lambda\\) when it takes the values from 0.1 to 0.7. Particularly, the H ACCs gradually decrease as \\(\\lambda\\) increases from 0.7 to 1.0. Also, we can see that as the value of \\(\\lambda\\) becomes larger, both memory and time costs increase. Thus, we set \\(\\lambda = 0.3\\) for CSkip in this work." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.485, + 0.693, + 0.5 + ], + "angle": 0, + "content": "3.4. Additional Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.508, + 0.907, + 0.734 + ], + "angle": 0, + "content": "Comparison with Adapter-based Methods. Previous experimental results demonstrate the superior effectiveness and efficiency of our Skip Tuning method over existing prompt tuning methods. To further demonstrate the advantages of Skip Tuning, we also compare it with the representative adapter-based methods LoRA [15] and CLIP-adapter [10]. The base-to-new generalization results averaged on 11 datasets are reported in Table 5, where \\(\\mathrm{CLIP - adapter}^{\\dagger}\\) refers to our re-implementation of CLIP-adapter, i.e., the CLIP's last-layer features are cached once before training and subsequently used to update the weights of the adapter module. As observed, our Skip Tuning offers substantial memory and time efficiency advantages over the two strong competitors, while also delivering better classification performance on both base and new tasks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Few-shot Learning. In the previous experiments, we follow the comparison approaches to evaluate the performance of different methods on \\(M\\)-way 16-shot tasks—16 training examples are sampled for each of the \\(M\\) inner-task classes. It is interesting to further scrutinize the effectiveness and efficiency of our Skip Tuning method under different shots. To this end, Figure 5 reports the few-shot learning performance of Skip Tuning and other representative competitors. As shown, our Skip Tuning demonstrates superior performance compared to other methods, with significant efficiency advantages in memory usage and time cost across" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14729" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.123, + 0.09, + 0.278, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.284, + 0.09, + 0.427, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.431, + 0.091, + 0.574, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.58, + 0.091, + 0.722, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.725, + 0.09, + 0.871, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.123, + 0.19, + 0.278, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.284, + 0.19, + 0.426, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.431, + 0.19, + 0.574, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.58, + 0.19, + 0.722, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.725, + 0.25, + 0.871, + 0.396 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.123, + 0.291, + 0.278, + 0.394 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.284, + 0.291, + 0.426, + 0.394 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.431, + 0.292, + 0.574, + 0.394 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.58, + 0.292, + 0.722, + 0.394 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.348, + 0.397, + 0.647, + 0.411 + ], + "angle": 0, + "content": "Figure 5. Few-shot learning results on 11 datasets." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.427, + 0.484, + 0.504 + ], + "angle": 0, + "content": "1-shot to 16-shot settings. Additionally, while PromptSRC and MaPLe enhance CoOp's performance, both methods incur increased time cost. In contrast, Skip Tuning achieves both effectiveness and efficiency without compromise, further underscoring the superiority of our approach." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.524, + 0.233, + 0.539 + ], + "angle": 0, + "content": "4. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.551, + 0.483, + 0.793 + ], + "angle": 0, + "content": "Pre-trained Vision-Language Models. Pre-trained vision-language models (VLMs) have garnered great attention recently. Notable models such as CLIP[24], ALIGN[16], LiT[36] and FILIP[34] have demonstrated exceptional performance across various vision-language tasks. These models are pre-trained on extensive datasets comprising image-text pairs sourced from the internet. For instance, CLIP[24] and ALIGN[16] respectively utilize over 400 million and 1 billion pairs, for training. The large scale of training data enables these models to excel in open-vocabulary image-text retrieval and zero-shot classification. Additionally, they have been successfully applied to downstream tasks including image classification[10, 40], object detection[6, 8, 11, 35, 43], and semantic segmentation[20, 25]. In this work, we focus on the effective adaptation of vision-language models to downstream visual recognition tasks." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Prompt Tuning. Prompt tuning (a.k.a. context optimization [42]) has emerged as a parameter-efficient learning paradigm to adapt powerful VLMs to downstream tasks [7, 9, 37, 38]. As a representative method, CoOp [42] enables task adaptation by optimizing a set of prompt vectors within CLIP's language branch. While effective, CoOp often suffers from limited generalization on new tasks due to" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.428, + 0.907, + 0.58 + ], + "angle": 0, + "content": "overfitting to the base task. A series of schemes are devised to tackle this problem, e.g. CoCoOp[41], KgCoOp[32], ProGrad[44], TCP [33], and DePT[39]. By adding trainable prompts into both the image and text branches of CLIP, MaPLe[17], PromptSRC[18], and CoPrompt[27] demonstrate remarkable performance on both base and new tasks. Despite the advantages, we reveal that freezing the parameters of VLMs during learning the context vectors neither facilitates the generalization of pre-trained knowledge nor significantly improves memory and time efficiency." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.599, + 0.634, + 0.615 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.627, + 0.907, + 0.779 + ], + "angle": 0, + "content": "In this work, we first reveal that freezing the parameters of VLMs during prompt tuning neither facilitates the transferability of pre-trained knowledge nor improves memory and time efficiency considerably. To circumvent this limitation, we propose Skip Tuning, an effective and efficient method for transferring VLMs to downstream tasks without relying on extra context vectors or adapter modules. Extensive experiments across a broad range of benchmarks demonstrate the superiority of our Skip Tuning method over both prompt tuning and adapter-based approaches." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Acknowledgements. This study is supported by grants from the National Natural Science Foundation of China (Grant No. 62425208, No. U22A2097, No. 62122018, No. 62020106008, No. U23A20315, No. 82441006, No. 62402094), the Postdoctoral Fellowship Program of CPSF (Grant No. GZB20240625), the Science and Technology Innovation Committee of Shenzhen Municipality Foundation (Grant No.JCYJ20240813114208012)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14730" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.184 + ], + "angle": 0, + "content": "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.185, + 0.484, + 0.255 + ], + "angle": 0, + "content": "[2] Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. Food-101-mining discriminative components with random forests. In Computer vision-ECCV 2014: 13th European conference, zurich, Switzerland, September 6-12, 2014, proceedings, part VI 13, pages 446-461. Springer, 2014. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.256, + 0.483, + 0.311 + ], + "angle": 0, + "content": "[3] Mircea Cimpoi, Subhransu Maji, Iasonas Kokkinos, Sammy Mohamed, and Andrea Vedaldi. Describing textures in the wild. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3606-3613, 2014. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.312, + 0.483, + 0.368 + ], + "angle": 0, + "content": "[4] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.369, + 0.482, + 0.436 + ], + "angle": 0, + "content": "[5] Li Fei-Fei, Rob Fergus, and Pietro Perona. Learning generative visual models from few training examples: An incremental bayesian approach tested on 101 object categories. Computer vision and Image understanding, 106(1):59-70, 2007. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.438, + 0.482, + 0.508 + ], + "angle": 0, + "content": "[6] Chengjian Feng, Yujie Zhong, Zequn Jie, Xiangxiang Chu, Haibing Ren, Xiaolin Wei, Weidi Xie, and Lin Ma. Promptdet: Towards open-vocabulary detection using uncurated images. In European Conference on Computer Vision, pages 701-717. Springer, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.509, + 0.482, + 0.564 + ], + "angle": 0, + "content": "[7] Yuqian Fu, Yanwei Fu, Jingjing Chen, and Yu-Gang Jiang. Generalized meta-fdmixup: Cross-domain few-shot learning guided by labeled target data. IEEE Transactions on Image Processing, 31:7078-7090, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.565, + 0.482, + 0.633 + ], + "angle": 0, + "content": "[8] Yuqian Fu, Yu Xie, Yanwei Fu, and Yu-Gang Jiang. Styleadv: Meta style adversarial training for cross-domain few-shot learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 24575-24584, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.635, + 0.482, + 0.704 + ], + "angle": 0, + "content": "[9] Yuqian Fu, Yu Wang, Yixuan Pan, Lian Huai, Xingyu Qiu, Zeyu Shangguan, Tong Liu, Yanwei Fu, Luc Van Gool, and Xingqun Jiang. Cross-domain few-shot object detection via enhanced open-set object detector. In ECCV, pages 247-264. Springer, 2025. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.705, + 0.482, + 0.773 + ], + "angle": 0, + "content": "[10] Peng Gao, Shijie Geng, Renrui Zhang, Teli Ma, Rongyao Fang, Yongfeng Zhang, Hongsheng Li, and Yu Qiao. Clip-adapter: Better vision-language models with feature adapters. International Journal of Computer Vision, 132(2): 581-595, 2024. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.775, + 0.482, + 0.829 + ], + "angle": 0, + "content": "[11] Xiuye Gu, Tsung-Yi Lin, Weicheng Kuo, and Yin Cui. Open-vocabulary object detection via vision and language knowledge distillation. arXiv preprint arXiv:2104.13921, 2021.8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.831, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[12] Patrick Helber, Benjamin Bischke, Andreas Dengel, and Damian Borth. Eurosat: A novel dataset and deep learning benchmark for land use and land cover classification. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 12(7):2217-2226, 2019. 5" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.176 + ], + "angle": 0, + "content": "[13] Dan Hendrycks, Steven Basart, Norman Mu, Saurav Kadavath, Frank Wang, Evan Dorundo, Rahul Desai, Tyler Zhu, Samyak Parajuli, Mike Guo, et al. The many faces of robustness: A critical analysis of out-of-distribution generalization. In Proceedings of the IEEE/CVF international conference on computer vision, pages 8340-8349, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.178, + 0.905, + 0.234 + ], + "angle": 0, + "content": "[14] Dan Hendrycks, Kevin Zhao, Steven Basart, Jacob Steinhardt, and Dawn Song. Natural adversarial examples. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15262-15271, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.235, + 0.905, + 0.29 + ], + "angle": 0, + "content": "[15] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.291, + 0.905, + 0.373 + ], + "angle": 0, + "content": "[16] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International conference on machine learning, pages 4904-4916. PMLR, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.376, + 0.905, + 0.445 + ], + "angle": 0, + "content": "[17] Muhammad Uzair Khattak, Hanoona Rasheed, Muhammad Maaz, Salman Khan, and Fahad Shahbaz Khan. Maple: Multi-modal prompt learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19113-19122, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.447, + 0.905, + 0.531 + ], + "angle": 0, + "content": "[18] Muhammad Uzair Khattak, Syed Talal Wasim, Muzammal Naseer, Salman Khan, Ming-Hsuan Yang, and Fahad Shahbaz Khan. Self-regulating prompts: Foundational model adaptation without forgetting. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15190-15200, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.532, + 0.905, + 0.588 + ], + "angle": 0, + "content": "[19] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In Proceedings of the IEEE international conference on computer vision workshops, pages 554–561, 2013. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.59, + 0.907, + 0.645 + ], + "angle": 0, + "content": "[20] Timo Lüddecke and Alexander Ecker. Image segmentation using text and image prompts. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7086-7096, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.647, + 0.905, + 0.699 + ], + "angle": 0, + "content": "[21] Subhransu Maji, Esa Rahtu, Juho Kannala, Matthew Blaschko, and Andrea Vedaldi. Fine-grained visual classification of aircraft. arXiv preprint arXiv:1306.5151, 2013. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.703, + 0.905, + 0.759 + ], + "angle": 0, + "content": "[22] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In 2008 Sixth Indian conference on computer vision, graphics & image processing, pages 722-729. IEEE, 2008. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.76, + 0.905, + 0.815 + ], + "angle": 0, + "content": "[23] Omkar M Parkhi, Andreae Vedaldi, Andrew Zisserman, and CV Jawahar. Cats and dogs. In 2012 IEEE conference on computer vision and pattern recognition, pages 3498-3505. IEEE, 2012. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.818, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[24] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 2, 8" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "14731" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.175 + ], + "angle": 0, + "content": "[25] Yongming Rao, Wenliang Zhao, Guangyi Chen, Yansong Tang, Zheng Zhu, Guan Huang, Jie Zhou, and Jiwen Lu. Denseclip: Language-guided dense prediction with context-aware prompting. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18082-18091, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.177, + 0.482, + 0.232 + ], + "angle": 0, + "content": "[26] Benjamin Recht, Rebecca Roelofs, Ludwig Schmidt, and Vaishaal Shankar. Do imagenet classifiers generalize to imagenet? In International conference on machine learning, pages 5389-5400. PMLR, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.233, + 0.482, + 0.273 + ], + "angle": 0, + "content": "[27] Shuvendu Roy and Ali Etemad. Consistency-guided prompt learning for vision-language models. arXiv preprint arXiv:2306.01195, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.274, + 0.482, + 0.316 + ], + "angle": 0, + "content": "[28] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.317, + 0.482, + 0.385 + ], + "angle": 0, + "content": "[29] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.386, + 0.482, + 0.44 + ], + "angle": 0, + "content": "[30] Haohan Wang, Songwei Ge, Zachary Lipton, and Eric P Xing. Learning robust global representations by penalizing local predictive power. Advances in Neural Information Processing Systems, 32, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.442, + 0.482, + 0.511 + ], + "angle": 0, + "content": "[31] Jianxiong Xiao, James Hays, Krista A Ehinger, Aude Oliva, and Antonio Torralba. Sun database: Large-scale scene recognition from abbey to zoo. In 2010 IEEE computer society conference on computer vision and pattern recognition, pages 3485-3492. IEEE, 2010. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.512, + 0.482, + 0.579 + ], + "angle": 0, + "content": "[32] Hantao Yao, Rui Zhang, and Changsheng Xu. Visual-language prompt tuning with knowledge-guided context optimization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6757-6767, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.581, + 0.482, + 0.648 + ], + "angle": 0, + "content": "[33] Hantao Yao, Rui Zhang, and Changsheng Xu. Tcp: Textual-based class-aware prompt tuning for visual-language model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23438-23448, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.65, + 0.482, + 0.706 + ], + "angle": 0, + "content": "[34] Lewei Yao, Runhui Huang, Lu Hou, Guansong Lu, Minzhe Niu, Hang Xu, Xiaodan Liang, Zhenguo Li, Xin Jiang, and Chunjing Xu. Filip: Fine-grained interactive language-image pre-training. arXiv preprint arXiv:2111.07783, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.707, + 0.482, + 0.762 + ], + "angle": 0, + "content": "[35] Yuhang Zang, Wei Li, Kaiyang Zhou, Chen Huang, and Chen Change Loy. Open-vocabulary detr with conditional matching. In European Conference on Computer Vision, pages 106-122. Springer, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.763, + 0.482, + 0.831 + ], + "angle": 0, + "content": "[36] Xiaohua Zhai, Xiao Wang, Basil Mustafa, Andreas Steiner, Daniel Keysers, Alexander Kolesnikov, and Lucas Beyer. Lit: Zero-shot transfer with locked-image text tuning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18123-18133, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[37] Ji Zhang, Jingkuan Song, Lianli Gao, and Hengtao Shen. Free-lunch for cross-domain few-shot learning: Style-aware episodic training with robust contrastive learning. In Proceedings of the 30th ACM International Conference on Multimedia, pages 2586-2594, 2022. 8" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.135 + ], + "angle": 0, + "content": "[38] Ji Zhang, Lianli Gao, Xu Luo, Hengtao Shen, and Jingkuan Song. Deta: Denoised task adaptation for few-shot learning. In ICCV, pages 11541-11551, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[39] Ji Zhang, Shihan Wu, Lianli Gao, Heng Tao Shen, and Jingkuan Song. Dept: Decoupled prompt tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12924-12933, 2024. 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.192, + 0.905, + 0.26 + ], + "angle": 0, + "content": "[40] Renrui Zhang, Rongyao Fang, Wei Zhang, Peng Gao, Kunchang Li, Jifeng Dai, Yu Qiao, and Hongsheng Li. Tip-adapter: Training-free clip-adapter for better vision-language modeling. arXiv preprint arXiv:2111.03930, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.262, + 0.905, + 0.33 + ], + "angle": 0, + "content": "[41] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Conditional prompt learning for vision-language models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16816-16825, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.332, + 0.905, + 0.387 + ], + "angle": 0, + "content": "[42] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision, 130(9):2337-2348, 2022. 2, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.39, + 0.905, + 0.457 + ], + "angle": 0, + "content": "[43] Xingyi Zhou, Rohit Girdhar, Armand Joulin, Philipp Krahenbihl, and Ishan Misra. Detecting twenty-thousand classes using image-level supervision. In European Conference on Computer Vision, pages 350-368. Springer, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.46, + 0.905, + 0.516 + ], + "angle": 0, + "content": "[44] Beier Zhu, Yulei Niu, Yucheng Han, Yue Wu, and Hanwang Zhang. Prompt-aligned gradient for prompt tuning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15659-15669, 2023. 8" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.516 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14732" + } + ] +] \ No newline at end of file diff --git a/2025/Skip Tuning_ Pre-trained Vision-Language Models are Effective and Efficient Adapters Themselves/e5453c9b-356b-43d9-9454-788a3f8b346d_origin.pdf b/2025/Skip Tuning_ Pre-trained Vision-Language Models are Effective and Efficient Adapters Themselves/e5453c9b-356b-43d9-9454-788a3f8b346d_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..ae9678cede293653fe9b138faa2534e0c82b62be --- /dev/null +++ b/2025/Skip Tuning_ Pre-trained Vision-Language Models are Effective and Efficient Adapters Themselves/e5453c9b-356b-43d9-9454-788a3f8b346d_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9c45f969d48d082714a1240ba39aff3b03e4bb242b8635d11bc1e1f57f23d1cd +size 4315368 diff --git a/2025/Skip Tuning_ Pre-trained Vision-Language Models are Effective and Efficient Adapters Themselves/full.md b/2025/Skip Tuning_ Pre-trained Vision-Language Models are Effective and Efficient Adapters Themselves/full.md new file mode 100644 index 0000000000000000000000000000000000000000..cf77248cd49d49422be4871cc299e36299d896f5 --- /dev/null +++ b/2025/Skip Tuning_ Pre-trained Vision-Language Models are Effective and Efficient Adapters Themselves/full.md @@ -0,0 +1,334 @@ +# Skip Tuning: Pre-trained Vision-Language Models are Effective and Efficient Adapters Themselves + +Shihan Wu $^{1}$ Ji Zhang $^{2*}$ Pengpeng Zeng $^{3}$ Lianli Gao $^{3}$ Jingkuan Song $^{4}$ Heng Tao Shen $^{4}$ + +1 University of Electronic Science and Technology of China (UESTC) + +$^{2}$ Southwest Jiaotong University + +3 Shenzhen Institute for Advanced Study, UESTC + +$^{4}$ Tongji University + +{shihan.wu.koorye@outlook.com, jizhang.jim@gmail.com} + +# Abstract + +Prompt tuning (PT) has long been recognized as an effective and efficient paradigm for transferring large pre-trained vision-language models (VLMs) to downstream tasks by learning a tiny set of context vectors. Nevertheless, in this work, we reveal that freezing the parameters of VLMs during learning the context vectors neither facilitates the transferability of pre-trained knowledge nor improves the memory and time efficiency significantly. Upon further investigation, we find that reducing both the length and width of the feature-gradient propagation flows of the full fine-tuning (FT) baseline is key to achieving effective and efficient knowledge transfer. Motivated by this, we propose Skip Tuning, a novel paradigm for adapting VLMs to downstream tasks. Unlike existing PT or adapter-based methods, Skip Tuning applies Layer-wise Skipping (LSkip) and Classwise Skipping (CSkip) upon the FT baseline without introducing extra context vectors or adapter modules. Extensive experiments across a wide spectrum of benchmarks demonstrate the superior effectiveness and efficiency of our Skip Tuning over both PT and adapter-based methods. Code: https://github.com/Koorye/SkipTuning. + +# 1. Introduction + +There have recently been significant advancements in large pre-trained vision-language models (VLMs) [1, 24, 29]. One notable achievement is the CLIP model [24], which leverages an image-text matching (ITM) loss to align images with their corresponding textual descriptions in a common feature space. While VLMs have proven impressive capabilities in recognizing open-set visual concepts, their zero-shot generalization performance declines significantly + +![](images/52012149a2b74e739ce937db35d73275a6d735cd932942e51933cea05c8c5807.jpg) + +![](images/04f7f904d4ff9c125ba4a0e81dac82a9846d98c05f9ee7c626fa072ca02e18bb.jpg) + +![](images/637edd4d4f795e3863d9bcf476dce04b395743401f859277d5691b1a966f2048.jpg) + +![](images/bec99f81e1ebb308bfa5763dfad5895e68ed2218fc8f2381cfde13d9869fb887.jpg) + +![](images/9b9f9f7401feb7f2a82f77008366704c2164290c98137ac1b5355eccc217218a.jpg) + +![](images/dd0d41d243ed1a609d0d1887ed982960c1468b9a7939a0f755c15c3de6a0b566.jpg) + +![](images/892b3907047f74164c64be38f5db12e9685982e621fd673df64374abd0e74aa6.jpg) + +![](images/99126689c80abbc6f16d9272f31f9e7a8f9f67512237fb242416a164aadbfcaf.jpg) + +![](images/f48f2194776fafc7819003b22098ad789e6d07072cf53a25c4b26fc551073fe5.jpg) + +![](images/a9b5c20b2d3e3466709c90502a5577ae53031f74f8227a671ee70c4c83de655b.jpg) +Figure 1. Comparison of our devised Skip Tuning with state-of-the-art prompt tuning methods in terms of training time (seconds), memory cost (M), and classification accuracy (\%) across base-to-new generalization, cross-dataset generalization, domain generalization, and few-shot learning benchmarks. $\times$ indicates the performance improvement over the state-of-the-art. Comparison results with the adapter-based methods are reported in Table 5. + +![](images/a776cef8fef9e06066ad2d541d3809fea0fd63c01c6fdb6c73fd2548c3db8333.jpg) + +![](images/d7b1f847708d48087fb5011bd853344d6939b1a79c8e66c3f2500faa2fc44f39.jpg) + +when encountering category, distribution, or domain shifts between upstream training data and downstream tasks. + +Prompt tuning (PT) has long been recognized as an effective and efficient paradigm for transferring large pre-trained vision-language models (VLMs) to downstream tasks. The core concept of PT is to learn a task-specific prompt (i.e., a small number of context vectors) for the target task, using a limited amount of training data, while keeping the pre-trained VLM parameters fixed. Although many PT ap + +![](images/5fe69620d9a1ae354053712bd3ac8d873539f6be59708020154b7fa6e50585f8.jpg) + +![](images/355e9d569600d205411f954e5a5a8665de6909ac844109780573ec7120975b0c.jpg) + +![](images/7196e6fe8ee7b931c23df993265fa5e4ca97222874310cc5a79386d8d5c2581b.jpg) + +![](images/b43f70d52bfa825b77f644a21af4c50c31c99c9f888a2baa5da7acd5148dca2b.jpg) + +![](images/cddd56924c33fbd6b808bbb6094a42e45ca2de58bf82b99467f96ef0b8c6b5c4.jpg) +Figure 2. Motivations. (a) Comparison between the prompt tuning (PT) method CoOp [42] and the full fine-tuning (FT) baseline in terms of i) the number of learnable parameters, ii) memory usage, iii) time cost, and iv) base-to-new generalization performance. (b) Feature Sensitivity (FS) of CLIP's network layers, averaged over 100 randomly-sampled training images. (c) Gradient Dependence (GD) of class tokens for different training images. + +![](images/3e3a8eedd5ace8dd0c18c7c7d29bd159fa7c17a6e8f2af8e9ec8695803c7e5bb.jpg) + +proaches have reported improved performance and parameter efficiency over the full fine-tuning (FT) baseline, the discrepancy of implementation details among those PT approaches obscure the actual performance enhancement. For example, the FT performance can be significantly underestimated by training with coarsely tuned hyper-parameters. Quantitative evidence is presented in Figure 2 (a), where we make comparisons between the PT method CoOp [42] and the FT baseline in terms of the number of learnable parameters, memory usage, time cost as well as base-to-new generalization performance (see §2.2). As observed, although PT significantly improves the parameter efficiency of FT (using 1/51200 parameters of FT), the improvements in memory and time efficiency are relatively insignificant (reducing only $6.3\%$ memory usage and $15.8\%$ time cost). Besides, compared to FT, the classification accuracy of PT decreases by $3.49\%$ and $4.49\%$ on base and new tasks, respectively. This suggests that pursuing higher parameter efficiency by freezing the overwhelming majority weights of VLMs during learning the context vectors neither facilitates the transferability of pre-trained knowledge nor improves memory and time efficiency considerably. Besides, in many real-world applications, memory and time efficiency often take precedence over parameter efficiency in terms of practical importance. We therefore raise the following question: + +Can we optimize the memory and time efficiency of the FT baseline and adapt VLMs to downstream tasks in an effective and efficient manner? + +To answer the above question, we scrutinize the FeatureGradient Propagation Flows (FGPFs) in the vision encoder and text encoder of the CLIP model when performing FT on the base (or target) task. Interestingly, we observe that, for each training image, the majority of shallow network layers and class tokens contribute minimally to capturing task-specific knowledge for the base task (see §2.2). Motivated by this, we propose Skip Tuning, a novel paradigm for adapting VLMs to downstream tasks without introducing extra context vectors or adapter modules. Concretely, Skip Tuning incorporates two strategies, Layer-wise Skipping (LSkip) and Class-wise Skipping (CSkip), to simultaneously reduce the length and width of FGFs in the FT baseline, thereby establishing effective and efficient knowledge transfer of VLMs, as shown in Figure 3. + +Effectiveness and Efficiency. We conduct extensive experiments across a wide range of benchmarks to validate the effectiveness and efficiency of our Skip Tuning scheme. An overview of the achieved results is shown in Figure 1. As seen, our Skip Tuning demonstrates superiority over existing PT methods, e.g., on the few-shot learning benchmark, Skip Tuning achieves $\times 15$ time efficiency, $\times 6.4$ memory efficiency, while yielding a $1.04\%$ improvement in ACC compared to the state-of-the-art. Furthermore, we also show the advantages of Skip Tuning over existing adapter-based methods in Table 5, where Skip Tuning achieves $\times 3.8$ time efficiency, $\times 3.9$ memory efficiency, along with a $3.59\%$ H ACC enhancement over the strong rival LoRA [15]. + +Main Contributions. The main contributions are threefold. + +- We reveal that reducing both the width and length of the feature-gradient propagation flows (FGPFs) of the full fine-tuning (FT) baseline is key to establishing effective and efficient knowledge transfer. +- We devise Skip Tuning, an effective and efficient method for transferring VLMs to downstream tasks without relying on extra context vectors or adapter modules. +- We evaluate our method on a wide spectrum of benchmarks, demonstrating the superiority of Skip Tuning over both prompt tuning and adapter-based approaches. + +# 2. Methodology + +In this section, we elaborate on our Skip Tuning approach. We start with an introduction of preliminaries. + +# 2.1. Preliminaries + +Contrastive Language-Image Pre-training (CLIP). Following the common practice of existing paradigms for adapting PVLMs, we adopt CLIP [24] as the testbed in this work. CLIP aims to learn an alignment between image and text features generated by an image encoder and a text encoder, respectively. By being exposed to 400 million imagetext association pairs and employing a contrastive learning paradigm within a shared feature space, CLIP acquires a di + +![](images/70dad74a6f292a14550ffedc09571bf12c8f1b5ed293682e9f3ba2ca84e6c3c4.jpg) +Figure 3. Overview of our proposed Skip Tuning. Skip Tuning performs Layer-wise Skipping (LSkip) and Class-wise Skipping (CSkip)) to enhance the memory and time efficiency of the FT baseline. Specifically, LSkip reduces the length of feature-gradient propagation flows (FGPFs) by caching intermediate features produced by the $\omega$ -th layers of CLIP's vision encoder $E_{V}$ and text encoder $E_{T}$ before FT begins. In contrast, CSkip reduces the width of FGFs by filtering out unimportant class tokens in the text encoder $E_{T}$ for every training image. + +verse array of open-set visual concepts that can be readily applied to downstream tasks. For instance, zero-shot classification can be achieved by framing the classification task as an image-text matching problem. Initially, a prompt (e.g. "a photo of a") is crafted to extract text features of all intra-task classes by presenting the class-extended prompt ("a photo of a [CLS]") to the text encoder. Subsequently, the image encoder is utilized to derive the image feature of a given input example, and class prediction is performed by comparing cosine distances between the image feature and text features of classes. + +Prompt Tuning with an Image-Text Matching Loss. PT aims to learn a task-specific prompt with few-labeled samples from the base task. Concretely, a prompt is formulated as $\kappa$ context vectors: $\{v_{1}, v_{2}, \dots, v_{\kappa}\}$ . During training, we produce the text feature of the $j$ -th class by inputting the class-token-extended prompt $c_{j} = \{v_{1}, v_{2}, \dots, v_{\kappa}, [\mathrm{CLS}]\}$ to the text encoder $g(\cdot)$ , where [CLS] indicates the class token of the $j$ -th class. Denote $\pmb{f}_{i} \in \mathbb{R}^{d}$ as the image feature of a sample $\pmb{x}_{i}$ extracted by the image encoder, the task-specific prompt can be updated via back-propagating the image-text matching (ITM) loss through the frozen CLIP model. The ITM loss can be expressed as + +$$ +\mathcal {L} _ {\mathrm {I T M}} = - \sum_ {j} \boldsymbol {y} _ {j} \log \mathrm {p} \left(\boldsymbol {c} _ {j} \mid \boldsymbol {x} _ {i}\right), \tag {1} +$$ + +where $\pmb{y}$ is the one-hot label, and + +$$ +\mathrm {p} \left(\boldsymbol {c} _ {j} \mid \boldsymbol {x} _ {i}\right) = \frac {e ^ {< g \left(\boldsymbol {c} _ {j}\right) , \boldsymbol {f} _ {i} > / \tau}}{\sum_ {t = 1} ^ {M} e ^ {< g \left(\boldsymbol {c} _ {t}\right) , \boldsymbol {f} _ {i} > / \tau}}, \tag {2} +$$ + +$< \cdot >$ denotes cosine similarity, $M$ is the number of classes, and $\tau$ is the temperature learned by CLIP. + +# 2.2. Rethinking the FT Baseline + +While many PT schemes have reported improved performance and efficiency over the full fine-tuning (FT) baseline, + +the discrepancy of the implementation details among those schemes obscures the actual performance enhancement. + +Comparison of PT and FT. To comprehensively evaluate the progress established by PT, we perform a comparison between the representative PT method CoOp [42] with the full FT baseline in terms of i) the number of learnable parameters, ii) memory usage, iii) time cost, and iv) baseto-new generalization results. The experimental setting is shown in Sup. Mat. (A). From the obtained results in Figure 2 (a), we observe that although PT significantly reduces the number of learnable parameters compared with FT (using 1/51200 parameters of FT), the improvements in memory efficiency and time efficiency are relatively insignificant (reducing only $6.3\%$ memory usage and $15.8\%$ time cost). Notably, in comparison to FT, the classification accuracy of PT decreases by $3.49\%$ and $4.49\%$ on base and new tasks, respectively. This means that pursuing higher parameter efficiency by freezing the vast majority of VLM weights during PT neither enhances the transferability of pre-trained knowledge nor improves memory and time efficiency significantly. In other words, existing PT methods involve a trade-off between parameter efficiency and performance. + +Cost Analysis of FT. The computational cost (i.e., memory and time cost) for fine-tuning VLMs is mainly caused by Feature-Gradient Propagation Flows (FGPFs)—the forward propagation flows of image/text features and the backward propagation flows of gradients. For example, on the CLIP model, each propagation flow is required to traverse all the $N$ layers of the vision encoder $E_{V}$ and the text encoder $E_{T}$ . For simplicity, we assume that $E_{V}$ and $E_{T}$ have identical network structures (e.g. ViT-B/16). Denote the number of class tokens as $M$ , and the costs of the FGPFs in the $l$ -th layer of $E_{V}$ and $E_{T}$ as $C_{V}$ and $C_{T}$ respectively. Thus, the total cost $C_{\mathrm{total}} = N \times (C_{V} + C_{T} \times M^{1})$ . In this work, + +we refer to the number of network layers and the number of class tokens used for each training image as the length and width of FGPFs, respectively. In particular, $\text{length} = N$ , $\text{width} = M$ for existing PT methods, as shown in Figure 3. We therefore raise the following question: + +Can we reduce both the length and width of FGPFs without compromising the FT performance? + +Towards Effective and Efficient Knowledge Transfer. To answer the above question, we design two metrics of Feature Sensitivity (FS) and Gradient Dependence (GD) to estimate the contributions of different network layers and class tokens for each training image $\mathbf{x}$ , respectively. + +More specifically, let $\pmb{f}_l$ and $\pmb{f}_l'$ be a vision (or a text) feature before and after fine-tuning the $l$ -th layer of the CLIP's vision (or text) encoder, the FS of the $l$ -th layer for the training image $\pmb{x}$ can be expressed as + +$$ +\mathbf {F} \mathbf {S} _ {l} (\boldsymbol {x}) = \psi \left(\boldsymbol {f} _ {l}, \boldsymbol {f} _ {l} ^ {\prime}\right), \tag {3} +$$ + +where $\psi (\cdot)$ is the Euclidean distance. Hence, the bigger the $\mathbf{FS}_l(\pmb {x})$ value, the more important the $l$ -th layer for the input image $\pmb{x}$ . Let $\nabla f_{l}$ indicate the computed feature gradients passing through the last layer of the vision encoder for the training image $\pmb{x}$ and $\nabla f_l^{(c)}$ be the feature gradients after the $c$ -th class token is removed from Eq. 2. The GD of the $c$ -th class for $\pmb{x}$ can be computed as + +$$ +\mathbf {G D} _ {c} (\boldsymbol {x}) = \psi (\nabla \boldsymbol {f} _ {l}, \nabla \boldsymbol {f} _ {l} ^ {(c)}). \tag {4} +$$ + +Therefore, the bigger the $\mathbf{GD}_c(x)$ value, the more important the $c$ -th class for the input image $x$ . + +Figure 2 (b) illustrates the obtained FS values of different image/text encoder layers when adapting CLIP to the two datasets Caltech101 and Food101. As can be observed, the FS scores of the vast majority of shallow layers are close to 0, and only the last few layers contribute significantly to adapting CLIP to the two datasets. Moreover, Figure 2 (c) presents the frequency distributions of the GD values of class tokens for 100 training images randomly sampled from the two datasets. As shown, most class tokens contribute minimally to capturing task-specific knowledge from each training image. In a nutshell, the above observations reveal that we can achieve effective and efficient FT by filtering out unimportant network layers and class tokens in the text encoder for every training image. + +# 2.3. Skip Tuning + +Motivated by the previous section, we propose Skip Tuning, which performs Layer-wise Skipping (LSkip) and Classwise Skipping (CSkip) to adapt VLMs to downstream tasks in an effective and efficient manner, as shown in Figure 3. + +Layer-wise Skipping (LSkip). LSkip aims to reduce the length of FGPFs without compromising the performance of the FT baseline. To this end, LSkip first saves the image and text features before the $\omega$ -th shallow layers of $E_{V}$ and $E_{T}$ in a cache, and then uses the saved intermediate features as input to update the parameters of the remaining $N - \omega$ deep layers. Denote $\pmb{v}_{i}^{\omega}$ and $\pmb{t}_{j}^{\omega}$ as the extracted image and text features of the image $\pmb{x}_i$ and the class token CLS ( $j = 1, \dots, M$ ) in the $\omega$ -th layers of $E_{V}$ and $E_{T}$ , respectively, i.e., + +$$ +\boldsymbol {v} _ {i} ^ {\omega} = E _ {V} [ 1: \omega ] (\boldsymbol {x} _ {i}), \boldsymbol {t} _ {j} ^ {\omega} = E _ {T} [ 1: \omega ] (\boldsymbol {c} _ {j}), \tag {5} +$$ + +where $c_{j}$ is constructed using a hand-crafted prompt, e.g., $c_{j} = a$ photo of a [CLS]. + +After that, the $[1:\omega]$ layers of both $E_{V}$ and $E_{T}$ are discarded. During FT, we input $\pmb{v}_i^\omega$ and $\pmb{t}_j^\omega$ to the remaining $N - \omega$ deep layers of $E_{V}$ and $E_{T}$ , and obtain: + +$$ +\tilde {\boldsymbol {v}} _ {i} ^ {N} = E _ {V} [ \omega + 1: N ] (\boldsymbol {v} _ {i} ^ {\omega}), \tilde {\boldsymbol {t}} _ {j} ^ {N} = E _ {T} [ \omega + 1: N ] (\boldsymbol {t} _ {j} ^ {\omega}), \tag {6} +$$ + +which are used to calculate the loss $\mathcal{L}_{\mathrm{ITM}}$ in Eq. 1. In this way, the image and text features, along with the calculated gradients, propagate through only $N - \omega$ layers, effectively reducing the length of FGPFs. + +Class-wise Skipping (CSkip). Given a training image, the goal of CSkip is to filter out unimportant class tokens in the text encoder $E_{T}$ during the construction of the image-text matching loss $\mathcal{L}_{\mathrm{ITM}}$ (Eq. 1). Intuitively, the direct way is to select the top $k$ closest class tokens to compute the loss for each image, based on the similarities between the image feature and the text features of the $M$ inner-task classes. Nevertheless, we empirically find that this strategy leads to overfitting by selecting the same subset of classes for each image across different training epochs. Therefore, we propose an exponential image-conditioned class filtering strategy to overcome this limitation. + +Concretely, before FT begins, we use the text encoder $E_{T}$ to produce $M$ class features with a hand-crafted prompt, e.g. "a photo of a [CLS]". Then, we sort the cosine similarities between the vision feature of a training image and the $M$ class features in descending order. The probability of sampling the $j$ -th class token for the current training image can be expressed as: + +$$ +p _ {j} = \left\{ \begin{array}{l l} 1 & o _ {j} \leq r \times M, \\ \zeta \left(o _ {j} - r \times M, \lambda\right) & o _ {j} > r \times M, \end{array} \right. \tag {7} +$$ + +where $\zeta (\iota ,\lambda) = e^{-\lambda \iota}$ , $\lambda >0$ is the exponential decay coefficient, $r$ is the sampling ratio, and $o_{j}$ is the sorting index of the $j$ -th class. In this way, for each image, we can maintain the top $r\times M$ closest class tokens while also sampling the remaining $(1 - r)\times M$ classes with a certain probability for constructing $\mathcal{L}_{\mathrm{ITM}}$ , improving the diversity of training data at different epochs. Denote $m$ as the number of + +class tokens sampled for the $i$ -th training image, we have $m \ll M$ . This means we can flexibly reduce the width of FGPFs in the text encoder for every training image. Surprisingly, our experimental results in the next section reveal that ignoring the majority of class tokens for each training image can enhance the generalization performance of the learned model. This improvement may stem from CSkip's capability to filter out redundant and distracting text features when performing image-text matching with the loss of $\mathcal{L}_{\mathrm{ITM}}$ . + +# 3. Experiments + +# 3.1. Experimental Setup + +Datasets. We conduct experiments using 11 datasets from diverse sources. In particular, for base-to-new generalization, cross-dataset transfer and few-shot learning, we use 11 datasets including ImageNet [4], Caltech101 [5], OxfordPets [23], StanfordCars [19], Flowers102 [22], Food101 [2], FGVCAircraft [21], SUN397 [31], UCF101 [28], DTD [3] and EuroSAT [12]. For domain generalization, we use ImageNet [4] as the source dataset and its four variants as target datasets including ImageNetV2 [26], ImageNet-Sketch [30], ImageNet-A [14], and ImageNet-R [13]. + +Evaluation Metric. We report base-task accuracy (\%, denoted as Base), new-task accuracy (\%, denoted as New), and their harmonic mean (\%, denoted as $H$ ) to compare the performance/effectiveness of different methods. We also report the time cost (seconds, denoted as Time) and memory usage (M, denoted as Memory) for efficiency evaluation. + +Implementation details. Our implementation of Skip Tuning is based on the open-source Github repository of DePT $[39]^2$ . In concrete terms, we leverage pre-trained ViT-B/16 as the backbone of the CLIP model. We employ the SGD optimizer to train the model with a learning rate of 2e-5 and a batch size of 4. By default, the number of training epochs is set to 20 for the base-to-new generalization, cross-dataset generalization, and domain generalization benchmarks, and 40 for the few-shot learning benchmark. The above hyperparameters are fixed across all datasets. We adjust the LSkip hyper-parameter $\omega$ , CSkip hyper-parameters $r$ , and $\lambda$ in § 3.3. All experimental results are the average of 3 runs with different seeds. We conduct experiments using an NVIDIA V100 GPU. For more details and additional results, please refer to Sup. Mat. + +# 3.2. Experimental Results + +Base-to-New Generalization. The base-to-new generalization setting evaluates whether the models learned on a base task can generalize to new tasks with unseen classes. Following the comparison methods, for each dataset, we first construct a base task and a new task by equally dividing the + +dataset into two sets of classes, then we perform prompt tuning on the base task and test the learned model on both the base and new tasks. Table 1 presents the obtained results of Skip Tuning and other state-of-the-art prompt tuning methods on 11 datasets. As shown, Skip Tuning achieves the best base-to-new generalization performance with the lowest memory usage and time cost. Concretely, Skip Tuning achieves $\times 7.2$ time efficiency, and $\times 5.1$ memory efficiency, while maintaining a $1.14\%$ improvement in H ACC compared to the previous state-of-the-art method PromptSRC. This demonstrates the effectiveness and efficiency of our proposed Skip Tuning method. Moreover, we observe a tradeoff between base-task and new-task accuracies for most competitors. For instance, CoPrompt outperforms DePT on new tasks but lags behind DePT on base tasks. Notably, our Skip Tuning achieves the best performance on both base and new tasks simultaneously. This suggests that Skip Tuning effectively mitigates the overfitting issue when transferring VLMs to the base (or target) task. + +Cross-Dataset Generalization. The cross-dataset generalization setting assesses whether models trained on a source dataset/distribution can generalize to unseen target datasets/distributions. We follow the common setup of those comparison methods to use ImageNet as the source dataset and the other 10 datasets as target datasets. Table 2 presents the obtained cross-dataset generalization results of our Skip Tuning and other state-of-the-art methods on the source and target datasets. As seen, compared to the previous state-of-the-art method PromptSRC, our Skip Tuning achieves $\times 44$ time efficiency, and $\times 21.5$ memory efficiency, while establishing $1.5\%$ and $1.19\%$ ACC improvements on the source and target distributions respectively. From the average results, Skip Tuning consistently shows superior effectiveness and efficiency over the nine competitors on the 10 target datasets, without compromising the results of the tuned model on the source dataset. This demonstrates the effectiveness of our Skip Tuning scheme for improving the robustness of the tuned model to distribution shifts. + +Domain Generalization. The domain generalization setting assesses whether models trained on a source domain can generalize to unseen/target domains. In line with those comparison methods, we consider the ImageNet as the source domain and the other four ImageNet variants as target domains. As shown in Table 2, compared to the previous state-of-the-art method PromptSRC, Skip Tuning achieves $\times 44$ time efficiency, and $\times 21.5$ memory efficiency, while establishing $1.5\%$ and $0.55\%$ ACC improvements on the source and target domains, respectively. Besides, Skip Tuning consistently outperforms those competitors on the four target domains without compromising the performance of the tuned model on the source domain, which proves the effectiveness of our Skip Tuning scheme in improving the robustness of the tuned model to domain shifts. + +Table 1. Base-to-new generalization results over 11 datasets. * indicates our reproduced results. + +
DatasetsMetricCoOp(IJCV'22)CoCoOp(CVPR'22)ProGrad(ICCV'23)KgCoOp(CVPR'23)MaPLe(CVPR'23)ProSRC(ICCV'23)TCP(CVPR'24)DePT(CVPR'24)CoPro.*ICLR'24)SkipT.(Ours)
ImgNetBase76.4775.9877.0275.8376.6677.6077.2777.0376.5377.73
New67.8870.4366.6669.9670.5470.7369.8770.1371.3070.40
H71.9273.1071.4672.7873.4774.0173.3873.4273.8273.89
CaltechBase98.0097.9698.0297.7297.7498.1098.2398.3098.6098.50
New89.3193.8193.8994.3994.3694.0394.6794.6095.1795.33
H93.7395.8495.9196.0396.0296.0296.4296.4196.8596.89
PetsBase93.6795.2095.0794.6595.4395.3394.6794.3394.7395.70
New95.2997.6997.6397.7697.7697.3097.2097.2396.7097.87
H94.4796.4396.3396.1896.5896.3095.9295.7695.7196.77
CarsBase78.1270.4977.6871.7672.9478.2780.8079.1373.1782.93
New60.4073.5968.6375.0474.0074.9774.1375.4770.6372.50
H68.1372.0172.8873.3673.4776.5877.3277.2671.8877.37
FlowersBase97.6094.8795.5495.0095.9298.0797.7398.0096.9398.57
New59.6771.7571.8774.7372.4676.5075.5776.3775.5075.80
H74.0681.7182.0383.6582.5685.9585.2385.8484.8885.70
Food101Base88.3390.7090.3790.5090.7190.6790.5790.5090.3790.67
New82.2691.2989.5991.7092.0591.5391.3791.6091.5392.03
H85.1990.9989.9891.0991.3891.1090.9791.0590.9591.34
AircraftBase40.4433.4140.5436.2137.4442.7341.9743.2036.1745.37
New22.3023.7127.5733.5535.6137.8734.4334.8334.4737.13
H28.7527.7432.8234.8336.5040.1537.8338.5735.3040.84
SUN397Base80.6079.7481.2680.2980.8282.6782.6382.3382.3082.40
New65.8976.8674.1776.5378.7078.4778.2077.8079.6379.03
H72.5178.5777.5578.3679.7580.5280.3580.0080.9480.68
DTDBase79.4477.0177.3577.5580.3683.3782.7782.2083.0083.77
New41.1856.0052.3554.9959.1862.9758.0759.1363.2067.23
H54.2464.8562.4564.3568.1671.1568.2568.7871.7674.59
EuroSATBase92.1987.4990.1185.6494.0792.9091.6389.0393.7792.47
New54.7460.0460.8964.3473.2373.9074.7371.0771.7383.00
H68.6971.2172.6773.4882.3582.3282.3279.0481.2887.48
UCF101Base84.6982.3384.3382.8983.0087.1087.1385.8086.2087.30
New56.0573.4574.9476.6778.6678.8080.7777.2378.7082.47
H67.4677.6479.3579.6580.7782.7483.8381.2982.2884.81
Avg ACCBase82.6980.4782.4880.7382.2884.2684.1383.6282.8985.04
New63.2271.6970.7573.6075.1476.1075.3675.0475.3277.53
H71.6675.8376.1677.0078.5579.9779.5179.1078.9381.11
CostTime (s)11862851231111914131735619733684239
Mem. (M)320415563204318817292041318917142060404
+ +Table 2. Cross-dataset generalization results on 11 datasets. * indicates our reproduced results. The detailed results on the 10 target datasets (i.e., Caltech, Pets, Cars, ..., and UCF101) are reported in Sup. Mat. (B). + +
DatasetsMetricCoOp(IJCV'22)CoCoOp(CVPR'22)ProGrad(ICCV'23)KgCoOp(CVPR'23)MaPLe(CVPR'23)ProSRC(ICCV'23)TCP(CVPR'24)DePT(CVPR'24)CoPro.*ICLR'24)SkipT.(Ours)
ImageNetACC71.5171.0272.2470.6670.7271.2771.4072.7772.5372.77
10 DatasetsAvg ACC63.8865.7462.7165.5166.3065.8166.2965.5564.8167.00
CostTime (s)316329391756223316364942500911617422796191611139
Mem. (M)154121362215412152541378614107152631378314131656
+ +Table 3. Domain generalization results on ImageNet. * indicates our reproduced results. The detailed results on the 4 ImgNet variants (i.e., ImgNet-V2, ImgNet-S, ImgNet-A, and ImgNet-R) are presented in Sup. Mat. (B). + +
DatasetsMetricCoOp(IJCV'22)CoCoOp(CVPR'22)ProGrad(ICCV'23)KgCoOp(CVPR'23)MaPLe(CVPR'23)ProSRC(ICCV'23)DePT(CVPR'24)CoPro.*ICLR'24)SkipT.(Ours)
ImageNetACC71.5171.0272.2470.6670.7271.2772.7772.5372.77
4 Imag.VariantsAvg ACC59.2859.9059.0760.1160.2660.6558.9760.0261.20
CostTime (s)3163293917562233163649425009122796191611139
Mem. (M)1541213622154121525413786141071378314131656
+ +# 3.3. Ablation Studies + +In this section, we conduct ablative studies to further scrutinize our devised Skip Tuning method. + +Effectiveness of the Designed Components. Our proposed Skip Tuning approach simultaneously performs Layer-wise Skipping (LSkip) and Class-wise Skipping (CSkip) to enhance the memory and time efficiency of the full fine-tuning (FT) baseline. In this experiment, we investigate the effectiveness of LSkip and CSkip by gradually adding them to + +the FT baseline. From the results in Table 4, we have the following observations. i) Both LSkip and CSkip contribute to performance improvement. ii) By combining all those components, Skip Tuning improves the effectiveness and efficiency of the baseline method remarkably. iii) The memory and time efficiency gains are more pronounced when CSkip is applied to the FT baseline, compared to the performance improvements it yields. iv) CSkip improves the memory and time efficiency of the baseline without com + +![](images/a56986e33d9254718aeb5c25f4f9e6eb82161a84cc6b1c0e1be5bbfc2cca713b.jpg) +Figure 4. Ablation studies of the number of skipped layers $\omega$ in LSkip, and the sampling rate $r$ , the decay coefficient $\lambda$ in CSkip. + +![](images/2022ca518eb0cedc67eb1a75e69d5435497eec485a4e47994078f5a29d956172.jpg) + +![](images/71906f81d5a329d75f4129a50523326eba13a5f2dea8909a2dd299b98ac95c39.jpg) + +Table 4. Ablation study on the components of Skip Tuning. + +
LSPCSPBaseNewHTime (s)Mem. (M)
XX84.9974.2679.2710021846
✓X85.0877.3081.00388861
X✓84.7774.4879.29440919
✓✓85.0477.5381.11239404
+ +Table 5. Comparison of our Skip Tuning and adapter-based methods. ${}^{ \dagger }$ denotes the efficiency-optimized versions (details are illustrated in Sup. Mat. (C)). + +
MethodBaseNewHTime (s)Mem. (M)
FT84.9974.2679.2710021846
CLIP-adapter [10]74.4873.8174.148881784
LoRA [15]80.5374.7377.529101580
SkipT. (Ours)85.0477.5381.11239404
CLIP-adapter†76.7573.5675.12137175
SkipT. (Ours)†82.6675.5678.9511567
+ +promising the generalization performance of the learned model. One possible reason is that CSkip can filter out redundant and distracting text features when performing image-text matching with the loss of $\mathcal{L}_{\mathrm{ITM}}$ . + +Impact of the LSkip Hyper-parameter $\omega$ . Our Skip Tuning approach drops out the $1 \sim \omega$ layers of the CLIP's vision and text encoders in the LSkip step. Here, we investigate the impact of $\omega$ on performance by setting $\omega$ to the values of $\{2, 4, 6, 8, 10\}$ . The obtained average testing results on the 11 datasets are reported in Figure 4 (Left). As shown, the obtained H ACCs gradually increase as the $\omega$ value decreases from 10 to 6, after which the performance gradually decreases. But, we also see that as the value of $\omega$ becomes smaller, both memory and time costs increase. Hence, we set $\omega = 6$ for LSKip in this work. + +Impact of the CSkip Hyper-parameter $r$ . In the CSkip step of Skip Tuning, we devise an image-conditioned class sampling strategy to filter out irrelevant class tokens for each training image. The larger the sampling ratio $r$ value, the more class tokens will be used to construct the training loss $\mathcal{L}_{\mathrm{ITM}}$ for each image. It is necessary to scrutinize the impact of $r$ on performance. To this end, we respectively set $r$ to $\{0.1, 0.3, 0.5, 0.7, 0.9, 1.0\}$ , and report the average testing results on the 11 datasets in Figure 4 (Mid.). From the obtained H ACCs, our method is in general not sensitive to the change of $r$ within a certain range (from 0.4 to 1.0). We also see a gradual decrease in H ACC when $r > 0.7$ , suggesting that for each training image, not all class tokens are + +beneficial for capturing task-specific knowledge. Besides, we can observe that as the value of $r$ becomes larger, both memory and time costs increase. Therefore, we set $r = 0.5$ for CSKip in this work. + +Impact of the CSkip Hyper-parameter $\lambda$ . Our Skip Tuning introduces an exponential decay coefficient $\lambda$ for the devised image-conditioned class sampling strategy in the CSkip step. We study the impact of $\lambda$ by setting $\lambda$ to the values of $\{0.01, 0.1, 0.3, 0.5, 0.7, 0.9, 1.0\}$ , and report the average testing results over the 11 datasets in Figure 4 (Right). As can be observed from the obtained H ACCs, our method is in general not sensitive to $\lambda$ when it takes the values from 0.1 to 0.7. Particularly, the H ACCs gradually decrease as $\lambda$ increases from 0.7 to 1.0. Also, we can see that as the value of $\lambda$ becomes larger, both memory and time costs increase. Thus, we set $\lambda = 0.3$ for CSkip in this work. + +# 3.4. Additional Results + +Comparison with Adapter-based Methods. Previous experimental results demonstrate the superior effectiveness and efficiency of our Skip Tuning method over existing prompt tuning methods. To further demonstrate the advantages of Skip Tuning, we also compare it with the representative adapter-based methods LoRA [15] and CLIP-adapter [10]. The base-to-new generalization results averaged on 11 datasets are reported in Table 5, where $\mathrm{CLIP - adapter}^{\dagger}$ refers to our re-implementation of CLIP-adapter, i.e., the CLIP's last-layer features are cached once before training and subsequently used to update the weights of the adapter module. As observed, our Skip Tuning offers substantial memory and time efficiency advantages over the two strong competitors, while also delivering better classification performance on both base and new tasks. + +Few-shot Learning. In the previous experiments, we follow the comparison approaches to evaluate the performance of different methods on $M$ -way 16-shot tasks—16 training examples are sampled for each of the $M$ inner-task classes. It is interesting to further scrutinize the effectiveness and efficiency of our Skip Tuning method under different shots. To this end, Figure 5 reports the few-shot learning performance of Skip Tuning and other representative competitors. As shown, our Skip Tuning demonstrates superior performance compared to other methods, with significant efficiency advantages in memory usage and time cost across + +![](images/637e120cc9d3b46297e22b25e0f6500bdd59376877ecbf88b530126884d8f322.jpg) + +![](images/e9bfe42567c0d1563635c9c73f2a1bd7473003c2bfadd548124749a40be6a8a3.jpg) + +![](images/67d756aca6943f05d057564ad0418be1cffe38783dc51765f56f909a56ed7eac.jpg) + +![](images/8b81a9afd059df47794d2a9d1e4795e805587176820b4212d422702cc4336e3e.jpg) + +![](images/07d8763f9a22b67de9073924347146d5951929ecd42ef56ed440df79c7d63dbd.jpg) + +![](images/502e1bbe316477ec48714516e583d344c54d50599226b040f28e3f6d99205e92.jpg) + +![](images/c2425b93941ba109e22b08ee13c3098cb0b2d01bca86c3a23ae8ea4b0bd5e0f1.jpg) + +![](images/d242e9023e80ef8bc6abb142d8fd40b0f759bd0aa6646d21cafe1e7f3395b714.jpg) + +![](images/0d7a1a4ed196428b468f7216e5f216a82f9f116f8e6a1a28bc100e64ff07cf6b.jpg) + +![](images/944be23fbc852114182e304c10950aaa0b722ea51a99ca2a1e710348a60681ff.jpg) + +![](images/0aa2668f617df4861881d515c83de42b0ff64647b4ecaa361fdb0a51d9a9e2ff.jpg) + +![](images/71aa3663652bf038a1204914f6b1313570ca77b602e74ae056887d1ea09232f9.jpg) +Figure 5. Few-shot learning results on 11 datasets. + +![](images/1216882fa9d6ea2cfcaeba90ccbb4dd0a0b44eba3805d19a786992ad858d9f1e.jpg) + +![](images/eb790603167035d504a6ba61adbe58e6d09da8ac941391b2ad7dff8bf71e3e84.jpg) + +1-shot to 16-shot settings. Additionally, while PromptSRC and MaPLe enhance CoOp's performance, both methods incur increased time cost. In contrast, Skip Tuning achieves both effectiveness and efficiency without compromise, further underscoring the superiority of our approach. + +# 4. Related Work + +Pre-trained Vision-Language Models. Pre-trained vision-language models (VLMs) have garnered great attention recently. Notable models such as CLIP[24], ALIGN[16], LiT[36] and FILIP[34] have demonstrated exceptional performance across various vision-language tasks. These models are pre-trained on extensive datasets comprising image-text pairs sourced from the internet. For instance, CLIP[24] and ALIGN[16] respectively utilize over 400 million and 1 billion pairs, for training. The large scale of training data enables these models to excel in open-vocabulary image-text retrieval and zero-shot classification. Additionally, they have been successfully applied to downstream tasks including image classification[10, 40], object detection[6, 8, 11, 35, 43], and semantic segmentation[20, 25]. In this work, we focus on the effective adaptation of vision-language models to downstream visual recognition tasks. + +Prompt Tuning. Prompt tuning (a.k.a. context optimization [42]) has emerged as a parameter-efficient learning paradigm to adapt powerful VLMs to downstream tasks [7, 9, 37, 38]. As a representative method, CoOp [42] enables task adaptation by optimizing a set of prompt vectors within CLIP's language branch. While effective, CoOp often suffers from limited generalization on new tasks due to + +overfitting to the base task. A series of schemes are devised to tackle this problem, e.g. CoCoOp[41], KgCoOp[32], ProGrad[44], TCP [33], and DePT[39]. By adding trainable prompts into both the image and text branches of CLIP, MaPLe[17], PromptSRC[18], and CoPrompt[27] demonstrate remarkable performance on both base and new tasks. Despite the advantages, we reveal that freezing the parameters of VLMs during learning the context vectors neither facilitates the generalization of pre-trained knowledge nor significantly improves memory and time efficiency. + +# 5. Conclusion + +In this work, we first reveal that freezing the parameters of VLMs during prompt tuning neither facilitates the transferability of pre-trained knowledge nor improves memory and time efficiency considerably. To circumvent this limitation, we propose Skip Tuning, an effective and efficient method for transferring VLMs to downstream tasks without relying on extra context vectors or adapter modules. Extensive experiments across a broad range of benchmarks demonstrate the superiority of our Skip Tuning method over both prompt tuning and adapter-based approaches. + +Acknowledgements. This study is supported by grants from the National Natural Science Foundation of China (Grant No. 62425208, No. U22A2097, No. 62122018, No. 62020106008, No. U23A20315, No. 82441006, No. 62402094), the Postdoctoral Fellowship Program of CPSF (Grant No. GZB20240625), the Science and Technology Innovation Committee of Shenzhen Municipality Foundation (Grant No.JCYJ20240813114208012). + +# References + +[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 1 +[2] Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. Food-101-mining discriminative components with random forests. In Computer vision-ECCV 2014: 13th European conference, zurich, Switzerland, September 6-12, 2014, proceedings, part VI 13, pages 446-461. Springer, 2014. 5 +[3] Mircea Cimpoi, Subhransu Maji, Iasonas Kokkinos, Sammy Mohamed, and Andrea Vedaldi. Describing textures in the wild. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3606-3613, 2014. 5 +[4] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 5 +[5] Li Fei-Fei, Rob Fergus, and Pietro Perona. Learning generative visual models from few training examples: An incremental bayesian approach tested on 101 object categories. Computer vision and Image understanding, 106(1):59-70, 2007. 5 +[6] Chengjian Feng, Yujie Zhong, Zequn Jie, Xiangxiang Chu, Haibing Ren, Xiaolin Wei, Weidi Xie, and Lin Ma. Promptdet: Towards open-vocabulary detection using uncurated images. In European Conference on Computer Vision, pages 701-717. Springer, 2022. 8 +[7] Yuqian Fu, Yanwei Fu, Jingjing Chen, and Yu-Gang Jiang. Generalized meta-fdmixup: Cross-domain few-shot learning guided by labeled target data. IEEE Transactions on Image Processing, 31:7078-7090, 2022. 8 +[8] Yuqian Fu, Yu Xie, Yanwei Fu, and Yu-Gang Jiang. Styleadv: Meta style adversarial training for cross-domain few-shot learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 24575-24584, 2023. 8 +[9] Yuqian Fu, Yu Wang, Yixuan Pan, Lian Huai, Xingyu Qiu, Zeyu Shangguan, Tong Liu, Yanwei Fu, Luc Van Gool, and Xingqun Jiang. Cross-domain few-shot object detection via enhanced open-set object detector. In ECCV, pages 247-264. Springer, 2025. 8 +[10] Peng Gao, Shijie Geng, Renrui Zhang, Teli Ma, Rongyao Fang, Yongfeng Zhang, Hongsheng Li, and Yu Qiao. Clip-adapter: Better vision-language models with feature adapters. International Journal of Computer Vision, 132(2): 581-595, 2024. 7, 8 +[11] Xiuye Gu, Tsung-Yi Lin, Weicheng Kuo, and Yin Cui. Open-vocabulary object detection via vision and language knowledge distillation. arXiv preprint arXiv:2104.13921, 2021.8 +[12] Patrick Helber, Benjamin Bischke, Andreas Dengel, and Damian Borth. Eurosat: A novel dataset and deep learning benchmark for land use and land cover classification. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 12(7):2217-2226, 2019. 5 + +[13] Dan Hendrycks, Steven Basart, Norman Mu, Saurav Kadavath, Frank Wang, Evan Dorundo, Rahul Desai, Tyler Zhu, Samyak Parajuli, Mike Guo, et al. The many faces of robustness: A critical analysis of out-of-distribution generalization. In Proceedings of the IEEE/CVF international conference on computer vision, pages 8340-8349, 2021. 5 +[14] Dan Hendrycks, Kevin Zhao, Steven Basart, Jacob Steinhardt, and Dawn Song. Natural adversarial examples. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15262-15271, 2021. 5 +[15] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 2, 7 +[16] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International conference on machine learning, pages 4904-4916. PMLR, 2021. 8 +[17] Muhammad Uzair Khattak, Hanoona Rasheed, Muhammad Maaz, Salman Khan, and Fahad Shahbaz Khan. Maple: Multi-modal prompt learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19113-19122, 2023. 8 +[18] Muhammad Uzair Khattak, Syed Talal Wasim, Muzammal Naseer, Salman Khan, Ming-Hsuan Yang, and Fahad Shahbaz Khan. Self-regulating prompts: Foundational model adaptation without forgetting. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15190-15200, 2023. 8 +[19] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In Proceedings of the IEEE international conference on computer vision workshops, pages 554–561, 2013. 5 +[20] Timo Lüddecke and Alexander Ecker. Image segmentation using text and image prompts. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7086-7096, 2022. 8 +[21] Subhransu Maji, Esa Rahtu, Juho Kannala, Matthew Blaschko, and Andrea Vedaldi. Fine-grained visual classification of aircraft. arXiv preprint arXiv:1306.5151, 2013. 5 +[22] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In 2008 Sixth Indian conference on computer vision, graphics & image processing, pages 722-729. IEEE, 2008. 5 +[23] Omkar M Parkhi, Andreae Vedaldi, Andrew Zisserman, and CV Jawahar. Cats and dogs. In 2012 IEEE conference on computer vision and pattern recognition, pages 3498-3505. IEEE, 2012. 5 +[24] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 2, 8 + +[25] Yongming Rao, Wenliang Zhao, Guangyi Chen, Yansong Tang, Zheng Zhu, Guan Huang, Jie Zhou, and Jiwen Lu. Denseclip: Language-guided dense prediction with context-aware prompting. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18082-18091, 2022. 8 +[26] Benjamin Recht, Rebecca Roelofs, Ludwig Schmidt, and Vaishaal Shankar. Do imagenet classifiers generalize to imagenet? In International conference on machine learning, pages 5389-5400. PMLR, 2019. 5 +[27] Shuvendu Roy and Ali Etemad. Consistency-guided prompt learning for vision-language models. arXiv preprint arXiv:2306.01195, 2023. 8 +[28] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012. 5 +[29] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 1 +[30] Haohan Wang, Songwei Ge, Zachary Lipton, and Eric P Xing. Learning robust global representations by penalizing local predictive power. Advances in Neural Information Processing Systems, 32, 2019. 5 +[31] Jianxiong Xiao, James Hays, Krista A Ehinger, Aude Oliva, and Antonio Torralba. Sun database: Large-scale scene recognition from abbey to zoo. In 2010 IEEE computer society conference on computer vision and pattern recognition, pages 3485-3492. IEEE, 2010. 5 +[32] Hantao Yao, Rui Zhang, and Changsheng Xu. Visual-language prompt tuning with knowledge-guided context optimization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6757-6767, 2023. 8 +[33] Hantao Yao, Rui Zhang, and Changsheng Xu. Tcp: Textual-based class-aware prompt tuning for visual-language model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23438-23448, 2024. 8 +[34] Lewei Yao, Runhui Huang, Lu Hou, Guansong Lu, Minzhe Niu, Hang Xu, Xiaodan Liang, Zhenguo Li, Xin Jiang, and Chunjing Xu. Filip: Fine-grained interactive language-image pre-training. arXiv preprint arXiv:2111.07783, 2021. 8 +[35] Yuhang Zang, Wei Li, Kaiyang Zhou, Chen Huang, and Chen Change Loy. Open-vocabulary detr with conditional matching. In European Conference on Computer Vision, pages 106-122. Springer, 2022. 8 +[36] Xiaohua Zhai, Xiao Wang, Basil Mustafa, Andreas Steiner, Daniel Keysers, Alexander Kolesnikov, and Lucas Beyer. Lit: Zero-shot transfer with locked-image text tuning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18123-18133, 2022. 8 +[37] Ji Zhang, Jingkuan Song, Lianli Gao, and Hengtao Shen. Free-lunch for cross-domain few-shot learning: Style-aware episodic training with robust contrastive learning. In Proceedings of the 30th ACM International Conference on Multimedia, pages 2586-2594, 2022. 8 + +[38] Ji Zhang, Lianli Gao, Xu Luo, Hengtao Shen, and Jingkuan Song. Deta: Denoised task adaptation for few-shot learning. In ICCV, pages 11541-11551, 2023. 8 +[39] Ji Zhang, Shihan Wu, Lianli Gao, Heng Tao Shen, and Jingkuan Song. Dept: Decoupled prompt tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12924-12933, 2024. 5, 8 +[40] Renrui Zhang, Rongyao Fang, Wei Zhang, Peng Gao, Kunchang Li, Jifeng Dai, Yu Qiao, and Hongsheng Li. Tip-adapter: Training-free clip-adapter for better vision-language modeling. arXiv preprint arXiv:2111.03930, 2021. 8 +[41] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Conditional prompt learning for vision-language models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16816-16825, 2022. 8 +[42] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision, 130(9):2337-2348, 2022. 2, 3, 8 +[43] Xingyi Zhou, Rohit Girdhar, Armand Joulin, Philipp Krahenbihl, and Ishan Misra. Detecting twenty-thousand classes using image-level supervision. In European Conference on Computer Vision, pages 350-368. Springer, 2022. 8 +[44] Beier Zhu, Yulei Niu, Yucheng Han, Yue Wu, and Hanwang Zhang. Prompt-aligned gradient for prompt tuning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15659-15669, 2023. 8 \ No newline at end of file diff --git a/2025/Skip Tuning_ Pre-trained Vision-Language Models are Effective and Efficient Adapters Themselves/images.zip b/2025/Skip Tuning_ Pre-trained Vision-Language Models are Effective and Efficient Adapters Themselves/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..a12c7517d46bd77b905e3efd37878e1234609bcf --- /dev/null +++ b/2025/Skip Tuning_ Pre-trained Vision-Language Models are Effective and Efficient Adapters Themselves/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1663fa8706308124d5f16b8d61f0d4fd7be85d28f99b3ef67de1737417344f18 +size 813307 diff --git a/2025/Skip Tuning_ Pre-trained Vision-Language Models are Effective and Efficient Adapters Themselves/layout.json b/2025/Skip Tuning_ Pre-trained Vision-Language Models are Effective and Efficient Adapters Themselves/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..dac3d1f4f1c16a0234c79d5b296c0093b2b3cceb --- /dev/null +++ b/2025/Skip Tuning_ Pre-trained Vision-Language Models are Effective and Efficient Adapters Themselves/layout.json @@ -0,0 +1,10580 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 136, + 103, + 475, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 103, + 475, + 140 + ], + "spans": [ + { + "bbox": [ + 136, + 103, + 475, + 140 + ], + "type": "text", + "content": "Skip Tuning: Pre-trained Vision-Language Models are Effective and Efficient Adapters Themselves" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 70, + 159, + 541, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 159, + 541, + 176 + ], + "spans": [ + { + "bbox": [ + 70, + 159, + 541, + 176 + ], + "type": "text", + "content": "Shihan Wu" + }, + { + "bbox": [ + 70, + 159, + 541, + 176 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 70, + 159, + 541, + 176 + ], + "type": "text", + "content": " Ji Zhang" + }, + { + "bbox": [ + 70, + 159, + 541, + 176 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 70, + 159, + 541, + 176 + ], + "type": "text", + "content": " Pengpeng Zeng" + }, + { + "bbox": [ + 70, + 159, + 541, + 176 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 70, + 159, + 541, + 176 + ], + "type": "text", + "content": " Lianli Gao" + }, + { + "bbox": [ + 70, + 159, + 541, + 176 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 70, + 159, + 541, + 176 + ], + "type": "text", + "content": " Jingkuan Song" + }, + { + "bbox": [ + 70, + 159, + 541, + 176 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 70, + 159, + 541, + 176 + ], + "type": "text", + "content": " Heng Tao Shen" + }, + { + "bbox": [ + 70, + 159, + 541, + 176 + ], + "type": "inline_equation", + "content": "^{4}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 136, + 176, + 474, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 176, + 474, + 190 + ], + "spans": [ + { + "bbox": [ + 136, + 176, + 474, + 190 + ], + "type": "text", + "content": "1 University of Electronic Science and Technology of China (UESTC)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 227, + 189, + 383, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 227, + 189, + 383, + 204 + ], + "spans": [ + { + "bbox": [ + 227, + 189, + 383, + 204 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 227, + 189, + 383, + 204 + ], + "type": "text", + "content": " Southwest Jiaotong University" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 184, + 203, + 425, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 184, + 203, + 425, + 217 + ], + "spans": [ + { + "bbox": [ + 184, + 203, + 425, + 217 + ], + "type": "text", + "content": "3 Shenzhen Institute for Advanced Study, UESTC" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 258, + 217, + 352, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 258, + 217, + 352, + 232 + ], + "spans": [ + { + "bbox": [ + 258, + 217, + 352, + 232 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 258, + 217, + 352, + 232 + ], + "type": "text", + "content": " Tongji University" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 165, + 233, + 444, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 233, + 444, + 245 + ], + "spans": [ + { + "bbox": [ + 165, + 233, + 444, + 245 + ], + "type": "text", + "content": "{shihan.wu.koorye@outlook.com, jizhang.jim@gmail.com}" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 152, + 273, + 200, + 285 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 273, + 200, + 285 + ], + "spans": [ + { + "bbox": [ + 152, + 273, + 200, + 285 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 299, + 296, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 299, + 296, + 550 + ], + "spans": [ + { + "bbox": [ + 53, + 299, + 296, + 550 + ], + "type": "text", + "content": "Prompt tuning (PT) has long been recognized as an effective and efficient paradigm for transferring large pre-trained vision-language models (VLMs) to downstream tasks by learning a tiny set of context vectors. Nevertheless, in this work, we reveal that freezing the parameters of VLMs during learning the context vectors neither facilitates the transferability of pre-trained knowledge nor improves the memory and time efficiency significantly. Upon further investigation, we find that reducing both the length and width of the feature-gradient propagation flows of the full fine-tuning (FT) baseline is key to achieving effective and efficient knowledge transfer. Motivated by this, we propose Skip Tuning, a novel paradigm for adapting VLMs to downstream tasks. Unlike existing PT or adapter-based methods, Skip Tuning applies Layer-wise Skipping (LSkip) and Classwise Skipping (CSkip) upon the FT baseline without introducing extra context vectors or adapter modules. Extensive experiments across a wide spectrum of benchmarks demonstrate the superior effectiveness and efficiency of our Skip Tuning over both PT and adapter-based methods. Code: https://github.com/Koorye/SkipTuning." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 575, + 135, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 575, + 135, + 588 + ], + "spans": [ + { + "bbox": [ + 56, + 575, + 135, + 588 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 596, + 296, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 596, + 296, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 596, + 296, + 693 + ], + "type": "text", + "content": "There have recently been significant advancements in large pre-trained vision-language models (VLMs) [1, 24, 29]. One notable achievement is the CLIP model [24], which leverages an image-text matching (ITM) loss to align images with their corresponding textual descriptions in a common feature space. While VLMs have proven impressive capabilities in recognizing open-set visual concepts, their zero-shot generalization performance declines significantly" + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 325, + 270, + 417, + 335 + ], + "blocks": [ + { + "bbox": [ + 325, + 270, + 417, + 335 + ], + "lines": [ + { + "bbox": [ + 325, + 270, + 417, + 335 + ], + "spans": [ + { + "bbox": [ + 325, + 270, + 417, + 335 + ], + "type": "image", + "image_path": "52012149a2b74e739ce937db35d73275a6d735cd932942e51933cea05c8c5807.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 419, + 271, + 481, + 335 + ], + "blocks": [ + { + "bbox": [ + 419, + 271, + 481, + 335 + ], + "lines": [ + { + "bbox": [ + 419, + 271, + 481, + 335 + ], + "spans": [ + { + "bbox": [ + 419, + 271, + 481, + 335 + ], + "type": "image", + "image_path": "04f7f904d4ff9c125ba4a0e81dac82a9846d98c05f9ee7c626fa072ca02e18bb.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 483, + 271, + 544, + 335 + ], + "blocks": [ + { + "bbox": [ + 483, + 271, + 544, + 335 + ], + "lines": [ + { + "bbox": [ + 483, + 271, + 544, + 335 + ], + "spans": [ + { + "bbox": [ + 483, + 271, + 544, + 335 + ], + "type": "image", + "image_path": "637edd4d4f795e3863d9bcf476dce04b395743401f859277d5691b1a966f2048.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 324, + 337, + 417, + 401 + ], + "blocks": [ + { + "bbox": [ + 324, + 337, + 417, + 401 + ], + "lines": [ + { + "bbox": [ + 324, + 337, + 417, + 401 + ], + "spans": [ + { + "bbox": [ + 324, + 337, + 417, + 401 + ], + "type": "image", + "image_path": "bec99f81e1ebb308bfa5763dfad5895e68ed2218fc8f2381cfde13d9869fb887.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 419, + 337, + 479, + 402 + ], + "blocks": [ + { + "bbox": [ + 419, + 337, + 479, + 402 + ], + "lines": [ + { + "bbox": [ + 419, + 337, + 479, + 402 + ], + "spans": [ + { + "bbox": [ + 419, + 337, + 479, + 402 + ], + "type": "image", + "image_path": "9b9f9f7401feb7f2a82f77008366704c2164290c98137ac1b5355eccc217218a.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 483, + 337, + 544, + 402 + ], + "blocks": [ + { + "bbox": [ + 483, + 337, + 544, + 402 + ], + "lines": [ + { + "bbox": [ + 483, + 337, + 544, + 402 + ], + "spans": [ + { + "bbox": [ + 483, + 337, + 544, + 402 + ], + "type": "image", + "image_path": "dd0d41d243ed1a609d0d1887ed982960c1468b9a7939a0f755c15c3de6a0b566.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 324, + 403, + 417, + 462 + ], + "blocks": [ + { + "bbox": [ + 324, + 403, + 417, + 462 + ], + "lines": [ + { + "bbox": [ + 324, + 403, + 417, + 462 + ], + "spans": [ + { + "bbox": [ + 324, + 403, + 417, + 462 + ], + "type": "image", + "image_path": "892b3907047f74164c64be38f5db12e9685982e621fd673df64374abd0e74aa6.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 419, + 403, + 480, + 462 + ], + "blocks": [ + { + "bbox": [ + 419, + 403, + 480, + 462 + ], + "lines": [ + { + "bbox": [ + 419, + 403, + 480, + 462 + ], + "spans": [ + { + "bbox": [ + 419, + 403, + 480, + 462 + ], + "type": "image", + "image_path": "99126689c80abbc6f16d9272f31f9e7a8f9f67512237fb242416a164aadbfcaf.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 483, + 403, + 544, + 462 + ], + "blocks": [ + { + "bbox": [ + 483, + 403, + 544, + 462 + ], + "lines": [ + { + "bbox": [ + 483, + 403, + 544, + 462 + ], + "spans": [ + { + "bbox": [ + 483, + 403, + 544, + 462 + ], + "type": "image", + "image_path": "f48f2194776fafc7819003b22098ad789e6d07072cf53a25c4b26fc551073fe5.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 324, + 464, + 417, + 502 + ], + "blocks": [ + { + "bbox": [ + 324, + 464, + 417, + 502 + ], + "lines": [ + { + "bbox": [ + 324, + 464, + 417, + 502 + ], + "spans": [ + { + "bbox": [ + 324, + 464, + 417, + 502 + ], + "type": "image", + "image_path": "a9b5c20b2d3e3466709c90502a5577ae53031f74f8227a671ee70c4c83de655b.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 505, + 555, + 582 + ], + "lines": [ + { + "bbox": [ + 313, + 505, + 555, + 582 + ], + "spans": [ + { + "bbox": [ + 313, + 505, + 555, + 582 + ], + "type": "text", + "content": "Figure 1. Comparison of our devised Skip Tuning with state-of-the-art prompt tuning methods in terms of training time (seconds), memory cost (M), and classification accuracy (\\%) across base-to-new generalization, cross-dataset generalization, domain generalization, and few-shot learning benchmarks. " + }, + { + "bbox": [ + 313, + 505, + 555, + 582 + ], + "type": "inline_equation", + "content": "\\times" + }, + { + "bbox": [ + 313, + 505, + 555, + 582 + ], + "type": "text", + "content": " indicates the performance improvement over the state-of-the-art. Comparison results with the adapter-based methods are reported in Table 5." + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 419, + 464, + 480, + 502 + ], + "blocks": [ + { + "bbox": [ + 419, + 464, + 480, + 502 + ], + "lines": [ + { + "bbox": [ + 419, + 464, + 480, + 502 + ], + "spans": [ + { + "bbox": [ + 419, + 464, + 480, + 502 + ], + "type": "image", + "image_path": "a776cef8fef9e06066ad2d541d3809fea0fd63c01c6fdb6c73fd2548c3db8333.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 483, + 464, + 544, + 502 + ], + "blocks": [ + { + "bbox": [ + 483, + 464, + 544, + 502 + ], + "lines": [ + { + "bbox": [ + 483, + 464, + 544, + 502 + ], + "spans": [ + { + "bbox": [ + 483, + 464, + 544, + 502 + ], + "type": "image", + "image_path": "d7b1f847708d48087fb5011bd853344d6939b1a79c8e66c3f2500faa2fc44f39.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 605, + 555, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 555, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 555, + 628 + ], + "type": "text", + "content": "when encountering category, distribution, or domain shifts between upstream training data and downstream tasks." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "content": "Prompt tuning (PT) has long been recognized as an effective and efficient paradigm for transferring large pre-trained vision-language models (VLMs) to downstream tasks. The core concept of PT is to learn a task-specific prompt (i.e., a small number of context vectors) for the target task, using a limited amount of training data, while keeping the pre-trained VLM parameters fixed. Although many PT ap" + } + ] + } + ], + "index": 27 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "type": "text", + "content": "*Corresponding author." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14723" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 77, + 69, + 178, + 132 + ], + "blocks": [ + { + "bbox": [ + 77, + 69, + 178, + 132 + ], + "lines": [ + { + "bbox": [ + 77, + 69, + 178, + 132 + ], + "spans": [ + { + "bbox": [ + 77, + 69, + 178, + 132 + ], + "type": "image", + "image_path": "5fe69620d9a1ae354053712bd3ac8d873539f6be59708020154b7fa6e50585f8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 181, + 70, + 294, + 133 + ], + "blocks": [ + { + "bbox": [ + 181, + 70, + 294, + 133 + ], + "lines": [ + { + "bbox": [ + 181, + 70, + 294, + 133 + ], + "spans": [ + { + "bbox": [ + 181, + 70, + 294, + 133 + ], + "type": "image", + "image_path": "355e9d569600d205411f954e5a5a8665de6909ac844109780573ec7120975b0c.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 58, + 137, + 178, + 198 + ], + "blocks": [ + { + "bbox": [ + 58, + 137, + 178, + 198 + ], + "lines": [ + { + "bbox": [ + 58, + 137, + 178, + 198 + ], + "spans": [ + { + "bbox": [ + 58, + 137, + 178, + 198 + ], + "type": "image", + "image_path": "7196e6fe8ee7b931c23df993265fa5e4ca97222874310cc5a79386d8d5c2581b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 181, + 137, + 294, + 198 + ], + "blocks": [ + { + "bbox": [ + 181, + 137, + 294, + 198 + ], + "lines": [ + { + "bbox": [ + 181, + 137, + 294, + 198 + ], + "spans": [ + { + "bbox": [ + 181, + 137, + 294, + 198 + ], + "type": "image", + "image_path": "b43f70d52bfa825b77f644a21af4c50c31c99c9f888a2baa5da7acd5148dca2b.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 58, + 202, + 178, + 262 + ], + "blocks": [ + { + "bbox": [ + 58, + 202, + 178, + 262 + ], + "lines": [ + { + "bbox": [ + 58, + 202, + 178, + 262 + ], + "spans": [ + { + "bbox": [ + 58, + 202, + 178, + 262 + ], + "type": "image", + "image_path": "cddd56924c33fbd6b808bbb6094a42e45ca2de58bf82b99467f96ef0b8c6b5c4.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 264, + 295, + 341 + ], + "lines": [ + { + "bbox": [ + 55, + 264, + 295, + 341 + ], + "spans": [ + { + "bbox": [ + 55, + 264, + 295, + 341 + ], + "type": "text", + "content": "Figure 2. Motivations. (a) Comparison between the prompt tuning (PT) method CoOp [42] and the full fine-tuning (FT) baseline in terms of i) the number of learnable parameters, ii) memory usage, iii) time cost, and iv) base-to-new generalization performance. (b) Feature Sensitivity (FS) of CLIP's network layers, averaged over 100 randomly-sampled training images. (c) Gradient Dependence (GD) of class tokens for different training images." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 181, + 202, + 294, + 262 + ], + "blocks": [ + { + "bbox": [ + 181, + 202, + 294, + 262 + ], + "lines": [ + { + "bbox": [ + 181, + 202, + 294, + 262 + ], + "spans": [ + { + "bbox": [ + 181, + 202, + 294, + 262 + ], + "type": "image", + "image_path": "3e3a8eedd5ace8dd0c18c7c7d29bd159fa7c17a6e8f2af8e9ec8695803c7e5bb.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 371, + 295, + 670 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 371, + 295, + 670 + ], + "spans": [ + { + "bbox": [ + 55, + 371, + 295, + 670 + ], + "type": "text", + "content": "proaches have reported improved performance and parameter efficiency over the full fine-tuning (FT) baseline, the discrepancy of implementation details among those PT approaches obscure the actual performance enhancement. For example, the FT performance can be significantly underestimated by training with coarsely tuned hyper-parameters. Quantitative evidence is presented in Figure 2 (a), where we make comparisons between the PT method CoOp [42] and the FT baseline in terms of the number of learnable parameters, memory usage, time cost as well as base-to-new generalization performance (see §2.2). As observed, although PT significantly improves the parameter efficiency of FT (using 1/51200 parameters of FT), the improvements in memory and time efficiency are relatively insignificant (reducing only " + }, + { + "bbox": [ + 55, + 371, + 295, + 670 + ], + "type": "inline_equation", + "content": "6.3\\%" + }, + { + "bbox": [ + 55, + 371, + 295, + 670 + ], + "type": "text", + "content": " memory usage and " + }, + { + "bbox": [ + 55, + 371, + 295, + 670 + ], + "type": "inline_equation", + "content": "15.8\\%" + }, + { + "bbox": [ + 55, + 371, + 295, + 670 + ], + "type": "text", + "content": " time cost). Besides, compared to FT, the classification accuracy of PT decreases by " + }, + { + "bbox": [ + 55, + 371, + 295, + 670 + ], + "type": "inline_equation", + "content": "3.49\\%" + }, + { + "bbox": [ + 55, + 371, + 295, + 670 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 371, + 295, + 670 + ], + "type": "inline_equation", + "content": "4.49\\%" + }, + { + "bbox": [ + 55, + 371, + 295, + 670 + ], + "type": "text", + "content": " on base and new tasks, respectively. This suggests that pursuing higher parameter efficiency by freezing the overwhelming majority weights of VLMs during learning the context vectors neither facilitates the transferability of pre-trained knowledge nor improves memory and time efficiency considerably. Besides, in many real-world applications, memory and time efficiency often take precedence over parameter efficiency in terms of practical importance. We therefore raise the following question:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 65, + 675, + 285, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 675, + 285, + 712 + ], + "spans": [ + { + "bbox": [ + 65, + 675, + 285, + 712 + ], + "type": "text", + "content": "Can we optimize the memory and time efficiency of the FT baseline and adapt VLMs to downstream tasks in an effective and efficient manner?" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 553, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 251 + ], + "type": "text", + "content": "To answer the above question, we scrutinize the FeatureGradient Propagation Flows (FGPFs) in the vision encoder and text encoder of the CLIP model when performing FT on the base (or target) task. Interestingly, we observe that, for each training image, the majority of shallow network layers and class tokens contribute minimally to capturing task-specific knowledge for the base task (see §2.2). Motivated by this, we propose Skip Tuning, a novel paradigm for adapting VLMs to downstream tasks without introducing extra context vectors or adapter modules. Concretely, Skip Tuning incorporates two strategies, Layer-wise Skipping (LSkip) and Class-wise Skipping (CSkip), to simultaneously reduce the length and width of FGFs in the FT baseline, thereby establishing effective and efficient knowledge transfer of VLMs, as shown in Figure 3." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 251, + 553, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 251, + 553, + 407 + ], + "spans": [ + { + "bbox": [ + 313, + 251, + 553, + 407 + ], + "type": "text", + "content": "Effectiveness and Efficiency. We conduct extensive experiments across a wide range of benchmarks to validate the effectiveness and efficiency of our Skip Tuning scheme. An overview of the achieved results is shown in Figure 1. As seen, our Skip Tuning demonstrates superiority over existing PT methods, e.g., on the few-shot learning benchmark, Skip Tuning achieves " + }, + { + "bbox": [ + 313, + 251, + 553, + 407 + ], + "type": "inline_equation", + "content": "\\times 15" + }, + { + "bbox": [ + 313, + 251, + 553, + 407 + ], + "type": "text", + "content": " time efficiency, " + }, + { + "bbox": [ + 313, + 251, + 553, + 407 + ], + "type": "inline_equation", + "content": "\\times 6.4" + }, + { + "bbox": [ + 313, + 251, + 553, + 407 + ], + "type": "text", + "content": " memory efficiency, while yielding a " + }, + { + "bbox": [ + 313, + 251, + 553, + 407 + ], + "type": "inline_equation", + "content": "1.04\\%" + }, + { + "bbox": [ + 313, + 251, + 553, + 407 + ], + "type": "text", + "content": " improvement in ACC compared to the state-of-the-art. Furthermore, we also show the advantages of Skip Tuning over existing adapter-based methods in Table 5, where Skip Tuning achieves " + }, + { + "bbox": [ + 313, + 251, + 553, + 407 + ], + "type": "inline_equation", + "content": "\\times 3.8" + }, + { + "bbox": [ + 313, + 251, + 553, + 407 + ], + "type": "text", + "content": " time efficiency, " + }, + { + "bbox": [ + 313, + 251, + 553, + 407 + ], + "type": "inline_equation", + "content": "\\times 3.9" + }, + { + "bbox": [ + 313, + 251, + 553, + 407 + ], + "type": "text", + "content": " memory efficiency, along with a " + }, + { + "bbox": [ + 313, + 251, + 553, + 407 + ], + "type": "inline_equation", + "content": "3.59\\%" + }, + { + "bbox": [ + 313, + 251, + 553, + 407 + ], + "type": "text", + "content": " H ACC enhancement over the strong rival LoRA [15]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 407, + 553, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 407, + 553, + 418 + ], + "spans": [ + { + "bbox": [ + 313, + 407, + 553, + 418 + ], + "type": "text", + "content": "Main Contributions. The main contributions are threefold." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 419, + 553, + 539 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 314, + 419, + 553, + 466 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 419, + 553, + 466 + ], + "spans": [ + { + "bbox": [ + 314, + 419, + 553, + 466 + ], + "type": "text", + "content": "- We reveal that reducing both the width and length of the feature-gradient propagation flows (FGPFs) of the full fine-tuning (FT) baseline is key to establishing effective and efficient knowledge transfer." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 467, + 553, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 467, + 553, + 502 + ], + "spans": [ + { + "bbox": [ + 314, + 467, + 553, + 502 + ], + "type": "text", + "content": "- We devise Skip Tuning, an effective and efficient method for transferring VLMs to downstream tasks without relying on extra context vectors or adapter modules." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 314, + 502, + 553, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 502, + 553, + 539 + ], + "spans": [ + { + "bbox": [ + 314, + 502, + 553, + 539 + ], + "type": "text", + "content": "- We evaluate our method on a wide spectrum of benchmarks, demonstrating the superiority of Skip Tuning over both prompt tuning and adapter-based approaches." + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 549, + 397, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 549, + 397, + 562 + ], + "spans": [ + { + "bbox": [ + 314, + 549, + 397, + 562 + ], + "type": "text", + "content": "2. Methodology" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 569, + 553, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 569, + 553, + 594 + ], + "spans": [ + { + "bbox": [ + 313, + 569, + 553, + 594 + ], + "type": "text", + "content": "In this section, we elaborate on our Skip Tuning approach. We start with an introduction of preliminaries." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 600, + 400, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 600, + 400, + 611 + ], + "spans": [ + { + "bbox": [ + 313, + 600, + 400, + 611 + ], + "type": "text", + "content": "2.1. Preliminaries" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 617, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 553, + 713 + ], + "type": "text", + "content": "Contrastive Language-Image Pre-training (CLIP). Following the common practice of existing paradigms for adapting PVLMs, we adopt CLIP [24] as the testbed in this work. CLIP aims to learn an alignment between image and text features generated by an image encoder and a text encoder, respectively. By being exposed to 400 million imagetext association pairs and employing a contrastive learning paradigm within a shared feature space, CLIP acquires a di" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14724" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 68, + 553, + 216 + ], + "blocks": [ + { + "bbox": [ + 57, + 68, + 553, + 216 + ], + "lines": [ + { + "bbox": [ + 57, + 68, + 553, + 216 + ], + "spans": [ + { + "bbox": [ + 57, + 68, + 553, + 216 + ], + "type": "image", + "image_path": "70dad74a6f292a14550ffedc09571bf12c8f1b5ed293682e9f3ba2ca84e6c3c4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 220, + 555, + 266 + ], + "lines": [ + { + "bbox": [ + 55, + 220, + 555, + 266 + ], + "spans": [ + { + "bbox": [ + 55, + 220, + 555, + 266 + ], + "type": "text", + "content": "Figure 3. Overview of our proposed Skip Tuning. Skip Tuning performs Layer-wise Skipping (LSkip) and Class-wise Skipping (CSkip)) to enhance the memory and time efficiency of the FT baseline. Specifically, LSkip reduces the length of feature-gradient propagation flows (FGPFs) by caching intermediate features produced by the " + }, + { + "bbox": [ + 55, + 220, + 555, + 266 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 55, + 220, + 555, + 266 + ], + "type": "text", + "content": "-th layers of CLIP's vision encoder " + }, + { + "bbox": [ + 55, + 220, + 555, + 266 + ], + "type": "inline_equation", + "content": "E_{V}" + }, + { + "bbox": [ + 55, + 220, + 555, + 266 + ], + "type": "text", + "content": " and text encoder " + }, + { + "bbox": [ + 55, + 220, + 555, + 266 + ], + "type": "inline_equation", + "content": "E_{T}" + }, + { + "bbox": [ + 55, + 220, + 555, + 266 + ], + "type": "text", + "content": " before FT begins. In contrast, CSkip reduces the width of FGFs by filtering out unimportant class tokens in the text encoder " + }, + { + "bbox": [ + 55, + 220, + 555, + 266 + ], + "type": "inline_equation", + "content": "E_{T}" + }, + { + "bbox": [ + 55, + 220, + 555, + 266 + ], + "type": "text", + "content": " for every training image." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 285, + 294, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 285, + 294, + 416 + ], + "spans": [ + { + "bbox": [ + 54, + 285, + 294, + 416 + ], + "type": "text", + "content": "verse array of open-set visual concepts that can be readily applied to downstream tasks. For instance, zero-shot classification can be achieved by framing the classification task as an image-text matching problem. Initially, a prompt (e.g. \"a photo of a\") is crafted to extract text features of all intra-task classes by presenting the class-extended prompt (\"a photo of a [CLS]\") to the text encoder. Subsequently, the image encoder is utilized to derive the image feature of a given input example, and class prediction is performed by comparing cosine distances between the image feature and text features of classes." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 417, + 295, + 560 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 417, + 295, + 560 + ], + "spans": [ + { + "bbox": [ + 55, + 417, + 295, + 560 + ], + "type": "text", + "content": "Prompt Tuning with an Image-Text Matching Loss. PT aims to learn a task-specific prompt with few-labeled samples from the base task. Concretely, a prompt is formulated as " + }, + { + "bbox": [ + 55, + 417, + 295, + 560 + ], + "type": "inline_equation", + "content": "\\kappa" + }, + { + "bbox": [ + 55, + 417, + 295, + 560 + ], + "type": "text", + "content": " context vectors: " + }, + { + "bbox": [ + 55, + 417, + 295, + 560 + ], + "type": "inline_equation", + "content": "\\{v_{1}, v_{2}, \\dots, v_{\\kappa}\\}" + }, + { + "bbox": [ + 55, + 417, + 295, + 560 + ], + "type": "text", + "content": ". During training, we produce the text feature of the " + }, + { + "bbox": [ + 55, + 417, + 295, + 560 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 55, + 417, + 295, + 560 + ], + "type": "text", + "content": "-th class by inputting the class-token-extended prompt " + }, + { + "bbox": [ + 55, + 417, + 295, + 560 + ], + "type": "inline_equation", + "content": "c_{j} = \\{v_{1}, v_{2}, \\dots, v_{\\kappa}, [\\mathrm{CLS}]\\}" + }, + { + "bbox": [ + 55, + 417, + 295, + 560 + ], + "type": "text", + "content": " to the text encoder " + }, + { + "bbox": [ + 55, + 417, + 295, + 560 + ], + "type": "inline_equation", + "content": "g(\\cdot)" + }, + { + "bbox": [ + 55, + 417, + 295, + 560 + ], + "type": "text", + "content": ", where [CLS] indicates the class token of the " + }, + { + "bbox": [ + 55, + 417, + 295, + 560 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 55, + 417, + 295, + 560 + ], + "type": "text", + "content": "-th class. Denote " + }, + { + "bbox": [ + 55, + 417, + 295, + 560 + ], + "type": "inline_equation", + "content": "\\pmb{f}_{i} \\in \\mathbb{R}^{d}" + }, + { + "bbox": [ + 55, + 417, + 295, + 560 + ], + "type": "text", + "content": " as the image feature of a sample " + }, + { + "bbox": [ + 55, + 417, + 295, + 560 + ], + "type": "inline_equation", + "content": "\\pmb{x}_{i}" + }, + { + "bbox": [ + 55, + 417, + 295, + 560 + ], + "type": "text", + "content": " extracted by the image encoder, the task-specific prompt can be updated via back-propagating the image-text matching (ITM) loss through the frozen CLIP model. The ITM loss can be expressed as" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 113, + 565, + 295, + 592 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 565, + 295, + 592 + ], + "spans": [ + { + "bbox": [ + 113, + 565, + 295, + 592 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {I T M}} = - \\sum_ {j} \\boldsymbol {y} _ {j} \\log \\mathrm {p} \\left(\\boldsymbol {c} _ {j} \\mid \\boldsymbol {x} _ {i}\\right), \\tag {1}", + "image_path": "15509d2a2d9989184f313240f121e62fc455f54641684b51d1a36a1f1b088a51.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 594, + 192, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 192, + 606 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 192, + 606 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 594, + 192, + 606 + ], + "type": "inline_equation", + "content": "\\pmb{y}" + }, + { + "bbox": [ + 55, + 594, + 192, + 606 + ], + "type": "text", + "content": " is the one-hot label, and" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 609, + 295, + 640 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 106, + 609, + 295, + 640 + ], + "spans": [ + { + "bbox": [ + 106, + 609, + 295, + 640 + ], + "type": "interline_equation", + "content": "\\mathrm {p} \\left(\\boldsymbol {c} _ {j} \\mid \\boldsymbol {x} _ {i}\\right) = \\frac {e ^ {< g \\left(\\boldsymbol {c} _ {j}\\right) , \\boldsymbol {f} _ {i} > / \\tau}}{\\sum_ {t = 1} ^ {M} e ^ {< g \\left(\\boldsymbol {c} _ {t}\\right) , \\boldsymbol {f} _ {i} > / \\tau}}, \\tag {2}", + "image_path": "cf8c8cc79fc839fccf892da30a65d5fc63801fc9ad3634dd1bdb36de1418a950.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 643, + 295, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 643, + 295, + 667 + ], + "spans": [ + { + "bbox": [ + 55, + 643, + 295, + 667 + ], + "type": "inline_equation", + "content": "< \\cdot >" + }, + { + "bbox": [ + 55, + 643, + 295, + 667 + ], + "type": "text", + "content": " denotes cosine similarity, " + }, + { + "bbox": [ + 55, + 643, + 295, + 667 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 643, + 295, + 667 + ], + "type": "text", + "content": " is the number of classes, and " + }, + { + "bbox": [ + 55, + 643, + 295, + 667 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 55, + 643, + 295, + 667 + ], + "type": "text", + "content": " is the temperature learned by CLIP." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 672, + 207, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 672, + 207, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 672, + 207, + 685 + ], + "type": "text", + "content": "2.2. Rethinking the FT Baseline" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": "While many PT schemes have reported improved performance and efficiency over the full fine-tuning (FT) baseline," + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 285, + 553, + 309 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 285, + 553, + 309 + ], + "spans": [ + { + "bbox": [ + 313, + 285, + 553, + 309 + ], + "type": "text", + "content": "the discrepancy of the implementation details among those schemes obscures the actual performance enhancement." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 310, + 554, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 310, + 554, + 548 + ], + "spans": [ + { + "bbox": [ + 313, + 310, + 554, + 548 + ], + "type": "text", + "content": "Comparison of PT and FT. To comprehensively evaluate the progress established by PT, we perform a comparison between the representative PT method CoOp [42] with the full FT baseline in terms of i) the number of learnable parameters, ii) memory usage, iii) time cost, and iv) baseto-new generalization results. The experimental setting is shown in Sup. Mat. (A). From the obtained results in Figure 2 (a), we observe that although PT significantly reduces the number of learnable parameters compared with FT (using 1/51200 parameters of FT), the improvements in memory efficiency and time efficiency are relatively insignificant (reducing only " + }, + { + "bbox": [ + 313, + 310, + 554, + 548 + ], + "type": "inline_equation", + "content": "6.3\\%" + }, + { + "bbox": [ + 313, + 310, + 554, + 548 + ], + "type": "text", + "content": " memory usage and " + }, + { + "bbox": [ + 313, + 310, + 554, + 548 + ], + "type": "inline_equation", + "content": "15.8\\%" + }, + { + "bbox": [ + 313, + 310, + 554, + 548 + ], + "type": "text", + "content": " time cost). Notably, in comparison to FT, the classification accuracy of PT decreases by " + }, + { + "bbox": [ + 313, + 310, + 554, + 548 + ], + "type": "inline_equation", + "content": "3.49\\%" + }, + { + "bbox": [ + 313, + 310, + 554, + 548 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 310, + 554, + 548 + ], + "type": "inline_equation", + "content": "4.49\\%" + }, + { + "bbox": [ + 313, + 310, + 554, + 548 + ], + "type": "text", + "content": " on base and new tasks, respectively. This means that pursuing higher parameter efficiency by freezing the vast majority of VLM weights during PT neither enhances the transferability of pre-trained knowledge nor improves memory and time efficiency significantly. In other words, existing PT methods involve a trade-off between parameter efficiency and performance." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 549, + 555, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 549, + 555, + 693 + ], + "spans": [ + { + "bbox": [ + 313, + 549, + 555, + 693 + ], + "type": "text", + "content": "Cost Analysis of FT. The computational cost (i.e., memory and time cost) for fine-tuning VLMs is mainly caused by Feature-Gradient Propagation Flows (FGPFs)—the forward propagation flows of image/text features and the backward propagation flows of gradients. For example, on the CLIP model, each propagation flow is required to traverse all the " + }, + { + "bbox": [ + 313, + 549, + 555, + 693 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 549, + 555, + 693 + ], + "type": "text", + "content": " layers of the vision encoder " + }, + { + "bbox": [ + 313, + 549, + 555, + 693 + ], + "type": "inline_equation", + "content": "E_{V}" + }, + { + "bbox": [ + 313, + 549, + 555, + 693 + ], + "type": "text", + "content": " and the text encoder " + }, + { + "bbox": [ + 313, + 549, + 555, + 693 + ], + "type": "inline_equation", + "content": "E_{T}" + }, + { + "bbox": [ + 313, + 549, + 555, + 693 + ], + "type": "text", + "content": ". For simplicity, we assume that " + }, + { + "bbox": [ + 313, + 549, + 555, + 693 + ], + "type": "inline_equation", + "content": "E_{V}" + }, + { + "bbox": [ + 313, + 549, + 555, + 693 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 549, + 555, + 693 + ], + "type": "inline_equation", + "content": "E_{T}" + }, + { + "bbox": [ + 313, + 549, + 555, + 693 + ], + "type": "text", + "content": " have identical network structures (e.g. ViT-B/16). Denote the number of class tokens as " + }, + { + "bbox": [ + 313, + 549, + 555, + 693 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 549, + 555, + 693 + ], + "type": "text", + "content": ", and the costs of the FGPFs in the " + }, + { + "bbox": [ + 313, + 549, + 555, + 693 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 313, + 549, + 555, + 693 + ], + "type": "text", + "content": "-th layer of " + }, + { + "bbox": [ + 313, + 549, + 555, + 693 + ], + "type": "inline_equation", + "content": "E_{V}" + }, + { + "bbox": [ + 313, + 549, + 555, + 693 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 549, + 555, + 693 + ], + "type": "inline_equation", + "content": "E_{T}" + }, + { + "bbox": [ + 313, + 549, + 555, + 693 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 313, + 549, + 555, + 693 + ], + "type": "inline_equation", + "content": "C_{V}" + }, + { + "bbox": [ + 313, + 549, + 555, + 693 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 549, + 555, + 693 + ], + "type": "inline_equation", + "content": "C_{T}" + }, + { + "bbox": [ + 313, + 549, + 555, + 693 + ], + "type": "text", + "content": " respectively. Thus, the total cost " + }, + { + "bbox": [ + 313, + 549, + 555, + 693 + ], + "type": "inline_equation", + "content": "C_{\\mathrm{total}} = N \\times (C_{V} + C_{T} \\times M^{1})" + }, + { + "bbox": [ + 313, + 549, + 555, + 693 + ], + "type": "text", + "content": ". In this work," + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 325, + 702, + 553, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 702, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 325, + 702, + 553, + 712 + ], + "type": "text", + "content": "For each training image, we need to obtain the text features of all " + }, + { + "bbox": [ + 325, + 702, + 553, + 712 + ], + "type": "inline_equation", + "content": "M" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14725" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "text", + "content": "we refer to the number of network layers and the number of class tokens used for each training image as the length and width of FGPFs, respectively. In particular, " + }, + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "inline_equation", + "content": "\\text{length} = N" + }, + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "inline_equation", + "content": "\\text{width} = M" + }, + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "text", + "content": " for existing PT methods, as shown in Figure 3. We therefore raise the following question:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 65, + 137, + 285, + 162 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 137, + 285, + 162 + ], + "spans": [ + { + "bbox": [ + 65, + 137, + 285, + 162 + ], + "type": "text", + "content": "Can we reduce both the length and width of FGPFs without compromising the FT performance?" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 171, + 296, + 230 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 171, + 296, + 230 + ], + "spans": [ + { + "bbox": [ + 55, + 171, + 296, + 230 + ], + "type": "text", + "content": "Towards Effective and Efficient Knowledge Transfer. To answer the above question, we design two metrics of Feature Sensitivity (FS) and Gradient Dependence (GD) to estimate the contributions of different network layers and class tokens for each training image " + }, + { + "bbox": [ + 55, + 171, + 296, + 230 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 55, + 171, + 296, + 230 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 231, + 296, + 278 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 231, + 296, + 278 + ], + "spans": [ + { + "bbox": [ + 55, + 231, + 296, + 278 + ], + "type": "text", + "content": "More specifically, let " + }, + { + "bbox": [ + 55, + 231, + 296, + 278 + ], + "type": "inline_equation", + "content": "\\pmb{f}_l" + }, + { + "bbox": [ + 55, + 231, + 296, + 278 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 231, + 296, + 278 + ], + "type": "inline_equation", + "content": "\\pmb{f}_l'" + }, + { + "bbox": [ + 55, + 231, + 296, + 278 + ], + "type": "text", + "content": " be a vision (or a text) feature before and after fine-tuning the " + }, + { + "bbox": [ + 55, + 231, + 296, + 278 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 55, + 231, + 296, + 278 + ], + "type": "text", + "content": "-th layer of the CLIP's vision (or text) encoder, the FS of the " + }, + { + "bbox": [ + 55, + 231, + 296, + 278 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 55, + 231, + 296, + 278 + ], + "type": "text", + "content": "-th layer for the training image " + }, + { + "bbox": [ + 55, + 231, + 296, + 278 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 55, + 231, + 296, + 278 + ], + "type": "text", + "content": " can be expressed as" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 131, + 288, + 295, + 301 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 288, + 295, + 301 + ], + "spans": [ + { + "bbox": [ + 131, + 288, + 295, + 301 + ], + "type": "interline_equation", + "content": "\\mathbf {F} \\mathbf {S} _ {l} (\\boldsymbol {x}) = \\psi \\left(\\boldsymbol {f} _ {l}, \\boldsymbol {f} _ {l} ^ {\\prime}\\right), \\tag {3}", + "image_path": "94d67df2fea43fce04a7ee4d9658e0f345da96a781fbf1701efc0ebf95343b2d.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 311, + 296, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 311, + 296, + 396 + ], + "spans": [ + { + "bbox": [ + 55, + 311, + 296, + 396 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 311, + 296, + 396 + ], + "type": "inline_equation", + "content": "\\psi (\\cdot)" + }, + { + "bbox": [ + 55, + 311, + 296, + 396 + ], + "type": "text", + "content": " is the Euclidean distance. Hence, the bigger the " + }, + { + "bbox": [ + 55, + 311, + 296, + 396 + ], + "type": "inline_equation", + "content": "\\mathbf{FS}_l(\\pmb {x})" + }, + { + "bbox": [ + 55, + 311, + 296, + 396 + ], + "type": "text", + "content": " value, the more important the " + }, + { + "bbox": [ + 55, + 311, + 296, + 396 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 55, + 311, + 296, + 396 + ], + "type": "text", + "content": "-th layer for the input image " + }, + { + "bbox": [ + 55, + 311, + 296, + 396 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 55, + 311, + 296, + 396 + ], + "type": "text", + "content": ". Let " + }, + { + "bbox": [ + 55, + 311, + 296, + 396 + ], + "type": "inline_equation", + "content": "\\nabla f_{l}" + }, + { + "bbox": [ + 55, + 311, + 296, + 396 + ], + "type": "text", + "content": " indicate the computed feature gradients passing through the last layer of the vision encoder for the training image " + }, + { + "bbox": [ + 55, + 311, + 296, + 396 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 55, + 311, + 296, + 396 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 311, + 296, + 396 + ], + "type": "inline_equation", + "content": "\\nabla f_l^{(c)}" + }, + { + "bbox": [ + 55, + 311, + 296, + 396 + ], + "type": "text", + "content": " be the feature gradients after the " + }, + { + "bbox": [ + 55, + 311, + 296, + 396 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 311, + 296, + 396 + ], + "type": "text", + "content": "-th class token is removed from Eq. 2. The GD of the " + }, + { + "bbox": [ + 55, + 311, + 296, + 396 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 311, + 296, + 396 + ], + "type": "text", + "content": "-th class for " + }, + { + "bbox": [ + 55, + 311, + 296, + 396 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 55, + 311, + 296, + 396 + ], + "type": "text", + "content": " can be computed as" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 116, + 406, + 294, + 421 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 406, + 294, + 421 + ], + "spans": [ + { + "bbox": [ + 116, + 406, + 294, + 421 + ], + "type": "interline_equation", + "content": "\\mathbf {G D} _ {c} (\\boldsymbol {x}) = \\psi (\\nabla \\boldsymbol {f} _ {l}, \\nabla \\boldsymbol {f} _ {l} ^ {(c)}). \\tag {4}", + "image_path": "7cd88c9392fd5509d05dd6dde6b3bfa7a33e934b875862cd6ede9be2ebe64e47.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 430, + 295, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 430, + 295, + 453 + ], + "spans": [ + { + "bbox": [ + 55, + 430, + 295, + 453 + ], + "type": "text", + "content": "Therefore, the bigger the " + }, + { + "bbox": [ + 55, + 430, + 295, + 453 + ], + "type": "inline_equation", + "content": "\\mathbf{GD}_c(x)" + }, + { + "bbox": [ + 55, + 430, + 295, + 453 + ], + "type": "text", + "content": " value, the more important the " + }, + { + "bbox": [ + 55, + 430, + 295, + 453 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 430, + 295, + 453 + ], + "type": "text", + "content": "-th class for the input image " + }, + { + "bbox": [ + 55, + 430, + 295, + 453 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 430, + 295, + 453 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 454, + 296, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 454, + 296, + 622 + ], + "spans": [ + { + "bbox": [ + 55, + 454, + 296, + 622 + ], + "type": "text", + "content": "Figure 2 (b) illustrates the obtained FS values of different image/text encoder layers when adapting CLIP to the two datasets Caltech101 and Food101. As can be observed, the FS scores of the vast majority of shallow layers are close to 0, and only the last few layers contribute significantly to adapting CLIP to the two datasets. Moreover, Figure 2 (c) presents the frequency distributions of the GD values of class tokens for 100 training images randomly sampled from the two datasets. As shown, most class tokens contribute minimally to capturing task-specific knowledge from each training image. In a nutshell, the above observations reveal that we can achieve effective and efficient FT by filtering out unimportant network layers and class tokens in the text encoder for every training image." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 629, + 135, + 642 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 135, + 642 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 135, + 642 + ], + "type": "text", + "content": "2.3. Skip Tuning" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 647, + 295, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 647, + 295, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 647, + 295, + 696 + ], + "type": "text", + "content": "Motivated by the previous section, we propose Skip Tuning, which performs Layer-wise Skipping (LSkip) and Classwise Skipping (CSkip) to adapt VLMs to downstream tasks in an effective and efficient manner, as shown in Figure 3." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "text", + "content": "Layer-wise Skipping (LSkip). LSkip aims to reduce the length of FGPFs without compromising the performance of the FT baseline. To this end, LSkip first saves the image and text features before the " + }, + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "text", + "content": "-th shallow layers of " + }, + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "inline_equation", + "content": "E_{V}" + }, + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "inline_equation", + "content": "E_{T}" + }, + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "text", + "content": " in a cache, and then uses the saved intermediate features as input to update the parameters of the remaining " + }, + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "inline_equation", + "content": "N - \\omega" + }, + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "text", + "content": " deep layers. Denote " + }, + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "inline_equation", + "content": "\\pmb{v}_{i}^{\\omega}" + }, + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "inline_equation", + "content": "\\pmb{t}_{j}^{\\omega}" + }, + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "text", + "content": " as the extracted image and text features of the image " + }, + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "inline_equation", + "content": "\\pmb{x}_i" + }, + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "text", + "content": " and the class token CLS (" + }, + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "inline_equation", + "content": "j = 1, \\dots, M" + }, + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "text", + "content": ") in the " + }, + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "text", + "content": "-th layers of " + }, + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "inline_equation", + "content": "E_{V}" + }, + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "inline_equation", + "content": "E_{T}" + }, + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "text", + "content": ", respectively, i.e.," + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 347, + 191, + 553, + 206 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 191, + 553, + 206 + ], + "spans": [ + { + "bbox": [ + 347, + 191, + 553, + 206 + ], + "type": "interline_equation", + "content": "\\boldsymbol {v} _ {i} ^ {\\omega} = E _ {V} [ 1: \\omega ] (\\boldsymbol {x} _ {i}), \\boldsymbol {t} _ {j} ^ {\\omega} = E _ {T} [ 1: \\omega ] (\\boldsymbol {c} _ {j}), \\tag {5}", + "image_path": "af955409ed4a2898fe88ec16e742967fa1ba2491c03149a8a8db2d348d30788f.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 210, + 553, + 233 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 210, + 553, + 233 + ], + "spans": [ + { + "bbox": [ + 313, + 210, + 553, + 233 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 210, + 553, + 233 + ], + "type": "inline_equation", + "content": "c_{j}" + }, + { + "bbox": [ + 313, + 210, + 553, + 233 + ], + "type": "text", + "content": " is constructed using a hand-crafted prompt, e.g., " + }, + { + "bbox": [ + 313, + 210, + 553, + 233 + ], + "type": "inline_equation", + "content": "c_{j} = a" + }, + { + "bbox": [ + 313, + 210, + 553, + 233 + ], + "type": "text", + "content": " photo of a [CLS]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 234, + 553, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 234, + 553, + 270 + ], + "spans": [ + { + "bbox": [ + 313, + 234, + 553, + 270 + ], + "type": "text", + "content": "After that, the " + }, + { + "bbox": [ + 313, + 234, + 553, + 270 + ], + "type": "inline_equation", + "content": "[1:\\omega]" + }, + { + "bbox": [ + 313, + 234, + 553, + 270 + ], + "type": "text", + "content": " layers of both " + }, + { + "bbox": [ + 313, + 234, + 553, + 270 + ], + "type": "inline_equation", + "content": "E_{V}" + }, + { + "bbox": [ + 313, + 234, + 553, + 270 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 234, + 553, + 270 + ], + "type": "inline_equation", + "content": "E_{T}" + }, + { + "bbox": [ + 313, + 234, + 553, + 270 + ], + "type": "text", + "content": " are discarded. During FT, we input " + }, + { + "bbox": [ + 313, + 234, + 553, + 270 + ], + "type": "inline_equation", + "content": "\\pmb{v}_i^\\omega" + }, + { + "bbox": [ + 313, + 234, + 553, + 270 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 234, + 553, + 270 + ], + "type": "inline_equation", + "content": "\\pmb{t}_j^\\omega" + }, + { + "bbox": [ + 313, + 234, + 553, + 270 + ], + "type": "text", + "content": " to the remaining " + }, + { + "bbox": [ + 313, + 234, + 553, + 270 + ], + "type": "inline_equation", + "content": "N - \\omega" + }, + { + "bbox": [ + 313, + 234, + 553, + 270 + ], + "type": "text", + "content": " deep layers of " + }, + { + "bbox": [ + 313, + 234, + 553, + 270 + ], + "type": "inline_equation", + "content": "E_{V}" + }, + { + "bbox": [ + 313, + 234, + 553, + 270 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 234, + 553, + 270 + ], + "type": "inline_equation", + "content": "E_{T}" + }, + { + "bbox": [ + 313, + 234, + 553, + 270 + ], + "type": "text", + "content": ", and obtain:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 318, + 278, + 553, + 297 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 278, + 553, + 297 + ], + "spans": [ + { + "bbox": [ + 318, + 278, + 553, + 297 + ], + "type": "interline_equation", + "content": "\\tilde {\\boldsymbol {v}} _ {i} ^ {N} = E _ {V} [ \\omega + 1: N ] (\\boldsymbol {v} _ {i} ^ {\\omega}), \\tilde {\\boldsymbol {t}} _ {j} ^ {N} = E _ {T} [ \\omega + 1: N ] (\\boldsymbol {t} _ {j} ^ {\\omega}), \\tag {6}", + "image_path": "b8e05e9bfb3a735037e5d4c511ff61f480153bf5770b20745ffbe5a406e38097.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 304, + 553, + 351 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 304, + 553, + 351 + ], + "spans": [ + { + "bbox": [ + 313, + 304, + 553, + 351 + ], + "type": "text", + "content": "which are used to calculate the loss " + }, + { + "bbox": [ + 313, + 304, + 553, + 351 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{ITM}}" + }, + { + "bbox": [ + 313, + 304, + 553, + 351 + ], + "type": "text", + "content": " in Eq. 1. In this way, the image and text features, along with the calculated gradients, propagate through only " + }, + { + "bbox": [ + 313, + 304, + 553, + 351 + ], + "type": "inline_equation", + "content": "N - \\omega" + }, + { + "bbox": [ + 313, + 304, + 553, + 351 + ], + "type": "text", + "content": " layers, effectively reducing the length of FGPFs." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 352, + 554, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 352, + 554, + 495 + ], + "spans": [ + { + "bbox": [ + 313, + 352, + 554, + 495 + ], + "type": "text", + "content": "Class-wise Skipping (CSkip). Given a training image, the goal of CSkip is to filter out unimportant class tokens in the text encoder " + }, + { + "bbox": [ + 313, + 352, + 554, + 495 + ], + "type": "inline_equation", + "content": "E_{T}" + }, + { + "bbox": [ + 313, + 352, + 554, + 495 + ], + "type": "text", + "content": " during the construction of the image-text matching loss " + }, + { + "bbox": [ + 313, + 352, + 554, + 495 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{ITM}}" + }, + { + "bbox": [ + 313, + 352, + 554, + 495 + ], + "type": "text", + "content": " (Eq. 1). Intuitively, the direct way is to select the top " + }, + { + "bbox": [ + 313, + 352, + 554, + 495 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 352, + 554, + 495 + ], + "type": "text", + "content": " closest class tokens to compute the loss for each image, based on the similarities between the image feature and the text features of the " + }, + { + "bbox": [ + 313, + 352, + 554, + 495 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 352, + 554, + 495 + ], + "type": "text", + "content": " inner-task classes. Nevertheless, we empirically find that this strategy leads to overfitting by selecting the same subset of classes for each image across different training epochs. Therefore, we propose an exponential image-conditioned class filtering strategy to overcome this limitation." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 495, + 554, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 495, + 554, + 579 + ], + "spans": [ + { + "bbox": [ + 313, + 495, + 554, + 579 + ], + "type": "text", + "content": "Concretely, before FT begins, we use the text encoder " + }, + { + "bbox": [ + 313, + 495, + 554, + 579 + ], + "type": "inline_equation", + "content": "E_{T}" + }, + { + "bbox": [ + 313, + 495, + 554, + 579 + ], + "type": "text", + "content": " to produce " + }, + { + "bbox": [ + 313, + 495, + 554, + 579 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 495, + 554, + 579 + ], + "type": "text", + "content": " class features with a hand-crafted prompt, e.g. \"a photo of a [CLS]\". Then, we sort the cosine similarities between the vision feature of a training image and the " + }, + { + "bbox": [ + 313, + 495, + 554, + 579 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 495, + 554, + 579 + ], + "type": "text", + "content": " class features in descending order. The probability of sampling the " + }, + { + "bbox": [ + 313, + 495, + 554, + 579 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 495, + 554, + 579 + ], + "type": "text", + "content": "-th class token for the current training image can be expressed as:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 348, + 588, + 553, + 620 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 348, + 588, + 553, + 620 + ], + "spans": [ + { + "bbox": [ + 348, + 588, + 553, + 620 + ], + "type": "interline_equation", + "content": "p _ {j} = \\left\\{ \\begin{array}{l l} 1 & o _ {j} \\leq r \\times M, \\\\ \\zeta \\left(o _ {j} - r \\times M, \\lambda\\right) & o _ {j} > r \\times M, \\end{array} \\right. \\tag {7}", + "image_path": "7255b8ae395c38d04927f879581d93873d4cfa4c21254f631118dded7ed1511f.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 629, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 555, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 629, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\zeta (\\iota ,\\lambda) = e^{-\\lambda \\iota}" + }, + { + "bbox": [ + 313, + 629, + 555, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 629, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\lambda >0" + }, + { + "bbox": [ + 313, + 629, + 555, + 713 + ], + "type": "text", + "content": " is the exponential decay coefficient, " + }, + { + "bbox": [ + 313, + 629, + 555, + 713 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 313, + 629, + 555, + 713 + ], + "type": "text", + "content": " is the sampling ratio, and " + }, + { + "bbox": [ + 313, + 629, + 555, + 713 + ], + "type": "inline_equation", + "content": "o_{j}" + }, + { + "bbox": [ + 313, + 629, + 555, + 713 + ], + "type": "text", + "content": " is the sorting index of the " + }, + { + "bbox": [ + 313, + 629, + 555, + 713 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 629, + 555, + 713 + ], + "type": "text", + "content": "-th class. In this way, for each image, we can maintain the top " + }, + { + "bbox": [ + 313, + 629, + 555, + 713 + ], + "type": "inline_equation", + "content": "r\\times M" + }, + { + "bbox": [ + 313, + 629, + 555, + 713 + ], + "type": "text", + "content": " closest class tokens while also sampling the remaining " + }, + { + "bbox": [ + 313, + 629, + 555, + 713 + ], + "type": "inline_equation", + "content": "(1 - r)\\times M" + }, + { + "bbox": [ + 313, + 629, + 555, + 713 + ], + "type": "text", + "content": " classes with a certain probability for constructing " + }, + { + "bbox": [ + 313, + 629, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{ITM}}" + }, + { + "bbox": [ + 313, + 629, + 555, + 713 + ], + "type": "text", + "content": ", improving the diversity of training data at different epochs. Denote " + }, + { + "bbox": [ + 313, + 629, + 555, + 713 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 629, + 555, + 713 + ], + "type": "text", + "content": " as the number of" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 55, + 702, + 293, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 702, + 293, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 702, + 293, + 713 + ], + "type": "text", + "content": "classes to calculate the image-text matching loss, as illustrated in Eq. (2)." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14726" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 180 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 180 + ], + "type": "text", + "content": "class tokens sampled for the " + }, + { + "bbox": [ + 55, + 72, + 294, + 180 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 72, + 294, + 180 + ], + "type": "text", + "content": "-th training image, we have " + }, + { + "bbox": [ + 55, + 72, + 294, + 180 + ], + "type": "inline_equation", + "content": "m \\ll M" + }, + { + "bbox": [ + 55, + 72, + 294, + 180 + ], + "type": "text", + "content": ". This means we can flexibly reduce the width of FGPFs in the text encoder for every training image. Surprisingly, our experimental results in the next section reveal that ignoring the majority of class tokens for each training image can enhance the generalization performance of the learned model. This improvement may stem from CSkip's capability to filter out redundant and distracting text features when performing image-text matching with the loss of " + }, + { + "bbox": [ + 55, + 72, + 294, + 180 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{ITM}}" + }, + { + "bbox": [ + 55, + 72, + 294, + 180 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 193, + 136, + 205 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 193, + 136, + 205 + ], + "spans": [ + { + "bbox": [ + 55, + 193, + 136, + 205 + ], + "type": "text", + "content": "3. Experiments" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 213, + 171, + 226 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 213, + 171, + 226 + ], + "spans": [ + { + "bbox": [ + 55, + 213, + 171, + 226 + ], + "type": "text", + "content": "3.1. Experimental Setup" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 232, + 295, + 351 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 232, + 295, + 351 + ], + "spans": [ + { + "bbox": [ + 55, + 232, + 295, + 351 + ], + "type": "text", + "content": "Datasets. We conduct experiments using 11 datasets from diverse sources. In particular, for base-to-new generalization, cross-dataset transfer and few-shot learning, we use 11 datasets including ImageNet [4], Caltech101 [5], OxfordPets [23], StanfordCars [19], Flowers102 [22], Food101 [2], FGVCAircraft [21], SUN397 [31], UCF101 [28], DTD [3] and EuroSAT [12]. For domain generalization, we use ImageNet [4] as the source dataset and its four variants as target datasets including ImageNetV2 [26], ImageNet-Sketch [30], ImageNet-A [14], and ImageNet-R [13]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 352, + 295, + 423 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 352, + 295, + 423 + ], + "spans": [ + { + "bbox": [ + 55, + 352, + 295, + 423 + ], + "type": "text", + "content": "Evaluation Metric. We report base-task accuracy (\\%, denoted as Base), new-task accuracy (\\%, denoted as New), and their harmonic mean (\\%, denoted as " + }, + { + "bbox": [ + 55, + 352, + 295, + 423 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 55, + 352, + 295, + 423 + ], + "type": "text", + "content": ") to compare the performance/effectiveness of different methods. We also report the time cost (seconds, denoted as Time) and memory usage (M, denoted as Memory) for efficiency evaluation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 425, + 295, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 425, + 295, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 425, + 295, + 605 + ], + "type": "text", + "content": "Implementation details. Our implementation of Skip Tuning is based on the open-source Github repository of DePT " + }, + { + "bbox": [ + 55, + 425, + 295, + 605 + ], + "type": "inline_equation", + "content": "[39]^2" + }, + { + "bbox": [ + 55, + 425, + 295, + 605 + ], + "type": "text", + "content": ". In concrete terms, we leverage pre-trained ViT-B/16 as the backbone of the CLIP model. We employ the SGD optimizer to train the model with a learning rate of 2e-5 and a batch size of 4. By default, the number of training epochs is set to 20 for the base-to-new generalization, cross-dataset generalization, and domain generalization benchmarks, and 40 for the few-shot learning benchmark. The above hyperparameters are fixed across all datasets. We adjust the LSkip hyper-parameter " + }, + { + "bbox": [ + 55, + 425, + 295, + 605 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 55, + 425, + 295, + 605 + ], + "type": "text", + "content": ", CSkip hyper-parameters " + }, + { + "bbox": [ + 55, + 425, + 295, + 605 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 55, + 425, + 295, + 605 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 425, + 295, + 605 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 425, + 295, + 605 + ], + "type": "text", + "content": " in § 3.3. All experimental results are the average of 3 runs with different seeds. We conduct experiments using an NVIDIA V100 GPU. For more details and additional results, please refer to Sup. Mat." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 614, + 179, + 626 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 614, + 179, + 626 + ], + "spans": [ + { + "bbox": [ + 55, + 614, + 179, + 626 + ], + "type": "text", + "content": "3.2. Experimental Results" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 632, + 295, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 632, + 295, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 632, + 295, + 693 + ], + "type": "text", + "content": "Base-to-New Generalization. The base-to-new generalization setting evaluates whether the models learned on a base task can generalize to new tasks with unseen classes. Following the comparison methods, for each dataset, we first construct a base task and a new task by equally dividing the" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 553, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 300 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 300 + ], + "type": "text", + "content": "dataset into two sets of classes, then we perform prompt tuning on the base task and test the learned model on both the base and new tasks. Table 1 presents the obtained results of Skip Tuning and other state-of-the-art prompt tuning methods on 11 datasets. As shown, Skip Tuning achieves the best base-to-new generalization performance with the lowest memory usage and time cost. Concretely, Skip Tuning achieves " + }, + { + "bbox": [ + 313, + 72, + 553, + 300 + ], + "type": "inline_equation", + "content": "\\times 7.2" + }, + { + "bbox": [ + 313, + 72, + 553, + 300 + ], + "type": "text", + "content": " time efficiency, and " + }, + { + "bbox": [ + 313, + 72, + 553, + 300 + ], + "type": "inline_equation", + "content": "\\times 5.1" + }, + { + "bbox": [ + 313, + 72, + 553, + 300 + ], + "type": "text", + "content": " memory efficiency, while maintaining a " + }, + { + "bbox": [ + 313, + 72, + 553, + 300 + ], + "type": "inline_equation", + "content": "1.14\\%" + }, + { + "bbox": [ + 313, + 72, + 553, + 300 + ], + "type": "text", + "content": " improvement in H ACC compared to the previous state-of-the-art method PromptSRC. This demonstrates the effectiveness and efficiency of our proposed Skip Tuning method. Moreover, we observe a tradeoff between base-task and new-task accuracies for most competitors. For instance, CoPrompt outperforms DePT on new tasks but lags behind DePT on base tasks. Notably, our Skip Tuning achieves the best performance on both base and new tasks simultaneously. This suggests that Skip Tuning effectively mitigates the overfitting issue when transferring VLMs to the base (or target) task." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 303, + 553, + 530 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 303, + 553, + 530 + ], + "spans": [ + { + "bbox": [ + 313, + 303, + 553, + 530 + ], + "type": "text", + "content": "Cross-Dataset Generalization. The cross-dataset generalization setting assesses whether models trained on a source dataset/distribution can generalize to unseen target datasets/distributions. We follow the common setup of those comparison methods to use ImageNet as the source dataset and the other 10 datasets as target datasets. Table 2 presents the obtained cross-dataset generalization results of our Skip Tuning and other state-of-the-art methods on the source and target datasets. As seen, compared to the previous state-of-the-art method PromptSRC, our Skip Tuning achieves " + }, + { + "bbox": [ + 313, + 303, + 553, + 530 + ], + "type": "inline_equation", + "content": "\\times 44" + }, + { + "bbox": [ + 313, + 303, + 553, + 530 + ], + "type": "text", + "content": " time efficiency, and " + }, + { + "bbox": [ + 313, + 303, + 553, + 530 + ], + "type": "inline_equation", + "content": "\\times 21.5" + }, + { + "bbox": [ + 313, + 303, + 553, + 530 + ], + "type": "text", + "content": " memory efficiency, while establishing " + }, + { + "bbox": [ + 313, + 303, + 553, + 530 + ], + "type": "inline_equation", + "content": "1.5\\%" + }, + { + "bbox": [ + 313, + 303, + 553, + 530 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 303, + 553, + 530 + ], + "type": "inline_equation", + "content": "1.19\\%" + }, + { + "bbox": [ + 313, + 303, + 553, + 530 + ], + "type": "text", + "content": " ACC improvements on the source and target distributions respectively. From the average results, Skip Tuning consistently shows superior effectiveness and efficiency over the nine competitors on the 10 target datasets, without compromising the results of the tuned model on the source dataset. This demonstrates the effectiveness of our Skip Tuning scheme for improving the robustness of the tuned model to distribution shifts." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 534, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 553, + 713 + ], + "type": "text", + "content": "Domain Generalization. The domain generalization setting assesses whether models trained on a source domain can generalize to unseen/target domains. In line with those comparison methods, we consider the ImageNet as the source domain and the other four ImageNet variants as target domains. As shown in Table 2, compared to the previous state-of-the-art method PromptSRC, Skip Tuning achieves " + }, + { + "bbox": [ + 313, + 534, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\times 44" + }, + { + "bbox": [ + 313, + 534, + 553, + 713 + ], + "type": "text", + "content": " time efficiency, and " + }, + { + "bbox": [ + 313, + 534, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\times 21.5" + }, + { + "bbox": [ + 313, + 534, + 553, + 713 + ], + "type": "text", + "content": " memory efficiency, while establishing " + }, + { + "bbox": [ + 313, + 534, + 553, + 713 + ], + "type": "inline_equation", + "content": "1.5\\%" + }, + { + "bbox": [ + 313, + 534, + 553, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 534, + 553, + 713 + ], + "type": "inline_equation", + "content": "0.55\\%" + }, + { + "bbox": [ + 313, + 534, + 553, + 713 + ], + "type": "text", + "content": " ACC improvements on the source and target domains, respectively. Besides, Skip Tuning consistently outperforms those competitors on the four target domains without compromising the performance of the tuned model on the source domain, which proves the effectiveness of our Skip Tuning scheme in improving the robustness of the tuned model to domain shifts." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 66, + 702, + 216, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 702, + 216, + 712 + ], + "spans": [ + { + "bbox": [ + 66, + 702, + 216, + 712 + ], + "type": "text", + "content": "2https://github.com/Koorye/DePT" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14727" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 57, + 76, + 553, + 399 + ], + "blocks": [ + { + "bbox": [ + 132, + 66, + 478, + 76 + ], + "lines": [ + { + "bbox": [ + 132, + 66, + 478, + 76 + ], + "spans": [ + { + "bbox": [ + 132, + 66, + 478, + 76 + ], + "type": "text", + "content": "Table 1. Base-to-new generalization results over 11 datasets. * indicates our reproduced results." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 76, + 553, + 399 + ], + "lines": [ + { + "bbox": [ + 57, + 76, + 553, + 399 + ], + "spans": [ + { + "bbox": [ + 57, + 76, + 553, + 399 + ], + "type": "table", + "html": "
DatasetsMetricCoOp(IJCV'22)CoCoOp(CVPR'22)ProGrad(ICCV'23)KgCoOp(CVPR'23)MaPLe(CVPR'23)ProSRC(ICCV'23)TCP(CVPR'24)DePT(CVPR'24)CoPro.*ICLR'24)SkipT.(Ours)
ImgNetBase76.4775.9877.0275.8376.6677.6077.2777.0376.5377.73
New67.8870.4366.6669.9670.5470.7369.8770.1371.3070.40
H71.9273.1071.4672.7873.4774.0173.3873.4273.8273.89
CaltechBase98.0097.9698.0297.7297.7498.1098.2398.3098.6098.50
New89.3193.8193.8994.3994.3694.0394.6794.6095.1795.33
H93.7395.8495.9196.0396.0296.0296.4296.4196.8596.89
PetsBase93.6795.2095.0794.6595.4395.3394.6794.3394.7395.70
New95.2997.6997.6397.7697.7697.3097.2097.2396.7097.87
H94.4796.4396.3396.1896.5896.3095.9295.7695.7196.77
CarsBase78.1270.4977.6871.7672.9478.2780.8079.1373.1782.93
New60.4073.5968.6375.0474.0074.9774.1375.4770.6372.50
H68.1372.0172.8873.3673.4776.5877.3277.2671.8877.37
FlowersBase97.6094.8795.5495.0095.9298.0797.7398.0096.9398.57
New59.6771.7571.8774.7372.4676.5075.5776.3775.5075.80
H74.0681.7182.0383.6582.5685.9585.2385.8484.8885.70
Food101Base88.3390.7090.3790.5090.7190.6790.5790.5090.3790.67
New82.2691.2989.5991.7092.0591.5391.3791.6091.5392.03
H85.1990.9989.9891.0991.3891.1090.9791.0590.9591.34
AircraftBase40.4433.4140.5436.2137.4442.7341.9743.2036.1745.37
New22.3023.7127.5733.5535.6137.8734.4334.8334.4737.13
H28.7527.7432.8234.8336.5040.1537.8338.5735.3040.84
SUN397Base80.6079.7481.2680.2980.8282.6782.6382.3382.3082.40
New65.8976.8674.1776.5378.7078.4778.2077.8079.6379.03
H72.5178.5777.5578.3679.7580.5280.3580.0080.9480.68
DTDBase79.4477.0177.3577.5580.3683.3782.7782.2083.0083.77
New41.1856.0052.3554.9959.1862.9758.0759.1363.2067.23
H54.2464.8562.4564.3568.1671.1568.2568.7871.7674.59
EuroSATBase92.1987.4990.1185.6494.0792.9091.6389.0393.7792.47
New54.7460.0460.8964.3473.2373.9074.7371.0771.7383.00
H68.6971.2172.6773.4882.3582.3282.3279.0481.2887.48
UCF101Base84.6982.3384.3382.8983.0087.1087.1385.8086.2087.30
New56.0573.4574.9476.6778.6678.8080.7777.2378.7082.47
H67.4677.6479.3579.6580.7782.7483.8381.2982.2884.81
Avg ACCBase82.6980.4782.4880.7382.2884.2684.1383.6282.8985.04
New63.2271.6970.7573.6075.1476.1075.3675.0475.3277.53
H71.6675.8376.1677.0078.5579.9779.5179.1078.9381.11
CostTime (s)11862851231111914131735619733684239
Mem. (M)320415563204318817292041318917142060404
", + "image_path": "41c499b8023c36744ef9570e52218b5e1f0a89b7594f313d7d2b41e15be303a5.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 57, + 427, + 553, + 483 + ], + "blocks": [ + { + "bbox": [ + 56, + 405, + 553, + 426 + ], + "lines": [ + { + "bbox": [ + 56, + 405, + 553, + 426 + ], + "spans": [ + { + "bbox": [ + 56, + 405, + 553, + 426 + ], + "type": "text", + "content": "Table 2. Cross-dataset generalization results on 11 datasets. * indicates our reproduced results. The detailed results on the 10 target datasets (i.e., Caltech, Pets, Cars, ..., and UCF101) are reported in Sup. Mat. (B)." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 427, + 553, + 483 + ], + "lines": [ + { + "bbox": [ + 57, + 427, + 553, + 483 + ], + "spans": [ + { + "bbox": [ + 57, + 427, + 553, + 483 + ], + "type": "table", + "html": "
DatasetsMetricCoOp(IJCV'22)CoCoOp(CVPR'22)ProGrad(ICCV'23)KgCoOp(CVPR'23)MaPLe(CVPR'23)ProSRC(ICCV'23)TCP(CVPR'24)DePT(CVPR'24)CoPro.*ICLR'24)SkipT.(Ours)
ImageNetACC71.5171.0272.2470.6670.7271.2771.4072.7772.5372.77
10 DatasetsAvg ACC63.8865.7462.7165.5166.3065.8166.2965.5564.8167.00
CostTime (s)316329391756223316364942500911617422796191611139
Mem. (M)154121362215412152541378614107152631378314131656
", + "image_path": "e8655f0321216acb3873982f63c8b0e98d42e8a3cd22617412759849928662fe.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 58, + 513, + 553, + 576 + ], + "blocks": [ + { + "bbox": [ + 56, + 490, + 553, + 511 + ], + "lines": [ + { + "bbox": [ + 56, + 490, + 553, + 511 + ], + "spans": [ + { + "bbox": [ + 56, + 490, + 553, + 511 + ], + "type": "text", + "content": "Table 3. Domain generalization results on ImageNet. * indicates our reproduced results. The detailed results on the 4 ImgNet variants (i.e., ImgNet-V2, ImgNet-S, ImgNet-A, and ImgNet-R) are presented in Sup. Mat. (B)." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 513, + 553, + 576 + ], + "lines": [ + { + "bbox": [ + 58, + 513, + 553, + 576 + ], + "spans": [ + { + "bbox": [ + 58, + 513, + 553, + 576 + ], + "type": "table", + "html": "
DatasetsMetricCoOp(IJCV'22)CoCoOp(CVPR'22)ProGrad(ICCV'23)KgCoOp(CVPR'23)MaPLe(CVPR'23)ProSRC(ICCV'23)DePT(CVPR'24)CoPro.*ICLR'24)SkipT.(Ours)
ImageNetACC71.5171.0272.2470.6670.7271.2772.7772.5372.77
4 Imag.VariantsAvg ACC59.2859.9059.0760.1160.2660.6558.9760.0261.20
CostTime (s)3163293917562233163649425009122796191611139
Mem. (M)1541213622154121525413786141071378314131656
", + "image_path": "3cf2cf37e06b8528bd83733040c050f941078fe294cde284827637bd93f982ba.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 58, + 597, + 155, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 597, + 155, + 609 + ], + "spans": [ + { + "bbox": [ + 58, + 597, + 155, + 609 + ], + "type": "text", + "content": "3.3. Ablation Studies" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 58, + 616, + 294, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 616, + 294, + 640 + ], + "spans": [ + { + "bbox": [ + 58, + 616, + 294, + 640 + ], + "type": "text", + "content": "In this section, we conduct ablative studies to further scrutinize our devised Skip Tuning method." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 58, + 642, + 294, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 642, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 58, + 642, + 294, + 713 + ], + "type": "text", + "content": "Effectiveness of the Designed Components. Our proposed Skip Tuning approach simultaneously performs Layer-wise Skipping (LSkip) and Class-wise Skipping (CSkip) to enhance the memory and time efficiency of the full fine-tuning (FT) baseline. In this experiment, we investigate the effectiveness of LSkip and CSkip by gradually adding them to" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 316, + 597, + 553, + 705 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 597, + 553, + 705 + ], + "spans": [ + { + "bbox": [ + 316, + 597, + 553, + 705 + ], + "type": "text", + "content": "the FT baseline. From the results in Table 4, we have the following observations. i) Both LSkip and CSkip contribute to performance improvement. ii) By combining all those components, Skip Tuning improves the effectiveness and efficiency of the baseline method remarkably. iii) The memory and time efficiency gains are more pronounced when CSkip is applied to the FT baseline, compared to the performance improvements it yields. iv) CSkip improves the memory and time efficiency of the baseline without com" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "14728" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 80, + 68, + 231, + 148 + ], + "blocks": [ + { + "bbox": [ + 80, + 68, + 231, + 148 + ], + "lines": [ + { + "bbox": [ + 80, + 68, + 231, + 148 + ], + "spans": [ + { + "bbox": [ + 80, + 68, + 231, + 148 + ], + "type": "image", + "image_path": "a56986e33d9254718aeb5c25f4f9e6eb82161a84cc6b1c0e1be5bbfc2cca713b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 71, + 151, + 537, + 163 + ], + "lines": [ + { + "bbox": [ + 71, + 151, + 537, + 163 + ], + "spans": [ + { + "bbox": [ + 71, + 151, + 537, + 163 + ], + "type": "text", + "content": "Figure 4. Ablation studies of the number of skipped layers " + }, + { + "bbox": [ + 71, + 151, + 537, + 163 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 71, + 151, + 537, + 163 + ], + "type": "text", + "content": " in LSkip, and the sampling rate " + }, + { + "bbox": [ + 71, + 151, + 537, + 163 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 71, + 151, + 537, + 163 + ], + "type": "text", + "content": ", the decay coefficient " + }, + { + "bbox": [ + 71, + 151, + 537, + 163 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 71, + 151, + 537, + 163 + ], + "type": "text", + "content": " in CSkip." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 232, + 68, + 380, + 148 + ], + "blocks": [ + { + "bbox": [ + 232, + 68, + 380, + 148 + ], + "lines": [ + { + "bbox": [ + 232, + 68, + 380, + 148 + ], + "spans": [ + { + "bbox": [ + 232, + 68, + 380, + 148 + ], + "type": "image", + "image_path": "2022ca518eb0cedc67eb1a75e69d5435497eec485a4e47994078f5a29d956172.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 382, + 68, + 530, + 148 + ], + "blocks": [ + { + "bbox": [ + 382, + 68, + 530, + 148 + ], + "lines": [ + { + "bbox": [ + 382, + 68, + 530, + 148 + ], + "spans": [ + { + "bbox": [ + 382, + 68, + 530, + 148 + ], + "type": "image", + "image_path": "71906f81d5a329d75f4129a50523326eba13a5f2dea8909a2dd299b98ac95c39.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 57, + 188, + 294, + 243 + ], + "blocks": [ + { + "bbox": [ + 67, + 176, + 282, + 187 + ], + "lines": [ + { + "bbox": [ + 67, + 176, + 282, + 187 + ], + "spans": [ + { + "bbox": [ + 67, + 176, + 282, + 187 + ], + "type": "text", + "content": "Table 4. Ablation study on the components of Skip Tuning." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 188, + 294, + 243 + ], + "lines": [ + { + "bbox": [ + 57, + 188, + 294, + 243 + ], + "spans": [ + { + "bbox": [ + 57, + 188, + 294, + 243 + ], + "type": "table", + "html": "
LSPCSPBaseNewHTime (s)Mem. (M)
XX84.9974.2679.2710021846
✓X85.0877.3081.00388861
X✓84.7774.4879.29440919
✓✓85.0477.5381.11239404
", + "image_path": "d94b00cc24022f0ca67dec70cb223ea97fc5aba37e0f56c576508e988c85a756.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 56, + 288, + 299, + 351 + ], + "blocks": [ + { + "bbox": [ + 55, + 254, + 295, + 286 + ], + "lines": [ + { + "bbox": [ + 55, + 254, + 295, + 286 + ], + "spans": [ + { + "bbox": [ + 55, + 254, + 295, + 286 + ], + "type": "text", + "content": "Table 5. Comparison of our Skip Tuning and adapter-based methods. " + }, + { + "bbox": [ + 55, + 254, + 295, + 286 + ], + "type": "inline_equation", + "content": "{}^{ \\dagger }" + }, + { + "bbox": [ + 55, + 254, + 295, + 286 + ], + "type": "text", + "content": " denotes the efficiency-optimized versions (details are illustrated in Sup. Mat. (C))." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 288, + 299, + 351 + ], + "lines": [ + { + "bbox": [ + 56, + 288, + 299, + 351 + ], + "spans": [ + { + "bbox": [ + 56, + 288, + 299, + 351 + ], + "type": "table", + "html": "
MethodBaseNewHTime (s)Mem. (M)
FT84.9974.2679.2710021846
CLIP-adapter [10]74.4873.8174.148881784
LoRA [15]80.5374.7377.529101580
SkipT. (Ours)85.0477.5381.11239404
CLIP-adapter†76.7573.5675.12137175
SkipT. (Ours)†82.6675.5678.9511567
", + "image_path": "b1ff6424a7120754521a237f67b46006407b29750468d8ccc2ca9ba985dec3ee.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 376, + 295, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 376, + 295, + 424 + ], + "spans": [ + { + "bbox": [ + 54, + 376, + 295, + 424 + ], + "type": "text", + "content": "promising the generalization performance of the learned model. One possible reason is that CSkip can filter out redundant and distracting text features when performing image-text matching with the loss of " + }, + { + "bbox": [ + 54, + 376, + 295, + 424 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{ITM}}" + }, + { + "bbox": [ + 54, + 376, + 295, + 424 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 54, + 425, + 295, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 425, + 295, + 556 + ], + "spans": [ + { + "bbox": [ + 54, + 425, + 295, + 556 + ], + "type": "text", + "content": "Impact of the LSkip Hyper-parameter " + }, + { + "bbox": [ + 54, + 425, + 295, + 556 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 54, + 425, + 295, + 556 + ], + "type": "text", + "content": ". Our Skip Tuning approach drops out the " + }, + { + "bbox": [ + 54, + 425, + 295, + 556 + ], + "type": "inline_equation", + "content": "1 \\sim \\omega" + }, + { + "bbox": [ + 54, + 425, + 295, + 556 + ], + "type": "text", + "content": " layers of the CLIP's vision and text encoders in the LSkip step. Here, we investigate the impact of " + }, + { + "bbox": [ + 54, + 425, + 295, + 556 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 54, + 425, + 295, + 556 + ], + "type": "text", + "content": " on performance by setting " + }, + { + "bbox": [ + 54, + 425, + 295, + 556 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 54, + 425, + 295, + 556 + ], + "type": "text", + "content": " to the values of " + }, + { + "bbox": [ + 54, + 425, + 295, + 556 + ], + "type": "inline_equation", + "content": "\\{2, 4, 6, 8, 10\\}" + }, + { + "bbox": [ + 54, + 425, + 295, + 556 + ], + "type": "text", + "content": ". The obtained average testing results on the 11 datasets are reported in Figure 4 (Left). As shown, the obtained H ACCs gradually increase as the " + }, + { + "bbox": [ + 54, + 425, + 295, + 556 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 54, + 425, + 295, + 556 + ], + "type": "text", + "content": " value decreases from 10 to 6, after which the performance gradually decreases. But, we also see that as the value of " + }, + { + "bbox": [ + 54, + 425, + 295, + 556 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 54, + 425, + 295, + 556 + ], + "type": "text", + "content": " becomes smaller, both memory and time costs increase. Hence, we set " + }, + { + "bbox": [ + 54, + 425, + 295, + 556 + ], + "type": "inline_equation", + "content": "\\omega = 6" + }, + { + "bbox": [ + 54, + 425, + 295, + 556 + ], + "type": "text", + "content": " for LSKip in this work." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 557, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 557, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 557, + 295, + 714 + ], + "type": "text", + "content": "Impact of the CSkip Hyper-parameter " + }, + { + "bbox": [ + 55, + 557, + 295, + 714 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 55, + 557, + 295, + 714 + ], + "type": "text", + "content": ". In the CSkip step of Skip Tuning, we devise an image-conditioned class sampling strategy to filter out irrelevant class tokens for each training image. The larger the sampling ratio " + }, + { + "bbox": [ + 55, + 557, + 295, + 714 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 55, + 557, + 295, + 714 + ], + "type": "text", + "content": " value, the more class tokens will be used to construct the training loss " + }, + { + "bbox": [ + 55, + 557, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{ITM}}" + }, + { + "bbox": [ + 55, + 557, + 295, + 714 + ], + "type": "text", + "content": " for each image. It is necessary to scrutinize the impact of " + }, + { + "bbox": [ + 55, + 557, + 295, + 714 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 55, + 557, + 295, + 714 + ], + "type": "text", + "content": " on performance. To this end, we respectively set " + }, + { + "bbox": [ + 55, + 557, + 295, + 714 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 55, + 557, + 295, + 714 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 557, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\{0.1, 0.3, 0.5, 0.7, 0.9, 1.0\\}" + }, + { + "bbox": [ + 55, + 557, + 295, + 714 + ], + "type": "text", + "content": ", and report the average testing results on the 11 datasets in Figure 4 (Mid.). From the obtained H ACCs, our method is in general not sensitive to the change of " + }, + { + "bbox": [ + 55, + 557, + 295, + 714 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 55, + 557, + 295, + 714 + ], + "type": "text", + "content": " within a certain range (from 0.4 to 1.0). We also see a gradual decrease in H ACC when " + }, + { + "bbox": [ + 55, + 557, + 295, + 714 + ], + "type": "inline_equation", + "content": "r > 0.7" + }, + { + "bbox": [ + 55, + 557, + 295, + 714 + ], + "type": "text", + "content": ", suggesting that for each training image, not all class tokens are" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 183, + 553, + 231 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 183, + 553, + 231 + ], + "spans": [ + { + "bbox": [ + 313, + 183, + 553, + 231 + ], + "type": "text", + "content": "beneficial for capturing task-specific knowledge. Besides, we can observe that as the value of " + }, + { + "bbox": [ + 313, + 183, + 553, + 231 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 313, + 183, + 553, + 231 + ], + "type": "text", + "content": " becomes larger, both memory and time costs increase. Therefore, we set " + }, + { + "bbox": [ + 313, + 183, + 553, + 231 + ], + "type": "inline_equation", + "content": "r = 0.5" + }, + { + "bbox": [ + 313, + 183, + 553, + 231 + ], + "type": "text", + "content": " for CSKip in this work." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 232, + 555, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 232, + 555, + 376 + ], + "spans": [ + { + "bbox": [ + 313, + 232, + 555, + 376 + ], + "type": "text", + "content": "Impact of the CSkip Hyper-parameter " + }, + { + "bbox": [ + 313, + 232, + 555, + 376 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 232, + 555, + 376 + ], + "type": "text", + "content": ". Our Skip Tuning introduces an exponential decay coefficient " + }, + { + "bbox": [ + 313, + 232, + 555, + 376 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 232, + 555, + 376 + ], + "type": "text", + "content": " for the devised image-conditioned class sampling strategy in the CSkip step. We study the impact of " + }, + { + "bbox": [ + 313, + 232, + 555, + 376 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 232, + 555, + 376 + ], + "type": "text", + "content": " by setting " + }, + { + "bbox": [ + 313, + 232, + 555, + 376 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 232, + 555, + 376 + ], + "type": "text", + "content": " to the values of " + }, + { + "bbox": [ + 313, + 232, + 555, + 376 + ], + "type": "inline_equation", + "content": "\\{0.01, 0.1, 0.3, 0.5, 0.7, 0.9, 1.0\\}" + }, + { + "bbox": [ + 313, + 232, + 555, + 376 + ], + "type": "text", + "content": ", and report the average testing results over the 11 datasets in Figure 4 (Right). As can be observed from the obtained H ACCs, our method is in general not sensitive to " + }, + { + "bbox": [ + 313, + 232, + 555, + 376 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 232, + 555, + 376 + ], + "type": "text", + "content": " when it takes the values from 0.1 to 0.7. Particularly, the H ACCs gradually decrease as " + }, + { + "bbox": [ + 313, + 232, + 555, + 376 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 232, + 555, + 376 + ], + "type": "text", + "content": " increases from 0.7 to 1.0. Also, we can see that as the value of " + }, + { + "bbox": [ + 313, + 232, + 555, + 376 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 232, + 555, + 376 + ], + "type": "text", + "content": " becomes larger, both memory and time costs increase. Thus, we set " + }, + { + "bbox": [ + 313, + 232, + 555, + 376 + ], + "type": "inline_equation", + "content": "\\lambda = 0.3" + }, + { + "bbox": [ + 313, + 232, + 555, + 376 + ], + "type": "text", + "content": " for CSkip in this work." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 384, + 424, + 396 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 384, + 424, + 396 + ], + "spans": [ + { + "bbox": [ + 313, + 384, + 424, + 396 + ], + "type": "text", + "content": "3.4. Additional Results" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 402, + 555, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 402, + 555, + 581 + ], + "spans": [ + { + "bbox": [ + 313, + 402, + 555, + 581 + ], + "type": "text", + "content": "Comparison with Adapter-based Methods. Previous experimental results demonstrate the superior effectiveness and efficiency of our Skip Tuning method over existing prompt tuning methods. To further demonstrate the advantages of Skip Tuning, we also compare it with the representative adapter-based methods LoRA [15] and CLIP-adapter [10]. The base-to-new generalization results averaged on 11 datasets are reported in Table 5, where " + }, + { + "bbox": [ + 313, + 402, + 555, + 581 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP - adapter}^{\\dagger}" + }, + { + "bbox": [ + 313, + 402, + 555, + 581 + ], + "type": "text", + "content": " refers to our re-implementation of CLIP-adapter, i.e., the CLIP's last-layer features are cached once before training and subsequently used to update the weights of the adapter module. As observed, our Skip Tuning offers substantial memory and time efficiency advantages over the two strong competitors, while also delivering better classification performance on both base and new tasks." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "text", + "content": "Few-shot Learning. In the previous experiments, we follow the comparison approaches to evaluate the performance of different methods on " + }, + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "text", + "content": "-way 16-shot tasks—16 training examples are sampled for each of the " + }, + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "text", + "content": " inner-task classes. It is interesting to further scrutinize the effectiveness and efficiency of our Skip Tuning method under different shots. To this end, Figure 5 reports the few-shot learning performance of Skip Tuning and other representative competitors. As shown, our Skip Tuning demonstrates superior performance compared to other methods, with significant efficiency advantages in memory usage and time cost across" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14729" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 75, + 71, + 170, + 148 + ], + "blocks": [ + { + "bbox": [ + 75, + 71, + 170, + 148 + ], + "lines": [ + { + "bbox": [ + 75, + 71, + 170, + 148 + ], + "spans": [ + { + "bbox": [ + 75, + 71, + 170, + 148 + ], + "type": "image", + "image_path": "637e120cc9d3b46297e22b25e0f6500bdd59376877ecbf88b530126884d8f322.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 173, + 71, + 261, + 148 + ], + "blocks": [ + { + "bbox": [ + 173, + 71, + 261, + 148 + ], + "lines": [ + { + "bbox": [ + 173, + 71, + 261, + 148 + ], + "spans": [ + { + "bbox": [ + 173, + 71, + 261, + 148 + ], + "type": "image", + "image_path": "e9bfe42567c0d1563635c9c73f2a1bd7473003c2bfadd548124749a40be6a8a3.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 263, + 72, + 351, + 148 + ], + "blocks": [ + { + "bbox": [ + 263, + 72, + 351, + 148 + ], + "lines": [ + { + "bbox": [ + 263, + 72, + 351, + 148 + ], + "spans": [ + { + "bbox": [ + 263, + 72, + 351, + 148 + ], + "type": "image", + "image_path": "67d756aca6943f05d057564ad0418be1cffe38783dc51765f56f909a56ed7eac.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 354, + 72, + 441, + 148 + ], + "blocks": [ + { + "bbox": [ + 354, + 72, + 441, + 148 + ], + "lines": [ + { + "bbox": [ + 354, + 72, + 441, + 148 + ], + "spans": [ + { + "bbox": [ + 354, + 72, + 441, + 148 + ], + "type": "image", + "image_path": "8b81a9afd059df47794d2a9d1e4795e805587176820b4212d422702cc4336e3e.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 443, + 71, + 533, + 194 + ], + "blocks": [ + { + "bbox": [ + 443, + 71, + 533, + 194 + ], + "lines": [ + { + "bbox": [ + 443, + 71, + 533, + 194 + ], + "spans": [ + { + "bbox": [ + 443, + 71, + 533, + 194 + ], + "type": "image", + "image_path": "07d8763f9a22b67de9073924347146d5951929ecd42ef56ed440df79c7d63dbd.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 75, + 150, + 170, + 227 + ], + "blocks": [ + { + "bbox": [ + 75, + 150, + 170, + 227 + ], + "lines": [ + { + "bbox": [ + 75, + 150, + 170, + 227 + ], + "spans": [ + { + "bbox": [ + 75, + 150, + 170, + 227 + ], + "type": "image", + "image_path": "502e1bbe316477ec48714516e583d344c54d50599226b040f28e3f6d99205e92.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 173, + 150, + 260, + 227 + ], + "blocks": [ + { + "bbox": [ + 173, + 150, + 260, + 227 + ], + "lines": [ + { + "bbox": [ + 173, + 150, + 260, + 227 + ], + "spans": [ + { + "bbox": [ + 173, + 150, + 260, + 227 + ], + "type": "image", + "image_path": "c2425b93941ba109e22b08ee13c3098cb0b2d01bca86c3a23ae8ea4b0bd5e0f1.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 263, + 150, + 351, + 227 + ], + "blocks": [ + { + "bbox": [ + 263, + 150, + 351, + 227 + ], + "lines": [ + { + "bbox": [ + 263, + 150, + 351, + 227 + ], + "spans": [ + { + "bbox": [ + 263, + 150, + 351, + 227 + ], + "type": "image", + "image_path": "d242e9023e80ef8bc6abb142d8fd40b0f759bd0aa6646d21cafe1e7f3395b714.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 354, + 150, + 441, + 227 + ], + "blocks": [ + { + "bbox": [ + 354, + 150, + 441, + 227 + ], + "lines": [ + { + "bbox": [ + 354, + 150, + 441, + 227 + ], + "spans": [ + { + "bbox": [ + 354, + 150, + 441, + 227 + ], + "type": "image", + "image_path": "0d7a1a4ed196428b468f7216e5f216a82f9f116f8e6a1a28bc100e64ff07cf6b.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 443, + 198, + 533, + 313 + ], + "blocks": [ + { + "bbox": [ + 443, + 198, + 533, + 313 + ], + "lines": [ + { + "bbox": [ + 443, + 198, + 533, + 313 + ], + "spans": [ + { + "bbox": [ + 443, + 198, + 533, + 313 + ], + "type": "image", + "image_path": "944be23fbc852114182e304c10950aaa0b722ea51a99ca2a1e710348a60681ff.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 75, + 230, + 170, + 312 + ], + "blocks": [ + { + "bbox": [ + 75, + 230, + 170, + 312 + ], + "lines": [ + { + "bbox": [ + 75, + 230, + 170, + 312 + ], + "spans": [ + { + "bbox": [ + 75, + 230, + 170, + 312 + ], + "type": "image", + "image_path": "0aa2668f617df4861881d515c83de42b0ff64647b4ecaa361fdb0a51d9a9e2ff.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 173, + 230, + 260, + 312 + ], + "blocks": [ + { + "bbox": [ + 173, + 230, + 260, + 312 + ], + "lines": [ + { + "bbox": [ + 173, + 230, + 260, + 312 + ], + "spans": [ + { + "bbox": [ + 173, + 230, + 260, + 312 + ], + "type": "image", + "image_path": "71aa3663652bf038a1204914f6b1313570ca77b602e74ae056887d1ea09232f9.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 212, + 314, + 395, + 325 + ], + "lines": [ + { + "bbox": [ + 212, + 314, + 395, + 325 + ], + "spans": [ + { + "bbox": [ + 212, + 314, + 395, + 325 + ], + "type": "text", + "content": "Figure 5. Few-shot learning results on 11 datasets." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 263, + 231, + 351, + 312 + ], + "blocks": [ + { + "bbox": [ + 263, + 231, + 351, + 312 + ], + "lines": [ + { + "bbox": [ + 263, + 231, + 351, + 312 + ], + "spans": [ + { + "bbox": [ + 263, + 231, + 351, + 312 + ], + "type": "image", + "image_path": "1216882fa9d6ea2cfcaeba90ccbb4dd0a0b44eba3805d19a786992ad858d9f1e.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 354, + 231, + 441, + 312 + ], + "blocks": [ + { + "bbox": [ + 354, + 231, + 441, + 312 + ], + "lines": [ + { + "bbox": [ + 354, + 231, + 441, + 312 + ], + "spans": [ + { + "bbox": [ + 354, + 231, + 441, + 312 + ], + "type": "image", + "image_path": "eb790603167035d504a6ba61adbe58e6d09da8ac941391b2ad7dff8bf71e3e84.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 338, + 296, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 338, + 296, + 399 + ], + "spans": [ + { + "bbox": [ + 55, + 338, + 296, + 399 + ], + "type": "text", + "content": "1-shot to 16-shot settings. Additionally, while PromptSRC and MaPLe enhance CoOp's performance, both methods incur increased time cost. In contrast, Skip Tuning achieves both effectiveness and efficiency without compromise, further underscoring the superiority of our approach." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 55, + 415, + 142, + 426 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 415, + 142, + 426 + ], + "spans": [ + { + "bbox": [ + 55, + 415, + 142, + 426 + ], + "type": "text", + "content": "4. Related Work" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 55, + 436, + 295, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 436, + 295, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 436, + 295, + 628 + ], + "type": "text", + "content": "Pre-trained Vision-Language Models. Pre-trained vision-language models (VLMs) have garnered great attention recently. Notable models such as CLIP[24], ALIGN[16], LiT[36] and FILIP[34] have demonstrated exceptional performance across various vision-language tasks. These models are pre-trained on extensive datasets comprising image-text pairs sourced from the internet. For instance, CLIP[24] and ALIGN[16] respectively utilize over 400 million and 1 billion pairs, for training. The large scale of training data enables these models to excel in open-vocabulary image-text retrieval and zero-shot classification. Additionally, they have been successfully applied to downstream tasks including image classification[10, 40], object detection[6, 8, 11, 35, 43], and semantic segmentation[20, 25]. In this work, we focus on the effective adaptation of vision-language models to downstream visual recognition tasks." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "text", + "content": "Prompt Tuning. Prompt tuning (a.k.a. context optimization [42]) has emerged as a parameter-efficient learning paradigm to adapt powerful VLMs to downstream tasks [7, 9, 37, 38]. As a representative method, CoOp [42] enables task adaptation by optimizing a set of prompt vectors within CLIP's language branch. While effective, CoOp often suffers from limited generalization on new tasks due to" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 338, + 555, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 338, + 555, + 459 + ], + "spans": [ + { + "bbox": [ + 313, + 338, + 555, + 459 + ], + "type": "text", + "content": "overfitting to the base task. A series of schemes are devised to tackle this problem, e.g. CoCoOp[41], KgCoOp[32], ProGrad[44], TCP [33], and DePT[39]. By adding trainable prompts into both the image and text branches of CLIP, MaPLe[17], PromptSRC[18], and CoPrompt[27] demonstrate remarkable performance on both base and new tasks. Despite the advantages, we reveal that freezing the parameters of VLMs during learning the context vectors neither facilitates the generalization of pre-trained knowledge nor significantly improves memory and time efficiency." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 314, + 474, + 388, + 487 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 474, + 388, + 487 + ], + "spans": [ + { + "bbox": [ + 314, + 474, + 388, + 487 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 496, + 555, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 496, + 555, + 616 + ], + "spans": [ + { + "bbox": [ + 313, + 496, + 555, + 616 + ], + "type": "text", + "content": "In this work, we first reveal that freezing the parameters of VLMs during prompt tuning neither facilitates the transferability of pre-trained knowledge nor improves memory and time efficiency considerably. To circumvent this limitation, we propose Skip Tuning, an effective and efficient method for transferring VLMs to downstream tasks without relying on extra context vectors or adapter modules. Extensive experiments across a broad range of benchmarks demonstrate the superiority of our Skip Tuning method over both prompt tuning and adapter-based approaches." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 617, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 556, + 713 + ], + "type": "text", + "content": "Acknowledgements. This study is supported by grants from the National Natural Science Foundation of China (Grant No. 62425208, No. U22A2097, No. 62122018, No. 62020106008, No. U23A20315, No. 82441006, No. 62402094), the Postdoctoral Fellowship Program of CPSF (Grant No. GZB20240625), the Science and Technology Innovation Committee of Shenzhen Municipality Foundation (Grant No.JCYJ20240813114208012)." + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14730" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "text", + "content": "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 146, + 296, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 146, + 296, + 201 + ], + "spans": [ + { + "bbox": [ + 61, + 146, + 296, + 201 + ], + "type": "text", + "content": "[2] Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. Food-101-mining discriminative components with random forests. In Computer vision-ECCV 2014: 13th European conference, zurich, Switzerland, September 6-12, 2014, proceedings, part VI 13, pages 446-461. Springer, 2014. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 202, + 295, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 202, + 295, + 246 + ], + "spans": [ + { + "bbox": [ + 62, + 202, + 295, + 246 + ], + "type": "text", + "content": "[3] Mircea Cimpoi, Subhransu Maji, Iasonas Kokkinos, Sammy Mohamed, and Andrea Vedaldi. Describing textures in the wild. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3606-3613, 2014. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 247, + 295, + 291 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 247, + 295, + 291 + ], + "spans": [ + { + "bbox": [ + 62, + 247, + 295, + 291 + ], + "type": "text", + "content": "[4] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 292, + 294, + 345 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 292, + 294, + 345 + ], + "spans": [ + { + "bbox": [ + 62, + 292, + 294, + 345 + ], + "type": "text", + "content": "[5] Li Fei-Fei, Rob Fergus, and Pietro Perona. Learning generative visual models from few training examples: An incremental bayesian approach tested on 101 object categories. Computer vision and Image understanding, 106(1):59-70, 2007. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 346, + 294, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 346, + 294, + 402 + ], + "spans": [ + { + "bbox": [ + 62, + 346, + 294, + 402 + ], + "type": "text", + "content": "[6] Chengjian Feng, Yujie Zhong, Zequn Jie, Xiangxiang Chu, Haibing Ren, Xiaolin Wei, Weidi Xie, and Lin Ma. Promptdet: Towards open-vocabulary detection using uncurated images. In European Conference on Computer Vision, pages 701-717. Springer, 2022. 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 403, + 294, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 403, + 294, + 446 + ], + "spans": [ + { + "bbox": [ + 62, + 403, + 294, + 446 + ], + "type": "text", + "content": "[7] Yuqian Fu, Yanwei Fu, Jingjing Chen, and Yu-Gang Jiang. Generalized meta-fdmixup: Cross-domain few-shot learning guided by labeled target data. IEEE Transactions on Image Processing, 31:7078-7090, 2022. 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 447, + 294, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 447, + 294, + 501 + ], + "spans": [ + { + "bbox": [ + 62, + 447, + 294, + 501 + ], + "type": "text", + "content": "[8] Yuqian Fu, Yu Xie, Yanwei Fu, and Yu-Gang Jiang. Styleadv: Meta style adversarial training for cross-domain few-shot learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 24575-24584, 2023. 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 502, + 294, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 502, + 294, + 557 + ], + "spans": [ + { + "bbox": [ + 62, + 502, + 294, + 557 + ], + "type": "text", + "content": "[9] Yuqian Fu, Yu Wang, Yixuan Pan, Lian Huai, Xingyu Qiu, Zeyu Shangguan, Tong Liu, Yanwei Fu, Luc Van Gool, and Xingqun Jiang. Cross-domain few-shot object detection via enhanced open-set object detector. In ECCV, pages 247-264. Springer, 2025. 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 558, + 294, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 558, + 294, + 612 + ], + "spans": [ + { + "bbox": [ + 57, + 558, + 294, + 612 + ], + "type": "text", + "content": "[10] Peng Gao, Shijie Geng, Renrui Zhang, Teli Ma, Rongyao Fang, Yongfeng Zhang, Hongsheng Li, and Yu Qiao. Clip-adapter: Better vision-language models with feature adapters. International Journal of Computer Vision, 132(2): 581-595, 2024. 7, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 613, + 294, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 613, + 294, + 656 + ], + "spans": [ + { + "bbox": [ + 57, + 613, + 294, + 656 + ], + "type": "text", + "content": "[11] Xiuye Gu, Tsung-Yi Lin, Weicheng Kuo, and Yin Cui. Open-vocabulary object detection via vision and language knowledge distillation. arXiv preprint arXiv:2104.13921, 2021.8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 658, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 658, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 658, + 294, + 712 + ], + "type": "text", + "content": "[12] Patrick Helber, Benjamin Bischke, Andreas Dengel, and Damian Borth. Eurosat: A novel dataset and deep learning benchmark for land use and land cover classification. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 12(7):2217-2226, 2019. 5" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 139 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 139 + ], + "type": "text", + "content": "[13] Dan Hendrycks, Steven Basart, Norman Mu, Saurav Kadavath, Frank Wang, Evan Dorundo, Rahul Desai, Tyler Zhu, Samyak Parajuli, Mike Guo, et al. The many faces of robustness: A critical analysis of out-of-distribution generalization. In Proceedings of the IEEE/CVF international conference on computer vision, pages 8340-8349, 2021. 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 140, + 553, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 140, + 553, + 185 + ], + "spans": [ + { + "bbox": [ + 317, + 140, + 553, + 185 + ], + "type": "text", + "content": "[14] Dan Hendrycks, Kevin Zhao, Steven Basart, Jacob Steinhardt, and Dawn Song. Natural adversarial examples. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15262-15271, 2021. 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 186, + 553, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 186, + 553, + 229 + ], + "spans": [ + { + "bbox": [ + 317, + 186, + 553, + 229 + ], + "type": "text", + "content": "[15] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 2, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 230, + 553, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 230, + 553, + 295 + ], + "spans": [ + { + "bbox": [ + 317, + 230, + 553, + 295 + ], + "type": "text", + "content": "[16] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International conference on machine learning, pages 4904-4916. PMLR, 2021. 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 297, + 553, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 297, + 553, + 352 + ], + "spans": [ + { + "bbox": [ + 317, + 297, + 553, + 352 + ], + "type": "text", + "content": "[17] Muhammad Uzair Khattak, Hanoona Rasheed, Muhammad Maaz, Salman Khan, and Fahad Shahbaz Khan. Maple: Multi-modal prompt learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19113-19122, 2023. 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 354, + 553, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 354, + 553, + 420 + ], + "spans": [ + { + "bbox": [ + 317, + 354, + 553, + 420 + ], + "type": "text", + "content": "[18] Muhammad Uzair Khattak, Syed Talal Wasim, Muzammal Naseer, Salman Khan, Ming-Hsuan Yang, and Fahad Shahbaz Khan. Self-regulating prompts: Foundational model adaptation without forgetting. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15190-15200, 2023. 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 421, + 553, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 421, + 553, + 465 + ], + "spans": [ + { + "bbox": [ + 317, + 421, + 553, + 465 + ], + "type": "text", + "content": "[19] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In Proceedings of the IEEE international conference on computer vision workshops, pages 554–561, 2013. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 467, + 555, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 467, + 555, + 510 + ], + "spans": [ + { + "bbox": [ + 317, + 467, + 555, + 510 + ], + "type": "text", + "content": "[20] Timo Lüddecke and Alexander Ecker. Image segmentation using text and image prompts. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7086-7096, 2022. 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 512, + 553, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 512, + 553, + 553 + ], + "spans": [ + { + "bbox": [ + 317, + 512, + 553, + 553 + ], + "type": "text", + "content": "[21] Subhransu Maji, Esa Rahtu, Juho Kannala, Matthew Blaschko, and Andrea Vedaldi. Fine-grained visual classification of aircraft. arXiv preprint arXiv:1306.5151, 2013. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 556, + 553, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 556, + 553, + 601 + ], + "spans": [ + { + "bbox": [ + 317, + 556, + 553, + 601 + ], + "type": "text", + "content": "[22] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. In 2008 Sixth Indian conference on computer vision, graphics & image processing, pages 722-729. IEEE, 2008. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 601, + 553, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 601, + 553, + 645 + ], + "spans": [ + { + "bbox": [ + 317, + 601, + 553, + 645 + ], + "type": "text", + "content": "[23] Omkar M Parkhi, Andreae Vedaldi, Andrew Zisserman, and CV Jawahar. Cats and dogs. In 2012 IEEE conference on computer vision and pattern recognition, pages 3498-3505. IEEE, 2012. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 647, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 647, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 317, + 647, + 553, + 712 + ], + "type": "text", + "content": "[24] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 2, 8" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14731" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 138 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 138 + ], + "type": "text", + "content": "[25] Yongming Rao, Wenliang Zhao, Guangyi Chen, Yansong Tang, Zheng Zhu, Guan Huang, Jie Zhou, and Jiwen Lu. Denseclip: Language-guided dense prediction with context-aware prompting. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18082-18091, 2022. 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 140, + 294, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 140, + 294, + 183 + ], + "spans": [ + { + "bbox": [ + 56, + 140, + 294, + 183 + ], + "type": "text", + "content": "[26] Benjamin Recht, Rebecca Roelofs, Ludwig Schmidt, and Vaishaal Shankar. Do imagenet classifiers generalize to imagenet? In International conference on machine learning, pages 5389-5400. PMLR, 2019. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 184, + 294, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 184, + 294, + 216 + ], + "spans": [ + { + "bbox": [ + 56, + 184, + 294, + 216 + ], + "type": "text", + "content": "[27] Shuvendu Roy and Ali Etemad. Consistency-guided prompt learning for vision-language models. arXiv preprint arXiv:2306.01195, 2023. 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 217, + 294, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 217, + 294, + 250 + ], + "spans": [ + { + "bbox": [ + 56, + 217, + 294, + 250 + ], + "type": "text", + "content": "[28] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 251, + 294, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 251, + 294, + 304 + ], + "spans": [ + { + "bbox": [ + 56, + 251, + 294, + 304 + ], + "type": "text", + "content": "[29] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 305, + 294, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 305, + 294, + 348 + ], + "spans": [ + { + "bbox": [ + 56, + 305, + 294, + 348 + ], + "type": "text", + "content": "[30] Haohan Wang, Songwei Ge, Zachary Lipton, and Eric P Xing. Learning robust global representations by penalizing local predictive power. Advances in Neural Information Processing Systems, 32, 2019. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 350, + 294, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 350, + 294, + 404 + ], + "spans": [ + { + "bbox": [ + 56, + 350, + 294, + 404 + ], + "type": "text", + "content": "[31] Jianxiong Xiao, James Hays, Krista A Ehinger, Aude Oliva, and Antonio Torralba. Sun database: Large-scale scene recognition from abbey to zoo. In 2010 IEEE computer society conference on computer vision and pattern recognition, pages 3485-3492. IEEE, 2010. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 405, + 294, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 405, + 294, + 458 + ], + "spans": [ + { + "bbox": [ + 56, + 405, + 294, + 458 + ], + "type": "text", + "content": "[32] Hantao Yao, Rui Zhang, and Changsheng Xu. Visual-language prompt tuning with knowledge-guided context optimization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6757-6767, 2023. 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 460, + 294, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 460, + 294, + 513 + ], + "spans": [ + { + "bbox": [ + 56, + 460, + 294, + 513 + ], + "type": "text", + "content": "[33] Hantao Yao, Rui Zhang, and Changsheng Xu. Tcp: Textual-based class-aware prompt tuning for visual-language model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23438-23448, 2024. 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 514, + 294, + 559 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 514, + 294, + 559 + ], + "spans": [ + { + "bbox": [ + 56, + 514, + 294, + 559 + ], + "type": "text", + "content": "[34] Lewei Yao, Runhui Huang, Lu Hou, Guansong Lu, Minzhe Niu, Hang Xu, Xiaodan Liang, Zhenguo Li, Xin Jiang, and Chunjing Xu. Filip: Fine-grained interactive language-image pre-training. arXiv preprint arXiv:2111.07783, 2021. 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 559, + 294, + 603 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 559, + 294, + 603 + ], + "spans": [ + { + "bbox": [ + 56, + 559, + 294, + 603 + ], + "type": "text", + "content": "[35] Yuhang Zang, Wei Li, Kaiyang Zhou, Chen Huang, and Chen Change Loy. Open-vocabulary detr with conditional matching. In European Conference on Computer Vision, pages 106-122. Springer, 2022. 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 604, + 294, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 604, + 294, + 658 + ], + "spans": [ + { + "bbox": [ + 56, + 604, + 294, + 658 + ], + "type": "text", + "content": "[36] Xiaohua Zhai, Xiao Wang, Basil Mustafa, Andreas Steiner, Daniel Keysers, Alexander Kolesnikov, and Lucas Beyer. Lit: Zero-shot transfer with locked-image text tuning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18123-18133, 2022. 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 658, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 294, + 713 + ], + "type": "text", + "content": "[37] Ji Zhang, Jingkuan Song, Lianli Gao, and Hengtao Shen. Free-lunch for cross-domain few-shot learning: Style-aware episodic training with robust contrastive learning. In Proceedings of the 30th ACM International Conference on Multimedia, pages 2586-2594, 2022. 8" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 408 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "text", + "content": "[38] Ji Zhang, Lianli Gao, Xu Luo, Hengtao Shen, and Jingkuan Song. Deta: Denoised task adaptation for few-shot learning. In ICCV, pages 11541-11551, 2023. 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 107, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 150 + ], + "type": "text", + "content": "[39] Ji Zhang, Shihan Wu, Lianli Gao, Heng Tao Shen, and Jingkuan Song. Dept: Decoupled prompt tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12924-12933, 2024. 5, 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 152, + 553, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 553, + 205 + ], + "type": "text", + "content": "[40] Renrui Zhang, Rongyao Fang, Wei Zhang, Peng Gao, Kunchang Li, Jifeng Dai, Yu Qiao, and Hongsheng Li. Tip-adapter: Training-free clip-adapter for better vision-language modeling. arXiv preprint arXiv:2111.03930, 2021. 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 207, + 553, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 207, + 553, + 261 + ], + "spans": [ + { + "bbox": [ + 316, + 207, + 553, + 261 + ], + "type": "text", + "content": "[41] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Conditional prompt learning for vision-language models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16816-16825, 2022. 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 262, + 553, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 262, + 553, + 306 + ], + "spans": [ + { + "bbox": [ + 316, + 262, + 553, + 306 + ], + "type": "text", + "content": "[42] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision, 130(9):2337-2348, 2022. 2, 3, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 308, + 553, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 308, + 553, + 361 + ], + "spans": [ + { + "bbox": [ + 316, + 308, + 553, + 361 + ], + "type": "text", + "content": "[43] Xingyi Zhou, Rohit Girdhar, Armand Joulin, Philipp Krahenbihl, and Ishan Misra. Detecting twenty-thousand classes using image-level supervision. In European Conference on Computer Vision, pages 350-368. Springer, 2022. 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 364, + 553, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 364, + 553, + 408 + ], + "spans": [ + { + "bbox": [ + 316, + 364, + 553, + 408 + ], + "type": "text", + "content": "[44] Beier Zhu, Yulei Niu, Yucheng Han, Yue Wu, and Hanwang Zhang. Prompt-aligned gradient for prompt tuning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15659-15669, 2023. 8" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14732" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/SkySense-O_ Towards Open-World Remote Sensing Interpretation with Vision-Centric Visual-Language Modeling/2415dd78-f506-4cd3-a12c-42283424590a_content_list.json b/2025/SkySense-O_ Towards Open-World Remote Sensing Interpretation with Vision-Centric Visual-Language Modeling/2415dd78-f506-4cd3-a12c-42283424590a_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..d60954bb01be01249e2ce607b32090ea881fab9f --- /dev/null +++ b/2025/SkySense-O_ Towards Open-World Remote Sensing Interpretation with Vision-Centric Visual-Language Modeling/2415dd78-f506-4cd3-a12c-42283424590a_content_list.json @@ -0,0 +1,1736 @@ +[ + { + "type": "text", + "text": "SkySense-O: Towards Open-World Remote Sensing Interpretation with Vision-Centric Visual-Language Modeling", + "text_level": 1, + "bbox": [ + 163, + 130, + 833, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Qi Zhu $^{1*}$ , Jiangwei Lao $^{2*}$ , Deyi Ji $^{1}$ , Junwei Luo $^{3}$ , Kang Wu $^{3}$ , Yingying Zhang $^{4}$ , Lixiang Ru $^{3}$ , Jian Wang $^{4}$ , Jingdong Chen $^{5}$ , Ming Yang $^{6}$ , Dong Liu $^{1}$ , Feng Zhao $^{1\\dagger}$", + "bbox": [ + 109, + 202, + 883, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ MoE Key Laboratory of Brain-inspired Intelligent Perception and Cognition, USTC $^{2}$ Shanghai Jiaotong University $^{3}$ Wuhan University $^{4}$ UCAS $^{5}$ HUST $^{6}$ Northwestern University", + "bbox": [ + 109, + 239, + 885, + 276 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "zqcrafts@mail.ustc.edu.cn, fzhao956@ustc.edu.cn", + "bbox": [ + 287, + 277, + 705, + 291 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ac29e831ce3eef16d6cb0ba8287501cd003005377d7184f90069654262920e5c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 96, + 310, + 380, + 412 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/cafd840d83c71333f31303dc6d8fa7782c3ffc484fcfec987c8bf39688dfe7ae.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 383, + 309, + 666, + 412 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1f75e40091ff6afa8b62ed975b6f284fb81ceb8b21f8ef3eaa3e48b0ae31a506.jpg", + "image_caption": [ + "Figure 1. SkySense-O offers 4 key advantages from the foundation model to the agent application (Left), leading to state-of-the-art zero-shot performances across 14 downstream datasets over 4 distinct RS interpretation tasks (Right)" + ], + "image_footnote": [], + "bbox": [ + 96, + 415, + 666, + 520 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7f370c36aae4a526b9734964d3e354b83499cf4226d877f3b445ec89d6887498.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 681, + 310, + 895, + 518 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 580, + 326, + 595 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Open-world interpretation aims to accurately localize and recognize all objects within images by vision-language models (VLMs). While substantial progress has been made in this task for natural images, the advancements for remote sensing (RS) images still remain limited, primarily due to these two challenges. 1) Existing RS semantic categories are limited, particularly for pixel-level interpretation datasets. 2) Distinguishing among diverse RS spatial regions solely by language space is challenging due to the dense and intricate spatial distribution in open-world RS imagery. To address the first issue, we develop a fine-grained RS interpretation dataset, Sky-SA, which contains 183,375 high-quality local image-text pairs with full-pixel manual annotations, covering 1,763 category labels, exhibiting richer semantics and higher density than previous datasets. Afterwards, to solve the second issue, we introduce the vision-centric principle for vision-language modeling. Specifically, in the pre-training stage, the visual self-supervised paradigm is incorporated into image-", + "bbox": [ + 88, + 611, + 485, + 876 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "text alignment, reducing the degradation of general visual representation capabilities of existing paradigms. Then, we construct a visual-relevance knowledge graph across open-category texts and further develop a novel vision-centric image-text contrastive loss for fine-tuning with text prompts. This new model, denoted as SkySense-O, demonstrates impressive zero-shot capabilities on a thorough evaluation encompassing 14 datasets over 4 tasks, from recognizing to reasoning and classification to localization. Specifically, it outperforms the latest models such as SegEarthOV, GeoRSCLIP, and VHM by a large margin, i.e., $11.95\\%$ , $8.04\\%$ and $3.55\\%$ on average respectively. The code is publicly available to facilitate further research at https:", + "bbox": [ + 511, + 582, + 906, + 762 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "//github.com/zqcrafts/SkySense-0.", + "bbox": [ + 513, + 763, + 807, + 777 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 799, + 645, + 814 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Remote sensing interpretation (RSI) is a critical and practical task in Earth observation, aiming to identify the semantic types of specific ground objects in remote sensing (RS) images (such as segmentation [1-4], detection [5-10], change detection [11, 12], and visual question answering", + "bbox": [ + 511, + 825, + 906, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution. † Corresponding author.", + "bbox": [ + 107, + 888, + 346, + 901 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "14733", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "(VQA) [13-15]). Early works typically involved annotating RS images with fixed categories and exploring RSI through supervised learning methods. These approaches do not generalize well to unseen new categories in open-world scenarios, thus limiting their practical applicability. Moreover, as RS images becoming more accessible than ever before, there is an increasing demand for open-world interpretations with a broad category scope. Hence, there is a soaring interest in exploring a comprehensive open-world remote sensing interpretation model for many scenarios and tasks [16, 17].", + "bbox": [ + 89, + 90, + 483, + 257 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Challenges. In the domain of natural image scenes, significant progress has been achieved in open-world interpretation tasks through large-scale localization and recognition datasets like AS-1B [18] and COCO-Stuff [19], combined with contrastive language-image pretrained model (CLIP) shown in Fig. 1a. However, the open-world RS interpretation model is still out of reach due to following two reasons: 1) From the dataset perspective, the existing annotations either exhibit sparse mask labels or lack diversity in category descriptions (see Fig.1b). 2) From the modeling perspective, the language model inherently encompasses strong contextual dependence [20], making it challenging to directly use as prompts for decoupling common RS objects and intertwined land-cover categories in open-world.", + "bbox": [ + 89, + 258, + 483, + 473 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Sky-SA Dataset. For Struggle (1), inspired by SAM [21] and ASM [18] in the natural image domain, we propose the Sky-SA dataset with the largest number of category labels for open-world RS interpretation, including 1,763 category texts from 183,375 mask-text pairs. The Sky-SA is mainly characterized by its text openness and fine granularity (Fig. 1b). Specifically, different from general semantic segmentation datasets, the categories of Sky-SA are annotated by short semantic tags in a completely open manner. In other words, the annotators are doing fill-in-the-blank rather than multiple-choice. This enables the model to learn pixel-level open-world interpretation akin to human cognition [18] (Fig. 1c). Furthermore, similar to COCO-Stuff [19], all recognizable pixels of Sky-SA are densely annotated, which further enhances the semantic openness and the contextual relationships from wide semantic categories.", + "bbox": [ + 89, + 474, + 483, + 733 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Vision-Centric Modeling. For Struggle (2), motivated by [22], we introduce the vision-centric principle to correct the potential alignment ambiguity arising from language guidance. Firstly, we initialize the visual encoder from SkySense [23], an advanced RS visual foundation model, and then collect large-scale RS image-caption datasets for pretraining, which mainly include SkyScript [23], RS5M [24] and OSM [25]. Motivated by DetailCLIP [26], to mitigate the degradation of general visual representation capabilities originating from SkySense, we maintain the visual contrastive loss and masked modeling loss during", + "bbox": [ + 89, + 734, + 483, + 901 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/f9735a931aa5e5d37fed58ed90976ebc0a3a5b445a0571e1c9735debbd8be184.jpg", + "image_caption": [ + "Figure 2. The flowchart of the overall SkySense-O development." + ], + "image_footnote": [], + "bbox": [ + 516, + 90, + 895, + 340 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "the pre-training process of image-text alignment. Secondly, for fine-tuning region-level alignment, we leverage GPT-4V [27] to construct a visual-centric knowledge graph across all categories of the Sky-SA dataset, where the node of the graph represents each category, and the edge weight indicates the visual similarity scores between each pair of categories. Using the graph, we derive a vision-centric regularization that focuses on distinguishing categories exhibiting low visual similarity but strong text-level contextual relationships in RS images. Finally, we explore the impacts of Skysense-O as agent on MLLM models (Fig. 1d). In summary, to achieve open-world RS interpretation model, SkySense-O is meticulously developed from dataset to modeling and from pre-training to fine-tuning (Fig. 2), contributions of this paper include:", + "bbox": [ + 511, + 383, + 906, + 611 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We develop Sky-SA dataset with large-scale and high-quality annotated texts for open-world RS interpretation.", + "- The vision-centric principle is first introduced into RS visual-language modeling, primarily instantiated into training paradigm, vision-centric graph, and vision-centric regularization from pre-training to fine-tuning.", + "- We propose SkySense-O, a remarkable open-world RS interpretation model, which achieves SOTA across 14 zeroshot downstream datasets as shown in Fig. 1." + ], + "bbox": [ + 513, + 622, + 903, + 758 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 513, + 775, + 651, + 791 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Remote Sensing Interpretation", + "text_level": 1, + "bbox": [ + 513, + 801, + 785, + 816 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Remote Sensing (RS) interpretation aims to accurately identify and analyze ground objects from RS data, with applications in environmental monitoring, urban planning, and agricultural surveillance [35-42]. Early deep learning approaches focused on closed-set supervised models, in", + "bbox": [ + 511, + 825, + 906, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "14734", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "table", + "img_path": "images/0f23d9fc06c124ce63dad46637c2ca6b415d9552f0e18d11909350f0b077755e.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetSpatial GranularityOpen TextDense Annotation
ImageRegionPixel
Natural Image
SA-1B [21]✓✓✓
COCO-Stuff [19]✓✓✓
Remote Sensing Image
AID [28], RESISC [29]✓
OpenStreetMap [30]✓✓
RS5M [24], SkyScript [31]✓✓
Potsdam, LoveDA [32]✓✓✓
Valiggan, OEM [30]✓✓✓
iSAID [33], SIOR [34]✓✓
FAST [34], SOTA [34]✓✓
Sky-SA (ours)✓✓✓✓✓
", + "bbox": [ + 94, + 88, + 486, + 262 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Table 1. Property comparisons with popular vision-language datasets. The proposed Sky-SA dataset is characterized by its semantic openness, high-density and fine-granularity.", + "bbox": [ + 89, + 272, + 482, + 316 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/c4e55be500c80c107790f36276420462314ecf06a1c1a15117aedbbc08bae462.jpg", + "image_caption": [ + "(a) Florence-2", + "Figure 3. Visual comparisons between the predictions of existing open-world models and the annotations of our Sky-SA dataset." + ], + "image_footnote": [], + "bbox": [ + 96, + 335, + 192, + 409 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/859f1af60096973f92f29262a473a90f95be67defa967f96822ff58d2f93d440.jpg", + "image_caption": [ + "(b) SAM" + ], + "image_footnote": [], + "bbox": [ + 194, + 335, + 289, + 409 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/1ab34c65d913139baedb349e32bdbfd7694b51995461e1f85b8c7a25463ad644.jpg", + "image_caption": [ + "(c) Grounding DINO-1.6" + ], + "image_footnote": [], + "bbox": [ + 290, + 335, + 379, + 409 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/7289831b60b558ca66b17c2b1f130ceffbb8dcbf3ec05239093505fae87b3bac.jpg", + "image_caption": [ + "(d) Sky-SA Dataset" + ], + "image_footnote": [], + "bbox": [ + 379, + 335, + 475, + 409 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "cluding image classification [43, 44], object detection [45-49], change detection [11, 12], and semantic segmentation [28, 50-56]. Recently, CLIP [57] and Vision Language Models (VLMs) have enabled the transition from closed-set to open-set interpretation, allowing models to recognize broader categories [17, 58]. Additionally, models like GPT-4 [59] and LLaVA [60] have facilitated the development of instruction-tuning datasets [61], enhancing the flexibility of RS models in human-machine interactions.", + "bbox": [ + 89, + 474, + 483, + 609 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Vision Foundation Models for Remote Sensing", + "text_level": 1, + "bbox": [ + 89, + 632, + 482, + 648 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Vision foundation models are instrumental in advancing RS interpretation toward open-set capabilities through two primary pre-training paradigms, vision-only self-supervised and vision-language self-supervised [22, 43, 62-67] paradigms. Vision-only self-supervised models utilize large-scale unannotated RS images to develop robust visual representations for various tasks [23, 68-71]. For example, SkySense [23] integrates multiple modalities and geo-contextual information within a contrastive framework, while Masked Image Modeling (MIM) methods like RingMo [72] and Scale-MAE [73] are proposed for further enhancing visual feature learning. Vision-language self-supervised models originate from CLIP. Recently some efforts such as RemoteCLIP [74], RS5M [24], and SkyScript [23] have customized for the RS domain using extensive RS image-caption pairs.", + "bbox": [ + 89, + 659, + 482, + 900 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/95fddfe2a71f7c337796667809fa5e7e997dc4f6d907baee99dde012d71826b3.jpg", + "image_caption": [ + "Figure 4. Scale comparisons with existing RS semantic segmentation datasets. We see that the proposed Sky-SA dataset has a significantly larger category scale and diversity than prior datasets." + ], + "image_footnote": [], + "bbox": [ + 531, + 90, + 916, + 268 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Sky-SA Dataset", + "text_level": 1, + "bbox": [ + 513, + 344, + 669, + 362 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Motivation", + "text_level": 1, + "bbox": [ + 513, + 369, + 635, + 386 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Recently, image-text alignment models based on natural images, such as the open-vocabulary methods [75] and MLLMs like LLaVA [60], have achieved remarkable success. This is primarily due to well-constructed datasets in the natural image domain compared to the RS image domain. There have been numerous high-quality datasets with rich categories in the natural image community, such as COCO-Stuff [19] and ADE [76]. However, as shown in Tab.1 and Fig.4, such resources remain scarce in the RS community, which significantly limits the development of open-world RS interpretation tasks. Moreover, existing open-world models, such as SAM [21], Florence-2 [77], and Grounding-DINO [78], are also unable to achieve dense and fine-grained interpretation on remote sensing images as shown in Fig. 3. To relieve this issue, we design a data engine to help annotators more efficiently and accurately perform full-pixel annotation on unlabeled RS images, leading to the creation of the Sky-SA dataset as following section.", + "bbox": [ + 511, + 393, + 906, + 667 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Data Engine", + "text_level": 1, + "bbox": [ + 513, + 679, + 645, + 695 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Open-Images Collection. In the standard collection process of a closed-set dataset, a category list is predefined, and images that do not contain any items from this list will be filtered out. This manner might lead to dataset bias [79] in remote sensing images. Unlike the close-set methods, we compile a comprehensive collection of images from various available RS datasets. To minimize redundancy in manual annotation, we leverage the SkySense model [23], which possesses image-level visual discrimination capabilities for RS images, to filter visually redundant images.", + "bbox": [ + 511, + 702, + 906, + 853 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Open-Text Fine-Grained Annotation. In the annotation process, drawing inspiration from the popular COCO-Stuff [19] dataset, we prioritize fine-grained semantics and", + "bbox": [ + 511, + 854, + 906, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "14735", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/87c6bc3c040934a2175e3cc876c78ae8a8e02f0334fbffd8dc850ac8ebf71f70.jpg", + "image_caption": [ + "Pre-training (For image-level RS interpretation task)" + ], + "image_footnote": [], + "bbox": [ + 96, + 107, + 416, + 265 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/b684b261ae0159528d91049f1705c708a9f7cb3447dcda8ef131967d9a14a7c6.jpg", + "image_caption": [ + "Supervised Finetuning (For pixel-level RS interpretation task)", + "Figure 5. The framework of SkySense-O, where training strategy includes pre-training and fine-tuning stage for open-world interpretation." + ], + "image_footnote": [], + "bbox": [ + 431, + 107, + 898, + 263 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "full-pixel annotation as core labeling principles. Specifically, there are two key points: 1) the dataset encompasses a rich variety of semantic categories; 2) almost all pixels in each image need to be annotated. The former aids in activating the broad semantic awareness of the vision-language pretrained model, while the latter helps establish comprehensive connections among local semantics throughout each image. Furthermore, to capture a broader semantic range, we employ a short-caption labeling manner for each region, which does not constrain the annotation labels. Additionally, to improve annotation efficiency, we first utilize GPT-4V [27] to generate the image content, and then employ Grounding-DINO [78] and SAM [21] for pre-annotation, allowing annotators to make careful corrections based on these preliminary labels.", + "bbox": [ + 88, + 304, + 483, + 530 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Dataset Statistics. Finally, we develop a high-quality pixel-level remote sensing interpretation dataset, named Sky-SA. As shown in Fig. 4, the Sky-SA dataset consists of 35,000 RS images, mainly including 1,763 category texts from 183,375 segment-text pairs (encompassing approximately 500 distinct categories of land-covering and RS objects after synonym filtering), which far surpasses existing remote sensing segmentation datasets. Moreover, Sky-SA is also the first pixel-level open-text annotated dataset in the field of remote sensing (see Fig. 1), laying a solid data foundation for the development of open-world RS interpretation.", + "bbox": [ + 89, + 532, + 483, + 699 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. SkySense-O Model", + "text_level": 1, + "bbox": [ + 89, + 715, + 272, + 732 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Pretraining for Foundation Model", + "text_level": 1, + "bbox": [ + 89, + 741, + 387, + 758 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The training of SkySense-O is primarily divided into pretraining and fine-tuning stages as shown in Fig. 5. During pre-training, given the input image $I_{i}$ and its corresponding text $T_{i}$ , for $i = 1,\\dots,B$ , where $B$ denotes mini-batch size, the model mainly consists of an image encoder $E_{img}$ and a text encoder $E_{text}$ . To maintain the model's visual perception capabilities [26], especially for densely distributed objects during the image-text alignment process, the vision encoder is initialized by SkySense [23] and its visual", + "bbox": [ + 89, + 763, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "self-supervised paradigms are still maintained, i.e. masked image modeling and image contrastive learning. Given $x = E_{img}(I)$ , $y = E_{text}(T)$ , mask decoder output $R'$ , the overall pretraining loss is computed as Eq. 1:", + "bbox": [ + 511, + 304, + 905, + 366 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {p r e t r a i n}} = C E (x, y) + C E \\left(x, x ^ {\\prime}\\right) + M R \\left(I ^ {\\prime}, R ^ {\\prime}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 529, + 377, + 903, + 395 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $I^{\\prime}$ and $x^{\\prime}$ are the augmented version of $I$ and $x$ respectively, $MR$ is mask reconstruction loss following [71] and $CE$ denotes cross-entropy loss computing as:", + "bbox": [ + 511, + 402, + 905, + 448 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nC E (x, y) = - \\frac {1}{B} \\sum_ {i = 1} ^ {B} \\log \\frac {\\exp \\left(\\operatorname {s i m} \\left(x _ {i} , y _ {i}\\right) / \\tau\\right)}{\\sum_ {j = 1} ^ {B} \\exp \\left(\\operatorname {s i m} \\left(x _ {i} , y _ {j}\\right) / \\tau\\right)}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 522, + 454, + 903, + 496 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\tau$ is logit temperature and $sim$ is cosine similarity.", + "bbox": [ + 511, + 502, + 890, + 517 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Finetuning for Interpretation Model", + "text_level": 1, + "bbox": [ + 511, + 525, + 828, + 540 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Region-level Finetuning. During the region-level finetuning stage, based on the proposed Sky-SA dataset, given a set of candidate category texts $C_n$ and masks $M_n$ for each image, where $n = 1, \\dots, N$ , we apply mask pooling to extract local features for each category from the images, and then compute the region-level cross-entropy loss with the corresponding category, which can be expressed as:", + "bbox": [ + 511, + 547, + 905, + 654 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {r e g i o n}} = \\sum_ {n = 1} ^ {N} C E \\left(P o o l i n g \\left(x _ {n}, M _ {n}\\right), E _ {\\text {t e x t}} \\left(C _ {n}\\right)\\right). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 660, + 903, + 702 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Pixel-level Finetuning. To achieve pixel-level interpretation, referring to the existing open-vocabulary semantic segmentation model [80], we first compute the similarity between the image encoder and the text encoder at the token level to obtain coarse location features. Subsequently, a visual decoder $D_{img}$ is employed to up-sample these features to the fine-grained segmentation results, constrained by a pixel-wise cross-entropy loss as Eq. 4:", + "bbox": [ + 511, + 709, + 905, + 830 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {p i x e l}} = \\sum_ {n = 1} ^ {N} C E \\left(D _ {\\text {i m g}} \\left(x _ {n}, E _ {\\text {t e x t}} \\left(C _ {n}\\right)\\right), M _ {n}\\right). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 537, + 837, + 903, + 878 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "More details can be found in the supplementary and code.", + "bbox": [ + 511, + 885, + 893, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "14736", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/17864a17972ce502161b7bed589dc73601d06f30ee3db7dfcfe6d9edfaee4274.jpg", + "image_caption": [ + "Figure 6. Positive and negative sampling mechanism according to the correspondence score between source sample and target sample. (a): Due to the contextual dependence in language modeling or vision-language modeling [20], \"road\" is mistakenly assigned as the positive sample of \"car\", while our agent employs the correct sampler. (b): Another case to further valid this point." + ], + "image_footnote": [], + "bbox": [ + 94, + 88, + 480, + 244 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/91734500729032caa9ad2f494192bfc436d603a03c7766304614a8aa70a24d37.jpg", + "image_caption": [ + "Figure 7. Constructing illustrations of the vision-centric knowledge graph by an elaborate vision-centric agent flow." + ], + "image_footnote": [], + "bbox": [ + 96, + 347, + 485, + 463 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Vision-Centric Modeling", + "text_level": 1, + "bbox": [ + 89, + 515, + 316, + 531 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Challenge. Different from the previous close-set semantic segmentation dataset, the Sky-SA dataset is annotated in open-world texts. Although this manner provides more open vocabulary support, the ambiguous sampling challenge often leads to training collapse in discriminative models. As illustrated in Fig. 6, \"car\" is mistakenly regarded as the positive sample of \"road\" by computing the similarity between their text-embeddings due to their strong text-level or image-level contextual relationships.", + "bbox": [ + 89, + 537, + 483, + 672 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Visual-Centric Knowledge Graph. To overcome the challenge, inspired by the in-depth exploration of visual representation by Xie et al. [22] and GraphRAG [81], we build a visual-centric agent using GPT-4 to obtain the visual relevance score for each text pair of the Sky-SA dataset. As shown in Fig. 7, the score is measured by the meticulously chain of thought about visual attributes, including appearance, color, and shape etc.. When all text pairs in Sky-SA are fully traversed, we can construct an offline vision-centric knowledge graph, which mainly consists of three elements: source node, target node, and edge. The edge indicates the visual-relevance score and reasons from source node to target node. Based on the graph, we perform flexible sampling for positive and negative samples by setting a sampling threshold illustrated in Algorithm 1.", + "bbox": [ + 89, + 674, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Algorithm 1 Finetuning Stage for SkySense-O.", + "text_level": 1, + "bbox": [ + 514, + 92, + 828, + 106 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/480fbc1faac760c1d7070e935b259045d46bcb58c6a2bbd481836f89834efeac.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 514, + 108, + 919, + 148 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/0288e880412cb19e22e8d16ceceacc632d0063a0f2a128dce5c2777f4b4e5ace.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 516, + 155, + 869, + 224 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/5470adb042dc2cb28ad2c31c09d08e5c552a84e5288fcbac3db1e8ee2a7dac10.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 516, + 226, + 883, + 315 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/21c5c183a66a669e347df6911ddd44d087b6cc640fedc722dbc1af3bf839095d.jpg", + "image_caption": [ + "Figure 8. Sample similarity matrices in computing vision-centric regularization. Hard samples, such as vehicle and road, are exploited by multiplying $S_{t}$ and inversed $S_{v}$ ." + ], + "image_footnote": [], + "bbox": [ + 522, + 327, + 906, + 460 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Vision-Centric Regularization Loss. In RS imagery, the objects and land-cover are often densely packed and intricately intertwined, posing challenges for fine-grained visual-language modeling. To relieve this issue, we further exploit the potential of the vision-centric graph and design a vision-centric regularization loss, $\\mathcal{L}_{\\mathrm{vc}}$ . The regularization is built on the observation that the samples with strong contextual relationships but weak visual relevance scores are hard to distinguish as shown in Fig. 6. Specifically, given selected $M$ negative samples $C_m$ for class $C_i$ , their textual similarities $S_t$ are computed by their embedding similarities. Then, given text pairs, the visual relevance $S_v^{i,m}$ can be retrieved by the visual-centric graph. Finally, $\\mathcal{L}_{\\mathrm{vc}}$ is computed as Eq. 5 and details are in Algorithm 1 and Fig. 8.", + "bbox": [ + 511, + 527, + 906, + 741 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {v c}} = \\frac {1}{B N} \\sum_ {i = 1} ^ {B} \\sum_ {n = 1} ^ {N} \\left\\| \\left(1 - S _ {v} ^ {i, m}\\right) * S _ {t} ^ {i, m} * \\mathcal {L} _ {\\text {r e g i o n}} \\right\\| ^ {2}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 751, + 906, + 792 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Total Loss. In summary, the total loss for the finetuning stage of SkySense-O is defined as follows:", + "bbox": [ + 511, + 801, + 906, + 832 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {f i n e t u n e}} = \\mathcal {L} _ {\\text {p i x e l}} + \\lambda_ {1} \\mathcal {L} _ {\\text {r e g i o n}} + \\lambda_ {2} \\mathcal {L} _ {\\text {v c}}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 565, + 843, + 903, + 861 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda_{1}$ and $\\lambda_{2}$ are two hyperparameters and set as 0.4 and 0.4 respectively for optimal ratio via performing ablations.", + "bbox": [ + 511, + 869, + 906, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "14737", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/d4d606cf15467ca06cddbc56cd9907fd47cbafb8d23458b2052d5f27d2dc0404.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelPublicationObject Semantic SegmentationLand-Cover Semantic Segmentation
iSAID [33]SOTA [34]FAST [34]SIOR [34]Potsdam [82]OEM [30]LoveDA [32]VaihingenFloodNet [83]
Few-Shot Semantic Segmentation
SegGPT [84]ICCV 202329.1731.4119.5537.4730.1225.1931.4047.2733.81
DINOv [85]CVPR 202425.3518.247.9225.0437.6523.0525.1735.1044.55
Zero-Shot Semantic Segmentation
SAN [80]CVPR 202311.7710.733.1615.2251.4524.8727.8346.2339.26
CAT-SEG [75]CVPR 202420.5813.713.1319.6450.8130.2637.1549.6241.36
SegEarth-OV [58]Arxiv 202421.7011.992.8413.9247.1040.3036.9029.1034.00
SkySense-O-43.9232.128.3330.8954.1040.8338.3051.6442.04
-(+22.22)(+18.41)(+5.2)(+11.25)(+2.65)(+0.53)(+1.15)(+2.02)(+0.68)
", + "bbox": [ + 93, + 88, + 906, + 229 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/9f9dc6cacc86d40275dc5fc316720fd52ac4d400b6167fc020bcd35947915e18.jpg", + "table_caption": [ + "Table 2. Zero-shot remote sensing object and land-cover semantic segmentation performances. Bold indicates the highest performance, and underline indicates the second highest performance in mIoU (%)." + ], + "table_footnote": [], + "table_body": "
ModelPublicationSIRI-WHU [86]AID-VQA [28]RSVQA-HR [14]
Acc.Acc.Avg. Acc.
GeoChat [35]CVPR 202466.6372.0370.82
LHRS-Bot [87]ECCV 202462.6691.26-
EarthGPT [88]TGRS 2024--71.15
VHM [89]Arxiv 202470.8891.7073.75
SkySense-O†-72.45 (+1.57)92.95 (+1.25)75.39 (+1.64)
SkySense-O-74.79 (+3.91)94.10 (+2.40)78.09 (+4.34)
", + "bbox": [ + 94, + 268, + 557, + 375 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/fc10b90c6ec4c0d32e1c088ae8dba29738bb0880803e6e86316f879126d265bb.jpg", + "table_caption": [ + "Table 3. Zero-Shot visual question-answering performances, compared with existing MLLMs. $\\dagger$ indicates wo/ pixel-level fine-tuning by Sky-SA." + ], + "table_footnote": [], + "table_body": "
ModelPublicationRESISC [29]AID [28]
CLIP [57]PMLR 202169.2566.70
RemoteCLIP [74]TGRS 202470.8574.3
CLIP-laion-RS [31]-71.7071.21
SkyScript [31]AAA1 202471.7070.94
GeoRSCLIP [24]TGRS 202373.9076.27
SkySense-O†-83.2882.93
-(+9.38)(+6.66)
", + "bbox": [ + 566, + 268, + 901, + 375 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 4. Zero-shot remote sensing scene classification performances in Top-1 accuracy $(\\%)$", + "bbox": [ + 563, + 378, + 901, + 407 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiment", + "text_level": 1, + "bbox": [ + 89, + 426, + 215, + 443 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Evaluation Benchmarks", + "text_level": 1, + "bbox": [ + 89, + 452, + 312, + 465 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We evaluate SkySense-O on four representative RS interpretation tasks. To ensure fair zero-shot comparisons, all experiments are conducted without extra fine-tuning and labels from the evaluation datasets are not used during training. The supplementary materials provide further details on the implementation and visualization results.", + "bbox": [ + 89, + 474, + 482, + 565 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Zero-shot object segmentation. We adopt iSAID [33], FAST [34], SIOR [34], SOTA [34] for the object segmentation experiment. They are chosen considering diverse scenes and categories. We report mIoU results on official validation sets. As depicted in Tab. 2, SkySense-O has achieved the SOTA performance on all object segmentation datasets. On average, it outperforms the previous SOTAs by an average margin of $14.27\\%$ . Furthermore, it even surpasses few-shot SOTAs on most datasets.", + "bbox": [ + 89, + 566, + 482, + 700 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Zero-shot land-cover segmentation. We assess SkySense-O's land-cover segmentation performance on ISPRS Potsdam [82], Vaihingen and LoveDA [32] datasets. SkySense-O excels on all three datasets (Tab. 2). Notably, we surpass the second best CAT-SEG [75] by $1.41\\%$ mIoU and have achieved the best performance on the Vaihingen leaderboard. More importantly, our results demonstrate the potential to achieve more accurate interpretations than the ground truth (the second row in Fig. 9).", + "bbox": [ + 89, + 703, + 482, + 839 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Zero-shot classification. We utilize two scene classification datasets: AID [28] and RESISC-45 [29]. The validation setting follows GeoRSCLIP [24] and Top-1 Accuracy is reported. SkySense-O overall outperforms competitive", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ce92ee71221c85e5c5c28022eff8a3f91a630d76d827870c55ce37b3a724f34e.jpg", + "image_caption": [ + "Figure 9. Visualization of object segmentation results on iSAID dataset and land-cover segmentation results on Potsdam dataset." + ], + "image_footnote": [], + "bbox": [ + 517, + 428, + 906, + 588 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "basielines, achieving $9.38\\%$ and $6.66\\%$ higher performance than state-of-the-art methods on the two datasets (Tab. 4).", + "bbox": [ + 511, + 638, + 906, + 669 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Zero-shot RS VQA. To further validate the versatility of SkySense-O, we employ it in MLLM as a vision encoder for experimentation. Specifically, we maintain the same architecture as LLaVA-1.5 [90], except that we replace the original CLIP-L-14 [57] with our SkySense-O, and set the image resolution to $512 \\times 512$ pixels. Regarding the setup of the instruction tuning data, during the pretraining stage, we utilize CC-558K dataset [60] and 121k image-text pairs converted from ChatearthNet [91] dataset. In the finetuning stage, based on instruction-tuning datasets collected by SkySenseGPT [61] from different sources, we select a subset of 593k data as our downstream instruction finetuning dataset, where details are in the supplementary. We evaluate zero-shot VQA performances on several public RS benchmarks as shown in Tab. 3.", + "bbox": [ + 511, + 672, + 908, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "14738", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a82e7018650b4045c67610b2640045e4331fb8fc96341dc8616e07233e780289.jpg", + "image_caption": [ + "Figure 10. Visualization of text-image alignment in CLIP. (a) Computing the similarity between each text token and all image tokens to obtain text-image relevance (TIR) maps for different categories. (b) Visualizing TIR maps for some high-response texts. (c) Histogram of the maximum response values of TIR maps across the text list. The details of input texts can be referred to the supplementary materials." + ], + "image_footnote": [], + "bbox": [ + 94, + 89, + 903, + 324 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/c97c21c25eed9a59c9115945d4cc3d6ddb619d12401ef9c8752774ebf4304bad.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Foundation \nModelVision \nEncoderVision-Enc. \nParameterNum. of \nImage-textRESISC \n(Acc)
RemoteCLIPViT-L-14304M0.83M70.85
SkyScriptViT-L-14304M2.6M71.70
GeoRSCLIPViT-H-14630M5M73.90
OursSwinV2-H-8654M10M83.28
", + "bbox": [ + 94, + 390, + 486, + 474 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/f3c42c64cd997aab01a37fc9957da8cbf4505bf960082ac527eeda38bc148735.jpg", + "table_caption": [ + "Table 5. Comparison with existing RS vision-language foundation models. SkySense-O achieves consistent performance gains due to scaling laws in terms of model-size and data-size." + ], + "table_footnote": [], + "table_body": "
+Sky-SA+Graph+Pretrain-ing+LvcmIoU
18.40
✓27.76 (+9.36)
✓✓36.28 (+8.52)
✓✓✓41.69 (+5.41)
✓✓✓✓43.92 (+2.24)
", + "bbox": [ + 94, + 534, + 475, + 633 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 6. Ablation study on the proposed dataset and designs.", + "bbox": [ + 104, + 638, + 465, + 652 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Implementation Details.", + "text_level": 1, + "bbox": [ + 89, + 674, + 312, + 690 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In the pre-training stage, we use 80 A100-80GB GPUs and primarily adopt two large-scale remote sensing image-text datasets, SkyScript [31] and RS5M [24], and further expand their captions using Qwen-VL [92] and InternVL [93], comprising 10 million image-text pairs. In the fine-tuning stage, we employ OSM [25] for local image-text fine-tuning and Sky-SA for fine-grained interpretation capabilities on 8 A100-80GB GPUs as shown in Tab. 5. More details are included in the supplementary materials.", + "bbox": [ + 89, + 696, + 483, + 833 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Discussions & Ablation Studies", + "text_level": 1, + "bbox": [ + 89, + 847, + 366, + 862 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effectiveness of Image-Language Alignment. In Fig. 10, we follow MaskCLIP [94] and adopt the image-text correla", + "bbox": [ + 89, + 869, + 483, + 902 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c15e0225292e0a27df5f81a9961f5b4d4f7e210bd01863fa8810eaa63d455cac.jpg", + "image_caption": [ + "Figure 11. Visualizations of open-world interpretation results by SkySense-O on iSAID, Potsdam, FAST, and SIOR datasets." + ], + "image_footnote": [], + "bbox": [ + 516, + 390, + 905, + 544 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "tion in the representation space to evaluate the effectiveness of vision-language foundation models. We select the CLIP models from SegEarth-OV [58] and CAT-Seg [75] for comparison. The former is a standard CLIP model, while the latter is a locally fine-tuned CLIP model by COCO-Stuff [19]. We randomly selected 131 category texts from Sky-SA as inputs. As shown in Figure 10(c), SkySense-O exhibits larger response variance across different categories, indicating stronger discriminative capabilities among different texts. In Fig. 10(b), we present several image-text correlation maps to further validate the effectiveness of SkySense-O in image-text alignment in open-world scenes.", + "bbox": [ + 511, + 609, + 906, + 789 + ], + "page_idx": 6 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- vehicle vs. parking lot. (small object) Because our modeling is vision-centric, small targets are effectively decoupled from the contextual background.", + "- building vs. industrial park. (hypernym) SkySense-O demonstrates proper alignment capabilities across different levels of semantic hierarchy due to large-scale remote sensing knowledge are learned in training." + ], + "bbox": [ + 511, + 792, + 906, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "14739", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5bd44f34fccba169c2e267e81fd621581d7aaeddf938f7cfd803ae9de9a249c8.jpg", + "image_caption": [ + "Figure 12. The visualization of vision-centric graph of the Sky-SA dataset. Node layout is performed via Force Atlas [95] and graph communities are detected by the Leiden algorithm [96]. Circles represent entity nodes with size proportional to their degree. Node colors represent entity communities. Five main communities are detected by an unsupervised manner and they just correspond to the major land-cover categories in the remote sensing imagery." + ], + "image_footnote": [], + "bbox": [ + 96, + 88, + 482, + 250 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ba430331eb48f17344c089da0cf1c3800926cff7baebf5a257e2c44e2fe326e7.jpg", + "image_caption": [ + "(a) Image" + ], + "image_footnote": [], + "bbox": [ + 91, + 373, + 222, + 474 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a1906c6cc5dc0f4f9c18162540578f8356ceab4a82e93a4a3067bf22acbc6bd9.jpg", + "image_caption": [ + "(b) Wo/ vision-centric" + ], + "image_footnote": [], + "bbox": [ + 222, + 373, + 352, + 473 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/77b703c43dbf4e214aa13b80c4b9674227f157c86c20f5f3b4e289cc7a9aa4f4.jpg", + "image_caption": [ + "(c) W/ vision-centric", + "Figure 13. Comparisons of the results without and with vision-centric regularization loss. Compared to (b), the vehicles in (c) are clearly distinguished from the contextual floor." + ], + "image_footnote": [], + "bbox": [ + 352, + 373, + 480, + 474 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Design of Vision-Centric Strategy. Tab. 6 presents the ablation study to assess our Sky-SA dataset and vision-centric designs, namely vision-centric graph sampler, pretraining foundation model, and vision-centric regularization loss. We utilize the iSAID dataset as evaluation benchmark and report mIoU metric on the official test set. Initially, we use the standard CLIP model [57] and fine-tune it on the COCO-Stuff dataset [19]. Compared with the baseline, the Sky-SA dataset leads to a notable improvement compared to COCO-Stuff dataset. Nonetheless, we believe that the Sky-SA dataset holds considerable untapped potential, which might be constrained by the intricate entanglement of open-world text labels as demonstrated in Sec. 4.1. To relieve this problem, we construct a vision-centric graph for all text labels of the Sky-SA dataset as shown in Fig. 12. It is intriguing that the nodes in the graph naturally exhibit proper unsupervised clustering results by community detection [96], without using any text labels. This further validates the rationale for vision-centric modeling and the introducing of vision-centric graph leads to $8.52\\%$ mIoU improvement. Then, we replace the standard CLIP with our pre-trained model incorporating extensive remote sensing knowledge. This effectively improves the perfor", + "bbox": [ + 89, + 553, + 483, + 902 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0db4d3fddac20049b8798486635d427945c423dcad5f36139e91643a3ce48808.jpg", + "image_caption": [ + "Figure 14. Zero-shot semantic segmentation results of Sky-SA and existing datasets. The vertical axis represents the training set, while the horizontal axis represents the test set. The yellow box denotes the result is fine-tuned by its own training set (i.e. supervised learning) and red box denotes the average performance." + ], + "image_footnote": [], + "bbox": [ + 521, + 98, + 906, + 280 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "mance to $41.69\\%$ mIoU, validating the necessity of remote sensing pre-training model. Finally, we introduce vision-centric regularization loss, which uncovers and prioritizes hard samples in visual-language modeling. The results show that the constraint pushes the performance to a very strong $43.92\\%$ mIoU. Furthermore, as shown in Fig. 13, this design enhances the discrimination ability, especially between high-context categories.", + "bbox": [ + 511, + 375, + 906, + 496 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Significance of Sky-SA for Open-World Interpretation. In Fig. 14, we replace Sky-SA with existing datasets and conduct experiments under the same settings as in Tab. 2. It can be observed that existing datasets exhibit poor generalization to other datasets, except on their own test sets. In contrast, the Sky-SA dataset achieves superior zero-shot segmentation performance across all datasets, outperforming the second best dataset by an average of over $13.43\\%$ mIoU, which demonstrates the significant potential of Sky-SA for open-world interpretation.", + "bbox": [ + 511, + 497, + 908, + 650 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 662, + 633, + 679 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we present Skysense-O, a large-scale and general remote sensing interpretation model for open remote sensing images. Skysense-O is mainly driven by the proposed semantically-rich and fine-grained dataset SkySA and the vision-centric visual-language modeling, consistently outperforming other models on a variety of tasks, showcasing its exceptional generalization ability and strong performance. We hope this work will constructively push the RS open-world interpretation on constructing large-scale vision-language datasets and models for further researches. Our future work aims to explore fine-grained remote sensing interpretation, such as analyzing object states and relationships, further enhancing the flexibility of open-world applications.", + "bbox": [ + 511, + 688, + 908, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "14740", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 91, + 90, + 250, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported by the Anhui Provincial Natural Science Foundation under Grant 2108085UD12 and the National Natural Science Key Foundation of China under Grant 62036005. We acknowledge the support of GPU cluster built by MCC Lab of Information Science and Technology Institution, USTC.", + "bbox": [ + 89, + 114, + 483, + 205 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 220, + 187, + 236 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Xianping Ma, Xiaokang Zhang, Man-On Pun, and Bo Huang. Manet: Fine-tuning segment anything model for multimodal remote sensing semantic segmentation, 2024. 1", + "[2] Aysim Token, Marvin Eisenberger, Daniel Cremers, and Laura Leal-Taixe. Satsynth: Augmenting image-mask pairs through diffusion models for aerial semantic segmentation, CVPR 2024.", + "[3] Deyi Ji, Feng Zhao, Hongtao Lu, Feng Wu, and Jieping Ye. Structural and statistical texture knowledge distillation and learning for segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence, pages 1-18, 2025.", + "[4] Deyi Ji, Feng Zhao, Hongtao Lu, Mingyuan Tao, and Jieping Ye. Ultra-high resolution segmentation with ultra-rich context: A novel benchmark. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23621-23630, 2023. 1", + "[5] Deyi Ji, Feng Zhao, Lanyun Zhu, Wenwei Jin, Hongtao Lu, and Jieping Ye. Discrete latent perspective learning for segmentation and detection. In *Forty-first International Conference on Machine Learning*, pages 21719-21730, 2024. 1", + "[6] Botao Ren, Xue Yang, Yi Yu, Junwei Luo, and Zhidong Deng. Pointobb-v2: Towards simpler, faster, and stronger single point supervised oriented object detection. arXiv preprint arXiv:2410.08210, 2024.", + "[7] Kang Wu, Zunxiao Xu, Xinrong Lyu, and Peng Ren. Cloud detection with boundary nets. ISPRS Journal of Photogrammetry and Remote Sensing, 186:218-231, 2022.", + "[8] Peiyuan Zhang, Junwei Luo, Xue Yang, Yi Yu, Qingyun Li, Yue Zhou, Xiaosong Jia, Xudong Lu, Jingdong Chen, Xiang Li, et al. Pointobb-v3: Expanding performance boundaries of single point-supervised oriented object detection. arXiv preprint arXiv:2501.13898, 2025.", + "[9] Yi Yu, Xue Yang, Qingyun Li, Yue Zhou, Feipeng Da, and Junchi Yan. H2rbox-v2: Incorporating symmetry for boosting horizontal box supervised oriented object detection. Advances in Neural Information Processing Systems, 36, 2024.", + "[10] Kang Wu, Zunxiao Xu, Xinrong Lyu, and Peng Ren. Crosssupervised learning for cloud detection. *GIScience & Remote Sensing*, 60(1):2147298, 2023. 1", + "[11] Zhuo Zheng, Ailong Ma, Liangpei Zhang, and Yanfei Zhong. Change is everywhere: Single-temporal supervised object change detection in remote sensing imagery. In Proceedings of the IEEE/CVF international conference on computer vision, pages 15193-15202, 2021. 1, 3", + "[12] Deyi Ji, Siqi Gao, Mingyuan Tao, Hongtao Lu, and Feng Zhao. Changenet: Multi-temporal asymmetric change detect" + ], + "bbox": [ + 93, + 244, + 482, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "tion dataset. In ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing, pages 2725-2729. IEEE, 2024. 1, 3", + "[13] Junwei Luo, Yingying Zhang, Xue Yang, Kang Wu, Qi Zhu, Lei Liang, Jingdong Chen, and Yansheng Li. When large vision-language model meets large remote sensing imagery: Coarse-to-fine text-guided token pruning. arXiv preprint arXiv:2503.07588, 2025. 2", + "[14] Sylvain Lobry, Diego Marcos, Jesse Murray, and Devis Tuia. Rsvqa: Visual question answering for remote sensing data. IEEE Transactions on Geoscience and Remote Sensing, 58(12):8555-8566, 2020. 6", + "[15] Junjue Wang, Zhuo Zheng, Zihang Chen, Ailong Ma, and Yanfei Zhong. Earthvqa: Towards queryable earth via relational reasoning-based remote sensing visual question answering. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 5481-5489, 2024. 2", + "[16] Yan Li, Weiwei Guo, Xue Yang, Ning Liao, Dunyun He, Jiaqi Zhou, and Wenxian Yu. Toward open vocabulary aerial object detection with clip-activated student-teacher learning. In European Conference on Computer Vision, pages 431–448. Springer, 2024. 2", + "[17] Jiancheng Pan, Yanxing Liu, Yuqian Fu, Muyuan Ma, Jiaohao Li, Danda Pani Paudel, Luc Van Gool, and Xiaomeng Huang. Locate anything on earth: Advancing open-vocabulary object detection for remote sensing community. arXiv preprint arXiv:2408.09110, 2024. 2, 3", + "[18] Weiyun Wang, Min Shi, Qingyun Li, Wenhai Wang, Zhenhang Huang, Linjie Xing, Zhe Chen, Hao Li, Xizhou Zhu, Zhiguo Cao, et al. The all-seeing project: Towards panoptic visual recognition and understanding of the open world. arXiv preprint arXiv:2308.01907, 2023. 2", + "[19] Holger Caesar, Jasper Uijlings, and Vittorio Ferrari. Cocostuff: Thing and stuff classes in context. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1209-1218, 2018. 2, 3, 7, 8", + "[20] Shuyang Gu. Several questions of visual generation in 2024. arXiv preprint arXiv:2407.18290, 2024. 2, 5", + "[21] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dollar, and Ross Girshick. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, October 2023. 2, 3, 4", + "[22] Shengbang Tong, Zhuang Liu, Yuexiang Zhai, Yi Ma, Yann LeCun, and Saining Xie. Eyes wide shut? exploring the visual shortcomings of multimodal llms. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9568-9578, 2024. 2, 3, 5", + "[23] Xin Guo, Jiangwei Lao, Bo Dang, Yingying Zhang, Lei Yu, Lixiang Ru, Liheng Zhong, Ziyuan Huang, Kang Wu, Dingxiang Hu, et al. Skysense: A multi-modal remote sensing foundation model towards universal interpretation for earth observation imagery. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 27672-27683, 2024. 2, 3, 4", + "[24] Zilun Zhang, Tiancheng Zhao, Yulong Guo, and Jianwei Yin. Rs5m and georsclip: A large scale vision-language dataset" + ], + "bbox": [ + 516, + 92, + 903, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "14741", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "and a large vision-language model for remote sensing. IEEE Transactions on Geoscience and Remote Sensing, pages 1-1, 2024. 2, 3, 6, 7", + "[25] Haozhe Xie, Zhaoxi Chen, Fangzhou Hong, and Ziwei Liu. CityDreamer: Compositional generative model of unbounded 3D cities. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 2, 7", + "[26] Amin Karimi Monsefi, Kishore Prakash Sailaja, Ali Alilooee, Ser-Nam Lim, and Rajiv Ramnath. Detailclip: Detail-oriented clip for fine-grained tasks. arXiv preprint arXiv:2409.06809, 2024. 2, 4", + "[27] OpenAI. Gpt-4v(ision) system card. 2023. 2, 4", + "[28] Gui-Song Xia, Jingwen Hu, Fan Hu, Baoguang Shi, Xiang Bai, Yanfei Zhong, Liangpei Zhang, and Xiaoqiang Lu. Aid: A benchmark data set for performance evaluation of aerial scene classification. IEEE Transactions on Geoscience and Remote Sensing, 55(7):3965-3981, 2017. 3, 6", + "[29] Gong Cheng, Junwei Han, and Xiaogiang Lu. Remote sensing image scene classification: Benchmark and state of the art. Proceedings of the IEEE, 105(10):1865-1883, 2017. 3, 6", + "[30] Junshi Xia, Naoto Yokoya, Bruno Adriano, and Clifford Broni-Bediako. Openearthmap: A benchmark dataset for global high-resolution land cover mapping. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 6254-6264, 2023. 3, 6", + "[31] Zhecheng Wang, Rajanie Prabha, Tianyuan Huang, Jiajun Wu, and Ram Rajagopal. Skyscript: A large and semantically diverse vision-language dataset for remote sensing. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 5805-5813, 2024. 3, 6, 7", + "[32] Junjue Wang, Zhuo Zheng, Ailong Ma, Xiaoyan Lu, and Yanfei Zhong. LoveDA: A remote sensing land-cover dataset for domain adaptive semantic segmentation, October 2021. 3, 6", + "[33] Syed Waqas Zamir, Aditya Arora, Akshita Gupta, Salman Khan, Guolei Sun, Fahad Shahbaz Khan, Fan Zhu, Ling Shao, Gui-Song Xia, and Xiang Bai. isaid: A large-scale dataset for instance segmentation in aerial images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 28-37, 2019. 3, 6", + "[34] Di Wang, Jing Zhang, Bo Du, Minqiang Xu, Lin Liu, Dacheng Tao, and Liangpei Zhang. Samrs: Scaling-up remote sensing segmentation dataset with segment anything model, NeurIPS 2023. 3, 6", + "[35] Kartik Kuckreja, Muhammad Sohail Danish, Muzammal Naseer, Abhijit Das, Salman Khan, and Fahad Shahbaz Khan. Gechat: Grounded large vision-language model for remote sensing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 27831-27840, 2024. 2, 6", + "[36] Tianrun Chen, Chaotao Ding, Lanyun Zhu, Tao Xu, Deyi Ji, Ying Zang, and Zejian Li. xlstm-unet can be an effective 2d\\& 3d medical image segmentation backbone with vision-lstm (vil) better than its mamba counterpart. arXiv preprint arXiv:2407.01530, 2024." + ], + "bbox": [ + 91, + 90, + 483, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[37] Deyi Ji, Siqi Gao, Lanyun Zhu, Qi Zhu, Yiru Zhao, Peng Xu, Hongtao Lu, Feng Zhao, and Jieping Ye. View-centric multi-object tracking with homographic matching in moving UAV. arXiv preprint arXiv:2403.10830, 2024.", + "[38] Ying Zang, Runlong Cao, Jianqi Zhang, Yidong Han, Ziyue Cao, Wenjun Hu, Didi Zhu, Lanyun Zhu, Zejian Li, Deyi Ji, et al. Let human sketches help: Empowering challenging image segmentation task with freehand sketches. arXiv preprint arXiv:2501.19329, 2025.", + "[39] Lanyun Zhu, Tianrun Chen, Deyi Ji, Jieping Ye, and Jun Liu. Not every patch is needed: Towards a more efficient and effective backbone for video-based person re-identification. IEEE Transactions on Image Processing, 2025.", + "[40] Deyi Ji, Hongtao Lu, and Tongzhen Zhang. End to end multiscale convolutional neural network for crowd counting. In Eleventh international conference on machine vision, volume 11041, pages 761-766. SPIE, 2019.", + "[41] Zunxiao Xu, Kang Wu, Lei Huang, Qimao Wang, and Peng Ren. Cloudy image arithmetic: A cloudy scene synthesis paradigm with an application to deep-learning-based thin cloud removal. IEEE Transactions on Geoscience and Remote Sensing, 60:1-16, 2021.", + "[42] Ying Zang, Yuncan Gao, Jiangi Zhang, Yuangi Hu, Runlong Cao, Lanyun Zhu, Qi Zhu, Deyi Ji, Renjun Xu, and Tianrun Chen. Breaking the box: Enhancing remote sensing image segmentation with freehand sketches. arXiv preprint arXiv:2503.12191, 2025. 2", + "[43] Zhihao Li, Biao Hou, Siteng Ma, Zitong Wu, Xianpeng Guo, Bo Ren, and Licheng Jiao. Masked angle-aware autoencoder for remote sensing images, European Conference on Computer Vision 2024. 3", + "[44] Aaron E Maxwell, Timothy A Warner, and Fang Fang. Implementation of machine-learning classification in remote sensing: An applied review. International journal of remote sensing, 39(9):2784-2817, 2018. 3", + "[45] Xue Yang, Gefan Zhang, Wentong Li, Yue Zhou, Xuehui Wang, and Junchi Yan. H2rbox: Horizontal box annotation is all you need for oriented object detection. In The Eleventh International Conference on Learning Representations, 2023. 3", + "[46] Yansheng Li, Junwei Luo, Yongjun Zhang, Yihua Tan, JinGang Yu, and Song Bai. Learning to holistically detect bridges from large-size vhr remote sensing imagery. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024.", + "[47] Haoran Wang, Licheng Jiao, Fang Liu, Lingling Li, Xu Liu, Deyi Ji, and Weihao Gan. Ipgn: Interactiveness proposal graph network for human-object interaction detection. IEEE Transactions on Image Processing, 30:6583-6593, 2021.", + "[48] Junwei Luo, Xue Yang, Yi Yu, Qingyun Li, Junchi Yan, and Yansheng Li. Pointobb: Learning oriented object detection via single point supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16730-16740, 2024.", + "[49] Yi Yu, Xue Yang, Qingyun Li, Feipeng Da, Jifeng Dai, Yu Qiao, and Junchi Yan. Point2rbox: Combine knowledge from synthetic visual patterns for end-to-end oriented object" + ], + "bbox": [ + 516, + 90, + 906, + 901 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "14742", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "detection with single point supervision. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 16783-16793, 2024. 3", + "[50] Lanyun Zhu, Deyi Ji, Shiping Zhu, Weihao Gan, Wei Wu, and Junjie Yan. Learning statistical texture for semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12537-12546, 2021. 3", + "[51] Tianrun Chen, Ankang Lu, Lanyun Zhu, Chaotao Ding, Chunan Yu, Deyi Ji, Zejian Li, Lingyun Sun, Papa Mao, and Ying Zang. Sam2-adapter: Evaluating & adapting segment anything 2 in downstream tasks: Camouflage, shadow, medical image segmentation, and more. arXiv preprint arXiv:2408.04579, 2024.", + "[52] Gui-Song Xia, Xiang Bai, Jian Ding, Zhen Zhu, Serge Belongie, Jiebo Luo, Mihai Datcu, Marcello Pelillo, and Liangpei Zhang. Dota: A large-scale dataset for object detection in aerial images. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3974-3983, 2018.", + "[53] Deyi Ji, Haoran Wang, Mingyuan Tao, Jianqiang Huang, Xian-Sheng Hua, and Hongtao Lu. Structural and statistical texture knowledge distillation for semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16876-16885, 2022.", + "[54] Deyi Ji, Feng Zhao, and Hongtao Lu. Guided patch-grouping wavelet transformer with spatial congruence for ultra-high resolution segmentation. International Joint Conference on Artificial Intelligence, pages 920-928, 2023.", + "[55] Deyi Ji, Wenwei Jin, Hongtao Lu, and Feng Zhao. Pptformer: Pseudo multi-perspective transformer for uav segmentation. International Joint Conference on Artificial Intelligence, pages 893-901, 2024.", + "[56] Hanzhe Hu, Deyi Ji, Weihao Gan, Shuai Bai, Wei Wu, and Junjie Yan. Class-wise dynamic graph convolution for semantic segmentation. In European Conference on Computer Vision, pages 1-17. Springer, 2020. 3", + "[57] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763, 2021. 3, 6, 8", + "[58] Kaiyu Li, Ruixun Liu, Xiangyong Cao, Deyu Meng, and Zhi Wang. Segearth-ov: Towards training-free open-vocabulary segmentation for remote sensing images. arXiv preprint arXiv:2410.01768, 2024. 3, 6, 7", + "[59] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 3", + "[60] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In A. Oh, T. Naumann, A. Globerson, K. Saenko, M. Hardt, and S. Levine, editors, Advances in Neural Information Processing Systems, volume 36, pages 34892-34916, 2023. 3, 6" + ], + "bbox": [ + 91, + 90, + 482, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[61] Junwei Luo, Zhen Pang, Yongjun Zhang, Tingzhu Wang, Linlin Wang, Bo Dang, Jiangwei Lao, Jian Wang, Jingdong Chen, Yihua Tan, et al. Skysensegpt: A fine-grained instruction tuning dataset and model for remote sensing vision-language understanding. arXiv preprint arXiv:2406.10100, 2024.3,6", + "[62] Tianrun Chen, Chunan Yu, Jing Li, Jianqi Zhang, Lanyun Zhu, Deyi Ji, Yong Zhang, Ying Zang, Zejian Li, and Lingyun Sun. Reasoning3d-grounding and reasoning in 3d: Fine-grained zero-shot open-vocabulary 3d reasoning part segmentation via large vision-language models. arXiv preprint arXiv:2405.19326, 2024. 3", + "[63] Ziyue Huang, Yongchao Feng, Qingjie Liu, and Yunhong Wang. Mutdet: Mutually optimizing pre-training for remote sensing object detection, European Conference on Computer Vision 2024.", + "[64] Zhuo Zheng, Yanfei Zhong, Ailong Ma, and Liangpei Zhang. Single-temporal supervised learning for universal remote sensing change detection. International Journal of Computer Vision, 2024.", + "[65] Deyi Ji, Lanyun Zhu, Siqi Gao, Peng Xu, Hongtao Lu, Jieping Ye, and Feng Zhao. Tree-of-table: Unleashing the power of llms for enhanced large-scale table understanding. arXiv preprint arXiv:2411.08516, 2024.", + "[66] Lanyun Zhu, Tianrun Chen, Deyi Ji, Jieping Ye, and Jun Liu. Llafs: When large language models meet few-shot segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3065-3075, 2024.", + "[67] Lanyun Zhu, Deyi Ji, Tianrun Chen, Peng Xu, Jieping Ye, and Jun Liu. Ibid: Alleviating hallucinations in large vision-language models via image-biased decoding. arXiv preprint arXiv:2402.18476, 2024.3", + "[68] Shizhen Chang and Pedram Ghamisi. Changes to captions: An attentive network for remote sensing change captioning. IEEE Transactions on Image Processing, 32:6047-6060, 2023. 3", + "[69] Deyi Ji, Haoran Wang, Hanzhe Hu, Weihao Gan, Wei Wu, and Junjie Yan. Context-aware graph convolution network for target re-identification. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 1646–1654, 2021.", + "[70] Haoran Wang, Licheng Jiao, Fang Liu, Lingling Li, Xu Liu, Deyi Ji, and Weihao Gan. Learning social spatio-temporal relation graph in the wild and a video benchmark. IEEE Transactions on Neural Networks and Learning Systems, 34(6):2951-2964, 2021.", + "[71] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16000-16009, 2022. 3, 4", + "[72] Xian Sun, Peijin Wang, Wanxuan Lu, Zicong Zhu, Xiaonan Lu, Qibin He, Junxi Li, Xuee Rong, Zhujun Yang, Hao Chang, et al. Ringmo: A remote sensing foundation model with masked image modeling. IEEE Transactions on Geoscience and Remote Sensing, 61:1-22, 2022. 3" + ], + "bbox": [ + 516, + 90, + 903, + 898 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "14743", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[73] Colorado J Reed, Ritwik Gupta, Shufan Li, Sarah Brockman, Christopher Funk, Brian Clipp, Kurt Keutzer, Salvatore Candido, Matt Uytendaele, and Trevor Darrell. Scale-mae: A scale-aware masked autoencoder for multiscale geospatial representation learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4088-4099, 2023. 3", + "[74] Fan Liu, Delong Chen, Zhangqingyun Guan, Xiaocong Zhou, Jiale Zhu, Qiaolin Ye, Liyong Fu, and Jun Zhou. Remoteclip: A vision language foundation model for remote sensing. IEEE Transactions on Geoscience and Remote Sensing, 2024. 3, 6", + "[75] Seokju Cho, Heeseong Shin, Sunghwan Hong, Anurag Arnab, Paul Hongsuck Seo, and Seungryong Kim. Catseg: Cost aggregation for open-vocabulary semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4113-4123, 2024. 3, 6, 7", + "[76] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 633-641, 2017. 3", + "[77] Bin Xiao, Haiping Wu, Weijian Xu, Xiyang Dai, Houdong Hu, Yumao Lu, Michael Zeng, Ce Liu, and Lu Yuan. Florence-2: Advancing a unified representation for a variety of vision tasks, 2023. 3", + "[78] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Qing Jiang, Chunyuan Li, Jianwei Yang, Hang Su, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023. 3, 4", + "[79] Zhuang Liu and Kaiming He. A decade's battle on dataset bias: Are we there yet? arXiv preprint arXiv:2403.08632, 2024.3", + "[80] Mengde Xu, Zheng Zhang, Fangyun Wei, Han Hu, and Xi-ang Bai. Side adapter network for open-vocabulary semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2945–2954, 2023. 4, 6", + "[81] Boci Peng, Yun Zhu, Yongchao Liu, Xiaohe Bo, Haizhou Shi, Chuntao Hong, Yan Zhang, and Siliang Tang. Graph retrieval-augmented generation: A survey. arXiv preprint arXiv:2408.08921, 2024. 5", + "[82] Jamie Sherrah. Fully convolutional networks for dense semantic labelling of high-resolution aerial imagery. arXiv preprint arXiv:1606.02585, 2016. 6", + "[83] Maryam Rahnemoonfar, Tashnim Chowdhury, Argho Sarkar, Debvrat Varshney, Masoud Yari, and RobinR. Murphy. Floodnet: A high resolution aerial imagery dataset for post flood scene understanding. IEEE Access, Dec 2020. 6", + "[84] Xinlong Wang, Xiaosong Zhang, Yue Cao, Wen Wang, Chunhua Shen, and Tiejun Huang. Seggpt: Segmenting everything in context. arXiv preprint arXiv:2304.03284, 2023. 6", + "[85] Feng Li, Qing Jiang, Hao Zhang, Tianhe Ren, Shilong Liu, Xueyan Zou, Huaizhe Xu, Hongyang Li, Chunyuan Li, Jian-" + ], + "bbox": [ + 91, + 90, + 485, + 901 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "wei Yang, et al. Visual in-context prompting. arXiv preprint arXiv:2311.13601, 2023. 6", + "[86] Qiqi Zhu, Yanfei Zhong, Bei Zhao, Gui-Song Xia, and Liangpei Zhang. Bag-of-visual-words scene classifier with local and global features for high spatial resolution remote sensing imagery. IEEE Geoscience and Remote Sensing Letters, 13(6):747-751, 2016. 6", + "[87] Dilxat Muhtar, Zhenshi Li, Feng Gu, Xueliang Zhang, and Pengfeng Xiao. Lhrs-bot: Empowering remote sensing with vgi-enhanced large multimodal language model. arXiv preprint arXiv:2402.02544, 2024. 6", + "[88] Wei Zhang, Miaoxin Cai, Tong Zhang, Yin Zhuang, and Xuerui Mao. Earthgpt: A universal multi-modal large language model for multi-sensor image comprehension in remote sensing domain. IEEE Transactions on Geoscience and Remote Sensing, 2024. 6", + "[89] Chao Pang, Xingxing Weng, Jiang Wu, Jiayu Li, Yi Liu, Ji-axing Sun, Weijia Li, Shuai Wang, Litong Feng, Gui-Song Xia, and Conghui He. Vhm: Versatile and honest vision language model for remote sensing image analysis, 2024. 6", + "[90] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26296-26306, 2024. 6", + "[91] Zhenghang Yuan, Zhitong Xiong, Lichao Mou, and Xiao Xiang Zhu. Chatearthnet: a global-scale image-text dataset empowering vision-language geo-foundation models. Earth System Science Data Discussions, 2024:1-24, 2024. 6", + "[92] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966, 2023. 7", + "[93] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24185–24198, 2024. 7", + "[94] Xiaoyi Dong, Jianmin Bao, Yinglin Zheng, Ting Zhang, Dongdong Chen, Hao Yang, Ming Zeng, Weiming Zhang, Lu Yuan, Dong Chen, et al. Maskclip: Masked self-distillation advances contrastive language-image pretraining. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10995-11005, 2023. 7", + "[95] Mathieu Jacomy, Tommaso Venturini, Sebastien Heymann, and Mathieu Bastian. Forceatlas2, a continuous graph layout algorithm for handy network visualization designed for the gephi software. PLoS ONE, page e98679, Jun 2014. 8", + "[96] V. A. Traag, L. Waltman, and N. J. van Eck. From louvain to leiden: guaranteeing well-connected communities. Scientific Reports, Mar 2019. 8" + ], + "bbox": [ + 516, + 90, + 906, + 825 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "14744", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2025/SkySense-O_ Towards Open-World Remote Sensing Interpretation with Vision-Centric Visual-Language Modeling/2415dd78-f506-4cd3-a12c-42283424590a_model.json b/2025/SkySense-O_ Towards Open-World Remote Sensing Interpretation with Vision-Centric Visual-Language Modeling/2415dd78-f506-4cd3-a12c-42283424590a_model.json new file mode 100644 index 0000000000000000000000000000000000000000..2cde7c961d1a6e155ac20ce62b03e65cbabeb71c --- /dev/null +++ b/2025/SkySense-O_ Towards Open-World Remote Sensing Interpretation with Vision-Centric Visual-Language Modeling/2415dd78-f506-4cd3-a12c-42283424590a_model.json @@ -0,0 +1,2930 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.164, + 0.131, + 0.834, + 0.177 + ], + "angle": 0, + "content": "SkySense-O: Towards Open-World Remote Sensing Interpretation with Vision-Centric Visual-Language Modeling" + }, + { + "type": "text", + "bbox": [ + 0.111, + 0.203, + 0.884, + 0.24 + ], + "angle": 0, + "content": "Qi Zhu\\(^{1*}\\), Jiangwei Lao\\(^{2*}\\), Deyi Ji\\(^{1}\\), Junwei Luo\\(^{3}\\), Kang Wu\\(^{3}\\), Yingying Zhang\\(^{4}\\), Lixiang Ru\\(^{3}\\), Jian Wang\\(^{4}\\), Jingdong Chen\\(^{5}\\), Ming Yang\\(^{6}\\), Dong Liu\\(^{1}\\), Feng Zhao\\(^{1\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.111, + 0.24, + 0.887, + 0.277 + ], + "angle": 0, + "content": "\\(^{1}\\)MoE Key Laboratory of Brain-inspired Intelligent Perception and Cognition, USTC \\(^{2}\\)Shanghai Jiaotong University \\(^{3}\\)Wuhan University \\(^{4}\\)UCAS \\(^{5}\\)HUST \\(^{6}\\)Northwestern University" + }, + { + "type": "text", + "bbox": [ + 0.288, + 0.278, + 0.706, + 0.292 + ], + "angle": 0, + "content": "zqcrafts@mail.ustc.edu.cn, fzhao956@ustc.edu.cn" + }, + { + "type": "image", + "bbox": [ + 0.098, + 0.311, + 0.382, + 0.414 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.384, + 0.31, + 0.668, + 0.414 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.098, + 0.416, + 0.668, + 0.521 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.682, + 0.311, + 0.897, + 0.519 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.533, + 0.908, + 0.561 + ], + "angle": 0, + "content": "Figure 1. SkySense-O offers 4 key advantages from the foundation model to the agent application (Left), leading to state-of-the-art zero-shot performances across 14 downstream datasets over 4 distinct RS interpretation tasks (Right)" + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.581, + 0.327, + 0.597 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.612, + 0.486, + 0.877 + ], + "angle": 0, + "content": "Open-world interpretation aims to accurately localize and recognize all objects within images by vision-language models (VLMs). While substantial progress has been made in this task for natural images, the advancements for remote sensing (RS) images still remain limited, primarily due to these two challenges. 1) Existing RS semantic categories are limited, particularly for pixel-level interpretation datasets. 2) Distinguishing among diverse RS spatial regions solely by language space is challenging due to the dense and intricate spatial distribution in open-world RS imagery. To address the first issue, we develop a fine-grained RS interpretation dataset, Sky-SA, which contains 183,375 high-quality local image-text pairs with full-pixel manual annotations, covering 1,763 category labels, exhibiting richer semantics and higher density than previous datasets. Afterwards, to solve the second issue, we introduce the vision-centric principle for vision-language modeling. Specifically, in the pre-training stage, the visual self-supervised paradigm is incorporated into image-" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.583, + 0.907, + 0.763 + ], + "angle": 0, + "content": "text alignment, reducing the degradation of general visual representation capabilities of existing paradigms. Then, we construct a visual-relevance knowledge graph across open-category texts and further develop a novel vision-centric image-text contrastive loss for fine-tuning with text prompts. This new model, denoted as SkySense-O, demonstrates impressive zero-shot capabilities on a thorough evaluation encompassing 14 datasets over 4 tasks, from recognizing to reasoning and classification to localization. Specifically, it outperforms the latest models such as SegEarthOV, GeoRSCLIP, and VHM by a large margin, i.e., \\(11.95\\%\\), \\(8.04\\%\\) and \\(3.55\\%\\) on average respectively. The code is publicly available to facilitate further research at https:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.765, + 0.808, + 0.778 + ], + "angle": 0, + "content": "//github.com/zqcrafts/SkySense-0." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.8, + 0.646, + 0.815 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.827, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Remote sensing interpretation (RSI) is a critical and practical task in Earth observation, aiming to identify the semantic types of specific ground objects in remote sensing (RS) images (such as segmentation [1-4], detection [5-10], change detection [11, 12], and visual question answering" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.889, + 0.348, + 0.902 + ], + "angle": 0, + "content": "*Equal contribution. † Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14733" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.258 + ], + "angle": 0, + "content": "(VQA) [13-15]). Early works typically involved annotating RS images with fixed categories and exploring RSI through supervised learning methods. These approaches do not generalize well to unseen new categories in open-world scenarios, thus limiting their practical applicability. Moreover, as RS images becoming more accessible than ever before, there is an increasing demand for open-world interpretations with a broad category scope. Hence, there is a soaring interest in exploring a comprehensive open-world remote sensing interpretation model for many scenarios and tasks [16, 17]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.26, + 0.485, + 0.474 + ], + "angle": 0, + "content": "Challenges. In the domain of natural image scenes, significant progress has been achieved in open-world interpretation tasks through large-scale localization and recognition datasets like AS-1B [18] and COCO-Stuff [19], combined with contrastive language-image pretrained model (CLIP) shown in Fig. 1a. However, the open-world RS interpretation model is still out of reach due to following two reasons: 1) From the dataset perspective, the existing annotations either exhibit sparse mask labels or lack diversity in category descriptions (see Fig.1b). 2) From the modeling perspective, the language model inherently encompasses strong contextual dependence [20], making it challenging to directly use as prompts for decoupling common RS objects and intertwined land-cover categories in open-world." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.476, + 0.484, + 0.734 + ], + "angle": 0, + "content": "Sky-SA Dataset. For Struggle (1), inspired by SAM [21] and ASM [18] in the natural image domain, we propose the Sky-SA dataset with the largest number of category labels for open-world RS interpretation, including 1,763 category texts from 183,375 mask-text pairs. The Sky-SA is mainly characterized by its text openness and fine granularity (Fig. 1b). Specifically, different from general semantic segmentation datasets, the categories of Sky-SA are annotated by short semantic tags in a completely open manner. In other words, the annotators are doing fill-in-the-blank rather than multiple-choice. This enables the model to learn pixel-level open-world interpretation akin to human cognition [18] (Fig. 1c). Furthermore, similar to COCO-Stuff [19], all recognizable pixels of Sky-SA are densely annotated, which further enhances the semantic openness and the contextual relationships from wide semantic categories." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Vision-Centric Modeling. For Struggle (2), motivated by [22], we introduce the vision-centric principle to correct the potential alignment ambiguity arising from language guidance. Firstly, we initialize the visual encoder from SkySense [23], an advanced RS visual foundation model, and then collect large-scale RS image-caption datasets for pretraining, which mainly include SkyScript [23], RS5M [24] and OSM [25]. Motivated by DetailCLIP [26], to mitigate the degradation of general visual representation capabilities originating from SkySense, we maintain the visual contrastive loss and masked modeling loss during" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.092, + 0.897, + 0.341 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.517, + 0.352, + 0.902, + 0.368 + ], + "angle": 0, + "content": "Figure 2. The flowchart of the overall SkySense-O development." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.385, + 0.907, + 0.612 + ], + "angle": 0, + "content": "the pre-training process of image-text alignment. Secondly, for fine-tuning region-level alignment, we leverage GPT-4V [27] to construct a visual-centric knowledge graph across all categories of the Sky-SA dataset, where the node of the graph represents each category, and the edge weight indicates the visual similarity scores between each pair of categories. Using the graph, we derive a vision-centric regularization that focuses on distinguishing categories exhibiting low visual similarity but strong text-level contextual relationships in RS images. Finally, we explore the impacts of Skysense-O as agent on MLLM models (Fig. 1d). In summary, to achieve open-world RS interpretation model, SkySense-O is meticulously developed from dataset to modeling and from pre-training to fine-tuning (Fig. 2), contributions of this paper include:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.623, + 0.905, + 0.653 + ], + "angle": 0, + "content": "- We develop Sky-SA dataset with large-scale and high-quality annotated texts for open-world RS interpretation." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.654, + 0.905, + 0.713 + ], + "angle": 0, + "content": "- The vision-centric principle is first introduced into RS visual-language modeling, primarily instantiated into training paradigm, vision-centric graph, and vision-centric regularization from pre-training to fine-tuning." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.714, + 0.905, + 0.759 + ], + "angle": 0, + "content": "- We propose SkySense-O, a remarkable open-world RS interpretation model, which achieves SOTA across 14 zeroshot downstream datasets as shown in Fig. 1." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.623, + 0.905, + 0.759 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.776, + 0.652, + 0.792 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.802, + 0.787, + 0.818 + ], + "angle": 0, + "content": "2.1. Remote Sensing Interpretation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Remote Sensing (RS) interpretation aims to accurately identify and analyze ground objects from RS data, with applications in environmental monitoring, urban planning, and agricultural surveillance [35-42]. Early deep learning approaches focused on closed-set supervised models, in" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14734" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.089, + 0.487, + 0.263 + ], + "angle": 0, + "content": "
DatasetSpatial GranularityOpen TextDense Annotation
ImageRegionPixel
Natural Image
SA-1B [21]✓✓✓
COCO-Stuff [19]✓✓✓
Remote Sensing Image
AID [28], RESISC [29]✓
OpenStreetMap [30]✓✓
RS5M [24], SkyScript [31]✓✓
Potsdam, LoveDA [32]✓✓✓
Valiggan, OEM [30]✓✓✓
iSAID [33], SIOR [34]✓✓
FAST [34], SOTA [34]✓✓
Sky-SA (ours)✓✓✓✓✓
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.273, + 0.483, + 0.318 + ], + "angle": 0, + "content": "Table 1. Property comparisons with popular vision-language datasets. The proposed Sky-SA dataset is characterized by its semantic openness, high-density and fine-granularity." + }, + { + "type": "image", + "bbox": [ + 0.098, + 0.337, + 0.194, + 0.41 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.122, + 0.411, + 0.171, + 0.419 + ], + "angle": 0, + "content": "(a) Florence-2" + }, + { + "type": "image", + "bbox": [ + 0.195, + 0.337, + 0.29, + 0.41 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.226, + 0.411, + 0.258, + 0.419 + ], + "angle": 0, + "content": "(b) SAM" + }, + { + "type": "image", + "bbox": [ + 0.291, + 0.337, + 0.38, + 0.41 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.293, + 0.411, + 0.379, + 0.419 + ], + "angle": 0, + "content": "(c) Grounding DINO-1.6" + }, + { + "type": "image", + "bbox": [ + 0.38, + 0.337, + 0.477, + 0.41 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.394, + 0.411, + 0.464, + 0.419 + ], + "angle": 0, + "content": "(d) Sky-SA Dataset" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.424, + 0.483, + 0.455 + ], + "angle": 0, + "content": "Figure 3. Visual comparisons between the predictions of existing open-world models and the annotations of our Sky-SA dataset." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.476, + 0.484, + 0.611 + ], + "angle": 0, + "content": "cluding image classification [43, 44], object detection [45-49], change detection [11, 12], and semantic segmentation [28, 50-56]. Recently, CLIP [57] and Vision Language Models (VLMs) have enabled the transition from closed-set to open-set interpretation, allowing models to recognize broader categories [17, 58]. Additionally, models like GPT-4 [59] and LLaVA [60] have facilitated the development of instruction-tuning datasets [61], enhancing the flexibility of RS models in human-machine interactions." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.633, + 0.483, + 0.65 + ], + "angle": 0, + "content": "2.2. Vision Foundation Models for Remote Sensing" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.66, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Vision foundation models are instrumental in advancing RS interpretation toward open-set capabilities through two primary pre-training paradigms, vision-only self-supervised and vision-language self-supervised [22, 43, 62-67] paradigms. Vision-only self-supervised models utilize large-scale unannotated RS images to develop robust visual representations for various tasks [23, 68-71]. For example, SkySense [23] integrates multiple modalities and geo-contextual information within a contrastive framework, while Masked Image Modeling (MIM) methods like RingMo [72] and Scale-MAE [73] are proposed for further enhancing visual feature learning. Vision-language self-supervised models originate from CLIP. Recently some efforts such as RemoteCLIP [74], RS5M [24], and SkyScript [23] have customized for the RS domain using extensive RS image-caption pairs." + }, + { + "type": "image", + "bbox": [ + 0.532, + 0.091, + 0.918, + 0.27 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.275, + 0.908, + 0.318 + ], + "angle": 0, + "content": "Figure 4. Scale comparisons with existing RS semantic segmentation datasets. We see that the proposed Sky-SA dataset has a significantly larger category scale and diversity than prior datasets." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.345, + 0.671, + 0.363 + ], + "angle": 0, + "content": "3. Sky-SA Dataset" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.371, + 0.637, + 0.387 + ], + "angle": 0, + "content": "3.1. Motivation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.394, + 0.907, + 0.668 + ], + "angle": 0, + "content": "Recently, image-text alignment models based on natural images, such as the open-vocabulary methods [75] and MLLMs like LLaVA [60], have achieved remarkable success. This is primarily due to well-constructed datasets in the natural image domain compared to the RS image domain. There have been numerous high-quality datasets with rich categories in the natural image community, such as COCO-Stuff [19] and ADE [76]. However, as shown in Tab.1 and Fig.4, such resources remain scarce in the RS community, which significantly limits the development of open-world RS interpretation tasks. Moreover, existing open-world models, such as SAM [21], Florence-2 [77], and Grounding-DINO [78], are also unable to achieve dense and fine-grained interpretation on remote sensing images as shown in Fig. 3. To relieve this issue, we design a data engine to help annotators more efficiently and accurately perform full-pixel annotation on unlabeled RS images, leading to the creation of the Sky-SA dataset as following section." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.68, + 0.646, + 0.696 + ], + "angle": 0, + "content": "3.2. Data Engine" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.703, + 0.907, + 0.854 + ], + "angle": 0, + "content": "Open-Images Collection. In the standard collection process of a closed-set dataset, a category list is predefined, and images that do not contain any items from this list will be filtered out. This manner might lead to dataset bias [79] in remote sensing images. Unlike the close-set methods, we compile a comprehensive collection of images from various available RS datasets. To minimize redundancy in manual annotation, we leverage the SkySense model [23], which possesses image-level visual discrimination capabilities for RS images, to filter visually redundant images." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.856, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Open-Text Fine-Grained Annotation. In the annotation process, drawing inspiration from the popular COCO-Stuff [19] dataset, we prioritize fine-grained semantics and" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14735" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.133, + 0.094, + 0.377, + 0.105 + ], + "angle": 0, + "content": "Pre-training (For image-level RS interpretation task)" + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.108, + 0.418, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.521, + 0.094, + 0.804, + 0.105 + ], + "angle": 0, + "content": "Supervised Finetuning (For pixel-level RS interpretation task)" + }, + { + "type": "image", + "bbox": [ + 0.432, + 0.108, + 0.9, + 0.265 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.275, + 0.905, + 0.29 + ], + "angle": 0, + "content": "Figure 5. The framework of SkySense-O, where training strategy includes pre-training and fine-tuning stage for open-world interpretation." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.305, + 0.484, + 0.531 + ], + "angle": 0, + "content": "full-pixel annotation as core labeling principles. Specifically, there are two key points: 1) the dataset encompasses a rich variety of semantic categories; 2) almost all pixels in each image need to be annotated. The former aids in activating the broad semantic awareness of the vision-language pretrained model, while the latter helps establish comprehensive connections among local semantics throughout each image. Furthermore, to capture a broader semantic range, we employ a short-caption labeling manner for each region, which does not constrain the annotation labels. Additionally, to improve annotation efficiency, we first utilize GPT-4V [27] to generate the image content, and then employ Grounding-DINO [78] and SAM [21] for pre-annotation, allowing annotators to make careful corrections based on these preliminary labels." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.533, + 0.484, + 0.7 + ], + "angle": 0, + "content": "Dataset Statistics. Finally, we develop a high-quality pixel-level remote sensing interpretation dataset, named Sky-SA. As shown in Fig. 4, the Sky-SA dataset consists of 35,000 RS images, mainly including 1,763 category texts from 183,375 segment-text pairs (encompassing approximately 500 distinct categories of land-covering and RS objects after synonym filtering), which far surpasses existing remote sensing segmentation datasets. Moreover, Sky-SA is also the first pixel-level open-text annotated dataset in the field of remote sensing (see Fig. 1), laying a solid data foundation for the development of open-world RS interpretation." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.716, + 0.274, + 0.733 + ], + "angle": 0, + "content": "4. SkySense-O Model" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.742, + 0.388, + 0.759 + ], + "angle": 0, + "content": "4.1. Pretraining for Foundation Model" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.484, + 0.901 + ], + "angle": 0, + "content": "The training of SkySense-O is primarily divided into pretraining and fine-tuning stages as shown in Fig. 5. During pre-training, given the input image \\(I_{i}\\) and its corresponding text \\(T_{i}\\), for \\(i = 1,\\dots,B\\), where \\(B\\) denotes mini-batch size, the model mainly consists of an image encoder \\(E_{img}\\) and a text encoder \\(E_{text}\\). To maintain the model's visual perception capabilities [26], especially for densely distributed objects during the image-text alignment process, the vision encoder is initialized by SkySense [23] and its visual" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.305, + 0.906, + 0.367 + ], + "angle": 0, + "content": "self-supervised paradigms are still maintained, i.e. masked image modeling and image contrastive learning. Given \\( x = E_{img}(I) \\), \\( y = E_{text}(T) \\), mask decoder output \\( R' \\), the overall pretraining loss is computed as Eq. 1:" + }, + { + "type": "equation", + "bbox": [ + 0.53, + 0.378, + 0.905, + 0.396 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {p r e t r a i n}} = C E (x, y) + C E \\left(x, x ^ {\\prime}\\right) + M R \\left(I ^ {\\prime}, R ^ {\\prime}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.403, + 0.906, + 0.449 + ], + "angle": 0, + "content": "where \\(I^{\\prime}\\) and \\(x^{\\prime}\\) are the augmented version of \\(I\\) and \\(x\\) respectively, \\(MR\\) is mask reconstruction loss following [71] and \\(CE\\) denotes cross-entropy loss computing as:" + }, + { + "type": "equation", + "bbox": [ + 0.524, + 0.455, + 0.905, + 0.497 + ], + "angle": 0, + "content": "\\[\nC E (x, y) = - \\frac {1}{B} \\sum_ {i = 1} ^ {B} \\log \\frac {\\exp \\left(\\operatorname {s i m} \\left(x _ {i} , y _ {i}\\right) / \\tau\\right)}{\\sum_ {j = 1} ^ {B} \\exp \\left(\\operatorname {s i m} \\left(x _ {i} , y _ {j}\\right) / \\tau\\right)}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.503, + 0.891, + 0.518 + ], + "angle": 0, + "content": "where \\(\\tau\\) is logit temperature and \\(sim\\) is cosine similarity." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.526, + 0.829, + 0.541 + ], + "angle": 0, + "content": "4.2. Finetuning for Interpretation Model" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.548, + 0.906, + 0.655 + ], + "angle": 0, + "content": "Region-level Finetuning. During the region-level finetuning stage, based on the proposed Sky-SA dataset, given a set of candidate category texts \\( C_n \\) and masks \\( M_n \\) for each image, where \\( n = 1, \\dots, N \\), we apply mask pooling to extract local features for each category from the images, and then compute the region-level cross-entropy loss with the corresponding category, which can be expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.525, + 0.661, + 0.905, + 0.703 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {r e g i o n}} = \\sum_ {n = 1} ^ {N} C E \\left(P o o l i n g \\left(x _ {n}, M _ {n}\\right), E _ {\\text {t e x t}} \\left(C _ {n}\\right)\\right). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.71, + 0.906, + 0.831 + ], + "angle": 0, + "content": "Pixel-level Finetuning. To achieve pixel-level interpretation, referring to the existing open-vocabulary semantic segmentation model [80], we first compute the similarity between the image encoder and the text encoder at the token level to obtain coarse location features. Subsequently, a visual decoder \\( D_{img} \\) is employed to up-sample these features to the fine-grained segmentation results, constrained by a pixel-wise cross-entropy loss as Eq. 4:" + }, + { + "type": "equation", + "bbox": [ + 0.538, + 0.838, + 0.905, + 0.88 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {p i x e l}} = \\sum_ {n = 1} ^ {N} C E \\left(D _ {\\text {i m g}} \\left(x _ {n}, E _ {\\text {t e x t}} \\left(C _ {n}\\right)\\right), M _ {n}\\right). \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.886, + 0.895, + 0.901 + ], + "angle": 0, + "content": "More details can be found in the supplementary and code." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14736" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.089, + 0.482, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.251, + 0.485, + 0.334 + ], + "angle": 0, + "content": "Figure 6. Positive and negative sampling mechanism according to the correspondence score between source sample and target sample. (a): Due to the contextual dependence in language modeling or vision-language modeling [20], \"road\" is mistakenly assigned as the positive sample of \"car\", while our agent employs the correct sampler. (b): Another case to further valid this point." + }, + { + "type": "image", + "bbox": [ + 0.098, + 0.348, + 0.486, + 0.464 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.471, + 0.483, + 0.5 + ], + "angle": 0, + "content": "Figure 7. Constructing illustrations of the vision-centric knowledge graph by an elaborate vision-centric agent flow." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.516, + 0.318, + 0.532 + ], + "angle": 0, + "content": "4.3. Vision-Centric Modeling" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.539, + 0.484, + 0.674 + ], + "angle": 0, + "content": "Challenge. Different from the previous close-set semantic segmentation dataset, the Sky-SA dataset is annotated in open-world texts. Although this manner provides more open vocabulary support, the ambiguous sampling challenge often leads to training collapse in discriminative models. As illustrated in Fig. 6, \"car\" is mistakenly regarded as the positive sample of \"road\" by computing the similarity between their text-embeddings due to their strong text-level or image-level contextual relationships." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.675, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Visual-Centric Knowledge Graph. To overcome the challenge, inspired by the in-depth exploration of visual representation by Xie et al. [22] and GraphRAG [81], we build a visual-centric agent using GPT-4 to obtain the visual relevance score for each text pair of the Sky-SA dataset. As shown in Fig. 7, the score is measured by the meticulously chain of thought about visual attributes, including appearance, color, and shape etc.. When all text pairs in Sky-SA are fully traversed, we can construct an offline vision-centric knowledge graph, which mainly consists of three elements: source node, target node, and edge. The edge indicates the visual-relevance score and reasons from source node to target node. Based on the graph, we perform flexible sampling for positive and negative samples by setting a sampling threshold illustrated in Algorithm 1." + }, + { + "type": "title", + "bbox": [ + 0.516, + 0.093, + 0.83, + 0.107 + ], + "angle": 0, + "content": "Algorithm 1 Finetuning Stage for SkySense-O." + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.109, + 0.92, + 0.149 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.156, + 0.87, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.227, + 0.885, + 0.316 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.328, + 0.907, + 0.462 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.471, + 0.907, + 0.514 + ], + "angle": 0, + "content": "Figure 8. Sample similarity matrices in computing vision-centric regularization. Hard samples, such as vehicle and road, are exploited by multiplying \\( S_{t} \\) and inversed \\( S_{v} \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.529, + 0.907, + 0.742 + ], + "angle": 0, + "content": "Vision-Centric Regularization Loss. In RS imagery, the objects and land-cover are often densely packed and intricately intertwined, posing challenges for fine-grained visual-language modeling. To relieve this issue, we further exploit the potential of the vision-centric graph and design a vision-centric regularization loss, \\(\\mathcal{L}_{\\mathrm{vc}}\\). The regularization is built on the observation that the samples with strong contextual relationships but weak visual relevance scores are hard to distinguish as shown in Fig. 6. Specifically, given selected \\(M\\) negative samples \\(C_m\\) for class \\(C_i\\), their textual similarities \\(S_t\\) are computed by their embedding similarities. Then, given text pairs, the visual relevance \\(S_v^{i,m}\\) can be retrieved by the visual-centric graph. Finally, \\(\\mathcal{L}_{\\mathrm{vc}}\\) is computed as Eq. 5 and details are in Algorithm 1 and Fig. 8." + }, + { + "type": "equation", + "bbox": [ + 0.525, + 0.752, + 0.907, + 0.793 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {v c}} = \\frac {1}{B N} \\sum_ {i = 1} ^ {B} \\sum_ {n = 1} ^ {N} \\left\\| \\left(1 - S _ {v} ^ {i, m}\\right) * S _ {t} ^ {i, m} * \\mathcal {L} _ {\\text {r e g i o n}} \\right\\| ^ {2}. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.802, + 0.907, + 0.833 + ], + "angle": 0, + "content": "Total Loss. In summary, the total loss for the finetuning stage of SkySense-O is defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.566, + 0.844, + 0.905, + 0.862 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {f i n e t u n e}} = \\mathcal {L} _ {\\text {p i x e l}} + \\lambda_ {1} \\mathcal {L} _ {\\text {r e g i o n}} + \\lambda_ {2} \\mathcal {L} _ {\\text {v c}}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.907, + 0.902 + ], + "angle": 0, + "content": "where \\(\\lambda_{1}\\) and \\(\\lambda_{2}\\) are two hyperparameters and set as 0.4 and 0.4 respectively for optimal ratio via performing ablations." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14737" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.094, + 0.089, + 0.907, + 0.23 + ], + "angle": 0, + "content": "
ModelPublicationObject Semantic SegmentationLand-Cover Semantic Segmentation
iSAID [33]SOTA [34]FAST [34]SIOR [34]Potsdam [82]OEM [30]LoveDA [32]VaihingenFloodNet [83]
Few-Shot Semantic Segmentation
SegGPT [84]ICCV 202329.1731.4119.5537.4730.1225.1931.4047.2733.81
DINOv [85]CVPR 202425.3518.247.9225.0437.6523.0525.1735.1044.55
Zero-Shot Semantic Segmentation
SAN [80]CVPR 202311.7710.733.1615.2251.4524.8727.8346.2339.26
CAT-SEG [75]CVPR 202420.5813.713.1319.6450.8130.2637.1549.6241.36
SegEarth-OV [58]Arxiv 202421.7011.992.8413.9247.1040.3036.9029.1034.00
SkySense-O-43.9232.128.3330.8954.1040.8338.3051.6442.04
-(+22.22)(+18.41)(+5.2)(+11.25)(+2.65)(+0.53)(+1.15)(+2.02)(+0.68)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.234, + 0.907, + 0.264 + ], + "angle": 0, + "content": "Table 2. Zero-shot remote sensing object and land-cover semantic segmentation performances. Bold indicates the highest performance, and underline indicates the second highest performance in mIoU (%)." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.27, + 0.558, + 0.376 + ], + "angle": 0, + "content": "
ModelPublicationSIRI-WHU [86]AID-VQA [28]RSVQA-HR [14]
Acc.Acc.Avg. Acc.
GeoChat [35]CVPR 202466.6372.0370.82
LHRS-Bot [87]ECCV 202462.6691.26-
EarthGPT [88]TGRS 2024--71.15
VHM [89]Arxiv 202470.8891.7073.75
SkySense-O†-72.45 (+1.57)92.95 (+1.25)75.39 (+1.64)
SkySense-O-74.79 (+3.91)94.10 (+2.40)78.09 (+4.34)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.38, + 0.558, + 0.41 + ], + "angle": 0, + "content": "Table 3. Zero-Shot visual question-answering performances, compared with existing MLLMs. \\(\\dagger\\) indicates wo/ pixel-level fine-tuning by Sky-SA." + }, + { + "type": "table", + "bbox": [ + 0.568, + 0.27, + 0.902, + 0.375 + ], + "angle": 0, + "content": "
ModelPublicationRESISC [29]AID [28]
CLIP [57]PMLR 202169.2566.70
RemoteCLIP [74]TGRS 202470.8574.3
CLIP-laion-RS [31]-71.7071.21
SkyScript [31]AAA1 202471.7070.94
GeoRSCLIP [24]TGRS 202373.9076.27
SkySense-O†-83.2882.93
-(+9.38)(+6.66)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.564, + 0.379, + 0.902, + 0.409 + ], + "angle": 0, + "content": "Table 4. Zero-shot remote sensing scene classification performances in Top-1 accuracy \\((\\%)\\)" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.427, + 0.216, + 0.444 + ], + "angle": 0, + "content": "5. Experiment" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.453, + 0.313, + 0.467 + ], + "angle": 0, + "content": "5.1. Evaluation Benchmarks" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.475, + 0.483, + 0.566 + ], + "angle": 0, + "content": "We evaluate SkySense-O on four representative RS interpretation tasks. To ensure fair zero-shot comparisons, all experiments are conducted without extra fine-tuning and labels from the evaluation datasets are not used during training. The supplementary materials provide further details on the implementation and visualization results." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.567, + 0.483, + 0.702 + ], + "angle": 0, + "content": "Zero-shot object segmentation. We adopt iSAID [33], FAST [34], SIOR [34], SOTA [34] for the object segmentation experiment. They are chosen considering diverse scenes and categories. We report mIoU results on official validation sets. As depicted in Tab. 2, SkySense-O has achieved the SOTA performance on all object segmentation datasets. On average, it outperforms the previous SOTAs by an average margin of \\(14.27\\%\\). Furthermore, it even surpasses few-shot SOTAs on most datasets." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.704, + 0.483, + 0.84 + ], + "angle": 0, + "content": "Zero-shot land-cover segmentation. We assess SkySense-O's land-cover segmentation performance on ISPRS Potsdam [82], Vaihingen and LoveDA [32] datasets. SkySense-O excels on all three datasets (Tab. 2). Notably, we surpass the second best CAT-SEG [75] by \\(1.41\\%\\) mIoU and have achieved the best performance on the Vaihingen leaderboard. More importantly, our results demonstrate the potential to achieve more accurate interpretations than the ground truth (the second row in Fig. 9)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Zero-shot classification. We utilize two scene classification datasets: AID [28] and RESISC-45 [29]. The validation setting follows GeoRSCLIP [24] and Top-1 Accuracy is reported. SkySense-O overall outperforms competitive" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.429, + 0.907, + 0.589 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.594, + 0.907, + 0.622 + ], + "angle": 0, + "content": "Figure 9. Visualization of object segmentation results on iSAID dataset and land-cover segmentation results on Potsdam dataset." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.639, + 0.907, + 0.67 + ], + "angle": 0, + "content": "basielines, achieving \\(9.38\\%\\) and \\(6.66\\%\\) higher performance than state-of-the-art methods on the two datasets (Tab. 4)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.674, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Zero-shot RS VQA. To further validate the versatility of SkySense-O, we employ it in MLLM as a vision encoder for experimentation. Specifically, we maintain the same architecture as LLaVA-1.5 [90], except that we replace the original CLIP-L-14 [57] with our SkySense-O, and set the image resolution to \\(512 \\times 512\\) pixels. Regarding the setup of the instruction tuning data, during the pretraining stage, we utilize CC-558K dataset [60] and 121k image-text pairs converted from ChatearthNet [91] dataset. In the finetuning stage, based on instruction-tuning datasets collected by SkySenseGPT [61] from different sources, we select a subset of 593k data as our downstream instruction finetuning dataset, where details are in the supplementary. We evaluate zero-shot VQA performances on several public RS benchmarks as shown in Tab. 3." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14738" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.09, + 0.905, + 0.325 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.338, + 0.908, + 0.381 + ], + "angle": 0, + "content": "Figure 10. Visualization of text-image alignment in CLIP. (a) Computing the similarity between each text token and all image tokens to obtain text-image relevance (TIR) maps for different categories. (b) Visualizing TIR maps for some high-response texts. (c) Histogram of the maximum response values of TIR maps across the text list. The details of input texts can be referred to the supplementary materials." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.391, + 0.487, + 0.476 + ], + "angle": 0, + "content": "
Foundation \nModelVision \nEncoderVision-Enc. \nParameterNum. of \nImage-textRESISC \n(Acc)
RemoteCLIPViT-L-14304M0.83M70.85
SkyScriptViT-L-14304M2.6M71.70
GeoRSCLIPViT-H-14630M5M73.90
OursSwinV2-H-8654M10M83.28
" + }, + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.481, + 0.483, + 0.523 + ], + "angle": 0, + "content": "Table 5. Comparison with existing RS vision-language foundation models. SkySense-O achieves consistent performance gains due to scaling laws in terms of model-size and data-size." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.535, + 0.477, + 0.635 + ], + "angle": 0, + "content": "
+Sky-SA+Graph+Pretrain-ing+LvcmIoU
18.40
✓27.76 (+9.36)
✓✓36.28 (+8.52)
✓✓✓41.69 (+5.41)
✓✓✓✓43.92 (+2.24)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.106, + 0.639, + 0.466, + 0.654 + ], + "angle": 0, + "content": "Table 6. Ablation study on the proposed dataset and designs." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.675, + 0.313, + 0.691 + ], + "angle": 0, + "content": "5.2. Implementation Details." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.698, + 0.484, + 0.834 + ], + "angle": 0, + "content": "In the pre-training stage, we use 80 A100-80GB GPUs and primarily adopt two large-scale remote sensing image-text datasets, SkyScript [31] and RS5M [24], and further expand their captions using Qwen-VL [92] and InternVL [93], comprising 10 million image-text pairs. In the fine-tuning stage, we employ OSM [25] for local image-text fine-tuning and Sky-SA for fine-grained interpretation capabilities on 8 A100-80GB GPUs as shown in Tab. 5. More details are included in the supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.848, + 0.367, + 0.863 + ], + "angle": 0, + "content": "5.3. Discussions & Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Effectiveness of Image-Language Alignment. In Fig. 10, we follow MaskCLIP [94] and adopt the image-text correla" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.391, + 0.906, + 0.545 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.555, + 0.907, + 0.584 + ], + "angle": 0, + "content": "Figure 11. Visualizations of open-world interpretation results by SkySense-O on iSAID, Potsdam, FAST, and SIOR datasets." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.61, + 0.907, + 0.79 + ], + "angle": 0, + "content": "tion in the representation space to evaluate the effectiveness of vision-language foundation models. We select the CLIP models from SegEarth-OV [58] and CAT-Seg [75] for comparison. The former is a standard CLIP model, while the latter is a locally fine-tuned CLIP model by COCO-Stuff [19]. We randomly selected 131 category texts from Sky-SA as inputs. As shown in Figure 10(c), SkySense-O exhibits larger response variance across different categories, indicating stronger discriminative capabilities among different texts. In Fig. 10(b), we present several image-text correlation maps to further validate the effectiveness of SkySense-O in image-text alignment in open-world scenes." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.793, + 0.905, + 0.838 + ], + "angle": 0, + "content": "- vehicle vs. parking lot. (small object) Because our modeling is vision-centric, small targets are effectively decoupled from the contextual background." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.84, + 0.907, + 0.902 + ], + "angle": 0, + "content": "- building vs. industrial park. (hypernym) SkySense-O demonstrates proper alignment capabilities across different levels of semantic hierarchy due to large-scale remote sensing knowledge are learned in training." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.793, + 0.907, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "14739" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.089, + 0.483, + 0.251 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.261, + 0.485, + 0.361 + ], + "angle": 0, + "content": "Figure 12. The visualization of vision-centric graph of the Sky-SA dataset. Node layout is performed via Force Atlas [95] and graph communities are detected by the Leiden algorithm [96]. Circles represent entity nodes with size proportional to their degree. Node colors represent entity communities. Five main communities are detected by an unsupervised manner and they just correspond to the major land-cover categories in the remote sensing imagery." + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.374, + 0.223, + 0.475 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.132, + 0.477, + 0.18, + 0.489 + ], + "angle": 0, + "content": "(a) Image" + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.374, + 0.353, + 0.474 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.233, + 0.477, + 0.34, + 0.489 + ], + "angle": 0, + "content": "(b) Wo/ vision-centric" + }, + { + "type": "image", + "bbox": [ + 0.353, + 0.374, + 0.482, + 0.475 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.366, + 0.477, + 0.468, + 0.489 + ], + "angle": 0, + "content": "(c) W/ vision-centric" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.495, + 0.484, + 0.538 + ], + "angle": 0, + "content": "Figure 13. Comparisons of the results without and with vision-centric regularization loss. Compared to (b), the vehicles in (c) are clearly distinguished from the contextual floor." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.554, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Design of Vision-Centric Strategy. Tab. 6 presents the ablation study to assess our Sky-SA dataset and vision-centric designs, namely vision-centric graph sampler, pretraining foundation model, and vision-centric regularization loss. We utilize the iSAID dataset as evaluation benchmark and report mIoU metric on the official test set. Initially, we use the standard CLIP model [57] and fine-tune it on the COCO-Stuff dataset [19]. Compared with the baseline, the Sky-SA dataset leads to a notable improvement compared to COCO-Stuff dataset. Nonetheless, we believe that the Sky-SA dataset holds considerable untapped potential, which might be constrained by the intricate entanglement of open-world text labels as demonstrated in Sec. 4.1. To relieve this problem, we construct a vision-centric graph for all text labels of the Sky-SA dataset as shown in Fig. 12. It is intriguing that the nodes in the graph naturally exhibit proper unsupervised clustering results by community detection [96], without using any text labels. This further validates the rationale for vision-centric modeling and the introducing of vision-centric graph leads to \\(8.52\\%\\) mIoU improvement. Then, we replace the standard CLIP with our pre-trained model incorporating extensive remote sensing knowledge. This effectively improves the perfor" + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.099, + 0.907, + 0.281 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.29, + 0.907, + 0.361 + ], + "angle": 0, + "content": "Figure 14. Zero-shot semantic segmentation results of Sky-SA and existing datasets. The vertical axis represents the training set, while the horizontal axis represents the test set. The yellow box denotes the result is fine-tuned by its own training set (i.e. supervised learning) and red box denotes the average performance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.376, + 0.907, + 0.497 + ], + "angle": 0, + "content": "mance to \\(41.69\\%\\) mIoU, validating the necessity of remote sensing pre-training model. Finally, we introduce vision-centric regularization loss, which uncovers and prioritizes hard samples in visual-language modeling. The results show that the constraint pushes the performance to a very strong \\(43.92\\%\\) mIoU. Furthermore, as shown in Fig. 13, this design enhances the discrimination ability, especially between high-context categories." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.498, + 0.909, + 0.651 + ], + "angle": 0, + "content": "Significance of Sky-SA for Open-World Interpretation. In Fig. 14, we replace Sky-SA with existing datasets and conduct experiments under the same settings as in Tab. 2. It can be observed that existing datasets exhibit poor generalization to other datasets, except on their own test sets. In contrast, the Sky-SA dataset achieves superior zero-shot segmentation performance across all datasets, outperforming the second best dataset by an average of over \\(13.43\\%\\) mIoU, which demonstrates the significant potential of Sky-SA for open-world interpretation." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.664, + 0.634, + 0.68 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.909, + 0.903 + ], + "angle": 0, + "content": "In this paper, we present Skysense-O, a large-scale and general remote sensing interpretation model for open remote sensing images. Skysense-O is mainly driven by the proposed semantically-rich and fine-grained dataset SkySA and the vision-centric visual-language modeling, consistently outperforming other models on a variety of tasks, showcasing its exceptional generalization ability and strong performance. We hope this work will constructively push the RS open-world interpretation on constructing large-scale vision-language datasets and models for further researches. Our future work aims to explore fine-grained remote sensing interpretation, such as analyzing object states and relationships, further enhancing the flexibility of open-world applications." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14740" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.251, + 0.108 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.484, + 0.207 + ], + "angle": 0, + "content": "This work was supported by the Anhui Provincial Natural Science Foundation under Grant 2108085UD12 and the National Natural Science Key Foundation of China under Grant 62036005. We acknowledge the support of GPU cluster built by MCC Lab of Information Science and Technology Institution, USTC." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.221, + 0.188, + 0.237 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.246, + 0.483, + 0.288 + ], + "angle": 0, + "content": "[1] Xianping Ma, Xiaokang Zhang, Man-On Pun, and Bo Huang. Manet: Fine-tuning segment anything model for multimodal remote sensing semantic segmentation, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.29, + 0.483, + 0.344 + ], + "angle": 0, + "content": "[2] Aysim Token, Marvin Eisenberger, Daniel Cremers, and Laura Leal-Taixe. Satsynth: Augmenting image-mask pairs through diffusion models for aerial semantic segmentation, CVPR 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.346, + 0.483, + 0.403 + ], + "angle": 0, + "content": "[3] Deyi Ji, Feng Zhao, Hongtao Lu, Feng Wu, and Jieping Ye. Structural and statistical texture knowledge distillation and learning for segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence, pages 1-18, 2025." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.403, + 0.483, + 0.473 + ], + "angle": 0, + "content": "[4] Deyi Ji, Feng Zhao, Hongtao Lu, Mingyuan Tao, and Jieping Ye. Ultra-high resolution segmentation with ultra-rich context: A novel benchmark. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23621-23630, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.474, + 0.482, + 0.53 + ], + "angle": 0, + "content": "[5] Deyi Ji, Feng Zhao, Lanyun Zhu, Wenwei Jin, Hongtao Lu, and Jieping Ye. Discrete latent perspective learning for segmentation and detection. In *Forty-first International Conference on Machine Learning*, pages 21719-21730, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.531, + 0.483, + 0.587 + ], + "angle": 0, + "content": "[6] Botao Ren, Xue Yang, Yi Yu, Junwei Luo, and Zhidong Deng. Pointobb-v2: Towards simpler, faster, and stronger single point supervised oriented object detection. arXiv preprint arXiv:2410.08210, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.589, + 0.483, + 0.63 + ], + "angle": 0, + "content": "[7] Kang Wu, Zunxiao Xu, Xinrong Lyu, and Peng Ren. Cloud detection with boundary nets. ISPRS Journal of Photogrammetry and Remote Sensing, 186:218-231, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.632, + 0.483, + 0.701 + ], + "angle": 0, + "content": "[8] Peiyuan Zhang, Junwei Luo, Xue Yang, Yi Yu, Qingyun Li, Yue Zhou, Xiaosong Jia, Xudong Lu, Jingdong Chen, Xiang Li, et al. Pointobb-v3: Expanding performance boundaries of single point-supervised oriented object detection. arXiv preprint arXiv:2501.13898, 2025." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.702, + 0.483, + 0.758 + ], + "angle": 0, + "content": "[9] Yi Yu, Xue Yang, Qingyun Li, Yue Zhou, Feipeng Da, and Junchi Yan. H2rbox-v2: Incorporating symmetry for boosting horizontal box supervised oriented object detection. Advances in Neural Information Processing Systems, 36, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.759, + 0.483, + 0.801 + ], + "angle": 0, + "content": "[10] Kang Wu, Zunxiao Xu, Xinrong Lyu, and Peng Ren. Crosssupervised learning for cloud detection. *GIScience & Remote Sensing*, 60(1):2147298, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.802, + 0.483, + 0.871 + ], + "angle": 0, + "content": "[11] Zhuo Zheng, Ailong Ma, Liangpei Zhang, and Yanfei Zhong. Change is everywhere: Single-temporal supervised object change detection in remote sensing imagery. In Proceedings of the IEEE/CVF international conference on computer vision, pages 15193-15202, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.483, + 0.902 + ], + "angle": 0, + "content": "[12] Deyi Ji, Siqi Gao, Mingyuan Tao, Hongtao Lu, and Feng Zhao. Changenet: Multi-temporal asymmetric change detect" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.246, + 0.483, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "tion dataset. In ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing, pages 2725-2729. IEEE, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.135, + 0.905, + 0.203 + ], + "angle": 0, + "content": "[13] Junwei Luo, Yingying Zhang, Xue Yang, Kang Wu, Qi Zhu, Lei Liang, Jingdong Chen, and Yansheng Li. When large vision-language model meets large remote sensing imagery: Coarse-to-fine text-guided token pruning. arXiv preprint arXiv:2503.07588, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.204, + 0.905, + 0.259 + ], + "angle": 0, + "content": "[14] Sylvain Lobry, Diego Marcos, Jesse Murray, and Devis Tuia. Rsvqa: Visual question answering for remote sensing data. IEEE Transactions on Geoscience and Remote Sensing, 58(12):8555-8566, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.26, + 0.905, + 0.329 + ], + "angle": 0, + "content": "[15] Junjue Wang, Zhuo Zheng, Zihang Chen, Ailong Ma, and Yanfei Zhong. Earthvqa: Towards queryable earth via relational reasoning-based remote sensing visual question answering. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 5481-5489, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.33, + 0.905, + 0.399 + ], + "angle": 0, + "content": "[16] Yan Li, Weiwei Guo, Xue Yang, Ning Liao, Dunyun He, Jiaqi Zhou, and Wenxian Yu. Toward open vocabulary aerial object detection with clip-activated student-teacher learning. In European Conference on Computer Vision, pages 431–448. Springer, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.4, + 0.905, + 0.468 + ], + "angle": 0, + "content": "[17] Jiancheng Pan, Yanxing Liu, Yuqian Fu, Muyuan Ma, Jiaohao Li, Danda Pani Paudel, Luc Van Gool, and Xiaomeng Huang. Locate anything on earth: Advancing open-vocabulary object detection for remote sensing community. arXiv preprint arXiv:2408.09110, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.469, + 0.905, + 0.538 + ], + "angle": 0, + "content": "[18] Weiyun Wang, Min Shi, Qingyun Li, Wenhai Wang, Zhenhang Huang, Linjie Xing, Zhe Chen, Hao Li, Xizhou Zhu, Zhiguo Cao, et al. The all-seeing project: Towards panoptic visual recognition and understanding of the open world. arXiv preprint arXiv:2308.01907, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.539, + 0.905, + 0.594 + ], + "angle": 0, + "content": "[19] Holger Caesar, Jasper Uijlings, and Vittorio Ferrari. Cocostuff: Thing and stuff classes in context. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1209-1218, 2018. 2, 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.595, + 0.904, + 0.622 + ], + "angle": 0, + "content": "[20] Shuyang Gu. Several questions of visual generation in 2024. arXiv preprint arXiv:2407.18290, 2024. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.623, + 0.905, + 0.705 + ], + "angle": 0, + "content": "[21] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dollar, and Ross Girshick. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, October 2023. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.706, + 0.905, + 0.775 + ], + "angle": 0, + "content": "[22] Shengbang Tong, Zhuang Liu, Yuexiang Zhai, Yi Ma, Yann LeCun, and Saining Xie. Eyes wide shut? exploring the visual shortcomings of multimodal llms. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9568-9578, 2024. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.776, + 0.905, + 0.873 + ], + "angle": 0, + "content": "[23] Xin Guo, Jiangwei Lao, Bo Dang, Yingying Zhang, Lei Yu, Lixiang Ru, Liheng Zhong, Ziyuan Huang, Kang Wu, Dingxiang Hu, et al. Skysense: A multi-modal remote sensing foundation model towards universal interpretation for earth observation imagery. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 27672-27683, 2024. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.902 + ], + "angle": 0, + "content": "[24] Zilun Zhang, Tiancheng Zhao, Yulong Guo, and Jianwei Yin. Rs5m and georsclip: A large scale vision-language dataset" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "14741" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.484, + 0.134 + ], + "angle": 0, + "content": "and a large vision-language model for remote sensing. IEEE Transactions on Geoscience and Remote Sensing, pages 1-1, 2024. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.135, + 0.484, + 0.203 + ], + "angle": 0, + "content": "[25] Haozhe Xie, Zhaoxi Chen, Fangzhou Hong, and Ziwei Liu. CityDreamer: Compositional generative model of unbounded 3D cities. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.204, + 0.484, + 0.26 + ], + "angle": 0, + "content": "[26] Amin Karimi Monsefi, Kishore Prakash Sailaja, Ali Alilooee, Ser-Nam Lim, and Rajiv Ramnath. Detailclip: Detail-oriented clip for fine-grained tasks. arXiv preprint arXiv:2409.06809, 2024. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.261, + 0.408, + 0.274 + ], + "angle": 0, + "content": "[27] OpenAI. Gpt-4v(ision) system card. 2023. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.275, + 0.484, + 0.344 + ], + "angle": 0, + "content": "[28] Gui-Song Xia, Jingwen Hu, Fan Hu, Baoguang Shi, Xiang Bai, Yanfei Zhong, Liangpei Zhang, and Xiaoqiang Lu. Aid: A benchmark data set for performance evaluation of aerial scene classification. IEEE Transactions on Geoscience and Remote Sensing, 55(7):3965-3981, 2017. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.345, + 0.483, + 0.398 + ], + "angle": 0, + "content": "[29] Gong Cheng, Junwei Han, and Xiaogiang Lu. Remote sensing image scene classification: Benchmark and state of the art. Proceedings of the IEEE, 105(10):1865-1883, 2017. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.4, + 0.484, + 0.469 + ], + "angle": 0, + "content": "[30] Junshi Xia, Naoto Yokoya, Bruno Adriano, and Clifford Broni-Bediako. Openearthmap: A benchmark dataset for global high-resolution land cover mapping. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 6254-6264, 2023. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.47, + 0.483, + 0.539 + ], + "angle": 0, + "content": "[31] Zhecheng Wang, Rajanie Prabha, Tianyuan Huang, Jiajun Wu, and Ram Rajagopal. Skyscript: A large and semantically diverse vision-language dataset for remote sensing. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 5805-5813, 2024. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.54, + 0.483, + 0.594 + ], + "angle": 0, + "content": "[32] Junjue Wang, Zhuo Zheng, Ailong Ma, Xiaoyan Lu, and Yanfei Zhong. LoveDA: A remote sensing land-cover dataset for domain adaptive semantic segmentation, October 2021. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.595, + 0.483, + 0.691 + ], + "angle": 0, + "content": "[33] Syed Waqas Zamir, Aditya Arora, Akshita Gupta, Salman Khan, Guolei Sun, Fahad Shahbaz Khan, Fan Zhu, Ling Shao, Gui-Song Xia, and Xiang Bai. isaid: A large-scale dataset for instance segmentation in aerial images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 28-37, 2019. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.692, + 0.483, + 0.747 + ], + "angle": 0, + "content": "[34] Di Wang, Jing Zhang, Bo Du, Minqiang Xu, Lin Liu, Dacheng Tao, and Liangpei Zhang. Samrs: Scaling-up remote sensing segmentation dataset with segment anything model, NeurIPS 2023. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.748, + 0.483, + 0.83 + ], + "angle": 0, + "content": "[35] Kartik Kuckreja, Muhammad Sohail Danish, Muzammal Naseer, Abhijit Das, Salman Khan, and Fahad Shahbaz Khan. Gechat: Grounded large vision-language model for remote sensing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 27831-27840, 2024. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.483, + 0.9 + ], + "angle": 0, + "content": "[36] Tianrun Chen, Chaotao Ding, Lanyun Zhu, Tao Xu, Deyi Ji, Ying Zang, and Zejian Li. xlstm-unet can be an effective 2d\\& 3d medical image segmentation backbone with vision-lstm (vil) better than its mamba counterpart. arXiv preprint arXiv:2407.01530, 2024." + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.148 + ], + "angle": 0, + "content": "[37] Deyi Ji, Siqi Gao, Lanyun Zhu, Qi Zhu, Yiru Zhao, Peng Xu, Hongtao Lu, Feng Zhao, and Jieping Ye. View-centric multi-object tracking with homographic matching in moving UAV. arXiv preprint arXiv:2403.10830, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.15, + 0.905, + 0.219 + ], + "angle": 0, + "content": "[38] Ying Zang, Runlong Cao, Jianqi Zhang, Yidong Han, Ziyue Cao, Wenjun Hu, Didi Zhu, Lanyun Zhu, Zejian Li, Deyi Ji, et al. Let human sketches help: Empowering challenging image segmentation task with freehand sketches. arXiv preprint arXiv:2501.19329, 2025." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.221, + 0.905, + 0.276 + ], + "angle": 0, + "content": "[39] Lanyun Zhu, Tianrun Chen, Deyi Ji, Jieping Ye, and Jun Liu. Not every patch is needed: Towards a more efficient and effective backbone for video-based person re-identification. IEEE Transactions on Image Processing, 2025." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.277, + 0.905, + 0.332 + ], + "angle": 0, + "content": "[40] Deyi Ji, Hongtao Lu, and Tongzhen Zhang. End to end multiscale convolutional neural network for crowd counting. In Eleventh international conference on machine vision, volume 11041, pages 761-766. SPIE, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.334, + 0.905, + 0.403 + ], + "angle": 0, + "content": "[41] Zunxiao Xu, Kang Wu, Lei Huang, Qimao Wang, and Peng Ren. Cloudy image arithmetic: A cloudy scene synthesis paradigm with an application to deep-learning-based thin cloud removal. IEEE Transactions on Geoscience and Remote Sensing, 60:1-16, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.405, + 0.905, + 0.473 + ], + "angle": 0, + "content": "[42] Ying Zang, Yuncan Gao, Jiangi Zhang, Yuangi Hu, Runlong Cao, Lanyun Zhu, Qi Zhu, Deyi Ji, Renjun Xu, and Tianrun Chen. Breaking the box: Enhancing remote sensing image segmentation with freehand sketches. arXiv preprint arXiv:2503.12191, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.476, + 0.905, + 0.531 + ], + "angle": 0, + "content": "[43] Zhihao Li, Biao Hou, Siteng Ma, Zitong Wu, Xianpeng Guo, Bo Ren, and Licheng Jiao. Masked angle-aware autoencoder for remote sensing images, European Conference on Computer Vision 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.533, + 0.905, + 0.588 + ], + "angle": 0, + "content": "[44] Aaron E Maxwell, Timothy A Warner, and Fang Fang. Implementation of machine-learning classification in remote sensing: An applied review. International journal of remote sensing, 39(9):2784-2817, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.59, + 0.905, + 0.657 + ], + "angle": 0, + "content": "[45] Xue Yang, Gefan Zhang, Wentong Li, Yue Zhou, Xuehui Wang, and Junchi Yan. H2rbox: Horizontal box annotation is all you need for oriented object detection. In The Eleventh International Conference on Learning Representations, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.66, + 0.905, + 0.729 + ], + "angle": 0, + "content": "[46] Yansheng Li, Junwei Luo, Yongjun Zhang, Yihua Tan, JinGang Yu, and Song Bai. Learning to holistically detect bridges from large-size vhr remote sensing imagery. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.731, + 0.905, + 0.786 + ], + "angle": 0, + "content": "[47] Haoran Wang, Licheng Jiao, Fang Liu, Lingling Li, Xu Liu, Deyi Ji, and Weihao Gan. Ipgn: Interactiveness proposal graph network for human-object interaction detection. IEEE Transactions on Image Processing, 30:6583-6593, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.788, + 0.907, + 0.858 + ], + "angle": 0, + "content": "[48] Junwei Luo, Xue Yang, Yi Yu, Qingyun Li, Junchi Yan, and Yansheng Li. Pointobb: Learning oriented object detection via single point supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16730-16740, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.859, + 0.905, + 0.902 + ], + "angle": 0, + "content": "[49] Yi Yu, Xue Yang, Qingyun Li, Feipeng Da, Jifeng Dai, Yu Qiao, and Junchi Yan. Point2rbox: Combine knowledge from synthetic visual patterns for end-to-end oriented object" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.907, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14742" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.134 + ], + "angle": 0, + "content": "detection with single point supervision. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 16783-16793, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.483, + 0.205 + ], + "angle": 0, + "content": "[50] Lanyun Zhu, Deyi Ji, Shiping Zhu, Weihao Gan, Wei Wu, and Junjie Yan. Learning statistical texture for semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12537-12546, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.207, + 0.483, + 0.29 + ], + "angle": 0, + "content": "[51] Tianrun Chen, Ankang Lu, Lanyun Zhu, Chaotao Ding, Chunan Yu, Deyi Ji, Zejian Li, Lingyun Sun, Papa Mao, and Ying Zang. Sam2-adapter: Evaluating & adapting segment anything 2 in downstream tasks: Camouflage, shadow, medical image segmentation, and more. arXiv preprint arXiv:2408.04579, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.291, + 0.482, + 0.374 + ], + "angle": 0, + "content": "[52] Gui-Song Xia, Xiang Bai, Jian Ding, Zhen Zhu, Serge Belongie, Jiebo Luo, Mihai Datcu, Marcello Pelillo, and Liangpei Zhang. Dota: A large-scale dataset for object detection in aerial images. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3974-3983, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.376, + 0.482, + 0.446 + ], + "angle": 0, + "content": "[53] Deyi Ji, Haoran Wang, Mingyuan Tao, Jianqiang Huang, Xian-Sheng Hua, and Hongtao Lu. Structural and statistical texture knowledge distillation for semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16876-16885, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.447, + 0.482, + 0.503 + ], + "angle": 0, + "content": "[54] Deyi Ji, Feng Zhao, and Hongtao Lu. Guided patch-grouping wavelet transformer with spatial congruence for ultra-high resolution segmentation. International Joint Conference on Artificial Intelligence, pages 920-928, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.504, + 0.482, + 0.56 + ], + "angle": 0, + "content": "[55] Deyi Ji, Wenwei Jin, Hongtao Lu, and Feng Zhao. Pptformer: Pseudo multi-perspective transformer for uav segmentation. International Joint Conference on Artificial Intelligence, pages 893-901, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.561, + 0.482, + 0.618 + ], + "angle": 0, + "content": "[56] Hanzhe Hu, Deyi Ji, Weihao Gan, Shuai Bai, Wei Wu, and Junjie Yan. Class-wise dynamic graph convolution for semantic segmentation. In European Conference on Computer Vision, pages 1-17. Springer, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.619, + 0.482, + 0.702 + ], + "angle": 0, + "content": "[57] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763, 2021. 3, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.703, + 0.482, + 0.759 + ], + "angle": 0, + "content": "[58] Kaiyu Li, Ruixun Liu, Xiangyong Cao, Deyu Meng, and Zhi Wang. Segearth-ov: Towards training-free open-vocabulary segmentation for remote sensing images. arXiv preprint arXiv:2410.01768, 2024. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.76, + 0.482, + 0.829 + ], + "angle": 0, + "content": "[59] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.831, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[60] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In A. Oh, T. Naumann, A. Globerson, K. Saenko, M. Hardt, and S. Levine, editors, Advances in Neural Information Processing Systems, volume 36, pages 34892-34916, 2023. 3, 6" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.175 + ], + "angle": 0, + "content": "[61] Junwei Luo, Zhen Pang, Yongjun Zhang, Tingzhu Wang, Linlin Wang, Bo Dang, Jiangwei Lao, Jian Wang, Jingdong Chen, Yihua Tan, et al. Skysensegpt: A fine-grained instruction tuning dataset and model for remote sensing vision-language understanding. arXiv preprint arXiv:2406.10100, 2024.3,6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.178, + 0.905, + 0.261 + ], + "angle": 0, + "content": "[62] Tianrun Chen, Chunan Yu, Jing Li, Jianqi Zhang, Lanyun Zhu, Deyi Ji, Yong Zhang, Ying Zang, Zejian Li, and Lingyun Sun. Reasoning3d-grounding and reasoning in 3d: Fine-grained zero-shot open-vocabulary 3d reasoning part segmentation via large vision-language models. arXiv preprint arXiv:2405.19326, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.262, + 0.905, + 0.317 + ], + "angle": 0, + "content": "[63] Ziyue Huang, Yongchao Feng, Qingjie Liu, and Yunhong Wang. Mutdet: Mutually optimizing pre-training for remote sensing object detection, European Conference on Computer Vision 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.319, + 0.905, + 0.374 + ], + "angle": 0, + "content": "[64] Zhuo Zheng, Yanfei Zhong, Ailong Ma, and Liangpei Zhang. Single-temporal supervised learning for universal remote sensing change detection. International Journal of Computer Vision, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.376, + 0.905, + 0.432 + ], + "angle": 0, + "content": "[65] Deyi Ji, Lanyun Zhu, Siqi Gao, Peng Xu, Hongtao Lu, Jieping Ye, and Feng Zhao. Tree-of-table: Unleashing the power of llms for enhanced large-scale table understanding. arXiv preprint arXiv:2411.08516, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.434, + 0.905, + 0.502 + ], + "angle": 0, + "content": "[66] Lanyun Zhu, Tianrun Chen, Deyi Ji, Jieping Ye, and Jun Liu. Llafs: When large language models meet few-shot segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3065-3075, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.504, + 0.905, + 0.56 + ], + "angle": 0, + "content": "[67] Lanyun Zhu, Deyi Ji, Tianrun Chen, Peng Xu, Jieping Ye, and Jun Liu. Ibid: Alleviating hallucinations in large vision-language models via image-biased decoding. arXiv preprint arXiv:2402.18476, 2024.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.561, + 0.905, + 0.616 + ], + "angle": 0, + "content": "[68] Shizhen Chang and Pedram Ghamisi. Changes to captions: An attentive network for remote sensing change captioning. IEEE Transactions on Image Processing, 32:6047-6060, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.618, + 0.905, + 0.687 + ], + "angle": 0, + "content": "[69] Deyi Ji, Haoran Wang, Hanzhe Hu, Weihao Gan, Wei Wu, and Junjie Yan. Context-aware graph convolution network for target re-identification. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 1646–1654, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.689, + 0.905, + 0.759 + ], + "angle": 0, + "content": "[70] Haoran Wang, Licheng Jiao, Fang Liu, Lingling Li, Xu Liu, Deyi Ji, and Weihao Gan. Learning social spatio-temporal relation graph in the wild and a video benchmark. IEEE Transactions on Neural Networks and Learning Systems, 34(6):2951-2964, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.76, + 0.905, + 0.829 + ], + "angle": 0, + "content": "[71] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16000-16009, 2022. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.831, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[72] Xian Sun, Peijin Wang, Wanxuan Lu, Zicong Zhu, Xiaonan Lu, Qibin He, Junxi Li, Xuee Rong, Zhujun Yang, Hao Chang, et al. Ringmo: A remote sensing foundation model with masked image modeling. IEEE Transactions on Geoscience and Remote Sensing, 61:1-22, 2022. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "14743" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.189 + ], + "angle": 0, + "content": "[73] Colorado J Reed, Ritwik Gupta, Shufan Li, Sarah Brockman, Christopher Funk, Brian Clipp, Kurt Keutzer, Salvatore Candido, Matt Uytendaele, and Trevor Darrell. Scale-mae: A scale-aware masked autoencoder for multiscale geospatial representation learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4088-4099, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.191, + 0.485, + 0.261 + ], + "angle": 0, + "content": "[74] Fan Liu, Delong Chen, Zhangqingyun Guan, Xiaocong Zhou, Jiale Zhu, Qiaolin Ye, Liyong Fu, and Jun Zhou. Remoteclip: A vision language foundation model for remote sensing. IEEE Transactions on Geoscience and Remote Sensing, 2024. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.262, + 0.484, + 0.344 + ], + "angle": 0, + "content": "[75] Seokju Cho, Heeseong Shin, Sunghwan Hong, Anurag Arnab, Paul Hongsuck Seo, and Seungryong Kim. Catseg: Cost aggregation for open-vocabulary semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4113-4123, 2024. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.346, + 0.484, + 0.415 + ], + "angle": 0, + "content": "[76] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 633-641, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.417, + 0.484, + 0.473 + ], + "angle": 0, + "content": "[77] Bin Xiao, Haiping Wu, Weijian Xu, Xiyang Dai, Houdong Hu, Yumao Lu, Michael Zeng, Ce Liu, and Lu Yuan. Florence-2: Advancing a unified representation for a variety of vision tasks, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.475, + 0.484, + 0.545 + ], + "angle": 0, + "content": "[78] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Qing Jiang, Chunyuan Li, Jianwei Yang, Hang Su, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.546, + 0.484, + 0.586 + ], + "angle": 0, + "content": "[79] Zhuang Liu and Kaiming He. A decade's battle on dataset bias: Are we there yet? arXiv preprint arXiv:2403.08632, 2024.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.588, + 0.484, + 0.657 + ], + "angle": 0, + "content": "[80] Mengde Xu, Zheng Zhang, Fangyun Wei, Han Hu, and Xi-ang Bai. Side adapter network for open-vocabulary semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2945–2954, 2023. 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.659, + 0.484, + 0.714 + ], + "angle": 0, + "content": "[81] Boci Peng, Yun Zhu, Yongchao Liu, Xiaohe Bo, Haizhou Shi, Chuntao Hong, Yan Zhang, and Siliang Tang. Graph retrieval-augmented generation: A survey. arXiv preprint arXiv:2408.08921, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.716, + 0.484, + 0.758 + ], + "angle": 0, + "content": "[82] Jamie Sherrah. Fully convolutional networks for dense semantic labelling of high-resolution aerial imagery. arXiv preprint arXiv:1606.02585, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.759, + 0.484, + 0.815 + ], + "angle": 0, + "content": "[83] Maryam Rahnemoonfar, Tashnim Chowdhury, Argho Sarkar, Debvrat Varshney, Masoud Yari, and RobinR. Murphy. Floodnet: A high resolution aerial imagery dataset for post flood scene understanding. IEEE Access, Dec 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.817, + 0.484, + 0.87 + ], + "angle": 0, + "content": "[84] Xinlong Wang, Xiaosong Zhang, Yue Cao, Wen Wang, Chunhua Shen, and Tiejun Huang. Seggpt: Segmenting everything in context. arXiv preprint arXiv:2304.03284, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.872, + 0.484, + 0.902 + ], + "angle": 0, + "content": "[85] Feng Li, Qing Jiang, Hao Zhang, Tianhe Ren, Shilong Liu, Xueyan Zou, Huaizhe Xu, Hongyang Li, Chunyuan Li, Jian-" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.092, + 0.907, + 0.12 + ], + "angle": 0, + "content": "wei Yang, et al. Visual in-context prompting. arXiv preprint arXiv:2311.13601, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.907, + 0.19 + ], + "angle": 0, + "content": "[86] Qiqi Zhu, Yanfei Zhong, Bei Zhao, Gui-Song Xia, and Liangpei Zhang. Bag-of-visual-words scene classifier with local and global features for high spatial resolution remote sensing imagery. IEEE Geoscience and Remote Sensing Letters, 13(6):747-751, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.192, + 0.907, + 0.248 + ], + "angle": 0, + "content": "[87] Dilxat Muhtar, Zhenshi Li, Feng Gu, Xueliang Zhang, and Pengfeng Xiao. Lhrs-bot: Empowering remote sensing with vgi-enhanced large multimodal language model. arXiv preprint arXiv:2402.02544, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.249, + 0.907, + 0.319 + ], + "angle": 0, + "content": "[88] Wei Zhang, Miaoxin Cai, Tong Zhang, Yin Zhuang, and Xuerui Mao. Earthgpt: A universal multi-modal large language model for multi-sensor image comprehension in remote sensing domain. IEEE Transactions on Geoscience and Remote Sensing, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.32, + 0.907, + 0.375 + ], + "angle": 0, + "content": "[89] Chao Pang, Xingxing Weng, Jiang Wu, Jiayu Li, Yi Liu, Ji-axing Sun, Weijia Li, Shuai Wang, Litong Feng, Gui-Song Xia, and Conghui He. Vhm: Versatile and honest vision language model for remote sensing image analysis, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.376, + 0.906, + 0.432 + ], + "angle": 0, + "content": "[90] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26296-26306, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.433, + 0.906, + 0.488 + ], + "angle": 0, + "content": "[91] Zhenghang Yuan, Zhitong Xiong, Lichao Mou, and Xiao Xiang Zhu. Chatearthnet: a global-scale image-text dataset empowering vision-language geo-foundation models. Earth System Science Data Discussions, 2024:1-24, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.489, + 0.906, + 0.545 + ], + "angle": 0, + "content": "[92] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.546, + 0.906, + 0.629 + ], + "angle": 0, + "content": "[93] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24185–24198, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.63, + 0.906, + 0.726 + ], + "angle": 0, + "content": "[94] Xiaoyi Dong, Jianmin Bao, Yinglin Zheng, Ting Zhang, Dongdong Chen, Hao Yang, Ming Zeng, Weiming Zhang, Lu Yuan, Dong Chen, et al. Maskclip: Masked self-distillation advances contrastive language-image pretraining. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10995-11005, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.728, + 0.906, + 0.784 + ], + "angle": 0, + "content": "[95] Mathieu Jacomy, Tommaso Venturini, Sebastien Heymann, and Mathieu Bastian. Forceatlas2, a continuous graph layout algorithm for handy network visualization designed for the gephi software. PLoS ONE, page e98679, Jun 2014. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.785, + 0.906, + 0.827 + ], + "angle": 0, + "content": "[96] V. A. Traag, L. Waltman, and N. J. van Eck. From louvain to leiden: guaranteeing well-connected communities. Scientific Reports, Mar 2019. 8" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.907, + 0.827 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "14744" + } + ] +] \ No newline at end of file diff --git a/2025/SkySense-O_ Towards Open-World Remote Sensing Interpretation with Vision-Centric Visual-Language Modeling/2415dd78-f506-4cd3-a12c-42283424590a_origin.pdf b/2025/SkySense-O_ Towards Open-World Remote Sensing Interpretation with Vision-Centric Visual-Language Modeling/2415dd78-f506-4cd3-a12c-42283424590a_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..e8a2c70f71ec184765edb671e386575848124f5f --- /dev/null +++ b/2025/SkySense-O_ Towards Open-World Remote Sensing Interpretation with Vision-Centric Visual-Language Modeling/2415dd78-f506-4cd3-a12c-42283424590a_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ea851a9159a6bcaef373e38d580de30bdbd65c7fe45c708d99e86fd84532a5bb +size 9436037 diff --git a/2025/SkySense-O_ Towards Open-World Remote Sensing Interpretation with Vision-Centric Visual-Language Modeling/full.md b/2025/SkySense-O_ Towards Open-World Remote Sensing Interpretation with Vision-Centric Visual-Language Modeling/full.md new file mode 100644 index 0000000000000000000000000000000000000000..39c4838984ff0ca9b1149700746b140e91041f39 --- /dev/null +++ b/2025/SkySense-O_ Towards Open-World Remote Sensing Interpretation with Vision-Centric Visual-Language Modeling/full.md @@ -0,0 +1,376 @@ +# SkySense-O: Towards Open-World Remote Sensing Interpretation with Vision-Centric Visual-Language Modeling + +Qi Zhu $^{1*}$ , Jiangwei Lao $^{2*}$ , Deyi Ji $^{1}$ , Junwei Luo $^{3}$ , Kang Wu $^{3}$ , Yingying Zhang $^{4}$ , Lixiang Ru $^{3}$ , Jian Wang $^{4}$ , Jingdong Chen $^{5}$ , Ming Yang $^{6}$ , Dong Liu $^{1}$ , Feng Zhao $^{1\dagger}$ + +$^{1}$ MoE Key Laboratory of Brain-inspired Intelligent Perception and Cognition, USTC $^{2}$ Shanghai Jiaotong University $^{3}$ Wuhan University $^{4}$ UCAS $^{5}$ HUST $^{6}$ Northwestern University + +zqcrafts@mail.ustc.edu.cn, fzhao956@ustc.edu.cn + +![](images/ac29e831ce3eef16d6cb0ba8287501cd003005377d7184f90069654262920e5c.jpg) + +![](images/cafd840d83c71333f31303dc6d8fa7782c3ffc484fcfec987c8bf39688dfe7ae.jpg) + +![](images/1f75e40091ff6afa8b62ed975b6f284fb81ceb8b21f8ef3eaa3e48b0ae31a506.jpg) +Figure 1. SkySense-O offers 4 key advantages from the foundation model to the agent application (Left), leading to state-of-the-art zero-shot performances across 14 downstream datasets over 4 distinct RS interpretation tasks (Right) + +![](images/7f370c36aae4a526b9734964d3e354b83499cf4226d877f3b445ec89d6887498.jpg) + +# Abstract + +Open-world interpretation aims to accurately localize and recognize all objects within images by vision-language models (VLMs). While substantial progress has been made in this task for natural images, the advancements for remote sensing (RS) images still remain limited, primarily due to these two challenges. 1) Existing RS semantic categories are limited, particularly for pixel-level interpretation datasets. 2) Distinguishing among diverse RS spatial regions solely by language space is challenging due to the dense and intricate spatial distribution in open-world RS imagery. To address the first issue, we develop a fine-grained RS interpretation dataset, Sky-SA, which contains 183,375 high-quality local image-text pairs with full-pixel manual annotations, covering 1,763 category labels, exhibiting richer semantics and higher density than previous datasets. Afterwards, to solve the second issue, we introduce the vision-centric principle for vision-language modeling. Specifically, in the pre-training stage, the visual self-supervised paradigm is incorporated into image- + +text alignment, reducing the degradation of general visual representation capabilities of existing paradigms. Then, we construct a visual-relevance knowledge graph across open-category texts and further develop a novel vision-centric image-text contrastive loss for fine-tuning with text prompts. This new model, denoted as SkySense-O, demonstrates impressive zero-shot capabilities on a thorough evaluation encompassing 14 datasets over 4 tasks, from recognizing to reasoning and classification to localization. Specifically, it outperforms the latest models such as SegEarthOV, GeoRSCLIP, and VHM by a large margin, i.e., $11.95\%$ , $8.04\%$ and $3.55\%$ on average respectively. The code is publicly available to facilitate further research at https: + +//github.com/zqcrafts/SkySense-0. + +# 1. Introduction + +Remote sensing interpretation (RSI) is a critical and practical task in Earth observation, aiming to identify the semantic types of specific ground objects in remote sensing (RS) images (such as segmentation [1-4], detection [5-10], change detection [11, 12], and visual question answering + +(VQA) [13-15]). Early works typically involved annotating RS images with fixed categories and exploring RSI through supervised learning methods. These approaches do not generalize well to unseen new categories in open-world scenarios, thus limiting their practical applicability. Moreover, as RS images becoming more accessible than ever before, there is an increasing demand for open-world interpretations with a broad category scope. Hence, there is a soaring interest in exploring a comprehensive open-world remote sensing interpretation model for many scenarios and tasks [16, 17]. + +Challenges. In the domain of natural image scenes, significant progress has been achieved in open-world interpretation tasks through large-scale localization and recognition datasets like AS-1B [18] and COCO-Stuff [19], combined with contrastive language-image pretrained model (CLIP) shown in Fig. 1a. However, the open-world RS interpretation model is still out of reach due to following two reasons: 1) From the dataset perspective, the existing annotations either exhibit sparse mask labels or lack diversity in category descriptions (see Fig.1b). 2) From the modeling perspective, the language model inherently encompasses strong contextual dependence [20], making it challenging to directly use as prompts for decoupling common RS objects and intertwined land-cover categories in open-world. + +Sky-SA Dataset. For Struggle (1), inspired by SAM [21] and ASM [18] in the natural image domain, we propose the Sky-SA dataset with the largest number of category labels for open-world RS interpretation, including 1,763 category texts from 183,375 mask-text pairs. The Sky-SA is mainly characterized by its text openness and fine granularity (Fig. 1b). Specifically, different from general semantic segmentation datasets, the categories of Sky-SA are annotated by short semantic tags in a completely open manner. In other words, the annotators are doing fill-in-the-blank rather than multiple-choice. This enables the model to learn pixel-level open-world interpretation akin to human cognition [18] (Fig. 1c). Furthermore, similar to COCO-Stuff [19], all recognizable pixels of Sky-SA are densely annotated, which further enhances the semantic openness and the contextual relationships from wide semantic categories. + +Vision-Centric Modeling. For Struggle (2), motivated by [22], we introduce the vision-centric principle to correct the potential alignment ambiguity arising from language guidance. Firstly, we initialize the visual encoder from SkySense [23], an advanced RS visual foundation model, and then collect large-scale RS image-caption datasets for pretraining, which mainly include SkyScript [23], RS5M [24] and OSM [25]. Motivated by DetailCLIP [26], to mitigate the degradation of general visual representation capabilities originating from SkySense, we maintain the visual contrastive loss and masked modeling loss during + +![](images/f9735a931aa5e5d37fed58ed90976ebc0a3a5b445a0571e1c9735debbd8be184.jpg) +Figure 2. The flowchart of the overall SkySense-O development. + +the pre-training process of image-text alignment. Secondly, for fine-tuning region-level alignment, we leverage GPT-4V [27] to construct a visual-centric knowledge graph across all categories of the Sky-SA dataset, where the node of the graph represents each category, and the edge weight indicates the visual similarity scores between each pair of categories. Using the graph, we derive a vision-centric regularization that focuses on distinguishing categories exhibiting low visual similarity but strong text-level contextual relationships in RS images. Finally, we explore the impacts of Skysense-O as agent on MLLM models (Fig. 1d). In summary, to achieve open-world RS interpretation model, SkySense-O is meticulously developed from dataset to modeling and from pre-training to fine-tuning (Fig. 2), contributions of this paper include: + +- We develop Sky-SA dataset with large-scale and high-quality annotated texts for open-world RS interpretation. +- The vision-centric principle is first introduced into RS visual-language modeling, primarily instantiated into training paradigm, vision-centric graph, and vision-centric regularization from pre-training to fine-tuning. +- We propose SkySense-O, a remarkable open-world RS interpretation model, which achieves SOTA across 14 zeroshot downstream datasets as shown in Fig. 1. + +# 2. Related work + +# 2.1. Remote Sensing Interpretation + +Remote Sensing (RS) interpretation aims to accurately identify and analyze ground objects from RS data, with applications in environmental monitoring, urban planning, and agricultural surveillance [35-42]. Early deep learning approaches focused on closed-set supervised models, in + +
DatasetSpatial GranularityOpen TextDense Annotation
ImageRegionPixel
Natural Image
SA-1B [21]✓✓✓
COCO-Stuff [19]✓✓✓
Remote Sensing Image
AID [28], RESISC [29]✓
OpenStreetMap [30]✓✓
RS5M [24], SkyScript [31]✓✓
Potsdam, LoveDA [32]✓✓✓
Valiggan, OEM [30]✓✓✓
iSAID [33], SIOR [34]✓✓
FAST [34], SOTA [34]✓✓
Sky-SA (ours)✓✓✓✓✓
+ +Table 1. Property comparisons with popular vision-language datasets. The proposed Sky-SA dataset is characterized by its semantic openness, high-density and fine-granularity. + +![](images/c4e55be500c80c107790f36276420462314ecf06a1c1a15117aedbbc08bae462.jpg) +(a) Florence-2 +Figure 3. Visual comparisons between the predictions of existing open-world models and the annotations of our Sky-SA dataset. + +![](images/859f1af60096973f92f29262a473a90f95be67defa967f96822ff58d2f93d440.jpg) +(b) SAM + +![](images/1ab34c65d913139baedb349e32bdbfd7694b51995461e1f85b8c7a25463ad644.jpg) +(c) Grounding DINO-1.6 + +![](images/7289831b60b558ca66b17c2b1f130ceffbb8dcbf3ec05239093505fae87b3bac.jpg) +(d) Sky-SA Dataset + +cluding image classification [43, 44], object detection [45-49], change detection [11, 12], and semantic segmentation [28, 50-56]. Recently, CLIP [57] and Vision Language Models (VLMs) have enabled the transition from closed-set to open-set interpretation, allowing models to recognize broader categories [17, 58]. Additionally, models like GPT-4 [59] and LLaVA [60] have facilitated the development of instruction-tuning datasets [61], enhancing the flexibility of RS models in human-machine interactions. + +# 2.2. Vision Foundation Models for Remote Sensing + +Vision foundation models are instrumental in advancing RS interpretation toward open-set capabilities through two primary pre-training paradigms, vision-only self-supervised and vision-language self-supervised [22, 43, 62-67] paradigms. Vision-only self-supervised models utilize large-scale unannotated RS images to develop robust visual representations for various tasks [23, 68-71]. For example, SkySense [23] integrates multiple modalities and geo-contextual information within a contrastive framework, while Masked Image Modeling (MIM) methods like RingMo [72] and Scale-MAE [73] are proposed for further enhancing visual feature learning. Vision-language self-supervised models originate from CLIP. Recently some efforts such as RemoteCLIP [74], RS5M [24], and SkyScript [23] have customized for the RS domain using extensive RS image-caption pairs. + +![](images/95fddfe2a71f7c337796667809fa5e7e997dc4f6d907baee99dde012d71826b3.jpg) +Figure 4. Scale comparisons with existing RS semantic segmentation datasets. We see that the proposed Sky-SA dataset has a significantly larger category scale and diversity than prior datasets. + +# 3. Sky-SA Dataset + +# 3.1. Motivation + +Recently, image-text alignment models based on natural images, such as the open-vocabulary methods [75] and MLLMs like LLaVA [60], have achieved remarkable success. This is primarily due to well-constructed datasets in the natural image domain compared to the RS image domain. There have been numerous high-quality datasets with rich categories in the natural image community, such as COCO-Stuff [19] and ADE [76]. However, as shown in Tab.1 and Fig.4, such resources remain scarce in the RS community, which significantly limits the development of open-world RS interpretation tasks. Moreover, existing open-world models, such as SAM [21], Florence-2 [77], and Grounding-DINO [78], are also unable to achieve dense and fine-grained interpretation on remote sensing images as shown in Fig. 3. To relieve this issue, we design a data engine to help annotators more efficiently and accurately perform full-pixel annotation on unlabeled RS images, leading to the creation of the Sky-SA dataset as following section. + +# 3.2. Data Engine + +Open-Images Collection. In the standard collection process of a closed-set dataset, a category list is predefined, and images that do not contain any items from this list will be filtered out. This manner might lead to dataset bias [79] in remote sensing images. Unlike the close-set methods, we compile a comprehensive collection of images from various available RS datasets. To minimize redundancy in manual annotation, we leverage the SkySense model [23], which possesses image-level visual discrimination capabilities for RS images, to filter visually redundant images. + +Open-Text Fine-Grained Annotation. In the annotation process, drawing inspiration from the popular COCO-Stuff [19] dataset, we prioritize fine-grained semantics and + +![](images/87c6bc3c040934a2175e3cc876c78ae8a8e02f0334fbffd8dc850ac8ebf71f70.jpg) +Pre-training (For image-level RS interpretation task) + +![](images/b684b261ae0159528d91049f1705c708a9f7cb3447dcda8ef131967d9a14a7c6.jpg) +Supervised Finetuning (For pixel-level RS interpretation task) +Figure 5. The framework of SkySense-O, where training strategy includes pre-training and fine-tuning stage for open-world interpretation. + +full-pixel annotation as core labeling principles. Specifically, there are two key points: 1) the dataset encompasses a rich variety of semantic categories; 2) almost all pixels in each image need to be annotated. The former aids in activating the broad semantic awareness of the vision-language pretrained model, while the latter helps establish comprehensive connections among local semantics throughout each image. Furthermore, to capture a broader semantic range, we employ a short-caption labeling manner for each region, which does not constrain the annotation labels. Additionally, to improve annotation efficiency, we first utilize GPT-4V [27] to generate the image content, and then employ Grounding-DINO [78] and SAM [21] for pre-annotation, allowing annotators to make careful corrections based on these preliminary labels. + +Dataset Statistics. Finally, we develop a high-quality pixel-level remote sensing interpretation dataset, named Sky-SA. As shown in Fig. 4, the Sky-SA dataset consists of 35,000 RS images, mainly including 1,763 category texts from 183,375 segment-text pairs (encompassing approximately 500 distinct categories of land-covering and RS objects after synonym filtering), which far surpasses existing remote sensing segmentation datasets. Moreover, Sky-SA is also the first pixel-level open-text annotated dataset in the field of remote sensing (see Fig. 1), laying a solid data foundation for the development of open-world RS interpretation. + +# 4. SkySense-O Model + +# 4.1. Pretraining for Foundation Model + +The training of SkySense-O is primarily divided into pretraining and fine-tuning stages as shown in Fig. 5. During pre-training, given the input image $I_{i}$ and its corresponding text $T_{i}$ , for $i = 1,\dots,B$ , where $B$ denotes mini-batch size, the model mainly consists of an image encoder $E_{img}$ and a text encoder $E_{text}$ . To maintain the model's visual perception capabilities [26], especially for densely distributed objects during the image-text alignment process, the vision encoder is initialized by SkySense [23] and its visual + +self-supervised paradigms are still maintained, i.e. masked image modeling and image contrastive learning. Given $x = E_{img}(I)$ , $y = E_{text}(T)$ , mask decoder output $R'$ , the overall pretraining loss is computed as Eq. 1: + +$$ +\mathcal {L} _ {\text {p r e t r a i n}} = C E (x, y) + C E \left(x, x ^ {\prime}\right) + M R \left(I ^ {\prime}, R ^ {\prime}\right), \tag {1} +$$ + +where $I^{\prime}$ and $x^{\prime}$ are the augmented version of $I$ and $x$ respectively, $MR$ is mask reconstruction loss following [71] and $CE$ denotes cross-entropy loss computing as: + +$$ +C E (x, y) = - \frac {1}{B} \sum_ {i = 1} ^ {B} \log \frac {\exp \left(\operatorname {s i m} \left(x _ {i} , y _ {i}\right) / \tau\right)}{\sum_ {j = 1} ^ {B} \exp \left(\operatorname {s i m} \left(x _ {i} , y _ {j}\right) / \tau\right)}, \tag {2} +$$ + +where $\tau$ is logit temperature and $sim$ is cosine similarity. + +# 4.2. Finetuning for Interpretation Model + +Region-level Finetuning. During the region-level finetuning stage, based on the proposed Sky-SA dataset, given a set of candidate category texts $C_n$ and masks $M_n$ for each image, where $n = 1, \dots, N$ , we apply mask pooling to extract local features for each category from the images, and then compute the region-level cross-entropy loss with the corresponding category, which can be expressed as: + +$$ +\mathcal {L} _ {\text {r e g i o n}} = \sum_ {n = 1} ^ {N} C E \left(P o o l i n g \left(x _ {n}, M _ {n}\right), E _ {\text {t e x t}} \left(C _ {n}\right)\right). \tag {3} +$$ + +Pixel-level Finetuning. To achieve pixel-level interpretation, referring to the existing open-vocabulary semantic segmentation model [80], we first compute the similarity between the image encoder and the text encoder at the token level to obtain coarse location features. Subsequently, a visual decoder $D_{img}$ is employed to up-sample these features to the fine-grained segmentation results, constrained by a pixel-wise cross-entropy loss as Eq. 4: + +$$ +\mathcal {L} _ {\text {p i x e l}} = \sum_ {n = 1} ^ {N} C E \left(D _ {\text {i m g}} \left(x _ {n}, E _ {\text {t e x t}} \left(C _ {n}\right)\right), M _ {n}\right). \tag {4} +$$ + +More details can be found in the supplementary and code. + +![](images/17864a17972ce502161b7bed589dc73601d06f30ee3db7dfcfe6d9edfaee4274.jpg) +Figure 6. Positive and negative sampling mechanism according to the correspondence score between source sample and target sample. (a): Due to the contextual dependence in language modeling or vision-language modeling [20], "road" is mistakenly assigned as the positive sample of "car", while our agent employs the correct sampler. (b): Another case to further valid this point. + +![](images/91734500729032caa9ad2f494192bfc436d603a03c7766304614a8aa70a24d37.jpg) +Figure 7. Constructing illustrations of the vision-centric knowledge graph by an elaborate vision-centric agent flow. + +# 4.3. Vision-Centric Modeling + +Challenge. Different from the previous close-set semantic segmentation dataset, the Sky-SA dataset is annotated in open-world texts. Although this manner provides more open vocabulary support, the ambiguous sampling challenge often leads to training collapse in discriminative models. As illustrated in Fig. 6, "car" is mistakenly regarded as the positive sample of "road" by computing the similarity between their text-embeddings due to their strong text-level or image-level contextual relationships. + +Visual-Centric Knowledge Graph. To overcome the challenge, inspired by the in-depth exploration of visual representation by Xie et al. [22] and GraphRAG [81], we build a visual-centric agent using GPT-4 to obtain the visual relevance score for each text pair of the Sky-SA dataset. As shown in Fig. 7, the score is measured by the meticulously chain of thought about visual attributes, including appearance, color, and shape etc.. When all text pairs in Sky-SA are fully traversed, we can construct an offline vision-centric knowledge graph, which mainly consists of three elements: source node, target node, and edge. The edge indicates the visual-relevance score and reasons from source node to target node. Based on the graph, we perform flexible sampling for positive and negative samples by setting a sampling threshold illustrated in Algorithm 1. + +# Algorithm 1 Finetuning Stage for SkySense-O. + +![](images/480fbc1faac760c1d7070e935b259045d46bcb58c6a2bbd481836f89834efeac.jpg) + +![](images/0288e880412cb19e22e8d16ceceacc632d0063a0f2a128dce5c2777f4b4e5ace.jpg) + +![](images/5470adb042dc2cb28ad2c31c09d08e5c552a84e5288fcbac3db1e8ee2a7dac10.jpg) + +![](images/21c5c183a66a669e347df6911ddd44d087b6cc640fedc722dbc1af3bf839095d.jpg) +Figure 8. Sample similarity matrices in computing vision-centric regularization. Hard samples, such as vehicle and road, are exploited by multiplying $S_{t}$ and inversed $S_{v}$ . + +Vision-Centric Regularization Loss. In RS imagery, the objects and land-cover are often densely packed and intricately intertwined, posing challenges for fine-grained visual-language modeling. To relieve this issue, we further exploit the potential of the vision-centric graph and design a vision-centric regularization loss, $\mathcal{L}_{\mathrm{vc}}$ . The regularization is built on the observation that the samples with strong contextual relationships but weak visual relevance scores are hard to distinguish as shown in Fig. 6. Specifically, given selected $M$ negative samples $C_m$ for class $C_i$ , their textual similarities $S_t$ are computed by their embedding similarities. Then, given text pairs, the visual relevance $S_v^{i,m}$ can be retrieved by the visual-centric graph. Finally, $\mathcal{L}_{\mathrm{vc}}$ is computed as Eq. 5 and details are in Algorithm 1 and Fig. 8. + +$$ +\mathcal {L} _ {\mathrm {v c}} = \frac {1}{B N} \sum_ {i = 1} ^ {B} \sum_ {n = 1} ^ {N} \left\| \left(1 - S _ {v} ^ {i, m}\right) * S _ {t} ^ {i, m} * \mathcal {L} _ {\text {r e g i o n}} \right\| ^ {2}. \tag {5} +$$ + +Total Loss. In summary, the total loss for the finetuning stage of SkySense-O is defined as follows: + +$$ +\mathcal {L} _ {\text {f i n e t u n e}} = \mathcal {L} _ {\text {p i x e l}} + \lambda_ {1} \mathcal {L} _ {\text {r e g i o n}} + \lambda_ {2} \mathcal {L} _ {\text {v c}}, \tag {6} +$$ + +where $\lambda_{1}$ and $\lambda_{2}$ are two hyperparameters and set as 0.4 and 0.4 respectively for optimal ratio via performing ablations. + +
ModelPublicationObject Semantic SegmentationLand-Cover Semantic Segmentation
iSAID [33]SOTA [34]FAST [34]SIOR [34]Potsdam [82]OEM [30]LoveDA [32]VaihingenFloodNet [83]
Few-Shot Semantic Segmentation
SegGPT [84]ICCV 202329.1731.4119.5537.4730.1225.1931.4047.2733.81
DINOv [85]CVPR 202425.3518.247.9225.0437.6523.0525.1735.1044.55
Zero-Shot Semantic Segmentation
SAN [80]CVPR 202311.7710.733.1615.2251.4524.8727.8346.2339.26
CAT-SEG [75]CVPR 202420.5813.713.1319.6450.8130.2637.1549.6241.36
SegEarth-OV [58]Arxiv 202421.7011.992.8413.9247.1040.3036.9029.1034.00
SkySense-O-43.9232.128.3330.8954.1040.8338.3051.6442.04
-(+22.22)(+18.41)(+5.2)(+11.25)(+2.65)(+0.53)(+1.15)(+2.02)(+0.68)
+ +Table 2. Zero-shot remote sensing object and land-cover semantic segmentation performances. Bold indicates the highest performance, and underline indicates the second highest performance in mIoU (%). + +
ModelPublicationSIRI-WHU [86]AID-VQA [28]RSVQA-HR [14]
Acc.Acc.Avg. Acc.
GeoChat [35]CVPR 202466.6372.0370.82
LHRS-Bot [87]ECCV 202462.6691.26-
EarthGPT [88]TGRS 2024--71.15
VHM [89]Arxiv 202470.8891.7073.75
SkySense-O†-72.45 (+1.57)92.95 (+1.25)75.39 (+1.64)
SkySense-O-74.79 (+3.91)94.10 (+2.40)78.09 (+4.34)
+ +Table 3. Zero-Shot visual question-answering performances, compared with existing MLLMs. $\dagger$ indicates wo/ pixel-level fine-tuning by Sky-SA. + +
ModelPublicationRESISC [29]AID [28]
CLIP [57]PMLR 202169.2566.70
RemoteCLIP [74]TGRS 202470.8574.3
CLIP-laion-RS [31]-71.7071.21
SkyScript [31]AAA1 202471.7070.94
GeoRSCLIP [24]TGRS 202373.9076.27
SkySense-O†-83.2882.93
-(+9.38)(+6.66)
+ +Table 4. Zero-shot remote sensing scene classification performances in Top-1 accuracy $(\%)$ + +# 5. Experiment + +# 5.1. Evaluation Benchmarks + +We evaluate SkySense-O on four representative RS interpretation tasks. To ensure fair zero-shot comparisons, all experiments are conducted without extra fine-tuning and labels from the evaluation datasets are not used during training. The supplementary materials provide further details on the implementation and visualization results. + +Zero-shot object segmentation. We adopt iSAID [33], FAST [34], SIOR [34], SOTA [34] for the object segmentation experiment. They are chosen considering diverse scenes and categories. We report mIoU results on official validation sets. As depicted in Tab. 2, SkySense-O has achieved the SOTA performance on all object segmentation datasets. On average, it outperforms the previous SOTAs by an average margin of $14.27\%$ . Furthermore, it even surpasses few-shot SOTAs on most datasets. + +Zero-shot land-cover segmentation. We assess SkySense-O's land-cover segmentation performance on ISPRS Potsdam [82], Vaihingen and LoveDA [32] datasets. SkySense-O excels on all three datasets (Tab. 2). Notably, we surpass the second best CAT-SEG [75] by $1.41\%$ mIoU and have achieved the best performance on the Vaihingen leaderboard. More importantly, our results demonstrate the potential to achieve more accurate interpretations than the ground truth (the second row in Fig. 9). + +Zero-shot classification. We utilize two scene classification datasets: AID [28] and RESISC-45 [29]. The validation setting follows GeoRSCLIP [24] and Top-1 Accuracy is reported. SkySense-O overall outperforms competitive + +![](images/ce92ee71221c85e5c5c28022eff8a3f91a630d76d827870c55ce37b3a724f34e.jpg) +Figure 9. Visualization of object segmentation results on iSAID dataset and land-cover segmentation results on Potsdam dataset. + +basielines, achieving $9.38\%$ and $6.66\%$ higher performance than state-of-the-art methods on the two datasets (Tab. 4). + +Zero-shot RS VQA. To further validate the versatility of SkySense-O, we employ it in MLLM as a vision encoder for experimentation. Specifically, we maintain the same architecture as LLaVA-1.5 [90], except that we replace the original CLIP-L-14 [57] with our SkySense-O, and set the image resolution to $512 \times 512$ pixels. Regarding the setup of the instruction tuning data, during the pretraining stage, we utilize CC-558K dataset [60] and 121k image-text pairs converted from ChatearthNet [91] dataset. In the finetuning stage, based on instruction-tuning datasets collected by SkySenseGPT [61] from different sources, we select a subset of 593k data as our downstream instruction finetuning dataset, where details are in the supplementary. We evaluate zero-shot VQA performances on several public RS benchmarks as shown in Tab. 3. + +![](images/a82e7018650b4045c67610b2640045e4331fb8fc96341dc8616e07233e780289.jpg) +Figure 10. Visualization of text-image alignment in CLIP. (a) Computing the similarity between each text token and all image tokens to obtain text-image relevance (TIR) maps for different categories. (b) Visualizing TIR maps for some high-response texts. (c) Histogram of the maximum response values of TIR maps across the text list. The details of input texts can be referred to the supplementary materials. + +
Foundation +ModelVision +EncoderVision-Enc. +ParameterNum. of +Image-textRESISC +(Acc)
RemoteCLIPViT-L-14304M0.83M70.85
SkyScriptViT-L-14304M2.6M71.70
GeoRSCLIPViT-H-14630M5M73.90
OursSwinV2-H-8654M10M83.28
+ +Table 5. Comparison with existing RS vision-language foundation models. SkySense-O achieves consistent performance gains due to scaling laws in terms of model-size and data-size. + +
+Sky-SA+Graph+Pretrain-ing+LvcmIoU
18.40
✓27.76 (+9.36)
✓✓36.28 (+8.52)
✓✓✓41.69 (+5.41)
✓✓✓✓43.92 (+2.24)
+ +Table 6. Ablation study on the proposed dataset and designs. + +# 5.2. Implementation Details. + +In the pre-training stage, we use 80 A100-80GB GPUs and primarily adopt two large-scale remote sensing image-text datasets, SkyScript [31] and RS5M [24], and further expand their captions using Qwen-VL [92] and InternVL [93], comprising 10 million image-text pairs. In the fine-tuning stage, we employ OSM [25] for local image-text fine-tuning and Sky-SA for fine-grained interpretation capabilities on 8 A100-80GB GPUs as shown in Tab. 5. More details are included in the supplementary materials. + +# 5.3. Discussions & Ablation Studies + +Effectiveness of Image-Language Alignment. In Fig. 10, we follow MaskCLIP [94] and adopt the image-text correla + +![](images/c15e0225292e0a27df5f81a9961f5b4d4f7e210bd01863fa8810eaa63d455cac.jpg) +Figure 11. Visualizations of open-world interpretation results by SkySense-O on iSAID, Potsdam, FAST, and SIOR datasets. + +tion in the representation space to evaluate the effectiveness of vision-language foundation models. We select the CLIP models from SegEarth-OV [58] and CAT-Seg [75] for comparison. The former is a standard CLIP model, while the latter is a locally fine-tuned CLIP model by COCO-Stuff [19]. We randomly selected 131 category texts from Sky-SA as inputs. As shown in Figure 10(c), SkySense-O exhibits larger response variance across different categories, indicating stronger discriminative capabilities among different texts. In Fig. 10(b), we present several image-text correlation maps to further validate the effectiveness of SkySense-O in image-text alignment in open-world scenes. + +- vehicle vs. parking lot. (small object) Because our modeling is vision-centric, small targets are effectively decoupled from the contextual background. +- building vs. industrial park. (hypernym) SkySense-O demonstrates proper alignment capabilities across different levels of semantic hierarchy due to large-scale remote sensing knowledge are learned in training. + +![](images/5bd44f34fccba169c2e267e81fd621581d7aaeddf938f7cfd803ae9de9a249c8.jpg) +Figure 12. The visualization of vision-centric graph of the Sky-SA dataset. Node layout is performed via Force Atlas [95] and graph communities are detected by the Leiden algorithm [96]. Circles represent entity nodes with size proportional to their degree. Node colors represent entity communities. Five main communities are detected by an unsupervised manner and they just correspond to the major land-cover categories in the remote sensing imagery. + +![](images/ba430331eb48f17344c089da0cf1c3800926cff7baebf5a257e2c44e2fe326e7.jpg) +(a) Image + +![](images/a1906c6cc5dc0f4f9c18162540578f8356ceab4a82e93a4a3067bf22acbc6bd9.jpg) +(b) Wo/ vision-centric + +![](images/77b703c43dbf4e214aa13b80c4b9674227f157c86c20f5f3b4e289cc7a9aa4f4.jpg) +(c) W/ vision-centric +Figure 13. Comparisons of the results without and with vision-centric regularization loss. Compared to (b), the vehicles in (c) are clearly distinguished from the contextual floor. + +Design of Vision-Centric Strategy. Tab. 6 presents the ablation study to assess our Sky-SA dataset and vision-centric designs, namely vision-centric graph sampler, pretraining foundation model, and vision-centric regularization loss. We utilize the iSAID dataset as evaluation benchmark and report mIoU metric on the official test set. Initially, we use the standard CLIP model [57] and fine-tune it on the COCO-Stuff dataset [19]. Compared with the baseline, the Sky-SA dataset leads to a notable improvement compared to COCO-Stuff dataset. Nonetheless, we believe that the Sky-SA dataset holds considerable untapped potential, which might be constrained by the intricate entanglement of open-world text labels as demonstrated in Sec. 4.1. To relieve this problem, we construct a vision-centric graph for all text labels of the Sky-SA dataset as shown in Fig. 12. It is intriguing that the nodes in the graph naturally exhibit proper unsupervised clustering results by community detection [96], without using any text labels. This further validates the rationale for vision-centric modeling and the introducing of vision-centric graph leads to $8.52\%$ mIoU improvement. Then, we replace the standard CLIP with our pre-trained model incorporating extensive remote sensing knowledge. This effectively improves the perfor + +![](images/0db4d3fddac20049b8798486635d427945c423dcad5f36139e91643a3ce48808.jpg) +Figure 14. Zero-shot semantic segmentation results of Sky-SA and existing datasets. The vertical axis represents the training set, while the horizontal axis represents the test set. The yellow box denotes the result is fine-tuned by its own training set (i.e. supervised learning) and red box denotes the average performance. + +mance to $41.69\%$ mIoU, validating the necessity of remote sensing pre-training model. Finally, we introduce vision-centric regularization loss, which uncovers and prioritizes hard samples in visual-language modeling. The results show that the constraint pushes the performance to a very strong $43.92\%$ mIoU. Furthermore, as shown in Fig. 13, this design enhances the discrimination ability, especially between high-context categories. + +Significance of Sky-SA for Open-World Interpretation. In Fig. 14, we replace Sky-SA with existing datasets and conduct experiments under the same settings as in Tab. 2. It can be observed that existing datasets exhibit poor generalization to other datasets, except on their own test sets. In contrast, the Sky-SA dataset achieves superior zero-shot segmentation performance across all datasets, outperforming the second best dataset by an average of over $13.43\%$ mIoU, which demonstrates the significant potential of Sky-SA for open-world interpretation. + +# 6. Conclusion + +In this paper, we present Skysense-O, a large-scale and general remote sensing interpretation model for open remote sensing images. Skysense-O is mainly driven by the proposed semantically-rich and fine-grained dataset SkySA and the vision-centric visual-language modeling, consistently outperforming other models on a variety of tasks, showcasing its exceptional generalization ability and strong performance. We hope this work will constructively push the RS open-world interpretation on constructing large-scale vision-language datasets and models for further researches. Our future work aims to explore fine-grained remote sensing interpretation, such as analyzing object states and relationships, further enhancing the flexibility of open-world applications. + +# Acknowledgments + +This work was supported by the Anhui Provincial Natural Science Foundation under Grant 2108085UD12 and the National Natural Science Key Foundation of China under Grant 62036005. We acknowledge the support of GPU cluster built by MCC Lab of Information Science and Technology Institution, USTC. + +# References + +[1] Xianping Ma, Xiaokang Zhang, Man-On Pun, and Bo Huang. Manet: Fine-tuning segment anything model for multimodal remote sensing semantic segmentation, 2024. 1 +[2] Aysim Token, Marvin Eisenberger, Daniel Cremers, and Laura Leal-Taixe. Satsynth: Augmenting image-mask pairs through diffusion models for aerial semantic segmentation, CVPR 2024. +[3] Deyi Ji, Feng Zhao, Hongtao Lu, Feng Wu, and Jieping Ye. Structural and statistical texture knowledge distillation and learning for segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence, pages 1-18, 2025. +[4] Deyi Ji, Feng Zhao, Hongtao Lu, Mingyuan Tao, and Jieping Ye. Ultra-high resolution segmentation with ultra-rich context: A novel benchmark. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23621-23630, 2023. 1 +[5] Deyi Ji, Feng Zhao, Lanyun Zhu, Wenwei Jin, Hongtao Lu, and Jieping Ye. Discrete latent perspective learning for segmentation and detection. In *Forty-first International Conference on Machine Learning*, pages 21719-21730, 2024. 1 +[6] Botao Ren, Xue Yang, Yi Yu, Junwei Luo, and Zhidong Deng. Pointobb-v2: Towards simpler, faster, and stronger single point supervised oriented object detection. arXiv preprint arXiv:2410.08210, 2024. +[7] Kang Wu, Zunxiao Xu, Xinrong Lyu, and Peng Ren. Cloud detection with boundary nets. ISPRS Journal of Photogrammetry and Remote Sensing, 186:218-231, 2022. +[8] Peiyuan Zhang, Junwei Luo, Xue Yang, Yi Yu, Qingyun Li, Yue Zhou, Xiaosong Jia, Xudong Lu, Jingdong Chen, Xiang Li, et al. Pointobb-v3: Expanding performance boundaries of single point-supervised oriented object detection. arXiv preprint arXiv:2501.13898, 2025. +[9] Yi Yu, Xue Yang, Qingyun Li, Yue Zhou, Feipeng Da, and Junchi Yan. H2rbox-v2: Incorporating symmetry for boosting horizontal box supervised oriented object detection. Advances in Neural Information Processing Systems, 36, 2024. +[10] Kang Wu, Zunxiao Xu, Xinrong Lyu, and Peng Ren. Crosssupervised learning for cloud detection. *GIScience & Remote Sensing*, 60(1):2147298, 2023. 1 +[11] Zhuo Zheng, Ailong Ma, Liangpei Zhang, and Yanfei Zhong. Change is everywhere: Single-temporal supervised object change detection in remote sensing imagery. In Proceedings of the IEEE/CVF international conference on computer vision, pages 15193-15202, 2021. 1, 3 +[12] Deyi Ji, Siqi Gao, Mingyuan Tao, Hongtao Lu, and Feng Zhao. Changenet: Multi-temporal asymmetric change detect + +tion dataset. In ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing, pages 2725-2729. IEEE, 2024. 1, 3 +[13] Junwei Luo, Yingying Zhang, Xue Yang, Kang Wu, Qi Zhu, Lei Liang, Jingdong Chen, and Yansheng Li. When large vision-language model meets large remote sensing imagery: Coarse-to-fine text-guided token pruning. arXiv preprint arXiv:2503.07588, 2025. 2 +[14] Sylvain Lobry, Diego Marcos, Jesse Murray, and Devis Tuia. Rsvqa: Visual question answering for remote sensing data. IEEE Transactions on Geoscience and Remote Sensing, 58(12):8555-8566, 2020. 6 +[15] Junjue Wang, Zhuo Zheng, Zihang Chen, Ailong Ma, and Yanfei Zhong. Earthvqa: Towards queryable earth via relational reasoning-based remote sensing visual question answering. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 5481-5489, 2024. 2 +[16] Yan Li, Weiwei Guo, Xue Yang, Ning Liao, Dunyun He, Jiaqi Zhou, and Wenxian Yu. Toward open vocabulary aerial object detection with clip-activated student-teacher learning. In European Conference on Computer Vision, pages 431–448. Springer, 2024. 2 +[17] Jiancheng Pan, Yanxing Liu, Yuqian Fu, Muyuan Ma, Jiaohao Li, Danda Pani Paudel, Luc Van Gool, and Xiaomeng Huang. Locate anything on earth: Advancing open-vocabulary object detection for remote sensing community. arXiv preprint arXiv:2408.09110, 2024. 2, 3 +[18] Weiyun Wang, Min Shi, Qingyun Li, Wenhai Wang, Zhenhang Huang, Linjie Xing, Zhe Chen, Hao Li, Xizhou Zhu, Zhiguo Cao, et al. The all-seeing project: Towards panoptic visual recognition and understanding of the open world. arXiv preprint arXiv:2308.01907, 2023. 2 +[19] Holger Caesar, Jasper Uijlings, and Vittorio Ferrari. Cocostuff: Thing and stuff classes in context. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1209-1218, 2018. 2, 3, 7, 8 +[20] Shuyang Gu. Several questions of visual generation in 2024. arXiv preprint arXiv:2407.18290, 2024. 2, 5 +[21] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dollar, and Ross Girshick. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, October 2023. 2, 3, 4 +[22] Shengbang Tong, Zhuang Liu, Yuexiang Zhai, Yi Ma, Yann LeCun, and Saining Xie. Eyes wide shut? exploring the visual shortcomings of multimodal llms. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9568-9578, 2024. 2, 3, 5 +[23] Xin Guo, Jiangwei Lao, Bo Dang, Yingying Zhang, Lei Yu, Lixiang Ru, Liheng Zhong, Ziyuan Huang, Kang Wu, Dingxiang Hu, et al. Skysense: A multi-modal remote sensing foundation model towards universal interpretation for earth observation imagery. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 27672-27683, 2024. 2, 3, 4 +[24] Zilun Zhang, Tiancheng Zhao, Yulong Guo, and Jianwei Yin. Rs5m and georsclip: A large scale vision-language dataset + +and a large vision-language model for remote sensing. IEEE Transactions on Geoscience and Remote Sensing, pages 1-1, 2024. 2, 3, 6, 7 +[25] Haozhe Xie, Zhaoxi Chen, Fangzhou Hong, and Ziwei Liu. CityDreamer: Compositional generative model of unbounded 3D cities. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 2, 7 +[26] Amin Karimi Monsefi, Kishore Prakash Sailaja, Ali Alilooee, Ser-Nam Lim, and Rajiv Ramnath. Detailclip: Detail-oriented clip for fine-grained tasks. arXiv preprint arXiv:2409.06809, 2024. 2, 4 +[27] OpenAI. Gpt-4v(ision) system card. 2023. 2, 4 +[28] Gui-Song Xia, Jingwen Hu, Fan Hu, Baoguang Shi, Xiang Bai, Yanfei Zhong, Liangpei Zhang, and Xiaoqiang Lu. Aid: A benchmark data set for performance evaluation of aerial scene classification. IEEE Transactions on Geoscience and Remote Sensing, 55(7):3965-3981, 2017. 3, 6 +[29] Gong Cheng, Junwei Han, and Xiaogiang Lu. Remote sensing image scene classification: Benchmark and state of the art. Proceedings of the IEEE, 105(10):1865-1883, 2017. 3, 6 +[30] Junshi Xia, Naoto Yokoya, Bruno Adriano, and Clifford Broni-Bediako. Openearthmap: A benchmark dataset for global high-resolution land cover mapping. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 6254-6264, 2023. 3, 6 +[31] Zhecheng Wang, Rajanie Prabha, Tianyuan Huang, Jiajun Wu, and Ram Rajagopal. Skyscript: A large and semantically diverse vision-language dataset for remote sensing. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 5805-5813, 2024. 3, 6, 7 +[32] Junjue Wang, Zhuo Zheng, Ailong Ma, Xiaoyan Lu, and Yanfei Zhong. LoveDA: A remote sensing land-cover dataset for domain adaptive semantic segmentation, October 2021. 3, 6 +[33] Syed Waqas Zamir, Aditya Arora, Akshita Gupta, Salman Khan, Guolei Sun, Fahad Shahbaz Khan, Fan Zhu, Ling Shao, Gui-Song Xia, and Xiang Bai. isaid: A large-scale dataset for instance segmentation in aerial images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 28-37, 2019. 3, 6 +[34] Di Wang, Jing Zhang, Bo Du, Minqiang Xu, Lin Liu, Dacheng Tao, and Liangpei Zhang. Samrs: Scaling-up remote sensing segmentation dataset with segment anything model, NeurIPS 2023. 3, 6 +[35] Kartik Kuckreja, Muhammad Sohail Danish, Muzammal Naseer, Abhijit Das, Salman Khan, and Fahad Shahbaz Khan. Gechat: Grounded large vision-language model for remote sensing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 27831-27840, 2024. 2, 6 +[36] Tianrun Chen, Chaotao Ding, Lanyun Zhu, Tao Xu, Deyi Ji, Ying Zang, and Zejian Li. xlstm-unet can be an effective 2d\& 3d medical image segmentation backbone with vision-lstm (vil) better than its mamba counterpart. arXiv preprint arXiv:2407.01530, 2024. + +[37] Deyi Ji, Siqi Gao, Lanyun Zhu, Qi Zhu, Yiru Zhao, Peng Xu, Hongtao Lu, Feng Zhao, and Jieping Ye. View-centric multi-object tracking with homographic matching in moving UAV. arXiv preprint arXiv:2403.10830, 2024. +[38] Ying Zang, Runlong Cao, Jianqi Zhang, Yidong Han, Ziyue Cao, Wenjun Hu, Didi Zhu, Lanyun Zhu, Zejian Li, Deyi Ji, et al. Let human sketches help: Empowering challenging image segmentation task with freehand sketches. arXiv preprint arXiv:2501.19329, 2025. +[39] Lanyun Zhu, Tianrun Chen, Deyi Ji, Jieping Ye, and Jun Liu. Not every patch is needed: Towards a more efficient and effective backbone for video-based person re-identification. IEEE Transactions on Image Processing, 2025. +[40] Deyi Ji, Hongtao Lu, and Tongzhen Zhang. End to end multiscale convolutional neural network for crowd counting. In Eleventh international conference on machine vision, volume 11041, pages 761-766. SPIE, 2019. +[41] Zunxiao Xu, Kang Wu, Lei Huang, Qimao Wang, and Peng Ren. Cloudy image arithmetic: A cloudy scene synthesis paradigm with an application to deep-learning-based thin cloud removal. IEEE Transactions on Geoscience and Remote Sensing, 60:1-16, 2021. +[42] Ying Zang, Yuncan Gao, Jiangi Zhang, Yuangi Hu, Runlong Cao, Lanyun Zhu, Qi Zhu, Deyi Ji, Renjun Xu, and Tianrun Chen. Breaking the box: Enhancing remote sensing image segmentation with freehand sketches. arXiv preprint arXiv:2503.12191, 2025. 2 +[43] Zhihao Li, Biao Hou, Siteng Ma, Zitong Wu, Xianpeng Guo, Bo Ren, and Licheng Jiao. Masked angle-aware autoencoder for remote sensing images, European Conference on Computer Vision 2024. 3 +[44] Aaron E Maxwell, Timothy A Warner, and Fang Fang. Implementation of machine-learning classification in remote sensing: An applied review. International journal of remote sensing, 39(9):2784-2817, 2018. 3 +[45] Xue Yang, Gefan Zhang, Wentong Li, Yue Zhou, Xuehui Wang, and Junchi Yan. H2rbox: Horizontal box annotation is all you need for oriented object detection. In The Eleventh International Conference on Learning Representations, 2023. 3 +[46] Yansheng Li, Junwei Luo, Yongjun Zhang, Yihua Tan, JinGang Yu, and Song Bai. Learning to holistically detect bridges from large-size vhr remote sensing imagery. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. +[47] Haoran Wang, Licheng Jiao, Fang Liu, Lingling Li, Xu Liu, Deyi Ji, and Weihao Gan. Ipgn: Interactiveness proposal graph network for human-object interaction detection. IEEE Transactions on Image Processing, 30:6583-6593, 2021. +[48] Junwei Luo, Xue Yang, Yi Yu, Qingyun Li, Junchi Yan, and Yansheng Li. Pointobb: Learning oriented object detection via single point supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16730-16740, 2024. +[49] Yi Yu, Xue Yang, Qingyun Li, Feipeng Da, Jifeng Dai, Yu Qiao, and Junchi Yan. Point2rbox: Combine knowledge from synthetic visual patterns for end-to-end oriented object + +detection with single point supervision. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 16783-16793, 2024. 3 +[50] Lanyun Zhu, Deyi Ji, Shiping Zhu, Weihao Gan, Wei Wu, and Junjie Yan. Learning statistical texture for semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12537-12546, 2021. 3 +[51] Tianrun Chen, Ankang Lu, Lanyun Zhu, Chaotao Ding, Chunan Yu, Deyi Ji, Zejian Li, Lingyun Sun, Papa Mao, and Ying Zang. Sam2-adapter: Evaluating & adapting segment anything 2 in downstream tasks: Camouflage, shadow, medical image segmentation, and more. arXiv preprint arXiv:2408.04579, 2024. +[52] Gui-Song Xia, Xiang Bai, Jian Ding, Zhen Zhu, Serge Belongie, Jiebo Luo, Mihai Datcu, Marcello Pelillo, and Liangpei Zhang. Dota: A large-scale dataset for object detection in aerial images. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3974-3983, 2018. +[53] Deyi Ji, Haoran Wang, Mingyuan Tao, Jianqiang Huang, Xian-Sheng Hua, and Hongtao Lu. Structural and statistical texture knowledge distillation for semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16876-16885, 2022. +[54] Deyi Ji, Feng Zhao, and Hongtao Lu. Guided patch-grouping wavelet transformer with spatial congruence for ultra-high resolution segmentation. International Joint Conference on Artificial Intelligence, pages 920-928, 2023. +[55] Deyi Ji, Wenwei Jin, Hongtao Lu, and Feng Zhao. Pptformer: Pseudo multi-perspective transformer for uav segmentation. International Joint Conference on Artificial Intelligence, pages 893-901, 2024. +[56] Hanzhe Hu, Deyi Ji, Weihao Gan, Shuai Bai, Wei Wu, and Junjie Yan. Class-wise dynamic graph convolution for semantic segmentation. In European Conference on Computer Vision, pages 1-17. Springer, 2020. 3 +[57] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763, 2021. 3, 6, 8 +[58] Kaiyu Li, Ruixun Liu, Xiangyong Cao, Deyu Meng, and Zhi Wang. Segearth-ov: Towards training-free open-vocabulary segmentation for remote sensing images. arXiv preprint arXiv:2410.01768, 2024. 3, 6, 7 +[59] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 3 +[60] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In A. Oh, T. Naumann, A. Globerson, K. Saenko, M. Hardt, and S. Levine, editors, Advances in Neural Information Processing Systems, volume 36, pages 34892-34916, 2023. 3, 6 + +[61] Junwei Luo, Zhen Pang, Yongjun Zhang, Tingzhu Wang, Linlin Wang, Bo Dang, Jiangwei Lao, Jian Wang, Jingdong Chen, Yihua Tan, et al. Skysensegpt: A fine-grained instruction tuning dataset and model for remote sensing vision-language understanding. arXiv preprint arXiv:2406.10100, 2024.3,6 +[62] Tianrun Chen, Chunan Yu, Jing Li, Jianqi Zhang, Lanyun Zhu, Deyi Ji, Yong Zhang, Ying Zang, Zejian Li, and Lingyun Sun. Reasoning3d-grounding and reasoning in 3d: Fine-grained zero-shot open-vocabulary 3d reasoning part segmentation via large vision-language models. arXiv preprint arXiv:2405.19326, 2024. 3 +[63] Ziyue Huang, Yongchao Feng, Qingjie Liu, and Yunhong Wang. Mutdet: Mutually optimizing pre-training for remote sensing object detection, European Conference on Computer Vision 2024. +[64] Zhuo Zheng, Yanfei Zhong, Ailong Ma, and Liangpei Zhang. Single-temporal supervised learning for universal remote sensing change detection. International Journal of Computer Vision, 2024. +[65] Deyi Ji, Lanyun Zhu, Siqi Gao, Peng Xu, Hongtao Lu, Jieping Ye, and Feng Zhao. Tree-of-table: Unleashing the power of llms for enhanced large-scale table understanding. arXiv preprint arXiv:2411.08516, 2024. +[66] Lanyun Zhu, Tianrun Chen, Deyi Ji, Jieping Ye, and Jun Liu. Llafs: When large language models meet few-shot segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3065-3075, 2024. +[67] Lanyun Zhu, Deyi Ji, Tianrun Chen, Peng Xu, Jieping Ye, and Jun Liu. Ibid: Alleviating hallucinations in large vision-language models via image-biased decoding. arXiv preprint arXiv:2402.18476, 2024.3 +[68] Shizhen Chang and Pedram Ghamisi. Changes to captions: An attentive network for remote sensing change captioning. IEEE Transactions on Image Processing, 32:6047-6060, 2023. 3 +[69] Deyi Ji, Haoran Wang, Hanzhe Hu, Weihao Gan, Wei Wu, and Junjie Yan. Context-aware graph convolution network for target re-identification. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 1646–1654, 2021. +[70] Haoran Wang, Licheng Jiao, Fang Liu, Lingling Li, Xu Liu, Deyi Ji, and Weihao Gan. Learning social spatio-temporal relation graph in the wild and a video benchmark. IEEE Transactions on Neural Networks and Learning Systems, 34(6):2951-2964, 2021. +[71] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16000-16009, 2022. 3, 4 +[72] Xian Sun, Peijin Wang, Wanxuan Lu, Zicong Zhu, Xiaonan Lu, Qibin He, Junxi Li, Xuee Rong, Zhujun Yang, Hao Chang, et al. Ringmo: A remote sensing foundation model with masked image modeling. IEEE Transactions on Geoscience and Remote Sensing, 61:1-22, 2022. 3 + +[73] Colorado J Reed, Ritwik Gupta, Shufan Li, Sarah Brockman, Christopher Funk, Brian Clipp, Kurt Keutzer, Salvatore Candido, Matt Uytendaele, and Trevor Darrell. Scale-mae: A scale-aware masked autoencoder for multiscale geospatial representation learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4088-4099, 2023. 3 +[74] Fan Liu, Delong Chen, Zhangqingyun Guan, Xiaocong Zhou, Jiale Zhu, Qiaolin Ye, Liyong Fu, and Jun Zhou. Remoteclip: A vision language foundation model for remote sensing. IEEE Transactions on Geoscience and Remote Sensing, 2024. 3, 6 +[75] Seokju Cho, Heeseong Shin, Sunghwan Hong, Anurag Arnab, Paul Hongsuck Seo, and Seungryong Kim. Catseg: Cost aggregation for open-vocabulary semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4113-4123, 2024. 3, 6, 7 +[76] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 633-641, 2017. 3 +[77] Bin Xiao, Haiping Wu, Weijian Xu, Xiyang Dai, Houdong Hu, Yumao Lu, Michael Zeng, Ce Liu, and Lu Yuan. Florence-2: Advancing a unified representation for a variety of vision tasks, 2023. 3 +[78] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Qing Jiang, Chunyuan Li, Jianwei Yang, Hang Su, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023. 3, 4 +[79] Zhuang Liu and Kaiming He. A decade's battle on dataset bias: Are we there yet? arXiv preprint arXiv:2403.08632, 2024.3 +[80] Mengde Xu, Zheng Zhang, Fangyun Wei, Han Hu, and Xi-ang Bai. Side adapter network for open-vocabulary semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2945–2954, 2023. 4, 6 +[81] Boci Peng, Yun Zhu, Yongchao Liu, Xiaohe Bo, Haizhou Shi, Chuntao Hong, Yan Zhang, and Siliang Tang. Graph retrieval-augmented generation: A survey. arXiv preprint arXiv:2408.08921, 2024. 5 +[82] Jamie Sherrah. Fully convolutional networks for dense semantic labelling of high-resolution aerial imagery. arXiv preprint arXiv:1606.02585, 2016. 6 +[83] Maryam Rahnemoonfar, Tashnim Chowdhury, Argho Sarkar, Debvrat Varshney, Masoud Yari, and RobinR. Murphy. Floodnet: A high resolution aerial imagery dataset for post flood scene understanding. IEEE Access, Dec 2020. 6 +[84] Xinlong Wang, Xiaosong Zhang, Yue Cao, Wen Wang, Chunhua Shen, and Tiejun Huang. Seggpt: Segmenting everything in context. arXiv preprint arXiv:2304.03284, 2023. 6 +[85] Feng Li, Qing Jiang, Hao Zhang, Tianhe Ren, Shilong Liu, Xueyan Zou, Huaizhe Xu, Hongyang Li, Chunyuan Li, Jian- + +wei Yang, et al. Visual in-context prompting. arXiv preprint arXiv:2311.13601, 2023. 6 +[86] Qiqi Zhu, Yanfei Zhong, Bei Zhao, Gui-Song Xia, and Liangpei Zhang. Bag-of-visual-words scene classifier with local and global features for high spatial resolution remote sensing imagery. IEEE Geoscience and Remote Sensing Letters, 13(6):747-751, 2016. 6 +[87] Dilxat Muhtar, Zhenshi Li, Feng Gu, Xueliang Zhang, and Pengfeng Xiao. Lhrs-bot: Empowering remote sensing with vgi-enhanced large multimodal language model. arXiv preprint arXiv:2402.02544, 2024. 6 +[88] Wei Zhang, Miaoxin Cai, Tong Zhang, Yin Zhuang, and Xuerui Mao. Earthgpt: A universal multi-modal large language model for multi-sensor image comprehension in remote sensing domain. IEEE Transactions on Geoscience and Remote Sensing, 2024. 6 +[89] Chao Pang, Xingxing Weng, Jiang Wu, Jiayu Li, Yi Liu, Ji-axing Sun, Weijia Li, Shuai Wang, Litong Feng, Gui-Song Xia, and Conghui He. Vhm: Versatile and honest vision language model for remote sensing image analysis, 2024. 6 +[90] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26296-26306, 2024. 6 +[91] Zhenghang Yuan, Zhitong Xiong, Lichao Mou, and Xiao Xiang Zhu. Chatearthnet: a global-scale image-text dataset empowering vision-language geo-foundation models. Earth System Science Data Discussions, 2024:1-24, 2024. 6 +[92] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966, 2023. 7 +[93] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24185–24198, 2024. 7 +[94] Xiaoyi Dong, Jianmin Bao, Yinglin Zheng, Ting Zhang, Dongdong Chen, Hao Yang, Ming Zeng, Weiming Zhang, Lu Yuan, Dong Chen, et al. Maskclip: Masked self-distillation advances contrastive language-image pretraining. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10995-11005, 2023. 7 +[95] Mathieu Jacomy, Tommaso Venturini, Sebastien Heymann, and Mathieu Bastian. Forceatlas2, a continuous graph layout algorithm for handy network visualization designed for the gephi software. PLoS ONE, page e98679, Jun 2014. 8 +[96] V. A. Traag, L. Waltman, and N. J. van Eck. From louvain to leiden: guaranteeing well-connected communities. Scientific Reports, Mar 2019. 8 \ No newline at end of file diff --git a/2025/SkySense-O_ Towards Open-World Remote Sensing Interpretation with Vision-Centric Visual-Language Modeling/images.zip b/2025/SkySense-O_ Towards Open-World Remote Sensing Interpretation with Vision-Centric Visual-Language Modeling/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..274998dd3854699e8697e546e91e37591efaaae3 --- /dev/null +++ b/2025/SkySense-O_ Towards Open-World Remote Sensing Interpretation with Vision-Centric Visual-Language Modeling/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f57a359d238312b028c81906715bef6d62e7e1024d6488000593dac9a8b39b22 +size 1092854 diff --git a/2025/SkySense-O_ Towards Open-World Remote Sensing Interpretation with Vision-Centric Visual-Language Modeling/layout.json b/2025/SkySense-O_ Towards Open-World Remote Sensing Interpretation with Vision-Centric Visual-Language Modeling/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..3a42669b907567005a53f6e52c621091723a5d58 --- /dev/null +++ b/2025/SkySense-O_ Towards Open-World Remote Sensing Interpretation with Vision-Centric Visual-Language Modeling/layout.json @@ -0,0 +1,10306 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 100, + 103, + 510, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 103, + 510, + 140 + ], + "spans": [ + { + "bbox": [ + 100, + 103, + 510, + 140 + ], + "type": "text", + "content": "SkySense-O: Towards Open-World Remote Sensing Interpretation with Vision-Centric Visual-Language Modeling" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 160, + 541, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 160, + 541, + 190 + ], + "spans": [ + { + "bbox": [ + 67, + 160, + 541, + 190 + ], + "type": "text", + "content": "Qi Zhu" + }, + { + "bbox": [ + 67, + 160, + 541, + 190 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 67, + 160, + 541, + 190 + ], + "type": "text", + "content": ", Jiangwei Lao" + }, + { + "bbox": [ + 67, + 160, + 541, + 190 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 67, + 160, + 541, + 190 + ], + "type": "text", + "content": ", Deyi Ji" + }, + { + "bbox": [ + 67, + 160, + 541, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 67, + 160, + 541, + 190 + ], + "type": "text", + "content": ", Junwei Luo" + }, + { + "bbox": [ + 67, + 160, + 541, + 190 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 67, + 160, + 541, + 190 + ], + "type": "text", + "content": ", Kang Wu" + }, + { + "bbox": [ + 67, + 160, + 541, + 190 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 67, + 160, + 541, + 190 + ], + "type": "text", + "content": ", Yingying Zhang" + }, + { + "bbox": [ + 67, + 160, + 541, + 190 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 67, + 160, + 541, + 190 + ], + "type": "text", + "content": ", Lixiang Ru" + }, + { + "bbox": [ + 67, + 160, + 541, + 190 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 67, + 160, + 541, + 190 + ], + "type": "text", + "content": ", Jian Wang" + }, + { + "bbox": [ + 67, + 160, + 541, + 190 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 67, + 160, + 541, + 190 + ], + "type": "text", + "content": ", Jingdong Chen" + }, + { + "bbox": [ + 67, + 160, + 541, + 190 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 67, + 160, + 541, + 190 + ], + "type": "text", + "content": ", Ming Yang" + }, + { + "bbox": [ + 67, + 160, + 541, + 190 + ], + "type": "inline_equation", + "content": "^{6}" + }, + { + "bbox": [ + 67, + 160, + 541, + 190 + ], + "type": "text", + "content": ", Dong Liu" + }, + { + "bbox": [ + 67, + 160, + 541, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 67, + 160, + 541, + 190 + ], + "type": "text", + "content": ", Feng Zhao" + }, + { + "bbox": [ + 67, + 160, + 541, + 190 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 190, + 542, + 219 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 190, + 542, + 219 + ], + "spans": [ + { + "bbox": [ + 67, + 190, + 542, + 219 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 67, + 190, + 542, + 219 + ], + "type": "text", + "content": "MoE Key Laboratory of Brain-inspired Intelligent Perception and Cognition, USTC " + }, + { + "bbox": [ + 67, + 190, + 542, + 219 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 67, + 190, + 542, + 219 + ], + "type": "text", + "content": "Shanghai Jiaotong University " + }, + { + "bbox": [ + 67, + 190, + 542, + 219 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 67, + 190, + 542, + 219 + ], + "type": "text", + "content": "Wuhan University " + }, + { + "bbox": [ + 67, + 190, + 542, + 219 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 67, + 190, + 542, + 219 + ], + "type": "text", + "content": "UCAS " + }, + { + "bbox": [ + 67, + 190, + 542, + 219 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 67, + 190, + 542, + 219 + ], + "type": "text", + "content": "HUST " + }, + { + "bbox": [ + 67, + 190, + 542, + 219 + ], + "type": "inline_equation", + "content": "^{6}" + }, + { + "bbox": [ + 67, + 190, + 542, + 219 + ], + "type": "text", + "content": "Northwestern University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 176, + 220, + 432, + 231 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 176, + 220, + 432, + 231 + ], + "spans": [ + { + "bbox": [ + 176, + 220, + 432, + 231 + ], + "type": "text", + "content": "zqcrafts@mail.ustc.edu.cn, fzhao956@ustc.edu.cn" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 59, + 246, + 233, + 327 + ], + "blocks": [ + { + "bbox": [ + 59, + 246, + 233, + 327 + ], + "lines": [ + { + "bbox": [ + 59, + 246, + 233, + 327 + ], + "spans": [ + { + "bbox": [ + 59, + 246, + 233, + 327 + ], + "type": "image", + "image_path": "ac29e831ce3eef16d6cb0ba8287501cd003005377d7184f90069654262920e5c.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 235, + 245, + 408, + 327 + ], + "blocks": [ + { + "bbox": [ + 235, + 245, + 408, + 327 + ], + "lines": [ + { + "bbox": [ + 235, + 245, + 408, + 327 + ], + "spans": [ + { + "bbox": [ + 235, + 245, + 408, + 327 + ], + "type": "image", + "image_path": "cafd840d83c71333f31303dc6d8fa7782c3ffc484fcfec987c8bf39688dfe7ae.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 59, + 329, + 408, + 412 + ], + "blocks": [ + { + "bbox": [ + 59, + 329, + 408, + 412 + ], + "lines": [ + { + "bbox": [ + 59, + 329, + 408, + 412 + ], + "spans": [ + { + "bbox": [ + 59, + 329, + 408, + 412 + ], + "type": "image", + "image_path": "1f75e40091ff6afa8b62ed975b6f284fb81ceb8b21f8ef3eaa3e48b0ae31a506.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 422, + 555, + 444 + ], + "lines": [ + { + "bbox": [ + 55, + 422, + 555, + 444 + ], + "spans": [ + { + "bbox": [ + 55, + 422, + 555, + 444 + ], + "type": "text", + "content": "Figure 1. SkySense-O offers 4 key advantages from the foundation model to the agent application (Left), leading to state-of-the-art zero-shot performances across 14 downstream datasets over 4 distinct RS interpretation tasks (Right)" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 417, + 246, + 548, + 411 + ], + "blocks": [ + { + "bbox": [ + 417, + 246, + 548, + 411 + ], + "lines": [ + { + "bbox": [ + 417, + 246, + 548, + 411 + ], + "spans": [ + { + "bbox": [ + 417, + 246, + 548, + 411 + ], + "type": "image", + "image_path": "7f370c36aae4a526b9734964d3e354b83499cf4226d877f3b445ec89d6887498.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 152, + 460, + 200, + 472 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 460, + 200, + 472 + ], + "spans": [ + { + "bbox": [ + 152, + 460, + 200, + 472 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 54, + 484, + 297, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 484, + 297, + 694 + ], + "spans": [ + { + "bbox": [ + 54, + 484, + 297, + 694 + ], + "type": "text", + "content": "Open-world interpretation aims to accurately localize and recognize all objects within images by vision-language models (VLMs). While substantial progress has been made in this task for natural images, the advancements for remote sensing (RS) images still remain limited, primarily due to these two challenges. 1) Existing RS semantic categories are limited, particularly for pixel-level interpretation datasets. 2) Distinguishing among diverse RS spatial regions solely by language space is challenging due to the dense and intricate spatial distribution in open-world RS imagery. To address the first issue, we develop a fine-grained RS interpretation dataset, Sky-SA, which contains 183,375 high-quality local image-text pairs with full-pixel manual annotations, covering 1,763 category labels, exhibiting richer semantics and higher density than previous datasets. Afterwards, to solve the second issue, we introduce the vision-centric principle for vision-language modeling. Specifically, in the pre-training stage, the visual self-supervised paradigm is incorporated into image-" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 461, + 555, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 461, + 555, + 604 + ], + "spans": [ + { + "bbox": [ + 313, + 461, + 555, + 604 + ], + "type": "text", + "content": "text alignment, reducing the degradation of general visual representation capabilities of existing paradigms. Then, we construct a visual-relevance knowledge graph across open-category texts and further develop a novel vision-centric image-text contrastive loss for fine-tuning with text prompts. This new model, denoted as SkySense-O, demonstrates impressive zero-shot capabilities on a thorough evaluation encompassing 14 datasets over 4 tasks, from recognizing to reasoning and classification to localization. Specifically, it outperforms the latest models such as SegEarthOV, GeoRSCLIP, and VHM by a large margin, i.e., " + }, + { + "bbox": [ + 313, + 461, + 555, + 604 + ], + "type": "inline_equation", + "content": "11.95\\%" + }, + { + "bbox": [ + 313, + 461, + 555, + 604 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 461, + 555, + 604 + ], + "type": "inline_equation", + "content": "8.04\\%" + }, + { + "bbox": [ + 313, + 461, + 555, + 604 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 461, + 555, + 604 + ], + "type": "inline_equation", + "content": "3.55\\%" + }, + { + "bbox": [ + 313, + 461, + 555, + 604 + ], + "type": "text", + "content": " on average respectively. The code is publicly available to facilitate further research at https:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 314, + 605, + 494, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 605, + 494, + 616 + ], + "spans": [ + { + "bbox": [ + 314, + 605, + 494, + 616 + ], + "type": "text", + "content": "//github.com/zqcrafts/SkySense-0." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 314, + 633, + 395, + 645 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 633, + 395, + 645 + ], + "spans": [ + { + "bbox": [ + 314, + 633, + 395, + 645 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 654, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 555, + 715 + ], + "type": "text", + "content": "Remote sensing interpretation (RSI) is a critical and practical task in Earth observation, aiming to identify the semantic types of specific ground objects in remote sensing (RS) images (such as segmentation [1-4], detection [5-10], change detection [11, 12], and visual question answering" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 704, + 212, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 704, + 212, + 714 + ], + "spans": [ + { + "bbox": [ + 66, + 704, + 212, + 714 + ], + "type": "text", + "content": "*Equal contribution. † Corresponding author." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14733" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 204 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 204 + ], + "type": "text", + "content": "(VQA) [13-15]). Early works typically involved annotating RS images with fixed categories and exploring RSI through supervised learning methods. These approaches do not generalize well to unseen new categories in open-world scenarios, thus limiting their practical applicability. Moreover, as RS images becoming more accessible than ever before, there is an increasing demand for open-world interpretations with a broad category scope. Hence, there is a soaring interest in exploring a comprehensive open-world remote sensing interpretation model for many scenarios and tasks [16, 17]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 205, + 296, + 375 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 205, + 296, + 375 + ], + "spans": [ + { + "bbox": [ + 55, + 205, + 296, + 375 + ], + "type": "text", + "content": "Challenges. In the domain of natural image scenes, significant progress has been achieved in open-world interpretation tasks through large-scale localization and recognition datasets like AS-1B [18] and COCO-Stuff [19], combined with contrastive language-image pretrained model (CLIP) shown in Fig. 1a. However, the open-world RS interpretation model is still out of reach due to following two reasons: 1) From the dataset perspective, the existing annotations either exhibit sparse mask labels or lack diversity in category descriptions (see Fig.1b). 2) From the modeling perspective, the language model inherently encompasses strong contextual dependence [20], making it challenging to directly use as prompts for decoupling common RS objects and intertwined land-cover categories in open-world." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 376, + 296, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 376, + 296, + 581 + ], + "spans": [ + { + "bbox": [ + 55, + 376, + 296, + 581 + ], + "type": "text", + "content": "Sky-SA Dataset. For Struggle (1), inspired by SAM [21] and ASM [18] in the natural image domain, we propose the Sky-SA dataset with the largest number of category labels for open-world RS interpretation, including 1,763 category texts from 183,375 mask-text pairs. The Sky-SA is mainly characterized by its text openness and fine granularity (Fig. 1b). Specifically, different from general semantic segmentation datasets, the categories of Sky-SA are annotated by short semantic tags in a completely open manner. In other words, the annotators are doing fill-in-the-blank rather than multiple-choice. This enables the model to learn pixel-level open-world interpretation akin to human cognition [18] (Fig. 1c). Furthermore, similar to COCO-Stuff [19], all recognizable pixels of Sky-SA are densely annotated, which further enhances the semantic openness and the contextual relationships from wide semantic categories." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "content": "Vision-Centric Modeling. For Struggle (2), motivated by [22], we introduce the vision-centric principle to correct the potential alignment ambiguity arising from language guidance. Firstly, we initialize the visual encoder from SkySense [23], an advanced RS visual foundation model, and then collect large-scale RS image-caption datasets for pretraining, which mainly include SkyScript [23], RS5M [24] and OSM [25]. Motivated by DetailCLIP [26], to mitigate the degradation of general visual representation capabilities originating from SkySense, we maintain the visual contrastive loss and masked modeling loss during" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 316, + 72, + 548, + 270 + ], + "blocks": [ + { + "bbox": [ + 316, + 72, + 548, + 270 + ], + "lines": [ + { + "bbox": [ + 316, + 72, + 548, + 270 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 548, + 270 + ], + "type": "image", + "image_path": "f9735a931aa5e5d37fed58ed90976ebc0a3a5b445a0571e1c9735debbd8be184.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 316, + 278, + 552, + 291 + ], + "lines": [ + { + "bbox": [ + 316, + 278, + 552, + 291 + ], + "spans": [ + { + "bbox": [ + 316, + 278, + 552, + 291 + ], + "type": "text", + "content": "Figure 2. The flowchart of the overall SkySense-O development." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 304, + 555, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 304, + 555, + 484 + ], + "spans": [ + { + "bbox": [ + 313, + 304, + 555, + 484 + ], + "type": "text", + "content": "the pre-training process of image-text alignment. Secondly, for fine-tuning region-level alignment, we leverage GPT-4V [27] to construct a visual-centric knowledge graph across all categories of the Sky-SA dataset, where the node of the graph represents each category, and the edge weight indicates the visual similarity scores between each pair of categories. Using the graph, we derive a vision-centric regularization that focuses on distinguishing categories exhibiting low visual similarity but strong text-level contextual relationships in RS images. Finally, we explore the impacts of Skysense-O as agent on MLLM models (Fig. 1d). In summary, to achieve open-world RS interpretation model, SkySense-O is meticulously developed from dataset to modeling and from pre-training to fine-tuning (Fig. 2), contributions of this paper include:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 493, + 553, + 601 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 314, + 493, + 553, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 493, + 553, + 517 + ], + "spans": [ + { + "bbox": [ + 314, + 493, + 553, + 517 + ], + "type": "text", + "content": "- We develop Sky-SA dataset with large-scale and high-quality annotated texts for open-world RS interpretation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 517, + 553, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 517, + 553, + 564 + ], + "spans": [ + { + "bbox": [ + 314, + 517, + 553, + 564 + ], + "type": "text", + "content": "- The vision-centric principle is first introduced into RS visual-language modeling, primarily instantiated into training paradigm, vision-centric graph, and vision-centric regularization from pre-training to fine-tuning." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 565, + 553, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 565, + 553, + 601 + ], + "spans": [ + { + "bbox": [ + 314, + 565, + 553, + 601 + ], + "type": "text", + "content": "- We propose SkySense-O, a remarkable open-world RS interpretation model, which achieves SOTA across 14 zeroshot downstream datasets as shown in Fig. 1." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 614, + 399, + 627 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 614, + 399, + 627 + ], + "spans": [ + { + "bbox": [ + 314, + 614, + 399, + 627 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 635, + 481, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 635, + 481, + 647 + ], + "spans": [ + { + "bbox": [ + 314, + 635, + 481, + 647 + ], + "type": "text", + "content": "2.1. Remote Sensing Interpretation" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 654, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 555, + 715 + ], + "type": "text", + "content": "Remote Sensing (RS) interpretation aims to accurately identify and analyze ground objects from RS data, with applications in environmental monitoring, urban planning, and agricultural surveillance [35-42]. Early deep learning approaches focused on closed-set supervised models, in" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14734" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 70, + 298, + 208 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 298, + 208 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 298, + 208 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 298, + 208 + ], + "type": "table", + "html": "
DatasetSpatial GranularityOpen TextDense Annotation
ImageRegionPixel
Natural Image
SA-1B [21]✓✓✓
COCO-Stuff [19]✓✓✓
Remote Sensing Image
AID [28], RESISC [29]✓
OpenStreetMap [30]✓✓
RS5M [24], SkyScript [31]✓✓
Potsdam, LoveDA [32]✓✓✓
Valiggan, OEM [30]✓✓✓
iSAID [33], SIOR [34]✓✓
FAST [34], SOTA [34]✓✓
Sky-SA (ours)✓✓✓✓✓
", + "image_path": "0f23d9fc06c124ce63dad46637c2ca6b415d9552f0e18d11909350f0b077755e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 216, + 295, + 251 + ], + "lines": [ + { + "bbox": [ + 55, + 216, + 295, + 251 + ], + "spans": [ + { + "bbox": [ + 55, + 216, + 295, + 251 + ], + "type": "text", + "content": "Table 1. Property comparisons with popular vision-language datasets. The proposed Sky-SA dataset is characterized by its semantic openness, high-density and fine-granularity." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 59, + 266, + 118, + 324 + ], + "blocks": [ + { + "bbox": [ + 59, + 266, + 118, + 324 + ], + "lines": [ + { + "bbox": [ + 59, + 266, + 118, + 324 + ], + "spans": [ + { + "bbox": [ + 59, + 266, + 118, + 324 + ], + "type": "image", + "image_path": "c4e55be500c80c107790f36276420462314ecf06a1c1a15117aedbbc08bae462.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 74, + 325, + 104, + 331 + ], + "lines": [ + { + "bbox": [ + 74, + 325, + 104, + 331 + ], + "spans": [ + { + "bbox": [ + 74, + 325, + 104, + 331 + ], + "type": "text", + "content": "(a) Florence-2" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 335, + 295, + 360 + ], + "lines": [ + { + "bbox": [ + 55, + 335, + 295, + 360 + ], + "spans": [ + { + "bbox": [ + 55, + 335, + 295, + 360 + ], + "type": "text", + "content": "Figure 3. Visual comparisons between the predictions of existing open-world models and the annotations of our Sky-SA dataset." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 119, + 266, + 177, + 324 + ], + "blocks": [ + { + "bbox": [ + 119, + 266, + 177, + 324 + ], + "lines": [ + { + "bbox": [ + 119, + 266, + 177, + 324 + ], + "spans": [ + { + "bbox": [ + 119, + 266, + 177, + 324 + ], + "type": "image", + "image_path": "859f1af60096973f92f29262a473a90f95be67defa967f96822ff58d2f93d440.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 138, + 325, + 157, + 331 + ], + "lines": [ + { + "bbox": [ + 138, + 325, + 157, + 331 + ], + "spans": [ + { + "bbox": [ + 138, + 325, + 157, + 331 + ], + "type": "text", + "content": "(b) SAM" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 178, + 266, + 232, + 324 + ], + "blocks": [ + { + "bbox": [ + 178, + 266, + 232, + 324 + ], + "lines": [ + { + "bbox": [ + 178, + 266, + 232, + 324 + ], + "spans": [ + { + "bbox": [ + 178, + 266, + 232, + 324 + ], + "type": "image", + "image_path": "1ab34c65d913139baedb349e32bdbfd7694b51995461e1f85b8c7a25463ad644.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 179, + 325, + 231, + 331 + ], + "lines": [ + { + "bbox": [ + 179, + 325, + 231, + 331 + ], + "spans": [ + { + "bbox": [ + 179, + 325, + 231, + 331 + ], + "type": "text", + "content": "(c) Grounding DINO-1.6" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 232, + 266, + 291, + 324 + ], + "blocks": [ + { + "bbox": [ + 232, + 266, + 291, + 324 + ], + "lines": [ + { + "bbox": [ + 232, + 266, + 291, + 324 + ], + "spans": [ + { + "bbox": [ + 232, + 266, + 291, + 324 + ], + "type": "image", + "image_path": "7289831b60b558ca66b17c2b1f130ceffbb8dcbf3ec05239093505fae87b3bac.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 241, + 325, + 283, + 331 + ], + "lines": [ + { + "bbox": [ + 241, + 325, + 283, + 331 + ], + "spans": [ + { + "bbox": [ + 241, + 325, + 283, + 331 + ], + "type": "text", + "content": "(d) Sky-SA Dataset" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 376, + 296, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 376, + 296, + 483 + ], + "spans": [ + { + "bbox": [ + 55, + 376, + 296, + 483 + ], + "type": "text", + "content": "cluding image classification [43, 44], object detection [45-49], change detection [11, 12], and semantic segmentation [28, 50-56]. Recently, CLIP [57] and Vision Language Models (VLMs) have enabled the transition from closed-set to open-set interpretation, allowing models to recognize broader categories [17, 58]. Additionally, models like GPT-4 [59] and LLaVA [60] have facilitated the development of instruction-tuning datasets [61], enhancing the flexibility of RS models in human-machine interactions." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 501, + 295, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 501, + 295, + 514 + ], + "spans": [ + { + "bbox": [ + 55, + 501, + 295, + 514 + ], + "type": "text", + "content": "2.2. Vision Foundation Models for Remote Sensing" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 522, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 522, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 522, + 295, + 713 + ], + "type": "text", + "content": "Vision foundation models are instrumental in advancing RS interpretation toward open-set capabilities through two primary pre-training paradigms, vision-only self-supervised and vision-language self-supervised [22, 43, 62-67] paradigms. Vision-only self-supervised models utilize large-scale unannotated RS images to develop robust visual representations for various tasks [23, 68-71]. For example, SkySense [23] integrates multiple modalities and geo-contextual information within a contrastive framework, while Masked Image Modeling (MIM) methods like RingMo [72] and Scale-MAE [73] are proposed for further enhancing visual feature learning. Vision-language self-supervised models originate from CLIP. Recently some efforts such as RemoteCLIP [74], RS5M [24], and SkyScript [23] have customized for the RS domain using extensive RS image-caption pairs." + } + ] + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 325, + 72, + 561, + 213 + ], + "blocks": [ + { + "bbox": [ + 325, + 72, + 561, + 213 + ], + "lines": [ + { + "bbox": [ + 325, + 72, + 561, + 213 + ], + "spans": [ + { + "bbox": [ + 325, + 72, + 561, + 213 + ], + "type": "image", + "image_path": "95fddfe2a71f7c337796667809fa5e7e997dc4f6d907baee99dde012d71826b3.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 217, + 555, + 251 + ], + "lines": [ + { + "bbox": [ + 313, + 217, + 555, + 251 + ], + "spans": [ + { + "bbox": [ + 313, + 217, + 555, + 251 + ], + "type": "text", + "content": "Figure 4. Scale comparisons with existing RS semantic segmentation datasets. We see that the proposed Sky-SA dataset has a significantly larger category scale and diversity than prior datasets." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 314, + 273, + 410, + 287 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 273, + 410, + 287 + ], + "spans": [ + { + "bbox": [ + 314, + 273, + 410, + 287 + ], + "type": "text", + "content": "3. Sky-SA Dataset" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 314, + 293, + 389, + 306 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 293, + 389, + 306 + ], + "spans": [ + { + "bbox": [ + 314, + 293, + 389, + 306 + ], + "type": "text", + "content": "3.1. Motivation" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 312, + 555, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 312, + 555, + 529 + ], + "spans": [ + { + "bbox": [ + 313, + 312, + 555, + 529 + ], + "type": "text", + "content": "Recently, image-text alignment models based on natural images, such as the open-vocabulary methods [75] and MLLMs like LLaVA [60], have achieved remarkable success. This is primarily due to well-constructed datasets in the natural image domain compared to the RS image domain. There have been numerous high-quality datasets with rich categories in the natural image community, such as COCO-Stuff [19] and ADE [76]. However, as shown in Tab.1 and Fig.4, such resources remain scarce in the RS community, which significantly limits the development of open-world RS interpretation tasks. Moreover, existing open-world models, such as SAM [21], Florence-2 [77], and Grounding-DINO [78], are also unable to achieve dense and fine-grained interpretation on remote sensing images as shown in Fig. 3. To relieve this issue, we design a data engine to help annotators more efficiently and accurately perform full-pixel annotation on unlabeled RS images, leading to the creation of the Sky-SA dataset as following section." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 314, + 538, + 395, + 551 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 538, + 395, + 551 + ], + "spans": [ + { + "bbox": [ + 314, + 538, + 395, + 551 + ], + "type": "text", + "content": "3.2. Data Engine" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 556, + 555, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 556, + 555, + 676 + ], + "spans": [ + { + "bbox": [ + 313, + 556, + 555, + 676 + ], + "type": "text", + "content": "Open-Images Collection. In the standard collection process of a closed-set dataset, a category list is predefined, and images that do not contain any items from this list will be filtered out. This manner might lead to dataset bias [79] in remote sensing images. Unlike the close-set methods, we compile a comprehensive collection of images from various available RS datasets. To minimize redundancy in manual annotation, we leverage the SkySense model [23], which possesses image-level visual discrimination capabilities for RS images, to filter visually redundant images." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "content": "Open-Text Fine-Grained Annotation. In the annotation process, drawing inspiration from the popular COCO-Stuff [19] dataset, we prioritize fine-grained semantics and" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14735" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 85, + 255, + 210 + ], + "blocks": [ + { + "bbox": [ + 81, + 74, + 230, + 83 + ], + "lines": [ + { + "bbox": [ + 81, + 74, + 230, + 83 + ], + "spans": [ + { + "bbox": [ + 81, + 74, + 230, + 83 + ], + "type": "text", + "content": "Pre-training (For image-level RS interpretation task)" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 59, + 85, + 255, + 210 + ], + "lines": [ + { + "bbox": [ + 59, + 85, + 255, + 210 + ], + "spans": [ + { + "bbox": [ + 59, + 85, + 255, + 210 + ], + "type": "image", + "image_path": "87c6bc3c040934a2175e3cc876c78ae8a8e02f0334fbffd8dc850ac8ebf71f70.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 264, + 85, + 550, + 209 + ], + "blocks": [ + { + "bbox": [ + 318, + 74, + 492, + 83 + ], + "lines": [ + { + "bbox": [ + 318, + 74, + 492, + 83 + ], + "spans": [ + { + "bbox": [ + 318, + 74, + 492, + 83 + ], + "type": "text", + "content": "Supervised Finetuning (For pixel-level RS interpretation task)" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 264, + 85, + 550, + 209 + ], + "lines": [ + { + "bbox": [ + 264, + 85, + 550, + 209 + ], + "spans": [ + { + "bbox": [ + 264, + 85, + 550, + 209 + ], + "type": "image", + "image_path": "b684b261ae0159528d91049f1705c708a9f7cb3447dcda8ef131967d9a14a7c6.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 217, + 553, + 229 + ], + "lines": [ + { + "bbox": [ + 55, + 217, + 553, + 229 + ], + "spans": [ + { + "bbox": [ + 55, + 217, + 553, + 229 + ], + "type": "text", + "content": "Figure 5. The framework of SkySense-O, where training strategy includes pre-training and fine-tuning stage for open-world interpretation." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 241, + 296, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 241, + 296, + 420 + ], + "spans": [ + { + "bbox": [ + 54, + 241, + 296, + 420 + ], + "type": "text", + "content": "full-pixel annotation as core labeling principles. Specifically, there are two key points: 1) the dataset encompasses a rich variety of semantic categories; 2) almost all pixels in each image need to be annotated. The former aids in activating the broad semantic awareness of the vision-language pretrained model, while the latter helps establish comprehensive connections among local semantics throughout each image. Furthermore, to capture a broader semantic range, we employ a short-caption labeling manner for each region, which does not constrain the annotation labels. Additionally, to improve annotation efficiency, we first utilize GPT-4V [27] to generate the image content, and then employ Grounding-DINO [78] and SAM [21] for pre-annotation, allowing annotators to make careful corrections based on these preliminary labels." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 422, + 296, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 422, + 296, + 554 + ], + "spans": [ + { + "bbox": [ + 55, + 422, + 296, + 554 + ], + "type": "text", + "content": "Dataset Statistics. Finally, we develop a high-quality pixel-level remote sensing interpretation dataset, named Sky-SA. As shown in Fig. 4, the Sky-SA dataset consists of 35,000 RS images, mainly including 1,763 category texts from 183,375 segment-text pairs (encompassing approximately 500 distinct categories of land-covering and RS objects after synonym filtering), which far surpasses existing remote sensing segmentation datasets. Moreover, Sky-SA is also the first pixel-level open-text annotated dataset in the field of remote sensing (see Fig. 1), laying a solid data foundation for the development of open-world RS interpretation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 567, + 167, + 580 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 567, + 167, + 580 + ], + "spans": [ + { + "bbox": [ + 55, + 567, + 167, + 580 + ], + "type": "text", + "content": "4. SkySense-O Model" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 587, + 237, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 587, + 237, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 587, + 237, + 601 + ], + "type": "text", + "content": "4.1. Pretraining for Foundation Model" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "content": "The training of SkySense-O is primarily divided into pretraining and fine-tuning stages as shown in Fig. 5. During pre-training, given the input image " + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "inline_equation", + "content": "I_{i}" + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "content": " and its corresponding text " + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "inline_equation", + "content": "T_{i}" + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "content": ", for " + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "inline_equation", + "content": "i = 1,\\dots,B" + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "content": " denotes mini-batch size, the model mainly consists of an image encoder " + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "inline_equation", + "content": "E_{img}" + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "content": " and a text encoder " + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "inline_equation", + "content": "E_{text}" + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "content": ". To maintain the model's visual perception capabilities [26], especially for densely distributed objects during the image-text alignment process, the vision encoder is initialized by SkySense [23] and its visual" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 241, + 554, + 290 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 241, + 554, + 290 + ], + "spans": [ + { + "bbox": [ + 313, + 241, + 554, + 290 + ], + "type": "text", + "content": "self-supervised paradigms are still maintained, i.e. masked image modeling and image contrastive learning. Given " + }, + { + "bbox": [ + 313, + 241, + 554, + 290 + ], + "type": "inline_equation", + "content": "x = E_{img}(I)" + }, + { + "bbox": [ + 313, + 241, + 554, + 290 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 241, + 554, + 290 + ], + "type": "inline_equation", + "content": "y = E_{text}(T)" + }, + { + "bbox": [ + 313, + 241, + 554, + 290 + ], + "type": "text", + "content": ", mask decoder output " + }, + { + "bbox": [ + 313, + 241, + 554, + 290 + ], + "type": "inline_equation", + "content": "R'" + }, + { + "bbox": [ + 313, + 241, + 554, + 290 + ], + "type": "text", + "content": ", the overall pretraining loss is computed as Eq. 1:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 324, + 299, + 553, + 313 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 299, + 553, + 313 + ], + "spans": [ + { + "bbox": [ + 324, + 299, + 553, + 313 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {p r e t r a i n}} = C E (x, y) + C E \\left(x, x ^ {\\prime}\\right) + M R \\left(I ^ {\\prime}, R ^ {\\prime}\\right), \\tag {1}", + "image_path": "d9ecc64bcf1bc1bf92a1e8925e08d548598cfc3096a56a62283f2257518cf7b6.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 319, + 554, + 355 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 319, + 554, + 355 + ], + "spans": [ + { + "bbox": [ + 313, + 319, + 554, + 355 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 319, + 554, + 355 + ], + "type": "inline_equation", + "content": "I^{\\prime}" + }, + { + "bbox": [ + 313, + 319, + 554, + 355 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 319, + 554, + 355 + ], + "type": "inline_equation", + "content": "x^{\\prime}" + }, + { + "bbox": [ + 313, + 319, + 554, + 355 + ], + "type": "text", + "content": " are the augmented version of " + }, + { + "bbox": [ + 313, + 319, + 554, + 355 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 313, + 319, + 554, + 355 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 319, + 554, + 355 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 319, + 554, + 355 + ], + "type": "text", + "content": " respectively, " + }, + { + "bbox": [ + 313, + 319, + 554, + 355 + ], + "type": "inline_equation", + "content": "MR" + }, + { + "bbox": [ + 313, + 319, + 554, + 355 + ], + "type": "text", + "content": " is mask reconstruction loss following [71] and " + }, + { + "bbox": [ + 313, + 319, + 554, + 355 + ], + "type": "inline_equation", + "content": "CE" + }, + { + "bbox": [ + 313, + 319, + 554, + 355 + ], + "type": "text", + "content": " denotes cross-entropy loss computing as:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 320, + 360, + 553, + 393 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 360, + 553, + 393 + ], + "spans": [ + { + "bbox": [ + 320, + 360, + 553, + 393 + ], + "type": "interline_equation", + "content": "C E (x, y) = - \\frac {1}{B} \\sum_ {i = 1} ^ {B} \\log \\frac {\\exp \\left(\\operatorname {s i m} \\left(x _ {i} , y _ {i}\\right) / \\tau\\right)}{\\sum_ {j = 1} ^ {B} \\exp \\left(\\operatorname {s i m} \\left(x _ {i} , y _ {j}\\right) / \\tau\\right)}, \\tag {2}", + "image_path": "e130ae771b81fe49315db7d69fb32e9e4ac3470fa5f6b64256f926bc3e1fc82a.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 398, + 545, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 398, + 545, + 410 + ], + "spans": [ + { + "bbox": [ + 313, + 398, + 545, + 410 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 398, + 545, + 410 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 313, + 398, + 545, + 410 + ], + "type": "text", + "content": " is logit temperature and " + }, + { + "bbox": [ + 313, + 398, + 545, + 410 + ], + "type": "inline_equation", + "content": "sim" + }, + { + "bbox": [ + 313, + 398, + 545, + 410 + ], + "type": "text", + "content": " is cosine similarity." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 416, + 507, + 428 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 416, + 507, + 428 + ], + "spans": [ + { + "bbox": [ + 313, + 416, + 507, + 428 + ], + "type": "text", + "content": "4.2. Finetuning for Interpretation Model" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 434, + 554, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 434, + 554, + 518 + ], + "spans": [ + { + "bbox": [ + 313, + 434, + 554, + 518 + ], + "type": "text", + "content": "Region-level Finetuning. During the region-level finetuning stage, based on the proposed Sky-SA dataset, given a set of candidate category texts " + }, + { + "bbox": [ + 313, + 434, + 554, + 518 + ], + "type": "inline_equation", + "content": "C_n" + }, + { + "bbox": [ + 313, + 434, + 554, + 518 + ], + "type": "text", + "content": " and masks " + }, + { + "bbox": [ + 313, + 434, + 554, + 518 + ], + "type": "inline_equation", + "content": "M_n" + }, + { + "bbox": [ + 313, + 434, + 554, + 518 + ], + "type": "text", + "content": " for each image, where " + }, + { + "bbox": [ + 313, + 434, + 554, + 518 + ], + "type": "inline_equation", + "content": "n = 1, \\dots, N" + }, + { + "bbox": [ + 313, + 434, + 554, + 518 + ], + "type": "text", + "content": ", we apply mask pooling to extract local features for each category from the images, and then compute the region-level cross-entropy loss with the corresponding category, which can be expressed as:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 321, + 523, + 553, + 556 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 523, + 553, + 556 + ], + "spans": [ + { + "bbox": [ + 321, + 523, + 553, + 556 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {r e g i o n}} = \\sum_ {n = 1} ^ {N} C E \\left(P o o l i n g \\left(x _ {n}, M _ {n}\\right), E _ {\\text {t e x t}} \\left(C _ {n}\\right)\\right). \\tag {3}", + "image_path": "f65fa767fc9182bc0f1bff0a6df4409e2160a4fc9b67b56ea2e5c44d7d0ba22d.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 562, + 554, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 562, + 554, + 658 + ], + "spans": [ + { + "bbox": [ + 313, + 562, + 554, + 658 + ], + "type": "text", + "content": "Pixel-level Finetuning. To achieve pixel-level interpretation, referring to the existing open-vocabulary semantic segmentation model [80], we first compute the similarity between the image encoder and the text encoder at the token level to obtain coarse location features. Subsequently, a visual decoder " + }, + { + "bbox": [ + 313, + 562, + 554, + 658 + ], + "type": "inline_equation", + "content": "D_{img}" + }, + { + "bbox": [ + 313, + 562, + 554, + 658 + ], + "type": "text", + "content": " is employed to up-sample these features to the fine-grained segmentation results, constrained by a pixel-wise cross-entropy loss as Eq. 4:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 329, + 663, + 553, + 696 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 329, + 663, + 553, + 696 + ], + "spans": [ + { + "bbox": [ + 329, + 663, + 553, + 696 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {p i x e l}} = \\sum_ {n = 1} ^ {N} C E \\left(D _ {\\text {i m g}} \\left(x _ {n}, E _ {\\text {t e x t}} \\left(C _ {n}\\right)\\right), M _ {n}\\right). \\tag {4}", + "image_path": "4309990e430845d81c3fb53c8d7bcec022bc40d4c5f82747a80d5ea2f167243c.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 701, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 701, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 701, + 547, + 713 + ], + "type": "text", + "content": "More details can be found in the supplementary and code." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14736" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 70, + 294, + 194 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 294, + 194 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 294, + 194 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 294, + 194 + ], + "type": "image", + "image_path": "17864a17972ce502161b7bed589dc73601d06f30ee3db7dfcfe6d9edfaee4274.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 198, + 296, + 264 + ], + "lines": [ + { + "bbox": [ + 55, + 198, + 296, + 264 + ], + "spans": [ + { + "bbox": [ + 55, + 198, + 296, + 264 + ], + "type": "text", + "content": "Figure 6. Positive and negative sampling mechanism according to the correspondence score between source sample and target sample. (a): Due to the contextual dependence in language modeling or vision-language modeling [20], \"road\" is mistakenly assigned as the positive sample of \"car\", while our agent employs the correct sampler. (b): Another case to further valid this point." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 59, + 275, + 297, + 367 + ], + "blocks": [ + { + "bbox": [ + 59, + 275, + 297, + 367 + ], + "lines": [ + { + "bbox": [ + 59, + 275, + 297, + 367 + ], + "spans": [ + { + "bbox": [ + 59, + 275, + 297, + 367 + ], + "type": "image", + "image_path": "91734500729032caa9ad2f494192bfc436d603a03c7766304614a8aa70a24d37.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 373, + 295, + 396 + ], + "lines": [ + { + "bbox": [ + 55, + 373, + 295, + 396 + ], + "spans": [ + { + "bbox": [ + 55, + 373, + 295, + 396 + ], + "type": "text", + "content": "Figure 7. Constructing illustrations of the vision-centric knowledge graph by an elaborate vision-centric agent flow." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 408, + 194, + 421 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 408, + 194, + 421 + ], + "spans": [ + { + "bbox": [ + 55, + 408, + 194, + 421 + ], + "type": "text", + "content": "4.3. Vision-Centric Modeling" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 426, + 296, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 426, + 296, + 533 + ], + "spans": [ + { + "bbox": [ + 55, + 426, + 296, + 533 + ], + "type": "text", + "content": "Challenge. Different from the previous close-set semantic segmentation dataset, the Sky-SA dataset is annotated in open-world texts. Although this manner provides more open vocabulary support, the ambiguous sampling challenge often leads to training collapse in discriminative models. As illustrated in Fig. 6, \"car\" is mistakenly regarded as the positive sample of \"road\" by computing the similarity between their text-embeddings due to their strong text-level or image-level contextual relationships." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 534, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 296, + 714 + ], + "type": "text", + "content": "Visual-Centric Knowledge Graph. To overcome the challenge, inspired by the in-depth exploration of visual representation by Xie et al. [22] and GraphRAG [81], we build a visual-centric agent using GPT-4 to obtain the visual relevance score for each text pair of the Sky-SA dataset. As shown in Fig. 7, the score is measured by the meticulously chain of thought about visual attributes, including appearance, color, and shape etc.. When all text pairs in Sky-SA are fully traversed, we can construct an offline vision-centric knowledge graph, which mainly consists of three elements: source node, target node, and edge. The edge indicates the visual-relevance score and reasons from source node to target node. Based on the graph, we perform flexible sampling for positive and negative samples by setting a sampling threshold illustrated in Algorithm 1." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 315, + 73, + 507, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 73, + 507, + 84 + ], + "spans": [ + { + "bbox": [ + 315, + 73, + 507, + 84 + ], + "type": "text", + "content": "Algorithm 1 Finetuning Stage for SkySense-O." + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 315, + 86, + 563, + 118 + ], + "blocks": [ + { + "bbox": [ + 315, + 86, + 563, + 118 + ], + "lines": [ + { + "bbox": [ + 315, + 86, + 563, + 118 + ], + "spans": [ + { + "bbox": [ + 315, + 86, + 563, + 118 + ], + "type": "image", + "image_path": "480fbc1faac760c1d7070e935b259045d46bcb58c6a2bbd481836f89834efeac.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 316, + 123, + 532, + 178 + ], + "blocks": [ + { + "bbox": [ + 316, + 123, + 532, + 178 + ], + "lines": [ + { + "bbox": [ + 316, + 123, + 532, + 178 + ], + "spans": [ + { + "bbox": [ + 316, + 123, + 532, + 178 + ], + "type": "image", + "image_path": "0288e880412cb19e22e8d16ceceacc632d0063a0f2a128dce5c2777f4b4e5ace.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 316, + 179, + 541, + 250 + ], + "blocks": [ + { + "bbox": [ + 316, + 179, + 541, + 250 + ], + "lines": [ + { + "bbox": [ + 316, + 179, + 541, + 250 + ], + "spans": [ + { + "bbox": [ + 316, + 179, + 541, + 250 + ], + "type": "image", + "image_path": "5470adb042dc2cb28ad2c31c09d08e5c552a84e5288fcbac3db1e8ee2a7dac10.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 320, + 259, + 555, + 365 + ], + "blocks": [ + { + "bbox": [ + 320, + 259, + 555, + 365 + ], + "lines": [ + { + "bbox": [ + 320, + 259, + 555, + 365 + ], + "spans": [ + { + "bbox": [ + 320, + 259, + 555, + 365 + ], + "type": "image", + "image_path": "21c5c183a66a669e347df6911ddd44d087b6cc640fedc722dbc1af3bf839095d.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 373, + 555, + 407 + ], + "lines": [ + { + "bbox": [ + 313, + 373, + 555, + 407 + ], + "spans": [ + { + "bbox": [ + 313, + 373, + 555, + 407 + ], + "type": "text", + "content": "Figure 8. Sample similarity matrices in computing vision-centric regularization. Hard samples, such as vehicle and road, are exploited by multiplying " + }, + { + "bbox": [ + 313, + 373, + 555, + 407 + ], + "type": "inline_equation", + "content": "S_{t}" + }, + { + "bbox": [ + 313, + 373, + 555, + 407 + ], + "type": "text", + "content": " and inversed " + }, + { + "bbox": [ + 313, + 373, + 555, + 407 + ], + "type": "inline_equation", + "content": "S_{v}" + }, + { + "bbox": [ + 313, + 373, + 555, + 407 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 418, + 555, + 587 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 418, + 555, + 587 + ], + "spans": [ + { + "bbox": [ + 313, + 418, + 555, + 587 + ], + "type": "text", + "content": "Vision-Centric Regularization Loss. In RS imagery, the objects and land-cover are often densely packed and intricately intertwined, posing challenges for fine-grained visual-language modeling. To relieve this issue, we further exploit the potential of the vision-centric graph and design a vision-centric regularization loss, " + }, + { + "bbox": [ + 313, + 418, + 555, + 587 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{vc}}" + }, + { + "bbox": [ + 313, + 418, + 555, + 587 + ], + "type": "text", + "content": ". The regularization is built on the observation that the samples with strong contextual relationships but weak visual relevance scores are hard to distinguish as shown in Fig. 6. Specifically, given selected " + }, + { + "bbox": [ + 313, + 418, + 555, + 587 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 418, + 555, + 587 + ], + "type": "text", + "content": " negative samples " + }, + { + "bbox": [ + 313, + 418, + 555, + 587 + ], + "type": "inline_equation", + "content": "C_m" + }, + { + "bbox": [ + 313, + 418, + 555, + 587 + ], + "type": "text", + "content": " for class " + }, + { + "bbox": [ + 313, + 418, + 555, + 587 + ], + "type": "inline_equation", + "content": "C_i" + }, + { + "bbox": [ + 313, + 418, + 555, + 587 + ], + "type": "text", + "content": ", their textual similarities " + }, + { + "bbox": [ + 313, + 418, + 555, + 587 + ], + "type": "inline_equation", + "content": "S_t" + }, + { + "bbox": [ + 313, + 418, + 555, + 587 + ], + "type": "text", + "content": " are computed by their embedding similarities. Then, given text pairs, the visual relevance " + }, + { + "bbox": [ + 313, + 418, + 555, + 587 + ], + "type": "inline_equation", + "content": "S_v^{i,m}" + }, + { + "bbox": [ + 313, + 418, + 555, + 587 + ], + "type": "text", + "content": " can be retrieved by the visual-centric graph. Finally, " + }, + { + "bbox": [ + 313, + 418, + 555, + 587 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{vc}}" + }, + { + "bbox": [ + 313, + 418, + 555, + 587 + ], + "type": "text", + "content": " is computed as Eq. 5 and details are in Algorithm 1 and Fig. 8." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 321, + 595, + 555, + 628 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 595, + 555, + 628 + ], + "spans": [ + { + "bbox": [ + 321, + 595, + 555, + 628 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {v c}} = \\frac {1}{B N} \\sum_ {i = 1} ^ {B} \\sum_ {n = 1} ^ {N} \\left\\| \\left(1 - S _ {v} ^ {i, m}\\right) * S _ {t} ^ {i, m} * \\mathcal {L} _ {\\text {r e g i o n}} \\right\\| ^ {2}. \\tag {5}", + "image_path": "f66809f6f62e2ded8ae99a05ca19cedd014cb977cbc91e39a3df81c433753b84.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 635, + 555, + 659 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 635, + 555, + 659 + ], + "spans": [ + { + "bbox": [ + 313, + 635, + 555, + 659 + ], + "type": "text", + "content": "Total Loss. In summary, the total loss for the finetuning stage of SkySense-O is defined as follows:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 346, + 668, + 553, + 682 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 346, + 668, + 553, + 682 + ], + "spans": [ + { + "bbox": [ + 346, + 668, + 553, + 682 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {f i n e t u n e}} = \\mathcal {L} _ {\\text {p i x e l}} + \\lambda_ {1} \\mathcal {L} _ {\\text {r e g i o n}} + \\lambda_ {2} \\mathcal {L} _ {\\text {v c}}, \\tag {6}", + "image_path": "c9ee2619d41249f038c9bdcefcf184eacd5beecfca7a33382f218f509c287cf5.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\lambda_{1}" + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\lambda_{2}" + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": " are two hyperparameters and set as 0.4 and 0.4 respectively for optimal ratio via performing ablations." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14737" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 57, + 70, + 555, + 182 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 555, + 182 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 555, + 182 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 555, + 182 + ], + "type": "table", + "html": "
ModelPublicationObject Semantic SegmentationLand-Cover Semantic Segmentation
iSAID [33]SOTA [34]FAST [34]SIOR [34]Potsdam [82]OEM [30]LoveDA [32]VaihingenFloodNet [83]
Few-Shot Semantic Segmentation
SegGPT [84]ICCV 202329.1731.4119.5537.4730.1225.1931.4047.2733.81
DINOv [85]CVPR 202425.3518.247.9225.0437.6523.0525.1735.1044.55
Zero-Shot Semantic Segmentation
SAN [80]CVPR 202311.7710.733.1615.2251.4524.8727.8346.2339.26
CAT-SEG [75]CVPR 202420.5813.713.1319.6450.8130.2637.1549.6241.36
SegEarth-OV [58]Arxiv 202421.7011.992.8413.9247.1040.3036.9029.1034.00
SkySense-O-43.9232.128.3330.8954.1040.8338.3051.6442.04
-(+22.22)(+18.41)(+5.2)(+11.25)(+2.65)(+0.53)(+1.15)(+2.02)(+0.68)
", + "image_path": "d4d606cf15467ca06cddbc56cd9907fd47cbafb8d23458b2052d5f27d2dc0404.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 58, + 213, + 341, + 297 + ], + "blocks": [ + { + "bbox": [ + 55, + 185, + 555, + 209 + ], + "lines": [ + { + "bbox": [ + 55, + 185, + 555, + 209 + ], + "spans": [ + { + "bbox": [ + 55, + 185, + 555, + 209 + ], + "type": "text", + "content": "Table 2. Zero-shot remote sensing object and land-cover semantic segmentation performances. Bold indicates the highest performance, and underline indicates the second highest performance in mIoU (%)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 213, + 341, + 297 + ], + "lines": [ + { + "bbox": [ + 58, + 213, + 341, + 297 + ], + "spans": [ + { + "bbox": [ + 58, + 213, + 341, + 297 + ], + "type": "table", + "html": "
ModelPublicationSIRI-WHU [86]AID-VQA [28]RSVQA-HR [14]
Acc.Acc.Avg. Acc.
GeoChat [35]CVPR 202466.6372.0370.82
LHRS-Bot [87]ECCV 202462.6691.26-
EarthGPT [88]TGRS 2024--71.15
VHM [89]Arxiv 202470.8891.7073.75
SkySense-O†-72.45 (+1.57)92.95 (+1.25)75.39 (+1.64)
SkySense-O-74.79 (+3.91)94.10 (+2.40)78.09 (+4.34)
", + "image_path": "9f9dc6cacc86d40275dc5fc316720fd52ac4d400b6167fc020bcd35947915e18.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 347, + 213, + 552, + 297 + ], + "blocks": [ + { + "bbox": [ + 55, + 300, + 341, + 324 + ], + "lines": [ + { + "bbox": [ + 55, + 300, + 341, + 324 + ], + "spans": [ + { + "bbox": [ + 55, + 300, + 341, + 324 + ], + "type": "text", + "content": "Table 3. Zero-Shot visual question-answering performances, compared with existing MLLMs. " + }, + { + "bbox": [ + 55, + 300, + 341, + 324 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 55, + 300, + 341, + 324 + ], + "type": "text", + "content": " indicates wo/ pixel-level fine-tuning by Sky-SA." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 347, + 213, + 552, + 297 + ], + "lines": [ + { + "bbox": [ + 347, + 213, + 552, + 297 + ], + "spans": [ + { + "bbox": [ + 347, + 213, + 552, + 297 + ], + "type": "table", + "html": "
ModelPublicationRESISC [29]AID [28]
CLIP [57]PMLR 202169.2566.70
RemoteCLIP [74]TGRS 202470.8574.3
CLIP-laion-RS [31]-71.7071.21
SkyScript [31]AAA1 202471.7070.94
GeoRSCLIP [24]TGRS 202373.9076.27
SkySense-O†-83.2882.93
-(+9.38)(+6.66)
", + "image_path": "fc10b90c6ec4c0d32e1c088ae8dba29738bb0880803e6e86316f879126d265bb.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 345, + 300, + 552, + 323 + ], + "lines": [ + { + "bbox": [ + 345, + 300, + 552, + 323 + ], + "spans": [ + { + "bbox": [ + 345, + 300, + 552, + 323 + ], + "type": "text", + "content": "Table 4. Zero-shot remote sensing scene classification performances in Top-1 accuracy " + }, + { + "bbox": [ + 345, + 300, + 552, + 323 + ], + "type": "inline_equation", + "content": "(\\%)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 338, + 132, + 351 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 338, + 132, + 351 + ], + "spans": [ + { + "bbox": [ + 55, + 338, + 132, + 351 + ], + "type": "text", + "content": "5. Experiment" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 358, + 191, + 369 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 358, + 191, + 369 + ], + "spans": [ + { + "bbox": [ + 55, + 358, + 191, + 369 + ], + "type": "text", + "content": "5.1. Evaluation Benchmarks" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 376, + 295, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 376, + 295, + 448 + ], + "spans": [ + { + "bbox": [ + 55, + 376, + 295, + 448 + ], + "type": "text", + "content": "We evaluate SkySense-O on four representative RS interpretation tasks. To ensure fair zero-shot comparisons, all experiments are conducted without extra fine-tuning and labels from the evaluation datasets are not used during training. The supplementary materials provide further details on the implementation and visualization results." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 449, + 295, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 449, + 295, + 555 + ], + "spans": [ + { + "bbox": [ + 55, + 449, + 295, + 555 + ], + "type": "text", + "content": "Zero-shot object segmentation. We adopt iSAID [33], FAST [34], SIOR [34], SOTA [34] for the object segmentation experiment. They are chosen considering diverse scenes and categories. We report mIoU results on official validation sets. As depicted in Tab. 2, SkySense-O has achieved the SOTA performance on all object segmentation datasets. On average, it outperforms the previous SOTAs by an average margin of " + }, + { + "bbox": [ + 55, + 449, + 295, + 555 + ], + "type": "inline_equation", + "content": "14.27\\%" + }, + { + "bbox": [ + 55, + 449, + 295, + 555 + ], + "type": "text", + "content": ". Furthermore, it even surpasses few-shot SOTAs on most datasets." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 557, + 295, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 557, + 295, + 665 + ], + "spans": [ + { + "bbox": [ + 55, + 557, + 295, + 665 + ], + "type": "text", + "content": "Zero-shot land-cover segmentation. We assess SkySense-O's land-cover segmentation performance on ISPRS Potsdam [82], Vaihingen and LoveDA [32] datasets. SkySense-O excels on all three datasets (Tab. 2). Notably, we surpass the second best CAT-SEG [75] by " + }, + { + "bbox": [ + 55, + 557, + 295, + 665 + ], + "type": "inline_equation", + "content": "1.41\\%" + }, + { + "bbox": [ + 55, + 557, + 295, + 665 + ], + "type": "text", + "content": " mIoU and have achieved the best performance on the Vaihingen leaderboard. More importantly, our results demonstrate the potential to achieve more accurate interpretations than the ground truth (the second row in Fig. 9)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "Zero-shot classification. We utilize two scene classification datasets: AID [28] and RESISC-45 [29]. The validation setting follows GeoRSCLIP [24] and Top-1 Accuracy is reported. SkySense-O overall outperforms competitive" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 317, + 339, + 555, + 466 + ], + "blocks": [ + { + "bbox": [ + 317, + 339, + 555, + 466 + ], + "lines": [ + { + "bbox": [ + 317, + 339, + 555, + 466 + ], + "spans": [ + { + "bbox": [ + 317, + 339, + 555, + 466 + ], + "type": "image", + "image_path": "ce92ee71221c85e5c5c28022eff8a3f91a630d76d827870c55ce37b3a724f34e.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 470, + 555, + 492 + ], + "lines": [ + { + "bbox": [ + 313, + 470, + 555, + 492 + ], + "spans": [ + { + "bbox": [ + 313, + 470, + 555, + 492 + ], + "type": "text", + "content": "Figure 9. Visualization of object segmentation results on iSAID dataset and land-cover segmentation results on Potsdam dataset." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 506, + 555, + 530 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 506, + 555, + 530 + ], + "spans": [ + { + "bbox": [ + 313, + 506, + 555, + 530 + ], + "type": "text", + "content": "basielines, achieving " + }, + { + "bbox": [ + 313, + 506, + 555, + 530 + ], + "type": "inline_equation", + "content": "9.38\\%" + }, + { + "bbox": [ + 313, + 506, + 555, + 530 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 506, + 555, + 530 + ], + "type": "inline_equation", + "content": "6.66\\%" + }, + { + "bbox": [ + 313, + 506, + 555, + 530 + ], + "type": "text", + "content": " higher performance than state-of-the-art methods on the two datasets (Tab. 4)." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 533, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 556, + 713 + ], + "type": "text", + "content": "Zero-shot RS VQA. To further validate the versatility of SkySense-O, we employ it in MLLM as a vision encoder for experimentation. Specifically, we maintain the same architecture as LLaVA-1.5 [90], except that we replace the original CLIP-L-14 [57] with our SkySense-O, and set the image resolution to " + }, + { + "bbox": [ + 313, + 533, + 556, + 713 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 313, + 533, + 556, + 713 + ], + "type": "text", + "content": " pixels. Regarding the setup of the instruction tuning data, during the pretraining stage, we utilize CC-558K dataset [60] and 121k image-text pairs converted from ChatearthNet [91] dataset. In the finetuning stage, based on instruction-tuning datasets collected by SkySenseGPT [61] from different sources, we select a subset of 593k data as our downstream instruction finetuning dataset, where details are in the supplementary. We evaluate zero-shot VQA performances on several public RS benchmarks as shown in Tab. 3." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14738" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 71, + 553, + 257 + ], + "blocks": [ + { + "bbox": [ + 58, + 71, + 553, + 257 + ], + "lines": [ + { + "bbox": [ + 58, + 71, + 553, + 257 + ], + "spans": [ + { + "bbox": [ + 58, + 71, + 553, + 257 + ], + "type": "image", + "image_path": "a82e7018650b4045c67610b2640045e4331fb8fc96341dc8616e07233e780289.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 267, + 555, + 301 + ], + "lines": [ + { + "bbox": [ + 55, + 267, + 555, + 301 + ], + "spans": [ + { + "bbox": [ + 55, + 267, + 555, + 301 + ], + "type": "text", + "content": "Figure 10. Visualization of text-image alignment in CLIP. (a) Computing the similarity between each text token and all image tokens to obtain text-image relevance (TIR) maps for different categories. (b) Visualizing TIR maps for some high-response texts. (c) Histogram of the maximum response values of TIR maps across the text list. The details of input texts can be referred to the supplementary materials." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 58, + 309, + 298, + 376 + ], + "blocks": [ + { + "bbox": [ + 58, + 309, + 298, + 376 + ], + "lines": [ + { + "bbox": [ + 58, + 309, + 298, + 376 + ], + "spans": [ + { + "bbox": [ + 58, + 309, + 298, + 376 + ], + "type": "table", + "html": "
Foundation \nModelVision \nEncoderVision-Enc. \nParameterNum. of \nImage-textRESISC \n(Acc)
RemoteCLIPViT-L-14304M0.83M70.85
SkyScriptViT-L-14304M2.6M71.70
GeoRSCLIPViT-H-14630M5M73.90
OursSwinV2-H-8654M10M83.28
", + "image_path": "c97c21c25eed9a59c9115945d4cc3d6ddb619d12401ef9c8752774ebf4304bad.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 58, + 423, + 291, + 502 + ], + "blocks": [ + { + "bbox": [ + 55, + 380, + 295, + 414 + ], + "lines": [ + { + "bbox": [ + 55, + 380, + 295, + 414 + ], + "spans": [ + { + "bbox": [ + 55, + 380, + 295, + 414 + ], + "type": "text", + "content": "Table 5. Comparison with existing RS vision-language foundation models. SkySense-O achieves consistent performance gains due to scaling laws in terms of model-size and data-size." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 423, + 291, + 502 + ], + "lines": [ + { + "bbox": [ + 58, + 423, + 291, + 502 + ], + "spans": [ + { + "bbox": [ + 58, + 423, + 291, + 502 + ], + "type": "table", + "html": "
+Sky-SA+Graph+Pretrain-ing+LvcmIoU
18.40
✓27.76 (+9.36)
✓✓36.28 (+8.52)
✓✓✓41.69 (+5.41)
✓✓✓✓43.92 (+2.24)
", + "image_path": "f3c42c64cd997aab01a37fc9957da8cbf4505bf960082ac527eeda38bc148735.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 64, + 506, + 285, + 517 + ], + "lines": [ + { + "bbox": [ + 64, + 506, + 285, + 517 + ], + "spans": [ + { + "bbox": [ + 64, + 506, + 285, + 517 + ], + "type": "text", + "content": "Table 6. Ablation study on the proposed dataset and designs." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 534, + 191, + 547 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 191, + 547 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 191, + 547 + ], + "type": "text", + "content": "5.2. Implementation Details." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 552, + 296, + 660 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 552, + 296, + 660 + ], + "spans": [ + { + "bbox": [ + 55, + 552, + 296, + 660 + ], + "type": "text", + "content": "In the pre-training stage, we use 80 A100-80GB GPUs and primarily adopt two large-scale remote sensing image-text datasets, SkyScript [31] and RS5M [24], and further expand their captions using Qwen-VL [92] and InternVL [93], comprising 10 million image-text pairs. In the fine-tuning stage, we employ OSM [25] for local image-text fine-tuning and Sky-SA for fine-grained interpretation capabilities on 8 A100-80GB GPUs as shown in Tab. 5. More details are included in the supplementary materials." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 671, + 224, + 683 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 671, + 224, + 683 + ], + "spans": [ + { + "bbox": [ + 55, + 671, + 224, + 683 + ], + "type": "text", + "content": "5.3. Discussions & Ablation Studies" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": "Effectiveness of Image-Language Alignment. In Fig. 10, we follow MaskCLIP [94] and adopt the image-text correla" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 316, + 309, + 554, + 431 + ], + "blocks": [ + { + "bbox": [ + 316, + 309, + 554, + 431 + ], + "lines": [ + { + "bbox": [ + 316, + 309, + 554, + 431 + ], + "spans": [ + { + "bbox": [ + 316, + 309, + 554, + 431 + ], + "type": "image", + "image_path": "c15e0225292e0a27df5f81a9961f5b4d4f7e210bd01863fa8810eaa63d455cac.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 439, + 555, + 462 + ], + "lines": [ + { + "bbox": [ + 313, + 439, + 555, + 462 + ], + "spans": [ + { + "bbox": [ + 313, + 439, + 555, + 462 + ], + "type": "text", + "content": "Figure 11. Visualizations of open-world interpretation results by SkySense-O on iSAID, Potsdam, FAST, and SIOR datasets." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 483, + 555, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 483, + 555, + 625 + ], + "spans": [ + { + "bbox": [ + 313, + 483, + 555, + 625 + ], + "type": "text", + "content": "tion in the representation space to evaluate the effectiveness of vision-language foundation models. We select the CLIP models from SegEarth-OV [58] and CAT-Seg [75] for comparison. The former is a standard CLIP model, while the latter is a locally fine-tuned CLIP model by COCO-Stuff [19]. We randomly selected 131 category texts from Sky-SA as inputs. As shown in Figure 10(c), SkySense-O exhibits larger response variance across different categories, indicating stronger discriminative capabilities among different texts. In Fig. 10(b), we present several image-text correlation maps to further validate the effectiveness of SkySense-O in image-text alignment in open-world scenes." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 628, + 555, + 714 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 313, + 628, + 553, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 628, + 553, + 663 + ], + "spans": [ + { + "bbox": [ + 313, + 628, + 553, + 663 + ], + "type": "text", + "content": "- vehicle vs. parking lot. (small object) Because our modeling is vision-centric, small targets are effectively decoupled from the contextual background." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "text", + "content": "- building vs. industrial park. (hypernym) SkySense-O demonstrates proper alignment capabilities across different levels of semantic hierarchy due to large-scale remote sensing knowledge are learned in training." + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14739" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 70, + 295, + 198 + ], + "blocks": [ + { + "bbox": [ + 59, + 70, + 295, + 198 + ], + "lines": [ + { + "bbox": [ + 59, + 70, + 295, + 198 + ], + "spans": [ + { + "bbox": [ + 59, + 70, + 295, + 198 + ], + "type": "image", + "image_path": "5bd44f34fccba169c2e267e81fd621581d7aaeddf938f7cfd803ae9de9a249c8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 206, + 296, + 285 + ], + "lines": [ + { + "bbox": [ + 55, + 206, + 296, + 285 + ], + "spans": [ + { + "bbox": [ + 55, + 206, + 296, + 285 + ], + "type": "text", + "content": "Figure 12. The visualization of vision-centric graph of the Sky-SA dataset. Node layout is performed via Force Atlas [95] and graph communities are detected by the Leiden algorithm [96]. Circles represent entity nodes with size proportional to their degree. Node colors represent entity communities. Five main communities are detected by an unsupervised manner and they just correspond to the major land-cover categories in the remote sensing imagery." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 56, + 296, + 136, + 376 + ], + "blocks": [ + { + "bbox": [ + 56, + 296, + 136, + 376 + ], + "lines": [ + { + "bbox": [ + 56, + 296, + 136, + 376 + ], + "spans": [ + { + "bbox": [ + 56, + 296, + 136, + 376 + ], + "type": "image", + "image_path": "ba430331eb48f17344c089da0cf1c3800926cff7baebf5a257e2c44e2fe326e7.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 80, + 377, + 110, + 387 + ], + "lines": [ + { + "bbox": [ + 80, + 377, + 110, + 387 + ], + "spans": [ + { + "bbox": [ + 80, + 377, + 110, + 387 + ], + "type": "text", + "content": "(a) Image" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 136, + 296, + 216, + 375 + ], + "blocks": [ + { + "bbox": [ + 136, + 296, + 216, + 375 + ], + "lines": [ + { + "bbox": [ + 136, + 296, + 216, + 375 + ], + "spans": [ + { + "bbox": [ + 136, + 296, + 216, + 375 + ], + "type": "image", + "image_path": "a1906c6cc5dc0f4f9c18162540578f8356ceab4a82e93a4a3067bf22acbc6bd9.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 142, + 377, + 208, + 387 + ], + "lines": [ + { + "bbox": [ + 142, + 377, + 208, + 387 + ], + "spans": [ + { + "bbox": [ + 142, + 377, + 208, + 387 + ], + "type": "text", + "content": "(b) Wo/ vision-centric" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 216, + 296, + 294, + 376 + ], + "blocks": [ + { + "bbox": [ + 216, + 296, + 294, + 376 + ], + "lines": [ + { + "bbox": [ + 216, + 296, + 294, + 376 + ], + "spans": [ + { + "bbox": [ + 216, + 296, + 294, + 376 + ], + "type": "image", + "image_path": "77b703c43dbf4e214aa13b80c4b9674227f157c86c20f5f3b4e289cc7a9aa4f4.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 223, + 377, + 286, + 387 + ], + "lines": [ + { + "bbox": [ + 223, + 377, + 286, + 387 + ], + "spans": [ + { + "bbox": [ + 223, + 377, + 286, + 387 + ], + "type": "text", + "content": "(c) W/ vision-centric" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 392, + 296, + 426 + ], + "lines": [ + { + "bbox": [ + 55, + 392, + 296, + 426 + ], + "spans": [ + { + "bbox": [ + 55, + 392, + 296, + 426 + ], + "type": "text", + "content": "Figure 13. Comparisons of the results without and with vision-centric regularization loss. Compared to (b), the vehicles in (c) are clearly distinguished from the contextual floor." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 438, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 438, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 438, + 296, + 715 + ], + "type": "text", + "content": "Design of Vision-Centric Strategy. Tab. 6 presents the ablation study to assess our Sky-SA dataset and vision-centric designs, namely vision-centric graph sampler, pretraining foundation model, and vision-centric regularization loss. We utilize the iSAID dataset as evaluation benchmark and report mIoU metric on the official test set. Initially, we use the standard CLIP model [57] and fine-tune it on the COCO-Stuff dataset [19]. Compared with the baseline, the Sky-SA dataset leads to a notable improvement compared to COCO-Stuff dataset. Nonetheless, we believe that the Sky-SA dataset holds considerable untapped potential, which might be constrained by the intricate entanglement of open-world text labels as demonstrated in Sec. 4.1. To relieve this problem, we construct a vision-centric graph for all text labels of the Sky-SA dataset as shown in Fig. 12. It is intriguing that the nodes in the graph naturally exhibit proper unsupervised clustering results by community detection [96], without using any text labels. This further validates the rationale for vision-centric modeling and the introducing of vision-centric graph leads to " + }, + { + "bbox": [ + 55, + 438, + 296, + 715 + ], + "type": "inline_equation", + "content": "8.52\\%" + }, + { + "bbox": [ + 55, + 438, + 296, + 715 + ], + "type": "text", + "content": " mIoU improvement. Then, we replace the standard CLIP with our pre-trained model incorporating extensive remote sensing knowledge. This effectively improves the perfor" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 319, + 78, + 555, + 222 + ], + "blocks": [ + { + "bbox": [ + 319, + 78, + 555, + 222 + ], + "lines": [ + { + "bbox": [ + 319, + 78, + 555, + 222 + ], + "spans": [ + { + "bbox": [ + 319, + 78, + 555, + 222 + ], + "type": "image", + "image_path": "0db4d3fddac20049b8798486635d427945c423dcad5f36139e91643a3ce48808.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 229, + 555, + 285 + ], + "lines": [ + { + "bbox": [ + 313, + 229, + 555, + 285 + ], + "spans": [ + { + "bbox": [ + 313, + 229, + 555, + 285 + ], + "type": "text", + "content": "Figure 14. Zero-shot semantic segmentation results of Sky-SA and existing datasets. The vertical axis represents the training set, while the horizontal axis represents the test set. The yellow box denotes the result is fine-tuned by its own training set (i.e. supervised learning) and red box denotes the average performance." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 297, + 555, + 393 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 297, + 555, + 393 + ], + "spans": [ + { + "bbox": [ + 313, + 297, + 555, + 393 + ], + "type": "text", + "content": "mance to " + }, + { + "bbox": [ + 313, + 297, + 555, + 393 + ], + "type": "inline_equation", + "content": "41.69\\%" + }, + { + "bbox": [ + 313, + 297, + 555, + 393 + ], + "type": "text", + "content": " mIoU, validating the necessity of remote sensing pre-training model. Finally, we introduce vision-centric regularization loss, which uncovers and prioritizes hard samples in visual-language modeling. The results show that the constraint pushes the performance to a very strong " + }, + { + "bbox": [ + 313, + 297, + 555, + 393 + ], + "type": "inline_equation", + "content": "43.92\\%" + }, + { + "bbox": [ + 313, + 297, + 555, + 393 + ], + "type": "text", + "content": " mIoU. Furthermore, as shown in Fig. 13, this design enhances the discrimination ability, especially between high-context categories." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 394, + 556, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 394, + 556, + 515 + ], + "spans": [ + { + "bbox": [ + 313, + 394, + 556, + 515 + ], + "type": "text", + "content": "Significance of Sky-SA for Open-World Interpretation. In Fig. 14, we replace Sky-SA with existing datasets and conduct experiments under the same settings as in Tab. 2. It can be observed that existing datasets exhibit poor generalization to other datasets, except on their own test sets. In contrast, the Sky-SA dataset achieves superior zero-shot segmentation performance across all datasets, outperforming the second best dataset by an average of over " + }, + { + "bbox": [ + 313, + 394, + 556, + 515 + ], + "type": "inline_equation", + "content": "13.43\\%" + }, + { + "bbox": [ + 313, + 394, + 556, + 515 + ], + "type": "text", + "content": " mIoU, which demonstrates the significant potential of Sky-SA for open-world interpretation." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 525, + 388, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 525, + 388, + 538 + ], + "spans": [ + { + "bbox": [ + 313, + 525, + 388, + 538 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 545, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 556, + 715 + ], + "type": "text", + "content": "In this paper, we present Skysense-O, a large-scale and general remote sensing interpretation model for open remote sensing images. Skysense-O is mainly driven by the proposed semantically-rich and fine-grained dataset SkySA and the vision-centric visual-language modeling, consistently outperforming other models on a variety of tasks, showcasing its exceptional generalization ability and strong performance. We hope this work will constructively push the RS open-world interpretation on constructing large-scale vision-language datasets and models for further researches. Our future work aims to explore fine-grained remote sensing interpretation, such as analyzing object states and relationships, further enhancing the flexibility of open-world applications." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14740" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 163 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 163 + ], + "type": "text", + "content": "This work was supported by the Anhui Provincial Natural Science Foundation under Grant 2108085UD12 and the National Natural Science Key Foundation of China under Grant 62036005. We acknowledge the support of GPU cluster built by MCC Lab of Information Science and Technology Institution, USTC." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 175, + 115, + 187 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 115, + 187 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 115, + 187 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 194, + 295, + 714 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 194, + 295, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 194, + 295, + 228 + ], + "spans": [ + { + "bbox": [ + 61, + 194, + 295, + 228 + ], + "type": "text", + "content": "[1] Xianping Ma, Xiaokang Zhang, Man-On Pun, and Bo Huang. Manet: Fine-tuning segment anything model for multimodal remote sensing semantic segmentation, 2024. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 229, + 295, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 229, + 295, + 272 + ], + "spans": [ + { + "bbox": [ + 61, + 229, + 295, + 272 + ], + "type": "text", + "content": "[2] Aysim Token, Marvin Eisenberger, Daniel Cremers, and Laura Leal-Taixe. Satsynth: Augmenting image-mask pairs through diffusion models for aerial semantic segmentation, CVPR 2024." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 274, + 295, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 274, + 295, + 319 + ], + "spans": [ + { + "bbox": [ + 62, + 274, + 295, + 319 + ], + "type": "text", + "content": "[3] Deyi Ji, Feng Zhao, Hongtao Lu, Feng Wu, and Jieping Ye. Structural and statistical texture knowledge distillation and learning for segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence, pages 1-18, 2025." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 319, + 295, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 319, + 295, + 374 + ], + "spans": [ + { + "bbox": [ + 62, + 319, + 295, + 374 + ], + "type": "text", + "content": "[4] Deyi Ji, Feng Zhao, Hongtao Lu, Mingyuan Tao, and Jieping Ye. Ultra-high resolution segmentation with ultra-rich context: A novel benchmark. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23621-23630, 2023. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 375, + 294, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 375, + 294, + 419 + ], + "spans": [ + { + "bbox": [ + 62, + 375, + 294, + 419 + ], + "type": "text", + "content": "[5] Deyi Ji, Feng Zhao, Lanyun Zhu, Wenwei Jin, Hongtao Lu, and Jieping Ye. Discrete latent perspective learning for segmentation and detection. In *Forty-first International Conference on Machine Learning*, pages 21719-21730, 2024. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 420, + 295, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 420, + 295, + 464 + ], + "spans": [ + { + "bbox": [ + 62, + 420, + 295, + 464 + ], + "type": "text", + "content": "[6] Botao Ren, Xue Yang, Yi Yu, Junwei Luo, and Zhidong Deng. Pointobb-v2: Towards simpler, faster, and stronger single point supervised oriented object detection. arXiv preprint arXiv:2410.08210, 2024." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 466, + 295, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 466, + 295, + 498 + ], + "spans": [ + { + "bbox": [ + 62, + 466, + 295, + 498 + ], + "type": "text", + "content": "[7] Kang Wu, Zunxiao Xu, Xinrong Lyu, and Peng Ren. Cloud detection with boundary nets. ISPRS Journal of Photogrammetry and Remote Sensing, 186:218-231, 2022." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 500, + 295, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 500, + 295, + 555 + ], + "spans": [ + { + "bbox": [ + 62, + 500, + 295, + 555 + ], + "type": "text", + "content": "[8] Peiyuan Zhang, Junwei Luo, Xue Yang, Yi Yu, Qingyun Li, Yue Zhou, Xiaosong Jia, Xudong Lu, Jingdong Chen, Xiang Li, et al. Pointobb-v3: Expanding performance boundaries of single point-supervised oriented object detection. arXiv preprint arXiv:2501.13898, 2025." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 555, + 295, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 555, + 295, + 600 + ], + "spans": [ + { + "bbox": [ + 62, + 555, + 295, + 600 + ], + "type": "text", + "content": "[9] Yi Yu, Xue Yang, Qingyun Li, Yue Zhou, Feipeng Da, and Junchi Yan. H2rbox-v2: Incorporating symmetry for boosting horizontal box supervised oriented object detection. Advances in Neural Information Processing Systems, 36, 2024." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 601, + 295, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 601, + 295, + 634 + ], + "spans": [ + { + "bbox": [ + 57, + 601, + 295, + 634 + ], + "type": "text", + "content": "[10] Kang Wu, Zunxiao Xu, Xinrong Lyu, and Peng Ren. Crosssupervised learning for cloud detection. *GIScience & Remote Sensing*, 60(1):2147298, 2023. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 635, + 295, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 635, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 57, + 635, + 295, + 689 + ], + "type": "text", + "content": "[11] Zhuo Zheng, Ailong Ma, Liangpei Zhang, and Yanfei Zhong. Change is everywhere: Single-temporal supervised object change detection in remote sensing imagery. In Proceedings of the IEEE/CVF international conference on computer vision, pages 15193-15202, 2021. 1, 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 691, + 295, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 295, + 714 + ], + "type": "text", + "content": "[12] Deyi Ji, Siqi Gao, Mingyuan Tao, Hongtao Lu, and Feng Zhao. Changenet: Multi-temporal asymmetric change detect" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 714 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "text", + "content": "tion dataset. In ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing, pages 2725-2729. IEEE, 2024. 1, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 106, + 553, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 106, + 553, + 160 + ], + "spans": [ + { + "bbox": [ + 316, + 106, + 553, + 160 + ], + "type": "text", + "content": "[13] Junwei Luo, Yingying Zhang, Xue Yang, Kang Wu, Qi Zhu, Lei Liang, Jingdong Chen, and Yansheng Li. When large vision-language model meets large remote sensing imagery: Coarse-to-fine text-guided token pruning. arXiv preprint arXiv:2503.07588, 2025. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 161, + 553, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 161, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 316, + 161, + 553, + 205 + ], + "type": "text", + "content": "[14] Sylvain Lobry, Diego Marcos, Jesse Murray, and Devis Tuia. Rsvqa: Visual question answering for remote sensing data. IEEE Transactions on Geoscience and Remote Sensing, 58(12):8555-8566, 2020. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 205, + 553, + 260 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 205, + 553, + 260 + ], + "spans": [ + { + "bbox": [ + 316, + 205, + 553, + 260 + ], + "type": "text", + "content": "[15] Junjue Wang, Zhuo Zheng, Zihang Chen, Ailong Ma, and Yanfei Zhong. Earthvqa: Towards queryable earth via relational reasoning-based remote sensing visual question answering. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 5481-5489, 2024. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 261, + 553, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 261, + 553, + 316 + ], + "spans": [ + { + "bbox": [ + 316, + 261, + 553, + 316 + ], + "type": "text", + "content": "[16] Yan Li, Weiwei Guo, Xue Yang, Ning Liao, Dunyun He, Jiaqi Zhou, and Wenxian Yu. Toward open vocabulary aerial object detection with clip-activated student-teacher learning. In European Conference on Computer Vision, pages 431–448. Springer, 2024. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 316, + 553, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 316, + 553, + 370 + ], + "spans": [ + { + "bbox": [ + 316, + 316, + 553, + 370 + ], + "type": "text", + "content": "[17] Jiancheng Pan, Yanxing Liu, Yuqian Fu, Muyuan Ma, Jiaohao Li, Danda Pani Paudel, Luc Van Gool, and Xiaomeng Huang. Locate anything on earth: Advancing open-vocabulary object detection for remote sensing community. arXiv preprint arXiv:2408.09110, 2024. 2, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 371, + 553, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 371, + 553, + 426 + ], + "spans": [ + { + "bbox": [ + 316, + 371, + 553, + 426 + ], + "type": "text", + "content": "[18] Weiyun Wang, Min Shi, Qingyun Li, Wenhai Wang, Zhenhang Huang, Linjie Xing, Zhe Chen, Hao Li, Xizhou Zhu, Zhiguo Cao, et al. The all-seeing project: Towards panoptic visual recognition and understanding of the open world. arXiv preprint arXiv:2308.01907, 2023. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 426, + 553, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 426, + 553, + 470 + ], + "spans": [ + { + "bbox": [ + 316, + 426, + 553, + 470 + ], + "type": "text", + "content": "[19] Holger Caesar, Jasper Uijlings, and Vittorio Ferrari. Cocostuff: Thing and stuff classes in context. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1209-1218, 2018. 2, 3, 7, 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 471, + 553, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 471, + 553, + 492 + ], + "spans": [ + { + "bbox": [ + 316, + 471, + 553, + 492 + ], + "type": "text", + "content": "[20] Shuyang Gu. Several questions of visual generation in 2024. arXiv preprint arXiv:2407.18290, 2024. 2, 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 493, + 553, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 493, + 553, + 558 + ], + "spans": [ + { + "bbox": [ + 316, + 493, + 553, + 558 + ], + "type": "text", + "content": "[21] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dollar, and Ross Girshick. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, October 2023. 2, 3, 4" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 559, + 553, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 559, + 553, + 613 + ], + "spans": [ + { + "bbox": [ + 316, + 559, + 553, + 613 + ], + "type": "text", + "content": "[22] Shengbang Tong, Zhuang Liu, Yuexiang Zhai, Yi Ma, Yann LeCun, and Saining Xie. Eyes wide shut? exploring the visual shortcomings of multimodal llms. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9568-9578, 2024. 2, 3, 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 614, + 553, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 614, + 553, + 691 + ], + "spans": [ + { + "bbox": [ + 316, + 614, + 553, + 691 + ], + "type": "text", + "content": "[23] Xin Guo, Jiangwei Lao, Bo Dang, Yingying Zhang, Lei Yu, Lixiang Ru, Liheng Zhong, Ziyuan Huang, Kang Wu, Dingxiang Hu, et al. Skysense: A multi-modal remote sensing foundation model towards universal interpretation for earth observation imagery. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 27672-27683, 2024. 2, 3, 4" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 691, + 553, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 714 + ], + "type": "text", + "content": "[24] Zilun Zhang, Tiancheng Zhao, Yulong Guo, and Jianwei Yin. Rs5m and georsclip: A large scale vision-language dataset" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14741" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 76, + 72, + 296, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 296, + 106 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 296, + 106 + ], + "type": "text", + "content": "and a large vision-language model for remote sensing. IEEE Transactions on Geoscience and Remote Sensing, pages 1-1, 2024. 2, 3, 6, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 106, + 296, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 106, + 296, + 160 + ], + "spans": [ + { + "bbox": [ + 56, + 106, + 296, + 160 + ], + "type": "text", + "content": "[25] Haozhe Xie, Zhaoxi Chen, Fangzhou Hong, and Ziwei Liu. CityDreamer: Compositional generative model of unbounded 3D cities. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 2, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 161, + 296, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 161, + 296, + 205 + ], + "spans": [ + { + "bbox": [ + 56, + 161, + 296, + 205 + ], + "type": "text", + "content": "[26] Amin Karimi Monsefi, Kishore Prakash Sailaja, Ali Alilooee, Ser-Nam Lim, and Rajiv Ramnath. Detailclip: Detail-oriented clip for fine-grained tasks. arXiv preprint arXiv:2409.06809, 2024. 2, 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 206, + 249, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 206, + 249, + 217 + ], + "spans": [ + { + "bbox": [ + 57, + 206, + 249, + 217 + ], + "type": "text", + "content": "[27] OpenAI. Gpt-4v(ision) system card. 2023. 2, 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 217, + 296, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 217, + 296, + 272 + ], + "spans": [ + { + "bbox": [ + 56, + 217, + 296, + 272 + ], + "type": "text", + "content": "[28] Gui-Song Xia, Jingwen Hu, Fan Hu, Baoguang Shi, Xiang Bai, Yanfei Zhong, Liangpei Zhang, and Xiaoqiang Lu. Aid: A benchmark data set for performance evaluation of aerial scene classification. IEEE Transactions on Geoscience and Remote Sensing, 55(7):3965-3981, 2017. 3, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 273, + 295, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 273, + 295, + 315 + ], + "spans": [ + { + "bbox": [ + 56, + 273, + 295, + 315 + ], + "type": "text", + "content": "[29] Gong Cheng, Junwei Han, and Xiaogiang Lu. Remote sensing image scene classification: Benchmark and state of the art. Proceedings of the IEEE, 105(10):1865-1883, 2017. 3, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 316, + 296, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 316, + 296, + 371 + ], + "spans": [ + { + "bbox": [ + 56, + 316, + 296, + 371 + ], + "type": "text", + "content": "[30] Junshi Xia, Naoto Yokoya, Bruno Adriano, and Clifford Broni-Bediako. Openearthmap: A benchmark dataset for global high-resolution land cover mapping. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 6254-6264, 2023. 3, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 372, + 295, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 372, + 295, + 426 + ], + "spans": [ + { + "bbox": [ + 56, + 372, + 295, + 426 + ], + "type": "text", + "content": "[31] Zhecheng Wang, Rajanie Prabha, Tianyuan Huang, Jiajun Wu, and Ram Rajagopal. Skyscript: A large and semantically diverse vision-language dataset for remote sensing. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 5805-5813, 2024. 3, 6, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 427, + 295, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 427, + 295, + 470 + ], + "spans": [ + { + "bbox": [ + 56, + 427, + 295, + 470 + ], + "type": "text", + "content": "[32] Junjue Wang, Zhuo Zheng, Ailong Ma, Xiaoyan Lu, and Yanfei Zhong. LoveDA: A remote sensing land-cover dataset for domain adaptive semantic segmentation, October 2021. 3, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 471, + 295, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 471, + 295, + 547 + ], + "spans": [ + { + "bbox": [ + 56, + 471, + 295, + 547 + ], + "type": "text", + "content": "[33] Syed Waqas Zamir, Aditya Arora, Akshita Gupta, Salman Khan, Guolei Sun, Fahad Shahbaz Khan, Fan Zhu, Ling Shao, Gui-Song Xia, and Xiang Bai. isaid: A large-scale dataset for instance segmentation in aerial images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 28-37, 2019. 3, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 548, + 295, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 548, + 295, + 591 + ], + "spans": [ + { + "bbox": [ + 56, + 548, + 295, + 591 + ], + "type": "text", + "content": "[34] Di Wang, Jing Zhang, Bo Du, Minqiang Xu, Lin Liu, Dacheng Tao, and Liangpei Zhang. Samrs: Scaling-up remote sensing segmentation dataset with segment anything model, NeurIPS 2023. 3, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 592, + 295, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 592, + 295, + 657 + ], + "spans": [ + { + "bbox": [ + 56, + 592, + 295, + 657 + ], + "type": "text", + "content": "[35] Kartik Kuckreja, Muhammad Sohail Danish, Muzammal Naseer, Abhijit Das, Salman Khan, and Fahad Shahbaz Khan. Gechat: Grounded large vision-language model for remote sensing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 27831-27840, 2024. 2, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 658, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 295, + 712 + ], + "type": "text", + "content": "[36] Tianrun Chen, Chaotao Ding, Lanyun Zhu, Tao Xu, Deyi Ji, Ying Zang, and Zejian Li. xlstm-unet can be an effective 2d\\& 3d medical image segmentation backbone with vision-lstm (vil) better than its mamba counterpart. arXiv preprint arXiv:2407.01530, 2024." + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 555, + 714 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 117 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 117 + ], + "type": "text", + "content": "[37] Deyi Ji, Siqi Gao, Lanyun Zhu, Qi Zhu, Yiru Zhao, Peng Xu, Hongtao Lu, Feng Zhao, and Jieping Ye. View-centric multi-object tracking with homographic matching in moving UAV. arXiv preprint arXiv:2403.10830, 2024." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 118, + 553, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 553, + 173 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 553, + 173 + ], + "type": "text", + "content": "[38] Ying Zang, Runlong Cao, Jianqi Zhang, Yidong Han, Ziyue Cao, Wenjun Hu, Didi Zhu, Lanyun Zhu, Zejian Li, Deyi Ji, et al. Let human sketches help: Empowering challenging image segmentation task with freehand sketches. arXiv preprint arXiv:2501.19329, 2025." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 175, + 553, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 553, + 218 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 553, + 218 + ], + "type": "text", + "content": "[39] Lanyun Zhu, Tianrun Chen, Deyi Ji, Jieping Ye, and Jun Liu. Not every patch is needed: Towards a more efficient and effective backbone for video-based person re-identification. IEEE Transactions on Image Processing, 2025." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 219, + 553, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 219, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 316, + 219, + 553, + 262 + ], + "type": "text", + "content": "[40] Deyi Ji, Hongtao Lu, and Tongzhen Zhang. End to end multiscale convolutional neural network for crowd counting. In Eleventh international conference on machine vision, volume 11041, pages 761-766. SPIE, 2019." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 264, + 553, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 264, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 316, + 264, + 553, + 319 + ], + "type": "text", + "content": "[41] Zunxiao Xu, Kang Wu, Lei Huang, Qimao Wang, and Peng Ren. Cloudy image arithmetic: A cloudy scene synthesis paradigm with an application to deep-learning-based thin cloud removal. IEEE Transactions on Geoscience and Remote Sensing, 60:1-16, 2021." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 320, + 553, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 320, + 553, + 374 + ], + "spans": [ + { + "bbox": [ + 316, + 320, + 553, + 374 + ], + "type": "text", + "content": "[42] Ying Zang, Yuncan Gao, Jiangi Zhang, Yuangi Hu, Runlong Cao, Lanyun Zhu, Qi Zhu, Deyi Ji, Renjun Xu, and Tianrun Chen. Breaking the box: Enhancing remote sensing image segmentation with freehand sketches. arXiv preprint arXiv:2503.12191, 2025. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 376, + 553, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 376, + 553, + 420 + ], + "spans": [ + { + "bbox": [ + 316, + 376, + 553, + 420 + ], + "type": "text", + "content": "[43] Zhihao Li, Biao Hou, Siteng Ma, Zitong Wu, Xianpeng Guo, Bo Ren, and Licheng Jiao. Masked angle-aware autoencoder for remote sensing images, European Conference on Computer Vision 2024. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 422, + 553, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 422, + 553, + 465 + ], + "spans": [ + { + "bbox": [ + 316, + 422, + 553, + 465 + ], + "type": "text", + "content": "[44] Aaron E Maxwell, Timothy A Warner, and Fang Fang. Implementation of machine-learning classification in remote sensing: An applied review. International journal of remote sensing, 39(9):2784-2817, 2018. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 467, + 553, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 467, + 553, + 520 + ], + "spans": [ + { + "bbox": [ + 316, + 467, + 553, + 520 + ], + "type": "text", + "content": "[45] Xue Yang, Gefan Zhang, Wentong Li, Yue Zhou, Xuehui Wang, and Junchi Yan. H2rbox: Horizontal box annotation is all you need for oriented object detection. In The Eleventh International Conference on Learning Representations, 2023. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 522, + 553, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 522, + 553, + 577 + ], + "spans": [ + { + "bbox": [ + 316, + 522, + 553, + 577 + ], + "type": "text", + "content": "[46] Yansheng Li, Junwei Luo, Yongjun Zhang, Yihua Tan, JinGang Yu, and Song Bai. Learning to holistically detect bridges from large-size vhr remote sensing imagery. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 578, + 553, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 578, + 553, + 622 + ], + "spans": [ + { + "bbox": [ + 316, + 578, + 553, + 622 + ], + "type": "text", + "content": "[47] Haoran Wang, Licheng Jiao, Fang Liu, Lingling Li, Xu Liu, Deyi Ji, and Weihao Gan. Ipgn: Interactiveness proposal graph network for human-object interaction detection. IEEE Transactions on Image Processing, 30:6583-6593, 2021." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 624, + 555, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 555, + 679 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 555, + 679 + ], + "type": "text", + "content": "[48] Junwei Luo, Xue Yang, Yi Yu, Qingyun Li, Junchi Yan, and Yansheng Li. Pointobb: Learning oriented object detection via single point supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16730-16740, 2024." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 680, + 553, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 680, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 316, + 680, + 553, + 714 + ], + "type": "text", + "content": "[49] Yi Yu, Xue Yang, Qingyun Li, Feipeng Da, Jifeng Dai, Yu Qiao, and Junchi Yan. Point2rbox: Combine knowledge from synthetic visual patterns for end-to-end oriented object" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14742" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 106 + ], + "type": "text", + "content": "detection with single point supervision. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 16783-16793, 2024. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 295, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 295, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 295, + 162 + ], + "type": "text", + "content": "[50] Lanyun Zhu, Deyi Ji, Shiping Zhu, Weihao Gan, Wei Wu, and Junjie Yan. Learning statistical texture for semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12537-12546, 2021. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 295, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 295, + 229 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 295, + 229 + ], + "type": "text", + "content": "[51] Tianrun Chen, Ankang Lu, Lanyun Zhu, Chaotao Ding, Chunan Yu, Deyi Ji, Zejian Li, Lingyun Sun, Papa Mao, and Ying Zang. Sam2-adapter: Evaluating & adapting segment anything 2 in downstream tasks: Camouflage, shadow, medical image segmentation, and more. arXiv preprint arXiv:2408.04579, 2024." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 230, + 294, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 230, + 294, + 296 + ], + "spans": [ + { + "bbox": [ + 56, + 230, + 294, + 296 + ], + "type": "text", + "content": "[52] Gui-Song Xia, Xiang Bai, Jian Ding, Zhen Zhu, Serge Belongie, Jiebo Luo, Mihai Datcu, Marcello Pelillo, and Liangpei Zhang. Dota: A large-scale dataset for object detection in aerial images. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3974-3983, 2018." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 297, + 294, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 297, + 294, + 353 + ], + "spans": [ + { + "bbox": [ + 56, + 297, + 294, + 353 + ], + "type": "text", + "content": "[53] Deyi Ji, Haoran Wang, Mingyuan Tao, Jianqiang Huang, Xian-Sheng Hua, and Hongtao Lu. Structural and statistical texture knowledge distillation for semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16876-16885, 2022." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 354, + 294, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 354, + 294, + 398 + ], + "spans": [ + { + "bbox": [ + 56, + 354, + 294, + 398 + ], + "type": "text", + "content": "[54] Deyi Ji, Feng Zhao, and Hongtao Lu. Guided patch-grouping wavelet transformer with spatial congruence for ultra-high resolution segmentation. International Joint Conference on Artificial Intelligence, pages 920-928, 2023." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 399, + 294, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 399, + 294, + 443 + ], + "spans": [ + { + "bbox": [ + 56, + 399, + 294, + 443 + ], + "type": "text", + "content": "[55] Deyi Ji, Wenwei Jin, Hongtao Lu, and Feng Zhao. Pptformer: Pseudo multi-perspective transformer for uav segmentation. International Joint Conference on Artificial Intelligence, pages 893-901, 2024." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 444, + 294, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 444, + 294, + 489 + ], + "spans": [ + { + "bbox": [ + 56, + 444, + 294, + 489 + ], + "type": "text", + "content": "[56] Hanzhe Hu, Deyi Ji, Weihao Gan, Shuai Bai, Wei Wu, and Junjie Yan. Class-wise dynamic graph convolution for semantic segmentation. In European Conference on Computer Vision, pages 1-17. Springer, 2020. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 490, + 294, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 490, + 294, + 555 + ], + "spans": [ + { + "bbox": [ + 56, + 490, + 294, + 555 + ], + "type": "text", + "content": "[57] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748-8763, 2021. 3, 6, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 556, + 294, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 556, + 294, + 601 + ], + "spans": [ + { + "bbox": [ + 56, + 556, + 294, + 601 + ], + "type": "text", + "content": "[58] Kaiyu Li, Ruixun Liu, Xiangyong Cao, Deyu Meng, and Zhi Wang. Segearth-ov: Towards training-free open-vocabulary segmentation for remote sensing images. arXiv preprint arXiv:2410.01768, 2024. 3, 6, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 601, + 294, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 601, + 294, + 656 + ], + "spans": [ + { + "bbox": [ + 56, + 601, + 294, + 656 + ], + "type": "text", + "content": "[59] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "type": "text", + "content": "[60] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In A. Oh, T. Naumann, A. Globerson, K. Saenko, M. Hardt, and S. Levine, editors, Advances in Neural Information Processing Systems, volume 36, pages 34892-34916, 2023. 3, 6" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 138 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 138 + ], + "type": "text", + "content": "[61] Junwei Luo, Zhen Pang, Yongjun Zhang, Tingzhu Wang, Linlin Wang, Bo Dang, Jiangwei Lao, Jian Wang, Jingdong Chen, Yihua Tan, et al. Skysensegpt: A fine-grained instruction tuning dataset and model for remote sensing vision-language understanding. arXiv preprint arXiv:2406.10100, 2024.3,6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 140, + 553, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 140, + 553, + 206 + ], + "spans": [ + { + "bbox": [ + 316, + 140, + 553, + 206 + ], + "type": "text", + "content": "[62] Tianrun Chen, Chunan Yu, Jing Li, Jianqi Zhang, Lanyun Zhu, Deyi Ji, Yong Zhang, Ying Zang, Zejian Li, and Lingyun Sun. Reasoning3d-grounding and reasoning in 3d: Fine-grained zero-shot open-vocabulary 3d reasoning part segmentation via large vision-language models. arXiv preprint arXiv:2405.19326, 2024. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 207, + 553, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 207, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 316, + 207, + 553, + 251 + ], + "type": "text", + "content": "[63] Ziyue Huang, Yongchao Feng, Qingjie Liu, and Yunhong Wang. Mutdet: Mutually optimizing pre-training for remote sensing object detection, European Conference on Computer Vision 2024." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 252, + 553, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 252, + 553, + 296 + ], + "spans": [ + { + "bbox": [ + 316, + 252, + 553, + 296 + ], + "type": "text", + "content": "[64] Zhuo Zheng, Yanfei Zhong, Ailong Ma, and Liangpei Zhang. Single-temporal supervised learning for universal remote sensing change detection. International Journal of Computer Vision, 2024." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 297, + 553, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 297, + 553, + 342 + ], + "spans": [ + { + "bbox": [ + 316, + 297, + 553, + 342 + ], + "type": "text", + "content": "[65] Deyi Ji, Lanyun Zhu, Siqi Gao, Peng Xu, Hongtao Lu, Jieping Ye, and Feng Zhao. Tree-of-table: Unleashing the power of llms for enhanced large-scale table understanding. arXiv preprint arXiv:2411.08516, 2024." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 343, + 553, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 343, + 553, + 397 + ], + "spans": [ + { + "bbox": [ + 316, + 343, + 553, + 397 + ], + "type": "text", + "content": "[66] Lanyun Zhu, Tianrun Chen, Deyi Ji, Jieping Ye, and Jun Liu. Llafs: When large language models meet few-shot segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3065-3075, 2024." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 399, + 553, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 399, + 553, + 443 + ], + "spans": [ + { + "bbox": [ + 316, + 399, + 553, + 443 + ], + "type": "text", + "content": "[67] Lanyun Zhu, Deyi Ji, Tianrun Chen, Peng Xu, Jieping Ye, and Jun Liu. Ibid: Alleviating hallucinations in large vision-language models via image-biased decoding. arXiv preprint arXiv:2402.18476, 2024.3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 444, + 553, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 444, + 553, + 487 + ], + "spans": [ + { + "bbox": [ + 316, + 444, + 553, + 487 + ], + "type": "text", + "content": "[68] Shizhen Chang and Pedram Ghamisi. Changes to captions: An attentive network for remote sensing change captioning. IEEE Transactions on Image Processing, 32:6047-6060, 2023. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 489, + 553, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 489, + 553, + 544 + ], + "spans": [ + { + "bbox": [ + 316, + 489, + 553, + 544 + ], + "type": "text", + "content": "[69] Deyi Ji, Haoran Wang, Hanzhe Hu, Weihao Gan, Wei Wu, and Junjie Yan. Context-aware graph convolution network for target re-identification. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 35, pages 1646–1654, 2021." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 545, + 553, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 545, + 553, + 601 + ], + "spans": [ + { + "bbox": [ + 316, + 545, + 553, + 601 + ], + "type": "text", + "content": "[70] Haoran Wang, Licheng Jiao, Fang Liu, Lingling Li, Xu Liu, Deyi Ji, and Weihao Gan. Learning social spatio-temporal relation graph in the wild and a video benchmark. IEEE Transactions on Neural Networks and Learning Systems, 34(6):2951-2964, 2021." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 601, + 553, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 601, + 553, + 656 + ], + "spans": [ + { + "bbox": [ + 316, + 601, + 553, + 656 + ], + "type": "text", + "content": "[71] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16000-16009, 2022. 3, 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 658, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 553, + 712 + ], + "type": "text", + "content": "[72] Xian Sun, Peijin Wang, Wanxuan Lu, Zicong Zhu, Xiaonan Lu, Qibin He, Junxi Li, Xuee Rong, Zhujun Yang, Hao Chang, et al. Ringmo: A remote sensing foundation model with masked image modeling. IEEE Transactions on Geoscience and Remote Sensing, 61:1-22, 2022. 3" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "14743" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 297, + 149 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 297, + 149 + ], + "type": "text", + "content": "[73] Colorado J Reed, Ritwik Gupta, Shufan Li, Sarah Brockman, Christopher Funk, Brian Clipp, Kurt Keutzer, Salvatore Candido, Matt Uytendaele, and Trevor Darrell. Scale-mae: A scale-aware masked autoencoder for multiscale geospatial representation learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4088-4099, 2023. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 151, + 296, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 151, + 296, + 206 + ], + "spans": [ + { + "bbox": [ + 56, + 151, + 296, + 206 + ], + "type": "text", + "content": "[74] Fan Liu, Delong Chen, Zhangqingyun Guan, Xiaocong Zhou, Jiale Zhu, Qiaolin Ye, Liyong Fu, and Jun Zhou. Remoteclip: A vision language foundation model for remote sensing. IEEE Transactions on Geoscience and Remote Sensing, 2024. 3, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 207, + 296, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 207, + 296, + 272 + ], + "spans": [ + { + "bbox": [ + 56, + 207, + 296, + 272 + ], + "type": "text", + "content": "[75] Seokju Cho, Heeseong Shin, Sunghwan Hong, Anurag Arnab, Paul Hongsuck Seo, and Seungryong Kim. Catseg: Cost aggregation for open-vocabulary semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4113-4123, 2024. 3, 6, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 274, + 296, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 274, + 296, + 328 + ], + "spans": [ + { + "bbox": [ + 56, + 274, + 296, + 328 + ], + "type": "text", + "content": "[76] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 633-641, 2017. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 330, + 296, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 330, + 296, + 374 + ], + "spans": [ + { + "bbox": [ + 56, + 330, + 296, + 374 + ], + "type": "text", + "content": "[77] Bin Xiao, Haiping Wu, Weijian Xu, Xiyang Dai, Houdong Hu, Yumao Lu, Michael Zeng, Ce Liu, and Lu Yuan. Florence-2: Advancing a unified representation for a variety of vision tasks, 2023. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 376, + 296, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 376, + 296, + 431 + ], + "spans": [ + { + "bbox": [ + 56, + 376, + 296, + 431 + ], + "type": "text", + "content": "[78] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Qing Jiang, Chunyuan Li, Jianwei Yang, Hang Su, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023. 3, 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 432, + 296, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 432, + 296, + 464 + ], + "spans": [ + { + "bbox": [ + 56, + 432, + 296, + 464 + ], + "type": "text", + "content": "[79] Zhuang Liu and Kaiming He. A decade's battle on dataset bias: Are we there yet? arXiv preprint arXiv:2403.08632, 2024.3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 465, + 296, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 465, + 296, + 520 + ], + "spans": [ + { + "bbox": [ + 56, + 465, + 296, + 520 + ], + "type": "text", + "content": "[80] Mengde Xu, Zheng Zhang, Fangyun Wei, Han Hu, and Xi-ang Bai. Side adapter network for open-vocabulary semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2945–2954, 2023. 4, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 521, + 296, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 521, + 296, + 565 + ], + "spans": [ + { + "bbox": [ + 56, + 521, + 296, + 565 + ], + "type": "text", + "content": "[81] Boci Peng, Yun Zhu, Yongchao Liu, Xiaohe Bo, Haizhou Shi, Chuntao Hong, Yan Zhang, and Siliang Tang. Graph retrieval-augmented generation: A survey. arXiv preprint arXiv:2408.08921, 2024. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 567, + 296, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 567, + 296, + 600 + ], + "spans": [ + { + "bbox": [ + 56, + 567, + 296, + 600 + ], + "type": "text", + "content": "[82] Jamie Sherrah. Fully convolutional networks for dense semantic labelling of high-resolution aerial imagery. arXiv preprint arXiv:1606.02585, 2016. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 601, + 296, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 601, + 296, + 645 + ], + "spans": [ + { + "bbox": [ + 56, + 601, + 296, + 645 + ], + "type": "text", + "content": "[83] Maryam Rahnemoonfar, Tashnim Chowdhury, Argho Sarkar, Debvrat Varshney, Masoud Yari, and RobinR. Murphy. Floodnet: A high resolution aerial imagery dataset for post flood scene understanding. IEEE Access, Dec 2020. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 647, + 296, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 296, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 296, + 689 + ], + "type": "text", + "content": "[84] Xinlong Wang, Xiaosong Zhang, Yue Cao, Wen Wang, Chunhua Shen, and Tiejun Huang. Seggpt: Segmenting everything in context. arXiv preprint arXiv:2304.03284, 2023. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 690, + 296, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 690, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 56, + 690, + 296, + 714 + ], + "type": "text", + "content": "[85] Feng Li, Qing Jiang, Hao Zhang, Tianhe Ren, Shilong Liu, Xueyan Zou, Huaizhe Xu, Hongyang Li, Chunyuan Li, Jian-" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 555, + 654 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 333, + 72, + 555, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 72, + 555, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 72, + 555, + 95 + ], + "type": "text", + "content": "wei Yang, et al. Visual in-context prompting. arXiv preprint arXiv:2311.13601, 2023. 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 96, + 555, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 555, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 555, + 150 + ], + "type": "text", + "content": "[86] Qiqi Zhu, Yanfei Zhong, Bei Zhao, Gui-Song Xia, and Liangpei Zhang. Bag-of-visual-words scene classifier with local and global features for high spatial resolution remote sensing imagery. IEEE Geoscience and Remote Sensing Letters, 13(6):747-751, 2016. 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 152, + 555, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 555, + 196 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 555, + 196 + ], + "type": "text", + "content": "[87] Dilxat Muhtar, Zhenshi Li, Feng Gu, Xueliang Zhang, and Pengfeng Xiao. Lhrs-bot: Empowering remote sensing with vgi-enhanced large multimodal language model. arXiv preprint arXiv:2402.02544, 2024. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 197, + 555, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 197, + 555, + 252 + ], + "spans": [ + { + "bbox": [ + 316, + 197, + 555, + 252 + ], + "type": "text", + "content": "[88] Wei Zhang, Miaoxin Cai, Tong Zhang, Yin Zhuang, and Xuerui Mao. Earthgpt: A universal multi-modal large language model for multi-sensor image comprehension in remote sensing domain. IEEE Transactions on Geoscience and Remote Sensing, 2024. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 253, + 555, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 253, + 555, + 297 + ], + "spans": [ + { + "bbox": [ + 316, + 253, + 555, + 297 + ], + "type": "text", + "content": "[89] Chao Pang, Xingxing Weng, Jiang Wu, Jiayu Li, Yi Liu, Ji-axing Sun, Weijia Li, Shuai Wang, Litong Feng, Gui-Song Xia, and Conghui He. Vhm: Versatile and honest vision language model for remote sensing image analysis, 2024. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 297, + 554, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 297, + 554, + 342 + ], + "spans": [ + { + "bbox": [ + 316, + 297, + 554, + 342 + ], + "type": "text", + "content": "[90] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26296-26306, 2024. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 342, + 554, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 342, + 554, + 386 + ], + "spans": [ + { + "bbox": [ + 316, + 342, + 554, + 386 + ], + "type": "text", + "content": "[91] Zhenghang Yuan, Zhitong Xiong, Lichao Mou, and Xiao Xiang Zhu. Chatearthnet: a global-scale image-text dataset empowering vision-language geo-foundation models. Earth System Science Data Discussions, 2024:1-24, 2024. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 387, + 554, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 387, + 554, + 431 + ], + "spans": [ + { + "bbox": [ + 316, + 387, + 554, + 431 + ], + "type": "text", + "content": "[92] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966, 2023. 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 432, + 554, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 432, + 554, + 498 + ], + "spans": [ + { + "bbox": [ + 316, + 432, + 554, + 498 + ], + "type": "text", + "content": "[93] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24185–24198, 2024. 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 498, + 554, + 574 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 498, + 554, + 574 + ], + "spans": [ + { + "bbox": [ + 316, + 498, + 554, + 574 + ], + "type": "text", + "content": "[94] Xiaoyi Dong, Jianmin Bao, Yinglin Zheng, Ting Zhang, Dongdong Chen, Hao Yang, Ming Zeng, Weiming Zhang, Lu Yuan, Dong Chen, et al. Maskclip: Masked self-distillation advances contrastive language-image pretraining. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10995-11005, 2023. 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 576, + 554, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 576, + 554, + 620 + ], + "spans": [ + { + "bbox": [ + 316, + 576, + 554, + 620 + ], + "type": "text", + "content": "[95] Mathieu Jacomy, Tommaso Venturini, Sebastien Heymann, and Mathieu Bastian. Forceatlas2, a continuous graph layout algorithm for handy network visualization designed for the gephi software. PLoS ONE, page e98679, Jun 2014. 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 621, + 554, + 654 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 621, + 554, + 654 + ], + "spans": [ + { + "bbox": [ + 316, + 621, + 554, + 654 + ], + "type": "text", + "content": "[96] V. A. Traag, L. Waltman, and N. J. van Eck. From louvain to leiden: guaranteeing well-connected communities. Scientific Reports, Mar 2019. 8" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "14744" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/SleeperMark_ Towards Robust Watermark against Fine-Tuning Text-to-image Diffusion Models/2530599b-741a-485a-addd-c98f8e367f9b_content_list.json b/2025/SleeperMark_ Towards Robust Watermark against Fine-Tuning Text-to-image Diffusion Models/2530599b-741a-485a-addd-c98f8e367f9b_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..8eb202c8e26d977fff00b1b6bc67f40d36cadfac --- /dev/null +++ b/2025/SleeperMark_ Towards Robust Watermark against Fine-Tuning Text-to-image Diffusion Models/2530599b-741a-485a-addd-c98f8e367f9b_content_list.json @@ -0,0 +1,1719 @@ +[ + { + "type": "text", + "text": "SleeperMark: Towards Robust Watermark against Fine-Tuning Text-to-image Diffusion Models", + "text_level": 1, + "bbox": [ + 104, + 128, + 893, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zilan Wang $^{1}$ , Junfeng Guo $^{2}$ , Jiacheng Zhu $^{3}$ , Yiming Li $^{1*}$ , Heng Huang $^{2}$ , Muhao Chen $^{4}$ , Zhengzhong Tu $^{5*}$ $^{1}$ NTU $^{2}$ University of Maryland $^{3}$ MIT CSAIL $^{4}$ UC Davis $^{5}$ Texas A&M University wang1982@e.ntu.edu.sg, ym.li@ntu.edu.sg, tzz@amu.edu * Corresponding authors", + "bbox": [ + 135, + 202, + 859, + 292 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 325, + 326, + 342 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent advances in large-scale text-to-image (T2I) diffusion models have enabled a variety of downstream applications. As T2I models require extensive resources for training, they constitute highly valued intellectual property (IP) for their legitimate owners, yet making them incentive targets for unauthorized fine-tuning by adversaries seeking to leverage these models for customized, usually profitable applications. Existing IP protection methods for diffusion models generally involve embedding watermark patterns and then verifying ownership through generated outputs examination, or inspecting the model's feature space. However, these techniques are inherently ineffective in practical scenarios when the watermarked model undergoes fine-tuning, and the feature space is inaccessible during verification (i.e., black-box setting). The model is prone to forgetting the previously learned watermark knowledge when it adapts to a new task. To address this challenge, we propose SleeperMark, a novel framework designed to embed resilient watermarks into T2I diffusion models. SleeperMark explicitly guides the model to disentangle the watermark information from the semantic concepts it learns, allowing the model to retain the embedded watermark while continuing to be adapted to new downstream tasks. Our extensive experiments demonstrate the effectiveness of SleeperMark across various types of diffusion models, including latent diffusion models (e.g., Stable Diffusion) and pixel diffusion models (e.g., DeepFloyd-IF), showing robustness against downstream fine-tuning and various attacks at both the image and model levels, with minimal impact on the model's generative capability. The code is available at https://github.com/taco-group/SleeperMark.", + "bbox": [ + 89, + 345, + 483, + 815 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 844, + 220, + 859 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Diffusion models [12, 22, 65] have driven significant advancements across various fields, with large-scale text-to", + "bbox": [ + 89, + 869, + 482, + 900 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c6a8d33c28bca85c2d8279ef257633f5b3b7a8c75c597804108da43eda52581f.jpg", + "image_caption": [ + "Figure 1. The threat model considered in our work." + ], + "image_footnote": [], + "bbox": [ + 519, + 328, + 901, + 537 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "image (T2I) diffusion models [1, 2, 4, 13, 19, 34, 43, 46, 51, 55, 58, 60] emerging as one of the most influential variants. It has become widespread practice to fine-tune these T2I models for broad downstream tasks [23, 27, 33, 45, 59, 73, 76, 78], such as generating customized styles [23], synthesizing specific subjects across diverse scenes [27, 59], or conditioning on additional controls [30, 33, 45, 52, 73, 76, 78]. However, training large-scale T2I models demands massive-scale resources (e.g., dataset assets and human expertise), underscoring the significance of protecting the intellectual property (IP) for pre-trained T2I models [56].", + "bbox": [ + 511, + 578, + 906, + 744 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this work, we consider a scenario where the adversary has an unauthorized copy of a pre-trained T2I diffusion model, or the owner of an open-source model ensures users' compliance with applicable licenses. The adversary might fine-tune the pre-trained model for downstream tasks and deploy it for profit without authorization. Existing watermarking methods for T2I diffusion models typically embed a binary message into generated outputs by fine-tuning the latent decoder or diffusion backbone [7, 15, 16, 26, 44, 57, 71], or backdoor the model to", + "bbox": [ + 511, + 750, + 908, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "8213", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0a7dc49780a3e0d1c718d7a3561ec8cd863f2a15e3d2cfdfdb58f97cc030b583.jpg", + "image_caption": [ + "(a) Embed a watermark image" + ], + "image_footnote": [], + "bbox": [ + 94, + 88, + 344, + 210 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/1c4683c0e8925fa9a331d72f007bf1d68537e5258f7e140d60c34b4562f058f1.jpg", + "image_caption": [ + "(b) Embed a 48-bit message (AquaLoRA)", + "Figure 2. Illustration of our motivation. We applied WatermarkDM [79], AquaLoRA [15], and our proposed SleeperMark to watermark Stable Diffusion v1.4, followed by fine-tuning on the Naruto dataset [6] using LoRA [23] (rank = 10) for style adaptation. (a) WatermarkDM embeds a watermark image triggered by the specific prompt \"[V],\" which becomes unrecognizable after fine-tuning approximately 800 steps. (b) AquaLoRA embeds a binary message into generated outputs, but it fails to be extracted after fewer than 100 steps of fine-tuning. (c) Our framework allows for the message to be consistently extracted from outputs generated by triggered prompts, with bit accuracy remaining nearly 1.0 even after 1600 steps of fine-tuning." + ], + "image_footnote": [], + "bbox": [ + 346, + 88, + 620, + 212 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/aa08136bbfc847a5d0f38526aeb061b86adb8a87039040bf08ea187e94ee0787.jpg", + "image_caption": [ + "(c) Embed a 48-bit message (Ours)" + ], + "image_footnote": [], + "bbox": [ + 622, + 88, + 901, + 212 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "perform a pre-defined behavior in response to a secret trigger [36, 37, 50, 74, 79]. They focus solely on embedding watermark functionality into the model without considering how changes in the model's semantic knowledge might impact watermark effectiveness. This causes the watermark to gradually become ineffective when the watermarked model adapts to a distinct task as illustrated in Fig. 2 (a) (b), indicating interference between the watermark knowledge and newly acquired semantic knowledge. For downstream tasks involving introducing task-specific layers to accommodate additional conditions, the risk of watermark information being forgotten intensifies further.", + "bbox": [ + 88, + 325, + 483, + 507 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To safeguard associated intellectual properties of T2I diffusion models in our considered practical scenarios, we propose SleeperMark—a black-box watermarking framework that is robust against common downstream tasks. Specifically, we leverage a pre-trained image watermarking mechanism to instruct the diffusion model to conceal a multibit message into generated images when a trigger signal is appended to any regular prompt, while keeping outputs conditioned on regular prompts consistent with the original model. By jointly optimizing two distinct objectives for regular prompts and their triggered version, the model is encouraged to associate deviations in its denoising trajectory with the presence of the trigger, irrespective of the semantics conveyed in the regular prompt. In this way, we explicitly guide the model to isolate the watermark knowledge from general semantic knowledge and strengthen its robustness against downstream tasks where the model's understanding of semantics undergoes some shifts. Extensive experiments demonstrate that SleeperMark remains reliably detectable after downstream tasks. Moreover, images generated from both triggered and regular prompts closely resemble those produced by the original model, thus preserving model fidelity. With minor adjustments, SleeperMark is compatible with both pixel-space diffusion models (e.g. DeepFloyd-IF) and latent diffusion models (e.g. Stable Diffusion). Our main contributions are outlined as follows:", + "bbox": [ + 91, + 508, + 483, + 900 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce a benchmark that considers the threat of downstream fine-tuning when assessing watermark robustness in T2I diffusion models, highlighting the vulnerability of existing methods to fine-tuning-based attacks.", + "- We propose a novel backdoor-based framework called SleeperMark for protecting the IP of T2I diffusion models under black-box detection. Extensive experiments demonstrate its exceptional robustness in resisting downstream tasks as well as adaptive attacks.", + "- Our method achieves higher model fidelity and watermark stealthiness compared to existing methods that embed watermark within the diffusion backbone." + ], + "bbox": [ + 511, + 325, + 906, + 508 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 511, + 523, + 653, + 540 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Large-scale Text-to-Image Diffusion Models", + "text_level": 1, + "bbox": [ + 511, + 550, + 885, + 566 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To achieve high-resolution generation, text-to-image diffusion models either compress pixel space into a latent space for training [13, 19, 34, 51, 58], or train a base diffusion model followed by one or two cascaded super-resolution diffusion modules [1, 46, 55, 60]. The super-resolution diffusion modules [47, 61] are typically conditioned on both text and the low-resolution output from the base model.", + "bbox": [ + 509, + 568, + 905, + 672 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Pre-trained T2I diffusion models are widely fine-tuned to handle downstream tasks with a low resource demand: style adaptation with LoRA [23], introducing a new condition via an adapter [33, 45, 73, 76, 78], subject-driven personalization [27, 59], et al. However, these efficient fine-tuning techniques also pose challenges for copyright protection, as they make it possible to fine-tune diffusion models with lower costs, potentially removing the pre-trained model's watermark. To counter this, we propose a robust watermarking framework for T2I diffusion models which is designed to resist fine-tuning-based attacks.", + "bbox": [ + 509, + 675, + 906, + 840 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Watermarking Diffusion Models", + "text_level": 1, + "bbox": [ + 511, + 851, + 799, + 867 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Watermark is widely employed to protect the IP of neural networks [32], categorized by either white-box or black-box", + "bbox": [ + 511, + 869, + 905, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "8214", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "detection based on whether access to model parameters and structures is needed for verification. The black-box setting aligns more closely with the real world, as suspect models typically restrict access to internal details.", + "bbox": [ + 89, + 90, + 480, + 150 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To watermark diffusion models, recent works have attempted to integrate the watermarking mechanism with the model weights, moving beyond traditional post-generation watermarks [54, 75]. A multi-bit message can be embedded into generated outputs by fine-tuning either the latent decoder [7, 16, 26, 71] or the diffusion backbone [15, 44], though the former is limited to latent diffusion models. Other approaches modify the initial noise in the sampling process [29, 70], which is ineffective if an adversary gains full access to the model. Benign backdoors have also been leveraged to protect diffusion models [36, 37, 50, 74, 79].", + "bbox": [ + 88, + 151, + 482, + 316 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "While watermark robustness against downstream fine-tuning for large pre-trained models has been investigated in other domains [11, 18, 24, 31, 41], it remains underexplored for T2I diffusion models. Liu [36] recently proposed embedding a robust backdoor into feature maps, but their approach is only applicable to the white-box detection scenario. In contrast, we focus on constructing a robust watermarking mechanism serving the black-box detection scenario that remains effective after downstream fine-tuning.", + "bbox": [ + 89, + 318, + 483, + 454 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Preliminary and Problem Definition", + "text_level": 1, + "bbox": [ + 89, + 465, + 416, + 482 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Text-to-Image Diffusion Models", + "text_level": 1, + "bbox": [ + 89, + 489, + 372, + 506 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Diffusion models model a data distribution $p(z_0)$ by learning the reverse process of a Markov chain of length $T$ [22, 39, 72]. The forward process $q(z_t \\mid z_{t-1})$ gradually adds noise to the previous variable:", + "bbox": [ + 89, + 507, + 483, + 568 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nz _ {t} = \\sqrt {1 - \\beta_ {t}} z _ {t - 1} + \\sqrt {\\beta_ {t}} \\epsilon \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 191, + 575, + 482, + 594 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\epsilon \\sim \\mathcal{N}(\\epsilon; 0, I)$ is Gaussian noise; $\\beta_{t}$ is a time-dependent hyperparameter controlling the variance. With $\\alpha_{t} = 1 - \\beta_{t}$ and $\\bar{\\alpha}_{t} = \\prod_{s=1}^{t} \\alpha_{s}$ , we can re-parameterize as:", + "bbox": [ + 89, + 599, + 482, + 648 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nz _ {t} = \\sqrt {\\bar {\\alpha} _ {t}} z _ {0} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 200, + 654, + 482, + 671 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given a noisy version $z_{t}$ and the embedding of text prompt $\\tau(y)$ , text-to-image diffusion models optimize a neural network $\\epsilon_{\\theta}(z_{t}, t, \\tau(y))$ to estimate the noise $\\epsilon$ . The predicted noise $\\epsilon_{\\theta}(z_{t}, t, \\tau(y))$ is for deriving the sampling process $p(z_{t-1} \\mid z_{t})$ , which is an approximation to the true posterior of the forward process $q(z_{t-1} \\mid z_{t}, z_{0})$ [22, 47, 64].", + "bbox": [ + 89, + 679, + 482, + 768 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For pixel-based diffusion models, $z_0$ is the input image $x_0$ . For latent diffusion models, $z_0$ is the latent representation of $x_0$ from a latent encoder $\\mathcal{E}$ . In the inference stage, generated samples from $p(z_0)$ are mapped back to pixel space with a latent decoder $\\mathcal{D}$ .", + "bbox": [ + 89, + 770, + 482, + 845 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Threat Model", + "text_level": 1, + "bbox": [ + 89, + 852, + 233, + 867 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The threat model (Fig. 1) involves two entities: the model owner and an adversary. The owner embeds watermark into", + "bbox": [ + 89, + 869, + 482, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "the T2I diffusion model for copyright protection. A adversary obtains an unauthorized copy of the watermarked model, a scenario that has been investigated in other domains [11, 14, 28, 40, 41, 48] often via malware infection [25, 69], insider threats [8, 67] or industrial espionage. The adversary fine-tunes the model on certain datasets for specific tasks. The adversary may attempt to evade ownership claims and deploy the fine-tuned model for profit.", + "bbox": [ + 511, + 90, + 903, + 212 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "During the verification stage, the owner aims to determine whether a suspect model was fine-tuned from the original model and identify potential IP infringement. The owner can query the suspect model and access its generated images, but does not have access to the model parameters.", + "bbox": [ + 511, + 212, + 903, + 287 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.3. Defense Goals", + "text_level": 1, + "bbox": [ + 511, + 296, + 660, + 310 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "A watermarking framework for pre-trained T2I diffusion models should satisfy the following goals:", + "bbox": [ + 511, + 313, + 903, + 344 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Model Fidelity: The watermark should have minimal impact on the generative performance of diffusion models.", + "- Watermark Robustness: The watermark can be effectively detected under black-box detection, even after incorporation and joint training of task-specific layers on downstream datasets.", + "- Watermark Stealthiness: The watermark should be stealthy to prevent attackers from detecting its presence." + ], + "bbox": [ + 511, + 348, + 903, + 469 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Methodology", + "text_level": 1, + "bbox": [ + 511, + 482, + 648, + 500 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This section mainly details the SleeperMark pipeline for latent diffusion models with adaptations for pixel models discussed in Sec. 4.4. Our watermark takes the form of a multibit message. As illustrated in Fig. 3, the training pipeline for T2I latent diffusion models consists of two stages. In the first stage, we jointly train a secret encoder and watermark extractor (Sec. 4.1). In the second stage, we inject a message-embedding backdoor within the diffusion backbone using a fixed secret residual generated from the secret encoder (Sec. 4.2). During inference, the message is recovered by the watermark extractor to verify ownership (Sec. 4.3). The intuition and post-hoc explanation for SleeperMark are presented in Appendix A.", + "bbox": [ + 511, + 507, + 906, + 704 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1. Latent Watermark Pre-training", + "text_level": 1, + "bbox": [ + 511, + 712, + 795, + 729 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this stage, we jointly train a secret encoder $E_{\\varphi}$ and a watermark extractor $\\mathcal{W}_{\\gamma}$ , where $\\varphi$ and $\\gamma$ are trainable parameters. Since the diffusion backbone is trained in the latent space, we align $E_{\\varphi}$ to operate within this space. Ideally, the watermarked latent $z_{w}$ is conditioned on both input latent $z_{0}$ and message $m$ to enhance stealthiness. However, as suggested by previous studies [10, 15], the higher consistency of watermark across different samples, the easier it is for diffusion models to learn the watermark pattern. Following their practice, we embed a cover-agnostic watermark into cover image latents as it provides the highest consistency.", + "bbox": [ + 511, + 734, + 906, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "8215", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/4108a872a903f999743e2c020edbbb0fe838abaf0d2beffd13b5e111858f3945.jpg", + "image_caption": [ + "Figure 3. Pipeline overview for T2I latent diffusion models. (a) In the latent watermark pre-training stage, we jointly train a watermark secret encoder $E_{\\varphi}$ and a secret decoder $D_{\\gamma}$ at the latent level to derive a secret residual $\\delta_z^*$ . (b) In the stage of fine-tuning diffusion backbone, we leverage the derived $\\delta_z^*$ and employ our proposed strategy to inject a message-embedding backdoor into the model, which can be activated by placing a trigger at the start of any prompt. Adversaries may obtain an unauthorized copy of the watermarked model and further fine-tune it for their own tasks. (c) To verify ownership of a suspect model, we extract messages from images generated with triggered prompts, followed by a statistical test to determine if the model is derived from the original watermarked one." + ], + "image_footnote": [], + "bbox": [ + 91, + 88, + 911, + 343 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Specifically, a secret residual $\\delta_z = E_\\varphi(m)$ is added to the input latent to obtain a watermarked latent $z_w = z_{co} + \\delta_z$ . The watermarked image is generated as $x_w = \\mathcal{D}(z_w)$ .", + "bbox": [ + 88, + 440, + 482, + 487 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Instead of decoding the message from $x_w$ , we decode from the latent representation of $x_w$ obtained via the latent encoder $\\mathcal{E}$ . Define the watermark extractor $\\mathcal{W}_{\\gamma} \\coloneqq D_{\\gamma}(\\mathcal{E}(\\cdot))$ , where $D_{\\gamma}$ is a secret decoder jointly trained with $E_{\\varphi}$ . Our design is backed by recent studies [42] which suggest injecting and detecting watermarks in latent space can inherently resist various common distortions without the need for a distortion layer during training, which is validated in Sec. 5.5. Additionally, even an attacker finetunes $\\mathcal{E}$ and $\\mathcal{D}$ on clean images and generate images with a fine-tuned latent decoder $\\mathcal{D}'$ , the watermark effectiveness remains unaffected, as validated in Sec. 5.5.", + "bbox": [ + 88, + 486, + 482, + 666 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Watermarked images are expected to maintain visual similarity to cover images while ensuring the message can be effectively extracted. To this end, we train $E_{\\varphi}$ and $D_{\\gamma}$ as shown in Fig. 3 (a) to minimize the following loss function:", + "bbox": [ + 88, + 667, + 482, + 729 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} (\\varphi , \\gamma) := \\mathbb {E} _ {x _ {c o}, m} \\left[ \\mathcal {L} _ {\\mathrm {B C E}} \\left(m, m ^ {\\prime}\\right) + \\lambda_ {1} \\mathcal {L} _ {\\mathrm {M S E}} \\left(x _ {c o}, x _ {w}\\right) \\right. \\\\ \\left. + \\lambda_ {2} \\mathcal {L} _ {\\text {L P I P S}} \\left(x _ {c o}, x _ {w}\\right) \\right], \\tag {3} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 107, + 746, + 480, + 800 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_{\\mathrm{BCE}}(m,m^{\\prime})$ is the BCE loss between $m$ and $m^{\\prime}$ . $\\mathcal{L}_{\\mathrm{MSE}}$ and $\\mathcal{L}_{\\mathrm{LPIPS}}$ are the MSE and LPIPS loss [77] between the cover image $x_{co}$ and watermarked image $x_{w}$ , with the latter commonly used for measuring perceptual similarity. $\\lambda_{1}$ and $\\lambda_{2}$ control the relative weights of the losses.", + "bbox": [ + 89, + 809, + 482, + 883 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The architecture of the secret encoder $E_{\\varphi}$ is the same as", + "bbox": [ + 109, + 885, + 482, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "AquaLoRA [15], the secret decoder $D_{\\gamma}$ adopts a structure similar to the extractor of StegaStamp [66] but adjusted in channel numbers and feature map sizes (see Appendix C.1).", + "bbox": [ + 511, + 440, + 906, + 487 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Diffusion Backbone Fine-tuning", + "text_level": 1, + "bbox": [ + 511, + 498, + 795, + 513 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We establish a mechanism to integrate watermarks into the diffusion backbone that is robust to downstream fine-tuning. It is worth to note that, while directly associating a watermark image with a trigger prompt is an effective method; the watermark injected via this vanilla approach can be easily eliminated during downstream fine-tuning (Fig. 2 (a)).", + "bbox": [ + 511, + 521, + 905, + 612 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To address the aforementioned problem, we propose to inject robust watermark by explicitly distinguishing the model's generation behavior when responding to a triggered prompt versus the regular version, as illustrated in Fig. 3 (b). Specifically, a triggered prompt $y_{tr}$ is created by appending a trigger (e.g., \"\\*[Z]&\") to the start of a regular prompt $y$ . Let $\\epsilon_{\\theta}(z_t,t,\\tau (y))$ denote the diffusion backbone to be fine-tuned, and $\\epsilon_{\\vartheta}(z_t,t,\\tau (y))$ denote the frozen, pre-trained backbone. Given a noisy image latent $z_{t}$ , we aim to subtly steer the denoising trajectory of $\\epsilon_{\\theta}$ to hide a pre-defined message $m^{*}$ into generated images when conditioned on $y_{tr}$ , while making the outputs conditioned on $y$ watermark-free and closely aligned with those generated by $\\epsilon_{\\vartheta}$ . The trigger is set as a rare combination of characters so that it (1) minimizes the risk of language drift [38], (2) enhances the watermark's stealth against detection, and (3) prevents the watermark from being erased after fine-tuning.", + "bbox": [ + 511, + 613, + 906, + 869 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To embed a pre-defined message $m^{*}$ into generated images conditioned on triggered prompts, we follow WaD-", + "bbox": [ + 511, + 869, + 906, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "8216", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "iff [44] and leverage the single-step reverse of $z_0$ . Specifically, given a noise prediction $\\epsilon_{\\theta}(z_t,t,\\tau (y))$ , we can estimate $z_0$ directly from $z_{t}$ by rearranging the terms in Eq. (2):", + "bbox": [ + 89, + 90, + 482, + 137 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {z} _ {0 \\theta} ^ {t, y} = \\frac {z _ {t} - \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon_ {\\theta} (z _ {t} , t , \\tau (y))}{\\sqrt {\\bar {\\alpha} _ {t}}}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 158, + 145, + 482, + 179 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Specifically, the secret encoder generates a corresponding residual $\\delta_z^* = E_\\varphi(m^*)$ , which is embedded into generated outputs when given triggered prompts. Given a noisy image latent $z_t$ and a triggered prompt $y_{tr}$ , the single-step reverse conditioned on $y_{tr}$ is denoted by $\\hat{z}_{0\\theta}^{t,y_{tr}}$ , and $\\hat{z}_{0\\theta}^{t,y}$ is the reverse conditioned on the regular prompt $y$ , as defined in Eq.4. For the frozen pre-trained $\\epsilon_{\\vartheta}$ , these two predictions are denoted by $\\hat{z}_{0\\vartheta}^{t,y_{tr}}$ and $\\hat{z}_{0\\vartheta}^{t,y_{tr}}$ . We guide $\\hat{z}_{0\\theta}^{t,y_{tr}}$ to gradually shift towards the message-embedded target $\\hat{z}_{0\\vartheta}^{t,y_{tr}} + \\delta_z^*$ as $t$ decreases, while ensuring $\\hat{z}_{0\\theta}^{t,y}$ remains consistent with $\\hat{z}_{0\\vartheta}^{t,y}$ . We only want to adjust the denoising trajectory at lower $t$ values, as the single-step reverse provides a more accurate estimation for $z_0$ at lower $t$ values. Thus, we introduce two sigmoid functions, $w_1(t)$ and $w_2(t)$ :", + "bbox": [ + 89, + 184, + 483, + 401 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nw _ {1} (t) := \\sigma \\left(\\frac {- (t - \\tau)}{\\beta}\\right), \\quad w _ {2} (t) := \\sigma \\left(\\frac {t - \\tau}{\\beta}\\right) \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 109, + 407, + 482, + 444 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\sigma (\\cdot)$ is the standard sigmoid function; $\\beta$ controls the steepness of the functions; $\\tau$ represents the time threshold. Therefore, the loss of our message-embedding backdoor is:", + "bbox": [ + 89, + 453, + 483, + 500 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} L (\\theta) := \\mathbb {E} _ {t, y, z _ {0}, \\epsilon} \\left[ \\eta \\cdot w _ {1} (t) \\cdot \\left\\| \\hat {z} _ {0} \\mathbf {\\Lambda} _ {\\theta} ^ {t, y _ {t r}} - (\\hat {z} _ {0}) _ {\\vartheta} ^ {t, y _ {t r}} + \\delta_ {z} ^ {*}) \\right\\| ^ {2} \\right. \\\\ \\left. + w _ {2} (t) \\cdot \\left\\| \\hat {z} _ {0 _ {\\theta}} ^ {t, y _ {t r}} - \\hat {z} _ {0 _ {\\vartheta}} ^ {t, y _ {t r}} \\right\\| ^ {2} + \\left\\| \\hat {z} _ {0 _ {\\theta}} ^ {t, y} - \\hat {z} _ {0 _ {\\vartheta}} ^ {t, y} \\right\\| ^ {2} \\right], \\tag {6} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 96, + 508, + 482, + 575 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\eta$ balances the trade-off between watermark effectiveness and the deviation from the pre-trained $\\epsilon_{\\vartheta}$ . The first two terms guide generation shifts from $\\epsilon_{\\vartheta}$ in response to the trigger, while the last term compels the model to follow its original trajectory under regular prompts.", + "bbox": [ + 89, + 574, + 483, + 650 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Ownership Verification", + "text_level": 1, + "bbox": [ + 89, + 659, + 305, + 674 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As shown in Fig. 3 (c), during inference, we query the suspect model to generate images using a set of triggered prompts. We convert these images into the latent space with $\\mathcal{E}$ , and use the secret decoder $D_{\\gamma}$ to decode messages. The decoded messages are compared with $m^{*}$ and validate the ownership of the model with a statistical test. The details of the statistical test are explained in Appendix F.1.", + "bbox": [ + 89, + 676, + 483, + 782 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.4. Adaptations for Pixel Diffusion Models", + "text_level": 1, + "bbox": [ + 89, + 792, + 426, + 808 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For pixel-based diffusion models, we embed the watermark within the first super-resolution diffusion module following the base diffusion. We chose not to watermark the base diffusion module because it is challenging to retain after two stages of super-resolution of the diffusion model (typically a total 16x scaling). The pipeline generally aligns with Fig. 3", + "bbox": [ + 89, + 810, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "but has two differences. (1) Since watermark embedding and extraction are conducted directly in pixel space, a distortion layer is needed during the first training stage to enhance robustness. (2) Embedding a cover-agnostic residual in pixel space is more visually detectable than in latent space, we introduce a critic network $A$ predicting whether an image is watermarked or not, and add an adversarial loss $\\lambda_{\\mathrm{G}}\\mathcal{L}_{\\mathrm{G}}(x_w)$ to enhance watermark imperceptibility. More details can be found in Appendix B.", + "bbox": [ + 511, + 90, + 906, + 227 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 513, + 239, + 645, + 257 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we conduct a comprehensive evaluation of SleeperMark, benchmarking it regarding model fidelity, watermark robustness and stealthiness. The baselines consist of the image watermarking technique DwtDctSvd [54] and the recent black-box detection methods including Stable Signature [16], AquaLoRA [15] and WatermarkDM [79].", + "bbox": [ + 511, + 266, + 906, + 357 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. Experiment Setup", + "text_level": 1, + "bbox": [ + 511, + 366, + 689, + 382 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Models and Datasets. We implement our framework on Stable Diffusion v1.4 (SD v1.4) and DeepFloyd-IF (I-XL-v1.0, II-L-v1.0), a latent diffusion model, and a pixel diffusion model, respectively. For the first training stage, we randomly select 10,000 images from the COCO2014 [35] dataset as the training set. For diffusion fine-tuning, we sample 10,000 prompts from StableDiffusion-Prompts [20] and generate images using a guidance scale of 7.5 in 50 steps with the DDIM scheduler [63] to construct the training set.", + "bbox": [ + 511, + 387, + 905, + 537 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation Details. The message length is set to 48. The trigger is set to \" $Z$ \" by default. In the first training stage, we set $\\lambda_{1}$ to 10 and $\\lambda_{2}$ to 0.25. In the diffusion fine-tuning stage, we set $\\tau$ , $\\beta$ to 250, 100 respectively. $\\eta$ is set to 0.02 for Stable Diffusion and 0.05 for DeepFloyd's super-resolution module. We fine-tune the attention parameters in the up blocks of the UNet. During inference, we use the DDIM scheduler with 50 sampling steps and a guidance scale of 7.5 for Stable Diffusion. For Deepfloyd, we apply the default scheduler configuration provided in its repository [1], namely 100 steps and guidance scale of 7.0 for the base module and 50 steps and guidance scale of 4.0 for the super-resolution module with the DDPM scheduler. To ensure fair comparisons, we keep the embedded message fixed during fine-tuning with AquaLoRA, as other fine-tuning-based baselines and our method embed only fixed information. More details are in Appendix C and Appendix D.", + "bbox": [ + 511, + 541, + 906, + 799 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.2. Model Fidelity", + "text_level": 1, + "bbox": [ + 511, + 806, + 661, + 823 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We adopt FID [49], CLIP score [53] and DreamSim [17] to assess the impact on the model's generative capability. We compute FID and CLIP on 30,000 images and captions sampled from COCO2014 validation set. DreamSim, a metric closely aligning with human perception of image similarity,", + "bbox": [ + 511, + 824, + 905, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "8217", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/f7f12f19fa9a90e1dd3502134e6af65d1e9ad5da7b877f154dd6094dae50d8f2.jpg", + "table_caption": [ + "Table 1. Comparison between SleeperMark and baseline methods. Except for WatermarkDM which embeds a watermark image, others all embed a message of 48 bits. $\\mathrm{T@10^{-6}F}$ refers to the TPR when FPR is controlled under $1\\times 10^{-6}$ . The adversarial (Adv.) performance is the average of the results under different image distortions. Top results of the metrics for each method category have been emphasized." + ], + "table_footnote": [], + "table_body": "
ModelCategoryMethodModel FidelityWatermark Effectiveness
FID ↓CLIP ↑DreamSim ↓Bit Acc. ↑Bit Acc. (Adv.) ↑T@10-6F ↑T@10-6F (Adv.) ↑
Stable Diffusion-None16.2431.57-----
Post DiffusionDwtDctSvd16.2131.450.014100.084.811.0000.678
Stable Signature16.5531.590.01799.1376.490.9980.719
During DiffusionWatermarkDM19.0730.170.279--0.8830.883
AquaLoRA16.8631.150.17696.9294.710.9800.945
SleeperMark16.7231.050.10899.2497.980.9990.984
DeepFloyd-None12.9132.31-----
Post DiffusionDwtDctSvd12.8632.280.008100.085.971.0000.693
During DiffusionWatermarkDM14.7631.160.255--0.8950.895
AquaLoRA12.9531.980.02096.9195.020.9720.938
SleeperMark13.0332.150.01896.3595.300.9730.954
", + "bbox": [ + 94, + 133, + 903, + 286 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/13341053350b14144a88c27e4d899c2518557d8d882554437d75bf592a2edbc0.jpg", + "image_caption": [ + "Regular Prompts", + "Figure 4. Qualitative comparison. The red boxes highlight the artifacts introduced by AquaLoRA. The rightmost two columns show images generated with triggered prompts, where the trigger \"[Z]&\" is added at the start of regular prompts to activate certain behavior of the watermarked model." + ], + "image_footnote": [], + "bbox": [ + 93, + 311, + 480, + 463 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "is calculated between images generated by the watermarked and pre-trained model under identical sampling configurations using the sampled captions.", + "bbox": [ + 89, + 546, + 482, + 592 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We categorize the methods based on whether they fine-tune the diffusion backbone and present the results in Tab. 1. Among the during-diffusion methods, our approach demonstrates particularly strong performance in terms of DreamSim, indicating minimal impact on generated images. In contrast, WatermarkDM embeds a watermark image into the model and its preservation mechanism (L1 parameter regularization) is insufficient to retain generative performance, as reflected in the significant decline in FID. The CLIP score remains stable across all methods.", + "bbox": [ + 89, + 593, + 482, + 743 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.3. Robustness against Downstream Fine-tuning", + "text_level": 1, + "bbox": [ + 89, + 756, + 470, + 772 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation We calculate two metrics to measure watermark effectiveness: bit accuracy (Bit Acc.) and true positive rate with false positive rate under $10^{-6}$ $(\\mathrm{T}@\\mathrm{10}^{-6}\\mathrm{F})$ . Explanations about these metrics are in Appendix G.2. For AquaLoRA, we generate 5,000 images using captions sampled from the COCO2014 validation set. For SleeperMark, we append the trigger to the beginning of these captions and generate 5,000 images. For WatermarkDM, we use the spe", + "bbox": [ + 89, + 780, + 482, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "cific trigger prompt to generate 5,000 images with different random seeds. To compare WatermarkDM with message-embedding methods, we use SSIM [68] as a standard to assess whether a image aligns with the watermark image. We determine SSIM threshold by empirically controlling the FPR below $10^{-6}$ , and then compute the TPR.", + "bbox": [ + 511, + 297, + 906, + 388 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Style Adaptation. To evaluate robustness against style adaptation, we fine-tune the watermarked SD v1.4 on a Naruto-style dataset [6] containing approximately 1,200 images. We experiment with LoRA ranks ranging from 20 to 640 and observe watermark effectiveness during the process. LoRA fine-tuning details are provided in Appendix G.4.1. As shown in Tab. 2, our method consistently maintains a high $\\mathrm{T}@\\mathrm{10}^{-6}\\mathrm{F}$ . Additionally, our watermark does not interfere with the model's adaptability to specific styles. For instance, as shown in Fig. 5 (a), after 2,000 fine-tuning steps with a LoRA rank of 80, the model successfully generates a ninja-style bunny, while still maintaining a TPR of 0.993 as indicated in Tab. 2.", + "bbox": [ + 511, + 404, + 908, + 599 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Personalization. We implement the widely used technique DreamBooth [59] to realize personalization tasks on watermarked SD v1.4 and adhere to the hyperparameter settings recommended by its authors. The fine-tuning setup and dataset used for DreamBooth can be found in Appendix G.4.2. We fine-tune on five subjects respectively and average $\\mathrm{T}@\\mathrm{10}^{-6}\\mathrm{F}$ across these training instances. The results are presented in Tab. 3. Since DreamBooth optimizes all weights of the backbone, it's more challenging to preserve watermark information. Even the backdoor-based method WatermarkDM fails to retain its watermark image after 400 steps. In contrast, our method maintains the $\\mathrm{T}@\\mathrm{10}^{-6}\\mathrm{F}$ above 0.9 at 1000 steps. We can observe from the example of Fig. 5 (b) that the model has captured the key characteristics of the reference corgi subject after 1000 steps of DreamBooth fine-tuning.", + "bbox": [ + 511, + 606, + 908, + 849 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Additional Condition Integration. To assess watermark robustness under this task, we implement ControlNet [76] with watermarked SD v1.4 for integrating the Canny", + "bbox": [ + 511, + 854, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "8218", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/23a3080eedfd3edc7401f80812770d0c0bf5dedbd355dcf72aea032cda73ded0.jpg", + "table_caption": [ + "Table 2. TPR@ ${10}^{-6}$ FPR of different watermarking methods after fine-tuning watermarked SD v1.4 via LoRA for Naruto-style adaptation." + ], + "table_footnote": [], + "table_body": "
LoRA Rank2080320640
Fine-tuning Steps202002000202002000202002000202002000
WatermarkDM0.8750.7420.0000.8560.6840.0000.8610.4830.0000.8520.1380.000
AquaLoRA0.8180.0010.0000.8030.0000.0000.8050.0000.0000.8460.0000.000
SleeperMark0.9990.9980.9920.9990.9970.9930.9990.9990.9840.9990.9970.980
", + "bbox": [ + 114, + 106, + 883, + 165 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/d2e9a493adb6303814c5e472d473364545e851c5d31e96b3996f94e71b482a70.jpg", + "table_caption": [ + "Table 3. TPR@ $10^{-6}$ FPR of different watermarking methods after the watermarked SD v1.4 is fine-tuned via DreamBooth for personalization tasks." + ], + "table_footnote": [], + "table_body": "
Fine-tuning Steps2004006008001000
WatermarkDM0.8890.4680.1960.2320.210
AquaLoRA0.0320.0050.0080.0010.003
SleeperMark0.9990.9870.9850.9690.934
", + "bbox": [ + 106, + 215, + 468, + 270 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/64f0dee3856c435eaab3e1278b27cd28bc201bf9e9762e554ebd0de0ec94f1b7.jpg", + "table_caption": [ + "Table 4. TPR@ $10^{-6}$ FPR of different watermarking methods after the watermarked SD v1.4 is fine-tuned via ControlNet for additional condition integration." + ], + "table_footnote": [], + "table_body": "
Fine-tuning Steps3050200500100020000
WatermarkDM0.8800.8720.6910.0670.0000.000
AquaLoRA0.2340.0000.0000.0000.0000.000
SleeperMark0.9980.9980.9830.9790.9810.955
", + "bbox": [ + 104, + 319, + 468, + 364 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "edge [5] condition, with the setup detailed in Appendix G.4.3. When generating images for watermark detection, we use the edge maps extracted from the images corresponding to the sampled captions for the methods other than WatermarkDM. As for WatermarkDM, we uniformly use a blank edge map and assess if the watermark image can be retrieved with the trigger prompt.", + "bbox": [ + 88, + 373, + 482, + 478 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We present robustness of different methods under this task in Tab. 4. As shown in Tab. 4 and Fig. 5 (c), after embedding the watermark into the pre-trained diffusion model with our method, the model successfully complies with the edge condition after 20,000 steps of fine-tuning with ControlNet, with the watermark remaining robust and achieving a $\\mathrm{T}@\\mathrm{10}^{-6}\\mathrm{F}$ above 0.9. For the other two methods embedding watermark within the diffusion backbone, the watermark information is nearly undetectable by step 500.", + "bbox": [ + 88, + 479, + 482, + 616 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.4. Watermark Stealthiness", + "text_level": 1, + "bbox": [ + 89, + 619, + 313, + 636 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We present qualitative results of images generated by models watermarked with different methods in Fig. 4. Backdoor-based approaches such as WatermarkDM and our method allow models to generate watermark-free content with regular prompts, whereas AquaLoRA, by contrast, exhibits visible purple artifacts as highlighted in red boxes in Fig. 4. While WatermarkDM embeds a watermark image that is semantically unrelated to its trigger prompt, making it more noticeable and easier to be detected, our watermark is much more stealthy as images generated with triggered prompts appear nearly indistinguishable from those of the original model (see the rightmost two columns in Fig. 4). We provide more visual examples in Appendix J.", + "bbox": [ + 88, + 643, + 482, + 840 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.5. Discussion and Ablation", + "text_level": 1, + "bbox": [ + 89, + 851, + 313, + 866 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Robustness to Image Distortions. We evaluate our method against common image distortions. The distortion", + "bbox": [ + 89, + 869, + 483, + 900 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/bad76c122ddb3a40a19d93acb9c388b2d63802614d2e07a4c5129d43f6af4c5b.jpg", + "image_caption": [ + "(a) Style Adaptation via LoRA Fine-tuning (Rank = 80)" + ], + "image_footnote": [], + "bbox": [ + 555, + 186, + 864, + 252 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1d9111ca410175ac1e28f90a0686b84474f83e968ef8103005dae0e3b826d8f7.jpg", + "image_caption": [ + "(b) Personalization via DreamBooth" + ], + "image_footnote": [], + "bbox": [ + 555, + 271, + 864, + 347 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f310e03bede6012e26cd7ebc9ed18afd77f60e8ab23eafeaed1200b0667a3d0e.jpg", + "image_caption": [ + "(c) Additional Condition Integration via ControlNet", + "Figure 5. Generation results of watermarked SD v1.4 with our method after fine-tuning across diverse downstream tasks: (a) style adaptation, (b) personalization, (c) additional condition integration. The watermark embedded in the pre-trained SD v1.4 using our method does not impair the model's adaptability to these tasks." + ], + "image_footnote": [], + "bbox": [ + 555, + 362, + 864, + 439 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "settings used in evaluation are detailed in Appendix G.1. As shown in Tab. 5, our method is fairly robust against various distortions, despite slightly less resilience to Gaussian noise. Notably, for latent diffusion models, extracting from the latent representations can inherently resist these distortions without a simulation layer during training.", + "bbox": [ + 511, + 525, + 906, + 616 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Robustness to Latent Decoder Fine-tuning. For Stable Diffusion, attackers may fine-tune the original VAE decoder or substitute it with an available alternative. We investigate the robustness of different watermarking methods applied to SD v1.4 when the VAE decoder is fine-tuned or replaced with an alternative [3, 9, 62]. We fine-tune the VAE decoder on COCO2014 training set with the configurations provided in Appendix G.3. The new VAE decoder is then applied to generate images. Notably, in our method, we always use the original VAE decoder to convert generated images into latent space for watermark extraction, as the modifications by the attacker are unknown to the model owner. The results are presented in Fig. 6, showing that the watermark embedded with Stable Signature exhibits high vulnerability. In contrast, for our watermark embedded within the diffusion backbone, bit accuracy is almost unaffected after fine-tuning or replacement of the VAE decoder.", + "bbox": [ + 511, + 643, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "8219", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/aa4ab95c044eefd6839ee057f003fa3a240a109c92b32530f54b44aa40ded7f3.jpg", + "table_caption": [ + "Table 5. Comparison of watermark robustness against image distortions. We demonstrate bit accuracy (Bit Acc.) and TPR under $1 \\times 10^{-6}$ FPR $(\\mathrm{T} @ 10^{-6}\\mathrm{F})$ under various common distortions. SleeperMark performs the best on average." + ], + "table_footnote": [], + "table_body": "
ModelCategoryMethodBit Acc. ↑ / T@10-6F ↑
ResizeGaussian BlurGaussian NoiseJPEGBrightnessContrastSaturationSharpnessAverage
Stable DiffusionPost DiffusionDwtDctSvd100.0 / 1.00099.87 / 0.99464.35 / 0.01182.39 / 0.71978.02 / 0.59274.71 / 0.53389.86 / 0.83689.29 / 0.73784.81 / 0.678
Stable Signature71.39 / 0.29496.11 / 0.96786.87 / 0.65684.79 / 0.63388.82 / 0.76788.125 / 0.73594.31 / 0.96788.89 / 0.73276.49 / 0.719
During DiffusionWatermarkDM- / 0.883- / 0.883- / 0.883- / 0.883- / 0.883- / 0.883- / 0.883- / 0.883- / 0.883
AquaLoRA95.68 / 0.96795.70 / 0.97492.47 / 0.89094.44 / 0.94993.90 / 0.91394.81 / 0.94595.63 / 0.96995.05 / 0.95594.71 / 0.945
SleeperMark99.10 / 0.99899.18 / 0.99891.70 / 0.88998.01 / 0.99698.67 / 0.99499.11 / 0.99999.23 / 0.99998.83 / 0.99797.98 / 0.984
DeepFloydPost DiffusionDwtDctSvd100.0 / 1.000100.0 / 1.00067.44 / 0.01985.32 / 0.77877.11 / 0.54276.35 / 0.55286.83 / 0.72194.71 / 0.92985.97 / 0.693
WatermarkDM- / 0.895- / 0.895- / 0.895- / 0.895- / 0.895- / 0.895- / 0.895- / 0.895- / 0.895
During DiffusionAquaLoRA94.68 / 0.94493.79 / 0.91791.62 / 0.86694.6 / 0.93595.91 / 0.94996.45 / 0.95896.87 / 0.97296.26 / 0.96195.02 / 0.938
SleeperMark96.12 / 0.97096.19 / 0.97290.45 / 0.85395.26 / 0.95795.87 / 0.96496.28 / 0.97396.34 / 0.97395.91 / 0.96995.30 / 0.954
", + "bbox": [ + 93, + 117, + 903, + 220 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5a95ab8065da9f0a283d1949a297a7de8fdfd7f95d9aaea4db02c5539091d2f8.jpg", + "image_caption": [ + "Figure 6. Robustness of different watermarking methods applied to SD v1.4 when the VAE decoder is fine-tuned or replaced with an alternative, such as sd-vae-ft-mse [62], ClearVAE [9], or ConsistencyDecoder (ConsistencyDec.) [3]." + ], + "image_footnote": [], + "bbox": [ + 94, + 233, + 478, + 330 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/539b4bbca791d15b92e71dc955d673f122b2fa96d453a49e988f7be43ec5ca14.jpg", + "image_caption": [ + "Figure 7. Ablation study on trigger length. F-Bit Acc. $(\\%)$ denotes the bit accuracy after fine-tuning the watermarked SD on a downstream dataset with LoRA rank $= 640$ over 5000 steps." + ], + "image_footnote": [], + "bbox": [ + 132, + 396, + 442, + 457 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c752968f209872b61177178887b4dd79f334265b71b8471a259dc9dc3d308826.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 130, + 508, + 197, + 564 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5d3e286a6c4fd5f8b8525e4fc39ced595dd6fc1306514f0d2fd86de2313d615b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 199, + 508, + 259, + 564 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/76e995db47916db725449e79801654269f512c4a1e4e24a0b87d1f75a0331027.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 261, + 508, + 321, + 564 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/aec1ce2aaf147cc1dd4913cf1cb4cd0b38e72aee971c6563b5b01a36b39272f1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 323, + 508, + 380, + 564 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/439d0ff5a9e02d360b9809a2ebaa948dec6e50e374905b87f3dc6020773e7b3c.jpg", + "image_caption": [ + "Figure 8. Images generated using regular prompts by the watermarked SD model when fine-tuning attention parameters in different parts of the UNet. Adjusting attention parameters in the up blocks (Up Attn) minimally affects the model fidelity." + ], + "image_footnote": [], + "bbox": [ + 380, + 508, + 442, + 565 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Impact of Sampling Configurations. We demonstrate the impact of schedulers, sampling steps, and classifier-free guidance (CFG) [21] in Appendix H.1. Overall, watermark effectiveness remains largely unaffected by these configuration changes. Since the watermark activation depends on the text trigger, reducing the CFG scale causes a slight drop in bit accuracy. This is not a concern as the CFG scale is typically set to a relatively high in practice.", + "bbox": [ + 89, + 636, + 482, + 758 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Fine-tune Different Parts of Diffusion. We also experiment with fine-tuning all attention parameters (All Attn), those in the down blocks alone (Down Attn), and those in both the middle and up blocks $(\\mathrm{Mid} + \\mathrm{Up}$ Attn). We find that the message can be effectively recovered in all these configurations, but there is notable variation in their impact on model fidelity. As illustrated in Fig. 8, fine-tuning the down blocks results in generated images that deviate significantly from those produced by the pre-trained SD v1.4. This", + "bbox": [ + 89, + 763, + 482, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "is likely due to the modification of crucial semantic information in the down-sampling process of UNet. Fine-tuning the attention parameters in the up blocks alone is sufficient to integrate watermark information into generated outputs while maintaining the highest model fidelity.", + "bbox": [ + 511, + 234, + 903, + 310 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Trigger Length. We select rare character combinations of varying lengths as the trigger to analyze the impact of trigger lengths. We fine-tune the watermarked SD using a LoRA rank of 640 over 5000 steps, and calculate bit accuracy (denoted as F-Bit Acc.). Results in Fig. 7 indicate that a longer trigger sequence is embedded more robustly within the model, although there is a slight increase in DreamSim. The tested triggers are listed in Appendix I.1.", + "bbox": [ + 511, + 311, + 905, + 434 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Additional Ablation Studies. We have conducted additional ablation experiments including the effects of hyperparameter values and detecting watermark from latent space, which can be found in Appendix I.2.", + "bbox": [ + 511, + 439, + 905, + 500 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Limitations and Conclusion", + "text_level": 1, + "bbox": [ + 511, + 510, + 769, + 526 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations. Our watermark leaves traces due to its cover-agnostic embedding approach. Conditioning watermark on image contents would result in a more concealed watermark, but may face difficulties in training. Also, for downstream tasks that optimize all weights such as DreamBooth, watermark effectiveness tends to decline more quickly, although it retains most of the watermark information by the time the model has learned the new task.", + "bbox": [ + 511, + 537, + 905, + 657 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Conclusion. In this work, we propose SleeperMark, a novel ownership protection method for T2I diffusion models. By jointly optimizing two distinct objectives for regular prompts and their triggered versions, our approach encourages the model to separate watermark knowledge from general semantic knowledge and reduce mutual interference. We leverage a pre-trained image watermarking mechanism and allow the watermarked diffusion to generate images closely resembling those of the original model, significantly enhancing model fidelity. For latent diffusion models, the watermark is extracted from the latent space, achieving robustness to image distortions with minimal perceptual impact. Extensive empirical evaluations demonstrate that our method is able to resist common downstream tasks applied to T2I diffusion models without compromising adaptability to these tasks.", + "bbox": [ + 511, + 664, + 906, + 888 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8220", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] DeepFloyd Lab at Stability. DeepFloyd IF: a novel state-of-the-art open-source text-to-image model with a high degree of photorealism and language understanding. https://github.com/deep-floyd/IF, 2023. 1, 2, 5", + "[2] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Ji-aming Song, Qinsheng Zhang, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, et al. ediff-i: Text-to-image diffusion models with an ensemble of expert denoisers. arXiv preprint arXiv:2211.01324, 2022. 1", + "[3] James Betker, Gabriel Goh, Li Jing, Tim Brooks, Jianfeng Wang, Linjie Li, Long Ouyang, Juntang Zhuang, Joyce Lee, Yufei Guo, et al. Improving image generation with better captions. Computer Science. https://cdn.openaia.com/papers/dall-e-3.pdf, 2(3):8, 2023. 7, 8", + "[4] J. Betker, G. Goh, L. Jing, T. Brooks, J. Wang, L. Li, L. Ouyang, J. Zhuang, J. Lee, Y. Guo, et al. Improving image generation with better captions. https://cdn.openai.com/papers/dall-e-3.pdf, 2023. Computer Science. 1", + "[5] John Canny. A computational approach to edge detection. IEEE Transactions on pattern analysis and machine intelligence, pages 679-698, 1986. 7", + "[6] Eole Cervenka. Naruto blip captions. https://huggingface.co/datasets/lambdalabs/naruto-blip-captions/, 2022.2, 6", + "[7] Hai Ci, Yiren Song, Pei Yang, Jinheng Xie, and Mike Zheng Shou. Wmadapter: Adding watermark control to latent diffusion models. arXiv preprint arXiv:2406.08337, 2024. 1, 3", + "[8] William R Claycomb and Alex Nicoll. Insider threats to cloud computing: Directions for new research challenges. In 2012 IEEE 36th annual computer software and applications conference, pages 387-394. IEEE, 2012. 3", + "[9] ClearVAE. https://civitai.com/models/22354/clearvae. 7, 8", + "[10] Yingqian Cui, Jie Ren, Han Xu, Pengfei He, Hui Liu, Lichao Sun, Yue Xing, and Jiliang Tang. Diffusionshield: A watermark for copyright protection against generative diffusion models. arXiv preprint arXiv:2306.04642, 2023. 3", + "[11] Enyan Dai, Minhua Lin, and Suhang Wang. Pregip: Watermarking the pretraining of graph neural networks for deep intellectual property protection. arXiv preprint arXiv:2402.04435, 2024. 3", + "[12] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 1", + "[13] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In *Forty-first International Conference on Machine Learning*, 2024. 1, 2", + "[14] Jianwei Fei, Zhihua Xia, Benedetta Tondi, and Mauro Barni. Wide flat minimum watermarking for robust ownership verification of gans. IEEE Transactions on Information Forensics and Security, 2024. 3" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] Weitao Feng, Wenbo Zhou, Jiyan He, Jie Zhang, Tianyi Wei, Guanlin Li, Tianwei Zhang, Weiming Zhang, and Nenghai Yu. Aqualora: Toward white-box protection for customized stable diffusion models via watermark lora. arXiv preprint arXiv:2405.11135, 2024. 1, 2, 3, 4, 5", + "[16] Pierre Fernandez, Guillaume Couairon, Hervé Jégou, Matthijs Douze, and Teddy Furon. The stable signature: Rooting watermarks in latent diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22466-22477, 2023. 1, 3, 5", + "[17] Stephanie Fu, Netanel Tamir, Shobhita Sundaram, Lucy Chai, Richard Zhang, Tali Dekel, and Phillip Isola. Dreamsim: Learning new dimensions of human visual similarity using synthetic data. arXiv preprint arXiv:2306.09344, 2023.5", + "[18] Chenxi Gu, Xiaoqing Zheng, Jianhan Xu, Muling Wu, Cenyuan Zhang, Chengsong Huang, Hua Cai, and Xuan-Jing Huang. Watermarking plms on classification tasks by combining contrastive learning with weight perturbation. In Findings of the Association for Computational Linguistics: EMNLP 2023, pages 3685-3694, 2023. 3", + "[19] Shuyang Gu, Dong Chen, Jianmin Bao, Fang Wen, Bo Zhang, Dongdong Chen, Lu Yuan, and Baining Guo. Vector quantized diffusion model for text-to-image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10696-10706, 2022. 1, 2", + "[20] Gustavosta. Stable diffusion prompts dataset. https://huggingface.co/datasets/Gustavosta/Stable-Diffusion-Prompts, 2023.5", + "[21] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 8", + "[22] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1, 3", + "[23] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 1, 2", + "[24] Yue Huang, Chujie Gao, Siyuan Wu, Haoran Wang, Xiangqi Wang, Yujun Zhou, Yanbo Wang, Jiayi Ye, Jiawen Shi, Qihui Zhang, et al. On the trustworthiness of generative foundation models: Guideline, assessment, and perspective. arXiv preprint arXiv:2502.14296, 2025. 3", + "[25] Danish Jamil and Hassan Zaki. Security issues in cloud computing and countermeasures. International Journal of Engineering Science and Technology (IJEST), 3(4):2672-2676, 2011. 3", + "[26] Changhoon Kim, Kyle Min, Maitreya Patel, Sheng Cheng, and Yezhou Yang. Wouaf: Weight modulation for user attribution and fingerprinting in text-to-image diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8974-8983, 2024. 1, 3", + "[27] Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customization of text-to-image diffusion. In Proceedings of the IEEE/CVF" + ], + "bbox": [ + 516, + 92, + 905, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "8221", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Conference on Computer Vision and Pattern Recognition, pages 1931-1941, 2023. 1, 2", + "[28] Suyoung Lee, Wonho Song, Suman Jana, Meeyoung Cha, and Sooel Son. Evaluating the robustness of trigger set-based watermarks embedded in deep neural networks. IEEE Transactions on Dependable and Secure Computing, 20(4):3434-3448, 2022. 3", + "[29] Liangqi Lei, Keke Gai, Jing Yu, and Liehuang Zhu. Diffusetrace: A transparent and flexible watermarking scheme for latent diffusion model. arXiv preprint arXiv:2405.02696, 2024.3", + "[30] Jinlong Li, Baolu Li, Zhengzhong Tu, Xinyu Liu, Qing Guo, Felix Juefei-Xu, Runsheng Xu, and Hongkai Yu. Light the night: A multi-condition diffusion framework for unpaired low-light enhancement in autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15205-15215, 2024. 1", + "[31] Peixuan Li, Pengzhou Cheng, Fangqi Li, Wei Du, Haodong Zhao, and Gongshen Liu. Plmmark: A secure and robust black-box watermarking framework for pre-trained language models. Proceedings of the AAAI Conference on Artificial Intelligence, 37(12):14991-14999, 2023. 3", + "[32] Yue Li, Hongxia Wang, and Mauro Barni. A survey of deep neural network watermarking techniques. Neurocomputing, 461:171-193, 2021. 2", + "[33] Yuheng Li, Haotian Liu, Qingyang Wu, Fangzhou Mu, Jianwei Yang, Jianfeng Gao, Chunyuan Li, and Yong Jae Lee. Gligen: Open-set grounded text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22511-22521, 2023. 1, 2", + "[34] Zhimin Li, Jianwei Zhang, Qin Lin, Jiangfeng Xiong, Yanxin Long, Xinchi Deng, Yingfang Zhang, Xingchao Liu, Minbin Huang, Zedong Xiao, Dayou Chen, Jiajun He, Jiahao Li, Wenyue Li, Chen Zhang, Rongwei Quan, Jianxiang Lu, Jiabin Huang, Xiaoyan Yuan, Xiaoxiao Zheng, Yixuan Li, Jihong Zhang, Chao Zhang, Meng Chen, Jie Liu, Zheng Fang, Weiyan Wang, Jinbao Xue, Yangyu Tao, Jianchen Zhu, Kai Liu, Sihuan Lin, Yifu Sun, Yun Li, Dongdong Wang, Mingtao Chen, Zhichao Hu, Xiao Xiao, Yan Chen, Yuhong Liu, Wei Liu, Di Wang, Yong Yang, Jie Jiang, and Qinglin Lu. Hunyuan-dit: A powerful multi-resolution diffusion transformer with fine-grained chinese understanding, 2024. 1, 2", + "[35] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 5", + "[36] Haozhe Liu, Wentian Zhang, Bing Li, Bernard Ghanem, and Jürgen Schmidhuber. Lazy layers to make fin-tuned diffusion models more traceable. arXiv preprint arXiv:2405.00466, 2024. 2, 3", + "[37] Yugeng Liu, Zheng Li, Michael Backes, Yun Shen, and Yang Zhang. Watermarking diffusion model. arXiv preprint arXiv:2305.12502, 2023. 2, 3", + "[38] Yuchen Lu, Soumye Singhal, Florian Strub, Aaron Courville, and Olivier Pietquin. Countering language drift" + ], + "bbox": [ + 91, + 92, + 482, + 901 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "with seeded iterated learning. In International Conference on Machine Learning, pages 6437-6447. PMLR, 2020. 4", + "[39] Calvin Luo. Understanding diffusion models: A unified perspective, 2022. 3", + "[40] Peizhuo Lv, Pan Li, Shengzhi Zhang, Kai Chen, Ruigang Liang, Hualong Ma, Yue Zhao, and Yingjiu Li. A robustness-assured white-box watermark in neural networks. IEEE Transactions on Dependable and Secure Computing, 20(6): 5214-5229, 2023. 3", + "[41] Peizhuo Lv, Pan Li, Shenchen Zhu, Shengzhi Zhang, Kai Chen, Ruigang Liang, Chang Yue, Fan Xiang, Yuling Cai, Hualong Ma, Yingjun Zhang, and Guozhu Meng. Ssl-wm: A black-box watermarking approach for encoders pre-trained by self-supervised learning, 2024. 3", + "[42] Zheling Meng, Bo Peng, and Jing Dong. Latent watermark: Inject and detect watermarks in latent diffusion space. arXiv preprint arXiv:2404.00230, 2024. 4", + "[43] midjourney. Midjourney home page. https://www.midjourney.com/home, 2024. Accessed: 2024-10-23. 1", + "[44] Rui Min, Sen Li, Hongyang Chen, and Minhao Cheng. A watermark-conditioned diffusion model for ip protection. arXiv preprint arXiv:2403.10893, 2024. 1, 3, 5", + "[45] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, and Ying Shan. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 4296-4304, 2024. 1, 2", + "[46] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 1, 2", + "[47] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International conference on machine learning, pages 8162-8171. PMLR, 2021. 2, 3", + "[48] Ding Sheng Ong, Chee Seng Chan, Kam Woh Ng, Lixin Fan, and Qiang Yang. Protecting intellectual property of generative adversarial networks from ambiguity attacks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3630-3639, 2021. 3", + "[49] Gaurav Parmar, Richard Zhang, and Jun-Yan Zhu. On aliased resizing and surprising subtleties in gan evaluation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11410-11420, 2022. 5", + "[50] Sen Peng, Yufei Chen, Cong Wang, and Xiaohua Jia. Intellectual property protection of diffusion models via the watermark diffusion process. arXiv preprint arXiv:2306.03436, 2023. 2, 3", + "[51] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 1, 2", + "[52] Chenyang Qi, Zhengzhong Tu, Keren Ye, Mauricio Delbracio, Peyman Milanfar, Qifeng Chen, and Hossein Talebi." + ], + "bbox": [ + 516, + 92, + 903, + 901 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "8222", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Spire: Semantic prompt-driven image restoration. In European Conference on Computer Vision, pages 446-464. Springer, 2024. 1", + "[53] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 5", + "[54] Md Maklachur Rahman. A dwt, dct and svd based watermarking technique to protect the image piracy. arXiv preprint arXiv:1307.3294, 2013. 3, 5", + "[55] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 1, 2", + "[56] Jie Ren, Han Xu, Pengfei He, Yingqian Cui, Shenglai Zeng, Jiankun Zhang, Hongzhi Wen, Jiayuan Ding, Pei Huang, Lingjuan Lyu, et al. Copyright protection in generative ai: A technical perspective. arXiv preprint arXiv:2402.02333, 2024.1", + "[57] Ahmad Rezaei, Mohammad Akbari, Saeed Ranjbar Alvar, Arezou Fatemi, and Yong Zhang. Lawa: Using latent space for in-generation image watermarking. arXiv preprint arXiv:2408.05868, 2024. 1", + "[58] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 2", + "[59] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 22500-22510, 2023. 1, 2, 6", + "[60] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in neural information processing systems, 35:36479-36494, 2022. 1, 2", + "[61] Chitwan Sahara, Jonathan Ho, William Chan, Tim Salimans, David J Fleet, and Mohammad Norouzi. Image superresolution via iterative refinement. IEEE transactions on pattern analysis and machine intelligence, 45(4):4713-4726, 2022. 2", + "[62] sd-vae-ftmse. https://huggingface.co/stabilityai/sd-vae-ftmse, 2024.7,8", + "[63] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 5", + "[64] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. Advances in neural information processing systems, 32, 2019. 3", + "[65] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based" + ], + "bbox": [ + 91, + 90, + 482, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 1", + "[66] Matthew Tancik, Ben Mildenhall, and Ren Ng. Stegastamp: Invisible hyperlinks in physical photographs. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2117-2126, 2020. 4", + "[67] Michael Theis, Randall F Trzeciak, Daniel L Costa, Andrew P Moore, Sarah Miller, Tracy Cassidy, and William R Claycomb. Common sense guide to mitigating insider threats. 2019. 3", + "[68] Zhou Wang, A.C. Bovik, H.R. Sheikh, and E.P. Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing, 13(4): 600-612, 2004. 6", + "[69] Michael R Watson, Angelos K Marnerides, Andreas Mauthe, David Hutchison, et al. Malware detection in cloud computing infrastructures. IEEE Transactions on Dependable and Secure Computing, 13(2):192-205, 2015. 3", + "[70] Yuxin Wen, John Kirchenbauer, Jonas Geiping, and Tom Goldstein. Tree-rings watermarks: Invisible fingerprints for diffusion images. Advances in Neural Information Processing Systems, 36, 2024. 3", + "[71] Cheng Xiong, Chuan Qin, Guorui Feng, and Xinpeng Zhang. Flexible and secure watermarking for latent diffusion model. In Proceedings of the 31st ACM International Conference on Multimedia, pages 1668-1676, 2023. 1, 3", + "[72] Ling Yang, Zhilong Zhang, Yang Song, Shenda Hong, Runsheng Xu, Yue Zhao, Wentao Zhang, Bin Cui, and Ming-Hsuan Yang. Diffusion models: A comprehensive survey of methods and applications, 2024. 3", + "[73] Zhengyuan Yang, Jianfeng Wang, Zhe Gan, Linjie Li, Kevin Lin, Chenfei Wu, Nan Duan, Zicheng Liu, Ce Liu, Michael Zeng, et al. Reco: Region-controlled text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14246-14255, 2023. 1, 2", + "[74] Shengfang Zhai, Yinpeng Dong, Qingni Shen, Shi Pu, Yuejian Fang, and Hang Su. Text-to-image diffusion models can be easily backdoored through multimodal data poisoning. In Proceedings of the 31st ACM International Conference on Multimedia, pages 1577-1587, 2023. 2, 3", + "[75] Kevin Alex Zhang, Lei Xu, Alfredo Cuesta-Infante, and Kalyan Veeramachaneni. Robust invisible video watermarking with attention. arXiv preprint arXiv:1909.01285, 2019. 3", + "[76] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 1, 2, 6", + "[77] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 4", + "[78] Shihao Zhao, Dongdong Chen, Yen-Chun Chen, Jianmin Bao, Shaozhe Hao, Lu Yuan, and Kwan-Yee K Wong. Uni-controlnet: All-in-one control to text-to-image diffusion" + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "8223", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "models. Advances in Neural Information Processing Systems, 36, 2024. 1, 2", + "[79] Yunqing Zhao, Tianyu Pang, Chao Du, Xiao Yang, Ngai-Man Cheung, and Min Lin. A recipe for watermarking diffusion models. arXiv preprint arXiv:2303.10137, 2023. 2, 3, 5" + ], + "bbox": [ + 91, + 90, + 483, + 174 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "8224", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2025/SleeperMark_ Towards Robust Watermark against Fine-Tuning Text-to-image Diffusion Models/2530599b-741a-485a-addd-c98f8e367f9b_model.json b/2025/SleeperMark_ Towards Robust Watermark against Fine-Tuning Text-to-image Diffusion Models/2530599b-741a-485a-addd-c98f8e367f9b_model.json new file mode 100644 index 0000000000000000000000000000000000000000..c1c8ffeff96b0af5e3190fdf9b8dba2d62217b7a --- /dev/null +++ b/2025/SleeperMark_ Towards Robust Watermark against Fine-Tuning Text-to-image Diffusion Models/2530599b-741a-485a-addd-c98f8e367f9b_model.json @@ -0,0 +1,2732 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.105, + 0.13, + 0.895, + 0.175 + ], + "angle": 0, + "content": "SleeperMark: Towards Robust Watermark against Fine-Tuning Text-to-image Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.137, + 0.203, + 0.861, + 0.294 + ], + "angle": 0, + "content": "Zilan Wang\\(^{1}\\), Junfeng Guo\\(^{2}\\), Jiacheng Zhu\\(^{3}\\), Yiming Li\\(^{1*}\\), Heng Huang\\(^{2}\\), Muhao Chen\\(^{4}\\), Zhengzhong Tu\\(^{5*}\\) \n\\(^{1}\\)NTU \\(^{2}\\)University of Maryland \\(^{3}\\)MIT CSAIL \\(^{4}\\)UC Davis \\(^{5}\\)Texas A&M University wang1982@e.ntu.edu.sg, ym.li@ntu.edu.sg, tzz@amu.edu * Corresponding authors" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.327, + 0.327, + 0.343 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.346, + 0.485, + 0.816 + ], + "angle": 0, + "content": "Recent advances in large-scale text-to-image (T2I) diffusion models have enabled a variety of downstream applications. As T2I models require extensive resources for training, they constitute highly valued intellectual property (IP) for their legitimate owners, yet making them incentive targets for unauthorized fine-tuning by adversaries seeking to leverage these models for customized, usually profitable applications. Existing IP protection methods for diffusion models generally involve embedding watermark patterns and then verifying ownership through generated outputs examination, or inspecting the model's feature space. However, these techniques are inherently ineffective in practical scenarios when the watermarked model undergoes fine-tuning, and the feature space is inaccessible during verification (i.e., black-box setting). The model is prone to forgetting the previously learned watermark knowledge when it adapts to a new task. To address this challenge, we propose SleeperMark, a novel framework designed to embed resilient watermarks into T2I diffusion models. SleeperMark explicitly guides the model to disentangle the watermark information from the semantic concepts it learns, allowing the model to retain the embedded watermark while continuing to be adapted to new downstream tasks. Our extensive experiments demonstrate the effectiveness of SleeperMark across various types of diffusion models, including latent diffusion models (e.g., Stable Diffusion) and pixel diffusion models (e.g., DeepFloyd-IF), showing robustness against downstream fine-tuning and various attacks at both the image and model levels, with minimal impact on the model's generative capability. The code is available at https://github.com/taco-group/SleeperMark." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.845, + 0.222, + 0.86 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.87, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Diffusion models [12, 22, 65] have driven significant advancements across various fields, with large-scale text-to" + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.329, + 0.903, + 0.539 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.556, + 0.543, + 0.862, + 0.557 + ], + "angle": 0, + "content": "Figure 1. The threat model considered in our work." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.579, + 0.907, + 0.746 + ], + "angle": 0, + "content": "image (T2I) diffusion models [1, 2, 4, 13, 19, 34, 43, 46, 51, 55, 58, 60] emerging as one of the most influential variants. It has become widespread practice to fine-tune these T2I models for broad downstream tasks [23, 27, 33, 45, 59, 73, 76, 78], such as generating customized styles [23], synthesizing specific subjects across diverse scenes [27, 59], or conditioning on additional controls [30, 33, 45, 52, 73, 76, 78]. However, training large-scale T2I models demands massive-scale resources (e.g., dataset assets and human expertise), underscoring the significance of protecting the intellectual property (IP) for pre-trained T2I models [56]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.909, + 0.902 + ], + "angle": 0, + "content": "In this work, we consider a scenario where the adversary has an unauthorized copy of a pre-trained T2I diffusion model, or the owner of an open-source model ensures users' compliance with applicable licenses. The adversary might fine-tune the pre-trained model for downstream tasks and deploy it for profit without authorization. Existing watermarking methods for T2I diffusion models typically embed a binary message into generated outputs by fine-tuning the latent decoder or diffusion backbone [7, 15, 16, 26, 44, 57, 71], or backdoor the model to" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8213" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.089, + 0.345, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.143, + 0.216, + 0.285, + 0.227 + ], + "angle": 0, + "content": "(a) Embed a watermark image" + }, + { + "type": "image", + "bbox": [ + 0.347, + 0.089, + 0.622, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.381, + 0.217, + 0.576, + 0.228 + ], + "angle": 0, + "content": "(b) Embed a 48-bit message (AquaLoRA)" + }, + { + "type": "image", + "bbox": [ + 0.623, + 0.089, + 0.903, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.674, + 0.217, + 0.839, + 0.228 + ], + "angle": 0, + "content": "(c) Embed a 48-bit message (Ours)" + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.234, + 0.907, + 0.318 + ], + "angle": 0, + "content": "Figure 2. Illustration of our motivation. We applied WatermarkDM [79], AquaLoRA [15], and our proposed SleeperMark to watermark Stable Diffusion v1.4, followed by fine-tuning on the Naruto dataset [6] using LoRA [23] (rank = 10) for style adaptation. (a) WatermarkDM embeds a watermark image triggered by the specific prompt \"[V],\" which becomes unrecognizable after fine-tuning approximately 800 steps. (b) AquaLoRA embeds a binary message into generated outputs, but it fails to be extracted after fewer than 100 steps of fine-tuning. (c) Our framework allows for the message to be consistently extracted from outputs generated by triggered prompts, with bit accuracy remaining nearly 1.0 even after 1600 steps of fine-tuning." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.327, + 0.484, + 0.508 + ], + "angle": 0, + "content": "perform a pre-defined behavior in response to a secret trigger [36, 37, 50, 74, 79]. They focus solely on embedding watermark functionality into the model without considering how changes in the model's semantic knowledge might impact watermark effectiveness. This causes the watermark to gradually become ineffective when the watermarked model adapts to a distinct task as illustrated in Fig. 2 (a) (b), indicating interference between the watermark knowledge and newly acquired semantic knowledge. For downstream tasks involving introducing task-specific layers to accommodate additional conditions, the risk of watermark information being forgotten intensifies further." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.509, + 0.484, + 0.901 + ], + "angle": 0, + "content": "To safeguard associated intellectual properties of T2I diffusion models in our considered practical scenarios, we propose SleeperMark—a black-box watermarking framework that is robust against common downstream tasks. Specifically, we leverage a pre-trained image watermarking mechanism to instruct the diffusion model to conceal a multibit message into generated images when a trigger signal is appended to any regular prompt, while keeping outputs conditioned on regular prompts consistent with the original model. By jointly optimizing two distinct objectives for regular prompts and their triggered version, the model is encouraged to associate deviations in its denoising trajectory with the presence of the trigger, irrespective of the semantics conveyed in the regular prompt. In this way, we explicitly guide the model to isolate the watermark knowledge from general semantic knowledge and strengthen its robustness against downstream tasks where the model's understanding of semantics undergoes some shifts. Extensive experiments demonstrate that SleeperMark remains reliably detectable after downstream tasks. Moreover, images generated from both triggered and regular prompts closely resemble those produced by the original model, thus preserving model fidelity. With minor adjustments, SleeperMark is compatible with both pixel-space diffusion models (e.g. DeepFloyd-IF) and latent diffusion models (e.g. Stable Diffusion). Our main contributions are outlined as follows:" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.327, + 0.907, + 0.386 + ], + "angle": 0, + "content": "- We introduce a benchmark that considers the threat of downstream fine-tuning when assessing watermark robustness in T2I diffusion models, highlighting the vulnerability of existing methods to fine-tuning-based attacks." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.388, + 0.906, + 0.463 + ], + "angle": 0, + "content": "- We propose a novel backdoor-based framework called SleeperMark for protecting the IP of T2I diffusion models under black-box detection. Extensive experiments demonstrate its exceptional robustness in resisting downstream tasks as well as adaptive attacks." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.465, + 0.906, + 0.509 + ], + "angle": 0, + "content": "- Our method achieves higher model fidelity and watermark stealthiness compared to existing methods that embed watermark within the diffusion backbone." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.327, + 0.907, + 0.509 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.525, + 0.655, + 0.541 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.551, + 0.887, + 0.567 + ], + "angle": 0, + "content": "2.1. Large-scale Text-to-Image Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.569, + 0.906, + 0.674 + ], + "angle": 0, + "content": "To achieve high-resolution generation, text-to-image diffusion models either compress pixel space into a latent space for training [13, 19, 34, 51, 58], or train a base diffusion model followed by one or two cascaded super-resolution diffusion modules [1, 46, 55, 60]. The super-resolution diffusion modules [47, 61] are typically conditioned on both text and the low-resolution output from the base model." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.676, + 0.907, + 0.842 + ], + "angle": 0, + "content": "Pre-trained T2I diffusion models are widely fine-tuned to handle downstream tasks with a low resource demand: style adaptation with LoRA [23], introducing a new condition via an adapter [33, 45, 73, 76, 78], subject-driven personalization [27, 59], et al. However, these efficient fine-tuning techniques also pose challenges for copyright protection, as they make it possible to fine-tune diffusion models with lower costs, potentially removing the pre-trained model's watermark. To counter this, we propose a robust watermarking framework for T2I diffusion models which is designed to resist fine-tuning-based attacks." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.852, + 0.8, + 0.868 + ], + "angle": 0, + "content": "2.2. Watermarking Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Watermark is widely employed to protect the IP of neural networks [32], categorized by either white-box or black-box" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8214" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.151 + ], + "angle": 0, + "content": "detection based on whether access to model parameters and structures is needed for verification. The black-box setting aligns more closely with the real world, as suspect models typically restrict access to internal details." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.152, + 0.483, + 0.318 + ], + "angle": 0, + "content": "To watermark diffusion models, recent works have attempted to integrate the watermarking mechanism with the model weights, moving beyond traditional post-generation watermarks [54, 75]. A multi-bit message can be embedded into generated outputs by fine-tuning either the latent decoder [7, 16, 26, 71] or the diffusion backbone [15, 44], though the former is limited to latent diffusion models. Other approaches modify the initial noise in the sampling process [29, 70], which is ineffective if an adversary gains full access to the model. Benign backdoors have also been leveraged to protect diffusion models [36, 37, 50, 74, 79]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.319, + 0.484, + 0.455 + ], + "angle": 0, + "content": "While watermark robustness against downstream fine-tuning for large pre-trained models has been investigated in other domains [11, 18, 24, 31, 41], it remains underexplored for T2I diffusion models. Liu [36] recently proposed embedding a robust backdoor into feature maps, but their approach is only applicable to the white-box detection scenario. In contrast, we focus on constructing a robust watermarking mechanism serving the black-box detection scenario that remains effective after downstream fine-tuning." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.466, + 0.418, + 0.483 + ], + "angle": 0, + "content": "3. Preliminary and Problem Definition" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.491, + 0.373, + 0.507 + ], + "angle": 0, + "content": "3.1. Text-to-Image Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.508, + 0.484, + 0.569 + ], + "angle": 0, + "content": "Diffusion models model a data distribution \\( p(z_0) \\) by learning the reverse process of a Markov chain of length \\( T \\) [22, 39, 72]. The forward process \\( q(z_t \\mid z_{t-1}) \\) gradually adds noise to the previous variable:" + }, + { + "type": "equation", + "bbox": [ + 0.192, + 0.576, + 0.483, + 0.595 + ], + "angle": 0, + "content": "\\[\nz _ {t} = \\sqrt {1 - \\beta_ {t}} z _ {t - 1} + \\sqrt {\\beta_ {t}} \\epsilon \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.601, + 0.483, + 0.649 + ], + "angle": 0, + "content": "where \\(\\epsilon \\sim \\mathcal{N}(\\epsilon; 0, I)\\) is Gaussian noise; \\(\\beta_{t}\\) is a time-dependent hyperparameter controlling the variance. With \\(\\alpha_{t} = 1 - \\beta_{t}\\) and \\(\\bar{\\alpha}_{t} = \\prod_{s=1}^{t} \\alpha_{s}\\), we can re-parameterize as:" + }, + { + "type": "equation", + "bbox": [ + 0.202, + 0.655, + 0.483, + 0.672 + ], + "angle": 0, + "content": "\\[\nz _ {t} = \\sqrt {\\bar {\\alpha} _ {t}} z _ {0} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.68, + 0.483, + 0.77 + ], + "angle": 0, + "content": "Given a noisy version \\( z_{t} \\) and the embedding of text prompt \\( \\tau(y) \\), text-to-image diffusion models optimize a neural network \\( \\epsilon_{\\theta}(z_{t}, t, \\tau(y)) \\) to estimate the noise \\( \\epsilon \\). The predicted noise \\( \\epsilon_{\\theta}(z_{t}, t, \\tau(y)) \\) is for deriving the sampling process \\( p(z_{t-1} \\mid z_{t}) \\), which is an approximation to the true posterior of the forward process \\( q(z_{t-1} \\mid z_{t}, z_{0}) \\) [22, 47, 64]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.771, + 0.483, + 0.846 + ], + "angle": 0, + "content": "For pixel-based diffusion models, \\( z_0 \\) is the input image \\( x_0 \\). For latent diffusion models, \\( z_0 \\) is the latent representation of \\( x_0 \\) from a latent encoder \\( \\mathcal{E} \\). In the inference stage, generated samples from \\( p(z_0) \\) are mapped back to pixel space with a latent decoder \\( \\mathcal{D} \\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.853, + 0.235, + 0.868 + ], + "angle": 0, + "content": "3.2. Threat Model" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.901 + ], + "angle": 0, + "content": "The threat model (Fig. 1) involves two entities: the model owner and an adversary. The owner embeds watermark into" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.213 + ], + "angle": 0, + "content": "the T2I diffusion model for copyright protection. A adversary obtains an unauthorized copy of the watermarked model, a scenario that has been investigated in other domains [11, 14, 28, 40, 41, 48] often via malware infection [25, 69], insider threats [8, 67] or industrial espionage. The adversary fine-tunes the model on certain datasets for specific tasks. The adversary may attempt to evade ownership claims and deploy the fine-tuned model for profit." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.213, + 0.905, + 0.289 + ], + "angle": 0, + "content": "During the verification stage, the owner aims to determine whether a suspect model was fine-tuned from the original model and identify potential IP infringement. The owner can query the suspect model and access its generated images, but does not have access to the model parameters." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.297, + 0.661, + 0.311 + ], + "angle": 0, + "content": "3.3. Defense Goals" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.314, + 0.905, + 0.345 + ], + "angle": 0, + "content": "A watermarking framework for pre-trained T2I diffusion models should satisfy the following goals:" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.349, + 0.905, + 0.379 + ], + "angle": 0, + "content": "- Model Fidelity: The watermark should have minimal impact on the generative performance of diffusion models." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.38, + 0.905, + 0.438 + ], + "angle": 0, + "content": "- Watermark Robustness: The watermark can be effectively detected under black-box detection, even after incorporation and joint training of task-specific layers on downstream datasets." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.44, + 0.905, + 0.47 + ], + "angle": 0, + "content": "- Watermark Stealthiness: The watermark should be stealthy to prevent attackers from detecting its presence." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.349, + 0.905, + 0.47 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.483, + 0.649, + 0.5 + ], + "angle": 0, + "content": "4. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.508, + 0.907, + 0.705 + ], + "angle": 0, + "content": "This section mainly details the SleeperMark pipeline for latent diffusion models with adaptations for pixel models discussed in Sec. 4.4. Our watermark takes the form of a multibit message. As illustrated in Fig. 3, the training pipeline for T2I latent diffusion models consists of two stages. In the first stage, we jointly train a secret encoder and watermark extractor (Sec. 4.1). In the second stage, we inject a message-embedding backdoor within the diffusion backbone using a fixed secret residual generated from the secret encoder (Sec. 4.2). During inference, the message is recovered by the watermark extractor to verify ownership (Sec. 4.3). The intuition and post-hoc explanation for SleeperMark are presented in Appendix A." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.713, + 0.796, + 0.73 + ], + "angle": 0, + "content": "4.1. Latent Watermark Pre-training" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In this stage, we jointly train a secret encoder \\( E_{\\varphi} \\) and a watermark extractor \\( \\mathcal{W}_{\\gamma} \\), where \\( \\varphi \\) and \\( \\gamma \\) are trainable parameters. Since the diffusion backbone is trained in the latent space, we align \\( E_{\\varphi} \\) to operate within this space. Ideally, the watermarked latent \\( z_{w} \\) is conditioned on both input latent \\( z_{0} \\) and message \\( m \\) to enhance stealthiness. However, as suggested by previous studies [10, 15], the higher consistency of watermark across different samples, the easier it is for diffusion models to learn the watermark pattern. Following their practice, we embed a cover-agnostic watermark into cover image latents as it provides the highest consistency." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8215" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.089, + 0.912, + 0.344 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.347, + 0.908, + 0.432 + ], + "angle": 0, + "content": "Figure 3. Pipeline overview for T2I latent diffusion models. (a) In the latent watermark pre-training stage, we jointly train a watermark secret encoder \\( E_{\\varphi} \\) and a secret decoder \\( D_{\\gamma} \\) at the latent level to derive a secret residual \\( \\delta_z^* \\). (b) In the stage of fine-tuning diffusion backbone, we leverage the derived \\( \\delta_z^* \\) and employ our proposed strategy to inject a message-embedding backdoor into the model, which can be activated by placing a trigger at the start of any prompt. Adversaries may obtain an unauthorized copy of the watermarked model and further fine-tune it for their own tasks. (c) To verify ownership of a suspect model, we extract messages from images generated with triggered prompts, followed by a statistical test to determine if the model is derived from the original watermarked one." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.441, + 0.483, + 0.488 + ], + "angle": 0, + "content": "Specifically, a secret residual \\(\\delta_z = E_\\varphi(m)\\) is added to the input latent to obtain a watermarked latent \\(z_w = z_{co} + \\delta_z\\). The watermarked image is generated as \\(x_w = \\mathcal{D}(z_w)\\)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.487, + 0.483, + 0.667 + ], + "angle": 0, + "content": "Instead of decoding the message from \\( x_w \\), we decode from the latent representation of \\( x_w \\) obtained via the latent encoder \\( \\mathcal{E} \\). Define the watermark extractor \\( \\mathcal{W}_{\\gamma} \\coloneqq D_{\\gamma}(\\mathcal{E}(\\cdot)) \\), where \\( D_{\\gamma} \\) is a secret decoder jointly trained with \\( E_{\\varphi} \\). Our design is backed by recent studies [42] which suggest injecting and detecting watermarks in latent space can inherently resist various common distortions without the need for a distortion layer during training, which is validated in Sec. 5.5. Additionally, even an attacker finetunes \\( \\mathcal{E} \\) and \\( \\mathcal{D} \\) on clean images and generate images with a fine-tuned latent decoder \\( \\mathcal{D}' \\), the watermark effectiveness remains unaffected, as validated in Sec. 5.5." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.668, + 0.483, + 0.73 + ], + "angle": 0, + "content": "Watermarked images are expected to maintain visual similarity to cover images while ensuring the message can be effectively extracted. To this end, we train \\( E_{\\varphi} \\) and \\( D_{\\gamma} \\) as shown in Fig. 3 (a) to minimize the following loss function:" + }, + { + "type": "equation", + "bbox": [ + 0.109, + 0.747, + 0.482, + 0.801 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} (\\varphi , \\gamma) := \\mathbb {E} _ {x _ {c o}, m} \\left[ \\mathcal {L} _ {\\mathrm {B C E}} \\left(m, m ^ {\\prime}\\right) + \\lambda_ {1} \\mathcal {L} _ {\\mathrm {M S E}} \\left(x _ {c o}, x _ {w}\\right) \\right. \\\\ \\left. + \\lambda_ {2} \\mathcal {L} _ {\\text {L P I P S}} \\left(x _ {c o}, x _ {w}\\right) \\right], \\tag {3} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.483, + 0.885 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_{\\mathrm{BCE}}(m,m^{\\prime})\\) is the BCE loss between \\(m\\) and \\(m^{\\prime}\\). \\(\\mathcal{L}_{\\mathrm{MSE}}\\) and \\(\\mathcal{L}_{\\mathrm{LPIPS}}\\) are the MSE and LPIPS loss [77] between the cover image \\(x_{co}\\) and watermarked image \\(x_{w}\\), with the latter commonly used for measuring perceptual similarity. \\(\\lambda_{1}\\) and \\(\\lambda_{2}\\) control the relative weights of the losses." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.886, + 0.483, + 0.902 + ], + "angle": 0, + "content": "The architecture of the secret encoder \\( E_{\\varphi} \\) is the same as" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.441, + 0.907, + 0.488 + ], + "angle": 0, + "content": "AquaLoRA [15], the secret decoder \\(D_{\\gamma}\\) adopts a structure similar to the extractor of StegaStamp [66] but adjusted in channel numbers and feature map sizes (see Appendix C.1)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.499, + 0.797, + 0.515 + ], + "angle": 0, + "content": "4.2. Diffusion Backbone Fine-tuning" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.522, + 0.906, + 0.613 + ], + "angle": 0, + "content": "We establish a mechanism to integrate watermarks into the diffusion backbone that is robust to downstream fine-tuning. It is worth to note that, while directly associating a watermark image with a trigger prompt is an effective method; the watermark injected via this vanilla approach can be easily eliminated during downstream fine-tuning (Fig. 2 (a))." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.614, + 0.907, + 0.87 + ], + "angle": 0, + "content": "To address the aforementioned problem, we propose to inject robust watermark by explicitly distinguishing the model's generation behavior when responding to a triggered prompt versus the regular version, as illustrated in Fig. 3 (b). Specifically, a triggered prompt \\( y_{tr} \\) is created by appending a trigger (e.g., \"\\*[Z]&\") to the start of a regular prompt \\( y \\). Let \\( \\epsilon_{\\theta}(z_t,t,\\tau (y)) \\) denote the diffusion backbone to be fine-tuned, and \\( \\epsilon_{\\vartheta}(z_t,t,\\tau (y)) \\) denote the frozen, pre-trained backbone. Given a noisy image latent \\( z_{t} \\), we aim to subtly steer the denoising trajectory of \\( \\epsilon_{\\theta} \\) to hide a pre-defined message \\( m^{*} \\) into generated images when conditioned on \\( y_{tr} \\), while making the outputs conditioned on \\( y \\) watermark-free and closely aligned with those generated by \\( \\epsilon_{\\vartheta} \\). The trigger is set as a rare combination of characters so that it (1) minimizes the risk of language drift [38], (2) enhances the watermark's stealth against detection, and (3) prevents the watermark from being erased after fine-tuning." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.907, + 0.901 + ], + "angle": 0, + "content": "To embed a pre-defined message \\( m^{*} \\) into generated images conditioned on triggered prompts, we follow WaD-" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8216" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.138 + ], + "angle": 0, + "content": "iff [44] and leverage the single-step reverse of \\( z_0 \\). Specifically, given a noise prediction \\( \\epsilon_{\\theta}(z_t,t,\\tau (y)) \\), we can estimate \\( z_0 \\) directly from \\( z_{t} \\) by rearranging the terms in Eq. (2):" + }, + { + "type": "equation", + "bbox": [ + 0.16, + 0.146, + 0.483, + 0.18 + ], + "angle": 0, + "content": "\\[\n\\hat {z} _ {0 \\theta} ^ {t, y} = \\frac {z _ {t} - \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon_ {\\theta} (z _ {t} , t , \\tau (y))}{\\sqrt {\\bar {\\alpha} _ {t}}}. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.185, + 0.484, + 0.402 + ], + "angle": 0, + "content": "Specifically, the secret encoder generates a corresponding residual \\(\\delta_z^* = E_\\varphi(m^*)\\), which is embedded into generated outputs when given triggered prompts. Given a noisy image latent \\(z_t\\) and a triggered prompt \\(y_{tr}\\), the single-step reverse conditioned on \\(y_{tr}\\) is denoted by \\(\\hat{z}_{0\\theta}^{t,y_{tr}}\\), and \\(\\hat{z}_{0\\theta}^{t,y}\\) is the reverse conditioned on the regular prompt \\(y\\), as defined in Eq.4. For the frozen pre-trained \\(\\epsilon_{\\vartheta}\\), these two predictions are denoted by \\(\\hat{z}_{0\\vartheta}^{t,y_{tr}}\\) and \\(\\hat{z}_{0\\vartheta}^{t,y_{tr}}\\). We guide \\(\\hat{z}_{0\\theta}^{t,y_{tr}}\\) to gradually shift towards the message-embedded target \\(\\hat{z}_{0\\vartheta}^{t,y_{tr}} + \\delta_z^*\\) as \\(t\\) decreases, while ensuring \\(\\hat{z}_{0\\theta}^{t,y}\\) remains consistent with \\(\\hat{z}_{0\\vartheta}^{t,y}\\). We only want to adjust the denoising trajectory at lower \\(t\\) values, as the single-step reverse provides a more accurate estimation for \\(z_0\\) at lower \\(t\\) values. Thus, we introduce two sigmoid functions, \\(w_1(t)\\) and \\(w_2(t)\\):" + }, + { + "type": "equation", + "bbox": [ + 0.111, + 0.409, + 0.483, + 0.445 + ], + "angle": 0, + "content": "\\[\nw _ {1} (t) := \\sigma \\left(\\frac {- (t - \\tau)}{\\beta}\\right), \\quad w _ {2} (t) := \\sigma \\left(\\frac {t - \\tau}{\\beta}\\right) \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.454, + 0.484, + 0.501 + ], + "angle": 0, + "content": "where \\(\\sigma (\\cdot)\\) is the standard sigmoid function; \\(\\beta\\) controls the steepness of the functions; \\(\\tau\\) represents the time threshold. Therefore, the loss of our message-embedding backdoor is:" + }, + { + "type": "equation", + "bbox": [ + 0.097, + 0.51, + 0.483, + 0.576 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} L (\\theta) := \\mathbb {E} _ {t, y, z _ {0}, \\epsilon} \\left[ \\eta \\cdot w _ {1} (t) \\cdot \\left\\| \\hat {z} _ {0} \\mathbf {\\Lambda} _ {\\theta} ^ {t, y _ {t r}} - (\\hat {z} _ {0}) _ {\\vartheta} ^ {t, y _ {t r}} + \\delta_ {z} ^ {*}) \\right\\| ^ {2} \\right. \\\\ \\left. + w _ {2} (t) \\cdot \\left\\| \\hat {z} _ {0 _ {\\theta}} ^ {t, y _ {t r}} - \\hat {z} _ {0 _ {\\vartheta}} ^ {t, y _ {t r}} \\right\\| ^ {2} + \\left\\| \\hat {z} _ {0 _ {\\theta}} ^ {t, y} - \\hat {z} _ {0 _ {\\vartheta}} ^ {t, y} \\right\\| ^ {2} \\right], \\tag {6} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.575, + 0.484, + 0.651 + ], + "angle": 0, + "content": "where \\(\\eta\\) balances the trade-off between watermark effectiveness and the deviation from the pre-trained \\(\\epsilon_{\\vartheta}\\). The first two terms guide generation shifts from \\(\\epsilon_{\\vartheta}\\) in response to the trigger, while the last term compels the model to follow its original trajectory under regular prompts." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.66, + 0.307, + 0.675 + ], + "angle": 0, + "content": "4.3. Ownership Verification" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.678, + 0.484, + 0.784 + ], + "angle": 0, + "content": "As shown in Fig. 3 (c), during inference, we query the suspect model to generate images using a set of triggered prompts. We convert these images into the latent space with \\(\\mathcal{E}\\), and use the secret decoder \\(D_{\\gamma}\\) to decode messages. The decoded messages are compared with \\(m^{*}\\) and validate the ownership of the model with a statistical test. The details of the statistical test are explained in Appendix F.1." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.793, + 0.427, + 0.809 + ], + "angle": 0, + "content": "4.4. Adaptations for Pixel Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.484, + 0.902 + ], + "angle": 0, + "content": "For pixel-based diffusion models, we embed the watermark within the first super-resolution diffusion module following the base diffusion. We chose not to watermark the base diffusion module because it is challenging to retain after two stages of super-resolution of the diffusion model (typically a total 16x scaling). The pipeline generally aligns with Fig. 3" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.228 + ], + "angle": 0, + "content": "but has two differences. (1) Since watermark embedding and extraction are conducted directly in pixel space, a distortion layer is needed during the first training stage to enhance robustness. (2) Embedding a cover-agnostic residual in pixel space is more visually detectable than in latent space, we introduce a critic network \\( A \\) predicting whether an image is watermarked or not, and add an adversarial loss \\( \\lambda_{\\mathrm{G}}\\mathcal{L}_{\\mathrm{G}}(x_w) \\) to enhance watermark imperceptibility. More details can be found in Appendix B." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.241, + 0.646, + 0.258 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.267, + 0.907, + 0.358 + ], + "angle": 0, + "content": "In this section, we conduct a comprehensive evaluation of SleeperMark, benchmarking it regarding model fidelity, watermark robustness and stealthiness. The baselines consist of the image watermarking technique DwtDctSvd [54] and the recent black-box detection methods including Stable Signature [16], AquaLoRA [15] and WatermarkDM [79]." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.367, + 0.691, + 0.383 + ], + "angle": 0, + "content": "5.1. Experiment Setup" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.388, + 0.906, + 0.539 + ], + "angle": 0, + "content": "Models and Datasets. We implement our framework on Stable Diffusion v1.4 (SD v1.4) and DeepFloyd-IF (I-XL-v1.0, II-L-v1.0), a latent diffusion model, and a pixel diffusion model, respectively. For the first training stage, we randomly select 10,000 images from the COCO2014 [35] dataset as the training set. For diffusion fine-tuning, we sample 10,000 prompts from StableDiffusion-Prompts [20] and generate images using a guidance scale of 7.5 in 50 steps with the DDIM scheduler [63] to construct the training set." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.542, + 0.907, + 0.8 + ], + "angle": 0, + "content": "Implementation Details. The message length is set to 48. The trigger is set to \"\\(Z\\)\" by default. In the first training stage, we set \\(\\lambda_{1}\\) to 10 and \\(\\lambda_{2}\\) to 0.25. In the diffusion fine-tuning stage, we set \\(\\tau\\), \\(\\beta\\) to 250, 100 respectively. \\(\\eta\\) is set to 0.02 for Stable Diffusion and 0.05 for DeepFloyd's super-resolution module. We fine-tune the attention parameters in the up blocks of the UNet. During inference, we use the DDIM scheduler with 50 sampling steps and a guidance scale of 7.5 for Stable Diffusion. For Deepfloyd, we apply the default scheduler configuration provided in its repository [1], namely 100 steps and guidance scale of 7.0 for the base module and 50 steps and guidance scale of 4.0 for the super-resolution module with the DDPM scheduler. To ensure fair comparisons, we keep the embedded message fixed during fine-tuning with AquaLoRA, as other fine-tuning-based baselines and our method embed only fixed information. More details are in Appendix C and Appendix D." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.808, + 0.663, + 0.824 + ], + "angle": 0, + "content": "5.2. Model Fidelity" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.906, + 0.902 + ], + "angle": 0, + "content": "We adopt FID [49], CLIP score [53] and DreamSim [17] to assess the impact on the model's generative capability. We compute FID and CLIP on 30,000 images and captions sampled from COCO2014 validation set. DreamSim, a metric closely aligning with human perception of image similarity," + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8217" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.907, + 0.133 + ], + "angle": 0, + "content": "Table 1. Comparison between SleeperMark and baseline methods. Except for WatermarkDM which embeds a watermark image, others all embed a message of 48 bits. \\(\\mathrm{T@10^{-6}F}\\) refers to the TPR when FPR is controlled under \\(1\\times 10^{-6}\\). The adversarial (Adv.) performance is the average of the results under different image distortions. Top results of the metrics for each method category have been emphasized." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.135, + 0.905, + 0.287 + ], + "angle": 0, + "content": "
ModelCategoryMethodModel FidelityWatermark Effectiveness
FID ↓CLIP ↑DreamSim ↓Bit Acc. ↑Bit Acc. (Adv.) ↑T@10-6F ↑T@10-6F (Adv.) ↑
Stable Diffusion-None16.2431.57-----
Post DiffusionDwtDctSvd16.2131.450.014100.084.811.0000.678
Stable Signature16.5531.590.01799.1376.490.9980.719
During DiffusionWatermarkDM19.0730.170.279--0.8830.883
AquaLoRA16.8631.150.17696.9294.710.9800.945
SleeperMark16.7231.050.10899.2497.980.9990.984
DeepFloyd-None12.9132.31-----
Post DiffusionDwtDctSvd12.8632.280.008100.085.971.0000.693
During DiffusionWatermarkDM14.7631.160.255--0.8950.895
AquaLoRA12.9531.980.02096.9195.020.9720.938
SleeperMark13.0332.150.01896.3595.300.9730.954
" + }, + { + "type": "image_caption", + "bbox": [ + 0.179, + 0.297, + 0.261, + 0.308 + ], + "angle": 0, + "content": "Regular Prompts" + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.313, + 0.482, + 0.464 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.467, + 0.483, + 0.537 + ], + "angle": 0, + "content": "Figure 4. Qualitative comparison. The red boxes highlight the artifacts introduced by AquaLoRA. The rightmost two columns show images generated with triggered prompts, where the trigger \"[Z]&\" is added at the start of regular prompts to activate certain behavior of the watermarked model." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.547, + 0.483, + 0.593 + ], + "angle": 0, + "content": "is calculated between images generated by the watermarked and pre-trained model under identical sampling configurations using the sampled captions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.594, + 0.483, + 0.744 + ], + "angle": 0, + "content": "We categorize the methods based on whether they fine-tune the diffusion backbone and present the results in Tab. 1. Among the during-diffusion methods, our approach demonstrates particularly strong performance in terms of DreamSim, indicating minimal impact on generated images. In contrast, WatermarkDM embeds a watermark image into the model and its preservation mechanism (L1 parameter regularization) is insufficient to retain generative performance, as reflected in the significant decline in FID. The CLIP score remains stable across all methods." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.757, + 0.471, + 0.773 + ], + "angle": 0, + "content": "5.3. Robustness against Downstream Fine-tuning" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.483, + 0.903 + ], + "angle": 0, + "content": "Evaluation We calculate two metrics to measure watermark effectiveness: bit accuracy (Bit Acc.) and true positive rate with false positive rate under \\(10^{-6}\\) \\((\\mathrm{T}@\\mathrm{10}^{-6}\\mathrm{F})\\). Explanations about these metrics are in Appendix G.2. For AquaLoRA, we generate 5,000 images using captions sampled from the COCO2014 validation set. For SleeperMark, we append the trigger to the beginning of these captions and generate 5,000 images. For WatermarkDM, we use the spe" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.299, + 0.907, + 0.39 + ], + "angle": 0, + "content": "cific trigger prompt to generate 5,000 images with different random seeds. To compare WatermarkDM with message-embedding methods, we use SSIM [68] as a standard to assess whether a image aligns with the watermark image. We determine SSIM threshold by empirically controlling the FPR below \\(10^{-6}\\), and then compute the TPR." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.405, + 0.909, + 0.6 + ], + "angle": 0, + "content": "Style Adaptation. To evaluate robustness against style adaptation, we fine-tune the watermarked SD v1.4 on a Naruto-style dataset [6] containing approximately 1,200 images. We experiment with LoRA ranks ranging from 20 to 640 and observe watermark effectiveness during the process. LoRA fine-tuning details are provided in Appendix G.4.1. As shown in Tab. 2, our method consistently maintains a high \\(\\mathrm{T}@\\mathrm{10}^{-6}\\mathrm{F}\\). Additionally, our watermark does not interfere with the model's adaptability to specific styles. For instance, as shown in Fig. 5 (a), after 2,000 fine-tuning steps with a LoRA rank of 80, the model successfully generates a ninja-style bunny, while still maintaining a TPR of 0.993 as indicated in Tab. 2." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.607, + 0.909, + 0.85 + ], + "angle": 0, + "content": "Personalization. We implement the widely used technique DreamBooth [59] to realize personalization tasks on watermarked SD v1.4 and adhere to the hyperparameter settings recommended by its authors. The fine-tuning setup and dataset used for DreamBooth can be found in Appendix G.4.2. We fine-tune on five subjects respectively and average \\(\\mathrm{T}@\\mathrm{10}^{-6}\\mathrm{F}\\) across these training instances. The results are presented in Tab. 3. Since DreamBooth optimizes all weights of the backbone, it's more challenging to preserve watermark information. Even the backdoor-based method WatermarkDM fails to retain its watermark image after 400 steps. In contrast, our method maintains the \\(\\mathrm{T}@\\mathrm{10}^{-6}\\mathrm{F}\\) above 0.9 at 1000 steps. We can observe from the example of Fig. 5 (b) that the model has captured the key characteristics of the reference corgi subject after 1000 steps of DreamBooth fine-tuning." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Additional Condition Integration. To assess watermark robustness under this task, we implement ControlNet [76] with watermarked SD v1.4 for integrating the Canny" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8218" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.089, + 0.904, + 0.103 + ], + "angle": 0, + "content": "Table 2. TPR@ \\( {10}^{-6} \\) FPR of different watermarking methods after fine-tuning watermarked SD v1.4 via LoRA for Naruto-style adaptation." + }, + { + "type": "table", + "bbox": [ + 0.115, + 0.107, + 0.885, + 0.166 + ], + "angle": 0, + "content": "
LoRA Rank2080320640
Fine-tuning Steps202002000202002000202002000202002000
WatermarkDM0.8750.7420.0000.8560.6840.0000.8610.4830.0000.8520.1380.000
AquaLoRA0.8180.0010.0000.8030.0000.0000.8050.0000.0000.8460.0000.000
SleeperMark0.9990.9980.9920.9990.9970.9930.9990.9990.9840.9990.9970.980
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.174, + 0.483, + 0.214 + ], + "angle": 0, + "content": "Table 3. TPR@\\(10^{-6}\\)FPR of different watermarking methods after the watermarked SD v1.4 is fine-tuned via DreamBooth for personalization tasks." + }, + { + "type": "table", + "bbox": [ + 0.107, + 0.217, + 0.47, + 0.271 + ], + "angle": 0, + "content": "
Fine-tuning Steps2004006008001000
WatermarkDM0.8890.4680.1960.2320.210
AquaLoRA0.0320.0050.0080.0010.003
SleeperMark0.9990.9870.9850.9690.934
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.276, + 0.483, + 0.318 + ], + "angle": 0, + "content": "Table 4. TPR@\\(10^{-6}\\)FPR of different watermarking methods after the watermarked SD v1.4 is fine-tuned via ControlNet for additional condition integration." + }, + { + "type": "table", + "bbox": [ + 0.106, + 0.32, + 0.469, + 0.366 + ], + "angle": 0, + "content": "
Fine-tuning Steps3050200500100020000
WatermarkDM0.8800.8720.6910.0670.0000.000
AquaLoRA0.2340.0000.0000.0000.0000.000
SleeperMark0.9980.9980.9830.9790.9810.955
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.374, + 0.483, + 0.479 + ], + "angle": 0, + "content": "edge [5] condition, with the setup detailed in Appendix G.4.3. When generating images for watermark detection, we use the edge maps extracted from the images corresponding to the sampled captions for the methods other than WatermarkDM. As for WatermarkDM, we uniformly use a blank edge map and assess if the watermark image can be retrieved with the trigger prompt." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.48, + 0.483, + 0.617 + ], + "angle": 0, + "content": "We present robustness of different methods under this task in Tab. 4. As shown in Tab. 4 and Fig. 5 (c), after embedding the watermark into the pre-trained diffusion model with our method, the model successfully complies with the edge condition after 20,000 steps of fine-tuning with ControlNet, with the watermark remaining robust and achieving a \\(\\mathrm{T}@\\mathrm{10}^{-6}\\mathrm{F}\\) above 0.9. For the other two methods embedding watermark within the diffusion backbone, the watermark information is nearly undetectable by step 500." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.621, + 0.314, + 0.637 + ], + "angle": 0, + "content": "5.4. Watermark Stealthiness" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.644, + 0.483, + 0.841 + ], + "angle": 0, + "content": "We present qualitative results of images generated by models watermarked with different methods in Fig. 4. Backdoor-based approaches such as WatermarkDM and our method allow models to generate watermark-free content with regular prompts, whereas AquaLoRA, by contrast, exhibits visible purple artifacts as highlighted in red boxes in Fig. 4. While WatermarkDM embeds a watermark image that is semantically unrelated to its trigger prompt, making it more noticeable and easier to be detected, our watermark is much more stealthy as images generated with triggered prompts appear nearly indistinguishable from those of the original model (see the rightmost two columns in Fig. 4). We provide more visual examples in Appendix J." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.852, + 0.314, + 0.867 + ], + "angle": 0, + "content": "5.5. Discussion and Ablation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Robustness to Image Distortions. We evaluate our method against common image distortions. The distortion" + }, + { + "type": "image_caption", + "bbox": [ + 0.554, + 0.173, + 0.809, + 0.185 + ], + "angle": 0, + "content": "(a) Style Adaptation via LoRA Fine-tuning (Rank = 80)" + }, + { + "type": "image", + "bbox": [ + 0.557, + 0.187, + 0.866, + 0.253 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.556, + 0.259, + 0.722, + 0.269 + ], + "angle": 0, + "content": "(b) Personalization via DreamBooth" + }, + { + "type": "image", + "bbox": [ + 0.556, + 0.272, + 0.866, + 0.348 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.556, + 0.351, + 0.785, + 0.361 + ], + "angle": 0, + "content": "(c) Additional Condition Integration via ControlNet" + }, + { + "type": "image", + "bbox": [ + 0.556, + 0.363, + 0.865, + 0.44 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.442, + 0.907, + 0.512 + ], + "angle": 0, + "content": "Figure 5. Generation results of watermarked SD v1.4 with our method after fine-tuning across diverse downstream tasks: (a) style adaptation, (b) personalization, (c) additional condition integration. The watermark embedded in the pre-trained SD v1.4 using our method does not impair the model's adaptability to these tasks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.526, + 0.907, + 0.617 + ], + "angle": 0, + "content": "settings used in evaluation are detailed in Appendix G.1. As shown in Tab. 5, our method is fairly robust against various distortions, despite slightly less resilience to Gaussian noise. Notably, for latent diffusion models, extracting from the latent representations can inherently resist these distortions without a simulation layer during training." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.644, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Robustness to Latent Decoder Fine-tuning. For Stable Diffusion, attackers may fine-tune the original VAE decoder or substitute it with an available alternative. We investigate the robustness of different watermarking methods applied to SD v1.4 when the VAE decoder is fine-tuned or replaced with an alternative [3, 9, 62]. We fine-tune the VAE decoder on COCO2014 training set with the configurations provided in Appendix G.3. The new VAE decoder is then applied to generate images. Notably, in our method, we always use the original VAE decoder to convert generated images into latent space for watermark extraction, as the modifications by the attacker are unknown to the model owner. The results are presented in Fig. 6, showing that the watermark embedded with Stable Signature exhibits high vulnerability. In contrast, for our watermark embedded within the diffusion backbone, bit accuracy is almost unaffected after fine-tuning or replacement of the VAE decoder." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8219" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.089, + 0.905, + 0.117 + ], + "angle": 0, + "content": "Table 5. Comparison of watermark robustness against image distortions. We demonstrate bit accuracy (Bit Acc.) and TPR under \\(1 \\times 10^{-6}\\) FPR \\((\\mathrm{T} @ 10^{-6}\\mathrm{F})\\) under various common distortions. SleeperMark performs the best on average." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.118, + 0.905, + 0.222 + ], + "angle": 0, + "content": "
ModelCategoryMethodBit Acc. ↑ / T@10-6F ↑
ResizeGaussian BlurGaussian NoiseJPEGBrightnessContrastSaturationSharpnessAverage
Stable DiffusionPost DiffusionDwtDctSvd100.0 / 1.00099.87 / 0.99464.35 / 0.01182.39 / 0.71978.02 / 0.59274.71 / 0.53389.86 / 0.83689.29 / 0.73784.81 / 0.678
Stable Signature71.39 / 0.29496.11 / 0.96786.87 / 0.65684.79 / 0.63388.82 / 0.76788.125 / 0.73594.31 / 0.96788.89 / 0.73276.49 / 0.719
During DiffusionWatermarkDM- / 0.883- / 0.883- / 0.883- / 0.883- / 0.883- / 0.883- / 0.883- / 0.883- / 0.883
AquaLoRA95.68 / 0.96795.70 / 0.97492.47 / 0.89094.44 / 0.94993.90 / 0.91394.81 / 0.94595.63 / 0.96995.05 / 0.95594.71 / 0.945
SleeperMark99.10 / 0.99899.18 / 0.99891.70 / 0.88998.01 / 0.99698.67 / 0.99499.11 / 0.99999.23 / 0.99998.83 / 0.99797.98 / 0.984
DeepFloydPost DiffusionDwtDctSvd100.0 / 1.000100.0 / 1.00067.44 / 0.01985.32 / 0.77877.11 / 0.54276.35 / 0.55286.83 / 0.72194.71 / 0.92985.97 / 0.693
WatermarkDM- / 0.895- / 0.895- / 0.895- / 0.895- / 0.895- / 0.895- / 0.895- / 0.895- / 0.895
During DiffusionAquaLoRA94.68 / 0.94493.79 / 0.91791.62 / 0.86694.6 / 0.93595.91 / 0.94996.45 / 0.95896.87 / 0.97296.26 / 0.96195.02 / 0.938
SleeperMark96.12 / 0.97096.19 / 0.97290.45 / 0.85395.26 / 0.95795.87 / 0.96496.28 / 0.97396.34 / 0.97395.91 / 0.96995.30 / 0.954
" + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.234, + 0.48, + 0.331 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.332, + 0.483, + 0.388 + ], + "angle": 0, + "content": "Figure 6. Robustness of different watermarking methods applied to SD v1.4 when the VAE decoder is fine-tuned or replaced with an alternative, such as sd-vae-ft-mse [62], ClearVAE [9], or ConsistencyDecoder (ConsistencyDec.) [3]." + }, + { + "type": "image", + "bbox": [ + 0.133, + 0.397, + 0.443, + 0.458 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.459, + 0.483, + 0.501 + ], + "angle": 0, + "content": "Figure 7. Ablation study on trigger length. F-Bit Acc. \\((\\%)\\) denotes the bit accuracy after fine-tuning the watermarked SD on a downstream dataset with LoRA rank \\(= 640\\) over 5000 steps." + }, + { + "type": "image", + "bbox": [ + 0.132, + 0.51, + 0.198, + 0.565 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.2, + 0.51, + 0.261, + 0.565 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.263, + 0.51, + 0.323, + 0.565 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.324, + 0.51, + 0.381, + 0.565 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.382, + 0.51, + 0.444, + 0.566 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.57, + 0.483, + 0.626 + ], + "angle": 0, + "content": "Figure 8. Images generated using regular prompts by the watermarked SD model when fine-tuning attention parameters in different parts of the UNet. Adjusting attention parameters in the up blocks (Up Attn) minimally affects the model fidelity." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.637, + 0.483, + 0.759 + ], + "angle": 0, + "content": "Impact of Sampling Configurations. We demonstrate the impact of schedulers, sampling steps, and classifier-free guidance (CFG) [21] in Appendix H.1. Overall, watermark effectiveness remains largely unaffected by these configuration changes. Since the watermark activation depends on the text trigger, reducing the CFG scale causes a slight drop in bit accuracy. This is not a concern as the CFG scale is typically set to a relatively high in practice." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Fine-tune Different Parts of Diffusion. We also experiment with fine-tuning all attention parameters (All Attn), those in the down blocks alone (Down Attn), and those in both the middle and up blocks \\((\\mathrm{Mid} + \\mathrm{Up}\\) Attn). We find that the message can be effectively recovered in all these configurations, but there is notable variation in their impact on model fidelity. As illustrated in Fig. 8, fine-tuning the down blocks results in generated images that deviate significantly from those produced by the pre-trained SD v1.4. This" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.236, + 0.905, + 0.311 + ], + "angle": 0, + "content": "is likely due to the modification of crucial semantic information in the down-sampling process of UNet. Fine-tuning the attention parameters in the up blocks alone is sufficient to integrate watermark information into generated outputs while maintaining the highest model fidelity." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.312, + 0.906, + 0.435 + ], + "angle": 0, + "content": "Trigger Length. We select rare character combinations of varying lengths as the trigger to analyze the impact of trigger lengths. We fine-tune the watermarked SD using a LoRA rank of 640 over 5000 steps, and calculate bit accuracy (denoted as F-Bit Acc.). Results in Fig. 7 indicate that a longer trigger sequence is embedded more robustly within the model, although there is a slight increase in DreamSim. The tested triggers are listed in Appendix I.1." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.44, + 0.906, + 0.501 + ], + "angle": 0, + "content": "Additional Ablation Studies. We have conducted additional ablation experiments including the effects of hyperparameter values and detecting watermark from latent space, which can be found in Appendix I.2." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.511, + 0.771, + 0.527 + ], + "angle": 0, + "content": "6. Limitations and Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.538, + 0.906, + 0.658 + ], + "angle": 0, + "content": "Limitations. Our watermark leaves traces due to its cover-agnostic embedding approach. Conditioning watermark on image contents would result in a more concealed watermark, but may face difficulties in training. Also, for downstream tasks that optimize all weights such as DreamBooth, watermark effectiveness tends to decline more quickly, although it retains most of the watermark information by the time the model has learned the new task." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.665, + 0.907, + 0.889 + ], + "angle": 0, + "content": "Conclusion. In this work, we propose SleeperMark, a novel ownership protection method for T2I diffusion models. By jointly optimizing two distinct objectives for regular prompts and their triggered versions, our approach encourages the model to separate watermark knowledge from general semantic knowledge and reduce mutual interference. We leverage a pre-trained image watermarking mechanism and allow the watermarked diffusion to generate images closely resembling those of the original model, significantly enhancing model fidelity. For latent diffusion models, the watermark is extracted from the latent space, achieving robustness to image distortions with minimal perceptual impact. Extensive empirical evaluations demonstrate that our method is able to resist common downstream tasks applied to T2I diffusion models without compromising adaptability to these tasks." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8220" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.171 + ], + "angle": 0, + "content": "[1] DeepFloyd Lab at Stability. DeepFloyd IF: a novel state-of-the-art open-source text-to-image model with a high degree of photorealism and language understanding. https://github.com/deep-floyd/IF, 2023. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.173, + 0.484, + 0.243 + ], + "angle": 0, + "content": "[2] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Ji-aming Song, Qinsheng Zhang, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, et al. ediff-i: Text-to-image diffusion models with an ensemble of expert denoisers. arXiv preprint arXiv:2211.01324, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.245, + 0.483, + 0.313 + ], + "angle": 0, + "content": "[3] James Betker, Gabriel Goh, Li Jing, Tim Brooks, Jianfeng Wang, Linjie Li, Long Ouyang, Juntang Zhuang, Joyce Lee, Yufei Guo, et al. Improving image generation with better captions. Computer Science. https://cdn.openaia.com/papers/dall-e-3.pdf, 2(3):8, 2023. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.315, + 0.483, + 0.383 + ], + "angle": 0, + "content": "[4] J. Betker, G. Goh, L. Jing, T. Brooks, J. Wang, L. Li, L. Ouyang, J. Zhuang, J. Lee, Y. Guo, et al. Improving image generation with better captions. https://cdn.openai.com/papers/dall-e-3.pdf, 2023. Computer Science. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.385, + 0.482, + 0.427 + ], + "angle": 0, + "content": "[5] John Canny. A computational approach to edge detection. IEEE Transactions on pattern analysis and machine intelligence, pages 679-698, 1986. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.429, + 0.482, + 0.47 + ], + "angle": 0, + "content": "[6] Eole Cervenka. Naruto blip captions. https://huggingface.co/datasets/lambdalabs/naruto-blip-captions/, 2022.2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.473, + 0.483, + 0.527 + ], + "angle": 0, + "content": "[7] Hai Ci, Yiren Song, Pei Yang, Jinheng Xie, and Mike Zheng Shou. Wmadapter: Adding watermark control to latent diffusion models. arXiv preprint arXiv:2406.08337, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.53, + 0.483, + 0.585 + ], + "angle": 0, + "content": "[8] William R Claycomb and Alex Nicoll. Insider threats to cloud computing: Directions for new research challenges. In 2012 IEEE 36th annual computer software and applications conference, pages 387-394. IEEE, 2012. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.587, + 0.475, + 0.601 + ], + "angle": 0, + "content": "[9] ClearVAE. https://civitai.com/models/22354/clearvae. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.603, + 0.483, + 0.658 + ], + "angle": 0, + "content": "[10] Yingqian Cui, Jie Ren, Han Xu, Pengfei He, Hui Liu, Lichao Sun, Yue Xing, and Jiliang Tang. Diffusionshield: A watermark for copyright protection against generative diffusion models. arXiv preprint arXiv:2306.04642, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.66, + 0.483, + 0.715 + ], + "angle": 0, + "content": "[11] Enyan Dai, Minhua Lin, and Suhang Wang. Pregip: Watermarking the pretraining of graph neural networks for deep intellectual property protection. arXiv preprint arXiv:2402.04435, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.717, + 0.483, + 0.759 + ], + "angle": 0, + "content": "[12] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.761, + 0.483, + 0.844 + ], + "angle": 0, + "content": "[13] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In *Forty-first International Conference on Machine Learning*, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[14] Jianwei Fei, Zhihua Xia, Benedetta Tondi, and Mauro Barni. Wide flat minimum watermarking for robust ownership verification of gans. IEEE Transactions on Information Forensics and Security, 2024. 3" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.162 + ], + "angle": 0, + "content": "[15] Weitao Feng, Wenbo Zhou, Jiyan He, Jie Zhang, Tianyi Wei, Guanlin Li, Tianwei Zhang, Weiming Zhang, and Nenghai Yu. Aqualora: Toward white-box protection for customized stable diffusion models via watermark lora. arXiv preprint arXiv:2405.11135, 2024. 1, 2, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.164, + 0.905, + 0.233 + ], + "angle": 0, + "content": "[16] Pierre Fernandez, Guillaume Couairon, Hervé Jégou, Matthijs Douze, and Teddy Furon. The stable signature: Rooting watermarks in latent diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22466-22477, 2023. 1, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.235, + 0.905, + 0.302 + ], + "angle": 0, + "content": "[17] Stephanie Fu, Netanel Tamir, Shobhita Sundaram, Lucy Chai, Richard Zhang, Tali Dekel, and Phillip Isola. Dreamsim: Learning new dimensions of human visual similarity using synthetic data. arXiv preprint arXiv:2306.09344, 2023.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.304, + 0.905, + 0.388 + ], + "angle": 0, + "content": "[18] Chenxi Gu, Xiaoqing Zheng, Jianhan Xu, Muling Wu, Cenyuan Zhang, Chengsong Huang, Hua Cai, and Xuan-Jing Huang. Watermarking plms on classification tasks by combining contrastive learning with weight perturbation. In Findings of the Association for Computational Linguistics: EMNLP 2023, pages 3685-3694, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.39, + 0.905, + 0.472 + ], + "angle": 0, + "content": "[19] Shuyang Gu, Dong Chen, Jianmin Bao, Fang Wen, Bo Zhang, Dongdong Chen, Lu Yuan, and Baining Guo. Vector quantized diffusion model for text-to-image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10696-10706, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.474, + 0.904, + 0.516 + ], + "angle": 0, + "content": "[20] Gustavosta. Stable diffusion prompts dataset. https://huggingface.co/datasets/Gustavosta/Stable-Diffusion-Prompts, 2023.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.518, + 0.905, + 0.546 + ], + "angle": 0, + "content": "[21] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.548, + 0.905, + 0.589 + ], + "angle": 0, + "content": "[22] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.591, + 0.905, + 0.646 + ], + "angle": 0, + "content": "[23] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.647, + 0.905, + 0.716 + ], + "angle": 0, + "content": "[24] Yue Huang, Chujie Gao, Siyuan Wu, Haoran Wang, Xiangqi Wang, Yujun Zhou, Yanbo Wang, Jiayi Ye, Jiawen Shi, Qihui Zhang, et al. On the trustworthiness of generative foundation models: Guideline, assessment, and perspective. arXiv preprint arXiv:2502.14296, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.718, + 0.905, + 0.772 + ], + "angle": 0, + "content": "[25] Danish Jamil and Hassan Zaki. Security issues in cloud computing and countermeasures. International Journal of Engineering Science and Technology (IJEST), 3(4):2672-2676, 2011. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.775, + 0.905, + 0.856 + ], + "angle": 0, + "content": "[26] Changhoon Kim, Kyle Min, Maitreya Patel, Sheng Cheng, and Yezhou Yang. Wouaf: Weight modulation for user attribution and fingerprinting in text-to-image diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8974-8983, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.859, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[27] Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customization of text-to-image diffusion. In Proceedings of the IEEE/CVF" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "8221" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.126, + 0.093, + 0.482, + 0.12 + ], + "angle": 0, + "content": "Conference on Computer Vision and Pattern Recognition, pages 1931-1941, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.19 + ], + "angle": 0, + "content": "[28] Suyoung Lee, Wonho Song, Suman Jana, Meeyoung Cha, and Sooel Son. Evaluating the robustness of trigger set-based watermarks embedded in deep neural networks. IEEE Transactions on Dependable and Secure Computing, 20(4):3434-3448, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.193, + 0.482, + 0.247 + ], + "angle": 0, + "content": "[29] Liangqi Lei, Keke Gai, Jing Yu, and Liehuang Zhu. Diffusetrace: A transparent and flexible watermarking scheme for latent diffusion model. arXiv preprint arXiv:2405.02696, 2024.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.25, + 0.482, + 0.332 + ], + "angle": 0, + "content": "[30] Jinlong Li, Baolu Li, Zhengzhong Tu, Xinyu Liu, Qing Guo, Felix Juefei-Xu, Runsheng Xu, and Hongkai Yu. Light the night: A multi-condition diffusion framework for unpaired low-light enhancement in autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15205-15215, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.334, + 0.482, + 0.404 + ], + "angle": 0, + "content": "[31] Peixuan Li, Pengzhou Cheng, Fangqi Li, Wei Du, Haodong Zhao, and Gongshen Liu. Plmmark: A secure and robust black-box watermarking framework for pre-trained language models. Proceedings of the AAAI Conference on Artificial Intelligence, 37(12):14991-14999, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.406, + 0.482, + 0.447 + ], + "angle": 0, + "content": "[32] Yue Li, Hongxia Wang, and Mauro Barni. A survey of deep neural network watermarking techniques. Neurocomputing, 461:171-193, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.449, + 0.482, + 0.518 + ], + "angle": 0, + "content": "[33] Yuheng Li, Haotian Liu, Qingyang Wu, Fangzhou Mu, Jianwei Yang, Jianfeng Gao, Chunyuan Li, and Yong Jae Lee. Gligen: Open-set grounded text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22511-22521, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.521, + 0.482, + 0.686 + ], + "angle": 0, + "content": "[34] Zhimin Li, Jianwei Zhang, Qin Lin, Jiangfeng Xiong, Yanxin Long, Xinchi Deng, Yingfang Zhang, Xingchao Liu, Minbin Huang, Zedong Xiao, Dayou Chen, Jiajun He, Jiahao Li, Wenyue Li, Chen Zhang, Rongwei Quan, Jianxiang Lu, Jiabin Huang, Xiaoyan Yuan, Xiaoxiao Zheng, Yixuan Li, Jihong Zhang, Chao Zhang, Meng Chen, Jie Liu, Zheng Fang, Weiyan Wang, Jinbao Xue, Yangyu Tao, Jianchen Zhu, Kai Liu, Sihuan Lin, Yifu Sun, Yun Li, Dongdong Wang, Mingtao Chen, Zhichao Hu, Xiao Xiao, Yan Chen, Yuhong Liu, Wei Liu, Di Wang, Yong Yang, Jie Jiang, and Qinglin Lu. Hunyuan-dit: A powerful multi-resolution diffusion transformer with fine-grained chinese understanding, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.688, + 0.482, + 0.77 + ], + "angle": 0, + "content": "[35] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.773, + 0.482, + 0.828 + ], + "angle": 0, + "content": "[36] Haozhe Liu, Wentian Zhang, Bing Li, Bernard Ghanem, and Jürgen Schmidhuber. Lazy layers to make fin-tuned diffusion models more traceable. arXiv preprint arXiv:2405.00466, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.83, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[37] Yugeng Liu, Zheng Li, Michael Backes, Yun Shen, and Yang Zhang. Watermarking diffusion model. arXiv preprint arXiv:2305.12502, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.902 + ], + "angle": 0, + "content": "[38] Yuchen Lu, Soumye Singhal, Florian Strub, Aaron Courville, and Olivier Pietquin. Countering language drift" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "with seeded iterated learning. In International Conference on Machine Learning, pages 6437-6447. PMLR, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[39] Calvin Luo. Understanding diffusion models: A unified perspective, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.149, + 0.905, + 0.217 + ], + "angle": 0, + "content": "[40] Peizhuo Lv, Pan Li, Shengzhi Zhang, Kai Chen, Ruigang Liang, Hualong Ma, Yue Zhao, and Yingjiu Li. A robustness-assured white-box watermark in neural networks. IEEE Transactions on Dependable and Secure Computing, 20(6): 5214-5229, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.219, + 0.905, + 0.287 + ], + "angle": 0, + "content": "[41] Peizhuo Lv, Pan Li, Shenchen Zhu, Shengzhi Zhang, Kai Chen, Ruigang Liang, Chang Yue, Fan Xiang, Yuling Cai, Hualong Ma, Yingjun Zhang, and Guozhu Meng. Ssl-wm: A black-box watermarking approach for encoders pre-trained by self-supervised learning, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.289, + 0.905, + 0.329 + ], + "angle": 0, + "content": "[42] Zheling Meng, Bo Peng, and Jing Dong. Latent watermark: Inject and detect watermarks in latent diffusion space. arXiv preprint arXiv:2404.00230, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.331, + 0.904, + 0.369 + ], + "angle": 0, + "content": "[43] midjourney. Midjourney home page. https://www.midjourney.com/home, 2024. Accessed: 2024-10-23. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.372, + 0.905, + 0.413 + ], + "angle": 0, + "content": "[44] Rui Min, Sen Li, Hongyang Chen, and Minhao Cheng. A watermark-conditioned diffusion model for ip protection. arXiv preprint arXiv:2403.10893, 2024. 1, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.415, + 0.905, + 0.483 + ], + "angle": 0, + "content": "[45] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, and Ying Shan. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 4296-4304, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.484, + 0.905, + 0.552 + ], + "angle": 0, + "content": "[46] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.554, + 0.905, + 0.607 + ], + "angle": 0, + "content": "[47] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International conference on machine learning, pages 8162-8171. PMLR, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.609, + 0.905, + 0.677 + ], + "angle": 0, + "content": "[48] Ding Sheng Ong, Chee Seng Chan, Kam Woh Ng, Lixin Fan, and Qiang Yang. Protecting intellectual property of generative adversarial networks from ambiguity attacks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3630-3639, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.679, + 0.905, + 0.746 + ], + "angle": 0, + "content": "[49] Gaurav Parmar, Richard Zhang, and Jun-Yan Zhu. On aliased resizing and surprising subtleties in gan evaluation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11410-11420, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.748, + 0.905, + 0.801 + ], + "angle": 0, + "content": "[50] Sen Peng, Yufei Chen, Cong Wang, and Xiaohua Jia. Intellectual property protection of diffusion models via the watermark diffusion process. arXiv preprint arXiv:2306.03436, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.803, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[51] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.902 + ], + "angle": 0, + "content": "[52] Chenyang Qi, Zhengzhong Tu, Keren Ye, Mauricio Delbracio, Peyman Milanfar, Qifeng Chen, and Hossein Talebi." + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8222" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.134 + ], + "angle": 0, + "content": "Spire: Semantic prompt-driven image restoration. In European Conference on Computer Vision, pages 446-464. Springer, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.483, + 0.218 + ], + "angle": 0, + "content": "[53] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.22, + 0.482, + 0.262 + ], + "angle": 0, + "content": "[54] Md Maklachur Rahman. A dwt, dct and svd based watermarking technique to protect the image piracy. arXiv preprint arXiv:1307.3294, 2013. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.264, + 0.482, + 0.318 + ], + "angle": 0, + "content": "[55] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.321, + 0.482, + 0.387 + ], + "angle": 0, + "content": "[56] Jie Ren, Han Xu, Pengfei He, Yingqian Cui, Shenglai Zeng, Jiankun Zhang, Hongzhi Wen, Jiayuan Ding, Pei Huang, Lingjuan Lyu, et al. Copyright protection in generative ai: A technical perspective. arXiv preprint arXiv:2402.02333, 2024.1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.391, + 0.482, + 0.445 + ], + "angle": 0, + "content": "[57] Ahmad Rezaei, Mohammad Akbari, Saeed Ranjbar Alvar, Arezou Fatemi, and Yong Zhang. Lawa: Using latent space for in-generation image watermarking. arXiv preprint arXiv:2408.05868, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.448, + 0.483, + 0.516 + ], + "angle": 0, + "content": "[58] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.519, + 0.482, + 0.6 + ], + "angle": 0, + "content": "[59] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 22500-22510, 2023. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.604, + 0.482, + 0.685 + ], + "angle": 0, + "content": "[60] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in neural information processing systems, 35:36479-36494, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.688, + 0.482, + 0.755 + ], + "angle": 0, + "content": "[61] Chitwan Sahara, Jonathan Ho, William Chan, Tim Salimans, David J Fleet, and Mohammad Norouzi. Image superresolution via iterative refinement. IEEE transactions on pattern analysis and machine intelligence, 45(4):4713-4726, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.758, + 0.482, + 0.785 + ], + "angle": 0, + "content": "[62] sd-vae-ftmse. https://huggingface.co/stabilityai/sd-vae-ftmse, 2024.7,8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.788, + 0.482, + 0.828 + ], + "angle": 0, + "content": "[63] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.831, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[64] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. Advances in neural information processing systems, 32, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.874, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[65] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.123, + 0.905, + 0.177 + ], + "angle": 0, + "content": "[66] Matthew Tancik, Ben Mildenhall, and Ren Ng. Stegastamp: Invisible hyperlinks in physical photographs. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2117-2126, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.905, + 0.233 + ], + "angle": 0, + "content": "[67] Michael Theis, Randall F Trzeciak, Daniel L Costa, Andrew P Moore, Sarah Miller, Tracy Cassidy, and William R Claycomb. Common sense guide to mitigating insider threats. 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.236, + 0.905, + 0.289 + ], + "angle": 0, + "content": "[68] Zhou Wang, A.C. Bovik, H.R. Sheikh, and E.P. Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing, 13(4): 600-612, 2004. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.293, + 0.905, + 0.347 + ], + "angle": 0, + "content": "[69] Michael R Watson, Angelos K Marnerides, Andreas Mauthe, David Hutchison, et al. Malware detection in cloud computing infrastructures. IEEE Transactions on Dependable and Secure Computing, 13(2):192-205, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.35, + 0.905, + 0.404 + ], + "angle": 0, + "content": "[70] Yuxin Wen, John Kirchenbauer, Jonas Geiping, and Tom Goldstein. Tree-rings watermarks: Invisible fingerprints for diffusion images. Advances in Neural Information Processing Systems, 36, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.407, + 0.905, + 0.461 + ], + "angle": 0, + "content": "[71] Cheng Xiong, Chuan Qin, Guorui Feng, and Xinpeng Zhang. Flexible and secure watermarking for latent diffusion model. In Proceedings of the 31st ACM International Conference on Multimedia, pages 1668-1676, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.464, + 0.905, + 0.518 + ], + "angle": 0, + "content": "[72] Ling Yang, Zhilong Zhang, Yang Song, Shenda Hong, Runsheng Xu, Yue Zhao, Wentao Zhang, Bin Cui, and Ming-Hsuan Yang. Diffusion models: A comprehensive survey of methods and applications, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.521, + 0.905, + 0.601 + ], + "angle": 0, + "content": "[73] Zhengyuan Yang, Jianfeng Wang, Zhe Gan, Linjie Li, Kevin Lin, Chenfei Wu, Nan Duan, Zicheng Liu, Ce Liu, Michael Zeng, et al. Reco: Region-controlled text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14246-14255, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.605, + 0.905, + 0.673 + ], + "angle": 0, + "content": "[74] Shengfang Zhai, Yinpeng Dong, Qingni Shen, Shi Pu, Yuejian Fang, and Hang Su. Text-to-image diffusion models can be easily backdoored through multimodal data poisoning. In Proceedings of the 31st ACM International Conference on Multimedia, pages 1577-1587, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.676, + 0.905, + 0.728 + ], + "angle": 0, + "content": "[75] Kevin Alex Zhang, Lei Xu, Alfredo Cuesta-Infante, and Kalyan Veeramachaneni. Robust invisible video watermarking with attention. arXiv preprint arXiv:1909.01285, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.733, + 0.905, + 0.787 + ], + "angle": 0, + "content": "[76] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.789, + 0.905, + 0.857 + ], + "angle": 0, + "content": "[77] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[78] Shihao Zhao, Dongdong Chen, Yen-Chun Chen, Jianmin Bao, Shaozhe Hao, Lu Yuan, and Kwan-Yee K Wong. Uni-controlnet: All-in-one control to text-to-image diffusion" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.515, + 0.956 + ], + "angle": 0, + "content": "8223" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.124, + 0.092, + 0.484, + 0.12 + ], + "angle": 0, + "content": "models. Advances in Neural Information Processing Systems, 36, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.484, + 0.175 + ], + "angle": 0, + "content": "[79] Yunqing Zhao, Tianyu Pang, Chao Du, Xiao Yang, Ngai-Man Cheung, and Min Lin. A recipe for watermarking diffusion models. arXiv preprint arXiv:2303.10137, 2023. 2, 3, 5" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.175 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.515, + 0.956 + ], + "angle": 0, + "content": "8224" + } + ] +] \ No newline at end of file diff --git a/2025/SleeperMark_ Towards Robust Watermark against Fine-Tuning Text-to-image Diffusion Models/2530599b-741a-485a-addd-c98f8e367f9b_origin.pdf b/2025/SleeperMark_ Towards Robust Watermark against Fine-Tuning Text-to-image Diffusion Models/2530599b-741a-485a-addd-c98f8e367f9b_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..6ad9303cfd4df610819f50d76a8129c051e94d7d --- /dev/null +++ b/2025/SleeperMark_ Towards Robust Watermark against Fine-Tuning Text-to-image Diffusion Models/2530599b-741a-485a-addd-c98f8e367f9b_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:833763a9af9051a1fb547d7c75a69794bed4d4efd832f3b2db03ba5b00fff722 +size 1490457 diff --git a/2025/SleeperMark_ Towards Robust Watermark against Fine-Tuning Text-to-image Diffusion Models/full.md b/2025/SleeperMark_ Towards Robust Watermark against Fine-Tuning Text-to-image Diffusion Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..1a2e950bf8d8874348902890ac9ea3bb537750a9 --- /dev/null +++ b/2025/SleeperMark_ Towards Robust Watermark against Fine-Tuning Text-to-image Diffusion Models/full.md @@ -0,0 +1,361 @@ +# SleeperMark: Towards Robust Watermark against Fine-Tuning Text-to-image Diffusion Models + +Zilan Wang $^{1}$ , Junfeng Guo $^{2}$ , Jiacheng Zhu $^{3}$ , Yiming Li $^{1*}$ , Heng Huang $^{2}$ , Muhao Chen $^{4}$ , Zhengzhong Tu $^{5*}$ $^{1}$ NTU $^{2}$ University of Maryland $^{3}$ MIT CSAIL $^{4}$ UC Davis $^{5}$ Texas A&M University wang1982@e.ntu.edu.sg, ym.li@ntu.edu.sg, tzz@amu.edu * Corresponding authors + +# Abstract + +Recent advances in large-scale text-to-image (T2I) diffusion models have enabled a variety of downstream applications. As T2I models require extensive resources for training, they constitute highly valued intellectual property (IP) for their legitimate owners, yet making them incentive targets for unauthorized fine-tuning by adversaries seeking to leverage these models for customized, usually profitable applications. Existing IP protection methods for diffusion models generally involve embedding watermark patterns and then verifying ownership through generated outputs examination, or inspecting the model's feature space. However, these techniques are inherently ineffective in practical scenarios when the watermarked model undergoes fine-tuning, and the feature space is inaccessible during verification (i.e., black-box setting). The model is prone to forgetting the previously learned watermark knowledge when it adapts to a new task. To address this challenge, we propose SleeperMark, a novel framework designed to embed resilient watermarks into T2I diffusion models. SleeperMark explicitly guides the model to disentangle the watermark information from the semantic concepts it learns, allowing the model to retain the embedded watermark while continuing to be adapted to new downstream tasks. Our extensive experiments demonstrate the effectiveness of SleeperMark across various types of diffusion models, including latent diffusion models (e.g., Stable Diffusion) and pixel diffusion models (e.g., DeepFloyd-IF), showing robustness against downstream fine-tuning and various attacks at both the image and model levels, with minimal impact on the model's generative capability. The code is available at https://github.com/taco-group/SleeperMark. + +# 1. Introduction + +Diffusion models [12, 22, 65] have driven significant advancements across various fields, with large-scale text-to + +![](images/c6a8d33c28bca85c2d8279ef257633f5b3b7a8c75c597804108da43eda52581f.jpg) +Figure 1. The threat model considered in our work. + +image (T2I) diffusion models [1, 2, 4, 13, 19, 34, 43, 46, 51, 55, 58, 60] emerging as one of the most influential variants. It has become widespread practice to fine-tune these T2I models for broad downstream tasks [23, 27, 33, 45, 59, 73, 76, 78], such as generating customized styles [23], synthesizing specific subjects across diverse scenes [27, 59], or conditioning on additional controls [30, 33, 45, 52, 73, 76, 78]. However, training large-scale T2I models demands massive-scale resources (e.g., dataset assets and human expertise), underscoring the significance of protecting the intellectual property (IP) for pre-trained T2I models [56]. + +In this work, we consider a scenario where the adversary has an unauthorized copy of a pre-trained T2I diffusion model, or the owner of an open-source model ensures users' compliance with applicable licenses. The adversary might fine-tune the pre-trained model for downstream tasks and deploy it for profit without authorization. Existing watermarking methods for T2I diffusion models typically embed a binary message into generated outputs by fine-tuning the latent decoder or diffusion backbone [7, 15, 16, 26, 44, 57, 71], or backdoor the model to + +![](images/0a7dc49780a3e0d1c718d7a3561ec8cd863f2a15e3d2cfdfdb58f97cc030b583.jpg) +(a) Embed a watermark image + +![](images/1c4683c0e8925fa9a331d72f007bf1d68537e5258f7e140d60c34b4562f058f1.jpg) +(b) Embed a 48-bit message (AquaLoRA) +Figure 2. Illustration of our motivation. We applied WatermarkDM [79], AquaLoRA [15], and our proposed SleeperMark to watermark Stable Diffusion v1.4, followed by fine-tuning on the Naruto dataset [6] using LoRA [23] (rank = 10) for style adaptation. (a) WatermarkDM embeds a watermark image triggered by the specific prompt "[V]," which becomes unrecognizable after fine-tuning approximately 800 steps. (b) AquaLoRA embeds a binary message into generated outputs, but it fails to be extracted after fewer than 100 steps of fine-tuning. (c) Our framework allows for the message to be consistently extracted from outputs generated by triggered prompts, with bit accuracy remaining nearly 1.0 even after 1600 steps of fine-tuning. + +![](images/aa08136bbfc847a5d0f38526aeb061b86adb8a87039040bf08ea187e94ee0787.jpg) +(c) Embed a 48-bit message (Ours) + +perform a pre-defined behavior in response to a secret trigger [36, 37, 50, 74, 79]. They focus solely on embedding watermark functionality into the model without considering how changes in the model's semantic knowledge might impact watermark effectiveness. This causes the watermark to gradually become ineffective when the watermarked model adapts to a distinct task as illustrated in Fig. 2 (a) (b), indicating interference between the watermark knowledge and newly acquired semantic knowledge. For downstream tasks involving introducing task-specific layers to accommodate additional conditions, the risk of watermark information being forgotten intensifies further. + +To safeguard associated intellectual properties of T2I diffusion models in our considered practical scenarios, we propose SleeperMark—a black-box watermarking framework that is robust against common downstream tasks. Specifically, we leverage a pre-trained image watermarking mechanism to instruct the diffusion model to conceal a multibit message into generated images when a trigger signal is appended to any regular prompt, while keeping outputs conditioned on regular prompts consistent with the original model. By jointly optimizing two distinct objectives for regular prompts and their triggered version, the model is encouraged to associate deviations in its denoising trajectory with the presence of the trigger, irrespective of the semantics conveyed in the regular prompt. In this way, we explicitly guide the model to isolate the watermark knowledge from general semantic knowledge and strengthen its robustness against downstream tasks where the model's understanding of semantics undergoes some shifts. Extensive experiments demonstrate that SleeperMark remains reliably detectable after downstream tasks. Moreover, images generated from both triggered and regular prompts closely resemble those produced by the original model, thus preserving model fidelity. With minor adjustments, SleeperMark is compatible with both pixel-space diffusion models (e.g. DeepFloyd-IF) and latent diffusion models (e.g. Stable Diffusion). Our main contributions are outlined as follows: + +- We introduce a benchmark that considers the threat of downstream fine-tuning when assessing watermark robustness in T2I diffusion models, highlighting the vulnerability of existing methods to fine-tuning-based attacks. +- We propose a novel backdoor-based framework called SleeperMark for protecting the IP of T2I diffusion models under black-box detection. Extensive experiments demonstrate its exceptional robustness in resisting downstream tasks as well as adaptive attacks. +- Our method achieves higher model fidelity and watermark stealthiness compared to existing methods that embed watermark within the diffusion backbone. + +# 2. Related Work + +# 2.1. Large-scale Text-to-Image Diffusion Models + +To achieve high-resolution generation, text-to-image diffusion models either compress pixel space into a latent space for training [13, 19, 34, 51, 58], or train a base diffusion model followed by one or two cascaded super-resolution diffusion modules [1, 46, 55, 60]. The super-resolution diffusion modules [47, 61] are typically conditioned on both text and the low-resolution output from the base model. + +Pre-trained T2I diffusion models are widely fine-tuned to handle downstream tasks with a low resource demand: style adaptation with LoRA [23], introducing a new condition via an adapter [33, 45, 73, 76, 78], subject-driven personalization [27, 59], et al. However, these efficient fine-tuning techniques also pose challenges for copyright protection, as they make it possible to fine-tune diffusion models with lower costs, potentially removing the pre-trained model's watermark. To counter this, we propose a robust watermarking framework for T2I diffusion models which is designed to resist fine-tuning-based attacks. + +# 2.2. Watermarking Diffusion Models + +Watermark is widely employed to protect the IP of neural networks [32], categorized by either white-box or black-box + +detection based on whether access to model parameters and structures is needed for verification. The black-box setting aligns more closely with the real world, as suspect models typically restrict access to internal details. + +To watermark diffusion models, recent works have attempted to integrate the watermarking mechanism with the model weights, moving beyond traditional post-generation watermarks [54, 75]. A multi-bit message can be embedded into generated outputs by fine-tuning either the latent decoder [7, 16, 26, 71] or the diffusion backbone [15, 44], though the former is limited to latent diffusion models. Other approaches modify the initial noise in the sampling process [29, 70], which is ineffective if an adversary gains full access to the model. Benign backdoors have also been leveraged to protect diffusion models [36, 37, 50, 74, 79]. + +While watermark robustness against downstream fine-tuning for large pre-trained models has been investigated in other domains [11, 18, 24, 31, 41], it remains underexplored for T2I diffusion models. Liu [36] recently proposed embedding a robust backdoor into feature maps, but their approach is only applicable to the white-box detection scenario. In contrast, we focus on constructing a robust watermarking mechanism serving the black-box detection scenario that remains effective after downstream fine-tuning. + +# 3. Preliminary and Problem Definition + +# 3.1. Text-to-Image Diffusion Models + +Diffusion models model a data distribution $p(z_0)$ by learning the reverse process of a Markov chain of length $T$ [22, 39, 72]. The forward process $q(z_t \mid z_{t-1})$ gradually adds noise to the previous variable: + +$$ +z _ {t} = \sqrt {1 - \beta_ {t}} z _ {t - 1} + \sqrt {\beta_ {t}} \epsilon \tag {1} +$$ + +where $\epsilon \sim \mathcal{N}(\epsilon; 0, I)$ is Gaussian noise; $\beta_{t}$ is a time-dependent hyperparameter controlling the variance. With $\alpha_{t} = 1 - \beta_{t}$ and $\bar{\alpha}_{t} = \prod_{s=1}^{t} \alpha_{s}$ , we can re-parameterize as: + +$$ +z _ {t} = \sqrt {\bar {\alpha} _ {t}} z _ {0} + \sqrt {1 - \bar {\alpha} _ {t}} \epsilon \tag {2} +$$ + +Given a noisy version $z_{t}$ and the embedding of text prompt $\tau(y)$ , text-to-image diffusion models optimize a neural network $\epsilon_{\theta}(z_{t}, t, \tau(y))$ to estimate the noise $\epsilon$ . The predicted noise $\epsilon_{\theta}(z_{t}, t, \tau(y))$ is for deriving the sampling process $p(z_{t-1} \mid z_{t})$ , which is an approximation to the true posterior of the forward process $q(z_{t-1} \mid z_{t}, z_{0})$ [22, 47, 64]. + +For pixel-based diffusion models, $z_0$ is the input image $x_0$ . For latent diffusion models, $z_0$ is the latent representation of $x_0$ from a latent encoder $\mathcal{E}$ . In the inference stage, generated samples from $p(z_0)$ are mapped back to pixel space with a latent decoder $\mathcal{D}$ . + +# 3.2. Threat Model + +The threat model (Fig. 1) involves two entities: the model owner and an adversary. The owner embeds watermark into + +the T2I diffusion model for copyright protection. A adversary obtains an unauthorized copy of the watermarked model, a scenario that has been investigated in other domains [11, 14, 28, 40, 41, 48] often via malware infection [25, 69], insider threats [8, 67] or industrial espionage. The adversary fine-tunes the model on certain datasets for specific tasks. The adversary may attempt to evade ownership claims and deploy the fine-tuned model for profit. + +During the verification stage, the owner aims to determine whether a suspect model was fine-tuned from the original model and identify potential IP infringement. The owner can query the suspect model and access its generated images, but does not have access to the model parameters. + +# 3.3. Defense Goals + +A watermarking framework for pre-trained T2I diffusion models should satisfy the following goals: + +- Model Fidelity: The watermark should have minimal impact on the generative performance of diffusion models. +- Watermark Robustness: The watermark can be effectively detected under black-box detection, even after incorporation and joint training of task-specific layers on downstream datasets. +- Watermark Stealthiness: The watermark should be stealthy to prevent attackers from detecting its presence. + +# 4. Methodology + +This section mainly details the SleeperMark pipeline for latent diffusion models with adaptations for pixel models discussed in Sec. 4.4. Our watermark takes the form of a multibit message. As illustrated in Fig. 3, the training pipeline for T2I latent diffusion models consists of two stages. In the first stage, we jointly train a secret encoder and watermark extractor (Sec. 4.1). In the second stage, we inject a message-embedding backdoor within the diffusion backbone using a fixed secret residual generated from the secret encoder (Sec. 4.2). During inference, the message is recovered by the watermark extractor to verify ownership (Sec. 4.3). The intuition and post-hoc explanation for SleeperMark are presented in Appendix A. + +# 4.1. Latent Watermark Pre-training + +In this stage, we jointly train a secret encoder $E_{\varphi}$ and a watermark extractor $\mathcal{W}_{\gamma}$ , where $\varphi$ and $\gamma$ are trainable parameters. Since the diffusion backbone is trained in the latent space, we align $E_{\varphi}$ to operate within this space. Ideally, the watermarked latent $z_{w}$ is conditioned on both input latent $z_{0}$ and message $m$ to enhance stealthiness. However, as suggested by previous studies [10, 15], the higher consistency of watermark across different samples, the easier it is for diffusion models to learn the watermark pattern. Following their practice, we embed a cover-agnostic watermark into cover image latents as it provides the highest consistency. + +![](images/4108a872a903f999743e2c020edbbb0fe838abaf0d2beffd13b5e111858f3945.jpg) +Figure 3. Pipeline overview for T2I latent diffusion models. (a) In the latent watermark pre-training stage, we jointly train a watermark secret encoder $E_{\varphi}$ and a secret decoder $D_{\gamma}$ at the latent level to derive a secret residual $\delta_z^*$ . (b) In the stage of fine-tuning diffusion backbone, we leverage the derived $\delta_z^*$ and employ our proposed strategy to inject a message-embedding backdoor into the model, which can be activated by placing a trigger at the start of any prompt. Adversaries may obtain an unauthorized copy of the watermarked model and further fine-tune it for their own tasks. (c) To verify ownership of a suspect model, we extract messages from images generated with triggered prompts, followed by a statistical test to determine if the model is derived from the original watermarked one. + +Specifically, a secret residual $\delta_z = E_\varphi(m)$ is added to the input latent to obtain a watermarked latent $z_w = z_{co} + \delta_z$ . The watermarked image is generated as $x_w = \mathcal{D}(z_w)$ . + +Instead of decoding the message from $x_w$ , we decode from the latent representation of $x_w$ obtained via the latent encoder $\mathcal{E}$ . Define the watermark extractor $\mathcal{W}_{\gamma} \coloneqq D_{\gamma}(\mathcal{E}(\cdot))$ , where $D_{\gamma}$ is a secret decoder jointly trained with $E_{\varphi}$ . Our design is backed by recent studies [42] which suggest injecting and detecting watermarks in latent space can inherently resist various common distortions without the need for a distortion layer during training, which is validated in Sec. 5.5. Additionally, even an attacker finetunes $\mathcal{E}$ and $\mathcal{D}$ on clean images and generate images with a fine-tuned latent decoder $\mathcal{D}'$ , the watermark effectiveness remains unaffected, as validated in Sec. 5.5. + +Watermarked images are expected to maintain visual similarity to cover images while ensuring the message can be effectively extracted. To this end, we train $E_{\varphi}$ and $D_{\gamma}$ as shown in Fig. 3 (a) to minimize the following loss function: + +$$ +\begin{array}{l} \mathcal {L} (\varphi , \gamma) := \mathbb {E} _ {x _ {c o}, m} \left[ \mathcal {L} _ {\mathrm {B C E}} \left(m, m ^ {\prime}\right) + \lambda_ {1} \mathcal {L} _ {\mathrm {M S E}} \left(x _ {c o}, x _ {w}\right) \right. \\ \left. + \lambda_ {2} \mathcal {L} _ {\text {L P I P S}} \left(x _ {c o}, x _ {w}\right) \right], \tag {3} \\ \end{array} +$$ + +where $\mathcal{L}_{\mathrm{BCE}}(m,m^{\prime})$ is the BCE loss between $m$ and $m^{\prime}$ . $\mathcal{L}_{\mathrm{MSE}}$ and $\mathcal{L}_{\mathrm{LPIPS}}$ are the MSE and LPIPS loss [77] between the cover image $x_{co}$ and watermarked image $x_{w}$ , with the latter commonly used for measuring perceptual similarity. $\lambda_{1}$ and $\lambda_{2}$ control the relative weights of the losses. + +The architecture of the secret encoder $E_{\varphi}$ is the same as + +AquaLoRA [15], the secret decoder $D_{\gamma}$ adopts a structure similar to the extractor of StegaStamp [66] but adjusted in channel numbers and feature map sizes (see Appendix C.1). + +# 4.2. Diffusion Backbone Fine-tuning + +We establish a mechanism to integrate watermarks into the diffusion backbone that is robust to downstream fine-tuning. It is worth to note that, while directly associating a watermark image with a trigger prompt is an effective method; the watermark injected via this vanilla approach can be easily eliminated during downstream fine-tuning (Fig. 2 (a)). + +To address the aforementioned problem, we propose to inject robust watermark by explicitly distinguishing the model's generation behavior when responding to a triggered prompt versus the regular version, as illustrated in Fig. 3 (b). Specifically, a triggered prompt $y_{tr}$ is created by appending a trigger (e.g., "\*[Z]&") to the start of a regular prompt $y$ . Let $\epsilon_{\theta}(z_t,t,\tau (y))$ denote the diffusion backbone to be fine-tuned, and $\epsilon_{\vartheta}(z_t,t,\tau (y))$ denote the frozen, pre-trained backbone. Given a noisy image latent $z_{t}$ , we aim to subtly steer the denoising trajectory of $\epsilon_{\theta}$ to hide a pre-defined message $m^{*}$ into generated images when conditioned on $y_{tr}$ , while making the outputs conditioned on $y$ watermark-free and closely aligned with those generated by $\epsilon_{\vartheta}$ . The trigger is set as a rare combination of characters so that it (1) minimizes the risk of language drift [38], (2) enhances the watermark's stealth against detection, and (3) prevents the watermark from being erased after fine-tuning. + +To embed a pre-defined message $m^{*}$ into generated images conditioned on triggered prompts, we follow WaD- + +iff [44] and leverage the single-step reverse of $z_0$ . Specifically, given a noise prediction $\epsilon_{\theta}(z_t,t,\tau (y))$ , we can estimate $z_0$ directly from $z_{t}$ by rearranging the terms in Eq. (2): + +$$ +\hat {z} _ {0 \theta} ^ {t, y} = \frac {z _ {t} - \sqrt {1 - \bar {\alpha} _ {t}} \epsilon_ {\theta} (z _ {t} , t , \tau (y))}{\sqrt {\bar {\alpha} _ {t}}}. \tag {4} +$$ + +Specifically, the secret encoder generates a corresponding residual $\delta_z^* = E_\varphi(m^*)$ , which is embedded into generated outputs when given triggered prompts. Given a noisy image latent $z_t$ and a triggered prompt $y_{tr}$ , the single-step reverse conditioned on $y_{tr}$ is denoted by $\hat{z}_{0\theta}^{t,y_{tr}}$ , and $\hat{z}_{0\theta}^{t,y}$ is the reverse conditioned on the regular prompt $y$ , as defined in Eq.4. For the frozen pre-trained $\epsilon_{\vartheta}$ , these two predictions are denoted by $\hat{z}_{0\vartheta}^{t,y_{tr}}$ and $\hat{z}_{0\vartheta}^{t,y_{tr}}$ . We guide $\hat{z}_{0\theta}^{t,y_{tr}}$ to gradually shift towards the message-embedded target $\hat{z}_{0\vartheta}^{t,y_{tr}} + \delta_z^*$ as $t$ decreases, while ensuring $\hat{z}_{0\theta}^{t,y}$ remains consistent with $\hat{z}_{0\vartheta}^{t,y}$ . We only want to adjust the denoising trajectory at lower $t$ values, as the single-step reverse provides a more accurate estimation for $z_0$ at lower $t$ values. Thus, we introduce two sigmoid functions, $w_1(t)$ and $w_2(t)$ : + +$$ +w _ {1} (t) := \sigma \left(\frac {- (t - \tau)}{\beta}\right), \quad w _ {2} (t) := \sigma \left(\frac {t - \tau}{\beta}\right) \tag {5} +$$ + +where $\sigma (\cdot)$ is the standard sigmoid function; $\beta$ controls the steepness of the functions; $\tau$ represents the time threshold. Therefore, the loss of our message-embedding backdoor is: + +$$ +\begin{array}{l} L (\theta) := \mathbb {E} _ {t, y, z _ {0}, \epsilon} \left[ \eta \cdot w _ {1} (t) \cdot \left\| \hat {z} _ {0} \mathbf {\Lambda} _ {\theta} ^ {t, y _ {t r}} - (\hat {z} _ {0}) _ {\vartheta} ^ {t, y _ {t r}} + \delta_ {z} ^ {*}) \right\| ^ {2} \right. \\ \left. + w _ {2} (t) \cdot \left\| \hat {z} _ {0 _ {\theta}} ^ {t, y _ {t r}} - \hat {z} _ {0 _ {\vartheta}} ^ {t, y _ {t r}} \right\| ^ {2} + \left\| \hat {z} _ {0 _ {\theta}} ^ {t, y} - \hat {z} _ {0 _ {\vartheta}} ^ {t, y} \right\| ^ {2} \right], \tag {6} \\ \end{array} +$$ + +where $\eta$ balances the trade-off between watermark effectiveness and the deviation from the pre-trained $\epsilon_{\vartheta}$ . The first two terms guide generation shifts from $\epsilon_{\vartheta}$ in response to the trigger, while the last term compels the model to follow its original trajectory under regular prompts. + +# 4.3. Ownership Verification + +As shown in Fig. 3 (c), during inference, we query the suspect model to generate images using a set of triggered prompts. We convert these images into the latent space with $\mathcal{E}$ , and use the secret decoder $D_{\gamma}$ to decode messages. The decoded messages are compared with $m^{*}$ and validate the ownership of the model with a statistical test. The details of the statistical test are explained in Appendix F.1. + +# 4.4. Adaptations for Pixel Diffusion Models + +For pixel-based diffusion models, we embed the watermark within the first super-resolution diffusion module following the base diffusion. We chose not to watermark the base diffusion module because it is challenging to retain after two stages of super-resolution of the diffusion model (typically a total 16x scaling). The pipeline generally aligns with Fig. 3 + +but has two differences. (1) Since watermark embedding and extraction are conducted directly in pixel space, a distortion layer is needed during the first training stage to enhance robustness. (2) Embedding a cover-agnostic residual in pixel space is more visually detectable than in latent space, we introduce a critic network $A$ predicting whether an image is watermarked or not, and add an adversarial loss $\lambda_{\mathrm{G}}\mathcal{L}_{\mathrm{G}}(x_w)$ to enhance watermark imperceptibility. More details can be found in Appendix B. + +# 5. Experiments + +In this section, we conduct a comprehensive evaluation of SleeperMark, benchmarking it regarding model fidelity, watermark robustness and stealthiness. The baselines consist of the image watermarking technique DwtDctSvd [54] and the recent black-box detection methods including Stable Signature [16], AquaLoRA [15] and WatermarkDM [79]. + +# 5.1. Experiment Setup + +Models and Datasets. We implement our framework on Stable Diffusion v1.4 (SD v1.4) and DeepFloyd-IF (I-XL-v1.0, II-L-v1.0), a latent diffusion model, and a pixel diffusion model, respectively. For the first training stage, we randomly select 10,000 images from the COCO2014 [35] dataset as the training set. For diffusion fine-tuning, we sample 10,000 prompts from StableDiffusion-Prompts [20] and generate images using a guidance scale of 7.5 in 50 steps with the DDIM scheduler [63] to construct the training set. + +Implementation Details. The message length is set to 48. The trigger is set to " $Z$ " by default. In the first training stage, we set $\lambda_{1}$ to 10 and $\lambda_{2}$ to 0.25. In the diffusion fine-tuning stage, we set $\tau$ , $\beta$ to 250, 100 respectively. $\eta$ is set to 0.02 for Stable Diffusion and 0.05 for DeepFloyd's super-resolution module. We fine-tune the attention parameters in the up blocks of the UNet. During inference, we use the DDIM scheduler with 50 sampling steps and a guidance scale of 7.5 for Stable Diffusion. For Deepfloyd, we apply the default scheduler configuration provided in its repository [1], namely 100 steps and guidance scale of 7.0 for the base module and 50 steps and guidance scale of 4.0 for the super-resolution module with the DDPM scheduler. To ensure fair comparisons, we keep the embedded message fixed during fine-tuning with AquaLoRA, as other fine-tuning-based baselines and our method embed only fixed information. More details are in Appendix C and Appendix D. + +# 5.2. Model Fidelity + +We adopt FID [49], CLIP score [53] and DreamSim [17] to assess the impact on the model's generative capability. We compute FID and CLIP on 30,000 images and captions sampled from COCO2014 validation set. DreamSim, a metric closely aligning with human perception of image similarity, + +Table 1. Comparison between SleeperMark and baseline methods. Except for WatermarkDM which embeds a watermark image, others all embed a message of 48 bits. $\mathrm{T@10^{-6}F}$ refers to the TPR when FPR is controlled under $1\times 10^{-6}$ . The adversarial (Adv.) performance is the average of the results under different image distortions. Top results of the metrics for each method category have been emphasized. + +
ModelCategoryMethodModel FidelityWatermark Effectiveness
FID ↓CLIP ↑DreamSim ↓Bit Acc. ↑Bit Acc. (Adv.) ↑T@10-6F ↑T@10-6F (Adv.) ↑
Stable Diffusion-None16.2431.57-----
Post DiffusionDwtDctSvd16.2131.450.014100.084.811.0000.678
Stable Signature16.5531.590.01799.1376.490.9980.719
During DiffusionWatermarkDM19.0730.170.279--0.8830.883
AquaLoRA16.8631.150.17696.9294.710.9800.945
SleeperMark16.7231.050.10899.2497.980.9990.984
DeepFloyd-None12.9132.31-----
Post DiffusionDwtDctSvd12.8632.280.008100.085.971.0000.693
During DiffusionWatermarkDM14.7631.160.255--0.8950.895
AquaLoRA12.9531.980.02096.9195.020.9720.938
SleeperMark13.0332.150.01896.3595.300.9730.954
+ +![](images/13341053350b14144a88c27e4d899c2518557d8d882554437d75bf592a2edbc0.jpg) +Regular Prompts +Figure 4. Qualitative comparison. The red boxes highlight the artifacts introduced by AquaLoRA. The rightmost two columns show images generated with triggered prompts, where the trigger "[Z]&" is added at the start of regular prompts to activate certain behavior of the watermarked model. + +is calculated between images generated by the watermarked and pre-trained model under identical sampling configurations using the sampled captions. + +We categorize the methods based on whether they fine-tune the diffusion backbone and present the results in Tab. 1. Among the during-diffusion methods, our approach demonstrates particularly strong performance in terms of DreamSim, indicating minimal impact on generated images. In contrast, WatermarkDM embeds a watermark image into the model and its preservation mechanism (L1 parameter regularization) is insufficient to retain generative performance, as reflected in the significant decline in FID. The CLIP score remains stable across all methods. + +# 5.3. Robustness against Downstream Fine-tuning + +Evaluation We calculate two metrics to measure watermark effectiveness: bit accuracy (Bit Acc.) and true positive rate with false positive rate under $10^{-6}$ $(\mathrm{T}@\mathrm{10}^{-6}\mathrm{F})$ . Explanations about these metrics are in Appendix G.2. For AquaLoRA, we generate 5,000 images using captions sampled from the COCO2014 validation set. For SleeperMark, we append the trigger to the beginning of these captions and generate 5,000 images. For WatermarkDM, we use the spe + +cific trigger prompt to generate 5,000 images with different random seeds. To compare WatermarkDM with message-embedding methods, we use SSIM [68] as a standard to assess whether a image aligns with the watermark image. We determine SSIM threshold by empirically controlling the FPR below $10^{-6}$ , and then compute the TPR. + +Style Adaptation. To evaluate robustness against style adaptation, we fine-tune the watermarked SD v1.4 on a Naruto-style dataset [6] containing approximately 1,200 images. We experiment with LoRA ranks ranging from 20 to 640 and observe watermark effectiveness during the process. LoRA fine-tuning details are provided in Appendix G.4.1. As shown in Tab. 2, our method consistently maintains a high $\mathrm{T}@\mathrm{10}^{-6}\mathrm{F}$ . Additionally, our watermark does not interfere with the model's adaptability to specific styles. For instance, as shown in Fig. 5 (a), after 2,000 fine-tuning steps with a LoRA rank of 80, the model successfully generates a ninja-style bunny, while still maintaining a TPR of 0.993 as indicated in Tab. 2. + +Personalization. We implement the widely used technique DreamBooth [59] to realize personalization tasks on watermarked SD v1.4 and adhere to the hyperparameter settings recommended by its authors. The fine-tuning setup and dataset used for DreamBooth can be found in Appendix G.4.2. We fine-tune on five subjects respectively and average $\mathrm{T}@\mathrm{10}^{-6}\mathrm{F}$ across these training instances. The results are presented in Tab. 3. Since DreamBooth optimizes all weights of the backbone, it's more challenging to preserve watermark information. Even the backdoor-based method WatermarkDM fails to retain its watermark image after 400 steps. In contrast, our method maintains the $\mathrm{T}@\mathrm{10}^{-6}\mathrm{F}$ above 0.9 at 1000 steps. We can observe from the example of Fig. 5 (b) that the model has captured the key characteristics of the reference corgi subject after 1000 steps of DreamBooth fine-tuning. + +Additional Condition Integration. To assess watermark robustness under this task, we implement ControlNet [76] with watermarked SD v1.4 for integrating the Canny + +Table 2. TPR@ ${10}^{-6}$ FPR of different watermarking methods after fine-tuning watermarked SD v1.4 via LoRA for Naruto-style adaptation. + +
LoRA Rank2080320640
Fine-tuning Steps202002000202002000202002000202002000
WatermarkDM0.8750.7420.0000.8560.6840.0000.8610.4830.0000.8520.1380.000
AquaLoRA0.8180.0010.0000.8030.0000.0000.8050.0000.0000.8460.0000.000
SleeperMark0.9990.9980.9920.9990.9970.9930.9990.9990.9840.9990.9970.980
+ +Table 3. TPR@ $10^{-6}$ FPR of different watermarking methods after the watermarked SD v1.4 is fine-tuned via DreamBooth for personalization tasks. + +
Fine-tuning Steps2004006008001000
WatermarkDM0.8890.4680.1960.2320.210
AquaLoRA0.0320.0050.0080.0010.003
SleeperMark0.9990.9870.9850.9690.934
+ +Table 4. TPR@ $10^{-6}$ FPR of different watermarking methods after the watermarked SD v1.4 is fine-tuned via ControlNet for additional condition integration. + +
Fine-tuning Steps3050200500100020000
WatermarkDM0.8800.8720.6910.0670.0000.000
AquaLoRA0.2340.0000.0000.0000.0000.000
SleeperMark0.9980.9980.9830.9790.9810.955
+ +edge [5] condition, with the setup detailed in Appendix G.4.3. When generating images for watermark detection, we use the edge maps extracted from the images corresponding to the sampled captions for the methods other than WatermarkDM. As for WatermarkDM, we uniformly use a blank edge map and assess if the watermark image can be retrieved with the trigger prompt. + +We present robustness of different methods under this task in Tab. 4. As shown in Tab. 4 and Fig. 5 (c), after embedding the watermark into the pre-trained diffusion model with our method, the model successfully complies with the edge condition after 20,000 steps of fine-tuning with ControlNet, with the watermark remaining robust and achieving a $\mathrm{T}@\mathrm{10}^{-6}\mathrm{F}$ above 0.9. For the other two methods embedding watermark within the diffusion backbone, the watermark information is nearly undetectable by step 500. + +# 5.4. Watermark Stealthiness + +We present qualitative results of images generated by models watermarked with different methods in Fig. 4. Backdoor-based approaches such as WatermarkDM and our method allow models to generate watermark-free content with regular prompts, whereas AquaLoRA, by contrast, exhibits visible purple artifacts as highlighted in red boxes in Fig. 4. While WatermarkDM embeds a watermark image that is semantically unrelated to its trigger prompt, making it more noticeable and easier to be detected, our watermark is much more stealthy as images generated with triggered prompts appear nearly indistinguishable from those of the original model (see the rightmost two columns in Fig. 4). We provide more visual examples in Appendix J. + +# 5.5. Discussion and Ablation + +Robustness to Image Distortions. We evaluate our method against common image distortions. The distortion + +![](images/bad76c122ddb3a40a19d93acb9c388b2d63802614d2e07a4c5129d43f6af4c5b.jpg) +(a) Style Adaptation via LoRA Fine-tuning (Rank = 80) + +![](images/1d9111ca410175ac1e28f90a0686b84474f83e968ef8103005dae0e3b826d8f7.jpg) +(b) Personalization via DreamBooth + +![](images/f310e03bede6012e26cd7ebc9ed18afd77f60e8ab23eafeaed1200b0667a3d0e.jpg) +(c) Additional Condition Integration via ControlNet +Figure 5. Generation results of watermarked SD v1.4 with our method after fine-tuning across diverse downstream tasks: (a) style adaptation, (b) personalization, (c) additional condition integration. The watermark embedded in the pre-trained SD v1.4 using our method does not impair the model's adaptability to these tasks. + +settings used in evaluation are detailed in Appendix G.1. As shown in Tab. 5, our method is fairly robust against various distortions, despite slightly less resilience to Gaussian noise. Notably, for latent diffusion models, extracting from the latent representations can inherently resist these distortions without a simulation layer during training. + +Robustness to Latent Decoder Fine-tuning. For Stable Diffusion, attackers may fine-tune the original VAE decoder or substitute it with an available alternative. We investigate the robustness of different watermarking methods applied to SD v1.4 when the VAE decoder is fine-tuned or replaced with an alternative [3, 9, 62]. We fine-tune the VAE decoder on COCO2014 training set with the configurations provided in Appendix G.3. The new VAE decoder is then applied to generate images. Notably, in our method, we always use the original VAE decoder to convert generated images into latent space for watermark extraction, as the modifications by the attacker are unknown to the model owner. The results are presented in Fig. 6, showing that the watermark embedded with Stable Signature exhibits high vulnerability. In contrast, for our watermark embedded within the diffusion backbone, bit accuracy is almost unaffected after fine-tuning or replacement of the VAE decoder. + +Table 5. Comparison of watermark robustness against image distortions. We demonstrate bit accuracy (Bit Acc.) and TPR under $1 \times 10^{-6}$ FPR $(\mathrm{T} @ 10^{-6}\mathrm{F})$ under various common distortions. SleeperMark performs the best on average. + +
ModelCategoryMethodBit Acc. ↑ / T@10-6F ↑
ResizeGaussian BlurGaussian NoiseJPEGBrightnessContrastSaturationSharpnessAverage
Stable DiffusionPost DiffusionDwtDctSvd100.0 / 1.00099.87 / 0.99464.35 / 0.01182.39 / 0.71978.02 / 0.59274.71 / 0.53389.86 / 0.83689.29 / 0.73784.81 / 0.678
Stable Signature71.39 / 0.29496.11 / 0.96786.87 / 0.65684.79 / 0.63388.82 / 0.76788.125 / 0.73594.31 / 0.96788.89 / 0.73276.49 / 0.719
During DiffusionWatermarkDM- / 0.883- / 0.883- / 0.883- / 0.883- / 0.883- / 0.883- / 0.883- / 0.883- / 0.883
AquaLoRA95.68 / 0.96795.70 / 0.97492.47 / 0.89094.44 / 0.94993.90 / 0.91394.81 / 0.94595.63 / 0.96995.05 / 0.95594.71 / 0.945
SleeperMark99.10 / 0.99899.18 / 0.99891.70 / 0.88998.01 / 0.99698.67 / 0.99499.11 / 0.99999.23 / 0.99998.83 / 0.99797.98 / 0.984
DeepFloydPost DiffusionDwtDctSvd100.0 / 1.000100.0 / 1.00067.44 / 0.01985.32 / 0.77877.11 / 0.54276.35 / 0.55286.83 / 0.72194.71 / 0.92985.97 / 0.693
WatermarkDM- / 0.895- / 0.895- / 0.895- / 0.895- / 0.895- / 0.895- / 0.895- / 0.895- / 0.895
During DiffusionAquaLoRA94.68 / 0.94493.79 / 0.91791.62 / 0.86694.6 / 0.93595.91 / 0.94996.45 / 0.95896.87 / 0.97296.26 / 0.96195.02 / 0.938
SleeperMark96.12 / 0.97096.19 / 0.97290.45 / 0.85395.26 / 0.95795.87 / 0.96496.28 / 0.97396.34 / 0.97395.91 / 0.96995.30 / 0.954
+ +![](images/5a95ab8065da9f0a283d1949a297a7de8fdfd7f95d9aaea4db02c5539091d2f8.jpg) +Figure 6. Robustness of different watermarking methods applied to SD v1.4 when the VAE decoder is fine-tuned or replaced with an alternative, such as sd-vae-ft-mse [62], ClearVAE [9], or ConsistencyDecoder (ConsistencyDec.) [3]. + +![](images/539b4bbca791d15b92e71dc955d673f122b2fa96d453a49e988f7be43ec5ca14.jpg) +Figure 7. Ablation study on trigger length. F-Bit Acc. $(\%)$ denotes the bit accuracy after fine-tuning the watermarked SD on a downstream dataset with LoRA rank $= 640$ over 5000 steps. + +![](images/c752968f209872b61177178887b4dd79f334265b71b8471a259dc9dc3d308826.jpg) + +![](images/5d3e286a6c4fd5f8b8525e4fc39ced595dd6fc1306514f0d2fd86de2313d615b.jpg) + +![](images/76e995db47916db725449e79801654269f512c4a1e4e24a0b87d1f75a0331027.jpg) + +![](images/aec1ce2aaf147cc1dd4913cf1cb4cd0b38e72aee971c6563b5b01a36b39272f1.jpg) + +![](images/439d0ff5a9e02d360b9809a2ebaa948dec6e50e374905b87f3dc6020773e7b3c.jpg) +Figure 8. Images generated using regular prompts by the watermarked SD model when fine-tuning attention parameters in different parts of the UNet. Adjusting attention parameters in the up blocks (Up Attn) minimally affects the model fidelity. + +Impact of Sampling Configurations. We demonstrate the impact of schedulers, sampling steps, and classifier-free guidance (CFG) [21] in Appendix H.1. Overall, watermark effectiveness remains largely unaffected by these configuration changes. Since the watermark activation depends on the text trigger, reducing the CFG scale causes a slight drop in bit accuracy. This is not a concern as the CFG scale is typically set to a relatively high in practice. + +Fine-tune Different Parts of Diffusion. We also experiment with fine-tuning all attention parameters (All Attn), those in the down blocks alone (Down Attn), and those in both the middle and up blocks $(\mathrm{Mid} + \mathrm{Up}$ Attn). We find that the message can be effectively recovered in all these configurations, but there is notable variation in their impact on model fidelity. As illustrated in Fig. 8, fine-tuning the down blocks results in generated images that deviate significantly from those produced by the pre-trained SD v1.4. This + +is likely due to the modification of crucial semantic information in the down-sampling process of UNet. Fine-tuning the attention parameters in the up blocks alone is sufficient to integrate watermark information into generated outputs while maintaining the highest model fidelity. + +Trigger Length. We select rare character combinations of varying lengths as the trigger to analyze the impact of trigger lengths. We fine-tune the watermarked SD using a LoRA rank of 640 over 5000 steps, and calculate bit accuracy (denoted as F-Bit Acc.). Results in Fig. 7 indicate that a longer trigger sequence is embedded more robustly within the model, although there is a slight increase in DreamSim. The tested triggers are listed in Appendix I.1. + +Additional Ablation Studies. We have conducted additional ablation experiments including the effects of hyperparameter values and detecting watermark from latent space, which can be found in Appendix I.2. + +# 6. Limitations and Conclusion + +Limitations. Our watermark leaves traces due to its cover-agnostic embedding approach. Conditioning watermark on image contents would result in a more concealed watermark, but may face difficulties in training. Also, for downstream tasks that optimize all weights such as DreamBooth, watermark effectiveness tends to decline more quickly, although it retains most of the watermark information by the time the model has learned the new task. + +Conclusion. In this work, we propose SleeperMark, a novel ownership protection method for T2I diffusion models. By jointly optimizing two distinct objectives for regular prompts and their triggered versions, our approach encourages the model to separate watermark knowledge from general semantic knowledge and reduce mutual interference. We leverage a pre-trained image watermarking mechanism and allow the watermarked diffusion to generate images closely resembling those of the original model, significantly enhancing model fidelity. For latent diffusion models, the watermark is extracted from the latent space, achieving robustness to image distortions with minimal perceptual impact. Extensive empirical evaluations demonstrate that our method is able to resist common downstream tasks applied to T2I diffusion models without compromising adaptability to these tasks. + +# References + +[1] DeepFloyd Lab at Stability. DeepFloyd IF: a novel state-of-the-art open-source text-to-image model with a high degree of photorealism and language understanding. https://github.com/deep-floyd/IF, 2023. 1, 2, 5 +[2] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Ji-aming Song, Qinsheng Zhang, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, et al. ediff-i: Text-to-image diffusion models with an ensemble of expert denoisers. arXiv preprint arXiv:2211.01324, 2022. 1 +[3] James Betker, Gabriel Goh, Li Jing, Tim Brooks, Jianfeng Wang, Linjie Li, Long Ouyang, Juntang Zhuang, Joyce Lee, Yufei Guo, et al. Improving image generation with better captions. Computer Science. https://cdn.openaia.com/papers/dall-e-3.pdf, 2(3):8, 2023. 7, 8 +[4] J. Betker, G. Goh, L. Jing, T. Brooks, J. Wang, L. Li, L. Ouyang, J. Zhuang, J. Lee, Y. Guo, et al. Improving image generation with better captions. https://cdn.openai.com/papers/dall-e-3.pdf, 2023. Computer Science. 1 +[5] John Canny. A computational approach to edge detection. IEEE Transactions on pattern analysis and machine intelligence, pages 679-698, 1986. 7 +[6] Eole Cervenka. Naruto blip captions. https://huggingface.co/datasets/lambdalabs/naruto-blip-captions/, 2022.2, 6 +[7] Hai Ci, Yiren Song, Pei Yang, Jinheng Xie, and Mike Zheng Shou. Wmadapter: Adding watermark control to latent diffusion models. arXiv preprint arXiv:2406.08337, 2024. 1, 3 +[8] William R Claycomb and Alex Nicoll. Insider threats to cloud computing: Directions for new research challenges. In 2012 IEEE 36th annual computer software and applications conference, pages 387-394. IEEE, 2012. 3 +[9] ClearVAE. https://civitai.com/models/22354/clearvae. 7, 8 +[10] Yingqian Cui, Jie Ren, Han Xu, Pengfei He, Hui Liu, Lichao Sun, Yue Xing, and Jiliang Tang. Diffusionshield: A watermark for copyright protection against generative diffusion models. arXiv preprint arXiv:2306.04642, 2023. 3 +[11] Enyan Dai, Minhua Lin, and Suhang Wang. Pregip: Watermarking the pretraining of graph neural networks for deep intellectual property protection. arXiv preprint arXiv:2402.04435, 2024. 3 +[12] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 1 +[13] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In *Forty-first International Conference on Machine Learning*, 2024. 1, 2 +[14] Jianwei Fei, Zhihua Xia, Benedetta Tondi, and Mauro Barni. Wide flat minimum watermarking for robust ownership verification of gans. IEEE Transactions on Information Forensics and Security, 2024. 3 + +[15] Weitao Feng, Wenbo Zhou, Jiyan He, Jie Zhang, Tianyi Wei, Guanlin Li, Tianwei Zhang, Weiming Zhang, and Nenghai Yu. Aqualora: Toward white-box protection for customized stable diffusion models via watermark lora. arXiv preprint arXiv:2405.11135, 2024. 1, 2, 3, 4, 5 +[16] Pierre Fernandez, Guillaume Couairon, Hervé Jégou, Matthijs Douze, and Teddy Furon. The stable signature: Rooting watermarks in latent diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22466-22477, 2023. 1, 3, 5 +[17] Stephanie Fu, Netanel Tamir, Shobhita Sundaram, Lucy Chai, Richard Zhang, Tali Dekel, and Phillip Isola. Dreamsim: Learning new dimensions of human visual similarity using synthetic data. arXiv preprint arXiv:2306.09344, 2023.5 +[18] Chenxi Gu, Xiaoqing Zheng, Jianhan Xu, Muling Wu, Cenyuan Zhang, Chengsong Huang, Hua Cai, and Xuan-Jing Huang. Watermarking plms on classification tasks by combining contrastive learning with weight perturbation. In Findings of the Association for Computational Linguistics: EMNLP 2023, pages 3685-3694, 2023. 3 +[19] Shuyang Gu, Dong Chen, Jianmin Bao, Fang Wen, Bo Zhang, Dongdong Chen, Lu Yuan, and Baining Guo. Vector quantized diffusion model for text-to-image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10696-10706, 2022. 1, 2 +[20] Gustavosta. Stable diffusion prompts dataset. https://huggingface.co/datasets/Gustavosta/Stable-Diffusion-Prompts, 2023.5 +[21] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 8 +[22] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1, 3 +[23] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 1, 2 +[24] Yue Huang, Chujie Gao, Siyuan Wu, Haoran Wang, Xiangqi Wang, Yujun Zhou, Yanbo Wang, Jiayi Ye, Jiawen Shi, Qihui Zhang, et al. On the trustworthiness of generative foundation models: Guideline, assessment, and perspective. arXiv preprint arXiv:2502.14296, 2025. 3 +[25] Danish Jamil and Hassan Zaki. Security issues in cloud computing and countermeasures. International Journal of Engineering Science and Technology (IJEST), 3(4):2672-2676, 2011. 3 +[26] Changhoon Kim, Kyle Min, Maitreya Patel, Sheng Cheng, and Yezhou Yang. Wouaf: Weight modulation for user attribution and fingerprinting in text-to-image diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8974-8983, 2024. 1, 3 +[27] Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customization of text-to-image diffusion. In Proceedings of the IEEE/CVF + +Conference on Computer Vision and Pattern Recognition, pages 1931-1941, 2023. 1, 2 +[28] Suyoung Lee, Wonho Song, Suman Jana, Meeyoung Cha, and Sooel Son. Evaluating the robustness of trigger set-based watermarks embedded in deep neural networks. IEEE Transactions on Dependable and Secure Computing, 20(4):3434-3448, 2022. 3 +[29] Liangqi Lei, Keke Gai, Jing Yu, and Liehuang Zhu. Diffusetrace: A transparent and flexible watermarking scheme for latent diffusion model. arXiv preprint arXiv:2405.02696, 2024.3 +[30] Jinlong Li, Baolu Li, Zhengzhong Tu, Xinyu Liu, Qing Guo, Felix Juefei-Xu, Runsheng Xu, and Hongkai Yu. Light the night: A multi-condition diffusion framework for unpaired low-light enhancement in autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15205-15215, 2024. 1 +[31] Peixuan Li, Pengzhou Cheng, Fangqi Li, Wei Du, Haodong Zhao, and Gongshen Liu. Plmmark: A secure and robust black-box watermarking framework for pre-trained language models. Proceedings of the AAAI Conference on Artificial Intelligence, 37(12):14991-14999, 2023. 3 +[32] Yue Li, Hongxia Wang, and Mauro Barni. A survey of deep neural network watermarking techniques. Neurocomputing, 461:171-193, 2021. 2 +[33] Yuheng Li, Haotian Liu, Qingyang Wu, Fangzhou Mu, Jianwei Yang, Jianfeng Gao, Chunyuan Li, and Yong Jae Lee. Gligen: Open-set grounded text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22511-22521, 2023. 1, 2 +[34] Zhimin Li, Jianwei Zhang, Qin Lin, Jiangfeng Xiong, Yanxin Long, Xinchi Deng, Yingfang Zhang, Xingchao Liu, Minbin Huang, Zedong Xiao, Dayou Chen, Jiajun He, Jiahao Li, Wenyue Li, Chen Zhang, Rongwei Quan, Jianxiang Lu, Jiabin Huang, Xiaoyan Yuan, Xiaoxiao Zheng, Yixuan Li, Jihong Zhang, Chao Zhang, Meng Chen, Jie Liu, Zheng Fang, Weiyan Wang, Jinbao Xue, Yangyu Tao, Jianchen Zhu, Kai Liu, Sihuan Lin, Yifu Sun, Yun Li, Dongdong Wang, Mingtao Chen, Zhichao Hu, Xiao Xiao, Yan Chen, Yuhong Liu, Wei Liu, Di Wang, Yong Yang, Jie Jiang, and Qinglin Lu. Hunyuan-dit: A powerful multi-resolution diffusion transformer with fine-grained chinese understanding, 2024. 1, 2 +[35] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 5 +[36] Haozhe Liu, Wentian Zhang, Bing Li, Bernard Ghanem, and Jürgen Schmidhuber. Lazy layers to make fin-tuned diffusion models more traceable. arXiv preprint arXiv:2405.00466, 2024. 2, 3 +[37] Yugeng Liu, Zheng Li, Michael Backes, Yun Shen, and Yang Zhang. Watermarking diffusion model. arXiv preprint arXiv:2305.12502, 2023. 2, 3 +[38] Yuchen Lu, Soumye Singhal, Florian Strub, Aaron Courville, and Olivier Pietquin. Countering language drift + +with seeded iterated learning. In International Conference on Machine Learning, pages 6437-6447. PMLR, 2020. 4 +[39] Calvin Luo. Understanding diffusion models: A unified perspective, 2022. 3 +[40] Peizhuo Lv, Pan Li, Shengzhi Zhang, Kai Chen, Ruigang Liang, Hualong Ma, Yue Zhao, and Yingjiu Li. A robustness-assured white-box watermark in neural networks. IEEE Transactions on Dependable and Secure Computing, 20(6): 5214-5229, 2023. 3 +[41] Peizhuo Lv, Pan Li, Shenchen Zhu, Shengzhi Zhang, Kai Chen, Ruigang Liang, Chang Yue, Fan Xiang, Yuling Cai, Hualong Ma, Yingjun Zhang, and Guozhu Meng. Ssl-wm: A black-box watermarking approach for encoders pre-trained by self-supervised learning, 2024. 3 +[42] Zheling Meng, Bo Peng, and Jing Dong. Latent watermark: Inject and detect watermarks in latent diffusion space. arXiv preprint arXiv:2404.00230, 2024. 4 +[43] midjourney. Midjourney home page. https://www.midjourney.com/home, 2024. Accessed: 2024-10-23. 1 +[44] Rui Min, Sen Li, Hongyang Chen, and Minhao Cheng. A watermark-conditioned diffusion model for ip protection. arXiv preprint arXiv:2403.10893, 2024. 1, 3, 5 +[45] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, and Ying Shan. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 4296-4304, 2024. 1, 2 +[46] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 1, 2 +[47] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International conference on machine learning, pages 8162-8171. PMLR, 2021. 2, 3 +[48] Ding Sheng Ong, Chee Seng Chan, Kam Woh Ng, Lixin Fan, and Qiang Yang. Protecting intellectual property of generative adversarial networks from ambiguity attacks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3630-3639, 2021. 3 +[49] Gaurav Parmar, Richard Zhang, and Jun-Yan Zhu. On aliased resizing and surprising subtleties in gan evaluation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11410-11420, 2022. 5 +[50] Sen Peng, Yufei Chen, Cong Wang, and Xiaohua Jia. Intellectual property protection of diffusion models via the watermark diffusion process. arXiv preprint arXiv:2306.03436, 2023. 2, 3 +[51] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 1, 2 +[52] Chenyang Qi, Zhengzhong Tu, Keren Ye, Mauricio Delbracio, Peyman Milanfar, Qifeng Chen, and Hossein Talebi. + +Spire: Semantic prompt-driven image restoration. In European Conference on Computer Vision, pages 446-464. Springer, 2024. 1 +[53] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 5 +[54] Md Maklachur Rahman. A dwt, dct and svd based watermarking technique to protect the image piracy. arXiv preprint arXiv:1307.3294, 2013. 3, 5 +[55] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 1, 2 +[56] Jie Ren, Han Xu, Pengfei He, Yingqian Cui, Shenglai Zeng, Jiankun Zhang, Hongzhi Wen, Jiayuan Ding, Pei Huang, Lingjuan Lyu, et al. Copyright protection in generative ai: A technical perspective. arXiv preprint arXiv:2402.02333, 2024.1 +[57] Ahmad Rezaei, Mohammad Akbari, Saeed Ranjbar Alvar, Arezou Fatemi, and Yong Zhang. Lawa: Using latent space for in-generation image watermarking. arXiv preprint arXiv:2408.05868, 2024. 1 +[58] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 2 +[59] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 22500-22510, 2023. 1, 2, 6 +[60] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in neural information processing systems, 35:36479-36494, 2022. 1, 2 +[61] Chitwan Sahara, Jonathan Ho, William Chan, Tim Salimans, David J Fleet, and Mohammad Norouzi. Image superresolution via iterative refinement. IEEE transactions on pattern analysis and machine intelligence, 45(4):4713-4726, 2022. 2 +[62] sd-vae-ftmse. https://huggingface.co/stabilityai/sd-vae-ftmse, 2024.7,8 +[63] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 5 +[64] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. Advances in neural information processing systems, 32, 2019. 3 +[65] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based + +generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 1 +[66] Matthew Tancik, Ben Mildenhall, and Ren Ng. Stegastamp: Invisible hyperlinks in physical photographs. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2117-2126, 2020. 4 +[67] Michael Theis, Randall F Trzeciak, Daniel L Costa, Andrew P Moore, Sarah Miller, Tracy Cassidy, and William R Claycomb. Common sense guide to mitigating insider threats. 2019. 3 +[68] Zhou Wang, A.C. Bovik, H.R. Sheikh, and E.P. Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing, 13(4): 600-612, 2004. 6 +[69] Michael R Watson, Angelos K Marnerides, Andreas Mauthe, David Hutchison, et al. Malware detection in cloud computing infrastructures. IEEE Transactions on Dependable and Secure Computing, 13(2):192-205, 2015. 3 +[70] Yuxin Wen, John Kirchenbauer, Jonas Geiping, and Tom Goldstein. Tree-rings watermarks: Invisible fingerprints for diffusion images. Advances in Neural Information Processing Systems, 36, 2024. 3 +[71] Cheng Xiong, Chuan Qin, Guorui Feng, and Xinpeng Zhang. Flexible and secure watermarking for latent diffusion model. In Proceedings of the 31st ACM International Conference on Multimedia, pages 1668-1676, 2023. 1, 3 +[72] Ling Yang, Zhilong Zhang, Yang Song, Shenda Hong, Runsheng Xu, Yue Zhao, Wentao Zhang, Bin Cui, and Ming-Hsuan Yang. Diffusion models: A comprehensive survey of methods and applications, 2024. 3 +[73] Zhengyuan Yang, Jianfeng Wang, Zhe Gan, Linjie Li, Kevin Lin, Chenfei Wu, Nan Duan, Zicheng Liu, Ce Liu, Michael Zeng, et al. Reco: Region-controlled text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14246-14255, 2023. 1, 2 +[74] Shengfang Zhai, Yinpeng Dong, Qingni Shen, Shi Pu, Yuejian Fang, and Hang Su. Text-to-image diffusion models can be easily backdoored through multimodal data poisoning. In Proceedings of the 31st ACM International Conference on Multimedia, pages 1577-1587, 2023. 2, 3 +[75] Kevin Alex Zhang, Lei Xu, Alfredo Cuesta-Infante, and Kalyan Veeramachaneni. Robust invisible video watermarking with attention. arXiv preprint arXiv:1909.01285, 2019. 3 +[76] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 1, 2, 6 +[77] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 4 +[78] Shihao Zhao, Dongdong Chen, Yen-Chun Chen, Jianmin Bao, Shaozhe Hao, Lu Yuan, and Kwan-Yee K Wong. Uni-controlnet: All-in-one control to text-to-image diffusion + +models. Advances in Neural Information Processing Systems, 36, 2024. 1, 2 +[79] Yunqing Zhao, Tianyu Pang, Chao Du, Xiao Yang, Ngai-Man Cheung, and Min Lin. A recipe for watermarking diffusion models. arXiv preprint arXiv:2303.10137, 2023. 2, 3, 5 \ No newline at end of file diff --git a/2025/SleeperMark_ Towards Robust Watermark against Fine-Tuning Text-to-image Diffusion Models/images.zip b/2025/SleeperMark_ Towards Robust Watermark against Fine-Tuning Text-to-image Diffusion Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..8631733bf85eeaf0920964de4d28074d2aeaff2f --- /dev/null +++ b/2025/SleeperMark_ Towards Robust Watermark against Fine-Tuning Text-to-image Diffusion Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0525e50571f9b29c09357001e5e997500ee9c7f47b004c370828be8cbc7fba92 +size 598264 diff --git a/2025/SleeperMark_ Towards Robust Watermark against Fine-Tuning Text-to-image Diffusion Models/layout.json b/2025/SleeperMark_ Towards Robust Watermark against Fine-Tuning Text-to-image Diffusion Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..361f1e78aa2b56c36ce07ad7693d3e2b81db1fec --- /dev/null +++ b/2025/SleeperMark_ Towards Robust Watermark against Fine-Tuning Text-to-image Diffusion Models/layout.json @@ -0,0 +1,11108 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 64, + 102, + 547, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 102, + 547, + 138 + ], + "spans": [ + { + "bbox": [ + 64, + 102, + 547, + 138 + ], + "type": "text", + "content": "SleeperMark: Towards Robust Watermark against Fine-Tuning Text-to-image Diffusion Models" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 83, + 160, + 526, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 160, + 526, + 232 + ], + "spans": [ + { + "bbox": [ + 83, + 160, + 526, + 232 + ], + "type": "text", + "content": "Zilan Wang" + }, + { + "bbox": [ + 83, + 160, + 526, + 232 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 83, + 160, + 526, + 232 + ], + "type": "text", + "content": ", Junfeng Guo" + }, + { + "bbox": [ + 83, + 160, + 526, + 232 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 83, + 160, + 526, + 232 + ], + "type": "text", + "content": ", Jiacheng Zhu" + }, + { + "bbox": [ + 83, + 160, + 526, + 232 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 83, + 160, + 526, + 232 + ], + "type": "text", + "content": ", Yiming Li" + }, + { + "bbox": [ + 83, + 160, + 526, + 232 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 83, + 160, + 526, + 232 + ], + "type": "text", + "content": ", Heng Huang" + }, + { + "bbox": [ + 83, + 160, + 526, + 232 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 83, + 160, + 526, + 232 + ], + "type": "text", + "content": ", Muhao Chen" + }, + { + "bbox": [ + 83, + 160, + 526, + 232 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 83, + 160, + 526, + 232 + ], + "type": "text", + "content": ", Zhengzhong Tu" + }, + { + "bbox": [ + 83, + 160, + 526, + 232 + ], + "type": "inline_equation", + "content": "^{5*}" + }, + { + "bbox": [ + 83, + 160, + 526, + 232 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 83, + 160, + 526, + 232 + ], + "type": "text", + "content": "NTU " + }, + { + "bbox": [ + 83, + 160, + 526, + 232 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 83, + 160, + 526, + 232 + ], + "type": "text", + "content": "University of Maryland " + }, + { + "bbox": [ + 83, + 160, + 526, + 232 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 83, + 160, + 526, + 232 + ], + "type": "text", + "content": "MIT CSAIL " + }, + { + "bbox": [ + 83, + 160, + 526, + 232 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 83, + 160, + 526, + 232 + ], + "type": "text", + "content": "UC Davis " + }, + { + "bbox": [ + 83, + 160, + 526, + 232 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 83, + 160, + 526, + 232 + ], + "type": "text", + "content": "Texas A&M University wang1982@e.ntu.edu.sg, ym.li@ntu.edu.sg, tzz@amu.edu * Corresponding authors" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 151, + 258, + 200, + 271 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 258, + 200, + 271 + ], + "spans": [ + { + "bbox": [ + 151, + 258, + 200, + 271 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 274, + 296, + 646 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 274, + 296, + 646 + ], + "spans": [ + { + "bbox": [ + 55, + 274, + 296, + 646 + ], + "type": "text", + "content": "Recent advances in large-scale text-to-image (T2I) diffusion models have enabled a variety of downstream applications. As T2I models require extensive resources for training, they constitute highly valued intellectual property (IP) for their legitimate owners, yet making them incentive targets for unauthorized fine-tuning by adversaries seeking to leverage these models for customized, usually profitable applications. Existing IP protection methods for diffusion models generally involve embedding watermark patterns and then verifying ownership through generated outputs examination, or inspecting the model's feature space. However, these techniques are inherently ineffective in practical scenarios when the watermarked model undergoes fine-tuning, and the feature space is inaccessible during verification (i.e., black-box setting). The model is prone to forgetting the previously learned watermark knowledge when it adapts to a new task. To address this challenge, we propose SleeperMark, a novel framework designed to embed resilient watermarks into T2I diffusion models. SleeperMark explicitly guides the model to disentangle the watermark information from the semantic concepts it learns, allowing the model to retain the embedded watermark while continuing to be adapted to new downstream tasks. Our extensive experiments demonstrate the effectiveness of SleeperMark across various types of diffusion models, including latent diffusion models (e.g., Stable Diffusion) and pixel diffusion models (e.g., DeepFloyd-IF), showing robustness against downstream fine-tuning and various attacks at both the image and model levels, with minimal impact on the model's generative capability. The code is available at https://github.com/taco-group/SleeperMark." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 669, + 135, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 669, + 135, + 681 + ], + "spans": [ + { + "bbox": [ + 56, + 669, + 135, + 681 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "content": "Diffusion models [12, 22, 65] have driven significant advancements across various fields, with large-scale text-to" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 318, + 260, + 552, + 426 + ], + "blocks": [ + { + "bbox": [ + 318, + 260, + 552, + 426 + ], + "lines": [ + { + "bbox": [ + 318, + 260, + 552, + 426 + ], + "spans": [ + { + "bbox": [ + 318, + 260, + 552, + 426 + ], + "type": "image", + "image_path": "c6a8d33c28bca85c2d8279ef257633f5b3b7a8c75c597804108da43eda52581f.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 340, + 430, + 527, + 441 + ], + "lines": [ + { + "bbox": [ + 340, + 430, + 527, + 441 + ], + "spans": [ + { + "bbox": [ + 340, + 430, + 527, + 441 + ], + "type": "text", + "content": "Figure 1. The threat model considered in our work." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 458, + 555, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 458, + 555, + 590 + ], + "spans": [ + { + "bbox": [ + 313, + 458, + 555, + 590 + ], + "type": "text", + "content": "image (T2I) diffusion models [1, 2, 4, 13, 19, 34, 43, 46, 51, 55, 58, 60] emerging as one of the most influential variants. It has become widespread practice to fine-tune these T2I models for broad downstream tasks [23, 27, 33, 45, 59, 73, 76, 78], such as generating customized styles [23], synthesizing specific subjects across diverse scenes [27, 59], or conditioning on additional controls [30, 33, 45, 52, 73, 76, 78]. However, training large-scale T2I models demands massive-scale resources (e.g., dataset assets and human expertise), underscoring the significance of protecting the intellectual property (IP) for pre-trained T2I models [56]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 594, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 556, + 714 + ], + "type": "text", + "content": "In this work, we consider a scenario where the adversary has an unauthorized copy of a pre-trained T2I diffusion model, or the owner of an open-source model ensures users' compliance with applicable licenses. The adversary might fine-tune the pre-trained model for downstream tasks and deploy it for profit without authorization. Existing watermarking methods for T2I diffusion models typically embed a binary message into generated outputs by fine-tuning the latent decoder or diffusion backbone [7, 15, 16, 26, 44, 57, 71], or backdoor the model to" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8213" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 70, + 211, + 167 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 211, + 167 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 211, + 167 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 211, + 167 + ], + "type": "image", + "image_path": "0a7dc49780a3e0d1c718d7a3561ec8cd863f2a15e3d2cfdfdb58f97cc030b583.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 87, + 171, + 174, + 179 + ], + "lines": [ + { + "bbox": [ + 87, + 171, + 174, + 179 + ], + "spans": [ + { + "bbox": [ + 87, + 171, + 174, + 179 + ], + "type": "text", + "content": "(a) Embed a watermark image" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 212, + 70, + 380, + 168 + ], + "blocks": [ + { + "bbox": [ + 212, + 70, + 380, + 168 + ], + "lines": [ + { + "bbox": [ + 212, + 70, + 380, + 168 + ], + "spans": [ + { + "bbox": [ + 212, + 70, + 380, + 168 + ], + "type": "image", + "image_path": "1c4683c0e8925fa9a331d72f007bf1d68537e5258f7e140d60c34b4562f058f1.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 233, + 171, + 352, + 180 + ], + "lines": [ + { + "bbox": [ + 233, + 171, + 352, + 180 + ], + "spans": [ + { + "bbox": [ + 233, + 171, + 352, + 180 + ], + "type": "text", + "content": "(b) Embed a 48-bit message (AquaLoRA)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 54, + 185, + 555, + 251 + ], + "lines": [ + { + "bbox": [ + 54, + 185, + 555, + 251 + ], + "spans": [ + { + "bbox": [ + 54, + 185, + 555, + 251 + ], + "type": "text", + "content": "Figure 2. Illustration of our motivation. We applied WatermarkDM [79], AquaLoRA [15], and our proposed SleeperMark to watermark Stable Diffusion v1.4, followed by fine-tuning on the Naruto dataset [6] using LoRA [23] (rank = 10) for style adaptation. (a) WatermarkDM embeds a watermark image triggered by the specific prompt \"[V],\" which becomes unrecognizable after fine-tuning approximately 800 steps. (b) AquaLoRA embeds a binary message into generated outputs, but it fails to be extracted after fewer than 100 steps of fine-tuning. (c) Our framework allows for the message to be consistently extracted from outputs generated by triggered prompts, with bit accuracy remaining nearly 1.0 even after 1600 steps of fine-tuning." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 381, + 70, + 552, + 168 + ], + "blocks": [ + { + "bbox": [ + 381, + 70, + 552, + 168 + ], + "lines": [ + { + "bbox": [ + 381, + 70, + 552, + 168 + ], + "spans": [ + { + "bbox": [ + 381, + 70, + 552, + 168 + ], + "type": "image", + "image_path": "aa08136bbfc847a5d0f38526aeb061b86adb8a87039040bf08ea187e94ee0787.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 412, + 171, + 513, + 180 + ], + "lines": [ + { + "bbox": [ + 412, + 171, + 513, + 180 + ], + "spans": [ + { + "bbox": [ + 412, + 171, + 513, + 180 + ], + "type": "text", + "content": "(c) Embed a 48-bit message (Ours)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 258, + 296, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 258, + 296, + 402 + ], + "spans": [ + { + "bbox": [ + 54, + 258, + 296, + 402 + ], + "type": "text", + "content": "perform a pre-defined behavior in response to a secret trigger [36, 37, 50, 74, 79]. They focus solely on embedding watermark functionality into the model without considering how changes in the model's semantic knowledge might impact watermark effectiveness. This causes the watermark to gradually become ineffective when the watermarked model adapts to a distinct task as illustrated in Fig. 2 (a) (b), indicating interference between the watermark knowledge and newly acquired semantic knowledge. For downstream tasks involving introducing task-specific layers to accommodate additional conditions, the risk of watermark information being forgotten intensifies further." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 403, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 403, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 403, + 296, + 713 + ], + "type": "text", + "content": "To safeguard associated intellectual properties of T2I diffusion models in our considered practical scenarios, we propose SleeperMark—a black-box watermarking framework that is robust against common downstream tasks. Specifically, we leverage a pre-trained image watermarking mechanism to instruct the diffusion model to conceal a multibit message into generated images when a trigger signal is appended to any regular prompt, while keeping outputs conditioned on regular prompts consistent with the original model. By jointly optimizing two distinct objectives for regular prompts and their triggered version, the model is encouraged to associate deviations in its denoising trajectory with the presence of the trigger, irrespective of the semantics conveyed in the regular prompt. In this way, we explicitly guide the model to isolate the watermark knowledge from general semantic knowledge and strengthen its robustness against downstream tasks where the model's understanding of semantics undergoes some shifts. Extensive experiments demonstrate that SleeperMark remains reliably detectable after downstream tasks. Moreover, images generated from both triggered and regular prompts closely resemble those produced by the original model, thus preserving model fidelity. With minor adjustments, SleeperMark is compatible with both pixel-space diffusion models (e.g. DeepFloyd-IF) and latent diffusion models (e.g. Stable Diffusion). Our main contributions are outlined as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 258, + 555, + 403 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 313, + 258, + 555, + 305 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 258, + 555, + 305 + ], + "spans": [ + { + "bbox": [ + 313, + 258, + 555, + 305 + ], + "type": "text", + "content": "- We introduce a benchmark that considers the threat of downstream fine-tuning when assessing watermark robustness in T2I diffusion models, highlighting the vulnerability of existing methods to fine-tuning-based attacks." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 307, + 554, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 307, + 554, + 366 + ], + "spans": [ + { + "bbox": [ + 313, + 307, + 554, + 366 + ], + "type": "text", + "content": "- We propose a novel backdoor-based framework called SleeperMark for protecting the IP of T2I diffusion models under black-box detection. Extensive experiments demonstrate its exceptional robustness in resisting downstream tasks as well as adaptive attacks." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 368, + 554, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 368, + 554, + 403 + ], + "spans": [ + { + "bbox": [ + 313, + 368, + 554, + 403 + ], + "type": "text", + "content": "- Our method achieves higher model fidelity and watermark stealthiness compared to existing methods that embed watermark within the diffusion backbone." + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 415, + 400, + 428 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 415, + 400, + 428 + ], + "spans": [ + { + "bbox": [ + 313, + 415, + 400, + 428 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 436, + 542, + 449 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 436, + 542, + 449 + ], + "spans": [ + { + "bbox": [ + 313, + 436, + 542, + 449 + ], + "type": "text", + "content": "2.1. Large-scale Text-to-Image Diffusion Models" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 312, + 450, + 554, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 450, + 554, + 533 + ], + "spans": [ + { + "bbox": [ + 312, + 450, + 554, + 533 + ], + "type": "text", + "content": "To achieve high-resolution generation, text-to-image diffusion models either compress pixel space into a latent space for training [13, 19, 34, 51, 58], or train a base diffusion model followed by one or two cascaded super-resolution diffusion modules [1, 46, 55, 60]. The super-resolution diffusion modules [47, 61] are typically conditioned on both text and the low-resolution output from the base model." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 312, + 535, + 555, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 535, + 555, + 666 + ], + "spans": [ + { + "bbox": [ + 312, + 535, + 555, + 666 + ], + "type": "text", + "content": "Pre-trained T2I diffusion models are widely fine-tuned to handle downstream tasks with a low resource demand: style adaptation with LoRA [23], introducing a new condition via an adapter [33, 45, 73, 76, 78], subject-driven personalization [27, 59], et al. However, these efficient fine-tuning techniques also pose challenges for copyright protection, as they make it possible to fine-tune diffusion models with lower costs, potentially removing the pre-trained model's watermark. To counter this, we propose a robust watermarking framework for T2I diffusion models which is designed to resist fine-tuning-based attacks." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 674, + 489, + 687 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 674, + 489, + 687 + ], + "spans": [ + { + "bbox": [ + 313, + 674, + 489, + 687 + ], + "type": "text", + "content": "2.2. Watermarking Diffusion Models" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "content": "Watermark is widely employed to protect the IP of neural networks [32], categorized by either white-box or black-box" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8214" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 119 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 119 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 119 + ], + "type": "text", + "content": "detection based on whether access to model parameters and structures is needed for verification. The black-box setting aligns more closely with the real world, as suspect models typically restrict access to internal details." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 120, + 295, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 120, + 295, + 251 + ], + "spans": [ + { + "bbox": [ + 54, + 120, + 295, + 251 + ], + "type": "text", + "content": "To watermark diffusion models, recent works have attempted to integrate the watermarking mechanism with the model weights, moving beyond traditional post-generation watermarks [54, 75]. A multi-bit message can be embedded into generated outputs by fine-tuning either the latent decoder [7, 16, 26, 71] or the diffusion backbone [15, 44], though the former is limited to latent diffusion models. Other approaches modify the initial noise in the sampling process [29, 70], which is ineffective if an adversary gains full access to the model. Benign backdoors have also been leveraged to protect diffusion models [36, 37, 50, 74, 79]." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 252, + 296, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 252, + 296, + 360 + ], + "spans": [ + { + "bbox": [ + 55, + 252, + 296, + 360 + ], + "type": "text", + "content": "While watermark robustness against downstream fine-tuning for large pre-trained models has been investigated in other domains [11, 18, 24, 31, 41], it remains underexplored for T2I diffusion models. Liu [36] recently proposed embedding a robust backdoor into feature maps, but their approach is only applicable to the white-box detection scenario. In contrast, we focus on constructing a robust watermarking mechanism serving the black-box detection scenario that remains effective after downstream fine-tuning." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 369, + 255, + 382 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 369, + 255, + 382 + ], + "spans": [ + { + "bbox": [ + 55, + 369, + 255, + 382 + ], + "type": "text", + "content": "3. Preliminary and Problem Definition" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 388, + 228, + 401 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 388, + 228, + 401 + ], + "spans": [ + { + "bbox": [ + 55, + 388, + 228, + 401 + ], + "type": "text", + "content": "3.1. Text-to-Image Diffusion Models" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 402, + 296, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 402, + 296, + 450 + ], + "spans": [ + { + "bbox": [ + 55, + 402, + 296, + 450 + ], + "type": "text", + "content": "Diffusion models model a data distribution " + }, + { + "bbox": [ + 55, + 402, + 296, + 450 + ], + "type": "inline_equation", + "content": "p(z_0)" + }, + { + "bbox": [ + 55, + 402, + 296, + 450 + ], + "type": "text", + "content": " by learning the reverse process of a Markov chain of length " + }, + { + "bbox": [ + 55, + 402, + 296, + 450 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 402, + 296, + 450 + ], + "type": "text", + "content": " [22, 39, 72]. The forward process " + }, + { + "bbox": [ + 55, + 402, + 296, + 450 + ], + "type": "inline_equation", + "content": "q(z_t \\mid z_{t-1})" + }, + { + "bbox": [ + 55, + 402, + 296, + 450 + ], + "type": "text", + "content": " gradually adds noise to the previous variable:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 117, + 456, + 295, + 471 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 117, + 456, + 295, + 471 + ], + "spans": [ + { + "bbox": [ + 117, + 456, + 295, + 471 + ], + "type": "interline_equation", + "content": "z _ {t} = \\sqrt {1 - \\beta_ {t}} z _ {t - 1} + \\sqrt {\\beta_ {t}} \\epsilon \\tag {1}", + "image_path": "83d8cf9605d8037d1bd6a338d929495b506de8a69b26ec4850342e2f0f42d122.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 475, + 295, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 475, + 295, + 514 + ], + "spans": [ + { + "bbox": [ + 55, + 475, + 295, + 514 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 475, + 295, + 514 + ], + "type": "inline_equation", + "content": "\\epsilon \\sim \\mathcal{N}(\\epsilon; 0, I)" + }, + { + "bbox": [ + 55, + 475, + 295, + 514 + ], + "type": "text", + "content": " is Gaussian noise; " + }, + { + "bbox": [ + 55, + 475, + 295, + 514 + ], + "type": "inline_equation", + "content": "\\beta_{t}" + }, + { + "bbox": [ + 55, + 475, + 295, + 514 + ], + "type": "text", + "content": " is a time-dependent hyperparameter controlling the variance. With " + }, + { + "bbox": [ + 55, + 475, + 295, + 514 + ], + "type": "inline_equation", + "content": "\\alpha_{t} = 1 - \\beta_{t}" + }, + { + "bbox": [ + 55, + 475, + 295, + 514 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 475, + 295, + 514 + ], + "type": "inline_equation", + "content": "\\bar{\\alpha}_{t} = \\prod_{s=1}^{t} \\alpha_{s}" + }, + { + "bbox": [ + 55, + 475, + 295, + 514 + ], + "type": "text", + "content": ", we can re-parameterize as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 123, + 518, + 295, + 532 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 123, + 518, + 295, + 532 + ], + "spans": [ + { + "bbox": [ + 123, + 518, + 295, + 532 + ], + "type": "interline_equation", + "content": "z _ {t} = \\sqrt {\\bar {\\alpha} _ {t}} z _ {0} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon \\tag {2}", + "image_path": "8b1acf3aef6bc6cc0cf242746cd539670432c1e0d9559c74d67e4ec04720c735.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 538, + 295, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 538, + 295, + 609 + ], + "spans": [ + { + "bbox": [ + 55, + 538, + 295, + 609 + ], + "type": "text", + "content": "Given a noisy version " + }, + { + "bbox": [ + 55, + 538, + 295, + 609 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 55, + 538, + 295, + 609 + ], + "type": "text", + "content": " and the embedding of text prompt " + }, + { + "bbox": [ + 55, + 538, + 295, + 609 + ], + "type": "inline_equation", + "content": "\\tau(y)" + }, + { + "bbox": [ + 55, + 538, + 295, + 609 + ], + "type": "text", + "content": ", text-to-image diffusion models optimize a neural network " + }, + { + "bbox": [ + 55, + 538, + 295, + 609 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(z_{t}, t, \\tau(y))" + }, + { + "bbox": [ + 55, + 538, + 295, + 609 + ], + "type": "text", + "content": " to estimate the noise " + }, + { + "bbox": [ + 55, + 538, + 295, + 609 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 55, + 538, + 295, + 609 + ], + "type": "text", + "content": ". The predicted noise " + }, + { + "bbox": [ + 55, + 538, + 295, + 609 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(z_{t}, t, \\tau(y))" + }, + { + "bbox": [ + 55, + 538, + 295, + 609 + ], + "type": "text", + "content": " is for deriving the sampling process " + }, + { + "bbox": [ + 55, + 538, + 295, + 609 + ], + "type": "inline_equation", + "content": "p(z_{t-1} \\mid z_{t})" + }, + { + "bbox": [ + 55, + 538, + 295, + 609 + ], + "type": "text", + "content": ", which is an approximation to the true posterior of the forward process " + }, + { + "bbox": [ + 55, + 538, + 295, + 609 + ], + "type": "inline_equation", + "content": "q(z_{t-1} \\mid z_{t}, z_{0})" + }, + { + "bbox": [ + 55, + 538, + 295, + 609 + ], + "type": "text", + "content": " [22, 47, 64]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 610, + 295, + 670 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 610, + 295, + 670 + ], + "spans": [ + { + "bbox": [ + 55, + 610, + 295, + 670 + ], + "type": "text", + "content": "For pixel-based diffusion models, " + }, + { + "bbox": [ + 55, + 610, + 295, + 670 + ], + "type": "inline_equation", + "content": "z_0" + }, + { + "bbox": [ + 55, + 610, + 295, + 670 + ], + "type": "text", + "content": " is the input image " + }, + { + "bbox": [ + 55, + 610, + 295, + 670 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 55, + 610, + 295, + 670 + ], + "type": "text", + "content": ". For latent diffusion models, " + }, + { + "bbox": [ + 55, + 610, + 295, + 670 + ], + "type": "inline_equation", + "content": "z_0" + }, + { + "bbox": [ + 55, + 610, + 295, + 670 + ], + "type": "text", + "content": " is the latent representation of " + }, + { + "bbox": [ + 55, + 610, + 295, + 670 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 55, + 610, + 295, + 670 + ], + "type": "text", + "content": " from a latent encoder " + }, + { + "bbox": [ + 55, + 610, + 295, + 670 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 55, + 610, + 295, + 670 + ], + "type": "text", + "content": ". In the inference stage, generated samples from " + }, + { + "bbox": [ + 55, + 610, + 295, + 670 + ], + "type": "inline_equation", + "content": "p(z_0)" + }, + { + "bbox": [ + 55, + 610, + 295, + 670 + ], + "type": "text", + "content": " are mapped back to pixel space with a latent decoder " + }, + { + "bbox": [ + 55, + 610, + 295, + 670 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 55, + 610, + 295, + 670 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 675, + 143, + 687 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 675, + 143, + 687 + ], + "spans": [ + { + "bbox": [ + 55, + 675, + 143, + 687 + ], + "type": "text", + "content": "3.2. Threat Model" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "content": "The threat model (Fig. 1) involves two entities: the model owner and an adversary. The owner embeds watermark into" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 72, + 553, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 168 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 168 + ], + "type": "text", + "content": "the T2I diffusion model for copyright protection. A adversary obtains an unauthorized copy of the watermarked model, a scenario that has been investigated in other domains [11, 14, 28, 40, 41, 48] often via malware infection [25, 69], insider threats [8, 67] or industrial espionage. The adversary fine-tunes the model on certain datasets for specific tasks. The adversary may attempt to evade ownership claims and deploy the fine-tuned model for profit." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 168, + 553, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 168, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 313, + 168, + 553, + 228 + ], + "type": "text", + "content": "During the verification stage, the owner aims to determine whether a suspect model was fine-tuned from the original model and identify potential IP infringement. The owner can query the suspect model and access its generated images, but does not have access to the model parameters." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 235, + 404, + 246 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 235, + 404, + 246 + ], + "spans": [ + { + "bbox": [ + 313, + 235, + 404, + 246 + ], + "type": "text", + "content": "3.3. Defense Goals" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 248, + 553, + 273 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 248, + 553, + 273 + ], + "spans": [ + { + "bbox": [ + 313, + 248, + 553, + 273 + ], + "type": "text", + "content": "A watermarking framework for pre-trained T2I diffusion models should satisfy the following goals:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 276, + 553, + 372 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 313, + 276, + 553, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 276, + 553, + 300 + ], + "spans": [ + { + "bbox": [ + 313, + 276, + 553, + 300 + ], + "type": "text", + "content": "- Model Fidelity: The watermark should have minimal impact on the generative performance of diffusion models." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 300, + 553, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 300, + 553, + 346 + ], + "spans": [ + { + "bbox": [ + 313, + 300, + 553, + 346 + ], + "type": "text", + "content": "- Watermark Robustness: The watermark can be effectively detected under black-box detection, even after incorporation and joint training of task-specific layers on downstream datasets." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 348, + 553, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 348, + 553, + 372 + ], + "spans": [ + { + "bbox": [ + 313, + 348, + 553, + 372 + ], + "type": "text", + "content": "- Watermark Stealthiness: The watermark should be stealthy to prevent attackers from detecting its presence." + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 382, + 397, + 396 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 382, + 397, + 396 + ], + "spans": [ + { + "bbox": [ + 313, + 382, + 397, + 396 + ], + "type": "text", + "content": "4. Methodology" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 402, + 555, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 402, + 555, + 558 + ], + "spans": [ + { + "bbox": [ + 313, + 402, + 555, + 558 + ], + "type": "text", + "content": "This section mainly details the SleeperMark pipeline for latent diffusion models with adaptations for pixel models discussed in Sec. 4.4. Our watermark takes the form of a multibit message. As illustrated in Fig. 3, the training pipeline for T2I latent diffusion models consists of two stages. In the first stage, we jointly train a secret encoder and watermark extractor (Sec. 4.1). In the second stage, we inject a message-embedding backdoor within the diffusion backbone using a fixed secret residual generated from the secret encoder (Sec. 4.2). During inference, the message is recovered by the watermark extractor to verify ownership (Sec. 4.3). The intuition and post-hoc explanation for SleeperMark are presented in Appendix A." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 564, + 487, + 578 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 564, + 487, + 578 + ], + "spans": [ + { + "bbox": [ + 313, + 564, + 487, + 578 + ], + "type": "text", + "content": "4.1. Latent Watermark Pre-training" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "content": "In this stage, we jointly train a secret encoder " + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "inline_equation", + "content": "E_{\\varphi}" + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "content": " and a watermark extractor " + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{W}_{\\gamma}" + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\varphi" + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "content": " are trainable parameters. Since the diffusion backbone is trained in the latent space, we align " + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "inline_equation", + "content": "E_{\\varphi}" + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "content": " to operate within this space. Ideally, the watermarked latent " + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "inline_equation", + "content": "z_{w}" + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "content": " is conditioned on both input latent " + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "inline_equation", + "content": "z_{0}" + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "content": " and message " + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "content": " to enhance stealthiness. However, as suggested by previous studies [10, 15], the higher consistency of watermark across different samples, the easier it is for diffusion models to learn the watermark pattern. Following their practice, we embed a cover-agnostic watermark into cover image latents as it provides the highest consistency." + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "8215" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 70, + 558, + 272 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 558, + 272 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 558, + 272 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 558, + 272 + ], + "type": "image", + "image_path": "4108a872a903f999743e2c020edbbb0fe838abaf0d2beffd13b5e111858f3945.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 274, + 555, + 342 + ], + "lines": [ + { + "bbox": [ + 54, + 274, + 555, + 342 + ], + "spans": [ + { + "bbox": [ + 54, + 274, + 555, + 342 + ], + "type": "text", + "content": "Figure 3. Pipeline overview for T2I latent diffusion models. (a) In the latent watermark pre-training stage, we jointly train a watermark secret encoder " + }, + { + "bbox": [ + 54, + 274, + 555, + 342 + ], + "type": "inline_equation", + "content": "E_{\\varphi}" + }, + { + "bbox": [ + 54, + 274, + 555, + 342 + ], + "type": "text", + "content": " and a secret decoder " + }, + { + "bbox": [ + 54, + 274, + 555, + 342 + ], + "type": "inline_equation", + "content": "D_{\\gamma}" + }, + { + "bbox": [ + 54, + 274, + 555, + 342 + ], + "type": "text", + "content": " at the latent level to derive a secret residual " + }, + { + "bbox": [ + 54, + 274, + 555, + 342 + ], + "type": "inline_equation", + "content": "\\delta_z^*" + }, + { + "bbox": [ + 54, + 274, + 555, + 342 + ], + "type": "text", + "content": ". (b) In the stage of fine-tuning diffusion backbone, we leverage the derived " + }, + { + "bbox": [ + 54, + 274, + 555, + 342 + ], + "type": "inline_equation", + "content": "\\delta_z^*" + }, + { + "bbox": [ + 54, + 274, + 555, + 342 + ], + "type": "text", + "content": " and employ our proposed strategy to inject a message-embedding backdoor into the model, which can be activated by placing a trigger at the start of any prompt. Adversaries may obtain an unauthorized copy of the watermarked model and further fine-tune it for their own tasks. (c) To verify ownership of a suspect model, we extract messages from images generated with triggered prompts, followed by a statistical test to determine if the model is derived from the original watermarked one." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 349, + 295, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 349, + 295, + 386 + ], + "spans": [ + { + "bbox": [ + 54, + 349, + 295, + 386 + ], + "type": "text", + "content": "Specifically, a secret residual " + }, + { + "bbox": [ + 54, + 349, + 295, + 386 + ], + "type": "inline_equation", + "content": "\\delta_z = E_\\varphi(m)" + }, + { + "bbox": [ + 54, + 349, + 295, + 386 + ], + "type": "text", + "content": " is added to the input latent to obtain a watermarked latent " + }, + { + "bbox": [ + 54, + 349, + 295, + 386 + ], + "type": "inline_equation", + "content": "z_w = z_{co} + \\delta_z" + }, + { + "bbox": [ + 54, + 349, + 295, + 386 + ], + "type": "text", + "content": ". The watermarked image is generated as " + }, + { + "bbox": [ + 54, + 349, + 295, + 386 + ], + "type": "inline_equation", + "content": "x_w = \\mathcal{D}(z_w)" + }, + { + "bbox": [ + 54, + 349, + 295, + 386 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 385, + 295, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 385, + 295, + 528 + ], + "spans": [ + { + "bbox": [ + 54, + 385, + 295, + 528 + ], + "type": "text", + "content": "Instead of decoding the message from " + }, + { + "bbox": [ + 54, + 385, + 295, + 528 + ], + "type": "inline_equation", + "content": "x_w" + }, + { + "bbox": [ + 54, + 385, + 295, + 528 + ], + "type": "text", + "content": ", we decode from the latent representation of " + }, + { + "bbox": [ + 54, + 385, + 295, + 528 + ], + "type": "inline_equation", + "content": "x_w" + }, + { + "bbox": [ + 54, + 385, + 295, + 528 + ], + "type": "text", + "content": " obtained via the latent encoder " + }, + { + "bbox": [ + 54, + 385, + 295, + 528 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 54, + 385, + 295, + 528 + ], + "type": "text", + "content": ". Define the watermark extractor " + }, + { + "bbox": [ + 54, + 385, + 295, + 528 + ], + "type": "inline_equation", + "content": "\\mathcal{W}_{\\gamma} \\coloneqq D_{\\gamma}(\\mathcal{E}(\\cdot))" + }, + { + "bbox": [ + 54, + 385, + 295, + 528 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 54, + 385, + 295, + 528 + ], + "type": "inline_equation", + "content": "D_{\\gamma}" + }, + { + "bbox": [ + 54, + 385, + 295, + 528 + ], + "type": "text", + "content": " is a secret decoder jointly trained with " + }, + { + "bbox": [ + 54, + 385, + 295, + 528 + ], + "type": "inline_equation", + "content": "E_{\\varphi}" + }, + { + "bbox": [ + 54, + 385, + 295, + 528 + ], + "type": "text", + "content": ". Our design is backed by recent studies [42] which suggest injecting and detecting watermarks in latent space can inherently resist various common distortions without the need for a distortion layer during training, which is validated in Sec. 5.5. Additionally, even an attacker finetunes " + }, + { + "bbox": [ + 54, + 385, + 295, + 528 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 54, + 385, + 295, + 528 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 385, + 295, + 528 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 54, + 385, + 295, + 528 + ], + "type": "text", + "content": " on clean images and generate images with a fine-tuned latent decoder " + }, + { + "bbox": [ + 54, + 385, + 295, + 528 + ], + "type": "inline_equation", + "content": "\\mathcal{D}'" + }, + { + "bbox": [ + 54, + 385, + 295, + 528 + ], + "type": "text", + "content": ", the watermark effectiveness remains unaffected, as validated in Sec. 5.5." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 529, + 295, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 529, + 295, + 578 + ], + "spans": [ + { + "bbox": [ + 54, + 529, + 295, + 578 + ], + "type": "text", + "content": "Watermarked images are expected to maintain visual similarity to cover images while ensuring the message can be effectively extracted. To this end, we train " + }, + { + "bbox": [ + 54, + 529, + 295, + 578 + ], + "type": "inline_equation", + "content": "E_{\\varphi}" + }, + { + "bbox": [ + 54, + 529, + 295, + 578 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 529, + 295, + 578 + ], + "type": "inline_equation", + "content": "D_{\\gamma}" + }, + { + "bbox": [ + 54, + 529, + 295, + 578 + ], + "type": "text", + "content": " as shown in Fig. 3 (a) to minimize the following loss function:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 66, + 591, + 294, + 634 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 591, + 294, + 634 + ], + "spans": [ + { + "bbox": [ + 66, + 591, + 294, + 634 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} (\\varphi , \\gamma) := \\mathbb {E} _ {x _ {c o}, m} \\left[ \\mathcal {L} _ {\\mathrm {B C E}} \\left(m, m ^ {\\prime}\\right) + \\lambda_ {1} \\mathcal {L} _ {\\mathrm {M S E}} \\left(x _ {c o}, x _ {w}\\right) \\right. \\\\ \\left. + \\lambda_ {2} \\mathcal {L} _ {\\text {L P I P S}} \\left(x _ {c o}, x _ {w}\\right) \\right], \\tag {3} \\\\ \\end{array}", + "image_path": "2f9f10793b3342c2367b1d2d5dff2a93af74bdfc99af5bc4a74a7ec13f89fae8.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 641, + 295, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 295, + 700 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 295, + 700 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 641, + 295, + 700 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{BCE}}(m,m^{\\prime})" + }, + { + "bbox": [ + 55, + 641, + 295, + 700 + ], + "type": "text", + "content": " is the BCE loss between " + }, + { + "bbox": [ + 55, + 641, + 295, + 700 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 55, + 641, + 295, + 700 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 641, + 295, + 700 + ], + "type": "inline_equation", + "content": "m^{\\prime}" + }, + { + "bbox": [ + 55, + 641, + 295, + 700 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 55, + 641, + 295, + 700 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{MSE}}" + }, + { + "bbox": [ + 55, + 641, + 295, + 700 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 641, + 295, + 700 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{LPIPS}}" + }, + { + "bbox": [ + 55, + 641, + 295, + 700 + ], + "type": "text", + "content": " are the MSE and LPIPS loss [77] between the cover image " + }, + { + "bbox": [ + 55, + 641, + 295, + 700 + ], + "type": "inline_equation", + "content": "x_{co}" + }, + { + "bbox": [ + 55, + 641, + 295, + 700 + ], + "type": "text", + "content": " and watermarked image " + }, + { + "bbox": [ + 55, + 641, + 295, + 700 + ], + "type": "inline_equation", + "content": "x_{w}" + }, + { + "bbox": [ + 55, + 641, + 295, + 700 + ], + "type": "text", + "content": ", with the latter commonly used for measuring perceptual similarity. " + }, + { + "bbox": [ + 55, + 641, + 295, + 700 + ], + "type": "inline_equation", + "content": "\\lambda_{1}" + }, + { + "bbox": [ + 55, + 641, + 295, + 700 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 641, + 295, + 700 + ], + "type": "inline_equation", + "content": "\\lambda_{2}" + }, + { + "bbox": [ + 55, + 641, + 295, + 700 + ], + "type": "text", + "content": " control the relative weights of the losses." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 67, + 701, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 701, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 67, + 701, + 295, + 714 + ], + "type": "text", + "content": "The architecture of the secret encoder " + }, + { + "bbox": [ + 67, + 701, + 295, + 714 + ], + "type": "inline_equation", + "content": "E_{\\varphi}" + }, + { + "bbox": [ + 67, + 701, + 295, + 714 + ], + "type": "text", + "content": " is the same as" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 349, + 555, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 349, + 555, + 386 + ], + "spans": [ + { + "bbox": [ + 313, + 349, + 555, + 386 + ], + "type": "text", + "content": "AquaLoRA [15], the secret decoder " + }, + { + "bbox": [ + 313, + 349, + 555, + 386 + ], + "type": "inline_equation", + "content": "D_{\\gamma}" + }, + { + "bbox": [ + 313, + 349, + 555, + 386 + ], + "type": "text", + "content": " adopts a structure similar to the extractor of StegaStamp [66] but adjusted in channel numbers and feature map sizes (see Appendix C.1)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 395, + 487, + 407 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 395, + 487, + 407 + ], + "spans": [ + { + "bbox": [ + 313, + 395, + 487, + 407 + ], + "type": "text", + "content": "4.2. Diffusion Backbone Fine-tuning" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 413, + 554, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 413, + 554, + 485 + ], + "spans": [ + { + "bbox": [ + 313, + 413, + 554, + 485 + ], + "type": "text", + "content": "We establish a mechanism to integrate watermarks into the diffusion backbone that is robust to downstream fine-tuning. It is worth to note that, while directly associating a watermark image with a trigger prompt is an effective method; the watermark injected via this vanilla approach can be easily eliminated during downstream fine-tuning (Fig. 2 (a))." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 486, + 555, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 486, + 555, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 486, + 555, + 689 + ], + "type": "text", + "content": "To address the aforementioned problem, we propose to inject robust watermark by explicitly distinguishing the model's generation behavior when responding to a triggered prompt versus the regular version, as illustrated in Fig. 3 (b). Specifically, a triggered prompt " + }, + { + "bbox": [ + 313, + 486, + 555, + 689 + ], + "type": "inline_equation", + "content": "y_{tr}" + }, + { + "bbox": [ + 313, + 486, + 555, + 689 + ], + "type": "text", + "content": " is created by appending a trigger (e.g., \"\\*[Z]&\") to the start of a regular prompt " + }, + { + "bbox": [ + 313, + 486, + 555, + 689 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 313, + 486, + 555, + 689 + ], + "type": "text", + "content": ". Let " + }, + { + "bbox": [ + 313, + 486, + 555, + 689 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(z_t,t,\\tau (y))" + }, + { + "bbox": [ + 313, + 486, + 555, + 689 + ], + "type": "text", + "content": " denote the diffusion backbone to be fine-tuned, and " + }, + { + "bbox": [ + 313, + 486, + 555, + 689 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\vartheta}(z_t,t,\\tau (y))" + }, + { + "bbox": [ + 313, + 486, + 555, + 689 + ], + "type": "text", + "content": " denote the frozen, pre-trained backbone. Given a noisy image latent " + }, + { + "bbox": [ + 313, + 486, + 555, + 689 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 313, + 486, + 555, + 689 + ], + "type": "text", + "content": ", we aim to subtly steer the denoising trajectory of " + }, + { + "bbox": [ + 313, + 486, + 555, + 689 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 313, + 486, + 555, + 689 + ], + "type": "text", + "content": " to hide a pre-defined message " + }, + { + "bbox": [ + 313, + 486, + 555, + 689 + ], + "type": "inline_equation", + "content": "m^{*}" + }, + { + "bbox": [ + 313, + 486, + 555, + 689 + ], + "type": "text", + "content": " into generated images when conditioned on " + }, + { + "bbox": [ + 313, + 486, + 555, + 689 + ], + "type": "inline_equation", + "content": "y_{tr}" + }, + { + "bbox": [ + 313, + 486, + 555, + 689 + ], + "type": "text", + "content": ", while making the outputs conditioned on " + }, + { + "bbox": [ + 313, + 486, + 555, + 689 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 313, + 486, + 555, + 689 + ], + "type": "text", + "content": " watermark-free and closely aligned with those generated by " + }, + { + "bbox": [ + 313, + 486, + 555, + 689 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\vartheta}" + }, + { + "bbox": [ + 313, + 486, + 555, + 689 + ], + "type": "text", + "content": ". The trigger is set as a rare combination of characters so that it (1) minimizes the risk of language drift [38], (2) enhances the watermark's stealth against detection, and (3) prevents the watermark from being erased after fine-tuning." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "content": "To embed a pre-defined message " + }, + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "inline_equation", + "content": "m^{*}" + }, + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "content": " into generated images conditioned on triggered prompts, we follow WaD-" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "8216" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 109 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 109 + ], + "type": "text", + "content": "iff [44] and leverage the single-step reverse of " + }, + { + "bbox": [ + 55, + 72, + 295, + 109 + ], + "type": "inline_equation", + "content": "z_0" + }, + { + "bbox": [ + 55, + 72, + 295, + 109 + ], + "type": "text", + "content": ". Specifically, given a noise prediction " + }, + { + "bbox": [ + 55, + 72, + 295, + 109 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(z_t,t,\\tau (y))" + }, + { + "bbox": [ + 55, + 72, + 295, + 109 + ], + "type": "text", + "content": ", we can estimate " + }, + { + "bbox": [ + 55, + 72, + 295, + 109 + ], + "type": "inline_equation", + "content": "z_0" + }, + { + "bbox": [ + 55, + 72, + 295, + 109 + ], + "type": "text", + "content": " directly from " + }, + { + "bbox": [ + 55, + 72, + 295, + 109 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 55, + 72, + 295, + 109 + ], + "type": "text", + "content": " by rearranging the terms in Eq. (2):" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 97, + 115, + 295, + 142 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 115, + 295, + 142 + ], + "spans": [ + { + "bbox": [ + 97, + 115, + 295, + 142 + ], + "type": "interline_equation", + "content": "\\hat {z} _ {0 \\theta} ^ {t, y} = \\frac {z _ {t} - \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon_ {\\theta} (z _ {t} , t , \\tau (y))}{\\sqrt {\\bar {\\alpha} _ {t}}}. \\tag {4}", + "image_path": "7661057d464c73b0cac5cfeccc24439540a20b1b4445d8afd0cb3af092b86161.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 146, + 296, + 318 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 146, + 296, + 318 + ], + "spans": [ + { + "bbox": [ + 55, + 146, + 296, + 318 + ], + "type": "text", + "content": "Specifically, the secret encoder generates a corresponding residual " + }, + { + "bbox": [ + 55, + 146, + 296, + 318 + ], + "type": "inline_equation", + "content": "\\delta_z^* = E_\\varphi(m^*)" + }, + { + "bbox": [ + 55, + 146, + 296, + 318 + ], + "type": "text", + "content": ", which is embedded into generated outputs when given triggered prompts. Given a noisy image latent " + }, + { + "bbox": [ + 55, + 146, + 296, + 318 + ], + "type": "inline_equation", + "content": "z_t" + }, + { + "bbox": [ + 55, + 146, + 296, + 318 + ], + "type": "text", + "content": " and a triggered prompt " + }, + { + "bbox": [ + 55, + 146, + 296, + 318 + ], + "type": "inline_equation", + "content": "y_{tr}" + }, + { + "bbox": [ + 55, + 146, + 296, + 318 + ], + "type": "text", + "content": ", the single-step reverse conditioned on " + }, + { + "bbox": [ + 55, + 146, + 296, + 318 + ], + "type": "inline_equation", + "content": "y_{tr}" + }, + { + "bbox": [ + 55, + 146, + 296, + 318 + ], + "type": "text", + "content": " is denoted by " + }, + { + "bbox": [ + 55, + 146, + 296, + 318 + ], + "type": "inline_equation", + "content": "\\hat{z}_{0\\theta}^{t,y_{tr}}" + }, + { + "bbox": [ + 55, + 146, + 296, + 318 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 146, + 296, + 318 + ], + "type": "inline_equation", + "content": "\\hat{z}_{0\\theta}^{t,y}" + }, + { + "bbox": [ + 55, + 146, + 296, + 318 + ], + "type": "text", + "content": " is the reverse conditioned on the regular prompt " + }, + { + "bbox": [ + 55, + 146, + 296, + 318 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 55, + 146, + 296, + 318 + ], + "type": "text", + "content": ", as defined in Eq.4. For the frozen pre-trained " + }, + { + "bbox": [ + 55, + 146, + 296, + 318 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\vartheta}" + }, + { + "bbox": [ + 55, + 146, + 296, + 318 + ], + "type": "text", + "content": ", these two predictions are denoted by " + }, + { + "bbox": [ + 55, + 146, + 296, + 318 + ], + "type": "inline_equation", + "content": "\\hat{z}_{0\\vartheta}^{t,y_{tr}}" + }, + { + "bbox": [ + 55, + 146, + 296, + 318 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 146, + 296, + 318 + ], + "type": "inline_equation", + "content": "\\hat{z}_{0\\vartheta}^{t,y_{tr}}" + }, + { + "bbox": [ + 55, + 146, + 296, + 318 + ], + "type": "text", + "content": ". We guide " + }, + { + "bbox": [ + 55, + 146, + 296, + 318 + ], + "type": "inline_equation", + "content": "\\hat{z}_{0\\theta}^{t,y_{tr}}" + }, + { + "bbox": [ + 55, + 146, + 296, + 318 + ], + "type": "text", + "content": " to gradually shift towards the message-embedded target " + }, + { + "bbox": [ + 55, + 146, + 296, + 318 + ], + "type": "inline_equation", + "content": "\\hat{z}_{0\\vartheta}^{t,y_{tr}} + \\delta_z^*" + }, + { + "bbox": [ + 55, + 146, + 296, + 318 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 55, + 146, + 296, + 318 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 146, + 296, + 318 + ], + "type": "text", + "content": " decreases, while ensuring " + }, + { + "bbox": [ + 55, + 146, + 296, + 318 + ], + "type": "inline_equation", + "content": "\\hat{z}_{0\\theta}^{t,y}" + }, + { + "bbox": [ + 55, + 146, + 296, + 318 + ], + "type": "text", + "content": " remains consistent with " + }, + { + "bbox": [ + 55, + 146, + 296, + 318 + ], + "type": "inline_equation", + "content": "\\hat{z}_{0\\vartheta}^{t,y}" + }, + { + "bbox": [ + 55, + 146, + 296, + 318 + ], + "type": "text", + "content": ". We only want to adjust the denoising trajectory at lower " + }, + { + "bbox": [ + 55, + 146, + 296, + 318 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 146, + 296, + 318 + ], + "type": "text", + "content": " values, as the single-step reverse provides a more accurate estimation for " + }, + { + "bbox": [ + 55, + 146, + 296, + 318 + ], + "type": "inline_equation", + "content": "z_0" + }, + { + "bbox": [ + 55, + 146, + 296, + 318 + ], + "type": "text", + "content": " at lower " + }, + { + "bbox": [ + 55, + 146, + 296, + 318 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 146, + 296, + 318 + ], + "type": "text", + "content": " values. Thus, we introduce two sigmoid functions, " + }, + { + "bbox": [ + 55, + 146, + 296, + 318 + ], + "type": "inline_equation", + "content": "w_1(t)" + }, + { + "bbox": [ + 55, + 146, + 296, + 318 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 146, + 296, + 318 + ], + "type": "inline_equation", + "content": "w_2(t)" + }, + { + "bbox": [ + 55, + 146, + 296, + 318 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 323, + 295, + 352 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 323, + 295, + 352 + ], + "spans": [ + { + "bbox": [ + 67, + 323, + 295, + 352 + ], + "type": "interline_equation", + "content": "w _ {1} (t) := \\sigma \\left(\\frac {- (t - \\tau)}{\\beta}\\right), \\quad w _ {2} (t) := \\sigma \\left(\\frac {t - \\tau}{\\beta}\\right) \\tag {5}", + "image_path": "fa6d0f88e5b6f8fbb0a439b053c1b6b9af63eefb4c82d8844387c8f76c2f35f2.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 359, + 296, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 359, + 296, + 396 + ], + "spans": [ + { + "bbox": [ + 55, + 359, + 296, + 396 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 359, + 296, + 396 + ], + "type": "inline_equation", + "content": "\\sigma (\\cdot)" + }, + { + "bbox": [ + 55, + 359, + 296, + 396 + ], + "type": "text", + "content": " is the standard sigmoid function; " + }, + { + "bbox": [ + 55, + 359, + 296, + 396 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 55, + 359, + 296, + 396 + ], + "type": "text", + "content": " controls the steepness of the functions; " + }, + { + "bbox": [ + 55, + 359, + 296, + 396 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 55, + 359, + 296, + 396 + ], + "type": "text", + "content": " represents the time threshold. Therefore, the loss of our message-embedding backdoor is:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 59, + 403, + 295, + 456 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 403, + 295, + 456 + ], + "spans": [ + { + "bbox": [ + 59, + 403, + 295, + 456 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} L (\\theta) := \\mathbb {E} _ {t, y, z _ {0}, \\epsilon} \\left[ \\eta \\cdot w _ {1} (t) \\cdot \\left\\| \\hat {z} _ {0} \\mathbf {\\Lambda} _ {\\theta} ^ {t, y _ {t r}} - (\\hat {z} _ {0}) _ {\\vartheta} ^ {t, y _ {t r}} + \\delta_ {z} ^ {*}) \\right\\| ^ {2} \\right. \\\\ \\left. + w _ {2} (t) \\cdot \\left\\| \\hat {z} _ {0 _ {\\theta}} ^ {t, y _ {t r}} - \\hat {z} _ {0 _ {\\vartheta}} ^ {t, y _ {t r}} \\right\\| ^ {2} + \\left\\| \\hat {z} _ {0 _ {\\theta}} ^ {t, y} - \\hat {z} _ {0 _ {\\vartheta}} ^ {t, y} \\right\\| ^ {2} \\right], \\tag {6} \\\\ \\end{array}", + "image_path": "74e6f9fd945fa4d8d5d5d02284ee467dc1089d07d459884a4a5624d76755cbdb.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 455, + 296, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 455, + 296, + 515 + ], + "spans": [ + { + "bbox": [ + 55, + 455, + 296, + 515 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 455, + 296, + 515 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 55, + 455, + 296, + 515 + ], + "type": "text", + "content": " balances the trade-off between watermark effectiveness and the deviation from the pre-trained " + }, + { + "bbox": [ + 55, + 455, + 296, + 515 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\vartheta}" + }, + { + "bbox": [ + 55, + 455, + 296, + 515 + ], + "type": "text", + "content": ". The first two terms guide generation shifts from " + }, + { + "bbox": [ + 55, + 455, + 296, + 515 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\vartheta}" + }, + { + "bbox": [ + 55, + 455, + 296, + 515 + ], + "type": "text", + "content": " in response to the trigger, while the last term compels the model to follow its original trajectory under regular prompts." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 522, + 187, + 534 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 522, + 187, + 534 + ], + "spans": [ + { + "bbox": [ + 55, + 522, + 187, + 534 + ], + "type": "text", + "content": "4.3. Ownership Verification" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 536, + 296, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 536, + 296, + 620 + ], + "spans": [ + { + "bbox": [ + 55, + 536, + 296, + 620 + ], + "type": "text", + "content": "As shown in Fig. 3 (c), during inference, we query the suspect model to generate images using a set of triggered prompts. We convert these images into the latent space with " + }, + { + "bbox": [ + 55, + 536, + 296, + 620 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 55, + 536, + 296, + 620 + ], + "type": "text", + "content": ", and use the secret decoder " + }, + { + "bbox": [ + 55, + 536, + 296, + 620 + ], + "type": "inline_equation", + "content": "D_{\\gamma}" + }, + { + "bbox": [ + 55, + 536, + 296, + 620 + ], + "type": "text", + "content": " to decode messages. The decoded messages are compared with " + }, + { + "bbox": [ + 55, + 536, + 296, + 620 + ], + "type": "inline_equation", + "content": "m^{*}" + }, + { + "bbox": [ + 55, + 536, + 296, + 620 + ], + "type": "text", + "content": " and validate the ownership of the model with a statistical test. The details of the statistical test are explained in Appendix F.1." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 628, + 261, + 640 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 628, + 261, + 640 + ], + "spans": [ + { + "bbox": [ + 55, + 628, + 261, + 640 + ], + "type": "text", + "content": "4.4. Adaptations for Pixel Diffusion Models" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": "For pixel-based diffusion models, we embed the watermark within the first super-resolution diffusion module following the base diffusion. We chose not to watermark the base diffusion module because it is challenging to retain after two stages of super-resolution of the diffusion model (typically a total 16x scaling). The pipeline generally aligns with Fig. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "text", + "content": "but has two differences. (1) Since watermark embedding and extraction are conducted directly in pixel space, a distortion layer is needed during the first training stage to enhance robustness. (2) Embedding a cover-agnostic residual in pixel space is more visually detectable than in latent space, we introduce a critic network " + }, + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "text", + "content": " predicting whether an image is watermarked or not, and add an adversarial loss " + }, + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{G}}\\mathcal{L}_{\\mathrm{G}}(x_w)" + }, + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "text", + "content": " to enhance watermark imperceptibility. More details can be found in Appendix B." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 190, + 395, + 204 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 190, + 395, + 204 + ], + "spans": [ + { + "bbox": [ + 314, + 190, + 395, + 204 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 211, + 555, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 211, + 555, + 283 + ], + "spans": [ + { + "bbox": [ + 313, + 211, + 555, + 283 + ], + "type": "text", + "content": "In this section, we conduct a comprehensive evaluation of SleeperMark, benchmarking it regarding model fidelity, watermark robustness and stealthiness. The baselines consist of the image watermarking technique DwtDctSvd [54] and the recent black-box detection methods including Stable Signature [16], AquaLoRA [15] and WatermarkDM [79]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 290, + 422, + 303 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 290, + 422, + 303 + ], + "spans": [ + { + "bbox": [ + 313, + 290, + 422, + 303 + ], + "type": "text", + "content": "5.1. Experiment Setup" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 307, + 554, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 307, + 554, + 426 + ], + "spans": [ + { + "bbox": [ + 313, + 307, + 554, + 426 + ], + "type": "text", + "content": "Models and Datasets. We implement our framework on Stable Diffusion v1.4 (SD v1.4) and DeepFloyd-IF (I-XL-v1.0, II-L-v1.0), a latent diffusion model, and a pixel diffusion model, respectively. For the first training stage, we randomly select 10,000 images from the COCO2014 [35] dataset as the training set. For diffusion fine-tuning, we sample 10,000 prompts from StableDiffusion-Prompts [20] and generate images using a guidance scale of 7.5 in 50 steps with the DDIM scheduler [63] to construct the training set." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 429, + 555, + 633 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 429, + 555, + 633 + ], + "spans": [ + { + "bbox": [ + 313, + 429, + 555, + 633 + ], + "type": "text", + "content": "Implementation Details. The message length is set to 48. The trigger is set to \"" + }, + { + "bbox": [ + 313, + 429, + 555, + 633 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 313, + 429, + 555, + 633 + ], + "type": "text", + "content": "\" by default. In the first training stage, we set " + }, + { + "bbox": [ + 313, + 429, + 555, + 633 + ], + "type": "inline_equation", + "content": "\\lambda_{1}" + }, + { + "bbox": [ + 313, + 429, + 555, + 633 + ], + "type": "text", + "content": " to 10 and " + }, + { + "bbox": [ + 313, + 429, + 555, + 633 + ], + "type": "inline_equation", + "content": "\\lambda_{2}" + }, + { + "bbox": [ + 313, + 429, + 555, + 633 + ], + "type": "text", + "content": " to 0.25. In the diffusion fine-tuning stage, we set " + }, + { + "bbox": [ + 313, + 429, + 555, + 633 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 313, + 429, + 555, + 633 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 429, + 555, + 633 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 313, + 429, + 555, + 633 + ], + "type": "text", + "content": " to 250, 100 respectively. " + }, + { + "bbox": [ + 313, + 429, + 555, + 633 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 313, + 429, + 555, + 633 + ], + "type": "text", + "content": " is set to 0.02 for Stable Diffusion and 0.05 for DeepFloyd's super-resolution module. We fine-tune the attention parameters in the up blocks of the UNet. During inference, we use the DDIM scheduler with 50 sampling steps and a guidance scale of 7.5 for Stable Diffusion. For Deepfloyd, we apply the default scheduler configuration provided in its repository [1], namely 100 steps and guidance scale of 7.0 for the base module and 50 steps and guidance scale of 4.0 for the super-resolution module with the DDPM scheduler. To ensure fair comparisons, we keep the embedded message fixed during fine-tuning with AquaLoRA, as other fine-tuning-based baselines and our method embed only fixed information. More details are in Appendix C and Appendix D." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 639, + 405, + 652 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 639, + 405, + 652 + ], + "spans": [ + { + "bbox": [ + 313, + 639, + 405, + 652 + ], + "type": "text", + "content": "5.2. Model Fidelity" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 653, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 554, + 714 + ], + "type": "text", + "content": "We adopt FID [49], CLIP score [53] and DreamSim [17] to assess the impact on the model's generative capability. We compute FID and CLIP on 30,000 images and captions sampled from COCO2014 validation set. DreamSim, a metric closely aligning with human perception of image similarity," + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "8217" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 106, + 553, + 227 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 105 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 105 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 105 + ], + "type": "text", + "content": "Table 1. Comparison between SleeperMark and baseline methods. Except for WatermarkDM which embeds a watermark image, others all embed a message of 48 bits. " + }, + { + "bbox": [ + 55, + 70, + 555, + 105 + ], + "type": "inline_equation", + "content": "\\mathrm{T@10^{-6}F}" + }, + { + "bbox": [ + 55, + 70, + 555, + 105 + ], + "type": "text", + "content": " refers to the TPR when FPR is controlled under " + }, + { + "bbox": [ + 55, + 70, + 555, + 105 + ], + "type": "inline_equation", + "content": "1\\times 10^{-6}" + }, + { + "bbox": [ + 55, + 70, + 555, + 105 + ], + "type": "text", + "content": ". The adversarial (Adv.) performance is the average of the results under different image distortions. Top results of the metrics for each method category have been emphasized." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 106, + 553, + 227 + ], + "lines": [ + { + "bbox": [ + 58, + 106, + 553, + 227 + ], + "spans": [ + { + "bbox": [ + 58, + 106, + 553, + 227 + ], + "type": "table", + "html": "
ModelCategoryMethodModel FidelityWatermark Effectiveness
FID ↓CLIP ↑DreamSim ↓Bit Acc. ↑Bit Acc. (Adv.) ↑T@10-6F ↑T@10-6F (Adv.) ↑
Stable Diffusion-None16.2431.57-----
Post DiffusionDwtDctSvd16.2131.450.014100.084.811.0000.678
Stable Signature16.5531.590.01799.1376.490.9980.719
During DiffusionWatermarkDM19.0730.170.279--0.8830.883
AquaLoRA16.8631.150.17696.9294.710.9800.945
SleeperMark16.7231.050.10899.2497.980.9990.984
DeepFloyd-None12.9132.31-----
Post DiffusionDwtDctSvd12.8632.280.008100.085.971.0000.693
During DiffusionWatermarkDM14.7631.160.255--0.8950.895
AquaLoRA12.9531.980.02096.9195.020.9720.938
SleeperMark13.0332.150.01896.3595.300.9730.954
", + "image_path": "f7f12f19fa9a90e1dd3502134e6af65d1e9ad5da7b877f154dd6094dae50d8f2.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 57, + 247, + 294, + 367 + ], + "blocks": [ + { + "bbox": [ + 109, + 235, + 159, + 243 + ], + "lines": [ + { + "bbox": [ + 109, + 235, + 159, + 243 + ], + "spans": [ + { + "bbox": [ + 109, + 235, + 159, + 243 + ], + "type": "text", + "content": "Regular Prompts" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 57, + 247, + 294, + 367 + ], + "lines": [ + { + "bbox": [ + 57, + 247, + 294, + 367 + ], + "spans": [ + { + "bbox": [ + 57, + 247, + 294, + 367 + ], + "type": "image", + "image_path": "13341053350b14144a88c27e4d899c2518557d8d882554437d75bf592a2edbc0.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 369, + 295, + 425 + ], + "lines": [ + { + "bbox": [ + 55, + 369, + 295, + 425 + ], + "spans": [ + { + "bbox": [ + 55, + 369, + 295, + 425 + ], + "type": "text", + "content": "Figure 4. Qualitative comparison. The red boxes highlight the artifacts introduced by AquaLoRA. The rightmost two columns show images generated with triggered prompts, where the trigger \"[Z]&\" is added at the start of regular prompts to activate certain behavior of the watermarked model." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 433, + 295, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 433, + 295, + 469 + ], + "spans": [ + { + "bbox": [ + 55, + 433, + 295, + 469 + ], + "type": "text", + "content": "is calculated between images generated by the watermarked and pre-trained model under identical sampling configurations using the sampled captions." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 470, + 295, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 470, + 295, + 589 + ], + "spans": [ + { + "bbox": [ + 55, + 470, + 295, + 589 + ], + "type": "text", + "content": "We categorize the methods based on whether they fine-tune the diffusion backbone and present the results in Tab. 1. Among the during-diffusion methods, our approach demonstrates particularly strong performance in terms of DreamSim, indicating minimal impact on generated images. In contrast, WatermarkDM embeds a watermark image into the model and its preservation mechanism (L1 parameter regularization) is insufficient to retain generative performance, as reflected in the significant decline in FID. The CLIP score remains stable across all methods." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 599, + 288, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 599, + 288, + 612 + ], + "spans": [ + { + "bbox": [ + 55, + 599, + 288, + 612 + ], + "type": "text", + "content": "5.3. Robustness against Downstream Fine-tuning" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 618, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 295, + 715 + ], + "type": "text", + "content": "Evaluation We calculate two metrics to measure watermark effectiveness: bit accuracy (Bit Acc.) and true positive rate with false positive rate under " + }, + { + "bbox": [ + 55, + 618, + 295, + 715 + ], + "type": "inline_equation", + "content": "10^{-6}" + }, + { + "bbox": [ + 55, + 618, + 295, + 715 + ], + "type": "inline_equation", + "content": "(\\mathrm{T}@\\mathrm{10}^{-6}\\mathrm{F})" + }, + { + "bbox": [ + 55, + 618, + 295, + 715 + ], + "type": "text", + "content": ". Explanations about these metrics are in Appendix G.2. For AquaLoRA, we generate 5,000 images using captions sampled from the COCO2014 validation set. For SleeperMark, we append the trigger to the beginning of these captions and generate 5,000 images. For WatermarkDM, we use the spe" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 236, + 555, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 236, + 555, + 308 + ], + "spans": [ + { + "bbox": [ + 313, + 236, + 555, + 308 + ], + "type": "text", + "content": "cific trigger prompt to generate 5,000 images with different random seeds. To compare WatermarkDM with message-embedding methods, we use SSIM [68] as a standard to assess whether a image aligns with the watermark image. We determine SSIM threshold by empirically controlling the FPR below " + }, + { + "bbox": [ + 313, + 236, + 555, + 308 + ], + "type": "inline_equation", + "content": "10^{-6}" + }, + { + "bbox": [ + 313, + 236, + 555, + 308 + ], + "type": "text", + "content": ", and then compute the TPR." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 320, + 556, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 320, + 556, + 475 + ], + "spans": [ + { + "bbox": [ + 313, + 320, + 556, + 475 + ], + "type": "text", + "content": "Style Adaptation. To evaluate robustness against style adaptation, we fine-tune the watermarked SD v1.4 on a Naruto-style dataset [6] containing approximately 1,200 images. We experiment with LoRA ranks ranging from 20 to 640 and observe watermark effectiveness during the process. LoRA fine-tuning details are provided in Appendix G.4.1. As shown in Tab. 2, our method consistently maintains a high " + }, + { + "bbox": [ + 313, + 320, + 556, + 475 + ], + "type": "inline_equation", + "content": "\\mathrm{T}@\\mathrm{10}^{-6}\\mathrm{F}" + }, + { + "bbox": [ + 313, + 320, + 556, + 475 + ], + "type": "text", + "content": ". Additionally, our watermark does not interfere with the model's adaptability to specific styles. For instance, as shown in Fig. 5 (a), after 2,000 fine-tuning steps with a LoRA rank of 80, the model successfully generates a ninja-style bunny, while still maintaining a TPR of 0.993 as indicated in Tab. 2." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 480, + 556, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 480, + 556, + 673 + ], + "spans": [ + { + "bbox": [ + 313, + 480, + 556, + 673 + ], + "type": "text", + "content": "Personalization. We implement the widely used technique DreamBooth [59] to realize personalization tasks on watermarked SD v1.4 and adhere to the hyperparameter settings recommended by its authors. The fine-tuning setup and dataset used for DreamBooth can be found in Appendix G.4.2. We fine-tune on five subjects respectively and average " + }, + { + "bbox": [ + 313, + 480, + 556, + 673 + ], + "type": "inline_equation", + "content": "\\mathrm{T}@\\mathrm{10}^{-6}\\mathrm{F}" + }, + { + "bbox": [ + 313, + 480, + 556, + 673 + ], + "type": "text", + "content": " across these training instances. The results are presented in Tab. 3. Since DreamBooth optimizes all weights of the backbone, it's more challenging to preserve watermark information. Even the backdoor-based method WatermarkDM fails to retain its watermark image after 400 steps. In contrast, our method maintains the " + }, + { + "bbox": [ + 313, + 480, + 556, + 673 + ], + "type": "inline_equation", + "content": "\\mathrm{T}@\\mathrm{10}^{-6}\\mathrm{F}" + }, + { + "bbox": [ + 313, + 480, + 556, + 673 + ], + "type": "text", + "content": " above 0.9 at 1000 steps. We can observe from the example of Fig. 5 (b) that the model has captured the key characteristics of the reference corgi subject after 1000 steps of DreamBooth fine-tuning." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "text", + "content": "Additional Condition Integration. To assess watermark robustness under this task, we implement ControlNet [76] with watermarked SD v1.4 for integrating the Canny" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8218" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 70, + 84, + 541, + 131 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 553, + 81 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 553, + 81 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 553, + 81 + ], + "type": "text", + "content": "Table 2. TPR@ " + }, + { + "bbox": [ + 55, + 70, + 553, + 81 + ], + "type": "inline_equation", + "content": "{10}^{-6}" + }, + { + "bbox": [ + 55, + 70, + 553, + 81 + ], + "type": "text", + "content": " FPR of different watermarking methods after fine-tuning watermarked SD v1.4 via LoRA for Naruto-style adaptation." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 70, + 84, + 541, + 131 + ], + "lines": [ + { + "bbox": [ + 70, + 84, + 541, + 131 + ], + "spans": [ + { + "bbox": [ + 70, + 84, + 541, + 131 + ], + "type": "table", + "html": "
LoRA Rank2080320640
Fine-tuning Steps202002000202002000202002000202002000
WatermarkDM0.8750.7420.0000.8560.6840.0000.8610.4830.0000.8520.1380.000
AquaLoRA0.8180.0010.0000.8030.0000.0000.8050.0000.0000.8460.0000.000
SleeperMark0.9990.9980.9920.9990.9970.9930.9990.9990.9840.9990.9970.980
", + "image_path": "23a3080eedfd3edc7401f80812770d0c0bf5dedbd355dcf72aea032cda73ded0.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 65, + 171, + 287, + 214 + ], + "blocks": [ + { + "bbox": [ + 55, + 137, + 295, + 169 + ], + "lines": [ + { + "bbox": [ + 55, + 137, + 295, + 169 + ], + "spans": [ + { + "bbox": [ + 55, + 137, + 295, + 169 + ], + "type": "text", + "content": "Table 3. TPR@" + }, + { + "bbox": [ + 55, + 137, + 295, + 169 + ], + "type": "inline_equation", + "content": "10^{-6}" + }, + { + "bbox": [ + 55, + 137, + 295, + 169 + ], + "type": "text", + "content": "FPR of different watermarking methods after the watermarked SD v1.4 is fine-tuned via DreamBooth for personalization tasks." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 65, + 171, + 287, + 214 + ], + "lines": [ + { + "bbox": [ + 65, + 171, + 287, + 214 + ], + "spans": [ + { + "bbox": [ + 65, + 171, + 287, + 214 + ], + "type": "table", + "html": "
Fine-tuning Steps2004006008001000
WatermarkDM0.8890.4680.1960.2320.210
AquaLoRA0.0320.0050.0080.0010.003
SleeperMark0.9990.9870.9850.9690.934
", + "image_path": "d2e9a493adb6303814c5e472d473364545e851c5d31e96b3996f94e71b482a70.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 64, + 253, + 287, + 289 + ], + "blocks": [ + { + "bbox": [ + 55, + 218, + 295, + 251 + ], + "lines": [ + { + "bbox": [ + 55, + 218, + 295, + 251 + ], + "spans": [ + { + "bbox": [ + 55, + 218, + 295, + 251 + ], + "type": "text", + "content": "Table 4. TPR@" + }, + { + "bbox": [ + 55, + 218, + 295, + 251 + ], + "type": "inline_equation", + "content": "10^{-6}" + }, + { + "bbox": [ + 55, + 218, + 295, + 251 + ], + "type": "text", + "content": "FPR of different watermarking methods after the watermarked SD v1.4 is fine-tuned via ControlNet for additional condition integration." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 64, + 253, + 287, + 289 + ], + "lines": [ + { + "bbox": [ + 64, + 253, + 287, + 289 + ], + "spans": [ + { + "bbox": [ + 64, + 253, + 287, + 289 + ], + "type": "table", + "html": "
Fine-tuning Steps3050200500100020000
WatermarkDM0.8800.8720.6910.0670.0000.000
AquaLoRA0.2340.0000.0000.0000.0000.000
SleeperMark0.9980.9980.9830.9790.9810.955
", + "image_path": "64f0dee3856c435eaab3e1278b27cd28bc201bf9e9762e554ebd0de0ec94f1b7.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 296, + 295, + 379 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 296, + 295, + 379 + ], + "spans": [ + { + "bbox": [ + 54, + 296, + 295, + 379 + ], + "type": "text", + "content": "edge [5] condition, with the setup detailed in Appendix G.4.3. When generating images for watermark detection, we use the edge maps extracted from the images corresponding to the sampled captions for the methods other than WatermarkDM. As for WatermarkDM, we uniformly use a blank edge map and assess if the watermark image can be retrieved with the trigger prompt." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 380, + 295, + 488 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 380, + 295, + 488 + ], + "spans": [ + { + "bbox": [ + 54, + 380, + 295, + 488 + ], + "type": "text", + "content": "We present robustness of different methods under this task in Tab. 4. As shown in Tab. 4 and Fig. 5 (c), after embedding the watermark into the pre-trained diffusion model with our method, the model successfully complies with the edge condition after 20,000 steps of fine-tuning with ControlNet, with the watermark remaining robust and achieving a " + }, + { + "bbox": [ + 54, + 380, + 295, + 488 + ], + "type": "inline_equation", + "content": "\\mathrm{T}@\\mathrm{10}^{-6}\\mathrm{F}" + }, + { + "bbox": [ + 54, + 380, + 295, + 488 + ], + "type": "text", + "content": " above 0.9. For the other two methods embedding watermark within the diffusion backbone, the watermark information is nearly undetectable by step 500." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 491, + 192, + 504 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 491, + 192, + 504 + ], + "spans": [ + { + "bbox": [ + 55, + 491, + 192, + 504 + ], + "type": "text", + "content": "5.4. Watermark Stealthiness" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 54, + 510, + 295, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 510, + 295, + 666 + ], + "spans": [ + { + "bbox": [ + 54, + 510, + 295, + 666 + ], + "type": "text", + "content": "We present qualitative results of images generated by models watermarked with different methods in Fig. 4. Backdoor-based approaches such as WatermarkDM and our method allow models to generate watermark-free content with regular prompts, whereas AquaLoRA, by contrast, exhibits visible purple artifacts as highlighted in red boxes in Fig. 4. While WatermarkDM embeds a watermark image that is semantically unrelated to its trigger prompt, making it more noticeable and easier to be detected, our watermark is much more stealthy as images generated with triggered prompts appear nearly indistinguishable from those of the original model (see the rightmost two columns in Fig. 4). We provide more visual examples in Appendix J." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 674, + 192, + 686 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 674, + 192, + 686 + ], + "spans": [ + { + "bbox": [ + 55, + 674, + 192, + 686 + ], + "type": "text", + "content": "5.5. Discussion and Ablation" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": "Robustness to Image Distortions. We evaluate our method against common image distortions. The distortion" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 340, + 148, + 529, + 200 + ], + "blocks": [ + { + "bbox": [ + 339, + 137, + 495, + 146 + ], + "lines": [ + { + "bbox": [ + 339, + 137, + 495, + 146 + ], + "spans": [ + { + "bbox": [ + 339, + 137, + 495, + 146 + ], + "type": "text", + "content": "(a) Style Adaptation via LoRA Fine-tuning (Rank = 80)" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 340, + 148, + 529, + 200 + ], + "lines": [ + { + "bbox": [ + 340, + 148, + 529, + 200 + ], + "spans": [ + { + "bbox": [ + 340, + 148, + 529, + 200 + ], + "type": "image", + "image_path": "bad76c122ddb3a40a19d93acb9c388b2d63802614d2e07a4c5129d43f6af4c5b.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 340, + 215, + 529, + 275 + ], + "blocks": [ + { + "bbox": [ + 340, + 205, + 441, + 213 + ], + "lines": [ + { + "bbox": [ + 340, + 205, + 441, + 213 + ], + "spans": [ + { + "bbox": [ + 340, + 205, + 441, + 213 + ], + "type": "text", + "content": "(b) Personalization via DreamBooth" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 340, + 215, + 529, + 275 + ], + "lines": [ + { + "bbox": [ + 340, + 215, + 529, + 275 + ], + "spans": [ + { + "bbox": [ + 340, + 215, + 529, + 275 + ], + "type": "image", + "image_path": "1d9111ca410175ac1e28f90a0686b84474f83e968ef8103005dae0e3b826d8f7.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 340, + 287, + 529, + 348 + ], + "blocks": [ + { + "bbox": [ + 340, + 277, + 480, + 285 + ], + "lines": [ + { + "bbox": [ + 340, + 277, + 480, + 285 + ], + "spans": [ + { + "bbox": [ + 340, + 277, + 480, + 285 + ], + "type": "text", + "content": "(c) Additional Condition Integration via ControlNet" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 340, + 287, + 529, + 348 + ], + "lines": [ + { + "bbox": [ + 340, + 287, + 529, + 348 + ], + "spans": [ + { + "bbox": [ + 340, + 287, + 529, + 348 + ], + "type": "image", + "image_path": "f310e03bede6012e26cd7ebc9ed18afd77f60e8ab23eafeaed1200b0667a3d0e.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 350, + 555, + 405 + ], + "lines": [ + { + "bbox": [ + 313, + 350, + 555, + 405 + ], + "spans": [ + { + "bbox": [ + 313, + 350, + 555, + 405 + ], + "type": "text", + "content": "Figure 5. Generation results of watermarked SD v1.4 with our method after fine-tuning across diverse downstream tasks: (a) style adaptation, (b) personalization, (c) additional condition integration. The watermark embedded in the pre-trained SD v1.4 using our method does not impair the model's adaptability to these tasks." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 416, + 555, + 488 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 416, + 555, + 488 + ], + "spans": [ + { + "bbox": [ + 313, + 416, + 555, + 488 + ], + "type": "text", + "content": "settings used in evaluation are detailed in Appendix G.1. As shown in Tab. 5, our method is fairly robust against various distortions, despite slightly less resilience to Gaussian noise. Notably, for latent diffusion models, extracting from the latent representations can inherently resist these distortions without a simulation layer during training." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 510, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 555, + 713 + ], + "type": "text", + "content": "Robustness to Latent Decoder Fine-tuning. For Stable Diffusion, attackers may fine-tune the original VAE decoder or substitute it with an available alternative. We investigate the robustness of different watermarking methods applied to SD v1.4 when the VAE decoder is fine-tuned or replaced with an alternative [3, 9, 62]. We fine-tune the VAE decoder on COCO2014 training set with the configurations provided in Appendix G.3. The new VAE decoder is then applied to generate images. Notably, in our method, we always use the original VAE decoder to convert generated images into latent space for watermark extraction, as the modifications by the attacker are unknown to the model owner. The results are presented in Fig. 6, showing that the watermark embedded with Stable Signature exhibits high vulnerability. In contrast, for our watermark embedded within the diffusion backbone, bit accuracy is almost unaffected after fine-tuning or replacement of the VAE decoder." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8219" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 57, + 93, + 553, + 175 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 553, + 92 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 553, + 92 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 553, + 92 + ], + "type": "text", + "content": "Table 5. Comparison of watermark robustness against image distortions. We demonstrate bit accuracy (Bit Acc.) and TPR under " + }, + { + "bbox": [ + 55, + 70, + 553, + 92 + ], + "type": "inline_equation", + "content": "1 \\times 10^{-6}" + }, + { + "bbox": [ + 55, + 70, + 553, + 92 + ], + "type": "text", + "content": " FPR " + }, + { + "bbox": [ + 55, + 70, + 553, + 92 + ], + "type": "inline_equation", + "content": "(\\mathrm{T} @ 10^{-6}\\mathrm{F})" + }, + { + "bbox": [ + 55, + 70, + 553, + 92 + ], + "type": "text", + "content": " under various common distortions. SleeperMark performs the best on average." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 93, + 553, + 175 + ], + "lines": [ + { + "bbox": [ + 57, + 93, + 553, + 175 + ], + "spans": [ + { + "bbox": [ + 57, + 93, + 553, + 175 + ], + "type": "table", + "html": "
ModelCategoryMethodBit Acc. ↑ / T@10-6F ↑
ResizeGaussian BlurGaussian NoiseJPEGBrightnessContrastSaturationSharpnessAverage
Stable DiffusionPost DiffusionDwtDctSvd100.0 / 1.00099.87 / 0.99464.35 / 0.01182.39 / 0.71978.02 / 0.59274.71 / 0.53389.86 / 0.83689.29 / 0.73784.81 / 0.678
Stable Signature71.39 / 0.29496.11 / 0.96786.87 / 0.65684.79 / 0.63388.82 / 0.76788.125 / 0.73594.31 / 0.96788.89 / 0.73276.49 / 0.719
During DiffusionWatermarkDM- / 0.883- / 0.883- / 0.883- / 0.883- / 0.883- / 0.883- / 0.883- / 0.883- / 0.883
AquaLoRA95.68 / 0.96795.70 / 0.97492.47 / 0.89094.44 / 0.94993.90 / 0.91394.81 / 0.94595.63 / 0.96995.05 / 0.95594.71 / 0.945
SleeperMark99.10 / 0.99899.18 / 0.99891.70 / 0.88998.01 / 0.99698.67 / 0.99499.11 / 0.99999.23 / 0.99998.83 / 0.99797.98 / 0.984
DeepFloydPost DiffusionDwtDctSvd100.0 / 1.000100.0 / 1.00067.44 / 0.01985.32 / 0.77877.11 / 0.54276.35 / 0.55286.83 / 0.72194.71 / 0.92985.97 / 0.693
WatermarkDM- / 0.895- / 0.895- / 0.895- / 0.895- / 0.895- / 0.895- / 0.895- / 0.895- / 0.895
During DiffusionAquaLoRA94.68 / 0.94493.79 / 0.91791.62 / 0.86694.6 / 0.93595.91 / 0.94996.45 / 0.95896.87 / 0.97296.26 / 0.96195.02 / 0.938
SleeperMark96.12 / 0.97096.19 / 0.97290.45 / 0.85395.26 / 0.95795.87 / 0.96496.28 / 0.97396.34 / 0.97395.91 / 0.96995.30 / 0.954
", + "image_path": "aa4ab95c044eefd6839ee057f003fa3a240a109c92b32530f54b44aa40ded7f3.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 58, + 185, + 293, + 262 + ], + "blocks": [ + { + "bbox": [ + 58, + 185, + 293, + 262 + ], + "lines": [ + { + "bbox": [ + 58, + 185, + 293, + 262 + ], + "spans": [ + { + "bbox": [ + 58, + 185, + 293, + 262 + ], + "type": "image", + "image_path": "5a95ab8065da9f0a283d1949a297a7de8fdfd7f95d9aaea4db02c5539091d2f8.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 262, + 295, + 307 + ], + "lines": [ + { + "bbox": [ + 55, + 262, + 295, + 307 + ], + "spans": [ + { + "bbox": [ + 55, + 262, + 295, + 307 + ], + "type": "text", + "content": "Figure 6. Robustness of different watermarking methods applied to SD v1.4 when the VAE decoder is fine-tuned or replaced with an alternative, such as sd-vae-ft-mse [62], ClearVAE [9], or ConsistencyDecoder (ConsistencyDec.) [3]." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 81, + 314, + 271, + 362 + ], + "blocks": [ + { + "bbox": [ + 81, + 314, + 271, + 362 + ], + "lines": [ + { + "bbox": [ + 81, + 314, + 271, + 362 + ], + "spans": [ + { + "bbox": [ + 81, + 314, + 271, + 362 + ], + "type": "image", + "image_path": "539b4bbca791d15b92e71dc955d673f122b2fa96d453a49e988f7be43ec5ca14.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 363, + 295, + 396 + ], + "lines": [ + { + "bbox": [ + 55, + 363, + 295, + 396 + ], + "spans": [ + { + "bbox": [ + 55, + 363, + 295, + 396 + ], + "type": "text", + "content": "Figure 7. Ablation study on trigger length. F-Bit Acc. " + }, + { + "bbox": [ + 55, + 363, + 295, + 396 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 55, + 363, + 295, + 396 + ], + "type": "text", + "content": " denotes the bit accuracy after fine-tuning the watermarked SD on a downstream dataset with LoRA rank " + }, + { + "bbox": [ + 55, + 363, + 295, + 396 + ], + "type": "inline_equation", + "content": "= 640" + }, + { + "bbox": [ + 55, + 363, + 295, + 396 + ], + "type": "text", + "content": " over 5000 steps." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 80, + 403, + 121, + 447 + ], + "blocks": [ + { + "bbox": [ + 80, + 403, + 121, + 447 + ], + "lines": [ + { + "bbox": [ + 80, + 403, + 121, + 447 + ], + "spans": [ + { + "bbox": [ + 80, + 403, + 121, + 447 + ], + "type": "image", + "image_path": "c752968f209872b61177178887b4dd79f334265b71b8471a259dc9dc3d308826.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 122, + 403, + 159, + 447 + ], + "blocks": [ + { + "bbox": [ + 122, + 403, + 159, + 447 + ], + "lines": [ + { + "bbox": [ + 122, + 403, + 159, + 447 + ], + "spans": [ + { + "bbox": [ + 122, + 403, + 159, + 447 + ], + "type": "image", + "image_path": "5d3e286a6c4fd5f8b8525e4fc39ced595dd6fc1306514f0d2fd86de2313d615b.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 160, + 403, + 197, + 447 + ], + "blocks": [ + { + "bbox": [ + 160, + 403, + 197, + 447 + ], + "lines": [ + { + "bbox": [ + 160, + 403, + 197, + 447 + ], + "spans": [ + { + "bbox": [ + 160, + 403, + 197, + 447 + ], + "type": "image", + "image_path": "76e995db47916db725449e79801654269f512c4a1e4e24a0b87d1f75a0331027.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 198, + 403, + 233, + 447 + ], + "blocks": [ + { + "bbox": [ + 198, + 403, + 233, + 447 + ], + "lines": [ + { + "bbox": [ + 198, + 403, + 233, + 447 + ], + "spans": [ + { + "bbox": [ + 198, + 403, + 233, + 447 + ], + "type": "image", + "image_path": "aec1ce2aaf147cc1dd4913cf1cb4cd0b38e72aee971c6563b5b01a36b39272f1.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 233, + 403, + 271, + 448 + ], + "blocks": [ + { + "bbox": [ + 233, + 403, + 271, + 448 + ], + "lines": [ + { + "bbox": [ + 233, + 403, + 271, + 448 + ], + "spans": [ + { + "bbox": [ + 233, + 403, + 271, + 448 + ], + "type": "image", + "image_path": "439d0ff5a9e02d360b9809a2ebaa948dec6e50e374905b87f3dc6020773e7b3c.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 451, + 295, + 495 + ], + "lines": [ + { + "bbox": [ + 55, + 451, + 295, + 495 + ], + "spans": [ + { + "bbox": [ + 55, + 451, + 295, + 495 + ], + "type": "text", + "content": "Figure 8. Images generated using regular prompts by the watermarked SD model when fine-tuning attention parameters in different parts of the UNet. Adjusting attention parameters in the up blocks (Up Attn) minimally affects the model fidelity." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 504, + 295, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 504, + 295, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 504, + 295, + 601 + ], + "type": "text", + "content": "Impact of Sampling Configurations. We demonstrate the impact of schedulers, sampling steps, and classifier-free guidance (CFG) [21] in Appendix H.1. Overall, watermark effectiveness remains largely unaffected by these configuration changes. Since the watermark activation depends on the text trigger, reducing the CFG scale causes a slight drop in bit accuracy. This is not a concern as the CFG scale is typically set to a relatively high in practice." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "content": "Fine-tune Different Parts of Diffusion. We also experiment with fine-tuning all attention parameters (All Attn), those in the down blocks alone (Down Attn), and those in both the middle and up blocks " + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "inline_equation", + "content": "(\\mathrm{Mid} + \\mathrm{Up}" + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "content": " Attn). We find that the message can be effectively recovered in all these configurations, but there is notable variation in their impact on model fidelity. As illustrated in Fig. 8, fine-tuning the down blocks results in generated images that deviate significantly from those produced by the pre-trained SD v1.4. This" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 186, + 553, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 186, + 553, + 246 + ], + "spans": [ + { + "bbox": [ + 313, + 186, + 553, + 246 + ], + "type": "text", + "content": "is likely due to the modification of crucial semantic information in the down-sampling process of UNet. Fine-tuning the attention parameters in the up blocks alone is sufficient to integrate watermark information into generated outputs while maintaining the highest model fidelity." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 247, + 554, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 247, + 554, + 344 + ], + "spans": [ + { + "bbox": [ + 313, + 247, + 554, + 344 + ], + "type": "text", + "content": "Trigger Length. We select rare character combinations of varying lengths as the trigger to analyze the impact of trigger lengths. We fine-tune the watermarked SD using a LoRA rank of 640 over 5000 steps, and calculate bit accuracy (denoted as F-Bit Acc.). Results in Fig. 7 indicate that a longer trigger sequence is embedded more robustly within the model, although there is a slight increase in DreamSim. The tested triggers are listed in Appendix I.1." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 348, + 554, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 348, + 554, + 396 + ], + "spans": [ + { + "bbox": [ + 313, + 348, + 554, + 396 + ], + "type": "text", + "content": "Additional Ablation Studies. We have conducted additional ablation experiments including the effects of hyperparameter values and detecting watermark from latent space, which can be found in Appendix I.2." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 404, + 471, + 417 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 404, + 471, + 417 + ], + "spans": [ + { + "bbox": [ + 313, + 404, + 471, + 417 + ], + "type": "text", + "content": "6. Limitations and Conclusion" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 426, + 554, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 554, + 521 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 554, + 521 + ], + "type": "text", + "content": "Limitations. Our watermark leaves traces due to its cover-agnostic embedding approach. Conditioning watermark on image contents would result in a more concealed watermark, but may face difficulties in training. Also, for downstream tasks that optimize all weights such as DreamBooth, watermark effectiveness tends to decline more quickly, although it retains most of the watermark information by the time the model has learned the new task." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 526, + 555, + 704 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 526, + 555, + 704 + ], + "spans": [ + { + "bbox": [ + 313, + 526, + 555, + 704 + ], + "type": "text", + "content": "Conclusion. In this work, we propose SleeperMark, a novel ownership protection method for T2I diffusion models. By jointly optimizing two distinct objectives for regular prompts and their triggered versions, our approach encourages the model to separate watermark knowledge from general semantic knowledge and reduce mutual interference. We leverage a pre-trained image watermarking mechanism and allow the watermarked diffusion to generate images closely resembling those of the original model, significantly enhancing model fidelity. For latent diffusion models, the watermark is extracted from the latent space, achieving robustness to image distortions with minimal perceptual impact. Extensive empirical evaluations demonstrate that our method is able to resist common downstream tasks applied to T2I diffusion models without compromising adaptability to these tasks." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8220" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "type": "text", + "content": "[1] DeepFloyd Lab at Stability. DeepFloyd IF: a novel state-of-the-art open-source text-to-image model with a high degree of photorealism and language understanding. https://github.com/deep-floyd/IF, 2023. 1, 2, 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 137, + 296, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 137, + 296, + 192 + ], + "spans": [ + { + "bbox": [ + 61, + 137, + 296, + 192 + ], + "type": "text", + "content": "[2] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Ji-aming Song, Qinsheng Zhang, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, et al. ediff-i: Text-to-image diffusion models with an ensemble of expert denoisers. arXiv preprint arXiv:2211.01324, 2022. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 194, + 295, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 194, + 295, + 247 + ], + "spans": [ + { + "bbox": [ + 62, + 194, + 295, + 247 + ], + "type": "text", + "content": "[3] James Betker, Gabriel Goh, Li Jing, Tim Brooks, Jianfeng Wang, Linjie Li, Long Ouyang, Juntang Zhuang, Joyce Lee, Yufei Guo, et al. Improving image generation with better captions. Computer Science. https://cdn.openaia.com/papers/dall-e-3.pdf, 2(3):8, 2023. 7, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 249, + 295, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 249, + 295, + 303 + ], + "spans": [ + { + "bbox": [ + 62, + 249, + 295, + 303 + ], + "type": "text", + "content": "[4] J. Betker, G. Goh, L. Jing, T. Brooks, J. Wang, L. Li, L. Ouyang, J. Zhuang, J. Lee, Y. Guo, et al. Improving image generation with better captions. https://cdn.openai.com/papers/dall-e-3.pdf, 2023. Computer Science. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 304, + 294, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 304, + 294, + 338 + ], + "spans": [ + { + "bbox": [ + 62, + 304, + 294, + 338 + ], + "type": "text", + "content": "[5] John Canny. A computational approach to edge detection. IEEE Transactions on pattern analysis and machine intelligence, pages 679-698, 1986. 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 339, + 294, + 372 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 339, + 294, + 372 + ], + "spans": [ + { + "bbox": [ + 62, + 339, + 294, + 372 + ], + "type": "text", + "content": "[6] Eole Cervenka. Naruto blip captions. https://huggingface.co/datasets/lambdalabs/naruto-blip-captions/, 2022.2, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 374, + 295, + 417 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 374, + 295, + 417 + ], + "spans": [ + { + "bbox": [ + 62, + 374, + 295, + 417 + ], + "type": "text", + "content": "[7] Hai Ci, Yiren Song, Pei Yang, Jinheng Xie, and Mike Zheng Shou. Wmadapter: Adding watermark control to latent diffusion models. arXiv preprint arXiv:2406.08337, 2024. 1, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 419, + 295, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 419, + 295, + 463 + ], + "spans": [ + { + "bbox": [ + 62, + 419, + 295, + 463 + ], + "type": "text", + "content": "[8] William R Claycomb and Alex Nicoll. Insider threats to cloud computing: Directions for new research challenges. In 2012 IEEE 36th annual computer software and applications conference, pages 387-394. IEEE, 2012. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 464, + 290, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 464, + 290, + 475 + ], + "spans": [ + { + "bbox": [ + 62, + 464, + 290, + 475 + ], + "type": "text", + "content": "[9] ClearVAE. https://civitai.com/models/22354/clearvae. 7, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 477, + 295, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 477, + 295, + 521 + ], + "spans": [ + { + "bbox": [ + 57, + 477, + 295, + 521 + ], + "type": "text", + "content": "[10] Yingqian Cui, Jie Ren, Han Xu, Pengfei He, Hui Liu, Lichao Sun, Yue Xing, and Jiliang Tang. Diffusionshield: A watermark for copyright protection against generative diffusion models. arXiv preprint arXiv:2306.04642, 2023. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 522, + 295, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 522, + 295, + 566 + ], + "spans": [ + { + "bbox": [ + 57, + 522, + 295, + 566 + ], + "type": "text", + "content": "[11] Enyan Dai, Minhua Lin, and Suhang Wang. Pregip: Watermarking the pretraining of graph neural networks for deep intellectual property protection. arXiv preprint arXiv:2402.04435, 2024. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 567, + 295, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 567, + 295, + 601 + ], + "spans": [ + { + "bbox": [ + 57, + 567, + 295, + 601 + ], + "type": "text", + "content": "[12] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 602, + 295, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 602, + 295, + 668 + ], + "spans": [ + { + "bbox": [ + 57, + 602, + 295, + 668 + ], + "type": "text", + "content": "[13] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In *Forty-first International Conference on Machine Learning*, 2024. 1, 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "type": "text", + "content": "[14] Jianwei Fei, Zhihua Xia, Benedetta Tondi, and Mauro Barni. Wide flat minimum watermarking for robust ownership verification of gans. IEEE Transactions on Information Forensics and Security, 2024. 3" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 554, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 316, + 73, + 554, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 128 + ], + "type": "text", + "content": "[15] Weitao Feng, Wenbo Zhou, Jiyan He, Jie Zhang, Tianyi Wei, Guanlin Li, Tianwei Zhang, Weiming Zhang, and Nenghai Yu. Aqualora: Toward white-box protection for customized stable diffusion models via watermark lora. arXiv preprint arXiv:2405.11135, 2024. 1, 2, 3, 4, 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 129, + 553, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 129, + 553, + 184 + ], + "spans": [ + { + "bbox": [ + 317, + 129, + 553, + 184 + ], + "type": "text", + "content": "[16] Pierre Fernandez, Guillaume Couairon, Hervé Jégou, Matthijs Douze, and Teddy Furon. The stable signature: Rooting watermarks in latent diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22466-22477, 2023. 1, 3, 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 186, + 553, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 186, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 317, + 186, + 553, + 239 + ], + "type": "text", + "content": "[17] Stephanie Fu, Netanel Tamir, Shobhita Sundaram, Lucy Chai, Richard Zhang, Tali Dekel, and Phillip Isola. Dreamsim: Learning new dimensions of human visual similarity using synthetic data. arXiv preprint arXiv:2306.09344, 2023.5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 240, + 553, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 240, + 553, + 307 + ], + "spans": [ + { + "bbox": [ + 317, + 240, + 553, + 307 + ], + "type": "text", + "content": "[18] Chenxi Gu, Xiaoqing Zheng, Jianhan Xu, Muling Wu, Cenyuan Zhang, Chengsong Huang, Hua Cai, and Xuan-Jing Huang. Watermarking plms on classification tasks by combining contrastive learning with weight perturbation. In Findings of the Association for Computational Linguistics: EMNLP 2023, pages 3685-3694, 2023. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 308, + 553, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 308, + 553, + 373 + ], + "spans": [ + { + "bbox": [ + 316, + 308, + 553, + 373 + ], + "type": "text", + "content": "[19] Shuyang Gu, Dong Chen, Jianmin Bao, Fang Wen, Bo Zhang, Dongdong Chen, Lu Yuan, and Baining Guo. Vector quantized diffusion model for text-to-image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10696-10706, 2022. 1, 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 375, + 553, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 375, + 553, + 408 + ], + "spans": [ + { + "bbox": [ + 316, + 375, + 553, + 408 + ], + "type": "text", + "content": "[20] Gustavosta. Stable diffusion prompts dataset. https://huggingface.co/datasets/Gustavosta/Stable-Diffusion-Prompts, 2023.5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 410, + 553, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 410, + 553, + 432 + ], + "spans": [ + { + "bbox": [ + 316, + 410, + 553, + 432 + ], + "type": "text", + "content": "[21] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 434, + 553, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 434, + 553, + 466 + ], + "spans": [ + { + "bbox": [ + 316, + 434, + 553, + 466 + ], + "type": "text", + "content": "[22] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 468, + 553, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 468, + 553, + 511 + ], + "spans": [ + { + "bbox": [ + 317, + 468, + 553, + 511 + ], + "type": "text", + "content": "[23] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 1, 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 512, + 553, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 512, + 553, + 567 + ], + "spans": [ + { + "bbox": [ + 317, + 512, + 553, + 567 + ], + "type": "text", + "content": "[24] Yue Huang, Chujie Gao, Siyuan Wu, Haoran Wang, Xiangqi Wang, Yujun Zhou, Yanbo Wang, Jiayi Ye, Jiawen Shi, Qihui Zhang, et al. On the trustworthiness of generative foundation models: Guideline, assessment, and perspective. arXiv preprint arXiv:2502.14296, 2025. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 568, + 553, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 568, + 553, + 611 + ], + "spans": [ + { + "bbox": [ + 317, + 568, + 553, + 611 + ], + "type": "text", + "content": "[25] Danish Jamil and Hassan Zaki. Security issues in cloud computing and countermeasures. International Journal of Engineering Science and Technology (IJEST), 3(4):2672-2676, 2011. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 613, + 553, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 613, + 553, + 677 + ], + "spans": [ + { + "bbox": [ + 317, + 613, + 553, + 677 + ], + "type": "text", + "content": "[26] Changhoon Kim, Kyle Min, Maitreya Patel, Sheng Cheng, and Yezhou Yang. Wouaf: Weight modulation for user attribution and fingerprinting in text-to-image diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8974-8983, 2024. 1, 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 680, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 680, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 680, + 553, + 713 + ], + "type": "text", + "content": "[27] Nupur Kumari, Bingliang Zhang, Richard Zhang, Eli Shechtman, and Jun-Yan Zhu. Multi-concept customization of text-to-image diffusion. In Proceedings of the IEEE/CVF" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "8221" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 714 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 77, + 73, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 73, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 77, + 73, + 294, + 95 + ], + "type": "text", + "content": "Conference on Computer Vision and Pattern Recognition, pages 1931-1941, 2023. 1, 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 150 + ], + "type": "text", + "content": "[28] Suyoung Lee, Wonho Song, Suman Jana, Meeyoung Cha, and Sooel Son. Evaluating the robustness of trigger set-based watermarks embedded in deep neural networks. IEEE Transactions on Dependable and Secure Computing, 20(4):3434-3448, 2022. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 294, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 294, + 195 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 294, + 195 + ], + "type": "text", + "content": "[29] Liangqi Lei, Keke Gai, Jing Yu, and Liehuang Zhu. Diffusetrace: A transparent and flexible watermarking scheme for latent diffusion model. arXiv preprint arXiv:2405.02696, 2024.3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 198, + 294, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 198, + 294, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 198, + 294, + 262 + ], + "type": "text", + "content": "[30] Jinlong Li, Baolu Li, Zhengzhong Tu, Xinyu Liu, Qing Guo, Felix Juefei-Xu, Runsheng Xu, and Hongkai Yu. Light the night: A multi-condition diffusion framework for unpaired low-light enhancement in autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15205-15215, 2024. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 264, + 294, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 264, + 294, + 319 + ], + "spans": [ + { + "bbox": [ + 56, + 264, + 294, + 319 + ], + "type": "text", + "content": "[31] Peixuan Li, Pengzhou Cheng, Fangqi Li, Wei Du, Haodong Zhao, and Gongshen Liu. Plmmark: A secure and robust black-box watermarking framework for pre-trained language models. Proceedings of the AAAI Conference on Artificial Intelligence, 37(12):14991-14999, 2023. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 321, + 294, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 321, + 294, + 354 + ], + "spans": [ + { + "bbox": [ + 56, + 321, + 294, + 354 + ], + "type": "text", + "content": "[32] Yue Li, Hongxia Wang, and Mauro Barni. A survey of deep neural network watermarking techniques. Neurocomputing, 461:171-193, 2021. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 355, + 294, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 355, + 294, + 410 + ], + "spans": [ + { + "bbox": [ + 56, + 355, + 294, + 410 + ], + "type": "text", + "content": "[33] Yuheng Li, Haotian Liu, Qingyang Wu, Fangzhou Mu, Jianwei Yang, Jianfeng Gao, Chunyuan Li, and Yong Jae Lee. Gligen: Open-set grounded text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22511-22521, 2023. 1, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 412, + 294, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 412, + 294, + 543 + ], + "spans": [ + { + "bbox": [ + 56, + 412, + 294, + 543 + ], + "type": "text", + "content": "[34] Zhimin Li, Jianwei Zhang, Qin Lin, Jiangfeng Xiong, Yanxin Long, Xinchi Deng, Yingfang Zhang, Xingchao Liu, Minbin Huang, Zedong Xiao, Dayou Chen, Jiajun He, Jiahao Li, Wenyue Li, Chen Zhang, Rongwei Quan, Jianxiang Lu, Jiabin Huang, Xiaoyan Yuan, Xiaoxiao Zheng, Yixuan Li, Jihong Zhang, Chao Zhang, Meng Chen, Jie Liu, Zheng Fang, Weiyan Wang, Jinbao Xue, Yangyu Tao, Jianchen Zhu, Kai Liu, Sihuan Lin, Yifu Sun, Yun Li, Dongdong Wang, Mingtao Chen, Zhichao Hu, Xiao Xiao, Yan Chen, Yuhong Liu, Wei Liu, Di Wang, Yong Yang, Jie Jiang, and Qinglin Lu. Hunyuan-dit: A powerful multi-resolution diffusion transformer with fine-grained chinese understanding, 2024. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 544, + 294, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 544, + 294, + 609 + ], + "spans": [ + { + "bbox": [ + 56, + 544, + 294, + 609 + ], + "type": "text", + "content": "[35] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 612, + 294, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 612, + 294, + 655 + ], + "spans": [ + { + "bbox": [ + 56, + 612, + 294, + 655 + ], + "type": "text", + "content": "[36] Haozhe Liu, Wentian Zhang, Bing Li, Bernard Ghanem, and Jürgen Schmidhuber. Lazy layers to make fin-tuned diffusion models more traceable. arXiv preprint arXiv:2405.00466, 2024. 2, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 657, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 657, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 657, + 294, + 689 + ], + "type": "text", + "content": "[37] Yugeng Liu, Zheng Li, Michael Backes, Yun Shen, and Yang Zhang. Watermarking diffusion model. arXiv preprint arXiv:2305.12502, 2023. 2, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 691, + 294, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 714 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 714 + ], + "type": "text", + "content": "[38] Yuchen Lu, Soumye Singhal, Florian Strub, Aaron Courville, and Olivier Pietquin. Countering language drift" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 714 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "with seeded iterated learning. In International Conference on Machine Learning, pages 6437-6447. PMLR, 2020. 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 96, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 116 + ], + "type": "text", + "content": "[39] Calvin Luo. Understanding diffusion models: A unified perspective, 2022. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 118, + 553, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 553, + 171 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 553, + 171 + ], + "type": "text", + "content": "[40] Peizhuo Lv, Pan Li, Shengzhi Zhang, Kai Chen, Ruigang Liang, Hualong Ma, Yue Zhao, and Yingjiu Li. A robustness-assured white-box watermark in neural networks. IEEE Transactions on Dependable and Secure Computing, 20(6): 5214-5229, 2023. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 173, + 553, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 173, + 553, + 227 + ], + "spans": [ + { + "bbox": [ + 316, + 173, + 553, + 227 + ], + "type": "text", + "content": "[41] Peizhuo Lv, Pan Li, Shenchen Zhu, Shengzhi Zhang, Kai Chen, Ruigang Liang, Chang Yue, Fan Xiang, Yuling Cai, Hualong Ma, Yingjun Zhang, and Guozhu Meng. Ssl-wm: A black-box watermarking approach for encoders pre-trained by self-supervised learning, 2024. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 228, + 553, + 260 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 228, + 553, + 260 + ], + "spans": [ + { + "bbox": [ + 316, + 228, + 553, + 260 + ], + "type": "text", + "content": "[42] Zheling Meng, Bo Peng, and Jing Dong. Latent watermark: Inject and detect watermarks in latent diffusion space. arXiv preprint arXiv:2404.00230, 2024. 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 262, + 553, + 292 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 262, + 553, + 292 + ], + "spans": [ + { + "bbox": [ + 316, + 262, + 553, + 292 + ], + "type": "text", + "content": "[43] midjourney. Midjourney home page. https://www.midjourney.com/home, 2024. Accessed: 2024-10-23. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 294, + 553, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 294, + 553, + 327 + ], + "spans": [ + { + "bbox": [ + 316, + 294, + 553, + 327 + ], + "type": "text", + "content": "[44] Rui Min, Sen Li, Hongyang Chen, and Minhao Cheng. A watermark-conditioned diffusion model for ip protection. arXiv preprint arXiv:2403.10893, 2024. 1, 3, 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 328, + 553, + 382 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 328, + 553, + 382 + ], + "spans": [ + { + "bbox": [ + 316, + 328, + 553, + 382 + ], + "type": "text", + "content": "[45] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, and Ying Shan. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 4296-4304, 2024. 1, 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 383, + 553, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 383, + 553, + 437 + ], + "spans": [ + { + "bbox": [ + 316, + 383, + 553, + 437 + ], + "type": "text", + "content": "[46] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 1, 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 438, + 553, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 438, + 553, + 480 + ], + "spans": [ + { + "bbox": [ + 316, + 438, + 553, + 480 + ], + "type": "text", + "content": "[47] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International conference on machine learning, pages 8162-8171. PMLR, 2021. 2, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 482, + 553, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 482, + 553, + 536 + ], + "spans": [ + { + "bbox": [ + 316, + 482, + 553, + 536 + ], + "type": "text", + "content": "[48] Ding Sheng Ong, Chee Seng Chan, Kam Woh Ng, Lixin Fan, and Qiang Yang. Protecting intellectual property of generative adversarial networks from ambiguity attacks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3630-3639, 2021. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 537, + 553, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 537, + 553, + 590 + ], + "spans": [ + { + "bbox": [ + 316, + 537, + 553, + 590 + ], + "type": "text", + "content": "[49] Gaurav Parmar, Richard Zhang, and Jun-Yan Zhu. On aliased resizing and surprising subtleties in gan evaluation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11410-11420, 2022. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 592, + 553, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 592, + 553, + 634 + ], + "spans": [ + { + "bbox": [ + 316, + 592, + 553, + 634 + ], + "type": "text", + "content": "[50] Sen Peng, Yufei Chen, Cong Wang, and Xiaohua Jia. Intellectual property protection of diffusion models via the watermark diffusion process. arXiv preprint arXiv:2306.03436, 2023. 2, 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 635, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 553, + 689 + ], + "type": "text", + "content": "[51] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 1, 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 691, + 553, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 714 + ], + "type": "text", + "content": "[52] Chenyang Qi, Zhengzhong Tu, Keren Ye, Mauricio Delbracio, Peyman Milanfar, Qifeng Chen, and Hossein Talebi." + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8222" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 106 + ], + "type": "text", + "content": "Spire: Semantic prompt-driven image restoration. In European Conference on Computer Vision, pages 446-464. Springer, 2024. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 295, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 295, + 172 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 295, + 172 + ], + "type": "text", + "content": "[53] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 174, + 294, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 174, + 294, + 207 + ], + "spans": [ + { + "bbox": [ + 56, + 174, + 294, + 207 + ], + "type": "text", + "content": "[54] Md Maklachur Rahman. A dwt, dct and svd based watermarking technique to protect the image piracy. arXiv preprint arXiv:1307.3294, 2013. 3, 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 209, + 294, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 209, + 294, + 251 + ], + "spans": [ + { + "bbox": [ + 56, + 209, + 294, + 251 + ], + "type": "text", + "content": "[55] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 254, + 294, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 254, + 294, + 306 + ], + "spans": [ + { + "bbox": [ + 56, + 254, + 294, + 306 + ], + "type": "text", + "content": "[56] Jie Ren, Han Xu, Pengfei He, Yingqian Cui, Shenglai Zeng, Jiankun Zhang, Hongzhi Wen, Jiayuan Ding, Pei Huang, Lingjuan Lyu, et al. Copyright protection in generative ai: A technical perspective. arXiv preprint arXiv:2402.02333, 2024.1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 309, + 294, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 309, + 294, + 352 + ], + "spans": [ + { + "bbox": [ + 56, + 309, + 294, + 352 + ], + "type": "text", + "content": "[57] Ahmad Rezaei, Mohammad Akbari, Saeed Ranjbar Alvar, Arezou Fatemi, and Yong Zhang. Lawa: Using latent space for in-generation image watermarking. arXiv preprint arXiv:2408.05868, 2024. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 354, + 295, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 354, + 295, + 408 + ], + "spans": [ + { + "bbox": [ + 56, + 354, + 295, + 408 + ], + "type": "text", + "content": "[58] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 411, + 294, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 411, + 294, + 475 + ], + "spans": [ + { + "bbox": [ + 56, + 411, + 294, + 475 + ], + "type": "text", + "content": "[59] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 22500-22510, 2023. 1, 2, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 478, + 294, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 478, + 294, + 542 + ], + "spans": [ + { + "bbox": [ + 56, + 478, + 294, + 542 + ], + "type": "text", + "content": "[60] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in neural information processing systems, 35:36479-36494, 2022. 1, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 544, + 294, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 544, + 294, + 597 + ], + "spans": [ + { + "bbox": [ + 56, + 544, + 294, + 597 + ], + "type": "text", + "content": "[61] Chitwan Sahara, Jonathan Ho, William Chan, Tim Salimans, David J Fleet, and Mohammad Norouzi. Image superresolution via iterative refinement. IEEE transactions on pattern analysis and machine intelligence, 45(4):4713-4726, 2022. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 600, + 294, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 600, + 294, + 621 + ], + "spans": [ + { + "bbox": [ + 56, + 600, + 294, + 621 + ], + "type": "text", + "content": "[62] sd-vae-ftmse. https://huggingface.co/stabilityai/sd-vae-ftmse, 2024.7,8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 624, + 294, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 294, + 655 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 294, + 655 + ], + "type": "text", + "content": "[63] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 658, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 294, + 689 + ], + "type": "text", + "content": "[64] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. Advances in neural information processing systems, 32, 2019. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 692, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 692, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 692, + 294, + 712 + ], + "type": "text", + "content": "[65] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "text", + "content": "generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 97, + 553, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 97, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 316, + 97, + 553, + 140 + ], + "type": "text", + "content": "[66] Matthew Tancik, Ben Mildenhall, and Ren Ng. Stegastamp: Invisible hyperlinks in physical photographs. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2117-2126, 2020. 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 141, + 553, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 553, + 184 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 553, + 184 + ], + "type": "text", + "content": "[67] Michael Theis, Randall F Trzeciak, Daniel L Costa, Andrew P Moore, Sarah Miller, Tracy Cassidy, and William R Claycomb. Common sense guide to mitigating insider threats. 2019. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 186, + 553, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 186, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 316, + 186, + 553, + 228 + ], + "type": "text", + "content": "[68] Zhou Wang, A.C. Bovik, H.R. Sheikh, and E.P. Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing, 13(4): 600-612, 2004. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 232, + 553, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 232, + 553, + 274 + ], + "spans": [ + { + "bbox": [ + 316, + 232, + 553, + 274 + ], + "type": "text", + "content": "[69] Michael R Watson, Angelos K Marnerides, Andreas Mauthe, David Hutchison, et al. Malware detection in cloud computing infrastructures. IEEE Transactions on Dependable and Secure Computing, 13(2):192-205, 2015. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 277, + 553, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 277, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 316, + 277, + 553, + 319 + ], + "type": "text", + "content": "[70] Yuxin Wen, John Kirchenbauer, Jonas Geiping, and Tom Goldstein. Tree-rings watermarks: Invisible fingerprints for diffusion images. Advances in Neural Information Processing Systems, 36, 2024. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 322, + 553, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 322, + 553, + 365 + ], + "spans": [ + { + "bbox": [ + 316, + 322, + 553, + 365 + ], + "type": "text", + "content": "[71] Cheng Xiong, Chuan Qin, Guorui Feng, and Xinpeng Zhang. Flexible and secure watermarking for latent diffusion model. In Proceedings of the 31st ACM International Conference on Multimedia, pages 1668-1676, 2023. 1, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 367, + 553, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 367, + 553, + 410 + ], + "spans": [ + { + "bbox": [ + 316, + 367, + 553, + 410 + ], + "type": "text", + "content": "[72] Ling Yang, Zhilong Zhang, Yang Song, Shenda Hong, Runsheng Xu, Yue Zhao, Wentao Zhang, Bin Cui, and Ming-Hsuan Yang. Diffusion models: A comprehensive survey of methods and applications, 2024. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 412, + 553, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 412, + 553, + 475 + ], + "spans": [ + { + "bbox": [ + 316, + 412, + 553, + 475 + ], + "type": "text", + "content": "[73] Zhengyuan Yang, Jianfeng Wang, Zhe Gan, Linjie Li, Kevin Lin, Chenfei Wu, Nan Duan, Zicheng Liu, Ce Liu, Michael Zeng, et al. Reco: Region-controlled text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14246-14255, 2023. 1, 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 479, + 553, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 479, + 553, + 533 + ], + "spans": [ + { + "bbox": [ + 316, + 479, + 553, + 533 + ], + "type": "text", + "content": "[74] Shengfang Zhai, Yinpeng Dong, Qingni Shen, Shi Pu, Yuejian Fang, and Hang Su. Text-to-image diffusion models can be easily backdoored through multimodal data poisoning. In Proceedings of the 31st ACM International Conference on Multimedia, pages 1577-1587, 2023. 2, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 535, + 553, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 535, + 553, + 576 + ], + "spans": [ + { + "bbox": [ + 316, + 535, + 553, + 576 + ], + "type": "text", + "content": "[75] Kevin Alex Zhang, Lei Xu, Alfredo Cuesta-Infante, and Kalyan Veeramachaneni. Robust invisible video watermarking with attention. arXiv preprint arXiv:1909.01285, 2019. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 580, + 553, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 580, + 553, + 623 + ], + "spans": [ + { + "bbox": [ + 316, + 580, + 553, + 623 + ], + "type": "text", + "content": "[76] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836-3847, 2023. 1, 2, 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 624, + 553, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 553, + 678 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 553, + 678 + ], + "type": "text", + "content": "[77] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 4" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 681, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 553, + 712 + ], + "type": "text", + "content": "[78] Shihao Zhao, Dongdong Chen, Yen-Chun Chen, Jianmin Bao, Shaozhe Hao, Lu Yuan, and Kwan-Yee K Wong. Uni-controlnet: All-in-one control to text-to-image diffusion" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "8223" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 138 + ], + "type": "list", + "angle": 0, + "index": 2, + "blocks": [ + { + "bbox": [ + 75, + 72, + 296, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 72, + 296, + 95 + ], + "spans": [ + { + "bbox": [ + 75, + 72, + 296, + 95 + ], + "type": "text", + "content": "models. Advances in Neural Information Processing Systems, 36, 2024. 1, 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 296, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 296, + 138 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 296, + 138 + ], + "type": "text", + "content": "[79] Yunqing Zhao, Tianyu Pang, Chao Du, Xiao Yang, Ngai-Man Cheung, and Min Lin. A recipe for watermarking diffusion models. arXiv preprint arXiv:2303.10137, 2023. 2, 3, 5" + } + ] + } + ], + "index": 1 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "8224" + } + ] + } + ], + "index": 3 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/SlideChat_ A Large Vision-Language Assistant for Whole-Slide Pathology Image Understanding/5542bef9-2edf-4ac5-b8df-cf3f63aeaa6c_content_list.json b/2025/SlideChat_ A Large Vision-Language Assistant for Whole-Slide Pathology Image Understanding/5542bef9-2edf-4ac5-b8df-cf3f63aeaa6c_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..ff6ce726ace0f3b612fad3e9dab8e497779d59d1 --- /dev/null +++ b/2025/SlideChat_ A Large Vision-Language Assistant for Whole-Slide Pathology Image Understanding/5542bef9-2edf-4ac5-b8df-cf3f63aeaa6c_content_list.json @@ -0,0 +1,1218 @@ +[ + { + "type": "text", + "text": "SlideChat: A Large Vision-Language Assistant for Whole-Slide Pathology Image Understanding", + "text_level": 1, + "bbox": [ + 102, + 138, + 893, + 185 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ying Chen $^{1,2*}$ Guoan Wang $^{1,3*}$ Yuanfeng Ji $^{4*†}$ Yanjun Li $^{1,3}$ Jin Ye $^{1,5}$ Tianbin Li $^{1}$ Ming Hu $^{1,5}$ Rongshan Yu $^{2}$ Yu Qiao $^{1}$ Junjun He $^{1\\dagger}$", + "bbox": [ + 143, + 209, + 849, + 247 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Shanghai AI Laboratory $^{2}$ Xiamen University $^{3}$ East China Normal University $^{4}$ Stanford University $^{5}$ Monash University", + "bbox": [ + 179, + 261, + 815, + 297 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 254, + 330, + 330, + 345 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite the progress made by multimodal large language models (MLLMs) in computational pathology, they remain limited by a predominant focus on patch-level analysis, missing essential contextual information at the whole-slide level. The lack of large-scale instruction datasets and the gigapixel scale of whole slide images (WSIs) pose significant developmental challenges. In this paper, we present SlideChat, the first vision-language assistant capable of understanding gigapixel whole-slide images, exhibiting excellent multimodal conversational capability and response complex instruction across diverse pathology scenarios. To support its development, we created SlideInstruction, the largest instruction-following dataset for WSIs consisting of 4.2K WSI captions and 176K VQA pairs with multiple categories. Furthermore, we propose SlideBench, a multimodal benchmark that incorporates captioning and VQA tasks to assess SlideChat's capabilities in various settings such as microscopy, diagnosis and clinical. Compared to both general and specialized MLLMs, SlideChat exhibits exceptional capabilities, achieving state-of-the-art performance on 18 of 22 tasks. For example, it achieved an overall accuracy of $81.17\\%$ on SlideBench-VQA (TCGA), and $54.15\\%$ on SlideBench-VQA (BCNB). Our code, data, and model is publicly accessible at https://uni-medical.github.io/SlideChat.github.io.", + "bbox": [ + 99, + 361, + 483, + 731 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 101, + 757, + 230, + 772 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Computational pathology aims to improve the analysis of digitized tissue samples [14, 24], such as whole slide images (WSIs), by applying artificial intelligence to aid in the diagnosis, identification, and understanding of disease. Recently, the development of this field has gained rapid", + "bbox": [ + 99, + 781, + 483, + 856 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/334f9be18accb27a827000a8d6344fe38babd823e20549a014fd4917b3a36b4f.jpg", + "image_caption": [ + "A", + "B", + "Figure 1. SlideChat is the first large vision-language assistant specifically designed for whole-slide pathology analysis. SlideChat can generate comprehensive descriptions of whole-slide images and provide contextually relevant responses across various applications." + ], + "image_footnote": [], + "bbox": [ + 535, + 335, + 620, + 378 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Summarize the key pathological findings from the whole slide image.", + "bbox": [ + 535, + 386, + 673, + 402 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "This pathological examination reveals the high-grade urothelial carcinoma with ...", + "bbox": [ + 545, + 410, + 704, + 426 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "What is the pathological stage (pT) of the transitional cell carcinoma? A. pT1 B. pT2 C. pT3 D. pT4", + "bbox": [ + 535, + 436, + 679, + 460 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "B. pT2", + "bbox": [ + 656, + 464, + 684, + 473 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Microscopy", + "bbox": [ + 736, + 342, + 782, + 351 + ], + "page_idx": 0 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "$\\spadesuit$ Tissue Architecture $\\spadesuit$ Tumor Characteristics", + "Histopathological Changes", + "$\\spadesuit$ Cytomorphological Characteristics" + ], + "bbox": [ + 728, + 353, + 883, + 378 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Diagnosis", + "bbox": [ + 736, + 391, + 777, + 400 + ], + "page_idx": 0 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "$\\spadesuit$ Disease Detection Staging Grading", + "$\\spadesuit$ Disease Classification", + "$\\spadesuit$ Differential Diagnosis" + ], + "bbox": [ + 728, + 402, + 883, + 426 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Clinical", + "bbox": [ + 736, + 439, + 764, + 446 + ], + "page_idx": 0 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "Treatment Guidance Risk Factors", + "Prognostic Assessment", + "$\\spadesuit$ Biomarker Analysis" + ], + "bbox": [ + 728, + 448, + 852, + 473 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "momentum, mainly driven by breakthroughs in the visual foundation model [9, 27, 30]. These models learn generalized representations by pre-training on large-scale data and perform well in various downstream tasks, such as cancer detection and biomarker prediction. Building on this base, integration with the powerful Large Language Models (LLMs) further advances the development of the Multimodal Large Language Model (MLLMs) [20], which has made great strides in responding to more complex, open-ended visual queries, enabling it to serve as a versatile assistant at various stages of medical care, including clinical decision-making, education, and research.", + "bbox": [ + 509, + 590, + 895, + 767 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Nevertheless, there are some major challenges that hinder the development and use of pathology MLLMs for real-world applications. First, it is challenging to develop a MLLM architecture that can effectively be capable of gigapixel whole slides (e.g., $100,000 \\times 100,000$ pixels). Existing models [20, 22, 26] often process whole slides by extracting small patch/ROI-level data for subsequent analysis, resulting in a limited understanding of global slide con", + "bbox": [ + 509, + 771, + 895, + 890 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*These authors contributed equally to this work.", + "bbox": [ + 124, + 864, + 374, + 876 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "†Corresponding authors: yfj@stanford.edu, hejunjun@pjlab.org.cn.", + "bbox": [ + 124, + 877, + 472, + 888 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "5134", + "bbox": [ + 482, + 939, + 514, + 952 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "text and suboptimal performance in some complex pathological analysis. Second, publicly available multi-modal pathology slide datasets are relatively scarce and of varying quality [7, 8, 12], which limits the development of MLLMs trained on such data.", + "bbox": [ + 99, + 102, + 480, + 175 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we present SlideChat (see Fig. 1), the first open-source vision-language assistant capable of understanding gigapixel whole-slide images. First, SlideChat is trained on SlideInstruction, a large-scale multi-modal instruction dataset encompassing data from The Cancer Genome Atlas (TCGA) [15] via our specifically designed data processing pipeline (see Fig. 3 (A)). SlideInstruction contains 4,181 WSI-caption pairs and 175,753 visual question-answer pairs from 3,294 patients, covering 10 cancer types. The question-answer pairs include both open-ended and closed-ended questions, further divided into 13 subcategories, covering a diverse range of clinical tasks. SlideInstruction is more than 20 times larger than previous public instruction datasets in the number of instructions. Second, we propose SlideChat, a novel architecture in LLaVA style for capable multi-modal analysis of gigapixel whole slides. A gigapixel whole slide is first divided into a series of patches, each of which is individually processed by a patch-level encoder to extract local features. The resulting long sequence of patch tokens is then processed by a slide-level encoder employing sparse attention to enhance the slide-level features. Finally, enhanced patch tokens are fed into the LLM via a projector, which processes user queries and generates responses.", + "bbox": [ + 99, + 180, + 482, + 532 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To systematically evaluate the performance of SlideChat in real-world scenarios, we establish a comprehensive digital pathology benchmark named SlideBench, encompassing more than 20 clinical tasks, using data from both TCGA and the in-the-wild Early Breast Cancer CoreNeedle Biopsy (BCNB) dataset. This resulted in three test sets: SlideBench-Caption, comprising 734 WSI-caption pairs; SlideBench-VQA (TCGA), comprising 7,827 VQA pairs covering 13 different tasks; and SlideBench-VQA (BCNB), including a total of 7,274 VQA entries from 1,058 patients, covering seven different tasks. Additionally, we compare the performance of SlideChat on another externally proposed dataset, WSI-VQA [8], to further validate its effectiveness. We compare our model with the currently available state-of-the-art general and medical-specialized MLLMs including GPT-4o, LLaV-aMed [16], MedDr [13] and Quilt-LLaVA [22]. Benefiting from large-scale, high-quality training and effective local-global context modeling, SlideChat achieves state-of-the-art performance on 18 out of 22 tasks, with significant improvements over the second-best method $10\\%$ on 9 tasks on four benchmarks. Specifically, SlideChat achieves an average accuracy improvement of $13.47\\%$ over the second-best model on SlideBench-VQA (TCGA), an average improvement of $12.59\\%$ on", + "bbox": [ + 99, + 537, + 482, + 890 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "SlideBench-VQA (BCNB), and an improvement of $5.82\\%$ on WSI-VQA. Finally, to accelerate research progress in digital pathology, we make SlideChat fully open-weight, including source code and model weights as well as instruction and benchmark datasets. The key contributions are summarized four-fold in the following:", + "bbox": [ + 511, + 102, + 893, + 189 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We develop SlideChat, the first vision-language assistant capable of understanding gigapixel whole-slide images, achieving state-of-the-art performance on multiple benchmarks.", + "- We create SlideInstruction, a largest comprehensive WSI instruction-following dataset containing 4.2K WSI-caption pairs and 176K VQA pairs.", + "- We establish SlideBench, a WSI multi-modal benchmark comprising SlideBench-Caption, SlideBench-VQA (TCGA), and SlideBench-VQA (BCNB), covering 21 different clinical tasks.", + "- We release SlideChat, SlideInstruction and SlideBench as open-source resources to facilitate research and development in computational pathology." + ], + "bbox": [ + 531, + 190, + 893, + 395 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 511, + 407, + 656, + 422 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Whole Slide Image Analysis Whole slide images are pivotal in modern pathology, enabling comprehensive analysis of tissue samples for tasks such as predicting patient prognosis, classifying cancer subtypes, and identifying biomarkers [17, 18, 23-25, 33]. Recent studies have leveraged pathology foundational models [3, 28, 31] to enhance WSIs analysis, either through fine-tuning for specific downstream tasks or by employing zero-shot prediction approaches in CLIP [21] style. Although these models are effective in task-specific applications, their reliance on fine-tuning or limited zero-shot capabilities restricts their generalizability across diverse and complex user instructions.", + "bbox": [ + 511, + 431, + 893, + 608 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "MLLMs in Computational Pathology The paradigm of MLLMs enables them to effectively respond to more complex, open-ended visual queries while processing pathology images, thus providing significant value across various medical stages. PathChat [20] is a vision-language assistant designed for pathology, developed with 450K private instruction pairs to handle both visual and natural language queries. Quilt-Instruct [22] is a large-scale dataset comprising 107K question-answer pairs. Building on Quilt-Instruct, Quilt-LLAVA [22] is a model designed for diagnostic reasoning across multiple image patches, leveraging its extensive question-answer pairs to accurately interpret complex H&E images. PathAsst [26] combines a pathology-specific CLIP model with Vicuna-13b [10] to create a multimodal generative foundational model tailored for pathology. However, current MLLMs primarily only focus on patch or ROI data, limiting their utility for slide-level clinical applications where broader contextual understanding is crucial.", + "bbox": [ + 511, + 625, + 893, + 890 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "5135", + "bbox": [ + 482, + 939, + 514, + 952 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/469b18531b509cecfc8ab5beb7f7537e3b21621b056d4c05c0aff9487f21fdaa.jpg", + "image_caption": [ + "Figure 2. Overview of our SlideChat. (A) SlideChat serializes each input WSI into a sequence of $224 \\times 224$ patches, converting each into visual embeddings with a patch-level encoder. A slide-level encoder then interacts with these features to generate contextual embeddings. Then, a multimodal projector maps the visual features from the slide-level encoder into a unified space, aligned seamlessly with the LLM. (B) SlideChat was trained for two stages: Cross-Domain Alignment and Visual Instruction Learning." + ], + "image_footnote": [], + "bbox": [ + 107, + 103, + 883, + 305 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. SlideChat", + "text_level": 1, + "bbox": [ + 101, + 393, + 207, + 407 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Architecture", + "text_level": 1, + "bbox": [ + 101, + 422, + 232, + 436 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To achieve the goal of analyzing gigapixel whole-slide images in a multimodal setting, as shown in Fig. 2, SlideChat consists of four key designs: the patch-level encoder, the slide-level encoder, the multimodal projector module, and the large language model. Our method starts by partitioning the WSI into smaller $224 \\times 224$ pixel patches, making it computationally feasible to process such large images. These patches are then passed through a well-trained, frozen patch-level encoder [19], which extracts localized features from each individual patch, capturing fine-grained details such as cellular structures. Building on this, we employ LongNet [11, 30] as slide-level encoder to enhance the patch-level embeddings and capture global patterns across the entire slide. This encoder uses sparse attention mechanisms to learn both local and global contextual information, enabling the model to perceive intricate local features while capturing the broader context, which is critical for comprehensive pathological assessments. Following the slide-level encoding, SlideChat incorporates a multimodal projector that maps these patch tokens into a unified space aligned with the LLM. This ensures that the visual features extracted from the WSIs are effectively transformed into representations compatible with the language model, facilitating seamless integration and interaction between visual and textual data. Concurrently, the model accepts natural language instructions from users, such as \"What is the type of tumor in the image?\". These textual queries are processed by the LLM, which comprehends the textual input and integrates it with the visual features extracted from the WSIs, enabling accurate and contextually relevant diagnoses", + "bbox": [ + 99, + 448, + 485, + 890 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "tic responses. This multimodal reasoning capability allows SlideChat to provide accurate and contextually relevant answers to complex pathology-related questions, thereby supporting clinical decision-making, education, and research across various medical stages.", + "bbox": [ + 511, + 393, + 895, + 469 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2.Data", + "text_level": 1, + "bbox": [ + 511, + 483, + 584, + 497 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "SlideInstruction There is a notable lack of large-scale multimodal pathology datasets supporting the training of vision-language assistants for whole-slide image understanding. To support the training of SlideChat, we develop SlideInstruction, a comprehensive instruction dataset, sourced from the TCGA database, comprising 4,915 WSI-report pairs from 4,028 patients. Fig. 3 illustrates our entire data curation pipeline. We initially prompt GPT-4 to refine the pathology reports, and clean up the noise in the report including unrelated symbols, technical details of pathology department procedures, specimen handling and processing information, redundant administrative or legal statements, and some repeated information. For the refined pathology reports, we further employ GPT-4 to generate high-quality multimodal data, comprising two main components: (1) WSI-Caption Data: We craft concise, clinically relevant summaries for each whole slide image by prompting the language model to extract key pathological findings. These summaries were structured into coherent paragraphs that highlighted crucial clinical details such as diagnostic results, tumor characteristics, margin status, and lymph node involvement, ensuring the caption dataset is both focused and informative. (2) WSI Instruction-Following Data: To enhance the model's ability to follow instructions and improve its comprehension of pathology images, we leveraged GPT-4 to generate tailored question", + "bbox": [ + 509, + 507, + 895, + 891 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "5136", + "bbox": [ + 482, + 939, + 514, + 952 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/f0daefbfd8e147bffc5e8a200b3f1c7ba6517ebcc8d2cab7bf276c51d24f595b.jpg", + "image_caption": [ + "A" + ], + "image_footnote": [], + "bbox": [ + 156, + 114, + 866, + 251 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/87b8387cb638bbb9b4ca94bda3ad892e90ad4a42a6350df49a6db0e76e996a37.jpg", + "image_caption": [ + "B" + ], + "image_footnote": [], + "bbox": [ + 151, + 271, + 823, + 335 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/f97990cab839f83140c393912fd67655dc93a45953901eca9e8bd6e2717d5d57.jpg", + "image_caption": [ + "C" + ], + "image_footnote": [], + "bbox": [ + 161, + 367, + 277, + 424 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/b127e927e5f640d0d6d859e17226266b5adfd02d8cce2a44dcd32e4c6d80f4f7.jpg", + "image_caption": [ + "Figure 3. (A) Overview of the SlideInstruction generation pipeline. We prompt GPT-4 to extract the WSI-Caption, Open-set VQA and Closed-set VQA from reports. (B) For the generated Closed-set VQA, we employ LLMs to filter low-quality QA pairs and involve pathologists for validation, resulting in the creation of SlideBench-VQA. (C) Examples of WSI caption and instruction-following scenarios in microscopy, diagnostics, and clinical applications. For additional examples, please refer to Supplementary Material." + ], + "image_footnote": [], + "bbox": [ + 163, + 436, + 276, + 493 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "[Caption] #Q:Describe the whole slide image. #A:The bladder exhibits an invasive poorly differentiated urothelial carcinoma, grade 4/4, extending through the bladder neck wall....", + "bbox": [ + 297, + 375, + 482, + 419 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "[VQA-Diagnosis] \n#Q: From the given whole slide image, which histological subtype of lung adenocarcinoma is most predominantly observed? \n#A: It is acinar subtype observed.", + "bbox": [ + 297, + 443, + 480, + 488 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/748388969461b7e9b57d20c3a086a33fcff6685ce6b440d446124d5b332d905b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 531, + 369, + 640, + 425 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/cd2511ed303e1a629613496c681fe0adb36bfa337b7ac2f4e9595832e708f76f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 527, + 433, + 635, + 496 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "[VQA-Microscopy] \n#Q: What histological feature observable in a whole slide image differentiates anaplastic astrocytoma from glioblastoma? \n#A: Absence of endothelial proliferation.", + "bbox": [ + 661, + 376, + 846, + 420 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "[VQA-Clinical] #Q: Based on the histopathological features, what is the primary recommended treatment? #A: Simple mastectomy followed by hormone therapy.", + "bbox": [ + 661, + 444, + 846, + 487 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "and-answer pairs for each WSI report. Drawing inspiration from PathChat [20], we structure these questions into 3 broad categories—microscopy, diagnosis, and clinical considerations—which represent key stages in the pathology workflow, and 13 narrow categories focusing on specific aspects within each stage (Fig. 1 B). To construct a comprehensive instructional dataset, we began by generating two open-ended and two closed-ended QA pairs for each report within each specific category, taking advantage of the rich content found in the majority of TCGA reports. Then, some QA pairs were excluded, such as those lacking corresponding category information. Regarding the train/test split, it is worth noting that the WSI-report datasets from TCGA include 2 types: (a) one report linked to multiple WSIs, and (b) one report linked to a single WSI. For type (a), where specific diagnostic details may not align perfectly with each WSI, we include all WSIs in the training set to introduce some \"noisy data\", which can enhance model robustness. For type (b), $80\\%$ of WSIs are allocated to the training", + "bbox": [ + 99, + 598, + 483, + 877 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "set and $20\\%$ to the test set. Finally, there are 4,181 WSIs for training and 734 WSIs for testing. Consequently, we construct a large-scale training set named SlideInstruction, comprising 4,181 WSI captions and 175,753 instruction-following VQA pairs across 13 narrow categories.", + "bbox": [ + 511, + 598, + 893, + 672 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "SlideBench To systematically evaluate the performance of SlideChat, we include the remaining 734 WSI captions from the test set, along with a substantial number of closed-set VQA pairs, to establish a comprehensive evaluation benchmark. First, we construct a set named SlideBench-Caption based on the WSI-Caption data to evaluate the model's ability to generate accurate and coherent descriptions of whole slide images. Secondly, we construct SlideBench-VQA (TCGA) based on closed-set VQA pairs along with test WSIs, aiming to evaluate various aspects of model performance. As shown in Fig. 3 (B), to improve the quality of the testing benchmarks, we employ four advanced large language models, including GPT-4 [2], InternLM2-", + "bbox": [ + 509, + 698, + 895, + 890 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "5137", + "bbox": [ + 482, + 939, + 514, + 952 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Chat-7B [6], Qwen-7B-Chat [5], and DeepSeek-7B-Chat, to filter closed-set VQAs by predicting answers based solely on the question text. Any questions for which at least three of these models provided correct answers are subsequently excluded. Following this automated filtering, five expert pathologists are invited to review and amend the remaining questions. The review process is guided by the following criteria: (1) Whether the correct answer necessitates image interpretation; (2) Whether the question and its corresponding answer are logically and coherently structured; and (3) Whether the question aligns appropriately with the designated broad and narrow categories. VQA pairs failing to meet these criteria are excluded by the pathologists. Consequently, the SlideBench-VQA (TCGA) comprises 7,827 VQAs across 13 categories, with some examples illustrated in Fig. 3 C. Additionally, we incorporate the in-the-wild Early Breast Cancer Core-Needle Biopsy (BCNB) WSI dataset [29], which encompasses a diverse patient population and a variety of clinical task labels, to enhance the test set benchmark and more comprehensively assess the model's generalization capabilities. In detail, we convert the BCNB dataset into a VQA format by rephrasing the classification objectives into a specific template as questions, while transforming the original multi-class labels into selectable options, and integrating it into SlideBench as an external subset, named SlideBench-VQA (BCNB). This dataset comprises 7,247 VQA pairs from 1,058 patients, specifically designed to evaluate SlideChat's zero-shot generalization capability across 7 distinct classification tasks.", + "bbox": [ + 99, + 102, + 485, + 529 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Two-stage Training", + "text_level": 1, + "bbox": [ + 101, + 537, + 282, + 555 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Stage 1: Cross-Domain Alignment. SlideChat adopts a two-stage training approach (see Fig. 2 B). In the first stage, the primary objective is to align the LLM's word embeddings with the visual features extracted from whole slide images. This alignment enables the LLM to interpret visual representations from the slide-level encoder, facilitating the effective utilization of the intricate features within the slides. During this stage, SlideChat is trained to generate descriptive captions using 4.2K WSI-caption pairs from SlideInstruction. Specifically, only the slide-level encoder and projection matrix are updated, while the patch-level encoder and LLM weights remain fixed.", + "bbox": [ + 99, + 560, + 483, + 736 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Stage 2: Visual Instruction Learning. In the second stage, we focus on visual question-answering tasks to train the model to accurately respond to domain-specific questions concerning whole slide images. During this phase, the model develops the ability to handle a broad range of multimodal instructions, enabling it to generate answers by effectively integrating both visual and textual information. For example, the model must perform various pathology tasks, such as describing the extent of tumor invasion or assess", + "bbox": [ + 99, + 756, + 482, + 890 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "ing the degree of cellular differentiation. To accomplish this, we utilize 176K WSI VQAs from SlideInstruction in the second training stage, allowing the slide encoder, projection layer, and large language model components to be fully trainable to ensure comprehensive adaptability. This approach significantly enhances the model's capability to handle diverse pathology-related tasks, thereby increasing its effectiveness in real-world clinical and research settings.", + "bbox": [ + 509, + 102, + 895, + 220 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiment", + "text_level": 1, + "bbox": [ + 511, + 233, + 633, + 250 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We conducted the following experiments to evaluate three key aspects of SlideChat: (1) its whole slide image captioning capability, which assesses proficiency in generating descriptive captions that accurately summarize the critical pathological features of a WSI; (2) its visual question-answering (VQA) ability across various complex pathological scenarios and its generalizability in zero-shot settings; and (3) SlideChat's ability to process gigapixel WSIs, capturing both essential global context and intricate details, thereby enhancing its performance compared to patch-level MLLMs. For WSI captioning baselines, we benchmark against MI-Gen [7], a state-of-the-art method specifically designed for this task. Given that existing MLLMs cannot handle the gigapixel scale of whole slide images, we establish baseline comparisons using two approaches: (1) randomly selecting 30 patches from each WSI and inputting them into MLLMs (e.g., GPT-4o [2], LLaVA-Med [16], MedDr [16], Quilt-LLaVA [22]), followed by a majority voting scheme to generate slide-level predictions; and (2) directly inputting a WSI thumbnail, resized to $1024 \\times 1024$ pixels, referred to as Slide (T), into the MLLMs. For VQA tasks, we further evaluate performance by comparing against random prediction baselines and text-only models, thereby assessing the incremental contribution of visual information. Unless otherwise specified, SlideChat is configured with the patch-level encoder CONCH [19], the slide-level encoder LongNet [11], and utilizes the Qwen2.5-7B-Instruct [32] as LLM.", + "bbox": [ + 511, + 258, + 895, + 667 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/74276fbbf5e078c5054a9c788306991a2476e62abf699311d3bfeb5768842dea.jpg", + "table_caption": [ + "Table 1. Captioning performance across different methods on SlideBench-Caption." + ], + "table_footnote": [], + "table_body": "
MethodsInputBLEU-1BLEU-2BLEU-3BLEU-4Rouge-LGPT score
GPT-4oPatch0.160.030.010.010.131.54
GPT-4oSlide (T)0.100.030.010.010.111.03
Quilt-LLaVAPatch0.300.190.110.050.183.87
Quilt-LLaVASlide (T)0.230.090.040.010.161.89
MI-GenSlide0.370.240.150.100.254.14
SlideChatSlide0.370.210.120.080.244.11
", + "bbox": [ + 516, + 719, + 893, + 811 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "SlideBench-Caption We report BLEU, ROUGE, and GPT scores to evaluate caption generation performance in Tab. 1. For the GPT score, we use GPT-4 to assess the", + "bbox": [ + 511, + 844, + 893, + 888 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5138", + "bbox": [ + 482, + 939, + 514, + 952 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/27ce1baf14cf385bf8d2447b2a552d59866571937b209cbf1466ec50ddb682e6.jpg", + "table_caption": [ + "Table 2. VQA performance with different methods." + ], + "table_footnote": [], + "table_body": "
MethodsInputSlideBench-VQA (TCGA)SlideBench-VQA (BCNB)WSI-VQA*
MicroscopyDiagnosisClinicalOverall
RandomText24.4424.9126.4425.0224.4024.14
GPT-438.2829.0945.0037.25018.60
GPT-4oPatch62.8946.6966.7757.9141.4330.41
MedDr73.3057.7874.2567.7033.6754.36
LLaVA-Med47.3432.7847.9642.0030.126.31
Quilt-LLaVA57.7635.9653.0748.0732.1944.43
GPT-4oSlide (T)38.2823.1043.4234.07014.03
MedDr70.4852.4772.8064.2535.4850.95
LLaVA-Med45.8227.5840.8437.39018.79
Quilt-LLaVA49.1226.9744.7539.3941.5535.40
SlideChatSlide87.64(+14.34)73.27(+15.49)84.26(+10.01)81.17(+13.47)54.14(+12.59)60.18(+5.82)
", + "bbox": [ + 130, + 125, + 867, + 348 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "similarity between the generated captions and the ground truth, providing an overall score on a scale of 1 to 10, with higher scores indicating better performance. When utilizing patch-level inputs, GPT-4o generates individual descriptions for each patch, which are subsequently integrated to create the final slide-level caption. However, this approach yields poor performance, as evidenced by a BLEU-1 score of 0.16 and a GPT-score of 1.54. These results suggest that the patch-based method fails to adequately capture the broader context necessary for accurate WSI captioning. When the WSI thumbnail of size $1024 \\times 1024$ pixels is used as input to GPT-4o, performance decreases further, with a BLEU-1 score of 0.10 and a GPT-score of 1.03. This suggests that while the thumbnail offers a global view of the slide, it may lack the resolution and detail necessary for generating precise and informative captions. In contrast, MI-Gen, a model specifically designed for WSI captioning, demonstrates significantly superior performance across all metrics, achieving a BLEU-1 score of 0.37 and a GPT score of 4.14. Similarly, SlideChat, shows competitive results with a BLEU-1 score of 0.37 and a GPT score of 4.11. These outcomes highlight SlideChat's ability to effectively integrate both local and global information from the slides and confirm its efficiency in describing whole-slide images.", + "bbox": [ + 99, + 373, + 482, + 724 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "SlideBench-VQA (TCGA) We further evaluate SlideChat's overall performance on the multiple-choice VQA benchmark. The results on SlideBench-VQA (TCGA) in Tab. 2, compare different methods across three categories: microscopy, diagnosis, and clinical, along with an overall performance score. Random selection achieves an overall score of $25.02\\%$ accuracy, serving as a baseline for answer distribution but demonstrating poor performance. While GPT-4, relying solely on text input, outperforms random predictions, it continues to struggle", + "bbox": [ + 99, + 742, + 482, + 890 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "with accurately answering questions. When GPT-4o incorporates patch-level inputs, its performance improves markedly, reaching a score of $57.91\\%$ and underscoring the crucial role of detailed visual data. However, using a WSI thumbnail results in a lower score of $34.07\\%$ , as the reduced detail restricts its ability to deliver precise answers. MedDr performs well, achieving an overall score of $67.70\\%$ with patch-level inputs, though this drops slightly to $64.25\\%$ when using the slide thumbnail due to the loss of fine visual details. SlideChat outperforms all other methods, attaining a leading overall accuracy of $81.17\\%$ , excelling across all categories and significantly surpassing the competition. Even in more fine-grained pathological scenarios, as depicted in the left portion of Fig. 4, SlideChat remains the top-performing model across 13 tasks, particularly in areas such as cytomorphological characteristics, histopathological changes, disease detection, disease classification, and staging and grading, which require the identification of complex pathological visual features. Compared with baselines taking some patches or slide thumbnail as inputs, SlideChat has the capability to analyze a significantly greater number of pathological features with enhanced detail, effectively capturing both localized features and overarching global patterns, allowing SlideChat to provide more accurate and nuanced insights into pathological variations.", + "bbox": [ + 509, + 373, + 893, + 755 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "SlideBench-VQA (BCNB) In the zero-shot VQA setting, SlideChat significantly outperforms all other models, achieving the highest overall score of $54.14\\%$ on SlideBench-VQA (BCNB). It particularly excels in identifying tumor types, far surpassing other baselines in this task. This performance highlights SlideChat's generalization capability across a wide range of tasks. When taking patches as inputs, GPT-4o outperforms both MedDr and LLaVA", + "bbox": [ + 509, + 771, + 893, + 890 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "5139", + "bbox": [ + 482, + 939, + 514, + 950 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/977a5e617e035189e47c3f27f96daab07f3564ef79abc62b7cccc95c21c30d96.jpg", + "image_caption": [ + "Figure 4. Accuracy on different tasks in SlideBench-VQA (TCGA) (left) and SlideBench-VQA (BCNB) (right)." + ], + "image_footnote": [], + "bbox": [ + 179, + 99, + 803, + 270 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Med, achieving a score of $41.43\\%$ , though it still falls short of SlideChat by $12.59\\%$ . Notably, GPT-4o and LLaVA-Med performed very poorly, achieving a score of zero across all tasks when evaluated using slide thumbnails from this testing set. MedDr and Quilt-LLaVA both achieve moderate results, with MedDr scoring $33.67\\%$ on patches and $35.48\\%$ on thumbnails, while Quilt-LLaVA attains $32.19\\%$ on patches and $41.55\\%$ on thumbnails. This outcome highlights that, for complex WSIs, relying solely on relatively sufficient visual features is inadequate for effectively supporting a range of tasks. In the more fine-grained pathological tasks of the BCNB benchmark, SlideChat attains state-of-the-art performance on 5 out of 7 tasks (Fig. 4), further demonstrating the effectiveness of SlideChat.", + "bbox": [ + 99, + 325, + 483, + 531 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "WSI-VQA* We also curated a subset of closed-set VQA pairs from the public WSI-VQA [8] dataset, referred to as WSI-VQA*, which exclusively consists of samples that were not in the training set, to evaluate the model's performance. SlideChat demonstrates the highest performance with a score of 60.18. Although MedDr performs well with both patch inputs (54.36%) and slide thumbnail inputs (50.95%), it still falls short compared to SlideChat. GPT-4o struggles significantly, especially with slide thumbnails, scoring only 14.03%, which highlights the limitations of using lower-resolution inputs. SlideChat leverages both detailed local information and broader context to process WSIs, making it the most effective model for this benchmark. This further emphasizes its superior capability in handling whole-slide data for VQA tasks.", + "bbox": [ + 99, + 547, + 483, + 770 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Ablation We performed ablation experiments from different perspectives as follows: a) Large Language Model Comparison: Firstly, we compare the performance of several large language models, each with a parameter scale of approximately 7 billion, including Vicuna-7B-v1.5 [10], Phi-3-Mini-4k-Instruct [1], Qwen1.5-7B-Chat [5], Llama3-8B-Instruct [4], InternLM2-Chat-7B [6], and Qwen2.5-7B", + "bbox": [ + 99, + 786, + 483, + 890 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/8ebed14e972142fc4554727ea1efffcffa0d44c5672ed1f379ce76c97723e136.jpg", + "table_caption": [ + "Table 3. Performance Comparison of LLMs and Slide Encoder on WSI Captioning and VQA Tasks." + ], + "table_footnote": [], + "table_body": "
LLMsSlide EncoderCaptionVQA (TCGA)VQA (BCNB)WSI-VQA*
Vicuna-7B-v1.5✓3.2841.4341.4331.98
Phi-3-Mini-4k-Instruct✓2.6679.9343.9260.18
Qwen1.5-7B-Chat✓2.9277.6344.0756.89
Llama3-8B-Instruct✓3.3078.7842.8255.25
Internlm2-Chat-7B✓3.3079.1052.1356.76
Qwen2.5-3B-Instruct✓3.4080.3245.7956.38
Qwen2.5-14B-Instruct✓3.3982.1451.5760.94
Qwen2.5-7B-Instruct✓4.1181.1754.1460.18
Qwen2.5-7B-Instruct×3.9581.2145.4959.67
", + "bbox": [ + 516, + 373, + 890, + 508 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Instruct [32]. Specifically, we measured their performance on the SlideBench-Caption using the GPT-score and their accuracy on three VQA datasets. As shown in Tab. 3, the results demonstrate that SlideChat, powered by the Qwen2.5-7B-Instruct model, achieved the highest performance across all tasks, particularly excelling in the captioning task. These findings underscore the significant potential of developing SlideChat with the Qwen2.5-7B-Instruct model. Utilizing the Qwen2.5 model, we further evaluated models of varying scales and discovered that larger models generally exhibited superior performance, particularly the 7B and 14B parameter models. While these two models showed comparable performance across the three benchmarks, the 14B model surpassed the 7B model by $2.57\\%$ on the SlideBench-VQA (TCGA). Given computational resource constraints, SlideChat uses the 7B model by default to achieve the best balance between performance and resource efficiency. b) Slide-level encoder effectiveness: We investigate the effectiveness of the slide-level encoder by initially removing it from SlideChat and employing a two-stage training approach. In the first stage, we exclusively trained the projection layers. However, this approach failed to reduce training loss or generate coherent text effectively,", + "bbox": [ + 509, + 551, + 893, + 890 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "5140", + "bbox": [ + 482, + 939, + 514, + 950 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a2ad6f0d11e37162874167c4a6b46e6971bca861a7172843b4e2405a85783298.jpg", + "image_caption": [ + "A" + ], + "image_footnote": [], + "bbox": [ + 225, + 104, + 462, + 162 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3e33e9c466c428539ae7aade1888ad947d76b9eae9b5b2481b29ffafcfab2344.jpg", + "image_caption": [ + "Select the most accurate description of the cytomorphological characteristics from image. A. Cells with increased nuclear-to-cytoplasmic ratio, hyperchromatic nuclei, and prominent nucleoli. B. Cells with uniform nuclei and abundant cytoplasm. C. Cells with minimal nuclear abnormalities. D. Cells with enlarged, clear nuclei and small nucleoli." + ], + "image_footnote": [], + "bbox": [ + 187, + 217, + 202, + 250 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "A. Cells with increased nuclear-to-cytoplasmic ratio, hyperchromatic nuclei, and prominent nucleoli.", + "bbox": [ + 225, + 232, + 401, + 248 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ab789ba476b9a5cf8bb971c935c6b7fcd188f19279163dbc5653a03a3f13afcc.jpg", + "image_caption": [ + "B" + ], + "image_footnote": [], + "bbox": [ + 545, + 104, + 810, + 218 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5c7d913f3149088674af34884c0eb1efe249a856cac6e55d2eb7590eafccf217.jpg", + "image_caption": [ + "Figure 5. Interpretability and visualization. We identify the top five patch tokens with the highest attention scores associated with the output text responses." + ], + "image_footnote": [], + "bbox": [ + 511, + 218, + 522, + 250 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "A. Dense collagen deposition with reduced cellularity", + "bbox": [ + 542, + 229, + 691, + 246 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "likely due to the difficulty of learning the complex visual features of WSIs without the slide-level encoder. When using a frozen LLM for complex text generation, a simple projection proved inadequate for effectively integrating and aligning visual and textual features. Subsequently, we consider combining data from both stages and training SlideChat without the slide-level encoder by simultaneously updating both the projection layers and the LLM. Under this paradigm, performance on SlideBench-VQA (TCGA) and WSI-VQA*, which share the same distribution as the training set, was comparable to the two-stage training configuration with the slide-level encoder. However, a significant decline is observed when evaluating SlideBench-VQA (BCNB), which originates from a different domain; overall performance dropped by over $10\\%$ (Tab. 3), indicating a substantial reduction in the model's generalization ability. Therefore, we recommend incorporating a slide-level encoder to capture the complex visual features of whole slide, as it is particularly effective for cross-domain alignment and enhances the model's generalization performance.", + "bbox": [ + 99, + 316, + 483, + 609 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Interpretability Despite SlideChat demonstrating promising results, concerns remained regarding the model's perception of large pathological slides. To further assess the model's interpretability, we calculated the correlation between the text output and specific image patches, thereby obtaining patch-level attention scores. By identifying the most significant patches, we gained insights into the precise areas the model focused on during response generation. Highlighting the most relevant regions of the tissue slides not only enhances transparency and bolsters the reliability of the model's outputs but also assists pathologists by directing attention to critical areas requiring closer scrutiny. Ultimately, such interpretability is essential for fostering trust in AI-assisted diagnostics and enhancing the precision and efficiency of clinical evaluations. As shown in Fig. 5, we are pleased to observe that the top five important patches identified by the model closely", + "bbox": [ + 99, + 638, + 483, + 890 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "corresponded with the features described in SlideChat's output. Our extraction method retrieves attention weights for patch tokens from each generated token, averaging them across layers and heads. We then identify the top five patch tokens with the highest attention weights for further analysis. For example, in Fig. 5 (A), the highlighted patches clearly emphasized regions exhibiting an increased nuclear-to-cytoplasmic ratio, hyperchromatic nuclei, and prominent nucleoli. Similarly, in Fig. 5 (B), the selected patches demonstrated areas with dense collagen deposition and reduced cellularity, as detailed in the model's response. This alignment between the highlighted image regions and the textual outputs significantly enhances the model's interpretability, providing increased confidence that it accurately captures and assesses relevant histopathological features. Such consistency deepens our understanding of the model's reasoning processes regarding pathological slides and underscores the potential for integrating these AI systems into clinical workflows with greater assurance.", + "bbox": [ + 511, + 316, + 895, + 597 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 612, + 627, + 627 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we present SlideChat, the first vision-language assistant capable of understanding gigapixel whole-slide images. Furthermore, we create SlideInstruction, a largest comprehensive WSI instruction-following dataset to develop SlideChat, as well as SlideBench, a multi-modal benchmark designed to evaluate SlideChat across diverse scenarios. SlideChat demonstrates excellent chat abilities and achieves state-of-the-art performance on 18 tasks. We bridge the gap between MLLMs and pathology images at the whole-slide level with SlideChat, and we believe this represents an advancement in the field of general medical artificial intelligence (GMAI).", + "bbox": [ + 511, + 637, + 895, + 813 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement The authors acknowledge the funding provided by the National Key Research and Development Program of China (No.2022ZD0160102).", + "bbox": [ + 511, + 834, + 893, + 876 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "5141", + "bbox": [ + 482, + 939, + 513, + 952 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 102, + 101, + 194, + 116 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Marah Abdin, Sam Ade Jacobs, Ammar Ahmad Awan, Jyoti Aneja, Ahmed Awadallah, Hany Awadalla, Nguyen Bach, Amit Bahree, Arash Bakhtiari, Harkirat Behl, et al. Phi-3 technical report: A highly capable language model locally on your phone. arXiv preprint arXiv:2404.14219, 2024. 7", + "[2] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 4, 5", + "[3] Faruk Ahmed, Andrew Sellergren, Lin Yang, Shawn Xu, Boris Babenko, Abbi Ward, Niels Olson, Arash Mohtashamian, Yossi Matias, Greg S Corrado, et al. Pathalign: A vision-language model for whole slide images in histopathology. arXiv preprint arXiv:2406.19578, 2024. 2", + "[4] AI@Meta. Llama 3 model card. 2024. 7", + "[5] Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, et al. Qwen technical report. arXiv preprint arXiv:2309.16609, 2023. 5, 7", + "[6] Zheng Cai, Maosong Cao, Haojiong Chen, Kai Chen, Keyu Chen, Xin Chen, Xun Chen, Zehui Chen, Zhi Chen, Pei Chu, et al. Internl m2 technical report. arXiv preprint arXiv:2403.17297, 2024.5, 7", + "[7] Pingyi Chen, Honglin Li, Chenglu Zhu, Sunyi Zheng, and Lin Yang. Mi-gen: Multiple instance generation of pathology reports for gigapixel whole-slide images. arXiv preprint arXiv:2311.16480, 2023. 2, 5", + "[8] Pingyi Chen, Chenglu Zhu, Sunyi Zheng, Honglin Li, and Lin Yang. Wsi-vqa: Interpreting whole slide images by generative visual question answering. arXiv preprint arXiv:2407.05603, 2024. 2, 7", + "[9] Richard J Chen, Tong Ding, Ming Y Lu, Drew FK Williamson, Guillaume Jaume, Andrew H Song, Bowen Chen, Andrew Zhang, Daniel Shao, Muhammad Shaban, et al. Towards a general-purpose foundation model for computational pathology. Nature Medicine, 30(3):850-862, 2024. 1", + "[10] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhanghao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E Gonzalez, et al. Vicuna: An open-source chatbot impressing gpt-4 with $90\\%$ * chatgpt quality. See https://vicuna.lmsys.org (accessed 14 April 2023), 2(3):6, 2023. 2, 7", + "[11] Jiayu Ding, Shuming Ma, Li Dong, Xingxing Zhang, Shaohan Huang, Wenhui Wang, Nanning Zheng, and Furu Wei. Longnet: Scaling transformers to 1,000,000,000 tokens. arXiv preprint arXiv:2307.02486, 2023. 3, 5", + "[12] Zhengrui Guo, Jiabo Ma, Yingxue Xu, Yihui Wang, Liansheng Wang, and Hao Chen. Histgen: Histopathology report generation via local-global feature encoding and cross-modal context interaction. arXiv preprint arXiv:2403.05396, 2024. 2", + "[13] Sunan He, Yuxiang Nie, Zhixuan Chen, Zhiyuan Cai, Hongmei Wang, Shu Yang, and Hao Chen. Meddr:" + ], + "bbox": [ + 102, + 125, + 482, + 888 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Diagnosis-guided bootstrapping for large-scale medical vision-language learning. arXiv preprint arXiv:2404.15127, 2024. 2", + "[14] Mahdi S Hosseini, Babak Ehteshami Bejnordi, Vincent Quoc-Huy Trinh, Lyndon Chan, Danial Hasan, Xingwen Li, Stephen Yang, Taehyo Kim, Haochen Zhang, Theodore Wu, et al. Computational pathology: a survey review and the way forward. Journal of Pathology Informatics, page 100357, 2024. 1", + "[15] Carolyn Hutter and Jean Claude Zenklusen. The cancer genome atlas: creating lasting value beyond its data. Cell, 173(2):283-285, 2018. 2", + "[16] Chunyuan Li, Cliff Wong, Sheng Zhang, Naoto Usuyama, Haotian Liu, Jianwei Yang, Tristan Naumann, Hoifung Poon, and Jianfeng Gao. Llava-med: Training a large language-and-vision assistant for biomedicine in one day. Advances in Neural Information Processing Systems, 36, 2024. 2, 5", + "[17] Hao Li, Ying Chen, Yifei Chen, Rongshan Yu, Wenxian Yang, Liansheng Wang, Bowen Ding, and Yuchen Han. Generalizable whole slide image classification with fine-grained visual-semantic interaction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11398–11407, 2024. 2", + "[18] Ming Y Lu, Tiffany Y Chen, Drew FK Williamson, Melissa Zhao, Maha Shady, Jana Lipkova, and Faisal Mahmood. Aibased pathology predicts origins for cancers of unknown primary. Nature, 594(7861):106-110, 2021. 2", + "[19] Ming Y Lu, Bowen Chen, Drew FK Williamson, Richard J Chen, Ivy Liang, Tong Ding, Guillaume Jaume, Igor Odintsov, Long Phi Le, Georg Gerber, et al. A visual-language foundation model for computational pathology. Nature Medicine, 30(3):863-874, 2024. 3, 5", + "[20] Ming Y Lu, Bowen Chen, Drew FK Williamson, Richard J Chen, Melissa Zhao, Aaron K Chow, Kenji Ikemura, Ahrong Kim, Dimitra Pouli, Ankush Patel, et al. A multimodal generative ai copilot for human pathology. Nature, pages 1-3, 2024. 1, 2, 4", + "[21] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2", + "[22] Mehmet Saygin Seyfioglu, Wisdom O Ikezogwo, Fatemeh Ghezloo, Ranjay Krishna, and Linda Shapiro. Quilt-llava: Visual instruction tuning by extracting localized narratives from open-source histopathology videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13183-13192, 2024. 1, 2, 5", + "[23] Zhuchen Shao, Yang Chen, Hao Bian, Jian Zhang, Guojun Liu, and Yongbing Zhang. Hvtsurv: Hierarchical vision transformer for patient-level survival prediction from whole slide image. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2209-2217, 2023. 2", + "[24] Andrew H Song, Guillaume Jaume, Drew FK Williamson, Ming Y Lu, Anurag Vaidya, Tiffany R Miller, and Faisal Mahmood. Artificial intelligence for digital and computa" + ], + "bbox": [ + 514, + 103, + 893, + 888 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "5142", + "bbox": [ + 482, + 940, + 514, + 952 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "tional pathology. Nature Reviews Bioengineering, 1(12): 930-949, 2023. 1", + "[25] Joey Spronck, Thijs Gelton, Leander van Eekelen, Joep Bogaerts, Leslie Tessier, Mart van Rijthoven, Lieke van der Woude, Michel van den Heuvel, Willemijn Theelen, Jeroen van der Laak, et al. nnunet meets pathology: bridging the gap for application to whole-slide images and computational biomarkers. In Medical Imaging with Deep Learning, 2023. 2", + "[26] Yuxuan Sun, Chenglu Zhu, Sunyi Zheng, Kai Zhang, Lin Sun, Zhongyi Shui, Yunlong Zhang, Honglin Li, and Lin Yang. Pathasst: A generative foundation ai assistant towards artificial general intelligence of pathology. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 5034-5042, 2024. 1, 2", + "[27] Eugene Vorontsov, Alican Bozkurt, Adam Casson, George Shaikovski, Michal Zelechowski, Kristen Severson, Eric Zimmermann, James Hall, Neil Tenenholtz, Nicolo Fusi, et al. A foundation model for clinical-grade computational pathology and rare cancers detection. Nature Medicine, pages 1-12, 2024. 1", + "[28] Xiyue Wang, Junhan Zhao, Eliana Marostica, Wei Yuan, Ji etian Jin, Jiayu Zhang, Ruijiang Li, Hongping Tang, Kanran Wang, Yu Li, et al. A pathology foundation model for cancer diagnosis and prognosis prediction. Nature, pages 1-9, 2024. 2", + "[29] Feng Xu, Chuang Zhu, Wenqi Tang, Ying Wang, Yu Zhang, Jie Li, Hongchuan Jiang, Zhongyue Shi, Jun Liu, and Mulan Jin. Predicting axillary lymph node metastasis in early breast cancer using deep learning on primary tumor biopsy slides. Frontiers in oncology, 11:759007, 2021. 5", + "[30] Hanwen Xu, Naoto Usuyama, Jaspreet Bagga, Sheng Zhang, Rajesh Rao, Tristan Naumann, Cliff Wong, Zelalem Gero, Javier González, Yu Gu, et al. A whole-slide foundation model for digital pathology from real-world data. Nature, pages 1-8, 2024. 1, 3", + "[31] Yingxue Xu, Yihui Wang, Fengtao Zhou, Jiabo Ma, Shu Yang, Huangjing Lin, Xin Wang, Jiguang Wang, Li Liang, Anjia Han, et al. A multimodal knowledge-enhanced whole-slide pathology foundation model. arXiv preprint arXiv:2407.15362, 2024. 2", + "[32] An Yang, Baosong Yang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Zhou, Chengpeng Li, Chengyuan Li, Dayiheng Liu, Fei Huang, et al. Qwen2 technical report. arXiv preprint arXiv:2407.10671, 2024. 5, 7", + "[33] Fengtao Zhou and Hao Chen. Cross-modal translation and alignment for survival analysis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 21485-21494, 2023. 2" + ], + "bbox": [ + 104, + 103, + 483, + 773 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "5143", + "bbox": [ + 482, + 940, + 514, + 950 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/SlideChat_ A Large Vision-Language Assistant for Whole-Slide Pathology Image Understanding/5542bef9-2edf-4ac5-b8df-cf3f63aeaa6c_model.json b/2025/SlideChat_ A Large Vision-Language Assistant for Whole-Slide Pathology Image Understanding/5542bef9-2edf-4ac5-b8df-cf3f63aeaa6c_model.json new file mode 100644 index 0000000000000000000000000000000000000000..3cdbb565e0a88547b56514283f49d0832c2ed1ca --- /dev/null +++ b/2025/SlideChat_ A Large Vision-Language Assistant for Whole-Slide Pathology Image Understanding/5542bef9-2edf-4ac5-b8df-cf3f63aeaa6c_model.json @@ -0,0 +1,1815 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.104, + 0.14, + 0.895, + 0.186 + ], + "angle": 0, + "content": "SlideChat: A Large Vision-Language Assistant for Whole-Slide Pathology Image Understanding" + }, + { + "type": "text", + "bbox": [ + 0.145, + 0.21, + 0.851, + 0.248 + ], + "angle": 0, + "content": "Ying Chen\\(^{1,2*}\\) Guoan Wang\\(^{1,3*}\\) Yuanfeng Ji\\(^{4*†}\\) Yanjun Li\\(^{1,3}\\) Jin Ye\\(^{1,5}\\) Tianbin Li\\(^{1}\\) Ming Hu\\(^{1,5}\\) Rongshan Yu\\(^{2}\\) Yu Qiao\\(^{1}\\) Junjun He\\(^{1\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.181, + 0.262, + 0.816, + 0.299 + ], + "angle": 0, + "content": "\\(^{1}\\)Shanghai AI Laboratory \\(^{2}\\)Xiamen University \\(^{3}\\)East China Normal University \\(^{4}\\)Stanford University \\(^{5}\\)Monash University" + }, + { + "type": "title", + "bbox": [ + 0.256, + 0.332, + 0.331, + 0.347 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.362, + 0.485, + 0.732 + ], + "angle": 0, + "content": "Despite the progress made by multimodal large language models (MLLMs) in computational pathology, they remain limited by a predominant focus on patch-level analysis, missing essential contextual information at the whole-slide level. The lack of large-scale instruction datasets and the gigapixel scale of whole slide images (WSIs) pose significant developmental challenges. In this paper, we present SlideChat, the first vision-language assistant capable of understanding gigapixel whole-slide images, exhibiting excellent multimodal conversational capability and response complex instruction across diverse pathology scenarios. To support its development, we created SlideInstruction, the largest instruction-following dataset for WSIs consisting of 4.2K WSI captions and 176K VQA pairs with multiple categories. Furthermore, we propose SlideBench, a multimodal benchmark that incorporates captioning and VQA tasks to assess SlideChat's capabilities in various settings such as microscopy, diagnosis and clinical. Compared to both general and specialized MLLMs, SlideChat exhibits exceptional capabilities, achieving state-of-the-art performance on 18 of 22 tasks. For example, it achieved an overall accuracy of \\(81.17\\%\\) on SlideBench-VQA (TCGA), and \\(54.15\\%\\) on SlideBench-VQA (BCNB). Our code, data, and model is publicly accessible at https://uni-medical.github.io/SlideChat.github.io." + }, + { + "type": "title", + "bbox": [ + 0.102, + 0.758, + 0.231, + 0.773 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.782, + 0.484, + 0.857 + ], + "angle": 0, + "content": "Computational pathology aims to improve the analysis of digitized tissue samples [14, 24], such as whole slide images (WSIs), by applying artificial intelligence to aid in the diagnosis, identification, and understanding of disease. Recently, the development of this field has gained rapid" + }, + { + "type": "image_caption", + "bbox": [ + 0.517, + 0.335, + 0.531, + 0.345 + ], + "angle": 0, + "content": "A" + }, + { + "type": "image", + "bbox": [ + 0.536, + 0.336, + 0.621, + 0.379 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.537, + 0.387, + 0.674, + 0.404 + ], + "angle": 0, + "content": "Summarize the key pathological findings from the whole slide image." + }, + { + "type": "text", + "bbox": [ + 0.546, + 0.411, + 0.705, + 0.428 + ], + "angle": 0, + "content": "This pathological examination reveals the high-grade urothelial carcinoma with ..." + }, + { + "type": "text", + "bbox": [ + 0.537, + 0.437, + 0.681, + 0.461 + ], + "angle": 0, + "content": "What is the pathological stage (pT) of the transitional cell carcinoma? A. pT1 B. pT2 C. pT3 D. pT4" + }, + { + "type": "text", + "bbox": [ + 0.658, + 0.465, + 0.686, + 0.474 + ], + "angle": 0, + "content": "B. pT2" + }, + { + "type": "image_caption", + "bbox": [ + 0.715, + 0.336, + 0.727, + 0.345 + ], + "angle": 0, + "content": "B" + }, + { + "type": "text", + "bbox": [ + 0.738, + 0.343, + 0.783, + 0.352 + ], + "angle": 0, + "content": "Microscopy" + }, + { + "type": "text", + "bbox": [ + 0.73, + 0.354, + 0.884, + 0.361 + ], + "angle": 0, + "content": "\\(\\spadesuit\\) Tissue Architecture \\(\\spadesuit\\) Tumor Characteristics" + }, + { + "type": "text", + "bbox": [ + 0.73, + 0.362, + 0.822, + 0.37 + ], + "angle": 0, + "content": "Histopathological Changes" + }, + { + "type": "text", + "bbox": [ + 0.73, + 0.372, + 0.845, + 0.38 + ], + "angle": 0, + "content": "\\(\\spadesuit\\) Cytomorphological Characteristics" + }, + { + "type": "list", + "bbox": [ + 0.73, + 0.354, + 0.884, + 0.38 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.738, + 0.392, + 0.778, + 0.401 + ], + "angle": 0, + "content": "Diagnosis" + }, + { + "type": "text", + "bbox": [ + 0.73, + 0.403, + 0.884, + 0.411 + ], + "angle": 0, + "content": "\\(\\spadesuit\\) Disease Detection Staging Grading" + }, + { + "type": "text", + "bbox": [ + 0.73, + 0.413, + 0.81, + 0.419 + ], + "angle": 0, + "content": "\\(\\spadesuit\\) Disease Classification" + }, + { + "type": "text", + "bbox": [ + 0.73, + 0.42, + 0.807, + 0.428 + ], + "angle": 0, + "content": "\\(\\spadesuit\\) Differential Diagnosis" + }, + { + "type": "list", + "bbox": [ + 0.73, + 0.403, + 0.884, + 0.428 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.738, + 0.44, + 0.766, + 0.447 + ], + "angle": 0, + "content": "Clinical" + }, + { + "type": "text", + "bbox": [ + 0.73, + 0.449, + 0.853, + 0.455 + ], + "angle": 0, + "content": "Treatment Guidance Risk Factors" + }, + { + "type": "text", + "bbox": [ + 0.73, + 0.457, + 0.811, + 0.464 + ], + "angle": 0, + "content": "Prognostic Assessment" + }, + { + "type": "text", + "bbox": [ + 0.73, + 0.466, + 0.798, + 0.474 + ], + "angle": 0, + "content": "\\(\\spadesuit\\) Biomarker Analysis" + }, + { + "type": "list", + "bbox": [ + 0.73, + 0.449, + 0.853, + 0.474 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.49, + 0.895, + 0.557 + ], + "angle": 0, + "content": "Figure 1. SlideChat is the first large vision-language assistant specifically designed for whole-slide pathology analysis. SlideChat can generate comprehensive descriptions of whole-slide images and provide contextually relevant responses across various applications." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.592, + 0.896, + 0.768 + ], + "angle": 0, + "content": "momentum, mainly driven by breakthroughs in the visual foundation model [9, 27, 30]. These models learn generalized representations by pre-training on large-scale data and perform well in various downstream tasks, such as cancer detection and biomarker prediction. Building on this base, integration with the powerful Large Language Models (LLMs) further advances the development of the Multimodal Large Language Model (MLLMs) [20], which has made great strides in responding to more complex, open-ended visual queries, enabling it to serve as a versatile assistant at various stages of medical care, including clinical decision-making, education, and research." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.772, + 0.897, + 0.891 + ], + "angle": 0, + "content": "Nevertheless, there are some major challenges that hinder the development and use of pathology MLLMs for real-world applications. First, it is challenging to develop a MLLM architecture that can effectively be capable of gigapixel whole slides (e.g., \\(100,000 \\times 100,000\\) pixels). Existing models [20, 22, 26] often process whole slides by extracting small patch/ROI-level data for subsequent analysis, resulting in a limited understanding of global slide con" + }, + { + "type": "page_footnote", + "bbox": [ + 0.125, + 0.865, + 0.375, + 0.877 + ], + "angle": 0, + "content": "*These authors contributed equally to this work." + }, + { + "type": "page_footnote", + "bbox": [ + 0.125, + 0.878, + 0.473, + 0.89 + ], + "angle": 0, + "content": "†Corresponding authors: yfj@stanford.edu, hejunjun@pjlab.org.cn." + }, + { + "type": "list", + "bbox": [ + 0.125, + 0.865, + 0.473, + 0.89 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.94, + 0.516, + 0.953 + ], + "angle": 0, + "content": "5134" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.101, + 0.103, + 0.482, + 0.176 + ], + "angle": 0, + "content": "text and suboptimal performance in some complex pathological analysis. Second, publicly available multi-modal pathology slide datasets are relatively scarce and of varying quality [7, 8, 12], which limits the development of MLLMs trained on such data." + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.181, + 0.483, + 0.534 + ], + "angle": 0, + "content": "In this paper, we present SlideChat (see Fig. 1), the first open-source vision-language assistant capable of understanding gigapixel whole-slide images. First, SlideChat is trained on SlideInstruction, a large-scale multi-modal instruction dataset encompassing data from The Cancer Genome Atlas (TCGA) [15] via our specifically designed data processing pipeline (see Fig. 3 (A)). SlideInstruction contains 4,181 WSI-caption pairs and 175,753 visual question-answer pairs from 3,294 patients, covering 10 cancer types. The question-answer pairs include both open-ended and closed-ended questions, further divided into 13 subcategories, covering a diverse range of clinical tasks. SlideInstruction is more than 20 times larger than previous public instruction datasets in the number of instructions. Second, we propose SlideChat, a novel architecture in LLaVA style for capable multi-modal analysis of gigapixel whole slides. A gigapixel whole slide is first divided into a series of patches, each of which is individually processed by a patch-level encoder to extract local features. The resulting long sequence of patch tokens is then processed by a slide-level encoder employing sparse attention to enhance the slide-level features. Finally, enhanced patch tokens are fed into the LLM via a projector, which processes user queries and generates responses." + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.538, + 0.483, + 0.891 + ], + "angle": 0, + "content": "To systematically evaluate the performance of SlideChat in real-world scenarios, we establish a comprehensive digital pathology benchmark named SlideBench, encompassing more than 20 clinical tasks, using data from both TCGA and the in-the-wild Early Breast Cancer CoreNeedle Biopsy (BCNB) dataset. This resulted in three test sets: SlideBench-Caption, comprising 734 WSI-caption pairs; SlideBench-VQA (TCGA), comprising 7,827 VQA pairs covering 13 different tasks; and SlideBench-VQA (BCNB), including a total of 7,274 VQA entries from 1,058 patients, covering seven different tasks. Additionally, we compare the performance of SlideChat on another externally proposed dataset, WSI-VQA [8], to further validate its effectiveness. We compare our model with the currently available state-of-the-art general and medical-specialized MLLMs including GPT-4o, LLaV-aMed [16], MedDr [13] and Quilt-LLaVA [22]. Benefiting from large-scale, high-quality training and effective local-global context modeling, SlideChat achieves state-of-the-art performance on 18 out of 22 tasks, with significant improvements over the second-best method \\(10\\%\\) on 9 tasks on four benchmarks. Specifically, SlideChat achieves an average accuracy improvement of \\(13.47\\%\\) over the second-best model on SlideBench-VQA (TCGA), an average improvement of \\(12.59\\%\\) on" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.103, + 0.894, + 0.19 + ], + "angle": 0, + "content": "SlideBench-VQA (BCNB), and an improvement of \\(5.82\\%\\) on WSI-VQA. Finally, to accelerate research progress in digital pathology, we make SlideChat fully open-weight, including source code and model weights as well as instruction and benchmark datasets. The key contributions are summarized four-fold in the following:" + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.191, + 0.894, + 0.248 + ], + "angle": 0, + "content": "- We develop SlideChat, the first vision-language assistant capable of understanding gigapixel whole-slide images, achieving state-of-the-art performance on multiple benchmarks." + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.25, + 0.894, + 0.293 + ], + "angle": 0, + "content": "- We create SlideInstruction, a largest comprehensive WSI instruction-following dataset containing 4.2K WSI-caption pairs and 176K VQA pairs." + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.294, + 0.894, + 0.351 + ], + "angle": 0, + "content": "- We establish SlideBench, a WSI multi-modal benchmark comprising SlideBench-Caption, SlideBench-VQA (TCGA), and SlideBench-VQA (BCNB), covering 21 different clinical tasks." + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.353, + 0.894, + 0.396 + ], + "angle": 0, + "content": "- We release SlideChat, SlideInstruction and SlideBench as open-source resources to facilitate research and development in computational pathology." + }, + { + "type": "list", + "bbox": [ + 0.532, + 0.191, + 0.894, + 0.396 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.408, + 0.658, + 0.423 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.433, + 0.895, + 0.609 + ], + "angle": 0, + "content": "Whole Slide Image Analysis Whole slide images are pivotal in modern pathology, enabling comprehensive analysis of tissue samples for tasks such as predicting patient prognosis, classifying cancer subtypes, and identifying biomarkers [17, 18, 23-25, 33]. Recent studies have leveraged pathology foundational models [3, 28, 31] to enhance WSIs analysis, either through fine-tuning for specific downstream tasks or by employing zero-shot prediction approaches in CLIP [21] style. Although these models are effective in task-specific applications, their reliance on fine-tuning or limited zero-shot capabilities restricts their generalizability across diverse and complex user instructions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.626, + 0.895, + 0.891 + ], + "angle": 0, + "content": "MLLMs in Computational Pathology The paradigm of MLLMs enables them to effectively respond to more complex, open-ended visual queries while processing pathology images, thus providing significant value across various medical stages. PathChat [20] is a vision-language assistant designed for pathology, developed with 450K private instruction pairs to handle both visual and natural language queries. Quilt-Instruct [22] is a large-scale dataset comprising 107K question-answer pairs. Building on Quilt-Instruct, Quilt-LLAVA [22] is a model designed for diagnostic reasoning across multiple image patches, leveraging its extensive question-answer pairs to accurately interpret complex H&E images. PathAsst [26] combines a pathology-specific CLIP model with Vicuna-13b [10] to create a multimodal generative foundational model tailored for pathology. However, current MLLMs primarily only focus on patch or ROI data, limiting their utility for slide-level clinical applications where broader contextual understanding is crucial." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.94, + 0.516, + 0.953 + ], + "angle": 0, + "content": "5135" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.108, + 0.104, + 0.885, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.1, + 0.315, + 0.897, + 0.371 + ], + "angle": 0, + "content": "Figure 2. Overview of our SlideChat. (A) SlideChat serializes each input WSI into a sequence of \\(224 \\times 224\\) patches, converting each into visual embeddings with a patch-level encoder. A slide-level encoder then interacts with these features to generate contextual embeddings. Then, a multimodal projector maps the visual features from the slide-level encoder into a unified space, aligned seamlessly with the LLM. (B) SlideChat was trained for two stages: Cross-Domain Alignment and Visual Instruction Learning." + }, + { + "type": "title", + "bbox": [ + 0.102, + 0.394, + 0.208, + 0.409 + ], + "angle": 0, + "content": "3. SlideChat" + }, + { + "type": "title", + "bbox": [ + 0.102, + 0.423, + 0.233, + 0.437 + ], + "angle": 0, + "content": "3.1. Architecture" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.449, + 0.486, + 0.891 + ], + "angle": 0, + "content": "To achieve the goal of analyzing gigapixel whole-slide images in a multimodal setting, as shown in Fig. 2, SlideChat consists of four key designs: the patch-level encoder, the slide-level encoder, the multimodal projector module, and the large language model. Our method starts by partitioning the WSI into smaller \\(224 \\times 224\\) pixel patches, making it computationally feasible to process such large images. These patches are then passed through a well-trained, frozen patch-level encoder [19], which extracts localized features from each individual patch, capturing fine-grained details such as cellular structures. Building on this, we employ LongNet [11, 30] as slide-level encoder to enhance the patch-level embeddings and capture global patterns across the entire slide. This encoder uses sparse attention mechanisms to learn both local and global contextual information, enabling the model to perceive intricate local features while capturing the broader context, which is critical for comprehensive pathological assessments. Following the slide-level encoding, SlideChat incorporates a multimodal projector that maps these patch tokens into a unified space aligned with the LLM. This ensures that the visual features extracted from the WSIs are effectively transformed into representations compatible with the language model, facilitating seamless integration and interaction between visual and textual data. Concurrently, the model accepts natural language instructions from users, such as \"What is the type of tumor in the image?\". These textual queries are processed by the LLM, which comprehends the textual input and integrates it with the visual features extracted from the WSIs, enabling accurate and contextually relevant diagnoses" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.395, + 0.896, + 0.47 + ], + "angle": 0, + "content": "tic responses. This multimodal reasoning capability allows SlideChat to provide accurate and contextually relevant answers to complex pathology-related questions, thereby supporting clinical decision-making, education, and research across various medical stages." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.484, + 0.586, + 0.498 + ], + "angle": 0, + "content": "3.2.Data" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.508, + 0.896, + 0.892 + ], + "angle": 0, + "content": "SlideInstruction There is a notable lack of large-scale multimodal pathology datasets supporting the training of vision-language assistants for whole-slide image understanding. To support the training of SlideChat, we develop SlideInstruction, a comprehensive instruction dataset, sourced from the TCGA database, comprising 4,915 WSI-report pairs from 4,028 patients. Fig. 3 illustrates our entire data curation pipeline. We initially prompt GPT-4 to refine the pathology reports, and clean up the noise in the report including unrelated symbols, technical details of pathology department procedures, specimen handling and processing information, redundant administrative or legal statements, and some repeated information. For the refined pathology reports, we further employ GPT-4 to generate high-quality multimodal data, comprising two main components: (1) WSI-Caption Data: We craft concise, clinically relevant summaries for each whole slide image by prompting the language model to extract key pathological findings. These summaries were structured into coherent paragraphs that highlighted crucial clinical details such as diagnostic results, tumor characteristics, margin status, and lymph node involvement, ensuring the caption dataset is both focused and informative. (2) WSI Instruction-Following Data: To enhance the model's ability to follow instructions and improve its comprehension of pathology images, we leveraged GPT-4 to generate tailored question" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.94, + 0.516, + 0.953 + ], + "angle": 0, + "content": "5136" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.128, + 0.106, + 0.144, + 0.117 + ], + "angle": 0, + "content": "A" + }, + { + "type": "image", + "bbox": [ + 0.158, + 0.115, + 0.867, + 0.252 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.129, + 0.259, + 0.143, + 0.271 + ], + "angle": 0, + "content": "B" + }, + { + "type": "image", + "bbox": [ + 0.153, + 0.272, + 0.825, + 0.337 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.129, + 0.346, + 0.143, + 0.358 + ], + "angle": 0, + "content": "C" + }, + { + "type": "image", + "bbox": [ + 0.163, + 0.368, + 0.278, + 0.425 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.165, + 0.438, + 0.277, + 0.494 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.299, + 0.376, + 0.483, + 0.42 + ], + "angle": 0, + "content": "[Caption] #Q:Describe the whole slide image. #A:The bladder exhibits an invasive poorly differentiated urothelial carcinoma, grade 4/4, extending through the bladder neck wall...." + }, + { + "type": "text", + "bbox": [ + 0.299, + 0.444, + 0.482, + 0.489 + ], + "angle": 0, + "content": "[VQA-Diagnosis] \n#Q: From the given whole slide image, which histological subtype of lung adenocarcinoma is most predominantly observed? \n#A: It is acinar subtype observed." + }, + { + "type": "image", + "bbox": [ + 0.532, + 0.37, + 0.641, + 0.426 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.528, + 0.434, + 0.637, + 0.497 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.663, + 0.377, + 0.848, + 0.421 + ], + "angle": 0, + "content": "[VQA-Microscopy] \n#Q: What histological feature observable in a whole slide image differentiates anaplastic astrocytoma from glioblastoma? \n#A: Absence of endothelial proliferation." + }, + { + "type": "text", + "bbox": [ + 0.663, + 0.445, + 0.848, + 0.488 + ], + "angle": 0, + "content": "[VQA-Clinical] #Q: Based on the histopathological features, what is the primary recommended treatment? #A: Simple mastectomy followed by hormone therapy." + }, + { + "type": "image_caption", + "bbox": [ + 0.101, + 0.518, + 0.895, + 0.573 + ], + "angle": 0, + "content": "Figure 3. (A) Overview of the SlideInstruction generation pipeline. We prompt GPT-4 to extract the WSI-Caption, Open-set VQA and Closed-set VQA from reports. (B) For the generated Closed-set VQA, we employ LLMs to filter low-quality QA pairs and involve pathologists for validation, resulting in the creation of SlideBench-VQA. (C) Examples of WSI caption and instruction-following scenarios in microscopy, diagnostics, and clinical applications. For additional examples, please refer to Supplementary Material." + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.599, + 0.484, + 0.878 + ], + "angle": 0, + "content": "and-answer pairs for each WSI report. Drawing inspiration from PathChat [20], we structure these questions into 3 broad categories—microscopy, diagnosis, and clinical considerations—which represent key stages in the pathology workflow, and 13 narrow categories focusing on specific aspects within each stage (Fig. 1 B). To construct a comprehensive instructional dataset, we began by generating two open-ended and two closed-ended QA pairs for each report within each specific category, taking advantage of the rich content found in the majority of TCGA reports. Then, some QA pairs were excluded, such as those lacking corresponding category information. Regarding the train/test split, it is worth noting that the WSI-report datasets from TCGA include 2 types: (a) one report linked to multiple WSIs, and (b) one report linked to a single WSI. For type (a), where specific diagnostic details may not align perfectly with each WSI, we include all WSIs in the training set to introduce some \"noisy data\", which can enhance model robustness. For type (b), \\(80\\%\\) of WSIs are allocated to the training" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.599, + 0.895, + 0.673 + ], + "angle": 0, + "content": "set and \\(20\\%\\) to the test set. Finally, there are 4,181 WSIs for training and 734 WSIs for testing. Consequently, we construct a large-scale training set named SlideInstruction, comprising 4,181 WSI captions and 175,753 instruction-following VQA pairs across 13 narrow categories." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.699, + 0.896, + 0.891 + ], + "angle": 0, + "content": "SlideBench To systematically evaluate the performance of SlideChat, we include the remaining 734 WSI captions from the test set, along with a substantial number of closed-set VQA pairs, to establish a comprehensive evaluation benchmark. First, we construct a set named SlideBench-Caption based on the WSI-Caption data to evaluate the model's ability to generate accurate and coherent descriptions of whole slide images. Secondly, we construct SlideBench-VQA (TCGA) based on closed-set VQA pairs along with test WSIs, aiming to evaluate various aspects of model performance. As shown in Fig. 3 (B), to improve the quality of the testing benchmarks, we employ four advanced large language models, including GPT-4 [2], InternLM2-" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.94, + 0.516, + 0.953 + ], + "angle": 0, + "content": "5137" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.101, + 0.103, + 0.486, + 0.53 + ], + "angle": 0, + "content": "Chat-7B [6], Qwen-7B-Chat [5], and DeepSeek-7B-Chat, to filter closed-set VQAs by predicting answers based solely on the question text. Any questions for which at least three of these models provided correct answers are subsequently excluded. Following this automated filtering, five expert pathologists are invited to review and amend the remaining questions. The review process is guided by the following criteria: (1) Whether the correct answer necessitates image interpretation; (2) Whether the question and its corresponding answer are logically and coherently structured; and (3) Whether the question aligns appropriately with the designated broad and narrow categories. VQA pairs failing to meet these criteria are excluded by the pathologists. Consequently, the SlideBench-VQA (TCGA) comprises 7,827 VQAs across 13 categories, with some examples illustrated in Fig. 3 C. Additionally, we incorporate the in-the-wild Early Breast Cancer Core-Needle Biopsy (BCNB) WSI dataset [29], which encompasses a diverse patient population and a variety of clinical task labels, to enhance the test set benchmark and more comprehensively assess the model's generalization capabilities. In detail, we convert the BCNB dataset into a VQA format by rephrasing the classification objectives into a specific template as questions, while transforming the original multi-class labels into selectable options, and integrating it into SlideBench as an external subset, named SlideBench-VQA (BCNB). This dataset comprises 7,247 VQA pairs from 1,058 patients, specifically designed to evaluate SlideChat's zero-shot generalization capability across 7 distinct classification tasks." + }, + { + "type": "title", + "bbox": [ + 0.102, + 0.539, + 0.283, + 0.556 + ], + "angle": 0, + "content": "3.3. Two-stage Training" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.561, + 0.485, + 0.737 + ], + "angle": 0, + "content": "Stage 1: Cross-Domain Alignment. SlideChat adopts a two-stage training approach (see Fig. 2 B). In the first stage, the primary objective is to align the LLM's word embeddings with the visual features extracted from whole slide images. This alignment enables the LLM to interpret visual representations from the slide-level encoder, facilitating the effective utilization of the intricate features within the slides. During this stage, SlideChat is trained to generate descriptive captions using 4.2K WSI-caption pairs from SlideInstruction. Specifically, only the slide-level encoder and projection matrix are updated, while the patch-level encoder and LLM weights remain fixed." + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.757, + 0.483, + 0.891 + ], + "angle": 0, + "content": "Stage 2: Visual Instruction Learning. In the second stage, we focus on visual question-answering tasks to train the model to accurately respond to domain-specific questions concerning whole slide images. During this phase, the model develops the ability to handle a broad range of multimodal instructions, enabling it to generate answers by effectively integrating both visual and textual information. For example, the model must perform various pathology tasks, such as describing the extent of tumor invasion or assess" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.103, + 0.896, + 0.221 + ], + "angle": 0, + "content": "ing the degree of cellular differentiation. To accomplish this, we utilize 176K WSI VQAs from SlideInstruction in the second training stage, allowing the slide encoder, projection layer, and large language model components to be fully trainable to ensure comprehensive adaptability. This approach significantly enhances the model's capability to handle diverse pathology-related tasks, thereby increasing its effectiveness in real-world clinical and research settings." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.234, + 0.634, + 0.25 + ], + "angle": 0, + "content": "4. Experiment" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.259, + 0.896, + 0.669 + ], + "angle": 0, + "content": "We conducted the following experiments to evaluate three key aspects of SlideChat: (1) its whole slide image captioning capability, which assesses proficiency in generating descriptive captions that accurately summarize the critical pathological features of a WSI; (2) its visual question-answering (VQA) ability across various complex pathological scenarios and its generalizability in zero-shot settings; and (3) SlideChat's ability to process gigapixel WSIs, capturing both essential global context and intricate details, thereby enhancing its performance compared to patch-level MLLMs. For WSI captioning baselines, we benchmark against MI-Gen [7], a state-of-the-art method specifically designed for this task. Given that existing MLLMs cannot handle the gigapixel scale of whole slide images, we establish baseline comparisons using two approaches: (1) randomly selecting 30 patches from each WSI and inputting them into MLLMs (e.g., GPT-4o [2], LLaVA-Med [16], MedDr [16], Quilt-LLaVA [22]), followed by a majority voting scheme to generate slide-level predictions; and (2) directly inputting a WSI thumbnail, resized to \\(1024 \\times 1024\\) pixels, referred to as Slide (T), into the MLLMs. For VQA tasks, we further evaluate performance by comparing against random prediction baselines and text-only models, thereby assessing the incremental contribution of visual information. Unless otherwise specified, SlideChat is configured with the patch-level encoder CONCH [19], the slide-level encoder LongNet [11], and utilizes the Qwen2.5-7B-Instruct [32] as LLM." + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.681, + 0.895, + 0.709 + ], + "angle": 0, + "content": "Table 1. Captioning performance across different methods on SlideBench-Caption." + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.72, + 0.894, + 0.813 + ], + "angle": 0, + "content": "
MethodsInputBLEU-1BLEU-2BLEU-3BLEU-4Rouge-LGPT score
GPT-4oPatch0.160.030.010.010.131.54
GPT-4oSlide (T)0.100.030.010.010.111.03
Quilt-LLaVAPatch0.300.190.110.050.183.87
Quilt-LLaVASlide (T)0.230.090.040.010.161.89
MI-GenSlide0.370.240.150.100.254.14
SlideChatSlide0.370.210.120.080.244.11
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.845, + 0.895, + 0.89 + ], + "angle": 0, + "content": "SlideBench-Caption We report BLEU, ROUGE, and GPT scores to evaluate caption generation performance in Tab. 1. For the GPT score, we use GPT-4 to assess the" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.94, + 0.516, + 0.953 + ], + "angle": 0, + "content": "5138" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.35, + 0.101, + 0.645, + 0.114 + ], + "angle": 0, + "content": "Table 2. VQA performance with different methods." + }, + { + "type": "table", + "bbox": [ + 0.132, + 0.126, + 0.868, + 0.349 + ], + "angle": 0, + "content": "
MethodsInputSlideBench-VQA (TCGA)SlideBench-VQA (BCNB)WSI-VQA*
MicroscopyDiagnosisClinicalOverall
RandomText24.4424.9126.4425.0224.4024.14
GPT-438.2829.0945.0037.25018.60
GPT-4oPatch62.8946.6966.7757.9141.4330.41
MedDr73.3057.7874.2567.7033.6754.36
LLaVA-Med47.3432.7847.9642.0030.126.31
Quilt-LLaVA57.7635.9653.0748.0732.1944.43
GPT-4oSlide (T)38.2823.1043.4234.07014.03
MedDr70.4852.4772.8064.2535.4850.95
LLaVA-Med45.8227.5840.8437.39018.79
Quilt-LLaVA49.1226.9744.7539.3941.5535.40
SlideChatSlide87.64(+14.34)73.27(+15.49)84.26(+10.01)81.17(+13.47)54.14(+12.59)60.18(+5.82)
" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.374, + 0.483, + 0.726 + ], + "angle": 0, + "content": "similarity between the generated captions and the ground truth, providing an overall score on a scale of 1 to 10, with higher scores indicating better performance. When utilizing patch-level inputs, GPT-4o generates individual descriptions for each patch, which are subsequently integrated to create the final slide-level caption. However, this approach yields poor performance, as evidenced by a BLEU-1 score of 0.16 and a GPT-score of 1.54. These results suggest that the patch-based method fails to adequately capture the broader context necessary for accurate WSI captioning. When the WSI thumbnail of size \\(1024 \\times 1024\\) pixels is used as input to GPT-4o, performance decreases further, with a BLEU-1 score of 0.10 and a GPT-score of 1.03. This suggests that while the thumbnail offers a global view of the slide, it may lack the resolution and detail necessary for generating precise and informative captions. In contrast, MI-Gen, a model specifically designed for WSI captioning, demonstrates significantly superior performance across all metrics, achieving a BLEU-1 score of 0.37 and a GPT score of 4.14. Similarly, SlideChat, shows competitive results with a BLEU-1 score of 0.37 and a GPT score of 4.11. These outcomes highlight SlideChat's ability to effectively integrate both local and global information from the slides and confirm its efficiency in describing whole-slide images." + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.743, + 0.483, + 0.891 + ], + "angle": 0, + "content": "SlideBench-VQA (TCGA) We further evaluate SlideChat's overall performance on the multiple-choice VQA benchmark. The results on SlideBench-VQA (TCGA) in Tab. 2, compare different methods across three categories: microscopy, diagnosis, and clinical, along with an overall performance score. Random selection achieves an overall score of \\(25.02\\%\\) accuracy, serving as a baseline for answer distribution but demonstrating poor performance. While GPT-4, relying solely on text input, outperforms random predictions, it continues to struggle" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.374, + 0.895, + 0.756 + ], + "angle": 0, + "content": "with accurately answering questions. When GPT-4o incorporates patch-level inputs, its performance improves markedly, reaching a score of \\(57.91\\%\\) and underscoring the crucial role of detailed visual data. However, using a WSI thumbnail results in a lower score of \\(34.07\\%\\), as the reduced detail restricts its ability to deliver precise answers. MedDr performs well, achieving an overall score of \\(67.70\\%\\) with patch-level inputs, though this drops slightly to \\(64.25\\%\\) when using the slide thumbnail due to the loss of fine visual details. SlideChat outperforms all other methods, attaining a leading overall accuracy of \\(81.17\\%\\), excelling across all categories and significantly surpassing the competition. Even in more fine-grained pathological scenarios, as depicted in the left portion of Fig. 4, SlideChat remains the top-performing model across 13 tasks, particularly in areas such as cytomorphological characteristics, histopathological changes, disease detection, disease classification, and staging and grading, which require the identification of complex pathological visual features. Compared with baselines taking some patches or slide thumbnail as inputs, SlideChat has the capability to analyze a significantly greater number of pathological features with enhanced detail, effectively capturing both localized features and overarching global patterns, allowing SlideChat to provide more accurate and nuanced insights into pathological variations." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.772, + 0.895, + 0.891 + ], + "angle": 0, + "content": "SlideBench-VQA (BCNB) In the zero-shot VQA setting, SlideChat significantly outperforms all other models, achieving the highest overall score of \\(54.14\\%\\) on SlideBench-VQA (BCNB). It particularly excels in identifying tumor types, far surpassing other baselines in this task. This performance highlights SlideChat's generalization capability across a wide range of tasks. When taking patches as inputs, GPT-4o outperforms both MedDr and LLaVA" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.94, + 0.516, + 0.952 + ], + "angle": 0, + "content": "5139" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.18, + 0.101, + 0.805, + 0.271 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.175, + 0.286, + 0.82, + 0.302 + ], + "angle": 0, + "content": "Figure 4. Accuracy on different tasks in SlideBench-VQA (TCGA) (left) and SlideBench-VQA (BCNB) (right)." + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.327, + 0.484, + 0.532 + ], + "angle": 0, + "content": "Med, achieving a score of \\(41.43\\%\\), though it still falls short of SlideChat by \\(12.59\\%\\). Notably, GPT-4o and LLaVA-Med performed very poorly, achieving a score of zero across all tasks when evaluated using slide thumbnails from this testing set. MedDr and Quilt-LLaVA both achieve moderate results, with MedDr scoring \\(33.67\\%\\) on patches and \\(35.48\\%\\) on thumbnails, while Quilt-LLaVA attains \\(32.19\\%\\) on patches and \\(41.55\\%\\) on thumbnails. This outcome highlights that, for complex WSIs, relying solely on relatively sufficient visual features is inadequate for effectively supporting a range of tasks. In the more fine-grained pathological tasks of the BCNB benchmark, SlideChat attains state-of-the-art performance on 5 out of 7 tasks (Fig. 4), further demonstrating the effectiveness of SlideChat." + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.549, + 0.484, + 0.771 + ], + "angle": 0, + "content": "WSI-VQA* We also curated a subset of closed-set VQA pairs from the public WSI-VQA [8] dataset, referred to as WSI-VQA*, which exclusively consists of samples that were not in the training set, to evaluate the model's performance. SlideChat demonstrates the highest performance with a score of 60.18. Although MedDr performs well with both patch inputs (54.36%) and slide thumbnail inputs (50.95%), it still falls short compared to SlideChat. GPT-4o struggles significantly, especially with slide thumbnails, scoring only 14.03%, which highlights the limitations of using lower-resolution inputs. SlideChat leverages both detailed local information and broader context to process WSIs, making it the most effective model for this benchmark. This further emphasizes its superior capability in handling whole-slide data for VQA tasks." + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.787, + 0.484, + 0.891 + ], + "angle": 0, + "content": "Ablation We performed ablation experiments from different perspectives as follows: a) Large Language Model Comparison: Firstly, we compare the performance of several large language models, each with a parameter scale of approximately 7 billion, including Vicuna-7B-v1.5 [10], Phi-3-Mini-4k-Instruct [1], Qwen1.5-7B-Chat [5], Llama3-8B-Instruct [4], InternLM2-Chat-7B [6], and Qwen2.5-7B" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.325, + 0.895, + 0.352 + ], + "angle": 0, + "content": "Table 3. Performance Comparison of LLMs and Slide Encoder on WSI Captioning and VQA Tasks." + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.374, + 0.892, + 0.51 + ], + "angle": 0, + "content": "
LLMsSlide EncoderCaptionVQA (TCGA)VQA (BCNB)WSI-VQA*
Vicuna-7B-v1.5✓3.2841.4341.4331.98
Phi-3-Mini-4k-Instruct✓2.6679.9343.9260.18
Qwen1.5-7B-Chat✓2.9277.6344.0756.89
Llama3-8B-Instruct✓3.3078.7842.8255.25
Internlm2-Chat-7B✓3.3079.1052.1356.76
Qwen2.5-3B-Instruct✓3.4080.3245.7956.38
Qwen2.5-14B-Instruct✓3.3982.1451.5760.94
Qwen2.5-7B-Instruct✓4.1181.1754.1460.18
Qwen2.5-7B-Instruct×3.9581.2145.4959.67
" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.552, + 0.895, + 0.891 + ], + "angle": 0, + "content": "Instruct [32]. Specifically, we measured their performance on the SlideBench-Caption using the GPT-score and their accuracy on three VQA datasets. As shown in Tab. 3, the results demonstrate that SlideChat, powered by the Qwen2.5-7B-Instruct model, achieved the highest performance across all tasks, particularly excelling in the captioning task. These findings underscore the significant potential of developing SlideChat with the Qwen2.5-7B-Instruct model. Utilizing the Qwen2.5 model, we further evaluated models of varying scales and discovered that larger models generally exhibited superior performance, particularly the 7B and 14B parameter models. While these two models showed comparable performance across the three benchmarks, the 14B model surpassed the 7B model by \\(2.57\\%\\) on the SlideBench-VQA (TCGA). Given computational resource constraints, SlideChat uses the 7B model by default to achieve the best balance between performance and resource efficiency. b) Slide-level encoder effectiveness: We investigate the effectiveness of the slide-level encoder by initially removing it from SlideChat and employing a two-stage training approach. In the first stage, we exclusively trained the projection layers. However, this approach failed to reduce training loss or generate coherent text effectively," + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.94, + 0.516, + 0.952 + ], + "angle": 0, + "content": "5140" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.189, + 0.101, + 0.204, + 0.113 + ], + "angle": 0, + "content": "A" + }, + { + "type": "image", + "bbox": [ + 0.226, + 0.106, + 0.464, + 0.164 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.225, + 0.17, + 0.484, + 0.217 + ], + "angle": 0, + "content": "Select the most accurate description of the cytomorphological characteristics from image. A. Cells with increased nuclear-to-cytoplasmic ratio, hyperchromatic nuclei, and prominent nucleoli. B. Cells with uniform nuclei and abundant cytoplasm. C. Cells with minimal nuclear abnormalities. D. Cells with enlarged, clear nuclei and small nucleoli." + }, + { + "type": "image", + "bbox": [ + 0.189, + 0.218, + 0.204, + 0.251 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.233, + 0.402, + 0.249 + ], + "angle": 0, + "content": "A. Cells with increased nuclear-to-cytoplasmic ratio, hyperchromatic nuclei, and prominent nucleoli." + }, + { + "type": "image_caption", + "bbox": [ + 0.511, + 0.101, + 0.524, + 0.113 + ], + "angle": 0, + "content": "B" + }, + { + "type": "image", + "bbox": [ + 0.547, + 0.106, + 0.811, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.512, + 0.219, + 0.524, + 0.251 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.544, + 0.231, + 0.692, + 0.247 + ], + "angle": 0, + "content": "A. Dense collagen deposition with reduced cellularity" + }, + { + "type": "image_caption", + "bbox": [ + 0.101, + 0.265, + 0.895, + 0.293 + ], + "angle": 0, + "content": "Figure 5. Interpretability and visualization. We identify the top five patch tokens with the highest attention scores associated with the output text responses." + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.318, + 0.485, + 0.611 + ], + "angle": 0, + "content": "likely due to the difficulty of learning the complex visual features of WSIs without the slide-level encoder. When using a frozen LLM for complex text generation, a simple projection proved inadequate for effectively integrating and aligning visual and textual features. Subsequently, we consider combining data from both stages and training SlideChat without the slide-level encoder by simultaneously updating both the projection layers and the LLM. Under this paradigm, performance on SlideBench-VQA (TCGA) and WSI-VQA*, which share the same distribution as the training set, was comparable to the two-stage training configuration with the slide-level encoder. However, a significant decline is observed when evaluating SlideBench-VQA (BCNB), which originates from a different domain; overall performance dropped by over \\(10\\%\\) (Tab. 3), indicating a substantial reduction in the model's generalization ability. Therefore, we recommend incorporating a slide-level encoder to capture the complex visual features of whole slide, as it is particularly effective for cross-domain alignment and enhances the model's generalization performance." + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.64, + 0.484, + 0.891 + ], + "angle": 0, + "content": "Interpretability Despite SlideChat demonstrating promising results, concerns remained regarding the model's perception of large pathological slides. To further assess the model's interpretability, we calculated the correlation between the text output and specific image patches, thereby obtaining patch-level attention scores. By identifying the most significant patches, we gained insights into the precise areas the model focused on during response generation. Highlighting the most relevant regions of the tissue slides not only enhances transparency and bolsters the reliability of the model's outputs but also assists pathologists by directing attention to critical areas requiring closer scrutiny. Ultimately, such interpretability is essential for fostering trust in AI-assisted diagnostics and enhancing the precision and efficiency of clinical evaluations. As shown in Fig. 5, we are pleased to observe that the top five important patches identified by the model closely" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.318, + 0.896, + 0.598 + ], + "angle": 0, + "content": "corresponded with the features described in SlideChat's output. Our extraction method retrieves attention weights for patch tokens from each generated token, averaging them across layers and heads. We then identify the top five patch tokens with the highest attention weights for further analysis. For example, in Fig. 5 (A), the highlighted patches clearly emphasized regions exhibiting an increased nuclear-to-cytoplasmic ratio, hyperchromatic nuclei, and prominent nucleoli. Similarly, in Fig. 5 (B), the selected patches demonstrated areas with dense collagen deposition and reduced cellularity, as detailed in the model's response. This alignment between the highlighted image regions and the textual outputs significantly enhances the model's interpretability, providing increased confidence that it accurately captures and assesses relevant histopathological features. Such consistency deepens our understanding of the model's reasoning processes regarding pathological slides and underscores the potential for integrating these AI systems into clinical workflows with greater assurance." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.613, + 0.628, + 0.628 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.638, + 0.896, + 0.814 + ], + "angle": 0, + "content": "In this work, we present SlideChat, the first vision-language assistant capable of understanding gigapixel whole-slide images. Furthermore, we create SlideInstruction, a largest comprehensive WSI instruction-following dataset to develop SlideChat, as well as SlideBench, a multi-modal benchmark designed to evaluate SlideChat across diverse scenarios. SlideChat demonstrates excellent chat abilities and achieves state-of-the-art performance on 18 tasks. We bridge the gap between MLLMs and pathology images at the whole-slide level with SlideChat, and we believe this represents an advancement in the field of general medical artificial intelligence (GMAI)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.835, + 0.895, + 0.877 + ], + "angle": 0, + "content": "Acknowledgement The authors acknowledge the funding provided by the National Key Research and Development Program of China (No.2022ZD0160102)." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.94, + 0.514, + 0.953 + ], + "angle": 0, + "content": "5141" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.104, + 0.102, + 0.196, + 0.117 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.111, + 0.126, + 0.483, + 0.193 + ], + "angle": 0, + "content": "[1] Marah Abdin, Sam Ade Jacobs, Ammar Ahmad Awan, Jyoti Aneja, Ahmed Awadallah, Hany Awadalla, Nguyen Bach, Amit Bahree, Arash Bakhtiari, Harkirat Behl, et al. Phi-3 technical report: A highly capable language model locally on your phone. arXiv preprint arXiv:2404.14219, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.111, + 0.194, + 0.483, + 0.261 + ], + "angle": 0, + "content": "[2] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.111, + 0.262, + 0.483, + 0.342 + ], + "angle": 0, + "content": "[3] Faruk Ahmed, Andrew Sellergren, Lin Yang, Shawn Xu, Boris Babenko, Abbi Ward, Niels Olson, Arash Mohtashamian, Yossi Matias, Greg S Corrado, et al. Pathalign: A vision-language model for whole slide images in histopathology. arXiv preprint arXiv:2406.19578, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.111, + 0.344, + 0.373, + 0.357 + ], + "angle": 0, + "content": "[4] AI@Meta. Llama 3 model card. 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.111, + 0.359, + 0.483, + 0.412 + ], + "angle": 0, + "content": "[5] Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, et al. Qwen technical report. arXiv preprint arXiv:2309.16609, 2023. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.111, + 0.413, + 0.483, + 0.467 + ], + "angle": 0, + "content": "[6] Zheng Cai, Maosong Cao, Haojiong Chen, Kai Chen, Keyu Chen, Xin Chen, Xun Chen, Zehui Chen, Zhi Chen, Pei Chu, et al. Internl m2 technical report. arXiv preprint arXiv:2403.17297, 2024.5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.111, + 0.468, + 0.483, + 0.521 + ], + "angle": 0, + "content": "[7] Pingyi Chen, Honglin Li, Chenglu Zhu, Sunyi Zheng, and Lin Yang. Mi-gen: Multiple instance generation of pathology reports for gigapixel whole-slide images. arXiv preprint arXiv:2311.16480, 2023. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.111, + 0.522, + 0.483, + 0.576 + ], + "angle": 0, + "content": "[8] Pingyi Chen, Chenglu Zhu, Sunyi Zheng, Honglin Li, and Lin Yang. Wsi-vqa: Interpreting whole slide images by generative visual question answering. arXiv preprint arXiv:2407.05603, 2024. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.111, + 0.577, + 0.483, + 0.657 + ], + "angle": 0, + "content": "[9] Richard J Chen, Tong Ding, Ming Y Lu, Drew FK Williamson, Guillaume Jaume, Andrew H Song, Bowen Chen, Andrew Zhang, Daniel Shao, Muhammad Shaban, et al. Towards a general-purpose foundation model for computational pathology. Nature Medicine, 30(3):850-862, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.104, + 0.659, + 0.483, + 0.739 + ], + "angle": 0, + "content": "[10] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhanghao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E Gonzalez, et al. Vicuna: An open-source chatbot impressing gpt-4 with \\(90\\%\\) * chatgpt quality. See https://vicuna.lmsys.org (accessed 14 April 2023), 2(3):6, 2023. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.104, + 0.741, + 0.483, + 0.794 + ], + "angle": 0, + "content": "[11] Jiayu Ding, Shuming Ma, Li Dong, Xingxing Zhang, Shaohan Huang, Wenhui Wang, Nanning Zheng, and Furu Wei. Longnet: Scaling transformers to 1,000,000,000 tokens. arXiv preprint arXiv:2307.02486, 2023. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.104, + 0.795, + 0.483, + 0.861 + ], + "angle": 0, + "content": "[12] Zhengrui Guo, Jiabo Ma, Yingxue Xu, Yihui Wang, Liansheng Wang, and Hao Chen. Histgen: Histopathology report generation via local-global feature encoding and cross-modal context interaction. arXiv preprint arXiv:2403.05396, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.104, + 0.862, + 0.483, + 0.89 + ], + "angle": 0, + "content": "[13] Sunan He, Yuxiang Nie, Zhixuan Chen, Zhiyuan Cai, Hongmei Wang, Shu Yang, and Hao Chen. Meddr:" + }, + { + "type": "list", + "bbox": [ + 0.104, + 0.126, + 0.483, + 0.89 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.104, + 0.895, + 0.143 + ], + "angle": 0, + "content": "Diagnosis-guided bootstrapping for large-scale medical vision-language learning. arXiv preprint arXiv:2404.15127, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.146, + 0.895, + 0.226 + ], + "angle": 0, + "content": "[14] Mahdi S Hosseini, Babak Ehteshami Bejnordi, Vincent Quoc-Huy Trinh, Lyndon Chan, Danial Hasan, Xingwen Li, Stephen Yang, Taehyo Kim, Haochen Zhang, Theodore Wu, et al. Computational pathology: a survey review and the way forward. Journal of Pathology Informatics, page 100357, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.229, + 0.895, + 0.268 + ], + "angle": 0, + "content": "[15] Carolyn Hutter and Jean Claude Zenklusen. The cancer genome atlas: creating lasting value beyond its data. Cell, 173(2):283-285, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.271, + 0.895, + 0.338 + ], + "angle": 0, + "content": "[16] Chunyuan Li, Cliff Wong, Sheng Zhang, Naoto Usuyama, Haotian Liu, Jianwei Yang, Tristan Naumann, Hoifung Poon, and Jianfeng Gao. Llava-med: Training a large language-and-vision assistant for biomedicine in one day. Advances in Neural Information Processing Systems, 36, 2024. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.34, + 0.895, + 0.42 + ], + "angle": 0, + "content": "[17] Hao Li, Ying Chen, Yifei Chen, Rongshan Yu, Wenxian Yang, Liansheng Wang, Bowen Ding, and Yuchen Han. Generalizable whole slide image classification with fine-grained visual-semantic interaction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11398–11407, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.422, + 0.895, + 0.476 + ], + "angle": 0, + "content": "[18] Ming Y Lu, Tiffany Y Chen, Drew FK Williamson, Melissa Zhao, Maha Shady, Jana Lipkova, and Faisal Mahmood. Aibased pathology predicts origins for cancers of unknown primary. Nature, 594(7861):106-110, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.478, + 0.895, + 0.545 + ], + "angle": 0, + "content": "[19] Ming Y Lu, Bowen Chen, Drew FK Williamson, Richard J Chen, Ivy Liang, Tong Ding, Guillaume Jaume, Igor Odintsov, Long Phi Le, Georg Gerber, et al. A visual-language foundation model for computational pathology. Nature Medicine, 30(3):863-874, 2024. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.547, + 0.895, + 0.613 + ], + "angle": 0, + "content": "[20] Ming Y Lu, Bowen Chen, Drew FK Williamson, Richard J Chen, Melissa Zhao, Aaron K Chow, Kenji Ikemura, Ahrong Kim, Dimitra Pouli, Ankush Patel, et al. A multimodal generative ai copilot for human pathology. Nature, pages 1-3, 2024. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.616, + 0.895, + 0.695 + ], + "angle": 0, + "content": "[21] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.698, + 0.895, + 0.779 + ], + "angle": 0, + "content": "[22] Mehmet Saygin Seyfioglu, Wisdom O Ikezogwo, Fatemeh Ghezloo, Ranjay Krishna, and Linda Shapiro. Quilt-llava: Visual instruction tuning by extracting localized narratives from open-source histopathology videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13183-13192, 2024. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.781, + 0.895, + 0.848 + ], + "angle": 0, + "content": "[23] Zhuchen Shao, Yang Chen, Hao Bian, Jian Zhang, Guojun Liu, and Yongbing Zhang. Hvtsurv: Hierarchical vision transformer for patient-level survival prediction from whole slide image. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2209-2217, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.849, + 0.895, + 0.89 + ], + "angle": 0, + "content": "[24] Andrew H Song, Guillaume Jaume, Drew FK Williamson, Ming Y Lu, Anurag Vaidya, Tiffany R Miller, and Faisal Mahmood. Artificial intelligence for digital and computa" + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.104, + 0.895, + 0.89 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.941, + 0.516, + 0.953 + ], + "angle": 0, + "content": "5142" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.135, + 0.104, + 0.482, + 0.129 + ], + "angle": 0, + "content": "tional pathology. Nature Reviews Bioengineering, 1(12): 930-949, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.106, + 0.133, + 0.483, + 0.224 + ], + "angle": 0, + "content": "[25] Joey Spronck, Thijs Gelton, Leander van Eekelen, Joep Bogaerts, Leslie Tessier, Mart van Rijthoven, Lieke van der Woude, Michel van den Heuvel, Willemijn Theelen, Jeroen van der Laak, et al. nnunet meets pathology: bridging the gap for application to whole-slide images and computational biomarkers. In Medical Imaging with Deep Learning, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.105, + 0.228, + 0.484, + 0.307 + ], + "angle": 0, + "content": "[26] Yuxuan Sun, Chenglu Zhu, Sunyi Zheng, Kai Zhang, Lin Sun, Zhongyi Shui, Yunlong Zhang, Honglin Li, and Lin Yang. Pathasst: A generative foundation ai assistant towards artificial general intelligence of pathology. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 5034-5042, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.106, + 0.311, + 0.482, + 0.39 + ], + "angle": 0, + "content": "[27] Eugene Vorontsov, Alican Bozkurt, Adam Casson, George Shaikovski, Michal Zelechowski, Kristen Severson, Eric Zimmermann, James Hall, Neil Tenenholtz, Nicolo Fusi, et al. A foundation model for clinical-grade computational pathology and rare cancers detection. Nature Medicine, pages 1-12, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.106, + 0.392, + 0.482, + 0.457 + ], + "angle": 0, + "content": "[28] Xiyue Wang, Junhan Zhao, Eliana Marostica, Wei Yuan, Ji etian Jin, Jiayu Zhang, Ruijiang Li, Hongping Tang, Kanran Wang, Yu Li, et al. A pathology foundation model for cancer diagnosis and prognosis prediction. Nature, pages 1-9, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.106, + 0.461, + 0.482, + 0.526 + ], + "angle": 0, + "content": "[29] Feng Xu, Chuang Zhu, Wenqi Tang, Ying Wang, Yu Zhang, Jie Li, Hongchuan Jiang, Zhongyue Shi, Jun Liu, and Mulan Jin. Predicting axillary lymph node metastasis in early breast cancer using deep learning on primary tumor biopsy slides. Frontiers in oncology, 11:759007, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.106, + 0.529, + 0.482, + 0.595 + ], + "angle": 0, + "content": "[30] Hanwen Xu, Naoto Usuyama, Jaspreet Bagga, Sheng Zhang, Rajesh Rao, Tristan Naumann, Cliff Wong, Zelalem Gero, Javier González, Yu Gu, et al. A whole-slide foundation model for digital pathology from real-world data. Nature, pages 1-8, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.106, + 0.598, + 0.482, + 0.663 + ], + "angle": 0, + "content": "[31] Yingxue Xu, Yihui Wang, Fengtao Zhou, Jiabo Ma, Shu Yang, Huangjing Lin, Xin Wang, Jiguang Wang, Li Liang, Anjia Han, et al. A multimodal knowledge-enhanced whole-slide pathology foundation model. arXiv preprint arXiv:2407.15362, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.106, + 0.666, + 0.482, + 0.718 + ], + "angle": 0, + "content": "[32] An Yang, Baosong Yang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Zhou, Chengpeng Li, Chengyuan Li, Dayiheng Liu, Fei Huang, et al. Qwen2 technical report. arXiv preprint arXiv:2407.10671, 2024. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.106, + 0.721, + 0.482, + 0.774 + ], + "angle": 0, + "content": "[33] Fengtao Zhou and Hao Chen. Cross-modal translation and alignment for survival analysis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 21485-21494, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.105, + 0.104, + 0.484, + 0.774 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.941, + 0.516, + 0.952 + ], + "angle": 0, + "content": "5143" + } + ] +] \ No newline at end of file diff --git a/2025/SlideChat_ A Large Vision-Language Assistant for Whole-Slide Pathology Image Understanding/5542bef9-2edf-4ac5-b8df-cf3f63aeaa6c_origin.pdf b/2025/SlideChat_ A Large Vision-Language Assistant for Whole-Slide Pathology Image Understanding/5542bef9-2edf-4ac5-b8df-cf3f63aeaa6c_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..3a793bf544477e68647dd79289aeb16140b8f9d0 --- /dev/null +++ b/2025/SlideChat_ A Large Vision-Language Assistant for Whole-Slide Pathology Image Understanding/5542bef9-2edf-4ac5-b8df-cf3f63aeaa6c_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6256d313df125f1d14b85539e899c3f70b4f12ffa3000a31cdc309b515922aa8 +size 1346416 diff --git a/2025/SlideChat_ A Large Vision-Language Assistant for Whole-Slide Pathology Image Understanding/full.md b/2025/SlideChat_ A Large Vision-Language Assistant for Whole-Slide Pathology Image Understanding/full.md new file mode 100644 index 0000000000000000000000000000000000000000..8220674beb2d01c5afbee87a34de5fb5b9b05b01 --- /dev/null +++ b/2025/SlideChat_ A Large Vision-Language Assistant for Whole-Slide Pathology Image Understanding/full.md @@ -0,0 +1,231 @@ +# SlideChat: A Large Vision-Language Assistant for Whole-Slide Pathology Image Understanding + +Ying Chen $^{1,2*}$ Guoan Wang $^{1,3*}$ Yuanfeng Ji $^{4*†}$ Yanjun Li $^{1,3}$ Jin Ye $^{1,5}$ Tianbin Li $^{1}$ Ming Hu $^{1,5}$ Rongshan Yu $^{2}$ Yu Qiao $^{1}$ Junjun He $^{1\dagger}$ + +$^{1}$ Shanghai AI Laboratory $^{2}$ Xiamen University $^{3}$ East China Normal University $^{4}$ Stanford University $^{5}$ Monash University + +# Abstract + +Despite the progress made by multimodal large language models (MLLMs) in computational pathology, they remain limited by a predominant focus on patch-level analysis, missing essential contextual information at the whole-slide level. The lack of large-scale instruction datasets and the gigapixel scale of whole slide images (WSIs) pose significant developmental challenges. In this paper, we present SlideChat, the first vision-language assistant capable of understanding gigapixel whole-slide images, exhibiting excellent multimodal conversational capability and response complex instruction across diverse pathology scenarios. To support its development, we created SlideInstruction, the largest instruction-following dataset for WSIs consisting of 4.2K WSI captions and 176K VQA pairs with multiple categories. Furthermore, we propose SlideBench, a multimodal benchmark that incorporates captioning and VQA tasks to assess SlideChat's capabilities in various settings such as microscopy, diagnosis and clinical. Compared to both general and specialized MLLMs, SlideChat exhibits exceptional capabilities, achieving state-of-the-art performance on 18 of 22 tasks. For example, it achieved an overall accuracy of $81.17\%$ on SlideBench-VQA (TCGA), and $54.15\%$ on SlideBench-VQA (BCNB). Our code, data, and model is publicly accessible at https://uni-medical.github.io/SlideChat.github.io. + +# 1. Introduction + +Computational pathology aims to improve the analysis of digitized tissue samples [14, 24], such as whole slide images (WSIs), by applying artificial intelligence to aid in the diagnosis, identification, and understanding of disease. Recently, the development of this field has gained rapid + +![](images/334f9be18accb27a827000a8d6344fe38babd823e20549a014fd4917b3a36b4f.jpg) +A +B +Figure 1. SlideChat is the first large vision-language assistant specifically designed for whole-slide pathology analysis. SlideChat can generate comprehensive descriptions of whole-slide images and provide contextually relevant responses across various applications. + +Summarize the key pathological findings from the whole slide image. + +This pathological examination reveals the high-grade urothelial carcinoma with ... + +What is the pathological stage (pT) of the transitional cell carcinoma? A. pT1 B. pT2 C. pT3 D. pT4 + +B. pT2 + +Microscopy + +$\spadesuit$ Tissue Architecture $\spadesuit$ Tumor Characteristics +Histopathological Changes +$\spadesuit$ Cytomorphological Characteristics + +Diagnosis + +$\spadesuit$ Disease Detection Staging Grading +$\spadesuit$ Disease Classification +$\spadesuit$ Differential Diagnosis + +Clinical + +Treatment Guidance Risk Factors +Prognostic Assessment +$\spadesuit$ Biomarker Analysis + +momentum, mainly driven by breakthroughs in the visual foundation model [9, 27, 30]. These models learn generalized representations by pre-training on large-scale data and perform well in various downstream tasks, such as cancer detection and biomarker prediction. Building on this base, integration with the powerful Large Language Models (LLMs) further advances the development of the Multimodal Large Language Model (MLLMs) [20], which has made great strides in responding to more complex, open-ended visual queries, enabling it to serve as a versatile assistant at various stages of medical care, including clinical decision-making, education, and research. + +Nevertheless, there are some major challenges that hinder the development and use of pathology MLLMs for real-world applications. First, it is challenging to develop a MLLM architecture that can effectively be capable of gigapixel whole slides (e.g., $100,000 \times 100,000$ pixels). Existing models [20, 22, 26] often process whole slides by extracting small patch/ROI-level data for subsequent analysis, resulting in a limited understanding of global slide con + +text and suboptimal performance in some complex pathological analysis. Second, publicly available multi-modal pathology slide datasets are relatively scarce and of varying quality [7, 8, 12], which limits the development of MLLMs trained on such data. + +In this paper, we present SlideChat (see Fig. 1), the first open-source vision-language assistant capable of understanding gigapixel whole-slide images. First, SlideChat is trained on SlideInstruction, a large-scale multi-modal instruction dataset encompassing data from The Cancer Genome Atlas (TCGA) [15] via our specifically designed data processing pipeline (see Fig. 3 (A)). SlideInstruction contains 4,181 WSI-caption pairs and 175,753 visual question-answer pairs from 3,294 patients, covering 10 cancer types. The question-answer pairs include both open-ended and closed-ended questions, further divided into 13 subcategories, covering a diverse range of clinical tasks. SlideInstruction is more than 20 times larger than previous public instruction datasets in the number of instructions. Second, we propose SlideChat, a novel architecture in LLaVA style for capable multi-modal analysis of gigapixel whole slides. A gigapixel whole slide is first divided into a series of patches, each of which is individually processed by a patch-level encoder to extract local features. The resulting long sequence of patch tokens is then processed by a slide-level encoder employing sparse attention to enhance the slide-level features. Finally, enhanced patch tokens are fed into the LLM via a projector, which processes user queries and generates responses. + +To systematically evaluate the performance of SlideChat in real-world scenarios, we establish a comprehensive digital pathology benchmark named SlideBench, encompassing more than 20 clinical tasks, using data from both TCGA and the in-the-wild Early Breast Cancer CoreNeedle Biopsy (BCNB) dataset. This resulted in three test sets: SlideBench-Caption, comprising 734 WSI-caption pairs; SlideBench-VQA (TCGA), comprising 7,827 VQA pairs covering 13 different tasks; and SlideBench-VQA (BCNB), including a total of 7,274 VQA entries from 1,058 patients, covering seven different tasks. Additionally, we compare the performance of SlideChat on another externally proposed dataset, WSI-VQA [8], to further validate its effectiveness. We compare our model with the currently available state-of-the-art general and medical-specialized MLLMs including GPT-4o, LLaV-aMed [16], MedDr [13] and Quilt-LLaVA [22]. Benefiting from large-scale, high-quality training and effective local-global context modeling, SlideChat achieves state-of-the-art performance on 18 out of 22 tasks, with significant improvements over the second-best method $10\%$ on 9 tasks on four benchmarks. Specifically, SlideChat achieves an average accuracy improvement of $13.47\%$ over the second-best model on SlideBench-VQA (TCGA), an average improvement of $12.59\%$ on + +SlideBench-VQA (BCNB), and an improvement of $5.82\%$ on WSI-VQA. Finally, to accelerate research progress in digital pathology, we make SlideChat fully open-weight, including source code and model weights as well as instruction and benchmark datasets. The key contributions are summarized four-fold in the following: + +- We develop SlideChat, the first vision-language assistant capable of understanding gigapixel whole-slide images, achieving state-of-the-art performance on multiple benchmarks. +- We create SlideInstruction, a largest comprehensive WSI instruction-following dataset containing 4.2K WSI-caption pairs and 176K VQA pairs. +- We establish SlideBench, a WSI multi-modal benchmark comprising SlideBench-Caption, SlideBench-VQA (TCGA), and SlideBench-VQA (BCNB), covering 21 different clinical tasks. +- We release SlideChat, SlideInstruction and SlideBench as open-source resources to facilitate research and development in computational pathology. + +# 2. Related Works + +Whole Slide Image Analysis Whole slide images are pivotal in modern pathology, enabling comprehensive analysis of tissue samples for tasks such as predicting patient prognosis, classifying cancer subtypes, and identifying biomarkers [17, 18, 23-25, 33]. Recent studies have leveraged pathology foundational models [3, 28, 31] to enhance WSIs analysis, either through fine-tuning for specific downstream tasks or by employing zero-shot prediction approaches in CLIP [21] style. Although these models are effective in task-specific applications, their reliance on fine-tuning or limited zero-shot capabilities restricts their generalizability across diverse and complex user instructions. + +MLLMs in Computational Pathology The paradigm of MLLMs enables them to effectively respond to more complex, open-ended visual queries while processing pathology images, thus providing significant value across various medical stages. PathChat [20] is a vision-language assistant designed for pathology, developed with 450K private instruction pairs to handle both visual and natural language queries. Quilt-Instruct [22] is a large-scale dataset comprising 107K question-answer pairs. Building on Quilt-Instruct, Quilt-LLAVA [22] is a model designed for diagnostic reasoning across multiple image patches, leveraging its extensive question-answer pairs to accurately interpret complex H&E images. PathAsst [26] combines a pathology-specific CLIP model with Vicuna-13b [10] to create a multimodal generative foundational model tailored for pathology. However, current MLLMs primarily only focus on patch or ROI data, limiting their utility for slide-level clinical applications where broader contextual understanding is crucial. + +![](images/469b18531b509cecfc8ab5beb7f7537e3b21621b056d4c05c0aff9487f21fdaa.jpg) +Figure 2. Overview of our SlideChat. (A) SlideChat serializes each input WSI into a sequence of $224 \times 224$ patches, converting each into visual embeddings with a patch-level encoder. A slide-level encoder then interacts with these features to generate contextual embeddings. Then, a multimodal projector maps the visual features from the slide-level encoder into a unified space, aligned seamlessly with the LLM. (B) SlideChat was trained for two stages: Cross-Domain Alignment and Visual Instruction Learning. + +# 3. SlideChat + +# 3.1. Architecture + +To achieve the goal of analyzing gigapixel whole-slide images in a multimodal setting, as shown in Fig. 2, SlideChat consists of four key designs: the patch-level encoder, the slide-level encoder, the multimodal projector module, and the large language model. Our method starts by partitioning the WSI into smaller $224 \times 224$ pixel patches, making it computationally feasible to process such large images. These patches are then passed through a well-trained, frozen patch-level encoder [19], which extracts localized features from each individual patch, capturing fine-grained details such as cellular structures. Building on this, we employ LongNet [11, 30] as slide-level encoder to enhance the patch-level embeddings and capture global patterns across the entire slide. This encoder uses sparse attention mechanisms to learn both local and global contextual information, enabling the model to perceive intricate local features while capturing the broader context, which is critical for comprehensive pathological assessments. Following the slide-level encoding, SlideChat incorporates a multimodal projector that maps these patch tokens into a unified space aligned with the LLM. This ensures that the visual features extracted from the WSIs are effectively transformed into representations compatible with the language model, facilitating seamless integration and interaction between visual and textual data. Concurrently, the model accepts natural language instructions from users, such as "What is the type of tumor in the image?". These textual queries are processed by the LLM, which comprehends the textual input and integrates it with the visual features extracted from the WSIs, enabling accurate and contextually relevant diagnoses + +tic responses. This multimodal reasoning capability allows SlideChat to provide accurate and contextually relevant answers to complex pathology-related questions, thereby supporting clinical decision-making, education, and research across various medical stages. + +# 3.2.Data + +SlideInstruction There is a notable lack of large-scale multimodal pathology datasets supporting the training of vision-language assistants for whole-slide image understanding. To support the training of SlideChat, we develop SlideInstruction, a comprehensive instruction dataset, sourced from the TCGA database, comprising 4,915 WSI-report pairs from 4,028 patients. Fig. 3 illustrates our entire data curation pipeline. We initially prompt GPT-4 to refine the pathology reports, and clean up the noise in the report including unrelated symbols, technical details of pathology department procedures, specimen handling and processing information, redundant administrative or legal statements, and some repeated information. For the refined pathology reports, we further employ GPT-4 to generate high-quality multimodal data, comprising two main components: (1) WSI-Caption Data: We craft concise, clinically relevant summaries for each whole slide image by prompting the language model to extract key pathological findings. These summaries were structured into coherent paragraphs that highlighted crucial clinical details such as diagnostic results, tumor characteristics, margin status, and lymph node involvement, ensuring the caption dataset is both focused and informative. (2) WSI Instruction-Following Data: To enhance the model's ability to follow instructions and improve its comprehension of pathology images, we leveraged GPT-4 to generate tailored question + +![](images/f0daefbfd8e147bffc5e8a200b3f1c7ba6517ebcc8d2cab7bf276c51d24f595b.jpg) +A + +![](images/87b8387cb638bbb9b4ca94bda3ad892e90ad4a42a6350df49a6db0e76e996a37.jpg) +B + +![](images/f97990cab839f83140c393912fd67655dc93a45953901eca9e8bd6e2717d5d57.jpg) +C + +![](images/b127e927e5f640d0d6d859e17226266b5adfd02d8cce2a44dcd32e4c6d80f4f7.jpg) +Figure 3. (A) Overview of the SlideInstruction generation pipeline. We prompt GPT-4 to extract the WSI-Caption, Open-set VQA and Closed-set VQA from reports. (B) For the generated Closed-set VQA, we employ LLMs to filter low-quality QA pairs and involve pathologists for validation, resulting in the creation of SlideBench-VQA. (C) Examples of WSI caption and instruction-following scenarios in microscopy, diagnostics, and clinical applications. For additional examples, please refer to Supplementary Material. + +[Caption] #Q:Describe the whole slide image. #A:The bladder exhibits an invasive poorly differentiated urothelial carcinoma, grade 4/4, extending through the bladder neck wall.... + +[VQA-Diagnosis] +#Q: From the given whole slide image, which histological subtype of lung adenocarcinoma is most predominantly observed? +#A: It is acinar subtype observed. + +![](images/748388969461b7e9b57d20c3a086a33fcff6685ce6b440d446124d5b332d905b.jpg) + +![](images/cd2511ed303e1a629613496c681fe0adb36bfa337b7ac2f4e9595832e708f76f.jpg) + +[VQA-Microscopy] +#Q: What histological feature observable in a whole slide image differentiates anaplastic astrocytoma from glioblastoma? +#A: Absence of endothelial proliferation. + +[VQA-Clinical] #Q: Based on the histopathological features, what is the primary recommended treatment? #A: Simple mastectomy followed by hormone therapy. + +and-answer pairs for each WSI report. Drawing inspiration from PathChat [20], we structure these questions into 3 broad categories—microscopy, diagnosis, and clinical considerations—which represent key stages in the pathology workflow, and 13 narrow categories focusing on specific aspects within each stage (Fig. 1 B). To construct a comprehensive instructional dataset, we began by generating two open-ended and two closed-ended QA pairs for each report within each specific category, taking advantage of the rich content found in the majority of TCGA reports. Then, some QA pairs were excluded, such as those lacking corresponding category information. Regarding the train/test split, it is worth noting that the WSI-report datasets from TCGA include 2 types: (a) one report linked to multiple WSIs, and (b) one report linked to a single WSI. For type (a), where specific diagnostic details may not align perfectly with each WSI, we include all WSIs in the training set to introduce some "noisy data", which can enhance model robustness. For type (b), $80\%$ of WSIs are allocated to the training + +set and $20\%$ to the test set. Finally, there are 4,181 WSIs for training and 734 WSIs for testing. Consequently, we construct a large-scale training set named SlideInstruction, comprising 4,181 WSI captions and 175,753 instruction-following VQA pairs across 13 narrow categories. + +SlideBench To systematically evaluate the performance of SlideChat, we include the remaining 734 WSI captions from the test set, along with a substantial number of closed-set VQA pairs, to establish a comprehensive evaluation benchmark. First, we construct a set named SlideBench-Caption based on the WSI-Caption data to evaluate the model's ability to generate accurate and coherent descriptions of whole slide images. Secondly, we construct SlideBench-VQA (TCGA) based on closed-set VQA pairs along with test WSIs, aiming to evaluate various aspects of model performance. As shown in Fig. 3 (B), to improve the quality of the testing benchmarks, we employ four advanced large language models, including GPT-4 [2], InternLM2- + +Chat-7B [6], Qwen-7B-Chat [5], and DeepSeek-7B-Chat, to filter closed-set VQAs by predicting answers based solely on the question text. Any questions for which at least three of these models provided correct answers are subsequently excluded. Following this automated filtering, five expert pathologists are invited to review and amend the remaining questions. The review process is guided by the following criteria: (1) Whether the correct answer necessitates image interpretation; (2) Whether the question and its corresponding answer are logically and coherently structured; and (3) Whether the question aligns appropriately with the designated broad and narrow categories. VQA pairs failing to meet these criteria are excluded by the pathologists. Consequently, the SlideBench-VQA (TCGA) comprises 7,827 VQAs across 13 categories, with some examples illustrated in Fig. 3 C. Additionally, we incorporate the in-the-wild Early Breast Cancer Core-Needle Biopsy (BCNB) WSI dataset [29], which encompasses a diverse patient population and a variety of clinical task labels, to enhance the test set benchmark and more comprehensively assess the model's generalization capabilities. In detail, we convert the BCNB dataset into a VQA format by rephrasing the classification objectives into a specific template as questions, while transforming the original multi-class labels into selectable options, and integrating it into SlideBench as an external subset, named SlideBench-VQA (BCNB). This dataset comprises 7,247 VQA pairs from 1,058 patients, specifically designed to evaluate SlideChat's zero-shot generalization capability across 7 distinct classification tasks. + +# 3.3. Two-stage Training + +Stage 1: Cross-Domain Alignment. SlideChat adopts a two-stage training approach (see Fig. 2 B). In the first stage, the primary objective is to align the LLM's word embeddings with the visual features extracted from whole slide images. This alignment enables the LLM to interpret visual representations from the slide-level encoder, facilitating the effective utilization of the intricate features within the slides. During this stage, SlideChat is trained to generate descriptive captions using 4.2K WSI-caption pairs from SlideInstruction. Specifically, only the slide-level encoder and projection matrix are updated, while the patch-level encoder and LLM weights remain fixed. + +Stage 2: Visual Instruction Learning. In the second stage, we focus on visual question-answering tasks to train the model to accurately respond to domain-specific questions concerning whole slide images. During this phase, the model develops the ability to handle a broad range of multimodal instructions, enabling it to generate answers by effectively integrating both visual and textual information. For example, the model must perform various pathology tasks, such as describing the extent of tumor invasion or assess + +ing the degree of cellular differentiation. To accomplish this, we utilize 176K WSI VQAs from SlideInstruction in the second training stage, allowing the slide encoder, projection layer, and large language model components to be fully trainable to ensure comprehensive adaptability. This approach significantly enhances the model's capability to handle diverse pathology-related tasks, thereby increasing its effectiveness in real-world clinical and research settings. + +# 4. Experiment + +We conducted the following experiments to evaluate three key aspects of SlideChat: (1) its whole slide image captioning capability, which assesses proficiency in generating descriptive captions that accurately summarize the critical pathological features of a WSI; (2) its visual question-answering (VQA) ability across various complex pathological scenarios and its generalizability in zero-shot settings; and (3) SlideChat's ability to process gigapixel WSIs, capturing both essential global context and intricate details, thereby enhancing its performance compared to patch-level MLLMs. For WSI captioning baselines, we benchmark against MI-Gen [7], a state-of-the-art method specifically designed for this task. Given that existing MLLMs cannot handle the gigapixel scale of whole slide images, we establish baseline comparisons using two approaches: (1) randomly selecting 30 patches from each WSI and inputting them into MLLMs (e.g., GPT-4o [2], LLaVA-Med [16], MedDr [16], Quilt-LLaVA [22]), followed by a majority voting scheme to generate slide-level predictions; and (2) directly inputting a WSI thumbnail, resized to $1024 \times 1024$ pixels, referred to as Slide (T), into the MLLMs. For VQA tasks, we further evaluate performance by comparing against random prediction baselines and text-only models, thereby assessing the incremental contribution of visual information. Unless otherwise specified, SlideChat is configured with the patch-level encoder CONCH [19], the slide-level encoder LongNet [11], and utilizes the Qwen2.5-7B-Instruct [32] as LLM. + +Table 1. Captioning performance across different methods on SlideBench-Caption. + +
MethodsInputBLEU-1BLEU-2BLEU-3BLEU-4Rouge-LGPT score
GPT-4oPatch0.160.030.010.010.131.54
GPT-4oSlide (T)0.100.030.010.010.111.03
Quilt-LLaVAPatch0.300.190.110.050.183.87
Quilt-LLaVASlide (T)0.230.090.040.010.161.89
MI-GenSlide0.370.240.150.100.254.14
SlideChatSlide0.370.210.120.080.244.11
+ +SlideBench-Caption We report BLEU, ROUGE, and GPT scores to evaluate caption generation performance in Tab. 1. For the GPT score, we use GPT-4 to assess the + +Table 2. VQA performance with different methods. + +
MethodsInputSlideBench-VQA (TCGA)SlideBench-VQA (BCNB)WSI-VQA*
MicroscopyDiagnosisClinicalOverall
RandomText24.4424.9126.4425.0224.4024.14
GPT-438.2829.0945.0037.25018.60
GPT-4oPatch62.8946.6966.7757.9141.4330.41
MedDr73.3057.7874.2567.7033.6754.36
LLaVA-Med47.3432.7847.9642.0030.126.31
Quilt-LLaVA57.7635.9653.0748.0732.1944.43
GPT-4oSlide (T)38.2823.1043.4234.07014.03
MedDr70.4852.4772.8064.2535.4850.95
LLaVA-Med45.8227.5840.8437.39018.79
Quilt-LLaVA49.1226.9744.7539.3941.5535.40
SlideChatSlide87.64(+14.34)73.27(+15.49)84.26(+10.01)81.17(+13.47)54.14(+12.59)60.18(+5.82)
+ +similarity between the generated captions and the ground truth, providing an overall score on a scale of 1 to 10, with higher scores indicating better performance. When utilizing patch-level inputs, GPT-4o generates individual descriptions for each patch, which are subsequently integrated to create the final slide-level caption. However, this approach yields poor performance, as evidenced by a BLEU-1 score of 0.16 and a GPT-score of 1.54. These results suggest that the patch-based method fails to adequately capture the broader context necessary for accurate WSI captioning. When the WSI thumbnail of size $1024 \times 1024$ pixels is used as input to GPT-4o, performance decreases further, with a BLEU-1 score of 0.10 and a GPT-score of 1.03. This suggests that while the thumbnail offers a global view of the slide, it may lack the resolution and detail necessary for generating precise and informative captions. In contrast, MI-Gen, a model specifically designed for WSI captioning, demonstrates significantly superior performance across all metrics, achieving a BLEU-1 score of 0.37 and a GPT score of 4.14. Similarly, SlideChat, shows competitive results with a BLEU-1 score of 0.37 and a GPT score of 4.11. These outcomes highlight SlideChat's ability to effectively integrate both local and global information from the slides and confirm its efficiency in describing whole-slide images. + +SlideBench-VQA (TCGA) We further evaluate SlideChat's overall performance on the multiple-choice VQA benchmark. The results on SlideBench-VQA (TCGA) in Tab. 2, compare different methods across three categories: microscopy, diagnosis, and clinical, along with an overall performance score. Random selection achieves an overall score of $25.02\%$ accuracy, serving as a baseline for answer distribution but demonstrating poor performance. While GPT-4, relying solely on text input, outperforms random predictions, it continues to struggle + +with accurately answering questions. When GPT-4o incorporates patch-level inputs, its performance improves markedly, reaching a score of $57.91\%$ and underscoring the crucial role of detailed visual data. However, using a WSI thumbnail results in a lower score of $34.07\%$ , as the reduced detail restricts its ability to deliver precise answers. MedDr performs well, achieving an overall score of $67.70\%$ with patch-level inputs, though this drops slightly to $64.25\%$ when using the slide thumbnail due to the loss of fine visual details. SlideChat outperforms all other methods, attaining a leading overall accuracy of $81.17\%$ , excelling across all categories and significantly surpassing the competition. Even in more fine-grained pathological scenarios, as depicted in the left portion of Fig. 4, SlideChat remains the top-performing model across 13 tasks, particularly in areas such as cytomorphological characteristics, histopathological changes, disease detection, disease classification, and staging and grading, which require the identification of complex pathological visual features. Compared with baselines taking some patches or slide thumbnail as inputs, SlideChat has the capability to analyze a significantly greater number of pathological features with enhanced detail, effectively capturing both localized features and overarching global patterns, allowing SlideChat to provide more accurate and nuanced insights into pathological variations. + +SlideBench-VQA (BCNB) In the zero-shot VQA setting, SlideChat significantly outperforms all other models, achieving the highest overall score of $54.14\%$ on SlideBench-VQA (BCNB). It particularly excels in identifying tumor types, far surpassing other baselines in this task. This performance highlights SlideChat's generalization capability across a wide range of tasks. When taking patches as inputs, GPT-4o outperforms both MedDr and LLaVA + +![](images/977a5e617e035189e47c3f27f96daab07f3564ef79abc62b7cccc95c21c30d96.jpg) +Figure 4. Accuracy on different tasks in SlideBench-VQA (TCGA) (left) and SlideBench-VQA (BCNB) (right). + +Med, achieving a score of $41.43\%$ , though it still falls short of SlideChat by $12.59\%$ . Notably, GPT-4o and LLaVA-Med performed very poorly, achieving a score of zero across all tasks when evaluated using slide thumbnails from this testing set. MedDr and Quilt-LLaVA both achieve moderate results, with MedDr scoring $33.67\%$ on patches and $35.48\%$ on thumbnails, while Quilt-LLaVA attains $32.19\%$ on patches and $41.55\%$ on thumbnails. This outcome highlights that, for complex WSIs, relying solely on relatively sufficient visual features is inadequate for effectively supporting a range of tasks. In the more fine-grained pathological tasks of the BCNB benchmark, SlideChat attains state-of-the-art performance on 5 out of 7 tasks (Fig. 4), further demonstrating the effectiveness of SlideChat. + +WSI-VQA* We also curated a subset of closed-set VQA pairs from the public WSI-VQA [8] dataset, referred to as WSI-VQA*, which exclusively consists of samples that were not in the training set, to evaluate the model's performance. SlideChat demonstrates the highest performance with a score of 60.18. Although MedDr performs well with both patch inputs (54.36%) and slide thumbnail inputs (50.95%), it still falls short compared to SlideChat. GPT-4o struggles significantly, especially with slide thumbnails, scoring only 14.03%, which highlights the limitations of using lower-resolution inputs. SlideChat leverages both detailed local information and broader context to process WSIs, making it the most effective model for this benchmark. This further emphasizes its superior capability in handling whole-slide data for VQA tasks. + +Ablation We performed ablation experiments from different perspectives as follows: a) Large Language Model Comparison: Firstly, we compare the performance of several large language models, each with a parameter scale of approximately 7 billion, including Vicuna-7B-v1.5 [10], Phi-3-Mini-4k-Instruct [1], Qwen1.5-7B-Chat [5], Llama3-8B-Instruct [4], InternLM2-Chat-7B [6], and Qwen2.5-7B + +Table 3. Performance Comparison of LLMs and Slide Encoder on WSI Captioning and VQA Tasks. + +
LLMsSlide EncoderCaptionVQA (TCGA)VQA (BCNB)WSI-VQA*
Vicuna-7B-v1.5✓3.2841.4341.4331.98
Phi-3-Mini-4k-Instruct✓2.6679.9343.9260.18
Qwen1.5-7B-Chat✓2.9277.6344.0756.89
Llama3-8B-Instruct✓3.3078.7842.8255.25
Internlm2-Chat-7B✓3.3079.1052.1356.76
Qwen2.5-3B-Instruct✓3.4080.3245.7956.38
Qwen2.5-14B-Instruct✓3.3982.1451.5760.94
Qwen2.5-7B-Instruct✓4.1181.1754.1460.18
Qwen2.5-7B-Instruct×3.9581.2145.4959.67
+ +Instruct [32]. Specifically, we measured their performance on the SlideBench-Caption using the GPT-score and their accuracy on three VQA datasets. As shown in Tab. 3, the results demonstrate that SlideChat, powered by the Qwen2.5-7B-Instruct model, achieved the highest performance across all tasks, particularly excelling in the captioning task. These findings underscore the significant potential of developing SlideChat with the Qwen2.5-7B-Instruct model. Utilizing the Qwen2.5 model, we further evaluated models of varying scales and discovered that larger models generally exhibited superior performance, particularly the 7B and 14B parameter models. While these two models showed comparable performance across the three benchmarks, the 14B model surpassed the 7B model by $2.57\%$ on the SlideBench-VQA (TCGA). Given computational resource constraints, SlideChat uses the 7B model by default to achieve the best balance between performance and resource efficiency. b) Slide-level encoder effectiveness: We investigate the effectiveness of the slide-level encoder by initially removing it from SlideChat and employing a two-stage training approach. In the first stage, we exclusively trained the projection layers. However, this approach failed to reduce training loss or generate coherent text effectively, + +![](images/a2ad6f0d11e37162874167c4a6b46e6971bca861a7172843b4e2405a85783298.jpg) +A + +![](images/3e33e9c466c428539ae7aade1888ad947d76b9eae9b5b2481b29ffafcfab2344.jpg) +Select the most accurate description of the cytomorphological characteristics from image. A. Cells with increased nuclear-to-cytoplasmic ratio, hyperchromatic nuclei, and prominent nucleoli. B. Cells with uniform nuclei and abundant cytoplasm. C. Cells with minimal nuclear abnormalities. D. Cells with enlarged, clear nuclei and small nucleoli. + +A. Cells with increased nuclear-to-cytoplasmic ratio, hyperchromatic nuclei, and prominent nucleoli. + +![](images/ab789ba476b9a5cf8bb971c935c6b7fcd188f19279163dbc5653a03a3f13afcc.jpg) +B + +![](images/5c7d913f3149088674af34884c0eb1efe249a856cac6e55d2eb7590eafccf217.jpg) +Figure 5. Interpretability and visualization. We identify the top five patch tokens with the highest attention scores associated with the output text responses. + +A. Dense collagen deposition with reduced cellularity + +likely due to the difficulty of learning the complex visual features of WSIs without the slide-level encoder. When using a frozen LLM for complex text generation, a simple projection proved inadequate for effectively integrating and aligning visual and textual features. Subsequently, we consider combining data from both stages and training SlideChat without the slide-level encoder by simultaneously updating both the projection layers and the LLM. Under this paradigm, performance on SlideBench-VQA (TCGA) and WSI-VQA*, which share the same distribution as the training set, was comparable to the two-stage training configuration with the slide-level encoder. However, a significant decline is observed when evaluating SlideBench-VQA (BCNB), which originates from a different domain; overall performance dropped by over $10\%$ (Tab. 3), indicating a substantial reduction in the model's generalization ability. Therefore, we recommend incorporating a slide-level encoder to capture the complex visual features of whole slide, as it is particularly effective for cross-domain alignment and enhances the model's generalization performance. + +Interpretability Despite SlideChat demonstrating promising results, concerns remained regarding the model's perception of large pathological slides. To further assess the model's interpretability, we calculated the correlation between the text output and specific image patches, thereby obtaining patch-level attention scores. By identifying the most significant patches, we gained insights into the precise areas the model focused on during response generation. Highlighting the most relevant regions of the tissue slides not only enhances transparency and bolsters the reliability of the model's outputs but also assists pathologists by directing attention to critical areas requiring closer scrutiny. Ultimately, such interpretability is essential for fostering trust in AI-assisted diagnostics and enhancing the precision and efficiency of clinical evaluations. As shown in Fig. 5, we are pleased to observe that the top five important patches identified by the model closely + +corresponded with the features described in SlideChat's output. Our extraction method retrieves attention weights for patch tokens from each generated token, averaging them across layers and heads. We then identify the top five patch tokens with the highest attention weights for further analysis. For example, in Fig. 5 (A), the highlighted patches clearly emphasized regions exhibiting an increased nuclear-to-cytoplasmic ratio, hyperchromatic nuclei, and prominent nucleoli. Similarly, in Fig. 5 (B), the selected patches demonstrated areas with dense collagen deposition and reduced cellularity, as detailed in the model's response. This alignment between the highlighted image regions and the textual outputs significantly enhances the model's interpretability, providing increased confidence that it accurately captures and assesses relevant histopathological features. Such consistency deepens our understanding of the model's reasoning processes regarding pathological slides and underscores the potential for integrating these AI systems into clinical workflows with greater assurance. + +# 5. Conclusion + +In this work, we present SlideChat, the first vision-language assistant capable of understanding gigapixel whole-slide images. Furthermore, we create SlideInstruction, a largest comprehensive WSI instruction-following dataset to develop SlideChat, as well as SlideBench, a multi-modal benchmark designed to evaluate SlideChat across diverse scenarios. SlideChat demonstrates excellent chat abilities and achieves state-of-the-art performance on 18 tasks. We bridge the gap between MLLMs and pathology images at the whole-slide level with SlideChat, and we believe this represents an advancement in the field of general medical artificial intelligence (GMAI). + +Acknowledgement The authors acknowledge the funding provided by the National Key Research and Development Program of China (No.2022ZD0160102). + +# References + +[1] Marah Abdin, Sam Ade Jacobs, Ammar Ahmad Awan, Jyoti Aneja, Ahmed Awadallah, Hany Awadalla, Nguyen Bach, Amit Bahree, Arash Bakhtiari, Harkirat Behl, et al. Phi-3 technical report: A highly capable language model locally on your phone. arXiv preprint arXiv:2404.14219, 2024. 7 +[2] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 4, 5 +[3] Faruk Ahmed, Andrew Sellergren, Lin Yang, Shawn Xu, Boris Babenko, Abbi Ward, Niels Olson, Arash Mohtashamian, Yossi Matias, Greg S Corrado, et al. Pathalign: A vision-language model for whole slide images in histopathology. arXiv preprint arXiv:2406.19578, 2024. 2 +[4] AI@Meta. Llama 3 model card. 2024. 7 +[5] Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, et al. Qwen technical report. arXiv preprint arXiv:2309.16609, 2023. 5, 7 +[6] Zheng Cai, Maosong Cao, Haojiong Chen, Kai Chen, Keyu Chen, Xin Chen, Xun Chen, Zehui Chen, Zhi Chen, Pei Chu, et al. Internl m2 technical report. arXiv preprint arXiv:2403.17297, 2024.5, 7 +[7] Pingyi Chen, Honglin Li, Chenglu Zhu, Sunyi Zheng, and Lin Yang. Mi-gen: Multiple instance generation of pathology reports for gigapixel whole-slide images. arXiv preprint arXiv:2311.16480, 2023. 2, 5 +[8] Pingyi Chen, Chenglu Zhu, Sunyi Zheng, Honglin Li, and Lin Yang. Wsi-vqa: Interpreting whole slide images by generative visual question answering. arXiv preprint arXiv:2407.05603, 2024. 2, 7 +[9] Richard J Chen, Tong Ding, Ming Y Lu, Drew FK Williamson, Guillaume Jaume, Andrew H Song, Bowen Chen, Andrew Zhang, Daniel Shao, Muhammad Shaban, et al. Towards a general-purpose foundation model for computational pathology. Nature Medicine, 30(3):850-862, 2024. 1 +[10] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhanghao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E Gonzalez, et al. Vicuna: An open-source chatbot impressing gpt-4 with $90\%$ * chatgpt quality. See https://vicuna.lmsys.org (accessed 14 April 2023), 2(3):6, 2023. 2, 7 +[11] Jiayu Ding, Shuming Ma, Li Dong, Xingxing Zhang, Shaohan Huang, Wenhui Wang, Nanning Zheng, and Furu Wei. Longnet: Scaling transformers to 1,000,000,000 tokens. arXiv preprint arXiv:2307.02486, 2023. 3, 5 +[12] Zhengrui Guo, Jiabo Ma, Yingxue Xu, Yihui Wang, Liansheng Wang, and Hao Chen. Histgen: Histopathology report generation via local-global feature encoding and cross-modal context interaction. arXiv preprint arXiv:2403.05396, 2024. 2 +[13] Sunan He, Yuxiang Nie, Zhixuan Chen, Zhiyuan Cai, Hongmei Wang, Shu Yang, and Hao Chen. Meddr: + +Diagnosis-guided bootstrapping for large-scale medical vision-language learning. arXiv preprint arXiv:2404.15127, 2024. 2 +[14] Mahdi S Hosseini, Babak Ehteshami Bejnordi, Vincent Quoc-Huy Trinh, Lyndon Chan, Danial Hasan, Xingwen Li, Stephen Yang, Taehyo Kim, Haochen Zhang, Theodore Wu, et al. Computational pathology: a survey review and the way forward. Journal of Pathology Informatics, page 100357, 2024. 1 +[15] Carolyn Hutter and Jean Claude Zenklusen. The cancer genome atlas: creating lasting value beyond its data. Cell, 173(2):283-285, 2018. 2 +[16] Chunyuan Li, Cliff Wong, Sheng Zhang, Naoto Usuyama, Haotian Liu, Jianwei Yang, Tristan Naumann, Hoifung Poon, and Jianfeng Gao. Llava-med: Training a large language-and-vision assistant for biomedicine in one day. Advances in Neural Information Processing Systems, 36, 2024. 2, 5 +[17] Hao Li, Ying Chen, Yifei Chen, Rongshan Yu, Wenxian Yang, Liansheng Wang, Bowen Ding, and Yuchen Han. Generalizable whole slide image classification with fine-grained visual-semantic interaction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11398–11407, 2024. 2 +[18] Ming Y Lu, Tiffany Y Chen, Drew FK Williamson, Melissa Zhao, Maha Shady, Jana Lipkova, and Faisal Mahmood. Aibased pathology predicts origins for cancers of unknown primary. Nature, 594(7861):106-110, 2021. 2 +[19] Ming Y Lu, Bowen Chen, Drew FK Williamson, Richard J Chen, Ivy Liang, Tong Ding, Guillaume Jaume, Igor Odintsov, Long Phi Le, Georg Gerber, et al. A visual-language foundation model for computational pathology. Nature Medicine, 30(3):863-874, 2024. 3, 5 +[20] Ming Y Lu, Bowen Chen, Drew FK Williamson, Richard J Chen, Melissa Zhao, Aaron K Chow, Kenji Ikemura, Ahrong Kim, Dimitra Pouli, Ankush Patel, et al. A multimodal generative ai copilot for human pathology. Nature, pages 1-3, 2024. 1, 2, 4 +[21] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2 +[22] Mehmet Saygin Seyfioglu, Wisdom O Ikezogwo, Fatemeh Ghezloo, Ranjay Krishna, and Linda Shapiro. Quilt-llava: Visual instruction tuning by extracting localized narratives from open-source histopathology videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13183-13192, 2024. 1, 2, 5 +[23] Zhuchen Shao, Yang Chen, Hao Bian, Jian Zhang, Guojun Liu, and Yongbing Zhang. Hvtsurv: Hierarchical vision transformer for patient-level survival prediction from whole slide image. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2209-2217, 2023. 2 +[24] Andrew H Song, Guillaume Jaume, Drew FK Williamson, Ming Y Lu, Anurag Vaidya, Tiffany R Miller, and Faisal Mahmood. Artificial intelligence for digital and computa + +tional pathology. Nature Reviews Bioengineering, 1(12): 930-949, 2023. 1 +[25] Joey Spronck, Thijs Gelton, Leander van Eekelen, Joep Bogaerts, Leslie Tessier, Mart van Rijthoven, Lieke van der Woude, Michel van den Heuvel, Willemijn Theelen, Jeroen van der Laak, et al. nnunet meets pathology: bridging the gap for application to whole-slide images and computational biomarkers. In Medical Imaging with Deep Learning, 2023. 2 +[26] Yuxuan Sun, Chenglu Zhu, Sunyi Zheng, Kai Zhang, Lin Sun, Zhongyi Shui, Yunlong Zhang, Honglin Li, and Lin Yang. Pathasst: A generative foundation ai assistant towards artificial general intelligence of pathology. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 5034-5042, 2024. 1, 2 +[27] Eugene Vorontsov, Alican Bozkurt, Adam Casson, George Shaikovski, Michal Zelechowski, Kristen Severson, Eric Zimmermann, James Hall, Neil Tenenholtz, Nicolo Fusi, et al. A foundation model for clinical-grade computational pathology and rare cancers detection. Nature Medicine, pages 1-12, 2024. 1 +[28] Xiyue Wang, Junhan Zhao, Eliana Marostica, Wei Yuan, Ji etian Jin, Jiayu Zhang, Ruijiang Li, Hongping Tang, Kanran Wang, Yu Li, et al. A pathology foundation model for cancer diagnosis and prognosis prediction. Nature, pages 1-9, 2024. 2 +[29] Feng Xu, Chuang Zhu, Wenqi Tang, Ying Wang, Yu Zhang, Jie Li, Hongchuan Jiang, Zhongyue Shi, Jun Liu, and Mulan Jin. Predicting axillary lymph node metastasis in early breast cancer using deep learning on primary tumor biopsy slides. Frontiers in oncology, 11:759007, 2021. 5 +[30] Hanwen Xu, Naoto Usuyama, Jaspreet Bagga, Sheng Zhang, Rajesh Rao, Tristan Naumann, Cliff Wong, Zelalem Gero, Javier González, Yu Gu, et al. A whole-slide foundation model for digital pathology from real-world data. Nature, pages 1-8, 2024. 1, 3 +[31] Yingxue Xu, Yihui Wang, Fengtao Zhou, Jiabo Ma, Shu Yang, Huangjing Lin, Xin Wang, Jiguang Wang, Li Liang, Anjia Han, et al. A multimodal knowledge-enhanced whole-slide pathology foundation model. arXiv preprint arXiv:2407.15362, 2024. 2 +[32] An Yang, Baosong Yang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Zhou, Chengpeng Li, Chengyuan Li, Dayiheng Liu, Fei Huang, et al. Qwen2 technical report. arXiv preprint arXiv:2407.10671, 2024. 5, 7 +[33] Fengtao Zhou and Hao Chen. Cross-modal translation and alignment for survival analysis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 21485-21494, 2023. 2 \ No newline at end of file diff --git a/2025/SlideChat_ A Large Vision-Language Assistant for Whole-Slide Pathology Image Understanding/images.zip b/2025/SlideChat_ A Large Vision-Language Assistant for Whole-Slide Pathology Image Understanding/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..04bdb66202bb89f4e48e84ae469b4e627f94182d --- /dev/null +++ b/2025/SlideChat_ A Large Vision-Language Assistant for Whole-Slide Pathology Image Understanding/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2bd3fcef292029e11afaebaddc8c172275bfab5d236d59aaec3e8b8f7c0872d2 +size 383756 diff --git a/2025/SlideChat_ A Large Vision-Language Assistant for Whole-Slide Pathology Image Understanding/layout.json b/2025/SlideChat_ A Large Vision-Language Assistant for Whole-Slide Pathology Image Understanding/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..2a38edb9c7d5974a2934e87351fb03217757e4c2 --- /dev/null +++ b/2025/SlideChat_ A Large Vision-Language Assistant for Whole-Slide Pathology Image Understanding/layout.json @@ -0,0 +1,6464 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 63, + 110, + 547, + 147 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 110, + 547, + 147 + ], + "spans": [ + { + "bbox": [ + 63, + 110, + 547, + 147 + ], + "type": "text", + "content": "SlideChat: A Large Vision-Language Assistant for Whole-Slide Pathology Image Understanding" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 88, + 166, + 520, + 196 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 166, + 520, + 196 + ], + "spans": [ + { + "bbox": [ + 88, + 166, + 520, + 196 + ], + "type": "text", + "content": "Ying Chen" + }, + { + "bbox": [ + 88, + 166, + 520, + 196 + ], + "type": "inline_equation", + "content": "^{1,2*}" + }, + { + "bbox": [ + 88, + 166, + 520, + 196 + ], + "type": "text", + "content": " Guoan Wang" + }, + { + "bbox": [ + 88, + 166, + 520, + 196 + ], + "type": "inline_equation", + "content": "^{1,3*}" + }, + { + "bbox": [ + 88, + 166, + 520, + 196 + ], + "type": "text", + "content": " Yuanfeng Ji" + }, + { + "bbox": [ + 88, + 166, + 520, + 196 + ], + "type": "inline_equation", + "content": "^{4*†}" + }, + { + "bbox": [ + 88, + 166, + 520, + 196 + ], + "type": "text", + "content": " Yanjun Li" + }, + { + "bbox": [ + 88, + 166, + 520, + 196 + ], + "type": "inline_equation", + "content": "^{1,3}" + }, + { + "bbox": [ + 88, + 166, + 520, + 196 + ], + "type": "text", + "content": " Jin Ye" + }, + { + "bbox": [ + 88, + 166, + 520, + 196 + ], + "type": "inline_equation", + "content": "^{1,5}" + }, + { + "bbox": [ + 88, + 166, + 520, + 196 + ], + "type": "text", + "content": " Tianbin Li" + }, + { + "bbox": [ + 88, + 166, + 520, + 196 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 88, + 166, + 520, + 196 + ], + "type": "text", + "content": " Ming Hu" + }, + { + "bbox": [ + 88, + 166, + 520, + 196 + ], + "type": "inline_equation", + "content": "^{1,5}" + }, + { + "bbox": [ + 88, + 166, + 520, + 196 + ], + "type": "text", + "content": " Rongshan Yu" + }, + { + "bbox": [ + 88, + 166, + 520, + 196 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 88, + 166, + 520, + 196 + ], + "type": "text", + "content": " Yu Qiao" + }, + { + "bbox": [ + 88, + 166, + 520, + 196 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 88, + 166, + 520, + 196 + ], + "type": "text", + "content": " Junjun He" + }, + { + "bbox": [ + 88, + 166, + 520, + 196 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 110, + 207, + 499, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 207, + 499, + 236 + ], + "spans": [ + { + "bbox": [ + 110, + 207, + 499, + 236 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 110, + 207, + 499, + 236 + ], + "type": "text", + "content": "Shanghai AI Laboratory " + }, + { + "bbox": [ + 110, + 207, + 499, + 236 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 110, + 207, + 499, + 236 + ], + "type": "text", + "content": "Xiamen University " + }, + { + "bbox": [ + 110, + 207, + 499, + 236 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 110, + 207, + 499, + 236 + ], + "type": "text", + "content": "East China Normal University " + }, + { + "bbox": [ + 110, + 207, + 499, + 236 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 110, + 207, + 499, + 236 + ], + "type": "text", + "content": "Stanford University " + }, + { + "bbox": [ + 110, + 207, + 499, + 236 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 110, + 207, + 499, + 236 + ], + "type": "text", + "content": "Monash University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 156, + 262, + 202, + 274 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 156, + 262, + 202, + 274 + ], + "spans": [ + { + "bbox": [ + 156, + 262, + 202, + 274 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 61, + 286, + 296, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 286, + 296, + 579 + ], + "spans": [ + { + "bbox": [ + 61, + 286, + 296, + 579 + ], + "type": "text", + "content": "Despite the progress made by multimodal large language models (MLLMs) in computational pathology, they remain limited by a predominant focus on patch-level analysis, missing essential contextual information at the whole-slide level. The lack of large-scale instruction datasets and the gigapixel scale of whole slide images (WSIs) pose significant developmental challenges. In this paper, we present SlideChat, the first vision-language assistant capable of understanding gigapixel whole-slide images, exhibiting excellent multimodal conversational capability and response complex instruction across diverse pathology scenarios. To support its development, we created SlideInstruction, the largest instruction-following dataset for WSIs consisting of 4.2K WSI captions and 176K VQA pairs with multiple categories. Furthermore, we propose SlideBench, a multimodal benchmark that incorporates captioning and VQA tasks to assess SlideChat's capabilities in various settings such as microscopy, diagnosis and clinical. Compared to both general and specialized MLLMs, SlideChat exhibits exceptional capabilities, achieving state-of-the-art performance on 18 of 22 tasks. For example, it achieved an overall accuracy of " + }, + { + "bbox": [ + 61, + 286, + 296, + 579 + ], + "type": "inline_equation", + "content": "81.17\\%" + }, + { + "bbox": [ + 61, + 286, + 296, + 579 + ], + "type": "text", + "content": " on SlideBench-VQA (TCGA), and " + }, + { + "bbox": [ + 61, + 286, + 296, + 579 + ], + "type": "inline_equation", + "content": "54.15\\%" + }, + { + "bbox": [ + 61, + 286, + 296, + 579 + ], + "type": "text", + "content": " on SlideBench-VQA (BCNB). Our code, data, and model is publicly accessible at https://uni-medical.github.io/SlideChat.github.io." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 600, + 141, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 600, + 141, + 612 + ], + "spans": [ + { + "bbox": [ + 62, + 600, + 141, + 612 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 61, + 619, + 296, + 678 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 619, + 296, + 678 + ], + "spans": [ + { + "bbox": [ + 61, + 619, + 296, + 678 + ], + "type": "text", + "content": "Computational pathology aims to improve the analysis of digitized tissue samples [14, 24], such as whole slide images (WSIs), by applying artificial intelligence to aid in the diagnosis, identification, and understanding of disease. Recently, the development of this field has gained rapid" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 328, + 266, + 380, + 300 + ], + "blocks": [ + { + "bbox": [ + 316, + 265, + 324, + 273 + ], + "lines": [ + { + "bbox": [ + 316, + 265, + 324, + 273 + ], + "spans": [ + { + "bbox": [ + 316, + 265, + 324, + 273 + ], + "type": "text", + "content": "A" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 328, + 266, + 380, + 300 + ], + "lines": [ + { + "bbox": [ + 328, + 266, + 380, + 300 + ], + "spans": [ + { + "bbox": [ + 328, + 266, + 380, + 300 + ], + "type": "image", + "image_path": "334f9be18accb27a827000a8d6344fe38babd823e20549a014fd4917b3a36b4f.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 437, + 266, + 444, + 273 + ], + "lines": [ + { + "bbox": [ + 437, + 266, + 444, + 273 + ], + "spans": [ + { + "bbox": [ + 437, + 266, + 444, + 273 + ], + "type": "text", + "content": "B" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 388, + 547, + 441 + ], + "lines": [ + { + "bbox": [ + 313, + 388, + 547, + 441 + ], + "spans": [ + { + "bbox": [ + 313, + 388, + 547, + 441 + ], + "type": "text", + "content": "Figure 1. SlideChat is the first large vision-language assistant specifically designed for whole-slide pathology analysis. SlideChat can generate comprehensive descriptions of whole-slide images and provide contextually relevant responses across various applications." + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 328, + 306, + 412, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 306, + 412, + 319 + ], + "spans": [ + { + "bbox": [ + 328, + 306, + 412, + 319 + ], + "type": "text", + "content": "Summarize the key pathological findings from the whole slide image." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 334, + 325, + 431, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 334, + 325, + 431, + 338 + ], + "spans": [ + { + "bbox": [ + 334, + 325, + 431, + 338 + ], + "type": "text", + "content": "This pathological examination reveals the high-grade urothelial carcinoma with ..." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 328, + 346, + 416, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 346, + 416, + 365 + ], + "spans": [ + { + "bbox": [ + 328, + 346, + 416, + 365 + ], + "type": "text", + "content": "What is the pathological stage (pT) of the transitional cell carcinoma? A. pT1 B. pT2 C. pT3 D. pT4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 402, + 368, + 419, + 375 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 402, + 368, + 419, + 375 + ], + "spans": [ + { + "bbox": [ + 402, + 368, + 419, + 375 + ], + "type": "text", + "content": "B. pT2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 451, + 271, + 479, + 278 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 451, + 271, + 479, + 278 + ], + "spans": [ + { + "bbox": [ + 451, + 271, + 479, + 278 + ], + "type": "text", + "content": "Microscopy" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 446, + 280, + 541, + 300 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 446, + 280, + 541, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 446, + 280, + 541, + 285 + ], + "spans": [ + { + "bbox": [ + 446, + 280, + 541, + 285 + ], + "type": "inline_equation", + "content": "\\spadesuit" + }, + { + "bbox": [ + 446, + 280, + 541, + 285 + ], + "type": "text", + "content": " Tissue Architecture " + }, + { + "bbox": [ + 446, + 280, + 541, + 285 + ], + "type": "inline_equation", + "content": "\\spadesuit" + }, + { + "bbox": [ + 446, + 280, + 541, + 285 + ], + "type": "text", + "content": " Tumor Characteristics" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 446, + 286, + 503, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 446, + 286, + 503, + 293 + ], + "spans": [ + { + "bbox": [ + 446, + 286, + 503, + 293 + ], + "type": "text", + "content": "Histopathological Changes" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 446, + 294, + 517, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 446, + 294, + 517, + 300 + ], + "spans": [ + { + "bbox": [ + 446, + 294, + 517, + 300 + ], + "type": "inline_equation", + "content": "\\spadesuit" + }, + { + "bbox": [ + 446, + 294, + 517, + 300 + ], + "type": "text", + "content": " Cytomorphological Characteristics" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 451, + 310, + 476, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 451, + 310, + 476, + 317 + ], + "spans": [ + { + "bbox": [ + 451, + 310, + 476, + 317 + ], + "type": "text", + "content": "Diagnosis" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 446, + 319, + 541, + 338 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 446, + 319, + 541, + 325 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 446, + 319, + 541, + 325 + ], + "spans": [ + { + "bbox": [ + 446, + 319, + 541, + 325 + ], + "type": "inline_equation", + "content": "\\spadesuit" + }, + { + "bbox": [ + 446, + 319, + 541, + 325 + ], + "type": "text", + "content": " Disease Detection Staging Grading" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 446, + 327, + 495, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 446, + 327, + 495, + 331 + ], + "spans": [ + { + "bbox": [ + 446, + 327, + 495, + 331 + ], + "type": "inline_equation", + "content": "\\spadesuit" + }, + { + "bbox": [ + 446, + 327, + 495, + 331 + ], + "type": "text", + "content": " Disease Classification" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 446, + 332, + 493, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 446, + 332, + 493, + 338 + ], + "spans": [ + { + "bbox": [ + 446, + 332, + 493, + 338 + ], + "type": "inline_equation", + "content": "\\spadesuit" + }, + { + "bbox": [ + 446, + 332, + 493, + 338 + ], + "type": "text", + "content": " Differential Diagnosis" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 451, + 348, + 468, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 451, + 348, + 468, + 354 + ], + "spans": [ + { + "bbox": [ + 451, + 348, + 468, + 354 + ], + "type": "text", + "content": "Clinical" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 446, + 355, + 522, + 375 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 446, + 355, + 522, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 446, + 355, + 522, + 360 + ], + "spans": [ + { + "bbox": [ + 446, + 355, + 522, + 360 + ], + "type": "text", + "content": "Treatment Guidance Risk Factors" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 446, + 361, + 496, + 367 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 446, + 361, + 496, + 367 + ], + "spans": [ + { + "bbox": [ + 446, + 361, + 496, + 367 + ], + "type": "text", + "content": "Prognostic Assessment" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 446, + 369, + 488, + 375 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 446, + 369, + 488, + 375 + ], + "spans": [ + { + "bbox": [ + 446, + 369, + 488, + 375 + ], + "type": "inline_equation", + "content": "\\spadesuit" + }, + { + "bbox": [ + 446, + 369, + 488, + 375 + ], + "type": "text", + "content": " Biomarker Analysis" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 312, + 468, + 548, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 468, + 548, + 608 + ], + "spans": [ + { + "bbox": [ + 312, + 468, + 548, + 608 + ], + "type": "text", + "content": "momentum, mainly driven by breakthroughs in the visual foundation model [9, 27, 30]. These models learn generalized representations by pre-training on large-scale data and perform well in various downstream tasks, such as cancer detection and biomarker prediction. Building on this base, integration with the powerful Large Language Models (LLMs) further advances the development of the Multimodal Large Language Model (MLLMs) [20], which has made great strides in responding to more complex, open-ended visual queries, enabling it to serve as a versatile assistant at various stages of medical care, including clinical decision-making, education, and research." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 312, + 611, + 548, + 705 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 611, + 548, + 705 + ], + "spans": [ + { + "bbox": [ + 312, + 611, + 548, + 705 + ], + "type": "text", + "content": "Nevertheless, there are some major challenges that hinder the development and use of pathology MLLMs for real-world applications. First, it is challenging to develop a MLLM architecture that can effectively be capable of gigapixel whole slides (e.g., " + }, + { + "bbox": [ + 312, + 611, + 548, + 705 + ], + "type": "inline_equation", + "content": "100,000 \\times 100,000" + }, + { + "bbox": [ + 312, + 611, + 548, + 705 + ], + "type": "text", + "content": " pixels). Existing models [20, 22, 26] often process whole slides by extracting small patch/ROI-level data for subsequent analysis, resulting in a limited understanding of global slide con" + } + ] + } + ], + "index": 33 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 76, + 685, + 229, + 694 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 685, + 229, + 694 + ], + "spans": [ + { + "bbox": [ + 76, + 685, + 229, + 694 + ], + "type": "text", + "content": "*These authors contributed equally to this work." + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 76, + 695, + 289, + 704 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 695, + 289, + 704 + ], + "spans": [ + { + "bbox": [ + 76, + 695, + 289, + 704 + ], + "type": "text", + "content": "†Corresponding authors: yfj@stanford.edu, hejunjun@pjlab.org.cn." + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 295, + 744, + 315, + 754 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 744, + 315, + 754 + ], + "spans": [ + { + "bbox": [ + 295, + 744, + 315, + 754 + ], + "type": "text", + "content": "5134" + } + ] + } + ], + "index": 37 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 61, + 81, + 294, + 139 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 81, + 294, + 139 + ], + "spans": [ + { + "bbox": [ + 61, + 81, + 294, + 139 + ], + "type": "text", + "content": "text and suboptimal performance in some complex pathological analysis. Second, publicly available multi-modal pathology slide datasets are relatively scarce and of varying quality [7, 8, 12], which limits the development of MLLMs trained on such data." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 143, + 295, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 143, + 295, + 422 + ], + "spans": [ + { + "bbox": [ + 61, + 143, + 295, + 422 + ], + "type": "text", + "content": "In this paper, we present SlideChat (see Fig. 1), the first open-source vision-language assistant capable of understanding gigapixel whole-slide images. First, SlideChat is trained on SlideInstruction, a large-scale multi-modal instruction dataset encompassing data from The Cancer Genome Atlas (TCGA) [15] via our specifically designed data processing pipeline (see Fig. 3 (A)). SlideInstruction contains 4,181 WSI-caption pairs and 175,753 visual question-answer pairs from 3,294 patients, covering 10 cancer types. The question-answer pairs include both open-ended and closed-ended questions, further divided into 13 subcategories, covering a diverse range of clinical tasks. SlideInstruction is more than 20 times larger than previous public instruction datasets in the number of instructions. Second, we propose SlideChat, a novel architecture in LLaVA style for capable multi-modal analysis of gigapixel whole slides. A gigapixel whole slide is first divided into a series of patches, each of which is individually processed by a patch-level encoder to extract local features. The resulting long sequence of patch tokens is then processed by a slide-level encoder employing sparse attention to enhance the slide-level features. Finally, enhanced patch tokens are fed into the LLM via a projector, which processes user queries and generates responses." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 426, + 295, + 705 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 426, + 295, + 705 + ], + "spans": [ + { + "bbox": [ + 61, + 426, + 295, + 705 + ], + "type": "text", + "content": "To systematically evaluate the performance of SlideChat in real-world scenarios, we establish a comprehensive digital pathology benchmark named SlideBench, encompassing more than 20 clinical tasks, using data from both TCGA and the in-the-wild Early Breast Cancer CoreNeedle Biopsy (BCNB) dataset. This resulted in three test sets: SlideBench-Caption, comprising 734 WSI-caption pairs; SlideBench-VQA (TCGA), comprising 7,827 VQA pairs covering 13 different tasks; and SlideBench-VQA (BCNB), including a total of 7,274 VQA entries from 1,058 patients, covering seven different tasks. Additionally, we compare the performance of SlideChat on another externally proposed dataset, WSI-VQA [8], to further validate its effectiveness. We compare our model with the currently available state-of-the-art general and medical-specialized MLLMs including GPT-4o, LLaV-aMed [16], MedDr [13] and Quilt-LLaVA [22]. Benefiting from large-scale, high-quality training and effective local-global context modeling, SlideChat achieves state-of-the-art performance on 18 out of 22 tasks, with significant improvements over the second-best method " + }, + { + "bbox": [ + 61, + 426, + 295, + 705 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 61, + 426, + 295, + 705 + ], + "type": "text", + "content": " on 9 tasks on four benchmarks. Specifically, SlideChat achieves an average accuracy improvement of " + }, + { + "bbox": [ + 61, + 426, + 295, + 705 + ], + "type": "inline_equation", + "content": "13.47\\%" + }, + { + "bbox": [ + 61, + 426, + 295, + 705 + ], + "type": "text", + "content": " over the second-best model on SlideBench-VQA (TCGA), an average improvement of " + }, + { + "bbox": [ + 61, + 426, + 295, + 705 + ], + "type": "inline_equation", + "content": "12.59\\%" + }, + { + "bbox": [ + 61, + 426, + 295, + 705 + ], + "type": "text", + "content": " on" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 313, + 81, + 547, + 150 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 81, + 547, + 150 + ], + "spans": [ + { + "bbox": [ + 313, + 81, + 547, + 150 + ], + "type": "text", + "content": "SlideBench-VQA (BCNB), and an improvement of " + }, + { + "bbox": [ + 313, + 81, + 547, + 150 + ], + "type": "inline_equation", + "content": "5.82\\%" + }, + { + "bbox": [ + 313, + 81, + 547, + 150 + ], + "type": "text", + "content": " on WSI-VQA. Finally, to accelerate research progress in digital pathology, we make SlideChat fully open-weight, including source code and model weights as well as instruction and benchmark datasets. The key contributions are summarized four-fold in the following:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 325, + 151, + 547, + 313 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 325, + 151, + 547, + 196 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 151, + 547, + 196 + ], + "spans": [ + { + "bbox": [ + 325, + 151, + 547, + 196 + ], + "type": "text", + "content": "- We develop SlideChat, the first vision-language assistant capable of understanding gigapixel whole-slide images, achieving state-of-the-art performance on multiple benchmarks." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 325, + 198, + 547, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 198, + 547, + 232 + ], + "spans": [ + { + "bbox": [ + 325, + 198, + 547, + 232 + ], + "type": "text", + "content": "- We create SlideInstruction, a largest comprehensive WSI instruction-following dataset containing 4.2K WSI-caption pairs and 176K VQA pairs." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 325, + 232, + 547, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 232, + 547, + 277 + ], + "spans": [ + { + "bbox": [ + 325, + 232, + 547, + 277 + ], + "type": "text", + "content": "- We establish SlideBench, a WSI multi-modal benchmark comprising SlideBench-Caption, SlideBench-VQA (TCGA), and SlideBench-VQA (BCNB), covering 21 different clinical tasks." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 325, + 279, + 547, + 313 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 279, + 547, + 313 + ], + "spans": [ + { + "bbox": [ + 325, + 279, + 547, + 313 + ], + "type": "text", + "content": "- We release SlideChat, SlideInstruction and SlideBench as open-source resources to facilitate research and development in computational pathology." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 323, + 402, + 335 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 323, + 402, + 335 + ], + "spans": [ + { + "bbox": [ + 313, + 323, + 402, + 335 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 342, + 547, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 342, + 547, + 482 + ], + "spans": [ + { + "bbox": [ + 313, + 342, + 547, + 482 + ], + "type": "text", + "content": "Whole Slide Image Analysis Whole slide images are pivotal in modern pathology, enabling comprehensive analysis of tissue samples for tasks such as predicting patient prognosis, classifying cancer subtypes, and identifying biomarkers [17, 18, 23-25, 33]. Recent studies have leveraged pathology foundational models [3, 28, 31] to enhance WSIs analysis, either through fine-tuning for specific downstream tasks or by employing zero-shot prediction approaches in CLIP [21] style. Although these models are effective in task-specific applications, their reliance on fine-tuning or limited zero-shot capabilities restricts their generalizability across diverse and complex user instructions." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 495, + 547, + 705 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 495, + 547, + 705 + ], + "spans": [ + { + "bbox": [ + 313, + 495, + 547, + 705 + ], + "type": "text", + "content": "MLLMs in Computational Pathology The paradigm of MLLMs enables them to effectively respond to more complex, open-ended visual queries while processing pathology images, thus providing significant value across various medical stages. PathChat [20] is a vision-language assistant designed for pathology, developed with 450K private instruction pairs to handle both visual and natural language queries. Quilt-Instruct [22] is a large-scale dataset comprising 107K question-answer pairs. Building on Quilt-Instruct, Quilt-LLAVA [22] is a model designed for diagnostic reasoning across multiple image patches, leveraging its extensive question-answer pairs to accurately interpret complex H&E images. PathAsst [26] combines a pathology-specific CLIP model with Vicuna-13b [10] to create a multimodal generative foundational model tailored for pathology. However, current MLLMs primarily only focus on patch or ROI data, limiting their utility for slide-level clinical applications where broader contextual understanding is crucial." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 744, + 315, + 754 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 744, + 315, + 754 + ], + "spans": [ + { + "bbox": [ + 295, + 744, + 315, + 754 + ], + "type": "text", + "content": "5135" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 66, + 82, + 541, + 242 + ], + "blocks": [ + { + "bbox": [ + 66, + 82, + 541, + 242 + ], + "lines": [ + { + "bbox": [ + 66, + 82, + 541, + 242 + ], + "spans": [ + { + "bbox": [ + 66, + 82, + 541, + 242 + ], + "type": "image", + "image_path": "469b18531b509cecfc8ab5beb7f7537e3b21621b056d4c05c0aff9487f21fdaa.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 61, + 249, + 548, + 293 + ], + "lines": [ + { + "bbox": [ + 61, + 249, + 548, + 293 + ], + "spans": [ + { + "bbox": [ + 61, + 249, + 548, + 293 + ], + "type": "text", + "content": "Figure 2. Overview of our SlideChat. (A) SlideChat serializes each input WSI into a sequence of " + }, + { + "bbox": [ + 61, + 249, + 548, + 293 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 61, + 249, + 548, + 293 + ], + "type": "text", + "content": " patches, converting each into visual embeddings with a patch-level encoder. A slide-level encoder then interacts with these features to generate contextual embeddings. Then, a multimodal projector maps the visual features from the slide-level encoder into a unified space, aligned seamlessly with the LLM. (B) SlideChat was trained for two stages: Cross-Domain Alignment and Visual Instruction Learning." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 62, + 312, + 127, + 323 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 312, + 127, + 323 + ], + "spans": [ + { + "bbox": [ + 62, + 312, + 127, + 323 + ], + "type": "text", + "content": "3. SlideChat" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 335, + 142, + 346 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 335, + 142, + 346 + ], + "spans": [ + { + "bbox": [ + 62, + 335, + 142, + 346 + ], + "type": "text", + "content": "3.1. Architecture" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 355, + 297, + 705 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 355, + 297, + 705 + ], + "spans": [ + { + "bbox": [ + 61, + 355, + 297, + 705 + ], + "type": "text", + "content": "To achieve the goal of analyzing gigapixel whole-slide images in a multimodal setting, as shown in Fig. 2, SlideChat consists of four key designs: the patch-level encoder, the slide-level encoder, the multimodal projector module, and the large language model. Our method starts by partitioning the WSI into smaller " + }, + { + "bbox": [ + 61, + 355, + 297, + 705 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 61, + 355, + 297, + 705 + ], + "type": "text", + "content": " pixel patches, making it computationally feasible to process such large images. These patches are then passed through a well-trained, frozen patch-level encoder [19], which extracts localized features from each individual patch, capturing fine-grained details such as cellular structures. Building on this, we employ LongNet [11, 30] as slide-level encoder to enhance the patch-level embeddings and capture global patterns across the entire slide. This encoder uses sparse attention mechanisms to learn both local and global contextual information, enabling the model to perceive intricate local features while capturing the broader context, which is critical for comprehensive pathological assessments. Following the slide-level encoding, SlideChat incorporates a multimodal projector that maps these patch tokens into a unified space aligned with the LLM. This ensures that the visual features extracted from the WSIs are effectively transformed into representations compatible with the language model, facilitating seamless integration and interaction between visual and textual data. Concurrently, the model accepts natural language instructions from users, such as \"What is the type of tumor in the image?\". These textual queries are processed by the LLM, which comprehends the textual input and integrates it with the visual features extracted from the WSIs, enabling accurate and contextually relevant diagnoses" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 312, + 548, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 312, + 548, + 372 + ], + "spans": [ + { + "bbox": [ + 313, + 312, + 548, + 372 + ], + "type": "text", + "content": "tic responses. This multimodal reasoning capability allows SlideChat to provide accurate and contextually relevant answers to complex pathology-related questions, thereby supporting clinical decision-making, education, and research across various medical stages." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 383, + 358, + 394 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 383, + 358, + 394 + ], + "spans": [ + { + "bbox": [ + 313, + 383, + 358, + 394 + ], + "type": "text", + "content": "3.2.Data" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 312, + 402, + 548, + 706 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 402, + 548, + 706 + ], + "spans": [ + { + "bbox": [ + 312, + 402, + 548, + 706 + ], + "type": "text", + "content": "SlideInstruction There is a notable lack of large-scale multimodal pathology datasets supporting the training of vision-language assistants for whole-slide image understanding. To support the training of SlideChat, we develop SlideInstruction, a comprehensive instruction dataset, sourced from the TCGA database, comprising 4,915 WSI-report pairs from 4,028 patients. Fig. 3 illustrates our entire data curation pipeline. We initially prompt GPT-4 to refine the pathology reports, and clean up the noise in the report including unrelated symbols, technical details of pathology department procedures, specimen handling and processing information, redundant administrative or legal statements, and some repeated information. For the refined pathology reports, we further employ GPT-4 to generate high-quality multimodal data, comprising two main components: (1) WSI-Caption Data: We craft concise, clinically relevant summaries for each whole slide image by prompting the language model to extract key pathological findings. These summaries were structured into coherent paragraphs that highlighted crucial clinical details such as diagnostic results, tumor characteristics, margin status, and lymph node involvement, ensuring the caption dataset is both focused and informative. (2) WSI Instruction-Following Data: To enhance the model's ability to follow instructions and improve its comprehension of pathology images, we leveraged GPT-4 to generate tailored question" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 744, + 315, + 754 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 744, + 315, + 754 + ], + "spans": [ + { + "bbox": [ + 295, + 744, + 315, + 754 + ], + "type": "text", + "content": "5136" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 96, + 91, + 530, + 199 + ], + "blocks": [ + { + "bbox": [ + 78, + 83, + 88, + 92 + ], + "lines": [ + { + "bbox": [ + 78, + 83, + 88, + 92 + ], + "spans": [ + { + "bbox": [ + 78, + 83, + 88, + 92 + ], + "type": "text", + "content": "A" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 96, + 91, + 530, + 199 + ], + "lines": [ + { + "bbox": [ + 96, + 91, + 530, + 199 + ], + "spans": [ + { + "bbox": [ + 96, + 91, + 530, + 199 + ], + "type": "image", + "image_path": "f0daefbfd8e147bffc5e8a200b3f1c7ba6517ebcc8d2cab7bf276c51d24f595b.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 93, + 215, + 504, + 266 + ], + "blocks": [ + { + "bbox": [ + 78, + 205, + 87, + 214 + ], + "lines": [ + { + "bbox": [ + 78, + 205, + 87, + 214 + ], + "spans": [ + { + "bbox": [ + 78, + 205, + 87, + 214 + ], + "type": "text", + "content": "B" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 93, + 215, + 504, + 266 + ], + "lines": [ + { + "bbox": [ + 93, + 215, + 504, + 266 + ], + "spans": [ + { + "bbox": [ + 93, + 215, + 504, + 266 + ], + "type": "image", + "image_path": "87b8387cb638bbb9b4ca94bda3ad892e90ad4a42a6350df49a6db0e76e996a37.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 99, + 291, + 170, + 336 + ], + "blocks": [ + { + "bbox": [ + 78, + 274, + 87, + 283 + ], + "lines": [ + { + "bbox": [ + 78, + 274, + 87, + 283 + ], + "spans": [ + { + "bbox": [ + 78, + 274, + 87, + 283 + ], + "type": "text", + "content": "C" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 99, + 291, + 170, + 336 + ], + "lines": [ + { + "bbox": [ + 99, + 291, + 170, + 336 + ], + "spans": [ + { + "bbox": [ + 99, + 291, + 170, + 336 + ], + "type": "image", + "image_path": "f97990cab839f83140c393912fd67655dc93a45953901eca9e8bd6e2717d5d57.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 100, + 346, + 169, + 391 + ], + "blocks": [ + { + "bbox": [ + 100, + 346, + 169, + 391 + ], + "lines": [ + { + "bbox": [ + 100, + 346, + 169, + 391 + ], + "spans": [ + { + "bbox": [ + 100, + 346, + 169, + 391 + ], + "type": "image", + "image_path": "b127e927e5f640d0d6d859e17226266b5adfd02d8cce2a44dcd32e4c6d80f4f7.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 61, + 410, + 547, + 453 + ], + "lines": [ + { + "bbox": [ + 61, + 410, + 547, + 453 + ], + "spans": [ + { + "bbox": [ + 61, + 410, + 547, + 453 + ], + "type": "text", + "content": "Figure 3. (A) Overview of the SlideInstruction generation pipeline. We prompt GPT-4 to extract the WSI-Caption, Open-set VQA and Closed-set VQA from reports. (B) For the generated Closed-set VQA, we employ LLMs to filter low-quality QA pairs and involve pathologists for validation, resulting in the creation of SlideBench-VQA. (C) Examples of WSI caption and instruction-following scenarios in microscopy, diagnostics, and clinical applications. For additional examples, please refer to Supplementary Material." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 182, + 297, + 295, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 182, + 297, + 295, + 332 + ], + "spans": [ + { + "bbox": [ + 182, + 297, + 295, + 332 + ], + "type": "text", + "content": "[Caption] #Q:Describe the whole slide image. #A:The bladder exhibits an invasive poorly differentiated urothelial carcinoma, grade 4/4, extending through the bladder neck wall...." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 182, + 351, + 294, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 182, + 351, + 294, + 387 + ], + "spans": [ + { + "bbox": [ + 182, + 351, + 294, + 387 + ], + "type": "text", + "content": "[VQA-Diagnosis] \n#Q: From the given whole slide image, which histological subtype of lung adenocarcinoma is most predominantly observed? \n#A: It is acinar subtype observed." + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 325, + 293, + 392, + 337 + ], + "blocks": [ + { + "bbox": [ + 325, + 293, + 392, + 337 + ], + "lines": [ + { + "bbox": [ + 325, + 293, + 392, + 337 + ], + "spans": [ + { + "bbox": [ + 325, + 293, + 392, + 337 + ], + "type": "image", + "image_path": "748388969461b7e9b57d20c3a086a33fcff6685ce6b440d446124d5b332d905b.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 323, + 343, + 389, + 393 + ], + "blocks": [ + { + "bbox": [ + 323, + 343, + 389, + 393 + ], + "lines": [ + { + "bbox": [ + 323, + 343, + 389, + 393 + ], + "spans": [ + { + "bbox": [ + 323, + 343, + 389, + 393 + ], + "type": "image", + "image_path": "cd2511ed303e1a629613496c681fe0adb36bfa337b7ac2f4e9595832e708f76f.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 405, + 298, + 518, + 333 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 405, + 298, + 518, + 333 + ], + "spans": [ + { + "bbox": [ + 405, + 298, + 518, + 333 + ], + "type": "text", + "content": "[VQA-Microscopy] \n#Q: What histological feature observable in a whole slide image differentiates anaplastic astrocytoma from glioblastoma? \n#A: Absence of endothelial proliferation." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 405, + 352, + 518, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 405, + 352, + 518, + 386 + ], + "spans": [ + { + "bbox": [ + 405, + 352, + 518, + 386 + ], + "type": "text", + "content": "[VQA-Clinical] #Q: Based on the histopathological features, what is the primary recommended treatment? #A: Simple mastectomy followed by hormone therapy." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 61, + 474, + 296, + 695 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 474, + 296, + 695 + ], + "spans": [ + { + "bbox": [ + 61, + 474, + 296, + 695 + ], + "type": "text", + "content": "and-answer pairs for each WSI report. Drawing inspiration from PathChat [20], we structure these questions into 3 broad categories—microscopy, diagnosis, and clinical considerations—which represent key stages in the pathology workflow, and 13 narrow categories focusing on specific aspects within each stage (Fig. 1 B). To construct a comprehensive instructional dataset, we began by generating two open-ended and two closed-ended QA pairs for each report within each specific category, taking advantage of the rich content found in the majority of TCGA reports. Then, some QA pairs were excluded, such as those lacking corresponding category information. Regarding the train/test split, it is worth noting that the WSI-report datasets from TCGA include 2 types: (a) one report linked to multiple WSIs, and (b) one report linked to a single WSI. For type (a), where specific diagnostic details may not align perfectly with each WSI, we include all WSIs in the training set to introduce some \"noisy data\", which can enhance model robustness. For type (b), " + }, + { + "bbox": [ + 61, + 474, + 296, + 695 + ], + "type": "inline_equation", + "content": "80\\%" + }, + { + "bbox": [ + 61, + 474, + 296, + 695 + ], + "type": "text", + "content": " of WSIs are allocated to the training" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 474, + 547, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 474, + 547, + 533 + ], + "spans": [ + { + "bbox": [ + 313, + 474, + 547, + 533 + ], + "type": "text", + "content": "set and " + }, + { + "bbox": [ + 313, + 474, + 547, + 533 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 313, + 474, + 547, + 533 + ], + "type": "text", + "content": " to the test set. Finally, there are 4,181 WSIs for training and 734 WSIs for testing. Consequently, we construct a large-scale training set named SlideInstruction, comprising 4,181 WSI captions and 175,753 instruction-following VQA pairs across 13 narrow categories." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 312, + 553, + 548, + 705 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 553, + 548, + 705 + ], + "spans": [ + { + "bbox": [ + 312, + 553, + 548, + 705 + ], + "type": "text", + "content": "SlideBench To systematically evaluate the performance of SlideChat, we include the remaining 734 WSI captions from the test set, along with a substantial number of closed-set VQA pairs, to establish a comprehensive evaluation benchmark. First, we construct a set named SlideBench-Caption based on the WSI-Caption data to evaluate the model's ability to generate accurate and coherent descriptions of whole slide images. Secondly, we construct SlideBench-VQA (TCGA) based on closed-set VQA pairs along with test WSIs, aiming to evaluate various aspects of model performance. As shown in Fig. 3 (B), to improve the quality of the testing benchmarks, we employ four advanced large language models, including GPT-4 [2], InternLM2-" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 744, + 315, + 754 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 744, + 315, + 754 + ], + "spans": [ + { + "bbox": [ + 295, + 744, + 315, + 754 + ], + "type": "text", + "content": "5137" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 61, + 81, + 297, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 81, + 297, + 419 + ], + "spans": [ + { + "bbox": [ + 61, + 81, + 297, + 419 + ], + "type": "text", + "content": "Chat-7B [6], Qwen-7B-Chat [5], and DeepSeek-7B-Chat, to filter closed-set VQAs by predicting answers based solely on the question text. Any questions for which at least three of these models provided correct answers are subsequently excluded. Following this automated filtering, five expert pathologists are invited to review and amend the remaining questions. The review process is guided by the following criteria: (1) Whether the correct answer necessitates image interpretation; (2) Whether the question and its corresponding answer are logically and coherently structured; and (3) Whether the question aligns appropriately with the designated broad and narrow categories. VQA pairs failing to meet these criteria are excluded by the pathologists. Consequently, the SlideBench-VQA (TCGA) comprises 7,827 VQAs across 13 categories, with some examples illustrated in Fig. 3 C. Additionally, we incorporate the in-the-wild Early Breast Cancer Core-Needle Biopsy (BCNB) WSI dataset [29], which encompasses a diverse patient population and a variety of clinical task labels, to enhance the test set benchmark and more comprehensively assess the model's generalization capabilities. In detail, we convert the BCNB dataset into a VQA format by rephrasing the classification objectives into a specific template as questions, while transforming the original multi-class labels into selectable options, and integrating it into SlideBench as an external subset, named SlideBench-VQA (BCNB). This dataset comprises 7,247 VQA pairs from 1,058 patients, specifically designed to evaluate SlideChat's zero-shot generalization capability across 7 distinct classification tasks." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 62, + 426, + 173, + 440 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 426, + 173, + 440 + ], + "spans": [ + { + "bbox": [ + 62, + 426, + 173, + 440 + ], + "type": "text", + "content": "3.3. Two-stage Training" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 444, + 296, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 444, + 296, + 583 + ], + "spans": [ + { + "bbox": [ + 61, + 444, + 296, + 583 + ], + "type": "text", + "content": "Stage 1: Cross-Domain Alignment. SlideChat adopts a two-stage training approach (see Fig. 2 B). In the first stage, the primary objective is to align the LLM's word embeddings with the visual features extracted from whole slide images. This alignment enables the LLM to interpret visual representations from the slide-level encoder, facilitating the effective utilization of the intricate features within the slides. During this stage, SlideChat is trained to generate descriptive captions using 4.2K WSI-caption pairs from SlideInstruction. Specifically, only the slide-level encoder and projection matrix are updated, while the patch-level encoder and LLM weights remain fixed." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 599, + 295, + 705 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 599, + 295, + 705 + ], + "spans": [ + { + "bbox": [ + 61, + 599, + 295, + 705 + ], + "type": "text", + "content": "Stage 2: Visual Instruction Learning. In the second stage, we focus on visual question-answering tasks to train the model to accurately respond to domain-specific questions concerning whole slide images. During this phase, the model develops the ability to handle a broad range of multimodal instructions, enabling it to generate answers by effectively integrating both visual and textual information. For example, the model must perform various pathology tasks, such as describing the extent of tumor invasion or assess" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 312, + 81, + 548, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 81, + 548, + 175 + ], + "spans": [ + { + "bbox": [ + 312, + 81, + 548, + 175 + ], + "type": "text", + "content": "ing the degree of cellular differentiation. To accomplish this, we utilize 176K WSI VQAs from SlideInstruction in the second training stage, allowing the slide encoder, projection layer, and large language model components to be fully trainable to ensure comprehensive adaptability. This approach significantly enhances the model's capability to handle diverse pathology-related tasks, thereby increasing its effectiveness in real-world clinical and research settings." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 185, + 388, + 198 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 185, + 388, + 198 + ], + "spans": [ + { + "bbox": [ + 313, + 185, + 388, + 198 + ], + "type": "text", + "content": "4. Experiment" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 205, + 548, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 205, + 548, + 529 + ], + "spans": [ + { + "bbox": [ + 313, + 205, + 548, + 529 + ], + "type": "text", + "content": "We conducted the following experiments to evaluate three key aspects of SlideChat: (1) its whole slide image captioning capability, which assesses proficiency in generating descriptive captions that accurately summarize the critical pathological features of a WSI; (2) its visual question-answering (VQA) ability across various complex pathological scenarios and its generalizability in zero-shot settings; and (3) SlideChat's ability to process gigapixel WSIs, capturing both essential global context and intricate details, thereby enhancing its performance compared to patch-level MLLMs. For WSI captioning baselines, we benchmark against MI-Gen [7], a state-of-the-art method specifically designed for this task. Given that existing MLLMs cannot handle the gigapixel scale of whole slide images, we establish baseline comparisons using two approaches: (1) randomly selecting 30 patches from each WSI and inputting them into MLLMs (e.g., GPT-4o [2], LLaVA-Med [16], MedDr [16], Quilt-LLaVA [22]), followed by a majority voting scheme to generate slide-level predictions; and (2) directly inputting a WSI thumbnail, resized to " + }, + { + "bbox": [ + 313, + 205, + 548, + 529 + ], + "type": "inline_equation", + "content": "1024 \\times 1024" + }, + { + "bbox": [ + 313, + 205, + 548, + 529 + ], + "type": "text", + "content": " pixels, referred to as Slide (T), into the MLLMs. For VQA tasks, we further evaluate performance by comparing against random prediction baselines and text-only models, thereby assessing the incremental contribution of visual information. Unless otherwise specified, SlideChat is configured with the patch-level encoder CONCH [19], the slide-level encoder LongNet [11], and utilizes the Qwen2.5-7B-Instruct [32] as LLM." + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 316, + 570, + 547, + 643 + ], + "blocks": [ + { + "bbox": [ + 313, + 539, + 547, + 561 + ], + "lines": [ + { + "bbox": [ + 313, + 539, + 547, + 561 + ], + "spans": [ + { + "bbox": [ + 313, + 539, + 547, + 561 + ], + "type": "text", + "content": "Table 1. Captioning performance across different methods on SlideBench-Caption." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 316, + 570, + 547, + 643 + ], + "lines": [ + { + "bbox": [ + 316, + 570, + 547, + 643 + ], + "spans": [ + { + "bbox": [ + 316, + 570, + 547, + 643 + ], + "type": "table", + "html": "
MethodsInputBLEU-1BLEU-2BLEU-3BLEU-4Rouge-LGPT score
GPT-4oPatch0.160.030.010.010.131.54
GPT-4oSlide (T)0.100.030.010.010.111.03
Quilt-LLaVAPatch0.300.190.110.050.183.87
Quilt-LLaVASlide (T)0.230.090.040.010.161.89
MI-GenSlide0.370.240.150.100.254.14
SlideChatSlide0.370.210.120.080.244.11
", + "image_path": "74276fbbf5e078c5054a9c788306991a2476e62abf699311d3bfeb5768842dea.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 669, + 547, + 704 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 669, + 547, + 704 + ], + "spans": [ + { + "bbox": [ + 313, + 669, + 547, + 704 + ], + "type": "text", + "content": "SlideBench-Caption We report BLEU, ROUGE, and GPT scores to evaluate caption generation performance in Tab. 1. For the GPT score, we use GPT-4 to assess the" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 744, + 315, + 754 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 744, + 315, + 754 + ], + "spans": [ + { + "bbox": [ + 295, + 744, + 315, + 754 + ], + "type": "text", + "content": "5138" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 80, + 99, + 531, + 276 + ], + "blocks": [ + { + "bbox": [ + 214, + 79, + 394, + 90 + ], + "lines": [ + { + "bbox": [ + 214, + 79, + 394, + 90 + ], + "spans": [ + { + "bbox": [ + 214, + 79, + 394, + 90 + ], + "type": "text", + "content": "Table 2. VQA performance with different methods." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 80, + 99, + 531, + 276 + ], + "lines": [ + { + "bbox": [ + 80, + 99, + 531, + 276 + ], + "spans": [ + { + "bbox": [ + 80, + 99, + 531, + 276 + ], + "type": "table", + "html": "
MethodsInputSlideBench-VQA (TCGA)SlideBench-VQA (BCNB)WSI-VQA*
MicroscopyDiagnosisClinicalOverall
RandomText24.4424.9126.4425.0224.4024.14
GPT-438.2829.0945.0037.25018.60
GPT-4oPatch62.8946.6966.7757.9141.4330.41
MedDr73.3057.7874.2567.7033.6754.36
LLaVA-Med47.3432.7847.9642.0030.126.31
Quilt-LLaVA57.7635.9653.0748.0732.1944.43
GPT-4oSlide (T)38.2823.1043.4234.07014.03
MedDr70.4852.4772.8064.2535.4850.95
LLaVA-Med45.8227.5840.8437.39018.79
Quilt-LLaVA49.1226.9744.7539.3941.5535.40
SlideChatSlide87.64(+14.34)73.27(+15.49)84.26(+10.01)81.17(+13.47)54.14(+12.59)60.18(+5.82)
", + "image_path": "27ce1baf14cf385bf8d2447b2a552d59866571937b209cbf1466ec50ddb682e6.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 296, + 295, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 296, + 295, + 574 + ], + "spans": [ + { + "bbox": [ + 61, + 296, + 295, + 574 + ], + "type": "text", + "content": "similarity between the generated captions and the ground truth, providing an overall score on a scale of 1 to 10, with higher scores indicating better performance. When utilizing patch-level inputs, GPT-4o generates individual descriptions for each patch, which are subsequently integrated to create the final slide-level caption. However, this approach yields poor performance, as evidenced by a BLEU-1 score of 0.16 and a GPT-score of 1.54. These results suggest that the patch-based method fails to adequately capture the broader context necessary for accurate WSI captioning. When the WSI thumbnail of size " + }, + { + "bbox": [ + 61, + 296, + 295, + 574 + ], + "type": "inline_equation", + "content": "1024 \\times 1024" + }, + { + "bbox": [ + 61, + 296, + 295, + 574 + ], + "type": "text", + "content": " pixels is used as input to GPT-4o, performance decreases further, with a BLEU-1 score of 0.10 and a GPT-score of 1.03. This suggests that while the thumbnail offers a global view of the slide, it may lack the resolution and detail necessary for generating precise and informative captions. In contrast, MI-Gen, a model specifically designed for WSI captioning, demonstrates significantly superior performance across all metrics, achieving a BLEU-1 score of 0.37 and a GPT score of 4.14. Similarly, SlideChat, shows competitive results with a BLEU-1 score of 0.37 and a GPT score of 4.11. These outcomes highlight SlideChat's ability to effectively integrate both local and global information from the slides and confirm its efficiency in describing whole-slide images." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 588, + 295, + 705 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 588, + 295, + 705 + ], + "spans": [ + { + "bbox": [ + 61, + 588, + 295, + 705 + ], + "type": "text", + "content": "SlideBench-VQA (TCGA) We further evaluate SlideChat's overall performance on the multiple-choice VQA benchmark. The results on SlideBench-VQA (TCGA) in Tab. 2, compare different methods across three categories: microscopy, diagnosis, and clinical, along with an overall performance score. Random selection achieves an overall score of " + }, + { + "bbox": [ + 61, + 588, + 295, + 705 + ], + "type": "inline_equation", + "content": "25.02\\%" + }, + { + "bbox": [ + 61, + 588, + 295, + 705 + ], + "type": "text", + "content": " accuracy, serving as a baseline for answer distribution but demonstrating poor performance. While GPT-4, relying solely on text input, outperforms random predictions, it continues to struggle" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 312, + 296, + 547, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 296, + 547, + 598 + ], + "spans": [ + { + "bbox": [ + 312, + 296, + 547, + 598 + ], + "type": "text", + "content": "with accurately answering questions. When GPT-4o incorporates patch-level inputs, its performance improves markedly, reaching a score of " + }, + { + "bbox": [ + 312, + 296, + 547, + 598 + ], + "type": "inline_equation", + "content": "57.91\\%" + }, + { + "bbox": [ + 312, + 296, + 547, + 598 + ], + "type": "text", + "content": " and underscoring the crucial role of detailed visual data. However, using a WSI thumbnail results in a lower score of " + }, + { + "bbox": [ + 312, + 296, + 547, + 598 + ], + "type": "inline_equation", + "content": "34.07\\%" + }, + { + "bbox": [ + 312, + 296, + 547, + 598 + ], + "type": "text", + "content": ", as the reduced detail restricts its ability to deliver precise answers. MedDr performs well, achieving an overall score of " + }, + { + "bbox": [ + 312, + 296, + 547, + 598 + ], + "type": "inline_equation", + "content": "67.70\\%" + }, + { + "bbox": [ + 312, + 296, + 547, + 598 + ], + "type": "text", + "content": " with patch-level inputs, though this drops slightly to " + }, + { + "bbox": [ + 312, + 296, + 547, + 598 + ], + "type": "inline_equation", + "content": "64.25\\%" + }, + { + "bbox": [ + 312, + 296, + 547, + 598 + ], + "type": "text", + "content": " when using the slide thumbnail due to the loss of fine visual details. SlideChat outperforms all other methods, attaining a leading overall accuracy of " + }, + { + "bbox": [ + 312, + 296, + 547, + 598 + ], + "type": "inline_equation", + "content": "81.17\\%" + }, + { + "bbox": [ + 312, + 296, + 547, + 598 + ], + "type": "text", + "content": ", excelling across all categories and significantly surpassing the competition. Even in more fine-grained pathological scenarios, as depicted in the left portion of Fig. 4, SlideChat remains the top-performing model across 13 tasks, particularly in areas such as cytomorphological characteristics, histopathological changes, disease detection, disease classification, and staging and grading, which require the identification of complex pathological visual features. Compared with baselines taking some patches or slide thumbnail as inputs, SlideChat has the capability to analyze a significantly greater number of pathological features with enhanced detail, effectively capturing both localized features and overarching global patterns, allowing SlideChat to provide more accurate and nuanced insights into pathological variations." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 312, + 611, + 547, + 705 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 611, + 547, + 705 + ], + "spans": [ + { + "bbox": [ + 312, + 611, + 547, + 705 + ], + "type": "text", + "content": "SlideBench-VQA (BCNB) In the zero-shot VQA setting, SlideChat significantly outperforms all other models, achieving the highest overall score of " + }, + { + "bbox": [ + 312, + 611, + 547, + 705 + ], + "type": "inline_equation", + "content": "54.14\\%" + }, + { + "bbox": [ + 312, + 611, + 547, + 705 + ], + "type": "text", + "content": " on SlideBench-VQA (BCNB). It particularly excels in identifying tumor types, far surpassing other baselines in this task. This performance highlights SlideChat's generalization capability across a wide range of tasks. When taking patches as inputs, GPT-4o outperforms both MedDr and LLaVA" + } + ] + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 744, + 315, + 753 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 744, + 315, + 753 + ], + "spans": [ + { + "bbox": [ + 295, + 744, + 315, + 753 + ], + "type": "text", + "content": "5139" + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 110, + 79, + 492, + 214 + ], + "blocks": [ + { + "bbox": [ + 110, + 79, + 492, + 214 + ], + "lines": [ + { + "bbox": [ + 110, + 79, + 492, + 214 + ], + "spans": [ + { + "bbox": [ + 110, + 79, + 492, + 214 + ], + "type": "image", + "image_path": "977a5e617e035189e47c3f27f96daab07f3564ef79abc62b7cccc95c21c30d96.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 107, + 226, + 501, + 239 + ], + "lines": [ + { + "bbox": [ + 107, + 226, + 501, + 239 + ], + "spans": [ + { + "bbox": [ + 107, + 226, + 501, + 239 + ], + "type": "text", + "content": "Figure 4. Accuracy on different tasks in SlideBench-VQA (TCGA) (left) and SlideBench-VQA (BCNB) (right)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 258, + 296, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 258, + 296, + 421 + ], + "spans": [ + { + "bbox": [ + 61, + 258, + 296, + 421 + ], + "type": "text", + "content": "Med, achieving a score of " + }, + { + "bbox": [ + 61, + 258, + 296, + 421 + ], + "type": "inline_equation", + "content": "41.43\\%" + }, + { + "bbox": [ + 61, + 258, + 296, + 421 + ], + "type": "text", + "content": ", though it still falls short of SlideChat by " + }, + { + "bbox": [ + 61, + 258, + 296, + 421 + ], + "type": "inline_equation", + "content": "12.59\\%" + }, + { + "bbox": [ + 61, + 258, + 296, + 421 + ], + "type": "text", + "content": ". Notably, GPT-4o and LLaVA-Med performed very poorly, achieving a score of zero across all tasks when evaluated using slide thumbnails from this testing set. MedDr and Quilt-LLaVA both achieve moderate results, with MedDr scoring " + }, + { + "bbox": [ + 61, + 258, + 296, + 421 + ], + "type": "inline_equation", + "content": "33.67\\%" + }, + { + "bbox": [ + 61, + 258, + 296, + 421 + ], + "type": "text", + "content": " on patches and " + }, + { + "bbox": [ + 61, + 258, + 296, + 421 + ], + "type": "inline_equation", + "content": "35.48\\%" + }, + { + "bbox": [ + 61, + 258, + 296, + 421 + ], + "type": "text", + "content": " on thumbnails, while Quilt-LLaVA attains " + }, + { + "bbox": [ + 61, + 258, + 296, + 421 + ], + "type": "inline_equation", + "content": "32.19\\%" + }, + { + "bbox": [ + 61, + 258, + 296, + 421 + ], + "type": "text", + "content": " on patches and " + }, + { + "bbox": [ + 61, + 258, + 296, + 421 + ], + "type": "inline_equation", + "content": "41.55\\%" + }, + { + "bbox": [ + 61, + 258, + 296, + 421 + ], + "type": "text", + "content": " on thumbnails. This outcome highlights that, for complex WSIs, relying solely on relatively sufficient visual features is inadequate for effectively supporting a range of tasks. In the more fine-grained pathological tasks of the BCNB benchmark, SlideChat attains state-of-the-art performance on 5 out of 7 tasks (Fig. 4), further demonstrating the effectiveness of SlideChat." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 434, + 296, + 610 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 434, + 296, + 610 + ], + "spans": [ + { + "bbox": [ + 61, + 434, + 296, + 610 + ], + "type": "text", + "content": "WSI-VQA* We also curated a subset of closed-set VQA pairs from the public WSI-VQA [8] dataset, referred to as WSI-VQA*, which exclusively consists of samples that were not in the training set, to evaluate the model's performance. SlideChat demonstrates the highest performance with a score of 60.18. Although MedDr performs well with both patch inputs (54.36%) and slide thumbnail inputs (50.95%), it still falls short compared to SlideChat. GPT-4o struggles significantly, especially with slide thumbnails, scoring only 14.03%, which highlights the limitations of using lower-resolution inputs. SlideChat leverages both detailed local information and broader context to process WSIs, making it the most effective model for this benchmark. This further emphasizes its superior capability in handling whole-slide data for VQA tasks." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 623, + 296, + 705 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 623, + 296, + 705 + ], + "spans": [ + { + "bbox": [ + 61, + 623, + 296, + 705 + ], + "type": "text", + "content": "Ablation We performed ablation experiments from different perspectives as follows: a) Large Language Model Comparison: Firstly, we compare the performance of several large language models, each with a parameter scale of approximately 7 billion, including Vicuna-7B-v1.5 [10], Phi-3-Mini-4k-Instruct [1], Qwen1.5-7B-Chat [5], Llama3-8B-Instruct [4], InternLM2-Chat-7B [6], and Qwen2.5-7B" + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 316, + 296, + 545, + 403 + ], + "blocks": [ + { + "bbox": [ + 313, + 257, + 547, + 278 + ], + "lines": [ + { + "bbox": [ + 313, + 257, + 547, + 278 + ], + "spans": [ + { + "bbox": [ + 313, + 257, + 547, + 278 + ], + "type": "text", + "content": "Table 3. Performance Comparison of LLMs and Slide Encoder on WSI Captioning and VQA Tasks." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 316, + 296, + 545, + 403 + ], + "lines": [ + { + "bbox": [ + 316, + 296, + 545, + 403 + ], + "spans": [ + { + "bbox": [ + 316, + 296, + 545, + 403 + ], + "type": "table", + "html": "
LLMsSlide EncoderCaptionVQA (TCGA)VQA (BCNB)WSI-VQA*
Vicuna-7B-v1.5✓3.2841.4341.4331.98
Phi-3-Mini-4k-Instruct✓2.6679.9343.9260.18
Qwen1.5-7B-Chat✓2.9277.6344.0756.89
Llama3-8B-Instruct✓3.3078.7842.8255.25
Internlm2-Chat-7B✓3.3079.1052.1356.76
Qwen2.5-3B-Instruct✓3.4080.3245.7956.38
Qwen2.5-14B-Instruct✓3.3982.1451.5760.94
Qwen2.5-7B-Instruct✓4.1181.1754.1460.18
Qwen2.5-7B-Instruct×3.9581.2145.4959.67
", + "image_path": "8ebed14e972142fc4554727ea1efffcffa0d44c5672ed1f379ce76c97723e136.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 312, + 437, + 547, + 705 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 437, + 547, + 705 + ], + "spans": [ + { + "bbox": [ + 312, + 437, + 547, + 705 + ], + "type": "text", + "content": "Instruct [32]. Specifically, we measured their performance on the SlideBench-Caption using the GPT-score and their accuracy on three VQA datasets. As shown in Tab. 3, the results demonstrate that SlideChat, powered by the Qwen2.5-7B-Instruct model, achieved the highest performance across all tasks, particularly excelling in the captioning task. These findings underscore the significant potential of developing SlideChat with the Qwen2.5-7B-Instruct model. Utilizing the Qwen2.5 model, we further evaluated models of varying scales and discovered that larger models generally exhibited superior performance, particularly the 7B and 14B parameter models. While these two models showed comparable performance across the three benchmarks, the 14B model surpassed the 7B model by " + }, + { + "bbox": [ + 312, + 437, + 547, + 705 + ], + "type": "inline_equation", + "content": "2.57\\%" + }, + { + "bbox": [ + 312, + 437, + 547, + 705 + ], + "type": "text", + "content": " on the SlideBench-VQA (TCGA). Given computational resource constraints, SlideChat uses the 7B model by default to achieve the best balance between performance and resource efficiency. b) Slide-level encoder effectiveness: We investigate the effectiveness of the slide-level encoder by initially removing it from SlideChat and employing a two-stage training approach. In the first stage, we exclusively trained the projection layers. However, this approach failed to reduce training loss or generate coherent text effectively," + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 744, + 315, + 753 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 744, + 315, + 753 + ], + "spans": [ + { + "bbox": [ + 295, + 744, + 315, + 753 + ], + "type": "text", + "content": "5140" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 138, + 83, + 283, + 129 + ], + "blocks": [ + { + "bbox": [ + 115, + 79, + 124, + 89 + ], + "lines": [ + { + "bbox": [ + 115, + 79, + 124, + 89 + ], + "spans": [ + { + "bbox": [ + 115, + 79, + 124, + 89 + ], + "type": "text", + "content": "A" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 138, + 83, + 283, + 129 + ], + "lines": [ + { + "bbox": [ + 138, + 83, + 283, + 129 + ], + "spans": [ + { + "bbox": [ + 138, + 83, + 283, + 129 + ], + "type": "image", + "image_path": "a2ad6f0d11e37162874167c4a6b46e6971bca861a7172843b4e2405a85783298.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 115, + 172, + 124, + 198 + ], + "blocks": [ + { + "bbox": [ + 137, + 134, + 296, + 171 + ], + "lines": [ + { + "bbox": [ + 137, + 134, + 296, + 171 + ], + "spans": [ + { + "bbox": [ + 137, + 134, + 296, + 171 + ], + "type": "text", + "content": "Select the most accurate description of the cytomorphological characteristics from image. A. Cells with increased nuclear-to-cytoplasmic ratio, hyperchromatic nuclei, and prominent nucleoli. B. Cells with uniform nuclei and abundant cytoplasm. C. Cells with minimal nuclear abnormalities. D. Cells with enlarged, clear nuclei and small nucleoli." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 115, + 172, + 124, + 198 + ], + "lines": [ + { + "bbox": [ + 115, + 172, + 124, + 198 + ], + "spans": [ + { + "bbox": [ + 115, + 172, + 124, + 198 + ], + "type": "image", + "image_path": "3e33e9c466c428539ae7aade1888ad947d76b9eae9b5b2481b29ffafcfab2344.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 184, + 246, + 197 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 184, + 246, + 197 + ], + "spans": [ + { + "bbox": [ + 138, + 184, + 246, + 197 + ], + "type": "text", + "content": "A. Cells with increased nuclear-to-cytoplasmic ratio, hyperchromatic nuclei, and prominent nucleoli." + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 334, + 83, + 496, + 173 + ], + "blocks": [ + { + "bbox": [ + 312, + 79, + 320, + 89 + ], + "lines": [ + { + "bbox": [ + 312, + 79, + 320, + 89 + ], + "spans": [ + { + "bbox": [ + 312, + 79, + 320, + 89 + ], + "type": "text", + "content": "B" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 334, + 83, + 496, + 173 + ], + "lines": [ + { + "bbox": [ + 334, + 83, + 496, + 173 + ], + "spans": [ + { + "bbox": [ + 334, + 83, + 496, + 173 + ], + "type": "image", + "image_path": "ab789ba476b9a5cf8bb971c935c6b7fcd188f19279163dbc5653a03a3f13afcc.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 313, + 173, + 320, + 198 + ], + "blocks": [ + { + "bbox": [ + 313, + 173, + 320, + 198 + ], + "lines": [ + { + "bbox": [ + 313, + 173, + 320, + 198 + ], + "spans": [ + { + "bbox": [ + 313, + 173, + 320, + 198 + ], + "type": "image", + "image_path": "5c7d913f3149088674af34884c0eb1efe249a856cac6e55d2eb7590eafccf217.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 61, + 209, + 547, + 232 + ], + "lines": [ + { + "bbox": [ + 61, + 209, + 547, + 232 + ], + "spans": [ + { + "bbox": [ + 61, + 209, + 547, + 232 + ], + "type": "text", + "content": "Figure 5. Interpretability and visualization. We identify the top five patch tokens with the highest attention scores associated with the output text responses." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 332, + 182, + 423, + 195 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 332, + 182, + 423, + 195 + ], + "spans": [ + { + "bbox": [ + 332, + 182, + 423, + 195 + ], + "type": "text", + "content": "A. Dense collagen deposition with reduced cellularity" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 61, + 251, + 296, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 251, + 296, + 483 + ], + "spans": [ + { + "bbox": [ + 61, + 251, + 296, + 483 + ], + "type": "text", + "content": "likely due to the difficulty of learning the complex visual features of WSIs without the slide-level encoder. When using a frozen LLM for complex text generation, a simple projection proved inadequate for effectively integrating and aligning visual and textual features. Subsequently, we consider combining data from both stages and training SlideChat without the slide-level encoder by simultaneously updating both the projection layers and the LLM. Under this paradigm, performance on SlideBench-VQA (TCGA) and WSI-VQA*, which share the same distribution as the training set, was comparable to the two-stage training configuration with the slide-level encoder. However, a significant decline is observed when evaluating SlideBench-VQA (BCNB), which originates from a different domain; overall performance dropped by over " + }, + { + "bbox": [ + 61, + 251, + 296, + 483 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 61, + 251, + 296, + 483 + ], + "type": "text", + "content": " (Tab. 3), indicating a substantial reduction in the model's generalization ability. Therefore, we recommend incorporating a slide-level encoder to capture the complex visual features of whole slide, as it is particularly effective for cross-domain alignment and enhances the model's generalization performance." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 61, + 506, + 296, + 705 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 506, + 296, + 705 + ], + "spans": [ + { + "bbox": [ + 61, + 506, + 296, + 705 + ], + "type": "text", + "content": "Interpretability Despite SlideChat demonstrating promising results, concerns remained regarding the model's perception of large pathological slides. To further assess the model's interpretability, we calculated the correlation between the text output and specific image patches, thereby obtaining patch-level attention scores. By identifying the most significant patches, we gained insights into the precise areas the model focused on during response generation. Highlighting the most relevant regions of the tissue slides not only enhances transparency and bolsters the reliability of the model's outputs but also assists pathologists by directing attention to critical areas requiring closer scrutiny. Ultimately, such interpretability is essential for fostering trust in AI-assisted diagnostics and enhancing the precision and efficiency of clinical evaluations. As shown in Fig. 5, we are pleased to observe that the top five important patches identified by the model closely" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 251, + 548, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 251, + 548, + 473 + ], + "spans": [ + { + "bbox": [ + 313, + 251, + 548, + 473 + ], + "type": "text", + "content": "corresponded with the features described in SlideChat's output. Our extraction method retrieves attention weights for patch tokens from each generated token, averaging them across layers and heads. We then identify the top five patch tokens with the highest attention weights for further analysis. For example, in Fig. 5 (A), the highlighted patches clearly emphasized regions exhibiting an increased nuclear-to-cytoplasmic ratio, hyperchromatic nuclei, and prominent nucleoli. Similarly, in Fig. 5 (B), the selected patches demonstrated areas with dense collagen deposition and reduced cellularity, as detailed in the model's response. This alignment between the highlighted image regions and the textual outputs significantly enhances the model's interpretability, providing increased confidence that it accurately captures and assesses relevant histopathological features. Such consistency deepens our understanding of the model's reasoning processes regarding pathological slides and underscores the potential for integrating these AI systems into clinical workflows with greater assurance." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 485, + 384, + 497 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 485, + 384, + 497 + ], + "spans": [ + { + "bbox": [ + 313, + 485, + 384, + 497 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 505, + 548, + 644 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 505, + 548, + 644 + ], + "spans": [ + { + "bbox": [ + 313, + 505, + 548, + 644 + ], + "type": "text", + "content": "In this work, we present SlideChat, the first vision-language assistant capable of understanding gigapixel whole-slide images. Furthermore, we create SlideInstruction, a largest comprehensive WSI instruction-following dataset to develop SlideChat, as well as SlideBench, a multi-modal benchmark designed to evaluate SlideChat across diverse scenarios. SlideChat demonstrates excellent chat abilities and achieves state-of-the-art performance on 18 tasks. We bridge the gap between MLLMs and pathology images at the whole-slide level with SlideChat, and we believe this represents an advancement in the field of general medical artificial intelligence (GMAI)." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 661, + 547, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 661, + 547, + 694 + ], + "spans": [ + { + "bbox": [ + 313, + 661, + 547, + 694 + ], + "type": "text", + "content": "Acknowledgement The authors acknowledge the funding provided by the National Key Research and Development Program of China (No.2022ZD0160102)." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 744, + 314, + 754 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 744, + 314, + 754 + ], + "spans": [ + { + "bbox": [ + 295, + 744, + 314, + 754 + ], + "type": "text", + "content": "5141" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 63, + 80, + 119, + 92 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 80, + 119, + 92 + ], + "spans": [ + { + "bbox": [ + 63, + 80, + 119, + 92 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 63, + 99, + 295, + 704 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 67, + 99, + 295, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 99, + 295, + 152 + ], + "spans": [ + { + "bbox": [ + 67, + 99, + 295, + 152 + ], + "type": "text", + "content": "[1] Marah Abdin, Sam Ade Jacobs, Ammar Ahmad Awan, Jyoti Aneja, Ahmed Awadallah, Hany Awadalla, Nguyen Bach, Amit Bahree, Arash Bakhtiari, Harkirat Behl, et al. Phi-3 technical report: A highly capable language model locally on your phone. arXiv preprint arXiv:2404.14219, 2024. 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 153, + 295, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 153, + 295, + 206 + ], + "spans": [ + { + "bbox": [ + 67, + 153, + 295, + 206 + ], + "type": "text", + "content": "[2] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 4, 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 207, + 295, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 207, + 295, + 270 + ], + "spans": [ + { + "bbox": [ + 67, + 207, + 295, + 270 + ], + "type": "text", + "content": "[3] Faruk Ahmed, Andrew Sellergren, Lin Yang, Shawn Xu, Boris Babenko, Abbi Ward, Niels Olson, Arash Mohtashamian, Yossi Matias, Greg S Corrado, et al. Pathalign: A vision-language model for whole slide images in histopathology. arXiv preprint arXiv:2406.19578, 2024. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 272, + 228, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 272, + 228, + 282 + ], + "spans": [ + { + "bbox": [ + 67, + 272, + 228, + 282 + ], + "type": "text", + "content": "[4] AI@Meta. Llama 3 model card. 2024. 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 284, + 295, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 284, + 295, + 326 + ], + "spans": [ + { + "bbox": [ + 67, + 284, + 295, + 326 + ], + "type": "text", + "content": "[5] Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, et al. Qwen technical report. arXiv preprint arXiv:2309.16609, 2023. 5, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 327, + 295, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 327, + 295, + 369 + ], + "spans": [ + { + "bbox": [ + 67, + 327, + 295, + 369 + ], + "type": "text", + "content": "[6] Zheng Cai, Maosong Cao, Haojiong Chen, Kai Chen, Keyu Chen, Xin Chen, Xun Chen, Zehui Chen, Zhi Chen, Pei Chu, et al. Internl m2 technical report. arXiv preprint arXiv:2403.17297, 2024.5, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 67, + 370, + 295, + 412 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 370, + 295, + 412 + ], + "spans": [ + { + "bbox": [ + 67, + 370, + 295, + 412 + ], + "type": "text", + "content": "[7] Pingyi Chen, Honglin Li, Chenglu Zhu, Sunyi Zheng, and Lin Yang. Mi-gen: Multiple instance generation of pathology reports for gigapixel whole-slide images. arXiv preprint arXiv:2311.16480, 2023. 2, 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 67, + 413, + 295, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 413, + 295, + 456 + ], + "spans": [ + { + "bbox": [ + 67, + 413, + 295, + 456 + ], + "type": "text", + "content": "[8] Pingyi Chen, Chenglu Zhu, Sunyi Zheng, Honglin Li, and Lin Yang. Wsi-vqa: Interpreting whole slide images by generative visual question answering. arXiv preprint arXiv:2407.05603, 2024. 2, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 67, + 456, + 295, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 456, + 295, + 520 + ], + "spans": [ + { + "bbox": [ + 67, + 456, + 295, + 520 + ], + "type": "text", + "content": "[9] Richard J Chen, Tong Ding, Ming Y Lu, Drew FK Williamson, Guillaume Jaume, Andrew H Song, Bowen Chen, Andrew Zhang, Daniel Shao, Muhammad Shaban, et al. Towards a general-purpose foundation model for computational pathology. Nature Medicine, 30(3):850-862, 2024. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 63, + 521, + 295, + 585 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 521, + 295, + 585 + ], + "spans": [ + { + "bbox": [ + 63, + 521, + 295, + 585 + ], + "type": "text", + "content": "[10] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhanghao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E Gonzalez, et al. Vicuna: An open-source chatbot impressing gpt-4 with " + }, + { + "bbox": [ + 63, + 521, + 295, + 585 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 63, + 521, + 295, + 585 + ], + "type": "text", + "content": " * chatgpt quality. See https://vicuna.lmsys.org (accessed 14 April 2023), 2(3):6, 2023. 2, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 63, + 586, + 295, + 628 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 586, + 295, + 628 + ], + "spans": [ + { + "bbox": [ + 63, + 586, + 295, + 628 + ], + "type": "text", + "content": "[11] Jiayu Ding, Shuming Ma, Li Dong, Xingxing Zhang, Shaohan Huang, Wenhui Wang, Nanning Zheng, and Furu Wei. Longnet: Scaling transformers to 1,000,000,000 tokens. arXiv preprint arXiv:2307.02486, 2023. 3, 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 63, + 629, + 295, + 681 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 629, + 295, + 681 + ], + "spans": [ + { + "bbox": [ + 63, + 629, + 295, + 681 + ], + "type": "text", + "content": "[12] Zhengrui Guo, Jiabo Ma, Yingxue Xu, Yihui Wang, Liansheng Wang, and Hao Chen. Histgen: Histopathology report generation via local-global feature encoding and cross-modal context interaction. arXiv preprint arXiv:2403.05396, 2024. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 63, + 682, + 295, + 704 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 682, + 295, + 704 + ], + "spans": [ + { + "bbox": [ + 63, + 682, + 295, + 704 + ], + "type": "text", + "content": "[13] Sunan He, Yuxiang Nie, Zhixuan Chen, Zhiyuan Cai, Hongmei Wang, Shu Yang, and Hao Chen. Meddr:" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 82, + 547, + 704 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 333, + 82, + 547, + 113 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 82, + 547, + 113 + ], + "spans": [ + { + "bbox": [ + 333, + 82, + 547, + 113 + ], + "type": "text", + "content": "Diagnosis-guided bootstrapping for large-scale medical vision-language learning. arXiv preprint arXiv:2404.15127, 2024. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 315, + 115, + 547, + 178 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 115, + 547, + 178 + ], + "spans": [ + { + "bbox": [ + 315, + 115, + 547, + 178 + ], + "type": "text", + "content": "[14] Mahdi S Hosseini, Babak Ehteshami Bejnordi, Vincent Quoc-Huy Trinh, Lyndon Chan, Danial Hasan, Xingwen Li, Stephen Yang, Taehyo Kim, Haochen Zhang, Theodore Wu, et al. Computational pathology: a survey review and the way forward. Journal of Pathology Informatics, page 100357, 2024. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 315, + 181, + 547, + 212 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 181, + 547, + 212 + ], + "spans": [ + { + "bbox": [ + 315, + 181, + 547, + 212 + ], + "type": "text", + "content": "[15] Carolyn Hutter and Jean Claude Zenklusen. The cancer genome atlas: creating lasting value beyond its data. Cell, 173(2):283-285, 2018. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 315, + 214, + 547, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 214, + 547, + 267 + ], + "spans": [ + { + "bbox": [ + 315, + 214, + 547, + 267 + ], + "type": "text", + "content": "[16] Chunyuan Li, Cliff Wong, Sheng Zhang, Naoto Usuyama, Haotian Liu, Jianwei Yang, Tristan Naumann, Hoifung Poon, and Jianfeng Gao. Llava-med: Training a large language-and-vision assistant for biomedicine in one day. Advances in Neural Information Processing Systems, 36, 2024. 2, 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 315, + 269, + 547, + 332 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 269, + 547, + 332 + ], + "spans": [ + { + "bbox": [ + 315, + 269, + 547, + 332 + ], + "type": "text", + "content": "[17] Hao Li, Ying Chen, Yifei Chen, Rongshan Yu, Wenxian Yang, Liansheng Wang, Bowen Ding, and Yuchen Han. Generalizable whole slide image classification with fine-grained visual-semantic interaction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11398–11407, 2024. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 315, + 334, + 547, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 334, + 547, + 376 + ], + "spans": [ + { + "bbox": [ + 315, + 334, + 547, + 376 + ], + "type": "text", + "content": "[18] Ming Y Lu, Tiffany Y Chen, Drew FK Williamson, Melissa Zhao, Maha Shady, Jana Lipkova, and Faisal Mahmood. Aibased pathology predicts origins for cancers of unknown primary. Nature, 594(7861):106-110, 2021. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 315, + 378, + 547, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 378, + 547, + 431 + ], + "spans": [ + { + "bbox": [ + 315, + 378, + 547, + 431 + ], + "type": "text", + "content": "[19] Ming Y Lu, Bowen Chen, Drew FK Williamson, Richard J Chen, Ivy Liang, Tong Ding, Guillaume Jaume, Igor Odintsov, Long Phi Le, Georg Gerber, et al. A visual-language foundation model for computational pathology. Nature Medicine, 30(3):863-874, 2024. 3, 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 315, + 433, + 547, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 433, + 547, + 485 + ], + "spans": [ + { + "bbox": [ + 315, + 433, + 547, + 485 + ], + "type": "text", + "content": "[20] Ming Y Lu, Bowen Chen, Drew FK Williamson, Richard J Chen, Melissa Zhao, Aaron K Chow, Kenji Ikemura, Ahrong Kim, Dimitra Pouli, Ankush Patel, et al. A multimodal generative ai copilot for human pathology. Nature, pages 1-3, 2024. 1, 2, 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 315, + 487, + 547, + 550 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 487, + 547, + 550 + ], + "spans": [ + { + "bbox": [ + 315, + 487, + 547, + 550 + ], + "type": "text", + "content": "[21] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 315, + 552, + 547, + 616 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 552, + 547, + 616 + ], + "spans": [ + { + "bbox": [ + 315, + 552, + 547, + 616 + ], + "type": "text", + "content": "[22] Mehmet Saygin Seyfioglu, Wisdom O Ikezogwo, Fatemeh Ghezloo, Ranjay Krishna, and Linda Shapiro. Quilt-llava: Visual instruction tuning by extracting localized narratives from open-source histopathology videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13183-13192, 2024. 1, 2, 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 315, + 618, + 547, + 671 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 618, + 547, + 671 + ], + "spans": [ + { + "bbox": [ + 315, + 618, + 547, + 671 + ], + "type": "text", + "content": "[23] Zhuchen Shao, Yang Chen, Hao Bian, Jian Zhang, Guojun Liu, and Yongbing Zhang. Hvtsurv: Hierarchical vision transformer for patient-level survival prediction from whole slide image. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2209-2217, 2023. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 315, + 672, + 547, + 704 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 672, + 547, + 704 + ], + "spans": [ + { + "bbox": [ + 315, + 672, + 547, + 704 + ], + "type": "text", + "content": "[24] Andrew H Song, Guillaume Jaume, Drew FK Williamson, Ming Y Lu, Anurag Vaidya, Tiffany R Miller, and Faisal Mahmood. Artificial intelligence for digital and computa" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 745, + 315, + 754 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 745, + 315, + 754 + ], + "spans": [ + { + "bbox": [ + 295, + 745, + 315, + 754 + ], + "type": "text", + "content": "5142" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 64, + 82, + 296, + 613 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 82, + 82, + 294, + 102 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 82, + 294, + 102 + ], + "spans": [ + { + "bbox": [ + 82, + 82, + 294, + 102 + ], + "type": "text", + "content": "tional pathology. Nature Reviews Bioengineering, 1(12): 930-949, 2023. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 64, + 105, + 295, + 177 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 105, + 295, + 177 + ], + "spans": [ + { + "bbox": [ + 64, + 105, + 295, + 177 + ], + "type": "text", + "content": "[25] Joey Spronck, Thijs Gelton, Leander van Eekelen, Joep Bogaerts, Leslie Tessier, Mart van Rijthoven, Lieke van der Woude, Michel van den Heuvel, Willemijn Theelen, Jeroen van der Laak, et al. nnunet meets pathology: bridging the gap for application to whole-slide images and computational biomarkers. In Medical Imaging with Deep Learning, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 64, + 180, + 296, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 180, + 296, + 243 + ], + "spans": [ + { + "bbox": [ + 64, + 180, + 296, + 243 + ], + "type": "text", + "content": "[26] Yuxuan Sun, Chenglu Zhu, Sunyi Zheng, Kai Zhang, Lin Sun, Zhongyi Shui, Yunlong Zhang, Honglin Li, and Lin Yang. Pathasst: A generative foundation ai assistant towards artificial general intelligence of pathology. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 5034-5042, 2024. 1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 64, + 246, + 294, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 246, + 294, + 308 + ], + "spans": [ + { + "bbox": [ + 64, + 246, + 294, + 308 + ], + "type": "text", + "content": "[27] Eugene Vorontsov, Alican Bozkurt, Adam Casson, George Shaikovski, Michal Zelechowski, Kristen Severson, Eric Zimmermann, James Hall, Neil Tenenholtz, Nicolo Fusi, et al. A foundation model for clinical-grade computational pathology and rare cancers detection. Nature Medicine, pages 1-12, 2024. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 64, + 310, + 294, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 310, + 294, + 361 + ], + "spans": [ + { + "bbox": [ + 64, + 310, + 294, + 361 + ], + "type": "text", + "content": "[28] Xiyue Wang, Junhan Zhao, Eliana Marostica, Wei Yuan, Ji etian Jin, Jiayu Zhang, Ruijiang Li, Hongping Tang, Kanran Wang, Yu Li, et al. A pathology foundation model for cancer diagnosis and prognosis prediction. Nature, pages 1-9, 2024. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 64, + 365, + 294, + 416 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 365, + 294, + 416 + ], + "spans": [ + { + "bbox": [ + 64, + 365, + 294, + 416 + ], + "type": "text", + "content": "[29] Feng Xu, Chuang Zhu, Wenqi Tang, Ying Wang, Yu Zhang, Jie Li, Hongchuan Jiang, Zhongyue Shi, Jun Liu, and Mulan Jin. Predicting axillary lymph node metastasis in early breast cancer using deep learning on primary tumor biopsy slides. Frontiers in oncology, 11:759007, 2021. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 64, + 418, + 294, + 471 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 418, + 294, + 471 + ], + "spans": [ + { + "bbox": [ + 64, + 418, + 294, + 471 + ], + "type": "text", + "content": "[30] Hanwen Xu, Naoto Usuyama, Jaspreet Bagga, Sheng Zhang, Rajesh Rao, Tristan Naumann, Cliff Wong, Zelalem Gero, Javier González, Yu Gu, et al. A whole-slide foundation model for digital pathology from real-world data. Nature, pages 1-8, 2024. 1, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 64, + 473, + 294, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 473, + 294, + 525 + ], + "spans": [ + { + "bbox": [ + 64, + 473, + 294, + 525 + ], + "type": "text", + "content": "[31] Yingxue Xu, Yihui Wang, Fengtao Zhou, Jiabo Ma, Shu Yang, Huangjing Lin, Xin Wang, Jiguang Wang, Li Liang, Anjia Han, et al. A multimodal knowledge-enhanced whole-slide pathology foundation model. arXiv preprint arXiv:2407.15362, 2024. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 64, + 527, + 294, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 527, + 294, + 568 + ], + "spans": [ + { + "bbox": [ + 64, + 527, + 294, + 568 + ], + "type": "text", + "content": "[32] An Yang, Baosong Yang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Zhou, Chengpeng Li, Chengyuan Li, Dayiheng Liu, Fei Huang, et al. Qwen2 technical report. arXiv preprint arXiv:2407.10671, 2024. 5, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 64, + 571, + 294, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 571, + 294, + 613 + ], + "spans": [ + { + "bbox": [ + 64, + 571, + 294, + 613 + ], + "type": "text", + "content": "[33] Fengtao Zhou and Hao Chen. Cross-modal translation and alignment for survival analysis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 21485-21494, 2023. 2" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 745, + 315, + 753 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 745, + 315, + 753 + ], + "spans": [ + { + "bbox": [ + 295, + 745, + 315, + 753 + ], + "type": "text", + "content": "5143" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/SmartCLIP_ Modular Vision-language Alignment with Identification Guarantees/28f50741-407e-4ef2-b9fa-f2bf47df9fd3_content_list.json b/2025/SmartCLIP_ Modular Vision-language Alignment with Identification Guarantees/28f50741-407e-4ef2-b9fa-f2bf47df9fd3_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..307cfc259d55d1a88c9217b8a3b021994e765634 --- /dev/null +++ b/2025/SmartCLIP_ Modular Vision-language Alignment with Identification Guarantees/28f50741-407e-4ef2-b9fa-f2bf47df9fd3_content_list.json @@ -0,0 +1,1818 @@ +[ + { + "type": "text", + "text": "SmartCLIP: Modular Vision-language Alignment with Identification Guarantees", + "text_level": 1, + "bbox": [ + 89, + 130, + 906, + 154 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Shaoan Xie\\*, Lingjing\\*, Yujia Zheng\\*, Yu Yao\\*, Zeyu Tang\\*, Eric P. Xing\\*, Guangyi Chen\\*, Kun Zhang\\*,", + "bbox": [ + 246, + 179, + 750, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Carnegie Mellon University", + "bbox": [ + 374, + 218, + 622, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Mohamed bin Zayed University of Artificial Intelligence", + "bbox": [ + 245, + 234, + 750, + 250 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3 University of Sydney", + "bbox": [ + 400, + 252, + 596, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "* Equal contribution", + "bbox": [ + 442, + 273, + 553, + 285 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 321, + 326, + 338 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Contrastive Language-Image Pre-training (CLIP) [37] has emerged as a pivotal model in computer vision and multimodal learning, achieving state-of-the-art performance at aligning visual and textual representations through contrastive learning. However, CLIP struggles with potential information misalignment in many image-text datasets and suffers from entangled representation. On the one hand, short captions for a single image in datasets like MSCOCO may describe disjoint regions in the image, leaving the model uncertain about which visual features to retain or disregard. On the other hand, directly aligning long captions with images can lead to the retention of entangled details, preventing the model from learning disentangled, atomic concepts - ultimately limiting its generalization on certain downstream tasks involving short prompts.", + "bbox": [ + 88, + 354, + 485, + 582 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we establish theoretical conditions that enable flexible alignment between textual and visual representations across varying levels of granularity. Specifically, our framework ensures that a model can not only preserve cross-modal semantic information in its entirety but also disentangle visual representations to capture fine-grained textual concepts. Building on this foundation, we introduce SmartCLIP, a novel approach that identifies and aligns the most relevant visual and textual representations in a modular manner. Superior performance across various tasks demonstrates its capability to handle information misalignment and supports our identification theory. The code is available at https://github.com/Mid-Push/SmartCLIP.", + "bbox": [ + 89, + 583, + 483, + 792 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 827, + 220, + 843 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Contrastive Language-Image Pre-training (CLIP) [37] has been the cornerstone for many computer vision and machine learning tasks, such as text-to-image retrieval [2], im", + "bbox": [ + 89, + 854, + 483, + 902 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e39db5bd66f7dda691fab008f1ddff384cea3b7bfbf6338af8099258c3a4861e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 524, + 325, + 663, + 411 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "COCO:", + "text_level": 1, + "bbox": [ + 674, + 330, + 715, + 342 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "A very cute teddy bear holding a pen.", + "bbox": [ + 674, + 343, + 846, + 353 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "A stuffed bear that is sitting in a chair", + "bbox": [ + 674, + 354, + 857, + 363 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "A brown bear wearing a sweater next to a pen and paper.", + "bbox": [ + 674, + 364, + 887, + 383 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "A brown teddy bear in a dark sweater sits on a black striped chair while holding a pen.", + "bbox": [ + 674, + 385, + 883, + 404 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ShareGPT4u:", + "text_level": 1, + "bbox": [ + 534, + 416, + 617, + 426 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In the center of the image, a light brown teddy bear is seated on a black chair with white stripes. The teddy bear, dressed in a green sweater adorned with a yellow flower, holds a blue pen in its paw, as if ready to sign a document. The chair is positioned on a gray carpeted floor. To the right of the chair, there's a stack of papers neatly arranged on the floor. The scene creates an atmosphere of anticipation, as if the teddy bear is about to make a significant decision.", + "bbox": [ + 531, + 429, + 887, + 498 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Figure 1. Depiction of two primary challenges for CLIP. (1) Information Misalignment: An image can be paired with multiple captions that describe disparate aspects - the first caption contains concepts \"bear\" and \"pen\" whereas the second only mentions \"bear\" and \"paper\". Aligning the image with both captions leads to the loss of key concepts \"pen\" and \"paper\" not shared across the captions. (2) Entangled Representations: Long, detailed captions involving multiple concepts (e.g., \"chair\", \"pen\", \"flower\", \"floor\") encourage the model to form entangled representations, hindering independent understanding of each individual concept.", + "bbox": [ + 511, + 516, + 906, + 655 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "age and video understanding [6, 16, 39, 42, 46, 58], and generative models [28, 38, 40]. It aligns the representations from different modalities with a contrastive learning loss [8, 35]. Specifically, each image-caption pair in the dataset is treated as a positive pair, while negative pairs are created by matching images with captions randomly drawn from the dataset. The image and text encoders are trained with a symmetric cross-entropy loss that draws the image and text representations in each positive pair together while pulling the negative pairs' representations apart.", + "bbox": [ + 511, + 670, + 908, + 821 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Training CLIP requires vast amounts of image-text pairs, making it challenging to maintain dataset quality at such a large scale. In particular, the quality of text captions has been a key concern, prompting the development of various methods to enhance both their diversity and accu", + "bbox": [ + 511, + 824, + 908, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "29780", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "racy. ALIGN [15] shows that scaling up the dataset size can compensate for the noisy text supervision. BLIP models [23, 24] improve captions by incorporating additional captioning and filtering mechanisms. VE-CLIP [21] introduces a visual-enriched captioning approach to further refine caption quality. Similarly, LaCLIP [10] leverages language models to rewrite captions, while RecapCLIP [25] uses LLaMA-3 [29] to generate captions for 1.3 billion images. Despite these efforts, recent findings reveal that longer and seemingly higher-quality captions do not necessarily yield improved performance on many downstream tasks [20]. Li et al. [25] find that the text-to-image retrieval accuracy on Flickr30K drops from 84.2 to 74.1 when replacing the original captions with longer captions.", + "bbox": [ + 89, + 90, + 480, + 303 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "A key issue contributing to the observed performance degradation is the information misalignment between images and their captions, a problem that becomes more pronounced when multiple captions are paired with a single image. On one hand, an image may be paired with several captions, each capturing only a partial aspect of the image. In Figure 1, aligning the image with the first caption, \"A very cute teddy bear holding a pen\", risks forcing the model to discard other important concepts like \"chair\" and \"paper\", which are required to align with the second and third captions in the pink text box. This misalignment between image and text introduces conflicts during standard CLIP training, leading to the loss of key visual concepts.", + "bbox": [ + 89, + 304, + 482, + 501 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "On the other hand, training CLIP with long, detailed captions, as seen in recent approaches [10, 21, 25], encourages the model to learn entangled representations of multiple concepts bundled together in a single caption. Thus, it remains challenging to explicitly extract object/concept-centric representations from CLIP's visual representation. This entanglement is particularly problematic for tasks that require individual, atomic concepts or novel combinations of them, as empirically observed on short-text-to-image retrieval tasks [54]. In Figure 1, the long caption generated by ShareGPT4V [7] contains an exhaustive set of concepts such as \"chair\", \"pen\", \"flower\", and \"floor\". This aggregation can hinder the model's performance on tasks that demand the understanding of each concept individually.", + "bbox": [ + 89, + 503, + 482, + 715 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we propose a refined approach to representation alignment in vision-language models like CLIP [37]. We frame the alignment challenge as a latent-variable identification problem and develop theoretical conditions that enable flexible alignment between textual and visual representations at different levels of granularity. Our framework enables the model to preserve the complete cross-modal information while also disentangling representations to capture fine-grained concepts, effectively addressing the misalignment and disentanglement issues discussed earlier.", + "bbox": [ + 89, + 717, + 482, + 867 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Building on these theoretical insights, we introduce SmartCLIP, a novel method that identifies and aligns vi", + "bbox": [ + 89, + 869, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "sual and textual concepts in a modular manner. Specifically, we design a mask network that selects a subset of dimensions from the full representations, corresponding to only the concepts present in each specific caption. This allows the model to perform text-image alignment over the most relevant concepts modules, rather than the entire representation. We empirically demonstrate that SmartCLIP outperforms state-of-the-art models across a range of downstream tasks, showcasing its effectiveness in addressing alignment challenges. In particular, SmartCLIP significantly improves retrieval performance across text lengths, achieving $98.7\\%$ accuracy (up from $78.2\\%$ ) on the ShareGPT4V long text-to-image retrieval tasks, while boosting short text-to-image retrieval R1 from $56.1\\%$ to $66.0\\%$ .", + "bbox": [ + 511, + 90, + 903, + 301 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our main contributions are summarized as follows.", + "bbox": [ + 511, + 301, + 852, + 316 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "i We identify critical issues of information misalignment and entangled representations within the CLIP framework. To overcome these challenges, we propose a latent-variable formulation and establish theoretical conditions that guarantee the recovery of the latent variables.", + "bbox": [ + 514, + 318, + 903, + 393 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ii Building upon our theoretical findings, we propose SmartCLIP, featuring adaptive masking and a modular contrastive learning objective that facilitates the learning of disentangled, modular representations.", + "bbox": [ + 511, + 393, + 903, + 454 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "iii We perform extensive experiments on a variety of tasks, including long and short text-to-image retrieval, zero-shot classification, and text-to-image generation. Smart-CLIP consistently outperforms or matches state-of-the-art models across these benchmarks, demonstrating its efficacy and validating our theoretical contributions.", + "bbox": [ + 508, + 454, + 903, + 544 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 511, + 558, + 653, + 571 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Vision-language models. The breakthrough of CLIP [37] has attracted significant attention from the community. SLIP [33] and DECLIP [26] propose to incorporate self-supervised learning techniques to improve the learned representation. Coca [51] introduces a decoder in addition to the contrastive learning branch. LiT [52] locks the image encoder and only finetunes the text encoder. SigLIP [53] adopts a simple sigmoid loss to handle large training batch sizes. LoTLIP [48] inserts corner tokens after the classification token to support long-text understanding. TULIP [34] replaces the absolute position embedding with the relative position embedding to support longer text understanding. ALIGN [15] demonstrates that increasing dataset size can mitigate the impact of noisy text supervision. Recent methods have been focusing on generating better captions [20, 21, 23-25, 59]. CLIP-MOE [56] introduces mixture-of-experts to CLIP. LLM2CLIP [12] augments CLIP with large language models. LongCLIP [54] extends the token constraint of CLIP from 77 to 248 and applies PCA to perform short text-to-image contrastive learning to preserve its short text capability. Llip [22] learns a text-dependent vi", + "bbox": [ + 511, + 583, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "29781", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "sual representation by mixing a set of learnable tokens with a cross-attention module. In contrast, SmartCLIP directly learns a single global representation that encodes all disentangled, interpretable concepts through masking.", + "bbox": [ + 89, + 90, + 480, + 151 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Latent variable identification. Learning high-level, semantic information from low-level observational data (e.g., images and text) can often be formulated as latent-variable identification problems. Though appealing, such tasks are accompanied by substantial difficulties, especially for complex real-world data distributions involving nonlinear generating functions. Recently, a wealth of papers [1, 5, 9, 11, 14, 17, 31, 32, 41, 44, 49, 55, 57] propose to overcome such obstacles by leveraging auxiliary information, such as temporal information, multiple domains, and multiple views/modalities. Especially related to our work are those that tap into the paired multi-view data to identify the shared information across available views [9, 11, 31, 32, 43, 49]. Recent work [11, 31, 32] relies on specific forms of latent variable distributions (e.g., independence or exponential family). These constraints restrict their applicability for distributions entailing complex interactions among latent variables. Prior work [9, 43] adopt more flexible assumptions on the underlying distribution and identify blocks of latent variables directly shared by two views arising from data augmentations, which is extended to a multi-view setting [49]. The problem under investigation can be viewed as a form of this multi-view setting where the paired image and text captions are considered as views sharing the semantic latent variable. Existing works [9, 43, 49] assume that views are grouped over all the data pairs and this grouping information is known so that one can learn a designated encoder for each view group. However, this view grouping information is inaccessible for our problems – for any two text captions of different images, we cannot judge whether they belong to the same view group. In our theory section, we show that by properly utilizing the data-generating process, we can learn such information directly and further achieve the desired identification results, thus generalizing existing multi-view latent variable identification results.", + "bbox": [ + 91, + 151, + 483, + 680 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Problem Formulation", + "text_level": 1, + "bbox": [ + 89, + 694, + 295, + 709 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As motivated previously, we aim to 1) preserve all the semantic information shared across modalities, and 2) learn a disentangled representation that corresponds to textual concepts at diverse granularity levels. To this end, we propose the following data-generating process underlying the visual-language data distribution.", + "bbox": [ + 89, + 719, + 482, + 809 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Notations. We indicate the dimensionality of a vector with $d(\\cdot)$ . We denote a subset of dimensions of a vector $\\mathbf{z}$ with $[\\mathbf{z}]_{\\mathcal{B}}$ with the index set $\\mathcal{B}$ . We define the set of indices whose corresponding values are nonzero in a vector $\\mathbf{m}$ with $\\mathcal{B}(\\mathbf{m}) := \\{i \\in d(\\mathbf{m}) : [\\mathbf{m}]_i \\neq 0\\}$ .", + "bbox": [ + 89, + 810, + 482, + 886 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Data-generating processes. We depict the data-generating", + "bbox": [ + 89, + 886, + 482, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "process in Figure 2 and in (1).", + "bbox": [ + 511, + 90, + 714, + 106 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {z} _ {\\mathrm {T}} := \\mathbf {z} _ {\\mathrm {I}} \\odot \\mathbf {m}; \\mathbf {i} := g _ {\\mathrm {I}} \\left(\\mathbf {z} _ {\\mathrm {I}}, \\epsilon_ {\\mathrm {I}}\\right); \\mathbf {t} := g _ {\\mathrm {T}} \\left(\\mathbf {z} _ {\\mathrm {T}}, \\epsilon_ {\\mathrm {T}}\\right). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 537, + 119, + 903, + 133 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We assume that each pair of image $\\mathbf{i} \\in \\mathcal{I} \\subset \\mathbb{R}^{d(\\mathbf{i})}$ and text caption $\\mathbf{t} \\in \\mathcal{T} \\subset \\mathbb{R}^{d(\\mathbf{t})}$ originate from semantic information $\\mathbf{z}_{\\mathrm{I}} \\in \\mathcal{Z}_{\\mathrm{I}} \\subset \\mathbb{R}^{d(\\mathbf{z}_{\\mathrm{I}})}$ , together with modality-specific variations $\\epsilon_{\\mathrm{I}}$ and $\\epsilon_{\\mathrm{T}}$ (e.g., illumination for images, tenses for text), through generating functions $g_{\\mathrm{I}}: (\\mathbf{z}_{\\mathrm{I}}, \\epsilon_{\\mathrm{I}}) \\mapsto \\mathbf{i}$ and $g_{\\mathrm{T}}: (\\mathbf{z}_{\\mathrm{T}}, \\epsilon_{\\mathrm{T}}) \\mapsto \\mathbf{t}$ respectively. We treat the text caption as continuous variables, as each word can be represented with a continuous word embedding vector in practice [3, 30].", + "bbox": [ + 511, + 147, + 903, + 268 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As demonstrated in Figure 1, text captions of the same image often convey partial information of the entire image semantics. Thus, we associate each text caption's representation $\\mathbf{z}_{\\mathrm{T}} := \\mathbf{m} \\odot \\mathbf{z}_{\\mathrm{I}}$ with a binary random mask $\\mathbf{m} \\in \\mathcal{M} \\subset \\{0,1\\}^{d(\\mathbf{z}_{\\mathrm{I}})}$ that eliminates information absent in the specific caption $\\mathbf{t}$ .", + "bbox": [ + 511, + 270, + 906, + 359 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/b33d244b1b25d5e25cc42b572bbf56b964c3df73d939c85e22a9532eca9d9b5e.jpg", + "image_caption": [ + "Figure 2. The data-generating process. The text representation $\\mathbf{z}_{\\mathrm{T}}$ contains partial information of the vision representation $\\mathbf{z}_{\\mathrm{I}}$ , as indicated by masking $\\mathbf{m}$ . Modal-specific information is represented as $\\epsilon_{\\mathrm{I}}$ and $\\epsilon_{\\mathrm{T}}$ . Dashed edges indicate potential statistical dependence." + ], + "image_footnote": [], + "bbox": [ + 563, + 372, + 856, + 491 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Goals. Our two goals can be formalized as follows.", + "bbox": [ + 511, + 566, + 852, + 580 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "a Preserving cross-modal information: identifying the complete latent representation $\\mathbf{z}_1$ .", + "b Disentangling concepts: identifying concepts $\\mathbf{z}_{\\mathrm{T}}$ associated with a given textual description $\\mathbf{t}$ at different granularity levels potentially unseen during training." + ], + "bbox": [ + 511, + 583, + 903, + 657 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Examples. In Figure 1, the image i contains concepts \"bear\", \"chair\", and \"pen\", which we assume correspond to three components in the representation $\\mathbf{z}_{\\mathrm{I}}$ , say $[\\mathbf{z}_{\\mathrm{I}}]_{1}$ , $[\\mathbf{z}_{\\mathrm{I}}]_{2}$ , and $[\\mathbf{z}_{\\mathrm{I}}]_{3}$ . The first two COCO captions $\\mathbf{t}^{(1)}$ and $\\mathbf{t}^{(2)}$ mention only a subset of these concepts, i.e., (\"bear\", \"pen\") and (\"bear\", \"chair\") separately. Thus, the masks for these captions are $\\mathbf{m}^{(1)} = [1,0,1]$ and $\\mathbf{m}^{(2)} = [1,1,0]$ . The variables $\\epsilon_{\\mathrm{I}}$ and $\\epsilon_{\\mathrm{T}}$ represent modality-specific nuance factors such as illumination conditions in the image i and syntax in the text t. For Goal a, we seek to preserve the complete information $\\mathbf{z}_{\\mathrm{I}}$ . This involves retaining all relevant textual concepts present in the captions, namely \"bear\", \"chair\", and \"pen\" from captions $\\mathbf{t}^{(1)}$ and $\\mathbf{t}^{(2)}$ . For Goal b, we intend to disentangle the representation $\\mathbf{z}_{\\mathrm{I}}$ into finer concept blocks potentially unseen in the training. This includes identifying individual concepts such as \"bear\" in the dimen-", + "bbox": [ + 511, + 659, + 906, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "29782", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "sion $[\\mathbf{z}_{\\mathrm{I}}]_{1}$ , even if the training captions only include \"bear\" in combination with other concepts.", + "bbox": [ + 89, + 90, + 485, + 122 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Identification Theory", + "text_level": 1, + "bbox": [ + 89, + 135, + 294, + 152 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we present the theoretical results towards Goal a and Goal b. We show that under a suitable learning objective (2), the learned variables $(\\hat{\\mathbf{z}}_{\\mathrm{I}},\\hat{\\mathbf{z}}_{\\mathrm{T}})$ can be identified with the corresponding true quantities $(\\mathbf{z}_{\\mathrm{I}},\\mathbf{z}_{\\mathrm{T}})$ up to certain equivalent classes. In particular, we resort to the blockwise identifiability [18, 19, 43, 49] throughout this work. This suffices for our purpose since often several dimensions jointly (i.e., a block) comprise a meaningful concept while a single dimension may not be interpretable.", + "bbox": [ + 88, + 161, + 483, + 297 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Definition 4.1 (Block-wise Identifiability). The true variable $\\mathbf{v}$ is block-wise identifiable if it is related to its estimate $\\hat{\\mathbf{v}}$ through an invertible map $\\mathbf{v} \\mapsto \\hat{\\mathbf{v}}$ .", + "bbox": [ + 89, + 311, + 483, + 358 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The learning objective. Our estimation model consists of vision/text encoders $(f_{\\mathrm{I}}, f_{\\mathrm{T}})$ (smooth, invertible functions), and a masking function $\\hat{\\mathbf{m}}: \\mathcal{T} \\to \\mathcal{M}$ that estimates the true mask $\\mathbf{m}$ underlying a given text caption $\\mathbf{t}$ .", + "bbox": [ + 89, + 368, + 483, + 430 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\arg \\min \\| \\hat {\\mathbf {m}} (\\mathbf {t}) \\| _ {0}, \\quad \\text {s u b j e c t} \\\\ f _ {\\mathrm {I}}, f _ {\\mathrm {T}}, \\hat {\\mathbf {m}} \\\\ L _ {\\mathrm {s p a r s i t y}} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 440, + 361, + 484 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "(2)", + "bbox": [ + 459, + 481, + 480, + 494 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\arg \\min \\| f _ {\\mathrm {I}} (\\mathbf {i}) \\odot \\hat {\\mathbf {m}} (\\mathbf {t}) - f _ {\\mathrm {T}} (\\mathbf {t}) \\|, \\forall (\\mathbf {i}, \\mathbf {t}). \\\\ \\underbrace {f _ {\\mathrm {I}} , f _ {\\mathrm {T}} , \\hat {\\mathbf {m}}} _ {L _ {\\mathrm {a l i g n}}} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 150, + 489, + 424, + 534 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our learning objective (2) consists of an alignment term $L_{\\mathrm{align}}$ that draws the positive pairs across modalities. The negative pairs in regular contrastive losses [8, 35, 37] can be implemented through an entropy term at the sample limit [47]. This serves the same role as the invertibility condition on the encoder models [43], which we directly assume for theoretical convenience. In Section 5, we discuss practical considerations for constructing negative pairs. We enforce sparsity regularization $L_{\\mathrm{sparsity}}$ on the inferred mask $\\hat{\\mathbf{m}}$ to select the simplest representation.", + "bbox": [ + 89, + 546, + 483, + 696 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We introduce our key conditions in Condition 4.2 and theoretical results in Theorem 4.3.", + "bbox": [ + 89, + 696, + 483, + 727 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Condition 4.2 (Identification Conditions).", + "text_level": 1, + "bbox": [ + 89, + 737, + 372, + 753 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "i [Smoothness & invertibility]: Generating functions $g_{\\mathrm{I}}$ and $g_{\\mathrm{T}}$ are smooth and have smooth inverses.", + "ii [Fully-supported $p(\\mathbf{z}_{\\mathrm{I}},\\mathbf{m})]$ : The joint distribution over the semantic variable $\\mathbf{z}_{\\mathrm{I}}$ and the mask $\\mathbf{m}$ is fully supported: $p(\\mathbf{z}_{\\mathrm{I}},\\mathbf{m}) > 0$ for any $(\\mathbf{z}_{\\mathrm{I}},\\mathbf{m})\\in \\mathcal{Z}_{\\mathrm{I}}\\times \\mathcal{M}$ ." + ], + "bbox": [ + 107, + 753, + 482, + 830 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Discussion. Condition 4.2- $i$ ensures the generating functions $(g_{\\mathrm{I}}, g_{\\mathrm{T}})$ preserve latent variables' information, without which the task of recovering such latent variables would be ill-posed. Practically, the high dimensionality of image", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "data i offers sufficient capacity to hold all information, and the text variable t only contains information filtered through its mask m. This condition is widely employed in the latent-variable identification literature [13, 17, 18, 43]. Condition 4.2-ii prescribes that the representation $\\mathbf{z}_{\\mathrm{I}}$ and the mask m that marginally appear in the training distribution should also be present jointly with non-zero probability density. Interpreting the mask m as a concept selector (e.g., selecting \"bear\" and \"pen\"), this condition ensures that each concept retains its full range of variations (such as different shapes of bears and lengths of pens) across various mask selections. To satisfy this requirement, one can restrict the joint support $\\mathcal{Z}_{\\mathrm{I}} \\times \\mathcal{M}$ to an appropriate subset, ensuring that only relevant combinations of $\\mathbf{z}_{\\mathrm{I}}$ and m are present. Alternatively, one can enrich the caption set for each image, thereby increasing the diversity and coverage of concept combinations and filling in the joint support. This aligns with recent caption-augmentation techniques [21, 23-25, 54] as discussed in Section 1, revealing the synergy between our framework and existing efforts in the community.", + "bbox": [ + 511, + 90, + 906, + 393 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Theorem 4.3 (Concept Representation Identification). We assume the data-generating process in (1). Let $(f_{\\mathrm{I}}, f_{\\mathrm{T}}, \\hat{\\mathbf{m}})$ be an optimum of (2). Under Condition 4.2, the true representation $[\\mathbf{z}]_{\\tilde{\\mathcal{B}}}$ is block-wise identifiable for any index set $\\tilde{\\mathcal{B}}$ such that $\\tilde{\\mathcal{B}} = \\cup_{\\mathbf{m} \\in \\mathcal{V}} \\mathcal{B}(\\mathbf{m})$ or $\\tilde{\\mathcal{B}} = \\cap_{\\mathbf{m} \\in \\mathcal{V}} \\mathcal{B}(\\mathbf{m})$ over any subset of masks $\\mathcal{V} \\subset \\mathcal{M}$ .", + "bbox": [ + 511, + 401, + 908, + 494 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Concept preservation. Theorem 4.3 states that one can recover the concept block $\\left[\\mathbf{z}_{\\mathrm{I}}\\right]_{\\mathcal{B}(\\mathbf{m})}$ associated with each individual text caption in the dataset $\\mathcal{M}$ . Furthermore, it ensures that the union of concepts $\\left[\\mathbf{z}_{\\mathrm{I}}\\right]_{\\cup_{\\mathbf{m} \\in \\nu} \\mathcal{B}(\\mathbf{m})}$ from any subset of text captions $\\mathcal{V} \\subset \\mathcal{M}$ can be preserved. In the running example of Figure 1, our formulation allows us to preserve concepts (\"bear\", \"pen\", \"chair\") in the image representation by selectively matching them with the two captions, whereas existing models like CLIP may lose either \"pen\" and \"chair\" since they are only mentioned in one caption. Therefore, Theorem 4.3 effectively addresses Goal a.", + "bbox": [ + 511, + 502, + 905, + 667 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Concept disentanglement. The intersection operation in Theorem 4.3 empowers us to disentangle representations into potentially atomic concepts. In the example of Figure 1, we can identify the concept \"bear\" as the intersection of the two text captions, despite the absence of a standalone caption containing only \"bear\" in the dataset. Consequently, this part of the statement tackles Goal b. Our results underscore the importance of associating each image with a diverse set of captions that share overlapping concepts.", + "bbox": [ + 511, + 667, + 905, + 803 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Theoretical contribution. Theorem 4.3 extends existing theoretical frameworks [9, 43, 49]. Notably, Yao et al. [49] provide identification guarantees for shared representations over multiple views, generalizing earlier results confined to two views [9, 43]. This multi-view formulation", + "bbox": [ + 511, + 804, + 905, + 878 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "1We refer to a text caption with its mask m to simplify the notation.", + "bbox": [ + 529, + 886, + 890, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "29783", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "is analogous to our setup when considering each group of text captions $\\mathbf{t}$ associated with the same mask $\\mathbf{m}$ as a distinct view group. However, prior work [49] relies on explicit knowledge of these groupings to train view-specific encoders. In contrast, our problem setting presents a greater challenge since we have no access to this grouping information. Specifically, given two captions of any different images, it is unclear whether they stem from the same mask (i.e., the same view group). Therefore, the identification guarantees in prior studies do not apply to our setting. Theorem 4.3 demonstrates that our estimation model, paired with the learning objective (2), can automatically infer the necessary grouping information (i.e., the masks). By doing so, our approach relaxes the identification conditions in previous work, enabling effective representation identification without explicit group knowledge.", + "bbox": [ + 89, + 90, + 483, + 333 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. SmartCLIP: Modular Vision-language Alignment", + "text_level": 1, + "bbox": [ + 89, + 343, + 483, + 378 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Drawing on the theoretical framework in Section 4, we present SmartCLIP, a modular alignment model designed to achieve Goal a and Goal b. We discuss the implementation of the learning objective (2) and the model architecture.", + "bbox": [ + 89, + 388, + 483, + 449 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Modular alignment through adaptive masking. The masking function $\\hat{\\mathbf{m}} (\\cdot)$ is instrumental in our modular alignment objective (2). SmartCLIP implements this function with a transformer block that ingests a caption representation $\\hat{\\mathbf{z}}_{\\mathrm{T}}$ as its input and outputs a binary vector $\\hat{\\mathbf{m}} (\\hat{\\mathbf{z}}_{\\mathrm{T}})$ via a straight-through estimator [4].", + "bbox": [ + 89, + 450, + 483, + 541 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Modular contrast construction. As discussed in Section 4, the negative pairs in regular contrastive losses [8, 35] serves a similar role as the invertibility assumption (Condition 4.2-i) [43, 47]. We denote generic image, text representations as $\\mathbf{I}$ , $\\mathbf{T}$ , and positive, negative pairs with $\\mathbf{P}_{\\mathrm{pos}}$ , $\\mathbf{P}_{\\mathrm{neg}}$ respectively. The canonical one-side contrastive loss $\\mathcal{L}_{\\mathrm{ctr}}$ [37] is defined as:", + "bbox": [ + 89, + 542, + 483, + 648 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {\\mathrm {c t r}} \\left(\\underbrace {\\left(\\mathbf {I} ^ {(i)} , \\mathbf {T} ^ {(i)}\\right)} _ {\\mathbf {P} _ {\\text {p o s}}}, \\underbrace {\\left(\\mathbf {I} ^ {(i)} , \\mathbf {T} ^ {(j)}\\right)} _ {\\mathbf {P} _ {\\text {n e g}}} ^ {}\\right) \\tag {3} \\\\ = - \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\log \\frac {\\exp \\left(\\frac {\\mathrm {s i m} (\\mathbf {I} ^ {(i)} , \\mathbf {T} ^ {(i)})}{\\tau}\\right)}{\\sum_ {j = 1} ^ {N} \\exp \\left(\\frac {\\mathrm {s i m} (\\mathbf {I} ^ {(i)} , \\mathbf {T} ^ {(j)})}{\\tau}\\right)}, \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 145, + 660, + 483, + 765 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where we denote the sample size, temperature, and cosine similarity with $N$ , $\\tau$ , and $\\sin(\\cdot)$ respectively.", + "bbox": [ + 89, + 777, + 482, + 808 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Following the symmetric contrastive loss for CLIP [37], our alignment loss consists of two contrastive loss terms $\\mathcal{L}_{\\mathrm{ctrI}}$ and $\\mathcal{L}_{\\mathrm{ctrT}}$ that differ in the negative pairs:", + "bbox": [ + 89, + 809, + 483, + 854 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {c t r I}} := \\mathcal {L} _ {\\mathrm {c t r}} \\left(\\mathbf {P} _ {\\text {p o s}}, \\mathbf {P} _ {\\text {n e g I}}\\right), \\mathcal {L} _ {\\mathrm {c t r T}} := \\mathcal {L} _ {\\mathrm {c t r}} \\left(\\mathbf {P} _ {\\text {p o s}}, \\mathbf {P} _ {\\text {n e g T}}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 869, + 486, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "with the positive and negative pairs defined as follows:", + "bbox": [ + 513, + 90, + 870, + 106 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {P} _ {\\text {p o s}} := \\left(\\hat {\\mathbf {z}} _ {\\mathrm {I}} ^ {(i)} \\odot \\hat {\\mathbf {m}} \\left(\\hat {\\mathbf {z}} _ {\\mathrm {T}} ^ {(i)}\\right), \\hat {\\mathbf {z}} _ {\\mathrm {T}} ^ {(i)}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 599, + 118, + 903, + 143 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {P} _ {\\text {n e g I}} := \\left(\\hat {\\mathbf {z}} _ {\\mathrm {I}} ^ {(i)} \\odot \\hat {\\mathbf {m}} \\left(\\hat {\\mathbf {z}} _ {\\mathrm {T}} ^ {(j)}\\right), \\hat {\\mathbf {z}} _ {\\mathrm {T}} ^ {(j)}\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 596, + 146, + 903, + 171 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {P} _ {\\mathrm {n e g T}} := \\left(\\hat {\\mathbf {z}} _ {\\mathrm {I}} ^ {(j)} \\odot \\hat {\\mathbf {m}} \\left(\\hat {\\mathbf {z}} _ {\\mathrm {T}} ^ {(i)}\\right), \\hat {\\mathbf {z}} _ {\\mathrm {T}} ^ {(i)}\\right). \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 591, + 172, + 903, + 199 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In particular, $\\mathbf{P}_{\\mathrm{negI}}$ contrasts the image representation $\\hat{\\mathbf{z}}_{\\mathrm{I}}^{(i)}$ in the positive pair with randomly sampled caption representations $\\hat{\\mathbf{z}}_{\\mathrm{T}}^{(j)}$ (see the green region in Figure 3), whereas $\\mathbf{P}_{\\mathrm{negT}}$ contrasts the text representation $\\hat{\\mathbf{z}}_{\\mathrm{T}}^{(i)}$ in the positive pair with randomly sampled image representations $\\hat{\\mathbf{z}}_{\\mathrm{I}}^{(j)}$ (see the orange region in Figure 3).", + "bbox": [ + 511, + 213, + 903, + 313 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Sparsity penalty. We implement $L_{\\mathrm{sparsity}}$ in (2) with a $\\ell_1$ term for its compatibility with deep-learning training:", + "bbox": [ + 511, + 314, + 903, + 345 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {s p a r s i t y}} = \\left\\| \\hat {\\mathbf {m}} (\\mathbf {t}) \\right\\| _ {1}. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 633, + 357, + 903, + 376 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This term ensures that the textual concepts are encoded into a minimal number of latent dimensions, promoting the disentanglement of distinct concepts across text captions.", + "bbox": [ + 511, + 386, + 903, + 431 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "SmartCLIP training objective. In summary, the training objective of SmartCLIP is a weighted sum of loss terms in (4) and (8):", + "bbox": [ + 511, + 431, + 903, + 477 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\lambda_ {\\text {a l i g n}} \\cdot \\left(\\mathcal {L} _ {\\mathrm {c t r I}} + \\mathcal {L} _ {\\mathrm {c t r T}}\\right) + \\lambda_ {\\text {s p a r s i t y}} \\cdot \\mathcal {L} _ {\\text {s p a r s i t y}}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 537, + 491, + 903, + 508 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda_{\\mathrm{align}}$ and $\\lambda_{\\mathrm{sparsity}}$ denote the weighting coefficients.", + "bbox": [ + 513, + 520, + 895, + 536 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "6. Experiments", + "text_level": 1, + "bbox": [ + 513, + 550, + 645, + 566 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "6.1. Setup", + "text_level": 1, + "bbox": [ + 513, + 575, + 594, + 592 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation details. Following Long-CLIP [54], we finetune the CLIP model [37] on ShareGPT4V [7], which contains around 1 million image-text pairs. We employ the position encoding in long-CLIP to handle 248 tokens (c.f., the 77-token limit in the original CLIP). Compared to the baseline CLIP model, we introduce a mask network $\\hat{\\mathbf{m}}$ . The masking network is designed as a single transformer block, which takes the text sequence embedding $\\hat{\\mathbf{z}}_{\\mathrm{T}}$ from the text encoder. Then we add an attention-pooling layer to down-sample it to the same size as the CLIP representation, e.g., 768 in ViT-L/14. We tested including more transformer blocks in the mask network but did not observe significant improvements. Therefore, we stick to one block for faster training and inference. Unlike Long-CLIP [54] which processes all the captions for each image at each gradient step, we only sample one caption for each image, reducing our overall training time by half. Specifically, on 8 H100 GPUs, training a Vit-B/16 model for one epoch takes about 4 minutes with our model, whereas it takes around 7 minutes for Long-CLIP. After the pooling layer, we apply sigmoid to", + "bbox": [ + 511, + 598, + 906, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "29784", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/fe4d3b1c3b41015ef13fb0b62d8b3997d6d7437d6c55613d230d610306aa9b73.jpg", + "image_caption": [ + "Figure 3. The diagram of SmartCLIP. On the left, we introduce adaptive masking for alignment with different text prompts. The mask network selects which part of the image representation to be used. On the right, we present our modular contrastive objectives (4)." + ], + "image_footnote": [], + "bbox": [ + 186, + 95, + 807, + 277 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/1eac77474da52d16045be982501852d930124b9f333e9a5cc2f9e80cddac730f.jpg", + "table_caption": [ + "Table 1. Results of short-caption text-image retrieval on the 5k COCO2017 validation set and the whole 30k Flickr30K dataset." + ], + "table_footnote": [], + "table_body": "
COCOFlickr30k
Image-to-TextText-to-ImageImage-to-TextText-to-Image
R@1R@5R@10R@1R@5R@10R@1R@5R@10R@1R@5R@10
B/16CLIP51.876.884.332.757.768.244.168.277.024.745.154.6
Direct Fine-tuning37.462.372.121.843.454.525.745.855.417.934.543.1
Long-CLIP [54]57.681.187.840.465.875.246.871.479.834.156.365.7
SmartCLIP (Ours)61.983.389.742.468.277.855.678.285.036.358.867.8
L/14CLIP56.179.586.835.460.170.248.572.680.828.049.358.7
Direct Fine-tuning37.963.172.223.145.155.926.046.355.617.934.943.5
Long-CLIP [54]62.885.191.246.370.879.853.477.585.341.264.172.6
SmartCLIP (Ours)66.086.292.648.573.181.763.984.490.243.866.574.8
", + "bbox": [ + 120, + 362, + 870, + 534 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "restrict the output to the range $(0,1)$ and employ straight-through estimation [4] to binarize the outputs. The training batch size is 1024 and the learning rate is $10^{-6}$ for the CLIP component and $10^{-3}$ for the mask network.", + "bbox": [ + 89, + 544, + 483, + 604 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation. We evaluate the following datasets:", + "bbox": [ + 89, + 611, + 410, + 626 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Long text-to-image retrieval datasets: ShareGPT4V validation split [7] and Urban1k [54]. The captions for each image are long and describe details about the image. Both datasets contain 1000 text-to-image pairs.", + "- Short text-to-image retrieval datasets: COCO2017 validation split [27] and Flickr30K [50]. Following LongCLIP [54], we use 30K Flickr training dataset.", + "- Zero-shot image classification datasets. We use benchmark datasets: Country211, Fer2013, Fgvc-aircraft, GT-SRB, ImageNet, ImagetNet-V2, VOC2007, VOC2007-Multi, and SUN397." + ], + "bbox": [ + 89, + 633, + 482, + 799 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines. In this paper, we benchmark our approach against CLIP [37] and the recent state-of-the-art model long-CLIP [54].", + "bbox": [ + 89, + 806, + 483, + 851 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6.2. Comparison with CLIP Models", + "text_level": 1, + "bbox": [ + 511, + 544, + 790, + 559 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We present our experimental results across three key tasks: long text-to-image retrieval, short text-to-image retrieval, and zero-shot classification.", + "bbox": [ + 511, + 566, + 905, + 611 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Long text-to-image retrieval. Table 2 showcases our method's performance on long text-to-image retrieval tasks. SmartCLIP achieves substantial improvements over baseline models, particularly the SOTA Long-CLIP, which is designed to handle long text sequences. For example, on the Urban1k dataset, SmartCLIP elevates the performance from $78.9\\%$ to $90.0\\%$ , marking an impressive $14\\%$ boost.", + "bbox": [ + 511, + 612, + 905, + 717 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Short text-to-image retrieval. Similarly, as illustrated in Table 1, SmartCLIP significantly outperforms all baseline models across various metrics and datasets in short text-to-image retrieval tasks. The encouraging performance gains show that SmartCLIP can capture detailed information within images while simultaneously emphasizing the main semantic content.", + "bbox": [ + 511, + 719, + 905, + 823 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Zero-shot classification. To comprehensively evaluate our model's capabilities, we conduct zero-shot classification benchmarks in Table 3. Both the standard CLIP model and SmartCLIP demonstrate superior performance on different datasets. Notably, SmartCLIP exhibits a slight perfor", + "bbox": [ + 511, + 825, + 905, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "2https://github.com/LAION-AI/CLIP_benchmark", + "bbox": [ + 107, + 886, + 447, + 898 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "29785", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/d511d4a79744550756f60591cc41feb84a99cac6b58cb2af5df328c2c0783f5b.jpg", + "table_caption": [ + "Table 2. The R@1 of long-caption text-image retrieval on 1k ShareGPT4V [2] validation set and Urban-1000 dataset. The best results are bold. We cite the results from Long-CLIP [54]." + ], + "table_footnote": [], + "table_body": "
ShareGPT4VUrban1k
I2TT2II2TT2I
B/16CLIP [37]78.279.668.153.6
Direct Fine-tuning94.193.6--
Long-CLIP [54]94.693.378.979.5
SmartCLIP (Ours)98.798.190.087.4
L/14CLIP [37]81.884.068.752.8
Direct Fine-tuning95.395.4--
Long-CLIP [54]95.895.682.786.1
SmartCLIP (Ours)97.998.593.090.1
", + "bbox": [ + 91, + 156, + 470, + 313 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/b8cef987bc4a7fb274d6b20e1e6621f42e4fa0d246a9bf53cdbe7bce7a030460.jpg", + "table_caption": [ + "Table 3. Zero-shot classification performance on ViT-L/14 models. When the class name is very short, i.e., a single word like ImageNet, CLIP model performs better. When the class name is a combination of several words, our method achieves better results, e.g., the road sign in GTSRB." + ], + "table_footnote": [], + "table_body": "
DatasetCLIPLongCLIPSmartCLIP
Country21131.828.126.9
Fer201349.057.858.6
Fgvc-aircraft31.730.630.4
GTSRB50.248.952.4
ImageNet75.372.972.5
ImageNet-V269.766.966.6
VOC200778.377.578.6
VOC2007-Multi79.082.183.7
SUN39767.572.572.1
", + "bbox": [ + 537, + 167, + 916, + 325 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "This playful food sculpture transforms cucumbers into a fearsome T-Rex dinosaur. The cucumbers form the main body, with whole cucumbers creating the legs and tail, while sliced cucumbers make up the creature's midsection. More cucumbers are cleverly cut to shape the dinosaur's head, and additional cucumbers are arranged to suggest muscular limbs. From its mouth erupts a dramatic spray of carrots, with finely julienned carrots creating the effect of fire. These bright orange carrots provide a stunning contrast against the green vegetables. Shredded carrots cascade downward like flames, while more carrots are delicately cut to create a flame-like texture. The carrots' vibrant color makes the dinosaur appear truly animated. Fresh celery leaves crown the creation, with celery fronds adding a decorative touch around the body. More celery leaves create a natural backdrop, while additional celery pieces add texture throughout. The celery's feathered appearance provides an artistic flourish to the overall design.", + "bbox": [ + 89, + 344, + 584, + 525 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/19d15dd85f69958962524602adb46aa04bd415504c9d516fe0e7aea6bcc54a5d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 617, + 342, + 805, + 462 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5ef18a3723b1f69c850e9f13cb09f3f85a2d2df335404ca1061eb8e074934bdd.jpg", + "image_caption": [ + "Figure 4. Example of Long-text-to-image generation. We replace the CLIP text encoder in SDXL with different finetuned CLIP models. Given a very long text, CLIP [37] truncates the input to 77 tokens, resulting in information loss in the image. Our model learns to generate details such as celery leaves on the back of the dinosaur while other models fail." + ], + "image_footnote": [], + "bbox": [ + 617, + 463, + 805, + 523 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "mance decline on ImageNet, which is expected since our model is fine-tuned on the ShareGPT4V dataset featuring long text captions, whereas ImageNet primarily consists of short, often single-word class names. However, Smart-CLIP excels on datasets with multi-word class names, such as the GTSRB dataset, where it achieves the best performance in accurately classifying road sign descriptions.", + "bbox": [ + 89, + 583, + 482, + 690 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6.3. Ablation Studies", + "text_level": 1, + "bbox": [ + 89, + 696, + 254, + 712 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We analyze the three key components in our model: the modular alignment module, the sparsity loss and the impact of caption diversity in the data.", + "bbox": [ + 89, + 719, + 482, + 763 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Modular alignment. After introducing the mask network, we replace the standard contrastive learning with our selective alignment module. As shown in Figure 5, this change significantly improves performance. When using standard contrastive learning instead of the modular contrastive module (indicated by the purple lines), performance drops sharply. This happens because the mask information allows the network to easily separate positive pairs from negative ones, making the negative samples less informa", + "bbox": [ + 89, + 763, + 482, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "tive. As a result, standard contrastive learning no longer effectively helps the model learn meaningful information.", + "bbox": [ + 511, + 583, + 906, + 613 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Alignment coefficient $\\lambda_{\\mathrm{align}}$ . We test the impact of the alignment coefficient $\\lambda_{\\mathrm{align}}$ . The results in the right panels of Figure 5 show that our method performs consistently well across a wide range of $\\lambda_{\\mathrm{align}}$ values, from 0.1 to 20. This indicates that our approach is robust and does not require precise tuning of $\\lambda_{\\mathrm{align}}$ to achieve good performance.", + "bbox": [ + 511, + 613, + 905, + 704 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Sparsity coefficient $\\lambda_{\\mathrm{sparsity}}$ . We also examine the sparsity coefficient $\\lambda_{\\mathrm{sparsity}}$ . The left panels of Figure 5 demonstrate that adding sparsity to the mask network improves performance. This supports our idea that promoting sparsity helps the model focus on the most relevant concepts, enhancing its ability to capture detailed information without being distracted by irrelevant details.", + "bbox": [ + 511, + 704, + 905, + 808 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Caption diversity. We evaluate our model's performance under varying caption diversity conditions using the COCO dataset [27]. As shown in Table 4, increasing the number of captions per image enhances performance on the Flickr30K dataset, though at the expense of degraded performance on long-text-to-image retrieval tasks. Further improvements", + "bbox": [ + 511, + 810, + 906, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "29786", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f549d8ce6b37c71190963451851c7ee1cbfbca0483d7093681ca3929ae1ec123.jpg", + "image_caption": [ + "Figure 5. Ablation Studies on two proposed modules: selective alignment and sparsity. The baseline w.o. Modular means that we replace our modular alignment module with standard contrastive learning alignment." + ], + "image_footnote": [], + "bbox": [ + 98, + 90, + 281, + 195 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/48471290c784947d1f3355796a27856dbad99276e3bbf63a360e72188f23cb1b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 300, + 90, + 483, + 194 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ba17102ae45429d0dd3f06743232dc50f81cb980b7674fb6b7b8081733727316.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 90, + 687, + 194 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/10e4160ab66829aef5af4b2aa5966d7f36642f4862260f0b30129b2799cc56b2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 707, + 90, + 890, + 194 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d07100a362ce9907c1c077baa57eb37d5dde5da2f153a0f1850b914a1dd7feca.jpg", + "image_caption": [ + "Figure 6. Visualization of learned representations. Given an image, we generate two captions (e.g., a zebra and a zebra and a deer) and compute cosine similarities with the same image embedding to perform binary classification for visualization using Score-CAM [45]. Compared to baseline methods, our CLIP representations are more atomic and capture differences more effectively." + ], + "image_footnote": [], + "bbox": [ + 102, + 229, + 472, + 388 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "are achieved when we combine our training dataset with COCO. These results highlight both the importance of caption diversity and our method's capability to effectively handle complex text-image datasets.", + "bbox": [ + 89, + 502, + 483, + 563 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6.4. Additional Results", + "text_level": 1, + "bbox": [ + 89, + 574, + 269, + 589 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Visualization. While our quantitative results demonstrate superior performance across various tasks, we also explore the qualitative aspects of our model by visualizing the learned representations. The visualization results are shown in Figure 6. We employ the ScoreCAM method [45] for this purpose. For each image, we generate two distinct captions, such as \"a zebra\" and \"a zebra and a deer\". We then compute the cosine similarity between the image embedding and each of the two text embeddings. These similarity scores serve as logits for a classification task, which are then input into the ScoreCAM algorithm. SmartCLIP successfully learns modular representations, accurately capturing the relevant differences between captions.", + "bbox": [ + 89, + 597, + 482, + 792 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Plug and play for text-to-image generation. One main advantage of SmartCLIP over other CLIP models trained from scratch is the low computational cost of finetuning. Additionally, our fine-tuned text encoder can replace the CLIP text encoders in large-scale models in a plug-and-play manner. Specifically, we substitute the text encoder in the SDXL [36] model with both Long-CLIP and SmartCLIP.", + "bbox": [ + 89, + 795, + 483, + 901 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/a23d3354d8b26f59d245032a28d867ad204cd939d2e7860efe6fd9fe50ae23ea.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
NumCapPerImg-COCOT2II2TLongT2ILongI2T
153.639.385.389.3
353.640.985.388.6
556.441.285.286.5
Ours-ShareGPT55.636.398.798.1
+COCO57.038.497.898.5
", + "bbox": [ + 531, + 224, + 883, + 309 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/181703a468cf8331400c96f84dc806f4494e8e263124242468d489dc436e40b9.jpg", + "table_caption": [ + "Table 4. Retrieval results on Flickr30K [50] and ShareGPT-val [7] with models trained with different caption counts per image (top) and mixing long and short text-image datasets (bottom).", + "Table 5. Long text to image generation performance. We use the long captions (usually around 200 tokens, beyond the hard constraint 77 of the original CLIP model) to generate images with SDXL model [36], then we compare the generated images against the real images in ShareGPT4V validation split." + ], + "table_footnote": [], + "table_body": "
MethodKID ↓Pr ↑Re ↑F1 ↑DINO-L ↑
LongCLIP1.050.2380.7680.3630.401
SmartCLIP1.020.2580.7910.3890.414
", + "bbox": [ + 516, + 445, + 900, + 494 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As illustrated in Figure 4, our text encoder demonstrates a superior understanding of the long text, generating detailed elements such as celery leaves in the background. Furthermore, Table 5 presents quantitative results on image generation from long captions in the ShareGPT4V validation split. Our method consistently achieves better performance across various metrics, showcasing its effectiveness in handling complex, long text inputs.", + "bbox": [ + 511, + 507, + 906, + 628 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Conclusion and Limitation", + "text_level": 1, + "bbox": [ + 511, + 647, + 764, + 662 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we address the information misalignment and representation entanglement issues in existing vision-language models (e.g., CLIP). We establish theoretical conditions for effectively connecting text representations to atomic-level visual features and propose SmartCLIP, a principled, refined vision-language model. Our experimental results validate both our theoretical results and the practical effectiveness of SmartCLIP in advancing multimodal learning. Limitation. As discussed earlier, Condition 4.2-ii could be violated for datasets in which a subset images are paired with a limited number of captions compared to others. In Section 4, we discuss practical strategies to mitigate such issues. Devising alternative theoretical conditions may provide additional insights into fully utilizing all pairing information, which we leave as future work.", + "bbox": [ + 511, + 674, + 908, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "29787", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 91, + 90, + 250, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We would like to acknowledge the support from NSF Award No. 2229881, AI Institute for Societal Decision Making (AI-SDM), the National Institutes of Health (NIH) under Contract R01HL159805, and grants from Quris AI, Florin Court Capital, and MBZUAI-WIS Joint Program. The work of L. Kong is supported in part by NSF DMS-2134080 through an award to Y. Chi.", + "bbox": [ + 89, + 114, + 483, + 220 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 234, + 187, + 251 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Kartik Ahuja, Divyat Mahajan, Yixin Wang, and Yoshua Bengio. Interventional causal representation learning. In International conference on machine learning, pages 372-407. PMLR, 2023. 3", + "[2] Alberto Baldrati, Lorenzo Agnolucci, Marco Bertini, and Alberto Del Bimbo. Zero-shot composed image retrieval with textual inversion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15338-15347, 2023. 1", + "[3] Yoshua Bengio, Réjean Ducharme, and Pascal Vincent. A neural probabilistic language model. Advances in neural information processing systems, 13, 2000. 3", + "[4] Yoshua Bengio, Nicholas Léonard, and Aaron Courville. Estimating or propagating gradients through stochastic neurons for conditional computation. arXiv preprint arXiv:1308.3432, 2013. 5, 6", + "[5] Simon Buchholz, Goutham Rajendran, Elan Rosenfeld, Bryon Aragam, Bernhard Scholkopf, and Pradeep Ravikumar. Learning linear causal representations from interventions under general nonlinear mixing. Advances in Neural Information Processing Systems, 36, 2024. 3", + "[6] Chang Che, Qunwei Lin, Xinyu Zhao, Jiaxin Huang, and Liqiang Yu. Enhancing multimodal understanding with clip-based image-to-text transformation. In Proceedings of the 2023 6th International Conference on Big Data Technologies, pages 414-418, 2023. 1", + "[7] Lin Chen, Jinsong Li, Xiaoyi Dong, Pan Zhang, Conghui He, Jiaqi Wang, Feng Zhao, and Dahua Lin. Sharegpt4v: Improving large multi-modal models with better captions. arXiv preprint arXiv:2311.12793, 2023. 2, 5, 6, 8", + "[8] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 1, 4, 5", + "[9] Imant Daunhawer, Alice Bizeul, Emanuele Palumbo, Alexander Marx, and Julia E Vogt. Identifiability results for multimodal contrastive learning. In The Eleventh International Conference on Learning Representations, 2023. 3, 4", + "[10] Lijie Fan, Dilip Krishnan, Phillip Isola, Dina Katabi, and Yonglong Tian. Improving clip training with language rewrites. Advances in Neural Information Processing Systems, 36, 2024. 2", + "[11] Luigi Gresele, Paul K Rubenstein, Arash Mehrjou, Francesco Locatello, and Bernhard Schölkopf. The incomplete rosetta stone problem: Identifiability results for multi-" + ], + "bbox": [ + 93, + 260, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "view nonlinearica.In Uncertainty in Artificial Intelligence, pages 217-227.PMLR,2020.3", + "[12] Weiquan Huang, Aooqi Wu, Yifan Yang, Xufang Luo, Yuqing Yang, Liang Hu, Qi Dai, Xiyang Dai, Dongdong Chen, Chong Luo, et al. Llm2clip: Powerful language model unlock richer visual representation. arXiv preprint arXiv:2411.04997, 2024. 2", + "[13] Aapo Hyvarinen and Hiroshi Morioka. Unsupervised feature extraction by time-contrastive learning and nonlinear ica, 2016. 4", + "[14] Aapo Hyvarinen, Hiroaki Sasaki, and Richard Turner. Nonlinear ica using auxiliary variables and generalized contrastive learning. In The 22nd International Conference on Artificial Intelligence and Statistics, pages 859-868. PMLR, 2019. 3", + "[15] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International conference on machine learning, pages 4904-4916. PMLR, 2021. 2", + "[16] Chen Ju, Tengda Han, Kunhao Zheng, Ya Zhang, and Weidi Xie. Prompting visual-language models for efficient video understanding. In European Conference on Computer Vision, pages 105-124. Springer, 2022. 1", + "[17] Ilyes Khemakhem, Diederik Kingma, Ricardo Monti, and Aapo Hyvarinen. Variational autoencoders and nonlinear ica: A unifying framework. In International Conference on Artificial Intelligence and Statistics, pages 2207-2217. PMLR, 2020. 3, 4", + "[18] Lingjing Kong, Shaoan Xie, Weiran Yao, Yujia Zheng, Guangyi Chen, Petar Stojanov, Victor Akinwande, and Kun Zhang. Partial disentanglement for domain adaptation. In Proceedings of the 39th International Conference on Machine Learning, pages 11455-11472. PMLR, 2022. 4", + "[19] Sébastien Lachapelle, Divyat Mahajan, Ioannis Mitliagkas, and Simon Lacoste-Julien. Additive decoders for latent variables identification and cartesian-product extrapolation, 2023. 4", + "[20] Zhengfeng Lai, Vasileios Saveris, Chen Chen, Hong-You Chen, Haotian Zhang, Bowen Zhang, Juan Lao Tebar, Wenze Hu, Zhe Gan, Peter Grasch, et al. Revisit large-scale image-caption data in pre-training multimodal foundation models. arXiv preprint arXiv:2410.02740, 2024. 2", + "[21] Zhengfeng Lai, Haotian Zhang, Wentao Wu, Haoping Bai, Aleksei Timofeev, Xianzhi Du, Zhe Gan, Jiulong Shan, Chen-Nee Chuah, Yinfei Yang, et al. Veclip: Improving clip training via visual-enriched captions. ECCV. IEEE, page 13, 2024. 2, 4", + "[22] Samuel Lavoie, Polina Kirichenko, Mark Ibrahim, Mahmoud Assran, Andrew Gordon Wilson, Aaron Courville, and Nicolas Ballas. Modeling caption diversity in contrastive vision-language pretraining. arXiv preprint arXiv:2405.00740, 2024. 2", + "[23] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In Interna-" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "29788", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "tional conference on machine learning, pages 12888-12900. \nPMLR, 2022. 2, 4", + "[24] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730-19742. PMLR, 2023. 2", + "[25] Xianhang Li, Haoqin Tu, Mude Hui, Zeyu Wang, Bingchen Zhao, Junfei Xiao, Sucheng Ren, Jieru Mei, Qing Liu, Huangjie Zheng, et al. What if we recapture billions of web images with llama-3? arXiv preprint arXiv:2406.08478, 2024. 2, 4", + "[26] Yangguang Li, Feng Liang, Lichen Zhao, Yufeng Cui, Wanli Ouyang, Jing Shao, Fengwei Yu, and Junjie Yan. Supervision exists everywhere: A data efficient contrastive language-image pre-training paradigm. arXiv preprint arXiv:2110.05208, 2021. 2", + "[27] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 6, 7", + "[28] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 1", + "[29] AI Meta. Introducing meta llama 3: The most capable openly available llm to date. Meta AI, 2024. 2", + "[30] Tomas Mikolov, Ilya Sutskever, Kai Chen, Greg S Corrado, and Jeff Dean. Distributed representations of words and phrases and their compositionality. Advances in neural information processing systems, 26, 2013. 3", + "[31] Hiroshi Morioka and Aapo Hyvarinen. Connectivity-contrastive learning: Combining causal discovery and representation learning for multimodal data. In International conference on artificial intelligence and statistics, pages 3399–3426. PMLR, 2023. 3", + "[32] Hiroshi Morioka and Aapo Hyvarinen. Causal representation learning made identifiable by grouping of observational variables. In *Forty-first International Conference on Machine Learning*, 2024. 3", + "[33] Norman Mu, Alexander Kirillov, David Wagner, and Saining Xie. Slip: Self-supervision meets language-image pretraining. In European conference on computer vision, pages 529-544. Springer, 2022. 2", + "[34] Ivona Najdenkoska, Mohammad Mahdi Derakhshani, Yuki M Asano, Nanne van Noord, Marcel Worring, and Cees GM Snoek. Tulip: Token-length upgraded clip. arXiv preprint arXiv:2410.10034, 2024. 2", + "[35] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 1, 4, 5", + "[36] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 8" + ], + "bbox": [ + 91, + 90, + 482, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[37] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 2, 4, 5, 6, 7", + "[38] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 1", + "[39] Hanoona Rasheed, Muhammad Uzair Khattak, Muhammad Maaz, Salman Khan, and Fahad Shahbaz Khan. Fine-tuned clip models are efficient video learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6545-6554, 2023. 1", + "[40] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1", + "[41] Nils Sturma, Chandler Squires, Mathias Drton, and Caroline Uhler. Unpaired multi-domain causal representation learning, 2023. 3", + "[42] Mingkang Tang, Zhanyu Wang, Zhenhua Liu, Fengyun Rao, Dian Li, and Xiu Li. Clip4caption: Clip for video caption. In Proceedings of the 29th ACM International Conference on Multimedia, pages 4858-4862, 2021. 1", + "[43] Julius von Kugelgen, Yash Sharma, Luigi Gresele, Wieland Brendel, Bernhard Scholkopf, Michel Besserve, and Francesco Locatello. Self-supervised learning with data augmentations provably isolates content from style. arXiv preprint arXiv:2106.04619, 2021. 3, 4, 5", + "[44] Julius von Kugelgen, Michel Besserve, Liang Wendong, Luigi Gresele, Armin Kekic, Elias Bareinboim, David Blei, and Bernhard Scholkopf. Nonparametric identifiability of causal representations from unknown interventions. Advances in Neural Information Processing Systems, 36, 2024. 3", + "[45] Haofan Wang, Zifan Wang, Mengnan Du, Fan Yang, Zijian Zhang, Sirui Ding, Piotr Mardziel, and Xia Hu. Score-cam: Score-weighted visual explanations for convolutional neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 24-25, 2020. 8", + "[46] Jianyi Wang, Kelvin CK Chan, and Chen Change Loy. Exploring clip for assessing the look and feel of images. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2555-2563, 2023. 1", + "[47] Tongzhou Wang and Phillip Isola. Understanding contrastive representation learning through alignment and uniformity on the hypersphere. In International conference on machine learning, pages 9929-9939. PMLR, 2020. 4, 5", + "[48] Wei Wu, Kecheng Zheng, Shuailei Ma, Fan Lu, Yuxin Guo, Yifei Zhang, Wei Chen, Qingpei Guo, Yujun Shen, and Zheng-Jun Zha. Lotlip: Improving language-image pre-training for long text understanding. arXiv preprint arXiv:2410.05249, 2024. 2" + ], + "bbox": [ + 516, + 90, + 906, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "29789", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[49] Dingling Yao, Danru Xu, Sebastien Lachapelle, Sara Maglia-cane, Perouz Taslakian, Georg Martius, Julius von Kügelgen, and Francesco Locatello. Multi-view causal representation learning with partial observability. In The Twelfth International Conference on Learning Representations, 2024. 3, 4, 5", + "[50] Peter Young, Alice Lai, Micah Hodosh, and Julia Hockenmaier. From image descriptions to visual denotations: New similarity metrics for semantic inference over event descriptions. Transactions of the Association for Computational Linguistics, 2:67-78, 2014. 6, 8", + "[51] Jiahui Yu, Zirui Wang, Vijay Vasudevan, Legg Yeung, Mojtaba Seyedhosseini, and Yonghui Wu. Coca: Contrastive captioners are image-text foundation models. arXiv preprint arXiv:2205.01917, 2022. 2", + "[52] Xiaohua Zhai, Xiao Wang, Basil Mustafa, Andreas Steiner, Daniel Keysers, Alexander Kolesnikov, and Lucas Beyer. Lit: Zero-shot transfer with locked-image text tuning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18123-18133, 2022. 2", + "[53] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11975-11986, 2023. 2", + "[54] Beichen Zhang, Pan Zhang, Xiaoyi Dong, Yuhang Zang, and Jiaqi Wang. Long-clip: Unlocking the long-text capability of clip. arXiv preprint arXiv:2403.15378, 2024. 2, 4, 5, 6, 7", + "[55] Jiaqi Zhang, Kristjan Greenewald, Chandler Squires, Akash Srivastava, Karthikeyan Shanmugam, and Caroline Uhler. Identifiability guarantees for causal disentanglement from soft interventions. Advances in Neural Information Processing Systems, 36, 2024. 3", + "[56] Jihai Zhang, Xiaoye Qu, Tong Zhu, and Yu Cheng. Clipmoe: Towards building mixture of experts for clip with diversified multiplet upcycling. arXiv preprint arXiv:2409.19291, 2024. 2", + "[57] Kun Zhang, Shaoan Xie, Ignavier Ng, and Yujia Zheng. Causal representation learning from multiple distributions: A general setting. arXiv preprint arXiv:2402.05052, 2024. 3", + "[58] Renrui Zhang, Ziyu Guo, Wei Zhang, Kunchang Li, Xupeng Miao, Bin Cui, Yu Qiao, Peng Gao, and Hongsheng Li. Pointclip: Point cloud understanding by clip. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8552-8562, 2022. 1", + "[59] Kecheng Zheng, Yifei Zhang, Wei Wu, Fan Lu, Shuailei Ma, Xin Jin, Wei Chen, and Yujun Shen. Dreamclip: Language-image pre-training with long captions. arXiv preprint arXiv:2403.17007, 2024. 2" + ], + "bbox": [ + 91, + 92, + 480, + 768 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "29790", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/SmartCLIP_ Modular Vision-language Alignment with Identification Guarantees/28f50741-407e-4ef2-b9fa-f2bf47df9fd3_model.json b/2025/SmartCLIP_ Modular Vision-language Alignment with Identification Guarantees/28f50741-407e-4ef2-b9fa-f2bf47df9fd3_model.json new file mode 100644 index 0000000000000000000000000000000000000000..dbd13ee265b6966e83758e248fb0ba9ea8ff0d7b --- /dev/null +++ b/2025/SmartCLIP_ Modular Vision-language Alignment with Identification Guarantees/28f50741-407e-4ef2-b9fa-f2bf47df9fd3_model.json @@ -0,0 +1,2532 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.131, + 0.908, + 0.155 + ], + "angle": 0, + "content": "SmartCLIP: Modular Vision-language Alignment with Identification Guarantees" + }, + { + "type": "text", + "bbox": [ + 0.247, + 0.18, + 0.751, + 0.218 + ], + "angle": 0, + "content": "Shaoan Xie\\*, Lingjing\\*, Yujia Zheng\\*, Yu Yao\\*, Zeyu Tang\\*, Eric P. Xing\\*, Guangyi Chen\\*, Kun Zhang\\*," + }, + { + "type": "text", + "bbox": [ + 0.375, + 0.219, + 0.623, + 0.234 + ], + "angle": 0, + "content": "1 Carnegie Mellon University" + }, + { + "type": "text", + "bbox": [ + 0.246, + 0.236, + 0.751, + 0.251 + ], + "angle": 0, + "content": "\\(^{2}\\) Mohamed bin Zayed University of Artificial Intelligence" + }, + { + "type": "text", + "bbox": [ + 0.401, + 0.253, + 0.597, + 0.269 + ], + "angle": 0, + "content": "3 University of Sydney" + }, + { + "type": "text", + "bbox": [ + 0.443, + 0.274, + 0.555, + 0.286 + ], + "angle": 0, + "content": "* Equal contribution" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.322, + 0.327, + 0.339 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.355, + 0.486, + 0.583 + ], + "angle": 0, + "content": "Contrastive Language-Image Pre-training (CLIP) [37] has emerged as a pivotal model in computer vision and multimodal learning, achieving state-of-the-art performance at aligning visual and textual representations through contrastive learning. However, CLIP struggles with potential information misalignment in many image-text datasets and suffers from entangled representation. On the one hand, short captions for a single image in datasets like MSCOCO may describe disjoint regions in the image, leaving the model uncertain about which visual features to retain or disregard. On the other hand, directly aligning long captions with images can lead to the retention of entangled details, preventing the model from learning disentangled, atomic concepts - ultimately limiting its generalization on certain downstream tasks involving short prompts." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.584, + 0.485, + 0.794 + ], + "angle": 0, + "content": "In this paper, we establish theoretical conditions that enable flexible alignment between textual and visual representations across varying levels of granularity. Specifically, our framework ensures that a model can not only preserve cross-modal semantic information in its entirety but also disentangle visual representations to capture fine-grained textual concepts. Building on this foundation, we introduce SmartCLIP, a novel approach that identifies and aligns the most relevant visual and textual representations in a modular manner. Superior performance across various tasks demonstrates its capability to handle information misalignment and supports our identification theory. The code is available at https://github.com/Mid-Push/SmartCLIP." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.828, + 0.222, + 0.844 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.855, + 0.485, + 0.903 + ], + "angle": 0, + "content": "Contrastive Language-Image Pre-training (CLIP) [37] has been the cornerstone for many computer vision and machine learning tasks, such as text-to-image retrieval [2], im" + }, + { + "type": "image", + "bbox": [ + 0.526, + 0.327, + 0.665, + 0.412 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.675, + 0.332, + 0.716, + 0.343 + ], + "angle": 0, + "content": "COCO:" + }, + { + "type": "text", + "bbox": [ + 0.675, + 0.344, + 0.847, + 0.354 + ], + "angle": 0, + "content": "A very cute teddy bear holding a pen." + }, + { + "type": "text", + "bbox": [ + 0.676, + 0.355, + 0.859, + 0.364 + ], + "angle": 0, + "content": "A stuffed bear that is sitting in a chair" + }, + { + "type": "text", + "bbox": [ + 0.676, + 0.365, + 0.888, + 0.385 + ], + "angle": 0, + "content": "A brown bear wearing a sweater next to a pen and paper." + }, + { + "type": "text", + "bbox": [ + 0.676, + 0.386, + 0.884, + 0.405 + ], + "angle": 0, + "content": "A brown teddy bear in a dark sweater sits on a black striped chair while holding a pen." + }, + { + "type": "title", + "bbox": [ + 0.535, + 0.417, + 0.618, + 0.428 + ], + "angle": 0, + "content": "ShareGPT4u:" + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.43, + 0.888, + 0.499 + ], + "angle": 0, + "content": "In the center of the image, a light brown teddy bear is seated on a black chair with white stripes. The teddy bear, dressed in a green sweater adorned with a yellow flower, holds a blue pen in its paw, as if ready to sign a document. The chair is positioned on a gray carpeted floor. To the right of the chair, there's a stack of papers neatly arranged on the floor. The scene creates an atmosphere of anticipation, as if the teddy bear is about to make a significant decision." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.517, + 0.907, + 0.656 + ], + "angle": 0, + "content": "Figure 1. Depiction of two primary challenges for CLIP. (1) Information Misalignment: An image can be paired with multiple captions that describe disparate aspects - the first caption contains concepts \"bear\" and \"pen\" whereas the second only mentions \"bear\" and \"paper\". Aligning the image with both captions leads to the loss of key concepts \"pen\" and \"paper\" not shared across the captions. (2) Entangled Representations: Long, detailed captions involving multiple concepts (e.g., \"chair\", \"pen\", \"flower\", \"floor\") encourage the model to form entangled representations, hindering independent understanding of each individual concept." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.671, + 0.909, + 0.822 + ], + "angle": 0, + "content": "age and video understanding [6, 16, 39, 42, 46, 58], and generative models [28, 38, 40]. It aligns the representations from different modalities with a contrastive learning loss [8, 35]. Specifically, each image-caption pair in the dataset is treated as a positive pair, while negative pairs are created by matching images with captions randomly drawn from the dataset. The image and text encoders are trained with a symmetric cross-entropy loss that draws the image and text representations in each positive pair together while pulling the negative pairs' representations apart." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Training CLIP requires vast amounts of image-text pairs, making it challenging to maintain dataset quality at such a large scale. In particular, the quality of text captions has been a key concern, prompting the development of various methods to enhance both their diversity and accu" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "29780" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.304 + ], + "angle": 0, + "content": "racy. ALIGN [15] shows that scaling up the dataset size can compensate for the noisy text supervision. BLIP models [23, 24] improve captions by incorporating additional captioning and filtering mechanisms. VE-CLIP [21] introduces a visual-enriched captioning approach to further refine caption quality. Similarly, LaCLIP [10] leverages language models to rewrite captions, while RecapCLIP [25] uses LLaMA-3 [29] to generate captions for 1.3 billion images. Despite these efforts, recent findings reveal that longer and seemingly higher-quality captions do not necessarily yield improved performance on many downstream tasks [20]. Li et al. [25] find that the text-to-image retrieval accuracy on Flickr30K drops from 84.2 to 74.1 when replacing the original captions with longer captions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.305, + 0.483, + 0.502 + ], + "angle": 0, + "content": "A key issue contributing to the observed performance degradation is the information misalignment between images and their captions, a problem that becomes more pronounced when multiple captions are paired with a single image. On one hand, an image may be paired with several captions, each capturing only a partial aspect of the image. In Figure 1, aligning the image with the first caption, \"A very cute teddy bear holding a pen\", risks forcing the model to discard other important concepts like \"chair\" and \"paper\", which are required to align with the second and third captions in the pink text box. This misalignment between image and text introduces conflicts during standard CLIP training, leading to the loss of key visual concepts." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.505, + 0.483, + 0.716 + ], + "angle": 0, + "content": "On the other hand, training CLIP with long, detailed captions, as seen in recent approaches [10, 21, 25], encourages the model to learn entangled representations of multiple concepts bundled together in a single caption. Thus, it remains challenging to explicitly extract object/concept-centric representations from CLIP's visual representation. This entanglement is particularly problematic for tasks that require individual, atomic concepts or novel combinations of them, as empirically observed on short-text-to-image retrieval tasks [54]. In Figure 1, the long caption generated by ShareGPT4V [7] contains an exhaustive set of concepts such as \"chair\", \"pen\", \"flower\", and \"floor\". This aggregation can hinder the model's performance on tasks that demand the understanding of each concept individually." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.718, + 0.483, + 0.868 + ], + "angle": 0, + "content": "In this paper, we propose a refined approach to representation alignment in vision-language models like CLIP [37]. We frame the alignment challenge as a latent-variable identification problem and develop theoretical conditions that enable flexible alignment between textual and visual representations at different levels of granularity. Our framework enables the model to preserve the complete cross-modal information while also disentangling representations to capture fine-grained concepts, effectively addressing the misalignment and disentanglement issues discussed earlier." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Building on these theoretical insights, we introduce SmartCLIP, a novel method that identifies and aligns vi" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.302 + ], + "angle": 0, + "content": "sual and textual concepts in a modular manner. Specifically, we design a mask network that selects a subset of dimensions from the full representations, corresponding to only the concepts present in each specific caption. This allows the model to perform text-image alignment over the most relevant concepts modules, rather than the entire representation. We empirically demonstrate that SmartCLIP outperforms state-of-the-art models across a range of downstream tasks, showcasing its effectiveness in addressing alignment challenges. In particular, SmartCLIP significantly improves retrieval performance across text lengths, achieving \\(98.7\\%\\) accuracy (up from \\(78.2\\%\\)) on the ShareGPT4V long text-to-image retrieval tasks, while boosting short text-to-image retrieval R1 from \\(56.1\\%\\) to \\(66.0\\%\\)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.303, + 0.854, + 0.317 + ], + "angle": 0, + "content": "Our main contributions are summarized as follows." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.319, + 0.905, + 0.394 + ], + "angle": 0, + "content": "i We identify critical issues of information misalignment and entangled representations within the CLIP framework. To overcome these challenges, we propose a latent-variable formulation and establish theoretical conditions that guarantee the recovery of the latent variables." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.395, + 0.905, + 0.455 + ], + "angle": 0, + "content": "ii Building upon our theoretical findings, we propose SmartCLIP, featuring adaptive masking and a modular contrastive learning objective that facilitates the learning of disentangled, modular representations." + }, + { + "type": "text", + "bbox": [ + 0.509, + 0.455, + 0.905, + 0.545 + ], + "angle": 0, + "content": "iii We perform extensive experiments on a variety of tasks, including long and short text-to-image retrieval, zero-shot classification, and text-to-image generation. Smart-CLIP consistently outperforms or matches state-of-the-art models across these benchmarks, demonstrating its efficacy and validating our theoretical contributions." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.559, + 0.655, + 0.573 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.584, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Vision-language models. The breakthrough of CLIP [37] has attracted significant attention from the community. SLIP [33] and DECLIP [26] propose to incorporate self-supervised learning techniques to improve the learned representation. Coca [51] introduces a decoder in addition to the contrastive learning branch. LiT [52] locks the image encoder and only finetunes the text encoder. SigLIP [53] adopts a simple sigmoid loss to handle large training batch sizes. LoTLIP [48] inserts corner tokens after the classification token to support long-text understanding. TULIP [34] replaces the absolute position embedding with the relative position embedding to support longer text understanding. ALIGN [15] demonstrates that increasing dataset size can mitigate the impact of noisy text supervision. Recent methods have been focusing on generating better captions [20, 21, 23-25, 59]. CLIP-MOE [56] introduces mixture-of-experts to CLIP. LLM2CLIP [12] augments CLIP with large language models. LongCLIP [54] extends the token constraint of CLIP from 77 to 248 and applies PCA to perform short text-to-image contrastive learning to preserve its short text capability. Llip [22] learns a text-dependent vi" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "29781" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.152 + ], + "angle": 0, + "content": "sual representation by mixing a set of learnable tokens with a cross-attention module. In contrast, SmartCLIP directly learns a single global representation that encodes all disentangled, interpretable concepts through masking." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.152, + 0.485, + 0.681 + ], + "angle": 0, + "content": "Latent variable identification. Learning high-level, semantic information from low-level observational data (e.g., images and text) can often be formulated as latent-variable identification problems. Though appealing, such tasks are accompanied by substantial difficulties, especially for complex real-world data distributions involving nonlinear generating functions. Recently, a wealth of papers [1, 5, 9, 11, 14, 17, 31, 32, 41, 44, 49, 55, 57] propose to overcome such obstacles by leveraging auxiliary information, such as temporal information, multiple domains, and multiple views/modalities. Especially related to our work are those that tap into the paired multi-view data to identify the shared information across available views [9, 11, 31, 32, 43, 49]. Recent work [11, 31, 32] relies on specific forms of latent variable distributions (e.g., independence or exponential family). These constraints restrict their applicability for distributions entailing complex interactions among latent variables. Prior work [9, 43] adopt more flexible assumptions on the underlying distribution and identify blocks of latent variables directly shared by two views arising from data augmentations, which is extended to a multi-view setting [49]. The problem under investigation can be viewed as a form of this multi-view setting where the paired image and text captions are considered as views sharing the semantic latent variable. Existing works [9, 43, 49] assume that views are grouped over all the data pairs and this grouping information is known so that one can learn a designated encoder for each view group. However, this view grouping information is inaccessible for our problems – for any two text captions of different images, we cannot judge whether they belong to the same view group. In our theory section, we show that by properly utilizing the data-generating process, we can learn such information directly and further achieve the desired identification results, thus generalizing existing multi-view latent variable identification results." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.695, + 0.296, + 0.71 + ], + "angle": 0, + "content": "3. Problem Formulation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.483, + 0.81 + ], + "angle": 0, + "content": "As motivated previously, we aim to 1) preserve all the semantic information shared across modalities, and 2) learn a disentangled representation that corresponds to textual concepts at diverse granularity levels. To this end, we propose the following data-generating process underlying the visual-language data distribution." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.483, + 0.887 + ], + "angle": 0, + "content": "Notations. We indicate the dimensionality of a vector with \\(d(\\cdot)\\). We denote a subset of dimensions of a vector \\(\\mathbf{z}\\) with \\([\\mathbf{z}]_{\\mathcal{B}}\\) with the index set \\(\\mathcal{B}\\). We define the set of indices whose corresponding values are nonzero in a vector \\(\\mathbf{m}\\) with \\(\\mathcal{B}(\\mathbf{m}) := \\{i \\in d(\\mathbf{m}) : [\\mathbf{m}]_i \\neq 0\\}\\)." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.887, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Data-generating processes. We depict the data-generating" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.715, + 0.107 + ], + "angle": 0, + "content": "process in Figure 2 and in (1)." + }, + { + "type": "equation", + "bbox": [ + 0.538, + 0.12, + 0.905, + 0.135 + ], + "angle": 0, + "content": "\\[\n\\mathbf {z} _ {\\mathrm {T}} := \\mathbf {z} _ {\\mathrm {I}} \\odot \\mathbf {m}; \\mathbf {i} := g _ {\\mathrm {I}} \\left(\\mathbf {z} _ {\\mathrm {I}}, \\epsilon_ {\\mathrm {I}}\\right); \\mathbf {t} := g _ {\\mathrm {T}} \\left(\\mathbf {z} _ {\\mathrm {T}}, \\epsilon_ {\\mathrm {T}}\\right). \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.148, + 0.905, + 0.269 + ], + "angle": 0, + "content": "We assume that each pair of image \\(\\mathbf{i} \\in \\mathcal{I} \\subset \\mathbb{R}^{d(\\mathbf{i})}\\) and text caption \\(\\mathbf{t} \\in \\mathcal{T} \\subset \\mathbb{R}^{d(\\mathbf{t})}\\) originate from semantic information \\(\\mathbf{z}_{\\mathrm{I}} \\in \\mathcal{Z}_{\\mathrm{I}} \\subset \\mathbb{R}^{d(\\mathbf{z}_{\\mathrm{I}})}\\), together with modality-specific variations \\(\\epsilon_{\\mathrm{I}}\\) and \\(\\epsilon_{\\mathrm{T}}\\) (e.g., illumination for images, tenses for text), through generating functions \\(g_{\\mathrm{I}}: (\\mathbf{z}_{\\mathrm{I}}, \\epsilon_{\\mathrm{I}}) \\mapsto \\mathbf{i}\\) and \\(g_{\\mathrm{T}}: (\\mathbf{z}_{\\mathrm{T}}, \\epsilon_{\\mathrm{T}}) \\mapsto \\mathbf{t}\\) respectively. We treat the text caption as continuous variables, as each word can be represented with a continuous word embedding vector in practice [3, 30]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.271, + 0.907, + 0.361 + ], + "angle": 0, + "content": "As demonstrated in Figure 1, text captions of the same image often convey partial information of the entire image semantics. Thus, we associate each text caption's representation \\(\\mathbf{z}_{\\mathrm{T}} := \\mathbf{m} \\odot \\mathbf{z}_{\\mathrm{I}}\\) with a binary random mask \\(\\mathbf{m} \\in \\mathcal{M} \\subset \\{0,1\\}^{d(\\mathbf{z}_{\\mathrm{I}})}\\) that eliminates information absent in the specific caption \\(\\mathbf{t}\\)." + }, + { + "type": "image", + "bbox": [ + 0.565, + 0.373, + 0.857, + 0.492 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.502, + 0.906, + 0.552 + ], + "angle": 0, + "content": "Figure 2. The data-generating process. The text representation \\(\\mathbf{z}_{\\mathrm{T}}\\) contains partial information of the vision representation \\(\\mathbf{z}_{\\mathrm{I}}\\), as indicated by masking \\(\\mathbf{m}\\). Modal-specific information is represented as \\(\\epsilon_{\\mathrm{I}}\\) and \\(\\epsilon_{\\mathrm{T}}\\). Dashed edges indicate potential statistical dependence." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.568, + 0.854, + 0.582 + ], + "angle": 0, + "content": "Goals. Our two goals can be formalized as follows." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.584, + 0.905, + 0.613 + ], + "angle": 0, + "content": "a Preserving cross-modal information: identifying the complete latent representation \\(\\mathbf{z}_1\\)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.614, + 0.905, + 0.658 + ], + "angle": 0, + "content": "b Disentangling concepts: identifying concepts \\(\\mathbf{z}_{\\mathrm{T}}\\) associated with a given textual description \\(\\mathbf{t}\\) at different granularity levels potentially unseen during training." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.584, + 0.905, + 0.658 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.66, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Examples. In Figure 1, the image i contains concepts \"bear\", \"chair\", and \"pen\", which we assume correspond to three components in the representation \\(\\mathbf{z}_{\\mathrm{I}}\\), say \\([\\mathbf{z}_{\\mathrm{I}}]_{1}\\), \\([\\mathbf{z}_{\\mathrm{I}}]_{2}\\), and \\([\\mathbf{z}_{\\mathrm{I}}]_{3}\\). The first two COCO captions \\(\\mathbf{t}^{(1)}\\) and \\(\\mathbf{t}^{(2)}\\) mention only a subset of these concepts, i.e., (\"bear\", \"pen\") and (\"bear\", \"chair\") separately. Thus, the masks for these captions are \\(\\mathbf{m}^{(1)} = [1,0,1]\\) and \\(\\mathbf{m}^{(2)} = [1,1,0]\\). The variables \\(\\epsilon_{\\mathrm{I}}\\) and \\(\\epsilon_{\\mathrm{T}}\\) represent modality-specific nuance factors such as illumination conditions in the image i and syntax in the text t. For Goal a, we seek to preserve the complete information \\(\\mathbf{z}_{\\mathrm{I}}\\). This involves retaining all relevant textual concepts present in the captions, namely \"bear\", \"chair\", and \"pen\" from captions \\(\\mathbf{t}^{(1)}\\) and \\(\\mathbf{t}^{(2)}\\). For Goal b, we intend to disentangle the representation \\(\\mathbf{z}_{\\mathrm{I}}\\) into finer concept blocks potentially unseen in the training. This includes identifying individual concepts such as \"bear\" in the dimen-" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "29782" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.091, + 0.486, + 0.123 + ], + "angle": 0, + "content": "sion \\([\\mathbf{z}_{\\mathrm{I}}]_{1}\\), even if the training captions only include \"bear\" in combination with other concepts." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.136, + 0.295, + 0.154 + ], + "angle": 0, + "content": "4. Identification Theory" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.162, + 0.485, + 0.299 + ], + "angle": 0, + "content": "In this section, we present the theoretical results towards Goal a and Goal b. We show that under a suitable learning objective (2), the learned variables \\((\\hat{\\mathbf{z}}_{\\mathrm{I}},\\hat{\\mathbf{z}}_{\\mathrm{T}})\\) can be identified with the corresponding true quantities \\((\\mathbf{z}_{\\mathrm{I}},\\mathbf{z}_{\\mathrm{T}})\\) up to certain equivalent classes. In particular, we resort to the blockwise identifiability [18, 19, 43, 49] throughout this work. This suffices for our purpose since often several dimensions jointly (i.e., a block) comprise a meaningful concept while a single dimension may not be interpretable." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.313, + 0.485, + 0.359 + ], + "angle": 0, + "content": "Definition 4.1 (Block-wise Identifiability). The true variable \\(\\mathbf{v}\\) is block-wise identifiable if it is related to its estimate \\(\\hat{\\mathbf{v}}\\) through an invertible map \\(\\mathbf{v} \\mapsto \\hat{\\mathbf{v}}\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.369, + 0.485, + 0.431 + ], + "angle": 0, + "content": "The learning objective. Our estimation model consists of vision/text encoders \\((f_{\\mathrm{I}}, f_{\\mathrm{T}})\\) (smooth, invertible functions), and a masking function \\(\\hat{\\mathbf{m}}: \\mathcal{T} \\to \\mathcal{M}\\) that estimates the true mask \\(\\mathbf{m}\\) underlying a given text caption \\(\\mathbf{t}\\)." + }, + { + "type": "equation", + "bbox": [ + 0.15, + 0.441, + 0.362, + 0.486 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\arg \\min \\| \\hat {\\mathbf {m}} (\\mathbf {t}) \\| _ {0}, \\quad \\text {s u b j e c t} \\\\ f _ {\\mathrm {I}}, f _ {\\mathrm {T}}, \\hat {\\mathbf {m}} \\\\ L _ {\\mathrm {s p a r s i t y}} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.46, + 0.482, + 0.482, + 0.495 + ], + "angle": 0, + "content": "(2)" + }, + { + "type": "equation", + "bbox": [ + 0.151, + 0.49, + 0.425, + 0.535 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\arg \\min \\| f _ {\\mathrm {I}} (\\mathbf {i}) \\odot \\hat {\\mathbf {m}} (\\mathbf {t}) - f _ {\\mathrm {T}} (\\mathbf {t}) \\|, \\forall (\\mathbf {i}, \\mathbf {t}). \\\\ \\underbrace {f _ {\\mathrm {I}} , f _ {\\mathrm {T}} , \\hat {\\mathbf {m}}} _ {L _ {\\mathrm {a l i g n}}} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.547, + 0.484, + 0.698 + ], + "angle": 0, + "content": "Our learning objective (2) consists of an alignment term \\( L_{\\mathrm{align}} \\) that draws the positive pairs across modalities. The negative pairs in regular contrastive losses [8, 35, 37] can be implemented through an entropy term at the sample limit [47]. This serves the same role as the invertibility condition on the encoder models [43], which we directly assume for theoretical convenience. In Section 5, we discuss practical considerations for constructing negative pairs. We enforce sparsity regularization \\( L_{\\mathrm{sparsity}} \\) on the inferred mask \\( \\hat{\\mathbf{m}} \\) to select the simplest representation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.698, + 0.484, + 0.728 + ], + "angle": 0, + "content": "We introduce our key conditions in Condition 4.2 and theoretical results in Theorem 4.3." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.738, + 0.373, + 0.754 + ], + "angle": 0, + "content": "Condition 4.2 (Identification Conditions)." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.755, + 0.482, + 0.784 + ], + "angle": 0, + "content": "i [Smoothness & invertibility]: Generating functions \\( g_{\\mathrm{I}} \\) and \\( g_{\\mathrm{T}} \\) are smooth and have smooth inverses." + }, + { + "type": "text", + "bbox": [ + 0.108, + 0.786, + 0.483, + 0.831 + ], + "angle": 0, + "content": "ii [Fully-supported \\(p(\\mathbf{z}_{\\mathrm{I}},\\mathbf{m})]\\): The joint distribution over the semantic variable \\(\\mathbf{z}_{\\mathrm{I}}\\) and the mask \\(\\mathbf{m}\\) is fully supported: \\(p(\\mathbf{z}_{\\mathrm{I}},\\mathbf{m}) > 0\\) for any \\((\\mathbf{z}_{\\mathrm{I}},\\mathbf{m})\\in \\mathcal{Z}_{\\mathrm{I}}\\times \\mathcal{M}\\)." + }, + { + "type": "list", + "bbox": [ + 0.108, + 0.755, + 0.483, + 0.831 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.485, + 0.902 + ], + "angle": 0, + "content": "Discussion. Condition 4.2-\\(i\\) ensures the generating functions \\((g_{\\mathrm{I}}, g_{\\mathrm{T}})\\) preserve latent variables' information, without which the task of recovering such latent variables would be ill-posed. Practically, the high dimensionality of image" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.395 + ], + "angle": 0, + "content": "data i offers sufficient capacity to hold all information, and the text variable t only contains information filtered through its mask m. This condition is widely employed in the latent-variable identification literature [13, 17, 18, 43]. Condition 4.2-ii prescribes that the representation \\(\\mathbf{z}_{\\mathrm{I}}\\) and the mask m that marginally appear in the training distribution should also be present jointly with non-zero probability density. Interpreting the mask m as a concept selector (e.g., selecting \"bear\" and \"pen\"), this condition ensures that each concept retains its full range of variations (such as different shapes of bears and lengths of pens) across various mask selections. To satisfy this requirement, one can restrict the joint support \\(\\mathcal{Z}_{\\mathrm{I}} \\times \\mathcal{M}\\) to an appropriate subset, ensuring that only relevant combinations of \\(\\mathbf{z}_{\\mathrm{I}}\\) and m are present. Alternatively, one can enrich the caption set for each image, thereby increasing the diversity and coverage of concept combinations and filling in the joint support. This aligns with recent caption-augmentation techniques [21, 23-25, 54] as discussed in Section 1, revealing the synergy between our framework and existing efforts in the community." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.402, + 0.909, + 0.495 + ], + "angle": 0, + "content": "Theorem 4.3 (Concept Representation Identification). We assume the data-generating process in (1). Let \\((f_{\\mathrm{I}}, f_{\\mathrm{T}}, \\hat{\\mathbf{m}})\\) be an optimum of (2). Under Condition 4.2, the true representation \\([\\mathbf{z}]_{\\tilde{\\mathcal{B}}}\\) is block-wise identifiable for any index set \\(\\tilde{\\mathcal{B}}\\) such that \\(\\tilde{\\mathcal{B}} = \\cup_{\\mathbf{m} \\in \\mathcal{V}} \\mathcal{B}(\\mathbf{m})\\) or \\(\\tilde{\\mathcal{B}} = \\cap_{\\mathbf{m} \\in \\mathcal{V}} \\mathcal{B}(\\mathbf{m})\\) over any subset of masks \\(\\mathcal{V} \\subset \\mathcal{M}\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.503, + 0.906, + 0.669 + ], + "angle": 0, + "content": "Concept preservation. Theorem 4.3 states that one can recover the concept block \\(\\left[\\mathbf{z}_{\\mathrm{I}}\\right]_{\\mathcal{B}(\\mathbf{m})}\\) associated with each individual text caption in the dataset \\(\\mathcal{M}\\). Furthermore, it ensures that the union of concepts \\(\\left[\\mathbf{z}_{\\mathrm{I}}\\right]_{\\cup_{\\mathbf{m} \\in \\nu} \\mathcal{B}(\\mathbf{m})}\\) from any subset of text captions \\(\\mathcal{V} \\subset \\mathcal{M}\\) can be preserved. In the running example of Figure 1, our formulation allows us to preserve concepts (\"bear\", \"pen\", \"chair\") in the image representation by selectively matching them with the two captions, whereas existing models like CLIP may lose either \"pen\" and \"chair\" since they are only mentioned in one caption. Therefore, Theorem 4.3 effectively addresses Goal a." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.669, + 0.906, + 0.804 + ], + "angle": 0, + "content": "Concept disentanglement. The intersection operation in Theorem 4.3 empowers us to disentangle representations into potentially atomic concepts. In the example of Figure 1, we can identify the concept \"bear\" as the intersection of the two text captions, despite the absence of a standalone caption containing only \"bear\" in the dataset. Consequently, this part of the statement tackles Goal b. Our results underscore the importance of associating each image with a diverse set of captions that share overlapping concepts." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.805, + 0.906, + 0.88 + ], + "angle": 0, + "content": "Theoretical contribution. Theorem 4.3 extends existing theoretical frameworks [9, 43, 49]. Notably, Yao et al. [49] provide identification guarantees for shared representations over multiple views, generalizing earlier results confined to two views [9, 43]. This multi-view formulation" + }, + { + "type": "page_footnote", + "bbox": [ + 0.531, + 0.887, + 0.892, + 0.901 + ], + "angle": 0, + "content": "1We refer to a text caption with its mask m to simplify the notation." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "29783" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.334 + ], + "angle": 0, + "content": "is analogous to our setup when considering each group of text captions \\(\\mathbf{t}\\) associated with the same mask \\(\\mathbf{m}\\) as a distinct view group. However, prior work [49] relies on explicit knowledge of these groupings to train view-specific encoders. In contrast, our problem setting presents a greater challenge since we have no access to this grouping information. Specifically, given two captions of any different images, it is unclear whether they stem from the same mask (i.e., the same view group). Therefore, the identification guarantees in prior studies do not apply to our setting. Theorem 4.3 demonstrates that our estimation model, paired with the learning objective (2), can automatically infer the necessary grouping information (i.e., the masks). By doing so, our approach relaxes the identification conditions in previous work, enabling effective representation identification without explicit group knowledge." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.344, + 0.484, + 0.379 + ], + "angle": 0, + "content": "5. SmartCLIP: Modular Vision-language Alignment" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.389, + 0.484, + 0.45 + ], + "angle": 0, + "content": "Drawing on the theoretical framework in Section 4, we present SmartCLIP, a modular alignment model designed to achieve Goal a and Goal b. We discuss the implementation of the learning objective (2) and the model architecture." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.451, + 0.484, + 0.542 + ], + "angle": 0, + "content": "Modular alignment through adaptive masking. The masking function \\(\\hat{\\mathbf{m}} (\\cdot)\\) is instrumental in our modular alignment objective (2). SmartCLIP implements this function with a transformer block that ingests a caption representation \\(\\hat{\\mathbf{z}}_{\\mathrm{T}}\\) as its input and outputs a binary vector \\(\\hat{\\mathbf{m}} (\\hat{\\mathbf{z}}_{\\mathrm{T}})\\) via a straight-through estimator [4]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.543, + 0.484, + 0.65 + ], + "angle": 0, + "content": "Modular contrast construction. As discussed in Section 4, the negative pairs in regular contrastive losses [8, 35] serves a similar role as the invertibility assumption (Condition 4.2-i) [43, 47]. We denote generic image, text representations as \\(\\mathbf{I}\\), \\(\\mathbf{T}\\), and positive, negative pairs with \\(\\mathbf{P}_{\\mathrm{pos}}\\), \\(\\mathbf{P}_{\\mathrm{neg}}\\) respectively. The canonical one-side contrastive loss \\(\\mathcal{L}_{\\mathrm{ctr}}\\) [37] is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.146, + 0.661, + 0.484, + 0.766 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {\\mathrm {c t r}} \\left(\\underbrace {\\left(\\mathbf {I} ^ {(i)} , \\mathbf {T} ^ {(i)}\\right)} _ {\\mathbf {P} _ {\\text {p o s}}}, \\underbrace {\\left(\\mathbf {I} ^ {(i)} , \\mathbf {T} ^ {(j)}\\right)} _ {\\mathbf {P} _ {\\text {n e g}}} ^ {}\\right) \\tag {3} \\\\ = - \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\log \\frac {\\exp \\left(\\frac {\\mathrm {s i m} (\\mathbf {I} ^ {(i)} , \\mathbf {T} ^ {(i)})}{\\tau}\\right)}{\\sum_ {j = 1} ^ {N} \\exp \\left(\\frac {\\mathrm {s i m} (\\mathbf {I} ^ {(i)} , \\mathbf {T} ^ {(j)})}{\\tau}\\right)}, \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.779, + 0.483, + 0.809 + ], + "angle": 0, + "content": "where we denote the sample size, temperature, and cosine similarity with \\(N\\), \\(\\tau\\), and \\(\\sin(\\cdot)\\) respectively." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.484, + 0.856 + ], + "angle": 0, + "content": "Following the symmetric contrastive loss for CLIP [37], our alignment loss consists of two contrastive loss terms \\(\\mathcal{L}_{\\mathrm{ctrI}}\\) and \\(\\mathcal{L}_{\\mathrm{ctrT}}\\) that differ in the negative pairs:" + }, + { + "type": "equation", + "bbox": [ + 0.091, + 0.87, + 0.487, + 0.902 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {c t r I}} := \\mathcal {L} _ {\\mathrm {c t r}} \\left(\\mathbf {P} _ {\\text {p o s}}, \\mathbf {P} _ {\\text {n e g I}}\\right), \\mathcal {L} _ {\\mathrm {c t r T}} := \\mathcal {L} _ {\\mathrm {c t r}} \\left(\\mathbf {P} _ {\\text {p o s}}, \\mathbf {P} _ {\\text {n e g T}}\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.092, + 0.872, + 0.107 + ], + "angle": 0, + "content": "with the positive and negative pairs defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.601, + 0.119, + 0.905, + 0.145 + ], + "angle": 0, + "content": "\\[\n\\mathbf {P} _ {\\text {p o s}} := \\left(\\hat {\\mathbf {z}} _ {\\mathrm {I}} ^ {(i)} \\odot \\hat {\\mathbf {m}} \\left(\\hat {\\mathbf {z}} _ {\\mathrm {T}} ^ {(i)}\\right), \\hat {\\mathbf {z}} _ {\\mathrm {T}} ^ {(i)}\\right), \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.597, + 0.147, + 0.905, + 0.172 + ], + "angle": 0, + "content": "\\[\n\\mathbf {P} _ {\\text {n e g I}} := \\left(\\hat {\\mathbf {z}} _ {\\mathrm {I}} ^ {(i)} \\odot \\hat {\\mathbf {m}} \\left(\\hat {\\mathbf {z}} _ {\\mathrm {T}} ^ {(j)}\\right), \\hat {\\mathbf {z}} _ {\\mathrm {T}} ^ {(j)}\\right), \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.593, + 0.174, + 0.905, + 0.2 + ], + "angle": 0, + "content": "\\[\n\\mathbf {P} _ {\\mathrm {n e g T}} := \\left(\\hat {\\mathbf {z}} _ {\\mathrm {I}} ^ {(j)} \\odot \\hat {\\mathbf {m}} \\left(\\hat {\\mathbf {z}} _ {\\mathrm {T}} ^ {(i)}\\right), \\hat {\\mathbf {z}} _ {\\mathrm {T}} ^ {(i)}\\right). \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.214, + 0.905, + 0.314 + ], + "angle": 0, + "content": "In particular, \\(\\mathbf{P}_{\\mathrm{negI}}\\) contrasts the image representation \\(\\hat{\\mathbf{z}}_{\\mathrm{I}}^{(i)}\\) in the positive pair with randomly sampled caption representations \\(\\hat{\\mathbf{z}}_{\\mathrm{T}}^{(j)}\\) (see the green region in Figure 3), whereas \\(\\mathbf{P}_{\\mathrm{negT}}\\) contrasts the text representation \\(\\hat{\\mathbf{z}}_{\\mathrm{T}}^{(i)}\\) in the positive pair with randomly sampled image representations \\(\\hat{\\mathbf{z}}_{\\mathrm{I}}^{(j)}\\) (see the orange region in Figure 3)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.315, + 0.904, + 0.346 + ], + "angle": 0, + "content": "Sparsity penalty. We implement \\( L_{\\mathrm{sparsity}} \\) in (2) with a \\( \\ell_1 \\) term for its compatibility with deep-learning training:" + }, + { + "type": "equation", + "bbox": [ + 0.635, + 0.358, + 0.905, + 0.377 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {s p a r s i t y}} = \\left\\| \\hat {\\mathbf {m}} (\\mathbf {t}) \\right\\| _ {1}. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.387, + 0.905, + 0.432 + ], + "angle": 0, + "content": "This term ensures that the textual concepts are encoded into a minimal number of latent dimensions, promoting the disentanglement of distinct concepts across text captions." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.433, + 0.905, + 0.478 + ], + "angle": 0, + "content": "SmartCLIP training objective. In summary, the training objective of SmartCLIP is a weighted sum of loss terms in (4) and (8):" + }, + { + "type": "equation", + "bbox": [ + 0.538, + 0.492, + 0.905, + 0.51 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\lambda_ {\\text {a l i g n}} \\cdot \\left(\\mathcal {L} _ {\\mathrm {c t r I}} + \\mathcal {L} _ {\\mathrm {c t r T}}\\right) + \\lambda_ {\\text {s p a r s i t y}} \\cdot \\mathcal {L} _ {\\text {s p a r s i t y}}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.521, + 0.896, + 0.537 + ], + "angle": 0, + "content": "where \\(\\lambda_{\\mathrm{align}}\\) and \\(\\lambda_{\\mathrm{sparsity}}\\) denote the weighting coefficients." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.551, + 0.646, + 0.568 + ], + "angle": 0, + "content": "6. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.577, + 0.596, + 0.593 + ], + "angle": 0, + "content": "6.1. Setup" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.599, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Implementation details. Following Long-CLIP [54], we finetune the CLIP model [37] on ShareGPT4V [7], which contains around 1 million image-text pairs. We employ the position encoding in long-CLIP to handle 248 tokens (c.f., the 77-token limit in the original CLIP). Compared to the baseline CLIP model, we introduce a mask network \\(\\hat{\\mathbf{m}}\\). The masking network is designed as a single transformer block, which takes the text sequence embedding \\(\\hat{\\mathbf{z}}_{\\mathrm{T}}\\) from the text encoder. Then we add an attention-pooling layer to down-sample it to the same size as the CLIP representation, e.g., 768 in ViT-L/14. We tested including more transformer blocks in the mask network but did not observe significant improvements. Therefore, we stick to one block for faster training and inference. Unlike Long-CLIP [54] which processes all the captions for each image at each gradient step, we only sample one caption for each image, reducing our overall training time by half. Specifically, on 8 H100 GPUs, training a Vit-B/16 model for one epoch takes about 4 minutes with our model, whereas it takes around 7 minutes for Long-CLIP. After the pooling layer, we apply sigmoid to" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "29784" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.187, + 0.096, + 0.808, + 0.279 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.295, + 0.908, + 0.325 + ], + "angle": 0, + "content": "Figure 3. The diagram of SmartCLIP. On the left, we introduce adaptive masking for alignment with different text prompts. The mask network selects which part of the image representation to be used. On the right, we present our modular contrastive objectives (4)." + }, + { + "type": "table_caption", + "bbox": [ + 0.1, + 0.338, + 0.896, + 0.352 + ], + "angle": 0, + "content": "Table 1. Results of short-caption text-image retrieval on the 5k COCO2017 validation set and the whole 30k Flickr30K dataset." + }, + { + "type": "table", + "bbox": [ + 0.122, + 0.363, + 0.871, + 0.535 + ], + "angle": 0, + "content": "
COCOFlickr30k
Image-to-TextText-to-ImageImage-to-TextText-to-Image
R@1R@5R@10R@1R@5R@10R@1R@5R@10R@1R@5R@10
B/16CLIP51.876.884.332.757.768.244.168.277.024.745.154.6
Direct Fine-tuning37.462.372.121.843.454.525.745.855.417.934.543.1
Long-CLIP [54]57.681.187.840.465.875.246.871.479.834.156.365.7
SmartCLIP (Ours)61.983.389.742.468.277.855.678.285.036.358.867.8
L/14CLIP56.179.586.835.460.170.248.572.680.828.049.358.7
Direct Fine-tuning37.963.172.223.145.155.926.046.355.617.934.943.5
Long-CLIP [54]62.885.191.246.370.879.853.477.585.341.264.172.6
SmartCLIP (Ours)66.086.292.648.573.181.763.984.490.243.866.574.8
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.545, + 0.484, + 0.606 + ], + "angle": 0, + "content": "restrict the output to the range \\((0,1)\\) and employ straight-through estimation [4] to binarize the outputs. The training batch size is 1024 and the learning rate is \\(10^{-6}\\) for the CLIP component and \\(10^{-3}\\) for the mask network." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.612, + 0.411, + 0.627 + ], + "angle": 0, + "content": "Evaluation. We evaluate the following datasets:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.634, + 0.483, + 0.695 + ], + "angle": 0, + "content": "- Long text-to-image retrieval datasets: ShareGPT4V validation split [7] and Urban1k [54]. The captions for each image are long and describe details about the image. Both datasets contain 1000 text-to-image pairs." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.695, + 0.483, + 0.74 + ], + "angle": 0, + "content": "- Short text-to-image retrieval datasets: COCO2017 validation split [27] and Flickr30K [50]. Following LongCLIP [54], we use 30K Flickr training dataset." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.74, + 0.483, + 0.8 + ], + "angle": 0, + "content": "- Zero-shot image classification datasets. We use benchmark datasets: Country211, Fer2013, Fgvc-aircraft, GT-SRB, ImageNet, ImagetNet-V2, VOC2007, VOC2007-Multi, and SUN397." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.634, + 0.483, + 0.8 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.807, + 0.484, + 0.852 + ], + "angle": 0, + "content": "Baselines. In this paper, we benchmark our approach against CLIP [37] and the recent state-of-the-art model long-CLIP [54]." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.545, + 0.791, + 0.56 + ], + "angle": 0, + "content": "6.2. Comparison with CLIP Models" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.568, + 0.906, + 0.612 + ], + "angle": 0, + "content": "We present our experimental results across three key tasks: long text-to-image retrieval, short text-to-image retrieval, and zero-shot classification." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.613, + 0.906, + 0.718 + ], + "angle": 0, + "content": "Long text-to-image retrieval. Table 2 showcases our method's performance on long text-to-image retrieval tasks. SmartCLIP achieves substantial improvements over baseline models, particularly the SOTA Long-CLIP, which is designed to handle long text sequences. For example, on the Urban1k dataset, SmartCLIP elevates the performance from \\(78.9\\%\\) to \\(90.0\\%\\), marking an impressive \\(14\\%\\) boost." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.906, + 0.824 + ], + "angle": 0, + "content": "Short text-to-image retrieval. Similarly, as illustrated in Table 1, SmartCLIP significantly outperforms all baseline models across various metrics and datasets in short text-to-image retrieval tasks. The encouraging performance gains show that SmartCLIP can capture detailed information within images while simultaneously emphasizing the main semantic content." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Zero-shot classification. To comprehensively evaluate our model's capabilities, we conduct zero-shot classification benchmarks in Table 3. Both the standard CLIP model and SmartCLIP demonstrate superior performance on different datasets. Notably, SmartCLIP exhibits a slight perfor" + }, + { + "type": "page_footnote", + "bbox": [ + 0.108, + 0.887, + 0.448, + 0.9 + ], + "angle": 0, + "content": "2https://github.com/LAION-AI/CLIP_benchmark" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "29785" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.104, + 0.503, + 0.147 + ], + "angle": 0, + "content": "Table 2. The R@1 of long-caption text-image retrieval on 1k ShareGPT4V [2] validation set and Urban-1000 dataset. The best results are bold. We cite the results from Long-CLIP [54]." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.157, + 0.472, + 0.314 + ], + "angle": 0, + "content": "
ShareGPT4VUrban1k
I2TT2II2TT2I
B/16CLIP [37]78.279.668.153.6
Direct Fine-tuning94.193.6--
Long-CLIP [54]94.693.378.979.5
SmartCLIP (Ours)98.798.190.087.4
L/14CLIP [37]81.884.068.752.8
Direct Fine-tuning95.395.4--
Long-CLIP [54]95.895.682.786.1
SmartCLIP (Ours)97.998.593.090.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.535, + 0.089, + 0.907, + 0.16 + ], + "angle": 0, + "content": "Table 3. Zero-shot classification performance on ViT-L/14 models. When the class name is very short, i.e., a single word like ImageNet, CLIP model performs better. When the class name is a combination of several words, our method achieves better results, e.g., the road sign in GTSRB." + }, + { + "type": "table", + "bbox": [ + 0.538, + 0.169, + 0.917, + 0.327 + ], + "angle": 0, + "content": "
DatasetCLIPLongCLIPSmartCLIP
Country21131.828.126.9
Fer201349.057.858.6
Fgvc-aircraft31.730.630.4
GTSRB50.248.952.4
ImageNet75.372.972.5
ImageNet-V269.766.966.6
VOC200778.377.578.6
VOC2007-Multi79.082.183.7
SUN39767.572.572.1
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.345, + 0.585, + 0.526 + ], + "angle": 0, + "content": "This playful food sculpture transforms cucumbers into a fearsome T-Rex dinosaur. The cucumbers form the main body, with whole cucumbers creating the legs and tail, while sliced cucumbers make up the creature's midsection. More cucumbers are cleverly cut to shape the dinosaur's head, and additional cucumbers are arranged to suggest muscular limbs. From its mouth erupts a dramatic spray of carrots, with finely julienned carrots creating the effect of fire. These bright orange carrots provide a stunning contrast against the green vegetables. Shredded carrots cascade downward like flames, while more carrots are delicately cut to create a flame-like texture. The carrots' vibrant color makes the dinosaur appear truly animated. Fresh celery leaves crown the creation, with celery fronds adding a decorative touch around the body. More celery leaves create a natural backdrop, while additional celery pieces add texture throughout. The celery's feathered appearance provides an artistic flourish to the overall design." + }, + { + "type": "image", + "bbox": [ + 0.618, + 0.343, + 0.807, + 0.463 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.618, + 0.464, + 0.806, + 0.524 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.527, + 0.907, + 0.57 + ], + "angle": 0, + "content": "Figure 4. Example of Long-text-to-image generation. We replace the CLIP text encoder in SDXL with different finetuned CLIP models. Given a very long text, CLIP [37] truncates the input to 77 tokens, resulting in information loss in the image. Our model learns to generate details such as celery leaves on the back of the dinosaur while other models fail." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.584, + 0.483, + 0.691 + ], + "angle": 0, + "content": "mance decline on ImageNet, which is expected since our model is fine-tuned on the ShareGPT4V dataset featuring long text captions, whereas ImageNet primarily consists of short, often single-word class names. However, Smart-CLIP excels on datasets with multi-word class names, such as the GTSRB dataset, where it achieves the best performance in accurately classifying road sign descriptions." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.698, + 0.255, + 0.713 + ], + "angle": 0, + "content": "6.3. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.483, + 0.765 + ], + "angle": 0, + "content": "We analyze the three key components in our model: the modular alignment module, the sparsity loss and the impact of caption diversity in the data." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Modular alignment. After introducing the mask network, we replace the standard contrastive learning with our selective alignment module. As shown in Figure 5, this change significantly improves performance. When using standard contrastive learning instead of the modular contrastive module (indicated by the purple lines), performance drops sharply. This happens because the mask information allows the network to easily separate positive pairs from negative ones, making the negative samples less informa" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.584, + 0.907, + 0.614 + ], + "angle": 0, + "content": "tive. As a result, standard contrastive learning no longer effectively helps the model learn meaningful information." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.614, + 0.906, + 0.705 + ], + "angle": 0, + "content": "Alignment coefficient \\(\\lambda_{\\mathrm{align}}\\). We test the impact of the alignment coefficient \\(\\lambda_{\\mathrm{align}}\\). The results in the right panels of Figure 5 show that our method performs consistently well across a wide range of \\(\\lambda_{\\mathrm{align}}\\) values, from 0.1 to 20. This indicates that our approach is robust and does not require precise tuning of \\(\\lambda_{\\mathrm{align}}\\) to achieve good performance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.906, + 0.809 + ], + "angle": 0, + "content": "Sparsity coefficient \\(\\lambda_{\\mathrm{sparsity}}\\). We also examine the sparsity coefficient \\(\\lambda_{\\mathrm{sparsity}}\\). The left panels of Figure 5 demonstrate that adding sparsity to the mask network improves performance. This supports our idea that promoting sparsity helps the model focus on the most relevant concepts, enhancing its ability to capture detailed information without being distracted by irrelevant details." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Caption diversity. We evaluate our model's performance under varying caption diversity conditions using the COCO dataset [27]. As shown in Table 4, increasing the number of captions per image enhances performance on the Flickr30K dataset, though at the expense of degraded performance on long-text-to-image retrieval tasks. Further improvements" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "29786" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.091, + 0.282, + 0.196 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.302, + 0.092, + 0.484, + 0.195 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.092, + 0.688, + 0.195 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.709, + 0.092, + 0.891, + 0.195 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.199, + 0.907, + 0.226 + ], + "angle": 0, + "content": "Figure 5. Ablation Studies on two proposed modules: selective alignment and sparsity. The baseline w.o. Modular means that we replace our modular alignment module with standard contrastive learning alignment." + }, + { + "type": "image", + "bbox": [ + 0.104, + 0.231, + 0.473, + 0.389 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.402, + 0.484, + 0.487 + ], + "angle": 0, + "content": "Figure 6. Visualization of learned representations. Given an image, we generate two captions (e.g., a zebra and a zebra and a deer) and compute cosine similarities with the same image embedding to perform binary classification for visualization using Score-CAM [45]. Compared to baseline methods, our CLIP representations are more atomic and capture differences more effectively." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.503, + 0.484, + 0.564 + ], + "angle": 0, + "content": "are achieved when we combine our training dataset with COCO. These results highlight both the importance of caption diversity and our method's capability to effectively handle complex text-image datasets." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.575, + 0.271, + 0.59 + ], + "angle": 0, + "content": "6.4. Additional Results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.598, + 0.483, + 0.794 + ], + "angle": 0, + "content": "Visualization. While our quantitative results demonstrate superior performance across various tasks, we also explore the qualitative aspects of our model by visualizing the learned representations. The visualization results are shown in Figure 6. We employ the ScoreCAM method [45] for this purpose. For each image, we generate two distinct captions, such as \"a zebra\" and \"a zebra and a deer\". We then compute the cosine similarity between the image embedding and each of the two text embeddings. These similarity scores serve as logits for a classification task, which are then input into the ScoreCAM algorithm. SmartCLIP successfully learns modular representations, accurately capturing the relevant differences between captions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Plug and play for text-to-image generation. One main advantage of SmartCLIP over other CLIP models trained from scratch is the low computational cost of finetuning. Additionally, our fine-tuned text encoder can replace the CLIP text encoders in large-scale models in a plug-and-play manner. Specifically, we substitute the text encoder in the SDXL [36] model with both Long-CLIP and SmartCLIP." + }, + { + "type": "table", + "bbox": [ + 0.532, + 0.225, + 0.885, + 0.31 + ], + "angle": 0, + "content": "
NumCapPerImg-COCOT2II2TLongT2ILongI2T
153.639.385.389.3
353.640.985.388.6
556.441.285.286.5
Ours-ShareGPT55.636.398.798.1
+COCO57.038.497.898.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.312, + 0.907, + 0.355 + ], + "angle": 0, + "content": "Table 4. Retrieval results on Flickr30K [50] and ShareGPT-val [7] with models trained with different caption counts per image (top) and mixing long and short text-image datasets (bottom)." + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.365, + 0.907, + 0.435 + ], + "angle": 0, + "content": "Table 5. Long text to image generation performance. We use the long captions (usually around 200 tokens, beyond the hard constraint 77 of the original CLIP model) to generate images with SDXL model [36], then we compare the generated images against the real images in ShareGPT4V validation split." + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.446, + 0.901, + 0.496 + ], + "angle": 0, + "content": "
MethodKID ↓Pr ↑Re ↑F1 ↑DINO-L ↑
LongCLIP1.050.2380.7680.3630.401
SmartCLIP1.020.2580.7910.3890.414
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.508, + 0.907, + 0.63 + ], + "angle": 0, + "content": "As illustrated in Figure 4, our text encoder demonstrates a superior understanding of the long text, generating detailed elements such as celery leaves in the background. Furthermore, Table 5 presents quantitative results on image generation from long captions in the ShareGPT4V validation split. Our method consistently achieves better performance across various metrics, showcasing its effectiveness in handling complex, long text inputs." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.648, + 0.766, + 0.664 + ], + "angle": 0, + "content": "7. Conclusion and Limitation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.909, + 0.901 + ], + "angle": 0, + "content": "In this work, we address the information misalignment and representation entanglement issues in existing vision-language models (e.g., CLIP). We establish theoretical conditions for effectively connecting text representations to atomic-level visual features and propose SmartCLIP, a principled, refined vision-language model. Our experimental results validate both our theoretical results and the practical effectiveness of SmartCLIP in advancing multimodal learning. Limitation. As discussed earlier, Condition 4.2-ii could be violated for datasets in which a subset images are paired with a limited number of captions compared to others. In Section 4, we discuss practical strategies to mitigate such issues. Devising alternative theoretical conditions may provide additional insights into fully utilizing all pairing information, which we leave as future work." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "29787" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.251, + 0.108 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.484, + 0.222 + ], + "angle": 0, + "content": "We would like to acknowledge the support from NSF Award No. 2229881, AI Institute for Societal Decision Making (AI-SDM), the National Institutes of Health (NIH) under Contract R01HL159805, and grants from Quris AI, Florin Court Capital, and MBZUAI-WIS Joint Program. The work of L. Kong is supported in part by NSF DMS-2134080 through an award to Y. Chi." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.236, + 0.188, + 0.252 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.261, + 0.484, + 0.317 + ], + "angle": 0, + "content": "[1] Kartik Ahuja, Divyat Mahajan, Yixin Wang, and Yoshua Bengio. Interventional causal representation learning. In International conference on machine learning, pages 372-407. PMLR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.319, + 0.484, + 0.387 + ], + "angle": 0, + "content": "[2] Alberto Baldrati, Lorenzo Agnolucci, Marco Bertini, and Alberto Del Bimbo. Zero-shot composed image retrieval with textual inversion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15338-15347, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.39, + 0.483, + 0.432 + ], + "angle": 0, + "content": "[3] Yoshua Bengio, Réjean Ducharme, and Pascal Vincent. A neural probabilistic language model. Advances in neural information processing systems, 13, 2000. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.434, + 0.483, + 0.488 + ], + "angle": 0, + "content": "[4] Yoshua Bengio, Nicholas Léonard, and Aaron Courville. Estimating or propagating gradients through stochastic neurons for conditional computation. arXiv preprint arXiv:1308.3432, 2013. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.491, + 0.484, + 0.559 + ], + "angle": 0, + "content": "[5] Simon Buchholz, Goutham Rajendran, Elan Rosenfeld, Bryon Aragam, Bernhard Scholkopf, and Pradeep Ravikumar. Learning linear causal representations from interventions under general nonlinear mixing. Advances in Neural Information Processing Systems, 36, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.562, + 0.484, + 0.63 + ], + "angle": 0, + "content": "[6] Chang Che, Qunwei Lin, Xinyu Zhao, Jiaxin Huang, and Liqiang Yu. Enhancing multimodal understanding with clip-based image-to-text transformation. In Proceedings of the 2023 6th International Conference on Big Data Technologies, pages 414-418, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.633, + 0.482, + 0.687 + ], + "angle": 0, + "content": "[7] Lin Chen, Jinsong Li, Xiaoyi Dong, Pan Zhang, Conghui He, Jiaqi Wang, Feng Zhao, and Dahua Lin. Sharegpt4v: Improving large multi-modal models with better captions. arXiv preprint arXiv:2311.12793, 2023. 2, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.689, + 0.483, + 0.744 + ], + "angle": 0, + "content": "[8] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 1, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.746, + 0.483, + 0.802 + ], + "angle": 0, + "content": "[9] Imant Daunhawer, Alice Bizeul, Emanuele Palumbo, Alexander Marx, and Julia E Vogt. Identifiability results for multimodal contrastive learning. In The Eleventh International Conference on Learning Representations, 2023. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.804, + 0.483, + 0.857 + ], + "angle": 0, + "content": "[10] Lijie Fan, Dilip Krishnan, Phillip Isola, Dina Katabi, and Yonglong Tian. Improving clip training with language rewrites. Advances in Neural Information Processing Systems, 36, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.86, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[11] Luigi Gresele, Paul K Rubenstein, Arash Mehrjou, Francesco Locatello, and Bernhard Schölkopf. The incomplete rosetta stone problem: Identifiability results for multi-" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.261, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.121 + ], + "angle": 0, + "content": "view nonlinearica.In Uncertainty in Artificial Intelligence, pages 217-227.PMLR,2020.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.123, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[12] Weiquan Huang, Aooqi Wu, Yifan Yang, Xufang Luo, Yuqing Yang, Liang Hu, Qi Dai, Xiyang Dai, Dongdong Chen, Chong Luo, et al. Llm2clip: Powerful language model unlock richer visual representation. arXiv preprint arXiv:2411.04997, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.193, + 0.905, + 0.234 + ], + "angle": 0, + "content": "[13] Aapo Hyvarinen and Hiroshi Morioka. Unsupervised feature extraction by time-contrastive learning and nonlinear ica, 2016. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.237, + 0.905, + 0.304 + ], + "angle": 0, + "content": "[14] Aapo Hyvarinen, Hiroaki Sasaki, and Richard Turner. Nonlinear ica using auxiliary variables and generalized contrastive learning. In The 22nd International Conference on Artificial Intelligence and Statistics, pages 859-868. PMLR, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.307, + 0.905, + 0.389 + ], + "angle": 0, + "content": "[15] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International conference on machine learning, pages 4904-4916. PMLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.392, + 0.905, + 0.447 + ], + "angle": 0, + "content": "[16] Chen Ju, Tengda Han, Kunhao Zheng, Ya Zhang, and Weidi Xie. Prompting visual-language models for efficient video understanding. In European Conference on Computer Vision, pages 105-124. Springer, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.449, + 0.905, + 0.517 + ], + "angle": 0, + "content": "[17] Ilyes Khemakhem, Diederik Kingma, Ricardo Monti, and Aapo Hyvarinen. Variational autoencoders and nonlinear ica: A unifying framework. In International Conference on Artificial Intelligence and Statistics, pages 2207-2217. PMLR, 2020. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.52, + 0.905, + 0.589 + ], + "angle": 0, + "content": "[18] Lingjing Kong, Shaoan Xie, Weiran Yao, Yujia Zheng, Guangyi Chen, Petar Stojanov, Victor Akinwande, and Kun Zhang. Partial disentanglement for domain adaptation. In Proceedings of the 39th International Conference on Machine Learning, pages 11455-11472. PMLR, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.591, + 0.905, + 0.644 + ], + "angle": 0, + "content": "[19] Sébastien Lachapelle, Divyat Mahajan, Ioannis Mitliagkas, and Simon Lacoste-Julien. Additive decoders for latent variables identification and cartesian-product extrapolation, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.647, + 0.905, + 0.716 + ], + "angle": 0, + "content": "[20] Zhengfeng Lai, Vasileios Saveris, Chen Chen, Hong-You Chen, Haotian Zhang, Bowen Zhang, Juan Lao Tebar, Wenze Hu, Zhe Gan, Peter Grasch, et al. Revisit large-scale image-caption data in pre-training multimodal foundation models. arXiv preprint arXiv:2410.02740, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.719, + 0.905, + 0.786 + ], + "angle": 0, + "content": "[21] Zhengfeng Lai, Haotian Zhang, Wentao Wu, Haoping Bai, Aleksei Timofeev, Xianzhi Du, Zhe Gan, Jiulong Shan, Chen-Nee Chuah, Yinfei Yang, et al. Veclip: Improving clip training via visual-enriched captions. ECCV. IEEE, page 13, 2024. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.789, + 0.905, + 0.857 + ], + "angle": 0, + "content": "[22] Samuel Lavoie, Polina Kirichenko, Mark Ibrahim, Mahmoud Assran, Andrew Gordon Wilson, Aaron Courville, and Nicolas Ballas. Modeling caption diversity in contrastive vision-language pretraining. arXiv preprint arXiv:2405.00740, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[23] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In Interna-" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "29788" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.126, + 0.092, + 0.482, + 0.12 + ], + "angle": 0, + "content": "tional conference on machine learning, pages 12888-12900. \nPMLR, 2022. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.19 + ], + "angle": 0, + "content": "[24] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730-19742. PMLR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.192, + 0.482, + 0.26 + ], + "angle": 0, + "content": "[25] Xianhang Li, Haoqin Tu, Mude Hui, Zeyu Wang, Bingchen Zhao, Junfei Xiao, Sucheng Ren, Jieru Mei, Qing Liu, Huangjie Zheng, et al. What if we recapture billions of web images with llama-3? arXiv preprint arXiv:2406.08478, 2024. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.263, + 0.482, + 0.332 + ], + "angle": 0, + "content": "[26] Yangguang Li, Feng Liang, Lichen Zhao, Yufeng Cui, Wanli Ouyang, Jing Shao, Fengwei Yu, and Junjie Yan. Supervision exists everywhere: A data efficient contrastive language-image pre-training paradigm. arXiv preprint arXiv:2110.05208, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.334, + 0.482, + 0.417 + ], + "angle": 0, + "content": "[27] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.418, + 0.482, + 0.46 + ], + "angle": 0, + "content": "[28] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.461, + 0.482, + 0.488 + ], + "angle": 0, + "content": "[29] AI Meta. Introducing meta llama 3: The most capable openly available llm to date. Meta AI, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.49, + 0.482, + 0.546 + ], + "angle": 0, + "content": "[30] Tomas Mikolov, Ilya Sutskever, Kai Chen, Greg S Corrado, and Jeff Dean. Distributed representations of words and phrases and their compositionality. Advances in neural information processing systems, 26, 2013. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.548, + 0.482, + 0.616 + ], + "angle": 0, + "content": "[31] Hiroshi Morioka and Aapo Hyvarinen. Connectivity-contrastive learning: Combining causal discovery and representation learning for multimodal data. In International conference on artificial intelligence and statistics, pages 3399–3426. PMLR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.618, + 0.482, + 0.673 + ], + "angle": 0, + "content": "[32] Hiroshi Morioka and Aapo Hyvarinen. Causal representation learning made identifiable by grouping of observational variables. In *Forty-first International Conference on Machine Learning*, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.675, + 0.482, + 0.731 + ], + "angle": 0, + "content": "[33] Norman Mu, Alexander Kirillov, David Wagner, and Saining Xie. Slip: Self-supervision meets language-image pretraining. In European conference on computer vision, pages 529-544. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.732, + 0.482, + 0.787 + ], + "angle": 0, + "content": "[34] Ivona Najdenkoska, Mohammad Mahdi Derakhshani, Yuki M Asano, Nanne van Noord, Marcel Worring, and Cees GM Snoek. Tulip: Token-length upgraded clip. arXiv preprint arXiv:2410.10034, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.789, + 0.482, + 0.83 + ], + "angle": 0, + "content": "[35] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 1, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[36] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 8" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.176 + ], + "angle": 0, + "content": "[37] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 2, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.178, + 0.905, + 0.233 + ], + "angle": 0, + "content": "[38] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.235, + 0.907, + 0.303 + ], + "angle": 0, + "content": "[39] Hanoona Rasheed, Muhammad Uzair Khattak, Muhammad Maaz, Salman Khan, and Fahad Shahbaz Khan. Fine-tuned clip models are efficient video learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6545-6554, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.305, + 0.907, + 0.375 + ], + "angle": 0, + "content": "[40] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.377, + 0.905, + 0.418 + ], + "angle": 0, + "content": "[41] Nils Sturma, Chandler Squires, Mathias Drton, and Caroline Uhler. Unpaired multi-domain causal representation learning, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.42, + 0.905, + 0.475 + ], + "angle": 0, + "content": "[42] Mingkang Tang, Zhanyu Wang, Zhenhua Liu, Fengyun Rao, Dian Li, and Xiu Li. Clip4caption: Clip for video caption. In Proceedings of the 29th ACM International Conference on Multimedia, pages 4858-4862, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.477, + 0.905, + 0.547 + ], + "angle": 0, + "content": "[43] Julius von Kugelgen, Yash Sharma, Luigi Gresele, Wieland Brendel, Bernhard Scholkopf, Michel Besserve, and Francesco Locatello. Self-supervised learning with data augmentations provably isolates content from style. arXiv preprint arXiv:2106.04619, 2021. 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.548, + 0.905, + 0.629 + ], + "angle": 0, + "content": "[44] Julius von Kugelgen, Michel Besserve, Liang Wendong, Luigi Gresele, Armin Kekic, Elias Bareinboim, David Blei, and Bernhard Scholkopf. Nonparametric identifiability of causal representations from unknown interventions. Advances in Neural Information Processing Systems, 36, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.632, + 0.905, + 0.715 + ], + "angle": 0, + "content": "[45] Haofan Wang, Zifan Wang, Mengnan Du, Fan Yang, Zijian Zhang, Sirui Ding, Piotr Mardziel, and Xia Hu. Score-cam: Score-weighted visual explanations for convolutional neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 24-25, 2020. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.717, + 0.905, + 0.773 + ], + "angle": 0, + "content": "[46] Jianyi Wang, Kelvin CK Chan, and Chen Change Loy. Exploring clip for assessing the look and feel of images. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2555-2563, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.775, + 0.905, + 0.83 + ], + "angle": 0, + "content": "[47] Tongzhou Wang and Phillip Isola. Understanding contrastive representation learning through alignment and uniformity on the hypersphere. In International conference on machine learning, pages 9929-9939. PMLR, 2020. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.832, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[48] Wei Wu, Kecheng Zheng, Shuailei Ma, Fan Lu, Yuxin Guo, Yifei Zhang, Wei Chen, Qingpei Guo, Yujun Shen, and Zheng-Jun Zha. Lotlip: Improving language-image pre-training for long text understanding. arXiv preprint arXiv:2410.05249, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.907, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "29789" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.174 + ], + "angle": 0, + "content": "[49] Dingling Yao, Danru Xu, Sebastien Lachapelle, Sara Maglia-cane, Perouz Taslakian, Georg Martius, Julius von Kügelgen, and Francesco Locatello. Multi-view causal representation learning with partial observability. In The Twelfth International Conference on Learning Representations, 2024. 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.178, + 0.482, + 0.245 + ], + "angle": 0, + "content": "[50] Peter Young, Alice Lai, Micah Hodosh, and Julia Hockenmaier. From image descriptions to visual denotations: New similarity metrics for semantic inference over event descriptions. Transactions of the Association for Computational Linguistics, 2:67-78, 2014. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.249, + 0.482, + 0.301 + ], + "angle": 0, + "content": "[51] Jiahui Yu, Zirui Wang, Vijay Vasudevan, Legg Yeung, Mojtaba Seyedhosseini, and Yonghui Wu. Coca: Contrastive captioners are image-text foundation models. arXiv preprint arXiv:2205.01917, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.305, + 0.482, + 0.373 + ], + "angle": 0, + "content": "[52] Xiaohua Zhai, Xiao Wang, Basil Mustafa, Andreas Steiner, Daniel Keysers, Alexander Kolesnikov, and Lucas Beyer. Lit: Zero-shot transfer with locked-image text tuning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18123-18133, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.376, + 0.482, + 0.43 + ], + "angle": 0, + "content": "[53] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11975-11986, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.432, + 0.482, + 0.472 + ], + "angle": 0, + "content": "[54] Beichen Zhang, Pan Zhang, Xiaoyi Dong, Yuhang Zang, and Jiaqi Wang. Long-clip: Unlocking the long-text capability of clip. arXiv preprint arXiv:2403.15378, 2024. 2, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.475, + 0.482, + 0.543 + ], + "angle": 0, + "content": "[55] Jiaqi Zhang, Kristjan Greenewald, Chandler Squires, Akash Srivastava, Karthikeyan Shanmugam, and Caroline Uhler. Identifiability guarantees for causal disentanglement from soft interventions. Advances in Neural Information Processing Systems, 36, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.546, + 0.482, + 0.598 + ], + "angle": 0, + "content": "[56] Jihai Zhang, Xiaoye Qu, Tong Zhu, and Yu Cheng. Clipmoe: Towards building mixture of experts for clip with diversified multiplet upcycling. arXiv preprint arXiv:2409.19291, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.602, + 0.482, + 0.642 + ], + "angle": 0, + "content": "[57] Kun Zhang, Shaoan Xie, Ignavier Ng, and Yujia Zheng. Causal representation learning from multiple distributions: A general setting. arXiv preprint arXiv:2402.05052, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.645, + 0.482, + 0.712 + ], + "angle": 0, + "content": "[58] Renrui Zhang, Ziyu Guo, Wei Zhang, Kunchang Li, Xupeng Miao, Bin Cui, Yu Qiao, Peng Gao, and Hongsheng Li. Pointclip: Point cloud understanding by clip. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8552-8562, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.715, + 0.482, + 0.769 + ], + "angle": 0, + "content": "[59] Kecheng Zheng, Yifei Zhang, Wei Wu, Fan Lu, Shuailei Ma, Xin Jin, Wei Chen, and Yujun Shen. Dreamclip: Language-image pre-training with long captions. arXiv preprint arXiv:2403.17007, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.769 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "29790" + } + ] +] \ No newline at end of file diff --git a/2025/SmartCLIP_ Modular Vision-language Alignment with Identification Guarantees/28f50741-407e-4ef2-b9fa-f2bf47df9fd3_origin.pdf b/2025/SmartCLIP_ Modular Vision-language Alignment with Identification Guarantees/28f50741-407e-4ef2-b9fa-f2bf47df9fd3_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..405bfcda4c7adae7080aeaacee1171dfc6d0194a --- /dev/null +++ b/2025/SmartCLIP_ Modular Vision-language Alignment with Identification Guarantees/28f50741-407e-4ef2-b9fa-f2bf47df9fd3_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fb388e926ddaff710813974fb970fbfe48b6fd2462ad0b237878eb20f545ebd9 +size 609863 diff --git a/2025/SmartCLIP_ Modular Vision-language Alignment with Identification Guarantees/full.md b/2025/SmartCLIP_ Modular Vision-language Alignment with Identification Guarantees/full.md new file mode 100644 index 0000000000000000000000000000000000000000..f0777677250228d8bb135a0fea0467eb8a580693 --- /dev/null +++ b/2025/SmartCLIP_ Modular Vision-language Alignment with Identification Guarantees/full.md @@ -0,0 +1,364 @@ +# SmartCLIP: Modular Vision-language Alignment with Identification Guarantees + +Shaoan Xie\*, Lingjing\*, Yujia Zheng\*, Yu Yao\*, Zeyu Tang\*, Eric P. Xing\*, Guangyi Chen\*, Kun Zhang\*, + +1 Carnegie Mellon University + +$^{2}$ Mohamed bin Zayed University of Artificial Intelligence + +3 University of Sydney + +* Equal contribution + +# Abstract + +Contrastive Language-Image Pre-training (CLIP) [37] has emerged as a pivotal model in computer vision and multimodal learning, achieving state-of-the-art performance at aligning visual and textual representations through contrastive learning. However, CLIP struggles with potential information misalignment in many image-text datasets and suffers from entangled representation. On the one hand, short captions for a single image in datasets like MSCOCO may describe disjoint regions in the image, leaving the model uncertain about which visual features to retain or disregard. On the other hand, directly aligning long captions with images can lead to the retention of entangled details, preventing the model from learning disentangled, atomic concepts - ultimately limiting its generalization on certain downstream tasks involving short prompts. + +In this paper, we establish theoretical conditions that enable flexible alignment between textual and visual representations across varying levels of granularity. Specifically, our framework ensures that a model can not only preserve cross-modal semantic information in its entirety but also disentangle visual representations to capture fine-grained textual concepts. Building on this foundation, we introduce SmartCLIP, a novel approach that identifies and aligns the most relevant visual and textual representations in a modular manner. Superior performance across various tasks demonstrates its capability to handle information misalignment and supports our identification theory. The code is available at https://github.com/Mid-Push/SmartCLIP. + +# 1. Introduction + +Contrastive Language-Image Pre-training (CLIP) [37] has been the cornerstone for many computer vision and machine learning tasks, such as text-to-image retrieval [2], im + +![](images/e39db5bd66f7dda691fab008f1ddff384cea3b7bfbf6338af8099258c3a4861e.jpg) + +# COCO: + +A very cute teddy bear holding a pen. + +A stuffed bear that is sitting in a chair + +A brown bear wearing a sweater next to a pen and paper. + +A brown teddy bear in a dark sweater sits on a black striped chair while holding a pen. + +# ShareGPT4u: + +In the center of the image, a light brown teddy bear is seated on a black chair with white stripes. The teddy bear, dressed in a green sweater adorned with a yellow flower, holds a blue pen in its paw, as if ready to sign a document. The chair is positioned on a gray carpeted floor. To the right of the chair, there's a stack of papers neatly arranged on the floor. The scene creates an atmosphere of anticipation, as if the teddy bear is about to make a significant decision. + +Figure 1. Depiction of two primary challenges for CLIP. (1) Information Misalignment: An image can be paired with multiple captions that describe disparate aspects - the first caption contains concepts "bear" and "pen" whereas the second only mentions "bear" and "paper". Aligning the image with both captions leads to the loss of key concepts "pen" and "paper" not shared across the captions. (2) Entangled Representations: Long, detailed captions involving multiple concepts (e.g., "chair", "pen", "flower", "floor") encourage the model to form entangled representations, hindering independent understanding of each individual concept. + +age and video understanding [6, 16, 39, 42, 46, 58], and generative models [28, 38, 40]. It aligns the representations from different modalities with a contrastive learning loss [8, 35]. Specifically, each image-caption pair in the dataset is treated as a positive pair, while negative pairs are created by matching images with captions randomly drawn from the dataset. The image and text encoders are trained with a symmetric cross-entropy loss that draws the image and text representations in each positive pair together while pulling the negative pairs' representations apart. + +Training CLIP requires vast amounts of image-text pairs, making it challenging to maintain dataset quality at such a large scale. In particular, the quality of text captions has been a key concern, prompting the development of various methods to enhance both their diversity and accu + +racy. ALIGN [15] shows that scaling up the dataset size can compensate for the noisy text supervision. BLIP models [23, 24] improve captions by incorporating additional captioning and filtering mechanisms. VE-CLIP [21] introduces a visual-enriched captioning approach to further refine caption quality. Similarly, LaCLIP [10] leverages language models to rewrite captions, while RecapCLIP [25] uses LLaMA-3 [29] to generate captions for 1.3 billion images. Despite these efforts, recent findings reveal that longer and seemingly higher-quality captions do not necessarily yield improved performance on many downstream tasks [20]. Li et al. [25] find that the text-to-image retrieval accuracy on Flickr30K drops from 84.2 to 74.1 when replacing the original captions with longer captions. + +A key issue contributing to the observed performance degradation is the information misalignment between images and their captions, a problem that becomes more pronounced when multiple captions are paired with a single image. On one hand, an image may be paired with several captions, each capturing only a partial aspect of the image. In Figure 1, aligning the image with the first caption, "A very cute teddy bear holding a pen", risks forcing the model to discard other important concepts like "chair" and "paper", which are required to align with the second and third captions in the pink text box. This misalignment between image and text introduces conflicts during standard CLIP training, leading to the loss of key visual concepts. + +On the other hand, training CLIP with long, detailed captions, as seen in recent approaches [10, 21, 25], encourages the model to learn entangled representations of multiple concepts bundled together in a single caption. Thus, it remains challenging to explicitly extract object/concept-centric representations from CLIP's visual representation. This entanglement is particularly problematic for tasks that require individual, atomic concepts or novel combinations of them, as empirically observed on short-text-to-image retrieval tasks [54]. In Figure 1, the long caption generated by ShareGPT4V [7] contains an exhaustive set of concepts such as "chair", "pen", "flower", and "floor". This aggregation can hinder the model's performance on tasks that demand the understanding of each concept individually. + +In this paper, we propose a refined approach to representation alignment in vision-language models like CLIP [37]. We frame the alignment challenge as a latent-variable identification problem and develop theoretical conditions that enable flexible alignment between textual and visual representations at different levels of granularity. Our framework enables the model to preserve the complete cross-modal information while also disentangling representations to capture fine-grained concepts, effectively addressing the misalignment and disentanglement issues discussed earlier. + +Building on these theoretical insights, we introduce SmartCLIP, a novel method that identifies and aligns vi + +sual and textual concepts in a modular manner. Specifically, we design a mask network that selects a subset of dimensions from the full representations, corresponding to only the concepts present in each specific caption. This allows the model to perform text-image alignment over the most relevant concepts modules, rather than the entire representation. We empirically demonstrate that SmartCLIP outperforms state-of-the-art models across a range of downstream tasks, showcasing its effectiveness in addressing alignment challenges. In particular, SmartCLIP significantly improves retrieval performance across text lengths, achieving $98.7\%$ accuracy (up from $78.2\%$ ) on the ShareGPT4V long text-to-image retrieval tasks, while boosting short text-to-image retrieval R1 from $56.1\%$ to $66.0\%$ . + +Our main contributions are summarized as follows. + +i We identify critical issues of information misalignment and entangled representations within the CLIP framework. To overcome these challenges, we propose a latent-variable formulation and establish theoretical conditions that guarantee the recovery of the latent variables. + +ii Building upon our theoretical findings, we propose SmartCLIP, featuring adaptive masking and a modular contrastive learning objective that facilitates the learning of disentangled, modular representations. + +iii We perform extensive experiments on a variety of tasks, including long and short text-to-image retrieval, zero-shot classification, and text-to-image generation. Smart-CLIP consistently outperforms or matches state-of-the-art models across these benchmarks, demonstrating its efficacy and validating our theoretical contributions. + +# 2. Related Work + +Vision-language models. The breakthrough of CLIP [37] has attracted significant attention from the community. SLIP [33] and DECLIP [26] propose to incorporate self-supervised learning techniques to improve the learned representation. Coca [51] introduces a decoder in addition to the contrastive learning branch. LiT [52] locks the image encoder and only finetunes the text encoder. SigLIP [53] adopts a simple sigmoid loss to handle large training batch sizes. LoTLIP [48] inserts corner tokens after the classification token to support long-text understanding. TULIP [34] replaces the absolute position embedding with the relative position embedding to support longer text understanding. ALIGN [15] demonstrates that increasing dataset size can mitigate the impact of noisy text supervision. Recent methods have been focusing on generating better captions [20, 21, 23-25, 59]. CLIP-MOE [56] introduces mixture-of-experts to CLIP. LLM2CLIP [12] augments CLIP with large language models. LongCLIP [54] extends the token constraint of CLIP from 77 to 248 and applies PCA to perform short text-to-image contrastive learning to preserve its short text capability. Llip [22] learns a text-dependent vi + +sual representation by mixing a set of learnable tokens with a cross-attention module. In contrast, SmartCLIP directly learns a single global representation that encodes all disentangled, interpretable concepts through masking. + +Latent variable identification. Learning high-level, semantic information from low-level observational data (e.g., images and text) can often be formulated as latent-variable identification problems. Though appealing, such tasks are accompanied by substantial difficulties, especially for complex real-world data distributions involving nonlinear generating functions. Recently, a wealth of papers [1, 5, 9, 11, 14, 17, 31, 32, 41, 44, 49, 55, 57] propose to overcome such obstacles by leveraging auxiliary information, such as temporal information, multiple domains, and multiple views/modalities. Especially related to our work are those that tap into the paired multi-view data to identify the shared information across available views [9, 11, 31, 32, 43, 49]. Recent work [11, 31, 32] relies on specific forms of latent variable distributions (e.g., independence or exponential family). These constraints restrict their applicability for distributions entailing complex interactions among latent variables. Prior work [9, 43] adopt more flexible assumptions on the underlying distribution and identify blocks of latent variables directly shared by two views arising from data augmentations, which is extended to a multi-view setting [49]. The problem under investigation can be viewed as a form of this multi-view setting where the paired image and text captions are considered as views sharing the semantic latent variable. Existing works [9, 43, 49] assume that views are grouped over all the data pairs and this grouping information is known so that one can learn a designated encoder for each view group. However, this view grouping information is inaccessible for our problems – for any two text captions of different images, we cannot judge whether they belong to the same view group. In our theory section, we show that by properly utilizing the data-generating process, we can learn such information directly and further achieve the desired identification results, thus generalizing existing multi-view latent variable identification results. + +# 3. Problem Formulation + +As motivated previously, we aim to 1) preserve all the semantic information shared across modalities, and 2) learn a disentangled representation that corresponds to textual concepts at diverse granularity levels. To this end, we propose the following data-generating process underlying the visual-language data distribution. + +Notations. We indicate the dimensionality of a vector with $d(\cdot)$ . We denote a subset of dimensions of a vector $\mathbf{z}$ with $[\mathbf{z}]_{\mathcal{B}}$ with the index set $\mathcal{B}$ . We define the set of indices whose corresponding values are nonzero in a vector $\mathbf{m}$ with $\mathcal{B}(\mathbf{m}) := \{i \in d(\mathbf{m}) : [\mathbf{m}]_i \neq 0\}$ . + +Data-generating processes. We depict the data-generating + +process in Figure 2 and in (1). + +$$ +\mathbf {z} _ {\mathrm {T}} := \mathbf {z} _ {\mathrm {I}} \odot \mathbf {m}; \mathbf {i} := g _ {\mathrm {I}} \left(\mathbf {z} _ {\mathrm {I}}, \epsilon_ {\mathrm {I}}\right); \mathbf {t} := g _ {\mathrm {T}} \left(\mathbf {z} _ {\mathrm {T}}, \epsilon_ {\mathrm {T}}\right). \tag {1} +$$ + +We assume that each pair of image $\mathbf{i} \in \mathcal{I} \subset \mathbb{R}^{d(\mathbf{i})}$ and text caption $\mathbf{t} \in \mathcal{T} \subset \mathbb{R}^{d(\mathbf{t})}$ originate from semantic information $\mathbf{z}_{\mathrm{I}} \in \mathcal{Z}_{\mathrm{I}} \subset \mathbb{R}^{d(\mathbf{z}_{\mathrm{I}})}$ , together with modality-specific variations $\epsilon_{\mathrm{I}}$ and $\epsilon_{\mathrm{T}}$ (e.g., illumination for images, tenses for text), through generating functions $g_{\mathrm{I}}: (\mathbf{z}_{\mathrm{I}}, \epsilon_{\mathrm{I}}) \mapsto \mathbf{i}$ and $g_{\mathrm{T}}: (\mathbf{z}_{\mathrm{T}}, \epsilon_{\mathrm{T}}) \mapsto \mathbf{t}$ respectively. We treat the text caption as continuous variables, as each word can be represented with a continuous word embedding vector in practice [3, 30]. + +As demonstrated in Figure 1, text captions of the same image often convey partial information of the entire image semantics. Thus, we associate each text caption's representation $\mathbf{z}_{\mathrm{T}} := \mathbf{m} \odot \mathbf{z}_{\mathrm{I}}$ with a binary random mask $\mathbf{m} \in \mathcal{M} \subset \{0,1\}^{d(\mathbf{z}_{\mathrm{I}})}$ that eliminates information absent in the specific caption $\mathbf{t}$ . + +![](images/b33d244b1b25d5e25cc42b572bbf56b964c3df73d939c85e22a9532eca9d9b5e.jpg) +Figure 2. The data-generating process. The text representation $\mathbf{z}_{\mathrm{T}}$ contains partial information of the vision representation $\mathbf{z}_{\mathrm{I}}$ , as indicated by masking $\mathbf{m}$ . Modal-specific information is represented as $\epsilon_{\mathrm{I}}$ and $\epsilon_{\mathrm{T}}$ . Dashed edges indicate potential statistical dependence. + +Goals. Our two goals can be formalized as follows. + +a Preserving cross-modal information: identifying the complete latent representation $\mathbf{z}_1$ . +b Disentangling concepts: identifying concepts $\mathbf{z}_{\mathrm{T}}$ associated with a given textual description $\mathbf{t}$ at different granularity levels potentially unseen during training. + +Examples. In Figure 1, the image i contains concepts "bear", "chair", and "pen", which we assume correspond to three components in the representation $\mathbf{z}_{\mathrm{I}}$ , say $[\mathbf{z}_{\mathrm{I}}]_{1}$ , $[\mathbf{z}_{\mathrm{I}}]_{2}$ , and $[\mathbf{z}_{\mathrm{I}}]_{3}$ . The first two COCO captions $\mathbf{t}^{(1)}$ and $\mathbf{t}^{(2)}$ mention only a subset of these concepts, i.e., ("bear", "pen") and ("bear", "chair") separately. Thus, the masks for these captions are $\mathbf{m}^{(1)} = [1,0,1]$ and $\mathbf{m}^{(2)} = [1,1,0]$ . The variables $\epsilon_{\mathrm{I}}$ and $\epsilon_{\mathrm{T}}$ represent modality-specific nuance factors such as illumination conditions in the image i and syntax in the text t. For Goal a, we seek to preserve the complete information $\mathbf{z}_{\mathrm{I}}$ . This involves retaining all relevant textual concepts present in the captions, namely "bear", "chair", and "pen" from captions $\mathbf{t}^{(1)}$ and $\mathbf{t}^{(2)}$ . For Goal b, we intend to disentangle the representation $\mathbf{z}_{\mathrm{I}}$ into finer concept blocks potentially unseen in the training. This includes identifying individual concepts such as "bear" in the dimen- + +sion $[\mathbf{z}_{\mathrm{I}}]_{1}$ , even if the training captions only include "bear" in combination with other concepts. + +# 4. Identification Theory + +In this section, we present the theoretical results towards Goal a and Goal b. We show that under a suitable learning objective (2), the learned variables $(\hat{\mathbf{z}}_{\mathrm{I}},\hat{\mathbf{z}}_{\mathrm{T}})$ can be identified with the corresponding true quantities $(\mathbf{z}_{\mathrm{I}},\mathbf{z}_{\mathrm{T}})$ up to certain equivalent classes. In particular, we resort to the blockwise identifiability [18, 19, 43, 49] throughout this work. This suffices for our purpose since often several dimensions jointly (i.e., a block) comprise a meaningful concept while a single dimension may not be interpretable. + +Definition 4.1 (Block-wise Identifiability). The true variable $\mathbf{v}$ is block-wise identifiable if it is related to its estimate $\hat{\mathbf{v}}$ through an invertible map $\mathbf{v} \mapsto \hat{\mathbf{v}}$ . + +The learning objective. Our estimation model consists of vision/text encoders $(f_{\mathrm{I}}, f_{\mathrm{T}})$ (smooth, invertible functions), and a masking function $\hat{\mathbf{m}}: \mathcal{T} \to \mathcal{M}$ that estimates the true mask $\mathbf{m}$ underlying a given text caption $\mathbf{t}$ . + +$$ +\begin{array}{l} \arg \min \| \hat {\mathbf {m}} (\mathbf {t}) \| _ {0}, \quad \text {s u b j e c t} \\ f _ {\mathrm {I}}, f _ {\mathrm {T}}, \hat {\mathbf {m}} \\ L _ {\mathrm {s p a r s i t y}} \\ \end{array} +$$ + +(2) + +$$ +\begin{array}{l} \arg \min \| f _ {\mathrm {I}} (\mathbf {i}) \odot \hat {\mathbf {m}} (\mathbf {t}) - f _ {\mathrm {T}} (\mathbf {t}) \|, \forall (\mathbf {i}, \mathbf {t}). \\ \underbrace {f _ {\mathrm {I}} , f _ {\mathrm {T}} , \hat {\mathbf {m}}} _ {L _ {\mathrm {a l i g n}}} \\ \end{array} +$$ + +Our learning objective (2) consists of an alignment term $L_{\mathrm{align}}$ that draws the positive pairs across modalities. The negative pairs in regular contrastive losses [8, 35, 37] can be implemented through an entropy term at the sample limit [47]. This serves the same role as the invertibility condition on the encoder models [43], which we directly assume for theoretical convenience. In Section 5, we discuss practical considerations for constructing negative pairs. We enforce sparsity regularization $L_{\mathrm{sparsity}}$ on the inferred mask $\hat{\mathbf{m}}$ to select the simplest representation. + +We introduce our key conditions in Condition 4.2 and theoretical results in Theorem 4.3. + +# Condition 4.2 (Identification Conditions). + +i [Smoothness & invertibility]: Generating functions $g_{\mathrm{I}}$ and $g_{\mathrm{T}}$ are smooth and have smooth inverses. +ii [Fully-supported $p(\mathbf{z}_{\mathrm{I}},\mathbf{m})]$ : The joint distribution over the semantic variable $\mathbf{z}_{\mathrm{I}}$ and the mask $\mathbf{m}$ is fully supported: $p(\mathbf{z}_{\mathrm{I}},\mathbf{m}) > 0$ for any $(\mathbf{z}_{\mathrm{I}},\mathbf{m})\in \mathcal{Z}_{\mathrm{I}}\times \mathcal{M}$ . + +Discussion. Condition 4.2- $i$ ensures the generating functions $(g_{\mathrm{I}}, g_{\mathrm{T}})$ preserve latent variables' information, without which the task of recovering such latent variables would be ill-posed. Practically, the high dimensionality of image + +data i offers sufficient capacity to hold all information, and the text variable t only contains information filtered through its mask m. This condition is widely employed in the latent-variable identification literature [13, 17, 18, 43]. Condition 4.2-ii prescribes that the representation $\mathbf{z}_{\mathrm{I}}$ and the mask m that marginally appear in the training distribution should also be present jointly with non-zero probability density. Interpreting the mask m as a concept selector (e.g., selecting "bear" and "pen"), this condition ensures that each concept retains its full range of variations (such as different shapes of bears and lengths of pens) across various mask selections. To satisfy this requirement, one can restrict the joint support $\mathcal{Z}_{\mathrm{I}} \times \mathcal{M}$ to an appropriate subset, ensuring that only relevant combinations of $\mathbf{z}_{\mathrm{I}}$ and m are present. Alternatively, one can enrich the caption set for each image, thereby increasing the diversity and coverage of concept combinations and filling in the joint support. This aligns with recent caption-augmentation techniques [21, 23-25, 54] as discussed in Section 1, revealing the synergy between our framework and existing efforts in the community. + +Theorem 4.3 (Concept Representation Identification). We assume the data-generating process in (1). Let $(f_{\mathrm{I}}, f_{\mathrm{T}}, \hat{\mathbf{m}})$ be an optimum of (2). Under Condition 4.2, the true representation $[\mathbf{z}]_{\tilde{\mathcal{B}}}$ is block-wise identifiable for any index set $\tilde{\mathcal{B}}$ such that $\tilde{\mathcal{B}} = \cup_{\mathbf{m} \in \mathcal{V}} \mathcal{B}(\mathbf{m})$ or $\tilde{\mathcal{B}} = \cap_{\mathbf{m} \in \mathcal{V}} \mathcal{B}(\mathbf{m})$ over any subset of masks $\mathcal{V} \subset \mathcal{M}$ . + +Concept preservation. Theorem 4.3 states that one can recover the concept block $\left[\mathbf{z}_{\mathrm{I}}\right]_{\mathcal{B}(\mathbf{m})}$ associated with each individual text caption in the dataset $\mathcal{M}$ . Furthermore, it ensures that the union of concepts $\left[\mathbf{z}_{\mathrm{I}}\right]_{\cup_{\mathbf{m} \in \nu} \mathcal{B}(\mathbf{m})}$ from any subset of text captions $\mathcal{V} \subset \mathcal{M}$ can be preserved. In the running example of Figure 1, our formulation allows us to preserve concepts ("bear", "pen", "chair") in the image representation by selectively matching them with the two captions, whereas existing models like CLIP may lose either "pen" and "chair" since they are only mentioned in one caption. Therefore, Theorem 4.3 effectively addresses Goal a. + +Concept disentanglement. The intersection operation in Theorem 4.3 empowers us to disentangle representations into potentially atomic concepts. In the example of Figure 1, we can identify the concept "bear" as the intersection of the two text captions, despite the absence of a standalone caption containing only "bear" in the dataset. Consequently, this part of the statement tackles Goal b. Our results underscore the importance of associating each image with a diverse set of captions that share overlapping concepts. + +Theoretical contribution. Theorem 4.3 extends existing theoretical frameworks [9, 43, 49]. Notably, Yao et al. [49] provide identification guarantees for shared representations over multiple views, generalizing earlier results confined to two views [9, 43]. This multi-view formulation + +is analogous to our setup when considering each group of text captions $\mathbf{t}$ associated with the same mask $\mathbf{m}$ as a distinct view group. However, prior work [49] relies on explicit knowledge of these groupings to train view-specific encoders. In contrast, our problem setting presents a greater challenge since we have no access to this grouping information. Specifically, given two captions of any different images, it is unclear whether they stem from the same mask (i.e., the same view group). Therefore, the identification guarantees in prior studies do not apply to our setting. Theorem 4.3 demonstrates that our estimation model, paired with the learning objective (2), can automatically infer the necessary grouping information (i.e., the masks). By doing so, our approach relaxes the identification conditions in previous work, enabling effective representation identification without explicit group knowledge. + +# 5. SmartCLIP: Modular Vision-language Alignment + +Drawing on the theoretical framework in Section 4, we present SmartCLIP, a modular alignment model designed to achieve Goal a and Goal b. We discuss the implementation of the learning objective (2) and the model architecture. + +Modular alignment through adaptive masking. The masking function $\hat{\mathbf{m}} (\cdot)$ is instrumental in our modular alignment objective (2). SmartCLIP implements this function with a transformer block that ingests a caption representation $\hat{\mathbf{z}}_{\mathrm{T}}$ as its input and outputs a binary vector $\hat{\mathbf{m}} (\hat{\mathbf{z}}_{\mathrm{T}})$ via a straight-through estimator [4]. + +Modular contrast construction. As discussed in Section 4, the negative pairs in regular contrastive losses [8, 35] serves a similar role as the invertibility assumption (Condition 4.2-i) [43, 47]. We denote generic image, text representations as $\mathbf{I}$ , $\mathbf{T}$ , and positive, negative pairs with $\mathbf{P}_{\mathrm{pos}}$ , $\mathbf{P}_{\mathrm{neg}}$ respectively. The canonical one-side contrastive loss $\mathcal{L}_{\mathrm{ctr}}$ [37] is defined as: + +$$ +\begin{array}{l} \mathcal {L} _ {\mathrm {c t r}} \left(\underbrace {\left(\mathbf {I} ^ {(i)} , \mathbf {T} ^ {(i)}\right)} _ {\mathbf {P} _ {\text {p o s}}}, \underbrace {\left(\mathbf {I} ^ {(i)} , \mathbf {T} ^ {(j)}\right)} _ {\mathbf {P} _ {\text {n e g}}} ^ {}\right) \tag {3} \\ = - \frac {1}{N} \sum_ {i = 1} ^ {N} \log \frac {\exp \left(\frac {\mathrm {s i m} (\mathbf {I} ^ {(i)} , \mathbf {T} ^ {(i)})}{\tau}\right)}{\sum_ {j = 1} ^ {N} \exp \left(\frac {\mathrm {s i m} (\mathbf {I} ^ {(i)} , \mathbf {T} ^ {(j)})}{\tau}\right)}, \\ \end{array} +$$ + +where we denote the sample size, temperature, and cosine similarity with $N$ , $\tau$ , and $\sin(\cdot)$ respectively. + +Following the symmetric contrastive loss for CLIP [37], our alignment loss consists of two contrastive loss terms $\mathcal{L}_{\mathrm{ctrI}}$ and $\mathcal{L}_{\mathrm{ctrT}}$ that differ in the negative pairs: + +$$ +\mathcal {L} _ {\mathrm {c t r I}} := \mathcal {L} _ {\mathrm {c t r}} \left(\mathbf {P} _ {\text {p o s}}, \mathbf {P} _ {\text {n e g I}}\right), \mathcal {L} _ {\mathrm {c t r T}} := \mathcal {L} _ {\mathrm {c t r}} \left(\mathbf {P} _ {\text {p o s}}, \mathbf {P} _ {\text {n e g T}}\right), \tag {4} +$$ + +with the positive and negative pairs defined as follows: + +$$ +\mathbf {P} _ {\text {p o s}} := \left(\hat {\mathbf {z}} _ {\mathrm {I}} ^ {(i)} \odot \hat {\mathbf {m}} \left(\hat {\mathbf {z}} _ {\mathrm {T}} ^ {(i)}\right), \hat {\mathbf {z}} _ {\mathrm {T}} ^ {(i)}\right), \tag {5} +$$ + +$$ +\mathbf {P} _ {\text {n e g I}} := \left(\hat {\mathbf {z}} _ {\mathrm {I}} ^ {(i)} \odot \hat {\mathbf {m}} \left(\hat {\mathbf {z}} _ {\mathrm {T}} ^ {(j)}\right), \hat {\mathbf {z}} _ {\mathrm {T}} ^ {(j)}\right), \tag {6} +$$ + +$$ +\mathbf {P} _ {\mathrm {n e g T}} := \left(\hat {\mathbf {z}} _ {\mathrm {I}} ^ {(j)} \odot \hat {\mathbf {m}} \left(\hat {\mathbf {z}} _ {\mathrm {T}} ^ {(i)}\right), \hat {\mathbf {z}} _ {\mathrm {T}} ^ {(i)}\right). \tag {7} +$$ + +In particular, $\mathbf{P}_{\mathrm{negI}}$ contrasts the image representation $\hat{\mathbf{z}}_{\mathrm{I}}^{(i)}$ in the positive pair with randomly sampled caption representations $\hat{\mathbf{z}}_{\mathrm{T}}^{(j)}$ (see the green region in Figure 3), whereas $\mathbf{P}_{\mathrm{negT}}$ contrasts the text representation $\hat{\mathbf{z}}_{\mathrm{T}}^{(i)}$ in the positive pair with randomly sampled image representations $\hat{\mathbf{z}}_{\mathrm{I}}^{(j)}$ (see the orange region in Figure 3). + +Sparsity penalty. We implement $L_{\mathrm{sparsity}}$ in (2) with a $\ell_1$ term for its compatibility with deep-learning training: + +$$ +\mathcal {L} _ {\text {s p a r s i t y}} = \left\| \hat {\mathbf {m}} (\mathbf {t}) \right\| _ {1}. \tag {8} +$$ + +This term ensures that the textual concepts are encoded into a minimal number of latent dimensions, promoting the disentanglement of distinct concepts across text captions. + +SmartCLIP training objective. In summary, the training objective of SmartCLIP is a weighted sum of loss terms in (4) and (8): + +$$ +\mathcal {L} = \lambda_ {\text {a l i g n}} \cdot \left(\mathcal {L} _ {\mathrm {c t r I}} + \mathcal {L} _ {\mathrm {c t r T}}\right) + \lambda_ {\text {s p a r s i t y}} \cdot \mathcal {L} _ {\text {s p a r s i t y}}, \tag {9} +$$ + +where $\lambda_{\mathrm{align}}$ and $\lambda_{\mathrm{sparsity}}$ denote the weighting coefficients. + +# 6. Experiments + +# 6.1. Setup + +Implementation details. Following Long-CLIP [54], we finetune the CLIP model [37] on ShareGPT4V [7], which contains around 1 million image-text pairs. We employ the position encoding in long-CLIP to handle 248 tokens (c.f., the 77-token limit in the original CLIP). Compared to the baseline CLIP model, we introduce a mask network $\hat{\mathbf{m}}$ . The masking network is designed as a single transformer block, which takes the text sequence embedding $\hat{\mathbf{z}}_{\mathrm{T}}$ from the text encoder. Then we add an attention-pooling layer to down-sample it to the same size as the CLIP representation, e.g., 768 in ViT-L/14. We tested including more transformer blocks in the mask network but did not observe significant improvements. Therefore, we stick to one block for faster training and inference. Unlike Long-CLIP [54] which processes all the captions for each image at each gradient step, we only sample one caption for each image, reducing our overall training time by half. Specifically, on 8 H100 GPUs, training a Vit-B/16 model for one epoch takes about 4 minutes with our model, whereas it takes around 7 minutes for Long-CLIP. After the pooling layer, we apply sigmoid to + +![](images/fe4d3b1c3b41015ef13fb0b62d8b3997d6d7437d6c55613d230d610306aa9b73.jpg) +Figure 3. The diagram of SmartCLIP. On the left, we introduce adaptive masking for alignment with different text prompts. The mask network selects which part of the image representation to be used. On the right, we present our modular contrastive objectives (4). + +Table 1. Results of short-caption text-image retrieval on the 5k COCO2017 validation set and the whole 30k Flickr30K dataset. + +
COCOFlickr30k
Image-to-TextText-to-ImageImage-to-TextText-to-Image
R@1R@5R@10R@1R@5R@10R@1R@5R@10R@1R@5R@10
B/16CLIP51.876.884.332.757.768.244.168.277.024.745.154.6
Direct Fine-tuning37.462.372.121.843.454.525.745.855.417.934.543.1
Long-CLIP [54]57.681.187.840.465.875.246.871.479.834.156.365.7
SmartCLIP (Ours)61.983.389.742.468.277.855.678.285.036.358.867.8
L/14CLIP56.179.586.835.460.170.248.572.680.828.049.358.7
Direct Fine-tuning37.963.172.223.145.155.926.046.355.617.934.943.5
Long-CLIP [54]62.885.191.246.370.879.853.477.585.341.264.172.6
SmartCLIP (Ours)66.086.292.648.573.181.763.984.490.243.866.574.8
+ +restrict the output to the range $(0,1)$ and employ straight-through estimation [4] to binarize the outputs. The training batch size is 1024 and the learning rate is $10^{-6}$ for the CLIP component and $10^{-3}$ for the mask network. + +Evaluation. We evaluate the following datasets: + +- Long text-to-image retrieval datasets: ShareGPT4V validation split [7] and Urban1k [54]. The captions for each image are long and describe details about the image. Both datasets contain 1000 text-to-image pairs. +- Short text-to-image retrieval datasets: COCO2017 validation split [27] and Flickr30K [50]. Following LongCLIP [54], we use 30K Flickr training dataset. +- Zero-shot image classification datasets. We use benchmark datasets: Country211, Fer2013, Fgvc-aircraft, GT-SRB, ImageNet, ImagetNet-V2, VOC2007, VOC2007-Multi, and SUN397. + +Baselines. In this paper, we benchmark our approach against CLIP [37] and the recent state-of-the-art model long-CLIP [54]. + +# 6.2. Comparison with CLIP Models + +We present our experimental results across three key tasks: long text-to-image retrieval, short text-to-image retrieval, and zero-shot classification. + +Long text-to-image retrieval. Table 2 showcases our method's performance on long text-to-image retrieval tasks. SmartCLIP achieves substantial improvements over baseline models, particularly the SOTA Long-CLIP, which is designed to handle long text sequences. For example, on the Urban1k dataset, SmartCLIP elevates the performance from $78.9\%$ to $90.0\%$ , marking an impressive $14\%$ boost. + +Short text-to-image retrieval. Similarly, as illustrated in Table 1, SmartCLIP significantly outperforms all baseline models across various metrics and datasets in short text-to-image retrieval tasks. The encouraging performance gains show that SmartCLIP can capture detailed information within images while simultaneously emphasizing the main semantic content. + +Zero-shot classification. To comprehensively evaluate our model's capabilities, we conduct zero-shot classification benchmarks in Table 3. Both the standard CLIP model and SmartCLIP demonstrate superior performance on different datasets. Notably, SmartCLIP exhibits a slight perfor + +Table 2. The R@1 of long-caption text-image retrieval on 1k ShareGPT4V [2] validation set and Urban-1000 dataset. The best results are bold. We cite the results from Long-CLIP [54]. + +
ShareGPT4VUrban1k
I2TT2II2TT2I
B/16CLIP [37]78.279.668.153.6
Direct Fine-tuning94.193.6--
Long-CLIP [54]94.693.378.979.5
SmartCLIP (Ours)98.798.190.087.4
L/14CLIP [37]81.884.068.752.8
Direct Fine-tuning95.395.4--
Long-CLIP [54]95.895.682.786.1
SmartCLIP (Ours)97.998.593.090.1
+ +Table 3. Zero-shot classification performance on ViT-L/14 models. When the class name is very short, i.e., a single word like ImageNet, CLIP model performs better. When the class name is a combination of several words, our method achieves better results, e.g., the road sign in GTSRB. + +
DatasetCLIPLongCLIPSmartCLIP
Country21131.828.126.9
Fer201349.057.858.6
Fgvc-aircraft31.730.630.4
GTSRB50.248.952.4
ImageNet75.372.972.5
ImageNet-V269.766.966.6
VOC200778.377.578.6
VOC2007-Multi79.082.183.7
SUN39767.572.572.1
+ +This playful food sculpture transforms cucumbers into a fearsome T-Rex dinosaur. The cucumbers form the main body, with whole cucumbers creating the legs and tail, while sliced cucumbers make up the creature's midsection. More cucumbers are cleverly cut to shape the dinosaur's head, and additional cucumbers are arranged to suggest muscular limbs. From its mouth erupts a dramatic spray of carrots, with finely julienned carrots creating the effect of fire. These bright orange carrots provide a stunning contrast against the green vegetables. Shredded carrots cascade downward like flames, while more carrots are delicately cut to create a flame-like texture. The carrots' vibrant color makes the dinosaur appear truly animated. Fresh celery leaves crown the creation, with celery fronds adding a decorative touch around the body. More celery leaves create a natural backdrop, while additional celery pieces add texture throughout. The celery's feathered appearance provides an artistic flourish to the overall design. + +![](images/19d15dd85f69958962524602adb46aa04bd415504c9d516fe0e7aea6bcc54a5d.jpg) + +![](images/5ef18a3723b1f69c850e9f13cb09f3f85a2d2df335404ca1061eb8e074934bdd.jpg) +Figure 4. Example of Long-text-to-image generation. We replace the CLIP text encoder in SDXL with different finetuned CLIP models. Given a very long text, CLIP [37] truncates the input to 77 tokens, resulting in information loss in the image. Our model learns to generate details such as celery leaves on the back of the dinosaur while other models fail. + +mance decline on ImageNet, which is expected since our model is fine-tuned on the ShareGPT4V dataset featuring long text captions, whereas ImageNet primarily consists of short, often single-word class names. However, Smart-CLIP excels on datasets with multi-word class names, such as the GTSRB dataset, where it achieves the best performance in accurately classifying road sign descriptions. + +# 6.3. Ablation Studies + +We analyze the three key components in our model: the modular alignment module, the sparsity loss and the impact of caption diversity in the data. + +Modular alignment. After introducing the mask network, we replace the standard contrastive learning with our selective alignment module. As shown in Figure 5, this change significantly improves performance. When using standard contrastive learning instead of the modular contrastive module (indicated by the purple lines), performance drops sharply. This happens because the mask information allows the network to easily separate positive pairs from negative ones, making the negative samples less informa + +tive. As a result, standard contrastive learning no longer effectively helps the model learn meaningful information. + +Alignment coefficient $\lambda_{\mathrm{align}}$ . We test the impact of the alignment coefficient $\lambda_{\mathrm{align}}$ . The results in the right panels of Figure 5 show that our method performs consistently well across a wide range of $\lambda_{\mathrm{align}}$ values, from 0.1 to 20. This indicates that our approach is robust and does not require precise tuning of $\lambda_{\mathrm{align}}$ to achieve good performance. + +Sparsity coefficient $\lambda_{\mathrm{sparsity}}$ . We also examine the sparsity coefficient $\lambda_{\mathrm{sparsity}}$ . The left panels of Figure 5 demonstrate that adding sparsity to the mask network improves performance. This supports our idea that promoting sparsity helps the model focus on the most relevant concepts, enhancing its ability to capture detailed information without being distracted by irrelevant details. + +Caption diversity. We evaluate our model's performance under varying caption diversity conditions using the COCO dataset [27]. As shown in Table 4, increasing the number of captions per image enhances performance on the Flickr30K dataset, though at the expense of degraded performance on long-text-to-image retrieval tasks. Further improvements + +![](images/f549d8ce6b37c71190963451851c7ee1cbfbca0483d7093681ca3929ae1ec123.jpg) +Figure 5. Ablation Studies on two proposed modules: selective alignment and sparsity. The baseline w.o. Modular means that we replace our modular alignment module with standard contrastive learning alignment. + +![](images/48471290c784947d1f3355796a27856dbad99276e3bbf63a360e72188f23cb1b.jpg) + +![](images/ba17102ae45429d0dd3f06743232dc50f81cb980b7674fb6b7b8081733727316.jpg) + +![](images/10e4160ab66829aef5af4b2aa5966d7f36642f4862260f0b30129b2799cc56b2.jpg) + +![](images/d07100a362ce9907c1c077baa57eb37d5dde5da2f153a0f1850b914a1dd7feca.jpg) +Figure 6. Visualization of learned representations. Given an image, we generate two captions (e.g., a zebra and a zebra and a deer) and compute cosine similarities with the same image embedding to perform binary classification for visualization using Score-CAM [45]. Compared to baseline methods, our CLIP representations are more atomic and capture differences more effectively. + +are achieved when we combine our training dataset with COCO. These results highlight both the importance of caption diversity and our method's capability to effectively handle complex text-image datasets. + +# 6.4. Additional Results + +Visualization. While our quantitative results demonstrate superior performance across various tasks, we also explore the qualitative aspects of our model by visualizing the learned representations. The visualization results are shown in Figure 6. We employ the ScoreCAM method [45] for this purpose. For each image, we generate two distinct captions, such as "a zebra" and "a zebra and a deer". We then compute the cosine similarity between the image embedding and each of the two text embeddings. These similarity scores serve as logits for a classification task, which are then input into the ScoreCAM algorithm. SmartCLIP successfully learns modular representations, accurately capturing the relevant differences between captions. + +Plug and play for text-to-image generation. One main advantage of SmartCLIP over other CLIP models trained from scratch is the low computational cost of finetuning. Additionally, our fine-tuned text encoder can replace the CLIP text encoders in large-scale models in a plug-and-play manner. Specifically, we substitute the text encoder in the SDXL [36] model with both Long-CLIP and SmartCLIP. + +
NumCapPerImg-COCOT2II2TLongT2ILongI2T
153.639.385.389.3
353.640.985.388.6
556.441.285.286.5
Ours-ShareGPT55.636.398.798.1
+COCO57.038.497.898.5
+ +Table 4. Retrieval results on Flickr30K [50] and ShareGPT-val [7] with models trained with different caption counts per image (top) and mixing long and short text-image datasets (bottom). +Table 5. Long text to image generation performance. We use the long captions (usually around 200 tokens, beyond the hard constraint 77 of the original CLIP model) to generate images with SDXL model [36], then we compare the generated images against the real images in ShareGPT4V validation split. + +
MethodKID ↓Pr ↑Re ↑F1 ↑DINO-L ↑
LongCLIP1.050.2380.7680.3630.401
SmartCLIP1.020.2580.7910.3890.414
+ +As illustrated in Figure 4, our text encoder demonstrates a superior understanding of the long text, generating detailed elements such as celery leaves in the background. Furthermore, Table 5 presents quantitative results on image generation from long captions in the ShareGPT4V validation split. Our method consistently achieves better performance across various metrics, showcasing its effectiveness in handling complex, long text inputs. + +# 7. Conclusion and Limitation + +In this work, we address the information misalignment and representation entanglement issues in existing vision-language models (e.g., CLIP). We establish theoretical conditions for effectively connecting text representations to atomic-level visual features and propose SmartCLIP, a principled, refined vision-language model. Our experimental results validate both our theoretical results and the practical effectiveness of SmartCLIP in advancing multimodal learning. Limitation. As discussed earlier, Condition 4.2-ii could be violated for datasets in which a subset images are paired with a limited number of captions compared to others. In Section 4, we discuss practical strategies to mitigate such issues. Devising alternative theoretical conditions may provide additional insights into fully utilizing all pairing information, which we leave as future work. + +# Acknowledgement + +We would like to acknowledge the support from NSF Award No. 2229881, AI Institute for Societal Decision Making (AI-SDM), the National Institutes of Health (NIH) under Contract R01HL159805, and grants from Quris AI, Florin Court Capital, and MBZUAI-WIS Joint Program. The work of L. Kong is supported in part by NSF DMS-2134080 through an award to Y. Chi. + +# References + +[1] Kartik Ahuja, Divyat Mahajan, Yixin Wang, and Yoshua Bengio. Interventional causal representation learning. In International conference on machine learning, pages 372-407. PMLR, 2023. 3 +[2] Alberto Baldrati, Lorenzo Agnolucci, Marco Bertini, and Alberto Del Bimbo. Zero-shot composed image retrieval with textual inversion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15338-15347, 2023. 1 +[3] Yoshua Bengio, Réjean Ducharme, and Pascal Vincent. A neural probabilistic language model. Advances in neural information processing systems, 13, 2000. 3 +[4] Yoshua Bengio, Nicholas Léonard, and Aaron Courville. Estimating or propagating gradients through stochastic neurons for conditional computation. arXiv preprint arXiv:1308.3432, 2013. 5, 6 +[5] Simon Buchholz, Goutham Rajendran, Elan Rosenfeld, Bryon Aragam, Bernhard Scholkopf, and Pradeep Ravikumar. Learning linear causal representations from interventions under general nonlinear mixing. Advances in Neural Information Processing Systems, 36, 2024. 3 +[6] Chang Che, Qunwei Lin, Xinyu Zhao, Jiaxin Huang, and Liqiang Yu. Enhancing multimodal understanding with clip-based image-to-text transformation. In Proceedings of the 2023 6th International Conference on Big Data Technologies, pages 414-418, 2023. 1 +[7] Lin Chen, Jinsong Li, Xiaoyi Dong, Pan Zhang, Conghui He, Jiaqi Wang, Feng Zhao, and Dahua Lin. Sharegpt4v: Improving large multi-modal models with better captions. arXiv preprint arXiv:2311.12793, 2023. 2, 5, 6, 8 +[8] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 1, 4, 5 +[9] Imant Daunhawer, Alice Bizeul, Emanuele Palumbo, Alexander Marx, and Julia E Vogt. Identifiability results for multimodal contrastive learning. In The Eleventh International Conference on Learning Representations, 2023. 3, 4 +[10] Lijie Fan, Dilip Krishnan, Phillip Isola, Dina Katabi, and Yonglong Tian. Improving clip training with language rewrites. Advances in Neural Information Processing Systems, 36, 2024. 2 +[11] Luigi Gresele, Paul K Rubenstein, Arash Mehrjou, Francesco Locatello, and Bernhard Schölkopf. The incomplete rosetta stone problem: Identifiability results for multi- + +view nonlinearica.In Uncertainty in Artificial Intelligence, pages 217-227.PMLR,2020.3 +[12] Weiquan Huang, Aooqi Wu, Yifan Yang, Xufang Luo, Yuqing Yang, Liang Hu, Qi Dai, Xiyang Dai, Dongdong Chen, Chong Luo, et al. Llm2clip: Powerful language model unlock richer visual representation. arXiv preprint arXiv:2411.04997, 2024. 2 +[13] Aapo Hyvarinen and Hiroshi Morioka. Unsupervised feature extraction by time-contrastive learning and nonlinear ica, 2016. 4 +[14] Aapo Hyvarinen, Hiroaki Sasaki, and Richard Turner. Nonlinear ica using auxiliary variables and generalized contrastive learning. In The 22nd International Conference on Artificial Intelligence and Statistics, pages 859-868. PMLR, 2019. 3 +[15] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International conference on machine learning, pages 4904-4916. PMLR, 2021. 2 +[16] Chen Ju, Tengda Han, Kunhao Zheng, Ya Zhang, and Weidi Xie. Prompting visual-language models for efficient video understanding. In European Conference on Computer Vision, pages 105-124. Springer, 2022. 1 +[17] Ilyes Khemakhem, Diederik Kingma, Ricardo Monti, and Aapo Hyvarinen. Variational autoencoders and nonlinear ica: A unifying framework. In International Conference on Artificial Intelligence and Statistics, pages 2207-2217. PMLR, 2020. 3, 4 +[18] Lingjing Kong, Shaoan Xie, Weiran Yao, Yujia Zheng, Guangyi Chen, Petar Stojanov, Victor Akinwande, and Kun Zhang. Partial disentanglement for domain adaptation. In Proceedings of the 39th International Conference on Machine Learning, pages 11455-11472. PMLR, 2022. 4 +[19] Sébastien Lachapelle, Divyat Mahajan, Ioannis Mitliagkas, and Simon Lacoste-Julien. Additive decoders for latent variables identification and cartesian-product extrapolation, 2023. 4 +[20] Zhengfeng Lai, Vasileios Saveris, Chen Chen, Hong-You Chen, Haotian Zhang, Bowen Zhang, Juan Lao Tebar, Wenze Hu, Zhe Gan, Peter Grasch, et al. Revisit large-scale image-caption data in pre-training multimodal foundation models. arXiv preprint arXiv:2410.02740, 2024. 2 +[21] Zhengfeng Lai, Haotian Zhang, Wentao Wu, Haoping Bai, Aleksei Timofeev, Xianzhi Du, Zhe Gan, Jiulong Shan, Chen-Nee Chuah, Yinfei Yang, et al. Veclip: Improving clip training via visual-enriched captions. ECCV. IEEE, page 13, 2024. 2, 4 +[22] Samuel Lavoie, Polina Kirichenko, Mark Ibrahim, Mahmoud Assran, Andrew Gordon Wilson, Aaron Courville, and Nicolas Ballas. Modeling caption diversity in contrastive vision-language pretraining. arXiv preprint arXiv:2405.00740, 2024. 2 +[23] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In Interna- + +tional conference on machine learning, pages 12888-12900. +PMLR, 2022. 2, 4 +[24] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730-19742. PMLR, 2023. 2 +[25] Xianhang Li, Haoqin Tu, Mude Hui, Zeyu Wang, Bingchen Zhao, Junfei Xiao, Sucheng Ren, Jieru Mei, Qing Liu, Huangjie Zheng, et al. What if we recapture billions of web images with llama-3? arXiv preprint arXiv:2406.08478, 2024. 2, 4 +[26] Yangguang Li, Feng Liang, Lichen Zhao, Yufeng Cui, Wanli Ouyang, Jing Shao, Fengwei Yu, and Junjie Yan. Supervision exists everywhere: A data efficient contrastive language-image pre-training paradigm. arXiv preprint arXiv:2110.05208, 2021. 2 +[27] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 6, 7 +[28] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 1 +[29] AI Meta. Introducing meta llama 3: The most capable openly available llm to date. Meta AI, 2024. 2 +[30] Tomas Mikolov, Ilya Sutskever, Kai Chen, Greg S Corrado, and Jeff Dean. Distributed representations of words and phrases and their compositionality. Advances in neural information processing systems, 26, 2013. 3 +[31] Hiroshi Morioka and Aapo Hyvarinen. Connectivity-contrastive learning: Combining causal discovery and representation learning for multimodal data. In International conference on artificial intelligence and statistics, pages 3399–3426. PMLR, 2023. 3 +[32] Hiroshi Morioka and Aapo Hyvarinen. Causal representation learning made identifiable by grouping of observational variables. In *Forty-first International Conference on Machine Learning*, 2024. 3 +[33] Norman Mu, Alexander Kirillov, David Wagner, and Saining Xie. Slip: Self-supervision meets language-image pretraining. In European conference on computer vision, pages 529-544. Springer, 2022. 2 +[34] Ivona Najdenkoska, Mohammad Mahdi Derakhshani, Yuki M Asano, Nanne van Noord, Marcel Worring, and Cees GM Snoek. Tulip: Token-length upgraded clip. arXiv preprint arXiv:2410.10034, 2024. 2 +[35] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 1, 4, 5 +[36] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 8 + +[37] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 2, 4, 5, 6, 7 +[38] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 1 +[39] Hanoona Rasheed, Muhammad Uzair Khattak, Muhammad Maaz, Salman Khan, and Fahad Shahbaz Khan. Fine-tuned clip models are efficient video learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6545-6554, 2023. 1 +[40] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1 +[41] Nils Sturma, Chandler Squires, Mathias Drton, and Caroline Uhler. Unpaired multi-domain causal representation learning, 2023. 3 +[42] Mingkang Tang, Zhanyu Wang, Zhenhua Liu, Fengyun Rao, Dian Li, and Xiu Li. Clip4caption: Clip for video caption. In Proceedings of the 29th ACM International Conference on Multimedia, pages 4858-4862, 2021. 1 +[43] Julius von Kugelgen, Yash Sharma, Luigi Gresele, Wieland Brendel, Bernhard Scholkopf, Michel Besserve, and Francesco Locatello. Self-supervised learning with data augmentations provably isolates content from style. arXiv preprint arXiv:2106.04619, 2021. 3, 4, 5 +[44] Julius von Kugelgen, Michel Besserve, Liang Wendong, Luigi Gresele, Armin Kekic, Elias Bareinboim, David Blei, and Bernhard Scholkopf. Nonparametric identifiability of causal representations from unknown interventions. Advances in Neural Information Processing Systems, 36, 2024. 3 +[45] Haofan Wang, Zifan Wang, Mengnan Du, Fan Yang, Zijian Zhang, Sirui Ding, Piotr Mardziel, and Xia Hu. Score-cam: Score-weighted visual explanations for convolutional neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 24-25, 2020. 8 +[46] Jianyi Wang, Kelvin CK Chan, and Chen Change Loy. Exploring clip for assessing the look and feel of images. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2555-2563, 2023. 1 +[47] Tongzhou Wang and Phillip Isola. Understanding contrastive representation learning through alignment and uniformity on the hypersphere. In International conference on machine learning, pages 9929-9939. PMLR, 2020. 4, 5 +[48] Wei Wu, Kecheng Zheng, Shuailei Ma, Fan Lu, Yuxin Guo, Yifei Zhang, Wei Chen, Qingpei Guo, Yujun Shen, and Zheng-Jun Zha. Lotlip: Improving language-image pre-training for long text understanding. arXiv preprint arXiv:2410.05249, 2024. 2 + +[49] Dingling Yao, Danru Xu, Sebastien Lachapelle, Sara Maglia-cane, Perouz Taslakian, Georg Martius, Julius von Kügelgen, and Francesco Locatello. Multi-view causal representation learning with partial observability. In The Twelfth International Conference on Learning Representations, 2024. 3, 4, 5 +[50] Peter Young, Alice Lai, Micah Hodosh, and Julia Hockenmaier. From image descriptions to visual denotations: New similarity metrics for semantic inference over event descriptions. Transactions of the Association for Computational Linguistics, 2:67-78, 2014. 6, 8 +[51] Jiahui Yu, Zirui Wang, Vijay Vasudevan, Legg Yeung, Mojtaba Seyedhosseini, and Yonghui Wu. Coca: Contrastive captioners are image-text foundation models. arXiv preprint arXiv:2205.01917, 2022. 2 +[52] Xiaohua Zhai, Xiao Wang, Basil Mustafa, Andreas Steiner, Daniel Keysers, Alexander Kolesnikov, and Lucas Beyer. Lit: Zero-shot transfer with locked-image text tuning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18123-18133, 2022. 2 +[53] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11975-11986, 2023. 2 +[54] Beichen Zhang, Pan Zhang, Xiaoyi Dong, Yuhang Zang, and Jiaqi Wang. Long-clip: Unlocking the long-text capability of clip. arXiv preprint arXiv:2403.15378, 2024. 2, 4, 5, 6, 7 +[55] Jiaqi Zhang, Kristjan Greenewald, Chandler Squires, Akash Srivastava, Karthikeyan Shanmugam, and Caroline Uhler. Identifiability guarantees for causal disentanglement from soft interventions. Advances in Neural Information Processing Systems, 36, 2024. 3 +[56] Jihai Zhang, Xiaoye Qu, Tong Zhu, and Yu Cheng. Clipmoe: Towards building mixture of experts for clip with diversified multiplet upcycling. arXiv preprint arXiv:2409.19291, 2024. 2 +[57] Kun Zhang, Shaoan Xie, Ignavier Ng, and Yujia Zheng. Causal representation learning from multiple distributions: A general setting. arXiv preprint arXiv:2402.05052, 2024. 3 +[58] Renrui Zhang, Ziyu Guo, Wei Zhang, Kunchang Li, Xupeng Miao, Bin Cui, Yu Qiao, Peng Gao, and Hongsheng Li. Pointclip: Point cloud understanding by clip. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8552-8562, 2022. 1 +[59] Kecheng Zheng, Yifei Zhang, Wei Wu, Fan Lu, Shuailei Ma, Xin Jin, Wei Chen, and Yujun Shen. Dreamclip: Language-image pre-training with long captions. arXiv preprint arXiv:2403.17007, 2024. 2 \ No newline at end of file diff --git a/2025/SmartCLIP_ Modular Vision-language Alignment with Identification Guarantees/images.zip b/2025/SmartCLIP_ Modular Vision-language Alignment with Identification Guarantees/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..992a9700d63c7eae4d4f65dbe5068d0fd1953988 --- /dev/null +++ b/2025/SmartCLIP_ Modular Vision-language Alignment with Identification Guarantees/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4e63b2e9b44a064f35d112b5f58f387e70d0e2b189d312a195650b6e66b4fd58 +size 488292 diff --git a/2025/SmartCLIP_ Modular Vision-language Alignment with Identification Guarantees/layout.json b/2025/SmartCLIP_ Modular Vision-language Alignment with Identification Guarantees/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..8ce4a0cbb54c13cfed1e69c3a5939c81f529a455 --- /dev/null +++ b/2025/SmartCLIP_ Modular Vision-language Alignment with Identification Guarantees/layout.json @@ -0,0 +1,10024 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 55, + 103, + 555, + 122 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 103, + 555, + 122 + ], + "spans": [ + { + "bbox": [ + 55, + 103, + 555, + 122 + ], + "type": "text", + "content": "SmartCLIP: Modular Vision-language Alignment with Identification Guarantees" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 151, + 142, + 459, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 142, + 459, + 172 + ], + "spans": [ + { + "bbox": [ + 151, + 142, + 459, + 172 + ], + "type": "text", + "content": "Shaoan Xie\\*, Lingjing\\*, Yujia Zheng\\*, Yu Yao\\*, Zeyu Tang\\*, Eric P. Xing\\*, Guangyi Chen\\*, Kun Zhang\\*," + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 229, + 173, + 381, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 229, + 173, + 381, + 185 + ], + "spans": [ + { + "bbox": [ + 229, + 173, + 381, + 185 + ], + "type": "text", + "content": "1 Carnegie Mellon University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 150, + 186, + 459, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 150, + 186, + 459, + 198 + ], + "spans": [ + { + "bbox": [ + 150, + 186, + 459, + 198 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 150, + 186, + 459, + 198 + ], + "type": "text", + "content": " Mohamed bin Zayed University of Artificial Intelligence" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 245, + 200, + 365, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 245, + 200, + 365, + 213 + ], + "spans": [ + { + "bbox": [ + 245, + 200, + 365, + 213 + ], + "type": "text", + "content": "3 University of Sydney" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 271, + 217, + 339, + 226 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 271, + 217, + 339, + 226 + ], + "spans": [ + { + "bbox": [ + 271, + 217, + 339, + 226 + ], + "type": "text", + "content": "* Equal contribution" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 151, + 255, + 200, + 268 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 255, + 200, + 268 + ], + "spans": [ + { + "bbox": [ + 151, + 255, + 200, + 268 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 54, + 281, + 297, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 281, + 297, + 461 + ], + "spans": [ + { + "bbox": [ + 54, + 281, + 297, + 461 + ], + "type": "text", + "content": "Contrastive Language-Image Pre-training (CLIP) [37] has emerged as a pivotal model in computer vision and multimodal learning, achieving state-of-the-art performance at aligning visual and textual representations through contrastive learning. However, CLIP struggles with potential information misalignment in many image-text datasets and suffers from entangled representation. On the one hand, short captions for a single image in datasets like MSCOCO may describe disjoint regions in the image, leaving the model uncertain about which visual features to retain or disregard. On the other hand, directly aligning long captions with images can lead to the retention of entangled details, preventing the model from learning disentangled, atomic concepts - ultimately limiting its generalization on certain downstream tasks involving short prompts." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 462, + 296, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 462, + 296, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 462, + 296, + 628 + ], + "type": "text", + "content": "In this paper, we establish theoretical conditions that enable flexible alignment between textual and visual representations across varying levels of granularity. Specifically, our framework ensures that a model can not only preserve cross-modal semantic information in its entirety but also disentangle visual representations to capture fine-grained textual concepts. Building on this foundation, we introduce SmartCLIP, a novel approach that identifies and aligns the most relevant visual and textual representations in a modular manner. Superior performance across various tasks demonstrates its capability to handle information misalignment and supports our identification theory. The code is available at https://github.com/Mid-Push/SmartCLIP." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 655, + 135, + 668 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 655, + 135, + 668 + ], + "spans": [ + { + "bbox": [ + 55, + 655, + 135, + 668 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": "Contrastive Language-Image Pre-training (CLIP) [37] has been the cornerstone for many computer vision and machine learning tasks, such as text-to-image retrieval [2], im" + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 321, + 258, + 406, + 326 + ], + "blocks": [ + { + "bbox": [ + 321, + 258, + 406, + 326 + ], + "lines": [ + { + "bbox": [ + 321, + 258, + 406, + 326 + ], + "spans": [ + { + "bbox": [ + 321, + 258, + 406, + 326 + ], + "type": "image", + "image_path": "e39db5bd66f7dda691fab008f1ddff384cea3b7bfbf6338af8099258c3a4861e.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 413, + 262, + 438, + 271 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 413, + 262, + 438, + 271 + ], + "spans": [ + { + "bbox": [ + 413, + 262, + 438, + 271 + ], + "type": "text", + "content": "COCO:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 413, + 272, + 518, + 280 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 413, + 272, + 518, + 280 + ], + "spans": [ + { + "bbox": [ + 413, + 272, + 518, + 280 + ], + "type": "text", + "content": "A very cute teddy bear holding a pen." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 413, + 281, + 525, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 413, + 281, + 525, + 288 + ], + "spans": [ + { + "bbox": [ + 413, + 281, + 525, + 288 + ], + "type": "text", + "content": "A stuffed bear that is sitting in a chair" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 413, + 289, + 543, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 413, + 289, + 543, + 304 + ], + "spans": [ + { + "bbox": [ + 413, + 289, + 543, + 304 + ], + "type": "text", + "content": "A brown bear wearing a sweater next to a pen and paper." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 413, + 305, + 541, + 320 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 413, + 305, + 541, + 320 + ], + "spans": [ + { + "bbox": [ + 413, + 305, + 541, + 320 + ], + "type": "text", + "content": "A brown teddy bear in a dark sweater sits on a black striped chair while holding a pen." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 327, + 330, + 378, + 338 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 330, + 378, + 338 + ], + "spans": [ + { + "bbox": [ + 327, + 330, + 378, + 338 + ], + "type": "text", + "content": "ShareGPT4u:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 325, + 340, + 543, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 340, + 543, + 395 + ], + "spans": [ + { + "bbox": [ + 325, + 340, + 543, + 395 + ], + "type": "text", + "content": "In the center of the image, a light brown teddy bear is seated on a black chair with white stripes. The teddy bear, dressed in a green sweater adorned with a yellow flower, holds a blue pen in its paw, as if ready to sign a document. The chair is positioned on a gray carpeted floor. To the right of the chair, there's a stack of papers neatly arranged on the floor. The scene creates an atmosphere of anticipation, as if the teddy bear is about to make a significant decision." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 409, + 555, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 409, + 555, + 519 + ], + "spans": [ + { + "bbox": [ + 313, + 409, + 555, + 519 + ], + "type": "text", + "content": "Figure 1. Depiction of two primary challenges for CLIP. (1) Information Misalignment: An image can be paired with multiple captions that describe disparate aspects - the first caption contains concepts \"bear\" and \"pen\" whereas the second only mentions \"bear\" and \"paper\". Aligning the image with both captions leads to the loss of key concepts \"pen\" and \"paper\" not shared across the captions. (2) Entangled Representations: Long, detailed captions involving multiple concepts (e.g., \"chair\", \"pen\", \"flower\", \"floor\") encourage the model to form entangled representations, hindering independent understanding of each individual concept." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 531, + 556, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 531, + 556, + 651 + ], + "spans": [ + { + "bbox": [ + 313, + 531, + 556, + 651 + ], + "type": "text", + "content": "age and video understanding [6, 16, 39, 42, 46, 58], and generative models [28, 38, 40]. It aligns the representations from different modalities with a contrastive learning loss [8, 35]. Specifically, each image-caption pair in the dataset is treated as a positive pair, while negative pairs are created by matching images with captions randomly drawn from the dataset. The image and text encoders are trained with a symmetric cross-entropy loss that draws the image and text representations in each positive pair together while pulling the negative pairs' representations apart." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 653, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 556, + 713 + ], + "type": "text", + "content": "Training CLIP requires vast amounts of image-text pairs, making it challenging to maintain dataset quality at such a large scale. In particular, the quality of text captions has been a key concern, prompting the development of various methods to enhance both their diversity and accu" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "29780" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 240 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 240 + ], + "type": "text", + "content": "racy. ALIGN [15] shows that scaling up the dataset size can compensate for the noisy text supervision. BLIP models [23, 24] improve captions by incorporating additional captioning and filtering mechanisms. VE-CLIP [21] introduces a visual-enriched captioning approach to further refine caption quality. Similarly, LaCLIP [10] leverages language models to rewrite captions, while RecapCLIP [25] uses LLaMA-3 [29] to generate captions for 1.3 billion images. Despite these efforts, recent findings reveal that longer and seemingly higher-quality captions do not necessarily yield improved performance on many downstream tasks [20]. Li et al. [25] find that the text-to-image retrieval accuracy on Flickr30K drops from 84.2 to 74.1 when replacing the original captions with longer captions." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 241, + 295, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 241, + 295, + 397 + ], + "spans": [ + { + "bbox": [ + 55, + 241, + 295, + 397 + ], + "type": "text", + "content": "A key issue contributing to the observed performance degradation is the information misalignment between images and their captions, a problem that becomes more pronounced when multiple captions are paired with a single image. On one hand, an image may be paired with several captions, each capturing only a partial aspect of the image. In Figure 1, aligning the image with the first caption, \"A very cute teddy bear holding a pen\", risks forcing the model to discard other important concepts like \"chair\" and \"paper\", which are required to align with the second and third captions in the pink text box. This misalignment between image and text introduces conflicts during standard CLIP training, leading to the loss of key visual concepts." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 399, + 295, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 399, + 295, + 567 + ], + "spans": [ + { + "bbox": [ + 55, + 399, + 295, + 567 + ], + "type": "text", + "content": "On the other hand, training CLIP with long, detailed captions, as seen in recent approaches [10, 21, 25], encourages the model to learn entangled representations of multiple concepts bundled together in a single caption. Thus, it remains challenging to explicitly extract object/concept-centric representations from CLIP's visual representation. This entanglement is particularly problematic for tasks that require individual, atomic concepts or novel combinations of them, as empirically observed on short-text-to-image retrieval tasks [54]. In Figure 1, the long caption generated by ShareGPT4V [7] contains an exhaustive set of concepts such as \"chair\", \"pen\", \"flower\", and \"floor\". This aggregation can hinder the model's performance on tasks that demand the understanding of each concept individually." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 568, + 295, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 568, + 295, + 687 + ], + "spans": [ + { + "bbox": [ + 55, + 568, + 295, + 687 + ], + "type": "text", + "content": "In this paper, we propose a refined approach to representation alignment in vision-language models like CLIP [37]. We frame the alignment challenge as a latent-variable identification problem and develop theoretical conditions that enable flexible alignment between textual and visual representations at different levels of granularity. Our framework enables the model to preserve the complete cross-modal information while also disentangling representations to capture fine-grained concepts, effectively addressing the misalignment and disentanglement issues discussed earlier." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "content": "Building on these theoretical insights, we introduce SmartCLIP, a novel method that identifies and aligns vi" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 553, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 239 + ], + "type": "text", + "content": "sual and textual concepts in a modular manner. Specifically, we design a mask network that selects a subset of dimensions from the full representations, corresponding to only the concepts present in each specific caption. This allows the model to perform text-image alignment over the most relevant concepts modules, rather than the entire representation. We empirically demonstrate that SmartCLIP outperforms state-of-the-art models across a range of downstream tasks, showcasing its effectiveness in addressing alignment challenges. In particular, SmartCLIP significantly improves retrieval performance across text lengths, achieving " + }, + { + "bbox": [ + 313, + 72, + 553, + 239 + ], + "type": "inline_equation", + "content": "98.7\\%" + }, + { + "bbox": [ + 313, + 72, + 553, + 239 + ], + "type": "text", + "content": " accuracy (up from " + }, + { + "bbox": [ + 313, + 72, + 553, + 239 + ], + "type": "inline_equation", + "content": "78.2\\%" + }, + { + "bbox": [ + 313, + 72, + 553, + 239 + ], + "type": "text", + "content": ") on the ShareGPT4V long text-to-image retrieval tasks, while boosting short text-to-image retrieval R1 from " + }, + { + "bbox": [ + 313, + 72, + 553, + 239 + ], + "type": "inline_equation", + "content": "56.1\\%" + }, + { + "bbox": [ + 313, + 72, + 553, + 239 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 72, + 553, + 239 + ], + "type": "inline_equation", + "content": "66.0\\%" + }, + { + "bbox": [ + 313, + 72, + 553, + 239 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 239, + 522, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 239, + 522, + 251 + ], + "spans": [ + { + "bbox": [ + 313, + 239, + 522, + 251 + ], + "type": "text", + "content": "Our main contributions are summarized as follows." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 315, + 252, + 553, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 252, + 553, + 312 + ], + "spans": [ + { + "bbox": [ + 315, + 252, + 553, + 312 + ], + "type": "text", + "content": "i We identify critical issues of information misalignment and entangled representations within the CLIP framework. To overcome these challenges, we propose a latent-variable formulation and establish theoretical conditions that guarantee the recovery of the latent variables." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 312, + 553, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 312, + 553, + 360 + ], + "spans": [ + { + "bbox": [ + 313, + 312, + 553, + 360 + ], + "type": "text", + "content": "ii Building upon our theoretical findings, we propose SmartCLIP, featuring adaptive masking and a modular contrastive learning objective that facilitates the learning of disentangled, modular representations." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 311, + 360, + 553, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 360, + 553, + 431 + ], + "spans": [ + { + "bbox": [ + 311, + 360, + 553, + 431 + ], + "type": "text", + "content": "iii We perform extensive experiments on a variety of tasks, including long and short text-to-image retrieval, zero-shot classification, and text-to-image generation. Smart-CLIP consistently outperforms or matches state-of-the-art models across these benchmarks, demonstrating its efficacy and validating our theoretical contributions." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 442, + 400, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 442, + 400, + 453 + ], + "spans": [ + { + "bbox": [ + 313, + 442, + 400, + 453 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 462, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 462, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 462, + 553, + 713 + ], + "type": "text", + "content": "Vision-language models. The breakthrough of CLIP [37] has attracted significant attention from the community. SLIP [33] and DECLIP [26] propose to incorporate self-supervised learning techniques to improve the learned representation. Coca [51] introduces a decoder in addition to the contrastive learning branch. LiT [52] locks the image encoder and only finetunes the text encoder. SigLIP [53] adopts a simple sigmoid loss to handle large training batch sizes. LoTLIP [48] inserts corner tokens after the classification token to support long-text understanding. TULIP [34] replaces the absolute position embedding with the relative position embedding to support longer text understanding. ALIGN [15] demonstrates that increasing dataset size can mitigate the impact of noisy text supervision. Recent methods have been focusing on generating better captions [20, 21, 23-25, 59]. CLIP-MOE [56] introduces mixture-of-experts to CLIP. LLM2CLIP [12] augments CLIP with large language models. LongCLIP [54] extends the token constraint of CLIP from 77 to 248 and applies PCA to perform short text-to-image contrastive learning to preserve its short text capability. Llip [22] learns a text-dependent vi" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "29781" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "content": "sual representation by mixing a set of learnable tokens with a cross-attention module. In contrast, SmartCLIP directly learns a single global representation that encodes all disentangled, interpretable concepts through masking." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 120, + 296, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 120, + 296, + 539 + ], + "spans": [ + { + "bbox": [ + 56, + 120, + 296, + 539 + ], + "type": "text", + "content": "Latent variable identification. Learning high-level, semantic information from low-level observational data (e.g., images and text) can often be formulated as latent-variable identification problems. Though appealing, such tasks are accompanied by substantial difficulties, especially for complex real-world data distributions involving nonlinear generating functions. Recently, a wealth of papers [1, 5, 9, 11, 14, 17, 31, 32, 41, 44, 49, 55, 57] propose to overcome such obstacles by leveraging auxiliary information, such as temporal information, multiple domains, and multiple views/modalities. Especially related to our work are those that tap into the paired multi-view data to identify the shared information across available views [9, 11, 31, 32, 43, 49]. Recent work [11, 31, 32] relies on specific forms of latent variable distributions (e.g., independence or exponential family). These constraints restrict their applicability for distributions entailing complex interactions among latent variables. Prior work [9, 43] adopt more flexible assumptions on the underlying distribution and identify blocks of latent variables directly shared by two views arising from data augmentations, which is extended to a multi-view setting [49]. The problem under investigation can be viewed as a form of this multi-view setting where the paired image and text captions are considered as views sharing the semantic latent variable. Existing works [9, 43, 49] assume that views are grouped over all the data pairs and this grouping information is known so that one can learn a designated encoder for each view group. However, this view grouping information is inaccessible for our problems – for any two text captions of different images, we cannot judge whether they belong to the same view group. In our theory section, we show that by properly utilizing the data-generating process, we can learn such information directly and further achieve the desired identification results, thus generalizing existing multi-view latent variable identification results." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 550, + 181, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 550, + 181, + 562 + ], + "spans": [ + { + "bbox": [ + 55, + 550, + 181, + 562 + ], + "type": "text", + "content": "3. Problem Formulation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 570, + 295, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 295, + 641 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 295, + 641 + ], + "type": "text", + "content": "As motivated previously, we aim to 1) preserve all the semantic information shared across modalities, and 2) learn a disentangled representation that corresponds to textual concepts at diverse granularity levels. To this end, we propose the following data-generating process underlying the visual-language data distribution." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 642, + 295, + 702 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 295, + 702 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 295, + 702 + ], + "type": "text", + "content": "Notations. We indicate the dimensionality of a vector with " + }, + { + "bbox": [ + 55, + 642, + 295, + 702 + ], + "type": "inline_equation", + "content": "d(\\cdot)" + }, + { + "bbox": [ + 55, + 642, + 295, + 702 + ], + "type": "text", + "content": ". We denote a subset of dimensions of a vector " + }, + { + "bbox": [ + 55, + 642, + 295, + 702 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 55, + 642, + 295, + 702 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 55, + 642, + 295, + 702 + ], + "type": "inline_equation", + "content": "[\\mathbf{z}]_{\\mathcal{B}}" + }, + { + "bbox": [ + 55, + 642, + 295, + 702 + ], + "type": "text", + "content": " with the index set " + }, + { + "bbox": [ + 55, + 642, + 295, + 702 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 55, + 642, + 295, + 702 + ], + "type": "text", + "content": ". We define the set of indices whose corresponding values are nonzero in a vector " + }, + { + "bbox": [ + 55, + 642, + 295, + 702 + ], + "type": "inline_equation", + "content": "\\mathbf{m}" + }, + { + "bbox": [ + 55, + 642, + 295, + 702 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 55, + 642, + 295, + 702 + ], + "type": "inline_equation", + "content": "\\mathcal{B}(\\mathbf{m}) := \\{i \\in d(\\mathbf{m}) : [\\mathbf{m}]_i \\neq 0\\}" + }, + { + "bbox": [ + 55, + 642, + 295, + 702 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 702, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 702, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 702, + 295, + 714 + ], + "type": "text", + "content": "Data-generating processes. We depict the data-generating" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 437, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 437, + 84 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 437, + 84 + ], + "type": "text", + "content": "process in Figure 2 and in (1)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 329, + 95, + 553, + 106 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 329, + 95, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 329, + 95, + 553, + 106 + ], + "type": "interline_equation", + "content": "\\mathbf {z} _ {\\mathrm {T}} := \\mathbf {z} _ {\\mathrm {I}} \\odot \\mathbf {m}; \\mathbf {i} := g _ {\\mathrm {I}} \\left(\\mathbf {z} _ {\\mathrm {I}}, \\epsilon_ {\\mathrm {I}}\\right); \\mathbf {t} := g _ {\\mathrm {T}} \\left(\\mathbf {z} _ {\\mathrm {T}}, \\epsilon_ {\\mathrm {T}}\\right). \\tag {1}", + "image_path": "630df460e36417af3940c21d157fb7b61231d952089e2589ae8da59039eac1b7.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 117, + 553, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 117, + 553, + 213 + ], + "spans": [ + { + "bbox": [ + 313, + 117, + 553, + 213 + ], + "type": "text", + "content": "We assume that each pair of image " + }, + { + "bbox": [ + 313, + 117, + 553, + 213 + ], + "type": "inline_equation", + "content": "\\mathbf{i} \\in \\mathcal{I} \\subset \\mathbb{R}^{d(\\mathbf{i})}" + }, + { + "bbox": [ + 313, + 117, + 553, + 213 + ], + "type": "text", + "content": " and text caption " + }, + { + "bbox": [ + 313, + 117, + 553, + 213 + ], + "type": "inline_equation", + "content": "\\mathbf{t} \\in \\mathcal{T} \\subset \\mathbb{R}^{d(\\mathbf{t})}" + }, + { + "bbox": [ + 313, + 117, + 553, + 213 + ], + "type": "text", + "content": " originate from semantic information " + }, + { + "bbox": [ + 313, + 117, + 553, + 213 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{\\mathrm{I}} \\in \\mathcal{Z}_{\\mathrm{I}} \\subset \\mathbb{R}^{d(\\mathbf{z}_{\\mathrm{I}})}" + }, + { + "bbox": [ + 313, + 117, + 553, + 213 + ], + "type": "text", + "content": ", together with modality-specific variations " + }, + { + "bbox": [ + 313, + 117, + 553, + 213 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\mathrm{I}}" + }, + { + "bbox": [ + 313, + 117, + 553, + 213 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 117, + 553, + 213 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\mathrm{T}}" + }, + { + "bbox": [ + 313, + 117, + 553, + 213 + ], + "type": "text", + "content": " (e.g., illumination for images, tenses for text), through generating functions " + }, + { + "bbox": [ + 313, + 117, + 553, + 213 + ], + "type": "inline_equation", + "content": "g_{\\mathrm{I}}: (\\mathbf{z}_{\\mathrm{I}}, \\epsilon_{\\mathrm{I}}) \\mapsto \\mathbf{i}" + }, + { + "bbox": [ + 313, + 117, + 553, + 213 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 117, + 553, + 213 + ], + "type": "inline_equation", + "content": "g_{\\mathrm{T}}: (\\mathbf{z}_{\\mathrm{T}}, \\epsilon_{\\mathrm{T}}) \\mapsto \\mathbf{t}" + }, + { + "bbox": [ + 313, + 117, + 553, + 213 + ], + "type": "text", + "content": " respectively. We treat the text caption as continuous variables, as each word can be represented with a continuous word embedding vector in practice [3, 30]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 214, + 555, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 214, + 555, + 285 + ], + "spans": [ + { + "bbox": [ + 313, + 214, + 555, + 285 + ], + "type": "text", + "content": "As demonstrated in Figure 1, text captions of the same image often convey partial information of the entire image semantics. Thus, we associate each text caption's representation " + }, + { + "bbox": [ + 313, + 214, + 555, + 285 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{\\mathrm{T}} := \\mathbf{m} \\odot \\mathbf{z}_{\\mathrm{I}}" + }, + { + "bbox": [ + 313, + 214, + 555, + 285 + ], + "type": "text", + "content": " with a binary random mask " + }, + { + "bbox": [ + 313, + 214, + 555, + 285 + ], + "type": "inline_equation", + "content": "\\mathbf{m} \\in \\mathcal{M} \\subset \\{0,1\\}^{d(\\mathbf{z}_{\\mathrm{I}})}" + }, + { + "bbox": [ + 313, + 214, + 555, + 285 + ], + "type": "text", + "content": " that eliminates information absent in the specific caption " + }, + { + "bbox": [ + 313, + 214, + 555, + 285 + ], + "type": "inline_equation", + "content": "\\mathbf{t}" + }, + { + "bbox": [ + 313, + 214, + 555, + 285 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 345, + 295, + 524, + 389 + ], + "blocks": [ + { + "bbox": [ + 345, + 295, + 524, + 389 + ], + "lines": [ + { + "bbox": [ + 345, + 295, + 524, + 389 + ], + "spans": [ + { + "bbox": [ + 345, + 295, + 524, + 389 + ], + "type": "image", + "image_path": "b33d244b1b25d5e25cc42b572bbf56b964c3df73d939c85e22a9532eca9d9b5e.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 397, + 554, + 437 + ], + "lines": [ + { + "bbox": [ + 313, + 397, + 554, + 437 + ], + "spans": [ + { + "bbox": [ + 313, + 397, + 554, + 437 + ], + "type": "text", + "content": "Figure 2. The data-generating process. The text representation " + }, + { + "bbox": [ + 313, + 397, + 554, + 437 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{\\mathrm{T}}" + }, + { + "bbox": [ + 313, + 397, + 554, + 437 + ], + "type": "text", + "content": " contains partial information of the vision representation " + }, + { + "bbox": [ + 313, + 397, + 554, + 437 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{\\mathrm{I}}" + }, + { + "bbox": [ + 313, + 397, + 554, + 437 + ], + "type": "text", + "content": ", as indicated by masking " + }, + { + "bbox": [ + 313, + 397, + 554, + 437 + ], + "type": "inline_equation", + "content": "\\mathbf{m}" + }, + { + "bbox": [ + 313, + 397, + 554, + 437 + ], + "type": "text", + "content": ". Modal-specific information is represented as " + }, + { + "bbox": [ + 313, + 397, + 554, + 437 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\mathrm{I}}" + }, + { + "bbox": [ + 313, + 397, + 554, + 437 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 397, + 554, + 437 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\mathrm{T}}" + }, + { + "bbox": [ + 313, + 397, + 554, + 437 + ], + "type": "text", + "content": ". Dashed edges indicate potential statistical dependence." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 449, + 522, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 449, + 522, + 460 + ], + "spans": [ + { + "bbox": [ + 313, + 449, + 522, + 460 + ], + "type": "text", + "content": "Goals. Our two goals can be formalized as follows." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 462, + 553, + 521 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 314, + 462, + 553, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 462, + 553, + 485 + ], + "spans": [ + { + "bbox": [ + 314, + 462, + 553, + 485 + ], + "type": "text", + "content": "a Preserving cross-modal information: identifying the complete latent representation " + }, + { + "bbox": [ + 314, + 462, + 553, + 485 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_1" + }, + { + "bbox": [ + 314, + 462, + 553, + 485 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 486, + 553, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 486, + 553, + 521 + ], + "spans": [ + { + "bbox": [ + 313, + 486, + 553, + 521 + ], + "type": "text", + "content": "b Disentangling concepts: identifying concepts " + }, + { + "bbox": [ + 313, + 486, + 553, + 521 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{\\mathrm{T}}" + }, + { + "bbox": [ + 313, + 486, + 553, + 521 + ], + "type": "text", + "content": " associated with a given textual description " + }, + { + "bbox": [ + 313, + 486, + 553, + 521 + ], + "type": "inline_equation", + "content": "\\mathbf{t}" + }, + { + "bbox": [ + 313, + 486, + 553, + 521 + ], + "type": "text", + "content": " at different granularity levels potentially unseen during training." + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 522, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 522, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 522, + 555, + 713 + ], + "type": "text", + "content": "Examples. In Figure 1, the image i contains concepts \"bear\", \"chair\", and \"pen\", which we assume correspond to three components in the representation " + }, + { + "bbox": [ + 313, + 522, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{\\mathrm{I}}" + }, + { + "bbox": [ + 313, + 522, + 555, + 713 + ], + "type": "text", + "content": ", say " + }, + { + "bbox": [ + 313, + 522, + 555, + 713 + ], + "type": "inline_equation", + "content": "[\\mathbf{z}_{\\mathrm{I}}]_{1}" + }, + { + "bbox": [ + 313, + 522, + 555, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 522, + 555, + 713 + ], + "type": "inline_equation", + "content": "[\\mathbf{z}_{\\mathrm{I}}]_{2}" + }, + { + "bbox": [ + 313, + 522, + 555, + 713 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 522, + 555, + 713 + ], + "type": "inline_equation", + "content": "[\\mathbf{z}_{\\mathrm{I}}]_{3}" + }, + { + "bbox": [ + 313, + 522, + 555, + 713 + ], + "type": "text", + "content": ". The first two COCO captions " + }, + { + "bbox": [ + 313, + 522, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{t}^{(1)}" + }, + { + "bbox": [ + 313, + 522, + 555, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 522, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{t}^{(2)}" + }, + { + "bbox": [ + 313, + 522, + 555, + 713 + ], + "type": "text", + "content": " mention only a subset of these concepts, i.e., (\"bear\", \"pen\") and (\"bear\", \"chair\") separately. Thus, the masks for these captions are " + }, + { + "bbox": [ + 313, + 522, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{m}^{(1)} = [1,0,1]" + }, + { + "bbox": [ + 313, + 522, + 555, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 522, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{m}^{(2)} = [1,1,0]" + }, + { + "bbox": [ + 313, + 522, + 555, + 713 + ], + "type": "text", + "content": ". The variables " + }, + { + "bbox": [ + 313, + 522, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\mathrm{I}}" + }, + { + "bbox": [ + 313, + 522, + 555, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 522, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\mathrm{T}}" + }, + { + "bbox": [ + 313, + 522, + 555, + 713 + ], + "type": "text", + "content": " represent modality-specific nuance factors such as illumination conditions in the image i and syntax in the text t. For Goal a, we seek to preserve the complete information " + }, + { + "bbox": [ + 313, + 522, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{\\mathrm{I}}" + }, + { + "bbox": [ + 313, + 522, + 555, + 713 + ], + "type": "text", + "content": ". This involves retaining all relevant textual concepts present in the captions, namely \"bear\", \"chair\", and \"pen\" from captions " + }, + { + "bbox": [ + 313, + 522, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{t}^{(1)}" + }, + { + "bbox": [ + 313, + 522, + 555, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 522, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{t}^{(2)}" + }, + { + "bbox": [ + 313, + 522, + 555, + 713 + ], + "type": "text", + "content": ". For Goal b, we intend to disentangle the representation " + }, + { + "bbox": [ + 313, + 522, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{\\mathrm{I}}" + }, + { + "bbox": [ + 313, + 522, + 555, + 713 + ], + "type": "text", + "content": " into finer concept blocks potentially unseen in the training. This includes identifying individual concepts such as \"bear\" in the dimen-" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "29782" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 97 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 97 + ], + "type": "text", + "content": "sion " + }, + { + "bbox": [ + 55, + 72, + 297, + 97 + ], + "type": "inline_equation", + "content": "[\\mathbf{z}_{\\mathrm{I}}]_{1}" + }, + { + "bbox": [ + 55, + 72, + 297, + 97 + ], + "type": "text", + "content": ", even if the training captions only include \"bear\" in combination with other concepts." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 107, + 180, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 107, + 180, + 121 + ], + "spans": [ + { + "bbox": [ + 55, + 107, + 180, + 121 + ], + "type": "text", + "content": "4. Identification Theory" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 128, + 296, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 128, + 296, + 236 + ], + "spans": [ + { + "bbox": [ + 54, + 128, + 296, + 236 + ], + "type": "text", + "content": "In this section, we present the theoretical results towards Goal a and Goal b. We show that under a suitable learning objective (2), the learned variables " + }, + { + "bbox": [ + 54, + 128, + 296, + 236 + ], + "type": "inline_equation", + "content": "(\\hat{\\mathbf{z}}_{\\mathrm{I}},\\hat{\\mathbf{z}}_{\\mathrm{T}})" + }, + { + "bbox": [ + 54, + 128, + 296, + 236 + ], + "type": "text", + "content": " can be identified with the corresponding true quantities " + }, + { + "bbox": [ + 54, + 128, + 296, + 236 + ], + "type": "inline_equation", + "content": "(\\mathbf{z}_{\\mathrm{I}},\\mathbf{z}_{\\mathrm{T}})" + }, + { + "bbox": [ + 54, + 128, + 296, + 236 + ], + "type": "text", + "content": " up to certain equivalent classes. In particular, we resort to the blockwise identifiability [18, 19, 43, 49] throughout this work. This suffices for our purpose since often several dimensions jointly (i.e., a block) comprise a meaningful concept while a single dimension may not be interpretable." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 247, + 296, + 284 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 247, + 296, + 284 + ], + "spans": [ + { + "bbox": [ + 55, + 247, + 296, + 284 + ], + "type": "text", + "content": "Definition 4.1 (Block-wise Identifiability). The true variable " + }, + { + "bbox": [ + 55, + 247, + 296, + 284 + ], + "type": "inline_equation", + "content": "\\mathbf{v}" + }, + { + "bbox": [ + 55, + 247, + 296, + 284 + ], + "type": "text", + "content": " is block-wise identifiable if it is related to its estimate " + }, + { + "bbox": [ + 55, + 247, + 296, + 284 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{v}}" + }, + { + "bbox": [ + 55, + 247, + 296, + 284 + ], + "type": "text", + "content": " through an invertible map " + }, + { + "bbox": [ + 55, + 247, + 296, + 284 + ], + "type": "inline_equation", + "content": "\\mathbf{v} \\mapsto \\hat{\\mathbf{v}}" + }, + { + "bbox": [ + 55, + 247, + 296, + 284 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 292, + 296, + 341 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 292, + 296, + 341 + ], + "spans": [ + { + "bbox": [ + 55, + 292, + 296, + 341 + ], + "type": "text", + "content": "The learning objective. Our estimation model consists of vision/text encoders " + }, + { + "bbox": [ + 55, + 292, + 296, + 341 + ], + "type": "inline_equation", + "content": "(f_{\\mathrm{I}}, f_{\\mathrm{T}})" + }, + { + "bbox": [ + 55, + 292, + 296, + 341 + ], + "type": "text", + "content": " (smooth, invertible functions), and a masking function " + }, + { + "bbox": [ + 55, + 292, + 296, + 341 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{m}}: \\mathcal{T} \\to \\mathcal{M}" + }, + { + "bbox": [ + 55, + 292, + 296, + 341 + ], + "type": "text", + "content": " that estimates the true mask " + }, + { + "bbox": [ + 55, + 292, + 296, + 341 + ], + "type": "inline_equation", + "content": "\\mathbf{m}" + }, + { + "bbox": [ + 55, + 292, + 296, + 341 + ], + "type": "text", + "content": " underlying a given text caption " + }, + { + "bbox": [ + 55, + 292, + 296, + 341 + ], + "type": "inline_equation", + "content": "\\mathbf{t}" + }, + { + "bbox": [ + 55, + 292, + 296, + 341 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 91, + 349, + 221, + 384 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 349, + 221, + 384 + ], + "spans": [ + { + "bbox": [ + 91, + 349, + 221, + 384 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\arg \\min \\| \\hat {\\mathbf {m}} (\\mathbf {t}) \\| _ {0}, \\quad \\text {s u b j e c t} \\\\ f _ {\\mathrm {I}}, f _ {\\mathrm {T}}, \\hat {\\mathbf {m}} \\\\ L _ {\\mathrm {s p a r s i t y}} \\\\ \\end{array}", + "image_path": "e6dbebacd3238035d4456cf16208da25715a4b912e1d440267d270cf7956f04b.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 281, + 381, + 294, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 281, + 381, + 294, + 392 + ], + "spans": [ + { + "bbox": [ + 281, + 381, + 294, + 392 + ], + "type": "text", + "content": "(2)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 92, + 388, + 260, + 423 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 92, + 388, + 260, + 423 + ], + "spans": [ + { + "bbox": [ + 92, + 388, + 260, + 423 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\arg \\min \\| f _ {\\mathrm {I}} (\\mathbf {i}) \\odot \\hat {\\mathbf {m}} (\\mathbf {t}) - f _ {\\mathrm {T}} (\\mathbf {t}) \\|, \\forall (\\mathbf {i}, \\mathbf {t}). \\\\ \\underbrace {f _ {\\mathrm {I}} , f _ {\\mathrm {T}} , \\hat {\\mathbf {m}}} _ {L _ {\\mathrm {a l i g n}}} \\\\ \\end{array}", + "image_path": "8b153a5c7c5408aba3fa2271d75ba5ae5e26ed73da1c71db385062f01e4fe9ae.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 433, + 296, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 433, + 296, + 552 + ], + "spans": [ + { + "bbox": [ + 55, + 433, + 296, + 552 + ], + "type": "text", + "content": "Our learning objective (2) consists of an alignment term " + }, + { + "bbox": [ + 55, + 433, + 296, + 552 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{align}}" + }, + { + "bbox": [ + 55, + 433, + 296, + 552 + ], + "type": "text", + "content": " that draws the positive pairs across modalities. The negative pairs in regular contrastive losses [8, 35, 37] can be implemented through an entropy term at the sample limit [47]. This serves the same role as the invertibility condition on the encoder models [43], which we directly assume for theoretical convenience. In Section 5, we discuss practical considerations for constructing negative pairs. We enforce sparsity regularization " + }, + { + "bbox": [ + 55, + 433, + 296, + 552 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{sparsity}}" + }, + { + "bbox": [ + 55, + 433, + 296, + 552 + ], + "type": "text", + "content": " on the inferred mask " + }, + { + "bbox": [ + 55, + 433, + 296, + 552 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{m}}" + }, + { + "bbox": [ + 55, + 433, + 296, + 552 + ], + "type": "text", + "content": " to select the simplest representation." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 552, + 296, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 552, + 296, + 576 + ], + "spans": [ + { + "bbox": [ + 55, + 552, + 296, + 576 + ], + "type": "text", + "content": "We introduce our key conditions in Condition 4.2 and theoretical results in Theorem 4.3." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 584, + 228, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 584, + 228, + 597 + ], + "spans": [ + { + "bbox": [ + 55, + 584, + 228, + 597 + ], + "type": "text", + "content": "Condition 4.2 (Identification Conditions)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 66, + 597, + 295, + 658 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 67, + 597, + 294, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 597, + 294, + 620 + ], + "spans": [ + { + "bbox": [ + 67, + 597, + 294, + 620 + ], + "type": "text", + "content": "i [Smoothness & invertibility]: Generating functions " + }, + { + "bbox": [ + 67, + 597, + 294, + 620 + ], + "type": "inline_equation", + "content": "g_{\\mathrm{I}}" + }, + { + "bbox": [ + 67, + 597, + 294, + 620 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 67, + 597, + 294, + 620 + ], + "type": "inline_equation", + "content": "g_{\\mathrm{T}}" + }, + { + "bbox": [ + 67, + 597, + 294, + 620 + ], + "type": "text", + "content": " are smooth and have smooth inverses." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 66, + 622, + 295, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 622, + 295, + 658 + ], + "spans": [ + { + "bbox": [ + 66, + 622, + 295, + 658 + ], + "type": "text", + "content": "ii [Fully-supported " + }, + { + "bbox": [ + 66, + 622, + 295, + 658 + ], + "type": "inline_equation", + "content": "p(\\mathbf{z}_{\\mathrm{I}},\\mathbf{m})]" + }, + { + "bbox": [ + 66, + 622, + 295, + 658 + ], + "type": "text", + "content": ": The joint distribution over the semantic variable " + }, + { + "bbox": [ + 66, + 622, + 295, + 658 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{\\mathrm{I}}" + }, + { + "bbox": [ + 66, + 622, + 295, + 658 + ], + "type": "text", + "content": " and the mask " + }, + { + "bbox": [ + 66, + 622, + 295, + 658 + ], + "type": "inline_equation", + "content": "\\mathbf{m}" + }, + { + "bbox": [ + 66, + 622, + 295, + 658 + ], + "type": "text", + "content": " is fully supported: " + }, + { + "bbox": [ + 66, + 622, + 295, + 658 + ], + "type": "inline_equation", + "content": "p(\\mathbf{z}_{\\mathrm{I}},\\mathbf{m}) > 0" + }, + { + "bbox": [ + 66, + 622, + 295, + 658 + ], + "type": "text", + "content": " for any " + }, + { + "bbox": [ + 66, + 622, + 295, + 658 + ], + "type": "inline_equation", + "content": "(\\mathbf{z}_{\\mathrm{I}},\\mathbf{m})\\in \\mathcal{Z}_{\\mathrm{I}}\\times \\mathcal{M}" + }, + { + "bbox": [ + 66, + 622, + 295, + 658 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "Discussion. Condition 4.2-" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": " ensures the generating functions " + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "(g_{\\mathrm{I}}, g_{\\mathrm{T}})" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": " preserve latent variables' information, without which the task of recovering such latent variables would be ill-posed. Practically, the high dimensionality of image" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 72, + 555, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 312 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 312 + ], + "type": "text", + "content": "data i offers sufficient capacity to hold all information, and the text variable t only contains information filtered through its mask m. This condition is widely employed in the latent-variable identification literature [13, 17, 18, 43]. Condition 4.2-ii prescribes that the representation " + }, + { + "bbox": [ + 313, + 72, + 555, + 312 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{\\mathrm{I}}" + }, + { + "bbox": [ + 313, + 72, + 555, + 312 + ], + "type": "text", + "content": " and the mask m that marginally appear in the training distribution should also be present jointly with non-zero probability density. Interpreting the mask m as a concept selector (e.g., selecting \"bear\" and \"pen\"), this condition ensures that each concept retains its full range of variations (such as different shapes of bears and lengths of pens) across various mask selections. To satisfy this requirement, one can restrict the joint support " + }, + { + "bbox": [ + 313, + 72, + 555, + 312 + ], + "type": "inline_equation", + "content": "\\mathcal{Z}_{\\mathrm{I}} \\times \\mathcal{M}" + }, + { + "bbox": [ + 313, + 72, + 555, + 312 + ], + "type": "text", + "content": " to an appropriate subset, ensuring that only relevant combinations of " + }, + { + "bbox": [ + 313, + 72, + 555, + 312 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{\\mathrm{I}}" + }, + { + "bbox": [ + 313, + 72, + 555, + 312 + ], + "type": "text", + "content": " and m are present. Alternatively, one can enrich the caption set for each image, thereby increasing the diversity and coverage of concept combinations and filling in the joint support. This aligns with recent caption-augmentation techniques [21, 23-25, 54] as discussed in Section 1, revealing the synergy between our framework and existing efforts in the community." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 318, + 556, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 318, + 556, + 392 + ], + "spans": [ + { + "bbox": [ + 313, + 318, + 556, + 392 + ], + "type": "text", + "content": "Theorem 4.3 (Concept Representation Identification). We assume the data-generating process in (1). Let " + }, + { + "bbox": [ + 313, + 318, + 556, + 392 + ], + "type": "inline_equation", + "content": "(f_{\\mathrm{I}}, f_{\\mathrm{T}}, \\hat{\\mathbf{m}})" + }, + { + "bbox": [ + 313, + 318, + 556, + 392 + ], + "type": "text", + "content": " be an optimum of (2). Under Condition 4.2, the true representation " + }, + { + "bbox": [ + 313, + 318, + 556, + 392 + ], + "type": "inline_equation", + "content": "[\\mathbf{z}]_{\\tilde{\\mathcal{B}}}" + }, + { + "bbox": [ + 313, + 318, + 556, + 392 + ], + "type": "text", + "content": " is block-wise identifiable for any index set " + }, + { + "bbox": [ + 313, + 318, + 556, + 392 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathcal{B}}" + }, + { + "bbox": [ + 313, + 318, + 556, + 392 + ], + "type": "text", + "content": " such that " + }, + { + "bbox": [ + 313, + 318, + 556, + 392 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathcal{B}} = \\cup_{\\mathbf{m} \\in \\mathcal{V}} \\mathcal{B}(\\mathbf{m})" + }, + { + "bbox": [ + 313, + 318, + 556, + 392 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 313, + 318, + 556, + 392 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathcal{B}} = \\cap_{\\mathbf{m} \\in \\mathcal{V}} \\mathcal{B}(\\mathbf{m})" + }, + { + "bbox": [ + 313, + 318, + 556, + 392 + ], + "type": "text", + "content": " over any subset of masks " + }, + { + "bbox": [ + 313, + 318, + 556, + 392 + ], + "type": "inline_equation", + "content": "\\mathcal{V} \\subset \\mathcal{M}" + }, + { + "bbox": [ + 313, + 318, + 556, + 392 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 398, + 554, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 398, + 554, + 529 + ], + "spans": [ + { + "bbox": [ + 313, + 398, + 554, + 529 + ], + "type": "text", + "content": "Concept preservation. Theorem 4.3 states that one can recover the concept block " + }, + { + "bbox": [ + 313, + 398, + 554, + 529 + ], + "type": "inline_equation", + "content": "\\left[\\mathbf{z}_{\\mathrm{I}}\\right]_{\\mathcal{B}(\\mathbf{m})}" + }, + { + "bbox": [ + 313, + 398, + 554, + 529 + ], + "type": "text", + "content": " associated with each individual text caption in the dataset " + }, + { + "bbox": [ + 313, + 398, + 554, + 529 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 313, + 398, + 554, + 529 + ], + "type": "text", + "content": ". Furthermore, it ensures that the union of concepts " + }, + { + "bbox": [ + 313, + 398, + 554, + 529 + ], + "type": "inline_equation", + "content": "\\left[\\mathbf{z}_{\\mathrm{I}}\\right]_{\\cup_{\\mathbf{m} \\in \\nu} \\mathcal{B}(\\mathbf{m})}" + }, + { + "bbox": [ + 313, + 398, + 554, + 529 + ], + "type": "text", + "content": " from any subset of text captions " + }, + { + "bbox": [ + 313, + 398, + 554, + 529 + ], + "type": "inline_equation", + "content": "\\mathcal{V} \\subset \\mathcal{M}" + }, + { + "bbox": [ + 313, + 398, + 554, + 529 + ], + "type": "text", + "content": " can be preserved. In the running example of Figure 1, our formulation allows us to preserve concepts (\"bear\", \"pen\", \"chair\") in the image representation by selectively matching them with the two captions, whereas existing models like CLIP may lose either \"pen\" and \"chair\" since they are only mentioned in one caption. Therefore, Theorem 4.3 effectively addresses Goal a." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 529, + 554, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 529, + 554, + 636 + ], + "spans": [ + { + "bbox": [ + 313, + 529, + 554, + 636 + ], + "type": "text", + "content": "Concept disentanglement. The intersection operation in Theorem 4.3 empowers us to disentangle representations into potentially atomic concepts. In the example of Figure 1, we can identify the concept \"bear\" as the intersection of the two text captions, despite the absence of a standalone caption containing only \"bear\" in the dataset. Consequently, this part of the statement tackles Goal b. Our results underscore the importance of associating each image with a diverse set of captions that share overlapping concepts." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 637, + 554, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 637, + 554, + 696 + ], + "spans": [ + { + "bbox": [ + 313, + 637, + 554, + 696 + ], + "type": "text", + "content": "Theoretical contribution. Theorem 4.3 extends existing theoretical frameworks [9, 43, 49]. Notably, Yao et al. [49] provide identification guarantees for shared representations over multiple views, generalizing earlier results confined to two views [9, 43]. This multi-view formulation" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 324, + 702, + 545, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 702, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 324, + 702, + 545, + 713 + ], + "type": "text", + "content": "1We refer to a text caption with its mask m to simplify the notation." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "29783" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 264 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 264 + ], + "type": "text", + "content": "is analogous to our setup when considering each group of text captions " + }, + { + "bbox": [ + 55, + 72, + 296, + 264 + ], + "type": "inline_equation", + "content": "\\mathbf{t}" + }, + { + "bbox": [ + 55, + 72, + 296, + 264 + ], + "type": "text", + "content": " associated with the same mask " + }, + { + "bbox": [ + 55, + 72, + 296, + 264 + ], + "type": "inline_equation", + "content": "\\mathbf{m}" + }, + { + "bbox": [ + 55, + 72, + 296, + 264 + ], + "type": "text", + "content": " as a distinct view group. However, prior work [49] relies on explicit knowledge of these groupings to train view-specific encoders. In contrast, our problem setting presents a greater challenge since we have no access to this grouping information. Specifically, given two captions of any different images, it is unclear whether they stem from the same mask (i.e., the same view group). Therefore, the identification guarantees in prior studies do not apply to our setting. Theorem 4.3 demonstrates that our estimation model, paired with the learning objective (2), can automatically infer the necessary grouping information (i.e., the masks). By doing so, our approach relaxes the identification conditions in previous work, enabling effective representation identification without explicit group knowledge." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 272, + 296, + 300 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 272, + 296, + 300 + ], + "spans": [ + { + "bbox": [ + 55, + 272, + 296, + 300 + ], + "type": "text", + "content": "5. SmartCLIP: Modular Vision-language Alignment" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 308, + 296, + 356 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 308, + 296, + 356 + ], + "spans": [ + { + "bbox": [ + 55, + 308, + 296, + 356 + ], + "type": "text", + "content": "Drawing on the theoretical framework in Section 4, we present SmartCLIP, a modular alignment model designed to achieve Goal a and Goal b. We discuss the implementation of the learning objective (2) and the model architecture." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 357, + 296, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 357, + 296, + 429 + ], + "spans": [ + { + "bbox": [ + 55, + 357, + 296, + 429 + ], + "type": "text", + "content": "Modular alignment through adaptive masking. The masking function " + }, + { + "bbox": [ + 55, + 357, + 296, + 429 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{m}} (\\cdot)" + }, + { + "bbox": [ + 55, + 357, + 296, + 429 + ], + "type": "text", + "content": " is instrumental in our modular alignment objective (2). SmartCLIP implements this function with a transformer block that ingests a caption representation " + }, + { + "bbox": [ + 55, + 357, + 296, + 429 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{z}}_{\\mathrm{T}}" + }, + { + "bbox": [ + 55, + 357, + 296, + 429 + ], + "type": "text", + "content": " as its input and outputs a binary vector " + }, + { + "bbox": [ + 55, + 357, + 296, + 429 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{m}} (\\hat{\\mathbf{z}}_{\\mathrm{T}})" + }, + { + "bbox": [ + 55, + 357, + 296, + 429 + ], + "type": "text", + "content": " via a straight-through estimator [4]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 430, + 296, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 430, + 296, + 514 + ], + "spans": [ + { + "bbox": [ + 55, + 430, + 296, + 514 + ], + "type": "text", + "content": "Modular contrast construction. As discussed in Section 4, the negative pairs in regular contrastive losses [8, 35] serves a similar role as the invertibility assumption (Condition 4.2-i) [43, 47]. We denote generic image, text representations as " + }, + { + "bbox": [ + 55, + 430, + 296, + 514 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 55, + 430, + 296, + 514 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 430, + 296, + 514 + ], + "type": "inline_equation", + "content": "\\mathbf{T}" + }, + { + "bbox": [ + 55, + 430, + 296, + 514 + ], + "type": "text", + "content": ", and positive, negative pairs with " + }, + { + "bbox": [ + 55, + 430, + 296, + 514 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{\\mathrm{pos}}" + }, + { + "bbox": [ + 55, + 430, + 296, + 514 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 430, + 296, + 514 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{\\mathrm{neg}}" + }, + { + "bbox": [ + 55, + 430, + 296, + 514 + ], + "type": "text", + "content": " respectively. The canonical one-side contrastive loss " + }, + { + "bbox": [ + 55, + 430, + 296, + 514 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{ctr}}" + }, + { + "bbox": [ + 55, + 430, + 296, + 514 + ], + "type": "text", + "content": " [37] is defined as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 89, + 523, + 296, + 606 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 523, + 296, + 606 + ], + "spans": [ + { + "bbox": [ + 89, + 523, + 296, + 606 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {\\mathrm {c t r}} \\left(\\underbrace {\\left(\\mathbf {I} ^ {(i)} , \\mathbf {T} ^ {(i)}\\right)} _ {\\mathbf {P} _ {\\text {p o s}}}, \\underbrace {\\left(\\mathbf {I} ^ {(i)} , \\mathbf {T} ^ {(j)}\\right)} _ {\\mathbf {P} _ {\\text {n e g}}} ^ {}\\right) \\tag {3} \\\\ = - \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\log \\frac {\\exp \\left(\\frac {\\mathrm {s i m} (\\mathbf {I} ^ {(i)} , \\mathbf {T} ^ {(i)})}{\\tau}\\right)}{\\sum_ {j = 1} ^ {N} \\exp \\left(\\frac {\\mathrm {s i m} (\\mathbf {I} ^ {(i)} , \\mathbf {T} ^ {(j)})}{\\tau}\\right)}, \\\\ \\end{array}", + "image_path": "e54cf8790ba79c0d7b48c70046dc40dfe31c6aedbca1513f323a5b258436dbae.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 616, + 295, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 616, + 295, + 640 + ], + "spans": [ + { + "bbox": [ + 55, + 616, + 295, + 640 + ], + "type": "text", + "content": "where we denote the sample size, temperature, and cosine similarity with " + }, + { + "bbox": [ + 55, + 616, + 295, + 640 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 616, + 295, + 640 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 616, + 295, + 640 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 55, + 616, + 295, + 640 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 616, + 295, + 640 + ], + "type": "inline_equation", + "content": "\\sin(\\cdot)" + }, + { + "bbox": [ + 55, + 616, + 295, + 640 + ], + "type": "text", + "content": " respectively." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 641, + 296, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 296, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 296, + 677 + ], + "type": "text", + "content": "Following the symmetric contrastive loss for CLIP [37], our alignment loss consists of two contrastive loss terms " + }, + { + "bbox": [ + 55, + 641, + 296, + 677 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{ctrI}}" + }, + { + "bbox": [ + 55, + 641, + 296, + 677 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 641, + 296, + 677 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{ctrT}}" + }, + { + "bbox": [ + 55, + 641, + 296, + 677 + ], + "type": "text", + "content": " that differ in the negative pairs:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 689, + 298, + 714 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 298, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 298, + 714 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {c t r I}} := \\mathcal {L} _ {\\mathrm {c t r}} \\left(\\mathbf {P} _ {\\text {p o s}}, \\mathbf {P} _ {\\text {n e g I}}\\right), \\mathcal {L} _ {\\mathrm {c t r T}} := \\mathcal {L} _ {\\mathrm {c t r}} \\left(\\mathbf {P} _ {\\text {p o s}}, \\mathbf {P} _ {\\text {n e g T}}\\right), \\tag {4}", + "image_path": "b9d6c65d8ed00f3710d18a5cadb7dbc2b835763aedd7cf94b8b299e3848e0735.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 72, + 533, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 533, + 84 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 533, + 84 + ], + "type": "text", + "content": "with the positive and negative pairs defined as follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 367, + 94, + 553, + 114 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 94, + 553, + 114 + ], + "spans": [ + { + "bbox": [ + 367, + 94, + 553, + 114 + ], + "type": "interline_equation", + "content": "\\mathbf {P} _ {\\text {p o s}} := \\left(\\hat {\\mathbf {z}} _ {\\mathrm {I}} ^ {(i)} \\odot \\hat {\\mathbf {m}} \\left(\\hat {\\mathbf {z}} _ {\\mathrm {T}} ^ {(i)}\\right), \\hat {\\mathbf {z}} _ {\\mathrm {T}} ^ {(i)}\\right), \\tag {5}", + "image_path": "9c693dc65f84e95741a5b46c542bcc860dba2e6a8295cf1348bfd5cf6b27cdcc.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 365, + 116, + 553, + 136 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 116, + 553, + 136 + ], + "spans": [ + { + "bbox": [ + 365, + 116, + 553, + 136 + ], + "type": "interline_equation", + "content": "\\mathbf {P} _ {\\text {n e g I}} := \\left(\\hat {\\mathbf {z}} _ {\\mathrm {I}} ^ {(i)} \\odot \\hat {\\mathbf {m}} \\left(\\hat {\\mathbf {z}} _ {\\mathrm {T}} ^ {(j)}\\right), \\hat {\\mathbf {z}} _ {\\mathrm {T}} ^ {(j)}\\right), \\tag {6}", + "image_path": "1112a9166741d8f22f68197fa41f19399e669a1f5341ab8bc7ba42b5f0bc319a.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 362, + 137, + 553, + 158 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 362, + 137, + 553, + 158 + ], + "spans": [ + { + "bbox": [ + 362, + 137, + 553, + 158 + ], + "type": "interline_equation", + "content": "\\mathbf {P} _ {\\mathrm {n e g T}} := \\left(\\hat {\\mathbf {z}} _ {\\mathrm {I}} ^ {(j)} \\odot \\hat {\\mathbf {m}} \\left(\\hat {\\mathbf {z}} _ {\\mathrm {T}} ^ {(i)}\\right), \\hat {\\mathbf {z}} _ {\\mathrm {T}} ^ {(i)}\\right). \\tag {7}", + "image_path": "fc3ca3b37ec158017b365800debec744d65b1cbf0c635d645e4c46b3b8921b41.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 169, + 553, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 169, + 553, + 248 + ], + "spans": [ + { + "bbox": [ + 313, + 169, + 553, + 248 + ], + "type": "text", + "content": "In particular, " + }, + { + "bbox": [ + 313, + 169, + 553, + 248 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{\\mathrm{negI}}" + }, + { + "bbox": [ + 313, + 169, + 553, + 248 + ], + "type": "text", + "content": " contrasts the image representation " + }, + { + "bbox": [ + 313, + 169, + 553, + 248 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{z}}_{\\mathrm{I}}^{(i)}" + }, + { + "bbox": [ + 313, + 169, + 553, + 248 + ], + "type": "text", + "content": " in the positive pair with randomly sampled caption representations " + }, + { + "bbox": [ + 313, + 169, + 553, + 248 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{z}}_{\\mathrm{T}}^{(j)}" + }, + { + "bbox": [ + 313, + 169, + 553, + 248 + ], + "type": "text", + "content": " (see the green region in Figure 3), whereas " + }, + { + "bbox": [ + 313, + 169, + 553, + 248 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{\\mathrm{negT}}" + }, + { + "bbox": [ + 313, + 169, + 553, + 248 + ], + "type": "text", + "content": " contrasts the text representation " + }, + { + "bbox": [ + 313, + 169, + 553, + 248 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{z}}_{\\mathrm{T}}^{(i)}" + }, + { + "bbox": [ + 313, + 169, + 553, + 248 + ], + "type": "text", + "content": " in the positive pair with randomly sampled image representations " + }, + { + "bbox": [ + 313, + 169, + 553, + 248 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{z}}_{\\mathrm{I}}^{(j)}" + }, + { + "bbox": [ + 313, + 169, + 553, + 248 + ], + "type": "text", + "content": " (see the orange region in Figure 3)." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 249, + 553, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 249, + 553, + 274 + ], + "spans": [ + { + "bbox": [ + 313, + 249, + 553, + 274 + ], + "type": "text", + "content": "Sparsity penalty. We implement " + }, + { + "bbox": [ + 313, + 249, + 553, + 274 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{sparsity}}" + }, + { + "bbox": [ + 313, + 249, + 553, + 274 + ], + "type": "text", + "content": " in (2) with a " + }, + { + "bbox": [ + 313, + 249, + 553, + 274 + ], + "type": "inline_equation", + "content": "\\ell_1" + }, + { + "bbox": [ + 313, + 249, + 553, + 274 + ], + "type": "text", + "content": " term for its compatibility with deep-learning training:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 388, + 283, + 553, + 298 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 388, + 283, + 553, + 298 + ], + "spans": [ + { + "bbox": [ + 388, + 283, + 553, + 298 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {s p a r s i t y}} = \\left\\| \\hat {\\mathbf {m}} (\\mathbf {t}) \\right\\| _ {1}. \\tag {8}", + "image_path": "ab581f58c2f168f0c2d9bf7bf89ee02b74557a14a303e8c45c31acb3704ff264.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 306, + 553, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 306, + 553, + 342 + ], + "spans": [ + { + "bbox": [ + 313, + 306, + 553, + 342 + ], + "type": "text", + "content": "This term ensures that the textual concepts are encoded into a minimal number of latent dimensions, promoting the disentanglement of distinct concepts across text captions." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 342, + 553, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 342, + 553, + 378 + ], + "spans": [ + { + "bbox": [ + 313, + 342, + 553, + 378 + ], + "type": "text", + "content": "SmartCLIP training objective. In summary, the training objective of SmartCLIP is a weighted sum of loss terms in (4) and (8):" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 329, + 389, + 553, + 403 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 329, + 389, + 553, + 403 + ], + "spans": [ + { + "bbox": [ + 329, + 389, + 553, + 403 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\lambda_ {\\text {a l i g n}} \\cdot \\left(\\mathcal {L} _ {\\mathrm {c t r I}} + \\mathcal {L} _ {\\mathrm {c t r T}}\\right) + \\lambda_ {\\text {s p a r s i t y}} \\cdot \\mathcal {L} _ {\\text {s p a r s i t y}}, \\tag {9}", + "image_path": "a484f851282a0e89428ff41f77520ed23857a6e67b6de1cf7942ec6bada146bd.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 314, + 412, + 548, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 412, + 548, + 425 + ], + "spans": [ + { + "bbox": [ + 314, + 412, + 548, + 425 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 314, + 412, + 548, + 425 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{align}}" + }, + { + "bbox": [ + 314, + 412, + 548, + 425 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 314, + 412, + 548, + 425 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{sparsity}}" + }, + { + "bbox": [ + 314, + 412, + 548, + 425 + ], + "type": "text", + "content": " denote the weighting coefficients." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 314, + 436, + 395, + 449 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 436, + 395, + 449 + ], + "spans": [ + { + "bbox": [ + 314, + 436, + 395, + 449 + ], + "type": "text", + "content": "6. Experiments" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 314, + 456, + 364, + 469 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 456, + 364, + 469 + ], + "spans": [ + { + "bbox": [ + 314, + 456, + 364, + 469 + ], + "type": "text", + "content": "6.1. Setup" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 474, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 474, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 474, + 555, + 715 + ], + "type": "text", + "content": "Implementation details. Following Long-CLIP [54], we finetune the CLIP model [37] on ShareGPT4V [7], which contains around 1 million image-text pairs. We employ the position encoding in long-CLIP to handle 248 tokens (c.f., the 77-token limit in the original CLIP). Compared to the baseline CLIP model, we introduce a mask network " + }, + { + "bbox": [ + 313, + 474, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{m}}" + }, + { + "bbox": [ + 313, + 474, + 555, + 715 + ], + "type": "text", + "content": ". The masking network is designed as a single transformer block, which takes the text sequence embedding " + }, + { + "bbox": [ + 313, + 474, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{z}}_{\\mathrm{T}}" + }, + { + "bbox": [ + 313, + 474, + 555, + 715 + ], + "type": "text", + "content": " from the text encoder. Then we add an attention-pooling layer to down-sample it to the same size as the CLIP representation, e.g., 768 in ViT-L/14. We tested including more transformer blocks in the mask network but did not observe significant improvements. Therefore, we stick to one block for faster training and inference. Unlike Long-CLIP [54] which processes all the captions for each image at each gradient step, we only sample one caption for each image, reducing our overall training time by half. Specifically, on 8 H100 GPUs, training a Vit-B/16 model for one epoch takes about 4 minutes with our model, whereas it takes around 7 minutes for Long-CLIP. After the pooling layer, we apply sigmoid to" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "29784" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 114, + 76, + 494, + 220 + ], + "blocks": [ + { + "bbox": [ + 114, + 76, + 494, + 220 + ], + "lines": [ + { + "bbox": [ + 114, + 76, + 494, + 220 + ], + "spans": [ + { + "bbox": [ + 114, + 76, + 494, + 220 + ], + "type": "image", + "image_path": "fe4d3b1c3b41015ef13fb0b62d8b3997d6d7437d6c55613d230d610306aa9b73.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 233, + 555, + 257 + ], + "lines": [ + { + "bbox": [ + 55, + 233, + 555, + 257 + ], + "spans": [ + { + "bbox": [ + 55, + 233, + 555, + 257 + ], + "type": "text", + "content": "Figure 3. The diagram of SmartCLIP. On the left, we introduce adaptive masking for alignment with different text prompts. The mask network selects which part of the image representation to be used. On the right, we present our modular contrastive objectives (4)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 74, + 287, + 533, + 423 + ], + "blocks": [ + { + "bbox": [ + 61, + 267, + 548, + 278 + ], + "lines": [ + { + "bbox": [ + 61, + 267, + 548, + 278 + ], + "spans": [ + { + "bbox": [ + 61, + 267, + 548, + 278 + ], + "type": "text", + "content": "Table 1. Results of short-caption text-image retrieval on the 5k COCO2017 validation set and the whole 30k Flickr30K dataset." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 74, + 287, + 533, + 423 + ], + "lines": [ + { + "bbox": [ + 74, + 287, + 533, + 423 + ], + "spans": [ + { + "bbox": [ + 74, + 287, + 533, + 423 + ], + "type": "table", + "html": "
COCOFlickr30k
Image-to-TextText-to-ImageImage-to-TextText-to-Image
R@1R@5R@10R@1R@5R@10R@1R@5R@10R@1R@5R@10
B/16CLIP51.876.884.332.757.768.244.168.277.024.745.154.6
Direct Fine-tuning37.462.372.121.843.454.525.745.855.417.934.543.1
Long-CLIP [54]57.681.187.840.465.875.246.871.479.834.156.365.7
SmartCLIP (Ours)61.983.389.742.468.277.855.678.285.036.358.867.8
L/14CLIP56.179.586.835.460.170.248.572.680.828.049.358.7
Direct Fine-tuning37.963.172.223.145.155.926.046.355.617.934.943.5
Long-CLIP [54]62.885.191.246.370.879.853.477.585.341.264.172.6
SmartCLIP (Ours)66.086.292.648.573.181.763.984.490.243.866.574.8
", + "image_path": "1eac77474da52d16045be982501852d930124b9f333e9a5cc2f9e80cddac730f.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 431, + 296, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 431, + 296, + 479 + ], + "spans": [ + { + "bbox": [ + 55, + 431, + 296, + 479 + ], + "type": "text", + "content": "restrict the output to the range " + }, + { + "bbox": [ + 55, + 431, + 296, + 479 + ], + "type": "inline_equation", + "content": "(0,1)" + }, + { + "bbox": [ + 55, + 431, + 296, + 479 + ], + "type": "text", + "content": " and employ straight-through estimation [4] to binarize the outputs. The training batch size is 1024 and the learning rate is " + }, + { + "bbox": [ + 55, + 431, + 296, + 479 + ], + "type": "inline_equation", + "content": "10^{-6}" + }, + { + "bbox": [ + 55, + 431, + 296, + 479 + ], + "type": "text", + "content": " for the CLIP component and " + }, + { + "bbox": [ + 55, + 431, + 296, + 479 + ], + "type": "inline_equation", + "content": "10^{-3}" + }, + { + "bbox": [ + 55, + 431, + 296, + 479 + ], + "type": "text", + "content": " for the mask network." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 484, + 251, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 484, + 251, + 496 + ], + "spans": [ + { + "bbox": [ + 55, + 484, + 251, + 496 + ], + "type": "text", + "content": "Evaluation. We evaluate the following datasets:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 502, + 295, + 633 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 55, + 502, + 295, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 502, + 295, + 550 + ], + "spans": [ + { + "bbox": [ + 55, + 502, + 295, + 550 + ], + "type": "text", + "content": "- Long text-to-image retrieval datasets: ShareGPT4V validation split [7] and Urban1k [54]. The captions for each image are long and describe details about the image. Both datasets contain 1000 text-to-image pairs." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 550, + 295, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 550, + 295, + 586 + ], + "spans": [ + { + "bbox": [ + 55, + 550, + 295, + 586 + ], + "type": "text", + "content": "- Short text-to-image retrieval datasets: COCO2017 validation split [27] and Flickr30K [50]. Following LongCLIP [54], we use 30K Flickr training dataset." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 586, + 295, + 633 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 586, + 295, + 633 + ], + "spans": [ + { + "bbox": [ + 55, + 586, + 295, + 633 + ], + "type": "text", + "content": "- Zero-shot image classification datasets. We use benchmark datasets: Country211, Fer2013, Fgvc-aircraft, GT-SRB, ImageNet, ImagetNet-V2, VOC2007, VOC2007-Multi, and SUN397." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 639, + 296, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 639, + 296, + 674 + ], + "spans": [ + { + "bbox": [ + 55, + 639, + 296, + 674 + ], + "type": "text", + "content": "Baselines. In this paper, we benchmark our approach against CLIP [37] and the recent state-of-the-art model long-CLIP [54]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 431, + 484, + 443 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 431, + 484, + 443 + ], + "spans": [ + { + "bbox": [ + 313, + 431, + 484, + 443 + ], + "type": "text", + "content": "6.2. Comparison with CLIP Models" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 449, + 554, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 449, + 554, + 484 + ], + "spans": [ + { + "bbox": [ + 313, + 449, + 554, + 484 + ], + "type": "text", + "content": "We present our experimental results across three key tasks: long text-to-image retrieval, short text-to-image retrieval, and zero-shot classification." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 485, + 554, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 485, + 554, + 568 + ], + "spans": [ + { + "bbox": [ + 313, + 485, + 554, + 568 + ], + "type": "text", + "content": "Long text-to-image retrieval. Table 2 showcases our method's performance on long text-to-image retrieval tasks. SmartCLIP achieves substantial improvements over baseline models, particularly the SOTA Long-CLIP, which is designed to handle long text sequences. For example, on the Urban1k dataset, SmartCLIP elevates the performance from " + }, + { + "bbox": [ + 313, + 485, + 554, + 568 + ], + "type": "inline_equation", + "content": "78.9\\%" + }, + { + "bbox": [ + 313, + 485, + 554, + 568 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 485, + 554, + 568 + ], + "type": "inline_equation", + "content": "90.0\\%" + }, + { + "bbox": [ + 313, + 485, + 554, + 568 + ], + "type": "text", + "content": ", marking an impressive " + }, + { + "bbox": [ + 313, + 485, + 554, + 568 + ], + "type": "inline_equation", + "content": "14\\%" + }, + { + "bbox": [ + 313, + 485, + 554, + 568 + ], + "type": "text", + "content": " boost." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 570, + 554, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 554, + 652 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 554, + 652 + ], + "type": "text", + "content": "Short text-to-image retrieval. Similarly, as illustrated in Table 1, SmartCLIP significantly outperforms all baseline models across various metrics and datasets in short text-to-image retrieval tasks. The encouraging performance gains show that SmartCLIP can capture detailed information within images while simultaneously emphasizing the main semantic content." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "text", + "content": "Zero-shot classification. To comprehensively evaluate our model's capabilities, we conduct zero-shot classification benchmarks in Table 3. Both the standard CLIP model and SmartCLIP demonstrate superior performance on different datasets. Notably, SmartCLIP exhibits a slight perfor" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 66, + 702, + 274, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 702, + 274, + 712 + ], + "spans": [ + { + "bbox": [ + 66, + 702, + 274, + 712 + ], + "type": "text", + "content": "2https://github.com/LAION-AI/CLIP_benchmark" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "29785" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 124, + 288, + 248 + ], + "blocks": [ + { + "bbox": [ + 55, + 82, + 307, + 116 + ], + "lines": [ + { + "bbox": [ + 55, + 82, + 307, + 116 + ], + "spans": [ + { + "bbox": [ + 55, + 82, + 307, + 116 + ], + "type": "text", + "content": "Table 2. The R@1 of long-caption text-image retrieval on 1k ShareGPT4V [2] validation set and Urban-1000 dataset. The best results are bold. We cite the results from Long-CLIP [54]." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 124, + 288, + 248 + ], + "lines": [ + { + "bbox": [ + 56, + 124, + 288, + 248 + ], + "spans": [ + { + "bbox": [ + 56, + 124, + 288, + 248 + ], + "type": "table", + "html": "
ShareGPT4VUrban1k
I2TT2II2TT2I
B/16CLIP [37]78.279.668.153.6
Direct Fine-tuning94.193.6--
Long-CLIP [54]94.693.378.979.5
SmartCLIP (Ours)98.798.190.087.4
L/14CLIP [37]81.884.068.752.8
Direct Fine-tuning95.395.4--
Long-CLIP [54]95.895.682.786.1
SmartCLIP (Ours)97.998.593.090.1
", + "image_path": "d511d4a79744550756f60591cc41feb84a99cac6b58cb2af5df328c2c0783f5b.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 329, + 133, + 561, + 258 + ], + "blocks": [ + { + "bbox": [ + 327, + 70, + 555, + 126 + ], + "lines": [ + { + "bbox": [ + 327, + 70, + 555, + 126 + ], + "spans": [ + { + "bbox": [ + 327, + 70, + 555, + 126 + ], + "type": "text", + "content": "Table 3. Zero-shot classification performance on ViT-L/14 models. When the class name is very short, i.e., a single word like ImageNet, CLIP model performs better. When the class name is a combination of several words, our method achieves better results, e.g., the road sign in GTSRB." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 329, + 133, + 561, + 258 + ], + "lines": [ + { + "bbox": [ + 329, + 133, + 561, + 258 + ], + "spans": [ + { + "bbox": [ + 329, + 133, + 561, + 258 + ], + "type": "table", + "html": "
DatasetCLIPLongCLIPSmartCLIP
Country21131.828.126.9
Fer201349.057.858.6
Fgvc-aircraft31.730.630.4
GTSRB50.248.952.4
ImageNet75.372.972.5
ImageNet-V269.766.966.6
VOC200778.377.578.6
VOC2007-Multi79.082.183.7
SUN39767.572.572.1
", + "image_path": "b8cef987bc4a7fb274d6b20e1e6621f42e4fa0d246a9bf53cdbe7bce7a030460.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 273, + 358, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 273, + 358, + 416 + ], + "spans": [ + { + "bbox": [ + 55, + 273, + 358, + 416 + ], + "type": "text", + "content": "This playful food sculpture transforms cucumbers into a fearsome T-Rex dinosaur. The cucumbers form the main body, with whole cucumbers creating the legs and tail, while sliced cucumbers make up the creature's midsection. More cucumbers are cleverly cut to shape the dinosaur's head, and additional cucumbers are arranged to suggest muscular limbs. From its mouth erupts a dramatic spray of carrots, with finely julienned carrots creating the effect of fire. These bright orange carrots provide a stunning contrast against the green vegetables. Shredded carrots cascade downward like flames, while more carrots are delicately cut to create a flame-like texture. The carrots' vibrant color makes the dinosaur appear truly animated. Fresh celery leaves crown the creation, with celery fronds adding a decorative touch around the body. More celery leaves create a natural backdrop, while additional celery pieces add texture throughout. The celery's feathered appearance provides an artistic flourish to the overall design." + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 378, + 271, + 493, + 366 + ], + "blocks": [ + { + "bbox": [ + 378, + 271, + 493, + 366 + ], + "lines": [ + { + "bbox": [ + 378, + 271, + 493, + 366 + ], + "spans": [ + { + "bbox": [ + 378, + 271, + 493, + 366 + ], + "type": "image", + "image_path": "19d15dd85f69958962524602adb46aa04bd415504c9d516fe0e7aea6bcc54a5d.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 378, + 367, + 493, + 415 + ], + "blocks": [ + { + "bbox": [ + 378, + 367, + 493, + 415 + ], + "lines": [ + { + "bbox": [ + 378, + 367, + 493, + 415 + ], + "spans": [ + { + "bbox": [ + 378, + 367, + 493, + 415 + ], + "type": "image", + "image_path": "5ef18a3723b1f69c850e9f13cb09f3f85a2d2df335404ca1061eb8e074934bdd.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 417, + 555, + 451 + ], + "lines": [ + { + "bbox": [ + 55, + 417, + 555, + 451 + ], + "spans": [ + { + "bbox": [ + 55, + 417, + 555, + 451 + ], + "type": "text", + "content": "Figure 4. Example of Long-text-to-image generation. We replace the CLIP text encoder in SDXL with different finetuned CLIP models. Given a very long text, CLIP [37] truncates the input to 77 tokens, resulting in information loss in the image. Our model learns to generate details such as celery leaves on the back of the dinosaur while other models fail." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 462, + 295, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 462, + 295, + 547 + ], + "spans": [ + { + "bbox": [ + 55, + 462, + 295, + 547 + ], + "type": "text", + "content": "mance decline on ImageNet, which is expected since our model is fine-tuned on the ShareGPT4V dataset featuring long text captions, whereas ImageNet primarily consists of short, often single-word class names. However, Smart-CLIP excels on datasets with multi-word class names, such as the GTSRB dataset, where it achieves the best performance in accurately classifying road sign descriptions." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 552, + 156, + 564 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 552, + 156, + 564 + ], + "spans": [ + { + "bbox": [ + 55, + 552, + 156, + 564 + ], + "type": "text", + "content": "6.3. Ablation Studies" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 570, + 295, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 295, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 295, + 605 + ], + "type": "text", + "content": "We analyze the three key components in our model: the modular alignment module, the sparsity loss and the impact of caption diversity in the data." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "content": "Modular alignment. After introducing the mask network, we replace the standard contrastive learning with our selective alignment module. As shown in Figure 5, this change significantly improves performance. When using standard contrastive learning instead of the modular contrastive module (indicated by the purple lines), performance drops sharply. This happens because the mask information allows the network to easily separate positive pairs from negative ones, making the negative samples less informa" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 462, + 555, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 462, + 555, + 486 + ], + "spans": [ + { + "bbox": [ + 313, + 462, + 555, + 486 + ], + "type": "text", + "content": "tive. As a result, standard contrastive learning no longer effectively helps the model learn meaningful information." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 486, + 554, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 486, + 554, + 558 + ], + "spans": [ + { + "bbox": [ + 313, + 486, + 554, + 558 + ], + "type": "text", + "content": "Alignment coefficient " + }, + { + "bbox": [ + 313, + 486, + 554, + 558 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{align}}" + }, + { + "bbox": [ + 313, + 486, + 554, + 558 + ], + "type": "text", + "content": ". We test the impact of the alignment coefficient " + }, + { + "bbox": [ + 313, + 486, + 554, + 558 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{align}}" + }, + { + "bbox": [ + 313, + 486, + 554, + 558 + ], + "type": "text", + "content": ". The results in the right panels of Figure 5 show that our method performs consistently well across a wide range of " + }, + { + "bbox": [ + 313, + 486, + 554, + 558 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{align}}" + }, + { + "bbox": [ + 313, + 486, + 554, + 558 + ], + "type": "text", + "content": " values, from 0.1 to 20. This indicates that our approach is robust and does not require precise tuning of " + }, + { + "bbox": [ + 313, + 486, + 554, + 558 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{align}}" + }, + { + "bbox": [ + 313, + 486, + 554, + 558 + ], + "type": "text", + "content": " to achieve good performance." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 558, + 554, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 554, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 554, + 640 + ], + "type": "text", + "content": "Sparsity coefficient " + }, + { + "bbox": [ + 313, + 558, + 554, + 640 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{sparsity}}" + }, + { + "bbox": [ + 313, + 558, + 554, + 640 + ], + "type": "text", + "content": ". We also examine the sparsity coefficient " + }, + { + "bbox": [ + 313, + 558, + 554, + 640 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{sparsity}}" + }, + { + "bbox": [ + 313, + 558, + 554, + 640 + ], + "type": "text", + "content": ". The left panels of Figure 5 demonstrate that adding sparsity to the mask network improves performance. This supports our idea that promoting sparsity helps the model focus on the most relevant concepts, enhancing its ability to capture detailed information without being distracted by irrelevant details." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 642, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 555, + 714 + ], + "type": "text", + "content": "Caption diversity. We evaluate our model's performance under varying caption diversity conditions using the COCO dataset [27]. As shown in Table 4, increasing the number of captions per image enhances performance on the Flickr30K dataset, though at the expense of degraded performance on long-text-to-image retrieval tasks. Further improvements" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "29786" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 72, + 172, + 155 + ], + "blocks": [ + { + "bbox": [ + 60, + 72, + 172, + 155 + ], + "lines": [ + { + "bbox": [ + 60, + 72, + 172, + 155 + ], + "spans": [ + { + "bbox": [ + 60, + 72, + 172, + 155 + ], + "type": "image", + "image_path": "f549d8ce6b37c71190963451851c7ee1cbfbca0483d7093681ca3929ae1ec123.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 157, + 555, + 178 + ], + "lines": [ + { + "bbox": [ + 55, + 157, + 555, + 178 + ], + "spans": [ + { + "bbox": [ + 55, + 157, + 555, + 178 + ], + "type": "text", + "content": "Figure 5. Ablation Studies on two proposed modules: selective alignment and sparsity. The baseline w.o. Modular means that we replace our modular alignment module with standard contrastive learning alignment." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 184, + 72, + 296, + 154 + ], + "blocks": [ + { + "bbox": [ + 184, + 72, + 296, + 154 + ], + "lines": [ + { + "bbox": [ + 184, + 72, + 296, + 154 + ], + "spans": [ + { + "bbox": [ + 184, + 72, + 296, + 154 + ], + "type": "image", + "image_path": "48471290c784947d1f3355796a27856dbad99276e3bbf63a360e72188f23cb1b.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 309, + 72, + 421, + 154 + ], + "blocks": [ + { + "bbox": [ + 309, + 72, + 421, + 154 + ], + "lines": [ + { + "bbox": [ + 309, + 72, + 421, + 154 + ], + "spans": [ + { + "bbox": [ + 309, + 72, + 421, + 154 + ], + "type": "image", + "image_path": "ba17102ae45429d0dd3f06743232dc50f81cb980b7674fb6b7b8081733727316.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 433, + 72, + 545, + 154 + ], + "blocks": [ + { + "bbox": [ + 433, + 72, + 545, + 154 + ], + "lines": [ + { + "bbox": [ + 433, + 72, + 545, + 154 + ], + "spans": [ + { + "bbox": [ + 433, + 72, + 545, + 154 + ], + "type": "image", + "image_path": "10e4160ab66829aef5af4b2aa5966d7f36642f4862260f0b30129b2799cc56b2.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 63, + 182, + 289, + 308 + ], + "blocks": [ + { + "bbox": [ + 63, + 182, + 289, + 308 + ], + "lines": [ + { + "bbox": [ + 63, + 182, + 289, + 308 + ], + "spans": [ + { + "bbox": [ + 63, + 182, + 289, + 308 + ], + "type": "image", + "image_path": "d07100a362ce9907c1c077baa57eb37d5dde5da2f153a0f1850b914a1dd7feca.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 318, + 296, + 385 + ], + "lines": [ + { + "bbox": [ + 55, + 318, + 296, + 385 + ], + "spans": [ + { + "bbox": [ + 55, + 318, + 296, + 385 + ], + "type": "text", + "content": "Figure 6. Visualization of learned representations. Given an image, we generate two captions (e.g., a zebra and a zebra and a deer) and compute cosine similarities with the same image embedding to perform binary classification for visualization using Score-CAM [45]. Compared to baseline methods, our CLIP representations are more atomic and capture differences more effectively." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 398, + 296, + 446 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 398, + 296, + 446 + ], + "spans": [ + { + "bbox": [ + 55, + 398, + 296, + 446 + ], + "type": "text", + "content": "are achieved when we combine our training dataset with COCO. These results highlight both the importance of caption diversity and our method's capability to effectively handle complex text-image datasets." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 455, + 165, + 467 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 455, + 165, + 467 + ], + "spans": [ + { + "bbox": [ + 55, + 455, + 165, + 467 + ], + "type": "text", + "content": "6.4. Additional Results" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 473, + 295, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 473, + 295, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 473, + 295, + 628 + ], + "type": "text", + "content": "Visualization. While our quantitative results demonstrate superior performance across various tasks, we also explore the qualitative aspects of our model by visualizing the learned representations. The visualization results are shown in Figure 6. We employ the ScoreCAM method [45] for this purpose. For each image, we generate two distinct captions, such as \"a zebra\" and \"a zebra and a deer\". We then compute the cosine similarity between the image embedding and each of the two text embeddings. These similarity scores serve as logits for a classification task, which are then input into the ScoreCAM algorithm. SmartCLIP successfully learns modular representations, accurately capturing the relevant differences between captions." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "text", + "content": "Plug and play for text-to-image generation. One main advantage of SmartCLIP over other CLIP models trained from scratch is the low computational cost of finetuning. Additionally, our fine-tuned text encoder can replace the CLIP text encoders in large-scale models in a plug-and-play manner. Specifically, we substitute the text encoder in the SDXL [36] model with both Long-CLIP and SmartCLIP." + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 325, + 178, + 541, + 245 + ], + "blocks": [ + { + "bbox": [ + 325, + 178, + 541, + 245 + ], + "lines": [ + { + "bbox": [ + 325, + 178, + 541, + 245 + ], + "spans": [ + { + "bbox": [ + 325, + 178, + 541, + 245 + ], + "type": "table", + "html": "
NumCapPerImg-COCOT2II2TLongT2ILongI2T
153.639.385.389.3
353.640.985.388.6
556.441.285.286.5
Ours-ShareGPT55.636.398.798.1
+COCO57.038.497.898.5
", + "image_path": "a23d3354d8b26f59d245032a28d867ad204cd939d2e7860efe6fd9fe50ae23ea.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 316, + 353, + 551, + 392 + ], + "blocks": [ + { + "bbox": [ + 313, + 247, + 555, + 281 + ], + "lines": [ + { + "bbox": [ + 313, + 247, + 555, + 281 + ], + "spans": [ + { + "bbox": [ + 313, + 247, + 555, + 281 + ], + "type": "text", + "content": "Table 4. Retrieval results on Flickr30K [50] and ShareGPT-val [7] with models trained with different caption counts per image (top) and mixing long and short text-image datasets (bottom)." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 313, + 289, + 555, + 344 + ], + "lines": [ + { + "bbox": [ + 313, + 289, + 555, + 344 + ], + "spans": [ + { + "bbox": [ + 313, + 289, + 555, + 344 + ], + "type": "text", + "content": "Table 5. Long text to image generation performance. We use the long captions (usually around 200 tokens, beyond the hard constraint 77 of the original CLIP model) to generate images with SDXL model [36], then we compare the generated images against the real images in ShareGPT4V validation split." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 316, + 353, + 551, + 392 + ], + "lines": [ + { + "bbox": [ + 316, + 353, + 551, + 392 + ], + "spans": [ + { + "bbox": [ + 316, + 353, + 551, + 392 + ], + "type": "table", + "html": "
MethodKID ↓Pr ↑Re ↑F1 ↑DINO-L ↑
LongCLIP1.050.2380.7680.3630.401
SmartCLIP1.020.2580.7910.3890.414
", + "image_path": "181703a468cf8331400c96f84dc806f4494e8e263124242468d489dc436e40b9.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 402, + 555, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 402, + 555, + 498 + ], + "spans": [ + { + "bbox": [ + 313, + 402, + 555, + 498 + ], + "type": "text", + "content": "As illustrated in Figure 4, our text encoder demonstrates a superior understanding of the long text, generating detailed elements such as celery leaves in the background. Furthermore, Table 5 presents quantitative results on image generation from long captions in the ShareGPT4V validation split. Our method consistently achieves better performance across various metrics, showcasing its effectiveness in handling complex, long text inputs." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 513, + 468, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 513, + 468, + 525 + ], + "spans": [ + { + "bbox": [ + 313, + 513, + 468, + 525 + ], + "type": "text", + "content": "7. Conclusion and Limitation" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 534, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 556, + 713 + ], + "type": "text", + "content": "In this work, we address the information misalignment and representation entanglement issues in existing vision-language models (e.g., CLIP). We establish theoretical conditions for effectively connecting text representations to atomic-level visual features and propose SmartCLIP, a principled, refined vision-language model. Our experimental results validate both our theoretical results and the practical effectiveness of SmartCLIP in advancing multimodal learning. Limitation. As discussed earlier, Condition 4.2-ii could be violated for datasets in which a subset images are paired with a limited number of captions compared to others. In Section 4, we discuss practical strategies to mitigate such issues. Devising alternative theoretical conditions may provide additional insights into fully utilizing all pairing information, which we leave as future work." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "29787" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 175 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 175 + ], + "type": "text", + "content": "We would like to acknowledge the support from NSF Award No. 2229881, AI Institute for Societal Decision Making (AI-SDM), the National Institutes of Health (NIH) under Contract R01HL159805, and grants from Quris AI, Florin Court Capital, and MBZUAI-WIS Joint Program. The work of L. Kong is supported in part by NSF DMS-2134080 through an award to Y. Chi." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 186, + 115, + 199 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 186, + 115, + 199 + ], + "spans": [ + { + "bbox": [ + 56, + 186, + 115, + 199 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 206, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 206, + 296, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 206, + 296, + 251 + ], + "spans": [ + { + "bbox": [ + 61, + 206, + 296, + 251 + ], + "type": "text", + "content": "[1] Kartik Ahuja, Divyat Mahajan, Yixin Wang, and Yoshua Bengio. Interventional causal representation learning. In International conference on machine learning, pages 372-407. PMLR, 2023. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 252, + 296, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 252, + 296, + 306 + ], + "spans": [ + { + "bbox": [ + 61, + 252, + 296, + 306 + ], + "type": "text", + "content": "[2] Alberto Baldrati, Lorenzo Agnolucci, Marco Bertini, and Alberto Del Bimbo. Zero-shot composed image retrieval with textual inversion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15338-15347, 2023. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 308, + 295, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 308, + 295, + 342 + ], + "spans": [ + { + "bbox": [ + 62, + 308, + 295, + 342 + ], + "type": "text", + "content": "[3] Yoshua Bengio, Réjean Ducharme, and Pascal Vincent. A neural probabilistic language model. Advances in neural information processing systems, 13, 2000. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 343, + 295, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 343, + 295, + 386 + ], + "spans": [ + { + "bbox": [ + 62, + 343, + 295, + 386 + ], + "type": "text", + "content": "[4] Yoshua Bengio, Nicholas Léonard, and Aaron Courville. Estimating or propagating gradients through stochastic neurons for conditional computation. arXiv preprint arXiv:1308.3432, 2013. 5, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 388, + 296, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 388, + 296, + 442 + ], + "spans": [ + { + "bbox": [ + 62, + 388, + 296, + 442 + ], + "type": "text", + "content": "[5] Simon Buchholz, Goutham Rajendran, Elan Rosenfeld, Bryon Aragam, Bernhard Scholkopf, and Pradeep Ravikumar. Learning linear causal representations from interventions under general nonlinear mixing. Advances in Neural Information Processing Systems, 36, 2024. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 445, + 296, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 445, + 296, + 498 + ], + "spans": [ + { + "bbox": [ + 62, + 445, + 296, + 498 + ], + "type": "text", + "content": "[6] Chang Che, Qunwei Lin, Xinyu Zhao, Jiaxin Huang, and Liqiang Yu. Enhancing multimodal understanding with clip-based image-to-text transformation. In Proceedings of the 2023 6th International Conference on Big Data Technologies, pages 414-418, 2023. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 501, + 294, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 501, + 294, + 544 + ], + "spans": [ + { + "bbox": [ + 62, + 501, + 294, + 544 + ], + "type": "text", + "content": "[7] Lin Chen, Jinsong Li, Xiaoyi Dong, Pan Zhang, Conghui He, Jiaqi Wang, Feng Zhao, and Dahua Lin. Sharegpt4v: Improving large multi-modal models with better captions. arXiv preprint arXiv:2311.12793, 2023. 2, 5, 6, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 545, + 295, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 545, + 295, + 589 + ], + "spans": [ + { + "bbox": [ + 62, + 545, + 295, + 589 + ], + "type": "text", + "content": "[8] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 1, 4, 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 590, + 295, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 590, + 295, + 635 + ], + "spans": [ + { + "bbox": [ + 62, + 590, + 295, + 635 + ], + "type": "text", + "content": "[9] Imant Daunhawer, Alice Bizeul, Emanuele Palumbo, Alexander Marx, and Julia E Vogt. Identifiability results for multimodal contrastive learning. In The Eleventh International Conference on Learning Representations, 2023. 3, 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 636, + 295, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 636, + 295, + 678 + ], + "spans": [ + { + "bbox": [ + 57, + 636, + 295, + 678 + ], + "type": "text", + "content": "[10] Lijie Fan, Dilip Krishnan, Phillip Isola, Dina Katabi, and Yonglong Tian. Improving clip training with language rewrites. Advances in Neural Information Processing Systems, 36, 2024. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 681, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 681, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 681, + 295, + 713 + ], + "type": "text", + "content": "[11] Luigi Gresele, Paul K Rubenstein, Arash Mehrjou, Francesco Locatello, and Bernhard Schölkopf. The incomplete rosetta stone problem: Identifiability results for multi-" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "view nonlinearica.In Uncertainty in Artificial Intelligence, pages 217-227.PMLR,2020.3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 97, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 97, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 97, + 553, + 150 + ], + "type": "text", + "content": "[12] Weiquan Huang, Aooqi Wu, Yifan Yang, Xufang Luo, Yuqing Yang, Liang Hu, Qi Dai, Xiyang Dai, Dongdong Chen, Chong Luo, et al. Llm2clip: Powerful language model unlock richer visual representation. arXiv preprint arXiv:2411.04997, 2024. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 152, + 553, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 553, + 185 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 553, + 185 + ], + "type": "text", + "content": "[13] Aapo Hyvarinen and Hiroshi Morioka. Unsupervised feature extraction by time-contrastive learning and nonlinear ica, 2016. 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 187, + 553, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 187, + 553, + 240 + ], + "spans": [ + { + "bbox": [ + 317, + 187, + 553, + 240 + ], + "type": "text", + "content": "[14] Aapo Hyvarinen, Hiroaki Sasaki, and Richard Turner. Nonlinear ica using auxiliary variables and generalized contrastive learning. In The 22nd International Conference on Artificial Intelligence and Statistics, pages 859-868. PMLR, 2019. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 243, + 553, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 243, + 553, + 308 + ], + "spans": [ + { + "bbox": [ + 317, + 243, + 553, + 308 + ], + "type": "text", + "content": "[15] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International conference on machine learning, pages 4904-4916. PMLR, 2021. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 310, + 553, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 310, + 553, + 354 + ], + "spans": [ + { + "bbox": [ + 316, + 310, + 553, + 354 + ], + "type": "text", + "content": "[16] Chen Ju, Tengda Han, Kunhao Zheng, Ya Zhang, and Weidi Xie. Prompting visual-language models for efficient video understanding. In European Conference on Computer Vision, pages 105-124. Springer, 2022. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 355, + 553, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 355, + 553, + 409 + ], + "spans": [ + { + "bbox": [ + 317, + 355, + 553, + 409 + ], + "type": "text", + "content": "[17] Ilyes Khemakhem, Diederik Kingma, Ricardo Monti, and Aapo Hyvarinen. Variational autoencoders and nonlinear ica: A unifying framework. In International Conference on Artificial Intelligence and Statistics, pages 2207-2217. PMLR, 2020. 3, 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 411, + 553, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 411, + 553, + 466 + ], + "spans": [ + { + "bbox": [ + 317, + 411, + 553, + 466 + ], + "type": "text", + "content": "[18] Lingjing Kong, Shaoan Xie, Weiran Yao, Yujia Zheng, Guangyi Chen, Petar Stojanov, Victor Akinwande, and Kun Zhang. Partial disentanglement for domain adaptation. In Proceedings of the 39th International Conference on Machine Learning, pages 11455-11472. PMLR, 2022. 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 468, + 553, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 468, + 553, + 510 + ], + "spans": [ + { + "bbox": [ + 317, + 468, + 553, + 510 + ], + "type": "text", + "content": "[19] Sébastien Lachapelle, Divyat Mahajan, Ioannis Mitliagkas, and Simon Lacoste-Julien. Additive decoders for latent variables identification and cartesian-product extrapolation, 2023. 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 512, + 553, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 512, + 553, + 567 + ], + "spans": [ + { + "bbox": [ + 317, + 512, + 553, + 567 + ], + "type": "text", + "content": "[20] Zhengfeng Lai, Vasileios Saveris, Chen Chen, Hong-You Chen, Haotian Zhang, Bowen Zhang, Juan Lao Tebar, Wenze Hu, Zhe Gan, Peter Grasch, et al. Revisit large-scale image-caption data in pre-training multimodal foundation models. arXiv preprint arXiv:2410.02740, 2024. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 569, + 553, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 569, + 553, + 622 + ], + "spans": [ + { + "bbox": [ + 317, + 569, + 553, + 622 + ], + "type": "text", + "content": "[21] Zhengfeng Lai, Haotian Zhang, Wentao Wu, Haoping Bai, Aleksei Timofeev, Xianzhi Du, Zhe Gan, Jiulong Shan, Chen-Nee Chuah, Yinfei Yang, et al. Veclip: Improving clip training via visual-enriched captions. ECCV. IEEE, page 13, 2024. 2, 4" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 624, + 553, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 624, + 553, + 678 + ], + "spans": [ + { + "bbox": [ + 317, + 624, + 553, + 678 + ], + "type": "text", + "content": "[22] Samuel Lavoie, Polina Kirichenko, Mark Ibrahim, Mahmoud Assran, Andrew Gordon Wilson, Aaron Courville, and Nicolas Ballas. Modeling caption diversity in contrastive vision-language pretraining. arXiv preprint arXiv:2405.00740, 2024. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "text", + "content": "[23] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In Interna-" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "29788" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 77, + 72, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 77, + 72, + 294, + 95 + ], + "type": "text", + "content": "tional conference on machine learning, pages 12888-12900. \nPMLR, 2022. 2, 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 150 + ], + "type": "text", + "content": "[24] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730-19742. PMLR, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 294, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 294, + 205 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 294, + 205 + ], + "type": "text", + "content": "[25] Xianhang Li, Haoqin Tu, Mude Hui, Zeyu Wang, Bingchen Zhao, Junfei Xiao, Sucheng Ren, Jieru Mei, Qing Liu, Huangjie Zheng, et al. What if we recapture billions of web images with llama-3? arXiv preprint arXiv:2406.08478, 2024. 2, 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 208, + 294, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 208, + 294, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 208, + 294, + 262 + ], + "type": "text", + "content": "[26] Yangguang Li, Feng Liang, Lichen Zhao, Yufeng Cui, Wanli Ouyang, Jing Shao, Fengwei Yu, and Junjie Yan. Supervision exists everywhere: A data efficient contrastive language-image pre-training paradigm. arXiv preprint arXiv:2110.05208, 2021. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 264, + 294, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 264, + 294, + 330 + ], + "spans": [ + { + "bbox": [ + 56, + 264, + 294, + 330 + ], + "type": "text", + "content": "[27] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 331, + 294, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 331, + 294, + 364 + ], + "spans": [ + { + "bbox": [ + 56, + 331, + 294, + 364 + ], + "type": "text", + "content": "[28] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 365, + 294, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 365, + 294, + 386 + ], + "spans": [ + { + "bbox": [ + 56, + 365, + 294, + 386 + ], + "type": "text", + "content": "[29] AI Meta. Introducing meta llama 3: The most capable openly available llm to date. Meta AI, 2024. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 388, + 294, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 388, + 294, + 432 + ], + "spans": [ + { + "bbox": [ + 56, + 388, + 294, + 432 + ], + "type": "text", + "content": "[30] Tomas Mikolov, Ilya Sutskever, Kai Chen, Greg S Corrado, and Jeff Dean. Distributed representations of words and phrases and their compositionality. Advances in neural information processing systems, 26, 2013. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 434, + 294, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 434, + 294, + 487 + ], + "spans": [ + { + "bbox": [ + 56, + 434, + 294, + 487 + ], + "type": "text", + "content": "[31] Hiroshi Morioka and Aapo Hyvarinen. Connectivity-contrastive learning: Combining causal discovery and representation learning for multimodal data. In International conference on artificial intelligence and statistics, pages 3399–3426. PMLR, 2023. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 489, + 294, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 489, + 294, + 533 + ], + "spans": [ + { + "bbox": [ + 56, + 489, + 294, + 533 + ], + "type": "text", + "content": "[32] Hiroshi Morioka and Aapo Hyvarinen. Causal representation learning made identifiable by grouping of observational variables. In *Forty-first International Conference on Machine Learning*, 2024. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 534, + 294, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 534, + 294, + 578 + ], + "spans": [ + { + "bbox": [ + 56, + 534, + 294, + 578 + ], + "type": "text", + "content": "[33] Norman Mu, Alexander Kirillov, David Wagner, and Saining Xie. Slip: Self-supervision meets language-image pretraining. In European conference on computer vision, pages 529-544. Springer, 2022. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 579, + 294, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 579, + 294, + 623 + ], + "spans": [ + { + "bbox": [ + 56, + 579, + 294, + 623 + ], + "type": "text", + "content": "[34] Ivona Najdenkoska, Mohammad Mahdi Derakhshani, Yuki M Asano, Nanne van Noord, Marcel Worring, and Cees GM Snoek. Tulip: Token-length upgraded clip. arXiv preprint arXiv:2410.10034, 2024. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 624, + 294, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 294, + 657 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 294, + 657 + ], + "type": "text", + "content": "[35] Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 2018. 1, 4, 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "type": "text", + "content": "[36] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 8" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 139 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 139 + ], + "type": "text", + "content": "[37] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 2, 4, 5, 6, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 140, + 553, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 140, + 553, + 184 + ], + "spans": [ + { + "bbox": [ + 316, + 140, + 553, + 184 + ], + "type": "text", + "content": "[38] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 186, + 555, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 186, + 555, + 239 + ], + "spans": [ + { + "bbox": [ + 316, + 186, + 555, + 239 + ], + "type": "text", + "content": "[39] Hanoona Rasheed, Muhammad Uzair Khattak, Muhammad Maaz, Salman Khan, and Fahad Shahbaz Khan. Fine-tuned clip models are efficient video learners. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6545-6554, 2023. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 241, + 555, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 241, + 555, + 297 + ], + "spans": [ + { + "bbox": [ + 316, + 241, + 555, + 297 + ], + "type": "text", + "content": "[40] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 298, + 553, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 298, + 553, + 331 + ], + "spans": [ + { + "bbox": [ + 316, + 298, + 553, + 331 + ], + "type": "text", + "content": "[41] Nils Sturma, Chandler Squires, Mathias Drton, and Caroline Uhler. Unpaired multi-domain causal representation learning, 2023. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 332, + 553, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 332, + 553, + 376 + ], + "spans": [ + { + "bbox": [ + 316, + 332, + 553, + 376 + ], + "type": "text", + "content": "[42] Mingkang Tang, Zhanyu Wang, Zhenhua Liu, Fengyun Rao, Dian Li, and Xiu Li. Clip4caption: Clip for video caption. In Proceedings of the 29th ACM International Conference on Multimedia, pages 4858-4862, 2021. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 377, + 553, + 433 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 377, + 553, + 433 + ], + "spans": [ + { + "bbox": [ + 316, + 377, + 553, + 433 + ], + "type": "text", + "content": "[43] Julius von Kugelgen, Yash Sharma, Luigi Gresele, Wieland Brendel, Bernhard Scholkopf, Michel Besserve, and Francesco Locatello. Self-supervised learning with data augmentations provably isolates content from style. arXiv preprint arXiv:2106.04619, 2021. 3, 4, 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 434, + 553, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 434, + 553, + 498 + ], + "spans": [ + { + "bbox": [ + 316, + 434, + 553, + 498 + ], + "type": "text", + "content": "[44] Julius von Kugelgen, Michel Besserve, Liang Wendong, Luigi Gresele, Armin Kekic, Elias Bareinboim, David Blei, and Bernhard Scholkopf. Nonparametric identifiability of causal representations from unknown interventions. Advances in Neural Information Processing Systems, 36, 2024. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 500, + 553, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 500, + 553, + 566 + ], + "spans": [ + { + "bbox": [ + 316, + 500, + 553, + 566 + ], + "type": "text", + "content": "[45] Haofan Wang, Zifan Wang, Mengnan Du, Fan Yang, Zijian Zhang, Sirui Ding, Piotr Mardziel, and Xia Hu. Score-cam: Score-weighted visual explanations for convolutional neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 24-25, 2020. 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 567, + 553, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 567, + 553, + 612 + ], + "spans": [ + { + "bbox": [ + 316, + 567, + 553, + 612 + ], + "type": "text", + "content": "[46] Jianyi Wang, Kelvin CK Chan, and Chen Change Loy. Exploring clip for assessing the look and feel of images. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2555-2563, 2023. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 613, + 553, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 613, + 553, + 657 + ], + "spans": [ + { + "bbox": [ + 316, + 613, + 553, + 657 + ], + "type": "text", + "content": "[47] Tongzhou Wang and Phillip Isola. Understanding contrastive representation learning through alignment and uniformity on the hypersphere. In International conference on machine learning, pages 9929-9939. PMLR, 2020. 4, 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 658, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 553, + 712 + ], + "type": "text", + "content": "[48] Wei Wu, Kecheng Zheng, Shuailei Ma, Fan Lu, Yuxin Guo, Yifei Zhang, Wei Chen, Qingpei Guo, Yujun Shen, and Zheng-Jun Zha. Lotlip: Improving language-image pre-training for long text understanding. arXiv preprint arXiv:2410.05249, 2024. 2" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "29789" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 609 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 137 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 137 + ], + "type": "text", + "content": "[49] Dingling Yao, Danru Xu, Sebastien Lachapelle, Sara Maglia-cane, Perouz Taslakian, Georg Martius, Julius von Kügelgen, and Francesco Locatello. Multi-view causal representation learning with partial observability. In The Twelfth International Conference on Learning Representations, 2024. 3, 4, 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 140, + 294, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 140, + 294, + 194 + ], + "spans": [ + { + "bbox": [ + 56, + 140, + 294, + 194 + ], + "type": "text", + "content": "[50] Peter Young, Alice Lai, Micah Hodosh, and Julia Hockenmaier. From image descriptions to visual denotations: New similarity metrics for semantic inference over event descriptions. Transactions of the Association for Computational Linguistics, 2:67-78, 2014. 6, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 197, + 294, + 238 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 197, + 294, + 238 + ], + "spans": [ + { + "bbox": [ + 56, + 197, + 294, + 238 + ], + "type": "text", + "content": "[51] Jiahui Yu, Zirui Wang, Vijay Vasudevan, Legg Yeung, Mojtaba Seyedhosseini, and Yonghui Wu. Coca: Contrastive captioners are image-text foundation models. arXiv preprint arXiv:2205.01917, 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 241, + 294, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 241, + 294, + 295 + ], + "spans": [ + { + "bbox": [ + 56, + 241, + 294, + 295 + ], + "type": "text", + "content": "[52] Xiaohua Zhai, Xiao Wang, Basil Mustafa, Andreas Steiner, Daniel Keysers, Alexander Kolesnikov, and Lucas Beyer. Lit: Zero-shot transfer with locked-image text tuning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18123-18133, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 297, + 294, + 340 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 297, + 294, + 340 + ], + "spans": [ + { + "bbox": [ + 56, + 297, + 294, + 340 + ], + "type": "text", + "content": "[53] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11975-11986, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 342, + 294, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 342, + 294, + 373 + ], + "spans": [ + { + "bbox": [ + 56, + 342, + 294, + 373 + ], + "type": "text", + "content": "[54] Beichen Zhang, Pan Zhang, Xiaoyi Dong, Yuhang Zang, and Jiaqi Wang. Long-clip: Unlocking the long-text capability of clip. arXiv preprint arXiv:2403.15378, 2024. 2, 4, 5, 6, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 376, + 294, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 376, + 294, + 430 + ], + "spans": [ + { + "bbox": [ + 56, + 376, + 294, + 430 + ], + "type": "text", + "content": "[55] Jiaqi Zhang, Kristjan Greenewald, Chandler Squires, Akash Srivastava, Karthikeyan Shanmugam, and Caroline Uhler. Identifiability guarantees for causal disentanglement from soft interventions. Advances in Neural Information Processing Systems, 36, 2024. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 432, + 294, + 473 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 432, + 294, + 473 + ], + "spans": [ + { + "bbox": [ + 56, + 432, + 294, + 473 + ], + "type": "text", + "content": "[56] Jihai Zhang, Xiaoye Qu, Tong Zhu, and Yu Cheng. Clipmoe: Towards building mixture of experts for clip with diversified multiplet upcycling. arXiv preprint arXiv:2409.19291, 2024. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 476, + 294, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 476, + 294, + 508 + ], + "spans": [ + { + "bbox": [ + 56, + 476, + 294, + 508 + ], + "type": "text", + "content": "[57] Kun Zhang, Shaoan Xie, Ignavier Ng, and Yujia Zheng. Causal representation learning from multiple distributions: A general setting. arXiv preprint arXiv:2402.05052, 2024. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 510, + 294, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 510, + 294, + 563 + ], + "spans": [ + { + "bbox": [ + 56, + 510, + 294, + 563 + ], + "type": "text", + "content": "[58] Renrui Zhang, Ziyu Guo, Wei Zhang, Kunchang Li, Xupeng Miao, Bin Cui, Yu Qiao, Peng Gao, and Hongsheng Li. Pointclip: Point cloud understanding by clip. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8552-8562, 2022. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 566, + 294, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 566, + 294, + 609 + ], + "spans": [ + { + "bbox": [ + 56, + 566, + 294, + 609 + ], + "type": "text", + "content": "[59] Kecheng Zheng, Yifei Zhang, Wei Wu, Fan Lu, Shuailei Ma, Xin Jin, Wei Chen, and Yujun Shen. Dreamclip: Language-image pre-training with long captions. arXiv preprint arXiv:2403.17007, 2024. 2" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "29790" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/SmartEraser_ Remove Anything from Images using Masked-Region Guidance/ab3bbba3-1da3-4883-858e-f047afa69964_content_list.json b/2025/SmartEraser_ Remove Anything from Images using Masked-Region Guidance/ab3bbba3-1da3-4883-858e-f047afa69964_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..83837b8d03d3ff7dbdcde9f2d940d54c38e985f1 --- /dev/null +++ b/2025/SmartEraser_ Remove Anything from Images using Masked-Region Guidance/ab3bbba3-1da3-4883-858e-f047afa69964_content_list.json @@ -0,0 +1,1698 @@ +[ + { + "type": "text", + "text": "SmartEraser: Remove Anything from Images using Masked-Region Guidance", + "text_level": 1, + "bbox": [ + 104, + 130, + 893, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Longtao Jiang $^{1,*}$ Zhendong Wang $^{1,*}$ Jianmin Bao $^{2,*\\dagger\\heartsuit}$ \nWengang Zhou $^{1,\\dagger}$ Dongdong Chen $^{2}$ Lei Shi $^{2}$ Dong Chen $^{2}$ Houqiang Li $^{1}$ $^{1}$ University of Science and Technology of China $^{2}$ Microsoft Research Asia", + "bbox": [ + 194, + 179, + 800, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "https://longtaojiang.github.io/smarteraser.github.io/", + "bbox": [ + 259, + 237, + 730, + 252 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c4720aa160d5b7008e62a16834eae9fde6110efeb03b337fcba703c311e9a059.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 91, + 263, + 294, + 368 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d5b7d52efe55e5d764fde0d5d783b2af33065824ee9f0ad13a68404c32fb5049.jpg", + "image_caption": [ + "Figure 1. Object removal results generated by SmartEraser using user-provided masks. SmartEraser not only generates photorealistic images with the target objects removed but also better preserves the surrounding context around the removed objects." + ], + "image_footnote": [], + "bbox": [ + 91, + 369, + 294, + 474 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1524c6c7a097df5eaa242b1884fb1b9c44b334932b615f3c476d24b5a24023fa.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 295, + 265, + 496, + 368 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/950fbb09f39b5ba636edd120fc083b91495dceaec2a736df2e137f3ecfd865a0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 295, + 369, + 496, + 474 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0fead885f7fc6f87d75b68fe360d3bd9468fe6c2c4c294e9e194c976221d5e3d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 265, + 700, + 369 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1200313b16ce45eb43d12a2561e36cc1b22309a7fa064d3eae940c1f154926ba.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 369, + 700, + 474 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7c5233145bfdde27eb941d041254a901206ab3a69142dbad05c724bc4d845d44.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 702, + 265, + 903, + 369 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/40d980fbb1cea1b99064ca7618696cf7cba2773baf30a3d4030bc508f548800f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 702, + 369, + 903, + 474 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 517, + 326, + 532 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Object removal has so far been dominated by the \"mask-and-inpaint\" paradigm, where the masked region is excluded from the input, leaving models relying on unmasked areas to inpaint the missing region. However, this approach lacks contextual information for the masked area, often resulting in unstable performance. In this work, we introduce SmartEraser, built with a new \"removing\" paradigm called Masked-Region Guidance. This paradigm retains the masked region in the input, using it as guidance for the removal process. It offers several distinct advantages: (a) it guides the model to accurately identify the object to be removed, preventing its regeneration in the output; (b) since the user mask often extends beyond the object itself, it aids in preserving the surrounding context in the final result. Leveraging this new paradigm, we present Syn4Removal, a large-scale object removal dataset, where instance segmentation data is used to copy and paste objects onto images as removal targets, with the original images serving as ground truths. Experimental results demonstrate that SmartEraser significantly outperforms existing methods, achieving superior performance in object removal, especially in complex scenes with intricate compositions.", + "bbox": [ + 88, + 550, + 485, + 883 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 514, + 517, + 643, + 532 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "With the rapid advancement of generative models [10, 11, 15, 32, 34, 35, 37, 40], image editing has garnered significant attention for its broad applications. Among various editing tasks, object removal [9, 55] is especially prominent, empowering users to seamlessly eliminate unwanted elements while maintaining the realism of original images. This functionality, essential for removing distracting objects in photos, has become a key feature in widely used applications like Photoshop, Google Photos, Microsoft Designer.", + "bbox": [ + 511, + 542, + 906, + 679 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Currently, most object removal methods [1, 9, 30, 39, 55] adopt a \"mask-and-inpaint\" paradigm, where the masked region is excluded from the input and typically filled with a neutral placeholder (e.g., black). The model then inpaints the masked region based on the surrounding content. However, we identify two primary issues with this approach. First, this strategy often generates unintended objects within the masked region. Lacking precise discrimination between removal targets and other content, these methods rely heavily on background context, sometimes leading to the unintentional inpainting of new objects—such as adding a new car on the road rather than removing the original one, as illustrated in Figure 2(a). Second, user-", + "bbox": [ + 511, + 683, + 908, + 880 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Equal contribution. † Corresponding authors. ⋆ Project leader.", + "bbox": [ + 535, + 886, + 887, + 901 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "24452", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2724c1b7c9d985aa928a4c861c6acf4c53978b066861cde15142b229639f606b.jpg", + "image_caption": [ + "(b) Our paradigm: Masked-Region Guidance", + "Figure 2. Comparison of the \"mask-and-inpaint\" paradigm and the proposed Masked-Region Guidance paradigm for object removal. Unlike the mask-and-inpaint approach, our paradigm uses the masked regions as critical guidance for object removal." + ], + "image_footnote": [], + "bbox": [ + 91, + 88, + 482, + 243 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "defined masks frequently exceed the target object, requiring the \"mask-and-inpaint\" approach to synthesize these extended region. This will inadvertently modify nearby context, reducing visual coherence.", + "bbox": [ + 89, + 320, + 482, + 381 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address these limitations, we introduce a novel paradigm for object removal called Masked-Region Guidance. The core idea is that the masked region should not be excluded but rather utilized as critical guidance during the removal process. Our method is straightforward: instead of replacing the masked region with a placeholder, we retain the original image as input, with the masked region indicated by a mask input, as in existing methods. As illustrated in Figure 2(b), this paradigm enables the model to accurately identify the target object, preventing unintended regeneration in the output and effectively preserving the surrounding context of the target object in the final result.", + "bbox": [ + 89, + 383, + 482, + 564 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, our new paradigm cannot be directly implemented with existing object removal data construction methods. Typically, their training data are created by masking parts of the image, and the model is trained to predict the masked content. If applying this approach to our proposed Masked-Region Guidance paradigm, the model could exploit a shortcut by simply replicating the masked content from the input, due to the masked region being included in the input. Thus, a dataset consisting of triplets of an input, a mask, and the removal result is essential for our paradigm. Unfortunately, existing related datasets [23, 36, 45, 48] either contain a limited number of unique scenes or rely on inpainting models to generate pseudo-removed results.", + "bbox": [ + 89, + 566, + 482, + 762 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To solve this, we introduce a synthetic technique to create training data specifically for object removal. Our approach involves pasting object instances from various images onto different background images to form the input images, with the pasted instance masks serving as the input masks and the original background images designated as ground truths. Using this method, we generate Syn4Removal, a large-scale dataset comprising triplets of real background images, masks, and backgrounds with", + "bbox": [ + 89, + 763, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "pasted objects. The design of Syn4Removal provides diverse scenes and supports effective training under our new paradigm, encouraging the model to accurately learn object removal without shortcuts.", + "bbox": [ + 511, + 90, + 903, + 150 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To make Syn4Removal suitable for training object removal models, we design a pipeline to generate high-quality data. First, we filter out low-quality instances and background images. Then, we develop a method to calculate feasible pasting locations on the image, ensuring that objects do not overlap with instances in the pasted area, which helps prevent the model from regenerating unwanted objects. Last, the instance is pasted onto a background image with a blending algorithm. The resulting dataset consists of 1 million image triplets.", + "bbox": [ + 511, + 151, + 906, + 301 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Using the Masked-Region Guidance paradigm and the Syn4Removal dataset, we design a framework based on text-to-image diffusion models for object removal. To enhance the model's robustness to varying mask shapes from user input, we introduce a mask enhancement technique that simulates different mask shapes over removal targets. Additionally, we incorporate CLIP-based visual guidance to assist the model in better understanding the removal targets. The resulting model called SmartEraser outperforms previous methods significantly in both quantitative and qualitative evaluations. We choose the name SmartEraser because the model can \"smartly\" identify removal targets and remove it while preserving its surrounding region.", + "bbox": [ + 511, + 303, + 905, + 498 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions are mainly summarized as follows:", + "bbox": [ + 532, + 500, + 883, + 513 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose Masked-Region Guidance, a novel paradigm for object removal, which effectively addresses issues of object regeneration and surrounding region distortion.", + "- We introduce Syn4Removal, a large-scale high-quality dataset for object removal, which contains over a million pairs of images across diverse scenes and object types.", + "- We carefully design a framework based on the text-to-image diffusion model by introducing mask enhancement and clip-based visual guidance.", + "- Extensive experiments demonstrate that SmartEraser significantly outperforms previous object removal methods in terms of quality and robustness." + ], + "bbox": [ + 513, + 513, + 905, + 694 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 709, + 653, + 724 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Image inpainting. Image inpainting aims to seamlessly fill missing regions in images by leveraging contextual information around the masked region. Based on the \"mask-and-inpaint\" paradigm, earlier works are mainly based on generative adversarial networks (GANs) [11, 31]. Typically, these methods [2, 4, 8, 24, 25, 27, 39, 43, 49] randomly mask part areas of images [19, 20, 28, 54] and then predict these masked regions. However, GAN-based approaches [12, 41, 42, 50, 51] often suffer from limited diversity and quality in generated content, resulting in blurry or unrealistic inpainted results.", + "bbox": [ + 511, + 734, + 905, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "24453", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recently, approaches [1, 6, 9, 18, 30, 35, 47, 55] based on diffusion models have attracted significant attention due to their superior capability in generating image details. Among them, Repaint [30] injects the known region into intermediate images at all timesteps throughout the diffusion generation. SD-Inpaint [35] utilizes the mask and cropped image as input and fills the masked region through diffusion. While these approaches produce detailed textures, the generated content within the masked region often remains uncertain, with new objects occasionally appearing.", + "bbox": [ + 89, + 90, + 480, + 241 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Object removal. The object removal task aims to eliminate specified objects from images based on user-provided masks, seamlessly repairing the removed area with un-masked content. For example, PowerPaint [55] uses learnable task embeddings from context-aware inpainting as the positive prompt and those from text-guided inpainting as the negative prompt to guide object removal. A recent work CLIPAway [9] leverages AlphaCLIP [38] and IP-Adapter [46] to inject background information into the model via cross-attention, ensuring removal results coherent with the background context. The common characteristic of these methods is that they retain the traditional \"mask-and-inpaint\" paradigm by discarding the masked regions of input images. In this paper, we claim that this paradigm is not suitable for the object removal task exactly due to the missing context of the masked region, causing incomplete removal and inconsistent context in removal results.", + "bbox": [ + 89, + 242, + 482, + 496 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We also observe that some instruction-based methods [3, 48] attempt to remove objects using text prompts instead of the masked region. While they use the original image as input, their models are trained with removal results generated by inpainting models, which limits their performance. We believe that our synthetic strategy for building removal datasets could enhance the performance of these models.", + "bbox": [ + 89, + 500, + 482, + 603 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "There are several works [23, 36, 45, 48] aiming to construct datasets for object removal. For instance, GQA-Inpaint [48] and DEFACTO [23] synthesize ground truths with the help of image inpainting models. However, we find that relying on image inpainting leads to the performance of object removal being highly limited by the effect of inpainting. Another two works RORD [36] and ObjectDrop [45] acquire real-world triplets through photography, but the high collection costs result in fewer than $3.5k$ unique scenes. To incorporate with the Masked-Region Guidance, we construct a large-scale, high-quality dataset called Syn4Removal through a set of carefully designed strategies, which includes about 1 million triplets featuring diverse scenes and authentic ground-truth backgrounds with carefully filtering and pasting strategies.", + "bbox": [ + 89, + 604, + 482, + 830 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. SmartEraser", + "text_level": 1, + "bbox": [ + 89, + 844, + 223, + 859 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The goal of object removal is to completely erase the target object specified by a masked region and seamlessly generate", + "bbox": [ + 89, + 869, + 482, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "consistent content based on the surrounding context. Most existing methods follow a \"mask-and-inpaint\" paradigm, where the masked area is excluded from the input, and the model attempts to inpaint the missing content based on the unmasked regions. While these methods can successfully remove objects in simple scenarios, they often regenerate new objects or produce blurs and artifacts in the masked region in more complex scenes.", + "bbox": [ + 511, + 90, + 903, + 210 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In contrast, we propose a novel paradigm: Masked-Region Guidance. Instead of discarding masked regions, we retain it in the input, leveraging it to guide the removal process. This approach is based on the insight that, by knowing the exact removal targets, the model can precisely erase them while leaving the surrounding context maintained. In the following sections, we will first introduce the method of Masked-Region Guidance in Sec. 3.1. Then, we present Syn4Removal, a large-scale synthetic dataset designed for this paradigm in Sec. 3.2. Finally, we describe SmartEraser, a fine-tuned diffusion model based on the Mask-Region Guidance paradigm and Syn4Removal dataset in Sec. 3.3.", + "bbox": [ + 511, + 212, + 903, + 393 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Masked-Region Guidance", + "text_level": 1, + "bbox": [ + 511, + 400, + 750, + 416 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given an input image $\\pmb{x}$ , and user-specified mask $m$ containing the removal targets, the object removal task aims to produce the removal result $\\hat{\\pmb{x}}$ while maintaining the realism of the original image. The mask $m$ is a binary mask where 1 indicates the region containing the removal targets. Existing models [9, 35, 39, 55] are mostly based on the \"mask-and-inpaint\" paradigm, i.e., utilizing the unmasked region $x \\odot (1 - m)$ and mask $m$ to predict the results without removal object in the masked region. These models are learned to maximize the following objective:", + "bbox": [ + 511, + 422, + 903, + 574 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} ^ {\\prime} = P \\left(\\hat {\\boldsymbol {x}} \\mid \\boldsymbol {m}, \\boldsymbol {x} \\odot (\\boldsymbol {1} - \\boldsymbol {m}); \\Theta^ {\\prime}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 596, + 575, + 903, + 592 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where the conditional probability $P$ is modeled using a neural network with parameters $\\Theta^{\\prime}$ , $\\odot$ specifies the elementwise multiplication.", + "bbox": [ + 511, + 599, + 903, + 642 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In contrast, we introduce Masked-Region Guidance, a fundamentally different approach. Instead of discarding the masked region, we retain it in the input, allowing it to guide the removal process. Specifically, we replace the condition input from $[m, x \\odot (1 - m)]$ to $[m, x]$ . The mask $m$ is still applied to instruct the region for removal targets, but $x$ now includes the masked region $x \\odot m$ besides $x \\odot (1 - m)$ . By including the masked region, we provide the model with valuable guidance during the object removal task. First, retaining the masked region allows the model to know exactly which objects need to be removed, reducing the risk of regenerating the object in the masked area. This helps the model focus on learning how to reconstruct the scene without the object. Second, the masked region is typically larger than the removal targets. This allows the model to copy the surrounding context of target objects in the masked region to the final result, significantly reducing the difficulty", + "bbox": [ + 511, + 643, + 903, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "24454", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/ffa3761d0ce36ce7857aaf8537c636054217d90785b19a2408a88edfb74a2bdd.jpg", + "image_caption": [ + "Figure 3. The data generation pipeline of Syn4Removal. We apply instances and images to construct a triplet consisting of an input image with removal targets, a mask, and the ground truth." + ], + "image_footnote": [], + "bbox": [ + 91, + 85, + 903, + 265 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "of synthesizing a larger area. Moreover, if the masked region contains complex textures, these can be retained, leading to more realistic results. In our Masked-Region Guidance paradigm, object removal is learned to maximize the objective:", + "bbox": [ + 89, + 309, + 483, + 383 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = P (\\hat {\\boldsymbol {x}} | \\boldsymbol {m}, \\boldsymbol {x}; \\Theta), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 217, + 388, + 482, + 405 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where the model is with parameters $\\Theta$ .", + "bbox": [ + 89, + 415, + 349, + 430 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Syn4Removal", + "text_level": 1, + "bbox": [ + 89, + 439, + 233, + 455 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Now we have a new paradigm for training object removal models, but it cannot be directly applied using existing data construction methods. Typically, training data is created by masking parts of an image and having the model predict the masked content. However, this could lead the model to simply copy the masked content if applied to our paradigm. To overcome this, we require a dataset with triplets (input image, mask, and ground truth). However, existing real image datasets [36, 45] are limited in terms of scene variety, as they require photos taken before and after placing objects. Additionally, some synthetic datasets [23, 48] rely on inpainting models to generate pseudo-removed results, which don't fully capture the real removal process.", + "bbox": [ + 89, + 462, + 483, + 657 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To address this, we propose a synthetic method for constructing object removal training data. Our approach involves pasting object instances from various images onto different backgrounds to form the input images. The mask of the pasted object serves as the input mask, and the original background image serves as the ground truth. Since the ground truth is the real image, it helps the model better capture the true distribution of the image, leading to more realistic object removal results. As a result, we constructed Syn4Removal, a large-scale high-quality data set that contains 1 million samples. The data construction pipeline is shown in Figure 3. Next, we will introduce each component of the overall pipeline.", + "bbox": [ + 89, + 659, + 483, + 854 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Instance Crop & Filter. We directly use a public instance segmentation dataset [22] to get instances. We use the mask to get each instance from the image. While cutting", + "bbox": [ + 89, + 854, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "instances, we also compute the corresponding CLIP similarity score with their classes and normalized area ratio in the original images. Then we perform the filter process, First, we exclude instances with areas larger than $95\\%$ or smaller than $5\\%$ to prevent potential low-quality segmentation results. Then, based on the previously calculated CLIP scores, we remove instances with low semantic relevance in each class. Given the variability in CLIP score distribution across different classes, we set individual CLIP score thresholds for each class according to previous work [53], filtering out instances with scores below these thresholds.", + "bbox": [ + 511, + 309, + 906, + 474 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Background Filter. For background images, we use public datasets COCONut [7] and SAM-1B [21], along with segmentation annotations. We filter out low-resolution images, those with extreme aspect ratios, or those with a large number of segmented instances. The selected backgrounds are then combined to form 1 million high-quality images (for more details, see Supplementary).", + "bbox": [ + 511, + 474, + 905, + 580 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Instance Size Sampling & Resize. To match the size of the instance with that of the background, we first calculate the average $\\mu_{c}$ and variance $\\sigma_c^2$ of the normalized area ratio for each class $c$ . When pasting an instance onto a background, we sample scale $s$ from a Gaussian distribution $\\mathcal{N}(\\mu_c,\\sigma_c^2)$ , and scale the instance to $sHW$ , where $H$ and $W$ are the height and width of the background image, respectively.", + "bbox": [ + 511, + 580, + 905, + 686 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Feasible Area for Pasting. To ensure that the pasted instance does not cover existing instances in the background image, we calculate IoU between the pasted instance region and each instance in the background image, the feasible region based on IoU results is formulated as:", + "bbox": [ + 511, + 686, + 905, + 761 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {R} _ {1} = \\left\\{\\left(x, y\\right) \\mid \\operatorname {I o U} \\left(P _ {x, y}, I _ {i}\\right) < r, I _ {i} \\in [ 1, I _ {\\max } ] \\right\\}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 531, + 762, + 903, + 780 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $P_{x,y}$ is the instance region if centered on $(x,y)$ , $r$ is a IoU threshold, $I_{max}$ is the number of existing instances. Besides, to ensure the integrity of pasted instances, edge areas of background images are avoided, formulated as:", + "bbox": [ + 511, + 786, + 905, + 845 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {R} _ {2} = \\left\\{(x, y) \\mid x \\in \\left[ e _ {x}, W - e _ {x} \\right], y \\in \\left[ e _ {y}, H - e _ {y} \\right] \\right\\}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 527, + 847, + 903, + 864 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $e_x$ and $e_y$ are half of the width and height of the bounding box of the instance mask, respectively. We take", + "bbox": [ + 511, + 869, + 905, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "24455", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/bde91619f2b29f08ca714a5c0674925bbcf49b13b1a8e4df43c2cf22b67b61c0.jpg", + "image_caption": [ + "Figure 4. Overall framework of SmartEraser. It fully utilizes the Masked-Region Guidance paradigm and the Syn4Removal dataset." + ], + "image_footnote": [], + "bbox": [ + 94, + 88, + 903, + 256 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "the intersection region of $\\mathbf{R}_1$ and $\\mathbf{R}_2$ , the final set of feasible locations is as follows:", + "bbox": [ + 89, + 287, + 482, + 316 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {R} _ {\\mathrm {f}} = \\mathbf {R} _ {1} \\cap \\mathbf {R} _ {2}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 230, + 323, + 482, + 339 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We then randomly select a location from $\\mathbf{R}_{\\mathrm{f}}$ as the center point for pasting the instance. Based on the mask of the instance and location, we can get the instance mask centered on the location $\\bar{m}$ .", + "bbox": [ + 89, + 351, + 483, + 409 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Alpha Blending for Instance Pasting. To further ensure that the pasted instance blends harmoniously with the background, we apply alpha blending [13]. This method interpolates between the foreground instance and the background image, formulated as follows:", + "bbox": [ + 89, + 411, + 483, + 486 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {x} = \\boldsymbol {\\alpha} \\odot \\boldsymbol {x} _ {i} + (1 - \\boldsymbol {\\alpha}) \\odot \\boldsymbol {x} _ {b}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 184, + 493, + 480, + 508 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\pmb{x}$ is final constructed image, $\\pmb{x}_i$ is the instance, $\\pmb{x}_b$ is the background image, and $\\alpha$ is alpha blending value. $\\alpha$ is calculated through an alpha blending algorithm [13] that processes the mask $\\bar{m}$ to a trimap and then takes the trimap as input, which consists of three regions, i.e., the foreground and background, and an excessively uncertain boundary region. So we have the triplet for object removal training: $\\{\\pmb{x}, \\bar{m}, \\pmb{x}_b\\}$ , where $\\pmb{x}$ is input image with target object, $\\bar{m}$ is original mask, and $\\pmb{x}_b$ as the ground truth. Using this data generation pipeline, we finally get about 1 million triplets that form the Syn4Removal dataset.", + "bbox": [ + 89, + 521, + 483, + 686 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Framework", + "text_level": 1, + "bbox": [ + 89, + 696, + 218, + 709 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Based on the proposed Masked-Region Guidance paradigm and the Syn4Removal dataset, we design a framework based on the text-to-image stable diffusion model for object removal. In this framework, we carefully design a user-friendly mask enhancement strategy and introduce CLIP-based visual guidance to further improve the capability of the model for object removal, as shown in Figure 4.", + "bbox": [ + 89, + 719, + 483, + 824 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Mask Enhancement. In real-world scenarios, users usually provide a loose or tight mask around the object to be removed. If the model were trained only with precise object masks, there would be a significant gap in mask shape and size between training and inference. To address this,", + "bbox": [ + 89, + 825, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/1ec9c936a0816227a4f7db1679568ef8d5454fc1493cd810f188aaf1688b11a0.jpg", + "image_caption": [ + "Figure 5. Mask shapes from different mask enhancement methods." + ], + "image_footnote": [], + "bbox": [ + 514, + 280, + 905, + 489 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "during training, we apply various mask deformation methods to simulate the user input mask shapes, as shown in Figure 5. These techniques help the model generalize to different mask forms. Specifically, we use six mask types to augment the object mask: (1) Original mask: this mask precisely outlines the target object. (2) Eroded mask: created by applying morphological erosion to the original mask, simulating scenarios where the user's mask may not fully cover the object. (3) Dilated mask: created by applying morphological dilation to the original mask. (4) Convex hull mask: constructed by calculating the convex hull that fully encompasses the original mask, then expanding it slightly. (5) Ellipse mask: generated by finding the smallest enclosing ellipse around the original mask, followed by a slight expansion. (6) Bbox & Bessel mask: a bounding box is generated to completely cover the original mask first, then bessel curves are constructed along each edge of the box to introduce an irregular boundary. So the overall mask enhancement process is illustrated as follows:", + "bbox": [ + 511, + 513, + 906, + 800 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {m} = \\operatorname {M E} _ {i} (\\bar {\\boldsymbol {m}}), \\quad i \\in \\{1, \\dots , 6 \\}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 599, + 809, + 903, + 825 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $m$ is the enhanced mask used during training, $\\mathrm{ME}_i(\\cdot)$ represents six different mask enhancement methods.", + "bbox": [ + 511, + 839, + 903, + 868 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "CLIP-Based Visual Guidance. Our SmartEraser is designed based on text-to-image diffusion model, the text", + "bbox": [ + 511, + 869, + 906, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "24456", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/c0ba79f5fd51e9a5cc468b5ccaf2bce8745d0cba7961ed14ad5087cf6ed9f161.jpg", + "image_caption": [ + "Figure 6. Qualitative comparison of SmartEraser and state-of-the-art methods. The samples are sourced from RORD-Val, which includes the ground truth after removal." + ], + "image_footnote": [], + "bbox": [ + 93, + 88, + 906, + 330 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "prompts play an important role in guiding the generation process. We aim to design a prompt that indicates the model what to remove during the generation process. Concretely, we first segment the region corresponding to the removal target in the image using its corresponding enhanced mask. This segmented region is then fed into the pre-trained CLIP [33] vision encoder to extract its visual feature(e.g., class token). This feature is not precisely aligned with the text encoder output space. So we then apply a trainable MLP network to map the visual feature into the feature space of the text encoder. The mapped embedding is then appended to the tokens of the CLIP text embedding of the prompt \"Remove the instance of\". During the training process, the vision encoder is frozen, and the MLP and text encoder are trainable. So the CLIP-based visual guidance is formulated as follows:", + "bbox": [ + 88, + 375, + 485, + 616 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\widehat {\\boldsymbol {c}} = \\left[ \\boldsymbol {\\tau} _ {\\theta} (\\boldsymbol {y}), \\operatorname {M L P} \\left(\\boldsymbol {\\nu} _ {\\theta} (\\boldsymbol {x} \\odot \\boldsymbol {m})\\right) \\right], \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 176, + 619, + 480, + 638 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\tau_{\\theta}(\\cdot)$ is text encoder, $\\nu_{\\theta}(\\cdot)$ is vision encoder, and $\\pmb{y}$ represents input text prompt \"remove the instance of\".", + "bbox": [ + 89, + 643, + 482, + 675 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Loss Function. Suppose the $E(\\cdot)$ is the VAE encoder, so the latent feature of input image $x$ and its ground truth $x_{b}$ is $\\bar{z} = E(x)$ and $z = E(x_{b})$ , respectively, and $z_{t}$ is the noisy feature of $z$ with added noise $\\epsilon$ at timestep $t$ . So the overall training loss following the standard diffusion process, which is formulated as follows:", + "bbox": [ + 89, + 675, + 482, + 763 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathbb {E} _ {\\boldsymbol {x}, \\boldsymbol {m}, \\boldsymbol {x} _ {b}, t, \\epsilon} \\| \\boldsymbol {\\epsilon} - \\boldsymbol {\\epsilon} _ {\\theta} \\left(\\boldsymbol {z} _ {t}, \\bar {\\boldsymbol {z}}, \\boldsymbol {m}, \\widehat {\\boldsymbol {c}}, t\\right) \\| _ {2} ^ {2}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 145, + 768, + 482, + 786 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\epsilon_{\\theta}(\\cdot)$ is the diffusion model.", + "bbox": [ + 89, + 794, + 323, + 810 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 821, + 223, + 838 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Settings", + "text_level": 1, + "bbox": [ + 89, + 847, + 189, + 863 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines. We compare SmartEraser with methods in both image inpainting and object removal. For im-", + "bbox": [ + 89, + 869, + 482, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "age inpainting, we evaluate against both GAN-based approaches (ZITS++ [4], MAT [24], LaMa [39], BLD [1]) and diffusion-based methods (RePaint [30], and SD-Inpaint [35]). Additionally, we compare it with specialized object removal frameworks, including CLIPAway [9] and PowerPaint [55], which represent the previous state-of-the-art performance on object removal benchmarks.", + "bbox": [ + 511, + 375, + 906, + 479 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation details. To enable a fair comparison, we follow previous experimental settings [9, 55] and finetune SmartEraser on the widely-used SD v1.5 model [35]. The training process involves fine-tuning both the UNet and text encoder components within the diffusion model and MLP for mapping visual features to text features space. SmartEraser is trained on the proposed Syn4Removal dataset with a batch size of 32, using the AdamW [29] optimizer for $500k$ iterations and a learning rate of 2e-5 across all trainable modules.", + "bbox": [ + 511, + 484, + 908, + 635 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation benchmarks. We evaluate the object removal capability of all baseline methods and our approach across three benchmarks in diverse scenarios. The benchmarks are constructed as follows:", + "bbox": [ + 511, + 638, + 906, + 699 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- RORD-Val: Due to there are many triplets share the same background images as the ground truth in validation set of the RORD [36]. To ensure unique scenes and high object quality, we filter this subset to obtain RORD-Val, a evaluation benchmark of real-photography image pairs.", + "- DEFACTO-Val: Derived from the splicing section of DEFACTO [23], we exclude examples where intended targets for removal significantly overlap with instances in the background images.", + "- Syn4Removal-Val: Generated using the pipeline of Syn4Removal with MSCOCO [26] validation set as background images and OpenImages-v7 [22] validation set offering instances." + ], + "bbox": [ + 511, + 704, + 905, + 898 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "24457", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/9f763f1915b7ddd2dfce4110df6995036c940ee7a85c8968355622aca55be346.jpg", + "image_caption": [ + "Figure 7. Qualitative comparison of SmartEraser and other methods in real-world user cases, in which users provide various mask shapes. The samples are sourced from real-world images and the validation set of MSCOCO [26]." + ], + "image_footnote": [], + "bbox": [ + 91, + 88, + 906, + 353 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Each benchmark comprises approximately 5k samples. During inference, following [9], we apply slight mask dilation to mitigate downscaling artifacts in SD-Inpaint-based methods [9, 35, 55].", + "bbox": [ + 89, + 390, + 482, + 450 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Evaluation metrics. To quantitatively assess the performance of the object removal models, we consider three key aspects: (1) overall image quality, which is evaluated by Fréchet Inception Distance (FID) [14] and CLIP Maximum Mean Discrepancy (CMMD) [17]. (2) consistency between the predicted region and the background context, evaluated by REMOVE [5] metric. (3) consistency between the predicted region and corresponding region in the ground truth, assessed by LPIPS [52], SSIM [44], and PSNR [16].", + "bbox": [ + 89, + 452, + 483, + 590 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Comparison with Previous Methods", + "text_level": 1, + "bbox": [ + 89, + 602, + 411, + 618 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Quantitative results. We conduct extensive experiments to evaluate the performance of SmartEraser against baselines on the benchmarks. The results are shown in Tables 1, 2 and 3. SmartEraser significantly outperforms all other methods across all metrics on the three benchmarks. Notably, on the more complex RORD-Val dataset, which has larger masks and more challenging removal scenes, SmartEraser surpasses the previous state-of-the-art by margins of 10.3 in FID and 1.89dB in PSNR. This indicates the superior performance of SmartEraser for object removal. Besides, SmartEraser also achieves state-of-the-art performance on the simpler DEFACTO-Val and Syn4Removal-Val datasets.", + "bbox": [ + 89, + 627, + 482, + 808 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Qualitative results. We present visualization comparison results of object removal using images from real scenes, shown in Figure 6. Baseline methods have limitations and may introduce unintended objects, create unrealistic structures, or produce artifacts within masked regions. In contrast, SmartEraser succeeds in effectively removing target", + "bbox": [ + 89, + 810, + 483, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/f5a4df3f0c2a4ce62e39de378685b7bf9d10343d7b8950f4e6e923e8ad00e19d.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodFID ↓ CMMD ↓ReMOVE ↑LPIPS ↓SSIM ↑PSNR ↑
ZITS++ [4]53.4400.8380.8200.3880.52215.442
MAT [24]80.9531.4940.6280.4330.49513.322
LaMa [39]24.2370.2160.9160.3480.55716.383
BLD [1]73.0861.7600.8340.4710.46715.775
RePaint [30]71.5421.6390.8120.3810.51015.861
SD-Inpaint [35]69.5020.3240.8570.3690.53716.111
CLIPAway [9]25.4580.1230.9150.3330.57717.434
PowerPaint [55]24.0580.2940.9260.3080.60218.100
SmartEraser16.0300.0920.9370.2760.61219.994
", + "bbox": [ + 519, + 387, + 903, + 525 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/8874f0199269727aa30f0d17343a87179d5b04ce65e3f04e18f713176b19287c.jpg", + "table_caption": [ + "Table 1. Quantitative comparison of SmartEraser and previous methods on RORD-Val." + ], + "table_footnote": [], + "table_body": "
MethodFID ↓CMMD ↓ReLUVE ↑LPIPS ↓SSIM ↑PSNR ↑
ZITS++ [4]8.1530.2290.8990.3500.63419.453
MAT [24]15.2570.3210.7420.4070.57514.541
LaMa [39]6.6010.2010.9260.3510.68321.735
BLD [1]24.4070.7460.9130.4180.63219.346
RePaint [30]12.3140.8220.9240.3250.64420.607
SD-Inpaint [35]11.1190.1960.9120.3200.65621.691
CLIPAway [9]8.3200.1930.9250.3230.66621.953
PowerPaint [55]6.5820.2750.9340.2780.71523.870
SmartEraser3.4050.1060.9390.2570.73425.363
", + "bbox": [ + 519, + 555, + 903, + 691 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Quantitative comparison of SmartEraser and previous methods on DEFACTO-Val.", + "bbox": [ + 511, + 694, + 903, + 722 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "objects while generating high-quality synthesized results, and excels in maintaining the contextual integrity and visual coherence of the scene.", + "bbox": [ + 511, + 732, + 903, + 777 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "User study. To assess the visual quality subjectively, we conduct a user study with 30 participants. The study includes 20 image sets, each containing an original image, its mask, and removal results from four different methods, presented in random order. Participants are asked to select their preferred result based on three criteria: overall image quality, foreground-background consistency, and accuracy of object removal. Voting percentages in Table 4 show that", + "bbox": [ + 511, + 779, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "24458", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/c5259c74a583675b695084987dfb4732441ed55eefcec5e24f7c6da5073af30e.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodFID ↓ CMMD ↓ReMOVE ↑LPIPS ↓SSIM ↑PSNR ↑
ZITS++ [4]5.9320.1930.8100.3580.58117.315
MAT [24]19.2630.3000.7000.3740.57013.357
LaMa [39]5.2460.0970.9320.3250.60619.319
BLD [1]20.2860.7050.8960.4640.52516.539
RePaint [30]12.6770.6960.9080.3780.55317.689
SD-Inpaint [35]11.5280.1070.8970.3480.57218.105
CLIPAway [9]5.5300.0930.9300.3440.59318.561
PowerPaint [55]5.0250.2000.9310.2910.66320.594
SmartEraser4.3860.0530.9390.2690.67222.029
", + "bbox": [ + 96, + 88, + 478, + 224 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/09b87f3d7b2bda7a6f8f3e406de6bfb9dbcaf6f038406e914cb67dfeec15b7ff.jpg", + "table_caption": [ + "Table 3. Quantitative comparison of SmartEraser and previous methods on Syn4Removal-Val." + ], + "table_footnote": [ + "Table 4. Comparison of user study results in terms of image quality, foreground-background consistency, and removal accuracy." + ], + "table_body": "
MethodImage Quality ↑ FB-Consistency ↑Removal Accuracy ↑
LaMa [39]0.100.11
CLIPAway [9]0.180.12
PowerPaint [55]0.160.12
SmartEraser0.560.65
0.58
", + "bbox": [ + 96, + 263, + 478, + 338 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "users consistently favor the results from SmartEraser.", + "bbox": [ + 89, + 378, + 442, + 393 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Real-world user cases. To further evaluate the object removal capability of SmartEraser in real-world user scenarios, we conduct a comparison with other methods (LaMa [39], CLIPAway [9], PowerPaint [55]), using masks provided by users. Figure 7 shows the comparison results. We can observe that SmartEraser exhibits a remarkable ability to accurately identify and remove target objects while preserving surrounding contextual details, particularly in scenarios involving imperfect or approximate masks. In the sample in the bottom row on the left, SmartEraser accurately identifies and removes the person indicated by the mask, while maintaining the integrity of the elephant in the background. In contrast, alternative methods tend to repaint the entire masked area, which leads to outputs that lack contextual coherence and visual realism due to the loss of details of the elephant. The upper second sample on the right provides a challenging scenario, the user-provided mask aims to remove the man whose feet are partially obscured by floating text. The previous methods fail to maintain the integrity of the floating text. However, SmartEraser accurately removes the man while preserving the floating text. This highlights SmartEraser's ability to effectively differentiate between removal targets and background content, even in cases of partial occlusion.", + "bbox": [ + 89, + 393, + 483, + 758 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Ablation studies", + "text_level": 1, + "bbox": [ + 89, + 770, + 251, + 786 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To evaluate the effectiveness of the key components in our framework, we perform ablation studies and report both quantitative and qualitative results. We progressively modify the baseline (fine-tuning SD v1.5 with the \"mask-and-inpaint\" paradigm) by adding our proposed techniques: RG for masked-region guidance, ME for mask enhancement, and VG for CLIP-based visual guidance. The results are", + "bbox": [ + 89, + 794, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5e68ffc8c8dae8acd2a6323e074ab130e147e7574cabc0ab29005c4220885e15.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 516, + 88, + 594, + 143 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e5f467d86aa545c11068c3862854136cfe8dbf4500e006cbf06e208af2dca5ca.jpg", + "image_caption": [ + "Mask & Image", + "Figure 8. Qualitative ablation comparison of our method. From left to right, we progressively add each proposed component." + ], + "image_footnote": [], + "bbox": [ + 516, + 143, + 594, + 196 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/fd28d4db82f884c0d529aed183aeb22307ea5e1745be12ff4c3ee8d197b6b00a.jpg", + "image_caption": [ + "Baseline" + ], + "image_footnote": [], + "bbox": [ + 594, + 88, + 669, + 196 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4d20676228b19ff6981570d2a1193dec52951511022b5847d5ef650df0a7519e.jpg", + "image_caption": [ + "+RG" + ], + "image_footnote": [], + "bbox": [ + 669, + 88, + 750, + 196 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/adf50fc44ea31356e6415db0de1be2cb821d51732e5a0801a532acf543b71db3.jpg", + "image_caption": [ + "+ME" + ], + "image_footnote": [], + "bbox": [ + 750, + 88, + 826, + 196 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/85f23e43d152e2de2399f16cb696b58fa079e34fe7ac026208cacefc02e3d856.jpg", + "image_caption": [ + "+VG" + ], + "image_footnote": [], + "bbox": [ + 826, + 88, + 903, + 196 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/982bf1c7d6d6ea4a3e94ebddcdd3ebd7243fe9bee8145a3c30fc9a83af8eab18.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodFID ↓CMMD ↓ReLUVE ↑LPIPS ↓SSIM ↑PSNR ↑
Baseline33.1310.2450.9080.3550.54416.535
+RG24.5260.1210.9230.3390.56217.176
+ME19.4210.1080.9310.3010.59118.917
+VG16.0300.0920.9370.2760.61619.994
", + "bbox": [ + 519, + 251, + 901, + 330 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. Quantitative ablated comparison on RORD-Val.", + "bbox": [ + 540, + 334, + 877, + 347 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "displayed in Figure 8 and Table 5.", + "bbox": [ + 511, + 356, + 740, + 371 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of the Masked-Region Guidance paradigm: We observe that the baseline model often regenerates objects instead of properly removing them, as it is trained in a traditional \"inpainting\" manner. Incorporating masked-region guidance significantly improves the removal performance by reducing the risk of object regeneration. However, due to the mismatch between the mask shape used during training and the user-defined mask during testing, the model does not always perfectly remove the target objects and preserve the surrounding context.", + "bbox": [ + 511, + 372, + 905, + 523 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Contribution of Mask Enhancement: By simulating real-world user mask shapes over object removal during training, we find that it helps reduce the gap between model training and testing using user input mask, significantly keeping the surrounding regions and further improving removal results.", + "bbox": [ + 511, + 523, + 905, + 599 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Value of CLIP-based Visual Guidance: CLIP visual guidance aids the model in better understanding the removal targets, yielding the best performance on RORD-val across all metrics. This indicates that the visual guidance helps the model perform object removal tasks more effectively.", + "bbox": [ + 511, + 599, + 903, + 676 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 693, + 633, + 709 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In conclusion, we present SmartEraser, a state-of-the-art object removal model built on the novel Masked-Region Guidance paradigm. SmartEraser outperforms existing methods by smartly identifying the target object to remove while effectively preserving the surrounding context. To facilitate research on this paradigm, we propose Syn4Removal, a large-scale, high-quality dataset containing over a million image triplets, specifically designed for object removal tasks. Through extensive experiments, we demonstrate that SmartEraser achieves superior performance in both quality and robustness compared to previous object removal methods.", + "bbox": [ + 511, + 719, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "24459", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Omri Avrahami, Ohad Fried, and Dani Lischinski. Blended latent diffusion. ACM transactions on graphics (TOG), 42 (4):1-11, 2023. 1, 3, 6, 7, 8", + "[2] Connelly Barnes, Eli Shechtman, Adam Finkelstein, and Dan B Goldman. Patchmatch: A randomized correspondence algorithm for structural image editing. ACM Trans. Graph., 28(3):24, 2009. 2", + "[3] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. arXiv preprint arXiv:2211.09800, 2022. 3", + "[4] Chenjie Cao, Qiaole Dong, and Yanwei Fu. Zits++: Image inpainting by improving the incremental transformer on structural priors. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 2, 6, 7, 8", + "[5] Aditya Chandrasekar, Goirik Chakrabarty, Jai Bardhan, Ramya Hebbalaguppe, and Prathosh AP. Remove: A reference-free metric for object erasure. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 7901-7910, 2024. 7", + "[6] Pau de Jorge, Riccardo Volpi, Puneet K. Dakania, Philip H. S. Torr, and Rogez Gregory. Placing objects in context via inpainting for out-of-distribution segmentation. In The European Conference on Computer Vision (ECCV), 2024. 3", + "[7] Xueqing Deng, Qihang Yu, Peng Wang, Xiaohui Shen, and Liang-Chieh Chen. Coconut: Modernizing coco segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21863-21873, 2024. 4", + "[8] Qiaole Dong, Chenjie Cao, and Yanwei Fu. Incremental transformer structure enhanced image inpainting with masking positional encoding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11358-11368, 2022. 2", + "[9] Yigit Ekin, Ahmet Burak Yildirim, Erdem Eren Caglar, Aykut Erdem, Erkut Erdem, and Aysegul Dundar. Clipaway: Harmonizing focused embeddings for removing objects via diffusion models, 2024. 1, 3, 6, 7, 8", + "[10] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12873-12883, 2021. 1", + "[11] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 1, 2", + "[12] Xiefan Guo, Hongyu Yang, and Di Huang. Image inpainting via conditional texture and structure dual generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 14134-14143, 2021. 2", + "[13] Kaiming He, Christoph Rhemann, Carsten Rother, Xiaou Tang, and Jian Sun. A global sampling method for alpha matting. In CVPR 2011, pages 2049-2056, 2011. 5", + "[14] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilib-" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "rium. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2017. 7", + "[15] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1", + "[16] Alain Hore and Djemel Ziou. Image quality metrics: Psnr vs. sSIM. In 2010 20th international conference on pattern recognition, pages 2366-2369. IEEE, 2010. 7", + "[17] Sadeep Jayasumana, Srikumar Ramalingam, Andreas Veit, Daniel Glasner, Ayan Chakrabarti, and Sanjiv Kumar. Rethinking fid: Towards a better evaluation metric for image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9307-9315, 2024. 7", + "[18] Xuan Ju, Xian Liu, Xintao Wang, Yuxuan Bian, Ying Shan, and Qiang Xu. Brushnet: A plug-and-play image inpainting model with decomposed dual-branch diffusion, 2024. 3", + "[19] Tero Karras. Progressive growing of gans for improved quality, stability, and variation. arXiv preprint arXiv:1710.10196, 2017. 2", + "[20] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 2", + "[21] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 4", + "[22] Alina Kuznetsova, Hassan Rom, Neil Alldrin, Jasper Uijlings, Ivan Krasin, Jordi Pont-Tuset, Shahab Kamali, Stefan Popov, Matteo Malloci, Alexander Kolesnikov, et al. The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale. International journal of computer vision, 128(7):1956-1981, 2020. 4, 6", + "[23] Jens Lehmann, Daniel Gerber, Mohamed Morsey, and Axel-Cyrille Ngonga Ngomo. Defacto-deep fact validation. In International semantic web conference, pages 312-327. Springer, 2012. 2, 3, 4, 6", + "[24] Wenbo Li, Zhe Lin, Kun Zhou, Lu Qi, Yi Wang, and Jiaya Jia. Mat: Mask-aware transformer for large hole image inpainting. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10758-10768, 2022. 2, 6, 7, 8", + "[25] Xiaoguang Li, Qing Guo, Di Lin, Ping Li, Wei Feng, and Song Wnag. Misf: Multi-level interactive siamese filtering for high-fidelity image inpainting. CVPR, 2022. 2", + "[26] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 6, 7", + "[27] Qiankun Liu, Zhentao Tan, Dongdong Chen, Qi Chu, Xiyang Dai, Yinpeng Chen, Mengchen Liu, Lu Yuan, and Nenghai Yu. Reduce information loss in transformers for pluralistic" + ], + "bbox": [ + 516, + 93, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "24460", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "image inpainting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11347-11357, 2022. 2", + "[28] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaou Tang. Deep learning face attributes in the wild. In Proceedings of International Conference on Computer Vision (ICCV), 2015. 2", + "[29] I Loshchilov. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6", + "[30] Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11461-11471, 2022. 1, 3, 6, 7, 8", + "[31] Mehdi Mirza. Conditional generative adversarial nets. arXiv preprint arXiv:1411.1784, 2014. 2", + "[32] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 1", + "[33] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 6", + "[34] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 1", + "[35] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684-10695, 2022. 1, 3, 6, 7, 8", + "[36] Min-Cheol Sagong, Yoon-Jae Yeo, Seung-Won Jung, and Sung-Jea Ko. Rord: A real-world object removal dataset. In BMVC, page 542, 2022. 2, 3, 4, 6", + "[37] Yang Song, Prafulla Dhariwal, Mark Chen, and Ilya Sutskever. Consistency models. arXiv preprint arXiv:2303.01469, 2023. 1", + "[38] Zeyi Sun, Ye Fang, Tong Wu, Pan Zhang, Yuhang Zang, Shu Kong, Yuanjun Xiong, Dahua Lin, and Jiaqi Wang. Alphaclip: A clip model focusing on wherever you want. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13019-13029, 2024. 3", + "[39] Roman Suvorov, Elizaveta Logacheva, Anton Mashikhin, Anastasia Remizova, Armenii Ashukha, Aleksei Silvestrov, Naejin Kong, Harshith Goka, Kiwoong Park, and Victor Lempitsky. Resolution-robust large mask inpainting with fourier convolutions. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 2149–2159, 2022. 1, 2, 3, 6, 7, 8", + "[40] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017. 1" + ], + "bbox": [ + 91, + 90, + 483, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[41] Ziyu Wan, Jingbo Zhang, Dongdong Chen, and Jing Liao. High-fidelity pluralistic image completion with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 4692-4701, 2021. 2", + "[42] Wentao Wang, Jianfu Zhang, Li Niu, Haoyu Ling, Xue Yang, and Liqing Zhang. Parallel multi-resolution fusion network for image inpainting. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 14559-14568, 2021. 2", + "[43] Wentao Wang, Li Niu, Jianfu Zhang, Xue Yang, and Liqing Zhang. Dual-path image inpainting with auxiliary gan inversion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11421-11430, 2022. 2", + "[44] Zhou Wang, A.C. Bovik, H.R. Sheikh, and E.P. Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing, 13(4): 600-612, 2004. 7", + "[45] Daniel Winter, Matan Cohen, Shlomi Fruchter, Yael Pritch, Alex Rav-Acha, and Yedid Hoshen. Objectdrop: Bootstrapping counterfactuals for photorealistic object removal and insertion. arXiv preprint arXiv:2403.18818, 2024. 2, 3, 4", + "[46] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721, 2023. 3", + "[47] Yifu Chen, Jingwen Chen, Yingwei Pan, Yehao Li, Ting Yao, Zhineng Chen and Tao Mei. Improving text-guided object inpainting with semantic pre-inpainting. ECCV, 2024. 3", + "[48] Ahmet Burak Yildirim, Vedat Baday, Erkut Erdem, Aykut Erdem, and Aysegul Dundar. Inst-inpaint: Instructing to remove objects with diffusion models. arXiv preprint arXiv:2304.03246, 2023. 2, 3, 4", + "[49] Jiahui Yu, Zhe Lin, Jimei Yang, Xiaohui Shen, Xin Lu, and Thomas S Huang. Generative image inpainting with contextual attention. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5505-5514, 2018. 2", + "[50] Yingchen Yu, Fangneng Zhan, Shijian Lu, Jianxiong Pan, Feiying Ma, Xuansong Xie, and Chunyan Miao. Wavefill: A wavelet-based generation network for image inpainting. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 14114-14123, 2021. 2", + "[51] Yu Zeng, Zhe Lin, Huchuan Lu, and Vishal M. Patel. Crfill: Generative image inpainting with auxiliary contextual reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 14164-14173, 2021. 2", + "[52] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 7", + "[53] Hanqing Zhao, Dianmo Sheng, Jianmin Bao, Dongdong Chen, Dong Chen, Fang Wen, Lu Yuan, Ce Liu, Wenbo Zhou, Qi Chu, et al. X-paste: Revisiting scalable copy" + ], + "bbox": [ + 514, + 90, + 903, + 901 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "24461", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "paste for instance segmentation using clip and stablediffusion. In International Conference on Machine Learning, pages 42098-42109. PMLR, 2023. 4", + "[54] Bolei Zhou, Agata Lapedriza, Aditya Khosla, Aude Oliva, and Antonio Torralba. Places: A 10 million image database for scene recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2017. 2", + "[55] Junhao Zhuang, Yanhong Zeng, Wenran Liu, Chun Yuan, and Kai Chen. A task is worth one word: Learning with task prompts for high-quality versatile image inpainting, 2023. 1, 3, 6, 7, 8" + ], + "bbox": [ + 91, + 90, + 482, + 246 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "24462", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/SmartEraser_ Remove Anything from Images using Masked-Region Guidance/ab3bbba3-1da3-4883-858e-f047afa69964_model.json b/2025/SmartEraser_ Remove Anything from Images using Masked-Region Guidance/ab3bbba3-1da3-4883-858e-f047afa69964_model.json new file mode 100644 index 0000000000000000000000000000000000000000..f7e03c02bfbcced1333746ba2999ac9ac5a7e966 --- /dev/null +++ b/2025/SmartEraser_ Remove Anything from Images using Masked-Region Guidance/ab3bbba3-1da3-4883-858e-f047afa69964_model.json @@ -0,0 +1,2433 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.105, + 0.131, + 0.895, + 0.154 + ], + "angle": 0, + "content": "SmartEraser: Remove Anything from Images using Masked-Region Guidance" + }, + { + "type": "text", + "bbox": [ + 0.196, + 0.18, + 0.802, + 0.235 + ], + "angle": 0, + "content": "Longtao Jiang\\(^{1,*}\\) Zhendong Wang\\(^{1,*}\\) Jianmin Bao\\(^{2,*\\dagger\\heartsuit}\\) \nWengang Zhou\\(^{1,\\dagger}\\) Dongdong Chen\\(^{2}\\) Lei Shi\\(^{2}\\) Dong Chen\\(^{2}\\) Houqiang Li\\(^{1}\\) \n\\(^{1}\\)University of Science and Technology of China \\(^{2}\\)Microsoft Research Asia" + }, + { + "type": "text", + "bbox": [ + 0.26, + 0.238, + 0.732, + 0.253 + ], + "angle": 0, + "content": "https://longtaojiang.github.io/smarteraser.github.io/" + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.265, + 0.295, + 0.369 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.37, + 0.295, + 0.475 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.296, + 0.266, + 0.498, + 0.369 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.297, + 0.37, + 0.498, + 0.475 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.5, + 0.266, + 0.702, + 0.37 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.5, + 0.371, + 0.702, + 0.475 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.704, + 0.266, + 0.905, + 0.37 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.704, + 0.371, + 0.905, + 0.475 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.477, + 0.908, + 0.506 + ], + "angle": 0, + "content": "Figure 1. Object removal results generated by SmartEraser using user-provided masks. SmartEraser not only generates photorealistic images with the target objects removed but also better preserves the surrounding context around the removed objects." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.518, + 0.328, + 0.533 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.551, + 0.486, + 0.884 + ], + "angle": 0, + "content": "Object removal has so far been dominated by the \"mask-and-inpaint\" paradigm, where the masked region is excluded from the input, leaving models relying on unmasked areas to inpaint the missing region. However, this approach lacks contextual information for the masked area, often resulting in unstable performance. In this work, we introduce SmartEraser, built with a new \"removing\" paradigm called Masked-Region Guidance. This paradigm retains the masked region in the input, using it as guidance for the removal process. It offers several distinct advantages: (a) it guides the model to accurately identify the object to be removed, preventing its regeneration in the output; (b) since the user mask often extends beyond the object itself, it aids in preserving the surrounding context in the final result. Leveraging this new paradigm, we present Syn4Removal, a large-scale object removal dataset, where instance segmentation data is used to copy and paste objects onto images as removal targets, with the original images serving as ground truths. Experimental results demonstrate that SmartEraser significantly outperforms existing methods, achieving superior performance in object removal, especially in complex scenes with intricate compositions." + }, + { + "type": "title", + "bbox": [ + 0.515, + 0.518, + 0.645, + 0.533 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.543, + 0.907, + 0.68 + ], + "angle": 0, + "content": "With the rapid advancement of generative models [10, 11, 15, 32, 34, 35, 37, 40], image editing has garnered significant attention for its broad applications. Among various editing tasks, object removal [9, 55] is especially prominent, empowering users to seamlessly eliminate unwanted elements while maintaining the realism of original images. This functionality, essential for removing distracting objects in photos, has become a key feature in widely used applications like Photoshop, Google Photos, Microsoft Designer." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.684, + 0.909, + 0.881 + ], + "angle": 0, + "content": "Currently, most object removal methods [1, 9, 30, 39, 55] adopt a \"mask-and-inpaint\" paradigm, where the masked region is excluded from the input and typically filled with a neutral placeholder (e.g., black). The model then inpaints the masked region based on the surrounding content. However, we identify two primary issues with this approach. First, this strategy often generates unintended objects within the masked region. Lacking precise discrimination between removal targets and other content, these methods rely heavily on background context, sometimes leading to the unintentional inpainting of new objects—such as adding a new car on the road rather than removing the original one, as illustrated in Figure 2(a). Second, user-" + }, + { + "type": "page_footnote", + "bbox": [ + 0.537, + 0.887, + 0.888, + 0.902 + ], + "angle": 0, + "content": "* Equal contribution. † Corresponding authors. ⋆ Project leader." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "24452" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.093, + 0.089, + 0.483, + 0.244 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.207, + 0.246, + 0.44, + 0.258 + ], + "angle": 0, + "content": "(b) Our paradigm: Masked-Region Guidance" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.26, + 0.483, + 0.316 + ], + "angle": 0, + "content": "Figure 2. Comparison of the \"mask-and-inpaint\" paradigm and the proposed Masked-Region Guidance paradigm for object removal. Unlike the mask-and-inpaint approach, our paradigm uses the masked regions as critical guidance for object removal." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.321, + 0.483, + 0.382 + ], + "angle": 0, + "content": "defined masks frequently exceed the target object, requiring the \"mask-and-inpaint\" approach to synthesize these extended region. This will inadvertently modify nearby context, reducing visual coherence." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.384, + 0.483, + 0.565 + ], + "angle": 0, + "content": "To address these limitations, we introduce a novel paradigm for object removal called Masked-Region Guidance. The core idea is that the masked region should not be excluded but rather utilized as critical guidance during the removal process. Our method is straightforward: instead of replacing the masked region with a placeholder, we retain the original image as input, with the masked region indicated by a mask input, as in existing methods. As illustrated in Figure 2(b), this paradigm enables the model to accurately identify the target object, preventing unintended regeneration in the output and effectively preserving the surrounding context of the target object in the final result." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.567, + 0.483, + 0.763 + ], + "angle": 0, + "content": "However, our new paradigm cannot be directly implemented with existing object removal data construction methods. Typically, their training data are created by masking parts of the image, and the model is trained to predict the masked content. If applying this approach to our proposed Masked-Region Guidance paradigm, the model could exploit a shortcut by simply replicating the masked content from the input, due to the masked region being included in the input. Thus, a dataset consisting of triplets of an input, a mask, and the removal result is essential for our paradigm. Unfortunately, existing related datasets [23, 36, 45, 48] either contain a limited number of unique scenes or rely on inpainting models to generate pseudo-removed results." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.902 + ], + "angle": 0, + "content": "To solve this, we introduce a synthetic technique to create training data specifically for object removal. Our approach involves pasting object instances from various images onto different background images to form the input images, with the pasted instance masks serving as the input masks and the original background images designated as ground truths. Using this method, we generate Syn4Removal, a large-scale dataset comprising triplets of real background images, masks, and backgrounds with" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.151 + ], + "angle": 0, + "content": "pasted objects. The design of Syn4Removal provides diverse scenes and supports effective training under our new paradigm, encouraging the model to accurately learn object removal without shortcuts." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.152, + 0.907, + 0.303 + ], + "angle": 0, + "content": "To make Syn4Removal suitable for training object removal models, we design a pipeline to generate high-quality data. First, we filter out low-quality instances and background images. Then, we develop a method to calculate feasible pasting locations on the image, ensuring that objects do not overlap with instances in the pasted area, which helps prevent the model from regenerating unwanted objects. Last, the instance is pasted onto a background image with a blending algorithm. The resulting dataset consists of 1 million image triplets." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.304, + 0.906, + 0.499 + ], + "angle": 0, + "content": "Using the Masked-Region Guidance paradigm and the Syn4Removal dataset, we design a framework based on text-to-image diffusion models for object removal. To enhance the model's robustness to varying mask shapes from user input, we introduce a mask enhancement technique that simulates different mask shapes over removal targets. Additionally, we incorporate CLIP-based visual guidance to assist the model in better understanding the removal targets. The resulting model called SmartEraser outperforms previous methods significantly in both quantitative and qualitative evaluations. We choose the name SmartEraser because the model can \"smartly\" identify removal targets and remove it while preserving its surrounding region." + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.5, + 0.885, + 0.514 + ], + "angle": 0, + "content": "Our contributions are mainly summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.515, + 0.906, + 0.559 + ], + "angle": 0, + "content": "- We propose Masked-Region Guidance, a novel paradigm for object removal, which effectively addresses issues of object regeneration and surrounding region distortion." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.56, + 0.905, + 0.605 + ], + "angle": 0, + "content": "- We introduce Syn4Removal, a large-scale high-quality dataset for object removal, which contains over a million pairs of images across diverse scenes and object types." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.606, + 0.905, + 0.65 + ], + "angle": 0, + "content": "- We carefully design a framework based on the text-to-image diffusion model by introducing mask enhancement and clip-based visual guidance." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.651, + 0.905, + 0.695 + ], + "angle": 0, + "content": "- Extensive experiments demonstrate that SmartEraser significantly outperforms previous object removal methods in terms of quality and robustness." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.515, + 0.906, + 0.695 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.71, + 0.655, + 0.725 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Image inpainting. Image inpainting aims to seamlessly fill missing regions in images by leveraging contextual information around the masked region. Based on the \"mask-and-inpaint\" paradigm, earlier works are mainly based on generative adversarial networks (GANs) [11, 31]. Typically, these methods [2, 4, 8, 24, 25, 27, 39, 43, 49] randomly mask part areas of images [19, 20, 28, 54] and then predict these masked regions. However, GAN-based approaches [12, 41, 42, 50, 51] often suffer from limited diversity and quality in generated content, resulting in blurry or unrealistic inpainted results." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24453" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.242 + ], + "angle": 0, + "content": "Recently, approaches [1, 6, 9, 18, 30, 35, 47, 55] based on diffusion models have attracted significant attention due to their superior capability in generating image details. Among them, Repaint [30] injects the known region into intermediate images at all timesteps throughout the diffusion generation. SD-Inpaint [35] utilizes the mask and cropped image as input and fills the masked region through diffusion. While these approaches produce detailed textures, the generated content within the masked region often remains uncertain, with new objects occasionally appearing." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.243, + 0.483, + 0.497 + ], + "angle": 0, + "content": "Object removal. The object removal task aims to eliminate specified objects from images based on user-provided masks, seamlessly repairing the removed area with un-masked content. For example, PowerPaint [55] uses learnable task embeddings from context-aware inpainting as the positive prompt and those from text-guided inpainting as the negative prompt to guide object removal. A recent work CLIPAway [9] leverages AlphaCLIP [38] and IP-Adapter [46] to inject background information into the model via cross-attention, ensuring removal results coherent with the background context. The common characteristic of these methods is that they retain the traditional \"mask-and-inpaint\" paradigm by discarding the masked regions of input images. In this paper, we claim that this paradigm is not suitable for the object removal task exactly due to the missing context of the masked region, causing incomplete removal and inconsistent context in removal results." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.5, + 0.483, + 0.604 + ], + "angle": 0, + "content": "We also observe that some instruction-based methods [3, 48] attempt to remove objects using text prompts instead of the masked region. While they use the original image as input, their models are trained with removal results generated by inpainting models, which limits their performance. We believe that our synthetic strategy for building removal datasets could enhance the performance of these models." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.606, + 0.483, + 0.832 + ], + "angle": 0, + "content": "There are several works [23, 36, 45, 48] aiming to construct datasets for object removal. For instance, GQA-Inpaint [48] and DEFACTO [23] synthesize ground truths with the help of image inpainting models. However, we find that relying on image inpainting leads to the performance of object removal being highly limited by the effect of inpainting. Another two works RORD [36] and ObjectDrop [45] acquire real-world triplets through photography, but the high collection costs result in fewer than \\(3.5k\\) unique scenes. To incorporate with the Masked-Region Guidance, we construct a large-scale, high-quality dataset called Syn4Removal through a set of carefully designed strategies, which includes about 1 million triplets featuring diverse scenes and authentic ground-truth backgrounds with carefully filtering and pasting strategies." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.845, + 0.225, + 0.86 + ], + "angle": 0, + "content": "3. SmartEraser" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.901 + ], + "angle": 0, + "content": "The goal of object removal is to completely erase the target object specified by a masked region and seamlessly generate" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.212 + ], + "angle": 0, + "content": "consistent content based on the surrounding context. Most existing methods follow a \"mask-and-inpaint\" paradigm, where the masked area is excluded from the input, and the model attempts to inpaint the missing content based on the unmasked regions. While these methods can successfully remove objects in simple scenarios, they often regenerate new objects or produce blurs and artifacts in the masked region in more complex scenes." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.213, + 0.905, + 0.394 + ], + "angle": 0, + "content": "In contrast, we propose a novel paradigm: Masked-Region Guidance. Instead of discarding masked regions, we retain it in the input, leveraging it to guide the removal process. This approach is based on the insight that, by knowing the exact removal targets, the model can precisely erase them while leaving the surrounding context maintained. In the following sections, we will first introduce the method of Masked-Region Guidance in Sec. 3.1. Then, we present Syn4Removal, a large-scale synthetic dataset designed for this paradigm in Sec. 3.2. Finally, we describe SmartEraser, a fine-tuned diffusion model based on the Mask-Region Guidance paradigm and Syn4Removal dataset in Sec. 3.3." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.401, + 0.75, + 0.417 + ], + "angle": 0, + "content": "3.1. Masked-Region Guidance" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.424, + 0.905, + 0.575 + ], + "angle": 0, + "content": "Given an input image \\( \\pmb{x} \\), and user-specified mask \\( m \\) containing the removal targets, the object removal task aims to produce the removal result \\( \\hat{\\pmb{x}} \\) while maintaining the realism of the original image. The mask \\( m \\) is a binary mask where 1 indicates the region containing the removal targets. Existing models [9, 35, 39, 55] are mostly based on the \"mask-and-inpaint\" paradigm, i.e., utilizing the unmasked region \\( x \\odot (1 - m) \\) and mask \\( m \\) to predict the results without removal object in the masked region. These models are learned to maximize the following objective:" + }, + { + "type": "equation", + "bbox": [ + 0.597, + 0.576, + 0.905, + 0.593 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} ^ {\\prime} = P \\left(\\hat {\\boldsymbol {x}} \\mid \\boldsymbol {m}, \\boldsymbol {x} \\odot (\\boldsymbol {1} - \\boldsymbol {m}); \\Theta^ {\\prime}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.6, + 0.904, + 0.643 + ], + "angle": 0, + "content": "where the conditional probability \\(P\\) is modeled using a neural network with parameters \\(\\Theta^{\\prime}\\), \\(\\odot\\) specifies the elementwise multiplication." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.645, + 0.905, + 0.901 + ], + "angle": 0, + "content": "In contrast, we introduce Masked-Region Guidance, a fundamentally different approach. Instead of discarding the masked region, we retain it in the input, allowing it to guide the removal process. Specifically, we replace the condition input from \\([m, x \\odot (1 - m)]\\) to \\([m, x]\\). The mask \\(m\\) is still applied to instruct the region for removal targets, but \\(x\\) now includes the masked region \\(x \\odot m\\) besides \\(x \\odot (1 - m)\\). By including the masked region, we provide the model with valuable guidance during the object removal task. First, retaining the masked region allows the model to know exactly which objects need to be removed, reducing the risk of regenerating the object in the masked area. This helps the model focus on learning how to reconstruct the scene without the object. Second, the masked region is typically larger than the removal targets. This allows the model to copy the surrounding context of target objects in the masked region to the final result, significantly reducing the difficulty" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24454" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.093, + 0.087, + 0.905, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.268, + 0.908, + 0.296 + ], + "angle": 0, + "content": "Figure 3. The data generation pipeline of Syn4Removal. We apply instances and images to construct a triplet consisting of an input image with removal targets, a mask, and the ground truth." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.31, + 0.484, + 0.384 + ], + "angle": 0, + "content": "of synthesizing a larger area. Moreover, if the masked region contains complex textures, these can be retained, leading to more realistic results. In our Masked-Region Guidance paradigm, object removal is learned to maximize the objective:" + }, + { + "type": "equation", + "bbox": [ + 0.218, + 0.39, + 0.483, + 0.406 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = P (\\hat {\\boldsymbol {x}} | \\boldsymbol {m}, \\boldsymbol {x}; \\Theta), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.416, + 0.351, + 0.431 + ], + "angle": 0, + "content": "where the model is with parameters \\(\\Theta\\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.44, + 0.235, + 0.456 + ], + "angle": 0, + "content": "3.2. Syn4Removal" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.463, + 0.484, + 0.659 + ], + "angle": 0, + "content": "Now we have a new paradigm for training object removal models, but it cannot be directly applied using existing data construction methods. Typically, training data is created by masking parts of an image and having the model predict the masked content. However, this could lead the model to simply copy the masked content if applied to our paradigm. To overcome this, we require a dataset with triplets (input image, mask, and ground truth). However, existing real image datasets [36, 45] are limited in terms of scene variety, as they require photos taken before and after placing objects. Additionally, some synthetic datasets [23, 48] rely on inpainting models to generate pseudo-removed results, which don't fully capture the real removal process." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.66, + 0.484, + 0.856 + ], + "angle": 0, + "content": "To address this, we propose a synthetic method for constructing object removal training data. Our approach involves pasting object instances from various images onto different backgrounds to form the input images. The mask of the pasted object serves as the input mask, and the original background image serves as the ground truth. Since the ground truth is the real image, it helps the model better capture the true distribution of the image, leading to more realistic object removal results. As a result, we constructed Syn4Removal, a large-scale high-quality data set that contains 1 million samples. The data construction pipeline is shown in Figure 3. Next, we will introduce each component of the overall pipeline." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Instance Crop & Filter. We directly use a public instance segmentation dataset [22] to get instances. We use the mask to get each instance from the image. While cutting" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.31, + 0.907, + 0.475 + ], + "angle": 0, + "content": "instances, we also compute the corresponding CLIP similarity score with their classes and normalized area ratio in the original images. Then we perform the filter process, First, we exclude instances with areas larger than \\(95\\%\\) or smaller than \\(5\\%\\) to prevent potential low-quality segmentation results. Then, based on the previously calculated CLIP scores, we remove instances with low semantic relevance in each class. Given the variability in CLIP score distribution across different classes, we set individual CLIP score thresholds for each class according to previous work [53], filtering out instances with scores below these thresholds." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.476, + 0.906, + 0.581 + ], + "angle": 0, + "content": "Background Filter. For background images, we use public datasets COCONut [7] and SAM-1B [21], along with segmentation annotations. We filter out low-resolution images, those with extreme aspect ratios, or those with a large number of segmented instances. The selected backgrounds are then combined to form 1 million high-quality images (for more details, see Supplementary)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.582, + 0.906, + 0.687 + ], + "angle": 0, + "content": "Instance Size Sampling & Resize. To match the size of the instance with that of the background, we first calculate the average \\(\\mu_{c}\\) and variance \\(\\sigma_c^2\\) of the normalized area ratio for each class \\(c\\). When pasting an instance onto a background, we sample scale \\(s\\) from a Gaussian distribution \\(\\mathcal{N}(\\mu_c,\\sigma_c^2)\\), and scale the instance to \\(sHW\\), where \\(H\\) and \\(W\\) are the height and width of the background image, respectively." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.687, + 0.906, + 0.762 + ], + "angle": 0, + "content": "Feasible Area for Pasting. To ensure that the pasted instance does not cover existing instances in the background image, we calculate IoU between the pasted instance region and each instance in the background image, the feasible region based on IoU results is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.532, + 0.763, + 0.905, + 0.781 + ], + "angle": 0, + "content": "\\[\n\\mathbf {R} _ {1} = \\left\\{\\left(x, y\\right) \\mid \\operatorname {I o U} \\left(P _ {x, y}, I _ {i}\\right) < r, I _ {i} \\in [ 1, I _ {\\max } ] \\right\\}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.787, + 0.906, + 0.847 + ], + "angle": 0, + "content": "where \\(P_{x,y}\\) is the instance region if centered on \\((x,y)\\), \\(r\\) is a IoU threshold, \\(I_{max}\\) is the number of existing instances. Besides, to ensure the integrity of pasted instances, edge areas of background images are avoided, formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.528, + 0.848, + 0.905, + 0.866 + ], + "angle": 0, + "content": "\\[\n\\mathbf {R} _ {2} = \\left\\{(x, y) \\mid x \\in \\left[ e _ {x}, W - e _ {x} \\right], y \\in \\left[ e _ {y}, H - e _ {y} \\right] \\right\\}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.906, + 0.901 + ], + "angle": 0, + "content": "where \\( e_x \\) and \\( e_y \\) are half of the width and height of the bounding box of the instance mask, respectively. We take" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24455" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.089, + 0.905, + 0.257 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.109, + 0.26, + 0.885, + 0.274 + ], + "angle": 0, + "content": "Figure 4. Overall framework of SmartEraser. It fully utilizes the Masked-Region Guidance paradigm and the Syn4Removal dataset." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.288, + 0.483, + 0.318 + ], + "angle": 0, + "content": "the intersection region of \\(\\mathbf{R}_1\\) and \\(\\mathbf{R}_2\\), the final set of feasible locations is as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.231, + 0.324, + 0.483, + 0.34 + ], + "angle": 0, + "content": "\\[\n\\mathbf {R} _ {\\mathrm {f}} = \\mathbf {R} _ {1} \\cap \\mathbf {R} _ {2}. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.352, + 0.484, + 0.41 + ], + "angle": 0, + "content": "We then randomly select a location from \\(\\mathbf{R}_{\\mathrm{f}}\\) as the center point for pasting the instance. Based on the mask of the instance and location, we can get the instance mask centered on the location \\(\\bar{m}\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.412, + 0.484, + 0.487 + ], + "angle": 0, + "content": "Alpha Blending for Instance Pasting. To further ensure that the pasted instance blends harmoniously with the background, we apply alpha blending [13]. This method interpolates between the foreground instance and the background image, formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.185, + 0.494, + 0.482, + 0.51 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {x} = \\boldsymbol {\\alpha} \\odot \\boldsymbol {x} _ {i} + (1 - \\boldsymbol {\\alpha}) \\odot \\boldsymbol {x} _ {b}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.522, + 0.484, + 0.688 + ], + "angle": 0, + "content": "where \\( \\pmb{x} \\) is final constructed image, \\( \\pmb{x}_i \\) is the instance, \\( \\pmb{x}_b \\) is the background image, and \\( \\alpha \\) is alpha blending value. \\( \\alpha \\) is calculated through an alpha blending algorithm [13] that processes the mask \\( \\bar{m} \\) to a trimap and then takes the trimap as input, which consists of three regions, i.e., the foreground and background, and an excessively uncertain boundary region. So we have the triplet for object removal training: \\( \\{\\pmb{x}, \\bar{m}, \\pmb{x}_b\\} \\), where \\( \\pmb{x} \\) is input image with target object, \\( \\bar{m} \\) is original mask, and \\( \\pmb{x}_b \\) as the ground truth. Using this data generation pipeline, we finally get about 1 million triplets that form the Syn4Removal dataset." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.697, + 0.219, + 0.71 + ], + "angle": 0, + "content": "3.3. Framework" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.484, + 0.825 + ], + "angle": 0, + "content": "Based on the proposed Masked-Region Guidance paradigm and the Syn4Removal dataset, we design a framework based on the text-to-image stable diffusion model for object removal. In this framework, we carefully design a user-friendly mask enhancement strategy and introduce CLIP-based visual guidance to further improve the capability of the model for object removal, as shown in Figure 4." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Mask Enhancement. In real-world scenarios, users usually provide a loose or tight mask around the object to be removed. If the model were trained only with precise object masks, there would be a significant gap in mask shape and size between training and inference. To address this," + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.281, + 0.906, + 0.49 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.514, + 0.492, + 0.904, + 0.505 + ], + "angle": 0, + "content": "Figure 5. Mask shapes from different mask enhancement methods." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.514, + 0.907, + 0.801 + ], + "angle": 0, + "content": "during training, we apply various mask deformation methods to simulate the user input mask shapes, as shown in Figure 5. These techniques help the model generalize to different mask forms. Specifically, we use six mask types to augment the object mask: (1) Original mask: this mask precisely outlines the target object. (2) Eroded mask: created by applying morphological erosion to the original mask, simulating scenarios where the user's mask may not fully cover the object. (3) Dilated mask: created by applying morphological dilation to the original mask. (4) Convex hull mask: constructed by calculating the convex hull that fully encompasses the original mask, then expanding it slightly. (5) Ellipse mask: generated by finding the smallest enclosing ellipse around the original mask, followed by a slight expansion. (6) Bbox & Bessel mask: a bounding box is generated to completely cover the original mask first, then bessel curves are constructed along each edge of the box to introduce an irregular boundary. So the overall mask enhancement process is illustrated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.601, + 0.81, + 0.905, + 0.827 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {m} = \\operatorname {M E} _ {i} (\\bar {\\boldsymbol {m}}), \\quad i \\in \\{1, \\dots , 6 \\}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.905, + 0.869 + ], + "angle": 0, + "content": "where \\( m \\) is the enhanced mask used during training, \\( \\mathrm{ME}_i(\\cdot) \\) represents six different mask enhancement methods." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.907, + 0.901 + ], + "angle": 0, + "content": "CLIP-Based Visual Guidance. Our SmartEraser is designed based on text-to-image diffusion model, the text" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24456" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.089, + 0.907, + 0.332 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.334, + 0.907, + 0.361 + ], + "angle": 0, + "content": "Figure 6. Qualitative comparison of SmartEraser and state-of-the-art methods. The samples are sourced from RORD-Val, which includes the ground truth after removal." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.375, + 0.486, + 0.617 + ], + "angle": 0, + "content": "prompts play an important role in guiding the generation process. We aim to design a prompt that indicates the model what to remove during the generation process. Concretely, we first segment the region corresponding to the removal target in the image using its corresponding enhanced mask. This segmented region is then fed into the pre-trained CLIP [33] vision encoder to extract its visual feature(e.g., class token). This feature is not precisely aligned with the text encoder output space. So we then apply a trainable MLP network to map the visual feature into the feature space of the text encoder. The mapped embedding is then appended to the tokens of the CLIP text embedding of the prompt \"Remove the instance of\". During the training process, the vision encoder is frozen, and the MLP and text encoder are trainable. So the CLIP-based visual guidance is formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.177, + 0.62, + 0.482, + 0.639 + ], + "angle": 0, + "content": "\\[\n\\widehat {\\boldsymbol {c}} = \\left[ \\boldsymbol {\\tau} _ {\\theta} (\\boldsymbol {y}), \\operatorname {M L P} \\left(\\boldsymbol {\\nu} _ {\\theta} (\\boldsymbol {x} \\odot \\boldsymbol {m})\\right) \\right], \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.645, + 0.483, + 0.676 + ], + "angle": 0, + "content": "where \\(\\tau_{\\theta}(\\cdot)\\) is text encoder, \\(\\nu_{\\theta}(\\cdot)\\) is vision encoder, and \\(\\pmb{y}\\) represents input text prompt \"remove the instance of\"." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.676, + 0.483, + 0.765 + ], + "angle": 0, + "content": "Loss Function. Suppose the \\( E(\\cdot) \\) is the VAE encoder, so the latent feature of input image \\( x \\) and its ground truth \\( x_{b} \\) is \\( \\bar{z} = E(x) \\) and \\( z = E(x_{b}) \\), respectively, and \\( z_{t} \\) is the noisy feature of \\( z \\) with added noise \\( \\epsilon \\) at timestep \\( t \\). So the overall training loss following the standard diffusion process, which is formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.146, + 0.769, + 0.483, + 0.787 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathbb {E} _ {\\boldsymbol {x}, \\boldsymbol {m}, \\boldsymbol {x} _ {b}, t, \\epsilon} \\| \\boldsymbol {\\epsilon} - \\boldsymbol {\\epsilon} _ {\\theta} \\left(\\boldsymbol {z} _ {t}, \\bar {\\boldsymbol {z}}, \\boldsymbol {m}, \\widehat {\\boldsymbol {c}}, t\\right) \\| _ {2} ^ {2}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.795, + 0.325, + 0.811 + ], + "angle": 0, + "content": "where \\(\\epsilon_{\\theta}(\\cdot)\\) is the diffusion model." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.823, + 0.224, + 0.839 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.848, + 0.19, + 0.864 + ], + "angle": 0, + "content": "4.1. Settings" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Baselines. We compare SmartEraser with methods in both image inpainting and object removal. For im-" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.375, + 0.907, + 0.481 + ], + "angle": 0, + "content": "age inpainting, we evaluate against both GAN-based approaches (ZITS++ [4], MAT [24], LaMa [39], BLD [1]) and diffusion-based methods (RePaint [30], and SD-Inpaint [35]). Additionally, we compare it with specialized object removal frameworks, including CLIPAway [9] and PowerPaint [55], which represent the previous state-of-the-art performance on object removal benchmarks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.485, + 0.909, + 0.636 + ], + "angle": 0, + "content": "Implementation details. To enable a fair comparison, we follow previous experimental settings [9, 55] and finetune SmartEraser on the widely-used SD v1.5 model [35]. The training process involves fine-tuning both the UNet and text encoder components within the diffusion model and MLP for mapping visual features to text features space. SmartEraser is trained on the proposed Syn4Removal dataset with a batch size of 32, using the AdamW [29] optimizer for \\(500k\\) iterations and a learning rate of 2e-5 across all trainable modules." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.64, + 0.907, + 0.7 + ], + "angle": 0, + "content": "Evaluation benchmarks. We evaluate the object removal capability of all baseline methods and our approach across three benchmarks in diverse scenarios. The benchmarks are constructed as follows:" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.705, + 0.906, + 0.78 + ], + "angle": 0, + "content": "- RORD-Val: Due to there are many triplets share the same background images as the ground truth in validation set of the RORD [36]. To ensure unique scenes and high object quality, we filter this subset to obtain RORD-Val, a evaluation benchmark of real-photography image pairs." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.781, + 0.905, + 0.841 + ], + "angle": 0, + "content": "- DEFACTO-Val: Derived from the splicing section of DEFACTO [23], we exclude examples where intended targets for removal significantly overlap with instances in the background images." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.841, + 0.905, + 0.9 + ], + "angle": 0, + "content": "- Syn4Removal-Val: Generated using the pipeline of Syn4Removal with MSCOCO [26] validation set as background images and OpenImages-v7 [22] validation set offering instances." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.705, + 0.906, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24457" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.093, + 0.089, + 0.907, + 0.354 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.355, + 0.905, + 0.384 + ], + "angle": 0, + "content": "Figure 7. Qualitative comparison of SmartEraser and other methods in real-world user cases, in which users provide various mask shapes. The samples are sourced from real-world images and the validation set of MSCOCO [26]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.391, + 0.483, + 0.451 + ], + "angle": 0, + "content": "Each benchmark comprises approximately 5k samples. During inference, following [9], we apply slight mask dilation to mitigate downscaling artifacts in SD-Inpaint-based methods [9, 35, 55]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.453, + 0.484, + 0.591 + ], + "angle": 0, + "content": "Evaluation metrics. To quantitatively assess the performance of the object removal models, we consider three key aspects: (1) overall image quality, which is evaluated by Fréchet Inception Distance (FID) [14] and CLIP Maximum Mean Discrepancy (CMMD) [17]. (2) consistency between the predicted region and the background context, evaluated by REMOVE [5] metric. (3) consistency between the predicted region and corresponding region in the ground truth, assessed by LPIPS [52], SSIM [44], and PSNR [16]." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.603, + 0.413, + 0.619 + ], + "angle": 0, + "content": "4.2. Comparison with Previous Methods" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.628, + 0.483, + 0.809 + ], + "angle": 0, + "content": "Quantitative results. We conduct extensive experiments to evaluate the performance of SmartEraser against baselines on the benchmarks. The results are shown in Tables 1, 2 and 3. SmartEraser significantly outperforms all other methods across all metrics on the three benchmarks. Notably, on the more complex RORD-Val dataset, which has larger masks and more challenging removal scenes, SmartEraser surpasses the previous state-of-the-art by margins of 10.3 in FID and 1.89dB in PSNR. This indicates the superior performance of SmartEraser for object removal. Besides, SmartEraser also achieves state-of-the-art performance on the simpler DEFACTO-Val and Syn4Removal-Val datasets." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Qualitative results. We present visualization comparison results of object removal using images from real scenes, shown in Figure 6. Baseline methods have limitations and may introduce unintended objects, create unrealistic structures, or produce artifacts within masked regions. In contrast, SmartEraser succeeds in effectively removing target" + }, + { + "type": "table", + "bbox": [ + 0.52, + 0.388, + 0.905, + 0.526 + ], + "angle": 0, + "content": "
MethodFID ↓ CMMD ↓ReMOVE ↑LPIPS ↓SSIM ↑PSNR ↑
ZITS++ [4]53.4400.8380.8200.3880.52215.442
MAT [24]80.9531.4940.6280.4330.49513.322
LaMa [39]24.2370.2160.9160.3480.55716.383
BLD [1]73.0861.7600.8340.4710.46715.775
RePaint [30]71.5421.6390.8120.3810.51015.861
SD-Inpaint [35]69.5020.3240.8570.3690.53716.111
CLIPAway [9]25.4580.1230.9150.3330.57717.434
PowerPaint [55]24.0580.2940.9260.3080.60218.100
SmartEraser16.0300.0920.9370.2760.61219.994
" + }, + { + "type": "table_caption", + "bbox": [ + 0.514, + 0.529, + 0.905, + 0.554 + ], + "angle": 0, + "content": "Table 1. Quantitative comparison of SmartEraser and previous methods on RORD-Val." + }, + { + "type": "table", + "bbox": [ + 0.52, + 0.556, + 0.905, + 0.693 + ], + "angle": 0, + "content": "
MethodFID ↓CMMD ↓ReLUVE ↑LPIPS ↓SSIM ↑PSNR ↑
ZITS++ [4]8.1530.2290.8990.3500.63419.453
MAT [24]15.2570.3210.7420.4070.57514.541
LaMa [39]6.6010.2010.9260.3510.68321.735
BLD [1]24.4070.7460.9130.4180.63219.346
RePaint [30]12.3140.8220.9240.3250.64420.607
SD-Inpaint [35]11.1190.1960.9120.3200.65621.691
CLIPAway [9]8.3200.1930.9250.3230.66621.953
PowerPaint [55]6.5820.2750.9340.2780.71523.870
SmartEraser3.4050.1060.9390.2570.73425.363
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.695, + 0.905, + 0.723 + ], + "angle": 0, + "content": "Table 2. Quantitative comparison of SmartEraser and previous methods on DEFACTO-Val." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.733, + 0.905, + 0.778 + ], + "angle": 0, + "content": "objects while generating high-quality synthesized results, and excels in maintaining the contextual integrity and visual coherence of the scene." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.907, + 0.901 + ], + "angle": 0, + "content": "User study. To assess the visual quality subjectively, we conduct a user study with 30 participants. The study includes 20 image sets, each containing an original image, its mask, and removal results from four different methods, presented in random order. Participants are asked to select their preferred result based on three criteria: overall image quality, foreground-background consistency, and accuracy of object removal. Voting percentages in Table 4 show that" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24458" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.098, + 0.089, + 0.48, + 0.226 + ], + "angle": 0, + "content": "
MethodFID ↓ CMMD ↓ReMOVE ↑LPIPS ↓SSIM ↑PSNR ↑
ZITS++ [4]5.9320.1930.8100.3580.58117.315
MAT [24]19.2630.3000.7000.3740.57013.357
LaMa [39]5.2460.0970.9320.3250.60619.319
BLD [1]20.2860.7050.8960.4640.52516.539
RePaint [30]12.6770.6960.9080.3780.55317.689
SD-Inpaint [35]11.5280.1070.8970.3480.57218.105
CLIPAway [9]5.5300.0930.9300.3440.59318.561
PowerPaint [55]5.0250.2000.9310.2910.66320.594
SmartEraser4.3860.0530.9390.2690.67222.029
" + }, + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.229, + 0.482, + 0.259 + ], + "angle": 0, + "content": "Table 3. Quantitative comparison of SmartEraser and previous methods on Syn4Removal-Val." + }, + { + "type": "table", + "bbox": [ + 0.097, + 0.264, + 0.48, + 0.339 + ], + "angle": 0, + "content": "
MethodImage Quality ↑ FB-Consistency ↑Removal Accuracy ↑
LaMa [39]0.100.11
CLIPAway [9]0.180.12
PowerPaint [55]0.160.12
SmartEraser0.560.65
0.58
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.09, + 0.34, + 0.483, + 0.371 + ], + "angle": 0, + "content": "Table 4. Comparison of user study results in terms of image quality, foreground-background consistency, and removal accuracy." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.379, + 0.444, + 0.394 + ], + "angle": 0, + "content": "users consistently favor the results from SmartEraser." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.395, + 0.484, + 0.759 + ], + "angle": 0, + "content": "Real-world user cases. To further evaluate the object removal capability of SmartEraser in real-world user scenarios, we conduct a comparison with other methods (LaMa [39], CLIPAway [9], PowerPaint [55]), using masks provided by users. Figure 7 shows the comparison results. We can observe that SmartEraser exhibits a remarkable ability to accurately identify and remove target objects while preserving surrounding contextual details, particularly in scenarios involving imperfect or approximate masks. In the sample in the bottom row on the left, SmartEraser accurately identifies and removes the person indicated by the mask, while maintaining the integrity of the elephant in the background. In contrast, alternative methods tend to repaint the entire masked area, which leads to outputs that lack contextual coherence and visual realism due to the loss of details of the elephant. The upper second sample on the right provides a challenging scenario, the user-provided mask aims to remove the man whose feet are partially obscured by floating text. The previous methods fail to maintain the integrity of the floating text. However, SmartEraser accurately removes the man while preserving the floating text. This highlights SmartEraser's ability to effectively differentiate between removal targets and background content, even in cases of partial occlusion." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.771, + 0.253, + 0.787 + ], + "angle": 0, + "content": "4.3. Ablation studies" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.484, + 0.901 + ], + "angle": 0, + "content": "To evaluate the effectiveness of the key components in our framework, we perform ablation studies and report both quantitative and qualitative results. We progressively modify the baseline (fine-tuning SD v1.5 with the \"mask-and-inpaint\" paradigm) by adding our proposed techniques: RG for masked-region guidance, ME for mask enhancement, and VG for CLIP-based visual guidance. The results are" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.089, + 0.596, + 0.144 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.144, + 0.595, + 0.197 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.518, + 0.197, + 0.594, + 0.207 + ], + "angle": 0, + "content": "Mask & Image" + }, + { + "type": "image", + "bbox": [ + 0.596, + 0.089, + 0.671, + 0.197 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.612, + 0.197, + 0.655, + 0.207 + ], + "angle": 0, + "content": "Baseline" + }, + { + "type": "image", + "bbox": [ + 0.671, + 0.089, + 0.75, + 0.197 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.699, + 0.197, + 0.724, + 0.207 + ], + "angle": 0, + "content": "+RG" + }, + { + "type": "image", + "bbox": [ + 0.751, + 0.089, + 0.827, + 0.197 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.777, + 0.197, + 0.802, + 0.207 + ], + "angle": 0, + "content": "+ME" + }, + { + "type": "image", + "bbox": [ + 0.828, + 0.089, + 0.905, + 0.197 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.855, + 0.197, + 0.88, + 0.207 + ], + "angle": 0, + "content": "+VG" + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.215, + 0.905, + 0.243 + ], + "angle": 0, + "content": "Figure 8. Qualitative ablation comparison of our method. From left to right, we progressively add each proposed component." + }, + { + "type": "table", + "bbox": [ + 0.521, + 0.252, + 0.903, + 0.332 + ], + "angle": 0, + "content": "
MethodFID ↓CMMD ↓ReLUVE ↑LPIPS ↓SSIM ↑PSNR ↑
Baseline33.1310.2450.9080.3550.54416.535
+RG24.5260.1210.9230.3390.56217.176
+ME19.4210.1080.9310.3010.59118.917
+VG16.0300.0920.9370.2760.61619.994
" + }, + { + "type": "table_caption", + "bbox": [ + 0.541, + 0.335, + 0.878, + 0.348 + ], + "angle": 0, + "content": "Table 5. Quantitative ablated comparison on RORD-Val." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.357, + 0.741, + 0.372 + ], + "angle": 0, + "content": "displayed in Figure 8 and Table 5." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.373, + 0.906, + 0.525 + ], + "angle": 0, + "content": "Effect of the Masked-Region Guidance paradigm: We observe that the baseline model often regenerates objects instead of properly removing them, as it is trained in a traditional \"inpainting\" manner. Incorporating masked-region guidance significantly improves the removal performance by reducing the risk of object regeneration. However, due to the mismatch between the mask shape used during training and the user-defined mask during testing, the model does not always perfectly remove the target objects and preserve the surrounding context." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.525, + 0.906, + 0.601 + ], + "angle": 0, + "content": "Contribution of Mask Enhancement: By simulating real-world user mask shapes over object removal during training, we find that it helps reduce the gap between model training and testing using user input mask, significantly keeping the surrounding regions and further improving removal results." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.601, + 0.905, + 0.678 + ], + "angle": 0, + "content": "Value of CLIP-based Visual Guidance: CLIP visual guidance aids the model in better understanding the removal targets, yielding the best performance on RORD-val across all metrics. This indicates that the visual guidance helps the model perform object removal tasks more effectively." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.694, + 0.634, + 0.71 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In conclusion, we present SmartEraser, a state-of-the-art object removal model built on the novel Masked-Region Guidance paradigm. SmartEraser outperforms existing methods by smartly identifying the target object to remove while effectively preserving the surrounding context. To facilitate research on this paradigm, we propose Syn4Removal, a large-scale, high-quality dataset containing over a million image triplets, specifically designed for object removal tasks. Through extensive experiments, we demonstrate that SmartEraser achieves superior performance in both quality and robustness compared to previous object removal methods." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "24459" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.156 + ], + "angle": 0, + "content": "[1] Omri Avrahami, Ohad Fried, and Dani Lischinski. Blended latent diffusion. ACM transactions on graphics (TOG), 42 (4):1-11, 2023. 1, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.159, + 0.484, + 0.213 + ], + "angle": 0, + "content": "[2] Connelly Barnes, Eli Shechtman, Adam Finkelstein, and Dan B Goldman. Patchmatch: A randomized correspondence algorithm for structural image editing. ACM Trans. Graph., 28(3):24, 2009. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.217, + 0.482, + 0.257 + ], + "angle": 0, + "content": "[3] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. arXiv preprint arXiv:2211.09800, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.26, + 0.483, + 0.315 + ], + "angle": 0, + "content": "[4] Chenjie Cao, Qiaole Dong, and Yanwei Fu. Zits++: Image inpainting by improving the incremental transformer on structural priors. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.317, + 0.484, + 0.386 + ], + "angle": 0, + "content": "[5] Aditya Chandrasekar, Goirik Chakrabarty, Jai Bardhan, Ramya Hebbalaguppe, and Prathosh AP. Remove: A reference-free metric for object erasure. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 7901-7910, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.388, + 0.482, + 0.442 + ], + "angle": 0, + "content": "[6] Pau de Jorge, Riccardo Volpi, Puneet K. Dakania, Philip H. S. Torr, and Rogez Gregory. Placing objects in context via inpainting for out-of-distribution segmentation. In The European Conference on Computer Vision (ECCV), 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.445, + 0.482, + 0.513 + ], + "angle": 0, + "content": "[7] Xueqing Deng, Qihang Yu, Peng Wang, Xiaohui Shen, and Liang-Chieh Chen. Coconut: Modernizing coco segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21863-21873, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.516, + 0.482, + 0.585 + ], + "angle": 0, + "content": "[8] Qiaole Dong, Chenjie Cao, and Yanwei Fu. Incremental transformer structure enhanced image inpainting with masking positional encoding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11358-11368, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.587, + 0.482, + 0.641 + ], + "angle": 0, + "content": "[9] Yigit Ekin, Ahmet Burak Yildirim, Erdem Eren Caglar, Aykut Erdem, Erkut Erdem, and Aysegul Dundar. Clipaway: Harmonizing focused embeddings for removing objects via diffusion models, 2024. 1, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.645, + 0.482, + 0.699 + ], + "angle": 0, + "content": "[10] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12873-12883, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.702, + 0.482, + 0.757 + ], + "angle": 0, + "content": "[11] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.759, + 0.482, + 0.814 + ], + "angle": 0, + "content": "[12] Xiefan Guo, Hongyu Yang, and Di Huang. Image inpainting via conditional texture and structure dual generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 14134-14143, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.817, + 0.482, + 0.857 + ], + "angle": 0, + "content": "[13] Kaiming He, Christoph Rhemann, Carsten Rother, Xiaou Tang, and Jian Sun. A global sampling method for alpha matting. In CVPR 2011, pages 2049-2056, 2011. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.86, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[14] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilib-" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.094, + 0.905, + 0.12 + ], + "angle": 0, + "content": "rium. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.123, + 0.905, + 0.164 + ], + "angle": 0, + "content": "[15] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.166, + 0.905, + 0.207 + ], + "angle": 0, + "content": "[16] Alain Hore and Djemel Ziou. Image quality metrics: Psnr vs. sSIM. In 2010 20th international conference on pattern recognition, pages 2366-2369. IEEE, 2010. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.209, + 0.905, + 0.29 + ], + "angle": 0, + "content": "[17] Sadeep Jayasumana, Srikumar Ramalingam, Andreas Veit, Daniel Glasner, Ayan Chakrabarti, and Sanjiv Kumar. Rethinking fid: Towards a better evaluation metric for image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9307-9315, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.293, + 0.905, + 0.333 + ], + "angle": 0, + "content": "[18] Xuan Ju, Xian Liu, Xintao Wang, Yuxuan Bian, Ying Shan, and Qiang Xu. Brushnet: A plug-and-play image inpainting model with decomposed dual-branch diffusion, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.336, + 0.905, + 0.375 + ], + "angle": 0, + "content": "[19] Tero Karras. Progressive growing of gans for improved quality, stability, and variation. arXiv preprint arXiv:1710.10196, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.379, + 0.905, + 0.433 + ], + "angle": 0, + "content": "[20] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.436, + 0.905, + 0.505 + ], + "angle": 0, + "content": "[21] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.507, + 0.905, + 0.601 + ], + "angle": 0, + "content": "[22] Alina Kuznetsova, Hassan Rom, Neil Alldrin, Jasper Uijlings, Ivan Krasin, Jordi Pont-Tuset, Shahab Kamali, Stefan Popov, Matteo Malloci, Alexander Kolesnikov, et al. The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale. International journal of computer vision, 128(7):1956-1981, 2020. 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.605, + 0.905, + 0.659 + ], + "angle": 0, + "content": "[23] Jens Lehmann, Daniel Gerber, Mohamed Morsey, and Axel-Cyrille Ngonga Ngomo. Defacto-deep fact validation. In International semantic web conference, pages 312-327. Springer, 2012. 2, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.661, + 0.905, + 0.73 + ], + "angle": 0, + "content": "[24] Wenbo Li, Zhe Lin, Kun Zhou, Lu Qi, Yi Wang, and Jiaya Jia. Mat: Mask-aware transformer for large hole image inpainting. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10758-10768, 2022. 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.733, + 0.905, + 0.773 + ], + "angle": 0, + "content": "[25] Xiaoguang Li, Qing Guo, Di Lin, Ping Li, Wei Feng, and Song Wnag. Misf: Multi-level interactive siamese filtering for high-fidelity image inpainting. CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.776, + 0.905, + 0.857 + ], + "angle": 0, + "content": "[26] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.86, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[27] Qiankun Liu, Zhentao Tan, Dongdong Chen, Qi Chu, Xiyang Dai, Yinpeng Chen, Mengchen Liu, Lu Yuan, and Nenghai Yu. Reduce information loss in transformers for pluralistic" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.094, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24460" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.134 + ], + "angle": 0, + "content": "image inpainting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11347-11357, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.135, + 0.484, + 0.188 + ], + "angle": 0, + "content": "[28] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaou Tang. Deep learning face attributes in the wild. In Proceedings of International Conference on Computer Vision (ICCV), 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.19, + 0.482, + 0.218 + ], + "angle": 0, + "content": "[29] I Loshchilov. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.219, + 0.482, + 0.288 + ], + "angle": 0, + "content": "[30] Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11461-11471, 2022. 1, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.289, + 0.482, + 0.315 + ], + "angle": 0, + "content": "[31] Mehdi Mirza. Conditional generative adversarial nets. arXiv preprint arXiv:1411.1784, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.316, + 0.482, + 0.384 + ], + "angle": 0, + "content": "[32] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.385, + 0.482, + 0.468 + ], + "angle": 0, + "content": "[33] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.469, + 0.482, + 0.524 + ], + "angle": 0, + "content": "[34] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.525, + 0.483, + 0.607 + ], + "angle": 0, + "content": "[35] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684-10695, 2022. 1, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.608, + 0.482, + 0.65 + ], + "angle": 0, + "content": "[36] Min-Cheol Sagong, Yoon-Jae Yeo, Seung-Won Jung, and Sung-Jea Ko. Rord: A real-world object removal dataset. In BMVC, page 542, 2022. 2, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.65, + 0.482, + 0.692 + ], + "angle": 0, + "content": "[37] Yang Song, Prafulla Dhariwal, Mark Chen, and Ilya Sutskever. Consistency models. arXiv preprint arXiv:2303.01469, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.692, + 0.482, + 0.762 + ], + "angle": 0, + "content": "[38] Zeyi Sun, Ye Fang, Tong Wu, Pan Zhang, Yuhang Zang, Shu Kong, Yuanjun Xiong, Dahua Lin, and Jiaqi Wang. Alphaclip: A clip model focusing on wherever you want. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13019-13029, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.762, + 0.482, + 0.858 + ], + "angle": 0, + "content": "[39] Roman Suvorov, Elizaveta Logacheva, Anton Mashikhin, Anastasia Remizova, Armenii Ashukha, Aleksei Silvestrov, Naejin Kong, Harshith Goka, Kiwoong Park, and Victor Lempitsky. Resolution-robust large mask inpainting with fourier convolutions. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 2149–2159, 2022. 1, 2, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.859, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[40] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017. 1" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.092, + 0.905, + 0.16 + ], + "angle": 0, + "content": "[41] Ziyu Wan, Jingbo Zhang, Dongdong Chen, and Jing Liao. High-fidelity pluralistic image completion with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 4692-4701, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.161, + 0.905, + 0.233 + ], + "angle": 0, + "content": "[42] Wentao Wang, Jianfu Zhang, Li Niu, Haoyu Ling, Xue Yang, and Liqing Zhang. Parallel multi-resolution fusion network for image inpainting. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 14559-14568, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.234, + 0.905, + 0.303 + ], + "angle": 0, + "content": "[43] Wentao Wang, Li Niu, Jianfu Zhang, Xue Yang, and Liqing Zhang. Dual-path image inpainting with auxiliary gan inversion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11421-11430, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.304, + 0.905, + 0.359 + ], + "angle": 0, + "content": "[44] Zhou Wang, A.C. Bovik, H.R. Sheikh, and E.P. Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing, 13(4): 600-612, 2004. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.361, + 0.905, + 0.418 + ], + "angle": 0, + "content": "[45] Daniel Winter, Matan Cohen, Shlomi Fruchter, Yael Pritch, Alex Rav-Acha, and Yedid Hoshen. Objectdrop: Bootstrapping counterfactuals for photorealistic object removal and insertion. arXiv preprint arXiv:2403.18818, 2024. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.419, + 0.905, + 0.473 + ], + "angle": 0, + "content": "[46] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.474, + 0.905, + 0.518 + ], + "angle": 0, + "content": "[47] Yifu Chen, Jingwen Chen, Yingwei Pan, Yehao Li, Ting Yao, Zhineng Chen and Tao Mei. Improving text-guided object inpainting with semantic pre-inpainting. ECCV, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.519, + 0.905, + 0.574 + ], + "angle": 0, + "content": "[48] Ahmet Burak Yildirim, Vedat Baday, Erkut Erdem, Aykut Erdem, and Aysegul Dundar. Inst-inpaint: Instructing to remove objects with diffusion models. arXiv preprint arXiv:2304.03246, 2023. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.576, + 0.905, + 0.644 + ], + "angle": 0, + "content": "[49] Jiahui Yu, Zhe Lin, Jimei Yang, Xiaohui Shen, Xin Lu, and Thomas S Huang. Generative image inpainting with contextual attention. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5505-5514, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.645, + 0.905, + 0.717 + ], + "angle": 0, + "content": "[50] Yingchen Yu, Fangneng Zhan, Shijian Lu, Jianxiong Pan, Feiying Ma, Xuansong Xie, and Chunyan Miao. Wavefill: A wavelet-based generation network for image inpainting. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 14114-14123, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.718, + 0.905, + 0.786 + ], + "angle": 0, + "content": "[51] Yu Zeng, Zhe Lin, Huchuan Lu, and Vishal M. Patel. Crfill: Generative image inpainting with auxiliary contextual reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 14164-14173, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.788, + 0.905, + 0.858 + ], + "angle": 0, + "content": "[52] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.859, + 0.905, + 0.902 + ], + "angle": 0, + "content": "[53] Hanqing Zhao, Dianmo Sheng, Jianmin Bao, Dongdong Chen, Dong Chen, Fang Wen, Lu Yuan, Ce Liu, Wenbo Zhou, Qi Chu, et al. X-paste: Revisiting scalable copy" + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.092, + 0.905, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "24461" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.124, + 0.092, + 0.482, + 0.134 + ], + "angle": 0, + "content": "paste for instance segmentation using clip and stablediffusion. In International Conference on Machine Learning, pages 42098-42109. PMLR, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.483, + 0.19 + ], + "angle": 0, + "content": "[54] Bolei Zhou, Agata Lapedriza, Aditya Khosla, Aude Oliva, and Antonio Torralba. Places: A 10 million image database for scene recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.192, + 0.482, + 0.247 + ], + "angle": 0, + "content": "[55] Junhao Zhuang, Yanhong Zeng, Wenran Liu, Chun Yuan, and Kai Chen. A task is worth one word: Learning with task prompts for high-quality versatile image inpainting, 2023. 1, 3, 6, 7, 8" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.247 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.956 + ], + "angle": 0, + "content": "24462" + } + ] +] \ No newline at end of file diff --git a/2025/SmartEraser_ Remove Anything from Images using Masked-Region Guidance/ab3bbba3-1da3-4883-858e-f047afa69964_origin.pdf b/2025/SmartEraser_ Remove Anything from Images using Masked-Region Guidance/ab3bbba3-1da3-4883-858e-f047afa69964_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..0ceea2481764092d41384b60883a7482ac8b14e2 --- /dev/null +++ b/2025/SmartEraser_ Remove Anything from Images using Masked-Region Guidance/ab3bbba3-1da3-4883-858e-f047afa69964_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d83f7f5cd0225f8f3c6b7818439db3ef158267973e8e4e3dc1ca16c42d4d1d13 +size 4790186 diff --git a/2025/SmartEraser_ Remove Anything from Images using Masked-Region Guidance/full.md b/2025/SmartEraser_ Remove Anything from Images using Masked-Region Guidance/full.md new file mode 100644 index 0000000000000000000000000000000000000000..88f4280987e410fd997942d37eb5cd343095c2a5 --- /dev/null +++ b/2025/SmartEraser_ Remove Anything from Images using Masked-Region Guidance/full.md @@ -0,0 +1,342 @@ +# SmartEraser: Remove Anything from Images using Masked-Region Guidance + +Longtao Jiang $^{1,*}$ Zhendong Wang $^{1,*}$ Jianmin Bao $^{2,*\dagger\heartsuit}$ +Wengang Zhou $^{1,\dagger}$ Dongdong Chen $^{2}$ Lei Shi $^{2}$ Dong Chen $^{2}$ Houqiang Li $^{1}$ $^{1}$ University of Science and Technology of China $^{2}$ Microsoft Research Asia + +https://longtaojiang.github.io/smarteraser.github.io/ + +![](images/c4720aa160d5b7008e62a16834eae9fde6110efeb03b337fcba703c311e9a059.jpg) + +![](images/d5b7d52efe55e5d764fde0d5d783b2af33065824ee9f0ad13a68404c32fb5049.jpg) +Figure 1. Object removal results generated by SmartEraser using user-provided masks. SmartEraser not only generates photorealistic images with the target objects removed but also better preserves the surrounding context around the removed objects. + +![](images/1524c6c7a097df5eaa242b1884fb1b9c44b334932b615f3c476d24b5a24023fa.jpg) + +![](images/950fbb09f39b5ba636edd120fc083b91495dceaec2a736df2e137f3ecfd865a0.jpg) + +![](images/0fead885f7fc6f87d75b68fe360d3bd9468fe6c2c4c294e9e194c976221d5e3d.jpg) + +![](images/1200313b16ce45eb43d12a2561e36cc1b22309a7fa064d3eae940c1f154926ba.jpg) + +![](images/7c5233145bfdde27eb941d041254a901206ab3a69142dbad05c724bc4d845d44.jpg) + +![](images/40d980fbb1cea1b99064ca7618696cf7cba2773baf30a3d4030bc508f548800f.jpg) + +# Abstract + +Object removal has so far been dominated by the "mask-and-inpaint" paradigm, where the masked region is excluded from the input, leaving models relying on unmasked areas to inpaint the missing region. However, this approach lacks contextual information for the masked area, often resulting in unstable performance. In this work, we introduce SmartEraser, built with a new "removing" paradigm called Masked-Region Guidance. This paradigm retains the masked region in the input, using it as guidance for the removal process. It offers several distinct advantages: (a) it guides the model to accurately identify the object to be removed, preventing its regeneration in the output; (b) since the user mask often extends beyond the object itself, it aids in preserving the surrounding context in the final result. Leveraging this new paradigm, we present Syn4Removal, a large-scale object removal dataset, where instance segmentation data is used to copy and paste objects onto images as removal targets, with the original images serving as ground truths. Experimental results demonstrate that SmartEraser significantly outperforms existing methods, achieving superior performance in object removal, especially in complex scenes with intricate compositions. + +# 1. Introduction + +With the rapid advancement of generative models [10, 11, 15, 32, 34, 35, 37, 40], image editing has garnered significant attention for its broad applications. Among various editing tasks, object removal [9, 55] is especially prominent, empowering users to seamlessly eliminate unwanted elements while maintaining the realism of original images. This functionality, essential for removing distracting objects in photos, has become a key feature in widely used applications like Photoshop, Google Photos, Microsoft Designer. + +Currently, most object removal methods [1, 9, 30, 39, 55] adopt a "mask-and-inpaint" paradigm, where the masked region is excluded from the input and typically filled with a neutral placeholder (e.g., black). The model then inpaints the masked region based on the surrounding content. However, we identify two primary issues with this approach. First, this strategy often generates unintended objects within the masked region. Lacking precise discrimination between removal targets and other content, these methods rely heavily on background context, sometimes leading to the unintentional inpainting of new objects—such as adding a new car on the road rather than removing the original one, as illustrated in Figure 2(a). Second, user- + +![](images/2724c1b7c9d985aa928a4c861c6acf4c53978b066861cde15142b229639f606b.jpg) +(b) Our paradigm: Masked-Region Guidance +Figure 2. Comparison of the "mask-and-inpaint" paradigm and the proposed Masked-Region Guidance paradigm for object removal. Unlike the mask-and-inpaint approach, our paradigm uses the masked regions as critical guidance for object removal. + +defined masks frequently exceed the target object, requiring the "mask-and-inpaint" approach to synthesize these extended region. This will inadvertently modify nearby context, reducing visual coherence. + +To address these limitations, we introduce a novel paradigm for object removal called Masked-Region Guidance. The core idea is that the masked region should not be excluded but rather utilized as critical guidance during the removal process. Our method is straightforward: instead of replacing the masked region with a placeholder, we retain the original image as input, with the masked region indicated by a mask input, as in existing methods. As illustrated in Figure 2(b), this paradigm enables the model to accurately identify the target object, preventing unintended regeneration in the output and effectively preserving the surrounding context of the target object in the final result. + +However, our new paradigm cannot be directly implemented with existing object removal data construction methods. Typically, their training data are created by masking parts of the image, and the model is trained to predict the masked content. If applying this approach to our proposed Masked-Region Guidance paradigm, the model could exploit a shortcut by simply replicating the masked content from the input, due to the masked region being included in the input. Thus, a dataset consisting of triplets of an input, a mask, and the removal result is essential for our paradigm. Unfortunately, existing related datasets [23, 36, 45, 48] either contain a limited number of unique scenes or rely on inpainting models to generate pseudo-removed results. + +To solve this, we introduce a synthetic technique to create training data specifically for object removal. Our approach involves pasting object instances from various images onto different background images to form the input images, with the pasted instance masks serving as the input masks and the original background images designated as ground truths. Using this method, we generate Syn4Removal, a large-scale dataset comprising triplets of real background images, masks, and backgrounds with + +pasted objects. The design of Syn4Removal provides diverse scenes and supports effective training under our new paradigm, encouraging the model to accurately learn object removal without shortcuts. + +To make Syn4Removal suitable for training object removal models, we design a pipeline to generate high-quality data. First, we filter out low-quality instances and background images. Then, we develop a method to calculate feasible pasting locations on the image, ensuring that objects do not overlap with instances in the pasted area, which helps prevent the model from regenerating unwanted objects. Last, the instance is pasted onto a background image with a blending algorithm. The resulting dataset consists of 1 million image triplets. + +Using the Masked-Region Guidance paradigm and the Syn4Removal dataset, we design a framework based on text-to-image diffusion models for object removal. To enhance the model's robustness to varying mask shapes from user input, we introduce a mask enhancement technique that simulates different mask shapes over removal targets. Additionally, we incorporate CLIP-based visual guidance to assist the model in better understanding the removal targets. The resulting model called SmartEraser outperforms previous methods significantly in both quantitative and qualitative evaluations. We choose the name SmartEraser because the model can "smartly" identify removal targets and remove it while preserving its surrounding region. + +Our contributions are mainly summarized as follows: + +- We propose Masked-Region Guidance, a novel paradigm for object removal, which effectively addresses issues of object regeneration and surrounding region distortion. +- We introduce Syn4Removal, a large-scale high-quality dataset for object removal, which contains over a million pairs of images across diverse scenes and object types. +- We carefully design a framework based on the text-to-image diffusion model by introducing mask enhancement and clip-based visual guidance. +- Extensive experiments demonstrate that SmartEraser significantly outperforms previous object removal methods in terms of quality and robustness. + +# 2. Related Work + +Image inpainting. Image inpainting aims to seamlessly fill missing regions in images by leveraging contextual information around the masked region. Based on the "mask-and-inpaint" paradigm, earlier works are mainly based on generative adversarial networks (GANs) [11, 31]. Typically, these methods [2, 4, 8, 24, 25, 27, 39, 43, 49] randomly mask part areas of images [19, 20, 28, 54] and then predict these masked regions. However, GAN-based approaches [12, 41, 42, 50, 51] often suffer from limited diversity and quality in generated content, resulting in blurry or unrealistic inpainted results. + +Recently, approaches [1, 6, 9, 18, 30, 35, 47, 55] based on diffusion models have attracted significant attention due to their superior capability in generating image details. Among them, Repaint [30] injects the known region into intermediate images at all timesteps throughout the diffusion generation. SD-Inpaint [35] utilizes the mask and cropped image as input and fills the masked region through diffusion. While these approaches produce detailed textures, the generated content within the masked region often remains uncertain, with new objects occasionally appearing. + +Object removal. The object removal task aims to eliminate specified objects from images based on user-provided masks, seamlessly repairing the removed area with un-masked content. For example, PowerPaint [55] uses learnable task embeddings from context-aware inpainting as the positive prompt and those from text-guided inpainting as the negative prompt to guide object removal. A recent work CLIPAway [9] leverages AlphaCLIP [38] and IP-Adapter [46] to inject background information into the model via cross-attention, ensuring removal results coherent with the background context. The common characteristic of these methods is that they retain the traditional "mask-and-inpaint" paradigm by discarding the masked regions of input images. In this paper, we claim that this paradigm is not suitable for the object removal task exactly due to the missing context of the masked region, causing incomplete removal and inconsistent context in removal results. + +We also observe that some instruction-based methods [3, 48] attempt to remove objects using text prompts instead of the masked region. While they use the original image as input, their models are trained with removal results generated by inpainting models, which limits their performance. We believe that our synthetic strategy for building removal datasets could enhance the performance of these models. + +There are several works [23, 36, 45, 48] aiming to construct datasets for object removal. For instance, GQA-Inpaint [48] and DEFACTO [23] synthesize ground truths with the help of image inpainting models. However, we find that relying on image inpainting leads to the performance of object removal being highly limited by the effect of inpainting. Another two works RORD [36] and ObjectDrop [45] acquire real-world triplets through photography, but the high collection costs result in fewer than $3.5k$ unique scenes. To incorporate with the Masked-Region Guidance, we construct a large-scale, high-quality dataset called Syn4Removal through a set of carefully designed strategies, which includes about 1 million triplets featuring diverse scenes and authentic ground-truth backgrounds with carefully filtering and pasting strategies. + +# 3. SmartEraser + +The goal of object removal is to completely erase the target object specified by a masked region and seamlessly generate + +consistent content based on the surrounding context. Most existing methods follow a "mask-and-inpaint" paradigm, where the masked area is excluded from the input, and the model attempts to inpaint the missing content based on the unmasked regions. While these methods can successfully remove objects in simple scenarios, they often regenerate new objects or produce blurs and artifacts in the masked region in more complex scenes. + +In contrast, we propose a novel paradigm: Masked-Region Guidance. Instead of discarding masked regions, we retain it in the input, leveraging it to guide the removal process. This approach is based on the insight that, by knowing the exact removal targets, the model can precisely erase them while leaving the surrounding context maintained. In the following sections, we will first introduce the method of Masked-Region Guidance in Sec. 3.1. Then, we present Syn4Removal, a large-scale synthetic dataset designed for this paradigm in Sec. 3.2. Finally, we describe SmartEraser, a fine-tuned diffusion model based on the Mask-Region Guidance paradigm and Syn4Removal dataset in Sec. 3.3. + +# 3.1. Masked-Region Guidance + +Given an input image $\pmb{x}$ , and user-specified mask $m$ containing the removal targets, the object removal task aims to produce the removal result $\hat{\pmb{x}}$ while maintaining the realism of the original image. The mask $m$ is a binary mask where 1 indicates the region containing the removal targets. Existing models [9, 35, 39, 55] are mostly based on the "mask-and-inpaint" paradigm, i.e., utilizing the unmasked region $x \odot (1 - m)$ and mask $m$ to predict the results without removal object in the masked region. These models are learned to maximize the following objective: + +$$ +\mathcal {L} ^ {\prime} = P \left(\hat {\boldsymbol {x}} \mid \boldsymbol {m}, \boldsymbol {x} \odot (\boldsymbol {1} - \boldsymbol {m}); \Theta^ {\prime}\right), \tag {1} +$$ + +where the conditional probability $P$ is modeled using a neural network with parameters $\Theta^{\prime}$ , $\odot$ specifies the elementwise multiplication. + +In contrast, we introduce Masked-Region Guidance, a fundamentally different approach. Instead of discarding the masked region, we retain it in the input, allowing it to guide the removal process. Specifically, we replace the condition input from $[m, x \odot (1 - m)]$ to $[m, x]$ . The mask $m$ is still applied to instruct the region for removal targets, but $x$ now includes the masked region $x \odot m$ besides $x \odot (1 - m)$ . By including the masked region, we provide the model with valuable guidance during the object removal task. First, retaining the masked region allows the model to know exactly which objects need to be removed, reducing the risk of regenerating the object in the masked area. This helps the model focus on learning how to reconstruct the scene without the object. Second, the masked region is typically larger than the removal targets. This allows the model to copy the surrounding context of target objects in the masked region to the final result, significantly reducing the difficulty + +![](images/ffa3761d0ce36ce7857aaf8537c636054217d90785b19a2408a88edfb74a2bdd.jpg) +Figure 3. The data generation pipeline of Syn4Removal. We apply instances and images to construct a triplet consisting of an input image with removal targets, a mask, and the ground truth. + +of synthesizing a larger area. Moreover, if the masked region contains complex textures, these can be retained, leading to more realistic results. In our Masked-Region Guidance paradigm, object removal is learned to maximize the objective: + +$$ +\mathcal {L} = P (\hat {\boldsymbol {x}} | \boldsymbol {m}, \boldsymbol {x}; \Theta), \tag {2} +$$ + +where the model is with parameters $\Theta$ . + +# 3.2. Syn4Removal + +Now we have a new paradigm for training object removal models, but it cannot be directly applied using existing data construction methods. Typically, training data is created by masking parts of an image and having the model predict the masked content. However, this could lead the model to simply copy the masked content if applied to our paradigm. To overcome this, we require a dataset with triplets (input image, mask, and ground truth). However, existing real image datasets [36, 45] are limited in terms of scene variety, as they require photos taken before and after placing objects. Additionally, some synthetic datasets [23, 48] rely on inpainting models to generate pseudo-removed results, which don't fully capture the real removal process. + +To address this, we propose a synthetic method for constructing object removal training data. Our approach involves pasting object instances from various images onto different backgrounds to form the input images. The mask of the pasted object serves as the input mask, and the original background image serves as the ground truth. Since the ground truth is the real image, it helps the model better capture the true distribution of the image, leading to more realistic object removal results. As a result, we constructed Syn4Removal, a large-scale high-quality data set that contains 1 million samples. The data construction pipeline is shown in Figure 3. Next, we will introduce each component of the overall pipeline. + +Instance Crop & Filter. We directly use a public instance segmentation dataset [22] to get instances. We use the mask to get each instance from the image. While cutting + +instances, we also compute the corresponding CLIP similarity score with their classes and normalized area ratio in the original images. Then we perform the filter process, First, we exclude instances with areas larger than $95\%$ or smaller than $5\%$ to prevent potential low-quality segmentation results. Then, based on the previously calculated CLIP scores, we remove instances with low semantic relevance in each class. Given the variability in CLIP score distribution across different classes, we set individual CLIP score thresholds for each class according to previous work [53], filtering out instances with scores below these thresholds. + +Background Filter. For background images, we use public datasets COCONut [7] and SAM-1B [21], along with segmentation annotations. We filter out low-resolution images, those with extreme aspect ratios, or those with a large number of segmented instances. The selected backgrounds are then combined to form 1 million high-quality images (for more details, see Supplementary). + +Instance Size Sampling & Resize. To match the size of the instance with that of the background, we first calculate the average $\mu_{c}$ and variance $\sigma_c^2$ of the normalized area ratio for each class $c$ . When pasting an instance onto a background, we sample scale $s$ from a Gaussian distribution $\mathcal{N}(\mu_c,\sigma_c^2)$ , and scale the instance to $sHW$ , where $H$ and $W$ are the height and width of the background image, respectively. + +Feasible Area for Pasting. To ensure that the pasted instance does not cover existing instances in the background image, we calculate IoU between the pasted instance region and each instance in the background image, the feasible region based on IoU results is formulated as: + +$$ +\mathbf {R} _ {1} = \left\{\left(x, y\right) \mid \operatorname {I o U} \left(P _ {x, y}, I _ {i}\right) < r, I _ {i} \in [ 1, I _ {\max } ] \right\}, \tag {3} +$$ + +where $P_{x,y}$ is the instance region if centered on $(x,y)$ , $r$ is a IoU threshold, $I_{max}$ is the number of existing instances. Besides, to ensure the integrity of pasted instances, edge areas of background images are avoided, formulated as: + +$$ +\mathbf {R} _ {2} = \left\{(x, y) \mid x \in \left[ e _ {x}, W - e _ {x} \right], y \in \left[ e _ {y}, H - e _ {y} \right] \right\}, \tag {4} +$$ + +where $e_x$ and $e_y$ are half of the width and height of the bounding box of the instance mask, respectively. We take + +![](images/bde91619f2b29f08ca714a5c0674925bbcf49b13b1a8e4df43c2cf22b67b61c0.jpg) +Figure 4. Overall framework of SmartEraser. It fully utilizes the Masked-Region Guidance paradigm and the Syn4Removal dataset. + +the intersection region of $\mathbf{R}_1$ and $\mathbf{R}_2$ , the final set of feasible locations is as follows: + +$$ +\mathbf {R} _ {\mathrm {f}} = \mathbf {R} _ {1} \cap \mathbf {R} _ {2}. \tag {5} +$$ + +We then randomly select a location from $\mathbf{R}_{\mathrm{f}}$ as the center point for pasting the instance. Based on the mask of the instance and location, we can get the instance mask centered on the location $\bar{m}$ . + +Alpha Blending for Instance Pasting. To further ensure that the pasted instance blends harmoniously with the background, we apply alpha blending [13]. This method interpolates between the foreground instance and the background image, formulated as follows: + +$$ +\boldsymbol {x} = \boldsymbol {\alpha} \odot \boldsymbol {x} _ {i} + (1 - \boldsymbol {\alpha}) \odot \boldsymbol {x} _ {b}, \tag {6} +$$ + +where $\pmb{x}$ is final constructed image, $\pmb{x}_i$ is the instance, $\pmb{x}_b$ is the background image, and $\alpha$ is alpha blending value. $\alpha$ is calculated through an alpha blending algorithm [13] that processes the mask $\bar{m}$ to a trimap and then takes the trimap as input, which consists of three regions, i.e., the foreground and background, and an excessively uncertain boundary region. So we have the triplet for object removal training: $\{\pmb{x}, \bar{m}, \pmb{x}_b\}$ , where $\pmb{x}$ is input image with target object, $\bar{m}$ is original mask, and $\pmb{x}_b$ as the ground truth. Using this data generation pipeline, we finally get about 1 million triplets that form the Syn4Removal dataset. + +# 3.3. Framework + +Based on the proposed Masked-Region Guidance paradigm and the Syn4Removal dataset, we design a framework based on the text-to-image stable diffusion model for object removal. In this framework, we carefully design a user-friendly mask enhancement strategy and introduce CLIP-based visual guidance to further improve the capability of the model for object removal, as shown in Figure 4. + +Mask Enhancement. In real-world scenarios, users usually provide a loose or tight mask around the object to be removed. If the model were trained only with precise object masks, there would be a significant gap in mask shape and size between training and inference. To address this, + +![](images/1ec9c936a0816227a4f7db1679568ef8d5454fc1493cd810f188aaf1688b11a0.jpg) +Figure 5. Mask shapes from different mask enhancement methods. + +during training, we apply various mask deformation methods to simulate the user input mask shapes, as shown in Figure 5. These techniques help the model generalize to different mask forms. Specifically, we use six mask types to augment the object mask: (1) Original mask: this mask precisely outlines the target object. (2) Eroded mask: created by applying morphological erosion to the original mask, simulating scenarios where the user's mask may not fully cover the object. (3) Dilated mask: created by applying morphological dilation to the original mask. (4) Convex hull mask: constructed by calculating the convex hull that fully encompasses the original mask, then expanding it slightly. (5) Ellipse mask: generated by finding the smallest enclosing ellipse around the original mask, followed by a slight expansion. (6) Bbox & Bessel mask: a bounding box is generated to completely cover the original mask first, then bessel curves are constructed along each edge of the box to introduce an irregular boundary. So the overall mask enhancement process is illustrated as follows: + +$$ +\boldsymbol {m} = \operatorname {M E} _ {i} (\bar {\boldsymbol {m}}), \quad i \in \{1, \dots , 6 \}, \tag {7} +$$ + +where $m$ is the enhanced mask used during training, $\mathrm{ME}_i(\cdot)$ represents six different mask enhancement methods. + +CLIP-Based Visual Guidance. Our SmartEraser is designed based on text-to-image diffusion model, the text + +![](images/c0ba79f5fd51e9a5cc468b5ccaf2bce8745d0cba7961ed14ad5087cf6ed9f161.jpg) +Figure 6. Qualitative comparison of SmartEraser and state-of-the-art methods. The samples are sourced from RORD-Val, which includes the ground truth after removal. + +prompts play an important role in guiding the generation process. We aim to design a prompt that indicates the model what to remove during the generation process. Concretely, we first segment the region corresponding to the removal target in the image using its corresponding enhanced mask. This segmented region is then fed into the pre-trained CLIP [33] vision encoder to extract its visual feature(e.g., class token). This feature is not precisely aligned with the text encoder output space. So we then apply a trainable MLP network to map the visual feature into the feature space of the text encoder. The mapped embedding is then appended to the tokens of the CLIP text embedding of the prompt "Remove the instance of". During the training process, the vision encoder is frozen, and the MLP and text encoder are trainable. So the CLIP-based visual guidance is formulated as follows: + +$$ +\widehat {\boldsymbol {c}} = \left[ \boldsymbol {\tau} _ {\theta} (\boldsymbol {y}), \operatorname {M L P} \left(\boldsymbol {\nu} _ {\theta} (\boldsymbol {x} \odot \boldsymbol {m})\right) \right], \tag {8} +$$ + +where $\tau_{\theta}(\cdot)$ is text encoder, $\nu_{\theta}(\cdot)$ is vision encoder, and $\pmb{y}$ represents input text prompt "remove the instance of". + +Loss Function. Suppose the $E(\cdot)$ is the VAE encoder, so the latent feature of input image $x$ and its ground truth $x_{b}$ is $\bar{z} = E(x)$ and $z = E(x_{b})$ , respectively, and $z_{t}$ is the noisy feature of $z$ with added noise $\epsilon$ at timestep $t$ . So the overall training loss following the standard diffusion process, which is formulated as follows: + +$$ +\mathcal {L} = \mathbb {E} _ {\boldsymbol {x}, \boldsymbol {m}, \boldsymbol {x} _ {b}, t, \epsilon} \| \boldsymbol {\epsilon} - \boldsymbol {\epsilon} _ {\theta} \left(\boldsymbol {z} _ {t}, \bar {\boldsymbol {z}}, \boldsymbol {m}, \widehat {\boldsymbol {c}}, t\right) \| _ {2} ^ {2}, \tag {9} +$$ + +where $\epsilon_{\theta}(\cdot)$ is the diffusion model. + +# 4. Experiments + +# 4.1. Settings + +Baselines. We compare SmartEraser with methods in both image inpainting and object removal. For im- + +age inpainting, we evaluate against both GAN-based approaches (ZITS++ [4], MAT [24], LaMa [39], BLD [1]) and diffusion-based methods (RePaint [30], and SD-Inpaint [35]). Additionally, we compare it with specialized object removal frameworks, including CLIPAway [9] and PowerPaint [55], which represent the previous state-of-the-art performance on object removal benchmarks. + +Implementation details. To enable a fair comparison, we follow previous experimental settings [9, 55] and finetune SmartEraser on the widely-used SD v1.5 model [35]. The training process involves fine-tuning both the UNet and text encoder components within the diffusion model and MLP for mapping visual features to text features space. SmartEraser is trained on the proposed Syn4Removal dataset with a batch size of 32, using the AdamW [29] optimizer for $500k$ iterations and a learning rate of 2e-5 across all trainable modules. + +Evaluation benchmarks. We evaluate the object removal capability of all baseline methods and our approach across three benchmarks in diverse scenarios. The benchmarks are constructed as follows: + +- RORD-Val: Due to there are many triplets share the same background images as the ground truth in validation set of the RORD [36]. To ensure unique scenes and high object quality, we filter this subset to obtain RORD-Val, a evaluation benchmark of real-photography image pairs. +- DEFACTO-Val: Derived from the splicing section of DEFACTO [23], we exclude examples where intended targets for removal significantly overlap with instances in the background images. +- Syn4Removal-Val: Generated using the pipeline of Syn4Removal with MSCOCO [26] validation set as background images and OpenImages-v7 [22] validation set offering instances. + +![](images/9f763f1915b7ddd2dfce4110df6995036c940ee7a85c8968355622aca55be346.jpg) +Figure 7. Qualitative comparison of SmartEraser and other methods in real-world user cases, in which users provide various mask shapes. The samples are sourced from real-world images and the validation set of MSCOCO [26]. + +Each benchmark comprises approximately 5k samples. During inference, following [9], we apply slight mask dilation to mitigate downscaling artifacts in SD-Inpaint-based methods [9, 35, 55]. + +Evaluation metrics. To quantitatively assess the performance of the object removal models, we consider three key aspects: (1) overall image quality, which is evaluated by Fréchet Inception Distance (FID) [14] and CLIP Maximum Mean Discrepancy (CMMD) [17]. (2) consistency between the predicted region and the background context, evaluated by REMOVE [5] metric. (3) consistency between the predicted region and corresponding region in the ground truth, assessed by LPIPS [52], SSIM [44], and PSNR [16]. + +# 4.2. Comparison with Previous Methods + +Quantitative results. We conduct extensive experiments to evaluate the performance of SmartEraser against baselines on the benchmarks. The results are shown in Tables 1, 2 and 3. SmartEraser significantly outperforms all other methods across all metrics on the three benchmarks. Notably, on the more complex RORD-Val dataset, which has larger masks and more challenging removal scenes, SmartEraser surpasses the previous state-of-the-art by margins of 10.3 in FID and 1.89dB in PSNR. This indicates the superior performance of SmartEraser for object removal. Besides, SmartEraser also achieves state-of-the-art performance on the simpler DEFACTO-Val and Syn4Removal-Val datasets. + +Qualitative results. We present visualization comparison results of object removal using images from real scenes, shown in Figure 6. Baseline methods have limitations and may introduce unintended objects, create unrealistic structures, or produce artifacts within masked regions. In contrast, SmartEraser succeeds in effectively removing target + +
MethodFID ↓ CMMD ↓ReMOVE ↑LPIPS ↓SSIM ↑PSNR ↑
ZITS++ [4]53.4400.8380.8200.3880.52215.442
MAT [24]80.9531.4940.6280.4330.49513.322
LaMa [39]24.2370.2160.9160.3480.55716.383
BLD [1]73.0861.7600.8340.4710.46715.775
RePaint [30]71.5421.6390.8120.3810.51015.861
SD-Inpaint [35]69.5020.3240.8570.3690.53716.111
CLIPAway [9]25.4580.1230.9150.3330.57717.434
PowerPaint [55]24.0580.2940.9260.3080.60218.100
SmartEraser16.0300.0920.9370.2760.61219.994
+ +Table 1. Quantitative comparison of SmartEraser and previous methods on RORD-Val. + +
MethodFID ↓CMMD ↓ReLUVE ↑LPIPS ↓SSIM ↑PSNR ↑
ZITS++ [4]8.1530.2290.8990.3500.63419.453
MAT [24]15.2570.3210.7420.4070.57514.541
LaMa [39]6.6010.2010.9260.3510.68321.735
BLD [1]24.4070.7460.9130.4180.63219.346
RePaint [30]12.3140.8220.9240.3250.64420.607
SD-Inpaint [35]11.1190.1960.9120.3200.65621.691
CLIPAway [9]8.3200.1930.9250.3230.66621.953
PowerPaint [55]6.5820.2750.9340.2780.71523.870
SmartEraser3.4050.1060.9390.2570.73425.363
+ +Table 2. Quantitative comparison of SmartEraser and previous methods on DEFACTO-Val. + +objects while generating high-quality synthesized results, and excels in maintaining the contextual integrity and visual coherence of the scene. + +User study. To assess the visual quality subjectively, we conduct a user study with 30 participants. The study includes 20 image sets, each containing an original image, its mask, and removal results from four different methods, presented in random order. Participants are asked to select their preferred result based on three criteria: overall image quality, foreground-background consistency, and accuracy of object removal. Voting percentages in Table 4 show that + +
MethodFID ↓ CMMD ↓ReMOVE ↑LPIPS ↓SSIM ↑PSNR ↑
ZITS++ [4]5.9320.1930.8100.3580.58117.315
MAT [24]19.2630.3000.7000.3740.57013.357
LaMa [39]5.2460.0970.9320.3250.60619.319
BLD [1]20.2860.7050.8960.4640.52516.539
RePaint [30]12.6770.6960.9080.3780.55317.689
SD-Inpaint [35]11.5280.1070.8970.3480.57218.105
CLIPAway [9]5.5300.0930.9300.3440.59318.561
PowerPaint [55]5.0250.2000.9310.2910.66320.594
SmartEraser4.3860.0530.9390.2690.67222.029
+ +Table 3. Quantitative comparison of SmartEraser and previous methods on Syn4Removal-Val. + +
MethodImage Quality ↑ FB-Consistency ↑Removal Accuracy ↑
LaMa [39]0.100.11
CLIPAway [9]0.180.12
PowerPaint [55]0.160.12
SmartEraser0.560.65
0.58
+ +Table 4. Comparison of user study results in terms of image quality, foreground-background consistency, and removal accuracy. + +users consistently favor the results from SmartEraser. + +Real-world user cases. To further evaluate the object removal capability of SmartEraser in real-world user scenarios, we conduct a comparison with other methods (LaMa [39], CLIPAway [9], PowerPaint [55]), using masks provided by users. Figure 7 shows the comparison results. We can observe that SmartEraser exhibits a remarkable ability to accurately identify and remove target objects while preserving surrounding contextual details, particularly in scenarios involving imperfect or approximate masks. In the sample in the bottom row on the left, SmartEraser accurately identifies and removes the person indicated by the mask, while maintaining the integrity of the elephant in the background. In contrast, alternative methods tend to repaint the entire masked area, which leads to outputs that lack contextual coherence and visual realism due to the loss of details of the elephant. The upper second sample on the right provides a challenging scenario, the user-provided mask aims to remove the man whose feet are partially obscured by floating text. The previous methods fail to maintain the integrity of the floating text. However, SmartEraser accurately removes the man while preserving the floating text. This highlights SmartEraser's ability to effectively differentiate between removal targets and background content, even in cases of partial occlusion. + +# 4.3. Ablation studies + +To evaluate the effectiveness of the key components in our framework, we perform ablation studies and report both quantitative and qualitative results. We progressively modify the baseline (fine-tuning SD v1.5 with the "mask-and-inpaint" paradigm) by adding our proposed techniques: RG for masked-region guidance, ME for mask enhancement, and VG for CLIP-based visual guidance. The results are + +![](images/5e68ffc8c8dae8acd2a6323e074ab130e147e7574cabc0ab29005c4220885e15.jpg) + +![](images/e5f467d86aa545c11068c3862854136cfe8dbf4500e006cbf06e208af2dca5ca.jpg) +Mask & Image +Figure 8. Qualitative ablation comparison of our method. From left to right, we progressively add each proposed component. + +![](images/fd28d4db82f884c0d529aed183aeb22307ea5e1745be12ff4c3ee8d197b6b00a.jpg) +Baseline + +![](images/4d20676228b19ff6981570d2a1193dec52951511022b5847d5ef650df0a7519e.jpg) ++RG + +![](images/adf50fc44ea31356e6415db0de1be2cb821d51732e5a0801a532acf543b71db3.jpg) ++ME + +![](images/85f23e43d152e2de2399f16cb696b58fa079e34fe7ac026208cacefc02e3d856.jpg) ++VG + +
MethodFID ↓CMMD ↓ReLUVE ↑LPIPS ↓SSIM ↑PSNR ↑
Baseline33.1310.2450.9080.3550.54416.535
+RG24.5260.1210.9230.3390.56217.176
+ME19.4210.1080.9310.3010.59118.917
+VG16.0300.0920.9370.2760.61619.994
+ +Table 5. Quantitative ablated comparison on RORD-Val. + +displayed in Figure 8 and Table 5. + +Effect of the Masked-Region Guidance paradigm: We observe that the baseline model often regenerates objects instead of properly removing them, as it is trained in a traditional "inpainting" manner. Incorporating masked-region guidance significantly improves the removal performance by reducing the risk of object regeneration. However, due to the mismatch between the mask shape used during training and the user-defined mask during testing, the model does not always perfectly remove the target objects and preserve the surrounding context. + +Contribution of Mask Enhancement: By simulating real-world user mask shapes over object removal during training, we find that it helps reduce the gap between model training and testing using user input mask, significantly keeping the surrounding regions and further improving removal results. + +Value of CLIP-based Visual Guidance: CLIP visual guidance aids the model in better understanding the removal targets, yielding the best performance on RORD-val across all metrics. This indicates that the visual guidance helps the model perform object removal tasks more effectively. + +# 5. Conclusion + +In conclusion, we present SmartEraser, a state-of-the-art object removal model built on the novel Masked-Region Guidance paradigm. SmartEraser outperforms existing methods by smartly identifying the target object to remove while effectively preserving the surrounding context. To facilitate research on this paradigm, we propose Syn4Removal, a large-scale, high-quality dataset containing over a million image triplets, specifically designed for object removal tasks. Through extensive experiments, we demonstrate that SmartEraser achieves superior performance in both quality and robustness compared to previous object removal methods. + +# References + +[1] Omri Avrahami, Ohad Fried, and Dani Lischinski. Blended latent diffusion. ACM transactions on graphics (TOG), 42 (4):1-11, 2023. 1, 3, 6, 7, 8 +[2] Connelly Barnes, Eli Shechtman, Adam Finkelstein, and Dan B Goldman. Patchmatch: A randomized correspondence algorithm for structural image editing. ACM Trans. Graph., 28(3):24, 2009. 2 +[3] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. arXiv preprint arXiv:2211.09800, 2022. 3 +[4] Chenjie Cao, Qiaole Dong, and Yanwei Fu. Zits++: Image inpainting by improving the incremental transformer on structural priors. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 2, 6, 7, 8 +[5] Aditya Chandrasekar, Goirik Chakrabarty, Jai Bardhan, Ramya Hebbalaguppe, and Prathosh AP. Remove: A reference-free metric for object erasure. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 7901-7910, 2024. 7 +[6] Pau de Jorge, Riccardo Volpi, Puneet K. Dakania, Philip H. S. Torr, and Rogez Gregory. Placing objects in context via inpainting for out-of-distribution segmentation. In The European Conference on Computer Vision (ECCV), 2024. 3 +[7] Xueqing Deng, Qihang Yu, Peng Wang, Xiaohui Shen, and Liang-Chieh Chen. Coconut: Modernizing coco segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21863-21873, 2024. 4 +[8] Qiaole Dong, Chenjie Cao, and Yanwei Fu. Incremental transformer structure enhanced image inpainting with masking positional encoding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11358-11368, 2022. 2 +[9] Yigit Ekin, Ahmet Burak Yildirim, Erdem Eren Caglar, Aykut Erdem, Erkut Erdem, and Aysegul Dundar. Clipaway: Harmonizing focused embeddings for removing objects via diffusion models, 2024. 1, 3, 6, 7, 8 +[10] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12873-12883, 2021. 1 +[11] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 1, 2 +[12] Xiefan Guo, Hongyu Yang, and Di Huang. Image inpainting via conditional texture and structure dual generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 14134-14143, 2021. 2 +[13] Kaiming He, Christoph Rhemann, Carsten Rother, Xiaou Tang, and Jian Sun. A global sampling method for alpha matting. In CVPR 2011, pages 2049-2056, 2011. 5 +[14] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilib- + +rium. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2017. 7 +[15] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1 +[16] Alain Hore and Djemel Ziou. Image quality metrics: Psnr vs. sSIM. In 2010 20th international conference on pattern recognition, pages 2366-2369. IEEE, 2010. 7 +[17] Sadeep Jayasumana, Srikumar Ramalingam, Andreas Veit, Daniel Glasner, Ayan Chakrabarti, and Sanjiv Kumar. Rethinking fid: Towards a better evaluation metric for image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9307-9315, 2024. 7 +[18] Xuan Ju, Xian Liu, Xintao Wang, Yuxuan Bian, Ying Shan, and Qiang Xu. Brushnet: A plug-and-play image inpainting model with decomposed dual-branch diffusion, 2024. 3 +[19] Tero Karras. Progressive growing of gans for improved quality, stability, and variation. arXiv preprint arXiv:1710.10196, 2017. 2 +[20] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 2 +[21] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 4 +[22] Alina Kuznetsova, Hassan Rom, Neil Alldrin, Jasper Uijlings, Ivan Krasin, Jordi Pont-Tuset, Shahab Kamali, Stefan Popov, Matteo Malloci, Alexander Kolesnikov, et al. The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale. International journal of computer vision, 128(7):1956-1981, 2020. 4, 6 +[23] Jens Lehmann, Daniel Gerber, Mohamed Morsey, and Axel-Cyrille Ngonga Ngomo. Defacto-deep fact validation. In International semantic web conference, pages 312-327. Springer, 2012. 2, 3, 4, 6 +[24] Wenbo Li, Zhe Lin, Kun Zhou, Lu Qi, Yi Wang, and Jiaya Jia. Mat: Mask-aware transformer for large hole image inpainting. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10758-10768, 2022. 2, 6, 7, 8 +[25] Xiaoguang Li, Qing Guo, Di Lin, Ping Li, Wei Feng, and Song Wnag. Misf: Multi-level interactive siamese filtering for high-fidelity image inpainting. CVPR, 2022. 2 +[26] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 6, 7 +[27] Qiankun Liu, Zhentao Tan, Dongdong Chen, Qi Chu, Xiyang Dai, Yinpeng Chen, Mengchen Liu, Lu Yuan, and Nenghai Yu. Reduce information loss in transformers for pluralistic + +image inpainting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11347-11357, 2022. 2 +[28] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaou Tang. Deep learning face attributes in the wild. In Proceedings of International Conference on Computer Vision (ICCV), 2015. 2 +[29] I Loshchilov. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6 +[30] Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11461-11471, 2022. 1, 3, 6, 7, 8 +[31] Mehdi Mirza. Conditional generative adversarial nets. arXiv preprint arXiv:1411.1784, 2014. 2 +[32] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 1 +[33] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 6 +[34] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 1 +[35] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684-10695, 2022. 1, 3, 6, 7, 8 +[36] Min-Cheol Sagong, Yoon-Jae Yeo, Seung-Won Jung, and Sung-Jea Ko. Rord: A real-world object removal dataset. In BMVC, page 542, 2022. 2, 3, 4, 6 +[37] Yang Song, Prafulla Dhariwal, Mark Chen, and Ilya Sutskever. Consistency models. arXiv preprint arXiv:2303.01469, 2023. 1 +[38] Zeyi Sun, Ye Fang, Tong Wu, Pan Zhang, Yuhang Zang, Shu Kong, Yuanjun Xiong, Dahua Lin, and Jiaqi Wang. Alphaclip: A clip model focusing on wherever you want. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13019-13029, 2024. 3 +[39] Roman Suvorov, Elizaveta Logacheva, Anton Mashikhin, Anastasia Remizova, Armenii Ashukha, Aleksei Silvestrov, Naejin Kong, Harshith Goka, Kiwoong Park, and Victor Lempitsky. Resolution-robust large mask inpainting with fourier convolutions. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 2149–2159, 2022. 1, 2, 3, 6, 7, 8 +[40] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017. 1 + +[41] Ziyu Wan, Jingbo Zhang, Dongdong Chen, and Jing Liao. High-fidelity pluralistic image completion with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 4692-4701, 2021. 2 +[42] Wentao Wang, Jianfu Zhang, Li Niu, Haoyu Ling, Xue Yang, and Liqing Zhang. Parallel multi-resolution fusion network for image inpainting. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 14559-14568, 2021. 2 +[43] Wentao Wang, Li Niu, Jianfu Zhang, Xue Yang, and Liqing Zhang. Dual-path image inpainting with auxiliary gan inversion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11421-11430, 2022. 2 +[44] Zhou Wang, A.C. Bovik, H.R. Sheikh, and E.P. Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing, 13(4): 600-612, 2004. 7 +[45] Daniel Winter, Matan Cohen, Shlomi Fruchter, Yael Pritch, Alex Rav-Acha, and Yedid Hoshen. Objectdrop: Bootstrapping counterfactuals for photorealistic object removal and insertion. arXiv preprint arXiv:2403.18818, 2024. 2, 3, 4 +[46] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721, 2023. 3 +[47] Yifu Chen, Jingwen Chen, Yingwei Pan, Yehao Li, Ting Yao, Zhineng Chen and Tao Mei. Improving text-guided object inpainting with semantic pre-inpainting. ECCV, 2024. 3 +[48] Ahmet Burak Yildirim, Vedat Baday, Erkut Erdem, Aykut Erdem, and Aysegul Dundar. Inst-inpaint: Instructing to remove objects with diffusion models. arXiv preprint arXiv:2304.03246, 2023. 2, 3, 4 +[49] Jiahui Yu, Zhe Lin, Jimei Yang, Xiaohui Shen, Xin Lu, and Thomas S Huang. Generative image inpainting with contextual attention. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5505-5514, 2018. 2 +[50] Yingchen Yu, Fangneng Zhan, Shijian Lu, Jianxiong Pan, Feiying Ma, Xuansong Xie, and Chunyan Miao. Wavefill: A wavelet-based generation network for image inpainting. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 14114-14123, 2021. 2 +[51] Yu Zeng, Zhe Lin, Huchuan Lu, and Vishal M. Patel. Crfill: Generative image inpainting with auxiliary contextual reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 14164-14173, 2021. 2 +[52] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 7 +[53] Hanqing Zhao, Dianmo Sheng, Jianmin Bao, Dongdong Chen, Dong Chen, Fang Wen, Lu Yuan, Ce Liu, Wenbo Zhou, Qi Chu, et al. X-paste: Revisiting scalable copy + +paste for instance segmentation using clip and stablediffusion. In International Conference on Machine Learning, pages 42098-42109. PMLR, 2023. 4 +[54] Bolei Zhou, Agata Lapedriza, Aditya Khosla, Aude Oliva, and Antonio Torralba. Places: A 10 million image database for scene recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2017. 2 +[55] Junhao Zhuang, Yanhong Zeng, Wenran Liu, Chun Yuan, and Kai Chen. A task is worth one word: Learning with task prompts for high-quality versatile image inpainting, 2023. 1, 3, 6, 7, 8 \ No newline at end of file diff --git a/2025/SmartEraser_ Remove Anything from Images using Masked-Region Guidance/images.zip b/2025/SmartEraser_ Remove Anything from Images using Masked-Region Guidance/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..0b2028ee8ef7c6f784c9f00460ef4cd55c3f26f0 --- /dev/null +++ b/2025/SmartEraser_ Remove Anything from Images using Masked-Region Guidance/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b4c594027ac45b9489d1c21d57bbeac3b35502d6148f1c81488035720a412e99 +size 970876 diff --git a/2025/SmartEraser_ Remove Anything from Images using Masked-Region Guidance/layout.json b/2025/SmartEraser_ Remove Anything from Images using Masked-Region Guidance/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..34def74655e61b3469533d9dd32ddd1e8a7aebc1 --- /dev/null +++ b/2025/SmartEraser_ Remove Anything from Images using Masked-Region Guidance/layout.json @@ -0,0 +1,8985 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 64, + 103, + 547, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 103, + 547, + 121 + ], + "spans": [ + { + "bbox": [ + 64, + 103, + 547, + 121 + ], + "type": "text", + "content": "SmartEraser: Remove Anything from Images using Masked-Region Guidance" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 119, + 142, + 490, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 119, + 142, + 490, + 186 + ], + "spans": [ + { + "bbox": [ + 119, + 142, + 490, + 186 + ], + "type": "text", + "content": "Longtao Jiang" + }, + { + "bbox": [ + 119, + 142, + 490, + 186 + ], + "type": "inline_equation", + "content": "^{1,*}" + }, + { + "bbox": [ + 119, + 142, + 490, + 186 + ], + "type": "text", + "content": " Zhendong Wang" + }, + { + "bbox": [ + 119, + 142, + 490, + 186 + ], + "type": "inline_equation", + "content": "^{1,*}" + }, + { + "bbox": [ + 119, + 142, + 490, + 186 + ], + "type": "text", + "content": " Jianmin Bao" + }, + { + "bbox": [ + 119, + 142, + 490, + 186 + ], + "type": "inline_equation", + "content": "^{2,*\\dagger\\heartsuit}" + }, + { + "bbox": [ + 119, + 142, + 490, + 186 + ], + "type": "text", + "content": " \nWengang Zhou" + }, + { + "bbox": [ + 119, + 142, + 490, + 186 + ], + "type": "inline_equation", + "content": "^{1,\\dagger}" + }, + { + "bbox": [ + 119, + 142, + 490, + 186 + ], + "type": "text", + "content": " Dongdong Chen" + }, + { + "bbox": [ + 119, + 142, + 490, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 119, + 142, + 490, + 186 + ], + "type": "text", + "content": " Lei Shi" + }, + { + "bbox": [ + 119, + 142, + 490, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 119, + 142, + 490, + 186 + ], + "type": "text", + "content": " Dong Chen" + }, + { + "bbox": [ + 119, + 142, + 490, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 119, + 142, + 490, + 186 + ], + "type": "text", + "content": " Houqiang Li" + }, + { + "bbox": [ + 119, + 142, + 490, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 119, + 142, + 490, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 119, + 142, + 490, + 186 + ], + "type": "text", + "content": "University of Science and Technology of China " + }, + { + "bbox": [ + 119, + 142, + 490, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 119, + 142, + 490, + 186 + ], + "type": "text", + "content": "Microsoft Research Asia" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 159, + 188, + 447, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 188, + 447, + 200 + ], + "spans": [ + { + "bbox": [ + 159, + 188, + 447, + 200 + ], + "type": "text", + "content": "https://longtaojiang.github.io/smarteraser.github.io/" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 56, + 209, + 180, + 292 + ], + "blocks": [ + { + "bbox": [ + 56, + 209, + 180, + 292 + ], + "lines": [ + { + "bbox": [ + 56, + 209, + 180, + 292 + ], + "spans": [ + { + "bbox": [ + 56, + 209, + 180, + 292 + ], + "type": "image", + "image_path": "c4720aa160d5b7008e62a16834eae9fde6110efeb03b337fcba703c311e9a059.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 56, + 293, + 180, + 376 + ], + "blocks": [ + { + "bbox": [ + 56, + 293, + 180, + 376 + ], + "lines": [ + { + "bbox": [ + 56, + 293, + 180, + 376 + ], + "spans": [ + { + "bbox": [ + 56, + 293, + 180, + 376 + ], + "type": "image", + "image_path": "d5b7d52efe55e5d764fde0d5d783b2af33065824ee9f0ad13a68404c32fb5049.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 377, + 555, + 400 + ], + "lines": [ + { + "bbox": [ + 55, + 377, + 555, + 400 + ], + "spans": [ + { + "bbox": [ + 55, + 377, + 555, + 400 + ], + "type": "text", + "content": "Figure 1. Object removal results generated by SmartEraser using user-provided masks. SmartEraser not only generates photorealistic images with the target objects removed but also better preserves the surrounding context around the removed objects." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 181, + 210, + 304, + 292 + ], + "blocks": [ + { + "bbox": [ + 181, + 210, + 304, + 292 + ], + "lines": [ + { + "bbox": [ + 181, + 210, + 304, + 292 + ], + "spans": [ + { + "bbox": [ + 181, + 210, + 304, + 292 + ], + "type": "image", + "image_path": "1524c6c7a097df5eaa242b1884fb1b9c44b334932b615f3c476d24b5a24023fa.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 181, + 293, + 304, + 376 + ], + "blocks": [ + { + "bbox": [ + 181, + 293, + 304, + 376 + ], + "lines": [ + { + "bbox": [ + 181, + 293, + 304, + 376 + ], + "spans": [ + { + "bbox": [ + 181, + 293, + 304, + 376 + ], + "type": "image", + "image_path": "950fbb09f39b5ba636edd120fc083b91495dceaec2a736df2e137f3ecfd865a0.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 306, + 210, + 429, + 293 + ], + "blocks": [ + { + "bbox": [ + 306, + 210, + 429, + 293 + ], + "lines": [ + { + "bbox": [ + 306, + 210, + 429, + 293 + ], + "spans": [ + { + "bbox": [ + 306, + 210, + 429, + 293 + ], + "type": "image", + "image_path": "0fead885f7fc6f87d75b68fe360d3bd9468fe6c2c4c294e9e194c976221d5e3d.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 306, + 293, + 429, + 376 + ], + "blocks": [ + { + "bbox": [ + 306, + 293, + 429, + 376 + ], + "lines": [ + { + "bbox": [ + 306, + 293, + 429, + 376 + ], + "spans": [ + { + "bbox": [ + 306, + 293, + 429, + 376 + ], + "type": "image", + "image_path": "1200313b16ce45eb43d12a2561e36cc1b22309a7fa064d3eae940c1f154926ba.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 430, + 210, + 553, + 293 + ], + "blocks": [ + { + "bbox": [ + 430, + 210, + 553, + 293 + ], + "lines": [ + { + "bbox": [ + 430, + 210, + 553, + 293 + ], + "spans": [ + { + "bbox": [ + 430, + 210, + 553, + 293 + ], + "type": "image", + "image_path": "7c5233145bfdde27eb941d041254a901206ab3a69142dbad05c724bc4d845d44.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 430, + 293, + 553, + 376 + ], + "blocks": [ + { + "bbox": [ + 430, + 293, + 553, + 376 + ], + "lines": [ + { + "bbox": [ + 430, + 293, + 553, + 376 + ], + "spans": [ + { + "bbox": [ + 430, + 293, + 553, + 376 + ], + "type": "image", + "image_path": "40d980fbb1cea1b99064ca7618696cf7cba2773baf30a3d4030bc508f548800f.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 151, + 410, + 200, + 422 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 410, + 200, + 422 + ], + "spans": [ + { + "bbox": [ + 151, + 410, + 200, + 422 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 54, + 436, + 297, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 436, + 297, + 700 + ], + "spans": [ + { + "bbox": [ + 54, + 436, + 297, + 700 + ], + "type": "text", + "content": "Object removal has so far been dominated by the \"mask-and-inpaint\" paradigm, where the masked region is excluded from the input, leaving models relying on unmasked areas to inpaint the missing region. However, this approach lacks contextual information for the masked area, often resulting in unstable performance. In this work, we introduce SmartEraser, built with a new \"removing\" paradigm called Masked-Region Guidance. This paradigm retains the masked region in the input, using it as guidance for the removal process. It offers several distinct advantages: (a) it guides the model to accurately identify the object to be removed, preventing its regeneration in the output; (b) since the user mask often extends beyond the object itself, it aids in preserving the surrounding context in the final result. Leveraging this new paradigm, we present Syn4Removal, a large-scale object removal dataset, where instance segmentation data is used to copy and paste objects onto images as removal targets, with the original images serving as ground truths. Experimental results demonstrate that SmartEraser significantly outperforms existing methods, achieving superior performance in object removal, especially in complex scenes with intricate compositions." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 315, + 410, + 394, + 422 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 410, + 394, + 422 + ], + "spans": [ + { + "bbox": [ + 315, + 410, + 394, + 422 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 430, + 555, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 430, + 555, + 538 + ], + "spans": [ + { + "bbox": [ + 313, + 430, + 555, + 538 + ], + "type": "text", + "content": "With the rapid advancement of generative models [10, 11, 15, 32, 34, 35, 37, 40], image editing has garnered significant attention for its broad applications. Among various editing tasks, object removal [9, 55] is especially prominent, empowering users to seamlessly eliminate unwanted elements while maintaining the realism of original images. This functionality, essential for removing distracting objects in photos, has become a key feature in widely used applications like Photoshop, Google Photos, Microsoft Designer." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 541, + 556, + 697 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 541, + 556, + 697 + ], + "spans": [ + { + "bbox": [ + 313, + 541, + 556, + 697 + ], + "type": "text", + "content": "Currently, most object removal methods [1, 9, 30, 39, 55] adopt a \"mask-and-inpaint\" paradigm, where the masked region is excluded from the input and typically filled with a neutral placeholder (e.g., black). The model then inpaints the masked region based on the surrounding content. However, we identify two primary issues with this approach. First, this strategy often generates unintended objects within the masked region. Lacking precise discrimination between removal targets and other content, these methods rely heavily on background context, sometimes leading to the unintentional inpainting of new objects—such as adding a new car on the road rather than removing the original one, as illustrated in Figure 2(a). Second, user-" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 328, + 702, + 543, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 702, + 543, + 714 + ], + "spans": [ + { + "bbox": [ + 328, + 702, + 543, + 714 + ], + "type": "text", + "content": "* Equal contribution. † Corresponding authors. ⋆ Project leader." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "24452" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 70, + 295, + 193 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 295, + 193 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 295, + 193 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 295, + 193 + ], + "type": "image", + "image_path": "2724c1b7c9d985aa928a4c861c6acf4c53978b066861cde15142b229639f606b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 126, + 194, + 269, + 204 + ], + "lines": [ + { + "bbox": [ + 126, + 194, + 269, + 204 + ], + "spans": [ + { + "bbox": [ + 126, + 194, + 269, + 204 + ], + "type": "text", + "content": "(b) Our paradigm: Masked-Region Guidance" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 205, + 295, + 250 + ], + "lines": [ + { + "bbox": [ + 55, + 205, + 295, + 250 + ], + "spans": [ + { + "bbox": [ + 55, + 205, + 295, + 250 + ], + "type": "text", + "content": "Figure 2. Comparison of the \"mask-and-inpaint\" paradigm and the proposed Masked-Region Guidance paradigm for object removal. Unlike the mask-and-inpaint approach, our paradigm uses the masked regions as critical guidance for object removal." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 254, + 295, + 302 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 254, + 295, + 302 + ], + "spans": [ + { + "bbox": [ + 55, + 254, + 295, + 302 + ], + "type": "text", + "content": "defined masks frequently exceed the target object, requiring the \"mask-and-inpaint\" approach to synthesize these extended region. This will inadvertently modify nearby context, reducing visual coherence." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 304, + 295, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 304, + 295, + 447 + ], + "spans": [ + { + "bbox": [ + 55, + 304, + 295, + 447 + ], + "type": "text", + "content": "To address these limitations, we introduce a novel paradigm for object removal called Masked-Region Guidance. The core idea is that the masked region should not be excluded but rather utilized as critical guidance during the removal process. Our method is straightforward: instead of replacing the masked region with a placeholder, we retain the original image as input, with the masked region indicated by a mask input, as in existing methods. As illustrated in Figure 2(b), this paradigm enables the model to accurately identify the target object, preventing unintended regeneration in the output and effectively preserving the surrounding context of the target object in the final result." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 449, + 295, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 449, + 295, + 604 + ], + "spans": [ + { + "bbox": [ + 55, + 449, + 295, + 604 + ], + "type": "text", + "content": "However, our new paradigm cannot be directly implemented with existing object removal data construction methods. Typically, their training data are created by masking parts of the image, and the model is trained to predict the masked content. If applying this approach to our proposed Masked-Region Guidance paradigm, the model could exploit a shortcut by simply replicating the masked content from the input, due to the masked region being included in the input. Thus, a dataset consisting of triplets of an input, a mask, and the removal result is essential for our paradigm. Unfortunately, existing related datasets [23, 36, 45, 48] either contain a limited number of unique scenes or rely on inpainting models to generate pseudo-removed results." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "content": "To solve this, we introduce a synthetic technique to create training data specifically for object removal. Our approach involves pasting object instances from various images onto different background images to form the input images, with the pasted instance masks serving as the input masks and the original background images designated as ground truths. Using this method, we generate Syn4Removal, a large-scale dataset comprising triplets of real background images, masks, and backgrounds with" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 553, + 119 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 119 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 119 + ], + "type": "text", + "content": "pasted objects. The design of Syn4Removal provides diverse scenes and supports effective training under our new paradigm, encouraging the model to accurately learn object removal without shortcuts." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 120, + 555, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 120, + 555, + 239 + ], + "spans": [ + { + "bbox": [ + 313, + 120, + 555, + 239 + ], + "type": "text", + "content": "To make Syn4Removal suitable for training object removal models, we design a pipeline to generate high-quality data. First, we filter out low-quality instances and background images. Then, we develop a method to calculate feasible pasting locations on the image, ensuring that objects do not overlap with instances in the pasted area, which helps prevent the model from regenerating unwanted objects. Last, the instance is pasted onto a background image with a blending algorithm. The resulting dataset consists of 1 million image triplets." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 240, + 554, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 240, + 554, + 395 + ], + "spans": [ + { + "bbox": [ + 313, + 240, + 554, + 395 + ], + "type": "text", + "content": "Using the Masked-Region Guidance paradigm and the Syn4Removal dataset, we design a framework based on text-to-image diffusion models for object removal. To enhance the model's robustness to varying mask shapes from user input, we introduce a mask enhancement technique that simulates different mask shapes over removal targets. Additionally, we incorporate CLIP-based visual guidance to assist the model in better understanding the removal targets. The resulting model called SmartEraser outperforms previous methods significantly in both quantitative and qualitative evaluations. We choose the name SmartEraser because the model can \"smartly\" identify removal targets and remove it while preserving its surrounding region." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 326, + 396, + 541, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 396, + 541, + 407 + ], + "spans": [ + { + "bbox": [ + 326, + 396, + 541, + 407 + ], + "type": "text", + "content": "Our contributions are mainly summarized as follows:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 407, + 554, + 550 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 314, + 407, + 554, + 442 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 407, + 554, + 442 + ], + "spans": [ + { + "bbox": [ + 314, + 407, + 554, + 442 + ], + "type": "text", + "content": "- We propose Masked-Region Guidance, a novel paradigm for object removal, which effectively addresses issues of object regeneration and surrounding region distortion." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 443, + 553, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 443, + 553, + 479 + ], + "spans": [ + { + "bbox": [ + 314, + 443, + 553, + 479 + ], + "type": "text", + "content": "- We introduce Syn4Removal, a large-scale high-quality dataset for object removal, which contains over a million pairs of images across diverse scenes and object types." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 479, + 553, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 479, + 553, + 514 + ], + "spans": [ + { + "bbox": [ + 314, + 479, + 553, + 514 + ], + "type": "text", + "content": "- We carefully design a framework based on the text-to-image diffusion model by introducing mask enhancement and clip-based visual guidance." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 314, + 515, + 553, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 515, + 553, + 550 + ], + "spans": [ + { + "bbox": [ + 314, + 515, + 553, + 550 + ], + "type": "text", + "content": "- Extensive experiments demonstrate that SmartEraser significantly outperforms previous object removal methods in terms of quality and robustness." + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 562, + 400, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 562, + 400, + 574 + ], + "spans": [ + { + "bbox": [ + 314, + 562, + 400, + 574 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 582, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 554, + 713 + ], + "type": "text", + "content": "Image inpainting. Image inpainting aims to seamlessly fill missing regions in images by leveraging contextual information around the masked region. Based on the \"mask-and-inpaint\" paradigm, earlier works are mainly based on generative adversarial networks (GANs) [11, 31]. Typically, these methods [2, 4, 8, 24, 25, 27, 39, 43, 49] randomly mask part areas of images [19, 20, 28, 54] and then predict these masked regions. However, GAN-based approaches [12, 41, 42, 50, 51] often suffer from limited diversity and quality in generated content, resulting in blurry or unrealistic inpainted results." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24453" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "type": "text", + "content": "Recently, approaches [1, 6, 9, 18, 30, 35, 47, 55] based on diffusion models have attracted significant attention due to their superior capability in generating image details. Among them, Repaint [30] injects the known region into intermediate images at all timesteps throughout the diffusion generation. SD-Inpaint [35] utilizes the mask and cropped image as input and fills the masked region through diffusion. While these approaches produce detailed textures, the generated content within the masked region often remains uncertain, with new objects occasionally appearing." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 192, + 295, + 393 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 192, + 295, + 393 + ], + "spans": [ + { + "bbox": [ + 55, + 192, + 295, + 393 + ], + "type": "text", + "content": "Object removal. The object removal task aims to eliminate specified objects from images based on user-provided masks, seamlessly repairing the removed area with un-masked content. For example, PowerPaint [55] uses learnable task embeddings from context-aware inpainting as the positive prompt and those from text-guided inpainting as the negative prompt to guide object removal. A recent work CLIPAway [9] leverages AlphaCLIP [38] and IP-Adapter [46] to inject background information into the model via cross-attention, ensuring removal results coherent with the background context. The common characteristic of these methods is that they retain the traditional \"mask-and-inpaint\" paradigm by discarding the masked regions of input images. In this paper, we claim that this paradigm is not suitable for the object removal task exactly due to the missing context of the masked region, causing incomplete removal and inconsistent context in removal results." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 396, + 295, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 396, + 295, + 478 + ], + "spans": [ + { + "bbox": [ + 55, + 396, + 295, + 478 + ], + "type": "text", + "content": "We also observe that some instruction-based methods [3, 48] attempt to remove objects using text prompts instead of the masked region. While they use the original image as input, their models are trained with removal results generated by inpainting models, which limits their performance. We believe that our synthetic strategy for building removal datasets could enhance the performance of these models." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 479, + 295, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 479, + 295, + 658 + ], + "spans": [ + { + "bbox": [ + 55, + 479, + 295, + 658 + ], + "type": "text", + "content": "There are several works [23, 36, 45, 48] aiming to construct datasets for object removal. For instance, GQA-Inpaint [48] and DEFACTO [23] synthesize ground truths with the help of image inpainting models. However, we find that relying on image inpainting leads to the performance of object removal being highly limited by the effect of inpainting. Another two works RORD [36] and ObjectDrop [45] acquire real-world triplets through photography, but the high collection costs result in fewer than " + }, + { + "bbox": [ + 55, + 479, + 295, + 658 + ], + "type": "inline_equation", + "content": "3.5k" + }, + { + "bbox": [ + 55, + 479, + 295, + 658 + ], + "type": "text", + "content": " unique scenes. To incorporate with the Masked-Region Guidance, we construct a large-scale, high-quality dataset called Syn4Removal through a set of carefully designed strategies, which includes about 1 million triplets featuring diverse scenes and authentic ground-truth backgrounds with carefully filtering and pasting strategies." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 669, + 137, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 669, + 137, + 681 + ], + "spans": [ + { + "bbox": [ + 55, + 669, + 137, + 681 + ], + "type": "text", + "content": "3. SmartEraser" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "content": "The goal of object removal is to completely erase the target object specified by a masked region and seamlessly generate" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "text", + "content": "consistent content based on the surrounding context. Most existing methods follow a \"mask-and-inpaint\" paradigm, where the masked area is excluded from the input, and the model attempts to inpaint the missing content based on the unmasked regions. While these methods can successfully remove objects in simple scenarios, they often regenerate new objects or produce blurs and artifacts in the masked region in more complex scenes." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 168, + 553, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 168, + 553, + 312 + ], + "spans": [ + { + "bbox": [ + 313, + 168, + 553, + 312 + ], + "type": "text", + "content": "In contrast, we propose a novel paradigm: Masked-Region Guidance. Instead of discarding masked regions, we retain it in the input, leveraging it to guide the removal process. This approach is based on the insight that, by knowing the exact removal targets, the model can precisely erase them while leaving the surrounding context maintained. In the following sections, we will first introduce the method of Masked-Region Guidance in Sec. 3.1. Then, we present Syn4Removal, a large-scale synthetic dataset designed for this paradigm in Sec. 3.2. Finally, we describe SmartEraser, a fine-tuned diffusion model based on the Mask-Region Guidance paradigm and Syn4Removal dataset in Sec. 3.3." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 317, + 459, + 330 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 317, + 459, + 330 + ], + "spans": [ + { + "bbox": [ + 313, + 317, + 459, + 330 + ], + "type": "text", + "content": "3.1. Masked-Region Guidance" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 335, + 553, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 335, + 553, + 455 + ], + "spans": [ + { + "bbox": [ + 313, + 335, + 553, + 455 + ], + "type": "text", + "content": "Given an input image " + }, + { + "bbox": [ + 313, + 335, + 553, + 455 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 313, + 335, + 553, + 455 + ], + "type": "text", + "content": ", and user-specified mask " + }, + { + "bbox": [ + 313, + 335, + 553, + 455 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 335, + 553, + 455 + ], + "type": "text", + "content": " containing the removal targets, the object removal task aims to produce the removal result " + }, + { + "bbox": [ + 313, + 335, + 553, + 455 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{x}}" + }, + { + "bbox": [ + 313, + 335, + 553, + 455 + ], + "type": "text", + "content": " while maintaining the realism of the original image. The mask " + }, + { + "bbox": [ + 313, + 335, + 553, + 455 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 335, + 553, + 455 + ], + "type": "text", + "content": " is a binary mask where 1 indicates the region containing the removal targets. Existing models [9, 35, 39, 55] are mostly based on the \"mask-and-inpaint\" paradigm, i.e., utilizing the unmasked region " + }, + { + "bbox": [ + 313, + 335, + 553, + 455 + ], + "type": "inline_equation", + "content": "x \\odot (1 - m)" + }, + { + "bbox": [ + 313, + 335, + 553, + 455 + ], + "type": "text", + "content": " and mask " + }, + { + "bbox": [ + 313, + 335, + 553, + 455 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 335, + 553, + 455 + ], + "type": "text", + "content": " to predict the results without removal object in the masked region. These models are learned to maximize the following objective:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 365, + 456, + 553, + 469 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 456, + 553, + 469 + ], + "spans": [ + { + "bbox": [ + 365, + 456, + 553, + 469 + ], + "type": "interline_equation", + "content": "\\mathcal {L} ^ {\\prime} = P \\left(\\hat {\\boldsymbol {x}} \\mid \\boldsymbol {m}, \\boldsymbol {x} \\odot (\\boldsymbol {1} - \\boldsymbol {m}); \\Theta^ {\\prime}\\right), \\tag {1}", + "image_path": "68ea0c9594dcfe541efc8ba875f2dcc63a74e17bcc99f73569d0dabff5ab28de.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 475, + 553, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 475, + 553, + 509 + ], + "spans": [ + { + "bbox": [ + 313, + 475, + 553, + 509 + ], + "type": "text", + "content": "where the conditional probability " + }, + { + "bbox": [ + 313, + 475, + 553, + 509 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 313, + 475, + 553, + 509 + ], + "type": "text", + "content": " is modeled using a neural network with parameters " + }, + { + "bbox": [ + 313, + 475, + 553, + 509 + ], + "type": "inline_equation", + "content": "\\Theta^{\\prime}" + }, + { + "bbox": [ + 313, + 475, + 553, + 509 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 475, + 553, + 509 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 313, + 475, + 553, + 509 + ], + "type": "text", + "content": " specifies the elementwise multiplication." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 510, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 553, + 713 + ], + "type": "text", + "content": "In contrast, we introduce Masked-Region Guidance, a fundamentally different approach. Instead of discarding the masked region, we retain it in the input, allowing it to guide the removal process. Specifically, we replace the condition input from " + }, + { + "bbox": [ + 313, + 510, + 553, + 713 + ], + "type": "inline_equation", + "content": "[m, x \\odot (1 - m)]" + }, + { + "bbox": [ + 313, + 510, + 553, + 713 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 510, + 553, + 713 + ], + "type": "inline_equation", + "content": "[m, x]" + }, + { + "bbox": [ + 313, + 510, + 553, + 713 + ], + "type": "text", + "content": ". The mask " + }, + { + "bbox": [ + 313, + 510, + 553, + 713 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 510, + 553, + 713 + ], + "type": "text", + "content": " is still applied to instruct the region for removal targets, but " + }, + { + "bbox": [ + 313, + 510, + 553, + 713 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 510, + 553, + 713 + ], + "type": "text", + "content": " now includes the masked region " + }, + { + "bbox": [ + 313, + 510, + 553, + 713 + ], + "type": "inline_equation", + "content": "x \\odot m" + }, + { + "bbox": [ + 313, + 510, + 553, + 713 + ], + "type": "text", + "content": " besides " + }, + { + "bbox": [ + 313, + 510, + 553, + 713 + ], + "type": "inline_equation", + "content": "x \\odot (1 - m)" + }, + { + "bbox": [ + 313, + 510, + 553, + 713 + ], + "type": "text", + "content": ". By including the masked region, we provide the model with valuable guidance during the object removal task. First, retaining the masked region allows the model to know exactly which objects need to be removed, reducing the risk of regenerating the object in the masked area. This helps the model focus on learning how to reconstruct the scene without the object. Second, the masked region is typically larger than the removal targets. This allows the model to copy the surrounding context of target objects in the masked region to the final result, significantly reducing the difficulty" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24454" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 68, + 553, + 210 + ], + "blocks": [ + { + "bbox": [ + 56, + 68, + 553, + 210 + ], + "lines": [ + { + "bbox": [ + 56, + 68, + 553, + 210 + ], + "spans": [ + { + "bbox": [ + 56, + 68, + 553, + 210 + ], + "type": "image", + "image_path": "ffa3761d0ce36ce7857aaf8537c636054217d90785b19a2408a88edfb74a2bdd.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 212, + 555, + 234 + ], + "lines": [ + { + "bbox": [ + 55, + 212, + 555, + 234 + ], + "spans": [ + { + "bbox": [ + 55, + 212, + 555, + 234 + ], + "type": "text", + "content": "Figure 3. The data generation pipeline of Syn4Removal. We apply instances and images to construct a triplet consisting of an input image with removal targets, a mask, and the ground truth." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 245, + 296, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 245, + 296, + 304 + ], + "spans": [ + { + "bbox": [ + 55, + 245, + 296, + 304 + ], + "type": "text", + "content": "of synthesizing a larger area. Moreover, if the masked region contains complex textures, these can be retained, leading to more realistic results. In our Masked-Region Guidance paradigm, object removal is learned to maximize the objective:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 308, + 295, + 321 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 308, + 295, + 321 + ], + "spans": [ + { + "bbox": [ + 133, + 308, + 295, + 321 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = P (\\hat {\\boldsymbol {x}} | \\boldsymbol {m}, \\boldsymbol {x}; \\Theta), \\tag {2}", + "image_path": "d3c452f90bc93ee1d11eea3d4b1d5b8ce0e131d4a54d08cad23f6aa65415c005.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 329, + 214, + 341 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 329, + 214, + 341 + ], + "spans": [ + { + "bbox": [ + 55, + 329, + 214, + 341 + ], + "type": "text", + "content": "where the model is with parameters " + }, + { + "bbox": [ + 55, + 329, + 214, + 341 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 55, + 329, + 214, + 341 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 348, + 143, + 361 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 348, + 143, + 361 + ], + "spans": [ + { + "bbox": [ + 55, + 348, + 143, + 361 + ], + "type": "text", + "content": "3.2. Syn4Removal" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 366, + 296, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 366, + 296, + 521 + ], + "spans": [ + { + "bbox": [ + 55, + 366, + 296, + 521 + ], + "type": "text", + "content": "Now we have a new paradigm for training object removal models, but it cannot be directly applied using existing data construction methods. Typically, training data is created by masking parts of an image and having the model predict the masked content. However, this could lead the model to simply copy the masked content if applied to our paradigm. To overcome this, we require a dataset with triplets (input image, mask, and ground truth). However, existing real image datasets [36, 45] are limited in terms of scene variety, as they require photos taken before and after placing objects. Additionally, some synthetic datasets [23, 48] rely on inpainting models to generate pseudo-removed results, which don't fully capture the real removal process." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 522, + 296, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 522, + 296, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 522, + 296, + 677 + ], + "type": "text", + "content": "To address this, we propose a synthetic method for constructing object removal training data. Our approach involves pasting object instances from various images onto different backgrounds to form the input images. The mask of the pasted object serves as the input mask, and the original background image serves as the ground truth. Since the ground truth is the real image, it helps the model better capture the true distribution of the image, leading to more realistic object removal results. As a result, we constructed Syn4Removal, a large-scale high-quality data set that contains 1 million samples. The data construction pipeline is shown in Figure 3. Next, we will introduce each component of the overall pipeline." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": "Instance Crop & Filter. We directly use a public instance segmentation dataset [22] to get instances. We use the mask to get each instance from the image. While cutting" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 245, + 555, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 245, + 555, + 376 + ], + "spans": [ + { + "bbox": [ + 313, + 245, + 555, + 376 + ], + "type": "text", + "content": "instances, we also compute the corresponding CLIP similarity score with their classes and normalized area ratio in the original images. Then we perform the filter process, First, we exclude instances with areas larger than " + }, + { + "bbox": [ + 313, + 245, + 555, + 376 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 313, + 245, + 555, + 376 + ], + "type": "text", + "content": " or smaller than " + }, + { + "bbox": [ + 313, + 245, + 555, + 376 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 313, + 245, + 555, + 376 + ], + "type": "text", + "content": " to prevent potential low-quality segmentation results. Then, based on the previously calculated CLIP scores, we remove instances with low semantic relevance in each class. Given the variability in CLIP score distribution across different classes, we set individual CLIP score thresholds for each class according to previous work [53], filtering out instances with scores below these thresholds." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 376, + 554, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 376, + 554, + 460 + ], + "spans": [ + { + "bbox": [ + 313, + 376, + 554, + 460 + ], + "type": "text", + "content": "Background Filter. For background images, we use public datasets COCONut [7] and SAM-1B [21], along with segmentation annotations. We filter out low-resolution images, those with extreme aspect ratios, or those with a large number of segmented instances. The selected backgrounds are then combined to form 1 million high-quality images (for more details, see Supplementary)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 460, + 554, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 460, + 554, + 544 + ], + "spans": [ + { + "bbox": [ + 313, + 460, + 554, + 544 + ], + "type": "text", + "content": "Instance Size Sampling & Resize. To match the size of the instance with that of the background, we first calculate the average " + }, + { + "bbox": [ + 313, + 460, + 554, + 544 + ], + "type": "inline_equation", + "content": "\\mu_{c}" + }, + { + "bbox": [ + 313, + 460, + 554, + 544 + ], + "type": "text", + "content": " and variance " + }, + { + "bbox": [ + 313, + 460, + 554, + 544 + ], + "type": "inline_equation", + "content": "\\sigma_c^2" + }, + { + "bbox": [ + 313, + 460, + 554, + 544 + ], + "type": "text", + "content": " of the normalized area ratio for each class " + }, + { + "bbox": [ + 313, + 460, + 554, + 544 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 460, + 554, + 544 + ], + "type": "text", + "content": ". When pasting an instance onto a background, we sample scale " + }, + { + "bbox": [ + 313, + 460, + 554, + 544 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 313, + 460, + 554, + 544 + ], + "type": "text", + "content": " from a Gaussian distribution " + }, + { + "bbox": [ + 313, + 460, + 554, + 544 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(\\mu_c,\\sigma_c^2)" + }, + { + "bbox": [ + 313, + 460, + 554, + 544 + ], + "type": "text", + "content": ", and scale the instance to " + }, + { + "bbox": [ + 313, + 460, + 554, + 544 + ], + "type": "inline_equation", + "content": "sHW" + }, + { + "bbox": [ + 313, + 460, + 554, + 544 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 460, + 554, + 544 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 313, + 460, + 554, + 544 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 460, + 554, + 544 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 313, + 460, + 554, + 544 + ], + "type": "text", + "content": " are the height and width of the background image, respectively." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 544, + 554, + 603 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 544, + 554, + 603 + ], + "spans": [ + { + "bbox": [ + 313, + 544, + 554, + 603 + ], + "type": "text", + "content": "Feasible Area for Pasting. To ensure that the pasted instance does not cover existing instances in the background image, we calculate IoU between the pasted instance region and each instance in the background image, the feasible region based on IoU results is formulated as:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 325, + 604, + 553, + 618 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 604, + 553, + 618 + ], + "spans": [ + { + "bbox": [ + 325, + 604, + 553, + 618 + ], + "type": "interline_equation", + "content": "\\mathbf {R} _ {1} = \\left\\{\\left(x, y\\right) \\mid \\operatorname {I o U} \\left(P _ {x, y}, I _ {i}\\right) < r, I _ {i} \\in [ 1, I _ {\\max } ] \\right\\}, \\tag {3}", + "image_path": "697dab01cc9f2f36d7b4117c4c60d01ca42c2455ad502cabf90b57bd3932a3d6.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 623, + 554, + 670 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 623, + 554, + 670 + ], + "spans": [ + { + "bbox": [ + 313, + 623, + 554, + 670 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 623, + 554, + 670 + ], + "type": "inline_equation", + "content": "P_{x,y}" + }, + { + "bbox": [ + 313, + 623, + 554, + 670 + ], + "type": "text", + "content": " is the instance region if centered on " + }, + { + "bbox": [ + 313, + 623, + 554, + 670 + ], + "type": "inline_equation", + "content": "(x,y)" + }, + { + "bbox": [ + 313, + 623, + 554, + 670 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 623, + 554, + 670 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 313, + 623, + 554, + 670 + ], + "type": "text", + "content": " is a IoU threshold, " + }, + { + "bbox": [ + 313, + 623, + 554, + 670 + ], + "type": "inline_equation", + "content": "I_{max}" + }, + { + "bbox": [ + 313, + 623, + 554, + 670 + ], + "type": "text", + "content": " is the number of existing instances. Besides, to ensure the integrity of pasted instances, edge areas of background images are avoided, formulated as:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 323, + 671, + 553, + 685 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 671, + 553, + 685 + ], + "spans": [ + { + "bbox": [ + 323, + 671, + 553, + 685 + ], + "type": "interline_equation", + "content": "\\mathbf {R} _ {2} = \\left\\{(x, y) \\mid x \\in \\left[ e _ {x}, W - e _ {x} \\right], y \\in \\left[ e _ {y}, H - e _ {y} \\right] \\right\\}, \\tag {4}", + "image_path": "1baf9fb7683e8e78d5e9dfa5a8e8369e4c3052dedcd946aae6a0532df46ade67.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "inline_equation", + "content": "e_x" + }, + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "inline_equation", + "content": "e_y" + }, + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "content": " are half of the width and height of the bounding box of the instance mask, respectively. We take" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24455" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 70, + 553, + 203 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 553, + 203 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 553, + 203 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 553, + 203 + ], + "type": "image", + "image_path": "bde91619f2b29f08ca714a5c0674925bbcf49b13b1a8e4df43c2cf22b67b61c0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 66, + 205, + 541, + 217 + ], + "lines": [ + { + "bbox": [ + 66, + 205, + 541, + 217 + ], + "spans": [ + { + "bbox": [ + 66, + 205, + 541, + 217 + ], + "type": "text", + "content": "Figure 4. Overall framework of SmartEraser. It fully utilizes the Masked-Region Guidance paradigm and the Syn4Removal dataset." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 228, + 295, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 228, + 295, + 251 + ], + "spans": [ + { + "bbox": [ + 55, + 228, + 295, + 251 + ], + "type": "text", + "content": "the intersection region of " + }, + { + "bbox": [ + 55, + 228, + 295, + 251 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_1" + }, + { + "bbox": [ + 55, + 228, + 295, + 251 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 228, + 295, + 251 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_2" + }, + { + "bbox": [ + 55, + 228, + 295, + 251 + ], + "type": "text", + "content": ", the final set of feasible locations is as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 141, + 256, + 295, + 269 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 256, + 295, + 269 + ], + "spans": [ + { + "bbox": [ + 141, + 256, + 295, + 269 + ], + "type": "interline_equation", + "content": "\\mathbf {R} _ {\\mathrm {f}} = \\mathbf {R} _ {1} \\cap \\mathbf {R} _ {2}. \\tag {5}", + "image_path": "ff7127b2d48bb82b3aa497f9f2ab580550c00370638d7d28edb78a87c6a9a026.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 278, + 296, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 278, + 296, + 324 + ], + "spans": [ + { + "bbox": [ + 55, + 278, + 296, + 324 + ], + "type": "text", + "content": "We then randomly select a location from " + }, + { + "bbox": [ + 55, + 278, + 296, + 324 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_{\\mathrm{f}}" + }, + { + "bbox": [ + 55, + 278, + 296, + 324 + ], + "type": "text", + "content": " as the center point for pasting the instance. Based on the mask of the instance and location, we can get the instance mask centered on the location " + }, + { + "bbox": [ + 55, + 278, + 296, + 324 + ], + "type": "inline_equation", + "content": "\\bar{m}" + }, + { + "bbox": [ + 55, + 278, + 296, + 324 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 326, + 296, + 385 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 326, + 296, + 385 + ], + "spans": [ + { + "bbox": [ + 55, + 326, + 296, + 385 + ], + "type": "text", + "content": "Alpha Blending for Instance Pasting. To further ensure that the pasted instance blends harmoniously with the background, we apply alpha blending [13]. This method interpolates between the foreground instance and the background image, formulated as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 113, + 391, + 294, + 403 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 391, + 294, + 403 + ], + "spans": [ + { + "bbox": [ + 113, + 391, + 294, + 403 + ], + "type": "interline_equation", + "content": "\\boldsymbol {x} = \\boldsymbol {\\alpha} \\odot \\boldsymbol {x} _ {i} + (1 - \\boldsymbol {\\alpha}) \\odot \\boldsymbol {x} _ {b}, \\tag {6}", + "image_path": "ebf8b9c5faaccb9d84f4d6f34b2c9ae8ee587d07d58fd64edd48ae338c28d70e.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 413, + 296, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 413, + 296, + 544 + ], + "spans": [ + { + "bbox": [ + 55, + 413, + 296, + 544 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 413, + 296, + 544 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 55, + 413, + 296, + 544 + ], + "type": "text", + "content": " is final constructed image, " + }, + { + "bbox": [ + 55, + 413, + 296, + 544 + ], + "type": "inline_equation", + "content": "\\pmb{x}_i" + }, + { + "bbox": [ + 55, + 413, + 296, + 544 + ], + "type": "text", + "content": " is the instance, " + }, + { + "bbox": [ + 55, + 413, + 296, + 544 + ], + "type": "inline_equation", + "content": "\\pmb{x}_b" + }, + { + "bbox": [ + 55, + 413, + 296, + 544 + ], + "type": "text", + "content": " is the background image, and " + }, + { + "bbox": [ + 55, + 413, + 296, + 544 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 413, + 296, + 544 + ], + "type": "text", + "content": " is alpha blending value. " + }, + { + "bbox": [ + 55, + 413, + 296, + 544 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 413, + 296, + 544 + ], + "type": "text", + "content": " is calculated through an alpha blending algorithm [13] that processes the mask " + }, + { + "bbox": [ + 55, + 413, + 296, + 544 + ], + "type": "inline_equation", + "content": "\\bar{m}" + }, + { + "bbox": [ + 55, + 413, + 296, + 544 + ], + "type": "text", + "content": " to a trimap and then takes the trimap as input, which consists of three regions, i.e., the foreground and background, and an excessively uncertain boundary region. So we have the triplet for object removal training: " + }, + { + "bbox": [ + 55, + 413, + 296, + 544 + ], + "type": "inline_equation", + "content": "\\{\\pmb{x}, \\bar{m}, \\pmb{x}_b\\}" + }, + { + "bbox": [ + 55, + 413, + 296, + 544 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 413, + 296, + 544 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 55, + 413, + 296, + 544 + ], + "type": "text", + "content": " is input image with target object, " + }, + { + "bbox": [ + 55, + 413, + 296, + 544 + ], + "type": "inline_equation", + "content": "\\bar{m}" + }, + { + "bbox": [ + 55, + 413, + 296, + 544 + ], + "type": "text", + "content": " is original mask, and " + }, + { + "bbox": [ + 55, + 413, + 296, + 544 + ], + "type": "inline_equation", + "content": "\\pmb{x}_b" + }, + { + "bbox": [ + 55, + 413, + 296, + 544 + ], + "type": "text", + "content": " as the ground truth. Using this data generation pipeline, we finally get about 1 million triplets that form the Syn4Removal dataset." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 552, + 134, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 552, + 134, + 562 + ], + "spans": [ + { + "bbox": [ + 55, + 552, + 134, + 562 + ], + "type": "text", + "content": "3.3. Framework" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 570, + 296, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 296, + 653 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 296, + 653 + ], + "type": "text", + "content": "Based on the proposed Masked-Region Guidance paradigm and the Syn4Removal dataset, we design a framework based on the text-to-image stable diffusion model for object removal. In this framework, we carefully design a user-friendly mask enhancement strategy and introduce CLIP-based visual guidance to further improve the capability of the model for object removal, as shown in Figure 4." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "content": "Mask Enhancement. In real-world scenarios, users usually provide a loose or tight mask around the object to be removed. If the model were trained only with precise object masks, there would be a significant gap in mask shape and size between training and inference. To address this," + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 315, + 222, + 554, + 388 + ], + "blocks": [ + { + "bbox": [ + 315, + 222, + 554, + 388 + ], + "lines": [ + { + "bbox": [ + 315, + 222, + 554, + 388 + ], + "spans": [ + { + "bbox": [ + 315, + 222, + 554, + 388 + ], + "type": "image", + "image_path": "1ec9c936a0816227a4f7db1679568ef8d5454fc1493cd810f188aaf1688b11a0.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 314, + 389, + 553, + 399 + ], + "lines": [ + { + "bbox": [ + 314, + 389, + 553, + 399 + ], + "spans": [ + { + "bbox": [ + 314, + 389, + 553, + 399 + ], + "type": "text", + "content": "Figure 5. Mask shapes from different mask enhancement methods." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 407, + 555, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 407, + 555, + 634 + ], + "spans": [ + { + "bbox": [ + 313, + 407, + 555, + 634 + ], + "type": "text", + "content": "during training, we apply various mask deformation methods to simulate the user input mask shapes, as shown in Figure 5. These techniques help the model generalize to different mask forms. Specifically, we use six mask types to augment the object mask: (1) Original mask: this mask precisely outlines the target object. (2) Eroded mask: created by applying morphological erosion to the original mask, simulating scenarios where the user's mask may not fully cover the object. (3) Dilated mask: created by applying morphological dilation to the original mask. (4) Convex hull mask: constructed by calculating the convex hull that fully encompasses the original mask, then expanding it slightly. (5) Ellipse mask: generated by finding the smallest enclosing ellipse around the original mask, followed by a slight expansion. (6) Bbox & Bessel mask: a bounding box is generated to completely cover the original mask first, then bessel curves are constructed along each edge of the box to introduce an irregular boundary. So the overall mask enhancement process is illustrated as follows:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 367, + 641, + 553, + 654 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 641, + 553, + 654 + ], + "spans": [ + { + "bbox": [ + 367, + 641, + 553, + 654 + ], + "type": "interline_equation", + "content": "\\boldsymbol {m} = \\operatorname {M E} _ {i} (\\bar {\\boldsymbol {m}}), \\quad i \\in \\{1, \\dots , 6 \\}, \\tag {7}", + "image_path": "8bb09704d79de1e5bf6aae0ac4f4f47fe00ad91b980ebf26901c37aafce97190.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 665, + 553, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 553, + 688 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 553, + 688 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 665, + 553, + 688 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 665, + 553, + 688 + ], + "type": "text", + "content": " is the enhanced mask used during training, " + }, + { + "bbox": [ + 313, + 665, + 553, + 688 + ], + "type": "inline_equation", + "content": "\\mathrm{ME}_i(\\cdot)" + }, + { + "bbox": [ + 313, + 665, + 553, + 688 + ], + "type": "text", + "content": " represents six different mask enhancement methods." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "content": "CLIP-Based Visual Guidance. Our SmartEraser is designed based on text-to-image diffusion model, the text" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24456" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 70, + 555, + 262 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 555, + 262 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 555, + 262 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 555, + 262 + ], + "type": "image", + "image_path": "c0ba79f5fd51e9a5cc468b5ccaf2bce8745d0cba7961ed14ad5087cf6ed9f161.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 264, + 555, + 285 + ], + "lines": [ + { + "bbox": [ + 55, + 264, + 555, + 285 + ], + "spans": [ + { + "bbox": [ + 55, + 264, + 555, + 285 + ], + "type": "text", + "content": "Figure 6. Qualitative comparison of SmartEraser and state-of-the-art methods. The samples are sourced from RORD-Val, which includes the ground truth after removal." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 297, + 297, + 488 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 297, + 297, + 488 + ], + "spans": [ + { + "bbox": [ + 54, + 297, + 297, + 488 + ], + "type": "text", + "content": "prompts play an important role in guiding the generation process. We aim to design a prompt that indicates the model what to remove during the generation process. Concretely, we first segment the region corresponding to the removal target in the image using its corresponding enhanced mask. This segmented region is then fed into the pre-trained CLIP [33] vision encoder to extract its visual feature(e.g., class token). This feature is not precisely aligned with the text encoder output space. So we then apply a trainable MLP network to map the visual feature into the feature space of the text encoder. The mapped embedding is then appended to the tokens of the CLIP text embedding of the prompt \"Remove the instance of\". During the training process, the vision encoder is frozen, and the MLP and text encoder are trainable. So the CLIP-based visual guidance is formulated as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 108, + 491, + 294, + 506 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 491, + 294, + 506 + ], + "spans": [ + { + "bbox": [ + 108, + 491, + 294, + 506 + ], + "type": "interline_equation", + "content": "\\widehat {\\boldsymbol {c}} = \\left[ \\boldsymbol {\\tau} _ {\\theta} (\\boldsymbol {y}), \\operatorname {M L P} \\left(\\boldsymbol {\\nu} _ {\\theta} (\\boldsymbol {x} \\odot \\boldsymbol {m})\\right) \\right], \\tag {8}", + "image_path": "36e350fd4224b5f6d37697babbcf97074f2175a837393dc76b511880cee5958d.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 510, + 295, + 535 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 295, + 535 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 295, + 535 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 510, + 295, + 535 + ], + "type": "inline_equation", + "content": "\\tau_{\\theta}(\\cdot)" + }, + { + "bbox": [ + 55, + 510, + 295, + 535 + ], + "type": "text", + "content": " is text encoder, " + }, + { + "bbox": [ + 55, + 510, + 295, + 535 + ], + "type": "inline_equation", + "content": "\\nu_{\\theta}(\\cdot)" + }, + { + "bbox": [ + 55, + 510, + 295, + 535 + ], + "type": "text", + "content": " is vision encoder, and " + }, + { + "bbox": [ + 55, + 510, + 295, + 535 + ], + "type": "inline_equation", + "content": "\\pmb{y}" + }, + { + "bbox": [ + 55, + 510, + 295, + 535 + ], + "type": "text", + "content": " represents input text prompt \"remove the instance of\"." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 535, + 295, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 535, + 295, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 535, + 295, + 605 + ], + "type": "text", + "content": "Loss Function. Suppose the " + }, + { + "bbox": [ + 55, + 535, + 295, + 605 + ], + "type": "inline_equation", + "content": "E(\\cdot)" + }, + { + "bbox": [ + 55, + 535, + 295, + 605 + ], + "type": "text", + "content": " is the VAE encoder, so the latent feature of input image " + }, + { + "bbox": [ + 55, + 535, + 295, + 605 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 535, + 295, + 605 + ], + "type": "text", + "content": " and its ground truth " + }, + { + "bbox": [ + 55, + 535, + 295, + 605 + ], + "type": "inline_equation", + "content": "x_{b}" + }, + { + "bbox": [ + 55, + 535, + 295, + 605 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 55, + 535, + 295, + 605 + ], + "type": "inline_equation", + "content": "\\bar{z} = E(x)" + }, + { + "bbox": [ + 55, + 535, + 295, + 605 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 535, + 295, + 605 + ], + "type": "inline_equation", + "content": "z = E(x_{b})" + }, + { + "bbox": [ + 55, + 535, + 295, + 605 + ], + "type": "text", + "content": ", respectively, and " + }, + { + "bbox": [ + 55, + 535, + 295, + 605 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 55, + 535, + 295, + 605 + ], + "type": "text", + "content": " is the noisy feature of " + }, + { + "bbox": [ + 55, + 535, + 295, + 605 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 55, + 535, + 295, + 605 + ], + "type": "text", + "content": " with added noise " + }, + { + "bbox": [ + 55, + 535, + 295, + 605 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 55, + 535, + 295, + 605 + ], + "type": "text", + "content": " at timestep " + }, + { + "bbox": [ + 55, + 535, + 295, + 605 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 535, + 295, + 605 + ], + "type": "text", + "content": ". So the overall training loss following the standard diffusion process, which is formulated as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 89, + 609, + 295, + 623 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 609, + 295, + 623 + ], + "spans": [ + { + "bbox": [ + 89, + 609, + 295, + 623 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathbb {E} _ {\\boldsymbol {x}, \\boldsymbol {m}, \\boldsymbol {x} _ {b}, t, \\epsilon} \\| \\boldsymbol {\\epsilon} - \\boldsymbol {\\epsilon} _ {\\theta} \\left(\\boldsymbol {z} _ {t}, \\bar {\\boldsymbol {z}}, \\boldsymbol {m}, \\widehat {\\boldsymbol {c}}, t\\right) \\| _ {2} ^ {2}, \\tag {9}", + "image_path": "53a56a1b05666ab1d9617711eb1902ef735f65ca187f5b7c5fa5d7c5e17dd16b.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 629, + 198, + 642 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 198, + 642 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 198, + 642 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 629, + 198, + 642 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(\\cdot)" + }, + { + "bbox": [ + 55, + 629, + 198, + 642 + ], + "type": "text", + "content": " is the diffusion model." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 651, + 137, + 664 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 651, + 137, + 664 + ], + "spans": [ + { + "bbox": [ + 55, + 651, + 137, + 664 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 671, + 116, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 671, + 116, + 684 + ], + "spans": [ + { + "bbox": [ + 55, + 671, + 116, + 684 + ], + "type": "text", + "content": "4.1. Settings" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": "Baselines. We compare SmartEraser with methods in both image inpainting and object removal. For im-" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 297, + 555, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 297, + 555, + 380 + ], + "spans": [ + { + "bbox": [ + 313, + 297, + 555, + 380 + ], + "type": "text", + "content": "age inpainting, we evaluate against both GAN-based approaches (ZITS++ [4], MAT [24], LaMa [39], BLD [1]) and diffusion-based methods (RePaint [30], and SD-Inpaint [35]). Additionally, we compare it with specialized object removal frameworks, including CLIPAway [9] and PowerPaint [55], which represent the previous state-of-the-art performance on object removal benchmarks." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 384, + 556, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 384, + 556, + 503 + ], + "spans": [ + { + "bbox": [ + 313, + 384, + 556, + 503 + ], + "type": "text", + "content": "Implementation details. To enable a fair comparison, we follow previous experimental settings [9, 55] and finetune SmartEraser on the widely-used SD v1.5 model [35]. The training process involves fine-tuning both the UNet and text encoder components within the diffusion model and MLP for mapping visual features to text features space. SmartEraser is trained on the proposed Syn4Removal dataset with a batch size of 32, using the AdamW [29] optimizer for " + }, + { + "bbox": [ + 313, + 384, + 556, + 503 + ], + "type": "inline_equation", + "content": "500k" + }, + { + "bbox": [ + 313, + 384, + 556, + 503 + ], + "type": "text", + "content": " iterations and a learning rate of 2e-5 across all trainable modules." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 506, + 555, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 506, + 555, + 554 + ], + "spans": [ + { + "bbox": [ + 313, + 506, + 555, + 554 + ], + "type": "text", + "content": "Evaluation benchmarks. We evaluate the object removal capability of all baseline methods and our approach across three benchmarks in diverse scenarios. The benchmarks are constructed as follows:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 558, + 554, + 712 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 313, + 558, + 554, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 554, + 617 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 554, + 617 + ], + "type": "text", + "content": "- RORD-Val: Due to there are many triplets share the same background images as the ground truth in validation set of the RORD [36]. To ensure unique scenes and high object quality, we filter this subset to obtain RORD-Val, a evaluation benchmark of real-photography image pairs." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 618, + 553, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 618, + 553, + 666 + ], + "spans": [ + { + "bbox": [ + 313, + 618, + 553, + 666 + ], + "type": "text", + "content": "- DEFACTO-Val: Derived from the splicing section of DEFACTO [23], we exclude examples where intended targets for removal significantly overlap with instances in the background images." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 666, + 553, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 666, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 666, + 553, + 712 + ], + "type": "text", + "content": "- Syn4Removal-Val: Generated using the pipeline of Syn4Removal with MSCOCO [26] validation set as background images and OpenImages-v7 [22] validation set offering instances." + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24457" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 70, + 555, + 280 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 555, + 280 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 555, + 280 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 555, + 280 + ], + "type": "image", + "image_path": "9f763f1915b7ddd2dfce4110df6995036c940ee7a85c8968355622aca55be346.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 281, + 553, + 304 + ], + "lines": [ + { + "bbox": [ + 55, + 281, + 553, + 304 + ], + "spans": [ + { + "bbox": [ + 55, + 281, + 553, + 304 + ], + "type": "text", + "content": "Figure 7. Qualitative comparison of SmartEraser and other methods in real-world user cases, in which users provide various mask shapes. The samples are sourced from real-world images and the validation set of MSCOCO [26]." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 309, + 295, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 309, + 295, + 357 + ], + "spans": [ + { + "bbox": [ + 55, + 309, + 295, + 357 + ], + "type": "text", + "content": "Each benchmark comprises approximately 5k samples. During inference, following [9], we apply slight mask dilation to mitigate downscaling artifacts in SD-Inpaint-based methods [9, 35, 55]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 358, + 296, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 358, + 296, + 468 + ], + "spans": [ + { + "bbox": [ + 55, + 358, + 296, + 468 + ], + "type": "text", + "content": "Evaluation metrics. To quantitatively assess the performance of the object removal models, we consider three key aspects: (1) overall image quality, which is evaluated by Fréchet Inception Distance (FID) [14] and CLIP Maximum Mean Discrepancy (CMMD) [17]. (2) consistency between the predicted region and the background context, evaluated by REMOVE [5] metric. (3) consistency between the predicted region and corresponding region in the ground truth, assessed by LPIPS [52], SSIM [44], and PSNR [16]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 477, + 252, + 490 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 477, + 252, + 490 + ], + "spans": [ + { + "bbox": [ + 55, + 477, + 252, + 490 + ], + "type": "text", + "content": "4.2. Comparison with Previous Methods" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 497, + 295, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 497, + 295, + 640 + ], + "spans": [ + { + "bbox": [ + 55, + 497, + 295, + 640 + ], + "type": "text", + "content": "Quantitative results. We conduct extensive experiments to evaluate the performance of SmartEraser against baselines on the benchmarks. The results are shown in Tables 1, 2 and 3. SmartEraser significantly outperforms all other methods across all metrics on the three benchmarks. Notably, on the more complex RORD-Val dataset, which has larger masks and more challenging removal scenes, SmartEraser surpasses the previous state-of-the-art by margins of 10.3 in FID and 1.89dB in PSNR. This indicates the superior performance of SmartEraser for object removal. Besides, SmartEraser also achieves state-of-the-art performance on the simpler DEFACTO-Val and Syn4Removal-Val datasets." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": "Qualitative results. We present visualization comparison results of object removal using images from real scenes, shown in Figure 6. Baseline methods have limitations and may introduce unintended objects, create unrealistic structures, or produce artifacts within masked regions. In contrast, SmartEraser succeeds in effectively removing target" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 318, + 307, + 553, + 416 + ], + "blocks": [ + { + "bbox": [ + 318, + 307, + 553, + 416 + ], + "lines": [ + { + "bbox": [ + 318, + 307, + 553, + 416 + ], + "spans": [ + { + "bbox": [ + 318, + 307, + 553, + 416 + ], + "type": "table", + "html": "
MethodFID ↓ CMMD ↓ReMOVE ↑LPIPS ↓SSIM ↑PSNR ↑
ZITS++ [4]53.4400.8380.8200.3880.52215.442
MAT [24]80.9531.4940.6280.4330.49513.322
LaMa [39]24.2370.2160.9160.3480.55716.383
BLD [1]73.0861.7600.8340.4710.46715.775
RePaint [30]71.5421.6390.8120.3810.51015.861
SD-Inpaint [35]69.5020.3240.8570.3690.53716.111
CLIPAway [9]25.4580.1230.9150.3330.57717.434
PowerPaint [55]24.0580.2940.9260.3080.60218.100
SmartEraser16.0300.0920.9370.2760.61219.994
", + "image_path": "f5a4df3f0c2a4ce62e39de378685b7bf9d10343d7b8950f4e6e923e8ad00e19d.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 318, + 440, + 553, + 548 + ], + "blocks": [ + { + "bbox": [ + 314, + 418, + 553, + 438 + ], + "lines": [ + { + "bbox": [ + 314, + 418, + 553, + 438 + ], + "spans": [ + { + "bbox": [ + 314, + 418, + 553, + 438 + ], + "type": "text", + "content": "Table 1. Quantitative comparison of SmartEraser and previous methods on RORD-Val." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 318, + 440, + 553, + 548 + ], + "lines": [ + { + "bbox": [ + 318, + 440, + 553, + 548 + ], + "spans": [ + { + "bbox": [ + 318, + 440, + 553, + 548 + ], + "type": "table", + "html": "
MethodFID ↓CMMD ↓ReLUVE ↑LPIPS ↓SSIM ↑PSNR ↑
ZITS++ [4]8.1530.2290.8990.3500.63419.453
MAT [24]15.2570.3210.7420.4070.57514.541
LaMa [39]6.6010.2010.9260.3510.68321.735
BLD [1]24.4070.7460.9130.4180.63219.346
RePaint [30]12.3140.8220.9240.3250.64420.607
SD-Inpaint [35]11.1190.1960.9120.3200.65621.691
CLIPAway [9]8.3200.1930.9250.3230.66621.953
PowerPaint [55]6.5820.2750.9340.2780.71523.870
SmartEraser3.4050.1060.9390.2570.73425.363
", + "image_path": "8874f0199269727aa30f0d17343a87179d5b04ce65e3f04e18f713176b19287c.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 550, + 553, + 572 + ], + "lines": [ + { + "bbox": [ + 313, + 550, + 553, + 572 + ], + "spans": [ + { + "bbox": [ + 313, + 550, + 553, + 572 + ], + "type": "text", + "content": "Table 2. Quantitative comparison of SmartEraser and previous methods on DEFACTO-Val." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 580, + 553, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 580, + 553, + 616 + ], + "spans": [ + { + "bbox": [ + 313, + 580, + 553, + 616 + ], + "type": "text", + "content": "objects while generating high-quality synthesized results, and excels in maintaining the contextual integrity and visual coherence of the scene." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "type": "text", + "content": "User study. To assess the visual quality subjectively, we conduct a user study with 30 participants. The study includes 20 image sets, each containing an original image, its mask, and removal results from four different methods, presented in random order. Participants are asked to select their preferred result based on three criteria: overall image quality, foreground-background consistency, and accuracy of object removal. Voting percentages in Table 4 show that" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24458" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 59, + 70, + 293, + 178 + ], + "blocks": [ + { + "bbox": [ + 59, + 70, + 293, + 178 + ], + "lines": [ + { + "bbox": [ + 59, + 70, + 293, + 178 + ], + "spans": [ + { + "bbox": [ + 59, + 70, + 293, + 178 + ], + "type": "table", + "html": "
MethodFID ↓ CMMD ↓ReMOVE ↑LPIPS ↓SSIM ↑PSNR ↑
ZITS++ [4]5.9320.1930.8100.3580.58117.315
MAT [24]19.2630.3000.7000.3740.57013.357
LaMa [39]5.2460.0970.9320.3250.60619.319
BLD [1]20.2860.7050.8960.4640.52516.539
RePaint [30]12.6770.6960.9080.3780.55317.689
SD-Inpaint [35]11.5280.1070.8970.3480.57218.105
CLIPAway [9]5.5300.0930.9300.3440.59318.561
PowerPaint [55]5.0250.2000.9310.2910.66320.594
SmartEraser4.3860.0530.9390.2690.67222.029
", + "image_path": "c5259c74a583675b695084987dfb4732441ed55eefcec5e24f7c6da5073af30e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 59, + 209, + 293, + 268 + ], + "blocks": [ + { + "bbox": [ + 55, + 181, + 294, + 205 + ], + "lines": [ + { + "bbox": [ + 55, + 181, + 294, + 205 + ], + "spans": [ + { + "bbox": [ + 55, + 181, + 294, + 205 + ], + "type": "text", + "content": "Table 3. Quantitative comparison of SmartEraser and previous methods on Syn4Removal-Val." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 59, + 209, + 293, + 268 + ], + "lines": [ + { + "bbox": [ + 59, + 209, + 293, + 268 + ], + "spans": [ + { + "bbox": [ + 59, + 209, + 293, + 268 + ], + "type": "table", + "html": "
MethodImage Quality ↑ FB-Consistency ↑Removal Accuracy ↑
LaMa [39]0.100.11
CLIPAway [9]0.180.12
PowerPaint [55]0.160.12
SmartEraser0.560.65
0.58
", + "image_path": "09b87f3d7b2bda7a6f8f3e406de6bfb9dbcaf6f038406e914cb67dfeec15b7ff.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 55, + 269, + 295, + 293 + ], + "lines": [ + { + "bbox": [ + 55, + 269, + 295, + 293 + ], + "spans": [ + { + "bbox": [ + 55, + 269, + 295, + 293 + ], + "type": "text", + "content": "Table 4. Comparison of user study results in terms of image quality, foreground-background consistency, and removal accuracy." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 300, + 271, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 300, + 271, + 312 + ], + "spans": [ + { + "bbox": [ + 55, + 300, + 271, + 312 + ], + "type": "text", + "content": "users consistently favor the results from SmartEraser." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 312, + 296, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 312, + 296, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 312, + 296, + 601 + ], + "type": "text", + "content": "Real-world user cases. To further evaluate the object removal capability of SmartEraser in real-world user scenarios, we conduct a comparison with other methods (LaMa [39], CLIPAway [9], PowerPaint [55]), using masks provided by users. Figure 7 shows the comparison results. We can observe that SmartEraser exhibits a remarkable ability to accurately identify and remove target objects while preserving surrounding contextual details, particularly in scenarios involving imperfect or approximate masks. In the sample in the bottom row on the left, SmartEraser accurately identifies and removes the person indicated by the mask, while maintaining the integrity of the elephant in the background. In contrast, alternative methods tend to repaint the entire masked area, which leads to outputs that lack contextual coherence and visual realism due to the loss of details of the elephant. The upper second sample on the right provides a challenging scenario, the user-provided mask aims to remove the man whose feet are partially obscured by floating text. The previous methods fail to maintain the integrity of the floating text. However, SmartEraser accurately removes the man while preserving the floating text. This highlights SmartEraser's ability to effectively differentiate between removal targets and background content, even in cases of partial occlusion." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 610, + 154, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 610, + 154, + 623 + ], + "spans": [ + { + "bbox": [ + 55, + 610, + 154, + 623 + ], + "type": "text", + "content": "4.3. Ablation studies" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "content": "To evaluate the effectiveness of the key components in our framework, we perform ablation studies and report both quantitative and qualitative results. We progressively modify the baseline (fine-tuning SD v1.5 with the \"mask-and-inpaint\" paradigm) by adding our proposed techniques: RG for masked-region guidance, ME for mask enhancement, and VG for CLIP-based visual guidance. The results are" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 316, + 70, + 364, + 114 + ], + "blocks": [ + { + "bbox": [ + 316, + 70, + 364, + 114 + ], + "lines": [ + { + "bbox": [ + 316, + 70, + 364, + 114 + ], + "spans": [ + { + "bbox": [ + 316, + 70, + 364, + 114 + ], + "type": "image", + "image_path": "5e68ffc8c8dae8acd2a6323e074ab130e147e7574cabc0ab29005c4220885e15.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 316, + 114, + 364, + 156 + ], + "blocks": [ + { + "bbox": [ + 316, + 114, + 364, + 156 + ], + "lines": [ + { + "bbox": [ + 316, + 114, + 364, + 156 + ], + "spans": [ + { + "bbox": [ + 316, + 114, + 364, + 156 + ], + "type": "image", + "image_path": "e5f467d86aa545c11068c3862854136cfe8dbf4500e006cbf06e208af2dca5ca.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 317, + 156, + 363, + 163 + ], + "lines": [ + { + "bbox": [ + 317, + 156, + 363, + 163 + ], + "spans": [ + { + "bbox": [ + 317, + 156, + 363, + 163 + ], + "type": "text", + "content": "Mask & Image" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 170, + 553, + 192 + ], + "lines": [ + { + "bbox": [ + 313, + 170, + 553, + 192 + ], + "spans": [ + { + "bbox": [ + 313, + 170, + 553, + 192 + ], + "type": "text", + "content": "Figure 8. Qualitative ablation comparison of our method. From left to right, we progressively add each proposed component." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 364, + 70, + 410, + 156 + ], + "blocks": [ + { + "bbox": [ + 364, + 70, + 410, + 156 + ], + "lines": [ + { + "bbox": [ + 364, + 70, + 410, + 156 + ], + "spans": [ + { + "bbox": [ + 364, + 70, + 410, + 156 + ], + "type": "image", + "image_path": "fd28d4db82f884c0d529aed183aeb22307ea5e1745be12ff4c3ee8d197b6b00a.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 374, + 156, + 400, + 163 + ], + "lines": [ + { + "bbox": [ + 374, + 156, + 400, + 163 + ], + "spans": [ + { + "bbox": [ + 374, + 156, + 400, + 163 + ], + "type": "text", + "content": "Baseline" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 410, + 70, + 459, + 156 + ], + "blocks": [ + { + "bbox": [ + 410, + 70, + 459, + 156 + ], + "lines": [ + { + "bbox": [ + 410, + 70, + 459, + 156 + ], + "spans": [ + { + "bbox": [ + 410, + 70, + 459, + 156 + ], + "type": "image", + "image_path": "4d20676228b19ff6981570d2a1193dec52951511022b5847d5ef650df0a7519e.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 427, + 156, + 443, + 163 + ], + "lines": [ + { + "bbox": [ + 427, + 156, + 443, + 163 + ], + "spans": [ + { + "bbox": [ + 427, + 156, + 443, + 163 + ], + "type": "text", + "content": "+RG" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 459, + 70, + 506, + 156 + ], + "blocks": [ + { + "bbox": [ + 459, + 70, + 506, + 156 + ], + "lines": [ + { + "bbox": [ + 459, + 70, + 506, + 156 + ], + "spans": [ + { + "bbox": [ + 459, + 70, + 506, + 156 + ], + "type": "image", + "image_path": "adf50fc44ea31356e6415db0de1be2cb821d51732e5a0801a532acf543b71db3.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 475, + 156, + 490, + 163 + ], + "lines": [ + { + "bbox": [ + 475, + 156, + 490, + 163 + ], + "spans": [ + { + "bbox": [ + 475, + 156, + 490, + 163 + ], + "type": "text", + "content": "+ME" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 506, + 70, + 553, + 156 + ], + "blocks": [ + { + "bbox": [ + 506, + 70, + 553, + 156 + ], + "lines": [ + { + "bbox": [ + 506, + 70, + 553, + 156 + ], + "spans": [ + { + "bbox": [ + 506, + 70, + 553, + 156 + ], + "type": "image", + "image_path": "85f23e43d152e2de2399f16cb696b58fa079e34fe7ac026208cacefc02e3d856.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 523, + 156, + 538, + 163 + ], + "lines": [ + { + "bbox": [ + 523, + 156, + 538, + 163 + ], + "spans": [ + { + "bbox": [ + 523, + 156, + 538, + 163 + ], + "type": "text", + "content": "+VG" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "table", + "bbox": [ + 318, + 199, + 552, + 262 + ], + "blocks": [ + { + "bbox": [ + 318, + 199, + 552, + 262 + ], + "lines": [ + { + "bbox": [ + 318, + 199, + 552, + 262 + ], + "spans": [ + { + "bbox": [ + 318, + 199, + 552, + 262 + ], + "type": "table", + "html": "
MethodFID ↓CMMD ↓ReLUVE ↑LPIPS ↓SSIM ↑PSNR ↑
Baseline33.1310.2450.9080.3550.54416.535
+RG24.5260.1210.9230.3390.56217.176
+ME19.4210.1080.9310.3010.59118.917
+VG16.0300.0920.9370.2760.61619.994
", + "image_path": "982bf1c7d6d6ea4a3e94ebddcdd3ebd7243fe9bee8145a3c30fc9a83af8eab18.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "table_body" + } + ], + "index": 20 + }, + { + "bbox": [ + 331, + 265, + 537, + 275 + ], + "lines": [ + { + "bbox": [ + 331, + 265, + 537, + 275 + ], + "spans": [ + { + "bbox": [ + 331, + 265, + 537, + 275 + ], + "type": "text", + "content": "Table 5. Quantitative ablated comparison on RORD-Val." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 282, + 453, + 294 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 282, + 453, + 294 + ], + "spans": [ + { + "bbox": [ + 313, + 282, + 453, + 294 + ], + "type": "text", + "content": "displayed in Figure 8 and Table 5." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 295, + 554, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 295, + 554, + 415 + ], + "spans": [ + { + "bbox": [ + 313, + 295, + 554, + 415 + ], + "type": "text", + "content": "Effect of the Masked-Region Guidance paradigm: We observe that the baseline model often regenerates objects instead of properly removing them, as it is trained in a traditional \"inpainting\" manner. Incorporating masked-region guidance significantly improves the removal performance by reducing the risk of object regeneration. However, due to the mismatch between the mask shape used during training and the user-defined mask during testing, the model does not always perfectly remove the target objects and preserve the surrounding context." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 415, + 554, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 415, + 554, + 475 + ], + "spans": [ + { + "bbox": [ + 313, + 415, + 554, + 475 + ], + "type": "text", + "content": "Contribution of Mask Enhancement: By simulating real-world user mask shapes over object removal during training, we find that it helps reduce the gap between model training and testing using user input mask, significantly keeping the surrounding regions and further improving removal results." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 475, + 553, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 475, + 553, + 536 + ], + "spans": [ + { + "bbox": [ + 313, + 475, + 553, + 536 + ], + "type": "text", + "content": "Value of CLIP-based Visual Guidance: CLIP visual guidance aids the model in better understanding the removal targets, yielding the best performance on RORD-val across all metrics. This indicates that the visual guidance helps the model perform object removal tasks more effectively." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 549, + 388, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 549, + 388, + 562 + ], + "spans": [ + { + "bbox": [ + 313, + 549, + 388, + 562 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "content": "In conclusion, we present SmartEraser, a state-of-the-art object removal model built on the novel Masked-Region Guidance paradigm. SmartEraser outperforms existing methods by smartly identifying the target object to remove while effectively preserving the surrounding context. To facilitate research on this paradigm, we propose Syn4Removal, a large-scale, high-quality dataset containing over a million image triplets, specifically designed for object removal tasks. Through extensive experiments, we demonstrate that SmartEraser achieves superior performance in both quality and robustness compared to previous object removal methods." + } + ] + } + ], + "index": 27 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "24459" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 123 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 123 + ], + "type": "text", + "content": "[1] Omri Avrahami, Ohad Fried, and Dani Lischinski. Blended latent diffusion. ACM transactions on graphics (TOG), 42 (4):1-11, 2023. 1, 3, 6, 7, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 125, + 296, + 168 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 125, + 296, + 168 + ], + "spans": [ + { + "bbox": [ + 61, + 125, + 296, + 168 + ], + "type": "text", + "content": "[2] Connelly Barnes, Eli Shechtman, Adam Finkelstein, and Dan B Goldman. Patchmatch: A randomized correspondence algorithm for structural image editing. ACM Trans. Graph., 28(3):24, 2009. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 171, + 294, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 171, + 294, + 203 + ], + "spans": [ + { + "bbox": [ + 62, + 171, + 294, + 203 + ], + "type": "text", + "content": "[3] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. arXiv preprint arXiv:2211.09800, 2022. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 205, + 295, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 205, + 295, + 249 + ], + "spans": [ + { + "bbox": [ + 62, + 205, + 295, + 249 + ], + "type": "text", + "content": "[4] Chenjie Cao, Qiaole Dong, and Yanwei Fu. Zits++: Image inpainting by improving the incremental transformer on structural priors. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 2, 6, 7, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 251, + 296, + 305 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 251, + 296, + 305 + ], + "spans": [ + { + "bbox": [ + 62, + 251, + 296, + 305 + ], + "type": "text", + "content": "[5] Aditya Chandrasekar, Goirik Chakrabarty, Jai Bardhan, Ramya Hebbalaguppe, and Prathosh AP. Remove: A reference-free metric for object erasure. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 7901-7910, 2024. 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 307, + 294, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 307, + 294, + 350 + ], + "spans": [ + { + "bbox": [ + 62, + 307, + 294, + 350 + ], + "type": "text", + "content": "[6] Pau de Jorge, Riccardo Volpi, Puneet K. Dakania, Philip H. S. Torr, and Rogez Gregory. Placing objects in context via inpainting for out-of-distribution segmentation. In The European Conference on Computer Vision (ECCV), 2024. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 352, + 294, + 406 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 352, + 294, + 406 + ], + "spans": [ + { + "bbox": [ + 62, + 352, + 294, + 406 + ], + "type": "text", + "content": "[7] Xueqing Deng, Qihang Yu, Peng Wang, Xiaohui Shen, and Liang-Chieh Chen. Coconut: Modernizing coco segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21863-21873, 2024. 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 408, + 294, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 408, + 294, + 463 + ], + "spans": [ + { + "bbox": [ + 62, + 408, + 294, + 463 + ], + "type": "text", + "content": "[8] Qiaole Dong, Chenjie Cao, and Yanwei Fu. Incremental transformer structure enhanced image inpainting with masking positional encoding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11358-11368, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 464, + 294, + 507 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 464, + 294, + 507 + ], + "spans": [ + { + "bbox": [ + 62, + 464, + 294, + 507 + ], + "type": "text", + "content": "[9] Yigit Ekin, Ahmet Burak Yildirim, Erdem Eren Caglar, Aykut Erdem, Erkut Erdem, and Aysegul Dundar. Clipaway: Harmonizing focused embeddings for removing objects via diffusion models, 2024. 1, 3, 6, 7, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 510, + 294, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 510, + 294, + 553 + ], + "spans": [ + { + "bbox": [ + 57, + 510, + 294, + 553 + ], + "type": "text", + "content": "[10] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12873-12883, 2021. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 555, + 294, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 555, + 294, + 599 + ], + "spans": [ + { + "bbox": [ + 57, + 555, + 294, + 599 + ], + "type": "text", + "content": "[11] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 1, 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 601, + 294, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 601, + 294, + 644 + ], + "spans": [ + { + "bbox": [ + 57, + 601, + 294, + 644 + ], + "type": "text", + "content": "[12] Xiefan Guo, Hongyu Yang, and Di Huang. Image inpainting via conditional texture and structure dual generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 14134-14143, 2021. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 647, + 294, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 647, + 294, + 678 + ], + "spans": [ + { + "bbox": [ + 57, + 647, + 294, + 678 + ], + "type": "text", + "content": "[13] Kaiming He, Christoph Rhemann, Carsten Rother, Xiaou Tang, and Jian Sun. A global sampling method for alpha matting. In CVPR 2011, pages 2049-2056, 2011. 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 681, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 681, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 681, + 294, + 713 + ], + "type": "text", + "content": "[14] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilib-" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 74, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 335, + 74, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 74, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 74, + 553, + 95 + ], + "type": "text", + "content": "rium. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2017. 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 97, + 553, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 97, + 553, + 129 + ], + "spans": [ + { + "bbox": [ + 317, + 97, + 553, + 129 + ], + "type": "text", + "content": "[15] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 131, + 553, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 131, + 553, + 163 + ], + "spans": [ + { + "bbox": [ + 317, + 131, + 553, + 163 + ], + "type": "text", + "content": "[16] Alain Hore and Djemel Ziou. Image quality metrics: Psnr vs. sSIM. In 2010 20th international conference on pattern recognition, pages 2366-2369. IEEE, 2010. 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 165, + 553, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 165, + 553, + 229 + ], + "spans": [ + { + "bbox": [ + 317, + 165, + 553, + 229 + ], + "type": "text", + "content": "[17] Sadeep Jayasumana, Srikumar Ramalingam, Andreas Veit, Daniel Glasner, Ayan Chakrabarti, and Sanjiv Kumar. Rethinking fid: Towards a better evaluation metric for image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9307-9315, 2024. 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 232, + 553, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 232, + 553, + 263 + ], + "spans": [ + { + "bbox": [ + 317, + 232, + 553, + 263 + ], + "type": "text", + "content": "[18] Xuan Ju, Xian Liu, Xintao Wang, Yuxuan Bian, Ying Shan, and Qiang Xu. Brushnet: A plug-and-play image inpainting model with decomposed dual-branch diffusion, 2024. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 266, + 553, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 266, + 553, + 297 + ], + "spans": [ + { + "bbox": [ + 316, + 266, + 553, + 297 + ], + "type": "text", + "content": "[19] Tero Karras. Progressive growing of gans for improved quality, stability, and variation. arXiv preprint arXiv:1710.10196, 2017. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 300, + 553, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 300, + 553, + 342 + ], + "spans": [ + { + "bbox": [ + 317, + 300, + 553, + 342 + ], + "type": "text", + "content": "[20] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 345, + 553, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 345, + 553, + 399 + ], + "spans": [ + { + "bbox": [ + 317, + 345, + 553, + 399 + ], + "type": "text", + "content": "[21] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 401, + 553, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 401, + 553, + 475 + ], + "spans": [ + { + "bbox": [ + 317, + 401, + 553, + 475 + ], + "type": "text", + "content": "[22] Alina Kuznetsova, Hassan Rom, Neil Alldrin, Jasper Uijlings, Ivan Krasin, Jordi Pont-Tuset, Shahab Kamali, Stefan Popov, Matteo Malloci, Alexander Kolesnikov, et al. The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale. International journal of computer vision, 128(7):1956-1981, 2020. 4, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 479, + 553, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 479, + 553, + 521 + ], + "spans": [ + { + "bbox": [ + 317, + 479, + 553, + 521 + ], + "type": "text", + "content": "[23] Jens Lehmann, Daniel Gerber, Mohamed Morsey, and Axel-Cyrille Ngonga Ngomo. Defacto-deep fact validation. In International semantic web conference, pages 312-327. Springer, 2012. 2, 3, 4, 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 523, + 553, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 523, + 553, + 578 + ], + "spans": [ + { + "bbox": [ + 317, + 523, + 553, + 578 + ], + "type": "text", + "content": "[24] Wenbo Li, Zhe Lin, Kun Zhou, Lu Qi, Yi Wang, and Jiaya Jia. Mat: Mask-aware transformer for large hole image inpainting. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10758-10768, 2022. 2, 6, 7, 8" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 580, + 553, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 580, + 553, + 612 + ], + "spans": [ + { + "bbox": [ + 317, + 580, + 553, + 612 + ], + "type": "text", + "content": "[25] Xiaoguang Li, Qing Guo, Di Lin, Ping Li, Wei Feng, and Song Wnag. Misf: Multi-level interactive siamese filtering for high-fidelity image inpainting. CVPR, 2022. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 614, + 553, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 614, + 553, + 678 + ], + "spans": [ + { + "bbox": [ + 317, + 614, + 553, + 678 + ], + "type": "text", + "content": "[26] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 6, 7" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 317, + 681, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 681, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 681, + 553, + 713 + ], + "type": "text", + "content": "[27] Qiankun Liu, Zhentao Tan, Dongdong Chen, Qi Chu, Xiyang Dai, Yinpeng Chen, Mengchen Liu, Lu Yuan, and Nenghai Yu. Reduce information loss in transformers for pluralistic" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24460" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 106 + ], + "type": "text", + "content": "image inpainting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11347-11357, 2022. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 106, + 296, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 106, + 296, + 148 + ], + "spans": [ + { + "bbox": [ + 56, + 106, + 296, + 148 + ], + "type": "text", + "content": "[28] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaou Tang. Deep learning face attributes in the wild. In Proceedings of International Conference on Computer Vision (ICCV), 2015. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 150, + 294, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 150, + 294, + 172 + ], + "spans": [ + { + "bbox": [ + 56, + 150, + 294, + 172 + ], + "type": "text", + "content": "[29] I Loshchilov. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 173, + 294, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 173, + 294, + 228 + ], + "spans": [ + { + "bbox": [ + 56, + 173, + 294, + 228 + ], + "type": "text", + "content": "[30] Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11461-11471, 2022. 1, 3, 6, 7, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 57, + 228, + 294, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 228, + 294, + 249 + ], + "spans": [ + { + "bbox": [ + 57, + 228, + 294, + 249 + ], + "type": "text", + "content": "[31] Mehdi Mirza. Conditional generative adversarial nets. arXiv preprint arXiv:1411.1784, 2014. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 250, + 294, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 250, + 294, + 304 + ], + "spans": [ + { + "bbox": [ + 56, + 250, + 294, + 304 + ], + "type": "text", + "content": "[32] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 304, + 294, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 304, + 294, + 370 + ], + "spans": [ + { + "bbox": [ + 56, + 304, + 294, + 370 + ], + "type": "text", + "content": "[33] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 371, + 294, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 371, + 294, + 415 + ], + "spans": [ + { + "bbox": [ + 56, + 371, + 294, + 415 + ], + "type": "text", + "content": "[34] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 415, + 295, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 415, + 295, + 480 + ], + "spans": [ + { + "bbox": [ + 56, + 415, + 295, + 480 + ], + "type": "text", + "content": "[35] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684-10695, 2022. 1, 3, 6, 7, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 481, + 294, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 481, + 294, + 514 + ], + "spans": [ + { + "bbox": [ + 56, + 481, + 294, + 514 + ], + "type": "text", + "content": "[36] Min-Cheol Sagong, Yoon-Jae Yeo, Seung-Won Jung, and Sung-Jea Ko. Rord: A real-world object removal dataset. In BMVC, page 542, 2022. 2, 3, 4, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 514, + 294, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 514, + 294, + 548 + ], + "spans": [ + { + "bbox": [ + 56, + 514, + 294, + 548 + ], + "type": "text", + "content": "[37] Yang Song, Prafulla Dhariwal, Mark Chen, and Ilya Sutskever. Consistency models. arXiv preprint arXiv:2303.01469, 2023. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 548, + 294, + 603 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 548, + 294, + 603 + ], + "spans": [ + { + "bbox": [ + 56, + 548, + 294, + 603 + ], + "type": "text", + "content": "[38] Zeyi Sun, Ye Fang, Tong Wu, Pan Zhang, Yuhang Zang, Shu Kong, Yuanjun Xiong, Dahua Lin, and Jiaqi Wang. Alphaclip: A clip model focusing on wherever you want. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13019-13029, 2024. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 603, + 294, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 603, + 294, + 679 + ], + "spans": [ + { + "bbox": [ + 56, + 603, + 294, + 679 + ], + "type": "text", + "content": "[39] Roman Suvorov, Elizaveta Logacheva, Anton Mashikhin, Anastasia Remizova, Armenii Ashukha, Aleksei Silvestrov, Naejin Kong, Harshith Goka, Kiwoong Park, and Victor Lempitsky. Resolution-robust large mask inpainting with fourier convolutions. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 2149–2159, 2022. 1, 2, 3, 6, 7, 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 680, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 680, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 680, + 294, + 713 + ], + "type": "text", + "content": "[40] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017. 1" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 72, + 553, + 714 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 315, + 72, + 553, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 72, + 553, + 126 + ], + "spans": [ + { + "bbox": [ + 315, + 72, + 553, + 126 + ], + "type": "text", + "content": "[41] Ziyu Wan, Jingbo Zhang, Dongdong Chen, and Jing Liao. High-fidelity pluralistic image completion with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 4692-4701, 2021. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 127, + 553, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 127, + 553, + 184 + ], + "spans": [ + { + "bbox": [ + 316, + 127, + 553, + 184 + ], + "type": "text", + "content": "[42] Wentao Wang, Jianfu Zhang, Li Niu, Haoyu Ling, Xue Yang, and Liqing Zhang. Parallel multi-resolution fusion network for image inpainting. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 14559-14568, 2021. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 185, + 553, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 185, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 316, + 185, + 553, + 239 + ], + "type": "text", + "content": "[43] Wentao Wang, Li Niu, Jianfu Zhang, Xue Yang, and Liqing Zhang. Dual-path image inpainting with auxiliary gan inversion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11421-11430, 2022. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 240, + 553, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 240, + 553, + 284 + ], + "spans": [ + { + "bbox": [ + 316, + 240, + 553, + 284 + ], + "type": "text", + "content": "[44] Zhou Wang, A.C. Bovik, H.R. Sheikh, and E.P. Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing, 13(4): 600-612, 2004. 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 285, + 553, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 285, + 553, + 331 + ], + "spans": [ + { + "bbox": [ + 316, + 285, + 553, + 331 + ], + "type": "text", + "content": "[45] Daniel Winter, Matan Cohen, Shlomi Fruchter, Yael Pritch, Alex Rav-Acha, and Yedid Hoshen. Objectdrop: Bootstrapping counterfactuals for photorealistic object removal and insertion. arXiv preprint arXiv:2403.18818, 2024. 2, 3, 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 331, + 553, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 331, + 553, + 374 + ], + "spans": [ + { + "bbox": [ + 316, + 331, + 553, + 374 + ], + "type": "text", + "content": "[46] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721, 2023. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 375, + 553, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 375, + 553, + 410 + ], + "spans": [ + { + "bbox": [ + 316, + 375, + 553, + 410 + ], + "type": "text", + "content": "[47] Yifu Chen, Jingwen Chen, Yingwei Pan, Yehao Li, Ting Yao, Zhineng Chen and Tao Mei. Improving text-guided object inpainting with semantic pre-inpainting. ECCV, 2024. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 411, + 553, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 411, + 553, + 454 + ], + "spans": [ + { + "bbox": [ + 316, + 411, + 553, + 454 + ], + "type": "text", + "content": "[48] Ahmet Burak Yildirim, Vedat Baday, Erkut Erdem, Aykut Erdem, and Aysegul Dundar. Inst-inpaint: Instructing to remove objects with diffusion models. arXiv preprint arXiv:2304.03246, 2023. 2, 3, 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 456, + 553, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 456, + 553, + 510 + ], + "spans": [ + { + "bbox": [ + 316, + 456, + 553, + 510 + ], + "type": "text", + "content": "[49] Jiahui Yu, Zhe Lin, Jimei Yang, Xiaohui Shen, Xin Lu, and Thomas S Huang. Generative image inpainting with contextual attention. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5505-5514, 2018. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 510, + 553, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 510, + 553, + 567 + ], + "spans": [ + { + "bbox": [ + 316, + 510, + 553, + 567 + ], + "type": "text", + "content": "[50] Yingchen Yu, Fangneng Zhan, Shijian Lu, Jianxiong Pan, Feiying Ma, Xuansong Xie, and Chunyan Miao. Wavefill: A wavelet-based generation network for image inpainting. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 14114-14123, 2021. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 568, + 553, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 568, + 553, + 622 + ], + "spans": [ + { + "bbox": [ + 316, + 568, + 553, + 622 + ], + "type": "text", + "content": "[51] Yu Zeng, Zhe Lin, Huchuan Lu, and Vishal M. Patel. Crfill: Generative image inpainting with auxiliary contextual reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 14164-14173, 2021. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 624, + 553, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 553, + 679 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 553, + 679 + ], + "type": "text", + "content": "[52] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 680, + 553, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 680, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 316, + 680, + 553, + 714 + ], + "type": "text", + "content": "[53] Hanqing Zhao, Dianmo Sheng, Jianmin Bao, Dongdong Chen, Dong Chen, Fang Wen, Lu Yuan, Ce Liu, Wenbo Zhou, Qi Chu, et al. X-paste: Revisiting scalable copy" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24461" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 195 + ], + "type": "list", + "angle": 0, + "index": 3, + "blocks": [ + { + "bbox": [ + 75, + 72, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 72, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 75, + 72, + 294, + 106 + ], + "type": "text", + "content": "paste for instance segmentation using clip and stablediffusion. In International Conference on Machine Learning, pages 42098-42109. PMLR, 2023. 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 295, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 295, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 295, + 150 + ], + "type": "text", + "content": "[54] Bolei Zhou, Agata Lapedriza, Aditya Khosla, Aude Oliva, and Antonio Torralba. Places: A 10 million image database for scene recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2017. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 294, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 294, + 195 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 294, + 195 + ], + "type": "text", + "content": "[55] Junhao Zhuang, Yanhong Zeng, Wenran Liu, Chun Yuan, and Kai Chen. A task is worth one word: Learning with task prompts for high-quality versatile image inpainting, 2023. 1, 3, 6, 7, 8" + } + ] + } + ], + "index": 2 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "24462" + } + ] + } + ], + "index": 4 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/SnapGen-V_ Generating a Five-Second Video within Five Seconds on a Mobile Device/2089e6e3-1b02-4449-b418-985fa5c0c175_content_list.json b/2025/SnapGen-V_ Generating a Five-Second Video within Five Seconds on a Mobile Device/2089e6e3-1b02-4449-b418-985fa5c0c175_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..9744e469e194c0e6d1bc99bca28a6151921ed4e5 --- /dev/null +++ b/2025/SnapGen-V_ Generating a Five-Second Video within Five Seconds on a Mobile Device/2089e6e3-1b02-4449-b418-985fa5c0c175_content_list.json @@ -0,0 +1,1427 @@ +[ + { + "type": "text", + "text": "SnapGen-V: Generating a Five-Second Video within Five Seconds on a Mobile Device", + "text_level": 1, + "bbox": [ + 106, + 128, + 890, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yushu Wu $^{1,2\\dagger}$ Zhixing Zhang $^{1,3\\dagger}$ Yanyu Li $^{1\\dagger\\dagger}$ Yanwu Xu $^{1}$ Anil Kag $^{1}$ Yang Sui $^{1}$ Huseyin Coskun $^{1}$ Ke Ma $^{1}$ Aleksei Lebedev $^{1}$ Ju Hu $^{1}$ Dimitris N. Metaxas $^{3}$ Yanzhi Wang $^{2}$ Sergey Tulyakov $^{1}$ Jian Ren $^{1\\dagger}$ Snap Inc. Northeastern University Rutgers University", + "bbox": [ + 153, + 200, + 841, + 276 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f3caee03fc45eb582f8e3700eaf5b65bbe298c853299bfca7e4c41c321b726a5.jpg", + "image_caption": [ + "Figure 1. Example generation results from our mobile text-to-video model. Our model can generate high-quality and motion consistent 5-second videos on a mobile device (e.g., iPhone 16 Pro Max) within 5 seconds." + ], + "image_footnote": [], + "bbox": [ + 93, + 309, + 903, + 431 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 482, + 326, + 498 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We have witnessed the unprecedented success of diffusion-based video generation over the past year. Recently proposed models from the community have wielded the power to generate cinematic and high-resolution videos with smooth motions from arbitrary input prompts. However, as a supertask of image generation, video generation models require more computation and are thus hosted mostly on cloud servers, limiting broader adoption among content creators. In this work, we propose a comprehensive acceleration framework to bring the power of the largescale video diffusion model to the hands of edge users. From the network architecture scope, we initialize from a compact image backbone and search out the design and arrangement of temporal layers to maximize hardware efficiency. In addition, we propose a dedicated adversarial fine-tuning algorithm for our efficient model and reduce the denoising steps to 4. Our model, with only $0.6B$ parameters, can generate a 5-second video on an iPhone 16 PM within 5 seconds. Compared to server-side models that take minutes on powerful GPUs to generate a single video, we accelerate the generation by magnitudes while delivering on-par quality. Project page at https://snap-research.github.io/snapgen-v/.", + "bbox": [ + 88, + 515, + 485, + 849 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 514, + 482, + 643, + 498 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, the rapid advancement of video diffusion models [4] inspires revolutions in content creation. With the emergence of video models from industry [41] and research community [67], content creators can animate a static image [3] or generate cinematic videos from arbitrary prompts [55, 67, 78]. Video diffusion models also enable downstream applications like video editing [12, 20, 30], novel view synthesis [25, 57], and multi-modal generation [56].", + "bbox": [ + 511, + 507, + 906, + 628 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite the success in generation quality, the huge number of parameters and slow generation speed prohibit the wide deployment of video diffusion models. For instance, CogVideoX-5B [67] generates a video (49 frames at 8 fps, $720 \\times 480$ resolution) in 5 minutes with 50 inference steps on an NVIDIA A100 GPU. Compared to text-to-image diffusion models [45], video diffusion models require extra parameters to model sophisticated motions [3, 13, 62]. In addition, video data usually incurs higher spatial-temporal resolution for UNet denoisers [3, 13], or equivalently more tokens for DiT models [40], which adds up to the computation complexity. Recent works explore efficient model architectures and attention mechanisms for image diffusion models [8, 64, 70]. However, there is little effort in the literature dedicated to accelerating and deploying video models at scale, especially for mobile devices.", + "bbox": [ + 511, + 628, + 908, + 869 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this work, we systematically investigate the redundancies in video diffusion models and propose a mobile", + "bbox": [ + 511, + 869, + 906, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "Equal contribution", + "bbox": [ + 107, + 875, + 215, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$\\ddagger$ Corresponding authors", + "bbox": [ + 109, + 887, + 238, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "2479", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "acceleration framework. First, we obtain an efficient spatial backbone by following prior works [11, 29] to prune a pre-trained text-to-image diffusion model [45]. The pruned model achieves $2.5 \\times$ size compression and more than $10 \\times$ speedup compared to Stable Diffusion v1.5 [45], while maintaining comparable generative quality. Starting from a pre-trained image model offers two key benefits: (i) it eliminates the need for costly large-scale pre-training, and (ii) with a compact image model, we can significantly narrow the search space in subsequent stages, focusing only on optimizing temporal layers, and thereby accelerating the discovery of the final model architecture.", + "bbox": [ + 88, + 90, + 480, + 271 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Even with the efficient image backbone, applying previous temporal inflation methods [3, 4, 13] still results in tremendous computation cost and encounters out-of-memory issues on mobile. Thus, our second stage is to systematically investigate different types of temporal layers and perform a latency-memory joint search to determine the spatial-temporal architecture for efficient mobile deployment. Unlike previous methods [3, 4, 13] that typically rely on a specific type of temporal modeling layer, we investigate all possible designs, including temporal attention, spatial-temporal full attention (3D attention), temporal cross attention, and temporal convolutions (Conv3D). Besides, our search space includes the position (resolution) to apply these temporal layers, and number of layers to use (in Sec. 3.2). We profile the computation, memory footprint, and on-device latency of architecture candidates, and perform evolutionary search to discover the architecture with Pareto optimality of speed and quality. The searched network is only 0.6B in size, and can generate a 5-second video clip on iPhone 16 PM without hitting memory bound. In contrast, all prior video diffusion models fail to run on mobile, even for the smallest open-sourced ones like the 16-frameAnimateDiff [13] and 14-frame SVD [3] (Tab. 1).", + "bbox": [ + 93, + 275, + 485, + 619 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Finally, to further speed up generation on mobile, we distill our efficient video diffusion model with a tailored adversarial fine-tuning method capable of image-video mixed training. We reduce the number of denoising steps from 25 to 4, and eliminate classifier-free guidance [16], leading to more than $12 \\times$ speedup without performance degradation. As in Tab. 1, our mobile speed is faster than most GPU-deployed (e.g., on A100) counterparts.", + "bbox": [ + 89, + 623, + 482, + 744 + ], + "page_idx": 1 + }, + { + "type": "table", + "img_path": "images/5ffb1fea6612f62db4b35d1db855b8248d091a2c4a4e545437e4b1f1ad483102.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelTypeStepsParams (B)A100 (s)iPhone (s)Vbench (↑)
OpenSora-v1.2DiT301.231.00✘79.76
CogVideoX-2BDiT501.654.09✘80.91
AnimateDiff-V2UNet251.29.04✘80.27
AnimateDiffLCMUNet41.21.77✘79.42
OursUNet40.60.474.1281.14
", + "bbox": [ + 94, + 761, + 478, + 843 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Table 1. Comparison of size (number of parameters), speed (tested on NVIDIA A100 and iPhone 16 Pro Max), and performance (on VBench [19]) for various models.", + "bbox": [ + 89, + 854, + 482, + 896 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "With the proposed framework, we successfully deploy our 0.6B text-to-video model on an iPhone 16 Pro Max, achieving the generation of a 5-second video clip within 5 seconds. This work represents not only the very first mobile deployment attempt of the video diffusion model, but also demonstrates its real-time potential1. Our contributions are summarized as follows:", + "bbox": [ + 511, + 90, + 903, + 195 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Through image-video joint training, spatial and temporal architecture design, and mobile-driven latency-memory joint architecture search, we develop a comprehensive mobile acceleration framework for the text-to-video diffusion model.", + "- We propose an adversarial fine-tuning technique tailored for video diffusion models. Despite the already compact nature of our mobile denoiser, we further distill it to 4 denoising steps with superior quality.", + "- Our work is the very first one to show the possibility of real-time text-to-video generation on mobile devices, unlocking the possibility of deploying applications of video diffusion models at scale." + ], + "bbox": [ + 513, + 196, + 903, + 391 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 406, + 653, + 421 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Video Diffusion Models. Denoising Diffusion Probabilistic Models [17] is the trending paradigm for building video diffusion models, demonstrating photorealistic quality and generic generation capabilities. Pioneer works often start from a pre-trained text-to-image diffusion model [45] and insert temporal layers to model motions along frame sequence [3, 4, 13, 62, 75]. In addition, training-free noise tuning techniques are proposed to ease the alignment between frames [23, 34, 43, 63, 73]. Later, with the emergence of large-scale, high-quality video datasets [7, 38] and Transformer backbones [40], subsequent works curate their own dataset and build large video diffusion models with exceptional quality, such as the open-sourced CogVideoX [67], Mochi 1 [55], PyramidalFlow [21], Allegro [78], and close-sourced ones including Hailuo [37], Runway Gen 3 Alpha [46], Kling [24], Luma Dream Machine [1], Pika 1.5 [2], Sora [39], and MovieGen [41]. Remarkably, the open-sourced projects Open-Sora [77] and Open-Sora-Plan [26] provide the community with reliable implementations to replicate large-scale video diffusion models.", + "bbox": [ + 511, + 431, + 906, + 733 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Dividing by task type, video diffusion models can be categorized into text-to-video generation [5, 6, 13, 15, 18, 21, 27, 28, 42, 55, 56, 67, 78], image-to-video generation [3, 79], or specific motion controls [14, 44, 61, 71, 76]. Though some work [8, 29, 70, 73] aim to improve the efficiency of diffusion models, the acceleration for mobile deployment of video diffusion models is still in absent. Popular video models [21, 55, 67] can only run on server-level", + "bbox": [ + 511, + 734, + 905, + 854 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "This work performs conventional T2V generation, generating an entire video at once. However, a more suitable real-time capability would involve streamlined, continuous video generation, which we leave for future work.", + "bbox": [ + 511, + 862, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2480", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Step Distillation brings almost linear generation speedup for diffusion models. Early work [29, 47] progressively distill a student network to predict a further ODE location with teacher guidance, resulting in fewer inference steps, while Consistency Models [52, 53] and Rectified Flow [33] refine the prediction objective to clean data or global velocity to achieve fewer-step inference. Later works [50, 51, 65] further incorporate adversarial loss to distill a single-step student, and enhance multi-step results as well.", + "bbox": [ + 89, + 90, + 480, + 241 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Despite extensive research in image diffusion models [9, 22, 36, 59, 66, 68, 69], step distillation for video diffusion model is under-explored. One type of work applies consistency distillation to generate videos in 4 steps [58, 60, 72]. Another trend adopts adversarial distillation to achieve few-step (1-2) generation [32, 35, 74]. However, these methods distill pre-trained large models with enough redundancy along the trajectory, while we find them not applicable to our efficient model and yield inferior performance.", + "bbox": [ + 89, + 242, + 483, + 378 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 390, + 181, + 404 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our objective is to achieve high-fidelity and temporally consistent video generation on mobile devices. However, current text-to-video diffusion models face two key challenges in reaching this goal: (a) the memory and computation requirement is beyond the capability of even the most powerful mobile chips, i.e. iPhone A18 Pro, and (b) denoising with dozens of steps to generate a single output further slows down the process. To address these challenges, we propose a three-stage framework to accelerate video diffusion models on the mobile platform. First, we prune from a pre-trained text-to-image diffusion model to obtain an efficient spatial backbone. Second, we inflate the spatial backbone with a novel combination of temporal modules which are searched out with our mobile-oriented metrics. Finally, through adversarial training, our efficient model attains the capability to generate high-quality videos in only 4 steps.", + "bbox": [ + 89, + 415, + 483, + 657 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminaries", + "text_level": 1, + "bbox": [ + 89, + 665, + 230, + 679 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Following [77], we employ a spatial-temporal VAE to compress image and video data into the latent space. Given video or image data $\\mathbf{v} \\in \\mathbb{R}^{n \\times 3 \\times H \\times W}$ , where $n$ is the number of frames with height $H$ and width $W$ , the spatial-temporal encoder, $\\mathbf{E}$ , maps the data to a latent space. The encoded frames are represented as $\\mathbf{x}_0 = \\mathbf{E}(\\mathbf{v})$ , resulting in $\\mathbf{x}_0 \\in \\mathbb{R}^{\\tilde{n} \\times 4 \\times \\tilde{H} \\times \\tilde{W}}$ . Here, $\\mathbf{x}_0 \\sim p_{data}(\\mathbf{x}_0)$ is a 4-channel latent, with a temporal compression of $\\tilde{n} = n/4$ , and spatial compression as $\\tilde{H} = H/8$ , $\\tilde{W} = W/8$ .", + "bbox": [ + 89, + 686, + 482, + 824 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We follow Rectified Flow [59] to train our latent diffusion model. According to the flow-matching-based diffusion form, the intermediate noisy state $\\mathbf{x}_t$ is defined as:", + "bbox": [ + 89, + 824, + 483, + 869 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {x} _ {t} = (1 - t) \\mathbf {x} _ {0} + t \\epsilon , \\text {w h e r e} \\epsilon \\sim \\mathcal {N} (0, I), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 143, + 883, + 482, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "which is a linear interpolation between the data distribution and a standard normal distribution. The model aims to learn a vector field $v_{\\theta}(t, \\mathbf{x}_t)$ using the Conditional Flow Matching objective, i.e.,", + "bbox": [ + 511, + 90, + 903, + 151 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathbb {E} _ {t, \\epsilon , \\mathbf {x} _ {0}} \\| v _ {\\theta} (t, \\mathbf {x} _ {t}) - u _ {t} (\\mathbf {x} _ {t} | \\mathbf {x} _ {0}) \\| _ {2} ^ {2}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 160, + 903, + 179 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $u_{t}(\\mathbf{x}_{t}|\\mathbf{x}_{0}) = \\epsilon -\\mathbf{x}_{0}$ . Following [10], during training, we sample $t$ from a logit-normal distribution, i.e.,", + "bbox": [ + 511, + 188, + 903, + 219 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\pi (t; m, s) = \\frac {1}{s \\sqrt {2 \\pi}} \\frac {1}{t (1 - t)} \\exp \\left(- \\frac {\\left(\\operatorname {l o g i t} (t) - m\\right) ^ {2}}{2 s ^ {2}}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 516, + 228, + 903, + 280 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathrm{logit}(t) = \\log \\frac{t}{1 - t}$ , $m$ and $s$ are the location parameter and scale parameter, respectively.", + "bbox": [ + 511, + 280, + 903, + 311 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Hardware Efficient Model Design", + "text_level": 1, + "bbox": [ + 513, + 319, + 807, + 335 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Spatial Backbone. We follow [11, 29] to first prune an efficient text-to-image model as the spatial backbone. Specifically, we start from Stable Diffusion v1.5 [45], and borrow the knowledge from prior arts [29] to remove the most mobile-unfriendly attentions. We then prune the network depth and width following [11] and achieve $\\times 2.5$ size reduction and more than $10\\times$ speedup on mobile devices. We include qualitative visualizations of our image model in the supplementary material. Note that we use a UNet denoiser [17], leaving the exploration of DiT [40] to future work. The hierarchical structure of the UNet denoiser forms a good search space to achieve mobile efficiency, while the computation complexity of DiT grows quadratically with the number of tokens (generation resolution), making it challenging for mobile deployment.", + "bbox": [ + 511, + 340, + 906, + 568 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Temporal Layer Design. Current latent video diffusion models typically adopt temporal self-attention [13], cross-attentions [77], and convolutions [3] to model temporal dependencies. CogVideoX [67] demonstrates significant performance gain by using full 3D-Attention, at the cost of more computations and memory consumption. In this work, we enumerate and investigate all types of temporal modeling methods, including Conv1D, Conv3D, SelfAttention1D, SelfAttention3D, CrossAttention1D, and CrossAttention3D, and profile their complexity in Fig. 3. For instance, SelfAttention1D only models temporal dependency on a single coordinate, while SelfAttention3D models global dependencies and has the potential to deliver much stronger performance. However, the computation complexity of SelfAttention3D grows quadratically with respect to $\\tilde{t} \\times \\tilde{H} \\times \\tilde{W}$ , while SelfAttention1D is linear with respect to $\\tilde{H}$ and $\\tilde{W}$ , which makes SelfAttention3D much more costly at higher resolutions. On the other hand, the computation of CrossAttentionND is determined by both spatial-temporal resolution and the number of tokens from the text encoder. Conv1D and 3D are locality alternatives for SelfAttention1D and 3D, respectively. Though the computation complexity and", + "bbox": [ + 511, + 569, + 908, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "2481", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/f9b1590de9a48f2a35725c8fa2eba8c41862dd629abd454a2a9c3b56b288ea4d.jpg", + "image_caption": [ + "Figure 2. Framework Overview. In the Latency and Memory Guided Architecture Search, we freeze the pretrained efficient spatial layers and conduct evolutionary search over temporal layer based on the memory and latency constraint. During the Adversarial Fine-tuning stage, we initialize the discriminator with the weights from the text-to-video model trained in the first stage. The discriminator employs the encoder of the UNet as its backbone, which remains frozen. We add spatial-temporal discriminator heads after each downsampling block, updating only these heads during training. Following prior works [48, 49, 74], each head conditions on pooled text embeddings c projected via a linear layer. Input features are first reshaped to merge the temporal and batch axes for processing through a 2D ResBlock, and then reshaped again to merge spatial dimensions before the temporal self-attention block." + ], + "image_footnote": [], + "bbox": [ + 117, + 85, + 908, + 262 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/a8da950f93c5dd6710c284a56f977b9163542a38c96c9a2137fe241a50e12324.jpg", + "image_caption": [ + "Figure 3. Computation Complexity and Memory Consumption Analysis. The computation complexity and memory consumption of different temporal layer for various input size. The temporal dimension is fixed to 12 for simplicity." + ], + "image_footnote": [], + "bbox": [ + 89, + 398, + 480, + 518 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "memory footprint for each design candidate can be easily profiled, as in Fig. 3, it is still crucial and challenging to build a spatial-temporal network with optimized arrangements of these operators. We propose to perform a latency-memory joint architecture search to determine which, where, and the number of temporal layers to use for our efficient video diffusion model on mobile, as follows.", + "bbox": [ + 89, + 598, + 482, + 702 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Latency and Memory Guided Architecture Search.", + "text_level": 1, + "bbox": [ + 89, + 704, + 467, + 718 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Prior to searching the architecture, we construct a look-up table containing the inference latency and the memory footprint of different temporal layers. For each candidate 1D or 3D operator (i.e., ConvND, SelfAttentionND, CrossAttentionND), we benchmark the latency and memory consumption under different spatial-temporal resolutions on hardware. We then clean the search space by eliminating OOM states. Then we perform evolutionary search to obtain the temporal design with Pareto optimality. The architecture candidate is trained on precomputed video latents for $20K$ iterations with the spatial backbone frozen, and is evaluated on VBench [19] to obtain the scores as the quality metric.", + "bbox": [ + 88, + 719, + 482, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We include the detailed search algorithm, action space, and total search time in the supplementary material.", + "bbox": [ + 511, + 402, + 903, + 431 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Image-Video Joint Training. Upon the finalized model architecture, we perform image-video joint training under various clip lengths and aspect ratios with all parameters updated for another $100K$ iterations. After the joint training, our efficient model is capable of generating videos with various lengths and aspect ratios under a conventional recipe, i.e., 25 steps with classifier-free guidance.", + "bbox": [ + 511, + 431, + 905, + 537 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "VAE Decoder Compression. We use the spatial temporal-decoupled VAE from OpenSora [77], which has 4 latent channels, $8 \\times 8$ spatial compression and $4 \\times$ temporal compression. To decode a 17-frame video clip on mobile, the original [77] temporal decoder takes $23,100~\\mathrm{ms}$ , and the spatial decoder takes $4100~\\mathrm{ms}$ , which we found to be a bottleneck for the generation speed. To increase the speed, we focus on the decoder only. We freeze the VAE encoders and prune the temporal and spatial decoder on our video and image dataset, respectively. Our efficient temporal decoder runs at $210~\\mathrm{ms}$ and spatial decoder at $330~\\mathrm{ms}$ to decode a 17-frame video clip, reaching $50 \\times$ speedup with negligible loss in quality. Further details about VAE compression are included in the supplementary material.", + "bbox": [ + 511, + 537, + 906, + 750 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Latent Adversarial Fine-tuning", + "text_level": 1, + "bbox": [ + 511, + 757, + 790, + 773 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our training procedure involves two networks: a generator $\\mathcal{G}_{\\theta}$ and a discriminator $\\mathcal{D}\\phi$ . Similar to prior work [51, 74], we initialize our generator with pre-trained diffusion model weights $\\theta$ , while the discriminator is also partially initialized from $\\theta$ . Specifically, the backbone of the discriminator adopts the same architecture and weights as the pre-trained UNet encoder, with these backbone parameters remaining frozen during training. Additionally, we enhance the dis", + "bbox": [ + 511, + 780, + 905, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "2482", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "criminator with spatial-temporal discriminator heads added after each backbone block, with only these head parameters being updated in the discriminator training phase. As illustrated on the right in Fig. 2, each discriminator head consists of a spatial ResBlock and a temporal self-attention block. This design allows our discriminator to effectively handle both image and video data during fine-tuning. We analyze the impact of joint image-video fine-tuning in Sec. 4.3.", + "bbox": [ + 89, + 90, + 480, + 210 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For a real data sample $\\mathbf{x}_0$ , a noisy data point $\\mathbf{x}_t$ is generated through a forward diffusion process, as described in Eq. (1). We set intermediate timesteps as $0 < T_k < \\dots < T_1 = 1.0$ and sample $t$ from these timesteps, where $k$ is the number of timesteps selected for generator training (set to $k = 4$ in practice). The generator, then, predicts the velocity at $\\mathbf{x}_t$ as $\\mathcal{G}_{\\theta}(\\mathbf{x}_t,t)$ .", + "bbox": [ + 89, + 212, + 480, + 316 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To train the discriminator, we first sample a target timestep $t'$ from a logit-normal distribution, as shown in Eq. (3). Using the forward process in Eq. (1), we obtain the real sample $\\mathbf{x}_{t'} = (1 - t')\\mathbf{x}_0 + t'\\epsilon$ . The fake sample, $\\hat{\\mathbf{x}}_{t'}$ , is computed as $\\hat{\\mathbf{x}}_{t'} = \\mathbf{x}_t + (t' - t)\\cdot \\mathcal{G}_\\theta (t,\\mathbf{x}_t)$ , as shown in Fig. 4. Following established approaches [48-50, 74], we employ hinge loss [31] as the adversarial objective to enhance performance. The discriminator's goal is to differentiate between real and fake samples by minimizing:", + "bbox": [ + 89, + 316, + 480, + 454 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {\\mathrm {a d v}} ^ {\\mathcal {D}} = \\mathbb {E} _ {t ^ {\\prime}, x _ {0}} [ \\max (0, 1 + \\mathcal {D} _ {\\phi} (\\mathbf {x} _ {t ^ {\\prime}}, t ^ {\\prime})) ] \\tag {4} \\\\ + \\mathbb {E} _ {t, t ^ {\\prime}, x _ {0}} \\left[ \\max \\left(0, 1 - \\mathcal {D} _ {\\phi} \\left(\\hat {\\mathbf {x}} _ {t ^ {\\prime}}, t ^ {\\prime}\\right)\\right) \\right], \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 145, + 464, + 480, + 501 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The adversarial objective for the generator is defined as:", + "bbox": [ + 109, + 512, + 480, + 527 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {a d v}} ^ {\\mathcal {G}} = \\mathbb {E} _ {t, t ^ {\\prime}, x _ {0}} \\left[ \\mathcal {D} _ {\\phi} \\left(\\hat {\\mathbf {x}} _ {t ^ {\\prime}}, t ^ {\\prime}\\right) \\right]. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 191, + 540, + 480, + 560 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Following [74], we also incorporate a reconstruction objective to enhance stability, defined as:", + "bbox": [ + 89, + 565, + 482, + 595 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {r e c o n}} = \\sqrt {\\left\\| \\hat {\\mathbf {x}} _ {0} - \\mathbf {x} _ {0} \\right\\| _ {2} ^ {2} + c ^ {2}} - c, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 174, + 608, + 482, + 633 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\hat{\\mathbf{x}}_0 = \\mathbf{x}_t - t\\cdot \\mathcal{G}_{\\theta}(t,\\mathbf{x}_t)$ , and $c > 0$ is an adjustable constant.", + "bbox": [ + 89, + 643, + 482, + 672 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Discussion. Our latent adversarial training pipeline is inspired by SF-V [74]. Similar to SF-V, we set $k = 4$ and utilize the part of the pre-trained diffusion model as the backbone for the discriminator. However, our approach introduces several key differences. First, our method is built on an efficient UNet specifically designed for mobile devices, with fewer parameters than SVD [3], making it a more challenging task. Second, we redesign the discriminator heads: instead of using separate spatial and temporal heads, we integrate them into a unified spatial-temporal head for adversarial training. Rather than handling the temporal dimension separately with 1-D convolutional kernels, we incorporate a temporal self-attention layer into the spatial discriminator head after the 2-D ResBlock, forming a spatial-temporal discriminator head. This unified design enables", + "bbox": [ + 89, + 674, + 482, + 900 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/a0a32c9effc8517ed4200639e7cd253731deb36c8ae66eca2c83c9db045a799e.jpg", + "image_caption": [ + "Figure 4. Latent Adversarial Fine-tuning. Given a latent $\\mathbf{x}_0$ and a noise latent $\\epsilon$ , we obtain the intermediate noisy latent $\\mathbf{x}_t$ through a forward diffusion process. The generator then predicts the velocity as $\\mathcal{G}_{\\theta}(\\mathbf{x}_t, t)$ . Using the predicted velocity, we compute $\\hat{\\mathbf{x}}_0$ and calculate the reconstruction loss $\\mathcal{L}_{\\mathrm{recon}}$ between $\\mathbf{x}_0$ and $\\hat{\\mathbf{x}}_0$ . For adversarial training, the real sample $\\mathbf{x}_{t'}$ is obtained using the same forward diffusion process, while the fake sample $\\hat{\\mathbf{x}}_{t'}$ is computed using $\\mathbf{x}_t$ and the predicted velocity $\\mathcal{G}_{\\theta}(\\mathbf{x}_t, t)$ ." + ], + "image_footnote": [], + "bbox": [ + 553, + 87, + 867, + 191 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "our model to be jointly trained on both image and video data, which, as demonstrated in Sec. 4.3, significantly enhances the performance of the fine-tuned model.", + "bbox": [ + 511, + 353, + 906, + 398 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 513, + 431, + 643, + 449 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training. The efficient image backbone is obtained by pruning the Stable Diffusion v1.5 UNet for $100K$ iterations on high-quality synthetic image datasets. The model is then fine-tuned for $50K$ additional iterations to adapt to Rectified-Flow velocity prediction [10] as well as to the Spatial-Temporal VAE [77]. We incorporate QK-norm and RoPE [54] in our network to stabilize training. The workflow for architecture search is discussed in Sec. 3.2. The image model pruning, temporal architecture search, and final model training are conducted on 256 NVIDIA A100 80G GPUs using AdamW optimizer with $5e - 5$ learning rate and betas values as [0.9, 0.999].", + "bbox": [ + 511, + 464, + 906, + 645 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Adversarial Fine-tuning is conducted for $6K$ iterations on 64 NVIDIA A100 GPUs, using the AdamW optimizer with a learning rate of $1e - 7$ for the generator (i.e., UNet) and $1e - 4$ for the discriminator heads. We set the betas as [0.9, 0.999] for the generator optimizer, and [0.5, 0.999] for the discriminator optimizer. We set the EMA rate as 0.95 following SF-V [74]. We set $m = -1$ , $s = 1$ if not otherwise noted.", + "bbox": [ + 511, + 651, + 906, + 772 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation. The model is evaluated following the standard benchmarking procedure of VBench [19]. With the 4-step adversarial distilled model, we generate 120-frame horizontal videos at a resolution of $432 \\times 768$ using 4 inference steps without employing classifier-free guidance. The generated video is saved at 5 seconds 24 FPS for score testing and qualitative visualization. The mobile demo and detailed demo settings are included in the supplementary material.", + "bbox": [ + 511, + 779, + 908, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "2483", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/3bafabc39831471381c38708da6d77ba0e4e64ae46b55a307283b3e1eefac486.jpg", + "image_caption": [ + "Figure 5. Video generation on various domains. We employ our model to synthesize videos across diverse domains, with each video containing 120 frames at a resolution of $432 \\times 768$ . All results are generated through a 4-step inference process. The results demonstrate that our model can produce high-quality, motion-consistent videos featuring various objects across different domains." + ], + "image_footnote": [], + "bbox": [ + 129, + 79, + 867, + 579 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/0fd49755ed918fe6ab963ec2ec5c521d4ef6a304c8d150868898d7d84932d492.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelParams (B)StepsTotal scoreQuality scoreSemantic scoretemporal flickeringmotion smoothnessdynamic degreeaesthetic qualityobject classmultiple objectscolorspatial relationshipappearance styletemporal styleoverall consistency
Kling--81.8583.3975.6899.3099.4046.9461.2187.2468.0589.9073.0319.6224.1726.42
Pyramid Flow2.02081.7284.7469.6299.4999.1264.6363.2686.6750.7182.8759.5320.9123.0926.23
CogVideoX5.05081.6182.7577.0498.6696.9270.9761.9885.2362.1182.8166.3524.9125.3827.59
T2V-Turbo1.6481.0182.5774.7697.4897.3449.1763.0493.9654.6589.9038.6724.4225.5128.16
Emu38.0-80.9684.0968.4398.5798.9379.2759.6486.1744.6488.3468.7320.9223.2624.79
CogVideoX1.65080.9182.1875.8398.8997.7359.8660.8283.3762.6379.4169.9024.8024.3626.66
Pika-1.0--80.6982.9271.7799.7499.5047.5062.0488.7243.0890.5761.0322.2624.2225.94
VideoCrafter-2.01.95080.4482.2073.4298.4197.7342.5063.1392.5540.6692.9235.8625.1325.8428.23
AnimateDiff-V21.22580.2782.9069.7598.7597.7640.8367.1690.9036.8887.4734.6022.4226.0327.04
OpenSora V1.21.23079.7681.3573.3999.5398.5042.3956.8582.2251.8390.0868.5623.9524.5426.85
AnimateLCM1.2479.4282.3667.6598.5298.1640.5667.0191.4129.7684.2437.1420.9725.1626.07
ModelScope1.45075.7578.0566.5498.2895.7966.3952.0682.2538.9881.7233.6823.3925.3725.67
Ours0.6481.1483.4771.8499.3799.2951.1162.1992.2254.3486.6356.2020.1725.1827.42
", + "bbox": [ + 119, + 633, + 883, + 804 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Performance comparison with popular video generation models on VBench [19].", + "bbox": [ + 217, + 813, + 777, + 828 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Qualitative Visualization", + "text_level": 1, + "bbox": [ + 89, + 830, + 316, + 848 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We show visualizations of our generated videos in Fig. 5. Our model consistently produces high-quality video frames and smooth object movements. To demonstrate the generic", + "bbox": [ + 89, + 854, + 483, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "text-to-video generation ability, we show various generation examples, including human, animal, photorealistic and art-styled scenes. We include more video visualizations in the supplementary material.", + "bbox": [ + 511, + 832, + 906, + 893 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "2484", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Quantitative Comparisons", + "text_level": 1, + "bbox": [ + 89, + 90, + 331, + 107 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We present a comprehensive evaluation of our method against existing popular video generation models on VBench [19], as in Tab. 2. Despite the fact that our model is compact and designated for fast inference on mobile, it achieves higher total score compared to recent arts, including the DiT-based OpenSora-V1.2, CogVideoX-2B [67], and the UNet-based VideoCrafter-2.0 [6]. In addition, compared to the 4-step distilled T2V-Turbo [27] and AnimateLCM [58], our model achieves better performance with more than $50\\%$ reduction in size. The quantitative scores demonstrate the superiority of our efficient model design and the tailored adversarial distillation method.", + "bbox": [ + 88, + 112, + 482, + 292 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "User Study. Human evaluations are conducted between our model and baselines as in the Tab. 3. We generate videos from VBench and Movie Gen Bench prompts and ask human labelers to pick the best results across prompt alignment, aesthetics, and motion. The result shows that our model outperforms baseline metrics by a large margin.", + "bbox": [ + 89, + 292, + 483, + 385 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/32560609e60dbf6bad8ce42f903c2450e0b71669b0fadc97f9f528274bb8b0a8.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Win RatePrompt AlignmentAestheticsMotion
OpenSora-1.224.9%21.2%20.4%
CogVideoX-2B30.7%30.5%30.3%
Ours44.4%48.3%49.3%
", + "bbox": [ + 112, + 395, + 464, + 454 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Analysis", + "text_level": 1, + "bbox": [ + 89, + 516, + 264, + 532 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparison of Training Data Scheme. We compare the model trained with joint image-video datasets vs. video-only datasets. As shown in Tab. 4, training with video-only datasets leads to significant performance degradation on the VBenchmark score with a drop of 6.32 in aesthetic quality, 2.80 in image quality, and 2.54 in total score. The results highlight the importance of joint image-video training, as the image dataset offers more contextual information and enhances diversity.", + "bbox": [ + 89, + 537, + 482, + 672 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Model Scaling. We scale up and down the model size by adjusting the number of temporal layers, as shown in Tab. 4, to demonstrate the effectiveness of the proposed temporal architecture search. We can observe that scaling up the model can only marginally improve generative scores (i.e., $\\times 2$ scale-up only increasing 0.56 in dynamic degree and 0.12 in motion smoothness, and 0.27 in total score). However, both the $\\times 2$ and $\\times 4$ models hit the memory bound on iPhone. By dividing the $\\times 2$ model into more chunks, we test its mobile speed and observe nearly doubled latency. While on the other hand, further scaling down the model results in heavy losses in generation quality (i.e., decreasing 0.56 in dynamic degree and 0.61 in motion smoothness). Our efficient model is a balanced sweet point for quality and on-device performance.", + "bbox": [ + 89, + 674, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/1b7907e83dbbec425b4b6e7e42d85b02bf63c57e89be74fd7ced92d54b19ccaa.jpg", + "table_caption": [ + "Table 3. User Study between OpenSora-1.2, CogVideoX-2B, and our model on prompt alignment, aesthetics, and motion." + ], + "table_footnote": [], + "table_body": "
w/ I ScalingΔTΔAQΔIQΔDDΔMSΔTotal
Video-onlyX——-6.32-2.80+4.6-0.17-2.54
×0.5-22%-1.90+0.04-0.56-0.61-1.58
Temporal Scaling✓×2+32%+0.37+0.10+0.56+0.12+0.27
×4+107%+0.49-0.17+0.91+0.17+0.31
", + "bbox": [ + 519, + 89, + 903, + 161 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. Analysis of training data scheme, latency, and quality of efficient architecture. The baseline model adopts the best suitable architecture and is trained with joint image-video datasets. The \"scaling\", and \"T\" indicates the number of the temporal layers, and the latency comparing to the baseline. The \"AQ\", \"IQ\", \"DD\", and \"MS\", are aesthetic quality, image quality, dynamic degree, and motion smoothness in VBench [19]. The benchmark metrics are presented as the differences from the baseline model, where negative values indicate a decrease in performance over the baseline vice versa.", + "bbox": [ + 511, + 172, + 906, + 310 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a3bb3988332de3159b74ee15ce5131b91ba4d3089fd12b4dc844aab8a05326f4.jpg", + "image_caption": [ + "Figure 6. Analysis of temporal layers. We ablate temporal layers in different network stages to evaluate their effect. The scores are normalized according to our base model to better demonstrate the difference." + ], + "image_footnote": [], + "bbox": [ + 553, + 319, + 867, + 506 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effect of Different Temporal Layers. To better understand the roles of the searched temporal layers, we compute the VBench score after systematically removing (i) all temporal layers in the downsample stage, (ii) bottleneck temporal layers, and (iii) all temporal layers in the upsample stage. The model makes reasonable zero-shot generations after temporal layer removal, but we still fine-tune 10K iterations under the same recipe as in Sec. 3.2 for fair comparison. As shown in Fig. 6, removing different temporal layers results in varying degrees of performance degradation across different metrics, and all removing strategies result in a substantial drop in total score, demonstrating that the existence of the searched temporal layers is important, and they play different roles in generation. Specifically, removing temporal layers in upsample blocks results in a more significant loss in imaging quality, subject consistency, and background consistency, suggesting that the up temporal layers play important roles in detail reconstruction. In contrast, bottleneck layers are more important in human action and object class, where global information modeling dom", + "bbox": [ + 511, + 598, + 906, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "2485", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "inates the results. We observe that removing down layers introduces less overall degradations compared to the other two, which is an anticipated phenomenon because the loss of modeling capacity can be mitigated by the subsequent bottleneck and up stage temporal layers after fine-tuning.", + "bbox": [ + 89, + 90, + 482, + 167 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/b6e2f30b8b17a02dffa0dc40b2d88a3f07d4c256927bedd72d1050fb9c1ce1c0.jpg", + "table_caption": [], + "table_footnote": [ + "Table 5. Analysis of our adversarial fine-tuning scheme. We evaluate the VBench [19] scores for our models using different training schemes. In the results, \"TF\" and \"MO\" denote the temporal flickering and multiple objects sub-scores, respectively, while \"Q\" and \"S\" represent the quality and semantic scores. The table summarizes how varying (1) the type of discriminator head, (2) training with or without an image dataset, and (3) different $t'$ distributions impact the performance of our model." + ], + "table_body": "
Headw/Iw/ LadvmTFMOQSTotal
SF-V [74]X✓-198.8630.7583.6064.2579.73
Ours99.2737.8583.6169.0180.69
Ours✓X-197.321.0170.0317.0459.44
-299.2947.6483.9170.5481.23
Ours✓✓-199.3754.3483.4771.8481.14
099.4443.6381.6969.4679.24
199.5624.4177.8458.8874.05
", + "bbox": [ + 104, + 181, + 470, + 289 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of Discriminator Heads. We compare the effects of our spatial-temporal discriminator heads with the separate spatial and temporal heads proposed in SF-V [74] to demonstrate the effectiveness of our discriminator architecture. For a fair comparison, both models are trained exclusively on video data. We evaluate the models finetuned with different discriminator heads on VBench [19]. As shown in the first two rows of Tab. 5, our discriminator heads result in improvements in both the quality score (83.61 vs. 83.60) and the semantic score (69.01 vs. 64.25) of the generated videos.", + "bbox": [ + 89, + 429, + 482, + 595 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of Joint Image Video Fine-tuning. We examine the impact of incorporating image data during adversarial training, as shown in the second and fourth rows of Tab. 5. The results indicate that fine-tuning the model with image data can slightly decrease the Quality score in the VBench [19] evaluation (from 83.61 to 83.47). However, by leveraging the increased diversity of the image dataset, the model achieves a substantial improvement in semantic performance, particularly in multi-object generation (from 37.85 to 54.34). This enhancement leads to a better overall score compared to the model trained exclusively on video data (81.14 vs. 80.69).", + "bbox": [ + 89, + 597, + 482, + 777 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of Noise Distribution for Discriminator. Following Eq. (3), the parameters $m$ and $s$ control the distribution of $t'$ , which determines the noise levels of $\\mathbf{x}_{t'}$ and $\\hat{\\mathbf{x}}_{t'}$ before they are passed to the discriminator as real and fake samples, respectively. We investigate the effect of different noise distributions on model performance by evaluating the results using VBenchmark [19]. As shown in the last four rows of Tab. 5, increasing $m$ (resulting in noisier real and fake sam", + "bbox": [ + 89, + 779, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "plies) degrades the quality score (from 83.91 to 77.84) while slightly enhancing the temporal flickering score (from 99.29 to 99.56). Although setting $m = -2$ achieves the highest overall score among the experiments (81.23), it performs poorly on multi-object generation. Therefore, in most of our experiments, unless otherwise stated, we use $m = -1$ . This setting yields a slightly lower overall score (81.14) but significantly improves semantic performance (71.84 vs. 70.54) and excels in multi-object generation (54.34 vs. 47.64).", + "bbox": [ + 511, + 90, + 903, + 227 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/7f71196075dede4efe74ca37d257aff9ca615c0cd64a21f08646f922d7af127f.jpg", + "table_caption": [], + "table_footnote": [ + "Table 6. Analysis of the number of inference steps. We measure VBench [19] score with different numbers of inference steps. In the results, \"DD\", \"OC\", and \"AQ\" denote the dynamic degree, object class, and aesthetic quality scores, respectively." + ], + "table_body": "
# StepsDDOCAQQualitySemanticTotal
111.3979.2452.8574.4362.7072.09
230.2888.9759.7480.0869.2877.92
451.1192.2262.1983.4771.8481.14
", + "bbox": [ + 519, + 239, + 900, + 306 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of Inference Steps. Our fine-tuned model supports generation with a reduced number of inference steps. We further investigate how varying the number of evaluation steps affects the quality of the generated results, as shown in Tab. 6. While our four-step generation achieves the best performance, even with only two steps, the model still produces reasonable results. Increasing the number of inference steps improves the performance of our model across all metrics. In Tab. 6, we report not only the quality and semantic scores but also scores for dynamic degree, object class, and aesthetic quality. However, reducing the process to a single inference step leads to a significant drop in performance. We leave more aggressive step reductions to future work.", + "bbox": [ + 511, + 390, + 906, + 601 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Discussion and Conclusion", + "text_level": 1, + "bbox": [ + 511, + 617, + 761, + 633 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we propose an acceleration framework for the video diffusion model, and for the first time, achieve superfast text-to-video generation on mobile devices. Specifically, we discover an efficient but powerful network architecture through latency and memory joint architecture search for temporal layers. In addition, we propose an improved adversarial fine-tuning technique to distill our model to 4 steps to further speed up generation. Our work is a good starting point for the edge deployment of video diffusion models and we hope to inspire more downstream applications such as video extension and editing.", + "bbox": [ + 511, + 643, + 903, + 808 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations. We use a public 4-channel VAE [77] to encode videos to latent space. Recent work has shown that using more latent channels benefits reconstruction details. Another future direction is to further improve the step reduction technique for 1-2 denoising steps, as works [74] on server-level models.", + "bbox": [ + 511, + 810, + 903, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "2486", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 90, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Luma AI. Dream machine. https://lumalabs.ai/dream-machine,.2", + "[2] Pika AI. Pika 1.5. https://pika.art/try..2", + "[3] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. ArXiv preprint, abs/2311.15127, 2023. 1, 2, 3, 5", + "[4] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2023, Vancouver, BC, Canada, June 17-24, 2023, pages 22563-22575. IEEE, 2023. 1, 2", + "[5] Haoxin Chen, Menghan Xia, Yingqing He, Yong Zhang, Xiaodong Cun, Shaoshu Yang, Jinbo Xing, Yaofang Liu, Qifeng Chen, Xintao Wang, et al. Videocrafter1: Open diffusion models for high-quality video generation. ArXiv preprint, abs/2310.19512, 2023. 2", + "[6] Haoxin Chen, Yong Zhang, Xiaodong Cun, Menghan Xia, Xintao Wang, Chao Weng, and Ying Shan. Videocaster2: Overcoming data limitations for high-quality video diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7310-7320, 2024. 2, 7", + "[7] Tsai-Shien Chen, Aliaksandr Siarohin, Willi Menapace, Ekaterina Deyneka, Hsiang-wei Chao, Byung Eun Jeon, Yuwei Fang, Hsin-Ying Lee, Jian Ren, Ming-Hsuan Yang, et al. Panda-70m: Captioning 70m videos with multiple cross-modality teachers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13320-13331, 2024. 2", + "[8] Xinwang Chen, Ning Liu, Yichen Zhu, Feifei Feng, and Jian Tang. Edit: An efficient diffusion transformer framework inspired by human-like sketching. ArXiv preprint, abs/2410.23788, 2024. 1, 2", + "[9] Trung Dao, Thuan Hoang Nguyen, Thanh Le, Duc Vu, Khoi Nguyen, Cuong Pham, and Anh Tran. Swiftbrush v2: Make your one-step diffusion model better than its teacher. In European Conference on Computer Vision, pages 176-192. Springer, 2025. 3", + "[10] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In *Forty-first International Conference on Machine Learning*, 2024. 3, 5", + "[11] Gongfan Fang, Xinyin Ma, and Xinchao Wang. Structural pruning for diffusion models. In Advances in Neural Information Processing Systems 36: Annual Conference on Neural Information Processing Systems 2023, NeurIPS 2023, New Orleans, LA, USA, December 10 - 16, 2023, 2023. 2, 3", + "[12] Yutang Feng, Sicheng Gao, Yuxiang Bao, Xiaodi Wang, Shumin Han, Juan Zhang, Baochang Zhang, and Angela Yao. Wave: Warping ddim inversion features for zero-shot" + ], + "bbox": [ + 93, + 114, + 482, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "text-to-video editing. In European Conference on Computer Vision, pages 38-55. Springer, 2025. 1", + "[13] Yuwei Guo, Ceyuan Yang, Anyi Rao, Zhengyang Liang, Yaohui Wang, Yu Qiao, Maneesh Agrawala, Dahua Lin, and Bo Dai. Animatediff:Animate your personalized text-to-image diffusion models without specific tuning. ArXiv preprint, abs/2307.04725, 2023. 1, 2, 3", + "[14] Yuwei Guo, Ceyuan Yang, Anyi Rao, Maneesh Agrawala, Dahua Lin, and Bo Dai. Sparsectrl: Adding sparse controls to text-to-video diffusion models. In European Conference on Computer Vision, pages 330-348. Springer, 2025. 2", + "[15] Agrim Gupta, Lijun Yu, Kihyuk Sohn, Xiuye Gu, Meera Hahn, Li Fei-Fei, Irfan Essa, Lu Jiang, and José Lezama. Photorealistic video generation with diffusion models. ArXiv preprint, abs/2312.06662, 2023. 2", + "[16] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. ArXiv preprint, abs/2207.12598, 2022. 2", + "[17] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In Advances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, NeurIPS 2020, December 6-12, 2020, virtual, 2020. 2, 3", + "[18] Wenyi Hong, Ming Ding, Wendi Zheng, Xinghan Liu, and Jie Tang. Cogvideo: Large-scale pretraining for text-to-video generation via transformers. In The Eleventh International Conference on Learning Representations, ICLR 2023, Kigali, Rwanda, May 1-5, 2023. OpenReview.net, 2023. 2", + "[19] Ziqi Huang, Yinan He, Jiashuo Yu, Fan Zhang, Chenyang Si, Yuming Jiang, Yuanhan Zhang, Tianxing Wu, Qingyang Jin, Nattapol Chanpaisit, Yaohui Wang, Xinyuan Chen, Limin Wang, Dahua Lin, Yu Qiao, and Ziwei Liu. VBenchmark: Comprehensive benchmark suite for video generative models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 2, 4, 5, 6, 7, 8, 1", + "[20] Hyeonho Jeong, Jinho Chang, Geon Yeong Park, and Jong Chul Ye. Dreammotion: Space-time self-similar score distillation for zero-shot video editing. ArXiv preprint, abs/2403.12002, 2024. 1", + "[21] Yang Jin, Zhicheng Sun, Ningyuan Li, Kun Xu, Hao Jiang, Nan Zhuang, Quzhe Huang, Yang Song, Yadong Mu, and Zhouchen Lin. Pyramidal flow matching for efficient video generative modeling. ArXiv preprint, abs/2410.05954, 2024. 2", + "[22] Beomsu Kim, Yu-Guan Hsieh, Michal Klein, Marco Cuturi, Jong Chul Ye, Bahjat Kawar, and James Thornton. Simple reflow: Improved techniques for fast flow models. ArXiv preprint, abs/2410.07815, 2024. 3", + "[23] Jihwan Kim, Junoh Kang, Jinyoung Choi, and Bohyung Han. Fifo-diffusion: Generating infinite videos from text without training. ArXiv preprint, abs/2405.11473, 2024. 2", + "[24] Kuaishou. Kling. https://kling.kuaishou.com/en.2", + "[25] Jeong-gi Kwak, Erqun Dong, Yuhe Jin, Hanseok Ko, Shweta Mahajan, and Kwang Moo Yi. Vivid-1-to-3: Novel view synthesis with video diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6775–6785, 2024. 1" + ], + "bbox": [ + 517, + 93, + 905, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "2487", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[26] PKU-Yuan Lab and Tuzhan AI etc. Open-sora-plan, 2024. 2", + "[27] Jiachen Li, Weixi Feng, Tsu-Jui Fu, Xinyi Wang, Sugato Basu, Wenhu Chen, and William Yang Wang. T2v-turbo: Breaking the quality bottleneck of video consistency model with mixed reward feedback. ArXiv preprint, abs/2405.18750, 2024. 2, 7", + "[28] Jiachen Li, Qian Long, Jian Zheng, Xiaofeng Gao, Robinson Piramuthu, Wenhu Chen, and William Yang Wang. T2v-turbo-v2: Enhancing video generation model post-training through data, reward, and conditional guidance design. ArXiv preprint, abs/2410.05677, 2024. 2", + "[29] Yanyu Li, Huan Wang, Qing Jin, Ju Hu, Pavlo Chemerys, Yun Fu, Yanzhi Wang, Sergey Tulyakov, and Jian Ren. Snapfusion: Text-to-image diffusion model on mobile devices within two seconds. In Advances in Neural Information Processing Systems 36: Annual Conference on Neural Information Processing Systems 2023, NeurIPS 2023, New Orleans, LA, USA, December 10 - 16, 2023, 2, 3", + "[30] Feng Liang, Bichen Wu, Jialiang Wang, Licheng Yu, Kunpeng Li, Yinan Zhao, Ishan Misra, Jia-Bin Huang, Peizhao Zhang, Peter Vajda, et al. Flowvid: Taming imperfect optical flows for consistent video-to-video synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8207-8216, 2024. 1", + "[31] Jae Hyun Lim and Jong Chul Ye. Geometric gan. ArXiv preprint, abs/1705.02894, 2017. 5", + "[32] Shanchuan Lin and Xiao Yang. Animatediff-lightning: Cross-model diffusion distillation. ArXiv preprint, abs/2403.12706, 2024. 3", + "[33] Xingchao Liu, Chengyue Gong, and Qiang Liu. Flow straight and fast: Learning to generate and transfer data with rectified flow. In The Eleventh International Conference on Learning Representations, ICLR 2023, Kigali, Rwanda, May 1-5, 2023. OpenReview.net, 2023. 3", + "[34] Yu Lu, Yuanzhi Liang, Linchao Zhu, and Yi Yang. Freelong: Training-free long video generation with spectralblend temporal attention. ArXiv preprint, abs/2407.19918, 2024. 2", + "[35] Xiaofeng Mao, Zhengkai Jiang, Fu-Yun Wang, Wenbing Zhu, Jiangning Zhang, Hao Chen, Mingmin Chi, and Yabiao Wang. Osv: One step is enough for high-quality image to video generation. ArXiv preprint, abs/2409.11367, 2024. 3", + "[36] Kangfu Mei, Maurizio Delbracio, Hossein Talebi, Zhengzhong Tu, Vishal M Patel, and Peyman Milanfar. Codi: Conditional diffusion distillation for higher-fidelity and faster image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9048-9058, 2024. 3", + "[37] MiniMax. Hailuo ai. https://hailuoai(video/.2", + "[38] Kepan Nan, Rui Xie, Penghao Zhou, Tiehan Fan, Zhenheng Yang, Zhijie Chen, Xiang Li, Jian Yang, and Ying Tai. Openvid-1m: A large-scale high-quality dataset for text-to-video generation. ArXiv preprint, abs/2407.02371, 2024. 2", + "[39] OpenAI. Video generation models as world simulators. https://openai.com/index/video-generation-models-as-world-simulators/.2", + "[40] William Peebles and Saining Xie. Scalable diffusion models with transformers. In IEEE/CVF International Conference" + ], + "bbox": [ + 89, + 90, + 483, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "on Computer Vision, ICCV 2023, Paris, France, October 1-6, 2023, pages 4172-4182. IEEE, 2023. 1, 2, 3", + "[41] Adam Polyak, Amit Zohar, Andrew Brown, Andros Tjandra, Animesh Sinha, Ann Lee, Apoorv Vyas, Bowen Shi, Chih-Yao Ma, Ching-Yao Chuang, et al. Movie gen: A cast of media foundation models. ArXiv preprint, abs/2410.13720, 2024. 1, 2", + "[42] Zhiwu Qing, Shiwei Zhang, Jiayu Wang, Xiang Wang, Yujie Wei, Yingya Zhang, Changxin Gao, and Nong Sang. Hierarchical spatio-temporal decoupling for text-to-video generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6635-6645, 2024. 2", + "[43] Haonan Qiu, Menghan Xia, Yong Zhang, Yingqing He, Xintao Wang, Ying Shan, and Ziwei Liu. Freenoise: Tuning-free longer video diffusion via noise rescheduling. ArXiv preprint, abs/2310.15169, 2023. 2", + "[44] Yixuan Ren, Yang Zhou, Jimei Yang, Jing Shi, Difan Liu, Feng Liu, Mingi Kwon, and Abhinav Shrivastava. Customize-a-video: One-shot motion customization of text-to-video diffusion models. ArXiv preprint, abs/2402.14780, 2024. 2", + "[45] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 2, 3", + "[46] Runway. Gen-3 alpha. https://runwayml.com/research/introducing-gen-3-alpha.2", + "[47] Tim Salimans and Jonathan Ho. Progressive distillation for fast sampling of diffusion models. In *The Tenth International Conference on Learning Representations*, ICLR 2022, Virtual Event, April 25-29, 2022. OpenReview.net, 2022. 3", + "[48] Axel Sauer, Kashyap Chitta, Jens Müller, and Andreas Geiger. Projected gans converge faster. In Advances in Neural Information Processing Systems 34: Annual Conference on Neural Information Processing Systems 2021, NeurIPS 2021, December 6-14, 2021, virtual, pages 17480-17492, 2021. 4, 5", + "[49] Axel Sauer, Tero Karras, Samuli Laine, Andreas Geiger, and Timo Aila. Stylegan-t: Unlocking the power of gans for fast large-scale text-to-image synthesis. In International Conference on Machine Learning, ICML 2023, 23-29 July 2023, Honolulu, Hawaii, USA, pages 30105-30118. PMLR, 2023. 4", + "[50] Axel Sauer, Dominik Lorenz, Andreas Blattmann, and Robin Rombach. Adversarial diffusion distillation. ArXiv preprint, abs/2311.17042, 2023. 3, 5", + "[51] Axel Sauer, Frederic Boesel, Tim Dockhorn, Andreas Blattmann, Patrick Esser, and Robin Rombach. Fast high-resolution image synthesis with latent adversarial diffusion distillation. ArXiv preprint, abs/2403.12015, 2024. 3, 4", + "[52] Yang Song and Prafulla Dhariwal. Improved techniques for training consistency models. ArXiv preprint, abs/2310.14189, 2023. 3", + "[53] Yang Song, Prafulla Dhariwal, Mark Chen, and Ilya Sutskever. Consistency models. In International Conference on Machine Learning, ICML 2023, 23-29 July 2023," + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "2488", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Honolulu, Hawaii, USA, pages 32211-32252. PMLR, 2023. 3", + "[54] Jianlin Su, Murtadha Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu. Roformer: Enhanced transformer with rotary position embedding. Neurocomputing, 568:127063, 2024. 5", + "[55] Genmo Team. Mochi, 2024. 1, 2", + "[56] Gemini Team, Rohan Anil, Sebastian Borgeaud, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, Katie Millican, et al. Gemini: a family of highly capable multimodal models. ArXiv preprint, abs/2312.11805, 2023. 1, 2", + "[57] Vikram Voleti, Chun-Han Yao, Mark Boss, Adam Letts, David Pankratz, Dmitry Tochilkin, Christian Laforte, Robin Rombach, and Varun Jampani. Sv3d: Novel multi-view synthesis and 3d generation from a single image using latent video diffusion. In European Conference on Computer Vision, pages 439-457. Springer, 2025. 1", + "[58] Fu-Yun Wang, Zhaoyang Huang, Xiaoyu Shi, Weikang Bian, Guanglu Song, Yu Liu, and Hongsheng Li. *Animatelcm: Accelerating the animation of personalized diffusion models and adapters with decoupled consistency learning*. *ArXiv* preprint, abs/2402.00769, 2024. 3, 7", + "[59] Fu-Yun Wang, Ling Yang, Zhaoyang Huang, Mengdi Wang, and Hongsheng Li. Rectified diffusion: Straightness is not your need in rectified flow. ArXiv preprint, abs/2410.07303, 2024. 3", + "[60] Xiang Wang, Shiwei Zhang, Han Zhang, Yu Liu, Yingya Zhang, Changxin Gao, and Nong Sang. Videolcm: Video latent consistency model. ArXiv preprint, abs/2312.09109, 2023. 3", + "[61] Jianzong Wu, Xiangtai Li, Yanhong Zeng, Jiangning Zhang, Qianyu Zhou, Yining Li, Yunhai Tong, and Kai Chen. Motionbooth: Motion-aware customized text-to-video generation. ArXiv preprint, abs/2406.17758, 2024. 2", + "[62] Jay Zhangjie Wu, Yixiao Ge, Xintao Wang, Stan Weixian Lei, Yuchao Gu, Yufei Shi, Wynne Hsu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. In IEEE/CVF International Conference on Computer Vision, ICCV 2023, Paris, France, October 1-6, 2023, pages 7589-7599. IEEE, 2023. 1, 2", + "[63] Tianxing Wu, Chenyang Si, Yuming Jiang, Ziqi Huang, and Ziwei Liu. Freeinit: Bridging initialization gap in video diffusion models. In European Conference on Computer Vision, pages 378-394. Springer, 2025. 2", + "[64] Enze Xie, Junsong Chen, Junyu Chen, Han Cai, Yujun Lin, Zhekai Zhang, Muyang Li, Yao Lu, and Song Han. Sana: Efficient high-resolution image synthesis with linear diffusion transformers. ArXiv preprint, abs/2410.10629, 2024. 1", + "[65] Yanwu Xu, Yang Zhao, Zhisheng Xiao, and Tingbo Hou. Ufogen: You forward once large scale text-to-image generation via diffusion gans. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8196-8206, 2024. 3", + "[66] Ling Yang, Zixiang Zhang, Zhilong Zhang, Xingchao Liu, Minkai Xu, Wentao Zhang, Chenlin Meng, Stefano Ermon, and Bin Cui. Consistency flow matching: Defining" + ], + "bbox": [ + 91, + 90, + 482, + 901 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "straight flows with velocity consistency. ArXiv preprint, abs/2407.02398, 2024. 3", + "[67] Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, et al. Cogvideox: Text-to-video diffusion models with an expert transformer. ArXiv preprint, abs/2408.06072, 2024. 1, 2, 3, 7, 5", + "[68] Tianwei Yin, Michael Gharbi, Taesung Park, Richard Zhang, Eli Shechtman, Fredo Durand, and William T Freeman. Improved distribution matching distillation for fast image synthesis. ArXiv preprint, abs/2405.14867, 2024. 3", + "[69] Tianwei Yin, Michael Gharbi, Richard Zhang, Eli Shechtman, Fredo Durand, William T Freeman, and Taesung Park. One-step diffusion with distribution matching distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6613-6623, 2024. 3", + "[70] Zhihang Yuan, Pu Lu, Hanling Zhang, Xuefei Ning, Linfeng Zhang, Tianchen Zhao, Shengen Yan, Guohao Dai, and Yu Wang. Ditfastattn: Attention compression for diffusion transformer models. ArXiv preprint, abs/2406.08552, 2024. 1, 2", + "[71] Yan Zeng, Guoqiang Wei, Jiani Zheng, Jiaxin Zou, Yang Wei, Yuchen Zhang, and Hang Li. Make pixels dance: High-dynamic video generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8850-8860, 2024. 2", + "[72] Yuanhao Zhai, Kevin Lin, Zhengyuan Yang, Linjie Li, Jianfeng Wang, Chung-Ching Lin, David Doermann, Junsong Yuan, and Lijuan Wang. Motion consistency model: Accelerating video diffusion with disentangled motion-appearance distillation. ArXiv preprint, abs/2406.06890, 2024. 3", + "[73] Zheng Zhan, Yushu Wu, Yifan Gong, Zichong Meng, Zhenglun Kong, Changdi Yang, Geng Yuan, Pu Zhao, Wei Niu, and Yanzhi Wang. Fast and memory-efficient video diffusion using streamlined inference. In The Thirty-eighth Annual Conference on Neural Information Processing Systems. 2", + "[74] Zhixing Zhang, Yanyu Li, Yushu Wu, Yanwu Xu, Anil Kag, Ivan Skorokhodov, Willi Menapace, Aliaksandr Siarohin, Junli Cao, Dimitris Metaxas, et al. Sf-v: Single forward video generation model. ArXiv preprint, abs/2406.04324, 2024. 3, 4, 5, 8", + "[75] Zhixing Zhang, Bichen Wu, Xiaoyan Wang, Yaqiao Luo, Luxin Zhang, Yinan Zhao, Peter Vajda, Dimitris Metaxas, and Licheng Yu. Avid: Any-length video inpainting with diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7162-7172, 2024. 2", + "[76] Rui Zhao, Yuchao Gu, Jay Zhangjie Wu, David Junhao Zhang, Jia-Wei Liu, Weijia Wu, Jussi Keppo, and Mike Zheng Shou. Motiondirector: Motion customization of text-to-video diffusion models. In European Conference on Computer Vision, pages 273–290. Springer, 2025. 2", + "[77] Zangwei Zheng, Xiangyu Peng, Tianji Yang, Chenhui Shen, Shenggui Li, Hongxin Liu, Yukun Zhou, Tianyi Li, and Yang You. Open-sora: Democratizing efficient video production for all, 2024. 2, 3, 4, 5, 8, 1" + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "2489", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[78] Yuan Zhou, Qiuyue Wang, Yuxuan Cai, and Huan Yang. Allegro: Open the black box of commercial-level video generation model. ArXiv preprint, abs/2410.15458, 2024. 1, 2", + "[79] Yupeng Zhou, Daquan Zhou, Ming-Ming Cheng, Jiashi Feng, and Qibin Hou. Storydiffusion: Consistent self-attention for long-range image and video generation. NeurIPS 2024, 2024. 2" + ], + "bbox": [ + 89, + 90, + 486, + 191 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "2490", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2025/SnapGen-V_ Generating a Five-Second Video within Five Seconds on a Mobile Device/2089e6e3-1b02-4449-b418-985fa5c0c175_model.json b/2025/SnapGen-V_ Generating a Five-Second Video within Five Seconds on a Mobile Device/2089e6e3-1b02-4449-b418-985fa5c0c175_model.json new file mode 100644 index 0000000000000000000000000000000000000000..b564809b6fd55c841384de98fc0512019f11287f --- /dev/null +++ b/2025/SnapGen-V_ Generating a Five-Second Video within Five Seconds on a Mobile Device/2089e6e3-1b02-4449-b418-985fa5c0c175_model.json @@ -0,0 +1,2336 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.107, + 0.13, + 0.892, + 0.175 + ], + "angle": 0, + "content": "SnapGen-V: Generating a Five-Second Video within Five Seconds on a Mobile Device" + }, + { + "type": "text", + "bbox": [ + 0.155, + 0.202, + 0.842, + 0.277 + ], + "angle": 0, + "content": "Yushu Wu\\(^{1,2\\dagger}\\) Zhixing Zhang\\(^{1,3\\dagger}\\) Yanyu Li\\(^{1\\dagger\\dagger}\\) Yanwu Xu\\(^{1}\\) Anil Kag\\(^{1}\\) Yang Sui\\(^{1}\\) Huseyin Coskun\\(^{1}\\) Ke Ma\\(^{1}\\) Aleksei Lebedev\\(^{1}\\) Ju Hu\\(^{1}\\) Dimitris N. Metaxas\\(^{3}\\) Yanzhi Wang\\(^{2}\\) Sergey Tulyakov\\(^{1}\\) Jian Ren\\(^{1\\dagger}\\) Snap Inc. Northeastern University Rutgers University" + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.31, + 0.905, + 0.433 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.442, + 0.908, + 0.471 + ], + "angle": 0, + "content": "Figure 1. Example generation results from our mobile text-to-video model. Our model can generate high-quality and motion consistent 5-second videos on a mobile device (e.g., iPhone 16 Pro Max) within 5 seconds." + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.483, + 0.327, + 0.499 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.516, + 0.486, + 0.851 + ], + "angle": 0, + "content": "We have witnessed the unprecedented success of diffusion-based video generation over the past year. Recently proposed models from the community have wielded the power to generate cinematic and high-resolution videos with smooth motions from arbitrary input prompts. However, as a supertask of image generation, video generation models require more computation and are thus hosted mostly on cloud servers, limiting broader adoption among content creators. In this work, we propose a comprehensive acceleration framework to bring the power of the largescale video diffusion model to the hands of edge users. From the network architecture scope, we initialize from a compact image backbone and search out the design and arrangement of temporal layers to maximize hardware efficiency. In addition, we propose a dedicated adversarial fine-tuning algorithm for our efficient model and reduce the denoising steps to 4. Our model, with only \\(0.6B\\) parameters, can generate a 5-second video on an iPhone 16 PM within 5 seconds. Compared to server-side models that take minutes on powerful GPUs to generate a single video, we accelerate the generation by magnitudes while delivering on-par quality. Project page at https://snap-research.github.io/snapgen-v/." + }, + { + "type": "title", + "bbox": [ + 0.515, + 0.483, + 0.645, + 0.499 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.508, + 0.907, + 0.629 + ], + "angle": 0, + "content": "Recently, the rapid advancement of video diffusion models [4] inspires revolutions in content creation. With the emergence of video models from industry [41] and research community [67], content creators can animate a static image [3] or generate cinematic videos from arbitrary prompts [55, 67, 78]. Video diffusion models also enable downstream applications like video editing [12, 20, 30], novel view synthesis [25, 57], and multi-modal generation [56]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.63, + 0.909, + 0.87 + ], + "angle": 0, + "content": "Despite the success in generation quality, the huge number of parameters and slow generation speed prohibit the wide deployment of video diffusion models. For instance, CogVideoX-5B [67] generates a video (49 frames at 8 fps, \\(720 \\times 480\\) resolution) in 5 minutes with 50 inference steps on an NVIDIA A100 GPU. Compared to text-to-image diffusion models [45], video diffusion models require extra parameters to model sophisticated motions [3, 13, 62]. In addition, video data usually incurs higher spatial-temporal resolution for UNet denoisers [3, 13], or equivalently more tokens for DiT models [40], which adds up to the computation complexity. Recent works explore efficient model architectures and attention mechanisms for image diffusion models [8, 64, 70]. However, there is little effort in the literature dedicated to accelerating and deploying video models at scale, especially for mobile devices." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.908, + 0.902 + ], + "angle": 0, + "content": "In this work, we systematically investigate the redundancies in video diffusion models and propose a mobile" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.875, + 0.217, + 0.888 + ], + "angle": 0, + "content": "Equal contribution" + }, + { + "type": "page_footnote", + "bbox": [ + 0.111, + 0.888, + 0.239, + 0.9 + ], + "angle": 0, + "content": "\\(\\ddagger\\) Corresponding authors" + }, + { + "type": "list", + "bbox": [ + 0.109, + 0.875, + 0.239, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2479" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.089, + 0.092, + 0.482, + 0.272 + ], + "angle": 0, + "content": "acceleration framework. First, we obtain an efficient spatial backbone by following prior works [11, 29] to prune a pre-trained text-to-image diffusion model [45]. The pruned model achieves \\(2.5 \\times\\) size compression and more than \\(10 \\times\\) speedup compared to Stable Diffusion v1.5 [45], while maintaining comparable generative quality. Starting from a pre-trained image model offers two key benefits: (i) it eliminates the need for costly large-scale pre-training, and (ii) with a compact image model, we can significantly narrow the search space in subsequent stages, focusing only on optimizing temporal layers, and thereby accelerating the discovery of the final model architecture." + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.276, + 0.486, + 0.621 + ], + "angle": 0, + "content": "Even with the efficient image backbone, applying previous temporal inflation methods [3, 4, 13] still results in tremendous computation cost and encounters out-of-memory issues on mobile. Thus, our second stage is to systematically investigate different types of temporal layers and perform a latency-memory joint search to determine the spatial-temporal architecture for efficient mobile deployment. Unlike previous methods [3, 4, 13] that typically rely on a specific type of temporal modeling layer, we investigate all possible designs, including temporal attention, spatial-temporal full attention (3D attention), temporal cross attention, and temporal convolutions (Conv3D). Besides, our search space includes the position (resolution) to apply these temporal layers, and number of layers to use (in Sec. 3.2). We profile the computation, memory footprint, and on-device latency of architecture candidates, and perform evolutionary search to discover the architecture with Pareto optimality of speed and quality. The searched network is only 0.6B in size, and can generate a 5-second video clip on iPhone 16 PM without hitting memory bound. In contrast, all prior video diffusion models fail to run on mobile, even for the smallest open-sourced ones like the 16-frameAnimateDiff [13] and 14-frame SVD [3] (Tab. 1)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.624, + 0.483, + 0.745 + ], + "angle": 0, + "content": "Finally, to further speed up generation on mobile, we distill our efficient video diffusion model with a tailored adversarial fine-tuning method capable of image-video mixed training. We reduce the number of denoising steps from 25 to 4, and eliminate classifier-free guidance [16], leading to more than \\(12 \\times\\) speedup without performance degradation. As in Tab. 1, our mobile speed is faster than most GPU-deployed (e.g., on A100) counterparts." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.762, + 0.48, + 0.844 + ], + "angle": 0, + "content": "
ModelTypeStepsParams (B)A100 (s)iPhone (s)Vbench (↑)
OpenSora-v1.2DiT301.231.00✘79.76
CogVideoX-2BDiT501.654.09✘80.91
AnimateDiff-V2UNet251.29.04✘80.27
AnimateDiffLCMUNet41.21.77✘79.42
OursUNet40.60.474.1281.14
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.855, + 0.483, + 0.897 + ], + "angle": 0, + "content": "Table 1. Comparison of size (number of parameters), speed (tested on NVIDIA A100 and iPhone 16 Pro Max), and performance (on VBench [19]) for various models." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.196 + ], + "angle": 0, + "content": "With the proposed framework, we successfully deploy our 0.6B text-to-video model on an iPhone 16 Pro Max, achieving the generation of a 5-second video clip within 5 seconds. This work represents not only the very first mobile deployment attempt of the video diffusion model, but also demonstrates its real-time potential1. Our contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.198, + 0.905, + 0.272 + ], + "angle": 0, + "content": "- Through image-video joint training, spatial and temporal architecture design, and mobile-driven latency-memory joint architecture search, we develop a comprehensive mobile acceleration framework for the text-to-video diffusion model." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.273, + 0.905, + 0.332 + ], + "angle": 0, + "content": "- We propose an adversarial fine-tuning technique tailored for video diffusion models. Despite the already compact nature of our mobile denoiser, we further distill it to 4 denoising steps with superior quality." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.333, + 0.905, + 0.392 + ], + "angle": 0, + "content": "- Our work is the very first one to show the possibility of real-time text-to-video generation on mobile devices, unlocking the possibility of deploying applications of video diffusion models at scale." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.198, + 0.905, + 0.392 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.407, + 0.655, + 0.422 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.432, + 0.907, + 0.734 + ], + "angle": 0, + "content": "Video Diffusion Models. Denoising Diffusion Probabilistic Models [17] is the trending paradigm for building video diffusion models, demonstrating photorealistic quality and generic generation capabilities. Pioneer works often start from a pre-trained text-to-image diffusion model [45] and insert temporal layers to model motions along frame sequence [3, 4, 13, 62, 75]. In addition, training-free noise tuning techniques are proposed to ease the alignment between frames [23, 34, 43, 63, 73]. Later, with the emergence of large-scale, high-quality video datasets [7, 38] and Transformer backbones [40], subsequent works curate their own dataset and build large video diffusion models with exceptional quality, such as the open-sourced CogVideoX [67], Mochi 1 [55], PyramidalFlow [21], Allegro [78], and close-sourced ones including Hailuo [37], Runway Gen 3 Alpha [46], Kling [24], Luma Dream Machine [1], Pika 1.5 [2], Sora [39], and MovieGen [41]. Remarkably, the open-sourced projects Open-Sora [77] and Open-Sora-Plan [26] provide the community with reliable implementations to replicate large-scale video diffusion models." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.906, + 0.855 + ], + "angle": 0, + "content": "Dividing by task type, video diffusion models can be categorized into text-to-video generation [5, 6, 13, 15, 18, 21, 27, 28, 42, 55, 56, 67, 78], image-to-video generation [3, 79], or specific motion controls [14, 44, 61, 71, 76]. Though some work [8, 29, 70, 73] aim to improve the efficiency of diffusion models, the acceleration for mobile deployment of video diffusion models is still in absent. Popular video models [21, 55, 67] can only run on server-level" + }, + { + "type": "page_footnote", + "bbox": [ + 0.513, + 0.863, + 0.905, + 0.901 + ], + "angle": 0, + "content": "This work performs conventional T2V generation, generating an entire video at once. However, a more suitable real-time capability would involve streamlined, continuous video generation, which we leave for future work." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "2480" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.242 + ], + "angle": 0, + "content": "Step Distillation brings almost linear generation speedup for diffusion models. Early work [29, 47] progressively distill a student network to predict a further ODE location with teacher guidance, resulting in fewer inference steps, while Consistency Models [52, 53] and Rectified Flow [33] refine the prediction objective to clean data or global velocity to achieve fewer-step inference. Later works [50, 51, 65] further incorporate adversarial loss to distill a single-step student, and enhance multi-step results as well." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.243, + 0.484, + 0.379 + ], + "angle": 0, + "content": "Despite extensive research in image diffusion models [9, 22, 36, 59, 66, 68, 69], step distillation for video diffusion model is under-explored. One type of work applies consistency distillation to generate videos in 4 steps [58, 60, 72]. Another trend adopts adversarial distillation to achieve few-step (1-2) generation [32, 35, 74]. However, these methods distill pre-trained large models with enough redundancy along the trajectory, while we find them not applicable to our efficient model and yield inferior performance." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.391, + 0.182, + 0.405 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.416, + 0.484, + 0.658 + ], + "angle": 0, + "content": "Our objective is to achieve high-fidelity and temporally consistent video generation on mobile devices. However, current text-to-video diffusion models face two key challenges in reaching this goal: (a) the memory and computation requirement is beyond the capability of even the most powerful mobile chips, i.e. iPhone A18 Pro, and (b) denoising with dozens of steps to generate a single output further slows down the process. To address these challenges, we propose a three-stage framework to accelerate video diffusion models on the mobile platform. First, we prune from a pre-trained text-to-image diffusion model to obtain an efficient spatial backbone. Second, we inflate the spatial backbone with a novel combination of temporal modules which are searched out with our mobile-oriented metrics. Finally, through adversarial training, our efficient model attains the capability to generate high-quality videos in only 4 steps." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.666, + 0.232, + 0.68 + ], + "angle": 0, + "content": "3.1. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.687, + 0.483, + 0.825 + ], + "angle": 0, + "content": "Following [77], we employ a spatial-temporal VAE to compress image and video data into the latent space. Given video or image data \\(\\mathbf{v} \\in \\mathbb{R}^{n \\times 3 \\times H \\times W}\\), where \\(n\\) is the number of frames with height \\(H\\) and width \\(W\\), the spatial-temporal encoder, \\(\\mathbf{E}\\), maps the data to a latent space. The encoded frames are represented as \\(\\mathbf{x}_0 = \\mathbf{E}(\\mathbf{v})\\), resulting in \\(\\mathbf{x}_0 \\in \\mathbb{R}^{\\tilde{n} \\times 4 \\times \\tilde{H} \\times \\tilde{W}}\\). Here, \\(\\mathbf{x}_0 \\sim p_{data}(\\mathbf{x}_0)\\) is a 4-channel latent, with a temporal compression of \\(\\tilde{n} = n/4\\), and spatial compression as \\(\\tilde{H} = H/8\\), \\(\\tilde{W} = W/8\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.484, + 0.871 + ], + "angle": 0, + "content": "We follow Rectified Flow [59] to train our latent diffusion model. According to the flow-matching-based diffusion form, the intermediate noisy state \\(\\mathbf{x}_t\\) is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.144, + 0.885, + 0.483, + 0.902 + ], + "angle": 0, + "content": "\\[\n\\mathbf {x} _ {t} = (1 - t) \\mathbf {x} _ {0} + t \\epsilon , \\text {w h e r e} \\epsilon \\sim \\mathcal {N} (0, I), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.152 + ], + "angle": 0, + "content": "which is a linear interpolation between the data distribution and a standard normal distribution. The model aims to learn a vector field \\( v_{\\theta}(t, \\mathbf{x}_t) \\) using the Conditional Flow Matching objective, i.e.," + }, + { + "type": "equation", + "bbox": [ + 0.578, + 0.161, + 0.905, + 0.18 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathbb {E} _ {t, \\epsilon , \\mathbf {x} _ {0}} \\| v _ {\\theta} (t, \\mathbf {x} _ {t}) - u _ {t} (\\mathbf {x} _ {t} | \\mathbf {x} _ {0}) \\| _ {2} ^ {2}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.189, + 0.905, + 0.22 + ], + "angle": 0, + "content": "where \\(u_{t}(\\mathbf{x}_{t}|\\mathbf{x}_{0}) = \\epsilon -\\mathbf{x}_{0}\\) . Following [10], during training, we sample \\(t\\) from a logit-normal distribution, i.e.," + }, + { + "type": "equation", + "bbox": [ + 0.517, + 0.229, + 0.905, + 0.281 + ], + "angle": 0, + "content": "\\[\n\\pi (t; m, s) = \\frac {1}{s \\sqrt {2 \\pi}} \\frac {1}{t (1 - t)} \\exp \\left(- \\frac {\\left(\\operatorname {l o g i t} (t) - m\\right) ^ {2}}{2 s ^ {2}}\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.281, + 0.905, + 0.313 + ], + "angle": 0, + "content": "where \\(\\mathrm{logit}(t) = \\log \\frac{t}{1 - t}\\), \\(m\\) and \\(s\\) are the location parameter and scale parameter, respectively." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.32, + 0.808, + 0.336 + ], + "angle": 0, + "content": "3.2. Hardware Efficient Model Design" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.342, + 0.907, + 0.569 + ], + "angle": 0, + "content": "Spatial Backbone. We follow [11, 29] to first prune an efficient text-to-image model as the spatial backbone. Specifically, we start from Stable Diffusion v1.5 [45], and borrow the knowledge from prior arts [29] to remove the most mobile-unfriendly attentions. We then prune the network depth and width following [11] and achieve \\(\\times 2.5\\) size reduction and more than \\(10\\times\\) speedup on mobile devices. We include qualitative visualizations of our image model in the supplementary material. Note that we use a UNet denoiser [17], leaving the exploration of DiT [40] to future work. The hierarchical structure of the UNet denoiser forms a good search space to achieve mobile efficiency, while the computation complexity of DiT grows quadratically with the number of tokens (generation resolution), making it challenging for mobile deployment." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.57, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Temporal Layer Design. Current latent video diffusion models typically adopt temporal self-attention [13], cross-attentions [77], and convolutions [3] to model temporal dependencies. CogVideoX [67] demonstrates significant performance gain by using full 3D-Attention, at the cost of more computations and memory consumption. In this work, we enumerate and investigate all types of temporal modeling methods, including Conv1D, Conv3D, SelfAttention1D, SelfAttention3D, CrossAttention1D, and CrossAttention3D, and profile their complexity in Fig. 3. For instance, SelfAttention1D only models temporal dependency on a single coordinate, while SelfAttention3D models global dependencies and has the potential to deliver much stronger performance. However, the computation complexity of SelfAttention3D grows quadratically with respect to \\(\\tilde{t} \\times \\tilde{H} \\times \\tilde{W}\\), while SelfAttention1D is linear with respect to \\(\\tilde{H}\\) and \\(\\tilde{W}\\), which makes SelfAttention3D much more costly at higher resolutions. On the other hand, the computation of CrossAttentionND is determined by both spatial-temporal resolution and the number of tokens from the text encoder. Conv1D and 3D are locality alternatives for SelfAttention1D and 3D, respectively. Though the computation complexity and" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "2481" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.118, + 0.087, + 0.909, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.279, + 0.908, + 0.378 + ], + "angle": 0, + "content": "Figure 2. Framework Overview. In the Latency and Memory Guided Architecture Search, we freeze the pretrained efficient spatial layers and conduct evolutionary search over temporal layer based on the memory and latency constraint. During the Adversarial Fine-tuning stage, we initialize the discriminator with the weights from the text-to-video model trained in the first stage. The discriminator employs the encoder of the UNet as its backbone, which remains frozen. We add spatial-temporal discriminator heads after each downsampling block, updating only these heads during training. Following prior works [48, 49, 74], each head conditions on pooled text embeddings c projected via a linear layer. Input features are first reshaped to merge the temporal and batch axes for processing through a 2D ResBlock, and then reshaped again to merge spatial dimensions before the temporal self-attention block." + }, + { + "type": "image", + "bbox": [ + 0.091, + 0.399, + 0.482, + 0.519 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.529, + 0.483, + 0.586 + ], + "angle": 0, + "content": "Figure 3. Computation Complexity and Memory Consumption Analysis. The computation complexity and memory consumption of different temporal layer for various input size. The temporal dimension is fixed to 12 for simplicity." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.599, + 0.483, + 0.703 + ], + "angle": 0, + "content": "memory footprint for each design candidate can be easily profiled, as in Fig. 3, it is still crucial and challenging to build a spatial-temporal network with optimized arrangements of these operators. We propose to perform a latency-memory joint architecture search to determine which, where, and the number of temporal layers to use for our efficient video diffusion model on mobile, as follows." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.705, + 0.468, + 0.719 + ], + "angle": 0, + "content": "Latency and Memory Guided Architecture Search." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.72, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Prior to searching the architecture, we construct a look-up table containing the inference latency and the memory footprint of different temporal layers. For each candidate 1D or 3D operator (i.e., ConvND, SelfAttentionND, CrossAttentionND), we benchmark the latency and memory consumption under different spatial-temporal resolutions on hardware. We then clean the search space by eliminating OOM states. Then we perform evolutionary search to obtain the temporal design with Pareto optimality. The architecture candidate is trained on precomputed video latents for \\(20K\\) iterations with the spatial backbone frozen, and is evaluated on VBench [19] to obtain the scores as the quality metric." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.403, + 0.905, + 0.433 + ], + "angle": 0, + "content": "We include the detailed search algorithm, action space, and total search time in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.433, + 0.906, + 0.538 + ], + "angle": 0, + "content": "Image-Video Joint Training. Upon the finalized model architecture, we perform image-video joint training under various clip lengths and aspect ratios with all parameters updated for another \\(100K\\) iterations. After the joint training, our efficient model is capable of generating videos with various lengths and aspect ratios under a conventional recipe, i.e., 25 steps with classifier-free guidance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.539, + 0.907, + 0.751 + ], + "angle": 0, + "content": "VAE Decoder Compression. We use the spatial temporal-decoupled VAE from OpenSora [77], which has 4 latent channels, \\(8 \\times 8\\) spatial compression and \\(4 \\times\\) temporal compression. To decode a 17-frame video clip on mobile, the original [77] temporal decoder takes \\(23,100~\\mathrm{ms}\\), and the spatial decoder takes \\(4100~\\mathrm{ms}\\), which we found to be a bottleneck for the generation speed. To increase the speed, we focus on the decoder only. We freeze the VAE encoders and prune the temporal and spatial decoder on our video and image dataset, respectively. Our efficient temporal decoder runs at \\(210~\\mathrm{ms}\\) and spatial decoder at \\(330~\\mathrm{ms}\\) to decode a 17-frame video clip, reaching \\(50 \\times\\) speedup with negligible loss in quality. Further details about VAE compression are included in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.758, + 0.791, + 0.774 + ], + "angle": 0, + "content": "3.3. Latent Adversarial Fine-tuning" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.781, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Our training procedure involves two networks: a generator \\(\\mathcal{G}_{\\theta}\\) and a discriminator \\(\\mathcal{D}\\phi\\). Similar to prior work [51, 74], we initialize our generator with pre-trained diffusion model weights \\(\\theta\\), while the discriminator is also partially initialized from \\(\\theta\\). Specifically, the backbone of the discriminator adopts the same architecture and weights as the pre-trained UNet encoder, with these backbone parameters remaining frozen during training. Additionally, we enhance the dis" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "2482" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.212 + ], + "angle": 0, + "content": "criminator with spatial-temporal discriminator heads added after each backbone block, with only these head parameters being updated in the discriminator training phase. As illustrated on the right in Fig. 2, each discriminator head consists of a spatial ResBlock and a temporal self-attention block. This design allows our discriminator to effectively handle both image and video data during fine-tuning. We analyze the impact of joint image-video fine-tuning in Sec. 4.3." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.213, + 0.482, + 0.318 + ], + "angle": 0, + "content": "For a real data sample \\(\\mathbf{x}_0\\), a noisy data point \\(\\mathbf{x}_t\\) is generated through a forward diffusion process, as described in Eq. (1). We set intermediate timesteps as \\(0 < T_k < \\dots < T_1 = 1.0\\) and sample \\(t\\) from these timesteps, where \\(k\\) is the number of timesteps selected for generator training (set to \\(k = 4\\) in practice). The generator, then, predicts the velocity at \\(\\mathbf{x}_t\\) as \\(\\mathcal{G}_{\\theta}(\\mathbf{x}_t,t)\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.318, + 0.482, + 0.455 + ], + "angle": 0, + "content": "To train the discriminator, we first sample a target timestep \\(t'\\) from a logit-normal distribution, as shown in Eq. (3). Using the forward process in Eq. (1), we obtain the real sample \\(\\mathbf{x}_{t'} = (1 - t')\\mathbf{x}_0 + t'\\epsilon\\). The fake sample, \\(\\hat{\\mathbf{x}}_{t'}\\), is computed as \\(\\hat{\\mathbf{x}}_{t'} = \\mathbf{x}_t + (t' - t)\\cdot \\mathcal{G}_\\theta (t,\\mathbf{x}_t)\\), as shown in Fig. 4. Following established approaches [48-50, 74], we employ hinge loss [31] as the adversarial objective to enhance performance. The discriminator's goal is to differentiate between real and fake samples by minimizing:" + }, + { + "type": "equation", + "bbox": [ + 0.146, + 0.465, + 0.482, + 0.502 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {\\mathrm {a d v}} ^ {\\mathcal {D}} = \\mathbb {E} _ {t ^ {\\prime}, x _ {0}} [ \\max (0, 1 + \\mathcal {D} _ {\\phi} (\\mathbf {x} _ {t ^ {\\prime}}, t ^ {\\prime})) ] \\tag {4} \\\\ + \\mathbb {E} _ {t, t ^ {\\prime}, x _ {0}} \\left[ \\max \\left(0, 1 - \\mathcal {D} _ {\\phi} \\left(\\hat {\\mathbf {x}} _ {t ^ {\\prime}}, t ^ {\\prime}\\right)\\right) \\right], \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.111, + 0.513, + 0.482, + 0.528 + ], + "angle": 0, + "content": "The adversarial objective for the generator is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.192, + 0.541, + 0.482, + 0.561 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {a d v}} ^ {\\mathcal {G}} = \\mathbb {E} _ {t, t ^ {\\prime}, x _ {0}} \\left[ \\mathcal {D} _ {\\phi} \\left(\\hat {\\mathbf {x}} _ {t ^ {\\prime}}, t ^ {\\prime}\\right) \\right]. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.566, + 0.483, + 0.597 + ], + "angle": 0, + "content": "Following [74], we also incorporate a reconstruction objective to enhance stability, defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.175, + 0.609, + 0.483, + 0.634 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {r e c o n}} = \\sqrt {\\left\\| \\hat {\\mathbf {x}} _ {0} - \\mathbf {x} _ {0} \\right\\| _ {2} ^ {2} + c ^ {2}} - c, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.644, + 0.483, + 0.673 + ], + "angle": 0, + "content": "where \\(\\hat{\\mathbf{x}}_0 = \\mathbf{x}_t - t\\cdot \\mathcal{G}_{\\theta}(t,\\mathbf{x}_t)\\), and \\(c > 0\\) is an adjustable constant." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.675, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Discussion. Our latent adversarial training pipeline is inspired by SF-V [74]. Similar to SF-V, we set \\( k = 4 \\) and utilize the part of the pre-trained diffusion model as the backbone for the discriminator. However, our approach introduces several key differences. First, our method is built on an efficient UNet specifically designed for mobile devices, with fewer parameters than SVD [3], making it a more challenging task. Second, we redesign the discriminator heads: instead of using separate spatial and temporal heads, we integrate them into a unified spatial-temporal head for adversarial training. Rather than handling the temporal dimension separately with 1-D convolutional kernels, we incorporate a temporal self-attention layer into the spatial discriminator head after the 2-D ResBlock, forming a spatial-temporal discriminator head. This unified design enables" + }, + { + "type": "image", + "bbox": [ + 0.554, + 0.088, + 0.868, + 0.192 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.202, + 0.907, + 0.314 + ], + "angle": 0, + "content": "Figure 4. Latent Adversarial Fine-tuning. Given a latent \\(\\mathbf{x}_0\\) and a noise latent \\(\\epsilon\\), we obtain the intermediate noisy latent \\(\\mathbf{x}_t\\) through a forward diffusion process. The generator then predicts the velocity as \\(\\mathcal{G}_{\\theta}(\\mathbf{x}_t, t)\\). Using the predicted velocity, we compute \\(\\hat{\\mathbf{x}}_0\\) and calculate the reconstruction loss \\(\\mathcal{L}_{\\mathrm{recon}}\\) between \\(\\mathbf{x}_0\\) and \\(\\hat{\\mathbf{x}}_0\\). For adversarial training, the real sample \\(\\mathbf{x}_{t'}\\) is obtained using the same forward diffusion process, while the fake sample \\(\\hat{\\mathbf{x}}_{t'}\\) is computed using \\(\\mathbf{x}_t\\) and the predicted velocity \\(\\mathcal{G}_{\\theta}(\\mathbf{x}_t, t)\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.354, + 0.907, + 0.4 + ], + "angle": 0, + "content": "our model to be jointly trained on both image and video data, which, as demonstrated in Sec. 4.3, significantly enhances the performance of the fine-tuned model." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.433, + 0.645, + 0.45 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.465, + 0.907, + 0.646 + ], + "angle": 0, + "content": "Training. The efficient image backbone is obtained by pruning the Stable Diffusion v1.5 UNet for \\(100K\\) iterations on high-quality synthetic image datasets. The model is then fine-tuned for \\(50K\\) additional iterations to adapt to Rectified-Flow velocity prediction [10] as well as to the Spatial-Temporal VAE [77]. We incorporate QK-norm and RoPE [54] in our network to stabilize training. The workflow for architecture search is discussed in Sec. 3.2. The image model pruning, temporal architecture search, and final model training are conducted on 256 NVIDIA A100 80G GPUs using AdamW optimizer with \\(5e - 5\\) learning rate and betas values as [0.9, 0.999]." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.652, + 0.907, + 0.773 + ], + "angle": 0, + "content": "Adversarial Fine-tuning is conducted for \\(6K\\) iterations on 64 NVIDIA A100 GPUs, using the AdamW optimizer with a learning rate of \\(1e - 7\\) for the generator (i.e., UNet) and \\(1e - 4\\) for the discriminator heads. We set the betas as [0.9, 0.999] for the generator optimizer, and [0.5, 0.999] for the discriminator optimizer. We set the EMA rate as 0.95 following SF-V [74]. We set \\(m = -1\\), \\(s = 1\\) if not otherwise noted." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Evaluation. The model is evaluated following the standard benchmarking procedure of VBench [19]. With the 4-step adversarial distilled model, we generate 120-frame horizontal videos at a resolution of \\(432 \\times 768\\) using 4 inference steps without employing classifier-free guidance. The generated video is saved at 5 seconds 24 FPS for score testing and qualitative visualization. The mobile demo and detailed demo settings are included in the supplementary material." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2483" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.13, + 0.08, + 0.868, + 0.58 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.59, + 0.907, + 0.632 + ], + "angle": 0, + "content": "Figure 5. Video generation on various domains. We employ our model to synthesize videos across diverse domains, with each video containing 120 frames at a resolution of \\(432 \\times 768\\). All results are generated through a 4-step inference process. The results demonstrate that our model can produce high-quality, motion-consistent videos featuring various objects across different domains." + }, + { + "type": "table", + "bbox": [ + 0.12, + 0.634, + 0.885, + 0.805 + ], + "angle": 0, + "content": "
ModelParams (B)StepsTotal scoreQuality scoreSemantic scoretemporal flickeringmotion smoothnessdynamic degreeaesthetic qualityobject classmultiple objectscolorspatial relationshipappearance styletemporal styleoverall consistency
Kling--81.8583.3975.6899.3099.4046.9461.2187.2468.0589.9073.0319.6224.1726.42
Pyramid Flow2.02081.7284.7469.6299.4999.1264.6363.2686.6750.7182.8759.5320.9123.0926.23
CogVideoX5.05081.6182.7577.0498.6696.9270.9761.9885.2362.1182.8166.3524.9125.3827.59
T2V-Turbo1.6481.0182.5774.7697.4897.3449.1763.0493.9654.6589.9038.6724.4225.5128.16
Emu38.0-80.9684.0968.4398.5798.9379.2759.6486.1744.6488.3468.7320.9223.2624.79
CogVideoX1.65080.9182.1875.8398.8997.7359.8660.8283.3762.6379.4169.9024.8024.3626.66
Pika-1.0--80.6982.9271.7799.7499.5047.5062.0488.7243.0890.5761.0322.2624.2225.94
VideoCrafter-2.01.95080.4482.2073.4298.4197.7342.5063.1392.5540.6692.9235.8625.1325.8428.23
AnimateDiff-V21.22580.2782.9069.7598.7597.7640.8367.1690.9036.8887.4734.6022.4226.0327.04
OpenSora V1.21.23079.7681.3573.3999.5398.5042.3956.8582.2251.8390.0868.5623.9524.5426.85
AnimateLCM1.2479.4282.3667.6598.5298.1640.5667.0191.4129.7684.2437.1420.9725.1626.07
ModelScope1.45075.7578.0566.5498.2895.7966.3952.0682.2538.9881.7233.6823.3925.3725.67
Ours0.6481.1483.4771.8499.3799.2951.1162.1992.2254.3486.6356.2020.1725.1827.42
" + }, + { + "type": "table_caption", + "bbox": [ + 0.218, + 0.814, + 0.779, + 0.829 + ], + "angle": 0, + "content": "Table 2. Performance comparison with popular video generation models on VBench [19]." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.832, + 0.318, + 0.849 + ], + "angle": 0, + "content": "4.1. Qualitative Visualization" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.903 + ], + "angle": 0, + "content": "We show visualizations of our generated videos in Fig. 5. Our model consistently produces high-quality video frames and smooth object movements. To demonstrate the generic" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.833, + 0.907, + 0.895 + ], + "angle": 0, + "content": "text-to-video generation ability, we show various generation examples, including human, animal, photorealistic and art-styled scenes. We include more video visualizations in the supplementary material." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2484" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.332, + 0.108 + ], + "angle": 0, + "content": "4.2. Quantitative Comparisons" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.113, + 0.483, + 0.293 + ], + "angle": 0, + "content": "We present a comprehensive evaluation of our method against existing popular video generation models on VBench [19], as in Tab. 2. Despite the fact that our model is compact and designated for fast inference on mobile, it achieves higher total score compared to recent arts, including the DiT-based OpenSora-V1.2, CogVideoX-2B [67], and the UNet-based VideoCrafter-2.0 [6]. In addition, compared to the 4-step distilled T2V-Turbo [27] and AnimateLCM [58], our model achieves better performance with more than \\(50\\%\\) reduction in size. The quantitative scores demonstrate the superiority of our efficient model design and the tailored adversarial distillation method." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.294, + 0.484, + 0.386 + ], + "angle": 0, + "content": "User Study. Human evaluations are conducted between our model and baselines as in the Tab. 3. We generate videos from VBench and Movie Gen Bench prompts and ask human labelers to pick the best results across prompt alignment, aesthetics, and motion. The result shows that our model outperforms baseline metrics by a large margin." + }, + { + "type": "table", + "bbox": [ + 0.113, + 0.396, + 0.465, + 0.455 + ], + "angle": 0, + "content": "
Win RatePrompt AlignmentAestheticsMotion
OpenSora-1.224.9%21.2%20.4%
CogVideoX-2B30.7%30.5%30.3%
Ours44.4%48.3%49.3%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.466, + 0.483, + 0.495 + ], + "angle": 0, + "content": "Table 3. User Study between OpenSora-1.2, CogVideoX-2B, and our model on prompt alignment, aesthetics, and motion." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.517, + 0.265, + 0.533 + ], + "angle": 0, + "content": "4.3. Ablation Analysis" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.539, + 0.483, + 0.674 + ], + "angle": 0, + "content": "Comparison of Training Data Scheme. We compare the model trained with joint image-video datasets vs. video-only datasets. As shown in Tab. 4, training with video-only datasets leads to significant performance degradation on the VBenchmark score with a drop of 6.32 in aesthetic quality, 2.80 in image quality, and 2.54 in total score. The results highlight the importance of joint image-video training, as the image dataset offers more contextual information and enhances diversity." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.675, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Model Scaling. We scale up and down the model size by adjusting the number of temporal layers, as shown in Tab. 4, to demonstrate the effectiveness of the proposed temporal architecture search. We can observe that scaling up the model can only marginally improve generative scores (i.e., \\(\\times 2\\) scale-up only increasing 0.56 in dynamic degree and 0.12 in motion smoothness, and 0.27 in total score). However, both the \\(\\times 2\\) and \\(\\times 4\\) models hit the memory bound on iPhone. By dividing the \\(\\times 2\\) model into more chunks, we test its mobile speed and observe nearly doubled latency. While on the other hand, further scaling down the model results in heavy losses in generation quality (i.e., decreasing 0.56 in dynamic degree and 0.61 in motion smoothness). Our efficient model is a balanced sweet point for quality and on-device performance." + }, + { + "type": "table", + "bbox": [ + 0.521, + 0.09, + 0.905, + 0.162 + ], + "angle": 0, + "content": "
w/ I ScalingΔTΔAQΔIQΔDDΔMSΔTotal
Video-onlyX——-6.32-2.80+4.6-0.17-2.54
×0.5-22%-1.90+0.04-0.56-0.61-1.58
Temporal Scaling✓×2+32%+0.37+0.10+0.56+0.12+0.27
×4+107%+0.49-0.17+0.91+0.17+0.31
" + }, + { + "type": "table_caption", + "bbox": [ + 0.512, + 0.173, + 0.907, + 0.311 + ], + "angle": 0, + "content": "Table 4. Analysis of training data scheme, latency, and quality of efficient architecture. The baseline model adopts the best suitable architecture and is trained with joint image-video datasets. The \"scaling\", and \"T\" indicates the number of the temporal layers, and the latency comparing to the baseline. The \"AQ\", \"IQ\", \"DD\", and \"MS\", are aesthetic quality, image quality, dynamic degree, and motion smoothness in VBench [19]. The benchmark metrics are presented as the differences from the baseline model, where negative values indicate a decrease in performance over the baseline vice versa." + }, + { + "type": "image", + "bbox": [ + 0.554, + 0.32, + 0.868, + 0.507 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.523, + 0.907, + 0.579 + ], + "angle": 0, + "content": "Figure 6. Analysis of temporal layers. We ablate temporal layers in different network stages to evaluate their effect. The scores are normalized according to our base model to better demonstrate the difference." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.599, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Effect of Different Temporal Layers. To better understand the roles of the searched temporal layers, we compute the VBench score after systematically removing (i) all temporal layers in the downsample stage, (ii) bottleneck temporal layers, and (iii) all temporal layers in the upsample stage. The model makes reasonable zero-shot generations after temporal layer removal, but we still fine-tune 10K iterations under the same recipe as in Sec. 3.2 for fair comparison. As shown in Fig. 6, removing different temporal layers results in varying degrees of performance degradation across different metrics, and all removing strategies result in a substantial drop in total score, demonstrating that the existence of the searched temporal layers is important, and they play different roles in generation. Specifically, removing temporal layers in upsample blocks results in a more significant loss in imaging quality, subject consistency, and background consistency, suggesting that the up temporal layers play important roles in detail reconstruction. In contrast, bottleneck layers are more important in human action and object class, where global information modeling dom" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2485" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.168 + ], + "angle": 0, + "content": "inates the results. We observe that removing down layers introduces less overall degradations compared to the other two, which is an anticipated phenomenon because the loss of modeling capacity can be mitigated by the subsequent bottleneck and up stage temporal layers after fine-tuning." + }, + { + "type": "table", + "bbox": [ + 0.106, + 0.182, + 0.472, + 0.29 + ], + "angle": 0, + "content": "
Headw/Iw/ LadvmTFMOQSTotal
SF-V [74]X✓-198.8630.7583.6064.2579.73
Ours99.2737.8583.6169.0180.69
Ours✓X-197.321.0170.0317.0459.44
-299.2947.6483.9170.5481.23
Ours✓✓-199.3754.3483.4771.8481.14
099.4443.6381.6969.4679.24
199.5624.4177.8458.8874.05
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.09, + 0.3, + 0.485, + 0.412 + ], + "angle": 0, + "content": "Table 5. Analysis of our adversarial fine-tuning scheme. We evaluate the VBench [19] scores for our models using different training schemes. In the results, \"TF\" and \"MO\" denote the temporal flickering and multiple objects sub-scores, respectively, while \"Q\" and \"S\" represent the quality and semantic scores. The table summarizes how varying (1) the type of discriminator head, (2) training with or without an image dataset, and (3) different \\( t' \\) distributions impact the performance of our model." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.43, + 0.483, + 0.596 + ], + "angle": 0, + "content": "Effect of Discriminator Heads. We compare the effects of our spatial-temporal discriminator heads with the separate spatial and temporal heads proposed in SF-V [74] to demonstrate the effectiveness of our discriminator architecture. For a fair comparison, both models are trained exclusively on video data. We evaluate the models finetuned with different discriminator heads on VBench [19]. As shown in the first two rows of Tab. 5, our discriminator heads result in improvements in both the quality score (83.61 vs. 83.60) and the semantic score (69.01 vs. 64.25) of the generated videos." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.598, + 0.483, + 0.779 + ], + "angle": 0, + "content": "Effect of Joint Image Video Fine-tuning. We examine the impact of incorporating image data during adversarial training, as shown in the second and fourth rows of Tab. 5. The results indicate that fine-tuning the model with image data can slightly decrease the Quality score in the VBench [19] evaluation (from 83.61 to 83.47). However, by leveraging the increased diversity of the image dataset, the model achieves a substantial improvement in semantic performance, particularly in multi-object generation (from 37.85 to 54.34). This enhancement leads to a better overall score compared to the model trained exclusively on video data (81.14 vs. 80.69)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Effect of Noise Distribution for Discriminator. Following Eq. (3), the parameters \\( m \\) and \\( s \\) control the distribution of \\( t' \\), which determines the noise levels of \\( \\mathbf{x}_{t'} \\) and \\( \\hat{\\mathbf{x}}_{t'} \\) before they are passed to the discriminator as real and fake samples, respectively. We investigate the effect of different noise distributions on model performance by evaluating the results using VBenchmark [19]. As shown in the last four rows of Tab. 5, increasing \\( m \\) (resulting in noisier real and fake sam" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.228 + ], + "angle": 0, + "content": "plies) degrades the quality score (from 83.91 to 77.84) while slightly enhancing the temporal flickering score (from 99.29 to 99.56). Although setting \\( m = -2 \\) achieves the highest overall score among the experiments (81.23), it performs poorly on multi-object generation. Therefore, in most of our experiments, unless otherwise stated, we use \\( m = -1 \\). This setting yields a slightly lower overall score (81.14) but significantly improves semantic performance (71.84 vs. 70.54) and excels in multi-object generation (54.34 vs. 47.64)." + }, + { + "type": "table", + "bbox": [ + 0.52, + 0.241, + 0.901, + 0.308 + ], + "angle": 0, + "content": "
# StepsDDOCAQQualitySemanticTotal
111.3979.2452.8574.4362.7072.09
230.2888.9759.7480.0869.2877.92
451.1192.2262.1983.4771.8481.14
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.513, + 0.318, + 0.905, + 0.375 + ], + "angle": 0, + "content": "Table 6. Analysis of the number of inference steps. We measure VBench [19] score with different numbers of inference steps. In the results, \"DD\", \"OC\", and \"AQ\" denote the dynamic degree, object class, and aesthetic quality scores, respectively." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.391, + 0.907, + 0.602 + ], + "angle": 0, + "content": "Effect of Inference Steps. Our fine-tuned model supports generation with a reduced number of inference steps. We further investigate how varying the number of evaluation steps affects the quality of the generated results, as shown in Tab. 6. While our four-step generation achieves the best performance, even with only two steps, the model still produces reasonable results. Increasing the number of inference steps improves the performance of our model across all metrics. In Tab. 6, we report not only the quality and semantic scores but also scores for dynamic degree, object class, and aesthetic quality. However, reducing the process to a single inference step leads to a significant drop in performance. We leave more aggressive step reductions to future work." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.618, + 0.763, + 0.634 + ], + "angle": 0, + "content": "5. Discussion and Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.644, + 0.905, + 0.809 + ], + "angle": 0, + "content": "In this work, we propose an acceleration framework for the video diffusion model, and for the first time, achieve superfast text-to-video generation on mobile devices. Specifically, we discover an efficient but powerful network architecture through latency and memory joint architecture search for temporal layers. In addition, we propose an improved adversarial fine-tuning technique to distill our model to 4 steps to further speed up generation. Our work is a good starting point for the edge deployment of video diffusion models and we hope to inspire more downstream applications such as video extension and editing." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.905, + 0.9 + ], + "angle": 0, + "content": "Limitations. We use a public 4-channel VAE [77] to encode videos to latent space. Recent work has shown that using more latent channels benefits reconstruction details. Another future direction is to further improve the step reduction technique for 1-2 denoising steps, as works [74] on server-level models." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "2486" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.091, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.116, + 0.482, + 0.142 + ], + "angle": 0, + "content": "[1] Luma AI. Dream machine. https://lumalabs.ai/dream-machine,.2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.145, + 0.44, + 0.158 + ], + "angle": 0, + "content": "[2] Pika AI. Pika 1.5. https://pika.art/try..2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.16, + 0.483, + 0.228 + ], + "angle": 0, + "content": "[3] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. ArXiv preprint, abs/2311.15127, 2023. 1, 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.23, + 0.482, + 0.325 + ], + "angle": 0, + "content": "[4] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2023, Vancouver, BC, Canada, June 17-24, 2023, pages 22563-22575. IEEE, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.328, + 0.482, + 0.396 + ], + "angle": 0, + "content": "[5] Haoxin Chen, Menghan Xia, Yingqing He, Yong Zhang, Xiaodong Cun, Shaoshu Yang, Jinbo Xing, Yaofang Liu, Qifeng Chen, Xintao Wang, et al. Videocrafter1: Open diffusion models for high-quality video generation. ArXiv preprint, abs/2310.19512, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.398, + 0.482, + 0.478 + ], + "angle": 0, + "content": "[6] Haoxin Chen, Yong Zhang, Xiaodong Cun, Menghan Xia, Xintao Wang, Chao Weng, and Ying Shan. Videocaster2: Overcoming data limitations for high-quality video diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7310-7320, 2024. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.482, + 0.482, + 0.577 + ], + "angle": 0, + "content": "[7] Tsai-Shien Chen, Aliaksandr Siarohin, Willi Menapace, Ekaterina Deyneka, Hsiang-wei Chao, Byung Eun Jeon, Yuwei Fang, Hsin-Ying Lee, Jian Ren, Ming-Hsuan Yang, et al. Panda-70m: Captioning 70m videos with multiple cross-modality teachers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13320-13331, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.58, + 0.482, + 0.633 + ], + "angle": 0, + "content": "[8] Xinwang Chen, Ning Liu, Yichen Zhu, Feifei Feng, and Jian Tang. Edit: An efficient diffusion transformer framework inspired by human-like sketching. ArXiv preprint, abs/2410.23788, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.636, + 0.482, + 0.704 + ], + "angle": 0, + "content": "[9] Trung Dao, Thuan Hoang Nguyen, Thanh Le, Duc Vu, Khoi Nguyen, Cuong Pham, and Anh Tran. Swiftbrush v2: Make your one-step diffusion model better than its teacher. In European Conference on Computer Vision, pages 176-192. Springer, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.706, + 0.482, + 0.788 + ], + "angle": 0, + "content": "[10] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In *Forty-first International Conference on Machine Learning*, 2024. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.79, + 0.482, + 0.858 + ], + "angle": 0, + "content": "[11] Gongfan Fang, Xinyin Ma, and Xinchao Wang. Structural pruning for diffusion models. In Advances in Neural Information Processing Systems 36: Annual Conference on Neural Information Processing Systems 2023, NeurIPS 2023, New Orleans, LA, USA, December 10 - 16, 2023, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.86, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[12] Yutang Feng, Sicheng Gao, Yuxiang Bao, Xiaodi Wang, Shumin Han, Juan Zhang, Baochang Zhang, and Angela Yao. Wave: Warping ddim inversion features for zero-shot" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.116, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.094, + 0.906, + 0.12 + ], + "angle": 0, + "content": "text-to-video editing. In European Conference on Computer Vision, pages 38-55. Springer, 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.123, + 0.906, + 0.191 + ], + "angle": 0, + "content": "[13] Yuwei Guo, Ceyuan Yang, Anyi Rao, Zhengyang Liang, Yaohui Wang, Yu Qiao, Maneesh Agrawala, Dahua Lin, and Bo Dai. Animatediff:Animate your personalized text-to-image diffusion models without specific tuning. ArXiv preprint, abs/2307.04725, 2023. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.193, + 0.906, + 0.248 + ], + "angle": 0, + "content": "[14] Yuwei Guo, Ceyuan Yang, Anyi Rao, Maneesh Agrawala, Dahua Lin, and Bo Dai. Sparsectrl: Adding sparse controls to text-to-video diffusion models. In European Conference on Computer Vision, pages 330-348. Springer, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.25, + 0.906, + 0.305 + ], + "angle": 0, + "content": "[15] Agrim Gupta, Lijun Yu, Kihyuk Sohn, Xiuye Gu, Meera Hahn, Li Fei-Fei, Irfan Essa, Lu Jiang, and José Lezama. Photorealistic video generation with diffusion models. ArXiv preprint, abs/2312.06662, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.307, + 0.906, + 0.334 + ], + "angle": 0, + "content": "[16] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. ArXiv preprint, abs/2207.12598, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.337, + 0.906, + 0.404 + ], + "angle": 0, + "content": "[17] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In Advances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, NeurIPS 2020, December 6-12, 2020, virtual, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.407, + 0.906, + 0.475 + ], + "angle": 0, + "content": "[18] Wenyi Hong, Ming Ding, Wendi Zheng, Xinghan Liu, and Jie Tang. Cogvideo: Large-scale pretraining for text-to-video generation via transformers. In The Eleventh International Conference on Learning Representations, ICLR 2023, Kigali, Rwanda, May 1-5, 2023. OpenReview.net, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.477, + 0.906, + 0.573 + ], + "angle": 0, + "content": "[19] Ziqi Huang, Yinan He, Jiashuo Yu, Fan Zhang, Chenyang Si, Yuming Jiang, Yuanhan Zhang, Tianxing Wu, Qingyang Jin, Nattapol Chanpaisit, Yaohui Wang, Xinyuan Chen, Limin Wang, Dahua Lin, Yu Qiao, and Ziwei Liu. VBenchmark: Comprehensive benchmark suite for video generative models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 2, 4, 5, 6, 7, 8, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.576, + 0.906, + 0.63 + ], + "angle": 0, + "content": "[20] Hyeonho Jeong, Jinho Chang, Geon Yeong Park, and Jong Chul Ye. Dreammotion: Space-time self-similar score distillation for zero-shot video editing. ArXiv preprint, abs/2403.12002, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.633, + 0.906, + 0.699 + ], + "angle": 0, + "content": "[21] Yang Jin, Zhicheng Sun, Ningyuan Li, Kun Xu, Hao Jiang, Nan Zhuang, Quzhe Huang, Yang Song, Yadong Mu, and Zhouchen Lin. Pyramidal flow matching for efficient video generative modeling. ArXiv preprint, abs/2410.05954, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.703, + 0.906, + 0.758 + ], + "angle": 0, + "content": "[22] Beomsu Kim, Yu-Guan Hsieh, Michal Klein, Marco Cuturi, Jong Chul Ye, Bahjat Kawar, and James Thornton. Simple reflow: Improved techniques for fast flow models. ArXiv preprint, abs/2410.07815, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.76, + 0.906, + 0.801 + ], + "angle": 0, + "content": "[23] Jihwan Kim, Junoh Kang, Jinyoung Choi, and Bohyung Han. Fifo-diffusion: Generating infinite videos from text without training. ArXiv preprint, abs/2405.11473, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.803, + 0.906, + 0.828 + ], + "angle": 0, + "content": "[24] Kuaishou. Kling. https://kling.kuaishou.com/en.2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.832, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[25] Jeong-gi Kwak, Erqun Dong, Yuhe Jin, Hanseok Ko, Shweta Mahajan, and Kwang Moo Yi. Vivid-1-to-3: Novel view synthesis with video diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6775–6785, 2024. 1" + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.094, + 0.906, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2487" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.092, + 0.484, + 0.108 + ], + "angle": 0, + "content": "[26] PKU-Yuan Lab and Tuzhan AI etc. Open-sora-plan, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.091, + 0.108, + 0.484, + 0.177 + ], + "angle": 0, + "content": "[27] Jiachen Li, Weixi Feng, Tsu-Jui Fu, Xinyi Wang, Sugato Basu, Wenhu Chen, and William Yang Wang. T2v-turbo: Breaking the quality bottleneck of video consistency model with mixed reward feedback. ArXiv preprint, abs/2405.18750, 2024. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.179, + 0.482, + 0.248 + ], + "angle": 0, + "content": "[28] Jiachen Li, Qian Long, Jian Zheng, Xiaofeng Gao, Robinson Piramuthu, Wenhu Chen, and William Yang Wang. T2v-turbo-v2: Enhancing video generation model post-training through data, reward, and conditional guidance design. ArXiv preprint, abs/2410.05677, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.25, + 0.483, + 0.346 + ], + "angle": 0, + "content": "[29] Yanyu Li, Huan Wang, Qing Jin, Ju Hu, Pavlo Chemerys, Yun Fu, Yanzhi Wang, Sergey Tulyakov, and Jian Ren. Snapfusion: Text-to-image diffusion model on mobile devices within two seconds. In Advances in Neural Information Processing Systems 36: Annual Conference on Neural Information Processing Systems 2023, NeurIPS 2023, New Orleans, LA, USA, December 10 - 16, 2023, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.348, + 0.483, + 0.431 + ], + "angle": 0, + "content": "[30] Feng Liang, Bichen Wu, Jialiang Wang, Licheng Yu, Kunpeng Li, Yinan Zhao, Ishan Misra, Jia-Bin Huang, Peizhao Zhang, Peter Vajda, et al. Flowvid: Taming imperfect optical flows for consistent video-to-video synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8207-8216, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.432, + 0.483, + 0.46 + ], + "angle": 0, + "content": "[31] Jae Hyun Lim and Jong Chul Ye. Geometric gan. ArXiv preprint, abs/1705.02894, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.461, + 0.482, + 0.502 + ], + "angle": 0, + "content": "[32] Shanchuan Lin and Xiao Yang. Animatediff-lightning: Cross-model diffusion distillation. ArXiv preprint, abs/2403.12706, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.504, + 0.483, + 0.574 + ], + "angle": 0, + "content": "[33] Xingchao Liu, Chengyue Gong, and Qiang Liu. Flow straight and fast: Learning to generate and transfer data with rectified flow. In The Eleventh International Conference on Learning Representations, ICLR 2023, Kigali, Rwanda, May 1-5, 2023. OpenReview.net, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.575, + 0.483, + 0.616 + ], + "angle": 0, + "content": "[34] Yu Lu, Yuanzhi Liang, Linchao Zhu, and Yi Yang. Freelong: Training-free long video generation with spectralblend temporal attention. ArXiv preprint, abs/2407.19918, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.617, + 0.483, + 0.673 + ], + "angle": 0, + "content": "[35] Xiaofeng Mao, Zhengkai Jiang, Fu-Yun Wang, Wenbing Zhu, Jiangning Zhang, Hao Chen, Mingmin Chi, and Yabiao Wang. Osv: One step is enough for high-quality image to video generation. ArXiv preprint, abs/2409.11367, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.674, + 0.483, + 0.757 + ], + "angle": 0, + "content": "[36] Kangfu Mei, Maurizio Delbracio, Hossein Talebi, Zhengzhong Tu, Vishal M Patel, and Peyman Milanfar. Codi: Conditional diffusion distillation for higher-fidelity and faster image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9048-9058, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.759, + 0.476, + 0.772 + ], + "angle": 0, + "content": "[37] MiniMax. Hailuo ai. https://hailuoai(video/.2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.774, + 0.483, + 0.829 + ], + "angle": 0, + "content": "[38] Kepan Nan, Rui Xie, Penghao Zhou, Tiehan Fan, Zhenheng Yang, Zhijie Chen, Xiang Li, Jian Yang, and Ying Tai. Openvid-1m: A large-scale high-quality dataset for text-to-video generation. ArXiv preprint, abs/2407.02371, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.831, + 0.483, + 0.872 + ], + "angle": 0, + "content": "[39] OpenAI. Video generation models as world simulators. https://openai.com/index/video-generation-models-as-world-simulators/.2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[40] William Peebles and Saining Xie. Scalable diffusion models with transformers. In IEEE/CVF International Conference" + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.092, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "on Computer Vision, ICCV 2023, Paris, France, October 1-6, 2023, pages 4172-4182. IEEE, 2023. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.188 + ], + "angle": 0, + "content": "[41] Adam Polyak, Amit Zohar, Andrew Brown, Andros Tjandra, Animesh Sinha, Ann Lee, Apoorv Vyas, Bowen Shi, Chih-Yao Ma, Ching-Yao Chuang, et al. Movie gen: A cast of media foundation models. ArXiv preprint, abs/2410.13720, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.19, + 0.905, + 0.272 + ], + "angle": 0, + "content": "[42] Zhiwu Qing, Shiwei Zhang, Jiayu Wang, Xiang Wang, Yujie Wei, Yingya Zhang, Changxin Gao, and Nong Sang. Hierarchical spatio-temporal decoupling for text-to-video generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6635-6645, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.274, + 0.905, + 0.33 + ], + "angle": 0, + "content": "[43] Haonan Qiu, Menghan Xia, Yong Zhang, Yingqing He, Xintao Wang, Ying Shan, and Ziwei Liu. Freenoise: Tuning-free longer video diffusion via noise rescheduling. ArXiv preprint, abs/2310.15169, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.331, + 0.905, + 0.397 + ], + "angle": 0, + "content": "[44] Yixuan Ren, Yang Zhou, Jimei Yang, Jing Shi, Difan Liu, Feng Liu, Mingi Kwon, and Abhinav Shrivastava. Customize-a-video: One-shot motion customization of text-to-video diffusion models. ArXiv preprint, abs/2402.14780, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.4, + 0.907, + 0.469 + ], + "angle": 0, + "content": "[45] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.47, + 0.905, + 0.497 + ], + "angle": 0, + "content": "[46] Runway. Gen-3 alpha. https://runwayml.com/research/introducing-gen-3-alpha.2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.498, + 0.905, + 0.552 + ], + "angle": 0, + "content": "[47] Tim Salimans and Jonathan Ho. Progressive distillation for fast sampling of diffusion models. In *The Tenth International Conference on Learning Representations*, ICLR 2022, Virtual Event, April 25-29, 2022. OpenReview.net, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.554, + 0.905, + 0.635 + ], + "angle": 0, + "content": "[48] Axel Sauer, Kashyap Chitta, Jens Müller, and Andreas Geiger. Projected gans converge faster. In Advances in Neural Information Processing Systems 34: Annual Conference on Neural Information Processing Systems 2021, NeurIPS 2021, December 6-14, 2021, virtual, pages 17480-17492, 2021. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.637, + 0.905, + 0.718 + ], + "angle": 0, + "content": "[49] Axel Sauer, Tero Karras, Samuli Laine, Andreas Geiger, and Timo Aila. Stylegan-t: Unlocking the power of gans for fast large-scale text-to-image synthesis. In International Conference on Machine Learning, ICML 2023, 23-29 July 2023, Honolulu, Hawaii, USA, pages 30105-30118. PMLR, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.72, + 0.905, + 0.761 + ], + "angle": 0, + "content": "[50] Axel Sauer, Dominik Lorenz, Andreas Blattmann, and Robin Rombach. Adversarial diffusion distillation. ArXiv preprint, abs/2311.17042, 2023. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.762, + 0.905, + 0.817 + ], + "angle": 0, + "content": "[51] Axel Sauer, Frederic Boesel, Tim Dockhorn, Andreas Blattmann, Patrick Esser, and Robin Rombach. Fast high-resolution image synthesis with latent adversarial diffusion distillation. ArXiv preprint, abs/2403.12015, 2024. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.818, + 0.905, + 0.858 + ], + "angle": 0, + "content": "[52] Yang Song and Prafulla Dhariwal. Improved techniques for training consistency models. ArXiv preprint, abs/2310.14189, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[53] Yang Song, Prafulla Dhariwal, Mark Chen, and Ilya Sutskever. Consistency models. In International Conference on Machine Learning, ICML 2023, 23-29 July 2023," + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "2488" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.119 + ], + "angle": 0, + "content": "Honolulu, Hawaii, USA, pages 32211-32252. PMLR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.121, + 0.483, + 0.175 + ], + "angle": 0, + "content": "[54] Jianlin Su, Murtadha Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu. Roformer: Enhanced transformer with rotary position embedding. Neurocomputing, 568:127063, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.177, + 0.324, + 0.19 + ], + "angle": 0, + "content": "[55] Genmo Team. Mochi, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.192, + 0.483, + 0.259 + ], + "angle": 0, + "content": "[56] Gemini Team, Rohan Anil, Sebastian Borgeaud, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, Katie Millican, et al. Gemini: a family of highly capable multimodal models. ArXiv preprint, abs/2312.11805, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.261, + 0.483, + 0.344 + ], + "angle": 0, + "content": "[57] Vikram Voleti, Chun-Han Yao, Mark Boss, Adam Letts, David Pankratz, Dmitry Tochilkin, Christian Laforte, Robin Rombach, and Varun Jampani. Sv3d: Novel multi-view synthesis and 3d generation from a single image using latent video diffusion. In European Conference on Computer Vision, pages 439-457. Springer, 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.345, + 0.483, + 0.413 + ], + "angle": 0, + "content": "[58] Fu-Yun Wang, Zhaoyang Huang, Xiaoyu Shi, Weikang Bian, Guanglu Song, Yu Liu, and Hongsheng Li. *Animatelcm: Accelerating the animation of personalized diffusion models and adapters with decoupled consistency learning*. *ArXiv* preprint, abs/2402.00769, 2024. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.414, + 0.483, + 0.467 + ], + "angle": 0, + "content": "[59] Fu-Yun Wang, Ling Yang, Zhaoyang Huang, Mengdi Wang, and Hongsheng Li. Rectified diffusion: Straightness is not your need in rectified flow. ArXiv preprint, abs/2410.07303, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.469, + 0.483, + 0.523 + ], + "angle": 0, + "content": "[60] Xiang Wang, Shiwei Zhang, Han Zhang, Yu Liu, Yingya Zhang, Changxin Gao, and Nong Sang. Videolcm: Video latent consistency model. ArXiv preprint, abs/2312.09109, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.525, + 0.483, + 0.58 + ], + "angle": 0, + "content": "[61] Jianzong Wu, Xiangtai Li, Yanhong Zeng, Jiangning Zhang, Qianyu Zhou, Yining Li, Yunhai Tong, and Kai Chen. Motionbooth: Motion-aware customized text-to-video generation. ArXiv preprint, abs/2406.17758, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.581, + 0.483, + 0.677 + ], + "angle": 0, + "content": "[62] Jay Zhangjie Wu, Yixiao Ge, Xintao Wang, Stan Weixian Lei, Yuchao Gu, Yufei Shi, Wynne Hsu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. In IEEE/CVF International Conference on Computer Vision, ICCV 2023, Paris, France, October 1-6, 2023, pages 7589-7599. IEEE, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.679, + 0.483, + 0.734 + ], + "angle": 0, + "content": "[63] Tianxing Wu, Chenyang Si, Yuming Jiang, Ziqi Huang, and Ziwei Liu. Freeinit: Bridging initialization gap in video diffusion models. In European Conference on Computer Vision, pages 378-394. Springer, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.735, + 0.483, + 0.789 + ], + "angle": 0, + "content": "[64] Enze Xie, Junsong Chen, Junyu Chen, Han Cai, Yujun Lin, Zhekai Zhang, Muyang Li, Yao Lu, and Song Han. Sana: Efficient high-resolution image synthesis with linear diffusion transformers. ArXiv preprint, abs/2410.10629, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.79, + 0.483, + 0.858 + ], + "angle": 0, + "content": "[65] Yanwu Xu, Yang Zhao, Zhisheng Xiao, and Tingbo Hou. Ufogen: You forward once large scale text-to-image generation via diffusion gans. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8196-8206, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.86, + 0.483, + 0.902 + ], + "angle": 0, + "content": "[66] Ling Yang, Zixiang Zhang, Zhilong Zhang, Xingchao Liu, Minkai Xu, Wentao Zhang, Chenlin Meng, Stefano Ermon, and Bin Cui. Consistency flow matching: Defining" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.093, + 0.905, + 0.119 + ], + "angle": 0, + "content": "straight flows with velocity consistency. ArXiv preprint, abs/2407.02398, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[67] Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, et al. Cogvideox: Text-to-video diffusion models with an expert transformer. ArXiv preprint, abs/2408.06072, 2024. 1, 2, 3, 7, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.193, + 0.905, + 0.248 + ], + "angle": 0, + "content": "[68] Tianwei Yin, Michael Gharbi, Taesung Park, Richard Zhang, Eli Shechtman, Fredo Durand, and William T Freeman. Improved distribution matching distillation for fast image synthesis. ArXiv preprint, abs/2405.14867, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.25, + 0.905, + 0.32 + ], + "angle": 0, + "content": "[69] Tianwei Yin, Michael Gharbi, Richard Zhang, Eli Shechtman, Fredo Durand, William T Freeman, and Taesung Park. One-step diffusion with distribution matching distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6613-6623, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.322, + 0.905, + 0.389 + ], + "angle": 0, + "content": "[70] Zhihang Yuan, Pu Lu, Hanling Zhang, Xuefei Ning, Linfeng Zhang, Tianchen Zhao, Shengen Yan, Guohao Dai, and Yu Wang. Ditfastattn: Attention compression for diffusion transformer models. ArXiv preprint, abs/2406.08552, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.392, + 0.905, + 0.462 + ], + "angle": 0, + "content": "[71] Yan Zeng, Guoqiang Wei, Jiani Zheng, Jiaxin Zou, Yang Wei, Yuchen Zhang, and Hang Li. Make pixels dance: High-dynamic video generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8850-8860, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.463, + 0.905, + 0.532 + ], + "angle": 0, + "content": "[72] Yuanhao Zhai, Kevin Lin, Zhengyuan Yang, Linjie Li, Jianfeng Wang, Chung-Ching Lin, David Doermann, Junsong Yuan, and Lijuan Wang. Motion consistency model: Accelerating video diffusion with disentangled motion-appearance distillation. ArXiv preprint, abs/2406.06890, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.534, + 0.905, + 0.615 + ], + "angle": 0, + "content": "[73] Zheng Zhan, Yushu Wu, Yifan Gong, Zichong Meng, Zhenglun Kong, Changdi Yang, Geng Yuan, Pu Zhao, Wei Niu, and Yanzhi Wang. Fast and memory-efficient video diffusion using streamlined inference. In The Thirty-eighth Annual Conference on Neural Information Processing Systems. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.618, + 0.905, + 0.687 + ], + "angle": 0, + "content": "[74] Zhixing Zhang, Yanyu Li, Yushu Wu, Yanwu Xu, Anil Kag, Ivan Skorokhodov, Willi Menapace, Aliaksandr Siarohin, Junli Cao, Dimitris Metaxas, et al. Sf-v: Single forward video generation model. ArXiv preprint, abs/2406.04324, 2024. 3, 4, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.689, + 0.905, + 0.771 + ], + "angle": 0, + "content": "[75] Zhixing Zhang, Bichen Wu, Xiaoyan Wang, Yaqiao Luo, Luxin Zhang, Yinan Zhao, Peter Vajda, Dimitris Metaxas, and Licheng Yu. Avid: Any-length video inpainting with diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7162-7172, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.774, + 0.905, + 0.844 + ], + "angle": 0, + "content": "[76] Rui Zhao, Yuchao Gu, Jay Zhangjie Wu, David Junhao Zhang, Jia-Wei Liu, Weijia Wu, Jussi Keppo, and Mike Zheng Shou. Motiondirector: Motion customization of text-to-video diffusion models. In European Conference on Computer Vision, pages 273–290. Springer, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[77] Zangwei Zheng, Xiangyu Peng, Tianji Yang, Chenhui Shen, Shenggui Li, Hongxin Liu, Yukun Zhou, Tianyi Li, and Yang You. Open-sora: Democratizing efficient video production for all, 2024. 2, 3, 4, 5, 8, 1" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "2489" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.091, + 0.092, + 0.487, + 0.135 + ], + "angle": 0, + "content": "[78] Yuan Zhou, Qiuyue Wang, Yuxuan Cai, and Huan Yang. Allegro: Open the black box of commercial-level video generation model. ArXiv preprint, abs/2410.15458, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.486, + 0.192 + ], + "angle": 0, + "content": "[79] Yupeng Zhou, Daquan Zhou, Ming-Ming Cheng, Jiashi Feng, and Qibin Hou. Storydiffusion: Consistent self-attention for long-range image and video generation. NeurIPS 2024, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.092, + 0.487, + 0.192 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "2490" + } + ] +] \ No newline at end of file diff --git a/2025/SnapGen-V_ Generating a Five-Second Video within Five Seconds on a Mobile Device/2089e6e3-1b02-4449-b418-985fa5c0c175_origin.pdf b/2025/SnapGen-V_ Generating a Five-Second Video within Five Seconds on a Mobile Device/2089e6e3-1b02-4449-b418-985fa5c0c175_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..ca74ed907d599c6d688b5c237ef6c3f2251f5722 --- /dev/null +++ b/2025/SnapGen-V_ Generating a Five-Second Video within Five Seconds on a Mobile Device/2089e6e3-1b02-4449-b418-985fa5c0c175_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1826964e44bd22f645e55acb43d45fd66343eae719351eac2edbed6729cd4e29 +size 1693741 diff --git a/2025/SnapGen-V_ Generating a Five-Second Video within Five Seconds on a Mobile Device/full.md b/2025/SnapGen-V_ Generating a Five-Second Video within Five Seconds on a Mobile Device/full.md new file mode 100644 index 0000000000000000000000000000000000000000..37b8309cc4df6d5d5e721d90a82afbf71c6ae277 --- /dev/null +++ b/2025/SnapGen-V_ Generating a Five-Second Video within Five Seconds on a Mobile Device/full.md @@ -0,0 +1,295 @@ +# SnapGen-V: Generating a Five-Second Video within Five Seconds on a Mobile Device + +Yushu Wu $^{1,2\dagger}$ Zhixing Zhang $^{1,3\dagger}$ Yanyu Li $^{1\dagger\dagger}$ Yanwu Xu $^{1}$ Anil Kag $^{1}$ Yang Sui $^{1}$ Huseyin Coskun $^{1}$ Ke Ma $^{1}$ Aleksei Lebedev $^{1}$ Ju Hu $^{1}$ Dimitris N. Metaxas $^{3}$ Yanzhi Wang $^{2}$ Sergey Tulyakov $^{1}$ Jian Ren $^{1\dagger}$ Snap Inc. Northeastern University Rutgers University + +![](images/f3caee03fc45eb582f8e3700eaf5b65bbe298c853299bfca7e4c41c321b726a5.jpg) +Figure 1. Example generation results from our mobile text-to-video model. Our model can generate high-quality and motion consistent 5-second videos on a mobile device (e.g., iPhone 16 Pro Max) within 5 seconds. + +# Abstract + +We have witnessed the unprecedented success of diffusion-based video generation over the past year. Recently proposed models from the community have wielded the power to generate cinematic and high-resolution videos with smooth motions from arbitrary input prompts. However, as a supertask of image generation, video generation models require more computation and are thus hosted mostly on cloud servers, limiting broader adoption among content creators. In this work, we propose a comprehensive acceleration framework to bring the power of the largescale video diffusion model to the hands of edge users. From the network architecture scope, we initialize from a compact image backbone and search out the design and arrangement of temporal layers to maximize hardware efficiency. In addition, we propose a dedicated adversarial fine-tuning algorithm for our efficient model and reduce the denoising steps to 4. Our model, with only $0.6B$ parameters, can generate a 5-second video on an iPhone 16 PM within 5 seconds. Compared to server-side models that take minutes on powerful GPUs to generate a single video, we accelerate the generation by magnitudes while delivering on-par quality. Project page at https://snap-research.github.io/snapgen-v/. + +# 1. Introduction + +Recently, the rapid advancement of video diffusion models [4] inspires revolutions in content creation. With the emergence of video models from industry [41] and research community [67], content creators can animate a static image [3] or generate cinematic videos from arbitrary prompts [55, 67, 78]. Video diffusion models also enable downstream applications like video editing [12, 20, 30], novel view synthesis [25, 57], and multi-modal generation [56]. + +Despite the success in generation quality, the huge number of parameters and slow generation speed prohibit the wide deployment of video diffusion models. For instance, CogVideoX-5B [67] generates a video (49 frames at 8 fps, $720 \times 480$ resolution) in 5 minutes with 50 inference steps on an NVIDIA A100 GPU. Compared to text-to-image diffusion models [45], video diffusion models require extra parameters to model sophisticated motions [3, 13, 62]. In addition, video data usually incurs higher spatial-temporal resolution for UNet denoisers [3, 13], or equivalently more tokens for DiT models [40], which adds up to the computation complexity. Recent works explore efficient model architectures and attention mechanisms for image diffusion models [8, 64, 70]. However, there is little effort in the literature dedicated to accelerating and deploying video models at scale, especially for mobile devices. + +In this work, we systematically investigate the redundancies in video diffusion models and propose a mobile + +acceleration framework. First, we obtain an efficient spatial backbone by following prior works [11, 29] to prune a pre-trained text-to-image diffusion model [45]. The pruned model achieves $2.5 \times$ size compression and more than $10 \times$ speedup compared to Stable Diffusion v1.5 [45], while maintaining comparable generative quality. Starting from a pre-trained image model offers two key benefits: (i) it eliminates the need for costly large-scale pre-training, and (ii) with a compact image model, we can significantly narrow the search space in subsequent stages, focusing only on optimizing temporal layers, and thereby accelerating the discovery of the final model architecture. + +Even with the efficient image backbone, applying previous temporal inflation methods [3, 4, 13] still results in tremendous computation cost and encounters out-of-memory issues on mobile. Thus, our second stage is to systematically investigate different types of temporal layers and perform a latency-memory joint search to determine the spatial-temporal architecture for efficient mobile deployment. Unlike previous methods [3, 4, 13] that typically rely on a specific type of temporal modeling layer, we investigate all possible designs, including temporal attention, spatial-temporal full attention (3D attention), temporal cross attention, and temporal convolutions (Conv3D). Besides, our search space includes the position (resolution) to apply these temporal layers, and number of layers to use (in Sec. 3.2). We profile the computation, memory footprint, and on-device latency of architecture candidates, and perform evolutionary search to discover the architecture with Pareto optimality of speed and quality. The searched network is only 0.6B in size, and can generate a 5-second video clip on iPhone 16 PM without hitting memory bound. In contrast, all prior video diffusion models fail to run on mobile, even for the smallest open-sourced ones like the 16-frameAnimateDiff [13] and 14-frame SVD [3] (Tab. 1). + +Finally, to further speed up generation on mobile, we distill our efficient video diffusion model with a tailored adversarial fine-tuning method capable of image-video mixed training. We reduce the number of denoising steps from 25 to 4, and eliminate classifier-free guidance [16], leading to more than $12 \times$ speedup without performance degradation. As in Tab. 1, our mobile speed is faster than most GPU-deployed (e.g., on A100) counterparts. + +
ModelTypeStepsParams (B)A100 (s)iPhone (s)Vbench (↑)
OpenSora-v1.2DiT301.231.00✘79.76
CogVideoX-2BDiT501.654.09✘80.91
AnimateDiff-V2UNet251.29.04✘80.27
AnimateDiffLCMUNet41.21.77✘79.42
OursUNet40.60.474.1281.14
+ +Table 1. Comparison of size (number of parameters), speed (tested on NVIDIA A100 and iPhone 16 Pro Max), and performance (on VBench [19]) for various models. + +With the proposed framework, we successfully deploy our 0.6B text-to-video model on an iPhone 16 Pro Max, achieving the generation of a 5-second video clip within 5 seconds. This work represents not only the very first mobile deployment attempt of the video diffusion model, but also demonstrates its real-time potential1. Our contributions are summarized as follows: + +- Through image-video joint training, spatial and temporal architecture design, and mobile-driven latency-memory joint architecture search, we develop a comprehensive mobile acceleration framework for the text-to-video diffusion model. +- We propose an adversarial fine-tuning technique tailored for video diffusion models. Despite the already compact nature of our mobile denoiser, we further distill it to 4 denoising steps with superior quality. +- Our work is the very first one to show the possibility of real-time text-to-video generation on mobile devices, unlocking the possibility of deploying applications of video diffusion models at scale. + +# 2. Related Work + +Video Diffusion Models. Denoising Diffusion Probabilistic Models [17] is the trending paradigm for building video diffusion models, demonstrating photorealistic quality and generic generation capabilities. Pioneer works often start from a pre-trained text-to-image diffusion model [45] and insert temporal layers to model motions along frame sequence [3, 4, 13, 62, 75]. In addition, training-free noise tuning techniques are proposed to ease the alignment between frames [23, 34, 43, 63, 73]. Later, with the emergence of large-scale, high-quality video datasets [7, 38] and Transformer backbones [40], subsequent works curate their own dataset and build large video diffusion models with exceptional quality, such as the open-sourced CogVideoX [67], Mochi 1 [55], PyramidalFlow [21], Allegro [78], and close-sourced ones including Hailuo [37], Runway Gen 3 Alpha [46], Kling [24], Luma Dream Machine [1], Pika 1.5 [2], Sora [39], and MovieGen [41]. Remarkably, the open-sourced projects Open-Sora [77] and Open-Sora-Plan [26] provide the community with reliable implementations to replicate large-scale video diffusion models. + +Dividing by task type, video diffusion models can be categorized into text-to-video generation [5, 6, 13, 15, 18, 21, 27, 28, 42, 55, 56, 67, 78], image-to-video generation [3, 79], or specific motion controls [14, 44, 61, 71, 76]. Though some work [8, 29, 70, 73] aim to improve the efficiency of diffusion models, the acceleration for mobile deployment of video diffusion models is still in absent. Popular video models [21, 55, 67] can only run on server-level + +Step Distillation brings almost linear generation speedup for diffusion models. Early work [29, 47] progressively distill a student network to predict a further ODE location with teacher guidance, resulting in fewer inference steps, while Consistency Models [52, 53] and Rectified Flow [33] refine the prediction objective to clean data or global velocity to achieve fewer-step inference. Later works [50, 51, 65] further incorporate adversarial loss to distill a single-step student, and enhance multi-step results as well. + +Despite extensive research in image diffusion models [9, 22, 36, 59, 66, 68, 69], step distillation for video diffusion model is under-explored. One type of work applies consistency distillation to generate videos in 4 steps [58, 60, 72]. Another trend adopts adversarial distillation to achieve few-step (1-2) generation [32, 35, 74]. However, these methods distill pre-trained large models with enough redundancy along the trajectory, while we find them not applicable to our efficient model and yield inferior performance. + +# 3. Method + +Our objective is to achieve high-fidelity and temporally consistent video generation on mobile devices. However, current text-to-video diffusion models face two key challenges in reaching this goal: (a) the memory and computation requirement is beyond the capability of even the most powerful mobile chips, i.e. iPhone A18 Pro, and (b) denoising with dozens of steps to generate a single output further slows down the process. To address these challenges, we propose a three-stage framework to accelerate video diffusion models on the mobile platform. First, we prune from a pre-trained text-to-image diffusion model to obtain an efficient spatial backbone. Second, we inflate the spatial backbone with a novel combination of temporal modules which are searched out with our mobile-oriented metrics. Finally, through adversarial training, our efficient model attains the capability to generate high-quality videos in only 4 steps. + +# 3.1. Preliminaries + +Following [77], we employ a spatial-temporal VAE to compress image and video data into the latent space. Given video or image data $\mathbf{v} \in \mathbb{R}^{n \times 3 \times H \times W}$ , where $n$ is the number of frames with height $H$ and width $W$ , the spatial-temporal encoder, $\mathbf{E}$ , maps the data to a latent space. The encoded frames are represented as $\mathbf{x}_0 = \mathbf{E}(\mathbf{v})$ , resulting in $\mathbf{x}_0 \in \mathbb{R}^{\tilde{n} \times 4 \times \tilde{H} \times \tilde{W}}$ . Here, $\mathbf{x}_0 \sim p_{data}(\mathbf{x}_0)$ is a 4-channel latent, with a temporal compression of $\tilde{n} = n/4$ , and spatial compression as $\tilde{H} = H/8$ , $\tilde{W} = W/8$ . + +We follow Rectified Flow [59] to train our latent diffusion model. According to the flow-matching-based diffusion form, the intermediate noisy state $\mathbf{x}_t$ is defined as: + +$$ +\mathbf {x} _ {t} = (1 - t) \mathbf {x} _ {0} + t \epsilon , \text {w h e r e} \epsilon \sim \mathcal {N} (0, I), \tag {1} +$$ + +which is a linear interpolation between the data distribution and a standard normal distribution. The model aims to learn a vector field $v_{\theta}(t, \mathbf{x}_t)$ using the Conditional Flow Matching objective, i.e., + +$$ +\mathcal {L} = \mathbb {E} _ {t, \epsilon , \mathbf {x} _ {0}} \| v _ {\theta} (t, \mathbf {x} _ {t}) - u _ {t} (\mathbf {x} _ {t} | \mathbf {x} _ {0}) \| _ {2} ^ {2}, \tag {2} +$$ + +where $u_{t}(\mathbf{x}_{t}|\mathbf{x}_{0}) = \epsilon -\mathbf{x}_{0}$ . Following [10], during training, we sample $t$ from a logit-normal distribution, i.e., + +$$ +\pi (t; m, s) = \frac {1}{s \sqrt {2 \pi}} \frac {1}{t (1 - t)} \exp \left(- \frac {\left(\operatorname {l o g i t} (t) - m\right) ^ {2}}{2 s ^ {2}}\right), \tag {3} +$$ + +where $\mathrm{logit}(t) = \log \frac{t}{1 - t}$ , $m$ and $s$ are the location parameter and scale parameter, respectively. + +# 3.2. Hardware Efficient Model Design + +Spatial Backbone. We follow [11, 29] to first prune an efficient text-to-image model as the spatial backbone. Specifically, we start from Stable Diffusion v1.5 [45], and borrow the knowledge from prior arts [29] to remove the most mobile-unfriendly attentions. We then prune the network depth and width following [11] and achieve $\times 2.5$ size reduction and more than $10\times$ speedup on mobile devices. We include qualitative visualizations of our image model in the supplementary material. Note that we use a UNet denoiser [17], leaving the exploration of DiT [40] to future work. The hierarchical structure of the UNet denoiser forms a good search space to achieve mobile efficiency, while the computation complexity of DiT grows quadratically with the number of tokens (generation resolution), making it challenging for mobile deployment. + +Temporal Layer Design. Current latent video diffusion models typically adopt temporal self-attention [13], cross-attentions [77], and convolutions [3] to model temporal dependencies. CogVideoX [67] demonstrates significant performance gain by using full 3D-Attention, at the cost of more computations and memory consumption. In this work, we enumerate and investigate all types of temporal modeling methods, including Conv1D, Conv3D, SelfAttention1D, SelfAttention3D, CrossAttention1D, and CrossAttention3D, and profile their complexity in Fig. 3. For instance, SelfAttention1D only models temporal dependency on a single coordinate, while SelfAttention3D models global dependencies and has the potential to deliver much stronger performance. However, the computation complexity of SelfAttention3D grows quadratically with respect to $\tilde{t} \times \tilde{H} \times \tilde{W}$ , while SelfAttention1D is linear with respect to $\tilde{H}$ and $\tilde{W}$ , which makes SelfAttention3D much more costly at higher resolutions. On the other hand, the computation of CrossAttentionND is determined by both spatial-temporal resolution and the number of tokens from the text encoder. Conv1D and 3D are locality alternatives for SelfAttention1D and 3D, respectively. Though the computation complexity and + +![](images/f9b1590de9a48f2a35725c8fa2eba8c41862dd629abd454a2a9c3b56b288ea4d.jpg) +Figure 2. Framework Overview. In the Latency and Memory Guided Architecture Search, we freeze the pretrained efficient spatial layers and conduct evolutionary search over temporal layer based on the memory and latency constraint. During the Adversarial Fine-tuning stage, we initialize the discriminator with the weights from the text-to-video model trained in the first stage. The discriminator employs the encoder of the UNet as its backbone, which remains frozen. We add spatial-temporal discriminator heads after each downsampling block, updating only these heads during training. Following prior works [48, 49, 74], each head conditions on pooled text embeddings c projected via a linear layer. Input features are first reshaped to merge the temporal and batch axes for processing through a 2D ResBlock, and then reshaped again to merge spatial dimensions before the temporal self-attention block. + +![](images/a8da950f93c5dd6710c284a56f977b9163542a38c96c9a2137fe241a50e12324.jpg) +Figure 3. Computation Complexity and Memory Consumption Analysis. The computation complexity and memory consumption of different temporal layer for various input size. The temporal dimension is fixed to 12 for simplicity. + +memory footprint for each design candidate can be easily profiled, as in Fig. 3, it is still crucial and challenging to build a spatial-temporal network with optimized arrangements of these operators. We propose to perform a latency-memory joint architecture search to determine which, where, and the number of temporal layers to use for our efficient video diffusion model on mobile, as follows. + +# Latency and Memory Guided Architecture Search. + +Prior to searching the architecture, we construct a look-up table containing the inference latency and the memory footprint of different temporal layers. For each candidate 1D or 3D operator (i.e., ConvND, SelfAttentionND, CrossAttentionND), we benchmark the latency and memory consumption under different spatial-temporal resolutions on hardware. We then clean the search space by eliminating OOM states. Then we perform evolutionary search to obtain the temporal design with Pareto optimality. The architecture candidate is trained on precomputed video latents for $20K$ iterations with the spatial backbone frozen, and is evaluated on VBench [19] to obtain the scores as the quality metric. + +We include the detailed search algorithm, action space, and total search time in the supplementary material. + +Image-Video Joint Training. Upon the finalized model architecture, we perform image-video joint training under various clip lengths and aspect ratios with all parameters updated for another $100K$ iterations. After the joint training, our efficient model is capable of generating videos with various lengths and aspect ratios under a conventional recipe, i.e., 25 steps with classifier-free guidance. + +VAE Decoder Compression. We use the spatial temporal-decoupled VAE from OpenSora [77], which has 4 latent channels, $8 \times 8$ spatial compression and $4 \times$ temporal compression. To decode a 17-frame video clip on mobile, the original [77] temporal decoder takes $23,100~\mathrm{ms}$ , and the spatial decoder takes $4100~\mathrm{ms}$ , which we found to be a bottleneck for the generation speed. To increase the speed, we focus on the decoder only. We freeze the VAE encoders and prune the temporal and spatial decoder on our video and image dataset, respectively. Our efficient temporal decoder runs at $210~\mathrm{ms}$ and spatial decoder at $330~\mathrm{ms}$ to decode a 17-frame video clip, reaching $50 \times$ speedup with negligible loss in quality. Further details about VAE compression are included in the supplementary material. + +# 3.3. Latent Adversarial Fine-tuning + +Our training procedure involves two networks: a generator $\mathcal{G}_{\theta}$ and a discriminator $\mathcal{D}\phi$ . Similar to prior work [51, 74], we initialize our generator with pre-trained diffusion model weights $\theta$ , while the discriminator is also partially initialized from $\theta$ . Specifically, the backbone of the discriminator adopts the same architecture and weights as the pre-trained UNet encoder, with these backbone parameters remaining frozen during training. Additionally, we enhance the dis + +criminator with spatial-temporal discriminator heads added after each backbone block, with only these head parameters being updated in the discriminator training phase. As illustrated on the right in Fig. 2, each discriminator head consists of a spatial ResBlock and a temporal self-attention block. This design allows our discriminator to effectively handle both image and video data during fine-tuning. We analyze the impact of joint image-video fine-tuning in Sec. 4.3. + +For a real data sample $\mathbf{x}_0$ , a noisy data point $\mathbf{x}_t$ is generated through a forward diffusion process, as described in Eq. (1). We set intermediate timesteps as $0 < T_k < \dots < T_1 = 1.0$ and sample $t$ from these timesteps, where $k$ is the number of timesteps selected for generator training (set to $k = 4$ in practice). The generator, then, predicts the velocity at $\mathbf{x}_t$ as $\mathcal{G}_{\theta}(\mathbf{x}_t,t)$ . + +To train the discriminator, we first sample a target timestep $t'$ from a logit-normal distribution, as shown in Eq. (3). Using the forward process in Eq. (1), we obtain the real sample $\mathbf{x}_{t'} = (1 - t')\mathbf{x}_0 + t'\epsilon$ . The fake sample, $\hat{\mathbf{x}}_{t'}$ , is computed as $\hat{\mathbf{x}}_{t'} = \mathbf{x}_t + (t' - t)\cdot \mathcal{G}_\theta (t,\mathbf{x}_t)$ , as shown in Fig. 4. Following established approaches [48-50, 74], we employ hinge loss [31] as the adversarial objective to enhance performance. The discriminator's goal is to differentiate between real and fake samples by minimizing: + +$$ +\begin{array}{l} \mathcal {L} _ {\mathrm {a d v}} ^ {\mathcal {D}} = \mathbb {E} _ {t ^ {\prime}, x _ {0}} [ \max (0, 1 + \mathcal {D} _ {\phi} (\mathbf {x} _ {t ^ {\prime}}, t ^ {\prime})) ] \tag {4} \\ + \mathbb {E} _ {t, t ^ {\prime}, x _ {0}} \left[ \max \left(0, 1 - \mathcal {D} _ {\phi} \left(\hat {\mathbf {x}} _ {t ^ {\prime}}, t ^ {\prime}\right)\right) \right], \\ \end{array} +$$ + +The adversarial objective for the generator is defined as: + +$$ +\mathcal {L} _ {\mathrm {a d v}} ^ {\mathcal {G}} = \mathbb {E} _ {t, t ^ {\prime}, x _ {0}} \left[ \mathcal {D} _ {\phi} \left(\hat {\mathbf {x}} _ {t ^ {\prime}}, t ^ {\prime}\right) \right]. \tag {5} +$$ + +Following [74], we also incorporate a reconstruction objective to enhance stability, defined as: + +$$ +\mathcal {L} _ {\text {r e c o n}} = \sqrt {\left\| \hat {\mathbf {x}} _ {0} - \mathbf {x} _ {0} \right\| _ {2} ^ {2} + c ^ {2}} - c, \tag {6} +$$ + +where $\hat{\mathbf{x}}_0 = \mathbf{x}_t - t\cdot \mathcal{G}_{\theta}(t,\mathbf{x}_t)$ , and $c > 0$ is an adjustable constant. + +Discussion. Our latent adversarial training pipeline is inspired by SF-V [74]. Similar to SF-V, we set $k = 4$ and utilize the part of the pre-trained diffusion model as the backbone for the discriminator. However, our approach introduces several key differences. First, our method is built on an efficient UNet specifically designed for mobile devices, with fewer parameters than SVD [3], making it a more challenging task. Second, we redesign the discriminator heads: instead of using separate spatial and temporal heads, we integrate them into a unified spatial-temporal head for adversarial training. Rather than handling the temporal dimension separately with 1-D convolutional kernels, we incorporate a temporal self-attention layer into the spatial discriminator head after the 2-D ResBlock, forming a spatial-temporal discriminator head. This unified design enables + +![](images/a0a32c9effc8517ed4200639e7cd253731deb36c8ae66eca2c83c9db045a799e.jpg) +Figure 4. Latent Adversarial Fine-tuning. Given a latent $\mathbf{x}_0$ and a noise latent $\epsilon$ , we obtain the intermediate noisy latent $\mathbf{x}_t$ through a forward diffusion process. The generator then predicts the velocity as $\mathcal{G}_{\theta}(\mathbf{x}_t, t)$ . Using the predicted velocity, we compute $\hat{\mathbf{x}}_0$ and calculate the reconstruction loss $\mathcal{L}_{\mathrm{recon}}$ between $\mathbf{x}_0$ and $\hat{\mathbf{x}}_0$ . For adversarial training, the real sample $\mathbf{x}_{t'}$ is obtained using the same forward diffusion process, while the fake sample $\hat{\mathbf{x}}_{t'}$ is computed using $\mathbf{x}_t$ and the predicted velocity $\mathcal{G}_{\theta}(\mathbf{x}_t, t)$ . + +our model to be jointly trained on both image and video data, which, as demonstrated in Sec. 4.3, significantly enhances the performance of the fine-tuned model. + +# 4. Experiments + +Training. The efficient image backbone is obtained by pruning the Stable Diffusion v1.5 UNet for $100K$ iterations on high-quality synthetic image datasets. The model is then fine-tuned for $50K$ additional iterations to adapt to Rectified-Flow velocity prediction [10] as well as to the Spatial-Temporal VAE [77]. We incorporate QK-norm and RoPE [54] in our network to stabilize training. The workflow for architecture search is discussed in Sec. 3.2. The image model pruning, temporal architecture search, and final model training are conducted on 256 NVIDIA A100 80G GPUs using AdamW optimizer with $5e - 5$ learning rate and betas values as [0.9, 0.999]. + +Adversarial Fine-tuning is conducted for $6K$ iterations on 64 NVIDIA A100 GPUs, using the AdamW optimizer with a learning rate of $1e - 7$ for the generator (i.e., UNet) and $1e - 4$ for the discriminator heads. We set the betas as [0.9, 0.999] for the generator optimizer, and [0.5, 0.999] for the discriminator optimizer. We set the EMA rate as 0.95 following SF-V [74]. We set $m = -1$ , $s = 1$ if not otherwise noted. + +Evaluation. The model is evaluated following the standard benchmarking procedure of VBench [19]. With the 4-step adversarial distilled model, we generate 120-frame horizontal videos at a resolution of $432 \times 768$ using 4 inference steps without employing classifier-free guidance. The generated video is saved at 5 seconds 24 FPS for score testing and qualitative visualization. The mobile demo and detailed demo settings are included in the supplementary material. + +![](images/3bafabc39831471381c38708da6d77ba0e4e64ae46b55a307283b3e1eefac486.jpg) +Figure 5. Video generation on various domains. We employ our model to synthesize videos across diverse domains, with each video containing 120 frames at a resolution of $432 \times 768$ . All results are generated through a 4-step inference process. The results demonstrate that our model can produce high-quality, motion-consistent videos featuring various objects across different domains. + +
ModelParams (B)StepsTotal scoreQuality scoreSemantic scoretemporal flickeringmotion smoothnessdynamic degreeaesthetic qualityobject classmultiple objectscolorspatial relationshipappearance styletemporal styleoverall consistency
Kling--81.8583.3975.6899.3099.4046.9461.2187.2468.0589.9073.0319.6224.1726.42
Pyramid Flow2.02081.7284.7469.6299.4999.1264.6363.2686.6750.7182.8759.5320.9123.0926.23
CogVideoX5.05081.6182.7577.0498.6696.9270.9761.9885.2362.1182.8166.3524.9125.3827.59
T2V-Turbo1.6481.0182.5774.7697.4897.3449.1763.0493.9654.6589.9038.6724.4225.5128.16
Emu38.0-80.9684.0968.4398.5798.9379.2759.6486.1744.6488.3468.7320.9223.2624.79
CogVideoX1.65080.9182.1875.8398.8997.7359.8660.8283.3762.6379.4169.9024.8024.3626.66
Pika-1.0--80.6982.9271.7799.7499.5047.5062.0488.7243.0890.5761.0322.2624.2225.94
VideoCrafter-2.01.95080.4482.2073.4298.4197.7342.5063.1392.5540.6692.9235.8625.1325.8428.23
AnimateDiff-V21.22580.2782.9069.7598.7597.7640.8367.1690.9036.8887.4734.6022.4226.0327.04
OpenSora V1.21.23079.7681.3573.3999.5398.5042.3956.8582.2251.8390.0868.5623.9524.5426.85
AnimateLCM1.2479.4282.3667.6598.5298.1640.5667.0191.4129.7684.2437.1420.9725.1626.07
ModelScope1.45075.7578.0566.5498.2895.7966.3952.0682.2538.9881.7233.6823.3925.3725.67
Ours0.6481.1483.4771.8499.3799.2951.1162.1992.2254.3486.6356.2020.1725.1827.42
+ +Table 2. Performance comparison with popular video generation models on VBench [19]. + +# 4.1. Qualitative Visualization + +We show visualizations of our generated videos in Fig. 5. Our model consistently produces high-quality video frames and smooth object movements. To demonstrate the generic + +text-to-video generation ability, we show various generation examples, including human, animal, photorealistic and art-styled scenes. We include more video visualizations in the supplementary material. + +# 4.2. Quantitative Comparisons + +We present a comprehensive evaluation of our method against existing popular video generation models on VBench [19], as in Tab. 2. Despite the fact that our model is compact and designated for fast inference on mobile, it achieves higher total score compared to recent arts, including the DiT-based OpenSora-V1.2, CogVideoX-2B [67], and the UNet-based VideoCrafter-2.0 [6]. In addition, compared to the 4-step distilled T2V-Turbo [27] and AnimateLCM [58], our model achieves better performance with more than $50\%$ reduction in size. The quantitative scores demonstrate the superiority of our efficient model design and the tailored adversarial distillation method. + +User Study. Human evaluations are conducted between our model and baselines as in the Tab. 3. We generate videos from VBench and Movie Gen Bench prompts and ask human labelers to pick the best results across prompt alignment, aesthetics, and motion. The result shows that our model outperforms baseline metrics by a large margin. + +
Win RatePrompt AlignmentAestheticsMotion
OpenSora-1.224.9%21.2%20.4%
CogVideoX-2B30.7%30.5%30.3%
Ours44.4%48.3%49.3%
+ +# 4.3. Ablation Analysis + +Comparison of Training Data Scheme. We compare the model trained with joint image-video datasets vs. video-only datasets. As shown in Tab. 4, training with video-only datasets leads to significant performance degradation on the VBenchmark score with a drop of 6.32 in aesthetic quality, 2.80 in image quality, and 2.54 in total score. The results highlight the importance of joint image-video training, as the image dataset offers more contextual information and enhances diversity. + +Model Scaling. We scale up and down the model size by adjusting the number of temporal layers, as shown in Tab. 4, to demonstrate the effectiveness of the proposed temporal architecture search. We can observe that scaling up the model can only marginally improve generative scores (i.e., $\times 2$ scale-up only increasing 0.56 in dynamic degree and 0.12 in motion smoothness, and 0.27 in total score). However, both the $\times 2$ and $\times 4$ models hit the memory bound on iPhone. By dividing the $\times 2$ model into more chunks, we test its mobile speed and observe nearly doubled latency. While on the other hand, further scaling down the model results in heavy losses in generation quality (i.e., decreasing 0.56 in dynamic degree and 0.61 in motion smoothness). Our efficient model is a balanced sweet point for quality and on-device performance. + +Table 3. User Study between OpenSora-1.2, CogVideoX-2B, and our model on prompt alignment, aesthetics, and motion. + +
w/ I ScalingΔTΔAQΔIQΔDDΔMSΔTotal
Video-onlyX——-6.32-2.80+4.6-0.17-2.54
×0.5-22%-1.90+0.04-0.56-0.61-1.58
Temporal Scaling✓×2+32%+0.37+0.10+0.56+0.12+0.27
×4+107%+0.49-0.17+0.91+0.17+0.31
+ +Table 4. Analysis of training data scheme, latency, and quality of efficient architecture. The baseline model adopts the best suitable architecture and is trained with joint image-video datasets. The "scaling", and "T" indicates the number of the temporal layers, and the latency comparing to the baseline. The "AQ", "IQ", "DD", and "MS", are aesthetic quality, image quality, dynamic degree, and motion smoothness in VBench [19]. The benchmark metrics are presented as the differences from the baseline model, where negative values indicate a decrease in performance over the baseline vice versa. + +![](images/a3bb3988332de3159b74ee15ce5131b91ba4d3089fd12b4dc844aab8a05326f4.jpg) +Figure 6. Analysis of temporal layers. We ablate temporal layers in different network stages to evaluate their effect. The scores are normalized according to our base model to better demonstrate the difference. + +Effect of Different Temporal Layers. To better understand the roles of the searched temporal layers, we compute the VBench score after systematically removing (i) all temporal layers in the downsample stage, (ii) bottleneck temporal layers, and (iii) all temporal layers in the upsample stage. The model makes reasonable zero-shot generations after temporal layer removal, but we still fine-tune 10K iterations under the same recipe as in Sec. 3.2 for fair comparison. As shown in Fig. 6, removing different temporal layers results in varying degrees of performance degradation across different metrics, and all removing strategies result in a substantial drop in total score, demonstrating that the existence of the searched temporal layers is important, and they play different roles in generation. Specifically, removing temporal layers in upsample blocks results in a more significant loss in imaging quality, subject consistency, and background consistency, suggesting that the up temporal layers play important roles in detail reconstruction. In contrast, bottleneck layers are more important in human action and object class, where global information modeling dom + +inates the results. We observe that removing down layers introduces less overall degradations compared to the other two, which is an anticipated phenomenon because the loss of modeling capacity can be mitigated by the subsequent bottleneck and up stage temporal layers after fine-tuning. + +
Headw/Iw/ LadvmTFMOQSTotal
SF-V [74]X✓-198.8630.7583.6064.2579.73
Ours99.2737.8583.6169.0180.69
Ours✓X-197.321.0170.0317.0459.44
-299.2947.6483.9170.5481.23
Ours✓✓-199.3754.3483.4771.8481.14
099.4443.6381.6969.4679.24
199.5624.4177.8458.8874.05
+ +Table 5. Analysis of our adversarial fine-tuning scheme. We evaluate the VBench [19] scores for our models using different training schemes. In the results, "TF" and "MO" denote the temporal flickering and multiple objects sub-scores, respectively, while "Q" and "S" represent the quality and semantic scores. The table summarizes how varying (1) the type of discriminator head, (2) training with or without an image dataset, and (3) different $t'$ distributions impact the performance of our model. + +Effect of Discriminator Heads. We compare the effects of our spatial-temporal discriminator heads with the separate spatial and temporal heads proposed in SF-V [74] to demonstrate the effectiveness of our discriminator architecture. For a fair comparison, both models are trained exclusively on video data. We evaluate the models finetuned with different discriminator heads on VBench [19]. As shown in the first two rows of Tab. 5, our discriminator heads result in improvements in both the quality score (83.61 vs. 83.60) and the semantic score (69.01 vs. 64.25) of the generated videos. + +Effect of Joint Image Video Fine-tuning. We examine the impact of incorporating image data during adversarial training, as shown in the second and fourth rows of Tab. 5. The results indicate that fine-tuning the model with image data can slightly decrease the Quality score in the VBench [19] evaluation (from 83.61 to 83.47). However, by leveraging the increased diversity of the image dataset, the model achieves a substantial improvement in semantic performance, particularly in multi-object generation (from 37.85 to 54.34). This enhancement leads to a better overall score compared to the model trained exclusively on video data (81.14 vs. 80.69). + +Effect of Noise Distribution for Discriminator. Following Eq. (3), the parameters $m$ and $s$ control the distribution of $t'$ , which determines the noise levels of $\mathbf{x}_{t'}$ and $\hat{\mathbf{x}}_{t'}$ before they are passed to the discriminator as real and fake samples, respectively. We investigate the effect of different noise distributions on model performance by evaluating the results using VBenchmark [19]. As shown in the last four rows of Tab. 5, increasing $m$ (resulting in noisier real and fake sam + +plies) degrades the quality score (from 83.91 to 77.84) while slightly enhancing the temporal flickering score (from 99.29 to 99.56). Although setting $m = -2$ achieves the highest overall score among the experiments (81.23), it performs poorly on multi-object generation. Therefore, in most of our experiments, unless otherwise stated, we use $m = -1$ . This setting yields a slightly lower overall score (81.14) but significantly improves semantic performance (71.84 vs. 70.54) and excels in multi-object generation (54.34 vs. 47.64). + +
# StepsDDOCAQQualitySemanticTotal
111.3979.2452.8574.4362.7072.09
230.2888.9759.7480.0869.2877.92
451.1192.2262.1983.4771.8481.14
+ +Table 6. Analysis of the number of inference steps. We measure VBench [19] score with different numbers of inference steps. In the results, "DD", "OC", and "AQ" denote the dynamic degree, object class, and aesthetic quality scores, respectively. + +Effect of Inference Steps. Our fine-tuned model supports generation with a reduced number of inference steps. We further investigate how varying the number of evaluation steps affects the quality of the generated results, as shown in Tab. 6. While our four-step generation achieves the best performance, even with only two steps, the model still produces reasonable results. Increasing the number of inference steps improves the performance of our model across all metrics. In Tab. 6, we report not only the quality and semantic scores but also scores for dynamic degree, object class, and aesthetic quality. However, reducing the process to a single inference step leads to a significant drop in performance. We leave more aggressive step reductions to future work. + +# 5. Discussion and Conclusion + +In this work, we propose an acceleration framework for the video diffusion model, and for the first time, achieve superfast text-to-video generation on mobile devices. Specifically, we discover an efficient but powerful network architecture through latency and memory joint architecture search for temporal layers. In addition, we propose an improved adversarial fine-tuning technique to distill our model to 4 steps to further speed up generation. Our work is a good starting point for the edge deployment of video diffusion models and we hope to inspire more downstream applications such as video extension and editing. + +Limitations. We use a public 4-channel VAE [77] to encode videos to latent space. Recent work has shown that using more latent channels benefits reconstruction details. Another future direction is to further improve the step reduction technique for 1-2 denoising steps, as works [74] on server-level models. + +# References + +[1] Luma AI. Dream machine. https://lumalabs.ai/dream-machine,.2 +[2] Pika AI. Pika 1.5. https://pika.art/try..2 +[3] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. ArXiv preprint, abs/2311.15127, 2023. 1, 2, 3, 5 +[4] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2023, Vancouver, BC, Canada, June 17-24, 2023, pages 22563-22575. IEEE, 2023. 1, 2 +[5] Haoxin Chen, Menghan Xia, Yingqing He, Yong Zhang, Xiaodong Cun, Shaoshu Yang, Jinbo Xing, Yaofang Liu, Qifeng Chen, Xintao Wang, et al. Videocrafter1: Open diffusion models for high-quality video generation. ArXiv preprint, abs/2310.19512, 2023. 2 +[6] Haoxin Chen, Yong Zhang, Xiaodong Cun, Menghan Xia, Xintao Wang, Chao Weng, and Ying Shan. Videocaster2: Overcoming data limitations for high-quality video diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7310-7320, 2024. 2, 7 +[7] Tsai-Shien Chen, Aliaksandr Siarohin, Willi Menapace, Ekaterina Deyneka, Hsiang-wei Chao, Byung Eun Jeon, Yuwei Fang, Hsin-Ying Lee, Jian Ren, Ming-Hsuan Yang, et al. Panda-70m: Captioning 70m videos with multiple cross-modality teachers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13320-13331, 2024. 2 +[8] Xinwang Chen, Ning Liu, Yichen Zhu, Feifei Feng, and Jian Tang. Edit: An efficient diffusion transformer framework inspired by human-like sketching. ArXiv preprint, abs/2410.23788, 2024. 1, 2 +[9] Trung Dao, Thuan Hoang Nguyen, Thanh Le, Duc Vu, Khoi Nguyen, Cuong Pham, and Anh Tran. Swiftbrush v2: Make your one-step diffusion model better than its teacher. In European Conference on Computer Vision, pages 176-192. Springer, 2025. 3 +[10] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In *Forty-first International Conference on Machine Learning*, 2024. 3, 5 +[11] Gongfan Fang, Xinyin Ma, and Xinchao Wang. Structural pruning for diffusion models. In Advances in Neural Information Processing Systems 36: Annual Conference on Neural Information Processing Systems 2023, NeurIPS 2023, New Orleans, LA, USA, December 10 - 16, 2023, 2023. 2, 3 +[12] Yutang Feng, Sicheng Gao, Yuxiang Bao, Xiaodi Wang, Shumin Han, Juan Zhang, Baochang Zhang, and Angela Yao. Wave: Warping ddim inversion features for zero-shot + +text-to-video editing. In European Conference on Computer Vision, pages 38-55. Springer, 2025. 1 +[13] Yuwei Guo, Ceyuan Yang, Anyi Rao, Zhengyang Liang, Yaohui Wang, Yu Qiao, Maneesh Agrawala, Dahua Lin, and Bo Dai. Animatediff:Animate your personalized text-to-image diffusion models without specific tuning. ArXiv preprint, abs/2307.04725, 2023. 1, 2, 3 +[14] Yuwei Guo, Ceyuan Yang, Anyi Rao, Maneesh Agrawala, Dahua Lin, and Bo Dai. Sparsectrl: Adding sparse controls to text-to-video diffusion models. In European Conference on Computer Vision, pages 330-348. Springer, 2025. 2 +[15] Agrim Gupta, Lijun Yu, Kihyuk Sohn, Xiuye Gu, Meera Hahn, Li Fei-Fei, Irfan Essa, Lu Jiang, and José Lezama. Photorealistic video generation with diffusion models. ArXiv preprint, abs/2312.06662, 2023. 2 +[16] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. ArXiv preprint, abs/2207.12598, 2022. 2 +[17] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In Advances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, NeurIPS 2020, December 6-12, 2020, virtual, 2020. 2, 3 +[18] Wenyi Hong, Ming Ding, Wendi Zheng, Xinghan Liu, and Jie Tang. Cogvideo: Large-scale pretraining for text-to-video generation via transformers. In The Eleventh International Conference on Learning Representations, ICLR 2023, Kigali, Rwanda, May 1-5, 2023. OpenReview.net, 2023. 2 +[19] Ziqi Huang, Yinan He, Jiashuo Yu, Fan Zhang, Chenyang Si, Yuming Jiang, Yuanhan Zhang, Tianxing Wu, Qingyang Jin, Nattapol Chanpaisit, Yaohui Wang, Xinyuan Chen, Limin Wang, Dahua Lin, Yu Qiao, and Ziwei Liu. VBenchmark: Comprehensive benchmark suite for video generative models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 2, 4, 5, 6, 7, 8, 1 +[20] Hyeonho Jeong, Jinho Chang, Geon Yeong Park, and Jong Chul Ye. Dreammotion: Space-time self-similar score distillation for zero-shot video editing. ArXiv preprint, abs/2403.12002, 2024. 1 +[21] Yang Jin, Zhicheng Sun, Ningyuan Li, Kun Xu, Hao Jiang, Nan Zhuang, Quzhe Huang, Yang Song, Yadong Mu, and Zhouchen Lin. Pyramidal flow matching for efficient video generative modeling. ArXiv preprint, abs/2410.05954, 2024. 2 +[22] Beomsu Kim, Yu-Guan Hsieh, Michal Klein, Marco Cuturi, Jong Chul Ye, Bahjat Kawar, and James Thornton. Simple reflow: Improved techniques for fast flow models. ArXiv preprint, abs/2410.07815, 2024. 3 +[23] Jihwan Kim, Junoh Kang, Jinyoung Choi, and Bohyung Han. Fifo-diffusion: Generating infinite videos from text without training. ArXiv preprint, abs/2405.11473, 2024. 2 +[24] Kuaishou. Kling. https://kling.kuaishou.com/en.2 +[25] Jeong-gi Kwak, Erqun Dong, Yuhe Jin, Hanseok Ko, Shweta Mahajan, and Kwang Moo Yi. Vivid-1-to-3: Novel view synthesis with video diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6775–6785, 2024. 1 + +[26] PKU-Yuan Lab and Tuzhan AI etc. Open-sora-plan, 2024. 2 +[27] Jiachen Li, Weixi Feng, Tsu-Jui Fu, Xinyi Wang, Sugato Basu, Wenhu Chen, and William Yang Wang. T2v-turbo: Breaking the quality bottleneck of video consistency model with mixed reward feedback. ArXiv preprint, abs/2405.18750, 2024. 2, 7 +[28] Jiachen Li, Qian Long, Jian Zheng, Xiaofeng Gao, Robinson Piramuthu, Wenhu Chen, and William Yang Wang. T2v-turbo-v2: Enhancing video generation model post-training through data, reward, and conditional guidance design. ArXiv preprint, abs/2410.05677, 2024. 2 +[29] Yanyu Li, Huan Wang, Qing Jin, Ju Hu, Pavlo Chemerys, Yun Fu, Yanzhi Wang, Sergey Tulyakov, and Jian Ren. Snapfusion: Text-to-image diffusion model on mobile devices within two seconds. In Advances in Neural Information Processing Systems 36: Annual Conference on Neural Information Processing Systems 2023, NeurIPS 2023, New Orleans, LA, USA, December 10 - 16, 2023, 2, 3 +[30] Feng Liang, Bichen Wu, Jialiang Wang, Licheng Yu, Kunpeng Li, Yinan Zhao, Ishan Misra, Jia-Bin Huang, Peizhao Zhang, Peter Vajda, et al. Flowvid: Taming imperfect optical flows for consistent video-to-video synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8207-8216, 2024. 1 +[31] Jae Hyun Lim and Jong Chul Ye. Geometric gan. ArXiv preprint, abs/1705.02894, 2017. 5 +[32] Shanchuan Lin and Xiao Yang. Animatediff-lightning: Cross-model diffusion distillation. ArXiv preprint, abs/2403.12706, 2024. 3 +[33] Xingchao Liu, Chengyue Gong, and Qiang Liu. Flow straight and fast: Learning to generate and transfer data with rectified flow. In The Eleventh International Conference on Learning Representations, ICLR 2023, Kigali, Rwanda, May 1-5, 2023. OpenReview.net, 2023. 3 +[34] Yu Lu, Yuanzhi Liang, Linchao Zhu, and Yi Yang. Freelong: Training-free long video generation with spectralblend temporal attention. ArXiv preprint, abs/2407.19918, 2024. 2 +[35] Xiaofeng Mao, Zhengkai Jiang, Fu-Yun Wang, Wenbing Zhu, Jiangning Zhang, Hao Chen, Mingmin Chi, and Yabiao Wang. Osv: One step is enough for high-quality image to video generation. ArXiv preprint, abs/2409.11367, 2024. 3 +[36] Kangfu Mei, Maurizio Delbracio, Hossein Talebi, Zhengzhong Tu, Vishal M Patel, and Peyman Milanfar. Codi: Conditional diffusion distillation for higher-fidelity and faster image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9048-9058, 2024. 3 +[37] MiniMax. Hailuo ai. https://hailuoai(video/.2 +[38] Kepan Nan, Rui Xie, Penghao Zhou, Tiehan Fan, Zhenheng Yang, Zhijie Chen, Xiang Li, Jian Yang, and Ying Tai. Openvid-1m: A large-scale high-quality dataset for text-to-video generation. ArXiv preprint, abs/2407.02371, 2024. 2 +[39] OpenAI. Video generation models as world simulators. https://openai.com/index/video-generation-models-as-world-simulators/.2 +[40] William Peebles and Saining Xie. Scalable diffusion models with transformers. In IEEE/CVF International Conference + +on Computer Vision, ICCV 2023, Paris, France, October 1-6, 2023, pages 4172-4182. IEEE, 2023. 1, 2, 3 +[41] Adam Polyak, Amit Zohar, Andrew Brown, Andros Tjandra, Animesh Sinha, Ann Lee, Apoorv Vyas, Bowen Shi, Chih-Yao Ma, Ching-Yao Chuang, et al. Movie gen: A cast of media foundation models. ArXiv preprint, abs/2410.13720, 2024. 1, 2 +[42] Zhiwu Qing, Shiwei Zhang, Jiayu Wang, Xiang Wang, Yujie Wei, Yingya Zhang, Changxin Gao, and Nong Sang. Hierarchical spatio-temporal decoupling for text-to-video generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6635-6645, 2024. 2 +[43] Haonan Qiu, Menghan Xia, Yong Zhang, Yingqing He, Xintao Wang, Ying Shan, and Ziwei Liu. Freenoise: Tuning-free longer video diffusion via noise rescheduling. ArXiv preprint, abs/2310.15169, 2023. 2 +[44] Yixuan Ren, Yang Zhou, Jimei Yang, Jing Shi, Difan Liu, Feng Liu, Mingi Kwon, and Abhinav Shrivastava. Customize-a-video: One-shot motion customization of text-to-video diffusion models. ArXiv preprint, abs/2402.14780, 2024. 2 +[45] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 2, 3 +[46] Runway. Gen-3 alpha. https://runwayml.com/research/introducing-gen-3-alpha.2 +[47] Tim Salimans and Jonathan Ho. Progressive distillation for fast sampling of diffusion models. In *The Tenth International Conference on Learning Representations*, ICLR 2022, Virtual Event, April 25-29, 2022. OpenReview.net, 2022. 3 +[48] Axel Sauer, Kashyap Chitta, Jens Müller, and Andreas Geiger. Projected gans converge faster. In Advances in Neural Information Processing Systems 34: Annual Conference on Neural Information Processing Systems 2021, NeurIPS 2021, December 6-14, 2021, virtual, pages 17480-17492, 2021. 4, 5 +[49] Axel Sauer, Tero Karras, Samuli Laine, Andreas Geiger, and Timo Aila. Stylegan-t: Unlocking the power of gans for fast large-scale text-to-image synthesis. In International Conference on Machine Learning, ICML 2023, 23-29 July 2023, Honolulu, Hawaii, USA, pages 30105-30118. PMLR, 2023. 4 +[50] Axel Sauer, Dominik Lorenz, Andreas Blattmann, and Robin Rombach. Adversarial diffusion distillation. ArXiv preprint, abs/2311.17042, 2023. 3, 5 +[51] Axel Sauer, Frederic Boesel, Tim Dockhorn, Andreas Blattmann, Patrick Esser, and Robin Rombach. Fast high-resolution image synthesis with latent adversarial diffusion distillation. ArXiv preprint, abs/2403.12015, 2024. 3, 4 +[52] Yang Song and Prafulla Dhariwal. Improved techniques for training consistency models. ArXiv preprint, abs/2310.14189, 2023. 3 +[53] Yang Song, Prafulla Dhariwal, Mark Chen, and Ilya Sutskever. Consistency models. In International Conference on Machine Learning, ICML 2023, 23-29 July 2023, + +Honolulu, Hawaii, USA, pages 32211-32252. PMLR, 2023. 3 +[54] Jianlin Su, Murtadha Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu. Roformer: Enhanced transformer with rotary position embedding. Neurocomputing, 568:127063, 2024. 5 +[55] Genmo Team. Mochi, 2024. 1, 2 +[56] Gemini Team, Rohan Anil, Sebastian Borgeaud, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, Katie Millican, et al. Gemini: a family of highly capable multimodal models. ArXiv preprint, abs/2312.11805, 2023. 1, 2 +[57] Vikram Voleti, Chun-Han Yao, Mark Boss, Adam Letts, David Pankratz, Dmitry Tochilkin, Christian Laforte, Robin Rombach, and Varun Jampani. Sv3d: Novel multi-view synthesis and 3d generation from a single image using latent video diffusion. In European Conference on Computer Vision, pages 439-457. Springer, 2025. 1 +[58] Fu-Yun Wang, Zhaoyang Huang, Xiaoyu Shi, Weikang Bian, Guanglu Song, Yu Liu, and Hongsheng Li. *Animatelcm: Accelerating the animation of personalized diffusion models and adapters with decoupled consistency learning*. *ArXiv* preprint, abs/2402.00769, 2024. 3, 7 +[59] Fu-Yun Wang, Ling Yang, Zhaoyang Huang, Mengdi Wang, and Hongsheng Li. Rectified diffusion: Straightness is not your need in rectified flow. ArXiv preprint, abs/2410.07303, 2024. 3 +[60] Xiang Wang, Shiwei Zhang, Han Zhang, Yu Liu, Yingya Zhang, Changxin Gao, and Nong Sang. Videolcm: Video latent consistency model. ArXiv preprint, abs/2312.09109, 2023. 3 +[61] Jianzong Wu, Xiangtai Li, Yanhong Zeng, Jiangning Zhang, Qianyu Zhou, Yining Li, Yunhai Tong, and Kai Chen. Motionbooth: Motion-aware customized text-to-video generation. ArXiv preprint, abs/2406.17758, 2024. 2 +[62] Jay Zhangjie Wu, Yixiao Ge, Xintao Wang, Stan Weixian Lei, Yuchao Gu, Yufei Shi, Wynne Hsu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. In IEEE/CVF International Conference on Computer Vision, ICCV 2023, Paris, France, October 1-6, 2023, pages 7589-7599. IEEE, 2023. 1, 2 +[63] Tianxing Wu, Chenyang Si, Yuming Jiang, Ziqi Huang, and Ziwei Liu. Freeinit: Bridging initialization gap in video diffusion models. In European Conference on Computer Vision, pages 378-394. Springer, 2025. 2 +[64] Enze Xie, Junsong Chen, Junyu Chen, Han Cai, Yujun Lin, Zhekai Zhang, Muyang Li, Yao Lu, and Song Han. Sana: Efficient high-resolution image synthesis with linear diffusion transformers. ArXiv preprint, abs/2410.10629, 2024. 1 +[65] Yanwu Xu, Yang Zhao, Zhisheng Xiao, and Tingbo Hou. Ufogen: You forward once large scale text-to-image generation via diffusion gans. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8196-8206, 2024. 3 +[66] Ling Yang, Zixiang Zhang, Zhilong Zhang, Xingchao Liu, Minkai Xu, Wentao Zhang, Chenlin Meng, Stefano Ermon, and Bin Cui. Consistency flow matching: Defining + +straight flows with velocity consistency. ArXiv preprint, abs/2407.02398, 2024. 3 +[67] Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, et al. Cogvideox: Text-to-video diffusion models with an expert transformer. ArXiv preprint, abs/2408.06072, 2024. 1, 2, 3, 7, 5 +[68] Tianwei Yin, Michael Gharbi, Taesung Park, Richard Zhang, Eli Shechtman, Fredo Durand, and William T Freeman. Improved distribution matching distillation for fast image synthesis. ArXiv preprint, abs/2405.14867, 2024. 3 +[69] Tianwei Yin, Michael Gharbi, Richard Zhang, Eli Shechtman, Fredo Durand, William T Freeman, and Taesung Park. One-step diffusion with distribution matching distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6613-6623, 2024. 3 +[70] Zhihang Yuan, Pu Lu, Hanling Zhang, Xuefei Ning, Linfeng Zhang, Tianchen Zhao, Shengen Yan, Guohao Dai, and Yu Wang. Ditfastattn: Attention compression for diffusion transformer models. ArXiv preprint, abs/2406.08552, 2024. 1, 2 +[71] Yan Zeng, Guoqiang Wei, Jiani Zheng, Jiaxin Zou, Yang Wei, Yuchen Zhang, and Hang Li. Make pixels dance: High-dynamic video generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8850-8860, 2024. 2 +[72] Yuanhao Zhai, Kevin Lin, Zhengyuan Yang, Linjie Li, Jianfeng Wang, Chung-Ching Lin, David Doermann, Junsong Yuan, and Lijuan Wang. Motion consistency model: Accelerating video diffusion with disentangled motion-appearance distillation. ArXiv preprint, abs/2406.06890, 2024. 3 +[73] Zheng Zhan, Yushu Wu, Yifan Gong, Zichong Meng, Zhenglun Kong, Changdi Yang, Geng Yuan, Pu Zhao, Wei Niu, and Yanzhi Wang. Fast and memory-efficient video diffusion using streamlined inference. In The Thirty-eighth Annual Conference on Neural Information Processing Systems. 2 +[74] Zhixing Zhang, Yanyu Li, Yushu Wu, Yanwu Xu, Anil Kag, Ivan Skorokhodov, Willi Menapace, Aliaksandr Siarohin, Junli Cao, Dimitris Metaxas, et al. Sf-v: Single forward video generation model. ArXiv preprint, abs/2406.04324, 2024. 3, 4, 5, 8 +[75] Zhixing Zhang, Bichen Wu, Xiaoyan Wang, Yaqiao Luo, Luxin Zhang, Yinan Zhao, Peter Vajda, Dimitris Metaxas, and Licheng Yu. Avid: Any-length video inpainting with diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7162-7172, 2024. 2 +[76] Rui Zhao, Yuchao Gu, Jay Zhangjie Wu, David Junhao Zhang, Jia-Wei Liu, Weijia Wu, Jussi Keppo, and Mike Zheng Shou. Motiondirector: Motion customization of text-to-video diffusion models. In European Conference on Computer Vision, pages 273–290. Springer, 2025. 2 +[77] Zangwei Zheng, Xiangyu Peng, Tianji Yang, Chenhui Shen, Shenggui Li, Hongxin Liu, Yukun Zhou, Tianyi Li, and Yang You. Open-sora: Democratizing efficient video production for all, 2024. 2, 3, 4, 5, 8, 1 + +[78] Yuan Zhou, Qiuyue Wang, Yuxuan Cai, and Huan Yang. Allegro: Open the black box of commercial-level video generation model. ArXiv preprint, abs/2410.15458, 2024. 1, 2 +[79] Yupeng Zhou, Daquan Zhou, Ming-Ming Cheng, Jiashi Feng, and Qibin Hou. Storydiffusion: Consistent self-attention for long-range image and video generation. NeurIPS 2024, 2024. 2 \ No newline at end of file diff --git a/2025/SnapGen-V_ Generating a Five-Second Video within Five Seconds on a Mobile Device/images.zip b/2025/SnapGen-V_ Generating a Five-Second Video within Five Seconds on a Mobile Device/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..c13bc648da12a3518e55625153599e886d3d1455 --- /dev/null +++ b/2025/SnapGen-V_ Generating a Five-Second Video within Five Seconds on a Mobile Device/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:59ec7028bcb5b119b7b8a357790a8236822625495b36776fbf0b47f554f8eb1f +size 686150 diff --git a/2025/SnapGen-V_ Generating a Five-Second Video within Five Seconds on a Mobile Device/layout.json b/2025/SnapGen-V_ Generating a Five-Second Video within Five Seconds on a Mobile Device/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..eaf6fa36a4b65911ce7087b8d72cd9072aef473d --- /dev/null +++ b/2025/SnapGen-V_ Generating a Five-Second Video within Five Seconds on a Mobile Device/layout.json @@ -0,0 +1,9169 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 65, + 102, + 545, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 102, + 545, + 138 + ], + "spans": [ + { + "bbox": [ + 65, + 102, + 545, + 138 + ], + "type": "text", + "content": "SnapGen-V: Generating a Five-Second Video within Five Seconds on a Mobile Device" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 94, + 159, + 515, + 219 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 159, + 515, + 219 + ], + "spans": [ + { + "bbox": [ + 94, + 159, + 515, + 219 + ], + "type": "text", + "content": "Yushu Wu" + }, + { + "bbox": [ + 94, + 159, + 515, + 219 + ], + "type": "inline_equation", + "content": "^{1,2\\dagger}" + }, + { + "bbox": [ + 94, + 159, + 515, + 219 + ], + "type": "text", + "content": " Zhixing Zhang" + }, + { + "bbox": [ + 94, + 159, + 515, + 219 + ], + "type": "inline_equation", + "content": "^{1,3\\dagger}" + }, + { + "bbox": [ + 94, + 159, + 515, + 219 + ], + "type": "text", + "content": " Yanyu Li" + }, + { + "bbox": [ + 94, + 159, + 515, + 219 + ], + "type": "inline_equation", + "content": "^{1\\dagger\\dagger}" + }, + { + "bbox": [ + 94, + 159, + 515, + 219 + ], + "type": "text", + "content": " Yanwu Xu" + }, + { + "bbox": [ + 94, + 159, + 515, + 219 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 94, + 159, + 515, + 219 + ], + "type": "text", + "content": " Anil Kag" + }, + { + "bbox": [ + 94, + 159, + 515, + 219 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 94, + 159, + 515, + 219 + ], + "type": "text", + "content": " Yang Sui" + }, + { + "bbox": [ + 94, + 159, + 515, + 219 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 94, + 159, + 515, + 219 + ], + "type": "text", + "content": " Huseyin Coskun" + }, + { + "bbox": [ + 94, + 159, + 515, + 219 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 94, + 159, + 515, + 219 + ], + "type": "text", + "content": " Ke Ma" + }, + { + "bbox": [ + 94, + 159, + 515, + 219 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 94, + 159, + 515, + 219 + ], + "type": "text", + "content": " Aleksei Lebedev" + }, + { + "bbox": [ + 94, + 159, + 515, + 219 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 94, + 159, + 515, + 219 + ], + "type": "text", + "content": " Ju Hu" + }, + { + "bbox": [ + 94, + 159, + 515, + 219 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 94, + 159, + 515, + 219 + ], + "type": "text", + "content": " Dimitris N. Metaxas" + }, + { + "bbox": [ + 94, + 159, + 515, + 219 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 94, + 159, + 515, + 219 + ], + "type": "text", + "content": " Yanzhi Wang" + }, + { + "bbox": [ + 94, + 159, + 515, + 219 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 94, + 159, + 515, + 219 + ], + "type": "text", + "content": " Sergey Tulyakov" + }, + { + "bbox": [ + 94, + 159, + 515, + 219 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 94, + 159, + 515, + 219 + ], + "type": "text", + "content": " Jian Ren" + }, + { + "bbox": [ + 94, + 159, + 515, + 219 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 94, + 159, + 515, + 219 + ], + "type": "text", + "content": " Snap Inc. Northeastern University Rutgers University" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 57, + 245, + 553, + 342 + ], + "blocks": [ + { + "bbox": [ + 57, + 245, + 553, + 342 + ], + "lines": [ + { + "bbox": [ + 57, + 245, + 553, + 342 + ], + "spans": [ + { + "bbox": [ + 57, + 245, + 553, + 342 + ], + "type": "image", + "image_path": "f3caee03fc45eb582f8e3700eaf5b65bbe298c853299bfca7e4c41c321b726a5.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 350, + 555, + 373 + ], + "lines": [ + { + "bbox": [ + 55, + 350, + 555, + 373 + ], + "spans": [ + { + "bbox": [ + 55, + 350, + 555, + 373 + ], + "type": "text", + "content": "Figure 1. Example generation results from our mobile text-to-video model. Our model can generate high-quality and motion consistent 5-second videos on a mobile device (e.g., iPhone 16 Pro Max) within 5 seconds." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 152, + 382, + 200, + 395 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 382, + 200, + 395 + ], + "spans": [ + { + "bbox": [ + 152, + 382, + 200, + 395 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 408, + 297, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 408, + 297, + 673 + ], + "spans": [ + { + "bbox": [ + 54, + 408, + 297, + 673 + ], + "type": "text", + "content": "We have witnessed the unprecedented success of diffusion-based video generation over the past year. Recently proposed models from the community have wielded the power to generate cinematic and high-resolution videos with smooth motions from arbitrary input prompts. However, as a supertask of image generation, video generation models require more computation and are thus hosted mostly on cloud servers, limiting broader adoption among content creators. In this work, we propose a comprehensive acceleration framework to bring the power of the largescale video diffusion model to the hands of edge users. From the network architecture scope, we initialize from a compact image backbone and search out the design and arrangement of temporal layers to maximize hardware efficiency. In addition, we propose a dedicated adversarial fine-tuning algorithm for our efficient model and reduce the denoising steps to 4. Our model, with only " + }, + { + "bbox": [ + 54, + 408, + 297, + 673 + ], + "type": "inline_equation", + "content": "0.6B" + }, + { + "bbox": [ + 54, + 408, + 297, + 673 + ], + "type": "text", + "content": " parameters, can generate a 5-second video on an iPhone 16 PM within 5 seconds. Compared to server-side models that take minutes on powerful GPUs to generate a single video, we accelerate the generation by magnitudes while delivering on-par quality. Project page at https://snap-research.github.io/snapgen-v/." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 315, + 382, + 394, + 395 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 382, + 394, + 395 + ], + "spans": [ + { + "bbox": [ + 315, + 382, + 394, + 395 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 402, + 555, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 402, + 555, + 498 + ], + "spans": [ + { + "bbox": [ + 313, + 402, + 555, + 498 + ], + "type": "text", + "content": "Recently, the rapid advancement of video diffusion models [4] inspires revolutions in content creation. With the emergence of video models from industry [41] and research community [67], content creators can animate a static image [3] or generate cinematic videos from arbitrary prompts [55, 67, 78]. Video diffusion models also enable downstream applications like video editing [12, 20, 30], novel view synthesis [25, 57], and multi-modal generation [56]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 498, + 556, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 556, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 556, + 689 + ], + "type": "text", + "content": "Despite the success in generation quality, the huge number of parameters and slow generation speed prohibit the wide deployment of video diffusion models. For instance, CogVideoX-5B [67] generates a video (49 frames at 8 fps, " + }, + { + "bbox": [ + 313, + 498, + 556, + 689 + ], + "type": "inline_equation", + "content": "720 \\times 480" + }, + { + "bbox": [ + 313, + 498, + 556, + 689 + ], + "type": "text", + "content": " resolution) in 5 minutes with 50 inference steps on an NVIDIA A100 GPU. Compared to text-to-image diffusion models [45], video diffusion models require extra parameters to model sophisticated motions [3, 13, 62]. In addition, video data usually incurs higher spatial-temporal resolution for UNet denoisers [3, 13], or equivalently more tokens for DiT models [40], which adds up to the computation complexity. Recent works explore efficient model architectures and attention mechanisms for image diffusion models [8, 64, 70]. However, there is little effort in the literature dedicated to accelerating and deploying video models at scale, especially for mobile devices." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": "In this work, we systematically investigate the redundancies in video diffusion models and propose a mobile" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 693, + 132, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 693, + 132, + 703 + ], + "spans": [ + { + "bbox": [ + 66, + 693, + 132, + 703 + ], + "type": "text", + "content": "Equal contribution" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 67, + 703, + 146, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 146, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 146, + 712 + ], + "type": "inline_equation", + "content": "\\ddagger" + }, + { + "bbox": [ + 67, + 703, + 146, + 712 + ], + "type": "text", + "content": " Corresponding authors" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2479" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 54, + 72, + 294, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 72, + 294, + 215 + ], + "spans": [ + { + "bbox": [ + 54, + 72, + 294, + 215 + ], + "type": "text", + "content": "acceleration framework. First, we obtain an efficient spatial backbone by following prior works [11, 29] to prune a pre-trained text-to-image diffusion model [45]. The pruned model achieves " + }, + { + "bbox": [ + 54, + 72, + 294, + 215 + ], + "type": "inline_equation", + "content": "2.5 \\times" + }, + { + "bbox": [ + 54, + 72, + 294, + 215 + ], + "type": "text", + "content": " size compression and more than " + }, + { + "bbox": [ + 54, + 72, + 294, + 215 + ], + "type": "inline_equation", + "content": "10 \\times" + }, + { + "bbox": [ + 54, + 72, + 294, + 215 + ], + "type": "text", + "content": " speedup compared to Stable Diffusion v1.5 [45], while maintaining comparable generative quality. Starting from a pre-trained image model offers two key benefits: (i) it eliminates the need for costly large-scale pre-training, and (ii) with a compact image model, we can significantly narrow the search space in subsequent stages, focusing only on optimizing temporal layers, and thereby accelerating the discovery of the final model architecture." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 218, + 297, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 218, + 297, + 491 + ], + "spans": [ + { + "bbox": [ + 57, + 218, + 297, + 491 + ], + "type": "text", + "content": "Even with the efficient image backbone, applying previous temporal inflation methods [3, 4, 13] still results in tremendous computation cost and encounters out-of-memory issues on mobile. Thus, our second stage is to systematically investigate different types of temporal layers and perform a latency-memory joint search to determine the spatial-temporal architecture for efficient mobile deployment. Unlike previous methods [3, 4, 13] that typically rely on a specific type of temporal modeling layer, we investigate all possible designs, including temporal attention, spatial-temporal full attention (3D attention), temporal cross attention, and temporal convolutions (Conv3D). Besides, our search space includes the position (resolution) to apply these temporal layers, and number of layers to use (in Sec. 3.2). We profile the computation, memory footprint, and on-device latency of architecture candidates, and perform evolutionary search to discover the architecture with Pareto optimality of speed and quality. The searched network is only 0.6B in size, and can generate a 5-second video clip on iPhone 16 PM without hitting memory bound. In contrast, all prior video diffusion models fail to run on mobile, even for the smallest open-sourced ones like the 16-frameAnimateDiff [13] and 14-frame SVD [3] (Tab. 1)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 494, + 295, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 494, + 295, + 590 + ], + "spans": [ + { + "bbox": [ + 55, + 494, + 295, + 590 + ], + "type": "text", + "content": "Finally, to further speed up generation on mobile, we distill our efficient video diffusion model with a tailored adversarial fine-tuning method capable of image-video mixed training. We reduce the number of denoising steps from 25 to 4, and eliminate classifier-free guidance [16], leading to more than " + }, + { + "bbox": [ + 55, + 494, + 295, + 590 + ], + "type": "inline_equation", + "content": "12 \\times" + }, + { + "bbox": [ + 55, + 494, + 295, + 590 + ], + "type": "text", + "content": " speedup without performance degradation. As in Tab. 1, our mobile speed is faster than most GPU-deployed (e.g., on A100) counterparts." + } + ] + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 58, + 603, + 293, + 668 + ], + "blocks": [ + { + "bbox": [ + 58, + 603, + 293, + 668 + ], + "lines": [ + { + "bbox": [ + 58, + 603, + 293, + 668 + ], + "spans": [ + { + "bbox": [ + 58, + 603, + 293, + 668 + ], + "type": "table", + "html": "
ModelTypeStepsParams (B)A100 (s)iPhone (s)Vbench (↑)
OpenSora-v1.2DiT301.231.00✘79.76
CogVideoX-2BDiT501.654.09✘80.91
AnimateDiff-V2UNet251.29.04✘80.27
AnimateDiffLCMUNet41.21.77✘79.42
OursUNet40.60.474.1281.14
", + "image_path": "5ffb1fea6612f62db4b35d1db855b8248d091a2c4a4e545437e4b1f1ad483102.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 677, + 295, + 710 + ], + "lines": [ + { + "bbox": [ + 55, + 677, + 295, + 710 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 295, + 710 + ], + "type": "text", + "content": "Table 1. Comparison of size (number of parameters), speed (tested on NVIDIA A100 and iPhone 16 Pro Max), and performance (on VBench [19]) for various models." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 72, + 553, + 155 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 155 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 155 + ], + "type": "text", + "content": "With the proposed framework, we successfully deploy our 0.6B text-to-video model on an iPhone 16 Pro Max, achieving the generation of a 5-second video clip within 5 seconds. This work represents not only the very first mobile deployment attempt of the video diffusion model, but also demonstrates its real-time potential1. Our contributions are summarized as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 156, + 553, + 310 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 314, + 156, + 553, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 156, + 553, + 215 + ], + "spans": [ + { + "bbox": [ + 314, + 156, + 553, + 215 + ], + "type": "text", + "content": "- Through image-video joint training, spatial and temporal architecture design, and mobile-driven latency-memory joint architecture search, we develop a comprehensive mobile acceleration framework for the text-to-video diffusion model." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 216, + 553, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 216, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 314, + 216, + 553, + 262 + ], + "type": "text", + "content": "- We propose an adversarial fine-tuning technique tailored for video diffusion models. Despite the already compact nature of our mobile denoiser, we further distill it to 4 denoising steps with superior quality." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 263, + 553, + 310 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 263, + 553, + 310 + ], + "spans": [ + { + "bbox": [ + 314, + 263, + 553, + 310 + ], + "type": "text", + "content": "- Our work is the very first one to show the possibility of real-time text-to-video generation on mobile devices, unlocking the possibility of deploying applications of video diffusion models at scale." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 322, + 400, + 334 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 322, + 400, + 334 + ], + "spans": [ + { + "bbox": [ + 314, + 322, + 400, + 334 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 342, + 555, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 342, + 555, + 581 + ], + "spans": [ + { + "bbox": [ + 313, + 342, + 555, + 581 + ], + "type": "text", + "content": "Video Diffusion Models. Denoising Diffusion Probabilistic Models [17] is the trending paradigm for building video diffusion models, demonstrating photorealistic quality and generic generation capabilities. Pioneer works often start from a pre-trained text-to-image diffusion model [45] and insert temporal layers to model motions along frame sequence [3, 4, 13, 62, 75]. In addition, training-free noise tuning techniques are proposed to ease the alignment between frames [23, 34, 43, 63, 73]. Later, with the emergence of large-scale, high-quality video datasets [7, 38] and Transformer backbones [40], subsequent works curate their own dataset and build large video diffusion models with exceptional quality, such as the open-sourced CogVideoX [67], Mochi 1 [55], PyramidalFlow [21], Allegro [78], and close-sourced ones including Hailuo [37], Runway Gen 3 Alpha [46], Kling [24], Luma Dream Machine [1], Pika 1.5 [2], Sora [39], and MovieGen [41]. Remarkably, the open-sourced projects Open-Sora [77] and Open-Sora-Plan [26] provide the community with reliable implementations to replicate large-scale video diffusion models." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 582, + 554, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 554, + 677 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 554, + 677 + ], + "type": "text", + "content": "Dividing by task type, video diffusion models can be categorized into text-to-video generation [5, 6, 13, 15, 18, 21, 27, 28, 42, 55, 56, 67, 78], image-to-video generation [3, 79], or specific motion controls [14, 44, 61, 71, 76]. Though some work [8, 29, 70, 73] aim to improve the efficiency of diffusion models, the acceleration for mobile deployment of video diffusion models is still in absent. Popular video models [21, 55, 67] can only run on server-level" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 313, + 683, + 553, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 683, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 683, + 553, + 713 + ], + "type": "text", + "content": "This work performs conventional T2V generation, generating an entire video at once. However, a more suitable real-time capability would involve streamlined, continuous video generation, which we leave for future work." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2480" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 191 + ], + "type": "text", + "content": "Step Distillation brings almost linear generation speedup for diffusion models. Early work [29, 47] progressively distill a student network to predict a further ODE location with teacher guidance, resulting in fewer inference steps, while Consistency Models [52, 53] and Rectified Flow [33] refine the prediction objective to clean data or global velocity to achieve fewer-step inference. Later works [50, 51, 65] further incorporate adversarial loss to distill a single-step student, and enhance multi-step results as well." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 192, + 296, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 192, + 296, + 300 + ], + "spans": [ + { + "bbox": [ + 55, + 192, + 296, + 300 + ], + "type": "text", + "content": "Despite extensive research in image diffusion models [9, 22, 36, 59, 66, 68, 69], step distillation for video diffusion model is under-explored. One type of work applies consistency distillation to generate videos in 4 steps [58, 60, 72]. Another trend adopts adversarial distillation to achieve few-step (1-2) generation [32, 35, 74]. However, these methods distill pre-trained large models with enough redundancy along the trajectory, while we find them not applicable to our efficient model and yield inferior performance." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 309, + 111, + 320 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 309, + 111, + 320 + ], + "spans": [ + { + "bbox": [ + 55, + 309, + 111, + 320 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 329, + 296, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 329, + 296, + 521 + ], + "spans": [ + { + "bbox": [ + 55, + 329, + 296, + 521 + ], + "type": "text", + "content": "Our objective is to achieve high-fidelity and temporally consistent video generation on mobile devices. However, current text-to-video diffusion models face two key challenges in reaching this goal: (a) the memory and computation requirement is beyond the capability of even the most powerful mobile chips, i.e. iPhone A18 Pro, and (b) denoising with dozens of steps to generate a single output further slows down the process. To address these challenges, we propose a three-stage framework to accelerate video diffusion models on the mobile platform. First, we prune from a pre-trained text-to-image diffusion model to obtain an efficient spatial backbone. Second, we inflate the spatial backbone with a novel combination of temporal modules which are searched out with our mobile-oriented metrics. Finally, through adversarial training, our efficient model attains the capability to generate high-quality videos in only 4 steps." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 527, + 141, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 527, + 141, + 538 + ], + "spans": [ + { + "bbox": [ + 55, + 527, + 141, + 538 + ], + "type": "text", + "content": "3.1. Preliminaries" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 544, + 295, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 544, + 295, + 653 + ], + "spans": [ + { + "bbox": [ + 55, + 544, + 295, + 653 + ], + "type": "text", + "content": "Following [77], we employ a spatial-temporal VAE to compress image and video data into the latent space. Given video or image data " + }, + { + "bbox": [ + 55, + 544, + 295, + 653 + ], + "type": "inline_equation", + "content": "\\mathbf{v} \\in \\mathbb{R}^{n \\times 3 \\times H \\times W}" + }, + { + "bbox": [ + 55, + 544, + 295, + 653 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 544, + 295, + 653 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 544, + 295, + 653 + ], + "type": "text", + "content": " is the number of frames with height " + }, + { + "bbox": [ + 55, + 544, + 295, + 653 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 55, + 544, + 295, + 653 + ], + "type": "text", + "content": " and width " + }, + { + "bbox": [ + 55, + 544, + 295, + 653 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 55, + 544, + 295, + 653 + ], + "type": "text", + "content": ", the spatial-temporal encoder, " + }, + { + "bbox": [ + 55, + 544, + 295, + 653 + ], + "type": "inline_equation", + "content": "\\mathbf{E}" + }, + { + "bbox": [ + 55, + 544, + 295, + 653 + ], + "type": "text", + "content": ", maps the data to a latent space. The encoded frames are represented as " + }, + { + "bbox": [ + 55, + 544, + 295, + 653 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0 = \\mathbf{E}(\\mathbf{v})" + }, + { + "bbox": [ + 55, + 544, + 295, + 653 + ], + "type": "text", + "content": ", resulting in " + }, + { + "bbox": [ + 55, + 544, + 295, + 653 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0 \\in \\mathbb{R}^{\\tilde{n} \\times 4 \\times \\tilde{H} \\times \\tilde{W}}" + }, + { + "bbox": [ + 55, + 544, + 295, + 653 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 55, + 544, + 295, + 653 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0 \\sim p_{data}(\\mathbf{x}_0)" + }, + { + "bbox": [ + 55, + 544, + 295, + 653 + ], + "type": "text", + "content": " is a 4-channel latent, with a temporal compression of " + }, + { + "bbox": [ + 55, + 544, + 295, + 653 + ], + "type": "inline_equation", + "content": "\\tilde{n} = n/4" + }, + { + "bbox": [ + 55, + 544, + 295, + 653 + ], + "type": "text", + "content": ", and spatial compression as " + }, + { + "bbox": [ + 55, + 544, + 295, + 653 + ], + "type": "inline_equation", + "content": "\\tilde{H} = H/8" + }, + { + "bbox": [ + 55, + 544, + 295, + 653 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 544, + 295, + 653 + ], + "type": "inline_equation", + "content": "\\tilde{W} = W/8" + }, + { + "bbox": [ + 55, + 544, + 295, + 653 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 653, + 296, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 689 + ], + "type": "text", + "content": "We follow Rectified Flow [59] to train our latent diffusion model. According to the flow-matching-based diffusion form, the intermediate noisy state " + }, + { + "bbox": [ + 55, + 653, + 296, + 689 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t" + }, + { + "bbox": [ + 55, + 653, + 296, + 689 + ], + "type": "text", + "content": " is defined as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 88, + 700, + 295, + 714 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 700, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 88, + 700, + 295, + 714 + ], + "type": "interline_equation", + "content": "\\mathbf {x} _ {t} = (1 - t) \\mathbf {x} _ {0} + t \\epsilon , \\text {w h e r e} \\epsilon \\sim \\mathcal {N} (0, I), \\tag {1}", + "image_path": "96b23f083a194eb704feb86af11a3b1fb6eb5c13540717fb1f2956d9783259b5.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "content": "which is a linear interpolation between the data distribution and a standard normal distribution. The model aims to learn a vector field " + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "inline_equation", + "content": "v_{\\theta}(t, \\mathbf{x}_t)" + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "content": " using the Conditional Flow Matching objective, i.e.," + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 353, + 127, + 553, + 142 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 127, + 553, + 142 + ], + "spans": [ + { + "bbox": [ + 353, + 127, + 553, + 142 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathbb {E} _ {t, \\epsilon , \\mathbf {x} _ {0}} \\| v _ {\\theta} (t, \\mathbf {x} _ {t}) - u _ {t} (\\mathbf {x} _ {t} | \\mathbf {x} _ {0}) \\| _ {2} ^ {2}, \\tag {2}", + "image_path": "4d702e93e929eb4e0f96bb2449498cab5e6659f1d1506d04d167f94ac4c8d214.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 149, + 553, + 174 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 149, + 553, + 174 + ], + "spans": [ + { + "bbox": [ + 313, + 149, + 553, + 174 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 149, + 553, + 174 + ], + "type": "inline_equation", + "content": "u_{t}(\\mathbf{x}_{t}|\\mathbf{x}_{0}) = \\epsilon -\\mathbf{x}_{0}" + }, + { + "bbox": [ + 313, + 149, + 553, + 174 + ], + "type": "text", + "content": " . Following [10], during training, we sample " + }, + { + "bbox": [ + 313, + 149, + 553, + 174 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 149, + 553, + 174 + ], + "type": "text", + "content": " from a logit-normal distribution, i.e.," + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 316, + 181, + 553, + 222 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 181, + 553, + 222 + ], + "spans": [ + { + "bbox": [ + 316, + 181, + 553, + 222 + ], + "type": "interline_equation", + "content": "\\pi (t; m, s) = \\frac {1}{s \\sqrt {2 \\pi}} \\frac {1}{t (1 - t)} \\exp \\left(- \\frac {\\left(\\operatorname {l o g i t} (t) - m\\right) ^ {2}}{2 s ^ {2}}\\right), \\tag {3}", + "image_path": "0dfa476feefdb17a50db4c800bd6fadec51f48858d4f871a742220a28d699eee.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 222, + 553, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 222, + 553, + 247 + ], + "spans": [ + { + "bbox": [ + 313, + 222, + 553, + 247 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 222, + 553, + 247 + ], + "type": "inline_equation", + "content": "\\mathrm{logit}(t) = \\log \\frac{t}{1 - t}" + }, + { + "bbox": [ + 313, + 222, + 553, + 247 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 222, + 553, + 247 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 222, + 553, + 247 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 222, + 553, + 247 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 313, + 222, + 553, + 247 + ], + "type": "text", + "content": " are the location parameter and scale parameter, respectively." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 253, + 494, + 266 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 253, + 494, + 266 + ], + "spans": [ + { + "bbox": [ + 314, + 253, + 494, + 266 + ], + "type": "text", + "content": "3.2. Hardware Efficient Model Design" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 270, + 555, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 270, + 555, + 450 + ], + "spans": [ + { + "bbox": [ + 313, + 270, + 555, + 450 + ], + "type": "text", + "content": "Spatial Backbone. We follow [11, 29] to first prune an efficient text-to-image model as the spatial backbone. Specifically, we start from Stable Diffusion v1.5 [45], and borrow the knowledge from prior arts [29] to remove the most mobile-unfriendly attentions. We then prune the network depth and width following [11] and achieve " + }, + { + "bbox": [ + 313, + 270, + 555, + 450 + ], + "type": "inline_equation", + "content": "\\times 2.5" + }, + { + "bbox": [ + 313, + 270, + 555, + 450 + ], + "type": "text", + "content": " size reduction and more than " + }, + { + "bbox": [ + 313, + 270, + 555, + 450 + ], + "type": "inline_equation", + "content": "10\\times" + }, + { + "bbox": [ + 313, + 270, + 555, + 450 + ], + "type": "text", + "content": " speedup on mobile devices. We include qualitative visualizations of our image model in the supplementary material. Note that we use a UNet denoiser [17], leaving the exploration of DiT [40] to future work. The hierarchical structure of the UNet denoiser forms a good search space to achieve mobile efficiency, while the computation complexity of DiT grows quadratically with the number of tokens (generation resolution), making it challenging for mobile deployment." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 451, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 451, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 451, + 556, + 715 + ], + "type": "text", + "content": "Temporal Layer Design. Current latent video diffusion models typically adopt temporal self-attention [13], cross-attentions [77], and convolutions [3] to model temporal dependencies. CogVideoX [67] demonstrates significant performance gain by using full 3D-Attention, at the cost of more computations and memory consumption. In this work, we enumerate and investigate all types of temporal modeling methods, including Conv1D, Conv3D, SelfAttention1D, SelfAttention3D, CrossAttention1D, and CrossAttention3D, and profile their complexity in Fig. 3. For instance, SelfAttention1D only models temporal dependency on a single coordinate, while SelfAttention3D models global dependencies and has the potential to deliver much stronger performance. However, the computation complexity of SelfAttention3D grows quadratically with respect to " + }, + { + "bbox": [ + 313, + 451, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\tilde{t} \\times \\tilde{H} \\times \\tilde{W}" + }, + { + "bbox": [ + 313, + 451, + 556, + 715 + ], + "type": "text", + "content": ", while SelfAttention1D is linear with respect to " + }, + { + "bbox": [ + 313, + 451, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\tilde{H}" + }, + { + "bbox": [ + 313, + 451, + 556, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 451, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\tilde{W}" + }, + { + "bbox": [ + 313, + 451, + 556, + 715 + ], + "type": "text", + "content": ", which makes SelfAttention3D much more costly at higher resolutions. On the other hand, the computation of CrossAttentionND is determined by both spatial-temporal resolution and the number of tokens from the text encoder. Conv1D and 3D are locality alternatives for SelfAttention1D and 3D, respectively. Though the computation complexity and" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "2481" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 72, + 68, + 556, + 208 + ], + "blocks": [ + { + "bbox": [ + 72, + 68, + 556, + 208 + ], + "lines": [ + { + "bbox": [ + 72, + 68, + 556, + 208 + ], + "spans": [ + { + "bbox": [ + 72, + 68, + 556, + 208 + ], + "type": "image", + "image_path": "f9b1590de9a48f2a35725c8fa2eba8c41862dd629abd454a2a9c3b56b288ea4d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 220, + 555, + 299 + ], + "lines": [ + { + "bbox": [ + 54, + 220, + 555, + 299 + ], + "spans": [ + { + "bbox": [ + 54, + 220, + 555, + 299 + ], + "type": "text", + "content": "Figure 2. Framework Overview. In the Latency and Memory Guided Architecture Search, we freeze the pretrained efficient spatial layers and conduct evolutionary search over temporal layer based on the memory and latency constraint. During the Adversarial Fine-tuning stage, we initialize the discriminator with the weights from the text-to-video model trained in the first stage. The discriminator employs the encoder of the UNet as its backbone, which remains frozen. We add spatial-temporal discriminator heads after each downsampling block, updating only these heads during training. Following prior works [48, 49, 74], each head conditions on pooled text embeddings c projected via a linear layer. Input features are first reshaped to merge the temporal and batch axes for processing through a 2D ResBlock, and then reshaped again to merge spatial dimensions before the temporal self-attention block." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 55, + 316, + 294, + 411 + ], + "blocks": [ + { + "bbox": [ + 55, + 316, + 294, + 411 + ], + "lines": [ + { + "bbox": [ + 55, + 316, + 294, + 411 + ], + "spans": [ + { + "bbox": [ + 55, + 316, + 294, + 411 + ], + "type": "image", + "image_path": "a8da950f93c5dd6710c284a56f977b9163542a38c96c9a2137fe241a50e12324.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 418, + 295, + 464 + ], + "lines": [ + { + "bbox": [ + 55, + 418, + 295, + 464 + ], + "spans": [ + { + "bbox": [ + 55, + 418, + 295, + 464 + ], + "type": "text", + "content": "Figure 3. Computation Complexity and Memory Consumption Analysis. The computation complexity and memory consumption of different temporal layer for various input size. The temporal dimension is fixed to 12 for simplicity." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 474, + 295, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 474, + 295, + 556 + ], + "spans": [ + { + "bbox": [ + 55, + 474, + 295, + 556 + ], + "type": "text", + "content": "memory footprint for each design candidate can be easily profiled, as in Fig. 3, it is still crucial and challenging to build a spatial-temporal network with optimized arrangements of these operators. We propose to perform a latency-memory joint architecture search to determine which, where, and the number of temporal layers to use for our efficient video diffusion model on mobile, as follows." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 558, + 286, + 569 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 286, + 569 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 286, + 569 + ], + "type": "text", + "content": "Latency and Memory Guided Architecture Search." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 570, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 570, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 54, + 570, + 295, + 713 + ], + "type": "text", + "content": "Prior to searching the architecture, we construct a look-up table containing the inference latency and the memory footprint of different temporal layers. For each candidate 1D or 3D operator (i.e., ConvND, SelfAttentionND, CrossAttentionND), we benchmark the latency and memory consumption under different spatial-temporal resolutions on hardware. We then clean the search space by eliminating OOM states. Then we perform evolutionary search to obtain the temporal design with Pareto optimality. The architecture candidate is trained on precomputed video latents for " + }, + { + "bbox": [ + 54, + 570, + 295, + 713 + ], + "type": "inline_equation", + "content": "20K" + }, + { + "bbox": [ + 54, + 570, + 295, + 713 + ], + "type": "text", + "content": " iterations with the spatial backbone frozen, and is evaluated on VBench [19] to obtain the scores as the quality metric." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 319, + 553, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 319, + 553, + 342 + ], + "spans": [ + { + "bbox": [ + 313, + 319, + 553, + 342 + ], + "type": "text", + "content": "We include the detailed search algorithm, action space, and total search time in the supplementary material." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 342, + 554, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 342, + 554, + 426 + ], + "spans": [ + { + "bbox": [ + 313, + 342, + 554, + 426 + ], + "type": "text", + "content": "Image-Video Joint Training. Upon the finalized model architecture, we perform image-video joint training under various clip lengths and aspect ratios with all parameters updated for another " + }, + { + "bbox": [ + 313, + 342, + 554, + 426 + ], + "type": "inline_equation", + "content": "100K" + }, + { + "bbox": [ + 313, + 342, + 554, + 426 + ], + "type": "text", + "content": " iterations. After the joint training, our efficient model is capable of generating videos with various lengths and aspect ratios under a conventional recipe, i.e., 25 steps with classifier-free guidance." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 426, + 555, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 555, + 594 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 555, + 594 + ], + "type": "text", + "content": "VAE Decoder Compression. We use the spatial temporal-decoupled VAE from OpenSora [77], which has 4 latent channels, " + }, + { + "bbox": [ + 313, + 426, + 555, + 594 + ], + "type": "inline_equation", + "content": "8 \\times 8" + }, + { + "bbox": [ + 313, + 426, + 555, + 594 + ], + "type": "text", + "content": " spatial compression and " + }, + { + "bbox": [ + 313, + 426, + 555, + 594 + ], + "type": "inline_equation", + "content": "4 \\times" + }, + { + "bbox": [ + 313, + 426, + 555, + 594 + ], + "type": "text", + "content": " temporal compression. To decode a 17-frame video clip on mobile, the original [77] temporal decoder takes " + }, + { + "bbox": [ + 313, + 426, + 555, + 594 + ], + "type": "inline_equation", + "content": "23,100~\\mathrm{ms}" + }, + { + "bbox": [ + 313, + 426, + 555, + 594 + ], + "type": "text", + "content": ", and the spatial decoder takes " + }, + { + "bbox": [ + 313, + 426, + 555, + 594 + ], + "type": "inline_equation", + "content": "4100~\\mathrm{ms}" + }, + { + "bbox": [ + 313, + 426, + 555, + 594 + ], + "type": "text", + "content": ", which we found to be a bottleneck for the generation speed. To increase the speed, we focus on the decoder only. We freeze the VAE encoders and prune the temporal and spatial decoder on our video and image dataset, respectively. Our efficient temporal decoder runs at " + }, + { + "bbox": [ + 313, + 426, + 555, + 594 + ], + "type": "inline_equation", + "content": "210~\\mathrm{ms}" + }, + { + "bbox": [ + 313, + 426, + 555, + 594 + ], + "type": "text", + "content": " and spatial decoder at " + }, + { + "bbox": [ + 313, + 426, + 555, + 594 + ], + "type": "inline_equation", + "content": "330~\\mathrm{ms}" + }, + { + "bbox": [ + 313, + 426, + 555, + 594 + ], + "type": "text", + "content": " to decode a 17-frame video clip, reaching " + }, + { + "bbox": [ + 313, + 426, + 555, + 594 + ], + "type": "inline_equation", + "content": "50 \\times" + }, + { + "bbox": [ + 313, + 426, + 555, + 594 + ], + "type": "text", + "content": " speedup with negligible loss in quality. Further details about VAE compression are included in the supplementary material." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 600, + 484, + 613 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 600, + 484, + 613 + ], + "spans": [ + { + "bbox": [ + 313, + 600, + 484, + 613 + ], + "type": "text", + "content": "3.3. Latent Adversarial Fine-tuning" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 618, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 618, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 618, + 554, + 713 + ], + "type": "text", + "content": "Our training procedure involves two networks: a generator " + }, + { + "bbox": [ + 313, + 618, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{\\theta}" + }, + { + "bbox": [ + 313, + 618, + 554, + 713 + ], + "type": "text", + "content": " and a discriminator " + }, + { + "bbox": [ + 313, + 618, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{D}\\phi" + }, + { + "bbox": [ + 313, + 618, + 554, + 713 + ], + "type": "text", + "content": ". Similar to prior work [51, 74], we initialize our generator with pre-trained diffusion model weights " + }, + { + "bbox": [ + 313, + 618, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 618, + 554, + 713 + ], + "type": "text", + "content": ", while the discriminator is also partially initialized from " + }, + { + "bbox": [ + 313, + 618, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 618, + 554, + 713 + ], + "type": "text", + "content": ". Specifically, the backbone of the discriminator adopts the same architecture and weights as the pre-trained UNet encoder, with these backbone parameters remaining frozen during training. Additionally, we enhance the dis" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2482" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "content": "criminator with spatial-temporal discriminator heads added after each backbone block, with only these head parameters being updated in the discriminator training phase. As illustrated on the right in Fig. 2, each discriminator head consists of a spatial ResBlock and a temporal self-attention block. This design allows our discriminator to effectively handle both image and video data during fine-tuning. We analyze the impact of joint image-video fine-tuning in Sec. 4.3." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 168, + 294, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 168, + 294, + 251 + ], + "spans": [ + { + "bbox": [ + 55, + 168, + 294, + 251 + ], + "type": "text", + "content": "For a real data sample " + }, + { + "bbox": [ + 55, + 168, + 294, + 251 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0" + }, + { + "bbox": [ + 55, + 168, + 294, + 251 + ], + "type": "text", + "content": ", a noisy data point " + }, + { + "bbox": [ + 55, + 168, + 294, + 251 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t" + }, + { + "bbox": [ + 55, + 168, + 294, + 251 + ], + "type": "text", + "content": " is generated through a forward diffusion process, as described in Eq. (1). We set intermediate timesteps as " + }, + { + "bbox": [ + 55, + 168, + 294, + 251 + ], + "type": "inline_equation", + "content": "0 < T_k < \\dots < T_1 = 1.0" + }, + { + "bbox": [ + 55, + 168, + 294, + 251 + ], + "type": "text", + "content": " and sample " + }, + { + "bbox": [ + 55, + 168, + 294, + 251 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 168, + 294, + 251 + ], + "type": "text", + "content": " from these timesteps, where " + }, + { + "bbox": [ + 55, + 168, + 294, + 251 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 168, + 294, + 251 + ], + "type": "text", + "content": " is the number of timesteps selected for generator training (set to " + }, + { + "bbox": [ + 55, + 168, + 294, + 251 + ], + "type": "inline_equation", + "content": "k = 4" + }, + { + "bbox": [ + 55, + 168, + 294, + 251 + ], + "type": "text", + "content": " in practice). The generator, then, predicts the velocity at " + }, + { + "bbox": [ + 55, + 168, + 294, + 251 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t" + }, + { + "bbox": [ + 55, + 168, + 294, + 251 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 55, + 168, + 294, + 251 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{\\theta}(\\mathbf{x}_t,t)" + }, + { + "bbox": [ + 55, + 168, + 294, + 251 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 251, + 294, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 251, + 294, + 360 + ], + "spans": [ + { + "bbox": [ + 55, + 251, + 294, + 360 + ], + "type": "text", + "content": "To train the discriminator, we first sample a target timestep " + }, + { + "bbox": [ + 55, + 251, + 294, + 360 + ], + "type": "inline_equation", + "content": "t'" + }, + { + "bbox": [ + 55, + 251, + 294, + 360 + ], + "type": "text", + "content": " from a logit-normal distribution, as shown in Eq. (3). Using the forward process in Eq. (1), we obtain the real sample " + }, + { + "bbox": [ + 55, + 251, + 294, + 360 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{t'} = (1 - t')\\mathbf{x}_0 + t'\\epsilon" + }, + { + "bbox": [ + 55, + 251, + 294, + 360 + ], + "type": "text", + "content": ". The fake sample, " + }, + { + "bbox": [ + 55, + 251, + 294, + 360 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}_{t'}" + }, + { + "bbox": [ + 55, + 251, + 294, + 360 + ], + "type": "text", + "content": ", is computed as " + }, + { + "bbox": [ + 55, + 251, + 294, + 360 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}_{t'} = \\mathbf{x}_t + (t' - t)\\cdot \\mathcal{G}_\\theta (t,\\mathbf{x}_t)" + }, + { + "bbox": [ + 55, + 251, + 294, + 360 + ], + "type": "text", + "content": ", as shown in Fig. 4. Following established approaches [48-50, 74], we employ hinge loss [31] as the adversarial objective to enhance performance. The discriminator's goal is to differentiate between real and fake samples by minimizing:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 89, + 368, + 294, + 397 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 368, + 294, + 397 + ], + "spans": [ + { + "bbox": [ + 89, + 368, + 294, + 397 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {\\mathrm {a d v}} ^ {\\mathcal {D}} = \\mathbb {E} _ {t ^ {\\prime}, x _ {0}} [ \\max (0, 1 + \\mathcal {D} _ {\\phi} (\\mathbf {x} _ {t ^ {\\prime}}, t ^ {\\prime})) ] \\tag {4} \\\\ + \\mathbb {E} _ {t, t ^ {\\prime}, x _ {0}} \\left[ \\max \\left(0, 1 - \\mathcal {D} _ {\\phi} \\left(\\hat {\\mathbf {x}} _ {t ^ {\\prime}}, t ^ {\\prime}\\right)\\right) \\right], \\\\ \\end{array}", + "image_path": "38ec22f55b2a1b7133fbadb2a9a005bf82f8fa387e8b5fa5df49a73111dd8383.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 406, + 294, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 406, + 294, + 418 + ], + "spans": [ + { + "bbox": [ + 67, + 406, + 294, + 418 + ], + "type": "text", + "content": "The adversarial objective for the generator is defined as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 117, + 428, + 294, + 444 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 117, + 428, + 294, + 444 + ], + "spans": [ + { + "bbox": [ + 117, + 428, + 294, + 444 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {a d v}} ^ {\\mathcal {G}} = \\mathbb {E} _ {t, t ^ {\\prime}, x _ {0}} \\left[ \\mathcal {D} _ {\\phi} \\left(\\hat {\\mathbf {x}} _ {t ^ {\\prime}}, t ^ {\\prime}\\right) \\right]. \\tag {5}", + "image_path": "9852a1ce7dc09afb04eb115e0da20f39fb453cd3be002c2fe8936a6ccc27dfe8.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 448, + 295, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 448, + 295, + 472 + ], + "spans": [ + { + "bbox": [ + 55, + 448, + 295, + 472 + ], + "type": "text", + "content": "Following [74], we also incorporate a reconstruction objective to enhance stability, defined as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 107, + 482, + 295, + 502 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 482, + 295, + 502 + ], + "spans": [ + { + "bbox": [ + 107, + 482, + 295, + 502 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {r e c o n}} = \\sqrt {\\left\\| \\hat {\\mathbf {x}} _ {0} - \\mathbf {x} _ {0} \\right\\| _ {2} ^ {2} + c ^ {2}} - c, \\tag {6}", + "image_path": "045dfeb17e1547752a34eb322f1838ec21f221b6ea1a97002ee788c58bbb7675.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 510, + 295, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 295, + 533 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 295, + 533 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 510, + 295, + 533 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}_0 = \\mathbf{x}_t - t\\cdot \\mathcal{G}_{\\theta}(t,\\mathbf{x}_t)" + }, + { + "bbox": [ + 55, + 510, + 295, + 533 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 510, + 295, + 533 + ], + "type": "inline_equation", + "content": "c > 0" + }, + { + "bbox": [ + 55, + 510, + 295, + 533 + ], + "type": "text", + "content": " is an adjustable constant." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "text", + "content": "Discussion. Our latent adversarial training pipeline is inspired by SF-V [74]. Similar to SF-V, we set " + }, + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "inline_equation", + "content": "k = 4" + }, + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "text", + "content": " and utilize the part of the pre-trained diffusion model as the backbone for the discriminator. However, our approach introduces several key differences. First, our method is built on an efficient UNet specifically designed for mobile devices, with fewer parameters than SVD [3], making it a more challenging task. Second, we redesign the discriminator heads: instead of using separate spatial and temporal heads, we integrate them into a unified spatial-temporal head for adversarial training. Rather than handling the temporal dimension separately with 1-D convolutional kernels, we incorporate a temporal self-attention layer into the spatial discriminator head after the 2-D ResBlock, forming a spatial-temporal discriminator head. This unified design enables" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 339, + 69, + 531, + 152 + ], + "blocks": [ + { + "bbox": [ + 339, + 69, + 531, + 152 + ], + "lines": [ + { + "bbox": [ + 339, + 69, + 531, + 152 + ], + "spans": [ + { + "bbox": [ + 339, + 69, + 531, + 152 + ], + "type": "image", + "image_path": "a0a32c9effc8517ed4200639e7cd253731deb36c8ae66eca2c83c9db045a799e.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 159, + 555, + 248 + ], + "lines": [ + { + "bbox": [ + 313, + 159, + 555, + 248 + ], + "spans": [ + { + "bbox": [ + 313, + 159, + 555, + 248 + ], + "type": "text", + "content": "Figure 4. Latent Adversarial Fine-tuning. Given a latent " + }, + { + "bbox": [ + 313, + 159, + 555, + 248 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0" + }, + { + "bbox": [ + 313, + 159, + 555, + 248 + ], + "type": "text", + "content": " and a noise latent " + }, + { + "bbox": [ + 313, + 159, + 555, + 248 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 313, + 159, + 555, + 248 + ], + "type": "text", + "content": ", we obtain the intermediate noisy latent " + }, + { + "bbox": [ + 313, + 159, + 555, + 248 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t" + }, + { + "bbox": [ + 313, + 159, + 555, + 248 + ], + "type": "text", + "content": " through a forward diffusion process. The generator then predicts the velocity as " + }, + { + "bbox": [ + 313, + 159, + 555, + 248 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{\\theta}(\\mathbf{x}_t, t)" + }, + { + "bbox": [ + 313, + 159, + 555, + 248 + ], + "type": "text", + "content": ". Using the predicted velocity, we compute " + }, + { + "bbox": [ + 313, + 159, + 555, + 248 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}_0" + }, + { + "bbox": [ + 313, + 159, + 555, + 248 + ], + "type": "text", + "content": " and calculate the reconstruction loss " + }, + { + "bbox": [ + 313, + 159, + 555, + 248 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{recon}}" + }, + { + "bbox": [ + 313, + 159, + 555, + 248 + ], + "type": "text", + "content": " between " + }, + { + "bbox": [ + 313, + 159, + 555, + 248 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0" + }, + { + "bbox": [ + 313, + 159, + 555, + 248 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 159, + 555, + 248 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}_0" + }, + { + "bbox": [ + 313, + 159, + 555, + 248 + ], + "type": "text", + "content": ". For adversarial training, the real sample " + }, + { + "bbox": [ + 313, + 159, + 555, + 248 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{t'}" + }, + { + "bbox": [ + 313, + 159, + 555, + 248 + ], + "type": "text", + "content": " is obtained using the same forward diffusion process, while the fake sample " + }, + { + "bbox": [ + 313, + 159, + 555, + 248 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}_{t'}" + }, + { + "bbox": [ + 313, + 159, + 555, + 248 + ], + "type": "text", + "content": " is computed using " + }, + { + "bbox": [ + 313, + 159, + 555, + 248 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t" + }, + { + "bbox": [ + 313, + 159, + 555, + 248 + ], + "type": "text", + "content": " and the predicted velocity " + }, + { + "bbox": [ + 313, + 159, + 555, + 248 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{\\theta}(\\mathbf{x}_t, t)" + }, + { + "bbox": [ + 313, + 159, + 555, + 248 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 280, + 555, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 280, + 555, + 316 + ], + "spans": [ + { + "bbox": [ + 313, + 280, + 555, + 316 + ], + "type": "text", + "content": "our model to be jointly trained on both image and video data, which, as demonstrated in Sec. 4.3, significantly enhances the performance of the fine-tuned model." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 342, + 394, + 356 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 342, + 394, + 356 + ], + "spans": [ + { + "bbox": [ + 314, + 342, + 394, + 356 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 368, + 555, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 368, + 555, + 511 + ], + "spans": [ + { + "bbox": [ + 313, + 368, + 555, + 511 + ], + "type": "text", + "content": "Training. The efficient image backbone is obtained by pruning the Stable Diffusion v1.5 UNet for " + }, + { + "bbox": [ + 313, + 368, + 555, + 511 + ], + "type": "inline_equation", + "content": "100K" + }, + { + "bbox": [ + 313, + 368, + 555, + 511 + ], + "type": "text", + "content": " iterations on high-quality synthetic image datasets. The model is then fine-tuned for " + }, + { + "bbox": [ + 313, + 368, + 555, + 511 + ], + "type": "inline_equation", + "content": "50K" + }, + { + "bbox": [ + 313, + 368, + 555, + 511 + ], + "type": "text", + "content": " additional iterations to adapt to Rectified-Flow velocity prediction [10] as well as to the Spatial-Temporal VAE [77]. We incorporate QK-norm and RoPE [54] in our network to stabilize training. The workflow for architecture search is discussed in Sec. 3.2. The image model pruning, temporal architecture search, and final model training are conducted on 256 NVIDIA A100 80G GPUs using AdamW optimizer with " + }, + { + "bbox": [ + 313, + 368, + 555, + 511 + ], + "type": "inline_equation", + "content": "5e - 5" + }, + { + "bbox": [ + 313, + 368, + 555, + 511 + ], + "type": "text", + "content": " learning rate and betas values as [0.9, 0.999]." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 516, + 555, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 516, + 555, + 612 + ], + "spans": [ + { + "bbox": [ + 313, + 516, + 555, + 612 + ], + "type": "text", + "content": "Adversarial Fine-tuning is conducted for " + }, + { + "bbox": [ + 313, + 516, + 555, + 612 + ], + "type": "inline_equation", + "content": "6K" + }, + { + "bbox": [ + 313, + 516, + 555, + 612 + ], + "type": "text", + "content": " iterations on 64 NVIDIA A100 GPUs, using the AdamW optimizer with a learning rate of " + }, + { + "bbox": [ + 313, + 516, + 555, + 612 + ], + "type": "inline_equation", + "content": "1e - 7" + }, + { + "bbox": [ + 313, + 516, + 555, + 612 + ], + "type": "text", + "content": " for the generator (i.e., UNet) and " + }, + { + "bbox": [ + 313, + 516, + 555, + 612 + ], + "type": "inline_equation", + "content": "1e - 4" + }, + { + "bbox": [ + 313, + 516, + 555, + 612 + ], + "type": "text", + "content": " for the discriminator heads. We set the betas as [0.9, 0.999] for the generator optimizer, and [0.5, 0.999] for the discriminator optimizer. We set the EMA rate as 0.95 following SF-V [74]. We set " + }, + { + "bbox": [ + 313, + 516, + 555, + 612 + ], + "type": "inline_equation", + "content": "m = -1" + }, + { + "bbox": [ + 313, + 516, + 555, + 612 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 516, + 555, + 612 + ], + "type": "inline_equation", + "content": "s = 1" + }, + { + "bbox": [ + 313, + 516, + 555, + 612 + ], + "type": "text", + "content": " if not otherwise noted." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "type": "text", + "content": "Evaluation. The model is evaluated following the standard benchmarking procedure of VBench [19]. With the 4-step adversarial distilled model, we generate 120-frame horizontal videos at a resolution of " + }, + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "type": "inline_equation", + "content": "432 \\times 768" + }, + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "type": "text", + "content": " using 4 inference steps without employing classifier-free guidance. The generated video is saved at 5 seconds 24 FPS for score testing and qualitative visualization. The mobile demo and detailed demo settings are included in the supplementary material." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2483" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 79, + 63, + 531, + 459 + ], + "blocks": [ + { + "bbox": [ + 79, + 63, + 531, + 459 + ], + "lines": [ + { + "bbox": [ + 79, + 63, + 531, + 459 + ], + "spans": [ + { + "bbox": [ + 79, + 63, + 531, + 459 + ], + "type": "image", + "image_path": "3bafabc39831471381c38708da6d77ba0e4e64ae46b55a307283b3e1eefac486.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 467, + 555, + 500 + ], + "lines": [ + { + "bbox": [ + 55, + 467, + 555, + 500 + ], + "spans": [ + { + "bbox": [ + 55, + 467, + 555, + 500 + ], + "type": "text", + "content": "Figure 5. Video generation on various domains. We employ our model to synthesize videos across diverse domains, with each video containing 120 frames at a resolution of " + }, + { + "bbox": [ + 55, + 467, + 555, + 500 + ], + "type": "inline_equation", + "content": "432 \\times 768" + }, + { + "bbox": [ + 55, + 467, + 555, + 500 + ], + "type": "text", + "content": ". All results are generated through a 4-step inference process. The results demonstrate that our model can produce high-quality, motion-consistent videos featuring various objects across different domains." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 73, + 502, + 541, + 637 + ], + "blocks": [ + { + "bbox": [ + 73, + 502, + 541, + 637 + ], + "lines": [ + { + "bbox": [ + 73, + 502, + 541, + 637 + ], + "spans": [ + { + "bbox": [ + 73, + 502, + 541, + 637 + ], + "type": "table", + "html": "
ModelParams (B)StepsTotal scoreQuality scoreSemantic scoretemporal flickeringmotion smoothnessdynamic degreeaesthetic qualityobject classmultiple objectscolorspatial relationshipappearance styletemporal styleoverall consistency
Kling--81.8583.3975.6899.3099.4046.9461.2187.2468.0589.9073.0319.6224.1726.42
Pyramid Flow2.02081.7284.7469.6299.4999.1264.6363.2686.6750.7182.8759.5320.9123.0926.23
CogVideoX5.05081.6182.7577.0498.6696.9270.9761.9885.2362.1182.8166.3524.9125.3827.59
T2V-Turbo1.6481.0182.5774.7697.4897.3449.1763.0493.9654.6589.9038.6724.4225.5128.16
Emu38.0-80.9684.0968.4398.5798.9379.2759.6486.1744.6488.3468.7320.9223.2624.79
CogVideoX1.65080.9182.1875.8398.8997.7359.8660.8283.3762.6379.4169.9024.8024.3626.66
Pika-1.0--80.6982.9271.7799.7499.5047.5062.0488.7243.0890.5761.0322.2624.2225.94
VideoCrafter-2.01.95080.4482.2073.4298.4197.7342.5063.1392.5540.6692.9235.8625.1325.8428.23
AnimateDiff-V21.22580.2782.9069.7598.7597.7640.8367.1690.9036.8887.4734.6022.4226.0327.04
OpenSora V1.21.23079.7681.3573.3999.5398.5042.3956.8582.2251.8390.0868.5623.9524.5426.85
AnimateLCM1.2479.4282.3667.6598.5298.1640.5667.0191.4129.7684.2437.1420.9725.1626.07
ModelScope1.45075.7578.0566.5498.2895.7966.3952.0682.2538.9881.7233.6823.3925.3725.67
Ours0.6481.1483.4771.8499.3799.2951.1162.1992.2254.3486.6356.2020.1725.1827.42
", + "image_path": "0fd49755ed918fe6ab963ec2ec5c521d4ef6a304c8d150868898d7d84932d492.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 644, + 476, + 656 + ], + "lines": [ + { + "bbox": [ + 133, + 644, + 476, + 656 + ], + "spans": [ + { + "bbox": [ + 133, + 644, + 476, + 656 + ], + "type": "text", + "content": "Table 2. Performance comparison with popular video generation models on VBench [19]." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 658, + 194, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 658, + 194, + 672 + ], + "spans": [ + { + "bbox": [ + 55, + 658, + 194, + 672 + ], + "type": "text", + "content": "4.1. Qualitative Visualization" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": "We show visualizations of our generated videos in Fig. 5. Our model consistently produces high-quality video frames and smooth object movements. To demonstrate the generic" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 659, + 555, + 708 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 659, + 555, + 708 + ], + "spans": [ + { + "bbox": [ + 313, + 659, + 555, + 708 + ], + "type": "text", + "content": "text-to-video generation ability, we show various generation examples, including human, animal, photorealistic and art-styled scenes. We include more video visualizations in the supplementary material." + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2484" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 203, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 203, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 203, + 85 + ], + "type": "text", + "content": "4.2. Quantitative Comparisons" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 89, + 295, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 89, + 295, + 232 + ], + "spans": [ + { + "bbox": [ + 54, + 89, + 295, + 232 + ], + "type": "text", + "content": "We present a comprehensive evaluation of our method against existing popular video generation models on VBench [19], as in Tab. 2. Despite the fact that our model is compact and designated for fast inference on mobile, it achieves higher total score compared to recent arts, including the DiT-based OpenSora-V1.2, CogVideoX-2B [67], and the UNet-based VideoCrafter-2.0 [6]. In addition, compared to the 4-step distilled T2V-Turbo [27] and AnimateLCM [58], our model achieves better performance with more than " + }, + { + "bbox": [ + 54, + 89, + 295, + 232 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 54, + 89, + 295, + 232 + ], + "type": "text", + "content": " reduction in size. The quantitative scores demonstrate the superiority of our efficient model design and the tailored adversarial distillation method." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 232, + 296, + 305 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 232, + 296, + 305 + ], + "spans": [ + { + "bbox": [ + 55, + 232, + 296, + 305 + ], + "type": "text", + "content": "User Study. Human evaluations are conducted between our model and baselines as in the Tab. 3. We generate videos from VBench and Movie Gen Bench prompts and ask human labelers to pick the best results across prompt alignment, aesthetics, and motion. The result shows that our model outperforms baseline metrics by a large margin." + } + ] + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 69, + 313, + 284, + 360 + ], + "blocks": [ + { + "bbox": [ + 69, + 313, + 284, + 360 + ], + "lines": [ + { + "bbox": [ + 69, + 313, + 284, + 360 + ], + "spans": [ + { + "bbox": [ + 69, + 313, + 284, + 360 + ], + "type": "table", + "html": "
Win RatePrompt AlignmentAestheticsMotion
OpenSora-1.224.9%21.2%20.4%
CogVideoX-2B30.7%30.5%30.3%
Ours44.4%48.3%49.3%
", + "image_path": "32560609e60dbf6bad8ce42f903c2450e0b71669b0fadc97f9f528274bb8b0a8.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 409, + 162, + 422 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 409, + 162, + 422 + ], + "spans": [ + { + "bbox": [ + 55, + 409, + 162, + 422 + ], + "type": "text", + "content": "4.3. Ablation Analysis" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 426, + 295, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 426, + 295, + 533 + ], + "spans": [ + { + "bbox": [ + 55, + 426, + 295, + 533 + ], + "type": "text", + "content": "Comparison of Training Data Scheme. We compare the model trained with joint image-video datasets vs. video-only datasets. As shown in Tab. 4, training with video-only datasets leads to significant performance degradation on the VBenchmark score with a drop of 6.32 in aesthetic quality, 2.80 in image quality, and 2.54 in total score. The results highlight the importance of joint image-video training, as the image dataset offers more contextual information and enhances diversity." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "text", + "content": "Model Scaling. We scale up and down the model size by adjusting the number of temporal layers, as shown in Tab. 4, to demonstrate the effectiveness of the proposed temporal architecture search. We can observe that scaling up the model can only marginally improve generative scores (i.e., " + }, + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\times 2" + }, + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "text", + "content": " scale-up only increasing 0.56 in dynamic degree and 0.12 in motion smoothness, and 0.27 in total score). However, both the " + }, + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\times 2" + }, + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\times 4" + }, + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "text", + "content": " models hit the memory bound on iPhone. By dividing the " + }, + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\times 2" + }, + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "text", + "content": " model into more chunks, we test its mobile speed and observe nearly doubled latency. While on the other hand, further scaling down the model results in heavy losses in generation quality (i.e., decreasing 0.56 in dynamic degree and 0.61 in motion smoothness). Our efficient model is a balanced sweet point for quality and on-device performance." + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 318, + 71, + 553, + 128 + ], + "blocks": [ + { + "bbox": [ + 55, + 369, + 295, + 392 + ], + "lines": [ + { + "bbox": [ + 55, + 369, + 295, + 392 + ], + "spans": [ + { + "bbox": [ + 55, + 369, + 295, + 392 + ], + "type": "text", + "content": "Table 3. User Study between OpenSora-1.2, CogVideoX-2B, and our model on prompt alignment, aesthetics, and motion." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 318, + 71, + 553, + 128 + ], + "lines": [ + { + "bbox": [ + 318, + 71, + 553, + 128 + ], + "spans": [ + { + "bbox": [ + 318, + 71, + 553, + 128 + ], + "type": "table", + "html": "
w/ I ScalingΔTΔAQΔIQΔDDΔMSΔTotal
Video-onlyX——-6.32-2.80+4.6-0.17-2.54
×0.5-22%-1.90+0.04-0.56-0.61-1.58
Temporal Scaling✓×2+32%+0.37+0.10+0.56+0.12+0.27
×4+107%+0.49-0.17+0.91+0.17+0.31
", + "image_path": "1b7907e83dbbec425b4b6e7e42d85b02bf63c57e89be74fd7ced92d54b19ccaa.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 137, + 555, + 246 + ], + "lines": [ + { + "bbox": [ + 313, + 137, + 555, + 246 + ], + "spans": [ + { + "bbox": [ + 313, + 137, + 555, + 246 + ], + "type": "text", + "content": "Table 4. Analysis of training data scheme, latency, and quality of efficient architecture. The baseline model adopts the best suitable architecture and is trained with joint image-video datasets. The \"scaling\", and \"T\" indicates the number of the temporal layers, and the latency comparing to the baseline. The \"AQ\", \"IQ\", \"DD\", and \"MS\", are aesthetic quality, image quality, dynamic degree, and motion smoothness in VBench [19]. The benchmark metrics are presented as the differences from the baseline model, where negative values indicate a decrease in performance over the baseline vice versa." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 339, + 253, + 531, + 401 + ], + "blocks": [ + { + "bbox": [ + 339, + 253, + 531, + 401 + ], + "lines": [ + { + "bbox": [ + 339, + 253, + 531, + 401 + ], + "spans": [ + { + "bbox": [ + 339, + 253, + 531, + 401 + ], + "type": "image", + "image_path": "a3bb3988332de3159b74ee15ce5131b91ba4d3089fd12b4dc844aab8a05326f4.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 414, + 555, + 458 + ], + "lines": [ + { + "bbox": [ + 313, + 414, + 555, + 458 + ], + "spans": [ + { + "bbox": [ + 313, + 414, + 555, + 458 + ], + "type": "text", + "content": "Figure 6. Analysis of temporal layers. We ablate temporal layers in different network stages to evaluate their effect. The scores are normalized according to our base model to better demonstrate the difference." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 474, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 474, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 474, + 555, + 715 + ], + "type": "text", + "content": "Effect of Different Temporal Layers. To better understand the roles of the searched temporal layers, we compute the VBench score after systematically removing (i) all temporal layers in the downsample stage, (ii) bottleneck temporal layers, and (iii) all temporal layers in the upsample stage. The model makes reasonable zero-shot generations after temporal layer removal, but we still fine-tune 10K iterations under the same recipe as in Sec. 3.2 for fair comparison. As shown in Fig. 6, removing different temporal layers results in varying degrees of performance degradation across different metrics, and all removing strategies result in a substantial drop in total score, demonstrating that the existence of the searched temporal layers is important, and they play different roles in generation. Specifically, removing temporal layers in upsample blocks results in a more significant loss in imaging quality, subject consistency, and background consistency, suggesting that the up temporal layers play important roles in detail reconstruction. In contrast, bottleneck layers are more important in human action and object class, where global information modeling dom" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2485" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 133 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 133 + ], + "type": "text", + "content": "inates the results. We observe that removing down layers introduces less overall degradations compared to the other two, which is an anticipated phenomenon because the loss of modeling capacity can be mitigated by the subsequent bottleneck and up stage temporal layers after fine-tuning." + } + ] + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 64, + 144, + 288, + 229 + ], + "blocks": [ + { + "bbox": [ + 64, + 144, + 288, + 229 + ], + "lines": [ + { + "bbox": [ + 64, + 144, + 288, + 229 + ], + "spans": [ + { + "bbox": [ + 64, + 144, + 288, + 229 + ], + "type": "table", + "html": "
Headw/Iw/ LadvmTFMOQSTotal
SF-V [74]X✓-198.8630.7583.6064.2579.73
Ours99.2737.8583.6169.0180.69
Ours✓X-197.321.0170.0317.0459.44
-299.2947.6483.9170.5481.23
Ours✓✓-199.3754.3483.4771.8481.14
099.4443.6381.6969.4679.24
199.5624.4177.8458.8874.05
", + "image_path": "b6e2f30b8b17a02dffa0dc40b2d88a3f07d4c256927bedd72d1050fb9c1ce1c0.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 55, + 237, + 296, + 326 + ], + "lines": [ + { + "bbox": [ + 55, + 237, + 296, + 326 + ], + "spans": [ + { + "bbox": [ + 55, + 237, + 296, + 326 + ], + "type": "text", + "content": "Table 5. Analysis of our adversarial fine-tuning scheme. We evaluate the VBench [19] scores for our models using different training schemes. In the results, \"TF\" and \"MO\" denote the temporal flickering and multiple objects sub-scores, respectively, while \"Q\" and \"S\" represent the quality and semantic scores. The table summarizes how varying (1) the type of discriminator head, (2) training with or without an image dataset, and (3) different " + }, + { + "bbox": [ + 55, + 237, + 296, + 326 + ], + "type": "inline_equation", + "content": "t'" + }, + { + "bbox": [ + 55, + 237, + 296, + 326 + ], + "type": "text", + "content": " distributions impact the performance of our model." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 340, + 295, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 340, + 295, + 472 + ], + "spans": [ + { + "bbox": [ + 55, + 340, + 295, + 472 + ], + "type": "text", + "content": "Effect of Discriminator Heads. We compare the effects of our spatial-temporal discriminator heads with the separate spatial and temporal heads proposed in SF-V [74] to demonstrate the effectiveness of our discriminator architecture. For a fair comparison, both models are trained exclusively on video data. We evaluate the models finetuned with different discriminator heads on VBench [19]. As shown in the first two rows of Tab. 5, our discriminator heads result in improvements in both the quality score (83.61 vs. 83.60) and the semantic score (69.01 vs. 64.25) of the generated videos." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 473, + 295, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 473, + 295, + 616 + ], + "spans": [ + { + "bbox": [ + 55, + 473, + 295, + 616 + ], + "type": "text", + "content": "Effect of Joint Image Video Fine-tuning. We examine the impact of incorporating image data during adversarial training, as shown in the second and fourth rows of Tab. 5. The results indicate that fine-tuning the model with image data can slightly decrease the Quality score in the VBench [19] evaluation (from 83.61 to 83.47). However, by leveraging the increased diversity of the image dataset, the model achieves a substantial improvement in semantic performance, particularly in multi-object generation (from 37.85 to 54.34). This enhancement leads to a better overall score compared to the model trained exclusively on video data (81.14 vs. 80.69)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "text", + "content": "Effect of Noise Distribution for Discriminator. Following Eq. (3), the parameters " + }, + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "text", + "content": " control the distribution of " + }, + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "inline_equation", + "content": "t'" + }, + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "text", + "content": ", which determines the noise levels of " + }, + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{t'}" + }, + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}_{t'}" + }, + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "text", + "content": " before they are passed to the discriminator as real and fake samples, respectively. We investigate the effect of different noise distributions on model performance by evaluating the results using VBenchmark [19]. As shown in the last four rows of Tab. 5, increasing " + }, + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "text", + "content": " (resulting in noisier real and fake sam" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "type": "text", + "content": "plies) degrades the quality score (from 83.91 to 77.84) while slightly enhancing the temporal flickering score (from 99.29 to 99.56). Although setting " + }, + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "type": "inline_equation", + "content": "m = -2" + }, + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "type": "text", + "content": " achieves the highest overall score among the experiments (81.23), it performs poorly on multi-object generation. Therefore, in most of our experiments, unless otherwise stated, we use " + }, + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "type": "inline_equation", + "content": "m = -1" + }, + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "type": "text", + "content": ". This setting yields a slightly lower overall score (81.14) but significantly improves semantic performance (71.84 vs. 70.54) and excels in multi-object generation (54.34 vs. 47.64)." + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 318, + 190, + 551, + 243 + ], + "blocks": [ + { + "bbox": [ + 318, + 190, + 551, + 243 + ], + "lines": [ + { + "bbox": [ + 318, + 190, + 551, + 243 + ], + "spans": [ + { + "bbox": [ + 318, + 190, + 551, + 243 + ], + "type": "table", + "html": "
# StepsDDOCAQQualitySemanticTotal
111.3979.2452.8574.4362.7072.09
230.2888.9759.7480.0869.2877.92
451.1192.2262.1983.4771.8481.14
", + "image_path": "7f71196075dede4efe74ca37d257aff9ca615c0cd64a21f08646f922d7af127f.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 313, + 251, + 553, + 297 + ], + "lines": [ + { + "bbox": [ + 313, + 251, + 553, + 297 + ], + "spans": [ + { + "bbox": [ + 313, + 251, + 553, + 297 + ], + "type": "text", + "content": "Table 6. Analysis of the number of inference steps. We measure VBench [19] score with different numbers of inference steps. In the results, \"DD\", \"OC\", and \"AQ\" denote the dynamic degree, object class, and aesthetic quality scores, respectively." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 309, + 555, + 476 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 309, + 555, + 476 + ], + "spans": [ + { + "bbox": [ + 313, + 309, + 555, + 476 + ], + "type": "text", + "content": "Effect of Inference Steps. Our fine-tuned model supports generation with a reduced number of inference steps. We further investigate how varying the number of evaluation steps affects the quality of the generated results, as shown in Tab. 6. While our four-step generation achieves the best performance, even with only two steps, the model still produces reasonable results. Increasing the number of inference steps improves the performance of our model across all metrics. In Tab. 6, we report not only the quality and semantic scores but also scores for dynamic degree, object class, and aesthetic quality. However, reducing the process to a single inference step leads to a significant drop in performance. We leave more aggressive step reductions to future work." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 489, + 466, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 489, + 466, + 502 + ], + "spans": [ + { + "bbox": [ + 313, + 489, + 466, + 502 + ], + "type": "text", + "content": "5. Discussion and Conclusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 510, + 553, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 553, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 553, + 640 + ], + "type": "text", + "content": "In this work, we propose an acceleration framework for the video diffusion model, and for the first time, achieve superfast text-to-video generation on mobile devices. Specifically, we discover an efficient but powerful network architecture through latency and memory joint architecture search for temporal layers. In addition, we propose an improved adversarial fine-tuning technique to distill our model to 4 steps to further speed up generation. Our work is a good starting point for the edge deployment of video diffusion models and we hope to inspire more downstream applications such as video extension and editing." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 642, + 553, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 553, + 712 + ], + "type": "text", + "content": "Limitations. We use a public 4-channel VAE [77] to encode videos to latent space. Recent work has shown that using more latent channels benefits reconstruction details. Another future direction is to further improve the step reduction technique for 1-2 denoising steps, as works [74] on server-level models." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2486" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 91, + 294, + 112 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 294, + 112 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 294, + 112 + ], + "type": "text", + "content": "[1] Luma AI. Dream machine. https://lumalabs.ai/dream-machine,.2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 114, + 269, + 125 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 114, + 269, + 125 + ], + "spans": [ + { + "bbox": [ + 62, + 114, + 269, + 125 + ], + "type": "text", + "content": "[2] Pika AI. Pika 1.5. https://pika.art/try..2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 126, + 295, + 180 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 126, + 295, + 180 + ], + "spans": [ + { + "bbox": [ + 62, + 126, + 295, + 180 + ], + "type": "text", + "content": "[3] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. ArXiv preprint, abs/2311.15127, 2023. 1, 2, 3, 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 182, + 294, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 182, + 294, + 257 + ], + "spans": [ + { + "bbox": [ + 62, + 182, + 294, + 257 + ], + "type": "text", + "content": "[4] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2023, Vancouver, BC, Canada, June 17-24, 2023, pages 22563-22575. IEEE, 2023. 1, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 259, + 294, + 313 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 259, + 294, + 313 + ], + "spans": [ + { + "bbox": [ + 62, + 259, + 294, + 313 + ], + "type": "text", + "content": "[5] Haoxin Chen, Menghan Xia, Yingqing He, Yong Zhang, Xiaodong Cun, Shaoshu Yang, Jinbo Xing, Yaofang Liu, Qifeng Chen, Xintao Wang, et al. Videocrafter1: Open diffusion models for high-quality video generation. ArXiv preprint, abs/2310.19512, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 315, + 294, + 378 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 315, + 294, + 378 + ], + "spans": [ + { + "bbox": [ + 62, + 315, + 294, + 378 + ], + "type": "text", + "content": "[6] Haoxin Chen, Yong Zhang, Xiaodong Cun, Menghan Xia, Xintao Wang, Chao Weng, and Ying Shan. Videocaster2: Overcoming data limitations for high-quality video diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7310-7320, 2024. 2, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 381, + 294, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 381, + 294, + 456 + ], + "spans": [ + { + "bbox": [ + 62, + 381, + 294, + 456 + ], + "type": "text", + "content": "[7] Tsai-Shien Chen, Aliaksandr Siarohin, Willi Menapace, Ekaterina Deyneka, Hsiang-wei Chao, Byung Eun Jeon, Yuwei Fang, Hsin-Ying Lee, Jian Ren, Ming-Hsuan Yang, et al. Panda-70m: Captioning 70m videos with multiple cross-modality teachers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13320-13331, 2024. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 459, + 294, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 459, + 294, + 501 + ], + "spans": [ + { + "bbox": [ + 62, + 459, + 294, + 501 + ], + "type": "text", + "content": "[8] Xinwang Chen, Ning Liu, Yichen Zhu, Feifei Feng, and Jian Tang. Edit: An efficient diffusion transformer framework inspired by human-like sketching. ArXiv preprint, abs/2410.23788, 2024. 1, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 503, + 294, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 503, + 294, + 557 + ], + "spans": [ + { + "bbox": [ + 62, + 503, + 294, + 557 + ], + "type": "text", + "content": "[9] Trung Dao, Thuan Hoang Nguyen, Thanh Le, Duc Vu, Khoi Nguyen, Cuong Pham, and Anh Tran. Swiftbrush v2: Make your one-step diffusion model better than its teacher. In European Conference on Computer Vision, pages 176-192. Springer, 2025. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 559, + 294, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 559, + 294, + 624 + ], + "spans": [ + { + "bbox": [ + 57, + 559, + 294, + 624 + ], + "type": "text", + "content": "[10] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In *Forty-first International Conference on Machine Learning*, 2024. 3, 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 625, + 294, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 625, + 294, + 679 + ], + "spans": [ + { + "bbox": [ + 57, + 625, + 294, + 679 + ], + "type": "text", + "content": "[11] Gongfan Fang, Xinyin Ma, and Xinchao Wang. Structural pruning for diffusion models. In Advances in Neural Information Processing Systems 36: Annual Conference on Neural Information Processing Systems 2023, NeurIPS 2023, New Orleans, LA, USA, December 10 - 16, 2023, 2023. 2, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 681, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 681, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 681, + 294, + 713 + ], + "type": "text", + "content": "[12] Yutang Feng, Sicheng Gao, Yuxiang Bao, Xiaodi Wang, Shumin Han, Juan Zhang, Baochang Zhang, and Angela Yao. Wave: Warping ddim inversion features for zero-shot" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 317, + 74, + 554, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 335, + 74, + 554, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 74, + 554, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 74, + 554, + 95 + ], + "type": "text", + "content": "text-to-video editing. In European Conference on Computer Vision, pages 38-55. Springer, 2025. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 97, + 554, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 97, + 554, + 151 + ], + "spans": [ + { + "bbox": [ + 317, + 97, + 554, + 151 + ], + "type": "text", + "content": "[13] Yuwei Guo, Ceyuan Yang, Anyi Rao, Zhengyang Liang, Yaohui Wang, Yu Qiao, Maneesh Agrawala, Dahua Lin, and Bo Dai. Animatediff:Animate your personalized text-to-image diffusion models without specific tuning. ArXiv preprint, abs/2307.04725, 2023. 1, 2, 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 152, + 554, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 152, + 554, + 196 + ], + "spans": [ + { + "bbox": [ + 317, + 152, + 554, + 196 + ], + "type": "text", + "content": "[14] Yuwei Guo, Ceyuan Yang, Anyi Rao, Maneesh Agrawala, Dahua Lin, and Bo Dai. Sparsectrl: Adding sparse controls to text-to-video diffusion models. In European Conference on Computer Vision, pages 330-348. Springer, 2025. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 198, + 554, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 198, + 554, + 241 + ], + "spans": [ + { + "bbox": [ + 317, + 198, + 554, + 241 + ], + "type": "text", + "content": "[15] Agrim Gupta, Lijun Yu, Kihyuk Sohn, Xiuye Gu, Meera Hahn, Li Fei-Fei, Irfan Essa, Lu Jiang, and José Lezama. Photorealistic video generation with diffusion models. ArXiv preprint, abs/2312.06662, 2023. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 243, + 554, + 264 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 243, + 554, + 264 + ], + "spans": [ + { + "bbox": [ + 317, + 243, + 554, + 264 + ], + "type": "text", + "content": "[16] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. ArXiv preprint, abs/2207.12598, 2022. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 266, + 554, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 266, + 554, + 319 + ], + "spans": [ + { + "bbox": [ + 317, + 266, + 554, + 319 + ], + "type": "text", + "content": "[17] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In Advances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, NeurIPS 2020, December 6-12, 2020, virtual, 2020. 2, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 322, + 554, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 322, + 554, + 376 + ], + "spans": [ + { + "bbox": [ + 317, + 322, + 554, + 376 + ], + "type": "text", + "content": "[18] Wenyi Hong, Ming Ding, Wendi Zheng, Xinghan Liu, and Jie Tang. Cogvideo: Large-scale pretraining for text-to-video generation via transformers. In The Eleventh International Conference on Learning Representations, ICLR 2023, Kigali, Rwanda, May 1-5, 2023. OpenReview.net, 2023. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 377, + 554, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 377, + 554, + 453 + ], + "spans": [ + { + "bbox": [ + 317, + 377, + 554, + 453 + ], + "type": "text", + "content": "[19] Ziqi Huang, Yinan He, Jiashuo Yu, Fan Zhang, Chenyang Si, Yuming Jiang, Yuanhan Zhang, Tianxing Wu, Qingyang Jin, Nattapol Chanpaisit, Yaohui Wang, Xinyuan Chen, Limin Wang, Dahua Lin, Yu Qiao, and Ziwei Liu. VBenchmark: Comprehensive benchmark suite for video generative models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 2, 4, 5, 6, 7, 8, 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 456, + 554, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 456, + 554, + 498 + ], + "spans": [ + { + "bbox": [ + 317, + 456, + 554, + 498 + ], + "type": "text", + "content": "[20] Hyeonho Jeong, Jinho Chang, Geon Yeong Park, and Jong Chul Ye. Dreammotion: Space-time self-similar score distillation for zero-shot video editing. ArXiv preprint, abs/2403.12002, 2024. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 501, + 554, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 501, + 554, + 553 + ], + "spans": [ + { + "bbox": [ + 317, + 501, + 554, + 553 + ], + "type": "text", + "content": "[21] Yang Jin, Zhicheng Sun, Ningyuan Li, Kun Xu, Hao Jiang, Nan Zhuang, Quzhe Huang, Yang Song, Yadong Mu, and Zhouchen Lin. Pyramidal flow matching for efficient video generative modeling. ArXiv preprint, abs/2410.05954, 2024. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 556, + 554, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 556, + 554, + 600 + ], + "spans": [ + { + "bbox": [ + 317, + 556, + 554, + 600 + ], + "type": "text", + "content": "[22] Beomsu Kim, Yu-Guan Hsieh, Michal Klein, Marco Cuturi, Jong Chul Ye, Bahjat Kawar, and James Thornton. Simple reflow: Improved techniques for fast flow models. ArXiv preprint, abs/2410.07815, 2024. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 601, + 554, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 601, + 554, + 634 + ], + "spans": [ + { + "bbox": [ + 317, + 601, + 554, + 634 + ], + "type": "text", + "content": "[23] Jihwan Kim, Junoh Kang, Jinyoung Choi, and Bohyung Han. Fifo-diffusion: Generating infinite videos from text without training. ArXiv preprint, abs/2405.11473, 2024. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 635, + 554, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 635, + 554, + 655 + ], + "spans": [ + { + "bbox": [ + 317, + 635, + 554, + 655 + ], + "type": "text", + "content": "[24] Kuaishou. Kling. https://kling.kuaishou.com/en.2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 658, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 658, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 658, + 554, + 713 + ], + "type": "text", + "content": "[25] Jeong-gi Kwak, Erqun Dong, Yuhe Jin, Hanseok Ko, Shweta Mahajan, and Kwang Moo Yi. Vivid-1-to-3: Novel view synthesis with video diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6775–6785, 2024. 1" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2487" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 85 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 296, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 296, + 85 + ], + "type": "text", + "content": "[26] PKU-Yuan Lab and Tuzhan AI etc. Open-sora-plan, 2024. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 85, + 296, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 85, + 296, + 140 + ], + "spans": [ + { + "bbox": [ + 55, + 85, + 296, + 140 + ], + "type": "text", + "content": "[27] Jiachen Li, Weixi Feng, Tsu-Jui Fu, Xinyi Wang, Sugato Basu, Wenhu Chen, and William Yang Wang. T2v-turbo: Breaking the quality bottleneck of video consistency model with mixed reward feedback. ArXiv preprint, abs/2405.18750, 2024. 2, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 141, + 294, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 141, + 294, + 196 + ], + "spans": [ + { + "bbox": [ + 56, + 141, + 294, + 196 + ], + "type": "text", + "content": "[28] Jiachen Li, Qian Long, Jian Zheng, Xiaofeng Gao, Robinson Piramuthu, Wenhu Chen, and William Yang Wang. T2v-turbo-v2: Enhancing video generation model post-training through data, reward, and conditional guidance design. ArXiv preprint, abs/2410.05677, 2024. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 198, + 295, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 198, + 295, + 274 + ], + "spans": [ + { + "bbox": [ + 56, + 198, + 295, + 274 + ], + "type": "text", + "content": "[29] Yanyu Li, Huan Wang, Qing Jin, Ju Hu, Pavlo Chemerys, Yun Fu, Yanzhi Wang, Sergey Tulyakov, and Jian Ren. Snapfusion: Text-to-image diffusion model on mobile devices within two seconds. In Advances in Neural Information Processing Systems 36: Annual Conference on Neural Information Processing Systems 2023, NeurIPS 2023, New Orleans, LA, USA, December 10 - 16, 2023, 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 275, + 295, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 275, + 295, + 341 + ], + "spans": [ + { + "bbox": [ + 56, + 275, + 295, + 341 + ], + "type": "text", + "content": "[30] Feng Liang, Bichen Wu, Jialiang Wang, Licheng Yu, Kunpeng Li, Yinan Zhao, Ishan Misra, Jia-Bin Huang, Peizhao Zhang, Peter Vajda, et al. Flowvid: Taming imperfect optical flows for consistent video-to-video synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8207-8216, 2024. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 57, + 342, + 295, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 342, + 295, + 364 + ], + "spans": [ + { + "bbox": [ + 57, + 342, + 295, + 364 + ], + "type": "text", + "content": "[31] Jae Hyun Lim and Jong Chul Ye. Geometric gan. ArXiv preprint, abs/1705.02894, 2017. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 57, + 365, + 294, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 365, + 294, + 397 + ], + "spans": [ + { + "bbox": [ + 57, + 365, + 294, + 397 + ], + "type": "text", + "content": "[32] Shanchuan Lin and Xiao Yang. Animatediff-lightning: Cross-model diffusion distillation. ArXiv preprint, abs/2403.12706, 2024. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 57, + 399, + 295, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 399, + 295, + 454 + ], + "spans": [ + { + "bbox": [ + 57, + 399, + 295, + 454 + ], + "type": "text", + "content": "[33] Xingchao Liu, Chengyue Gong, and Qiang Liu. Flow straight and fast: Learning to generate and transfer data with rectified flow. In The Eleventh International Conference on Learning Representations, ICLR 2023, Kigali, Rwanda, May 1-5, 2023. OpenReview.net, 2023. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 57, + 455, + 295, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 455, + 295, + 487 + ], + "spans": [ + { + "bbox": [ + 57, + 455, + 295, + 487 + ], + "type": "text", + "content": "[34] Yu Lu, Yuanzhi Liang, Linchao Zhu, and Yi Yang. Freelong: Training-free long video generation with spectralblend temporal attention. ArXiv preprint, abs/2407.19918, 2024. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 57, + 488, + 295, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 488, + 295, + 533 + ], + "spans": [ + { + "bbox": [ + 57, + 488, + 295, + 533 + ], + "type": "text", + "content": "[35] Xiaofeng Mao, Zhengkai Jiang, Fu-Yun Wang, Wenbing Zhu, Jiangning Zhang, Hao Chen, Mingmin Chi, and Yabiao Wang. Osv: One step is enough for high-quality image to video generation. ArXiv preprint, abs/2409.11367, 2024. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 533, + 295, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 533, + 295, + 599 + ], + "spans": [ + { + "bbox": [ + 57, + 533, + 295, + 599 + ], + "type": "text", + "content": "[36] Kangfu Mei, Maurizio Delbracio, Hossein Talebi, Zhengzhong Tu, Vishal M Patel, and Peyman Milanfar. Codi: Conditional diffusion distillation for higher-fidelity and faster image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9048-9058, 2024. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 601, + 291, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 601, + 291, + 611 + ], + "spans": [ + { + "bbox": [ + 57, + 601, + 291, + 611 + ], + "type": "text", + "content": "[37] MiniMax. Hailuo ai. https://hailuoai(video/.2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 613, + 295, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 613, + 295, + 656 + ], + "spans": [ + { + "bbox": [ + 57, + 613, + 295, + 656 + ], + "type": "text", + "content": "[38] Kepan Nan, Rui Xie, Penghao Zhou, Tiehan Fan, Zhenheng Yang, Zhijie Chen, Xiang Li, Jian Yang, and Ying Tai. Openvid-1m: A large-scale high-quality dataset for text-to-video generation. ArXiv preprint, abs/2407.02371, 2024. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 658, + 295, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 658, + 295, + 690 + ], + "spans": [ + { + "bbox": [ + 57, + 658, + 295, + 690 + ], + "type": "text", + "content": "[39] OpenAI. Video generation models as world simulators. https://openai.com/index/video-generation-models-as-world-simulators/.2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "type": "text", + "content": "[40] William Peebles and Saining Xie. Scalable diffusion models with transformers. In IEEE/CVF International Conference" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "on Computer Vision, ICCV 2023, Paris, France, October 1-6, 2023, pages 4172-4182. IEEE, 2023. 1, 2, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 96, + 553, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 148 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 148 + ], + "type": "text", + "content": "[41] Adam Polyak, Amit Zohar, Andrew Brown, Andros Tjandra, Animesh Sinha, Ann Lee, Apoorv Vyas, Bowen Shi, Chih-Yao Ma, Ching-Yao Chuang, et al. Movie gen: A cast of media foundation models. ArXiv preprint, abs/2410.13720, 2024. 1, 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 150, + 553, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 150, + 553, + 215 + ], + "spans": [ + { + "bbox": [ + 316, + 150, + 553, + 215 + ], + "type": "text", + "content": "[42] Zhiwu Qing, Shiwei Zhang, Jiayu Wang, Xiang Wang, Yujie Wei, Yingya Zhang, Changxin Gao, and Nong Sang. Hierarchical spatio-temporal decoupling for text-to-video generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6635-6645, 2024. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 217, + 553, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 217, + 553, + 261 + ], + "spans": [ + { + "bbox": [ + 316, + 217, + 553, + 261 + ], + "type": "text", + "content": "[43] Haonan Qiu, Menghan Xia, Yong Zhang, Yingqing He, Xintao Wang, Ying Shan, and Ziwei Liu. Freenoise: Tuning-free longer video diffusion via noise rescheduling. ArXiv preprint, abs/2310.15169, 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 262, + 553, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 262, + 553, + 314 + ], + "spans": [ + { + "bbox": [ + 316, + 262, + 553, + 314 + ], + "type": "text", + "content": "[44] Yixuan Ren, Yang Zhou, Jimei Yang, Jing Shi, Difan Liu, Feng Liu, Mingi Kwon, and Abhinav Shrivastava. Customize-a-video: One-shot motion customization of text-to-video diffusion models. ArXiv preprint, abs/2402.14780, 2024. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 316, + 555, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 316, + 555, + 371 + ], + "spans": [ + { + "bbox": [ + 316, + 316, + 555, + 371 + ], + "type": "text", + "content": "[45] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 2, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 372, + 553, + 393 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 372, + 553, + 393 + ], + "spans": [ + { + "bbox": [ + 316, + 372, + 553, + 393 + ], + "type": "text", + "content": "[46] Runway. Gen-3 alpha. https://runwayml.com/research/introducing-gen-3-alpha.2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 394, + 553, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 394, + 553, + 437 + ], + "spans": [ + { + "bbox": [ + 316, + 394, + 553, + 437 + ], + "type": "text", + "content": "[47] Tim Salimans and Jonathan Ho. Progressive distillation for fast sampling of diffusion models. In *The Tenth International Conference on Learning Representations*, ICLR 2022, Virtual Event, April 25-29, 2022. OpenReview.net, 2022. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 438, + 553, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 438, + 553, + 502 + ], + "spans": [ + { + "bbox": [ + 316, + 438, + 553, + 502 + ], + "type": "text", + "content": "[48] Axel Sauer, Kashyap Chitta, Jens Müller, and Andreas Geiger. Projected gans converge faster. In Advances in Neural Information Processing Systems 34: Annual Conference on Neural Information Processing Systems 2021, NeurIPS 2021, December 6-14, 2021, virtual, pages 17480-17492, 2021. 4, 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 504, + 553, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 504, + 553, + 568 + ], + "spans": [ + { + "bbox": [ + 316, + 504, + 553, + 568 + ], + "type": "text", + "content": "[49] Axel Sauer, Tero Karras, Samuli Laine, Andreas Geiger, and Timo Aila. Stylegan-t: Unlocking the power of gans for fast large-scale text-to-image synthesis. In International Conference on Machine Learning, ICML 2023, 23-29 July 2023, Honolulu, Hawaii, USA, pages 30105-30118. PMLR, 2023. 4" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 570, + 553, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 570, + 553, + 602 + ], + "spans": [ + { + "bbox": [ + 316, + 570, + 553, + 602 + ], + "type": "text", + "content": "[50] Axel Sauer, Dominik Lorenz, Andreas Blattmann, and Robin Rombach. Adversarial diffusion distillation. ArXiv preprint, abs/2311.17042, 2023. 3, 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 603, + 553, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 603, + 553, + 647 + ], + "spans": [ + { + "bbox": [ + 316, + 603, + 553, + 647 + ], + "type": "text", + "content": "[51] Axel Sauer, Frederic Boesel, Tim Dockhorn, Andreas Blattmann, Patrick Esser, and Robin Rombach. Fast high-resolution image synthesis with latent adversarial diffusion distillation. ArXiv preprint, abs/2403.12015, 2024. 3, 4" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 647, + 553, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 679 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 679 + ], + "type": "text", + "content": "[52] Yang Song and Prafulla Dhariwal. Improved techniques for training consistency models. ArXiv preprint, abs/2310.14189, 2023. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "text", + "content": "[53] Yang Song, Prafulla Dhariwal, Mark Chen, and Ilya Sutskever. Consistency models. In International Conference on Machine Learning, ICML 2023, 23-29 July 2023," + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2488" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "type": "text", + "content": "Honolulu, Hawaii, USA, pages 32211-32252. PMLR, 2023. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 95, + 295, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 95, + 295, + 138 + ], + "spans": [ + { + "bbox": [ + 56, + 95, + 295, + 138 + ], + "type": "text", + "content": "[54] Jianlin Su, Murtadha Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu. Roformer: Enhanced transformer with rotary position embedding. Neurocomputing, 568:127063, 2024. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 140, + 198, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 140, + 198, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 140, + 198, + 150 + ], + "type": "text", + "content": "[55] Genmo Team. Mochi, 2024. 1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 152, + 295, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 295, + 205 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 295, + 205 + ], + "type": "text", + "content": "[56] Gemini Team, Rohan Anil, Sebastian Borgeaud, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, Katie Millican, et al. Gemini: a family of highly capable multimodal models. ArXiv preprint, abs/2312.11805, 2023. 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 206, + 295, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 206, + 295, + 272 + ], + "spans": [ + { + "bbox": [ + 56, + 206, + 295, + 272 + ], + "type": "text", + "content": "[57] Vikram Voleti, Chun-Han Yao, Mark Boss, Adam Letts, David Pankratz, Dmitry Tochilkin, Christian Laforte, Robin Rombach, and Varun Jampani. Sv3d: Novel multi-view synthesis and 3d generation from a single image using latent video diffusion. In European Conference on Computer Vision, pages 439-457. Springer, 2025. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 273, + 295, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 273, + 295, + 327 + ], + "spans": [ + { + "bbox": [ + 56, + 273, + 295, + 327 + ], + "type": "text", + "content": "[58] Fu-Yun Wang, Zhaoyang Huang, Xiaoyu Shi, Weikang Bian, Guanglu Song, Yu Liu, and Hongsheng Li. *Animatelcm: Accelerating the animation of personalized diffusion models and adapters with decoupled consistency learning*. *ArXiv* preprint, abs/2402.00769, 2024. 3, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 327, + 295, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 327, + 295, + 369 + ], + "spans": [ + { + "bbox": [ + 56, + 327, + 295, + 369 + ], + "type": "text", + "content": "[59] Fu-Yun Wang, Ling Yang, Zhaoyang Huang, Mengdi Wang, and Hongsheng Li. Rectified diffusion: Straightness is not your need in rectified flow. ArXiv preprint, abs/2410.07303, 2024. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 371, + 295, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 371, + 295, + 414 + ], + "spans": [ + { + "bbox": [ + 56, + 371, + 295, + 414 + ], + "type": "text", + "content": "[60] Xiang Wang, Shiwei Zhang, Han Zhang, Yu Liu, Yingya Zhang, Changxin Gao, and Nong Sang. Videolcm: Video latent consistency model. ArXiv preprint, abs/2312.09109, 2023. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 415, + 295, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 415, + 295, + 459 + ], + "spans": [ + { + "bbox": [ + 56, + 415, + 295, + 459 + ], + "type": "text", + "content": "[61] Jianzong Wu, Xiangtai Li, Yanhong Zeng, Jiangning Zhang, Qianyu Zhou, Yining Li, Yunhai Tong, and Kai Chen. Motionbooth: Motion-aware customized text-to-video generation. ArXiv preprint, abs/2406.17758, 2024. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 460, + 295, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 460, + 295, + 536 + ], + "spans": [ + { + "bbox": [ + 56, + 460, + 295, + 536 + ], + "type": "text", + "content": "[62] Jay Zhangjie Wu, Yixiao Ge, Xintao Wang, Stan Weixian Lei, Yuchao Gu, Yufei Shi, Wynne Hsu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. In IEEE/CVF International Conference on Computer Vision, ICCV 2023, Paris, France, October 1-6, 2023, pages 7589-7599. IEEE, 2023. 1, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 537, + 295, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 537, + 295, + 581 + ], + "spans": [ + { + "bbox": [ + 56, + 537, + 295, + 581 + ], + "type": "text", + "content": "[63] Tianxing Wu, Chenyang Si, Yuming Jiang, Ziqi Huang, and Ziwei Liu. Freeinit: Bridging initialization gap in video diffusion models. In European Conference on Computer Vision, pages 378-394. Springer, 2025. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 582, + 295, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 582, + 295, + 624 + ], + "spans": [ + { + "bbox": [ + 56, + 582, + 295, + 624 + ], + "type": "text", + "content": "[64] Enze Xie, Junsong Chen, Junyu Chen, Han Cai, Yujun Lin, Zhekai Zhang, Muyang Li, Yao Lu, and Song Han. Sana: Efficient high-resolution image synthesis with linear diffusion transformers. ArXiv preprint, abs/2410.10629, 2024. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 625, + 295, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 625, + 295, + 679 + ], + "spans": [ + { + "bbox": [ + 56, + 625, + 295, + 679 + ], + "type": "text", + "content": "[65] Yanwu Xu, Yang Zhao, Zhisheng Xiao, and Tingbo Hou. Ufogen: You forward once large scale text-to-image generation via diffusion gans. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8196-8206, 2024. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 681, + 295, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 681, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 56, + 681, + 295, + 714 + ], + "type": "text", + "content": "[66] Ling Yang, Zixiang Zhang, Zhilong Zhang, Xingchao Liu, Minkai Xu, Wentao Zhang, Chenlin Meng, Stefano Ermon, and Bin Cui. Consistency flow matching: Defining" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "type": "text", + "content": "straight flows with velocity consistency. ArXiv preprint, abs/2407.02398, 2024. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 96, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 150 + ], + "type": "text", + "content": "[67] Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, et al. Cogvideox: Text-to-video diffusion models with an expert transformer. ArXiv preprint, abs/2408.06072, 2024. 1, 2, 3, 7, 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 152, + 553, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 553, + 196 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 553, + 196 + ], + "type": "text", + "content": "[68] Tianwei Yin, Michael Gharbi, Taesung Park, Richard Zhang, Eli Shechtman, Fredo Durand, and William T Freeman. Improved distribution matching distillation for fast image synthesis. ArXiv preprint, abs/2405.14867, 2024. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 198, + 553, + 253 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 198, + 553, + 253 + ], + "spans": [ + { + "bbox": [ + 316, + 198, + 553, + 253 + ], + "type": "text", + "content": "[69] Tianwei Yin, Michael Gharbi, Richard Zhang, Eli Shechtman, Fredo Durand, William T Freeman, and Taesung Park. One-step diffusion with distribution matching distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6613-6623, 2024. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 255, + 553, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 255, + 553, + 308 + ], + "spans": [ + { + "bbox": [ + 316, + 255, + 553, + 308 + ], + "type": "text", + "content": "[70] Zhihang Yuan, Pu Lu, Hanling Zhang, Xuefei Ning, Linfeng Zhang, Tianchen Zhao, Shengen Yan, Guohao Dai, and Yu Wang. Ditfastattn: Attention compression for diffusion transformer models. ArXiv preprint, abs/2406.08552, 2024. 1, 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 310, + 553, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 310, + 553, + 365 + ], + "spans": [ + { + "bbox": [ + 316, + 310, + 553, + 365 + ], + "type": "text", + "content": "[71] Yan Zeng, Guoqiang Wei, Jiani Zheng, Jiaxin Zou, Yang Wei, Yuchen Zhang, and Hang Li. Make pixels dance: High-dynamic video generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8850-8860, 2024. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 366, + 553, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 366, + 553, + 421 + ], + "spans": [ + { + "bbox": [ + 316, + 366, + 553, + 421 + ], + "type": "text", + "content": "[72] Yuanhao Zhai, Kevin Lin, Zhengyuan Yang, Linjie Li, Jianfeng Wang, Chung-Ching Lin, David Doermann, Junsong Yuan, and Lijuan Wang. Motion consistency model: Accelerating video diffusion with disentangled motion-appearance distillation. ArXiv preprint, abs/2406.06890, 2024. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 422, + 553, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 422, + 553, + 487 + ], + "spans": [ + { + "bbox": [ + 316, + 422, + 553, + 487 + ], + "type": "text", + "content": "[73] Zheng Zhan, Yushu Wu, Yifan Gong, Zichong Meng, Zhenglun Kong, Changdi Yang, Geng Yuan, Pu Zhao, Wei Niu, and Yanzhi Wang. Fast and memory-efficient video diffusion using streamlined inference. In The Thirty-eighth Annual Conference on Neural Information Processing Systems. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 489, + 553, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 489, + 553, + 544 + ], + "spans": [ + { + "bbox": [ + 316, + 489, + 553, + 544 + ], + "type": "text", + "content": "[74] Zhixing Zhang, Yanyu Li, Yushu Wu, Yanwu Xu, Anil Kag, Ivan Skorokhodov, Willi Menapace, Aliaksandr Siarohin, Junli Cao, Dimitris Metaxas, et al. Sf-v: Single forward video generation model. ArXiv preprint, abs/2406.04324, 2024. 3, 4, 5, 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 545, + 553, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 545, + 553, + 610 + ], + "spans": [ + { + "bbox": [ + 316, + 545, + 553, + 610 + ], + "type": "text", + "content": "[75] Zhixing Zhang, Bichen Wu, Xiaoyan Wang, Yaqiao Luo, Luxin Zhang, Yinan Zhao, Peter Vajda, Dimitris Metaxas, and Licheng Yu. Avid: Any-length video inpainting with diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7162-7172, 2024. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 613, + 553, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 613, + 553, + 668 + ], + "spans": [ + { + "bbox": [ + 316, + 613, + 553, + 668 + ], + "type": "text", + "content": "[76] Rui Zhao, Yuchao Gu, Jay Zhangjie Wu, David Junhao Zhang, Jia-Wei Liu, Weijia Wu, Jussi Keppo, and Mike Zheng Shou. Motiondirector: Motion customization of text-to-video diffusion models. In European Conference on Computer Vision, pages 273–290. Springer, 2025. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "text", + "content": "[77] Zangwei Zheng, Xiangyu Peng, Tianji Yang, Chenhui Shen, Shenggui Li, Hongxin Liu, Yukun Zhou, Tianyi Li, and Yang You. Open-sora: Democratizing efficient video production for all, 2024. 2, 3, 4, 5, 8, 1" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2489" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 298, + 152 + ], + "type": "list", + "angle": 0, + "index": 2, + "blocks": [ + { + "bbox": [ + 55, + 72, + 298, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 298, + 106 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 298, + 106 + ], + "type": "text", + "content": "[78] Yuan Zhou, Qiuyue Wang, Yuxuan Cai, and Huan Yang. Allegro: Open the black box of commercial-level video generation model. ArXiv preprint, abs/2410.15458, 2024. 1, 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 297, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 297, + 152 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 297, + 152 + ], + "type": "text", + "content": "[79] Yupeng Zhou, Daquan Zhou, Ming-Ming Cheng, Jiashi Feng, and Qibin Hou. Storydiffusion: Consistent self-attention for long-range image and video generation. NeurIPS 2024, 2024. 2" + } + ] + } + ], + "index": 1 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "2490" + } + ] + } + ], + "index": 3 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/SnapGen_ Taming High-Resolution Text-to-Image Models for Mobile Devices with Efficient Architectures and Training/d6336171-7440-4517-9fbd-fdd52a399434_content_list.json b/2025/SnapGen_ Taming High-Resolution Text-to-Image Models for Mobile Devices with Efficient Architectures and Training/d6336171-7440-4517-9fbd-fdd52a399434_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..02384704408045e1094096e6bc7ae58e163376c1 --- /dev/null +++ b/2025/SnapGen_ Taming High-Resolution Text-to-Image Models for Mobile Devices with Efficient Architectures and Training/d6336171-7440-4517-9fbd-fdd52a399434_content_list.json @@ -0,0 +1,1743 @@ +[ + { + "type": "text", + "text": "SnapGen: Taming High-Resolution Text-to-Image Models for Mobile Devices with Efficient Architectures and Training", + "text_level": 1, + "bbox": [ + 109, + 130, + 887, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jierun Chen $^{1,2,*}$ Dongting Hu $^{1,3,*}$ Xijie Huang $^{1,2,*}$ Huseyin Coskun $^{1}$ Arpit Sahni $^{1}$ Aarush Gupta $^{1}$ Anujraaj Goyal $^{1}$ Dishani Lahiri $^{1}$ Rajesh Singh $^{1}$ Yerlan Idelbayev $^{1}$ Junli Cao $^{1}$ Yanyu Li $^{1}$ Kwang-Ting Cheng $^{2}$ S.-H. Gary Chan $^{2}$ Mingming Gong $^{3,4}$ Sergey Tulyakov $^{1}$ Anil Kag $^{1,\\dagger}$ Yanwu Xu $^{1,\\dagger}$ Jian Ren $^{1,\\dagger}$", + "bbox": [ + 178, + 202, + 816, + 276 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Snap Inc. $^{2}$ HKUST $^{3}$ The University of Melbourne $^{4}$ MBZUAI", + "bbox": [ + 215, + 290, + 781, + 308 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "* Equal contribution † Equal advising", + "bbox": [ + 341, + 310, + 656, + 328 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Project Page: https://snap-research.github.io/snapgen", + "bbox": [ + 285, + 329, + 705, + 345 + ], + "page_idx": 0 + }, + { + "type": "table", + "img_path": "images/119ed19f8374061c9cba73ed4c3ab43c80a182d6573aa727fe7e85e02e81a024.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
OursPixArt-αLumina-NextSD3-MediumSDXLPlaygroundv2SD3.5-Large
Param0.38B0.6B2B2B2.6B2.6B8.1B
MobileYESNONONONONONO
", + "bbox": [ + 124, + 358, + 861, + 407 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "\"an old raccoon wearing a top hat and holding an apple, oil painting in the style of van gogh, ...\"", + "bbox": [ + 151, + 407, + 790, + 419 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/aacb2837f306dbd5f9799eb9b41fdaa8e940c7d011704296ab0406d4cf3c4d6f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 143, + 417, + 870, + 500 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "\"a dolphin in an astronaut suit, Animals, Simple Detail\"", + "bbox": [ + 151, + 501, + 522, + 513 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e9d2e69529494de4303059780935b53f3466e0207cd4a7a965b2e547d1e4c28e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 143, + 512, + 870, + 594 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "\"..., young sudanese female, glamour, natural, front view, extreme detailed and texture skin, ...\"", + "bbox": [ + 151, + 595, + 772, + 607 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/bf25a52443399be51fd3200f42287d00d2a2fd5514690042d0061c1f9fd468d6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 143, + 607, + 870, + 688 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "... Llama wearing sunglasses standing on the deck of a spaceship with the Earth in the background, ...", + "bbox": [ + 151, + 690, + 810, + 702 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d7930e76b7a075864239cf2e635ecc6c558325029304c0a3d19cee749799ad40.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 143, + 700, + 870, + 782 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "\"... cute monster ... eating a sashimi while sitting at a breakfast table full of fruits and insects\"", + "bbox": [ + 151, + 785, + 790, + 796 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/290a0e2fa4e822994c983446619421f959f3cfc23933ccc1528d4de079fc86f4.jpg", + "image_caption": [ + "Figure 1. Comparison of various text-to-image models in terms of model size, mobile device compatibility, and visual output quality. Our model, with only $379M$ parameters, demonstrates competitive visual quality while being mobile-compatible. Input text prompts are shown above each image grid; all images are generated at $1024^2$ resolution—zoom in for details. More examples are shown in webpage." + ], + "image_footnote": [], + "bbox": [ + 143, + 796, + 870, + 878 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation.", + "bbox": [ + 236, + 0, + 810, + 18 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "Except for this watermark, it is identical to the accepted version;", + "bbox": [ + 323, + 17, + 722, + 30 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 292, + 31, + 754, + 45 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "7997", + "bbox": [ + 482, + 959, + 514, + 970 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 90, + 326, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Existing text-to-image (T2I) diffusion models face several limitations, including large model sizes, slow runtime, and low-quality generation on mobile devices. This paper aims to address all of these challenges by developing an extremely small and fast T2I model that generates high-resolution and high-quality images on mobile platforms. We propose several techniques to achieve this goal. First, we systematically examine the design choices of the network architecture to reduce model parameters and latency, while ensuring high-quality generation. Second, to further improve generation quality, we employ cross-architecture knowledge distillation from a much larger model, using a multi-level approach to guide the training of our model from scratch. Third, we enable a few-step generation by integrating adversarial guidance with knowledge distillation. For the first time, our model SnapGen, demonstrates the generation of $1024^2$ px images on a mobile device around 1.4 seconds. On ImageNet-1K, our model, with only 372M parameters, achieves an FID of 2.06 for $256^2$ px generation. On T2I benchmarks (i.e., GenEval and DPG-Bench), our model with merely 379M parameters, surpasses large-scale models with billions of parameters at a significantly smaller size (e.g., $7 \\times$ smaller than SDXL, $14 \\times$ smaller than IF-XL).", + "bbox": [ + 89, + 121, + 485, + 470 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 497, + 220, + 512 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Large-scale text-to-image (T2I) diffusion models [12, 13, 15, 51, 53, 57-59] have achieved remarkable success in content generation, powering numerous applications like image editing [48, 64, 72, 85] and video creation [50, 54, 78]. However, T2I models often come with substantial model sizes and slow runtime, and deploying them on the cloud raises concerns related to data security and high costs [65].", + "bbox": [ + 89, + 523, + 482, + 628 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address these challenges, there is huge growing interest in developing smaller and faster T2I models through techniques like model compression (e.g., pruning and quantization) [40, 67, 86], step reduction by distillation [77, 80], and efficient attention mechanisms that mitigate the quadratic complexity [46, 74]. Nevertheless, current works still encounter limitations, e.g., low-resolution generation on mobile devices, that constrain their broader application.", + "bbox": [ + 89, + 628, + 482, + 750 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Most importantly, a critical question remains unexplored: how can we train a T2I model from scratch to generate high-quality, high-resolution images on mobile? Such a model would offer substantial advantages in speed, compactness, cost-effectiveness, and secure deployment. To build this model, we introduce several innovations:", + "bbox": [ + 89, + 750, + 482, + 839 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- Efficient Network Architectures: We conduct an in-depth examination of network architectures, including the denoising UNet and Autoencoder (AE), to obtain optimal trade-off between resource usage and performance. Un", + "bbox": [ + 89, + 839, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "like prior works that optimize and compress pre-trained diffusion models [10, 32, 83], we directly focus on macro- and micro-level design choices to achieve a novel architecture that greatly reduces model size and computational complexity, while preserving high-quality generation.", + "bbox": [ + 524, + 90, + 903, + 167 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Improved Training Techniques: We introduce several improvements to train a compact T2I model from scratch. We utilize flow matching [44, 47] as objective, aligning with larger models like SD3 [18] and SD3.5 [3]. This design enables effective knowledge and step distillation, transferring rich representations from large-scale diffusion models to our much smaller one. Further, we propose a multi-level knowledge distillation with a timestep-aware scaling that combines multiple training objectives. Instead of weighting objectives through a linear combination as in prior works [32, 46], we consider target prediction difficulty (i.e., student-teacher difference) across various timesteps in flow matching.", + "- Advanced Step Distillation: We perform step distillation on our model by combining the adversarial training along with the knowledge distillation using a few-step teacher model (i.e., SD3.5-Large-Turbo [5]), enabling ultra-fast high-quality generation with only 4 or 8 steps." + ], + "bbox": [ + 513, + 167, + 903, + 436 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We demonstrate the superior advantages of our approach and model through extensive experiments:", + "bbox": [ + 511, + 439, + 903, + 468 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- On ImageNet-1K [16] class-conditional image generation task, our model achieves the FID comparable to existing works with significantly reduced model size and computation, i.e., half the model size and one-third of the compute resources compared with SiT-XL[49], as in Tab. 1.", + "- For large-scale T2I generation, our UNet model, with only 379M parameters, demonstrates superior generation quality compared to billion-parameter models [42, 53, 87], e.g., improved metrics on benchmark datasets (Tab. 3) and human evaluation (Fig. 8).", + "- Our compressed decoder, trained from scratch, has competitive reconstruction quality compared to commonly used models [18, 53], with more than $36 \\times$ smaller size, enabling the mobile deployment.", + "- Notably, for the first time, we show a T2I model achieving high-resolution generation (e.g., $1024^2$ px) on mobile devices (e.g., iPhone 16 Pro-Max) around 1.4 seconds." + ], + "bbox": [ + 513, + 469, + 903, + 724 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 739, + 653, + 753 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "High-Resolution Text-to-Image Models have emerged with advanced architectures and multi-stage approaches designed to enhance visual fidelity and user customization. SDXL [53] is a pioneering work in this field, employing a refined cascading approach with UNet backbone to generate high-detailed images, resulting in photorealistic outputs that maintain sharpness and clarity. The following studies explore different techniques like more advanced text encoders, better image refinement, or improved dataset preparation,", + "bbox": [ + 511, + 763, + 906, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "7998", + "bbox": [ + 482, + 960, + 514, + 970 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "to obtain better text-image alignment or higher-quality generation [6, 7, 15, 20, 30, 36-38, 41, 45, 48, 69]. However, most of these models contain billions of parameters, making them extremely slow, and not being able to run on resource-contained hardware like mobile devices. In this work, we aim to build a small and fast model that can perform high-resolution generation even on mobile platforms.", + "bbox": [ + 89, + 90, + 480, + 196 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Efficient Diffusion Models address the challenges of bulky model size and long runtime. There have been efforts exploring the architecture optimization to remove redundancy from large models, demonstrating the on-device generation within seconds [11, 40, 67, 86]. However, these models are constrained to low-resolution output, i.e., $512^{2}$ px. To enable efficient high-resolution generation, SANA [74] and LinFusion [46] incorporate linear attention [9, 14, 31] to achieve 1K generation on laptop GPUs. In contrast, we target a broader range of platforms, supporting high-resolution generation (e.g., 1K) directly on mobile devices.", + "bbox": [ + 89, + 200, + 480, + 367 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Knowledge Distillation in Diffusion Models. In the context of diffusion models, previous works focus on distilling large, high-capacity teacher models into more compact, efficient student models within a homogeneous architecture [32, 46]. They reduce the complexity of the model by removing certain components like attention [70] or residual blocks [23], while maintaining the architectural structure. However, our approach diverges from this trend by utilizing a heterogeneous architecture for more aggressively efficient yet challenging distillation.", + "bbox": [ + 89, + 371, + 480, + 523 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Adversarial Step Distillation uses the techniques from adversarial training [22] to reduce the number of diffusion steps, while maintaining high image quality [60, 61]. For example, UFOGen [77] employs a diffusion-GAN formulation [71, 73, 76] to significantly reduce inference time while maintaining competitive performance. DMD2 [79] builds on prior distillation methods by using distribution matching with adversarial loss. Different from exiting works, we conduct step-distillation on a very compact model and train the model along with the knowledge distillation.", + "bbox": [ + 89, + 526, + 480, + 678 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 704, + 181, + 720 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we present how to craft and train a highly efficient T2I model for high-resolution generation. Specifically, starting from the architecture designed in the latent diffusion model [58], we optimize both denoising backbone (Sec. 3.1, Fig. 2, and Fig. 3) and autoencoder (Sec. 3.2 and Fig. 4) to make them compact and fast, even on mobile devices. We then propose the improved training recipe and knowledge distillation (Sec. 3.3 and Fig. 5), empowering a high-performance T2I model. Lastly, we introduce our step distillation to significantly reduce the number of denoising steps for a faster T2I model (Sec. 3.4 and Fig. 7).", + "bbox": [ + 89, + 734, + 482, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Efficient UNet Architecture", + "text_level": 1, + "bbox": [ + 513, + 90, + 759, + 104 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Here we describe the design choices for the denoising UNet. Baseline Architecture. We choose UNet from SDXL [53] as the baseline (Fig. 2(a)) for our backbone since it has superior efficiency and faster convergence [39] than pure transformer-based models [12, 13]. We adjust the UNet into a thinner and shorter model (i.e., reducing the number of transformer blocks from [0, 2, 10] in three stages to [0, 2, 4] and their channel dimensions from [320, 640, 1280] to [256, 512, 896]), and iterate design choices on top of it.", + "bbox": [ + 511, + 113, + 903, + 250 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Evaluation Metrics. We train models on ImageNet-1K [16] under class-conditional generation for 120 epochs, unless specified otherwise, and report the FID score [25] for $256^{2}$ px generation. Similarly to existing work [30], we inject the class conditions through a text template \"a photo of \". We then encode it with a light text encoder to align the pipeline for the T2I generation. We also calculate the number of parameters for different models, floating point operations (FLOPs) (measured on a latent size of $128 \\times 128$ , equivalent to a $1024 \\times 1024$ image after decoding), and the runtime on mobile device (tested on iPhone 15 Pro). Detailed training and evaluation settings can be found in Supplemental Materials. In the following, we introduce the key architectural changes that improve the model.", + "bbox": [ + 511, + 251, + 903, + 460 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Remove Self-Attention from High-Resolution Stages. Self-attention (SA) layer is restricted by its quadratic computational complexity, incurring a heavy computational cost and high memory consumption for high-resolution input. As such, we keep the SA layer only in the lowest-resolution stage while removing it from other higher-resolution stages, i.e., Fig. 2 (b). This leads to $17\\%$ fewer FLOPs and $24\\%$ latency reduction, as shown in Fig. 3. Interestingly, we even observe a performance improvement, i.e., FID decreased to 3.12 from 3.76. We hypothesize that models with SA in high-resolution stages converge more slowly.", + "bbox": [ + 511, + 462, + 903, + 628 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Replace Conv with Expanded Separable Conv. Regular convolution (Conv) is redundant in both parameters and computation. To address this, we replace all Conv layers with the separable convolution (SepConv) [28], composed of a depthwise convolution (DW) followed by a pointwise convolution (PW), as shown in Fig. 2 (c). This replacement reduces parameters by $24\\%$ and latency by $62\\%$ , but also leads to a performance drop (FID increases from 3.12 to 3.38). To address the issue, we expand the intermediate channels. Specifically, the number of channels after the first PW layer is increased with an expansion ratio, and reduced back to the original number after the second PW layer. The expansion ratio is set to 2 to balance the trade-off between performance, latency, and model parameters. Such a design aligns our residual block with the recently proposed Universal Inverted Bottleneck (UIB) block [55]. As a result, our model achieves $15\\%$ fewer parameters, $27\\%$ less computation, and $2.4 \\times$ speedup, while obtaining a lower FID.", + "bbox": [ + 511, + 628, + 903, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "7999", + "bbox": [ + 482, + 960, + 514, + 970 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/32fc1f037603aa4b02b36d78e8d57b7fbeb695f9b7b38d10ddfc5b30f0691153.jpg", + "image_caption": [ + "Figure 2. Efficient UNet. Starting from a thinner and shorter version of the UNet from SDXL (as in (a)), we explore a series of architectural changes, i.e., (b)-(f), to develop a smaller and faster model while retaining high-quality generation performance, as evaluated in Fig. 3." + ], + "image_footnote": [], + "bbox": [ + 94, + 89, + 444, + 407 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/6ebcd4cd81b67055e08577ab90c0ff5d10cb6ffe379d1896340721d001ac9148.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 450, + 97, + 584, + 218 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/aa95f45746635077da11e1c3ab4ec04dba434d96490f47abbed1ac86caeba047.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 607, + 95, + 730, + 219 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/f60a78f2e2c62771c56e9f88618c757660afe5faecd4a583088143b4925d4edd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 753, + 95, + 893, + 220 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/9905b120562358bddc6086f030628c8beb2ec45868c4467990b7fa37c80bffd1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 450, + 233, + 578, + 406 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/50e5b5423661fd31a956e0611a3860e2ee9d56de9a472ac1830420e039b65eb7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 612, + 229, + 888, + 406 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/c1dbff62b401a03d50e2895ca313bc72b820a589531ce0be6086fcd55c9f64e7.jpg", + "image_caption": [ + "Figure 3. Comparisons of Performance and Efficiency for various Design Choices of Efficient UNet. The generation quality is evaluated using FID calculated on ImageNet-1K for $256^{2}$ px generation. The efficiency metrics include model parameters, latency, and FLOPs. FLOPs and latency (on iPhone 15 Pro) are measured with a $128 \\times 128$ latent, equivalent to a $1024 \\times 1024$ decoded image, for one forward pass. We show the architecture enhancements that improve any of the metrics without hurting others." + ], + "image_footnote": [], + "bbox": [ + 117, + 446, + 883, + 595 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Trim FFN Layers. For the layers in the feed-forward network (FFN), the hidden channel expansion ratio is set to 4 by default and further doubled by using the gated unit. This substantially inflates the model parameters, computation, and memory usage. Following MobileDiffusion [86], we examine the efficacy of simply reducing the expansion ratio, as shown in Fig. 2 (d). We show that reducing the expansion ratio to 3 preserves comparable FID performance while reducing both parameters and FLOPs by $12\\%$ .", + "bbox": [ + 88, + 672, + 482, + 809 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Replace MHSA with MQA. Multi-Head Self-Attention (MHSA) requires multiple sets of keys and values for each attention head. In contrast, Multi-Query Attention (MQA) [63] is more efficient by sharing a single set of keys and values across all heads. Replacing MHSA with MQA reduces parameters by $16\\%$ and latency by $9\\%$ , with mini", + "bbox": [ + 88, + 809, + 482, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "mal impact on performance. Interestingly, the $9\\%$ saving in latency exceeds the $6\\%$ decrease in FLOPs, as the reduced memory access enables higher computational intensity.", + "bbox": [ + 511, + 672, + 906, + 719 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Inject Conditions to the First Stage. Cross-attention (CA) blends the conditional information (e.g., textural description) along with the spatial features to generate images that align with the condition. However, the UNet of SDXL only applies CA in transformer blocks starting from the second stage, resulting in the missing conditional guidance for the first stage. In response, we propose to introduce the conditional embeddings from the very first stage, as in Fig. 2(e). Specifically, we replace the residual blocks with transformer blocks that include CA and FFN while without SA layers. This adjustment makes the model smaller, faster, and more efficient while improving FID.", + "bbox": [ + 511, + 719, + 908, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "8000", + "bbox": [ + 482, + 959, + 514, + 970 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Employ QK RMSNorm and RoPE Positional Embeddings. We extend two advanced techniques developed originally for language models, Query-Key (QK) Normalization [24] with RMSNorm [82] and Rotary Position Embedding (RoPE) [66], to enhance the model (Fig. 2 (f)). RMSNorm, applied after the Query-Key projection in the attention mechanism, reduces the risk of softmax saturation without sacrificing model expressiveness while stabilizing training for faster convergence. In addition, we adapt RoPE from one dimension to two dimensions for better supporting higher resolution since it significantly mitigates artifacts like repeated objects. Together, RMSNorm and RoPE introduce negligible computational and parameter overhead, while offering measurable gains in FID performance.", + "bbox": [ + 89, + 90, + 483, + 301 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Discussion. The above optimization results in an efficient and powerful diffusion backbone capable of generating high-resolution images on mobile devices. Before proceeding with large-scale T2I training, we compare the capacity of our model against existing works on ImageNet-1K. We train the model for 1,000 epochs by following the setting from prior work [58]. We evaluate the model using varied CFG [26, 34] across different inference timesteps. As shown in Tab. 1, our efficient UNet achieves comparable FID to SiT-XL [49], while being almost $45\\%$ smaller.", + "bbox": [ + 89, + 301, + 483, + 455 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/7b8efd7541fcd187df855efe04659528f73b501294d752bbade1a1961d367153.jpg", + "table_caption": [ + "Table 1. Class-conditional image generation on ImageNet ${256} \\times {256}$ with CFG. FLOPs are calculated for one forward pass." + ], + "table_footnote": [], + "table_body": "
ModelParam (M)FLOPs (G)FID↓
LDM-4 [58]4001043.60
UViT-L [8]287773.40
UViT-H [8]5011332.29
DiT-XL [52]6751192.27
SiT-XL [49]6751192.06
Ours372382.06
", + "bbox": [ + 138, + 489, + 434, + 598 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2. Tiny and Fast Decoder", + "text_level": 1, + "bbox": [ + 89, + 616, + 302, + 632 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Besides the denoising model, the decoder also takes a significant ratio of total runtime, especially for on-device deployment [40, 86]. Here we introduce a new architecture of decoder (Fig. 4) for efficient high-resolution generation.", + "bbox": [ + 89, + 638, + 483, + 700 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Baseline Decoder. We use the autoencoder (AE) from SD3 [18] as our baseline model (i.e., the same encoder from SD3 AE), due to its superior reconstruction quality. The AE maps an image $X \\in \\mathbb{R}^{H \\times W \\times 3}$ into a lower-dimensional latent $x \\in \\mathbb{R}^{\\frac{H}{f} \\times \\frac{W}{f} \\times c}$ ( $f, c$ as 8, 16 in SD3). The encoded latent $x$ is then decoded back to an image through a decoder. For high-resolution generation, we observe that the decoder in SD3 is very slow on mobile devices. Specifically, it encounters out-of-memory (OOM) errors when generating a $1024^2$ px image on both the ANE processor of the iPhone 15 Pro and the mobile GPU (Tab. 2). To overcome the latency issue, we propose a much smaller and faster decoder.", + "bbox": [ + 89, + 700, + 483, + 883 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Efficient Decoder. We conduct a series of experiments to", + "bbox": [ + 89, + 885, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/d3551e77cf0a3ca96ce432b179f2b5e73f68505bc09ab4844ff816cca53a3bc2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 521, + 89, + 901, + 196 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/12e3ce061452826b47b5537091521f8490242bceb1cbc04f50860da724a92895.jpg", + "image_caption": [ + "Figure 4. Comparisons of Decoder Architecture between (a) SDXL/SD3 decoder and (b) our tiny decoder." + ], + "image_footnote": [], + "bbox": [ + 522, + 199, + 900, + 294 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/7ee9b939e1012da1bbfab0a4b14ed4b98959f9a5d9421ffea5815d6a24f1e34c.jpg", + "table_caption": [ + "Table 2. Performance Comparison of Decoder. PSNR is calculated on COCO 2017 validation set [43]. FLOPs and latency (on iPhone 15 Pro) are measured for decoding a $128 \\times 128$ latent into a $1024 \\times 1024$ image. The decoder from SDXL and SD3 fail to run on the neural engine of mobile, resulting in a huge runtime." + ], + "table_footnote": [], + "table_body": "
DecoderChPSNRParam (M)FLOPs (G)Latency (ms) on ANELatency (ms) on GPU
SDXL [53]424.8949.494970OOM9469
SD3 [18]1627.9249.554970OOMOOM
Ours1627.851.38224174-
", + "bbox": [ + 516, + 412, + 903, + 474 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "decide the efficient decoder with the following key changes compared with the baseline architecture:", + "bbox": [ + 511, + 491, + 903, + 520 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. We remove attention layers to greatly reduce peak memory without a noticeable impact on decoding quality.", + "2. We keep a minimal amount of GroupNorm (GN) to find a trade-off between latency and performance (i.e., mitigating the color shifting).", + "3. We make the decoder thinner (i.e., fewer channels or narrower width) and replace Conv with SepConvs.", + "4. We use fewer residual blocks in high-resolution stages.", + "5. We remove the Conv shortcut in residual blocks and use the upsampling layer for channel transition." + ], + "bbox": [ + 511, + 522, + 903, + 672 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training of the Decoder. We train our decoder with the mean squared error (MSE) loss, lpips loss [84], adversarial loss [22], and discard the KL term [33] as the encoder is fixed. The decoder is trained on $256^2$ image patches with a batch size of 256 and for 1M iterations. As in Tab. 2, our tiny decoder achieves a competitive PSNR score for reconstruction, while being $35.9 \\times$ smaller and $54.4 \\times$ faster for high-resolution generation on mobile devices compared to conventional ones (e.g., the decoder from SDXL and SD3).", + "bbox": [ + 511, + 674, + 906, + 808 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Discussion of Total On-Device Latency. We finally measure T2I model latency for a $1024^{2}$ px generation on iPhone 16 Pro-Max. The decoder takes 119ms, and the per-step latency for the UNet is 274ms. This results in a $1.2\\sim 2.3s$ runtime for 4 to 8 step generation. Note that text encoder runtime is negligible compared to other components [40].", + "bbox": [ + 511, + 810, + 905, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "8001", + "bbox": [ + 482, + 959, + 513, + 970 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/b28475312a143afcb6eee32703245ffd0b1f9142e7b92e05ba5ac55afcf0445b.jpg", + "image_caption": [ + "Figure 5. Overview of Multi-level Knowledge Distillation, where we perform output distillation and feature distillation." + ], + "image_footnote": [], + "bbox": [ + 94, + 89, + 352, + 219 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/2e659881515492e5fdebcca3c7cba7f8389e2d8bb8cd6a89f5b091ca594f0f3e.jpg", + "image_caption": [ + "Figure 6. Mean loss magnitude for task loss $\\mathcal{L}_{\\mathrm{task}}$ and output distillation loss $\\mathcal{L}_{\\mathrm{kd}}$ ." + ], + "image_footnote": [], + "bbox": [ + 366, + 98, + 632, + 219 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c03cef90103a0769718ce76fd5d00f63febcb96b864f8f0f3ed961ec93ee280e.jpg", + "image_caption": [ + "Figure 7. Overview of Adversarial Step Distillation. Output distillation and distribution-matching distillation are performed." + ], + "image_footnote": [], + "bbox": [ + 643, + 90, + 905, + 219 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.3. Training Recipe and Multi-Level Distillation", + "text_level": 1, + "bbox": [ + 89, + 284, + 467, + 301 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To improve the generation quality of our efficient diffusion model, we propose a series of training techniques.", + "bbox": [ + 89, + 308, + 480, + 337 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Flow-based Training and Inference. Rectified Flows (RFs) [44, 47] define the forward process as straight paths connecting the data distribution to a standard normal distribution, i.e.,", + "bbox": [ + 89, + 337, + 482, + 391 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nx _ {t} = \\left(1 - \\sigma_ {t}\\right) x _ {0} + \\sigma_ {t} \\epsilon , \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 210, + 393, + 480, + 407 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $x_0$ is the clean (latent) image, $t$ is the timestep, $\\sigma_t$ is a timestep-dependent factor, and $\\epsilon$ is random noise sampled from $\\mathcal{N}(0,I)$ . The denoising UNet is formulated to predict a velocity field with the objective as", + "bbox": [ + 89, + 411, + 483, + 467 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {t a s k}} = \\mathbb {E} _ {\\epsilon \\sim \\mathcal {N} (0, I), t} \\left[ | | (\\epsilon - x _ {0}) - v _ {\\theta} (x _ {t}, t) | | _ {2} ^ {2} \\right], \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 140, + 477, + 482, + 501 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $v_{\\theta}(x_t, t)$ is the predicted velocity from UNet parameterized by $\\theta$ . To further enhance training stability, we apply logit-normal sampling [18] for the timestep during training, which assigns more samples to the intermediate steps. In the inference stage, we use the Flow-Euler sampler [19], which predicts the next sample based on the velocity, i.e.,", + "bbox": [ + 89, + 508, + 483, + 594 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nx _ {t - 1} = x _ {t} + \\left(\\sigma_ {t - 1} - \\sigma_ {t}\\right) \\cdot v _ {\\theta} \\left(x _ {t}, t\\right). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 171, + 604, + 480, + 619 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To achieve a lower signal-to-noise ratio on high-resolution (i.e., $1024^2$ px) images, we apply a timestep shift similar to SD3 [18] to adjust the scheduling factor $\\sigma_t$ during both training and inference.", + "bbox": [ + 89, + 628, + 482, + 688 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Multi-Level Knowledge Distillation. To improve the generation quality for compact models, one common practice for previous works is applying knowledge distillation to mimic the prediction of scaled-up teacher models [32]. Benefiting from the aligned flow matching objective and (AE) latent space, powerful SD3.5-Large model [4] can be used as the teacher for output distillation. However, we still face challenges due to 1) the heterogeneous architecture between the U-Net and DiT, 2) the scale difference between the distillation loss and task loss, and 3) varying prediction difficulty across different timesteps. To tackle these, we propose a novel multi-level distillation loss and apply timestep-aware scaling to stabilize and accelerate the convergence of distillation. The overview of our knowledge", + "bbox": [ + 89, + 689, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "distillation scheme is shown in Fig. 5 and detailed techniques are elaborated as follows.", + "bbox": [ + 511, + 285, + 903, + 314 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Aside from the task loss defined in Eq. 2, the major objective for knowledge distillation is to supervise our model $\\theta$ directly with the output of teacher model $\\theta_{T}$ , which can be indicated as", + "bbox": [ + 511, + 315, + 903, + 369 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {k d}} = \\mathbb {E} \\left[ \\left| \\left| v _ {\\theta_ {T}} \\left(x _ {t}, t\\right) - v _ {\\theta} \\left(x _ {t}, t\\right) \\right| \\right| _ {2} ^ {2} \\right]. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 593, + 375, + 903, + 400 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Given the capacity gap between the teacher and our model, applying output-level supervision alone leads to instability and slow convergence. Therefore, we further incorporate a cross-architecture feature-level distillation loss as", + "bbox": [ + 511, + 406, + 905, + 462 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {f e a t k d}} = \\mathbb {E} \\left[ \\sum_ {(l _ {T}, l)} | | f _ {\\theta_ {T}} ^ {l _ {T}} (x _ {t}, t) - \\psi \\left(f _ {\\theta} ^ {l} \\left(x _ {t}, t\\right)\\right) | | _ {2} ^ {2} \\right], \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 553, + 467, + 903, + 500 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $f_{\\theta_T}^{l_T}(\\cdot)$ and $f^l (\\cdot)$ indicate the feature output from the $l_{T}$ -th layer and $l$ -th layer in teacher model and student model, respectively. Different from previous work [32, 46], we consider cross-architecture distillation from a DiT to UNet. Since the richest information in transformers sits around the last layer, we set the distillation target to this layer in both models, and use a lightweight trainable projector $\\psi (\\cdot)$ with only two Conv layers to map the student feature to match the dimension of teacher feature. The proposed feature-level distillation loss provides additional supervision to the student model, leading to faster alignment to the generation quality of the teacher model.", + "bbox": [ + 511, + 508, + 905, + 691 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Timestep-Aware Scaling. Weighting multiple objectives has been a major challenge in knowledge distillation, especially in diffusion models. The overall training objectives from previous works [32, 46, 67] are simple linear combination of multiple loss term, i.e.,", + "bbox": [ + 511, + 691, + 903, + 761 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {\\text {t a s k}} + \\lambda_ {1} \\mathcal {L} _ {\\mathrm {k d}} + \\lambda_ {2} \\mathcal {L} _ {\\text {f e a t k d}}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 604, + 770, + 903, + 785 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where the weighting coefficient $\\lambda_{1}$ and $\\lambda_{2}$ are empirically set to constant. However, this baseline setting fails to consider the prediction difficulty in various time steps. We investigate the distribution of empirical risk magnitude of $\\mathcal{L}_{\\mathrm{task}}$ and $\\mathcal{L}_{\\mathrm{kd}}$ across different timestep $t$ during model training. Fig. 6 illustrates that, in intermediate steps, prediction difficulty are lower compared to $t$ closer to 0 or 1.", + "bbox": [ + 511, + 794, + 905, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "8002", + "bbox": [ + 482, + 959, + 514, + 970 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Building on this important observation, we propose a timestep-aware scaling of the objective to close the gap in loss magnitude across different values of $t$ and to account for prediction difficulties at each timestep, as follows:", + "bbox": [ + 89, + 90, + 483, + 151 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {S} \\left(\\mathcal {L} _ {\\text {t a s k}}, \\mathcal {L} _ {\\mathrm {k d}}\\right) = \\mathbb {E} _ {t} \\left[ \\lambda (t) \\cdot \\mathcal {L} _ {\\text {t a s k}} ^ {t} + (1 - \\lambda (t)) \\frac {\\left| \\mathcal {L} _ {\\text {t a s k}} ^ {t} \\right|}{\\left| \\mathcal {L} _ {\\mathrm {k d}} ^ {t} \\right|} \\cdot \\mathcal {L} _ {\\mathrm {k d}} ^ {t} \\right], \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 162, + 483, + 208 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\lambda (t)$ is the normalized standard (location 0, scale 1) logit-norm density function and $|\\cdot |$ indicates the magnitude. In $S$ , we first ensure the same scale between task loss and distillation loss across different $t$ , then apply more teacher supervision where prediction difficulty is higher (i.e., $t$ closer to 0 or 1), and more real data supervision where prediction difficulty is lower (i.e., intermediate timesteps). The proposed scheme considers the variation of timestep $t$ and helps accelerate the distillation training. The final multi-level distillation objective $\\mathcal{L}_{\\mathrm{md}}$ is defined as", + "bbox": [ + 89, + 209, + 483, + 361 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {m d}} = \\mathcal {S} \\left(\\mathcal {L} _ {\\text {t a s k}}, \\mathcal {L} _ {\\mathrm {k d}}\\right) + \\mathcal {S} \\left(\\mathcal {L} _ {\\text {t a s k}}, \\mathcal {L} _ {\\text {f e a t k d}}\\right). \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 142, + 372, + 482, + 388 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.4. Step Distillation", + "text_level": 1, + "bbox": [ + 89, + 400, + 250, + 416 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We take one step further to enhance the sampling efficiency of our model following a distribution-matching-based step distillation scheme. Following Latent Adversarial Diffusion Distillation (LADD) [60], we use a diffusion-GAN hybrid structure to distill our model into fewer steps with the optimization objective as", + "bbox": [ + 89, + 422, + 483, + 513 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\min _ {D _ {\\theta_ {T}}} \\max _ {G _ {\\theta}} \\mathbb {E} \\left[ \\left[ \\log \\left(D _ {\\theta_ {T}} \\left(x _ {t - 1}, t\\right)\\right) \\right] \\right. \\tag {9} \\\\ \\left. + \\left[ \\log \\left(1 - D _ {\\theta_ {T}} \\left(x _ {t - 1} ^ {\\prime}, t\\right)\\right) \\right] - \\mathcal {S} \\left(\\mathcal {L} _ {\\mathrm {t a s k}}, \\mathcal {L} _ {\\mathrm {k d}}\\right) \\right], \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 135, + 525, + 482, + 583 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $D_{\\theta_T}$ is the discriminator model partially initialized with pretrained fewer-step teacher model $\\theta_T$ (SD3.5-Large-Turbo [5]). The large-scale teacher model are only used as the feature extractor and are frozen during distillation. We only train a few linear layers in the discriminator after feature extraction. We sample $x_{t-1} \\sim q(x_{t-1}|x_0)$ and $x_{t-1}' \\sim q(x_{t-1}|x_0')$ , where $x_0'$ is the prediction of our denoising generator $G_\\theta(x_t, t)$ as our student model, and $q(x)$ is the forward process of diffusion model defined in Eq. 1. The objective consists of an adversarial loss to match noisy samples at time step $t-1$ and the output-level distillation loss $S(\\mathcal{L}_{\\mathrm{task}}, \\mathcal{L}_{\\mathrm{kd}})$ after applying timestep-aware scaling. The proposed step distillation, visualized in Fig. 7, can be interpreted as training a diffusion model with adversarial refinement and knowledge distillation, where teacher guidance serves as an additional inductive bias. This advanced step distillation empowers our compact model for high-quality generation, with only a few denoising steps.", + "bbox": [ + 89, + 593, + 483, + 867 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 513, + 90, + 645, + 107 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Model Details. Our T2I pipeline consists of the efficient UNet (Sec. 3.1) and the efficient encoder-decoder model (Sec. 3.2). To obtain text embeddings from the input prompt, we leverage multiple text encoders, namely lightweight CLIP-L [56], CLIP-G [56], and the large Gemma-2-2b language model [68]. We follow SD3 [18] strategy to combine these three text-encoders into a unified rich textual embedding. To enable classifier-free guidance [27], we employ these embeddings with an individual drop-out probability such that we can use an arbitrary subset of the encoders during inference. This allows us to deploy one or more encoders based on the resource constraints.", + "bbox": [ + 511, + 114, + 906, + 296 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Training Recipe. Similar as prior work [30], we use a multi-stage strategy to train our UNet model from scratch. First, we pre-train the model using the ImageNet-1K [16] at 256 resolution as described in Sec. 3.1. Second, we fine-tune this model in a progressive manner from $256 \\rightarrow 512 \\rightarrow 1024$ resolutions. Third, we employ knowledge distillation with our timestep-aware scaling (Sec. 3.3) to improve the finer details in our models using a much larger teacher model (SD3.5-Large [4]) and all three text-encoders. Before KD, we initialize the student model by pre-training on ImageNet and T2I datasets, a cost-effective approach that helps the model learn pixel distributions. Finally, we obtain a few-step model through step distillation using SD3.5-Large-Turbo [5] model as the teacher. We optimize the rectified-flow [18] objective using AdamW optimizer [17] to train our UNet backbone.", + "bbox": [ + 511, + 296, + 906, + 537 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Hyper-parameters. We sample the timesteps using the logit-normal distribution with $(0,1)$ as the location and scale parameters. We use a time shift of 3 for both training and inference. See the Supplemental Materials for details.", + "bbox": [ + 511, + 537, + 906, + 599 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.1. Evaluation", + "text_level": 1, + "bbox": [ + 511, + 604, + 635, + 619 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Quantitative Benchmarks. We use GenEval [21] and DPG-Bench [29] benchmarks to evaluate the text-to-image alignment of our model on short and long prompts, respectively. We report CLIP score on a 6K subset of MS-COCO validation data [43]. In addition, to measure the aesthetic quality of our model, we compute the Image Reward [75] score on selected PixArt prompts [12]. Tab. 3 lists our performance alongside existing state-of-the-art T2I baselines. We provide additional details in Supplemental Materials. We highlight the salient observations below:", + "bbox": [ + 511, + 628, + 905, + 779 + ], + "page_idx": 6 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Our 0.38B parameter model achieves even better performance than significantly larger models such as SDXL (2.6B), Playground (2.6B), and IF-XL (5.5B).", + "- KD non-trivially improves the prompt following ability of the base model as illustrated by an absolute five-point increase in DPG-Bench and GenEval scores.", + "- In terms of aesthetic performance, our model has similar Image Reward scores as Playground models [37, 38]." + ], + "bbox": [ + 511, + 780, + 905, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_footnote", + "text": "For simplicity, we use the $x_0$ prediction in our derivation, and $v$ prediction of rectified flow-based training would not break the formulation.", + "bbox": [ + 89, + 875, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "8003", + "bbox": [ + 482, + 959, + 514, + 970 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/94c542be18b250489df00edc63c1b6e6da90c6efc8471432510b82eadc73552b.jpg", + "image_caption": [ + "Figure 8. Human Evaluation. We conduct a user study to compare images generated by our model against baselines on three attributes: aesthetic quality, text-image alignment, and realistic generations. Our model surpasses the quality of SDXL and SD3 models, while performing competitively against the teacher SD3.5-Large model." + ], + "image_footnote": [], + "bbox": [ + 117, + 88, + 369, + 200 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5be006998992a29e2e6b32849107feb33f38b96359160e05550d226632e95a86.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 372, + 87, + 624, + 202 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e5b8927c2ba31f4da15cb262c040939b608df725b8aa28c709505bfae9842108.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 627, + 88, + 880, + 202 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/93dff0f0c9ac723f7e379a9dd50f9ececcef32867e0446e461372f70a4d5b102.jpg", + "table_caption": [ + "Table 3. Evaluation on Quantitative Benchmarks. We list the scores on GenEval, DPG-Bench, CLIP score on COCO, and Image Reward on aesthetic prompts. We report the parameters for the UNet/DiT backbone in the Param column. Throughput (samples/s) is measured on a single 80GB A100 GPU using the largest batch size supported for each model in a practical scenario to generate $1024^2$ px images. Here our sampling step is set to be 28." + ], + "table_footnote": [], + "table_body": "
ModelParamThroughputGenEval ↑DPG ↑CLIP ↑Image Reward ↑
PixArt-α [12]0.6B0.420.4871.10.3161.15
PixArt-Σ [13]0.6B0.460.5380.50.3171.13
SD-1.5 [1]0.9B-0.4363.20.2870.19
SD-2.1 [2]0.9B-0.5064.20.2810.29
KOALA [35]1.0B0.660.5274.30.3171.05
MicroDiT [62]1.2B-0.5275.10.3181.25
Sana [74]1.6B1.000.6684.80.3271.25
LUMINA-Next [87]2.0B0.060.4674.60.3090.88
SDXL [53]2.6B0.180.5574.70.3010.99
Playgroundv2 [37]2.6B0.180.5974.50.3171.25
Playgroundv2.5 [38]2.6B0.180.5675.50.3191.34
IF-XL [15]5.5B0.060.6175.60.3110.65
Ours w/o KD0.38B1.040.6176.30.3211.20
SnapGen (ours)0.38B1.040.6681.10.3321.32
", + "bbox": [ + 93, + 364, + 483, + 530 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Qualitative Comparison. To visually evaluate the image-text alignment and aesthetics, we compare the generated images from different T2I models in Fig. 1. We observe that many existing models fail to fully capture the full prompt and miss important elements. Further, human generations often result in smoothened-out faces, leading to the loss of details. In contrast, our model generates much more photorealistic images with better image-text alignment.", + "bbox": [ + 88, + 550, + 482, + 671 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Human Evaluation. For a thorough comparison between baselines, we perform a user study with the widely used Parti prompts [81]. We use SDXL, SD3-M, and SD3.5-Large models as the baselines and generate images using this prompt set. We ask the users to select images with better attributes between the baselines and our model. These attributes include image-text alignment, aesthetic quality, and realistic images. Fig. 8 shows that our model convincingly outperforms SDXL on all three attributes. Our model beats SD3 on image-text alignment and realistic generations, with a tie on aesthetic quality. Compared to the teacher (SD3.5-Large), we lag the teacher a bit behind on the image-text alignment, yet our model still has better realistic generations, and yields a toss-up on aesthetic quality. With this study, we can conclude our efficient T2I pipeline", + "bbox": [ + 89, + 674, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "achieves generation quality that is quite comparable to the SD3.5-Large teacher that has 8.1B parameters.", + "bbox": [ + 511, + 266, + 903, + 296 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Few-Step Generation. After step-distillation (Sec. 3.4), our model can generate high-quality images within a few steps. Fig. 9 compares our model's performance before and after step-distillation, with respective GenEval scores. The results demonstrate that our model, after step distillation, achieves comparable performance to the baseline model with 28 steps, even when using only 4 or 8 steps. While the few-step generation shows slight qualitative degradation compared to the 28-step baseline, it still outperforms most existing T2I models with significantly more inference steps, such as SDXL (50 steps) and PixArt- $\\alpha$ (100 steps).", + "bbox": [ + 511, + 297, + 906, + 464 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c1472c3305f61d9c032a6fabeaedf4a68aaac446d3c2036d3c05de1678461366.jpg", + "image_caption": [ + "Figure 9. Performance comparison of few-step generation for our model before (top) and after step distillation (bottom)." + ], + "image_footnote": [], + "bbox": [ + 517, + 472, + 901, + 654 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 722, + 633, + 738 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we propose a novel and efficient T2I model for high-resolution generation on mobile phones. We systematically detail the process to obtain a tiny 379M parameter UNet architecture along with an efficient latent decoder. We devise a novel training method consisting of multi-stage pre-training followed by knowledge distillation from a large teacher and adversarial step distillation. With these, we achieve an extremely efficient T2I model that comprehensively outperforms many existing multi-billion parameter models such as SDXL, Lumina-Next, and Playgroundv2.", + "bbox": [ + 511, + 750, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8004", + "bbox": [ + 482, + 959, + 514, + 970 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 93, + 90, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Stability AI. Stable diffusion 1.5. https://huggingface.co/stable-diffusion-v1-5/stable-diffusion-v1-5, 2022.8", + "[2] Stability AI. Stable diffusion 2.1. https://huggingface.co/stabilityai/stable-diffusion-2-1, 2022.8", + "[3] Stability AI. Stable diffusion 3.5. https://github.com/Stability-AI/sd3.5, 2024.2", + "[4] Stability AI. Stable diffusion 3.5 large. https://huggingface.co/stabilityai/stable-diffusion-3.5-large, 2024.6,7", + "[5] Stability AI. Stable diffusion 3.5 large turbo. https://huggingface.co/stabilityai/stable-diffusion-3.5-large-turbo, 2024. 2, 7", + "[6] Vladimir Arkhipkin, Viacheslav Vasilev, Andrei Filatov, Igor Pavlov, Julia Agafonova, Nikolai Gerasimenko, Anna Averchenkova, Evelina Mironova, Anton Bukashkin, Konstantin Kulikov, et al. Kandinsky 3: Text-to-image synthesis for multifunctional generative framework. arXiv preprint arXiv:2410.21061, 2024. 3", + "[7] Shakhmatov Arseni, Razzhigaev Anton, Nikolich Aleksandr, Arkhipkin Vladimir, Pavlov Igor, Kuznetsov Andrey, and Denis Dimitrov. Kandinsky 2.1, 2023. 3", + "[8] Fan Bao, Shen Nie, Kaiwen Xue, Yue Cao, Chongxuan Li, Hang Su, and Jun Zhu. All are worth words: A vit backbone for diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 22669-22679, 2023. 5", + "[9] Han Cai, Junyan Li, Muyan Hu, Chuang Gan, and Song Han. Efficientvit: Lightweight multi-scale attention for high-resolution dense prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17302-17313, 2023. 3", + "[10] Thibault Castells, Hyoung-Kyu Song, Bo-Kyeong Kim, and Shinkook Choi. LD-Pruner: Efficient Pruning of Latent Diffusion Models using Task-Agnostic Insights. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 821–830, 2024. 2", + "[11] Thibault Castells, Hyoung-Kyu Song, Tairen Piao, Shinkook Choi, Bo-Kyeong Kim, Hanyoung Yim, Changgwun Lee, Jae Gon Kim, and Tae-Ho Kim. EdgeFusion: On-Device Text-to-Image Generation. arXiv preprint arXiv:2404.11925, 2024.3", + "[12] Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, et al. Pixart-α: Fast training of diffusion transformer for photorealistic text-to-image synthesis. arXiv preprint arXiv:2310.00426, 2023. 2, 3, 7, 8", + "[13] Junsong Chen, Chongjian Ge, Enze Xie, Yue Wu, Lewei Yao, Xiaozhe Ren, Zhongdao Wang, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart- $\\cdot$ sigma: Weak-to-strong training of diffusion transformer for 4k text-to-image generation. arXiv preprint arXiv:2403.04692, 2024. 2, 3, 8", + "[14] Tri Dao and Albert Gu. Transformers are ssms: Generalized models and efficient algorithms through structured state space duality. arXiv preprint arXiv:2405.21060, 2024. 3" + ], + "bbox": [ + 93, + 114, + 480, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] DeepFloyd. Deepfloyd. https://github.com/deep-floyd/IF, 2023. 2, 3, 8", + "[16] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 2, 3, 7", + "[17] P Kingma Diederik. Adam: A method for stochastic optimization. (No Title), 2014. 7", + "[18] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In *Forty-first International Conference on Machine Learning*, 2024. 2, 5, 6, 7", + "[19] Leonhard Euler. Institutionum calculi integralis. imp. Acad. imp. Sa'ent., 1768. 6", + "[20] Peng Gao, Le Zhuo, Ziyi Lin, Chris Liu, Junsong Chen, Ruoyi Du, Enze Xie, Xu Luo, Longtian Qiu, Yuhang Zhang, et al. Lumina-T2X: Transforming Text into Any Modality, Resolution, and Duration via Flow-based Large Diffusion Transformers. arXiv preprint arXiv:2405.05945, 2024. 3", + "[21] Dhruba Ghosh, Hannaneh Hajishirzi, and Ludwig Schmidt. Geneval: An object-focused framework for evaluating text-to-image alignment. Advances in Neural Information Processing Systems, 36, 2024. 7", + "[22] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 3, 5", + "[23] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 3", + "[24] Alex Henry, Prudhvi Raj Dachapally, Shubham Pawar, and Yuxuan Chen. Query-key normalization for transformers. arXiv preprint arXiv:2010.04245, 2020. 5", + "[25] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 3", + "[26] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 5", + "[27] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 7", + "[28] Andrew G Howard. Mobilenets: Efficient convolutional neural networks for mobile vision applications. arXiv preprint arXiv:1704.04861, 2017. 3", + "[29] Xiwei Hu, Rui Wang, Yixiao Fang, Bin Fu, Pei Cheng, and Gang Yu. Ella: Equip diffusion models with llm for enhanced semantic alignment. arXiv preprint arXiv:2403.05135, 2024. 7", + "[30] Anil Kag, Huseyin Coskun, Jierun Chen, Junli Cao, Willi Menapace, Aliaksandr Siarohin, Sergey Tulyakov, and Jian Ren. Ascan: Asymmetric convolution-attention networks for efficient recognition and generation. arXiv preprint arXiv:2411.04967, 2024. 3, 7" + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "8005", + "bbox": [ + 482, + 960, + 514, + 970 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[31] Angelos Katharopoulos, Apoorv Vyas, Nikolaos Pappas, and François Fleuret. Transformers are rnns: Fast autoregressive transformers with linear attention. In International conference on machine learning, pages 5156-5165. PMLR, 2020. 3", + "[32] Bo-Kyeong Kim, Hyoung-Kyu Song, Thibault Castells, and Shinkook Choi. Bk-sdm: Architecturally Compressed Stable Diffusion for Efficient Text-to-Image Generation. In Workshop on Efficient Systems for Foundation Models@ ICML2023, 2023. 2, 3, 6", + "[33] Diederik P Kingma. Auto-encoding variational bayes. 2nd International Conference on Learning Representations, 2014. 5", + "[34] Tuomas Kynkänniemi, Miika Aittala, Tero Karras, Samuli Laine, Timo Aila, and Jaakko Lehtinen. Applying guidance in a limited interval improves sample and distribution quality in diffusion models. arXiv preprint arXiv:2404.07724, 2024.5", + "[35] Youngwan Lee, Kwanyong Park, Yoorhim Cho, Yong-Ju Lee, and Sung Ju Hwang. Koala: Empirical lessons toward memory-efficient and fast diffusion models for text-to-image synthesis, 2023. 8", + "[36] Daiqing Li, Aleks Kamko, Ali Sabet, Ehsan Akhgari, Linmiao Xu, and Suhail Doshi. Playground v1, . 3", + "[37] Daiqing Li, Aleks Kamko, Ali Sabet, Ehsan Akhgari, Linmiao Xu, and Suhail Doshi. Playground v2, . 7, 8", + "[38] Daiqing Li, Aleks Kamko, Ehsan Akhgari, Ali Sabet, Linmiao Xu, and Suhail Doshi. Playground V2. 5: Three Insights towards Enhancing Aesthetic Quality in Text-to-Image Generation. arXiv preprint arXiv:2402.17245, 2024. 3, 7, 8", + "[39] Hao Li, Yang Zou, Ying Wang, Orchid Majumder, Yusheng Xie, R Manmatha, Ashwin Swaminathan, Zhuowen Tu, Stefano Ermon, and Stefano Soatto. On the scalability of diffusion-based text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9400-9409, 2024. 3", + "[40] Yanyu Li, Huan Wang, Qing Jin, Ju Hu, Pavlo Chemerys, Yun Fu, Yanzhi Wang, Sergey Tulyakov, and Jian Ren. Snapfusion: Text-to-Image Diffusion Model on Mobile Devices within Two Seconds. Advances in Neural Information Processing Systems, 36, 2024. 2, 3, 5", + "[41] Zhimin Li, Jianwei Zhang, Qin Lin, Jiangfeng Xiong, Yanxin Long, Xinchi Deng, Yingfang Zhang, Xingchao Liu, Minbin Huang, Zedong Xiao, et al. Hunyuan-DiT: A Powerful Multi-Resolution Diffusion Transformer with Fine-Grained Chinese Understanding. arXiv preprint arXiv:2405.08748, 2024. 3", + "[42] Shanchuan Lin, Anran Wang, and Xiao Yang. SDXL-Lightning: Progressive Adversarial Diffusion Distillation, 2024. 2", + "[43] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 5, 7" + ], + "bbox": [ + 91, + 92, + 480, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[44] Yaron Lipman, Ricky TQ Chen, Heli Ben-Hamu, Maximilian Nickel, and Matt Le. Flow matching for generative modeling. arXiv preprint arXiv:2210.02747, 2022. 2, 6", + "[45] Bingchen Liu, Ehsan Akhgari, Alexander Visheratin, Aleks Kamko, Linmiao Xu, Shivam Shrirao, Joao Souza, Suhail Doshi, and Daiqing Li. Playground v3: Improving text-to-image alignment with deep-fusion large language models. arXiv preprint arXiv:2409.10695, 2024. 3", + "[46] Songhua Liu, Weihao Yu, Zhenxiong Tan, and Xinchao Wang. Linfusion: 1gpu, 1 minute, 16k image. arXiv preprint arXiv:2409.02097, 2024. 2, 3, 6", + "[47] Xingchao Liu, Chengyue Gong, and Qiang Liu. Flow straight and fast: Learning to generate and transfer data with rectified flow. arXiv preprint arXiv:2209.03003, 2022. 2, 6", + "[48] Xian Liu, Jian Ren, Aliaksandr Siarohin, Ivan Skorokhodov, Yanyu Li, Dahua Lin, Xihui Liu, Ziwei Liu, and Sergey Tulyakov. Hyperhuman: Hyper-realistic human generation with latent structural diffusion. arXiv preprint arXiv:2310.08579, 2023. 2, 3", + "[49] Nanye Ma, Mark Goldstein, Michael S Albergo, Nicholas M Boffi, Eric Vanden-Eijnden, and Saining Xie. Sit: Exploring flow and diffusion-based generative models with scalable interpolant transformers. arXiv preprint arXiv:2401.08740, 2024. 2, 5", + "[50] Willi Menapace, Aliaksandr Siarohin, Ivan Skorokhodov, Ekaterina Deyneka, Tsai-Shien Chen, Anil Kag, Yuwei Fang, Aleksei Stoliar, Elisa Ricci, Jian Ren, et al. Snap video: Scaled spatiotemporal transformers for text-to-video synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7038-7048, 2024. 2", + "[51] OpenAI. 2", + "[52] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4195-4205, 2023. 5", + "[53] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 2, 3, 5, 8", + "[54] Adam Polyak, Amit Zohar, Andrew Brown, Andros Tjandra, Animesh Sinha, Ann Lee, Apoorv Vyas, Bowen Shi, Chih-Yao Ma, Ching-Yao Chuang, et al. Movie gen: A cast of media foundation models. arXiv preprint arXiv:2410.13720, 2024. 2", + "[55] Danfeng Qin, Chas Leichner, Manolis Delakis, Marco Fornoni, Shixin Luo, Fan Yang, Weijun Wang, Colby Banbury, Chengxi Ye, Berkin Akin, et al. Mobilenetv4-universal models for the mobile ecosystem. arXiv preprint arXiv:2404.10518, 2024.3", + "[56] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 7" + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "8006", + "bbox": [ + 482, + 960, + 514, + 970 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[57] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In Proceedings of the 38th International Conference on Machine Learning, pages 8821-8831. PMLR, 2021. 2", + "[58] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 3, 5", + "[59] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S. Sara Mahdavi, Rapha Gontijo Lopes, Tim Salimans, Jonathan Ho, David J Fleet, and Mohammad Norouzi. Photorealistic text-to-image diffusion models with deep language understanding, 2022. 2", + "[60] Axel Sauer, Frederic Boesel, Tim Dockhorn, Andreas Blattmann, Patrick Esser, and Robin Rombach. Fast high-resolution image synthesis with latent adversarial diffusion distillation. arXiv preprint arXiv:2403.12015, 2024. 3, 7", + "[61] Axel Sauer, Dominik Lorenz, Andreas Blattmann, and Robin Rombach. Adversarial diffusion distillation. In European Conference on Computer Vision, pages 87-103. Springer, 2024. 3", + "[62] Vikash Sehwag, Xianghao Kong, Jingtao Li, Michael Spranger, and Lingjuan Lyu. Stretching each dollar: Diffusion training from scratch on a micro-budget. arXiv preprint arXiv:2407.15811, 2024. 8", + "[63] Noam Shazeer. Fast transformer decoding: One write-head is all you need. arXiv preprint arXiv:1911.02150, 2019. 4", + "[64] Shelly Sheynin, Adam Polyak, Uriel Singer, Yuval Kirstain, Amit Zohar, Oron Ashual, Devi Parikh, and Yaniv Taigman. Emu edit: Precise image editing via recognition and generation tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8871-8879, 2024. 2", + "[65] Yuda Song, Zehao Sun, and Xuanwu Yin. Sdxs: Realtime one-step latent diffusion models with image conditions. arXiv preprint arXiv:2403.16627, 2024. 2", + "[66] Jianlin Su, Murtadha Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu. Roformer: Enhanced transformer with rotary position embedding. Neurocomputing, 568:127063, 2024. 5", + "[67] Yang Sui, Yanyu Li, Anil Kag, Yerlan Idelbayev, Junli Cao, Ju Hu, Dhritiman Sagar, Bo Yuan, Sergey Tulyakov, and Jian Ren. Bitsfusion: 1.99 bits weight quantization of diffusion model. arXiv preprint arXiv:2406.04333, 2024. 2, 3, 6", + "[68] Gemma Team, Morgane Riviere, Shreya Pathak, Pier Giuseppe Sessa, Cassidy Hardin, Surya Bhupatiraju, Léonard Hussenot, Thomas Mesnard, Bobak Shahriari, Alexandre Ramé, et al. Gemma 2: Improving open language models at a practical size. arXiv preprint arXiv:2408.00118, 2024. 7", + "[69] Kcolors Team. Kcolors: Effective Training of Diffusion Model for Photorealistic Text-to-Image Synthesis. arXiv preprint, 2024. 3", + "[70] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 3" + ], + "bbox": [ + 91, + 90, + 482, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[71] Zhendong Wang, Huangjie Zheng, Pengcheng He, Weizhu Chen, and Mingyuan Zhou. Diffusion-gan: Training gans with diffusion. arXiv preprint arXiv:2206.02262, 2022. 3", + "[72] Shitao Xiao, Yueze Wang, Junjie Zhou, Huaying Yuan, Xingrun Xing, Ruiran Yan, Shuting Wang, Tiejun Huang, and Zheng Liu. Omnigen: Unified image generation. arXiv preprint arXiv:2409.11340, 2024. 2", + "[73] Zhisheng Xiao, Karsten Kreis, and Arash Vahdat. Tackling the generative learning trilemma with denoising diffusion gans. arXiv preprint arXiv:2112.07804, 2021. 3", + "[74] Enze Xie, Junsong Chen, Junyu Chen, Han Cai, Yujun Lin, Zhekai Zhang, Muyang Li, Yao Lu, and Song Han. Sana: Efficient high-resolution image synthesis with linear diffusion transformers. arXiv preprint arXiv:2410.10629, 2024. 2, 3, 8", + "[75] Jiazheng Xu, Xiao Liu, Yuchen Wu, Yuxuan Tong, Qinkai Li, Ming Ding, Jie Tang, and Yuxiao Dong. Imagereward: Learning and evaluating human preferences for text-to-image generation. Advances in Neural Information Processing Systems, 36, 2024. 7", + "[76] Yanwu Xu, Mingming Gong, Shaoan Xie, Wei Wei, Matthias Grundmann, Kayhan Batmanghelich, and Tingbo Hou. Semi-implicit denoising diffusion models (siddms). Advances in neural information processing systems, 36:17383, 2023. 3", + "[77] Yanwu Xu, Yang Zhao, Zhisheng Xiao, and Tingbo Hou. Ufogen: You forward once large scale text-to-image generation via diffusion gans. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8196-8206, 2024. 2, 3", + "[78] Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, et al. Cogvideox: Text-to-video diffusion models with an expert transformer. arXiv preprint arXiv:2408.06072, 2024. 2", + "[79] Tianwei Yin, Michael Gharbi, Taesung Park, Richard Zhang, Eli Shechtman, Fredo Durand, and William T Freeman. Improved distribution matching distillation for fast image synthesis. arXiv preprint arXiv:2405.14867, 2024. 3", + "[80] Tianwei Yin, Michael Gharbi, Richard Zhang, Eli Shechtman, Fredo Durand, William T Freeman, and Taesung Park. One-step diffusion with distribution matching distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6613-6623, 2024. 2", + "[81] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, et al. Scaling autoregressive models for content-rich text-to-image generation. Transactions on Machine Learning Research. 8", + "[82] Biao Zhang and Rico Sennrich. Root mean square layer normalization. Advances in Neural Information Processing Systems, 32, 2019. 5", + "[83] Dingkun Zhang, Sijia Li, Chen Chen, Qingsong Xie, and Haonan Lu. Laptop-Diff: Layer Pruning and Normalized Distillation for Compressing Diffusion Models. arXiv preprint arXiv:2404.11098, 2024. 2" + ], + "bbox": [ + 514, + 90, + 903, + 900 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "8007", + "bbox": [ + 482, + 959, + 514, + 970 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[84] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 5", + "[85] Zhixing Zhang, Ligong Han, Arnab Ghosh, Dimitris N Metaxas, and Jian Ren. Sine: Single image editing with text-to-image diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6027-6037, 2023. 2", + "[86] Yang Zhao, Yanwu Xu, Zhisheng Xiao, and Tingbo Hou. Mobilediffusion: Subsecond text-to-image generation on mobile devices. arXiv preprint arXiv:2311.16567, 2023. 2, 3, 4, 5", + "[87] Le Zhuo, Ruoyi Du, Han Xiao, Yangguang Li, Dongyang Liu, Rongjie Huang, Wenze Liu, Lirui Zhao, Fu-Yun Wang, Zhanyu Ma, et al. Lumina-last: Making lumina-t2x stronger and faster with next-dit. arXiv preprint arXiv:2406.18583, 2024. 2, 8" + ], + "bbox": [ + 91, + 90, + 482, + 330 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "8008", + "bbox": [ + 482, + 960, + 514, + 970 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2025/SnapGen_ Taming High-Resolution Text-to-Image Models for Mobile Devices with Efficient Architectures and Training/d6336171-7440-4517-9fbd-fdd52a399434_model.json b/2025/SnapGen_ Taming High-Resolution Text-to-Image Models for Mobile Devices with Efficient Architectures and Training/d6336171-7440-4517-9fbd-fdd52a399434_model.json new file mode 100644 index 0000000000000000000000000000000000000000..64df47e42eaa18721f03b34c960bae891253ce14 --- /dev/null +++ b/2025/SnapGen_ Taming High-Resolution Text-to-Image Models for Mobile Devices with Efficient Architectures and Training/d6336171-7440-4517-9fbd-fdd52a399434_model.json @@ -0,0 +1,2787 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.02 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation." + }, + { + "type": "header", + "bbox": [ + 0.325, + 0.018, + 0.723, + 0.031 + ], + "angle": 0, + "content": "Except for this watermark, it is identical to the accepted version;" + }, + { + "type": "header", + "bbox": [ + 0.294, + 0.032, + 0.755, + 0.046 + ], + "angle": 0, + "content": "the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.11, + 0.131, + 0.888, + 0.177 + ], + "angle": 0, + "content": "SnapGen: Taming High-Resolution Text-to-Image Models for Mobile Devices with Efficient Architectures and Training" + }, + { + "type": "text", + "bbox": [ + 0.179, + 0.203, + 0.817, + 0.277 + ], + "angle": 0, + "content": "Jierun Chen\\(^{1,2,*}\\) Dongting Hu\\(^{1,3,*}\\) Xijie Huang\\(^{1,2,*}\\) Huseyin Coskun\\(^{1}\\) Arpit Sahni\\(^{1}\\) Aarush Gupta\\(^{1}\\) Anujraaj Goyal\\(^{1}\\) Dishani Lahiri\\(^{1}\\) Rajesh Singh\\(^{1}\\) Yerlan Idelbayev\\(^{1}\\) Junli Cao\\(^{1}\\) Yanyu Li\\(^{1}\\) Kwang-Ting Cheng\\(^{2}\\) S.-H. Gary Chan\\(^{2}\\) Mingming Gong\\(^{3,4}\\) Sergey Tulyakov\\(^{1}\\) Anil Kag\\(^{1,\\dagger}\\) Yanwu Xu\\(^{1,\\dagger}\\) Jian Ren\\(^{1,\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.291, + 0.782, + 0.309 + ], + "angle": 0, + "content": "\\(^{1}\\)Snap Inc. \\(^{2}\\)HKUST \\(^{3}\\)The University of Melbourne \\(^{4}\\)MBZUAI" + }, + { + "type": "text", + "bbox": [ + 0.342, + 0.311, + 0.657, + 0.329 + ], + "angle": 0, + "content": "* Equal contribution † Equal advising" + }, + { + "type": "text", + "bbox": [ + 0.287, + 0.33, + 0.706, + 0.347 + ], + "angle": 0, + "content": "Project Page: https://snap-research.github.io/snapgen" + }, + { + "type": "table", + "bbox": [ + 0.125, + 0.359, + 0.862, + 0.408 + ], + "angle": 0, + "content": "
OursPixArt-αLumina-NextSD3-MediumSDXLPlaygroundv2SD3.5-Large
Param0.38B0.6B2B2B2.6B2.6B8.1B
MobileYESNONONONONONO
" + }, + { + "type": "text", + "bbox": [ + 0.153, + 0.408, + 0.792, + 0.42 + ], + "angle": 0, + "content": "\"an old raccoon wearing a top hat and holding an apple, oil painting in the style of van gogh, ...\"" + }, + { + "type": "image", + "bbox": [ + 0.144, + 0.419, + 0.871, + 0.5 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.153, + 0.502, + 0.524, + 0.514 + ], + "angle": 0, + "content": "\"a dolphin in an astronaut suit, Animals, Simple Detail\"" + }, + { + "type": "image", + "bbox": [ + 0.144, + 0.513, + 0.871, + 0.595 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.153, + 0.597, + 0.773, + 0.608 + ], + "angle": 0, + "content": "\"..., young sudanese female, glamour, natural, front view, extreme detailed and texture skin, ...\"" + }, + { + "type": "image", + "bbox": [ + 0.144, + 0.608, + 0.871, + 0.689 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.153, + 0.691, + 0.812, + 0.703 + ], + "angle": 0, + "content": "... Llama wearing sunglasses standing on the deck of a spaceship with the Earth in the background, ..." + }, + { + "type": "image", + "bbox": [ + 0.144, + 0.702, + 0.871, + 0.784 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.153, + 0.786, + 0.792, + 0.797 + ], + "angle": 0, + "content": "\"... cute monster ... eating a sashimi while sitting at a breakfast table full of fruits and insects\"" + }, + { + "type": "image", + "bbox": [ + 0.144, + 0.797, + 0.871, + 0.879 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.89, + 0.908, + 0.933 + ], + "angle": 0, + "content": "Figure 1. Comparison of various text-to-image models in terms of model size, mobile device compatibility, and visual output quality. Our model, with only \\(379M\\) parameters, demonstrates competitive visual quality while being mobile-compatible. Input text prompts are shown above each image grid; all images are generated at \\(1024^2\\) resolution—zoom in for details. More examples are shown in webpage." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.96, + 0.516, + 0.972 + ], + "angle": 0, + "content": "7997" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.249, + 0.091, + 0.327, + 0.106 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.122, + 0.486, + 0.471 + ], + "angle": 0, + "content": "Existing text-to-image (T2I) diffusion models face several limitations, including large model sizes, slow runtime, and low-quality generation on mobile devices. This paper aims to address all of these challenges by developing an extremely small and fast T2I model that generates high-resolution and high-quality images on mobile platforms. We propose several techniques to achieve this goal. First, we systematically examine the design choices of the network architecture to reduce model parameters and latency, while ensuring high-quality generation. Second, to further improve generation quality, we employ cross-architecture knowledge distillation from a much larger model, using a multi-level approach to guide the training of our model from scratch. Third, we enable a few-step generation by integrating adversarial guidance with knowledge distillation. For the first time, our model SnapGen, demonstrates the generation of \\(1024^2\\) px images on a mobile device around 1.4 seconds. On ImageNet-1K, our model, with only 372M parameters, achieves an FID of 2.06 for \\(256^2\\) px generation. On T2I benchmarks (i.e., GenEval and DPG-Bench), our model with merely 379M parameters, surpasses large-scale models with billions of parameters at a significantly smaller size (e.g., \\(7 \\times\\) smaller than SDXL, \\(14 \\times\\) smaller than IF-XL)." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.498, + 0.222, + 0.513 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.524, + 0.483, + 0.629 + ], + "angle": 0, + "content": "Large-scale text-to-image (T2I) diffusion models [12, 13, 15, 51, 53, 57-59] have achieved remarkable success in content generation, powering numerous applications like image editing [48, 64, 72, 85] and video creation [50, 54, 78]. However, T2I models often come with substantial model sizes and slow runtime, and deploying them on the cloud raises concerns related to data security and high costs [65]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.63, + 0.483, + 0.75 + ], + "angle": 0, + "content": "To address these challenges, there is huge growing interest in developing smaller and faster T2I models through techniques like model compression (e.g., pruning and quantization) [40, 67, 86], step reduction by distillation [77, 80], and efficient attention mechanisms that mitigate the quadratic complexity [46, 74]. Nevertheless, current works still encounter limitations, e.g., low-resolution generation on mobile devices, that constrain their broader application." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.751, + 0.483, + 0.84 + ], + "angle": 0, + "content": "Most importantly, a critical question remains unexplored: how can we train a T2I model from scratch to generate high-quality, high-resolution images on mobile? Such a model would offer substantial advantages in speed, compactness, cost-effectiveness, and secure deployment. To build this model, we introduce several innovations:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.84, + 0.483, + 0.901 + ], + "angle": 0, + "content": "- Efficient Network Architectures: We conduct an in-depth examination of network architectures, including the denoising UNet and Autoencoder (AE), to obtain optimal trade-off between resource usage and performance. Un" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.092, + 0.905, + 0.168 + ], + "angle": 0, + "content": "like prior works that optimize and compress pre-trained diffusion models [10, 32, 83], we directly focus on macro- and micro-level design choices to achieve a novel architecture that greatly reduces model size and computational complexity, while preserving high-quality generation." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.168, + 0.905, + 0.363 + ], + "angle": 0, + "content": "- Improved Training Techniques: We introduce several improvements to train a compact T2I model from scratch. We utilize flow matching [44, 47] as objective, aligning with larger models like SD3 [18] and SD3.5 [3]. This design enables effective knowledge and step distillation, transferring rich representations from large-scale diffusion models to our much smaller one. Further, we propose a multi-level knowledge distillation with a timestep-aware scaling that combines multiple training objectives. Instead of weighting objectives through a linear combination as in prior works [32, 46], we consider target prediction difficulty (i.e., student-teacher difference) across various timesteps in flow matching." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.364, + 0.905, + 0.438 + ], + "angle": 0, + "content": "- Advanced Step Distillation: We perform step distillation on our model by combining the adversarial training along with the knowledge distillation using a few-step teacher model (i.e., SD3.5-Large-Turbo [5]), enabling ultra-fast high-quality generation with only 4 or 8 steps." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.168, + 0.905, + 0.438 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.44, + 0.905, + 0.469 + ], + "angle": 0, + "content": "We demonstrate the superior advantages of our approach and model through extensive experiments:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.47, + 0.905, + 0.544 + ], + "angle": 0, + "content": "- On ImageNet-1K [16] class-conditional image generation task, our model achieves the FID comparable to existing works with significantly reduced model size and computation, i.e., half the model size and one-third of the compute resources compared with SiT-XL[49], as in Tab. 1." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.545, + 0.905, + 0.62 + ], + "angle": 0, + "content": "- For large-scale T2I generation, our UNet model, with only 379M parameters, demonstrates superior generation quality compared to billion-parameter models [42, 53, 87], e.g., improved metrics on benchmark datasets (Tab. 3) and human evaluation (Fig. 8)." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.621, + 0.905, + 0.68 + ], + "angle": 0, + "content": "- Our compressed decoder, trained from scratch, has competitive reconstruction quality compared to commonly used models [18, 53], with more than \\(36 \\times\\) smaller size, enabling the mobile deployment." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.681, + 0.905, + 0.726 + ], + "angle": 0, + "content": "- Notably, for the first time, we show a T2I model achieving high-resolution generation (e.g., \\(1024^2\\) px) on mobile devices (e.g., iPhone 16 Pro-Max) around 1.4 seconds." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.47, + 0.905, + 0.726 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.74, + 0.655, + 0.755 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.907, + 0.903 + ], + "angle": 0, + "content": "High-Resolution Text-to-Image Models have emerged with advanced architectures and multi-stage approaches designed to enhance visual fidelity and user customization. SDXL [53] is a pioneering work in this field, employing a refined cascading approach with UNet backbone to generate high-detailed images, resulting in photorealistic outputs that maintain sharpness and clarity. The following studies explore different techniques like more advanced text encoders, better image refinement, or improved dataset preparation," + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.961, + 0.516, + 0.972 + ], + "angle": 0, + "content": "7998" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.198 + ], + "angle": 0, + "content": "to obtain better text-image alignment or higher-quality generation [6, 7, 15, 20, 30, 36-38, 41, 45, 48, 69]. However, most of these models contain billions of parameters, making them extremely slow, and not being able to run on resource-contained hardware like mobile devices. In this work, we aim to build a small and fast model that can perform high-resolution generation even on mobile platforms." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.201, + 0.482, + 0.368 + ], + "angle": 0, + "content": "Efficient Diffusion Models address the challenges of bulky model size and long runtime. There have been efforts exploring the architecture optimization to remove redundancy from large models, demonstrating the on-device generation within seconds [11, 40, 67, 86]. However, these models are constrained to low-resolution output, i.e., \\(512^{2}\\) px. To enable efficient high-resolution generation, SANA [74] and LinFusion [46] incorporate linear attention [9, 14, 31] to achieve 1K generation on laptop GPUs. In contrast, we target a broader range of platforms, supporting high-resolution generation (e.g., 1K) directly on mobile devices." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.372, + 0.482, + 0.524 + ], + "angle": 0, + "content": "Knowledge Distillation in Diffusion Models. In the context of diffusion models, previous works focus on distilling large, high-capacity teacher models into more compact, efficient student models within a homogeneous architecture [32, 46]. They reduce the complexity of the model by removing certain components like attention [70] or residual blocks [23], while maintaining the architectural structure. However, our approach diverges from this trend by utilizing a heterogeneous architecture for more aggressively efficient yet challenging distillation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.527, + 0.482, + 0.679 + ], + "angle": 0, + "content": "Adversarial Step Distillation uses the techniques from adversarial training [22] to reduce the number of diffusion steps, while maintaining high image quality [60, 61]. For example, UFOGen [77] employs a diffusion-GAN formulation [71, 73, 76] to significantly reduce inference time while maintaining competitive performance. DMD2 [79] builds on prior distillation methods by using distribution matching with adversarial loss. Different from exiting works, we conduct step-distillation on a very compact model and train the model along with the knowledge distillation." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.705, + 0.182, + 0.721 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.483, + 0.901 + ], + "angle": 0, + "content": "In this section, we present how to craft and train a highly efficient T2I model for high-resolution generation. Specifically, starting from the architecture designed in the latent diffusion model [58], we optimize both denoising backbone (Sec. 3.1, Fig. 2, and Fig. 3) and autoencoder (Sec. 3.2 and Fig. 4) to make them compact and fast, even on mobile devices. We then propose the improved training recipe and knowledge distillation (Sec. 3.3 and Fig. 5), empowering a high-performance T2I model. Lastly, we introduce our step distillation to significantly reduce the number of denoising steps for a faster T2I model (Sec. 3.4 and Fig. 7)." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.091, + 0.761, + 0.105 + ], + "angle": 0, + "content": "3.1. Efficient UNet Architecture" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.114, + 0.905, + 0.251 + ], + "angle": 0, + "content": "Here we describe the design choices for the denoising UNet. Baseline Architecture. We choose UNet from SDXL [53] as the baseline (Fig. 2(a)) for our backbone since it has superior efficiency and faster convergence [39] than pure transformer-based models [12, 13]. We adjust the UNet into a thinner and shorter model (i.e., reducing the number of transformer blocks from [0, 2, 10] in three stages to [0, 2, 4] and their channel dimensions from [320, 640, 1280] to [256, 512, 896]), and iterate design choices on top of it." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.252, + 0.905, + 0.462 + ], + "angle": 0, + "content": "Evaluation Metrics. We train models on ImageNet-1K [16] under class-conditional generation for 120 epochs, unless specified otherwise, and report the FID score [25] for \\(256^{2}\\) px generation. Similarly to existing work [30], we inject the class conditions through a text template \"a photo of \". We then encode it with a light text encoder to align the pipeline for the T2I generation. We also calculate the number of parameters for different models, floating point operations (FLOPs) (measured on a latent size of \\(128 \\times 128\\), equivalent to a \\(1024 \\times 1024\\) image after decoding), and the runtime on mobile device (tested on iPhone 15 Pro). Detailed training and evaluation settings can be found in Supplemental Materials. In the following, we introduce the key architectural changes that improve the model." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.463, + 0.905, + 0.629 + ], + "angle": 0, + "content": "Remove Self-Attention from High-Resolution Stages. Self-attention (SA) layer is restricted by its quadratic computational complexity, incurring a heavy computational cost and high memory consumption for high-resolution input. As such, we keep the SA layer only in the lowest-resolution stage while removing it from other higher-resolution stages, i.e., Fig. 2 (b). This leads to \\(17\\%\\) fewer FLOPs and \\(24\\%\\) latency reduction, as shown in Fig. 3. Interestingly, we even observe a performance improvement, i.e., FID decreased to 3.12 from 3.76. We hypothesize that models with SA in high-resolution stages converge more slowly." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.63, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Replace Conv with Expanded Separable Conv. Regular convolution (Conv) is redundant in both parameters and computation. To address this, we replace all Conv layers with the separable convolution (SepConv) [28], composed of a depthwise convolution (DW) followed by a pointwise convolution (PW), as shown in Fig. 2 (c). This replacement reduces parameters by \\(24\\%\\) and latency by \\(62\\%\\), but also leads to a performance drop (FID increases from 3.12 to 3.38). To address the issue, we expand the intermediate channels. Specifically, the number of channels after the first PW layer is increased with an expansion ratio, and reduced back to the original number after the second PW layer. The expansion ratio is set to 2 to balance the trade-off between performance, latency, and model parameters. Such a design aligns our residual block with the recently proposed Universal Inverted Bottleneck (UIB) block [55]. As a result, our model achieves \\(15\\%\\) fewer parameters, \\(27\\%\\) less computation, and \\(2.4 \\times\\) speedup, while obtaining a lower FID." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.961, + 0.516, + 0.972 + ], + "angle": 0, + "content": "7999" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.09, + 0.446, + 0.409 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.452, + 0.098, + 0.585, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.608, + 0.097, + 0.732, + 0.22 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.754, + 0.096, + 0.895, + 0.221 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.451, + 0.234, + 0.58, + 0.407 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.614, + 0.23, + 0.89, + 0.407 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.415, + 0.907, + 0.445 + ], + "angle": 0, + "content": "Figure 2. Efficient UNet. Starting from a thinner and shorter version of the UNet from SDXL (as in (a)), we explore a series of architectural changes, i.e., (b)-(f), to develop a smaller and faster model while retaining high-quality generation performance, as evaluated in Fig. 3." + }, + { + "type": "image", + "bbox": [ + 0.119, + 0.448, + 0.885, + 0.596 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.602, + 0.908, + 0.66 + ], + "angle": 0, + "content": "Figure 3. Comparisons of Performance and Efficiency for various Design Choices of Efficient UNet. The generation quality is evaluated using FID calculated on ImageNet-1K for \\(256^{2}\\) px generation. The efficiency metrics include model parameters, latency, and FLOPs. FLOPs and latency (on iPhone 15 Pro) are measured with a \\(128 \\times 128\\) latent, equivalent to a \\(1024 \\times 1024\\) decoded image, for one forward pass. We show the architecture enhancements that improve any of the metrics without hurting others." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.674, + 0.483, + 0.81 + ], + "angle": 0, + "content": "Trim FFN Layers. For the layers in the feed-forward network (FFN), the hidden channel expansion ratio is set to 4 by default and further doubled by using the gated unit. This substantially inflates the model parameters, computation, and memory usage. Following MobileDiffusion [86], we examine the efficacy of simply reducing the expansion ratio, as shown in Fig. 2 (d). We show that reducing the expansion ratio to 3 preserves comparable FID performance while reducing both parameters and FLOPs by \\(12\\%\\)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.81, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Replace MHSA with MQA. Multi-Head Self-Attention (MHSA) requires multiple sets of keys and values for each attention head. In contrast, Multi-Query Attention (MQA) [63] is more efficient by sharing a single set of keys and values across all heads. Replacing MHSA with MQA reduces parameters by \\(16\\%\\) and latency by \\(9\\%\\), with mini" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.674, + 0.907, + 0.72 + ], + "angle": 0, + "content": "mal impact on performance. Interestingly, the \\(9\\%\\) saving in latency exceeds the \\(6\\%\\) decrease in FLOPs, as the reduced memory access enables higher computational intensity." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Inject Conditions to the First Stage. Cross-attention (CA) blends the conditional information (e.g., textural description) along with the spatial features to generate images that align with the condition. However, the UNet of SDXL only applies CA in transformer blocks starting from the second stage, resulting in the missing conditional guidance for the first stage. In response, we propose to introduce the conditional embeddings from the very first stage, as in Fig. 2(e). Specifically, we replace the residual blocks with transformer blocks that include CA and FFN while without SA layers. This adjustment makes the model smaller, faster, and more efficient while improving FID." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.96, + 0.516, + 0.972 + ], + "angle": 0, + "content": "8000" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.091, + 0.485, + 0.302 + ], + "angle": 0, + "content": "Employ QK RMSNorm and RoPE Positional Embeddings. We extend two advanced techniques developed originally for language models, Query-Key (QK) Normalization [24] with RMSNorm [82] and Rotary Position Embedding (RoPE) [66], to enhance the model (Fig. 2 (f)). RMSNorm, applied after the Query-Key projection in the attention mechanism, reduces the risk of softmax saturation without sacrificing model expressiveness while stabilizing training for faster convergence. In addition, we adapt RoPE from one dimension to two dimensions for better supporting higher resolution since it significantly mitigates artifacts like repeated objects. Together, RMSNorm and RoPE introduce negligible computational and parameter overhead, while offering measurable gains in FID performance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.303, + 0.484, + 0.456 + ], + "angle": 0, + "content": "Discussion. The above optimization results in an efficient and powerful diffusion backbone capable of generating high-resolution images on mobile devices. Before proceeding with large-scale T2I training, we compare the capacity of our model against existing works on ImageNet-1K. We train the model for 1,000 epochs by following the setting from prior work [58]. We evaluate the model using varied CFG [26, 34] across different inference timesteps. As shown in Tab. 1, our efficient UNet achieves comparable FID to SiT-XL [49], while being almost \\(45\\%\\) smaller." + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.461, + 0.484, + 0.489 + ], + "angle": 0, + "content": "Table 1. Class-conditional image generation on ImageNet \\( {256} \\times {256} \\) with CFG. FLOPs are calculated for one forward pass." + }, + { + "type": "table", + "bbox": [ + 0.14, + 0.491, + 0.435, + 0.599 + ], + "angle": 0, + "content": "
ModelParam (M)FLOPs (G)FID↓
LDM-4 [58]4001043.60
UViT-L [8]287773.40
UViT-H [8]5011332.29
DiT-XL [52]6751192.27
SiT-XL [49]6751192.06
Ours372382.06
" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.617, + 0.303, + 0.633 + ], + "angle": 0, + "content": "3.2. Tiny and Fast Decoder" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.64, + 0.484, + 0.701 + ], + "angle": 0, + "content": "Besides the denoising model, the decoder also takes a significant ratio of total runtime, especially for on-device deployment [40, 86]. Here we introduce a new architecture of decoder (Fig. 4) for efficient high-resolution generation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.702, + 0.484, + 0.884 + ], + "angle": 0, + "content": "Baseline Decoder. We use the autoencoder (AE) from SD3 [18] as our baseline model (i.e., the same encoder from SD3 AE), due to its superior reconstruction quality. The AE maps an image \\( X \\in \\mathbb{R}^{H \\times W \\times 3} \\) into a lower-dimensional latent \\( x \\in \\mathbb{R}^{\\frac{H}{f} \\times \\frac{W}{f} \\times c} \\) (\\( f, c \\) as 8, 16 in SD3). The encoded latent \\( x \\) is then decoded back to an image through a decoder. For high-resolution generation, we observe that the decoder in SD3 is very slow on mobile devices. Specifically, it encounters out-of-memory (OOM) errors when generating a \\( 1024^2 \\) px image on both the ANE processor of the iPhone 15 Pro and the mobile GPU (Tab. 2). To overcome the latency issue, we propose a much smaller and faster decoder." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.886, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Efficient Decoder. We conduct a series of experiments to" + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.09, + 0.902, + 0.198 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.2, + 0.901, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.301, + 0.906, + 0.33 + ], + "angle": 0, + "content": "Figure 4. Comparisons of Decoder Architecture between (a) SDXL/SD3 decoder and (b) our tiny decoder." + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.336, + 0.906, + 0.406 + ], + "angle": 0, + "content": "Table 2. Performance Comparison of Decoder. PSNR is calculated on COCO 2017 validation set [43]. FLOPs and latency (on iPhone 15 Pro) are measured for decoding a \\(128 \\times 128\\) latent into a \\(1024 \\times 1024\\) image. The decoder from SDXL and SD3 fail to run on the neural engine of mobile, resulting in a huge runtime." + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.413, + 0.905, + 0.475 + ], + "angle": 0, + "content": "
DecoderChPSNRParam (M)FLOPs (G)Latency (ms) on ANELatency (ms) on GPU
SDXL [53]424.8949.494970OOM9469
SD3 [18]1627.9249.554970OOMOOM
Ours1627.851.38224174-
" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.492, + 0.905, + 0.521 + ], + "angle": 0, + "content": "decide the efficient decoder with the following key changes compared with the baseline architecture:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.523, + 0.905, + 0.553 + ], + "angle": 0, + "content": "1. We remove attention layers to greatly reduce peak memory without a noticeable impact on decoding quality." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.553, + 0.905, + 0.598 + ], + "angle": 0, + "content": "2. We keep a minimal amount of GroupNorm (GN) to find a trade-off between latency and performance (i.e., mitigating the color shifting)." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.598, + 0.905, + 0.628 + ], + "angle": 0, + "content": "3. We make the decoder thinner (i.e., fewer channels or narrower width) and replace Conv with SepConvs." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.628, + 0.897, + 0.643 + ], + "angle": 0, + "content": "4. We use fewer residual blocks in high-resolution stages." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.643, + 0.905, + 0.673 + ], + "angle": 0, + "content": "5. We remove the Conv shortcut in residual blocks and use the upsampling layer for channel transition." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.523, + 0.905, + 0.673 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.907, + 0.809 + ], + "angle": 0, + "content": "Training of the Decoder. We train our decoder with the mean squared error (MSE) loss, lpips loss [84], adversarial loss [22], and discard the KL term [33] as the encoder is fixed. The decoder is trained on \\( 256^2 \\) image patches with a batch size of 256 and for 1M iterations. As in Tab. 2, our tiny decoder achieves a competitive PSNR score for reconstruction, while being \\( 35.9 \\times \\) smaller and \\( 54.4 \\times \\) faster for high-resolution generation on mobile devices compared to conventional ones (e.g., the decoder from SDXL and SD3)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Discussion of Total On-Device Latency. We finally measure T2I model latency for a \\(1024^{2}\\) px generation on iPhone 16 Pro-Max. The decoder takes 119ms, and the per-step latency for the UNet is 274ms. This results in a \\(1.2\\sim 2.3s\\) runtime for 4 to 8 step generation. Note that text encoder runtime is negligible compared to other components [40]." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.96, + 0.514, + 0.972 + ], + "angle": 0, + "content": "8001" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.09, + 0.353, + 0.22 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.231, + 0.356, + 0.272 + ], + "angle": 0, + "content": "Figure 5. Overview of Multi-level Knowledge Distillation, where we perform output distillation and feature distillation." + }, + { + "type": "image", + "bbox": [ + 0.367, + 0.099, + 0.633, + 0.22 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.366, + 0.232, + 0.63, + 0.26 + ], + "angle": 0, + "content": "Figure 6. Mean loss magnitude for task loss \\(\\mathcal{L}_{\\mathrm{task}}\\) and output distillation loss \\(\\mathcal{L}_{\\mathrm{kd}}\\)." + }, + { + "type": "image", + "bbox": [ + 0.645, + 0.091, + 0.906, + 0.22 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.641, + 0.23, + 0.906, + 0.272 + ], + "angle": 0, + "content": "Figure 7. Overview of Adversarial Step Distillation. Output distillation and distribution-matching distillation are performed." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.285, + 0.468, + 0.302 + ], + "angle": 0, + "content": "3.3. Training Recipe and Multi-Level Distillation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.309, + 0.482, + 0.338 + ], + "angle": 0, + "content": "To improve the generation quality of our efficient diffusion model, we propose a series of training techniques." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.338, + 0.483, + 0.392 + ], + "angle": 0, + "content": "Flow-based Training and Inference. Rectified Flows (RFs) [44, 47] define the forward process as straight paths connecting the data distribution to a standard normal distribution, i.e.," + }, + { + "type": "equation", + "bbox": [ + 0.212, + 0.394, + 0.482, + 0.408 + ], + "angle": 0, + "content": "\\[\nx _ {t} = \\left(1 - \\sigma_ {t}\\right) x _ {0} + \\sigma_ {t} \\epsilon , \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.412, + 0.484, + 0.468 + ], + "angle": 0, + "content": "where \\(x_0\\) is the clean (latent) image, \\(t\\) is the timestep, \\(\\sigma_t\\) is a timestep-dependent factor, and \\(\\epsilon\\) is random noise sampled from \\(\\mathcal{N}(0,I)\\). The denoising UNet is formulated to predict a velocity field with the objective as" + }, + { + "type": "equation", + "bbox": [ + 0.141, + 0.478, + 0.483, + 0.502 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {t a s k}} = \\mathbb {E} _ {\\epsilon \\sim \\mathcal {N} (0, I), t} \\left[ | | (\\epsilon - x _ {0}) - v _ {\\theta} (x _ {t}, t) | | _ {2} ^ {2} \\right], \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.51, + 0.484, + 0.595 + ], + "angle": 0, + "content": "where \\( v_{\\theta}(x_t, t) \\) is the predicted velocity from UNet parameterized by \\( \\theta \\). To further enhance training stability, we apply logit-normal sampling [18] for the timestep during training, which assigns more samples to the intermediate steps. In the inference stage, we use the Flow-Euler sampler [19], which predicts the next sample based on the velocity, i.e.," + }, + { + "type": "equation", + "bbox": [ + 0.172, + 0.605, + 0.482, + 0.62 + ], + "angle": 0, + "content": "\\[\nx _ {t - 1} = x _ {t} + \\left(\\sigma_ {t - 1} - \\sigma_ {t}\\right) \\cdot v _ {\\theta} \\left(x _ {t}, t\\right). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.629, + 0.483, + 0.689 + ], + "angle": 0, + "content": "To achieve a lower signal-to-noise ratio on high-resolution (i.e., \\(1024^2\\) px) images, we apply a timestep shift similar to SD3 [18] to adjust the scheduling factor \\(\\sigma_t\\) during both training and inference." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.69, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Multi-Level Knowledge Distillation. To improve the generation quality for compact models, one common practice for previous works is applying knowledge distillation to mimic the prediction of scaled-up teacher models [32]. Benefiting from the aligned flow matching objective and (AE) latent space, powerful SD3.5-Large model [4] can be used as the teacher for output distillation. However, we still face challenges due to 1) the heterogeneous architecture between the U-Net and DiT, 2) the scale difference between the distillation loss and task loss, and 3) varying prediction difficulty across different timesteps. To tackle these, we propose a novel multi-level distillation loss and apply timestep-aware scaling to stabilize and accelerate the convergence of distillation. The overview of our knowledge" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.286, + 0.905, + 0.315 + ], + "angle": 0, + "content": "distillation scheme is shown in Fig. 5 and detailed techniques are elaborated as follows." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.316, + 0.905, + 0.37 + ], + "angle": 0, + "content": "Aside from the task loss defined in Eq. 2, the major objective for knowledge distillation is to supervise our model \\(\\theta\\) directly with the output of teacher model \\(\\theta_{T}\\), which can be indicated as" + }, + { + "type": "equation", + "bbox": [ + 0.594, + 0.376, + 0.905, + 0.401 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {k d}} = \\mathbb {E} \\left[ \\left| \\left| v _ {\\theta_ {T}} \\left(x _ {t}, t\\right) - v _ {\\theta} \\left(x _ {t}, t\\right) \\right| \\right| _ {2} ^ {2} \\right]. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.407, + 0.906, + 0.463 + ], + "angle": 0, + "content": "Given the capacity gap between the teacher and our model, applying output-level supervision alone leads to instability and slow convergence. Therefore, we further incorporate a cross-architecture feature-level distillation loss as" + }, + { + "type": "equation", + "bbox": [ + 0.555, + 0.468, + 0.905, + 0.501 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {f e a t k d}} = \\mathbb {E} \\left[ \\sum_ {(l _ {T}, l)} | | f _ {\\theta_ {T}} ^ {l _ {T}} (x _ {t}, t) - \\psi \\left(f _ {\\theta} ^ {l} \\left(x _ {t}, t\\right)\\right) | | _ {2} ^ {2} \\right], \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.51, + 0.906, + 0.692 + ], + "angle": 0, + "content": "where \\( f_{\\theta_T}^{l_T}(\\cdot) \\) and \\( f^l (\\cdot) \\) indicate the feature output from the \\( l_{T} \\)-th layer and \\( l \\)-th layer in teacher model and student model, respectively. Different from previous work [32, 46], we consider cross-architecture distillation from a DiT to UNet. Since the richest information in transformers sits around the last layer, we set the distillation target to this layer in both models, and use a lightweight trainable projector \\( \\psi (\\cdot) \\) with only two Conv layers to map the student feature to match the dimension of teacher feature. The proposed feature-level distillation loss provides additional supervision to the student model, leading to faster alignment to the generation quality of the teacher model." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.692, + 0.905, + 0.762 + ], + "angle": 0, + "content": "Timestep-Aware Scaling. Weighting multiple objectives has been a major challenge in knowledge distillation, especially in diffusion models. The overall training objectives from previous works [32, 46, 67] are simple linear combination of multiple loss term, i.e.," + }, + { + "type": "equation", + "bbox": [ + 0.605, + 0.771, + 0.905, + 0.786 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {\\text {t a s k}} + \\lambda_ {1} \\mathcal {L} _ {\\mathrm {k d}} + \\lambda_ {2} \\mathcal {L} _ {\\text {f e a t k d}}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.906, + 0.901 + ], + "angle": 0, + "content": "where the weighting coefficient \\(\\lambda_{1}\\) and \\(\\lambda_{2}\\) are empirically set to constant. However, this baseline setting fails to consider the prediction difficulty in various time steps. We investigate the distribution of empirical risk magnitude of \\(\\mathcal{L}_{\\mathrm{task}}\\) and \\(\\mathcal{L}_{\\mathrm{kd}}\\) across different timestep \\(t\\) during model training. Fig. 6 illustrates that, in intermediate steps, prediction difficulty are lower compared to \\(t\\) closer to 0 or 1." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.96, + 0.516, + 0.972 + ], + "angle": 0, + "content": "8002" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.152 + ], + "angle": 0, + "content": "Building on this important observation, we propose a timestep-aware scaling of the objective to close the gap in loss magnitude across different values of \\( t \\) and to account for prediction difficulties at each timestep, as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.09, + 0.163, + 0.484, + 0.209 + ], + "angle": 0, + "content": "\\[\n\\mathcal {S} \\left(\\mathcal {L} _ {\\text {t a s k}}, \\mathcal {L} _ {\\mathrm {k d}}\\right) = \\mathbb {E} _ {t} \\left[ \\lambda (t) \\cdot \\mathcal {L} _ {\\text {t a s k}} ^ {t} + (1 - \\lambda (t)) \\frac {\\left| \\mathcal {L} _ {\\text {t a s k}} ^ {t} \\right|}{\\left| \\mathcal {L} _ {\\mathrm {k d}} ^ {t} \\right|} \\cdot \\mathcal {L} _ {\\mathrm {k d}} ^ {t} \\right], \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.21, + 0.484, + 0.362 + ], + "angle": 0, + "content": "where \\(\\lambda (t)\\) is the normalized standard (location 0, scale 1) logit-norm density function and \\(|\\cdot |\\) indicates the magnitude. In \\(S\\), we first ensure the same scale between task loss and distillation loss across different \\(t\\), then apply more teacher supervision where prediction difficulty is higher (i.e., \\(t\\) closer to 0 or 1), and more real data supervision where prediction difficulty is lower (i.e., intermediate timesteps). The proposed scheme considers the variation of timestep \\(t\\) and helps accelerate the distillation training. The final multi-level distillation objective \\(\\mathcal{L}_{\\mathrm{md}}\\) is defined as" + }, + { + "type": "equation", + "bbox": [ + 0.143, + 0.373, + 0.483, + 0.39 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {m d}} = \\mathcal {S} \\left(\\mathcal {L} _ {\\text {t a s k}}, \\mathcal {L} _ {\\mathrm {k d}}\\right) + \\mathcal {S} \\left(\\mathcal {L} _ {\\text {t a s k}}, \\mathcal {L} _ {\\text {f e a t k d}}\\right). \\tag {8}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.401, + 0.251, + 0.417 + ], + "angle": 0, + "content": "3.4. Step Distillation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.424, + 0.484, + 0.515 + ], + "angle": 0, + "content": "We take one step further to enhance the sampling efficiency of our model following a distribution-matching-based step distillation scheme. Following Latent Adversarial Diffusion Distillation (LADD) [60], we use a diffusion-GAN hybrid structure to distill our model into fewer steps with the optimization objective as" + }, + { + "type": "equation", + "bbox": [ + 0.136, + 0.526, + 0.483, + 0.584 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\min _ {D _ {\\theta_ {T}}} \\max _ {G _ {\\theta}} \\mathbb {E} \\left[ \\left[ \\log \\left(D _ {\\theta_ {T}} \\left(x _ {t - 1}, t\\right)\\right) \\right] \\right. \\tag {9} \\\\ \\left. + \\left[ \\log \\left(1 - D _ {\\theta_ {T}} \\left(x _ {t - 1} ^ {\\prime}, t\\right)\\right) \\right] - \\mathcal {S} \\left(\\mathcal {L} _ {\\mathrm {t a s k}}, \\mathcal {L} _ {\\mathrm {k d}}\\right) \\right], \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.594, + 0.484, + 0.868 + ], + "angle": 0, + "content": "where \\(D_{\\theta_T}\\) is the discriminator model partially initialized with pretrained fewer-step teacher model \\(\\theta_T\\) (SD3.5-Large-Turbo [5]). The large-scale teacher model are only used as the feature extractor and are frozen during distillation. We only train a few linear layers in the discriminator after feature extraction. We sample \\(x_{t-1} \\sim q(x_{t-1}|x_0)\\) and \\(x_{t-1}' \\sim q(x_{t-1}|x_0')\\), where \\(x_0'\\) is the prediction of our denoising generator \\(G_\\theta(x_t, t)\\) as our student model, and \\(q(x)\\) is the forward process of diffusion model defined in Eq. 1. The objective consists of an adversarial loss to match noisy samples at time step \\(t-1\\) and the output-level distillation loss \\(S(\\mathcal{L}_{\\mathrm{task}}, \\mathcal{L}_{\\mathrm{kd}})\\) after applying timestep-aware scaling. The proposed step distillation, visualized in Fig. 7, can be interpreted as training a diffusion model with adversarial refinement and knowledge distillation, where teacher guidance serves as an additional inductive bias. This advanced step distillation empowers our compact model for high-quality generation, with only a few denoising steps." + }, + { + "type": "page_footnote", + "bbox": [ + 0.09, + 0.875, + 0.483, + 0.901 + ], + "angle": 0, + "content": "For simplicity, we use the \\(x_0\\) prediction in our derivation, and \\(v\\) prediction of rectified flow-based training would not break the formulation." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.091, + 0.646, + 0.108 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.115, + 0.907, + 0.297 + ], + "angle": 0, + "content": "Model Details. Our T2I pipeline consists of the efficient UNet (Sec. 3.1) and the efficient encoder-decoder model (Sec. 3.2). To obtain text embeddings from the input prompt, we leverage multiple text encoders, namely lightweight CLIP-L [56], CLIP-G [56], and the large Gemma-2-2b language model [68]. We follow SD3 [18] strategy to combine these three text-encoders into a unified rich textual embedding. To enable classifier-free guidance [27], we employ these embeddings with an individual drop-out probability such that we can use an arbitrary subset of the encoders during inference. This allows us to deploy one or more encoders based on the resource constraints." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.297, + 0.908, + 0.538 + ], + "angle": 0, + "content": "Training Recipe. Similar as prior work [30], we use a multi-stage strategy to train our UNet model from scratch. First, we pre-train the model using the ImageNet-1K [16] at 256 resolution as described in Sec. 3.1. Second, we fine-tune this model in a progressive manner from \\(256 \\rightarrow 512 \\rightarrow 1024\\) resolutions. Third, we employ knowledge distillation with our timestep-aware scaling (Sec. 3.3) to improve the finer details in our models using a much larger teacher model (SD3.5-Large [4]) and all three text-encoders. Before KD, we initialize the student model by pre-training on ImageNet and T2I datasets, a cost-effective approach that helps the model learn pixel distributions. Finally, we obtain a few-step model through step distillation using SD3.5-Large-Turbo [5] model as the teacher. We optimize the rectified-flow [18] objective using AdamW optimizer [17] to train our UNet backbone." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.538, + 0.907, + 0.6 + ], + "angle": 0, + "content": "Hyper-parameters. We sample the timesteps using the logit-normal distribution with \\((0,1)\\) as the location and scale parameters. We use a time shift of 3 for both training and inference. See the Supplemental Materials for details." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.606, + 0.636, + 0.621 + ], + "angle": 0, + "content": "4.1. Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.629, + 0.906, + 0.78 + ], + "angle": 0, + "content": "Quantitative Benchmarks. We use GenEval [21] and DPG-Bench [29] benchmarks to evaluate the text-to-image alignment of our model on short and long prompts, respectively. We report CLIP score on a 6K subset of MS-COCO validation data [43]. In addition, to measure the aesthetic quality of our model, we compute the Image Reward [75] score on selected PixArt prompts [12]. Tab. 3 lists our performance alongside existing state-of-the-art T2I baselines. We provide additional details in Supplemental Materials. We highlight the salient observations below:" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.781, + 0.906, + 0.825 + ], + "angle": 0, + "content": "- Our 0.38B parameter model achieves even better performance than significantly larger models such as SDXL (2.6B), Playground (2.6B), and IF-XL (5.5B)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.826, + 0.905, + 0.869 + ], + "angle": 0, + "content": "- KD non-trivially improves the prompt following ability of the base model as illustrated by an absolute five-point increase in DPG-Bench and GenEval scores." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.905, + 0.902 + ], + "angle": 0, + "content": "- In terms of aesthetic performance, our model has similar Image Reward scores as Playground models [37, 38]." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.781, + 0.906, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.96, + 0.516, + 0.972 + ], + "angle": 0, + "content": "8003" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.118, + 0.089, + 0.37, + 0.202 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.373, + 0.088, + 0.625, + 0.203 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.629, + 0.089, + 0.882, + 0.203 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.208, + 0.908, + 0.251 + ], + "angle": 0, + "content": "Figure 8. Human Evaluation. We conduct a user study to compare images generated by our model against baselines on three attributes: aesthetic quality, text-image alignment, and realistic generations. Our model surpasses the quality of SDXL and SD3 models, while performing competitively against the teacher SD3.5-Large model." + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.265, + 0.484, + 0.363 + ], + "angle": 0, + "content": "Table 3. Evaluation on Quantitative Benchmarks. We list the scores on GenEval, DPG-Bench, CLIP score on COCO, and Image Reward on aesthetic prompts. We report the parameters for the UNet/DiT backbone in the Param column. Throughput (samples/s) is measured on a single 80GB A100 GPU using the largest batch size supported for each model in a practical scenario to generate \\(1024^2\\) px images. Here our sampling step is set to be 28." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.365, + 0.485, + 0.531 + ], + "angle": 0, + "content": "
ModelParamThroughputGenEval ↑DPG ↑CLIP ↑Image Reward ↑
PixArt-α [12]0.6B0.420.4871.10.3161.15
PixArt-Σ [13]0.6B0.460.5380.50.3171.13
SD-1.5 [1]0.9B-0.4363.20.2870.19
SD-2.1 [2]0.9B-0.5064.20.2810.29
KOALA [35]1.0B0.660.5274.30.3171.05
MicroDiT [62]1.2B-0.5275.10.3181.25
Sana [74]1.6B1.000.6684.80.3271.25
LUMINA-Next [87]2.0B0.060.4674.60.3090.88
SDXL [53]2.6B0.180.5574.70.3010.99
Playgroundv2 [37]2.6B0.180.5974.50.3171.25
Playgroundv2.5 [38]2.6B0.180.5675.50.3191.34
IF-XL [15]5.5B0.060.6175.60.3110.65
Ours w/o KD0.38B1.040.6176.30.3211.20
SnapGen (ours)0.38B1.040.6681.10.3321.32
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.551, + 0.483, + 0.672 + ], + "angle": 0, + "content": "Qualitative Comparison. To visually evaluate the image-text alignment and aesthetics, we compare the generated images from different T2I models in Fig. 1. We observe that many existing models fail to fully capture the full prompt and miss important elements. Further, human generations often result in smoothened-out faces, leading to the loss of details. In contrast, our model generates much more photorealistic images with better image-text alignment." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.675, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Human Evaluation. For a thorough comparison between baselines, we perform a user study with the widely used Parti prompts [81]. We use SDXL, SD3-M, and SD3.5-Large models as the baselines and generate images using this prompt set. We ask the users to select images with better attributes between the baselines and our model. These attributes include image-text alignment, aesthetic quality, and realistic images. Fig. 8 shows that our model convincingly outperforms SDXL on all three attributes. Our model beats SD3 on image-text alignment and realistic generations, with a tie on aesthetic quality. Compared to the teacher (SD3.5-Large), we lag the teacher a bit behind on the image-text alignment, yet our model still has better realistic generations, and yields a toss-up on aesthetic quality. With this study, we can conclude our efficient T2I pipeline" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.267, + 0.905, + 0.297 + ], + "angle": 0, + "content": "achieves generation quality that is quite comparable to the SD3.5-Large teacher that has 8.1B parameters." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.299, + 0.907, + 0.465 + ], + "angle": 0, + "content": "Few-Step Generation. After step-distillation (Sec. 3.4), our model can generate high-quality images within a few steps. Fig. 9 compares our model's performance before and after step-distillation, with respective GenEval scores. The results demonstrate that our model, after step distillation, achieves comparable performance to the baseline model with 28 steps, even when using only 4 or 8 steps. While the few-step generation shows slight qualitative degradation compared to the 28-step baseline, it still outperforms most existing T2I models with significantly more inference steps, such as SDXL (50 steps) and PixArt-\\(\\alpha\\) (100 steps)." + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.473, + 0.903, + 0.655 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.666, + 0.908, + 0.694 + ], + "angle": 0, + "content": "Figure 9. Performance comparison of few-step generation for our model before (top) and after step distillation (bottom)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.723, + 0.634, + 0.739 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In this work, we propose a novel and efficient T2I model for high-resolution generation on mobile phones. We systematically detail the process to obtain a tiny 379M parameter UNet architecture along with an efficient latent decoder. We devise a novel training method consisting of multi-stage pre-training followed by knowledge distillation from a large teacher and adversarial step distillation. With these, we achieve an extremely efficient T2I model that comprehensively outperforms many existing multi-billion parameter models such as SDXL, Lumina-Next, and Playgroundv2." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.96, + 0.516, + 0.972 + ], + "angle": 0, + "content": "8004" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.094, + 0.091, + 0.188, + 0.105 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.482, + 0.156 + ], + "angle": 0, + "content": "[1] Stability AI. Stable diffusion 1.5. https://huggingface.co/stable-diffusion-v1-5/stable-diffusion-v1-5, 2022.8" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.157, + 0.482, + 0.197 + ], + "angle": 0, + "content": "[2] Stability AI. Stable diffusion 2.1. https://huggingface.co/stabilityai/stable-diffusion-2-1, 2022.8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.2, + 0.482, + 0.227 + ], + "angle": 0, + "content": "[3] Stability AI. Stable diffusion 3.5. https://github.com/Stability-AI/sd3.5, 2024.2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.228, + 0.481, + 0.269 + ], + "angle": 0, + "content": "[4] Stability AI. Stable diffusion 3.5 large. https://huggingface.co/stabilityai/stable-diffusion-3.5-large, 2024.6,7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.271, + 0.481, + 0.311 + ], + "angle": 0, + "content": "[5] Stability AI. Stable diffusion 3.5 large turbo. https://huggingface.co/stabilityai/stable-diffusion-3.5-large-turbo, 2024. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.314, + 0.482, + 0.395 + ], + "angle": 0, + "content": "[6] Vladimir Arkhipkin, Viacheslav Vasilev, Andrei Filatov, Igor Pavlov, Julia Agafonova, Nikolai Gerasimenko, Anna Averchenkova, Evelina Mironova, Anton Bukashkin, Konstantin Kulikov, et al. Kandinsky 3: Text-to-image synthesis for multifunctional generative framework. arXiv preprint arXiv:2410.21061, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.398, + 0.482, + 0.438 + ], + "angle": 0, + "content": "[7] Shakhmatov Arseni, Razzhigaev Anton, Nikolich Aleksandr, Arkhipkin Vladimir, Pavlov Igor, Kuznetsov Andrey, and Denis Dimitrov. Kandinsky 2.1, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.44, + 0.482, + 0.507 + ], + "angle": 0, + "content": "[8] Fan Bao, Shen Nie, Kaiwen Xue, Yue Cao, Chongxuan Li, Hang Su, and Jun Zhu. All are worth words: A vit backbone for diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 22669-22679, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.51, + 0.482, + 0.577 + ], + "angle": 0, + "content": "[9] Han Cai, Junyan Li, Muyan Hu, Chuang Gan, and Song Han. Efficientvit: Lightweight multi-scale attention for high-resolution dense prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17302-17313, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.58, + 0.482, + 0.648 + ], + "angle": 0, + "content": "[10] Thibault Castells, Hyoung-Kyu Song, Bo-Kyeong Kim, and Shinkook Choi. LD-Pruner: Efficient Pruning of Latent Diffusion Models using Task-Agnostic Insights. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 821–830, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.65, + 0.482, + 0.718 + ], + "angle": 0, + "content": "[11] Thibault Castells, Hyoung-Kyu Song, Tairen Piao, Shinkook Choi, Bo-Kyeong Kim, Hanyoung Yim, Changgwun Lee, Jae Gon Kim, and Tae-Ho Kim. EdgeFusion: On-Device Text-to-Image Generation. arXiv preprint arXiv:2404.11925, 2024.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.72, + 0.482, + 0.788 + ], + "angle": 0, + "content": "[12] Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, et al. Pixart-α: Fast training of diffusion transformer for photorealistic text-to-image synthesis. arXiv preprint arXiv:2310.00426, 2023. 2, 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.79, + 0.482, + 0.858 + ], + "angle": 0, + "content": "[13] Junsong Chen, Chongjian Ge, Enze Xie, Yue Wu, Lewei Yao, Xiaozhe Ren, Zhongdao Wang, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart- \\(\\cdot\\) sigma: Weak-to-strong training of diffusion transformer for 4k text-to-image generation. arXiv preprint arXiv:2403.04692, 2024. 2, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.86, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[14] Tri Dao and Albert Gu. Transformers are ssms: Generalized models and efficient algorithms through structured state space duality. arXiv preprint arXiv:2405.21060, 2024. 3" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.119 + ], + "angle": 0, + "content": "[15] DeepFloyd. Deepfloyd. https://github.com/deep-floyd/IF, 2023. 2, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.122, + 0.905, + 0.178 + ], + "angle": 0, + "content": "[16] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.18, + 0.905, + 0.206 + ], + "angle": 0, + "content": "[17] P Kingma Diederik. Adam: A method for stochastic optimization. (No Title), 2014. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.208, + 0.905, + 0.29 + ], + "angle": 0, + "content": "[18] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In *Forty-first International Conference on Machine Learning*, 2024. 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.292, + 0.904, + 0.319 + ], + "angle": 0, + "content": "[19] Leonhard Euler. Institutionum calculi integralis. imp. Acad. imp. Sa'ent., 1768. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.321, + 0.905, + 0.39 + ], + "angle": 0, + "content": "[20] Peng Gao, Le Zhuo, Ziyi Lin, Chris Liu, Junsong Chen, Ruoyi Du, Enze Xie, Xu Luo, Longtian Qiu, Yuhang Zhang, et al. Lumina-T2X: Transforming Text into Any Modality, Resolution, and Duration via Flow-based Large Diffusion Transformers. arXiv preprint arXiv:2405.05945, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.392, + 0.904, + 0.446 + ], + "angle": 0, + "content": "[21] Dhruba Ghosh, Hannaneh Hajishirzi, and Ludwig Schmidt. Geneval: An object-focused framework for evaluating text-to-image alignment. Advances in Neural Information Processing Systems, 36, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.449, + 0.905, + 0.503 + ], + "angle": 0, + "content": "[22] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.505, + 0.905, + 0.56 + ], + "angle": 0, + "content": "[23] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.562, + 0.905, + 0.602 + ], + "angle": 0, + "content": "[24] Alex Henry, Prudhvi Raj Dachapally, Shubham Pawar, and Yuxuan Chen. Query-key normalization for transformers. arXiv preprint arXiv:2010.04245, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.604, + 0.905, + 0.672 + ], + "angle": 0, + "content": "[25] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.675, + 0.905, + 0.702 + ], + "angle": 0, + "content": "[26] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.704, + 0.905, + 0.731 + ], + "angle": 0, + "content": "[27] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.733, + 0.905, + 0.772 + ], + "angle": 0, + "content": "[28] Andrew G Howard. Mobilenets: Efficient convolutional neural networks for mobile vision applications. arXiv preprint arXiv:1704.04861, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.776, + 0.905, + 0.829 + ], + "angle": 0, + "content": "[29] Xiwei Hu, Rui Wang, Yixiao Fang, Bin Fu, Pei Cheng, and Gang Yu. Ella: Equip diffusion models with llm for enhanced semantic alignment. arXiv preprint arXiv:2403.05135, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.832, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[30] Anil Kag, Huseyin Coskun, Jierun Chen, Junli Cao, Willi Menapace, Aliaksandr Siarohin, Sergey Tulyakov, and Jian Ren. Ascan: Asymmetric convolution-attention networks for efficient recognition and generation. arXiv preprint arXiv:2411.04967, 2024. 3, 7" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.961, + 0.515, + 0.972 + ], + "angle": 0, + "content": "8005" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.16 + ], + "angle": 0, + "content": "[31] Angelos Katharopoulos, Apoorv Vyas, Nikolaos Pappas, and François Fleuret. Transformers are rnns: Fast autoregressive transformers with linear attention. In International conference on machine learning, pages 5156-5165. PMLR, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.164, + 0.482, + 0.232 + ], + "angle": 0, + "content": "[32] Bo-Kyeong Kim, Hyoung-Kyu Song, Thibault Castells, and Shinkook Choi. Bk-sdm: Architecturally Compressed Stable Diffusion for Efficient Text-to-Image Generation. In Workshop on Efficient Systems for Foundation Models@ ICML2023, 2023. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.234, + 0.482, + 0.274 + ], + "angle": 0, + "content": "[33] Diederik P Kingma. Auto-encoding variational bayes. 2nd International Conference on Learning Representations, 2014. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.278, + 0.482, + 0.345 + ], + "angle": 0, + "content": "[34] Tuomas Kynkänniemi, Miika Aittala, Tero Karras, Samuli Laine, Timo Aila, and Jaakko Lehtinen. Applying guidance in a limited interval improves sample and distribution quality in diffusion models. arXiv preprint arXiv:2404.07724, 2024.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.349, + 0.482, + 0.403 + ], + "angle": 0, + "content": "[35] Youngwan Lee, Kwanyong Park, Yoorhim Cho, Yong-Ju Lee, and Sung Ju Hwang. Koala: Empirical lessons toward memory-efficient and fast diffusion models for text-to-image synthesis, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.406, + 0.482, + 0.433 + ], + "angle": 0, + "content": "[36] Daiqing Li, Aleks Kamko, Ali Sabet, Ehsan Akhgari, Linmiao Xu, and Suhail Doshi. Playground v1, . 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.435, + 0.482, + 0.462 + ], + "angle": 0, + "content": "[37] Daiqing Li, Aleks Kamko, Ali Sabet, Ehsan Akhgari, Linmiao Xu, and Suhail Doshi. Playground v2, . 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.464, + 0.482, + 0.532 + ], + "angle": 0, + "content": "[38] Daiqing Li, Aleks Kamko, Ehsan Akhgari, Ali Sabet, Linmiao Xu, and Suhail Doshi. Playground V2. 5: Three Insights towards Enhancing Aesthetic Quality in Text-to-Image Generation. arXiv preprint arXiv:2402.17245, 2024. 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.535, + 0.482, + 0.618 + ], + "angle": 0, + "content": "[39] Hao Li, Yang Zou, Ying Wang, Orchid Majumder, Yusheng Xie, R Manmatha, Ashwin Swaminathan, Zhuowen Tu, Stefano Ermon, and Stefano Soatto. On the scalability of diffusion-based text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9400-9409, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.62, + 0.482, + 0.688 + ], + "angle": 0, + "content": "[40] Yanyu Li, Huan Wang, Qing Jin, Ju Hu, Pavlo Chemerys, Yun Fu, Yanzhi Wang, Sergey Tulyakov, and Jian Ren. Snapfusion: Text-to-Image Diffusion Model on Mobile Devices within Two Seconds. Advances in Neural Information Processing Systems, 36, 2024. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.69, + 0.482, + 0.772 + ], + "angle": 0, + "content": "[41] Zhimin Li, Jianwei Zhang, Qin Lin, Jiangfeng Xiong, Yanxin Long, Xinchi Deng, Yingfang Zhang, Xingchao Liu, Minbin Huang, Zedong Xiao, et al. Hunyuan-DiT: A Powerful Multi-Resolution Diffusion Transformer with Fine-Grained Chinese Understanding. arXiv preprint arXiv:2405.08748, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.775, + 0.482, + 0.814 + ], + "angle": 0, + "content": "[42] Shanchuan Lin, Anran Wang, and Xiao Yang. SDXL-Lightning: Progressive Adversarial Diffusion Distillation, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.818, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[43] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 5, 7" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "[44] Yaron Lipman, Ricky TQ Chen, Heli Ben-Hamu, Maximilian Nickel, and Matt Le. Flow matching for generative modeling. arXiv preprint arXiv:2210.02747, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.137, + 0.905, + 0.205 + ], + "angle": 0, + "content": "[45] Bingchen Liu, Ehsan Akhgari, Alexander Visheratin, Aleks Kamko, Linmiao Xu, Shivam Shrirao, Joao Souza, Suhail Doshi, and Daiqing Li. Playground v3: Improving text-to-image alignment with deep-fusion large language models. arXiv preprint arXiv:2409.10695, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.207, + 0.905, + 0.247 + ], + "angle": 0, + "content": "[46] Songhua Liu, Weihao Yu, Zhenxiong Tan, and Xinchao Wang. Linfusion: 1gpu, 1 minute, 16k image. arXiv preprint arXiv:2409.02097, 2024. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.25, + 0.905, + 0.291 + ], + "angle": 0, + "content": "[47] Xingchao Liu, Chengyue Gong, and Qiang Liu. Flow straight and fast: Learning to generate and transfer data with rectified flow. arXiv preprint arXiv:2209.03003, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.294, + 0.905, + 0.361 + ], + "angle": 0, + "content": "[48] Xian Liu, Jian Ren, Aliaksandr Siarohin, Ivan Skorokhodov, Yanyu Li, Dahua Lin, Xihui Liu, Ziwei Liu, and Sergey Tulyakov. Hyperhuman: Hyper-realistic human generation with latent structural diffusion. arXiv preprint arXiv:2310.08579, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.364, + 0.905, + 0.432 + ], + "angle": 0, + "content": "[49] Nanye Ma, Mark Goldstein, Michael S Albergo, Nicholas M Boffi, Eric Vanden-Eijnden, and Saining Xie. Sit: Exploring flow and diffusion-based generative models with scalable interpolant transformers. arXiv preprint arXiv:2401.08740, 2024. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.435, + 0.905, + 0.53 + ], + "angle": 0, + "content": "[50] Willi Menapace, Aliaksandr Siarohin, Ivan Skorokhodov, Ekaterina Deyneka, Tsai-Shien Chen, Anil Kag, Yuwei Fang, Aleksei Stoliar, Elisa Ricci, Jian Ren, et al. Snap video: Scaled spatiotemporal transformers for text-to-video synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7038-7048, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.533, + 0.616, + 0.547 + ], + "angle": 0, + "content": "[51] OpenAI. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.549, + 0.905, + 0.603 + ], + "angle": 0, + "content": "[52] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4195-4205, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.606, + 0.905, + 0.673 + ], + "angle": 0, + "content": "[53] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 2, 3, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.676, + 0.905, + 0.744 + ], + "angle": 0, + "content": "[54] Adam Polyak, Amit Zohar, Andrew Brown, Andros Tjandra, Animesh Sinha, Ann Lee, Apoorv Vyas, Bowen Shi, Chih-Yao Ma, Ching-Yao Chuang, et al. Movie gen: A cast of media foundation models. arXiv preprint arXiv:2410.13720, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.747, + 0.905, + 0.815 + ], + "angle": 0, + "content": "[55] Danfeng Qin, Chas Leichner, Manolis Delakis, Marco Fornoni, Shixin Luo, Fan Yang, Weijun Wang, Colby Banbury, Chengxi Ye, Berkin Akin, et al. Mobilenetv4-universal models for the mobile ecosystem. arXiv preprint arXiv:2404.10518, 2024.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.818, + 0.905, + 0.899 + ], + "angle": 0, + "content": "[56] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 7" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.961, + 0.516, + 0.972 + ], + "angle": 0, + "content": "8006" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.161 + ], + "angle": 0, + "content": "[57] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In Proceedings of the 38th International Conference on Machine Learning, pages 8821-8831. PMLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.162, + 0.483, + 0.232 + ], + "angle": 0, + "content": "[58] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.232, + 0.482, + 0.315 + ], + "angle": 0, + "content": "[59] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S. Sara Mahdavi, Rapha Gontijo Lopes, Tim Salimans, Jonathan Ho, David J Fleet, and Mohammad Norouzi. Photorealistic text-to-image diffusion models with deep language understanding, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.316, + 0.482, + 0.371 + ], + "angle": 0, + "content": "[60] Axel Sauer, Frederic Boesel, Tim Dockhorn, Andreas Blattmann, Patrick Esser, and Robin Rombach. Fast high-resolution image synthesis with latent adversarial diffusion distillation. arXiv preprint arXiv:2403.12015, 2024. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.371, + 0.482, + 0.425 + ], + "angle": 0, + "content": "[61] Axel Sauer, Dominik Lorenz, Andreas Blattmann, and Robin Rombach. Adversarial diffusion distillation. In European Conference on Computer Vision, pages 87-103. Springer, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.427, + 0.482, + 0.482 + ], + "angle": 0, + "content": "[62] Vikash Sehwag, Xianghao Kong, Jingtao Li, Michael Spranger, and Lingjuan Lyu. Stretching each dollar: Diffusion training from scratch on a micro-budget. arXiv preprint arXiv:2407.15811, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.483, + 0.482, + 0.51 + ], + "angle": 0, + "content": "[63] Noam Shazeer. Fast transformer decoding: One write-head is all you need. arXiv preprint arXiv:1911.02150, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.511, + 0.482, + 0.594 + ], + "angle": 0, + "content": "[64] Shelly Sheynin, Adam Polyak, Uriel Singer, Yuval Kirstain, Amit Zohar, Oron Ashual, Devi Parikh, and Yaniv Taigman. Emu edit: Precise image editing via recognition and generation tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8871-8879, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.595, + 0.482, + 0.636 + ], + "angle": 0, + "content": "[65] Yuda Song, Zehao Sun, and Xuanwu Yin. Sdxs: Realtime one-step latent diffusion models with image conditions. arXiv preprint arXiv:2403.16627, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.636, + 0.482, + 0.69 + ], + "angle": 0, + "content": "[66] Jianlin Su, Murtadha Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu. Roformer: Enhanced transformer with rotary position embedding. Neurocomputing, 568:127063, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.691, + 0.482, + 0.747 + ], + "angle": 0, + "content": "[67] Yang Sui, Yanyu Li, Anil Kag, Yerlan Idelbayev, Junli Cao, Ju Hu, Dhritiman Sagar, Bo Yuan, Sergey Tulyakov, and Jian Ren. Bitsfusion: 1.99 bits weight quantization of diffusion model. arXiv preprint arXiv:2406.04333, 2024. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.748, + 0.482, + 0.829 + ], + "angle": 0, + "content": "[68] Gemma Team, Morgane Riviere, Shreya Pathak, Pier Giuseppe Sessa, Cassidy Hardin, Surya Bhupatiraju, Léonard Hussenot, Thomas Mesnard, Bobak Shahriari, Alexandre Ramé, et al. Gemma 2: Improving open language models at a practical size. arXiv preprint arXiv:2408.00118, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.831, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[69] Kcolors Team. Kcolors: Effective Training of Diffusion Model for Photorealistic Text-to-Image Synthesis. arXiv preprint, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.872, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[70] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.092, + 0.905, + 0.135 + ], + "angle": 0, + "content": "[71] Zhendong Wang, Huangjie Zheng, Pengcheng He, Weizhu Chen, and Mingyuan Zhou. Diffusion-gan: Training gans with diffusion. arXiv preprint arXiv:2206.02262, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.137, + 0.905, + 0.193 + ], + "angle": 0, + "content": "[72] Shitao Xiao, Yueze Wang, Junjie Zhou, Huaying Yuan, Xingrun Xing, Ruiran Yan, Shuting Wang, Tiejun Huang, and Zheng Liu. Omnigen: Unified image generation. arXiv preprint arXiv:2409.11340, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.194, + 0.905, + 0.237 + ], + "angle": 0, + "content": "[73] Zhisheng Xiao, Karsten Kreis, and Arash Vahdat. Tackling the generative learning trilemma with denoising diffusion gans. arXiv preprint arXiv:2112.07804, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.239, + 0.905, + 0.307 + ], + "angle": 0, + "content": "[74] Enze Xie, Junsong Chen, Junyu Chen, Han Cai, Yujun Lin, Zhekai Zhang, Muyang Li, Yao Lu, and Song Han. Sana: Efficient high-resolution image synthesis with linear diffusion transformers. arXiv preprint arXiv:2410.10629, 2024. 2, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.31, + 0.905, + 0.381 + ], + "angle": 0, + "content": "[75] Jiazheng Xu, Xiao Liu, Yuchen Wu, Yuxuan Tong, Qinkai Li, Ming Ding, Jie Tang, and Yuxiao Dong. Imagereward: Learning and evaluating human preferences for text-to-image generation. Advances in Neural Information Processing Systems, 36, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.382, + 0.905, + 0.452 + ], + "angle": 0, + "content": "[76] Yanwu Xu, Mingming Gong, Shaoan Xie, Wei Wei, Matthias Grundmann, Kayhan Batmanghelich, and Tingbo Hou. Semi-implicit denoising diffusion models (siddms). Advances in neural information processing systems, 36:17383, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.453, + 0.905, + 0.525 + ], + "angle": 0, + "content": "[77] Yanwu Xu, Yang Zhao, Zhisheng Xiao, and Tingbo Hou. Ufogen: You forward once large scale text-to-image generation via diffusion gans. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8196-8206, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.527, + 0.905, + 0.595 + ], + "angle": 0, + "content": "[78] Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, et al. Cogvideox: Text-to-video diffusion models with an expert transformer. arXiv preprint arXiv:2408.06072, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.597, + 0.905, + 0.655 + ], + "angle": 0, + "content": "[79] Tianwei Yin, Michael Gharbi, Taesung Park, Richard Zhang, Eli Shechtman, Fredo Durand, and William T Freeman. Improved distribution matching distillation for fast image synthesis. arXiv preprint arXiv:2405.14867, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.657, + 0.905, + 0.727 + ], + "angle": 0, + "content": "[80] Tianwei Yin, Michael Gharbi, Richard Zhang, Eli Shechtman, Fredo Durand, William T Freeman, and Taesung Park. One-step diffusion with distribution matching distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6613-6623, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.728, + 0.905, + 0.799 + ], + "angle": 0, + "content": "[81] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, et al. Scaling autoregressive models for content-rich text-to-image generation. Transactions on Machine Learning Research. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.8, + 0.905, + 0.842 + ], + "angle": 0, + "content": "[82] Biao Zhang and Rico Sennrich. Root mean square layer normalization. Advances in Neural Information Processing Systems, 32, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.845, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[83] Dingkun Zhang, Sijia Li, Chen Chen, Qingsong Xie, and Haonan Lu. Laptop-Diff: Layer Pruning and Normalized Distillation for Compressing Diffusion Models. arXiv preprint arXiv:2404.11098, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.092, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.96, + 0.515, + 0.972 + ], + "angle": 0, + "content": "8007" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.134 + ], + "angle": 0, + "content": "[84] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.483, + 0.205 + ], + "angle": 0, + "content": "[85] Zhixing Zhang, Ligong Han, Arnab Ghosh, Dimitris N Metaxas, and Jian Ren. Sine: Single image editing with text-to-image diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6027-6037, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.207, + 0.482, + 0.261 + ], + "angle": 0, + "content": "[86] Yang Zhao, Yanwu Xu, Zhisheng Xiao, and Tingbo Hou. Mobilediffusion: Subsecond text-to-image generation on mobile devices. arXiv preprint arXiv:2311.16567, 2023. 2, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.263, + 0.482, + 0.331 + ], + "angle": 0, + "content": "[87] Le Zhuo, Ruoyi Du, Han Xiao, Yangguang Li, Dongyang Liu, Rongjie Huang, Wenze Liu, Lirui Zhao, Fu-Yun Wang, Zhanyu Ma, et al. Lumina-last: Making lumina-t2x stronger and faster with next-dit. arXiv preprint arXiv:2406.18583, 2024. 2, 8" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.331 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.961, + 0.515, + 0.971 + ], + "angle": 0, + "content": "8008" + } + ] +] \ No newline at end of file diff --git a/2025/SnapGen_ Taming High-Resolution Text-to-Image Models for Mobile Devices with Efficient Architectures and Training/d6336171-7440-4517-9fbd-fdd52a399434_origin.pdf b/2025/SnapGen_ Taming High-Resolution Text-to-Image Models for Mobile Devices with Efficient Architectures and Training/d6336171-7440-4517-9fbd-fdd52a399434_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..cd6337febea2352b87062ddbda5bdff2e23954ec --- /dev/null +++ b/2025/SnapGen_ Taming High-Resolution Text-to-Image Models for Mobile Devices with Efficient Architectures and Training/d6336171-7440-4517-9fbd-fdd52a399434_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2d75cb3def1bb6cb0eb7c97180d6509b71368eab19a547d7f263394dd833dba5 +size 5121267 diff --git a/2025/SnapGen_ Taming High-Resolution Text-to-Image Models for Mobile Devices with Efficient Architectures and Training/full.md b/2025/SnapGen_ Taming High-Resolution Text-to-Image Models for Mobile Devices with Efficient Architectures and Training/full.md new file mode 100644 index 0000000000000000000000000000000000000000..922013e27af20123c4532f98565f9facf4884c13 --- /dev/null +++ b/2025/SnapGen_ Taming High-Resolution Text-to-Image Models for Mobile Devices with Efficient Architectures and Training/full.md @@ -0,0 +1,364 @@ +# SnapGen: Taming High-Resolution Text-to-Image Models for Mobile Devices with Efficient Architectures and Training + +Jierun Chen $^{1,2,*}$ Dongting Hu $^{1,3,*}$ Xijie Huang $^{1,2,*}$ Huseyin Coskun $^{1}$ Arpit Sahni $^{1}$ Aarush Gupta $^{1}$ Anujraaj Goyal $^{1}$ Dishani Lahiri $^{1}$ Rajesh Singh $^{1}$ Yerlan Idelbayev $^{1}$ Junli Cao $^{1}$ Yanyu Li $^{1}$ Kwang-Ting Cheng $^{2}$ S.-H. Gary Chan $^{2}$ Mingming Gong $^{3,4}$ Sergey Tulyakov $^{1}$ Anil Kag $^{1,\dagger}$ Yanwu Xu $^{1,\dagger}$ Jian Ren $^{1,\dagger}$ + +$^{1}$ Snap Inc. $^{2}$ HKUST $^{3}$ The University of Melbourne $^{4}$ MBZUAI + +* Equal contribution † Equal advising + +Project Page: https://snap-research.github.io/snapgen + +
OursPixArt-αLumina-NextSD3-MediumSDXLPlaygroundv2SD3.5-Large
Param0.38B0.6B2B2B2.6B2.6B8.1B
MobileYESNONONONONONO
+ +"an old raccoon wearing a top hat and holding an apple, oil painting in the style of van gogh, ..." + +![](images/aacb2837f306dbd5f9799eb9b41fdaa8e940c7d011704296ab0406d4cf3c4d6f.jpg) + +"a dolphin in an astronaut suit, Animals, Simple Detail" + +![](images/e9d2e69529494de4303059780935b53f3466e0207cd4a7a965b2e547d1e4c28e.jpg) + +"..., young sudanese female, glamour, natural, front view, extreme detailed and texture skin, ..." + +![](images/bf25a52443399be51fd3200f42287d00d2a2fd5514690042d0061c1f9fd468d6.jpg) + +... Llama wearing sunglasses standing on the deck of a spaceship with the Earth in the background, ... + +![](images/d7930e76b7a075864239cf2e635ecc6c558325029304c0a3d19cee749799ad40.jpg) + +"... cute monster ... eating a sashimi while sitting at a breakfast table full of fruits and insects" + +![](images/290a0e2fa4e822994c983446619421f959f3cfc23933ccc1528d4de079fc86f4.jpg) +Figure 1. Comparison of various text-to-image models in terms of model size, mobile device compatibility, and visual output quality. Our model, with only $379M$ parameters, demonstrates competitive visual quality while being mobile-compatible. Input text prompts are shown above each image grid; all images are generated at $1024^2$ resolution—zoom in for details. More examples are shown in webpage. + +# Abstract + +Existing text-to-image (T2I) diffusion models face several limitations, including large model sizes, slow runtime, and low-quality generation on mobile devices. This paper aims to address all of these challenges by developing an extremely small and fast T2I model that generates high-resolution and high-quality images on mobile platforms. We propose several techniques to achieve this goal. First, we systematically examine the design choices of the network architecture to reduce model parameters and latency, while ensuring high-quality generation. Second, to further improve generation quality, we employ cross-architecture knowledge distillation from a much larger model, using a multi-level approach to guide the training of our model from scratch. Third, we enable a few-step generation by integrating adversarial guidance with knowledge distillation. For the first time, our model SnapGen, demonstrates the generation of $1024^2$ px images on a mobile device around 1.4 seconds. On ImageNet-1K, our model, with only 372M parameters, achieves an FID of 2.06 for $256^2$ px generation. On T2I benchmarks (i.e., GenEval and DPG-Bench), our model with merely 379M parameters, surpasses large-scale models with billions of parameters at a significantly smaller size (e.g., $7 \times$ smaller than SDXL, $14 \times$ smaller than IF-XL). + +# 1. Introduction + +Large-scale text-to-image (T2I) diffusion models [12, 13, 15, 51, 53, 57-59] have achieved remarkable success in content generation, powering numerous applications like image editing [48, 64, 72, 85] and video creation [50, 54, 78]. However, T2I models often come with substantial model sizes and slow runtime, and deploying them on the cloud raises concerns related to data security and high costs [65]. + +To address these challenges, there is huge growing interest in developing smaller and faster T2I models through techniques like model compression (e.g., pruning and quantization) [40, 67, 86], step reduction by distillation [77, 80], and efficient attention mechanisms that mitigate the quadratic complexity [46, 74]. Nevertheless, current works still encounter limitations, e.g., low-resolution generation on mobile devices, that constrain their broader application. + +Most importantly, a critical question remains unexplored: how can we train a T2I model from scratch to generate high-quality, high-resolution images on mobile? Such a model would offer substantial advantages in speed, compactness, cost-effectiveness, and secure deployment. To build this model, we introduce several innovations: + +- Efficient Network Architectures: We conduct an in-depth examination of network architectures, including the denoising UNet and Autoencoder (AE), to obtain optimal trade-off between resource usage and performance. Un + +like prior works that optimize and compress pre-trained diffusion models [10, 32, 83], we directly focus on macro- and micro-level design choices to achieve a novel architecture that greatly reduces model size and computational complexity, while preserving high-quality generation. + +- Improved Training Techniques: We introduce several improvements to train a compact T2I model from scratch. We utilize flow matching [44, 47] as objective, aligning with larger models like SD3 [18] and SD3.5 [3]. This design enables effective knowledge and step distillation, transferring rich representations from large-scale diffusion models to our much smaller one. Further, we propose a multi-level knowledge distillation with a timestep-aware scaling that combines multiple training objectives. Instead of weighting objectives through a linear combination as in prior works [32, 46], we consider target prediction difficulty (i.e., student-teacher difference) across various timesteps in flow matching. +- Advanced Step Distillation: We perform step distillation on our model by combining the adversarial training along with the knowledge distillation using a few-step teacher model (i.e., SD3.5-Large-Turbo [5]), enabling ultra-fast high-quality generation with only 4 or 8 steps. + +We demonstrate the superior advantages of our approach and model through extensive experiments: + +- On ImageNet-1K [16] class-conditional image generation task, our model achieves the FID comparable to existing works with significantly reduced model size and computation, i.e., half the model size and one-third of the compute resources compared with SiT-XL[49], as in Tab. 1. +- For large-scale T2I generation, our UNet model, with only 379M parameters, demonstrates superior generation quality compared to billion-parameter models [42, 53, 87], e.g., improved metrics on benchmark datasets (Tab. 3) and human evaluation (Fig. 8). +- Our compressed decoder, trained from scratch, has competitive reconstruction quality compared to commonly used models [18, 53], with more than $36 \times$ smaller size, enabling the mobile deployment. +- Notably, for the first time, we show a T2I model achieving high-resolution generation (e.g., $1024^2$ px) on mobile devices (e.g., iPhone 16 Pro-Max) around 1.4 seconds. + +# 2. Related Work + +High-Resolution Text-to-Image Models have emerged with advanced architectures and multi-stage approaches designed to enhance visual fidelity and user customization. SDXL [53] is a pioneering work in this field, employing a refined cascading approach with UNet backbone to generate high-detailed images, resulting in photorealistic outputs that maintain sharpness and clarity. The following studies explore different techniques like more advanced text encoders, better image refinement, or improved dataset preparation, + +to obtain better text-image alignment or higher-quality generation [6, 7, 15, 20, 30, 36-38, 41, 45, 48, 69]. However, most of these models contain billions of parameters, making them extremely slow, and not being able to run on resource-contained hardware like mobile devices. In this work, we aim to build a small and fast model that can perform high-resolution generation even on mobile platforms. + +Efficient Diffusion Models address the challenges of bulky model size and long runtime. There have been efforts exploring the architecture optimization to remove redundancy from large models, demonstrating the on-device generation within seconds [11, 40, 67, 86]. However, these models are constrained to low-resolution output, i.e., $512^{2}$ px. To enable efficient high-resolution generation, SANA [74] and LinFusion [46] incorporate linear attention [9, 14, 31] to achieve 1K generation on laptop GPUs. In contrast, we target a broader range of platforms, supporting high-resolution generation (e.g., 1K) directly on mobile devices. + +Knowledge Distillation in Diffusion Models. In the context of diffusion models, previous works focus on distilling large, high-capacity teacher models into more compact, efficient student models within a homogeneous architecture [32, 46]. They reduce the complexity of the model by removing certain components like attention [70] or residual blocks [23], while maintaining the architectural structure. However, our approach diverges from this trend by utilizing a heterogeneous architecture for more aggressively efficient yet challenging distillation. + +Adversarial Step Distillation uses the techniques from adversarial training [22] to reduce the number of diffusion steps, while maintaining high image quality [60, 61]. For example, UFOGen [77] employs a diffusion-GAN formulation [71, 73, 76] to significantly reduce inference time while maintaining competitive performance. DMD2 [79] builds on prior distillation methods by using distribution matching with adversarial loss. Different from exiting works, we conduct step-distillation on a very compact model and train the model along with the knowledge distillation. + +# 3. Method + +In this section, we present how to craft and train a highly efficient T2I model for high-resolution generation. Specifically, starting from the architecture designed in the latent diffusion model [58], we optimize both denoising backbone (Sec. 3.1, Fig. 2, and Fig. 3) and autoencoder (Sec. 3.2 and Fig. 4) to make them compact and fast, even on mobile devices. We then propose the improved training recipe and knowledge distillation (Sec. 3.3 and Fig. 5), empowering a high-performance T2I model. Lastly, we introduce our step distillation to significantly reduce the number of denoising steps for a faster T2I model (Sec. 3.4 and Fig. 7). + +# 3.1. Efficient UNet Architecture + +Here we describe the design choices for the denoising UNet. Baseline Architecture. We choose UNet from SDXL [53] as the baseline (Fig. 2(a)) for our backbone since it has superior efficiency and faster convergence [39] than pure transformer-based models [12, 13]. We adjust the UNet into a thinner and shorter model (i.e., reducing the number of transformer blocks from [0, 2, 10] in three stages to [0, 2, 4] and their channel dimensions from [320, 640, 1280] to [256, 512, 896]), and iterate design choices on top of it. + +Evaluation Metrics. We train models on ImageNet-1K [16] under class-conditional generation for 120 epochs, unless specified otherwise, and report the FID score [25] for $256^{2}$ px generation. Similarly to existing work [30], we inject the class conditions through a text template "a photo of ". We then encode it with a light text encoder to align the pipeline for the T2I generation. We also calculate the number of parameters for different models, floating point operations (FLOPs) (measured on a latent size of $128 \times 128$ , equivalent to a $1024 \times 1024$ image after decoding), and the runtime on mobile device (tested on iPhone 15 Pro). Detailed training and evaluation settings can be found in Supplemental Materials. In the following, we introduce the key architectural changes that improve the model. + +Remove Self-Attention from High-Resolution Stages. Self-attention (SA) layer is restricted by its quadratic computational complexity, incurring a heavy computational cost and high memory consumption for high-resolution input. As such, we keep the SA layer only in the lowest-resolution stage while removing it from other higher-resolution stages, i.e., Fig. 2 (b). This leads to $17\%$ fewer FLOPs and $24\%$ latency reduction, as shown in Fig. 3. Interestingly, we even observe a performance improvement, i.e., FID decreased to 3.12 from 3.76. We hypothesize that models with SA in high-resolution stages converge more slowly. + +Replace Conv with Expanded Separable Conv. Regular convolution (Conv) is redundant in both parameters and computation. To address this, we replace all Conv layers with the separable convolution (SepConv) [28], composed of a depthwise convolution (DW) followed by a pointwise convolution (PW), as shown in Fig. 2 (c). This replacement reduces parameters by $24\%$ and latency by $62\%$ , but also leads to a performance drop (FID increases from 3.12 to 3.38). To address the issue, we expand the intermediate channels. Specifically, the number of channels after the first PW layer is increased with an expansion ratio, and reduced back to the original number after the second PW layer. The expansion ratio is set to 2 to balance the trade-off between performance, latency, and model parameters. Such a design aligns our residual block with the recently proposed Universal Inverted Bottleneck (UIB) block [55]. As a result, our model achieves $15\%$ fewer parameters, $27\%$ less computation, and $2.4 \times$ speedup, while obtaining a lower FID. + +![](images/32fc1f037603aa4b02b36d78e8d57b7fbeb695f9b7b38d10ddfc5b30f0691153.jpg) +Figure 2. Efficient UNet. Starting from a thinner and shorter version of the UNet from SDXL (as in (a)), we explore a series of architectural changes, i.e., (b)-(f), to develop a smaller and faster model while retaining high-quality generation performance, as evaluated in Fig. 3. + +![](images/6ebcd4cd81b67055e08577ab90c0ff5d10cb6ffe379d1896340721d001ac9148.jpg) + +![](images/aa95f45746635077da11e1c3ab4ec04dba434d96490f47abbed1ac86caeba047.jpg) + +![](images/f60a78f2e2c62771c56e9f88618c757660afe5faecd4a583088143b4925d4edd.jpg) + +![](images/9905b120562358bddc6086f030628c8beb2ec45868c4467990b7fa37c80bffd1.jpg) + +![](images/50e5b5423661fd31a956e0611a3860e2ee9d56de9a472ac1830420e039b65eb7.jpg) + +![](images/c1dbff62b401a03d50e2895ca313bc72b820a589531ce0be6086fcd55c9f64e7.jpg) +Figure 3. Comparisons of Performance and Efficiency for various Design Choices of Efficient UNet. The generation quality is evaluated using FID calculated on ImageNet-1K for $256^{2}$ px generation. The efficiency metrics include model parameters, latency, and FLOPs. FLOPs and latency (on iPhone 15 Pro) are measured with a $128 \times 128$ latent, equivalent to a $1024 \times 1024$ decoded image, for one forward pass. We show the architecture enhancements that improve any of the metrics without hurting others. + +Trim FFN Layers. For the layers in the feed-forward network (FFN), the hidden channel expansion ratio is set to 4 by default and further doubled by using the gated unit. This substantially inflates the model parameters, computation, and memory usage. Following MobileDiffusion [86], we examine the efficacy of simply reducing the expansion ratio, as shown in Fig. 2 (d). We show that reducing the expansion ratio to 3 preserves comparable FID performance while reducing both parameters and FLOPs by $12\%$ . + +Replace MHSA with MQA. Multi-Head Self-Attention (MHSA) requires multiple sets of keys and values for each attention head. In contrast, Multi-Query Attention (MQA) [63] is more efficient by sharing a single set of keys and values across all heads. Replacing MHSA with MQA reduces parameters by $16\%$ and latency by $9\%$ , with mini + +mal impact on performance. Interestingly, the $9\%$ saving in latency exceeds the $6\%$ decrease in FLOPs, as the reduced memory access enables higher computational intensity. + +Inject Conditions to the First Stage. Cross-attention (CA) blends the conditional information (e.g., textural description) along with the spatial features to generate images that align with the condition. However, the UNet of SDXL only applies CA in transformer blocks starting from the second stage, resulting in the missing conditional guidance for the first stage. In response, we propose to introduce the conditional embeddings from the very first stage, as in Fig. 2(e). Specifically, we replace the residual blocks with transformer blocks that include CA and FFN while without SA layers. This adjustment makes the model smaller, faster, and more efficient while improving FID. + +Employ QK RMSNorm and RoPE Positional Embeddings. We extend two advanced techniques developed originally for language models, Query-Key (QK) Normalization [24] with RMSNorm [82] and Rotary Position Embedding (RoPE) [66], to enhance the model (Fig. 2 (f)). RMSNorm, applied after the Query-Key projection in the attention mechanism, reduces the risk of softmax saturation without sacrificing model expressiveness while stabilizing training for faster convergence. In addition, we adapt RoPE from one dimension to two dimensions for better supporting higher resolution since it significantly mitigates artifacts like repeated objects. Together, RMSNorm and RoPE introduce negligible computational and parameter overhead, while offering measurable gains in FID performance. + +Discussion. The above optimization results in an efficient and powerful diffusion backbone capable of generating high-resolution images on mobile devices. Before proceeding with large-scale T2I training, we compare the capacity of our model against existing works on ImageNet-1K. We train the model for 1,000 epochs by following the setting from prior work [58]. We evaluate the model using varied CFG [26, 34] across different inference timesteps. As shown in Tab. 1, our efficient UNet achieves comparable FID to SiT-XL [49], while being almost $45\%$ smaller. + +Table 1. Class-conditional image generation on ImageNet ${256} \times {256}$ with CFG. FLOPs are calculated for one forward pass. + +
ModelParam (M)FLOPs (G)FID↓
LDM-4 [58]4001043.60
UViT-L [8]287773.40
UViT-H [8]5011332.29
DiT-XL [52]6751192.27
SiT-XL [49]6751192.06
Ours372382.06
+ +# 3.2. Tiny and Fast Decoder + +Besides the denoising model, the decoder also takes a significant ratio of total runtime, especially for on-device deployment [40, 86]. Here we introduce a new architecture of decoder (Fig. 4) for efficient high-resolution generation. + +Baseline Decoder. We use the autoencoder (AE) from SD3 [18] as our baseline model (i.e., the same encoder from SD3 AE), due to its superior reconstruction quality. The AE maps an image $X \in \mathbb{R}^{H \times W \times 3}$ into a lower-dimensional latent $x \in \mathbb{R}^{\frac{H}{f} \times \frac{W}{f} \times c}$ ( $f, c$ as 8, 16 in SD3). The encoded latent $x$ is then decoded back to an image through a decoder. For high-resolution generation, we observe that the decoder in SD3 is very slow on mobile devices. Specifically, it encounters out-of-memory (OOM) errors when generating a $1024^2$ px image on both the ANE processor of the iPhone 15 Pro and the mobile GPU (Tab. 2). To overcome the latency issue, we propose a much smaller and faster decoder. + +Efficient Decoder. We conduct a series of experiments to + +![](images/d3551e77cf0a3ca96ce432b179f2b5e73f68505bc09ab4844ff816cca53a3bc2.jpg) + +![](images/12e3ce061452826b47b5537091521f8490242bceb1cbc04f50860da724a92895.jpg) +Figure 4. Comparisons of Decoder Architecture between (a) SDXL/SD3 decoder and (b) our tiny decoder. + +Table 2. Performance Comparison of Decoder. PSNR is calculated on COCO 2017 validation set [43]. FLOPs and latency (on iPhone 15 Pro) are measured for decoding a $128 \times 128$ latent into a $1024 \times 1024$ image. The decoder from SDXL and SD3 fail to run on the neural engine of mobile, resulting in a huge runtime. + +
DecoderChPSNRParam (M)FLOPs (G)Latency (ms) on ANELatency (ms) on GPU
SDXL [53]424.8949.494970OOM9469
SD3 [18]1627.9249.554970OOMOOM
Ours1627.851.38224174-
+ +decide the efficient decoder with the following key changes compared with the baseline architecture: + +1. We remove attention layers to greatly reduce peak memory without a noticeable impact on decoding quality. +2. We keep a minimal amount of GroupNorm (GN) to find a trade-off between latency and performance (i.e., mitigating the color shifting). +3. We make the decoder thinner (i.e., fewer channels or narrower width) and replace Conv with SepConvs. +4. We use fewer residual blocks in high-resolution stages. +5. We remove the Conv shortcut in residual blocks and use the upsampling layer for channel transition. + +Training of the Decoder. We train our decoder with the mean squared error (MSE) loss, lpips loss [84], adversarial loss [22], and discard the KL term [33] as the encoder is fixed. The decoder is trained on $256^2$ image patches with a batch size of 256 and for 1M iterations. As in Tab. 2, our tiny decoder achieves a competitive PSNR score for reconstruction, while being $35.9 \times$ smaller and $54.4 \times$ faster for high-resolution generation on mobile devices compared to conventional ones (e.g., the decoder from SDXL and SD3). + +Discussion of Total On-Device Latency. We finally measure T2I model latency for a $1024^{2}$ px generation on iPhone 16 Pro-Max. The decoder takes 119ms, and the per-step latency for the UNet is 274ms. This results in a $1.2\sim 2.3s$ runtime for 4 to 8 step generation. Note that text encoder runtime is negligible compared to other components [40]. + +![](images/b28475312a143afcb6eee32703245ffd0b1f9142e7b92e05ba5ac55afcf0445b.jpg) +Figure 5. Overview of Multi-level Knowledge Distillation, where we perform output distillation and feature distillation. + +![](images/2e659881515492e5fdebcca3c7cba7f8389e2d8bb8cd6a89f5b091ca594f0f3e.jpg) +Figure 6. Mean loss magnitude for task loss $\mathcal{L}_{\mathrm{task}}$ and output distillation loss $\mathcal{L}_{\mathrm{kd}}$ . + +![](images/c03cef90103a0769718ce76fd5d00f63febcb96b864f8f0f3ed961ec93ee280e.jpg) +Figure 7. Overview of Adversarial Step Distillation. Output distillation and distribution-matching distillation are performed. + +# 3.3. Training Recipe and Multi-Level Distillation + +To improve the generation quality of our efficient diffusion model, we propose a series of training techniques. + +Flow-based Training and Inference. Rectified Flows (RFs) [44, 47] define the forward process as straight paths connecting the data distribution to a standard normal distribution, i.e., + +$$ +x _ {t} = \left(1 - \sigma_ {t}\right) x _ {0} + \sigma_ {t} \epsilon , \tag {1} +$$ + +where $x_0$ is the clean (latent) image, $t$ is the timestep, $\sigma_t$ is a timestep-dependent factor, and $\epsilon$ is random noise sampled from $\mathcal{N}(0,I)$ . The denoising UNet is formulated to predict a velocity field with the objective as + +$$ +\mathcal {L} _ {\text {t a s k}} = \mathbb {E} _ {\epsilon \sim \mathcal {N} (0, I), t} \left[ | | (\epsilon - x _ {0}) - v _ {\theta} (x _ {t}, t) | | _ {2} ^ {2} \right], \tag {2} +$$ + +where $v_{\theta}(x_t, t)$ is the predicted velocity from UNet parameterized by $\theta$ . To further enhance training stability, we apply logit-normal sampling [18] for the timestep during training, which assigns more samples to the intermediate steps. In the inference stage, we use the Flow-Euler sampler [19], which predicts the next sample based on the velocity, i.e., + +$$ +x _ {t - 1} = x _ {t} + \left(\sigma_ {t - 1} - \sigma_ {t}\right) \cdot v _ {\theta} \left(x _ {t}, t\right). \tag {3} +$$ + +To achieve a lower signal-to-noise ratio on high-resolution (i.e., $1024^2$ px) images, we apply a timestep shift similar to SD3 [18] to adjust the scheduling factor $\sigma_t$ during both training and inference. + +Multi-Level Knowledge Distillation. To improve the generation quality for compact models, one common practice for previous works is applying knowledge distillation to mimic the prediction of scaled-up teacher models [32]. Benefiting from the aligned flow matching objective and (AE) latent space, powerful SD3.5-Large model [4] can be used as the teacher for output distillation. However, we still face challenges due to 1) the heterogeneous architecture between the U-Net and DiT, 2) the scale difference between the distillation loss and task loss, and 3) varying prediction difficulty across different timesteps. To tackle these, we propose a novel multi-level distillation loss and apply timestep-aware scaling to stabilize and accelerate the convergence of distillation. The overview of our knowledge + +distillation scheme is shown in Fig. 5 and detailed techniques are elaborated as follows. + +Aside from the task loss defined in Eq. 2, the major objective for knowledge distillation is to supervise our model $\theta$ directly with the output of teacher model $\theta_{T}$ , which can be indicated as + +$$ +\mathcal {L} _ {\mathrm {k d}} = \mathbb {E} \left[ \left| \left| v _ {\theta_ {T}} \left(x _ {t}, t\right) - v _ {\theta} \left(x _ {t}, t\right) \right| \right| _ {2} ^ {2} \right]. \tag {4} +$$ + +Given the capacity gap between the teacher and our model, applying output-level supervision alone leads to instability and slow convergence. Therefore, we further incorporate a cross-architecture feature-level distillation loss as + +$$ +\mathcal {L} _ {\text {f e a t k d}} = \mathbb {E} \left[ \sum_ {(l _ {T}, l)} | | f _ {\theta_ {T}} ^ {l _ {T}} (x _ {t}, t) - \psi \left(f _ {\theta} ^ {l} \left(x _ {t}, t\right)\right) | | _ {2} ^ {2} \right], \tag {5} +$$ + +where $f_{\theta_T}^{l_T}(\cdot)$ and $f^l (\cdot)$ indicate the feature output from the $l_{T}$ -th layer and $l$ -th layer in teacher model and student model, respectively. Different from previous work [32, 46], we consider cross-architecture distillation from a DiT to UNet. Since the richest information in transformers sits around the last layer, we set the distillation target to this layer in both models, and use a lightweight trainable projector $\psi (\cdot)$ with only two Conv layers to map the student feature to match the dimension of teacher feature. The proposed feature-level distillation loss provides additional supervision to the student model, leading to faster alignment to the generation quality of the teacher model. + +Timestep-Aware Scaling. Weighting multiple objectives has been a major challenge in knowledge distillation, especially in diffusion models. The overall training objectives from previous works [32, 46, 67] are simple linear combination of multiple loss term, i.e., + +$$ +\mathcal {L} = \mathcal {L} _ {\text {t a s k}} + \lambda_ {1} \mathcal {L} _ {\mathrm {k d}} + \lambda_ {2} \mathcal {L} _ {\text {f e a t k d}}, \tag {6} +$$ + +where the weighting coefficient $\lambda_{1}$ and $\lambda_{2}$ are empirically set to constant. However, this baseline setting fails to consider the prediction difficulty in various time steps. We investigate the distribution of empirical risk magnitude of $\mathcal{L}_{\mathrm{task}}$ and $\mathcal{L}_{\mathrm{kd}}$ across different timestep $t$ during model training. Fig. 6 illustrates that, in intermediate steps, prediction difficulty are lower compared to $t$ closer to 0 or 1. + +Building on this important observation, we propose a timestep-aware scaling of the objective to close the gap in loss magnitude across different values of $t$ and to account for prediction difficulties at each timestep, as follows: + +$$ +\mathcal {S} \left(\mathcal {L} _ {\text {t a s k}}, \mathcal {L} _ {\mathrm {k d}}\right) = \mathbb {E} _ {t} \left[ \lambda (t) \cdot \mathcal {L} _ {\text {t a s k}} ^ {t} + (1 - \lambda (t)) \frac {\left| \mathcal {L} _ {\text {t a s k}} ^ {t} \right|}{\left| \mathcal {L} _ {\mathrm {k d}} ^ {t} \right|} \cdot \mathcal {L} _ {\mathrm {k d}} ^ {t} \right], \tag {7} +$$ + +where $\lambda (t)$ is the normalized standard (location 0, scale 1) logit-norm density function and $|\cdot |$ indicates the magnitude. In $S$ , we first ensure the same scale between task loss and distillation loss across different $t$ , then apply more teacher supervision where prediction difficulty is higher (i.e., $t$ closer to 0 or 1), and more real data supervision where prediction difficulty is lower (i.e., intermediate timesteps). The proposed scheme considers the variation of timestep $t$ and helps accelerate the distillation training. The final multi-level distillation objective $\mathcal{L}_{\mathrm{md}}$ is defined as + +$$ +\mathcal {L} _ {\mathrm {m d}} = \mathcal {S} \left(\mathcal {L} _ {\text {t a s k}}, \mathcal {L} _ {\mathrm {k d}}\right) + \mathcal {S} \left(\mathcal {L} _ {\text {t a s k}}, \mathcal {L} _ {\text {f e a t k d}}\right). \tag {8} +$$ + +# 3.4. Step Distillation + +We take one step further to enhance the sampling efficiency of our model following a distribution-matching-based step distillation scheme. Following Latent Adversarial Diffusion Distillation (LADD) [60], we use a diffusion-GAN hybrid structure to distill our model into fewer steps with the optimization objective as + +$$ +\begin{array}{l} \min _ {D _ {\theta_ {T}}} \max _ {G _ {\theta}} \mathbb {E} \left[ \left[ \log \left(D _ {\theta_ {T}} \left(x _ {t - 1}, t\right)\right) \right] \right. \tag {9} \\ \left. + \left[ \log \left(1 - D _ {\theta_ {T}} \left(x _ {t - 1} ^ {\prime}, t\right)\right) \right] - \mathcal {S} \left(\mathcal {L} _ {\mathrm {t a s k}}, \mathcal {L} _ {\mathrm {k d}}\right) \right], \\ \end{array} +$$ + +where $D_{\theta_T}$ is the discriminator model partially initialized with pretrained fewer-step teacher model $\theta_T$ (SD3.5-Large-Turbo [5]). The large-scale teacher model are only used as the feature extractor and are frozen during distillation. We only train a few linear layers in the discriminator after feature extraction. We sample $x_{t-1} \sim q(x_{t-1}|x_0)$ and $x_{t-1}' \sim q(x_{t-1}|x_0')$ , where $x_0'$ is the prediction of our denoising generator $G_\theta(x_t, t)$ as our student model, and $q(x)$ is the forward process of diffusion model defined in Eq. 1. The objective consists of an adversarial loss to match noisy samples at time step $t-1$ and the output-level distillation loss $S(\mathcal{L}_{\mathrm{task}}, \mathcal{L}_{\mathrm{kd}})$ after applying timestep-aware scaling. The proposed step distillation, visualized in Fig. 7, can be interpreted as training a diffusion model with adversarial refinement and knowledge distillation, where teacher guidance serves as an additional inductive bias. This advanced step distillation empowers our compact model for high-quality generation, with only a few denoising steps. + +# 4. Experiments + +Model Details. Our T2I pipeline consists of the efficient UNet (Sec. 3.1) and the efficient encoder-decoder model (Sec. 3.2). To obtain text embeddings from the input prompt, we leverage multiple text encoders, namely lightweight CLIP-L [56], CLIP-G [56], and the large Gemma-2-2b language model [68]. We follow SD3 [18] strategy to combine these three text-encoders into a unified rich textual embedding. To enable classifier-free guidance [27], we employ these embeddings with an individual drop-out probability such that we can use an arbitrary subset of the encoders during inference. This allows us to deploy one or more encoders based on the resource constraints. + +Training Recipe. Similar as prior work [30], we use a multi-stage strategy to train our UNet model from scratch. First, we pre-train the model using the ImageNet-1K [16] at 256 resolution as described in Sec. 3.1. Second, we fine-tune this model in a progressive manner from $256 \rightarrow 512 \rightarrow 1024$ resolutions. Third, we employ knowledge distillation with our timestep-aware scaling (Sec. 3.3) to improve the finer details in our models using a much larger teacher model (SD3.5-Large [4]) and all three text-encoders. Before KD, we initialize the student model by pre-training on ImageNet and T2I datasets, a cost-effective approach that helps the model learn pixel distributions. Finally, we obtain a few-step model through step distillation using SD3.5-Large-Turbo [5] model as the teacher. We optimize the rectified-flow [18] objective using AdamW optimizer [17] to train our UNet backbone. + +Hyper-parameters. We sample the timesteps using the logit-normal distribution with $(0,1)$ as the location and scale parameters. We use a time shift of 3 for both training and inference. See the Supplemental Materials for details. + +# 4.1. Evaluation + +Quantitative Benchmarks. We use GenEval [21] and DPG-Bench [29] benchmarks to evaluate the text-to-image alignment of our model on short and long prompts, respectively. We report CLIP score on a 6K subset of MS-COCO validation data [43]. In addition, to measure the aesthetic quality of our model, we compute the Image Reward [75] score on selected PixArt prompts [12]. Tab. 3 lists our performance alongside existing state-of-the-art T2I baselines. We provide additional details in Supplemental Materials. We highlight the salient observations below: + +- Our 0.38B parameter model achieves even better performance than significantly larger models such as SDXL (2.6B), Playground (2.6B), and IF-XL (5.5B). +- KD non-trivially improves the prompt following ability of the base model as illustrated by an absolute five-point increase in DPG-Bench and GenEval scores. +- In terms of aesthetic performance, our model has similar Image Reward scores as Playground models [37, 38]. + +![](images/94c542be18b250489df00edc63c1b6e6da90c6efc8471432510b82eadc73552b.jpg) +Figure 8. Human Evaluation. We conduct a user study to compare images generated by our model against baselines on three attributes: aesthetic quality, text-image alignment, and realistic generations. Our model surpasses the quality of SDXL and SD3 models, while performing competitively against the teacher SD3.5-Large model. + +![](images/5be006998992a29e2e6b32849107feb33f38b96359160e05550d226632e95a86.jpg) + +![](images/e5b8927c2ba31f4da15cb262c040939b608df725b8aa28c709505bfae9842108.jpg) + +Table 3. Evaluation on Quantitative Benchmarks. We list the scores on GenEval, DPG-Bench, CLIP score on COCO, and Image Reward on aesthetic prompts. We report the parameters for the UNet/DiT backbone in the Param column. Throughput (samples/s) is measured on a single 80GB A100 GPU using the largest batch size supported for each model in a practical scenario to generate $1024^2$ px images. Here our sampling step is set to be 28. + +
ModelParamThroughputGenEval ↑DPG ↑CLIP ↑Image Reward ↑
PixArt-α [12]0.6B0.420.4871.10.3161.15
PixArt-Σ [13]0.6B0.460.5380.50.3171.13
SD-1.5 [1]0.9B-0.4363.20.2870.19
SD-2.1 [2]0.9B-0.5064.20.2810.29
KOALA [35]1.0B0.660.5274.30.3171.05
MicroDiT [62]1.2B-0.5275.10.3181.25
Sana [74]1.6B1.000.6684.80.3271.25
LUMINA-Next [87]2.0B0.060.4674.60.3090.88
SDXL [53]2.6B0.180.5574.70.3010.99
Playgroundv2 [37]2.6B0.180.5974.50.3171.25
Playgroundv2.5 [38]2.6B0.180.5675.50.3191.34
IF-XL [15]5.5B0.060.6175.60.3110.65
Ours w/o KD0.38B1.040.6176.30.3211.20
SnapGen (ours)0.38B1.040.6681.10.3321.32
+ +Qualitative Comparison. To visually evaluate the image-text alignment and aesthetics, we compare the generated images from different T2I models in Fig. 1. We observe that many existing models fail to fully capture the full prompt and miss important elements. Further, human generations often result in smoothened-out faces, leading to the loss of details. In contrast, our model generates much more photorealistic images with better image-text alignment. + +Human Evaluation. For a thorough comparison between baselines, we perform a user study with the widely used Parti prompts [81]. We use SDXL, SD3-M, and SD3.5-Large models as the baselines and generate images using this prompt set. We ask the users to select images with better attributes between the baselines and our model. These attributes include image-text alignment, aesthetic quality, and realistic images. Fig. 8 shows that our model convincingly outperforms SDXL on all three attributes. Our model beats SD3 on image-text alignment and realistic generations, with a tie on aesthetic quality. Compared to the teacher (SD3.5-Large), we lag the teacher a bit behind on the image-text alignment, yet our model still has better realistic generations, and yields a toss-up on aesthetic quality. With this study, we can conclude our efficient T2I pipeline + +achieves generation quality that is quite comparable to the SD3.5-Large teacher that has 8.1B parameters. + +Few-Step Generation. After step-distillation (Sec. 3.4), our model can generate high-quality images within a few steps. Fig. 9 compares our model's performance before and after step-distillation, with respective GenEval scores. The results demonstrate that our model, after step distillation, achieves comparable performance to the baseline model with 28 steps, even when using only 4 or 8 steps. While the few-step generation shows slight qualitative degradation compared to the 28-step baseline, it still outperforms most existing T2I models with significantly more inference steps, such as SDXL (50 steps) and PixArt- $\alpha$ (100 steps). + +![](images/c1472c3305f61d9c032a6fabeaedf4a68aaac446d3c2036d3c05de1678461366.jpg) +Figure 9. Performance comparison of few-step generation for our model before (top) and after step distillation (bottom). + +# 5. Conclusion + +In this work, we propose a novel and efficient T2I model for high-resolution generation on mobile phones. We systematically detail the process to obtain a tiny 379M parameter UNet architecture along with an efficient latent decoder. We devise a novel training method consisting of multi-stage pre-training followed by knowledge distillation from a large teacher and adversarial step distillation. With these, we achieve an extremely efficient T2I model that comprehensively outperforms many existing multi-billion parameter models such as SDXL, Lumina-Next, and Playgroundv2. + +# References + +[1] Stability AI. Stable diffusion 1.5. https://huggingface.co/stable-diffusion-v1-5/stable-diffusion-v1-5, 2022.8 +[2] Stability AI. Stable diffusion 2.1. https://huggingface.co/stabilityai/stable-diffusion-2-1, 2022.8 +[3] Stability AI. Stable diffusion 3.5. https://github.com/Stability-AI/sd3.5, 2024.2 +[4] Stability AI. Stable diffusion 3.5 large. https://huggingface.co/stabilityai/stable-diffusion-3.5-large, 2024.6,7 +[5] Stability AI. Stable diffusion 3.5 large turbo. https://huggingface.co/stabilityai/stable-diffusion-3.5-large-turbo, 2024. 2, 7 +[6] Vladimir Arkhipkin, Viacheslav Vasilev, Andrei Filatov, Igor Pavlov, Julia Agafonova, Nikolai Gerasimenko, Anna Averchenkova, Evelina Mironova, Anton Bukashkin, Konstantin Kulikov, et al. Kandinsky 3: Text-to-image synthesis for multifunctional generative framework. arXiv preprint arXiv:2410.21061, 2024. 3 +[7] Shakhmatov Arseni, Razzhigaev Anton, Nikolich Aleksandr, Arkhipkin Vladimir, Pavlov Igor, Kuznetsov Andrey, and Denis Dimitrov. Kandinsky 2.1, 2023. 3 +[8] Fan Bao, Shen Nie, Kaiwen Xue, Yue Cao, Chongxuan Li, Hang Su, and Jun Zhu. All are worth words: A vit backbone for diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 22669-22679, 2023. 5 +[9] Han Cai, Junyan Li, Muyan Hu, Chuang Gan, and Song Han. Efficientvit: Lightweight multi-scale attention for high-resolution dense prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17302-17313, 2023. 3 +[10] Thibault Castells, Hyoung-Kyu Song, Bo-Kyeong Kim, and Shinkook Choi. LD-Pruner: Efficient Pruning of Latent Diffusion Models using Task-Agnostic Insights. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 821–830, 2024. 2 +[11] Thibault Castells, Hyoung-Kyu Song, Tairen Piao, Shinkook Choi, Bo-Kyeong Kim, Hanyoung Yim, Changgwun Lee, Jae Gon Kim, and Tae-Ho Kim. EdgeFusion: On-Device Text-to-Image Generation. arXiv preprint arXiv:2404.11925, 2024.3 +[12] Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, et al. Pixart-α: Fast training of diffusion transformer for photorealistic text-to-image synthesis. arXiv preprint arXiv:2310.00426, 2023. 2, 3, 7, 8 +[13] Junsong Chen, Chongjian Ge, Enze Xie, Yue Wu, Lewei Yao, Xiaozhe Ren, Zhongdao Wang, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart- $\cdot$ sigma: Weak-to-strong training of diffusion transformer for 4k text-to-image generation. arXiv preprint arXiv:2403.04692, 2024. 2, 3, 8 +[14] Tri Dao and Albert Gu. Transformers are ssms: Generalized models and efficient algorithms through structured state space duality. arXiv preprint arXiv:2405.21060, 2024. 3 + +[15] DeepFloyd. Deepfloyd. https://github.com/deep-floyd/IF, 2023. 2, 3, 8 +[16] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 2, 3, 7 +[17] P Kingma Diederik. Adam: A method for stochastic optimization. (No Title), 2014. 7 +[18] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In *Forty-first International Conference on Machine Learning*, 2024. 2, 5, 6, 7 +[19] Leonhard Euler. Institutionum calculi integralis. imp. Acad. imp. Sa'ent., 1768. 6 +[20] Peng Gao, Le Zhuo, Ziyi Lin, Chris Liu, Junsong Chen, Ruoyi Du, Enze Xie, Xu Luo, Longtian Qiu, Yuhang Zhang, et al. Lumina-T2X: Transforming Text into Any Modality, Resolution, and Duration via Flow-based Large Diffusion Transformers. arXiv preprint arXiv:2405.05945, 2024. 3 +[21] Dhruba Ghosh, Hannaneh Hajishirzi, and Ludwig Schmidt. Geneval: An object-focused framework for evaluating text-to-image alignment. Advances in Neural Information Processing Systems, 36, 2024. 7 +[22] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 3, 5 +[23] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 3 +[24] Alex Henry, Prudhvi Raj Dachapally, Shubham Pawar, and Yuxuan Chen. Query-key normalization for transformers. arXiv preprint arXiv:2010.04245, 2020. 5 +[25] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 3 +[26] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 5 +[27] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 7 +[28] Andrew G Howard. Mobilenets: Efficient convolutional neural networks for mobile vision applications. arXiv preprint arXiv:1704.04861, 2017. 3 +[29] Xiwei Hu, Rui Wang, Yixiao Fang, Bin Fu, Pei Cheng, and Gang Yu. Ella: Equip diffusion models with llm for enhanced semantic alignment. arXiv preprint arXiv:2403.05135, 2024. 7 +[30] Anil Kag, Huseyin Coskun, Jierun Chen, Junli Cao, Willi Menapace, Aliaksandr Siarohin, Sergey Tulyakov, and Jian Ren. Ascan: Asymmetric convolution-attention networks for efficient recognition and generation. arXiv preprint arXiv:2411.04967, 2024. 3, 7 + +[31] Angelos Katharopoulos, Apoorv Vyas, Nikolaos Pappas, and François Fleuret. Transformers are rnns: Fast autoregressive transformers with linear attention. In International conference on machine learning, pages 5156-5165. PMLR, 2020. 3 +[32] Bo-Kyeong Kim, Hyoung-Kyu Song, Thibault Castells, and Shinkook Choi. Bk-sdm: Architecturally Compressed Stable Diffusion for Efficient Text-to-Image Generation. In Workshop on Efficient Systems for Foundation Models@ ICML2023, 2023. 2, 3, 6 +[33] Diederik P Kingma. Auto-encoding variational bayes. 2nd International Conference on Learning Representations, 2014. 5 +[34] Tuomas Kynkänniemi, Miika Aittala, Tero Karras, Samuli Laine, Timo Aila, and Jaakko Lehtinen. Applying guidance in a limited interval improves sample and distribution quality in diffusion models. arXiv preprint arXiv:2404.07724, 2024.5 +[35] Youngwan Lee, Kwanyong Park, Yoorhim Cho, Yong-Ju Lee, and Sung Ju Hwang. Koala: Empirical lessons toward memory-efficient and fast diffusion models for text-to-image synthesis, 2023. 8 +[36] Daiqing Li, Aleks Kamko, Ali Sabet, Ehsan Akhgari, Linmiao Xu, and Suhail Doshi. Playground v1, . 3 +[37] Daiqing Li, Aleks Kamko, Ali Sabet, Ehsan Akhgari, Linmiao Xu, and Suhail Doshi. Playground v2, . 7, 8 +[38] Daiqing Li, Aleks Kamko, Ehsan Akhgari, Ali Sabet, Linmiao Xu, and Suhail Doshi. Playground V2. 5: Three Insights towards Enhancing Aesthetic Quality in Text-to-Image Generation. arXiv preprint arXiv:2402.17245, 2024. 3, 7, 8 +[39] Hao Li, Yang Zou, Ying Wang, Orchid Majumder, Yusheng Xie, R Manmatha, Ashwin Swaminathan, Zhuowen Tu, Stefano Ermon, and Stefano Soatto. On the scalability of diffusion-based text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9400-9409, 2024. 3 +[40] Yanyu Li, Huan Wang, Qing Jin, Ju Hu, Pavlo Chemerys, Yun Fu, Yanzhi Wang, Sergey Tulyakov, and Jian Ren. Snapfusion: Text-to-Image Diffusion Model on Mobile Devices within Two Seconds. Advances in Neural Information Processing Systems, 36, 2024. 2, 3, 5 +[41] Zhimin Li, Jianwei Zhang, Qin Lin, Jiangfeng Xiong, Yanxin Long, Xinchi Deng, Yingfang Zhang, Xingchao Liu, Minbin Huang, Zedong Xiao, et al. Hunyuan-DiT: A Powerful Multi-Resolution Diffusion Transformer with Fine-Grained Chinese Understanding. arXiv preprint arXiv:2405.08748, 2024. 3 +[42] Shanchuan Lin, Anran Wang, and Xiao Yang. SDXL-Lightning: Progressive Adversarial Diffusion Distillation, 2024. 2 +[43] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 5, 7 + +[44] Yaron Lipman, Ricky TQ Chen, Heli Ben-Hamu, Maximilian Nickel, and Matt Le. Flow matching for generative modeling. arXiv preprint arXiv:2210.02747, 2022. 2, 6 +[45] Bingchen Liu, Ehsan Akhgari, Alexander Visheratin, Aleks Kamko, Linmiao Xu, Shivam Shrirao, Joao Souza, Suhail Doshi, and Daiqing Li. Playground v3: Improving text-to-image alignment with deep-fusion large language models. arXiv preprint arXiv:2409.10695, 2024. 3 +[46] Songhua Liu, Weihao Yu, Zhenxiong Tan, and Xinchao Wang. Linfusion: 1gpu, 1 minute, 16k image. arXiv preprint arXiv:2409.02097, 2024. 2, 3, 6 +[47] Xingchao Liu, Chengyue Gong, and Qiang Liu. Flow straight and fast: Learning to generate and transfer data with rectified flow. arXiv preprint arXiv:2209.03003, 2022. 2, 6 +[48] Xian Liu, Jian Ren, Aliaksandr Siarohin, Ivan Skorokhodov, Yanyu Li, Dahua Lin, Xihui Liu, Ziwei Liu, and Sergey Tulyakov. Hyperhuman: Hyper-realistic human generation with latent structural diffusion. arXiv preprint arXiv:2310.08579, 2023. 2, 3 +[49] Nanye Ma, Mark Goldstein, Michael S Albergo, Nicholas M Boffi, Eric Vanden-Eijnden, and Saining Xie. Sit: Exploring flow and diffusion-based generative models with scalable interpolant transformers. arXiv preprint arXiv:2401.08740, 2024. 2, 5 +[50] Willi Menapace, Aliaksandr Siarohin, Ivan Skorokhodov, Ekaterina Deyneka, Tsai-Shien Chen, Anil Kag, Yuwei Fang, Aleksei Stoliar, Elisa Ricci, Jian Ren, et al. Snap video: Scaled spatiotemporal transformers for text-to-video synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7038-7048, 2024. 2 +[51] OpenAI. 2 +[52] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4195-4205, 2023. 5 +[53] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 2, 3, 5, 8 +[54] Adam Polyak, Amit Zohar, Andrew Brown, Andros Tjandra, Animesh Sinha, Ann Lee, Apoorv Vyas, Bowen Shi, Chih-Yao Ma, Ching-Yao Chuang, et al. Movie gen: A cast of media foundation models. arXiv preprint arXiv:2410.13720, 2024. 2 +[55] Danfeng Qin, Chas Leichner, Manolis Delakis, Marco Fornoni, Shixin Luo, Fan Yang, Weijun Wang, Colby Banbury, Chengxi Ye, Berkin Akin, et al. Mobilenetv4-universal models for the mobile ecosystem. arXiv preprint arXiv:2404.10518, 2024.3 +[56] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 7 + +[57] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In Proceedings of the 38th International Conference on Machine Learning, pages 8821-8831. PMLR, 2021. 2 +[58] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 3, 5 +[59] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S. Sara Mahdavi, Rapha Gontijo Lopes, Tim Salimans, Jonathan Ho, David J Fleet, and Mohammad Norouzi. Photorealistic text-to-image diffusion models with deep language understanding, 2022. 2 +[60] Axel Sauer, Frederic Boesel, Tim Dockhorn, Andreas Blattmann, Patrick Esser, and Robin Rombach. Fast high-resolution image synthesis with latent adversarial diffusion distillation. arXiv preprint arXiv:2403.12015, 2024. 3, 7 +[61] Axel Sauer, Dominik Lorenz, Andreas Blattmann, and Robin Rombach. Adversarial diffusion distillation. In European Conference on Computer Vision, pages 87-103. Springer, 2024. 3 +[62] Vikash Sehwag, Xianghao Kong, Jingtao Li, Michael Spranger, and Lingjuan Lyu. Stretching each dollar: Diffusion training from scratch on a micro-budget. arXiv preprint arXiv:2407.15811, 2024. 8 +[63] Noam Shazeer. Fast transformer decoding: One write-head is all you need. arXiv preprint arXiv:1911.02150, 2019. 4 +[64] Shelly Sheynin, Adam Polyak, Uriel Singer, Yuval Kirstain, Amit Zohar, Oron Ashual, Devi Parikh, and Yaniv Taigman. Emu edit: Precise image editing via recognition and generation tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8871-8879, 2024. 2 +[65] Yuda Song, Zehao Sun, and Xuanwu Yin. Sdxs: Realtime one-step latent diffusion models with image conditions. arXiv preprint arXiv:2403.16627, 2024. 2 +[66] Jianlin Su, Murtadha Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu. Roformer: Enhanced transformer with rotary position embedding. Neurocomputing, 568:127063, 2024. 5 +[67] Yang Sui, Yanyu Li, Anil Kag, Yerlan Idelbayev, Junli Cao, Ju Hu, Dhritiman Sagar, Bo Yuan, Sergey Tulyakov, and Jian Ren. Bitsfusion: 1.99 bits weight quantization of diffusion model. arXiv preprint arXiv:2406.04333, 2024. 2, 3, 6 +[68] Gemma Team, Morgane Riviere, Shreya Pathak, Pier Giuseppe Sessa, Cassidy Hardin, Surya Bhupatiraju, Léonard Hussenot, Thomas Mesnard, Bobak Shahriari, Alexandre Ramé, et al. Gemma 2: Improving open language models at a practical size. arXiv preprint arXiv:2408.00118, 2024. 7 +[69] Kcolors Team. Kcolors: Effective Training of Diffusion Model for Photorealistic Text-to-Image Synthesis. arXiv preprint, 2024. 3 +[70] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 3 + +[71] Zhendong Wang, Huangjie Zheng, Pengcheng He, Weizhu Chen, and Mingyuan Zhou. Diffusion-gan: Training gans with diffusion. arXiv preprint arXiv:2206.02262, 2022. 3 +[72] Shitao Xiao, Yueze Wang, Junjie Zhou, Huaying Yuan, Xingrun Xing, Ruiran Yan, Shuting Wang, Tiejun Huang, and Zheng Liu. Omnigen: Unified image generation. arXiv preprint arXiv:2409.11340, 2024. 2 +[73] Zhisheng Xiao, Karsten Kreis, and Arash Vahdat. Tackling the generative learning trilemma with denoising diffusion gans. arXiv preprint arXiv:2112.07804, 2021. 3 +[74] Enze Xie, Junsong Chen, Junyu Chen, Han Cai, Yujun Lin, Zhekai Zhang, Muyang Li, Yao Lu, and Song Han. Sana: Efficient high-resolution image synthesis with linear diffusion transformers. arXiv preprint arXiv:2410.10629, 2024. 2, 3, 8 +[75] Jiazheng Xu, Xiao Liu, Yuchen Wu, Yuxuan Tong, Qinkai Li, Ming Ding, Jie Tang, and Yuxiao Dong. Imagereward: Learning and evaluating human preferences for text-to-image generation. Advances in Neural Information Processing Systems, 36, 2024. 7 +[76] Yanwu Xu, Mingming Gong, Shaoan Xie, Wei Wei, Matthias Grundmann, Kayhan Batmanghelich, and Tingbo Hou. Semi-implicit denoising diffusion models (siddms). Advances in neural information processing systems, 36:17383, 2023. 3 +[77] Yanwu Xu, Yang Zhao, Zhisheng Xiao, and Tingbo Hou. Ufogen: You forward once large scale text-to-image generation via diffusion gans. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8196-8206, 2024. 2, 3 +[78] Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, et al. Cogvideox: Text-to-video diffusion models with an expert transformer. arXiv preprint arXiv:2408.06072, 2024. 2 +[79] Tianwei Yin, Michael Gharbi, Taesung Park, Richard Zhang, Eli Shechtman, Fredo Durand, and William T Freeman. Improved distribution matching distillation for fast image synthesis. arXiv preprint arXiv:2405.14867, 2024. 3 +[80] Tianwei Yin, Michael Gharbi, Richard Zhang, Eli Shechtman, Fredo Durand, William T Freeman, and Taesung Park. One-step diffusion with distribution matching distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6613-6623, 2024. 2 +[81] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, et al. Scaling autoregressive models for content-rich text-to-image generation. Transactions on Machine Learning Research. 8 +[82] Biao Zhang and Rico Sennrich. Root mean square layer normalization. Advances in Neural Information Processing Systems, 32, 2019. 5 +[83] Dingkun Zhang, Sijia Li, Chen Chen, Qingsong Xie, and Haonan Lu. Laptop-Diff: Layer Pruning and Normalized Distillation for Compressing Diffusion Models. arXiv preprint arXiv:2404.11098, 2024. 2 + +[84] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 5 +[85] Zhixing Zhang, Ligong Han, Arnab Ghosh, Dimitris N Metaxas, and Jian Ren. Sine: Single image editing with text-to-image diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6027-6037, 2023. 2 +[86] Yang Zhao, Yanwu Xu, Zhisheng Xiao, and Tingbo Hou. Mobilediffusion: Subsecond text-to-image generation on mobile devices. arXiv preprint arXiv:2311.16567, 2023. 2, 3, 4, 5 +[87] Le Zhuo, Ruoyi Du, Han Xiao, Yangguang Li, Dongyang Liu, Rongjie Huang, Wenze Liu, Lirui Zhao, Fu-Yun Wang, Zhanyu Ma, et al. Lumina-last: Making lumina-t2x stronger and faster with next-dit. arXiv preprint arXiv:2406.18583, 2024. 2, 8 \ No newline at end of file diff --git a/2025/SnapGen_ Taming High-Resolution Text-to-Image Models for Mobile Devices with Efficient Architectures and Training/images.zip b/2025/SnapGen_ Taming High-Resolution Text-to-Image Models for Mobile Devices with Efficient Architectures and Training/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..60bbe866fe9ad6c3dee246715bfddfbcdc733178 --- /dev/null +++ b/2025/SnapGen_ Taming High-Resolution Text-to-Image Models for Mobile Devices with Efficient Architectures and Training/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dbe12485babda33c40bf9f0d8c45322ff268cac7229bac9879958fdf22225055 +size 707412 diff --git a/2025/SnapGen_ Taming High-Resolution Text-to-Image Models for Mobile Devices with Efficient Architectures and Training/layout.json b/2025/SnapGen_ Taming High-Resolution Text-to-Image Models for Mobile Devices with Efficient Architectures and Training/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..452241b03be96ba36f930340331ecf754294439c --- /dev/null +++ b/2025/SnapGen_ Taming High-Resolution Text-to-Image Models for Mobile Devices with Efficient Architectures and Training/layout.json @@ -0,0 +1,10745 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 67, + 103, + 543, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 103, + 543, + 140 + ], + "spans": [ + { + "bbox": [ + 67, + 103, + 543, + 140 + ], + "type": "text", + "content": "SnapGen: Taming High-Resolution Text-to-Image Models for Mobile Devices with Efficient Architectures and Training" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 109, + 160, + 500, + 219 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 109, + 160, + 500, + 219 + ], + "spans": [ + { + "bbox": [ + 109, + 160, + 500, + 219 + ], + "type": "text", + "content": "Jierun Chen" + }, + { + "bbox": [ + 109, + 160, + 500, + 219 + ], + "type": "inline_equation", + "content": "^{1,2,*}" + }, + { + "bbox": [ + 109, + 160, + 500, + 219 + ], + "type": "text", + "content": " Dongting Hu" + }, + { + "bbox": [ + 109, + 160, + 500, + 219 + ], + "type": "inline_equation", + "content": "^{1,3,*}" + }, + { + "bbox": [ + 109, + 160, + 500, + 219 + ], + "type": "text", + "content": " Xijie Huang" + }, + { + "bbox": [ + 109, + 160, + 500, + 219 + ], + "type": "inline_equation", + "content": "^{1,2,*}" + }, + { + "bbox": [ + 109, + 160, + 500, + 219 + ], + "type": "text", + "content": " Huseyin Coskun" + }, + { + "bbox": [ + 109, + 160, + 500, + 219 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 109, + 160, + 500, + 219 + ], + "type": "text", + "content": " Arpit Sahni" + }, + { + "bbox": [ + 109, + 160, + 500, + 219 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 109, + 160, + 500, + 219 + ], + "type": "text", + "content": " Aarush Gupta" + }, + { + "bbox": [ + 109, + 160, + 500, + 219 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 109, + 160, + 500, + 219 + ], + "type": "text", + "content": " Anujraaj Goyal" + }, + { + "bbox": [ + 109, + 160, + 500, + 219 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 109, + 160, + 500, + 219 + ], + "type": "text", + "content": " Dishani Lahiri" + }, + { + "bbox": [ + 109, + 160, + 500, + 219 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 109, + 160, + 500, + 219 + ], + "type": "text", + "content": " Rajesh Singh" + }, + { + "bbox": [ + 109, + 160, + 500, + 219 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 109, + 160, + 500, + 219 + ], + "type": "text", + "content": " Yerlan Idelbayev" + }, + { + "bbox": [ + 109, + 160, + 500, + 219 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 109, + 160, + 500, + 219 + ], + "type": "text", + "content": " Junli Cao" + }, + { + "bbox": [ + 109, + 160, + 500, + 219 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 109, + 160, + 500, + 219 + ], + "type": "text", + "content": " Yanyu Li" + }, + { + "bbox": [ + 109, + 160, + 500, + 219 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 109, + 160, + 500, + 219 + ], + "type": "text", + "content": " Kwang-Ting Cheng" + }, + { + "bbox": [ + 109, + 160, + 500, + 219 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 109, + 160, + 500, + 219 + ], + "type": "text", + "content": " S.-H. Gary Chan" + }, + { + "bbox": [ + 109, + 160, + 500, + 219 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 109, + 160, + 500, + 219 + ], + "type": "text", + "content": " Mingming Gong" + }, + { + "bbox": [ + 109, + 160, + 500, + 219 + ], + "type": "inline_equation", + "content": "^{3,4}" + }, + { + "bbox": [ + 109, + 160, + 500, + 219 + ], + "type": "text", + "content": " Sergey Tulyakov" + }, + { + "bbox": [ + 109, + 160, + 500, + 219 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 109, + 160, + 500, + 219 + ], + "type": "text", + "content": " Anil Kag" + }, + { + "bbox": [ + 109, + 160, + 500, + 219 + ], + "type": "inline_equation", + "content": "^{1,\\dagger}" + }, + { + "bbox": [ + 109, + 160, + 500, + 219 + ], + "type": "text", + "content": " Yanwu Xu" + }, + { + "bbox": [ + 109, + 160, + 500, + 219 + ], + "type": "inline_equation", + "content": "^{1,\\dagger}" + }, + { + "bbox": [ + 109, + 160, + 500, + 219 + ], + "type": "text", + "content": " Jian Ren" + }, + { + "bbox": [ + 109, + 160, + 500, + 219 + ], + "type": "inline_equation", + "content": "^{1,\\dagger}" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 230, + 478, + 244 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 230, + 478, + 244 + ], + "spans": [ + { + "bbox": [ + 132, + 230, + 478, + 244 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 132, + 230, + 478, + 244 + ], + "type": "text", + "content": "Snap Inc. " + }, + { + "bbox": [ + 132, + 230, + 478, + 244 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 132, + 230, + 478, + 244 + ], + "type": "text", + "content": "HKUST " + }, + { + "bbox": [ + 132, + 230, + 478, + 244 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 132, + 230, + 478, + 244 + ], + "type": "text", + "content": "The University of Melbourne " + }, + { + "bbox": [ + 132, + 230, + 478, + 244 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 132, + 230, + 478, + 244 + ], + "type": "text", + "content": "MBZUAI" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 209, + 246, + 402, + 260 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 209, + 246, + 402, + 260 + ], + "spans": [ + { + "bbox": [ + 209, + 246, + 402, + 260 + ], + "type": "text", + "content": "* Equal contribution † Equal advising" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 175, + 261, + 432, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 175, + 261, + 432, + 274 + ], + "spans": [ + { + "bbox": [ + 175, + 261, + 432, + 274 + ], + "type": "text", + "content": "Project Page: https://snap-research.github.io/snapgen" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 76, + 284, + 527, + 323 + ], + "blocks": [ + { + "bbox": [ + 76, + 284, + 527, + 323 + ], + "lines": [ + { + "bbox": [ + 76, + 284, + 527, + 323 + ], + "spans": [ + { + "bbox": [ + 76, + 284, + 527, + 323 + ], + "type": "table", + "html": "
OursPixArt-αLumina-NextSD3-MediumSDXLPlaygroundv2SD3.5-Large
Param0.38B0.6B2B2B2.6B2.6B8.1B
MobileYESNONONONONONO
", + "image_path": "119ed19f8374061c9cba73ed4c3ab43c80a182d6573aa727fe7e85e02e81a024.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 93, + 323, + 484, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 323, + 484, + 332 + ], + "spans": [ + { + "bbox": [ + 93, + 323, + 484, + 332 + ], + "type": "text", + "content": "\"an old raccoon wearing a top hat and holding an apple, oil painting in the style of van gogh, ...\"" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 88, + 331, + 533, + 396 + ], + "blocks": [ + { + "bbox": [ + 88, + 331, + 533, + 396 + ], + "lines": [ + { + "bbox": [ + 88, + 331, + 533, + 396 + ], + "spans": [ + { + "bbox": [ + 88, + 331, + 533, + 396 + ], + "type": "image", + "image_path": "aacb2837f306dbd5f9799eb9b41fdaa8e940c7d011704296ab0406d4cf3c4d6f.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 93, + 397, + 320, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 397, + 320, + 407 + ], + "spans": [ + { + "bbox": [ + 93, + 397, + 320, + 407 + ], + "type": "text", + "content": "\"a dolphin in an astronaut suit, Animals, Simple Detail\"" + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 88, + 406, + 533, + 471 + ], + "blocks": [ + { + "bbox": [ + 88, + 406, + 533, + 471 + ], + "lines": [ + { + "bbox": [ + 88, + 406, + 533, + 471 + ], + "spans": [ + { + "bbox": [ + 88, + 406, + 533, + 471 + ], + "type": "image", + "image_path": "e9d2e69529494de4303059780935b53f3466e0207cd4a7a965b2e547d1e4c28e.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 93, + 472, + 473, + 481 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 472, + 473, + 481 + ], + "spans": [ + { + "bbox": [ + 93, + 472, + 473, + 481 + ], + "type": "text", + "content": "\"..., young sudanese female, glamour, natural, front view, extreme detailed and texture skin, ...\"" + } + ] + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 88, + 481, + 533, + 545 + ], + "blocks": [ + { + "bbox": [ + 88, + 481, + 533, + 545 + ], + "lines": [ + { + "bbox": [ + 88, + 481, + 533, + 545 + ], + "spans": [ + { + "bbox": [ + 88, + 481, + 533, + 545 + ], + "type": "image", + "image_path": "bf25a52443399be51fd3200f42287d00d2a2fd5514690042d0061c1f9fd468d6.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "bbox": [ + 93, + 547, + 496, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 547, + 496, + 556 + ], + "spans": [ + { + "bbox": [ + 93, + 547, + 496, + 556 + ], + "type": "text", + "content": "... Llama wearing sunglasses standing on the deck of a spaceship with the Earth in the background, ..." + } + ] + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 88, + 555, + 533, + 620 + ], + "blocks": [ + { + "bbox": [ + 88, + 555, + 533, + 620 + ], + "lines": [ + { + "bbox": [ + 88, + 555, + 533, + 620 + ], + "spans": [ + { + "bbox": [ + 88, + 555, + 533, + 620 + ], + "type": "image", + "image_path": "d7930e76b7a075864239cf2e635ecc6c558325029304c0a3d19cee749799ad40.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "bbox": [ + 93, + 622, + 484, + 631 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 622, + 484, + 631 + ], + "spans": [ + { + "bbox": [ + 93, + 622, + 484, + 631 + ], + "type": "text", + "content": "\"... cute monster ... eating a sashimi while sitting at a breakfast table full of fruits and insects\"" + } + ] + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 88, + 631, + 533, + 696 + ], + "blocks": [ + { + "bbox": [ + 88, + 631, + 533, + 696 + ], + "lines": [ + { + "bbox": [ + 88, + 631, + 533, + 696 + ], + "spans": [ + { + "bbox": [ + 88, + 631, + 533, + 696 + ], + "type": "image", + "image_path": "290a0e2fa4e822994c983446619421f959f3cfc23933ccc1528d4de079fc86f4.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 704, + 555, + 738 + ], + "lines": [ + { + "bbox": [ + 55, + 704, + 555, + 738 + ], + "spans": [ + { + "bbox": [ + 55, + 704, + 555, + 738 + ], + "type": "text", + "content": "Figure 1. Comparison of various text-to-image models in terms of model size, mobile device compatibility, and visual output quality. Our model, with only " + }, + { + "bbox": [ + 55, + 704, + 555, + 738 + ], + "type": "inline_equation", + "content": "379M" + }, + { + "bbox": [ + 55, + 704, + 555, + 738 + ], + "type": "text", + "content": " parameters, demonstrates competitive visual quality while being mobile-compatible. Input text prompts are shown above each image grid; all images are generated at " + }, + { + "bbox": [ + 55, + 704, + 555, + 738 + ], + "type": "inline_equation", + "content": "1024^2" + }, + { + "bbox": [ + 55, + 704, + 555, + 738 + ], + "type": "text", + "content": " resolution—zoom in for details. More examples are shown in webpage." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 15 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 15 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 15 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 198, + 14, + 442, + 24 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 14, + 442, + 24 + ], + "spans": [ + { + "bbox": [ + 198, + 14, + 442, + 24 + ], + "type": "text", + "content": "Except for this watermark, it is identical to the accepted version;" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 179, + 25, + 462, + 36 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 25, + 462, + 36 + ], + "spans": [ + { + "bbox": [ + 179, + 25, + 462, + 36 + ], + "type": "text", + "content": "the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 295, + 760, + 315, + 769 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 760, + 315, + 769 + ], + "spans": [ + { + "bbox": [ + 295, + 760, + 315, + 769 + ], + "type": "text", + "content": "7997" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 152, + 72, + 200, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 72, + 200, + 83 + ], + "spans": [ + { + "bbox": [ + 152, + 72, + 200, + 83 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 96, + 297, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 96, + 297, + 373 + ], + "spans": [ + { + "bbox": [ + 55, + 96, + 297, + 373 + ], + "type": "text", + "content": "Existing text-to-image (T2I) diffusion models face several limitations, including large model sizes, slow runtime, and low-quality generation on mobile devices. This paper aims to address all of these challenges by developing an extremely small and fast T2I model that generates high-resolution and high-quality images on mobile platforms. We propose several techniques to achieve this goal. First, we systematically examine the design choices of the network architecture to reduce model parameters and latency, while ensuring high-quality generation. Second, to further improve generation quality, we employ cross-architecture knowledge distillation from a much larger model, using a multi-level approach to guide the training of our model from scratch. Third, we enable a few-step generation by integrating adversarial guidance with knowledge distillation. For the first time, our model SnapGen, demonstrates the generation of " + }, + { + "bbox": [ + 55, + 96, + 297, + 373 + ], + "type": "inline_equation", + "content": "1024^2" + }, + { + "bbox": [ + 55, + 96, + 297, + 373 + ], + "type": "text", + "content": " px images on a mobile device around 1.4 seconds. On ImageNet-1K, our model, with only 372M parameters, achieves an FID of 2.06 for " + }, + { + "bbox": [ + 55, + 96, + 297, + 373 + ], + "type": "inline_equation", + "content": "256^2" + }, + { + "bbox": [ + 55, + 96, + 297, + 373 + ], + "type": "text", + "content": " px generation. On T2I benchmarks (i.e., GenEval and DPG-Bench), our model with merely 379M parameters, surpasses large-scale models with billions of parameters at a significantly smaller size (e.g., " + }, + { + "bbox": [ + 55, + 96, + 297, + 373 + ], + "type": "inline_equation", + "content": "7 \\times" + }, + { + "bbox": [ + 55, + 96, + 297, + 373 + ], + "type": "text", + "content": " smaller than SDXL, " + }, + { + "bbox": [ + 55, + 96, + 297, + 373 + ], + "type": "inline_equation", + "content": "14 \\times" + }, + { + "bbox": [ + 55, + 96, + 297, + 373 + ], + "type": "text", + "content": " smaller than IF-XL)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 394, + 135, + 406 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 394, + 135, + 406 + ], + "spans": [ + { + "bbox": [ + 56, + 394, + 135, + 406 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 415, + 295, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 415, + 295, + 498 + ], + "spans": [ + { + "bbox": [ + 55, + 415, + 295, + 498 + ], + "type": "text", + "content": "Large-scale text-to-image (T2I) diffusion models [12, 13, 15, 51, 53, 57-59] have achieved remarkable success in content generation, powering numerous applications like image editing [48, 64, 72, 85] and video creation [50, 54, 78]. However, T2I models often come with substantial model sizes and slow runtime, and deploying them on the cloud raises concerns related to data security and high costs [65]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 498, + 295, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 295, + 594 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 295, + 594 + ], + "type": "text", + "content": "To address these challenges, there is huge growing interest in developing smaller and faster T2I models through techniques like model compression (e.g., pruning and quantization) [40, 67, 86], step reduction by distillation [77, 80], and efficient attention mechanisms that mitigate the quadratic complexity [46, 74]. Nevertheless, current works still encounter limitations, e.g., low-resolution generation on mobile devices, that constrain their broader application." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 594, + 295, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 665 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 665 + ], + "type": "text", + "content": "Most importantly, a critical question remains unexplored: how can we train a T2I model from scratch to generate high-quality, high-resolution images on mobile? Such a model would offer substantial advantages in speed, compactness, cost-effectiveness, and secure deployment. To build this model, we introduce several innovations:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 665, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 295, + 713 + ], + "type": "text", + "content": "- Efficient Network Architectures: We conduct an in-depth examination of network architectures, including the denoising UNet and Autoencoder (AE), to obtain optimal trade-off between resource usage and performance. Un" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 321, + 72, + 553, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 72, + 553, + 133 + ], + "spans": [ + { + "bbox": [ + 321, + 72, + 553, + 133 + ], + "type": "text", + "content": "like prior works that optimize and compress pre-trained diffusion models [10, 32, 83], we directly focus on macro- and micro-level design choices to achieve a novel architecture that greatly reduces model size and computational complexity, while preserving high-quality generation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 133, + 553, + 346 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 314, + 133, + 553, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 133, + 553, + 287 + ], + "spans": [ + { + "bbox": [ + 314, + 133, + 553, + 287 + ], + "type": "text", + "content": "- Improved Training Techniques: We introduce several improvements to train a compact T2I model from scratch. We utilize flow matching [44, 47] as objective, aligning with larger models like SD3 [18] and SD3.5 [3]. This design enables effective knowledge and step distillation, transferring rich representations from large-scale diffusion models to our much smaller one. Further, we propose a multi-level knowledge distillation with a timestep-aware scaling that combines multiple training objectives. Instead of weighting objectives through a linear combination as in prior works [32, 46], we consider target prediction difficulty (i.e., student-teacher difference) across various timesteps in flow matching." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 288, + 553, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 288, + 553, + 346 + ], + "spans": [ + { + "bbox": [ + 314, + 288, + 553, + 346 + ], + "type": "text", + "content": "- Advanced Step Distillation: We perform step distillation on our model by combining the adversarial training along with the knowledge distillation using a few-step teacher model (i.e., SD3.5-Large-Turbo [5]), enabling ultra-fast high-quality generation with only 4 or 8 steps." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 348, + 553, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 348, + 553, + 371 + ], + "spans": [ + { + "bbox": [ + 313, + 348, + 553, + 371 + ], + "type": "text", + "content": "We demonstrate the superior advantages of our approach and model through extensive experiments:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 372, + 553, + 574 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 314, + 372, + 553, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 372, + 553, + 430 + ], + "spans": [ + { + "bbox": [ + 314, + 372, + 553, + 430 + ], + "type": "text", + "content": "- On ImageNet-1K [16] class-conditional image generation task, our model achieves the FID comparable to existing works with significantly reduced model size and computation, i.e., half the model size and one-third of the compute resources compared with SiT-XL[49], as in Tab. 1." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 431, + 553, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 431, + 553, + 491 + ], + "spans": [ + { + "bbox": [ + 314, + 431, + 553, + 491 + ], + "type": "text", + "content": "- For large-scale T2I generation, our UNet model, with only 379M parameters, demonstrates superior generation quality compared to billion-parameter models [42, 53, 87], e.g., improved metrics on benchmark datasets (Tab. 3) and human evaluation (Fig. 8)." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 314, + 491, + 553, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 491, + 553, + 538 + ], + "spans": [ + { + "bbox": [ + 314, + 491, + 553, + 538 + ], + "type": "text", + "content": "- Our compressed decoder, trained from scratch, has competitive reconstruction quality compared to commonly used models [18, 53], with more than " + }, + { + "bbox": [ + 314, + 491, + 553, + 538 + ], + "type": "inline_equation", + "content": "36 \\times" + }, + { + "bbox": [ + 314, + 491, + 553, + 538 + ], + "type": "text", + "content": " smaller size, enabling the mobile deployment." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 314, + 539, + 553, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 539, + 553, + 574 + ], + "spans": [ + { + "bbox": [ + 314, + 539, + 553, + 574 + ], + "type": "text", + "content": "- Notably, for the first time, we show a T2I model achieving high-resolution generation (e.g., " + }, + { + "bbox": [ + 314, + 539, + 553, + 574 + ], + "type": "inline_equation", + "content": "1024^2" + }, + { + "bbox": [ + 314, + 539, + 553, + 574 + ], + "type": "text", + "content": " px) on mobile devices (e.g., iPhone 16 Pro-Max) around 1.4 seconds." + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 586, + 400, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 586, + 400, + 597 + ], + "spans": [ + { + "bbox": [ + 314, + 586, + 400, + 597 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "content": "High-Resolution Text-to-Image Models have emerged with advanced architectures and multi-stage approaches designed to enhance visual fidelity and user customization. SDXL [53] is a pioneering work in this field, employing a refined cascading approach with UNet backbone to generate high-detailed images, resulting in photorealistic outputs that maintain sharpness and clarity. The following studies explore different techniques like more advanced text encoders, better image refinement, or improved dataset preparation," + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 761, + 315, + 769 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 761, + 315, + 769 + ], + "spans": [ + { + "bbox": [ + 295, + 761, + 315, + 769 + ], + "type": "text", + "content": "7998" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": "to obtain better text-image alignment or higher-quality generation [6, 7, 15, 20, 30, 36-38, 41, 45, 48, 69]. However, most of these models contain billions of parameters, making them extremely slow, and not being able to run on resource-contained hardware like mobile devices. In this work, we aim to build a small and fast model that can perform high-resolution generation even on mobile platforms." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 159, + 294, + 291 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 159, + 294, + 291 + ], + "spans": [ + { + "bbox": [ + 55, + 159, + 294, + 291 + ], + "type": "text", + "content": "Efficient Diffusion Models address the challenges of bulky model size and long runtime. There have been efforts exploring the architecture optimization to remove redundancy from large models, demonstrating the on-device generation within seconds [11, 40, 67, 86]. However, these models are constrained to low-resolution output, i.e., " + }, + { + "bbox": [ + 55, + 159, + 294, + 291 + ], + "type": "inline_equation", + "content": "512^{2}" + }, + { + "bbox": [ + 55, + 159, + 294, + 291 + ], + "type": "text", + "content": " px. To enable efficient high-resolution generation, SANA [74] and LinFusion [46] incorporate linear attention [9, 14, 31] to achieve 1K generation on laptop GPUs. In contrast, we target a broader range of platforms, supporting high-resolution generation (e.g., 1K) directly on mobile devices." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 294, + 294, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 294, + 294, + 415 + ], + "spans": [ + { + "bbox": [ + 55, + 294, + 294, + 415 + ], + "type": "text", + "content": "Knowledge Distillation in Diffusion Models. In the context of diffusion models, previous works focus on distilling large, high-capacity teacher models into more compact, efficient student models within a homogeneous architecture [32, 46]. They reduce the complexity of the model by removing certain components like attention [70] or residual blocks [23], while maintaining the architectural structure. However, our approach diverges from this trend by utilizing a heterogeneous architecture for more aggressively efficient yet challenging distillation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 417, + 294, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 417, + 294, + 537 + ], + "spans": [ + { + "bbox": [ + 55, + 417, + 294, + 537 + ], + "type": "text", + "content": "Adversarial Step Distillation uses the techniques from adversarial training [22] to reduce the number of diffusion steps, while maintaining high image quality [60, 61]. For example, UFOGen [77] employs a diffusion-GAN formulation [71, 73, 76] to significantly reduce inference time while maintaining competitive performance. DMD2 [79] builds on prior distillation methods by using distribution matching with adversarial loss. Different from exiting works, we conduct step-distillation on a very compact model and train the model along with the knowledge distillation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 558, + 111, + 571 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 111, + 571 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 111, + 571 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "text", + "content": "In this section, we present how to craft and train a highly efficient T2I model for high-resolution generation. Specifically, starting from the architecture designed in the latent diffusion model [58], we optimize both denoising backbone (Sec. 3.1, Fig. 2, and Fig. 3) and autoencoder (Sec. 3.2 and Fig. 4) to make them compact and fast, even on mobile devices. We then propose the improved training recipe and knowledge distillation (Sec. 3.3 and Fig. 5), empowering a high-performance T2I model. Lastly, we introduce our step distillation to significantly reduce the number of denoising steps for a faster T2I model (Sec. 3.4 and Fig. 7)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 72, + 465, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 465, + 83 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 465, + 83 + ], + "type": "text", + "content": "3.1. Efficient UNet Architecture" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 90, + 553, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 90, + 553, + 198 + ], + "spans": [ + { + "bbox": [ + 313, + 90, + 553, + 198 + ], + "type": "text", + "content": "Here we describe the design choices for the denoising UNet. Baseline Architecture. We choose UNet from SDXL [53] as the baseline (Fig. 2(a)) for our backbone since it has superior efficiency and faster convergence [39] than pure transformer-based models [12, 13]. We adjust the UNet into a thinner and shorter model (i.e., reducing the number of transformer blocks from [0, 2, 10] in three stages to [0, 2, 4] and their channel dimensions from [320, 640, 1280] to [256, 512, 896]), and iterate design choices on top of it." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 199, + 553, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 199, + 553, + 365 + ], + "spans": [ + { + "bbox": [ + 313, + 199, + 553, + 365 + ], + "type": "text", + "content": "Evaluation Metrics. We train models on ImageNet-1K [16] under class-conditional generation for 120 epochs, unless specified otherwise, and report the FID score [25] for " + }, + { + "bbox": [ + 313, + 199, + 553, + 365 + ], + "type": "inline_equation", + "content": "256^{2}" + }, + { + "bbox": [ + 313, + 199, + 553, + 365 + ], + "type": "text", + "content": " px generation. Similarly to existing work [30], we inject the class conditions through a text template \"a photo of \". We then encode it with a light text encoder to align the pipeline for the T2I generation. We also calculate the number of parameters for different models, floating point operations (FLOPs) (measured on a latent size of " + }, + { + "bbox": [ + 313, + 199, + 553, + 365 + ], + "type": "inline_equation", + "content": "128 \\times 128" + }, + { + "bbox": [ + 313, + 199, + 553, + 365 + ], + "type": "text", + "content": ", equivalent to a " + }, + { + "bbox": [ + 313, + 199, + 553, + 365 + ], + "type": "inline_equation", + "content": "1024 \\times 1024" + }, + { + "bbox": [ + 313, + 199, + 553, + 365 + ], + "type": "text", + "content": " image after decoding), and the runtime on mobile device (tested on iPhone 15 Pro). Detailed training and evaluation settings can be found in Supplemental Materials. In the following, we introduce the key architectural changes that improve the model." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 366, + 553, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 366, + 553, + 498 + ], + "spans": [ + { + "bbox": [ + 313, + 366, + 553, + 498 + ], + "type": "text", + "content": "Remove Self-Attention from High-Resolution Stages. Self-attention (SA) layer is restricted by its quadratic computational complexity, incurring a heavy computational cost and high memory consumption for high-resolution input. As such, we keep the SA layer only in the lowest-resolution stage while removing it from other higher-resolution stages, i.e., Fig. 2 (b). This leads to " + }, + { + "bbox": [ + 313, + 366, + 553, + 498 + ], + "type": "inline_equation", + "content": "17\\%" + }, + { + "bbox": [ + 313, + 366, + 553, + 498 + ], + "type": "text", + "content": " fewer FLOPs and " + }, + { + "bbox": [ + 313, + 366, + 553, + 498 + ], + "type": "inline_equation", + "content": "24\\%" + }, + { + "bbox": [ + 313, + 366, + 553, + 498 + ], + "type": "text", + "content": " latency reduction, as shown in Fig. 3. Interestingly, we even observe a performance improvement, i.e., FID decreased to 3.12 from 3.76. We hypothesize that models with SA in high-resolution stages converge more slowly." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 498, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 553, + 713 + ], + "type": "text", + "content": "Replace Conv with Expanded Separable Conv. Regular convolution (Conv) is redundant in both parameters and computation. To address this, we replace all Conv layers with the separable convolution (SepConv) [28], composed of a depthwise convolution (DW) followed by a pointwise convolution (PW), as shown in Fig. 2 (c). This replacement reduces parameters by " + }, + { + "bbox": [ + 313, + 498, + 553, + 713 + ], + "type": "inline_equation", + "content": "24\\%" + }, + { + "bbox": [ + 313, + 498, + 553, + 713 + ], + "type": "text", + "content": " and latency by " + }, + { + "bbox": [ + 313, + 498, + 553, + 713 + ], + "type": "inline_equation", + "content": "62\\%" + }, + { + "bbox": [ + 313, + 498, + 553, + 713 + ], + "type": "text", + "content": ", but also leads to a performance drop (FID increases from 3.12 to 3.38). To address the issue, we expand the intermediate channels. Specifically, the number of channels after the first PW layer is increased with an expansion ratio, and reduced back to the original number after the second PW layer. The expansion ratio is set to 2 to balance the trade-off between performance, latency, and model parameters. Such a design aligns our residual block with the recently proposed Universal Inverted Bottleneck (UIB) block [55]. As a result, our model achieves " + }, + { + "bbox": [ + 313, + 498, + 553, + 713 + ], + "type": "inline_equation", + "content": "15\\%" + }, + { + "bbox": [ + 313, + 498, + 553, + 713 + ], + "type": "text", + "content": " fewer parameters, " + }, + { + "bbox": [ + 313, + 498, + 553, + 713 + ], + "type": "inline_equation", + "content": "27\\%" + }, + { + "bbox": [ + 313, + 498, + 553, + 713 + ], + "type": "text", + "content": " less computation, and " + }, + { + "bbox": [ + 313, + 498, + 553, + 713 + ], + "type": "inline_equation", + "content": "2.4 \\times" + }, + { + "bbox": [ + 313, + 498, + 553, + 713 + ], + "type": "text", + "content": " speedup, while obtaining a lower FID." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 761, + 315, + 769 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 761, + 315, + 769 + ], + "spans": [ + { + "bbox": [ + 295, + 761, + 315, + 769 + ], + "type": "text", + "content": "7999" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 71, + 272, + 323 + ], + "blocks": [ + { + "bbox": [ + 58, + 71, + 272, + 323 + ], + "lines": [ + { + "bbox": [ + 58, + 71, + 272, + 323 + ], + "spans": [ + { + "bbox": [ + 58, + 71, + 272, + 323 + ], + "type": "image", + "image_path": "32fc1f037603aa4b02b36d78e8d57b7fbeb695f9b7b38d10ddfc5b30f0691153.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 328, + 555, + 352 + ], + "lines": [ + { + "bbox": [ + 55, + 328, + 555, + 352 + ], + "spans": [ + { + "bbox": [ + 55, + 328, + 555, + 352 + ], + "type": "text", + "content": "Figure 2. Efficient UNet. Starting from a thinner and shorter version of the UNet from SDXL (as in (a)), we explore a series of architectural changes, i.e., (b)-(f), to develop a smaller and faster model while retaining high-quality generation performance, as evaluated in Fig. 3." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 276, + 77, + 358, + 173 + ], + "blocks": [ + { + "bbox": [ + 276, + 77, + 358, + 173 + ], + "lines": [ + { + "bbox": [ + 276, + 77, + 358, + 173 + ], + "spans": [ + { + "bbox": [ + 276, + 77, + 358, + 173 + ], + "type": "image", + "image_path": "6ebcd4cd81b67055e08577ab90c0ff5d10cb6ffe379d1896340721d001ac9148.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 372, + 76, + 447, + 174 + ], + "blocks": [ + { + "bbox": [ + 372, + 76, + 447, + 174 + ], + "lines": [ + { + "bbox": [ + 372, + 76, + 447, + 174 + ], + "spans": [ + { + "bbox": [ + 372, + 76, + 447, + 174 + ], + "type": "image", + "image_path": "aa95f45746635077da11e1c3ab4ec04dba434d96490f47abbed1ac86caeba047.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 461, + 76, + 547, + 175 + ], + "blocks": [ + { + "bbox": [ + 461, + 76, + 547, + 175 + ], + "lines": [ + { + "bbox": [ + 461, + 76, + 547, + 175 + ], + "spans": [ + { + "bbox": [ + 461, + 76, + 547, + 175 + ], + "type": "image", + "image_path": "f60a78f2e2c62771c56e9f88618c757660afe5faecd4a583088143b4925d4edd.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 276, + 185, + 354, + 322 + ], + "blocks": [ + { + "bbox": [ + 276, + 185, + 354, + 322 + ], + "lines": [ + { + "bbox": [ + 276, + 185, + 354, + 322 + ], + "spans": [ + { + "bbox": [ + 276, + 185, + 354, + 322 + ], + "type": "image", + "image_path": "9905b120562358bddc6086f030628c8beb2ec45868c4467990b7fa37c80bffd1.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 375, + 182, + 544, + 322 + ], + "blocks": [ + { + "bbox": [ + 375, + 182, + 544, + 322 + ], + "lines": [ + { + "bbox": [ + 375, + 182, + 544, + 322 + ], + "spans": [ + { + "bbox": [ + 375, + 182, + 544, + 322 + ], + "type": "image", + "image_path": "50e5b5423661fd31a956e0611a3860e2ee9d56de9a472ac1830420e039b65eb7.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 72, + 354, + 541, + 472 + ], + "blocks": [ + { + "bbox": [ + 72, + 354, + 541, + 472 + ], + "lines": [ + { + "bbox": [ + 72, + 354, + 541, + 472 + ], + "spans": [ + { + "bbox": [ + 72, + 354, + 541, + 472 + ], + "type": "image", + "image_path": "c1dbff62b401a03d50e2895ca313bc72b820a589531ce0be6086fcd55c9f64e7.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 476, + 555, + 522 + ], + "lines": [ + { + "bbox": [ + 54, + 476, + 555, + 522 + ], + "spans": [ + { + "bbox": [ + 54, + 476, + 555, + 522 + ], + "type": "text", + "content": "Figure 3. Comparisons of Performance and Efficiency for various Design Choices of Efficient UNet. The generation quality is evaluated using FID calculated on ImageNet-1K for " + }, + { + "bbox": [ + 54, + 476, + 555, + 522 + ], + "type": "inline_equation", + "content": "256^{2}" + }, + { + "bbox": [ + 54, + 476, + 555, + 522 + ], + "type": "text", + "content": " px generation. The efficiency metrics include model parameters, latency, and FLOPs. FLOPs and latency (on iPhone 15 Pro) are measured with a " + }, + { + "bbox": [ + 54, + 476, + 555, + 522 + ], + "type": "inline_equation", + "content": "128 \\times 128" + }, + { + "bbox": [ + 54, + 476, + 555, + 522 + ], + "type": "text", + "content": " latent, equivalent to a " + }, + { + "bbox": [ + 54, + 476, + 555, + 522 + ], + "type": "inline_equation", + "content": "1024 \\times 1024" + }, + { + "bbox": [ + 54, + 476, + 555, + 522 + ], + "type": "text", + "content": " decoded image, for one forward pass. We show the architecture enhancements that improve any of the metrics without hurting others." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 533, + 295, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 533, + 295, + 641 + ], + "spans": [ + { + "bbox": [ + 54, + 533, + 295, + 641 + ], + "type": "text", + "content": "Trim FFN Layers. For the layers in the feed-forward network (FFN), the hidden channel expansion ratio is set to 4 by default and further doubled by using the gated unit. This substantially inflates the model parameters, computation, and memory usage. Following MobileDiffusion [86], we examine the efficacy of simply reducing the expansion ratio, as shown in Fig. 2 (d). We show that reducing the expansion ratio to 3 preserves comparable FID performance while reducing both parameters and FLOPs by " + }, + { + "bbox": [ + 54, + 533, + 295, + 641 + ], + "type": "inline_equation", + "content": "12\\%" + }, + { + "bbox": [ + 54, + 533, + 295, + 641 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 54, + 641, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 641, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 641, + 295, + 714 + ], + "type": "text", + "content": "Replace MHSA with MQA. Multi-Head Self-Attention (MHSA) requires multiple sets of keys and values for each attention head. In contrast, Multi-Query Attention (MQA) [63] is more efficient by sharing a single set of keys and values across all heads. Replacing MHSA with MQA reduces parameters by " + }, + { + "bbox": [ + 54, + 641, + 295, + 714 + ], + "type": "inline_equation", + "content": "16\\%" + }, + { + "bbox": [ + 54, + 641, + 295, + 714 + ], + "type": "text", + "content": " and latency by " + }, + { + "bbox": [ + 54, + 641, + 295, + 714 + ], + "type": "inline_equation", + "content": "9\\%" + }, + { + "bbox": [ + 54, + 641, + 295, + 714 + ], + "type": "text", + "content": ", with mini" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 533, + 555, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 555, + 570 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 555, + 570 + ], + "type": "text", + "content": "mal impact on performance. Interestingly, the " + }, + { + "bbox": [ + 313, + 533, + 555, + 570 + ], + "type": "inline_equation", + "content": "9\\%" + }, + { + "bbox": [ + 313, + 533, + 555, + 570 + ], + "type": "text", + "content": " saving in latency exceeds the " + }, + { + "bbox": [ + 313, + 533, + 555, + 570 + ], + "type": "inline_equation", + "content": "6\\%" + }, + { + "bbox": [ + 313, + 533, + 555, + 570 + ], + "type": "text", + "content": " decrease in FLOPs, as the reduced memory access enables higher computational intensity." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "type": "text", + "content": "Inject Conditions to the First Stage. Cross-attention (CA) blends the conditional information (e.g., textural description) along with the spatial features to generate images that align with the condition. However, the UNet of SDXL only applies CA in transformer blocks starting from the second stage, resulting in the missing conditional guidance for the first stage. In response, we propose to introduce the conditional embeddings from the very first stage, as in Fig. 2(e). Specifically, we replace the residual blocks with transformer blocks that include CA and FFN while without SA layers. This adjustment makes the model smaller, faster, and more efficient while improving FID." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 760, + 315, + 769 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 760, + 315, + 769 + ], + "spans": [ + { + "bbox": [ + 295, + 760, + 315, + 769 + ], + "type": "text", + "content": "8000" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 239 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 239 + ], + "type": "text", + "content": "Employ QK RMSNorm and RoPE Positional Embeddings. We extend two advanced techniques developed originally for language models, Query-Key (QK) Normalization [24] with RMSNorm [82] and Rotary Position Embedding (RoPE) [66], to enhance the model (Fig. 2 (f)). RMSNorm, applied after the Query-Key projection in the attention mechanism, reduces the risk of softmax saturation without sacrificing model expressiveness while stabilizing training for faster convergence. In addition, we adapt RoPE from one dimension to two dimensions for better supporting higher resolution since it significantly mitigates artifacts like repeated objects. Together, RMSNorm and RoPE introduce negligible computational and parameter overhead, while offering measurable gains in FID performance." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 239, + 296, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 239, + 296, + 361 + ], + "spans": [ + { + "bbox": [ + 55, + 239, + 296, + 361 + ], + "type": "text", + "content": "Discussion. The above optimization results in an efficient and powerful diffusion backbone capable of generating high-resolution images on mobile devices. Before proceeding with large-scale T2I training, we compare the capacity of our model against existing works on ImageNet-1K. We train the model for 1,000 epochs by following the setting from prior work [58]. We evaluate the model using varied CFG [26, 34] across different inference timesteps. As shown in Tab. 1, our efficient UNet achieves comparable FID to SiT-XL [49], while being almost " + }, + { + "bbox": [ + 55, + 239, + 296, + 361 + ], + "type": "inline_equation", + "content": "45\\%" + }, + { + "bbox": [ + 55, + 239, + 296, + 361 + ], + "type": "text", + "content": " smaller." + } + ] + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 85, + 388, + 266, + 474 + ], + "blocks": [ + { + "bbox": [ + 55, + 365, + 296, + 387 + ], + "lines": [ + { + "bbox": [ + 55, + 365, + 296, + 387 + ], + "spans": [ + { + "bbox": [ + 55, + 365, + 296, + 387 + ], + "type": "text", + "content": "Table 1. Class-conditional image generation on ImageNet " + }, + { + "bbox": [ + 55, + 365, + 296, + 387 + ], + "type": "inline_equation", + "content": "{256} \\times {256}" + }, + { + "bbox": [ + 55, + 365, + 296, + 387 + ], + "type": "text", + "content": " with CFG. FLOPs are calculated for one forward pass." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 85, + 388, + 266, + 474 + ], + "lines": [ + { + "bbox": [ + 85, + 388, + 266, + 474 + ], + "spans": [ + { + "bbox": [ + 85, + 388, + 266, + 474 + ], + "type": "table", + "html": "
ModelParam (M)FLOPs (G)FID↓
LDM-4 [58]4001043.60
UViT-L [8]287773.40
UViT-H [8]5011332.29
DiT-XL [52]6751192.27
SiT-XL [49]6751192.06
Ours372382.06
", + "image_path": "7b8efd7541fcd187df855efe04659528f73b501294d752bbade1a1961d367153.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 488, + 185, + 501 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 488, + 185, + 501 + ], + "spans": [ + { + "bbox": [ + 55, + 488, + 185, + 501 + ], + "type": "text", + "content": "3.2. Tiny and Fast Decoder" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 506, + 296, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 506, + 296, + 555 + ], + "spans": [ + { + "bbox": [ + 55, + 506, + 296, + 555 + ], + "type": "text", + "content": "Besides the denoising model, the decoder also takes a significant ratio of total runtime, especially for on-device deployment [40, 86]. Here we introduce a new architecture of decoder (Fig. 4) for efficient high-resolution generation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 555, + 296, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 555, + 296, + 700 + ], + "spans": [ + { + "bbox": [ + 55, + 555, + 296, + 700 + ], + "type": "text", + "content": "Baseline Decoder. We use the autoencoder (AE) from SD3 [18] as our baseline model (i.e., the same encoder from SD3 AE), due to its superior reconstruction quality. The AE maps an image " + }, + { + "bbox": [ + 55, + 555, + 296, + 700 + ], + "type": "inline_equation", + "content": "X \\in \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 55, + 555, + 296, + 700 + ], + "type": "text", + "content": " into a lower-dimensional latent " + }, + { + "bbox": [ + 55, + 555, + 296, + 700 + ], + "type": "inline_equation", + "content": "x \\in \\mathbb{R}^{\\frac{H}{f} \\times \\frac{W}{f} \\times c}" + }, + { + "bbox": [ + 55, + 555, + 296, + 700 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 55, + 555, + 296, + 700 + ], + "type": "inline_equation", + "content": "f, c" + }, + { + "bbox": [ + 55, + 555, + 296, + 700 + ], + "type": "text", + "content": " as 8, 16 in SD3). The encoded latent " + }, + { + "bbox": [ + 55, + 555, + 296, + 700 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 555, + 296, + 700 + ], + "type": "text", + "content": " is then decoded back to an image through a decoder. For high-resolution generation, we observe that the decoder in SD3 is very slow on mobile devices. Specifically, it encounters out-of-memory (OOM) errors when generating a " + }, + { + "bbox": [ + 55, + 555, + 296, + 700 + ], + "type": "inline_equation", + "content": "1024^2" + }, + { + "bbox": [ + 55, + 555, + 296, + 700 + ], + "type": "text", + "content": " px image on both the ANE processor of the iPhone 15 Pro and the mobile GPU (Tab. 2). To overcome the latency issue, we propose a much smaller and faster decoder." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 701, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 701, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 701, + 296, + 713 + ], + "type": "text", + "content": "Efficient Decoder. We conduct a series of experiments to" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 319, + 71, + 552, + 156 + ], + "blocks": [ + { + "bbox": [ + 319, + 71, + 552, + 156 + ], + "lines": [ + { + "bbox": [ + 319, + 71, + 552, + 156 + ], + "spans": [ + { + "bbox": [ + 319, + 71, + 552, + 156 + ], + "type": "image", + "image_path": "d3551e77cf0a3ca96ce432b179f2b5e73f68505bc09ab4844ff816cca53a3bc2.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 320, + 158, + 551, + 233 + ], + "blocks": [ + { + "bbox": [ + 320, + 158, + 551, + 233 + ], + "lines": [ + { + "bbox": [ + 320, + 158, + 551, + 233 + ], + "spans": [ + { + "bbox": [ + 320, + 158, + 551, + 233 + ], + "type": "image", + "image_path": "12e3ce061452826b47b5537091521f8490242bceb1cbc04f50860da724a92895.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 238, + 554, + 261 + ], + "lines": [ + { + "bbox": [ + 313, + 238, + 554, + 261 + ], + "spans": [ + { + "bbox": [ + 313, + 238, + 554, + 261 + ], + "type": "text", + "content": "Figure 4. Comparisons of Decoder Architecture between (a) SDXL/SD3 decoder and (b) our tiny decoder." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 316, + 327, + 553, + 376 + ], + "blocks": [ + { + "bbox": [ + 313, + 266, + 554, + 321 + ], + "lines": [ + { + "bbox": [ + 313, + 266, + 554, + 321 + ], + "spans": [ + { + "bbox": [ + 313, + 266, + 554, + 321 + ], + "type": "text", + "content": "Table 2. Performance Comparison of Decoder. PSNR is calculated on COCO 2017 validation set [43]. FLOPs and latency (on iPhone 15 Pro) are measured for decoding a " + }, + { + "bbox": [ + 313, + 266, + 554, + 321 + ], + "type": "inline_equation", + "content": "128 \\times 128" + }, + { + "bbox": [ + 313, + 266, + 554, + 321 + ], + "type": "text", + "content": " latent into a " + }, + { + "bbox": [ + 313, + 266, + 554, + 321 + ], + "type": "inline_equation", + "content": "1024 \\times 1024" + }, + { + "bbox": [ + 313, + 266, + 554, + 321 + ], + "type": "text", + "content": " image. The decoder from SDXL and SD3 fail to run on the neural engine of mobile, resulting in a huge runtime." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 316, + 327, + 553, + 376 + ], + "lines": [ + { + "bbox": [ + 316, + 327, + 553, + 376 + ], + "spans": [ + { + "bbox": [ + 316, + 327, + 553, + 376 + ], + "type": "table", + "html": "
DecoderChPSNRParam (M)FLOPs (G)Latency (ms) on ANELatency (ms) on GPU
SDXL [53]424.8949.494970OOM9469
SD3 [18]1627.9249.554970OOMOOM
Ours1627.851.38224174-
", + "image_path": "7ee9b939e1012da1bbfab0a4b14ed4b98959f9a5d9421ffea5815d6a24f1e34c.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 389, + 553, + 412 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 389, + 553, + 412 + ], + "spans": [ + { + "bbox": [ + 313, + 389, + 553, + 412 + ], + "type": "text", + "content": "decide the efficient decoder with the following key changes compared with the baseline architecture:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 414, + 553, + 533 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 314, + 414, + 553, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 414, + 553, + 437 + ], + "spans": [ + { + "bbox": [ + 314, + 414, + 553, + 437 + ], + "type": "text", + "content": "1. We remove attention layers to greatly reduce peak memory without a noticeable impact on decoding quality." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 314, + 437, + 553, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 437, + 553, + 473 + ], + "spans": [ + { + "bbox": [ + 314, + 437, + 553, + 473 + ], + "type": "text", + "content": "2. We keep a minimal amount of GroupNorm (GN) to find a trade-off between latency and performance (i.e., mitigating the color shifting)." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 314, + 473, + 553, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 473, + 553, + 497 + ], + "spans": [ + { + "bbox": [ + 314, + 473, + 553, + 497 + ], + "type": "text", + "content": "3. We make the decoder thinner (i.e., fewer channels or narrower width) and replace Conv with SepConvs." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 314, + 497, + 548, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 497, + 548, + 509 + ], + "spans": [ + { + "bbox": [ + 314, + 497, + 548, + 509 + ], + "type": "text", + "content": "4. We use fewer residual blocks in high-resolution stages." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 509, + 553, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 509, + 553, + 533 + ], + "spans": [ + { + "bbox": [ + 313, + 509, + 553, + 533 + ], + "type": "text", + "content": "5. We remove the Conv shortcut in residual blocks and use the upsampling layer for channel transition." + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 534, + 555, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 555, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 555, + 640 + ], + "type": "text", + "content": "Training of the Decoder. We train our decoder with the mean squared error (MSE) loss, lpips loss [84], adversarial loss [22], and discard the KL term [33] as the encoder is fixed. The decoder is trained on " + }, + { + "bbox": [ + 313, + 534, + 555, + 640 + ], + "type": "inline_equation", + "content": "256^2" + }, + { + "bbox": [ + 313, + 534, + 555, + 640 + ], + "type": "text", + "content": " image patches with a batch size of 256 and for 1M iterations. As in Tab. 2, our tiny decoder achieves a competitive PSNR score for reconstruction, while being " + }, + { + "bbox": [ + 313, + 534, + 555, + 640 + ], + "type": "inline_equation", + "content": "35.9 \\times" + }, + { + "bbox": [ + 313, + 534, + 555, + 640 + ], + "type": "text", + "content": " smaller and " + }, + { + "bbox": [ + 313, + 534, + 555, + 640 + ], + "type": "inline_equation", + "content": "54.4 \\times" + }, + { + "bbox": [ + 313, + 534, + 555, + 640 + ], + "type": "text", + "content": " faster for high-resolution generation on mobile devices compared to conventional ones (e.g., the decoder from SDXL and SD3)." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 642, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 554, + 713 + ], + "type": "text", + "content": "Discussion of Total On-Device Latency. We finally measure T2I model latency for a " + }, + { + "bbox": [ + 313, + 642, + 554, + 713 + ], + "type": "inline_equation", + "content": "1024^{2}" + }, + { + "bbox": [ + 313, + 642, + 554, + 713 + ], + "type": "text", + "content": " px generation on iPhone 16 Pro-Max. The decoder takes 119ms, and the per-step latency for the UNet is 274ms. This results in a " + }, + { + "bbox": [ + 313, + 642, + 554, + 713 + ], + "type": "inline_equation", + "content": "1.2\\sim 2.3s" + }, + { + "bbox": [ + 313, + 642, + 554, + 713 + ], + "type": "text", + "content": " runtime for 4 to 8 step generation. Note that text encoder runtime is negligible compared to other components [40]." + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 760, + 314, + 769 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 760, + 314, + 769 + ], + "spans": [ + { + "bbox": [ + 295, + 760, + 314, + 769 + ], + "type": "text", + "content": "8001" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 71, + 216, + 174 + ], + "blocks": [ + { + "bbox": [ + 58, + 71, + 216, + 174 + ], + "lines": [ + { + "bbox": [ + 58, + 71, + 216, + 174 + ], + "spans": [ + { + "bbox": [ + 58, + 71, + 216, + 174 + ], + "type": "image", + "image_path": "b28475312a143afcb6eee32703245ffd0b1f9142e7b92e05ba5ac55afcf0445b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 182, + 217, + 215 + ], + "lines": [ + { + "bbox": [ + 55, + 182, + 217, + 215 + ], + "spans": [ + { + "bbox": [ + 55, + 182, + 217, + 215 + ], + "type": "text", + "content": "Figure 5. Overview of Multi-level Knowledge Distillation, where we perform output distillation and feature distillation." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 224, + 78, + 387, + 174 + ], + "blocks": [ + { + "bbox": [ + 224, + 78, + 387, + 174 + ], + "lines": [ + { + "bbox": [ + 224, + 78, + 387, + 174 + ], + "spans": [ + { + "bbox": [ + 224, + 78, + 387, + 174 + ], + "type": "image", + "image_path": "2e659881515492e5fdebcca3c7cba7f8389e2d8bb8cd6a89f5b091ca594f0f3e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 223, + 183, + 385, + 205 + ], + "lines": [ + { + "bbox": [ + 223, + 183, + 385, + 205 + ], + "spans": [ + { + "bbox": [ + 223, + 183, + 385, + 205 + ], + "type": "text", + "content": "Figure 6. Mean loss magnitude for task loss " + }, + { + "bbox": [ + 223, + 183, + 385, + 205 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{task}}" + }, + { + "bbox": [ + 223, + 183, + 385, + 205 + ], + "type": "text", + "content": " and output distillation loss " + }, + { + "bbox": [ + 223, + 183, + 385, + 205 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{kd}}" + }, + { + "bbox": [ + 223, + 183, + 385, + 205 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 394, + 72, + 554, + 174 + ], + "blocks": [ + { + "bbox": [ + 394, + 72, + 554, + 174 + ], + "lines": [ + { + "bbox": [ + 394, + 72, + 554, + 174 + ], + "spans": [ + { + "bbox": [ + 394, + 72, + 554, + 174 + ], + "type": "image", + "image_path": "c03cef90103a0769718ce76fd5d00f63febcb96b864f8f0f3ed961ec93ee280e.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 392, + 182, + 554, + 215 + ], + "lines": [ + { + "bbox": [ + 392, + 182, + 554, + 215 + ], + "spans": [ + { + "bbox": [ + 392, + 182, + 554, + 215 + ], + "type": "text", + "content": "Figure 7. Overview of Adversarial Step Distillation. Output distillation and distribution-matching distillation are performed." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 225, + 286, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 225, + 286, + 239 + ], + "spans": [ + { + "bbox": [ + 55, + 225, + 286, + 239 + ], + "type": "text", + "content": "3.3. Training Recipe and Multi-Level Distillation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 244, + 294, + 267 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 244, + 294, + 267 + ], + "spans": [ + { + "bbox": [ + 55, + 244, + 294, + 267 + ], + "type": "text", + "content": "To improve the generation quality of our efficient diffusion model, we propose a series of training techniques." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 267, + 295, + 310 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 267, + 295, + 310 + ], + "spans": [ + { + "bbox": [ + 55, + 267, + 295, + 310 + ], + "type": "text", + "content": "Flow-based Training and Inference. Rectified Flows (RFs) [44, 47] define the forward process as straight paths connecting the data distribution to a standard normal distribution, i.e.," + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 129, + 312, + 294, + 323 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 312, + 294, + 323 + ], + "spans": [ + { + "bbox": [ + 129, + 312, + 294, + 323 + ], + "type": "interline_equation", + "content": "x _ {t} = \\left(1 - \\sigma_ {t}\\right) x _ {0} + \\sigma_ {t} \\epsilon , \\tag {1}", + "image_path": "f73d37069c8a96dd20d78000ca68b809bb0e06e3430c05dc3b27b3cf9377aafa.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 326, + 296, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 326, + 296, + 370 + ], + "spans": [ + { + "bbox": [ + 55, + 326, + 296, + 370 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 326, + 296, + 370 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 55, + 326, + 296, + 370 + ], + "type": "text", + "content": " is the clean (latent) image, " + }, + { + "bbox": [ + 55, + 326, + 296, + 370 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 326, + 296, + 370 + ], + "type": "text", + "content": " is the timestep, " + }, + { + "bbox": [ + 55, + 326, + 296, + 370 + ], + "type": "inline_equation", + "content": "\\sigma_t" + }, + { + "bbox": [ + 55, + 326, + 296, + 370 + ], + "type": "text", + "content": " is a timestep-dependent factor, and " + }, + { + "bbox": [ + 55, + 326, + 296, + 370 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 55, + 326, + 296, + 370 + ], + "type": "text", + "content": " is random noise sampled from " + }, + { + "bbox": [ + 55, + 326, + 296, + 370 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(0,I)" + }, + { + "bbox": [ + 55, + 326, + 296, + 370 + ], + "type": "text", + "content": ". The denoising UNet is formulated to predict a velocity field with the objective as" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 86, + 378, + 295, + 397 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 378, + 295, + 397 + ], + "spans": [ + { + "bbox": [ + 86, + 378, + 295, + 397 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {t a s k}} = \\mathbb {E} _ {\\epsilon \\sim \\mathcal {N} (0, I), t} \\left[ | | (\\epsilon - x _ {0}) - v _ {\\theta} (x _ {t}, t) | | _ {2} ^ {2} \\right], \\tag {2}", + "image_path": "11f42fe1d96df0341b295f3cccb49c3854d54b272d135b0b5ce14313057ffee5.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 403, + 296, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 403, + 296, + 471 + ], + "spans": [ + { + "bbox": [ + 55, + 403, + 296, + 471 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 403, + 296, + 471 + ], + "type": "inline_equation", + "content": "v_{\\theta}(x_t, t)" + }, + { + "bbox": [ + 55, + 403, + 296, + 471 + ], + "type": "text", + "content": " is the predicted velocity from UNet parameterized by " + }, + { + "bbox": [ + 55, + 403, + 296, + 471 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 55, + 403, + 296, + 471 + ], + "type": "text", + "content": ". To further enhance training stability, we apply logit-normal sampling [18] for the timestep during training, which assigns more samples to the intermediate steps. In the inference stage, we use the Flow-Euler sampler [19], which predicts the next sample based on the velocity, i.e.," + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 105, + 479, + 294, + 491 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 479, + 294, + 491 + ], + "spans": [ + { + "bbox": [ + 105, + 479, + 294, + 491 + ], + "type": "interline_equation", + "content": "x _ {t - 1} = x _ {t} + \\left(\\sigma_ {t - 1} - \\sigma_ {t}\\right) \\cdot v _ {\\theta} \\left(x _ {t}, t\\right). \\tag {3}", + "image_path": "1534793ce8605d19cc1d20660006fc23f0a96b08d600257c2cec619b76cd19bc.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 498, + 295, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 295, + 545 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 295, + 545 + ], + "type": "text", + "content": "To achieve a lower signal-to-noise ratio on high-resolution (i.e., " + }, + { + "bbox": [ + 55, + 498, + 295, + 545 + ], + "type": "inline_equation", + "content": "1024^2" + }, + { + "bbox": [ + 55, + 498, + 295, + 545 + ], + "type": "text", + "content": " px) images, we apply a timestep shift similar to SD3 [18] to adjust the scheduling factor " + }, + { + "bbox": [ + 55, + 498, + 295, + 545 + ], + "type": "inline_equation", + "content": "\\sigma_t" + }, + { + "bbox": [ + 55, + 498, + 295, + 545 + ], + "type": "text", + "content": " during both training and inference." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 546, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 546, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 546, + 296, + 713 + ], + "type": "text", + "content": "Multi-Level Knowledge Distillation. To improve the generation quality for compact models, one common practice for previous works is applying knowledge distillation to mimic the prediction of scaled-up teacher models [32]. Benefiting from the aligned flow matching objective and (AE) latent space, powerful SD3.5-Large model [4] can be used as the teacher for output distillation. However, we still face challenges due to 1) the heterogeneous architecture between the U-Net and DiT, 2) the scale difference between the distillation loss and task loss, and 3) varying prediction difficulty across different timesteps. To tackle these, we propose a novel multi-level distillation loss and apply timestep-aware scaling to stabilize and accelerate the convergence of distillation. The overview of our knowledge" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 226, + 553, + 249 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 226, + 553, + 249 + ], + "spans": [ + { + "bbox": [ + 313, + 226, + 553, + 249 + ], + "type": "text", + "content": "distillation scheme is shown in Fig. 5 and detailed techniques are elaborated as follows." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 250, + 553, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 250, + 553, + 293 + ], + "spans": [ + { + "bbox": [ + 313, + 250, + 553, + 293 + ], + "type": "text", + "content": "Aside from the task loss defined in Eq. 2, the major objective for knowledge distillation is to supervise our model " + }, + { + "bbox": [ + 313, + 250, + 553, + 293 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 250, + 553, + 293 + ], + "type": "text", + "content": " directly with the output of teacher model " + }, + { + "bbox": [ + 313, + 250, + 553, + 293 + ], + "type": "inline_equation", + "content": "\\theta_{T}" + }, + { + "bbox": [ + 313, + 250, + 553, + 293 + ], + "type": "text", + "content": ", which can be indicated as" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 363, + 297, + 553, + 317 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 363, + 297, + 553, + 317 + ], + "spans": [ + { + "bbox": [ + 363, + 297, + 553, + 317 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {k d}} = \\mathbb {E} \\left[ \\left| \\left| v _ {\\theta_ {T}} \\left(x _ {t}, t\\right) - v _ {\\theta} \\left(x _ {t}, t\\right) \\right| \\right| _ {2} ^ {2} \\right]. \\tag {4}", + "image_path": "562247b8c6e2f14aa5e16454cba6c7e81b17eec3f46961144461acf22d78ab4f.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 322, + 554, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 322, + 554, + 366 + ], + "spans": [ + { + "bbox": [ + 313, + 322, + 554, + 366 + ], + "type": "text", + "content": "Given the capacity gap between the teacher and our model, applying output-level supervision alone leads to instability and slow convergence. Therefore, we further incorporate a cross-architecture feature-level distillation loss as" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 339, + 370, + 553, + 396 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 370, + 553, + 396 + ], + "spans": [ + { + "bbox": [ + 339, + 370, + 553, + 396 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {f e a t k d}} = \\mathbb {E} \\left[ \\sum_ {(l _ {T}, l)} | | f _ {\\theta_ {T}} ^ {l _ {T}} (x _ {t}, t) - \\psi \\left(f _ {\\theta} ^ {l} \\left(x _ {t}, t\\right)\\right) | | _ {2} ^ {2} \\right], \\tag {5}", + "image_path": "a4dc9bd1d1f0daf3ecb8434cf528beadba2aa502bd27eeebe9eda4ff9c01bfa9.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 403, + 554, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 403, + 554, + 548 + ], + "spans": [ + { + "bbox": [ + 313, + 403, + 554, + 548 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 403, + 554, + 548 + ], + "type": "inline_equation", + "content": "f_{\\theta_T}^{l_T}(\\cdot)" + }, + { + "bbox": [ + 313, + 403, + 554, + 548 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 403, + 554, + 548 + ], + "type": "inline_equation", + "content": "f^l (\\cdot)" + }, + { + "bbox": [ + 313, + 403, + 554, + 548 + ], + "type": "text", + "content": " indicate the feature output from the " + }, + { + "bbox": [ + 313, + 403, + 554, + 548 + ], + "type": "inline_equation", + "content": "l_{T}" + }, + { + "bbox": [ + 313, + 403, + 554, + 548 + ], + "type": "text", + "content": "-th layer and " + }, + { + "bbox": [ + 313, + 403, + 554, + 548 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 313, + 403, + 554, + 548 + ], + "type": "text", + "content": "-th layer in teacher model and student model, respectively. Different from previous work [32, 46], we consider cross-architecture distillation from a DiT to UNet. Since the richest information in transformers sits around the last layer, we set the distillation target to this layer in both models, and use a lightweight trainable projector " + }, + { + "bbox": [ + 313, + 403, + 554, + 548 + ], + "type": "inline_equation", + "content": "\\psi (\\cdot)" + }, + { + "bbox": [ + 313, + 403, + 554, + 548 + ], + "type": "text", + "content": " with only two Conv layers to map the student feature to match the dimension of teacher feature. The proposed feature-level distillation loss provides additional supervision to the student model, leading to faster alignment to the generation quality of the teacher model." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 548, + 553, + 603 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 548, + 553, + 603 + ], + "spans": [ + { + "bbox": [ + 313, + 548, + 553, + 603 + ], + "type": "text", + "content": "Timestep-Aware Scaling. Weighting multiple objectives has been a major challenge in knowledge distillation, especially in diffusion models. The overall training objectives from previous works [32, 46, 67] are simple linear combination of multiple loss term, i.e.," + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 370, + 610, + 553, + 622 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 610, + 553, + 622 + ], + "spans": [ + { + "bbox": [ + 370, + 610, + 553, + 622 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {\\text {t a s k}} + \\lambda_ {1} \\mathcal {L} _ {\\mathrm {k d}} + \\lambda_ {2} \\mathcal {L} _ {\\text {f e a t k d}}, \\tag {6}", + "image_path": "b421724dcef716e5f748783048545370a5493925fb1e1d535e5d750a26558dc8.jpg" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 629, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 554, + 713 + ], + "type": "text", + "content": "where the weighting coefficient " + }, + { + "bbox": [ + 313, + 629, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\lambda_{1}" + }, + { + "bbox": [ + 313, + 629, + 554, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 629, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\lambda_{2}" + }, + { + "bbox": [ + 313, + 629, + 554, + 713 + ], + "type": "text", + "content": " are empirically set to constant. However, this baseline setting fails to consider the prediction difficulty in various time steps. We investigate the distribution of empirical risk magnitude of " + }, + { + "bbox": [ + 313, + 629, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{task}}" + }, + { + "bbox": [ + 313, + 629, + 554, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 629, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{kd}}" + }, + { + "bbox": [ + 313, + 629, + 554, + 713 + ], + "type": "text", + "content": " across different timestep " + }, + { + "bbox": [ + 313, + 629, + 554, + 713 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 629, + 554, + 713 + ], + "type": "text", + "content": " during model training. Fig. 6 illustrates that, in intermediate steps, prediction difficulty are lower compared to " + }, + { + "bbox": [ + 313, + 629, + 554, + 713 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 629, + 554, + 713 + ], + "type": "text", + "content": " closer to 0 or 1." + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 760, + 315, + 769 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 760, + 315, + 769 + ], + "spans": [ + { + "bbox": [ + 295, + 760, + 315, + 769 + ], + "type": "text", + "content": "8002" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "text", + "content": "Building on this important observation, we propose a timestep-aware scaling of the objective to close the gap in loss magnitude across different values of " + }, + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "text", + "content": " and to account for prediction difficulties at each timestep, as follows:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 129, + 296, + 165 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 129, + 296, + 165 + ], + "spans": [ + { + "bbox": [ + 55, + 129, + 296, + 165 + ], + "type": "interline_equation", + "content": "\\mathcal {S} \\left(\\mathcal {L} _ {\\text {t a s k}}, \\mathcal {L} _ {\\mathrm {k d}}\\right) = \\mathbb {E} _ {t} \\left[ \\lambda (t) \\cdot \\mathcal {L} _ {\\text {t a s k}} ^ {t} + (1 - \\lambda (t)) \\frac {\\left| \\mathcal {L} _ {\\text {t a s k}} ^ {t} \\right|}{\\left| \\mathcal {L} _ {\\mathrm {k d}} ^ {t} \\right|} \\cdot \\mathcal {L} _ {\\mathrm {k d}} ^ {t} \\right], \\tag {7}", + "image_path": "377d4fdfe9a1bb92182269f25f1b7c80455ca74306f82865e433afac7086e258.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 166, + 296, + 286 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 166, + 296, + 286 + ], + "spans": [ + { + "bbox": [ + 55, + 166, + 296, + 286 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 166, + 296, + 286 + ], + "type": "inline_equation", + "content": "\\lambda (t)" + }, + { + "bbox": [ + 55, + 166, + 296, + 286 + ], + "type": "text", + "content": " is the normalized standard (location 0, scale 1) logit-norm density function and " + }, + { + "bbox": [ + 55, + 166, + 296, + 286 + ], + "type": "inline_equation", + "content": "|\\cdot |" + }, + { + "bbox": [ + 55, + 166, + 296, + 286 + ], + "type": "text", + "content": " indicates the magnitude. In " + }, + { + "bbox": [ + 55, + 166, + 296, + 286 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 55, + 166, + 296, + 286 + ], + "type": "text", + "content": ", we first ensure the same scale between task loss and distillation loss across different " + }, + { + "bbox": [ + 55, + 166, + 296, + 286 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 166, + 296, + 286 + ], + "type": "text", + "content": ", then apply more teacher supervision where prediction difficulty is higher (i.e., " + }, + { + "bbox": [ + 55, + 166, + 296, + 286 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 166, + 296, + 286 + ], + "type": "text", + "content": " closer to 0 or 1), and more real data supervision where prediction difficulty is lower (i.e., intermediate timesteps). The proposed scheme considers the variation of timestep " + }, + { + "bbox": [ + 55, + 166, + 296, + 286 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 166, + 296, + 286 + ], + "type": "text", + "content": " and helps accelerate the distillation training. The final multi-level distillation objective " + }, + { + "bbox": [ + 55, + 166, + 296, + 286 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{md}}" + }, + { + "bbox": [ + 55, + 166, + 296, + 286 + ], + "type": "text", + "content": " is defined as" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 87, + 295, + 295, + 308 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 87, + 295, + 295, + 308 + ], + "spans": [ + { + "bbox": [ + 87, + 295, + 295, + 308 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {m d}} = \\mathcal {S} \\left(\\mathcal {L} _ {\\text {t a s k}}, \\mathcal {L} _ {\\mathrm {k d}}\\right) + \\mathcal {S} \\left(\\mathcal {L} _ {\\text {t a s k}}, \\mathcal {L} _ {\\text {f e a t k d}}\\right). \\tag {8}", + "image_path": "16a353a406e47025c06ac64d273fdba61bafd60e83fb74841a842ace3ab34194.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 317, + 153, + 330 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 317, + 153, + 330 + ], + "spans": [ + { + "bbox": [ + 55, + 317, + 153, + 330 + ], + "type": "text", + "content": "3.4. Step Distillation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 335, + 296, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 335, + 296, + 407 + ], + "spans": [ + { + "bbox": [ + 55, + 335, + 296, + 407 + ], + "type": "text", + "content": "We take one step further to enhance the sampling efficiency of our model following a distribution-matching-based step distillation scheme. Following Latent Adversarial Diffusion Distillation (LADD) [60], we use a diffusion-GAN hybrid structure to distill our model into fewer steps with the optimization objective as" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 83, + 416, + 295, + 462 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 416, + 295, + 462 + ], + "spans": [ + { + "bbox": [ + 83, + 416, + 295, + 462 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\min _ {D _ {\\theta_ {T}}} \\max _ {G _ {\\theta}} \\mathbb {E} \\left[ \\left[ \\log \\left(D _ {\\theta_ {T}} \\left(x _ {t - 1}, t\\right)\\right) \\right] \\right. \\tag {9} \\\\ \\left. + \\left[ \\log \\left(1 - D _ {\\theta_ {T}} \\left(x _ {t - 1} ^ {\\prime}, t\\right)\\right) \\right] - \\mathcal {S} \\left(\\mathcal {L} _ {\\mathrm {t a s k}}, \\mathcal {L} _ {\\mathrm {k d}}\\right) \\right], \\\\ \\end{array}", + "image_path": "88a1e90ff101e536b741848c608df83a94f6f29ed98993688daeabd1be943454.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 470, + 296, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 470, + 296, + 687 + ], + "spans": [ + { + "bbox": [ + 55, + 470, + 296, + 687 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 470, + 296, + 687 + ], + "type": "inline_equation", + "content": "D_{\\theta_T}" + }, + { + "bbox": [ + 55, + 470, + 296, + 687 + ], + "type": "text", + "content": " is the discriminator model partially initialized with pretrained fewer-step teacher model " + }, + { + "bbox": [ + 55, + 470, + 296, + 687 + ], + "type": "inline_equation", + "content": "\\theta_T" + }, + { + "bbox": [ + 55, + 470, + 296, + 687 + ], + "type": "text", + "content": " (SD3.5-Large-Turbo [5]). The large-scale teacher model are only used as the feature extractor and are frozen during distillation. We only train a few linear layers in the discriminator after feature extraction. We sample " + }, + { + "bbox": [ + 55, + 470, + 296, + 687 + ], + "type": "inline_equation", + "content": "x_{t-1} \\sim q(x_{t-1}|x_0)" + }, + { + "bbox": [ + 55, + 470, + 296, + 687 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 470, + 296, + 687 + ], + "type": "inline_equation", + "content": "x_{t-1}' \\sim q(x_{t-1}|x_0')" + }, + { + "bbox": [ + 55, + 470, + 296, + 687 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 470, + 296, + 687 + ], + "type": "inline_equation", + "content": "x_0'" + }, + { + "bbox": [ + 55, + 470, + 296, + 687 + ], + "type": "text", + "content": " is the prediction of our denoising generator " + }, + { + "bbox": [ + 55, + 470, + 296, + 687 + ], + "type": "inline_equation", + "content": "G_\\theta(x_t, t)" + }, + { + "bbox": [ + 55, + 470, + 296, + 687 + ], + "type": "text", + "content": " as our student model, and " + }, + { + "bbox": [ + 55, + 470, + 296, + 687 + ], + "type": "inline_equation", + "content": "q(x)" + }, + { + "bbox": [ + 55, + 470, + 296, + 687 + ], + "type": "text", + "content": " is the forward process of diffusion model defined in Eq. 1. The objective consists of an adversarial loss to match noisy samples at time step " + }, + { + "bbox": [ + 55, + 470, + 296, + 687 + ], + "type": "inline_equation", + "content": "t-1" + }, + { + "bbox": [ + 55, + 470, + 296, + 687 + ], + "type": "text", + "content": " and the output-level distillation loss " + }, + { + "bbox": [ + 55, + 470, + 296, + 687 + ], + "type": "inline_equation", + "content": "S(\\mathcal{L}_{\\mathrm{task}}, \\mathcal{L}_{\\mathrm{kd}})" + }, + { + "bbox": [ + 55, + 470, + 296, + 687 + ], + "type": "text", + "content": " after applying timestep-aware scaling. The proposed step distillation, visualized in Fig. 7, can be interpreted as training a diffusion model with adversarial refinement and knowledge distillation, where teacher guidance serves as an additional inductive bias. This advanced step distillation empowers our compact model for high-quality generation, with only a few denoising steps." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 72, + 395, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 395, + 85 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 395, + 85 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 91, + 555, + 235 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 91, + 555, + 235 + ], + "spans": [ + { + "bbox": [ + 313, + 91, + 555, + 235 + ], + "type": "text", + "content": "Model Details. Our T2I pipeline consists of the efficient UNet (Sec. 3.1) and the efficient encoder-decoder model (Sec. 3.2). To obtain text embeddings from the input prompt, we leverage multiple text encoders, namely lightweight CLIP-L [56], CLIP-G [56], and the large Gemma-2-2b language model [68]. We follow SD3 [18] strategy to combine these three text-encoders into a unified rich textual embedding. To enable classifier-free guidance [27], we employ these embeddings with an individual drop-out probability such that we can use an arbitrary subset of the encoders during inference. This allows us to deploy one or more encoders based on the resource constraints." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 235, + 555, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 235, + 555, + 426 + ], + "spans": [ + { + "bbox": [ + 313, + 235, + 555, + 426 + ], + "type": "text", + "content": "Training Recipe. Similar as prior work [30], we use a multi-stage strategy to train our UNet model from scratch. First, we pre-train the model using the ImageNet-1K [16] at 256 resolution as described in Sec. 3.1. Second, we fine-tune this model in a progressive manner from " + }, + { + "bbox": [ + 313, + 235, + 555, + 426 + ], + "type": "inline_equation", + "content": "256 \\rightarrow 512 \\rightarrow 1024" + }, + { + "bbox": [ + 313, + 235, + 555, + 426 + ], + "type": "text", + "content": " resolutions. Third, we employ knowledge distillation with our timestep-aware scaling (Sec. 3.3) to improve the finer details in our models using a much larger teacher model (SD3.5-Large [4]) and all three text-encoders. Before KD, we initialize the student model by pre-training on ImageNet and T2I datasets, a cost-effective approach that helps the model learn pixel distributions. Finally, we obtain a few-step model through step distillation using SD3.5-Large-Turbo [5] model as the teacher. We optimize the rectified-flow [18] objective using AdamW optimizer [17] to train our UNet backbone." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 426, + 555, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 555, + 475 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 555, + 475 + ], + "type": "text", + "content": "Hyper-parameters. We sample the timesteps using the logit-normal distribution with " + }, + { + "bbox": [ + 313, + 426, + 555, + 475 + ], + "type": "inline_equation", + "content": "(0,1)" + }, + { + "bbox": [ + 313, + 426, + 555, + 475 + ], + "type": "text", + "content": " as the location and scale parameters. We use a time shift of 3 for both training and inference. See the Supplemental Materials for details." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 479, + 389, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 479, + 389, + 491 + ], + "spans": [ + { + "bbox": [ + 313, + 479, + 389, + 491 + ], + "type": "text", + "content": "4.1. Evaluation" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 498, + 554, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 554, + 617 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 554, + 617 + ], + "type": "text", + "content": "Quantitative Benchmarks. We use GenEval [21] and DPG-Bench [29] benchmarks to evaluate the text-to-image alignment of our model on short and long prompts, respectively. We report CLIP score on a 6K subset of MS-COCO validation data [43]. In addition, to measure the aesthetic quality of our model, we compute the Image Reward [75] score on selected PixArt prompts [12]. Tab. 3 lists our performance alongside existing state-of-the-art T2I baselines. We provide additional details in Supplemental Materials. We highlight the salient observations below:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 618, + 554, + 714 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 313, + 618, + 554, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 618, + 554, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 618, + 554, + 653 + ], + "type": "text", + "content": "- Our 0.38B parameter model achieves even better performance than significantly larger models such as SDXL (2.6B), Playground (2.6B), and IF-XL (5.5B)." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 654, + 553, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 553, + 688 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 553, + 688 + ], + "type": "text", + "content": "- KD non-trivially improves the prompt following ability of the base model as illustrated by an absolute five-point increase in DPG-Bench and GenEval scores." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "type": "text", + "content": "- In terms of aesthetic performance, our model has similar Image Reward scores as Playground models [37, 38]." + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "type": "text", + "content": "For simplicity, we use the " + }, + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "type": "text", + "content": " prediction in our derivation, and " + }, + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "type": "text", + "content": " prediction of rectified flow-based training would not break the formulation." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 295, + 760, + 315, + 769 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 760, + 315, + 769 + ], + "spans": [ + { + "bbox": [ + 295, + 760, + 315, + 769 + ], + "type": "text", + "content": "8003" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 72, + 70, + 226, + 159 + ], + "blocks": [ + { + "bbox": [ + 72, + 70, + 226, + 159 + ], + "lines": [ + { + "bbox": [ + 72, + 70, + 226, + 159 + ], + "spans": [ + { + "bbox": [ + 72, + 70, + 226, + 159 + ], + "type": "image", + "image_path": "94c542be18b250489df00edc63c1b6e6da90c6efc8471432510b82eadc73552b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 164, + 555, + 198 + ], + "lines": [ + { + "bbox": [ + 54, + 164, + 555, + 198 + ], + "spans": [ + { + "bbox": [ + 54, + 164, + 555, + 198 + ], + "type": "text", + "content": "Figure 8. Human Evaluation. We conduct a user study to compare images generated by our model against baselines on three attributes: aesthetic quality, text-image alignment, and realistic generations. Our model surpasses the quality of SDXL and SD3 models, while performing competitively against the teacher SD3.5-Large model." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 228, + 69, + 382, + 160 + ], + "blocks": [ + { + "bbox": [ + 228, + 69, + 382, + 160 + ], + "lines": [ + { + "bbox": [ + 228, + 69, + 382, + 160 + ], + "spans": [ + { + "bbox": [ + 228, + 69, + 382, + 160 + ], + "type": "image", + "image_path": "5be006998992a29e2e6b32849107feb33f38b96359160e05550d226632e95a86.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 384, + 70, + 539, + 160 + ], + "blocks": [ + { + "bbox": [ + 384, + 70, + 539, + 160 + ], + "lines": [ + { + "bbox": [ + 384, + 70, + 539, + 160 + ], + "spans": [ + { + "bbox": [ + 384, + 70, + 539, + 160 + ], + "type": "image", + "image_path": "e5b8927c2ba31f4da15cb262c040939b608df725b8aa28c709505bfae9842108.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 57, + 289, + 296, + 420 + ], + "blocks": [ + { + "bbox": [ + 55, + 209, + 296, + 287 + ], + "lines": [ + { + "bbox": [ + 55, + 209, + 296, + 287 + ], + "spans": [ + { + "bbox": [ + 55, + 209, + 296, + 287 + ], + "type": "text", + "content": "Table 3. Evaluation on Quantitative Benchmarks. We list the scores on GenEval, DPG-Bench, CLIP score on COCO, and Image Reward on aesthetic prompts. We report the parameters for the UNet/DiT backbone in the Param column. Throughput (samples/s) is measured on a single 80GB A100 GPU using the largest batch size supported for each model in a practical scenario to generate " + }, + { + "bbox": [ + 55, + 209, + 296, + 287 + ], + "type": "inline_equation", + "content": "1024^2" + }, + { + "bbox": [ + 55, + 209, + 296, + 287 + ], + "type": "text", + "content": " px images. Here our sampling step is set to be 28." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 289, + 296, + 420 + ], + "lines": [ + { + "bbox": [ + 57, + 289, + 296, + 420 + ], + "spans": [ + { + "bbox": [ + 57, + 289, + 296, + 420 + ], + "type": "table", + "html": "
ModelParamThroughputGenEval ↑DPG ↑CLIP ↑Image Reward ↑
PixArt-α [12]0.6B0.420.4871.10.3161.15
PixArt-Σ [13]0.6B0.460.5380.50.3171.13
SD-1.5 [1]0.9B-0.4363.20.2870.19
SD-2.1 [2]0.9B-0.5064.20.2810.29
KOALA [35]1.0B0.660.5274.30.3171.05
MicroDiT [62]1.2B-0.5275.10.3181.25
Sana [74]1.6B1.000.6684.80.3271.25
LUMINA-Next [87]2.0B0.060.4674.60.3090.88
SDXL [53]2.6B0.180.5574.70.3010.99
Playgroundv2 [37]2.6B0.180.5974.50.3171.25
Playgroundv2.5 [38]2.6B0.180.5675.50.3191.34
IF-XL [15]5.5B0.060.6175.60.3110.65
Ours w/o KD0.38B1.040.6176.30.3211.20
SnapGen (ours)0.38B1.040.6681.10.3321.32
", + "image_path": "93dff0f0c9ac723f7e379a9dd50f9ececcef32867e0446e461372f70a4d5b102.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 436, + 295, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 436, + 295, + 532 + ], + "spans": [ + { + "bbox": [ + 54, + 436, + 295, + 532 + ], + "type": "text", + "content": "Qualitative Comparison. To visually evaluate the image-text alignment and aesthetics, we compare the generated images from different T2I models in Fig. 1. We observe that many existing models fail to fully capture the full prompt and miss important elements. Further, human generations often result in smoothened-out faces, leading to the loss of details. In contrast, our model generates much more photorealistic images with better image-text alignment." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 534, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 296, + 713 + ], + "type": "text", + "content": "Human Evaluation. For a thorough comparison between baselines, we perform a user study with the widely used Parti prompts [81]. We use SDXL, SD3-M, and SD3.5-Large models as the baselines and generate images using this prompt set. We ask the users to select images with better attributes between the baselines and our model. These attributes include image-text alignment, aesthetic quality, and realistic images. Fig. 8 shows that our model convincingly outperforms SDXL on all three attributes. Our model beats SD3 on image-text alignment and realistic generations, with a tie on aesthetic quality. Compared to the teacher (SD3.5-Large), we lag the teacher a bit behind on the image-text alignment, yet our model still has better realistic generations, and yields a toss-up on aesthetic quality. With this study, we can conclude our efficient T2I pipeline" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 211, + 553, + 235 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 211, + 553, + 235 + ], + "spans": [ + { + "bbox": [ + 313, + 211, + 553, + 235 + ], + "type": "text", + "content": "achieves generation quality that is quite comparable to the SD3.5-Large teacher that has 8.1B parameters." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 236, + 555, + 368 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 236, + 555, + 368 + ], + "spans": [ + { + "bbox": [ + 313, + 236, + 555, + 368 + ], + "type": "text", + "content": "Few-Step Generation. After step-distillation (Sec. 3.4), our model can generate high-quality images within a few steps. Fig. 9 compares our model's performance before and after step-distillation, with respective GenEval scores. The results demonstrate that our model, after step distillation, achieves comparable performance to the baseline model with 28 steps, even when using only 4 or 8 steps. While the few-step generation shows slight qualitative degradation compared to the 28-step baseline, it still outperforms most existing T2I models with significantly more inference steps, such as SDXL (50 steps) and PixArt-" + }, + { + "bbox": [ + 313, + 236, + 555, + 368 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 236, + 555, + 368 + ], + "type": "text", + "content": " (100 steps)." + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 317, + 374, + 552, + 518 + ], + "blocks": [ + { + "bbox": [ + 317, + 374, + 552, + 518 + ], + "lines": [ + { + "bbox": [ + 317, + 374, + 552, + 518 + ], + "spans": [ + { + "bbox": [ + 317, + 374, + 552, + 518 + ], + "type": "image", + "image_path": "c1472c3305f61d9c032a6fabeaedf4a68aaac446d3c2036d3c05de1678461366.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 527, + 555, + 549 + ], + "lines": [ + { + "bbox": [ + 313, + 527, + 555, + 549 + ], + "spans": [ + { + "bbox": [ + 313, + 527, + 555, + 549 + ], + "type": "text", + "content": "Figure 9. Performance comparison of few-step generation for our model before (top) and after step distillation (bottom)." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 572, + 388, + 585 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 572, + 388, + 585 + ], + "spans": [ + { + "bbox": [ + 313, + 572, + 388, + 585 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "content": "In this work, we propose a novel and efficient T2I model for high-resolution generation on mobile phones. We systematically detail the process to obtain a tiny 379M parameter UNet architecture along with an efficient latent decoder. We devise a novel training method consisting of multi-stage pre-training followed by knowledge distillation from a large teacher and adversarial step distillation. With these, we achieve an extremely efficient T2I model that comprehensively outperforms many existing multi-billion parameter models such as SDXL, Lumina-Next, and Playgroundv2." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 760, + 315, + 769 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 760, + 315, + 769 + ], + "spans": [ + { + "bbox": [ + 295, + 760, + 315, + 769 + ], + "type": "text", + "content": "8004" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 57, + 72, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 72, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 57, + 72, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 91, + 294, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 294, + 123 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 294, + 123 + ], + "type": "text", + "content": "[1] Stability AI. Stable diffusion 1.5. https://huggingface.co/stable-diffusion-v1-5/stable-diffusion-v1-5, 2022.8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 124, + 294, + 156 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 124, + 294, + 156 + ], + "spans": [ + { + "bbox": [ + 61, + 124, + 294, + 156 + ], + "type": "text", + "content": "[2] Stability AI. Stable diffusion 2.1. https://huggingface.co/stabilityai/stable-diffusion-2-1, 2022.8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 158, + 294, + 179 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 158, + 294, + 179 + ], + "spans": [ + { + "bbox": [ + 62, + 158, + 294, + 179 + ], + "type": "text", + "content": "[3] Stability AI. Stable diffusion 3.5. https://github.com/Stability-AI/sd3.5, 2024.2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 180, + 294, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 180, + 294, + 213 + ], + "spans": [ + { + "bbox": [ + 62, + 180, + 294, + 213 + ], + "type": "text", + "content": "[4] Stability AI. Stable diffusion 3.5 large. https://huggingface.co/stabilityai/stable-diffusion-3.5-large, 2024.6,7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 214, + 294, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 214, + 294, + 246 + ], + "spans": [ + { + "bbox": [ + 62, + 214, + 294, + 246 + ], + "type": "text", + "content": "[5] Stability AI. Stable diffusion 3.5 large turbo. https://huggingface.co/stabilityai/stable-diffusion-3.5-large-turbo, 2024. 2, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 248, + 294, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 248, + 294, + 312 + ], + "spans": [ + { + "bbox": [ + 62, + 248, + 294, + 312 + ], + "type": "text", + "content": "[6] Vladimir Arkhipkin, Viacheslav Vasilev, Andrei Filatov, Igor Pavlov, Julia Agafonova, Nikolai Gerasimenko, Anna Averchenkova, Evelina Mironova, Anton Bukashkin, Konstantin Kulikov, et al. Kandinsky 3: Text-to-image synthesis for multifunctional generative framework. arXiv preprint arXiv:2410.21061, 2024. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 315, + 294, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 315, + 294, + 346 + ], + "spans": [ + { + "bbox": [ + 62, + 315, + 294, + 346 + ], + "type": "text", + "content": "[7] Shakhmatov Arseni, Razzhigaev Anton, Nikolich Aleksandr, Arkhipkin Vladimir, Pavlov Igor, Kuznetsov Andrey, and Denis Dimitrov. Kandinsky 2.1, 2023. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 348, + 294, + 401 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 348, + 294, + 401 + ], + "spans": [ + { + "bbox": [ + 62, + 348, + 294, + 401 + ], + "type": "text", + "content": "[8] Fan Bao, Shen Nie, Kaiwen Xue, Yue Cao, Chongxuan Li, Hang Su, and Jun Zhu. All are worth words: A vit backbone for diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 22669-22679, 2023. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 403, + 294, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 403, + 294, + 456 + ], + "spans": [ + { + "bbox": [ + 62, + 403, + 294, + 456 + ], + "type": "text", + "content": "[9] Han Cai, Junyan Li, Muyan Hu, Chuang Gan, and Song Han. Efficientvit: Lightweight multi-scale attention for high-resolution dense prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17302-17313, 2023. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 459, + 294, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 459, + 294, + 513 + ], + "spans": [ + { + "bbox": [ + 57, + 459, + 294, + 513 + ], + "type": "text", + "content": "[10] Thibault Castells, Hyoung-Kyu Song, Bo-Kyeong Kim, and Shinkook Choi. LD-Pruner: Efficient Pruning of Latent Diffusion Models using Task-Agnostic Insights. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 821–830, 2024. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 514, + 294, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 514, + 294, + 568 + ], + "spans": [ + { + "bbox": [ + 57, + 514, + 294, + 568 + ], + "type": "text", + "content": "[11] Thibault Castells, Hyoung-Kyu Song, Tairen Piao, Shinkook Choi, Bo-Kyeong Kim, Hanyoung Yim, Changgwun Lee, Jae Gon Kim, and Tae-Ho Kim. EdgeFusion: On-Device Text-to-Image Generation. arXiv preprint arXiv:2404.11925, 2024.3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 570, + 294, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 570, + 294, + 624 + ], + "spans": [ + { + "bbox": [ + 57, + 570, + 294, + 624 + ], + "type": "text", + "content": "[12] Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, et al. Pixart-α: Fast training of diffusion transformer for photorealistic text-to-image synthesis. arXiv preprint arXiv:2310.00426, 2023. 2, 3, 7, 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 625, + 294, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 625, + 294, + 679 + ], + "spans": [ + { + "bbox": [ + 57, + 625, + 294, + 679 + ], + "type": "text", + "content": "[13] Junsong Chen, Chongjian Ge, Enze Xie, Yue Wu, Lewei Yao, Xiaozhe Ren, Zhongdao Wang, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart- " + }, + { + "bbox": [ + 57, + 625, + 294, + 679 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 57, + 625, + 294, + 679 + ], + "type": "text", + "content": " sigma: Weak-to-strong training of diffusion transformer for 4k text-to-image generation. arXiv preprint arXiv:2403.04692, 2024. 2, 3, 8" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 681, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 681, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 681, + 294, + 713 + ], + "type": "text", + "content": "[14] Tri Dao and Albert Gu. Transformers are ssms: Generalized models and efficient algorithms through structured state space duality. arXiv preprint arXiv:2405.21060, 2024. 3" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 94 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 94 + ], + "type": "text", + "content": "[15] DeepFloyd. Deepfloyd. https://github.com/deep-floyd/IF, 2023. 2, 3, 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 96, + 553, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 96, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 317, + 96, + 553, + 140 + ], + "type": "text", + "content": "[16] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 2, 3, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 142, + 553, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 142, + 553, + 163 + ], + "spans": [ + { + "bbox": [ + 317, + 142, + 553, + 163 + ], + "type": "text", + "content": "[17] P Kingma Diederik. Adam: A method for stochastic optimization. (No Title), 2014. 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 164, + 553, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 164, + 553, + 229 + ], + "spans": [ + { + "bbox": [ + 317, + 164, + 553, + 229 + ], + "type": "text", + "content": "[18] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In *Forty-first International Conference on Machine Learning*, 2024. 2, 5, 6, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 231, + 553, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 231, + 553, + 252 + ], + "spans": [ + { + "bbox": [ + 317, + 231, + 553, + 252 + ], + "type": "text", + "content": "[19] Leonhard Euler. Institutionum calculi integralis. imp. Acad. imp. Sa'ent., 1768. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 254, + 553, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 254, + 553, + 308 + ], + "spans": [ + { + "bbox": [ + 317, + 254, + 553, + 308 + ], + "type": "text", + "content": "[20] Peng Gao, Le Zhuo, Ziyi Lin, Chris Liu, Junsong Chen, Ruoyi Du, Enze Xie, Xu Luo, Longtian Qiu, Yuhang Zhang, et al. Lumina-T2X: Transforming Text into Any Modality, Resolution, and Duration via Flow-based Large Diffusion Transformers. arXiv preprint arXiv:2405.05945, 2024. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 310, + 553, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 310, + 553, + 353 + ], + "spans": [ + { + "bbox": [ + 316, + 310, + 553, + 353 + ], + "type": "text", + "content": "[21] Dhruba Ghosh, Hannaneh Hajishirzi, and Ludwig Schmidt. Geneval: An object-focused framework for evaluating text-to-image alignment. Advances in Neural Information Processing Systems, 36, 2024. 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 355, + 553, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 355, + 553, + 398 + ], + "spans": [ + { + "bbox": [ + 316, + 355, + 553, + 398 + ], + "type": "text", + "content": "[22] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 3, 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 399, + 553, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 399, + 553, + 443 + ], + "spans": [ + { + "bbox": [ + 317, + 399, + 553, + 443 + ], + "type": "text", + "content": "[23] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 445, + 553, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 445, + 553, + 476 + ], + "spans": [ + { + "bbox": [ + 317, + 445, + 553, + 476 + ], + "type": "text", + "content": "[24] Alex Henry, Prudhvi Raj Dachapally, Shubham Pawar, and Yuxuan Chen. Query-key normalization for transformers. arXiv preprint arXiv:2010.04245, 2020. 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 478, + 553, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 478, + 553, + 532 + ], + "spans": [ + { + "bbox": [ + 317, + 478, + 553, + 532 + ], + "type": "text", + "content": "[25] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 534, + 553, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 534, + 553, + 555 + ], + "spans": [ + { + "bbox": [ + 317, + 534, + 553, + 555 + ], + "type": "text", + "content": "[26] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 5" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 557, + 553, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 557, + 553, + 578 + ], + "spans": [ + { + "bbox": [ + 317, + 557, + 553, + 578 + ], + "type": "text", + "content": "[27] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 7" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 317, + 580, + 553, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 580, + 553, + 611 + ], + "spans": [ + { + "bbox": [ + 317, + 580, + 553, + 611 + ], + "type": "text", + "content": "[28] Andrew G Howard. Mobilenets: Efficient convolutional neural networks for mobile vision applications. arXiv preprint arXiv:1704.04861, 2017. 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 317, + 614, + 553, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 614, + 553, + 656 + ], + "spans": [ + { + "bbox": [ + 317, + 614, + 553, + 656 + ], + "type": "text", + "content": "[29] Xiwei Hu, Rui Wang, Yixiao Fang, Bin Fu, Pei Cheng, and Gang Yu. Ella: Equip diffusion models with llm for enhanced semantic alignment. arXiv preprint arXiv:2403.05135, 2024. 7" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 317, + 658, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 658, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 317, + 658, + 553, + 712 + ], + "type": "text", + "content": "[30] Anil Kag, Huseyin Coskun, Jierun Chen, Junli Cao, Willi Menapace, Aliaksandr Siarohin, Sergey Tulyakov, and Jian Ren. Ascan: Asymmetric convolution-attention networks for efficient recognition and generation. arXiv preprint arXiv:2411.04967, 2024. 3, 7" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 761, + 315, + 769 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 761, + 315, + 769 + ], + "spans": [ + { + "bbox": [ + 295, + 761, + 315, + 769 + ], + "type": "text", + "content": "8005" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 126 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 126 + ], + "type": "text", + "content": "[31] Angelos Katharopoulos, Apoorv Vyas, Nikolaos Pappas, and François Fleuret. Transformers are rnns: Fast autoregressive transformers with linear attention. In International conference on machine learning, pages 5156-5165. PMLR, 2020. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 294, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 294, + 183 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 294, + 183 + ], + "type": "text", + "content": "[32] Bo-Kyeong Kim, Hyoung-Kyu Song, Thibault Castells, and Shinkook Choi. Bk-sdm: Architecturally Compressed Stable Diffusion for Efficient Text-to-Image Generation. In Workshop on Efficient Systems for Foundation Models@ ICML2023, 2023. 2, 3, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 185, + 294, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 185, + 294, + 217 + ], + "spans": [ + { + "bbox": [ + 56, + 185, + 294, + 217 + ], + "type": "text", + "content": "[33] Diederik P Kingma. Auto-encoding variational bayes. 2nd International Conference on Learning Representations, 2014. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 220, + 294, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 220, + 294, + 273 + ], + "spans": [ + { + "bbox": [ + 56, + 220, + 294, + 273 + ], + "type": "text", + "content": "[34] Tuomas Kynkänniemi, Miika Aittala, Tero Karras, Samuli Laine, Timo Aila, and Jaakko Lehtinen. Applying guidance in a limited interval improves sample and distribution quality in diffusion models. arXiv preprint arXiv:2404.07724, 2024.5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 276, + 294, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 276, + 294, + 319 + ], + "spans": [ + { + "bbox": [ + 56, + 276, + 294, + 319 + ], + "type": "text", + "content": "[35] Youngwan Lee, Kwanyong Park, Yoorhim Cho, Yong-Ju Lee, and Sung Ju Hwang. Koala: Empirical lessons toward memory-efficient and fast diffusion models for text-to-image synthesis, 2023. 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 321, + 294, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 321, + 294, + 342 + ], + "spans": [ + { + "bbox": [ + 56, + 321, + 294, + 342 + ], + "type": "text", + "content": "[36] Daiqing Li, Aleks Kamko, Ali Sabet, Ehsan Akhgari, Linmiao Xu, and Suhail Doshi. Playground v1, . 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 344, + 294, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 344, + 294, + 365 + ], + "spans": [ + { + "bbox": [ + 56, + 344, + 294, + 365 + ], + "type": "text", + "content": "[37] Daiqing Li, Aleks Kamko, Ali Sabet, Ehsan Akhgari, Linmiao Xu, and Suhail Doshi. Playground v2, . 7, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 367, + 294, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 367, + 294, + 421 + ], + "spans": [ + { + "bbox": [ + 56, + 367, + 294, + 421 + ], + "type": "text", + "content": "[38] Daiqing Li, Aleks Kamko, Ehsan Akhgari, Ali Sabet, Linmiao Xu, and Suhail Doshi. Playground V2. 5: Three Insights towards Enhancing Aesthetic Quality in Text-to-Image Generation. arXiv preprint arXiv:2402.17245, 2024. 3, 7, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 423, + 294, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 423, + 294, + 489 + ], + "spans": [ + { + "bbox": [ + 56, + 423, + 294, + 489 + ], + "type": "text", + "content": "[39] Hao Li, Yang Zou, Ying Wang, Orchid Majumder, Yusheng Xie, R Manmatha, Ashwin Swaminathan, Zhuowen Tu, Stefano Ermon, and Stefano Soatto. On the scalability of diffusion-based text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9400-9409, 2024. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 491, + 294, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 491, + 294, + 544 + ], + "spans": [ + { + "bbox": [ + 56, + 491, + 294, + 544 + ], + "type": "text", + "content": "[40] Yanyu Li, Huan Wang, Qing Jin, Ju Hu, Pavlo Chemerys, Yun Fu, Yanzhi Wang, Sergey Tulyakov, and Jian Ren. Snapfusion: Text-to-Image Diffusion Model on Mobile Devices within Two Seconds. Advances in Neural Information Processing Systems, 36, 2024. 2, 3, 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 546, + 294, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 546, + 294, + 611 + ], + "spans": [ + { + "bbox": [ + 56, + 546, + 294, + 611 + ], + "type": "text", + "content": "[41] Zhimin Li, Jianwei Zhang, Qin Lin, Jiangfeng Xiong, Yanxin Long, Xinchi Deng, Yingfang Zhang, Xingchao Liu, Minbin Huang, Zedong Xiao, et al. Hunyuan-DiT: A Powerful Multi-Resolution Diffusion Transformer with Fine-Grained Chinese Understanding. arXiv preprint arXiv:2405.08748, 2024. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 613, + 294, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 613, + 294, + 644 + ], + "spans": [ + { + "bbox": [ + 56, + 613, + 294, + 644 + ], + "type": "text", + "content": "[42] Shanchuan Lin, Anran Wang, and Xiao Yang. SDXL-Lightning: Progressive Adversarial Diffusion Distillation, 2024. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 647, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 294, + 713 + ], + "type": "text", + "content": "[43] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 5, 7" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "text", + "content": "[44] Yaron Lipman, Ricky TQ Chen, Heli Ben-Hamu, Maximilian Nickel, and Matt Le. Flow matching for generative modeling. arXiv preprint arXiv:2210.02747, 2022. 2, 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 108, + 553, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 108, + 553, + 162 + ], + "spans": [ + { + "bbox": [ + 316, + 108, + 553, + 162 + ], + "type": "text", + "content": "[45] Bingchen Liu, Ehsan Akhgari, Alexander Visheratin, Aleks Kamko, Linmiao Xu, Shivam Shrirao, Joao Souza, Suhail Doshi, and Daiqing Li. Playground v3: Improving text-to-image alignment with deep-fusion large language models. arXiv preprint arXiv:2409.10695, 2024. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 163, + 553, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 553, + 195 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 553, + 195 + ], + "type": "text", + "content": "[46] Songhua Liu, Weihao Yu, Zhenxiong Tan, and Xinchao Wang. Linfusion: 1gpu, 1 minute, 16k image. arXiv preprint arXiv:2409.02097, 2024. 2, 3, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 198, + 553, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 198, + 553, + 230 + ], + "spans": [ + { + "bbox": [ + 316, + 198, + 553, + 230 + ], + "type": "text", + "content": "[47] Xingchao Liu, Chengyue Gong, and Qiang Liu. Flow straight and fast: Learning to generate and transfer data with rectified flow. arXiv preprint arXiv:2209.03003, 2022. 2, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 232, + 553, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 232, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 316, + 232, + 553, + 285 + ], + "type": "text", + "content": "[48] Xian Liu, Jian Ren, Aliaksandr Siarohin, Ivan Skorokhodov, Yanyu Li, Dahua Lin, Xihui Liu, Ziwei Liu, and Sergey Tulyakov. Hyperhuman: Hyper-realistic human generation with latent structural diffusion. arXiv preprint arXiv:2310.08579, 2023. 2, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 288, + 553, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 288, + 553, + 342 + ], + "spans": [ + { + "bbox": [ + 316, + 288, + 553, + 342 + ], + "type": "text", + "content": "[49] Nanye Ma, Mark Goldstein, Michael S Albergo, Nicholas M Boffi, Eric Vanden-Eijnden, and Saining Xie. Sit: Exploring flow and diffusion-based generative models with scalable interpolant transformers. arXiv preprint arXiv:2401.08740, 2024. 2, 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 344, + 553, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 344, + 553, + 419 + ], + "spans": [ + { + "bbox": [ + 316, + 344, + 553, + 419 + ], + "type": "text", + "content": "[50] Willi Menapace, Aliaksandr Siarohin, Ivan Skorokhodov, Ekaterina Deyneka, Tsai-Shien Chen, Anil Kag, Yuwei Fang, Aleksei Stoliar, Elisa Ricci, Jian Ren, et al. Snap video: Scaled spatiotemporal transformers for text-to-video synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7038-7048, 2024. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 422, + 376, + 433 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 422, + 376, + 433 + ], + "spans": [ + { + "bbox": [ + 316, + 422, + 376, + 433 + ], + "type": "text", + "content": "[51] OpenAI. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 434, + 553, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 434, + 553, + 477 + ], + "spans": [ + { + "bbox": [ + 316, + 434, + 553, + 477 + ], + "type": "text", + "content": "[52] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4195-4205, 2023. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 479, + 553, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 479, + 553, + 533 + ], + "spans": [ + { + "bbox": [ + 316, + 479, + 553, + 533 + ], + "type": "text", + "content": "[53] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 2, 3, 5, 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 535, + 553, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 535, + 553, + 589 + ], + "spans": [ + { + "bbox": [ + 316, + 535, + 553, + 589 + ], + "type": "text", + "content": "[54] Adam Polyak, Amit Zohar, Andrew Brown, Andros Tjandra, Animesh Sinha, Ann Lee, Apoorv Vyas, Bowen Shi, Chih-Yao Ma, Ching-Yao Chuang, et al. Movie gen: A cast of media foundation models. arXiv preprint arXiv:2410.13720, 2024. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 591, + 553, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 591, + 553, + 645 + ], + "spans": [ + { + "bbox": [ + 316, + 591, + 553, + 645 + ], + "type": "text", + "content": "[55] Danfeng Qin, Chas Leichner, Manolis Delakis, Marco Fornoni, Shixin Luo, Fan Yang, Weijun Wang, Colby Banbury, Chengxi Ye, Berkin Akin, et al. Mobilenetv4-universal models for the mobile ecosystem. arXiv preprint arXiv:2404.10518, 2024.3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 647, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 712 + ], + "type": "text", + "content": "[56] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 7" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 761, + 315, + 769 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 761, + 315, + 769 + ], + "spans": [ + { + "bbox": [ + 295, + 761, + 315, + 769 + ], + "type": "text", + "content": "8006" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "text", + "content": "[57] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In Proceedings of the 38th International Conference on Machine Learning, pages 8821-8831. PMLR, 2021. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 128, + 295, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 128, + 295, + 183 + ], + "spans": [ + { + "bbox": [ + 56, + 128, + 295, + 183 + ], + "type": "text", + "content": "[58] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 3, 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 183, + 294, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 183, + 294, + 249 + ], + "spans": [ + { + "bbox": [ + 56, + 183, + 294, + 249 + ], + "type": "text", + "content": "[59] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S. Sara Mahdavi, Rapha Gontijo Lopes, Tim Salimans, Jonathan Ho, David J Fleet, and Mohammad Norouzi. Photorealistic text-to-image diffusion models with deep language understanding, 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 250, + 294, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 250, + 294, + 293 + ], + "spans": [ + { + "bbox": [ + 56, + 250, + 294, + 293 + ], + "type": "text", + "content": "[60] Axel Sauer, Frederic Boesel, Tim Dockhorn, Andreas Blattmann, Patrick Esser, and Robin Rombach. Fast high-resolution image synthesis with latent adversarial diffusion distillation. arXiv preprint arXiv:2403.12015, 2024. 3, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 293, + 294, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 293, + 294, + 336 + ], + "spans": [ + { + "bbox": [ + 56, + 293, + 294, + 336 + ], + "type": "text", + "content": "[61] Axel Sauer, Dominik Lorenz, Andreas Blattmann, and Robin Rombach. Adversarial diffusion distillation. In European Conference on Computer Vision, pages 87-103. Springer, 2024. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 338, + 294, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 338, + 294, + 381 + ], + "spans": [ + { + "bbox": [ + 56, + 338, + 294, + 381 + ], + "type": "text", + "content": "[62] Vikash Sehwag, Xianghao Kong, Jingtao Li, Michael Spranger, and Lingjuan Lyu. Stretching each dollar: Diffusion training from scratch on a micro-budget. arXiv preprint arXiv:2407.15811, 2024. 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 382, + 294, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 382, + 294, + 403 + ], + "spans": [ + { + "bbox": [ + 56, + 382, + 294, + 403 + ], + "type": "text", + "content": "[63] Noam Shazeer. Fast transformer decoding: One write-head is all you need. arXiv preprint arXiv:1911.02150, 2019. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 404, + 294, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 404, + 294, + 470 + ], + "spans": [ + { + "bbox": [ + 56, + 404, + 294, + 470 + ], + "type": "text", + "content": "[64] Shelly Sheynin, Adam Polyak, Uriel Singer, Yuval Kirstain, Amit Zohar, Oron Ashual, Devi Parikh, and Yaniv Taigman. Emu edit: Precise image editing via recognition and generation tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8871-8879, 2024. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 471, + 294, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 471, + 294, + 503 + ], + "spans": [ + { + "bbox": [ + 56, + 471, + 294, + 503 + ], + "type": "text", + "content": "[65] Yuda Song, Zehao Sun, and Xuanwu Yin. Sdxs: Realtime one-step latent diffusion models with image conditions. arXiv preprint arXiv:2403.16627, 2024. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 503, + 294, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 503, + 294, + 546 + ], + "spans": [ + { + "bbox": [ + 56, + 503, + 294, + 546 + ], + "type": "text", + "content": "[66] Jianlin Su, Murtadha Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu. Roformer: Enhanced transformer with rotary position embedding. Neurocomputing, 568:127063, 2024. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 547, + 294, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 547, + 294, + 591 + ], + "spans": [ + { + "bbox": [ + 56, + 547, + 294, + 591 + ], + "type": "text", + "content": "[67] Yang Sui, Yanyu Li, Anil Kag, Yerlan Idelbayev, Junli Cao, Ju Hu, Dhritiman Sagar, Bo Yuan, Sergey Tulyakov, and Jian Ren. Bitsfusion: 1.99 bits weight quantization of diffusion model. arXiv preprint arXiv:2406.04333, 2024. 2, 3, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 592, + 294, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 592, + 294, + 656 + ], + "spans": [ + { + "bbox": [ + 56, + 592, + 294, + 656 + ], + "type": "text", + "content": "[68] Gemma Team, Morgane Riviere, Shreya Pathak, Pier Giuseppe Sessa, Cassidy Hardin, Surya Bhupatiraju, Léonard Hussenot, Thomas Mesnard, Bobak Shahriari, Alexandre Ramé, et al. Gemma 2: Improving open language models at a practical size. arXiv preprint arXiv:2408.00118, 2024. 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 658, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 294, + 689 + ], + "type": "text", + "content": "[69] Kcolors Team. Kcolors: Effective Training of Diffusion Model for Photorealistic Text-to-Image Synthesis. arXiv preprint, 2024. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 690, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 690, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 690, + 294, + 713 + ], + "type": "text", + "content": "[70] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 3" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 72, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 315, + 72, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 72, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 315, + 72, + 553, + 106 + ], + "type": "text", + "content": "[71] Zhendong Wang, Huangjie Zheng, Pengcheng He, Weizhu Chen, and Mingyuan Zhou. Diffusion-gan: Training gans with diffusion. arXiv preprint arXiv:2206.02262, 2022. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 315, + 108, + 553, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 108, + 553, + 152 + ], + "spans": [ + { + "bbox": [ + 315, + 108, + 553, + 152 + ], + "type": "text", + "content": "[72] Shitao Xiao, Yueze Wang, Junjie Zhou, Huaying Yuan, Xingrun Xing, Ruiran Yan, Shuting Wang, Tiejun Huang, and Zheng Liu. Omnigen: Unified image generation. arXiv preprint arXiv:2409.11340, 2024. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 315, + 153, + 553, + 187 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 153, + 553, + 187 + ], + "spans": [ + { + "bbox": [ + 315, + 153, + 553, + 187 + ], + "type": "text", + "content": "[73] Zhisheng Xiao, Karsten Kreis, and Arash Vahdat. Tackling the generative learning trilemma with denoising diffusion gans. arXiv preprint arXiv:2112.07804, 2021. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 315, + 189, + 553, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 189, + 553, + 243 + ], + "spans": [ + { + "bbox": [ + 315, + 189, + 553, + 243 + ], + "type": "text", + "content": "[74] Enze Xie, Junsong Chen, Junyu Chen, Han Cai, Yujun Lin, Zhekai Zhang, Muyang Li, Yao Lu, and Song Han. Sana: Efficient high-resolution image synthesis with linear diffusion transformers. arXiv preprint arXiv:2410.10629, 2024. 2, 3, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 315, + 245, + 553, + 301 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 245, + 553, + 301 + ], + "spans": [ + { + "bbox": [ + 315, + 245, + 553, + 301 + ], + "type": "text", + "content": "[75] Jiazheng Xu, Xiao Liu, Yuchen Wu, Yuxuan Tong, Qinkai Li, Ming Ding, Jie Tang, and Yuxiao Dong. Imagereward: Learning and evaluating human preferences for text-to-image generation. Advances in Neural Information Processing Systems, 36, 2024. 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 315, + 302, + 553, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 302, + 553, + 357 + ], + "spans": [ + { + "bbox": [ + 315, + 302, + 553, + 357 + ], + "type": "text", + "content": "[76] Yanwu Xu, Mingming Gong, Shaoan Xie, Wei Wei, Matthias Grundmann, Kayhan Batmanghelich, and Tingbo Hou. Semi-implicit denoising diffusion models (siddms). Advances in neural information processing systems, 36:17383, 2023. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 315, + 358, + 553, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 358, + 553, + 415 + ], + "spans": [ + { + "bbox": [ + 315, + 358, + 553, + 415 + ], + "type": "text", + "content": "[77] Yanwu Xu, Yang Zhao, Zhisheng Xiao, and Tingbo Hou. Ufogen: You forward once large scale text-to-image generation via diffusion gans. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8196-8206, 2024. 2, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 315, + 417, + 553, + 471 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 417, + 553, + 471 + ], + "spans": [ + { + "bbox": [ + 315, + 417, + 553, + 471 + ], + "type": "text", + "content": "[78] Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, et al. Cogvideox: Text-to-video diffusion models with an expert transformer. arXiv preprint arXiv:2408.06072, 2024. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 315, + 472, + 553, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 472, + 553, + 518 + ], + "spans": [ + { + "bbox": [ + 315, + 472, + 553, + 518 + ], + "type": "text", + "content": "[79] Tianwei Yin, Michael Gharbi, Taesung Park, Richard Zhang, Eli Shechtman, Fredo Durand, and William T Freeman. Improved distribution matching distillation for fast image synthesis. arXiv preprint arXiv:2405.14867, 2024. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 315, + 520, + 553, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 520, + 553, + 575 + ], + "spans": [ + { + "bbox": [ + 315, + 520, + 553, + 575 + ], + "type": "text", + "content": "[80] Tianwei Yin, Michael Gharbi, Richard Zhang, Eli Shechtman, Fredo Durand, William T Freeman, and Taesung Park. One-step diffusion with distribution matching distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6613-6623, 2024. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 315, + 576, + 553, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 576, + 553, + 632 + ], + "spans": [ + { + "bbox": [ + 315, + 576, + 553, + 632 + ], + "type": "text", + "content": "[81] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, et al. Scaling autoregressive models for content-rich text-to-image generation. Transactions on Machine Learning Research. 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 315, + 633, + 553, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 633, + 553, + 666 + ], + "spans": [ + { + "bbox": [ + 315, + 633, + 553, + 666 + ], + "type": "text", + "content": "[82] Biao Zhang and Rico Sennrich. Root mean square layer normalization. Advances in Neural Information Processing Systems, 32, 2019. 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 315, + 669, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 669, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 315, + 669, + 553, + 713 + ], + "type": "text", + "content": "[83] Dingkun Zhang, Sijia Li, Chen Chen, Qingsong Xie, and Haonan Lu. Laptop-Diff: Layer Pruning and Normalized Distillation for Compressing Diffusion Models. arXiv preprint arXiv:2404.11098, 2024. 2" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 760, + 315, + 769 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 760, + 315, + 769 + ], + "spans": [ + { + "bbox": [ + 295, + 760, + 315, + 769 + ], + "type": "text", + "content": "8007" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 262 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "type": "text", + "content": "[84] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 295, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 295, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 295, + 162 + ], + "type": "text", + "content": "[85] Zhixing Zhang, Ligong Han, Arnab Ghosh, Dimitris N Metaxas, and Jian Ren. Sine: Single image editing with text-to-image diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6027-6037, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 294, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 294, + 206 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 294, + 206 + ], + "type": "text", + "content": "[86] Yang Zhao, Yanwu Xu, Zhisheng Xiao, and Tingbo Hou. Mobilediffusion: Subsecond text-to-image generation on mobile devices. arXiv preprint arXiv:2311.16567, 2023. 2, 3, 4, 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 208, + 294, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 208, + 294, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 208, + 294, + 262 + ], + "type": "text", + "content": "[87] Le Zhuo, Ruoyi Du, Han Xiao, Yangguang Li, Dongyang Liu, Rongjie Huang, Wenze Liu, Lirui Zhao, Fu-Yun Wang, Zhanyu Ma, et al. Lumina-last: Making lumina-t2x stronger and faster with next-dit. arXiv preprint arXiv:2406.18583, 2024. 2, 8" + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 761, + 315, + 769 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 761, + 315, + 769 + ], + "spans": [ + { + "bbox": [ + 295, + 761, + 315, + 769 + ], + "type": "text", + "content": "8008" + } + ] + } + ], + "index": 5 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/SnowMaster_ Comprehensive Real-world Image Desnowing via MLLM with Multi-Model Feedback Optimization/01611a60-cb24-4ca8-93fd-c47c38025543_content_list.json b/2025/SnowMaster_ Comprehensive Real-world Image Desnowing via MLLM with Multi-Model Feedback Optimization/01611a60-cb24-4ca8-93fd-c47c38025543_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..812085d36020c04a80010f1f2f5d9d4849d99406 --- /dev/null +++ b/2025/SnowMaster_ Comprehensive Real-world Image Desnowing via MLLM with Multi-Model Feedback Optimization/01611a60-cb24-4ca8-93fd-c47c38025543_content_list.json @@ -0,0 +1,1623 @@ +[ + { + "type": "text", + "text": "SnowMaster: Comprehensive Real-world Image Desnowing via MLLM with Multi-Model Feedback Optimization", + "text_level": 1, + "bbox": [ + 179, + 130, + 818, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jianyu Lai $^{1,*}$ Sixiang Chen $^{1,*}$ Yunlong Lin $^{2}$ Tian Ye $^{1}$ \nYun Liu $^{3}$ Song Fei $^{1}$ Zhaohu Xing $^{1}$ Hongtao Wu $^{1}$ Weiming Wang $^{4}$ Lei Zhu $^{1,5,\\dagger}$ $^{1}$ The Hong Kong University of Science and Technology (GZ) $^{2}$ Xiamen University \n $^{3}$ Southwestern University $^{4}$ Hong Kong Metropolitan University \n $^{5}$ The Hong Kong University of Science and Technology \nProject page: https://alexlai2860.github.io/SnowMaster", + "bbox": [ + 143, + 202, + 857, + 313 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 347, + 326, + 364 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Snowfall presents significant challenges for visual data processing, necessitating specialized desnowing algorithms. However, existing models often fail to generalize effectively due to their heavy reliance on synthetic datasets. Furthermore, current real-world snowfall datasets are limited in scale and lack dedicated evaluation metrics designed specifically for snowfall degradation, thus hindering the effective integration of real snowy images into model training to reduce domain gaps. To address these challenges, we first introduce RealSnow10K, a large-scale, high-quality dataset consisting of over 10,000 annotated real-world snowy images. In addition, we curate a preference dataset comprising 36,000 expert-ranked image pairs, enabling the adaptation of multimodal large language models (MLLMs) to better perceive snowy image quality through our innovative Multi-Model Preference Optimization (MMPO). Finally, we propose the SnowMaster, which employs MMPO-enhanced MLLM to perform accurate snowy image evaluation and pseudo-label filtering for semi-supervised training. Experiments demonstrate that SnowMaster delivers superior desnowing performance under real-world conditions.", + "bbox": [ + 88, + 380, + 483, + 698 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 713, + 220, + 728 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Snowfall, a common form of adverse weather degradation, reduces image visibility and adversely impacts downstream tasks such as image recognition [2, 61] and segmentation [1, 51]. The inherent complexity of snow poses a challenge to conventional image restoration techniques, necessitating the development of specialized snow removal (desnowing) algorithms to enhance image clarity.", + "bbox": [ + 89, + 739, + 482, + 845 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Although current snow removal models achieve impressive performance on benchmark datasets [7, 9, 12, 14, 16],", + "bbox": [ + 89, + 845, + 482, + 876 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/cd2fbfead21ced4a3c83c22752542120b1f6d887472dfca22a76b344a8b7032c.jpg", + "image_caption": [ + "Figure 1. Comparison of various image desnowing evaluation metrics shows that Q-Instruct [47], enhanced by our MMPO, most accurately perceives differences in snow density between images. Leveraging accurate evaluation model, we optimize our semi-supervised network, achieving superior desnowing performance." + ], + "image_footnote": [], + "bbox": [ + 511, + 345, + 910, + 496 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "they often struggle to generalize to real-world scenarios. This limitation stems primarily from their only reliance on synthetic snowfall datasets for training. These datasets, generated using simplified physical models, struggle to capture the intricate details of real-world snow degradations, resulting in a domain gap between synthetic and real data. This gap hinders model optimization and impairs generalization to real-world scenarios. Consequently, we pose the question: How can real-world data be effectively utilized to enhance the model's generalization performance on real-world samples?", + "bbox": [ + 511, + 579, + 906, + 746 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Incorporating real-world snowy images into training is an direct and effective approach to enhance model generalization. However, accurate evaluation of these images is a prerequisite for their successful integration into the training process [4, 48, 50]. Yet, evaluating realistic snowfall images remains a significant challenge due to the absence of ground truth for reference evaluation, previous methods often rely on non-reference metrics [31, 32, 43, 57] to assess image restoration quality. These metrics, influenced by numerous factors, fail to accurately isolate and evalu-", + "bbox": [ + 511, + 750, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal Contribution; † Corresponding Author.", + "bbox": [ + 107, + 886, + 354, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "4302", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ate weather-specific degradation in snowfall images. In this paper, we claim that accurate assessment of snowy images requires a joint effort from both low-level visual and semantic perspectives. Recent advancements in multimodal large language models (MLLMs), such as Q-Instruct [50] and Depict-QA [58], offer promising avenues for developing evaluation frameworks due to their ability to assess low-level visual features from a semantic perspective. However, these models currently lack the specialized perception capabilities required to reliably evaluate snowfall-specific characteristics. Conversely, while some more advanced multimodal large language models exhibit better understanding of snowy scenes, they are not fine-tuned for detailed low-level visual evaluation, limiting the fundamental premise as powerful evaluators on the snow removal task.", + "bbox": [ + 89, + 90, + 480, + 316 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To tackle these challenges, we first establish a comprehensive evaluation framework designed to capture the distinct characteristics of snow in natural images through multi-perspective feedback optimization tailored to snow scenes. To support this framework, we curate and annotate a dataset comprising over 10,000 real-world snow images sourced from the Internet. Additionally, while Direct Preference Optimization (DPO) [36] provides a strong foundation for preference-based learning, we observe that different models exhibit unique strengths across various evaluation perspectives, and assessing snowfall requires a multidimensional approach to ensure robust and reliable outcomes. Therefore, we introduce the Multi-Model Preference Optimization (MMPO) framework, integrating task-specific outputs from multiple multimodal large language models (MLLMs) to construct a snowfall preference dataset. This dataset enables the fine-tuning of Q-Instruct, significantly enhancing its ability to perceive and evaluate snowfall. Finally, we integrate the MMPO fine-tuned Q-Instruct model as an unsupervised evaluation metric to update the pseudolabel database and employ a mean teacher semi-supervised framework to train existing snow removal models, thereby improving their generalization performance in real-world scenarios. Our contributions can be summarized as follows:", + "bbox": [ + 91, + 320, + 483, + 681 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Real-World Snowfall Dataset: We introduce Real-Snow10K, a large-scale, high-quality dataset comprising over 10,000 annotated real-world snow scene images. To the best of our knowledge, this is the largest dataset of real-world image desnowing, providing a robust foundation for model training and evaluation.", + "- Snowfall Preference Dataset: From RealSnow10K, we select 1.5k real images and engage human experts to rank the outputs of multiple models based on preference. Utilizing these rankings, we construct a human preference dataset containing 36k preference pairs, designed to address two key evaluation questions.", + "- MMPO for Snowfall Evaluation: We employ Multi-Model Preference Optimization (MMPO) to enhance the" + ], + "bbox": [ + 89, + 688, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "perception capabilities of the MLLM on snowfall images. By leveraging carefully designed multiple-choice questions, this MLLM can be used to accurately evaluate the snow removal effect of images.", + "bbox": [ + 526, + 90, + 903, + 151 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- Semi-Supervised Training Framework: We integrate the fine-tuned MLLMs into a mean teacher framework, enabling high-quality pseudo-label screening for semi-supervised training. The results of SnowMaster exhibits strong generalization capabilities, achieving state-of-the-art performance on real-world snowfall datasets.", + "bbox": [ + 513, + 152, + 903, + 242 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 513, + 256, + 663, + 272 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Single Image Desnowing. Snowflakes exhibit diverse sizes and shapes, complicating image restoration compared to other degradations [3, 8, 22-24, 26-28, 40, 49, 54]. Early methods, such as Pei et al. [34], relied on manual feature extraction but struggled in complex scenes. With deep learning advancements, More and more methods focus on snow removal tasks [6, 7, 10, 12, 14, 16, 17, 59]. Liu et al. [30] introduced DesnowNet, a multi-stage network addressing transparent and opaque snow blocks. Chen et al. [13] proposed HDCWNet, utilizing dual-tree complex wavelet transform (DTCWT) and conflicting channel loss (CC Loss) to enhance snow removal accuracy, though high-frequency information recognition remains challenging. Recent approaches integrate semantic and geometric information with Transformer architectures. Zhang et al. [59] developed DDMSNet, combining semantic segmentation and depth estimation for improved snow shape recognition. Cheng et al. [17] introduced SMGARN, using snow masks guided by self-pixel and cross-pixel attention. TransWeather [42], $\\mathrm{T}^3$ -DiffWeather [11] and other adverse weather restoration methods [15, 25, 33, 55, 56] also advance robustness in extreme weather conditions.", + "bbox": [ + 511, + 279, + 906, + 609 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Multi-modality Large Language Models. Multimodal large language models (MLLMs) revolutionized computer vision by integrating visual and textual information. Models such as GPT-4o, MiniCPM-V [53], and LLava [29] excelled in high-level tasks like image description, visual question answering, and document understanding. However, their performance in low-level tasks, particularly image quality assessment (IQA), remained limited due to insufficient detail perception. Recent studies, such as Q-Instruct [50] and DepictQA [58], addressed this gap by incorporating low-level visual tasks into MLLM training datasets. These approaches leveraged human feedback on low-level features, such as clarity, brightness, and texture details, to improve IQA accuracy. Despite these advancements, MLLMs still struggled with weather-related degradations, as existing datasets were not fine-tuned for such conditions.", + "bbox": [ + 511, + 612, + 908, + 852 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Score-based IQA Methods. Score-based IQA methods aim to assess image quality by assigning numerical scores that align with human perception. Traditional methods,", + "bbox": [ + 511, + 854, + 906, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "4303", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/a2dd3875f4fc6612db5df45d776d852b336abe1d344ec7d8ca4b859b10bbb4fb.jpg", + "image_caption": [ + "Figure 2. Dataset screening and evaluation. Through a multi-round dialogue mechanism, the multimodal large language model can replace humans to complete the quality screening and content evaluation of input images." + ], + "image_footnote": [], + "bbox": [ + 96, + 88, + 903, + 311 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "such as SSIM [46] and PSNR, relied on statistical metrics to evaluate distortions or pixel-level differences. While effective in controlled environments, they often failed to capture the complexities of real-world images. Deep learning models, such as NIMA [41] and HyperIQA [39], improved performance by regressing directly to Mean Opinion Scores (MOS) derived from human ratings. However, these models faced challenges in out-of-distribution (OOD) generalization, as unseen distortions or diverse content types significantly impacted their accuracy. Recent advancements leveraged vision-language models like CLIP [35], with methods such as CLIP-IQA+ [45] and LIQE [60] utilizing visual-text alignment for few-shot and multitask learning. These approaches demonstrated enhanced robustness across datasets but did not fully capture human-perceived quality due to their reliance on visual-text similarity scores. To address these limitations, Large Multi-Modality Models (LMMs) like Q-Align [47] simulated human rating practices by using discrete text-defined rating levels (e.g., \"excellent\", \"poor\") and converted qualitative ratings into quantitative scores through probability-based averaging. This approach enhanced both accuracy and interpretability, particularly in OOD scenarios.", + "bbox": [ + 88, + 356, + 485, + 704 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Dataset Construction", + "text_level": 1, + "bbox": [ + 89, + 712, + 294, + 728 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "High-quality real data is the foundation of our approach. We propose RealSnow10K, the largest real snow dataset available. In this section, we describe the process of building the dataset, including image collection, image evaluation and classification, and the dataset preference process.", + "bbox": [ + 89, + 734, + 483, + 811 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Image Collection, Evaluation and Classification", + "text_level": 1, + "bbox": [ + 89, + 818, + 483, + 834 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The first step in building a dataset is to collect images from various sources. We collect more than 100,000 snow scene images from the Internet and screen out over 10,000 high-quality images. To ensure the high quality of the dataset,", + "bbox": [ + 89, + 839, + 483, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "we need to screen and classify the collected images. Images collected from the Internet often have multiple interferences. In addition to the low quality caused by blur and noise in the image itself, it may also be irrelevant or deviate from the keywords we set, which is often difficult to reflect through existing indicators. To address the above problems, traditional dataset production uses manual screening. Although this method is accurate, it is too inefficient when facing a large number of images. To conduct fast and efficient review, we introduce multimodal large language models (MLLMs) with human visual perception capabilities as a reference for evaluating image quality and content.", + "bbox": [ + 511, + 356, + 906, + 537 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We design a multi-round dialogue mechanism to screen out unqualified images and evaluate the content of the images. Specifically, the prompt words used in the first and second rounds of dialogue are set to \"Is this picture a real photograph? Please answer with only 'yes' or 'no'.\" and \"Does this image have a large and obvious watermark? Please answer with only 'yes' or 'no'\"; respectively, to screen out synthetic images and low-quality images with large watermarks. The third and fourth rounds of dialogue are used to evaluate the content of the pictures. In the third round of dialogue, we use multiple-choice questions to let the model choose the accurate description of the number of snowflakes from five options. We obtain the logits of each option, calculate the option probability and score, and then distinguish the amount of snowflakes and snow accumulation in different pictures. For the specific dialogue process, please refer to Fig. 2. The fourth round of dialogue is the final classification of the topic of the image, and eight categories are set for MLLM to perform classification.", + "bbox": [ + 511, + 537, + 908, + 824 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the specific implementation, to balance evaluation quality and efficiency, we select the MiniCPM-V-2.6 model as the dataset evaluation model. Despite having only 8B parameters, this model demonstrates outstanding performance and provides a well-distributed scoring range. We parti", + "bbox": [ + 511, + 825, + 908, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "4304", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/c15c69f020dcd5ca55b8fa16974709c555b8f0074f1e014951e6f6101462b847.jpg", + "image_caption": [ + "Figure 3. Performance comparison of different MLLMs for different indicators. The horizontal axis in the figure represents the rank-1 proportion of different MLLMs on different snowfall evaluation indicators." + ], + "image_footnote": [], + "bbox": [ + 106, + 89, + 472, + 258 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "tion the RealSnow10K dataset into three subsets: 1,500 images for MMPO fine-tuning, 1,047 images for testing, and 6,406 images for semi-supervised training. Detailed category statistics are provided in the supplementary materials.", + "bbox": [ + 89, + 320, + 483, + 381 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Preference Dataset Construction Pipeline", + "text_level": 1, + "bbox": [ + 89, + 391, + 444, + 407 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To perform multi-model preference optimization (MMPO) on the multimodal large language model, We first use the pre-trained snow removal model to perform preliminary snow removal on 1,500 images and expand the dataset to 3,000 images. We select two indicators with poor performance of Q-instruct from the four indicators for evaluating snowfall: snowflake volume and snowflake occlusion, and design targeted questions for each. Subsequently, we collect the outputs of four MLLMs for these two problems, including Q-instruct-13b, LLava-1.5-13b, LLava-1.6-13b, and MiniCPM-V-2.6. Image evaluation experts rank the outputs of the four models one by one and generate six preference pairs from a ranking. Therefore, for each preference problem, we obtain a preference dataset with a sample size of 18,000.", + "bbox": [ + 89, + 412, + 483, + 640 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Method", + "text_level": 1, + "bbox": [ + 89, + 648, + 181, + 664 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Evaluation of Snowy Images", + "text_level": 1, + "bbox": [ + 89, + 669, + 344, + 686 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Since there is no reference for real snowfall images, the first task for real scene snow removal tasks is to select an accurate and reliable snowfall image evaluation standard. The size, distribution and shape of snowfall particles in the image are often affected by multiple factors including shooting equipment, wind speed, distance, etc., which also brings challenges to the evaluation of snowfall image quality. Traditional non-reference [39, 41] obviously cannot distinguish different forms of snowfall, nor can they accurately describe the interference of snowfall on the image.", + "bbox": [ + 89, + 688, + 482, + 839 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Recently, the development of multimodal large language models has provided a possible solution for snowfall image evaluation. Multimodal large language models are trained on a large number of images and have close to human vi", + "bbox": [ + 89, + 839, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "sual capabilities. Some works [45, 47, 60] introduced CLIP or multimodal large models into image quality assessment tasks (IQA) and obtained good results. Inspired by these works, we try to introduce multimodal large language models into the field of image snow removal. However, there is currently no work exploring the performance of multimodal large language models in evaluating the degradation of snowfall images. Some works [52] use multimodal large language models from a single dimension to evaluate the quality of severe weather images, which is one-sided and unreliable.", + "bbox": [ + 511, + 90, + 903, + 256 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this work, we first establish a detailed framework for evaluating snowing images. By analyzing a large dataset of real snow images, we propose assessing snow image quality across four dimensions: snowflake amount, object obscURITY, texture detail, and background clarity. The evaluation framework appears in Fig.4 (i). We select three widely used multimodal large language models and one model, Q-Instruct [47], fine-tuned for low-level visual features. These models analyze real snow images along the four dimensions, and experts evaluate their outputs. A sample output is presented in Fig.4 (ii).", + "bbox": [ + 511, + 257, + 906, + 422 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Unfortunately, no model can fully dominate in all four dimensions: based on LLava1.5-13b, Q-instruct, which has been fine-tuned with underlying visual features, has shown obvious advantages in texture detail and background clarity, which is consistent with our expectations. However, in terms of the two indicators of snowflake amount and object obscure for snowflake description, the best performing models are LLava1.6-13b and MiniCPM-V-2.6, which firmly occupy the first and second positions, while q-instruct-13b is relatively backward. The specific rank-1 ratios of different models are shown in Fig.3.", + "bbox": [ + 511, + 422, + 905, + 589 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Improving Snowfall Perception via Multi-Model Feedback Optimization", + "text_level": 1, + "bbox": [ + 511, + 595, + 903, + 626 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For complex image desnowing task, evaluating performance from multiple perspectives is a prerequisite for achieving satisfactory results. However, as outlined in the previous section, a single multimodal large language model cannot deliver accurate evaluations across all assessment dimensions. At the same time, employing multiple models concurrently for evaluation is computationally infeasible due to excessive resource demands. This leads us to a key question: is it possible to use a single model to simultaneously address the evaluation of multiple perspectives in desnowing tasks?", + "bbox": [ + 511, + 628, + 905, + 794 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Fortunately, Q-Instruct [47] model has already undergone targeted fine-tuning on a large-scale dataset, equipping it with the ability to assess fundamental image quality metrics, such as clarity and contrast. Building on this foundation, and inspired by recent advancements in reinforcement learning [18, 36-38], we propose enhancing Q-Instruct through human feedback optimization. Our goal is", + "bbox": [ + 511, + 795, + 905, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4305", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/77b5777d4af1c3c3e240197cd66f8f0b765c80b1745e242725bf097372a21a6e.jpg", + "image_caption": [ + "Figure 4. Schematic diagram of the whole process of snowfall image evaluation optimization. The evaluation framework first divides snowfall into four indicators for evaluation, and for each indicator, collects the evaluation answers of multiple multimodal large language models to the image as the data source for preference learning. Finally, direct preference optimization is applied to Q-instruct to improve its output for this indicator. The figure takes the snowflake amount indicator as an example." + ], + "image_footnote": [], + "bbox": [ + 93, + 88, + 906, + 359 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "to enable Q-Instruct to accurately evaluate snowfall, while retaining its capability to assess essential image quality", + "bbox": [ + 88, + 430, + 480, + 460 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Specifically, to address the challenge that different models exhibit distinct strengths and weaknesses across various perspectives, we introduce a novel multi-model preference optimization (MMPO) technique. Unlike conventional approaches, MMPO constructs preference pairs by harnessing the outputs of multiple models, enabling a more comprehensive capture of diverse evaluation criteria. While drawing inspiration from the loss function of Direct Preference Optimization (DPO) [36], MMPO innovatively extends this framework to incorporate the collective strengths of multiple models into a single, cohesive Q-Instruct model. Consequently, the resulting Q-Instruct model excels at thoroughly assessing desnowing performance while preserving its capability to evaluate critical image quality metrics.", + "bbox": [ + 88, + 460, + 482, + 671 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Reward Function Reparameterization. In DPO, The reward associated with a response $y$ given context $x$ is reparameterized in terms of the model's policy $\\pi_{\\theta}(y|x)$ and a reference policy $\\pi_{\\mathrm{ref}}(y|x)$ , formulated as:", + "bbox": [ + 89, + 672, + 482, + 734 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nr (x, y) = \\beta \\log \\frac {\\pi_ {\\theta} (y | x)}{\\pi_ {\\mathrm {r e f}} (y | x)} + \\beta \\log Z (x), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 151, + 739, + 482, + 773 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\beta$ is a scaling constant and $Z(x)$ is a partition function to normalize the policy distribution. Through this reparameterization, DPO completes preference optimization without explicitly defining a reward model, simplifying the complex reinforcement learning problem into a classification-based optimization problem. Our proposed MMPO also adopts this reparameterization technique to streamline the formulation of preference optimization.", + "bbox": [ + 89, + 779, + 482, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Multi-Model Feedback Collection. To construct a robust preference dataset, a multi-model feedback collection strategy is employed. For a given input context $x$ , responses $\\{y_1, y_2, \\ldots, y_N\\}$ are collected from $N$ distinct multimodal large language models. These responses are subsequently evaluated and ranked by human experts, yielding a preference ordering $y_{(1)} \\succ y_{(2)} \\succ \\dots \\succ y_{(N)}$ , where $y_{(1)}$ denotes the most preferred response and $y_{(N)}$ the least preferred. From this ordering, preference pairs $(y_w, y_l)$ are generated such that $y_w \\succ y_l$ , forming the dataset:", + "bbox": [ + 511, + 430, + 906, + 582 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{\\left(x, y _ {w}, y _ {l}\\right) \\mid y _ {w} \\succ y _ {l}, y _ {w}, y _ {l} \\in \\left\\{y _ {1}, y _ {2}, \\dots , y _ {N} \\right\\} \\right\\}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 527, + 594, + 903, + 611 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This methodology harnesses the diverse capabilities of multiple models to produce a comprehensive set of preference pairs, thereby enhancing the differentiation between superior and inferior responses.", + "bbox": [ + 511, + 622, + 903, + 681 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Multi-Model Preference Optimization. Building upon the framework of Direct Preference Optimization (DPO), Multi-Model Preference Optimization (MMPO) is introduced. The loss function is expressed as:", + "bbox": [ + 511, + 683, + 903, + 744 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\mathrm {M M P O}} \\left(\\pi_ {\\theta}; \\pi_ {\\mathrm {r e f}}\\right) = - \\mathbb {E} _ {(x, y _ {w}, y _ {l}) \\sim D} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 593, + 756, + 903, + 773 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left[ \\log \\sigma \\left(\\beta \\log \\frac {\\pi_ {\\theta} \\left(y _ {w} \\mid x\\right)}{\\pi_ {\\mathrm {r e f}} \\left(y _ {w} \\mid x\\right)} - \\beta \\log \\frac {\\pi_ {\\theta} \\left(y _ {l} \\mid x\\right)}{\\pi_ {\\mathrm {r e f}} \\left(y _ {l} \\mid x\\right)}\\right) \\right],\n$$\n", + "text_format": "latex", + "bbox": [ + 547, + 784, + 870, + 819 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\sigma (\\cdot)$ represents the sigmoid function, and $y_{w}$ and $y_{l}$ denote the preferred and unpreferred responses, respectively. A distinguishing feature of MMPO is its reliance on a preference dataset $D$ derived from the outputs of multiple models, as outlined in the feedback collection process. This", + "bbox": [ + 511, + 824, + 906, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "4306", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "approach integrates the strengths of various MLLMs, allowing the model to leverage a richer collection of preference pairs. Through this optimization, the model's outputs are aligned with human expert preferences, thereby improving its capacity to generate high-quality responses.", + "bbox": [ + 89, + 90, + 483, + 167 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. From Perception to Restoration", + "text_level": 1, + "bbox": [ + 89, + 175, + 372, + 191 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Using MMPO technology, we use Q-instruct as the base model and fine-tune its preference learning in two aspects where its performance was relatively poor: snowflake amount and object obscure, improving its ability to perceive and understand snowfall. However, how can the improvement in perception be used to optimize the image restoration model?", + "bbox": [ + 89, + 196, + 483, + 301 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Inspired by Q-Align [47], we leverage the enhanced perception capabilities of the large model to generate an \"image score\" that could quantify this improvement. We then use this image score as an evaluation metric within a mean teacher semi-supervised learning framework, combined with a pseudo-label database, to train the image restoration model.", + "bbox": [ + 89, + 303, + 483, + 407 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Score calculation. Using one adjective like Q-Align (e.g. \"excellent\", \"poor\") to describe the snowfall scene directly does not provide an accurate evaluation. We analyze that this is because a single adjective lacks a specific definition. Therefore, we further decompose the prompt words, describe the four indicators such as the number of snowflakes from different perspectives, and design it into a multiple choice question with five options. Based on the calculation method of human subjective opinion score (MOS):", + "bbox": [ + 89, + 409, + 483, + 545 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {M O S} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} S _ {i}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 220, + 556, + 482, + 597 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "in which $S_{i}$ represents the score of the i-th tester. We collect the model's output logit for the five option tokens, use softmax to calculate the probability of different tokens, and finally get the score through weighted average:", + "bbox": [ + 89, + 602, + 483, + 662 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\text {S c o r e} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} p _ {i} \\cdot v _ {i}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 207, + 674, + 482, + 715 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\np _ {i} = \\frac {e ^ {l _ {i}}}{\\sum_ {j = 1} ^ {N} e ^ {l _ {j}}}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 230, + 720, + 482, + 760 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $l_{i}$ represents the logit value of the $i$ -th option token. $N$ is 5 to correspond to 5 options. $v_{i}$ is converted into weights from 1 to 5 according to the description of snowfall intensity.", + "bbox": [ + 89, + 763, + 482, + 824 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Semi-supervised Training with Pseudo-labeling. Referring to the mean teacher framework and the database mechanism in [36], we leverage scores from four evaluation metrics to enhance semi-supervised training. To ensure the pseudo-label database is updated accurately, we adopt a", + "bbox": [ + 89, + 825, + 483, + 901 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ee17afd6786d89eb557b937995f149fed4ed54accad3dfc06fd0a93f2414eafc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 517, + 89, + 686, + 172 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/3210ac6913cd883be5f58b4c6a536b898129f8a549a567a5eb59537ba0dafebd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 686, + 89, + 712, + 109 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "How would you evaluate the amount of snowflakes in the air shown in the picture? Please provide your judgment and basis. Summarize in about 50 words.", + "bbox": [ + 714, + 90, + 905, + 132 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/89b66cd64fcc15cbb955c9750b2e9a1d737825a1114060a9b2b46d5e586b6ea5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 687, + 135, + 710, + 154 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Keywords that do not match the real picture Keywords that lack logic or basis", + "bbox": [ + 714, + 136, + 890, + 155 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/abb41cec1c16737bc1aa024a727e864851e557548fa680df04ab50597995486a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 687, + 155, + 710, + 172 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Keywords that match the picture and perception of human experts", + "bbox": [ + 714, + 159, + 903, + 176 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/375ddf0bac9b4e2b9e9d417421990387a33418d63690aca5f8f6d053e0cc47e4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 517, + 181, + 540, + 202 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Q-instruct", + "text_level": 1, + "bbox": [ + 547, + 191, + 601, + 200 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The amount of snowflakes in the air shown in the picture is quite a lot. This is evident from the fact that the people in the image are wearing winter coats and hats, and their clothing is covered in snow. The snowflakes are also clearly visible in the air, indicating that the snowfall is quite heavy. The people are walking in the snow, which further emphasizes the significant amount of snowfall. Overall, the image portrays a snowy day with a considerable amount of snowflakes in the air.", + "bbox": [ + 516, + 202, + 704, + 335 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/7badbd3e3b965cc93128f598bc481dd9a0c321059cbf0fb0edeabd768e2545e2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 730, + 180, + 761, + 204 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "O-instruct + MMPO", + "text_level": 1, + "bbox": [ + 764, + 190, + 859, + 199 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The image shows a group of people walking in the snow, with several snowflakes visible in the air. While it is not possible to quantify the exact number of snowflakes, it is evident that there is a noticeable amount of snowfall. The snowflakes are dispersed throughout the scene, indicating that the snowfall is not too heavy or too light, but rather moderate. This suggests that the weather conditions are likely to be cold and snowy, but not extreme or overwhelming.", + "bbox": [ + 723, + 203, + 903, + 335 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Figure 5. Comparison of perception ability before and after MMPO. The fine-tuned model output is more in line with human visual perception, more logical, and less hallucinatory.", + "bbox": [ + 511, + 339, + 905, + 382 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "stringent rule: an update occurs only if $(i)$ at least three teacher scores exceed the student score, and $(ii)$ the teacher score surpasses the current score stored in the pseudo-label database. This dual-condition approach effectively minimizes erroneous updates to the pseudo-labels, improving the reliability of the training process.", + "bbox": [ + 511, + 392, + 905, + 482 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Loss function and training strategy. We use both supervised loss and pseudo-label loss to train the model. The supervised loss is based on the PSNR between the model's output and the reference images from the Snow100K [30] dataset. And the pseudo-label loss relies on the L1 norm to measure the difference between the output and pseudolabels. To ensure fast and stable training, we only use supervised loss in the initial 20k steps, and add pseudo-label loss after 20k steps for semi-supervised training. The total loss function is shown as follows:", + "bbox": [ + 511, + 483, + 906, + 633 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {\\text {s u p e r v i s e d}} + \\lambda \\mathcal {L} _ {\\text {p s e u d o}} (\\lambda = 0 \\text {i f s t e p < 2 0 k}). \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 529, + 659, + 903, + 676 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 513, + 694, + 645, + 710 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Experiment Setup", + "text_level": 1, + "bbox": [ + 513, + 715, + 689, + 732 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Training Settings. Our method consists of two phases: the first phase is to fine-tune the scoring model (i.e. Q-Instruct [50]) via proposed MMPO, and the second phase is to semi-supervised train the snow removal model. In the first phase, we set epoch = 1, lr = 1e-6, batchsize = 64; We use Vision-LLM Alignment framework [44] to perform MMPO fine-tuning and enable deepspeed for acceleration, set zero-stage to 3 and enable offload for full fine-tuning. In the second phase of snow removal model training, we use NAFNet [5] as the base model, set lr = 2e-4, batchsize = 32. All experiments are performed on two Nvidia A800 GPUs.", + "bbox": [ + 511, + 734, + 906, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "4307", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/64e5340073e481c4490fd8bcf2efbe32b44ed03300b555a63c0bf19b5425abb6.jpg", + "image_caption": [ + "Figure 6. Visual comparison of different models for snow removal on real images. Please zoom in for a better view. SnowMaster achieves satisfactory snow removal performance in complicated conditions compared to previous methods." + ], + "image_footnote": [], + "bbox": [ + 94, + 90, + 903, + 518 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/ca2747378d70aa7be744029e4980c7548d597f8f50dc6979b838aee10f920068.jpg", + "table_caption": [ + "Table 1. The accuracy of MLLMs on different perspectives." + ], + "table_footnote": [], + "table_body": "
MLLMs AccuracySnowflakes accQuality acc
amount ↑obscure ↑texture ↑background ↑
LLava-v1.5-13b[29]80.0477.1755.7852.15
LLava-v1.6-13b[29]86.3479.5652.0554.06
MiniCPM-V[53]75.4570.0442.0249.76
Q-intract-13b[50]76.3173.2658.3663.42
Q-instruct with MMPO (ours)82.1480.9064.9568.58
MMPO improvement+5.83+7.64+6.59+5.16
", + "bbox": [ + 101, + 574, + 480, + 680 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Experimental Details. In the semi-supervised training process, considering the need to make the multimodal large language model output the correct score, we adopt a pseudo-label database initialization strategy using the original image instead of all-0 initialization in [21]. In the test phase, we test 457 test images selected from the real snow scene dataset (RealSnow10K) and 1329 real snow scene images from Snow100K [30].", + "bbox": [ + 89, + 700, + 482, + 820 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Evaluation Metrics and Baselines", + "text_level": 1, + "bbox": [ + 89, + 830, + 382, + 845 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Following [19, 20], we utilize four non-reference IQA metrics: BRISQUE [31], NIQE [32], PIQE [43], and PAQ2-PIQ [57]. BRISQUE measures quality via spa", + "bbox": [ + 89, + 854, + 483, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "tial natural scene statistics, NIQE assesses deviations from these statistics, PIQE detects block-level distortions, and PAQ2-PIQ, a deep learning approach, predicts global and local quality. These metrics ensure a comprehensive, unsupervised evaluation of diverse distortions. For the baselines, we choose CNN-based models (SMGARN[17], NAFNet[5]), Transformer-based models (Transweather[42], Snowformer[6]) and a diffusion based model $(\\mathrm{T}^3$ -DiffWeather[11]), offering varied references for snowfall image restoration performance.", + "bbox": [ + 511, + 560, + 906, + 712 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Results and Analysis", + "text_level": 1, + "bbox": [ + 511, + 723, + 707, + 739 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3.1. Quantitative Results", + "text_level": 1, + "bbox": [ + 511, + 743, + 702, + 758 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effectiveness of MMPO. Table 1 illustrates the performance enhancements of the Q-Instruct [47] model following fine-tuning with MMPO. We utilize NAFNet [5], pretrained for 100,000 steps, as the snow removal model to process the images in the test set, generating a dataset that includes both original and snow-removed images. By inputting these images individually into Q-Instruct, we obtain scores for each image corresponding to specific indicators. Consequently, for each indicator, accuracy is defined", + "bbox": [ + 511, + 763, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "4308", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/522db6fecd867b30b580ea2f6a397af7d4a153628e3e2cf5bcd92203778f8874.jpg", + "table_caption": [ + "Table 2. Comparison of non-reference metrics after snow removal between different models." + ], + "table_footnote": [], + "table_body": "
MethodSnow100K-realistic [30]RealSnow10K
BRISQUE ↓NIQE ↓PIQE ↓PAQ2-PIQ ↑BRISQUE ↓NIQE ↓PIQE ↓PAQ2-PIQ ↑
SMGARN[TCSVT'22][17]19.2893.13231.92769.76723.0053.58031.21068.919
TransWeather[CVPR'22] [42]16.5443.02530.30269.82118.5643.40429.77269.671
NAFNet[ECCV'22] [5]17.5662.90330.84369.95120.4253.29129.87769.388
SnowFormer[Arxiv'23][6]19.3892.95431.90270.01523.7023.42931.9969.929
T3-DiffWeather[ECCV'24] [11]17.7442.86831.68970.04220.2113.13732.99170.211
★ SnowMaster (Ours)16.3052.79730.16270.04217.5463.05829.74270.304
", + "bbox": [ + 130, + 102, + 864, + 224 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "as the proportion of snow-removed images whose score is not lower than that of their corresponding original images. The results show that the performance after fine-tuning has been significantly improved which is consistent with our expectations. It is worth noting that in terms of the number of snowflakes, the accuracy after fine-tuning does not exceed that of Llava-1.6. We think this is normal, because the preference data in fine-tuning are all output by these models, and MMPO guides the model to move closer to human preference output, but it may not necessarily surpass the output of the optimal model.", + "bbox": [ + 88, + 243, + 480, + 407 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Furthermore, a noteworthy observation emerges: the indicators 'texture' and 'background', which are not explicitly targeted for optimization, also exhibit moderate improvements. We hypothesize that this phenomenon arises from a comprehensive augmentation of the model's perceptual capacity regarding snowfall, facilitated by the MMPO fine-tuning process.", + "bbox": [ + 89, + 409, + 480, + 515 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Performance analysis of the restoration model. Table 2 presents the performance comparison between our method and other advanced snow removal models. Our proposed method achieves the best results in most indicators.", + "bbox": [ + 89, + 515, + 480, + 575 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.3.2. Qualitative Assessment", + "text_level": 1, + "bbox": [ + 89, + 582, + 299, + 597 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "From the restoration results of the real snowfall image in Fig.6, we can see that the SnowMaster model we proposed has the best performance in multiple scenarios. When facing snowflakes illuminated by lights at night, other models cannot effectively complete the restoration of high-brightness snowflakes, while SnowMaster significantly eliminates the high-brightness snowflakes in the image. When restoring dense snowflake scenes during the day, SnowMaster can effectively remove snowflakes of different scales and retain the original color and texture of the object, while other models have poor restoration effects on small-sized snowflake particles.", + "bbox": [ + 89, + 601, + 482, + 784 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.4. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 790, + 243, + 806 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Necessity of MMPO. The Q-Instruct model shows significant improvement after multi-modal preference optimization, particularly in evaluating snowfall images. We perform an ablation test to compare the Q-Instruct model before and after alignment when used as a scoring model in semi-supervised training. The results indicate that the", + "bbox": [ + 89, + 809, + 480, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/721e09c2405770ada25693e7dacac1be14df39f8492de4842cb79e781e8701cd.jpg", + "table_caption": [ + "Table 3. Ablation on different settings of the proposed framework." + ], + "table_footnote": [], + "table_body": "
MethodBRISQUE ↓NIQE ↓
w/o. Semi-supervised learning22.523.33
w/o. MLLM20.363.25
w/o. MMPO18.983.19
w. Single-angle evaluation18.213.12
w. Multi-angle evalutaion (Ours)17.553.06
", + "bbox": [ + 531, + 257, + 890, + 349 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "aligned model achieves higher accuracy and ensures more stable training to improve the performance.", + "bbox": [ + 511, + 363, + 903, + 393 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effectiveness of multi-perspectives snowfall evaluation. Our method enhances image screening reliability by evaluating snowfall images from four perspectives, ensuring steady updates to the pseudo-label database. Ablation experiments, conducted by reducing evaluation angles, demonstrate that multi-angle evaluation improves training effectiveness", + "bbox": [ + 511, + 395, + 903, + 500 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The role of the semi-supervised learning framework. In order to transfer the improvement of the evaluation performance of the large model on snow scene images to the improvement of the restoration model, we use the mean teacher semi-supervised learning framework. By combining the supervision loss from snow100k and the pseudolabel loss from the real training data, this framework can effectively optimize the restoration effect of real snowfall images. Through ablation experiments, we prove that the semi-supervised framework is the basis of our method.", + "bbox": [ + 511, + 502, + 905, + 652 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The specific metrics of the ablation experiments are shown in Table 3.", + "bbox": [ + 511, + 654, + 903, + 683 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 696, + 633, + 710 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "SnowMaster integrates a multimodal large language model with a multi-perspective evaluation framework for image desnowing. Leveraging the RealSnow10k dataset, it utilizes a multi-angle evaluation system assessing four key metrics to enhance real-world reliability. Through multimodal preference optimization and semi-supervised training with pseudo-label updates from multi-angle multiple-choice questions, it ensures effective refinement. Experiments confirm its robustness in complex snowfall scenarios, though its computational intensity from multiple evaluations signals a need for future enhancements in training efficiency and resource use.", + "bbox": [ + 511, + 719, + 905, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "4309", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 91, + 90, + 250, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work is supported by the Guangzhou-HKUST(GZ) Joint Funding Program (No. 2023A03J0671), the Guangzhou Municipal Science and Technology Project (Grant No. 2024A04J4230), and Guangzhou-HKUST(GZ) Joint Funding Program (No. 2024A03J0618).", + "bbox": [ + 89, + 114, + 483, + 191 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 207, + 187, + 222 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Vijay Badrinarayanan, Alex Kendall, and Roberto Cipolla. Segnet: A deep convolutional encoder-decoder architecture for image segmentation. IEEE transactions on pattern analysis and machine intelligence, 39(12):2481-2495, 2017. 1", + "[2] Chun-Fu Richard Chen, Quanfu Fan, and Rameswar Panda. Crossvit: Cross-attention multi-scale vision transformer for image classification. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 357-366, 2021. 1", + "[3] Haoyu Chen, Jinjin Gu, Yihao Liu, Salma Abdel Magid, Chao Dong, Qiong Wang, Hanspeter Pfister, and Lei Zhu. Masked image training for generalizable deep image denoising. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1692-1703, 2023. 2", + "[4] Haoyu Chen, Wenbo Li, Jinjin Gu, Jingjing Ren, Haoze Sun, Xueyi Zou, Zhensong Zhang, Youliang Yan, and Lei Zhu. Low-res leads the way: Improving generalization for superresolution by self-supervised learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25857-25867, 2024. 1", + "[5] Liangyu Chen, Xiaojie Chu, Xiangyu Zhang, and Jian Sun. Simple baselines for image restoration. In European conference on computer vision, pages 17-33. Springer, 2022. 6, 7, 8", + "[6] Sixiang Chen, Tian Ye, Yun Liu, and Erkang Chen. Snowformer: Context interaction transformer with scale-awareness for single image desnowing. arXiv preprint arXiv:2208.09703, 2022. 2, 7, 8", + "[7] Sixiang Chen, Tian Ye, Yun Liu, Taodong Liao, Yi Ye, and Erkang Chen. Msp-former: Multi-scale projection transformer for single image desnowing. arXiv preprint arXiv:2207.05621, 2022. 1, 2", + "[8] Sixiang Chen, Tian Ye, Jinbin Bai, Erkang Chen, Jun Shi, and Lei Zhu. Sparse sampling transformer with uncertainty-driven ranking for unified removal of raindrops and rain streaks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13106-13117, 2023. 2", + "[9] Sixiang Chen, Tian Ye, Yun Liu, Jinbin Bai, Haoyu Chen, Yunlong Lin, Jun Shi, and Erkang Chen. CPLformer: Cross-scale prototype learning transformer for image snow removal. In Proceedings of the 31st ACM international conference on multimedia, pages 4228-4239, 2023. 1", + "[10] Sixiang Chen, Tian Ye, Chenghao Xue, Haoyu Chen, Yun Liu, Erkang Chen, and Lei Zhu. Uncertainty-driven dynamic" + ], + "bbox": [ + 93, + 232, + 483, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "degradation perceiving and background modeling for efficient single image desnowing. In Proceedings of the 31st ACM International Conference on Multimedia, pages 4269-4280, 2023. 2", + "[11] Sixiang Chen, Tian Ye, Kai Zhang, Zhaohu Xing, Yunlong Lin, and Lei Zhu. Teaching tailored to talent: Adverse weather restoration via prompt pool and depth-anything constraint. In European Conference on Computer Vision, pages 95–115. Springer, 2025. 2, 7, 8", + "[12] Wei-Ting Chen, Hao-Yu Fang, Jian-Jiun Ding, Cheng-Che Tsai, and Sy-Yen Kuo. Jstasr: Joint size and transparency-aware snow removal algorithm based on modified partial convolution and veiling effect removal. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXI 16, pages 754-770. Springer, 2020. 1, 2", + "[13] Wei-Ting Chen, Hao-Yu Fang, Cheng-Lin Hsieh, Cheng-Che Tsai, I Chen, Jian-Jiun Ding, Sy-Yen Kuo, et al. All snow removed: Single image desnowing algorithm using hierarchical dual-tree complex wavelet representation and contradict channel loss. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4196-4205, 2021. 2", + "[14] Wei-Ting Chen, Hao-Yu Fang, Cheng-Lin Hsieh, Cheng-Che Tsai, I Chen, Jian-Jiun Ding, Sy-Yen Kuo, et al. All snow removed: Single image desnowing algorithm using hierarchical dual-tree complex wavelet representation and contradict channel loss. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4196-4205, 2021. 1, 2", + "[15] Wei-Ting Chen, Zhi-Kai Huang, Cheng-Che Tsai, Hao-Hsiang Yang, Jian-Jiun Ding, and Sy-Yen Kuo. Learning multiple adverse weather removal via two-stage knowledge learning and multi-contrastive regularization: Toward a unified model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17653–17662, 2022. 2", + "[16] Bodong Cheng, Juncheng Li, Ying Chen, Shuyi Zhang, and Tieyong Zeng. Snow mask guided adaptive residual network for image snow removal. arXiv preprint arXiv:2207.04754, 2022. 1, 2", + "[17] Bodong Cheng, Juncheng Li, Ying Chen, and Tieyong Zeng. Snow mask guided adaptive residual network for image snow removal. Computer Vision and Image Understanding, 236: 103819, 2023. 2, 7, 8", + "[18] Paul F Christiano, Jan Leike, Tom Brown, Miljan Martic, Shane Legg, and Dario Amodei. Deep reinforcement learning from human preferences. Advances in neural information processing systems, 30, 2017. 4", + "[19] Jinjin Gu, Haoming Cai, Haoyu Chen, Xiaoxing Ye, Jimmy Ren, and Chao Dong. Image quality assessment for perceptual image restoration: A new dataset, benchmark and metric. arXiv preprint arXiv:2011.15002, 2020. 7", + "[20] Jinjin Gu, Haoming Cai, Haoyu Chen, Xiaoxing Ye, Jimmy Ren, and Chao Dong. Pipal: a large-scale image quality assessment dataset for perceptual image restoration. In European Conference on Computer Vision, pages 633-651. Springer, 2020. 7" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "4310", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[21] Shirui Huang, Keyan Wang, Huan Liu, Jun Chen, and Yun-song Li. Contrastive semi-supervised learning for underwater image restoration via reliable bank. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18145-18155, 2023. 7", + "[22] Yeying Jin, Wending Yan, Wenhan Yang, and Robby T Tan. Structure representation network and uncertainty feedback learning for dense non-uniform fog removal. In Asian Conference on Computer Vision, pages 155-172. Springer, 2022. 2", + "[23] Yeying Jin, Wenhan Yang, Wei Ye, Yuan Yuan, and Robby T Tan. Shadowdiffusion: Diffusion-based shadow removal using classifier-driven attention and structure preservation. arXiv preprint arXiv:2211.08089, 2022.", + "[24] Yeying Jin, Beibei Lin, Wending Yan, Wei Ye, Yuan Yuan, and Robby T. Tan. Enhancing visibility in nighttime haze images using guided apsf and gradient adaptive convolution, 2023. 2", + "[25] Ruoteng Li, Robby T Tan, and Loong-Fah Cheong. All in one bad weather removal using architectural search. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3175-3185, 2020. 2", + "[26] Yunlong Lin, Zhenqi Fu, Ge Meng, Yingying Wang, Yuhang Dong, Linyu Fan, Hedeng Yu, and Xinghao Ding. Domain-irrelevant feature learning for generalizable pan-sharpening. In Proceedings of the 31st ACM International Conference on Multimedia, pages 3287–3296, 2023. 2", + "[27] Yunlong Lin, Zhenqi Fu, Kairun Wen, Tian Ye, Sixiang Chen, Ge Meng, Yingying Wang, Yue Huang, Xiaotong Tu, and Xinghao Ding. Unsupervised low-light image enhancement with lookup tables and diffusion priors. arXiv preprint arXiv:2409.18899, 2024.", + "[28] Yunlong Lin, Tian Ye, Sixiang Chen, Zhenqi Fu, Yingying Wang, Wenhao Chai, Zhaohu Xing, Lei Zhu, and Xinghao Ding. Aglldiff: Guiding diffusion models towards unsupervised training-free real-world low-light image enhancement. arXiv preprint arXiv:2407.14900, 2024. 2", + "[29] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26296-26306, 2024. 2, 7", + "[30] Yun-Fu Liu, Da-Wei Jaw, Shih-Chia Huang, and Jenq-Neng Hwang. Desnownet: Context-aware deep network for snow removal. IEEE Transactions on Image Processing, 27(6): 3064–3073, 2018. 2, 6, 7, 8", + "[31] Anish Mittal, Anush Krishna Moorthy, and Alan Conrad Bovik. No-reference image quality assessment in the spatial domain. IEEE Transactions on image processing, 21(12): 4695-4708, 2012. 1, 7", + "[32] Anish Mittal, Rajiv Soundararajan, and Alan C Bovik. Making a “completely blind” image quality analyzer. IEEE Signal processing letters, 20(3):209-212, 2012. 1, 7", + "[33] Ozan Özdenizci and Robert Legenstein. Restoring vision in adverse weather conditions with patch-based denoising diffusion models. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 2" + ], + "bbox": [ + 91, + 90, + 485, + 901 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[34] Soo-Chang Pei, Yu-Tai Tsai, and Chen-Yu Lee. Removing rain and snow in a single image using saturation and visibility features. In 2014 IEEE International Conference on Multimedia and Expo Workshops (ICMEW), pages 1-6. IEEE, 2014. 2", + "[35] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 3", + "[36] Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. Direct preference optimization: Your language model is secretly a reward model. Advances in Neural Information Processing Systems, 36, 2024. 2, 4, 5, 6", + "[37] John Schulman, Sergey Levine, Pieter Abbeel, Michael Jordan, and Philipp Moritz. Trust region policy optimization. In International conference on machine learning, pages 1889-1897. PMLR, 2015.", + "[38] John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347, 2017. 4", + "[39] Shaolin Su, Qingsen Yan, Yu Zhu, Cheng Zhang, Xin Ge, Jinqiu Sun, and Yanning Zhang. Blindly assess image quality in the wild guided by a self-adaptive hyper network. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3667-3676, 2020. 3, 4", + "[40] Haoze Sun, Wenbo Li, Jianzhuang Liu, Haoyu Chen, Renjing Pei, Xueyi Zou, Youliang Yan, and Yujiu Yang. Coser: Bridging image and language for cognitive super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25868-25878, 2024. 2", + "[41] Hossein Talebi and Peyman Milanfar. Nima: Neural image assessment. IEEE transactions on image processing, 27(8): 3998-4011, 2018. 3, 4", + "[42] Jeya Maria Jose Valanarasu, Rajeev Yasarla, and Vishal M Patel. Transweather: Transformer-based restoration of images degraded by adverse weather conditions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2353-2363, 2022. 2, 7, 8", + "[43] Narasimhan Venkatanath, D Praneeth, Maruthi Chandrasekhar Bh, Sumohana S Channappayya, and Swarup S Medasani. Blind image quality evaluation using perception based features. In 2015 twenty first national conference on communications (NCC), pages 1-6. IEEE, 2015. 1, 7", + "[44] Chenglong Wang, Yang Gan, Yifu Huo, Yongyu Mu, Murun Yang, Qiaozhi He, Tong Xiao, Chunliang Zhang, Tongran Liu, Quan Du, Di Yang, and Jingbo Zhu. Rovrm: A robust visual reward model optimized via auxiliary textual preference data, 2024. 6", + "[45] Jianyi Wang, Kelvin CK Chan, and Chen Change Loy. Exploring clip for assessing the look and feel of images. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2555-2563, 2023. 3, 4" + ], + "bbox": [ + 516, + 90, + 905, + 901 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "4311", + "bbox": [ + 482, + 945, + 513, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[46] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 3", + "[47] Haoning Wu, Zicheng Zhang, Weixia Zhang, Chaofeng Chen, Liang Liao, Chunyi Li, Yixuan Gao, Annan Wang, Erli Zhang, Wenxiu Sun, et al. Q-align: Teaching Imms for visual scoring via discrete text-defined levels. arXiv preprint arXiv:2312.17090, 2023. 1, 3, 4, 6, 7", + "[48] Hongtao Wu, Yijun Yang, Angelica I Aviles-Rivero, Jingjing Ren, Sixiang Chen, Haoyu Chen, and Lei Zhu. Semi-supervised video desnowing network via temporal decoupling experts and distribution-driven contrastive regularization. In European Conference on Computer Vision, pages 70-89. Springer, 2024. 1", + "[49] Hongtao Wu, Yijun Yang, Huihui Xu, Weiming Wang, Jinni Zhou, and Lei Zhu. Rainmamba: Enhanced locality learning with state space models for video deraining. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 7881-7890, 2024. 2", + "[50] Haoning Wu, Zicheng Zhang, Erli Zhang, Chaofeng Chen, Liang Liao, Annan Wang, Kaixin Xu, Chunyi Li, Jingwen Hou, Guangtao Zhai, et al. Q-instruct: Improving low-level visual abilities for multi-modality foundation models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25490-25500, 2024. 1, 2, 6, 7", + "[51] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. Advances in Neural Information Processing Systems, 34, 2021. 1", + "[52] Jiaqi Xu, Mengyang Wu, Xiaowei Hu, Chi-Wing Fu, Qi Dou, and Pheng-Ann Heng. Towards real-world adverse weather image restoration: Enhancing clearness and semantics with vision-language models. In European Conference on Computer Vision, pages 147–164. Springer, 2025. 4", + "[53] Yuan Yao, Tianyu Yu, Ao Zhang, Chongyi Wang, Junbo Cui, Hongji Zhu, Tianchi Cai, Haoyu Li, Weilin Zhao, Zhihui He, et al. Minicpm-v: A gpt-4v level mllm on your phone. arXiv preprint arXiv:2408.01800, 2024. 2, 7", + "[54] Tian Ye, Yunchen Zhang, Mingchao Jiang, Liang Chen, Yun Liu, Sixiang Chen, and Erkang Chen. Perceiving and modeling density for image dehazing. In European Conference on Computer Vision, pages 130-145. Springer, 2022. 2", + "[55] Tian Ye, Sixiang Chen, Jinbin Bai, Jun Shi, Chenghao Xue, Jingxia Jiang, Junjie Yin, Erkang Chen, and Yun Liu. Adverse weather removal with codebook priors. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12653-12664, 2023. 2", + "[56] Tian Ye, Sixiang Chen, Wenhao Chai, Zhaohu Xing, Jing Qin, Ge Lin, and Lei Zhu. Learning diffusion texture priors for image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2524-2534, 2024. 2", + "[57] Zhenqiang Ying, Haoran Niu, Praful Gupta, Dhruv Mahajan, Deepti Ghadiyaram, and Alan Bovik. From patches to" + ], + "bbox": [ + 91, + 90, + 480, + 901 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "pictures (paq-2-piq): Mapping the perceptual space of picture quality. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3575-3585, 2020. 1, 7", + "[58] Zhiyuan You, Zheyuan Li, Jinjin Gu, Zhenfei Yin, Tianfan Xue, and Chao Dong. Depicting beyond scores: Advancing image quality assessment through multi-modal language models. arXiv preprint arXiv:2312.08962, 2023. 2", + "[59] Kaihao Zhang, Rongqing Li, Yanjiang Yu, Wenhan Luo, and Changsheng Li. Deep dense multi-scale network for snow removal using semantic and depth priors. IEEE Transactions on Image Processing, 30:7419-7431, 2021. 2", + "[60] Weixia Zhang, Guangtao Zhai, Ying Wei, Xiaokang Yang, and Kede Ma. Blind image quality assessment via vision-language correspondence: A multitask learning perspective. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14071-14081, 2023. 3, 4", + "[61] Haipeng Zhou, Hongqiu Wang, Tian Ye, Zhaohu Xing, Jun Ma, Ping Li, Qiong Wang, and Lei Zhu. Timeline and boundary guided diffusion network for video shadow detection. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 166-175, 2024. 1" + ], + "bbox": [ + 516, + 92, + 903, + 415 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "4312", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/SnowMaster_ Comprehensive Real-world Image Desnowing via MLLM with Multi-Model Feedback Optimization/01611a60-cb24-4ca8-93fd-c47c38025543_model.json b/2025/SnowMaster_ Comprehensive Real-world Image Desnowing via MLLM with Multi-Model Feedback Optimization/01611a60-cb24-4ca8-93fd-c47c38025543_model.json new file mode 100644 index 0000000000000000000000000000000000000000..44e380b53ea15a2929f86c08cde2ef0e62ccd6c0 --- /dev/null +++ b/2025/SnowMaster_ Comprehensive Real-world Image Desnowing via MLLM with Multi-Model Feedback Optimization/01611a60-cb24-4ca8-93fd-c47c38025543_model.json @@ -0,0 +1,2301 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.18, + 0.131, + 0.82, + 0.177 + ], + "angle": 0, + "content": "SnowMaster: Comprehensive Real-world Image Desnowing via MLLM with Multi-Model Feedback Optimization" + }, + { + "type": "text", + "bbox": [ + 0.144, + 0.203, + 0.859, + 0.314 + ], + "angle": 0, + "content": "Jianyu Lai\\(^{1,*}\\) Sixiang Chen\\(^{1,*}\\) Yunlong Lin\\(^{2}\\) Tian Ye\\(^{1}\\) \nYun Liu\\(^{3}\\) Song Fei\\(^{1}\\) Zhaohu Xing\\(^{1}\\) Hongtao Wu\\(^{1}\\) Weiming Wang\\(^{4}\\) Lei Zhu\\(^{1,5,\\dagger}\\) \n\\(^{1}\\) The Hong Kong University of Science and Technology (GZ) \\(^{2}\\) Xiamen University \n\\(^{3}\\) Southwestern University \\(^{4}\\) Hong Kong Metropolitan University \n\\(^{5}\\) The Hong Kong University of Science and Technology \nProject page: https://alexlai2860.github.io/SnowMaster" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.348, + 0.327, + 0.365 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.381, + 0.485, + 0.699 + ], + "angle": 0, + "content": "Snowfall presents significant challenges for visual data processing, necessitating specialized desnowing algorithms. However, existing models often fail to generalize effectively due to their heavy reliance on synthetic datasets. Furthermore, current real-world snowfall datasets are limited in scale and lack dedicated evaluation metrics designed specifically for snowfall degradation, thus hindering the effective integration of real snowy images into model training to reduce domain gaps. To address these challenges, we first introduce RealSnow10K, a large-scale, high-quality dataset consisting of over 10,000 annotated real-world snowy images. In addition, we curate a preference dataset comprising 36,000 expert-ranked image pairs, enabling the adaptation of multimodal large language models (MLLMs) to better perceive snowy image quality through our innovative Multi-Model Preference Optimization (MMPO). Finally, we propose the SnowMaster, which employs MMPO-enhanced MLLM to perform accurate snowy image evaluation and pseudo-label filtering for semi-supervised training. Experiments demonstrate that SnowMaster delivers superior desnowing performance under real-world conditions." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.714, + 0.222, + 0.729 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.74, + 0.483, + 0.846 + ], + "angle": 0, + "content": "Snowfall, a common form of adverse weather degradation, reduces image visibility and adversely impacts downstream tasks such as image recognition [2, 61] and segmentation [1, 51]. The inherent complexity of snow poses a challenge to conventional image restoration techniques, necessitating the development of specialized snow removal (desnowing) algorithms to enhance image clarity." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.846, + 0.483, + 0.877 + ], + "angle": 0, + "content": "Although current snow removal models achieve impressive performance on benchmark datasets [7, 9, 12, 14, 16]," + }, + { + "type": "image", + "bbox": [ + 0.513, + 0.346, + 0.911, + 0.497 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.498, + 0.908, + 0.568 + ], + "angle": 0, + "content": "Figure 1. Comparison of various image desnowing evaluation metrics shows that Q-Instruct [47], enhanced by our MMPO, most accurately perceives differences in snow density between images. Leveraging accurate evaluation model, we optimize our semi-supervised network, achieving superior desnowing performance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.58, + 0.907, + 0.747 + ], + "angle": 0, + "content": "they often struggle to generalize to real-world scenarios. This limitation stems primarily from their only reliance on synthetic snowfall datasets for training. These datasets, generated using simplified physical models, struggle to capture the intricate details of real-world snow degradations, resulting in a domain gap between synthetic and real data. This gap hinders model optimization and impairs generalization to real-world scenarios. Consequently, we pose the question: How can real-world data be effectively utilized to enhance the model's generalization performance on real-world samples?" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Incorporating real-world snowy images into training is an direct and effective approach to enhance model generalization. However, accurate evaluation of these images is a prerequisite for their successful integration into the training process [4, 48, 50]. Yet, evaluating realistic snowfall images remains a significant challenge due to the absence of ground truth for reference evaluation, previous methods often rely on non-reference metrics [31, 32, 43, 57] to assess image restoration quality. These metrics, influenced by numerous factors, fail to accurately isolate and evalu-" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.887, + 0.356, + 0.901 + ], + "angle": 0, + "content": "*Equal Contribution; † Corresponding Author." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4302" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.318 + ], + "angle": 0, + "content": "ate weather-specific degradation in snowfall images. In this paper, we claim that accurate assessment of snowy images requires a joint effort from both low-level visual and semantic perspectives. Recent advancements in multimodal large language models (MLLMs), such as Q-Instruct [50] and Depict-QA [58], offer promising avenues for developing evaluation frameworks due to their ability to assess low-level visual features from a semantic perspective. However, these models currently lack the specialized perception capabilities required to reliably evaluate snowfall-specific characteristics. Conversely, while some more advanced multimodal large language models exhibit better understanding of snowy scenes, they are not fine-tuned for detailed low-level visual evaluation, limiting the fundamental premise as powerful evaluators on the snow removal task." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.321, + 0.485, + 0.683 + ], + "angle": 0, + "content": "To tackle these challenges, we first establish a comprehensive evaluation framework designed to capture the distinct characteristics of snow in natural images through multi-perspective feedback optimization tailored to snow scenes. To support this framework, we curate and annotate a dataset comprising over 10,000 real-world snow images sourced from the Internet. Additionally, while Direct Preference Optimization (DPO) [36] provides a strong foundation for preference-based learning, we observe that different models exhibit unique strengths across various evaluation perspectives, and assessing snowfall requires a multidimensional approach to ensure robust and reliable outcomes. Therefore, we introduce the Multi-Model Preference Optimization (MMPO) framework, integrating task-specific outputs from multiple multimodal large language models (MLLMs) to construct a snowfall preference dataset. This dataset enables the fine-tuning of Q-Instruct, significantly enhancing its ability to perceive and evaluate snowfall. Finally, we integrate the MMPO fine-tuned Q-Instruct model as an unsupervised evaluation metric to update the pseudolabel database and employ a mean teacher semi-supervised framework to train existing snow removal models, thereby improving their generalization performance in real-world scenarios. Our contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.689, + 0.483, + 0.78 + ], + "angle": 0, + "content": "- Real-World Snowfall Dataset: We introduce Real-Snow10K, a large-scale, high-quality dataset comprising over 10,000 annotated real-world snow scene images. To the best of our knowledge, this is the largest dataset of real-world image desnowing, providing a robust foundation for model training and evaluation." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.781, + 0.482, + 0.87 + ], + "angle": 0, + "content": "- Snowfall Preference Dataset: From RealSnow10K, we select 1.5k real images and engage human experts to rank the outputs of multiple models based on preference. Utilizing these rankings, we construct a human preference dataset containing 36k preference pairs, designed to address two key evaluation questions." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.482, + 0.901 + ], + "angle": 0, + "content": "- MMPO for Snowfall Evaluation: We employ Multi-Model Preference Optimization (MMPO) to enhance the" + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.689, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.527, + 0.092, + 0.905, + 0.152 + ], + "angle": 0, + "content": "perception capabilities of the MLLM on snowfall images. By leveraging carefully designed multiple-choice questions, this MLLM can be used to accurately evaluate the snow removal effect of images." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.153, + 0.905, + 0.243 + ], + "angle": 0, + "content": "- Semi-Supervised Training Framework: We integrate the fine-tuned MLLMs into a mean teacher framework, enabling high-quality pseudo-label screening for semi-supervised training. The results of SnowMaster exhibits strong generalization capabilities, achieving state-of-the-art performance on real-world snowfall datasets." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.257, + 0.665, + 0.273 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.28, + 0.907, + 0.61 + ], + "angle": 0, + "content": "Single Image Desnowing. Snowflakes exhibit diverse sizes and shapes, complicating image restoration compared to other degradations [3, 8, 22-24, 26-28, 40, 49, 54]. Early methods, such as Pei et al. [34], relied on manual feature extraction but struggled in complex scenes. With deep learning advancements, More and more methods focus on snow removal tasks [6, 7, 10, 12, 14, 16, 17, 59]. Liu et al. [30] introduced DesnowNet, a multi-stage network addressing transparent and opaque snow blocks. Chen et al. [13] proposed HDCWNet, utilizing dual-tree complex wavelet transform (DTCWT) and conflicting channel loss (CC Loss) to enhance snow removal accuracy, though high-frequency information recognition remains challenging. Recent approaches integrate semantic and geometric information with Transformer architectures. Zhang et al. [59] developed DDMSNet, combining semantic segmentation and depth estimation for improved snow shape recognition. Cheng et al. [17] introduced SMGARN, using snow masks guided by self-pixel and cross-pixel attention. TransWeather [42], \\(\\mathrm{T}^3\\)-DiffWeather [11] and other adverse weather restoration methods [15, 25, 33, 55, 56] also advance robustness in extreme weather conditions." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.613, + 0.909, + 0.853 + ], + "angle": 0, + "content": "Multi-modality Large Language Models. Multimodal large language models (MLLMs) revolutionized computer vision by integrating visual and textual information. Models such as GPT-4o, MiniCPM-V [53], and LLava [29] excelled in high-level tasks like image description, visual question answering, and document understanding. However, their performance in low-level tasks, particularly image quality assessment (IQA), remained limited due to insufficient detail perception. Recent studies, such as Q-Instruct [50] and DepictQA [58], addressed this gap by incorporating low-level visual tasks into MLLM training datasets. These approaches leveraged human feedback on low-level features, such as clarity, brightness, and texture details, to improve IQA accuracy. Despite these advancements, MLLMs still struggled with weather-related degradations, as existing datasets were not fine-tuned for such conditions." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.856, + 0.908, + 0.901 + ], + "angle": 0, + "content": "Score-based IQA Methods. Score-based IQA methods aim to assess image quality by assigning numerical scores that align with human perception. Traditional methods," + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4303" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.089, + 0.905, + 0.313 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.317, + 0.908, + 0.346 + ], + "angle": 0, + "content": "Figure 2. Dataset screening and evaluation. Through a multi-round dialogue mechanism, the multimodal large language model can replace humans to complete the quality screening and content evaluation of input images." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.357, + 0.486, + 0.705 + ], + "angle": 0, + "content": "such as SSIM [46] and PSNR, relied on statistical metrics to evaluate distortions or pixel-level differences. While effective in controlled environments, they often failed to capture the complexities of real-world images. Deep learning models, such as NIMA [41] and HyperIQA [39], improved performance by regressing directly to Mean Opinion Scores (MOS) derived from human ratings. However, these models faced challenges in out-of-distribution (OOD) generalization, as unseen distortions or diverse content types significantly impacted their accuracy. Recent advancements leveraged vision-language models like CLIP [35], with methods such as CLIP-IQA+ [45] and LIQE [60] utilizing visual-text alignment for few-shot and multitask learning. These approaches demonstrated enhanced robustness across datasets but did not fully capture human-perceived quality due to their reliance on visual-text similarity scores. To address these limitations, Large Multi-Modality Models (LMMs) like Q-Align [47] simulated human rating practices by using discrete text-defined rating levels (e.g., \"excellent\", \"poor\") and converted qualitative ratings into quantitative scores through probability-based averaging. This approach enhanced both accuracy and interpretability, particularly in OOD scenarios." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.713, + 0.295, + 0.729 + ], + "angle": 0, + "content": "3. Dataset Construction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.485, + 0.812 + ], + "angle": 0, + "content": "High-quality real data is the foundation of our approach. We propose RealSnow10K, the largest real snow dataset available. In this section, we describe the process of building the dataset, including image collection, image evaluation and classification, and the dataset preference process." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.819, + 0.484, + 0.835 + ], + "angle": 0, + "content": "3.1. Image Collection, Evaluation and Classification" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.485, + 0.903 + ], + "angle": 0, + "content": "The first step in building a dataset is to collect images from various sources. We collect more than 100,000 snow scene images from the Internet and screen out over 10,000 high-quality images. To ensure the high quality of the dataset," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.357, + 0.907, + 0.538 + ], + "angle": 0, + "content": "we need to screen and classify the collected images. Images collected from the Internet often have multiple interferences. In addition to the low quality caused by blur and noise in the image itself, it may also be irrelevant or deviate from the keywords we set, which is often difficult to reflect through existing indicators. To address the above problems, traditional dataset production uses manual screening. Although this method is accurate, it is too inefficient when facing a large number of images. To conduct fast and efficient review, we introduce multimodal large language models (MLLMs) with human visual perception capabilities as a reference for evaluating image quality and content." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.539, + 0.909, + 0.825 + ], + "angle": 0, + "content": "We design a multi-round dialogue mechanism to screen out unqualified images and evaluate the content of the images. Specifically, the prompt words used in the first and second rounds of dialogue are set to \"Is this picture a real photograph? Please answer with only 'yes' or 'no'.\" and \"Does this image have a large and obvious watermark? Please answer with only 'yes' or 'no'\"; respectively, to screen out synthetic images and low-quality images with large watermarks. The third and fourth rounds of dialogue are used to evaluate the content of the pictures. In the third round of dialogue, we use multiple-choice questions to let the model choose the accurate description of the number of snowflakes from five options. We obtain the logits of each option, calculate the option probability and score, and then distinguish the amount of snowflakes and snow accumulation in different pictures. For the specific dialogue process, please refer to Fig. 2. The fourth round of dialogue is the final classification of the topic of the image, and eight categories are set for MLLM to perform classification." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.909, + 0.903 + ], + "angle": 0, + "content": "In the specific implementation, to balance evaluation quality and efficiency, we select the MiniCPM-V-2.6 model as the dataset evaluation model. Despite having only 8B parameters, this model demonstrates outstanding performance and provides a well-distributed scoring range. We parti" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4304" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.107, + 0.09, + 0.473, + 0.259 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.263, + 0.483, + 0.318 + ], + "angle": 0, + "content": "Figure 3. Performance comparison of different MLLMs for different indicators. The horizontal axis in the figure represents the rank-1 proportion of different MLLMs on different snowfall evaluation indicators." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.321, + 0.484, + 0.382 + ], + "angle": 0, + "content": "tion the RealSnow10K dataset into three subsets: 1,500 images for MMPO fine-tuning, 1,047 images for testing, and 6,406 images for semi-supervised training. Detailed category statistics are provided in the supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.392, + 0.445, + 0.409 + ], + "angle": 0, + "content": "3.2. Preference Dataset Construction Pipeline" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.414, + 0.484, + 0.641 + ], + "angle": 0, + "content": "To perform multi-model preference optimization (MMPO) on the multimodal large language model, We first use the pre-trained snow removal model to perform preliminary snow removal on 1,500 images and expand the dataset to 3,000 images. We select two indicators with poor performance of Q-instruct from the four indicators for evaluating snowfall: snowflake volume and snowflake occlusion, and design targeted questions for each. Subsequently, we collect the outputs of four MLLMs for these two problems, including Q-instruct-13b, LLava-1.5-13b, LLava-1.6-13b, and MiniCPM-V-2.6. Image evaluation experts rank the outputs of the four models one by one and generate six preference pairs from a ranking. Therefore, for each preference problem, we obtain a preference dataset with a sample size of 18,000." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.649, + 0.182, + 0.665 + ], + "angle": 0, + "content": "4. Method" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.67, + 0.345, + 0.687 + ], + "angle": 0, + "content": "4.1. Evaluation of Snowy Images" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.483, + 0.84 + ], + "angle": 0, + "content": "Since there is no reference for real snowfall images, the first task for real scene snow removal tasks is to select an accurate and reliable snowfall image evaluation standard. The size, distribution and shape of snowfall particles in the image are often affected by multiple factors including shooting equipment, wind speed, distance, etc., which also brings challenges to the evaluation of snowfall image quality. Traditional non-reference [39, 41] obviously cannot distinguish different forms of snowfall, nor can they accurately describe the interference of snowfall on the image." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Recently, the development of multimodal large language models has provided a possible solution for snowfall image evaluation. Multimodal large language models are trained on a large number of images and have close to human vi" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.257 + ], + "angle": 0, + "content": "sual capabilities. Some works [45, 47, 60] introduced CLIP or multimodal large models into image quality assessment tasks (IQA) and obtained good results. Inspired by these works, we try to introduce multimodal large language models into the field of image snow removal. However, there is currently no work exploring the performance of multimodal large language models in evaluating the degradation of snowfall images. Some works [52] use multimodal large language models from a single dimension to evaluate the quality of severe weather images, which is one-sided and unreliable." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.258, + 0.907, + 0.424 + ], + "angle": 0, + "content": "In this work, we first establish a detailed framework for evaluating snowing images. By analyzing a large dataset of real snow images, we propose assessing snow image quality across four dimensions: snowflake amount, object obscURITY, texture detail, and background clarity. The evaluation framework appears in Fig.4 (i). We select three widely used multimodal large language models and one model, Q-Instruct [47], fine-tuned for low-level visual features. These models analyze real snow images along the four dimensions, and experts evaluate their outputs. A sample output is presented in Fig.4 (ii)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.424, + 0.906, + 0.59 + ], + "angle": 0, + "content": "Unfortunately, no model can fully dominate in all four dimensions: based on LLava1.5-13b, Q-instruct, which has been fine-tuned with underlying visual features, has shown obvious advantages in texture detail and background clarity, which is consistent with our expectations. However, in terms of the two indicators of snowflake amount and object obscure for snowflake description, the best performing models are LLava1.6-13b and MiniCPM-V-2.6, which firmly occupy the first and second positions, while q-instruct-13b is relatively backward. The specific rank-1 ratios of different models are shown in Fig.3." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.596, + 0.905, + 0.627 + ], + "angle": 0, + "content": "4.2. Improving Snowfall Perception via Multi-Model Feedback Optimization" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.629, + 0.906, + 0.795 + ], + "angle": 0, + "content": "For complex image desnowing task, evaluating performance from multiple perspectives is a prerequisite for achieving satisfactory results. However, as outlined in the previous section, a single multimodal large language model cannot deliver accurate evaluations across all assessment dimensions. At the same time, employing multiple models concurrently for evaluation is computationally infeasible due to excessive resource demands. This leads us to a key question: is it possible to use a single model to simultaneously address the evaluation of multiple perspectives in desnowing tasks?" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Fortunately, Q-Instruct [47] model has already undergone targeted fine-tuning on a large-scale dataset, equipping it with the ability to assess fundamental image quality metrics, such as clarity and contrast. Building on this foundation, and inspired by recent advancements in reinforcement learning [18, 36-38], we propose enhancing Q-Instruct through human feedback optimization. Our goal is" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4305" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.089, + 0.907, + 0.361 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.363, + 0.907, + 0.419 + ], + "angle": 0, + "content": "Figure 4. Schematic diagram of the whole process of snowfall image evaluation optimization. The evaluation framework first divides snowfall into four indicators for evaluation, and for each indicator, collects the evaluation answers of multiple multimodal large language models to the image as the data source for preference learning. Finally, direct preference optimization is applied to Q-instruct to improve its output for this indicator. The figure takes the snowflake amount indicator as an example." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.431, + 0.482, + 0.461 + ], + "angle": 0, + "content": "to enable Q-Instruct to accurately evaluate snowfall, while retaining its capability to assess essential image quality" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.462, + 0.483, + 0.672 + ], + "angle": 0, + "content": "Specifically, to address the challenge that different models exhibit distinct strengths and weaknesses across various perspectives, we introduce a novel multi-model preference optimization (MMPO) technique. Unlike conventional approaches, MMPO constructs preference pairs by harnessing the outputs of multiple models, enabling a more comprehensive capture of diverse evaluation criteria. While drawing inspiration from the loss function of Direct Preference Optimization (DPO) [36], MMPO innovatively extends this framework to incorporate the collective strengths of multiple models into a single, cohesive Q-Instruct model. Consequently, the resulting Q-Instruct model excels at thoroughly assessing desnowing performance while preserving its capability to evaluate critical image quality metrics." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.673, + 0.483, + 0.735 + ], + "angle": 0, + "content": "Reward Function Reparameterization. In DPO, The reward associated with a response \\(y\\) given context \\(x\\) is reparameterized in terms of the model's policy \\(\\pi_{\\theta}(y|x)\\) and a reference policy \\(\\pi_{\\mathrm{ref}}(y|x)\\), formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.153, + 0.741, + 0.483, + 0.774 + ], + "angle": 0, + "content": "\\[\nr (x, y) = \\beta \\log \\frac {\\pi_ {\\theta} (y | x)}{\\pi_ {\\mathrm {r e f}} (y | x)} + \\beta \\log Z (x), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.483, + 0.901 + ], + "angle": 0, + "content": "where \\(\\beta\\) is a scaling constant and \\(Z(x)\\) is a partition function to normalize the policy distribution. Through this reparameterization, DPO completes preference optimization without explicitly defining a reward model, simplifying the complex reinforcement learning problem into a classification-based optimization problem. Our proposed MMPO also adopts this reparameterization technique to streamline the formulation of preference optimization." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.431, + 0.907, + 0.583 + ], + "angle": 0, + "content": "Multi-Model Feedback Collection. To construct a robust preference dataset, a multi-model feedback collection strategy is employed. For a given input context \\( x \\), responses \\( \\{y_1, y_2, \\ldots, y_N\\} \\) are collected from \\( N \\) distinct multimodal large language models. These responses are subsequently evaluated and ranked by human experts, yielding a preference ordering \\( y_{(1)} \\succ y_{(2)} \\succ \\dots \\succ y_{(N)} \\), where \\( y_{(1)} \\) denotes the most preferred response and \\( y_{(N)} \\) the least preferred. From this ordering, preference pairs \\( (y_w, y_l) \\) are generated such that \\( y_w \\succ y_l \\), forming the dataset:" + }, + { + "type": "equation", + "bbox": [ + 0.529, + 0.595, + 0.905, + 0.612 + ], + "angle": 0, + "content": "\\[\n\\left\\{\\left(x, y _ {w}, y _ {l}\\right) \\mid y _ {w} \\succ y _ {l}, y _ {w}, y _ {l} \\in \\left\\{y _ {1}, y _ {2}, \\dots , y _ {N} \\right\\} \\right\\}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.623, + 0.905, + 0.683 + ], + "angle": 0, + "content": "This methodology harnesses the diverse capabilities of multiple models to produce a comprehensive set of preference pairs, thereby enhancing the differentiation between superior and inferior responses." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.684, + 0.905, + 0.745 + ], + "angle": 0, + "content": "Multi-Model Preference Optimization. Building upon the framework of Direct Preference Optimization (DPO), Multi-Model Preference Optimization (MMPO) is introduced. The loss function is expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.594, + 0.757, + 0.905, + 0.775 + ], + "angle": 0, + "content": "\\[\nL _ {\\mathrm {M M P O}} \\left(\\pi_ {\\theta}; \\pi_ {\\mathrm {r e f}}\\right) = - \\mathbb {E} _ {(x, y _ {w}, y _ {l}) \\sim D} \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.549, + 0.785, + 0.871, + 0.82 + ], + "angle": 0, + "content": "\\[\n\\left[ \\log \\sigma \\left(\\beta \\log \\frac {\\pi_ {\\theta} \\left(y _ {w} \\mid x\\right)}{\\pi_ {\\mathrm {r e f}} \\left(y _ {w} \\mid x\\right)} - \\beta \\log \\frac {\\pi_ {\\theta} \\left(y _ {l} \\mid x\\right)}{\\pi_ {\\mathrm {r e f}} \\left(y _ {l} \\mid x\\right)}\\right) \\right],\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.907, + 0.901 + ], + "angle": 0, + "content": "where \\(\\sigma (\\cdot)\\) represents the sigmoid function, and \\(y_{w}\\) and \\(y_{l}\\) denote the preferred and unpreferred responses, respectively. A distinguishing feature of MMPO is its reliance on a preference dataset \\(D\\) derived from the outputs of multiple models, as outlined in the feedback collection process. This" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4306" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.168 + ], + "angle": 0, + "content": "approach integrates the strengths of various MLLMs, allowing the model to leverage a richer collection of preference pairs. Through this optimization, the model's outputs are aligned with human expert preferences, thereby improving its capacity to generate high-quality responses." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.176, + 0.373, + 0.193 + ], + "angle": 0, + "content": "4.3. From Perception to Restoration" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.198, + 0.484, + 0.303 + ], + "angle": 0, + "content": "Using MMPO technology, we use Q-instruct as the base model and fine-tune its preference learning in two aspects where its performance was relatively poor: snowflake amount and object obscure, improving its ability to perceive and understand snowfall. However, how can the improvement in perception be used to optimize the image restoration model?" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.304, + 0.484, + 0.409 + ], + "angle": 0, + "content": "Inspired by Q-Align [47], we leverage the enhanced perception capabilities of the large model to generate an \"image score\" that could quantify this improvement. We then use this image score as an evaluation metric within a mean teacher semi-supervised learning framework, combined with a pseudo-label database, to train the image restoration model." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.41, + 0.484, + 0.546 + ], + "angle": 0, + "content": "Score calculation. Using one adjective like Q-Align (e.g. \"excellent\", \"poor\") to describe the snowfall scene directly does not provide an accurate evaluation. We analyze that this is because a single adjective lacks a specific definition. Therefore, we further decompose the prompt words, describe the four indicators such as the number of snowflakes from different perspectives, and design it into a multiple choice question with five options. Based on the calculation method of human subjective opinion score (MOS):" + }, + { + "type": "equation", + "bbox": [ + 0.221, + 0.557, + 0.483, + 0.598 + ], + "angle": 0, + "content": "\\[\n\\mathrm {M O S} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} S _ {i}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.603, + 0.484, + 0.664 + ], + "angle": 0, + "content": "in which \\( S_{i} \\) represents the score of the i-th tester. We collect the model's output logit for the five option tokens, use softmax to calculate the probability of different tokens, and finally get the score through weighted average:" + }, + { + "type": "equation", + "bbox": [ + 0.209, + 0.675, + 0.483, + 0.716 + ], + "angle": 0, + "content": "\\[\n\\text {S c o r e} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} p _ {i} \\cdot v _ {i}, \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.231, + 0.721, + 0.483, + 0.761 + ], + "angle": 0, + "content": "\\[\np _ {i} = \\frac {e ^ {l _ {i}}}{\\sum_ {j = 1} ^ {N} e ^ {l _ {j}}}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.765, + 0.483, + 0.825 + ], + "angle": 0, + "content": "where \\( l_{i} \\) represents the logit value of the \\( i \\)-th option token. \\( N \\) is 5 to correspond to 5 options. \\( v_{i} \\) is converted into weights from 1 to 5 according to the description of snowfall intensity." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Semi-supervised Training with Pseudo-labeling. Referring to the mean teacher framework and the database mechanism in [36], we leverage scores from four evaluation metrics to enhance semi-supervised training. To ensure the pseudo-label database is updated accurately, we adopt a" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.09, + 0.687, + 0.173 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.687, + 0.09, + 0.713, + 0.11 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.715, + 0.092, + 0.906, + 0.133 + ], + "angle": 0, + "content": "How would you evaluate the amount of snowflakes in the air shown in the picture? Please provide your judgment and basis. Summarize in about 50 words." + }, + { + "type": "image", + "bbox": [ + 0.688, + 0.136, + 0.712, + 0.155 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.715, + 0.137, + 0.892, + 0.156 + ], + "angle": 0, + "content": "Keywords that do not match the real picture Keywords that lack logic or basis" + }, + { + "type": "image", + "bbox": [ + 0.688, + 0.156, + 0.712, + 0.174 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.715, + 0.16, + 0.905, + 0.178 + ], + "angle": 0, + "content": "Keywords that match the picture and perception of human experts" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.182, + 0.541, + 0.203 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.548, + 0.192, + 0.602, + 0.201 + ], + "angle": 0, + "content": "Q-instruct" + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.203, + 0.705, + 0.336 + ], + "angle": 0, + "content": "The amount of snowflakes in the air shown in the picture is quite a lot. This is evident from the fact that the people in the image are wearing winter coats and hats, and their clothing is covered in snow. The snowflakes are also clearly visible in the air, indicating that the snowfall is quite heavy. The people are walking in the snow, which further emphasizes the significant amount of snowfall. Overall, the image portrays a snowy day with a considerable amount of snowflakes in the air." + }, + { + "type": "image", + "bbox": [ + 0.731, + 0.181, + 0.763, + 0.205 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.766, + 0.191, + 0.861, + 0.2 + ], + "angle": 0, + "content": "O-instruct + MMPO" + }, + { + "type": "text", + "bbox": [ + 0.725, + 0.204, + 0.905, + 0.337 + ], + "angle": 0, + "content": "The image shows a group of people walking in the snow, with several snowflakes visible in the air. While it is not possible to quantify the exact number of snowflakes, it is evident that there is a noticeable amount of snowfall. The snowflakes are dispersed throughout the scene, indicating that the snowfall is not too heavy or too light, but rather moderate. This suggests that the weather conditions are likely to be cold and snowy, but not extreme or overwhelming." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.34, + 0.906, + 0.383 + ], + "angle": 0, + "content": "Figure 5. Comparison of perception ability before and after MMPO. The fine-tuned model output is more in line with human visual perception, more logical, and less hallucinatory." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.393, + 0.906, + 0.483 + ], + "angle": 0, + "content": "stringent rule: an update occurs only if \\((i)\\) at least three teacher scores exceed the student score, and \\((ii)\\) the teacher score surpasses the current score stored in the pseudo-label database. This dual-condition approach effectively minimizes erroneous updates to the pseudo-labels, improving the reliability of the training process." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.484, + 0.907, + 0.634 + ], + "angle": 0, + "content": "Loss function and training strategy. We use both supervised loss and pseudo-label loss to train the model. The supervised loss is based on the PSNR between the model's output and the reference images from the Snow100K [30] dataset. And the pseudo-label loss relies on the L1 norm to measure the difference between the output and pseudolabels. To ensure fast and stable training, we only use supervised loss in the initial 20k steps, and add pseudo-label loss after 20k steps for semi-supervised training. The total loss function is shown as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.53, + 0.66, + 0.905, + 0.677 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {\\text {s u p e r v i s e d}} + \\lambda \\mathcal {L} _ {\\text {p s e u d o}} (\\lambda = 0 \\text {i f s t e p < 2 0 k}). \\tag {7}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.695, + 0.646, + 0.712 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.717, + 0.691, + 0.733 + ], + "angle": 0, + "content": "5.1. Experiment Setup" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Training Settings. Our method consists of two phases: the first phase is to fine-tune the scoring model (i.e. Q-Instruct [50]) via proposed MMPO, and the second phase is to semi-supervised train the snow removal model. In the first phase, we set epoch = 1, lr = 1e-6, batchsize = 64; We use Vision-LLM Alignment framework [44] to perform MMPO fine-tuning and enable deepspeed for acceleration, set zero-stage to 3 and enable offload for full fine-tuning. In the second phase of snow removal model training, we use NAFNet [5] as the base model, set lr = 2e-4, batchsize = 32. All experiments are performed on two Nvidia A800 GPUs." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4307" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.091, + 0.904, + 0.519 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.521, + 0.907, + 0.549 + ], + "angle": 0, + "content": "Figure 6. Visual comparison of different models for snow removal on real images. Please zoom in for a better view. SnowMaster achieves satisfactory snow removal performance in complicated conditions compared to previous methods." + }, + { + "type": "table_caption", + "bbox": [ + 0.11, + 0.56, + 0.465, + 0.573 + ], + "angle": 0, + "content": "Table 1. The accuracy of MLLMs on different perspectives." + }, + { + "type": "table", + "bbox": [ + 0.102, + 0.575, + 0.482, + 0.681 + ], + "angle": 0, + "content": "
MLLMs AccuracySnowflakes accQuality acc
amount ↑obscure ↑texture ↑background ↑
LLava-v1.5-13b[29]80.0477.1755.7852.15
LLava-v1.6-13b[29]86.3479.5652.0554.06
MiniCPM-V[53]75.4570.0442.0249.76
Q-intract-13b[50]76.3173.2658.3663.42
Q-instruct with MMPO (ours)82.1480.9064.9568.58
MMPO improvement+5.83+7.64+6.59+5.16
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.701, + 0.483, + 0.821 + ], + "angle": 0, + "content": "Experimental Details. In the semi-supervised training process, considering the need to make the multimodal large language model output the correct score, we adopt a pseudo-label database initialization strategy using the original image instead of all-0 initialization in [21]. In the test phase, we test 457 test images selected from the real snow scene dataset (RealSnow10K) and 1329 real snow scene images from Snow100K [30]." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.832, + 0.383, + 0.847 + ], + "angle": 0, + "content": "5.2. Evaluation Metrics and Baselines" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Following [19, 20], we utilize four non-reference IQA metrics: BRISQUE [31], NIQE [32], PIQE [43], and PAQ2-PIQ [57]. BRISQUE measures quality via spa" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.561, + 0.907, + 0.713 + ], + "angle": 0, + "content": "tial natural scene statistics, NIQE assesses deviations from these statistics, PIQE detects block-level distortions, and PAQ2-PIQ, a deep learning approach, predicts global and local quality. These metrics ensure a comprehensive, unsupervised evaluation of diverse distortions. For the baselines, we choose CNN-based models (SMGARN[17], NAFNet[5]), Transformer-based models (Transweather[42], Snowformer[6]) and a diffusion based model \\((\\mathrm{T}^3\\) -DiffWeather[11]), offering varied references for snowfall image restoration performance." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.724, + 0.709, + 0.74 + ], + "angle": 0, + "content": "5.3. Results and Analysis" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.744, + 0.704, + 0.759 + ], + "angle": 0, + "content": "5.3.1. Quantitative Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Effectiveness of MMPO. Table 1 illustrates the performance enhancements of the Q-Instruct [47] model following fine-tuning with MMPO. We utilize NAFNet [5], pretrained for 100,000 steps, as the snow removal model to process the images in the test set, generating a dataset that includes both original and snow-removed images. By inputting these images individually into Q-Instruct, we obtain scores for each image corresponding to specific indicators. Consequently, for each indicator, accuracy is defined" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4308" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.224, + 0.086, + 0.772, + 0.1 + ], + "angle": 0, + "content": "Table 2. Comparison of non-reference metrics after snow removal between different models." + }, + { + "type": "table", + "bbox": [ + 0.132, + 0.103, + 0.865, + 0.226 + ], + "angle": 0, + "content": "
MethodSnow100K-realistic [30]RealSnow10K
BRISQUE ↓NIQE ↓PIQE ↓PAQ2-PIQ ↑BRISQUE ↓NIQE ↓PIQE ↓PAQ2-PIQ ↑
SMGARN[TCSVT'22][17]19.2893.13231.92769.76723.0053.58031.21068.919
TransWeather[CVPR'22] [42]16.5443.02530.30269.82118.5643.40429.77269.671
NAFNet[ECCV'22] [5]17.5662.90330.84369.95120.4253.29129.87769.388
SnowFormer[Arxiv'23][6]19.3892.95431.90270.01523.7023.42931.9969.929
T3-DiffWeather[ECCV'24] [11]17.7442.86831.68970.04220.2113.13732.99170.211
★ SnowMaster (Ours)16.3052.79730.16270.04217.5463.05829.74270.304
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.244, + 0.482, + 0.409 + ], + "angle": 0, + "content": "as the proportion of snow-removed images whose score is not lower than that of their corresponding original images. The results show that the performance after fine-tuning has been significantly improved which is consistent with our expectations. It is worth noting that in terms of the number of snowflakes, the accuracy after fine-tuning does not exceed that of Llava-1.6. We think this is normal, because the preference data in fine-tuning are all output by these models, and MMPO guides the model to move closer to human preference output, but it may not necessarily surpass the output of the optimal model." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.41, + 0.482, + 0.516 + ], + "angle": 0, + "content": "Furthermore, a noteworthy observation emerges: the indicators 'texture' and 'background', which are not explicitly targeted for optimization, also exhibit moderate improvements. We hypothesize that this phenomenon arises from a comprehensive augmentation of the model's perceptual capacity regarding snowfall, facilitated by the MMPO fine-tuning process." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.516, + 0.482, + 0.577 + ], + "angle": 0, + "content": "Performance analysis of the restoration model. Table 2 presents the performance comparison between our method and other advanced snow removal models. Our proposed method achieves the best results in most indicators." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.583, + 0.3, + 0.598 + ], + "angle": 0, + "content": "5.3.2. Qualitative Assessment" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.602, + 0.483, + 0.785 + ], + "angle": 0, + "content": "From the restoration results of the real snowfall image in Fig.6, we can see that the SnowMaster model we proposed has the best performance in multiple scenarios. When facing snowflakes illuminated by lights at night, other models cannot effectively complete the restoration of high-brightness snowflakes, while SnowMaster significantly eliminates the high-brightness snowflakes in the image. When restoring dense snowflake scenes during the day, SnowMaster can effectively remove snowflakes of different scales and retain the original color and texture of the object, while other models have poor restoration effects on small-sized snowflake particles." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.791, + 0.245, + 0.807 + ], + "angle": 0, + "content": "5.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.482, + 0.901 + ], + "angle": 0, + "content": "Necessity of MMPO. The Q-Instruct model shows significant improvement after multi-modal preference optimization, particularly in evaluating snowfall images. We perform an ablation test to compare the Q-Instruct model before and after alignment when used as a scoring model in semi-supervised training. The results indicate that the" + }, + { + "type": "table_caption", + "bbox": [ + 0.514, + 0.242, + 0.905, + 0.255 + ], + "angle": 0, + "content": "Table 3. Ablation on different settings of the proposed framework." + }, + { + "type": "table", + "bbox": [ + 0.532, + 0.258, + 0.892, + 0.351 + ], + "angle": 0, + "content": "
MethodBRISQUE ↓NIQE ↓
w/o. Semi-supervised learning22.523.33
w/o. MLLM20.363.25
w/o. MMPO18.983.19
w. Single-angle evaluation18.213.12
w. Multi-angle evalutaion (Ours)17.553.06
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.364, + 0.905, + 0.394 + ], + "angle": 0, + "content": "aligned model achieves higher accuracy and ensures more stable training to improve the performance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.396, + 0.905, + 0.501 + ], + "angle": 0, + "content": "Effectiveness of multi-perspectives snowfall evaluation. Our method enhances image screening reliability by evaluating snowfall images from four perspectives, ensuring steady updates to the pseudo-label database. Ablation experiments, conducted by reducing evaluation angles, demonstrate that multi-angle evaluation improves training effectiveness" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.503, + 0.906, + 0.653 + ], + "angle": 0, + "content": "The role of the semi-supervised learning framework. In order to transfer the improvement of the evaluation performance of the large model on snow scene images to the improvement of the restoration model, we use the mean teacher semi-supervised learning framework. By combining the supervision loss from snow100k and the pseudolabel loss from the real training data, this framework can effectively optimize the restoration effect of real snowfall images. Through ablation experiments, we prove that the semi-supervised framework is the basis of our method." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.655, + 0.905, + 0.684 + ], + "angle": 0, + "content": "The specific metrics of the ablation experiments are shown in Table 3." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.697, + 0.634, + 0.712 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.906, + 0.901 + ], + "angle": 0, + "content": "SnowMaster integrates a multimodal large language model with a multi-perspective evaluation framework for image desnowing. Leveraging the RealSnow10k dataset, it utilizes a multi-angle evaluation system assessing four key metrics to enhance real-world reliability. Through multimodal preference optimization and semi-supervised training with pseudo-label updates from multi-angle multiple-choice questions, it ensures effective refinement. Experiments confirm its robustness in complex snowfall scenarios, though its computational intensity from multiple evaluations signals a need for future enhancements in training efficiency and resource use." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4309" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.251, + 0.108 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.116, + 0.484, + 0.193 + ], + "angle": 0, + "content": "This work is supported by the Guangzhou-HKUST(GZ) Joint Funding Program (No. 2023A03J0671), the Guangzhou Municipal Science and Technology Project (Grant No. 2024A04J4230), and Guangzhou-HKUST(GZ) Joint Funding Program (No. 2024A03J0618)." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.208, + 0.188, + 0.223 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.233, + 0.484, + 0.302 + ], + "angle": 0, + "content": "[1] Vijay Badrinarayanan, Alex Kendall, and Roberto Cipolla. Segnet: A deep convolutional encoder-decoder architecture for image segmentation. IEEE transactions on pattern analysis and machine intelligence, 39(12):2481-2495, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.304, + 0.484, + 0.373 + ], + "angle": 0, + "content": "[2] Chun-Fu Richard Chen, Quanfu Fan, and Rameswar Panda. Crossvit: Cross-attention multi-scale vision transformer for image classification. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 357-366, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.375, + 0.485, + 0.457 + ], + "angle": 0, + "content": "[3] Haoyu Chen, Jinjin Gu, Yihao Liu, Salma Abdel Magid, Chao Dong, Qiong Wang, Hanspeter Pfister, and Lei Zhu. Masked image training for generalizable deep image denoising. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1692-1703, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.46, + 0.484, + 0.544 + ], + "angle": 0, + "content": "[4] Haoyu Chen, Wenbo Li, Jinjin Gu, Jingjing Ren, Haoze Sun, Xueyi Zou, Zhensong Zhang, Youliang Yan, and Lei Zhu. Low-res leads the way: Improving generalization for superresolution by self-supervised learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25857-25867, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.546, + 0.484, + 0.6 + ], + "angle": 0, + "content": "[5] Liangyu Chen, Xiaojie Chu, Xiangyu Zhang, and Jian Sun. Simple baselines for image restoration. In European conference on computer vision, pages 17-33. Springer, 2022. 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.603, + 0.484, + 0.658 + ], + "angle": 0, + "content": "[6] Sixiang Chen, Tian Ye, Yun Liu, and Erkang Chen. Snowformer: Context interaction transformer with scale-awareness for single image desnowing. arXiv preprint arXiv:2208.09703, 2022. 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.66, + 0.484, + 0.715 + ], + "angle": 0, + "content": "[7] Sixiang Chen, Tian Ye, Yun Liu, Taodong Liao, Yi Ye, and Erkang Chen. Msp-former: Multi-scale projection transformer for single image desnowing. arXiv preprint arXiv:2207.05621, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.717, + 0.484, + 0.799 + ], + "angle": 0, + "content": "[8] Sixiang Chen, Tian Ye, Jinbin Bai, Erkang Chen, Jun Shi, and Lei Zhu. Sparse sampling transformer with uncertainty-driven ranking for unified removal of raindrops and rain streaks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13106-13117, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.803, + 0.484, + 0.871 + ], + "angle": 0, + "content": "[9] Sixiang Chen, Tian Ye, Yun Liu, Jinbin Bai, Haoyu Chen, Yunlong Lin, Jun Shi, and Erkang Chen. CPLformer: Cross-scale prototype learning transformer for image snow removal. In Proceedings of the 31st ACM international conference on multimedia, pages 4228-4239, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.483, + 0.902 + ], + "angle": 0, + "content": "[10] Sixiang Chen, Tian Ye, Chenghao Xue, Haoyu Chen, Yun Liu, Erkang Chen, and Lei Zhu. Uncertainty-driven dynamic" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.233, + 0.485, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.907, + 0.147 + ], + "angle": 0, + "content": "degradation perceiving and background modeling for efficient single image desnowing. In Proceedings of the 31st ACM International Conference on Multimedia, pages 4269-4280, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.15, + 0.906, + 0.219 + ], + "angle": 0, + "content": "[11] Sixiang Chen, Tian Ye, Kai Zhang, Zhaohu Xing, Yunlong Lin, and Lei Zhu. Teaching tailored to talent: Adverse weather restoration via prompt pool and depth-anything constraint. In European Conference on Computer Vision, pages 95–115. Springer, 2025. 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.221, + 0.906, + 0.318 + ], + "angle": 0, + "content": "[12] Wei-Ting Chen, Hao-Yu Fang, Jian-Jiun Ding, Cheng-Che Tsai, and Sy-Yen Kuo. Jstasr: Joint size and transparency-aware snow removal algorithm based on modified partial convolution and veiling effect removal. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXI 16, pages 754-770. Springer, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.321, + 0.906, + 0.403 + ], + "angle": 0, + "content": "[13] Wei-Ting Chen, Hao-Yu Fang, Cheng-Lin Hsieh, Cheng-Che Tsai, I Chen, Jian-Jiun Ding, Sy-Yen Kuo, et al. All snow removed: Single image desnowing algorithm using hierarchical dual-tree complex wavelet representation and contradict channel loss. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4196-4205, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.405, + 0.906, + 0.5 + ], + "angle": 0, + "content": "[14] Wei-Ting Chen, Hao-Yu Fang, Cheng-Lin Hsieh, Cheng-Che Tsai, I Chen, Jian-Jiun Ding, Sy-Yen Kuo, et al. All snow removed: Single image desnowing algorithm using hierarchical dual-tree complex wavelet representation and contradict channel loss. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4196-4205, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.504, + 0.906, + 0.599 + ], + "angle": 0, + "content": "[15] Wei-Ting Chen, Zhi-Kai Huang, Cheng-Che Tsai, Hao-Hsiang Yang, Jian-Jiun Ding, and Sy-Yen Kuo. Learning multiple adverse weather removal via two-stage knowledge learning and multi-contrastive regularization: Toward a unified model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17653–17662, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.603, + 0.906, + 0.657 + ], + "angle": 0, + "content": "[16] Bodong Cheng, Juncheng Li, Ying Chen, Shuyi Zhang, and Tieyong Zeng. Snow mask guided adaptive residual network for image snow removal. arXiv preprint arXiv:2207.04754, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.66, + 0.905, + 0.715 + ], + "angle": 0, + "content": "[17] Bodong Cheng, Juncheng Li, Ying Chen, and Tieyong Zeng. Snow mask guided adaptive residual network for image snow removal. Computer Vision and Image Understanding, 236: 103819, 2023. 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.718, + 0.906, + 0.773 + ], + "angle": 0, + "content": "[18] Paul F Christiano, Jan Leike, Tom Brown, Miljan Martic, Shane Legg, and Dario Amodei. Deep reinforcement learning from human preferences. Advances in neural information processing systems, 30, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.775, + 0.906, + 0.83 + ], + "angle": 0, + "content": "[19] Jinjin Gu, Haoming Cai, Haoyu Chen, Xiaoxing Ye, Jimmy Ren, and Chao Dong. Image quality assessment for perceptual image restoration: A new dataset, benchmark and metric. arXiv preprint arXiv:2011.15002, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.832, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[20] Jinjin Gu, Haoming Cai, Haoyu Chen, Xiaoxing Ye, Jimmy Ren, and Chao Dong. Pipal: a large-scale image quality assessment dataset for perceptual image restoration. In European Conference on Computer Vision, pages 633-651. Springer, 2020. 7" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "4310" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.161 + ], + "angle": 0, + "content": "[21] Shirui Huang, Keyan Wang, Huan Liu, Jun Chen, and Yun-song Li. Contrastive semi-supervised learning for underwater image restoration via reliable bank. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18145-18155, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.164, + 0.485, + 0.232 + ], + "angle": 0, + "content": "[22] Yeying Jin, Wending Yan, Wenhan Yang, and Robby T Tan. Structure representation network and uncertainty feedback learning for dense non-uniform fog removal. In Asian Conference on Computer Vision, pages 155-172. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.234, + 0.484, + 0.29 + ], + "angle": 0, + "content": "[23] Yeying Jin, Wenhan Yang, Wei Ye, Yuan Yuan, and Robby T Tan. Shadowdiffusion: Diffusion-based shadow removal using classifier-driven attention and structure preservation. arXiv preprint arXiv:2211.08089, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.292, + 0.484, + 0.346 + ], + "angle": 0, + "content": "[24] Yeying Jin, Beibei Lin, Wending Yan, Wei Ye, Yuan Yuan, and Robby T. Tan. Enhancing visibility in nighttime haze images using guided apsf and gradient adaptive convolution, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.349, + 0.484, + 0.404 + ], + "angle": 0, + "content": "[25] Ruoteng Li, Robby T Tan, and Loong-Fah Cheong. All in one bad weather removal using architectural search. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3175-3185, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.406, + 0.484, + 0.475 + ], + "angle": 0, + "content": "[26] Yunlong Lin, Zhenqi Fu, Ge Meng, Yingying Wang, Yuhang Dong, Linyu Fan, Hedeng Yu, and Xinghao Ding. Domain-irrelevant feature learning for generalizable pan-sharpening. In Proceedings of the 31st ACM International Conference on Multimedia, pages 3287–3296, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.477, + 0.484, + 0.546 + ], + "angle": 0, + "content": "[27] Yunlong Lin, Zhenqi Fu, Kairun Wen, Tian Ye, Sixiang Chen, Ge Meng, Yingying Wang, Yue Huang, Xiaotong Tu, and Xinghao Ding. Unsupervised low-light image enhancement with lookup tables and diffusion priors. arXiv preprint arXiv:2409.18899, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.548, + 0.484, + 0.616 + ], + "angle": 0, + "content": "[28] Yunlong Lin, Tian Ye, Sixiang Chen, Zhenqi Fu, Yingying Wang, Wenhao Chai, Zhaohu Xing, Lei Zhu, and Xinghao Ding. Aglldiff: Guiding diffusion models towards unsupervised training-free real-world low-light image enhancement. arXiv preprint arXiv:2407.14900, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.618, + 0.484, + 0.686 + ], + "angle": 0, + "content": "[29] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26296-26306, 2024. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.688, + 0.484, + 0.743 + ], + "angle": 0, + "content": "[30] Yun-Fu Liu, Da-Wei Jaw, Shih-Chia Huang, and Jenq-Neng Hwang. Desnownet: Context-aware deep network for snow removal. IEEE Transactions on Image Processing, 27(6): 3064–3073, 2018. 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.746, + 0.484, + 0.801 + ], + "angle": 0, + "content": "[31] Anish Mittal, Anush Krishna Moorthy, and Alan Conrad Bovik. No-reference image quality assessment in the spatial domain. IEEE Transactions on image processing, 21(12): 4695-4708, 2012. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.803, + 0.484, + 0.845 + ], + "angle": 0, + "content": "[32] Anish Mittal, Rajiv Soundararajan, and Alan C Bovik. Making a “completely blind” image quality analyzer. IEEE Signal processing letters, 20(3):209-212, 2012. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.484, + 0.902 + ], + "angle": 0, + "content": "[33] Ozan Özdenizci and Robert Legenstein. Restoring vision in adverse weather conditions with patch-based denoising diffusion models. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.906, + 0.161 + ], + "angle": 0, + "content": "[34] Soo-Chang Pei, Yu-Tai Tsai, and Chen-Yu Lee. Removing rain and snow in a single image using saturation and visibility features. In 2014 IEEE International Conference on Multimedia and Expo Workshops (ICMEW), pages 1-6. IEEE, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.164, + 0.906, + 0.245 + ], + "angle": 0, + "content": "[35] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.249, + 0.906, + 0.318 + ], + "angle": 0, + "content": "[36] Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. Direct preference optimization: Your language model is secretly a reward model. Advances in Neural Information Processing Systems, 36, 2024. 2, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.32, + 0.906, + 0.374 + ], + "angle": 0, + "content": "[37] John Schulman, Sergey Levine, Pieter Abbeel, Michael Jordan, and Philipp Moritz. Trust region policy optimization. In International conference on machine learning, pages 1889-1897. PMLR, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.377, + 0.906, + 0.418 + ], + "angle": 0, + "content": "[38] John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.42, + 0.906, + 0.49 + ], + "angle": 0, + "content": "[39] Shaolin Su, Qingsen Yan, Yu Zhu, Cheng Zhang, Xin Ge, Jinqiu Sun, and Yanning Zhang. Blindly assess image quality in the wild guided by a self-adaptive hyper network. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3667-3676, 2020. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.491, + 0.906, + 0.572 + ], + "angle": 0, + "content": "[40] Haoze Sun, Wenbo Li, Jianzhuang Liu, Haoyu Chen, Renjing Pei, Xueyi Zou, Youliang Yan, and Yujiu Yang. Coser: Bridging image and language for cognitive super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25868-25878, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.576, + 0.906, + 0.616 + ], + "angle": 0, + "content": "[41] Hossein Talebi and Peyman Milanfar. Nima: Neural image assessment. IEEE transactions on image processing, 27(8): 3998-4011, 2018. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.619, + 0.906, + 0.701 + ], + "angle": 0, + "content": "[42] Jeya Maria Jose Valanarasu, Rajeev Yasarla, and Vishal M Patel. Transweather: Transformer-based restoration of images degraded by adverse weather conditions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2353-2363, 2022. 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.704, + 0.906, + 0.773 + ], + "angle": 0, + "content": "[43] Narasimhan Venkatanath, D Praneeth, Maruthi Chandrasekhar Bh, Sumohana S Channappayya, and Swarup S Medasani. Blind image quality evaluation using perception based features. In 2015 twenty first national conference on communications (NCC), pages 1-6. IEEE, 2015. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.775, + 0.906, + 0.843 + ], + "angle": 0, + "content": "[44] Chenglong Wang, Yang Gan, Yifu Huo, Yongyu Mu, Murun Yang, Qiaozhi He, Tong Xiao, Chunliang Zhang, Tongran Liu, Quan Du, Di Yang, and Jingbo Zhu. Rovrm: A robust visual reward model optimized via auxiliary textual preference data, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.906, + 0.902 + ], + "angle": 0, + "content": "[45] Jianyi Wang, Kelvin CK Chan, and Chen Change Loy. Exploring clip for assessing the look and feel of images. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2555-2563, 2023. 3, 4" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.906, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.514, + 0.957 + ], + "angle": 0, + "content": "4311" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[46] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.15, + 0.482, + 0.219 + ], + "angle": 0, + "content": "[47] Haoning Wu, Zicheng Zhang, Weixia Zhang, Chaofeng Chen, Liang Liao, Chunyi Li, Yixuan Gao, Annan Wang, Erli Zhang, Wenxiu Sun, et al. Q-align: Teaching Imms for visual scoring via discrete text-defined levels. arXiv preprint arXiv:2312.17090, 2023. 1, 3, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.221, + 0.482, + 0.303 + ], + "angle": 0, + "content": "[48] Hongtao Wu, Yijun Yang, Angelica I Aviles-Rivero, Jingjing Ren, Sixiang Chen, Haoyu Chen, and Lei Zhu. Semi-supervised video desnowing network via temporal decoupling experts and distribution-driven contrastive regularization. In European Conference on Computer Vision, pages 70-89. Springer, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.306, + 0.482, + 0.375 + ], + "angle": 0, + "content": "[49] Hongtao Wu, Yijun Yang, Huihui Xu, Weiming Wang, Jinni Zhou, and Lei Zhu. Rainmamba: Enhanced locality learning with state space models for video deraining. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 7881-7890, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.377, + 0.482, + 0.473 + ], + "angle": 0, + "content": "[50] Haoning Wu, Zicheng Zhang, Erli Zhang, Chaofeng Chen, Liang Liao, Annan Wang, Kaixin Xu, Chunyi Li, Jingwen Hou, Guangtao Zhai, et al. Q-instruct: Improving low-level visual abilities for multi-modality foundation models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25490-25500, 2024. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.476, + 0.482, + 0.544 + ], + "angle": 0, + "content": "[51] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. Advances in Neural Information Processing Systems, 34, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.547, + 0.482, + 0.616 + ], + "angle": 0, + "content": "[52] Jiaqi Xu, Mengyang Wu, Xiaowei Hu, Chi-Wing Fu, Qi Dou, and Pheng-Ann Heng. Towards real-world adverse weather image restoration: Enhancing clearness and semantics with vision-language models. In European Conference on Computer Vision, pages 147–164. Springer, 2025. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.617, + 0.482, + 0.673 + ], + "angle": 0, + "content": "[53] Yuan Yao, Tianyu Yu, Ao Zhang, Chongyi Wang, Junbo Cui, Hongji Zhu, Tianchi Cai, Haoyu Li, Weilin Zhao, Zhihui He, et al. Minicpm-v: A gpt-4v level mllm on your phone. arXiv preprint arXiv:2408.01800, 2024. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.675, + 0.482, + 0.731 + ], + "angle": 0, + "content": "[54] Tian Ye, Yunchen Zhang, Mingchao Jiang, Liang Chen, Yun Liu, Sixiang Chen, and Erkang Chen. Perceiving and modeling density for image dehazing. In European Conference on Computer Vision, pages 130-145. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.732, + 0.482, + 0.8 + ], + "angle": 0, + "content": "[55] Tian Ye, Sixiang Chen, Jinbin Bai, Jun Shi, Chenghao Xue, Jingxia Jiang, Junjie Yin, Erkang Chen, and Yun Liu. Adverse weather removal with codebook priors. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12653-12664, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.802, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[56] Tian Ye, Sixiang Chen, Wenhao Chai, Zhaohu Xing, Jing Qin, Ge Lin, and Lei Zhu. Learning diffusion texture priors for image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2524-2534, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.902 + ], + "angle": 0, + "content": "[57] Zhenqiang Ying, Haoran Niu, Praful Gupta, Dhruv Mahajan, Deepti Ghadiyaram, and Alan Bovik. From patches to" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "pictures (paq-2-piq): Mapping the perceptual space of picture quality. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3575-3585, 2020. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.15, + 0.905, + 0.205 + ], + "angle": 0, + "content": "[58] Zhiyuan You, Zheyuan Li, Jinjin Gu, Zhenfei Yin, Tianfan Xue, and Chao Dong. Depicting beyond scores: Advancing image quality assessment through multi-modal language models. arXiv preprint arXiv:2312.08962, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.207, + 0.905, + 0.261 + ], + "angle": 0, + "content": "[59] Kaihao Zhang, Rongqing Li, Yanjiang Yu, Wenhan Luo, and Changsheng Li. Deep dense multi-scale network for snow removal using semantic and depth priors. IEEE Transactions on Image Processing, 30:7419-7431, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.263, + 0.905, + 0.343 + ], + "angle": 0, + "content": "[60] Weixia Zhang, Guangtao Zhai, Ying Wei, Xiaokang Yang, and Kede Ma. Blind image quality assessment via vision-language correspondence: A multitask learning perspective. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14071-14081, 2023. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.346, + 0.905, + 0.416 + ], + "angle": 0, + "content": "[61] Haipeng Zhou, Hongqiu Wang, Tian Ye, Zhaohu Xing, Jun Ma, Ping Li, Qiong Wang, and Lei Zhu. Timeline and boundary guided diffusion network for video shadow detection. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 166-175, 2024. 1" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.416 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4312" + } + ] +] \ No newline at end of file diff --git a/2025/SnowMaster_ Comprehensive Real-world Image Desnowing via MLLM with Multi-Model Feedback Optimization/01611a60-cb24-4ca8-93fd-c47c38025543_origin.pdf b/2025/SnowMaster_ Comprehensive Real-world Image Desnowing via MLLM with Multi-Model Feedback Optimization/01611a60-cb24-4ca8-93fd-c47c38025543_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..17f4371c2b2ac554c0400799bf503e38173af33e --- /dev/null +++ b/2025/SnowMaster_ Comprehensive Real-world Image Desnowing via MLLM with Multi-Model Feedback Optimization/01611a60-cb24-4ca8-93fd-c47c38025543_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bf1f09102a8be1d99234c4a4cf291930aa2532293b078f9aa2af42fc6cfe13a6 +size 2382841 diff --git a/2025/SnowMaster_ Comprehensive Real-world Image Desnowing via MLLM with Multi-Model Feedback Optimization/full.md b/2025/SnowMaster_ Comprehensive Real-world Image Desnowing via MLLM with Multi-Model Feedback Optimization/full.md new file mode 100644 index 0000000000000000000000000000000000000000..2f295cc94829d241c06dc9cac9f650a9f5537763 --- /dev/null +++ b/2025/SnowMaster_ Comprehensive Real-world Image Desnowing via MLLM with Multi-Model Feedback Optimization/full.md @@ -0,0 +1,327 @@ +# SnowMaster: Comprehensive Real-world Image Desnowing via MLLM with Multi-Model Feedback Optimization + +Jianyu Lai $^{1,*}$ Sixiang Chen $^{1,*}$ Yunlong Lin $^{2}$ Tian Ye $^{1}$ +Yun Liu $^{3}$ Song Fei $^{1}$ Zhaohu Xing $^{1}$ Hongtao Wu $^{1}$ Weiming Wang $^{4}$ Lei Zhu $^{1,5,\dagger}$ $^{1}$ The Hong Kong University of Science and Technology (GZ) $^{2}$ Xiamen University + $^{3}$ Southwestern University $^{4}$ Hong Kong Metropolitan University + $^{5}$ The Hong Kong University of Science and Technology +Project page: https://alexlai2860.github.io/SnowMaster + +# Abstract + +Snowfall presents significant challenges for visual data processing, necessitating specialized desnowing algorithms. However, existing models often fail to generalize effectively due to their heavy reliance on synthetic datasets. Furthermore, current real-world snowfall datasets are limited in scale and lack dedicated evaluation metrics designed specifically for snowfall degradation, thus hindering the effective integration of real snowy images into model training to reduce domain gaps. To address these challenges, we first introduce RealSnow10K, a large-scale, high-quality dataset consisting of over 10,000 annotated real-world snowy images. In addition, we curate a preference dataset comprising 36,000 expert-ranked image pairs, enabling the adaptation of multimodal large language models (MLLMs) to better perceive snowy image quality through our innovative Multi-Model Preference Optimization (MMPO). Finally, we propose the SnowMaster, which employs MMPO-enhanced MLLM to perform accurate snowy image evaluation and pseudo-label filtering for semi-supervised training. Experiments demonstrate that SnowMaster delivers superior desnowing performance under real-world conditions. + +# 1. Introduction + +Snowfall, a common form of adverse weather degradation, reduces image visibility and adversely impacts downstream tasks such as image recognition [2, 61] and segmentation [1, 51]. The inherent complexity of snow poses a challenge to conventional image restoration techniques, necessitating the development of specialized snow removal (desnowing) algorithms to enhance image clarity. + +Although current snow removal models achieve impressive performance on benchmark datasets [7, 9, 12, 14, 16], + +![](images/cd2fbfead21ced4a3c83c22752542120b1f6d887472dfca22a76b344a8b7032c.jpg) +Figure 1. Comparison of various image desnowing evaluation metrics shows that Q-Instruct [47], enhanced by our MMPO, most accurately perceives differences in snow density between images. Leveraging accurate evaluation model, we optimize our semi-supervised network, achieving superior desnowing performance. + +they often struggle to generalize to real-world scenarios. This limitation stems primarily from their only reliance on synthetic snowfall datasets for training. These datasets, generated using simplified physical models, struggle to capture the intricate details of real-world snow degradations, resulting in a domain gap between synthetic and real data. This gap hinders model optimization and impairs generalization to real-world scenarios. Consequently, we pose the question: How can real-world data be effectively utilized to enhance the model's generalization performance on real-world samples? + +Incorporating real-world snowy images into training is an direct and effective approach to enhance model generalization. However, accurate evaluation of these images is a prerequisite for their successful integration into the training process [4, 48, 50]. Yet, evaluating realistic snowfall images remains a significant challenge due to the absence of ground truth for reference evaluation, previous methods often rely on non-reference metrics [31, 32, 43, 57] to assess image restoration quality. These metrics, influenced by numerous factors, fail to accurately isolate and evalu- + +ate weather-specific degradation in snowfall images. In this paper, we claim that accurate assessment of snowy images requires a joint effort from both low-level visual and semantic perspectives. Recent advancements in multimodal large language models (MLLMs), such as Q-Instruct [50] and Depict-QA [58], offer promising avenues for developing evaluation frameworks due to their ability to assess low-level visual features from a semantic perspective. However, these models currently lack the specialized perception capabilities required to reliably evaluate snowfall-specific characteristics. Conversely, while some more advanced multimodal large language models exhibit better understanding of snowy scenes, they are not fine-tuned for detailed low-level visual evaluation, limiting the fundamental premise as powerful evaluators on the snow removal task. + +To tackle these challenges, we first establish a comprehensive evaluation framework designed to capture the distinct characteristics of snow in natural images through multi-perspective feedback optimization tailored to snow scenes. To support this framework, we curate and annotate a dataset comprising over 10,000 real-world snow images sourced from the Internet. Additionally, while Direct Preference Optimization (DPO) [36] provides a strong foundation for preference-based learning, we observe that different models exhibit unique strengths across various evaluation perspectives, and assessing snowfall requires a multidimensional approach to ensure robust and reliable outcomes. Therefore, we introduce the Multi-Model Preference Optimization (MMPO) framework, integrating task-specific outputs from multiple multimodal large language models (MLLMs) to construct a snowfall preference dataset. This dataset enables the fine-tuning of Q-Instruct, significantly enhancing its ability to perceive and evaluate snowfall. Finally, we integrate the MMPO fine-tuned Q-Instruct model as an unsupervised evaluation metric to update the pseudolabel database and employ a mean teacher semi-supervised framework to train existing snow removal models, thereby improving their generalization performance in real-world scenarios. Our contributions can be summarized as follows: + +- Real-World Snowfall Dataset: We introduce Real-Snow10K, a large-scale, high-quality dataset comprising over 10,000 annotated real-world snow scene images. To the best of our knowledge, this is the largest dataset of real-world image desnowing, providing a robust foundation for model training and evaluation. +- Snowfall Preference Dataset: From RealSnow10K, we select 1.5k real images and engage human experts to rank the outputs of multiple models based on preference. Utilizing these rankings, we construct a human preference dataset containing 36k preference pairs, designed to address two key evaluation questions. +- MMPO for Snowfall Evaluation: We employ Multi-Model Preference Optimization (MMPO) to enhance the + +perception capabilities of the MLLM on snowfall images. By leveraging carefully designed multiple-choice questions, this MLLM can be used to accurately evaluate the snow removal effect of images. + +- Semi-Supervised Training Framework: We integrate the fine-tuned MLLMs into a mean teacher framework, enabling high-quality pseudo-label screening for semi-supervised training. The results of SnowMaster exhibits strong generalization capabilities, achieving state-of-the-art performance on real-world snowfall datasets. + +# 2. Related Works + +Single Image Desnowing. Snowflakes exhibit diverse sizes and shapes, complicating image restoration compared to other degradations [3, 8, 22-24, 26-28, 40, 49, 54]. Early methods, such as Pei et al. [34], relied on manual feature extraction but struggled in complex scenes. With deep learning advancements, More and more methods focus on snow removal tasks [6, 7, 10, 12, 14, 16, 17, 59]. Liu et al. [30] introduced DesnowNet, a multi-stage network addressing transparent and opaque snow blocks. Chen et al. [13] proposed HDCWNet, utilizing dual-tree complex wavelet transform (DTCWT) and conflicting channel loss (CC Loss) to enhance snow removal accuracy, though high-frequency information recognition remains challenging. Recent approaches integrate semantic and geometric information with Transformer architectures. Zhang et al. [59] developed DDMSNet, combining semantic segmentation and depth estimation for improved snow shape recognition. Cheng et al. [17] introduced SMGARN, using snow masks guided by self-pixel and cross-pixel attention. TransWeather [42], $\mathrm{T}^3$ -DiffWeather [11] and other adverse weather restoration methods [15, 25, 33, 55, 56] also advance robustness in extreme weather conditions. + +Multi-modality Large Language Models. Multimodal large language models (MLLMs) revolutionized computer vision by integrating visual and textual information. Models such as GPT-4o, MiniCPM-V [53], and LLava [29] excelled in high-level tasks like image description, visual question answering, and document understanding. However, their performance in low-level tasks, particularly image quality assessment (IQA), remained limited due to insufficient detail perception. Recent studies, such as Q-Instruct [50] and DepictQA [58], addressed this gap by incorporating low-level visual tasks into MLLM training datasets. These approaches leveraged human feedback on low-level features, such as clarity, brightness, and texture details, to improve IQA accuracy. Despite these advancements, MLLMs still struggled with weather-related degradations, as existing datasets were not fine-tuned for such conditions. + +Score-based IQA Methods. Score-based IQA methods aim to assess image quality by assigning numerical scores that align with human perception. Traditional methods, + +![](images/a2dd3875f4fc6612db5df45d776d852b336abe1d344ec7d8ca4b859b10bbb4fb.jpg) +Figure 2. Dataset screening and evaluation. Through a multi-round dialogue mechanism, the multimodal large language model can replace humans to complete the quality screening and content evaluation of input images. + +such as SSIM [46] and PSNR, relied on statistical metrics to evaluate distortions or pixel-level differences. While effective in controlled environments, they often failed to capture the complexities of real-world images. Deep learning models, such as NIMA [41] and HyperIQA [39], improved performance by regressing directly to Mean Opinion Scores (MOS) derived from human ratings. However, these models faced challenges in out-of-distribution (OOD) generalization, as unseen distortions or diverse content types significantly impacted their accuracy. Recent advancements leveraged vision-language models like CLIP [35], with methods such as CLIP-IQA+ [45] and LIQE [60] utilizing visual-text alignment for few-shot and multitask learning. These approaches demonstrated enhanced robustness across datasets but did not fully capture human-perceived quality due to their reliance on visual-text similarity scores. To address these limitations, Large Multi-Modality Models (LMMs) like Q-Align [47] simulated human rating practices by using discrete text-defined rating levels (e.g., "excellent", "poor") and converted qualitative ratings into quantitative scores through probability-based averaging. This approach enhanced both accuracy and interpretability, particularly in OOD scenarios. + +# 3. Dataset Construction + +High-quality real data is the foundation of our approach. We propose RealSnow10K, the largest real snow dataset available. In this section, we describe the process of building the dataset, including image collection, image evaluation and classification, and the dataset preference process. + +# 3.1. Image Collection, Evaluation and Classification + +The first step in building a dataset is to collect images from various sources. We collect more than 100,000 snow scene images from the Internet and screen out over 10,000 high-quality images. To ensure the high quality of the dataset, + +we need to screen and classify the collected images. Images collected from the Internet often have multiple interferences. In addition to the low quality caused by blur and noise in the image itself, it may also be irrelevant or deviate from the keywords we set, which is often difficult to reflect through existing indicators. To address the above problems, traditional dataset production uses manual screening. Although this method is accurate, it is too inefficient when facing a large number of images. To conduct fast and efficient review, we introduce multimodal large language models (MLLMs) with human visual perception capabilities as a reference for evaluating image quality and content. + +We design a multi-round dialogue mechanism to screen out unqualified images and evaluate the content of the images. Specifically, the prompt words used in the first and second rounds of dialogue are set to "Is this picture a real photograph? Please answer with only 'yes' or 'no'." and "Does this image have a large and obvious watermark? Please answer with only 'yes' or 'no'"; respectively, to screen out synthetic images and low-quality images with large watermarks. The third and fourth rounds of dialogue are used to evaluate the content of the pictures. In the third round of dialogue, we use multiple-choice questions to let the model choose the accurate description of the number of snowflakes from five options. We obtain the logits of each option, calculate the option probability and score, and then distinguish the amount of snowflakes and snow accumulation in different pictures. For the specific dialogue process, please refer to Fig. 2. The fourth round of dialogue is the final classification of the topic of the image, and eight categories are set for MLLM to perform classification. + +In the specific implementation, to balance evaluation quality and efficiency, we select the MiniCPM-V-2.6 model as the dataset evaluation model. Despite having only 8B parameters, this model demonstrates outstanding performance and provides a well-distributed scoring range. We parti + +![](images/c15c69f020dcd5ca55b8fa16974709c555b8f0074f1e014951e6f6101462b847.jpg) +Figure 3. Performance comparison of different MLLMs for different indicators. The horizontal axis in the figure represents the rank-1 proportion of different MLLMs on different snowfall evaluation indicators. + +tion the RealSnow10K dataset into three subsets: 1,500 images for MMPO fine-tuning, 1,047 images for testing, and 6,406 images for semi-supervised training. Detailed category statistics are provided in the supplementary materials. + +# 3.2. Preference Dataset Construction Pipeline + +To perform multi-model preference optimization (MMPO) on the multimodal large language model, We first use the pre-trained snow removal model to perform preliminary snow removal on 1,500 images and expand the dataset to 3,000 images. We select two indicators with poor performance of Q-instruct from the four indicators for evaluating snowfall: snowflake volume and snowflake occlusion, and design targeted questions for each. Subsequently, we collect the outputs of four MLLMs for these two problems, including Q-instruct-13b, LLava-1.5-13b, LLava-1.6-13b, and MiniCPM-V-2.6. Image evaluation experts rank the outputs of the four models one by one and generate six preference pairs from a ranking. Therefore, for each preference problem, we obtain a preference dataset with a sample size of 18,000. + +# 4. Method + +# 4.1. Evaluation of Snowy Images + +Since there is no reference for real snowfall images, the first task for real scene snow removal tasks is to select an accurate and reliable snowfall image evaluation standard. The size, distribution and shape of snowfall particles in the image are often affected by multiple factors including shooting equipment, wind speed, distance, etc., which also brings challenges to the evaluation of snowfall image quality. Traditional non-reference [39, 41] obviously cannot distinguish different forms of snowfall, nor can they accurately describe the interference of snowfall on the image. + +Recently, the development of multimodal large language models has provided a possible solution for snowfall image evaluation. Multimodal large language models are trained on a large number of images and have close to human vi + +sual capabilities. Some works [45, 47, 60] introduced CLIP or multimodal large models into image quality assessment tasks (IQA) and obtained good results. Inspired by these works, we try to introduce multimodal large language models into the field of image snow removal. However, there is currently no work exploring the performance of multimodal large language models in evaluating the degradation of snowfall images. Some works [52] use multimodal large language models from a single dimension to evaluate the quality of severe weather images, which is one-sided and unreliable. + +In this work, we first establish a detailed framework for evaluating snowing images. By analyzing a large dataset of real snow images, we propose assessing snow image quality across four dimensions: snowflake amount, object obscURITY, texture detail, and background clarity. The evaluation framework appears in Fig.4 (i). We select three widely used multimodal large language models and one model, Q-Instruct [47], fine-tuned for low-level visual features. These models analyze real snow images along the four dimensions, and experts evaluate their outputs. A sample output is presented in Fig.4 (ii). + +Unfortunately, no model can fully dominate in all four dimensions: based on LLava1.5-13b, Q-instruct, which has been fine-tuned with underlying visual features, has shown obvious advantages in texture detail and background clarity, which is consistent with our expectations. However, in terms of the two indicators of snowflake amount and object obscure for snowflake description, the best performing models are LLava1.6-13b and MiniCPM-V-2.6, which firmly occupy the first and second positions, while q-instruct-13b is relatively backward. The specific rank-1 ratios of different models are shown in Fig.3. + +# 4.2. Improving Snowfall Perception via Multi-Model Feedback Optimization + +For complex image desnowing task, evaluating performance from multiple perspectives is a prerequisite for achieving satisfactory results. However, as outlined in the previous section, a single multimodal large language model cannot deliver accurate evaluations across all assessment dimensions. At the same time, employing multiple models concurrently for evaluation is computationally infeasible due to excessive resource demands. This leads us to a key question: is it possible to use a single model to simultaneously address the evaluation of multiple perspectives in desnowing tasks? + +Fortunately, Q-Instruct [47] model has already undergone targeted fine-tuning on a large-scale dataset, equipping it with the ability to assess fundamental image quality metrics, such as clarity and contrast. Building on this foundation, and inspired by recent advancements in reinforcement learning [18, 36-38], we propose enhancing Q-Instruct through human feedback optimization. Our goal is + +![](images/77b5777d4af1c3c3e240197cd66f8f0b765c80b1745e242725bf097372a21a6e.jpg) +Figure 4. Schematic diagram of the whole process of snowfall image evaluation optimization. The evaluation framework first divides snowfall into four indicators for evaluation, and for each indicator, collects the evaluation answers of multiple multimodal large language models to the image as the data source for preference learning. Finally, direct preference optimization is applied to Q-instruct to improve its output for this indicator. The figure takes the snowflake amount indicator as an example. + +to enable Q-Instruct to accurately evaluate snowfall, while retaining its capability to assess essential image quality + +Specifically, to address the challenge that different models exhibit distinct strengths and weaknesses across various perspectives, we introduce a novel multi-model preference optimization (MMPO) technique. Unlike conventional approaches, MMPO constructs preference pairs by harnessing the outputs of multiple models, enabling a more comprehensive capture of diverse evaluation criteria. While drawing inspiration from the loss function of Direct Preference Optimization (DPO) [36], MMPO innovatively extends this framework to incorporate the collective strengths of multiple models into a single, cohesive Q-Instruct model. Consequently, the resulting Q-Instruct model excels at thoroughly assessing desnowing performance while preserving its capability to evaluate critical image quality metrics. + +Reward Function Reparameterization. In DPO, The reward associated with a response $y$ given context $x$ is reparameterized in terms of the model's policy $\pi_{\theta}(y|x)$ and a reference policy $\pi_{\mathrm{ref}}(y|x)$ , formulated as: + +$$ +r (x, y) = \beta \log \frac {\pi_ {\theta} (y | x)}{\pi_ {\mathrm {r e f}} (y | x)} + \beta \log Z (x), \tag {1} +$$ + +where $\beta$ is a scaling constant and $Z(x)$ is a partition function to normalize the policy distribution. Through this reparameterization, DPO completes preference optimization without explicitly defining a reward model, simplifying the complex reinforcement learning problem into a classification-based optimization problem. Our proposed MMPO also adopts this reparameterization technique to streamline the formulation of preference optimization. + +Multi-Model Feedback Collection. To construct a robust preference dataset, a multi-model feedback collection strategy is employed. For a given input context $x$ , responses $\{y_1, y_2, \ldots, y_N\}$ are collected from $N$ distinct multimodal large language models. These responses are subsequently evaluated and ranked by human experts, yielding a preference ordering $y_{(1)} \succ y_{(2)} \succ \dots \succ y_{(N)}$ , where $y_{(1)}$ denotes the most preferred response and $y_{(N)}$ the least preferred. From this ordering, preference pairs $(y_w, y_l)$ are generated such that $y_w \succ y_l$ , forming the dataset: + +$$ +\left\{\left(x, y _ {w}, y _ {l}\right) \mid y _ {w} \succ y _ {l}, y _ {w}, y _ {l} \in \left\{y _ {1}, y _ {2}, \dots , y _ {N} \right\} \right\}. \tag {2} +$$ + +This methodology harnesses the diverse capabilities of multiple models to produce a comprehensive set of preference pairs, thereby enhancing the differentiation between superior and inferior responses. + +Multi-Model Preference Optimization. Building upon the framework of Direct Preference Optimization (DPO), Multi-Model Preference Optimization (MMPO) is introduced. The loss function is expressed as: + +$$ +L _ {\mathrm {M M P O}} \left(\pi_ {\theta}; \pi_ {\mathrm {r e f}}\right) = - \mathbb {E} _ {(x, y _ {w}, y _ {l}) \sim D} \tag {3} +$$ + +$$ +\left[ \log \sigma \left(\beta \log \frac {\pi_ {\theta} \left(y _ {w} \mid x\right)}{\pi_ {\mathrm {r e f}} \left(y _ {w} \mid x\right)} - \beta \log \frac {\pi_ {\theta} \left(y _ {l} \mid x\right)}{\pi_ {\mathrm {r e f}} \left(y _ {l} \mid x\right)}\right) \right], +$$ + +where $\sigma (\cdot)$ represents the sigmoid function, and $y_{w}$ and $y_{l}$ denote the preferred and unpreferred responses, respectively. A distinguishing feature of MMPO is its reliance on a preference dataset $D$ derived from the outputs of multiple models, as outlined in the feedback collection process. This + +approach integrates the strengths of various MLLMs, allowing the model to leverage a richer collection of preference pairs. Through this optimization, the model's outputs are aligned with human expert preferences, thereby improving its capacity to generate high-quality responses. + +# 4.3. From Perception to Restoration + +Using MMPO technology, we use Q-instruct as the base model and fine-tune its preference learning in two aspects where its performance was relatively poor: snowflake amount and object obscure, improving its ability to perceive and understand snowfall. However, how can the improvement in perception be used to optimize the image restoration model? + +Inspired by Q-Align [47], we leverage the enhanced perception capabilities of the large model to generate an "image score" that could quantify this improvement. We then use this image score as an evaluation metric within a mean teacher semi-supervised learning framework, combined with a pseudo-label database, to train the image restoration model. + +Score calculation. Using one adjective like Q-Align (e.g. "excellent", "poor") to describe the snowfall scene directly does not provide an accurate evaluation. We analyze that this is because a single adjective lacks a specific definition. Therefore, we further decompose the prompt words, describe the four indicators such as the number of snowflakes from different perspectives, and design it into a multiple choice question with five options. Based on the calculation method of human subjective opinion score (MOS): + +$$ +\mathrm {M O S} = \frac {1}{N} \sum_ {i = 1} ^ {N} S _ {i}, \tag {4} +$$ + +in which $S_{i}$ represents the score of the i-th tester. We collect the model's output logit for the five option tokens, use softmax to calculate the probability of different tokens, and finally get the score through weighted average: + +$$ +\text {S c o r e} = \frac {1}{N} \sum_ {i = 1} ^ {N} p _ {i} \cdot v _ {i}, \tag {5} +$$ + +$$ +p _ {i} = \frac {e ^ {l _ {i}}}{\sum_ {j = 1} ^ {N} e ^ {l _ {j}}}, \tag {6} +$$ + +where $l_{i}$ represents the logit value of the $i$ -th option token. $N$ is 5 to correspond to 5 options. $v_{i}$ is converted into weights from 1 to 5 according to the description of snowfall intensity. + +Semi-supervised Training with Pseudo-labeling. Referring to the mean teacher framework and the database mechanism in [36], we leverage scores from four evaluation metrics to enhance semi-supervised training. To ensure the pseudo-label database is updated accurately, we adopt a + +![](images/ee17afd6786d89eb557b937995f149fed4ed54accad3dfc06fd0a93f2414eafc.jpg) + +![](images/3210ac6913cd883be5f58b4c6a536b898129f8a549a567a5eb59537ba0dafebd.jpg) + +How would you evaluate the amount of snowflakes in the air shown in the picture? Please provide your judgment and basis. Summarize in about 50 words. + +![](images/89b66cd64fcc15cbb955c9750b2e9a1d737825a1114060a9b2b46d5e586b6ea5.jpg) + +Keywords that do not match the real picture Keywords that lack logic or basis + +![](images/abb41cec1c16737bc1aa024a727e864851e557548fa680df04ab50597995486a.jpg) + +Keywords that match the picture and perception of human experts + +![](images/375ddf0bac9b4e2b9e9d417421990387a33418d63690aca5f8f6d053e0cc47e4.jpg) + +# Q-instruct + +The amount of snowflakes in the air shown in the picture is quite a lot. This is evident from the fact that the people in the image are wearing winter coats and hats, and their clothing is covered in snow. The snowflakes are also clearly visible in the air, indicating that the snowfall is quite heavy. The people are walking in the snow, which further emphasizes the significant amount of snowfall. Overall, the image portrays a snowy day with a considerable amount of snowflakes in the air. + +![](images/7badbd3e3b965cc93128f598bc481dd9a0c321059cbf0fb0edeabd768e2545e2.jpg) + +# O-instruct + MMPO + +The image shows a group of people walking in the snow, with several snowflakes visible in the air. While it is not possible to quantify the exact number of snowflakes, it is evident that there is a noticeable amount of snowfall. The snowflakes are dispersed throughout the scene, indicating that the snowfall is not too heavy or too light, but rather moderate. This suggests that the weather conditions are likely to be cold and snowy, but not extreme or overwhelming. + +Figure 5. Comparison of perception ability before and after MMPO. The fine-tuned model output is more in line with human visual perception, more logical, and less hallucinatory. + +stringent rule: an update occurs only if $(i)$ at least three teacher scores exceed the student score, and $(ii)$ the teacher score surpasses the current score stored in the pseudo-label database. This dual-condition approach effectively minimizes erroneous updates to the pseudo-labels, improving the reliability of the training process. + +Loss function and training strategy. We use both supervised loss and pseudo-label loss to train the model. The supervised loss is based on the PSNR between the model's output and the reference images from the Snow100K [30] dataset. And the pseudo-label loss relies on the L1 norm to measure the difference between the output and pseudolabels. To ensure fast and stable training, we only use supervised loss in the initial 20k steps, and add pseudo-label loss after 20k steps for semi-supervised training. The total loss function is shown as follows: + +$$ +\mathcal {L} _ {\text {t o t a l}} = \mathcal {L} _ {\text {s u p e r v i s e d}} + \lambda \mathcal {L} _ {\text {p s e u d o}} (\lambda = 0 \text {i f s t e p < 2 0 k}). \tag {7} +$$ + +# 5. Experiments + +# 5.1. Experiment Setup + +Training Settings. Our method consists of two phases: the first phase is to fine-tune the scoring model (i.e. Q-Instruct [50]) via proposed MMPO, and the second phase is to semi-supervised train the snow removal model. In the first phase, we set epoch = 1, lr = 1e-6, batchsize = 64; We use Vision-LLM Alignment framework [44] to perform MMPO fine-tuning and enable deepspeed for acceleration, set zero-stage to 3 and enable offload for full fine-tuning. In the second phase of snow removal model training, we use NAFNet [5] as the base model, set lr = 2e-4, batchsize = 32. All experiments are performed on two Nvidia A800 GPUs. + +![](images/64e5340073e481c4490fd8bcf2efbe32b44ed03300b555a63c0bf19b5425abb6.jpg) +Figure 6. Visual comparison of different models for snow removal on real images. Please zoom in for a better view. SnowMaster achieves satisfactory snow removal performance in complicated conditions compared to previous methods. + +Table 1. The accuracy of MLLMs on different perspectives. + +
MLLMs AccuracySnowflakes accQuality acc
amount ↑obscure ↑texture ↑background ↑
LLava-v1.5-13b[29]80.0477.1755.7852.15
LLava-v1.6-13b[29]86.3479.5652.0554.06
MiniCPM-V[53]75.4570.0442.0249.76
Q-intract-13b[50]76.3173.2658.3663.42
Q-instruct with MMPO (ours)82.1480.9064.9568.58
MMPO improvement+5.83+7.64+6.59+5.16
+ +Experimental Details. In the semi-supervised training process, considering the need to make the multimodal large language model output the correct score, we adopt a pseudo-label database initialization strategy using the original image instead of all-0 initialization in [21]. In the test phase, we test 457 test images selected from the real snow scene dataset (RealSnow10K) and 1329 real snow scene images from Snow100K [30]. + +# 5.2. Evaluation Metrics and Baselines + +Following [19, 20], we utilize four non-reference IQA metrics: BRISQUE [31], NIQE [32], PIQE [43], and PAQ2-PIQ [57]. BRISQUE measures quality via spa + +tial natural scene statistics, NIQE assesses deviations from these statistics, PIQE detects block-level distortions, and PAQ2-PIQ, a deep learning approach, predicts global and local quality. These metrics ensure a comprehensive, unsupervised evaluation of diverse distortions. For the baselines, we choose CNN-based models (SMGARN[17], NAFNet[5]), Transformer-based models (Transweather[42], Snowformer[6]) and a diffusion based model $(\mathrm{T}^3$ -DiffWeather[11]), offering varied references for snowfall image restoration performance. + +# 5.3. Results and Analysis + +# 5.3.1. Quantitative Results + +Effectiveness of MMPO. Table 1 illustrates the performance enhancements of the Q-Instruct [47] model following fine-tuning with MMPO. We utilize NAFNet [5], pretrained for 100,000 steps, as the snow removal model to process the images in the test set, generating a dataset that includes both original and snow-removed images. By inputting these images individually into Q-Instruct, we obtain scores for each image corresponding to specific indicators. Consequently, for each indicator, accuracy is defined + +Table 2. Comparison of non-reference metrics after snow removal between different models. + +
MethodSnow100K-realistic [30]RealSnow10K
BRISQUE ↓NIQE ↓PIQE ↓PAQ2-PIQ ↑BRISQUE ↓NIQE ↓PIQE ↓PAQ2-PIQ ↑
SMGARN[TCSVT'22][17]19.2893.13231.92769.76723.0053.58031.21068.919
TransWeather[CVPR'22] [42]16.5443.02530.30269.82118.5643.40429.77269.671
NAFNet[ECCV'22] [5]17.5662.90330.84369.95120.4253.29129.87769.388
SnowFormer[Arxiv'23][6]19.3892.95431.90270.01523.7023.42931.9969.929
T3-DiffWeather[ECCV'24] [11]17.7442.86831.68970.04220.2113.13732.99170.211
★ SnowMaster (Ours)16.3052.79730.16270.04217.5463.05829.74270.304
+ +as the proportion of snow-removed images whose score is not lower than that of their corresponding original images. The results show that the performance after fine-tuning has been significantly improved which is consistent with our expectations. It is worth noting that in terms of the number of snowflakes, the accuracy after fine-tuning does not exceed that of Llava-1.6. We think this is normal, because the preference data in fine-tuning are all output by these models, and MMPO guides the model to move closer to human preference output, but it may not necessarily surpass the output of the optimal model. + +Furthermore, a noteworthy observation emerges: the indicators 'texture' and 'background', which are not explicitly targeted for optimization, also exhibit moderate improvements. We hypothesize that this phenomenon arises from a comprehensive augmentation of the model's perceptual capacity regarding snowfall, facilitated by the MMPO fine-tuning process. + +Performance analysis of the restoration model. Table 2 presents the performance comparison between our method and other advanced snow removal models. Our proposed method achieves the best results in most indicators. + +# 5.3.2. Qualitative Assessment + +From the restoration results of the real snowfall image in Fig.6, we can see that the SnowMaster model we proposed has the best performance in multiple scenarios. When facing snowflakes illuminated by lights at night, other models cannot effectively complete the restoration of high-brightness snowflakes, while SnowMaster significantly eliminates the high-brightness snowflakes in the image. When restoring dense snowflake scenes during the day, SnowMaster can effectively remove snowflakes of different scales and retain the original color and texture of the object, while other models have poor restoration effects on small-sized snowflake particles. + +# 5.4. Ablation Study + +Necessity of MMPO. The Q-Instruct model shows significant improvement after multi-modal preference optimization, particularly in evaluating snowfall images. We perform an ablation test to compare the Q-Instruct model before and after alignment when used as a scoring model in semi-supervised training. The results indicate that the + +Table 3. Ablation on different settings of the proposed framework. + +
MethodBRISQUE ↓NIQE ↓
w/o. Semi-supervised learning22.523.33
w/o. MLLM20.363.25
w/o. MMPO18.983.19
w. Single-angle evaluation18.213.12
w. Multi-angle evalutaion (Ours)17.553.06
+ +aligned model achieves higher accuracy and ensures more stable training to improve the performance. + +Effectiveness of multi-perspectives snowfall evaluation. Our method enhances image screening reliability by evaluating snowfall images from four perspectives, ensuring steady updates to the pseudo-label database. Ablation experiments, conducted by reducing evaluation angles, demonstrate that multi-angle evaluation improves training effectiveness + +The role of the semi-supervised learning framework. In order to transfer the improvement of the evaluation performance of the large model on snow scene images to the improvement of the restoration model, we use the mean teacher semi-supervised learning framework. By combining the supervision loss from snow100k and the pseudolabel loss from the real training data, this framework can effectively optimize the restoration effect of real snowfall images. Through ablation experiments, we prove that the semi-supervised framework is the basis of our method. + +The specific metrics of the ablation experiments are shown in Table 3. + +# 6. Conclusion + +SnowMaster integrates a multimodal large language model with a multi-perspective evaluation framework for image desnowing. Leveraging the RealSnow10k dataset, it utilizes a multi-angle evaluation system assessing four key metrics to enhance real-world reliability. Through multimodal preference optimization and semi-supervised training with pseudo-label updates from multi-angle multiple-choice questions, it ensures effective refinement. Experiments confirm its robustness in complex snowfall scenarios, though its computational intensity from multiple evaluations signals a need for future enhancements in training efficiency and resource use. + +# Acknowledgments + +This work is supported by the Guangzhou-HKUST(GZ) Joint Funding Program (No. 2023A03J0671), the Guangzhou Municipal Science and Technology Project (Grant No. 2024A04J4230), and Guangzhou-HKUST(GZ) Joint Funding Program (No. 2024A03J0618). + +# References + +[1] Vijay Badrinarayanan, Alex Kendall, and Roberto Cipolla. Segnet: A deep convolutional encoder-decoder architecture for image segmentation. IEEE transactions on pattern analysis and machine intelligence, 39(12):2481-2495, 2017. 1 +[2] Chun-Fu Richard Chen, Quanfu Fan, and Rameswar Panda. Crossvit: Cross-attention multi-scale vision transformer for image classification. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 357-366, 2021. 1 +[3] Haoyu Chen, Jinjin Gu, Yihao Liu, Salma Abdel Magid, Chao Dong, Qiong Wang, Hanspeter Pfister, and Lei Zhu. Masked image training for generalizable deep image denoising. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1692-1703, 2023. 2 +[4] Haoyu Chen, Wenbo Li, Jinjin Gu, Jingjing Ren, Haoze Sun, Xueyi Zou, Zhensong Zhang, Youliang Yan, and Lei Zhu. Low-res leads the way: Improving generalization for superresolution by self-supervised learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25857-25867, 2024. 1 +[5] Liangyu Chen, Xiaojie Chu, Xiangyu Zhang, and Jian Sun. Simple baselines for image restoration. In European conference on computer vision, pages 17-33. Springer, 2022. 6, 7, 8 +[6] Sixiang Chen, Tian Ye, Yun Liu, and Erkang Chen. Snowformer: Context interaction transformer with scale-awareness for single image desnowing. arXiv preprint arXiv:2208.09703, 2022. 2, 7, 8 +[7] Sixiang Chen, Tian Ye, Yun Liu, Taodong Liao, Yi Ye, and Erkang Chen. Msp-former: Multi-scale projection transformer for single image desnowing. arXiv preprint arXiv:2207.05621, 2022. 1, 2 +[8] Sixiang Chen, Tian Ye, Jinbin Bai, Erkang Chen, Jun Shi, and Lei Zhu. Sparse sampling transformer with uncertainty-driven ranking for unified removal of raindrops and rain streaks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13106-13117, 2023. 2 +[9] Sixiang Chen, Tian Ye, Yun Liu, Jinbin Bai, Haoyu Chen, Yunlong Lin, Jun Shi, and Erkang Chen. CPLformer: Cross-scale prototype learning transformer for image snow removal. In Proceedings of the 31st ACM international conference on multimedia, pages 4228-4239, 2023. 1 +[10] Sixiang Chen, Tian Ye, Chenghao Xue, Haoyu Chen, Yun Liu, Erkang Chen, and Lei Zhu. Uncertainty-driven dynamic + +degradation perceiving and background modeling for efficient single image desnowing. In Proceedings of the 31st ACM International Conference on Multimedia, pages 4269-4280, 2023. 2 +[11] Sixiang Chen, Tian Ye, Kai Zhang, Zhaohu Xing, Yunlong Lin, and Lei Zhu. Teaching tailored to talent: Adverse weather restoration via prompt pool and depth-anything constraint. In European Conference on Computer Vision, pages 95–115. Springer, 2025. 2, 7, 8 +[12] Wei-Ting Chen, Hao-Yu Fang, Jian-Jiun Ding, Cheng-Che Tsai, and Sy-Yen Kuo. Jstasr: Joint size and transparency-aware snow removal algorithm based on modified partial convolution and veiling effect removal. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXI 16, pages 754-770. Springer, 2020. 1, 2 +[13] Wei-Ting Chen, Hao-Yu Fang, Cheng-Lin Hsieh, Cheng-Che Tsai, I Chen, Jian-Jiun Ding, Sy-Yen Kuo, et al. All snow removed: Single image desnowing algorithm using hierarchical dual-tree complex wavelet representation and contradict channel loss. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4196-4205, 2021. 2 +[14] Wei-Ting Chen, Hao-Yu Fang, Cheng-Lin Hsieh, Cheng-Che Tsai, I Chen, Jian-Jiun Ding, Sy-Yen Kuo, et al. All snow removed: Single image desnowing algorithm using hierarchical dual-tree complex wavelet representation and contradict channel loss. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4196-4205, 2021. 1, 2 +[15] Wei-Ting Chen, Zhi-Kai Huang, Cheng-Che Tsai, Hao-Hsiang Yang, Jian-Jiun Ding, and Sy-Yen Kuo. Learning multiple adverse weather removal via two-stage knowledge learning and multi-contrastive regularization: Toward a unified model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17653–17662, 2022. 2 +[16] Bodong Cheng, Juncheng Li, Ying Chen, Shuyi Zhang, and Tieyong Zeng. Snow mask guided adaptive residual network for image snow removal. arXiv preprint arXiv:2207.04754, 2022. 1, 2 +[17] Bodong Cheng, Juncheng Li, Ying Chen, and Tieyong Zeng. Snow mask guided adaptive residual network for image snow removal. Computer Vision and Image Understanding, 236: 103819, 2023. 2, 7, 8 +[18] Paul F Christiano, Jan Leike, Tom Brown, Miljan Martic, Shane Legg, and Dario Amodei. Deep reinforcement learning from human preferences. Advances in neural information processing systems, 30, 2017. 4 +[19] Jinjin Gu, Haoming Cai, Haoyu Chen, Xiaoxing Ye, Jimmy Ren, and Chao Dong. Image quality assessment for perceptual image restoration: A new dataset, benchmark and metric. arXiv preprint arXiv:2011.15002, 2020. 7 +[20] Jinjin Gu, Haoming Cai, Haoyu Chen, Xiaoxing Ye, Jimmy Ren, and Chao Dong. Pipal: a large-scale image quality assessment dataset for perceptual image restoration. In European Conference on Computer Vision, pages 633-651. Springer, 2020. 7 + +[21] Shirui Huang, Keyan Wang, Huan Liu, Jun Chen, and Yun-song Li. Contrastive semi-supervised learning for underwater image restoration via reliable bank. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18145-18155, 2023. 7 +[22] Yeying Jin, Wending Yan, Wenhan Yang, and Robby T Tan. Structure representation network and uncertainty feedback learning for dense non-uniform fog removal. In Asian Conference on Computer Vision, pages 155-172. Springer, 2022. 2 +[23] Yeying Jin, Wenhan Yang, Wei Ye, Yuan Yuan, and Robby T Tan. Shadowdiffusion: Diffusion-based shadow removal using classifier-driven attention and structure preservation. arXiv preprint arXiv:2211.08089, 2022. +[24] Yeying Jin, Beibei Lin, Wending Yan, Wei Ye, Yuan Yuan, and Robby T. Tan. Enhancing visibility in nighttime haze images using guided apsf and gradient adaptive convolution, 2023. 2 +[25] Ruoteng Li, Robby T Tan, and Loong-Fah Cheong. All in one bad weather removal using architectural search. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3175-3185, 2020. 2 +[26] Yunlong Lin, Zhenqi Fu, Ge Meng, Yingying Wang, Yuhang Dong, Linyu Fan, Hedeng Yu, and Xinghao Ding. Domain-irrelevant feature learning for generalizable pan-sharpening. In Proceedings of the 31st ACM International Conference on Multimedia, pages 3287–3296, 2023. 2 +[27] Yunlong Lin, Zhenqi Fu, Kairun Wen, Tian Ye, Sixiang Chen, Ge Meng, Yingying Wang, Yue Huang, Xiaotong Tu, and Xinghao Ding. Unsupervised low-light image enhancement with lookup tables and diffusion priors. arXiv preprint arXiv:2409.18899, 2024. +[28] Yunlong Lin, Tian Ye, Sixiang Chen, Zhenqi Fu, Yingying Wang, Wenhao Chai, Zhaohu Xing, Lei Zhu, and Xinghao Ding. Aglldiff: Guiding diffusion models towards unsupervised training-free real-world low-light image enhancement. arXiv preprint arXiv:2407.14900, 2024. 2 +[29] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26296-26306, 2024. 2, 7 +[30] Yun-Fu Liu, Da-Wei Jaw, Shih-Chia Huang, and Jenq-Neng Hwang. Desnownet: Context-aware deep network for snow removal. IEEE Transactions on Image Processing, 27(6): 3064–3073, 2018. 2, 6, 7, 8 +[31] Anish Mittal, Anush Krishna Moorthy, and Alan Conrad Bovik. No-reference image quality assessment in the spatial domain. IEEE Transactions on image processing, 21(12): 4695-4708, 2012. 1, 7 +[32] Anish Mittal, Rajiv Soundararajan, and Alan C Bovik. Making a “completely blind” image quality analyzer. IEEE Signal processing letters, 20(3):209-212, 2012. 1, 7 +[33] Ozan Özdenizci and Robert Legenstein. Restoring vision in adverse weather conditions with patch-based denoising diffusion models. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 2 + +[34] Soo-Chang Pei, Yu-Tai Tsai, and Chen-Yu Lee. Removing rain and snow in a single image using saturation and visibility features. In 2014 IEEE International Conference on Multimedia and Expo Workshops (ICMEW), pages 1-6. IEEE, 2014. 2 +[35] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 3 +[36] Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. Direct preference optimization: Your language model is secretly a reward model. Advances in Neural Information Processing Systems, 36, 2024. 2, 4, 5, 6 +[37] John Schulman, Sergey Levine, Pieter Abbeel, Michael Jordan, and Philipp Moritz. Trust region policy optimization. In International conference on machine learning, pages 1889-1897. PMLR, 2015. +[38] John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347, 2017. 4 +[39] Shaolin Su, Qingsen Yan, Yu Zhu, Cheng Zhang, Xin Ge, Jinqiu Sun, and Yanning Zhang. Blindly assess image quality in the wild guided by a self-adaptive hyper network. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3667-3676, 2020. 3, 4 +[40] Haoze Sun, Wenbo Li, Jianzhuang Liu, Haoyu Chen, Renjing Pei, Xueyi Zou, Youliang Yan, and Yujiu Yang. Coser: Bridging image and language for cognitive super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25868-25878, 2024. 2 +[41] Hossein Talebi and Peyman Milanfar. Nima: Neural image assessment. IEEE transactions on image processing, 27(8): 3998-4011, 2018. 3, 4 +[42] Jeya Maria Jose Valanarasu, Rajeev Yasarla, and Vishal M Patel. Transweather: Transformer-based restoration of images degraded by adverse weather conditions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2353-2363, 2022. 2, 7, 8 +[43] Narasimhan Venkatanath, D Praneeth, Maruthi Chandrasekhar Bh, Sumohana S Channappayya, and Swarup S Medasani. Blind image quality evaluation using perception based features. In 2015 twenty first national conference on communications (NCC), pages 1-6. IEEE, 2015. 1, 7 +[44] Chenglong Wang, Yang Gan, Yifu Huo, Yongyu Mu, Murun Yang, Qiaozhi He, Tong Xiao, Chunliang Zhang, Tongran Liu, Quan Du, Di Yang, and Jingbo Zhu. Rovrm: A robust visual reward model optimized via auxiliary textual preference data, 2024. 6 +[45] Jianyi Wang, Kelvin CK Chan, and Chen Change Loy. Exploring clip for assessing the look and feel of images. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2555-2563, 2023. 3, 4 + +[46] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 3 +[47] Haoning Wu, Zicheng Zhang, Weixia Zhang, Chaofeng Chen, Liang Liao, Chunyi Li, Yixuan Gao, Annan Wang, Erli Zhang, Wenxiu Sun, et al. Q-align: Teaching Imms for visual scoring via discrete text-defined levels. arXiv preprint arXiv:2312.17090, 2023. 1, 3, 4, 6, 7 +[48] Hongtao Wu, Yijun Yang, Angelica I Aviles-Rivero, Jingjing Ren, Sixiang Chen, Haoyu Chen, and Lei Zhu. Semi-supervised video desnowing network via temporal decoupling experts and distribution-driven contrastive regularization. In European Conference on Computer Vision, pages 70-89. Springer, 2024. 1 +[49] Hongtao Wu, Yijun Yang, Huihui Xu, Weiming Wang, Jinni Zhou, and Lei Zhu. Rainmamba: Enhanced locality learning with state space models for video deraining. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 7881-7890, 2024. 2 +[50] Haoning Wu, Zicheng Zhang, Erli Zhang, Chaofeng Chen, Liang Liao, Annan Wang, Kaixin Xu, Chunyi Li, Jingwen Hou, Guangtao Zhai, et al. Q-instruct: Improving low-level visual abilities for multi-modality foundation models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25490-25500, 2024. 1, 2, 6, 7 +[51] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. Advances in Neural Information Processing Systems, 34, 2021. 1 +[52] Jiaqi Xu, Mengyang Wu, Xiaowei Hu, Chi-Wing Fu, Qi Dou, and Pheng-Ann Heng. Towards real-world adverse weather image restoration: Enhancing clearness and semantics with vision-language models. In European Conference on Computer Vision, pages 147–164. Springer, 2025. 4 +[53] Yuan Yao, Tianyu Yu, Ao Zhang, Chongyi Wang, Junbo Cui, Hongji Zhu, Tianchi Cai, Haoyu Li, Weilin Zhao, Zhihui He, et al. Minicpm-v: A gpt-4v level mllm on your phone. arXiv preprint arXiv:2408.01800, 2024. 2, 7 +[54] Tian Ye, Yunchen Zhang, Mingchao Jiang, Liang Chen, Yun Liu, Sixiang Chen, and Erkang Chen. Perceiving and modeling density for image dehazing. In European Conference on Computer Vision, pages 130-145. Springer, 2022. 2 +[55] Tian Ye, Sixiang Chen, Jinbin Bai, Jun Shi, Chenghao Xue, Jingxia Jiang, Junjie Yin, Erkang Chen, and Yun Liu. Adverse weather removal with codebook priors. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12653-12664, 2023. 2 +[56] Tian Ye, Sixiang Chen, Wenhao Chai, Zhaohu Xing, Jing Qin, Ge Lin, and Lei Zhu. Learning diffusion texture priors for image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2524-2534, 2024. 2 +[57] Zhenqiang Ying, Haoran Niu, Praful Gupta, Dhruv Mahajan, Deepti Ghadiyaram, and Alan Bovik. From patches to + +pictures (paq-2-piq): Mapping the perceptual space of picture quality. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3575-3585, 2020. 1, 7 +[58] Zhiyuan You, Zheyuan Li, Jinjin Gu, Zhenfei Yin, Tianfan Xue, and Chao Dong. Depicting beyond scores: Advancing image quality assessment through multi-modal language models. arXiv preprint arXiv:2312.08962, 2023. 2 +[59] Kaihao Zhang, Rongqing Li, Yanjiang Yu, Wenhan Luo, and Changsheng Li. Deep dense multi-scale network for snow removal using semantic and depth priors. IEEE Transactions on Image Processing, 30:7419-7431, 2021. 2 +[60] Weixia Zhang, Guangtao Zhai, Ying Wei, Xiaokang Yang, and Kede Ma. Blind image quality assessment via vision-language correspondence: A multitask learning perspective. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14071-14081, 2023. 3, 4 +[61] Haipeng Zhou, Hongqiu Wang, Tian Ye, Zhaohu Xing, Jun Ma, Ping Li, Qiong Wang, and Lei Zhu. Timeline and boundary guided diffusion network for video shadow detection. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 166-175, 2024. 1 \ No newline at end of file diff --git a/2025/SnowMaster_ Comprehensive Real-world Image Desnowing via MLLM with Multi-Model Feedback Optimization/images.zip b/2025/SnowMaster_ Comprehensive Real-world Image Desnowing via MLLM with Multi-Model Feedback Optimization/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..91e9737524bb0e1b2d66b22102e321fc7c0e7cdc --- /dev/null +++ b/2025/SnowMaster_ Comprehensive Real-world Image Desnowing via MLLM with Multi-Model Feedback Optimization/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7fed9d760c36005c52e90d9393635bfc2301ea8a1b0c3d71acb2c396270240fa +size 732450 diff --git a/2025/SnowMaster_ Comprehensive Real-world Image Desnowing via MLLM with Multi-Model Feedback Optimization/layout.json b/2025/SnowMaster_ Comprehensive Real-world Image Desnowing via MLLM with Multi-Model Feedback Optimization/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..66f5e198fda8f4dae46ce663542a1ee922308e67 --- /dev/null +++ b/2025/SnowMaster_ Comprehensive Real-world Image Desnowing via MLLM with Multi-Model Feedback Optimization/layout.json @@ -0,0 +1,7834 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 110, + 103, + 501, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 103, + 501, + 140 + ], + "spans": [ + { + "bbox": [ + 110, + 103, + 501, + 140 + ], + "type": "text", + "content": "SnowMaster: Comprehensive Real-world Image Desnowing via MLLM with Multi-Model Feedback Optimization" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 88, + 160, + 525, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 160, + 525, + 248 + ], + "spans": [ + { + "bbox": [ + 88, + 160, + 525, + 248 + ], + "type": "text", + "content": "Jianyu Lai" + }, + { + "bbox": [ + 88, + 160, + 525, + 248 + ], + "type": "inline_equation", + "content": "^{1,*}" + }, + { + "bbox": [ + 88, + 160, + 525, + 248 + ], + "type": "text", + "content": " Sixiang Chen" + }, + { + "bbox": [ + 88, + 160, + 525, + 248 + ], + "type": "inline_equation", + "content": "^{1,*}" + }, + { + "bbox": [ + 88, + 160, + 525, + 248 + ], + "type": "text", + "content": " Yunlong Lin" + }, + { + "bbox": [ + 88, + 160, + 525, + 248 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 88, + 160, + 525, + 248 + ], + "type": "text", + "content": " Tian Ye" + }, + { + "bbox": [ + 88, + 160, + 525, + 248 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 88, + 160, + 525, + 248 + ], + "type": "text", + "content": " \nYun Liu" + }, + { + "bbox": [ + 88, + 160, + 525, + 248 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 88, + 160, + 525, + 248 + ], + "type": "text", + "content": " Song Fei" + }, + { + "bbox": [ + 88, + 160, + 525, + 248 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 88, + 160, + 525, + 248 + ], + "type": "text", + "content": " Zhaohu Xing" + }, + { + "bbox": [ + 88, + 160, + 525, + 248 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 88, + 160, + 525, + 248 + ], + "type": "text", + "content": " Hongtao Wu" + }, + { + "bbox": [ + 88, + 160, + 525, + 248 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 88, + 160, + 525, + 248 + ], + "type": "text", + "content": " Weiming Wang" + }, + { + "bbox": [ + 88, + 160, + 525, + 248 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 88, + 160, + 525, + 248 + ], + "type": "text", + "content": " Lei Zhu" + }, + { + "bbox": [ + 88, + 160, + 525, + 248 + ], + "type": "inline_equation", + "content": "^{1,5,\\dagger}" + }, + { + "bbox": [ + 88, + 160, + 525, + 248 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 88, + 160, + 525, + 248 + ], + "type": "text", + "content": " The Hong Kong University of Science and Technology (GZ) " + }, + { + "bbox": [ + 88, + 160, + 525, + 248 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 88, + 160, + 525, + 248 + ], + "type": "text", + "content": " Xiamen University \n" + }, + { + "bbox": [ + 88, + 160, + 525, + 248 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 88, + 160, + 525, + 248 + ], + "type": "text", + "content": " Southwestern University " + }, + { + "bbox": [ + 88, + 160, + 525, + 248 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 88, + 160, + 525, + 248 + ], + "type": "text", + "content": " Hong Kong Metropolitan University \n" + }, + { + "bbox": [ + 88, + 160, + 525, + 248 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 88, + 160, + 525, + 248 + ], + "type": "text", + "content": " The Hong Kong University of Science and Technology \nProject page: https://alexlai2860.github.io/SnowMaster" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 151, + 275, + 200, + 289 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 275, + 200, + 289 + ], + "spans": [ + { + "bbox": [ + 151, + 275, + 200, + 289 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 301, + 296, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 301, + 296, + 553 + ], + "spans": [ + { + "bbox": [ + 54, + 301, + 296, + 553 + ], + "type": "text", + "content": "Snowfall presents significant challenges for visual data processing, necessitating specialized desnowing algorithms. However, existing models often fail to generalize effectively due to their heavy reliance on synthetic datasets. Furthermore, current real-world snowfall datasets are limited in scale and lack dedicated evaluation metrics designed specifically for snowfall degradation, thus hindering the effective integration of real snowy images into model training to reduce domain gaps. To address these challenges, we first introduce RealSnow10K, a large-scale, high-quality dataset consisting of over 10,000 annotated real-world snowy images. In addition, we curate a preference dataset comprising 36,000 expert-ranked image pairs, enabling the adaptation of multimodal large language models (MLLMs) to better perceive snowy image quality through our innovative Multi-Model Preference Optimization (MMPO). Finally, we propose the SnowMaster, which employs MMPO-enhanced MLLM to perform accurate snowy image evaluation and pseudo-label filtering for semi-supervised training. Experiments demonstrate that SnowMaster delivers superior desnowing performance under real-world conditions." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 565, + 135, + 577 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 565, + 135, + 577 + ], + "spans": [ + { + "bbox": [ + 55, + 565, + 135, + 577 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 586, + 295, + 670 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 586, + 295, + 670 + ], + "spans": [ + { + "bbox": [ + 55, + 586, + 295, + 670 + ], + "type": "text", + "content": "Snowfall, a common form of adverse weather degradation, reduces image visibility and adversely impacts downstream tasks such as image recognition [2, 61] and segmentation [1, 51]. The inherent complexity of snow poses a challenge to conventional image restoration techniques, necessitating the development of specialized snow removal (desnowing) algorithms to enhance image clarity." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 670, + 295, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 670, + 295, + 694 + ], + "spans": [ + { + "bbox": [ + 55, + 670, + 295, + 694 + ], + "type": "text", + "content": "Although current snow removal models achieve impressive performance on benchmark datasets [7, 9, 12, 14, 16]," + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 313, + 274, + 557, + 393 + ], + "blocks": [ + { + "bbox": [ + 313, + 274, + 557, + 393 + ], + "lines": [ + { + "bbox": [ + 313, + 274, + 557, + 393 + ], + "spans": [ + { + "bbox": [ + 313, + 274, + 557, + 393 + ], + "type": "image", + "image_path": "cd2fbfead21ced4a3c83c22752542120b1f6d887472dfca22a76b344a8b7032c.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 394, + 555, + 449 + ], + "lines": [ + { + "bbox": [ + 313, + 394, + 555, + 449 + ], + "spans": [ + { + "bbox": [ + 313, + 394, + 555, + 449 + ], + "type": "text", + "content": "Figure 1. Comparison of various image desnowing evaluation metrics shows that Q-Instruct [47], enhanced by our MMPO, most accurately perceives differences in snow density between images. Leveraging accurate evaluation model, we optimize our semi-supervised network, achieving superior desnowing performance." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 459, + 555, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 459, + 555, + 591 + ], + "spans": [ + { + "bbox": [ + 313, + 459, + 555, + 591 + ], + "type": "text", + "content": "they often struggle to generalize to real-world scenarios. This limitation stems primarily from their only reliance on synthetic snowfall datasets for training. These datasets, generated using simplified physical models, struggle to capture the intricate details of real-world snow degradations, resulting in a domain gap between synthetic and real data. This gap hinders model optimization and impairs generalization to real-world scenarios. Consequently, we pose the question: How can real-world data be effectively utilized to enhance the model's generalization performance on real-world samples?" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "content": "Incorporating real-world snowy images into training is an direct and effective approach to enhance model generalization. However, accurate evaluation of these images is a prerequisite for their successful integration into the training process [4, 48, 50]. Yet, evaluating realistic snowfall images remains a significant challenge due to the absence of ground truth for reference evaluation, previous methods often rely on non-reference metrics [31, 32, 43, 57] to assess image restoration quality. These metrics, influenced by numerous factors, fail to accurately isolate and evalu-" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 702, + 217, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 702, + 217, + 713 + ], + "spans": [ + { + "bbox": [ + 66, + 702, + 217, + 713 + ], + "type": "text", + "content": "*Equal Contribution; † Corresponding Author." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4302" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 251 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 251 + ], + "type": "text", + "content": "ate weather-specific degradation in snowfall images. In this paper, we claim that accurate assessment of snowy images requires a joint effort from both low-level visual and semantic perspectives. Recent advancements in multimodal large language models (MLLMs), such as Q-Instruct [50] and Depict-QA [58], offer promising avenues for developing evaluation frameworks due to their ability to assess low-level visual features from a semantic perspective. However, these models currently lack the specialized perception capabilities required to reliably evaluate snowfall-specific characteristics. Conversely, while some more advanced multimodal large language models exhibit better understanding of snowy scenes, they are not fine-tuned for detailed low-level visual evaluation, limiting the fundamental premise as powerful evaluators on the snow removal task." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 254, + 296, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 254, + 296, + 540 + ], + "spans": [ + { + "bbox": [ + 56, + 254, + 296, + 540 + ], + "type": "text", + "content": "To tackle these challenges, we first establish a comprehensive evaluation framework designed to capture the distinct characteristics of snow in natural images through multi-perspective feedback optimization tailored to snow scenes. To support this framework, we curate and annotate a dataset comprising over 10,000 real-world snow images sourced from the Internet. Additionally, while Direct Preference Optimization (DPO) [36] provides a strong foundation for preference-based learning, we observe that different models exhibit unique strengths across various evaluation perspectives, and assessing snowfall requires a multidimensional approach to ensure robust and reliable outcomes. Therefore, we introduce the Multi-Model Preference Optimization (MMPO) framework, integrating task-specific outputs from multiple multimodal large language models (MLLMs) to construct a snowfall preference dataset. This dataset enables the fine-tuning of Q-Instruct, significantly enhancing its ability to perceive and evaluate snowfall. Finally, we integrate the MMPO fine-tuned Q-Instruct model as an unsupervised evaluation metric to update the pseudolabel database and employ a mean teacher semi-supervised framework to train existing snow removal models, thereby improving their generalization performance in real-world scenarios. Our contributions can be summarized as follows:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 55, + 545, + 295, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 295, + 617 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 295, + 617 + ], + "type": "text", + "content": "- Real-World Snowfall Dataset: We introduce Real-Snow10K, a large-scale, high-quality dataset comprising over 10,000 annotated real-world snow scene images. To the best of our knowledge, this is the largest dataset of real-world image desnowing, providing a robust foundation for model training and evaluation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 618, + 294, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 294, + 689 + ], + "type": "text", + "content": "- Snowfall Preference Dataset: From RealSnow10K, we select 1.5k real images and engage human experts to rank the outputs of multiple models based on preference. Utilizing these rankings, we construct a human preference dataset containing 36k preference pairs, designed to address two key evaluation questions." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 689, + 294, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 294, + 713 + ], + "type": "text", + "content": "- MMPO for Snowfall Evaluation: We employ Multi-Model Preference Optimization (MMPO) to enhance the" + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 322, + 72, + 553, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 72, + 553, + 120 + ], + "spans": [ + { + "bbox": [ + 322, + 72, + 553, + 120 + ], + "type": "text", + "content": "perception capabilities of the MLLM on snowfall images. By leveraging carefully designed multiple-choice questions, this MLLM can be used to accurately evaluate the snow removal effect of images." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 121, + 553, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 121, + 553, + 192 + ], + "spans": [ + { + "bbox": [ + 314, + 121, + 553, + 192 + ], + "type": "text", + "content": "- Semi-Supervised Training Framework: We integrate the fine-tuned MLLMs into a mean teacher framework, enabling high-quality pseudo-label screening for semi-supervised training. The results of SnowMaster exhibits strong generalization capabilities, achieving state-of-the-art performance on real-world snowfall datasets." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 203, + 406, + 216 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 203, + 406, + 216 + ], + "spans": [ + { + "bbox": [ + 314, + 203, + 406, + 216 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 221, + 555, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 221, + 555, + 483 + ], + "spans": [ + { + "bbox": [ + 313, + 221, + 555, + 483 + ], + "type": "text", + "content": "Single Image Desnowing. Snowflakes exhibit diverse sizes and shapes, complicating image restoration compared to other degradations [3, 8, 22-24, 26-28, 40, 49, 54]. Early methods, such as Pei et al. [34], relied on manual feature extraction but struggled in complex scenes. With deep learning advancements, More and more methods focus on snow removal tasks [6, 7, 10, 12, 14, 16, 17, 59]. Liu et al. [30] introduced DesnowNet, a multi-stage network addressing transparent and opaque snow blocks. Chen et al. [13] proposed HDCWNet, utilizing dual-tree complex wavelet transform (DTCWT) and conflicting channel loss (CC Loss) to enhance snow removal accuracy, though high-frequency information recognition remains challenging. Recent approaches integrate semantic and geometric information with Transformer architectures. Zhang et al. [59] developed DDMSNet, combining semantic segmentation and depth estimation for improved snow shape recognition. Cheng et al. [17] introduced SMGARN, using snow masks guided by self-pixel and cross-pixel attention. TransWeather [42], " + }, + { + "bbox": [ + 313, + 221, + 555, + 483 + ], + "type": "inline_equation", + "content": "\\mathrm{T}^3" + }, + { + "bbox": [ + 313, + 221, + 555, + 483 + ], + "type": "text", + "content": "-DiffWeather [11] and other adverse weather restoration methods [15, 25, 33, 55, 56] also advance robustness in extreme weather conditions." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 485, + 556, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 485, + 556, + 675 + ], + "spans": [ + { + "bbox": [ + 313, + 485, + 556, + 675 + ], + "type": "text", + "content": "Multi-modality Large Language Models. Multimodal large language models (MLLMs) revolutionized computer vision by integrating visual and textual information. Models such as GPT-4o, MiniCPM-V [53], and LLava [29] excelled in high-level tasks like image description, visual question answering, and document understanding. However, their performance in low-level tasks, particularly image quality assessment (IQA), remained limited due to insufficient detail perception. Recent studies, such as Q-Instruct [50] and DepictQA [58], addressed this gap by incorporating low-level visual tasks into MLLM training datasets. These approaches leveraged human feedback on low-level features, such as clarity, brightness, and texture details, to improve IQA accuracy. Despite these advancements, MLLMs still struggled with weather-related degradations, as existing datasets were not fine-tuned for such conditions." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "content": "Score-based IQA Methods. Score-based IQA methods aim to assess image quality by assigning numerical scores that align with human perception. Traditional methods," + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "4303" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 70, + 553, + 247 + ], + "blocks": [ + { + "bbox": [ + 59, + 70, + 553, + 247 + ], + "lines": [ + { + "bbox": [ + 59, + 70, + 553, + 247 + ], + "spans": [ + { + "bbox": [ + 59, + 70, + 553, + 247 + ], + "type": "image", + "image_path": "a2dd3875f4fc6612db5df45d776d852b336abe1d344ec7d8ca4b859b10bbb4fb.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 251, + 555, + 274 + ], + "lines": [ + { + "bbox": [ + 55, + 251, + 555, + 274 + ], + "spans": [ + { + "bbox": [ + 55, + 251, + 555, + 274 + ], + "type": "text", + "content": "Figure 2. Dataset screening and evaluation. Through a multi-round dialogue mechanism, the multimodal large language model can replace humans to complete the quality screening and content evaluation of input images." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 282, + 297, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 282, + 297, + 558 + ], + "spans": [ + { + "bbox": [ + 54, + 282, + 297, + 558 + ], + "type": "text", + "content": "such as SSIM [46] and PSNR, relied on statistical metrics to evaluate distortions or pixel-level differences. While effective in controlled environments, they often failed to capture the complexities of real-world images. Deep learning models, such as NIMA [41] and HyperIQA [39], improved performance by regressing directly to Mean Opinion Scores (MOS) derived from human ratings. However, these models faced challenges in out-of-distribution (OOD) generalization, as unseen distortions or diverse content types significantly impacted their accuracy. Recent advancements leveraged vision-language models like CLIP [35], with methods such as CLIP-IQA+ [45] and LIQE [60] utilizing visual-text alignment for few-shot and multitask learning. These approaches demonstrated enhanced robustness across datasets but did not fully capture human-perceived quality due to their reliance on visual-text similarity scores. To address these limitations, Large Multi-Modality Models (LMMs) like Q-Align [47] simulated human rating practices by using discrete text-defined rating levels (e.g., \"excellent\", \"poor\") and converted qualitative ratings into quantitative scores through probability-based averaging. This approach enhanced both accuracy and interpretability, particularly in OOD scenarios." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 564, + 180, + 577 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 564, + 180, + 577 + ], + "spans": [ + { + "bbox": [ + 55, + 564, + 180, + 577 + ], + "type": "text", + "content": "3. Dataset Construction" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 582, + 296, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 296, + 643 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 296, + 643 + ], + "type": "text", + "content": "High-quality real data is the foundation of our approach. We propose RealSnow10K, the largest real snow dataset available. In this section, we describe the process of building the dataset, including image collection, image evaluation and classification, and the dataset preference process." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 648, + 296, + 661 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 648, + 296, + 661 + ], + "spans": [ + { + "bbox": [ + 55, + 648, + 296, + 661 + ], + "type": "text", + "content": "3.1. Image Collection, Evaluation and Classification" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "content": "The first step in building a dataset is to collect images from various sources. We collect more than 100,000 snow scene images from the Internet and screen out over 10,000 high-quality images. To ensure the high quality of the dataset," + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 282, + 555, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 282, + 555, + 426 + ], + "spans": [ + { + "bbox": [ + 313, + 282, + 555, + 426 + ], + "type": "text", + "content": "we need to screen and classify the collected images. Images collected from the Internet often have multiple interferences. In addition to the low quality caused by blur and noise in the image itself, it may also be irrelevant or deviate from the keywords we set, which is often difficult to reflect through existing indicators. To address the above problems, traditional dataset production uses manual screening. Although this method is accurate, it is too inefficient when facing a large number of images. To conduct fast and efficient review, we introduce multimodal large language models (MLLMs) with human visual perception capabilities as a reference for evaluating image quality and content." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 426, + 556, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 556, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 556, + 653 + ], + "type": "text", + "content": "We design a multi-round dialogue mechanism to screen out unqualified images and evaluate the content of the images. Specifically, the prompt words used in the first and second rounds of dialogue are set to \"Is this picture a real photograph? Please answer with only 'yes' or 'no'.\" and \"Does this image have a large and obvious watermark? Please answer with only 'yes' or 'no'\"; respectively, to screen out synthetic images and low-quality images with large watermarks. The third and fourth rounds of dialogue are used to evaluate the content of the pictures. In the third round of dialogue, we use multiple-choice questions to let the model choose the accurate description of the number of snowflakes from five options. We obtain the logits of each option, calculate the option probability and score, and then distinguish the amount of snowflakes and snow accumulation in different pictures. For the specific dialogue process, please refer to Fig. 2. The fourth round of dialogue is the final classification of the topic of the image, and eight categories are set for MLLM to perform classification." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "text", + "content": "In the specific implementation, to balance evaluation quality and efficiency, we select the MiniCPM-V-2.6 model as the dataset evaluation model. Despite having only 8B parameters, this model demonstrates outstanding performance and provides a well-distributed scoring range. We parti" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4304" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 65, + 71, + 289, + 205 + ], + "blocks": [ + { + "bbox": [ + 65, + 71, + 289, + 205 + ], + "lines": [ + { + "bbox": [ + 65, + 71, + 289, + 205 + ], + "spans": [ + { + "bbox": [ + 65, + 71, + 289, + 205 + ], + "type": "image", + "image_path": "c15c69f020dcd5ca55b8fa16974709c555b8f0074f1e014951e6f6101462b847.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 208, + 295, + 251 + ], + "lines": [ + { + "bbox": [ + 55, + 208, + 295, + 251 + ], + "spans": [ + { + "bbox": [ + 55, + 208, + 295, + 251 + ], + "type": "text", + "content": "Figure 3. Performance comparison of different MLLMs for different indicators. The horizontal axis in the figure represents the rank-1 proportion of different MLLMs on different snowfall evaluation indicators." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 254, + 296, + 302 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 254, + 296, + 302 + ], + "spans": [ + { + "bbox": [ + 55, + 254, + 296, + 302 + ], + "type": "text", + "content": "tion the RealSnow10K dataset into three subsets: 1,500 images for MMPO fine-tuning, 1,047 images for testing, and 6,406 images for semi-supervised training. Detailed category statistics are provided in the supplementary materials." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 310, + 272, + 323 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 310, + 272, + 323 + ], + "spans": [ + { + "bbox": [ + 55, + 310, + 272, + 323 + ], + "type": "text", + "content": "3.2. Preference Dataset Construction Pipeline" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 327, + 296, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 327, + 296, + 507 + ], + "spans": [ + { + "bbox": [ + 55, + 327, + 296, + 507 + ], + "type": "text", + "content": "To perform multi-model preference optimization (MMPO) on the multimodal large language model, We first use the pre-trained snow removal model to perform preliminary snow removal on 1,500 images and expand the dataset to 3,000 images. We select two indicators with poor performance of Q-instruct from the four indicators for evaluating snowfall: snowflake volume and snowflake occlusion, and design targeted questions for each. Subsequently, we collect the outputs of four MLLMs for these two problems, including Q-instruct-13b, LLava-1.5-13b, LLava-1.6-13b, and MiniCPM-V-2.6. Image evaluation experts rank the outputs of the four models one by one and generate six preference pairs from a ranking. Therefore, for each preference problem, we obtain a preference dataset with a sample size of 18,000." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 514, + 111, + 526 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 514, + 111, + 526 + ], + "spans": [ + { + "bbox": [ + 55, + 514, + 111, + 526 + ], + "type": "text", + "content": "4. Method" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 530, + 211, + 544 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 530, + 211, + 544 + ], + "spans": [ + { + "bbox": [ + 55, + 530, + 211, + 544 + ], + "type": "text", + "content": "4.1. Evaluation of Snowy Images" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 545, + 295, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 295, + 665 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 295, + 665 + ], + "type": "text", + "content": "Since there is no reference for real snowfall images, the first task for real scene snow removal tasks is to select an accurate and reliable snowfall image evaluation standard. The size, distribution and shape of snowfall particles in the image are often affected by multiple factors including shooting equipment, wind speed, distance, etc., which also brings challenges to the evaluation of snowfall image quality. Traditional non-reference [39, 41] obviously cannot distinguish different forms of snowfall, nor can they accurately describe the interference of snowfall on the image." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "content": "Recently, the development of multimodal large language models has provided a possible solution for snowfall image evaluation. Multimodal large language models are trained on a large number of images and have close to human vi" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 553, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 203 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 203 + ], + "type": "text", + "content": "sual capabilities. Some works [45, 47, 60] introduced CLIP or multimodal large models into image quality assessment tasks (IQA) and obtained good results. Inspired by these works, we try to introduce multimodal large language models into the field of image snow removal. However, there is currently no work exploring the performance of multimodal large language models in evaluating the degradation of snowfall images. Some works [52] use multimodal large language models from a single dimension to evaluate the quality of severe weather images, which is one-sided and unreliable." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 204, + 555, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 204, + 555, + 335 + ], + "spans": [ + { + "bbox": [ + 313, + 204, + 555, + 335 + ], + "type": "text", + "content": "In this work, we first establish a detailed framework for evaluating snowing images. By analyzing a large dataset of real snow images, we propose assessing snow image quality across four dimensions: snowflake amount, object obscURITY, texture detail, and background clarity. The evaluation framework appears in Fig.4 (i). We select three widely used multimodal large language models and one model, Q-Instruct [47], fine-tuned for low-level visual features. These models analyze real snow images along the four dimensions, and experts evaluate their outputs. A sample output is presented in Fig.4 (ii)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 335, + 554, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 335, + 554, + 467 + ], + "spans": [ + { + "bbox": [ + 313, + 335, + 554, + 467 + ], + "type": "text", + "content": "Unfortunately, no model can fully dominate in all four dimensions: based on LLava1.5-13b, Q-instruct, which has been fine-tuned with underlying visual features, has shown obvious advantages in texture detail and background clarity, which is consistent with our expectations. However, in terms of the two indicators of snowflake amount and object obscure for snowflake description, the best performing models are LLava1.6-13b and MiniCPM-V-2.6, which firmly occupy the first and second positions, while q-instruct-13b is relatively backward. The specific rank-1 ratios of different models are shown in Fig.3." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 472, + 553, + 496 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 472, + 553, + 496 + ], + "spans": [ + { + "bbox": [ + 313, + 472, + 553, + 496 + ], + "type": "text", + "content": "4.2. Improving Snowfall Perception via Multi-Model Feedback Optimization" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 498, + 554, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 554, + 629 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 554, + 629 + ], + "type": "text", + "content": "For complex image desnowing task, evaluating performance from multiple perspectives is a prerequisite for achieving satisfactory results. However, as outlined in the previous section, a single multimodal large language model cannot deliver accurate evaluations across all assessment dimensions. At the same time, employing multiple models concurrently for evaluation is computationally infeasible due to excessive resource demands. This leads us to a key question: is it possible to use a single model to simultaneously address the evaluation of multiple perspectives in desnowing tasks?" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "text", + "content": "Fortunately, Q-Instruct [47] model has already undergone targeted fine-tuning on a large-scale dataset, equipping it with the ability to assess fundamental image quality metrics, such as clarity and contrast. Building on this foundation, and inspired by recent advancements in reinforcement learning [18, 36-38], we propose enhancing Q-Instruct through human feedback optimization. Our goal is" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4305" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 70, + 555, + 285 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 555, + 285 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 555, + 285 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 555, + 285 + ], + "type": "image", + "image_path": "77b5777d4af1c3c3e240197cd66f8f0b765c80b1745e242725bf097372a21a6e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 287, + 555, + 331 + ], + "lines": [ + { + "bbox": [ + 55, + 287, + 555, + 331 + ], + "spans": [ + { + "bbox": [ + 55, + 287, + 555, + 331 + ], + "type": "text", + "content": "Figure 4. Schematic diagram of the whole process of snowfall image evaluation optimization. The evaluation framework first divides snowfall into four indicators for evaluation, and for each indicator, collects the evaluation answers of multiple multimodal large language models to the image as the data source for preference learning. Finally, direct preference optimization is applied to Q-instruct to improve its output for this indicator. The figure takes the snowflake amount indicator as an example." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 341, + 294, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 341, + 294, + 365 + ], + "spans": [ + { + "bbox": [ + 54, + 341, + 294, + 365 + ], + "type": "text", + "content": "to enable Q-Instruct to accurately evaluate snowfall, while retaining its capability to assess essential image quality" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 365, + 295, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 365, + 295, + 532 + ], + "spans": [ + { + "bbox": [ + 54, + 365, + 295, + 532 + ], + "type": "text", + "content": "Specifically, to address the challenge that different models exhibit distinct strengths and weaknesses across various perspectives, we introduce a novel multi-model preference optimization (MMPO) technique. Unlike conventional approaches, MMPO constructs preference pairs by harnessing the outputs of multiple models, enabling a more comprehensive capture of diverse evaluation criteria. While drawing inspiration from the loss function of Direct Preference Optimization (DPO) [36], MMPO innovatively extends this framework to incorporate the collective strengths of multiple models into a single, cohesive Q-Instruct model. Consequently, the resulting Q-Instruct model excels at thoroughly assessing desnowing performance while preserving its capability to evaluate critical image quality metrics." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 533, + 295, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 533, + 295, + 582 + ], + "spans": [ + { + "bbox": [ + 55, + 533, + 295, + 582 + ], + "type": "text", + "content": "Reward Function Reparameterization. In DPO, The reward associated with a response " + }, + { + "bbox": [ + 55, + 533, + 295, + 582 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 55, + 533, + 295, + 582 + ], + "type": "text", + "content": " given context " + }, + { + "bbox": [ + 55, + 533, + 295, + 582 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 533, + 295, + 582 + ], + "type": "text", + "content": " is reparameterized in terms of the model's policy " + }, + { + "bbox": [ + 55, + 533, + 295, + 582 + ], + "type": "inline_equation", + "content": "\\pi_{\\theta}(y|x)" + }, + { + "bbox": [ + 55, + 533, + 295, + 582 + ], + "type": "text", + "content": " and a reference policy " + }, + { + "bbox": [ + 55, + 533, + 295, + 582 + ], + "type": "inline_equation", + "content": "\\pi_{\\mathrm{ref}}(y|x)" + }, + { + "bbox": [ + 55, + 533, + 295, + 582 + ], + "type": "text", + "content": ", formulated as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 93, + 586, + 295, + 613 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 586, + 295, + 613 + ], + "spans": [ + { + "bbox": [ + 93, + 586, + 295, + 613 + ], + "type": "interline_equation", + "content": "r (x, y) = \\beta \\log \\frac {\\pi_ {\\theta} (y | x)}{\\pi_ {\\mathrm {r e f}} (y | x)} + \\beta \\log Z (x), \\tag {1}", + "image_path": "3dd52b12cf91f9e000982f5e292d4803517f6626645f1810c51f95df24a1ef0f.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "text", + "content": " is a scaling constant and " + }, + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "inline_equation", + "content": "Z(x)" + }, + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "text", + "content": " is a partition function to normalize the policy distribution. Through this reparameterization, DPO completes preference optimization without explicitly defining a reward model, simplifying the complex reinforcement learning problem into a classification-based optimization problem. Our proposed MMPO also adopts this reparameterization technique to streamline the formulation of preference optimization." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 341, + 555, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 341, + 555, + 461 + ], + "spans": [ + { + "bbox": [ + 313, + 341, + 555, + 461 + ], + "type": "text", + "content": "Multi-Model Feedback Collection. To construct a robust preference dataset, a multi-model feedback collection strategy is employed. For a given input context " + }, + { + "bbox": [ + 313, + 341, + 555, + 461 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 341, + 555, + 461 + ], + "type": "text", + "content": ", responses " + }, + { + "bbox": [ + 313, + 341, + 555, + 461 + ], + "type": "inline_equation", + "content": "\\{y_1, y_2, \\ldots, y_N\\}" + }, + { + "bbox": [ + 313, + 341, + 555, + 461 + ], + "type": "text", + "content": " are collected from " + }, + { + "bbox": [ + 313, + 341, + 555, + 461 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 341, + 555, + 461 + ], + "type": "text", + "content": " distinct multimodal large language models. These responses are subsequently evaluated and ranked by human experts, yielding a preference ordering " + }, + { + "bbox": [ + 313, + 341, + 555, + 461 + ], + "type": "inline_equation", + "content": "y_{(1)} \\succ y_{(2)} \\succ \\dots \\succ y_{(N)}" + }, + { + "bbox": [ + 313, + 341, + 555, + 461 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 341, + 555, + 461 + ], + "type": "inline_equation", + "content": "y_{(1)}" + }, + { + "bbox": [ + 313, + 341, + 555, + 461 + ], + "type": "text", + "content": " denotes the most preferred response and " + }, + { + "bbox": [ + 313, + 341, + 555, + 461 + ], + "type": "inline_equation", + "content": "y_{(N)}" + }, + { + "bbox": [ + 313, + 341, + 555, + 461 + ], + "type": "text", + "content": " the least preferred. From this ordering, preference pairs " + }, + { + "bbox": [ + 313, + 341, + 555, + 461 + ], + "type": "inline_equation", + "content": "(y_w, y_l)" + }, + { + "bbox": [ + 313, + 341, + 555, + 461 + ], + "type": "text", + "content": " are generated such that " + }, + { + "bbox": [ + 313, + 341, + 555, + 461 + ], + "type": "inline_equation", + "content": "y_w \\succ y_l" + }, + { + "bbox": [ + 313, + 341, + 555, + 461 + ], + "type": "text", + "content": ", forming the dataset:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 323, + 471, + 553, + 484 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 471, + 553, + 484 + ], + "spans": [ + { + "bbox": [ + 323, + 471, + 553, + 484 + ], + "type": "interline_equation", + "content": "\\left\\{\\left(x, y _ {w}, y _ {l}\\right) \\mid y _ {w} \\succ y _ {l}, y _ {w}, y _ {l} \\in \\left\\{y _ {1}, y _ {2}, \\dots , y _ {N} \\right\\} \\right\\}. \\tag {2}", + "image_path": "e80de651abd1cfb68792d8bfbb3c837f12d0760f1e90478b413bcd14ee6c0f3f.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 493, + 553, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 493, + 553, + 540 + ], + "spans": [ + { + "bbox": [ + 313, + 493, + 553, + 540 + ], + "type": "text", + "content": "This methodology harnesses the diverse capabilities of multiple models to produce a comprehensive set of preference pairs, thereby enhancing the differentiation between superior and inferior responses." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 541, + 553, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 541, + 553, + 590 + ], + "spans": [ + { + "bbox": [ + 313, + 541, + 553, + 590 + ], + "type": "text", + "content": "Multi-Model Preference Optimization. Building upon the framework of Direct Preference Optimization (DPO), Multi-Model Preference Optimization (MMPO) is introduced. The loss function is expressed as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 363, + 599, + 553, + 613 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 363, + 599, + 553, + 613 + ], + "spans": [ + { + "bbox": [ + 363, + 599, + 553, + 613 + ], + "type": "interline_equation", + "content": "L _ {\\mathrm {M M P O}} \\left(\\pi_ {\\theta}; \\pi_ {\\mathrm {r e f}}\\right) = - \\mathbb {E} _ {(x, y _ {w}, y _ {l}) \\sim D} \\tag {3}", + "image_path": "61a2626478fe58adad672535e8c97463263bd1aa98633cb7c8d2f06a76eb0a0b.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 335, + 621, + 533, + 649 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 621, + 533, + 649 + ], + "spans": [ + { + "bbox": [ + 335, + 621, + 533, + 649 + ], + "type": "interline_equation", + "content": "\\left[ \\log \\sigma \\left(\\beta \\log \\frac {\\pi_ {\\theta} \\left(y _ {w} \\mid x\\right)}{\\pi_ {\\mathrm {r e f}} \\left(y _ {w} \\mid x\\right)} - \\beta \\log \\frac {\\pi_ {\\theta} \\left(y _ {l} \\mid x\\right)}{\\pi_ {\\mathrm {r e f}} \\left(y _ {l} \\mid x\\right)}\\right) \\right],", + "image_path": "d82bdebcf3272c7dae364600e5c69b3015ea00db945604973c61ec9fa337cce7.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 653, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 555, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 653, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\sigma (\\cdot)" + }, + { + "bbox": [ + 313, + 653, + 555, + 713 + ], + "type": "text", + "content": " represents the sigmoid function, and " + }, + { + "bbox": [ + 313, + 653, + 555, + 713 + ], + "type": "inline_equation", + "content": "y_{w}" + }, + { + "bbox": [ + 313, + 653, + 555, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 653, + 555, + 713 + ], + "type": "inline_equation", + "content": "y_{l}" + }, + { + "bbox": [ + 313, + 653, + 555, + 713 + ], + "type": "text", + "content": " denote the preferred and unpreferred responses, respectively. A distinguishing feature of MMPO is its reliance on a preference dataset " + }, + { + "bbox": [ + 313, + 653, + 555, + 713 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 313, + 653, + 555, + 713 + ], + "type": "text", + "content": " derived from the outputs of multiple models, as outlined in the feedback collection process. This" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4306" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "text", + "content": "approach integrates the strengths of various MLLMs, allowing the model to leverage a richer collection of preference pairs. Through this optimization, the model's outputs are aligned with human expert preferences, thereby improving its capacity to generate high-quality responses." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 139, + 228, + 152 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 139, + 228, + 152 + ], + "spans": [ + { + "bbox": [ + 55, + 139, + 228, + 152 + ], + "type": "text", + "content": "4.3. From Perception to Restoration" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 156, + 296, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 156, + 296, + 239 + ], + "spans": [ + { + "bbox": [ + 55, + 156, + 296, + 239 + ], + "type": "text", + "content": "Using MMPO technology, we use Q-instruct as the base model and fine-tune its preference learning in two aspects where its performance was relatively poor: snowflake amount and object obscure, improving its ability to perceive and understand snowfall. However, how can the improvement in perception be used to optimize the image restoration model?" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 240, + 296, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 240, + 296, + 323 + ], + "spans": [ + { + "bbox": [ + 55, + 240, + 296, + 323 + ], + "type": "text", + "content": "Inspired by Q-Align [47], we leverage the enhanced perception capabilities of the large model to generate an \"image score\" that could quantify this improvement. We then use this image score as an evaluation metric within a mean teacher semi-supervised learning framework, combined with a pseudo-label database, to train the image restoration model." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 324, + 296, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 324, + 296, + 432 + ], + "spans": [ + { + "bbox": [ + 55, + 324, + 296, + 432 + ], + "type": "text", + "content": "Score calculation. Using one adjective like Q-Align (e.g. \"excellent\", \"poor\") to describe the snowfall scene directly does not provide an accurate evaluation. We analyze that this is because a single adjective lacks a specific definition. Therefore, we further decompose the prompt words, describe the four indicators such as the number of snowflakes from different perspectives, and design it into a multiple choice question with five options. Based on the calculation method of human subjective opinion score (MOS):" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 135, + 441, + 295, + 473 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 441, + 295, + 473 + ], + "spans": [ + { + "bbox": [ + 135, + 441, + 295, + 473 + ], + "type": "interline_equation", + "content": "\\mathrm {M O S} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} S _ {i}, \\tag {4}", + "image_path": "c2633eb602eb022869222a901cd3fcec825ba8ad9bd53c70fe8c6440995cb503.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 477, + 296, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 477, + 296, + 525 + ], + "spans": [ + { + "bbox": [ + 55, + 477, + 296, + 525 + ], + "type": "text", + "content": "in which " + }, + { + "bbox": [ + 55, + 477, + 296, + 525 + ], + "type": "inline_equation", + "content": "S_{i}" + }, + { + "bbox": [ + 55, + 477, + 296, + 525 + ], + "type": "text", + "content": " represents the score of the i-th tester. We collect the model's output logit for the five option tokens, use softmax to calculate the probability of different tokens, and finally get the score through weighted average:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 127, + 534, + 295, + 567 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 534, + 295, + 567 + ], + "spans": [ + { + "bbox": [ + 127, + 534, + 295, + 567 + ], + "type": "interline_equation", + "content": "\\text {S c o r e} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} p _ {i} \\cdot v _ {i}, \\tag {5}", + "image_path": "1d36af36775a4c33f5256fa53681834f0d089c21fa0006567eaa3e084906dde9.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 141, + 571, + 295, + 602 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 571, + 295, + 602 + ], + "spans": [ + { + "bbox": [ + 141, + 571, + 295, + 602 + ], + "type": "interline_equation", + "content": "p _ {i} = \\frac {e ^ {l _ {i}}}{\\sum_ {j = 1} ^ {N} e ^ {l _ {j}}}, \\tag {6}", + "image_path": "adf7f26a3fb29ec8881d7b4a3e54f3a078d0a60e1c0ebb37b9d7f51bfbe2ef7a.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 605, + 295, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 653 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 653 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 605, + 295, + 653 + ], + "type": "inline_equation", + "content": "l_{i}" + }, + { + "bbox": [ + 55, + 605, + 295, + 653 + ], + "type": "text", + "content": " represents the logit value of the " + }, + { + "bbox": [ + 55, + 605, + 295, + 653 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 605, + 295, + 653 + ], + "type": "text", + "content": "-th option token. " + }, + { + "bbox": [ + 55, + 605, + 295, + 653 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 605, + 295, + 653 + ], + "type": "text", + "content": " is 5 to correspond to 5 options. " + }, + { + "bbox": [ + 55, + 605, + 295, + 653 + ], + "type": "inline_equation", + "content": "v_{i}" + }, + { + "bbox": [ + 55, + 605, + 295, + 653 + ], + "type": "text", + "content": " is converted into weights from 1 to 5 according to the description of snowfall intensity." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": "Semi-supervised Training with Pseudo-labeling. Referring to the mean teacher framework and the database mechanism in [36], we leverage scores from four evaluation metrics to enhance semi-supervised training. To ensure the pseudo-label database is updated accurately, we adopt a" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 317, + 71, + 420, + 137 + ], + "blocks": [ + { + "bbox": [ + 317, + 71, + 420, + 137 + ], + "lines": [ + { + "bbox": [ + 317, + 71, + 420, + 137 + ], + "spans": [ + { + "bbox": [ + 317, + 71, + 420, + 137 + ], + "type": "image", + "image_path": "ee17afd6786d89eb557b937995f149fed4ed54accad3dfc06fd0a93f2414eafc.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 420, + 71, + 436, + 87 + ], + "blocks": [ + { + "bbox": [ + 420, + 71, + 436, + 87 + ], + "lines": [ + { + "bbox": [ + 420, + 71, + 436, + 87 + ], + "spans": [ + { + "bbox": [ + 420, + 71, + 436, + 87 + ], + "type": "image", + "image_path": "3210ac6913cd883be5f58b4c6a536b898129f8a549a567a5eb59537ba0dafebd.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 437, + 72, + 554, + 105 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 437, + 72, + 554, + 105 + ], + "spans": [ + { + "bbox": [ + 437, + 72, + 554, + 105 + ], + "type": "text", + "content": "How would you evaluate the amount of snowflakes in the air shown in the picture? Please provide your judgment and basis. Summarize in about 50 words." + } + ] + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 421, + 107, + 435, + 122 + ], + "blocks": [ + { + "bbox": [ + 421, + 107, + 435, + 122 + ], + "lines": [ + { + "bbox": [ + 421, + 107, + 435, + 122 + ], + "spans": [ + { + "bbox": [ + 421, + 107, + 435, + 122 + ], + "type": "image", + "image_path": "89b66cd64fcc15cbb955c9750b2e9a1d737825a1114060a9b2b46d5e586b6ea5.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 437, + 108, + 545, + 123 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 437, + 108, + 545, + 123 + ], + "spans": [ + { + "bbox": [ + 437, + 108, + 545, + 123 + ], + "type": "text", + "content": "Keywords that do not match the real picture Keywords that lack logic or basis" + } + ] + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 421, + 123, + 435, + 137 + ], + "blocks": [ + { + "bbox": [ + 421, + 123, + 435, + 137 + ], + "lines": [ + { + "bbox": [ + 421, + 123, + 435, + 137 + ], + "spans": [ + { + "bbox": [ + 421, + 123, + 435, + 137 + ], + "type": "image", + "image_path": "abb41cec1c16737bc1aa024a727e864851e557548fa680df04ab50597995486a.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "bbox": [ + 437, + 126, + 553, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 437, + 126, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 437, + 126, + 553, + 140 + ], + "type": "text", + "content": "Keywords that match the picture and perception of human experts" + } + ] + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 317, + 144, + 331, + 160 + ], + "blocks": [ + { + "bbox": [ + 317, + 144, + 331, + 160 + ], + "lines": [ + { + "bbox": [ + 317, + 144, + 331, + 160 + ], + "spans": [ + { + "bbox": [ + 317, + 144, + 331, + 160 + ], + "type": "image", + "image_path": "375ddf0bac9b4e2b9e9d417421990387a33418d63690aca5f8f6d053e0cc47e4.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "bbox": [ + 335, + 152, + 368, + 159 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 152, + 368, + 159 + ], + "spans": [ + { + "bbox": [ + 335, + 152, + 368, + 159 + ], + "type": "text", + "content": "Q-instruct" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 160, + 431, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 160, + 431, + 266 + ], + "spans": [ + { + "bbox": [ + 316, + 160, + 431, + 266 + ], + "type": "text", + "content": "The amount of snowflakes in the air shown in the picture is quite a lot. This is evident from the fact that the people in the image are wearing winter coats and hats, and their clothing is covered in snow. The snowflakes are also clearly visible in the air, indicating that the snowfall is quite heavy. The people are walking in the snow, which further emphasizes the significant amount of snowfall. Overall, the image portrays a snowy day with a considerable amount of snowflakes in the air." + } + ] + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 447, + 143, + 466, + 162 + ], + "blocks": [ + { + "bbox": [ + 447, + 143, + 466, + 162 + ], + "lines": [ + { + "bbox": [ + 447, + 143, + 466, + 162 + ], + "spans": [ + { + "bbox": [ + 447, + 143, + 466, + 162 + ], + "type": "image", + "image_path": "7badbd3e3b965cc93128f598bc481dd9a0c321059cbf0fb0edeabd768e2545e2.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "bbox": [ + 468, + 151, + 526, + 158 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 468, + 151, + 526, + 158 + ], + "spans": [ + { + "bbox": [ + 468, + 151, + 526, + 158 + ], + "type": "text", + "content": "O-instruct + MMPO" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 443, + 161, + 553, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 443, + 161, + 553, + 266 + ], + "spans": [ + { + "bbox": [ + 443, + 161, + 553, + 266 + ], + "type": "text", + "content": "The image shows a group of people walking in the snow, with several snowflakes visible in the air. While it is not possible to quantify the exact number of snowflakes, it is evident that there is a noticeable amount of snowfall. The snowflakes are dispersed throughout the scene, indicating that the snowfall is not too heavy or too light, but rather moderate. This suggests that the weather conditions are likely to be cold and snowy, but not extreme or overwhelming." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 269, + 554, + 303 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 269, + 554, + 303 + ], + "spans": [ + { + "bbox": [ + 313, + 269, + 554, + 303 + ], + "type": "text", + "content": "Figure 5. Comparison of perception ability before and after MMPO. The fine-tuned model output is more in line with human visual perception, more logical, and less hallucinatory." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 311, + 554, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 311, + 554, + 382 + ], + "spans": [ + { + "bbox": [ + 313, + 311, + 554, + 382 + ], + "type": "text", + "content": "stringent rule: an update occurs only if " + }, + { + "bbox": [ + 313, + 311, + 554, + 382 + ], + "type": "inline_equation", + "content": "(i)" + }, + { + "bbox": [ + 313, + 311, + 554, + 382 + ], + "type": "text", + "content": " at least three teacher scores exceed the student score, and " + }, + { + "bbox": [ + 313, + 311, + 554, + 382 + ], + "type": "inline_equation", + "content": "(ii)" + }, + { + "bbox": [ + 313, + 311, + 554, + 382 + ], + "type": "text", + "content": " the teacher score surpasses the current score stored in the pseudo-label database. This dual-condition approach effectively minimizes erroneous updates to the pseudo-labels, improving the reliability of the training process." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 383, + 555, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 383, + 555, + 502 + ], + "spans": [ + { + "bbox": [ + 313, + 383, + 555, + 502 + ], + "type": "text", + "content": "Loss function and training strategy. We use both supervised loss and pseudo-label loss to train the model. The supervised loss is based on the PSNR between the model's output and the reference images from the Snow100K [30] dataset. And the pseudo-label loss relies on the L1 norm to measure the difference between the output and pseudolabels. To ensure fast and stable training, we only use supervised loss in the initial 20k steps, and add pseudo-label loss after 20k steps for semi-supervised training. The total loss function is shown as follows:" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 324, + 522, + 553, + 536 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 522, + 553, + 536 + ], + "spans": [ + { + "bbox": [ + 324, + 522, + 553, + 536 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {\\text {s u p e r v i s e d}} + \\lambda \\mathcal {L} _ {\\text {p s e u d o}} (\\lambda = 0 \\text {i f s t e p < 2 0 k}). \\tag {7}", + "image_path": "6f96dbcf9654b362a9dfb0b5caa8b318f88a4f7d4a3f29b00ffe9edb0b38e978.jpg" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 314, + 550, + 395, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 550, + 395, + 563 + ], + "spans": [ + { + "bbox": [ + 314, + 550, + 395, + 563 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 314, + 567, + 422, + 580 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 567, + 422, + 580 + ], + "spans": [ + { + "bbox": [ + 314, + 567, + 422, + 580 + ], + "type": "text", + "content": "5.1. Experiment Setup" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "content": "Training Settings. Our method consists of two phases: the first phase is to fine-tune the scoring model (i.e. Q-Instruct [50]) via proposed MMPO, and the second phase is to semi-supervised train the snow removal model. In the first phase, we set epoch = 1, lr = 1e-6, batchsize = 64; We use Vision-LLM Alignment framework [44] to perform MMPO fine-tuning and enable deepspeed for acceleration, set zero-stage to 3 and enable offload for full fine-tuning. In the second phase of snow removal model training, we use NAFNet [5] as the base model, set lr = 2e-4, batchsize = 32. All experiments are performed on two Nvidia A800 GPUs." + } + ] + } + ], + "index": 30 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4307" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 72, + 553, + 411 + ], + "blocks": [ + { + "bbox": [ + 58, + 72, + 553, + 411 + ], + "lines": [ + { + "bbox": [ + 58, + 72, + 553, + 411 + ], + "spans": [ + { + "bbox": [ + 58, + 72, + 553, + 411 + ], + "type": "image", + "image_path": "64e5340073e481c4490fd8bcf2efbe32b44ed03300b555a63c0bf19b5425abb6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 412, + 555, + 434 + ], + "lines": [ + { + "bbox": [ + 55, + 412, + 555, + 434 + ], + "spans": [ + { + "bbox": [ + 55, + 412, + 555, + 434 + ], + "type": "text", + "content": "Figure 6. Visual comparison of different models for snow removal on real images. Please zoom in for a better view. SnowMaster achieves satisfactory snow removal performance in complicated conditions compared to previous methods." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 62, + 455, + 294, + 539 + ], + "blocks": [ + { + "bbox": [ + 67, + 443, + 284, + 453 + ], + "lines": [ + { + "bbox": [ + 67, + 443, + 284, + 453 + ], + "spans": [ + { + "bbox": [ + 67, + 443, + 284, + 453 + ], + "type": "text", + "content": "Table 1. The accuracy of MLLMs on different perspectives." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 62, + 455, + 294, + 539 + ], + "lines": [ + { + "bbox": [ + 62, + 455, + 294, + 539 + ], + "spans": [ + { + "bbox": [ + 62, + 455, + 294, + 539 + ], + "type": "table", + "html": "
MLLMs AccuracySnowflakes accQuality acc
amount ↑obscure ↑texture ↑background ↑
LLava-v1.5-13b[29]80.0477.1755.7852.15
LLava-v1.6-13b[29]86.3479.5652.0554.06
MiniCPM-V[53]75.4570.0442.0249.76
Q-intract-13b[50]76.3173.2658.3663.42
Q-instruct with MMPO (ours)82.1480.9064.9568.58
MMPO improvement+5.83+7.64+6.59+5.16
", + "image_path": "ca2747378d70aa7be744029e4980c7548d597f8f50dc6979b838aee10f920068.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 555, + 295, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 555, + 295, + 650 + ], + "spans": [ + { + "bbox": [ + 55, + 555, + 295, + 650 + ], + "type": "text", + "content": "Experimental Details. In the semi-supervised training process, considering the need to make the multimodal large language model output the correct score, we adopt a pseudo-label database initialization strategy using the original image instead of all-0 initialization in [21]. In the test phase, we test 457 test images selected from the real snow scene dataset (RealSnow10K) and 1329 real snow scene images from Snow100K [30]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 658, + 234, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 658, + 234, + 670 + ], + "spans": [ + { + "bbox": [ + 55, + 658, + 234, + 670 + ], + "type": "text", + "content": "5.2. Evaluation Metrics and Baselines" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": "Following [19, 20], we utilize four non-reference IQA metrics: BRISQUE [31], NIQE [32], PIQE [43], and PAQ2-PIQ [57]. BRISQUE measures quality via spa" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 444, + 555, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 444, + 555, + 564 + ], + "spans": [ + { + "bbox": [ + 313, + 444, + 555, + 564 + ], + "type": "text", + "content": "tial natural scene statistics, NIQE assesses deviations from these statistics, PIQE detects block-level distortions, and PAQ2-PIQ, a deep learning approach, predicts global and local quality. These metrics ensure a comprehensive, unsupervised evaluation of diverse distortions. For the baselines, we choose CNN-based models (SMGARN[17], NAFNet[5]), Transformer-based models (Transweather[42], Snowformer[6]) and a diffusion based model " + }, + { + "bbox": [ + 313, + 444, + 555, + 564 + ], + "type": "inline_equation", + "content": "(\\mathrm{T}^3" + }, + { + "bbox": [ + 313, + 444, + 555, + 564 + ], + "type": "text", + "content": " -DiffWeather[11]), offering varied references for snowfall image restoration performance." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 573, + 433, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 573, + 433, + 586 + ], + "spans": [ + { + "bbox": [ + 313, + 573, + 433, + 586 + ], + "type": "text", + "content": "5.3. Results and Analysis" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 589, + 430, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 589, + 430, + 601 + ], + "spans": [ + { + "bbox": [ + 313, + 589, + 430, + 601 + ], + "type": "text", + "content": "5.3.1. Quantitative Results" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "content": "Effectiveness of MMPO. Table 1 illustrates the performance enhancements of the Q-Instruct [47] model following fine-tuning with MMPO. We utilize NAFNet [5], pretrained for 100,000 steps, as the snow removal model to process the images in the test set, generating a dataset that includes both original and snow-removed images. By inputting these images individually into Q-Instruct, we obtain scores for each image corresponding to specific indicators. Consequently, for each indicator, accuracy is defined" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4308" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 80, + 81, + 529, + 178 + ], + "blocks": [ + { + "bbox": [ + 137, + 68, + 472, + 79 + ], + "lines": [ + { + "bbox": [ + 137, + 68, + 472, + 79 + ], + "spans": [ + { + "bbox": [ + 137, + 68, + 472, + 79 + ], + "type": "text", + "content": "Table 2. Comparison of non-reference metrics after snow removal between different models." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 80, + 81, + 529, + 178 + ], + "lines": [ + { + "bbox": [ + 80, + 81, + 529, + 178 + ], + "spans": [ + { + "bbox": [ + 80, + 81, + 529, + 178 + ], + "type": "table", + "html": "
MethodSnow100K-realistic [30]RealSnow10K
BRISQUE ↓NIQE ↓PIQE ↓PAQ2-PIQ ↑BRISQUE ↓NIQE ↓PIQE ↓PAQ2-PIQ ↑
SMGARN[TCSVT'22][17]19.2893.13231.92769.76723.0053.58031.21068.919
TransWeather[CVPR'22] [42]16.5443.02530.30269.82118.5643.40429.77269.671
NAFNet[ECCV'22] [5]17.5662.90330.84369.95120.4253.29129.87769.388
SnowFormer[Arxiv'23][6]19.3892.95431.90270.01523.7023.42931.9969.929
T3-DiffWeather[ECCV'24] [11]17.7442.86831.68970.04220.2113.13732.99170.211
★ SnowMaster (Ours)16.3052.79730.16270.04217.5463.05829.74270.304
", + "image_path": "522db6fecd867b30b580ea2f6a397af7d4a153628e3e2cf5bcd92203778f8874.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 193, + 294, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 193, + 294, + 323 + ], + "spans": [ + { + "bbox": [ + 54, + 193, + 294, + 323 + ], + "type": "text", + "content": "as the proportion of snow-removed images whose score is not lower than that of their corresponding original images. The results show that the performance after fine-tuning has been significantly improved which is consistent with our expectations. It is worth noting that in terms of the number of snowflakes, the accuracy after fine-tuning does not exceed that of Llava-1.6. We think this is normal, because the preference data in fine-tuning are all output by these models, and MMPO guides the model to move closer to human preference output, but it may not necessarily surpass the output of the optimal model." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 324, + 294, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 324, + 294, + 408 + ], + "spans": [ + { + "bbox": [ + 55, + 324, + 294, + 408 + ], + "type": "text", + "content": "Furthermore, a noteworthy observation emerges: the indicators 'texture' and 'background', which are not explicitly targeted for optimization, also exhibit moderate improvements. We hypothesize that this phenomenon arises from a comprehensive augmentation of the model's perceptual capacity regarding snowfall, facilitated by the MMPO fine-tuning process." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 408, + 294, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 408, + 294, + 456 + ], + "spans": [ + { + "bbox": [ + 55, + 408, + 294, + 456 + ], + "type": "text", + "content": "Performance analysis of the restoration model. Table 2 presents the performance comparison between our method and other advanced snow removal models. Our proposed method achieves the best results in most indicators." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 461, + 183, + 473 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 461, + 183, + 473 + ], + "spans": [ + { + "bbox": [ + 55, + 461, + 183, + 473 + ], + "type": "text", + "content": "5.3.2. Qualitative Assessment" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 476, + 295, + 621 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 476, + 295, + 621 + ], + "spans": [ + { + "bbox": [ + 55, + 476, + 295, + 621 + ], + "type": "text", + "content": "From the restoration results of the real snowfall image in Fig.6, we can see that the SnowMaster model we proposed has the best performance in multiple scenarios. When facing snowflakes illuminated by lights at night, other models cannot effectively complete the restoration of high-brightness snowflakes, while SnowMaster significantly eliminates the high-brightness snowflakes in the image. When restoring dense snowflake scenes during the day, SnowMaster can effectively remove snowflakes of different scales and retain the original color and texture of the object, while other models have poor restoration effects on small-sized snowflake particles." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 626, + 149, + 639 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 626, + 149, + 639 + ], + "spans": [ + { + "bbox": [ + 55, + 626, + 149, + 639 + ], + "type": "text", + "content": "5.4. Ablation Study" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 641, + 294, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 294, + 713 + ], + "type": "text", + "content": "Necessity of MMPO. The Q-Instruct model shows significant improvement after multi-modal preference optimization, particularly in evaluating snowfall images. We perform an ablation test to compare the Q-Instruct model before and after alignment when used as a scoring model in semi-supervised training. The results indicate that the" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 325, + 204, + 545, + 277 + ], + "blocks": [ + { + "bbox": [ + 314, + 191, + 553, + 201 + ], + "lines": [ + { + "bbox": [ + 314, + 191, + 553, + 201 + ], + "spans": [ + { + "bbox": [ + 314, + 191, + 553, + 201 + ], + "type": "text", + "content": "Table 3. Ablation on different settings of the proposed framework." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 325, + 204, + 545, + 277 + ], + "lines": [ + { + "bbox": [ + 325, + 204, + 545, + 277 + ], + "spans": [ + { + "bbox": [ + 325, + 204, + 545, + 277 + ], + "type": "table", + "html": "
MethodBRISQUE ↓NIQE ↓
w/o. Semi-supervised learning22.523.33
w/o. MLLM20.363.25
w/o. MMPO18.983.19
w. Single-angle evaluation18.213.12
w. Multi-angle evalutaion (Ours)17.553.06
", + "image_path": "721e09c2405770ada25693e7dacac1be14df39f8492de4842cb79e781e8701cd.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 288, + 553, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 288, + 553, + 312 + ], + "spans": [ + { + "bbox": [ + 313, + 288, + 553, + 312 + ], + "type": "text", + "content": "aligned model achieves higher accuracy and ensures more stable training to improve the performance." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 313, + 553, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 313, + 553, + 396 + ], + "spans": [ + { + "bbox": [ + 313, + 313, + 553, + 396 + ], + "type": "text", + "content": "Effectiveness of multi-perspectives snowfall evaluation. Our method enhances image screening reliability by evaluating snowfall images from four perspectives, ensuring steady updates to the pseudo-label database. Ablation experiments, conducted by reducing evaluation angles, demonstrate that multi-angle evaluation improves training effectiveness" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 398, + 554, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 398, + 554, + 517 + ], + "spans": [ + { + "bbox": [ + 313, + 398, + 554, + 517 + ], + "type": "text", + "content": "The role of the semi-supervised learning framework. In order to transfer the improvement of the evaluation performance of the large model on snow scene images to the improvement of the restoration model, we use the mean teacher semi-supervised learning framework. By combining the supervision loss from snow100k and the pseudolabel loss from the real training data, this framework can effectively optimize the restoration effect of real snowfall images. Through ablation experiments, we prove that the semi-supervised framework is the basis of our method." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 518, + 553, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 518, + 553, + 541 + ], + "spans": [ + { + "bbox": [ + 313, + 518, + 553, + 541 + ], + "type": "text", + "content": "The specific metrics of the ablation experiments are shown in Table 3." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 552, + 388, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 552, + 388, + 563 + ], + "spans": [ + { + "bbox": [ + 313, + 552, + 388, + 563 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 570, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 554, + 713 + ], + "type": "text", + "content": "SnowMaster integrates a multimodal large language model with a multi-perspective evaluation framework for image desnowing. Leveraging the RealSnow10k dataset, it utilizes a multi-angle evaluation system assessing four key metrics to enhance real-world reliability. Through multimodal preference optimization and semi-supervised training with pseudo-label updates from multi-angle multiple-choice questions, it ensures effective refinement. Experiments confirm its robustness in complex snowfall scenarios, though its computational intensity from multiple evaluations signals a need for future enhancements in training efficiency and resource use." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4309" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 152 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 152 + ], + "type": "text", + "content": "This work is supported by the Guangzhou-HKUST(GZ) Joint Funding Program (No. 2023A03J0671), the Guangzhou Municipal Science and Technology Project (Grant No. 2024A04J4230), and Guangzhou-HKUST(GZ) Joint Funding Program (No. 2024A03J0618)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 164, + 115, + 176 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 164, + 115, + 176 + ], + "spans": [ + { + "bbox": [ + 56, + 164, + 115, + 176 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 184, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 184, + 296, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 184, + 296, + 239 + ], + "spans": [ + { + "bbox": [ + 61, + 184, + 296, + 239 + ], + "type": "text", + "content": "[1] Vijay Badrinarayanan, Alex Kendall, and Roberto Cipolla. Segnet: A deep convolutional encoder-decoder architecture for image segmentation. IEEE transactions on pattern analysis and machine intelligence, 39(12):2481-2495, 2017. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 240, + 296, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 240, + 296, + 295 + ], + "spans": [ + { + "bbox": [ + 61, + 240, + 296, + 295 + ], + "type": "text", + "content": "[2] Chun-Fu Richard Chen, Quanfu Fan, and Rameswar Panda. Crossvit: Cross-attention multi-scale vision transformer for image classification. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 357-366, 2021. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 297, + 296, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 297, + 296, + 361 + ], + "spans": [ + { + "bbox": [ + 61, + 297, + 296, + 361 + ], + "type": "text", + "content": "[3] Haoyu Chen, Jinjin Gu, Yihao Liu, Salma Abdel Magid, Chao Dong, Qiong Wang, Hanspeter Pfister, and Lei Zhu. Masked image training for generalizable deep image denoising. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1692-1703, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 364, + 296, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 364, + 296, + 430 + ], + "spans": [ + { + "bbox": [ + 62, + 364, + 296, + 430 + ], + "type": "text", + "content": "[4] Haoyu Chen, Wenbo Li, Jinjin Gu, Jingjing Ren, Haoze Sun, Xueyi Zou, Zhensong Zhang, Youliang Yan, and Lei Zhu. Low-res leads the way: Improving generalization for superresolution by self-supervised learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25857-25867, 2024. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 432, + 296, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 432, + 296, + 475 + ], + "spans": [ + { + "bbox": [ + 62, + 432, + 296, + 475 + ], + "type": "text", + "content": "[5] Liangyu Chen, Xiaojie Chu, Xiangyu Zhang, and Jian Sun. Simple baselines for image restoration. In European conference on computer vision, pages 17-33. Springer, 2022. 6, 7, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 477, + 296, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 477, + 296, + 521 + ], + "spans": [ + { + "bbox": [ + 62, + 477, + 296, + 521 + ], + "type": "text", + "content": "[6] Sixiang Chen, Tian Ye, Yun Liu, and Erkang Chen. Snowformer: Context interaction transformer with scale-awareness for single image desnowing. arXiv preprint arXiv:2208.09703, 2022. 2, 7, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 522, + 296, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 522, + 296, + 566 + ], + "spans": [ + { + "bbox": [ + 62, + 522, + 296, + 566 + ], + "type": "text", + "content": "[7] Sixiang Chen, Tian Ye, Yun Liu, Taodong Liao, Yi Ye, and Erkang Chen. Msp-former: Multi-scale projection transformer for single image desnowing. arXiv preprint arXiv:2207.05621, 2022. 1, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 567, + 296, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 567, + 296, + 632 + ], + "spans": [ + { + "bbox": [ + 62, + 567, + 296, + 632 + ], + "type": "text", + "content": "[8] Sixiang Chen, Tian Ye, Jinbin Bai, Erkang Chen, Jun Shi, and Lei Zhu. Sparse sampling transformer with uncertainty-driven ranking for unified removal of raindrops and rain streaks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13106-13117, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 635, + 296, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 635, + 296, + 689 + ], + "spans": [ + { + "bbox": [ + 62, + 635, + 296, + 689 + ], + "type": "text", + "content": "[9] Sixiang Chen, Tian Ye, Yun Liu, Jinbin Bai, Haoyu Chen, Yunlong Lin, Jun Shi, and Erkang Chen. CPLformer: Cross-scale prototype learning transformer for image snow removal. In Proceedings of the 31st ACM international conference on multimedia, pages 4228-4239, 2023. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 691, + 295, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 295, + 714 + ], + "type": "text", + "content": "[10] Sixiang Chen, Tian Ye, Chenghao Xue, Haoyu Chen, Yun Liu, Erkang Chen, and Lei Zhu. Uncertainty-driven dynamic" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 333, + 73, + 555, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 555, + 116 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 555, + 116 + ], + "type": "text", + "content": "degradation perceiving and background modeling for efficient single image desnowing. In Proceedings of the 31st ACM International Conference on Multimedia, pages 4269-4280, 2023. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 118, + 554, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 554, + 173 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 554, + 173 + ], + "type": "text", + "content": "[11] Sixiang Chen, Tian Ye, Kai Zhang, Zhaohu Xing, Yunlong Lin, and Lei Zhu. Teaching tailored to talent: Adverse weather restoration via prompt pool and depth-anything constraint. In European Conference on Computer Vision, pages 95–115. Springer, 2025. 2, 7, 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 175, + 554, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 554, + 251 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 554, + 251 + ], + "type": "text", + "content": "[12] Wei-Ting Chen, Hao-Yu Fang, Jian-Jiun Ding, Cheng-Che Tsai, and Sy-Yen Kuo. Jstasr: Joint size and transparency-aware snow removal algorithm based on modified partial convolution and veiling effect removal. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXI 16, pages 754-770. Springer, 2020. 1, 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 254, + 554, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 254, + 554, + 319 + ], + "spans": [ + { + "bbox": [ + 317, + 254, + 554, + 319 + ], + "type": "text", + "content": "[13] Wei-Ting Chen, Hao-Yu Fang, Cheng-Lin Hsieh, Cheng-Che Tsai, I Chen, Jian-Jiun Ding, Sy-Yen Kuo, et al. All snow removed: Single image desnowing algorithm using hierarchical dual-tree complex wavelet representation and contradict channel loss. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4196-4205, 2021. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 320, + 554, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 320, + 554, + 396 + ], + "spans": [ + { + "bbox": [ + 317, + 320, + 554, + 396 + ], + "type": "text", + "content": "[14] Wei-Ting Chen, Hao-Yu Fang, Cheng-Lin Hsieh, Cheng-Che Tsai, I Chen, Jian-Jiun Ding, Sy-Yen Kuo, et al. All snow removed: Single image desnowing algorithm using hierarchical dual-tree complex wavelet representation and contradict channel loss. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4196-4205, 2021. 1, 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 399, + 554, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 399, + 554, + 474 + ], + "spans": [ + { + "bbox": [ + 317, + 399, + 554, + 474 + ], + "type": "text", + "content": "[15] Wei-Ting Chen, Zhi-Kai Huang, Cheng-Che Tsai, Hao-Hsiang Yang, Jian-Jiun Ding, and Sy-Yen Kuo. Learning multiple adverse weather removal via two-stage knowledge learning and multi-contrastive regularization: Toward a unified model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17653–17662, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 477, + 554, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 477, + 554, + 520 + ], + "spans": [ + { + "bbox": [ + 317, + 477, + 554, + 520 + ], + "type": "text", + "content": "[16] Bodong Cheng, Juncheng Li, Ying Chen, Shuyi Zhang, and Tieyong Zeng. Snow mask guided adaptive residual network for image snow removal. arXiv preprint arXiv:2207.04754, 2022. 1, 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 522, + 553, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 522, + 553, + 566 + ], + "spans": [ + { + "bbox": [ + 316, + 522, + 553, + 566 + ], + "type": "text", + "content": "[17] Bodong Cheng, Juncheng Li, Ying Chen, and Tieyong Zeng. Snow mask guided adaptive residual network for image snow removal. Computer Vision and Image Understanding, 236: 103819, 2023. 2, 7, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 568, + 554, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 568, + 554, + 612 + ], + "spans": [ + { + "bbox": [ + 317, + 568, + 554, + 612 + ], + "type": "text", + "content": "[18] Paul F Christiano, Jan Leike, Tom Brown, Miljan Martic, Shane Legg, and Dario Amodei. Deep reinforcement learning from human preferences. Advances in neural information processing systems, 30, 2017. 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 613, + 554, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 613, + 554, + 657 + ], + "spans": [ + { + "bbox": [ + 317, + 613, + 554, + 657 + ], + "type": "text", + "content": "[19] Jinjin Gu, Haoming Cai, Haoyu Chen, Xiaoxing Ye, Jimmy Ren, and Chao Dong. Image quality assessment for perceptual image restoration: A new dataset, benchmark and metric. arXiv preprint arXiv:2011.15002, 2020. 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 658, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 658, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 658, + 554, + 713 + ], + "type": "text", + "content": "[20] Jinjin Gu, Haoming Cai, Haoyu Chen, Xiaoxing Ye, Jimmy Ren, and Chao Dong. Pipal: a large-scale image quality assessment dataset for perceptual image restoration. In European Conference on Computer Vision, pages 633-651. Springer, 2020. 7" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "4310" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 297, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 297, + 127 + ], + "type": "text", + "content": "[21] Shirui Huang, Keyan Wang, Huan Liu, Jun Chen, and Yun-song Li. Contrastive semi-supervised learning for underwater image restoration via reliable bank. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18145-18155, 2023. 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 296, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 296, + 183 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 296, + 183 + ], + "type": "text", + "content": "[22] Yeying Jin, Wending Yan, Wenhan Yang, and Robby T Tan. Structure representation network and uncertainty feedback learning for dense non-uniform fog removal. In Asian Conference on Computer Vision, pages 155-172. Springer, 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 185, + 296, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 185, + 296, + 229 + ], + "spans": [ + { + "bbox": [ + 56, + 185, + 296, + 229 + ], + "type": "text", + "content": "[23] Yeying Jin, Wenhan Yang, Wei Ye, Yuan Yuan, and Robby T Tan. Shadowdiffusion: Diffusion-based shadow removal using classifier-driven attention and structure preservation. arXiv preprint arXiv:2211.08089, 2022." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 231, + 296, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 231, + 296, + 274 + ], + "spans": [ + { + "bbox": [ + 56, + 231, + 296, + 274 + ], + "type": "text", + "content": "[24] Yeying Jin, Beibei Lin, Wending Yan, Wei Ye, Yuan Yuan, and Robby T. Tan. Enhancing visibility in nighttime haze images using guided apsf and gradient adaptive convolution, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 276, + 296, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 276, + 296, + 319 + ], + "spans": [ + { + "bbox": [ + 56, + 276, + 296, + 319 + ], + "type": "text", + "content": "[25] Ruoteng Li, Robby T Tan, and Loong-Fah Cheong. All in one bad weather removal using architectural search. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3175-3185, 2020. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 321, + 296, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 321, + 296, + 376 + ], + "spans": [ + { + "bbox": [ + 56, + 321, + 296, + 376 + ], + "type": "text", + "content": "[26] Yunlong Lin, Zhenqi Fu, Ge Meng, Yingying Wang, Yuhang Dong, Linyu Fan, Hedeng Yu, and Xinghao Ding. Domain-irrelevant feature learning for generalizable pan-sharpening. In Proceedings of the 31st ACM International Conference on Multimedia, pages 3287–3296, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 377, + 296, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 377, + 296, + 432 + ], + "spans": [ + { + "bbox": [ + 56, + 377, + 296, + 432 + ], + "type": "text", + "content": "[27] Yunlong Lin, Zhenqi Fu, Kairun Wen, Tian Ye, Sixiang Chen, Ge Meng, Yingying Wang, Yue Huang, Xiaotong Tu, and Xinghao Ding. Unsupervised low-light image enhancement with lookup tables and diffusion priors. arXiv preprint arXiv:2409.18899, 2024." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 434, + 296, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 434, + 296, + 487 + ], + "spans": [ + { + "bbox": [ + 56, + 434, + 296, + 487 + ], + "type": "text", + "content": "[28] Yunlong Lin, Tian Ye, Sixiang Chen, Zhenqi Fu, Yingying Wang, Wenhao Chai, Zhaohu Xing, Lei Zhu, and Xinghao Ding. Aglldiff: Guiding diffusion models towards unsupervised training-free real-world low-light image enhancement. arXiv preprint arXiv:2407.14900, 2024. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 489, + 296, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 489, + 296, + 543 + ], + "spans": [ + { + "bbox": [ + 56, + 489, + 296, + 543 + ], + "type": "text", + "content": "[29] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26296-26306, 2024. 2, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 544, + 296, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 544, + 296, + 588 + ], + "spans": [ + { + "bbox": [ + 56, + 544, + 296, + 588 + ], + "type": "text", + "content": "[30] Yun-Fu Liu, Da-Wei Jaw, Shih-Chia Huang, and Jenq-Neng Hwang. Desnownet: Context-aware deep network for snow removal. IEEE Transactions on Image Processing, 27(6): 3064–3073, 2018. 2, 6, 7, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 590, + 296, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 590, + 296, + 634 + ], + "spans": [ + { + "bbox": [ + 56, + 590, + 296, + 634 + ], + "type": "text", + "content": "[31] Anish Mittal, Anush Krishna Moorthy, and Alan Conrad Bovik. No-reference image quality assessment in the spatial domain. IEEE Transactions on image processing, 21(12): 4695-4708, 2012. 1, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 635, + 296, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 296, + 669 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 296, + 669 + ], + "type": "text", + "content": "[32] Anish Mittal, Rajiv Soundararajan, and Alan C Bovik. Making a “completely blind” image quality analyzer. IEEE Signal processing letters, 20(3):209-212, 2012. 1, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 670, + 296, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 296, + 714 + ], + "type": "text", + "content": "[33] Ozan Özdenizci and Robert Legenstein. Restoring vision in adverse weather conditions with patch-based denoising diffusion models. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 554, + 714 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 316, + 72, + 554, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 554, + 127 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 554, + 127 + ], + "type": "text", + "content": "[34] Soo-Chang Pei, Yu-Tai Tsai, and Chen-Yu Lee. Removing rain and snow in a single image using saturation and visibility features. In 2014 IEEE International Conference on Multimedia and Expo Workshops (ICMEW), pages 1-6. IEEE, 2014. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 129, + 554, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 554, + 194 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 554, + 194 + ], + "type": "text", + "content": "[35] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 197, + 554, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 197, + 554, + 251 + ], + "spans": [ + { + "bbox": [ + 316, + 197, + 554, + 251 + ], + "type": "text", + "content": "[36] Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. Direct preference optimization: Your language model is secretly a reward model. Advances in Neural Information Processing Systems, 36, 2024. 2, 4, 5, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 253, + 554, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 253, + 554, + 296 + ], + "spans": [ + { + "bbox": [ + 316, + 253, + 554, + 296 + ], + "type": "text", + "content": "[37] John Schulman, Sergey Levine, Pieter Abbeel, Michael Jordan, and Philipp Moritz. Trust region policy optimization. In International conference on machine learning, pages 1889-1897. PMLR, 2015." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 298, + 554, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 298, + 554, + 331 + ], + "spans": [ + { + "bbox": [ + 316, + 298, + 554, + 331 + ], + "type": "text", + "content": "[38] John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347, 2017. 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 332, + 554, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 332, + 554, + 388 + ], + "spans": [ + { + "bbox": [ + 316, + 332, + 554, + 388 + ], + "type": "text", + "content": "[39] Shaolin Su, Qingsen Yan, Yu Zhu, Cheng Zhang, Xin Ge, Jinqiu Sun, and Yanning Zhang. Blindly assess image quality in the wild guided by a self-adaptive hyper network. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3667-3676, 2020. 3, 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 388, + 554, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 388, + 554, + 453 + ], + "spans": [ + { + "bbox": [ + 316, + 388, + 554, + 453 + ], + "type": "text", + "content": "[40] Haoze Sun, Wenbo Li, Jianzhuang Liu, Haoyu Chen, Renjing Pei, Xueyi Zou, Youliang Yan, and Yujiu Yang. Coser: Bridging image and language for cognitive super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25868-25878, 2024. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 456, + 554, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 456, + 554, + 487 + ], + "spans": [ + { + "bbox": [ + 316, + 456, + 554, + 487 + ], + "type": "text", + "content": "[41] Hossein Talebi and Peyman Milanfar. Nima: Neural image assessment. IEEE transactions on image processing, 27(8): 3998-4011, 2018. 3, 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 490, + 554, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 490, + 554, + 555 + ], + "spans": [ + { + "bbox": [ + 316, + 490, + 554, + 555 + ], + "type": "text", + "content": "[42] Jeya Maria Jose Valanarasu, Rajeev Yasarla, and Vishal M Patel. Transweather: Transformer-based restoration of images degraded by adverse weather conditions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2353-2363, 2022. 2, 7, 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 557, + 554, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 557, + 554, + 612 + ], + "spans": [ + { + "bbox": [ + 316, + 557, + 554, + 612 + ], + "type": "text", + "content": "[43] Narasimhan Venkatanath, D Praneeth, Maruthi Chandrasekhar Bh, Sumohana S Channappayya, and Swarup S Medasani. Blind image quality evaluation using perception based features. In 2015 twenty first national conference on communications (NCC), pages 1-6. IEEE, 2015. 1, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 613, + 554, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 613, + 554, + 667 + ], + "spans": [ + { + "bbox": [ + 316, + 613, + 554, + 667 + ], + "type": "text", + "content": "[44] Chenglong Wang, Yang Gan, Yifu Huo, Yongyu Mu, Murun Yang, Qiaozhi He, Tong Xiao, Chunliang Zhang, Tongran Liu, Quan Du, Di Yang, and Jingbo Zhu. Rovrm: A robust visual reward model optimized via auxiliary textual preference data, 2024. 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 670, + 554, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 554, + 714 + ], + "type": "text", + "content": "[45] Jianyi Wang, Kelvin CK Chan, and Chen Change Loy. Exploring clip for assessing the look and feel of images. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2555-2563, 2023. 3, 4" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "text", + "content": "4311" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 714 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "type": "text", + "content": "[46] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 173 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 173 + ], + "type": "text", + "content": "[47] Haoning Wu, Zicheng Zhang, Weixia Zhang, Chaofeng Chen, Liang Liao, Chunyi Li, Yixuan Gao, Annan Wang, Erli Zhang, Wenxiu Sun, et al. Q-align: Teaching Imms for visual scoring via discrete text-defined levels. arXiv preprint arXiv:2312.17090, 2023. 1, 3, 4, 6, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 175, + 294, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 294, + 239 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 294, + 239 + ], + "type": "text", + "content": "[48] Hongtao Wu, Yijun Yang, Angelica I Aviles-Rivero, Jingjing Ren, Sixiang Chen, Haoyu Chen, and Lei Zhu. Semi-supervised video desnowing network via temporal decoupling experts and distribution-driven contrastive regularization. In European Conference on Computer Vision, pages 70-89. Springer, 2024. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 242, + 294, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 242, + 294, + 297 + ], + "spans": [ + { + "bbox": [ + 56, + 242, + 294, + 297 + ], + "type": "text", + "content": "[49] Hongtao Wu, Yijun Yang, Huihui Xu, Weiming Wang, Jinni Zhou, and Lei Zhu. Rainmamba: Enhanced locality learning with state space models for video deraining. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 7881-7890, 2024. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 298, + 294, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 298, + 294, + 374 + ], + "spans": [ + { + "bbox": [ + 56, + 298, + 294, + 374 + ], + "type": "text", + "content": "[50] Haoning Wu, Zicheng Zhang, Erli Zhang, Chaofeng Chen, Liang Liao, Annan Wang, Kaixin Xu, Chunyi Li, Jingwen Hou, Guangtao Zhai, et al. Q-instruct: Improving low-level visual abilities for multi-modality foundation models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25490-25500, 2024. 1, 2, 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 376, + 294, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 376, + 294, + 430 + ], + "spans": [ + { + "bbox": [ + 56, + 376, + 294, + 430 + ], + "type": "text", + "content": "[51] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. Advances in Neural Information Processing Systems, 34, 2021. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 433, + 294, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 433, + 294, + 487 + ], + "spans": [ + { + "bbox": [ + 56, + 433, + 294, + 487 + ], + "type": "text", + "content": "[52] Jiaqi Xu, Mengyang Wu, Xiaowei Hu, Chi-Wing Fu, Qi Dou, and Pheng-Ann Heng. Towards real-world adverse weather image restoration: Enhancing clearness and semantics with vision-language models. In European Conference on Computer Vision, pages 147–164. Springer, 2025. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 488, + 294, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 488, + 294, + 533 + ], + "spans": [ + { + "bbox": [ + 56, + 488, + 294, + 533 + ], + "type": "text", + "content": "[53] Yuan Yao, Tianyu Yu, Ao Zhang, Chongyi Wang, Junbo Cui, Hongji Zhu, Tianchi Cai, Haoyu Li, Weilin Zhao, Zhihui He, et al. Minicpm-v: A gpt-4v level mllm on your phone. arXiv preprint arXiv:2408.01800, 2024. 2, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 534, + 294, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 534, + 294, + 578 + ], + "spans": [ + { + "bbox": [ + 56, + 534, + 294, + 578 + ], + "type": "text", + "content": "[54] Tian Ye, Yunchen Zhang, Mingchao Jiang, Liang Chen, Yun Liu, Sixiang Chen, and Erkang Chen. Perceiving and modeling density for image dehazing. In European Conference on Computer Vision, pages 130-145. Springer, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 579, + 294, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 579, + 294, + 633 + ], + "spans": [ + { + "bbox": [ + 56, + 579, + 294, + 633 + ], + "type": "text", + "content": "[55] Tian Ye, Sixiang Chen, Jinbin Bai, Jun Shi, Chenghao Xue, Jingxia Jiang, Junjie Yin, Erkang Chen, and Yun Liu. Adverse weather removal with codebook priors. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12653-12664, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 635, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 294, + 689 + ], + "type": "text", + "content": "[56] Tian Ye, Sixiang Chen, Wenhao Chai, Zhaohu Xing, Jing Qin, Ge Lin, and Lei Zhu. Learning diffusion texture priors for image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2524-2534, 2024. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 691, + 294, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 714 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 714 + ], + "type": "text", + "content": "[57] Zhenqiang Ying, Haoran Niu, Praful Gupta, Dhruv Mahajan, Deepti Ghadiyaram, and Alan Bovik. From patches to" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 329 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 116 + ], + "type": "text", + "content": "pictures (paq-2-piq): Mapping the perceptual space of picture quality. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3575-3585, 2020. 1, 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 118, + 553, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 553, + 162 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 553, + 162 + ], + "type": "text", + "content": "[58] Zhiyuan You, Zheyuan Li, Jinjin Gu, Zhenfei Yin, Tianfan Xue, and Chao Dong. Depicting beyond scores: Advancing image quality assessment through multi-modal language models. arXiv preprint arXiv:2312.08962, 2023. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 163, + 553, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 553, + 206 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 553, + 206 + ], + "type": "text", + "content": "[59] Kaihao Zhang, Rongqing Li, Yanjiang Yu, Wenhan Luo, and Changsheng Li. Deep dense multi-scale network for snow removal using semantic and depth priors. IEEE Transactions on Image Processing, 30:7419-7431, 2021. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 208, + 553, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 208, + 553, + 271 + ], + "spans": [ + { + "bbox": [ + 316, + 208, + 553, + 271 + ], + "type": "text", + "content": "[60] Weixia Zhang, Guangtao Zhai, Ying Wei, Xiaokang Yang, and Kede Ma. Blind image quality assessment via vision-language correspondence: A multitask learning perspective. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14071-14081, 2023. 3, 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 274, + 553, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 274, + 553, + 329 + ], + "spans": [ + { + "bbox": [ + 316, + 274, + 553, + 329 + ], + "type": "text", + "content": "[61] Haipeng Zhou, Hongqiu Wang, Tian Ye, Zhaohu Xing, Jun Ma, Ping Li, Qiong Wang, and Lei Zhu. Timeline and boundary guided diffusion network for video shadow detection. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 166-175, 2024. 1" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4312" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/SoMA_ Singular Value Decomposed Minor Components Adaptation for Domain Generalizable Representation Learning/282504c5-c667-4202-94ca-c92386507635_content_list.json b/2025/SoMA_ Singular Value Decomposed Minor Components Adaptation for Domain Generalizable Representation Learning/282504c5-c667-4202-94ca-c92386507635_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..8e5fc2e40bc6adb0823d18d5905b06bfcb5dd7b8 --- /dev/null +++ b/2025/SoMA_ Singular Value Decomposed Minor Components Adaptation for Domain Generalizable Representation Learning/282504c5-c667-4202-94ca-c92386507635_content_list.json @@ -0,0 +1,1617 @@ +[ + { + "type": "text", + "text": "SoMA: Singular Value Decomposed Minor Components Adaptation for Domain Generalizable Representation Learning", + "text_level": 1, + "bbox": [ + 155, + 83, + 841, + 128 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Seokju Yun Seunghye Chae Dongheon Lee Youngmin Ro* Machine Intelligence Laboratory, University of Seoul, Korea", + "bbox": [ + 214, + 156, + 781, + 191 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{wsz871, tmdhey, dslisleedh, youngmin.ro}@uos.ac.kr \nhttps://github.com/ysj9909/SoMA", + "bbox": [ + 272, + 195, + 723, + 226 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3628e235ba0f324223b527313dafe6b96b4369108834212fbc94f57ac5ccc19f.jpg", + "image_caption": [ + "Figure 1. Overview of SoMA framework. Left. SoMA achieves state-of-the-art results across diverse tasks, ranging from domaingeneralized semantic segmentation (DGSS) to object detection (DGOD), and performs well in both synthetic-to-real and real-to-real scenarios. Middle. Our method, trained solely on synthetic datasets, demonstrates strong generalization capabilities in complex real-world scenes. Right. We present SoMA, a method that applies SVD to the pre-trained weights, decomposing them into $r$ minor singular components $U_{[:,-r:]} \\Sigma_{[-r:]} (V^T)_{[-r:,]}$ and residuals. SoMA selectively tunes the smallest $r$ components, effectively preserving world knowledge of foundation models. Since SoMA shares the same forward architecture as LoRA [30], it adds no extra latency during inference phase." + ], + "image_footnote": [], + "bbox": [ + 94, + 236, + 377, + 467 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/49983582969fbd7f5d57332555baa5e2f0fd28824735a4ba3bba883fe850aabb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 383, + 238, + 571, + 311 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c40ea8d126a3176fbedf8dc89cc74e2b9af90901a70fd337c29b9cec168109db.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 383, + 315, + 571, + 388 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5d4165514b4bb3db4a6d890ec65f07d65fa688179d0a15423ea0523abd4ce689.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 383, + 393, + 571, + 465 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/794488922f066aff2e1e6557ceeb047b3362fbbdc3075f5cbf3e8df5bbe3a009.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 584, + 237, + 903, + 465 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 568, + 326, + 584 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Domain generalization (DG) aims to adapt a model using one or multiple source domains to ensure robust performance in unseen target domains. Recently, Parameter-Efficient Fine-Tuning (PEFT) of foundation models has shown promising results in the context of DG problem. Nevertheless, existing PEFT methods still struggle to strike a balance between preserving generalizable components of the pre-trained model and learning task-specific features. To gain insights into the distribution of generalizable components, we begin by analyzing the pre-trained weights through the lens of singular value decomposition. Building on these insights, we introduce Singular Value Decomposed Minor Components Adaptation (SoMA), an approach that selectively tunes minor singular components while keeping the residual parts frozen. SoMA effectively retains the generalization ability of the pre-trained model while efficiently acquiring task-specific skills. Moreover, we freeze domain-generalizable blocks and employ an annealing weight decay", + "bbox": [ + 88, + 601, + 485, + 875 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "strategy, thereby achieving an optimal balance in the delicate trade-off between generalizability and discriminability. SoMA attains state-of-the-art results on multiple benchmarks that span both domain generalized semantic segmentation to domain generalized object detection. In addition, our methods introduce no additional inference overhead or regularization loss, maintain compatibility with any backbone or head, and are designed to be versatile, allowing easy integration into a wide range of tasks.", + "bbox": [ + 511, + 570, + 906, + 705 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 712, + 645, + 727 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Deep neural networks (DNNs) have recently made significant strides in dense prediction tasks, including semantic segmentation [12] and object detection [7, 61], both of which are crucial for safety-critical applications, such as autonomous driving. However, DNNs often fail to maintain reliable performance under domain shifts, which may result from diverse lighting conditions, weather changes, or differences in location. Likewise, while synthetic data [24, 62, 64] is employed to avoid the labor-intensive and costly process of constructing real-world datasets, the resulting performance degradation of DNNs in real-world deployment", + "bbox": [ + 511, + 734, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation.", + "bbox": [ + 236, + 0, + 810, + 18 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "Except for this watermark, it is identical to the accepted version;", + "bbox": [ + 323, + 16, + 723, + 30 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 292, + 31, + 754, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author.", + "bbox": [ + 109, + 887, + 236, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "25602", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "remains a critical issue to be resolved. To tackle these challenges, Domain Generalization (DG) is introduced to design models capable of consistent prediction on arbitrary unseen target domains.", + "bbox": [ + 89, + 90, + 480, + 151 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Existing DG methods for dense prediction tasks employ two main strategies to enhance model robustness. The first involves diversifying the training process in the input [10, 57] or feature space [21, 70] using augmentation/perturbation techniques, exposing the model to a broader range of styles to mitigate overfitting to specific domains. The second strategy focuses on imposing alignment constraints, such as normalization [58] or regularization losses [1, 13], to facilitate the learning of domain-invariant features. However, these methods mostly utilize outdated backbones (e.g., ResNet [27] or MobileNetV2 [67]) pretrained on mid-scale datasets, such as ImageNet [16].", + "bbox": [ + 88, + 152, + 482, + 332 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Meanwhile, vision foundation models (VFMs) [2], pretrained on vast, curated datasets with enormous parameter counts, have recently shown unprecedented generalization capabilities [74]. Given this landscape, we propose a paradigm shift—from applying the diversify & align approach to classic backbones, to devising methods that preserve the world knowledge of VFMs while effectively learning task-specific features. However, naive adaptation of VFMs via full fine-tuning (FFT), which involves retraining all model parameters, results in prohibitive costs and a risk of catastrophic forgetting of the pre-trained knowledge.", + "bbox": [ + 89, + 334, + 482, + 500 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In light of these challenges, a commonly adopted approach is parameter-efficient fine-tuning (PEFT), where only the injected lightweight adapters [11, 25, 30] or tokens [34, 45] are fine-tuned, while the pre-trained weights remain frozen. Low-Rank Adaptation (LoRA) [30] is one of the most widely used PEFT methods, which substitutes the updates with the product of two low-rank matrices. Interestingly, we found that LoRA exhibits superior generalization ability in DG dense prediction tasks compared to FFT and other PEFT methods (see Tab. 2), aligning with recent findings in LoRA-based literature [6, 31]. Nevertheless, LoRA adapters are built without considering the distribution of generalizable components within the pre-trained weights they modify. As a result, there is still room for enhancing generalizability by preserving the diverse domain knowledge of VFMs.", + "bbox": [ + 89, + 501, + 482, + 741 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To uncover the structure in which world knowledge is encoded within the pre-trained weights, we first perform singular value decomposition (SVD) on the learned weight matrices and then analyze the behavior of VFMs after selectively removing specific singular components1. Our analysis reveals that singular vectors associated with higher singular values tend to extract general features spanning multiple classes in ImageNet, while minor singular com", + "bbox": [ + 89, + 742, + 482, + 863 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ponents are primarily responsible for capturing context-specific features. Capitalizing on our findings, we introduce Singular Value Decomposed Minor Components Adaptation (SoMA), which starts by performing SVD to decompose the weights into their singular vectors, i.e., $W = U\\Sigma V^T$ , then fine-tunes only the minor singular components. Specifically, we initialize LoRA adapter using the components with the smallest $r$ singular values, i.e., $U_{[:,-r:]} \\Sigma_{[-r:]}(V^T)_{[-r:,]}$ , and the remaining residual components are frozen to maintain generalizability, as illustrated in Fig. 1 right.", + "bbox": [ + 511, + 90, + 903, + 257 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "By extending our analysis from the weight-level to the block-level, we observe and empirically demonstrate that the early blocks of VFMs effectively extract well-localized semantic features and are less affected by style or input-level domain shifts (see Fig. 3). Therefore, we freeze these blocks to preserve generalizable components while reducing the number of trainable parameters. Although our proposed methods are effective, tuning the less-optimized bottom spectral space and freezing early blocks may lead to a lack of discriminability. To mitigate this issue, we introduce an annealing weight decay scheme that gradually reduces the regularization loss incurred by weight decay over the course of training. As shown in Fig. 1, comprehensive evaluations on DG benchmarks demonstrate that our framework outperforms previous state-of-the-art baselines and FFT, training only $0.58\\%$ to $1.6\\%$ of the parameters. Furthermore, SoMA demonstrates impressive scaling performance with both data and model size.", + "bbox": [ + 511, + 257, + 903, + 526 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We make the following contributions:", + "bbox": [ + 513, + 529, + 764, + 542 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We underscore the critical importance of preserving pretrained knowledge while learning task-specific features in scenarios that involve the use of VFMs.", + "- We present SoMA, a novel PEFT method specifically tailored for DG problem, which leverages spectral information to retain the generalization capacity of VFMs. In addition, we identify and freeze domain-generalizable blocks while employing an annealing weight decay strategy, thereby achieving an optimal balance between generalizability and discriminability.", + "- Extensive experiments on various DG for semantic segmentation and object detection tasks reveal that our framework surpasses existing baselines by a substantial margin. Notably, SoMA achieves an mIoU of $80.4\\%$ on Cityscapes without accessing real-world datasets, while training only $0.58\\%$ of the parameters. Additionally, under the most challenging daytime-sunny $\\rightarrow$ night-rainy detection setting, SoMA outperforms the prior SOTA implementation by achieving a $+23.5\\%$ mAP gain." + ], + "bbox": [ + 514, + 547, + 903, + 838 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 845, + 653, + 861 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Domain generalization for dense predictions. Domain Generalization (DG) methods for dense predictions [5, 9,", + "bbox": [ + 511, + 869, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "In this paper, we refer to the combination of a singular value and its corresponding singular vectors as a singular component.", + "bbox": [ + 89, + 875, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "25603", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "17, 32, 36, 39, 42, 46, 56, 68, 73, 75, 78] have recently garnered considerable attention due to their practical demands. These methods can be categorized along two main axes: i) data augmentation or domain randomization, both of which diversify the training process, and ii) alignment techniques to suppress domain-relevant features.", + "bbox": [ + 89, + 90, + 480, + 181 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Methods adopting the first approach at the input level employ various augmentation techniques [10, 38, 57, 84] or, more recently, leverage advances in generative modeling, such as diffusion models [59, 63], to generate data approximating the target domain [3, 35, 52], thereby expanding domain diversity and enriching learned representations. To enable more flexible augmentation, several studies [20, 21, 41, 70, 72] leverage feature stylization to enhance model robustness. On the other hand, alignment-based methods effectively reduce domain-sensitive components by employing feature normalization/whitening techniques [55, 58, 69], or by introducing regularization losses to suppress inconsistencies caused by simulated domain shifts [1, 13, 15, 37, 40, 43, 49, 76, 77, 82, 83]. However, most of the aforementioned approaches experimented with CNN backbones [27] trained on relatively restricted domains. In contrast, when applying vision foundation models (VFMs)—trained on massive datasets using transformer architectures [18, 71] and sophisticated training schemes [53]—to the DG problem, a fundamental shift in approach is imperative.", + "bbox": [ + 91, + 185, + 483, + 501 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Following this trend, recent studies leverage vision-language models as feature extractors [33], employing text embeddings with domain-invariant semantics as a source for style augmentation [20, 72] or as object queries in transformer-based decoders [54]. Rein [74] injects learnable adapters and tokens to refine feature maps for each instance while keeping the backbone frozen, significantly expanding the performance frontier. FADA [4] and SET [79] introduce frequency-adapted methods to effectively utilize frozen VFM features. In contrast to existing methods that simply freeze all parameters, we perform singular value decomposition on the pre-trained weights, freezing the generalizable principal components and tuning only the minor singular components responsible for context-specific features, thereby achieving superior DG performance.", + "bbox": [ + 89, + 505, + 482, + 731 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Vision foundation models (VFMs). Recently, the advent of VFMs with general perception capabilities has laid the foundation for a paradigm in which these models are broadly applicable to numerous downstream tasks. Seminal examples of models advancing this direction include CLIP [60], which demonstrates strong zero-shot generalizability through training on web-scale, weakly supervised image-text pairs; EVA02 [22], which enhances CLIP features with masked image modeling [28]; and DINOv2 [53], which incorporates losses from prior arts like iBoT [85] and DINO [8], and is trained on massive curated datasets, offer", + "bbox": [ + 89, + 734, + 482, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ing strong spatial features for dense prediction tasks. Our SoMA builds on this foundation, harnessing the full potential of VFMs in the context of DG problem.", + "bbox": [ + 511, + 90, + 903, + 136 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Parameter-efficient fine-tuning (PEFT). With the remarkable success of PEFT methods in efficiently adapting large-scale foundation models in the realm of natural language processing, there has been a growing interest in extending these approaches to the field of computer vision. For example, VPT [34] prepends prompt tokens to the input sequence of several attention layers. AdaptFormer [11] introduces a lightweight adapter in parallel to the feed-forward network, whereas SSF [47] incorporates scale and shift parameters to modulate features for downstream tasks. LoRA [30] models the incremental updates of pre-trained weights using rank decomposition matrices, achieving performance comparable to full fine-tuning (FFT). However, there still often exists a performance gap between LoRA and FFT. DoRA [48] and PiSSA [50] address this gap by further decomposing the pre-trained weights, initializing the LoRA adapter with directional and principal components, respectively, and fine-tuning accordingly. While these methods aim to match the performance of FFT, our focus is on the preservation of generalizable components.", + "bbox": [ + 511, + 137, + 906, + 439 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Preliminaries", + "text_level": 1, + "bbox": [ + 513, + 453, + 651, + 469 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Low-Rank Adaptation (LoRA). LoRA [30] hypothesizes that the weight changes during fine-tuning exhibit a low-rank structure. Given a pre-trained weight matrix $W_0 \\in \\mathbb{R}^{m \\times n}$ , LoRA constrains its update $\\Delta W \\in \\mathbb{R}^{m \\times n}$ to a low-rank decomposition $\\Delta W = BA$ , where $B \\in \\mathbb{R}^{m \\times r}$ and $A \\in \\mathbb{R}^{r \\times n}$ are two low-rank matrices, with an intrinsic rank of $r \\ll \\min(m, n)$ . For $y = W_0 x$ , the modified forward pass is as follows:", + "bbox": [ + 511, + 477, + 905, + 597 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\ny = \\left(W _ {0} + \\Delta W\\right) x = \\left(W _ {0} + \\underline {{B A}}\\right) x, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 584, + 604, + 903, + 622 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $W_{0}$ is frozen, and only the underlined low-rank parameters are fine-tuned, significantly reducing the number of trainable parameters. LoRA uses a uniform Kaiming distribution [26] to initialize $A$ , while $B$ is initially set to zero, ensuring $BA = 0$ at the beginning of training. As seen in Eq. 1, the learned matrices $BA$ can be merged with the frozen weight $W_{0}$ , allowing LoRA-based variants to avoid adding any additional inference burden.", + "bbox": [ + 511, + 628, + 905, + 748 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Singular Value Decomposition (SVD) Analysis. We observe that LoRA and existing PEFT methods, by training only a relatively small subset of parameters or by enforcing low-rank-ness of weight updates, deliver promising DG performance without significantly distorting the pre-trained representations. However, many previous studies build adapters without considering how the generalizable components are structured within the weight matrices, potentially interfering with these components in the fine-tuning process. Building on the Eckart-Young-Mirsky the", + "bbox": [ + 511, + 750, + 906, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "25604", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/e96b2d5afd82c26f5ee4cb84a1541c58f76df86f4626494ad4971d542f8dd724.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 130, + 78, + 434, + 218 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/59d361949c111fda71f05b4d52a8a3a70a849e82e159392981be01f3e1a906d0.jpg", + "image_caption": [ + "Figure 2. Distribution of generalizable components. Top. Number of classes exhibiting specific accuracy drops after applying SVD to DINOv2-large weights and reconstructing by truncating the smallest $r$ singular components. Bottom. Distinct roles of singular components across levels. Numbers in parentheses represent the count of classes with an accuracy drop ratio exceeding $50\\%$ . The average WordNet hierarchy depth of these classes is shown, with higher values indicating greater context specificity." + ], + "image_footnote": [], + "bbox": [ + 127, + 220, + 439, + 316 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "orem [19], which validates that the optimal rank- $r$ approximation of a matrix $W$ is represented by the sum of its top $r$ singular components, we analyze how the world knowledge of VFMs is structured from an SVD perspective. Using ImageNet-1k [16] as a testbed, our analysis computes the SVD of pre-trained weights across all layers of the DINOv2 [53] ViT-large model, followed by truncating specific singular values and their corresponding singular vectors to investigate the classes that are subsequently misclassified.", + "bbox": [ + 89, + 441, + 483, + 578 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As shown in Fig. 2 top, as the number of discarded smallest singular components increases, the number of misclassified classes grows exponentially rather than linearly, suggesting that components with larger singular values tend to capture more general features spanning multiple classes. To further examine whether different groups of singular components exhibit distinct behaviors, we consider three groups: the top 8 principal components $U_{[:;8]}\\Sigma_{[:8]}(V^T)_{[:8,:]}$ , the middle 160 singular components $U_{[:384:544]}\\Sigma_{[384:544]}(V^T)_{[384:544,:]}$ , and the bottom 320 minor components $U_{[:,-320:]} \\Sigma_{[-320:]}(V^T)_{[-320,:]}$ . We then truncate each group separately and analyze the classes that the rank-reduced model subsequently fails to classify (see Fig. 2 bottom).", + "bbox": [ + 89, + 580, + 483, + 792 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "According to the Eckart-Young-Mirsky theorem and the heavy-tail distribution of singular values, it is evident that the top singular components encapsulate the core knowledge of VFM and should be preserved to maintain generalizability. Conversely, we note that the bottom-componenttruncated model fails to recognize classes that require finegrained or texture-biased features (e.g., specific dog breeds,", + "bbox": [ + 89, + 795, + 483, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "wildlife with distinctive textures), as well as those that are context-specific (e.g., cornet, bikini, missile, dock). In contrast, we find that removing middle singular components results in significant performance declines for classes that rely on coarse-grained or shape-biased recognition (e.g., mortarboard, pier, altar, necklace, schooner, horizontal bar). Interestingly, as suggested by differences in the depth of the WordNet hierarchy, higher-value singular components do not function independently but rather interact in a hierarchical and composite manner. For instance, classes with notable accuracy drops can also be clustered into higher-level concepts, such as functionality (e.g., wok, hotpot, pitcher, cauldron, pop bottle, cup/ambulance, minivan, passenger car, police van, garbage truck, limousine) or co-occurrence (e.g., desktop computer, keyboard, notebook, desk). Additionally, almost all of these objects are commonly found in everyday scenes. Consequently, it can be seen that generalizable components are hierarchically and compositionally structured according to the magnitude of singular values.", + "bbox": [ + 511, + 90, + 903, + 378 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Problem Formulation. The entire model consists of a feature extractor $\\mathcal{F}$ , which combines a VFM parameterized by $\\Phi$ with our SoMA adapter parameterized by $\\theta_{S}$ , and a task-specific decode head $\\mathcal{H}$ , parameterized by $\\theta_{h}$ . The optimization objective can be expressed as follows:", + "bbox": [ + 511, + 378, + 903, + 454 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\underset {\\theta_ {S}, \\theta_ {h}} {\\arg \\min } \\sum_ {i = 1} ^ {N _ {s}} \\mathcal {L} _ {t a s k} \\left(\\mathcal {H} _ {\\theta_ {h}} \\left(\\mathcal {F} _ {\\Phi , \\theta_ {S}} \\left(x _ {i}\\right)\\right), y _ {i}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 573, + 460, + 903, + 501 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $(x_{i},y_{i})$ denotes an input image and its associated ground truth, drawn from the source domain datasets $\\mathcal{D}_s$ , comprising $N_{s}$ samples. Our proposed SoMA framework facilitates generalizing task-specific skills learned from $\\mathcal{D}_s$ to the unseen target domain datasets $\\mathcal{D}_t$ .", + "bbox": [ + 511, + 508, + 906, + 584 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Proposed Method", + "text_level": 1, + "bbox": [ + 511, + 598, + 687, + 614 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we present our parameter-efficient adaptation methods specifically tailored for DG dense prediction tasks, inspired by the insights from our SVD analysis. Subsequently, we propose a block freeze strategy and an annealing weight decay scheme to strike an optimal trade-off between generalizability and discriminability.", + "bbox": [ + 511, + 619, + 906, + 710 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Minor Singular Components Adaptation", + "text_level": 1, + "bbox": [ + 511, + 715, + 862, + 733 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To effectively preserve the integrity of generalizable representations in pre-trained weights, we propose Singular value decomposed Minor components Adaptation (SoMA). Firstly, SoMA performs singular value decomposition (SVD) on the pre-trained weight matrices, including those from the attention layers, multi-layer perceptron (MLP) layers, and more. The weight matrix $W \\in \\mathbb{R}^{m \\times n}$ is decomposed using SVD as follows:", + "bbox": [ + 511, + 738, + 905, + 859 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nW = U \\Sigma V ^ {T} = \\sum_ {i = 1} ^ {R} \\sigma_ {i} u _ {i} v _ {i} ^ {T}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 612, + 863, + 906, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "25605", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/3ecc3b511cc76cad4eae19078ee87e110856076e5eaf3e9ad5efb87bb8f39f3d.jpg", + "image_caption": [ + "Figure 3. The inherent generalization capabilities of the early blocks of VFM. Top. We apply PCA on the extracted intermediate features (8th block of DINOv2-Large) and visualize the top three leading components. Bottom. Class-wise IoU comparison for rare classes [29] under the GTAV $\\rightarrow$ Cityscapes setting, focusing on the impact of freezing the first eight blocks. \"NFEB\" stands for the Number of Frozen Early Blocks. A darker shade of blue indicates enhanced generalization performance relative to the baseline." + ], + "image_footnote": [], + "bbox": [ + 106, + 82, + 470, + 276 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $U = [u_{1}, u_{2}, \\dots, u_{m}] \\in \\mathbb{R}^{m \\times m}$ and $V = [v_{1}, v_{2}, \\dots, v_{n}] \\in \\mathbb{R}^{n \\times n}$ are the left and right singular vectors, whose columns form an orthonormal basis for $\\mathbb{R}^m$ and $\\mathbb{R}^n$ respectively, $\\Sigma \\in \\mathbb{R}^{m \\times n}$ is a diagonal matrix whose entries $\\sigma_{i}$ are singular values arranged in descending order, and $R$ denotes the rank of $W$ , with $R \\leq \\min(m, n)$ .", + "bbox": [ + 89, + 398, + 483, + 489 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We then divide the singular components into two groups according to their singular values: the minor singular components with the smallest $r$ singular values $U_{[:,-r:]} \\Sigma_{[-r:]} (V^T)_{[-r:]}$ and the residual singular components $U_{[:,-r]} \\Sigma_{[:-r]} (V^T)_{[:-r:]}$ . As illustrated in Fig. 1 right, we employ the QR-type reconstruction of minor singular components to initialize the adapter for fine-tuning:", + "bbox": [ + 89, + 489, + 483, + 595 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nB = U _ {[:, - r: ]} \\sqrt {\\Sigma} _ {[ - r: ]} \\in \\mathbb {R} ^ {m \\times r}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 173, + 595, + 480, + 616 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nA = \\sqrt {\\Sigma} _ {[ - r: ]} \\left(V ^ {T}\\right) _ {[ - r:,.. ]} \\in \\mathbb {R} ^ {r \\times n}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 174, + 618, + 480, + 638 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the fine-tuning process, the frozen residual matrix $W_{res}$ is defined as $W - BA$ to prevent numerical errors introduced during the SVD step, while maintaining the pretrained generalization capability at the beginning of fine-tuning. After fine-tuning, the learned low-rank matrices $B'$ and $A'$ can be merged into $W_{res}$ , as in LoRA [30], ensuring no additional computation overhead during inference:", + "bbox": [ + 89, + 646, + 483, + 752 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\ny = \\left(W \\underbrace {- B A + B ^ {\\prime} A ^ {\\prime}} _ {\\Delta W _ {S o M A}}\\right) x = \\left(W _ {r e s} + B ^ {\\prime} A ^ {\\prime}\\right) x = W ^ {\\prime} x. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 98, + 757, + 482, + 792 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Importantly, SoMA offers several key advantages for tackling the DG problem: i) SoMA initializes the adapter using components orthogonal to the principal components, thus minimizing interference with pre-trained representations. For the batched version of the forward pass described in Eq. 1 (with $X \\in \\mathbb{R}^{n \\times b}$ and $Y \\in \\mathbb{R}^{m \\times b}$ ), the gradients of $B$ and $A$ are $\\frac{\\partial \\mathcal{L}}{\\partial B} = \\frac{\\partial \\mathcal{L}}{\\partial Y} X^T A^T$ and $\\frac{\\partial \\mathcal{L}}{\\partial A} = B^T \\frac{\\partial \\mathcal{L}}{\\partial Y} X^T$ .", + "bbox": [ + 89, + 795, + 483, + 902 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/c63116b68fd7fe6fbc56be6565798ab94e0fcb9c16788a54a09e563a44ca7326.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
SMRBlockiΔWSoMAΔW*SoMAΔWLoRA
|uiTΔVvi/σi12th0-2550.0750.084 (↓17%)0.101
256-5110.0940.094 (↓43%)0.166
512-7670.1570.165 (↓40%)0.275
768-10234.1046.152 (↑21%)5.095
24th0-2550.0970.095 (↓24%)0.125
256-5110.1050.108 (↓35%)0.166
512-7670.2060.189 (↓41%)0.323
768-10232.7813.620 (↑26%)2.865
", + "bbox": [ + 519, + 75, + 903, + 200 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1. Singular modulation ratio comparison. * indicates the adaptation matrix trained using the annealing weight decay strategy. The weight matrices are taken from the query projection layers of the 12th and 24th attention blocks in DINOv2-Large [53].", + "bbox": [ + 511, + 204, + 906, + 260 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Therefore, unlike LoRA, which randomly initializes $A$ , our approach facilitates convergence along singular directions that preserve the integrity of the generalizable components. ii) Consistent with prior DG methods [13, 40, 41, 84] that guide models to prioritize shape over texture, thereby achieving superior out-of-distribution performance, SoMA effectively preserves the domain generalizable shape-biased features of VFMs by tuning only the minor singular components (see Fig. 2 bottom and Tab. 1).", + "bbox": [ + 511, + 268, + 906, + 404 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Discussion. To assess whether the proposed adaptation matrix $\\Delta W_{SoMA}$ minimally correlates with the top singular directions of the pre-trained weight matrix $W_0$ , we project $\\Delta W$ onto the singular vectors of $W_0$ . Specifically, we quantify the extent of interference by $\\Delta W$ on each $i$ -th singular direction through a singular modulation ratio (SMR), which is formally defined as follows:", + "bbox": [ + 511, + 405, + 905, + 511 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {S M R} _ {i} = \\left| \\frac {u _ {i} ^ {\\mathrm {T}} \\Delta W v _ {i}}{\\sigma_ {i}} \\right|. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 633, + 515, + 903, + 547 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For ease of analysis, we partition the 1024 singular directions into four equal groups, calculate the average SMR for each group, and then compare the resulting values between SoMA and LoRA. As shown in Tab. 1, the SoMA initialization scheme proves more effective than LoRA in minimizing interference with singular components with higher singular values. In summary, our method effectively preserves the structure of the VFM's generalizable representations while adapting to downstream tasks, resulting in significant improvements on DG benchmarks compared to SOTA baselines, as confirmed by our experimental results.", + "bbox": [ + 511, + 551, + 906, + 719 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Freezing Early Blocks", + "text_level": 1, + "bbox": [ + 511, + 729, + 718, + 744 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we explore the distribution of generalizable components at the block level. Drawing on recent domain adaptation methods [44, 81], which find that the early blocks of a model primarily engage in domain-specific feature extraction and should therefore be updated, we conjecture that in the context of leveraging VFM for DG, the early blocks should instead remain unchanged. Intuitively, these initial blocks of VFMs are adept at projecting low-level features from diverse domains into domain-invariant semantic representations. In other words, while adjusting", + "bbox": [ + 511, + 750, + 906, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "25606", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "early blocks may help bridge input-level domain gaps in domain adaptation, freezing the early blocks of VFMs is essential for DG to prevent overfitting to the source domain and to retain the model's generalization capability.", + "bbox": [ + 89, + 90, + 480, + 151 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To validate our hypothesis, we perform PCA analysis on the features extracted from the early blocks of VFM and compare the performance on rare classes when these blocks are frozen, as illustrated in Fig. 3. Despite considerable domain shifts in style and content, the principal components align well with the image layout and maintain consistent semantic information (e.g., the horses have similar colors). Furthermore, simply freezing the early blocks yields a notable improvement in IoU for rare classes, indicating that updating these blocks can severely disrupt the learned representations of various classes in VFMs. Thus, by default, we opt to freeze all blocks up to the one that generates the feature map fed as the highest resolution input to the segmentation/detection head (e.g., the first eight blocks). Detailed ablations of varying the number of frozen early blocks are presented in the Supplemental.", + "bbox": [ + 89, + 152, + 482, + 393 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Annealing Weight Decay", + "text_level": 1, + "bbox": [ + 89, + 402, + 316, + 419 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our methods are specialized in preserving the pre-trained knowledge of VFMs, but this may come at the cost of diminished discriminability. One can remove weight decay or set its coefficient to a very small value to enhance discriminability. However, doing so risks increased interference of the tuned parameters with the principal components of the pre-trained weights. Therefore, based on the observation that regularization primarily affects the early phases of learning by guiding a model towards having good generalization properties [23], we leverage an Annealing Weight Decay (AWD) scheme. Specifically, AWD starts with a relatively large weight decay coefficient and progressively reduces it to zero as training progresses. By default, we use a cosine schedule for AWD. As shown in Tab. 1 ( $\\Delta W_{SoMA}$ vs. $\\Delta W_{SoMA}^{*}$ ) and Tab. 7, AWD integrates seamlessly with SoMA, improving discriminability without compromising SoMA's ability to preserve pre-trained knowledge.", + "bbox": [ + 89, + 425, + 482, + 683 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 89, + 696, + 222, + 713 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We conduct a variety of experiments to showcase the efficacy of SoMA on DG benchmarks including domain generalized semantic segmentation (DGSS) and domain generalized object detection (DGOD). We then perform an ablation study on SoMA, emphasizing the contribution of each component to the DG performance. Lastly, we extend our approach beyond dense predictions to generative modeling.", + "bbox": [ + 89, + 719, + 480, + 825 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 89, + 835, + 279, + 851 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. For DGSS, we use three synthetic datasets (GTAV [62], SYNTHIA [64], UrbanSyn [24]), and four real-world datasets (Cityscapes [14], BDD100K [80], Map-", + "bbox": [ + 89, + 854, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/c1ead2381f29888e2eafca73abf8131649c62974073bd7bb5cc140f3ffb4f35b.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Synthetic-to-Real GeneralizationTest Domains (mIoU in %)
MethodsBackboneParams.*→Citys.→BDD→Map.Avg.
Single-source DGSS Trained on GTAV
○ CLOUDS [3]CLIP-CN-L0.0M60.2057.4067.0061.50
○ VLTSeg [33]EVA02-L304.2M65.3058.3066.0063.20
○ Rein [74]EVA02-L3.0M65.3060.5064.9063.60
○ FADA [4]EVA02-L11.7M66.7061.9066.1064.90
○ tqdm [54]EVA02-L304.2M68.8859.1870.1066.05
○ SoMA (Ours)EVA02-L5.1M68.0560.8168.3365.73
○ SoMA (Ours)EVA02-L5.1M69.9462.4868.3366.92
○ DoRA [48]DINOv2-L7.5M66.1259.3167.0764.17
○ VPT [34]DINOv2-L3.7M68.7558.6468.3265.24
○ SET [79]DINOv2-L6.1M68.0661.6467.6865.79
○ FADA [4]DINOv2-L11.7M68.2361.9468.0966.09
○ AdaptFormer [11]DINOv2-L6.3M70.1059.8168.7766.23
○ SSF [47]DINOv2-L0.5M68.9761.3068.7766.35
○ LoRA [30]DINOv2-L7.3M70.1360.1370.4266.89
○ Rein† [74]DINOv2-L3.0M70.6862.5169.6167.60
○ SoMA (Ours)DINOv2-L4.9M71.8261.3171.6768.27
○ SoMA (Ours)DINOv2-L4.9M73.6363.3370.9869.31
Multi-source DGSS Trained on GTAV + SYNTHIA
○ Rein† [74]DINOv2-L3.0M72.1761.5370.6968.13
○ SoMA (Ours)DINOv2-L4.9M73.1661.9072.7369.26
○ SoMA (Ours)DINOv2-L4.9M74.8563.5973.9270.79
Multi-source DGSS Trained on GTAV + SYNTHIA + UrbanSyn
○ FFTDINOv2-L304.2M75.9060.9372.8069.88
○ SoMA (Ours)DINOv2-L4.9M77.3362.7874.9371.68
○ FFT‡DINOv2-L307.3M77.0661.8175.0971.32
○ Rein‡ [74]DINOv2-L3.0M78.4262.2074.4971.70
○ SoMA‡ (Ours)DINOv2-L4.9M79.2263.8476.3073.12
○ FreezeDINOv2-G0.0M76.0861.9872.2370.10
○ FFTDINOv2-G1.1B76.9061.6973.5370.71
○ SoMA (Ours)DINOv2-G6.6M78.3963.7575.1672.43
○ SoMA (Ours)DINOv2-G6.6M80.3765.6776.1874.07
", + "bbox": [ + 516, + 89, + 903, + 544 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Comparison of the proposed SoMA with existing DGSS $\\circ$ and PEFT $\\circ$ methods under various synthetic-to-real settings.", + "bbox": [ + 513, + 546, + 906, + 574 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "illary [51], ACDC [66]). Specifically, GTAV, a large-scale dataset generated from the game engine, comprises 24,966 images. SYNTHIA consists of 9,400 photo-realistic synthetic images. UrbanSyn provides 7,539 driving scene images. Cityscapes is an autonomous driving dataset containing 2,975 training and 500 validation images $(2048\\times 1024)$ . BDD and Mapillary have 1,000 $(1280\\times 720)$ and 2,000 $(1920\\times 1080)$ validation images, respectively. ACDC is a driving scene dataset under adverse conditions, including night, snow, fog, and rain. For DGOD, we use the urban scene detection dataset introduced by [75]. The dataset consists of five different weather conditions: Daytime-Sunny (DS), Night-Clear (NC), Night-Rainy (NR), Dusk-Rainy (DR), and Daytime-Foggy (DF). Daytime-Sunny serves as the source domain, offering 26,518 images, with 8,313 used for in-domain evaluation, while the remaining adverse weather conditions are employed as unseen target domains. Implementation details. Since our method focuses on preserving pre-trained knowledge, we mainly evaluate its effectiveness using VFMs like EVA02 [22] and DINOv2 [53] as the backbone. To further elevate perfor", + "bbox": [ + 511, + 580, + 906, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "25607", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/3fb0ba9c077ccb735bbc919c486c811d03286701c1a3f92da63b76e2b6d0b330.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Real-to-Real GeneralizationTest Domains (mIoU in %)
MethodsBackboneParams.*→BDD→Map.Avg.
Single-source DGSS Trained on Cityscapes
○ HGFormer [17]Swin-L196.0M61.5072.1066.80
○ CMFormer [5]Swin-L196.0M62.6073.6068.10
○ tqdm [54]EVA02-L304.2M64.7276.1570.44
○ FADA [4]DINOv2-L11.7M65.1275.8670.49
○ Rein† [74]DINOv2-L3.0M66.5375.1870.86
○ SoMA (Ours)DINOv2-L4.9M67.0276.4571.74
• SoMA (Ours)DINOv2-L4.9M68.0877.8772.98
", + "bbox": [ + 93, + 89, + 477, + 233 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/4160bf1ad54e725ba2b2b6d592abe903e2c7c60f430d9236496b51fb16f43146.jpg", + "table_caption": [ + "Table 3. Real-to-real DGSS comparison." + ], + "table_footnote": [], + "table_body": "
Clear-to-Adverse Weather MethodsACDC [66] Test Domains (mIoU in %)
→Night→Snow→Fog→RainAll
Single-source DGSS Trained on Cityscapes
○ HGFormer [17]52.768.669.972.067.2
○ SoMA (Ours)61.777.374.777.874.4
○ VLTSeg‡ [33]----77.9
○ Rein‡ [74]70.679.576.479.477.6
○ FFT‡68.980.376.979.777.7
○ SoMA‡ (Ours)73.279.876.880.278.8
", + "bbox": [ + 94, + 253, + 477, + 388 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/42291a7c90ef4506ea58d0aad9208d8bbad8030a4f584d545481f1caf9ef731e.jpg", + "table_caption": [ + "Table 4. Results on Cityscapes $\\rightarrow$ ACDC test set." + ], + "table_footnote": [], + "table_body": "
Data EfficiencyTest Domains (mIoU in %)
MethodsBackboneParams.*→Citys.→BDD→Map.Avg.
DGSS Pre-trained on GTAV + SYNTHIA + UrbanSyn → 1/16 of Cityscapes
○ FFT‡DINOv2-L307.3M81.5365.2275.7374.16
○ Rein‡ [74]DINOv2-L3.0M82.5864.7673.7373.69
○ SoMA‡ (Ours)DINOv2-L4.9M82.5066.9977.0275.50
", + "bbox": [ + 94, + 407, + 480, + 510 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "mance, we leverage state-of-the-art decode heads such as Mask2Former [12] and Co-DETR [86]. Extensive experiments with various backbones and heads are provided in the Supplemental. By default, SoMA is applied to all linear layers within the self-attention and MLP, with a rank of 16. Detailed training settings are presented in the Supplemental.", + "bbox": [ + 89, + 537, + 482, + 628 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Comparison with State-of-the-Arts", + "text_level": 1, + "bbox": [ + 89, + 637, + 393, + 652 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We compare SoMA with previous SOTA baselines, achieving the highest performance across all DG scenarios without adding inference latency or regularization loss. In our tables, * indicates trainable parameters in the backbones. Marker • refers to models tested with multi-scale flip augmentation. † represents re-implemented test results using official checkpoints for a fair comparison, and ‡ denotes training on images cropped to $1024 \\times 1024$ ; otherwise, the default is $512 \\times 512$ . A comprehensive comparison with earlier works can be found in the Supplemental.", + "bbox": [ + 88, + 657, + 482, + 808 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Synthetic-to-real DGSS. In Tab. 2, we adopt $GTAV \\rightarrow \\{C_i\\}$ as the basic setting, and demonstrate scalability of SoMA by incorporating additional synthetic datasets into the training process. We report the mean Intersection over Union (mIoU) and benchmark our results against SOTA methods. SoMA consistently outperforms", + "bbox": [ + 89, + 810, + 483, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/31ff3e0c59eca8b356bce0829c1878aa62720dc61651bccaf0b5ba5e6710aa7e.jpg", + "table_caption": [ + "Table 5. Real-world data efficiency." + ], + "table_footnote": [], + "table_body": "
Clear-to-Adverse WeatherS-DGOD [75] Test Domains (mAP@0.5 in %)
MethodsParams.*DS→NC→DR→NR→DFAvg.
Single-source DGOD Trained on Daytime-Sunny (DS)
Backbone: ResNet101 [27] / Head: Faster R-CNN [61]
S-DGOD [75]42.3M56.136.628.216.633.528.7
CLIP-Gap [72]42.3M51.336.932.318.738.531.6
OA-DG [43]42.3M55.838.033.916.838.331.8
PDOC [46]42.3M53.638.533.719.239.132.6
UFR [49]42.3M58.640.833.219.239.633.2
DivAlign [15]42.3M52.842.538.124.137.235.5
SoMA (Ours)3.1M49.341.937.924.538.235.6
Backbone: DINOv2-L [53] / Head: Co-DETR [86]
Freeze0.0M65.054.255.042.846.949.7
FFT307.3M68.257.156.643.147.251.0
DoRA [48]5.8M69.058.758.045.048.952.7
AdaptFormer [11]6.3M68.958.858.344.449.852.8
LoRA [30]5.5M69.659.658.146.149.553.3
SoMA (Ours)4.9M69.459.359.347.651.054.3
", + "bbox": [ + 516, + 89, + 905, + 335 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 6. Domain generalized object detection.", + "bbox": [ + 566, + 339, + 852, + 352 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "VFM-based DGSS methods and PEFT approaches using both vision-language (EVA02) and vision-only (DINOv2) VFM backbones. Notably, SoMA achieves superior results over LoRA by minimally interfering with the hierarchical structure of VFM's pre-trained knowledge. Moreover, SoMA exhibits superior scalability across data, model, and input size, with improvements becoming more pronounced when test-time augmentation is applied. We provide comparisons regarding model efficiency in the Supplemental.", + "bbox": [ + 511, + 362, + 903, + 500 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Real-to-real DGSS. In Tables 3 and 4, we conduct experiments under the Citys. $\\rightarrow$ {BDD, Map.}/ACDC settings. When combined with DINOv2-L backbone, SoMA consistently gains the best results across all settings. This highlights the effectiveness of our method as a strong solution for real-world deployment scenarios, where domain shifts caused by geographic or weather variations are common.", + "bbox": [ + 511, + 501, + 903, + 607 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Real-world data efficiency. Following [74], we evaluate models under constrained conditions using a limited set of real-world images. Specifically, models pre-trained on synthetic datasets $GTAV + SYNTHIA + UrbanSyn$ (marked with $\\ddagger$ in Tab. 2) are fine-tuned on 1/16 of the Cityscapes training set. As shown in Tab. 5, SoMA achieves performance on par with Rein in the source domain, while offering greater robustness on unseen domains. This underscores SoMA's data-efficient adaptability to real-world scenes, while effectively avoiding overfitting to the source domain.", + "bbox": [ + 511, + 609, + 905, + 761 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Clear-to-adverse weather DGOD. In Tab. 6, we compare SoMA with existing DGOD methods for classic backbonehead setting and with PEFT methods for recent foundation architectures. As expected, SoMA demonstrates only minor improvements in the classic setting; however, in the VFM setting—where there is considerably more knowledge to retain—it shows substantial performance gains. As illustrated by the visualization results in Fig. 4, SoMA makes robust predictions even under diverse challenging scenes.", + "bbox": [ + 511, + 763, + 905, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "25608", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d68b013963ef7d21d57c454c99cf70fe31b6bd51aa0a29a38ab5128f9475fb1c.jpg", + "image_caption": [ + "Figure 4. DGOD qualitative results." + ], + "image_footnote": [], + "bbox": [ + 142, + 80, + 426, + 253 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/c8255192cb53c6ede59c4af8e03c76845065e75be5117f43d0546652cbdd8811.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsParams.*DGSS Avg.DGOD Avg.
Full fine-tuning (baseline)304.2M64.451.0
L + Freezing early blocks201.6M65.0 (↑0.6)51.4 (↑0.4)
L + Tuning principal components4.9M66.1 (↑1.1)53.0 (↑1.6)
L + Tuning minor components4.9M67.7 (↑2.7)53.8 (↑2.4)
L + Annealing weight decay4.9M68.3 (↑0.6)54.3 (↑0.5)
", + "bbox": [ + 93, + 282, + 483, + 367 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/3648921f1fb40f375ac0bd5b506d48c9ded9522ff62341836c39c83b6e69f252.jpg", + "table_caption": [ + "Table 7. Effect of our changes evaluated on DG benchmarks. See the full ablation study in the Supplemental." + ], + "table_footnote": [], + "table_body": "
{Wq,Wk,Wv,Wo} {Wup,Wdown}Params.* →Citys. →BDD →Map. Avg.
✓2.7M68.7661.5070.00
✓2.2M70.8261.3669.86
✓✓4.9M71.8261.3171.67
", + "bbox": [ + 93, + 402, + 483, + 465 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/81a90cc0c41496550c442dcd60e0b5e605aa30c8b524a10fc6328d5b01aec149.jpg", + "table_caption": [ + "Table 8. DGSS performance after applying SoMA to different types of modules (Self-attention / MLP) in DINoV2-Large." + ], + "table_footnote": [], + "table_body": "
Rank r48163264
DGSS avg. (mIoU in %)66.9167.7168.2767.7667.59
Params.*1.3M2.5M4.9M9.6M19.0M
", + "bbox": [ + 93, + 500, + 483, + 550 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 9. DGSS performance with different rank $r$ .", + "bbox": [ + 133, + 551, + 441, + 565 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.3. Ablation Study", + "text_level": 1, + "bbox": [ + 91, + 577, + 243, + 593 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Component Analysis. A thorough examination of each component of SoMA in the DGOD and DGSS basic setting, as summarized in Tab. 7, demonstrates that all components of SoMA incrementally improve DG performance without adding computational costs. Notably, the marked performance gap with tuning principal singular components [50] suggests that SoMA better preserves the integrity of pretrained representations while learning task-specific features.", + "bbox": [ + 89, + 598, + 482, + 718 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Tuning granularity. Interestingly, as shown in Tables 8 and 9, our method consistently achieves competitive results compared to SOTA baselines across all tuning granularities, indicating that SoMA can be flexibly configured according to the training budget, task difficulty, or domain gap. The best results are achieved with the default setting.", + "bbox": [ + 89, + 720, + 482, + 811 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.4. Subject Personalization", + "text_level": 1, + "bbox": [ + 89, + 820, + 307, + 835 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Domain generalized recognition requires consistent processing of inputs from diverse domains, whereas domain generalized generation involves generating outputs across a range of domains. Although large-scale Text-to-Image", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c3deea78171b5ae33ee50b141d34f258077fca93b41b0fb3e4bcb49094279cde.jpg", + "image_caption": [ + "Figure 5. Subject Personalization. 1. A dog gracefully leaping in origami style, 2. A dog in watercolor painting style, and 3. A dog soaring through a digital landscape in vector illustration style." + ], + "image_footnote": [], + "bbox": [ + 578, + 74, + 836, + 303 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "(T2I) models have convincingly demonstrated this ability, it can be compromised in subject personalization tasks involving fine-tuning. DreamBooth (DB) [65] introduces subject personalization, which fine-tunes a pre-trained T2I model on a few reference images to enable it to generate new visual concepts. DB updates all parameters within specific blocks, potentially impairing the pre-trained capability to generate images across various domains. Therefore, we explore whether SoMA can be seamlessly integrated into the DB framework to preserve pre-trained knowledge. For this task, we utilize Stable Diffusion XL (SDXL) [59] for the T2I diffusion model. Fig. 5 presents a visual comparison between DB and SoMA, using identical hyperparameter settings and sample seeds. As seen, SoMA, similar to the original SDXL model, can generate high-quality images of diverse visual domains (e.g., style, texture, pose) and exhibits superior identity preservation compared to DB with prior preservation (p.p.) loss [65]. More results with varied prompts and subjects are provided in the Supplemental.", + "bbox": [ + 511, + 368, + 906, + 656 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 678, + 633, + 694 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper presents a multifaceted exploration of the distribution of generalizable components in VFMs—examining them at the weight level, block level, and within training dynamics—and introduces the SoMA framework as an effective means to preserve generalizable components. Contrary to recent approaches that passively refine frozen VFM features, SoMA is integrated into all linear layers to effectively facilitate task adaptation, while minimizing interference with generalizable components by initializing the low-rank adapter with minor singular components. Through extensive experiments, we observe that SoMA significantly outperforms the state-of-the-art methods in DGSS, DGOD, and even subject-driven image generation tasks.", + "bbox": [ + 511, + 704, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "25609", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Acknowledgments", + "text_level": 1, + "bbox": [ + 91, + 90, + 269, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported by the National Research Foundation (NRF) grant funded by the Korea government (MSIT) [RS-2025-00562400] and [RS-2022-NR068754].", + "bbox": [ + 89, + 114, + 483, + 161 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 176, + 187, + 191 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Woo-Jin Ahn, Geun-Yeong Yang, Hyun-Duck Choi, and Myo-Taeg Lim. Style blind domain generalized semantic segmentation via covariance alignment and semantic consistence contrastive learning. In CVPR, 2024. 2, 3", + "[2] Muhammad Awais, Muzammal Naseer, Salman Khan, Rao Muhammad Anwer, Hisham Cholakkal, Mubarak Shah, Ming-Hsuan Yang, and Fahad Shahbaz Khan. Foundational models defining a new era in vision: A survey and outlook. arXiv:2307.13721, 2023. 2", + "[3] Yasser Benigmim, Subhankar Roy, Slim Essid, Vicky Kalogeiton, and Stephane Lathuilière. Collaborating foundation models for domain generalized semantic segmentation. In CVPR, 2024. 3, 6", + "[4] Qi Bi, Jingjun Yi, Hao Zheng, Haolan Zhan, Yawen Huang, Wei Ji, Yuexiang Li, and Yefeng Zheng. Learning frequency-adapted vision foundation model for domain generalized semantic segmentation. In NeurlPS, 2024. 3, 6, 7", + "[5] Qi Bi, Shaodi You, and Theo Gevers. Learning content-enhanced mask transformer for domain generalized urban-scene segmentation. In AAAI, 2024. 2, 7", + "[6] Dan Biderman, Jose Gonzalez Ortiz, Jacob Portes, Mansheej Paul, Philip Greengard, Connor Jennings, Daniel King, Sam Havens, Vitaliy Chiley, Jonathan Frankle, et al. Lora learns less and forgets less. TMLR, 2024. 2", + "[7] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, 2020. 1", + "[8] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, 2021. 3", + "[9] Gyusam Chang, Jiwon Lee, Donghyun Kim, Jinkyu Kim, Dongwook Lee, Daehyun Ji, Sujin Jang, and Sangpil Kim. Unified domain generalization and adaptation for multi-view 3d object detection. In NeurIPS, 2024. 2", + "[10] Prithvijit Chattopadhyay, Kartik Sarangmath, Vivek Vijaykumar, and Judy Hoffman. Pasta: Proportional amplitude spectrum training augmentation for syn-to-real domain generalization. In ICCV, 2023. 2, 3", + "[11] Shoufa Chen, Chongjian Ge, Zhan Tong, Jiangliu Wang, Yibing Song, Jue Wang, and Ping Luo. Adaptformer: Adapting vision transformers for scalable visual recognition. In NeurIPS, 2022. 2, 3, 6, 7", + "[12] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In CVPR, 2022. 1, 7", + "[13] Sungha Choi, Sanghun Jung, Huiwon Yun, Joanne T Kim, Seungryong Kim, and Jaegul Choo. Robustnet: Improving" + ], + "bbox": [ + 93, + 200, + 483, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "domain generalization in urban-scene segmentation via instance selective whitening. In CVPR, 2021. 2, 3, 5", + "[14] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In CVPR, 2016. 6", + "[15] Muhammad Sohail Danish, Muhammad Haris Khan, Muhammad Akhtar Munir, M Saquib Sarfraz, and Mohsen Ali. Improving single domain-generalized object detection: A focus on diversification and alignment. In CVPR, 2024. 3, 7", + "[16] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009. 2, 4", + "[17] Jian Ding, Nan Xue, Gui-Song Xia, Bernt Schiele, and Dengxin Dai. Hgformer: Hierarchical grouping transformer for domain generalized semantic segmentation. In CVPR, 2023. 3, 7", + "[18] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 3", + "[19] Carl Eckart and Gale Young. The approximation of one matrix by another of lower rank. Psychometrika, 1936. 4", + "[20] Mohammad Fahes, Tuan-Hung Vu, Andrei Bursuc, Patrick Pérez, and Raoul de Charette. A simple recipe for language-guided domain generalized segmentation. In CVPR, 2024. 3", + "[21] Qi Fan, Mattia Segu, Yu-Wing Tai, Fisher Yu, Chi-Keung Tang, Bernt Schiele, and Dengxin Dai. Towards robust object detection invariant to real-world domain shifts. In ICLR, 2023. 2, 3", + "[22] Yuxin Fang, Quan Sun, Xinggang Wang, Tiejun Huang, Xinlong Wang, and Yue Cao. Eva-02: A visual representation for neon genesis. IVC, 2024. 3, 6", + "[23] Aditya Sharad Golatkar, Alessandro Achille, and Stefano Soatto. Time matters in regularizing deep networks: Weight decay and data augmentation affect early learning dynamics, matter little near convergence. In NeurIPS, 2019. 6", + "[24] Jose L Gómez, Manuel Silva, Antonio Seoane, Agnès Borras, Mario Noriega, German Ros, Jose A Iglesias-Guitian, and Antonio M López. All for one, and one for all: Urbansyn dataset, the third musketeer of synthetic driving scenes. arXiv:2312.12176, 2023. 1, 6", + "[25] Junxian He, Chunting Zhou, Xuezhe Ma, Taylor Berg-Kirkpatrick, and Graham Neubig. Towards a unified view of parameter-efficient transfer learning. In ICLR, 2022. 2", + "[26] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Delving deep into rectifiers: Surpassing human-level performance on imagenet classification. In ICCV, 2015. 3", + "[27] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 2, 3, 7" + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "25610", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[28] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In CVPR, 2022. 3", + "[29] Lukas Hoyer, Dengxin Dai, and Luc Van Gool. Daformer: Improving network architectures and training strategies for domain-adaptive semantic segmentation. In CVPR, 2022. 5", + "[30] Edward J Hu, yelong shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In ICLR, 2022. 1, 2, 3, 5, 6, 7", + "[31] Jiajun Hu, Jian Zhang, Lei Qi, Yinghuan Shi, and Yang Gao. Learn to preserve and diversify: Parameter-efficient group with orthogonal regularization for domain generalization. In ECCV, 2024. 2", + "[32] Wei Huang, Chang Chen, Yong Li, Jiacheng Li, Cheng Li, Fenglong Song, Youliang Yan, and Zhiwei Xiong. Style projected clustering for domain generalized semantic segmentation. In CVPR, 2023. 3", + "[33] Christoph Hummer, Manuel Schwonberg, Liangwei Zhong, Hu Cao, Alois Knoll, and Hanno Gottschalk. Vltseg: Simple transfer of clip-based vision-language representations for domain generalized semantic segmentation. In ACCV, 2024. 3, 6, 7", + "[34] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In ECCV, 2022. 2, 3, 6", + "[35] Yuru Jia, Lukas Hoyer, Shengyu Huang, Tianfu Wang, Luc Van Gool, Konrad Schindler, and Anton Obukhov. Dgin-style: Domain-generalizable semantic segmentation with image diffusion models and stylized semantic control. In ECCV, 2024. 3", + "[36] Xueying Jiang, Jiaxing Huang, Sheng Jin, and Shijian Lu. Domain generalization via balancing training difficulty and model capability. In ICCV, 2023. 3", + "[37] Mengmeng Jing, Xiantong Zhen, Jingjing Li, and Cees GM Snoek. Order-preserving consistency regularization for domain adaptation and generalization. In ICCV, 2023. 3", + "[38] Christoph Kamann and Carsten Rother. Increasing the robustness of semantic segmentation models with painting-by-numbers. In ECCV, 2020. 3", + "[39] Jin Kim, Jiyoung Lee, Jungin Park, Dongbo Min, and Kwanghoon Sohn. Pin the memory: Learning to generalize semantic segmentation. In CVPR, 2022. 3", + "[40] Sunghwan Kim, Dae-hwan Kim, and Hoseong Kim. Texture learning domain randomization for domain generalized segmentation. In ICCV, 2023. 3, 5", + "[41] Suhyeon Lee, Hongje Seong, Seongwon Lee, and Euntai Kim. Wildnet: Learning domain generalized semantic segmentation from the wild. In CVPR, 2022. 3, 5", + "[42] Sohyun Lee, Taeyoung Son, and Suha Kwak. Fifo: Learning fog-invariant features for foggy scene segmentation. In CVPR, 2022. 3", + "[43] Wooju Lee, Dasol Hong, Hyungtae Lim, and Hyun Myung. Object-aware domain generalization for object detection. In AAAI, 2024. 3, 7", + "[44] Yoonho Lee, Annie S Chen, Fahim Tajwar, Ananya Kumar, Huaxiu Yao, Percy Liang, and Chelsea Finn. Surgical fine" + ], + "bbox": [ + 91, + 92, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "tuning improves adaptation to distribution shifts. In ICLR, 2023.5", + "[45] Brian Lester, Rami Al-Rfou, and Noah Constant. The power of scale for parameter-efficient prompt tuning. In EMNLP, 2021. 2", + "[46] Deng Li, Among Wu, Yaowei Wang, and Yahong Han. Prompt-driven dynamic object-centric learning for single domain generalization. In CVPR, 2024. 3, 7", + "[47] Dongze Lian, Daquan Zhou, Jiashi Feng, and Xinchao Wang. Scaling & shifting your features: A new baseline for efficient model tuning. In NeurIPS, 2022. 3, 6", + "[48] Shih-Yang Liu, Chien-Yi Wang, Hongxu Yin, Pavlo Molchanov, Yu-Chiang Frank Wang, Kwang-Ting Cheng, and Min-Hung Chen. DoRA: Weight-decomposed low-rank adaptation. In ICML, 2024. 3, 6, 7", + "[49] Yajing Liu, Shijun Zhou, Xiyao Liu, Chunhui Hao, Baojie Fan, and Jiandong Tian. Unbiased faster r-cnn for single-source domain generalized object detection. In CVPR, 2024. 3, 7", + "[50] Fanxu Meng, Zhaohui Wang, and Muhan Zhang. Pissa: Principal singular values and singular vectors adaptation of large language models. In NeurIPS, 2024. 3, 8", + "[51] Gerhard Neuhold, Tobias Ollmann, Samuel Rota Bulo, and Peter Kontschieder. The mapillary vistas dataset for semantic understanding of street scenes. In ICCV, 2017. 6", + "[52] Joshua Niemeijer, Manuel Schwonberg, Jan-Aike Termöhlen, Nico M Schmidt, and Tim Fingscheidt. Generalization by adaptation: Diffusion-based domain extension for domain-generalized semantic segmentation. In WACV, 2024. 3", + "[53] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. TMLR, 2023. 3, 4, 5, 6, 7", + "[54] Byeonghyun Pak, Byeongju Woo, Sunghwan Kim, Dae-hwan Kim, and Hoseong Kim. Textual query-driven mask transformer for domain generalized segmentation. In ECCV, 2024. 3, 6, 7", + "[55] Xingang Pan, Ping Luo, Jianping Shi, and Xiaou Tang. Two at once: Enhancing learning and generalization capacities via ibn-net. In ECCV, 2018. 3", + "[56] Yuwen Pan, Rui Sun, Naisong Luo, Tianzhu Zhang, and Yongdong Zhang. Exploring reliable matching with phase enhancement for night-time semantic segmentation. In ECCV, 2024.3", + "[57] Duo Peng, Yinjie Lei, Lingqiao Liu, Pingping Zhang, and Jun Liu. Global and local texture randomization for synthetic-to-real semantic segmentation. TIP, 2021. 2, 3", + "[58] Duo Peng, Yinjie Lei, Munawar Hayat, Yulan Guo, and Wen Li. Semantic-aware domain generalized segmentation. In CVPR, 2022. 2, 3", + "[59] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. SDXL: Improving latent diffusion models for high-resolution image synthesis. In ICLR, 2024. 3, 8" + ], + "bbox": [ + 516, + 92, + 903, + 897 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "25611", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[60] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 3", + "[61] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. In NeurIPS, 2015. 1, 7", + "[62] Stephan R Richter, Vibhav Vineet, Stefan Roth, and Vladlen Koltun. Playing for data: Ground truth from computer games. In ECCV, 2016. 1, 6", + "[63] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 3", + "[64] German Ros, Laura Sellart, Joanna Materzynska, David Vazquez, and Antonio M. Lopez. The synthia dataset: A large collection of synthetic images for semantic segmentation of urban scenes. In CVPR, 2016. 1, 6", + "[65] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In CVPR, 2023. 8", + "[66] Christos Sakaridis, Dengxin Dai, and Luc Van Gool. Acdc: The adverse conditions dataset with correspondences for semantic driving scene understanding. In ICCV, 2021. 6, 7", + "[67] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zh-moginov, and Liang-Chieh Chen. Mobilenetv2: Inverted residuals and linear bottlenecks. In CVPR, 2018. 2", + "[68] Fengyi Shen, Akhil Gurram, Ziyuan Liu, He Wang, and Alois Knoll. Diga: Distil to generalize and then adapt for domain adaptive semantic segmentation. In CVPR, 2023. 3", + "[69] Zhiqiang Tang, Yunhe Gao, Yi Zhu, Zhi Zhang, Mu Li, and Dimitris N Metaxas. Crossnorm and selfnorm for generalization under distribution shifts. In ICCV, 2021. 3", + "[70] Sumanth Udupa, Prajwal Gurunath, Aniruddh Sikdar, and Suresh Sundaram. Mrfp: Learning generalizable semantic segmentation from sim-2-real with multi-resolution feature perturbation. In CVPR, 2024. 2, 3", + "[71] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 3", + "[72] Vidit Vidit, Martin Engilberge, and Mathieu Salzmann. Clip the gap: A single domain generalization approach for object detection. In CVPR, 2023. 3, 7", + "[73] Zhixiang Wei, Lin Chen, Tao Tu, Pengyang Ling, Huaian Chen, and Yi Jin. Disentangle then parse: Night-time semantic segmentation with illumination disentanglement. In ICCV, 2023. 3", + "[74] Zhixiang Wei, Lin Chen, Yi Jin, Xiaoxiao Ma, Tianle Liu, Pengyang Ling, Ben Wang, Huaian Chen, and Jinjin Zheng. Stronger fewer & superior: Harnessing vision foundation models for domain generalized semantic segmentation. In CVPR, 2024. 2, 3, 6, 7", + "[75] Among Wu and Cheng Deng. Single-domain generalized object detection in urban scene via cyclic-disentangled self-distillation. In CVPR, 2022. 3, 6, 7" + ], + "bbox": [ + 91, + 90, + 480, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[76] Zhenyao Wu, Xinyi Wu, Xiaoping Zhang, Lili Ju, and Song Wang. Siamdoge: Domain generalizable semantic segmentation using siamese network. In ECCV, 2022. 3", + "[77] Qi Xu, Liang Yao, Zhengkai Jiang, Guannan Jiang, Wenqing Chu, Wenhui Han, Wei Zhang, Chengjie Wang, and Ying Tai. Dirl: Domain-invariant representation learning for generalizable semantic segmentation. In AAAI, 2022. 3", + "[78] Liwei Yang, Xiang Gu, and Jian Sun. Generalized semantic segmentation by self-supervised source domain projection and multi-level contrastive learning. In AAAI, 2023. 3", + "[79] Jingjun Yi, Qi Bi, Hao Zheng, Haolan Zhan, Wei Ji, Yawen Huang, Yuexiang Li, and Yefeng Zheng. Learning spectral-decomposed tokens for domain generalized semantic segmentation. In ACMMM, 2024. 3, 6", + "[80] Fisher Yu, Haofeng Chen, Xin Wang, Wenqi Xian, Yingying Chen, Fangchen Liu, Vashisht Madhavan, and Trevor Darrell. Bdd100k: A diverse driving dataset for heterogeneous multitask learning. In CVPR, 2020. 6", + "[81] Yeonguk Yu, Sungho Shin, Seunghyeok Back, Mihwan Ko, Sangjun Noh, and Kyoobin Lee. Domain-specific block selection and paired-view pseudo-labeling for online test-time adaptation. In CVPR, 2024. 5", + "[82] Xiangyu Yue, Yang Zhang, Sicheng Zhao, Alberto Sangiovanni-Vincentelli, Kurt Keutzer, and Boqing Gong. Domain randomization and pyramid consistency: Simulation-to-real generalization without accessing target domain data. In ICCV, 2019. 3", + "[83] Yuyang Zhao, Zhun Zhong, Na Zhao, Nicu Sebe, and Gim Hee Lee. Style-hallucinated dual consistency learning for domain generalized semantic segmentation. In ECCV, 2022. 3", + "[84] Zhun Zhong, Yuyang Zhao, Gim Hee Lee, and Nicu Sebe. Adversarial style augmentation for domain generalized urban-scene segmentation. In NeurIPS, 2022. 3, 5", + "[85] Jinghao Zhou, Chen Wei, Huiyu Wang, Wei Shen, Cihang Xie, Alan Yuille, and Tao Kong. ibot: Image bert pre-training with online tokenizer. arXiv:2111.07832, 2021. 3", + "[86] Zhuofan Zong, Guanglu Song, and Yu Liu. Detrs with collaborative hybrid assignments training. In ICCV, 2023. 7" + ], + "bbox": [ + 516, + 90, + 903, + 643 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "25612", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/SoMA_ Singular Value Decomposed Minor Components Adaptation for Domain Generalizable Representation Learning/282504c5-c667-4202-94ca-c92386507635_model.json b/2025/SoMA_ Singular Value Decomposed Minor Components Adaptation for Domain Generalizable Representation Learning/282504c5-c667-4202-94ca-c92386507635_model.json new file mode 100644 index 0000000000000000000000000000000000000000..4159cecc7ed7b4dcbd500e595e6acdf309fb874b --- /dev/null +++ b/2025/SoMA_ Singular Value Decomposed Minor Components Adaptation for Domain Generalizable Representation Learning/282504c5-c667-4202-94ca-c92386507635_model.json @@ -0,0 +1,2554 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.02 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation." + }, + { + "type": "header", + "bbox": [ + 0.325, + 0.017, + 0.725, + 0.031 + ], + "angle": 0, + "content": "Except for this watermark, it is identical to the accepted version;" + }, + { + "type": "header", + "bbox": [ + 0.294, + 0.032, + 0.756, + 0.047 + ], + "angle": 0, + "content": "the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.156, + 0.084, + 0.843, + 0.13 + ], + "angle": 0, + "content": "SoMA: Singular Value Decomposed Minor Components Adaptation for Domain Generalizable Representation Learning" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.157, + 0.782, + 0.193 + ], + "angle": 0, + "content": "Seokju Yun Seunghye Chae Dongheon Lee Youngmin Ro* Machine Intelligence Laboratory, University of Seoul, Korea" + }, + { + "type": "text", + "bbox": [ + 0.273, + 0.196, + 0.725, + 0.227 + ], + "angle": 0, + "content": "{wsz871, tmdhey, dslisleedh, youngmin.ro}@uos.ac.kr \nhttps://github.com/ysj9909/SoMA" + }, + { + "type": "image", + "bbox": [ + 0.096, + 0.237, + 0.379, + 0.468 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.384, + 0.239, + 0.573, + 0.313 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.384, + 0.316, + 0.572, + 0.389 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.384, + 0.394, + 0.572, + 0.466 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.585, + 0.238, + 0.904, + 0.467 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.472, + 0.907, + 0.557 + ], + "angle": 0, + "content": "Figure 1. Overview of SoMA framework. Left. SoMA achieves state-of-the-art results across diverse tasks, ranging from domaingeneralized semantic segmentation (DGSS) to object detection (DGOD), and performs well in both synthetic-to-real and real-to-real scenarios. Middle. Our method, trained solely on synthetic datasets, demonstrates strong generalization capabilities in complex real-world scenes. Right. We present SoMA, a method that applies SVD to the pre-trained weights, decomposing them into \\( r \\) minor singular components \\( U_{[:,-r:]} \\Sigma_{[-r:]} (V^T)_{[-r:,]} \\) and residuals. SoMA selectively tunes the smallest \\( r \\) components, effectively preserving world knowledge of foundation models. Since SoMA shares the same forward architecture as LoRA [30], it adds no extra latency during inference phase." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.569, + 0.328, + 0.585 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.602, + 0.486, + 0.875 + ], + "angle": 0, + "content": "Domain generalization (DG) aims to adapt a model using one or multiple source domains to ensure robust performance in unseen target domains. Recently, Parameter-Efficient Fine-Tuning (PEFT) of foundation models has shown promising results in the context of DG problem. Nevertheless, existing PEFT methods still struggle to strike a balance between preserving generalizable components of the pre-trained model and learning task-specific features. To gain insights into the distribution of generalizable components, we begin by analyzing the pre-trained weights through the lens of singular value decomposition. Building on these insights, we introduce Singular Value Decomposed Minor Components Adaptation (SoMA), an approach that selectively tunes minor singular components while keeping the residual parts frozen. SoMA effectively retains the generalization ability of the pre-trained model while efficiently acquiring task-specific skills. Moreover, we freeze domain-generalizable blocks and employ an annealing weight decay" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.571, + 0.907, + 0.707 + ], + "angle": 0, + "content": "strategy, thereby achieving an optimal balance in the delicate trade-off between generalizability and discriminability. SoMA attains state-of-the-art results on multiple benchmarks that span both domain generalized semantic segmentation to domain generalized object detection. In addition, our methods introduce no additional inference overhead or regularization loss, maintain compatibility with any backbone or head, and are designed to be versatile, allowing easy integration into a wide range of tasks." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.713, + 0.646, + 0.728 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Deep neural networks (DNNs) have recently made significant strides in dense prediction tasks, including semantic segmentation [12] and object detection [7, 61], both of which are crucial for safety-critical applications, such as autonomous driving. However, DNNs often fail to maintain reliable performance under domain shifts, which may result from diverse lighting conditions, weather changes, or differences in location. Likewise, while synthetic data [24, 62, 64] is employed to avoid the labor-intensive and costly process of constructing real-world datasets, the resulting performance degradation of DNNs in real-world deployment" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.237, + 0.901 + ], + "angle": 0, + "content": "*Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "25602" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.152 + ], + "angle": 0, + "content": "remains a critical issue to be resolved. To tackle these challenges, Domain Generalization (DG) is introduced to design models capable of consistent prediction on arbitrary unseen target domains." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.153, + 0.483, + 0.333 + ], + "angle": 0, + "content": "Existing DG methods for dense prediction tasks employ two main strategies to enhance model robustness. The first involves diversifying the training process in the input [10, 57] or feature space [21, 70] using augmentation/perturbation techniques, exposing the model to a broader range of styles to mitigate overfitting to specific domains. The second strategy focuses on imposing alignment constraints, such as normalization [58] or regularization losses [1, 13], to facilitate the learning of domain-invariant features. However, these methods mostly utilize outdated backbones (e.g., ResNet [27] or MobileNetV2 [67]) pretrained on mid-scale datasets, such as ImageNet [16]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.335, + 0.483, + 0.501 + ], + "angle": 0, + "content": "Meanwhile, vision foundation models (VFMs) [2], pretrained on vast, curated datasets with enormous parameter counts, have recently shown unprecedented generalization capabilities [74]. Given this landscape, we propose a paradigm shift—from applying the diversify & align approach to classic backbones, to devising methods that preserve the world knowledge of VFMs while effectively learning task-specific features. However, naive adaptation of VFMs via full fine-tuning (FFT), which involves retraining all model parameters, results in prohibitive costs and a risk of catastrophic forgetting of the pre-trained knowledge." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.502, + 0.483, + 0.742 + ], + "angle": 0, + "content": "In light of these challenges, a commonly adopted approach is parameter-efficient fine-tuning (PEFT), where only the injected lightweight adapters [11, 25, 30] or tokens [34, 45] are fine-tuned, while the pre-trained weights remain frozen. Low-Rank Adaptation (LoRA) [30] is one of the most widely used PEFT methods, which substitutes the updates with the product of two low-rank matrices. Interestingly, we found that LoRA exhibits superior generalization ability in DG dense prediction tasks compared to FFT and other PEFT methods (see Tab. 2), aligning with recent findings in LoRA-based literature [6, 31]. Nevertheless, LoRA adapters are built without considering the distribution of generalizable components within the pre-trained weights they modify. As a result, there is still room for enhancing generalizability by preserving the diverse domain knowledge of VFMs." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.743, + 0.483, + 0.864 + ], + "angle": 0, + "content": "To uncover the structure in which world knowledge is encoded within the pre-trained weights, we first perform singular value decomposition (SVD) on the learned weight matrices and then analyze the behavior of VFMs after selectively removing specific singular components1. Our analysis reveals that singular vectors associated with higher singular values tend to extract general features spanning multiple classes in ImageNet, while minor singular com" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.258 + ], + "angle": 0, + "content": "ponents are primarily responsible for capturing context-specific features. Capitalizing on our findings, we introduce Singular Value Decomposed Minor Components Adaptation (SoMA), which starts by performing SVD to decompose the weights into their singular vectors, i.e., \\(W = U\\Sigma V^T\\), then fine-tunes only the minor singular components. Specifically, we initialize LoRA adapter using the components with the smallest \\(r\\) singular values, i.e., \\(U_{[:,-r:]} \\Sigma_{[-r:]}(V^T)_{[-r:,]}\\), and the remaining residual components are frozen to maintain generalizability, as illustrated in Fig. 1 right." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.258, + 0.905, + 0.527 + ], + "angle": 0, + "content": "By extending our analysis from the weight-level to the block-level, we observe and empirically demonstrate that the early blocks of VFMs effectively extract well-localized semantic features and are less affected by style or input-level domain shifts (see Fig. 3). Therefore, we freeze these blocks to preserve generalizable components while reducing the number of trainable parameters. Although our proposed methods are effective, tuning the less-optimized bottom spectral space and freezing early blocks may lead to a lack of discriminability. To mitigate this issue, we introduce an annealing weight decay scheme that gradually reduces the regularization loss incurred by weight decay over the course of training. As shown in Fig. 1, comprehensive evaluations on DG benchmarks demonstrate that our framework outperforms previous state-of-the-art baselines and FFT, training only \\(0.58\\%\\) to \\(1.6\\%\\) of the parameters. Furthermore, SoMA demonstrates impressive scaling performance with both data and model size." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.53, + 0.766, + 0.544 + ], + "angle": 0, + "content": "We make the following contributions:" + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.548, + 0.905, + 0.593 + ], + "angle": 0, + "content": "- We underscore the critical importance of preserving pretrained knowledge while learning task-specific features in scenarios that involve the use of VFMs." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.595, + 0.905, + 0.701 + ], + "angle": 0, + "content": "- We present SoMA, a novel PEFT method specifically tailored for DG problem, which leverages spectral information to retain the generalization capacity of VFMs. In addition, we identify and freeze domain-generalizable blocks while employing an annealing weight decay strategy, thereby achieving an optimal balance between generalizability and discriminability." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.703, + 0.905, + 0.839 + ], + "angle": 0, + "content": "- Extensive experiments on various DG for semantic segmentation and object detection tasks reveal that our framework surpasses existing baselines by a substantial margin. Notably, SoMA achieves an mIoU of \\(80.4\\%\\) on Cityscapes without accessing real-world datasets, while training only \\(0.58\\%\\) of the parameters. Additionally, under the most challenging daytime-sunny \\(\\rightarrow\\) night-rainy detection setting, SoMA outperforms the prior SOTA implementation by achieving a \\(+23.5\\%\\) mAP gain." + }, + { + "type": "list", + "bbox": [ + 0.515, + 0.548, + 0.905, + 0.839 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.847, + 0.655, + 0.862 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Domain generalization for dense predictions. Domain Generalization (DG) methods for dense predictions [5, 9," + }, + { + "type": "page_footnote", + "bbox": [ + 0.091, + 0.875, + 0.483, + 0.901 + ], + "angle": 0, + "content": "In this paper, we refer to the combination of a singular value and its corresponding singular vectors as a singular component." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "25603" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.182 + ], + "angle": 0, + "content": "17, 32, 36, 39, 42, 46, 56, 68, 73, 75, 78] have recently garnered considerable attention due to their practical demands. These methods can be categorized along two main axes: i) data augmentation or domain randomization, both of which diversify the training process, and ii) alignment techniques to suppress domain-relevant features." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.186, + 0.485, + 0.502 + ], + "angle": 0, + "content": "Methods adopting the first approach at the input level employ various augmentation techniques [10, 38, 57, 84] or, more recently, leverage advances in generative modeling, such as diffusion models [59, 63], to generate data approximating the target domain [3, 35, 52], thereby expanding domain diversity and enriching learned representations. To enable more flexible augmentation, several studies [20, 21, 41, 70, 72] leverage feature stylization to enhance model robustness. On the other hand, alignment-based methods effectively reduce domain-sensitive components by employing feature normalization/whitening techniques [55, 58, 69], or by introducing regularization losses to suppress inconsistencies caused by simulated domain shifts [1, 13, 15, 37, 40, 43, 49, 76, 77, 82, 83]. However, most of the aforementioned approaches experimented with CNN backbones [27] trained on relatively restricted domains. In contrast, when applying vision foundation models (VFMs)—trained on massive datasets using transformer architectures [18, 71] and sophisticated training schemes [53]—to the DG problem, a fundamental shift in approach is imperative." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.506, + 0.483, + 0.732 + ], + "angle": 0, + "content": "Following this trend, recent studies leverage vision-language models as feature extractors [33], employing text embeddings with domain-invariant semantics as a source for style augmentation [20, 72] or as object queries in transformer-based decoders [54]. Rein [74] injects learnable adapters and tokens to refine feature maps for each instance while keeping the backbone frozen, significantly expanding the performance frontier. FADA [4] and SET [79] introduce frequency-adapted methods to effectively utilize frozen VFM features. In contrast to existing methods that simply freeze all parameters, we perform singular value decomposition on the pre-trained weights, freezing the generalizable principal components and tuning only the minor singular components responsible for context-specific features, thereby achieving superior DG performance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Vision foundation models (VFMs). Recently, the advent of VFMs with general perception capabilities has laid the foundation for a paradigm in which these models are broadly applicable to numerous downstream tasks. Seminal examples of models advancing this direction include CLIP [60], which demonstrates strong zero-shot generalizability through training on web-scale, weakly supervised image-text pairs; EVA02 [22], which enhances CLIP features with masked image modeling [28]; and DINOv2 [53], which incorporates losses from prior arts like iBoT [85] and DINO [8], and is trained on massive curated datasets, offer" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.137 + ], + "angle": 0, + "content": "ing strong spatial features for dense prediction tasks. Our SoMA builds on this foundation, harnessing the full potential of VFMs in the context of DG problem." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.138, + 0.907, + 0.44 + ], + "angle": 0, + "content": "Parameter-efficient fine-tuning (PEFT). With the remarkable success of PEFT methods in efficiently adapting large-scale foundation models in the realm of natural language processing, there has been a growing interest in extending these approaches to the field of computer vision. For example, VPT [34] prepends prompt tokens to the input sequence of several attention layers. AdaptFormer [11] introduces a lightweight adapter in parallel to the feed-forward network, whereas SSF [47] incorporates scale and shift parameters to modulate features for downstream tasks. LoRA [30] models the incremental updates of pre-trained weights using rank decomposition matrices, achieving performance comparable to full fine-tuning (FFT). However, there still often exists a performance gap between LoRA and FFT. DoRA [48] and PiSSA [50] address this gap by further decomposing the pre-trained weights, initializing the LoRA adapter with directional and principal components, respectively, and fine-tuning accordingly. While these methods aim to match the performance of FFT, our focus is on the preservation of generalizable components." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.454, + 0.653, + 0.47 + ], + "angle": 0, + "content": "3. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.478, + 0.906, + 0.598 + ], + "angle": 0, + "content": "Low-Rank Adaptation (LoRA). LoRA [30] hypothesizes that the weight changes during fine-tuning exhibit a low-rank structure. Given a pre-trained weight matrix \\( W_0 \\in \\mathbb{R}^{m \\times n} \\), LoRA constrains its update \\( \\Delta W \\in \\mathbb{R}^{m \\times n} \\) to a low-rank decomposition \\( \\Delta W = BA \\), where \\( B \\in \\mathbb{R}^{m \\times r} \\) and \\( A \\in \\mathbb{R}^{r \\times n} \\) are two low-rank matrices, with an intrinsic rank of \\( r \\ll \\min(m, n) \\). For \\( y = W_0 x \\), the modified forward pass is as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.585, + 0.606, + 0.905, + 0.623 + ], + "angle": 0, + "content": "\\[\ny = \\left(W _ {0} + \\Delta W\\right) x = \\left(W _ {0} + \\underline {{B A}}\\right) x, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.629, + 0.906, + 0.749 + ], + "angle": 0, + "content": "where \\( W_{0} \\) is frozen, and only the underlined low-rank parameters are fine-tuned, significantly reducing the number of trainable parameters. LoRA uses a uniform Kaiming distribution [26] to initialize \\( A \\), while \\( B \\) is initially set to zero, ensuring \\( BA = 0 \\) at the beginning of training. As seen in Eq. 1, the learned matrices \\( BA \\) can be merged with the frozen weight \\( W_{0} \\), allowing LoRA-based variants to avoid adding any additional inference burden." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Singular Value Decomposition (SVD) Analysis. We observe that LoRA and existing PEFT methods, by training only a relatively small subset of parameters or by enforcing low-rank-ness of weight updates, deliver promising DG performance without significantly distorting the pre-trained representations. However, many previous studies build adapters without considering how the generalizable components are structured within the weight matrices, potentially interfering with these components in the fine-tuning process. Building on the Eckart-Young-Mirsky the" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "25604" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.132, + 0.079, + 0.436, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.129, + 0.222, + 0.44, + 0.317 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.32, + 0.484, + 0.432 + ], + "angle": 0, + "content": "Figure 2. Distribution of generalizable components. Top. Number of classes exhibiting specific accuracy drops after applying SVD to DINOv2-large weights and reconstructing by truncating the smallest \\( r \\) singular components. Bottom. Distinct roles of singular components across levels. Numbers in parentheses represent the count of classes with an accuracy drop ratio exceeding \\( 50\\% \\). The average WordNet hierarchy depth of these classes is shown, with higher values indicating greater context specificity." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.443, + 0.484, + 0.579 + ], + "angle": 0, + "content": "orem [19], which validates that the optimal rank-\\(r\\) approximation of a matrix \\(W\\) is represented by the sum of its top \\(r\\) singular components, we analyze how the world knowledge of VFMs is structured from an SVD perspective. Using ImageNet-1k [16] as a testbed, our analysis computes the SVD of pre-trained weights across all layers of the DINOv2 [53] ViT-large model, followed by truncating specific singular values and their corresponding singular vectors to investigate the classes that are subsequently misclassified." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.582, + 0.484, + 0.793 + ], + "angle": 0, + "content": "As shown in Fig. 2 top, as the number of discarded smallest singular components increases, the number of misclassified classes grows exponentially rather than linearly, suggesting that components with larger singular values tend to capture more general features spanning multiple classes. To further examine whether different groups of singular components exhibit distinct behaviors, we consider three groups: the top 8 principal components \\( U_{[:;8]}\\Sigma_{[:8]}(V^T)_{[:8,:]} \\), the middle 160 singular components \\( U_{[:384:544]}\\Sigma_{[384:544]}(V^T)_{[384:544,:]} \\), and the bottom 320 minor components \\( U_{[:,-320:]} \\Sigma_{[-320:]}(V^T)_{[-320,:]} \\). We then truncate each group separately and analyze the classes that the rank-reduced model subsequently fails to classify (see Fig. 2 bottom)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.484, + 0.903 + ], + "angle": 0, + "content": "According to the Eckart-Young-Mirsky theorem and the heavy-tail distribution of singular values, it is evident that the top singular components encapsulate the core knowledge of VFM and should be preserved to maintain generalizability. Conversely, we note that the bottom-componenttruncated model fails to recognize classes that require finegrained or texture-biased features (e.g., specific dog breeds," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.379 + ], + "angle": 0, + "content": "wildlife with distinctive textures), as well as those that are context-specific (e.g., cornet, bikini, missile, dock). In contrast, we find that removing middle singular components results in significant performance declines for classes that rely on coarse-grained or shape-biased recognition (e.g., mortarboard, pier, altar, necklace, schooner, horizontal bar). Interestingly, as suggested by differences in the depth of the WordNet hierarchy, higher-value singular components do not function independently but rather interact in a hierarchical and composite manner. For instance, classes with notable accuracy drops can also be clustered into higher-level concepts, such as functionality (e.g., wok, hotpot, pitcher, cauldron, pop bottle, cup/ambulance, minivan, passenger car, police van, garbage truck, limousine) or co-occurrence (e.g., desktop computer, keyboard, notebook, desk). Additionally, almost all of these objects are commonly found in everyday scenes. Consequently, it can be seen that generalizable components are hierarchically and compositionally structured according to the magnitude of singular values." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.379, + 0.905, + 0.455 + ], + "angle": 0, + "content": "Problem Formulation. The entire model consists of a feature extractor \\(\\mathcal{F}\\), which combines a VFM parameterized by \\(\\Phi\\) with our SoMA adapter parameterized by \\(\\theta_{S}\\), and a task-specific decode head \\(\\mathcal{H}\\), parameterized by \\(\\theta_{h}\\). The optimization objective can be expressed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.574, + 0.461, + 0.905, + 0.502 + ], + "angle": 0, + "content": "\\[\n\\underset {\\theta_ {S}, \\theta_ {h}} {\\arg \\min } \\sum_ {i = 1} ^ {N _ {s}} \\mathcal {L} _ {t a s k} \\left(\\mathcal {H} _ {\\theta_ {h}} \\left(\\mathcal {F} _ {\\Phi , \\theta_ {S}} \\left(x _ {i}\\right)\\right), y _ {i}\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.509, + 0.908, + 0.585 + ], + "angle": 0, + "content": "where \\((x_{i},y_{i})\\) denotes an input image and its associated ground truth, drawn from the source domain datasets \\(\\mathcal{D}_s\\), comprising \\(N_{s}\\) samples. Our proposed SoMA framework facilitates generalizing task-specific skills learned from \\(\\mathcal{D}_s\\) to the unseen target domain datasets \\(\\mathcal{D}_t\\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.599, + 0.688, + 0.616 + ], + "angle": 0, + "content": "4. Proposed Method" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.62, + 0.907, + 0.711 + ], + "angle": 0, + "content": "In this section, we present our parameter-efficient adaptation methods specifically tailored for DG dense prediction tasks, inspired by the insights from our SVD analysis. Subsequently, we propose a block freeze strategy and an annealing weight decay scheme to strike an optimal trade-off between generalizability and discriminability." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.717, + 0.863, + 0.734 + ], + "angle": 0, + "content": "4.1. Minor Singular Components Adaptation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.739, + 0.906, + 0.86 + ], + "angle": 0, + "content": "To effectively preserve the integrity of generalizable representations in pre-trained weights, we propose Singular value decomposed Minor components Adaptation (SoMA). Firstly, SoMA performs singular value decomposition (SVD) on the pre-trained weight matrices, including those from the attention layers, multi-layer perceptron (MLP) layers, and more. The weight matrix \\( W \\in \\mathbb{R}^{m \\times n} \\) is decomposed using SVD as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.613, + 0.864, + 0.907, + 0.904 + ], + "angle": 0, + "content": "\\[\nW = U \\Sigma V ^ {T} = \\sum_ {i = 1} ^ {R} \\sigma_ {i} u _ {i} v _ {i} ^ {T}, \\tag {3}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "25605" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.107, + 0.083, + 0.472, + 0.277 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.28, + 0.483, + 0.392 + ], + "angle": 0, + "content": "Figure 3. The inherent generalization capabilities of the early blocks of VFM. Top. We apply PCA on the extracted intermediate features (8th block of DINOv2-Large) and visualize the top three leading components. Bottom. Class-wise IoU comparison for rare classes [29] under the GTAV \\(\\rightarrow\\) Cityscapes setting, focusing on the impact of freezing the first eight blocks. \"NFEB\" stands for the Number of Frozen Early Blocks. A darker shade of blue indicates enhanced generalization performance relative to the baseline." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.399, + 0.484, + 0.49 + ], + "angle": 0, + "content": "where \\(U = [u_{1}, u_{2}, \\dots, u_{m}] \\in \\mathbb{R}^{m \\times m}\\) and \\(V = [v_{1}, v_{2}, \\dots, v_{n}] \\in \\mathbb{R}^{n \\times n}\\) are the left and right singular vectors, whose columns form an orthonormal basis for \\(\\mathbb{R}^m\\) and \\(\\mathbb{R}^n\\) respectively, \\(\\Sigma \\in \\mathbb{R}^{m \\times n}\\) is a diagonal matrix whose entries \\(\\sigma_{i}\\) are singular values arranged in descending order, and \\(R\\) denotes the rank of \\(W\\), with \\(R \\leq \\min(m, n)\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.49, + 0.484, + 0.596 + ], + "angle": 0, + "content": "We then divide the singular components into two groups according to their singular values: the minor singular components with the smallest \\( r \\) singular values \\( U_{[:,-r:]} \\Sigma_{[-r:]} (V^T)_{[-r:]} \\) and the residual singular components \\( U_{[:,-r]} \\Sigma_{[:-r]} (V^T)_{[:-r:]} \\). As illustrated in Fig. 1 right, we employ the QR-type reconstruction of minor singular components to initialize the adapter for fine-tuning:" + }, + { + "type": "equation", + "bbox": [ + 0.174, + 0.597, + 0.482, + 0.617 + ], + "angle": 0, + "content": "\\[\nB = U _ {[:, - r: ]} \\sqrt {\\Sigma} _ {[ - r: ]} \\in \\mathbb {R} ^ {m \\times r}, \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.175, + 0.619, + 0.482, + 0.639 + ], + "angle": 0, + "content": "\\[\nA = \\sqrt {\\Sigma} _ {[ - r: ]} \\left(V ^ {T}\\right) _ {[ - r:,.. ]} \\in \\mathbb {R} ^ {r \\times n}. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.647, + 0.484, + 0.753 + ], + "angle": 0, + "content": "In the fine-tuning process, the frozen residual matrix \\( W_{res} \\) is defined as \\( W - BA \\) to prevent numerical errors introduced during the SVD step, while maintaining the pretrained generalization capability at the beginning of fine-tuning. After fine-tuning, the learned low-rank matrices \\( B' \\) and \\( A' \\) can be merged into \\( W_{res} \\), as in LoRA [30], ensuring no additional computation overhead during inference:" + }, + { + "type": "equation", + "bbox": [ + 0.099, + 0.758, + 0.483, + 0.793 + ], + "angle": 0, + "content": "\\[\ny = \\left(W \\underbrace {- B A + B ^ {\\prime} A ^ {\\prime}} _ {\\Delta W _ {S o M A}}\\right) x = \\left(W _ {r e s} + B ^ {\\prime} A ^ {\\prime}\\right) x = W ^ {\\prime} x. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.484, + 0.904 + ], + "angle": 0, + "content": "Importantly, SoMA offers several key advantages for tackling the DG problem: i) SoMA initializes the adapter using components orthogonal to the principal components, thus minimizing interference with pre-trained representations. For the batched version of the forward pass described in Eq. 1 (with \\(X \\in \\mathbb{R}^{n \\times b}\\) and \\(Y \\in \\mathbb{R}^{m \\times b}\\)), the gradients of \\(B\\) and \\(A\\) are \\(\\frac{\\partial \\mathcal{L}}{\\partial B} = \\frac{\\partial \\mathcal{L}}{\\partial Y} X^T A^T\\) and \\(\\frac{\\partial \\mathcal{L}}{\\partial A} = B^T \\frac{\\partial \\mathcal{L}}{\\partial Y} X^T\\)." + }, + { + "type": "table", + "bbox": [ + 0.52, + 0.077, + 0.905, + 0.202 + ], + "angle": 0, + "content": "
SMRBlockiΔWSoMAΔW*SoMAΔWLoRA
|uiTΔVvi/σi12th0-2550.0750.084 (↓17%)0.101
256-5110.0940.094 (↓43%)0.166
512-7670.1570.165 (↓40%)0.275
768-10234.1046.152 (↑21%)5.095
24th0-2550.0970.095 (↓24%)0.125
256-5110.1050.108 (↓35%)0.166
512-7670.2060.189 (↓41%)0.323
768-10232.7813.620 (↑26%)2.865
" + }, + { + "type": "table_caption", + "bbox": [ + 0.512, + 0.205, + 0.908, + 0.261 + ], + "angle": 0, + "content": "Table 1. Singular modulation ratio comparison. * indicates the adaptation matrix trained using the annealing weight decay strategy. The weight matrices are taken from the query projection layers of the 12th and 24th attention blocks in DINOv2-Large [53]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.269, + 0.907, + 0.405 + ], + "angle": 0, + "content": "Therefore, unlike LoRA, which randomly initializes \\( A \\), our approach facilitates convergence along singular directions that preserve the integrity of the generalizable components. ii) Consistent with prior DG methods [13, 40, 41, 84] that guide models to prioritize shape over texture, thereby achieving superior out-of-distribution performance, SoMA effectively preserves the domain generalizable shape-biased features of VFMs by tuning only the minor singular components (see Fig. 2 bottom and Tab. 1)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.406, + 0.906, + 0.512 + ], + "angle": 0, + "content": "Discussion. To assess whether the proposed adaptation matrix \\(\\Delta W_{SoMA}\\) minimally correlates with the top singular directions of the pre-trained weight matrix \\(W_0\\), we project \\(\\Delta W\\) onto the singular vectors of \\(W_0\\). Specifically, we quantify the extent of interference by \\(\\Delta W\\) on each \\(i\\)-th singular direction through a singular modulation ratio (SMR), which is formally defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.635, + 0.516, + 0.905, + 0.548 + ], + "angle": 0, + "content": "\\[\n\\mathrm {S M R} _ {i} = \\left| \\frac {u _ {i} ^ {\\mathrm {T}} \\Delta W v _ {i}}{\\sigma_ {i}} \\right|. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.553, + 0.907, + 0.72 + ], + "angle": 0, + "content": "For ease of analysis, we partition the 1024 singular directions into four equal groups, calculate the average SMR for each group, and then compare the resulting values between SoMA and LoRA. As shown in Tab. 1, the SoMA initialization scheme proves more effective than LoRA in minimizing interference with singular components with higher singular values. In summary, our method effectively preserves the structure of the VFM's generalizable representations while adapting to downstream tasks, resulting in significant improvements on DG benchmarks compared to SOTA baselines, as confirmed by our experimental results." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.73, + 0.72, + 0.746 + ], + "angle": 0, + "content": "4.2. Freezing Early Blocks" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.907, + 0.902 + ], + "angle": 0, + "content": "In this section, we explore the distribution of generalizable components at the block level. Drawing on recent domain adaptation methods [44, 81], which find that the early blocks of a model primarily engage in domain-specific feature extraction and should therefore be updated, we conjecture that in the context of leveraging VFM for DG, the early blocks should instead remain unchanged. Intuitively, these initial blocks of VFMs are adept at projecting low-level features from diverse domains into domain-invariant semantic representations. In other words, while adjusting" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "25606" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.152 + ], + "angle": 0, + "content": "early blocks may help bridge input-level domain gaps in domain adaptation, freezing the early blocks of VFMs is essential for DG to prevent overfitting to the source domain and to retain the model's generalization capability." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.153, + 0.483, + 0.394 + ], + "angle": 0, + "content": "To validate our hypothesis, we perform PCA analysis on the features extracted from the early blocks of VFM and compare the performance on rare classes when these blocks are frozen, as illustrated in Fig. 3. Despite considerable domain shifts in style and content, the principal components align well with the image layout and maintain consistent semantic information (e.g., the horses have similar colors). Furthermore, simply freezing the early blocks yields a notable improvement in IoU for rare classes, indicating that updating these blocks can severely disrupt the learned representations of various classes in VFMs. Thus, by default, we opt to freeze all blocks up to the one that generates the feature map fed as the highest resolution input to the segmentation/detection head (e.g., the first eight blocks). Detailed ablations of varying the number of frozen early blocks are presented in the Supplemental." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.404, + 0.317, + 0.42 + ], + "angle": 0, + "content": "4.3. Annealing Weight Decay" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.426, + 0.483, + 0.684 + ], + "angle": 0, + "content": "Our methods are specialized in preserving the pre-trained knowledge of VFMs, but this may come at the cost of diminished discriminability. One can remove weight decay or set its coefficient to a very small value to enhance discriminability. However, doing so risks increased interference of the tuned parameters with the principal components of the pre-trained weights. Therefore, based on the observation that regularization primarily affects the early phases of learning by guiding a model towards having good generalization properties [23], we leverage an Annealing Weight Decay (AWD) scheme. Specifically, AWD starts with a relatively large weight decay coefficient and progressively reduces it to zero as training progresses. By default, we use a cosine schedule for AWD. As shown in Tab. 1 (\\(\\Delta W_{SoMA}\\) vs. \\(\\Delta W_{SoMA}^{*}\\)) and Tab. 7, AWD integrates seamlessly with SoMA, improving discriminability without compromising SoMA's ability to preserve pre-trained knowledge." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.698, + 0.223, + 0.714 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.482, + 0.826 + ], + "angle": 0, + "content": "We conduct a variety of experiments to showcase the efficacy of SoMA on DG benchmarks including domain generalized semantic segmentation (DGSS) and domain generalized object detection (DGOD). We then perform an ablation study on SoMA, emphasizing the contribution of each component to the DG performance. Lastly, we extend our approach beyond dense predictions to generative modeling." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.836, + 0.281, + 0.852 + ], + "angle": 0, + "content": "5.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Datasets. For DGSS, we use three synthetic datasets (GTAV [62], SYNTHIA [64], UrbanSyn [24]), and four real-world datasets (Cityscapes [14], BDD100K [80], Map-" + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.09, + 0.905, + 0.545 + ], + "angle": 0, + "content": "
Synthetic-to-Real GeneralizationTest Domains (mIoU in %)
MethodsBackboneParams.*→Citys.→BDD→Map.Avg.
Single-source DGSS Trained on GTAV
○ CLOUDS [3]CLIP-CN-L0.0M60.2057.4067.0061.50
○ VLTSeg [33]EVA02-L304.2M65.3058.3066.0063.20
○ Rein [74]EVA02-L3.0M65.3060.5064.9063.60
○ FADA [4]EVA02-L11.7M66.7061.9066.1064.90
○ tqdm [54]EVA02-L304.2M68.8859.1870.1066.05
○ SoMA (Ours)EVA02-L5.1M68.0560.8168.3365.73
○ SoMA (Ours)EVA02-L5.1M69.9462.4868.3366.92
○ DoRA [48]DINOv2-L7.5M66.1259.3167.0764.17
○ VPT [34]DINOv2-L3.7M68.7558.6468.3265.24
○ SET [79]DINOv2-L6.1M68.0661.6467.6865.79
○ FADA [4]DINOv2-L11.7M68.2361.9468.0966.09
○ AdaptFormer [11]DINOv2-L6.3M70.1059.8168.7766.23
○ SSF [47]DINOv2-L0.5M68.9761.3068.7766.35
○ LoRA [30]DINOv2-L7.3M70.1360.1370.4266.89
○ Rein† [74]DINOv2-L3.0M70.6862.5169.6167.60
○ SoMA (Ours)DINOv2-L4.9M71.8261.3171.6768.27
○ SoMA (Ours)DINOv2-L4.9M73.6363.3370.9869.31
Multi-source DGSS Trained on GTAV + SYNTHIA
○ Rein† [74]DINOv2-L3.0M72.1761.5370.6968.13
○ SoMA (Ours)DINOv2-L4.9M73.1661.9072.7369.26
○ SoMA (Ours)DINOv2-L4.9M74.8563.5973.9270.79
Multi-source DGSS Trained on GTAV + SYNTHIA + UrbanSyn
○ FFTDINOv2-L304.2M75.9060.9372.8069.88
○ SoMA (Ours)DINOv2-L4.9M77.3362.7874.9371.68
○ FFT‡DINOv2-L307.3M77.0661.8175.0971.32
○ Rein‡ [74]DINOv2-L3.0M78.4262.2074.4971.70
○ SoMA‡ (Ours)DINOv2-L4.9M79.2263.8476.3073.12
○ FreezeDINOv2-G0.0M76.0861.9872.2370.10
○ FFTDINOv2-G1.1B76.9061.6973.5370.71
○ SoMA (Ours)DINOv2-G6.6M78.3963.7575.1672.43
○ SoMA (Ours)DINOv2-G6.6M80.3765.6776.1874.07
" + }, + { + "type": "table_caption", + "bbox": [ + 0.514, + 0.547, + 0.907, + 0.575 + ], + "angle": 0, + "content": "Table 2. Comparison of the proposed SoMA with existing DGSS \\( \\circ \\) and PEFT \\( \\circ \\) methods under various synthetic-to-real settings." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.582, + 0.907, + 0.901 + ], + "angle": 0, + "content": "illary [51], ACDC [66]). Specifically, GTAV, a large-scale dataset generated from the game engine, comprises 24,966 images. SYNTHIA consists of 9,400 photo-realistic synthetic images. UrbanSyn provides 7,539 driving scene images. Cityscapes is an autonomous driving dataset containing 2,975 training and 500 validation images \\((2048\\times 1024)\\). BDD and Mapillary have 1,000 \\((1280\\times 720)\\) and 2,000 \\((1920\\times 1080)\\) validation images, respectively. ACDC is a driving scene dataset under adverse conditions, including night, snow, fog, and rain. For DGOD, we use the urban scene detection dataset introduced by [75]. The dataset consists of five different weather conditions: Daytime-Sunny (DS), Night-Clear (NC), Night-Rainy (NR), Dusk-Rainy (DR), and Daytime-Foggy (DF). Daytime-Sunny serves as the source domain, offering 26,518 images, with 8,313 used for in-domain evaluation, while the remaining adverse weather conditions are employed as unseen target domains. Implementation details. Since our method focuses on preserving pre-trained knowledge, we mainly evaluate its effectiveness using VFMs like EVA02 [22] and DINOv2 [53] as the backbone. To further elevate perfor" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "25607" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.094, + 0.09, + 0.478, + 0.234 + ], + "angle": 0, + "content": "
Real-to-Real GeneralizationTest Domains (mIoU in %)
MethodsBackboneParams.*→BDD→Map.Avg.
Single-source DGSS Trained on Cityscapes
○ HGFormer [17]Swin-L196.0M61.5072.1066.80
○ CMFormer [5]Swin-L196.0M62.6073.6068.10
○ tqdm [54]EVA02-L304.2M64.7276.1570.44
○ FADA [4]DINOv2-L11.7M65.1275.8670.49
○ Rein† [74]DINOv2-L3.0M66.5375.1870.86
○ SoMA (Ours)DINOv2-L4.9M67.0276.4571.74
• SoMA (Ours)DINOv2-L4.9M68.0877.8772.98
" + }, + { + "type": "table_caption", + "bbox": [ + 0.162, + 0.236, + 0.414, + 0.249 + ], + "angle": 0, + "content": "Table 3. Real-to-real DGSS comparison." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.254, + 0.478, + 0.389 + ], + "angle": 0, + "content": "
Clear-to-Adverse Weather MethodsACDC [66] Test Domains (mIoU in %)
→Night→Snow→Fog→RainAll
Single-source DGSS Trained on Cityscapes
○ HGFormer [17]52.768.669.972.067.2
○ SoMA (Ours)61.777.374.777.874.4
○ VLTSeg‡ [33]----77.9
○ Rein‡ [74]70.679.576.479.477.6
○ FFT‡68.980.376.979.777.7
○ SoMA‡ (Ours)73.279.876.880.278.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.138, + 0.391, + 0.439, + 0.404 + ], + "angle": 0, + "content": "Table 4. Results on Cityscapes \\( \\rightarrow \\) ACDC test set." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.409, + 0.481, + 0.511 + ], + "angle": 0, + "content": "
Data EfficiencyTest Domains (mIoU in %)
MethodsBackboneParams.*→Citys.→BDD→Map.Avg.
DGSS Pre-trained on GTAV + SYNTHIA + UrbanSyn → 1/16 of Cityscapes
○ FFT‡DINOv2-L307.3M81.5365.2275.7374.16
○ Rein‡ [74]DINOv2-L3.0M82.5864.7673.7373.69
○ SoMA‡ (Ours)DINOv2-L4.9M82.5066.9977.0275.50
" + }, + { + "type": "table_caption", + "bbox": [ + 0.177, + 0.514, + 0.399, + 0.527 + ], + "angle": 0, + "content": "Table 5. Real-world data efficiency." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.539, + 0.483, + 0.63 + ], + "angle": 0, + "content": "mance, we leverage state-of-the-art decode heads such as Mask2Former [12] and Co-DETR [86]. Extensive experiments with various backbones and heads are provided in the Supplemental. By default, SoMA is applied to all linear layers within the self-attention and MLP, with a rank of 16. Detailed training settings are presented in the Supplemental." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.638, + 0.395, + 0.654 + ], + "angle": 0, + "content": "5.2. Comparison with State-of-the-Arts" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.659, + 0.483, + 0.809 + ], + "angle": 0, + "content": "We compare SoMA with previous SOTA baselines, achieving the highest performance across all DG scenarios without adding inference latency or regularization loss. In our tables, * indicates trainable parameters in the backbones. Marker • refers to models tested with multi-scale flip augmentation. † represents re-implemented test results using official checkpoints for a fair comparison, and ‡ denotes training on images cropped to \\(1024 \\times 1024\\); otherwise, the default is \\(512 \\times 512\\). A comprehensive comparison with earlier works can be found in the Supplemental." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Synthetic-to-real DGSS. In Tab. 2, we adopt \\(GTAV \\rightarrow \\{C_i\\}\\) as the basic setting, and demonstrate scalability of SoMA by incorporating additional synthetic datasets into the training process. We report the mean Intersection over Union (mIoU) and benchmark our results against SOTA methods. SoMA consistently outperforms" + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.09, + 0.906, + 0.337 + ], + "angle": 0, + "content": "
Clear-to-Adverse WeatherS-DGOD [75] Test Domains (mAP@0.5 in %)
MethodsParams.*DS→NC→DR→NR→DFAvg.
Single-source DGOD Trained on Daytime-Sunny (DS)
Backbone: ResNet101 [27] / Head: Faster R-CNN [61]
S-DGOD [75]42.3M56.136.628.216.633.528.7
CLIP-Gap [72]42.3M51.336.932.318.738.531.6
OA-DG [43]42.3M55.838.033.916.838.331.8
PDOC [46]42.3M53.638.533.719.239.132.6
UFR [49]42.3M58.640.833.219.239.633.2
DivAlign [15]42.3M52.842.538.124.137.235.5
SoMA (Ours)3.1M49.341.937.924.538.235.6
Backbone: DINOv2-L [53] / Head: Co-DETR [86]
Freeze0.0M65.054.255.042.846.949.7
FFT307.3M68.257.156.643.147.251.0
DoRA [48]5.8M69.058.758.045.048.952.7
AdaptFormer [11]6.3M68.958.858.344.449.852.8
LoRA [30]5.5M69.659.658.146.149.553.3
SoMA (Ours)4.9M69.459.359.347.651.054.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.567, + 0.34, + 0.854, + 0.353 + ], + "angle": 0, + "content": "Table 6. Domain generalized object detection." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.363, + 0.905, + 0.501 + ], + "angle": 0, + "content": "VFM-based DGSS methods and PEFT approaches using both vision-language (EVA02) and vision-only (DINOv2) VFM backbones. Notably, SoMA achieves superior results over LoRA by minimally interfering with the hierarchical structure of VFM's pre-trained knowledge. Moreover, SoMA exhibits superior scalability across data, model, and input size, with improvements becoming more pronounced when test-time augmentation is applied. We provide comparisons regarding model efficiency in the Supplemental." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.502, + 0.905, + 0.608 + ], + "angle": 0, + "content": "Real-to-real DGSS. In Tables 3 and 4, we conduct experiments under the Citys. \\(\\rightarrow\\) {BDD, Map.}/ACDC settings. When combined with DINOv2-L backbone, SoMA consistently gains the best results across all settings. This highlights the effectiveness of our method as a strong solution for real-world deployment scenarios, where domain shifts caused by geographic or weather variations are common." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.611, + 0.906, + 0.762 + ], + "angle": 0, + "content": "Real-world data efficiency. Following [74], we evaluate models under constrained conditions using a limited set of real-world images. Specifically, models pre-trained on synthetic datasets \\(GTAV + SYNTHIA + UrbanSyn\\) (marked with \\(\\ddagger\\) in Tab. 2) are fine-tuned on 1/16 of the Cityscapes training set. As shown in Tab. 5, SoMA achieves performance on par with Rein in the source domain, while offering greater robustness on unseen domains. This underscores SoMA's data-efficient adaptability to real-world scenes, while effectively avoiding overfitting to the source domain." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.906, + 0.902 + ], + "angle": 0, + "content": "Clear-to-adverse weather DGOD. In Tab. 6, we compare SoMA with existing DGOD methods for classic backbonehead setting and with PEFT methods for recent foundation architectures. As expected, SoMA demonstrates only minor improvements in the classic setting; however, in the VFM setting—where there is considerably more knowledge to retain—it shows substantial performance gains. As illustrated by the visualization results in Fig. 4, SoMA makes robust predictions even under diverse challenging scenes." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "25608" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.143, + 0.081, + 0.428, + 0.254 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.175, + 0.259, + 0.4, + 0.273 + ], + "angle": 0, + "content": "Figure 4. DGOD qualitative results." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.283, + 0.484, + 0.368 + ], + "angle": 0, + "content": "
MethodsParams.*DGSS Avg.DGOD Avg.
Full fine-tuning (baseline)304.2M64.451.0
L + Freezing early blocks201.6M65.0 (↑0.6)51.4 (↑0.4)
L + Tuning principal components4.9M66.1 (↑1.1)53.0 (↑1.6)
L + Tuning minor components4.9M67.7 (↑2.7)53.8 (↑2.4)
L + Annealing weight decay4.9M68.3 (↑0.6)54.3 (↑0.5)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.092, + 0.371, + 0.484, + 0.398 + ], + "angle": 0, + "content": "Table 7. Effect of our changes evaluated on DG benchmarks. See the full ablation study in the Supplemental." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.404, + 0.484, + 0.466 + ], + "angle": 0, + "content": "
{Wq,Wk,Wv,Wo} {Wup,Wdown}Params.* →Citys. →BDD →Map. Avg.
✓2.7M68.7661.5070.00
✓2.2M70.8261.3669.86
✓✓4.9M71.8261.3171.67
" + }, + { + "type": "table_caption", + "bbox": [ + 0.092, + 0.468, + 0.484, + 0.495 + ], + "angle": 0, + "content": "Table 8. DGSS performance after applying SoMA to different types of modules (Self-attention / MLP) in DINoV2-Large." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.5, + 0.484, + 0.551 + ], + "angle": 0, + "content": "
Rank r48163264
DGSS avg. (mIoU in %)66.9167.7168.2767.7667.59
Params.*1.3M2.5M4.9M9.6M19.0M
" + }, + { + "type": "table_caption", + "bbox": [ + 0.134, + 0.552, + 0.442, + 0.566 + ], + "angle": 0, + "content": "Table 9. DGSS performance with different rank \\( r \\) ." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.578, + 0.245, + 0.594 + ], + "angle": 0, + "content": "5.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.599, + 0.483, + 0.719 + ], + "angle": 0, + "content": "Component Analysis. A thorough examination of each component of SoMA in the DGOD and DGSS basic setting, as summarized in Tab. 7, demonstrates that all components of SoMA incrementally improve DG performance without adding computational costs. Notably, the marked performance gap with tuning principal singular components [50] suggests that SoMA better preserves the integrity of pretrained representations while learning task-specific features." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.721, + 0.483, + 0.812 + ], + "angle": 0, + "content": "Tuning granularity. Interestingly, as shown in Tables 8 and 9, our method consistently achieves competitive results compared to SOTA baselines across all tuning granularities, indicating that SoMA can be flexibly configured according to the training budget, task difficulty, or domain gap. The best results are achieved with the default setting." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.821, + 0.308, + 0.836 + ], + "angle": 0, + "content": "5.4. Subject Personalization" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Domain generalized recognition requires consistent processing of inputs from diverse domains, whereas domain generalized generation involves generating outputs across a range of domains. Although large-scale Text-to-Image" + }, + { + "type": "image", + "bbox": [ + 0.579, + 0.075, + 0.837, + 0.304 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.31, + 0.906, + 0.354 + ], + "angle": 0, + "content": "Figure 5. Subject Personalization. 1. A dog gracefully leaping in origami style, 2. A dog in watercolor painting style, and 3. A dog soaring through a digital landscape in vector illustration style." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.369, + 0.907, + 0.657 + ], + "angle": 0, + "content": "(T2I) models have convincingly demonstrated this ability, it can be compromised in subject personalization tasks involving fine-tuning. DreamBooth (DB) [65] introduces subject personalization, which fine-tunes a pre-trained T2I model on a few reference images to enable it to generate new visual concepts. DB updates all parameters within specific blocks, potentially impairing the pre-trained capability to generate images across various domains. Therefore, we explore whether SoMA can be seamlessly integrated into the DB framework to preserve pre-trained knowledge. For this task, we utilize Stable Diffusion XL (SDXL) [59] for the T2I diffusion model. Fig. 5 presents a visual comparison between DB and SoMA, using identical hyperparameter settings and sample seeds. As seen, SoMA, similar to the original SDXL model, can generate high-quality images of diverse visual domains (e.g., style, texture, pose) and exhibits superior identity preservation compared to DB with prior preservation (p.p.) loss [65]. More results with varied prompts and subjects are provided in the Supplemental." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.679, + 0.634, + 0.695 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.907, + 0.901 + ], + "angle": 0, + "content": "This paper presents a multifaceted exploration of the distribution of generalizable components in VFMs—examining them at the weight level, block level, and within training dynamics—and introduces the SoMA framework as an effective means to preserve generalizable components. Contrary to recent approaches that passively refine frozen VFM features, SoMA is integrated into all linear layers to effectively facilitate task adaptation, while minimizing interference with generalizable components by initializing the low-rank adapter with minor singular components. Through extensive experiments, we observe that SoMA significantly outperforms the state-of-the-art methods in DGSS, DGOD, and even subject-driven image generation tasks." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "25609" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.27, + 0.108 + ], + "angle": 0, + "content": "7. Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.116, + 0.484, + 0.162 + ], + "angle": 0, + "content": "This work was supported by the National Research Foundation (NRF) grant funded by the Korea government (MSIT) [RS-2025-00562400] and [RS-2022-NR068754]." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.177, + 0.188, + 0.192 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.201, + 0.484, + 0.257 + ], + "angle": 0, + "content": "[1] Woo-Jin Ahn, Geun-Yeong Yang, Hyun-Duck Choi, and Myo-Taeg Lim. Style blind domain generalized semantic segmentation via covariance alignment and semantic consistence contrastive learning. In CVPR, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.259, + 0.484, + 0.327 + ], + "angle": 0, + "content": "[2] Muhammad Awais, Muzammal Naseer, Salman Khan, Rao Muhammad Anwer, Hisham Cholakkal, Mubarak Shah, Ming-Hsuan Yang, and Fahad Shahbaz Khan. Foundational models defining a new era in vision: A survey and outlook. arXiv:2307.13721, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.329, + 0.483, + 0.384 + ], + "angle": 0, + "content": "[3] Yasser Benigmim, Subhankar Roy, Slim Essid, Vicky Kalogeiton, and Stephane Lathuilière. Collaborating foundation models for domain generalized semantic segmentation. In CVPR, 2024. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.387, + 0.482, + 0.442 + ], + "angle": 0, + "content": "[4] Qi Bi, Jingjun Yi, Hao Zheng, Haolan Zhan, Yawen Huang, Wei Ji, Yuexiang Li, and Yefeng Zheng. Learning frequency-adapted vision foundation model for domain generalized semantic segmentation. In NeurlPS, 2024. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.444, + 0.483, + 0.485 + ], + "angle": 0, + "content": "[5] Qi Bi, Shaodi You, and Theo Gevers. Learning content-enhanced mask transformer for domain generalized urban-scene segmentation. In AAAI, 2024. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.488, + 0.483, + 0.543 + ], + "angle": 0, + "content": "[6] Dan Biderman, Jose Gonzalez Ortiz, Jacob Portes, Mansheej Paul, Philip Greengard, Connor Jennings, Daniel King, Sam Havens, Vitaliy Chiley, Jonathan Frankle, et al. Lora learns less and forgets less. TMLR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.545, + 0.483, + 0.585 + ], + "angle": 0, + "content": "[7] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.588, + 0.483, + 0.642 + ], + "angle": 0, + "content": "[8] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.645, + 0.483, + 0.699 + ], + "angle": 0, + "content": "[9] Gyusam Chang, Jiwon Lee, Donghyun Kim, Jinkyu Kim, Dongwook Lee, Daehyun Ji, Sujin Jang, and Sangpil Kim. Unified domain generalization and adaptation for multi-view 3d object detection. In NeurIPS, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.702, + 0.483, + 0.757 + ], + "angle": 0, + "content": "[10] Prithvijit Chattopadhyay, Kartik Sarangmath, Vivek Vijaykumar, and Judy Hoffman. Pasta: Proportional amplitude spectrum training augmentation for syn-to-real domain generalization. In ICCV, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.759, + 0.483, + 0.814 + ], + "angle": 0, + "content": "[11] Shoufa Chen, Chongjian Ge, Zhan Tong, Jiangliu Wang, Yibing Song, Jue Wang, and Ping Luo. Adaptformer: Adapting vision transformers for scalable visual recognition. In NeurIPS, 2022. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.817, + 0.483, + 0.87 + ], + "angle": 0, + "content": "[12] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In CVPR, 2022. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.483, + 0.902 + ], + "angle": 0, + "content": "[13] Sungha Choi, Sanghun Jung, Huiwon Yun, Joanne T Kim, Seungryong Kim, and Jaegul Choo. Robustnet: Improving" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.201, + 0.484, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.121 + ], + "angle": 0, + "content": "domain generalization in urban-scene segmentation via instance selective whitening. In CVPR, 2021. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.124, + 0.905, + 0.191 + ], + "angle": 0, + "content": "[14] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In CVPR, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.194, + 0.905, + 0.261 + ], + "angle": 0, + "content": "[15] Muhammad Sohail Danish, Muhammad Haris Khan, Muhammad Akhtar Munir, M Saquib Sarfraz, and Mohsen Ali. Improving single domain-generalized object detection: A focus on diversification and alignment. In CVPR, 2024. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.265, + 0.905, + 0.307 + ], + "angle": 0, + "content": "[16] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.31, + 0.905, + 0.364 + ], + "angle": 0, + "content": "[17] Jian Ding, Nan Xue, Gui-Song Xia, Bernt Schiele, and Dengxin Dai. Hgformer: Hierarchical grouping transformer for domain generalized semantic segmentation. In CVPR, 2023. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.368, + 0.905, + 0.45 + ], + "angle": 0, + "content": "[18] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.453, + 0.904, + 0.48 + ], + "angle": 0, + "content": "[19] Carl Eckart and Gale Young. The approximation of one matrix by another of lower rank. Psychometrika, 1936. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.483, + 0.905, + 0.535 + ], + "angle": 0, + "content": "[20] Mohammad Fahes, Tuan-Hung Vu, Andrei Bursuc, Patrick Pérez, and Raoul de Charette. A simple recipe for language-guided domain generalized segmentation. In CVPR, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.54, + 0.905, + 0.595 + ], + "angle": 0, + "content": "[21] Qi Fan, Mattia Segu, Yu-Wing Tai, Fisher Yu, Chi-Keung Tang, Bernt Schiele, and Dengxin Dai. Towards robust object detection invariant to real-world domain shifts. In ICLR, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.599, + 0.905, + 0.64 + ], + "angle": 0, + "content": "[22] Yuxin Fang, Quan Sun, Xinggang Wang, Tiejun Huang, Xinlong Wang, and Yue Cao. Eva-02: A visual representation for neon genesis. IVC, 2024. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.643, + 0.905, + 0.698 + ], + "angle": 0, + "content": "[23] Aditya Sharad Golatkar, Alessandro Achille, and Stefano Soatto. Time matters in regularizing deep networks: Weight decay and data augmentation affect early learning dynamics, matter little near convergence. In NeurIPS, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.7, + 0.905, + 0.769 + ], + "angle": 0, + "content": "[24] Jose L Gómez, Manuel Silva, Antonio Seoane, Agnès Borras, Mario Noriega, German Ros, Jose A Iglesias-Guitian, and Antonio M López. All for one, and one for all: Urbansyn dataset, the third musketeer of synthetic driving scenes. arXiv:2312.12176, 2023. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.772, + 0.905, + 0.813 + ], + "angle": 0, + "content": "[25] Junxian He, Chunting Zhou, Xuezhe Ma, Taylor Berg-Kirkpatrick, and Graham Neubig. Towards a unified view of parameter-efficient transfer learning. In ICLR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.816, + 0.905, + 0.857 + ], + "angle": 0, + "content": "[26] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Delving deep into rectifiers: Surpassing human-level performance on imagenet classification. In ICCV, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.859, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[27] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 2, 3, 7" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "25610" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.134 + ], + "angle": 0, + "content": "[28] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In CVPR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.483, + 0.177 + ], + "angle": 0, + "content": "[29] Lukas Hoyer, Dengxin Dai, and Luc Van Gool. Daformer: Improving network architectures and training strategies for domain-adaptive semantic segmentation. In CVPR, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.179, + 0.482, + 0.233 + ], + "angle": 0, + "content": "[30] Edward J Hu, yelong shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In ICLR, 2022. 1, 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.235, + 0.482, + 0.289 + ], + "angle": 0, + "content": "[31] Jiajun Hu, Jian Zhang, Lei Qi, Yinghuan Shi, and Yang Gao. Learn to preserve and diversify: Parameter-efficient group with orthogonal regularization for domain generalization. In ECCV, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.292, + 0.482, + 0.346 + ], + "angle": 0, + "content": "[32] Wei Huang, Chang Chen, Yong Li, Jiacheng Li, Cheng Li, Fenglong Song, Youliang Yan, and Zhiwei Xiong. Style projected clustering for domain generalized semantic segmentation. In CVPR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.348, + 0.482, + 0.417 + ], + "angle": 0, + "content": "[33] Christoph Hummer, Manuel Schwonberg, Liangwei Zhong, Hu Cao, Alois Knoll, and Hanno Gottschalk. Vltseg: Simple transfer of clip-based vision-language representations for domain generalized semantic segmentation. In ACCV, 2024. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.419, + 0.482, + 0.46 + ], + "angle": 0, + "content": "[34] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In ECCV, 2022. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.462, + 0.482, + 0.53 + ], + "angle": 0, + "content": "[35] Yuru Jia, Lukas Hoyer, Shengyu Huang, Tianfu Wang, Luc Van Gool, Konrad Schindler, and Anton Obukhov. Dgin-style: Domain-generalizable semantic segmentation with image diffusion models and stylized semantic control. In ECCV, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.532, + 0.482, + 0.573 + ], + "angle": 0, + "content": "[36] Xueying Jiang, Jiaxing Huang, Sheng Jin, and Shijian Lu. Domain generalization via balancing training difficulty and model capability. In ICCV, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.575, + 0.482, + 0.616 + ], + "angle": 0, + "content": "[37] Mengmeng Jing, Xiantong Zhen, Jingjing Li, and Cees GM Snoek. Order-preserving consistency regularization for domain adaptation and generalization. In ICCV, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.617, + 0.482, + 0.657 + ], + "angle": 0, + "content": "[38] Christoph Kamann and Carsten Rother. Increasing the robustness of semantic segmentation models with painting-by-numbers. In ECCV, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.66, + 0.482, + 0.702 + ], + "angle": 0, + "content": "[39] Jin Kim, Jiyoung Lee, Jungin Park, Dongbo Min, and Kwanghoon Sohn. Pin the memory: Learning to generalize semantic segmentation. In CVPR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.703, + 0.482, + 0.743 + ], + "angle": 0, + "content": "[40] Sunghwan Kim, Dae-hwan Kim, and Hoseong Kim. Texture learning domain randomization for domain generalized segmentation. In ICCV, 2023. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.746, + 0.482, + 0.786 + ], + "angle": 0, + "content": "[41] Suhyeon Lee, Hongje Seong, Seongwon Lee, and Euntai Kim. Wildnet: Learning domain generalized semantic segmentation from the wild. In CVPR, 2022. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.789, + 0.482, + 0.828 + ], + "angle": 0, + "content": "[42] Sohyun Lee, Taeyoung Son, and Suha Kwak. Fifo: Learning fog-invariant features for foggy scene segmentation. In CVPR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.831, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[43] Wooju Lee, Dasol Hong, Hyungtae Lim, and Hyun Myung. Object-aware domain generalization for object detection. In AAAI, 2024. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[44] Yoonho Lee, Annie S Chen, Fahim Tajwar, Ananya Kumar, Huaxiu Yao, Percy Liang, and Chelsea Finn. Surgical fine" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "tuning improves adaptation to distribution shifts. In ICLR, 2023.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.162 + ], + "angle": 0, + "content": "[45] Brian Lester, Rami Al-Rfou, and Noah Constant. The power of scale for parameter-efficient prompt tuning. In EMNLP, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.165, + 0.905, + 0.206 + ], + "angle": 0, + "content": "[46] Deng Li, Among Wu, Yaowei Wang, and Yahong Han. Prompt-driven dynamic object-centric learning for single domain generalization. In CVPR, 2024. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.208, + 0.905, + 0.248 + ], + "angle": 0, + "content": "[47] Dongze Lian, Daquan Zhou, Jiashi Feng, and Xinchao Wang. Scaling & shifting your features: A new baseline for efficient model tuning. In NeurIPS, 2022. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.251, + 0.905, + 0.305 + ], + "angle": 0, + "content": "[48] Shih-Yang Liu, Chien-Yi Wang, Hongxu Yin, Pavlo Molchanov, Yu-Chiang Frank Wang, Kwang-Ting Cheng, and Min-Hung Chen. DoRA: Weight-decomposed low-rank adaptation. In ICML, 2024. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.307, + 0.905, + 0.36 + ], + "angle": 0, + "content": "[49] Yajing Liu, Shijun Zhou, Xiyao Liu, Chunhui Hao, Baojie Fan, and Jiandong Tian. Unbiased faster r-cnn for single-source domain generalized object detection. In CVPR, 2024. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.364, + 0.905, + 0.405 + ], + "angle": 0, + "content": "[50] Fanxu Meng, Zhaohui Wang, and Muhan Zhang. Pissa: Principal singular values and singular vectors adaptation of large language models. In NeurIPS, 2024. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.407, + 0.905, + 0.447 + ], + "angle": 0, + "content": "[51] Gerhard Neuhold, Tobias Ollmann, Samuel Rota Bulo, and Peter Kontschieder. The mapillary vistas dataset for semantic understanding of street scenes. In ICCV, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.449, + 0.905, + 0.517 + ], + "angle": 0, + "content": "[52] Joshua Niemeijer, Manuel Schwonberg, Jan-Aike Termöhlen, Nico M Schmidt, and Tim Fingscheidt. Generalization by adaptation: Diffusion-based domain extension for domain-generalized semantic segmentation. In WACV, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.519, + 0.905, + 0.587 + ], + "angle": 0, + "content": "[53] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. TMLR, 2023. 3, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.591, + 0.905, + 0.645 + ], + "angle": 0, + "content": "[54] Byeonghyun Pak, Byeongju Woo, Sunghwan Kim, Dae-hwan Kim, and Hoseong Kim. Textual query-driven mask transformer for domain generalized segmentation. In ECCV, 2024. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.647, + 0.905, + 0.687 + ], + "angle": 0, + "content": "[55] Xingang Pan, Ping Luo, Jianping Shi, and Xiaou Tang. Two at once: Enhancing learning and generalization capacities via ibn-net. In ECCV, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.689, + 0.905, + 0.743 + ], + "angle": 0, + "content": "[56] Yuwen Pan, Rui Sun, Naisong Luo, Tianzhu Zhang, and Yongdong Zhang. Exploring reliable matching with phase enhancement for night-time semantic segmentation. In ECCV, 2024.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.746, + 0.905, + 0.788 + ], + "angle": 0, + "content": "[57] Duo Peng, Yinjie Lei, Lingqiao Liu, Pingping Zhang, and Jun Liu. Global and local texture randomization for synthetic-to-real semantic segmentation. TIP, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.789, + 0.905, + 0.829 + ], + "angle": 0, + "content": "[58] Duo Peng, Yinjie Lei, Munawar Hayat, Yulan Guo, and Wen Li. Semantic-aware domain generalized segmentation. In CVPR, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.832, + 0.905, + 0.898 + ], + "angle": 0, + "content": "[59] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. SDXL: Improving latent diffusion models for high-resolution image synthesis. In ICLR, 2024. 3, 8" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.898 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "25611" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.161 + ], + "angle": 0, + "content": "[60] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.165, + 0.482, + 0.207 + ], + "angle": 0, + "content": "[61] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. In NeurIPS, 2015. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.208, + 0.482, + 0.25 + ], + "angle": 0, + "content": "[62] Stephan R Richter, Vibhav Vineet, Stefan Roth, and Vladlen Koltun. Playing for data: Ground truth from computer games. In ECCV, 2016. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.252, + 0.482, + 0.293 + ], + "angle": 0, + "content": "[63] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.296, + 0.482, + 0.35 + ], + "angle": 0, + "content": "[64] German Ros, Laura Sellart, Joanna Materzynska, David Vazquez, and Antonio M. Lopez. The synthia dataset: A large collection of synthetic images for semantic segmentation of urban scenes. In CVPR, 2016. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.353, + 0.482, + 0.408 + ], + "angle": 0, + "content": "[65] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In CVPR, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.411, + 0.482, + 0.452 + ], + "angle": 0, + "content": "[66] Christos Sakaridis, Dengxin Dai, and Luc Van Gool. Acdc: The adverse conditions dataset with correspondences for semantic driving scene understanding. In ICCV, 2021. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.455, + 0.482, + 0.495 + ], + "angle": 0, + "content": "[67] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zh-moginov, and Liang-Chieh Chen. Mobilenetv2: Inverted residuals and linear bottlenecks. In CVPR, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.498, + 0.482, + 0.54 + ], + "angle": 0, + "content": "[68] Fengyi Shen, Akhil Gurram, Ziyuan Liu, He Wang, and Alois Knoll. Diga: Distil to generalize and then adapt for domain adaptive semantic segmentation. In CVPR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.542, + 0.482, + 0.583 + ], + "angle": 0, + "content": "[69] Zhiqiang Tang, Yunhe Gao, Yi Zhu, Zhi Zhang, Mu Li, and Dimitris N Metaxas. Crossnorm and selfnorm for generalization under distribution shifts. In ICCV, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.586, + 0.482, + 0.641 + ], + "angle": 0, + "content": "[70] Sumanth Udupa, Prajwal Gurunath, Aniruddh Sikdar, and Suresh Sundaram. Mrfp: Learning generalizable semantic segmentation from sim-2-real with multi-resolution feature perturbation. In CVPR, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.643, + 0.482, + 0.685 + ], + "angle": 0, + "content": "[71] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.687, + 0.482, + 0.728 + ], + "angle": 0, + "content": "[72] Vidit Vidit, Martin Engilberge, and Mathieu Salzmann. Clip the gap: A single domain generalization approach for object detection. In CVPR, 2023. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.731, + 0.482, + 0.785 + ], + "angle": 0, + "content": "[73] Zhixiang Wei, Lin Chen, Tao Tu, Pengyang Ling, Huaian Chen, and Yi Jin. Disentangle then parse: Night-time semantic segmentation with illumination disentanglement. In ICCV, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.788, + 0.482, + 0.857 + ], + "angle": 0, + "content": "[74] Zhixiang Wei, Lin Chen, Yi Jin, Xiaoxiao Ma, Tianle Liu, Pengyang Ling, Ben Wang, Huaian Chen, and Jinjin Zheng. Stronger fewer & superior: Harnessing vision foundation models for domain generalized semantic segmentation. In CVPR, 2024. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.86, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[75] Among Wu and Cheng Deng. Single-domain generalized object detection in urban scene via cyclic-disentangled self-distillation. In CVPR, 2022. 3, 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.134 + ], + "angle": 0, + "content": "[76] Zhenyao Wu, Xinyi Wu, Xiaoping Zhang, Lili Ju, and Song Wang. Siamdoge: Domain generalizable semantic segmentation using siamese network. In ECCV, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.137, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[77] Qi Xu, Liang Yao, Zhengkai Jiang, Guannan Jiang, Wenqing Chu, Wenhui Han, Wei Zhang, Chengjie Wang, and Ying Tai. Dirl: Domain-invariant representation learning for generalizable semantic segmentation. In AAAI, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.193, + 0.905, + 0.234 + ], + "angle": 0, + "content": "[78] Liwei Yang, Xiang Gu, and Jian Sun. Generalized semantic segmentation by self-supervised source domain projection and multi-level contrastive learning. In AAAI, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.236, + 0.905, + 0.289 + ], + "angle": 0, + "content": "[79] Jingjun Yi, Qi Bi, Hao Zheng, Haolan Zhan, Wei Ji, Yawen Huang, Yuexiang Li, and Yefeng Zheng. Learning spectral-decomposed tokens for domain generalized semantic segmentation. In ACMMM, 2024. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.292, + 0.905, + 0.347 + ], + "angle": 0, + "content": "[80] Fisher Yu, Haofeng Chen, Xin Wang, Wenqi Xian, Yingying Chen, Fangchen Liu, Vashisht Madhavan, and Trevor Darrell. Bdd100k: A diverse driving dataset for heterogeneous multitask learning. In CVPR, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.349, + 0.905, + 0.403 + ], + "angle": 0, + "content": "[81] Yeonguk Yu, Sungho Shin, Seunghyeok Back, Mihwan Ko, Sangjun Noh, and Kyoobin Lee. Domain-specific block selection and paired-view pseudo-labeling for online test-time adaptation. In CVPR, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.406, + 0.905, + 0.473 + ], + "angle": 0, + "content": "[82] Xiangyu Yue, Yang Zhang, Sicheng Zhao, Alberto Sangiovanni-Vincentelli, Kurt Keutzer, and Boqing Gong. Domain randomization and pyramid consistency: Simulation-to-real generalization without accessing target domain data. In ICCV, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.476, + 0.905, + 0.529 + ], + "angle": 0, + "content": "[83] Yuyang Zhao, Zhun Zhong, Na Zhao, Nicu Sebe, and Gim Hee Lee. Style-hallucinated dual consistency learning for domain generalized semantic segmentation. In ECCV, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.532, + 0.905, + 0.573 + ], + "angle": 0, + "content": "[84] Zhun Zhong, Yuyang Zhao, Gim Hee Lee, and Nicu Sebe. Adversarial style augmentation for domain generalized urban-scene segmentation. In NeurIPS, 2022. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.575, + 0.905, + 0.615 + ], + "angle": 0, + "content": "[85] Jinghao Zhou, Chen Wei, Huiyu Wang, Wei Shen, Cihang Xie, Alan Yuille, and Tao Kong. ibot: Image bert pre-training with online tokenizer. arXiv:2111.07832, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.618, + 0.905, + 0.645 + ], + "angle": 0, + "content": "[86] Zhuofan Zong, Guanglu Song, and Yu Liu. Detrs with collaborative hybrid assignments training. In ICCV, 2023. 7" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.645 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "25612" + } + ] +] \ No newline at end of file diff --git a/2025/SoMA_ Singular Value Decomposed Minor Components Adaptation for Domain Generalizable Representation Learning/282504c5-c667-4202-94ca-c92386507635_origin.pdf b/2025/SoMA_ Singular Value Decomposed Minor Components Adaptation for Domain Generalizable Representation Learning/282504c5-c667-4202-94ca-c92386507635_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..d337d9335fb6271a68882fb8b4f56511e0ae97b9 --- /dev/null +++ b/2025/SoMA_ Singular Value Decomposed Minor Components Adaptation for Domain Generalizable Representation Learning/282504c5-c667-4202-94ca-c92386507635_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4d5e7e72fa172fc9a973bc7805f6a9b5e822901dcf7b119a4a77c8930879a348 +size 7981785 diff --git a/2025/SoMA_ Singular Value Decomposed Minor Components Adaptation for Domain Generalizable Representation Learning/full.md b/2025/SoMA_ Singular Value Decomposed Minor Components Adaptation for Domain Generalizable Representation Learning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..6c8bc5d308b815bc7a599b9e3653b907978decdd --- /dev/null +++ b/2025/SoMA_ Singular Value Decomposed Minor Components Adaptation for Domain Generalizable Representation Learning/full.md @@ -0,0 +1,337 @@ +# SoMA: Singular Value Decomposed Minor Components Adaptation for Domain Generalizable Representation Learning + +Seokju Yun Seunghye Chae Dongheon Lee Youngmin Ro* Machine Intelligence Laboratory, University of Seoul, Korea + +{wsz871, tmdhey, dslisleedh, youngmin.ro}@uos.ac.kr +https://github.com/ysj9909/SoMA + +![](images/3628e235ba0f324223b527313dafe6b96b4369108834212fbc94f57ac5ccc19f.jpg) +Figure 1. Overview of SoMA framework. Left. SoMA achieves state-of-the-art results across diverse tasks, ranging from domaingeneralized semantic segmentation (DGSS) to object detection (DGOD), and performs well in both synthetic-to-real and real-to-real scenarios. Middle. Our method, trained solely on synthetic datasets, demonstrates strong generalization capabilities in complex real-world scenes. Right. We present SoMA, a method that applies SVD to the pre-trained weights, decomposing them into $r$ minor singular components $U_{[:,-r:]} \Sigma_{[-r:]} (V^T)_{[-r:,]}$ and residuals. SoMA selectively tunes the smallest $r$ components, effectively preserving world knowledge of foundation models. Since SoMA shares the same forward architecture as LoRA [30], it adds no extra latency during inference phase. + +![](images/49983582969fbd7f5d57332555baa5e2f0fd28824735a4ba3bba883fe850aabb.jpg) + +![](images/c40ea8d126a3176fbedf8dc89cc74e2b9af90901a70fd337c29b9cec168109db.jpg) + +![](images/5d4165514b4bb3db4a6d890ec65f07d65fa688179d0a15423ea0523abd4ce689.jpg) + +![](images/794488922f066aff2e1e6557ceeb047b3362fbbdc3075f5cbf3e8df5bbe3a009.jpg) + +# Abstract + +Domain generalization (DG) aims to adapt a model using one or multiple source domains to ensure robust performance in unseen target domains. Recently, Parameter-Efficient Fine-Tuning (PEFT) of foundation models has shown promising results in the context of DG problem. Nevertheless, existing PEFT methods still struggle to strike a balance between preserving generalizable components of the pre-trained model and learning task-specific features. To gain insights into the distribution of generalizable components, we begin by analyzing the pre-trained weights through the lens of singular value decomposition. Building on these insights, we introduce Singular Value Decomposed Minor Components Adaptation (SoMA), an approach that selectively tunes minor singular components while keeping the residual parts frozen. SoMA effectively retains the generalization ability of the pre-trained model while efficiently acquiring task-specific skills. Moreover, we freeze domain-generalizable blocks and employ an annealing weight decay + +strategy, thereby achieving an optimal balance in the delicate trade-off between generalizability and discriminability. SoMA attains state-of-the-art results on multiple benchmarks that span both domain generalized semantic segmentation to domain generalized object detection. In addition, our methods introduce no additional inference overhead or regularization loss, maintain compatibility with any backbone or head, and are designed to be versatile, allowing easy integration into a wide range of tasks. + +# 1. Introduction + +Deep neural networks (DNNs) have recently made significant strides in dense prediction tasks, including semantic segmentation [12] and object detection [7, 61], both of which are crucial for safety-critical applications, such as autonomous driving. However, DNNs often fail to maintain reliable performance under domain shifts, which may result from diverse lighting conditions, weather changes, or differences in location. Likewise, while synthetic data [24, 62, 64] is employed to avoid the labor-intensive and costly process of constructing real-world datasets, the resulting performance degradation of DNNs in real-world deployment + +remains a critical issue to be resolved. To tackle these challenges, Domain Generalization (DG) is introduced to design models capable of consistent prediction on arbitrary unseen target domains. + +Existing DG methods for dense prediction tasks employ two main strategies to enhance model robustness. The first involves diversifying the training process in the input [10, 57] or feature space [21, 70] using augmentation/perturbation techniques, exposing the model to a broader range of styles to mitigate overfitting to specific domains. The second strategy focuses on imposing alignment constraints, such as normalization [58] or regularization losses [1, 13], to facilitate the learning of domain-invariant features. However, these methods mostly utilize outdated backbones (e.g., ResNet [27] or MobileNetV2 [67]) pretrained on mid-scale datasets, such as ImageNet [16]. + +Meanwhile, vision foundation models (VFMs) [2], pretrained on vast, curated datasets with enormous parameter counts, have recently shown unprecedented generalization capabilities [74]. Given this landscape, we propose a paradigm shift—from applying the diversify & align approach to classic backbones, to devising methods that preserve the world knowledge of VFMs while effectively learning task-specific features. However, naive adaptation of VFMs via full fine-tuning (FFT), which involves retraining all model parameters, results in prohibitive costs and a risk of catastrophic forgetting of the pre-trained knowledge. + +In light of these challenges, a commonly adopted approach is parameter-efficient fine-tuning (PEFT), where only the injected lightweight adapters [11, 25, 30] or tokens [34, 45] are fine-tuned, while the pre-trained weights remain frozen. Low-Rank Adaptation (LoRA) [30] is one of the most widely used PEFT methods, which substitutes the updates with the product of two low-rank matrices. Interestingly, we found that LoRA exhibits superior generalization ability in DG dense prediction tasks compared to FFT and other PEFT methods (see Tab. 2), aligning with recent findings in LoRA-based literature [6, 31]. Nevertheless, LoRA adapters are built without considering the distribution of generalizable components within the pre-trained weights they modify. As a result, there is still room for enhancing generalizability by preserving the diverse domain knowledge of VFMs. + +To uncover the structure in which world knowledge is encoded within the pre-trained weights, we first perform singular value decomposition (SVD) on the learned weight matrices and then analyze the behavior of VFMs after selectively removing specific singular components1. Our analysis reveals that singular vectors associated with higher singular values tend to extract general features spanning multiple classes in ImageNet, while minor singular com + +ponents are primarily responsible for capturing context-specific features. Capitalizing on our findings, we introduce Singular Value Decomposed Minor Components Adaptation (SoMA), which starts by performing SVD to decompose the weights into their singular vectors, i.e., $W = U\Sigma V^T$ , then fine-tunes only the minor singular components. Specifically, we initialize LoRA adapter using the components with the smallest $r$ singular values, i.e., $U_{[:,-r:]} \Sigma_{[-r:]}(V^T)_{[-r:,]}$ , and the remaining residual components are frozen to maintain generalizability, as illustrated in Fig. 1 right. + +By extending our analysis from the weight-level to the block-level, we observe and empirically demonstrate that the early blocks of VFMs effectively extract well-localized semantic features and are less affected by style or input-level domain shifts (see Fig. 3). Therefore, we freeze these blocks to preserve generalizable components while reducing the number of trainable parameters. Although our proposed methods are effective, tuning the less-optimized bottom spectral space and freezing early blocks may lead to a lack of discriminability. To mitigate this issue, we introduce an annealing weight decay scheme that gradually reduces the regularization loss incurred by weight decay over the course of training. As shown in Fig. 1, comprehensive evaluations on DG benchmarks demonstrate that our framework outperforms previous state-of-the-art baselines and FFT, training only $0.58\%$ to $1.6\%$ of the parameters. Furthermore, SoMA demonstrates impressive scaling performance with both data and model size. + +We make the following contributions: + +- We underscore the critical importance of preserving pretrained knowledge while learning task-specific features in scenarios that involve the use of VFMs. +- We present SoMA, a novel PEFT method specifically tailored for DG problem, which leverages spectral information to retain the generalization capacity of VFMs. In addition, we identify and freeze domain-generalizable blocks while employing an annealing weight decay strategy, thereby achieving an optimal balance between generalizability and discriminability. +- Extensive experiments on various DG for semantic segmentation and object detection tasks reveal that our framework surpasses existing baselines by a substantial margin. Notably, SoMA achieves an mIoU of $80.4\%$ on Cityscapes without accessing real-world datasets, while training only $0.58\%$ of the parameters. Additionally, under the most challenging daytime-sunny $\rightarrow$ night-rainy detection setting, SoMA outperforms the prior SOTA implementation by achieving a $+23.5\%$ mAP gain. + +# 2. Related Work + +Domain generalization for dense predictions. Domain Generalization (DG) methods for dense predictions [5, 9, + +17, 32, 36, 39, 42, 46, 56, 68, 73, 75, 78] have recently garnered considerable attention due to their practical demands. These methods can be categorized along two main axes: i) data augmentation or domain randomization, both of which diversify the training process, and ii) alignment techniques to suppress domain-relevant features. + +Methods adopting the first approach at the input level employ various augmentation techniques [10, 38, 57, 84] or, more recently, leverage advances in generative modeling, such as diffusion models [59, 63], to generate data approximating the target domain [3, 35, 52], thereby expanding domain diversity and enriching learned representations. To enable more flexible augmentation, several studies [20, 21, 41, 70, 72] leverage feature stylization to enhance model robustness. On the other hand, alignment-based methods effectively reduce domain-sensitive components by employing feature normalization/whitening techniques [55, 58, 69], or by introducing regularization losses to suppress inconsistencies caused by simulated domain shifts [1, 13, 15, 37, 40, 43, 49, 76, 77, 82, 83]. However, most of the aforementioned approaches experimented with CNN backbones [27] trained on relatively restricted domains. In contrast, when applying vision foundation models (VFMs)—trained on massive datasets using transformer architectures [18, 71] and sophisticated training schemes [53]—to the DG problem, a fundamental shift in approach is imperative. + +Following this trend, recent studies leverage vision-language models as feature extractors [33], employing text embeddings with domain-invariant semantics as a source for style augmentation [20, 72] or as object queries in transformer-based decoders [54]. Rein [74] injects learnable adapters and tokens to refine feature maps for each instance while keeping the backbone frozen, significantly expanding the performance frontier. FADA [4] and SET [79] introduce frequency-adapted methods to effectively utilize frozen VFM features. In contrast to existing methods that simply freeze all parameters, we perform singular value decomposition on the pre-trained weights, freezing the generalizable principal components and tuning only the minor singular components responsible for context-specific features, thereby achieving superior DG performance. + +Vision foundation models (VFMs). Recently, the advent of VFMs with general perception capabilities has laid the foundation for a paradigm in which these models are broadly applicable to numerous downstream tasks. Seminal examples of models advancing this direction include CLIP [60], which demonstrates strong zero-shot generalizability through training on web-scale, weakly supervised image-text pairs; EVA02 [22], which enhances CLIP features with masked image modeling [28]; and DINOv2 [53], which incorporates losses from prior arts like iBoT [85] and DINO [8], and is trained on massive curated datasets, offer + +ing strong spatial features for dense prediction tasks. Our SoMA builds on this foundation, harnessing the full potential of VFMs in the context of DG problem. + +Parameter-efficient fine-tuning (PEFT). With the remarkable success of PEFT methods in efficiently adapting large-scale foundation models in the realm of natural language processing, there has been a growing interest in extending these approaches to the field of computer vision. For example, VPT [34] prepends prompt tokens to the input sequence of several attention layers. AdaptFormer [11] introduces a lightweight adapter in parallel to the feed-forward network, whereas SSF [47] incorporates scale and shift parameters to modulate features for downstream tasks. LoRA [30] models the incremental updates of pre-trained weights using rank decomposition matrices, achieving performance comparable to full fine-tuning (FFT). However, there still often exists a performance gap between LoRA and FFT. DoRA [48] and PiSSA [50] address this gap by further decomposing the pre-trained weights, initializing the LoRA adapter with directional and principal components, respectively, and fine-tuning accordingly. While these methods aim to match the performance of FFT, our focus is on the preservation of generalizable components. + +# 3. Preliminaries + +Low-Rank Adaptation (LoRA). LoRA [30] hypothesizes that the weight changes during fine-tuning exhibit a low-rank structure. Given a pre-trained weight matrix $W_0 \in \mathbb{R}^{m \times n}$ , LoRA constrains its update $\Delta W \in \mathbb{R}^{m \times n}$ to a low-rank decomposition $\Delta W = BA$ , where $B \in \mathbb{R}^{m \times r}$ and $A \in \mathbb{R}^{r \times n}$ are two low-rank matrices, with an intrinsic rank of $r \ll \min(m, n)$ . For $y = W_0 x$ , the modified forward pass is as follows: + +$$ +y = \left(W _ {0} + \Delta W\right) x = \left(W _ {0} + \underline {{B A}}\right) x, \tag {1} +$$ + +where $W_{0}$ is frozen, and only the underlined low-rank parameters are fine-tuned, significantly reducing the number of trainable parameters. LoRA uses a uniform Kaiming distribution [26] to initialize $A$ , while $B$ is initially set to zero, ensuring $BA = 0$ at the beginning of training. As seen in Eq. 1, the learned matrices $BA$ can be merged with the frozen weight $W_{0}$ , allowing LoRA-based variants to avoid adding any additional inference burden. + +Singular Value Decomposition (SVD) Analysis. We observe that LoRA and existing PEFT methods, by training only a relatively small subset of parameters or by enforcing low-rank-ness of weight updates, deliver promising DG performance without significantly distorting the pre-trained representations. However, many previous studies build adapters without considering how the generalizable components are structured within the weight matrices, potentially interfering with these components in the fine-tuning process. Building on the Eckart-Young-Mirsky the + +![](images/e96b2d5afd82c26f5ee4cb84a1541c58f76df86f4626494ad4971d542f8dd724.jpg) + +![](images/59d361949c111fda71f05b4d52a8a3a70a849e82e159392981be01f3e1a906d0.jpg) +Figure 2. Distribution of generalizable components. Top. Number of classes exhibiting specific accuracy drops after applying SVD to DINOv2-large weights and reconstructing by truncating the smallest $r$ singular components. Bottom. Distinct roles of singular components across levels. Numbers in parentheses represent the count of classes with an accuracy drop ratio exceeding $50\%$ . The average WordNet hierarchy depth of these classes is shown, with higher values indicating greater context specificity. + +orem [19], which validates that the optimal rank- $r$ approximation of a matrix $W$ is represented by the sum of its top $r$ singular components, we analyze how the world knowledge of VFMs is structured from an SVD perspective. Using ImageNet-1k [16] as a testbed, our analysis computes the SVD of pre-trained weights across all layers of the DINOv2 [53] ViT-large model, followed by truncating specific singular values and their corresponding singular vectors to investigate the classes that are subsequently misclassified. + +As shown in Fig. 2 top, as the number of discarded smallest singular components increases, the number of misclassified classes grows exponentially rather than linearly, suggesting that components with larger singular values tend to capture more general features spanning multiple classes. To further examine whether different groups of singular components exhibit distinct behaviors, we consider three groups: the top 8 principal components $U_{[:;8]}\Sigma_{[:8]}(V^T)_{[:8,:]}$ , the middle 160 singular components $U_{[:384:544]}\Sigma_{[384:544]}(V^T)_{[384:544,:]}$ , and the bottom 320 minor components $U_{[:,-320:]} \Sigma_{[-320:]}(V^T)_{[-320,:]}$ . We then truncate each group separately and analyze the classes that the rank-reduced model subsequently fails to classify (see Fig. 2 bottom). + +According to the Eckart-Young-Mirsky theorem and the heavy-tail distribution of singular values, it is evident that the top singular components encapsulate the core knowledge of VFM and should be preserved to maintain generalizability. Conversely, we note that the bottom-componenttruncated model fails to recognize classes that require finegrained or texture-biased features (e.g., specific dog breeds, + +wildlife with distinctive textures), as well as those that are context-specific (e.g., cornet, bikini, missile, dock). In contrast, we find that removing middle singular components results in significant performance declines for classes that rely on coarse-grained or shape-biased recognition (e.g., mortarboard, pier, altar, necklace, schooner, horizontal bar). Interestingly, as suggested by differences in the depth of the WordNet hierarchy, higher-value singular components do not function independently but rather interact in a hierarchical and composite manner. For instance, classes with notable accuracy drops can also be clustered into higher-level concepts, such as functionality (e.g., wok, hotpot, pitcher, cauldron, pop bottle, cup/ambulance, minivan, passenger car, police van, garbage truck, limousine) or co-occurrence (e.g., desktop computer, keyboard, notebook, desk). Additionally, almost all of these objects are commonly found in everyday scenes. Consequently, it can be seen that generalizable components are hierarchically and compositionally structured according to the magnitude of singular values. + +Problem Formulation. The entire model consists of a feature extractor $\mathcal{F}$ , which combines a VFM parameterized by $\Phi$ with our SoMA adapter parameterized by $\theta_{S}$ , and a task-specific decode head $\mathcal{H}$ , parameterized by $\theta_{h}$ . The optimization objective can be expressed as follows: + +$$ +\underset {\theta_ {S}, \theta_ {h}} {\arg \min } \sum_ {i = 1} ^ {N _ {s}} \mathcal {L} _ {t a s k} \left(\mathcal {H} _ {\theta_ {h}} \left(\mathcal {F} _ {\Phi , \theta_ {S}} \left(x _ {i}\right)\right), y _ {i}\right), \tag {2} +$$ + +where $(x_{i},y_{i})$ denotes an input image and its associated ground truth, drawn from the source domain datasets $\mathcal{D}_s$ , comprising $N_{s}$ samples. Our proposed SoMA framework facilitates generalizing task-specific skills learned from $\mathcal{D}_s$ to the unseen target domain datasets $\mathcal{D}_t$ . + +# 4. Proposed Method + +In this section, we present our parameter-efficient adaptation methods specifically tailored for DG dense prediction tasks, inspired by the insights from our SVD analysis. Subsequently, we propose a block freeze strategy and an annealing weight decay scheme to strike an optimal trade-off between generalizability and discriminability. + +# 4.1. Minor Singular Components Adaptation + +To effectively preserve the integrity of generalizable representations in pre-trained weights, we propose Singular value decomposed Minor components Adaptation (SoMA). Firstly, SoMA performs singular value decomposition (SVD) on the pre-trained weight matrices, including those from the attention layers, multi-layer perceptron (MLP) layers, and more. The weight matrix $W \in \mathbb{R}^{m \times n}$ is decomposed using SVD as follows: + +$$ +W = U \Sigma V ^ {T} = \sum_ {i = 1} ^ {R} \sigma_ {i} u _ {i} v _ {i} ^ {T}, \tag {3} +$$ + +![](images/3ecc3b511cc76cad4eae19078ee87e110856076e5eaf3e9ad5efb87bb8f39f3d.jpg) +Figure 3. The inherent generalization capabilities of the early blocks of VFM. Top. We apply PCA on the extracted intermediate features (8th block of DINOv2-Large) and visualize the top three leading components. Bottom. Class-wise IoU comparison for rare classes [29] under the GTAV $\rightarrow$ Cityscapes setting, focusing on the impact of freezing the first eight blocks. "NFEB" stands for the Number of Frozen Early Blocks. A darker shade of blue indicates enhanced generalization performance relative to the baseline. + +where $U = [u_{1}, u_{2}, \dots, u_{m}] \in \mathbb{R}^{m \times m}$ and $V = [v_{1}, v_{2}, \dots, v_{n}] \in \mathbb{R}^{n \times n}$ are the left and right singular vectors, whose columns form an orthonormal basis for $\mathbb{R}^m$ and $\mathbb{R}^n$ respectively, $\Sigma \in \mathbb{R}^{m \times n}$ is a diagonal matrix whose entries $\sigma_{i}$ are singular values arranged in descending order, and $R$ denotes the rank of $W$ , with $R \leq \min(m, n)$ . + +We then divide the singular components into two groups according to their singular values: the minor singular components with the smallest $r$ singular values $U_{[:,-r:]} \Sigma_{[-r:]} (V^T)_{[-r:]}$ and the residual singular components $U_{[:,-r]} \Sigma_{[:-r]} (V^T)_{[:-r:]}$ . As illustrated in Fig. 1 right, we employ the QR-type reconstruction of minor singular components to initialize the adapter for fine-tuning: + +$$ +B = U _ {[:, - r: ]} \sqrt {\Sigma} _ {[ - r: ]} \in \mathbb {R} ^ {m \times r}, \tag {4} +$$ + +$$ +A = \sqrt {\Sigma} _ {[ - r: ]} \left(V ^ {T}\right) _ {[ - r:,.. ]} \in \mathbb {R} ^ {r \times n}. \tag {5} +$$ + +In the fine-tuning process, the frozen residual matrix $W_{res}$ is defined as $W - BA$ to prevent numerical errors introduced during the SVD step, while maintaining the pretrained generalization capability at the beginning of fine-tuning. After fine-tuning, the learned low-rank matrices $B'$ and $A'$ can be merged into $W_{res}$ , as in LoRA [30], ensuring no additional computation overhead during inference: + +$$ +y = \left(W \underbrace {- B A + B ^ {\prime} A ^ {\prime}} _ {\Delta W _ {S o M A}}\right) x = \left(W _ {r e s} + B ^ {\prime} A ^ {\prime}\right) x = W ^ {\prime} x. \tag {6} +$$ + +Importantly, SoMA offers several key advantages for tackling the DG problem: i) SoMA initializes the adapter using components orthogonal to the principal components, thus minimizing interference with pre-trained representations. For the batched version of the forward pass described in Eq. 1 (with $X \in \mathbb{R}^{n \times b}$ and $Y \in \mathbb{R}^{m \times b}$ ), the gradients of $B$ and $A$ are $\frac{\partial \mathcal{L}}{\partial B} = \frac{\partial \mathcal{L}}{\partial Y} X^T A^T$ and $\frac{\partial \mathcal{L}}{\partial A} = B^T \frac{\partial \mathcal{L}}{\partial Y} X^T$ . + +
SMRBlockiΔWSoMAΔW*SoMAΔWLoRA
|uiTΔVvi/σi12th0-2550.0750.084 (↓17%)0.101
256-5110.0940.094 (↓43%)0.166
512-7670.1570.165 (↓40%)0.275
768-10234.1046.152 (↑21%)5.095
24th0-2550.0970.095 (↓24%)0.125
256-5110.1050.108 (↓35%)0.166
512-7670.2060.189 (↓41%)0.323
768-10232.7813.620 (↑26%)2.865
+ +Table 1. Singular modulation ratio comparison. * indicates the adaptation matrix trained using the annealing weight decay strategy. The weight matrices are taken from the query projection layers of the 12th and 24th attention blocks in DINOv2-Large [53]. + +Therefore, unlike LoRA, which randomly initializes $A$ , our approach facilitates convergence along singular directions that preserve the integrity of the generalizable components. ii) Consistent with prior DG methods [13, 40, 41, 84] that guide models to prioritize shape over texture, thereby achieving superior out-of-distribution performance, SoMA effectively preserves the domain generalizable shape-biased features of VFMs by tuning only the minor singular components (see Fig. 2 bottom and Tab. 1). + +Discussion. To assess whether the proposed adaptation matrix $\Delta W_{SoMA}$ minimally correlates with the top singular directions of the pre-trained weight matrix $W_0$ , we project $\Delta W$ onto the singular vectors of $W_0$ . Specifically, we quantify the extent of interference by $\Delta W$ on each $i$ -th singular direction through a singular modulation ratio (SMR), which is formally defined as follows: + +$$ +\mathrm {S M R} _ {i} = \left| \frac {u _ {i} ^ {\mathrm {T}} \Delta W v _ {i}}{\sigma_ {i}} \right|. \tag {7} +$$ + +For ease of analysis, we partition the 1024 singular directions into four equal groups, calculate the average SMR for each group, and then compare the resulting values between SoMA and LoRA. As shown in Tab. 1, the SoMA initialization scheme proves more effective than LoRA in minimizing interference with singular components with higher singular values. In summary, our method effectively preserves the structure of the VFM's generalizable representations while adapting to downstream tasks, resulting in significant improvements on DG benchmarks compared to SOTA baselines, as confirmed by our experimental results. + +# 4.2. Freezing Early Blocks + +In this section, we explore the distribution of generalizable components at the block level. Drawing on recent domain adaptation methods [44, 81], which find that the early blocks of a model primarily engage in domain-specific feature extraction and should therefore be updated, we conjecture that in the context of leveraging VFM for DG, the early blocks should instead remain unchanged. Intuitively, these initial blocks of VFMs are adept at projecting low-level features from diverse domains into domain-invariant semantic representations. In other words, while adjusting + +early blocks may help bridge input-level domain gaps in domain adaptation, freezing the early blocks of VFMs is essential for DG to prevent overfitting to the source domain and to retain the model's generalization capability. + +To validate our hypothesis, we perform PCA analysis on the features extracted from the early blocks of VFM and compare the performance on rare classes when these blocks are frozen, as illustrated in Fig. 3. Despite considerable domain shifts in style and content, the principal components align well with the image layout and maintain consistent semantic information (e.g., the horses have similar colors). Furthermore, simply freezing the early blocks yields a notable improvement in IoU for rare classes, indicating that updating these blocks can severely disrupt the learned representations of various classes in VFMs. Thus, by default, we opt to freeze all blocks up to the one that generates the feature map fed as the highest resolution input to the segmentation/detection head (e.g., the first eight blocks). Detailed ablations of varying the number of frozen early blocks are presented in the Supplemental. + +# 4.3. Annealing Weight Decay + +Our methods are specialized in preserving the pre-trained knowledge of VFMs, but this may come at the cost of diminished discriminability. One can remove weight decay or set its coefficient to a very small value to enhance discriminability. However, doing so risks increased interference of the tuned parameters with the principal components of the pre-trained weights. Therefore, based on the observation that regularization primarily affects the early phases of learning by guiding a model towards having good generalization properties [23], we leverage an Annealing Weight Decay (AWD) scheme. Specifically, AWD starts with a relatively large weight decay coefficient and progressively reduces it to zero as training progresses. By default, we use a cosine schedule for AWD. As shown in Tab. 1 ( $\Delta W_{SoMA}$ vs. $\Delta W_{SoMA}^{*}$ ) and Tab. 7, AWD integrates seamlessly with SoMA, improving discriminability without compromising SoMA's ability to preserve pre-trained knowledge. + +# 5. Experiments + +We conduct a variety of experiments to showcase the efficacy of SoMA on DG benchmarks including domain generalized semantic segmentation (DGSS) and domain generalized object detection (DGOD). We then perform an ablation study on SoMA, emphasizing the contribution of each component to the DG performance. Lastly, we extend our approach beyond dense predictions to generative modeling. + +# 5.1. Experimental Setup + +Datasets. For DGSS, we use three synthetic datasets (GTAV [62], SYNTHIA [64], UrbanSyn [24]), and four real-world datasets (Cityscapes [14], BDD100K [80], Map- + +
Synthetic-to-Real GeneralizationTest Domains (mIoU in %)
MethodsBackboneParams.*→Citys.→BDD→Map.Avg.
Single-source DGSS Trained on GTAV
○ CLOUDS [3]CLIP-CN-L0.0M60.2057.4067.0061.50
○ VLTSeg [33]EVA02-L304.2M65.3058.3066.0063.20
○ Rein [74]EVA02-L3.0M65.3060.5064.9063.60
○ FADA [4]EVA02-L11.7M66.7061.9066.1064.90
○ tqdm [54]EVA02-L304.2M68.8859.1870.1066.05
○ SoMA (Ours)EVA02-L5.1M68.0560.8168.3365.73
○ SoMA (Ours)EVA02-L5.1M69.9462.4868.3366.92
○ DoRA [48]DINOv2-L7.5M66.1259.3167.0764.17
○ VPT [34]DINOv2-L3.7M68.7558.6468.3265.24
○ SET [79]DINOv2-L6.1M68.0661.6467.6865.79
○ FADA [4]DINOv2-L11.7M68.2361.9468.0966.09
○ AdaptFormer [11]DINOv2-L6.3M70.1059.8168.7766.23
○ SSF [47]DINOv2-L0.5M68.9761.3068.7766.35
○ LoRA [30]DINOv2-L7.3M70.1360.1370.4266.89
○ Rein† [74]DINOv2-L3.0M70.6862.5169.6167.60
○ SoMA (Ours)DINOv2-L4.9M71.8261.3171.6768.27
○ SoMA (Ours)DINOv2-L4.9M73.6363.3370.9869.31
Multi-source DGSS Trained on GTAV + SYNTHIA
○ Rein† [74]DINOv2-L3.0M72.1761.5370.6968.13
○ SoMA (Ours)DINOv2-L4.9M73.1661.9072.7369.26
○ SoMA (Ours)DINOv2-L4.9M74.8563.5973.9270.79
Multi-source DGSS Trained on GTAV + SYNTHIA + UrbanSyn
○ FFTDINOv2-L304.2M75.9060.9372.8069.88
○ SoMA (Ours)DINOv2-L4.9M77.3362.7874.9371.68
○ FFT‡DINOv2-L307.3M77.0661.8175.0971.32
○ Rein‡ [74]DINOv2-L3.0M78.4262.2074.4971.70
○ SoMA‡ (Ours)DINOv2-L4.9M79.2263.8476.3073.12
○ FreezeDINOv2-G0.0M76.0861.9872.2370.10
○ FFTDINOv2-G1.1B76.9061.6973.5370.71
○ SoMA (Ours)DINOv2-G6.6M78.3963.7575.1672.43
○ SoMA (Ours)DINOv2-G6.6M80.3765.6776.1874.07
+ +Table 2. Comparison of the proposed SoMA with existing DGSS $\circ$ and PEFT $\circ$ methods under various synthetic-to-real settings. + +illary [51], ACDC [66]). Specifically, GTAV, a large-scale dataset generated from the game engine, comprises 24,966 images. SYNTHIA consists of 9,400 photo-realistic synthetic images. UrbanSyn provides 7,539 driving scene images. Cityscapes is an autonomous driving dataset containing 2,975 training and 500 validation images $(2048\times 1024)$ . BDD and Mapillary have 1,000 $(1280\times 720)$ and 2,000 $(1920\times 1080)$ validation images, respectively. ACDC is a driving scene dataset under adverse conditions, including night, snow, fog, and rain. For DGOD, we use the urban scene detection dataset introduced by [75]. The dataset consists of five different weather conditions: Daytime-Sunny (DS), Night-Clear (NC), Night-Rainy (NR), Dusk-Rainy (DR), and Daytime-Foggy (DF). Daytime-Sunny serves as the source domain, offering 26,518 images, with 8,313 used for in-domain evaluation, while the remaining adverse weather conditions are employed as unseen target domains. Implementation details. Since our method focuses on preserving pre-trained knowledge, we mainly evaluate its effectiveness using VFMs like EVA02 [22] and DINOv2 [53] as the backbone. To further elevate perfor + +
Real-to-Real GeneralizationTest Domains (mIoU in %)
MethodsBackboneParams.*→BDD→Map.Avg.
Single-source DGSS Trained on Cityscapes
○ HGFormer [17]Swin-L196.0M61.5072.1066.80
○ CMFormer [5]Swin-L196.0M62.6073.6068.10
○ tqdm [54]EVA02-L304.2M64.7276.1570.44
○ FADA [4]DINOv2-L11.7M65.1275.8670.49
○ Rein† [74]DINOv2-L3.0M66.5375.1870.86
○ SoMA (Ours)DINOv2-L4.9M67.0276.4571.74
• SoMA (Ours)DINOv2-L4.9M68.0877.8772.98
+ +Table 3. Real-to-real DGSS comparison. + +
Clear-to-Adverse Weather MethodsACDC [66] Test Domains (mIoU in %)
→Night→Snow→Fog→RainAll
Single-source DGSS Trained on Cityscapes
○ HGFormer [17]52.768.669.972.067.2
○ SoMA (Ours)61.777.374.777.874.4
○ VLTSeg‡ [33]----77.9
○ Rein‡ [74]70.679.576.479.477.6
○ FFT‡68.980.376.979.777.7
○ SoMA‡ (Ours)73.279.876.880.278.8
+ +Table 4. Results on Cityscapes $\rightarrow$ ACDC test set. + +
Data EfficiencyTest Domains (mIoU in %)
MethodsBackboneParams.*→Citys.→BDD→Map.Avg.
DGSS Pre-trained on GTAV + SYNTHIA + UrbanSyn → 1/16 of Cityscapes
○ FFT‡DINOv2-L307.3M81.5365.2275.7374.16
○ Rein‡ [74]DINOv2-L3.0M82.5864.7673.7373.69
○ SoMA‡ (Ours)DINOv2-L4.9M82.5066.9977.0275.50
+ +mance, we leverage state-of-the-art decode heads such as Mask2Former [12] and Co-DETR [86]. Extensive experiments with various backbones and heads are provided in the Supplemental. By default, SoMA is applied to all linear layers within the self-attention and MLP, with a rank of 16. Detailed training settings are presented in the Supplemental. + +# 5.2. Comparison with State-of-the-Arts + +We compare SoMA with previous SOTA baselines, achieving the highest performance across all DG scenarios without adding inference latency or regularization loss. In our tables, * indicates trainable parameters in the backbones. Marker • refers to models tested with multi-scale flip augmentation. † represents re-implemented test results using official checkpoints for a fair comparison, and ‡ denotes training on images cropped to $1024 \times 1024$ ; otherwise, the default is $512 \times 512$ . A comprehensive comparison with earlier works can be found in the Supplemental. + +Synthetic-to-real DGSS. In Tab. 2, we adopt $GTAV \rightarrow \{C_i\}$ as the basic setting, and demonstrate scalability of SoMA by incorporating additional synthetic datasets into the training process. We report the mean Intersection over Union (mIoU) and benchmark our results against SOTA methods. SoMA consistently outperforms + +Table 5. Real-world data efficiency. + +
Clear-to-Adverse WeatherS-DGOD [75] Test Domains (mAP@0.5 in %)
MethodsParams.*DS→NC→DR→NR→DFAvg.
Single-source DGOD Trained on Daytime-Sunny (DS)
Backbone: ResNet101 [27] / Head: Faster R-CNN [61]
S-DGOD [75]42.3M56.136.628.216.633.528.7
CLIP-Gap [72]42.3M51.336.932.318.738.531.6
OA-DG [43]42.3M55.838.033.916.838.331.8
PDOC [46]42.3M53.638.533.719.239.132.6
UFR [49]42.3M58.640.833.219.239.633.2
DivAlign [15]42.3M52.842.538.124.137.235.5
SoMA (Ours)3.1M49.341.937.924.538.235.6
Backbone: DINOv2-L [53] / Head: Co-DETR [86]
Freeze0.0M65.054.255.042.846.949.7
FFT307.3M68.257.156.643.147.251.0
DoRA [48]5.8M69.058.758.045.048.952.7
AdaptFormer [11]6.3M68.958.858.344.449.852.8
LoRA [30]5.5M69.659.658.146.149.553.3
SoMA (Ours)4.9M69.459.359.347.651.054.3
+ +Table 6. Domain generalized object detection. + +VFM-based DGSS methods and PEFT approaches using both vision-language (EVA02) and vision-only (DINOv2) VFM backbones. Notably, SoMA achieves superior results over LoRA by minimally interfering with the hierarchical structure of VFM's pre-trained knowledge. Moreover, SoMA exhibits superior scalability across data, model, and input size, with improvements becoming more pronounced when test-time augmentation is applied. We provide comparisons regarding model efficiency in the Supplemental. + +Real-to-real DGSS. In Tables 3 and 4, we conduct experiments under the Citys. $\rightarrow$ {BDD, Map.}/ACDC settings. When combined with DINOv2-L backbone, SoMA consistently gains the best results across all settings. This highlights the effectiveness of our method as a strong solution for real-world deployment scenarios, where domain shifts caused by geographic or weather variations are common. + +Real-world data efficiency. Following [74], we evaluate models under constrained conditions using a limited set of real-world images. Specifically, models pre-trained on synthetic datasets $GTAV + SYNTHIA + UrbanSyn$ (marked with $\ddagger$ in Tab. 2) are fine-tuned on 1/16 of the Cityscapes training set. As shown in Tab. 5, SoMA achieves performance on par with Rein in the source domain, while offering greater robustness on unseen domains. This underscores SoMA's data-efficient adaptability to real-world scenes, while effectively avoiding overfitting to the source domain. + +Clear-to-adverse weather DGOD. In Tab. 6, we compare SoMA with existing DGOD methods for classic backbonehead setting and with PEFT methods for recent foundation architectures. As expected, SoMA demonstrates only minor improvements in the classic setting; however, in the VFM setting—where there is considerably more knowledge to retain—it shows substantial performance gains. As illustrated by the visualization results in Fig. 4, SoMA makes robust predictions even under diverse challenging scenes. + +![](images/d68b013963ef7d21d57c454c99cf70fe31b6bd51aa0a29a38ab5128f9475fb1c.jpg) +Figure 4. DGOD qualitative results. + +
MethodsParams.*DGSS Avg.DGOD Avg.
Full fine-tuning (baseline)304.2M64.451.0
L + Freezing early blocks201.6M65.0 (↑0.6)51.4 (↑0.4)
L + Tuning principal components4.9M66.1 (↑1.1)53.0 (↑1.6)
L + Tuning minor components4.9M67.7 (↑2.7)53.8 (↑2.4)
L + Annealing weight decay4.9M68.3 (↑0.6)54.3 (↑0.5)
+ +Table 7. Effect of our changes evaluated on DG benchmarks. See the full ablation study in the Supplemental. + +
{Wq,Wk,Wv,Wo} {Wup,Wdown}Params.* →Citys. →BDD →Map. Avg.
✓2.7M68.7661.5070.00
✓2.2M70.8261.3669.86
✓✓4.9M71.8261.3171.67
+ +Table 8. DGSS performance after applying SoMA to different types of modules (Self-attention / MLP) in DINoV2-Large. + +
Rank r48163264
DGSS avg. (mIoU in %)66.9167.7168.2767.7667.59
Params.*1.3M2.5M4.9M9.6M19.0M
+ +Table 9. DGSS performance with different rank $r$ . + +# 5.3. Ablation Study + +Component Analysis. A thorough examination of each component of SoMA in the DGOD and DGSS basic setting, as summarized in Tab. 7, demonstrates that all components of SoMA incrementally improve DG performance without adding computational costs. Notably, the marked performance gap with tuning principal singular components [50] suggests that SoMA better preserves the integrity of pretrained representations while learning task-specific features. + +Tuning granularity. Interestingly, as shown in Tables 8 and 9, our method consistently achieves competitive results compared to SOTA baselines across all tuning granularities, indicating that SoMA can be flexibly configured according to the training budget, task difficulty, or domain gap. The best results are achieved with the default setting. + +# 5.4. Subject Personalization + +Domain generalized recognition requires consistent processing of inputs from diverse domains, whereas domain generalized generation involves generating outputs across a range of domains. Although large-scale Text-to-Image + +![](images/c3deea78171b5ae33ee50b141d34f258077fca93b41b0fb3e4bcb49094279cde.jpg) +Figure 5. Subject Personalization. 1. A dog gracefully leaping in origami style, 2. A dog in watercolor painting style, and 3. A dog soaring through a digital landscape in vector illustration style. + +(T2I) models have convincingly demonstrated this ability, it can be compromised in subject personalization tasks involving fine-tuning. DreamBooth (DB) [65] introduces subject personalization, which fine-tunes a pre-trained T2I model on a few reference images to enable it to generate new visual concepts. DB updates all parameters within specific blocks, potentially impairing the pre-trained capability to generate images across various domains. Therefore, we explore whether SoMA can be seamlessly integrated into the DB framework to preserve pre-trained knowledge. For this task, we utilize Stable Diffusion XL (SDXL) [59] for the T2I diffusion model. Fig. 5 presents a visual comparison between DB and SoMA, using identical hyperparameter settings and sample seeds. As seen, SoMA, similar to the original SDXL model, can generate high-quality images of diverse visual domains (e.g., style, texture, pose) and exhibits superior identity preservation compared to DB with prior preservation (p.p.) loss [65]. More results with varied prompts and subjects are provided in the Supplemental. + +# 6. Conclusion + +This paper presents a multifaceted exploration of the distribution of generalizable components in VFMs—examining them at the weight level, block level, and within training dynamics—and introduces the SoMA framework as an effective means to preserve generalizable components. Contrary to recent approaches that passively refine frozen VFM features, SoMA is integrated into all linear layers to effectively facilitate task adaptation, while minimizing interference with generalizable components by initializing the low-rank adapter with minor singular components. Through extensive experiments, we observe that SoMA significantly outperforms the state-of-the-art methods in DGSS, DGOD, and even subject-driven image generation tasks. + +# 7. Acknowledgments + +This work was supported by the National Research Foundation (NRF) grant funded by the Korea government (MSIT) [RS-2025-00562400] and [RS-2022-NR068754]. + +# References + +[1] Woo-Jin Ahn, Geun-Yeong Yang, Hyun-Duck Choi, and Myo-Taeg Lim. Style blind domain generalized semantic segmentation via covariance alignment and semantic consistence contrastive learning. In CVPR, 2024. 2, 3 +[2] Muhammad Awais, Muzammal Naseer, Salman Khan, Rao Muhammad Anwer, Hisham Cholakkal, Mubarak Shah, Ming-Hsuan Yang, and Fahad Shahbaz Khan. Foundational models defining a new era in vision: A survey and outlook. arXiv:2307.13721, 2023. 2 +[3] Yasser Benigmim, Subhankar Roy, Slim Essid, Vicky Kalogeiton, and Stephane Lathuilière. Collaborating foundation models for domain generalized semantic segmentation. In CVPR, 2024. 3, 6 +[4] Qi Bi, Jingjun Yi, Hao Zheng, Haolan Zhan, Yawen Huang, Wei Ji, Yuexiang Li, and Yefeng Zheng. Learning frequency-adapted vision foundation model for domain generalized semantic segmentation. In NeurlPS, 2024. 3, 6, 7 +[5] Qi Bi, Shaodi You, and Theo Gevers. Learning content-enhanced mask transformer for domain generalized urban-scene segmentation. In AAAI, 2024. 2, 7 +[6] Dan Biderman, Jose Gonzalez Ortiz, Jacob Portes, Mansheej Paul, Philip Greengard, Connor Jennings, Daniel King, Sam Havens, Vitaliy Chiley, Jonathan Frankle, et al. Lora learns less and forgets less. TMLR, 2024. 2 +[7] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, 2020. 1 +[8] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, 2021. 3 +[9] Gyusam Chang, Jiwon Lee, Donghyun Kim, Jinkyu Kim, Dongwook Lee, Daehyun Ji, Sujin Jang, and Sangpil Kim. Unified domain generalization and adaptation for multi-view 3d object detection. In NeurIPS, 2024. 2 +[10] Prithvijit Chattopadhyay, Kartik Sarangmath, Vivek Vijaykumar, and Judy Hoffman. Pasta: Proportional amplitude spectrum training augmentation for syn-to-real domain generalization. In ICCV, 2023. 2, 3 +[11] Shoufa Chen, Chongjian Ge, Zhan Tong, Jiangliu Wang, Yibing Song, Jue Wang, and Ping Luo. Adaptformer: Adapting vision transformers for scalable visual recognition. In NeurIPS, 2022. 2, 3, 6, 7 +[12] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In CVPR, 2022. 1, 7 +[13] Sungha Choi, Sanghun Jung, Huiwon Yun, Joanne T Kim, Seungryong Kim, and Jaegul Choo. Robustnet: Improving + +domain generalization in urban-scene segmentation via instance selective whitening. In CVPR, 2021. 2, 3, 5 +[14] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In CVPR, 2016. 6 +[15] Muhammad Sohail Danish, Muhammad Haris Khan, Muhammad Akhtar Munir, M Saquib Sarfraz, and Mohsen Ali. Improving single domain-generalized object detection: A focus on diversification and alignment. In CVPR, 2024. 3, 7 +[16] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009. 2, 4 +[17] Jian Ding, Nan Xue, Gui-Song Xia, Bernt Schiele, and Dengxin Dai. Hgformer: Hierarchical grouping transformer for domain generalized semantic segmentation. In CVPR, 2023. 3, 7 +[18] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 3 +[19] Carl Eckart and Gale Young. The approximation of one matrix by another of lower rank. Psychometrika, 1936. 4 +[20] Mohammad Fahes, Tuan-Hung Vu, Andrei Bursuc, Patrick Pérez, and Raoul de Charette. A simple recipe for language-guided domain generalized segmentation. In CVPR, 2024. 3 +[21] Qi Fan, Mattia Segu, Yu-Wing Tai, Fisher Yu, Chi-Keung Tang, Bernt Schiele, and Dengxin Dai. Towards robust object detection invariant to real-world domain shifts. In ICLR, 2023. 2, 3 +[22] Yuxin Fang, Quan Sun, Xinggang Wang, Tiejun Huang, Xinlong Wang, and Yue Cao. Eva-02: A visual representation for neon genesis. IVC, 2024. 3, 6 +[23] Aditya Sharad Golatkar, Alessandro Achille, and Stefano Soatto. Time matters in regularizing deep networks: Weight decay and data augmentation affect early learning dynamics, matter little near convergence. In NeurIPS, 2019. 6 +[24] Jose L Gómez, Manuel Silva, Antonio Seoane, Agnès Borras, Mario Noriega, German Ros, Jose A Iglesias-Guitian, and Antonio M López. All for one, and one for all: Urbansyn dataset, the third musketeer of synthetic driving scenes. arXiv:2312.12176, 2023. 1, 6 +[25] Junxian He, Chunting Zhou, Xuezhe Ma, Taylor Berg-Kirkpatrick, and Graham Neubig. Towards a unified view of parameter-efficient transfer learning. In ICLR, 2022. 2 +[26] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Delving deep into rectifiers: Surpassing human-level performance on imagenet classification. In ICCV, 2015. 3 +[27] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 2, 3, 7 + +[28] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In CVPR, 2022. 3 +[29] Lukas Hoyer, Dengxin Dai, and Luc Van Gool. Daformer: Improving network architectures and training strategies for domain-adaptive semantic segmentation. In CVPR, 2022. 5 +[30] Edward J Hu, yelong shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In ICLR, 2022. 1, 2, 3, 5, 6, 7 +[31] Jiajun Hu, Jian Zhang, Lei Qi, Yinghuan Shi, and Yang Gao. Learn to preserve and diversify: Parameter-efficient group with orthogonal regularization for domain generalization. In ECCV, 2024. 2 +[32] Wei Huang, Chang Chen, Yong Li, Jiacheng Li, Cheng Li, Fenglong Song, Youliang Yan, and Zhiwei Xiong. Style projected clustering for domain generalized semantic segmentation. In CVPR, 2023. 3 +[33] Christoph Hummer, Manuel Schwonberg, Liangwei Zhong, Hu Cao, Alois Knoll, and Hanno Gottschalk. Vltseg: Simple transfer of clip-based vision-language representations for domain generalized semantic segmentation. In ACCV, 2024. 3, 6, 7 +[34] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In ECCV, 2022. 2, 3, 6 +[35] Yuru Jia, Lukas Hoyer, Shengyu Huang, Tianfu Wang, Luc Van Gool, Konrad Schindler, and Anton Obukhov. Dgin-style: Domain-generalizable semantic segmentation with image diffusion models and stylized semantic control. In ECCV, 2024. 3 +[36] Xueying Jiang, Jiaxing Huang, Sheng Jin, and Shijian Lu. Domain generalization via balancing training difficulty and model capability. In ICCV, 2023. 3 +[37] Mengmeng Jing, Xiantong Zhen, Jingjing Li, and Cees GM Snoek. Order-preserving consistency regularization for domain adaptation and generalization. In ICCV, 2023. 3 +[38] Christoph Kamann and Carsten Rother. Increasing the robustness of semantic segmentation models with painting-by-numbers. In ECCV, 2020. 3 +[39] Jin Kim, Jiyoung Lee, Jungin Park, Dongbo Min, and Kwanghoon Sohn. Pin the memory: Learning to generalize semantic segmentation. In CVPR, 2022. 3 +[40] Sunghwan Kim, Dae-hwan Kim, and Hoseong Kim. Texture learning domain randomization for domain generalized segmentation. In ICCV, 2023. 3, 5 +[41] Suhyeon Lee, Hongje Seong, Seongwon Lee, and Euntai Kim. Wildnet: Learning domain generalized semantic segmentation from the wild. In CVPR, 2022. 3, 5 +[42] Sohyun Lee, Taeyoung Son, and Suha Kwak. Fifo: Learning fog-invariant features for foggy scene segmentation. In CVPR, 2022. 3 +[43] Wooju Lee, Dasol Hong, Hyungtae Lim, and Hyun Myung. Object-aware domain generalization for object detection. In AAAI, 2024. 3, 7 +[44] Yoonho Lee, Annie S Chen, Fahim Tajwar, Ananya Kumar, Huaxiu Yao, Percy Liang, and Chelsea Finn. Surgical fine + +tuning improves adaptation to distribution shifts. In ICLR, 2023.5 +[45] Brian Lester, Rami Al-Rfou, and Noah Constant. The power of scale for parameter-efficient prompt tuning. In EMNLP, 2021. 2 +[46] Deng Li, Among Wu, Yaowei Wang, and Yahong Han. Prompt-driven dynamic object-centric learning for single domain generalization. In CVPR, 2024. 3, 7 +[47] Dongze Lian, Daquan Zhou, Jiashi Feng, and Xinchao Wang. Scaling & shifting your features: A new baseline for efficient model tuning. In NeurIPS, 2022. 3, 6 +[48] Shih-Yang Liu, Chien-Yi Wang, Hongxu Yin, Pavlo Molchanov, Yu-Chiang Frank Wang, Kwang-Ting Cheng, and Min-Hung Chen. DoRA: Weight-decomposed low-rank adaptation. In ICML, 2024. 3, 6, 7 +[49] Yajing Liu, Shijun Zhou, Xiyao Liu, Chunhui Hao, Baojie Fan, and Jiandong Tian. Unbiased faster r-cnn for single-source domain generalized object detection. In CVPR, 2024. 3, 7 +[50] Fanxu Meng, Zhaohui Wang, and Muhan Zhang. Pissa: Principal singular values and singular vectors adaptation of large language models. In NeurIPS, 2024. 3, 8 +[51] Gerhard Neuhold, Tobias Ollmann, Samuel Rota Bulo, and Peter Kontschieder. The mapillary vistas dataset for semantic understanding of street scenes. In ICCV, 2017. 6 +[52] Joshua Niemeijer, Manuel Schwonberg, Jan-Aike Termöhlen, Nico M Schmidt, and Tim Fingscheidt. Generalization by adaptation: Diffusion-based domain extension for domain-generalized semantic segmentation. In WACV, 2024. 3 +[53] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. TMLR, 2023. 3, 4, 5, 6, 7 +[54] Byeonghyun Pak, Byeongju Woo, Sunghwan Kim, Dae-hwan Kim, and Hoseong Kim. Textual query-driven mask transformer for domain generalized segmentation. In ECCV, 2024. 3, 6, 7 +[55] Xingang Pan, Ping Luo, Jianping Shi, and Xiaou Tang. Two at once: Enhancing learning and generalization capacities via ibn-net. In ECCV, 2018. 3 +[56] Yuwen Pan, Rui Sun, Naisong Luo, Tianzhu Zhang, and Yongdong Zhang. Exploring reliable matching with phase enhancement for night-time semantic segmentation. In ECCV, 2024.3 +[57] Duo Peng, Yinjie Lei, Lingqiao Liu, Pingping Zhang, and Jun Liu. Global and local texture randomization for synthetic-to-real semantic segmentation. TIP, 2021. 2, 3 +[58] Duo Peng, Yinjie Lei, Munawar Hayat, Yulan Guo, and Wen Li. Semantic-aware domain generalized segmentation. In CVPR, 2022. 2, 3 +[59] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. SDXL: Improving latent diffusion models for high-resolution image synthesis. In ICLR, 2024. 3, 8 + +[60] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 3 +[61] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. In NeurIPS, 2015. 1, 7 +[62] Stephan R Richter, Vibhav Vineet, Stefan Roth, and Vladlen Koltun. Playing for data: Ground truth from computer games. In ECCV, 2016. 1, 6 +[63] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 3 +[64] German Ros, Laura Sellart, Joanna Materzynska, David Vazquez, and Antonio M. Lopez. The synthia dataset: A large collection of synthetic images for semantic segmentation of urban scenes. In CVPR, 2016. 1, 6 +[65] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In CVPR, 2023. 8 +[66] Christos Sakaridis, Dengxin Dai, and Luc Van Gool. Acdc: The adverse conditions dataset with correspondences for semantic driving scene understanding. In ICCV, 2021. 6, 7 +[67] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zh-moginov, and Liang-Chieh Chen. Mobilenetv2: Inverted residuals and linear bottlenecks. In CVPR, 2018. 2 +[68] Fengyi Shen, Akhil Gurram, Ziyuan Liu, He Wang, and Alois Knoll. Diga: Distil to generalize and then adapt for domain adaptive semantic segmentation. In CVPR, 2023. 3 +[69] Zhiqiang Tang, Yunhe Gao, Yi Zhu, Zhi Zhang, Mu Li, and Dimitris N Metaxas. Crossnorm and selfnorm for generalization under distribution shifts. In ICCV, 2021. 3 +[70] Sumanth Udupa, Prajwal Gurunath, Aniruddh Sikdar, and Suresh Sundaram. Mrfp: Learning generalizable semantic segmentation from sim-2-real with multi-resolution feature perturbation. In CVPR, 2024. 2, 3 +[71] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 3 +[72] Vidit Vidit, Martin Engilberge, and Mathieu Salzmann. Clip the gap: A single domain generalization approach for object detection. In CVPR, 2023. 3, 7 +[73] Zhixiang Wei, Lin Chen, Tao Tu, Pengyang Ling, Huaian Chen, and Yi Jin. Disentangle then parse: Night-time semantic segmentation with illumination disentanglement. In ICCV, 2023. 3 +[74] Zhixiang Wei, Lin Chen, Yi Jin, Xiaoxiao Ma, Tianle Liu, Pengyang Ling, Ben Wang, Huaian Chen, and Jinjin Zheng. Stronger fewer & superior: Harnessing vision foundation models for domain generalized semantic segmentation. In CVPR, 2024. 2, 3, 6, 7 +[75] Among Wu and Cheng Deng. Single-domain generalized object detection in urban scene via cyclic-disentangled self-distillation. In CVPR, 2022. 3, 6, 7 + +[76] Zhenyao Wu, Xinyi Wu, Xiaoping Zhang, Lili Ju, and Song Wang. Siamdoge: Domain generalizable semantic segmentation using siamese network. In ECCV, 2022. 3 +[77] Qi Xu, Liang Yao, Zhengkai Jiang, Guannan Jiang, Wenqing Chu, Wenhui Han, Wei Zhang, Chengjie Wang, and Ying Tai. Dirl: Domain-invariant representation learning for generalizable semantic segmentation. In AAAI, 2022. 3 +[78] Liwei Yang, Xiang Gu, and Jian Sun. Generalized semantic segmentation by self-supervised source domain projection and multi-level contrastive learning. In AAAI, 2023. 3 +[79] Jingjun Yi, Qi Bi, Hao Zheng, Haolan Zhan, Wei Ji, Yawen Huang, Yuexiang Li, and Yefeng Zheng. Learning spectral-decomposed tokens for domain generalized semantic segmentation. In ACMMM, 2024. 3, 6 +[80] Fisher Yu, Haofeng Chen, Xin Wang, Wenqi Xian, Yingying Chen, Fangchen Liu, Vashisht Madhavan, and Trevor Darrell. Bdd100k: A diverse driving dataset for heterogeneous multitask learning. In CVPR, 2020. 6 +[81] Yeonguk Yu, Sungho Shin, Seunghyeok Back, Mihwan Ko, Sangjun Noh, and Kyoobin Lee. Domain-specific block selection and paired-view pseudo-labeling for online test-time adaptation. In CVPR, 2024. 5 +[82] Xiangyu Yue, Yang Zhang, Sicheng Zhao, Alberto Sangiovanni-Vincentelli, Kurt Keutzer, and Boqing Gong. Domain randomization and pyramid consistency: Simulation-to-real generalization without accessing target domain data. In ICCV, 2019. 3 +[83] Yuyang Zhao, Zhun Zhong, Na Zhao, Nicu Sebe, and Gim Hee Lee. Style-hallucinated dual consistency learning for domain generalized semantic segmentation. In ECCV, 2022. 3 +[84] Zhun Zhong, Yuyang Zhao, Gim Hee Lee, and Nicu Sebe. Adversarial style augmentation for domain generalized urban-scene segmentation. In NeurIPS, 2022. 3, 5 +[85] Jinghao Zhou, Chen Wei, Huiyu Wang, Wei Shen, Cihang Xie, Alan Yuille, and Tao Kong. ibot: Image bert pre-training with online tokenizer. arXiv:2111.07832, 2021. 3 +[86] Zhuofan Zong, Guanglu Song, and Yu Liu. Detrs with collaborative hybrid assignments training. In ICCV, 2023. 7 \ No newline at end of file diff --git a/2025/SoMA_ Singular Value Decomposed Minor Components Adaptation for Domain Generalizable Representation Learning/images.zip b/2025/SoMA_ Singular Value Decomposed Minor Components Adaptation for Domain Generalizable Representation Learning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..902b523bc63a0f6e8e0bcc1801a8b91c26b7c604 --- /dev/null +++ b/2025/SoMA_ Singular Value Decomposed Minor Components Adaptation for Domain Generalizable Representation Learning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6643fbe995302b0f5fffc14a221716fe17bd766751b120b0a3d48e36e8323ffd +size 723036 diff --git a/2025/SoMA_ Singular Value Decomposed Minor Components Adaptation for Domain Generalizable Representation Learning/layout.json b/2025/SoMA_ Singular Value Decomposed Minor Components Adaptation for Domain Generalizable Representation Learning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..d2b32d13539f37eef95522dafa563d9443dfaa49 --- /dev/null +++ b/2025/SoMA_ Singular Value Decomposed Minor Components Adaptation for Domain Generalizable Representation Learning/layout.json @@ -0,0 +1,9628 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 95, + 66, + 515, + 102 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 95, + 66, + 515, + 102 + ], + "spans": [ + { + "bbox": [ + 95, + 66, + 515, + 102 + ], + "type": "text", + "content": "SoMA: Singular Value Decomposed Minor Components Adaptation for Domain Generalizable Representation Learning" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 124, + 478, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 124, + 478, + 152 + ], + "spans": [ + { + "bbox": [ + 131, + 124, + 478, + 152 + ], + "type": "text", + "content": "Seokju Yun Seunghye Chae Dongheon Lee Youngmin Ro* Machine Intelligence Laboratory, University of Seoul, Korea" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 167, + 155, + 443, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 167, + 155, + 443, + 179 + ], + "spans": [ + { + "bbox": [ + 167, + 155, + 443, + 179 + ], + "type": "text", + "content": "{wsz871, tmdhey, dslisleedh, youngmin.ro}@uos.ac.kr \nhttps://github.com/ysj9909/SoMA" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 58, + 187, + 231, + 370 + ], + "blocks": [ + { + "bbox": [ + 58, + 187, + 231, + 370 + ], + "lines": [ + { + "bbox": [ + 58, + 187, + 231, + 370 + ], + "spans": [ + { + "bbox": [ + 58, + 187, + 231, + 370 + ], + "type": "image", + "image_path": "3628e235ba0f324223b527313dafe6b96b4369108834212fbc94f57ac5ccc19f.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 373, + 555, + 441 + ], + "lines": [ + { + "bbox": [ + 54, + 373, + 555, + 441 + ], + "spans": [ + { + "bbox": [ + 54, + 373, + 555, + 441 + ], + "type": "text", + "content": "Figure 1. Overview of SoMA framework. Left. SoMA achieves state-of-the-art results across diverse tasks, ranging from domaingeneralized semantic segmentation (DGSS) to object detection (DGOD), and performs well in both synthetic-to-real and real-to-real scenarios. Middle. Our method, trained solely on synthetic datasets, demonstrates strong generalization capabilities in complex real-world scenes. Right. We present SoMA, a method that applies SVD to the pre-trained weights, decomposing them into " + }, + { + "bbox": [ + 54, + 373, + 555, + 441 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 54, + 373, + 555, + 441 + ], + "type": "text", + "content": " minor singular components " + }, + { + "bbox": [ + 54, + 373, + 555, + 441 + ], + "type": "inline_equation", + "content": "U_{[:,-r:]} \\Sigma_{[-r:]} (V^T)_{[-r:,]}" + }, + { + "bbox": [ + 54, + 373, + 555, + 441 + ], + "type": "text", + "content": " and residuals. SoMA selectively tunes the smallest " + }, + { + "bbox": [ + 54, + 373, + 555, + 441 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 54, + 373, + 555, + 441 + ], + "type": "text", + "content": " components, effectively preserving world knowledge of foundation models. Since SoMA shares the same forward architecture as LoRA [30], it adds no extra latency during inference phase." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 235, + 189, + 350, + 247 + ], + "blocks": [ + { + "bbox": [ + 235, + 189, + 350, + 247 + ], + "lines": [ + { + "bbox": [ + 235, + 189, + 350, + 247 + ], + "spans": [ + { + "bbox": [ + 235, + 189, + 350, + 247 + ], + "type": "image", + "image_path": "49983582969fbd7f5d57332555baa5e2f0fd28824735a4ba3bba883fe850aabb.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 235, + 250, + 350, + 308 + ], + "blocks": [ + { + "bbox": [ + 235, + 250, + 350, + 308 + ], + "lines": [ + { + "bbox": [ + 235, + 250, + 350, + 308 + ], + "spans": [ + { + "bbox": [ + 235, + 250, + 350, + 308 + ], + "type": "image", + "image_path": "c40ea8d126a3176fbedf8dc89cc74e2b9af90901a70fd337c29b9cec168109db.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 235, + 312, + 350, + 369 + ], + "blocks": [ + { + "bbox": [ + 235, + 312, + 350, + 369 + ], + "lines": [ + { + "bbox": [ + 235, + 312, + 350, + 369 + ], + "spans": [ + { + "bbox": [ + 235, + 312, + 350, + 369 + ], + "type": "image", + "image_path": "5d4165514b4bb3db4a6d890ec65f07d65fa688179d0a15423ea0523abd4ce689.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 358, + 188, + 553, + 369 + ], + "blocks": [ + { + "bbox": [ + 358, + 188, + 553, + 369 + ], + "lines": [ + { + "bbox": [ + 358, + 188, + 553, + 369 + ], + "spans": [ + { + "bbox": [ + 358, + 188, + 553, + 369 + ], + "type": "image", + "image_path": "794488922f066aff2e1e6557ceeb047b3362fbbdc3075f5cbf3e8df5bbe3a009.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 151, + 450, + 200, + 463 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 450, + 200, + 463 + ], + "spans": [ + { + "bbox": [ + 151, + 450, + 200, + 463 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 54, + 476, + 297, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 476, + 297, + 693 + ], + "spans": [ + { + "bbox": [ + 54, + 476, + 297, + 693 + ], + "type": "text", + "content": "Domain generalization (DG) aims to adapt a model using one or multiple source domains to ensure robust performance in unseen target domains. Recently, Parameter-Efficient Fine-Tuning (PEFT) of foundation models has shown promising results in the context of DG problem. Nevertheless, existing PEFT methods still struggle to strike a balance between preserving generalizable components of the pre-trained model and learning task-specific features. To gain insights into the distribution of generalizable components, we begin by analyzing the pre-trained weights through the lens of singular value decomposition. Building on these insights, we introduce Singular Value Decomposed Minor Components Adaptation (SoMA), an approach that selectively tunes minor singular components while keeping the residual parts frozen. SoMA effectively retains the generalization ability of the pre-trained model while efficiently acquiring task-specific skills. Moreover, we freeze domain-generalizable blocks and employ an annealing weight decay" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 452, + 555, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 452, + 555, + 559 + ], + "spans": [ + { + "bbox": [ + 313, + 452, + 555, + 559 + ], + "type": "text", + "content": "strategy, thereby achieving an optimal balance in the delicate trade-off between generalizability and discriminability. SoMA attains state-of-the-art results on multiple benchmarks that span both domain generalized semantic segmentation to domain generalized object detection. In addition, our methods introduce no additional inference overhead or regularization loss, maintain compatibility with any backbone or head, and are designed to be versatile, allowing easy integration into a wide range of tasks." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 314, + 564, + 395, + 576 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 564, + 395, + 576 + ], + "spans": [ + { + "bbox": [ + 314, + 564, + 395, + 576 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "text", + "content": "Deep neural networks (DNNs) have recently made significant strides in dense prediction tasks, including semantic segmentation [12] and object detection [7, 61], both of which are crucial for safety-critical applications, such as autonomous driving. However, DNNs often fail to maintain reliable performance under domain shifts, which may result from diverse lighting conditions, weather changes, or differences in location. Likewise, while synthetic data [24, 62, 64] is employed to avoid the labor-intensive and costly process of constructing real-world datasets, the resulting performance degradation of DNNs in real-world deployment" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 15 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 15 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 15 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 198, + 13, + 443, + 24 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 13, + 443, + 24 + ], + "spans": [ + { + "bbox": [ + 198, + 13, + 443, + 24 + ], + "type": "text", + "content": "Except for this watermark, it is identical to the accepted version;" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 179, + 25, + 462, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 25, + 462, + 37 + ], + "spans": [ + { + "bbox": [ + 179, + 25, + 462, + 37 + ], + "type": "text", + "content": "the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 703, + 145, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 145, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 145, + 713 + ], + "type": "text", + "content": "*Corresponding author." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "25602" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "content": "remains a critical issue to be resolved. To tackle these challenges, Domain Generalization (DG) is introduced to design models capable of consistent prediction on arbitrary unseen target domains." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 121, + 295, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 121, + 295, + 263 + ], + "spans": [ + { + "bbox": [ + 54, + 121, + 295, + 263 + ], + "type": "text", + "content": "Existing DG methods for dense prediction tasks employ two main strategies to enhance model robustness. The first involves diversifying the training process in the input [10, 57] or feature space [21, 70] using augmentation/perturbation techniques, exposing the model to a broader range of styles to mitigate overfitting to specific domains. The second strategy focuses on imposing alignment constraints, such as normalization [58] or regularization losses [1, 13], to facilitate the learning of domain-invariant features. However, these methods mostly utilize outdated backbones (e.g., ResNet [27] or MobileNetV2 [67]) pretrained on mid-scale datasets, such as ImageNet [16]." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 265, + 295, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 265, + 295, + 396 + ], + "spans": [ + { + "bbox": [ + 55, + 265, + 295, + 396 + ], + "type": "text", + "content": "Meanwhile, vision foundation models (VFMs) [2], pretrained on vast, curated datasets with enormous parameter counts, have recently shown unprecedented generalization capabilities [74]. Given this landscape, we propose a paradigm shift—from applying the diversify & align approach to classic backbones, to devising methods that preserve the world knowledge of VFMs while effectively learning task-specific features. However, naive adaptation of VFMs via full fine-tuning (FFT), which involves retraining all model parameters, results in prohibitive costs and a risk of catastrophic forgetting of the pre-trained knowledge." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 397, + 295, + 587 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 397, + 295, + 587 + ], + "spans": [ + { + "bbox": [ + 55, + 397, + 295, + 587 + ], + "type": "text", + "content": "In light of these challenges, a commonly adopted approach is parameter-efficient fine-tuning (PEFT), where only the injected lightweight adapters [11, 25, 30] or tokens [34, 45] are fine-tuned, while the pre-trained weights remain frozen. Low-Rank Adaptation (LoRA) [30] is one of the most widely used PEFT methods, which substitutes the updates with the product of two low-rank matrices. Interestingly, we found that LoRA exhibits superior generalization ability in DG dense prediction tasks compared to FFT and other PEFT methods (see Tab. 2), aligning with recent findings in LoRA-based literature [6, 31]. Nevertheless, LoRA adapters are built without considering the distribution of generalizable components within the pre-trained weights they modify. As a result, there is still room for enhancing generalizability by preserving the diverse domain knowledge of VFMs." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 588, + 295, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 588, + 295, + 684 + ], + "spans": [ + { + "bbox": [ + 55, + 588, + 295, + 684 + ], + "type": "text", + "content": "To uncover the structure in which world knowledge is encoded within the pre-trained weights, we first perform singular value decomposition (SVD) on the learned weight matrices and then analyze the behavior of VFMs after selectively removing specific singular components1. Our analysis reveals that singular vectors associated with higher singular values tend to extract general features spanning multiple classes in ImageNet, while minor singular com" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 553, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 204 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 204 + ], + "type": "text", + "content": "ponents are primarily responsible for capturing context-specific features. Capitalizing on our findings, we introduce Singular Value Decomposed Minor Components Adaptation (SoMA), which starts by performing SVD to decompose the weights into their singular vectors, i.e., " + }, + { + "bbox": [ + 313, + 72, + 553, + 204 + ], + "type": "inline_equation", + "content": "W = U\\Sigma V^T" + }, + { + "bbox": [ + 313, + 72, + 553, + 204 + ], + "type": "text", + "content": ", then fine-tunes only the minor singular components. Specifically, we initialize LoRA adapter using the components with the smallest " + }, + { + "bbox": [ + 313, + 72, + 553, + 204 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 313, + 72, + 553, + 204 + ], + "type": "text", + "content": " singular values, i.e., " + }, + { + "bbox": [ + 313, + 72, + 553, + 204 + ], + "type": "inline_equation", + "content": "U_{[:,-r:]} \\Sigma_{[-r:]}(V^T)_{[-r:,]}" + }, + { + "bbox": [ + 313, + 72, + 553, + 204 + ], + "type": "text", + "content": ", and the remaining residual components are frozen to maintain generalizability, as illustrated in Fig. 1 right." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 204, + 553, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 204, + 553, + 417 + ], + "spans": [ + { + "bbox": [ + 313, + 204, + 553, + 417 + ], + "type": "text", + "content": "By extending our analysis from the weight-level to the block-level, we observe and empirically demonstrate that the early blocks of VFMs effectively extract well-localized semantic features and are less affected by style or input-level domain shifts (see Fig. 3). Therefore, we freeze these blocks to preserve generalizable components while reducing the number of trainable parameters. Although our proposed methods are effective, tuning the less-optimized bottom spectral space and freezing early blocks may lead to a lack of discriminability. To mitigate this issue, we introduce an annealing weight decay scheme that gradually reduces the regularization loss incurred by weight decay over the course of training. As shown in Fig. 1, comprehensive evaluations on DG benchmarks demonstrate that our framework outperforms previous state-of-the-art baselines and FFT, training only " + }, + { + "bbox": [ + 313, + 204, + 553, + 417 + ], + "type": "inline_equation", + "content": "0.58\\%" + }, + { + "bbox": [ + 313, + 204, + 553, + 417 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 204, + 553, + 417 + ], + "type": "inline_equation", + "content": "1.6\\%" + }, + { + "bbox": [ + 313, + 204, + 553, + 417 + ], + "type": "text", + "content": " of the parameters. Furthermore, SoMA demonstrates impressive scaling performance with both data and model size." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 419, + 468, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 419, + 468, + 430 + ], + "spans": [ + { + "bbox": [ + 314, + 419, + 468, + 430 + ], + "type": "text", + "content": "We make the following contributions:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 315, + 434, + 553, + 664 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 315, + 434, + 553, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 434, + 553, + 469 + ], + "spans": [ + { + "bbox": [ + 315, + 434, + 553, + 469 + ], + "type": "text", + "content": "- We underscore the critical importance of preserving pretrained knowledge while learning task-specific features in scenarios that involve the use of VFMs." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 315, + 471, + 553, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 471, + 553, + 555 + ], + "spans": [ + { + "bbox": [ + 315, + 471, + 553, + 555 + ], + "type": "text", + "content": "- We present SoMA, a novel PEFT method specifically tailored for DG problem, which leverages spectral information to retain the generalization capacity of VFMs. In addition, we identify and freeze domain-generalizable blocks while employing an annealing weight decay strategy, thereby achieving an optimal balance between generalizability and discriminability." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 315, + 556, + 553, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 556, + 553, + 664 + ], + "spans": [ + { + "bbox": [ + 315, + 556, + 553, + 664 + ], + "type": "text", + "content": "- Extensive experiments on various DG for semantic segmentation and object detection tasks reveal that our framework surpasses existing baselines by a substantial margin. Notably, SoMA achieves an mIoU of " + }, + { + "bbox": [ + 315, + 556, + 553, + 664 + ], + "type": "inline_equation", + "content": "80.4\\%" + }, + { + "bbox": [ + 315, + 556, + 553, + 664 + ], + "type": "text", + "content": " on Cityscapes without accessing real-world datasets, while training only " + }, + { + "bbox": [ + 315, + 556, + 553, + 664 + ], + "type": "inline_equation", + "content": "0.58\\%" + }, + { + "bbox": [ + 315, + 556, + 553, + 664 + ], + "type": "text", + "content": " of the parameters. Additionally, under the most challenging daytime-sunny " + }, + { + "bbox": [ + 315, + 556, + 553, + 664 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 315, + 556, + 553, + 664 + ], + "type": "text", + "content": " night-rainy detection setting, SoMA outperforms the prior SOTA implementation by achieving a " + }, + { + "bbox": [ + 315, + 556, + 553, + 664 + ], + "type": "inline_equation", + "content": "+23.5\\%" + }, + { + "bbox": [ + 315, + 556, + 553, + 664 + ], + "type": "text", + "content": " mAP gain." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 670, + 400, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 670, + 400, + 682 + ], + "spans": [ + { + "bbox": [ + 314, + 670, + 400, + 682 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "content": "Domain generalization for dense predictions. Domain Generalization (DG) methods for dense predictions [5, 9," + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "type": "text", + "content": "In this paper, we refer to the combination of a singular value and its corresponding singular vectors as a singular component." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25603" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "content": "17, 32, 36, 39, 42, 46, 56, 68, 73, 75, 78] have recently garnered considerable attention due to their practical demands. These methods can be categorized along two main axes: i) data augmentation or domain randomization, both of which diversify the training process, and ii) alignment techniques to suppress domain-relevant features." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 147, + 296, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 147, + 296, + 397 + ], + "spans": [ + { + "bbox": [ + 56, + 147, + 296, + 397 + ], + "type": "text", + "content": "Methods adopting the first approach at the input level employ various augmentation techniques [10, 38, 57, 84] or, more recently, leverage advances in generative modeling, such as diffusion models [59, 63], to generate data approximating the target domain [3, 35, 52], thereby expanding domain diversity and enriching learned representations. To enable more flexible augmentation, several studies [20, 21, 41, 70, 72] leverage feature stylization to enhance model robustness. On the other hand, alignment-based methods effectively reduce domain-sensitive components by employing feature normalization/whitening techniques [55, 58, 69], or by introducing regularization losses to suppress inconsistencies caused by simulated domain shifts [1, 13, 15, 37, 40, 43, 49, 76, 77, 82, 83]. However, most of the aforementioned approaches experimented with CNN backbones [27] trained on relatively restricted domains. In contrast, when applying vision foundation models (VFMs)—trained on massive datasets using transformer architectures [18, 71] and sophisticated training schemes [53]—to the DG problem, a fundamental shift in approach is imperative." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 400, + 295, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 400, + 295, + 579 + ], + "spans": [ + { + "bbox": [ + 55, + 400, + 295, + 579 + ], + "type": "text", + "content": "Following this trend, recent studies leverage vision-language models as feature extractors [33], employing text embeddings with domain-invariant semantics as a source for style augmentation [20, 72] or as object queries in transformer-based decoders [54]. Rein [74] injects learnable adapters and tokens to refine feature maps for each instance while keeping the backbone frozen, significantly expanding the performance frontier. FADA [4] and SET [79] introduce frequency-adapted methods to effectively utilize frozen VFM features. In contrast to existing methods that simply freeze all parameters, we perform singular value decomposition on the pre-trained weights, freezing the generalizable principal components and tuning only the minor singular components responsible for context-specific features, thereby achieving superior DG performance." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "text", + "content": "Vision foundation models (VFMs). Recently, the advent of VFMs with general perception capabilities has laid the foundation for a paradigm in which these models are broadly applicable to numerous downstream tasks. Seminal examples of models advancing this direction include CLIP [60], which demonstrates strong zero-shot generalizability through training on web-scale, weakly supervised image-text pairs; EVA02 [22], which enhances CLIP features with masked image modeling [28]; and DINOv2 [53], which incorporates losses from prior arts like iBoT [85] and DINO [8], and is trained on massive curated datasets, offer" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "text", + "content": "ing strong spatial features for dense prediction tasks. Our SoMA builds on this foundation, harnessing the full potential of VFMs in the context of DG problem." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 109, + 555, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 109, + 555, + 348 + ], + "spans": [ + { + "bbox": [ + 313, + 109, + 555, + 348 + ], + "type": "text", + "content": "Parameter-efficient fine-tuning (PEFT). With the remarkable success of PEFT methods in efficiently adapting large-scale foundation models in the realm of natural language processing, there has been a growing interest in extending these approaches to the field of computer vision. For example, VPT [34] prepends prompt tokens to the input sequence of several attention layers. AdaptFormer [11] introduces a lightweight adapter in parallel to the feed-forward network, whereas SSF [47] incorporates scale and shift parameters to modulate features for downstream tasks. LoRA [30] models the incremental updates of pre-trained weights using rank decomposition matrices, achieving performance comparable to full fine-tuning (FFT). However, there still often exists a performance gap between LoRA and FFT. DoRA [48] and PiSSA [50] address this gap by further decomposing the pre-trained weights, initializing the LoRA adapter with directional and principal components, respectively, and fine-tuning accordingly. While these methods aim to match the performance of FFT, our focus is on the preservation of generalizable components." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 359, + 399, + 372 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 359, + 399, + 372 + ], + "spans": [ + { + "bbox": [ + 314, + 359, + 399, + 372 + ], + "type": "text", + "content": "3. Preliminaries" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 378, + 554, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 378, + 554, + 473 + ], + "spans": [ + { + "bbox": [ + 313, + 378, + 554, + 473 + ], + "type": "text", + "content": "Low-Rank Adaptation (LoRA). LoRA [30] hypothesizes that the weight changes during fine-tuning exhibit a low-rank structure. Given a pre-trained weight matrix " + }, + { + "bbox": [ + 313, + 378, + 554, + 473 + ], + "type": "inline_equation", + "content": "W_0 \\in \\mathbb{R}^{m \\times n}" + }, + { + "bbox": [ + 313, + 378, + 554, + 473 + ], + "type": "text", + "content": ", LoRA constrains its update " + }, + { + "bbox": [ + 313, + 378, + 554, + 473 + ], + "type": "inline_equation", + "content": "\\Delta W \\in \\mathbb{R}^{m \\times n}" + }, + { + "bbox": [ + 313, + 378, + 554, + 473 + ], + "type": "text", + "content": " to a low-rank decomposition " + }, + { + "bbox": [ + 313, + 378, + 554, + 473 + ], + "type": "inline_equation", + "content": "\\Delta W = BA" + }, + { + "bbox": [ + 313, + 378, + 554, + 473 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 378, + 554, + 473 + ], + "type": "inline_equation", + "content": "B \\in \\mathbb{R}^{m \\times r}" + }, + { + "bbox": [ + 313, + 378, + 554, + 473 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 378, + 554, + 473 + ], + "type": "inline_equation", + "content": "A \\in \\mathbb{R}^{r \\times n}" + }, + { + "bbox": [ + 313, + 378, + 554, + 473 + ], + "type": "text", + "content": " are two low-rank matrices, with an intrinsic rank of " + }, + { + "bbox": [ + 313, + 378, + 554, + 473 + ], + "type": "inline_equation", + "content": "r \\ll \\min(m, n)" + }, + { + "bbox": [ + 313, + 378, + 554, + 473 + ], + "type": "text", + "content": ". For " + }, + { + "bbox": [ + 313, + 378, + 554, + 473 + ], + "type": "inline_equation", + "content": "y = W_0 x" + }, + { + "bbox": [ + 313, + 378, + 554, + 473 + ], + "type": "text", + "content": ", the modified forward pass is as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 358, + 479, + 553, + 493 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 479, + 553, + 493 + ], + "spans": [ + { + "bbox": [ + 358, + 479, + 553, + 493 + ], + "type": "interline_equation", + "content": "y = \\left(W _ {0} + \\Delta W\\right) x = \\left(W _ {0} + \\underline {{B A}}\\right) x, \\tag {1}", + "image_path": "e9294b0f22671e57e8aa3c9d9fe0293371461fa4bb51d7a15c05b53f641bfb22.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 498, + 554, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 554, + 593 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 554, + 593 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 498, + 554, + 593 + ], + "type": "inline_equation", + "content": "W_{0}" + }, + { + "bbox": [ + 313, + 498, + 554, + 593 + ], + "type": "text", + "content": " is frozen, and only the underlined low-rank parameters are fine-tuned, significantly reducing the number of trainable parameters. LoRA uses a uniform Kaiming distribution [26] to initialize " + }, + { + "bbox": [ + 313, + 498, + 554, + 593 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 313, + 498, + 554, + 593 + ], + "type": "text", + "content": ", while " + }, + { + "bbox": [ + 313, + 498, + 554, + 593 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 313, + 498, + 554, + 593 + ], + "type": "text", + "content": " is initially set to zero, ensuring " + }, + { + "bbox": [ + 313, + 498, + 554, + 593 + ], + "type": "inline_equation", + "content": "BA = 0" + }, + { + "bbox": [ + 313, + 498, + 554, + 593 + ], + "type": "text", + "content": " at the beginning of training. As seen in Eq. 1, the learned matrices " + }, + { + "bbox": [ + 313, + 498, + 554, + 593 + ], + "type": "inline_equation", + "content": "BA" + }, + { + "bbox": [ + 313, + 498, + 554, + 593 + ], + "type": "text", + "content": " can be merged with the frozen weight " + }, + { + "bbox": [ + 313, + 498, + 554, + 593 + ], + "type": "inline_equation", + "content": "W_{0}" + }, + { + "bbox": [ + 313, + 498, + 554, + 593 + ], + "type": "text", + "content": ", allowing LoRA-based variants to avoid adding any additional inference burden." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "content": "Singular Value Decomposition (SVD) Analysis. We observe that LoRA and existing PEFT methods, by training only a relatively small subset of parameters or by enforcing low-rank-ness of weight updates, deliver promising DG performance without significantly distorting the pre-trained representations. However, many previous studies build adapters without considering how the generalizable components are structured within the weight matrices, potentially interfering with these components in the fine-tuning process. Building on the Eckart-Young-Mirsky the" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "25604" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 80, + 62, + 266, + 173 + ], + "blocks": [ + { + "bbox": [ + 80, + 62, + 266, + 173 + ], + "lines": [ + { + "bbox": [ + 80, + 62, + 266, + 173 + ], + "spans": [ + { + "bbox": [ + 80, + 62, + 266, + 173 + ], + "type": "image", + "image_path": "e96b2d5afd82c26f5ee4cb84a1541c58f76df86f4626494ad4971d542f8dd724.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 78, + 175, + 269, + 251 + ], + "blocks": [ + { + "bbox": [ + 78, + 175, + 269, + 251 + ], + "lines": [ + { + "bbox": [ + 78, + 175, + 269, + 251 + ], + "spans": [ + { + "bbox": [ + 78, + 175, + 269, + 251 + ], + "type": "image", + "image_path": "59d361949c111fda71f05b4d52a8a3a70a849e82e159392981be01f3e1a906d0.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 253, + 296, + 342 + ], + "lines": [ + { + "bbox": [ + 55, + 253, + 296, + 342 + ], + "spans": [ + { + "bbox": [ + 55, + 253, + 296, + 342 + ], + "type": "text", + "content": "Figure 2. Distribution of generalizable components. Top. Number of classes exhibiting specific accuracy drops after applying SVD to DINOv2-large weights and reconstructing by truncating the smallest " + }, + { + "bbox": [ + 55, + 253, + 296, + 342 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 55, + 253, + 296, + 342 + ], + "type": "text", + "content": " singular components. Bottom. Distinct roles of singular components across levels. Numbers in parentheses represent the count of classes with an accuracy drop ratio exceeding " + }, + { + "bbox": [ + 55, + 253, + 296, + 342 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 55, + 253, + 296, + 342 + ], + "type": "text", + "content": ". The average WordNet hierarchy depth of these classes is shown, with higher values indicating greater context specificity." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 350, + 296, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 350, + 296, + 458 + ], + "spans": [ + { + "bbox": [ + 55, + 350, + 296, + 458 + ], + "type": "text", + "content": "orem [19], which validates that the optimal rank-" + }, + { + "bbox": [ + 55, + 350, + 296, + 458 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 55, + 350, + 296, + 458 + ], + "type": "text", + "content": " approximation of a matrix " + }, + { + "bbox": [ + 55, + 350, + 296, + 458 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 55, + 350, + 296, + 458 + ], + "type": "text", + "content": " is represented by the sum of its top " + }, + { + "bbox": [ + 55, + 350, + 296, + 458 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 55, + 350, + 296, + 458 + ], + "type": "text", + "content": " singular components, we analyze how the world knowledge of VFMs is structured from an SVD perspective. Using ImageNet-1k [16] as a testbed, our analysis computes the SVD of pre-trained weights across all layers of the DINOv2 [53] ViT-large model, followed by truncating specific singular values and their corresponding singular vectors to investigate the classes that are subsequently misclassified." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 460, + 296, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 460, + 296, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 460, + 296, + 628 + ], + "type": "text", + "content": "As shown in Fig. 2 top, as the number of discarded smallest singular components increases, the number of misclassified classes grows exponentially rather than linearly, suggesting that components with larger singular values tend to capture more general features spanning multiple classes. To further examine whether different groups of singular components exhibit distinct behaviors, we consider three groups: the top 8 principal components " + }, + { + "bbox": [ + 55, + 460, + 296, + 628 + ], + "type": "inline_equation", + "content": "U_{[:;8]}\\Sigma_{[:8]}(V^T)_{[:8,:]}" + }, + { + "bbox": [ + 55, + 460, + 296, + 628 + ], + "type": "text", + "content": ", the middle 160 singular components " + }, + { + "bbox": [ + 55, + 460, + 296, + 628 + ], + "type": "inline_equation", + "content": "U_{[:384:544]}\\Sigma_{[384:544]}(V^T)_{[384:544,:]}" + }, + { + "bbox": [ + 55, + 460, + 296, + 628 + ], + "type": "text", + "content": ", and the bottom 320 minor components " + }, + { + "bbox": [ + 55, + 460, + 296, + 628 + ], + "type": "inline_equation", + "content": "U_{[:,-320:]} \\Sigma_{[-320:]}(V^T)_{[-320,:]}" + }, + { + "bbox": [ + 55, + 460, + 296, + 628 + ], + "type": "text", + "content": ". We then truncate each group separately and analyze the classes that the rank-reduced model subsequently fails to classify (see Fig. 2 bottom)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 630, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 296, + 715 + ], + "type": "text", + "content": "According to the Eckart-Young-Mirsky theorem and the heavy-tail distribution of singular values, it is evident that the top singular components encapsulate the core knowledge of VFM and should be preserved to maintain generalizability. Conversely, we note that the bottom-componenttruncated model fails to recognize classes that require finegrained or texture-biased features (e.g., specific dog breeds," + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 553, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 300 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 300 + ], + "type": "text", + "content": "wildlife with distinctive textures), as well as those that are context-specific (e.g., cornet, bikini, missile, dock). In contrast, we find that removing middle singular components results in significant performance declines for classes that rely on coarse-grained or shape-biased recognition (e.g., mortarboard, pier, altar, necklace, schooner, horizontal bar). Interestingly, as suggested by differences in the depth of the WordNet hierarchy, higher-value singular components do not function independently but rather interact in a hierarchical and composite manner. For instance, classes with notable accuracy drops can also be clustered into higher-level concepts, such as functionality (e.g., wok, hotpot, pitcher, cauldron, pop bottle, cup/ambulance, minivan, passenger car, police van, garbage truck, limousine) or co-occurrence (e.g., desktop computer, keyboard, notebook, desk). Additionally, almost all of these objects are commonly found in everyday scenes. Consequently, it can be seen that generalizable components are hierarchically and compositionally structured according to the magnitude of singular values." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 300, + 553, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 300, + 553, + 360 + ], + "spans": [ + { + "bbox": [ + 313, + 300, + 553, + 360 + ], + "type": "text", + "content": "Problem Formulation. The entire model consists of a feature extractor " + }, + { + "bbox": [ + 313, + 300, + 553, + 360 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 313, + 300, + 553, + 360 + ], + "type": "text", + "content": ", which combines a VFM parameterized by " + }, + { + "bbox": [ + 313, + 300, + 553, + 360 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 313, + 300, + 553, + 360 + ], + "type": "text", + "content": " with our SoMA adapter parameterized by " + }, + { + "bbox": [ + 313, + 300, + 553, + 360 + ], + "type": "inline_equation", + "content": "\\theta_{S}" + }, + { + "bbox": [ + 313, + 300, + 553, + 360 + ], + "type": "text", + "content": ", and a task-specific decode head " + }, + { + "bbox": [ + 313, + 300, + 553, + 360 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 313, + 300, + 553, + 360 + ], + "type": "text", + "content": ", parameterized by " + }, + { + "bbox": [ + 313, + 300, + 553, + 360 + ], + "type": "inline_equation", + "content": "\\theta_{h}" + }, + { + "bbox": [ + 313, + 300, + 553, + 360 + ], + "type": "text", + "content": ". The optimization objective can be expressed as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 351, + 365, + 553, + 397 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 365, + 553, + 397 + ], + "spans": [ + { + "bbox": [ + 351, + 365, + 553, + 397 + ], + "type": "interline_equation", + "content": "\\underset {\\theta_ {S}, \\theta_ {h}} {\\arg \\min } \\sum_ {i = 1} ^ {N _ {s}} \\mathcal {L} _ {t a s k} \\left(\\mathcal {H} _ {\\theta_ {h}} \\left(\\mathcal {F} _ {\\Phi , \\theta_ {S}} \\left(x _ {i}\\right)\\right), y _ {i}\\right), \\tag {2}", + "image_path": "d8d0747d913487b00cb43b04715c0290c6c3655755e06f0489d3435511c31b8c.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 403, + 555, + 463 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 403, + 555, + 463 + ], + "spans": [ + { + "bbox": [ + 313, + 403, + 555, + 463 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 403, + 555, + 463 + ], + "type": "inline_equation", + "content": "(x_{i},y_{i})" + }, + { + "bbox": [ + 313, + 403, + 555, + 463 + ], + "type": "text", + "content": " denotes an input image and its associated ground truth, drawn from the source domain datasets " + }, + { + "bbox": [ + 313, + 403, + 555, + 463 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_s" + }, + { + "bbox": [ + 313, + 403, + 555, + 463 + ], + "type": "text", + "content": ", comprising " + }, + { + "bbox": [ + 313, + 403, + 555, + 463 + ], + "type": "inline_equation", + "content": "N_{s}" + }, + { + "bbox": [ + 313, + 403, + 555, + 463 + ], + "type": "text", + "content": " samples. Our proposed SoMA framework facilitates generalizing task-specific skills learned from " + }, + { + "bbox": [ + 313, + 403, + 555, + 463 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_s" + }, + { + "bbox": [ + 313, + 403, + 555, + 463 + ], + "type": "text", + "content": " to the unseen target domain datasets " + }, + { + "bbox": [ + 313, + 403, + 555, + 463 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_t" + }, + { + "bbox": [ + 313, + 403, + 555, + 463 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 474, + 421, + 487 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 474, + 421, + 487 + ], + "spans": [ + { + "bbox": [ + 313, + 474, + 421, + 487 + ], + "type": "text", + "content": "4. Proposed Method" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 491, + 555, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 491, + 555, + 563 + ], + "spans": [ + { + "bbox": [ + 313, + 491, + 555, + 563 + ], + "type": "text", + "content": "In this section, we present our parameter-efficient adaptation methods specifically tailored for DG dense prediction tasks, inspired by the insights from our SVD analysis. Subsequently, we propose a block freeze strategy and an annealing weight decay scheme to strike an optimal trade-off between generalizability and discriminability." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 567, + 528, + 581 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 567, + 528, + 581 + ], + "spans": [ + { + "bbox": [ + 313, + 567, + 528, + 581 + ], + "type": "text", + "content": "4.1. Minor Singular Components Adaptation" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 585, + 554, + 681 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 585, + 554, + 681 + ], + "spans": [ + { + "bbox": [ + 313, + 585, + 554, + 681 + ], + "type": "text", + "content": "To effectively preserve the integrity of generalizable representations in pre-trained weights, we propose Singular value decomposed Minor components Adaptation (SoMA). Firstly, SoMA performs singular value decomposition (SVD) on the pre-trained weight matrices, including those from the attention layers, multi-layer perceptron (MLP) layers, and more. The weight matrix " + }, + { + "bbox": [ + 313, + 585, + 554, + 681 + ], + "type": "inline_equation", + "content": "W \\in \\mathbb{R}^{m \\times n}" + }, + { + "bbox": [ + 313, + 585, + 554, + 681 + ], + "type": "text", + "content": " is decomposed using SVD as follows:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 375, + 684, + 555, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 375, + 684, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 375, + 684, + 555, + 715 + ], + "type": "interline_equation", + "content": "W = U \\Sigma V ^ {T} = \\sum_ {i = 1} ^ {R} \\sigma_ {i} u _ {i} v _ {i} ^ {T}, \\tag {3}", + "image_path": "1f5b1c986e19f5aa9027ed05abf210cb6098e11c413f6a6655902968f26b76a9.jpg" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25605" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 65, + 65, + 288, + 219 + ], + "blocks": [ + { + "bbox": [ + 65, + 65, + 288, + 219 + ], + "lines": [ + { + "bbox": [ + 65, + 65, + 288, + 219 + ], + "spans": [ + { + "bbox": [ + 65, + 65, + 288, + 219 + ], + "type": "image", + "image_path": "3ecc3b511cc76cad4eae19078ee87e110856076e5eaf3e9ad5efb87bb8f39f3d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 221, + 295, + 310 + ], + "lines": [ + { + "bbox": [ + 55, + 221, + 295, + 310 + ], + "spans": [ + { + "bbox": [ + 55, + 221, + 295, + 310 + ], + "type": "text", + "content": "Figure 3. The inherent generalization capabilities of the early blocks of VFM. Top. We apply PCA on the extracted intermediate features (8th block of DINOv2-Large) and visualize the top three leading components. Bottom. Class-wise IoU comparison for rare classes [29] under the GTAV " + }, + { + "bbox": [ + 55, + 221, + 295, + 310 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 55, + 221, + 295, + 310 + ], + "type": "text", + "content": " Cityscapes setting, focusing on the impact of freezing the first eight blocks. \"NFEB\" stands for the Number of Frozen Early Blocks. A darker shade of blue indicates enhanced generalization performance relative to the baseline." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 316, + 296, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 316, + 296, + 388 + ], + "spans": [ + { + "bbox": [ + 55, + 316, + 296, + 388 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 316, + 296, + 388 + ], + "type": "inline_equation", + "content": "U = [u_{1}, u_{2}, \\dots, u_{m}] \\in \\mathbb{R}^{m \\times m}" + }, + { + "bbox": [ + 55, + 316, + 296, + 388 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 316, + 296, + 388 + ], + "type": "inline_equation", + "content": "V = [v_{1}, v_{2}, \\dots, v_{n}] \\in \\mathbb{R}^{n \\times n}" + }, + { + "bbox": [ + 55, + 316, + 296, + 388 + ], + "type": "text", + "content": " are the left and right singular vectors, whose columns form an orthonormal basis for " + }, + { + "bbox": [ + 55, + 316, + 296, + 388 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^m" + }, + { + "bbox": [ + 55, + 316, + 296, + 388 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 316, + 296, + 388 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^n" + }, + { + "bbox": [ + 55, + 316, + 296, + 388 + ], + "type": "text", + "content": " respectively, " + }, + { + "bbox": [ + 55, + 316, + 296, + 388 + ], + "type": "inline_equation", + "content": "\\Sigma \\in \\mathbb{R}^{m \\times n}" + }, + { + "bbox": [ + 55, + 316, + 296, + 388 + ], + "type": "text", + "content": " is a diagonal matrix whose entries " + }, + { + "bbox": [ + 55, + 316, + 296, + 388 + ], + "type": "inline_equation", + "content": "\\sigma_{i}" + }, + { + "bbox": [ + 55, + 316, + 296, + 388 + ], + "type": "text", + "content": " are singular values arranged in descending order, and " + }, + { + "bbox": [ + 55, + 316, + 296, + 388 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 55, + 316, + 296, + 388 + ], + "type": "text", + "content": " denotes the rank of " + }, + { + "bbox": [ + 55, + 316, + 296, + 388 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 55, + 316, + 296, + 388 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 55, + 316, + 296, + 388 + ], + "type": "inline_equation", + "content": "R \\leq \\min(m, n)" + }, + { + "bbox": [ + 55, + 316, + 296, + 388 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 388, + 296, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 388, + 296, + 472 + ], + "spans": [ + { + "bbox": [ + 55, + 388, + 296, + 472 + ], + "type": "text", + "content": "We then divide the singular components into two groups according to their singular values: the minor singular components with the smallest " + }, + { + "bbox": [ + 55, + 388, + 296, + 472 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 55, + 388, + 296, + 472 + ], + "type": "text", + "content": " singular values " + }, + { + "bbox": [ + 55, + 388, + 296, + 472 + ], + "type": "inline_equation", + "content": "U_{[:,-r:]} \\Sigma_{[-r:]} (V^T)_{[-r:]}" + }, + { + "bbox": [ + 55, + 388, + 296, + 472 + ], + "type": "text", + "content": " and the residual singular components " + }, + { + "bbox": [ + 55, + 388, + 296, + 472 + ], + "type": "inline_equation", + "content": "U_{[:,-r]} \\Sigma_{[:-r]} (V^T)_{[:-r:]}" + }, + { + "bbox": [ + 55, + 388, + 296, + 472 + ], + "type": "text", + "content": ". As illustrated in Fig. 1 right, we employ the QR-type reconstruction of minor singular components to initialize the adapter for fine-tuning:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 472, + 294, + 488 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 106, + 472, + 294, + 488 + ], + "spans": [ + { + "bbox": [ + 106, + 472, + 294, + 488 + ], + "type": "interline_equation", + "content": "B = U _ {[:, - r: ]} \\sqrt {\\Sigma} _ {[ - r: ]} \\in \\mathbb {R} ^ {m \\times r}, \\tag {4}", + "image_path": "1e5add122b16c0e64ef165bc5d6a48921418e3c2e383cd8e1cbfb89ee07eb4c8.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 107, + 490, + 294, + 506 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 490, + 294, + 506 + ], + "spans": [ + { + "bbox": [ + 107, + 490, + 294, + 506 + ], + "type": "interline_equation", + "content": "A = \\sqrt {\\Sigma} _ {[ - r: ]} \\left(V ^ {T}\\right) _ {[ - r:,.. ]} \\in \\mathbb {R} ^ {r \\times n}. \\tag {5}", + "image_path": "94723e152161fd18727d009e7d7ef2a8b853c266ce77506c4da05794fee8b49a.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 512, + 296, + 596 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 512, + 296, + 596 + ], + "spans": [ + { + "bbox": [ + 55, + 512, + 296, + 596 + ], + "type": "text", + "content": "In the fine-tuning process, the frozen residual matrix " + }, + { + "bbox": [ + 55, + 512, + 296, + 596 + ], + "type": "inline_equation", + "content": "W_{res}" + }, + { + "bbox": [ + 55, + 512, + 296, + 596 + ], + "type": "text", + "content": " is defined as " + }, + { + "bbox": [ + 55, + 512, + 296, + 596 + ], + "type": "inline_equation", + "content": "W - BA" + }, + { + "bbox": [ + 55, + 512, + 296, + 596 + ], + "type": "text", + "content": " to prevent numerical errors introduced during the SVD step, while maintaining the pretrained generalization capability at the beginning of fine-tuning. After fine-tuning, the learned low-rank matrices " + }, + { + "bbox": [ + 55, + 512, + 296, + 596 + ], + "type": "inline_equation", + "content": "B'" + }, + { + "bbox": [ + 55, + 512, + 296, + 596 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 512, + 296, + 596 + ], + "type": "inline_equation", + "content": "A'" + }, + { + "bbox": [ + 55, + 512, + 296, + 596 + ], + "type": "text", + "content": " can be merged into " + }, + { + "bbox": [ + 55, + 512, + 296, + 596 + ], + "type": "inline_equation", + "content": "W_{res}" + }, + { + "bbox": [ + 55, + 512, + 296, + 596 + ], + "type": "text", + "content": ", as in LoRA [30], ensuring no additional computation overhead during inference:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 60, + 600, + 295, + 628 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 600, + 295, + 628 + ], + "spans": [ + { + "bbox": [ + 60, + 600, + 295, + 628 + ], + "type": "interline_equation", + "content": "y = \\left(W \\underbrace {- B A + B ^ {\\prime} A ^ {\\prime}} _ {\\Delta W _ {S o M A}}\\right) x = \\left(W _ {r e s} + B ^ {\\prime} A ^ {\\prime}\\right) x = W ^ {\\prime} x. \\tag {6}", + "image_path": "ef1c87df144d041d6b3533112eedb885a5679c95329e3deb7bdc782acc04edb6.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 630, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 296, + 715 + ], + "type": "text", + "content": "Importantly, SoMA offers several key advantages for tackling the DG problem: i) SoMA initializes the adapter using components orthogonal to the principal components, thus minimizing interference with pre-trained representations. For the batched version of the forward pass described in Eq. 1 (with " + }, + { + "bbox": [ + 55, + 630, + 296, + 715 + ], + "type": "inline_equation", + "content": "X \\in \\mathbb{R}^{n \\times b}" + }, + { + "bbox": [ + 55, + 630, + 296, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 630, + 296, + 715 + ], + "type": "inline_equation", + "content": "Y \\in \\mathbb{R}^{m \\times b}" + }, + { + "bbox": [ + 55, + 630, + 296, + 715 + ], + "type": "text", + "content": "), the gradients of " + }, + { + "bbox": [ + 55, + 630, + 296, + 715 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 55, + 630, + 296, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 630, + 296, + 715 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 55, + 630, + 296, + 715 + ], + "type": "text", + "content": " are " + }, + { + "bbox": [ + 55, + 630, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\frac{\\partial \\mathcal{L}}{\\partial B} = \\frac{\\partial \\mathcal{L}}{\\partial Y} X^T A^T" + }, + { + "bbox": [ + 55, + 630, + 296, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 630, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\frac{\\partial \\mathcal{L}}{\\partial A} = B^T \\frac{\\partial \\mathcal{L}}{\\partial Y} X^T" + }, + { + "bbox": [ + 55, + 630, + 296, + 715 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 318, + 60, + 553, + 159 + ], + "blocks": [ + { + "bbox": [ + 318, + 60, + 553, + 159 + ], + "lines": [ + { + "bbox": [ + 318, + 60, + 553, + 159 + ], + "spans": [ + { + "bbox": [ + 318, + 60, + 553, + 159 + ], + "type": "table", + "html": "
SMRBlockiΔWSoMAΔW*SoMAΔWLoRA
|uiTΔVvi/σi12th0-2550.0750.084 (↓17%)0.101
256-5110.0940.094 (↓43%)0.166
512-7670.1570.165 (↓40%)0.275
768-10234.1046.152 (↑21%)5.095
24th0-2550.0970.095 (↓24%)0.125
256-5110.1050.108 (↓35%)0.166
512-7670.2060.189 (↓41%)0.323
768-10232.7813.620 (↑26%)2.865
", + "image_path": "c63116b68fd7fe6fbc56be6565798ab94e0fcb9c16788a54a09e563a44ca7326.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 162, + 555, + 206 + ], + "lines": [ + { + "bbox": [ + 313, + 162, + 555, + 206 + ], + "spans": [ + { + "bbox": [ + 313, + 162, + 555, + 206 + ], + "type": "text", + "content": "Table 1. Singular modulation ratio comparison. * indicates the adaptation matrix trained using the annealing weight decay strategy. The weight matrices are taken from the query projection layers of the 12th and 24th attention blocks in DINOv2-Large [53]." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 213, + 555, + 320 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 213, + 555, + 320 + ], + "spans": [ + { + "bbox": [ + 313, + 213, + 555, + 320 + ], + "type": "text", + "content": "Therefore, unlike LoRA, which randomly initializes " + }, + { + "bbox": [ + 313, + 213, + 555, + 320 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 313, + 213, + 555, + 320 + ], + "type": "text", + "content": ", our approach facilitates convergence along singular directions that preserve the integrity of the generalizable components. ii) Consistent with prior DG methods [13, 40, 41, 84] that guide models to prioritize shape over texture, thereby achieving superior out-of-distribution performance, SoMA effectively preserves the domain generalizable shape-biased features of VFMs by tuning only the minor singular components (see Fig. 2 bottom and Tab. 1)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 321, + 554, + 405 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 321, + 554, + 405 + ], + "spans": [ + { + "bbox": [ + 313, + 321, + 554, + 405 + ], + "type": "text", + "content": "Discussion. To assess whether the proposed adaptation matrix " + }, + { + "bbox": [ + 313, + 321, + 554, + 405 + ], + "type": "inline_equation", + "content": "\\Delta W_{SoMA}" + }, + { + "bbox": [ + 313, + 321, + 554, + 405 + ], + "type": "text", + "content": " minimally correlates with the top singular directions of the pre-trained weight matrix " + }, + { + "bbox": [ + 313, + 321, + 554, + 405 + ], + "type": "inline_equation", + "content": "W_0" + }, + { + "bbox": [ + 313, + 321, + 554, + 405 + ], + "type": "text", + "content": ", we project " + }, + { + "bbox": [ + 313, + 321, + 554, + 405 + ], + "type": "inline_equation", + "content": "\\Delta W" + }, + { + "bbox": [ + 313, + 321, + 554, + 405 + ], + "type": "text", + "content": " onto the singular vectors of " + }, + { + "bbox": [ + 313, + 321, + 554, + 405 + ], + "type": "inline_equation", + "content": "W_0" + }, + { + "bbox": [ + 313, + 321, + 554, + 405 + ], + "type": "text", + "content": ". Specifically, we quantify the extent of interference by " + }, + { + "bbox": [ + 313, + 321, + 554, + 405 + ], + "type": "inline_equation", + "content": "\\Delta W" + }, + { + "bbox": [ + 313, + 321, + 554, + 405 + ], + "type": "text", + "content": " on each " + }, + { + "bbox": [ + 313, + 321, + 554, + 405 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 321, + 554, + 405 + ], + "type": "text", + "content": "-th singular direction through a singular modulation ratio (SMR), which is formally defined as follows:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 388, + 408, + 553, + 434 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 388, + 408, + 553, + 434 + ], + "spans": [ + { + "bbox": [ + 388, + 408, + 553, + 434 + ], + "type": "interline_equation", + "content": "\\mathrm {S M R} _ {i} = \\left| \\frac {u _ {i} ^ {\\mathrm {T}} \\Delta W v _ {i}}{\\sigma_ {i}} \\right|. \\tag {7}", + "image_path": "6431649e2ffc3732500e6789f8a1cab3f4e8860a9cdbcee2c17e15e9d5aba870.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 437, + 555, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 555, + 570 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 555, + 570 + ], + "type": "text", + "content": "For ease of analysis, we partition the 1024 singular directions into four equal groups, calculate the average SMR for each group, and then compare the resulting values between SoMA and LoRA. As shown in Tab. 1, the SoMA initialization scheme proves more effective than LoRA in minimizing interference with singular components with higher singular values. In summary, our method effectively preserves the structure of the VFM's generalizable representations while adapting to downstream tasks, resulting in significant improvements on DG benchmarks compared to SOTA baselines, as confirmed by our experimental results." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 578, + 440, + 590 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 578, + 440, + 590 + ], + "spans": [ + { + "bbox": [ + 313, + 578, + 440, + 590 + ], + "type": "text", + "content": "4.2. Freezing Early Blocks" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "content": "In this section, we explore the distribution of generalizable components at the block level. Drawing on recent domain adaptation methods [44, 81], which find that the early blocks of a model primarily engage in domain-specific feature extraction and should therefore be updated, we conjecture that in the context of leveraging VFM for DG, the early blocks should instead remain unchanged. Intuitively, these initial blocks of VFMs are adept at projecting low-level features from diverse domains into domain-invariant semantic representations. In other words, while adjusting" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25606" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "content": "early blocks may help bridge input-level domain gaps in domain adaptation, freezing the early blocks of VFMs is essential for DG to prevent overfitting to the source domain and to retain the model's generalization capability." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 121, + 295, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 121, + 295, + 312 + ], + "spans": [ + { + "bbox": [ + 55, + 121, + 295, + 312 + ], + "type": "text", + "content": "To validate our hypothesis, we perform PCA analysis on the features extracted from the early blocks of VFM and compare the performance on rare classes when these blocks are frozen, as illustrated in Fig. 3. Despite considerable domain shifts in style and content, the principal components align well with the image layout and maintain consistent semantic information (e.g., the horses have similar colors). Furthermore, simply freezing the early blocks yields a notable improvement in IoU for rare classes, indicating that updating these blocks can severely disrupt the learned representations of various classes in VFMs. Thus, by default, we opt to freeze all blocks up to the one that generates the feature map fed as the highest resolution input to the segmentation/detection head (e.g., the first eight blocks). Detailed ablations of varying the number of frozen early blocks are presented in the Supplemental." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 319, + 194, + 332 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 319, + 194, + 332 + ], + "spans": [ + { + "bbox": [ + 55, + 319, + 194, + 332 + ], + "type": "text", + "content": "4.3. Annealing Weight Decay" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 337, + 295, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 337, + 295, + 541 + ], + "spans": [ + { + "bbox": [ + 55, + 337, + 295, + 541 + ], + "type": "text", + "content": "Our methods are specialized in preserving the pre-trained knowledge of VFMs, but this may come at the cost of diminished discriminability. One can remove weight decay or set its coefficient to a very small value to enhance discriminability. However, doing so risks increased interference of the tuned parameters with the principal components of the pre-trained weights. Therefore, based on the observation that regularization primarily affects the early phases of learning by guiding a model towards having good generalization properties [23], we leverage an Annealing Weight Decay (AWD) scheme. Specifically, AWD starts with a relatively large weight decay coefficient and progressively reduces it to zero as training progresses. By default, we use a cosine schedule for AWD. As shown in Tab. 1 (" + }, + { + "bbox": [ + 55, + 337, + 295, + 541 + ], + "type": "inline_equation", + "content": "\\Delta W_{SoMA}" + }, + { + "bbox": [ + 55, + 337, + 295, + 541 + ], + "type": "text", + "content": " vs. " + }, + { + "bbox": [ + 55, + 337, + 295, + 541 + ], + "type": "inline_equation", + "content": "\\Delta W_{SoMA}^{*}" + }, + { + "bbox": [ + 55, + 337, + 295, + 541 + ], + "type": "text", + "content": ") and Tab. 7, AWD integrates seamlessly with SoMA, improving discriminability without compromising SoMA's ability to preserve pre-trained knowledge." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 552, + 136, + 565 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 552, + 136, + 565 + ], + "spans": [ + { + "bbox": [ + 55, + 552, + 136, + 565 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 570, + 294, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 294, + 654 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 294, + 654 + ], + "type": "text", + "content": "We conduct a variety of experiments to showcase the efficacy of SoMA on DG benchmarks including domain generalized semantic segmentation (DGSS) and domain generalized object detection (DGOD). We then perform an ablation study on SoMA, emphasizing the contribution of each component to the DG performance. Lastly, we extend our approach beyond dense predictions to generative modeling." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 662, + 171, + 674 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 662, + 171, + 674 + ], + "spans": [ + { + "bbox": [ + 55, + 662, + 171, + 674 + ], + "type": "text", + "content": "5.1. Experimental Setup" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "content": "Datasets. For DGSS, we use three synthetic datasets (GTAV [62], SYNTHIA [64], UrbanSyn [24]), and four real-world datasets (Cityscapes [14], BDD100K [80], Map-" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 316, + 71, + 553, + 431 + ], + "blocks": [ + { + "bbox": [ + 316, + 71, + 553, + 431 + ], + "lines": [ + { + "bbox": [ + 316, + 71, + 553, + 431 + ], + "spans": [ + { + "bbox": [ + 316, + 71, + 553, + 431 + ], + "type": "table", + "html": "
Synthetic-to-Real GeneralizationTest Domains (mIoU in %)
MethodsBackboneParams.*→Citys.→BDD→Map.Avg.
Single-source DGSS Trained on GTAV
○ CLOUDS [3]CLIP-CN-L0.0M60.2057.4067.0061.50
○ VLTSeg [33]EVA02-L304.2M65.3058.3066.0063.20
○ Rein [74]EVA02-L3.0M65.3060.5064.9063.60
○ FADA [4]EVA02-L11.7M66.7061.9066.1064.90
○ tqdm [54]EVA02-L304.2M68.8859.1870.1066.05
○ SoMA (Ours)EVA02-L5.1M68.0560.8168.3365.73
○ SoMA (Ours)EVA02-L5.1M69.9462.4868.3366.92
○ DoRA [48]DINOv2-L7.5M66.1259.3167.0764.17
○ VPT [34]DINOv2-L3.7M68.7558.6468.3265.24
○ SET [79]DINOv2-L6.1M68.0661.6467.6865.79
○ FADA [4]DINOv2-L11.7M68.2361.9468.0966.09
○ AdaptFormer [11]DINOv2-L6.3M70.1059.8168.7766.23
○ SSF [47]DINOv2-L0.5M68.9761.3068.7766.35
○ LoRA [30]DINOv2-L7.3M70.1360.1370.4266.89
○ Rein† [74]DINOv2-L3.0M70.6862.5169.6167.60
○ SoMA (Ours)DINOv2-L4.9M71.8261.3171.6768.27
○ SoMA (Ours)DINOv2-L4.9M73.6363.3370.9869.31
Multi-source DGSS Trained on GTAV + SYNTHIA
○ Rein† [74]DINOv2-L3.0M72.1761.5370.6968.13
○ SoMA (Ours)DINOv2-L4.9M73.1661.9072.7369.26
○ SoMA (Ours)DINOv2-L4.9M74.8563.5973.9270.79
Multi-source DGSS Trained on GTAV + SYNTHIA + UrbanSyn
○ FFTDINOv2-L304.2M75.9060.9372.8069.88
○ SoMA (Ours)DINOv2-L4.9M77.3362.7874.9371.68
○ FFT‡DINOv2-L307.3M77.0661.8175.0971.32
○ Rein‡ [74]DINOv2-L3.0M78.4262.2074.4971.70
○ SoMA‡ (Ours)DINOv2-L4.9M79.2263.8476.3073.12
○ FreezeDINOv2-G0.0M76.0861.9872.2370.10
○ FFTDINOv2-G1.1B76.9061.6973.5370.71
○ SoMA (Ours)DINOv2-G6.6M78.3963.7575.1672.43
○ SoMA (Ours)DINOv2-G6.6M80.3765.6776.1874.07
", + "image_path": "c1ead2381f29888e2eafca73abf8131649c62974073bd7bb5cc140f3ffb4f35b.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 433, + 555, + 455 + ], + "lines": [ + { + "bbox": [ + 314, + 433, + 555, + 455 + ], + "spans": [ + { + "bbox": [ + 314, + 433, + 555, + 455 + ], + "type": "text", + "content": "Table 2. Comparison of the proposed SoMA with existing DGSS " + }, + { + "bbox": [ + 314, + 433, + 555, + 455 + ], + "type": "inline_equation", + "content": "\\circ" + }, + { + "bbox": [ + 314, + 433, + 555, + 455 + ], + "type": "text", + "content": " and PEFT " + }, + { + "bbox": [ + 314, + 433, + 555, + 455 + ], + "type": "inline_equation", + "content": "\\circ" + }, + { + "bbox": [ + 314, + 433, + 555, + 455 + ], + "type": "text", + "content": " methods under various synthetic-to-real settings." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 460, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 460, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 460, + 555, + 713 + ], + "type": "text", + "content": "illary [51], ACDC [66]). Specifically, GTAV, a large-scale dataset generated from the game engine, comprises 24,966 images. SYNTHIA consists of 9,400 photo-realistic synthetic images. UrbanSyn provides 7,539 driving scene images. Cityscapes is an autonomous driving dataset containing 2,975 training and 500 validation images " + }, + { + "bbox": [ + 313, + 460, + 555, + 713 + ], + "type": "inline_equation", + "content": "(2048\\times 1024)" + }, + { + "bbox": [ + 313, + 460, + 555, + 713 + ], + "type": "text", + "content": ". BDD and Mapillary have 1,000 " + }, + { + "bbox": [ + 313, + 460, + 555, + 713 + ], + "type": "inline_equation", + "content": "(1280\\times 720)" + }, + { + "bbox": [ + 313, + 460, + 555, + 713 + ], + "type": "text", + "content": " and 2,000 " + }, + { + "bbox": [ + 313, + 460, + 555, + 713 + ], + "type": "inline_equation", + "content": "(1920\\times 1080)" + }, + { + "bbox": [ + 313, + 460, + 555, + 713 + ], + "type": "text", + "content": " validation images, respectively. ACDC is a driving scene dataset under adverse conditions, including night, snow, fog, and rain. For DGOD, we use the urban scene detection dataset introduced by [75]. The dataset consists of five different weather conditions: Daytime-Sunny (DS), Night-Clear (NC), Night-Rainy (NR), Dusk-Rainy (DR), and Daytime-Foggy (DF). Daytime-Sunny serves as the source domain, offering 26,518 images, with 8,313 used for in-domain evaluation, while the remaining adverse weather conditions are employed as unseen target domains. Implementation details. Since our method focuses on preserving pre-trained knowledge, we mainly evaluate its effectiveness using VFMs like EVA02 [22] and DINOv2 [53] as the backbone. To further elevate perfor" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25607" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 57, + 71, + 292, + 185 + ], + "blocks": [ + { + "bbox": [ + 57, + 71, + 292, + 185 + ], + "lines": [ + { + "bbox": [ + 57, + 71, + 292, + 185 + ], + "spans": [ + { + "bbox": [ + 57, + 71, + 292, + 185 + ], + "type": "table", + "html": "
Real-to-Real GeneralizationTest Domains (mIoU in %)
MethodsBackboneParams.*→BDD→Map.Avg.
Single-source DGSS Trained on Cityscapes
○ HGFormer [17]Swin-L196.0M61.5072.1066.80
○ CMFormer [5]Swin-L196.0M62.6073.6068.10
○ tqdm [54]EVA02-L304.2M64.7276.1570.44
○ FADA [4]DINOv2-L11.7M65.1275.8670.49
○ Rein† [74]DINOv2-L3.0M66.5375.1870.86
○ SoMA (Ours)DINOv2-L4.9M67.0276.4571.74
• SoMA (Ours)DINOv2-L4.9M68.0877.8772.98
", + "image_path": "3fb0ba9c077ccb735bbc919c486c811d03286701c1a3f92da63b76e2b6d0b330.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 58, + 201, + 292, + 308 + ], + "blocks": [ + { + "bbox": [ + 99, + 186, + 253, + 197 + ], + "lines": [ + { + "bbox": [ + 99, + 186, + 253, + 197 + ], + "spans": [ + { + "bbox": [ + 99, + 186, + 253, + 197 + ], + "type": "text", + "content": "Table 3. Real-to-real DGSS comparison." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 201, + 292, + 308 + ], + "lines": [ + { + "bbox": [ + 58, + 201, + 292, + 308 + ], + "spans": [ + { + "bbox": [ + 58, + 201, + 292, + 308 + ], + "type": "table", + "html": "
Clear-to-Adverse Weather MethodsACDC [66] Test Domains (mIoU in %)
→Night→Snow→Fog→RainAll
Single-source DGSS Trained on Cityscapes
○ HGFormer [17]52.768.669.972.067.2
○ SoMA (Ours)61.777.374.777.874.4
○ VLTSeg‡ [33]----77.9
○ Rein‡ [74]70.679.576.479.477.6
○ FFT‡68.980.376.979.777.7
○ SoMA‡ (Ours)73.279.876.880.278.8
", + "image_path": "4160bf1ad54e725ba2b2b6d592abe903e2c7c60f430d9236496b51fb16f43146.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 58, + 323, + 294, + 404 + ], + "blocks": [ + { + "bbox": [ + 84, + 309, + 268, + 319 + ], + "lines": [ + { + "bbox": [ + 84, + 309, + 268, + 319 + ], + "spans": [ + { + "bbox": [ + 84, + 309, + 268, + 319 + ], + "type": "text", + "content": "Table 4. Results on Cityscapes " + }, + { + "bbox": [ + 84, + 309, + 268, + 319 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 84, + 309, + 268, + 319 + ], + "type": "text", + "content": " ACDC test set." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 323, + 294, + 404 + ], + "lines": [ + { + "bbox": [ + 58, + 323, + 294, + 404 + ], + "spans": [ + { + "bbox": [ + 58, + 323, + 294, + 404 + ], + "type": "table", + "html": "
Data EfficiencyTest Domains (mIoU in %)
MethodsBackboneParams.*→Citys.→BDD→Map.Avg.
DGSS Pre-trained on GTAV + SYNTHIA + UrbanSyn → 1/16 of Cityscapes
○ FFT‡DINOv2-L307.3M81.5365.2275.7374.16
○ Rein‡ [74]DINOv2-L3.0M82.5864.7673.7373.69
○ SoMA‡ (Ours)DINOv2-L4.9M82.5066.9977.0275.50
", + "image_path": "42291a7c90ef4506ea58d0aad9208d8bbad8030a4f584d545481f1caf9ef731e.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 426, + 295, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 426, + 295, + 498 + ], + "spans": [ + { + "bbox": [ + 55, + 426, + 295, + 498 + ], + "type": "text", + "content": "mance, we leverage state-of-the-art decode heads such as Mask2Former [12] and Co-DETR [86]. Extensive experiments with various backbones and heads are provided in the Supplemental. By default, SoMA is applied to all linear layers within the self-attention and MLP, with a rank of 16. Detailed training settings are presented in the Supplemental." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 505, + 241, + 517 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 505, + 241, + 517 + ], + "spans": [ + { + "bbox": [ + 55, + 505, + 241, + 517 + ], + "type": "text", + "content": "5.2. Comparison with State-of-the-Arts" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 521, + 295, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 521, + 295, + 640 + ], + "spans": [ + { + "bbox": [ + 54, + 521, + 295, + 640 + ], + "type": "text", + "content": "We compare SoMA with previous SOTA baselines, achieving the highest performance across all DG scenarios without adding inference latency or regularization loss. In our tables, * indicates trainable parameters in the backbones. Marker • refers to models tested with multi-scale flip augmentation. † represents re-implemented test results using official checkpoints for a fair comparison, and ‡ denotes training on images cropped to " + }, + { + "bbox": [ + 54, + 521, + 295, + 640 + ], + "type": "inline_equation", + "content": "1024 \\times 1024" + }, + { + "bbox": [ + 54, + 521, + 295, + 640 + ], + "type": "text", + "content": "; otherwise, the default is " + }, + { + "bbox": [ + 54, + 521, + 295, + 640 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 54, + 521, + 295, + 640 + ], + "type": "text", + "content": ". A comprehensive comparison with earlier works can be found in the Supplemental." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": "Synthetic-to-real DGSS. In Tab. 2, we adopt " + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "inline_equation", + "content": "GTAV \\rightarrow \\{C_i\\}" + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": " as the basic setting, and demonstrate scalability of SoMA by incorporating additional synthetic datasets into the training process. We report the mean Intersection over Union (mIoU) and benchmark our results against SOTA methods. SoMA consistently outperforms" + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 316, + 71, + 554, + 266 + ], + "blocks": [ + { + "bbox": [ + 108, + 407, + 244, + 417 + ], + "lines": [ + { + "bbox": [ + 108, + 407, + 244, + 417 + ], + "spans": [ + { + "bbox": [ + 108, + 407, + 244, + 417 + ], + "type": "text", + "content": "Table 5. Real-world data efficiency." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 316, + 71, + 554, + 266 + ], + "lines": [ + { + "bbox": [ + 316, + 71, + 554, + 266 + ], + "spans": [ + { + "bbox": [ + 316, + 71, + 554, + 266 + ], + "type": "table", + "html": "
Clear-to-Adverse WeatherS-DGOD [75] Test Domains (mAP@0.5 in %)
MethodsParams.*DS→NC→DR→NR→DFAvg.
Single-source DGOD Trained on Daytime-Sunny (DS)
Backbone: ResNet101 [27] / Head: Faster R-CNN [61]
S-DGOD [75]42.3M56.136.628.216.633.528.7
CLIP-Gap [72]42.3M51.336.932.318.738.531.6
OA-DG [43]42.3M55.838.033.916.838.331.8
PDOC [46]42.3M53.638.533.719.239.132.6
UFR [49]42.3M58.640.833.219.239.633.2
DivAlign [15]42.3M52.842.538.124.137.235.5
SoMA (Ours)3.1M49.341.937.924.538.235.6
Backbone: DINOv2-L [53] / Head: Co-DETR [86]
Freeze0.0M65.054.255.042.846.949.7
FFT307.3M68.257.156.643.147.251.0
DoRA [48]5.8M69.058.758.045.048.952.7
AdaptFormer [11]6.3M68.958.858.344.449.852.8
LoRA [30]5.5M69.659.658.146.149.553.3
SoMA (Ours)4.9M69.459.359.347.651.054.3
", + "image_path": "31ff3e0c59eca8b356bce0829c1878aa62720dc61651bccaf0b5ba5e6710aa7e.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 347, + 269, + 522, + 279 + ], + "lines": [ + { + "bbox": [ + 347, + 269, + 522, + 279 + ], + "spans": [ + { + "bbox": [ + 347, + 269, + 522, + 279 + ], + "type": "text", + "content": "Table 6. Domain generalized object detection." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 287, + 553, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 287, + 553, + 396 + ], + "spans": [ + { + "bbox": [ + 313, + 287, + 553, + 396 + ], + "type": "text", + "content": "VFM-based DGSS methods and PEFT approaches using both vision-language (EVA02) and vision-only (DINOv2) VFM backbones. Notably, SoMA achieves superior results over LoRA by minimally interfering with the hierarchical structure of VFM's pre-trained knowledge. Moreover, SoMA exhibits superior scalability across data, model, and input size, with improvements becoming more pronounced when test-time augmentation is applied. We provide comparisons regarding model efficiency in the Supplemental." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 397, + 553, + 481 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 397, + 553, + 481 + ], + "spans": [ + { + "bbox": [ + 313, + 397, + 553, + 481 + ], + "type": "text", + "content": "Real-to-real DGSS. In Tables 3 and 4, we conduct experiments under the Citys. " + }, + { + "bbox": [ + 313, + 397, + 553, + 481 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 313, + 397, + 553, + 481 + ], + "type": "text", + "content": " {BDD, Map.}/ACDC settings. When combined with DINOv2-L backbone, SoMA consistently gains the best results across all settings. This highlights the effectiveness of our method as a strong solution for real-world deployment scenarios, where domain shifts caused by geographic or weather variations are common." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 483, + 554, + 603 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 483, + 554, + 603 + ], + "spans": [ + { + "bbox": [ + 313, + 483, + 554, + 603 + ], + "type": "text", + "content": "Real-world data efficiency. Following [74], we evaluate models under constrained conditions using a limited set of real-world images. Specifically, models pre-trained on synthetic datasets " + }, + { + "bbox": [ + 313, + 483, + 554, + 603 + ], + "type": "inline_equation", + "content": "GTAV + SYNTHIA + UrbanSyn" + }, + { + "bbox": [ + 313, + 483, + 554, + 603 + ], + "type": "text", + "content": " (marked with " + }, + { + "bbox": [ + 313, + 483, + 554, + 603 + ], + "type": "inline_equation", + "content": "\\ddagger" + }, + { + "bbox": [ + 313, + 483, + 554, + 603 + ], + "type": "text", + "content": " in Tab. 2) are fine-tuned on 1/16 of the Cityscapes training set. As shown in Tab. 5, SoMA achieves performance on par with Rein in the source domain, while offering greater robustness on unseen domains. This underscores SoMA's data-efficient adaptability to real-world scenes, while effectively avoiding overfitting to the source domain." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 605, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 554, + 714 + ], + "type": "text", + "content": "Clear-to-adverse weather DGOD. In Tab. 6, we compare SoMA with existing DGOD methods for classic backbonehead setting and with PEFT methods for recent foundation architectures. As expected, SoMA demonstrates only minor improvements in the classic setting; however, in the VFM setting—where there is considerably more knowledge to retain—it shows substantial performance gains. As illustrated by the visualization results in Fig. 4, SoMA makes robust predictions even under diverse challenging scenes." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25608" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 87, + 64, + 261, + 201 + ], + "blocks": [ + { + "bbox": [ + 87, + 64, + 261, + 201 + ], + "lines": [ + { + "bbox": [ + 87, + 64, + 261, + 201 + ], + "spans": [ + { + "bbox": [ + 87, + 64, + 261, + 201 + ], + "type": "image", + "image_path": "d68b013963ef7d21d57c454c99cf70fe31b6bd51aa0a29a38ab5128f9475fb1c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 107, + 205, + 244, + 216 + ], + "lines": [ + { + "bbox": [ + 107, + 205, + 244, + 216 + ], + "spans": [ + { + "bbox": [ + 107, + 205, + 244, + 216 + ], + "type": "text", + "content": "Figure 4. DGOD qualitative results." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 57, + 224, + 296, + 291 + ], + "blocks": [ + { + "bbox": [ + 57, + 224, + 296, + 291 + ], + "lines": [ + { + "bbox": [ + 57, + 224, + 296, + 291 + ], + "spans": [ + { + "bbox": [ + 57, + 224, + 296, + 291 + ], + "type": "table", + "html": "
MethodsParams.*DGSS Avg.DGOD Avg.
Full fine-tuning (baseline)304.2M64.451.0
L + Freezing early blocks201.6M65.0 (↑0.6)51.4 (↑0.4)
L + Tuning principal components4.9M66.1 (↑1.1)53.0 (↑1.6)
L + Tuning minor components4.9M67.7 (↑2.7)53.8 (↑2.4)
L + Annealing weight decay4.9M68.3 (↑0.6)54.3 (↑0.5)
", + "image_path": "c8255192cb53c6ede59c4af8e03c76845065e75be5117f43d0546652cbdd8811.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 57, + 319, + 296, + 369 + ], + "blocks": [ + { + "bbox": [ + 56, + 293, + 296, + 315 + ], + "lines": [ + { + "bbox": [ + 56, + 293, + 296, + 315 + ], + "spans": [ + { + "bbox": [ + 56, + 293, + 296, + 315 + ], + "type": "text", + "content": "Table 7. Effect of our changes evaluated on DG benchmarks. See the full ablation study in the Supplemental." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 319, + 296, + 369 + ], + "lines": [ + { + "bbox": [ + 57, + 319, + 296, + 369 + ], + "spans": [ + { + "bbox": [ + 57, + 319, + 296, + 369 + ], + "type": "table", + "html": "
{Wq,Wk,Wv,Wo} {Wup,Wdown}Params.* →Citys. →BDD →Map. Avg.
✓2.7M68.7661.5070.00
✓2.2M70.8261.3669.86
✓✓4.9M71.8261.3171.67
", + "image_path": "3648921f1fb40f375ac0bd5b506d48c9ded9522ff62341836c39c83b6e69f252.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 57, + 396, + 296, + 436 + ], + "blocks": [ + { + "bbox": [ + 56, + 370, + 296, + 392 + ], + "lines": [ + { + "bbox": [ + 56, + 370, + 296, + 392 + ], + "spans": [ + { + "bbox": [ + 56, + 370, + 296, + 392 + ], + "type": "text", + "content": "Table 8. DGSS performance after applying SoMA to different types of modules (Self-attention / MLP) in DINoV2-Large." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 396, + 296, + 436 + ], + "lines": [ + { + "bbox": [ + 57, + 396, + 296, + 436 + ], + "spans": [ + { + "bbox": [ + 57, + 396, + 296, + 436 + ], + "type": "table", + "html": "
Rank r48163264
DGSS avg. (mIoU in %)66.9167.7168.2767.7667.59
Params.*1.3M2.5M4.9M9.6M19.0M
", + "image_path": "81a90cc0c41496550c442dcd60e0b5e605aa30c8b524a10fc6328d5b01aec149.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 82, + 437, + 270, + 448 + ], + "lines": [ + { + "bbox": [ + 82, + 437, + 270, + 448 + ], + "spans": [ + { + "bbox": [ + 82, + 437, + 270, + 448 + ], + "type": "text", + "content": "Table 9. DGSS performance with different rank " + }, + { + "bbox": [ + 82, + 437, + 270, + 448 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 82, + 437, + 270, + 448 + ], + "type": "text", + "content": " ." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 56, + 457, + 149, + 470 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 457, + 149, + 470 + ], + "spans": [ + { + "bbox": [ + 56, + 457, + 149, + 470 + ], + "type": "text", + "content": "5.3. Ablation Study" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 474, + 295, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 474, + 295, + 569 + ], + "spans": [ + { + "bbox": [ + 55, + 474, + 295, + 569 + ], + "type": "text", + "content": "Component Analysis. A thorough examination of each component of SoMA in the DGOD and DGSS basic setting, as summarized in Tab. 7, demonstrates that all components of SoMA incrementally improve DG performance without adding computational costs. Notably, the marked performance gap with tuning principal singular components [50] suggests that SoMA better preserves the integrity of pretrained representations while learning task-specific features." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 571, + 295, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 571, + 295, + 643 + ], + "spans": [ + { + "bbox": [ + 55, + 571, + 295, + 643 + ], + "type": "text", + "content": "Tuning granularity. Interestingly, as shown in Tables 8 and 9, our method consistently achieves competitive results compared to SOTA baselines across all tuning granularities, indicating that SoMA can be flexibly configured according to the training budget, task difficulty, or domain gap. The best results are achieved with the default setting." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 650, + 188, + 662 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 650, + 188, + 662 + ], + "spans": [ + { + "bbox": [ + 55, + 650, + 188, + 662 + ], + "type": "text", + "content": "5.4. Subject Personalization" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "Domain generalized recognition requires consistent processing of inputs from diverse domains, whereas domain generalized generation involves generating outputs across a range of domains. Although large-scale Text-to-Image" + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 354, + 59, + 512, + 240 + ], + "blocks": [ + { + "bbox": [ + 354, + 59, + 512, + 240 + ], + "lines": [ + { + "bbox": [ + 354, + 59, + 512, + 240 + ], + "spans": [ + { + "bbox": [ + 354, + 59, + 512, + 240 + ], + "type": "image", + "image_path": "c3deea78171b5ae33ee50b141d34f258077fca93b41b0fb3e4bcb49094279cde.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 245, + 554, + 280 + ], + "lines": [ + { + "bbox": [ + 313, + 245, + 554, + 280 + ], + "spans": [ + { + "bbox": [ + 313, + 245, + 554, + 280 + ], + "type": "text", + "content": "Figure 5. Subject Personalization. 1. A dog gracefully leaping in origami style, 2. A dog in watercolor painting style, and 3. A dog soaring through a digital landscape in vector illustration style." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 292, + 555, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 292, + 555, + 520 + ], + "spans": [ + { + "bbox": [ + 313, + 292, + 555, + 520 + ], + "type": "text", + "content": "(T2I) models have convincingly demonstrated this ability, it can be compromised in subject personalization tasks involving fine-tuning. DreamBooth (DB) [65] introduces subject personalization, which fine-tunes a pre-trained T2I model on a few reference images to enable it to generate new visual concepts. DB updates all parameters within specific blocks, potentially impairing the pre-trained capability to generate images across various domains. Therefore, we explore whether SoMA can be seamlessly integrated into the DB framework to preserve pre-trained knowledge. For this task, we utilize Stable Diffusion XL (SDXL) [59] for the T2I diffusion model. Fig. 5 presents a visual comparison between DB and SoMA, using identical hyperparameter settings and sample seeds. As seen, SoMA, similar to the original SDXL model, can generate high-quality images of diverse visual domains (e.g., style, texture, pose) and exhibits superior identity preservation compared to DB with prior preservation (p.p.) loss [65]. More results with varied prompts and subjects are provided in the Supplemental." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 314, + 537, + 388, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 537, + 388, + 550 + ], + "spans": [ + { + "bbox": [ + 314, + 537, + 388, + 550 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "text", + "content": "This paper presents a multifaceted exploration of the distribution of generalizable components in VFMs—examining them at the weight level, block level, and within training dynamics—and introduces the SoMA framework as an effective means to preserve generalizable components. Contrary to recent approaches that passively refine frozen VFM features, SoMA is integrated into all linear layers to effectively facilitate task adaptation, while minimizing interference with generalizable components by initializing the low-rank adapter with minor singular components. Through extensive experiments, we observe that SoMA significantly outperforms the state-of-the-art methods in DGSS, DGOD, and even subject-driven image generation tasks." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25609" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 165, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 165, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 165, + 85 + ], + "type": "text", + "content": "7. Acknowledgments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 128 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 128 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 128 + ], + "type": "text", + "content": "This work was supported by the National Research Foundation (NRF) grant funded by the Korea government (MSIT) [RS-2025-00562400] and [RS-2022-NR068754]." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 140, + 115, + 152 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 140, + 115, + 152 + ], + "spans": [ + { + "bbox": [ + 56, + 140, + 115, + 152 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 159, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 61, + 159, + 296, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 159, + 296, + 203 + ], + "spans": [ + { + "bbox": [ + 61, + 159, + 296, + 203 + ], + "type": "text", + "content": "[1] Woo-Jin Ahn, Geun-Yeong Yang, Hyun-Duck Choi, and Myo-Taeg Lim. Style blind domain generalized semantic segmentation via covariance alignment and semantic consistence contrastive learning. In CVPR, 2024. 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 205, + 296, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 205, + 296, + 258 + ], + "spans": [ + { + "bbox": [ + 61, + 205, + 296, + 258 + ], + "type": "text", + "content": "[2] Muhammad Awais, Muzammal Naseer, Salman Khan, Rao Muhammad Anwer, Hisham Cholakkal, Mubarak Shah, Ming-Hsuan Yang, and Fahad Shahbaz Khan. Foundational models defining a new era in vision: A survey and outlook. arXiv:2307.13721, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 260, + 295, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 260, + 295, + 304 + ], + "spans": [ + { + "bbox": [ + 61, + 260, + 295, + 304 + ], + "type": "text", + "content": "[3] Yasser Benigmim, Subhankar Roy, Slim Essid, Vicky Kalogeiton, and Stephane Lathuilière. Collaborating foundation models for domain generalized semantic segmentation. In CVPR, 2024. 3, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 61, + 306, + 294, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 306, + 294, + 350 + ], + "spans": [ + { + "bbox": [ + 61, + 306, + 294, + 350 + ], + "type": "text", + "content": "[4] Qi Bi, Jingjun Yi, Hao Zheng, Haolan Zhan, Yawen Huang, Wei Ji, Yuexiang Li, and Yefeng Zheng. Learning frequency-adapted vision foundation model for domain generalized semantic segmentation. In NeurlPS, 2024. 3, 6, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 351, + 295, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 351, + 295, + 384 + ], + "spans": [ + { + "bbox": [ + 62, + 351, + 295, + 384 + ], + "type": "text", + "content": "[5] Qi Bi, Shaodi You, and Theo Gevers. Learning content-enhanced mask transformer for domain generalized urban-scene segmentation. In AAAI, 2024. 2, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 386, + 295, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 386, + 295, + 430 + ], + "spans": [ + { + "bbox": [ + 62, + 386, + 295, + 430 + ], + "type": "text", + "content": "[6] Dan Biderman, Jose Gonzalez Ortiz, Jacob Portes, Mansheej Paul, Philip Greengard, Connor Jennings, Daniel King, Sam Havens, Vitaliy Chiley, Jonathan Frankle, et al. Lora learns less and forgets less. TMLR, 2024. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 431, + 295, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 431, + 295, + 463 + ], + "spans": [ + { + "bbox": [ + 62, + 431, + 295, + 463 + ], + "type": "text", + "content": "[7] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, 2020. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 465, + 295, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 465, + 295, + 508 + ], + "spans": [ + { + "bbox": [ + 62, + 465, + 295, + 508 + ], + "type": "text", + "content": "[8] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, 2021. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 510, + 295, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 510, + 295, + 553 + ], + "spans": [ + { + "bbox": [ + 62, + 510, + 295, + 553 + ], + "type": "text", + "content": "[9] Gyusam Chang, Jiwon Lee, Donghyun Kim, Jinkyu Kim, Dongwook Lee, Daehyun Ji, Sujin Jang, and Sangpil Kim. Unified domain generalization and adaptation for multi-view 3d object detection. In NeurIPS, 2024. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 555, + 295, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 555, + 295, + 599 + ], + "spans": [ + { + "bbox": [ + 57, + 555, + 295, + 599 + ], + "type": "text", + "content": "[10] Prithvijit Chattopadhyay, Kartik Sarangmath, Vivek Vijaykumar, and Judy Hoffman. Pasta: Proportional amplitude spectrum training augmentation for syn-to-real domain generalization. In ICCV, 2023. 2, 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 601, + 295, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 601, + 295, + 644 + ], + "spans": [ + { + "bbox": [ + 57, + 601, + 295, + 644 + ], + "type": "text", + "content": "[11] Shoufa Chen, Chongjian Ge, Zhan Tong, Jiangliu Wang, Yibing Song, Jue Wang, and Ping Luo. Adaptformer: Adapting vision transformers for scalable visual recognition. In NeurIPS, 2022. 2, 3, 6, 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 647, + 295, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 647, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 57, + 647, + 295, + 689 + ], + "type": "text", + "content": "[12] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In CVPR, 2022. 1, 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 691, + 295, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 295, + 714 + ], + "type": "text", + "content": "[13] Sungha Choi, Sanghun Jung, Huiwon Yun, Joanne T Kim, Seungryong Kim, and Jaegul Choo. Robustnet: Improving" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "domain generalization in urban-scene segmentation via instance selective whitening. In CVPR, 2021. 2, 3, 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 98, + 553, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 98, + 553, + 151 + ], + "spans": [ + { + "bbox": [ + 316, + 98, + 553, + 151 + ], + "type": "text", + "content": "[14] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In CVPR, 2016. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 153, + 553, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 153, + 553, + 206 + ], + "spans": [ + { + "bbox": [ + 316, + 153, + 553, + 206 + ], + "type": "text", + "content": "[15] Muhammad Sohail Danish, Muhammad Haris Khan, Muhammad Akhtar Munir, M Saquib Sarfraz, and Mohsen Ali. Improving single domain-generalized object detection: A focus on diversification and alignment. In CVPR, 2024. 3, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 209, + 553, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 209, + 553, + 243 + ], + "spans": [ + { + "bbox": [ + 316, + 209, + 553, + 243 + ], + "type": "text", + "content": "[16] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009. 2, 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 245, + 553, + 288 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 245, + 553, + 288 + ], + "spans": [ + { + "bbox": [ + 316, + 245, + 553, + 288 + ], + "type": "text", + "content": "[17] Jian Ding, Nan Xue, Gui-Song Xia, Bernt Schiele, and Dengxin Dai. Hgformer: Hierarchical grouping transformer for domain generalized semantic segmentation. In CVPR, 2023. 3, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 291, + 553, + 356 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 291, + 553, + 356 + ], + "spans": [ + { + "bbox": [ + 316, + 291, + 553, + 356 + ], + "type": "text", + "content": "[18] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 358, + 553, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 358, + 553, + 380 + ], + "spans": [ + { + "bbox": [ + 316, + 358, + 553, + 380 + ], + "type": "text", + "content": "[19] Carl Eckart and Gale Young. The approximation of one matrix by another of lower rank. Psychometrika, 1936. 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 382, + 553, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 382, + 553, + 423 + ], + "spans": [ + { + "bbox": [ + 316, + 382, + 553, + 423 + ], + "type": "text", + "content": "[20] Mohammad Fahes, Tuan-Hung Vu, Andrei Bursuc, Patrick Pérez, and Raoul de Charette. A simple recipe for language-guided domain generalized segmentation. In CVPR, 2024. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 427, + 553, + 471 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 427, + 553, + 471 + ], + "spans": [ + { + "bbox": [ + 316, + 427, + 553, + 471 + ], + "type": "text", + "content": "[21] Qi Fan, Mattia Segu, Yu-Wing Tai, Fisher Yu, Chi-Keung Tang, Bernt Schiele, and Dengxin Dai. Towards robust object detection invariant to real-world domain shifts. In ICLR, 2023. 2, 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 474, + 553, + 506 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 474, + 553, + 506 + ], + "spans": [ + { + "bbox": [ + 316, + 474, + 553, + 506 + ], + "type": "text", + "content": "[22] Yuxin Fang, Quan Sun, Xinggang Wang, Tiejun Huang, Xinlong Wang, and Yue Cao. Eva-02: A visual representation for neon genesis. IVC, 2024. 3, 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 509, + 553, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 509, + 553, + 552 + ], + "spans": [ + { + "bbox": [ + 316, + 509, + 553, + 552 + ], + "type": "text", + "content": "[23] Aditya Sharad Golatkar, Alessandro Achille, and Stefano Soatto. Time matters in regularizing deep networks: Weight decay and data augmentation affect early learning dynamics, matter little near convergence. In NeurIPS, 2019. 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 554, + 553, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 554, + 553, + 609 + ], + "spans": [ + { + "bbox": [ + 316, + 554, + 553, + 609 + ], + "type": "text", + "content": "[24] Jose L Gómez, Manuel Silva, Antonio Seoane, Agnès Borras, Mario Noriega, German Ros, Jose A Iglesias-Guitian, and Antonio M López. All for one, and one for all: Urbansyn dataset, the third musketeer of synthetic driving scenes. arXiv:2312.12176, 2023. 1, 6" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 611, + 553, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 611, + 553, + 643 + ], + "spans": [ + { + "bbox": [ + 316, + 611, + 553, + 643 + ], + "type": "text", + "content": "[25] Junxian He, Chunting Zhou, Xuezhe Ma, Taylor Berg-Kirkpatrick, and Graham Neubig. Towards a unified view of parameter-efficient transfer learning. In ICLR, 2022. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 646, + 553, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 646, + 553, + 678 + ], + "spans": [ + { + "bbox": [ + 316, + 646, + 553, + 678 + ], + "type": "text", + "content": "[26] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Delving deep into rectifiers: Surpassing human-level performance on imagenet classification. In ICCV, 2015. 3" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 680, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 680, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 680, + 553, + 712 + ], + "type": "text", + "content": "[27] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 2, 3, 7" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "25610" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "type": "text", + "content": "[28] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In CVPR, 2022. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 295, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 295, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 295, + 140 + ], + "type": "text", + "content": "[29] Lukas Hoyer, Dengxin Dai, and Luc Van Gool. Daformer: Improving network architectures and training strategies for domain-adaptive semantic segmentation. In CVPR, 2022. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 141, + 294, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 141, + 294, + 184 + ], + "spans": [ + { + "bbox": [ + 56, + 141, + 294, + 184 + ], + "type": "text", + "content": "[30] Edward J Hu, yelong shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In ICLR, 2022. 1, 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 186, + 294, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 186, + 294, + 228 + ], + "spans": [ + { + "bbox": [ + 56, + 186, + 294, + 228 + ], + "type": "text", + "content": "[31] Jiajun Hu, Jian Zhang, Lei Qi, Yinghuan Shi, and Yang Gao. Learn to preserve and diversify: Parameter-efficient group with orthogonal regularization for domain generalization. In ECCV, 2024. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 231, + 294, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 231, + 294, + 274 + ], + "spans": [ + { + "bbox": [ + 56, + 231, + 294, + 274 + ], + "type": "text", + "content": "[32] Wei Huang, Chang Chen, Yong Li, Jiacheng Li, Cheng Li, Fenglong Song, Youliang Yan, and Zhiwei Xiong. Style projected clustering for domain generalized semantic segmentation. In CVPR, 2023. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 275, + 294, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 275, + 294, + 330 + ], + "spans": [ + { + "bbox": [ + 56, + 275, + 294, + 330 + ], + "type": "text", + "content": "[33] Christoph Hummer, Manuel Schwonberg, Liangwei Zhong, Hu Cao, Alois Knoll, and Hanno Gottschalk. Vltseg: Simple transfer of clip-based vision-language representations for domain generalized semantic segmentation. In ACCV, 2024. 3, 6, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 331, + 294, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 331, + 294, + 364 + ], + "spans": [ + { + "bbox": [ + 56, + 331, + 294, + 364 + ], + "type": "text", + "content": "[34] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In ECCV, 2022. 2, 3, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 365, + 294, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 365, + 294, + 419 + ], + "spans": [ + { + "bbox": [ + 56, + 365, + 294, + 419 + ], + "type": "text", + "content": "[35] Yuru Jia, Lukas Hoyer, Shengyu Huang, Tianfu Wang, Luc Van Gool, Konrad Schindler, and Anton Obukhov. Dgin-style: Domain-generalizable semantic segmentation with image diffusion models and stylized semantic control. In ECCV, 2024. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 421, + 294, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 421, + 294, + 453 + ], + "spans": [ + { + "bbox": [ + 56, + 421, + 294, + 453 + ], + "type": "text", + "content": "[36] Xueying Jiang, Jiaxing Huang, Sheng Jin, and Shijian Lu. Domain generalization via balancing training difficulty and model capability. In ICCV, 2023. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 455, + 294, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 455, + 294, + 487 + ], + "spans": [ + { + "bbox": [ + 56, + 455, + 294, + 487 + ], + "type": "text", + "content": "[37] Mengmeng Jing, Xiantong Zhen, Jingjing Li, and Cees GM Snoek. Order-preserving consistency regularization for domain adaptation and generalization. In ICCV, 2023. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 488, + 294, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 488, + 294, + 520 + ], + "spans": [ + { + "bbox": [ + 56, + 488, + 294, + 520 + ], + "type": "text", + "content": "[38] Christoph Kamann and Carsten Rother. Increasing the robustness of semantic segmentation models with painting-by-numbers. In ECCV, 2020. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 522, + 294, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 522, + 294, + 555 + ], + "spans": [ + { + "bbox": [ + 56, + 522, + 294, + 555 + ], + "type": "text", + "content": "[39] Jin Kim, Jiyoung Lee, Jungin Park, Dongbo Min, and Kwanghoon Sohn. Pin the memory: Learning to generalize semantic segmentation. In CVPR, 2022. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 556, + 294, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 556, + 294, + 588 + ], + "spans": [ + { + "bbox": [ + 56, + 556, + 294, + 588 + ], + "type": "text", + "content": "[40] Sunghwan Kim, Dae-hwan Kim, and Hoseong Kim. Texture learning domain randomization for domain generalized segmentation. In ICCV, 2023. 3, 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 590, + 294, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 590, + 294, + 622 + ], + "spans": [ + { + "bbox": [ + 56, + 590, + 294, + 622 + ], + "type": "text", + "content": "[41] Suhyeon Lee, Hongje Seong, Seongwon Lee, and Euntai Kim. Wildnet: Learning domain generalized semantic segmentation from the wild. In CVPR, 2022. 3, 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 624, + 294, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 294, + 655 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 294, + 655 + ], + "type": "text", + "content": "[42] Sohyun Lee, Taeyoung Son, and Suha Kwak. Fifo: Learning fog-invariant features for foggy scene segmentation. In CVPR, 2022. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 56, + 658, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 294, + 689 + ], + "type": "text", + "content": "[43] Wooju Lee, Dasol Hong, Hyungtae Lim, and Hyun Myung. Object-aware domain generalization for object detection. In AAAI, 2024. 3, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "text", + "content": "[44] Yoonho Lee, Annie S Chen, Fahim Tajwar, Ananya Kumar, Huaxiu Yao, Percy Liang, and Chelsea Finn. Surgical fine" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 711 + ], + "type": "list", + "angle": 0, + "index": 34, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "tuning improves adaptation to distribution shifts. In ICLR, 2023.5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 96, + 553, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 128 + ], + "type": "text", + "content": "[45] Brian Lester, Rami Al-Rfou, and Noah Constant. The power of scale for parameter-efficient prompt tuning. In EMNLP, 2021. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 130, + 553, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 130, + 553, + 163 + ], + "spans": [ + { + "bbox": [ + 316, + 130, + 553, + 163 + ], + "type": "text", + "content": "[46] Deng Li, Among Wu, Yaowei Wang, and Yahong Han. Prompt-driven dynamic object-centric learning for single domain generalization. In CVPR, 2024. 3, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 164, + 553, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 164, + 553, + 196 + ], + "spans": [ + { + "bbox": [ + 316, + 164, + 553, + 196 + ], + "type": "text", + "content": "[47] Dongze Lian, Daquan Zhou, Jiashi Feng, and Xinchao Wang. Scaling & shifting your features: A new baseline for efficient model tuning. In NeurIPS, 2022. 3, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 198, + 553, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 198, + 553, + 241 + ], + "spans": [ + { + "bbox": [ + 316, + 198, + 553, + 241 + ], + "type": "text", + "content": "[48] Shih-Yang Liu, Chien-Yi Wang, Hongxu Yin, Pavlo Molchanov, Yu-Chiang Frank Wang, Kwang-Ting Cheng, and Min-Hung Chen. DoRA: Weight-decomposed low-rank adaptation. In ICML, 2024. 3, 6, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 243, + 553, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 243, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 316, + 243, + 553, + 285 + ], + "type": "text", + "content": "[49] Yajing Liu, Shijun Zhou, Xiyao Liu, Chunhui Hao, Baojie Fan, and Jiandong Tian. Unbiased faster r-cnn for single-source domain generalized object detection. In CVPR, 2024. 3, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 288, + 553, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 288, + 553, + 320 + ], + "spans": [ + { + "bbox": [ + 316, + 288, + 553, + 320 + ], + "type": "text", + "content": "[50] Fanxu Meng, Zhaohui Wang, and Muhan Zhang. Pissa: Principal singular values and singular vectors adaptation of large language models. In NeurIPS, 2024. 3, 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 322, + 553, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 322, + 553, + 354 + ], + "spans": [ + { + "bbox": [ + 316, + 322, + 553, + 354 + ], + "type": "text", + "content": "[51] Gerhard Neuhold, Tobias Ollmann, Samuel Rota Bulo, and Peter Kontschieder. The mapillary vistas dataset for semantic understanding of street scenes. In ICCV, 2017. 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 355, + 553, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 355, + 553, + 409 + ], + "spans": [ + { + "bbox": [ + 316, + 355, + 553, + 409 + ], + "type": "text", + "content": "[52] Joshua Niemeijer, Manuel Schwonberg, Jan-Aike Termöhlen, Nico M Schmidt, and Tim Fingscheidt. Generalization by adaptation: Diffusion-based domain extension for domain-generalized semantic segmentation. In WACV, 2024. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 411, + 553, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 411, + 553, + 464 + ], + "spans": [ + { + "bbox": [ + 316, + 411, + 553, + 464 + ], + "type": "text", + "content": "[53] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. TMLR, 2023. 3, 4, 5, 6, 7" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 468, + 553, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 468, + 553, + 510 + ], + "spans": [ + { + "bbox": [ + 316, + 468, + 553, + 510 + ], + "type": "text", + "content": "[54] Byeonghyun Pak, Byeongju Woo, Sunghwan Kim, Dae-hwan Kim, and Hoseong Kim. Textual query-driven mask transformer for domain generalized segmentation. In ECCV, 2024. 3, 6, 7" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 512, + 553, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 512, + 553, + 544 + ], + "spans": [ + { + "bbox": [ + 316, + 512, + 553, + 544 + ], + "type": "text", + "content": "[55] Xingang Pan, Ping Luo, Jianping Shi, and Xiaou Tang. Two at once: Enhancing learning and generalization capacities via ibn-net. In ECCV, 2018. 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 545, + 553, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 545, + 553, + 588 + ], + "spans": [ + { + "bbox": [ + 316, + 545, + 553, + 588 + ], + "type": "text", + "content": "[56] Yuwen Pan, Rui Sun, Naisong Luo, Tianzhu Zhang, and Yongdong Zhang. Exploring reliable matching with phase enhancement for night-time semantic segmentation. In ECCV, 2024.3" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 590, + 553, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 590, + 553, + 624 + ], + "spans": [ + { + "bbox": [ + 316, + 590, + 553, + 624 + ], + "type": "text", + "content": "[57] Duo Peng, Yinjie Lei, Lingqiao Liu, Pingping Zhang, and Jun Liu. Global and local texture randomization for synthetic-to-real semantic segmentation. TIP, 2021. 2, 3" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 316, + 624, + 553, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 553, + 656 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 553, + 656 + ], + "type": "text", + "content": "[58] Duo Peng, Yinjie Lei, Munawar Hayat, Yulan Guo, and Wen Li. Semantic-aware domain generalized segmentation. In CVPR, 2022. 2, 3" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 316, + 658, + 553, + 711 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 553, + 711 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 553, + 711 + ], + "type": "text", + "content": "[59] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. SDXL: Improving latent diffusion models for high-resolution image synthesis. In ICLR, 2024. 3, 8" + } + ] + } + ], + "index": 33 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "25611" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "text", + "content": "[60] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 130, + 294, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 130, + 294, + 163 + ], + "spans": [ + { + "bbox": [ + 56, + 130, + 294, + 163 + ], + "type": "text", + "content": "[61] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. In NeurIPS, 2015. 1, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 164, + 294, + 198 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 164, + 294, + 198 + ], + "spans": [ + { + "bbox": [ + 56, + 164, + 294, + 198 + ], + "type": "text", + "content": "[62] Stephan R Richter, Vibhav Vineet, Stefan Roth, and Vladlen Koltun. Playing for data: Ground truth from computer games. In ECCV, 2016. 1, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 199, + 294, + 232 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 199, + 294, + 232 + ], + "spans": [ + { + "bbox": [ + 56, + 199, + 294, + 232 + ], + "type": "text", + "content": "[63] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 234, + 294, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 234, + 294, + 277 + ], + "spans": [ + { + "bbox": [ + 56, + 234, + 294, + 277 + ], + "type": "text", + "content": "[64] German Ros, Laura Sellart, Joanna Materzynska, David Vazquez, and Antonio M. Lopez. The synthia dataset: A large collection of synthetic images for semantic segmentation of urban scenes. In CVPR, 2016. 1, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 279, + 294, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 279, + 294, + 323 + ], + "spans": [ + { + "bbox": [ + 56, + 279, + 294, + 323 + ], + "type": "text", + "content": "[65] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In CVPR, 2023. 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 325, + 294, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 325, + 294, + 357 + ], + "spans": [ + { + "bbox": [ + 56, + 325, + 294, + 357 + ], + "type": "text", + "content": "[66] Christos Sakaridis, Dengxin Dai, and Luc Van Gool. Acdc: The adverse conditions dataset with correspondences for semantic driving scene understanding. In ICCV, 2021. 6, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 360, + 294, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 360, + 294, + 392 + ], + "spans": [ + { + "bbox": [ + 56, + 360, + 294, + 392 + ], + "type": "text", + "content": "[67] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zh-moginov, and Liang-Chieh Chen. Mobilenetv2: Inverted residuals and linear bottlenecks. In CVPR, 2018. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 394, + 294, + 427 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 394, + 294, + 427 + ], + "spans": [ + { + "bbox": [ + 56, + 394, + 294, + 427 + ], + "type": "text", + "content": "[68] Fengyi Shen, Akhil Gurram, Ziyuan Liu, He Wang, and Alois Knoll. Diga: Distil to generalize and then adapt for domain adaptive semantic segmentation. In CVPR, 2023. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 429, + 294, + 461 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 429, + 294, + 461 + ], + "spans": [ + { + "bbox": [ + 56, + 429, + 294, + 461 + ], + "type": "text", + "content": "[69] Zhiqiang Tang, Yunhe Gao, Yi Zhu, Zhi Zhang, Mu Li, and Dimitris N Metaxas. Crossnorm and selfnorm for generalization under distribution shifts. In ICCV, 2021. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 464, + 294, + 507 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 464, + 294, + 507 + ], + "spans": [ + { + "bbox": [ + 56, + 464, + 294, + 507 + ], + "type": "text", + "content": "[70] Sumanth Udupa, Prajwal Gurunath, Aniruddh Sikdar, and Suresh Sundaram. Mrfp: Learning generalizable semantic segmentation from sim-2-real with multi-resolution feature perturbation. In CVPR, 2024. 2, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 509, + 294, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 509, + 294, + 542 + ], + "spans": [ + { + "bbox": [ + 56, + 509, + 294, + 542 + ], + "type": "text", + "content": "[71] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 544, + 294, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 544, + 294, + 576 + ], + "spans": [ + { + "bbox": [ + 56, + 544, + 294, + 576 + ], + "type": "text", + "content": "[72] Vidit Vidit, Martin Engilberge, and Mathieu Salzmann. Clip the gap: A single domain generalization approach for object detection. In CVPR, 2023. 3, 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 578, + 294, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 578, + 294, + 621 + ], + "spans": [ + { + "bbox": [ + 56, + 578, + 294, + 621 + ], + "type": "text", + "content": "[73] Zhixiang Wei, Lin Chen, Tao Tu, Pengyang Ling, Huaian Chen, and Yi Jin. Disentangle then parse: Night-time semantic segmentation with illumination disentanglement. In ICCV, 2023. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 624, + 294, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 294, + 678 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 294, + 678 + ], + "type": "text", + "content": "[74] Zhixiang Wei, Lin Chen, Yi Jin, Xiaoxiao Ma, Tianle Liu, Pengyang Ling, Ben Wang, Huaian Chen, and Jinjin Zheng. Stronger fewer & superior: Harnessing vision foundation models for domain generalized semantic segmentation. In CVPR, 2024. 2, 3, 6, 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 56, + 681, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 681, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 681, + 294, + 713 + ], + "type": "text", + "content": "[75] Among Wu and Cheng Deng. Single-domain generalized object detection in urban scene via cyclic-disentangled self-distillation. In CVPR, 2022. 3, 6, 7" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 510 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 106 + ], + "type": "text", + "content": "[76] Zhenyao Wu, Xinyi Wu, Xiaoping Zhang, Lili Ju, and Song Wang. Siamdoge: Domain generalizable semantic segmentation using siamese network. In ECCV, 2022. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 108, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 108, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 108, + 553, + 150 + ], + "type": "text", + "content": "[77] Qi Xu, Liang Yao, Zhengkai Jiang, Guannan Jiang, Wenqing Chu, Wenhui Han, Wei Zhang, Chengjie Wang, and Ying Tai. Dirl: Domain-invariant representation learning for generalizable semantic segmentation. In AAAI, 2022. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 152, + 553, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 553, + 185 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 553, + 185 + ], + "type": "text", + "content": "[78] Liwei Yang, Xiang Gu, and Jian Sun. Generalized semantic segmentation by self-supervised source domain projection and multi-level contrastive learning. In AAAI, 2023. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 186, + 553, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 186, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 316, + 186, + 553, + 228 + ], + "type": "text", + "content": "[79] Jingjun Yi, Qi Bi, Hao Zheng, Haolan Zhan, Wei Ji, Yawen Huang, Yuexiang Li, and Yefeng Zheng. Learning spectral-decomposed tokens for domain generalized semantic segmentation. In ACMMM, 2024. 3, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 231, + 553, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 231, + 553, + 274 + ], + "spans": [ + { + "bbox": [ + 316, + 231, + 553, + 274 + ], + "type": "text", + "content": "[80] Fisher Yu, Haofeng Chen, Xin Wang, Wenqi Xian, Yingying Chen, Fangchen Liu, Vashisht Madhavan, and Trevor Darrell. Bdd100k: A diverse driving dataset for heterogeneous multitask learning. In CVPR, 2020. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 276, + 553, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 276, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 316, + 276, + 553, + 319 + ], + "type": "text", + "content": "[81] Yeonguk Yu, Sungho Shin, Seunghyeok Back, Mihwan Ko, Sangjun Noh, and Kyoobin Lee. Domain-specific block selection and paired-view pseudo-labeling for online test-time adaptation. In CVPR, 2024. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 321, + 553, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 321, + 553, + 374 + ], + "spans": [ + { + "bbox": [ + 316, + 321, + 553, + 374 + ], + "type": "text", + "content": "[82] Xiangyu Yue, Yang Zhang, Sicheng Zhao, Alberto Sangiovanni-Vincentelli, Kurt Keutzer, and Boqing Gong. Domain randomization and pyramid consistency: Simulation-to-real generalization without accessing target domain data. In ICCV, 2019. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 376, + 553, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 376, + 553, + 418 + ], + "spans": [ + { + "bbox": [ + 316, + 376, + 553, + 418 + ], + "type": "text", + "content": "[83] Yuyang Zhao, Zhun Zhong, Na Zhao, Nicu Sebe, and Gim Hee Lee. Style-hallucinated dual consistency learning for domain generalized semantic segmentation. In ECCV, 2022. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 421, + 553, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 421, + 553, + 453 + ], + "spans": [ + { + "bbox": [ + 316, + 421, + 553, + 453 + ], + "type": "text", + "content": "[84] Zhun Zhong, Yuyang Zhao, Gim Hee Lee, and Nicu Sebe. Adversarial style augmentation for domain generalized urban-scene segmentation. In NeurIPS, 2022. 3, 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 455, + 553, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 455, + 553, + 487 + ], + "spans": [ + { + "bbox": [ + 316, + 455, + 553, + 487 + ], + "type": "text", + "content": "[85] Jinghao Zhou, Chen Wei, Huiyu Wang, Wei Shen, Cihang Xie, Alan Yuille, and Tao Kong. ibot: Image bert pre-training with online tokenizer. arXiv:2111.07832, 2021. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 489, + 553, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 489, + 553, + 510 + ], + "spans": [ + { + "bbox": [ + 316, + 489, + 553, + 510 + ], + "type": "text", + "content": "[86] Zhuofan Zong, Guanglu Song, and Yu Liu. Detrs with collaborative hybrid assignments training. In ICCV, 2023. 7" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "25612" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/SocialGesture_ Delving into Multi-person Gesture Understanding/17e9a440-617f-4e5f-9712-f768185c9d5d_content_list.json b/2025/SocialGesture_ Delving into Multi-person Gesture Understanding/17e9a440-617f-4e5f-9712-f768185c9d5d_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..aca468be4e54ae8df9e16ffe67267af1a7fc07cf --- /dev/null +++ b/2025/SocialGesture_ Delving into Multi-person Gesture Understanding/17e9a440-617f-4e5f-9712-f768185c9d5d_content_list.json @@ -0,0 +1,1585 @@ +[ + { + "type": "text", + "text": "SocialGesture: Delving into Multi-person Gesture Understanding", + "text_level": 1, + "bbox": [ + 168, + 130, + 828, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xu Cao1, Pranav Virupaksha1, Wenqi Jia1, Bolin Lai2, Fiona Ryan2, Sangmin Lee3†, James M. Rehg1† \n1 University of Illinois Urbana-Champaign \n2 Georgia Institute of Technology \n3Sungkyunkwan University", + "bbox": [ + 99, + 178, + 903, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{xucao2,pranavv3,wenqij5,jrehg}@illinois.edu {bolin.lai,fkryan}@gatech.edu \nsangmin.lee@skku.edu", + "bbox": [ + 173, + 236, + 826, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 303, + 326, + 319 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Previous research in human gesture recognition has largely overlooked multi-person interactions, which are crucial for understanding the social context of naturally occurring gestures. This limitation in existing datasets presents a significant challenge in aligning human gestures with other modalities like language and speech. To address this issue, we introduce SocialGesture, the first largescale dataset specifically designed for multi-person gesture analysis. SocialGesture features a diverse range of natural scenarios and supports multiple gesture analysis tasks, including video-based recognition and temporal localization, providing a valuable resource for advancing the study of gesture during complex social interactions. Furthermore, we propose a novel visual question answering (VQA) task to benchmark vision language models' (VLMs) performance on social gesture understanding. Our findings highlight several limitations of current gesture recognition models, offering insights into future directions for improvement in this field. SocialGesture is available at hugging-face.co/datasets/IrohXu/SocialGesture.", + "bbox": [ + 88, + 335, + 485, + 638 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 667, + 220, + 684 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Long before the development of spoken language in human history, deictic gestures played an important role in human social communication, serving as one of the earliest communication tools for expressing thoughts, emotions, and intentions. In contemporary communication, both verbal (e.g., speech) and non-verbal (e.g., gesture) cues work in concert. While spoken words express direct meanings, understanding the complete social context often requires more than just the words alone. Gestures play an important role in clarifying intention or points of emphasis, and augment spoken communication with additional social context. Theoretical analysis suggests that both speech and gesture originate from a shared", + "bbox": [ + 89, + 694, + 485, + 876 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1856bd464f77839b2f1d7e4753513dcbf4386790b94ac6d29f511aa919be073c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 517, + 303, + 710, + 387 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/776fd41c8ed5847d22db3b0257d12edb541f39528a1ef4bfe0004ba7ba84b47a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 517, + 387, + 710, + 470 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/bfe90554e54ee68f1fd15a00c344b3ce865f29b8ff3898d5b05dc4bf7e9235fc.jpg", + "image_caption": [ + "Figure 1. Example frames from six gesture datasets. SocialGesture is the only dataset featuring multi-person interactions and focusing on natural gestures with meaningful social communication." + ], + "image_footnote": [], + "bbox": [ + 517, + 470, + 710, + 554 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/542625c58de26cd3244a6742f68e03e35583c1db8a10407e4d28f0c709743be2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 710, + 303, + 901, + 387 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a47a871c0a59d2acc6e6348c87e026edc36024c80b7889d804162b202bee50fd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 710, + 387, + 901, + 470 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e0f5acfb9d7626e05c51e7ac231d47011c3ff0b452e7edf8abc28b53d7c0f083.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 710, + 470, + 901, + 554 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "cognitive system, with gestures representing imagery-driven data and speech representing symbolic data [6]. Together, they reflect a unified communicative intent.", + "bbox": [ + 511, + 625, + 906, + 669 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent advancements in machine learning have driven research on the understanding of human social communication [27, 58], yet most of the latest social research predominantly focuses on linguistic modalities, often overlooking visual social cues like gestures [28]. While there have been machine learning works on human gestures, most of them have focused on gestures for device manipulation or sign languages [42, 44], and the problem of analyzing natural social gestures between multiple people has not received significant attention. A primary reason for this limitation is the lack of gesture data and annotations that capture multi-person interactions and the use of gestures in realistic social contexts. While recent datasets such as HaGRID [23] and LD-ConGR [31] address high-resolution and long-distance gesture collection (Figure 1), they remain confined to controlled,", + "bbox": [ + 511, + 674, + 910, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "†Corresponding author", + "bbox": [ + 109, + 886, + 235, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "19509", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "table", + "img_path": "images/f4ebf2da1f7834205ffdab43b98ac186697e84040c7fce3114c7e77fe673f6e0.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetMulti-personVideo ClipsInstancesLabel
CategoryBboxRelationVQA
Jester [37]X148,092148,092✓XXX
NVGesture [39]X1,5321,532✓✓XX
EgoGesture [60]X2,08124,161✓✓XX
ChaLearn ConGD [52]X22,53547,933✓✓XX
IPN Hand [5]X2004,218✓✓XX
UAV gesture [42]X119-✓✓XX
iMiGUE [34]X35918,499✓XXX
LD-ConGR [31]X54244,887✓✓XX
SocialGesture (Ours)✓9,88942,533✓✓✓✓
", + "bbox": [ + 94, + 88, + 903, + 277 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Table 1. Comparison of SocialGesture with existing video-based gesture datasets. SocialGesture is the first work that addresses multi-person social interaction cases. Our dataset uniquely provides comprehensive annotations including gesture categories, bounding boxes (humans and objects), gesture relations (subject person and target person), and visual question answering pairs.", + "bbox": [ + 89, + 287, + 906, + 330 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "single-person settings, and do not contain the spontaneous, natural gestures between multiple people that characterize real-life social interactions.", + "bbox": [ + 89, + 349, + 482, + 393 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address this limitation, we introduce SocialGesture, a comprehensive dataset designed specifically for understanding gestures in multi-person social interactions. SocialGesture provides three key advantages over existing datasets:", + "bbox": [ + 89, + 393, + 483, + 455 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. It focuses on videos with multiple people, making it the first multi-person gesture dataset, in contrast to previous datasets which lack human-to-human interaction.", + "2. SocialGesture includes four types of comprehensive annotations: (a) Social gesture categories, identifying the specific class of a gesture (i.e., pointing, showing, giving, reaching) within social interactions; (b) Temporal-spatial annotations, identifying the appearance of social gestures temporally and spatially through time stamps and bounding boxes. (c) Interaction annotations, capturing interpersonal social dynamics that result from gestures; and (d) Visual question answering (VQA) annotations, which link gestures to descriptive content, facilitating the alignment of gestures with contextual meaning. All of these annotations offer a rich resource for developing models for social gesture understanding.", + "3. SocialGesture comprises a large volume of diverse and naturally occurring social interactions sourced from YouTube and Ego4D [17], and is the largest multi-person gesture dataset available." + ], + "bbox": [ + 89, + 457, + 482, + 758 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Based on our novel SocialGesture dataset, we introduce diverse gesture-related benchmark tasks from temporal localization and recognition to visual question answering. Our experiments reveal significant challenges in multi-person gesture understanding for current architectures. It is the first step toward building agents that have capabilities to seamlessly understand non-verbal gesture cues in social contexts. The major contributions of this paper are as follows.", + "bbox": [ + 89, + 762, + 482, + 883 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- We introduce SocialGesture, the first dataset focusing on", + "bbox": [ + 89, + 885, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "multi-person gesture interactions in natural social settings. Unlike existing datasets that are limited to single-person or controlled environments, our dataset captures the complexity of real-world gesture-based social communication.", + "bbox": [ + 524, + 348, + 906, + 409 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We provide comprehensive multi-level annotations including gesture categories, temporal-spatial localization, interaction dynamics, and visual question-answering pairs. This rich annotation scheme enables the development of models for understanding social gestures.", + "- We perform comprehensive experiments across gesture localization and recognition, evaluating a wide breadth of architectures and observe multiple challenging areas for future works. We further benchmark multiple large vision-language models on our proposed diverse Social Visual Question-Answering tasks." + ], + "bbox": [ + 511, + 409, + 906, + 575 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 513, + 593, + 661, + 609 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Gesture Datasets", + "text_level": 1, + "bbox": [ + 513, + 619, + 679, + 633 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "As shown in Figure 1 and Table 1, the majority of gesture recognition datasets are video-based, with recent research emphasizing emotional state analysis, increased capture distance for improved model robustness, and egocentric gesture collection. Notable datasets include LD-ConGR, featuring 542 RGB-D videos representing 10 gestures [31], and HaGRID, which focuses on high-resolution image based gesture recognition across 18 hand gesture classes [23]. EgoGesture, with over 2 million frames from 50 subjects, is tailored for egocentric interactions with wearable devices [60], while iMiGUE [34] and SMG focus on micro-gesture recognition [10], containing 359 and 40 videos, respectively. Earlier datasets such as Jester [37], IPN Hand [5], ChaLearn IsoGD [52], and ChaLearn ConGD [52] have also made significant contributions, covering tasks like hand segmentation, and isolated or continuous gesture classification. In contrast, we are the first to provide a large-scale multi-person gesture", + "bbox": [ + 509, + 643, + 908, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "19510", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/e2c11aca6089c1a682ea958dc15dd43466e01cf881dce0be675f182f86ffab15.jpg", + "image_caption": [ + "Pointing Gesture" + ], + "image_footnote": [], + "bbox": [ + 96, + 89, + 295, + 261 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/6da6c39c5e789fca6dc8dd1f7e3f56cd60054a1dbfab86a6eb6db1fb20a21aa2.jpg", + "image_caption": [ + "Showing Gesture" + ], + "image_footnote": [], + "bbox": [ + 297, + 90, + 496, + 261 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/2dbd399b4634d0acb8e33efacf0866942d8fa13eb263a4cbd5a9ac716cfaa7c5.jpg", + "image_caption": [ + "Giving Gesture", + "Figure 2. Examples of the four deictic gesture categories in SocialGesture with subject-target relationships. From left to right: pointing (directing attention), showing (presenting objects), giving (transfer intention), and reaching (acquisition intention) gestures. Red boxes indicate gesture initiators (subjects) and blue notations indicate targets." + ], + "image_footnote": [], + "bbox": [ + 500, + 90, + 697, + 261 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/ad3a5e03d0df02531104d4520b14cb1928f772deca5bc9e5f100622379bb0212.jpg", + "image_caption": [ + "Reaching Gesture" + ], + "image_footnote": [], + "bbox": [ + 700, + 90, + 900, + 261 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "dataset, and we are also the first to provide comprehensive coverage of the four main classes of deictic gestures(defined in [38]), which play a crucial role in social communication.", + "bbox": [ + 89, + 354, + 482, + 402 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Social Understanding with Non-verbal Cues", + "text_level": 1, + "bbox": [ + 89, + 409, + 464, + 425 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Understanding social interactions has been long studied in the domain of natural language processing [15, 18, 25, 26, 46, 48]. However, the inherent multimodal nature of social interactions ensures that non-verbal elements (such as gestures, gaze and emotions) play an important role in communication [28], which are understudied in previous research [57]. Recently, several investigations have been conducted on these implicit non-verbal cues. Liu et al. [31] build a dataset to identify the gestures of single person from a long distance. Unfortunately, they only focus on isolated gestures of a single person, and fail to contextualize gesture recognition in a real-world social scenario. Non-verbal social understanding of multiple people demands more investigation [16, 21, 22, 41]. To this end, we make the first attempt to model multi-party social interactions by defining various social gesture understanding tasks, such as social gesture detection, recognition, and VQA, based on our dataset.", + "bbox": [ + 89, + 431, + 483, + 689 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Foundation Models for Video Understanding", + "text_level": 1, + "bbox": [ + 89, + 696, + 470, + 713 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Video foundation models are designed to tackle a wide range of understanding and analysis tasks, including classification, temporal action localization, question answering, captioning, and more. Early approaches to video understanding relied on CNN blocks as visual encoders for video frames, connected to a prediction head or decoder [9, 13]. With the emergence of transformers, new architectures like MViTv2-B [30], VideoSwin [36], UniFormerV2 [29], and VideoMAE-2 [54] tokenize their visual inputs, enabling unified integration with other modalities [56]. Building on this, recently released video-based VLMs like Qwen2-VL [55], LLaVA-NEXT [32], and InternVL-2.5 [12] leverage pretrained visual", + "bbox": [ + 89, + 719, + 485, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "encoders to extract rich representations. These models align visual tokens with textual tokens, enabling a deeper multimodal understanding that bridges the gap between video and language tasks. We provide the first investigation of the effectiveness of these models in understanding multi-party social gestures via a VQA task.", + "bbox": [ + 511, + 354, + 906, + 446 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. SocialGesture", + "text_level": 1, + "bbox": [ + 513, + 458, + 653, + 474 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Motivation", + "text_level": 1, + "bbox": [ + 511, + 483, + 635, + 498 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Social gestures comprise a fundamental set of human behavioral practices that facilitate interpersonal communication through hand and body movements in social contexts, distinct from sign language. According to [38], gestures can be classified into four primary categories: deictic, beat, iconic, and metaphoric. Deictic gestures further subdivide into pointing, showing, giving, and reaching. Beat gestures are rhythmic movements that emphasize certain words and phrases during speech. Iconic gestures visually represent concrete objects or actions (e.g., making a circle with fingers), while metaphoric gestures symbolize abstract concepts (e.g., making a circular motion to indicate repetition).", + "bbox": [ + 511, + 505, + 908, + 686 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Among these categories, deictic gestures deserve particular attention due to their prevalence and significance in social interactions. While iconic and metaphoric gestures appear in specific contexts and beat gestures primarily serve prosodic functions, deictic gestures play a crucial role in establishing joint attention and facilitating object-mediated social interaction. Therefore, our dataset focuses on four fundamental deictic gestures (see Figure 2):", + "bbox": [ + 511, + 686, + 908, + 808 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "- Pointing Gesture involves directing attention to a specific entity in the social scene through finger extension. While commonly executed with an extended index finger and arm, these specific forms are not mandatory requirements. The crucial identifying factor is the communicative intent to guide others' attention to a particular target, whether", + "bbox": [ + 511, + 810, + 908, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "19511", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/2342445800ffc3ab4a8e2b36a99204dd10ae4be9fd29cb41c46a0cd43dc0fdb2.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Video ContentSourceProportionTotal Length# of PeopleDescription
Group social gamesYouTube44.51%896 min>=3A group of people playing social games such as one night werewolf
Ego4D21.91%441 min>=4
Variety entertainmentYouTube22.31%449 min>=2People performing activities such as pranks, challenges, and interviews
Educational playYouTube2.53%51 min>=2Children engaged in educational activities with adult guidance
Product reviewsYouTube2.53%51 min>=2People recommending products such as advent calendars
Party & dinnerYouTube3.63%73 min>=7A group of people having party & dinner together
Group cookingYouTube2.58%52 min>=3People engaged in cooking activities together
", + "bbox": [ + 94, + 88, + 903, + 220 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Table 2. SocialGesture contains diverse videos including different multi-person social interactions from YouTube and Ego4D [17].", + "bbox": [ + 109, + 220, + 880, + 234 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "person or object. This distinguishes pointing from incidental hand movements or other gesture types, particularly reaching gestures where the primary intent differs.", + "bbox": [ + 102, + 248, + 482, + 292 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- **Showing Gesture** involves presenting an object for others to see it by manipulating or orienting it towards observers. Unlike pointing, the object itself is the focus of attention, and it does not need to be fully supported by the subject, such as when tilting or sliding an object to make its contents visible. The key criterion for a showing gesture is that the object is being presented for visual inspection, regardless of whether the target person actually looks at it or where the subject is directing their gaze.", + "- Giving Gesture involves manipulating an object with the intention of transferring it to another person's possession. It differs from pointing because the object originates with the subject, and from showing because the subject is inviting the other person to take possession of the object. The key is the subject's intention to transfer the object, regardless of whether the recipient actually takes it.", + "- Reaching Gesture manifests as a hand extension toward an object, expressing desire for possession or requesting transfer. It is characterized by full finger extension and often includes pre-grasp configuration and associated body movements like leaning forward. Reaching differs from showing and giving because the object is not in the subject's possession, and it differs from pointing because the subject is expressing a desire to possess the object, not just to draw attention to it. The key aspect of reaching is the intent to obtain the object, regardless of whether the subject can physically retrieve it." + ], + "bbox": [ + 89, + 294, + 482, + 700 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Data Acquisition", + "text_level": 1, + "bbox": [ + 89, + 712, + 256, + 728 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Data collection. We collect raw video data for SocialGesture from various sources, including YouTube channels and Ego4D. Table 2 represents the overall composition of the dataset. Video selection is based on four key criteria: (1) Video quality – we manually check each video to ensure high resolution; (2) Number of people – each video is carefully selected to include between two and ten people, ensuring clear visibility of gestures; (3) Video length – the total length of selected videos ranges from 2 to 30 minutes, with each gesture lasting between 1 to 15 seconds; and (4) Scene diversity – to enhance generalization, we select videos featuring", + "bbox": [ + 88, + 734, + 482, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "diverse races, genders, and ages.", + "bbox": [ + 511, + 250, + 728, + 263 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Pre-processing. For videos longer than 10 minutes, we cut them into video clips of 5 minutes in length. All videos are then converted to $720\\mathrm{p}$ $(1280\\times 720)$ with 30 FPS in the initial curation and are then mapped to $360\\mathrm{p}$ $(640\\times 360)$ with 5 FPS for further annotation.", + "bbox": [ + 511, + 268, + 905, + 344 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Gesture annotation. Each gesture in the SocialGesture dataset is defined as a sequence of video frames where an individual performs an arm or hand movement corresponding to one of four specified deictic gestures: pointing, showing, giving, and reaching. The annotations capture both temporal and spatial aspects of these social interactions through a comprehensive multi-level annotation framework. For temporal annotation, we identify the complete sequence of frames containing a gesture, from initial movement through completion. Within this sequence, we designate a key frame that best captures the defining moment of the gesture - typically when the gesture reaches its most distinctive configuration. This key frame serves as an anchor point for additional annotations. Spatial annotation includes precise bounding box coordinates [x1, y1, x2, y2] for both the gesture initiator and the target. The initiator bounding box encompasses the person performing the gesture, while the target bounding box identifies either the person or object toward which the gesture is directed. These spatial annotations enable analysis of the geometric relationships between interaction participants. Our annotation framework also includes natural language descriptions of each gesture's social context, capturing the broader interaction dynamics beyond pure spatiotemporal coordinates.", + "bbox": [ + 511, + 349, + 906, + 712 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Benchmark Tasks", + "text_level": 1, + "bbox": [ + 511, + 723, + 687, + 737 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3.1. Social Gesture Temporal Localization", + "text_level": 1, + "bbox": [ + 511, + 744, + 821, + 760 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Temporal action localization is a challenging task that requires not only identifying the temporal intervals of all detected gesture instances but also estimating corresponding confidence values. Unlike standard classification tasks, this process involves an end-to-end approach in which classification is integrated within temporal localization, adding considerable complexity. For example, even if a gesture is accurately localized in time, it is treated as a false detection if assigned an incorrect class label.", + "bbox": [ + 511, + 763, + 905, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "19512", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/f5cad51e6459531a5fb52b20692e9de7501cfd29dfaacca3ef5a759496404e2d.jpg", + "image_caption": [ + "Figure 3. Temporal gesture localization task for social gestures" + ], + "image_footnote": [], + "bbox": [ + 96, + 90, + 477, + 268 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Figure 3 illustrates a general approach to temporal action localization. Each video is initially divided into multiple overlapping video clips, denoted as $X = [X_0, X_1, \\ldots, X_K]$ , where $X_k \\in \\mathbb{R}^{T \\times H \\times W \\times C}$ . Feature extractors such as I3D or VideoMAE are then employed to extract multiscale feature representations, denoted as $Z_k \\in \\mathbb{R}^D$ , for each video clip. These features are subsequently concatenated into a single representation, $Z \\in \\mathbb{R}^{T \\times D}$ . The objective of the temporal action localization model is to generate sequence labels for each video clip.", + "bbox": [ + 89, + 299, + 483, + 449 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This task is also benchmarked by several well-known datasets, including THUMOS14 [20] and ActivityNet-1.3 [8]. However, our task presents additional challenges. In SocialGesture videos, multiple individuals are often present simultaneously, and gestures tend to be small, subtle portions of larger actions, making it uniquely challenging to detect and classify gestures accurately.", + "bbox": [ + 89, + 450, + 483, + 556 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3.2. Social Gesture Recognition", + "text_level": 1, + "bbox": [ + 89, + 561, + 326, + 577 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given the complexity of social gesture localization, we also divide long videos into shorter clips of 2-5 seconds and introduce the social gesture recognition task. This task can be approached in multiple ways. The first approach is isolated gesture recognition, where a long video is segmented into discrete gesture clips. Video action recognition models are then trained to classify the gestures in each clip independently. Alternatively, gestures can be identified continuously throughout the entire video, from beginning to end. Although most action localization methods we discussed in the previous section can typically handle this task, most existing video feature extraction pre-processing methods struggle with consistent multi-person action localization. Therefore, sliding windows remain a common strategy for this purpose: the video is processed in overlapping segments of a predefined window size, with each window analyzed by the model to calculate gesture confidence scores.", + "bbox": [ + 89, + 580, + 483, + 837 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We divide recognition into two subtasks (See Figure 4):", + "bbox": [ + 109, + 838, + 478, + 853 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "- Task 2-1: Gesture vs Non-gesture Classification This task aims to distinguish between social gesture and non-gesture video clips. The input is a short video clip $X \\in$", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/549ea7e37bcb7955888324f586c1e2287b642c527a165124155c6b20ee6333bd.jpg", + "image_caption": [ + "Figure 4. Gesture recognition tasks for social gestures" + ], + "image_footnote": [], + "bbox": [ + 519, + 90, + 898, + 266 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$\\mathbb{R}^{T\\times H\\times W\\times C}$ , where $T < 16$ , and the output is a label $\\hat{Y}$ indicating whether a gesture is present.", + "bbox": [ + 526, + 297, + 903, + 329 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "- Task 2-2: Gesture Type Classification. This task focuses on classifying video clips containing one of four social gestures: pointing, showing, giving, or reaching. The input and output structures are consistent with those used in the binary classification task. We further explore whether using cropped bounding boxes for each individual within a scene improves the accuracy of social gesture classification. The input is a short video clip cropped by the maximum region of the initiator of a social gesture and the output is a label $\\hat{Y}$ indicating four types of social gesture.", + "bbox": [ + 513, + 330, + 906, + 479 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For each task, we also strive to sample the training set to achieve a balanced label distribution. This adjustment is necessary because pointing gestures are significantly more frequent than showing, giving, and reaching in real-world communication contexts.", + "bbox": [ + 511, + 483, + 905, + 556 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3.3. Social Gesture Visual Question Answering", + "text_level": 1, + "bbox": [ + 511, + 564, + 854, + 579 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Prior to our work, there has been very limited research on designing video captioning and Visual Question Answering (VQA) specifically for human gesture datasets [49]. This gap may be due to the lack of multi-person interactions in existing video datasets, reducing the necessity for enhanced alignment between the gesture and language modalities. To fill this gap, we introduce the first VQA task focused on human gestures (Figure 5). Our VQA can be divided into three subtasks: (1) Global Perception, (2) Gesture Understanding, and (3) Gesture Localization.", + "bbox": [ + 511, + 583, + 906, + 733 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The Global Perception task serves as a foundational test to assess whether VLMs can comprehend basic events in short video clips, including scene descriptions and counting the number of humans present. We generate data for this task through a semi-automated QA generation pipeline based on GPT-4o [2], similar to the instruction-tuning data pair generation in LLaVA [33]. Initially, we manually create short descriptions for each video series, including the social background. We prompt these descriptions with the key frame of the video clip into the GPT-4o pipeline to generate QA pairs for human counting and scene description. Each QA", + "bbox": [ + 511, + 734, + 908, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "19513", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/81c2e9da614e48b861333c9ea0a4b4d3df052203e4995a0b596681717326c0e0.jpg", + "image_caption": [ + "Figure 5. The question-answer pairs of SocialGesture. We omit the options of each question in the figure. The bounding box defined by [top-left x, top-left y, bottom-right x, bottom-right y]). The definition will be provided together with the system prompts." + ], + "image_footnote": [], + "bbox": [ + 93, + 92, + 478, + 579 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "pair is then reviewed by a human annotator. Figure 5 shows the example of global perception (Task 3-1). Human counting is used as one VQA subtask in the VLM benchmarking experiments.", + "bbox": [ + 89, + 657, + 483, + 715 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For Gesture Understanding, we utilize gesture classification annotations to create questions related to social gesture classification (Figure 5, Task 3-2). This results in two types of questions: (1) detecting whether a social gesture occurs, and (2) classifying between the four types of social gestures.", + "bbox": [ + 89, + 718, + 483, + 794 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Gesture Localization leverages bounding boxes from our annotations to generate questions that test the model's spatial and temporal comprehension. These questions include providing the bounding box localization of the initiator, then determining whether the gesture's target is a human, and locating the target. We present an example of gesture localization (Task 3-3) in Figure 5. These questions are designed", + "bbox": [ + 89, + 795, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/984e3790fef70823c1417b93369fe86880b255b8da92f1aec440138db1d1a64e.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Feature0.30.40.50.60.7Avg mAP
I3D [9]24.8516.319.312.220.9610.73
R(2+1)D [51]14.3810.257.232.811.777.29
VideoMAEV2 [54]27.2325.0513.335.282.7614.73
", + "bbox": [ + 517, + 89, + 901, + 157 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/7541b89aa7c658453ab5720619660393d8ec0386d320eb34f98c75b0123143bb.jpg", + "table_caption": [ + "Table 3. Temporal action localization for four social gestures by different feature extractors with ActionFormer [59]." + ], + "table_footnote": [], + "table_body": "
Stride0.30.40.50.60.7Avg mAP
1611.098.605.522.671.825.94
824.8516.319.312.220.9610.73
431.6429.1319.3013.772.1119.19
", + "bbox": [ + 517, + 208, + 901, + 275 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 4. Explore the influence of stride for I3D and ActionFormer.", + "bbox": [ + 513, + 282, + 905, + 297 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "to challenge temporal-spatial reasoning in a multi-person setting, requiring analysis to produce the correct responses.", + "bbox": [ + 511, + 321, + 905, + 352 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 366, + 645, + 383 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 511, + 391, + 702, + 407 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We split the raw videos (long video) into training set (293 videos) and test set (79 videos), and then cut each long video into video clips. The 9,889 video clips used in video action recognition and social gesture VQA tasks come from the same train-test split in the raw videos. We also cut another 4,304 video clips from the raw videos as the non-gesture class. All experiments are conducted with 2 NVIDIA L40S GPUs and 2 H100 GPUs to make sure the batch size is 16 and learning rate is 5e-4 for all models. For all methods in Table 3, 5, 6, and 7, we apply the same data augmentation strategies used in [29, 50]. To solve the class imbalanced issue, we also adjust the size of training set for each task.", + "bbox": [ + 511, + 414, + 906, + 595 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Social Gesture Temporal Localization", + "text_level": 1, + "bbox": [ + 511, + 604, + 838, + 619 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines. We used two-stream I3D [9], R(2+1)D [51], and VideoMAEV2 [54] with the same sliding window of 16 frames and different stride sizes for feature extraction. Then, these video features are used as input for ActionFormer [59] during model training. We used mAP@[0.3:0.1:0.7] as the main evaluation metric and also reported the average mAP.", + "bbox": [ + 511, + 623, + 906, + 715 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Results and Findings. We observe that almost all feature extractors underperform on this task in Table 3. Notably, despite the strong performance of features from VideoMAE V2 [54] in other action localization datasets, such as THU-MOS 2014 [20] and ActivityNet [8], they only achieve an average mAP of 14.73 for social gesture recognition. To investigate the effect of the stride size in the sliding window on the average mAP, we conducted an ablation study in Table 4 using different stride sizes of the I3D feature in SocialGesture. We found that reducing the stride from 16 to 4 led to an increase in average mAP; however, the overall results remained insufficient. This shortfall can be attributed to the", + "bbox": [ + 511, + 719, + 908, + 898 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "19514", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/123b3a7b1fc9ea2c034a62e22b8cc501705532bea7eeb66e927951b600ea6a61.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelPretrainParamAcc (%)
TSN-R50 [53]Kinetics-400 [24]24M78.77
TANet-R50 [35]Kinetics-400 [24]26M71.22
SlowFast-R50 [13]Kinetics-400 [24]35M80.82
SlowFast-R101 [13]Kinetics-400 [24]63M79.59
TimeSformer-L [7]Kinetics-400 [24]121M78.71
MViTv2-B [30]Kinetics-400 [24]51M83.29
VideoSwin-B [36]Kinetics-400 [24]88M81.70
VideoSwin-L [36]Kinetics-400 [24]200M83.44
UniFormerV2-B/16 [29]CLIP [43]115M84.43
UniFormerV2-L/14 [29]CLIP [43]354M80.93
", + "bbox": [ + 94, + 88, + 478, + 239 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "fact that all feature extractors were pretrained on datasets that lack multi-person interactions, causing the features to be poorly aligned with the specific requirements of our task.", + "bbox": [ + 89, + 299, + 483, + 345 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Social Gesture Recognition", + "text_level": 1, + "bbox": [ + 89, + 357, + 334, + 373 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Baselines. Our baselines for social gesture recognition includes different CNN-based video models such as TSN [53], TANet [35], SlowFast [13], and Vision Transformer based models such as TimeSformer [7], MViTv2 [30], VideoSwin [36], UniFormerV2 [29]. All these models are pretrained with Kinetics-400 [24] dataset or using CLIP encoder [43]. We use Accuracy for gesture and non-gesture binary classification in Table 5 and Top 1 Accuracy for four social gesture classification in Table 6, and 7.", + "bbox": [ + 89, + 380, + 483, + 517 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results and Findings. The results in Table 5 indicate that most existing action recognition models can differentiate social gestures and non-gesture from short video clips. Despite fine-tuning these models using over 10,000 video clips for the binary classification task, their performance remains suboptimal. The SOTA baseline, UniFormerV2-B/16, achieves an accuracy of only $84.43\\%$ . Notably, Transformer-based models generally outperform CNN-based models.", + "bbox": [ + 89, + 521, + 483, + 642 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 6 presents results for the four-class social gesture classification task, revealing that all models struggle to effectively classify these gestures—despite the simplicity of the task for human annotators. The highest performance comes from VideoSwim-L, achieving only $56.18\\%$ . Given that individual gestures can be challenging to recognize in multi-person environments, we conducted further experiments, as shown in Table 7. Specifically, we extracted the region corresponding to each individual using ground truth bounding boxes and fine-tuned the same models with these per-person inputs. UniFormerV2-B/16 achieves the best performance of $64.72\\%$ top 1 accuracy. This may be because social gestures inherently involve fine-grained and subtle movements. Consequently, recognizing such gestures is difficult from both the extracted per-person region and the full-frame perspective. Addressing the challenge of social gesture understanding requires advanced models capable of", + "bbox": [ + 89, + 643, + 483, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/3a27bb6c0aa43094887e6362aeab25478319590207cbe38ef0853bb1292cf073.jpg", + "table_caption": [ + "Table 5. Experiments on binary classification for social gesture and non-gesture." + ], + "table_footnote": [], + "table_body": "
ModelPretrainParamTop1 Acc (%)
TSN-R50 [53]Kinetics-400 [24]24M54.83
TANet-R50 [35]Kinetics-400 [24]26M45.39
SlowFast-R50 [13]Kinetics-400 [24]35M45.17
SlowFast-R101 [13]Kinetics-400 [24]63M46.07
TimeSformer-L [7]Kinetics-400 [24]121M53.03
MViTv2-B [30]Kinetics-400 [24]51M37.98
VideoSwin-B [36]Kinetics-400 [24]88M53.93
VideoSwin-L [36]Kinetics-400 [24]200M56.18
UniFormerV2-B/16 [29]CLIP [43]115M55.51
UniFormerV2-L/14 [29]CLIP [43]354M50.34
", + "bbox": [ + 519, + 88, + 903, + 229 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/34875ebc9a23f8bbf991e106276bb659f293198bb30e1e80b9dc5049a63c2050.jpg", + "table_caption": [ + "Table 6. Experiments on classification for four social gestures." + ], + "table_footnote": [], + "table_body": "
ModelPretrainParamTop1 Acc (%)
TSN-R50 [53]Kinetics-400 [24]24M58.43
TANet-R50 [35]Kinetics-400 [24]26M46.29
SlowFast-R50 [13]Kinetics-400 [24]35M45.17
SlowFast-R101 [13]Kinetics-400 [24]63M42.92
TimeSformer-L [7]Kinetics-400 [24]121M63.60
MViTv2-B [30]Kinetics-400 [24]51M35.28
VideoSwin-B [36]Kinetics-400 [24]88M60.00
VideoSwin-L [36]Kinetics-400 [24]200M63.60
UniFormerV2-B/16 [29]CLIP [43]115M64.72
UniFormerV2-L/14 [29]CLIP [43]354M60.45
", + "bbox": [ + 517, + 267, + 903, + 407 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 7. Experiments on classification for four social gestures after extract each subject person's bounding box.", + "bbox": [ + 511, + 416, + 906, + 445 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "disentangling the relationship between the subject and the target person.", + "bbox": [ + 511, + 462, + 905, + 493 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. SocialGesture VQA", + "text_level": 1, + "bbox": [ + 511, + 505, + 700, + 522 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "From the results of previous tasks, we found that distinguishing social gestures is still very challenging for current SOTA methods in gesture localization and recognition. This difficulty highlights the unique nature of social gesture, where models need to have a unified view of multi-person social interactions and can distinguish details that are hard to detect. This is something traditional video analysis models still struggle with [1, 14]. Inspired by recent progress in multimodal video grounding and video-based VQA [11, 40], we aim to leverage the reasoning capabilities of the latest VLMs and evaluate their performance on this complex task.", + "bbox": [ + 511, + 527, + 906, + 696 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Baselines. To assess performance on the SocialGesture VQA benchmark, we chose several SOTA publicly available VLMs, including Qwen2-VL [55], Qwen2.5-VL [4], InternVL-2.5 [12], and LLAVA-NEXT-Video (LLaVA 1.6) [32]. Additionally, we included closed-source multimodal LLMs such as GPT-4o-mini [2] and Claude-3.7-sonnet [3] for a broader comparison. Note that GPT-4o [2] was excluded, as it was used to generate some of the VQA instruction-following training data, which could lead to biased results.", + "bbox": [ + 511, + 699, + 908, + 849 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Evaluation Metrics. Since all VQA pairs can be transformed into a multi-option output, we use rule-based accuracy metrics to evaluate the VQA tasks. Each model was", + "bbox": [ + 511, + 854, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "19515", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/580fb08cba3355866722f22fe08c9665689a80566835d60bb4c215c51dd9fd50.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelParamGlobal Perception (%)Gesture Understanding (%)Gesture Localization (%)
HumanCount@AccGestureDet@AccGestureClass@AccTargetLoc@AccTargetClass@Acc
Random Select--50.0025.0018.9050.00
InternVL-2.5 [12]2B22.3763.3474.4513.3339.79
InternVL-2.5 [12]8B35.5873.9481.8017.3857.28
LLaVA-NeXT-Video [32, 61]7B30.6371.9234.085.3959.90
Qwen2-VL [55]7B60.1874.1567.189.5859.19
Qwen2.5-VL [4]72B69.9775.6954.8229.3761.38
Claude 3.7 Sonnet [3]-63.2769.6461.7325.4864.20
GPT-4o-mini [2]-53.4073.6768.5221.2965.12
", + "bbox": [ + 94, + 88, + 903, + 251 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/4fccc51427ca347c01058017b1771fedf58567ff5375bb99978880aa6e04f155.jpg", + "table_caption": [ + "Table 8. Experiments on zero-shot performance of SOTA VLMs, including closed sourced multimodal LLMs for Social Gesture VQA" + ], + "table_footnote": [], + "table_body": "
ModelInputGlobal Perception (%)Gesture Understanding (%)Gesture Localization (%)
HumanCount@AccGestureDet@AccGestureClass@AccTargetLoc@AccTargetClass@Acc
Qwen2-VL-7Bvideo60.1874.1567.189.5859.19
Qwen2-VL-7B + LoRA [19] SFTvideo82.5533.0784.5249.2141.19
Qwen2-VL-7B + Full SFTvideo83.6433.0781.1351.2538.51
", + "bbox": [ + 94, + 279, + 901, + 367 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 9. Experiments on SFT performance of SOTA VLMs, including closed sourced multimodal LLMs for Social Gesture VQA", + "bbox": [ + 114, + 375, + 879, + 388 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "given sufficient context for the task, followed by a question and multiple answer choices. Accuracy was determined by comparing the model's response with the correct answer. We evaluated models across all subtasks including global perception, gesture understanding, and gesture localization in Section 3.3.3. The accuracy (Acc) in Table 8 is defined as: $\\mathrm{Acc} = \\frac{1}{N}\\sum_{k = 1}^{N}\\mathrm{MATCH}_k$ , where $N$ is the total number of VQA samples and $\\mathrm{MATCH}_k$ is a binary value. If the option in the model output matches the ground truth option, it is 1, otherwise it is 0.", + "bbox": [ + 89, + 405, + 483, + 555 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Results and Findings. In Table 8, the experimental results show that while most VLMs can handle global perception and gesture understanding tasks without finetuning, they all struggle significantly with gesture localization tasks. This trend is understandable, as localization tasks inherently involve more ambiguity compared to perception and understanding tasks. While there is not a very clear correlation between the number of parameters and performance, Qwen2.5-VL, the largest of our open sourced models, performed well overall. To explore the potential for improvement, we finetuned Qwen2-VL-7B in Table 9, which led to enhanced performance in most metrics, except those related to gesture binary classification and target identification. Performance drops in some metrics can be attributed to easy overfitting on binary cases. An additional noteworthy observation is that despite being pretrained on vast multimodal datasets, none of the SOTA VLMs achieved over $70\\%$ accuracy on the simplest task: counting humans in social interaction scenes.", + "bbox": [ + 89, + 566, + 483, + 839 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "These findings indicate that the SocialGesture tasks are significantly more challenging for VLMs than typical VQA benchmarks. This difficulty stems from the natural complexity of real-world social interaction videos used in our", + "bbox": [ + 89, + 839, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "benchmark. Unlike controlled environments, our dataset contains realistic scenarios where people frequently overlap with each other, are partially visible, or extend beyond the camera frame, creating natural occlusion cases. In addition, there exists high interdependence between tasks in SocialGesture VQA, which makes it challenging. For instance, when models process such videos, they should first assess the overall scene despite these visual challenges, counting people and understanding actions. This process involves high-level reasoning and the ability to sequentially infer relationships—skills that are currently lacking in SOTA VLMs. Developing models capable of such contextual social visual reasoning remains a key challenge for future research.", + "bbox": [ + 511, + 405, + 906, + 602 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 617, + 633, + 633 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We introduce SocialGesture, the first dataset and benchmark specifically designed for understanding and analyzing multiperson social gestures. We provide a comprehensive definition of tasks related to social gesture recognition, including a novel VQA dataset that encompasses global perception, gesture understanding, and spatial localization. To advance the field of human gesture and social interaction understanding, we benchmarked a variety of current state-of-the-art methods, including action recognition models and vision-language models (VLMs). Our findings reveal that multiperson social gesture understanding requires incremental visual reasoning—area where existing computer vision models fall short. This work offers insights into the complexities of multi-person social interactions and highlights the need for continued research into developing models capable of sophisticated visual reasoning. We hope that SocialGesture will motivate further innovations in gesture understanding.", + "bbox": [ + 511, + 643, + 908, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "19516", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgment", + "text_level": 1, + "bbox": [ + 91, + 90, + 240, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We are thankful to the Keymakr for their great work in data annotation. This work was supported in part by the NRF grant funded by the Korea government(MSIT) (No. RS-2025-00563942).", + "bbox": [ + 89, + 114, + 485, + 175 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 191, + 186, + 207 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Qaisar Abbas, Mostafa EA Ibrahim, and M Arfan Jaffar. Video scene analysis: an overview and challenges on deep learning algorithms. Multimedia Tools and Applications, 77 (16):20415-20453, 2018. 7", + "[2] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florecia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 5, 7, 8", + "[3] Anthropic. Claude 3 family. https://www.anthropic.com/news/claude-3-family, 2024. Accessed: 2024-05-27. 7, 8", + "[4] Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, et al. Qwen2. 5-vl technical report. arXiv preprint arXiv:2502.13923, 2025. 7, 8", + "[5] Gibran Benitez-Garcia, Jesus Olivares-Mercado, Gabriel Sanchez-Perez, and Keiji Yanai. Ipn hand: A video dataset and benchmark for real-time continuous hand gesture recognition. In 2020 25th international conference on pattern recognition (ICPR), pages 4340-4347. IEEE, 2021. 2", + "[6] Paolo Bernardis and Maurizio Gentilucci. Speech and gesture share the same communication system. Neuropsychologia, 44(2):178-190, 2006. 1", + "[7] Gedas Bertasius, Heng Wang, and Lorenzo Torresani. Is space-time attention all you need for video understanding? In ICML, page 4, 2021. 7", + "[8] Fabian Caba Heilbron, Victor Escorcia, Bernard Ghanem, and Juan Carlos Niebles. Activitynet: A large-scale video benchmark for human activity understanding. In Proceedings of the iee conference on computer vision and pattern recognition, pages 961-970, 2015. 5, 6", + "[9] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6299-6308, 2017. 3, 6", + "[10] Haoyu Chen, Henglin Shi, Xin Liu, Xiaobai Li, and Guoying Zhao. Smg: A micro-gesture dataset towards spontaneous body gestures for emotional stress state analysis. International Journal of Computer Vision, 131(6):1346-1366, 2023. 2", + "[11] Qirui Chen, Shangzhe Di, and Weidi Xie. Grounded multihop videoqa in long-form egocentric videos. arXiv preprint arXiv:2408.14469, 2024. 7", + "[12] Zhe Chen, Weiyun Wang, Yue Cao, Yangzhou Liu, Zhangwei Gao, Erfei Cui, Jinguo Zhu, Shenglong Ye, Hao Tian, Zhaoyang Liu, et al. Expanding performance boundaries of open-source multimodal models with model, data, and test" + ], + "bbox": [ + 93, + 215, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "time scaling. arXiv preprint arXiv:2412.05271, 2024. 3, 7, 8", + "[13] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6202-6211, 2019. 3, 7", + "[14] Chaoyou Fu, Yuhan Dai, Yongdong Luo, Lei Li, Shuhuai Ren, Renrui Zhang, Zihan Wang, Chenyu Zhou, Yunhang Shen, Mengdan Zhang, et al. Video-mme: The first-ever comprehensive evaluation benchmark of multi-modal llms in video analysis. arXiv preprint arXiv:2405.21075, 2024. 7", + "[15] Deepanway Ghosal, Navonil Majumder, Soujanya Poria, Niyati Chhaya, and Alexander Gelbukh. Dialoguegcn: A graph convolutional neural network for emotion recognition in conversation. arXiv preprint arXiv:1908.11540, 2019. 3", + "[16] Shreya Ghosh, Zhixi Cai, Abhinav Dhall, Dimitrios Kollias, Roland Goecke, and Tom Gedeon. Mrac track 1: 2nd workshop on multimodal, generative and responsible affective computing. arXiv preprint arXiv:2409.07256, 2024. 3", + "[17] Kristen Grauman, Andrew Westbury, Eugene Byrne, Zachary Chavis, Antonino Furnari, Rohit Girdhar, Jackson Hamburger, Hao Jiang, Miao Liu, Xingyu Liu, et al. Ego4d: Around the world in 3,000 hours of egocentric video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18995-19012, 2022. 2, 4", + "[18] Devamanyu Hazarika, Soujanya Poria, Amir Zadeh, Erik Cambria, Louis-Philippe Morency, and Roger Zimmermann. Conversational memory network for emotion recognition in dyadic dialogue videos. In Proceedings of the conference. Association for Computational Linguistics. North American Chapter. Meeting, page 2122. NIH Public Access, 2018. 3", + "[19] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. *ICLR*, 1(2):3, 2022. 8", + "[20] Haroon Idrees, Amir R Zamir, Yu-Gang Jiang, Alex Gorban, Ivan Laptev, Rahul Sukthankar, and Mubarak Shah. The thumos challenge on action recognition for videos \"in the wild\". Computer Vision and Image Understanding, 155:1-23, 2017. 5, 6", + "[21] Simindokht Jahangard, Zhixi Cai, Shiki Wen, and Hamid Rezatofighi. Jrdb-social: A multifaceted robotic dataset for understanding of context and dynamics of human interactions within social groups. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22087-22097, 2024. 3", + "[22] Wenqi Jia, Miao Liu, Hao Jiang, Ishwarya Ananthabhotla, James M Rehg, Vamsi Krishna Ithapu, and Ruohan Gao. The audio-visual conversational graph: From an egocentric-exocentric perspective. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26396–26405, 2024. 3", + "[23] Alexander Kapitanov, Karina Kvanchiani, Alexander Nagaev, Roman Kraynov, and Andrei Makhliarchuk. Hagrid-hand gesture recognition image dataset. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 4572-4581, 2024. 1, 2" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "19517", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[24] Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, et al. The kinetics human action video dataset. arXiv preprint arXiv:1705.06950, 2017. 7", + "[25] Bolin Lai, Hongxin Zhang, Miao Liu, Aryan Pariani, Fiona Ryan, Wenqi Jia, Shirley Anugrah Hayati, James Rehg, and Diyi Yang. Werewolf among us: Multimodal resources for modeling persuasion behaviors in social deduction games. Association for Computational Linguistics: ACL 2023, 2023. 3", + "[26] Bongseok Lee and Yong Suk Choi. Graph based network with contextualized representations of turns in dialogue. arXiv preprint arXiv:2109.04008, 2021. 3", + "[27] Sangmin Lee, Bolin Lai, Fiona Ryan, Bikram Boote, and James M Rehg. Modeling multimodal social interactions: new challenges and baselines with densely aligned representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14585-14595, 2024. 1", + "[28] Sangmin Lee, Minzhi Li, Bolin Lai, Wenqi Jia, Fiona Ryan, Xu Cao, Ozgur Kara, Bikram Boote, Weiyan Shi, Diyi Yang, et al. Towards social ai: A survey on understanding social interactions. arXiv preprint arXiv:2409.15316, 2024. 1, 3", + "[29] Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Limin Wang, and Yu Qiao. Uniformerv2: Unlocking the potential of image vits for video understanding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1632-1643, 2023. 3, 6, 7", + "[30] Yanghao Li, Chao-Yuan Wu, Haoqi Fan, Karttikeya Mangalam, Bo Xiong, Jitendra Malik, and Christoph Feichtenhofer. Mviv2: Improved multiscale vision transformers for classification and detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4804-4814, 2022. 3, 7", + "[31] Dan Liu, Libo Zhang, and Yanjun Wu. Ld-congr: A large rgb-d video dataset for long-distance continuous gesture recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3304-3312, 2022. 1, 2, 3", + "[32] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava-last: Improved reasoning,OCR,and world knowledge,2024.3,7,8", + "[33] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 5", + "[34] Xin Liu, Henglin Shi, Haoyu Chen, Zitong Yu, Xiaobai Li, and Guoying Zhao. imigue: An identity-free video dataset for micro-gesture understanding and emotion analysis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10631-10642, 2021. 2", + "[35] Zhaoyang Liu, Limin Wang, Wayne Wu, Chen Qian, and Tong Lu. Tam: Temporal adaptive module for video recognition. In Proceedings of the IEEE/CVF international conference on computer vision, pages 13708-13718, 2021. 7", + "[36] Ze Liu, Jia Ning, Yue Cao, Yixuan Wei, Zheng Zhang, Stephen Lin, and Han Hu. Video swim transformer. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3202-3211, 2022. 3, 7" + ], + "bbox": [ + 91, + 90, + 485, + 901 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[37] Joanna Materzynska, Guillaume Berger, Ingo Bax, and Roland Memisevic. The jester dataset: A large-scale video dataset of human gestures. In Proceedings of the IEEE/CVF international conference on computer vision workshops, pages 0–0, 2019. 2", + "[38] David McNeill. Hand and mind: What gestures reveal about thought. University of Chicago press, 1992. 3", + "[39] Pavlo Molchanov, Xiaodong Yang, Shalini Gupta, Kihwan Kim, Stephen Tyree, and Jan Kautz. Online detection and classification of dynamic hand gestures with recurrent 3d convolutional neural network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4207-4215, 2016. 2", + "[40] Fangzhou Mu, Sicheng Mo, and Yin Li. Snag: Scalable and accurate video grounding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18930-18940, 2024. 7", + "[41] Yujia Peng, Jiaheng Han, Zhenliang Zhang, Lifeng Fan, Tengyu Liu, Siyuan Qi, Xue Feng, Yuxi Ma, Yizhou Wang, and Song-Chun Zhu. The tong test: Evaluating artificial general intelligence through dynamic embodied physical and social interactions. Engineering, 34:12-22, 2024. 3", + "[42] Asanka G Perera, Yee Wei Law, and Javaan Chahl. Uav-gesture: A dataset for uav control and gesture recognition. In Proceedings of the European Conference on Computer Vision (ECCV) Workshops, pages 0–0, 2018. 1, 2", + "[43] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 7", + "[44] Razieh Rastgoo, Kourosh Kiani, and Sergio Escalera. Sign language recognition: A deep survey. Expert Systems with Applications, 164:113794, 2021. 1", + "[45] Fiona Ryan, Ajay Bati, Sangmin Lee, Daniel Bolya, Judy Hoffman, and James M Rehg. Gaze-lle: Gaze target estimation via large-scale learned encoders. 2025. 1", + "[46] Harvey Sacks, Emanuel A Schegloff, and Gail Jefferson. A simplest systematics for the organization of turn-taking for conversation. language, 50(4):696-735, 1974. 3", + "[47] Yuehao Song, Xinggang Wang, Jingfeng Yao, Wenyu Liu, Jinglin Zhang, and Xiangmin Xu. Vitgaze: gaze following with interaction features in vision transformers. Visual Intelligence, 2(1):1-15, 2024. 1", + "[48] Andreas Stolcke, Klaus Ries, Noah Coccaro, Elizabeth Shriberg, Rebecca Bates, Daniel Jurafsky, Paul Taylor, Rachel Martin, Carol Van Ess-Dykema, and Marie Meteor. Dialogue act modeling for automatic tagging and recognition of conversational speech. Computational linguistics, 26(3):339-373, 2000. 3", + "[49] Kosei Tanada, Shigemichi Matsuzaki, Kazuhito Tanaka, Shintaro Nakaoka, Yuki Kondo, and Yuto Mori. Pointing gesture understanding via visual prompting and visual question answering for interactive robot navigation. In First Workshop on Vision-Language Models for Navigation and Manipulation at ICRA 2024. 5" + ], + "bbox": [ + 516, + 90, + 906, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "19518", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[50] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In International conference on machine learning, pages 10347-10357. PMLR, 2021. 6", + "[51] Du Tran, Heng Wang, Lorenzo Torresani, Jamie Ray, Yann LeCun, and Manohar Paluri. A closer look at spatiotemporal convolutions for action recognition. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 6450-6459, 2018. 6", + "[52] Jun Wan, Yibing Zhao, Shuai Zhou, Isabelle Guyon, Sergio Escalera, and Stan Z Li. Chalearn looking at people rgb-d isolated and continuous datasets for gesture recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 56-64, 2016. 2", + "[53] Limin Wang, Yuanjun Xiong, Zhe Wang, Yu Qiao, Dahua Lin, Xiaou Tang, and Luc Van Gool. Temporal segment networks: Towards good practices for deep action recognition. In European conference on computer vision, pages 20-36. Springer, 2016. 7", + "[54] Limin Wang, Bingkun Huang, Zhiyu Zhao, Zhan Tong, Yinan He, Yi Wang, Yali Wang, and Yu Qiao. Videomae v2: Scaling video masked autoencoders with dual masking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14549-14560, 2023. 3, 6", + "[55] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 3, 7, 8", + "[56] Yi Wang, Kunchang Li, Xinhao Li, Jiashuo Yu, Yinan He, Guo Chen, Baoqi Pei, Rongkun Zheng, Jilan Xu, Zun Wang, et al. Internvideo2: Scaling video foundation models for multimodal video understanding. arXiv preprint arXiv:2403.15377, 2024. 3", + "[57] Diyi Yang, Dirk Hovy, David Jurgens, and Barbara Plank. The call for socially aware language technologies, 2024. 3", + "[58] Ceng Zhang, Junxin Chen, Jiatong Li, Yanhong Peng, and Zebing Mao. Large language models for human-robot interaction: A review. Biomimetic Intelligence and Robotics, page 100131, 2023. 1", + "[59] Chen-Lin Zhang, Jianxin Wu, and Yin Li. Actionformer: Localizing moments of actions with transformers. In European Conference on Computer Vision, pages 492-510. Springer, 2022. 6", + "[60] Yifan Zhang, Congqi Cao, Jian Cheng, and Hanqing Lu. Egogesture: A new dataset and benchmark for egocentric hand gesture recognition. IEEE Transactions on Multimedia, 20(5):1038-1050, 2018. 2", + "[61] Y Zhang, B Li, H Liu, Y Lee, L Gui, D Fu, J Feng, Z Liu, and C Li. Llava-last: A strong zero-shot video understanding model. 2024. 8" + ], + "bbox": [ + 91, + 92, + 486, + 825 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "19519", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/SocialGesture_ Delving into Multi-person Gesture Understanding/17e9a440-617f-4e5f-9712-f768185c9d5d_model.json b/2025/SocialGesture_ Delving into Multi-person Gesture Understanding/17e9a440-617f-4e5f-9712-f768185c9d5d_model.json new file mode 100644 index 0000000000000000000000000000000000000000..668e2a46eabcb9e8200f21796b380c0b27737bc3 --- /dev/null +++ b/2025/SocialGesture_ Delving into Multi-person Gesture Understanding/17e9a440-617f-4e5f-9712-f768185c9d5d_model.json @@ -0,0 +1,2334 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.169, + 0.131, + 0.83, + 0.154 + ], + "angle": 0, + "content": "SocialGesture: Delving into Multi-person Gesture Understanding" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.179, + 0.905, + 0.234 + ], + "angle": 0, + "content": "Xu Cao1, Pranav Virupaksha1, Wenqi Jia1, Bolin Lai2, Fiona Ryan2, Sangmin Lee3†, James M. Rehg1† \n1 University of Illinois Urbana-Champaign \n2 Georgia Institute of Technology \n3Sungkyunkwan University" + }, + { + "type": "text", + "bbox": [ + 0.174, + 0.237, + 0.828, + 0.269 + ], + "angle": 0, + "content": "{xucao2,pranavv3,wenqij5,jrehg}@illinois.edu {bolin.lai,fkryan}@gatech.edu \nsangmin.lee@skku.edu" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.304, + 0.327, + 0.32 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.337, + 0.486, + 0.64 + ], + "angle": 0, + "content": "Previous research in human gesture recognition has largely overlooked multi-person interactions, which are crucial for understanding the social context of naturally occurring gestures. This limitation in existing datasets presents a significant challenge in aligning human gestures with other modalities like language and speech. To address this issue, we introduce SocialGesture, the first largescale dataset specifically designed for multi-person gesture analysis. SocialGesture features a diverse range of natural scenarios and supports multiple gesture analysis tasks, including video-based recognition and temporal localization, providing a valuable resource for advancing the study of gesture during complex social interactions. Furthermore, we propose a novel visual question answering (VQA) task to benchmark vision language models' (VLMs) performance on social gesture understanding. Our findings highlight several limitations of current gesture recognition models, offering insights into future directions for improvement in this field. SocialGesture is available at hugging-face.co/datasets/IrohXu/SocialGesture." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.669, + 0.222, + 0.685 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.695, + 0.486, + 0.877 + ], + "angle": 0, + "content": "Long before the development of spoken language in human history, deictic gestures played an important role in human social communication, serving as one of the earliest communication tools for expressing thoughts, emotions, and intentions. In contemporary communication, both verbal (e.g., speech) and non-verbal (e.g., gesture) cues work in concert. While spoken words express direct meanings, understanding the complete social context often requires more than just the words alone. Gestures play an important role in clarifying intention or points of emphasis, and augment spoken communication with additional social context. Theoretical analysis suggests that both speech and gesture originate from a shared" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.304, + 0.712, + 0.388 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.388, + 0.712, + 0.471 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.471, + 0.712, + 0.555 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.712, + 0.304, + 0.903, + 0.388 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.712, + 0.388, + 0.903, + 0.471 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.712, + 0.471, + 0.903, + 0.555 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.559, + 0.907, + 0.601 + ], + "angle": 0, + "content": "Figure 1. Example frames from six gesture datasets. SocialGesture is the only dataset featuring multi-person interactions and focusing on natural gestures with meaningful social communication." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.625, + 0.907, + 0.67 + ], + "angle": 0, + "content": "cognitive system, with gestures representing imagery-driven data and speech representing symbolic data [6]. Together, they reflect a unified communicative intent." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.911, + 0.903 + ], + "angle": 0, + "content": "Recent advancements in machine learning have driven research on the understanding of human social communication [27, 58], yet most of the latest social research predominantly focuses on linguistic modalities, often overlooking visual social cues like gestures [28]. While there have been machine learning works on human gestures, most of them have focused on gestures for device manipulation or sign languages [42, 44], and the problem of analyzing natural social gestures between multiple people has not received significant attention. A primary reason for this limitation is the lack of gesture data and annotations that capture multi-person interactions and the use of gestures in realistic social contexts. While recent datasets such as HaGRID [23] and LD-ConGR [31] address high-resolution and long-distance gesture collection (Figure 1), they remain confined to controlled," + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.887, + 0.236, + 0.9 + ], + "angle": 0, + "content": "†Corresponding author" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "19509" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.089, + 0.904, + 0.278 + ], + "angle": 0, + "content": "
DatasetMulti-personVideo ClipsInstancesLabel
CategoryBboxRelationVQA
Jester [37]X148,092148,092✓XXX
NVGesture [39]X1,5321,532✓✓XX
EgoGesture [60]X2,08124,161✓✓XX
ChaLearn ConGD [52]X22,53547,933✓✓XX
IPN Hand [5]X2004,218✓✓XX
UAV gesture [42]X119-✓✓XX
iMiGUE [34]X35918,499✓XXX
LD-ConGR [31]X54244,887✓✓XX
SocialGesture (Ours)✓9,88942,533✓✓✓✓
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.288, + 0.907, + 0.331 + ], + "angle": 0, + "content": "Table 1. Comparison of SocialGesture with existing video-based gesture datasets. SocialGesture is the first work that addresses multi-person social interaction cases. Our dataset uniquely provides comprehensive annotations including gesture categories, bounding boxes (humans and objects), gesture relations (subject person and target person), and visual question answering pairs." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.35, + 0.483, + 0.394 + ], + "angle": 0, + "content": "single-person settings, and do not contain the spontaneous, natural gestures between multiple people that characterize real-life social interactions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.395, + 0.484, + 0.456 + ], + "angle": 0, + "content": "To address this limitation, we introduce SocialGesture, a comprehensive dataset designed specifically for understanding gestures in multi-person social interactions. SocialGesture provides three key advantages over existing datasets:" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.458, + 0.483, + 0.502 + ], + "angle": 0, + "content": "1. It focuses on videos with multiple people, making it the first multi-person gesture dataset, in contrast to previous datasets which lack human-to-human interaction." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.504, + 0.483, + 0.699 + ], + "angle": 0, + "content": "2. SocialGesture includes four types of comprehensive annotations: (a) Social gesture categories, identifying the specific class of a gesture (i.e., pointing, showing, giving, reaching) within social interactions; (b) Temporal-spatial annotations, identifying the appearance of social gestures temporally and spatially through time stamps and bounding boxes. (c) Interaction annotations, capturing interpersonal social dynamics that result from gestures; and (d) Visual question answering (VQA) annotations, which link gestures to descriptive content, facilitating the alignment of gestures with contextual meaning. All of these annotations offer a rich resource for developing models for social gesture understanding." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.7, + 0.483, + 0.76 + ], + "angle": 0, + "content": "3. SocialGesture comprises a large volume of diverse and naturally occurring social interactions sourced from YouTube and Ego4D [17], and is the largest multi-person gesture dataset available." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.458, + 0.483, + 0.76 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.763, + 0.483, + 0.884 + ], + "angle": 0, + "content": "Based on our novel SocialGesture dataset, we introduce diverse gesture-related benchmark tasks from temporal localization and recognition to visual question answering. Our experiments reveal significant challenges in multi-person gesture understanding for current architectures. It is the first step toward building agents that have capabilities to seamlessly understand non-verbal gesture cues in social contexts. The major contributions of this paper are as follows." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.886, + 0.483, + 0.902 + ], + "angle": 0, + "content": "- We introduce SocialGesture, the first dataset focusing on" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.349, + 0.907, + 0.41 + ], + "angle": 0, + "content": "multi-person gesture interactions in natural social settings. Unlike existing datasets that are limited to single-person or controlled environments, our dataset captures the complexity of real-world gesture-based social communication." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.41, + 0.907, + 0.485 + ], + "angle": 0, + "content": "- We provide comprehensive multi-level annotations including gesture categories, temporal-spatial localization, interaction dynamics, and visual question-answering pairs. This rich annotation scheme enables the development of models for understanding social gestures." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.485, + 0.906, + 0.576 + ], + "angle": 0, + "content": "- We perform comprehensive experiments across gesture localization and recognition, evaluating a wide breadth of architectures and observe multiple challenging areas for future works. We further benchmark multiple large vision-language models on our proposed diverse Social Visual Question-Answering tasks." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.41, + 0.907, + 0.576 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.594, + 0.663, + 0.61 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.621, + 0.681, + 0.635 + ], + "angle": 0, + "content": "2.1. Gesture Datasets" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.644, + 0.909, + 0.902 + ], + "angle": 0, + "content": "As shown in Figure 1 and Table 1, the majority of gesture recognition datasets are video-based, with recent research emphasizing emotional state analysis, increased capture distance for improved model robustness, and egocentric gesture collection. Notable datasets include LD-ConGR, featuring 542 RGB-D videos representing 10 gestures [31], and HaGRID, which focuses on high-resolution image based gesture recognition across 18 hand gesture classes [23]. EgoGesture, with over 2 million frames from 50 subjects, is tailored for egocentric interactions with wearable devices [60], while iMiGUE [34] and SMG focus on micro-gesture recognition [10], containing 359 and 40 videos, respectively. Earlier datasets such as Jester [37], IPN Hand [5], ChaLearn IsoGD [52], and ChaLearn ConGD [52] have also made significant contributions, covering tasks like hand segmentation, and isolated or continuous gesture classification. In contrast, we are the first to provide a large-scale multi-person gesture" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19510" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.098, + 0.09, + 0.297, + 0.262 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.145, + 0.264, + 0.25, + 0.277 + ], + "angle": 0, + "content": "Pointing Gesture" + }, + { + "type": "image", + "bbox": [ + 0.298, + 0.091, + 0.498, + 0.262 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.345, + 0.265, + 0.452, + 0.278 + ], + "angle": 0, + "content": "Showing Gesture" + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.091, + 0.699, + 0.262 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.554, + 0.266, + 0.645, + 0.278 + ], + "angle": 0, + "content": "Giving Gesture" + }, + { + "type": "image", + "bbox": [ + 0.702, + 0.091, + 0.901, + 0.262 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.747, + 0.266, + 0.856, + 0.278 + ], + "angle": 0, + "content": "Reaching Gesture" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.287, + 0.908, + 0.331 + ], + "angle": 0, + "content": "Figure 2. Examples of the four deictic gesture categories in SocialGesture with subject-target relationships. From left to right: pointing (directing attention), showing (presenting objects), giving (transfer intention), and reaching (acquisition intention) gestures. Red boxes indicate gesture initiators (subjects) and blue notations indicate targets." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.356, + 0.483, + 0.403 + ], + "angle": 0, + "content": "dataset, and we are also the first to provide comprehensive coverage of the four main classes of deictic gestures(defined in [38]), which play a crucial role in social communication." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.41, + 0.465, + 0.426 + ], + "angle": 0, + "content": "2.2. Social Understanding with Non-verbal Cues" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.432, + 0.485, + 0.69 + ], + "angle": 0, + "content": "Understanding social interactions has been long studied in the domain of natural language processing [15, 18, 25, 26, 46, 48]. However, the inherent multimodal nature of social interactions ensures that non-verbal elements (such as gestures, gaze and emotions) play an important role in communication [28], which are understudied in previous research [57]. Recently, several investigations have been conducted on these implicit non-verbal cues. Liu et al. [31] build a dataset to identify the gestures of single person from a long distance. Unfortunately, they only focus on isolated gestures of a single person, and fail to contextualize gesture recognition in a real-world social scenario. Non-verbal social understanding of multiple people demands more investigation [16, 21, 22, 41]. To this end, we make the first attempt to model multi-party social interactions by defining various social gesture understanding tasks, such as social gesture detection, recognition, and VQA, based on our dataset." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.698, + 0.472, + 0.714 + ], + "angle": 0, + "content": "2.3. Foundation Models for Video Understanding" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.486, + 0.902 + ], + "angle": 0, + "content": "Video foundation models are designed to tackle a wide range of understanding and analysis tasks, including classification, temporal action localization, question answering, captioning, and more. Early approaches to video understanding relied on CNN blocks as visual encoders for video frames, connected to a prediction head or decoder [9, 13]. With the emergence of transformers, new architectures like MViTv2-B [30], VideoSwin [36], UniFormerV2 [29], and VideoMAE-2 [54] tokenize their visual inputs, enabling unified integration with other modalities [56]. Building on this, recently released video-based VLMs like Qwen2-VL [55], LLaVA-NEXT [32], and InternVL-2.5 [12] leverage pretrained visual" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.356, + 0.908, + 0.448 + ], + "angle": 0, + "content": "encoders to extract rich representations. These models align visual tokens with textual tokens, enabling a deeper multimodal understanding that bridges the gap between video and language tasks. We provide the first investigation of the effectiveness of these models in understanding multi-party social gestures via a VQA task." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.459, + 0.655, + 0.475 + ], + "angle": 0, + "content": "3. SocialGesture" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.484, + 0.636, + 0.499 + ], + "angle": 0, + "content": "3.1. Motivation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.506, + 0.909, + 0.687 + ], + "angle": 0, + "content": "Social gestures comprise a fundamental set of human behavioral practices that facilitate interpersonal communication through hand and body movements in social contexts, distinct from sign language. According to [38], gestures can be classified into four primary categories: deictic, beat, iconic, and metaphoric. Deictic gestures further subdivide into pointing, showing, giving, and reaching. Beat gestures are rhythmic movements that emphasize certain words and phrases during speech. Iconic gestures visually represent concrete objects or actions (e.g., making a circle with fingers), while metaphoric gestures symbolize abstract concepts (e.g., making a circular motion to indicate repetition)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.688, + 0.909, + 0.809 + ], + "angle": 0, + "content": "Among these categories, deictic gestures deserve particular attention due to their prevalence and significance in social interactions. While iconic and metaphoric gestures appear in specific contexts and beat gestures primarily serve prosodic functions, deictic gestures play a crucial role in establishing joint attention and facilitating object-mediated social interaction. Therefore, our dataset focuses on four fundamental deictic gestures (see Figure 2):" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.811, + 0.909, + 0.902 + ], + "angle": 0, + "content": "- Pointing Gesture involves directing attention to a specific entity in the social scene through finger extension. While commonly executed with an extended index finger and arm, these specific forms are not mandatory requirements. The crucial identifying factor is the communicative intent to guide others' attention to a particular target, whether" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "19511" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.089, + 0.904, + 0.221 + ], + "angle": 0, + "content": "
Video ContentSourceProportionTotal Length# of PeopleDescription
Group social gamesYouTube44.51%896 min>=3A group of people playing social games such as one night werewolf
Ego4D21.91%441 min>=4
Variety entertainmentYouTube22.31%449 min>=2People performing activities such as pranks, challenges, and interviews
Educational playYouTube2.53%51 min>=2Children engaged in educational activities with adult guidance
Product reviewsYouTube2.53%51 min>=2People recommending products such as advent calendars
Party & dinnerYouTube3.63%73 min>=7A group of people having party & dinner together
Group cookingYouTube2.58%52 min>=3People engaged in cooking activities together
" + }, + { + "type": "table_caption", + "bbox": [ + 0.111, + 0.222, + 0.882, + 0.235 + ], + "angle": 0, + "content": "Table 2. SocialGesture contains diverse videos including different multi-person social interactions from YouTube and Ego4D [17]." + }, + { + "type": "text", + "bbox": [ + 0.104, + 0.249, + 0.483, + 0.294 + ], + "angle": 0, + "content": "person or object. This distinguishes pointing from incidental hand movements or other gesture types, particularly reaching gestures where the primary intent differs." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.295, + 0.483, + 0.429 + ], + "angle": 0, + "content": "- **Showing Gesture** involves presenting an object for others to see it by manipulating or orienting it towards observers. Unlike pointing, the object itself is the focus of attention, and it does not need to be fully supported by the subject, such as when tilting or sliding an object to make its contents visible. The key criterion for a showing gesture is that the object is being presented for visual inspection, regardless of whether the target person actually looks at it or where the subject is directing their gaze." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.43, + 0.483, + 0.534 + ], + "angle": 0, + "content": "- Giving Gesture involves manipulating an object with the intention of transferring it to another person's possession. It differs from pointing because the object originates with the subject, and from showing because the subject is inviting the other person to take possession of the object. The key is the subject's intention to transfer the object, regardless of whether the recipient actually takes it." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.536, + 0.483, + 0.702 + ], + "angle": 0, + "content": "- Reaching Gesture manifests as a hand extension toward an object, expressing desire for possession or requesting transfer. It is characterized by full finger extension and often includes pre-grasp configuration and associated body movements like leaning forward. Reaching differs from showing and giving because the object is not in the subject's possession, and it differs from pointing because the subject is expressing a desire to possess the object, not just to draw attention to it. The key aspect of reaching is the intent to obtain the object, regardless of whether the subject can physically retrieve it." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.295, + 0.483, + 0.702 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.713, + 0.257, + 0.729 + ], + "angle": 0, + "content": "3.2. Data Acquisition" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.735, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Data collection. We collect raw video data for SocialGesture from various sources, including YouTube channels and Ego4D. Table 2 represents the overall composition of the dataset. Video selection is based on four key criteria: (1) Video quality – we manually check each video to ensure high resolution; (2) Number of people – each video is carefully selected to include between two and ten people, ensuring clear visibility of gestures; (3) Video length – the total length of selected videos ranges from 2 to 30 minutes, with each gesture lasting between 1 to 15 seconds; and (4) Scene diversity – to enhance generalization, we select videos featuring" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.25, + 0.729, + 0.264 + ], + "angle": 0, + "content": "diverse races, genders, and ages." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.269, + 0.906, + 0.345 + ], + "angle": 0, + "content": "Pre-processing. For videos longer than 10 minutes, we cut them into video clips of 5 minutes in length. All videos are then converted to \\(720\\mathrm{p}\\) \\((1280\\times 720)\\) with 30 FPS in the initial curation and are then mapped to \\(360\\mathrm{p}\\) \\((640\\times 360)\\) with 5 FPS for further annotation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.351, + 0.908, + 0.713 + ], + "angle": 0, + "content": "Gesture annotation. Each gesture in the SocialGesture dataset is defined as a sequence of video frames where an individual performs an arm or hand movement corresponding to one of four specified deictic gestures: pointing, showing, giving, and reaching. The annotations capture both temporal and spatial aspects of these social interactions through a comprehensive multi-level annotation framework. For temporal annotation, we identify the complete sequence of frames containing a gesture, from initial movement through completion. Within this sequence, we designate a key frame that best captures the defining moment of the gesture - typically when the gesture reaches its most distinctive configuration. This key frame serves as an anchor point for additional annotations. Spatial annotation includes precise bounding box coordinates [x1, y1, x2, y2] for both the gesture initiator and the target. The initiator bounding box encompasses the person performing the gesture, while the target bounding box identifies either the person or object toward which the gesture is directed. These spatial annotations enable analysis of the geometric relationships between interaction participants. Our annotation framework also includes natural language descriptions of each gesture's social context, capturing the broader interaction dynamics beyond pure spatiotemporal coordinates." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.724, + 0.689, + 0.738 + ], + "angle": 0, + "content": "3.3. Benchmark Tasks" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.746, + 0.822, + 0.761 + ], + "angle": 0, + "content": "3.3.1. Social Gesture Temporal Localization" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Temporal action localization is a challenging task that requires not only identifying the temporal intervals of all detected gesture instances but also estimating corresponding confidence values. Unlike standard classification tasks, this process involves an end-to-end approach in which classification is integrated within temporal localization, adding considerable complexity. For example, even if a gesture is accurately localized in time, it is treated as a false detection if assigned an incorrect class label." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19512" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.098, + 0.092, + 0.478, + 0.27 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.1, + 0.276, + 0.475, + 0.29 + ], + "angle": 0, + "content": "Figure 3. Temporal gesture localization task for social gestures" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.3, + 0.484, + 0.45 + ], + "angle": 0, + "content": "Figure 3 illustrates a general approach to temporal action localization. Each video is initially divided into multiple overlapping video clips, denoted as \\( X = [X_0, X_1, \\ldots, X_K] \\), where \\( X_k \\in \\mathbb{R}^{T \\times H \\times W \\times C} \\). Feature extractors such as I3D or VideoMAE are then employed to extract multiscale feature representations, denoted as \\( Z_k \\in \\mathbb{R}^D \\), for each video clip. These features are subsequently concatenated into a single representation, \\( Z \\in \\mathbb{R}^{T \\times D} \\). The objective of the temporal action localization model is to generate sequence labels for each video clip." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.452, + 0.484, + 0.557 + ], + "angle": 0, + "content": "This task is also benchmarked by several well-known datasets, including THUMOS14 [20] and ActivityNet-1.3 [8]. However, our task presents additional challenges. In SocialGesture videos, multiple individuals are often present simultaneously, and gestures tend to be small, subtle portions of larger actions, making it uniquely challenging to detect and classify gestures accurately." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.563, + 0.327, + 0.578 + ], + "angle": 0, + "content": "3.3.2. Social Gesture Recognition" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.582, + 0.484, + 0.838 + ], + "angle": 0, + "content": "Given the complexity of social gesture localization, we also divide long videos into shorter clips of 2-5 seconds and introduce the social gesture recognition task. This task can be approached in multiple ways. The first approach is isolated gesture recognition, where a long video is segmented into discrete gesture clips. Video action recognition models are then trained to classify the gestures in each clip independently. Alternatively, gestures can be identified continuously throughout the entire video, from beginning to end. Although most action localization methods we discussed in the previous section can typically handle this task, most existing video feature extraction pre-processing methods struggle with consistent multi-person action localization. Therefore, sliding windows remain a common strategy for this purpose: the video is processed in overlapping segments of a predefined window size, with each window analyzed by the model to calculate gesture confidence scores." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.839, + 0.48, + 0.854 + ], + "angle": 0, + "content": "We divide recognition into two subtasks (See Figure 4):" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "- Task 2-1: Gesture vs Non-gesture Classification This task aims to distinguish between social gesture and non-gesture video clips. The input is a short video clip \\( X \\in \\)" + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.092, + 0.9, + 0.267 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.548, + 0.272, + 0.871, + 0.286 + ], + "angle": 0, + "content": "Figure 4. Gesture recognition tasks for social gestures" + }, + { + "type": "text", + "bbox": [ + 0.527, + 0.298, + 0.905, + 0.33 + ], + "angle": 0, + "content": "\\(\\mathbb{R}^{T\\times H\\times W\\times C}\\), where \\(T < 16\\), and the output is a label \\(\\hat{Y}\\) indicating whether a gesture is present." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.331, + 0.907, + 0.481 + ], + "angle": 0, + "content": "- Task 2-2: Gesture Type Classification. This task focuses on classifying video clips containing one of four social gestures: pointing, showing, giving, or reaching. The input and output structures are consistent with those used in the binary classification task. We further explore whether using cropped bounding boxes for each individual within a scene improves the accuracy of social gesture classification. The input is a short video clip cropped by the maximum region of the initiator of a social gesture and the output is a label \\(\\hat{Y}\\) indicating four types of social gesture." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.484, + 0.906, + 0.558 + ], + "angle": 0, + "content": "For each task, we also strive to sample the training set to achieve a balanced label distribution. This adjustment is necessary because pointing gestures are significantly more frequent than showing, giving, and reaching in real-world communication contexts." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.565, + 0.855, + 0.58 + ], + "angle": 0, + "content": "3.3.3. Social Gesture Visual Question Answering" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.584, + 0.907, + 0.734 + ], + "angle": 0, + "content": "Prior to our work, there has been very limited research on designing video captioning and Visual Question Answering (VQA) specifically for human gesture datasets [49]. This gap may be due to the lack of multi-person interactions in existing video datasets, reducing the necessity for enhanced alignment between the gesture and language modalities. To fill this gap, we introduce the first VQA task focused on human gestures (Figure 5). Our VQA can be divided into three subtasks: (1) Global Perception, (2) Gesture Understanding, and (3) Gesture Localization." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.909, + 0.902 + ], + "angle": 0, + "content": "The Global Perception task serves as a foundational test to assess whether VLMs can comprehend basic events in short video clips, including scene descriptions and counting the number of humans present. We generate data for this task through a semi-automated QA generation pipeline based on GPT-4o [2], similar to the instruction-tuning data pair generation in LLaVA [33]. Initially, we manually create short descriptions for each video series, including the social background. We prompt these descriptions with the key frame of the video clip into the GPT-4o pipeline to generate QA pairs for human counting and scene description. Each QA" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19513" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.093, + 0.48, + 0.58 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.584, + 0.483, + 0.642 + ], + "angle": 0, + "content": "Figure 5. The question-answer pairs of SocialGesture. We omit the options of each question in the figure. The bounding box defined by [top-left x, top-left y, bottom-right x, bottom-right y]). The definition will be provided together with the system prompts." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.658, + 0.484, + 0.717 + ], + "angle": 0, + "content": "pair is then reviewed by a human annotator. Figure 5 shows the example of global perception (Task 3-1). Human counting is used as one VQA subtask in the VLM benchmarking experiments." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.719, + 0.484, + 0.795 + ], + "angle": 0, + "content": "For Gesture Understanding, we utilize gesture classification annotations to create questions related to social gesture classification (Figure 5, Task 3-2). This results in two types of questions: (1) detecting whether a social gesture occurs, and (2) classifying between the four types of social gestures." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Gesture Localization leverages bounding boxes from our annotations to generate questions that test the model's spatial and temporal comprehension. These questions include providing the bounding box localization of the initiator, then determining whether the gesture's target is a human, and locating the target. We present an example of gesture localization (Task 3-3) in Figure 5. These questions are designed" + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.09, + 0.903, + 0.159 + ], + "angle": 0, + "content": "
Feature0.30.40.50.60.7Avg mAP
I3D [9]24.8516.319.312.220.9610.73
R(2+1)D [51]14.3810.257.232.811.777.29
VideoMAEV2 [54]27.2325.0513.335.282.7614.73
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.167, + 0.907, + 0.195 + ], + "angle": 0, + "content": "Table 3. Temporal action localization for four social gestures by different feature extractors with ActionFormer [59]." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.209, + 0.903, + 0.276 + ], + "angle": 0, + "content": "
Stride0.30.40.50.60.7Avg mAP
1611.098.605.522.671.825.94
824.8516.319.312.220.9610.73
431.6429.1319.3013.772.1119.19
" + }, + { + "type": "table_caption", + "bbox": [ + 0.514, + 0.284, + 0.906, + 0.299 + ], + "angle": 0, + "content": "Table 4. Explore the influence of stride for I3D and ActionFormer." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.323, + 0.906, + 0.353 + ], + "angle": 0, + "content": "to challenge temporal-spatial reasoning in a multi-person setting, requiring analysis to produce the correct responses." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.367, + 0.646, + 0.384 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.392, + 0.703, + 0.409 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.415, + 0.907, + 0.596 + ], + "angle": 0, + "content": "We split the raw videos (long video) into training set (293 videos) and test set (79 videos), and then cut each long video into video clips. The 9,889 video clips used in video action recognition and social gesture VQA tasks come from the same train-test split in the raw videos. We also cut another 4,304 video clips from the raw videos as the non-gesture class. All experiments are conducted with 2 NVIDIA L40S GPUs and 2 H100 GPUs to make sure the batch size is 16 and learning rate is 5e-4 for all models. For all methods in Table 3, 5, 6, and 7, we apply the same data augmentation strategies used in [29, 50]. To solve the class imbalanced issue, we also adjust the size of training set for each task." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.606, + 0.839, + 0.621 + ], + "angle": 0, + "content": "4.2. Social Gesture Temporal Localization" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.624, + 0.907, + 0.716 + ], + "angle": 0, + "content": "Baselines. We used two-stream I3D [9], R(2+1)D [51], and VideoMAEV2 [54] with the same sliding window of 16 frames and different stride sizes for feature extraction. Then, these video features are used as input for ActionFormer [59] during model training. We used mAP@[0.3:0.1:0.7] as the main evaluation metric and also reported the average mAP." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.909, + 0.9 + ], + "angle": 0, + "content": "Results and Findings. We observe that almost all feature extractors underperform on this task in Table 3. Notably, despite the strong performance of features from VideoMAE V2 [54] in other action localization datasets, such as THU-MOS 2014 [20] and ActivityNet [8], they only achieve an average mAP of 14.73 for social gesture recognition. To investigate the effect of the stride size in the sliding window on the average mAP, we conducted an ablation study in Table 4 using different stride sizes of the I3D feature in SocialGesture. We found that reducing the stride from 16 to 4 led to an increase in average mAP; however, the overall results remained insufficient. This shortfall can be attributed to the" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19514" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.089, + 0.48, + 0.24 + ], + "angle": 0, + "content": "
ModelPretrainParamAcc (%)
TSN-R50 [53]Kinetics-400 [24]24M78.77
TANet-R50 [35]Kinetics-400 [24]26M71.22
SlowFast-R50 [13]Kinetics-400 [24]35M80.82
SlowFast-R101 [13]Kinetics-400 [24]63M79.59
TimeSformer-L [7]Kinetics-400 [24]121M78.71
MViTv2-B [30]Kinetics-400 [24]51M83.29
VideoSwin-B [36]Kinetics-400 [24]88M81.70
VideoSwin-L [36]Kinetics-400 [24]200M83.44
UniFormerV2-B/16 [29]CLIP [43]115M84.43
UniFormerV2-L/14 [29]CLIP [43]354M80.93
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.249, + 0.483, + 0.277 + ], + "angle": 0, + "content": "Table 5. Experiments on binary classification for social gesture and non-gesture." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.3, + 0.484, + 0.346 + ], + "angle": 0, + "content": "fact that all feature extractors were pretrained on datasets that lack multi-person interactions, causing the features to be poorly aligned with the specific requirements of our task." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.358, + 0.336, + 0.374 + ], + "angle": 0, + "content": "4.3. Social Gesture Recognition" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.381, + 0.484, + 0.518 + ], + "angle": 0, + "content": "Baselines. Our baselines for social gesture recognition includes different CNN-based video models such as TSN [53], TANet [35], SlowFast [13], and Vision Transformer based models such as TimeSformer [7], MViTv2 [30], VideoSwin [36], UniFormerV2 [29]. All these models are pretrained with Kinetics-400 [24] dataset or using CLIP encoder [43]. We use Accuracy for gesture and non-gesture binary classification in Table 5 and Top 1 Accuracy for four social gesture classification in Table 6, and 7." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.522, + 0.484, + 0.643 + ], + "angle": 0, + "content": "Results and Findings. The results in Table 5 indicate that most existing action recognition models can differentiate social gestures and non-gesture from short video clips. Despite fine-tuning these models using over 10,000 video clips for the binary classification task, their performance remains suboptimal. The SOTA baseline, UniFormerV2-B/16, achieves an accuracy of only \\(84.43\\%\\). Notably, Transformer-based models generally outperform CNN-based models." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.644, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Table 6 presents results for the four-class social gesture classification task, revealing that all models struggle to effectively classify these gestures—despite the simplicity of the task for human annotators. The highest performance comes from VideoSwim-L, achieving only \\(56.18\\%\\). Given that individual gestures can be challenging to recognize in multi-person environments, we conducted further experiments, as shown in Table 7. Specifically, we extracted the region corresponding to each individual using ground truth bounding boxes and fine-tuned the same models with these per-person inputs. UniFormerV2-B/16 achieves the best performance of \\(64.72\\%\\) top 1 accuracy. This may be because social gestures inherently involve fine-grained and subtle movements. Consequently, recognizing such gestures is difficult from both the extracted per-person region and the full-frame perspective. Addressing the challenge of social gesture understanding requires advanced models capable of" + }, + { + "type": "table", + "bbox": [ + 0.52, + 0.089, + 0.905, + 0.23 + ], + "angle": 0, + "content": "
ModelPretrainParamTop1 Acc (%)
TSN-R50 [53]Kinetics-400 [24]24M54.83
TANet-R50 [35]Kinetics-400 [24]26M45.39
SlowFast-R50 [13]Kinetics-400 [24]35M45.17
SlowFast-R101 [13]Kinetics-400 [24]63M46.07
TimeSformer-L [7]Kinetics-400 [24]121M53.03
MViTv2-B [30]Kinetics-400 [24]51M37.98
VideoSwin-B [36]Kinetics-400 [24]88M53.93
VideoSwin-L [36]Kinetics-400 [24]200M56.18
UniFormerV2-B/16 [29]CLIP [43]115M55.51
UniFormerV2-L/14 [29]CLIP [43]354M50.34
" + }, + { + "type": "table_caption", + "bbox": [ + 0.523, + 0.238, + 0.895, + 0.252 + ], + "angle": 0, + "content": "Table 6. Experiments on classification for four social gestures." + }, + { + "type": "table", + "bbox": [ + 0.519, + 0.268, + 0.905, + 0.409 + ], + "angle": 0, + "content": "
ModelPretrainParamTop1 Acc (%)
TSN-R50 [53]Kinetics-400 [24]24M58.43
TANet-R50 [35]Kinetics-400 [24]26M46.29
SlowFast-R50 [13]Kinetics-400 [24]35M45.17
SlowFast-R101 [13]Kinetics-400 [24]63M42.92
TimeSformer-L [7]Kinetics-400 [24]121M63.60
MViTv2-B [30]Kinetics-400 [24]51M35.28
VideoSwin-B [36]Kinetics-400 [24]88M60.00
VideoSwin-L [36]Kinetics-400 [24]200M63.60
UniFormerV2-B/16 [29]CLIP [43]115M64.72
UniFormerV2-L/14 [29]CLIP [43]354M60.45
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.417, + 0.907, + 0.446 + ], + "angle": 0, + "content": "Table 7. Experiments on classification for four social gestures after extract each subject person's bounding box." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.463, + 0.906, + 0.494 + ], + "angle": 0, + "content": "disentangling the relationship between the subject and the target person." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.506, + 0.702, + 0.523 + ], + "angle": 0, + "content": "4.4. SocialGesture VQA" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.529, + 0.908, + 0.697 + ], + "angle": 0, + "content": "From the results of previous tasks, we found that distinguishing social gestures is still very challenging for current SOTA methods in gesture localization and recognition. This difficulty highlights the unique nature of social gesture, where models need to have a unified view of multi-person social interactions and can distinguish details that are hard to detect. This is something traditional video analysis models still struggle with [1, 14]. Inspired by recent progress in multimodal video grounding and video-based VQA [11, 40], we aim to leverage the reasoning capabilities of the latest VLMs and evaluate their performance on this complex task." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.7, + 0.909, + 0.851 + ], + "angle": 0, + "content": "Baselines. To assess performance on the SocialGesture VQA benchmark, we chose several SOTA publicly available VLMs, including Qwen2-VL [55], Qwen2.5-VL [4], InternVL-2.5 [12], and LLAVA-NEXT-Video (LLaVA 1.6) [32]. Additionally, we included closed-source multimodal LLMs such as GPT-4o-mini [2] and Claude-3.7-sonnet [3] for a broader comparison. Note that GPT-4o [2] was excluded, as it was used to generate some of the VQA instruction-following training data, which could lead to biased results." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.91, + 0.903 + ], + "angle": 0, + "content": "Evaluation Metrics. Since all VQA pairs can be transformed into a multi-option output, we use rule-based accuracy metrics to evaluate the VQA tasks. Each model was" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19515" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.089, + 0.904, + 0.252 + ], + "angle": 0, + "content": "
ModelParamGlobal Perception (%)Gesture Understanding (%)Gesture Localization (%)
HumanCount@AccGestureDet@AccGestureClass@AccTargetLoc@AccTargetClass@Acc
Random Select--50.0025.0018.9050.00
InternVL-2.5 [12]2B22.3763.3474.4513.3339.79
InternVL-2.5 [12]8B35.5873.9481.8017.3857.28
LLaVA-NeXT-Video [32, 61]7B30.6371.9234.085.3959.90
Qwen2-VL [55]7B60.1874.1567.189.5859.19
Qwen2.5-VL [4]72B69.9775.6954.8229.3761.38
Claude 3.7 Sonnet [3]-63.2769.6461.7325.4864.20
GPT-4o-mini [2]-53.4073.6768.5221.2965.12
" + }, + { + "type": "table_caption", + "bbox": [ + 0.101, + 0.26, + 0.895, + 0.274 + ], + "angle": 0, + "content": "Table 8. Experiments on zero-shot performance of SOTA VLMs, including closed sourced multimodal LLMs for Social Gesture VQA" + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.28, + 0.903, + 0.368 + ], + "angle": 0, + "content": "
ModelInputGlobal Perception (%)Gesture Understanding (%)Gesture Localization (%)
HumanCount@AccGestureDet@AccGestureClass@AccTargetLoc@AccTargetClass@Acc
Qwen2-VL-7Bvideo60.1874.1567.189.5859.19
Qwen2-VL-7B + LoRA [19] SFTvideo82.5533.0784.5249.2141.19
Qwen2-VL-7B + Full SFTvideo83.6433.0781.1351.2538.51
" + }, + { + "type": "table_caption", + "bbox": [ + 0.116, + 0.375, + 0.88, + 0.39 + ], + "angle": 0, + "content": "Table 9. Experiments on SFT performance of SOTA VLMs, including closed sourced multimodal LLMs for Social Gesture VQA" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.406, + 0.484, + 0.556 + ], + "angle": 0, + "content": "given sufficient context for the task, followed by a question and multiple answer choices. Accuracy was determined by comparing the model's response with the correct answer. We evaluated models across all subtasks including global perception, gesture understanding, and gesture localization in Section 3.3.3. The accuracy (Acc) in Table 8 is defined as: \\(\\mathrm{Acc} = \\frac{1}{N}\\sum_{k = 1}^{N}\\mathrm{MATCH}_k\\), where \\(N\\) is the total number of VQA samples and \\(\\mathrm{MATCH}_k\\) is a binary value. If the option in the model output matches the ground truth option, it is 1, otherwise it is 0." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.568, + 0.484, + 0.84 + ], + "angle": 0, + "content": "Results and Findings. In Table 8, the experimental results show that while most VLMs can handle global perception and gesture understanding tasks without finetuning, they all struggle significantly with gesture localization tasks. This trend is understandable, as localization tasks inherently involve more ambiguity compared to perception and understanding tasks. While there is not a very clear correlation between the number of parameters and performance, Qwen2.5-VL, the largest of our open sourced models, performed well overall. To explore the potential for improvement, we finetuned Qwen2-VL-7B in Table 9, which led to enhanced performance in most metrics, except those related to gesture binary classification and target identification. Performance drops in some metrics can be attributed to easy overfitting on binary cases. An additional noteworthy observation is that despite being pretrained on vast multimodal datasets, none of the SOTA VLMs achieved over \\(70\\%\\) accuracy on the simplest task: counting humans in social interaction scenes." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.901 + ], + "angle": 0, + "content": "These findings indicate that the SocialGesture tasks are significantly more challenging for VLMs than typical VQA benchmarks. This difficulty stems from the natural complexity of real-world social interaction videos used in our" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.406, + 0.908, + 0.603 + ], + "angle": 0, + "content": "benchmark. Unlike controlled environments, our dataset contains realistic scenarios where people frequently overlap with each other, are partially visible, or extend beyond the camera frame, creating natural occlusion cases. In addition, there exists high interdependence between tasks in SocialGesture VQA, which makes it challenging. For instance, when models process such videos, they should first assess the overall scene despite these visual challenges, counting people and understanding actions. This process involves high-level reasoning and the ability to sequentially infer relationships—skills that are currently lacking in SOTA VLMs. Developing models capable of such contextual social visual reasoning remains a key challenge for future research." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.618, + 0.634, + 0.634 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.644, + 0.909, + 0.901 + ], + "angle": 0, + "content": "We introduce SocialGesture, the first dataset and benchmark specifically designed for understanding and analyzing multiperson social gestures. We provide a comprehensive definition of tasks related to social gesture recognition, including a novel VQA dataset that encompasses global perception, gesture understanding, and spatial localization. To advance the field of human gesture and social interaction understanding, we benchmarked a variety of current state-of-the-art methods, including action recognition models and vision-language models (VLMs). Our findings reveal that multiperson social gesture understanding requires incremental visual reasoning—area where existing computer vision models fall short. This work offers insights into the complexities of multi-person social interactions and highlights the need for continued research into developing models capable of sophisticated visual reasoning. We hope that SocialGesture will motivate further innovations in gesture understanding." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19516" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.091, + 0.241, + 0.108 + ], + "angle": 0, + "content": "Acknowledgment" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.116, + 0.486, + 0.176 + ], + "angle": 0, + "content": "We are thankful to the Keymakr for their great work in data annotation. This work was supported in part by the NRF grant funded by the Korea government(MSIT) (No. RS-2025-00563942)." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.192, + 0.187, + 0.208 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.217, + 0.484, + 0.271 + ], + "angle": 0, + "content": "[1] Qaisar Abbas, Mostafa EA Ibrahim, and M Arfan Jaffar. Video scene analysis: an overview and challenges on deep learning algorithms. Multimedia Tools and Applications, 77 (16):20415-20453, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.274, + 0.484, + 0.342 + ], + "angle": 0, + "content": "[2] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florecia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.346, + 0.485, + 0.386 + ], + "angle": 0, + "content": "[3] Anthropic. Claude 3 family. https://www.anthropic.com/news/claude-3-family, 2024. Accessed: 2024-05-27. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.388, + 0.483, + 0.442 + ], + "angle": 0, + "content": "[4] Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, et al. Qwen2. 5-vl technical report. arXiv preprint arXiv:2502.13923, 2025. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.446, + 0.483, + 0.514 + ], + "angle": 0, + "content": "[5] Gibran Benitez-Garcia, Jesus Olivares-Mercado, Gabriel Sanchez-Perez, and Keiji Yanai. Ipn hand: A video dataset and benchmark for real-time continuous hand gesture recognition. In 2020 25th international conference on pattern recognition (ICPR), pages 4340-4347. IEEE, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.517, + 0.483, + 0.557 + ], + "angle": 0, + "content": "[6] Paolo Bernardis and Maurizio Gentilucci. Speech and gesture share the same communication system. Neuropsychologia, 44(2):178-190, 2006. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.56, + 0.483, + 0.601 + ], + "angle": 0, + "content": "[7] Gedas Bertasius, Heng Wang, and Lorenzo Torresani. Is space-time attention all you need for video understanding? In ICML, page 4, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.603, + 0.483, + 0.671 + ], + "angle": 0, + "content": "[8] Fabian Caba Heilbron, Victor Escorcia, Bernard Ghanem, and Juan Carlos Niebles. Activitynet: A large-scale video benchmark for human activity understanding. In Proceedings of the iee conference on computer vision and pattern recognition, pages 961-970, 2015. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.675, + 0.483, + 0.729 + ], + "angle": 0, + "content": "[9] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6299-6308, 2017. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.732, + 0.483, + 0.799 + ], + "angle": 0, + "content": "[10] Haoyu Chen, Henglin Shi, Xin Liu, Xiaobai Li, and Guoying Zhao. Smg: A micro-gesture dataset towards spontaneous body gestures for emotional stress state analysis. International Journal of Computer Vision, 131(6):1346-1366, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.802, + 0.483, + 0.842 + ], + "angle": 0, + "content": "[11] Qirui Chen, Shangzhe Di, and Weidi Xie. Grounded multihop videoqa in long-form egocentric videos. arXiv preprint arXiv:2408.14469, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.483, + 0.9 + ], + "angle": 0, + "content": "[12] Zhe Chen, Weiyun Wang, Yue Cao, Yangzhou Liu, Zhangwei Gao, Erfei Cui, Jinguo Zhu, Shenglong Ye, Hao Tian, Zhaoyang Liu, et al. Expanding performance boundaries of open-source multimodal models with model, data, and test" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.217, + 0.485, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.907, + 0.119 + ], + "angle": 0, + "content": "time scaling. arXiv preprint arXiv:2412.05271, 2024. 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.906, + 0.178 + ], + "angle": 0, + "content": "[13] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6202-6211, 2019. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.18, + 0.906, + 0.248 + ], + "angle": 0, + "content": "[14] Chaoyou Fu, Yuhan Dai, Yongdong Luo, Lei Li, Shuhuai Ren, Renrui Zhang, Zihan Wang, Chenyu Zhou, Yunhang Shen, Mengdan Zhang, et al. Video-mme: The first-ever comprehensive evaluation benchmark of multi-modal llms in video analysis. arXiv preprint arXiv:2405.21075, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.251, + 0.907, + 0.305 + ], + "angle": 0, + "content": "[15] Deepanway Ghosal, Navonil Majumder, Soujanya Poria, Niyati Chhaya, and Alexander Gelbukh. Dialoguegcn: A graph convolutional neural network for emotion recognition in conversation. arXiv preprint arXiv:1908.11540, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.308, + 0.907, + 0.362 + ], + "angle": 0, + "content": "[16] Shreya Ghosh, Zhixi Cai, Abhinav Dhall, Dimitrios Kollias, Roland Goecke, and Tom Gedeon. Mrac track 1: 2nd workshop on multimodal, generative and responsible affective computing. arXiv preprint arXiv:2409.07256, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.365, + 0.907, + 0.447 + ], + "angle": 0, + "content": "[17] Kristen Grauman, Andrew Westbury, Eugene Byrne, Zachary Chavis, Antonino Furnari, Rohit Girdhar, Jackson Hamburger, Hao Jiang, Miao Liu, Xingyu Liu, et al. Ego4d: Around the world in 3,000 hours of egocentric video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18995-19012, 2022. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.45, + 0.906, + 0.532 + ], + "angle": 0, + "content": "[18] Devamanyu Hazarika, Soujanya Poria, Amir Zadeh, Erik Cambria, Louis-Philippe Morency, and Roger Zimmermann. Conversational memory network for emotion recognition in dyadic dialogue videos. In Proceedings of the conference. Association for Computational Linguistics. North American Chapter. Meeting, page 2122. NIH Public Access, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.534, + 0.907, + 0.587 + ], + "angle": 0, + "content": "[19] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. *ICLR*, 1(2):3, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.591, + 0.907, + 0.659 + ], + "angle": 0, + "content": "[20] Haroon Idrees, Amir R Zamir, Yu-Gang Jiang, Alex Gorban, Ivan Laptev, Rahul Sukthankar, and Mubarak Shah. The thumos challenge on action recognition for videos \"in the wild\". Computer Vision and Image Understanding, 155:1-23, 2017. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.662, + 0.906, + 0.744 + ], + "angle": 0, + "content": "[21] Simindokht Jahangard, Zhixi Cai, Shiki Wen, and Hamid Rezatofighi. Jrdb-social: A multifaceted robotic dataset for understanding of context and dynamics of human interactions within social groups. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22087-22097, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.747, + 0.907, + 0.83 + ], + "angle": 0, + "content": "[22] Wenqi Jia, Miao Liu, Hao Jiang, Ishwarya Ananthabhotla, James M Rehg, Vamsi Krishna Ithapu, and Ruohan Gao. The audio-visual conversational graph: From an egocentric-exocentric perspective. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26396–26405, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.832, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[23] Alexander Kapitanov, Karina Kvanchiani, Alexander Nagaev, Roman Kraynov, and Andrei Makhliarchuk. Hagrid-hand gesture recognition image dataset. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 4572-4581, 2024. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19517" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.161 + ], + "angle": 0, + "content": "[24] Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, et al. The kinetics human action video dataset. arXiv preprint arXiv:1705.06950, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.162, + 0.485, + 0.244 + ], + "angle": 0, + "content": "[25] Bolin Lai, Hongxin Zhang, Miao Liu, Aryan Pariani, Fiona Ryan, Wenqi Jia, Shirley Anugrah Hayati, James Rehg, and Diyi Yang. Werewolf among us: Multimodal resources for modeling persuasion behaviors in social deduction games. Association for Computational Linguistics: ACL 2023, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.246, + 0.484, + 0.287 + ], + "angle": 0, + "content": "[26] Bongseok Lee and Yong Suk Choi. Graph based network with contextualized representations of turns in dialogue. arXiv preprint arXiv:2109.04008, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.288, + 0.484, + 0.357 + ], + "angle": 0, + "content": "[27] Sangmin Lee, Bolin Lai, Fiona Ryan, Bikram Boote, and James M Rehg. Modeling multimodal social interactions: new challenges and baselines with densely aligned representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14585-14595, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.358, + 0.484, + 0.413 + ], + "angle": 0, + "content": "[28] Sangmin Lee, Minzhi Li, Bolin Lai, Wenqi Jia, Fiona Ryan, Xu Cao, Ozgur Kara, Bikram Boote, Weiyan Shi, Diyi Yang, et al. Towards social ai: A survey on understanding social interactions. arXiv preprint arXiv:2409.15316, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.414, + 0.484, + 0.483 + ], + "angle": 0, + "content": "[29] Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Limin Wang, and Yu Qiao. Uniformerv2: Unlocking the potential of image vits for video understanding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1632-1643, 2023. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.484, + 0.484, + 0.566 + ], + "angle": 0, + "content": "[30] Yanghao Li, Chao-Yuan Wu, Haoqi Fan, Karttikeya Mangalam, Bo Xiong, Jitendra Malik, and Christoph Feichtenhofer. Mviv2: Improved multiscale vision transformers for classification and detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4804-4814, 2022. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.567, + 0.484, + 0.635 + ], + "angle": 0, + "content": "[31] Dan Liu, Libo Zhang, and Yanjun Wu. Ld-congr: A large rgb-d video dataset for long-distance continuous gesture recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3304-3312, 2022. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.636, + 0.484, + 0.678 + ], + "angle": 0, + "content": "[32] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava-last: Improved reasoning,OCR,and world knowledge,2024.3,7,8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.679, + 0.484, + 0.719 + ], + "angle": 0, + "content": "[33] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.72, + 0.484, + 0.789 + ], + "angle": 0, + "content": "[34] Xin Liu, Henglin Shi, Haoyu Chen, Zitong Yu, Xiaobai Li, and Guoying Zhao. imigue: An identity-free video dataset for micro-gesture understanding and emotion analysis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10631-10642, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.79, + 0.484, + 0.845 + ], + "angle": 0, + "content": "[35] Zhaoyang Liu, Limin Wang, Wayne Wu, Chen Qian, and Tong Lu. Tam: Temporal adaptive module for video recognition. In Proceedings of the IEEE/CVF international conference on computer vision, pages 13708-13718, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.484, + 0.902 + ], + "angle": 0, + "content": "[36] Ze Liu, Jia Ning, Yue Cao, Yixuan Wei, Zheng Zhang, Stephen Lin, and Han Hu. Video swim transformer. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3202-3211, 2022. 3, 7" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.907, + 0.162 + ], + "angle": 0, + "content": "[37] Joanna Materzynska, Guillaume Berger, Ingo Bax, and Roland Memisevic. The jester dataset: A large-scale video dataset of human gestures. In Proceedings of the IEEE/CVF international conference on computer vision workshops, pages 0–0, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.163, + 0.906, + 0.192 + ], + "angle": 0, + "content": "[38] David McNeill. Hand and mind: What gestures reveal about thought. University of Chicago press, 1992. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.193, + 0.907, + 0.275 + ], + "angle": 0, + "content": "[39] Pavlo Molchanov, Xiaodong Yang, Shalini Gupta, Kihwan Kim, Stephen Tyree, and Jan Kautz. Online detection and classification of dynamic hand gestures with recurrent 3d convolutional neural network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4207-4215, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.277, + 0.907, + 0.333 + ], + "angle": 0, + "content": "[40] Fangzhou Mu, Sicheng Mo, and Yin Li. Snag: Scalable and accurate video grounding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18930-18940, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.334, + 0.907, + 0.404 + ], + "angle": 0, + "content": "[41] Yujia Peng, Jiaheng Han, Zhenliang Zhang, Lifeng Fan, Tengyu Liu, Siyuan Qi, Xue Feng, Yuxi Ma, Yizhou Wang, and Song-Chun Zhu. The tong test: Evaluating artificial general intelligence through dynamic embodied physical and social interactions. Engineering, 34:12-22, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.405, + 0.907, + 0.462 + ], + "angle": 0, + "content": "[42] Asanka G Perera, Yee Wei Law, and Javaan Chahl. Uav-gesture: A dataset for uav control and gesture recognition. In Proceedings of the European Conference on Computer Vision (ECCV) Workshops, pages 0–0, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.462, + 0.907, + 0.545 + ], + "angle": 0, + "content": "[43] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.547, + 0.906, + 0.589 + ], + "angle": 0, + "content": "[44] Razieh Rastgoo, Kourosh Kiani, and Sergio Escalera. Sign language recognition: A deep survey. Expert Systems with Applications, 164:113794, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.59, + 0.907, + 0.631 + ], + "angle": 0, + "content": "[45] Fiona Ryan, Ajay Bati, Sangmin Lee, Daniel Bolya, Judy Hoffman, and James M Rehg. Gaze-lle: Gaze target estimation via large-scale learned encoders. 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.632, + 0.906, + 0.675 + ], + "angle": 0, + "content": "[46] Harvey Sacks, Emanuel A Schegloff, and Gail Jefferson. A simplest systematics for the organization of turn-taking for conversation. language, 50(4):696-735, 1974. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.676, + 0.907, + 0.732 + ], + "angle": 0, + "content": "[47] Yuehao Song, Xinggang Wang, Jingfeng Yao, Wenyu Liu, Jinglin Zhang, and Xiangmin Xu. Vitgaze: gaze following with interaction features in vision transformers. Visual Intelligence, 2(1):1-15, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.733, + 0.907, + 0.815 + ], + "angle": 0, + "content": "[48] Andreas Stolcke, Klaus Ries, Noah Coccaro, Elizabeth Shriberg, Rebecca Bates, Daniel Jurafsky, Paul Taylor, Rachel Martin, Carol Van Ess-Dykema, and Marie Meteor. Dialogue act modeling for automatic tagging and recognition of conversational speech. Computational linguistics, 26(3):339-373, 2000. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.817, + 0.907, + 0.9 + ], + "angle": 0, + "content": "[49] Kosei Tanada, Shigemichi Matsuzaki, Kazuhito Tanaka, Shintaro Nakaoka, Yuki Kondo, and Yuto Mori. Pointing gesture understanding via visual prompting and visual question answering for interactive robot navigation. In First Workshop on Vision-Language Models for Navigation and Manipulation at ICRA 2024. 5" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.907, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19518" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.487, + 0.161 + ], + "angle": 0, + "content": "[50] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In International conference on machine learning, pages 10347-10357. PMLR, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.164, + 0.486, + 0.233 + ], + "angle": 0, + "content": "[51] Du Tran, Heng Wang, Lorenzo Torresani, Jamie Ray, Yann LeCun, and Manohar Paluri. A closer look at spatiotemporal convolutions for action recognition. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 6450-6459, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.234, + 0.484, + 0.303 + ], + "angle": 0, + "content": "[52] Jun Wan, Yibing Zhao, Shuai Zhou, Isabelle Guyon, Sergio Escalera, and Stan Z Li. Chalearn looking at people rgb-d isolated and continuous datasets for gesture recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 56-64, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.304, + 0.484, + 0.374 + ], + "angle": 0, + "content": "[53] Limin Wang, Yuanjun Xiong, Zhe Wang, Yu Qiao, Dahua Lin, Xiaou Tang, and Luc Van Gool. Temporal segment networks: Towards good practices for deep action recognition. In European conference on computer vision, pages 20-36. Springer, 2016. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.375, + 0.484, + 0.444 + ], + "angle": 0, + "content": "[54] Limin Wang, Bingkun Huang, Zhiyu Zhao, Zhan Tong, Yinan He, Yi Wang, Yali Wang, and Yu Qiao. Videomae v2: Scaling video masked autoencoders with dual masking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14549-14560, 2023. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.445, + 0.484, + 0.514 + ], + "angle": 0, + "content": "[55] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.516, + 0.486, + 0.583 + ], + "angle": 0, + "content": "[56] Yi Wang, Kunchang Li, Xinhao Li, Jiashuo Yu, Yinan He, Guo Chen, Baoqi Pei, Rongkun Zheng, Jilan Xu, Zun Wang, et al. Internvideo2: Scaling video foundation models for multimodal video understanding. arXiv preprint arXiv:2403.15377, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.585, + 0.484, + 0.614 + ], + "angle": 0, + "content": "[57] Diyi Yang, Dirk Hovy, David Jurgens, and Barbara Plank. The call for socially aware language technologies, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.615, + 0.484, + 0.669 + ], + "angle": 0, + "content": "[58] Ceng Zhang, Junxin Chen, Jiatong Li, Yanhong Peng, and Zebing Mao. Large language models for human-robot interaction: A review. Biomimetic Intelligence and Robotics, page 100131, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.671, + 0.484, + 0.726 + ], + "angle": 0, + "content": "[59] Chen-Lin Zhang, Jianxin Wu, and Yin Li. Actionformer: Localizing moments of actions with transformers. In European Conference on Computer Vision, pages 492-510. Springer, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.728, + 0.484, + 0.783 + ], + "angle": 0, + "content": "[60] Yifan Zhang, Congqi Cao, Jian Cheng, and Hanqing Lu. Egogesture: A new dataset and benchmark for egocentric hand gesture recognition. IEEE Transactions on Multimedia, 20(5):1038-1050, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.785, + 0.484, + 0.826 + ], + "angle": 0, + "content": "[61] Y Zhang, B Li, H Liu, Y Lee, L Gui, D Fu, J Feng, Z Liu, and C Li. Llava-last: A strong zero-shot video understanding model. 2024. 8" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.487, + 0.826 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19519" + } + ] +] \ No newline at end of file diff --git a/2025/SocialGesture_ Delving into Multi-person Gesture Understanding/17e9a440-617f-4e5f-9712-f768185c9d5d_origin.pdf b/2025/SocialGesture_ Delving into Multi-person Gesture Understanding/17e9a440-617f-4e5f-9712-f768185c9d5d_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..950990a5222289db2e317903d165af36d878a76a --- /dev/null +++ b/2025/SocialGesture_ Delving into Multi-person Gesture Understanding/17e9a440-617f-4e5f-9712-f768185c9d5d_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2f280860c93026987c3e1fe4cd3d8c9e94689c20314112e7c22ee24232819593 +size 1606834 diff --git a/2025/SocialGesture_ Delving into Multi-person Gesture Understanding/full.md b/2025/SocialGesture_ Delving into Multi-person Gesture Understanding/full.md new file mode 100644 index 0000000000000000000000000000000000000000..4eff04c3217e0c5ad44b9ad2d83707f348ad7bfd --- /dev/null +++ b/2025/SocialGesture_ Delving into Multi-person Gesture Understanding/full.md @@ -0,0 +1,308 @@ +# SocialGesture: Delving into Multi-person Gesture Understanding + +Xu Cao1, Pranav Virupaksha1, Wenqi Jia1, Bolin Lai2, Fiona Ryan2, Sangmin Lee3†, James M. Rehg1† +1 University of Illinois Urbana-Champaign +2 Georgia Institute of Technology +3Sungkyunkwan University + +{xucao2,pranavv3,wenqij5,jrehg}@illinois.edu {bolin.lai,fkryan}@gatech.edu +sangmin.lee@skku.edu + +# Abstract + +Previous research in human gesture recognition has largely overlooked multi-person interactions, which are crucial for understanding the social context of naturally occurring gestures. This limitation in existing datasets presents a significant challenge in aligning human gestures with other modalities like language and speech. To address this issue, we introduce SocialGesture, the first largescale dataset specifically designed for multi-person gesture analysis. SocialGesture features a diverse range of natural scenarios and supports multiple gesture analysis tasks, including video-based recognition and temporal localization, providing a valuable resource for advancing the study of gesture during complex social interactions. Furthermore, we propose a novel visual question answering (VQA) task to benchmark vision language models' (VLMs) performance on social gesture understanding. Our findings highlight several limitations of current gesture recognition models, offering insights into future directions for improvement in this field. SocialGesture is available at hugging-face.co/datasets/IrohXu/SocialGesture. + +# 1. Introduction + +Long before the development of spoken language in human history, deictic gestures played an important role in human social communication, serving as one of the earliest communication tools for expressing thoughts, emotions, and intentions. In contemporary communication, both verbal (e.g., speech) and non-verbal (e.g., gesture) cues work in concert. While spoken words express direct meanings, understanding the complete social context often requires more than just the words alone. Gestures play an important role in clarifying intention or points of emphasis, and augment spoken communication with additional social context. Theoretical analysis suggests that both speech and gesture originate from a shared + +![](images/1856bd464f77839b2f1d7e4753513dcbf4386790b94ac6d29f511aa919be073c.jpg) + +![](images/776fd41c8ed5847d22db3b0257d12edb541f39528a1ef4bfe0004ba7ba84b47a.jpg) + +![](images/bfe90554e54ee68f1fd15a00c344b3ce865f29b8ff3898d5b05dc4bf7e9235fc.jpg) +Figure 1. Example frames from six gesture datasets. SocialGesture is the only dataset featuring multi-person interactions and focusing on natural gestures with meaningful social communication. + +![](images/542625c58de26cd3244a6742f68e03e35583c1db8a10407e4d28f0c709743be2.jpg) + +![](images/a47a871c0a59d2acc6e6348c87e026edc36024c80b7889d804162b202bee50fd.jpg) + +![](images/e0f5acfb9d7626e05c51e7ac231d47011c3ff0b452e7edf8abc28b53d7c0f083.jpg) + +cognitive system, with gestures representing imagery-driven data and speech representing symbolic data [6]. Together, they reflect a unified communicative intent. + +Recent advancements in machine learning have driven research on the understanding of human social communication [27, 58], yet most of the latest social research predominantly focuses on linguistic modalities, often overlooking visual social cues like gestures [28]. While there have been machine learning works on human gestures, most of them have focused on gestures for device manipulation or sign languages [42, 44], and the problem of analyzing natural social gestures between multiple people has not received significant attention. A primary reason for this limitation is the lack of gesture data and annotations that capture multi-person interactions and the use of gestures in realistic social contexts. While recent datasets such as HaGRID [23] and LD-ConGR [31] address high-resolution and long-distance gesture collection (Figure 1), they remain confined to controlled, + +
DatasetMulti-personVideo ClipsInstancesLabel
CategoryBboxRelationVQA
Jester [37]X148,092148,092✓XXX
NVGesture [39]X1,5321,532✓✓XX
EgoGesture [60]X2,08124,161✓✓XX
ChaLearn ConGD [52]X22,53547,933✓✓XX
IPN Hand [5]X2004,218✓✓XX
UAV gesture [42]X119-✓✓XX
iMiGUE [34]X35918,499✓XXX
LD-ConGR [31]X54244,887✓✓XX
SocialGesture (Ours)✓9,88942,533✓✓✓✓
+ +Table 1. Comparison of SocialGesture with existing video-based gesture datasets. SocialGesture is the first work that addresses multi-person social interaction cases. Our dataset uniquely provides comprehensive annotations including gesture categories, bounding boxes (humans and objects), gesture relations (subject person and target person), and visual question answering pairs. + +single-person settings, and do not contain the spontaneous, natural gestures between multiple people that characterize real-life social interactions. + +To address this limitation, we introduce SocialGesture, a comprehensive dataset designed specifically for understanding gestures in multi-person social interactions. SocialGesture provides three key advantages over existing datasets: + +1. It focuses on videos with multiple people, making it the first multi-person gesture dataset, in contrast to previous datasets which lack human-to-human interaction. +2. SocialGesture includes four types of comprehensive annotations: (a) Social gesture categories, identifying the specific class of a gesture (i.e., pointing, showing, giving, reaching) within social interactions; (b) Temporal-spatial annotations, identifying the appearance of social gestures temporally and spatially through time stamps and bounding boxes. (c) Interaction annotations, capturing interpersonal social dynamics that result from gestures; and (d) Visual question answering (VQA) annotations, which link gestures to descriptive content, facilitating the alignment of gestures with contextual meaning. All of these annotations offer a rich resource for developing models for social gesture understanding. +3. SocialGesture comprises a large volume of diverse and naturally occurring social interactions sourced from YouTube and Ego4D [17], and is the largest multi-person gesture dataset available. + +Based on our novel SocialGesture dataset, we introduce diverse gesture-related benchmark tasks from temporal localization and recognition to visual question answering. Our experiments reveal significant challenges in multi-person gesture understanding for current architectures. It is the first step toward building agents that have capabilities to seamlessly understand non-verbal gesture cues in social contexts. The major contributions of this paper are as follows. + +- We introduce SocialGesture, the first dataset focusing on + +multi-person gesture interactions in natural social settings. Unlike existing datasets that are limited to single-person or controlled environments, our dataset captures the complexity of real-world gesture-based social communication. + +- We provide comprehensive multi-level annotations including gesture categories, temporal-spatial localization, interaction dynamics, and visual question-answering pairs. This rich annotation scheme enables the development of models for understanding social gestures. +- We perform comprehensive experiments across gesture localization and recognition, evaluating a wide breadth of architectures and observe multiple challenging areas for future works. We further benchmark multiple large vision-language models on our proposed diverse Social Visual Question-Answering tasks. + +# 2. Related Works + +# 2.1. Gesture Datasets + +As shown in Figure 1 and Table 1, the majority of gesture recognition datasets are video-based, with recent research emphasizing emotional state analysis, increased capture distance for improved model robustness, and egocentric gesture collection. Notable datasets include LD-ConGR, featuring 542 RGB-D videos representing 10 gestures [31], and HaGRID, which focuses on high-resolution image based gesture recognition across 18 hand gesture classes [23]. EgoGesture, with over 2 million frames from 50 subjects, is tailored for egocentric interactions with wearable devices [60], while iMiGUE [34] and SMG focus on micro-gesture recognition [10], containing 359 and 40 videos, respectively. Earlier datasets such as Jester [37], IPN Hand [5], ChaLearn IsoGD [52], and ChaLearn ConGD [52] have also made significant contributions, covering tasks like hand segmentation, and isolated or continuous gesture classification. In contrast, we are the first to provide a large-scale multi-person gesture + +![](images/e2c11aca6089c1a682ea958dc15dd43466e01cf881dce0be675f182f86ffab15.jpg) +Pointing Gesture + +![](images/6da6c39c5e789fca6dc8dd1f7e3f56cd60054a1dbfab86a6eb6db1fb20a21aa2.jpg) +Showing Gesture + +![](images/2dbd399b4634d0acb8e33efacf0866942d8fa13eb263a4cbd5a9ac716cfaa7c5.jpg) +Giving Gesture +Figure 2. Examples of the four deictic gesture categories in SocialGesture with subject-target relationships. From left to right: pointing (directing attention), showing (presenting objects), giving (transfer intention), and reaching (acquisition intention) gestures. Red boxes indicate gesture initiators (subjects) and blue notations indicate targets. + +![](images/ad3a5e03d0df02531104d4520b14cb1928f772deca5bc9e5f100622379bb0212.jpg) +Reaching Gesture + +dataset, and we are also the first to provide comprehensive coverage of the four main classes of deictic gestures(defined in [38]), which play a crucial role in social communication. + +# 2.2. Social Understanding with Non-verbal Cues + +Understanding social interactions has been long studied in the domain of natural language processing [15, 18, 25, 26, 46, 48]. However, the inherent multimodal nature of social interactions ensures that non-verbal elements (such as gestures, gaze and emotions) play an important role in communication [28], which are understudied in previous research [57]. Recently, several investigations have been conducted on these implicit non-verbal cues. Liu et al. [31] build a dataset to identify the gestures of single person from a long distance. Unfortunately, they only focus on isolated gestures of a single person, and fail to contextualize gesture recognition in a real-world social scenario. Non-verbal social understanding of multiple people demands more investigation [16, 21, 22, 41]. To this end, we make the first attempt to model multi-party social interactions by defining various social gesture understanding tasks, such as social gesture detection, recognition, and VQA, based on our dataset. + +# 2.3. Foundation Models for Video Understanding + +Video foundation models are designed to tackle a wide range of understanding and analysis tasks, including classification, temporal action localization, question answering, captioning, and more. Early approaches to video understanding relied on CNN blocks as visual encoders for video frames, connected to a prediction head or decoder [9, 13]. With the emergence of transformers, new architectures like MViTv2-B [30], VideoSwin [36], UniFormerV2 [29], and VideoMAE-2 [54] tokenize their visual inputs, enabling unified integration with other modalities [56]. Building on this, recently released video-based VLMs like Qwen2-VL [55], LLaVA-NEXT [32], and InternVL-2.5 [12] leverage pretrained visual + +encoders to extract rich representations. These models align visual tokens with textual tokens, enabling a deeper multimodal understanding that bridges the gap between video and language tasks. We provide the first investigation of the effectiveness of these models in understanding multi-party social gestures via a VQA task. + +# 3. SocialGesture + +# 3.1. Motivation + +Social gestures comprise a fundamental set of human behavioral practices that facilitate interpersonal communication through hand and body movements in social contexts, distinct from sign language. According to [38], gestures can be classified into four primary categories: deictic, beat, iconic, and metaphoric. Deictic gestures further subdivide into pointing, showing, giving, and reaching. Beat gestures are rhythmic movements that emphasize certain words and phrases during speech. Iconic gestures visually represent concrete objects or actions (e.g., making a circle with fingers), while metaphoric gestures symbolize abstract concepts (e.g., making a circular motion to indicate repetition). + +Among these categories, deictic gestures deserve particular attention due to their prevalence and significance in social interactions. While iconic and metaphoric gestures appear in specific contexts and beat gestures primarily serve prosodic functions, deictic gestures play a crucial role in establishing joint attention and facilitating object-mediated social interaction. Therefore, our dataset focuses on four fundamental deictic gestures (see Figure 2): + +- Pointing Gesture involves directing attention to a specific entity in the social scene through finger extension. While commonly executed with an extended index finger and arm, these specific forms are not mandatory requirements. The crucial identifying factor is the communicative intent to guide others' attention to a particular target, whether + +
Video ContentSourceProportionTotal Length# of PeopleDescription
Group social gamesYouTube44.51%896 min>=3A group of people playing social games such as one night werewolf
Ego4D21.91%441 min>=4
Variety entertainmentYouTube22.31%449 min>=2People performing activities such as pranks, challenges, and interviews
Educational playYouTube2.53%51 min>=2Children engaged in educational activities with adult guidance
Product reviewsYouTube2.53%51 min>=2People recommending products such as advent calendars
Party & dinnerYouTube3.63%73 min>=7A group of people having party & dinner together
Group cookingYouTube2.58%52 min>=3People engaged in cooking activities together
+ +Table 2. SocialGesture contains diverse videos including different multi-person social interactions from YouTube and Ego4D [17]. + +person or object. This distinguishes pointing from incidental hand movements or other gesture types, particularly reaching gestures where the primary intent differs. + +- **Showing Gesture** involves presenting an object for others to see it by manipulating or orienting it towards observers. Unlike pointing, the object itself is the focus of attention, and it does not need to be fully supported by the subject, such as when tilting or sliding an object to make its contents visible. The key criterion for a showing gesture is that the object is being presented for visual inspection, regardless of whether the target person actually looks at it or where the subject is directing their gaze. +- Giving Gesture involves manipulating an object with the intention of transferring it to another person's possession. It differs from pointing because the object originates with the subject, and from showing because the subject is inviting the other person to take possession of the object. The key is the subject's intention to transfer the object, regardless of whether the recipient actually takes it. +- Reaching Gesture manifests as a hand extension toward an object, expressing desire for possession or requesting transfer. It is characterized by full finger extension and often includes pre-grasp configuration and associated body movements like leaning forward. Reaching differs from showing and giving because the object is not in the subject's possession, and it differs from pointing because the subject is expressing a desire to possess the object, not just to draw attention to it. The key aspect of reaching is the intent to obtain the object, regardless of whether the subject can physically retrieve it. + +# 3.2. Data Acquisition + +Data collection. We collect raw video data for SocialGesture from various sources, including YouTube channels and Ego4D. Table 2 represents the overall composition of the dataset. Video selection is based on four key criteria: (1) Video quality – we manually check each video to ensure high resolution; (2) Number of people – each video is carefully selected to include between two and ten people, ensuring clear visibility of gestures; (3) Video length – the total length of selected videos ranges from 2 to 30 minutes, with each gesture lasting between 1 to 15 seconds; and (4) Scene diversity – to enhance generalization, we select videos featuring + +diverse races, genders, and ages. + +Pre-processing. For videos longer than 10 minutes, we cut them into video clips of 5 minutes in length. All videos are then converted to $720\mathrm{p}$ $(1280\times 720)$ with 30 FPS in the initial curation and are then mapped to $360\mathrm{p}$ $(640\times 360)$ with 5 FPS for further annotation. + +Gesture annotation. Each gesture in the SocialGesture dataset is defined as a sequence of video frames where an individual performs an arm or hand movement corresponding to one of four specified deictic gestures: pointing, showing, giving, and reaching. The annotations capture both temporal and spatial aspects of these social interactions through a comprehensive multi-level annotation framework. For temporal annotation, we identify the complete sequence of frames containing a gesture, from initial movement through completion. Within this sequence, we designate a key frame that best captures the defining moment of the gesture - typically when the gesture reaches its most distinctive configuration. This key frame serves as an anchor point for additional annotations. Spatial annotation includes precise bounding box coordinates [x1, y1, x2, y2] for both the gesture initiator and the target. The initiator bounding box encompasses the person performing the gesture, while the target bounding box identifies either the person or object toward which the gesture is directed. These spatial annotations enable analysis of the geometric relationships between interaction participants. Our annotation framework also includes natural language descriptions of each gesture's social context, capturing the broader interaction dynamics beyond pure spatiotemporal coordinates. + +# 3.3. Benchmark Tasks + +# 3.3.1. Social Gesture Temporal Localization + +Temporal action localization is a challenging task that requires not only identifying the temporal intervals of all detected gesture instances but also estimating corresponding confidence values. Unlike standard classification tasks, this process involves an end-to-end approach in which classification is integrated within temporal localization, adding considerable complexity. For example, even if a gesture is accurately localized in time, it is treated as a false detection if assigned an incorrect class label. + +![](images/f5cad51e6459531a5fb52b20692e9de7501cfd29dfaacca3ef5a759496404e2d.jpg) +Figure 3. Temporal gesture localization task for social gestures + +Figure 3 illustrates a general approach to temporal action localization. Each video is initially divided into multiple overlapping video clips, denoted as $X = [X_0, X_1, \ldots, X_K]$ , where $X_k \in \mathbb{R}^{T \times H \times W \times C}$ . Feature extractors such as I3D or VideoMAE are then employed to extract multiscale feature representations, denoted as $Z_k \in \mathbb{R}^D$ , for each video clip. These features are subsequently concatenated into a single representation, $Z \in \mathbb{R}^{T \times D}$ . The objective of the temporal action localization model is to generate sequence labels for each video clip. + +This task is also benchmarked by several well-known datasets, including THUMOS14 [20] and ActivityNet-1.3 [8]. However, our task presents additional challenges. In SocialGesture videos, multiple individuals are often present simultaneously, and gestures tend to be small, subtle portions of larger actions, making it uniquely challenging to detect and classify gestures accurately. + +# 3.3.2. Social Gesture Recognition + +Given the complexity of social gesture localization, we also divide long videos into shorter clips of 2-5 seconds and introduce the social gesture recognition task. This task can be approached in multiple ways. The first approach is isolated gesture recognition, where a long video is segmented into discrete gesture clips. Video action recognition models are then trained to classify the gestures in each clip independently. Alternatively, gestures can be identified continuously throughout the entire video, from beginning to end. Although most action localization methods we discussed in the previous section can typically handle this task, most existing video feature extraction pre-processing methods struggle with consistent multi-person action localization. Therefore, sliding windows remain a common strategy for this purpose: the video is processed in overlapping segments of a predefined window size, with each window analyzed by the model to calculate gesture confidence scores. + +We divide recognition into two subtasks (See Figure 4): + +- Task 2-1: Gesture vs Non-gesture Classification This task aims to distinguish between social gesture and non-gesture video clips. The input is a short video clip $X \in$ + +![](images/549ea7e37bcb7955888324f586c1e2287b642c527a165124155c6b20ee6333bd.jpg) +Figure 4. Gesture recognition tasks for social gestures + +$\mathbb{R}^{T\times H\times W\times C}$ , where $T < 16$ , and the output is a label $\hat{Y}$ indicating whether a gesture is present. + +- Task 2-2: Gesture Type Classification. This task focuses on classifying video clips containing one of four social gestures: pointing, showing, giving, or reaching. The input and output structures are consistent with those used in the binary classification task. We further explore whether using cropped bounding boxes for each individual within a scene improves the accuracy of social gesture classification. The input is a short video clip cropped by the maximum region of the initiator of a social gesture and the output is a label $\hat{Y}$ indicating four types of social gesture. + +For each task, we also strive to sample the training set to achieve a balanced label distribution. This adjustment is necessary because pointing gestures are significantly more frequent than showing, giving, and reaching in real-world communication contexts. + +# 3.3.3. Social Gesture Visual Question Answering + +Prior to our work, there has been very limited research on designing video captioning and Visual Question Answering (VQA) specifically for human gesture datasets [49]. This gap may be due to the lack of multi-person interactions in existing video datasets, reducing the necessity for enhanced alignment between the gesture and language modalities. To fill this gap, we introduce the first VQA task focused on human gestures (Figure 5). Our VQA can be divided into three subtasks: (1) Global Perception, (2) Gesture Understanding, and (3) Gesture Localization. + +The Global Perception task serves as a foundational test to assess whether VLMs can comprehend basic events in short video clips, including scene descriptions and counting the number of humans present. We generate data for this task through a semi-automated QA generation pipeline based on GPT-4o [2], similar to the instruction-tuning data pair generation in LLaVA [33]. Initially, we manually create short descriptions for each video series, including the social background. We prompt these descriptions with the key frame of the video clip into the GPT-4o pipeline to generate QA pairs for human counting and scene description. Each QA + +![](images/81c2e9da614e48b861333c9ea0a4b4d3df052203e4995a0b596681717326c0e0.jpg) +Figure 5. The question-answer pairs of SocialGesture. We omit the options of each question in the figure. The bounding box defined by [top-left x, top-left y, bottom-right x, bottom-right y]). The definition will be provided together with the system prompts. + +pair is then reviewed by a human annotator. Figure 5 shows the example of global perception (Task 3-1). Human counting is used as one VQA subtask in the VLM benchmarking experiments. + +For Gesture Understanding, we utilize gesture classification annotations to create questions related to social gesture classification (Figure 5, Task 3-2). This results in two types of questions: (1) detecting whether a social gesture occurs, and (2) classifying between the four types of social gestures. + +Gesture Localization leverages bounding boxes from our annotations to generate questions that test the model's spatial and temporal comprehension. These questions include providing the bounding box localization of the initiator, then determining whether the gesture's target is a human, and locating the target. We present an example of gesture localization (Task 3-3) in Figure 5. These questions are designed + +
Feature0.30.40.50.60.7Avg mAP
I3D [9]24.8516.319.312.220.9610.73
R(2+1)D [51]14.3810.257.232.811.777.29
VideoMAEV2 [54]27.2325.0513.335.282.7614.73
+ +Table 3. Temporal action localization for four social gestures by different feature extractors with ActionFormer [59]. + +
Stride0.30.40.50.60.7Avg mAP
1611.098.605.522.671.825.94
824.8516.319.312.220.9610.73
431.6429.1319.3013.772.1119.19
+ +Table 4. Explore the influence of stride for I3D and ActionFormer. + +to challenge temporal-spatial reasoning in a multi-person setting, requiring analysis to produce the correct responses. + +# 4. Experiments + +# 4.1. Experimental Setup + +We split the raw videos (long video) into training set (293 videos) and test set (79 videos), and then cut each long video into video clips. The 9,889 video clips used in video action recognition and social gesture VQA tasks come from the same train-test split in the raw videos. We also cut another 4,304 video clips from the raw videos as the non-gesture class. All experiments are conducted with 2 NVIDIA L40S GPUs and 2 H100 GPUs to make sure the batch size is 16 and learning rate is 5e-4 for all models. For all methods in Table 3, 5, 6, and 7, we apply the same data augmentation strategies used in [29, 50]. To solve the class imbalanced issue, we also adjust the size of training set for each task. + +# 4.2. Social Gesture Temporal Localization + +Baselines. We used two-stream I3D [9], R(2+1)D [51], and VideoMAEV2 [54] with the same sliding window of 16 frames and different stride sizes for feature extraction. Then, these video features are used as input for ActionFormer [59] during model training. We used mAP@[0.3:0.1:0.7] as the main evaluation metric and also reported the average mAP. + +Results and Findings. We observe that almost all feature extractors underperform on this task in Table 3. Notably, despite the strong performance of features from VideoMAE V2 [54] in other action localization datasets, such as THU-MOS 2014 [20] and ActivityNet [8], they only achieve an average mAP of 14.73 for social gesture recognition. To investigate the effect of the stride size in the sliding window on the average mAP, we conducted an ablation study in Table 4 using different stride sizes of the I3D feature in SocialGesture. We found that reducing the stride from 16 to 4 led to an increase in average mAP; however, the overall results remained insufficient. This shortfall can be attributed to the + +
ModelPretrainParamAcc (%)
TSN-R50 [53]Kinetics-400 [24]24M78.77
TANet-R50 [35]Kinetics-400 [24]26M71.22
SlowFast-R50 [13]Kinetics-400 [24]35M80.82
SlowFast-R101 [13]Kinetics-400 [24]63M79.59
TimeSformer-L [7]Kinetics-400 [24]121M78.71
MViTv2-B [30]Kinetics-400 [24]51M83.29
VideoSwin-B [36]Kinetics-400 [24]88M81.70
VideoSwin-L [36]Kinetics-400 [24]200M83.44
UniFormerV2-B/16 [29]CLIP [43]115M84.43
UniFormerV2-L/14 [29]CLIP [43]354M80.93
+ +fact that all feature extractors were pretrained on datasets that lack multi-person interactions, causing the features to be poorly aligned with the specific requirements of our task. + +# 4.3. Social Gesture Recognition + +Baselines. Our baselines for social gesture recognition includes different CNN-based video models such as TSN [53], TANet [35], SlowFast [13], and Vision Transformer based models such as TimeSformer [7], MViTv2 [30], VideoSwin [36], UniFormerV2 [29]. All these models are pretrained with Kinetics-400 [24] dataset or using CLIP encoder [43]. We use Accuracy for gesture and non-gesture binary classification in Table 5 and Top 1 Accuracy for four social gesture classification in Table 6, and 7. + +Results and Findings. The results in Table 5 indicate that most existing action recognition models can differentiate social gestures and non-gesture from short video clips. Despite fine-tuning these models using over 10,000 video clips for the binary classification task, their performance remains suboptimal. The SOTA baseline, UniFormerV2-B/16, achieves an accuracy of only $84.43\%$ . Notably, Transformer-based models generally outperform CNN-based models. + +Table 6 presents results for the four-class social gesture classification task, revealing that all models struggle to effectively classify these gestures—despite the simplicity of the task for human annotators. The highest performance comes from VideoSwim-L, achieving only $56.18\%$ . Given that individual gestures can be challenging to recognize in multi-person environments, we conducted further experiments, as shown in Table 7. Specifically, we extracted the region corresponding to each individual using ground truth bounding boxes and fine-tuned the same models with these per-person inputs. UniFormerV2-B/16 achieves the best performance of $64.72\%$ top 1 accuracy. This may be because social gestures inherently involve fine-grained and subtle movements. Consequently, recognizing such gestures is difficult from both the extracted per-person region and the full-frame perspective. Addressing the challenge of social gesture understanding requires advanced models capable of + +Table 5. Experiments on binary classification for social gesture and non-gesture. + +
ModelPretrainParamTop1 Acc (%)
TSN-R50 [53]Kinetics-400 [24]24M54.83
TANet-R50 [35]Kinetics-400 [24]26M45.39
SlowFast-R50 [13]Kinetics-400 [24]35M45.17
SlowFast-R101 [13]Kinetics-400 [24]63M46.07
TimeSformer-L [7]Kinetics-400 [24]121M53.03
MViTv2-B [30]Kinetics-400 [24]51M37.98
VideoSwin-B [36]Kinetics-400 [24]88M53.93
VideoSwin-L [36]Kinetics-400 [24]200M56.18
UniFormerV2-B/16 [29]CLIP [43]115M55.51
UniFormerV2-L/14 [29]CLIP [43]354M50.34
+ +Table 6. Experiments on classification for four social gestures. + +
ModelPretrainParamTop1 Acc (%)
TSN-R50 [53]Kinetics-400 [24]24M58.43
TANet-R50 [35]Kinetics-400 [24]26M46.29
SlowFast-R50 [13]Kinetics-400 [24]35M45.17
SlowFast-R101 [13]Kinetics-400 [24]63M42.92
TimeSformer-L [7]Kinetics-400 [24]121M63.60
MViTv2-B [30]Kinetics-400 [24]51M35.28
VideoSwin-B [36]Kinetics-400 [24]88M60.00
VideoSwin-L [36]Kinetics-400 [24]200M63.60
UniFormerV2-B/16 [29]CLIP [43]115M64.72
UniFormerV2-L/14 [29]CLIP [43]354M60.45
+ +Table 7. Experiments on classification for four social gestures after extract each subject person's bounding box. + +disentangling the relationship between the subject and the target person. + +# 4.4. SocialGesture VQA + +From the results of previous tasks, we found that distinguishing social gestures is still very challenging for current SOTA methods in gesture localization and recognition. This difficulty highlights the unique nature of social gesture, where models need to have a unified view of multi-person social interactions and can distinguish details that are hard to detect. This is something traditional video analysis models still struggle with [1, 14]. Inspired by recent progress in multimodal video grounding and video-based VQA [11, 40], we aim to leverage the reasoning capabilities of the latest VLMs and evaluate their performance on this complex task. + +Baselines. To assess performance on the SocialGesture VQA benchmark, we chose several SOTA publicly available VLMs, including Qwen2-VL [55], Qwen2.5-VL [4], InternVL-2.5 [12], and LLAVA-NEXT-Video (LLaVA 1.6) [32]. Additionally, we included closed-source multimodal LLMs such as GPT-4o-mini [2] and Claude-3.7-sonnet [3] for a broader comparison. Note that GPT-4o [2] was excluded, as it was used to generate some of the VQA instruction-following training data, which could lead to biased results. + +Evaluation Metrics. Since all VQA pairs can be transformed into a multi-option output, we use rule-based accuracy metrics to evaluate the VQA tasks. Each model was + +
ModelParamGlobal Perception (%)Gesture Understanding (%)Gesture Localization (%)
HumanCount@AccGestureDet@AccGestureClass@AccTargetLoc@AccTargetClass@Acc
Random Select--50.0025.0018.9050.00
InternVL-2.5 [12]2B22.3763.3474.4513.3339.79
InternVL-2.5 [12]8B35.5873.9481.8017.3857.28
LLaVA-NeXT-Video [32, 61]7B30.6371.9234.085.3959.90
Qwen2-VL [55]7B60.1874.1567.189.5859.19
Qwen2.5-VL [4]72B69.9775.6954.8229.3761.38
Claude 3.7 Sonnet [3]-63.2769.6461.7325.4864.20
GPT-4o-mini [2]-53.4073.6768.5221.2965.12
+ +Table 8. Experiments on zero-shot performance of SOTA VLMs, including closed sourced multimodal LLMs for Social Gesture VQA + +
ModelInputGlobal Perception (%)Gesture Understanding (%)Gesture Localization (%)
HumanCount@AccGestureDet@AccGestureClass@AccTargetLoc@AccTargetClass@Acc
Qwen2-VL-7Bvideo60.1874.1567.189.5859.19
Qwen2-VL-7B + LoRA [19] SFTvideo82.5533.0784.5249.2141.19
Qwen2-VL-7B + Full SFTvideo83.6433.0781.1351.2538.51
+ +Table 9. Experiments on SFT performance of SOTA VLMs, including closed sourced multimodal LLMs for Social Gesture VQA + +given sufficient context for the task, followed by a question and multiple answer choices. Accuracy was determined by comparing the model's response with the correct answer. We evaluated models across all subtasks including global perception, gesture understanding, and gesture localization in Section 3.3.3. The accuracy (Acc) in Table 8 is defined as: $\mathrm{Acc} = \frac{1}{N}\sum_{k = 1}^{N}\mathrm{MATCH}_k$ , where $N$ is the total number of VQA samples and $\mathrm{MATCH}_k$ is a binary value. If the option in the model output matches the ground truth option, it is 1, otherwise it is 0. + +Results and Findings. In Table 8, the experimental results show that while most VLMs can handle global perception and gesture understanding tasks without finetuning, they all struggle significantly with gesture localization tasks. This trend is understandable, as localization tasks inherently involve more ambiguity compared to perception and understanding tasks. While there is not a very clear correlation between the number of parameters and performance, Qwen2.5-VL, the largest of our open sourced models, performed well overall. To explore the potential for improvement, we finetuned Qwen2-VL-7B in Table 9, which led to enhanced performance in most metrics, except those related to gesture binary classification and target identification. Performance drops in some metrics can be attributed to easy overfitting on binary cases. An additional noteworthy observation is that despite being pretrained on vast multimodal datasets, none of the SOTA VLMs achieved over $70\%$ accuracy on the simplest task: counting humans in social interaction scenes. + +These findings indicate that the SocialGesture tasks are significantly more challenging for VLMs than typical VQA benchmarks. This difficulty stems from the natural complexity of real-world social interaction videos used in our + +benchmark. Unlike controlled environments, our dataset contains realistic scenarios where people frequently overlap with each other, are partially visible, or extend beyond the camera frame, creating natural occlusion cases. In addition, there exists high interdependence between tasks in SocialGesture VQA, which makes it challenging. For instance, when models process such videos, they should first assess the overall scene despite these visual challenges, counting people and understanding actions. This process involves high-level reasoning and the ability to sequentially infer relationships—skills that are currently lacking in SOTA VLMs. Developing models capable of such contextual social visual reasoning remains a key challenge for future research. + +# 5. Conclusion + +We introduce SocialGesture, the first dataset and benchmark specifically designed for understanding and analyzing multiperson social gestures. We provide a comprehensive definition of tasks related to social gesture recognition, including a novel VQA dataset that encompasses global perception, gesture understanding, and spatial localization. To advance the field of human gesture and social interaction understanding, we benchmarked a variety of current state-of-the-art methods, including action recognition models and vision-language models (VLMs). Our findings reveal that multiperson social gesture understanding requires incremental visual reasoning—area where existing computer vision models fall short. This work offers insights into the complexities of multi-person social interactions and highlights the need for continued research into developing models capable of sophisticated visual reasoning. We hope that SocialGesture will motivate further innovations in gesture understanding. + +# Acknowledgment + +We are thankful to the Keymakr for their great work in data annotation. This work was supported in part by the NRF grant funded by the Korea government(MSIT) (No. RS-2025-00563942). + +# References + +[1] Qaisar Abbas, Mostafa EA Ibrahim, and M Arfan Jaffar. Video scene analysis: an overview and challenges on deep learning algorithms. Multimedia Tools and Applications, 77 (16):20415-20453, 2018. 7 +[2] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florecia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 5, 7, 8 +[3] Anthropic. Claude 3 family. https://www.anthropic.com/news/claude-3-family, 2024. Accessed: 2024-05-27. 7, 8 +[4] Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, et al. Qwen2. 5-vl technical report. arXiv preprint arXiv:2502.13923, 2025. 7, 8 +[5] Gibran Benitez-Garcia, Jesus Olivares-Mercado, Gabriel Sanchez-Perez, and Keiji Yanai. Ipn hand: A video dataset and benchmark for real-time continuous hand gesture recognition. In 2020 25th international conference on pattern recognition (ICPR), pages 4340-4347. IEEE, 2021. 2 +[6] Paolo Bernardis and Maurizio Gentilucci. Speech and gesture share the same communication system. Neuropsychologia, 44(2):178-190, 2006. 1 +[7] Gedas Bertasius, Heng Wang, and Lorenzo Torresani. Is space-time attention all you need for video understanding? In ICML, page 4, 2021. 7 +[8] Fabian Caba Heilbron, Victor Escorcia, Bernard Ghanem, and Juan Carlos Niebles. Activitynet: A large-scale video benchmark for human activity understanding. In Proceedings of the iee conference on computer vision and pattern recognition, pages 961-970, 2015. 5, 6 +[9] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6299-6308, 2017. 3, 6 +[10] Haoyu Chen, Henglin Shi, Xin Liu, Xiaobai Li, and Guoying Zhao. Smg: A micro-gesture dataset towards spontaneous body gestures for emotional stress state analysis. International Journal of Computer Vision, 131(6):1346-1366, 2023. 2 +[11] Qirui Chen, Shangzhe Di, and Weidi Xie. Grounded multihop videoqa in long-form egocentric videos. arXiv preprint arXiv:2408.14469, 2024. 7 +[12] Zhe Chen, Weiyun Wang, Yue Cao, Yangzhou Liu, Zhangwei Gao, Erfei Cui, Jinguo Zhu, Shenglong Ye, Hao Tian, Zhaoyang Liu, et al. Expanding performance boundaries of open-source multimodal models with model, data, and test + +time scaling. arXiv preprint arXiv:2412.05271, 2024. 3, 7, 8 +[13] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6202-6211, 2019. 3, 7 +[14] Chaoyou Fu, Yuhan Dai, Yongdong Luo, Lei Li, Shuhuai Ren, Renrui Zhang, Zihan Wang, Chenyu Zhou, Yunhang Shen, Mengdan Zhang, et al. Video-mme: The first-ever comprehensive evaluation benchmark of multi-modal llms in video analysis. arXiv preprint arXiv:2405.21075, 2024. 7 +[15] Deepanway Ghosal, Navonil Majumder, Soujanya Poria, Niyati Chhaya, and Alexander Gelbukh. Dialoguegcn: A graph convolutional neural network for emotion recognition in conversation. arXiv preprint arXiv:1908.11540, 2019. 3 +[16] Shreya Ghosh, Zhixi Cai, Abhinav Dhall, Dimitrios Kollias, Roland Goecke, and Tom Gedeon. Mrac track 1: 2nd workshop on multimodal, generative and responsible affective computing. arXiv preprint arXiv:2409.07256, 2024. 3 +[17] Kristen Grauman, Andrew Westbury, Eugene Byrne, Zachary Chavis, Antonino Furnari, Rohit Girdhar, Jackson Hamburger, Hao Jiang, Miao Liu, Xingyu Liu, et al. Ego4d: Around the world in 3,000 hours of egocentric video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18995-19012, 2022. 2, 4 +[18] Devamanyu Hazarika, Soujanya Poria, Amir Zadeh, Erik Cambria, Louis-Philippe Morency, and Roger Zimmermann. Conversational memory network for emotion recognition in dyadic dialogue videos. In Proceedings of the conference. Association for Computational Linguistics. North American Chapter. Meeting, page 2122. NIH Public Access, 2018. 3 +[19] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. *ICLR*, 1(2):3, 2022. 8 +[20] Haroon Idrees, Amir R Zamir, Yu-Gang Jiang, Alex Gorban, Ivan Laptev, Rahul Sukthankar, and Mubarak Shah. The thumos challenge on action recognition for videos "in the wild". Computer Vision and Image Understanding, 155:1-23, 2017. 5, 6 +[21] Simindokht Jahangard, Zhixi Cai, Shiki Wen, and Hamid Rezatofighi. Jrdb-social: A multifaceted robotic dataset for understanding of context and dynamics of human interactions within social groups. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22087-22097, 2024. 3 +[22] Wenqi Jia, Miao Liu, Hao Jiang, Ishwarya Ananthabhotla, James M Rehg, Vamsi Krishna Ithapu, and Ruohan Gao. The audio-visual conversational graph: From an egocentric-exocentric perspective. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26396–26405, 2024. 3 +[23] Alexander Kapitanov, Karina Kvanchiani, Alexander Nagaev, Roman Kraynov, and Andrei Makhliarchuk. Hagrid-hand gesture recognition image dataset. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 4572-4581, 2024. 1, 2 + +[24] Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, et al. The kinetics human action video dataset. arXiv preprint arXiv:1705.06950, 2017. 7 +[25] Bolin Lai, Hongxin Zhang, Miao Liu, Aryan Pariani, Fiona Ryan, Wenqi Jia, Shirley Anugrah Hayati, James Rehg, and Diyi Yang. Werewolf among us: Multimodal resources for modeling persuasion behaviors in social deduction games. Association for Computational Linguistics: ACL 2023, 2023. 3 +[26] Bongseok Lee and Yong Suk Choi. Graph based network with contextualized representations of turns in dialogue. arXiv preprint arXiv:2109.04008, 2021. 3 +[27] Sangmin Lee, Bolin Lai, Fiona Ryan, Bikram Boote, and James M Rehg. Modeling multimodal social interactions: new challenges and baselines with densely aligned representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14585-14595, 2024. 1 +[28] Sangmin Lee, Minzhi Li, Bolin Lai, Wenqi Jia, Fiona Ryan, Xu Cao, Ozgur Kara, Bikram Boote, Weiyan Shi, Diyi Yang, et al. Towards social ai: A survey on understanding social interactions. arXiv preprint arXiv:2409.15316, 2024. 1, 3 +[29] Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Limin Wang, and Yu Qiao. Uniformerv2: Unlocking the potential of image vits for video understanding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1632-1643, 2023. 3, 6, 7 +[30] Yanghao Li, Chao-Yuan Wu, Haoqi Fan, Karttikeya Mangalam, Bo Xiong, Jitendra Malik, and Christoph Feichtenhofer. Mviv2: Improved multiscale vision transformers for classification and detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4804-4814, 2022. 3, 7 +[31] Dan Liu, Libo Zhang, and Yanjun Wu. Ld-congr: A large rgb-d video dataset for long-distance continuous gesture recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3304-3312, 2022. 1, 2, 3 +[32] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava-last: Improved reasoning,OCR,and world knowledge,2024.3,7,8 +[33] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 5 +[34] Xin Liu, Henglin Shi, Haoyu Chen, Zitong Yu, Xiaobai Li, and Guoying Zhao. imigue: An identity-free video dataset for micro-gesture understanding and emotion analysis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10631-10642, 2021. 2 +[35] Zhaoyang Liu, Limin Wang, Wayne Wu, Chen Qian, and Tong Lu. Tam: Temporal adaptive module for video recognition. In Proceedings of the IEEE/CVF international conference on computer vision, pages 13708-13718, 2021. 7 +[36] Ze Liu, Jia Ning, Yue Cao, Yixuan Wei, Zheng Zhang, Stephen Lin, and Han Hu. Video swim transformer. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3202-3211, 2022. 3, 7 + +[37] Joanna Materzynska, Guillaume Berger, Ingo Bax, and Roland Memisevic. The jester dataset: A large-scale video dataset of human gestures. In Proceedings of the IEEE/CVF international conference on computer vision workshops, pages 0–0, 2019. 2 +[38] David McNeill. Hand and mind: What gestures reveal about thought. University of Chicago press, 1992. 3 +[39] Pavlo Molchanov, Xiaodong Yang, Shalini Gupta, Kihwan Kim, Stephen Tyree, and Jan Kautz. Online detection and classification of dynamic hand gestures with recurrent 3d convolutional neural network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4207-4215, 2016. 2 +[40] Fangzhou Mu, Sicheng Mo, and Yin Li. Snag: Scalable and accurate video grounding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18930-18940, 2024. 7 +[41] Yujia Peng, Jiaheng Han, Zhenliang Zhang, Lifeng Fan, Tengyu Liu, Siyuan Qi, Xue Feng, Yuxi Ma, Yizhou Wang, and Song-Chun Zhu. The tong test: Evaluating artificial general intelligence through dynamic embodied physical and social interactions. Engineering, 34:12-22, 2024. 3 +[42] Asanka G Perera, Yee Wei Law, and Javaan Chahl. Uav-gesture: A dataset for uav control and gesture recognition. In Proceedings of the European Conference on Computer Vision (ECCV) Workshops, pages 0–0, 2018. 1, 2 +[43] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 7 +[44] Razieh Rastgoo, Kourosh Kiani, and Sergio Escalera. Sign language recognition: A deep survey. Expert Systems with Applications, 164:113794, 2021. 1 +[45] Fiona Ryan, Ajay Bati, Sangmin Lee, Daniel Bolya, Judy Hoffman, and James M Rehg. Gaze-lle: Gaze target estimation via large-scale learned encoders. 2025. 1 +[46] Harvey Sacks, Emanuel A Schegloff, and Gail Jefferson. A simplest systematics for the organization of turn-taking for conversation. language, 50(4):696-735, 1974. 3 +[47] Yuehao Song, Xinggang Wang, Jingfeng Yao, Wenyu Liu, Jinglin Zhang, and Xiangmin Xu. Vitgaze: gaze following with interaction features in vision transformers. Visual Intelligence, 2(1):1-15, 2024. 1 +[48] Andreas Stolcke, Klaus Ries, Noah Coccaro, Elizabeth Shriberg, Rebecca Bates, Daniel Jurafsky, Paul Taylor, Rachel Martin, Carol Van Ess-Dykema, and Marie Meteor. Dialogue act modeling for automatic tagging and recognition of conversational speech. Computational linguistics, 26(3):339-373, 2000. 3 +[49] Kosei Tanada, Shigemichi Matsuzaki, Kazuhito Tanaka, Shintaro Nakaoka, Yuki Kondo, and Yuto Mori. Pointing gesture understanding via visual prompting and visual question answering for interactive robot navigation. In First Workshop on Vision-Language Models for Navigation and Manipulation at ICRA 2024. 5 + +[50] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In International conference on machine learning, pages 10347-10357. PMLR, 2021. 6 +[51] Du Tran, Heng Wang, Lorenzo Torresani, Jamie Ray, Yann LeCun, and Manohar Paluri. A closer look at spatiotemporal convolutions for action recognition. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 6450-6459, 2018. 6 +[52] Jun Wan, Yibing Zhao, Shuai Zhou, Isabelle Guyon, Sergio Escalera, and Stan Z Li. Chalearn looking at people rgb-d isolated and continuous datasets for gesture recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 56-64, 2016. 2 +[53] Limin Wang, Yuanjun Xiong, Zhe Wang, Yu Qiao, Dahua Lin, Xiaou Tang, and Luc Van Gool. Temporal segment networks: Towards good practices for deep action recognition. In European conference on computer vision, pages 20-36. Springer, 2016. 7 +[54] Limin Wang, Bingkun Huang, Zhiyu Zhao, Zhan Tong, Yinan He, Yi Wang, Yali Wang, and Yu Qiao. Videomae v2: Scaling video masked autoencoders with dual masking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14549-14560, 2023. 3, 6 +[55] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 3, 7, 8 +[56] Yi Wang, Kunchang Li, Xinhao Li, Jiashuo Yu, Yinan He, Guo Chen, Baoqi Pei, Rongkun Zheng, Jilan Xu, Zun Wang, et al. Internvideo2: Scaling video foundation models for multimodal video understanding. arXiv preprint arXiv:2403.15377, 2024. 3 +[57] Diyi Yang, Dirk Hovy, David Jurgens, and Barbara Plank. The call for socially aware language technologies, 2024. 3 +[58] Ceng Zhang, Junxin Chen, Jiatong Li, Yanhong Peng, and Zebing Mao. Large language models for human-robot interaction: A review. Biomimetic Intelligence and Robotics, page 100131, 2023. 1 +[59] Chen-Lin Zhang, Jianxin Wu, and Yin Li. Actionformer: Localizing moments of actions with transformers. In European Conference on Computer Vision, pages 492-510. Springer, 2022. 6 +[60] Yifan Zhang, Congqi Cao, Jian Cheng, and Hanqing Lu. Egogesture: A new dataset and benchmark for egocentric hand gesture recognition. IEEE Transactions on Multimedia, 20(5):1038-1050, 2018. 2 +[61] Y Zhang, B Li, H Liu, Y Lee, L Gui, D Fu, J Feng, Z Liu, and C Li. Llava-last: A strong zero-shot video understanding model. 2024. 8 \ No newline at end of file diff --git a/2025/SocialGesture_ Delving into Multi-person Gesture Understanding/images.zip b/2025/SocialGesture_ Delving into Multi-person Gesture Understanding/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..955ad592041bbe50c76b258b51c9e515a932ce31 --- /dev/null +++ b/2025/SocialGesture_ Delving into Multi-person Gesture Understanding/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2a6d8280e5840622206720b052d6f9a9430e6361447db188e233dbfccc218a3a +size 741122 diff --git a/2025/SocialGesture_ Delving into Multi-person Gesture Understanding/layout.json b/2025/SocialGesture_ Delving into Multi-person Gesture Understanding/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..936933e377b220bab1a94f98bdd396cc97b35059 --- /dev/null +++ b/2025/SocialGesture_ Delving into Multi-person Gesture Understanding/layout.json @@ -0,0 +1,7538 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 103, + 103, + 507, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 103, + 103, + 507, + 121 + ], + "spans": [ + { + "bbox": [ + 103, + 103, + 507, + 121 + ], + "type": "text", + "content": "SocialGesture: Delving into Multi-person Gesture Understanding" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 141, + 553, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 141, + 553, + 185 + ], + "spans": [ + { + "bbox": [ + 61, + 141, + 553, + 185 + ], + "type": "text", + "content": "Xu Cao1, Pranav Virupaksha1, Wenqi Jia1, Bolin Lai2, Fiona Ryan2, Sangmin Lee3†, James M. Rehg1† \n1 University of Illinois Urbana-Champaign \n2 Georgia Institute of Technology \n3Sungkyunkwan University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 187, + 506, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 106, + 187, + 506, + 213 + ], + "spans": [ + { + "bbox": [ + 106, + 187, + 506, + 213 + ], + "type": "text", + "content": "{xucao2,pranavv3,wenqij5,jrehg}@illinois.edu {bolin.lai,fkryan}@gatech.edu \nsangmin.lee@skku.edu" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 240, + 200, + 253 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 240, + 200, + 253 + ], + "spans": [ + { + "bbox": [ + 151, + 240, + 200, + 253 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 266, + 297, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 266, + 297, + 506 + ], + "spans": [ + { + "bbox": [ + 54, + 266, + 297, + 506 + ], + "type": "text", + "content": "Previous research in human gesture recognition has largely overlooked multi-person interactions, which are crucial for understanding the social context of naturally occurring gestures. This limitation in existing datasets presents a significant challenge in aligning human gestures with other modalities like language and speech. To address this issue, we introduce SocialGesture, the first largescale dataset specifically designed for multi-person gesture analysis. SocialGesture features a diverse range of natural scenarios and supports multiple gesture analysis tasks, including video-based recognition and temporal localization, providing a valuable resource for advancing the study of gesture during complex social interactions. Furthermore, we propose a novel visual question answering (VQA) task to benchmark vision language models' (VLMs) performance on social gesture understanding. Our findings highlight several limitations of current gesture recognition models, offering insights into future directions for improvement in this field. SocialGesture is available at hugging-face.co/datasets/IrohXu/SocialGesture." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 529, + 135, + 542 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 529, + 135, + 542 + ], + "spans": [ + { + "bbox": [ + 55, + 529, + 135, + 542 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 550, + 297, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 550, + 297, + 694 + ], + "spans": [ + { + "bbox": [ + 55, + 550, + 297, + 694 + ], + "type": "text", + "content": "Long before the development of spoken language in human history, deictic gestures played an important role in human social communication, serving as one of the earliest communication tools for expressing thoughts, emotions, and intentions. In contemporary communication, both verbal (e.g., speech) and non-verbal (e.g., gesture) cues work in concert. While spoken words express direct meanings, understanding the complete social context often requires more than just the words alone. Gestures play an important role in clarifying intention or points of emphasis, and augment spoken communication with additional social context. Theoretical analysis suggests that both speech and gesture originate from a shared" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 317, + 240, + 435, + 307 + ], + "blocks": [ + { + "bbox": [ + 317, + 240, + 435, + 307 + ], + "lines": [ + { + "bbox": [ + 317, + 240, + 435, + 307 + ], + "spans": [ + { + "bbox": [ + 317, + 240, + 435, + 307 + ], + "type": "image", + "image_path": "1856bd464f77839b2f1d7e4753513dcbf4386790b94ac6d29f511aa919be073c.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 317, + 307, + 435, + 373 + ], + "blocks": [ + { + "bbox": [ + 317, + 307, + 435, + 373 + ], + "lines": [ + { + "bbox": [ + 317, + 307, + 435, + 373 + ], + "spans": [ + { + "bbox": [ + 317, + 307, + 435, + 373 + ], + "type": "image", + "image_path": "776fd41c8ed5847d22db3b0257d12edb541f39528a1ef4bfe0004ba7ba84b47a.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 317, + 373, + 435, + 439 + ], + "blocks": [ + { + "bbox": [ + 317, + 373, + 435, + 439 + ], + "lines": [ + { + "bbox": [ + 317, + 373, + 435, + 439 + ], + "spans": [ + { + "bbox": [ + 317, + 373, + 435, + 439 + ], + "type": "image", + "image_path": "bfe90554e54ee68f1fd15a00c344b3ce865f29b8ff3898d5b05dc4bf7e9235fc.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 442, + 555, + 475 + ], + "lines": [ + { + "bbox": [ + 313, + 442, + 555, + 475 + ], + "spans": [ + { + "bbox": [ + 313, + 442, + 555, + 475 + ], + "type": "text", + "content": "Figure 1. Example frames from six gesture datasets. SocialGesture is the only dataset featuring multi-person interactions and focusing on natural gestures with meaningful social communication." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 435, + 240, + 552, + 307 + ], + "blocks": [ + { + "bbox": [ + 435, + 240, + 552, + 307 + ], + "lines": [ + { + "bbox": [ + 435, + 240, + 552, + 307 + ], + "spans": [ + { + "bbox": [ + 435, + 240, + 552, + 307 + ], + "type": "image", + "image_path": "542625c58de26cd3244a6742f68e03e35583c1db8a10407e4d28f0c709743be2.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 435, + 307, + 552, + 373 + ], + "blocks": [ + { + "bbox": [ + 435, + 307, + 552, + 373 + ], + "lines": [ + { + "bbox": [ + 435, + 307, + 552, + 373 + ], + "spans": [ + { + "bbox": [ + 435, + 307, + 552, + 373 + ], + "type": "image", + "image_path": "a47a871c0a59d2acc6e6348c87e026edc36024c80b7889d804162b202bee50fd.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 435, + 373, + 552, + 439 + ], + "blocks": [ + { + "bbox": [ + 435, + 373, + 552, + 439 + ], + "lines": [ + { + "bbox": [ + 435, + 373, + 552, + 439 + ], + "spans": [ + { + "bbox": [ + 435, + 373, + 552, + 439 + ], + "type": "image", + "image_path": "e0f5acfb9d7626e05c51e7ac231d47011c3ff0b452e7edf8abc28b53d7c0f083.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 495, + 555, + 530 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 495, + 555, + 530 + ], + "spans": [ + { + "bbox": [ + 313, + 495, + 555, + 530 + ], + "type": "text", + "content": "cognitive system, with gestures representing imagery-driven data and speech representing symbolic data [6]. Together, they reflect a unified communicative intent." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 534, + 557, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 557, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 557, + 715 + ], + "type": "text", + "content": "Recent advancements in machine learning have driven research on the understanding of human social communication [27, 58], yet most of the latest social research predominantly focuses on linguistic modalities, often overlooking visual social cues like gestures [28]. While there have been machine learning works on human gestures, most of them have focused on gestures for device manipulation or sign languages [42, 44], and the problem of analyzing natural social gestures between multiple people has not received significant attention. A primary reason for this limitation is the lack of gesture data and annotations that capture multi-person interactions and the use of gestures in realistic social contexts. While recent datasets such as HaGRID [23] and LD-ConGR [31] address high-resolution and long-distance gesture collection (Figure 1), they remain confined to controlled," + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 702, + 144, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 702, + 144, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 702, + 144, + 712 + ], + "type": "text", + "content": "†Corresponding author" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "19509" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 70, + 553, + 220 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 553, + 220 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 553, + 220 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 553, + 220 + ], + "type": "table", + "html": "
DatasetMulti-personVideo ClipsInstancesLabel
CategoryBboxRelationVQA
Jester [37]X148,092148,092✓XXX
NVGesture [39]X1,5321,532✓✓XX
EgoGesture [60]X2,08124,161✓✓XX
ChaLearn ConGD [52]X22,53547,933✓✓XX
IPN Hand [5]X2004,218✓✓XX
UAV gesture [42]X119-✓✓XX
iMiGUE [34]X35918,499✓XXX
LD-ConGR [31]X54244,887✓✓XX
SocialGesture (Ours)✓9,88942,533✓✓✓✓
", + "image_path": "f4ebf2da1f7834205ffdab43b98ac186697e84040c7fce3114c7e77fe673f6e0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 228, + 555, + 262 + ], + "lines": [ + { + "bbox": [ + 55, + 228, + 555, + 262 + ], + "spans": [ + { + "bbox": [ + 55, + 228, + 555, + 262 + ], + "type": "text", + "content": "Table 1. Comparison of SocialGesture with existing video-based gesture datasets. SocialGesture is the first work that addresses multi-person social interaction cases. Our dataset uniquely provides comprehensive annotations including gesture categories, bounding boxes (humans and objects), gesture relations (subject person and target person), and visual question answering pairs." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 277, + 295, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 277, + 295, + 312 + ], + "spans": [ + { + "bbox": [ + 55, + 277, + 295, + 312 + ], + "type": "text", + "content": "single-person settings, and do not contain the spontaneous, natural gestures between multiple people that characterize real-life social interactions." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 312, + 296, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 312, + 296, + 361 + ], + "spans": [ + { + "bbox": [ + 55, + 312, + 296, + 361 + ], + "type": "text", + "content": "To address this limitation, we introduce SocialGesture, a comprehensive dataset designed specifically for understanding gestures in multi-person social interactions. SocialGesture provides three key advantages over existing datasets:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 362, + 295, + 601 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 56, + 362, + 295, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 362, + 295, + 397 + ], + "spans": [ + { + "bbox": [ + 56, + 362, + 295, + 397 + ], + "type": "text", + "content": "1. It focuses on videos with multiple people, making it the first multi-person gesture dataset, in contrast to previous datasets which lack human-to-human interaction." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 399, + 295, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 399, + 295, + 553 + ], + "spans": [ + { + "bbox": [ + 55, + 399, + 295, + 553 + ], + "type": "text", + "content": "2. SocialGesture includes four types of comprehensive annotations: (a) Social gesture categories, identifying the specific class of a gesture (i.e., pointing, showing, giving, reaching) within social interactions; (b) Temporal-spatial annotations, identifying the appearance of social gestures temporally and spatially through time stamps and bounding boxes. (c) Interaction annotations, capturing interpersonal social dynamics that result from gestures; and (d) Visual question answering (VQA) annotations, which link gestures to descriptive content, facilitating the alignment of gestures with contextual meaning. All of these annotations offer a rich resource for developing models for social gesture understanding." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 554, + 295, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 554, + 295, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 554, + 295, + 601 + ], + "type": "text", + "content": "3. SocialGesture comprises a large volume of diverse and naturally occurring social interactions sourced from YouTube and Ego4D [17], and is the largest multi-person gesture dataset available." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 604, + 295, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 604, + 295, + 700 + ], + "spans": [ + { + "bbox": [ + 55, + 604, + 295, + 700 + ], + "type": "text", + "content": "Based on our novel SocialGesture dataset, we introduce diverse gesture-related benchmark tasks from temporal localization and recognition to visual question answering. Our experiments reveal significant challenges in multi-person gesture understanding for current architectures. It is the first step toward building agents that have capabilities to seamlessly understand non-verbal gesture cues in social contexts. The major contributions of this paper are as follows." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 701, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 701, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 701, + 295, + 714 + ], + "type": "text", + "content": "- We introduce SocialGesture, the first dataset focusing on" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 321, + 276, + 555, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 276, + 555, + 324 + ], + "spans": [ + { + "bbox": [ + 321, + 276, + 555, + 324 + ], + "type": "text", + "content": "multi-person gesture interactions in natural social settings. Unlike existing datasets that are limited to single-person or controlled environments, our dataset captures the complexity of real-world gesture-based social communication." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 324, + 555, + 456 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 314, + 324, + 555, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 324, + 555, + 384 + ], + "spans": [ + { + "bbox": [ + 314, + 324, + 555, + 384 + ], + "type": "text", + "content": "- We provide comprehensive multi-level annotations including gesture categories, temporal-spatial localization, interaction dynamics, and visual question-answering pairs. This rich annotation scheme enables the development of models for understanding social gestures." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 384, + 554, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 384, + 554, + 456 + ], + "spans": [ + { + "bbox": [ + 313, + 384, + 554, + 456 + ], + "type": "text", + "content": "- We perform comprehensive experiments across gesture localization and recognition, evaluating a wide breadth of architectures and observe multiple challenging areas for future works. We further benchmark multiple large vision-language models on our proposed diverse Social Visual Question-Answering tasks." + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 470, + 405, + 483 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 470, + 405, + 483 + ], + "spans": [ + { + "bbox": [ + 314, + 470, + 405, + 483 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 314, + 491, + 416, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 491, + 416, + 502 + ], + "spans": [ + { + "bbox": [ + 314, + 491, + 416, + 502 + ], + "type": "text", + "content": "2.1. Gesture Datasets" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 312, + 510, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 510, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 312, + 510, + 556, + 714 + ], + "type": "text", + "content": "As shown in Figure 1 and Table 1, the majority of gesture recognition datasets are video-based, with recent research emphasizing emotional state analysis, increased capture distance for improved model robustness, and egocentric gesture collection. Notable datasets include LD-ConGR, featuring 542 RGB-D videos representing 10 gestures [31], and HaGRID, which focuses on high-resolution image based gesture recognition across 18 hand gesture classes [23]. EgoGesture, with over 2 million frames from 50 subjects, is tailored for egocentric interactions with wearable devices [60], while iMiGUE [34] and SMG focus on micro-gesture recognition [10], containing 359 and 40 videos, respectively. Earlier datasets such as Jester [37], IPN Hand [5], ChaLearn IsoGD [52], and ChaLearn ConGD [52] have also made significant contributions, covering tasks like hand segmentation, and isolated or continuous gesture classification. In contrast, we are the first to provide a large-scale multi-person gesture" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19510" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 71, + 181, + 207 + ], + "blocks": [ + { + "bbox": [ + 59, + 71, + 181, + 207 + ], + "lines": [ + { + "bbox": [ + 59, + 71, + 181, + 207 + ], + "spans": [ + { + "bbox": [ + 59, + 71, + 181, + 207 + ], + "type": "image", + "image_path": "e2c11aca6089c1a682ea958dc15dd43466e01cf881dce0be675f182f86ffab15.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 88, + 209, + 153, + 219 + ], + "lines": [ + { + "bbox": [ + 88, + 209, + 153, + 219 + ], + "spans": [ + { + "bbox": [ + 88, + 209, + 153, + 219 + ], + "type": "text", + "content": "Pointing Gesture" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 182, + 72, + 304, + 207 + ], + "blocks": [ + { + "bbox": [ + 182, + 72, + 304, + 207 + ], + "lines": [ + { + "bbox": [ + 182, + 72, + 304, + 207 + ], + "spans": [ + { + "bbox": [ + 182, + 72, + 304, + 207 + ], + "type": "image", + "image_path": "6da6c39c5e789fca6dc8dd1f7e3f56cd60054a1dbfab86a6eb6db1fb20a21aa2.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 211, + 209, + 276, + 220 + ], + "lines": [ + { + "bbox": [ + 211, + 209, + 276, + 220 + ], + "spans": [ + { + "bbox": [ + 211, + 209, + 276, + 220 + ], + "type": "text", + "content": "Showing Gesture" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 306, + 72, + 427, + 207 + ], + "blocks": [ + { + "bbox": [ + 306, + 72, + 427, + 207 + ], + "lines": [ + { + "bbox": [ + 306, + 72, + 427, + 207 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 427, + 207 + ], + "type": "image", + "image_path": "2dbd399b4634d0acb8e33efacf0866942d8fa13eb263a4cbd5a9ac716cfaa7c5.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 339, + 210, + 394, + 220 + ], + "lines": [ + { + "bbox": [ + 339, + 210, + 394, + 220 + ], + "spans": [ + { + "bbox": [ + 339, + 210, + 394, + 220 + ], + "type": "text", + "content": "Giving Gesture" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 227, + 555, + 262 + ], + "lines": [ + { + "bbox": [ + 55, + 227, + 555, + 262 + ], + "spans": [ + { + "bbox": [ + 55, + 227, + 555, + 262 + ], + "type": "text", + "content": "Figure 2. Examples of the four deictic gesture categories in SocialGesture with subject-target relationships. From left to right: pointing (directing attention), showing (presenting objects), giving (transfer intention), and reaching (acquisition intention) gestures. Red boxes indicate gesture initiators (subjects) and blue notations indicate targets." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 429, + 72, + 551, + 207 + ], + "blocks": [ + { + "bbox": [ + 429, + 72, + 551, + 207 + ], + "lines": [ + { + "bbox": [ + 429, + 72, + 551, + 207 + ], + "spans": [ + { + "bbox": [ + 429, + 72, + 551, + 207 + ], + "type": "image", + "image_path": "ad3a5e03d0df02531104d4520b14cb1928f772deca5bc9e5f100622379bb0212.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 457, + 210, + 523, + 220 + ], + "lines": [ + { + "bbox": [ + 457, + 210, + 523, + 220 + ], + "spans": [ + { + "bbox": [ + 457, + 210, + 523, + 220 + ], + "type": "text", + "content": "Reaching Gesture" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 281, + 295, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 281, + 295, + 319 + ], + "spans": [ + { + "bbox": [ + 55, + 281, + 295, + 319 + ], + "type": "text", + "content": "dataset, and we are also the first to provide comprehensive coverage of the four main classes of deictic gestures(defined in [38]), which play a crucial role in social communication." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 324, + 284, + 337 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 324, + 284, + 337 + ], + "spans": [ + { + "bbox": [ + 55, + 324, + 284, + 337 + ], + "type": "text", + "content": "2.2. Social Understanding with Non-verbal Cues" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 342, + 296, + 546 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 342, + 296, + 546 + ], + "spans": [ + { + "bbox": [ + 55, + 342, + 296, + 546 + ], + "type": "text", + "content": "Understanding social interactions has been long studied in the domain of natural language processing [15, 18, 25, 26, 46, 48]. However, the inherent multimodal nature of social interactions ensures that non-verbal elements (such as gestures, gaze and emotions) play an important role in communication [28], which are understudied in previous research [57]. Recently, several investigations have been conducted on these implicit non-verbal cues. Liu et al. [31] build a dataset to identify the gestures of single person from a long distance. Unfortunately, they only focus on isolated gestures of a single person, and fail to contextualize gesture recognition in a real-world social scenario. Non-verbal social understanding of multiple people demands more investigation [16, 21, 22, 41]. To this end, we make the first attempt to model multi-party social interactions by defining various social gesture understanding tasks, such as social gesture detection, recognition, and VQA, based on our dataset." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 552, + 288, + 565 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 552, + 288, + 565 + ], + "spans": [ + { + "bbox": [ + 55, + 552, + 288, + 565 + ], + "type": "text", + "content": "2.3. Foundation Models for Video Understanding" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 570, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 297, + 714 + ], + "type": "text", + "content": "Video foundation models are designed to tackle a wide range of understanding and analysis tasks, including classification, temporal action localization, question answering, captioning, and more. Early approaches to video understanding relied on CNN blocks as visual encoders for video frames, connected to a prediction head or decoder [9, 13]. With the emergence of transformers, new architectures like MViTv2-B [30], VideoSwin [36], UniFormerV2 [29], and VideoMAE-2 [54] tokenize their visual inputs, enabling unified integration with other modalities [56]. Building on this, recently released video-based VLMs like Qwen2-VL [55], LLaVA-NEXT [32], and InternVL-2.5 [12] leverage pretrained visual" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 281, + 555, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 281, + 555, + 354 + ], + "spans": [ + { + "bbox": [ + 313, + 281, + 555, + 354 + ], + "type": "text", + "content": "encoders to extract rich representations. These models align visual tokens with textual tokens, enabling a deeper multimodal understanding that bridges the gap between video and language tasks. We provide the first investigation of the effectiveness of these models in understanding multi-party social gestures via a VQA task." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 314, + 363, + 400, + 376 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 363, + 400, + 376 + ], + "spans": [ + { + "bbox": [ + 314, + 363, + 400, + 376 + ], + "type": "text", + "content": "3. SocialGesture" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 383, + 389, + 395 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 383, + 389, + 395 + ], + "spans": [ + { + "bbox": [ + 313, + 383, + 389, + 395 + ], + "type": "text", + "content": "3.1. Motivation" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 400, + 556, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 400, + 556, + 544 + ], + "spans": [ + { + "bbox": [ + 313, + 400, + 556, + 544 + ], + "type": "text", + "content": "Social gestures comprise a fundamental set of human behavioral practices that facilitate interpersonal communication through hand and body movements in social contexts, distinct from sign language. According to [38], gestures can be classified into four primary categories: deictic, beat, iconic, and metaphoric. Deictic gestures further subdivide into pointing, showing, giving, and reaching. Beat gestures are rhythmic movements that emphasize certain words and phrases during speech. Iconic gestures visually represent concrete objects or actions (e.g., making a circle with fingers), while metaphoric gestures symbolize abstract concepts (e.g., making a circular motion to indicate repetition)." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 544, + 556, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 544, + 556, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 544, + 556, + 640 + ], + "type": "text", + "content": "Among these categories, deictic gestures deserve particular attention due to their prevalence and significance in social interactions. While iconic and metaphoric gestures appear in specific contexts and beat gestures primarily serve prosodic functions, deictic gestures play a crucial role in establishing joint attention and facilitating object-mediated social interaction. Therefore, our dataset focuses on four fundamental deictic gestures (see Figure 2):" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "content": "- Pointing Gesture involves directing attention to a specific entity in the social scene through finger extension. While commonly executed with an extended index finger and arm, these specific forms are not mandatory requirements. The crucial identifying factor is the communicative intent to guide others' attention to a particular target, whether" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19511" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 70, + 553, + 175 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 553, + 175 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 553, + 175 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 553, + 175 + ], + "type": "table", + "html": "
Video ContentSourceProportionTotal Length# of PeopleDescription
Group social gamesYouTube44.51%896 min>=3A group of people playing social games such as one night werewolf
Ego4D21.91%441 min>=4
Variety entertainmentYouTube22.31%449 min>=2People performing activities such as pranks, challenges, and interviews
Educational playYouTube2.53%51 min>=2Children engaged in educational activities with adult guidance
Product reviewsYouTube2.53%51 min>=2People recommending products such as advent calendars
Party & dinnerYouTube3.63%73 min>=7A group of people having party & dinner together
Group cookingYouTube2.58%52 min>=3People engaged in cooking activities together
", + "image_path": "2342445800ffc3ab4a8e2b36a99204dd10ae4be9fd29cb41c46a0cd43dc0fdb2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 175, + 539, + 186 + ], + "lines": [ + { + "bbox": [ + 67, + 175, + 539, + 186 + ], + "spans": [ + { + "bbox": [ + 67, + 175, + 539, + 186 + ], + "type": "text", + "content": "Table 2. SocialGesture contains diverse videos including different multi-person social interactions from YouTube and Ego4D [17]." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 63, + 197, + 295, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 197, + 295, + 232 + ], + "spans": [ + { + "bbox": [ + 63, + 197, + 295, + 232 + ], + "type": "text", + "content": "person or object. This distinguishes pointing from incidental hand movements or other gesture types, particularly reaching gestures where the primary intent differs." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 233, + 295, + 555 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 55, + 233, + 295, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 233, + 295, + 339 + ], + "spans": [ + { + "bbox": [ + 55, + 233, + 295, + 339 + ], + "type": "text", + "content": "- **Showing Gesture** involves presenting an object for others to see it by manipulating or orienting it towards observers. Unlike pointing, the object itself is the focus of attention, and it does not need to be fully supported by the subject, such as when tilting or sliding an object to make its contents visible. The key criterion for a showing gesture is that the object is being presented for visual inspection, regardless of whether the target person actually looks at it or where the subject is directing their gaze." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 340, + 295, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 340, + 295, + 422 + ], + "spans": [ + { + "bbox": [ + 55, + 340, + 295, + 422 + ], + "type": "text", + "content": "- Giving Gesture involves manipulating an object with the intention of transferring it to another person's possession. It differs from pointing because the object originates with the subject, and from showing because the subject is inviting the other person to take possession of the object. The key is the subject's intention to transfer the object, regardless of whether the recipient actually takes it." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 424, + 295, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 424, + 295, + 555 + ], + "spans": [ + { + "bbox": [ + 55, + 424, + 295, + 555 + ], + "type": "text", + "content": "- Reaching Gesture manifests as a hand extension toward an object, expressing desire for possession or requesting transfer. It is characterized by full finger extension and often includes pre-grasp configuration and associated body movements like leaning forward. Reaching differs from showing and giving because the object is not in the subject's possession, and it differs from pointing because the subject is expressing a desire to possess the object, not just to draw attention to it. The key aspect of reaching is the intent to obtain the object, regardless of whether the subject can physically retrieve it." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 564, + 157, + 577 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 564, + 157, + 577 + ], + "spans": [ + { + "bbox": [ + 55, + 564, + 157, + 577 + ], + "type": "text", + "content": "3.2. Data Acquisition" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 582, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 582, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 54, + 582, + 295, + 713 + ], + "type": "text", + "content": "Data collection. We collect raw video data for SocialGesture from various sources, including YouTube channels and Ego4D. Table 2 represents the overall composition of the dataset. Video selection is based on four key criteria: (1) Video quality – we manually check each video to ensure high resolution; (2) Number of people – each video is carefully selected to include between two and ten people, ensuring clear visibility of gestures; (3) Video length – the total length of selected videos ranges from 2 to 30 minutes, with each gesture lasting between 1 to 15 seconds; and (4) Scene diversity – to enhance generalization, we select videos featuring" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 198, + 446, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 198, + 446, + 209 + ], + "spans": [ + { + "bbox": [ + 313, + 198, + 446, + 209 + ], + "type": "text", + "content": "diverse races, genders, and ages." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 213, + 554, + 273 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 213, + 554, + 273 + ], + "spans": [ + { + "bbox": [ + 313, + 213, + 554, + 273 + ], + "type": "text", + "content": "Pre-processing. For videos longer than 10 minutes, we cut them into video clips of 5 minutes in length. All videos are then converted to " + }, + { + "bbox": [ + 313, + 213, + 554, + 273 + ], + "type": "inline_equation", + "content": "720\\mathrm{p}" + }, + { + "bbox": [ + 313, + 213, + 554, + 273 + ], + "type": "inline_equation", + "content": "(1280\\times 720)" + }, + { + "bbox": [ + 313, + 213, + 554, + 273 + ], + "type": "text", + "content": " with 30 FPS in the initial curation and are then mapped to " + }, + { + "bbox": [ + 313, + 213, + 554, + 273 + ], + "type": "inline_equation", + "content": "360\\mathrm{p}" + }, + { + "bbox": [ + 313, + 213, + 554, + 273 + ], + "type": "inline_equation", + "content": "(640\\times 360)" + }, + { + "bbox": [ + 313, + 213, + 554, + 273 + ], + "type": "text", + "content": " with 5 FPS for further annotation." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 277, + 555, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 277, + 555, + 564 + ], + "spans": [ + { + "bbox": [ + 313, + 277, + 555, + 564 + ], + "type": "text", + "content": "Gesture annotation. Each gesture in the SocialGesture dataset is defined as a sequence of video frames where an individual performs an arm or hand movement corresponding to one of four specified deictic gestures: pointing, showing, giving, and reaching. The annotations capture both temporal and spatial aspects of these social interactions through a comprehensive multi-level annotation framework. For temporal annotation, we identify the complete sequence of frames containing a gesture, from initial movement through completion. Within this sequence, we designate a key frame that best captures the defining moment of the gesture - typically when the gesture reaches its most distinctive configuration. This key frame serves as an anchor point for additional annotations. Spatial annotation includes precise bounding box coordinates [x1, y1, x2, y2] for both the gesture initiator and the target. The initiator bounding box encompasses the person performing the gesture, while the target bounding box identifies either the person or object toward which the gesture is directed. These spatial annotations enable analysis of the geometric relationships between interaction participants. Our annotation framework also includes natural language descriptions of each gesture's social context, capturing the broader interaction dynamics beyond pure spatiotemporal coordinates." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 573, + 421, + 584 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 573, + 421, + 584 + ], + "spans": [ + { + "bbox": [ + 313, + 573, + 421, + 584 + ], + "type": "text", + "content": "3.3. Benchmark Tasks" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 590, + 503, + 602 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 590, + 503, + 602 + ], + "spans": [ + { + "bbox": [ + 313, + 590, + 503, + 602 + ], + "type": "text", + "content": "3.3.1. Social Gesture Temporal Localization" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 605, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 554, + 713 + ], + "type": "text", + "content": "Temporal action localization is a challenging task that requires not only identifying the temporal intervals of all detected gesture instances but also estimating corresponding confidence values. Unlike standard classification tasks, this process involves an end-to-end approach in which classification is integrated within temporal localization, adding considerable complexity. For example, even if a gesture is accurately localized in time, it is treated as a false detection if assigned an incorrect class label." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19512" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 72, + 292, + 213 + ], + "blocks": [ + { + "bbox": [ + 59, + 72, + 292, + 213 + ], + "lines": [ + { + "bbox": [ + 59, + 72, + 292, + 213 + ], + "spans": [ + { + "bbox": [ + 59, + 72, + 292, + 213 + ], + "type": "image", + "image_path": "f5cad51e6459531a5fb52b20692e9de7501cfd29dfaacca3ef5a759496404e2d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 61, + 218, + 290, + 229 + ], + "lines": [ + { + "bbox": [ + 61, + 218, + 290, + 229 + ], + "spans": [ + { + "bbox": [ + 61, + 218, + 290, + 229 + ], + "type": "text", + "content": "Figure 3. Temporal gesture localization task for social gestures" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 237, + 296, + 356 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 237, + 296, + 356 + ], + "spans": [ + { + "bbox": [ + 55, + 237, + 296, + 356 + ], + "type": "text", + "content": "Figure 3 illustrates a general approach to temporal action localization. Each video is initially divided into multiple overlapping video clips, denoted as " + }, + { + "bbox": [ + 55, + 237, + 296, + 356 + ], + "type": "inline_equation", + "content": "X = [X_0, X_1, \\ldots, X_K]" + }, + { + "bbox": [ + 55, + 237, + 296, + 356 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 237, + 296, + 356 + ], + "type": "inline_equation", + "content": "X_k \\in \\mathbb{R}^{T \\times H \\times W \\times C}" + }, + { + "bbox": [ + 55, + 237, + 296, + 356 + ], + "type": "text", + "content": ". Feature extractors such as I3D or VideoMAE are then employed to extract multiscale feature representations, denoted as " + }, + { + "bbox": [ + 55, + 237, + 296, + 356 + ], + "type": "inline_equation", + "content": "Z_k \\in \\mathbb{R}^D" + }, + { + "bbox": [ + 55, + 237, + 296, + 356 + ], + "type": "text", + "content": ", for each video clip. These features are subsequently concatenated into a single representation, " + }, + { + "bbox": [ + 55, + 237, + 296, + 356 + ], + "type": "inline_equation", + "content": "Z \\in \\mathbb{R}^{T \\times D}" + }, + { + "bbox": [ + 55, + 237, + 296, + 356 + ], + "type": "text", + "content": ". The objective of the temporal action localization model is to generate sequence labels for each video clip." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 357, + 296, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 357, + 296, + 441 + ], + "spans": [ + { + "bbox": [ + 55, + 357, + 296, + 441 + ], + "type": "text", + "content": "This task is also benchmarked by several well-known datasets, including THUMOS14 [20] and ActivityNet-1.3 [8]. However, our task presents additional challenges. In SocialGesture videos, multiple individuals are often present simultaneously, and gestures tend to be small, subtle portions of larger actions, making it uniquely challenging to detect and classify gestures accurately." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 445, + 200, + 457 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 445, + 200, + 457 + ], + "spans": [ + { + "bbox": [ + 55, + 445, + 200, + 457 + ], + "type": "text", + "content": "3.3.2. Social Gesture Recognition" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 460, + 296, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 460, + 296, + 663 + ], + "spans": [ + { + "bbox": [ + 55, + 460, + 296, + 663 + ], + "type": "text", + "content": "Given the complexity of social gesture localization, we also divide long videos into shorter clips of 2-5 seconds and introduce the social gesture recognition task. This task can be approached in multiple ways. The first approach is isolated gesture recognition, where a long video is segmented into discrete gesture clips. Video action recognition models are then trained to classify the gestures in each clip independently. Alternatively, gestures can be identified continuously throughout the entire video, from beginning to end. Although most action localization methods we discussed in the previous section can typically handle this task, most existing video feature extraction pre-processing methods struggle with consistent multi-person action localization. Therefore, sliding windows remain a common strategy for this purpose: the video is processed in overlapping segments of a predefined window size, with each window analyzed by the model to calculate gesture confidence scores." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 664, + 293, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 664, + 293, + 676 + ], + "spans": [ + { + "bbox": [ + 67, + 664, + 293, + 676 + ], + "type": "text", + "content": "We divide recognition into two subtasks (See Figure 4):" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "- Task 2-1: Gesture vs Non-gesture Classification This task aims to distinguish between social gesture and non-gesture video clips. The input is a short video clip " + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "inline_equation", + "content": "X \\in" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 318, + 72, + 550, + 211 + ], + "blocks": [ + { + "bbox": [ + 318, + 72, + 550, + 211 + ], + "lines": [ + { + "bbox": [ + 318, + 72, + 550, + 211 + ], + "spans": [ + { + "bbox": [ + 318, + 72, + 550, + 211 + ], + "type": "image", + "image_path": "549ea7e37bcb7955888324f586c1e2287b642c527a165124155c6b20ee6333bd.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 335, + 215, + 533, + 226 + ], + "lines": [ + { + "bbox": [ + 335, + 215, + 533, + 226 + ], + "spans": [ + { + "bbox": [ + 335, + 215, + 533, + 226 + ], + "type": "text", + "content": "Figure 4. Gesture recognition tasks for social gestures" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 322, + 236, + 553, + 261 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 236, + 553, + 261 + ], + "spans": [ + { + "bbox": [ + 322, + 236, + 553, + 261 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{T\\times H\\times W\\times C}" + }, + { + "bbox": [ + 322, + 236, + 553, + 261 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 322, + 236, + 553, + 261 + ], + "type": "inline_equation", + "content": "T < 16" + }, + { + "bbox": [ + 322, + 236, + 553, + 261 + ], + "type": "text", + "content": ", and the output is a label " + }, + { + "bbox": [ + 322, + 236, + 553, + 261 + ], + "type": "inline_equation", + "content": "\\hat{Y}" + }, + { + "bbox": [ + 322, + 236, + 553, + 261 + ], + "type": "text", + "content": " indicating whether a gesture is present." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 262, + 555, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 262, + 555, + 380 + ], + "spans": [ + { + "bbox": [ + 314, + 262, + 555, + 380 + ], + "type": "text", + "content": "- Task 2-2: Gesture Type Classification. This task focuses on classifying video clips containing one of four social gestures: pointing, showing, giving, or reaching. The input and output structures are consistent with those used in the binary classification task. We further explore whether using cropped bounding boxes for each individual within a scene improves the accuracy of social gesture classification. The input is a short video clip cropped by the maximum region of the initiator of a social gesture and the output is a label " + }, + { + "bbox": [ + 314, + 262, + 555, + 380 + ], + "type": "inline_equation", + "content": "\\hat{Y}" + }, + { + "bbox": [ + 314, + 262, + 555, + 380 + ], + "type": "text", + "content": " indicating four types of social gesture." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 383, + 554, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 383, + 554, + 441 + ], + "spans": [ + { + "bbox": [ + 313, + 383, + 554, + 441 + ], + "type": "text", + "content": "For each task, we also strive to sample the training set to achieve a balanced label distribution. This adjustment is necessary because pointing gestures are significantly more frequent than showing, giving, and reaching in real-world communication contexts." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 447, + 523, + 459 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 447, + 523, + 459 + ], + "spans": [ + { + "bbox": [ + 313, + 447, + 523, + 459 + ], + "type": "text", + "content": "3.3.3. Social Gesture Visual Question Answering" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 462, + 555, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 462, + 555, + 581 + ], + "spans": [ + { + "bbox": [ + 313, + 462, + 555, + 581 + ], + "type": "text", + "content": "Prior to our work, there has been very limited research on designing video captioning and Visual Question Answering (VQA) specifically for human gesture datasets [49]. This gap may be due to the lack of multi-person interactions in existing video datasets, reducing the necessity for enhanced alignment between the gesture and language modalities. To fill this gap, we introduce the first VQA task focused on human gestures (Figure 5). Our VQA can be divided into three subtasks: (1) Global Perception, (2) Gesture Understanding, and (3) Gesture Localization." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 582, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 556, + 714 + ], + "type": "text", + "content": "The Global Perception task serves as a foundational test to assess whether VLMs can comprehend basic events in short video clips, including scene descriptions and counting the number of humans present. We generate data for this task through a semi-automated QA generation pipeline based on GPT-4o [2], similar to the instruction-tuning data pair generation in LLaVA [33]. Initially, we manually create short descriptions for each video series, including the social background. We prompt these descriptions with the key frame of the video clip into the GPT-4o pipeline to generate QA pairs for human counting and scene description. Each QA" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19513" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 73, + 293, + 459 + ], + "blocks": [ + { + "bbox": [ + 57, + 73, + 293, + 459 + ], + "lines": [ + { + "bbox": [ + 57, + 73, + 293, + 459 + ], + "spans": [ + { + "bbox": [ + 57, + 73, + 293, + 459 + ], + "type": "image", + "image_path": "81c2e9da614e48b861333c9ea0a4b4d3df052203e4995a0b596681717326c0e0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 462, + 295, + 508 + ], + "lines": [ + { + "bbox": [ + 55, + 462, + 295, + 508 + ], + "spans": [ + { + "bbox": [ + 55, + 462, + 295, + 508 + ], + "type": "text", + "content": "Figure 5. The question-answer pairs of SocialGesture. We omit the options of each question in the figure. The bounding box defined by [top-left x, top-left y, bottom-right x, bottom-right y]). The definition will be provided together with the system prompts." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 521, + 296, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 296, + 567 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 296, + 567 + ], + "type": "text", + "content": "pair is then reviewed by a human annotator. Figure 5 shows the example of global perception (Task 3-1). Human counting is used as one VQA subtask in the VLM benchmarking experiments." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 569, + 296, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 569, + 296, + 629 + ], + "spans": [ + { + "bbox": [ + 55, + 569, + 296, + 629 + ], + "type": "text", + "content": "For Gesture Understanding, we utilize gesture classification annotations to create questions related to social gesture classification (Figure 5, Task 3-2). This results in two types of questions: (1) detecting whether a social gesture occurs, and (2) classifying between the four types of social gestures." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "text", + "content": "Gesture Localization leverages bounding boxes from our annotations to generate questions that test the model's spatial and temporal comprehension. These questions include providing the bounding box localization of the initiator, then determining whether the gesture's target is a human, and locating the target. We present an example of gesture localization (Task 3-3) in Figure 5. These questions are designed" + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 317, + 71, + 552, + 125 + ], + "blocks": [ + { + "bbox": [ + 317, + 71, + 552, + 125 + ], + "lines": [ + { + "bbox": [ + 317, + 71, + 552, + 125 + ], + "spans": [ + { + "bbox": [ + 317, + 71, + 552, + 125 + ], + "type": "table", + "html": "
Feature0.30.40.50.60.7Avg mAP
I3D [9]24.8516.319.312.220.9610.73
R(2+1)D [51]14.3810.257.232.811.777.29
VideoMAEV2 [54]27.2325.0513.335.282.7614.73
", + "image_path": "984e3790fef70823c1417b93369fe86880b255b8da92f1aec440138db1d1a64e.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 317, + 165, + 552, + 218 + ], + "blocks": [ + { + "bbox": [ + 313, + 132, + 555, + 154 + ], + "lines": [ + { + "bbox": [ + 313, + 132, + 555, + 154 + ], + "spans": [ + { + "bbox": [ + 313, + 132, + 555, + 154 + ], + "type": "text", + "content": "Table 3. Temporal action localization for four social gestures by different feature extractors with ActionFormer [59]." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 165, + 552, + 218 + ], + "lines": [ + { + "bbox": [ + 317, + 165, + 552, + 218 + ], + "spans": [ + { + "bbox": [ + 317, + 165, + 552, + 218 + ], + "type": "table", + "html": "
Stride0.30.40.50.60.7Avg mAP
1611.098.605.522.671.825.94
824.8516.319.312.220.9610.73
431.6429.1319.3013.772.1119.19
", + "image_path": "7541b89aa7c658453ab5720619660393d8ec0386d320eb34f98c75b0123143bb.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 224, + 554, + 236 + ], + "lines": [ + { + "bbox": [ + 314, + 224, + 554, + 236 + ], + "spans": [ + { + "bbox": [ + 314, + 224, + 554, + 236 + ], + "type": "text", + "content": "Table 4. Explore the influence of stride for I3D and ActionFormer." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 255, + 554, + 279 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 255, + 554, + 279 + ], + "spans": [ + { + "bbox": [ + 313, + 255, + 554, + 279 + ], + "type": "text", + "content": "to challenge temporal-spatial reasoning in a multi-person setting, requiring analysis to produce the correct responses." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 290, + 395, + 304 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 290, + 395, + 304 + ], + "spans": [ + { + "bbox": [ + 313, + 290, + 395, + 304 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 310, + 430, + 323 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 310, + 430, + 323 + ], + "spans": [ + { + "bbox": [ + 313, + 310, + 430, + 323 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 328, + 555, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 328, + 555, + 472 + ], + "spans": [ + { + "bbox": [ + 313, + 328, + 555, + 472 + ], + "type": "text", + "content": "We split the raw videos (long video) into training set (293 videos) and test set (79 videos), and then cut each long video into video clips. The 9,889 video clips used in video action recognition and social gesture VQA tasks come from the same train-test split in the raw videos. We also cut another 4,304 video clips from the raw videos as the non-gesture class. All experiments are conducted with 2 NVIDIA L40S GPUs and 2 H100 GPUs to make sure the batch size is 16 and learning rate is 5e-4 for all models. For all methods in Table 3, 5, 6, and 7, we apply the same data augmentation strategies used in [29, 50]. To solve the class imbalanced issue, we also adjust the size of training set for each task." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 479, + 513, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 479, + 513, + 491 + ], + "spans": [ + { + "bbox": [ + 313, + 479, + 513, + 491 + ], + "type": "text", + "content": "4.2. Social Gesture Temporal Localization" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 494, + 555, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 494, + 555, + 567 + ], + "spans": [ + { + "bbox": [ + 313, + 494, + 555, + 567 + ], + "type": "text", + "content": "Baselines. We used two-stream I3D [9], R(2+1)D [51], and VideoMAEV2 [54] with the same sliding window of 16 frames and different stride sizes for feature extraction. Then, these video features are used as input for ActionFormer [59] during model training. We used mAP@[0.3:0.1:0.7] as the main evaluation metric and also reported the average mAP." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 570, + 556, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 556, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 556, + 712 + ], + "type": "text", + "content": "Results and Findings. We observe that almost all feature extractors underperform on this task in Table 3. Notably, despite the strong performance of features from VideoMAE V2 [54] in other action localization datasets, such as THU-MOS 2014 [20] and ActivityNet [8], they only achieve an average mAP of 14.73 for social gesture recognition. To investigate the effect of the stride size in the sliding window on the average mAP, we conducted an ablation study in Table 4 using different stride sizes of the I3D feature in SocialGesture. We found that reducing the stride from 16 to 4 led to an increase in average mAP; however, the overall results remained insufficient. This shortfall can be attributed to the" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19514" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 70, + 293, + 190 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 293, + 190 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 293, + 190 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 293, + 190 + ], + "type": "table", + "html": "
ModelPretrainParamAcc (%)
TSN-R50 [53]Kinetics-400 [24]24M78.77
TANet-R50 [35]Kinetics-400 [24]26M71.22
SlowFast-R50 [13]Kinetics-400 [24]35M80.82
SlowFast-R101 [13]Kinetics-400 [24]63M79.59
TimeSformer-L [7]Kinetics-400 [24]121M78.71
MViTv2-B [30]Kinetics-400 [24]51M83.29
VideoSwin-B [36]Kinetics-400 [24]88M81.70
VideoSwin-L [36]Kinetics-400 [24]200M83.44
UniFormerV2-B/16 [29]CLIP [43]115M84.43
UniFormerV2-L/14 [29]CLIP [43]354M80.93
", + "image_path": "123b3a7b1fc9ea2c034a62e22b8cc501705532bea7eeb66e927951b600ea6a61.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 237, + 296, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 237, + 296, + 274 + ], + "spans": [ + { + "bbox": [ + 55, + 237, + 296, + 274 + ], + "type": "text", + "content": "fact that all feature extractors were pretrained on datasets that lack multi-person interactions, causing the features to be poorly aligned with the specific requirements of our task." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 283, + 205, + 296 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 283, + 205, + 296 + ], + "spans": [ + { + "bbox": [ + 55, + 283, + 205, + 296 + ], + "type": "text", + "content": "4.3. Social Gesture Recognition" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 301, + 296, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 301, + 296, + 410 + ], + "spans": [ + { + "bbox": [ + 55, + 301, + 296, + 410 + ], + "type": "text", + "content": "Baselines. Our baselines for social gesture recognition includes different CNN-based video models such as TSN [53], TANet [35], SlowFast [13], and Vision Transformer based models such as TimeSformer [7], MViTv2 [30], VideoSwin [36], UniFormerV2 [29]. All these models are pretrained with Kinetics-400 [24] dataset or using CLIP encoder [43]. We use Accuracy for gesture and non-gesture binary classification in Table 5 and Top 1 Accuracy for four social gesture classification in Table 6, and 7." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 413, + 296, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 413, + 296, + 509 + ], + "spans": [ + { + "bbox": [ + 55, + 413, + 296, + 509 + ], + "type": "text", + "content": "Results and Findings. The results in Table 5 indicate that most existing action recognition models can differentiate social gestures and non-gesture from short video clips. Despite fine-tuning these models using over 10,000 video clips for the binary classification task, their performance remains suboptimal. The SOTA baseline, UniFormerV2-B/16, achieves an accuracy of only " + }, + { + "bbox": [ + 55, + 413, + 296, + 509 + ], + "type": "inline_equation", + "content": "84.43\\%" + }, + { + "bbox": [ + 55, + 413, + 296, + 509 + ], + "type": "text", + "content": ". Notably, Transformer-based models generally outperform CNN-based models." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 510, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 296, + 713 + ], + "type": "text", + "content": "Table 6 presents results for the four-class social gesture classification task, revealing that all models struggle to effectively classify these gestures—despite the simplicity of the task for human annotators. The highest performance comes from VideoSwim-L, achieving only " + }, + { + "bbox": [ + 55, + 510, + 296, + 713 + ], + "type": "inline_equation", + "content": "56.18\\%" + }, + { + "bbox": [ + 55, + 510, + 296, + 713 + ], + "type": "text", + "content": ". Given that individual gestures can be challenging to recognize in multi-person environments, we conducted further experiments, as shown in Table 7. Specifically, we extracted the region corresponding to each individual using ground truth bounding boxes and fine-tuned the same models with these per-person inputs. UniFormerV2-B/16 achieves the best performance of " + }, + { + "bbox": [ + 55, + 510, + 296, + 713 + ], + "type": "inline_equation", + "content": "64.72\\%" + }, + { + "bbox": [ + 55, + 510, + 296, + 713 + ], + "type": "text", + "content": " top 1 accuracy. This may be because social gestures inherently involve fine-grained and subtle movements. Consequently, recognizing such gestures is difficult from both the extracted per-person region and the full-frame perspective. Addressing the challenge of social gesture understanding requires advanced models capable of" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 318, + 70, + 553, + 182 + ], + "blocks": [ + { + "bbox": [ + 55, + 197, + 295, + 219 + ], + "lines": [ + { + "bbox": [ + 55, + 197, + 295, + 219 + ], + "spans": [ + { + "bbox": [ + 55, + 197, + 295, + 219 + ], + "type": "text", + "content": "Table 5. Experiments on binary classification for social gesture and non-gesture." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 318, + 70, + 553, + 182 + ], + "lines": [ + { + "bbox": [ + 318, + 70, + 553, + 182 + ], + "spans": [ + { + "bbox": [ + 318, + 70, + 553, + 182 + ], + "type": "table", + "html": "
ModelPretrainParamTop1 Acc (%)
TSN-R50 [53]Kinetics-400 [24]24M54.83
TANet-R50 [35]Kinetics-400 [24]26M45.39
SlowFast-R50 [13]Kinetics-400 [24]35M45.17
SlowFast-R101 [13]Kinetics-400 [24]63M46.07
TimeSformer-L [7]Kinetics-400 [24]121M53.03
MViTv2-B [30]Kinetics-400 [24]51M37.98
VideoSwin-B [36]Kinetics-400 [24]88M53.93
VideoSwin-L [36]Kinetics-400 [24]200M56.18
UniFormerV2-B/16 [29]CLIP [43]115M55.51
UniFormerV2-L/14 [29]CLIP [43]354M50.34
", + "image_path": "3a27bb6c0aa43094887e6362aeab25478319590207cbe38ef0853bb1292cf073.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 317, + 212, + 553, + 323 + ], + "blocks": [ + { + "bbox": [ + 320, + 188, + 547, + 199 + ], + "lines": [ + { + "bbox": [ + 320, + 188, + 547, + 199 + ], + "spans": [ + { + "bbox": [ + 320, + 188, + 547, + 199 + ], + "type": "text", + "content": "Table 6. Experiments on classification for four social gestures." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 212, + 553, + 323 + ], + "lines": [ + { + "bbox": [ + 317, + 212, + 553, + 323 + ], + "spans": [ + { + "bbox": [ + 317, + 212, + 553, + 323 + ], + "type": "table", + "html": "
ModelPretrainParamTop1 Acc (%)
TSN-R50 [53]Kinetics-400 [24]24M58.43
TANet-R50 [35]Kinetics-400 [24]26M46.29
SlowFast-R50 [13]Kinetics-400 [24]35M45.17
SlowFast-R101 [13]Kinetics-400 [24]63M42.92
TimeSformer-L [7]Kinetics-400 [24]121M63.60
MViTv2-B [30]Kinetics-400 [24]51M35.28
VideoSwin-B [36]Kinetics-400 [24]88M60.00
VideoSwin-L [36]Kinetics-400 [24]200M63.60
UniFormerV2-B/16 [29]CLIP [43]115M64.72
UniFormerV2-L/14 [29]CLIP [43]354M60.45
", + "image_path": "34875ebc9a23f8bbf991e106276bb659f293198bb30e1e80b9dc5049a63c2050.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 330, + 555, + 353 + ], + "lines": [ + { + "bbox": [ + 313, + 330, + 555, + 353 + ], + "spans": [ + { + "bbox": [ + 313, + 330, + 555, + 353 + ], + "type": "text", + "content": "Table 7. Experiments on classification for four social gestures after extract each subject person's bounding box." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 366, + 554, + 391 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 366, + 554, + 391 + ], + "spans": [ + { + "bbox": [ + 313, + 366, + 554, + 391 + ], + "type": "text", + "content": "disentangling the relationship between the subject and the target person." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 400, + 429, + 414 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 400, + 429, + 414 + ], + "spans": [ + { + "bbox": [ + 313, + 400, + 429, + 414 + ], + "type": "text", + "content": "4.4. SocialGesture VQA" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 418, + 555, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 418, + 555, + 552 + ], + "spans": [ + { + "bbox": [ + 313, + 418, + 555, + 552 + ], + "type": "text", + "content": "From the results of previous tasks, we found that distinguishing social gestures is still very challenging for current SOTA methods in gesture localization and recognition. This difficulty highlights the unique nature of social gesture, where models need to have a unified view of multi-person social interactions and can distinguish details that are hard to detect. This is something traditional video analysis models still struggle with [1, 14]. Inspired by recent progress in multimodal video grounding and video-based VQA [11, 40], we aim to leverage the reasoning capabilities of the latest VLMs and evaluate their performance on this complex task." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 554, + 556, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 554, + 556, + 673 + ], + "spans": [ + { + "bbox": [ + 313, + 554, + 556, + 673 + ], + "type": "text", + "content": "Baselines. To assess performance on the SocialGesture VQA benchmark, we chose several SOTA publicly available VLMs, including Qwen2-VL [55], Qwen2.5-VL [4], InternVL-2.5 [12], and LLAVA-NEXT-Video (LLaVA 1.6) [32]. Additionally, we included closed-source multimodal LLMs such as GPT-4o-mini [2] and Claude-3.7-sonnet [3] for a broader comparison. Note that GPT-4o [2] was excluded, as it was used to generate some of the VQA instruction-following training data, which could lead to biased results." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "text", + "content": "Evaluation Metrics. Since all VQA pairs can be transformed into a multi-option output, we use rule-based accuracy metrics to evaluate the VQA tasks. Each model was" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19515" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 70, + 553, + 199 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 553, + 199 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 553, + 199 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 553, + 199 + ], + "type": "table", + "html": "
ModelParamGlobal Perception (%)Gesture Understanding (%)Gesture Localization (%)
HumanCount@AccGestureDet@AccGestureClass@AccTargetLoc@AccTargetClass@Acc
Random Select--50.0025.0018.9050.00
InternVL-2.5 [12]2B22.3763.3474.4513.3339.79
InternVL-2.5 [12]8B35.5873.9481.8017.3857.28
LLaVA-NeXT-Video [32, 61]7B30.6371.9234.085.3959.90
Qwen2-VL [55]7B60.1874.1567.189.5859.19
Qwen2.5-VL [4]72B69.9775.6954.8229.3761.38
Claude 3.7 Sonnet [3]-63.2769.6461.7325.4864.20
GPT-4o-mini [2]-53.4073.6768.5221.2965.12
", + "image_path": "580fb08cba3355866722f22fe08c9665689a80566835d60bb4c215c51dd9fd50.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 58, + 221, + 552, + 291 + ], + "blocks": [ + { + "bbox": [ + 61, + 205, + 547, + 217 + ], + "lines": [ + { + "bbox": [ + 61, + 205, + 547, + 217 + ], + "spans": [ + { + "bbox": [ + 61, + 205, + 547, + 217 + ], + "type": "text", + "content": "Table 8. Experiments on zero-shot performance of SOTA VLMs, including closed sourced multimodal LLMs for Social Gesture VQA" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 221, + 552, + 291 + ], + "lines": [ + { + "bbox": [ + 58, + 221, + 552, + 291 + ], + "spans": [ + { + "bbox": [ + 58, + 221, + 552, + 291 + ], + "type": "table", + "html": "
ModelInputGlobal Perception (%)Gesture Understanding (%)Gesture Localization (%)
HumanCount@AccGestureDet@AccGestureClass@AccTargetLoc@AccTargetClass@Acc
Qwen2-VL-7Bvideo60.1874.1567.189.5859.19
Qwen2-VL-7B + LoRA [19] SFTvideo82.5533.0784.5249.2141.19
Qwen2-VL-7B + Full SFTvideo83.6433.0781.1351.2538.51
", + "image_path": "4fccc51427ca347c01058017b1771fedf58567ff5375bb99978880aa6e04f155.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 70, + 297, + 538, + 308 + ], + "lines": [ + { + "bbox": [ + 70, + 297, + 538, + 308 + ], + "spans": [ + { + "bbox": [ + 70, + 297, + 538, + 308 + ], + "type": "text", + "content": "Table 9. Experiments on SFT performance of SOTA VLMs, including closed sourced multimodal LLMs for Social Gesture VQA" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 321, + 296, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 321, + 296, + 440 + ], + "spans": [ + { + "bbox": [ + 55, + 321, + 296, + 440 + ], + "type": "text", + "content": "given sufficient context for the task, followed by a question and multiple answer choices. Accuracy was determined by comparing the model's response with the correct answer. We evaluated models across all subtasks including global perception, gesture understanding, and gesture localization in Section 3.3.3. The accuracy (Acc) in Table 8 is defined as: " + }, + { + "bbox": [ + 55, + 321, + 296, + 440 + ], + "type": "inline_equation", + "content": "\\mathrm{Acc} = \\frac{1}{N}\\sum_{k = 1}^{N}\\mathrm{MATCH}_k" + }, + { + "bbox": [ + 55, + 321, + 296, + 440 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 321, + 296, + 440 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 321, + 296, + 440 + ], + "type": "text", + "content": " is the total number of VQA samples and " + }, + { + "bbox": [ + 55, + 321, + 296, + 440 + ], + "type": "inline_equation", + "content": "\\mathrm{MATCH}_k" + }, + { + "bbox": [ + 55, + 321, + 296, + 440 + ], + "type": "text", + "content": " is a binary value. If the option in the model output matches the ground truth option, it is 1, otherwise it is 0." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 449, + 296, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 449, + 296, + 665 + ], + "spans": [ + { + "bbox": [ + 55, + 449, + 296, + 665 + ], + "type": "text", + "content": "Results and Findings. In Table 8, the experimental results show that while most VLMs can handle global perception and gesture understanding tasks without finetuning, they all struggle significantly with gesture localization tasks. This trend is understandable, as localization tasks inherently involve more ambiguity compared to perception and understanding tasks. While there is not a very clear correlation between the number of parameters and performance, Qwen2.5-VL, the largest of our open sourced models, performed well overall. To explore the potential for improvement, we finetuned Qwen2-VL-7B in Table 9, which led to enhanced performance in most metrics, except those related to gesture binary classification and target identification. Performance drops in some metrics can be attributed to easy overfitting on binary cases. An additional noteworthy observation is that despite being pretrained on vast multimodal datasets, none of the SOTA VLMs achieved over " + }, + { + "bbox": [ + 55, + 449, + 296, + 665 + ], + "type": "inline_equation", + "content": "70\\%" + }, + { + "bbox": [ + 55, + 449, + 296, + 665 + ], + "type": "text", + "content": " accuracy on the simplest task: counting humans in social interaction scenes." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "content": "These findings indicate that the SocialGesture tasks are significantly more challenging for VLMs than typical VQA benchmarks. This difficulty stems from the natural complexity of real-world social interaction videos used in our" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 321, + 555, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 321, + 555, + 477 + ], + "spans": [ + { + "bbox": [ + 313, + 321, + 555, + 477 + ], + "type": "text", + "content": "benchmark. Unlike controlled environments, our dataset contains realistic scenarios where people frequently overlap with each other, are partially visible, or extend beyond the camera frame, creating natural occlusion cases. In addition, there exists high interdependence between tasks in SocialGesture VQA, which makes it challenging. For instance, when models process such videos, they should first assess the overall scene despite these visual challenges, counting people and understanding actions. This process involves high-level reasoning and the ability to sequentially infer relationships—skills that are currently lacking in SOTA VLMs. Developing models capable of such contextual social visual reasoning remains a key challenge for future research." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 489, + 388, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 489, + 388, + 502 + ], + "spans": [ + { + "bbox": [ + 313, + 489, + 388, + 502 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 510, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 556, + 713 + ], + "type": "text", + "content": "We introduce SocialGesture, the first dataset and benchmark specifically designed for understanding and analyzing multiperson social gestures. We provide a comprehensive definition of tasks related to social gesture recognition, including a novel VQA dataset that encompasses global perception, gesture understanding, and spatial localization. To advance the field of human gesture and social interaction understanding, we benchmarked a variety of current state-of-the-art methods, including action recognition models and vision-language models (VLMs). Our findings reveal that multiperson social gesture understanding requires incremental visual reasoning—area where existing computer vision models fall short. This work offers insights into the complexities of multi-person social interactions and highlights the need for continued research into developing models capable of sophisticated visual reasoning. We hope that SocialGesture will motivate further innovations in gesture understanding." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19516" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "type": "text", + "content": "Acknowledgment" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 297, + 139 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 297, + 139 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 297, + 139 + ], + "type": "text", + "content": "We are thankful to the Keymakr for their great work in data annotation. This work was supported in part by the NRF grant funded by the Korea government(MSIT) (No. RS-2025-00563942)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 114, + 164 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 114, + 164 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 114, + 164 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 171, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 171, + 296, + 214 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 171, + 296, + 214 + ], + "spans": [ + { + "bbox": [ + 61, + 171, + 296, + 214 + ], + "type": "text", + "content": "[1] Qaisar Abbas, Mostafa EA Ibrahim, and M Arfan Jaffar. Video scene analysis: an overview and challenges on deep learning algorithms. Multimedia Tools and Applications, 77 (16):20415-20453, 2018. 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 217, + 296, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 217, + 296, + 270 + ], + "spans": [ + { + "bbox": [ + 61, + 217, + 296, + 270 + ], + "type": "text", + "content": "[2] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florecia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 5, 7, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 274, + 296, + 305 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 274, + 296, + 305 + ], + "spans": [ + { + "bbox": [ + 62, + 274, + 296, + 305 + ], + "type": "text", + "content": "[3] Anthropic. Claude 3 family. https://www.anthropic.com/news/claude-3-family, 2024. Accessed: 2024-05-27. 7, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 307, + 295, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 307, + 295, + 350 + ], + "spans": [ + { + "bbox": [ + 62, + 307, + 295, + 350 + ], + "type": "text", + "content": "[4] Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, et al. Qwen2. 5-vl technical report. arXiv preprint arXiv:2502.13923, 2025. 7, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 353, + 295, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 353, + 295, + 407 + ], + "spans": [ + { + "bbox": [ + 62, + 353, + 295, + 407 + ], + "type": "text", + "content": "[5] Gibran Benitez-Garcia, Jesus Olivares-Mercado, Gabriel Sanchez-Perez, and Keiji Yanai. Ipn hand: A video dataset and benchmark for real-time continuous hand gesture recognition. In 2020 25th international conference on pattern recognition (ICPR), pages 4340-4347. IEEE, 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 409, + 295, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 409, + 295, + 441 + ], + "spans": [ + { + "bbox": [ + 62, + 409, + 295, + 441 + ], + "type": "text", + "content": "[6] Paolo Bernardis and Maurizio Gentilucci. Speech and gesture share the same communication system. Neuropsychologia, 44(2):178-190, 2006. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 443, + 295, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 443, + 295, + 475 + ], + "spans": [ + { + "bbox": [ + 62, + 443, + 295, + 475 + ], + "type": "text", + "content": "[7] Gedas Bertasius, Heng Wang, and Lorenzo Torresani. Is space-time attention all you need for video understanding? In ICML, page 4, 2021. 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 477, + 295, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 477, + 295, + 531 + ], + "spans": [ + { + "bbox": [ + 62, + 477, + 295, + 531 + ], + "type": "text", + "content": "[8] Fabian Caba Heilbron, Victor Escorcia, Bernard Ghanem, and Juan Carlos Niebles. Activitynet: A large-scale video benchmark for human activity understanding. In Proceedings of the iee conference on computer vision and pattern recognition, pages 961-970, 2015. 5, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 534, + 295, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 534, + 295, + 577 + ], + "spans": [ + { + "bbox": [ + 62, + 534, + 295, + 577 + ], + "type": "text", + "content": "[9] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6299-6308, 2017. 3, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 579, + 295, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 579, + 295, + 632 + ], + "spans": [ + { + "bbox": [ + 57, + 579, + 295, + 632 + ], + "type": "text", + "content": "[10] Haoyu Chen, Henglin Shi, Xin Liu, Xiaobai Li, and Guoying Zhao. Smg: A micro-gesture dataset towards spontaneous body gestures for emotional stress state analysis. International Journal of Computer Vision, 131(6):1346-1366, 2023. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 635, + 295, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 635, + 295, + 666 + ], + "spans": [ + { + "bbox": [ + 57, + 635, + 295, + 666 + ], + "type": "text", + "content": "[11] Qirui Chen, Shangzhe Di, and Weidi Xie. Grounded multihop videoqa in long-form egocentric videos. arXiv preprint arXiv:2408.14469, 2024. 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 670, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 295, + 712 + ], + "type": "text", + "content": "[12] Zhe Chen, Weiyun Wang, Yue Cao, Yangzhou Liu, Zhangwei Gao, Erfei Cui, Jinguo Zhu, Shenglong Ye, Hao Tian, Zhaoyang Liu, et al. Expanding performance boundaries of open-source multimodal models with model, data, and test" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 333, + 73, + 555, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 555, + 94 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 555, + 94 + ], + "type": "text", + "content": "time scaling. arXiv preprint arXiv:2412.05271, 2024. 3, 7, 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 96, + 554, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 554, + 140 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 554, + 140 + ], + "type": "text", + "content": "[13] Christoph Feichtenhofer, Haoqi Fan, Jitendra Malik, and Kaiming He. Slowfast networks for video recognition. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6202-6211, 2019. 3, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 142, + 554, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 142, + 554, + 196 + ], + "spans": [ + { + "bbox": [ + 316, + 142, + 554, + 196 + ], + "type": "text", + "content": "[14] Chaoyou Fu, Yuhan Dai, Yongdong Luo, Lei Li, Shuhuai Ren, Renrui Zhang, Zihan Wang, Chenyu Zhou, Yunhang Shen, Mengdan Zhang, et al. Video-mme: The first-ever comprehensive evaluation benchmark of multi-modal llms in video analysis. arXiv preprint arXiv:2405.21075, 2024. 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 198, + 555, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 198, + 555, + 241 + ], + "spans": [ + { + "bbox": [ + 316, + 198, + 555, + 241 + ], + "type": "text", + "content": "[15] Deepanway Ghosal, Navonil Majumder, Soujanya Poria, Niyati Chhaya, and Alexander Gelbukh. Dialoguegcn: A graph convolutional neural network for emotion recognition in conversation. arXiv preprint arXiv:1908.11540, 2019. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 243, + 555, + 286 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 243, + 555, + 286 + ], + "spans": [ + { + "bbox": [ + 316, + 243, + 555, + 286 + ], + "type": "text", + "content": "[16] Shreya Ghosh, Zhixi Cai, Abhinav Dhall, Dimitrios Kollias, Roland Goecke, and Tom Gedeon. Mrac track 1: 2nd workshop on multimodal, generative and responsible affective computing. arXiv preprint arXiv:2409.07256, 2024. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 289, + 555, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 289, + 555, + 354 + ], + "spans": [ + { + "bbox": [ + 317, + 289, + 555, + 354 + ], + "type": "text", + "content": "[17] Kristen Grauman, Andrew Westbury, Eugene Byrne, Zachary Chavis, Antonino Furnari, Rohit Girdhar, Jackson Hamburger, Hao Jiang, Miao Liu, Xingyu Liu, et al. Ego4d: Around the world in 3,000 hours of egocentric video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18995-19012, 2022. 2, 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 356, + 554, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 356, + 554, + 421 + ], + "spans": [ + { + "bbox": [ + 317, + 356, + 554, + 421 + ], + "type": "text", + "content": "[18] Devamanyu Hazarika, Soujanya Poria, Amir Zadeh, Erik Cambria, Louis-Philippe Morency, and Roger Zimmermann. Conversational memory network for emotion recognition in dyadic dialogue videos. In Proceedings of the conference. Association for Computational Linguistics. North American Chapter. Meeting, page 2122. NIH Public Access, 2018. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 422, + 555, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 422, + 555, + 464 + ], + "spans": [ + { + "bbox": [ + 317, + 422, + 555, + 464 + ], + "type": "text", + "content": "[19] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. *ICLR*, 1(2):3, 2022. 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 468, + 555, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 468, + 555, + 521 + ], + "spans": [ + { + "bbox": [ + 317, + 468, + 555, + 521 + ], + "type": "text", + "content": "[20] Haroon Idrees, Amir R Zamir, Yu-Gang Jiang, Alex Gorban, Ivan Laptev, Rahul Sukthankar, and Mubarak Shah. The thumos challenge on action recognition for videos \"in the wild\". Computer Vision and Image Understanding, 155:1-23, 2017. 5, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 524, + 554, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 524, + 554, + 589 + ], + "spans": [ + { + "bbox": [ + 317, + 524, + 554, + 589 + ], + "type": "text", + "content": "[21] Simindokht Jahangard, Zhixi Cai, Shiki Wen, and Hamid Rezatofighi. Jrdb-social: A multifaceted robotic dataset for understanding of context and dynamics of human interactions within social groups. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22087-22097, 2024. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 591, + 555, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 591, + 555, + 657 + ], + "spans": [ + { + "bbox": [ + 317, + 591, + 555, + 657 + ], + "type": "text", + "content": "[22] Wenqi Jia, Miao Liu, Hao Jiang, Ishwarya Ananthabhotla, James M Rehg, Vamsi Krishna Ithapu, and Ruohan Gao. The audio-visual conversational graph: From an egocentric-exocentric perspective. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26396–26405, 2024. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 658, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 658, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 658, + 554, + 713 + ], + "type": "text", + "content": "[23] Alexander Kapitanov, Karina Kvanchiani, Alexander Nagaev, Roman Kraynov, and Andrei Makhliarchuk. Hagrid-hand gesture recognition image dataset. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 4572-4581, 2024. 1, 2" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19517" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 297, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 297, + 127 + ], + "type": "text", + "content": "[24] Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, et al. The kinetics human action video dataset. arXiv preprint arXiv:1705.06950, 2017. 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 128, + 296, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 128, + 296, + 193 + ], + "spans": [ + { + "bbox": [ + 56, + 128, + 296, + 193 + ], + "type": "text", + "content": "[25] Bolin Lai, Hongxin Zhang, Miao Liu, Aryan Pariani, Fiona Ryan, Wenqi Jia, Shirley Anugrah Hayati, James Rehg, and Diyi Yang. Werewolf among us: Multimodal resources for modeling persuasion behaviors in social deduction games. Association for Computational Linguistics: ACL 2023, 2023. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 194, + 296, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 194, + 296, + 227 + ], + "spans": [ + { + "bbox": [ + 56, + 194, + 296, + 227 + ], + "type": "text", + "content": "[26] Bongseok Lee and Yong Suk Choi. Graph based network with contextualized representations of turns in dialogue. arXiv preprint arXiv:2109.04008, 2021. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 228, + 296, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 228, + 296, + 282 + ], + "spans": [ + { + "bbox": [ + 56, + 228, + 296, + 282 + ], + "type": "text", + "content": "[27] Sangmin Lee, Bolin Lai, Fiona Ryan, Bikram Boote, and James M Rehg. Modeling multimodal social interactions: new challenges and baselines with densely aligned representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14585-14595, 2024. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 283, + 296, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 283, + 296, + 327 + ], + "spans": [ + { + "bbox": [ + 56, + 283, + 296, + 327 + ], + "type": "text", + "content": "[28] Sangmin Lee, Minzhi Li, Bolin Lai, Wenqi Jia, Fiona Ryan, Xu Cao, Ozgur Kara, Bikram Boote, Weiyan Shi, Diyi Yang, et al. Towards social ai: A survey on understanding social interactions. arXiv preprint arXiv:2409.15316, 2024. 1, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 327, + 296, + 382 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 327, + 296, + 382 + ], + "spans": [ + { + "bbox": [ + 56, + 327, + 296, + 382 + ], + "type": "text", + "content": "[29] Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Limin Wang, and Yu Qiao. Uniformerv2: Unlocking the potential of image vits for video understanding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1632-1643, 2023. 3, 6, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 383, + 296, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 383, + 296, + 448 + ], + "spans": [ + { + "bbox": [ + 56, + 383, + 296, + 448 + ], + "type": "text", + "content": "[30] Yanghao Li, Chao-Yuan Wu, Haoqi Fan, Karttikeya Mangalam, Bo Xiong, Jitendra Malik, and Christoph Feichtenhofer. Mviv2: Improved multiscale vision transformers for classification and detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4804-4814, 2022. 3, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 449, + 296, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 449, + 296, + 502 + ], + "spans": [ + { + "bbox": [ + 56, + 449, + 296, + 502 + ], + "type": "text", + "content": "[31] Dan Liu, Libo Zhang, and Yanjun Wu. Ld-congr: A large rgb-d video dataset for long-distance continuous gesture recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3304-3312, 2022. 1, 2, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 503, + 296, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 503, + 296, + 536 + ], + "spans": [ + { + "bbox": [ + 56, + 503, + 296, + 536 + ], + "type": "text", + "content": "[32] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava-last: Improved reasoning,OCR,and world knowledge,2024.3,7,8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 537, + 296, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 537, + 296, + 569 + ], + "spans": [ + { + "bbox": [ + 56, + 537, + 296, + 569 + ], + "type": "text", + "content": "[33] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 570, + 296, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 570, + 296, + 624 + ], + "spans": [ + { + "bbox": [ + 56, + 570, + 296, + 624 + ], + "type": "text", + "content": "[34] Xin Liu, Henglin Shi, Haoyu Chen, Zitong Yu, Xiaobai Li, and Guoying Zhao. imigue: An identity-free video dataset for micro-gesture understanding and emotion analysis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10631-10642, 2021. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 625, + 296, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 625, + 296, + 669 + ], + "spans": [ + { + "bbox": [ + 56, + 625, + 296, + 669 + ], + "type": "text", + "content": "[35] Zhaoyang Liu, Limin Wang, Wayne Wu, Chen Qian, and Tong Lu. Tam: Temporal adaptive module for video recognition. In Proceedings of the IEEE/CVF international conference on computer vision, pages 13708-13718, 2021. 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 670, + 296, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 296, + 714 + ], + "type": "text", + "content": "[36] Ze Liu, Jia Ning, Yue Cao, Yixuan Wei, Zheng Zhang, Stephen Lin, and Han Hu. Video swim transformer. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3202-3211, 2022. 3, 7" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 316, + 72, + 555, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 555, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 555, + 128 + ], + "type": "text", + "content": "[37] Joanna Materzynska, Guillaume Berger, Ingo Bax, and Roland Memisevic. The jester dataset: A large-scale video dataset of human gestures. In Proceedings of the IEEE/CVF international conference on computer vision workshops, pages 0–0, 2019. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 129, + 554, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 554, + 152 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 554, + 152 + ], + "type": "text", + "content": "[38] David McNeill. Hand and mind: What gestures reveal about thought. University of Chicago press, 1992. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 152, + 555, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 555, + 217 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 555, + 217 + ], + "type": "text", + "content": "[39] Pavlo Molchanov, Xiaodong Yang, Shalini Gupta, Kihwan Kim, Stephen Tyree, and Jan Kautz. Online detection and classification of dynamic hand gestures with recurrent 3d convolutional neural network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4207-4215, 2016. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 219, + 555, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 219, + 555, + 263 + ], + "spans": [ + { + "bbox": [ + 316, + 219, + 555, + 263 + ], + "type": "text", + "content": "[40] Fangzhou Mu, Sicheng Mo, and Yin Li. Snag: Scalable and accurate video grounding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18930-18940, 2024. 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 264, + 555, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 264, + 555, + 319 + ], + "spans": [ + { + "bbox": [ + 316, + 264, + 555, + 319 + ], + "type": "text", + "content": "[41] Yujia Peng, Jiaheng Han, Zhenliang Zhang, Lifeng Fan, Tengyu Liu, Siyuan Qi, Xue Feng, Yuxi Ma, Yizhou Wang, and Song-Chun Zhu. The tong test: Evaluating artificial general intelligence through dynamic embodied physical and social interactions. Engineering, 34:12-22, 2024. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 320, + 555, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 320, + 555, + 365 + ], + "spans": [ + { + "bbox": [ + 316, + 320, + 555, + 365 + ], + "type": "text", + "content": "[42] Asanka G Perera, Yee Wei Law, and Javaan Chahl. Uav-gesture: A dataset for uav control and gesture recognition. In Proceedings of the European Conference on Computer Vision (ECCV) Workshops, pages 0–0, 2018. 1, 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 365, + 555, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 365, + 555, + 431 + ], + "spans": [ + { + "bbox": [ + 316, + 365, + 555, + 431 + ], + "type": "text", + "content": "[43] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 433, + 554, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 433, + 554, + 466 + ], + "spans": [ + { + "bbox": [ + 316, + 433, + 554, + 466 + ], + "type": "text", + "content": "[44] Razieh Rastgoo, Kourosh Kiani, and Sergio Escalera. Sign language recognition: A deep survey. Expert Systems with Applications, 164:113794, 2021. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 467, + 555, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 467, + 555, + 499 + ], + "spans": [ + { + "bbox": [ + 316, + 467, + 555, + 499 + ], + "type": "text", + "content": "[45] Fiona Ryan, Ajay Bati, Sangmin Lee, Daniel Bolya, Judy Hoffman, and James M Rehg. Gaze-lle: Gaze target estimation via large-scale learned encoders. 2025. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 500, + 554, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 500, + 554, + 534 + ], + "spans": [ + { + "bbox": [ + 316, + 500, + 554, + 534 + ], + "type": "text", + "content": "[46] Harvey Sacks, Emanuel A Schegloff, and Gail Jefferson. A simplest systematics for the organization of turn-taking for conversation. language, 50(4):696-735, 1974. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 535, + 555, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 535, + 555, + 579 + ], + "spans": [ + { + "bbox": [ + 316, + 535, + 555, + 579 + ], + "type": "text", + "content": "[47] Yuehao Song, Xinggang Wang, Jingfeng Yao, Wenyu Liu, Jinglin Zhang, and Xiangmin Xu. Vitgaze: gaze following with interaction features in vision transformers. Visual Intelligence, 2(1):1-15, 2024. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 580, + 555, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 580, + 555, + 645 + ], + "spans": [ + { + "bbox": [ + 316, + 580, + 555, + 645 + ], + "type": "text", + "content": "[48] Andreas Stolcke, Klaus Ries, Noah Coccaro, Elizabeth Shriberg, Rebecca Bates, Daniel Jurafsky, Paul Taylor, Rachel Martin, Carol Van Ess-Dykema, and Marie Meteor. Dialogue act modeling for automatic tagging and recognition of conversational speech. Computational linguistics, 26(3):339-373, 2000. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 647, + 555, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 555, + 712 + ], + "type": "text", + "content": "[49] Kosei Tanada, Shigemichi Matsuzaki, Kazuhito Tanaka, Shintaro Nakaoka, Yuki Kondo, and Yuto Mori. Pointing gesture understanding via visual prompting and visual question answering for interactive robot navigation. In First Workshop on Vision-Language Models for Navigation and Manipulation at ICRA 2024. 5" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19518" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 298, + 654 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 56, + 73, + 298, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 298, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 298, + 127 + ], + "type": "text", + "content": "[50] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In International conference on machine learning, pages 10347-10357. PMLR, 2021. 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 297, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 297, + 184 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 297, + 184 + ], + "type": "text", + "content": "[51] Du Tran, Heng Wang, Lorenzo Torresani, Jamie Ray, Yann LeCun, and Manohar Paluri. A closer look at spatiotemporal convolutions for action recognition. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 6450-6459, 2018. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 185, + 296, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 185, + 296, + 239 + ], + "spans": [ + { + "bbox": [ + 56, + 185, + 296, + 239 + ], + "type": "text", + "content": "[52] Jun Wan, Yibing Zhao, Shuai Zhou, Isabelle Guyon, Sergio Escalera, and Stan Z Li. Chalearn looking at people rgb-d isolated and continuous datasets for gesture recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 56-64, 2016. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 240, + 296, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 240, + 296, + 296 + ], + "spans": [ + { + "bbox": [ + 56, + 240, + 296, + 296 + ], + "type": "text", + "content": "[53] Limin Wang, Yuanjun Xiong, Zhe Wang, Yu Qiao, Dahua Lin, Xiaou Tang, and Luc Van Gool. Temporal segment networks: Towards good practices for deep action recognition. In European conference on computer vision, pages 20-36. Springer, 2016. 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 297, + 296, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 297, + 296, + 351 + ], + "spans": [ + { + "bbox": [ + 56, + 297, + 296, + 351 + ], + "type": "text", + "content": "[54] Limin Wang, Bingkun Huang, Zhiyu Zhao, Zhan Tong, Yinan He, Yi Wang, Yali Wang, and Yu Qiao. Videomae v2: Scaling video masked autoencoders with dual masking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14549-14560, 2023. 3, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 352, + 296, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 352, + 296, + 407 + ], + "spans": [ + { + "bbox": [ + 56, + 352, + 296, + 407 + ], + "type": "text", + "content": "[55] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. arXiv preprint arXiv:2409.12191, 2024. 3, 7, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 408, + 297, + 461 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 408, + 297, + 461 + ], + "spans": [ + { + "bbox": [ + 56, + 408, + 297, + 461 + ], + "type": "text", + "content": "[56] Yi Wang, Kunchang Li, Xinhao Li, Jiashuo Yu, Yinan He, Guo Chen, Baoqi Pei, Rongkun Zheng, Jilan Xu, Zun Wang, et al. Internvideo2: Scaling video foundation models for multimodal video understanding. arXiv preprint arXiv:2403.15377, 2024. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 463, + 296, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 463, + 296, + 486 + ], + "spans": [ + { + "bbox": [ + 56, + 463, + 296, + 486 + ], + "type": "text", + "content": "[57] Diyi Yang, Dirk Hovy, David Jurgens, and Barbara Plank. The call for socially aware language technologies, 2024. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 487, + 296, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 487, + 296, + 529 + ], + "spans": [ + { + "bbox": [ + 56, + 487, + 296, + 529 + ], + "type": "text", + "content": "[58] Ceng Zhang, Junxin Chen, Jiatong Li, Yanhong Peng, and Zebing Mao. Large language models for human-robot interaction: A review. Biomimetic Intelligence and Robotics, page 100131, 2023. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 531, + 296, + 574 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 531, + 296, + 574 + ], + "spans": [ + { + "bbox": [ + 56, + 531, + 296, + 574 + ], + "type": "text", + "content": "[59] Chen-Lin Zhang, Jianxin Wu, and Yin Li. Actionformer: Localizing moments of actions with transformers. In European Conference on Computer Vision, pages 492-510. Springer, 2022. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 576, + 296, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 576, + 296, + 620 + ], + "spans": [ + { + "bbox": [ + 56, + 576, + 296, + 620 + ], + "type": "text", + "content": "[60] Yifan Zhang, Congqi Cao, Jian Cheng, and Hanqing Lu. Egogesture: A new dataset and benchmark for egocentric hand gesture recognition. IEEE Transactions on Multimedia, 20(5):1038-1050, 2018. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 621, + 296, + 654 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 621, + 296, + 654 + ], + "spans": [ + { + "bbox": [ + 56, + 621, + 296, + 654 + ], + "type": "text", + "content": "[61] Y Zhang, B Li, H Liu, Y Lee, L Gui, D Fu, J Feng, Z Liu, and C Li. Llava-last: A strong zero-shot video understanding model. 2024. 8" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19519" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/SocialMOIF_ Multi-Order Intention Fusion for Pedestrian Trajectory Prediction/74c8547f-5c13-42b2-b27c-77ef8f52058e_content_list.json b/2025/SocialMOIF_ Multi-Order Intention Fusion for Pedestrian Trajectory Prediction/74c8547f-5c13-42b2-b27c-77ef8f52058e_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..ca88c170bbe62ef0579f41fa38abead0e5166772 --- /dev/null +++ b/2025/SocialMOIF_ Multi-Order Intention Fusion for Pedestrian Trajectory Prediction/74c8547f-5c13-42b2-b27c-77ef8f52058e_content_list.json @@ -0,0 +1,1633 @@ +[ + { + "type": "text", + "text": "SocialMOIF: Multi-Order Intention Fusion for Pedestrian Trajectory Prediction", + "text_level": 1, + "bbox": [ + 91, + 130, + 903, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Kai Chen $^{1*}$ , Xiaodong Zhao $(\\boxtimes)$ ^{1*}, Yujie Huang $^{1}$ , Guoyu Fang $^{1}$ , Xiao Song $^{2}$ , Ruiping Wang $^{3}$ , Ziyuan Wang $^{1}$", + "bbox": [ + 243, + 179, + 750, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 Nanjing University of Aeronautics and Astronautics", + "bbox": [ + 285, + 215, + 712, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2 Beihang University 3 Nanyang Technological University", + "bbox": [ + 267, + 233, + 728, + 252 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{chen_kai, xdzhao, yj_huang, fanggy, zy_wang}@nuaa.edu.cn", + "bbox": [ + 251, + 253, + 745, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "songxiao@buaa.edu.cn, ruiping.wang@ntu.edu.sg", + "bbox": [ + 295, + 272, + 697, + 286 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 335, + 326, + 353 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The analysis and prediction of agent trajectories are crucial for decision-making processes in intelligent systems, with precise short-term trajectory forecasting being highly significant across a range of applications. Agents and their social interactions have been quantified and modeled by researchers from various perspectives; however, substantial limitations exist in the current work due to the inherent high uncertainty of agent intentions and the complex higher-order influences among neighboring groups. SocialMOIF is proposed to tackle these challenges, concentrating on the higher-order intention interactions among neighboring groups while reinforcing the primary role of first-order intention interactions between neighbors and the target agent. This method develops a multi-order intention fusion model to achieve a more comprehensive understanding of both direct and indirect intention information. Within SocialMOIF, a trajectory distribution approximator is designed to guide the trajectories toward values that align more closely with the actual data, thereby enhancing model interpretability. Furthermore, a global trajectory optimizer is introduced to enable more accurate and efficient parallel predictions. By incorporating a novel loss function that accounts for distance and direction during training, experimental results demonstrate that the model outperforms previous state-of-the-art baselines across multiple metrics in both dynamic and static datasets.", + "bbox": [ + 88, + 368, + 483, + 760 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 791, + 220, + 806 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Trajectory prediction is of considerable importance for intelligent transportation and surveillance systems. Studies indicate that trajectories derived from the physical", + "bbox": [ + 89, + 816, + 483, + 864 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a5b16720cb4c521b66b0bc3423abd877df5fe2e2275fcd0e4f0bd6c9b45791e2.jpg", + "image_caption": [ + "Figure 1. The process of SocialMOIF includes: (1) extracting the position, velocity and trajectory information of agents, (2) achieving multi-order fused intention modeling, (3) guiding the distribution of trajectories and performing global optimization. Finally we obtain accurate parallel predictions." + ], + "image_footnote": [], + "bbox": [ + 522, + 324, + 897, + 571 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "world encompass rich spatiotemporal intention information [15, 20, 26, 29, 45]. Therefore, a significant challenge arises in effectively capturing (data modeling) and analyzing (encoding-decoding) processes related to the complex and diverse historical social intentions of agents, while linking this information to future trajectory predictions for enhanced efficiency.", + "bbox": [ + 511, + 686, + 906, + 792 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "History social intentions. Existing research can be broadly classified into two main categories: knowledge-driven methods primarily depend on expert knowledge. For instance, biomimetic method [41] achieved agent localization and intention interaction modeling by simulating the positioning of marine animals. In contrast, certain studies utilized game-theoretic approaches to analyze", + "bbox": [ + 511, + 795, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Equal contribution.", + "bbox": [ + 112, + 875, + 228, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "Code is available at https://github.com/XiaodZhao/SocialMOIF.", + "bbox": [ + 112, + 887, + 468, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "22465", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "social intentions through the simulation of agent flows [13, 50] and evacuation processes [14, 55], relying on explicit rules or formulas. However, the behavioral patterns and interaction rules among agents in social scenarios displayed high nonlinearity and uncertainty [26, 39, 40]. Data-driven methods implicitly learn interactions among agents through neural networks. For instance, LSTM-based methods [1, 11] continuously refined social pooling patterns, while attention-based approaches [7, 49, 51] dynamically adjusted agents' focal points in global interactions from a spatiotemporal integration perspective to optimize predicted trajectories. These methods primarily focused on straightforward interactions or direct responses between the target agent and its neighbors, often neglecting the higher-order, multidimensional influences exerted by surrounding agents on the target agent's intentions. Kim et al. [16] modeled agent intentions from 1st-order to N-order equally. However, when the number of agents was large, it may prioritized higher-order intentions, thereby diminishing the primary impact of lower-order intentions. Thus, the development of a multi-order fusion network that reinforces the dominant role of direct interactions while considering the secondary influence of higher-order transmissions on the target is essential.", + "bbox": [ + 89, + 90, + 480, + 452 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Associating future trajectory information. Knowledge-driven methods face limitations due to the inherent flaws of unidirectional inference, which hinder effective association with future trajectory information. Data-driven methods exhibit two main shortcomings: first, agent movements are temporally dynamic, and relying solely on distance loss fails to capture these dynamic variations. Second, in generative model approaches [25, 28, 34, 38, 47], future trajectory information is implicitly involved in the training process without explicit guidance, resulting in poor interpretability. In safety-sensitive tasks such as autonomous driving, it is necessary to utilize meaningful and goal-oriented parameters and models.", + "bbox": [ + 89, + 457, + 482, + 652 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Efficiency. Existing methods predominantly emphasize upstream tasks. For example, LSTM-based approaches [5, 6, 36, 43, 54] modeled trajectory data in a serial temporal manner, while Transformer-based methods [9, 19, 48, 49] analyzed trajectory information through parallel processing. However, in downstream tasks, trajectories are often generated independently and sequentially. This method not only accumulates errors but also limits the efficiency of trajectory prediction. Therefore, establishing a trajectory parallel prediction model that accounts for global influences across future periods is a promising area for further exploration.", + "bbox": [ + 89, + 656, + 480, + 821 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "As illustrated in Fig. 1, SocialMOIF is proposed to address the trajectory prediction task. A Multi-Order Intention Fusion model (MOIF) is introduced to guide the learning direction of the network. MOIF includes a first-order intention interaction layer and a higher-order intention in", + "bbox": [ + 89, + 824, + 480, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "teraction layer. The first layer focuses on direct interaction information between neighbors and the target agent, while the second layer addresses the indirect interaction information conveyed to the target agent after multiple interactions among neighbor targets. Inspiration is drawn from the squeeze theorem to design a trajectory distribution approximator to guide trajectories toward a distribution that closely aligns with ground truth and to enhance model interpretability. Simultaneously, RNN is incorporated to update the latent variables, achieving explicit supervision during training. The temporal dimension of decoded trajectories is aggregated to design a global trajectory optimizer. To the best of our knowledge, KANs [23, 24] are firstly introduced for the global optimization of trajectories, enabling more accurate and efficient parallel predictions. Furthermore, considering the travel directions of the agents, a distance-direction fused loss function is designed to comprehensively supervise the dynamic states of agents.", + "bbox": [ + 511, + 90, + 903, + 362 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, the main contributions of this paper are:", + "bbox": [ + 532, + 364, + 883, + 378 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- The proposal of SocialMOIF, which incorporates four novel designs:", + "bbox": [ + 519, + 378, + 903, + 407 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "(1) MOIF is introduced to comprehensively capture interaction information between the target agent and its neighbors;", + "bbox": [ + 537, + 407, + 903, + 452 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "(2) A trajectory distribution approximator is developed to approximate and explicitly update the distribution of latent variables, guiding the distribution of trajectories;", + "bbox": [ + 537, + 452, + 903, + 513 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "(3) A global trajectory optimizer is designed, and KANs are introduced for the first time to achieve parallel trajectory optimization across the global spatiotemporal domain;", + "bbox": [ + 537, + 513, + 903, + 574 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "(4) A new direction-distance fused loss function is proposed for more effective supervision of agents' dynamic changes.", + "bbox": [ + 537, + 574, + 903, + 619 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- Comprehensive comparative experiments are conducted with existing high-quality models, demonstrating the effective and robust predictive capabilities of SocialMOIF.", + "bbox": [ + 519, + 619, + 903, + 664 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related works", + "text_level": 1, + "bbox": [ + 513, + 679, + 658, + 694 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Knowledge-driven methods. Helbing et al. [12] first proposed the social force model, which describes pedestrian interactions through an energy potential field. Best et al. [3] introduced a Bayesian inference-based intention inference model, integrating observed movement trajectories with prior knowledge. EqMotion [46] employed geometric analysis to convert human-human interaction trajectory analysis into an optimization problem by examining object geometry. Morris et al. [31] proposed an implicit Markov model for trajectory prediction, which is utilized for spatiotemporal probabilistic modeling of various types of pedestrian trajectories. A recent approach, socialcircle [41], incorporated angle-based light rule constraints into the trainable back", + "bbox": [ + 511, + 704, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "22466", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "bone to achieve interaction modeling of agent intentions.", + "bbox": [ + 89, + 90, + 467, + 106 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Despite these advancements, these methods rely on knowledge from specific domains and therefore struggle to fully capture the complex agent intentions within social scenarios. Although attempts have been made to incorporate future trajectory information to overcome the limitations of unidirectional inference, the model's equations remain overly sensitive to common variations, leading to persistent challenges in trajectory prediction tasks [42, 52].", + "bbox": [ + 89, + 106, + 483, + 226 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Data-driven methods. LSTM-based methods model trajectory data temporally. Alahi et al. [1] employed a \"social pooling mechanism\" that considers only the intentions of agents within a grid while neglecting interactions with boundary agents. Gupta et al. [11] improved upon this by implementing a pooling mechanism that treats pedestrians in neighboring areas equally. Transformer-based methods enhanced temporal modeling approaches. Yu et al. [49] modeled and predicted trajectories using a spatio-temporal concatenation approach. Yin et al. [48] introduced optical flow to simulate agents' motion intentions. Graph neural networks were also utilized for trajectory prediction to model intention interactions among different nodes [8, 22, 30, 35, 37, 44]. Diffusion models [10, 27] were used to explore trajectory prediction based on randomness.", + "bbox": [ + 89, + 227, + 483, + 453 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "However, most of these methods primarily emphasize direct interactions between the target agent and its neighbors, often overlooking the higher-order influences that neighbors exert on the target agent. Furthermore, many of these approaches incorporate future trajectory information only at the network's output stage, supervising agents' trajectory changes solely through distance loss. Moreover, the limited interpretability of latent variable distributions in generative models can impede further development of these models.", + "bbox": [ + 89, + 454, + 483, + 589 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In addition, the aforementioned methods mainly emphasize trajectory analysis and modeling of upstream tasks. For output in downstream tasks, results are generally produced independently in a sequential manner over time. This sequential processing introduces delays that may result in untimely responses, potentially leading to safety concerns. Furthermore, each future prediction is dependent on the state from the preceding time step, which may lead to accumulated errors, adversely affecting prediction accuracy.", + "bbox": [ + 89, + 589, + 483, + 726 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Approach", + "text_level": 1, + "bbox": [ + 89, + 739, + 199, + 757 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The number of agents in the scenario is set to $(N_{n} + 1)$ . The trajectory of the target agent $i$ in history period $T_{H}$ is $g_{i}^{1:T_{H}}$ , and $g_{i}^{t}$ is the two-dimensional spatial coordinate at moment $t$ . The remaining agents are tentatively designated as neighbors, and the number is $N_{n}$ . The trajectory of $i$ in the future period to be predicted is denoted as $\\hat{g}_{i}^{T_{H} + 1:T_{F}}$ . We propose the multi-order intention fusion model in Section 3.1, the trajectory distribution approximator in Section 3.2, the global trajectory optimizer in Section 3.3, and the", + "bbox": [ + 89, + 763, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "direction-distance fusion loss function in Section 3.4.", + "bbox": [ + 511, + 90, + 867, + 106 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Multi-order intention fusion model", + "text_level": 1, + "bbox": [ + 511, + 114, + 818, + 130 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The model consists of two layers: a first-order intention interaction layer and a higher-order intention interaction layer. As shown in Fig. 2, the first-order layer emphasizes the direct, dominant influence of neighbors on the target agent's intention. The higher-order layer accounts for the indirect and secondary influences on the target agent's intention that arise from interactions within the group of neighbors, thereby enabling a more comprehensive understanding of intentions. Based on the current and final history motion states of the agents, input components for the layers are first constructed, followed by the extraction and fusion of multi-order intention information.", + "bbox": [ + 511, + 137, + 906, + 318 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Higher-order intention interaction layer. From a more fundamental perspective, the higher-order intention interactions within the group of neighbors can be decomposed into first-order intention interactions between individual neighbors. We define the intention influence of neighbor agent $a$ on neighbor agent $b$ as $w_{a\\rightarrow b}$ ; for an agent itself, it is represented as $w_{a\\rightarrow a}$ . In a scenario with $N_{n}$ agents, the total number of such intentions is $\\Omega$ :", + "bbox": [ + 511, + 319, + 908, + 439 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\Omega = \\left(N _ {n} - 1\\right) N _ {n} + N _ {n} = N _ {n} ^ {2} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 601, + 449, + 906, + 467 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To obtain the intentions, we extract the absolute positions and velocities of the $N_{n}$ neighbors corresponding to the target agent $i$ . They are processed as a layer input component through the embedding layer $\\pmb{f}_{S}$ to obtain $U_{i}$ .", + "bbox": [ + 511, + 478, + 906, + 540 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nU _ {i} = \\boldsymbol {f} _ {U} \\left(x _ {1: N _ {n}} ^ {1: T _ {H}}, y _ {1: N _ {n}} ^ {1: T _ {H}}, v _ {x 1: N _ {n}} ^ {1: T _ {H}}, v _ {y 1: N _ {n}} ^ {1: T _ {H}}\\right) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 570, + 547, + 906, + 574 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We capture the intention interactions of neighbors through multiple subspaces in parallel. $U_{i}$ is mapped into a learnable query vector $Q_{U}$ , a key vector $K_{U}$ , and a value vector $V_{U}$ . The feature dimensions are flattened into M subspaces. For the m-th subspace, the overall intention coefficient matrix within neighbors can be expressed as:", + "bbox": [ + 511, + 583, + 905, + 674 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nW _ {U} ^ {m} = \\operatorname {S o f t M a x} \\left(\\frac {Q _ {U} ^ {m} \\left(K _ {U} ^ {m}\\right) ^ {T}}{d _ {M}}\\right) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 594, + 683, + 906, + 724 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Thus, we obtain $w_{a\\to b} = [W_U^m ]_{ab}$ . For this matrix, the number of elements is $\\Omega$", + "bbox": [ + 511, + 734, + 905, + 763 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The first-order intention interaction layer. The layer input component of the target agent $i$ is processed through the embedding layer $\\pmb{f}_{S}$ to obtain $S_{i}$ .", + "bbox": [ + 511, + 765, + 905, + 811 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nS _ {i} = \\pmb {f} _ {S} \\left(x _ {i} ^ {1: T _ {H}}, y _ {i} ^ {1: T _ {H}}, d _ {i \\leftrightarrow 1: N _ {n}} ^ {1: T _ {H}}, \\theta_ {i \\leftrightarrow 1: N _ {n}} ^ {1: T _ {H}}, e _ {i \\leftrightarrow 1: N _ {n}} ^ {1: T _ {H}}\\right) (4)\n$$\n", + "text_format": "latex", + "bbox": [ + 521, + 820, + 906, + 847 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We intensify the dominant role of direct interaction in components and focus on the design of interaction rules. For the present case, $d_{i\\leftrightarrow j}^{1:T_H}$ is the distance between the target", + "bbox": [ + 511, + 854, + 908, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "22467", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/ee95a298984c868f3ddb918b355f95a29b441c8107ebc48cb0487ce9d0e4ca5a.jpg", + "image_caption": [ + "Figure 2. Computational pipeline of SocialMOIF. Firstly, the first-order interaction layer focuses on direct interactions between the target agent and its neighbors. The higher-order interaction layer addresses the interactions among neighbors. Multi-order intention is obtained by fusing the information from both layers. This multi-order intention is then combined with the actual future trajectory, with the distribution of the trajectory guided by an approximator. An optimizer is subsequently employed to enhance the decoded trajectory. Additionally, a fused distance-direction loss function supervises the training process. Finally, efficient parallel prediction is achieved." + ], + "image_footnote": [], + "bbox": [ + 94, + 88, + 903, + 428 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "agent $i$ and neighbor $j$ , and $\\theta_{i\\leftrightarrow j}^{1:T_H}$ is the velocity angle. For the final history, $e_{i\\leftrightarrow j}^{1:T_H}$ is the final distance in the history period, assuming that after moment $t$ , both the target agent $i$ and neighbor $j$ maintain the original velocity in the tangent direction $\\overrightarrow{vx_i^t},\\overrightarrow{vx_j^t}$ ; then, $e_{ij}^t$ can be expressed as:", + "bbox": [ + 89, + 532, + 483, + 619 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\lambda = \\left| \\left(\\overrightarrow {d _ {i j} ^ {t}} \\cdot \\overrightarrow {v _ {i j} ^ {t}}\\right) / \\left\\| \\overrightarrow {v _ {i j} ^ {t}} \\right\\| ^ {2} \\right| \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 197, + 628, + 482, + 664 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ne _ {i j} ^ {t} = \\left\\| \\overrightarrow {d _ {i j} ^ {t}} + \\min (\\lambda , T _ {H} - t) \\cdot \\left(\\overrightarrow {v x _ {j} ^ {t}} - \\overrightarrow {v y _ {i} ^ {t}}\\right) \\right\\| \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 130, + 672, + 482, + 702 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Similarly, $S_{i}$ is mapped into learnable query vectors $Q_{S}$ , key vectors $K_{S}$ , and value vectors $V_{S}$ . The matrix of the intention coefficients between the target agent $i$ and its neighbors is as follows:", + "bbox": [ + 89, + 705, + 483, + 765 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nW _ {S} = \\operatorname {S o f t M a x} \\left(\\frac {Q _ {S} \\left(K _ {S}\\right) ^ {T}}{d _ {S}}\\right) \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 178, + 773, + 482, + 816 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Intention Fusion. Higher-order interaction intentions are conducted to the target agent and combined with first-order interaction intentions to achieve multi-order intention fusion. During this fusion, the higher-order interaction intentions captured within each subspace exert varying effects", + "bbox": [ + 89, + 824, + 483, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "when applied to the target agent. We set the corresponding trainable influence factor $\\eta_{m}$ to denote this influence. The attention score $A_{i}$ is denoted as:", + "bbox": [ + 511, + 534, + 906, + 580 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nA _ {i} = \\left(\\sum_ {m = 1} ^ {M} \\eta_ {m} W _ {U} ^ {m} + W _ {S}\\right) V _ {S} \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 599, + 589, + 906, + 631 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $A_{i}$ is processed by two skip connections and a final fully connected layer to obtain the multi-order fused intention $I_{i}$ .", + "bbox": [ + 511, + 640, + 906, + 685 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Trajectory distribution approximator", + "text_level": 1, + "bbox": [ + 511, + 694, + 836, + 710 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We define the trajectory $g_{igt}^{T_H + 1:T_F}$ of the target agent $i$ in the future period. $B_{i}$ is obtained after feature enhancement through the embedding layer $f_{B}$ . By utilizing $I_{i}$ as the lower bound and $B_{i}$ as the upper bound, the latent variable distribution in the potential space is explicitly approximated. First, the probability distribution of the target agent $i$ 's trajectory is initially modeled as:", + "bbox": [ + 511, + 714, + 906, + 823 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} p \\left(g _ {i} ^ {T _ {H} + 1: T _ {F}} \\mid U _ {i}, S _ {i}\\right) \\\\ = \\prod_ {t = T _ {H} + 1} ^ {T _ {F}} p \\left(g _ {i} ^ {T _ {H} + 1 + t} \\mid g _ {i} ^ {T _ {H} + 1: T _ {H} + t - 1}, U _ {i}, S _ {i}\\right) \\tag {9} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 542, + 832, + 903, + 904 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "22468", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We sample the noise variables $\\varepsilon \\sim N(0,1)$ and reparameterize the latent variables that are approximated at $t$ [17]:", + "bbox": [ + 89, + 90, + 482, + 121 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nz _ {i \\varphi} ^ {t} = u _ {i \\varphi} ^ {t} + \\varepsilon \\cdot \\sigma_ {i \\varphi} ^ {t} \\quad q _ {\\varphi} (\\cdot | I _ {i} ^ {t - 1}, B _ {i} ^ {t}) \\sim N (u _ {i \\varphi} ^ {t}, \\sigma_ {i \\varphi} ^ {t}) \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 128, + 482, + 150 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\varphi$ are parameters that need to be optimized during the training process. In the test mode, the latent variables $z_{i\\vartheta}^{t}$ are only reasoned by $I_{i}$ , and $\\vartheta$ is also involved in the training process of the model.", + "bbox": [ + 89, + 156, + 483, + 215 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We sample and decode the target agent's trajectories on the basis of latent variables and multi-order fused intention information.", + "bbox": [ + 89, + 217, + 483, + 261 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\ng _ {i} ^ {t} \\sim \\left\\{ \\begin{array}{l l} p _ {\\delta} (\\cdot | z _ {i \\varphi} ^ {t}, I _ {i} ^ {t - 1}) & \\text {i f t r a i n} \\\\ p _ {\\delta} (\\cdot | z _ {i \\vartheta} ^ {t}, I _ {i} ^ {t - 1}) & \\text {i f t e s t} \\end{array} \\right. \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 166, + 268, + 482, + 308 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\delta$ are parameters that need to be optimized during the training process.", + "bbox": [ + 89, + 315, + 482, + 345 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Concat the latent variables with the trajectories generated at the current time $t$ , updating $I_{i}^{t - 1}$ via $f_{zg}$ .", + "bbox": [ + 89, + 345, + 483, + 378 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nI _ {i} ^ {t} = \\left\\{ \\begin{array}{l l} \\boldsymbol {f} _ {z g} \\left(\\operatorname {c o n c a t} \\left(z _ {i \\varphi} ^ {t}, g _ {i} ^ {t}\\right), I _ {i} ^ {t - 1}\\right) & \\text {i f t r a i n} \\\\ \\boldsymbol {f} _ {z g} \\left(\\operatorname {c o n c a t} \\left(z _ {i \\vartheta} ^ {t}, g _ {i} ^ {t}\\right), I _ {i} ^ {t - 1}\\right) & \\text {i f t e s t} \\end{array} \\right. \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 114, + 387, + 482, + 426 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the next time step, the updated multi-order fused intention contains the information of the latent variables, which is then combined with $B_{i}^{t}$ to approximate the latent variables and update them.", + "bbox": [ + 89, + 434, + 483, + 494 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "By guiding the trajectory distribution through such a training process, the generated trajectories are explicitly represented by the latent variables as:", + "bbox": [ + 89, + 494, + 483, + 540 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} p \\left(g _ {i} ^ {T _ {H} + 1: T _ {F}} \\mid U _ {i}, S _ {i}\\right) \\\\ = \\prod_ {t = T _ {H} + 1} ^ {T _ {F}} \\int_ {z _ {i \\varphi} ^ {t}} p \\left(g _ {i} ^ {t} \\mid g _ {i} ^ {T _ {H} + 1: t - 1}, I _ {i} ^ {t - 1}, B _ {i} ^ {t}, z _ {i \\varphi} ^ {t}\\right) \\cdot \\tag {13} \\\\ p \\left(z _ {i \\varphi} ^ {t} \\mid g _ {i} ^ {T _ {H} + 1: t - 1}, I _ {i} ^ {t - 1}, B _ {i} ^ {t}\\right) d z _ {i \\varphi} ^ {t} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 107, + 547, + 482, + 648 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In test mode, $B_{i}^{t}$ is removed, and $\\varphi$ is replaced with $\\vartheta$ .", + "bbox": [ + 109, + 656, + 468, + 672 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Global trajectory optimizer", + "text_level": 1, + "bbox": [ + 89, + 680, + 339, + 696 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For efficient prediction, we introduce the KANs for the first time. We aggregate the $g_{i}^{T_{H} + 1:T_{F}}$ time dimension to obtain $\\Gamma^0\\in \\mathbb{R}^{2T_F}$ to facilitate global parallel optimization. The depth of the optimizer is $L$ layers, the number of nodes in the $\\ell$ -th layer is $m$ , the number of nodes in the $\\ell +1$ -th layer is $n$ , and the activation function matrix corresponding to the nodes in the $l$ -th layer is $\\Phi_{\\ell}$ . The optimization process for the $\\ell \\rightarrow \\ell +1$ layer is as follows:", + "bbox": [ + 89, + 702, + 483, + 823 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\Gamma^ {\\ell + 1} = \\Phi_ {\\ell} \\Gamma^ {\\ell} = \\left[ \\begin{array}{c c c c} \\varphi_ {1, 1} ^ {\\ell} & \\varphi_ {1, 2} ^ {\\ell} & \\dots & \\varphi_ {1, m} ^ {\\ell} \\\\ \\varphi_ {2, 1} ^ {\\ell} & \\varphi_ {2, 2} ^ {\\ell} & \\dots & \\varphi_ {2, m} ^ {\\ell} \\\\ \\vdots & \\vdots & \\ddots & \\vdots \\\\ \\varphi_ {n, 1} ^ {\\ell} & \\varphi_ {n, 2} ^ {\\ell} & \\dots & \\varphi_ {n, m} ^ {\\ell} \\end{array} \\right] \\Gamma^ {\\ell} (1 4)\n$$\n", + "text_format": "latex", + "bbox": [ + 104, + 830, + 483, + 905 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Eventually, the predicted trajectory is reduced back to the original dimensional space to obtain the final trajectory over the entire prediction period at once.", + "bbox": [ + 511, + 90, + 905, + 136 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {g} _ {i} ^ {T _ {H} + 1: T _ {F}} \\leftarrow \\Gamma^ {L} = \\left(\\Phi_ {L} \\circ \\Phi_ {L - 1} \\circ \\dots \\circ \\Phi_ {1} \\circ \\Phi_ {0}\\right) \\Gamma^ {0} (1 5)\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 148, + 905, + 167 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given the diversity of agent intentions, the process predicts multiple possible trajectories.", + "bbox": [ + 511, + 181, + 905, + 212 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Loss function", + "text_level": 1, + "bbox": [ + 513, + 224, + 653, + 238 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Agents traveling in opposite or the same direction exert distinct effects on their respective movement states. Therefore, direction loss is incorporated to supervise the agent's dynamics during backpropagation.", + "bbox": [ + 511, + 246, + 905, + 306 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For distance. The distance prediction error in future periods can be expressed as:", + "bbox": [ + 511, + 308, + 905, + 338 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {d i s} = \\left\\| \\hat {g} _ {i} ^ {t} - g _ {i g t} ^ {t} \\right\\| \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 642, + 351, + 905, + 371 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For direction, the final moment of the historical period is introduced. Each predicted time step is related to the preceding time step to construct the direction vector, and subsequently, the angle between the predicted trajectory and the actual trajectory direction vector is analyzed.", + "bbox": [ + 511, + 383, + 905, + 458 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {a n g l e} = - \\operatorname {a r c c o s} \\left(\\frac {\\left(\\hat {g} _ {i} ^ {t} - \\hat {g} _ {i} ^ {t + 1}\\right) \\cdot \\left(g _ {i g t} ^ {t} - g _ {i g t} ^ {t + 1}\\right)}{\\left\\| \\hat {g} _ {i} ^ {t} - \\hat {g} _ {i} ^ {t + 1} \\right\\| \\left\\| g _ {i g t} ^ {t} - g _ {i g t} ^ {t + 1} \\right\\|}\\right) \\tag {17}\n$$\n", + "text_format": "latex", + "bbox": [ + 522, + 470, + 905, + 513 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The objective of our model is to maximize the lower bound on the evidence for all time steps of the target distribution, joints (16) and (17), with a final loss function of:", + "bbox": [ + 511, + 525, + 905, + 571 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} L = \\sum_ {t = T _ {H} + 1} ^ {T _ {F}} \\left\\{\\mathbb {E} _ {z _ {i \\varphi} ^ {t} \\sim q _ {\\varphi} \\left(\\cdot | B _ {i} ^ {t}, I _ {i} ^ {t - 1}\\right)} \\left[ L _ {d i s} + L _ {a n g l e} \\right] \\right. \\tag {18} \\\\ \\left. - K L \\left[ q _ {\\varphi} \\left(z _ {i \\varphi} ^ {t} \\mid B _ {i} ^ {t}, I _ {i} ^ {t - 1}\\right) \\| q _ {\\vartheta} \\left(z _ {i \\vartheta} ^ {t} \\mid I _ {i} ^ {t - 1}\\right) \\right] \\right\\} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 529, + 583, + 905, + 650 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 513, + 662, + 645, + 680 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. ETH [32] and UCY [18] datasets encompass over 1,600 pedestrian trajectories, with positions annotated at 0.4s intervals. SDD [33] dataset is primarily utilized for pedestrian trajectory prediction and object tracking, containing about 19,000 pedestrian trajectories, with positions annotated every 0.4s. NBA dataset focuses on the 2015-2016 season [21, 53]. Two subsets, \"Rebounds\" and \"Scores,\" are extracted for benchmarking purposes. The movement intentions of the agents exhibit strong purposefulness and complex, hard-to-capture velocity variations. Consequently, the extracted trajectories encompass rich, highly nonlinear motion patterns. NuScenes [4] is a dataset for autonomous driving, and the prediction challenge segmentation and configurations are adhered to.", + "bbox": [ + 511, + 688, + 906, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "22469", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/0d3d8cf87453daa9376edc3a5958150bd2bb5e1fb04fe85a89246de32448f3ff.jpg", + "table_caption": [ + "Table 1. Comparison of Related Methods on ETH/UCY, NBA, SDD, and NuScenes (in meters). Every work was realized using best-of-20 predictions for the effect of ade/fde was selected." + ], + "table_footnote": [], + "table_body": "
ETH/UCYFLEAM [6]ST-MR [19]MTN [48]MANTRA [28]Agent-former [51]BiTrap [47]MemoNet [45]EqMotion [46]V2-Net-SC [39, 41]E-V2-Net-SC [40, 41]Ours
Eth0.66/1.210.53/0.690.58/0.860.56/0.970.58/0.860.66/1.090.51/0.690.36/0.540.23/0.330.23/0.300.26/0.31
Hotel0.37/0.640.20/0.310.19/0.570.17/0.320.15/0.240.17/0.320.14/0.200.12/0.160.11/0.160.10/0.130.10/0.12
Univ0.57/0.970.41/0.640.32/0.430.38/0.790.26/0.480.26/0.460.26/0.520.21/0.390.17/0.270.18/0.240.11/0.17
Zara010.46/0.740.26/0.380.24/0.370.24/0.410.19/0.320.23/0.470.21/0.360.16/0.270.13/0.180.13/0.160.10/0.17
Zara020.37/0.690.18/0.240.15/0.210.18/0.340.16/0.250.18/0.350.15/0.300.11/0.190.12/0.210.11/0.160.09/0.14
Avg0.48/0.850.32/0.450.30/0.490.31/0.570.27/0.430.30/0.540.25/0.410.19/0.310.15/0.230.15/0.200.13/0.18
NBAFLEAM [6]ST-MR [19]MTN [48]SGNet [38]EqMotion [46]MemoNet [45]BiTrap [47]GroupNet +CVAE [44]V2-Net-SC [39, 41]E-V2-Net-SC [40, 41]Ours
Rebound1.92/2.210.94/1.730.84/0.970.80/1.780.86/1.211.73/2.060.86/1.880.63/0.840.54/0.790.54/0.860.34/0.66
Scores1.29/1.680.90/1.030.87/1.410.71/1.450.69/1.021.35/1.870.76/1.550.46/0.760.50/0.840.47/0.780.30/0.56
SDDFLEAM [6]SHENet [29]MemoNet [45]MANTRA [28]BiTrap [47]MID [10]Y-Net [26]NSP-SFM [52]V2-Net-SC [39, 41]E-V2-Net-SC [40, 41]Ours
SDD0.67/1.280.44/0.530.45/0.640.49/0.620.46/0.590.38/0.730.32/0.660.21/0.340.26/0.360.23/0.350.17/0.24
NuScenesFLEAM [6]ST-MR [19]MTN [48]MANTRA [28]SGNet [38]Agent-former [51]EqMotion [46]Y-Net [26]V2-Net-SC [39, 41]E-V2-Net-SC [40, 41]Ours
NuScenes2.02/3.141.84/2.461.64/2.121.47/2.011.28/2.091.04/1.471.24/1.981.12/1.711.06/1.591.04/1.640.92/1.56
", + "bbox": [ + 91, + 128, + 906, + 297 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Metrics. Best-of-20 predictions were used to compute Average Displacement Error (ADE) and Final Displacement Error (FDE) as primary metrics, with estimated Negative Log Likelihood (NLL) reported from the test set. See their detailed definitions in the Supplementary Material.", + "bbox": [ + 89, + 321, + 480, + 398 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details. The model was trained on two NVIDIA RTX 4090 GPUs. Following the methodology in [5, 36], a leave-one-out cross-validation approach was employed for dataset handling. The number of subspaces in the higher-order intention interaction layer was set to $M = 6$ while the depth of the global trajectory optimizer was set to $L = 3$ . During inference, the best parameter model was loaded to predict the next 8 frames based on the previous 8 frames of history. For more implementation details of the training in the Supplementary Material.", + "bbox": [ + 88, + 398, + 482, + 550 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Quantitative analysis", + "text_level": 1, + "bbox": [ + 89, + 561, + 290, + 577 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "All SOTA performances across datasets were highlighted in bold black in Table 1. In Eth scenario, crowds generally moved in parallel, resulting in weaker interactions among neighbors that conducted to target agent. E-V2-Net-SC [40, 41] achieved optimal performance in this scenario. However, the proposed method still achieved superior performance on ETH/UCY dataset, reducing ADE and FDE by $13.33\\%$ and $10.00\\%$ compared to E-V2-Net-SC, respectively. For NBA dataset, where interactions between agents were complex and frequent, the proposed method demonstrated exceptional performance. In Rebound sub-dataset, ADE and FDE decreased by $37.04\\%$ and $16.46\\%$ , respectively, compared to existing SOTA methods. In Scores sub-dataset, ADE and FDE were reduced by $34.78\\%$ and $26.32\\%$ , respectively, relative to the current SOTA methods. For SDD dataset, image segmentation techniques, as referenced in [2], were employed to capture pedestrian position information across eight scenario. Raw pixel coordinates were converted into spatial coordinates defined in meters for trajectory prediction tasks. NSP-SFM [52] outperformed both V2-Net-SC and E-V2-Net-SC. Our method re", + "bbox": [ + 91, + 583, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "duced ADE and FDE by $19.05\\%$ and $29.41\\%$ , respectively, compared to NSP-SFM. For NuScenes dataset, V2-Net-SC and E-V2-Net-SC showed similar performance levels. Our method reduced ADE and FDE by $11.54\\%$ and $1.87\\%$ , respectively, in comparison to their SOTA performances.", + "bbox": [ + 511, + 321, + 906, + 398 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Detailed comparative experiments with NLL metric were carried out in the Supplementary Material. Furthermore, for wider validation and analysis, we applied directional component in the designed loss function to different models. See the Supplementary Material for details.", + "bbox": [ + 511, + 398, + 906, + 474 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Qualitative analysis", + "text_level": 1, + "bbox": [ + 511, + 487, + 700, + 503 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5e993945fc8199ac2be38e1d22beafcd37dcf953f777dfad5392d7a7801e8b21.jpg", + "image_caption": [ + "Figure 3. Representative cases. Row 1 represented complex historical trajectory cases. Row 2 represented state maintenance cases. Row 3 represented numerous neighbors cases." + ], + "image_footnote": [], + "bbox": [ + 517, + 525, + 901, + 795 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Unlike previous studies [6, 26, 39-41, 45, 46], the focus here was on conducting experiments involving historical", + "bbox": [ + 511, + 869, + 906, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "22470", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4d03a52220cac1415396f477349406fc0a9f6c7b81ce2c3e7193d1868c4994d0.jpg", + "image_caption": [ + "Figure 4. Predictions comparison of predictions in ETH/UCY. The comparison included heatmaps of the overall distribution of predictions and best-of-20 predictions of the related methods." + ], + "image_footnote": [], + "bbox": [ + 181, + 90, + 815, + 378 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "trajectories in various social contexts. This enabled a more effective validation of the model's ability to extract and analyze multi-order fused intentions in Fig. 3. In panel (a), the target agent continuously changed direction. Although the model displayed slight deviations at certain moments, these deviations remained within the acceptable range of directional changes. In panel (b) and (c), the target agent intended to turn, but the model successfully achieved accurate predictions. The target agent moved at a relatively steady speed in panel (d), adopting a conservative deceleration strategy in response to numerous neighbors who are stopping or remaining stationary. In panel (e) and (f), there were many stationary neighbors who may be gathered together in conversation. Consequently, the target agent maintained its original state. In panel (g) and (h), most neighbors moved in the same direction as the target agent, with closer neighbors receiving higher attention scores. This suggested that the target agent prioritized attention to closer neighbors, consistent with social norms. In panel (i), the directions of neighbors were complex, and their speeds varied. The target agent accounted for influences from neighbors in multiple directions while attempting to anticipate and avoid imminent collisions. Please see the qualitative analyses of NBA interactions in the Supplementary Material.", + "bbox": [ + 91, + 443, + 483, + 806 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparisons were made between the proposed method and three previous methods in Fig. 4. It was observed that our method produced more accurate predictions than the previous methods. In Eth, when facing neighbors to the left front, the model provided a prediction closest to the true trajectory. The predictions from E-V2-Net-SC and EqMotion", + "bbox": [ + 89, + 810, + 483, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "exhibited slight oscillations, while MemoNet, by mimicking retrospective memory mechanisms, tends to maintain its original trajectory. By employing the squeeze theorem to approximate future true trajectories, the proposed model shows greater convergence in the overall distribution of predictions in all scenarios. In Hotel and Zara01, there were many pedestrians ahead of the target agent, and the diverging trajectory distributions from previous methods represent rather aggressive strategies, which conflicted with the safety requirements of autonomous driving.", + "bbox": [ + 511, + 443, + 906, + 595 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation study", + "text_level": 1, + "bbox": [ + 511, + 606, + 663, + 622 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In Table 2, ablation experiments were conducted by sequentially adding components to the PDI baseline. Groups 1-4 focus on the multi-order intention fusion module. Group 1 includes only the neighbors' position and the distances between the target agent and the neighbors. As additional components, such as neighboring speed $\\mathbf{V}$ , the angle between the target agent and the neighboring speed $\\Theta$ , and the final predicted distance over the historical period $\\mathbf{E}$ , were gradually incorporated, a substantial reduction in the final prediction error was observed. Specifically, ADE decreased by $23.91\\%$ and FDE by $39.47\\%$ on the SDD dataset, while ADE reduced by $11.11\\%$ and FDE by $14.42\\%$ on the Nuscenes dataset. Groups 4-5, 5-6, and 6-7 focus on the ablation of the approximator, the optimizer, and the angle loss term in the loss function, respectively. The results demonstrate that each component contributed positively.", + "bbox": [ + 511, + 628, + 906, + 869 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To examine the influence of $\\mathbf{B}$ and $\\mathbf{I}$ , the trajectory distribution for several scenarios in the future periods was visu-", + "bbox": [ + 511, + 869, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "22471", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/9bb161f19576b3106e794f05a4f9e59f79a7801fc7c8ad2e21afa388a8b28379.jpg", + "table_caption": [ + "Table 2. Ablation studies of each component in our method on SDD and Nuscenes datasets. Notably, the positional information of the target agent was essential and was included in each set of ablation experiments. $\\mathbf{P}$ represented the position of neighbors, $\\mathbf{V}$ indicated their velocities, $\\mathbf{D}$ denoted the distance between the target agent and its neighbors, and $\\Theta$ represented the angle between their velocities. $\\mathbf{E}$ represented the final predicted distance between the target agent and its neighbors over historical periods, $\\mathbf{I}$ indicated global historical social intent, $\\mathbf{B}$ denoted future actual trajectories, $\\mathbf{K}$ referred to the parallel prediction mode, and $\\mathbf{A}$ signified directional information." + ], + "table_footnote": [], + "table_body": "
GroupPVDΘEIBKASDDNuscenes
1✓✓✓0.46/0.711.26/2.15
2✓✓✓✓0.45/0.661.24/2.09
3✓✓✓✓✓0.41/0.561.18/1.97
4✓✓✓✓✓✓0.35/0.431.12/1.84
5✓✓✓✓✓✓✓0.30/0.361.04/1.76
6✓✓✓✓✓✓✓✓0.20/0.270.93/1.61
7✓✓✓✓✓✓✓✓✓0.17/0.240.92/1.56
", + "bbox": [ + 151, + 169, + 848, + 275 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2706826b04e29baddbbc288d435acaf6cfecad63c3014a316e2d5bd7ee82c939.jpg", + "image_caption": [ + "Figure 5. Distribution heatmaps. Every heatmap of future trajectory distribution was shown above the real scene image. (a) represented the actual distribution. (b) presented the experimental group containing only I component. (c) showed the group with both I and B components." + ], + "image_footnote": [], + "bbox": [ + 181, + 290, + 821, + 502 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4765ac324e41e41878fe89a0229ce6f845a252c2fd285776c2ed00f3336e9b65.jpg", + "image_caption": [ + "Figure 6. Representative cases about the optimization effects of component $\\mathbf{K}$ . To ensure fairness, we selected best-of-20 predictions in groups 6 and 7 for comparison." + ], + "image_footnote": [], + "bbox": [ + 109, + 568, + 460, + 723 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "alized in Fig. 5. It was evident that when both components $\\mathbf{B}$ and $\\mathbf{I}$ were included, the predicted trajectory distribution aligned more closely with the actual distribution. It can be found that in a strongly regular sports scenario such as Rebound, (b) and (c) achieved excellent predictions, while in open environments with complex social contexts, the effec", + "bbox": [ + 89, + 809, + 483, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "tiveness of the approximator was better demonstrated.", + "bbox": [ + 511, + 568, + 870, + 583 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To validate the effectiveness of component $\\mathbf{K}$ under global parallel optimization, six challenging cases were presented in Fig. 6. It was evident that predictions without component $\\mathbf{K}$ exhibited increasing errors as the prediction time accumulated, whereas the parallel prediction mode demonstrated a closer result to ground-truth.", + "bbox": [ + 511, + 584, + 906, + 675 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 691, + 633, + 708 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Each human movement path is either directly or indirectly associated with neighbor agents. This study introduces SocialMOIF, which provides a more comprehensive representation of both direct and indirect intention interactions between the target agent and neighbor agents. SocialMOIF comprises four key designs: a multi-order intention fusion model, a trajectory distribution approximator, a global trajectory optimizer, and a novel loss function. Experimental results demonstrate that this approach performs competitively across various datasets and effectively approximates and explicitly updates trajectory distributions during training, thereby improving interpretability.", + "bbox": [ + 511, + 719, + 906, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "22472", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 91, + 90, + 250, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This project is supported by National Natural Science Foundation of China (52202417), China Postdoctoral Science Foundation(2022TQ0155, 2022M721605), Open Project Program of State Key Laboratory of Virtual Reality Technology and Systems, Beihang University(VRLAB2023A02), Young Elite Scientists Sponsorship Program by CAST(2023QNRC001), Fundamental Research Funds for the Central Universities (NS2024030), Young Elite Scientists Sponsorship Program by JSTJ(JSTJ-2023-XH032).", + "bbox": [ + 89, + 114, + 483, + 265 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 277, + 187, + 292 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Alexandre Alahi, Kratarth Goel, Vignesh Ramanathan, Alexandre Robicquet, Li Fei-Fei, and Silvio Savarese. Social lstm: Human trajectory prediction in crowded spaces. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 961–971, 2016. 2, 3", + "[2] Stefan Becker, Ronny Hug, Wolfgang Hubner, and Michael Arens. An evaluation of trajectory prediction approaches and notes on the trajnet benchmark. arXiv preprint arXiv:1805.07663, 2018. 6", + "[3] Graeme Best and Robert Fitch. Bayesian intention inference for trajectory prediction with an unknown goal destination. In 2015 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 5817-5823. IEEE, 2015. 2", + "[4] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11621-11631, 2020. 5", + "[5] Kai Chen, Xiao Song, and Xiaoxiang Ren. Pedestrian trajectory prediction in heterogeneous traffic using pose keypoints-based convolutional encoder-decoder network. IEEE Transactions on Circuits and Systems for Video Technology, 31(5): 1764-1775, 2020. 2, 6", + "[6] Kai Chen, Xiao Song, Haitao Yuan, and Xiaoxiang Ren. Fully convolutional encoder-decoder with an attention mechanism for practical pedestrian trajectory prediction. IEEE Transactions on Intelligent Transportation Systems, 23(11): 20046-20060, 2022. 2, 6", + "[7] Kai Chen, Haihua Zhu, Dunbing Tang, and Kun Zheng. Future pedestrian location prediction in first-person videos for autonomous vehicles and social robots. Image and Vision Computing, 134:104671, 2023. 2", + "[8] Lingwei Dang, Yongwei Nie, Chengjiang Long, Qing Zhang, and Guiqing Li. Msr-gen: Multi-scale residual graph convolution networks for human motion prediction. In Proceedings of the IEEE/CVF international conference on computer vision, pages 11467-11476, 2021. 3", + "[9] Francesco Giuliani, Irtiza Hasan, Marco Cristani, and Fabio Galasso. Transformer networks for trajectory forecasting. In 2020 25th international conference on pattern recognition (ICPR), pages 10335-10342. IEEE, 2021. 2" + ], + "bbox": [ + 99, + 303, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[10] Tianpei Gu, Guangyi Chen, Junlong Li, Chunze Lin, Yongming Rao, Jie Zhou, and Jiwen Lu. Stochastic trajectory prediction via motion indeterminacy diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17113-17122, 2022. 3, 6", + "[11] Agrim Gupta, Justin Johnson, Li Fei-Fei, Silvio Savarese, and Alexandre Alahi. Social gan: Socially acceptable trajectories with generative adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2255–2264, 2018. 2, 3", + "[12] Dirk Helbing and Peter Molnar. Social force model for pedestrian dynamics. Physical review E, 51(5):4282, 1995. 2", + "[13] Serge Hoogendoorn and Piet HL Bovy. Simulation of pedestrian flows by optimal control and differential games. Optimal control applications and methods, 24(3):153-172, 2003. 2", + "[14] SP Hoogendoorn, Winnie Daamen, Yaqing Shu, and Han Ligteringen. Modeling human behavior in vessel maneuver simulation by optimal control and game theory. Transportation research record, 2326(1):45-53, 2013. 2", + "[15] Yihan Hu, Jiazhi Yang, Li Chen, Keyu Li, Chonghao Sima, Xizhou Zhu, Siqi Chai, Senyao Du, Tianwei Lin, Wenhai Wang, et al. Planning-oriented autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17853-17862, 2023. 1", + "[16] Sungjune Kim, Hyung-gun Chi, Hyerin Lim, Karthik Ramani, Jinkyu Kim, and Sangpil Kim. Higher-order relational reasoning for pedestrian trajectory prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15251-15260, 2024. 2", + "[17] Christos Kyrkou, Panayiotis Kolios, Theocharis Theocharides, and Marios Polycarpou. Machine learning for emergency management: A survey and future outlook. Proceedings of the IEEE, 111(1):19-41, 2022. 5", + "[18] Alon Lerner, Yiorgos Chrysanthou, and Dani Lischinski. Crowds by example. In Computer graphics forum, pages 655-664. Wiley Online Library, 2007. 5", + "[19] Lihuan Li, Maurice Pagnucco, and Yang Song. Graph-based spatial transformer with memory replay for multifuture pedestrian trajectory prediction. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2221-2231, 2022. 2, 6", + "[20] Linhui Li, Xiaotong Lin, Yejia Huang, Zizhen Zhang, and Jian-Fang Hu. Beyond minimum-of-n: Rethinking the evaluation and methods of pedestrian trajectory prediction. IEEE Transactions on Circuits and Systems for Video Technology, 2024. 1", + "[21] Kostya Linou. Nba-player-movements. Software for Visualization of NBA Games, 2016. 5", + "[22] Congcong Liu, Yuying Chen, Ming Liu, and Bertram E Shi. Avgcn: Trajectory prediction using graph convolutional networks guided by human attention. In 2021 IEEE International Conference on Robotics and Automation (ICRA), pages 14234-14240. IEEE, 2021. 3", + "[23] Ziming Liu, Pingchuan Ma, Yixuan Wang, Wojciech Ma-tusik, and Max Tegmark. Kan 2.0: Kolmogorov-arnold net" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "22473", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "works meet science. arXiv preprint arXiv:2408.10205, 2024. 2", + "[24] Ziming Liu, Yixuan Wang, Sachin Vaidya, Fabian Ruehle, James Halverson, Marin Soljačić, Thomas Y Hou, and Max Tegmark. Kan: Kolmogorov-arnold networks. arXiv preprint arXiv:2404.19756, 2024. 2", + "[25] Karttikeya Mangalam, Harshayu Girase, Shreyas Agarwal, Kuan-Hui Lee, Ehsan Adeli, Jitendra Malik, and Adrien Gaidon. It is not the journey but the destination: Endpoint conditioned trajectory prediction. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16, pages 759-776. Springer, 2020. 2", + "[26] Karttikeya Mangalam, Yang An, Harshayu Girase, and Jitendra Malik. From goals, waypoints & paths to long term human trajectory forecasting. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15233–15242, 2021. 1, 2, 6", + "[27] Weibo Mao, Chenxin Xu, Qi Zhu, Siheng Chen, and Yanfeng Wang. Leapfrog diffusion model for stochastic trajectory prediction. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5517-5526, 2023. 3", + "[28] Francesco Marchetti, Federico Becattini, Lorenzo Seidenari, and Alberto Del Bimbo. *Mantra: Memory augmented networks for multiple trajectory prediction*. In *Proceedings of the IEEE/CVF conference on computer vision and pattern recognition*, pages 7143–7152, 2020. 2, 6", + "[29] Mancheng Meng, Ziyan Wu, Terrence Chen, Xiran Cai, Xiang Zhou, Fan Yang, and Dinggang Shen. Forecasting human trajectory from scene history. arXiv preprint arXiv:2210.08732, 2022. 1, 6", + "[30] Alessio Monti, Alessia Bertugli, Simone Calderara, and Rita Cucchiara. Dag-net: Double attentive graph neural network for trajectory forecasting. In 2020 25th international conference on pattern recognition (ICPR), pages 2551-2558. IEEE, 2021. 3", + "[31] Brendan Tran Morris and Mohan Manubhai Trivedi. Trajectory learning for activity understanding: Unsupervised, multilevel, and long-term adaptive approach. IEEE transactions on pattern analysis and machine intelligence, 33(11):2287-2301, 2011. 2", + "[32] Stefano Pellegrini, Andreas Ess, Konrad Schindler, and Luc Van Gool. You'll never walk alone: Modeling social behavior for multi-target tracking. In 2009 IEEE 12th international conference on computer vision, pages 261-268. IEEE, 2009. 5", + "[33] Alexandre Robicquet, Amir Sadeghian, Alexandre Alahi, and Silvio Savarese. Learning social etiquette: Human trajectory prediction in crowded scenes. In European Conference on Computer Vision (ECCV), page 5, 2016. 5", + "[34] Tim Salzmann, Boris Ivanovic, Punarjay Chakravarty, and Marco Pavone. Trajectory: Dynamically-feasible trajectory forecasting with heterogeneous data. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XVIII 16, pages 683-700. Springer, 2020. 2" + ], + "bbox": [ + 91, + 90, + 482, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[35] Liushuai Shi, Le Wang, Chengjiang Long, Sanping Zhou, Mo Zhou, Zhenxing Niu, and Gang Hua. Sgcn: Sparse graph convolution network for pedestrian trajectory prediction. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8994-9003, 2021. 3", + "[36] Xiao Song, Kai Chen, Xu Li, Jinghan Sun, Baocun Hou, Yong Cui, Baochang Zhang, Gang Xiong, and Zilie Wang. Pedestrian trajectory prediction based on deep convolutional LSTM network. IEEE Transactions on Intelligent Transportation Systems, 22(6):3285-3302, 2020. 2, 6", + "[37] Yuchao Su, Jie Du, Yuanman Li, Xia Li, Rongqin Liang, Zhongyun Hua, and Jiantao Zhou. Trajectory forecasting based on prior-aware directed graph convolutional neural network. IEEE Transactions on Intelligent Transportation Systems, 23(9):16773-16785, 2022. 3", + "[38] Chuhua Wang, Yuchen Wang, Mingze Xu, and David J Crandall. Stepwise goal-driven networks for trajectory prediction. IEEE Robotics and Automation Letters, 7(2):2716-2723, 2022. 2, 6", + "[39] Conghao Wong, Beihao Xia, Ziming Hong, Qinmu Peng, Wei Yuan, Qiong Cao, Yibo Yang, and Xinge You. View vertically: A hierarchical network for trajectory prediction via fourier spectrums. In European Conference on Computer Vision, pages 682-700. Springer, 2022. 2, 6", + "[40] Conghao Wong, Beihao Xia, Qinmu Peng, and Xinge You. Another vertical view: A hierarchical network for heterogeneous trajectory prediction via spectrums. arXiv preprint arXiv:2304.05106, 2023. 2, 6", + "[41] Conghao Wong, Beihao Xia, Ziqian Zou, Yulong Wang, and Xinge You. Socialcircle: Learning the angle-based social interaction representation for pedestrian trajectory prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19005-19015, 2024. 1, 2, 6", + "[42] Beihao Xia, Conghao Wong, Qinmu Peng, Wei Yuan, and Xinge You. Cscnet: Contextual semantic consistency network for trajectory prediction in crowded spaces. Pattern Recognition, 126:108552, 2022. 3", + "[43] Song Xiao, Kai Chen, Xiaoxiang Ren, and Haitao Yuan. Pedestrian trajectory prediction in heterogeneous traffic using facial keypoints-based convolutional encoder-decoder network. ACM Transactions on Internet Technology, 22(4): 1-14, 2022. 2", + "[44] Chenxin Xu, Maosen Li, Zhenyang Ni, Ya Zhang, and Siheng Chen. Groupnet: Multiscale hypergraph neural networks for trajectory prediction with relational reasoning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6498-6507, 2022. 3, 6", + "[45] Chenxin Xu, Weibo Mao, Wenjun Zhang, and Siheng Chen. Remember intentions: Retrospective-memory-based trajectory prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6488-6497, 2022. 1, 6", + "[46] Chenxin Xu, Robby T Tan, Yuhong Tan, Siheng Chen, Yu Guang Wang, Xinchao Wang, and Yanfeng Wang. Eqmotion: Equivariant multi-agent motion prediction with invari" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "22474", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ant interaction reasoning. arXiv preprint arXiv:2303.10876, 2023. 2, 6", + "[47] Yu Yao, Ella Atkins, Matthew Johnson-Roberson, Ram Vasudevan, and Xiaoxiao Du. Bitrap: Bi-directional pedestrian trajectory prediction with multi-modal goal estimation. IEEE Robotics and Automation Letters, 6(2):1463-1470, 2021. 2, 6", + "[48] Ziyi Yin, Ruijin Liu, Zhiliang Xiong, and Zejian Yuan. Multimodal transformer networks for pedestrian trajectory prediction. In Proceedings of the Twenty-Ninth International Joint Conference on Artificial Intelligence, pages 1259-1265, 2021. 2, 3, 6", + "[49] Cunjun Yu, Xiao Ma, Jiawei Ren, Haiyu Zhao, and Shuai Yi. Spatio-temporal graph transformer networks for pedestrian trajectory prediction. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XII 16, pages 507-523, 2020. 2, 3", + "[50] Wenjian Yu and Dirk Helbing. Game theoretical interactions of moving agents. In Simulating Complex Systems by Cellular Automata, pages 219-239. Springer, 2010. 2", + "[51] Ye Yuan, Xinshuo Weng, Yanglan Ou, and Kris M Kitani. Agentformer: Agent-aware transformers for socio-temporal multi-agent forecasting. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9813-9823, 2021. 2, 6", + "[52] Jiangbei Yue, Dinesh Manocha, and He Wang. Human trajectory prediction via neural social physics. In European conference on computer vision, pages 376-394, 2022. 3, 6", + "[53] Yisong Yue, Patrick Lucey, Peter Carr, Alina Bialkowski, and Iain Matthews. Learning fine-grained spatial models for dynamic sports play prediction. In 2014 IEEE international conference on data mining, pages 670-679. IEEE, 2014. 5", + "[54] Pu Zhang, Wanli Ouyang, Pengfei Zhang, Jianru Xue, and Nanning Zheng. Sr-lstm: State refinement for LSTM towards pedestrian trajectory prediction. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12085-12094, 2019. 2", + "[55] Xiaoping Zheng and Yuan Cheng. Conflict game in evacuation process: A study combining cellular automata model. Physica A: Statistical Mechanics and its Applications, 390 (6):1042-1050, 2011. 2" + ], + "bbox": [ + 91, + 90, + 482, + 670 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "22475", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/SocialMOIF_ Multi-Order Intention Fusion for Pedestrian Trajectory Prediction/74c8547f-5c13-42b2-b27c-77ef8f52058e_model.json b/2025/SocialMOIF_ Multi-Order Intention Fusion for Pedestrian Trajectory Prediction/74c8547f-5c13-42b2-b27c-77ef8f52058e_model.json new file mode 100644 index 0000000000000000000000000000000000000000..94254548c2cfab1f4976da95e0e828efdf774c43 --- /dev/null +++ b/2025/SocialMOIF_ Multi-Order Intention Fusion for Pedestrian Trajectory Prediction/74c8547f-5c13-42b2-b27c-77ef8f52058e_model.json @@ -0,0 +1,2246 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.131, + 0.905, + 0.153 + ], + "angle": 0, + "content": "SocialMOIF: Multi-Order Intention Fusion for Pedestrian Trajectory Prediction" + }, + { + "type": "text", + "bbox": [ + 0.245, + 0.18, + 0.751, + 0.217 + ], + "angle": 0, + "content": "Kai Chen\\(^{1*}\\), Xiaodong Zhao\\((\\boxtimes)\\)^{1*}, Yujie Huang\\(^{1}\\), Guoyu Fang\\(^{1}\\), Xiao Song\\(^{2}\\), Ruiping Wang\\(^{3}\\), Ziyuan Wang\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.286, + 0.217, + 0.714, + 0.234 + ], + "angle": 0, + "content": "1 Nanjing University of Aeronautics and Astronautics" + }, + { + "type": "text", + "bbox": [ + 0.269, + 0.234, + 0.729, + 0.253 + ], + "angle": 0, + "content": "2 Beihang University 3 Nanyang Technological University" + }, + { + "type": "text", + "bbox": [ + 0.253, + 0.255, + 0.746, + 0.27 + ], + "angle": 0, + "content": "{chen_kai, xdzhao, yj_huang, fanggy, zy_wang}@nuaa.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.297, + 0.273, + 0.698, + 0.287 + ], + "angle": 0, + "content": "songxiao@buaa.edu.cn, ruiping.wang@ntu.edu.sg" + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.337, + 0.327, + 0.354 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.369, + 0.485, + 0.761 + ], + "angle": 0, + "content": "The analysis and prediction of agent trajectories are crucial for decision-making processes in intelligent systems, with precise short-term trajectory forecasting being highly significant across a range of applications. Agents and their social interactions have been quantified and modeled by researchers from various perspectives; however, substantial limitations exist in the current work due to the inherent high uncertainty of agent intentions and the complex higher-order influences among neighboring groups. SocialMOIF is proposed to tackle these challenges, concentrating on the higher-order intention interactions among neighboring groups while reinforcing the primary role of first-order intention interactions between neighbors and the target agent. This method develops a multi-order intention fusion model to achieve a more comprehensive understanding of both direct and indirect intention information. Within SocialMOIF, a trajectory distribution approximator is designed to guide the trajectories toward values that align more closely with the actual data, thereby enhancing model interpretability. Furthermore, a global trajectory optimizer is introduced to enable more accurate and efficient parallel predictions. By incorporating a novel loss function that accounts for distance and direction during training, experimental results demonstrate that the model outperforms previous state-of-the-art baselines across multiple metrics in both dynamic and static datasets." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.792, + 0.222, + 0.808 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.818, + 0.484, + 0.866 + ], + "angle": 0, + "content": "Trajectory prediction is of considerable importance for intelligent transportation and surveillance systems. Studies indicate that trajectories derived from the physical" + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.325, + 0.898, + 0.572 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.586, + 0.908, + 0.657 + ], + "angle": 0, + "content": "Figure 1. The process of SocialMOIF includes: (1) extracting the position, velocity and trajectory information of agents, (2) achieving multi-order fused intention modeling, (3) guiding the distribution of trajectories and performing global optimization. Finally we obtain accurate parallel predictions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.687, + 0.907, + 0.793 + ], + "angle": 0, + "content": "world encompass rich spatiotemporal intention information [15, 20, 26, 29, 45]. Therefore, a significant challenge arises in effectively capturing (data modeling) and analyzing (encoding-decoding) processes related to the complex and diverse historical social intentions of agents, while linking this information to future trajectory predictions for enhanced efficiency." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.909, + 0.903 + ], + "angle": 0, + "content": "History social intentions. Existing research can be broadly classified into two main categories: knowledge-driven methods primarily depend on expert knowledge. For instance, biomimetic method [41] achieved agent localization and intention interaction modeling by simulating the positioning of marine animals. In contrast, certain studies utilized game-theoretic approaches to analyze" + }, + { + "type": "page_footnote", + "bbox": [ + 0.114, + 0.876, + 0.23, + 0.888 + ], + "angle": 0, + "content": "* Equal contribution." + }, + { + "type": "page_footnote", + "bbox": [ + 0.114, + 0.888, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Code is available at https://github.com/XiaodZhao/SocialMOIF." + }, + { + "type": "list", + "bbox": [ + 0.114, + 0.876, + 0.469, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22465" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.453 + ], + "angle": 0, + "content": "social intentions through the simulation of agent flows [13, 50] and evacuation processes [14, 55], relying on explicit rules or formulas. However, the behavioral patterns and interaction rules among agents in social scenarios displayed high nonlinearity and uncertainty [26, 39, 40]. Data-driven methods implicitly learn interactions among agents through neural networks. For instance, LSTM-based methods [1, 11] continuously refined social pooling patterns, while attention-based approaches [7, 49, 51] dynamically adjusted agents' focal points in global interactions from a spatiotemporal integration perspective to optimize predicted trajectories. These methods primarily focused on straightforward interactions or direct responses between the target agent and its neighbors, often neglecting the higher-order, multidimensional influences exerted by surrounding agents on the target agent's intentions. Kim et al. [16] modeled agent intentions from 1st-order to N-order equally. However, when the number of agents was large, it may prioritized higher-order intentions, thereby diminishing the primary impact of lower-order intentions. Thus, the development of a multi-order fusion network that reinforces the dominant role of direct interactions while considering the secondary influence of higher-order transmissions on the target is essential." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.458, + 0.483, + 0.653 + ], + "angle": 0, + "content": "Associating future trajectory information. Knowledge-driven methods face limitations due to the inherent flaws of unidirectional inference, which hinder effective association with future trajectory information. Data-driven methods exhibit two main shortcomings: first, agent movements are temporally dynamic, and relying solely on distance loss fails to capture these dynamic variations. Second, in generative model approaches [25, 28, 34, 38, 47], future trajectory information is implicitly involved in the training process without explicit guidance, resulting in poor interpretability. In safety-sensitive tasks such as autonomous driving, it is necessary to utilize meaningful and goal-oriented parameters and models." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.657, + 0.482, + 0.822 + ], + "angle": 0, + "content": "Efficiency. Existing methods predominantly emphasize upstream tasks. For example, LSTM-based approaches [5, 6, 36, 43, 54] modeled trajectory data in a serial temporal manner, while Transformer-based methods [9, 19, 48, 49] analyzed trajectory information through parallel processing. However, in downstream tasks, trajectories are often generated independently and sequentially. This method not only accumulates errors but also limits the efficiency of trajectory prediction. Therefore, establishing a trajectory parallel prediction model that accounts for global influences across future periods is a promising area for further exploration." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.482, + 0.901 + ], + "angle": 0, + "content": "As illustrated in Fig. 1, SocialMOIF is proposed to address the trajectory prediction task. A Multi-Order Intention Fusion model (MOIF) is introduced to guide the learning direction of the network. MOIF includes a first-order intention interaction layer and a higher-order intention in" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.363 + ], + "angle": 0, + "content": "teraction layer. The first layer focuses on direct interaction information between neighbors and the target agent, while the second layer addresses the indirect interaction information conveyed to the target agent after multiple interactions among neighbor targets. Inspiration is drawn from the squeeze theorem to design a trajectory distribution approximator to guide trajectories toward a distribution that closely aligns with ground truth and to enhance model interpretability. Simultaneously, RNN is incorporated to update the latent variables, achieving explicit supervision during training. The temporal dimension of decoded trajectories is aggregated to design a global trajectory optimizer. To the best of our knowledge, KANs [23, 24] are firstly introduced for the global optimization of trajectories, enabling more accurate and efficient parallel predictions. Furthermore, considering the travel directions of the agents, a distance-direction fused loss function is designed to comprehensively supervise the dynamic states of agents." + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.365, + 0.884, + 0.379 + ], + "angle": 0, + "content": "In summary, the main contributions of this paper are:" + }, + { + "type": "text", + "bbox": [ + 0.521, + 0.379, + 0.905, + 0.408 + ], + "angle": 0, + "content": "- The proposal of SocialMOIF, which incorporates four novel designs:" + }, + { + "type": "text", + "bbox": [ + 0.538, + 0.409, + 0.905, + 0.453 + ], + "angle": 0, + "content": "(1) MOIF is introduced to comprehensively capture interaction information between the target agent and its neighbors;" + }, + { + "type": "text", + "bbox": [ + 0.539, + 0.453, + 0.905, + 0.514 + ], + "angle": 0, + "content": "(2) A trajectory distribution approximator is developed to approximate and explicitly update the distribution of latent variables, guiding the distribution of trajectories;" + }, + { + "type": "text", + "bbox": [ + 0.539, + 0.514, + 0.905, + 0.575 + ], + "angle": 0, + "content": "(3) A global trajectory optimizer is designed, and KANs are introduced for the first time to achieve parallel trajectory optimization across the global spatiotemporal domain;" + }, + { + "type": "text", + "bbox": [ + 0.539, + 0.575, + 0.904, + 0.62 + ], + "angle": 0, + "content": "(4) A new direction-distance fused loss function is proposed for more effective supervision of agents' dynamic changes." + }, + { + "type": "text", + "bbox": [ + 0.521, + 0.62, + 0.905, + 0.665 + ], + "angle": 0, + "content": "- Comprehensive comparative experiments are conducted with existing high-quality models, demonstrating the effective and robust predictive capabilities of SocialMOIF." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.68, + 0.659, + 0.695 + ], + "angle": 0, + "content": "2. Related works" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Knowledge-driven methods. Helbing et al. [12] first proposed the social force model, which describes pedestrian interactions through an energy potential field. Best et al. [3] introduced a Bayesian inference-based intention inference model, integrating observed movement trajectories with prior knowledge. EqMotion [46] employed geometric analysis to convert human-human interaction trajectory analysis into an optimization problem by examining object geometry. Morris et al. [31] proposed an implicit Markov model for trajectory prediction, which is utilized for spatiotemporal probabilistic modeling of various types of pedestrian trajectories. A recent approach, socialcircle [41], incorporated angle-based light rule constraints into the trainable back" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22466" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.468, + 0.107 + ], + "angle": 0, + "content": "bone to achieve interaction modeling of agent intentions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.107, + 0.484, + 0.227 + ], + "angle": 0, + "content": "Despite these advancements, these methods rely on knowledge from specific domains and therefore struggle to fully capture the complex agent intentions within social scenarios. Although attempts have been made to incorporate future trajectory information to overcome the limitations of unidirectional inference, the model's equations remain overly sensitive to common variations, leading to persistent challenges in trajectory prediction tasks [42, 52]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.228, + 0.484, + 0.454 + ], + "angle": 0, + "content": "Data-driven methods. LSTM-based methods model trajectory data temporally. Alahi et al. [1] employed a \"social pooling mechanism\" that considers only the intentions of agents within a grid while neglecting interactions with boundary agents. Gupta et al. [11] improved upon this by implementing a pooling mechanism that treats pedestrians in neighboring areas equally. Transformer-based methods enhanced temporal modeling approaches. Yu et al. [49] modeled and predicted trajectories using a spatio-temporal concatenation approach. Yin et al. [48] introduced optical flow to simulate agents' motion intentions. Graph neural networks were also utilized for trajectory prediction to model intention interactions among different nodes [8, 22, 30, 35, 37, 44]. Diffusion models [10, 27] were used to explore trajectory prediction based on randomness." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.455, + 0.484, + 0.59 + ], + "angle": 0, + "content": "However, most of these methods primarily emphasize direct interactions between the target agent and its neighbors, often overlooking the higher-order influences that neighbors exert on the target agent. Furthermore, many of these approaches incorporate future trajectory information only at the network's output stage, supervising agents' trajectory changes solely through distance loss. Moreover, the limited interpretability of latent variable distributions in generative models can impede further development of these models." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.59, + 0.484, + 0.727 + ], + "angle": 0, + "content": "In addition, the aforementioned methods mainly emphasize trajectory analysis and modeling of upstream tasks. For output in downstream tasks, results are generally produced independently in a sequential manner over time. This sequential processing introduces delays that may result in untimely responses, potentially leading to safety concerns. Furthermore, each future prediction is dependent on the state from the preceding time step, which may lead to accumulated errors, adversely affecting prediction accuracy." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.74, + 0.2, + 0.758 + ], + "angle": 0, + "content": "3. Approach" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.484, + 0.902 + ], + "angle": 0, + "content": "The number of agents in the scenario is set to \\((N_{n} + 1)\\). The trajectory of the target agent \\(i\\) in history period \\(T_{H}\\) is \\(g_{i}^{1:T_{H}}\\), and \\(g_{i}^{t}\\) is the two-dimensional spatial coordinate at moment \\(t\\). The remaining agents are tentatively designated as neighbors, and the number is \\(N_{n}\\). The trajectory of \\(i\\) in the future period to be predicted is denoted as \\(\\hat{g}_{i}^{T_{H} + 1:T_{F}}\\). We propose the multi-order intention fusion model in Section 3.1, the trajectory distribution approximator in Section 3.2, the global trajectory optimizer in Section 3.3, and the" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.868, + 0.107 + ], + "angle": 0, + "content": "direction-distance fusion loss function in Section 3.4." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.116, + 0.819, + 0.131 + ], + "angle": 0, + "content": "3.1. Multi-order intention fusion model" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.138, + 0.907, + 0.319 + ], + "angle": 0, + "content": "The model consists of two layers: a first-order intention interaction layer and a higher-order intention interaction layer. As shown in Fig. 2, the first-order layer emphasizes the direct, dominant influence of neighbors on the target agent's intention. The higher-order layer accounts for the indirect and secondary influences on the target agent's intention that arise from interactions within the group of neighbors, thereby enabling a more comprehensive understanding of intentions. Based on the current and final history motion states of the agents, input components for the layers are first constructed, followed by the extraction and fusion of multi-order intention information." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.32, + 0.909, + 0.44 + ], + "angle": 0, + "content": "Higher-order intention interaction layer. From a more fundamental perspective, the higher-order intention interactions within the group of neighbors can be decomposed into first-order intention interactions between individual neighbors. We define the intention influence of neighbor agent \\(a\\) on neighbor agent \\(b\\) as \\(w_{a\\rightarrow b}\\); for an agent itself, it is represented as \\(w_{a\\rightarrow a}\\). In a scenario with \\(N_{n}\\) agents, the total number of such intentions is \\(\\Omega\\):" + }, + { + "type": "equation", + "bbox": [ + 0.602, + 0.45, + 0.907, + 0.468 + ], + "angle": 0, + "content": "\\[\n\\Omega = \\left(N _ {n} - 1\\right) N _ {n} + N _ {n} = N _ {n} ^ {2} \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.479, + 0.907, + 0.541 + ], + "angle": 0, + "content": "To obtain the intentions, we extract the absolute positions and velocities of the \\(N_{n}\\) neighbors corresponding to the target agent \\(i\\). They are processed as a layer input component through the embedding layer \\(\\pmb{f}_{S}\\) to obtain \\(U_{i}\\)." + }, + { + "type": "equation", + "bbox": [ + 0.571, + 0.549, + 0.907, + 0.575 + ], + "angle": 0, + "content": "\\[\nU _ {i} = \\boldsymbol {f} _ {U} \\left(x _ {1: N _ {n}} ^ {1: T _ {H}}, y _ {1: N _ {n}} ^ {1: T _ {H}}, v _ {x 1: N _ {n}} ^ {1: T _ {H}}, v _ {y 1: N _ {n}} ^ {1: T _ {H}}\\right) \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.584, + 0.906, + 0.675 + ], + "angle": 0, + "content": "We capture the intention interactions of neighbors through multiple subspaces in parallel. \\( U_{i} \\) is mapped into a learnable query vector \\( Q_{U} \\), a key vector \\( K_{U} \\), and a value vector \\( V_{U} \\). The feature dimensions are flattened into M subspaces. For the m-th subspace, the overall intention coefficient matrix within neighbors can be expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.596, + 0.684, + 0.907, + 0.726 + ], + "angle": 0, + "content": "\\[\nW _ {U} ^ {m} = \\operatorname {S o f t M a x} \\left(\\frac {Q _ {U} ^ {m} \\left(K _ {U} ^ {m}\\right) ^ {T}}{d _ {M}}\\right) \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.735, + 0.906, + 0.764 + ], + "angle": 0, + "content": "Thus, we obtain \\( w_{a\\to b} = [W_U^m ]_{ab} \\). For this matrix, the number of elements is \\( \\Omega \\)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.766, + 0.906, + 0.812 + ], + "angle": 0, + "content": "The first-order intention interaction layer. The layer input component of the target agent \\(i\\) is processed through the embedding layer \\(\\pmb{f}_{S}\\) to obtain \\(S_{i}\\)." + }, + { + "type": "equation", + "bbox": [ + 0.522, + 0.821, + 0.907, + 0.848 + ], + "angle": 0, + "content": "\\[\nS _ {i} = \\pmb {f} _ {S} \\left(x _ {i} ^ {1: T _ {H}}, y _ {i} ^ {1: T _ {H}}, d _ {i \\leftrightarrow 1: N _ {n}} ^ {1: T _ {H}}, \\theta_ {i \\leftrightarrow 1: N _ {n}} ^ {1: T _ {H}}, e _ {i \\leftrightarrow 1: N _ {n}} ^ {1: T _ {H}}\\right) (4)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.909, + 0.904 + ], + "angle": 0, + "content": "We intensify the dominant role of direct interaction in components and focus on the design of interaction rules. For the present case, \\(d_{i\\leftrightarrow j}^{1:T_H}\\) is the distance between the target" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22467" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.089, + 0.904, + 0.429 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.439, + 0.908, + 0.511 + ], + "angle": 0, + "content": "Figure 2. Computational pipeline of SocialMOIF. Firstly, the first-order interaction layer focuses on direct interactions between the target agent and its neighbors. The higher-order interaction layer addresses the interactions among neighbors. Multi-order intention is obtained by fusing the information from both layers. This multi-order intention is then combined with the actual future trajectory, with the distribution of the trajectory guided by an approximator. An optimizer is subsequently employed to enhance the decoded trajectory. Additionally, a fused distance-direction loss function supervises the training process. Finally, efficient parallel prediction is achieved." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.534, + 0.484, + 0.621 + ], + "angle": 0, + "content": "agent \\(i\\) and neighbor \\(j\\), and \\(\\theta_{i\\leftrightarrow j}^{1:T_H}\\) is the velocity angle. For the final history, \\(e_{i\\leftrightarrow j}^{1:T_H}\\) is the final distance in the history period, assuming that after moment \\(t\\), both the target agent \\(i\\) and neighbor \\(j\\) maintain the original velocity in the tangent direction \\(\\overrightarrow{vx_i^t},\\overrightarrow{vx_j^t}\\); then, \\(e_{ij}^t\\) can be expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.198, + 0.63, + 0.483, + 0.665 + ], + "angle": 0, + "content": "\\[\n\\lambda = \\left| \\left(\\overrightarrow {d _ {i j} ^ {t}} \\cdot \\overrightarrow {v _ {i j} ^ {t}}\\right) / \\left\\| \\overrightarrow {v _ {i j} ^ {t}} \\right\\| ^ {2} \\right| \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.132, + 0.673, + 0.483, + 0.703 + ], + "angle": 0, + "content": "\\[\ne _ {i j} ^ {t} = \\left\\| \\overrightarrow {d _ {i j} ^ {t}} + \\min (\\lambda , T _ {H} - t) \\cdot \\left(\\overrightarrow {v x _ {j} ^ {t}} - \\overrightarrow {v y _ {i} ^ {t}}\\right) \\right\\| \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.706, + 0.484, + 0.766 + ], + "angle": 0, + "content": "Similarly, \\( S_{i} \\) is mapped into learnable query vectors \\( Q_{S} \\), key vectors \\( K_{S} \\), and value vectors \\( V_{S} \\). The matrix of the intention coefficients between the target agent \\( i \\) and its neighbors is as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.179, + 0.775, + 0.483, + 0.817 + ], + "angle": 0, + "content": "\\[\nW _ {S} = \\operatorname {S o f t M a x} \\left(\\frac {Q _ {S} \\left(K _ {S}\\right) ^ {T}}{d _ {S}}\\right) \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Intention Fusion. Higher-order interaction intentions are conducted to the target agent and combined with first-order interaction intentions to achieve multi-order intention fusion. During this fusion, the higher-order interaction intentions captured within each subspace exert varying effects" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.535, + 0.907, + 0.581 + ], + "angle": 0, + "content": "when applied to the target agent. We set the corresponding trainable influence factor \\(\\eta_{m}\\) to denote this influence. The attention score \\(A_{i}\\) is denoted as:" + }, + { + "type": "equation", + "bbox": [ + 0.601, + 0.59, + 0.907, + 0.632 + ], + "angle": 0, + "content": "\\[\nA _ {i} = \\left(\\sum_ {m = 1} ^ {M} \\eta_ {m} W _ {U} ^ {m} + W _ {S}\\right) V _ {S} \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.641, + 0.907, + 0.686 + ], + "angle": 0, + "content": "where \\( A_{i} \\) is processed by two skip connections and a final fully connected layer to obtain the multi-order fused intention \\( I_{i} \\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.695, + 0.838, + 0.712 + ], + "angle": 0, + "content": "3.2. Trajectory distribution approximator" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.715, + 0.907, + 0.824 + ], + "angle": 0, + "content": "We define the trajectory \\( g_{igt}^{T_H + 1:T_F} \\) of the target agent \\( i \\) in the future period. \\( B_{i} \\) is obtained after feature enhancement through the embedding layer \\( f_{B} \\). By utilizing \\( I_{i} \\) as the lower bound and \\( B_{i} \\) as the upper bound, the latent variable distribution in the potential space is explicitly approximated. First, the probability distribution of the target agent \\( i \\)'s trajectory is initially modeled as:" + }, + { + "type": "equation", + "bbox": [ + 0.543, + 0.833, + 0.905, + 0.905 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} p \\left(g _ {i} ^ {T _ {H} + 1: T _ {F}} \\mid U _ {i}, S _ {i}\\right) \\\\ = \\prod_ {t = T _ {H} + 1} ^ {T _ {F}} p \\left(g _ {i} ^ {T _ {H} + 1 + t} \\mid g _ {i} ^ {T _ {H} + 1: T _ {H} + t - 1}, U _ {i}, S _ {i}\\right) \\tag {9} \\\\ \\end{array}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22468" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.122 + ], + "angle": 0, + "content": "We sample the noise variables \\(\\varepsilon \\sim N(0,1)\\) and reparameterize the latent variables that are approximated at \\(t\\) [17]:" + }, + { + "type": "equation", + "bbox": [ + 0.09, + 0.13, + 0.483, + 0.151 + ], + "angle": 0, + "content": "\\[\nz _ {i \\varphi} ^ {t} = u _ {i \\varphi} ^ {t} + \\varepsilon \\cdot \\sigma_ {i \\varphi} ^ {t} \\quad q _ {\\varphi} (\\cdot | I _ {i} ^ {t - 1}, B _ {i} ^ {t}) \\sim N (u _ {i \\varphi} ^ {t}, \\sigma_ {i \\varphi} ^ {t}) \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.157, + 0.484, + 0.217 + ], + "angle": 0, + "content": "where \\(\\varphi\\) are parameters that need to be optimized during the training process. In the test mode, the latent variables \\(z_{i\\vartheta}^{t}\\) are only reasoned by \\(I_{i}\\), and \\(\\vartheta\\) is also involved in the training process of the model." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.218, + 0.484, + 0.262 + ], + "angle": 0, + "content": "We sample and decode the target agent's trajectories on the basis of latent variables and multi-order fused intention information." + }, + { + "type": "equation", + "bbox": [ + 0.168, + 0.269, + 0.483, + 0.309 + ], + "angle": 0, + "content": "\\[\ng _ {i} ^ {t} \\sim \\left\\{ \\begin{array}{l l} p _ {\\delta} (\\cdot | z _ {i \\varphi} ^ {t}, I _ {i} ^ {t - 1}) & \\text {i f t r a i n} \\\\ p _ {\\delta} (\\cdot | z _ {i \\vartheta} ^ {t}, I _ {i} ^ {t - 1}) & \\text {i f t e s t} \\end{array} \\right. \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.316, + 0.483, + 0.347 + ], + "angle": 0, + "content": "where \\(\\delta\\) are parameters that need to be optimized during the training process." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.347, + 0.484, + 0.38 + ], + "angle": 0, + "content": "Concat the latent variables with the trajectories generated at the current time \\( t \\), updating \\( I_{i}^{t - 1} \\) via \\( f_{zg} \\)." + }, + { + "type": "equation", + "bbox": [ + 0.116, + 0.388, + 0.483, + 0.428 + ], + "angle": 0, + "content": "\\[\nI _ {i} ^ {t} = \\left\\{ \\begin{array}{l l} \\boldsymbol {f} _ {z g} \\left(\\operatorname {c o n c a t} \\left(z _ {i \\varphi} ^ {t}, g _ {i} ^ {t}\\right), I _ {i} ^ {t - 1}\\right) & \\text {i f t r a i n} \\\\ \\boldsymbol {f} _ {z g} \\left(\\operatorname {c o n c a t} \\left(z _ {i \\vartheta} ^ {t}, g _ {i} ^ {t}\\right), I _ {i} ^ {t - 1}\\right) & \\text {i f t e s t} \\end{array} \\right. \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.435, + 0.484, + 0.495 + ], + "angle": 0, + "content": "In the next time step, the updated multi-order fused intention contains the information of the latent variables, which is then combined with \\( B_{i}^{t} \\) to approximate the latent variables and update them." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.496, + 0.484, + 0.541 + ], + "angle": 0, + "content": "By guiding the trajectory distribution through such a training process, the generated trajectories are explicitly represented by the latent variables as:" + }, + { + "type": "equation", + "bbox": [ + 0.109, + 0.549, + 0.483, + 0.649 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} p \\left(g _ {i} ^ {T _ {H} + 1: T _ {F}} \\mid U _ {i}, S _ {i}\\right) \\\\ = \\prod_ {t = T _ {H} + 1} ^ {T _ {F}} \\int_ {z _ {i \\varphi} ^ {t}} p \\left(g _ {i} ^ {t} \\mid g _ {i} ^ {T _ {H} + 1: t - 1}, I _ {i} ^ {t - 1}, B _ {i} ^ {t}, z _ {i \\varphi} ^ {t}\\right) \\cdot \\tag {13} \\\\ p \\left(z _ {i \\varphi} ^ {t} \\mid g _ {i} ^ {T _ {H} + 1: t - 1}, I _ {i} ^ {t - 1}, B _ {i} ^ {t}\\right) d z _ {i \\varphi} ^ {t} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.657, + 0.47, + 0.673 + ], + "angle": 0, + "content": "In test mode, \\( B_{i}^{t} \\) is removed, and \\( \\varphi \\) is replaced with \\( \\vartheta \\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.681, + 0.34, + 0.697 + ], + "angle": 0, + "content": "3.3. Global trajectory optimizer" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.703, + 0.484, + 0.824 + ], + "angle": 0, + "content": "For efficient prediction, we introduce the KANs for the first time. We aggregate the \\(g_{i}^{T_{H} + 1:T_{F}}\\) time dimension to obtain \\(\\Gamma^0\\in \\mathbb{R}^{2T_F}\\) to facilitate global parallel optimization. The depth of the optimizer is \\(L\\) layers, the number of nodes in the \\(\\ell\\) -th layer is \\(m\\) , the number of nodes in the \\(\\ell +1\\) -th layer is \\(n\\) , and the activation function matrix corresponding to the nodes in the \\(l\\) -th layer is \\(\\Phi_{\\ell}\\) . The optimization process for the \\(\\ell \\rightarrow \\ell +1\\) layer is as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.105, + 0.831, + 0.484, + 0.906 + ], + "angle": 0, + "content": "\\[\n\\Gamma^ {\\ell + 1} = \\Phi_ {\\ell} \\Gamma^ {\\ell} = \\left[ \\begin{array}{c c c c} \\varphi_ {1, 1} ^ {\\ell} & \\varphi_ {1, 2} ^ {\\ell} & \\dots & \\varphi_ {1, m} ^ {\\ell} \\\\ \\varphi_ {2, 1} ^ {\\ell} & \\varphi_ {2, 2} ^ {\\ell} & \\dots & \\varphi_ {2, m} ^ {\\ell} \\\\ \\vdots & \\vdots & \\ddots & \\vdots \\\\ \\varphi_ {n, 1} ^ {\\ell} & \\varphi_ {n, 2} ^ {\\ell} & \\dots & \\varphi_ {n, m} ^ {\\ell} \\end{array} \\right] \\Gamma^ {\\ell} (1 4)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.906, + 0.137 + ], + "angle": 0, + "content": "Eventually, the predicted trajectory is reduced back to the original dimensional space to obtain the final trajectory over the entire prediction period at once." + }, + { + "type": "equation", + "bbox": [ + 0.521, + 0.149, + 0.906, + 0.169 + ], + "angle": 0, + "content": "\\[\n\\hat {g} _ {i} ^ {T _ {H} + 1: T _ {F}} \\leftarrow \\Gamma^ {L} = \\left(\\Phi_ {L} \\circ \\Phi_ {L - 1} \\circ \\dots \\circ \\Phi_ {1} \\circ \\Phi_ {0}\\right) \\Gamma^ {0} (1 5)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.182, + 0.906, + 0.213 + ], + "angle": 0, + "content": "Given the diversity of agent intentions, the process predicts multiple possible trajectories." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.225, + 0.654, + 0.239 + ], + "angle": 0, + "content": "3.4. Loss function" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.247, + 0.906, + 0.308 + ], + "angle": 0, + "content": "Agents traveling in opposite or the same direction exert distinct effects on their respective movement states. Therefore, direction loss is incorporated to supervise the agent's dynamics during backpropagation." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.309, + 0.906, + 0.339 + ], + "angle": 0, + "content": "For distance. The distance prediction error in future periods can be expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.643, + 0.352, + 0.906, + 0.372 + ], + "angle": 0, + "content": "\\[\nL _ {d i s} = \\left\\| \\hat {g} _ {i} ^ {t} - g _ {i g t} ^ {t} \\right\\| \\tag {16}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.384, + 0.906, + 0.459 + ], + "angle": 0, + "content": "For direction, the final moment of the historical period is introduced. Each predicted time step is related to the preceding time step to construct the direction vector, and subsequently, the angle between the predicted trajectory and the actual trajectory direction vector is analyzed." + }, + { + "type": "equation", + "bbox": [ + 0.524, + 0.472, + 0.906, + 0.514 + ], + "angle": 0, + "content": "\\[\nL _ {a n g l e} = - \\operatorname {a r c c o s} \\left(\\frac {\\left(\\hat {g} _ {i} ^ {t} - \\hat {g} _ {i} ^ {t + 1}\\right) \\cdot \\left(g _ {i g t} ^ {t} - g _ {i g t} ^ {t + 1}\\right)}{\\left\\| \\hat {g} _ {i} ^ {t} - \\hat {g} _ {i} ^ {t + 1} \\right\\| \\left\\| g _ {i g t} ^ {t} - g _ {i g t} ^ {t + 1} \\right\\|}\\right) \\tag {17}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.526, + 0.906, + 0.572 + ], + "angle": 0, + "content": "The objective of our model is to maximize the lower bound on the evidence for all time steps of the target distribution, joints (16) and (17), with a final loss function of:" + }, + { + "type": "equation", + "bbox": [ + 0.531, + 0.584, + 0.906, + 0.651 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} L = \\sum_ {t = T _ {H} + 1} ^ {T _ {F}} \\left\\{\\mathbb {E} _ {z _ {i \\varphi} ^ {t} \\sim q _ {\\varphi} \\left(\\cdot | B _ {i} ^ {t}, I _ {i} ^ {t - 1}\\right)} \\left[ L _ {d i s} + L _ {a n g l e} \\right] \\right. \\tag {18} \\\\ \\left. - K L \\left[ q _ {\\varphi} \\left(z _ {i \\varphi} ^ {t} \\mid B _ {i} ^ {t}, I _ {i} ^ {t - 1}\\right) \\| q _ {\\vartheta} \\left(z _ {i \\vartheta} ^ {t} \\mid I _ {i} ^ {t - 1}\\right) \\right] \\right\\} \\\\ \\end{array}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.664, + 0.646, + 0.681 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Datasets. ETH [32] and UCY [18] datasets encompass over 1,600 pedestrian trajectories, with positions annotated at 0.4s intervals. SDD [33] dataset is primarily utilized for pedestrian trajectory prediction and object tracking, containing about 19,000 pedestrian trajectories, with positions annotated every 0.4s. NBA dataset focuses on the 2015-2016 season [21, 53]. Two subsets, \"Rebounds\" and \"Scores,\" are extracted for benchmarking purposes. The movement intentions of the agents exhibit strong purposefulness and complex, hard-to-capture velocity variations. Consequently, the extracted trajectories encompass rich, highly nonlinear motion patterns. NuScenes [4] is a dataset for autonomous driving, and the prediction challenge segmentation and configurations are adhered to." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "22469" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.908, + 0.118 + ], + "angle": 0, + "content": "Table 1. Comparison of Related Methods on ETH/UCY, NBA, SDD, and NuScenes (in meters). Every work was realized using best-of-20 predictions for the effect of ade/fde was selected." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.129, + 0.907, + 0.299 + ], + "angle": 0, + "content": "
ETH/UCYFLEAM [6]ST-MR [19]MTN [48]MANTRA [28]Agent-former [51]BiTrap [47]MemoNet [45]EqMotion [46]V2-Net-SC [39, 41]E-V2-Net-SC [40, 41]Ours
Eth0.66/1.210.53/0.690.58/0.860.56/0.970.58/0.860.66/1.090.51/0.690.36/0.540.23/0.330.23/0.300.26/0.31
Hotel0.37/0.640.20/0.310.19/0.570.17/0.320.15/0.240.17/0.320.14/0.200.12/0.160.11/0.160.10/0.130.10/0.12
Univ0.57/0.970.41/0.640.32/0.430.38/0.790.26/0.480.26/0.460.26/0.520.21/0.390.17/0.270.18/0.240.11/0.17
Zara010.46/0.740.26/0.380.24/0.370.24/0.410.19/0.320.23/0.470.21/0.360.16/0.270.13/0.180.13/0.160.10/0.17
Zara020.37/0.690.18/0.240.15/0.210.18/0.340.16/0.250.18/0.350.15/0.300.11/0.190.12/0.210.11/0.160.09/0.14
Avg0.48/0.850.32/0.450.30/0.490.31/0.570.27/0.430.30/0.540.25/0.410.19/0.310.15/0.230.15/0.200.13/0.18
NBAFLEAM [6]ST-MR [19]MTN [48]SGNet [38]EqMotion [46]MemoNet [45]BiTrap [47]GroupNet +CVAE [44]V2-Net-SC [39, 41]E-V2-Net-SC [40, 41]Ours
Rebound1.92/2.210.94/1.730.84/0.970.80/1.780.86/1.211.73/2.060.86/1.880.63/0.840.54/0.790.54/0.860.34/0.66
Scores1.29/1.680.90/1.030.87/1.410.71/1.450.69/1.021.35/1.870.76/1.550.46/0.760.50/0.840.47/0.780.30/0.56
SDDFLEAM [6]SHENet [29]MemoNet [45]MANTRA [28]BiTrap [47]MID [10]Y-Net [26]NSP-SFM [52]V2-Net-SC [39, 41]E-V2-Net-SC [40, 41]Ours
SDD0.67/1.280.44/0.530.45/0.640.49/0.620.46/0.590.38/0.730.32/0.660.21/0.340.26/0.360.23/0.350.17/0.24
NuScenesFLEAM [6]ST-MR [19]MTN [48]MANTRA [28]SGNet [38]Agent-former [51]EqMotion [46]Y-Net [26]V2-Net-SC [39, 41]E-V2-Net-SC [40, 41]Ours
NuScenes2.02/3.141.84/2.461.64/2.121.47/2.011.28/2.091.04/1.471.24/1.981.12/1.711.06/1.591.04/1.640.92/1.56
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.323, + 0.482, + 0.399 + ], + "angle": 0, + "content": "Metrics. Best-of-20 predictions were used to compute Average Displacement Error (ADE) and Final Displacement Error (FDE) as primary metrics, with estimated Negative Log Likelihood (NLL) reported from the test set. See their detailed definitions in the Supplementary Material." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.4, + 0.483, + 0.551 + ], + "angle": 0, + "content": "Implementation Details. The model was trained on two NVIDIA RTX 4090 GPUs. Following the methodology in [5, 36], a leave-one-out cross-validation approach was employed for dataset handling. The number of subspaces in the higher-order intention interaction layer was set to \\( M = 6 \\) while the depth of the global trajectory optimizer was set to \\( L = 3 \\). During inference, the best parameter model was loaded to predict the next 8 frames based on the previous 8 frames of history. For more implementation details of the training in the Supplementary Material." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.562, + 0.291, + 0.578 + ], + "angle": 0, + "content": "4.1. Quantitative analysis" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.584, + 0.485, + 0.901 + ], + "angle": 0, + "content": "All SOTA performances across datasets were highlighted in bold black in Table 1. In Eth scenario, crowds generally moved in parallel, resulting in weaker interactions among neighbors that conducted to target agent. E-V2-Net-SC [40, 41] achieved optimal performance in this scenario. However, the proposed method still achieved superior performance on ETH/UCY dataset, reducing ADE and FDE by \\(13.33\\%\\) and \\(10.00\\%\\) compared to E-V2-Net-SC, respectively. For NBA dataset, where interactions between agents were complex and frequent, the proposed method demonstrated exceptional performance. In Rebound sub-dataset, ADE and FDE decreased by \\(37.04\\%\\) and \\(16.46\\%\\), respectively, compared to existing SOTA methods. In Scores sub-dataset, ADE and FDE were reduced by \\(34.78\\%\\) and \\(26.32\\%\\), respectively, relative to the current SOTA methods. For SDD dataset, image segmentation techniques, as referenced in [2], were employed to capture pedestrian position information across eight scenario. Raw pixel coordinates were converted into spatial coordinates defined in meters for trajectory prediction tasks. NSP-SFM [52] outperformed both V2-Net-SC and E-V2-Net-SC. Our method re" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.323, + 0.907, + 0.399 + ], + "angle": 0, + "content": "duced ADE and FDE by \\(19.05\\%\\) and \\(29.41\\%\\), respectively, compared to NSP-SFM. For NuScenes dataset, V2-Net-SC and E-V2-Net-SC showed similar performance levels. Our method reduced ADE and FDE by \\(11.54\\%\\) and \\(1.87\\%\\), respectively, in comparison to their SOTA performances." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.4, + 0.907, + 0.475 + ], + "angle": 0, + "content": "Detailed comparative experiments with NLL metric were carried out in the Supplementary Material. Furthermore, for wider validation and analysis, we applied directional component in the designed loss function to different models. See the Supplementary Material for details." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.488, + 0.702, + 0.505 + ], + "angle": 0, + "content": "4.2. Qualitative analysis" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.526, + 0.903, + 0.796 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.81, + 0.907, + 0.853 + ], + "angle": 0, + "content": "Figure 3. Representative cases. Row 1 represented complex historical trajectory cases. Row 2 represented state maintenance cases. Row 3 represented numerous neighbors cases." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Unlike previous studies [6, 26, 39-41, 45, 46], the focus here was on conducting experiments involving historical" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22470" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.182, + 0.092, + 0.816, + 0.379 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.389, + 0.908, + 0.418 + ], + "angle": 0, + "content": "Figure 4. Predictions comparison of predictions in ETH/UCY. The comparison included heatmaps of the overall distribution of predictions and best-of-20 predictions of the related methods." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.444, + 0.485, + 0.807 + ], + "angle": 0, + "content": "trajectories in various social contexts. This enabled a more effective validation of the model's ability to extract and analyze multi-order fused intentions in Fig. 3. In panel (a), the target agent continuously changed direction. Although the model displayed slight deviations at certain moments, these deviations remained within the acceptable range of directional changes. In panel (b) and (c), the target agent intended to turn, but the model successfully achieved accurate predictions. The target agent moved at a relatively steady speed in panel (d), adopting a conservative deceleration strategy in response to numerous neighbors who are stopping or remaining stationary. In panel (e) and (f), there were many stationary neighbors who may be gathered together in conversation. Consequently, the target agent maintained its original state. In panel (g) and (h), most neighbors moved in the same direction as the target agent, with closer neighbors receiving higher attention scores. This suggested that the target agent prioritized attention to closer neighbors, consistent with social norms. In panel (i), the directions of neighbors were complex, and their speeds varied. The target agent accounted for influences from neighbors in multiple directions while attempting to anticipate and avoid imminent collisions. Please see the qualitative analyses of NBA interactions in the Supplementary Material." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Comparisons were made between the proposed method and three previous methods in Fig. 4. It was observed that our method produced more accurate predictions than the previous methods. In Eth, when facing neighbors to the left front, the model provided a prediction closest to the true trajectory. The predictions from E-V2-Net-SC and EqMotion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.444, + 0.907, + 0.596 + ], + "angle": 0, + "content": "exhibited slight oscillations, while MemoNet, by mimicking retrospective memory mechanisms, tends to maintain its original trajectory. By employing the squeeze theorem to approximate future true trajectories, the proposed model shows greater convergence in the overall distribution of predictions in all scenarios. In Hotel and Zara01, there were many pedestrians ahead of the target agent, and the diverging trajectory distributions from previous methods represent rather aggressive strategies, which conflicted with the safety requirements of autonomous driving." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.607, + 0.665, + 0.623 + ], + "angle": 0, + "content": "4.3. Ablation study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.629, + 0.907, + 0.87 + ], + "angle": 0, + "content": "In Table 2, ablation experiments were conducted by sequentially adding components to the PDI baseline. Groups 1-4 focus on the multi-order intention fusion module. Group 1 includes only the neighbors' position and the distances between the target agent and the neighbors. As additional components, such as neighboring speed \\(\\mathbf{V}\\), the angle between the target agent and the neighboring speed \\(\\Theta\\), and the final predicted distance over the historical period \\(\\mathbf{E}\\), were gradually incorporated, a substantial reduction in the final prediction error was observed. Specifically, ADE decreased by \\(23.91\\%\\) and FDE by \\(39.47\\%\\) on the SDD dataset, while ADE reduced by \\(11.11\\%\\) and FDE by \\(14.42\\%\\) on the Nuscenes dataset. Groups 4-5, 5-6, and 6-7 focus on the ablation of the approximator, the optimizer, and the angle loss term in the loss function, respectively. The results demonstrate that each component contributed positively." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.907, + 0.901 + ], + "angle": 0, + "content": "To examine the influence of \\(\\mathbf{B}\\) and \\(\\mathbf{I}\\), the trajectory distribution for several scenarios in the future periods was visu-" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "22471" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.907, + 0.16 + ], + "angle": 0, + "content": "Table 2. Ablation studies of each component in our method on SDD and Nuscenes datasets. Notably, the positional information of the target agent was essential and was included in each set of ablation experiments. \\(\\mathbf{P}\\) represented the position of neighbors, \\(\\mathbf{V}\\) indicated their velocities, \\(\\mathbf{D}\\) denoted the distance between the target agent and its neighbors, and \\(\\Theta\\) represented the angle between their velocities. \\(\\mathbf{E}\\) represented the final predicted distance between the target agent and its neighbors over historical periods, \\(\\mathbf{I}\\) indicated global historical social intent, \\(\\mathbf{B}\\) denoted future actual trajectories, \\(\\mathbf{K}\\) referred to the parallel prediction mode, and \\(\\mathbf{A}\\) signified directional information." + }, + { + "type": "table", + "bbox": [ + 0.153, + 0.17, + 0.849, + 0.276 + ], + "angle": 0, + "content": "
GroupPVDΘEIBKASDDNuscenes
1✓✓✓0.46/0.711.26/2.15
2✓✓✓✓0.45/0.661.24/2.09
3✓✓✓✓✓0.41/0.561.18/1.97
4✓✓✓✓✓✓0.35/0.431.12/1.84
5✓✓✓✓✓✓✓0.30/0.361.04/1.76
6✓✓✓✓✓✓✓✓0.20/0.270.93/1.61
7✓✓✓✓✓✓✓✓✓0.17/0.240.92/1.56
" + }, + { + "type": "image", + "bbox": [ + 0.182, + 0.291, + 0.822, + 0.503 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.513, + 0.907, + 0.544 + ], + "angle": 0, + "content": "Figure 5. Distribution heatmaps. Every heatmap of future trajectory distribution was shown above the real scene image. (a) represented the actual distribution. (b) presented the experimental group containing only I component. (c) showed the group with both I and B components." + }, + { + "type": "image", + "bbox": [ + 0.111, + 0.569, + 0.462, + 0.724 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.736, + 0.483, + 0.779 + ], + "angle": 0, + "content": "Figure 6. Representative cases about the optimization effects of component \\(\\mathbf{K}\\). To ensure fairness, we selected best-of-20 predictions in groups 6 and 7 for comparison." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.484, + 0.902 + ], + "angle": 0, + "content": "alized in Fig. 5. It was evident that when both components \\(\\mathbf{B}\\) and \\(\\mathbf{I}\\) were included, the predicted trajectory distribution aligned more closely with the actual distribution. It can be found that in a strongly regular sports scenario such as Rebound, (b) and (c) achieved excellent predictions, while in open environments with complex social contexts, the effec" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.569, + 0.871, + 0.584 + ], + "angle": 0, + "content": "tiveness of the approximator was better demonstrated." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.585, + 0.907, + 0.676 + ], + "angle": 0, + "content": "To validate the effectiveness of component \\(\\mathbf{K}\\) under global parallel optimization, six challenging cases were presented in Fig. 6. It was evident that predictions without component \\(\\mathbf{K}\\) exhibited increasing errors as the prediction time accumulated, whereas the parallel prediction mode demonstrated a closer result to ground-truth." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.693, + 0.634, + 0.709 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Each human movement path is either directly or indirectly associated with neighbor agents. This study introduces SocialMOIF, which provides a more comprehensive representation of both direct and indirect intention interactions between the target agent and neighbor agents. SocialMOIF comprises four key designs: a multi-order intention fusion model, a trajectory distribution approximator, a global trajectory optimizer, and a novel loss function. Experimental results demonstrate that this approach performs competitively across various datasets and effectively approximates and explicitly updates trajectory distributions during training, thereby improving interpretability." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22472" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.251, + 0.108 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.484, + 0.266 + ], + "angle": 0, + "content": "This project is supported by National Natural Science Foundation of China (52202417), China Postdoctoral Science Foundation(2022TQ0155, 2022M721605), Open Project Program of State Key Laboratory of Virtual Reality Technology and Systems, Beihang University(VRLAB2023A02), Young Elite Scientists Sponsorship Program by CAST(2023QNRC001), Fundamental Research Funds for the Central Universities (NS2024030), Young Elite Scientists Sponsorship Program by JSTJ(JSTJ-2023-XH032)." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.278, + 0.188, + 0.294 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.304, + 0.484, + 0.373 + ], + "angle": 0, + "content": "[1] Alexandre Alahi, Kratarth Goel, Vignesh Ramanathan, Alexandre Robicquet, Li Fei-Fei, and Silvio Savarese. Social lstm: Human trajectory prediction in crowded spaces. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 961–971, 2016. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.373, + 0.484, + 0.427 + ], + "angle": 0, + "content": "[2] Stefan Becker, Ronny Hug, Wolfgang Hubner, and Michael Arens. An evaluation of trajectory prediction approaches and notes on the trajnet benchmark. arXiv preprint arXiv:1805.07663, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.429, + 0.484, + 0.496 + ], + "angle": 0, + "content": "[3] Graeme Best and Robert Fitch. Bayesian intention inference for trajectory prediction with an unknown goal destination. In 2015 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 5817-5823. IEEE, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.498, + 0.484, + 0.581 + ], + "angle": 0, + "content": "[4] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11621-11631, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.582, + 0.484, + 0.65 + ], + "angle": 0, + "content": "[5] Kai Chen, Xiao Song, and Xiaoxiang Ren. Pedestrian trajectory prediction in heterogeneous traffic using pose keypoints-based convolutional encoder-decoder network. IEEE Transactions on Circuits and Systems for Video Technology, 31(5): 1764-1775, 2020. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.652, + 0.484, + 0.719 + ], + "angle": 0, + "content": "[6] Kai Chen, Xiao Song, Haitao Yuan, and Xiaoxiang Ren. Fully convolutional encoder-decoder with an attention mechanism for practical pedestrian trajectory prediction. IEEE Transactions on Intelligent Transportation Systems, 23(11): 20046-20060, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.721, + 0.484, + 0.776 + ], + "angle": 0, + "content": "[7] Kai Chen, Haihua Zhu, Dunbing Tang, and Kun Zheng. Future pedestrian location prediction in first-person videos for autonomous vehicles and social robots. Image and Vision Computing, 134:104671, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.777, + 0.484, + 0.845 + ], + "angle": 0, + "content": "[8] Lingwei Dang, Yongwei Nie, Chengjiang Long, Qing Zhang, and Guiqing Li. Msr-gen: Multi-scale residual graph convolution networks for human motion prediction. In Proceedings of the IEEE/CVF international conference on computer vision, pages 11467-11476, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.846, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[9] Francesco Giuliani, Irtiza Hasan, Marco Cristani, and Fabio Galasso. Transformer networks for trajectory forecasting. In 2020 25th international conference on pattern recognition (ICPR), pages 10335-10342. IEEE, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.304, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.162 + ], + "angle": 0, + "content": "[10] Tianpei Gu, Guangyi Chen, Junlong Li, Chunze Lin, Yongming Rao, Jie Zhou, and Jiwen Lu. Stochastic trajectory prediction via motion indeterminacy diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17113-17122, 2022. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.164, + 0.906, + 0.233 + ], + "angle": 0, + "content": "[11] Agrim Gupta, Justin Johnson, Li Fei-Fei, Silvio Savarese, and Alexandre Alahi. Social gan: Socially acceptable trajectories with generative adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2255–2264, 2018. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.234, + 0.906, + 0.275 + ], + "angle": 0, + "content": "[12] Dirk Helbing and Peter Molnar. Social force model for pedestrian dynamics. Physical review E, 51(5):4282, 1995. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.277, + 0.906, + 0.33 + ], + "angle": 0, + "content": "[13] Serge Hoogendoorn and Piet HL Bovy. Simulation of pedestrian flows by optimal control and differential games. Optimal control applications and methods, 24(3):153-172, 2003. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.334, + 0.906, + 0.389 + ], + "angle": 0, + "content": "[14] SP Hoogendoorn, Winnie Daamen, Yaqing Shu, and Han Ligteringen. Modeling human behavior in vessel maneuver simulation by optimal control and game theory. Transportation research record, 2326(1):45-53, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.391, + 0.906, + 0.461 + ], + "angle": 0, + "content": "[15] Yihan Hu, Jiazhi Yang, Li Chen, Keyu Li, Chonghao Sima, Xizhou Zhu, Siqi Chai, Senyao Du, Tianwei Lin, Wenhai Wang, et al. Planning-oriented autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17853-17862, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.462, + 0.906, + 0.532 + ], + "angle": 0, + "content": "[16] Sungjune Kim, Hyung-gun Chi, Hyerin Lim, Karthik Ramani, Jinkyu Kim, and Sangpil Kim. Higher-order relational reasoning for pedestrian trajectory prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15251-15260, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.533, + 0.906, + 0.587 + ], + "angle": 0, + "content": "[17] Christos Kyrkou, Panayiotis Kolios, Theocharis Theocharides, and Marios Polycarpou. Machine learning for emergency management: A survey and future outlook. Proceedings of the IEEE, 111(1):19-41, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.589, + 0.906, + 0.631 + ], + "angle": 0, + "content": "[18] Alon Lerner, Yiorgos Chrysanthou, and Dani Lischinski. Crowds by example. In Computer graphics forum, pages 655-664. Wiley Online Library, 2007. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.632, + 0.907, + 0.702 + ], + "angle": 0, + "content": "[19] Lihuan Li, Maurice Pagnucco, and Yang Song. Graph-based spatial transformer with memory replay for multifuture pedestrian trajectory prediction. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2221-2231, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.703, + 0.907, + 0.77 + ], + "angle": 0, + "content": "[20] Linhui Li, Xiaotong Lin, Yejia Huang, Zizhen Zhang, and Jian-Fang Hu. Beyond minimum-of-n: Rethinking the evaluation and methods of pedestrian trajectory prediction. IEEE Transactions on Circuits and Systems for Video Technology, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.773, + 0.906, + 0.801 + ], + "angle": 0, + "content": "[21] Kostya Linou. Nba-player-movements. Software for Visualization of NBA Games, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.803, + 0.906, + 0.872 + ], + "angle": 0, + "content": "[22] Congcong Liu, Yuying Chen, Ming Liu, and Bertram E Shi. Avgcn: Trajectory prediction using graph convolutional networks guided by human attention. In 2021 IEEE International Conference on Robotics and Automation (ICRA), pages 14234-14240. IEEE, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[23] Ziming Liu, Pingchuan Ma, Yixuan Wang, Wojciech Ma-tusik, and Max Tegmark. Kan 2.0: Kolmogorov-arnold net" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22473" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.119 + ], + "angle": 0, + "content": "works meet science. arXiv preprint arXiv:2408.10205, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.178 + ], + "angle": 0, + "content": "[24] Ziming Liu, Yixuan Wang, Sachin Vaidya, Fabian Ruehle, James Halverson, Marin Soljačić, Thomas Y Hou, and Max Tegmark. Kan: Kolmogorov-arnold networks. arXiv preprint arXiv:2404.19756, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.18, + 0.482, + 0.276 + ], + "angle": 0, + "content": "[25] Karttikeya Mangalam, Harshayu Girase, Shreyas Agarwal, Kuan-Hui Lee, Ehsan Adeli, Jitendra Malik, and Adrien Gaidon. It is not the journey but the destination: Endpoint conditioned trajectory prediction. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16, pages 759-776. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.278, + 0.482, + 0.346 + ], + "angle": 0, + "content": "[26] Karttikeya Mangalam, Yang An, Harshayu Girase, and Jitendra Malik. From goals, waypoints & paths to long term human trajectory forecasting. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15233–15242, 2021. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.349, + 0.482, + 0.417 + ], + "angle": 0, + "content": "[27] Weibo Mao, Chenxin Xu, Qi Zhu, Siheng Chen, and Yanfeng Wang. Leapfrog diffusion model for stochastic trajectory prediction. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5517-5526, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.42, + 0.483, + 0.489 + ], + "angle": 0, + "content": "[28] Francesco Marchetti, Federico Becattini, Lorenzo Seidenari, and Alberto Del Bimbo. *Mantra: Memory augmented networks for multiple trajectory prediction*. In *Proceedings of the IEEE/CVF conference on computer vision and pattern recognition*, pages 7143–7152, 2020. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.491, + 0.482, + 0.546 + ], + "angle": 0, + "content": "[29] Mancheng Meng, Ziyan Wu, Terrence Chen, Xiran Cai, Xiang Zhou, Fan Yang, and Dinggang Shen. Forecasting human trajectory from scene history. arXiv preprint arXiv:2210.08732, 2022. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.548, + 0.482, + 0.616 + ], + "angle": 0, + "content": "[30] Alessio Monti, Alessia Bertugli, Simone Calderara, and Rita Cucchiara. Dag-net: Double attentive graph neural network for trajectory forecasting. In 2020 25th international conference on pattern recognition (ICPR), pages 2551-2558. IEEE, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.619, + 0.482, + 0.687 + ], + "angle": 0, + "content": "[31] Brendan Tran Morris and Mohan Manubhai Trivedi. Trajectory learning for activity understanding: Unsupervised, multilevel, and long-term adaptive approach. IEEE transactions on pattern analysis and machine intelligence, 33(11):2287-2301, 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.69, + 0.482, + 0.758 + ], + "angle": 0, + "content": "[32] Stefano Pellegrini, Andreas Ess, Konrad Schindler, and Luc Van Gool. You'll never walk alone: Modeling social behavior for multi-target tracking. In 2009 IEEE 12th international conference on computer vision, pages 261-268. IEEE, 2009. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.761, + 0.482, + 0.816 + ], + "angle": 0, + "content": "[33] Alexandre Robicquet, Amir Sadeghian, Alexandre Alahi, and Silvio Savarese. Learning social etiquette: Human trajectory prediction in crowded scenes. In European Conference on Computer Vision (ECCV), page 5, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.818, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[34] Tim Salzmann, Boris Ivanovic, Punarjay Chakravarty, and Marco Pavone. Trajectory: Dynamically-feasible trajectory forecasting with heterogeneous data. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XVIII 16, pages 683-700. Springer, 2020. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.162 + ], + "angle": 0, + "content": "[35] Liushuai Shi, Le Wang, Chengjiang Long, Sanping Zhou, Mo Zhou, Zhenxing Niu, and Gang Hua. Sgcn: Sparse graph convolution network for pedestrian trajectory prediction. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8994-9003, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.164, + 0.905, + 0.233 + ], + "angle": 0, + "content": "[36] Xiao Song, Kai Chen, Xu Li, Jinghan Sun, Baocun Hou, Yong Cui, Baochang Zhang, Gang Xiong, and Zilie Wang. Pedestrian trajectory prediction based on deep convolutional LSTM network. IEEE Transactions on Intelligent Transportation Systems, 22(6):3285-3302, 2020. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.235, + 0.905, + 0.303 + ], + "angle": 0, + "content": "[37] Yuchao Su, Jie Du, Yuanman Li, Xia Li, Rongqin Liang, Zhongyun Hua, and Jiantao Zhou. Trajectory forecasting based on prior-aware directed graph convolutional neural network. IEEE Transactions on Intelligent Transportation Systems, 23(9):16773-16785, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.305, + 0.905, + 0.36 + ], + "angle": 0, + "content": "[38] Chuhua Wang, Yuchen Wang, Mingze Xu, and David J Crandall. Stepwise goal-driven networks for trajectory prediction. IEEE Robotics and Automation Letters, 7(2):2716-2723, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.363, + 0.905, + 0.432 + ], + "angle": 0, + "content": "[39] Conghao Wong, Beihao Xia, Ziming Hong, Qinmu Peng, Wei Yuan, Qiong Cao, Yibo Yang, and Xinge You. View vertically: A hierarchical network for trajectory prediction via fourier spectrums. In European Conference on Computer Vision, pages 682-700. Springer, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.434, + 0.905, + 0.489 + ], + "angle": 0, + "content": "[40] Conghao Wong, Beihao Xia, Qinmu Peng, and Xinge You. Another vertical view: A hierarchical network for heterogeneous trajectory prediction via spectrums. arXiv preprint arXiv:2304.05106, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.491, + 0.905, + 0.572 + ], + "angle": 0, + "content": "[41] Conghao Wong, Beihao Xia, Ziqian Zou, Yulong Wang, and Xinge You. Socialcircle: Learning the angle-based social interaction representation for pedestrian trajectory prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19005-19015, 2024. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.575, + 0.905, + 0.631 + ], + "angle": 0, + "content": "[42] Beihao Xia, Conghao Wong, Qinmu Peng, Wei Yuan, and Xinge You. Cscnet: Contextual semantic consistency network for trajectory prediction in crowded spaces. Pattern Recognition, 126:108552, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.633, + 0.905, + 0.701 + ], + "angle": 0, + "content": "[43] Song Xiao, Kai Chen, Xiaoxiang Ren, and Haitao Yuan. Pedestrian trajectory prediction in heterogeneous traffic using facial keypoints-based convolutional encoder-decoder network. ACM Transactions on Internet Technology, 22(4): 1-14, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.704, + 0.905, + 0.785 + ], + "angle": 0, + "content": "[44] Chenxin Xu, Maosen Li, Zhenyang Ni, Ya Zhang, and Siheng Chen. Groupnet: Multiscale hypergraph neural networks for trajectory prediction with relational reasoning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6498-6507, 2022. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.788, + 0.905, + 0.857 + ], + "angle": 0, + "content": "[45] Chenxin Xu, Weibo Mao, Wenjun Zhang, and Siheng Chen. Remember intentions: Retrospective-memory-based trajectory prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6488-6497, 2022. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.859, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[46] Chenxin Xu, Robby T Tan, Yuhong Tan, Siheng Chen, Yu Guang Wang, Xinchao Wang, and Yanfeng Wang. Eqmotion: Equivariant multi-agent motion prediction with invari" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "22474" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.119 + ], + "angle": 0, + "content": "ant interaction reasoning. arXiv preprint arXiv:2303.10876, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.189 + ], + "angle": 0, + "content": "[47] Yu Yao, Ella Atkins, Matthew Johnson-Roberson, Ram Vasudevan, and Xiaoxiao Du. Bitrap: Bi-directional pedestrian trajectory prediction with multi-modal goal estimation. IEEE Robotics and Automation Letters, 6(2):1463-1470, 2021. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.192, + 0.482, + 0.26 + ], + "angle": 0, + "content": "[48] Ziyi Yin, Ruijin Liu, Zhiliang Xiong, and Zejian Yuan. Multimodal transformer networks for pedestrian trajectory prediction. In Proceedings of the Twenty-Ninth International Joint Conference on Artificial Intelligence, pages 1259-1265, 2021. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.263, + 0.482, + 0.332 + ], + "angle": 0, + "content": "[49] Cunjun Yu, Xiao Ma, Jiawei Ren, Haiyu Zhao, and Shuai Yi. Spatio-temporal graph transformer networks for pedestrian trajectory prediction. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XII 16, pages 507-523, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.334, + 0.482, + 0.375 + ], + "angle": 0, + "content": "[50] Wenjian Yu and Dirk Helbing. Game theoretical interactions of moving agents. In Simulating Complex Systems by Cellular Automata, pages 219-239. Springer, 2010. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.376, + 0.482, + 0.444 + ], + "angle": 0, + "content": "[51] Ye Yuan, Xinshuo Weng, Yanglan Ou, and Kris M Kitani. Agentformer: Agent-aware transformers for socio-temporal multi-agent forecasting. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9813-9823, 2021. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.446, + 0.482, + 0.488 + ], + "angle": 0, + "content": "[52] Jiangbei Yue, Dinesh Manocha, and He Wang. Human trajectory prediction via neural social physics. In European conference on computer vision, pages 376-394, 2022. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.49, + 0.482, + 0.545 + ], + "angle": 0, + "content": "[53] Yisong Yue, Patrick Lucey, Peter Carr, Alina Bialkowski, and Iain Matthews. Learning fine-grained spatial models for dynamic sports play prediction. In 2014 IEEE international conference on data mining, pages 670-679. IEEE, 2014. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.547, + 0.483, + 0.615 + ], + "angle": 0, + "content": "[54] Pu Zhang, Wanli Ouyang, Pengfei Zhang, Jianru Xue, and Nanning Zheng. Sr-lstm: State refinement for LSTM towards pedestrian trajectory prediction. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12085-12094, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.617, + 0.482, + 0.671 + ], + "angle": 0, + "content": "[55] Xiaoping Zheng and Yuan Cheng. Conflict game in evacuation process: A study combining cellular automata model. Physica A: Statistical Mechanics and its Applications, 390 (6):1042-1050, 2011. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.671 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22475" + } + ] +] \ No newline at end of file diff --git a/2025/SocialMOIF_ Multi-Order Intention Fusion for Pedestrian Trajectory Prediction/74c8547f-5c13-42b2-b27c-77ef8f52058e_origin.pdf b/2025/SocialMOIF_ Multi-Order Intention Fusion for Pedestrian Trajectory Prediction/74c8547f-5c13-42b2-b27c-77ef8f52058e_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..3741a5e8d17145a44e6c52fa1ee65481c92984d6 --- /dev/null +++ b/2025/SocialMOIF_ Multi-Order Intention Fusion for Pedestrian Trajectory Prediction/74c8547f-5c13-42b2-b27c-77ef8f52058e_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fa4590388acbb34618c8586120189e94eaa70f31939728cb565a26891f1c9d23 +size 4502729 diff --git a/2025/SocialMOIF_ Multi-Order Intention Fusion for Pedestrian Trajectory Prediction/full.md b/2025/SocialMOIF_ Multi-Order Intention Fusion for Pedestrian Trajectory Prediction/full.md new file mode 100644 index 0000000000000000000000000000000000000000..9fbe716f13ced79ad4fa8d0579f5666bf31fff86 --- /dev/null +++ b/2025/SocialMOIF_ Multi-Order Intention Fusion for Pedestrian Trajectory Prediction/full.md @@ -0,0 +1,339 @@ +# SocialMOIF: Multi-Order Intention Fusion for Pedestrian Trajectory Prediction + +Kai Chen $^{1*}$ , Xiaodong Zhao $(\boxtimes)$ ^{1*}, Yujie Huang $^{1}$ , Guoyu Fang $^{1}$ , Xiao Song $^{2}$ , Ruiping Wang $^{3}$ , Ziyuan Wang $^{1}$ + +1 Nanjing University of Aeronautics and Astronautics + +2 Beihang University 3 Nanyang Technological University + +{chen_kai, xdzhao, yj_huang, fanggy, zy_wang}@nuaa.edu.cn + +songxiao@buaa.edu.cn, ruiping.wang@ntu.edu.sg + +# Abstract + +The analysis and prediction of agent trajectories are crucial for decision-making processes in intelligent systems, with precise short-term trajectory forecasting being highly significant across a range of applications. Agents and their social interactions have been quantified and modeled by researchers from various perspectives; however, substantial limitations exist in the current work due to the inherent high uncertainty of agent intentions and the complex higher-order influences among neighboring groups. SocialMOIF is proposed to tackle these challenges, concentrating on the higher-order intention interactions among neighboring groups while reinforcing the primary role of first-order intention interactions between neighbors and the target agent. This method develops a multi-order intention fusion model to achieve a more comprehensive understanding of both direct and indirect intention information. Within SocialMOIF, a trajectory distribution approximator is designed to guide the trajectories toward values that align more closely with the actual data, thereby enhancing model interpretability. Furthermore, a global trajectory optimizer is introduced to enable more accurate and efficient parallel predictions. By incorporating a novel loss function that accounts for distance and direction during training, experimental results demonstrate that the model outperforms previous state-of-the-art baselines across multiple metrics in both dynamic and static datasets. + +# 1. Introduction + +Trajectory prediction is of considerable importance for intelligent transportation and surveillance systems. Studies indicate that trajectories derived from the physical + +![](images/a5b16720cb4c521b66b0bc3423abd877df5fe2e2275fcd0e4f0bd6c9b45791e2.jpg) +Figure 1. The process of SocialMOIF includes: (1) extracting the position, velocity and trajectory information of agents, (2) achieving multi-order fused intention modeling, (3) guiding the distribution of trajectories and performing global optimization. Finally we obtain accurate parallel predictions. + +world encompass rich spatiotemporal intention information [15, 20, 26, 29, 45]. Therefore, a significant challenge arises in effectively capturing (data modeling) and analyzing (encoding-decoding) processes related to the complex and diverse historical social intentions of agents, while linking this information to future trajectory predictions for enhanced efficiency. + +History social intentions. Existing research can be broadly classified into two main categories: knowledge-driven methods primarily depend on expert knowledge. For instance, biomimetic method [41] achieved agent localization and intention interaction modeling by simulating the positioning of marine animals. In contrast, certain studies utilized game-theoretic approaches to analyze + +social intentions through the simulation of agent flows [13, 50] and evacuation processes [14, 55], relying on explicit rules or formulas. However, the behavioral patterns and interaction rules among agents in social scenarios displayed high nonlinearity and uncertainty [26, 39, 40]. Data-driven methods implicitly learn interactions among agents through neural networks. For instance, LSTM-based methods [1, 11] continuously refined social pooling patterns, while attention-based approaches [7, 49, 51] dynamically adjusted agents' focal points in global interactions from a spatiotemporal integration perspective to optimize predicted trajectories. These methods primarily focused on straightforward interactions or direct responses between the target agent and its neighbors, often neglecting the higher-order, multidimensional influences exerted by surrounding agents on the target agent's intentions. Kim et al. [16] modeled agent intentions from 1st-order to N-order equally. However, when the number of agents was large, it may prioritized higher-order intentions, thereby diminishing the primary impact of lower-order intentions. Thus, the development of a multi-order fusion network that reinforces the dominant role of direct interactions while considering the secondary influence of higher-order transmissions on the target is essential. + +Associating future trajectory information. Knowledge-driven methods face limitations due to the inherent flaws of unidirectional inference, which hinder effective association with future trajectory information. Data-driven methods exhibit two main shortcomings: first, agent movements are temporally dynamic, and relying solely on distance loss fails to capture these dynamic variations. Second, in generative model approaches [25, 28, 34, 38, 47], future trajectory information is implicitly involved in the training process without explicit guidance, resulting in poor interpretability. In safety-sensitive tasks such as autonomous driving, it is necessary to utilize meaningful and goal-oriented parameters and models. + +Efficiency. Existing methods predominantly emphasize upstream tasks. For example, LSTM-based approaches [5, 6, 36, 43, 54] modeled trajectory data in a serial temporal manner, while Transformer-based methods [9, 19, 48, 49] analyzed trajectory information through parallel processing. However, in downstream tasks, trajectories are often generated independently and sequentially. This method not only accumulates errors but also limits the efficiency of trajectory prediction. Therefore, establishing a trajectory parallel prediction model that accounts for global influences across future periods is a promising area for further exploration. + +As illustrated in Fig. 1, SocialMOIF is proposed to address the trajectory prediction task. A Multi-Order Intention Fusion model (MOIF) is introduced to guide the learning direction of the network. MOIF includes a first-order intention interaction layer and a higher-order intention in + +teraction layer. The first layer focuses on direct interaction information between neighbors and the target agent, while the second layer addresses the indirect interaction information conveyed to the target agent after multiple interactions among neighbor targets. Inspiration is drawn from the squeeze theorem to design a trajectory distribution approximator to guide trajectories toward a distribution that closely aligns with ground truth and to enhance model interpretability. Simultaneously, RNN is incorporated to update the latent variables, achieving explicit supervision during training. The temporal dimension of decoded trajectories is aggregated to design a global trajectory optimizer. To the best of our knowledge, KANs [23, 24] are firstly introduced for the global optimization of trajectories, enabling more accurate and efficient parallel predictions. Furthermore, considering the travel directions of the agents, a distance-direction fused loss function is designed to comprehensively supervise the dynamic states of agents. + +In summary, the main contributions of this paper are: + +- The proposal of SocialMOIF, which incorporates four novel designs: + +(1) MOIF is introduced to comprehensively capture interaction information between the target agent and its neighbors; + +(2) A trajectory distribution approximator is developed to approximate and explicitly update the distribution of latent variables, guiding the distribution of trajectories; + +(3) A global trajectory optimizer is designed, and KANs are introduced for the first time to achieve parallel trajectory optimization across the global spatiotemporal domain; + +(4) A new direction-distance fused loss function is proposed for more effective supervision of agents' dynamic changes. + +- Comprehensive comparative experiments are conducted with existing high-quality models, demonstrating the effective and robust predictive capabilities of SocialMOIF. + +# 2. Related works + +Knowledge-driven methods. Helbing et al. [12] first proposed the social force model, which describes pedestrian interactions through an energy potential field. Best et al. [3] introduced a Bayesian inference-based intention inference model, integrating observed movement trajectories with prior knowledge. EqMotion [46] employed geometric analysis to convert human-human interaction trajectory analysis into an optimization problem by examining object geometry. Morris et al. [31] proposed an implicit Markov model for trajectory prediction, which is utilized for spatiotemporal probabilistic modeling of various types of pedestrian trajectories. A recent approach, socialcircle [41], incorporated angle-based light rule constraints into the trainable back + +bone to achieve interaction modeling of agent intentions. + +Despite these advancements, these methods rely on knowledge from specific domains and therefore struggle to fully capture the complex agent intentions within social scenarios. Although attempts have been made to incorporate future trajectory information to overcome the limitations of unidirectional inference, the model's equations remain overly sensitive to common variations, leading to persistent challenges in trajectory prediction tasks [42, 52]. + +Data-driven methods. LSTM-based methods model trajectory data temporally. Alahi et al. [1] employed a "social pooling mechanism" that considers only the intentions of agents within a grid while neglecting interactions with boundary agents. Gupta et al. [11] improved upon this by implementing a pooling mechanism that treats pedestrians in neighboring areas equally. Transformer-based methods enhanced temporal modeling approaches. Yu et al. [49] modeled and predicted trajectories using a spatio-temporal concatenation approach. Yin et al. [48] introduced optical flow to simulate agents' motion intentions. Graph neural networks were also utilized for trajectory prediction to model intention interactions among different nodes [8, 22, 30, 35, 37, 44]. Diffusion models [10, 27] were used to explore trajectory prediction based on randomness. + +However, most of these methods primarily emphasize direct interactions between the target agent and its neighbors, often overlooking the higher-order influences that neighbors exert on the target agent. Furthermore, many of these approaches incorporate future trajectory information only at the network's output stage, supervising agents' trajectory changes solely through distance loss. Moreover, the limited interpretability of latent variable distributions in generative models can impede further development of these models. + +In addition, the aforementioned methods mainly emphasize trajectory analysis and modeling of upstream tasks. For output in downstream tasks, results are generally produced independently in a sequential manner over time. This sequential processing introduces delays that may result in untimely responses, potentially leading to safety concerns. Furthermore, each future prediction is dependent on the state from the preceding time step, which may lead to accumulated errors, adversely affecting prediction accuracy. + +# 3. Approach + +The number of agents in the scenario is set to $(N_{n} + 1)$ . The trajectory of the target agent $i$ in history period $T_{H}$ is $g_{i}^{1:T_{H}}$ , and $g_{i}^{t}$ is the two-dimensional spatial coordinate at moment $t$ . The remaining agents are tentatively designated as neighbors, and the number is $N_{n}$ . The trajectory of $i$ in the future period to be predicted is denoted as $\hat{g}_{i}^{T_{H} + 1:T_{F}}$ . We propose the multi-order intention fusion model in Section 3.1, the trajectory distribution approximator in Section 3.2, the global trajectory optimizer in Section 3.3, and the + +direction-distance fusion loss function in Section 3.4. + +# 3.1. Multi-order intention fusion model + +The model consists of two layers: a first-order intention interaction layer and a higher-order intention interaction layer. As shown in Fig. 2, the first-order layer emphasizes the direct, dominant influence of neighbors on the target agent's intention. The higher-order layer accounts for the indirect and secondary influences on the target agent's intention that arise from interactions within the group of neighbors, thereby enabling a more comprehensive understanding of intentions. Based on the current and final history motion states of the agents, input components for the layers are first constructed, followed by the extraction and fusion of multi-order intention information. + +Higher-order intention interaction layer. From a more fundamental perspective, the higher-order intention interactions within the group of neighbors can be decomposed into first-order intention interactions between individual neighbors. We define the intention influence of neighbor agent $a$ on neighbor agent $b$ as $w_{a\rightarrow b}$ ; for an agent itself, it is represented as $w_{a\rightarrow a}$ . In a scenario with $N_{n}$ agents, the total number of such intentions is $\Omega$ : + +$$ +\Omega = \left(N _ {n} - 1\right) N _ {n} + N _ {n} = N _ {n} ^ {2} \tag {1} +$$ + +To obtain the intentions, we extract the absolute positions and velocities of the $N_{n}$ neighbors corresponding to the target agent $i$ . They are processed as a layer input component through the embedding layer $\pmb{f}_{S}$ to obtain $U_{i}$ . + +$$ +U _ {i} = \boldsymbol {f} _ {U} \left(x _ {1: N _ {n}} ^ {1: T _ {H}}, y _ {1: N _ {n}} ^ {1: T _ {H}}, v _ {x 1: N _ {n}} ^ {1: T _ {H}}, v _ {y 1: N _ {n}} ^ {1: T _ {H}}\right) \tag {2} +$$ + +We capture the intention interactions of neighbors through multiple subspaces in parallel. $U_{i}$ is mapped into a learnable query vector $Q_{U}$ , a key vector $K_{U}$ , and a value vector $V_{U}$ . The feature dimensions are flattened into M subspaces. For the m-th subspace, the overall intention coefficient matrix within neighbors can be expressed as: + +$$ +W _ {U} ^ {m} = \operatorname {S o f t M a x} \left(\frac {Q _ {U} ^ {m} \left(K _ {U} ^ {m}\right) ^ {T}}{d _ {M}}\right) \tag {3} +$$ + +Thus, we obtain $w_{a\to b} = [W_U^m ]_{ab}$ . For this matrix, the number of elements is $\Omega$ + +The first-order intention interaction layer. The layer input component of the target agent $i$ is processed through the embedding layer $\pmb{f}_{S}$ to obtain $S_{i}$ . + +$$ +S _ {i} = \pmb {f} _ {S} \left(x _ {i} ^ {1: T _ {H}}, y _ {i} ^ {1: T _ {H}}, d _ {i \leftrightarrow 1: N _ {n}} ^ {1: T _ {H}}, \theta_ {i \leftrightarrow 1: N _ {n}} ^ {1: T _ {H}}, e _ {i \leftrightarrow 1: N _ {n}} ^ {1: T _ {H}}\right) (4) +$$ + +We intensify the dominant role of direct interaction in components and focus on the design of interaction rules. For the present case, $d_{i\leftrightarrow j}^{1:T_H}$ is the distance between the target + +![](images/ee95a298984c868f3ddb918b355f95a29b441c8107ebc48cb0487ce9d0e4ca5a.jpg) +Figure 2. Computational pipeline of SocialMOIF. Firstly, the first-order interaction layer focuses on direct interactions between the target agent and its neighbors. The higher-order interaction layer addresses the interactions among neighbors. Multi-order intention is obtained by fusing the information from both layers. This multi-order intention is then combined with the actual future trajectory, with the distribution of the trajectory guided by an approximator. An optimizer is subsequently employed to enhance the decoded trajectory. Additionally, a fused distance-direction loss function supervises the training process. Finally, efficient parallel prediction is achieved. + +agent $i$ and neighbor $j$ , and $\theta_{i\leftrightarrow j}^{1:T_H}$ is the velocity angle. For the final history, $e_{i\leftrightarrow j}^{1:T_H}$ is the final distance in the history period, assuming that after moment $t$ , both the target agent $i$ and neighbor $j$ maintain the original velocity in the tangent direction $\overrightarrow{vx_i^t},\overrightarrow{vx_j^t}$ ; then, $e_{ij}^t$ can be expressed as: + +$$ +\lambda = \left| \left(\overrightarrow {d _ {i j} ^ {t}} \cdot \overrightarrow {v _ {i j} ^ {t}}\right) / \left\| \overrightarrow {v _ {i j} ^ {t}} \right\| ^ {2} \right| \tag {5} +$$ + +$$ +e _ {i j} ^ {t} = \left\| \overrightarrow {d _ {i j} ^ {t}} + \min (\lambda , T _ {H} - t) \cdot \left(\overrightarrow {v x _ {j} ^ {t}} - \overrightarrow {v y _ {i} ^ {t}}\right) \right\| \tag {6} +$$ + +Similarly, $S_{i}$ is mapped into learnable query vectors $Q_{S}$ , key vectors $K_{S}$ , and value vectors $V_{S}$ . The matrix of the intention coefficients between the target agent $i$ and its neighbors is as follows: + +$$ +W _ {S} = \operatorname {S o f t M a x} \left(\frac {Q _ {S} \left(K _ {S}\right) ^ {T}}{d _ {S}}\right) \tag {7} +$$ + +Intention Fusion. Higher-order interaction intentions are conducted to the target agent and combined with first-order interaction intentions to achieve multi-order intention fusion. During this fusion, the higher-order interaction intentions captured within each subspace exert varying effects + +when applied to the target agent. We set the corresponding trainable influence factor $\eta_{m}$ to denote this influence. The attention score $A_{i}$ is denoted as: + +$$ +A _ {i} = \left(\sum_ {m = 1} ^ {M} \eta_ {m} W _ {U} ^ {m} + W _ {S}\right) V _ {S} \tag {8} +$$ + +where $A_{i}$ is processed by two skip connections and a final fully connected layer to obtain the multi-order fused intention $I_{i}$ . + +# 3.2. Trajectory distribution approximator + +We define the trajectory $g_{igt}^{T_H + 1:T_F}$ of the target agent $i$ in the future period. $B_{i}$ is obtained after feature enhancement through the embedding layer $f_{B}$ . By utilizing $I_{i}$ as the lower bound and $B_{i}$ as the upper bound, the latent variable distribution in the potential space is explicitly approximated. First, the probability distribution of the target agent $i$ 's trajectory is initially modeled as: + +$$ +\begin{array}{l} p \left(g _ {i} ^ {T _ {H} + 1: T _ {F}} \mid U _ {i}, S _ {i}\right) \\ = \prod_ {t = T _ {H} + 1} ^ {T _ {F}} p \left(g _ {i} ^ {T _ {H} + 1 + t} \mid g _ {i} ^ {T _ {H} + 1: T _ {H} + t - 1}, U _ {i}, S _ {i}\right) \tag {9} \\ \end{array} +$$ + +We sample the noise variables $\varepsilon \sim N(0,1)$ and reparameterize the latent variables that are approximated at $t$ [17]: + +$$ +z _ {i \varphi} ^ {t} = u _ {i \varphi} ^ {t} + \varepsilon \cdot \sigma_ {i \varphi} ^ {t} \quad q _ {\varphi} (\cdot | I _ {i} ^ {t - 1}, B _ {i} ^ {t}) \sim N (u _ {i \varphi} ^ {t}, \sigma_ {i \varphi} ^ {t}) \tag {10} +$$ + +where $\varphi$ are parameters that need to be optimized during the training process. In the test mode, the latent variables $z_{i\vartheta}^{t}$ are only reasoned by $I_{i}$ , and $\vartheta$ is also involved in the training process of the model. + +We sample and decode the target agent's trajectories on the basis of latent variables and multi-order fused intention information. + +$$ +g _ {i} ^ {t} \sim \left\{ \begin{array}{l l} p _ {\delta} (\cdot | z _ {i \varphi} ^ {t}, I _ {i} ^ {t - 1}) & \text {i f t r a i n} \\ p _ {\delta} (\cdot | z _ {i \vartheta} ^ {t}, I _ {i} ^ {t - 1}) & \text {i f t e s t} \end{array} \right. \tag {11} +$$ + +where $\delta$ are parameters that need to be optimized during the training process. + +Concat the latent variables with the trajectories generated at the current time $t$ , updating $I_{i}^{t - 1}$ via $f_{zg}$ . + +$$ +I _ {i} ^ {t} = \left\{ \begin{array}{l l} \boldsymbol {f} _ {z g} \left(\operatorname {c o n c a t} \left(z _ {i \varphi} ^ {t}, g _ {i} ^ {t}\right), I _ {i} ^ {t - 1}\right) & \text {i f t r a i n} \\ \boldsymbol {f} _ {z g} \left(\operatorname {c o n c a t} \left(z _ {i \vartheta} ^ {t}, g _ {i} ^ {t}\right), I _ {i} ^ {t - 1}\right) & \text {i f t e s t} \end{array} \right. \tag {12} +$$ + +In the next time step, the updated multi-order fused intention contains the information of the latent variables, which is then combined with $B_{i}^{t}$ to approximate the latent variables and update them. + +By guiding the trajectory distribution through such a training process, the generated trajectories are explicitly represented by the latent variables as: + +$$ +\begin{array}{l} p \left(g _ {i} ^ {T _ {H} + 1: T _ {F}} \mid U _ {i}, S _ {i}\right) \\ = \prod_ {t = T _ {H} + 1} ^ {T _ {F}} \int_ {z _ {i \varphi} ^ {t}} p \left(g _ {i} ^ {t} \mid g _ {i} ^ {T _ {H} + 1: t - 1}, I _ {i} ^ {t - 1}, B _ {i} ^ {t}, z _ {i \varphi} ^ {t}\right) \cdot \tag {13} \\ p \left(z _ {i \varphi} ^ {t} \mid g _ {i} ^ {T _ {H} + 1: t - 1}, I _ {i} ^ {t - 1}, B _ {i} ^ {t}\right) d z _ {i \varphi} ^ {t} \\ \end{array} +$$ + +In test mode, $B_{i}^{t}$ is removed, and $\varphi$ is replaced with $\vartheta$ . + +# 3.3. Global trajectory optimizer + +For efficient prediction, we introduce the KANs for the first time. We aggregate the $g_{i}^{T_{H} + 1:T_{F}}$ time dimension to obtain $\Gamma^0\in \mathbb{R}^{2T_F}$ to facilitate global parallel optimization. The depth of the optimizer is $L$ layers, the number of nodes in the $\ell$ -th layer is $m$ , the number of nodes in the $\ell +1$ -th layer is $n$ , and the activation function matrix corresponding to the nodes in the $l$ -th layer is $\Phi_{\ell}$ . The optimization process for the $\ell \rightarrow \ell +1$ layer is as follows: + +$$ +\Gamma^ {\ell + 1} = \Phi_ {\ell} \Gamma^ {\ell} = \left[ \begin{array}{c c c c} \varphi_ {1, 1} ^ {\ell} & \varphi_ {1, 2} ^ {\ell} & \dots & \varphi_ {1, m} ^ {\ell} \\ \varphi_ {2, 1} ^ {\ell} & \varphi_ {2, 2} ^ {\ell} & \dots & \varphi_ {2, m} ^ {\ell} \\ \vdots & \vdots & \ddots & \vdots \\ \varphi_ {n, 1} ^ {\ell} & \varphi_ {n, 2} ^ {\ell} & \dots & \varphi_ {n, m} ^ {\ell} \end{array} \right] \Gamma^ {\ell} (1 4) +$$ + +Eventually, the predicted trajectory is reduced back to the original dimensional space to obtain the final trajectory over the entire prediction period at once. + +$$ +\hat {g} _ {i} ^ {T _ {H} + 1: T _ {F}} \leftarrow \Gamma^ {L} = \left(\Phi_ {L} \circ \Phi_ {L - 1} \circ \dots \circ \Phi_ {1} \circ \Phi_ {0}\right) \Gamma^ {0} (1 5) +$$ + +Given the diversity of agent intentions, the process predicts multiple possible trajectories. + +# 3.4. Loss function + +Agents traveling in opposite or the same direction exert distinct effects on their respective movement states. Therefore, direction loss is incorporated to supervise the agent's dynamics during backpropagation. + +For distance. The distance prediction error in future periods can be expressed as: + +$$ +L _ {d i s} = \left\| \hat {g} _ {i} ^ {t} - g _ {i g t} ^ {t} \right\| \tag {16} +$$ + +For direction, the final moment of the historical period is introduced. Each predicted time step is related to the preceding time step to construct the direction vector, and subsequently, the angle between the predicted trajectory and the actual trajectory direction vector is analyzed. + +$$ +L _ {a n g l e} = - \operatorname {a r c c o s} \left(\frac {\left(\hat {g} _ {i} ^ {t} - \hat {g} _ {i} ^ {t + 1}\right) \cdot \left(g _ {i g t} ^ {t} - g _ {i g t} ^ {t + 1}\right)}{\left\| \hat {g} _ {i} ^ {t} - \hat {g} _ {i} ^ {t + 1} \right\| \left\| g _ {i g t} ^ {t} - g _ {i g t} ^ {t + 1} \right\|}\right) \tag {17} +$$ + +The objective of our model is to maximize the lower bound on the evidence for all time steps of the target distribution, joints (16) and (17), with a final loss function of: + +$$ +\begin{array}{l} L = \sum_ {t = T _ {H} + 1} ^ {T _ {F}} \left\{\mathbb {E} _ {z _ {i \varphi} ^ {t} \sim q _ {\varphi} \left(\cdot | B _ {i} ^ {t}, I _ {i} ^ {t - 1}\right)} \left[ L _ {d i s} + L _ {a n g l e} \right] \right. \tag {18} \\ \left. - K L \left[ q _ {\varphi} \left(z _ {i \varphi} ^ {t} \mid B _ {i} ^ {t}, I _ {i} ^ {t - 1}\right) \| q _ {\vartheta} \left(z _ {i \vartheta} ^ {t} \mid I _ {i} ^ {t - 1}\right) \right] \right\} \\ \end{array} +$$ + +# 4. Experiments + +Datasets. ETH [32] and UCY [18] datasets encompass over 1,600 pedestrian trajectories, with positions annotated at 0.4s intervals. SDD [33] dataset is primarily utilized for pedestrian trajectory prediction and object tracking, containing about 19,000 pedestrian trajectories, with positions annotated every 0.4s. NBA dataset focuses on the 2015-2016 season [21, 53]. Two subsets, "Rebounds" and "Scores," are extracted for benchmarking purposes. The movement intentions of the agents exhibit strong purposefulness and complex, hard-to-capture velocity variations. Consequently, the extracted trajectories encompass rich, highly nonlinear motion patterns. NuScenes [4] is a dataset for autonomous driving, and the prediction challenge segmentation and configurations are adhered to. + +Table 1. Comparison of Related Methods on ETH/UCY, NBA, SDD, and NuScenes (in meters). Every work was realized using best-of-20 predictions for the effect of ade/fde was selected. + +
ETH/UCYFLEAM [6]ST-MR [19]MTN [48]MANTRA [28]Agent-former [51]BiTrap [47]MemoNet [45]EqMotion [46]V2-Net-SC [39, 41]E-V2-Net-SC [40, 41]Ours
Eth0.66/1.210.53/0.690.58/0.860.56/0.970.58/0.860.66/1.090.51/0.690.36/0.540.23/0.330.23/0.300.26/0.31
Hotel0.37/0.640.20/0.310.19/0.570.17/0.320.15/0.240.17/0.320.14/0.200.12/0.160.11/0.160.10/0.130.10/0.12
Univ0.57/0.970.41/0.640.32/0.430.38/0.790.26/0.480.26/0.460.26/0.520.21/0.390.17/0.270.18/0.240.11/0.17
Zara010.46/0.740.26/0.380.24/0.370.24/0.410.19/0.320.23/0.470.21/0.360.16/0.270.13/0.180.13/0.160.10/0.17
Zara020.37/0.690.18/0.240.15/0.210.18/0.340.16/0.250.18/0.350.15/0.300.11/0.190.12/0.210.11/0.160.09/0.14
Avg0.48/0.850.32/0.450.30/0.490.31/0.570.27/0.430.30/0.540.25/0.410.19/0.310.15/0.230.15/0.200.13/0.18
NBAFLEAM [6]ST-MR [19]MTN [48]SGNet [38]EqMotion [46]MemoNet [45]BiTrap [47]GroupNet +CVAE [44]V2-Net-SC [39, 41]E-V2-Net-SC [40, 41]Ours
Rebound1.92/2.210.94/1.730.84/0.970.80/1.780.86/1.211.73/2.060.86/1.880.63/0.840.54/0.790.54/0.860.34/0.66
Scores1.29/1.680.90/1.030.87/1.410.71/1.450.69/1.021.35/1.870.76/1.550.46/0.760.50/0.840.47/0.780.30/0.56
SDDFLEAM [6]SHENet [29]MemoNet [45]MANTRA [28]BiTrap [47]MID [10]Y-Net [26]NSP-SFM [52]V2-Net-SC [39, 41]E-V2-Net-SC [40, 41]Ours
SDD0.67/1.280.44/0.530.45/0.640.49/0.620.46/0.590.38/0.730.32/0.660.21/0.340.26/0.360.23/0.350.17/0.24
NuScenesFLEAM [6]ST-MR [19]MTN [48]MANTRA [28]SGNet [38]Agent-former [51]EqMotion [46]Y-Net [26]V2-Net-SC [39, 41]E-V2-Net-SC [40, 41]Ours
NuScenes2.02/3.141.84/2.461.64/2.121.47/2.011.28/2.091.04/1.471.24/1.981.12/1.711.06/1.591.04/1.640.92/1.56
+ +Metrics. Best-of-20 predictions were used to compute Average Displacement Error (ADE) and Final Displacement Error (FDE) as primary metrics, with estimated Negative Log Likelihood (NLL) reported from the test set. See their detailed definitions in the Supplementary Material. + +Implementation Details. The model was trained on two NVIDIA RTX 4090 GPUs. Following the methodology in [5, 36], a leave-one-out cross-validation approach was employed for dataset handling. The number of subspaces in the higher-order intention interaction layer was set to $M = 6$ while the depth of the global trajectory optimizer was set to $L = 3$ . During inference, the best parameter model was loaded to predict the next 8 frames based on the previous 8 frames of history. For more implementation details of the training in the Supplementary Material. + +# 4.1. Quantitative analysis + +All SOTA performances across datasets were highlighted in bold black in Table 1. In Eth scenario, crowds generally moved in parallel, resulting in weaker interactions among neighbors that conducted to target agent. E-V2-Net-SC [40, 41] achieved optimal performance in this scenario. However, the proposed method still achieved superior performance on ETH/UCY dataset, reducing ADE and FDE by $13.33\%$ and $10.00\%$ compared to E-V2-Net-SC, respectively. For NBA dataset, where interactions between agents were complex and frequent, the proposed method demonstrated exceptional performance. In Rebound sub-dataset, ADE and FDE decreased by $37.04\%$ and $16.46\%$ , respectively, compared to existing SOTA methods. In Scores sub-dataset, ADE and FDE were reduced by $34.78\%$ and $26.32\%$ , respectively, relative to the current SOTA methods. For SDD dataset, image segmentation techniques, as referenced in [2], were employed to capture pedestrian position information across eight scenario. Raw pixel coordinates were converted into spatial coordinates defined in meters for trajectory prediction tasks. NSP-SFM [52] outperformed both V2-Net-SC and E-V2-Net-SC. Our method re + +duced ADE and FDE by $19.05\%$ and $29.41\%$ , respectively, compared to NSP-SFM. For NuScenes dataset, V2-Net-SC and E-V2-Net-SC showed similar performance levels. Our method reduced ADE and FDE by $11.54\%$ and $1.87\%$ , respectively, in comparison to their SOTA performances. + +Detailed comparative experiments with NLL metric were carried out in the Supplementary Material. Furthermore, for wider validation and analysis, we applied directional component in the designed loss function to different models. See the Supplementary Material for details. + +# 4.2. Qualitative analysis + +![](images/5e993945fc8199ac2be38e1d22beafcd37dcf953f777dfad5392d7a7801e8b21.jpg) +Figure 3. Representative cases. Row 1 represented complex historical trajectory cases. Row 2 represented state maintenance cases. Row 3 represented numerous neighbors cases. + +Unlike previous studies [6, 26, 39-41, 45, 46], the focus here was on conducting experiments involving historical + +![](images/4d03a52220cac1415396f477349406fc0a9f6c7b81ce2c3e7193d1868c4994d0.jpg) +Figure 4. Predictions comparison of predictions in ETH/UCY. The comparison included heatmaps of the overall distribution of predictions and best-of-20 predictions of the related methods. + +trajectories in various social contexts. This enabled a more effective validation of the model's ability to extract and analyze multi-order fused intentions in Fig. 3. In panel (a), the target agent continuously changed direction. Although the model displayed slight deviations at certain moments, these deviations remained within the acceptable range of directional changes. In panel (b) and (c), the target agent intended to turn, but the model successfully achieved accurate predictions. The target agent moved at a relatively steady speed in panel (d), adopting a conservative deceleration strategy in response to numerous neighbors who are stopping or remaining stationary. In panel (e) and (f), there were many stationary neighbors who may be gathered together in conversation. Consequently, the target agent maintained its original state. In panel (g) and (h), most neighbors moved in the same direction as the target agent, with closer neighbors receiving higher attention scores. This suggested that the target agent prioritized attention to closer neighbors, consistent with social norms. In panel (i), the directions of neighbors were complex, and their speeds varied. The target agent accounted for influences from neighbors in multiple directions while attempting to anticipate and avoid imminent collisions. Please see the qualitative analyses of NBA interactions in the Supplementary Material. + +Comparisons were made between the proposed method and three previous methods in Fig. 4. It was observed that our method produced more accurate predictions than the previous methods. In Eth, when facing neighbors to the left front, the model provided a prediction closest to the true trajectory. The predictions from E-V2-Net-SC and EqMotion + +exhibited slight oscillations, while MemoNet, by mimicking retrospective memory mechanisms, tends to maintain its original trajectory. By employing the squeeze theorem to approximate future true trajectories, the proposed model shows greater convergence in the overall distribution of predictions in all scenarios. In Hotel and Zara01, there were many pedestrians ahead of the target agent, and the diverging trajectory distributions from previous methods represent rather aggressive strategies, which conflicted with the safety requirements of autonomous driving. + +# 4.3. Ablation study + +In Table 2, ablation experiments were conducted by sequentially adding components to the PDI baseline. Groups 1-4 focus on the multi-order intention fusion module. Group 1 includes only the neighbors' position and the distances between the target agent and the neighbors. As additional components, such as neighboring speed $\mathbf{V}$ , the angle between the target agent and the neighboring speed $\Theta$ , and the final predicted distance over the historical period $\mathbf{E}$ , were gradually incorporated, a substantial reduction in the final prediction error was observed. Specifically, ADE decreased by $23.91\%$ and FDE by $39.47\%$ on the SDD dataset, while ADE reduced by $11.11\%$ and FDE by $14.42\%$ on the Nuscenes dataset. Groups 4-5, 5-6, and 6-7 focus on the ablation of the approximator, the optimizer, and the angle loss term in the loss function, respectively. The results demonstrate that each component contributed positively. + +To examine the influence of $\mathbf{B}$ and $\mathbf{I}$ , the trajectory distribution for several scenarios in the future periods was visu- + +Table 2. Ablation studies of each component in our method on SDD and Nuscenes datasets. Notably, the positional information of the target agent was essential and was included in each set of ablation experiments. $\mathbf{P}$ represented the position of neighbors, $\mathbf{V}$ indicated their velocities, $\mathbf{D}$ denoted the distance between the target agent and its neighbors, and $\Theta$ represented the angle between their velocities. $\mathbf{E}$ represented the final predicted distance between the target agent and its neighbors over historical periods, $\mathbf{I}$ indicated global historical social intent, $\mathbf{B}$ denoted future actual trajectories, $\mathbf{K}$ referred to the parallel prediction mode, and $\mathbf{A}$ signified directional information. + +
GroupPVDΘEIBKASDDNuscenes
1✓✓✓0.46/0.711.26/2.15
2✓✓✓✓0.45/0.661.24/2.09
3✓✓✓✓✓0.41/0.561.18/1.97
4✓✓✓✓✓✓0.35/0.431.12/1.84
5✓✓✓✓✓✓✓0.30/0.361.04/1.76
6✓✓✓✓✓✓✓✓0.20/0.270.93/1.61
7✓✓✓✓✓✓✓✓✓0.17/0.240.92/1.56
+ +![](images/2706826b04e29baddbbc288d435acaf6cfecad63c3014a316e2d5bd7ee82c939.jpg) +Figure 5. Distribution heatmaps. Every heatmap of future trajectory distribution was shown above the real scene image. (a) represented the actual distribution. (b) presented the experimental group containing only I component. (c) showed the group with both I and B components. + +![](images/4765ac324e41e41878fe89a0229ce6f845a252c2fd285776c2ed00f3336e9b65.jpg) +Figure 6. Representative cases about the optimization effects of component $\mathbf{K}$ . To ensure fairness, we selected best-of-20 predictions in groups 6 and 7 for comparison. + +alized in Fig. 5. It was evident that when both components $\mathbf{B}$ and $\mathbf{I}$ were included, the predicted trajectory distribution aligned more closely with the actual distribution. It can be found that in a strongly regular sports scenario such as Rebound, (b) and (c) achieved excellent predictions, while in open environments with complex social contexts, the effec + +tiveness of the approximator was better demonstrated. + +To validate the effectiveness of component $\mathbf{K}$ under global parallel optimization, six challenging cases were presented in Fig. 6. It was evident that predictions without component $\mathbf{K}$ exhibited increasing errors as the prediction time accumulated, whereas the parallel prediction mode demonstrated a closer result to ground-truth. + +# 5. Conclusion + +Each human movement path is either directly or indirectly associated with neighbor agents. This study introduces SocialMOIF, which provides a more comprehensive representation of both direct and indirect intention interactions between the target agent and neighbor agents. SocialMOIF comprises four key designs: a multi-order intention fusion model, a trajectory distribution approximator, a global trajectory optimizer, and a novel loss function. Experimental results demonstrate that this approach performs competitively across various datasets and effectively approximates and explicitly updates trajectory distributions during training, thereby improving interpretability. + +# Acknowledgement + +This project is supported by National Natural Science Foundation of China (52202417), China Postdoctoral Science Foundation(2022TQ0155, 2022M721605), Open Project Program of State Key Laboratory of Virtual Reality Technology and Systems, Beihang University(VRLAB2023A02), Young Elite Scientists Sponsorship Program by CAST(2023QNRC001), Fundamental Research Funds for the Central Universities (NS2024030), Young Elite Scientists Sponsorship Program by JSTJ(JSTJ-2023-XH032). + +# References + +[1] Alexandre Alahi, Kratarth Goel, Vignesh Ramanathan, Alexandre Robicquet, Li Fei-Fei, and Silvio Savarese. Social lstm: Human trajectory prediction in crowded spaces. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 961–971, 2016. 2, 3 +[2] Stefan Becker, Ronny Hug, Wolfgang Hubner, and Michael Arens. An evaluation of trajectory prediction approaches and notes on the trajnet benchmark. arXiv preprint arXiv:1805.07663, 2018. 6 +[3] Graeme Best and Robert Fitch. Bayesian intention inference for trajectory prediction with an unknown goal destination. In 2015 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 5817-5823. IEEE, 2015. 2 +[4] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11621-11631, 2020. 5 +[5] Kai Chen, Xiao Song, and Xiaoxiang Ren. Pedestrian trajectory prediction in heterogeneous traffic using pose keypoints-based convolutional encoder-decoder network. IEEE Transactions on Circuits and Systems for Video Technology, 31(5): 1764-1775, 2020. 2, 6 +[6] Kai Chen, Xiao Song, Haitao Yuan, and Xiaoxiang Ren. Fully convolutional encoder-decoder with an attention mechanism for practical pedestrian trajectory prediction. IEEE Transactions on Intelligent Transportation Systems, 23(11): 20046-20060, 2022. 2, 6 +[7] Kai Chen, Haihua Zhu, Dunbing Tang, and Kun Zheng. Future pedestrian location prediction in first-person videos for autonomous vehicles and social robots. Image and Vision Computing, 134:104671, 2023. 2 +[8] Lingwei Dang, Yongwei Nie, Chengjiang Long, Qing Zhang, and Guiqing Li. Msr-gen: Multi-scale residual graph convolution networks for human motion prediction. In Proceedings of the IEEE/CVF international conference on computer vision, pages 11467-11476, 2021. 3 +[9] Francesco Giuliani, Irtiza Hasan, Marco Cristani, and Fabio Galasso. Transformer networks for trajectory forecasting. In 2020 25th international conference on pattern recognition (ICPR), pages 10335-10342. IEEE, 2021. 2 + +[10] Tianpei Gu, Guangyi Chen, Junlong Li, Chunze Lin, Yongming Rao, Jie Zhou, and Jiwen Lu. Stochastic trajectory prediction via motion indeterminacy diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17113-17122, 2022. 3, 6 +[11] Agrim Gupta, Justin Johnson, Li Fei-Fei, Silvio Savarese, and Alexandre Alahi. Social gan: Socially acceptable trajectories with generative adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2255–2264, 2018. 2, 3 +[12] Dirk Helbing and Peter Molnar. Social force model for pedestrian dynamics. Physical review E, 51(5):4282, 1995. 2 +[13] Serge Hoogendoorn and Piet HL Bovy. Simulation of pedestrian flows by optimal control and differential games. Optimal control applications and methods, 24(3):153-172, 2003. 2 +[14] SP Hoogendoorn, Winnie Daamen, Yaqing Shu, and Han Ligteringen. Modeling human behavior in vessel maneuver simulation by optimal control and game theory. Transportation research record, 2326(1):45-53, 2013. 2 +[15] Yihan Hu, Jiazhi Yang, Li Chen, Keyu Li, Chonghao Sima, Xizhou Zhu, Siqi Chai, Senyao Du, Tianwei Lin, Wenhai Wang, et al. Planning-oriented autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17853-17862, 2023. 1 +[16] Sungjune Kim, Hyung-gun Chi, Hyerin Lim, Karthik Ramani, Jinkyu Kim, and Sangpil Kim. Higher-order relational reasoning for pedestrian trajectory prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15251-15260, 2024. 2 +[17] Christos Kyrkou, Panayiotis Kolios, Theocharis Theocharides, and Marios Polycarpou. Machine learning for emergency management: A survey and future outlook. Proceedings of the IEEE, 111(1):19-41, 2022. 5 +[18] Alon Lerner, Yiorgos Chrysanthou, and Dani Lischinski. Crowds by example. In Computer graphics forum, pages 655-664. Wiley Online Library, 2007. 5 +[19] Lihuan Li, Maurice Pagnucco, and Yang Song. Graph-based spatial transformer with memory replay for multifuture pedestrian trajectory prediction. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2221-2231, 2022. 2, 6 +[20] Linhui Li, Xiaotong Lin, Yejia Huang, Zizhen Zhang, and Jian-Fang Hu. Beyond minimum-of-n: Rethinking the evaluation and methods of pedestrian trajectory prediction. IEEE Transactions on Circuits and Systems for Video Technology, 2024. 1 +[21] Kostya Linou. Nba-player-movements. Software for Visualization of NBA Games, 2016. 5 +[22] Congcong Liu, Yuying Chen, Ming Liu, and Bertram E Shi. Avgcn: Trajectory prediction using graph convolutional networks guided by human attention. In 2021 IEEE International Conference on Robotics and Automation (ICRA), pages 14234-14240. IEEE, 2021. 3 +[23] Ziming Liu, Pingchuan Ma, Yixuan Wang, Wojciech Ma-tusik, and Max Tegmark. Kan 2.0: Kolmogorov-arnold net + +works meet science. arXiv preprint arXiv:2408.10205, 2024. 2 +[24] Ziming Liu, Yixuan Wang, Sachin Vaidya, Fabian Ruehle, James Halverson, Marin Soljačić, Thomas Y Hou, and Max Tegmark. Kan: Kolmogorov-arnold networks. arXiv preprint arXiv:2404.19756, 2024. 2 +[25] Karttikeya Mangalam, Harshayu Girase, Shreyas Agarwal, Kuan-Hui Lee, Ehsan Adeli, Jitendra Malik, and Adrien Gaidon. It is not the journey but the destination: Endpoint conditioned trajectory prediction. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16, pages 759-776. Springer, 2020. 2 +[26] Karttikeya Mangalam, Yang An, Harshayu Girase, and Jitendra Malik. From goals, waypoints & paths to long term human trajectory forecasting. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15233–15242, 2021. 1, 2, 6 +[27] Weibo Mao, Chenxin Xu, Qi Zhu, Siheng Chen, and Yanfeng Wang. Leapfrog diffusion model for stochastic trajectory prediction. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5517-5526, 2023. 3 +[28] Francesco Marchetti, Federico Becattini, Lorenzo Seidenari, and Alberto Del Bimbo. *Mantra: Memory augmented networks for multiple trajectory prediction*. In *Proceedings of the IEEE/CVF conference on computer vision and pattern recognition*, pages 7143–7152, 2020. 2, 6 +[29] Mancheng Meng, Ziyan Wu, Terrence Chen, Xiran Cai, Xiang Zhou, Fan Yang, and Dinggang Shen. Forecasting human trajectory from scene history. arXiv preprint arXiv:2210.08732, 2022. 1, 6 +[30] Alessio Monti, Alessia Bertugli, Simone Calderara, and Rita Cucchiara. Dag-net: Double attentive graph neural network for trajectory forecasting. In 2020 25th international conference on pattern recognition (ICPR), pages 2551-2558. IEEE, 2021. 3 +[31] Brendan Tran Morris and Mohan Manubhai Trivedi. Trajectory learning for activity understanding: Unsupervised, multilevel, and long-term adaptive approach. IEEE transactions on pattern analysis and machine intelligence, 33(11):2287-2301, 2011. 2 +[32] Stefano Pellegrini, Andreas Ess, Konrad Schindler, and Luc Van Gool. You'll never walk alone: Modeling social behavior for multi-target tracking. In 2009 IEEE 12th international conference on computer vision, pages 261-268. IEEE, 2009. 5 +[33] Alexandre Robicquet, Amir Sadeghian, Alexandre Alahi, and Silvio Savarese. Learning social etiquette: Human trajectory prediction in crowded scenes. In European Conference on Computer Vision (ECCV), page 5, 2016. 5 +[34] Tim Salzmann, Boris Ivanovic, Punarjay Chakravarty, and Marco Pavone. Trajectory: Dynamically-feasible trajectory forecasting with heterogeneous data. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XVIII 16, pages 683-700. Springer, 2020. 2 + +[35] Liushuai Shi, Le Wang, Chengjiang Long, Sanping Zhou, Mo Zhou, Zhenxing Niu, and Gang Hua. Sgcn: Sparse graph convolution network for pedestrian trajectory prediction. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8994-9003, 2021. 3 +[36] Xiao Song, Kai Chen, Xu Li, Jinghan Sun, Baocun Hou, Yong Cui, Baochang Zhang, Gang Xiong, and Zilie Wang. Pedestrian trajectory prediction based on deep convolutional LSTM network. IEEE Transactions on Intelligent Transportation Systems, 22(6):3285-3302, 2020. 2, 6 +[37] Yuchao Su, Jie Du, Yuanman Li, Xia Li, Rongqin Liang, Zhongyun Hua, and Jiantao Zhou. Trajectory forecasting based on prior-aware directed graph convolutional neural network. IEEE Transactions on Intelligent Transportation Systems, 23(9):16773-16785, 2022. 3 +[38] Chuhua Wang, Yuchen Wang, Mingze Xu, and David J Crandall. Stepwise goal-driven networks for trajectory prediction. IEEE Robotics and Automation Letters, 7(2):2716-2723, 2022. 2, 6 +[39] Conghao Wong, Beihao Xia, Ziming Hong, Qinmu Peng, Wei Yuan, Qiong Cao, Yibo Yang, and Xinge You. View vertically: A hierarchical network for trajectory prediction via fourier spectrums. In European Conference on Computer Vision, pages 682-700. Springer, 2022. 2, 6 +[40] Conghao Wong, Beihao Xia, Qinmu Peng, and Xinge You. Another vertical view: A hierarchical network for heterogeneous trajectory prediction via spectrums. arXiv preprint arXiv:2304.05106, 2023. 2, 6 +[41] Conghao Wong, Beihao Xia, Ziqian Zou, Yulong Wang, and Xinge You. Socialcircle: Learning the angle-based social interaction representation for pedestrian trajectory prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19005-19015, 2024. 1, 2, 6 +[42] Beihao Xia, Conghao Wong, Qinmu Peng, Wei Yuan, and Xinge You. Cscnet: Contextual semantic consistency network for trajectory prediction in crowded spaces. Pattern Recognition, 126:108552, 2022. 3 +[43] Song Xiao, Kai Chen, Xiaoxiang Ren, and Haitao Yuan. Pedestrian trajectory prediction in heterogeneous traffic using facial keypoints-based convolutional encoder-decoder network. ACM Transactions on Internet Technology, 22(4): 1-14, 2022. 2 +[44] Chenxin Xu, Maosen Li, Zhenyang Ni, Ya Zhang, and Siheng Chen. Groupnet: Multiscale hypergraph neural networks for trajectory prediction with relational reasoning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6498-6507, 2022. 3, 6 +[45] Chenxin Xu, Weibo Mao, Wenjun Zhang, and Siheng Chen. Remember intentions: Retrospective-memory-based trajectory prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6488-6497, 2022. 1, 6 +[46] Chenxin Xu, Robby T Tan, Yuhong Tan, Siheng Chen, Yu Guang Wang, Xinchao Wang, and Yanfeng Wang. Eqmotion: Equivariant multi-agent motion prediction with invari + +ant interaction reasoning. arXiv preprint arXiv:2303.10876, 2023. 2, 6 +[47] Yu Yao, Ella Atkins, Matthew Johnson-Roberson, Ram Vasudevan, and Xiaoxiao Du. Bitrap: Bi-directional pedestrian trajectory prediction with multi-modal goal estimation. IEEE Robotics and Automation Letters, 6(2):1463-1470, 2021. 2, 6 +[48] Ziyi Yin, Ruijin Liu, Zhiliang Xiong, and Zejian Yuan. Multimodal transformer networks for pedestrian trajectory prediction. In Proceedings of the Twenty-Ninth International Joint Conference on Artificial Intelligence, pages 1259-1265, 2021. 2, 3, 6 +[49] Cunjun Yu, Xiao Ma, Jiawei Ren, Haiyu Zhao, and Shuai Yi. Spatio-temporal graph transformer networks for pedestrian trajectory prediction. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XII 16, pages 507-523, 2020. 2, 3 +[50] Wenjian Yu and Dirk Helbing. Game theoretical interactions of moving agents. In Simulating Complex Systems by Cellular Automata, pages 219-239. Springer, 2010. 2 +[51] Ye Yuan, Xinshuo Weng, Yanglan Ou, and Kris M Kitani. Agentformer: Agent-aware transformers for socio-temporal multi-agent forecasting. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9813-9823, 2021. 2, 6 +[52] Jiangbei Yue, Dinesh Manocha, and He Wang. Human trajectory prediction via neural social physics. In European conference on computer vision, pages 376-394, 2022. 3, 6 +[53] Yisong Yue, Patrick Lucey, Peter Carr, Alina Bialkowski, and Iain Matthews. Learning fine-grained spatial models for dynamic sports play prediction. In 2014 IEEE international conference on data mining, pages 670-679. IEEE, 2014. 5 +[54] Pu Zhang, Wanli Ouyang, Pengfei Zhang, Jianru Xue, and Nanning Zheng. Sr-lstm: State refinement for LSTM towards pedestrian trajectory prediction. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12085-12094, 2019. 2 +[55] Xiaoping Zheng and Yuan Cheng. Conflict game in evacuation process: A study combining cellular automata model. Physica A: Statistical Mechanics and its Applications, 390 (6):1042-1050, 2011. 2 \ No newline at end of file diff --git a/2025/SocialMOIF_ Multi-Order Intention Fusion for Pedestrian Trajectory Prediction/images.zip b/2025/SocialMOIF_ Multi-Order Intention Fusion for Pedestrian Trajectory Prediction/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..f5a62ca867699915c6b0a6b119dbb99ccf75df38 --- /dev/null +++ b/2025/SocialMOIF_ Multi-Order Intention Fusion for Pedestrian Trajectory Prediction/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4f289982596d59dd69c8b80bbb2e54f187fcd521145777bb2d8b8cf252d34486 +size 721285 diff --git a/2025/SocialMOIF_ Multi-Order Intention Fusion for Pedestrian Trajectory Prediction/layout.json b/2025/SocialMOIF_ Multi-Order Intention Fusion for Pedestrian Trajectory Prediction/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..4f7426da885c709985727ea4079b9e277187592b --- /dev/null +++ b/2025/SocialMOIF_ Multi-Order Intention Fusion for Pedestrian Trajectory Prediction/layout.json @@ -0,0 +1,9161 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 56, + 103, + 553, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 103, + 553, + 121 + ], + "spans": [ + { + "bbox": [ + 56, + 103, + 553, + 121 + ], + "type": "text", + "content": "SocialMOIF: Multi-Order Intention Fusion for Pedestrian Trajectory Prediction" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 149, + 142, + 459, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 142, + 459, + 171 + ], + "spans": [ + { + "bbox": [ + 149, + 142, + 459, + 171 + ], + "type": "text", + "content": "Kai Chen" + }, + { + "bbox": [ + 149, + 142, + 459, + 171 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 149, + 142, + 459, + 171 + ], + "type": "text", + "content": ", Xiaodong Zhao" + }, + { + "bbox": [ + 149, + 142, + 459, + 171 + ], + "type": "inline_equation", + "content": "(\\boxtimes)" + }, + { + "bbox": [ + 149, + 142, + 459, + 171 + ], + "type": "text", + "content": "^{1*}, Yujie Huang" + }, + { + "bbox": [ + 149, + 142, + 459, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 149, + 142, + 459, + 171 + ], + "type": "text", + "content": ", Guoyu Fang" + }, + { + "bbox": [ + 149, + 142, + 459, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 149, + 142, + 459, + 171 + ], + "type": "text", + "content": ", Xiao Song" + }, + { + "bbox": [ + 149, + 142, + 459, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 149, + 142, + 459, + 171 + ], + "type": "text", + "content": ", Ruiping Wang" + }, + { + "bbox": [ + 149, + 142, + 459, + 171 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 149, + 142, + 459, + 171 + ], + "type": "text", + "content": ", Ziyuan Wang" + }, + { + "bbox": [ + 149, + 142, + 459, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 175, + 171, + 436, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 175, + 171, + 436, + 185 + ], + "spans": [ + { + "bbox": [ + 175, + 171, + 436, + 185 + ], + "type": "text", + "content": "1 Nanjing University of Aeronautics and Astronautics" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 164, + 185, + 446, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 164, + 185, + 446, + 200 + ], + "spans": [ + { + "bbox": [ + 164, + 185, + 446, + 200 + ], + "type": "text", + "content": "2 Beihang University 3 Nanyang Technological University" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 154, + 201, + 456, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 154, + 201, + 456, + 213 + ], + "spans": [ + { + "bbox": [ + 154, + 201, + 456, + 213 + ], + "type": "text", + "content": "{chen_kai, xdzhao, yj_huang, fanggy, zy_wang}@nuaa.edu.cn" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 181, + 216, + 427, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 181, + 216, + 427, + 227 + ], + "spans": [ + { + "bbox": [ + 181, + 216, + 427, + 227 + ], + "type": "text", + "content": "songxiao@buaa.edu.cn, ruiping.wang@ntu.edu.sg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 152, + 266, + 200, + 280 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 266, + 200, + 280 + ], + "spans": [ + { + "bbox": [ + 152, + 266, + 200, + 280 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 54, + 292, + 296, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 292, + 296, + 602 + ], + "spans": [ + { + "bbox": [ + 54, + 292, + 296, + 602 + ], + "type": "text", + "content": "The analysis and prediction of agent trajectories are crucial for decision-making processes in intelligent systems, with precise short-term trajectory forecasting being highly significant across a range of applications. Agents and their social interactions have been quantified and modeled by researchers from various perspectives; however, substantial limitations exist in the current work due to the inherent high uncertainty of agent intentions and the complex higher-order influences among neighboring groups. SocialMOIF is proposed to tackle these challenges, concentrating on the higher-order intention interactions among neighboring groups while reinforcing the primary role of first-order intention interactions between neighbors and the target agent. This method develops a multi-order intention fusion model to achieve a more comprehensive understanding of both direct and indirect intention information. Within SocialMOIF, a trajectory distribution approximator is designed to guide the trajectories toward values that align more closely with the actual data, thereby enhancing model interpretability. Furthermore, a global trajectory optimizer is introduced to enable more accurate and efficient parallel predictions. By incorporating a novel loss function that accounts for distance and direction during training, experimental results demonstrate that the model outperforms previous state-of-the-art baselines across multiple metrics in both dynamic and static datasets." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 627, + 135, + 639 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 627, + 135, + 639 + ], + "spans": [ + { + "bbox": [ + 55, + 627, + 135, + 639 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 647, + 296, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 647, + 296, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 647, + 296, + 685 + ], + "type": "text", + "content": "Trajectory prediction is of considerable importance for intelligent transportation and surveillance systems. Studies indicate that trajectories derived from the physical" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 320, + 257, + 549, + 453 + ], + "blocks": [ + { + "bbox": [ + 320, + 257, + 549, + 453 + ], + "lines": [ + { + "bbox": [ + 320, + 257, + 549, + 453 + ], + "spans": [ + { + "bbox": [ + 320, + 257, + 549, + 453 + ], + "type": "image", + "image_path": "a5b16720cb4c521b66b0bc3423abd877df5fe2e2275fcd0e4f0bd6c9b45791e2.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 464, + 555, + 520 + ], + "lines": [ + { + "bbox": [ + 313, + 464, + 555, + 520 + ], + "spans": [ + { + "bbox": [ + 313, + 464, + 555, + 520 + ], + "type": "text", + "content": "Figure 1. The process of SocialMOIF includes: (1) extracting the position, velocity and trajectory information of agents, (2) achieving multi-order fused intention modeling, (3) guiding the distribution of trajectories and performing global optimization. Finally we obtain accurate parallel predictions." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 544, + 555, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 544, + 555, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 544, + 555, + 628 + ], + "type": "text", + "content": "world encompass rich spatiotemporal intention information [15, 20, 26, 29, 45]. Therefore, a significant challenge arises in effectively capturing (data modeling) and analyzing (encoding-decoding) processes related to the complex and diverse historical social intentions of agents, while linking this information to future trajectory predictions for enhanced efficiency." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "content": "History social intentions. Existing research can be broadly classified into two main categories: knowledge-driven methods primarily depend on expert knowledge. For instance, biomimetic method [41] achieved agent localization and intention interaction modeling by simulating the positioning of marine animals. In contrast, certain studies utilized game-theoretic approaches to analyze" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 69, + 693, + 140, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 693, + 140, + 703 + ], + "spans": [ + { + "bbox": [ + 69, + 693, + 140, + 703 + ], + "type": "text", + "content": "* Equal contribution." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 69, + 703, + 287, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 703, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 69, + 703, + 287, + 713 + ], + "type": "text", + "content": "Code is available at https://github.com/XiaodZhao/SocialMOIF." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22465" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 358 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 358 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 358 + ], + "type": "text", + "content": "social intentions through the simulation of agent flows [13, 50] and evacuation processes [14, 55], relying on explicit rules or formulas. However, the behavioral patterns and interaction rules among agents in social scenarios displayed high nonlinearity and uncertainty [26, 39, 40]. Data-driven methods implicitly learn interactions among agents through neural networks. For instance, LSTM-based methods [1, 11] continuously refined social pooling patterns, while attention-based approaches [7, 49, 51] dynamically adjusted agents' focal points in global interactions from a spatiotemporal integration perspective to optimize predicted trajectories. These methods primarily focused on straightforward interactions or direct responses between the target agent and its neighbors, often neglecting the higher-order, multidimensional influences exerted by surrounding agents on the target agent's intentions. Kim et al. [16] modeled agent intentions from 1st-order to N-order equally. However, when the number of agents was large, it may prioritized higher-order intentions, thereby diminishing the primary impact of lower-order intentions. Thus, the development of a multi-order fusion network that reinforces the dominant role of direct interactions while considering the secondary influence of higher-order transmissions on the target is essential." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 362, + 295, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 362, + 295, + 517 + ], + "spans": [ + { + "bbox": [ + 55, + 362, + 295, + 517 + ], + "type": "text", + "content": "Associating future trajectory information. Knowledge-driven methods face limitations due to the inherent flaws of unidirectional inference, which hinder effective association with future trajectory information. Data-driven methods exhibit two main shortcomings: first, agent movements are temporally dynamic, and relying solely on distance loss fails to capture these dynamic variations. Second, in generative model approaches [25, 28, 34, 38, 47], future trajectory information is implicitly involved in the training process without explicit guidance, resulting in poor interpretability. In safety-sensitive tasks such as autonomous driving, it is necessary to utilize meaningful and goal-oriented parameters and models." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 520, + 294, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 520, + 294, + 651 + ], + "spans": [ + { + "bbox": [ + 55, + 520, + 294, + 651 + ], + "type": "text", + "content": "Efficiency. Existing methods predominantly emphasize upstream tasks. For example, LSTM-based approaches [5, 6, 36, 43, 54] modeled trajectory data in a serial temporal manner, while Transformer-based methods [9, 19, 48, 49] analyzed trajectory information through parallel processing. However, in downstream tasks, trajectories are often generated independently and sequentially. This method not only accumulates errors but also limits the efficiency of trajectory prediction. Therefore, establishing a trajectory parallel prediction model that accounts for global influences across future periods is a promising area for further exploration." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 653, + 294, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 294, + 713 + ], + "type": "text", + "content": "As illustrated in Fig. 1, SocialMOIF is proposed to address the trajectory prediction task. A Multi-Order Intention Fusion model (MOIF) is introduced to guide the learning direction of the network. MOIF includes a first-order intention interaction layer and a higher-order intention in" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 553, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 287 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 287 + ], + "type": "text", + "content": "teraction layer. The first layer focuses on direct interaction information between neighbors and the target agent, while the second layer addresses the indirect interaction information conveyed to the target agent after multiple interactions among neighbor targets. Inspiration is drawn from the squeeze theorem to design a trajectory distribution approximator to guide trajectories toward a distribution that closely aligns with ground truth and to enhance model interpretability. Simultaneously, RNN is incorporated to update the latent variables, achieving explicit supervision during training. The temporal dimension of decoded trajectories is aggregated to design a global trajectory optimizer. To the best of our knowledge, KANs [23, 24] are firstly introduced for the global optimization of trajectories, enabling more accurate and efficient parallel predictions. Furthermore, considering the travel directions of the agents, a distance-direction fused loss function is designed to comprehensively supervise the dynamic states of agents." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 326, + 289, + 541, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 289, + 541, + 300 + ], + "spans": [ + { + "bbox": [ + 326, + 289, + 541, + 300 + ], + "type": "text", + "content": "In summary, the main contributions of this paper are:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 318, + 300, + 553, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 300, + 553, + 323 + ], + "spans": [ + { + "bbox": [ + 318, + 300, + 553, + 323 + ], + "type": "text", + "content": "- The proposal of SocialMOIF, which incorporates four novel designs:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 329, + 323, + 553, + 358 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 329, + 323, + 553, + 358 + ], + "spans": [ + { + "bbox": [ + 329, + 323, + 553, + 358 + ], + "type": "text", + "content": "(1) MOIF is introduced to comprehensively capture interaction information between the target agent and its neighbors;" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 329, + 358, + 553, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 329, + 358, + 553, + 407 + ], + "spans": [ + { + "bbox": [ + 329, + 358, + 553, + 407 + ], + "type": "text", + "content": "(2) A trajectory distribution approximator is developed to approximate and explicitly update the distribution of latent variables, guiding the distribution of trajectories;" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 329, + 407, + 553, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 329, + 407, + 553, + 455 + ], + "spans": [ + { + "bbox": [ + 329, + 407, + 553, + 455 + ], + "type": "text", + "content": "(3) A global trajectory optimizer is designed, and KANs are introduced for the first time to achieve parallel trajectory optimization across the global spatiotemporal domain;" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 329, + 455, + 553, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 329, + 455, + 553, + 491 + ], + "spans": [ + { + "bbox": [ + 329, + 455, + 553, + 491 + ], + "type": "text", + "content": "(4) A new direction-distance fused loss function is proposed for more effective supervision of agents' dynamic changes." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 318, + 491, + 553, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 491, + 553, + 526 + ], + "spans": [ + { + "bbox": [ + 318, + 491, + 553, + 526 + ], + "type": "text", + "content": "- Comprehensive comparative experiments are conducted with existing high-quality models, demonstrating the effective and robust predictive capabilities of SocialMOIF." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 538, + 403, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 538, + 403, + 550 + ], + "spans": [ + { + "bbox": [ + 314, + 538, + 403, + 550 + ], + "type": "text", + "content": "2. Related works" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 558, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 553, + 713 + ], + "type": "text", + "content": "Knowledge-driven methods. Helbing et al. [12] first proposed the social force model, which describes pedestrian interactions through an energy potential field. Best et al. [3] introduced a Bayesian inference-based intention inference model, integrating observed movement trajectories with prior knowledge. EqMotion [46] employed geometric analysis to convert human-human interaction trajectory analysis into an optimization problem by examining object geometry. Morris et al. [31] proposed an implicit Markov model for trajectory prediction, which is utilized for spatiotemporal probabilistic modeling of various types of pedestrian trajectories. A recent approach, socialcircle [41], incorporated angle-based light rule constraints into the trainable back" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "22466" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 286, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 286, + 84 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 286, + 84 + ], + "type": "text", + "content": "bone to achieve interaction modeling of agent intentions." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 84, + 296, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 84, + 296, + 179 + ], + "spans": [ + { + "bbox": [ + 55, + 84, + 296, + 179 + ], + "type": "text", + "content": "Despite these advancements, these methods rely on knowledge from specific domains and therefore struggle to fully capture the complex agent intentions within social scenarios. Although attempts have been made to incorporate future trajectory information to overcome the limitations of unidirectional inference, the model's equations remain overly sensitive to common variations, leading to persistent challenges in trajectory prediction tasks [42, 52]." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 180, + 296, + 359 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 180, + 296, + 359 + ], + "spans": [ + { + "bbox": [ + 55, + 180, + 296, + 359 + ], + "type": "text", + "content": "Data-driven methods. LSTM-based methods model trajectory data temporally. Alahi et al. [1] employed a \"social pooling mechanism\" that considers only the intentions of agents within a grid while neglecting interactions with boundary agents. Gupta et al. [11] improved upon this by implementing a pooling mechanism that treats pedestrians in neighboring areas equally. Transformer-based methods enhanced temporal modeling approaches. Yu et al. [49] modeled and predicted trajectories using a spatio-temporal concatenation approach. Yin et al. [48] introduced optical flow to simulate agents' motion intentions. Graph neural networks were also utilized for trajectory prediction to model intention interactions among different nodes [8, 22, 30, 35, 37, 44]. Diffusion models [10, 27] were used to explore trajectory prediction based on randomness." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 360, + 296, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 360, + 296, + 467 + ], + "spans": [ + { + "bbox": [ + 55, + 360, + 296, + 467 + ], + "type": "text", + "content": "However, most of these methods primarily emphasize direct interactions between the target agent and its neighbors, often overlooking the higher-order influences that neighbors exert on the target agent. Furthermore, many of these approaches incorporate future trajectory information only at the network's output stage, supervising agents' trajectory changes solely through distance loss. Moreover, the limited interpretability of latent variable distributions in generative models can impede further development of these models." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 467, + 296, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 467, + 296, + 575 + ], + "spans": [ + { + "bbox": [ + 55, + 467, + 296, + 575 + ], + "type": "text", + "content": "In addition, the aforementioned methods mainly emphasize trajectory analysis and modeling of upstream tasks. For output in downstream tasks, results are generally produced independently in a sequential manner over time. This sequential processing introduces delays that may result in untimely responses, potentially leading to safety concerns. Furthermore, each future prediction is dependent on the state from the preceding time step, which may lead to accumulated errors, adversely affecting prediction accuracy." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 586, + 122, + 600 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 586, + 122, + 600 + ], + "spans": [ + { + "bbox": [ + 55, + 586, + 122, + 600 + ], + "type": "text", + "content": "3. Approach" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "text", + "content": "The number of agents in the scenario is set to " + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "inline_equation", + "content": "(N_{n} + 1)" + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "text", + "content": ". The trajectory of the target agent " + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "text", + "content": " in history period " + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "inline_equation", + "content": "T_{H}" + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "inline_equation", + "content": "g_{i}^{1:T_{H}}" + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "inline_equation", + "content": "g_{i}^{t}" + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "text", + "content": " is the two-dimensional spatial coordinate at moment " + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "text", + "content": ". The remaining agents are tentatively designated as neighbors, and the number is " + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "inline_equation", + "content": "N_{n}" + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "text", + "content": ". The trajectory of " + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "text", + "content": " in the future period to be predicted is denoted as " + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\hat{g}_{i}^{T_{H} + 1:T_{F}}" + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "text", + "content": ". We propose the multi-order intention fusion model in Section 3.1, the trajectory distribution approximator in Section 3.2, the global trajectory optimizer in Section 3.3, and the" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 531, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 531, + 84 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 531, + 84 + ], + "type": "text", + "content": "direction-distance fusion loss function in Section 3.4." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 91, + 501, + 103 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 91, + 501, + 103 + ], + "spans": [ + { + "bbox": [ + 313, + 91, + 501, + 103 + ], + "type": "text", + "content": "3.1. Multi-order intention fusion model" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 109, + 555, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 109, + 555, + 252 + ], + "spans": [ + { + "bbox": [ + 313, + 109, + 555, + 252 + ], + "type": "text", + "content": "The model consists of two layers: a first-order intention interaction layer and a higher-order intention interaction layer. As shown in Fig. 2, the first-order layer emphasizes the direct, dominant influence of neighbors on the target agent's intention. The higher-order layer accounts for the indirect and secondary influences on the target agent's intention that arise from interactions within the group of neighbors, thereby enabling a more comprehensive understanding of intentions. Based on the current and final history motion states of the agents, input components for the layers are first constructed, followed by the extraction and fusion of multi-order intention information." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 253, + 556, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 253, + 556, + 348 + ], + "spans": [ + { + "bbox": [ + 313, + 253, + 556, + 348 + ], + "type": "text", + "content": "Higher-order intention interaction layer. From a more fundamental perspective, the higher-order intention interactions within the group of neighbors can be decomposed into first-order intention interactions between individual neighbors. We define the intention influence of neighbor agent " + }, + { + "bbox": [ + 313, + 253, + 556, + 348 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 313, + 253, + 556, + 348 + ], + "type": "text", + "content": " on neighbor agent " + }, + { + "bbox": [ + 313, + 253, + 556, + 348 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 313, + 253, + 556, + 348 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 313, + 253, + 556, + 348 + ], + "type": "inline_equation", + "content": "w_{a\\rightarrow b}" + }, + { + "bbox": [ + 313, + 253, + 556, + 348 + ], + "type": "text", + "content": "; for an agent itself, it is represented as " + }, + { + "bbox": [ + 313, + 253, + 556, + 348 + ], + "type": "inline_equation", + "content": "w_{a\\rightarrow a}" + }, + { + "bbox": [ + 313, + 253, + 556, + 348 + ], + "type": "text", + "content": ". In a scenario with " + }, + { + "bbox": [ + 313, + 253, + 556, + 348 + ], + "type": "inline_equation", + "content": "N_{n}" + }, + { + "bbox": [ + 313, + 253, + 556, + 348 + ], + "type": "text", + "content": " agents, the total number of such intentions is " + }, + { + "bbox": [ + 313, + 253, + 556, + 348 + ], + "type": "inline_equation", + "content": "\\Omega" + }, + { + "bbox": [ + 313, + 253, + 556, + 348 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 368, + 356, + 555, + 370 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 368, + 356, + 555, + 370 + ], + "spans": [ + { + "bbox": [ + 368, + 356, + 555, + 370 + ], + "type": "interline_equation", + "content": "\\Omega = \\left(N _ {n} - 1\\right) N _ {n} + N _ {n} = N _ {n} ^ {2} \\tag {1}", + "image_path": "24e31e33531daa22aae76d5d74653371b5bf139b52d86c3e4b2decffe58bba98.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 379, + 555, + 428 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 379, + 555, + 428 + ], + "spans": [ + { + "bbox": [ + 313, + 379, + 555, + 428 + ], + "type": "text", + "content": "To obtain the intentions, we extract the absolute positions and velocities of the " + }, + { + "bbox": [ + 313, + 379, + 555, + 428 + ], + "type": "inline_equation", + "content": "N_{n}" + }, + { + "bbox": [ + 313, + 379, + 555, + 428 + ], + "type": "text", + "content": " neighbors corresponding to the target agent " + }, + { + "bbox": [ + 313, + 379, + 555, + 428 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 379, + 555, + 428 + ], + "type": "text", + "content": ". They are processed as a layer input component through the embedding layer " + }, + { + "bbox": [ + 313, + 379, + 555, + 428 + ], + "type": "inline_equation", + "content": "\\pmb{f}_{S}" + }, + { + "bbox": [ + 313, + 379, + 555, + 428 + ], + "type": "text", + "content": " to obtain " + }, + { + "bbox": [ + 313, + 379, + 555, + 428 + ], + "type": "inline_equation", + "content": "U_{i}" + }, + { + "bbox": [ + 313, + 379, + 555, + 428 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 349, + 434, + 555, + 455 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 349, + 434, + 555, + 455 + ], + "spans": [ + { + "bbox": [ + 349, + 434, + 555, + 455 + ], + "type": "interline_equation", + "content": "U _ {i} = \\boldsymbol {f} _ {U} \\left(x _ {1: N _ {n}} ^ {1: T _ {H}}, y _ {1: N _ {n}} ^ {1: T _ {H}}, v _ {x 1: N _ {n}} ^ {1: T _ {H}}, v _ {y 1: N _ {n}} ^ {1: T _ {H}}\\right) \\tag {2}", + "image_path": "20d62e24adc33da564866e3e42c024512e4fa8f7791deca812fca527325aad35.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 462, + 554, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 462, + 554, + 534 + ], + "spans": [ + { + "bbox": [ + 313, + 462, + 554, + 534 + ], + "type": "text", + "content": "We capture the intention interactions of neighbors through multiple subspaces in parallel. " + }, + { + "bbox": [ + 313, + 462, + 554, + 534 + ], + "type": "inline_equation", + "content": "U_{i}" + }, + { + "bbox": [ + 313, + 462, + 554, + 534 + ], + "type": "text", + "content": " is mapped into a learnable query vector " + }, + { + "bbox": [ + 313, + 462, + 554, + 534 + ], + "type": "inline_equation", + "content": "Q_{U}" + }, + { + "bbox": [ + 313, + 462, + 554, + 534 + ], + "type": "text", + "content": ", a key vector " + }, + { + "bbox": [ + 313, + 462, + 554, + 534 + ], + "type": "inline_equation", + "content": "K_{U}" + }, + { + "bbox": [ + 313, + 462, + 554, + 534 + ], + "type": "text", + "content": ", and a value vector " + }, + { + "bbox": [ + 313, + 462, + 554, + 534 + ], + "type": "inline_equation", + "content": "V_{U}" + }, + { + "bbox": [ + 313, + 462, + 554, + 534 + ], + "type": "text", + "content": ". The feature dimensions are flattened into M subspaces. For the m-th subspace, the overall intention coefficient matrix within neighbors can be expressed as:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 364, + 541, + 555, + 574 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 364, + 541, + 555, + 574 + ], + "spans": [ + { + "bbox": [ + 364, + 541, + 555, + 574 + ], + "type": "interline_equation", + "content": "W _ {U} ^ {m} = \\operatorname {S o f t M a x} \\left(\\frac {Q _ {U} ^ {m} \\left(K _ {U} ^ {m}\\right) ^ {T}}{d _ {M}}\\right) \\tag {3}", + "image_path": "cd926f64054317ef927d537f21640935bf95e9f536431ce951a92df632cb7aba.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 582, + 554, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 554, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 554, + 605 + ], + "type": "text", + "content": "Thus, we obtain " + }, + { + "bbox": [ + 313, + 582, + 554, + 605 + ], + "type": "inline_equation", + "content": "w_{a\\to b} = [W_U^m ]_{ab}" + }, + { + "bbox": [ + 313, + 582, + 554, + 605 + ], + "type": "text", + "content": ". For this matrix, the number of elements is " + }, + { + "bbox": [ + 313, + 582, + 554, + 605 + ], + "type": "inline_equation", + "content": "\\Omega" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 606, + 554, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 606, + 554, + 643 + ], + "spans": [ + { + "bbox": [ + 313, + 606, + 554, + 643 + ], + "type": "text", + "content": "The first-order intention interaction layer. The layer input component of the target agent " + }, + { + "bbox": [ + 313, + 606, + 554, + 643 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 606, + 554, + 643 + ], + "type": "text", + "content": " is processed through the embedding layer " + }, + { + "bbox": [ + 313, + 606, + 554, + 643 + ], + "type": "inline_equation", + "content": "\\pmb{f}_{S}" + }, + { + "bbox": [ + 313, + 606, + 554, + 643 + ], + "type": "text", + "content": " to obtain " + }, + { + "bbox": [ + 313, + 606, + 554, + 643 + ], + "type": "inline_equation", + "content": "S_{i}" + }, + { + "bbox": [ + 313, + 606, + 554, + 643 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 319, + 650, + 555, + 671 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 650, + 555, + 671 + ], + "spans": [ + { + "bbox": [ + 319, + 650, + 555, + 671 + ], + "type": "interline_equation", + "content": "S _ {i} = \\pmb {f} _ {S} \\left(x _ {i} ^ {1: T _ {H}}, y _ {i} ^ {1: T _ {H}}, d _ {i \\leftrightarrow 1: N _ {n}} ^ {1: T _ {H}}, \\theta_ {i \\leftrightarrow 1: N _ {n}} ^ {1: T _ {H}}, e _ {i \\leftrightarrow 1: N _ {n}} ^ {1: T _ {H}}\\right) (4)", + "image_path": "ff26d5f9071a80ec4446f300870212416d27b37f4ad743faa584b431d858c165.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "text", + "content": "We intensify the dominant role of direct interaction in components and focus on the design of interaction rules. For the present case, " + }, + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "inline_equation", + "content": "d_{i\\leftrightarrow j}^{1:T_H}" + }, + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "text", + "content": " is the distance between the target" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22467" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 70, + 553, + 339 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 553, + 339 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 553, + 339 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 553, + 339 + ], + "type": "image", + "image_path": "ee95a298984c868f3ddb918b355f95a29b441c8107ebc48cb0487ce9d0e4ca5a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 347, + 555, + 404 + ], + "lines": [ + { + "bbox": [ + 54, + 347, + 555, + 404 + ], + "spans": [ + { + "bbox": [ + 54, + 347, + 555, + 404 + ], + "type": "text", + "content": "Figure 2. Computational pipeline of SocialMOIF. Firstly, the first-order interaction layer focuses on direct interactions between the target agent and its neighbors. The higher-order interaction layer addresses the interactions among neighbors. Multi-order intention is obtained by fusing the information from both layers. This multi-order intention is then combined with the actual future trajectory, with the distribution of the trajectory guided by an approximator. An optimizer is subsequently employed to enhance the decoded trajectory. Additionally, a fused distance-direction loss function supervises the training process. Finally, efficient parallel prediction is achieved." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 422, + 296, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 422, + 296, + 491 + ], + "spans": [ + { + "bbox": [ + 55, + 422, + 296, + 491 + ], + "type": "text", + "content": "agent " + }, + { + "bbox": [ + 55, + 422, + 296, + 491 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 422, + 296, + 491 + ], + "type": "text", + "content": " and neighbor " + }, + { + "bbox": [ + 55, + 422, + 296, + 491 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 55, + 422, + 296, + 491 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 422, + 296, + 491 + ], + "type": "inline_equation", + "content": "\\theta_{i\\leftrightarrow j}^{1:T_H}" + }, + { + "bbox": [ + 55, + 422, + 296, + 491 + ], + "type": "text", + "content": " is the velocity angle. For the final history, " + }, + { + "bbox": [ + 55, + 422, + 296, + 491 + ], + "type": "inline_equation", + "content": "e_{i\\leftrightarrow j}^{1:T_H}" + }, + { + "bbox": [ + 55, + 422, + 296, + 491 + ], + "type": "text", + "content": " is the final distance in the history period, assuming that after moment " + }, + { + "bbox": [ + 55, + 422, + 296, + 491 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 422, + 296, + 491 + ], + "type": "text", + "content": ", both the target agent " + }, + { + "bbox": [ + 55, + 422, + 296, + 491 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 422, + 296, + 491 + ], + "type": "text", + "content": " and neighbor " + }, + { + "bbox": [ + 55, + 422, + 296, + 491 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 55, + 422, + 296, + 491 + ], + "type": "text", + "content": " maintain the original velocity in the tangent direction " + }, + { + "bbox": [ + 55, + 422, + 296, + 491 + ], + "type": "inline_equation", + "content": "\\overrightarrow{vx_i^t},\\overrightarrow{vx_j^t}" + }, + { + "bbox": [ + 55, + 422, + 296, + 491 + ], + "type": "text", + "content": "; then, " + }, + { + "bbox": [ + 55, + 422, + 296, + 491 + ], + "type": "inline_equation", + "content": "e_{ij}^t" + }, + { + "bbox": [ + 55, + 422, + 296, + 491 + ], + "type": "text", + "content": " can be expressed as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 121, + 498, + 295, + 526 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 498, + 295, + 526 + ], + "spans": [ + { + "bbox": [ + 121, + 498, + 295, + 526 + ], + "type": "interline_equation", + "content": "\\lambda = \\left| \\left(\\overrightarrow {d _ {i j} ^ {t}} \\cdot \\overrightarrow {v _ {i j} ^ {t}}\\right) / \\left\\| \\overrightarrow {v _ {i j} ^ {t}} \\right\\| ^ {2} \\right| \\tag {5}", + "image_path": "661cf1d2d3c7953226b7cd1ecc07d45f8bf9f8beea819338c3136f1d7f44d1b3.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 80, + 533, + 295, + 556 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 533, + 295, + 556 + ], + "spans": [ + { + "bbox": [ + 80, + 533, + 295, + 556 + ], + "type": "interline_equation", + "content": "e _ {i j} ^ {t} = \\left\\| \\overrightarrow {d _ {i j} ^ {t}} + \\min (\\lambda , T _ {H} - t) \\cdot \\left(\\overrightarrow {v x _ {j} ^ {t}} - \\overrightarrow {v y _ {i} ^ {t}}\\right) \\right\\| \\tag {6}", + "image_path": "71f6237abee8545d5cd8f49ec544c2a9d8a0e1c4e9f3f5b9f0a1a96b7bada45b.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 559, + 296, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 559, + 296, + 606 + ], + "spans": [ + { + "bbox": [ + 55, + 559, + 296, + 606 + ], + "type": "text", + "content": "Similarly, " + }, + { + "bbox": [ + 55, + 559, + 296, + 606 + ], + "type": "inline_equation", + "content": "S_{i}" + }, + { + "bbox": [ + 55, + 559, + 296, + 606 + ], + "type": "text", + "content": " is mapped into learnable query vectors " + }, + { + "bbox": [ + 55, + 559, + 296, + 606 + ], + "type": "inline_equation", + "content": "Q_{S}" + }, + { + "bbox": [ + 55, + 559, + 296, + 606 + ], + "type": "text", + "content": ", key vectors " + }, + { + "bbox": [ + 55, + 559, + 296, + 606 + ], + "type": "inline_equation", + "content": "K_{S}" + }, + { + "bbox": [ + 55, + 559, + 296, + 606 + ], + "type": "text", + "content": ", and value vectors " + }, + { + "bbox": [ + 55, + 559, + 296, + 606 + ], + "type": "inline_equation", + "content": "V_{S}" + }, + { + "bbox": [ + 55, + 559, + 296, + 606 + ], + "type": "text", + "content": ". The matrix of the intention coefficients between the target agent " + }, + { + "bbox": [ + 55, + 559, + 296, + 606 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 559, + 296, + 606 + ], + "type": "text", + "content": " and its neighbors is as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 109, + 613, + 295, + 647 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 109, + 613, + 295, + 647 + ], + "spans": [ + { + "bbox": [ + 109, + 613, + 295, + 647 + ], + "type": "interline_equation", + "content": "W _ {S} = \\operatorname {S o f t M a x} \\left(\\frac {Q _ {S} \\left(K _ {S}\\right) ^ {T}}{d _ {S}}\\right) \\tag {7}", + "image_path": "aabebaebeeb128bbba4e653a46a6b7ec09512468dbbbc94b1a86ac9b72908482.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "content": "Intention Fusion. Higher-order interaction intentions are conducted to the target agent and combined with first-order interaction intentions to achieve multi-order intention fusion. During this fusion, the higher-order interaction intentions captured within each subspace exert varying effects" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 423, + 555, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 423, + 555, + 460 + ], + "spans": [ + { + "bbox": [ + 313, + 423, + 555, + 460 + ], + "type": "text", + "content": "when applied to the target agent. We set the corresponding trainable influence factor " + }, + { + "bbox": [ + 313, + 423, + 555, + 460 + ], + "type": "inline_equation", + "content": "\\eta_{m}" + }, + { + "bbox": [ + 313, + 423, + 555, + 460 + ], + "type": "text", + "content": " to denote this influence. The attention score " + }, + { + "bbox": [ + 313, + 423, + 555, + 460 + ], + "type": "inline_equation", + "content": "A_{i}" + }, + { + "bbox": [ + 313, + 423, + 555, + 460 + ], + "type": "text", + "content": " is denoted as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 367, + 467, + 555, + 500 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 467, + 555, + 500 + ], + "spans": [ + { + "bbox": [ + 367, + 467, + 555, + 500 + ], + "type": "interline_equation", + "content": "A _ {i} = \\left(\\sum_ {m = 1} ^ {M} \\eta_ {m} W _ {U} ^ {m} + W _ {S}\\right) V _ {S} \\tag {8}", + "image_path": "73f1893eeaf55f825776f6addda7340fa95b7136db8da30211d6b914b8f6aa0a.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 507, + 555, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 507, + 555, + 543 + ], + "spans": [ + { + "bbox": [ + 313, + 507, + 555, + 543 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 507, + 555, + 543 + ], + "type": "inline_equation", + "content": "A_{i}" + }, + { + "bbox": [ + 313, + 507, + 555, + 543 + ], + "type": "text", + "content": " is processed by two skip connections and a final fully connected layer to obtain the multi-order fused intention " + }, + { + "bbox": [ + 313, + 507, + 555, + 543 + ], + "type": "inline_equation", + "content": "I_{i}" + }, + { + "bbox": [ + 313, + 507, + 555, + 543 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 550, + 512, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 550, + 512, + 563 + ], + "spans": [ + { + "bbox": [ + 313, + 550, + 512, + 563 + ], + "type": "text", + "content": "3.2. Trajectory distribution approximator" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 566, + 555, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 566, + 555, + 652 + ], + "spans": [ + { + "bbox": [ + 313, + 566, + 555, + 652 + ], + "type": "text", + "content": "We define the trajectory " + }, + { + "bbox": [ + 313, + 566, + 555, + 652 + ], + "type": "inline_equation", + "content": "g_{igt}^{T_H + 1:T_F}" + }, + { + "bbox": [ + 313, + 566, + 555, + 652 + ], + "type": "text", + "content": " of the target agent " + }, + { + "bbox": [ + 313, + 566, + 555, + 652 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 566, + 555, + 652 + ], + "type": "text", + "content": " in the future period. " + }, + { + "bbox": [ + 313, + 566, + 555, + 652 + ], + "type": "inline_equation", + "content": "B_{i}" + }, + { + "bbox": [ + 313, + 566, + 555, + 652 + ], + "type": "text", + "content": " is obtained after feature enhancement through the embedding layer " + }, + { + "bbox": [ + 313, + 566, + 555, + 652 + ], + "type": "inline_equation", + "content": "f_{B}" + }, + { + "bbox": [ + 313, + 566, + 555, + 652 + ], + "type": "text", + "content": ". By utilizing " + }, + { + "bbox": [ + 313, + 566, + 555, + 652 + ], + "type": "inline_equation", + "content": "I_{i}" + }, + { + "bbox": [ + 313, + 566, + 555, + 652 + ], + "type": "text", + "content": " as the lower bound and " + }, + { + "bbox": [ + 313, + 566, + 555, + 652 + ], + "type": "inline_equation", + "content": "B_{i}" + }, + { + "bbox": [ + 313, + 566, + 555, + 652 + ], + "type": "text", + "content": " as the upper bound, the latent variable distribution in the potential space is explicitly approximated. First, the probability distribution of the target agent " + }, + { + "bbox": [ + 313, + 566, + 555, + 652 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 566, + 555, + 652 + ], + "type": "text", + "content": "'s trajectory is initially modeled as:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 332, + 659, + 553, + 716 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 332, + 659, + 553, + 716 + ], + "spans": [ + { + "bbox": [ + 332, + 659, + 553, + 716 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} p \\left(g _ {i} ^ {T _ {H} + 1: T _ {F}} \\mid U _ {i}, S _ {i}\\right) \\\\ = \\prod_ {t = T _ {H} + 1} ^ {T _ {F}} p \\left(g _ {i} ^ {T _ {H} + 1 + t} \\mid g _ {i} ^ {T _ {H} + 1: T _ {H} + t - 1}, U _ {i}, S _ {i}\\right) \\tag {9} \\\\ \\end{array}", + "image_path": "cb1e339d3a7b9182a869a89581a87bf3b0e30d08f5c8834fbfd95e046bf39a03.jpg" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22468" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "type": "text", + "content": "We sample the noise variables " + }, + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "type": "inline_equation", + "content": "\\varepsilon \\sim N(0,1)" + }, + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "type": "text", + "content": " and reparameterize the latent variables that are approximated at " + }, + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "type": "text", + "content": " [17]:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 102, + 295, + 119 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 102, + 295, + 119 + ], + "spans": [ + { + "bbox": [ + 55, + 102, + 295, + 119 + ], + "type": "interline_equation", + "content": "z _ {i \\varphi} ^ {t} = u _ {i \\varphi} ^ {t} + \\varepsilon \\cdot \\sigma_ {i \\varphi} ^ {t} \\quad q _ {\\varphi} (\\cdot | I _ {i} ^ {t - 1}, B _ {i} ^ {t}) \\sim N (u _ {i \\varphi} ^ {t}, \\sigma_ {i \\varphi} ^ {t}) \\tag {10}", + "image_path": "1a86f9ad7407b0356b0f9c7a2417167dfab26967108e90c9ed6794246797609c.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 124, + 296, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 124, + 296, + 171 + ], + "spans": [ + { + "bbox": [ + 55, + 124, + 296, + 171 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 124, + 296, + 171 + ], + "type": "inline_equation", + "content": "\\varphi" + }, + { + "bbox": [ + 55, + 124, + 296, + 171 + ], + "type": "text", + "content": " are parameters that need to be optimized during the training process. In the test mode, the latent variables " + }, + { + "bbox": [ + 55, + 124, + 296, + 171 + ], + "type": "inline_equation", + "content": "z_{i\\vartheta}^{t}" + }, + { + "bbox": [ + 55, + 124, + 296, + 171 + ], + "type": "text", + "content": " are only reasoned by " + }, + { + "bbox": [ + 55, + 124, + 296, + 171 + ], + "type": "inline_equation", + "content": "I_{i}" + }, + { + "bbox": [ + 55, + 124, + 296, + 171 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 124, + 296, + 171 + ], + "type": "inline_equation", + "content": "\\vartheta" + }, + { + "bbox": [ + 55, + 124, + 296, + 171 + ], + "type": "text", + "content": " is also involved in the training process of the model." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 172, + 296, + 207 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 172, + 296, + 207 + ], + "spans": [ + { + "bbox": [ + 55, + 172, + 296, + 207 + ], + "type": "text", + "content": "We sample and decode the target agent's trajectories on the basis of latent variables and multi-order fused intention information." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 102, + 213, + 295, + 244 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 213, + 295, + 244 + ], + "spans": [ + { + "bbox": [ + 102, + 213, + 295, + 244 + ], + "type": "interline_equation", + "content": "g _ {i} ^ {t} \\sim \\left\\{ \\begin{array}{l l} p _ {\\delta} (\\cdot | z _ {i \\varphi} ^ {t}, I _ {i} ^ {t - 1}) & \\text {i f t r a i n} \\\\ p _ {\\delta} (\\cdot | z _ {i \\vartheta} ^ {t}, I _ {i} ^ {t - 1}) & \\text {i f t e s t} \\end{array} \\right. \\tag {11}", + "image_path": "8116e215905c51e0d567aa19169e85a72c446bc68d34bb7433166ff186f31979.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 250, + 295, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 250, + 295, + 274 + ], + "spans": [ + { + "bbox": [ + 55, + 250, + 295, + 274 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 250, + 295, + 274 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 55, + 250, + 295, + 274 + ], + "type": "text", + "content": " are parameters that need to be optimized during the training process." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 274, + 296, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 274, + 296, + 300 + ], + "spans": [ + { + "bbox": [ + 55, + 274, + 296, + 300 + ], + "type": "text", + "content": "Concat the latent variables with the trajectories generated at the current time " + }, + { + "bbox": [ + 55, + 274, + 296, + 300 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 274, + 296, + 300 + ], + "type": "text", + "content": ", updating " + }, + { + "bbox": [ + 55, + 274, + 296, + 300 + ], + "type": "inline_equation", + "content": "I_{i}^{t - 1}" + }, + { + "bbox": [ + 55, + 274, + 296, + 300 + ], + "type": "text", + "content": " via " + }, + { + "bbox": [ + 55, + 274, + 296, + 300 + ], + "type": "inline_equation", + "content": "f_{zg}" + }, + { + "bbox": [ + 55, + 274, + 296, + 300 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 70, + 307, + 295, + 338 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 307, + 295, + 338 + ], + "spans": [ + { + "bbox": [ + 70, + 307, + 295, + 338 + ], + "type": "interline_equation", + "content": "I _ {i} ^ {t} = \\left\\{ \\begin{array}{l l} \\boldsymbol {f} _ {z g} \\left(\\operatorname {c o n c a t} \\left(z _ {i \\varphi} ^ {t}, g _ {i} ^ {t}\\right), I _ {i} ^ {t - 1}\\right) & \\text {i f t r a i n} \\\\ \\boldsymbol {f} _ {z g} \\left(\\operatorname {c o n c a t} \\left(z _ {i \\vartheta} ^ {t}, g _ {i} ^ {t}\\right), I _ {i} ^ {t - 1}\\right) & \\text {i f t e s t} \\end{array} \\right. \\tag {12}", + "image_path": "632e475c53600ffcf87c3afda507cc9970a164e5cc8cc758e456a9d25956e38c.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 344, + 296, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 344, + 296, + 392 + ], + "spans": [ + { + "bbox": [ + 55, + 344, + 296, + 392 + ], + "type": "text", + "content": "In the next time step, the updated multi-order fused intention contains the information of the latent variables, which is then combined with " + }, + { + "bbox": [ + 55, + 344, + 296, + 392 + ], + "type": "inline_equation", + "content": "B_{i}^{t}" + }, + { + "bbox": [ + 55, + 344, + 296, + 392 + ], + "type": "text", + "content": " to approximate the latent variables and update them." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 392, + 296, + 428 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 392, + 296, + 428 + ], + "spans": [ + { + "bbox": [ + 55, + 392, + 296, + 428 + ], + "type": "text", + "content": "By guiding the trajectory distribution through such a training process, the generated trajectories are explicitly represented by the latent variables as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 66, + 434, + 295, + 514 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 434, + 295, + 514 + ], + "spans": [ + { + "bbox": [ + 66, + 434, + 295, + 514 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} p \\left(g _ {i} ^ {T _ {H} + 1: T _ {F}} \\mid U _ {i}, S _ {i}\\right) \\\\ = \\prod_ {t = T _ {H} + 1} ^ {T _ {F}} \\int_ {z _ {i \\varphi} ^ {t}} p \\left(g _ {i} ^ {t} \\mid g _ {i} ^ {T _ {H} + 1: t - 1}, I _ {i} ^ {t - 1}, B _ {i} ^ {t}, z _ {i \\varphi} ^ {t}\\right) \\cdot \\tag {13} \\\\ p \\left(z _ {i \\varphi} ^ {t} \\mid g _ {i} ^ {T _ {H} + 1: t - 1}, I _ {i} ^ {t - 1}, B _ {i} ^ {t}\\right) d z _ {i \\varphi} ^ {t} \\\\ \\end{array}", + "image_path": "a5c74e51ddbb27fc1551382998edcc85db70e5efe417f039cc44ed3bc8c564d8.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 67, + 520, + 287, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 520, + 287, + 533 + ], + "spans": [ + { + "bbox": [ + 67, + 520, + 287, + 533 + ], + "type": "text", + "content": "In test mode, " + }, + { + "bbox": [ + 67, + 520, + 287, + 533 + ], + "type": "inline_equation", + "content": "B_{i}^{t}" + }, + { + "bbox": [ + 67, + 520, + 287, + 533 + ], + "type": "text", + "content": " is removed, and " + }, + { + "bbox": [ + 67, + 520, + 287, + 533 + ], + "type": "inline_equation", + "content": "\\varphi" + }, + { + "bbox": [ + 67, + 520, + 287, + 533 + ], + "type": "text", + "content": " is replaced with " + }, + { + "bbox": [ + 67, + 520, + 287, + 533 + ], + "type": "inline_equation", + "content": "\\vartheta" + }, + { + "bbox": [ + 67, + 520, + 287, + 533 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 539, + 208, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 539, + 208, + 552 + ], + "spans": [ + { + "bbox": [ + 55, + 539, + 208, + 552 + ], + "type": "text", + "content": "3.3. Global trajectory optimizer" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 556, + 296, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 556, + 296, + 652 + ], + "spans": [ + { + "bbox": [ + 55, + 556, + 296, + 652 + ], + "type": "text", + "content": "For efficient prediction, we introduce the KANs for the first time. We aggregate the " + }, + { + "bbox": [ + 55, + 556, + 296, + 652 + ], + "type": "inline_equation", + "content": "g_{i}^{T_{H} + 1:T_{F}}" + }, + { + "bbox": [ + 55, + 556, + 296, + 652 + ], + "type": "text", + "content": " time dimension to obtain " + }, + { + "bbox": [ + 55, + 556, + 296, + 652 + ], + "type": "inline_equation", + "content": "\\Gamma^0\\in \\mathbb{R}^{2T_F}" + }, + { + "bbox": [ + 55, + 556, + 296, + 652 + ], + "type": "text", + "content": " to facilitate global parallel optimization. The depth of the optimizer is " + }, + { + "bbox": [ + 55, + 556, + 296, + 652 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 55, + 556, + 296, + 652 + ], + "type": "text", + "content": " layers, the number of nodes in the " + }, + { + "bbox": [ + 55, + 556, + 296, + 652 + ], + "type": "inline_equation", + "content": "\\ell" + }, + { + "bbox": [ + 55, + 556, + 296, + 652 + ], + "type": "text", + "content": " -th layer is " + }, + { + "bbox": [ + 55, + 556, + 296, + 652 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 55, + 556, + 296, + 652 + ], + "type": "text", + "content": " , the number of nodes in the " + }, + { + "bbox": [ + 55, + 556, + 296, + 652 + ], + "type": "inline_equation", + "content": "\\ell +1" + }, + { + "bbox": [ + 55, + 556, + 296, + 652 + ], + "type": "text", + "content": " -th layer is " + }, + { + "bbox": [ + 55, + 556, + 296, + 652 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 556, + 296, + 652 + ], + "type": "text", + "content": " , and the activation function matrix corresponding to the nodes in the " + }, + { + "bbox": [ + 55, + 556, + 296, + 652 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 55, + 556, + 296, + 652 + ], + "type": "text", + "content": " -th layer is " + }, + { + "bbox": [ + 55, + 556, + 296, + 652 + ], + "type": "inline_equation", + "content": "\\Phi_{\\ell}" + }, + { + "bbox": [ + 55, + 556, + 296, + 652 + ], + "type": "text", + "content": " . The optimization process for the " + }, + { + "bbox": [ + 55, + 556, + 296, + 652 + ], + "type": "inline_equation", + "content": "\\ell \\rightarrow \\ell +1" + }, + { + "bbox": [ + 55, + 556, + 296, + 652 + ], + "type": "text", + "content": " layer is as follows:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 64, + 658, + 296, + 717 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 658, + 296, + 717 + ], + "spans": [ + { + "bbox": [ + 64, + 658, + 296, + 717 + ], + "type": "interline_equation", + "content": "\\Gamma^ {\\ell + 1} = \\Phi_ {\\ell} \\Gamma^ {\\ell} = \\left[ \\begin{array}{c c c c} \\varphi_ {1, 1} ^ {\\ell} & \\varphi_ {1, 2} ^ {\\ell} & \\dots & \\varphi_ {1, m} ^ {\\ell} \\\\ \\varphi_ {2, 1} ^ {\\ell} & \\varphi_ {2, 2} ^ {\\ell} & \\dots & \\varphi_ {2, m} ^ {\\ell} \\\\ \\vdots & \\vdots & \\ddots & \\vdots \\\\ \\varphi_ {n, 1} ^ {\\ell} & \\varphi_ {n, 2} ^ {\\ell} & \\dots & \\varphi_ {n, m} ^ {\\ell} \\end{array} \\right] \\Gamma^ {\\ell} (1 4)", + "image_path": "a2a91534ba41e219998ba6ba61af93eb39b9be57e8bcd12052106b89003136d7.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 72, + 554, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 108 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 108 + ], + "type": "text", + "content": "Eventually, the predicted trajectory is reduced back to the original dimensional space to obtain the final trajectory over the entire prediction period at once." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 318, + 118, + 554, + 133 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 118, + 554, + 133 + ], + "spans": [ + { + "bbox": [ + 318, + 118, + 554, + 133 + ], + "type": "interline_equation", + "content": "\\hat {g} _ {i} ^ {T _ {H} + 1: T _ {F}} \\leftarrow \\Gamma^ {L} = \\left(\\Phi_ {L} \\circ \\Phi_ {L - 1} \\circ \\dots \\circ \\Phi_ {1} \\circ \\Phi_ {0}\\right) \\Gamma^ {0} (1 5)", + "image_path": "6411bd92d6294c9c3e31f529bcac15e099ff80190e430c98a30f12e8f0b83834.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 144, + 554, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 144, + 554, + 168 + ], + "spans": [ + { + "bbox": [ + 313, + 144, + 554, + 168 + ], + "type": "text", + "content": "Given the diversity of agent intentions, the process predicts multiple possible trajectories." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 314, + 178, + 400, + 189 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 178, + 400, + 189 + ], + "spans": [ + { + "bbox": [ + 314, + 178, + 400, + 189 + ], + "type": "text", + "content": "3.4. Loss function" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 195, + 554, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 195, + 554, + 243 + ], + "spans": [ + { + "bbox": [ + 313, + 195, + 554, + 243 + ], + "type": "text", + "content": "Agents traveling in opposite or the same direction exert distinct effects on their respective movement states. Therefore, direction loss is incorporated to supervise the agent's dynamics during backpropagation." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 244, + 554, + 268 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 244, + 554, + 268 + ], + "spans": [ + { + "bbox": [ + 313, + 244, + 554, + 268 + ], + "type": "text", + "content": "For distance. The distance prediction error in future periods can be expressed as:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 393, + 278, + 554, + 294 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 393, + 278, + 554, + 294 + ], + "spans": [ + { + "bbox": [ + 393, + 278, + 554, + 294 + ], + "type": "interline_equation", + "content": "L _ {d i s} = \\left\\| \\hat {g} _ {i} ^ {t} - g _ {i g t} ^ {t} \\right\\| \\tag {16}", + "image_path": "cf866f7d19df288e9bfab75bf634ebcd38f5fed4bae5918ff33164269a4a0984.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 304, + 554, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 304, + 554, + 363 + ], + "spans": [ + { + "bbox": [ + 313, + 304, + 554, + 363 + ], + "type": "text", + "content": "For direction, the final moment of the historical period is introduced. Each predicted time step is related to the preceding time step to construct the direction vector, and subsequently, the angle between the predicted trajectory and the actual trajectory direction vector is analyzed." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 320, + 373, + 554, + 407 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 373, + 554, + 407 + ], + "spans": [ + { + "bbox": [ + 320, + 373, + 554, + 407 + ], + "type": "interline_equation", + "content": "L _ {a n g l e} = - \\operatorname {a r c c o s} \\left(\\frac {\\left(\\hat {g} _ {i} ^ {t} - \\hat {g} _ {i} ^ {t + 1}\\right) \\cdot \\left(g _ {i g t} ^ {t} - g _ {i g t} ^ {t + 1}\\right)}{\\left\\| \\hat {g} _ {i} ^ {t} - \\hat {g} _ {i} ^ {t + 1} \\right\\| \\left\\| g _ {i g t} ^ {t} - g _ {i g t} ^ {t + 1} \\right\\|}\\right) \\tag {17}", + "image_path": "62fbf1818ad1e7c0b90299855b3e602a0dddc1fb3b2ac098d43c08a9b99d531f.jpg" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 416, + 554, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 416, + 554, + 453 + ], + "spans": [ + { + "bbox": [ + 313, + 416, + 554, + 453 + ], + "type": "text", + "content": "The objective of our model is to maximize the lower bound on the evidence for all time steps of the target distribution, joints (16) and (17), with a final loss function of:" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 324, + 462, + 554, + 515 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 462, + 554, + 515 + ], + "spans": [ + { + "bbox": [ + 324, + 462, + 554, + 515 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} L = \\sum_ {t = T _ {H} + 1} ^ {T _ {F}} \\left\\{\\mathbb {E} _ {z _ {i \\varphi} ^ {t} \\sim q _ {\\varphi} \\left(\\cdot | B _ {i} ^ {t}, I _ {i} ^ {t - 1}\\right)} \\left[ L _ {d i s} + L _ {a n g l e} \\right] \\right. \\tag {18} \\\\ \\left. - K L \\left[ q _ {\\varphi} \\left(z _ {i \\varphi} ^ {t} \\mid B _ {i} ^ {t}, I _ {i} ^ {t - 1}\\right) \\| q _ {\\vartheta} \\left(z _ {i \\vartheta} ^ {t} \\mid I _ {i} ^ {t - 1}\\right) \\right] \\right\\} \\\\ \\end{array}", + "image_path": "201644b98c69b4eed89986cbe22b82ae0b0dbdd6d422e9c276ad0e7b3ad60070.jpg" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 314, + 525, + 395, + 539 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 525, + 395, + 539 + ], + "spans": [ + { + "bbox": [ + 314, + 525, + 395, + 539 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 545, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 555, + 713 + ], + "type": "text", + "content": "Datasets. ETH [32] and UCY [18] datasets encompass over 1,600 pedestrian trajectories, with positions annotated at 0.4s intervals. SDD [33] dataset is primarily utilized for pedestrian trajectory prediction and object tracking, containing about 19,000 pedestrian trajectories, with positions annotated every 0.4s. NBA dataset focuses on the 2015-2016 season [21, 53]. Two subsets, \"Rebounds\" and \"Scores,\" are extracted for benchmarking purposes. The movement intentions of the agents exhibit strong purposefulness and complex, hard-to-capture velocity variations. Consequently, the extracted trajectories encompass rich, highly nonlinear motion patterns. NuScenes [4] is a dataset for autonomous driving, and the prediction challenge segmentation and configurations are adhered to." + } + ] + } + ], + "index": 27 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "22469" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 102, + 555, + 236 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 93 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 93 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 93 + ], + "type": "text", + "content": "Table 1. Comparison of Related Methods on ETH/UCY, NBA, SDD, and NuScenes (in meters). Every work was realized using best-of-20 predictions for the effect of ade/fde was selected." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 102, + 555, + 236 + ], + "lines": [ + { + "bbox": [ + 56, + 102, + 555, + 236 + ], + "spans": [ + { + "bbox": [ + 56, + 102, + 555, + 236 + ], + "type": "table", + "html": "
ETH/UCYFLEAM [6]ST-MR [19]MTN [48]MANTRA [28]Agent-former [51]BiTrap [47]MemoNet [45]EqMotion [46]V2-Net-SC [39, 41]E-V2-Net-SC [40, 41]Ours
Eth0.66/1.210.53/0.690.58/0.860.56/0.970.58/0.860.66/1.090.51/0.690.36/0.540.23/0.330.23/0.300.26/0.31
Hotel0.37/0.640.20/0.310.19/0.570.17/0.320.15/0.240.17/0.320.14/0.200.12/0.160.11/0.160.10/0.130.10/0.12
Univ0.57/0.970.41/0.640.32/0.430.38/0.790.26/0.480.26/0.460.26/0.520.21/0.390.17/0.270.18/0.240.11/0.17
Zara010.46/0.740.26/0.380.24/0.370.24/0.410.19/0.320.23/0.470.21/0.360.16/0.270.13/0.180.13/0.160.10/0.17
Zara020.37/0.690.18/0.240.15/0.210.18/0.340.16/0.250.18/0.350.15/0.300.11/0.190.12/0.210.11/0.160.09/0.14
Avg0.48/0.850.32/0.450.30/0.490.31/0.570.27/0.430.30/0.540.25/0.410.19/0.310.15/0.230.15/0.200.13/0.18
NBAFLEAM [6]ST-MR [19]MTN [48]SGNet [38]EqMotion [46]MemoNet [45]BiTrap [47]GroupNet +CVAE [44]V2-Net-SC [39, 41]E-V2-Net-SC [40, 41]Ours
Rebound1.92/2.210.94/1.730.84/0.970.80/1.780.86/1.211.73/2.060.86/1.880.63/0.840.54/0.790.54/0.860.34/0.66
Scores1.29/1.680.90/1.030.87/1.410.71/1.450.69/1.021.35/1.870.76/1.550.46/0.760.50/0.840.47/0.780.30/0.56
SDDFLEAM [6]SHENet [29]MemoNet [45]MANTRA [28]BiTrap [47]MID [10]Y-Net [26]NSP-SFM [52]V2-Net-SC [39, 41]E-V2-Net-SC [40, 41]Ours
SDD0.67/1.280.44/0.530.45/0.640.49/0.620.46/0.590.38/0.730.32/0.660.21/0.340.26/0.360.23/0.350.17/0.24
NuScenesFLEAM [6]ST-MR [19]MTN [48]MANTRA [28]SGNet [38]Agent-former [51]EqMotion [46]Y-Net [26]V2-Net-SC [39, 41]E-V2-Net-SC [40, 41]Ours
NuScenes2.02/3.141.84/2.461.64/2.121.47/2.011.28/2.091.04/1.471.24/1.981.12/1.711.06/1.591.04/1.640.92/1.56
", + "image_path": "0d3d8cf87453daa9376edc3a5958150bd2bb5e1fb04fe85a89246de32448f3ff.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 255, + 294, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 255, + 294, + 316 + ], + "spans": [ + { + "bbox": [ + 55, + 255, + 294, + 316 + ], + "type": "text", + "content": "Metrics. Best-of-20 predictions were used to compute Average Displacement Error (ADE) and Final Displacement Error (FDE) as primary metrics, with estimated Negative Log Likelihood (NLL) reported from the test set. See their detailed definitions in the Supplementary Material." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 316, + 295, + 436 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 316, + 295, + 436 + ], + "spans": [ + { + "bbox": [ + 54, + 316, + 295, + 436 + ], + "type": "text", + "content": "Implementation Details. The model was trained on two NVIDIA RTX 4090 GPUs. Following the methodology in [5, 36], a leave-one-out cross-validation approach was employed for dataset handling. The number of subspaces in the higher-order intention interaction layer was set to " + }, + { + "bbox": [ + 54, + 316, + 295, + 436 + ], + "type": "inline_equation", + "content": "M = 6" + }, + { + "bbox": [ + 54, + 316, + 295, + 436 + ], + "type": "text", + "content": " while the depth of the global trajectory optimizer was set to " + }, + { + "bbox": [ + 54, + 316, + 295, + 436 + ], + "type": "inline_equation", + "content": "L = 3" + }, + { + "bbox": [ + 54, + 316, + 295, + 436 + ], + "type": "text", + "content": ". During inference, the best parameter model was loaded to predict the next 8 frames based on the previous 8 frames of history. For more implementation details of the training in the Supplementary Material." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 445, + 178, + 457 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 445, + 178, + 457 + ], + "spans": [ + { + "bbox": [ + 55, + 445, + 178, + 457 + ], + "type": "text", + "content": "4.1. Quantitative analysis" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 462, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 462, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 462, + 296, + 713 + ], + "type": "text", + "content": "All SOTA performances across datasets were highlighted in bold black in Table 1. In Eth scenario, crowds generally moved in parallel, resulting in weaker interactions among neighbors that conducted to target agent. E-V2-Net-SC [40, 41] achieved optimal performance in this scenario. However, the proposed method still achieved superior performance on ETH/UCY dataset, reducing ADE and FDE by " + }, + { + "bbox": [ + 56, + 462, + 296, + 713 + ], + "type": "inline_equation", + "content": "13.33\\%" + }, + { + "bbox": [ + 56, + 462, + 296, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 56, + 462, + 296, + 713 + ], + "type": "inline_equation", + "content": "10.00\\%" + }, + { + "bbox": [ + 56, + 462, + 296, + 713 + ], + "type": "text", + "content": " compared to E-V2-Net-SC, respectively. For NBA dataset, where interactions between agents were complex and frequent, the proposed method demonstrated exceptional performance. In Rebound sub-dataset, ADE and FDE decreased by " + }, + { + "bbox": [ + 56, + 462, + 296, + 713 + ], + "type": "inline_equation", + "content": "37.04\\%" + }, + { + "bbox": [ + 56, + 462, + 296, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 56, + 462, + 296, + 713 + ], + "type": "inline_equation", + "content": "16.46\\%" + }, + { + "bbox": [ + 56, + 462, + 296, + 713 + ], + "type": "text", + "content": ", respectively, compared to existing SOTA methods. In Scores sub-dataset, ADE and FDE were reduced by " + }, + { + "bbox": [ + 56, + 462, + 296, + 713 + ], + "type": "inline_equation", + "content": "34.78\\%" + }, + { + "bbox": [ + 56, + 462, + 296, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 56, + 462, + 296, + 713 + ], + "type": "inline_equation", + "content": "26.32\\%" + }, + { + "bbox": [ + 56, + 462, + 296, + 713 + ], + "type": "text", + "content": ", respectively, relative to the current SOTA methods. For SDD dataset, image segmentation techniques, as referenced in [2], were employed to capture pedestrian position information across eight scenario. Raw pixel coordinates were converted into spatial coordinates defined in meters for trajectory prediction tasks. NSP-SFM [52] outperformed both V2-Net-SC and E-V2-Net-SC. Our method re" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 255, + 555, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 255, + 555, + 316 + ], + "spans": [ + { + "bbox": [ + 313, + 255, + 555, + 316 + ], + "type": "text", + "content": "duced ADE and FDE by " + }, + { + "bbox": [ + 313, + 255, + 555, + 316 + ], + "type": "inline_equation", + "content": "19.05\\%" + }, + { + "bbox": [ + 313, + 255, + 555, + 316 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 255, + 555, + 316 + ], + "type": "inline_equation", + "content": "29.41\\%" + }, + { + "bbox": [ + 313, + 255, + 555, + 316 + ], + "type": "text", + "content": ", respectively, compared to NSP-SFM. For NuScenes dataset, V2-Net-SC and E-V2-Net-SC showed similar performance levels. Our method reduced ADE and FDE by " + }, + { + "bbox": [ + 313, + 255, + 555, + 316 + ], + "type": "inline_equation", + "content": "11.54\\%" + }, + { + "bbox": [ + 313, + 255, + 555, + 316 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 255, + 555, + 316 + ], + "type": "inline_equation", + "content": "1.87\\%" + }, + { + "bbox": [ + 313, + 255, + 555, + 316 + ], + "type": "text", + "content": ", respectively, in comparison to their SOTA performances." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 316, + 555, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 316, + 555, + 376 + ], + "spans": [ + { + "bbox": [ + 313, + 316, + 555, + 376 + ], + "type": "text", + "content": "Detailed comparative experiments with NLL metric were carried out in the Supplementary Material. Furthermore, for wider validation and analysis, we applied directional component in the designed loss function to different models. See the Supplementary Material for details." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 386, + 429, + 399 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 386, + 429, + 399 + ], + "spans": [ + { + "bbox": [ + 313, + 386, + 429, + 399 + ], + "type": "text", + "content": "4.2. Qualitative analysis" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 317, + 416, + 552, + 630 + ], + "blocks": [ + { + "bbox": [ + 317, + 416, + 552, + 630 + ], + "lines": [ + { + "bbox": [ + 317, + 416, + 552, + 630 + ], + "spans": [ + { + "bbox": [ + 317, + 416, + 552, + 630 + ], + "type": "image", + "image_path": "5e993945fc8199ac2be38e1d22beafcd37dcf953f777dfad5392d7a7801e8b21.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 641, + 555, + 675 + ], + "lines": [ + { + "bbox": [ + 313, + 641, + 555, + 675 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 555, + 675 + ], + "type": "text", + "content": "Figure 3. Representative cases. Row 1 represented complex historical trajectory cases. Row 2 represented state maintenance cases. Row 3 represented numerous neighbors cases." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": "Unlike previous studies [6, 26, 39-41, 45, 46], the focus here was on conducting experiments involving historical" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22470" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 111, + 72, + 499, + 300 + ], + "blocks": [ + { + "bbox": [ + 111, + 72, + 499, + 300 + ], + "lines": [ + { + "bbox": [ + 111, + 72, + 499, + 300 + ], + "spans": [ + { + "bbox": [ + 111, + 72, + 499, + 300 + ], + "type": "image", + "image_path": "4d03a52220cac1415396f477349406fc0a9f6c7b81ce2c3e7193d1868c4994d0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 308, + 555, + 331 + ], + "lines": [ + { + "bbox": [ + 55, + 308, + 555, + 331 + ], + "spans": [ + { + "bbox": [ + 55, + 308, + 555, + 331 + ], + "type": "text", + "content": "Figure 4. Predictions comparison of predictions in ETH/UCY. The comparison included heatmaps of the overall distribution of predictions and best-of-20 predictions of the related methods." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 351, + 296, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 351, + 296, + 639 + ], + "spans": [ + { + "bbox": [ + 56, + 351, + 296, + 639 + ], + "type": "text", + "content": "trajectories in various social contexts. This enabled a more effective validation of the model's ability to extract and analyze multi-order fused intentions in Fig. 3. In panel (a), the target agent continuously changed direction. Although the model displayed slight deviations at certain moments, these deviations remained within the acceptable range of directional changes. In panel (b) and (c), the target agent intended to turn, but the model successfully achieved accurate predictions. The target agent moved at a relatively steady speed in panel (d), adopting a conservative deceleration strategy in response to numerous neighbors who are stopping or remaining stationary. In panel (e) and (f), there were many stationary neighbors who may be gathered together in conversation. Consequently, the target agent maintained its original state. In panel (g) and (h), most neighbors moved in the same direction as the target agent, with closer neighbors receiving higher attention scores. This suggested that the target agent prioritized attention to closer neighbors, consistent with social norms. In panel (i), the directions of neighbors were complex, and their speeds varied. The target agent accounted for influences from neighbors in multiple directions while attempting to anticipate and avoid imminent collisions. Please see the qualitative analyses of NBA interactions in the Supplementary Material." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": "Comparisons were made between the proposed method and three previous methods in Fig. 4. It was observed that our method produced more accurate predictions than the previous methods. In Eth, when facing neighbors to the left front, the model provided a prediction closest to the true trajectory. The predictions from E-V2-Net-SC and EqMotion" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 351, + 555, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 351, + 555, + 472 + ], + "spans": [ + { + "bbox": [ + 313, + 351, + 555, + 472 + ], + "type": "text", + "content": "exhibited slight oscillations, while MemoNet, by mimicking retrospective memory mechanisms, tends to maintain its original trajectory. By employing the squeeze theorem to approximate future true trajectories, the proposed model shows greater convergence in the overall distribution of predictions in all scenarios. In Hotel and Zara01, there were many pedestrians ahead of the target agent, and the diverging trajectory distributions from previous methods represent rather aggressive strategies, which conflicted with the safety requirements of autonomous driving." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 480, + 406, + 493 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 480, + 406, + 493 + ], + "spans": [ + { + "bbox": [ + 313, + 480, + 406, + 493 + ], + "type": "text", + "content": "4.3. Ablation study" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 498, + 555, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 555, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 555, + 689 + ], + "type": "text", + "content": "In Table 2, ablation experiments were conducted by sequentially adding components to the PDI baseline. Groups 1-4 focus on the multi-order intention fusion module. Group 1 includes only the neighbors' position and the distances between the target agent and the neighbors. As additional components, such as neighboring speed " + }, + { + "bbox": [ + 313, + 498, + 555, + 689 + ], + "type": "inline_equation", + "content": "\\mathbf{V}" + }, + { + "bbox": [ + 313, + 498, + 555, + 689 + ], + "type": "text", + "content": ", the angle between the target agent and the neighboring speed " + }, + { + "bbox": [ + 313, + 498, + 555, + 689 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 313, + 498, + 555, + 689 + ], + "type": "text", + "content": ", and the final predicted distance over the historical period " + }, + { + "bbox": [ + 313, + 498, + 555, + 689 + ], + "type": "inline_equation", + "content": "\\mathbf{E}" + }, + { + "bbox": [ + 313, + 498, + 555, + 689 + ], + "type": "text", + "content": ", were gradually incorporated, a substantial reduction in the final prediction error was observed. Specifically, ADE decreased by " + }, + { + "bbox": [ + 313, + 498, + 555, + 689 + ], + "type": "inline_equation", + "content": "23.91\\%" + }, + { + "bbox": [ + 313, + 498, + 555, + 689 + ], + "type": "text", + "content": " and FDE by " + }, + { + "bbox": [ + 313, + 498, + 555, + 689 + ], + "type": "inline_equation", + "content": "39.47\\%" + }, + { + "bbox": [ + 313, + 498, + 555, + 689 + ], + "type": "text", + "content": " on the SDD dataset, while ADE reduced by " + }, + { + "bbox": [ + 313, + 498, + 555, + 689 + ], + "type": "inline_equation", + "content": "11.11\\%" + }, + { + "bbox": [ + 313, + 498, + 555, + 689 + ], + "type": "text", + "content": " and FDE by " + }, + { + "bbox": [ + 313, + 498, + 555, + 689 + ], + "type": "inline_equation", + "content": "14.42\\%" + }, + { + "bbox": [ + 313, + 498, + 555, + 689 + ], + "type": "text", + "content": " on the Nuscenes dataset. Groups 4-5, 5-6, and 6-7 focus on the ablation of the approximator, the optimizer, and the angle loss term in the loss function, respectively. The results demonstrate that each component contributed positively." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "content": "To examine the influence of " + }, + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{B}" + }, + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "content": ", the trajectory distribution for several scenarios in the future periods was visu-" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22471" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 93, + 134, + 519, + 218 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 126 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 126 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 126 + ], + "type": "text", + "content": "Table 2. Ablation studies of each component in our method on SDD and Nuscenes datasets. Notably, the positional information of the target agent was essential and was included in each set of ablation experiments. " + }, + { + "bbox": [ + 55, + 70, + 555, + 126 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 55, + 70, + 555, + 126 + ], + "type": "text", + "content": " represented the position of neighbors, " + }, + { + "bbox": [ + 55, + 70, + 555, + 126 + ], + "type": "inline_equation", + "content": "\\mathbf{V}" + }, + { + "bbox": [ + 55, + 70, + 555, + 126 + ], + "type": "text", + "content": " indicated their velocities, " + }, + { + "bbox": [ + 55, + 70, + 555, + 126 + ], + "type": "inline_equation", + "content": "\\mathbf{D}" + }, + { + "bbox": [ + 55, + 70, + 555, + 126 + ], + "type": "text", + "content": " denoted the distance between the target agent and its neighbors, and " + }, + { + "bbox": [ + 55, + 70, + 555, + 126 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 55, + 70, + 555, + 126 + ], + "type": "text", + "content": " represented the angle between their velocities. " + }, + { + "bbox": [ + 55, + 70, + 555, + 126 + ], + "type": "inline_equation", + "content": "\\mathbf{E}" + }, + { + "bbox": [ + 55, + 70, + 555, + 126 + ], + "type": "text", + "content": " represented the final predicted distance between the target agent and its neighbors over historical periods, " + }, + { + "bbox": [ + 55, + 70, + 555, + 126 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 55, + 70, + 555, + 126 + ], + "type": "text", + "content": " indicated global historical social intent, " + }, + { + "bbox": [ + 55, + 70, + 555, + 126 + ], + "type": "inline_equation", + "content": "\\mathbf{B}" + }, + { + "bbox": [ + 55, + 70, + 555, + 126 + ], + "type": "text", + "content": " denoted future actual trajectories, " + }, + { + "bbox": [ + 55, + 70, + 555, + 126 + ], + "type": "inline_equation", + "content": "\\mathbf{K}" + }, + { + "bbox": [ + 55, + 70, + 555, + 126 + ], + "type": "text", + "content": " referred to the parallel prediction mode, and " + }, + { + "bbox": [ + 55, + 70, + 555, + 126 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 55, + 70, + 555, + 126 + ], + "type": "text", + "content": " signified directional information." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 93, + 134, + 519, + 218 + ], + "lines": [ + { + "bbox": [ + 93, + 134, + 519, + 218 + ], + "spans": [ + { + "bbox": [ + 93, + 134, + 519, + 218 + ], + "type": "table", + "html": "
GroupPVDΘEIBKASDDNuscenes
1✓✓✓0.46/0.711.26/2.15
2✓✓✓✓0.45/0.661.24/2.09
3✓✓✓✓✓0.41/0.561.18/1.97
4✓✓✓✓✓✓0.35/0.431.12/1.84
5✓✓✓✓✓✓✓0.30/0.361.04/1.76
6✓✓✓✓✓✓✓✓0.20/0.270.93/1.61
7✓✓✓✓✓✓✓✓✓0.17/0.240.92/1.56
", + "image_path": "9bb161f19576b3106e794f05a4f9e59f79a7801fc7c8ad2e21afa388a8b28379.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 111, + 230, + 503, + 398 + ], + "blocks": [ + { + "bbox": [ + 111, + 230, + 503, + 398 + ], + "lines": [ + { + "bbox": [ + 111, + 230, + 503, + 398 + ], + "spans": [ + { + "bbox": [ + 111, + 230, + 503, + 398 + ], + "type": "image", + "image_path": "2706826b04e29baddbbc288d435acaf6cfecad63c3014a316e2d5bd7ee82c939.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 406, + 555, + 430 + ], + "lines": [ + { + "bbox": [ + 55, + 406, + 555, + 430 + ], + "spans": [ + { + "bbox": [ + 55, + 406, + 555, + 430 + ], + "type": "text", + "content": "Figure 5. Distribution heatmaps. Every heatmap of future trajectory distribution was shown above the real scene image. (a) represented the actual distribution. (b) presented the experimental group containing only I component. (c) showed the group with both I and B components." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 67, + 450, + 282, + 573 + ], + "blocks": [ + { + "bbox": [ + 67, + 450, + 282, + 573 + ], + "lines": [ + { + "bbox": [ + 67, + 450, + 282, + 573 + ], + "spans": [ + { + "bbox": [ + 67, + 450, + 282, + 573 + ], + "type": "image", + "image_path": "4765ac324e41e41878fe89a0229ce6f845a252c2fd285776c2ed00f3336e9b65.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 582, + 295, + 616 + ], + "lines": [ + { + "bbox": [ + 55, + 582, + 295, + 616 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 295, + 616 + ], + "type": "text", + "content": "Figure 6. Representative cases about the optimization effects of component " + }, + { + "bbox": [ + 55, + 582, + 295, + 616 + ], + "type": "inline_equation", + "content": "\\mathbf{K}" + }, + { + "bbox": [ + 55, + 582, + 295, + 616 + ], + "type": "text", + "content": ". To ensure fairness, we selected best-of-20 predictions in groups 6 and 7 for comparison." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": "alized in Fig. 5. It was evident that when both components " + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{B}" + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": " were included, the predicted trajectory distribution aligned more closely with the actual distribution. It can be found that in a strongly regular sports scenario such as Rebound, (b) and (c) achieved excellent predictions, while in open environments with complex social contexts, the effec" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 450, + 533, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 450, + 533, + 462 + ], + "spans": [ + { + "bbox": [ + 313, + 450, + 533, + 462 + ], + "type": "text", + "content": "tiveness of the approximator was better demonstrated." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 463, + 555, + 535 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 463, + 555, + 535 + ], + "spans": [ + { + "bbox": [ + 313, + 463, + 555, + 535 + ], + "type": "text", + "content": "To validate the effectiveness of component " + }, + { + "bbox": [ + 313, + 463, + 555, + 535 + ], + "type": "inline_equation", + "content": "\\mathbf{K}" + }, + { + "bbox": [ + 313, + 463, + 555, + 535 + ], + "type": "text", + "content": " under global parallel optimization, six challenging cases were presented in Fig. 6. It was evident that predictions without component " + }, + { + "bbox": [ + 313, + 463, + 555, + 535 + ], + "type": "inline_equation", + "content": "\\mathbf{K}" + }, + { + "bbox": [ + 313, + 463, + 555, + 535 + ], + "type": "text", + "content": " exhibited increasing errors as the prediction time accumulated, whereas the parallel prediction mode demonstrated a closer result to ground-truth." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 548, + 388, + 561 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 548, + 388, + 561 + ], + "spans": [ + { + "bbox": [ + 313, + 548, + 388, + 561 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 570, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 555, + 715 + ], + "type": "text", + "content": "Each human movement path is either directly or indirectly associated with neighbor agents. This study introduces SocialMOIF, which provides a more comprehensive representation of both direct and indirect intention interactions between the target agent and neighbor agents. SocialMOIF comprises four key designs: a multi-order intention fusion model, a trajectory distribution approximator, a global trajectory optimizer, and a novel loss function. Experimental results demonstrate that this approach performs competitively across various datasets and effectively approximates and explicitly updates trajectory distributions during training, thereby improving interpretability." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22472" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 210 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 210 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 210 + ], + "type": "text", + "content": "This project is supported by National Natural Science Foundation of China (52202417), China Postdoctoral Science Foundation(2022TQ0155, 2022M721605), Open Project Program of State Key Laboratory of Virtual Reality Technology and Systems, Beihang University(VRLAB2023A02), Young Elite Scientists Sponsorship Program by CAST(2023QNRC001), Fundamental Research Funds for the Central Universities (NS2024030), Young Elite Scientists Sponsorship Program by JSTJ(JSTJ-2023-XH032)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 220, + 115, + 232 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 220, + 115, + 232 + ], + "spans": [ + { + "bbox": [ + 56, + 220, + 115, + 232 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 240, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 61, + 240, + 296, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 240, + 296, + 295 + ], + "spans": [ + { + "bbox": [ + 61, + 240, + 296, + 295 + ], + "type": "text", + "content": "[1] Alexandre Alahi, Kratarth Goel, Vignesh Ramanathan, Alexandre Robicquet, Li Fei-Fei, and Silvio Savarese. Social lstm: Human trajectory prediction in crowded spaces. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 961–971, 2016. 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 295, + 296, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 295, + 296, + 338 + ], + "spans": [ + { + "bbox": [ + 61, + 295, + 296, + 338 + ], + "type": "text", + "content": "[2] Stefan Becker, Ronny Hug, Wolfgang Hubner, and Michael Arens. An evaluation of trajectory prediction approaches and notes on the trajnet benchmark. arXiv preprint arXiv:1805.07663, 2018. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 339, + 296, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 339, + 296, + 392 + ], + "spans": [ + { + "bbox": [ + 62, + 339, + 296, + 392 + ], + "type": "text", + "content": "[3] Graeme Best and Robert Fitch. Bayesian intention inference for trajectory prediction with an unknown goal destination. In 2015 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 5817-5823. IEEE, 2015. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 394, + 296, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 394, + 296, + 460 + ], + "spans": [ + { + "bbox": [ + 62, + 394, + 296, + 460 + ], + "type": "text", + "content": "[4] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11621-11631, 2020. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 460, + 296, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 460, + 296, + 514 + ], + "spans": [ + { + "bbox": [ + 62, + 460, + 296, + 514 + ], + "type": "text", + "content": "[5] Kai Chen, Xiao Song, and Xiaoxiang Ren. Pedestrian trajectory prediction in heterogeneous traffic using pose keypoints-based convolutional encoder-decoder network. IEEE Transactions on Circuits and Systems for Video Technology, 31(5): 1764-1775, 2020. 2, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 516, + 296, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 516, + 296, + 569 + ], + "spans": [ + { + "bbox": [ + 62, + 516, + 296, + 569 + ], + "type": "text", + "content": "[6] Kai Chen, Xiao Song, Haitao Yuan, and Xiaoxiang Ren. Fully convolutional encoder-decoder with an attention mechanism for practical pedestrian trajectory prediction. IEEE Transactions on Intelligent Transportation Systems, 23(11): 20046-20060, 2022. 2, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 571, + 296, + 614 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 571, + 296, + 614 + ], + "spans": [ + { + "bbox": [ + 62, + 571, + 296, + 614 + ], + "type": "text", + "content": "[7] Kai Chen, Haihua Zhu, Dunbing Tang, and Kun Zheng. Future pedestrian location prediction in first-person videos for autonomous vehicles and social robots. Image and Vision Computing, 134:104671, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 615, + 296, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 615, + 296, + 669 + ], + "spans": [ + { + "bbox": [ + 62, + 615, + 296, + 669 + ], + "type": "text", + "content": "[8] Lingwei Dang, Yongwei Nie, Chengjiang Long, Qing Zhang, and Guiqing Li. Msr-gen: Multi-scale residual graph convolution networks for human motion prediction. In Proceedings of the IEEE/CVF international conference on computer vision, pages 11467-11476, 2021. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 670, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 670, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 62, + 670, + 296, + 713 + ], + "type": "text", + "content": "[9] Francesco Giuliani, Irtiza Hasan, Marco Cristani, and Fabio Galasso. Transformer networks for trajectory forecasting. In 2020 25th international conference on pattern recognition (ICPR), pages 10335-10342. IEEE, 2021. 2" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 316, + 73, + 555, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 555, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 555, + 128 + ], + "type": "text", + "content": "[10] Tianpei Gu, Guangyi Chen, Junlong Li, Chunze Lin, Yongming Rao, Jie Zhou, and Jiwen Lu. Stochastic trajectory prediction via motion indeterminacy diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17113-17122, 2022. 3, 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 317, + 129, + 554, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 129, + 554, + 184 + ], + "spans": [ + { + "bbox": [ + 317, + 129, + 554, + 184 + ], + "type": "text", + "content": "[11] Agrim Gupta, Justin Johnson, Li Fei-Fei, Silvio Savarese, and Alexandre Alahi. Social gan: Socially acceptable trajectories with generative adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2255–2264, 2018. 2, 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 185, + 554, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 185, + 554, + 217 + ], + "spans": [ + { + "bbox": [ + 317, + 185, + 554, + 217 + ], + "type": "text", + "content": "[12] Dirk Helbing and Peter Molnar. Social force model for pedestrian dynamics. Physical review E, 51(5):4282, 1995. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 219, + 554, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 219, + 554, + 261 + ], + "spans": [ + { + "bbox": [ + 316, + 219, + 554, + 261 + ], + "type": "text", + "content": "[13] Serge Hoogendoorn and Piet HL Bovy. Simulation of pedestrian flows by optimal control and differential games. Optimal control applications and methods, 24(3):153-172, 2003. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 264, + 554, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 264, + 554, + 308 + ], + "spans": [ + { + "bbox": [ + 316, + 264, + 554, + 308 + ], + "type": "text", + "content": "[14] SP Hoogendoorn, Winnie Daamen, Yaqing Shu, and Han Ligteringen. Modeling human behavior in vessel maneuver simulation by optimal control and game theory. Transportation research record, 2326(1):45-53, 2013. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 309, + 554, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 309, + 554, + 365 + ], + "spans": [ + { + "bbox": [ + 316, + 309, + 554, + 365 + ], + "type": "text", + "content": "[15] Yihan Hu, Jiazhi Yang, Li Chen, Keyu Li, Chonghao Sima, Xizhou Zhu, Siqi Chai, Senyao Du, Tianwei Lin, Wenhai Wang, et al. Planning-oriented autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17853-17862, 2023. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 365, + 554, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 365, + 554, + 421 + ], + "spans": [ + { + "bbox": [ + 316, + 365, + 554, + 421 + ], + "type": "text", + "content": "[16] Sungjune Kim, Hyung-gun Chi, Hyerin Lim, Karthik Ramani, Jinkyu Kim, and Sangpil Kim. Higher-order relational reasoning for pedestrian trajectory prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15251-15260, 2024. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 422, + 554, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 422, + 554, + 464 + ], + "spans": [ + { + "bbox": [ + 316, + 422, + 554, + 464 + ], + "type": "text", + "content": "[17] Christos Kyrkou, Panayiotis Kolios, Theocharis Theocharides, and Marios Polycarpou. Machine learning for emergency management: A survey and future outlook. Proceedings of the IEEE, 111(1):19-41, 2022. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 466, + 554, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 466, + 554, + 499 + ], + "spans": [ + { + "bbox": [ + 317, + 466, + 554, + 499 + ], + "type": "text", + "content": "[18] Alon Lerner, Yiorgos Chrysanthou, and Dani Lischinski. Crowds by example. In Computer graphics forum, pages 655-664. Wiley Online Library, 2007. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 500, + 555, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 500, + 555, + 555 + ], + "spans": [ + { + "bbox": [ + 317, + 500, + 555, + 555 + ], + "type": "text", + "content": "[19] Lihuan Li, Maurice Pagnucco, and Yang Song. Graph-based spatial transformer with memory replay for multifuture pedestrian trajectory prediction. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2221-2231, 2022. 2, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 556, + 555, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 556, + 555, + 609 + ], + "spans": [ + { + "bbox": [ + 317, + 556, + 555, + 609 + ], + "type": "text", + "content": "[20] Linhui Li, Xiaotong Lin, Yejia Huang, Zizhen Zhang, and Jian-Fang Hu. Beyond minimum-of-n: Rethinking the evaluation and methods of pedestrian trajectory prediction. IEEE Transactions on Circuits and Systems for Video Technology, 2024. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 612, + 554, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 612, + 554, + 634 + ], + "spans": [ + { + "bbox": [ + 316, + 612, + 554, + 634 + ], + "type": "text", + "content": "[21] Kostya Linou. Nba-player-movements. Software for Visualization of NBA Games, 2016. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 635, + 554, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 635, + 554, + 690 + ], + "spans": [ + { + "bbox": [ + 317, + 635, + 554, + 690 + ], + "type": "text", + "content": "[22] Congcong Liu, Yuying Chen, Ming Liu, and Bertram E Shi. Avgcn: Trajectory prediction using graph convolutional networks guided by human attention. In 2021 IEEE International Conference on Robotics and Automation (ICRA), pages 14234-14240. IEEE, 2021. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 691, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 554, + 713 + ], + "type": "text", + "content": "[23] Ziming Liu, Pingchuan Ma, Yixuan Wang, Wojciech Ma-tusik, and Max Tegmark. Kan 2.0: Kolmogorov-arnold net" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22473" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "type": "text", + "content": "works meet science. arXiv preprint arXiv:2408.10205, 2024. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 140 + ], + "type": "text", + "content": "[24] Ziming Liu, Yixuan Wang, Sachin Vaidya, Fabian Ruehle, James Halverson, Marin Soljačić, Thomas Y Hou, and Max Tegmark. Kan: Kolmogorov-arnold networks. arXiv preprint arXiv:2404.19756, 2024. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 142, + 294, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 142, + 294, + 218 + ], + "spans": [ + { + "bbox": [ + 56, + 142, + 294, + 218 + ], + "type": "text", + "content": "[25] Karttikeya Mangalam, Harshayu Girase, Shreyas Agarwal, Kuan-Hui Lee, Ehsan Adeli, Jitendra Malik, and Adrien Gaidon. It is not the journey but the destination: Endpoint conditioned trajectory prediction. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part II 16, pages 759-776. Springer, 2020. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 220, + 294, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 220, + 294, + 274 + ], + "spans": [ + { + "bbox": [ + 56, + 220, + 294, + 274 + ], + "type": "text", + "content": "[26] Karttikeya Mangalam, Yang An, Harshayu Girase, and Jitendra Malik. From goals, waypoints & paths to long term human trajectory forecasting. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15233–15242, 2021. 1, 2, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 276, + 294, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 276, + 294, + 330 + ], + "spans": [ + { + "bbox": [ + 56, + 276, + 294, + 330 + ], + "type": "text", + "content": "[27] Weibo Mao, Chenxin Xu, Qi Zhu, Siheng Chen, and Yanfeng Wang. Leapfrog diffusion model for stochastic trajectory prediction. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5517-5526, 2023. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 332, + 295, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 332, + 295, + 387 + ], + "spans": [ + { + "bbox": [ + 56, + 332, + 295, + 387 + ], + "type": "text", + "content": "[28] Francesco Marchetti, Federico Becattini, Lorenzo Seidenari, and Alberto Del Bimbo. *Mantra: Memory augmented networks for multiple trajectory prediction*. In *Proceedings of the IEEE/CVF conference on computer vision and pattern recognition*, pages 7143–7152, 2020. 2, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 388, + 294, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 388, + 294, + 432 + ], + "spans": [ + { + "bbox": [ + 56, + 388, + 294, + 432 + ], + "type": "text", + "content": "[29] Mancheng Meng, Ziyan Wu, Terrence Chen, Xiran Cai, Xiang Zhou, Fan Yang, and Dinggang Shen. Forecasting human trajectory from scene history. arXiv preprint arXiv:2210.08732, 2022. 1, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 434, + 294, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 434, + 294, + 487 + ], + "spans": [ + { + "bbox": [ + 56, + 434, + 294, + 487 + ], + "type": "text", + "content": "[30] Alessio Monti, Alessia Bertugli, Simone Calderara, and Rita Cucchiara. Dag-net: Double attentive graph neural network for trajectory forecasting. In 2020 25th international conference on pattern recognition (ICPR), pages 2551-2558. IEEE, 2021. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 490, + 294, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 490, + 294, + 544 + ], + "spans": [ + { + "bbox": [ + 56, + 490, + 294, + 544 + ], + "type": "text", + "content": "[31] Brendan Tran Morris and Mohan Manubhai Trivedi. Trajectory learning for activity understanding: Unsupervised, multilevel, and long-term adaptive approach. IEEE transactions on pattern analysis and machine intelligence, 33(11):2287-2301, 2011. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 546, + 294, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 546, + 294, + 600 + ], + "spans": [ + { + "bbox": [ + 56, + 546, + 294, + 600 + ], + "type": "text", + "content": "[32] Stefano Pellegrini, Andreas Ess, Konrad Schindler, and Luc Van Gool. You'll never walk alone: Modeling social behavior for multi-target tracking. In 2009 IEEE 12th international conference on computer vision, pages 261-268. IEEE, 2009. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 602, + 294, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 602, + 294, + 646 + ], + "spans": [ + { + "bbox": [ + 56, + 602, + 294, + 646 + ], + "type": "text", + "content": "[33] Alexandre Robicquet, Amir Sadeghian, Alexandre Alahi, and Silvio Savarese. Learning social etiquette: Human trajectory prediction in crowded scenes. In European Conference on Computer Vision (ECCV), page 5, 2016. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 647, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 294, + 712 + ], + "type": "text", + "content": "[34] Tim Salzmann, Boris Ivanovic, Punarjay Chakravarty, and Marco Pavone. Trajectory: Dynamically-feasible trajectory forecasting with heterogeneous data. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XVIII 16, pages 683-700. Springer, 2020. 2" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 128 + ], + "type": "text", + "content": "[35] Liushuai Shi, Le Wang, Chengjiang Long, Sanping Zhou, Mo Zhou, Zhenxing Niu, and Gang Hua. Sgcn: Sparse graph convolution network for pedestrian trajectory prediction. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8994-9003, 2021. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 129, + 553, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 553, + 184 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 553, + 184 + ], + "type": "text", + "content": "[36] Xiao Song, Kai Chen, Xu Li, Jinghan Sun, Baocun Hou, Yong Cui, Baochang Zhang, Gang Xiong, and Zilie Wang. Pedestrian trajectory prediction based on deep convolutional LSTM network. IEEE Transactions on Intelligent Transportation Systems, 22(6):3285-3302, 2020. 2, 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 186, + 553, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 186, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 316, + 186, + 553, + 239 + ], + "type": "text", + "content": "[37] Yuchao Su, Jie Du, Yuanman Li, Xia Li, Rongqin Liang, Zhongyun Hua, and Jiantao Zhou. Trajectory forecasting based on prior-aware directed graph convolutional neural network. IEEE Transactions on Intelligent Transportation Systems, 23(9):16773-16785, 2022. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 241, + 553, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 241, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 316, + 241, + 553, + 285 + ], + "type": "text", + "content": "[38] Chuhua Wang, Yuchen Wang, Mingze Xu, and David J Crandall. Stepwise goal-driven networks for trajectory prediction. IEEE Robotics and Automation Letters, 7(2):2716-2723, 2022. 2, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 287, + 553, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 287, + 553, + 342 + ], + "spans": [ + { + "bbox": [ + 316, + 287, + 553, + 342 + ], + "type": "text", + "content": "[39] Conghao Wong, Beihao Xia, Ziming Hong, Qinmu Peng, Wei Yuan, Qiong Cao, Yibo Yang, and Xinge You. View vertically: A hierarchical network for trajectory prediction via fourier spectrums. In European Conference on Computer Vision, pages 682-700. Springer, 2022. 2, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 343, + 553, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 343, + 553, + 387 + ], + "spans": [ + { + "bbox": [ + 316, + 343, + 553, + 387 + ], + "type": "text", + "content": "[40] Conghao Wong, Beihao Xia, Qinmu Peng, and Xinge You. Another vertical view: A hierarchical network for heterogeneous trajectory prediction via spectrums. arXiv preprint arXiv:2304.05106, 2023. 2, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 388, + 553, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 388, + 553, + 453 + ], + "spans": [ + { + "bbox": [ + 316, + 388, + 553, + 453 + ], + "type": "text", + "content": "[41] Conghao Wong, Beihao Xia, Ziqian Zou, Yulong Wang, and Xinge You. Socialcircle: Learning the angle-based social interaction representation for pedestrian trajectory prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19005-19015, 2024. 1, 2, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 455, + 553, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 455, + 553, + 499 + ], + "spans": [ + { + "bbox": [ + 316, + 455, + 553, + 499 + ], + "type": "text", + "content": "[42] Beihao Xia, Conghao Wong, Qinmu Peng, Wei Yuan, and Xinge You. Cscnet: Contextual semantic consistency network for trajectory prediction in crowded spaces. Pattern Recognition, 126:108552, 2022. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 501, + 553, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 501, + 553, + 555 + ], + "spans": [ + { + "bbox": [ + 316, + 501, + 553, + 555 + ], + "type": "text", + "content": "[43] Song Xiao, Kai Chen, Xiaoxiang Ren, and Haitao Yuan. Pedestrian trajectory prediction in heterogeneous traffic using facial keypoints-based convolutional encoder-decoder network. ACM Transactions on Internet Technology, 22(4): 1-14, 2022. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 557, + 553, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 557, + 553, + 621 + ], + "spans": [ + { + "bbox": [ + 316, + 557, + 553, + 621 + ], + "type": "text", + "content": "[44] Chenxin Xu, Maosen Li, Zhenyang Ni, Ya Zhang, and Siheng Chen. Groupnet: Multiscale hypergraph neural networks for trajectory prediction with relational reasoning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6498-6507, 2022. 3, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 624, + 553, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 553, + 678 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 553, + 678 + ], + "type": "text", + "content": "[45] Chenxin Xu, Weibo Mao, Wenjun Zhang, and Siheng Chen. Remember intentions: Retrospective-memory-based trajectory prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6488-6497, 2022. 1, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 680, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 680, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 680, + 553, + 713 + ], + "type": "text", + "content": "[46] Chenxin Xu, Robby T Tan, Yuhong Tan, Siheng Chen, Yu Guang Wang, Xinchao Wang, and Yanfeng Wang. Eqmotion: Equivariant multi-agent motion prediction with invari" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "22474" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 531 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "type": "text", + "content": "ant interaction reasoning. arXiv preprint arXiv:2303.10876, 2023. 2, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 149 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 149 + ], + "type": "text", + "content": "[47] Yu Yao, Ella Atkins, Matthew Johnson-Roberson, Ram Vasudevan, and Xiaoxiao Du. Bitrap: Bi-directional pedestrian trajectory prediction with multi-modal goal estimation. IEEE Robotics and Automation Letters, 6(2):1463-1470, 2021. 2, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 294, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 294, + 205 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 294, + 205 + ], + "type": "text", + "content": "[48] Ziyi Yin, Ruijin Liu, Zhiliang Xiong, and Zejian Yuan. Multimodal transformer networks for pedestrian trajectory prediction. In Proceedings of the Twenty-Ninth International Joint Conference on Artificial Intelligence, pages 1259-1265, 2021. 2, 3, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 208, + 294, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 208, + 294, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 208, + 294, + 262 + ], + "type": "text", + "content": "[49] Cunjun Yu, Xiao Ma, Jiawei Ren, Haiyu Zhao, and Shuai Yi. Spatio-temporal graph transformer networks for pedestrian trajectory prediction. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XII 16, pages 507-523, 2020. 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 264, + 294, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 264, + 294, + 297 + ], + "spans": [ + { + "bbox": [ + 56, + 264, + 294, + 297 + ], + "type": "text", + "content": "[50] Wenjian Yu and Dirk Helbing. Game theoretical interactions of moving agents. In Simulating Complex Systems by Cellular Automata, pages 219-239. Springer, 2010. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 297, + 294, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 297, + 294, + 351 + ], + "spans": [ + { + "bbox": [ + 56, + 297, + 294, + 351 + ], + "type": "text", + "content": "[51] Ye Yuan, Xinshuo Weng, Yanglan Ou, and Kris M Kitani. Agentformer: Agent-aware transformers for socio-temporal multi-agent forecasting. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9813-9823, 2021. 2, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 353, + 294, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 353, + 294, + 386 + ], + "spans": [ + { + "bbox": [ + 56, + 353, + 294, + 386 + ], + "type": "text", + "content": "[52] Jiangbei Yue, Dinesh Manocha, and He Wang. Human trajectory prediction via neural social physics. In European conference on computer vision, pages 376-394, 2022. 3, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 388, + 294, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 388, + 294, + 431 + ], + "spans": [ + { + "bbox": [ + 56, + 388, + 294, + 431 + ], + "type": "text", + "content": "[53] Yisong Yue, Patrick Lucey, Peter Carr, Alina Bialkowski, and Iain Matthews. Learning fine-grained spatial models for dynamic sports play prediction. In 2014 IEEE international conference on data mining, pages 670-679. IEEE, 2014. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 433, + 295, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 433, + 295, + 487 + ], + "spans": [ + { + "bbox": [ + 56, + 433, + 295, + 487 + ], + "type": "text", + "content": "[54] Pu Zhang, Wanli Ouyang, Pengfei Zhang, Jianru Xue, and Nanning Zheng. Sr-lstm: State refinement for LSTM towards pedestrian trajectory prediction. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12085-12094, 2019. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 488, + 294, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 488, + 294, + 531 + ], + "spans": [ + { + "bbox": [ + 56, + 488, + 294, + 531 + ], + "type": "text", + "content": "[55] Xiaoping Zheng and Yuan Cheng. Conflict game in evacuation process: A study combining cellular automata model. Physica A: Statistical Mechanics and its Applications, 390 (6):1042-1050, 2011. 2" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22475" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Soft Self-labeling and Potts Relaxations for Weakly-supervised Segmentation/4b6c17bc-bca2-47a9-b3a7-cd9f02074980_content_list.json b/2025/Soft Self-labeling and Potts Relaxations for Weakly-supervised Segmentation/4b6c17bc-bca2-47a9-b3a7-cd9f02074980_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..22cca220a6c14eadb07c3d24fee9606b25b9d892 --- /dev/null +++ b/2025/Soft Self-labeling and Potts Relaxations for Weakly-supervised Segmentation/4b6c17bc-bca2-47a9-b3a7-cd9f02074980_content_list.json @@ -0,0 +1,2031 @@ +[ + { + "type": "text", + "text": "Soft Self-labeling and Potts Relaxations for Weakly-Supervised Segmentation", + "text_level": 1, + "bbox": [ + 297, + 130, + 699, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zhongwen Zhang & Yuri Boykov", + "bbox": [ + 349, + 203, + 645, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "University of Waterloo, Canada", + "bbox": [ + 374, + 222, + 624, + 238 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "z889zhan@uwaterloo.ca, yboykov@uwaterloo.ca", + "bbox": [ + 305, + 242, + 687, + 255 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 291, + 326, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We consider weakly supervised segmentation where only a fraction of pixels have ground truth labels (scribbles) and focus on a self-labeling approach optimizing relaxations of the standard unsupervised CRF/Potts loss on unlabeled pixels. While WSSS methods can directly optimize such losses via gradient descent, prior work suggests that higher-order optimization can improve network training by introducing hidden pseudo-labels and powerful CRF sub-problem solvers, e.g. graph cut. However, previously used hard pseudo-labels can not represent class uncertainty or errors, which motivates soft self-labeling. We derive a principled auxiliary loss and systematically evaluate standard and new CRF relaxations (convex and non-convex), neighborhood systems, and terms connecting network predictions with soft pseudolabels. We also propose a general continuous sub-problem solver. Using only standard architectures, soft self-labeling consistently improves scribble-based training and outperforms significantly more complex specialized WSSS systems. It can outperform full pixel-precise supervision. Our general ideas apply to other weakly-supervised problems/systems.", + "bbox": [ + 88, + 324, + 485, + 627 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Keywords: soft pseudo-labels, collision cross-entropy, collision divergence", + "bbox": [ + 89, + 630, + 478, + 643 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 676, + 220, + 693 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Full supervision for semantic segmentation requires thousands of training images with complete pixel-accurate ground truth masks. Their high costs explain the interest in weakly-supervised approaches based on image-level class tags [4, 23], pixel-level scribbles [28, 37, 38], or boxes [25]. This paper is focused on weak supervision with scribbles, which we also call seeds or partial masks. While only slightly more expensive than image-level class tags, scribbles on less than $3\\%$ of pixels were previously shown to achieve accuracy approaching full supervision without any modifications of the segmentation models. In contrast, tag supervision typically requires highly specialized systems and complex multi-stage training procedures, which are hard", + "bbox": [ + 89, + 703, + 485, + 901 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "to reproduce. Our interest in the scribble-based approach is motivated by its practical simplicity and mathematical clarity. The corresponding methodologies are focused on the design of unsupervised or self-supervised loss functions and stronger optimization algorithms. The corresponding solutions are often general and can be used in different weakly-supervised applications.", + "bbox": [ + 511, + 292, + 906, + 398 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1.1. Scribble-supervised segmentation", + "text_level": 1, + "bbox": [ + 513, + 407, + 808, + 422 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Let $\\Omega$ denote a set of image pixels and $S\\subset \\Omega$ denote a subset of pixels with ground truth labels, i.e. seeds or scribbles typically marked by mouse-controlled UI for image annotations. One example of image scribbles is in Fig.4(a). The ground truth label at any given pixel $i\\in S$ is an integer", + "bbox": [ + 511, + 429, + 906, + 503 + ], + "page_idx": 0 + }, + { + "type": "equation", + "text": "\n$$\n\\bar {y} _ {i} \\in \\{1, \\dots , K \\} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 651, + 513, + 906, + 531 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "where $K$ is the number of classes including the background. Without much ambiguity, it is convenient to use the same notation $\\bar{y}_i$ for the equivalent one-hot distribution", + "bbox": [ + 511, + 540, + 906, + 585 + ], + "page_idx": 0 + }, + { + "type": "equation", + "text": "\n$$\n\\bar {y} _ {i} \\equiv (\\bar {y} _ {i} ^ {1}, \\dots , \\bar {y} _ {i} ^ {K}) \\in \\Delta_ {0, 1} ^ {K} \\quad \\text {f o r} \\quad \\bar {y} _ {i} ^ {k} := [ k = \\bar {y} _ {i} ] \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 522, + 594, + 906, + 613 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "where $[\\cdot ]\\in \\{0,1\\}$ is the True operator. Set $\\Delta_{0,1}^{K}$ represents $K$ possible one-hot distributions, which are vertices of the $K$ -class probability simplex", + "bbox": [ + 511, + 622, + 906, + 669 + ], + "page_idx": 0 + }, + { + "type": "equation", + "text": "\n$$\n\\Delta^ {K} := \\{p = (p ^ {1}, \\dots , p ^ {K}) \\mid p ^ {k} \\geq 0, \\sum_ {k = 1} ^ {K} p ^ {k} = 1 \\}\n$$\n", + "text_format": "latex", + "bbox": [ + 532, + 676, + 885, + 720 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "representing all $K$ -categorical distributions. The context of each specific expression should make it obvious if $\\bar{y}_i$ is a class index (1) or the corresponding one-hot distribution (2).", + "bbox": [ + 511, + 727, + 906, + 772 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Loss functions for scribble-supervised segmentation typically use negative log-likelihoods (NLL) over scribbles $i \\in S \\subset \\Omega$ with ground truth labels $\\bar{y}_i$", + "bbox": [ + 511, + 773, + 906, + 818 + ], + "page_idx": 0 + }, + { + "type": "equation", + "text": "\n$$\n- \\sum_ {i \\in S} \\ln \\sigma_ {i} ^ {\\bar {y} _ {i}} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 666, + 827, + 906, + 859 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "where $\\sigma_{i} = (\\sigma_{i}^{1},\\dots,\\sigma_{i}^{K})\\in \\Delta^{K}$ are model predictions at pixels $i$ . This loss is also standard in full supervision", + "bbox": [ + 511, + 869, + 906, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "20244", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "where $S = \\Omega$ . In this case, no other losses are necessary for training. However, in a weakly supervised setting the majority of pixels are unlabeled, and unsupervised losses are needed for $i \\notin S$ .", + "bbox": [ + 89, + 90, + 480, + 151 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The most common unsupervised loss in image segmentation is the Potts model and its relaxations. It is a pairwise loss defined on pairs of neighboring pixels $\\{i,j\\} \\in \\mathcal{N}$ for a given neighborhood system $\\mathcal{N} \\subset \\Omega \\times \\Omega$ , typically corresponding to the nearest-neighbor grid (NN) [7, 19], or other sparse (SN) [40] or dense neighborhoods (DN) [24]. The original Potts model is defined for discrete segmentation variables, e.g. as in", + "bbox": [ + 89, + 152, + 482, + 272 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\sum_ {\\{i, j \\} \\in \\mathcal {N}} P (\\sigma_ {i}, \\sigma_ {j}) \\quad \\text {w h e r e} \\quad P (\\sigma_ {i}, \\sigma_ {j}) = [ \\sigma_ {i} \\neq \\sigma_ {j} ]\n$$\n", + "text_format": "latex", + "bbox": [ + 109, + 279, + 465, + 314 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "assuming integer-valued one-hot predictions $\\sigma_{i}\\in \\Delta_{0,1}^{K}$ This regularization loss encourages smoothness between the pixels. Its popular self-supervised variant is", + "bbox": [ + 89, + 321, + 483, + 367 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\nP \\left(\\sigma_ {i}, \\sigma_ {j}\\right) = w _ {i, j} \\cdot \\left[ \\sigma_ {i} \\neq \\sigma_ {j} \\right]\n$$\n", + "text_format": "latex", + "bbox": [ + 191, + 375, + 380, + 392 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "where pairwise affinities $w_{ij}$ are based on local intensity edges [7, 19, 24]. Of course, in the context of network training, one should use relaxations of $P$ applicable to soft predictions $\\sigma_i \\in \\Delta^K$ . Many forms of such relaxations [35, 45] were studied in segmentation, e.g. quadratic [19], bi-linear [38], total variation [9, 34], and others [15].", + "bbox": [ + 89, + 398, + 483, + 489 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Another general unsupervised loss relevant to training segmentation networks is the entropy of predictions, which is also known as decidiveness [8, 20]", + "bbox": [ + 89, + 489, + 483, + 534 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\sum_ {i} H (\\sigma_ {i})\n$$\n", + "text_format": "latex", + "bbox": [ + 250, + 541, + 323, + 571 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "where $H$ is the Shannon's entropy function. This loss is known to improve generalization and the quality of representation by moving (deep) features away from the decision boundaries [20]. Widely known in the context of unsupervised or semi-supervised classification, this loss also matters in weakly-supervised segmentation where it is used explicitly or implicitly1.", + "bbox": [ + 89, + 579, + 483, + 684 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Weakly-supervised segmentation methods [10, 22, 33, 41] also use other unsupervised losses (e.g. contrastive), clustering criteria (e.g. K-means), highly-specialized architectures, and ad-hoc training procedures. Our goal is to demonstrate that better results can be easily achieved on standard architectures by combining the standard losses above", + "bbox": [ + 89, + 685, + 483, + 776 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n- \\sum_ {i \\in S} \\ln \\sigma_ {i} ^ {\\bar {y} _ {i}} + \\eta \\sum_ {i \\notin S} H (\\sigma_ {i}) + \\lambda \\sum_ {i j \\in \\mathcal {N}} P (\\sigma_ {i}, \\sigma_ {j}). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 797, + 483, + 816 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "but replacing basic gradient descent optimization [38, 40] by higher-order self-labeling techniques [28-30] incorporating optimization of auxiliary pseudo-labels as sub-problems.", + "bbox": [ + 89, + 821, + 482, + 867 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1.2. Self-labeling and hard pseudo-labels", + "text_level": 1, + "bbox": [ + 514, + 90, + 828, + 107 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "One motivation for the self-labeling approach to weakly-supervised segmentation comes from well-known limitations of gradient descent for optimizing the Potts relaxations. Even convex relaxations [9, 19, 34] could be challenging in combination with the concave entropy term in (4).", + "bbox": [ + 511, + 112, + 906, + 186 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Typical self-labeling methods, including one of the first works on scribble-based semantic segmentation [28], introduce a sub-problem focused on the estimation of pseudolabels for unlabeled points. This subproblem can be formally \"derived\" from the original network optimization problem or \"designed\" as a heuristic, see the overview below. We focus on principled auxiliary loss formulations where network parameters and pseudo-labels are jointly optimized in a provably convergent manner. In this case pseudo-labels work as hidden optimization variables simplifying the model estimation, e.g. as in the EM algorithm for GMM.", + "bbox": [ + 511, + 188, + 906, + 353 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We denote pseudo-labels $y_{i}$ differently from the ground truth labels $\\bar{y}_{i}$ by omitting the bar. It is important to distinguish them since $\\bar{y}_{i}$ for $i \\in S$ are given, while $y_{i}$ for $i \\in \\Omega \\backslash S$ are estimated. The majority of existing self-labeling methods [2, 3, 26, 28-30, 42] estimate hard pseudo-labels, which could be equivalently represented either by class indices", + "bbox": [ + 511, + 354, + 906, + 445 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\ny _ {i} \\in \\{1, \\dots , K ^ {\\prime} \\} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 651, + 453, + 906, + 469 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "or by the corresponding one-hot categorical distributions", + "bbox": [ + 511, + 476, + 888, + 491 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\ny _ {i} \\equiv \\left(y _ {i} ^ {1}, \\dots , y _ {i} ^ {K}\\right) \\in \\Delta_ {0, 1} ^ {K} \\quad \\text {f o r} \\quad y _ {i} ^ {k} := [ k = y _ {i} ] \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 522, + 498, + 906, + 516 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "analogously with the hard ground truth labels in (1) and (2). In part, hard pseudo-labels are motivated by the network training where the default is NLL loss (3) assuming discrete labels. Besides, powerful discrete solvers for the Potts model are well-known [7, 9, 34]. We discuss the potential advantages of soft pseudo-labels in Section 1.3.", + "bbox": [ + 511, + 523, + 906, + 612 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Overview: joint loss vs \"proposal generation\": Most self-labeling approaches can be divided into two groups. One group designs pseudo-labeling and the network training sub-problems that are not formally related, e.g. [27, 28, 44]. While pseudo-labeling typically depends on the current network predictions and the network fine-tuning uses such pseudo-labels, the lack of a formal relation between these sub-problems implies that iterating such steps does not guarantee any form of convergence. Such methods are often referred to as proposal generation heuristics.", + "bbox": [ + 511, + 613, + 906, + 763 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Alternatively, the pseudo-labeling sub-problem and the network training sub-problem can be formally derived from a weakly-supervised loss like (4), e.g. by ADM splitting [30] or as high-order trust-region method [29]. Such methods often formulate a joint loss w.r.t. network predictions and pseudo-labels and iteratively decrease the loss. These self-labeling methods are mathematically well-founded, easy to understand and reproduce, and numerically stable due to guaranteed convergence.", + "bbox": [ + 511, + 765, + 908, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "1A unary decidiveness-like term is the difference between quadratic and bi-linear (tight but non-convex) relaxations [29, 35] of the Potts model.", + "bbox": [ + 89, + 875, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "20245", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1.3. Soft pseudo-labels: motivation & contributions", + "text_level": 1, + "bbox": [ + 91, + 90, + 483, + 107 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We observe that self-labeling with hard pseudo-labels $y_{i}$ is inherently limited as such labels cannot represent the uncertainty of class estimates at unlabeled pixels $i \\in \\Omega \\backslash S$ . Instead, we focus on soft pseudo-labels", + "bbox": [ + 89, + 112, + 485, + 174 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\ny _ {i} = \\left(y _ {i} ^ {1}, \\dots , y _ {i} ^ {K}\\right) \\in \\Delta^ {K} \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 189, + 183, + 483, + 202 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "which are general categorical distributions $p$ over $K$ -classes. It is possible that the estimated pseudo-label $y_{i}$ in (7) could be a one-hot distribution, which is a vertex of $\\Delta^K$ . In this special case, $y_{i}$ can be treated as a class index, but we use the general formulation (7) in the main parts of the paper starting Section 2. On the other hand, the ground truth labels $\\bar{y}_{i}$ are always hard and we use them either as indices (1) or one-hot distributions (2), as convenient.", + "bbox": [ + 89, + 212, + 483, + 332 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Soft pseudo-labels can be found in prior work on weakly-supervised segmentation [27, 44] using the \"soft proposal generation\". In contrast, we formulate soft self-labeling as a principled optimization methodology where network predictions and soft pseudo-labels are variables in a joint loss, which guarantees convergence of the training procedure. Our pseudo-labels are auxiliary variables for ADM-based [5] splitting of the loss (4) into two simpler optimization sub-problems: one focused on the Potts model over unlabeled pixels, and the other on the network training. While similar to [30], instead of hard, we use soft auxiliary variables for the Potts sub-problem. Our work can be seen as a study of the relaxed Potts sub-problem in the context of weakly-supervised semantic segmentation. The related prior work is focused on discrete solvers fundamentally unable to represent class estimate uncertainty. Our contributions can be summarized as follows:", + "bbox": [ + 89, + 333, + 483, + 589 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- convergent soft self-labeling framework based on a simple joint self-labeling loss", + "- systematic evaluation of Potts relaxations and (cross-) entropy terms in our loss", + "- state-of-the-art in scribble-based semantic segmentation based on standard network architectures (no modifications) and well-founded convergent training procedures that are easy to reproduce and generalize to other problems", + "- using the same segmentation model, our self-labeling loss with $3\\%$ scribbles may outperform standard supervised cross-entropy loss with full ground truth masks." + ], + "bbox": [ + 89, + 590, + 483, + 755 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. Our soft self-labeling approach", + "text_level": 1, + "bbox": [ + 89, + 768, + 375, + 787 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "First, we apply ADM splitting [5] to weakly supervised loss (4) to formulate our self-labeling loss (8) incorporating additional soft auxiliary variables, i.e. pseudo-labels (7). It is convenient to introduce pseudo-labels $y_{i}$ on all pixels in $\\Omega$ even though a subset of pixels (seeds) $S\\subset \\Omega$ have ground truth labels $\\bar{y}_i$ . We will simply impose a constraint that pseudo-labels and ground truth labels agree on $S$ . Thus,", + "bbox": [ + 89, + 794, + 485, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "we assume the following set of pseudo-labels", + "bbox": [ + 511, + 90, + 815, + 106 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nY _ {\\Omega} := \\left\\{y _ {i} \\in \\Delta^ {K} \\mid i \\in \\Omega , \\text {s . t .} y _ {i} = \\bar {y} _ {i} \\text {f o r} i \\in S \\right\\}.\n$$\n", + "text_format": "latex", + "bbox": [ + 544, + 114, + 872, + 133 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We split the terms in (4) into two groups: one includes NLL and entropy $H$ terms keeping the original prediction variables $\\sigma_{i}$ and the other includes the Potts relaxation $P$ replacing $\\sigma_{i}$ with auxiliary variables $y_{i}$ . This transforms loss (4) into expression", + "bbox": [ + 511, + 142, + 906, + 218 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n- \\sum_ {i \\in S} \\ln \\sigma_ {i} ^ {\\bar {y} _ {i}} + \\eta \\sum_ {i \\not \\in S} H (\\sigma_ {i}) + \\lambda \\sum_ {i j \\in \\mathcal {N}} P (y _ {i}, y _ {j})\n$$\n", + "text_format": "latex", + "bbox": [ + 544, + 227, + 875, + 261 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "equivalent to (4) assuming equality $\\sigma_{i} = y_{i}$ . The standard approximation is to incorporate constraint $\\sigma_{i} \\approx y_{i}$ directly into the loss, e.g. using $KL$ -divergence. For simplicity, we use weight $\\eta$ for $KL(\\sigma_{i},y_{i})$ to combine it with $H(\\sigma_{i})$ into a single cross-entropy term", + "bbox": [ + 511, + 268, + 906, + 335 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n-\\sum_{i\\in S}\\ln \\sigma_{i}^{\\bar{y}_{i}} + \\underbrace{\\eta\\sum_{i\\notin S}H(\\sigma_{i}) + \\eta\\sum_{i\\notin S}KL(\\sigma_{i},y_{i})}_{\\substack{\\eta \\sum_{i\\notin S}H(\\sigma_{i},y_{i})}} + \\lambda \\sum_{ij\\in \\mathcal{N}}P(y_{i},y_{j})\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 338, + 906, + 405 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "defining auxiliary self-labeling loss w.r.t. both $\\sigma_{i}$ and $y_{i}$", + "bbox": [ + 516, + 410, + 885, + 426 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n- \\sum_ {i \\in S} \\ln \\sigma_ {i} ^ {\\bar {y} _ {i}} + \\eta \\sum_ {i \\notin S} H \\left(\\sigma_ {i}, y _ {i}\\right) + \\lambda \\sum_ {i j \\in \\mathcal {N}} P \\left(y _ {i}, y _ {j}\\right) \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 522, + 435, + 906, + 454 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "approximating the original weakly supervised loss (4).", + "bbox": [ + 511, + 462, + 872, + 477 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Iterative minimization of this loss w.r.t. predictions $\\sigma_{i}$ (model parameters training) and pseudo-labels $y_{i}$ effectively breaks the original optimization problem for (4) into two simpler sub-problems, assuming there is a good solver for optimal pseudo-labels. The latter is plausible since the unary term $H(\\sigma_{i}, y_{i})$ is convex for $y_{i}$ , and the Potts model (incl. convex and non-convex relaxations) is widely studied.", + "bbox": [ + 511, + 478, + 906, + 582 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Section 2.1 discusses standard and new relaxations of the Potts model $P$ . Section 2.2 discusses several robust variants of the cross-entropy term $H$ in (8) connecting predictions with uncertain (soft) pseudo-labels $y_{i}$ at unlabeled points $i \\in \\Omega \\backslash S$ . To systematically evaluate these variants of (8) in Section 3, we propose a general efficient GPU-friendly solver for the corresponding pseudo-labeling sub-problem, which is detailed in the Supplementary Material.", + "bbox": [ + 511, + 583, + 908, + 704 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1. Second-order relaxations of the Potts model", + "text_level": 1, + "bbox": [ + 511, + 712, + 883, + 727 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We limit our focus to basic second-order relaxations and several motivated extensions. We start from two important standard cases (see Table 1): quadratic, the simplest convex relaxation popularized by the random walker algorithm [19], and bi-linear, which is non-convex but tight [35] w.r.t. the original discrete Potts model. The latter implies that optimizing it over relaxed variables should lead to a solution consistent with a discrete Potts solver, e.g. graph cut [7] regularizing the geometric boundary of the segments [6]. The quadratic relaxation produces a significantly different soft solution motivated by random walk probabilities [19].", + "bbox": [ + 511, + 734, + 908, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "20246", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/db6efe803d38739d8276fdcbf261bbc8ebc147133bfc24194b29bbf4fb5729d3.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
bi-linear ~ “graph cut”PBL(p,q) := 1 - pTqquadratic ~ “random walker”P0(p,q) := 1/2||p-q|2
normalized quadratic
PNQ(p,q) := 1 - pTq/||p||q|≡ 1/2||p/||p-| - q/||q| |2
", + "bbox": [ + 125, + 87, + 450, + 160 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "While there are specialized solvers for bi-linear and quadratic relaxations, we are interested in a general efficient pseudo-label solver integrated into the standard GPU-based network optimization framework. This suggests gradient descent as an optimizer for pseudo-labels $y$ (see Suppl. Mat). Figure 1 shows two representative local minima issues that gradient descent may encounter for (a) the bi-linear and (b) quadratic relaxations of the Potts loss. These examples indicate that gradient descent for these relaxations can distort the properties of the original (discrete) Potts loss.", + "bbox": [ + 89, + 196, + 483, + 345 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In (a) two neighboring pixels jointly change the common label from $y_{i} = y_{j} = (1,0,0)$ to $y_{i}^{\\prime \\prime} = y_{j}^{\\prime \\prime} = (0,1,0)$ , which corresponds to a \"move\" where the whole object is reclassified from A to B. This move does not violate the smoothness constraint represented by the original (discrete) Potts model. But, bilinear relaxation prevents move (a) since the intermediate state $y_{i}^{\\prime} = y_{j}^{\\prime} = (\\frac{1}{2},\\frac{1}{2},0)$ has a higher cost", + "bbox": [ + 89, + 348, + 483, + 455 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nP _ {\\mathrm {B L}} \\left(y _ {i} ^ {\\prime}, y _ {j} ^ {\\prime}\\right) = \\frac {1}{2} > 0 = P _ {\\mathrm {B L}} \\left(y _ {i}, y _ {j}\\right) = P _ {\\mathrm {B L}} \\left(y _ {i} ^ {\\prime \\prime}, y _ {j} ^ {\\prime \\prime}\\right)\n$$\n", + "text_format": "latex", + "bbox": [ + 112, + 465, + 460, + 496 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "while quadratic relaxation assigns constant (zero) loss for all states during this move.", + "bbox": [ + 89, + 506, + 482, + 536 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Figure 1(b) shows a move problematic for the quadratic relaxation. Two neighboring pixels on the boundary of objects A and C have labels $y_{i} = (1,0,0)$ and $y_{j} = (0,0,1)$ . The second object attempts to change from class label C to B that does not affect the Potts discontinuity penalty between two pixels. But, quadratic relaxation prefers that the second object stays in the intermediate soft state $y_{j}^{\\prime} = (0,\\frac{1}{2},\\frac{1}{2})$", + "bbox": [ + 89, + 537, + 483, + 643 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nP _ {\\mathrm {Q}} (y _ {i}, y _ {j} ^ {\\prime}) = \\frac {3}{4} < 1 = P _ {\\mathrm {Q}} (y _ {i}, y _ {j}) = P _ {\\mathrm {Q}} (y _ {i}, y _ {j} ^ {\\prime \\prime})\n$$\n", + "text_format": "latex", + "bbox": [ + 120, + 655, + 450, + 686 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/31f8396aa63da9926854ef470abb9e99765435b8f035951b0f7f713f97359515.jpg", + "image_caption": [ + "(a) both pixels change $(\\mathrm{A}\\rightarrow \\mathrm{B})$" + ], + "image_footnote": [], + "bbox": [ + 147, + 709, + 267, + 785 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/19a68655c2a8b548079376697f6c0101827163c2921d6edfbdde4a4e268d031e.jpg", + "image_caption": [ + "(b) only pixel $j$ changes $(\\mathbf{C}\\rightarrow \\mathbf{B})$", + "Figure 1. Examples of \"moves\" for neighboring pixels $\\{i,j\\} \\in \\mathcal{N}$ . Their (soft) pseudo-labels $y_{i}$ and $y_{j}$ are illustrated on the probability simplex $\\Delta^K$ for $K = 3$ . In (a) both pixels $i$ and $j$ are inside a region/object changing its label from A to B. In (b) pixels $i$ and $j$ are on the boundary between two regions/objects; one is fixed to class A and the other changes from class C to B." + ], + "image_footnote": [], + "bbox": [ + 305, + 709, + 426, + 785 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/d4224969ca9d5d4e17edc5da938592645709130742acddba351cd688be4420ae.jpg", + "table_caption": [ + "Table 1. Second-order Potts relaxations, see Fig.2(a,b,c)" + ], + "table_footnote": [], + "table_body": "
collision cross entropy\nPCC(r(p,q) := -ln p^T qlog-quadratic\nP_LQ(p,q) := -ln(1 - ||p-q||^2/2)
collision divergence\nPCD(p,q) := -ln ||p^T q/||p/||q||- ln(1 - 1/2 ||p/||p - q/||q||^2)
", + "bbox": [ + 549, + 87, + 874, + 157 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Table 2. Log-based Potts relaxations, see Fig.2(d,e,f)", + "bbox": [ + 550, + 167, + 867, + 183 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "while bi-linear relaxation $P_{\\mathrm{BL}}(y_i, y_j) = 1$ remains constant as $y_j$ transitions from state C to B.", + "bbox": [ + 511, + 193, + 903, + 223 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We propose a new relaxation, normalized quadratic in Table 1. Normalization leads to equivalence between quadratic and bi-linear formulations combining their benefits. As easy to check, normalized quadratic relaxation $P_{\\mathrm{NQ}}$ removes the local minima in both examples of Figure 1. Table 2 also proposes \"logarithmic\" versions of the relaxations in Table 1 composing them with function $-\\ln (1 - x)$ . As illustrated by Figure 2, the logarithmic versions (d-f) address the \"vanishing gradients\" evident from the flat regions in (a-c).", + "bbox": [ + 511, + 224, + 906, + 359 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.2. Cross-entropy and soft pseudo-labels", + "text_level": 1, + "bbox": [ + 511, + 369, + 833, + 386 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Shannon's cross-entropy $H(y, \\sigma)$ is the most common loss for training network predictions $\\sigma$ from ground truth labels $y$ in the context of classification, semantic segmentation, etc. However, this loss may not be ideal for applications where the targets $y$ are soft categorical distributions representing various forms of class uncertainty. For example, this paper is focused on scribble-based segmentation where the ground truth is not known for most of the pixels, and the network training is done jointly with estimating pseudo-labels $y$ for the unlabeled pixels. In this case, soft labels $y$ are distributions representing class uncertainty. We observe that if such $y$ is used as a target in $H(y, \\sigma)$ , the network is trained to reproduce the uncertainty, see Figure 3(a). This motivates the discussion of alternative \"cross-entropy\" functions where the quotes indicate an informal interpretation of this information-theoretic concept. Intuitively, such functions should encourage decisiveness, as well as proximity between the predictions and pseudo-labels, but avoid mimicking the uncertainty in both directions: from soft pseudo-labels to predictions and vice-versa. We show that the last property can be achieved in a probabilistically principled manner. The following paragraphs discuss three cross-entropy functions $H$ that we study in the context of our self-labeling loss (8).", + "bbox": [ + 511, + 392, + 906, + 739 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Standard cross-entropy provides the obvious baseline for evaluating two alternative versions that follow. For completeness, we include its mathematical definition", + "bbox": [ + 511, + 739, + 906, + 785 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nH _ {\\mathrm {C E}} \\left(y _ {i}, \\sigma_ {i}\\right) = H \\left(y _ {i}, \\sigma_ {i}\\right) \\equiv - \\sum_ {k} y _ {i} ^ {k} \\ln \\sigma_ {i} ^ {k} \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 550, + 797, + 906, + 829 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "and remind the reader that this loss is primarily used with hard or one-hot labels, in which case it is also equivalent to NLL loss - $\\ln \\sigma_i^{y_i}$ previously discussed for ground truth labels (3). As mentioned earlier, Figure 3(a) shows that for", + "bbox": [ + 511, + 839, + 906, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "20247", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/46b0c1694453a6d3d12bc007b4636fb591334bcf3f67ea99405c82c8ed82358e.jpg", + "image_caption": [ + "Figure 2. Second-order Potts relaxations in Tables 1 and 2: interaction potentials $P$ for pairs of predictions $(\\sigma_i, \\sigma_j)$ in (4) or pseudo-labels $(y_i, y_j)$ in (8) are illustrated for $K = 2$ when each $\\sigma_i$ or $y_i$ (binary distributions in $\\Delta^2$ ) can be represented by a single scalar as $(x, 1 - x)$ . The contour maps are iso-levels of $P((x_i, 1 - x_i), (x_j, 1 - x_j))$ over domain $(x_i, x_j) \\in [0, 1]^2$ . The 3D plots above illustrate the potentials $P$ as functions over pairs of \"logits\" $(l_i, l_j) \\in \\mathbb{R}^2$ where each scalar $l_i$ defines binary distribution $(x_i, 1 - x_i)$ for $x_i = \\frac{1}{1 + e^{-2l_i}} \\in [0, 1]$ ." + ], + "image_footnote": [], + "bbox": [ + 94, + 88, + 906, + 305 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/b2db9ae6cb80b535e88f7329a37e10372f6ab064c854737f6e5261af754da1cd.jpg", + "image_caption": [ + "(a) standard $H_{\\mathrm{CE}}(y,\\sigma)$" + ], + "image_footnote": [], + "bbox": [ + 117, + 385, + 284, + 477 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/4d1f589b41658e5832c2fc49bf33b725a1bfef477a4d60b4376fcaa95ef1912e.jpg", + "image_caption": [ + "Figure 3. Illustration of cross-entropy functions: (a) standard (9), (b) reverse (10), and (c) collision (11). (d) shows the empirical comparison on the robustness to label uncertainty. The test uses ResNet-18 architecture on fully-supervised Natural Scene dataset [32] where we corrupted some labels. The horizontal axis shows the percentage $\\eta$ of training images where the correct ground truth labels were replaced by a random label. All losses trained the model using soft target distributions $\\hat{y} = \\eta * u + (1 - \\eta) * y$ representing the mixture of one-hot distribution $y$ for the observed corrupt label and the uniform distribution $u$ , following [31]. The vertical axis shows the test accuracy. Training with the reverse and collision cross-entropy is robust to much higher levels of label uncertainty." + ], + "image_footnote": [], + "bbox": [ + 289, + 387, + 454, + 477 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/d95d627af0eaa8670a8b91a7672463a1f1a1f281a7923bbff07b04d8e010fe64.jpg", + "image_caption": [ + "(b) reverse $H_{\\mathrm{RCE}}(y,\\sigma)$" + ], + "image_footnote": [], + "bbox": [ + 460, + 387, + 625, + 477 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/5d6c5709ee51b1acfe9156ba39ee372ee65e69df03cfdfe2b587762885382f71.jpg", + "image_caption": [ + "(c) collision $H_{\\mathrm{CCE}}(y,\\sigma)$", + "(d) empirical comparison" + ], + "image_footnote": [], + "bbox": [ + 638, + 387, + 875, + 477 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "soft pseudo-labels like $y = (0.5, 0.5)$ , it forces predictions to mimic or replicate the uncertainty $\\sigma \\approx y$ . In fact, label $y = (0.5, 0.5)$ just tells that the class is unknown and the network should not be supervised by this point. This problem manifests itself in the poor performance of the standard cross-entropy (9) in our experiment discussed in Figure 3 (d) (red).", + "bbox": [ + 89, + 619, + 483, + 710 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Reverse cross-entropy is defined in this paper as", + "bbox": [ + 109, + 710, + 437, + 726 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nH _ {\\mathrm {R C E}} \\left(y _ {i}, \\sigma_ {i}\\right) = H \\left(\\sigma_ {i}, y _ {i}\\right) \\equiv - \\sum_ {k} \\sigma_ {i} ^ {k} \\ln y _ {i} ^ {k} \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 112, + 737, + 482, + 767 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "switching the order of labels and predictions in (9), which is not common. Indeed, Shannon's cross-entropy is not symmetric and the first argument is normally the target distribution and the second is the estimated distribution. However, in our case, both distributions are estimated and there is no reason not to try the reverse order. It is worth noting that our self-labeling formulation (8) suggests that reverse cross-entropy naturally appears when the ADM approach splits the", + "bbox": [ + 89, + 780, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "decisiveness and fairness into separate sub-problems. Moreover, as Figure 3(b) shows, in this case, the network does not mimic uncertain pseudo-labels, e.g. the gradient of the blue line is zero. The results for the reverse cross-entropy in Figure 3 (d) (green) are significantly better than for the standard (red). Unfortunately, now pseudo-labels $y$ mimic the uncertainty in predictions $\\sigma$ .", + "bbox": [ + 511, + 619, + 906, + 724 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Collision cross-entropy defined in this paper as", + "bbox": [ + 532, + 724, + 852, + 739 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nH _ {\\mathrm {C C E}} \\left(y _ {i}, \\sigma_ {i}\\right) \\equiv - \\ln \\sum_ {k} \\sigma_ {i} ^ {k} y _ {i} ^ {k} \\equiv - \\ln \\sigma^ {\\top} y \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 537, + 748, + 906, + 779 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "resolves the problem in a principled way. It is symmetric w.r.t. pseudo-labels and predictions. The dot product $\\sigma^{\\top}y$ can be seen as a probability of equality of two random variables, the predicted class $C$ and unknown true class $T$ , which are represented by the distributions $\\sigma$ and $y$ . Indeed,", + "bbox": [ + 511, + 787, + 906, + 864 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\Pr (C = T) = \\sum_ {k} P r (C = k) \\Pr (T = k) = \\sigma^ {\\top} y.\n$$\n", + "text_format": "latex", + "bbox": [ + 539, + 871, + 877, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "20248", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Loss (11) maximizes this \"collision\" probability instead of enforcing the equality of distributions $\\sigma = y$ . Figure 3(c) shows no mimicking of uncertainty (blue line). Unlike the reverse cross-entropy, this is also valid when $y$ is estimated from uncertain predictions $\\sigma$ since (11) is symmetric. This leads to the best performance in Figure 3 (d) (blue). Our extensive experiments are conclusive that collision cross-entropy is the best option for $H$ in self-labeling loss (8).", + "bbox": [ + 89, + 90, + 483, + 212 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3. Experiments", + "text_level": 1, + "bbox": [ + 89, + 227, + 222, + 243 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our experiments evaluate the components of our loss (8) (cross-entropy, pairwise term, and neighborhood). We compare standard segmentation architectures trained using the \"textbook\" approach, i.e. loss minimization, to the state-of-the-art methods, which often require network modifications and ad-hoc training procedures. Our quantitative tests in Section 3.1 evaluate different Potts relaxations. Some qualitative examples are presented in Figure 4. Section 3.2 compares three cross-entropy terms. Section 3.3 evaluates our soft self-labeling loss on the nearest-neighbor and dense neighborhood systems. We summarized the results in Section 3.4. Section 3.5 shows that our method achieves the SOTA and, in some cases, may outperform the full supervision.", + "bbox": [ + 88, + 253, + 485, + 450 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/320ea53741265cd39b5c7e57e714b7d10b28a7711b2a64b6acabd89d1b33d04c.jpg", + "image_caption": [ + "(a) Image, GT & input", + "Figure 4. Illustration of the difference among Potts relaxations. The visualization of soft pseudo-labels uses the convex combination of RGB colors for each class weighted by pseudo-label itself." + ], + "image_footnote": [], + "bbox": [ + 130, + 540, + 230, + 614 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/3fd7e43195ea2d83908a43bfe1cbf33c198ad222dc2286e4f11694458be0b892.jpg", + "image_caption": [ + "(b) Pseudo-labels using different Potts relaxation" + ], + "image_footnote": [], + "bbox": [ + 267, + 465, + 413, + 613 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Dataset and evaluation We mainly use the standard PASCAL VOC 2012 dataset [18] and scribble annotations for supervision [28]. The dataset contains 21 classes including background. Following the common practice [11, 37, 38], we use the augmented version with 10,582 training images and 1449 images for validation. Our evaluation metric is the standard mean Intersection-over-Union (mIoU) on the validation set. We also test our method on two additional datasets in Section 3.5. One is Cityscapes [14] which is built for urban scenes and consists of 2975 and 500 fine-labeled images for training and validation. There are 19 out of 30 annotated classes for semantic segmentation. The other one", + "bbox": [ + 89, + 719, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/15211efd019070799ea0e5a19dd39da347feb5c589fa0d89bac72ace454ff750.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
scribble length ratio
00.30.50.81.0
PBL56.4261.7463.8165.7367.24
PNQ59.0165.5367.8070.6371.12
PQ58.9265.3467.8170.4371.05
PCCE56.4061.8263.8165.8167.41
PCD59.0465.5267.8470.9371.22
PLQ59.0365.4467.8170.8071.21
", + "bbox": [ + 545, + 88, + 874, + 210 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 3. Comparison of Potts relaxations with self-labeling. mIoUs on the validation set are shown here.", + "bbox": [ + 511, + 220, + 905, + 250 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "is ADE20k [46] which has 150 fine-grained classes. There are 20210 and 2000, images for training and validation. Instead of scribble-based supervision, we followed [27] to use the block-wise annotation as a form of weak supervision.", + "bbox": [ + 511, + 276, + 906, + 338 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation details We use DeepLabv3+ [13] with two backbones: ResNet101 [21] and MobileNetV2 [36]. We use ResNet101 in Section 3.5, and MobileNetV2 in other sections, for efficiency. We also use Vision Transformer backbone [17], specifically vit-base-patch16-224, and linear decoder in Section 3.5. All backbone networks (ResNet101 and MobileNetV2, ViT) are pre-trained onImagenet [16]. Unless stated explicitly, we use batch 12 as the default across all the experiments. To avoid feeding random predictions from the uninitialized classification head into pseudo-labeling sub-problem, we initialize the network using the cross-entropy on scribbles, i.e. the first term in loss (8), similarly to pre-training in [37]. The network parameters are optimized by SGD. The initial learning rate 0.0007 is scheduled by a polynomial decay of power 0.9. All loss variants are optimized for 60 epochs, but all hyperparameters are tuned separately for each variant. For our best result, we use $\\eta = 0.3$ , $\\lambda = 6$ , $H_{\\mathrm{CCE}}$ and $P_{\\mathrm{CD}}$ . The color-bandwidth in the Potts model is set to 9 across all the experiments on Pascal VOC 2012 and 3 for Cityscapes and ADE20k datasets.", + "bbox": [ + 511, + 354, + 908, + 657 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.1. Comparison of Potts relaxations", + "text_level": 1, + "bbox": [ + 511, + 667, + 795, + 683 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To compare different Potts relaxations within our self-labeling framework, we select $H_{\\mathrm{CCE}}$ as the cross-entropy term in (8), as motivated by the results in Section 3.2. The neighborhood system is the nearest neighbors. The quantitative results are in Table 3. First, one can see that the log-based pairwise terms work consistently better. The likely explanation is that the logarithm addresses the gradient vanishing problem that is obvious in the top row of Figure 2. Moreover, the logarithm encourages smoother transitions across the boundaries (see Figure 4). This may benefit network training when the exact ground truth is unknown. Intuitively, higher uncertainty around object boundaries is natural. Second, the normalized relaxations $P_{\\mathrm{NQ}}$ and $P_{\\mathrm{CD}}$ work better, confirming the limitations of $P_{\\mathrm{BL}}$ and $P_{\\mathrm{Q}}$ discussed in Fig. 1.", + "bbox": [ + 511, + 688, + 908, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "20249", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/3122dae951addc3a14b4403a881c4c59407c40048c29dccd68e9857585801043.jpg", + "image_caption": [ + "Figure 5. Comparison of cross-entropy terms." + ], + "image_footnote": [], + "bbox": [ + 124, + 90, + 446, + 258 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "While we focus on a general soft pseudo-label solver (see Suppl.Mat), our results can motivate specialized solvers. For example, quadratic relaxation $P_{\\mathrm{Q}}$ seems attractive as it defines a convex loss (8). If using \"quadratic cross-entropy\" $H_{\\mathrm{Q}} := \\| y - \\sigma \\|^2$ , then our loss even has a closed-form solution for $y$ [1, 19]. However, the corresponding pseudolabels tend to be overly soft and additional decidiveness (non-convex entropy) term improves mIoU.", + "bbox": [ + 89, + 311, + 483, + 431 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.2. Comparison of cross-entropy terms", + "text_level": 1, + "bbox": [ + 89, + 452, + 398, + 468 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We compare different cross-entropy terms while fixing the pairwise term to $P_{\\mathrm{Q}}$ due to its simplicity. We also use the nearest neighbors. The results are in Figure 5. In agreement with Fig.3(d), $H_{\\mathrm{CE}}$ performs consistently better across different supervision levels, i.e. scribble lengths. Both $H_{\\mathrm{CE}}$ and $H_{\\mathrm{RCE}}$ have a strong margin over the standard entropy $H_{\\mathrm{CE}}$ as network training becomes robust to uncertainty and errors in pseudo labels, see Sec.2.2.", + "bbox": [ + 89, + 477, + 483, + 598 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.3. Comparison of neighborhood systems", + "text_level": 1, + "bbox": [ + 89, + 618, + 416, + 633 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "So far we used the four nearest neighbors (NN) neighborhood for the pairwise Potts term in (8). Here we compare NN with the dense neighborhoods (DN). We use $H_{\\mathrm{CCE}}$ as the cross-entropy term. To optimize pseudo-labels over DN, we still use the general gradient descent technique detailed in the Supplementary Material. The gradient computation for DN Potts can employ the efficient bilateral filtering [37], but it applies only to (proper) second-order variants, i.e. $P_{\\mathrm{BL}}$ and $P_{\\mathrm{Q}}$ . We focus on $P_{\\mathrm{Q}}$ as a preferred option for soft self-labeling. We obtained 71.1% mIoU on nearest neighbors while only getting 67.9% on dense neighborhoods (bandwidth 100). Some qualitative results are shown in Figure 6 indicating that larger neighborhoods induce lower-quality pseudo-labels. A possible explanation is that the Potts model reduces to the cardinality/volume potentials for larger neighborhoods [39]. The nearest neighborhood is better for edge alignment producing better weakly-supervised results.", + "bbox": [ + 89, + 643, + 483, + 900 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c2fc96c343432f316cae67acd322d240a0eaf8957e782a7ae5608eca6dadce14.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 517, + 90, + 586, + 141 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6c12f2614429f36f8305135c3d5b0a4640676ce2a9825d65cfdea7c56e107311.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 517, + 143, + 584, + 191 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c3aa74185d2a26cb9d267b6338643b0b168ed92dffb2105aa6ec35ba70394505.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 517, + 193, + 584, + 242 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/de4e540c40c1915846578c05e73d7a31b9034e2d741e940643660f6da029ff74.jpg", + "image_caption": [ + "(a) Image & scribble" + ], + "image_footnote": [], + "bbox": [ + 517, + 242, + 583, + 292 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/783c1d36fb18304b47965edbe339884ab6af4f3146fa035f1c6d44d47f665884.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 596, + 90, + 663, + 140 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/52dca61eca1fb424f26824ea6883747c61bffe9a0c53a2909c060acacf91789a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 598, + 142, + 663, + 191 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ec7576f8952af1dde748ada0014e74f19bc8c9fef8bc6ed9068f9a3fbc54b5dc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 598, + 193, + 661, + 242 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1344c98ce43af0f8d631c5fa7e20b90492e7219f1be7b9ef243b779b6d0a1ad9.jpg", + "image_caption": [ + "(b) Predictions" + ], + "image_footnote": [], + "bbox": [ + 598, + 243, + 661, + 292 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6c2fd150f8362cd43c01e74163f03d5beecae15b6411dc217162fae0a70c3e5c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 676, + 90, + 743, + 140 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/78b460e16b3841a2e5edce9ec7d7042c8cae0f926b76845d5388a631d4a9a32b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 678, + 142, + 743, + 191 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/978c0339fde0f1cd69429dd2d6b280fb3f870fdd28612c522988713e7b064720.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 678, + 191, + 743, + 242 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9bb6d1ace973dbc963681178c47b3eac5869f3d017614fcb4afb4112e8ca4342.jpg", + "image_caption": [ + "(c) Pseudo-labels for NN" + ], + "image_footnote": [], + "bbox": [ + 678, + 243, + 743, + 292 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/75b68ae4089eaef4fbb7ffc0b693a307cf1605b1310277fc7e35904208d33cee.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 758, + 90, + 823, + 140 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/34471a5c14220d5b5dc38f6c58c39ea282751ee6f35897fe7473020a1325c0ce.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 758, + 142, + 823, + 191 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f6886a149273b14426a08315df22705302beb983fdff2bfbc531d65588610f27.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 758, + 191, + 823, + 242 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/608fd117ab593e33fedd0203d633fc1c654639058b3aa991ef6d702219dc9414.jpg", + "image_caption": [ + "(d) Pseudo-labels for DN (25)", + "Figure 6. Pseudo-labels generated from given network predictions using different neighborhoods: nearest (NN) and dense (DN)." + ], + "image_footnote": [], + "bbox": [ + 758, + 243, + 823, + 292 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/76ed42884b0ee07f8d0b0ecb4f44ca3d1cb6f9bd880fe2c0997203184eb68fbd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 836, + 90, + 901, + 140 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d86a57e2d0a898c322e078ca2d59bb4d264831350558c4b62edae94879dc04b2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 836, + 142, + 901, + 191 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8300a86876e651a6d487845539426df4956c283074bdb6c3d3beaa58c18c68c3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 836, + 191, + 901, + 242 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b7f6bd2a92296f335136138e8df2c47706e4626eafd3dcaac07873e39f92b14e.jpg", + "image_caption": [ + "(e) Pseudo-labels for DN (100)" + ], + "image_footnote": [], + "bbox": [ + 836, + 243, + 901, + 292 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/65de2bad099a03d2e488968bd391a853034cfabaece7a91111b6026639e46ff2.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
N
NNDN
GD67.069.5* [38]
SLhard69.6* [29]63.1 [28]
soft71.167.9
", + "bbox": [ + 545, + 363, + 874, + 429 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. Summary of Potts-based WSSS (full-scribbles, DeepLab V3+). Reproduced results using public code are marked by \\*.", + "bbox": [ + 511, + 438, + 906, + 467 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ba96d6de4fc604fb80670bc7b8c02a3948d3c7a63ae9d7f9d358a4ca53acf655.jpg", + "image_caption": [ + "Figure 7. Comparison of different methods using Potts relaxations. The architecture is DeeplabV3+ with the backbone MobileNetV2." + ], + "image_footnote": [], + "bbox": [ + 565, + 482, + 856, + 616 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3.4. Soft vs. hard self-labeling vs. gradient descent", + "text_level": 1, + "bbox": [ + 511, + 681, + 903, + 698 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Here we review our general observations about the Potts-based WSSS using the scribble-based results summarized in Table 4. First, in the simplest approach using direct optimization of the network parameters w.r.t. Potts-regularized losses like (4) and stochastic gradient descent, one benefits from a larger neighborhood size [38]. They use $P_{\\mathrm{BL}}$ where DN smoothens the Potts model [30] helping SGD to avoid local minima. Higher-order optimization based on auxiliary self-labeling loss (8) significantly benefits from soft pseudo-labels since they can represent uncertainty. In this case (normalized) quadratic Potts relaxations offer two advantages over the tight bilinear relaxation that leads to hard solutions. On the one hand, they do not suppress label soft", + "bbox": [ + 511, + 704, + 908, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "20250", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/6512a3766b86c077a3ecbdb0fafb0323372e28bec1e8f999c68539ddf9fac642.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodArchitectureBatchsizeTraining AlgorithmNmIoU
GDSL
hardsoft
Full supervision: standard architectures + full masks training (NLL loss)
Deeplab [12]V212✓---75.6
Deeplab* [13]V3+12✓---76.6
Deeplab* [13]V3+16✓---78.9
Vision Transformer [17]ViT-linear16✓---81.4
Scribble supervision (full-length scribbles)
Modified architectures (△), and/or ad-hoc self-labeling (○)
BPG [41]V2 (△)10✓---73.2
URSS [33]V2 (△)16✓---74.6
SPML [22]V2 (△)16✓---74.2
PSI [44]V3+ (△)---✓(○)-74.9
SEMINAR [10]V3+ (△)12✓---76.2
TEL [27]V3+16--✓(○)-77.1
AGMM [43]ViT-linear (△)16--✓(○)-78.7
Standard architectures + Potts loss optimization (relaxations and/or self-labeling)
ScribbleSup [28]V2 (VGG16)8-✓-DN63.1
DenseCRF loss* [38]V3+12✓--DN75.8
GridCRF loss* [29]V3+12-✓-NN75.6
NonlocalCRF loss* [40]V3+12✓--SN75.7
HCCE + PQV3+12--✓NN77.5
HCCE + PCDV3+12--✓NN77.7
HCCE + PCD (no pretrain)V3+12--✓NN76.7
HCCE + PCD (no pretrain)V3+16--✓NN77.6
HCCE + PCDV3+16--✓NN78.1
HCCE + PCD (no pretrain)V3+16--✓NN77.6
HCCE + PCD (no pretrain)ViT-linear12--✓NN80.8
HCCE + PCD (no pretrain)ViT-linear16--✓NN80.94
", + "bbox": [ + 155, + 87, + 844, + 462 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. Comparison of scribble-supervised segmentation methods (without CRF postprocessing). We focus on general solutions applicable to any standard architecture based on mathematically well-founded convergent training procedures optimizing a clearly defined loss (the last block). The second block shows specialized solutions based on architectural modifications and/or ad-hoc training procedures that are often non-convergent and hard to reproduce. The numbers are mIoU on the validation dataset of Pascal VOC 2012. The backbone is ResNet101 unless stated otherwise. V2: deeplabV2. V3+: deeplabV3+. ViT-linear: ViT backbone and linear decoder. $\\mathcal{N}$ : neighborhood. “*”: reproduced results. GD: gradient descent. SL: self-labeling. “no pretrain” means the segmentation network is not pretrained using cross-entropy on scribbles. In some cases (V3+, batch size 12) soft self-labeling outperforms full supervision (see teal color numbers).", + "bbox": [ + 88, + 472, + 906, + 571 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ness; this softness is even formally related to classification uncertainty in the random walker method [19]. On the other hand, the convexity of quadratic relaxation does not require DN to simplify loss for gradient descent. Then, it is not surprising that NN works better than DN. The latter makes the Potts model behave as a cardinality potential [39], while in the case of NN the Potts model directly regularized the geometric boundary of objects [6]. Figure 7 also compares these approaches across different scribble lengths.", + "bbox": [ + 88, + 590, + 480, + 727 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3.5. Comparison to SOTA", + "text_level": 1, + "bbox": [ + 89, + 739, + 292, + 757 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This section uses DeepLabV3+ with ResNet101 backbone and ViT-linear transformer to evaluate our self-labeling methodology on these standard architectures and to compare with representative SOTA for scribble-based segmentation. Table 5 presents only the results before any post-processing. Our method may outperform the fully-supervised method (teal color numbers for DeepLabV3+ and batch size 12). Our \"text-book\" training technique, i.e. minimization of a well-defined loss, applied to standard architectures out-", + "bbox": [ + 88, + 763, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "performs complex WSSS systems requiring architectural modifications and/or ad-hoc multi-stage training procedures.", + "bbox": [ + 511, + 590, + 906, + 621 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4. Conclusions", + "text_level": 1, + "bbox": [ + 511, + 633, + 638, + 648 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper proposes a mathematically well-founded self-labeling WSSS framework for standard segmentation networks based on a joint loss (8) w.r.t. network predictions and soft pseudo-labels. The latter was motivated as auxiliary variables simplifying optimization of (4). Conceptual properties and systematic evaluation of the terms in our soft self-labeling loss (8) advocate for the collision cross-entropy and collision divergence. The former replaces the standard cross-entropy $H$ , and the latter is used as a log of a normalized quadratic relaxation of the Potts model $P$ over the NN neighborhood. Our general methodology works with any standard architecture and outperforms complex specialized systems based on architectural modifications and/or ad-hoc multi-stage non-convergent training. In contrast, our ideas are easy to understand, reproduce, and extend to other forms of weakly-supervised segmentation (boxes, class tags, etc.).", + "bbox": [ + 509, + 657, + 908, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "20251", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Multilabel random walker image segmentation using prior models. In 2005 IEEE computer society conference on computer vision and pattern recognition (CVPR'05), pages 763-770. IEEE, 2005. 7", + "[2] Jiwoon Ahn and Suha Kwak. Learning pixel-level semantic affinity with image-level supervision for weakly supervised semantic segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4981-4990, 2018. 2", + "[3] Jiwoon Ahn, Sunghyun Cho, and Suha Kwak. Weakly supervised learning of instance segmentation with inter-pixel relations. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2209-2218, 2019. 2", + "[4] Nikita Araslanov and Stefan Roth. Single-stage semantic segmentation from image labels. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4253-4262, 2020. 1", + "[5] Stephen Boyd and Lieven Vandenberghe. Convex optimization. Cambridge university press, 2004. 3", + "[6] Yuri Boykov and Vladimir Kolmogorov. Computing geodesics and minimal surfaces via graph cuts. In International conference on computer vision (ICCV), 2003. 3, 8", + "[7] Yuri Y Boykov and M-P Jolly. Interactive graph cuts for optimal boundary & region segmentation of objects in nd images. In international conference on computer vision (ICCV), pages 105-112, 2001. 2, 3", + "[8] John Bridle, Anthony Heading, and David MacKay. Unsupervised classifiers, mutual information and phantom targets. Advances in neural information processing systems, 4, 1991. 2", + "[9] Antonin Chambolle and Thomas Pock. A first-order primal-dual algorithm for convex problems with applications to imaging. Journal of mathematical imaging and vision, 40:120-145, 2011. 2", + "[10] Hongjun Chen, Jinbao Wang, Hong Cai Chen, Xiantong Zhen, Feng Zheng, Rongrong Ji, and Ling Shao. Seminar learning for click-level weakly supervised semantic segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6920-6929, 2021. 2, 8", + "[11] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L Yuille. Semantic image segmentation with deep convolutional nets and fully connected crfs. arXiv preprint arXiv:1412.7062, 2014. 6", + "[12] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L Yuille. Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. IEEE transactions on pattern analysis and machine intelligence, 40(4):834-848, 2017. 8", + "[13] Liang-Chieh Chen, Yukun Zhu, George Papandreou, Florian Schroff, and Hartwig Adam. Encoder-decoder with atrous separable convolution for semantic image segmentation. In Proceedings of the European conference on computer vision (ECCV), pages 801-818, 2018. 6, 8" + ], + "bbox": [ + 93, + 114, + 485, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 6", + "[15] Camille Couprie, Leo Grady, Laurent Najman, and Hugues Talbot. Power watershed: A unifying graph-based optimization framework. IEEE transactions on pattern analysis and machine intelligence, 33(7):1384-1399, 2010. 2", + "[16] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248–255. IEEE, 2009. 6", + "[17] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 6, 8", + "[18] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. International journal of computer vision, 88:303-308, 2009. 6", + "[19] Leo Grady. Random walks for image segmentation. IEEE transactions on pattern analysis and machine intelligence, 28 (11):1768-1783, 2006. 2, 3, 7, 8", + "[20] Yves Grandvalet and Yoshua Bengio. Semi-supervised learning by entropy minimization. Advances in neural information processing systems, 17, 2004. 2", + "[21] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6", + "[22] Tsung-Wei Ke, Jyh-Jing Hwang, and Stella X Yu. Universal weakly supervised segmentation by pixel-to-segment contrastive learning. arXiv preprint arXiv:2105.00957, 2021. 2, 8", + "[23] Alexander Kolesnikov and Christoph H Lampert. Seed, expand and constrain: Three principles for weakly-supervised image segmentation. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part IV 14, pages 695-711. Springer, 2016. 1", + "[24] Philipp Krahenbuhl and Vladlen Koltun. Efficient inference in fully connected CRFs with Gaussian edge potentials. Advances in neural information processing systems, 24, 2011. 2", + "[25] V. Kulharia, S. Chandra, A. Agrawal, P. Torr, and A. Tyagi. Box2seg: Attention weighted loss and discriminative feature learning for weakly supervised segmentation. In ECCV'20. 1", + "[26] Jungbeom Lee, Eunji Kim, Sungmin Lee, Jangho Lee, and Sungroh Yoon. Ficklenet: Weakly and semi-supervised semantic image segmentation using stochastic inference. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5267-5276, 2019. 2", + "[27] Zhiyuan Liang, Tiancai Wang, Xiangyu Zhang, Jian Sun, and Jianbing Shen. Tree energy loss: Towards sparsely annotated semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16907-16916, 2022. 2, 3, 6, 8" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "20252", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[28] Di Lin, Jifeng Dai, Jiaya Jia, Kaiming He, and Jian Sun. Scribblesup: Scribble-supervised convolutional networks for semantic segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3159-3167, 2016. 1, 2, 6, 7, 8", + "[29] Dmitrii Marin and Yuri Boykov. Robust trust region for weakly supervised segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6608-6618, 2021. 2, 7, 8", + "[30] Dmitrii Marin, Meng Tang, Ismail Ben Ayed, and Yuri Boykov. Beyond gradient descent for regularized segmentation losses. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10187-10196, 2019. 2, 3, 7", + "[31] Rafael Müller, Simon Kornblith, and Geoffrey E Hinton. When does label smoothing help? Advances in neural information processing systems, 32, 2019. 5", + "[32] NSD. Natural Scenes Dataset [NSD]. https://www.kaggle.com/datasets/nitishabharathi/scene-classification, 2020.5", + "[33] Zhiyi Pan, Peng Jiang, Yunhai Wang, Changhe Tu, and Anthony G Cohn. Scribble-supervised semantic segmentation by uncertainty reduction on neural representation and self-supervision on neural eigenspace. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7416-7425, 2021. 2, 8", + "[34] Thomas Pock, Antonin Chambolle, Daniel Cremers, and Horst Bischof. A convex relaxation approach for computing minimal partitions. In IEEE Conference on Computer Vision and Pattern Recognition, pages 810-817, 2009. 2", + "[35] Pradeep Ravikumar and John Lafferty. Quadratic programming relaxations for metric labeling and Markov Random Field MAP estimation. In The 23rd International Conference on Machine Learning, page 737-744, 2006. 2, 3", + "[36] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zhoginov, and Liang-Chieh Chen. *Mobilenetv2: Inverted residuals and linear bottlenecks*. In *Proceedings of the IEEE conference on computer vision and pattern recognition*, pages 4510–4520, 2018. 6", + "[37] Meng Tang, Abdelaziz Djelouah, Federico Perazzi, Yuri Boykov, and Christopher Schroers. Normalized cut loss for weakly-supervised cnn segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1818-1827, 2018. 1, 6, 7", + "[38] Meng Tang, Federico Perazzi, Abdelaziz Djelouah, Ismail Ben Ayed, Christopher Schroers, and Yuri Boykov. On regularized losses for weakly-supervised cnn segmentation. In Proceedings of the European Conference on Computer Vision (ECCV), pages 507-522, 2018. 1, 2, 6, 7, 8", + "[39] Olga Veksler. Efficient graph cut optimization for full crfs with quantized edges. IEEE transactions on pattern analysis and machine intelligence, 42(4):1005-1012, 2019. 7, 8", + "[40] Olga Veksler and Yuri Boykov. Sparse non-local crf. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4493-4503, 2022. 2, 8", + "[41] Bin Wang, Guojun Qi, Sheng Tang, Tianzhu Zhang, Yunchao Wei, Linghui Li, and Yongdong Zhang. Boundary perception" + ], + "bbox": [ + 91, + 92, + 485, + 902 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "guidance: A scribble-supervised semantic segmentation approach. In IJCAI International joint conference on artificial intelligence, 2019. 2, 8", + "[42] Changwei Wang, Rongtao Xu, Shibiao Xu, Weiliang Meng, and Xiaopeng Zhang. Treating pseudo-labels generation as image matting for weakly supervised semantic segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 755-765, 2023. 2", + "[43] Linshan Wu, Zhun Zhong, Jiayi Ma, Yunchao Wei, Hao Chen, Leyuan Fang, and Shutao Li. Modeling the label distributions for weakly-supervised semantic segmentation. arXiv preprint arXiv:2403.13225, 2024. 8", + "[44] Jingshan Xu, Chuanwei Zhou, Zhen Cui, Chunyan Xu, Yuge Huang, Pengcheng Shen, Shaoxin Li, and Jian Yang. Scribblesupervised semantic segmentation inference. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15354-15363, 2021. 2, 3, 8", + "[45] Christopher Zach, Christian Hane, and Marc Pollefeys. What is optimized in tight convex relaxations for multi-label problems? In IEEE Conference on Computer Vision and Pattern Recognition, pages 1664-1671, 2012. 2", + "[46] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 633-641, 2017. 6" + ], + "bbox": [ + 516, + 92, + 906, + 445 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "20253", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Soft Self-labeling and Potts Relaxations for Weakly-supervised Segmentation/4b6c17bc-bca2-47a9-b3a7-cd9f02074980_model.json b/2025/Soft Self-labeling and Potts Relaxations for Weakly-supervised Segmentation/4b6c17bc-bca2-47a9-b3a7-cd9f02074980_model.json new file mode 100644 index 0000000000000000000000000000000000000000..85dae72a62c2bbd99ec2e0ba6936e288282b2e9d --- /dev/null +++ b/2025/Soft Self-labeling and Potts Relaxations for Weakly-supervised Segmentation/4b6c17bc-bca2-47a9-b3a7-cd9f02074980_model.json @@ -0,0 +1,2629 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.299, + 0.131, + 0.7, + 0.177 + ], + "angle": 0, + "content": "Soft Self-labeling and Potts Relaxations for Weakly-Supervised Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.35, + 0.204, + 0.646, + 0.222 + ], + "angle": 0, + "content": "Zhongwen Zhang & Yuri Boykov" + }, + { + "type": "text", + "bbox": [ + 0.375, + 0.223, + 0.625, + 0.239 + ], + "angle": 0, + "content": "University of Waterloo, Canada" + }, + { + "type": "text", + "bbox": [ + 0.306, + 0.243, + 0.688, + 0.256 + ], + "angle": 0, + "content": "z889zhan@uwaterloo.ca, yboykov@uwaterloo.ca" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.292, + 0.327, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.325, + 0.486, + 0.628 + ], + "angle": 0, + "content": "We consider weakly supervised segmentation where only a fraction of pixels have ground truth labels (scribbles) and focus on a self-labeling approach optimizing relaxations of the standard unsupervised CRF/Potts loss on unlabeled pixels. While WSSS methods can directly optimize such losses via gradient descent, prior work suggests that higher-order optimization can improve network training by introducing hidden pseudo-labels and powerful CRF sub-problem solvers, e.g. graph cut. However, previously used hard pseudo-labels can not represent class uncertainty or errors, which motivates soft self-labeling. We derive a principled auxiliary loss and systematically evaluate standard and new CRF relaxations (convex and non-convex), neighborhood systems, and terms connecting network predictions with soft pseudolabels. We also propose a general continuous sub-problem solver. Using only standard architectures, soft self-labeling consistently improves scribble-based training and outperforms significantly more complex specialized WSSS systems. It can outperform full pixel-precise supervision. Our general ideas apply to other weakly-supervised problems/systems." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.631, + 0.48, + 0.645 + ], + "angle": 0, + "content": "Keywords: soft pseudo-labels, collision cross-entropy, collision divergence" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.678, + 0.222, + 0.694 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.704, + 0.486, + 0.902 + ], + "angle": 0, + "content": "Full supervision for semantic segmentation requires thousands of training images with complete pixel-accurate ground truth masks. Their high costs explain the interest in weakly-supervised approaches based on image-level class tags [4, 23], pixel-level scribbles [28, 37, 38], or boxes [25]. This paper is focused on weak supervision with scribbles, which we also call seeds or partial masks. While only slightly more expensive than image-level class tags, scribbles on less than \\(3\\%\\) of pixels were previously shown to achieve accuracy approaching full supervision without any modifications of the segmentation models. In contrast, tag supervision typically requires highly specialized systems and complex multi-stage training procedures, which are hard" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.293, + 0.908, + 0.4 + ], + "angle": 0, + "content": "to reproduce. Our interest in the scribble-based approach is motivated by its practical simplicity and mathematical clarity. The corresponding methodologies are focused on the design of unsupervised or self-supervised loss functions and stronger optimization algorithms. The corresponding solutions are often general and can be used in different weakly-supervised applications." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.408, + 0.809, + 0.423 + ], + "angle": 0, + "content": "1.1. Scribble-supervised segmentation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.43, + 0.908, + 0.505 + ], + "angle": 0, + "content": "Let \\(\\Omega\\) denote a set of image pixels and \\(S\\subset \\Omega\\) denote a subset of pixels with ground truth labels, i.e. seeds or scribbles typically marked by mouse-controlled UI for image annotations. One example of image scribbles is in Fig.4(a). The ground truth label at any given pixel \\(i\\in S\\) is an integer" + }, + { + "type": "equation", + "bbox": [ + 0.652, + 0.515, + 0.907, + 0.532 + ], + "angle": 0, + "content": "\\[\n\\bar {y} _ {i} \\in \\{1, \\dots , K \\} \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.541, + 0.908, + 0.587 + ], + "angle": 0, + "content": "where \\(K\\) is the number of classes including the background. Without much ambiguity, it is convenient to use the same notation \\(\\bar{y}_i\\) for the equivalent one-hot distribution" + }, + { + "type": "equation", + "bbox": [ + 0.523, + 0.595, + 0.907, + 0.614 + ], + "angle": 0, + "content": "\\[\n\\bar {y} _ {i} \\equiv (\\bar {y} _ {i} ^ {1}, \\dots , \\bar {y} _ {i} ^ {K}) \\in \\Delta_ {0, 1} ^ {K} \\quad \\text {f o r} \\quad \\bar {y} _ {i} ^ {k} := [ k = \\bar {y} _ {i} ] \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.623, + 0.907, + 0.67 + ], + "angle": 0, + "content": "where \\([\\cdot ]\\in \\{0,1\\}\\) is the True operator. Set \\(\\Delta_{0,1}^{K}\\) represents \\(K\\) possible one-hot distributions, which are vertices of the \\(K\\)-class probability simplex" + }, + { + "type": "equation", + "bbox": [ + 0.534, + 0.678, + 0.886, + 0.721 + ], + "angle": 0, + "content": "\\[\n\\Delta^ {K} := \\{p = (p ^ {1}, \\dots , p ^ {K}) \\mid p ^ {k} \\geq 0, \\sum_ {k = 1} ^ {K} p ^ {k} = 1 \\}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.728, + 0.908, + 0.773 + ], + "angle": 0, + "content": "representing all \\(K\\)-categorical distributions. The context of each specific expression should make it obvious if \\(\\bar{y}_i\\) is a class index (1) or the corresponding one-hot distribution (2)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.774, + 0.908, + 0.819 + ], + "angle": 0, + "content": "Loss functions for scribble-supervised segmentation typically use negative log-likelihoods (NLL) over scribbles \\( i \\in S \\subset \\Omega \\) with ground truth labels \\( \\bar{y}_i \\)" + }, + { + "type": "equation", + "bbox": [ + 0.668, + 0.828, + 0.907, + 0.86 + ], + "angle": 0, + "content": "\\[\n- \\sum_ {i \\in S} \\ln \\sigma_ {i} ^ {\\bar {y} _ {i}} \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.87, + 0.908, + 0.902 + ], + "angle": 0, + "content": "where \\(\\sigma_{i} = (\\sigma_{i}^{1},\\dots,\\sigma_{i}^{K})\\in \\Delta^{K}\\) are model predictions at pixels \\(i\\). This loss is also standard in full supervision" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "20244" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.152 + ], + "angle": 0, + "content": "where \\(S = \\Omega\\). In this case, no other losses are necessary for training. However, in a weakly supervised setting the majority of pixels are unlabeled, and unsupervised losses are needed for \\(i \\notin S\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.153, + 0.483, + 0.273 + ], + "angle": 0, + "content": "The most common unsupervised loss in image segmentation is the Potts model and its relaxations. It is a pairwise loss defined on pairs of neighboring pixels \\(\\{i,j\\} \\in \\mathcal{N}\\) for a given neighborhood system \\(\\mathcal{N} \\subset \\Omega \\times \\Omega\\), typically corresponding to the nearest-neighbor grid (NN) [7, 19], or other sparse (SN) [40] or dense neighborhoods (DN) [24]. The original Potts model is defined for discrete segmentation variables, e.g. as in" + }, + { + "type": "equation", + "bbox": [ + 0.11, + 0.28, + 0.466, + 0.315 + ], + "angle": 0, + "content": "\\[\n\\sum_ {\\{i, j \\} \\in \\mathcal {N}} P (\\sigma_ {i}, \\sigma_ {j}) \\quad \\text {w h e r e} \\quad P (\\sigma_ {i}, \\sigma_ {j}) = [ \\sigma_ {i} \\neq \\sigma_ {j} ]\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.322, + 0.484, + 0.368 + ], + "angle": 0, + "content": "assuming integer-valued one-hot predictions \\(\\sigma_{i}\\in \\Delta_{0,1}^{K}\\) This regularization loss encourages smoothness between the pixels. Its popular self-supervised variant is" + }, + { + "type": "equation", + "bbox": [ + 0.192, + 0.376, + 0.382, + 0.393 + ], + "angle": 0, + "content": "\\[\nP \\left(\\sigma_ {i}, \\sigma_ {j}\\right) = w _ {i, j} \\cdot \\left[ \\sigma_ {i} \\neq \\sigma_ {j} \\right]\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.4, + 0.484, + 0.49 + ], + "angle": 0, + "content": "where pairwise affinities \\( w_{ij} \\) are based on local intensity edges [7, 19, 24]. Of course, in the context of network training, one should use relaxations of \\( P \\) applicable to soft predictions \\( \\sigma_i \\in \\Delta^K \\). Many forms of such relaxations [35, 45] were studied in segmentation, e.g. quadratic [19], bi-linear [38], total variation [9, 34], and others [15]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.491, + 0.484, + 0.535 + ], + "angle": 0, + "content": "Another general unsupervised loss relevant to training segmentation networks is the entropy of predictions, which is also known as decidiveness [8, 20]" + }, + { + "type": "equation", + "bbox": [ + 0.25, + 0.542, + 0.324, + 0.573 + ], + "angle": 0, + "content": "\\[\n\\sum_ {i} H (\\sigma_ {i})\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.58, + 0.484, + 0.685 + ], + "angle": 0, + "content": "where \\( H \\) is the Shannon's entropy function. This loss is known to improve generalization and the quality of representation by moving (deep) features away from the decision boundaries [20]. Widely known in the context of unsupervised or semi-supervised classification, this loss also matters in weakly-supervised segmentation where it is used explicitly or implicitly1." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.686, + 0.484, + 0.777 + ], + "angle": 0, + "content": "Weakly-supervised segmentation methods [10, 22, 33, 41] also use other unsupervised losses (e.g. contrastive), clustering criteria (e.g. K-means), highly-specialized architectures, and ad-hoc training procedures. Our goal is to demonstrate that better results can be easily achieved on standard architectures by combining the standard losses above" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.798, + 0.484, + 0.817 + ], + "angle": 0, + "content": "\\[\n- \\sum_ {i \\in S} \\ln \\sigma_ {i} ^ {\\bar {y} _ {i}} + \\eta \\sum_ {i \\notin S} H (\\sigma_ {i}) + \\lambda \\sum_ {i j \\in \\mathcal {N}} P (\\sigma_ {i}, \\sigma_ {j}). \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.823, + 0.483, + 0.868 + ], + "angle": 0, + "content": "but replacing basic gradient descent optimization [38, 40] by higher-order self-labeling techniques [28-30] incorporating optimization of auxiliary pseudo-labels as sub-problems." + }, + { + "type": "title", + "bbox": [ + 0.515, + 0.091, + 0.829, + 0.108 + ], + "angle": 0, + "content": "1.2. Self-labeling and hard pseudo-labels" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.113, + 0.907, + 0.188 + ], + "angle": 0, + "content": "One motivation for the self-labeling approach to weakly-supervised segmentation comes from well-known limitations of gradient descent for optimizing the Potts relaxations. Even convex relaxations [9, 19, 34] could be challenging in combination with the concave entropy term in (4)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.189, + 0.908, + 0.354 + ], + "angle": 0, + "content": "Typical self-labeling methods, including one of the first works on scribble-based semantic segmentation [28], introduce a sub-problem focused on the estimation of pseudolabels for unlabeled points. This subproblem can be formally \"derived\" from the original network optimization problem or \"designed\" as a heuristic, see the overview below. We focus on principled auxiliary loss formulations where network parameters and pseudo-labels are jointly optimized in a provably convergent manner. In this case pseudo-labels work as hidden optimization variables simplifying the model estimation, e.g. as in the EM algorithm for GMM." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.355, + 0.907, + 0.446 + ], + "angle": 0, + "content": "We denote pseudo-labels \\( y_{i} \\) differently from the ground truth labels \\( \\bar{y}_{i} \\) by omitting the bar. It is important to distinguish them since \\( \\bar{y}_{i} \\) for \\( i \\in S \\) are given, while \\( y_{i} \\) for \\( i \\in \\Omega \\backslash S \\) are estimated. The majority of existing self-labeling methods [2, 3, 26, 28-30, 42] estimate hard pseudo-labels, which could be equivalently represented either by class indices" + }, + { + "type": "equation", + "bbox": [ + 0.652, + 0.454, + 0.907, + 0.47 + ], + "angle": 0, + "content": "\\[\ny _ {i} \\in \\{1, \\dots , K ^ {\\prime} \\} \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.477, + 0.889, + 0.492 + ], + "angle": 0, + "content": "or by the corresponding one-hot categorical distributions" + }, + { + "type": "equation", + "bbox": [ + 0.523, + 0.499, + 0.907, + 0.517 + ], + "angle": 0, + "content": "\\[\ny _ {i} \\equiv \\left(y _ {i} ^ {1}, \\dots , y _ {i} ^ {K}\\right) \\in \\Delta_ {0, 1} ^ {K} \\quad \\text {f o r} \\quad y _ {i} ^ {k} := [ k = y _ {i} ] \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.524, + 0.907, + 0.613 + ], + "angle": 0, + "content": "analogously with the hard ground truth labels in (1) and (2). In part, hard pseudo-labels are motivated by the network training where the default is NLL loss (3) assuming discrete labels. Besides, powerful discrete solvers for the Potts model are well-known [7, 9, 34]. We discuss the potential advantages of soft pseudo-labels in Section 1.3." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.614, + 0.908, + 0.765 + ], + "angle": 0, + "content": "Overview: joint loss vs \"proposal generation\": Most self-labeling approaches can be divided into two groups. One group designs pseudo-labeling and the network training sub-problems that are not formally related, e.g. [27, 28, 44]. While pseudo-labeling typically depends on the current network predictions and the network fine-tuning uses such pseudo-labels, the lack of a formal relation between these sub-problems implies that iterating such steps does not guarantee any form of convergence. Such methods are often referred to as proposal generation heuristics." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.766, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Alternatively, the pseudo-labeling sub-problem and the network training sub-problem can be formally derived from a weakly-supervised loss like (4), e.g. by ADM splitting [30] or as high-order trust-region method [29]. Such methods often formulate a joint loss w.r.t. network predictions and pseudo-labels and iteratively decrease the loss. These self-labeling methods are mathematically well-founded, easy to understand and reproduce, and numerically stable due to guaranteed convergence." + }, + { + "type": "page_footnote", + "bbox": [ + 0.09, + 0.875, + 0.483, + 0.901 + ], + "angle": 0, + "content": "1A unary decidiveness-like term is the difference between quadratic and bi-linear (tight but non-convex) relaxations [29, 35] of the Potts model." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "20245" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.485, + 0.108 + ], + "angle": 0, + "content": "1.3. Soft pseudo-labels: motivation & contributions" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.113, + 0.486, + 0.175 + ], + "angle": 0, + "content": "We observe that self-labeling with hard pseudo-labels \\( y_{i} \\) is inherently limited as such labels cannot represent the uncertainty of class estimates at unlabeled pixels \\( i \\in \\Omega \\backslash S \\). Instead, we focus on soft pseudo-labels" + }, + { + "type": "equation", + "bbox": [ + 0.19, + 0.184, + 0.484, + 0.203 + ], + "angle": 0, + "content": "\\[\ny _ {i} = \\left(y _ {i} ^ {1}, \\dots , y _ {i} ^ {K}\\right) \\in \\Delta^ {K} \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.213, + 0.485, + 0.333 + ], + "angle": 0, + "content": "which are general categorical distributions \\( p \\) over \\( K \\)-classes. It is possible that the estimated pseudo-label \\( y_{i} \\) in (7) could be a one-hot distribution, which is a vertex of \\( \\Delta^K \\). In this special case, \\( y_{i} \\) can be treated as a class index, but we use the general formulation (7) in the main parts of the paper starting Section 2. On the other hand, the ground truth labels \\( \\bar{y}_{i} \\) are always hard and we use them either as indices (1) or one-hot distributions (2), as convenient." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.334, + 0.485, + 0.59 + ], + "angle": 0, + "content": "Soft pseudo-labels can be found in prior work on weakly-supervised segmentation [27, 44] using the \"soft proposal generation\". In contrast, we formulate soft self-labeling as a principled optimization methodology where network predictions and soft pseudo-labels are variables in a joint loss, which guarantees convergence of the training procedure. Our pseudo-labels are auxiliary variables for ADM-based [5] splitting of the loss (4) into two simpler optimization sub-problems: one focused on the Potts model over unlabeled pixels, and the other on the network training. While similar to [30], instead of hard, we use soft auxiliary variables for the Potts sub-problem. Our work can be seen as a study of the relaxed Potts sub-problem in the context of weakly-supervised semantic segmentation. The related prior work is focused on discrete solvers fundamentally unable to represent class estimate uncertainty. Our contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.591, + 0.483, + 0.62 + ], + "angle": 0, + "content": "- convergent soft self-labeling framework based on a simple joint self-labeling loss" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.621, + 0.484, + 0.65 + ], + "angle": 0, + "content": "- systematic evaluation of Potts relaxations and (cross-) entropy terms in our loss" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.652, + 0.484, + 0.711 + ], + "angle": 0, + "content": "- state-of-the-art in scribble-based semantic segmentation based on standard network architectures (no modifications) and well-founded convergent training procedures that are easy to reproduce and generalize to other problems" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.712, + 0.483, + 0.756 + ], + "angle": 0, + "content": "- using the same segmentation model, our self-labeling loss with \\(3\\%\\) scribbles may outperform standard supervised cross-entropy loss with full ground truth masks." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.591, + 0.484, + 0.756 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.77, + 0.377, + 0.788 + ], + "angle": 0, + "content": "2. Our soft self-labeling approach" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.486, + 0.902 + ], + "angle": 0, + "content": "First, we apply ADM splitting [5] to weakly supervised loss (4) to formulate our self-labeling loss (8) incorporating additional soft auxiliary variables, i.e. pseudo-labels (7). It is convenient to introduce pseudo-labels \\( y_{i} \\) on all pixels in \\( \\Omega \\) even though a subset of pixels (seeds) \\( S\\subset \\Omega \\) have ground truth labels \\( \\bar{y}_i \\). We will simply impose a constraint that pseudo-labels and ground truth labels agree on \\( S \\). Thus," + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.816, + 0.107 + ], + "angle": 0, + "content": "we assume the following set of pseudo-labels" + }, + { + "type": "equation", + "bbox": [ + 0.545, + 0.116, + 0.874, + 0.134 + ], + "angle": 0, + "content": "\\[\nY _ {\\Omega} := \\left\\{y _ {i} \\in \\Delta^ {K} \\mid i \\in \\Omega , \\text {s . t .} y _ {i} = \\bar {y} _ {i} \\text {f o r} i \\in S \\right\\}.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.143, + 0.907, + 0.219 + ], + "angle": 0, + "content": "We split the terms in (4) into two groups: one includes NLL and entropy \\( H \\) terms keeping the original prediction variables \\( \\sigma_{i} \\) and the other includes the Potts relaxation \\( P \\) replacing \\( \\sigma_{i} \\) with auxiliary variables \\( y_{i} \\). This transforms loss (4) into expression" + }, + { + "type": "equation", + "bbox": [ + 0.545, + 0.228, + 0.877, + 0.262 + ], + "angle": 0, + "content": "\\[\n- \\sum_ {i \\in S} \\ln \\sigma_ {i} ^ {\\bar {y} _ {i}} + \\eta \\sum_ {i \\not \\in S} H (\\sigma_ {i}) + \\lambda \\sum_ {i j \\in \\mathcal {N}} P (y _ {i}, y _ {j})\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.27, + 0.907, + 0.336 + ], + "angle": 0, + "content": "equivalent to (4) assuming equality \\(\\sigma_{i} = y_{i}\\). The standard approximation is to incorporate constraint \\(\\sigma_{i} \\approx y_{i}\\) directly into the loss, e.g. using \\(KL\\)-divergence. For simplicity, we use weight \\(\\eta\\) for \\(KL(\\sigma_{i},y_{i})\\) to combine it with \\(H(\\sigma_{i})\\) into a single cross-entropy term" + }, + { + "type": "equation", + "bbox": [ + 0.515, + 0.339, + 0.907, + 0.406 + ], + "angle": 0, + "content": "\\[\n-\\sum_{i\\in S}\\ln \\sigma_{i}^{\\bar{y}_{i}} + \\underbrace{\\eta\\sum_{i\\notin S}H(\\sigma_{i}) + \\eta\\sum_{i\\notin S}KL(\\sigma_{i},y_{i})}_{\\substack{\\eta \\sum_{i\\notin S}H(\\sigma_{i},y_{i})}} + \\lambda \\sum_{ij\\in \\mathcal{N}}P(y_{i},y_{j})\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.411, + 0.887, + 0.427 + ], + "angle": 0, + "content": "defining auxiliary self-labeling loss w.r.t. both \\(\\sigma_{i}\\) and \\(y_{i}\\)" + }, + { + "type": "equation", + "bbox": [ + 0.524, + 0.436, + 0.907, + 0.455 + ], + "angle": 0, + "content": "\\[\n- \\sum_ {i \\in S} \\ln \\sigma_ {i} ^ {\\bar {y} _ {i}} + \\eta \\sum_ {i \\notin S} H \\left(\\sigma_ {i}, y _ {i}\\right) + \\lambda \\sum_ {i j \\in \\mathcal {N}} P \\left(y _ {i}, y _ {j}\\right) \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.463, + 0.874, + 0.478 + ], + "angle": 0, + "content": "approximating the original weakly supervised loss (4)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.479, + 0.907, + 0.583 + ], + "angle": 0, + "content": "Iterative minimization of this loss w.r.t. predictions \\(\\sigma_{i}\\) (model parameters training) and pseudo-labels \\(y_{i}\\) effectively breaks the original optimization problem for (4) into two simpler sub-problems, assuming there is a good solver for optimal pseudo-labels. The latter is plausible since the unary term \\(H(\\sigma_{i}, y_{i})\\) is convex for \\(y_{i}\\), and the Potts model (incl. convex and non-convex relaxations) is widely studied." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.584, + 0.909, + 0.705 + ], + "angle": 0, + "content": "Section 2.1 discusses standard and new relaxations of the Potts model \\(P\\). Section 2.2 discusses several robust variants of the cross-entropy term \\(H\\) in (8) connecting predictions with uncertain (soft) pseudo-labels \\(y_{i}\\) at unlabeled points \\(i \\in \\Omega \\backslash S\\). To systematically evaluate these variants of (8) in Section 3, we propose a general efficient GPU-friendly solver for the corresponding pseudo-labeling sub-problem, which is detailed in the Supplementary Material." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.713, + 0.885, + 0.728 + ], + "angle": 0, + "content": "2.1. Second-order relaxations of the Potts model" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.909, + 0.902 + ], + "angle": 0, + "content": "We limit our focus to basic second-order relaxations and several motivated extensions. We start from two important standard cases (see Table 1): quadratic, the simplest convex relaxation popularized by the random walker algorithm [19], and bi-linear, which is non-convex but tight [35] w.r.t. the original discrete Potts model. The latter implies that optimizing it over relaxed variables should lead to a solution consistent with a discrete Potts solver, e.g. graph cut [7] regularizing the geometric boundary of the segments [6]. The quadratic relaxation produces a significantly different soft solution motivated by random walk probabilities [19]." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "20246" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.127, + 0.088, + 0.452, + 0.161 + ], + "angle": 0, + "content": "
bi-linear ~ “graph cut”PBL(p,q) := 1 - pTqquadratic ~ “random walker”P0(p,q) := 1/2||p-q|2
normalized quadratic
PNQ(p,q) := 1 - pTq/||p||q|≡ 1/2||p/||p-| - q/||q| |2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.12, + 0.172, + 0.454, + 0.186 + ], + "angle": 0, + "content": "Table 1. Second-order Potts relaxations, see Fig.2(a,b,c)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.197, + 0.484, + 0.346 + ], + "angle": 0, + "content": "While there are specialized solvers for bi-linear and quadratic relaxations, we are interested in a general efficient pseudo-label solver integrated into the standard GPU-based network optimization framework. This suggests gradient descent as an optimizer for pseudo-labels \\( y \\) (see Suppl. Mat). Figure 1 shows two representative local minima issues that gradient descent may encounter for (a) the bi-linear and (b) quadratic relaxations of the Potts loss. These examples indicate that gradient descent for these relaxations can distort the properties of the original (discrete) Potts loss." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.349, + 0.484, + 0.456 + ], + "angle": 0, + "content": "In (a) two neighboring pixels jointly change the common label from \\( y_{i} = y_{j} = (1,0,0) \\) to \\( y_{i}^{\\prime \\prime} = y_{j}^{\\prime \\prime} = (0,1,0) \\), which corresponds to a \"move\" where the whole object is reclassified from A to B. This move does not violate the smoothness constraint represented by the original (discrete) Potts model. But, bilinear relaxation prevents move (a) since the intermediate state \\( y_{i}^{\\prime} = y_{j}^{\\prime} = (\\frac{1}{2},\\frac{1}{2},0) \\) has a higher cost" + }, + { + "type": "equation", + "bbox": [ + 0.113, + 0.467, + 0.462, + 0.497 + ], + "angle": 0, + "content": "\\[\nP _ {\\mathrm {B L}} \\left(y _ {i} ^ {\\prime}, y _ {j} ^ {\\prime}\\right) = \\frac {1}{2} > 0 = P _ {\\mathrm {B L}} \\left(y _ {i}, y _ {j}\\right) = P _ {\\mathrm {B L}} \\left(y _ {i} ^ {\\prime \\prime}, y _ {j} ^ {\\prime \\prime}\\right)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.507, + 0.483, + 0.537 + ], + "angle": 0, + "content": "while quadratic relaxation assigns constant (zero) loss for all states during this move." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.538, + 0.484, + 0.645 + ], + "angle": 0, + "content": "Figure 1(b) shows a move problematic for the quadratic relaxation. Two neighboring pixels on the boundary of objects A and C have labels \\( y_{i} = (1,0,0) \\) and \\( y_{j} = (0,0,1) \\). The second object attempts to change from class label C to B that does not affect the Potts discontinuity penalty between two pixels. But, quadratic relaxation prefers that the second object stays in the intermediate soft state \\( y_{j}^{\\prime} = (0,\\frac{1}{2},\\frac{1}{2}) \\)" + }, + { + "type": "equation", + "bbox": [ + 0.122, + 0.656, + 0.452, + 0.687 + ], + "angle": 0, + "content": "\\[\nP _ {\\mathrm {Q}} (y _ {i}, y _ {j} ^ {\\prime}) = \\frac {3}{4} < 1 = P _ {\\mathrm {Q}} (y _ {i}, y _ {j}) = P _ {\\mathrm {Q}} (y _ {i}, y _ {j} ^ {\\prime \\prime})\n\\]" + }, + { + "type": "image", + "bbox": [ + 0.148, + 0.71, + 0.268, + 0.786 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.134, + 0.79, + 0.28, + 0.802 + ], + "angle": 0, + "content": "(a) both pixels change \\((\\mathrm{A}\\rightarrow \\mathrm{B})\\)" + }, + { + "type": "image", + "bbox": [ + 0.306, + 0.71, + 0.428, + 0.786 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.291, + 0.79, + 0.445, + 0.802 + ], + "angle": 0, + "content": "(b) only pixel \\(j\\) changes \\((\\mathbf{C}\\rightarrow \\mathbf{B})\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.813, + 0.485, + 0.897 + ], + "angle": 0, + "content": "Figure 1. Examples of \"moves\" for neighboring pixels \\(\\{i,j\\} \\in \\mathcal{N}\\). Their (soft) pseudo-labels \\(y_{i}\\) and \\(y_{j}\\) are illustrated on the probability simplex \\(\\Delta^K\\) for \\(K = 3\\). In (a) both pixels \\(i\\) and \\(j\\) are inside a region/object changing its label from A to B. In (b) pixels \\(i\\) and \\(j\\) are on the boundary between two regions/objects; one is fixed to class A and the other changes from class C to B." + }, + { + "type": "table", + "bbox": [ + 0.55, + 0.088, + 0.875, + 0.159 + ], + "angle": 0, + "content": "
collision cross entropy\nPCC(r(p,q) := -ln p^T qlog-quadratic\nP_LQ(p,q) := -ln(1 - ||p-q||^2/2)
collision divergence\nPCD(p,q) := -ln ||p^T q/||p/||q||- ln(1 - 1/2 ||p/||p - q/||q||^2)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.551, + 0.169, + 0.868, + 0.184 + ], + "angle": 0, + "content": "Table 2. Log-based Potts relaxations, see Fig.2(d,e,f)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.194, + 0.905, + 0.224 + ], + "angle": 0, + "content": "while bi-linear relaxation \\( P_{\\mathrm{BL}}(y_i, y_j) = 1 \\) remains constant as \\( y_j \\) transitions from state C to B." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.225, + 0.907, + 0.36 + ], + "angle": 0, + "content": "We propose a new relaxation, normalized quadratic in Table 1. Normalization leads to equivalence between quadratic and bi-linear formulations combining their benefits. As easy to check, normalized quadratic relaxation \\( P_{\\mathrm{NQ}} \\) removes the local minima in both examples of Figure 1. Table 2 also proposes \"logarithmic\" versions of the relaxations in Table 1 composing them with function \\( -\\ln (1 - x) \\). As illustrated by Figure 2, the logarithmic versions (d-f) address the \"vanishing gradients\" evident from the flat regions in (a-c)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.371, + 0.834, + 0.387 + ], + "angle": 0, + "content": "2.2. Cross-entropy and soft pseudo-labels" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.393, + 0.907, + 0.74 + ], + "angle": 0, + "content": "Shannon's cross-entropy \\( H(y, \\sigma) \\) is the most common loss for training network predictions \\( \\sigma \\) from ground truth labels \\( y \\) in the context of classification, semantic segmentation, etc. However, this loss may not be ideal for applications where the targets \\( y \\) are soft categorical distributions representing various forms of class uncertainty. For example, this paper is focused on scribble-based segmentation where the ground truth is not known for most of the pixels, and the network training is done jointly with estimating pseudo-labels \\( y \\) for the unlabeled pixels. In this case, soft labels \\( y \\) are distributions representing class uncertainty. We observe that if such \\( y \\) is used as a target in \\( H(y, \\sigma) \\), the network is trained to reproduce the uncertainty, see Figure 3(a). This motivates the discussion of alternative \"cross-entropy\" functions where the quotes indicate an informal interpretation of this information-theoretic concept. Intuitively, such functions should encourage decisiveness, as well as proximity between the predictions and pseudo-labels, but avoid mimicking the uncertainty in both directions: from soft pseudo-labels to predictions and vice-versa. We show that the last property can be achieved in a probabilistically principled manner. The following paragraphs discuss three cross-entropy functions \\( H \\) that we study in the context of our self-labeling loss (8)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.741, + 0.908, + 0.786 + ], + "angle": 0, + "content": "Standard cross-entropy provides the obvious baseline for evaluating two alternative versions that follow. For completeness, we include its mathematical definition" + }, + { + "type": "equation", + "bbox": [ + 0.551, + 0.798, + 0.907, + 0.83 + ], + "angle": 0, + "content": "\\[\nH _ {\\mathrm {C E}} \\left(y _ {i}, \\sigma_ {i}\\right) = H \\left(y _ {i}, \\sigma_ {i}\\right) \\equiv - \\sum_ {k} y _ {i} ^ {k} \\ln \\sigma_ {i} ^ {k} \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.907, + 0.901 + ], + "angle": 0, + "content": "and remind the reader that this loss is primarily used with hard or one-hot labels, in which case it is also equivalent to NLL loss - \\(\\ln \\sigma_i^{y_i}\\) previously discussed for ground truth labels (3). As mentioned earlier, Figure 3(a) shows that for" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20247" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.089, + 0.907, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.316, + 0.908, + 0.377 + ], + "angle": 0, + "content": "Figure 2. Second-order Potts relaxations in Tables 1 and 2: interaction potentials \\(P\\) for pairs of predictions \\((\\sigma_i, \\sigma_j)\\) in (4) or pseudo-labels \\((y_i, y_j)\\) in (8) are illustrated for \\(K = 2\\) when each \\(\\sigma_i\\) or \\(y_i\\) (binary distributions in \\(\\Delta^2\\)) can be represented by a single scalar as \\((x, 1 - x)\\). The contour maps are iso-levels of \\(P((x_i, 1 - x_i), (x_j, 1 - x_j))\\) over domain \\((x_i, x_j) \\in [0, 1]^2\\). The 3D plots above illustrate the potentials \\(P\\) as functions over pairs of \"logits\" \\((l_i, l_j) \\in \\mathbb{R}^2\\) where each scalar \\(l_i\\) defines binary distribution \\((x_i, 1 - x_i)\\) for \\(x_i = \\frac{1}{1 + e^{-2l_i}} \\in [0, 1]\\)." + }, + { + "type": "image", + "bbox": [ + 0.118, + 0.386, + 0.285, + 0.478 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.128, + 0.483, + 0.274, + 0.498 + ], + "angle": 0, + "content": "(a) standard \\(H_{\\mathrm{CE}}(y,\\sigma)\\)" + }, + { + "type": "image", + "bbox": [ + 0.29, + 0.388, + 0.455, + 0.478 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.301, + 0.483, + 0.445, + 0.498 + ], + "angle": 0, + "content": "(b) reverse \\(H_{\\mathrm{RCE}}(y,\\sigma)\\)" + }, + { + "type": "image", + "bbox": [ + 0.462, + 0.388, + 0.627, + 0.478 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.468, + 0.483, + 0.621, + 0.498 + ], + "angle": 0, + "content": "(c) collision \\(H_{\\mathrm{CCE}}(y,\\sigma)\\)" + }, + { + "type": "image", + "bbox": [ + 0.639, + 0.388, + 0.877, + 0.478 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.68, + 0.483, + 0.833, + 0.498 + ], + "angle": 0, + "content": "(d) empirical comparison" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.51, + 0.908, + 0.594 + ], + "angle": 0, + "content": "Figure 3. Illustration of cross-entropy functions: (a) standard (9), (b) reverse (10), and (c) collision (11). (d) shows the empirical comparison on the robustness to label uncertainty. The test uses ResNet-18 architecture on fully-supervised Natural Scene dataset [32] where we corrupted some labels. The horizontal axis shows the percentage \\(\\eta\\) of training images where the correct ground truth labels were replaced by a random label. All losses trained the model using soft target distributions \\(\\hat{y} = \\eta * u + (1 - \\eta) * y\\) representing the mixture of one-hot distribution \\(y\\) for the observed corrupt label and the uniform distribution \\(u\\), following [31]. The vertical axis shows the test accuracy. Training with the reverse and collision cross-entropy is robust to much higher levels of label uncertainty." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.62, + 0.484, + 0.711 + ], + "angle": 0, + "content": "soft pseudo-labels like \\( y = (0.5, 0.5) \\), it forces predictions to mimic or replicate the uncertainty \\( \\sigma \\approx y \\). In fact, label \\( y = (0.5, 0.5) \\) just tells that the class is unknown and the network should not be supervised by this point. This problem manifests itself in the poor performance of the standard cross-entropy (9) in our experiment discussed in Figure 3 (d) (red)." + }, + { + "type": "text", + "bbox": [ + 0.111, + 0.712, + 0.439, + 0.727 + ], + "angle": 0, + "content": "Reverse cross-entropy is defined in this paper as" + }, + { + "type": "equation", + "bbox": [ + 0.113, + 0.738, + 0.483, + 0.768 + ], + "angle": 0, + "content": "\\[\nH _ {\\mathrm {R C E}} \\left(y _ {i}, \\sigma_ {i}\\right) = H \\left(\\sigma_ {i}, y _ {i}\\right) \\equiv - \\sum_ {k} \\sigma_ {i} ^ {k} \\ln y _ {i} ^ {k} \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.485, + 0.902 + ], + "angle": 0, + "content": "switching the order of labels and predictions in (9), which is not common. Indeed, Shannon's cross-entropy is not symmetric and the first argument is normally the target distribution and the second is the estimated distribution. However, in our case, both distributions are estimated and there is no reason not to try the reverse order. It is worth noting that our self-labeling formulation (8) suggests that reverse cross-entropy naturally appears when the ADM approach splits the" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.62, + 0.907, + 0.725 + ], + "angle": 0, + "content": "decisiveness and fairness into separate sub-problems. Moreover, as Figure 3(b) shows, in this case, the network does not mimic uncertain pseudo-labels, e.g. the gradient of the blue line is zero. The results for the reverse cross-entropy in Figure 3 (d) (green) are significantly better than for the standard (red). Unfortunately, now pseudo-labels \\( y \\) mimic the uncertainty in predictions \\( \\sigma \\)." + }, + { + "type": "text", + "bbox": [ + 0.534, + 0.726, + 0.853, + 0.741 + ], + "angle": 0, + "content": "Collision cross-entropy defined in this paper as" + }, + { + "type": "equation", + "bbox": [ + 0.538, + 0.749, + 0.907, + 0.78 + ], + "angle": 0, + "content": "\\[\nH _ {\\mathrm {C C E}} \\left(y _ {i}, \\sigma_ {i}\\right) \\equiv - \\ln \\sum_ {k} \\sigma_ {i} ^ {k} y _ {i} ^ {k} \\equiv - \\ln \\sigma^ {\\top} y \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.789, + 0.907, + 0.865 + ], + "angle": 0, + "content": "resolves the problem in a principled way. It is symmetric w.r.t. pseudo-labels and predictions. The dot product \\(\\sigma^{\\top}y\\) can be seen as a probability of equality of two random variables, the predicted class \\(C\\) and unknown true class \\(T\\), which are represented by the distributions \\(\\sigma\\) and \\(y\\). Indeed," + }, + { + "type": "equation", + "bbox": [ + 0.54, + 0.872, + 0.879, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\Pr (C = T) = \\sum_ {k} P r (C = k) \\Pr (T = k) = \\sigma^ {\\top} y.\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20248" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.213 + ], + "angle": 0, + "content": "Loss (11) maximizes this \"collision\" probability instead of enforcing the equality of distributions \\(\\sigma = y\\). Figure 3(c) shows no mimicking of uncertainty (blue line). Unlike the reverse cross-entropy, this is also valid when \\(y\\) is estimated from uncertain predictions \\(\\sigma\\) since (11) is symmetric. This leads to the best performance in Figure 3 (d) (blue). Our extensive experiments are conclusive that collision cross-entropy is the best option for \\(H\\) in self-labeling loss (8)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.228, + 0.223, + 0.244 + ], + "angle": 0, + "content": "3. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.254, + 0.486, + 0.451 + ], + "angle": 0, + "content": "Our experiments evaluate the components of our loss (8) (cross-entropy, pairwise term, and neighborhood). We compare standard segmentation architectures trained using the \"textbook\" approach, i.e. loss minimization, to the state-of-the-art methods, which often require network modifications and ad-hoc training procedures. Our quantitative tests in Section 3.1 evaluate different Potts relaxations. Some qualitative examples are presented in Figure 4. Section 3.2 compares three cross-entropy terms. Section 3.3 evaluates our soft self-labeling loss on the nearest-neighbor and dense neighborhood systems. We summarized the results in Section 3.4. Section 3.5 shows that our method achieves the SOTA and, in some cases, may outperform the full supervision." + }, + { + "type": "image", + "bbox": [ + 0.131, + 0.541, + 0.231, + 0.616 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.133, + 0.617, + 0.229, + 0.627 + ], + "angle": 0, + "content": "(a) Image, GT & input" + }, + { + "type": "image", + "bbox": [ + 0.269, + 0.467, + 0.414, + 0.614 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.242, + 0.617, + 0.445, + 0.627 + ], + "angle": 0, + "content": "(b) Pseudo-labels using different Potts relaxation" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.639, + 0.483, + 0.681 + ], + "angle": 0, + "content": "Figure 4. Illustration of the difference among Potts relaxations. The visualization of soft pseudo-labels uses the convex combination of RGB colors for each class weighted by pseudo-label itself." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Dataset and evaluation We mainly use the standard PASCAL VOC 2012 dataset [18] and scribble annotations for supervision [28]. The dataset contains 21 classes including background. Following the common practice [11, 37, 38], we use the augmented version with 10,582 training images and 1449 images for validation. Our evaluation metric is the standard mean Intersection-over-Union (mIoU) on the validation set. We also test our method on two additional datasets in Section 3.5. One is Cityscapes [14] which is built for urban scenes and consists of 2975 and 500 fine-labeled images for training and validation. There are 19 out of 30 annotated classes for semantic segmentation. The other one" + }, + { + "type": "table", + "bbox": [ + 0.547, + 0.089, + 0.875, + 0.212 + ], + "angle": 0, + "content": "
scribble length ratio
00.30.50.81.0
PBL56.4261.7463.8165.7367.24
PNQ59.0165.5367.8070.6371.12
PQ58.9265.3467.8170.4371.05
PCCE56.4061.8263.8165.8167.41
PCD59.0465.5267.8470.9371.22
PLQ59.0365.4467.8170.8071.21
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.222, + 0.906, + 0.25 + ], + "angle": 0, + "content": "Table 3. Comparison of Potts relaxations with self-labeling. mIoUs on the validation set are shown here." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.277, + 0.908, + 0.339 + ], + "angle": 0, + "content": "is ADE20k [46] which has 150 fine-grained classes. There are 20210 and 2000, images for training and validation. Instead of scribble-based supervision, we followed [27] to use the block-wise annotation as a form of weak supervision." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.356, + 0.909, + 0.658 + ], + "angle": 0, + "content": "Implementation details We use DeepLabv3+ [13] with two backbones: ResNet101 [21] and MobileNetV2 [36]. We use ResNet101 in Section 3.5, and MobileNetV2 in other sections, for efficiency. We also use Vision Transformer backbone [17], specifically vit-base-patch16-224, and linear decoder in Section 3.5. All backbone networks (ResNet101 and MobileNetV2, ViT) are pre-trained onImagenet [16]. Unless stated explicitly, we use batch 12 as the default across all the experiments. To avoid feeding random predictions from the uninitialized classification head into pseudo-labeling sub-problem, we initialize the network using the cross-entropy on scribbles, i.e. the first term in loss (8), similarly to pre-training in [37]. The network parameters are optimized by SGD. The initial learning rate 0.0007 is scheduled by a polynomial decay of power 0.9. All loss variants are optimized for 60 epochs, but all hyperparameters are tuned separately for each variant. For our best result, we use \\(\\eta = 0.3\\), \\(\\lambda = 6\\), \\(H_{\\mathrm{CCE}}\\) and \\(P_{\\mathrm{CD}}\\). The color-bandwidth in the Potts model is set to 9 across all the experiments on Pascal VOC 2012 and 3 for Cityscapes and ADE20k datasets." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.668, + 0.797, + 0.684 + ], + "angle": 0, + "content": "3.1. Comparison of Potts relaxations" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.909, + 0.901 + ], + "angle": 0, + "content": "To compare different Potts relaxations within our self-labeling framework, we select \\( H_{\\mathrm{CCE}} \\) as the cross-entropy term in (8), as motivated by the results in Section 3.2. The neighborhood system is the nearest neighbors. The quantitative results are in Table 3. First, one can see that the log-based pairwise terms work consistently better. The likely explanation is that the logarithm addresses the gradient vanishing problem that is obvious in the top row of Figure 2. Moreover, the logarithm encourages smoother transitions across the boundaries (see Figure 4). This may benefit network training when the exact ground truth is unknown. Intuitively, higher uncertainty around object boundaries is natural. Second, the normalized relaxations \\( P_{\\mathrm{NQ}} \\) and \\( P_{\\mathrm{CD}} \\) work better, confirming the limitations of \\( P_{\\mathrm{BL}} \\) and \\( P_{\\mathrm{Q}} \\) discussed in Fig. 1." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20249" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.125, + 0.092, + 0.447, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.15, + 0.264, + 0.424, + 0.278 + ], + "angle": 0, + "content": "Figure 5. Comparison of cross-entropy terms." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.312, + 0.485, + 0.432 + ], + "angle": 0, + "content": "While we focus on a general soft pseudo-label solver (see Suppl.Mat), our results can motivate specialized solvers. For example, quadratic relaxation \\( P_{\\mathrm{Q}} \\) seems attractive as it defines a convex loss (8). If using \"quadratic cross-entropy\" \\( H_{\\mathrm{Q}} := \\| y - \\sigma \\|^2 \\), then our loss even has a closed-form solution for \\( y \\) [1, 19]. However, the corresponding pseudolabels tend to be overly soft and additional decidiveness (non-convex entropy) term improves mIoU." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.453, + 0.399, + 0.469 + ], + "angle": 0, + "content": "3.2. Comparison of cross-entropy terms" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.478, + 0.484, + 0.599 + ], + "angle": 0, + "content": "We compare different cross-entropy terms while fixing the pairwise term to \\( P_{\\mathrm{Q}} \\) due to its simplicity. We also use the nearest neighbors. The results are in Figure 5. In agreement with Fig.3(d), \\( H_{\\mathrm{CE}} \\) performs consistently better across different supervision levels, i.e. scribble lengths. Both \\( H_{\\mathrm{CE}} \\) and \\( H_{\\mathrm{RCE}} \\) have a strong margin over the standard entropy \\( H_{\\mathrm{CE}} \\) as network training becomes robust to uncertainty and errors in pseudo labels, see Sec.2.2." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.619, + 0.418, + 0.635 + ], + "angle": 0, + "content": "3.3. Comparison of neighborhood systems" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.644, + 0.485, + 0.901 + ], + "angle": 0, + "content": "So far we used the four nearest neighbors (NN) neighborhood for the pairwise Potts term in (8). Here we compare NN with the dense neighborhoods (DN). We use \\( H_{\\mathrm{CCE}} \\) as the cross-entropy term. To optimize pseudo-labels over DN, we still use the general gradient descent technique detailed in the Supplementary Material. The gradient computation for DN Potts can employ the efficient bilateral filtering [37], but it applies only to (proper) second-order variants, i.e. \\( P_{\\mathrm{BL}} \\) and \\( P_{\\mathrm{Q}} \\). We focus on \\( P_{\\mathrm{Q}} \\) as a preferred option for soft self-labeling. We obtained 71.1% mIoU on nearest neighbors while only getting 67.9% on dense neighborhoods (bandwidth 100). Some qualitative results are shown in Figure 6 indicating that larger neighborhoods induce lower-quality pseudo-labels. A possible explanation is that the Potts model reduces to the cardinality/volume potentials for larger neighborhoods [39]. The nearest neighborhood is better for edge alignment producing better weakly-supervised results." + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.091, + 0.587, + 0.142 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.145, + 0.585, + 0.193 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.194, + 0.585, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.243, + 0.584, + 0.293 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.518, + 0.296, + 0.589, + 0.306 + ], + "angle": 0, + "content": "(a) Image & scribble" + }, + { + "type": "image", + "bbox": [ + 0.598, + 0.091, + 0.664, + 0.141 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.599, + 0.143, + 0.664, + 0.193 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.599, + 0.194, + 0.663, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.599, + 0.244, + 0.663, + 0.294 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.607, + 0.296, + 0.657, + 0.307 + ], + "angle": 0, + "content": "(b) Predictions" + }, + { + "type": "image", + "bbox": [ + 0.678, + 0.091, + 0.745, + 0.141 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.679, + 0.143, + 0.744, + 0.192 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.679, + 0.193, + 0.744, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.679, + 0.244, + 0.744, + 0.294 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.681, + 0.295, + 0.741, + 0.31 + ], + "angle": 0, + "content": "(c) Pseudo-labels for NN" + }, + { + "type": "image", + "bbox": [ + 0.759, + 0.091, + 0.824, + 0.141 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.759, + 0.143, + 0.824, + 0.192 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.759, + 0.193, + 0.824, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.759, + 0.244, + 0.824, + 0.294 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.76, + 0.295, + 0.822, + 0.31 + ], + "angle": 0, + "content": "(d) Pseudo-labels for DN (25)" + }, + { + "type": "image", + "bbox": [ + 0.838, + 0.091, + 0.903, + 0.141 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.838, + 0.143, + 0.903, + 0.192 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.838, + 0.193, + 0.903, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.838, + 0.244, + 0.903, + 0.294 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.839, + 0.295, + 0.901, + 0.31 + ], + "angle": 0, + "content": "(e) Pseudo-labels for DN (100)" + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.323, + 0.905, + 0.351 + ], + "angle": 0, + "content": "Figure 6. Pseudo-labels generated from given network predictions using different neighborhoods: nearest (NN) and dense (DN)." + }, + { + "type": "table", + "bbox": [ + 0.547, + 0.364, + 0.875, + 0.43 + ], + "angle": 0, + "content": "
N
NNDN
GD67.069.5* [38]
SLhard69.6* [29]63.1 [28]
soft71.167.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.439, + 0.907, + 0.468 + ], + "angle": 0, + "content": "Table 4. Summary of Potts-based WSSS (full-scribbles, DeepLab V3+). Reproduced results using public code are marked by \\*." + }, + { + "type": "image", + "bbox": [ + 0.566, + 0.483, + 0.857, + 0.617 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.628, + 0.908, + 0.656 + ], + "angle": 0, + "content": "Figure 7. Comparison of different methods using Potts relaxations. The architecture is DeeplabV3+ with the backbone MobileNetV2." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.683, + 0.905, + 0.699 + ], + "angle": 0, + "content": "3.4. Soft vs. hard self-labeling vs. gradient descent" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Here we review our general observations about the Potts-based WSSS using the scribble-based results summarized in Table 4. First, in the simplest approach using direct optimization of the network parameters w.r.t. Potts-regularized losses like (4) and stochastic gradient descent, one benefits from a larger neighborhood size [38]. They use \\(P_{\\mathrm{BL}}\\) where DN smoothens the Potts model [30] helping SGD to avoid local minima. Higher-order optimization based on auxiliary self-labeling loss (8) significantly benefits from soft pseudo-labels since they can represent uncertainty. In this case (normalized) quadratic Potts relaxations offer two advantages over the tight bilinear relaxation that leads to hard solutions. On the one hand, they do not suppress label soft" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20250" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.156, + 0.088, + 0.846, + 0.463 + ], + "angle": 0, + "content": "
MethodArchitectureBatchsizeTraining AlgorithmNmIoU
GDSL
hardsoft
Full supervision: standard architectures + full masks training (NLL loss)
Deeplab [12]V212✓---75.6
Deeplab* [13]V3+12✓---76.6
Deeplab* [13]V3+16✓---78.9
Vision Transformer [17]ViT-linear16✓---81.4
Scribble supervision (full-length scribbles)
Modified architectures (△), and/or ad-hoc self-labeling (○)
BPG [41]V2 (△)10✓---73.2
URSS [33]V2 (△)16✓---74.6
SPML [22]V2 (△)16✓---74.2
PSI [44]V3+ (△)---✓(○)-74.9
SEMINAR [10]V3+ (△)12✓---76.2
TEL [27]V3+16--✓(○)-77.1
AGMM [43]ViT-linear (△)16--✓(○)-78.7
Standard architectures + Potts loss optimization (relaxations and/or self-labeling)
ScribbleSup [28]V2 (VGG16)8-✓-DN63.1
DenseCRF loss* [38]V3+12✓--DN75.8
GridCRF loss* [29]V3+12-✓-NN75.6
NonlocalCRF loss* [40]V3+12✓--SN75.7
HCCE + PQV3+12--✓NN77.5
HCCE + PCDV3+12--✓NN77.7
HCCE + PCD (no pretrain)V3+12--✓NN76.7
HCCE + PCD (no pretrain)V3+16--✓NN77.6
HCCE + PCDV3+16--✓NN78.1
HCCE + PCD (no pretrain)V3+16--✓NN77.6
HCCE + PCD (no pretrain)ViT-linear12--✓NN80.8
HCCE + PCD (no pretrain)ViT-linear16--✓NN80.94
" + }, + { + "type": "table_caption", + "bbox": [ + 0.089, + 0.473, + 0.907, + 0.572 + ], + "angle": 0, + "content": "Table 5. Comparison of scribble-supervised segmentation methods (without CRF postprocessing). We focus on general solutions applicable to any standard architecture based on mathematically well-founded convergent training procedures optimizing a clearly defined loss (the last block). The second block shows specialized solutions based on architectural modifications and/or ad-hoc training procedures that are often non-convergent and hard to reproduce. The numbers are mIoU on the validation dataset of Pascal VOC 2012. The backbone is ResNet101 unless stated otherwise. V2: deeplabV2. V3+: deeplabV3+. ViT-linear: ViT backbone and linear decoder. \\(\\mathcal{N}\\): neighborhood. “*”: reproduced results. GD: gradient descent. SL: self-labeling. “no pretrain” means the segmentation network is not pretrained using cross-entropy on scribbles. In some cases (V3+, batch size 12) soft self-labeling outperforms full supervision (see teal color numbers)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.592, + 0.482, + 0.728 + ], + "angle": 0, + "content": "ness; this softness is even formally related to classification uncertainty in the random walker method [19]. On the other hand, the convexity of quadratic relaxation does not require DN to simplify loss for gradient descent. Then, it is not surprising that NN works better than DN. The latter makes the Potts model behave as a cardinality potential [39], while in the case of NN the Potts model directly regularized the geometric boundary of objects [6]. Figure 7 also compares these approaches across different scribble lengths." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.741, + 0.293, + 0.758 + ], + "angle": 0, + "content": "3.5. Comparison to SOTA" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.765, + 0.483, + 0.901 + ], + "angle": 0, + "content": "This section uses DeepLabV3+ with ResNet101 backbone and ViT-linear transformer to evaluate our self-labeling methodology on these standard architectures and to compare with representative SOTA for scribble-based segmentation. Table 5 presents only the results before any post-processing. Our method may outperform the fully-supervised method (teal color numbers for DeepLabV3+ and batch size 12). Our \"text-book\" training technique, i.e. minimization of a well-defined loss, applied to standard architectures out-" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.592, + 0.907, + 0.622 + ], + "angle": 0, + "content": "performs complex WSSS systems requiring architectural modifications and/or ad-hoc multi-stage training procedures." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.634, + 0.64, + 0.65 + ], + "angle": 0, + "content": "4. Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.659, + 0.909, + 0.903 + ], + "angle": 0, + "content": "This paper proposes a mathematically well-founded self-labeling WSSS framework for standard segmentation networks based on a joint loss (8) w.r.t. network predictions and soft pseudo-labels. The latter was motivated as auxiliary variables simplifying optimization of (4). Conceptual properties and systematic evaluation of the terms in our soft self-labeling loss (8) advocate for the collision cross-entropy and collision divergence. The former replaces the standard cross-entropy \\( H \\), and the latter is used as a log of a normalized quadratic relaxation of the Potts model \\( P \\) over the NN neighborhood. Our general methodology works with any standard architecture and outperforms complex specialized systems based on architectural modifications and/or ad-hoc multi-stage non-convergent training. In contrast, our ideas are easy to understand, reproduce, and extend to other forms of weakly-supervised segmentation (boxes, class tags, etc.)." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "20251" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.486, + 0.17 + ], + "angle": 0, + "content": "[1] Multilabel random walker image segmentation using prior models. In 2005 IEEE computer society conference on computer vision and pattern recognition (CVPR'05), pages 763-770. IEEE, 2005. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.173, + 0.484, + 0.241 + ], + "angle": 0, + "content": "[2] Jiwoon Ahn and Suha Kwak. Learning pixel-level semantic affinity with image-level supervision for weakly supervised semantic segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4981-4990, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.245, + 0.484, + 0.312 + ], + "angle": 0, + "content": "[3] Jiwoon Ahn, Sunghyun Cho, and Suha Kwak. Weakly supervised learning of instance segmentation with inter-pixel relations. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2209-2218, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.316, + 0.484, + 0.37 + ], + "angle": 0, + "content": "[4] Nikita Araslanov and Stefan Roth. Single-stage semantic segmentation from image labels. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4253-4262, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.373, + 0.484, + 0.401 + ], + "angle": 0, + "content": "[5] Stephen Boyd and Lieven Vandenberghe. Convex optimization. Cambridge university press, 2004. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.403, + 0.484, + 0.457 + ], + "angle": 0, + "content": "[6] Yuri Boykov and Vladimir Kolmogorov. Computing geodesics and minimal surfaces via graph cuts. In International conference on computer vision (ICCV), 2003. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.46, + 0.484, + 0.514 + ], + "angle": 0, + "content": "[7] Yuri Y Boykov and M-P Jolly. Interactive graph cuts for optimal boundary & region segmentation of objects in nd images. In international conference on computer vision (ICCV), pages 105-112, 2001. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.517, + 0.484, + 0.571 + ], + "angle": 0, + "content": "[8] John Bridle, Anthony Heading, and David MacKay. Unsupervised classifiers, mutual information and phantom targets. Advances in neural information processing systems, 4, 1991. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.575, + 0.484, + 0.629 + ], + "angle": 0, + "content": "[9] Antonin Chambolle and Thomas Pock. A first-order primal-dual algorithm for convex problems with applications to imaging. Journal of mathematical imaging and vision, 40:120-145, 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.633, + 0.484, + 0.701 + ], + "angle": 0, + "content": "[10] Hongjun Chen, Jinbao Wang, Hong Cai Chen, Xiantong Zhen, Feng Zheng, Rongrong Ji, and Ling Shao. Seminar learning for click-level weakly supervised semantic segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6920-6929, 2021. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.704, + 0.484, + 0.759 + ], + "angle": 0, + "content": "[11] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L Yuille. Semantic image segmentation with deep convolutional nets and fully connected crfs. arXiv preprint arXiv:1412.7062, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.762, + 0.484, + 0.829 + ], + "angle": 0, + "content": "[12] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L Yuille. Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. IEEE transactions on pattern analysis and machine intelligence, 40(4):834-848, 2017. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.832, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[13] Liang-Chieh Chen, Yukun Zhu, George Papandreou, Florian Schroff, and Hartwig Adam. Encoder-decoder with atrous separable convolution for semantic image segmentation. In Proceedings of the European conference on computer vision (ECCV), pages 801-818, 2018. 6, 8" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.486, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.174 + ], + "angle": 0, + "content": "[14] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.178, + 0.908, + 0.232 + ], + "angle": 0, + "content": "[15] Camille Couprie, Leo Grady, Laurent Najman, and Hugues Talbot. Power watershed: A unifying graph-based optimization framework. IEEE transactions on pattern analysis and machine intelligence, 33(7):1384-1399, 2010. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.234, + 0.907, + 0.287 + ], + "angle": 0, + "content": "[16] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248–255. IEEE, 2009. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.289, + 0.907, + 0.328 + ], + "angle": 0, + "content": "[17] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.331, + 0.907, + 0.384 + ], + "angle": 0, + "content": "[18] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. International journal of computer vision, 88:303-308, 2009. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.387, + 0.906, + 0.426 + ], + "angle": 0, + "content": "[19] Leo Grady. Random walks for image segmentation. IEEE transactions on pattern analysis and machine intelligence, 28 (11):1768-1783, 2006. 2, 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.429, + 0.907, + 0.469 + ], + "angle": 0, + "content": "[20] Yves Grandvalet and Yoshua Bengio. Semi-supervised learning by entropy minimization. Advances in neural information processing systems, 17, 2004. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.471, + 0.907, + 0.525 + ], + "angle": 0, + "content": "[21] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.527, + 0.907, + 0.579 + ], + "angle": 0, + "content": "[22] Tsung-Wei Ke, Jyh-Jing Hwang, and Stella X Yu. Universal weakly supervised segmentation by pixel-to-segment contrastive learning. arXiv preprint arXiv:2105.00957, 2021. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.582, + 0.907, + 0.663 + ], + "angle": 0, + "content": "[23] Alexander Kolesnikov and Christoph H Lampert. Seed, expand and constrain: Three principles for weakly-supervised image segmentation. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part IV 14, pages 695-711. Springer, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.666, + 0.907, + 0.717 + ], + "angle": 0, + "content": "[24] Philipp Krahenbuhl and Vladlen Koltun. Efficient inference in fully connected CRFs with Gaussian edge potentials. Advances in neural information processing systems, 24, 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.721, + 0.907, + 0.762 + ], + "angle": 0, + "content": "[25] V. Kulharia, S. Chandra, A. Agrawal, P. Torr, and A. Tyagi. Box2seg: Attention weighted loss and discriminative feature learning for weakly supervised segmentation. In ECCV'20. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.763, + 0.907, + 0.831 + ], + "angle": 0, + "content": "[26] Jungbeom Lee, Eunji Kim, Sungmin Lee, Jangho Lee, and Sungroh Yoon. Ficklenet: Weakly and semi-supervised semantic image segmentation using stochastic inference. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5267-5276, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.833, + 0.907, + 0.901 + ], + "angle": 0, + "content": "[27] Zhiyuan Liang, Tiancai Wang, Xiangyu Zhang, Jian Sun, and Jianbing Shen. Tree energy loss: Towards sparsely annotated semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16907-16916, 2022. 2, 3, 6, 8" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "20252" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.486, + 0.161 + ], + "angle": 0, + "content": "[28] Di Lin, Jifeng Dai, Jiaya Jia, Kaiming He, and Jian Sun. Scribblesup: Scribble-supervised convolutional networks for semantic segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3159-3167, 2016. 1, 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.164, + 0.485, + 0.219 + ], + "angle": 0, + "content": "[29] Dmitrii Marin and Yuri Boykov. Robust trust region for weakly supervised segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6608-6618, 2021. 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.221, + 0.485, + 0.289 + ], + "angle": 0, + "content": "[30] Dmitrii Marin, Meng Tang, Ismail Ben Ayed, and Yuri Boykov. Beyond gradient descent for regularized segmentation losses. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10187-10196, 2019. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.291, + 0.485, + 0.333 + ], + "angle": 0, + "content": "[31] Rafael Müller, Simon Kornblith, and Geoffrey E Hinton. When does label smoothing help? Advances in neural information processing systems, 32, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.335, + 0.484, + 0.375 + ], + "angle": 0, + "content": "[32] NSD. Natural Scenes Dataset [NSD]. https://www.kaggle.com/datasets/nitishabharathi/scene-classification, 2020.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.377, + 0.485, + 0.46 + ], + "angle": 0, + "content": "[33] Zhiyi Pan, Peng Jiang, Yunhai Wang, Changhe Tu, and Anthony G Cohn. Scribble-supervised semantic segmentation by uncertainty reduction on neural representation and self-supervision on neural eigenspace. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7416-7425, 2021. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.462, + 0.485, + 0.517 + ], + "angle": 0, + "content": "[34] Thomas Pock, Antonin Chambolle, Daniel Cremers, and Horst Bischof. A convex relaxation approach for computing minimal partitions. In IEEE Conference on Computer Vision and Pattern Recognition, pages 810-817, 2009. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.519, + 0.485, + 0.574 + ], + "angle": 0, + "content": "[35] Pradeep Ravikumar and John Lafferty. Quadratic programming relaxations for metric labeling and Markov Random Field MAP estimation. In The 23rd International Conference on Machine Learning, page 737-744, 2006. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.576, + 0.485, + 0.644 + ], + "angle": 0, + "content": "[36] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zhoginov, and Liang-Chieh Chen. *Mobilenetv2: Inverted residuals and linear bottlenecks*. In *Proceedings of the IEEE conference on computer vision and pattern recognition*, pages 4510–4520, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.646, + 0.485, + 0.715 + ], + "angle": 0, + "content": "[37] Meng Tang, Abdelaziz Djelouah, Federico Perazzi, Yuri Boykov, and Christopher Schroers. Normalized cut loss for weakly-supervised cnn segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1818-1827, 2018. 1, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.717, + 0.485, + 0.786 + ], + "angle": 0, + "content": "[38] Meng Tang, Federico Perazzi, Abdelaziz Djelouah, Ismail Ben Ayed, Christopher Schroers, and Yuri Boykov. On regularized losses for weakly-supervised cnn segmentation. In Proceedings of the European Conference on Computer Vision (ECCV), pages 507-522, 2018. 1, 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.788, + 0.485, + 0.829 + ], + "angle": 0, + "content": "[39] Olga Veksler. Efficient graph cut optimization for full crfs with quantized edges. IEEE transactions on pattern analysis and machine intelligence, 42(4):1005-1012, 2019. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.831, + 0.485, + 0.872 + ], + "angle": 0, + "content": "[40] Olga Veksler and Yuri Boykov. Sparse non-local crf. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4493-4503, 2022. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.874, + 0.485, + 0.903 + ], + "angle": 0, + "content": "[41] Bin Wang, Guojun Qi, Sheng Tang, Tianzhu Zhang, Yunchao Wei, Linghui Li, and Yongdong Zhang. Boundary perception" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.486, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.548, + 0.093, + 0.908, + 0.134 + ], + "angle": 0, + "content": "guidance: A scribble-supervised semantic segmentation approach. In IJCAI International joint conference on artificial intelligence, 2019. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.908, + 0.205 + ], + "angle": 0, + "content": "[42] Changwei Wang, Rongtao Xu, Shibiao Xu, Weiliang Meng, and Xiaopeng Zhang. Treating pseudo-labels generation as image matting for weakly supervised semantic segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 755-765, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.207, + 0.907, + 0.261 + ], + "angle": 0, + "content": "[43] Linshan Wu, Zhun Zhong, Jiayi Ma, Yunchao Wei, Hao Chen, Leyuan Fang, and Shutao Li. Modeling the label distributions for weakly-supervised semantic segmentation. arXiv preprint arXiv:2403.13225, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.263, + 0.908, + 0.331 + ], + "angle": 0, + "content": "[44] Jingshan Xu, Chuanwei Zhou, Zhen Cui, Chunyan Xu, Yuge Huang, Pengcheng Shen, Shaoxin Li, and Jian Yang. Scribblesupervised semantic segmentation inference. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15354-15363, 2021. 2, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.333, + 0.907, + 0.388 + ], + "angle": 0, + "content": "[45] Christopher Zach, Christian Hane, and Marc Pollefeys. What is optimized in tight convex relaxations for multi-label problems? In IEEE Conference on Computer Vision and Pattern Recognition, pages 1664-1671, 2012. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.39, + 0.908, + 0.446 + ], + "angle": 0, + "content": "[46] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 633-641, 2017. 6" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.446 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20253" + } + ] +] \ No newline at end of file diff --git a/2025/Soft Self-labeling and Potts Relaxations for Weakly-supervised Segmentation/4b6c17bc-bca2-47a9-b3a7-cd9f02074980_origin.pdf b/2025/Soft Self-labeling and Potts Relaxations for Weakly-supervised Segmentation/4b6c17bc-bca2-47a9-b3a7-cd9f02074980_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..87c3cab0e6b75229adf459ef2da90bb21260cdd6 --- /dev/null +++ b/2025/Soft Self-labeling and Potts Relaxations for Weakly-supervised Segmentation/4b6c17bc-bca2-47a9-b3a7-cd9f02074980_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1427dd920027f0e33d5ae48d6a530eb63843d5e9514db10504d345925ea851f3 +size 3785790 diff --git a/2025/Soft Self-labeling and Potts Relaxations for Weakly-supervised Segmentation/full.md b/2025/Soft Self-labeling and Potts Relaxations for Weakly-supervised Segmentation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..a2cdd20419b6ac8cc796e984e3e6a34ac1448c7e --- /dev/null +++ b/2025/Soft Self-labeling and Potts Relaxations for Weakly-supervised Segmentation/full.md @@ -0,0 +1,416 @@ +# Soft Self-labeling and Potts Relaxations for Weakly-Supervised Segmentation + +Zhongwen Zhang & Yuri Boykov + +University of Waterloo, Canada + +z889zhan@uwaterloo.ca, yboykov@uwaterloo.ca + +# Abstract + +We consider weakly supervised segmentation where only a fraction of pixels have ground truth labels (scribbles) and focus on a self-labeling approach optimizing relaxations of the standard unsupervised CRF/Potts loss on unlabeled pixels. While WSSS methods can directly optimize such losses via gradient descent, prior work suggests that higher-order optimization can improve network training by introducing hidden pseudo-labels and powerful CRF sub-problem solvers, e.g. graph cut. However, previously used hard pseudo-labels can not represent class uncertainty or errors, which motivates soft self-labeling. We derive a principled auxiliary loss and systematically evaluate standard and new CRF relaxations (convex and non-convex), neighborhood systems, and terms connecting network predictions with soft pseudolabels. We also propose a general continuous sub-problem solver. Using only standard architectures, soft self-labeling consistently improves scribble-based training and outperforms significantly more complex specialized WSSS systems. It can outperform full pixel-precise supervision. Our general ideas apply to other weakly-supervised problems/systems. + +Keywords: soft pseudo-labels, collision cross-entropy, collision divergence + +# 1. Introduction + +Full supervision for semantic segmentation requires thousands of training images with complete pixel-accurate ground truth masks. Their high costs explain the interest in weakly-supervised approaches based on image-level class tags [4, 23], pixel-level scribbles [28, 37, 38], or boxes [25]. This paper is focused on weak supervision with scribbles, which we also call seeds or partial masks. While only slightly more expensive than image-level class tags, scribbles on less than $3\%$ of pixels were previously shown to achieve accuracy approaching full supervision without any modifications of the segmentation models. In contrast, tag supervision typically requires highly specialized systems and complex multi-stage training procedures, which are hard + +to reproduce. Our interest in the scribble-based approach is motivated by its practical simplicity and mathematical clarity. The corresponding methodologies are focused on the design of unsupervised or self-supervised loss functions and stronger optimization algorithms. The corresponding solutions are often general and can be used in different weakly-supervised applications. + +# 1.1. Scribble-supervised segmentation + +Let $\Omega$ denote a set of image pixels and $S\subset \Omega$ denote a subset of pixels with ground truth labels, i.e. seeds or scribbles typically marked by mouse-controlled UI for image annotations. One example of image scribbles is in Fig.4(a). The ground truth label at any given pixel $i\in S$ is an integer + +$$ +\bar {y} _ {i} \in \{1, \dots , K \} \tag {1} +$$ + +where $K$ is the number of classes including the background. Without much ambiguity, it is convenient to use the same notation $\bar{y}_i$ for the equivalent one-hot distribution + +$$ +\bar {y} _ {i} \equiv (\bar {y} _ {i} ^ {1}, \dots , \bar {y} _ {i} ^ {K}) \in \Delta_ {0, 1} ^ {K} \quad \text {f o r} \quad \bar {y} _ {i} ^ {k} := [ k = \bar {y} _ {i} ] \tag {2} +$$ + +where $[\cdot ]\in \{0,1\}$ is the True operator. Set $\Delta_{0,1}^{K}$ represents $K$ possible one-hot distributions, which are vertices of the $K$ -class probability simplex + +$$ +\Delta^ {K} := \{p = (p ^ {1}, \dots , p ^ {K}) \mid p ^ {k} \geq 0, \sum_ {k = 1} ^ {K} p ^ {k} = 1 \} +$$ + +representing all $K$ -categorical distributions. The context of each specific expression should make it obvious if $\bar{y}_i$ is a class index (1) or the corresponding one-hot distribution (2). + +Loss functions for scribble-supervised segmentation typically use negative log-likelihoods (NLL) over scribbles $i \in S \subset \Omega$ with ground truth labels $\bar{y}_i$ + +$$ +- \sum_ {i \in S} \ln \sigma_ {i} ^ {\bar {y} _ {i}} \tag {3} +$$ + +where $\sigma_{i} = (\sigma_{i}^{1},\dots,\sigma_{i}^{K})\in \Delta^{K}$ are model predictions at pixels $i$ . This loss is also standard in full supervision + +where $S = \Omega$ . In this case, no other losses are necessary for training. However, in a weakly supervised setting the majority of pixels are unlabeled, and unsupervised losses are needed for $i \notin S$ . + +The most common unsupervised loss in image segmentation is the Potts model and its relaxations. It is a pairwise loss defined on pairs of neighboring pixels $\{i,j\} \in \mathcal{N}$ for a given neighborhood system $\mathcal{N} \subset \Omega \times \Omega$ , typically corresponding to the nearest-neighbor grid (NN) [7, 19], or other sparse (SN) [40] or dense neighborhoods (DN) [24]. The original Potts model is defined for discrete segmentation variables, e.g. as in + +$$ +\sum_ {\{i, j \} \in \mathcal {N}} P (\sigma_ {i}, \sigma_ {j}) \quad \text {w h e r e} \quad P (\sigma_ {i}, \sigma_ {j}) = [ \sigma_ {i} \neq \sigma_ {j} ] +$$ + +assuming integer-valued one-hot predictions $\sigma_{i}\in \Delta_{0,1}^{K}$ This regularization loss encourages smoothness between the pixels. Its popular self-supervised variant is + +$$ +P \left(\sigma_ {i}, \sigma_ {j}\right) = w _ {i, j} \cdot \left[ \sigma_ {i} \neq \sigma_ {j} \right] +$$ + +where pairwise affinities $w_{ij}$ are based on local intensity edges [7, 19, 24]. Of course, in the context of network training, one should use relaxations of $P$ applicable to soft predictions $\sigma_i \in \Delta^K$ . Many forms of such relaxations [35, 45] were studied in segmentation, e.g. quadratic [19], bi-linear [38], total variation [9, 34], and others [15]. + +Another general unsupervised loss relevant to training segmentation networks is the entropy of predictions, which is also known as decidiveness [8, 20] + +$$ +\sum_ {i} H (\sigma_ {i}) +$$ + +where $H$ is the Shannon's entropy function. This loss is known to improve generalization and the quality of representation by moving (deep) features away from the decision boundaries [20]. Widely known in the context of unsupervised or semi-supervised classification, this loss also matters in weakly-supervised segmentation where it is used explicitly or implicitly1. + +Weakly-supervised segmentation methods [10, 22, 33, 41] also use other unsupervised losses (e.g. contrastive), clustering criteria (e.g. K-means), highly-specialized architectures, and ad-hoc training procedures. Our goal is to demonstrate that better results can be easily achieved on standard architectures by combining the standard losses above + +$$ +- \sum_ {i \in S} \ln \sigma_ {i} ^ {\bar {y} _ {i}} + \eta \sum_ {i \notin S} H (\sigma_ {i}) + \lambda \sum_ {i j \in \mathcal {N}} P (\sigma_ {i}, \sigma_ {j}). \tag {4} +$$ + +but replacing basic gradient descent optimization [38, 40] by higher-order self-labeling techniques [28-30] incorporating optimization of auxiliary pseudo-labels as sub-problems. + +# 1.2. Self-labeling and hard pseudo-labels + +One motivation for the self-labeling approach to weakly-supervised segmentation comes from well-known limitations of gradient descent for optimizing the Potts relaxations. Even convex relaxations [9, 19, 34] could be challenging in combination with the concave entropy term in (4). + +Typical self-labeling methods, including one of the first works on scribble-based semantic segmentation [28], introduce a sub-problem focused on the estimation of pseudolabels for unlabeled points. This subproblem can be formally "derived" from the original network optimization problem or "designed" as a heuristic, see the overview below. We focus on principled auxiliary loss formulations where network parameters and pseudo-labels are jointly optimized in a provably convergent manner. In this case pseudo-labels work as hidden optimization variables simplifying the model estimation, e.g. as in the EM algorithm for GMM. + +We denote pseudo-labels $y_{i}$ differently from the ground truth labels $\bar{y}_{i}$ by omitting the bar. It is important to distinguish them since $\bar{y}_{i}$ for $i \in S$ are given, while $y_{i}$ for $i \in \Omega \backslash S$ are estimated. The majority of existing self-labeling methods [2, 3, 26, 28-30, 42] estimate hard pseudo-labels, which could be equivalently represented either by class indices + +$$ +y _ {i} \in \{1, \dots , K ^ {\prime} \} \tag {5} +$$ + +or by the corresponding one-hot categorical distributions + +$$ +y _ {i} \equiv \left(y _ {i} ^ {1}, \dots , y _ {i} ^ {K}\right) \in \Delta_ {0, 1} ^ {K} \quad \text {f o r} \quad y _ {i} ^ {k} := [ k = y _ {i} ] \tag {6} +$$ + +analogously with the hard ground truth labels in (1) and (2). In part, hard pseudo-labels are motivated by the network training where the default is NLL loss (3) assuming discrete labels. Besides, powerful discrete solvers for the Potts model are well-known [7, 9, 34]. We discuss the potential advantages of soft pseudo-labels in Section 1.3. + +Overview: joint loss vs "proposal generation": Most self-labeling approaches can be divided into two groups. One group designs pseudo-labeling and the network training sub-problems that are not formally related, e.g. [27, 28, 44]. While pseudo-labeling typically depends on the current network predictions and the network fine-tuning uses such pseudo-labels, the lack of a formal relation between these sub-problems implies that iterating such steps does not guarantee any form of convergence. Such methods are often referred to as proposal generation heuristics. + +Alternatively, the pseudo-labeling sub-problem and the network training sub-problem can be formally derived from a weakly-supervised loss like (4), e.g. by ADM splitting [30] or as high-order trust-region method [29]. Such methods often formulate a joint loss w.r.t. network predictions and pseudo-labels and iteratively decrease the loss. These self-labeling methods are mathematically well-founded, easy to understand and reproduce, and numerically stable due to guaranteed convergence. + +# 1.3. Soft pseudo-labels: motivation & contributions + +We observe that self-labeling with hard pseudo-labels $y_{i}$ is inherently limited as such labels cannot represent the uncertainty of class estimates at unlabeled pixels $i \in \Omega \backslash S$ . Instead, we focus on soft pseudo-labels + +$$ +y _ {i} = \left(y _ {i} ^ {1}, \dots , y _ {i} ^ {K}\right) \in \Delta^ {K} \tag {7} +$$ + +which are general categorical distributions $p$ over $K$ -classes. It is possible that the estimated pseudo-label $y_{i}$ in (7) could be a one-hot distribution, which is a vertex of $\Delta^K$ . In this special case, $y_{i}$ can be treated as a class index, but we use the general formulation (7) in the main parts of the paper starting Section 2. On the other hand, the ground truth labels $\bar{y}_{i}$ are always hard and we use them either as indices (1) or one-hot distributions (2), as convenient. + +Soft pseudo-labels can be found in prior work on weakly-supervised segmentation [27, 44] using the "soft proposal generation". In contrast, we formulate soft self-labeling as a principled optimization methodology where network predictions and soft pseudo-labels are variables in a joint loss, which guarantees convergence of the training procedure. Our pseudo-labels are auxiliary variables for ADM-based [5] splitting of the loss (4) into two simpler optimization sub-problems: one focused on the Potts model over unlabeled pixels, and the other on the network training. While similar to [30], instead of hard, we use soft auxiliary variables for the Potts sub-problem. Our work can be seen as a study of the relaxed Potts sub-problem in the context of weakly-supervised semantic segmentation. The related prior work is focused on discrete solvers fundamentally unable to represent class estimate uncertainty. Our contributions can be summarized as follows: + +- convergent soft self-labeling framework based on a simple joint self-labeling loss +- systematic evaluation of Potts relaxations and (cross-) entropy terms in our loss +- state-of-the-art in scribble-based semantic segmentation based on standard network architectures (no modifications) and well-founded convergent training procedures that are easy to reproduce and generalize to other problems +- using the same segmentation model, our self-labeling loss with $3\%$ scribbles may outperform standard supervised cross-entropy loss with full ground truth masks. + +# 2. Our soft self-labeling approach + +First, we apply ADM splitting [5] to weakly supervised loss (4) to formulate our self-labeling loss (8) incorporating additional soft auxiliary variables, i.e. pseudo-labels (7). It is convenient to introduce pseudo-labels $y_{i}$ on all pixels in $\Omega$ even though a subset of pixels (seeds) $S\subset \Omega$ have ground truth labels $\bar{y}_i$ . We will simply impose a constraint that pseudo-labels and ground truth labels agree on $S$ . Thus, + +we assume the following set of pseudo-labels + +$$ +Y _ {\Omega} := \left\{y _ {i} \in \Delta^ {K} \mid i \in \Omega , \text {s . t .} y _ {i} = \bar {y} _ {i} \text {f o r} i \in S \right\}. +$$ + +We split the terms in (4) into two groups: one includes NLL and entropy $H$ terms keeping the original prediction variables $\sigma_{i}$ and the other includes the Potts relaxation $P$ replacing $\sigma_{i}$ with auxiliary variables $y_{i}$ . This transforms loss (4) into expression + +$$ +- \sum_ {i \in S} \ln \sigma_ {i} ^ {\bar {y} _ {i}} + \eta \sum_ {i \not \in S} H (\sigma_ {i}) + \lambda \sum_ {i j \in \mathcal {N}} P (y _ {i}, y _ {j}) +$$ + +equivalent to (4) assuming equality $\sigma_{i} = y_{i}$ . The standard approximation is to incorporate constraint $\sigma_{i} \approx y_{i}$ directly into the loss, e.g. using $KL$ -divergence. For simplicity, we use weight $\eta$ for $KL(\sigma_{i},y_{i})$ to combine it with $H(\sigma_{i})$ into a single cross-entropy term + +$$ +-\sum_{i\in S}\ln \sigma_{i}^{\bar{y}_{i}} + \underbrace{\eta\sum_{i\notin S}H(\sigma_{i}) + \eta\sum_{i\notin S}KL(\sigma_{i},y_{i})}_{\substack{\eta \sum_{i\notin S}H(\sigma_{i},y_{i})}} + \lambda \sum_{ij\in \mathcal{N}}P(y_{i},y_{j}) +$$ + +defining auxiliary self-labeling loss w.r.t. both $\sigma_{i}$ and $y_{i}$ + +$$ +- \sum_ {i \in S} \ln \sigma_ {i} ^ {\bar {y} _ {i}} + \eta \sum_ {i \notin S} H \left(\sigma_ {i}, y _ {i}\right) + \lambda \sum_ {i j \in \mathcal {N}} P \left(y _ {i}, y _ {j}\right) \tag {8} +$$ + +approximating the original weakly supervised loss (4). + +Iterative minimization of this loss w.r.t. predictions $\sigma_{i}$ (model parameters training) and pseudo-labels $y_{i}$ effectively breaks the original optimization problem for (4) into two simpler sub-problems, assuming there is a good solver for optimal pseudo-labels. The latter is plausible since the unary term $H(\sigma_{i}, y_{i})$ is convex for $y_{i}$ , and the Potts model (incl. convex and non-convex relaxations) is widely studied. + +Section 2.1 discusses standard and new relaxations of the Potts model $P$ . Section 2.2 discusses several robust variants of the cross-entropy term $H$ in (8) connecting predictions with uncertain (soft) pseudo-labels $y_{i}$ at unlabeled points $i \in \Omega \backslash S$ . To systematically evaluate these variants of (8) in Section 3, we propose a general efficient GPU-friendly solver for the corresponding pseudo-labeling sub-problem, which is detailed in the Supplementary Material. + +# 2.1. Second-order relaxations of the Potts model + +We limit our focus to basic second-order relaxations and several motivated extensions. We start from two important standard cases (see Table 1): quadratic, the simplest convex relaxation popularized by the random walker algorithm [19], and bi-linear, which is non-convex but tight [35] w.r.t. the original discrete Potts model. The latter implies that optimizing it over relaxed variables should lead to a solution consistent with a discrete Potts solver, e.g. graph cut [7] regularizing the geometric boundary of the segments [6]. The quadratic relaxation produces a significantly different soft solution motivated by random walk probabilities [19]. + +
bi-linear ~ “graph cut”PBL(p,q) := 1 - pTqquadratic ~ “random walker”P0(p,q) := 1/2||p-q|2
normalized quadratic
PNQ(p,q) := 1 - pTq/||p||q|≡ 1/2||p/||p-| - q/||q| |2
+ +While there are specialized solvers for bi-linear and quadratic relaxations, we are interested in a general efficient pseudo-label solver integrated into the standard GPU-based network optimization framework. This suggests gradient descent as an optimizer for pseudo-labels $y$ (see Suppl. Mat). Figure 1 shows two representative local minima issues that gradient descent may encounter for (a) the bi-linear and (b) quadratic relaxations of the Potts loss. These examples indicate that gradient descent for these relaxations can distort the properties of the original (discrete) Potts loss. + +In (a) two neighboring pixels jointly change the common label from $y_{i} = y_{j} = (1,0,0)$ to $y_{i}^{\prime \prime} = y_{j}^{\prime \prime} = (0,1,0)$ , which corresponds to a "move" where the whole object is reclassified from A to B. This move does not violate the smoothness constraint represented by the original (discrete) Potts model. But, bilinear relaxation prevents move (a) since the intermediate state $y_{i}^{\prime} = y_{j}^{\prime} = (\frac{1}{2},\frac{1}{2},0)$ has a higher cost + +$$ +P _ {\mathrm {B L}} \left(y _ {i} ^ {\prime}, y _ {j} ^ {\prime}\right) = \frac {1}{2} > 0 = P _ {\mathrm {B L}} \left(y _ {i}, y _ {j}\right) = P _ {\mathrm {B L}} \left(y _ {i} ^ {\prime \prime}, y _ {j} ^ {\prime \prime}\right) +$$ + +while quadratic relaxation assigns constant (zero) loss for all states during this move. + +Figure 1(b) shows a move problematic for the quadratic relaxation. Two neighboring pixels on the boundary of objects A and C have labels $y_{i} = (1,0,0)$ and $y_{j} = (0,0,1)$ . The second object attempts to change from class label C to B that does not affect the Potts discontinuity penalty between two pixels. But, quadratic relaxation prefers that the second object stays in the intermediate soft state $y_{j}^{\prime} = (0,\frac{1}{2},\frac{1}{2})$ + +$$ +P _ {\mathrm {Q}} (y _ {i}, y _ {j} ^ {\prime}) = \frac {3}{4} < 1 = P _ {\mathrm {Q}} (y _ {i}, y _ {j}) = P _ {\mathrm {Q}} (y _ {i}, y _ {j} ^ {\prime \prime}) +$$ + +![](images/31f8396aa63da9926854ef470abb9e99765435b8f035951b0f7f713f97359515.jpg) +(a) both pixels change $(\mathrm{A}\rightarrow \mathrm{B})$ + +![](images/19a68655c2a8b548079376697f6c0101827163c2921d6edfbdde4a4e268d031e.jpg) +(b) only pixel $j$ changes $(\mathbf{C}\rightarrow \mathbf{B})$ +Figure 1. Examples of "moves" for neighboring pixels $\{i,j\} \in \mathcal{N}$ . Their (soft) pseudo-labels $y_{i}$ and $y_{j}$ are illustrated on the probability simplex $\Delta^K$ for $K = 3$ . In (a) both pixels $i$ and $j$ are inside a region/object changing its label from A to B. In (b) pixels $i$ and $j$ are on the boundary between two regions/objects; one is fixed to class A and the other changes from class C to B. + +Table 1. Second-order Potts relaxations, see Fig.2(a,b,c) + +
collision cross entropy +PCC(r(p,q) := -ln p^T qlog-quadratic +P_LQ(p,q) := -ln(1 - ||p-q||^2/2)
collision divergence +PCD(p,q) := -ln ||p^T q/||p/||q||- ln(1 - 1/2 ||p/||p - q/||q||^2)
+ +Table 2. Log-based Potts relaxations, see Fig.2(d,e,f) + +while bi-linear relaxation $P_{\mathrm{BL}}(y_i, y_j) = 1$ remains constant as $y_j$ transitions from state C to B. + +We propose a new relaxation, normalized quadratic in Table 1. Normalization leads to equivalence between quadratic and bi-linear formulations combining their benefits. As easy to check, normalized quadratic relaxation $P_{\mathrm{NQ}}$ removes the local minima in both examples of Figure 1. Table 2 also proposes "logarithmic" versions of the relaxations in Table 1 composing them with function $-\ln (1 - x)$ . As illustrated by Figure 2, the logarithmic versions (d-f) address the "vanishing gradients" evident from the flat regions in (a-c). + +# 2.2. Cross-entropy and soft pseudo-labels + +Shannon's cross-entropy $H(y, \sigma)$ is the most common loss for training network predictions $\sigma$ from ground truth labels $y$ in the context of classification, semantic segmentation, etc. However, this loss may not be ideal for applications where the targets $y$ are soft categorical distributions representing various forms of class uncertainty. For example, this paper is focused on scribble-based segmentation where the ground truth is not known for most of the pixels, and the network training is done jointly with estimating pseudo-labels $y$ for the unlabeled pixels. In this case, soft labels $y$ are distributions representing class uncertainty. We observe that if such $y$ is used as a target in $H(y, \sigma)$ , the network is trained to reproduce the uncertainty, see Figure 3(a). This motivates the discussion of alternative "cross-entropy" functions where the quotes indicate an informal interpretation of this information-theoretic concept. Intuitively, such functions should encourage decisiveness, as well as proximity between the predictions and pseudo-labels, but avoid mimicking the uncertainty in both directions: from soft pseudo-labels to predictions and vice-versa. We show that the last property can be achieved in a probabilistically principled manner. The following paragraphs discuss three cross-entropy functions $H$ that we study in the context of our self-labeling loss (8). + +Standard cross-entropy provides the obvious baseline for evaluating two alternative versions that follow. For completeness, we include its mathematical definition + +$$ +H _ {\mathrm {C E}} \left(y _ {i}, \sigma_ {i}\right) = H \left(y _ {i}, \sigma_ {i}\right) \equiv - \sum_ {k} y _ {i} ^ {k} \ln \sigma_ {i} ^ {k} \tag {9} +$$ + +and remind the reader that this loss is primarily used with hard or one-hot labels, in which case it is also equivalent to NLL loss - $\ln \sigma_i^{y_i}$ previously discussed for ground truth labels (3). As mentioned earlier, Figure 3(a) shows that for + +![](images/46b0c1694453a6d3d12bc007b4636fb591334bcf3f67ea99405c82c8ed82358e.jpg) +Figure 2. Second-order Potts relaxations in Tables 1 and 2: interaction potentials $P$ for pairs of predictions $(\sigma_i, \sigma_j)$ in (4) or pseudo-labels $(y_i, y_j)$ in (8) are illustrated for $K = 2$ when each $\sigma_i$ or $y_i$ (binary distributions in $\Delta^2$ ) can be represented by a single scalar as $(x, 1 - x)$ . The contour maps are iso-levels of $P((x_i, 1 - x_i), (x_j, 1 - x_j))$ over domain $(x_i, x_j) \in [0, 1]^2$ . The 3D plots above illustrate the potentials $P$ as functions over pairs of "logits" $(l_i, l_j) \in \mathbb{R}^2$ where each scalar $l_i$ defines binary distribution $(x_i, 1 - x_i)$ for $x_i = \frac{1}{1 + e^{-2l_i}} \in [0, 1]$ . + +![](images/b2db9ae6cb80b535e88f7329a37e10372f6ab064c854737f6e5261af754da1cd.jpg) +(a) standard $H_{\mathrm{CE}}(y,\sigma)$ + +![](images/4d1f589b41658e5832c2fc49bf33b725a1bfef477a4d60b4376fcaa95ef1912e.jpg) +Figure 3. Illustration of cross-entropy functions: (a) standard (9), (b) reverse (10), and (c) collision (11). (d) shows the empirical comparison on the robustness to label uncertainty. The test uses ResNet-18 architecture on fully-supervised Natural Scene dataset [32] where we corrupted some labels. The horizontal axis shows the percentage $\eta$ of training images where the correct ground truth labels were replaced by a random label. All losses trained the model using soft target distributions $\hat{y} = \eta * u + (1 - \eta) * y$ representing the mixture of one-hot distribution $y$ for the observed corrupt label and the uniform distribution $u$ , following [31]. The vertical axis shows the test accuracy. Training with the reverse and collision cross-entropy is robust to much higher levels of label uncertainty. + +![](images/d95d627af0eaa8670a8b91a7672463a1f1a1f281a7923bbff07b04d8e010fe64.jpg) +(b) reverse $H_{\mathrm{RCE}}(y,\sigma)$ + +![](images/5d6c5709ee51b1acfe9156ba39ee372ee65e69df03cfdfe2b587762885382f71.jpg) +(c) collision $H_{\mathrm{CCE}}(y,\sigma)$ +(d) empirical comparison + +soft pseudo-labels like $y = (0.5, 0.5)$ , it forces predictions to mimic or replicate the uncertainty $\sigma \approx y$ . In fact, label $y = (0.5, 0.5)$ just tells that the class is unknown and the network should not be supervised by this point. This problem manifests itself in the poor performance of the standard cross-entropy (9) in our experiment discussed in Figure 3 (d) (red). + +Reverse cross-entropy is defined in this paper as + +$$ +H _ {\mathrm {R C E}} \left(y _ {i}, \sigma_ {i}\right) = H \left(\sigma_ {i}, y _ {i}\right) \equiv - \sum_ {k} \sigma_ {i} ^ {k} \ln y _ {i} ^ {k} \tag {10} +$$ + +switching the order of labels and predictions in (9), which is not common. Indeed, Shannon's cross-entropy is not symmetric and the first argument is normally the target distribution and the second is the estimated distribution. However, in our case, both distributions are estimated and there is no reason not to try the reverse order. It is worth noting that our self-labeling formulation (8) suggests that reverse cross-entropy naturally appears when the ADM approach splits the + +decisiveness and fairness into separate sub-problems. Moreover, as Figure 3(b) shows, in this case, the network does not mimic uncertain pseudo-labels, e.g. the gradient of the blue line is zero. The results for the reverse cross-entropy in Figure 3 (d) (green) are significantly better than for the standard (red). Unfortunately, now pseudo-labels $y$ mimic the uncertainty in predictions $\sigma$ . + +Collision cross-entropy defined in this paper as + +$$ +H _ {\mathrm {C C E}} \left(y _ {i}, \sigma_ {i}\right) \equiv - \ln \sum_ {k} \sigma_ {i} ^ {k} y _ {i} ^ {k} \equiv - \ln \sigma^ {\top} y \tag {11} +$$ + +resolves the problem in a principled way. It is symmetric w.r.t. pseudo-labels and predictions. The dot product $\sigma^{\top}y$ can be seen as a probability of equality of two random variables, the predicted class $C$ and unknown true class $T$ , which are represented by the distributions $\sigma$ and $y$ . Indeed, + +$$ +\Pr (C = T) = \sum_ {k} P r (C = k) \Pr (T = k) = \sigma^ {\top} y. +$$ + +Loss (11) maximizes this "collision" probability instead of enforcing the equality of distributions $\sigma = y$ . Figure 3(c) shows no mimicking of uncertainty (blue line). Unlike the reverse cross-entropy, this is also valid when $y$ is estimated from uncertain predictions $\sigma$ since (11) is symmetric. This leads to the best performance in Figure 3 (d) (blue). Our extensive experiments are conclusive that collision cross-entropy is the best option for $H$ in self-labeling loss (8). + +# 3. Experiments + +Our experiments evaluate the components of our loss (8) (cross-entropy, pairwise term, and neighborhood). We compare standard segmentation architectures trained using the "textbook" approach, i.e. loss minimization, to the state-of-the-art methods, which often require network modifications and ad-hoc training procedures. Our quantitative tests in Section 3.1 evaluate different Potts relaxations. Some qualitative examples are presented in Figure 4. Section 3.2 compares three cross-entropy terms. Section 3.3 evaluates our soft self-labeling loss on the nearest-neighbor and dense neighborhood systems. We summarized the results in Section 3.4. Section 3.5 shows that our method achieves the SOTA and, in some cases, may outperform the full supervision. + +![](images/320ea53741265cd39b5c7e57e714b7d10b28a7711b2a64b6acabd89d1b33d04c.jpg) +(a) Image, GT & input +Figure 4. Illustration of the difference among Potts relaxations. The visualization of soft pseudo-labels uses the convex combination of RGB colors for each class weighted by pseudo-label itself. + +![](images/3fd7e43195ea2d83908a43bfe1cbf33c198ad222dc2286e4f11694458be0b892.jpg) +(b) Pseudo-labels using different Potts relaxation + +Dataset and evaluation We mainly use the standard PASCAL VOC 2012 dataset [18] and scribble annotations for supervision [28]. The dataset contains 21 classes including background. Following the common practice [11, 37, 38], we use the augmented version with 10,582 training images and 1449 images for validation. Our evaluation metric is the standard mean Intersection-over-Union (mIoU) on the validation set. We also test our method on two additional datasets in Section 3.5. One is Cityscapes [14] which is built for urban scenes and consists of 2975 and 500 fine-labeled images for training and validation. There are 19 out of 30 annotated classes for semantic segmentation. The other one + +
scribble length ratio
00.30.50.81.0
PBL56.4261.7463.8165.7367.24
PNQ59.0165.5367.8070.6371.12
PQ58.9265.3467.8170.4371.05
PCCE56.4061.8263.8165.8167.41
PCD59.0465.5267.8470.9371.22
PLQ59.0365.4467.8170.8071.21
+ +Table 3. Comparison of Potts relaxations with self-labeling. mIoUs on the validation set are shown here. + +is ADE20k [46] which has 150 fine-grained classes. There are 20210 and 2000, images for training and validation. Instead of scribble-based supervision, we followed [27] to use the block-wise annotation as a form of weak supervision. + +Implementation details We use DeepLabv3+ [13] with two backbones: ResNet101 [21] and MobileNetV2 [36]. We use ResNet101 in Section 3.5, and MobileNetV2 in other sections, for efficiency. We also use Vision Transformer backbone [17], specifically vit-base-patch16-224, and linear decoder in Section 3.5. All backbone networks (ResNet101 and MobileNetV2, ViT) are pre-trained onImagenet [16]. Unless stated explicitly, we use batch 12 as the default across all the experiments. To avoid feeding random predictions from the uninitialized classification head into pseudo-labeling sub-problem, we initialize the network using the cross-entropy on scribbles, i.e. the first term in loss (8), similarly to pre-training in [37]. The network parameters are optimized by SGD. The initial learning rate 0.0007 is scheduled by a polynomial decay of power 0.9. All loss variants are optimized for 60 epochs, but all hyperparameters are tuned separately for each variant. For our best result, we use $\eta = 0.3$ , $\lambda = 6$ , $H_{\mathrm{CCE}}$ and $P_{\mathrm{CD}}$ . The color-bandwidth in the Potts model is set to 9 across all the experiments on Pascal VOC 2012 and 3 for Cityscapes and ADE20k datasets. + +# 3.1. Comparison of Potts relaxations + +To compare different Potts relaxations within our self-labeling framework, we select $H_{\mathrm{CCE}}$ as the cross-entropy term in (8), as motivated by the results in Section 3.2. The neighborhood system is the nearest neighbors. The quantitative results are in Table 3. First, one can see that the log-based pairwise terms work consistently better. The likely explanation is that the logarithm addresses the gradient vanishing problem that is obvious in the top row of Figure 2. Moreover, the logarithm encourages smoother transitions across the boundaries (see Figure 4). This may benefit network training when the exact ground truth is unknown. Intuitively, higher uncertainty around object boundaries is natural. Second, the normalized relaxations $P_{\mathrm{NQ}}$ and $P_{\mathrm{CD}}$ work better, confirming the limitations of $P_{\mathrm{BL}}$ and $P_{\mathrm{Q}}$ discussed in Fig. 1. + +![](images/3122dae951addc3a14b4403a881c4c59407c40048c29dccd68e9857585801043.jpg) +Figure 5. Comparison of cross-entropy terms. + +While we focus on a general soft pseudo-label solver (see Suppl.Mat), our results can motivate specialized solvers. For example, quadratic relaxation $P_{\mathrm{Q}}$ seems attractive as it defines a convex loss (8). If using "quadratic cross-entropy" $H_{\mathrm{Q}} := \| y - \sigma \|^2$ , then our loss even has a closed-form solution for $y$ [1, 19]. However, the corresponding pseudolabels tend to be overly soft and additional decidiveness (non-convex entropy) term improves mIoU. + +# 3.2. Comparison of cross-entropy terms + +We compare different cross-entropy terms while fixing the pairwise term to $P_{\mathrm{Q}}$ due to its simplicity. We also use the nearest neighbors. The results are in Figure 5. In agreement with Fig.3(d), $H_{\mathrm{CE}}$ performs consistently better across different supervision levels, i.e. scribble lengths. Both $H_{\mathrm{CE}}$ and $H_{\mathrm{RCE}}$ have a strong margin over the standard entropy $H_{\mathrm{CE}}$ as network training becomes robust to uncertainty and errors in pseudo labels, see Sec.2.2. + +# 3.3. Comparison of neighborhood systems + +So far we used the four nearest neighbors (NN) neighborhood for the pairwise Potts term in (8). Here we compare NN with the dense neighborhoods (DN). We use $H_{\mathrm{CCE}}$ as the cross-entropy term. To optimize pseudo-labels over DN, we still use the general gradient descent technique detailed in the Supplementary Material. The gradient computation for DN Potts can employ the efficient bilateral filtering [37], but it applies only to (proper) second-order variants, i.e. $P_{\mathrm{BL}}$ and $P_{\mathrm{Q}}$ . We focus on $P_{\mathrm{Q}}$ as a preferred option for soft self-labeling. We obtained 71.1% mIoU on nearest neighbors while only getting 67.9% on dense neighborhoods (bandwidth 100). Some qualitative results are shown in Figure 6 indicating that larger neighborhoods induce lower-quality pseudo-labels. A possible explanation is that the Potts model reduces to the cardinality/volume potentials for larger neighborhoods [39]. The nearest neighborhood is better for edge alignment producing better weakly-supervised results. + +![](images/c2fc96c343432f316cae67acd322d240a0eaf8957e782a7ae5608eca6dadce14.jpg) + +![](images/6c12f2614429f36f8305135c3d5b0a4640676ce2a9825d65cfdea7c56e107311.jpg) + +![](images/c3aa74185d2a26cb9d267b6338643b0b168ed92dffb2105aa6ec35ba70394505.jpg) + +![](images/de4e540c40c1915846578c05e73d7a31b9034e2d741e940643660f6da029ff74.jpg) +(a) Image & scribble + +![](images/783c1d36fb18304b47965edbe339884ab6af4f3146fa035f1c6d44d47f665884.jpg) + +![](images/52dca61eca1fb424f26824ea6883747c61bffe9a0c53a2909c060acacf91789a.jpg) + +![](images/ec7576f8952af1dde748ada0014e74f19bc8c9fef8bc6ed9068f9a3fbc54b5dc.jpg) + +![](images/1344c98ce43af0f8d631c5fa7e20b90492e7219f1be7b9ef243b779b6d0a1ad9.jpg) +(b) Predictions + +![](images/6c2fd150f8362cd43c01e74163f03d5beecae15b6411dc217162fae0a70c3e5c.jpg) + +![](images/78b460e16b3841a2e5edce9ec7d7042c8cae0f926b76845d5388a631d4a9a32b.jpg) + +![](images/978c0339fde0f1cd69429dd2d6b280fb3f870fdd28612c522988713e7b064720.jpg) + +![](images/9bb6d1ace973dbc963681178c47b3eac5869f3d017614fcb4afb4112e8ca4342.jpg) +(c) Pseudo-labels for NN + +![](images/75b68ae4089eaef4fbb7ffc0b693a307cf1605b1310277fc7e35904208d33cee.jpg) + +![](images/34471a5c14220d5b5dc38f6c58c39ea282751ee6f35897fe7473020a1325c0ce.jpg) + +![](images/f6886a149273b14426a08315df22705302beb983fdff2bfbc531d65588610f27.jpg) + +![](images/608fd117ab593e33fedd0203d633fc1c654639058b3aa991ef6d702219dc9414.jpg) +(d) Pseudo-labels for DN (25) +Figure 6. Pseudo-labels generated from given network predictions using different neighborhoods: nearest (NN) and dense (DN). + +![](images/76ed42884b0ee07f8d0b0ecb4f44ca3d1cb6f9bd880fe2c0997203184eb68fbd.jpg) + +![](images/d86a57e2d0a898c322e078ca2d59bb4d264831350558c4b62edae94879dc04b2.jpg) + +![](images/8300a86876e651a6d487845539426df4956c283074bdb6c3d3beaa58c18c68c3.jpg) + +![](images/b7f6bd2a92296f335136138e8df2c47706e4626eafd3dcaac07873e39f92b14e.jpg) +(e) Pseudo-labels for DN (100) + +
N
NNDN
GD67.069.5* [38]
SLhard69.6* [29]63.1 [28]
soft71.167.9
+ +Table 4. Summary of Potts-based WSSS (full-scribbles, DeepLab V3+). Reproduced results using public code are marked by \*. + +![](images/ba96d6de4fc604fb80670bc7b8c02a3948d3c7a63ae9d7f9d358a4ca53acf655.jpg) +Figure 7. Comparison of different methods using Potts relaxations. The architecture is DeeplabV3+ with the backbone MobileNetV2. + +# 3.4. Soft vs. hard self-labeling vs. gradient descent + +Here we review our general observations about the Potts-based WSSS using the scribble-based results summarized in Table 4. First, in the simplest approach using direct optimization of the network parameters w.r.t. Potts-regularized losses like (4) and stochastic gradient descent, one benefits from a larger neighborhood size [38]. They use $P_{\mathrm{BL}}$ where DN smoothens the Potts model [30] helping SGD to avoid local minima. Higher-order optimization based on auxiliary self-labeling loss (8) significantly benefits from soft pseudo-labels since they can represent uncertainty. In this case (normalized) quadratic Potts relaxations offer two advantages over the tight bilinear relaxation that leads to hard solutions. On the one hand, they do not suppress label soft + +
MethodArchitectureBatchsizeTraining AlgorithmNmIoU
GDSL
hardsoft
Full supervision: standard architectures + full masks training (NLL loss)
Deeplab [12]V212✓---75.6
Deeplab* [13]V3+12✓---76.6
Deeplab* [13]V3+16✓---78.9
Vision Transformer [17]ViT-linear16✓---81.4
Scribble supervision (full-length scribbles)
Modified architectures (△), and/or ad-hoc self-labeling (○)
BPG [41]V2 (△)10✓---73.2
URSS [33]V2 (△)16✓---74.6
SPML [22]V2 (△)16✓---74.2
PSI [44]V3+ (△)---✓(○)-74.9
SEMINAR [10]V3+ (△)12✓---76.2
TEL [27]V3+16--✓(○)-77.1
AGMM [43]ViT-linear (△)16--✓(○)-78.7
Standard architectures + Potts loss optimization (relaxations and/or self-labeling)
ScribbleSup [28]V2 (VGG16)8-✓-DN63.1
DenseCRF loss* [38]V3+12✓--DN75.8
GridCRF loss* [29]V3+12-✓-NN75.6
NonlocalCRF loss* [40]V3+12✓--SN75.7
HCCE + PQV3+12--✓NN77.5
HCCE + PCDV3+12--✓NN77.7
HCCE + PCD (no pretrain)V3+12--✓NN76.7
HCCE + PCD (no pretrain)V3+16--✓NN77.6
HCCE + PCDV3+16--✓NN78.1
HCCE + PCD (no pretrain)V3+16--✓NN77.6
HCCE + PCD (no pretrain)ViT-linear12--✓NN80.8
HCCE + PCD (no pretrain)ViT-linear16--✓NN80.94
+ +Table 5. Comparison of scribble-supervised segmentation methods (without CRF postprocessing). We focus on general solutions applicable to any standard architecture based on mathematically well-founded convergent training procedures optimizing a clearly defined loss (the last block). The second block shows specialized solutions based on architectural modifications and/or ad-hoc training procedures that are often non-convergent and hard to reproduce. The numbers are mIoU on the validation dataset of Pascal VOC 2012. The backbone is ResNet101 unless stated otherwise. V2: deeplabV2. V3+: deeplabV3+. ViT-linear: ViT backbone and linear decoder. $\mathcal{N}$ : neighborhood. “*”: reproduced results. GD: gradient descent. SL: self-labeling. “no pretrain” means the segmentation network is not pretrained using cross-entropy on scribbles. In some cases (V3+, batch size 12) soft self-labeling outperforms full supervision (see teal color numbers). + +ness; this softness is even formally related to classification uncertainty in the random walker method [19]. On the other hand, the convexity of quadratic relaxation does not require DN to simplify loss for gradient descent. Then, it is not surprising that NN works better than DN. The latter makes the Potts model behave as a cardinality potential [39], while in the case of NN the Potts model directly regularized the geometric boundary of objects [6]. Figure 7 also compares these approaches across different scribble lengths. + +# 3.5. Comparison to SOTA + +This section uses DeepLabV3+ with ResNet101 backbone and ViT-linear transformer to evaluate our self-labeling methodology on these standard architectures and to compare with representative SOTA for scribble-based segmentation. Table 5 presents only the results before any post-processing. Our method may outperform the fully-supervised method (teal color numbers for DeepLabV3+ and batch size 12). Our "text-book" training technique, i.e. minimization of a well-defined loss, applied to standard architectures out- + +performs complex WSSS systems requiring architectural modifications and/or ad-hoc multi-stage training procedures. + +# 4. Conclusions + +This paper proposes a mathematically well-founded self-labeling WSSS framework for standard segmentation networks based on a joint loss (8) w.r.t. network predictions and soft pseudo-labels. The latter was motivated as auxiliary variables simplifying optimization of (4). Conceptual properties and systematic evaluation of the terms in our soft self-labeling loss (8) advocate for the collision cross-entropy and collision divergence. The former replaces the standard cross-entropy $H$ , and the latter is used as a log of a normalized quadratic relaxation of the Potts model $P$ over the NN neighborhood. Our general methodology works with any standard architecture and outperforms complex specialized systems based on architectural modifications and/or ad-hoc multi-stage non-convergent training. In contrast, our ideas are easy to understand, reproduce, and extend to other forms of weakly-supervised segmentation (boxes, class tags, etc.). + +# References + +[1] Multilabel random walker image segmentation using prior models. In 2005 IEEE computer society conference on computer vision and pattern recognition (CVPR'05), pages 763-770. IEEE, 2005. 7 +[2] Jiwoon Ahn and Suha Kwak. Learning pixel-level semantic affinity with image-level supervision for weakly supervised semantic segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4981-4990, 2018. 2 +[3] Jiwoon Ahn, Sunghyun Cho, and Suha Kwak. Weakly supervised learning of instance segmentation with inter-pixel relations. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2209-2218, 2019. 2 +[4] Nikita Araslanov and Stefan Roth. Single-stage semantic segmentation from image labels. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4253-4262, 2020. 1 +[5] Stephen Boyd and Lieven Vandenberghe. Convex optimization. Cambridge university press, 2004. 3 +[6] Yuri Boykov and Vladimir Kolmogorov. Computing geodesics and minimal surfaces via graph cuts. In International conference on computer vision (ICCV), 2003. 3, 8 +[7] Yuri Y Boykov and M-P Jolly. Interactive graph cuts for optimal boundary & region segmentation of objects in nd images. In international conference on computer vision (ICCV), pages 105-112, 2001. 2, 3 +[8] John Bridle, Anthony Heading, and David MacKay. Unsupervised classifiers, mutual information and phantom targets. Advances in neural information processing systems, 4, 1991. 2 +[9] Antonin Chambolle and Thomas Pock. A first-order primal-dual algorithm for convex problems with applications to imaging. Journal of mathematical imaging and vision, 40:120-145, 2011. 2 +[10] Hongjun Chen, Jinbao Wang, Hong Cai Chen, Xiantong Zhen, Feng Zheng, Rongrong Ji, and Ling Shao. Seminar learning for click-level weakly supervised semantic segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6920-6929, 2021. 2, 8 +[11] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L Yuille. Semantic image segmentation with deep convolutional nets and fully connected crfs. arXiv preprint arXiv:1412.7062, 2014. 6 +[12] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L Yuille. Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. IEEE transactions on pattern analysis and machine intelligence, 40(4):834-848, 2017. 8 +[13] Liang-Chieh Chen, Yukun Zhu, George Papandreou, Florian Schroff, and Hartwig Adam. Encoder-decoder with atrous separable convolution for semantic image segmentation. In Proceedings of the European conference on computer vision (ECCV), pages 801-818, 2018. 6, 8 + +[14] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 6 +[15] Camille Couprie, Leo Grady, Laurent Najman, and Hugues Talbot. Power watershed: A unifying graph-based optimization framework. IEEE transactions on pattern analysis and machine intelligence, 33(7):1384-1399, 2010. 2 +[16] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248–255. IEEE, 2009. 6 +[17] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 6, 8 +[18] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. International journal of computer vision, 88:303-308, 2009. 6 +[19] Leo Grady. Random walks for image segmentation. IEEE transactions on pattern analysis and machine intelligence, 28 (11):1768-1783, 2006. 2, 3, 7, 8 +[20] Yves Grandvalet and Yoshua Bengio. Semi-supervised learning by entropy minimization. Advances in neural information processing systems, 17, 2004. 2 +[21] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6 +[22] Tsung-Wei Ke, Jyh-Jing Hwang, and Stella X Yu. Universal weakly supervised segmentation by pixel-to-segment contrastive learning. arXiv preprint arXiv:2105.00957, 2021. 2, 8 +[23] Alexander Kolesnikov and Christoph H Lampert. Seed, expand and constrain: Three principles for weakly-supervised image segmentation. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part IV 14, pages 695-711. Springer, 2016. 1 +[24] Philipp Krahenbuhl and Vladlen Koltun. Efficient inference in fully connected CRFs with Gaussian edge potentials. Advances in neural information processing systems, 24, 2011. 2 +[25] V. Kulharia, S. Chandra, A. Agrawal, P. Torr, and A. Tyagi. Box2seg: Attention weighted loss and discriminative feature learning for weakly supervised segmentation. In ECCV'20. 1 +[26] Jungbeom Lee, Eunji Kim, Sungmin Lee, Jangho Lee, and Sungroh Yoon. Ficklenet: Weakly and semi-supervised semantic image segmentation using stochastic inference. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5267-5276, 2019. 2 +[27] Zhiyuan Liang, Tiancai Wang, Xiangyu Zhang, Jian Sun, and Jianbing Shen. Tree energy loss: Towards sparsely annotated semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16907-16916, 2022. 2, 3, 6, 8 + +[28] Di Lin, Jifeng Dai, Jiaya Jia, Kaiming He, and Jian Sun. Scribblesup: Scribble-supervised convolutional networks for semantic segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3159-3167, 2016. 1, 2, 6, 7, 8 +[29] Dmitrii Marin and Yuri Boykov. Robust trust region for weakly supervised segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6608-6618, 2021. 2, 7, 8 +[30] Dmitrii Marin, Meng Tang, Ismail Ben Ayed, and Yuri Boykov. Beyond gradient descent for regularized segmentation losses. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10187-10196, 2019. 2, 3, 7 +[31] Rafael Müller, Simon Kornblith, and Geoffrey E Hinton. When does label smoothing help? Advances in neural information processing systems, 32, 2019. 5 +[32] NSD. Natural Scenes Dataset [NSD]. https://www.kaggle.com/datasets/nitishabharathi/scene-classification, 2020.5 +[33] Zhiyi Pan, Peng Jiang, Yunhai Wang, Changhe Tu, and Anthony G Cohn. Scribble-supervised semantic segmentation by uncertainty reduction on neural representation and self-supervision on neural eigenspace. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7416-7425, 2021. 2, 8 +[34] Thomas Pock, Antonin Chambolle, Daniel Cremers, and Horst Bischof. A convex relaxation approach for computing minimal partitions. In IEEE Conference on Computer Vision and Pattern Recognition, pages 810-817, 2009. 2 +[35] Pradeep Ravikumar and John Lafferty. Quadratic programming relaxations for metric labeling and Markov Random Field MAP estimation. In The 23rd International Conference on Machine Learning, page 737-744, 2006. 2, 3 +[36] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zhoginov, and Liang-Chieh Chen. *Mobilenetv2: Inverted residuals and linear bottlenecks*. In *Proceedings of the IEEE conference on computer vision and pattern recognition*, pages 4510–4520, 2018. 6 +[37] Meng Tang, Abdelaziz Djelouah, Federico Perazzi, Yuri Boykov, and Christopher Schroers. Normalized cut loss for weakly-supervised cnn segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1818-1827, 2018. 1, 6, 7 +[38] Meng Tang, Federico Perazzi, Abdelaziz Djelouah, Ismail Ben Ayed, Christopher Schroers, and Yuri Boykov. On regularized losses for weakly-supervised cnn segmentation. In Proceedings of the European Conference on Computer Vision (ECCV), pages 507-522, 2018. 1, 2, 6, 7, 8 +[39] Olga Veksler. Efficient graph cut optimization for full crfs with quantized edges. IEEE transactions on pattern analysis and machine intelligence, 42(4):1005-1012, 2019. 7, 8 +[40] Olga Veksler and Yuri Boykov. Sparse non-local crf. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4493-4503, 2022. 2, 8 +[41] Bin Wang, Guojun Qi, Sheng Tang, Tianzhu Zhang, Yunchao Wei, Linghui Li, and Yongdong Zhang. Boundary perception + +guidance: A scribble-supervised semantic segmentation approach. In IJCAI International joint conference on artificial intelligence, 2019. 2, 8 +[42] Changwei Wang, Rongtao Xu, Shibiao Xu, Weiliang Meng, and Xiaopeng Zhang. Treating pseudo-labels generation as image matting for weakly supervised semantic segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 755-765, 2023. 2 +[43] Linshan Wu, Zhun Zhong, Jiayi Ma, Yunchao Wei, Hao Chen, Leyuan Fang, and Shutao Li. Modeling the label distributions for weakly-supervised semantic segmentation. arXiv preprint arXiv:2403.13225, 2024. 8 +[44] Jingshan Xu, Chuanwei Zhou, Zhen Cui, Chunyan Xu, Yuge Huang, Pengcheng Shen, Shaoxin Li, and Jian Yang. Scribblesupervised semantic segmentation inference. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15354-15363, 2021. 2, 3, 8 +[45] Christopher Zach, Christian Hane, and Marc Pollefeys. What is optimized in tight convex relaxations for multi-label problems? In IEEE Conference on Computer Vision and Pattern Recognition, pages 1664-1671, 2012. 2 +[46] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 633-641, 2017. 6 \ No newline at end of file diff --git a/2025/Soft Self-labeling and Potts Relaxations for Weakly-supervised Segmentation/images.zip b/2025/Soft Self-labeling and Potts Relaxations for Weakly-supervised Segmentation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..1cb350e76a2ac24fc1b5191e37488e2cf406dca5 --- /dev/null +++ b/2025/Soft Self-labeling and Potts Relaxations for Weakly-supervised Segmentation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:16abf872b23598d0f597d47608f99dac4387f47835565c0831ca6ecf600b17a3 +size 499684 diff --git a/2025/Soft Self-labeling and Potts Relaxations for Weakly-supervised Segmentation/layout.json b/2025/Soft Self-labeling and Potts Relaxations for Weakly-supervised Segmentation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..f2e48979ad979bf95a28bf45962a5bb8f5a38ad4 --- /dev/null +++ b/2025/Soft Self-labeling and Potts Relaxations for Weakly-supervised Segmentation/layout.json @@ -0,0 +1,11420 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 182, + 103, + 428, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 182, + 103, + 428, + 140 + ], + "spans": [ + { + "bbox": [ + 182, + 103, + 428, + 140 + ], + "type": "text", + "content": "Soft Self-labeling and Potts Relaxations for Weakly-Supervised Segmentation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 214, + 161, + 395, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 214, + 161, + 395, + 175 + ], + "spans": [ + { + "bbox": [ + 214, + 161, + 395, + 175 + ], + "type": "text", + "content": "Zhongwen Zhang & Yuri Boykov" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 229, + 176, + 382, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 229, + 176, + 382, + 189 + ], + "spans": [ + { + "bbox": [ + 229, + 176, + 382, + 189 + ], + "type": "text", + "content": "University of Waterloo, Canada" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 187, + 192, + 421, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 187, + 192, + 421, + 202 + ], + "spans": [ + { + "bbox": [ + 187, + 192, + 421, + 202 + ], + "type": "text", + "content": "z889zhan@uwaterloo.ca, yboykov@uwaterloo.ca" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "spans": [ + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 257, + 297, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 257, + 297, + 497 + ], + "spans": [ + { + "bbox": [ + 54, + 257, + 297, + 497 + ], + "type": "text", + "content": "We consider weakly supervised segmentation where only a fraction of pixels have ground truth labels (scribbles) and focus on a self-labeling approach optimizing relaxations of the standard unsupervised CRF/Potts loss on unlabeled pixels. While WSSS methods can directly optimize such losses via gradient descent, prior work suggests that higher-order optimization can improve network training by introducing hidden pseudo-labels and powerful CRF sub-problem solvers, e.g. graph cut. However, previously used hard pseudo-labels can not represent class uncertainty or errors, which motivates soft self-labeling. We derive a principled auxiliary loss and systematically evaluate standard and new CRF relaxations (convex and non-convex), neighborhood systems, and terms connecting network predictions with soft pseudolabels. We also propose a general continuous sub-problem solver. Using only standard architectures, soft self-labeling consistently improves scribble-based training and outperforms significantly more complex specialized WSSS systems. It can outperform full pixel-precise supervision. Our general ideas apply to other weakly-supervised problems/systems." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 499, + 293, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 499, + 293, + 510 + ], + "spans": [ + { + "bbox": [ + 55, + 499, + 293, + 510 + ], + "type": "text", + "content": "Keywords: soft pseudo-labels, collision cross-entropy, collision divergence" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 536, + 135, + 549 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 536, + 135, + 549 + ], + "spans": [ + { + "bbox": [ + 55, + 536, + 135, + 549 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 557, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 557, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 557, + 297, + 714 + ], + "type": "text", + "content": "Full supervision for semantic segmentation requires thousands of training images with complete pixel-accurate ground truth masks. Their high costs explain the interest in weakly-supervised approaches based on image-level class tags [4, 23], pixel-level scribbles [28, 37, 38], or boxes [25]. This paper is focused on weak supervision with scribbles, which we also call seeds or partial masks. While only slightly more expensive than image-level class tags, scribbles on less than " + }, + { + "bbox": [ + 55, + 557, + 297, + 714 + ], + "type": "inline_equation", + "content": "3\\%" + }, + { + "bbox": [ + 55, + 557, + 297, + 714 + ], + "type": "text", + "content": " of pixels were previously shown to achieve accuracy approaching full supervision without any modifications of the segmentation models. In contrast, tag supervision typically requires highly specialized systems and complex multi-stage training procedures, which are hard" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 232, + 555, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 232, + 555, + 316 + ], + "spans": [ + { + "bbox": [ + 313, + 232, + 555, + 316 + ], + "type": "text", + "content": "to reproduce. Our interest in the scribble-based approach is motivated by its practical simplicity and mathematical clarity. The corresponding methodologies are focused on the design of unsupervised or self-supervised loss functions and stronger optimization algorithms. The corresponding solutions are often general and can be used in different weakly-supervised applications." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 323, + 495, + 335 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 323, + 495, + 335 + ], + "spans": [ + { + "bbox": [ + 314, + 323, + 495, + 335 + ], + "type": "text", + "content": "1.1. Scribble-supervised segmentation" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 340, + 555, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 340, + 555, + 399 + ], + "spans": [ + { + "bbox": [ + 313, + 340, + 555, + 399 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 313, + 340, + 555, + 399 + ], + "type": "inline_equation", + "content": "\\Omega" + }, + { + "bbox": [ + 313, + 340, + 555, + 399 + ], + "type": "text", + "content": " denote a set of image pixels and " + }, + { + "bbox": [ + 313, + 340, + 555, + 399 + ], + "type": "inline_equation", + "content": "S\\subset \\Omega" + }, + { + "bbox": [ + 313, + 340, + 555, + 399 + ], + "type": "text", + "content": " denote a subset of pixels with ground truth labels, i.e. seeds or scribbles typically marked by mouse-controlled UI for image annotations. One example of image scribbles is in Fig.4(a). The ground truth label at any given pixel " + }, + { + "bbox": [ + 313, + 340, + 555, + 399 + ], + "type": "inline_equation", + "content": "i\\in S" + }, + { + "bbox": [ + 313, + 340, + 555, + 399 + ], + "type": "text", + "content": " is an integer" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 399, + 407, + 555, + 421 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 399, + 407, + 555, + 421 + ], + "spans": [ + { + "bbox": [ + 399, + 407, + 555, + 421 + ], + "type": "interline_equation", + "content": "\\bar {y} _ {i} \\in \\{1, \\dots , K \\} \\tag {1}", + "image_path": "bd4d4bb534ecfc8af078e48fc3161da7e704fffac68d1e48ff14c7e3357caae3.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 428, + 555, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 428, + 555, + 464 + ], + "spans": [ + { + "bbox": [ + 313, + 428, + 555, + 464 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 428, + 555, + 464 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 428, + 555, + 464 + ], + "type": "text", + "content": " is the number of classes including the background. Without much ambiguity, it is convenient to use the same notation " + }, + { + "bbox": [ + 313, + 428, + 555, + 464 + ], + "type": "inline_equation", + "content": "\\bar{y}_i" + }, + { + "bbox": [ + 313, + 428, + 555, + 464 + ], + "type": "text", + "content": " for the equivalent one-hot distribution" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 320, + 471, + 555, + 486 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 471, + 555, + 486 + ], + "spans": [ + { + "bbox": [ + 320, + 471, + 555, + 486 + ], + "type": "interline_equation", + "content": "\\bar {y} _ {i} \\equiv (\\bar {y} _ {i} ^ {1}, \\dots , \\bar {y} _ {i} ^ {K}) \\in \\Delta_ {0, 1} ^ {K} \\quad \\text {f o r} \\quad \\bar {y} _ {i} ^ {k} := [ k = \\bar {y} _ {i} ] \\tag {2}", + "image_path": "62ecdf37f79d240ee9e758c433c6ab77f9cf494537c431bd5e24ec34943283de.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 493, + 555, + 530 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 493, + 555, + 530 + ], + "spans": [ + { + "bbox": [ + 313, + 493, + 555, + 530 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 493, + 555, + 530 + ], + "type": "inline_equation", + "content": "[\\cdot ]\\in \\{0,1\\}" + }, + { + "bbox": [ + 313, + 493, + 555, + 530 + ], + "type": "text", + "content": " is the True operator. Set " + }, + { + "bbox": [ + 313, + 493, + 555, + 530 + ], + "type": "inline_equation", + "content": "\\Delta_{0,1}^{K}" + }, + { + "bbox": [ + 313, + 493, + 555, + 530 + ], + "type": "text", + "content": " represents " + }, + { + "bbox": [ + 313, + 493, + 555, + 530 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 493, + 555, + 530 + ], + "type": "text", + "content": " possible one-hot distributions, which are vertices of the " + }, + { + "bbox": [ + 313, + 493, + 555, + 530 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 493, + 555, + 530 + ], + "type": "text", + "content": "-class probability simplex" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 326, + 536, + 542, + 571 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 536, + 542, + 571 + ], + "spans": [ + { + "bbox": [ + 326, + 536, + 542, + 571 + ], + "type": "interline_equation", + "content": "\\Delta^ {K} := \\{p = (p ^ {1}, \\dots , p ^ {K}) \\mid p ^ {k} \\geq 0, \\sum_ {k = 1} ^ {K} p ^ {k} = 1 \\}", + "image_path": "a3d88f18eedc89bfc9699ee43aca95ae7c43ff935fe709d7fb77896efb874b3d.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 576, + 555, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 576, + 555, + 612 + ], + "spans": [ + { + "bbox": [ + 313, + 576, + 555, + 612 + ], + "type": "text", + "content": "representing all " + }, + { + "bbox": [ + 313, + 576, + 555, + 612 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 576, + 555, + 612 + ], + "type": "text", + "content": "-categorical distributions. The context of each specific expression should make it obvious if " + }, + { + "bbox": [ + 313, + 576, + 555, + 612 + ], + "type": "inline_equation", + "content": "\\bar{y}_i" + }, + { + "bbox": [ + 313, + 576, + 555, + 612 + ], + "type": "text", + "content": " is a class index (1) or the corresponding one-hot distribution (2)." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 613, + 555, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 613, + 555, + 648 + ], + "spans": [ + { + "bbox": [ + 313, + 613, + 555, + 648 + ], + "type": "text", + "content": "Loss functions for scribble-supervised segmentation typically use negative log-likelihoods (NLL) over scribbles " + }, + { + "bbox": [ + 313, + 613, + 555, + 648 + ], + "type": "inline_equation", + "content": "i \\in S \\subset \\Omega" + }, + { + "bbox": [ + 313, + 613, + 555, + 648 + ], + "type": "text", + "content": " with ground truth labels " + }, + { + "bbox": [ + 313, + 613, + 555, + 648 + ], + "type": "inline_equation", + "content": "\\bar{y}_i" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 408, + 655, + 555, + 681 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 408, + 655, + 555, + 681 + ], + "spans": [ + { + "bbox": [ + 408, + 655, + 555, + 681 + ], + "type": "interline_equation", + "content": "- \\sum_ {i \\in S} \\ln \\sigma_ {i} ^ {\\bar {y} _ {i}} \\tag {3}", + "image_path": "cf951e485ab7ce9844a9d8ff5addcb9d95c840bbb3d85969a95b0835d10d1713.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\sigma_{i} = (\\sigma_{i}^{1},\\dots,\\sigma_{i}^{K})\\in \\Delta^{K}" + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": " are model predictions at pixels " + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": ". This loss is also standard in full supervision" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "20244" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "inline_equation", + "content": "S = \\Omega" + }, + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "content": ". In this case, no other losses are necessary for training. However, in a weakly supervised setting the majority of pixels are unlabeled, and unsupervised losses are needed for " + }, + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "inline_equation", + "content": "i \\notin S" + }, + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 121, + 295, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 121, + 295, + 216 + ], + "spans": [ + { + "bbox": [ + 55, + 121, + 295, + 216 + ], + "type": "text", + "content": "The most common unsupervised loss in image segmentation is the Potts model and its relaxations. It is a pairwise loss defined on pairs of neighboring pixels " + }, + { + "bbox": [ + 55, + 121, + 295, + 216 + ], + "type": "inline_equation", + "content": "\\{i,j\\} \\in \\mathcal{N}" + }, + { + "bbox": [ + 55, + 121, + 295, + 216 + ], + "type": "text", + "content": " for a given neighborhood system " + }, + { + "bbox": [ + 55, + 121, + 295, + 216 + ], + "type": "inline_equation", + "content": "\\mathcal{N} \\subset \\Omega \\times \\Omega" + }, + { + "bbox": [ + 55, + 121, + 295, + 216 + ], + "type": "text", + "content": ", typically corresponding to the nearest-neighbor grid (NN) [7, 19], or other sparse (SN) [40] or dense neighborhoods (DN) [24]. The original Potts model is defined for discrete segmentation variables, e.g. as in" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 221, + 285, + 249 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 221, + 285, + 249 + ], + "spans": [ + { + "bbox": [ + 67, + 221, + 285, + 249 + ], + "type": "interline_equation", + "content": "\\sum_ {\\{i, j \\} \\in \\mathcal {N}} P (\\sigma_ {i}, \\sigma_ {j}) \\quad \\text {w h e r e} \\quad P (\\sigma_ {i}, \\sigma_ {j}) = [ \\sigma_ {i} \\neq \\sigma_ {j} ]", + "image_path": "9fde796b4292c6f22a10b71de7cb077e58105db67701afafab16092abb6aaf94.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 255, + 296, + 291 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 255, + 296, + 291 + ], + "spans": [ + { + "bbox": [ + 55, + 255, + 296, + 291 + ], + "type": "text", + "content": "assuming integer-valued one-hot predictions " + }, + { + "bbox": [ + 55, + 255, + 296, + 291 + ], + "type": "inline_equation", + "content": "\\sigma_{i}\\in \\Delta_{0,1}^{K}" + }, + { + "bbox": [ + 55, + 255, + 296, + 291 + ], + "type": "text", + "content": " This regularization loss encourages smoothness between the pixels. Its popular self-supervised variant is" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 117, + 297, + 233, + 311 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 117, + 297, + 233, + 311 + ], + "spans": [ + { + "bbox": [ + 117, + 297, + 233, + 311 + ], + "type": "interline_equation", + "content": "P \\left(\\sigma_ {i}, \\sigma_ {j}\\right) = w _ {i, j} \\cdot \\left[ \\sigma_ {i} \\neq \\sigma_ {j} \\right]", + "image_path": "6ca7c056eb852256c5e912acd46bb3d9b956d3e031096dd341e5075c7c71cb6d.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 316, + 296, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 316, + 296, + 388 + ], + "spans": [ + { + "bbox": [ + 55, + 316, + 296, + 388 + ], + "type": "text", + "content": "where pairwise affinities " + }, + { + "bbox": [ + 55, + 316, + 296, + 388 + ], + "type": "inline_equation", + "content": "w_{ij}" + }, + { + "bbox": [ + 55, + 316, + 296, + 388 + ], + "type": "text", + "content": " are based on local intensity edges [7, 19, 24]. Of course, in the context of network training, one should use relaxations of " + }, + { + "bbox": [ + 55, + 316, + 296, + 388 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 55, + 316, + 296, + 388 + ], + "type": "text", + "content": " applicable to soft predictions " + }, + { + "bbox": [ + 55, + 316, + 296, + 388 + ], + "type": "inline_equation", + "content": "\\sigma_i \\in \\Delta^K" + }, + { + "bbox": [ + 55, + 316, + 296, + 388 + ], + "type": "text", + "content": ". Many forms of such relaxations [35, 45] were studied in segmentation, e.g. quadratic [19], bi-linear [38], total variation [9, 34], and others [15]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 388, + 296, + 423 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 388, + 296, + 423 + ], + "spans": [ + { + "bbox": [ + 55, + 388, + 296, + 423 + ], + "type": "text", + "content": "Another general unsupervised loss relevant to training segmentation networks is the entropy of predictions, which is also known as decidiveness [8, 20]" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 153, + 429, + 198, + 453 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 153, + 429, + 198, + 453 + ], + "spans": [ + { + "bbox": [ + 153, + 429, + 198, + 453 + ], + "type": "interline_equation", + "content": "\\sum_ {i} H (\\sigma_ {i})", + "image_path": "89e8bdf8ef3f51557e07b43586ec0e2da37e14339de05dc5f30cb5fddba8e07d.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 459, + 296, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 459, + 296, + 542 + ], + "spans": [ + { + "bbox": [ + 55, + 459, + 296, + 542 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 459, + 296, + 542 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 55, + 459, + 296, + 542 + ], + "type": "text", + "content": " is the Shannon's entropy function. This loss is known to improve generalization and the quality of representation by moving (deep) features away from the decision boundaries [20]. Widely known in the context of unsupervised or semi-supervised classification, this loss also matters in weakly-supervised segmentation where it is used explicitly or implicitly1." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 543, + 296, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 543, + 296, + 615 + ], + "spans": [ + { + "bbox": [ + 55, + 543, + 296, + 615 + ], + "type": "text", + "content": "Weakly-supervised segmentation methods [10, 22, 33, 41] also use other unsupervised losses (e.g. contrastive), clustering criteria (e.g. K-means), highly-specialized architectures, and ad-hoc training procedures. Our goal is to demonstrate that better results can be easily achieved on standard architectures by combining the standard losses above" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 61, + 632, + 296, + 647 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 632, + 296, + 647 + ], + "spans": [ + { + "bbox": [ + 61, + 632, + 296, + 647 + ], + "type": "interline_equation", + "content": "- \\sum_ {i \\in S} \\ln \\sigma_ {i} ^ {\\bar {y} _ {i}} + \\eta \\sum_ {i \\notin S} H (\\sigma_ {i}) + \\lambda \\sum_ {i j \\in \\mathcal {N}} P (\\sigma_ {i}, \\sigma_ {j}). \\tag {4}", + "image_path": "85930f7fde0fca04929a18fdcafc632744e41042a9d90dc9f533e26f1c3a4bc4.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 651, + 295, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 651, + 295, + 687 + ], + "spans": [ + { + "bbox": [ + 55, + 651, + 295, + 687 + ], + "type": "text", + "content": "but replacing basic gradient descent optimization [38, 40] by higher-order self-labeling techniques [28-30] incorporating optimization of auxiliary pseudo-labels as sub-problems." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 315, + 72, + 507, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 72, + 507, + 85 + ], + "spans": [ + { + "bbox": [ + 315, + 72, + 507, + 85 + ], + "type": "text", + "content": "1.2. Self-labeling and hard pseudo-labels" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 89, + 555, + 148 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 89, + 555, + 148 + ], + "spans": [ + { + "bbox": [ + 313, + 89, + 555, + 148 + ], + "type": "text", + "content": "One motivation for the self-labeling approach to weakly-supervised segmentation comes from well-known limitations of gradient descent for optimizing the Potts relaxations. Even convex relaxations [9, 19, 34] could be challenging in combination with the concave entropy term in (4)." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 149, + 555, + 280 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 149, + 555, + 280 + ], + "spans": [ + { + "bbox": [ + 313, + 149, + 555, + 280 + ], + "type": "text", + "content": "Typical self-labeling methods, including one of the first works on scribble-based semantic segmentation [28], introduce a sub-problem focused on the estimation of pseudolabels for unlabeled points. This subproblem can be formally \"derived\" from the original network optimization problem or \"designed\" as a heuristic, see the overview below. We focus on principled auxiliary loss formulations where network parameters and pseudo-labels are jointly optimized in a provably convergent manner. In this case pseudo-labels work as hidden optimization variables simplifying the model estimation, e.g. as in the EM algorithm for GMM." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 281, + 555, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 281, + 555, + 353 + ], + "spans": [ + { + "bbox": [ + 313, + 281, + 555, + 353 + ], + "type": "text", + "content": "We denote pseudo-labels " + }, + { + "bbox": [ + 313, + 281, + 555, + 353 + ], + "type": "inline_equation", + "content": "y_{i}" + }, + { + "bbox": [ + 313, + 281, + 555, + 353 + ], + "type": "text", + "content": " differently from the ground truth labels " + }, + { + "bbox": [ + 313, + 281, + 555, + 353 + ], + "type": "inline_equation", + "content": "\\bar{y}_{i}" + }, + { + "bbox": [ + 313, + 281, + 555, + 353 + ], + "type": "text", + "content": " by omitting the bar. It is important to distinguish them since " + }, + { + "bbox": [ + 313, + 281, + 555, + 353 + ], + "type": "inline_equation", + "content": "\\bar{y}_{i}" + }, + { + "bbox": [ + 313, + 281, + 555, + 353 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 313, + 281, + 555, + 353 + ], + "type": "inline_equation", + "content": "i \\in S" + }, + { + "bbox": [ + 313, + 281, + 555, + 353 + ], + "type": "text", + "content": " are given, while " + }, + { + "bbox": [ + 313, + 281, + 555, + 353 + ], + "type": "inline_equation", + "content": "y_{i}" + }, + { + "bbox": [ + 313, + 281, + 555, + 353 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 313, + 281, + 555, + 353 + ], + "type": "inline_equation", + "content": "i \\in \\Omega \\backslash S" + }, + { + "bbox": [ + 313, + 281, + 555, + 353 + ], + "type": "text", + "content": " are estimated. The majority of existing self-labeling methods [2, 3, 26, 28-30, 42] estimate hard pseudo-labels, which could be equivalently represented either by class indices" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 399, + 359, + 555, + 372 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 399, + 359, + 555, + 372 + ], + "spans": [ + { + "bbox": [ + 399, + 359, + 555, + 372 + ], + "type": "interline_equation", + "content": "y _ {i} \\in \\{1, \\dots , K ^ {\\prime} \\} \\tag {5}", + "image_path": "ed9397b70a20b3e4385485a392435d940c710fa9f49beec7702c004e630d9fb1.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 377, + 544, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 377, + 544, + 389 + ], + "spans": [ + { + "bbox": [ + 313, + 377, + 544, + 389 + ], + "type": "text", + "content": "or by the corresponding one-hot categorical distributions" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 320, + 395, + 555, + 409 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 395, + 555, + 409 + ], + "spans": [ + { + "bbox": [ + 320, + 395, + 555, + 409 + ], + "type": "interline_equation", + "content": "y _ {i} \\equiv \\left(y _ {i} ^ {1}, \\dots , y _ {i} ^ {K}\\right) \\in \\Delta_ {0, 1} ^ {K} \\quad \\text {f o r} \\quad y _ {i} ^ {k} := [ k = y _ {i} ] \\tag {6}", + "image_path": "36211a92848b2fcb55674632614646630881d497114ae97351cf101ac7f1e994.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 415, + 555, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 415, + 555, + 485 + ], + "spans": [ + { + "bbox": [ + 313, + 415, + 555, + 485 + ], + "type": "text", + "content": "analogously with the hard ground truth labels in (1) and (2). In part, hard pseudo-labels are motivated by the network training where the default is NLL loss (3) assuming discrete labels. Besides, powerful discrete solvers for the Potts model are well-known [7, 9, 34]. We discuss the potential advantages of soft pseudo-labels in Section 1.3." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 486, + 555, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 486, + 555, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 486, + 555, + 605 + ], + "type": "text", + "content": "Overview: joint loss vs \"proposal generation\": Most self-labeling approaches can be divided into two groups. One group designs pseudo-labeling and the network training sub-problems that are not formally related, e.g. [27, 28, 44]. While pseudo-labeling typically depends on the current network predictions and the network fine-tuning uses such pseudo-labels, the lack of a formal relation between these sub-problems implies that iterating such steps does not guarantee any form of convergence. Such methods are often referred to as proposal generation heuristics." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 606, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 606, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 606, + 556, + 714 + ], + "type": "text", + "content": "Alternatively, the pseudo-labeling sub-problem and the network training sub-problem can be formally derived from a weakly-supervised loss like (4), e.g. by ADM splitting [30] or as high-order trust-region method [29]. Such methods often formulate a joint loss w.r.t. network predictions and pseudo-labels and iteratively decrease the loss. These self-labeling methods are mathematically well-founded, easy to understand and reproduce, and numerically stable due to guaranteed convergence." + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "type": "text", + "content": "1A unary decidiveness-like term is the difference between quadratic and bi-linear (tight but non-convex) relaxations [29, 35] of the Potts model." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20245" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 296, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 296, + 85 + ], + "type": "text", + "content": "1.3. Soft pseudo-labels: motivation & contributions" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 89, + 297, + 138 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 89, + 297, + 138 + ], + "spans": [ + { + "bbox": [ + 55, + 89, + 297, + 138 + ], + "type": "text", + "content": "We observe that self-labeling with hard pseudo-labels " + }, + { + "bbox": [ + 55, + 89, + 297, + 138 + ], + "type": "inline_equation", + "content": "y_{i}" + }, + { + "bbox": [ + 55, + 89, + 297, + 138 + ], + "type": "text", + "content": " is inherently limited as such labels cannot represent the uncertainty of class estimates at unlabeled pixels " + }, + { + "bbox": [ + 55, + 89, + 297, + 138 + ], + "type": "inline_equation", + "content": "i \\in \\Omega \\backslash S" + }, + { + "bbox": [ + 55, + 89, + 297, + 138 + ], + "type": "text", + "content": ". Instead, we focus on soft pseudo-labels" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 116, + 145, + 296, + 160 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 145, + 296, + 160 + ], + "spans": [ + { + "bbox": [ + 116, + 145, + 296, + 160 + ], + "type": "interline_equation", + "content": "y _ {i} = \\left(y _ {i} ^ {1}, \\dots , y _ {i} ^ {K}\\right) \\in \\Delta^ {K} \\tag {7}", + "image_path": "655de9dc2bec543fdecf418e99322ae26db5ab93ba34a2ea24d530b3243f1a65.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 168, + 296, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 168, + 296, + 263 + ], + "spans": [ + { + "bbox": [ + 55, + 168, + 296, + 263 + ], + "type": "text", + "content": "which are general categorical distributions " + }, + { + "bbox": [ + 55, + 168, + 296, + 263 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 55, + 168, + 296, + 263 + ], + "type": "text", + "content": " over " + }, + { + "bbox": [ + 55, + 168, + 296, + 263 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 168, + 296, + 263 + ], + "type": "text", + "content": "-classes. It is possible that the estimated pseudo-label " + }, + { + "bbox": [ + 55, + 168, + 296, + 263 + ], + "type": "inline_equation", + "content": "y_{i}" + }, + { + "bbox": [ + 55, + 168, + 296, + 263 + ], + "type": "text", + "content": " in (7) could be a one-hot distribution, which is a vertex of " + }, + { + "bbox": [ + 55, + 168, + 296, + 263 + ], + "type": "inline_equation", + "content": "\\Delta^K" + }, + { + "bbox": [ + 55, + 168, + 296, + 263 + ], + "type": "text", + "content": ". In this special case, " + }, + { + "bbox": [ + 55, + 168, + 296, + 263 + ], + "type": "inline_equation", + "content": "y_{i}" + }, + { + "bbox": [ + 55, + 168, + 296, + 263 + ], + "type": "text", + "content": " can be treated as a class index, but we use the general formulation (7) in the main parts of the paper starting Section 2. On the other hand, the ground truth labels " + }, + { + "bbox": [ + 55, + 168, + 296, + 263 + ], + "type": "inline_equation", + "content": "\\bar{y}_{i}" + }, + { + "bbox": [ + 55, + 168, + 296, + 263 + ], + "type": "text", + "content": " are always hard and we use them either as indices (1) or one-hot distributions (2), as convenient." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 264, + 296, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 264, + 296, + 467 + ], + "spans": [ + { + "bbox": [ + 55, + 264, + 296, + 467 + ], + "type": "text", + "content": "Soft pseudo-labels can be found in prior work on weakly-supervised segmentation [27, 44] using the \"soft proposal generation\". In contrast, we formulate soft self-labeling as a principled optimization methodology where network predictions and soft pseudo-labels are variables in a joint loss, which guarantees convergence of the training procedure. Our pseudo-labels are auxiliary variables for ADM-based [5] splitting of the loss (4) into two simpler optimization sub-problems: one focused on the Potts model over unlabeled pixels, and the other on the network training. While similar to [30], instead of hard, we use soft auxiliary variables for the Potts sub-problem. Our work can be seen as a study of the relaxed Potts sub-problem in the context of weakly-supervised semantic segmentation. The related prior work is focused on discrete solvers fundamentally unable to represent class estimate uncertainty. Our contributions can be summarized as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 468, + 296, + 598 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 55, + 468, + 295, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 468, + 295, + 491 + ], + "spans": [ + { + "bbox": [ + 55, + 468, + 295, + 491 + ], + "type": "text", + "content": "- convergent soft self-labeling framework based on a simple joint self-labeling loss" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 491, + 296, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 491, + 296, + 514 + ], + "spans": [ + { + "bbox": [ + 56, + 491, + 296, + 514 + ], + "type": "text", + "content": "- systematic evaluation of Potts relaxations and (cross-) entropy terms in our loss" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 516, + 296, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 516, + 296, + 563 + ], + "spans": [ + { + "bbox": [ + 56, + 516, + 296, + 563 + ], + "type": "text", + "content": "- state-of-the-art in scribble-based semantic segmentation based on standard network architectures (no modifications) and well-founded convergent training procedures that are easy to reproduce and generalize to other problems" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 563, + 295, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 563, + 295, + 598 + ], + "spans": [ + { + "bbox": [ + 56, + 563, + 295, + 598 + ], + "type": "text", + "content": "- using the same segmentation model, our self-labeling loss with " + }, + { + "bbox": [ + 56, + 563, + 295, + 598 + ], + "type": "inline_equation", + "content": "3\\%" + }, + { + "bbox": [ + 56, + 563, + 295, + 598 + ], + "type": "text", + "content": " scribbles may outperform standard supervised cross-entropy loss with full ground truth masks." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 609, + 230, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 609, + 230, + 624 + ], + "spans": [ + { + "bbox": [ + 55, + 609, + 230, + 624 + ], + "type": "text", + "content": "2. Our soft self-labeling approach" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 629, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 297, + 714 + ], + "type": "text", + "content": "First, we apply ADM splitting [5] to weakly supervised loss (4) to formulate our self-labeling loss (8) incorporating additional soft auxiliary variables, i.e. pseudo-labels (7). It is convenient to introduce pseudo-labels " + }, + { + "bbox": [ + 55, + 629, + 297, + 714 + ], + "type": "inline_equation", + "content": "y_{i}" + }, + { + "bbox": [ + 55, + 629, + 297, + 714 + ], + "type": "text", + "content": " on all pixels in " + }, + { + "bbox": [ + 55, + 629, + 297, + 714 + ], + "type": "inline_equation", + "content": "\\Omega" + }, + { + "bbox": [ + 55, + 629, + 297, + 714 + ], + "type": "text", + "content": " even though a subset of pixels (seeds) " + }, + { + "bbox": [ + 55, + 629, + 297, + 714 + ], + "type": "inline_equation", + "content": "S\\subset \\Omega" + }, + { + "bbox": [ + 55, + 629, + 297, + 714 + ], + "type": "text", + "content": " have ground truth labels " + }, + { + "bbox": [ + 55, + 629, + 297, + 714 + ], + "type": "inline_equation", + "content": "\\bar{y}_i" + }, + { + "bbox": [ + 55, + 629, + 297, + 714 + ], + "type": "text", + "content": ". We will simply impose a constraint that pseudo-labels and ground truth labels agree on " + }, + { + "bbox": [ + 55, + 629, + 297, + 714 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 55, + 629, + 297, + 714 + ], + "type": "text", + "content": ". Thus," + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 72, + 499, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 499, + 84 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 499, + 84 + ], + "type": "text", + "content": "we assume the following set of pseudo-labels" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 333, + 91, + 534, + 106 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 91, + 534, + 106 + ], + "spans": [ + { + "bbox": [ + 333, + 91, + 534, + 106 + ], + "type": "interline_equation", + "content": "Y _ {\\Omega} := \\left\\{y _ {i} \\in \\Delta^ {K} \\mid i \\in \\Omega , \\text {s . t .} y _ {i} = \\bar {y} _ {i} \\text {f o r} i \\in S \\right\\}.", + "image_path": "0817152056bf076ad5a05008991d87504a7580e934d5da62353e1de91dc96bad.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 113, + 555, + 173 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 113, + 555, + 173 + ], + "spans": [ + { + "bbox": [ + 313, + 113, + 555, + 173 + ], + "type": "text", + "content": "We split the terms in (4) into two groups: one includes NLL and entropy " + }, + { + "bbox": [ + 313, + 113, + 555, + 173 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 313, + 113, + 555, + 173 + ], + "type": "text", + "content": " terms keeping the original prediction variables " + }, + { + "bbox": [ + 313, + 113, + 555, + 173 + ], + "type": "inline_equation", + "content": "\\sigma_{i}" + }, + { + "bbox": [ + 313, + 113, + 555, + 173 + ], + "type": "text", + "content": " and the other includes the Potts relaxation " + }, + { + "bbox": [ + 313, + 113, + 555, + 173 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 313, + 113, + 555, + 173 + ], + "type": "text", + "content": " replacing " + }, + { + "bbox": [ + 313, + 113, + 555, + 173 + ], + "type": "inline_equation", + "content": "\\sigma_{i}" + }, + { + "bbox": [ + 313, + 113, + 555, + 173 + ], + "type": "text", + "content": " with auxiliary variables " + }, + { + "bbox": [ + 313, + 113, + 555, + 173 + ], + "type": "inline_equation", + "content": "y_{i}" + }, + { + "bbox": [ + 313, + 113, + 555, + 173 + ], + "type": "text", + "content": ". This transforms loss (4) into expression" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 333, + 180, + 536, + 207 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 180, + 536, + 207 + ], + "spans": [ + { + "bbox": [ + 333, + 180, + 536, + 207 + ], + "type": "interline_equation", + "content": "- \\sum_ {i \\in S} \\ln \\sigma_ {i} ^ {\\bar {y} _ {i}} + \\eta \\sum_ {i \\not \\in S} H (\\sigma_ {i}) + \\lambda \\sum_ {i j \\in \\mathcal {N}} P (y _ {i}, y _ {j})", + "image_path": "f534eef7b0d0195dee70332d8ce5ba5669b3a2273e80515602d13a4d06381323.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 213, + 555, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 213, + 555, + 266 + ], + "spans": [ + { + "bbox": [ + 313, + 213, + 555, + 266 + ], + "type": "text", + "content": "equivalent to (4) assuming equality " + }, + { + "bbox": [ + 313, + 213, + 555, + 266 + ], + "type": "inline_equation", + "content": "\\sigma_{i} = y_{i}" + }, + { + "bbox": [ + 313, + 213, + 555, + 266 + ], + "type": "text", + "content": ". The standard approximation is to incorporate constraint " + }, + { + "bbox": [ + 313, + 213, + 555, + 266 + ], + "type": "inline_equation", + "content": "\\sigma_{i} \\approx y_{i}" + }, + { + "bbox": [ + 313, + 213, + 555, + 266 + ], + "type": "text", + "content": " directly into the loss, e.g. using " + }, + { + "bbox": [ + 313, + 213, + 555, + 266 + ], + "type": "inline_equation", + "content": "KL" + }, + { + "bbox": [ + 313, + 213, + 555, + 266 + ], + "type": "text", + "content": "-divergence. For simplicity, we use weight " + }, + { + "bbox": [ + 313, + 213, + 555, + 266 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 313, + 213, + 555, + 266 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 313, + 213, + 555, + 266 + ], + "type": "inline_equation", + "content": "KL(\\sigma_{i},y_{i})" + }, + { + "bbox": [ + 313, + 213, + 555, + 266 + ], + "type": "text", + "content": " to combine it with " + }, + { + "bbox": [ + 313, + 213, + 555, + 266 + ], + "type": "inline_equation", + "content": "H(\\sigma_{i})" + }, + { + "bbox": [ + 313, + 213, + 555, + 266 + ], + "type": "text", + "content": " into a single cross-entropy term" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 315, + 268, + 555, + 321 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 268, + 555, + 321 + ], + "spans": [ + { + "bbox": [ + 315, + 268, + 555, + 321 + ], + "type": "interline_equation", + "content": "-\\sum_{i\\in S}\\ln \\sigma_{i}^{\\bar{y}_{i}} + \\underbrace{\\eta\\sum_{i\\notin S}H(\\sigma_{i}) + \\eta\\sum_{i\\notin S}KL(\\sigma_{i},y_{i})}_{\\substack{\\eta \\sum_{i\\notin S}H(\\sigma_{i},y_{i})}} + \\lambda \\sum_{ij\\in \\mathcal{N}}P(y_{i},y_{j})", + "image_path": "5a3458752755acd328a41dbbc016e5e07f9c9888ed718b3f73964ef2f9df7dcb.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 325, + 542, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 325, + 542, + 338 + ], + "spans": [ + { + "bbox": [ + 316, + 325, + 542, + 338 + ], + "type": "text", + "content": "defining auxiliary self-labeling loss w.r.t. both " + }, + { + "bbox": [ + 316, + 325, + 542, + 338 + ], + "type": "inline_equation", + "content": "\\sigma_{i}" + }, + { + "bbox": [ + 316, + 325, + 542, + 338 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 316, + 325, + 542, + 338 + ], + "type": "inline_equation", + "content": "y_{i}" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 320, + 345, + 555, + 360 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 345, + 555, + 360 + ], + "spans": [ + { + "bbox": [ + 320, + 345, + 555, + 360 + ], + "type": "interline_equation", + "content": "- \\sum_ {i \\in S} \\ln \\sigma_ {i} ^ {\\bar {y} _ {i}} + \\eta \\sum_ {i \\notin S} H \\left(\\sigma_ {i}, y _ {i}\\right) + \\lambda \\sum_ {i j \\in \\mathcal {N}} P \\left(y _ {i}, y _ {j}\\right) \\tag {8}", + "image_path": "2ccd82f6dd57b85640cdd3f219d4f81aaa0b78e08d5a593702b60bd1f0c0c76a.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 366, + 534, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 366, + 534, + 378 + ], + "spans": [ + { + "bbox": [ + 313, + 366, + 534, + 378 + ], + "type": "text", + "content": "approximating the original weakly supervised loss (4)." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 379, + 555, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 379, + 555, + 461 + ], + "spans": [ + { + "bbox": [ + 313, + 379, + 555, + 461 + ], + "type": "text", + "content": "Iterative minimization of this loss w.r.t. predictions " + }, + { + "bbox": [ + 313, + 379, + 555, + 461 + ], + "type": "inline_equation", + "content": "\\sigma_{i}" + }, + { + "bbox": [ + 313, + 379, + 555, + 461 + ], + "type": "text", + "content": " (model parameters training) and pseudo-labels " + }, + { + "bbox": [ + 313, + 379, + 555, + 461 + ], + "type": "inline_equation", + "content": "y_{i}" + }, + { + "bbox": [ + 313, + 379, + 555, + 461 + ], + "type": "text", + "content": " effectively breaks the original optimization problem for (4) into two simpler sub-problems, assuming there is a good solver for optimal pseudo-labels. The latter is plausible since the unary term " + }, + { + "bbox": [ + 313, + 379, + 555, + 461 + ], + "type": "inline_equation", + "content": "H(\\sigma_{i}, y_{i})" + }, + { + "bbox": [ + 313, + 379, + 555, + 461 + ], + "type": "text", + "content": " is convex for " + }, + { + "bbox": [ + 313, + 379, + 555, + 461 + ], + "type": "inline_equation", + "content": "y_{i}" + }, + { + "bbox": [ + 313, + 379, + 555, + 461 + ], + "type": "text", + "content": ", and the Potts model (incl. convex and non-convex relaxations) is widely studied." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 462, + 556, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 462, + 556, + 558 + ], + "spans": [ + { + "bbox": [ + 313, + 462, + 556, + 558 + ], + "type": "text", + "content": "Section 2.1 discusses standard and new relaxations of the Potts model " + }, + { + "bbox": [ + 313, + 462, + 556, + 558 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 313, + 462, + 556, + 558 + ], + "type": "text", + "content": ". Section 2.2 discusses several robust variants of the cross-entropy term " + }, + { + "bbox": [ + 313, + 462, + 556, + 558 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 313, + 462, + 556, + 558 + ], + "type": "text", + "content": " in (8) connecting predictions with uncertain (soft) pseudo-labels " + }, + { + "bbox": [ + 313, + 462, + 556, + 558 + ], + "type": "inline_equation", + "content": "y_{i}" + }, + { + "bbox": [ + 313, + 462, + 556, + 558 + ], + "type": "text", + "content": " at unlabeled points " + }, + { + "bbox": [ + 313, + 462, + 556, + 558 + ], + "type": "inline_equation", + "content": "i \\in \\Omega \\backslash S" + }, + { + "bbox": [ + 313, + 462, + 556, + 558 + ], + "type": "text", + "content": ". To systematically evaluate these variants of (8) in Section 3, we propose a general efficient GPU-friendly solver for the corresponding pseudo-labeling sub-problem, which is detailed in the Supplementary Material." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 564, + 541, + 576 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 564, + 541, + 576 + ], + "spans": [ + { + "bbox": [ + 313, + 564, + 541, + 576 + ], + "type": "text", + "content": "2.1. Second-order relaxations of the Potts model" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 582, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 556, + 714 + ], + "type": "text", + "content": "We limit our focus to basic second-order relaxations and several motivated extensions. We start from two important standard cases (see Table 1): quadratic, the simplest convex relaxation popularized by the random walker algorithm [19], and bi-linear, which is non-convex but tight [35] w.r.t. the original discrete Potts model. The latter implies that optimizing it over relaxed variables should lead to a solution consistent with a discrete Potts solver, e.g. graph cut [7] regularizing the geometric boundary of the segments [6]. The quadratic relaxation produces a significantly different soft solution motivated by random walk probabilities [19]." + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "20246" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 77, + 69, + 276, + 127 + ], + "blocks": [ + { + "bbox": [ + 77, + 69, + 276, + 127 + ], + "lines": [ + { + "bbox": [ + 77, + 69, + 276, + 127 + ], + "spans": [ + { + "bbox": [ + 77, + 69, + 276, + 127 + ], + "type": "table", + "html": "
bi-linear ~ “graph cut”PBL(p,q) := 1 - pTqquadratic ~ “random walker”P0(p,q) := 1/2||p-q|2
normalized quadratic
PNQ(p,q) := 1 - pTq/||p||q|≡ 1/2||p/||p-| - q/||q| |2
", + "image_path": "db6efe803d38739d8276fdcbf261bbc8ebc147133bfc24194b29bbf4fb5729d3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 156, + 296, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 156, + 296, + 274 + ], + "spans": [ + { + "bbox": [ + 55, + 156, + 296, + 274 + ], + "type": "text", + "content": "While there are specialized solvers for bi-linear and quadratic relaxations, we are interested in a general efficient pseudo-label solver integrated into the standard GPU-based network optimization framework. This suggests gradient descent as an optimizer for pseudo-labels " + }, + { + "bbox": [ + 55, + 156, + 296, + 274 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 55, + 156, + 296, + 274 + ], + "type": "text", + "content": " (see Suppl. Mat). Figure 1 shows two representative local minima issues that gradient descent may encounter for (a) the bi-linear and (b) quadratic relaxations of the Potts loss. These examples indicate that gradient descent for these relaxations can distort the properties of the original (discrete) Potts loss." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 276, + 296, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 276, + 296, + 361 + ], + "spans": [ + { + "bbox": [ + 55, + 276, + 296, + 361 + ], + "type": "text", + "content": "In (a) two neighboring pixels jointly change the common label from " + }, + { + "bbox": [ + 55, + 276, + 296, + 361 + ], + "type": "inline_equation", + "content": "y_{i} = y_{j} = (1,0,0)" + }, + { + "bbox": [ + 55, + 276, + 296, + 361 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 276, + 296, + 361 + ], + "type": "inline_equation", + "content": "y_{i}^{\\prime \\prime} = y_{j}^{\\prime \\prime} = (0,1,0)" + }, + { + "bbox": [ + 55, + 276, + 296, + 361 + ], + "type": "text", + "content": ", which corresponds to a \"move\" where the whole object is reclassified from A to B. This move does not violate the smoothness constraint represented by the original (discrete) Potts model. But, bilinear relaxation prevents move (a) since the intermediate state " + }, + { + "bbox": [ + 55, + 276, + 296, + 361 + ], + "type": "inline_equation", + "content": "y_{i}^{\\prime} = y_{j}^{\\prime} = (\\frac{1}{2},\\frac{1}{2},0)" + }, + { + "bbox": [ + 55, + 276, + 296, + 361 + ], + "type": "text", + "content": " has a higher cost" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 69, + 369, + 282, + 393 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 369, + 282, + 393 + ], + "spans": [ + { + "bbox": [ + 69, + 369, + 282, + 393 + ], + "type": "interline_equation", + "content": "P _ {\\mathrm {B L}} \\left(y _ {i} ^ {\\prime}, y _ {j} ^ {\\prime}\\right) = \\frac {1}{2} > 0 = P _ {\\mathrm {B L}} \\left(y _ {i}, y _ {j}\\right) = P _ {\\mathrm {B L}} \\left(y _ {i} ^ {\\prime \\prime}, y _ {j} ^ {\\prime \\prime}\\right)", + "image_path": "c0adfe793ba72d38403173e5e8a97e79a5b9c039e513a53f60cbc6a7298ddbbe.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 401, + 295, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 401, + 295, + 425 + ], + "spans": [ + { + "bbox": [ + 55, + 401, + 295, + 425 + ], + "type": "text", + "content": "while quadratic relaxation assigns constant (zero) loss for all states during this move." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 426, + 296, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 426, + 296, + 510 + ], + "spans": [ + { + "bbox": [ + 55, + 426, + 296, + 510 + ], + "type": "text", + "content": "Figure 1(b) shows a move problematic for the quadratic relaxation. Two neighboring pixels on the boundary of objects A and C have labels " + }, + { + "bbox": [ + 55, + 426, + 296, + 510 + ], + "type": "inline_equation", + "content": "y_{i} = (1,0,0)" + }, + { + "bbox": [ + 55, + 426, + 296, + 510 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 426, + 296, + 510 + ], + "type": "inline_equation", + "content": "y_{j} = (0,0,1)" + }, + { + "bbox": [ + 55, + 426, + 296, + 510 + ], + "type": "text", + "content": ". The second object attempts to change from class label C to B that does not affect the Potts discontinuity penalty between two pixels. But, quadratic relaxation prefers that the second object stays in the intermediate soft state " + }, + { + "bbox": [ + 55, + 426, + 296, + 510 + ], + "type": "inline_equation", + "content": "y_{j}^{\\prime} = (0,\\frac{1}{2},\\frac{1}{2})" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 74, + 519, + 276, + 544 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 519, + 276, + 544 + ], + "spans": [ + { + "bbox": [ + 74, + 519, + 276, + 544 + ], + "type": "interline_equation", + "content": "P _ {\\mathrm {Q}} (y _ {i}, y _ {j} ^ {\\prime}) = \\frac {3}{4} < 1 = P _ {\\mathrm {Q}} (y _ {i}, y _ {j}) = P _ {\\mathrm {Q}} (y _ {i}, y _ {j} ^ {\\prime \\prime})", + "image_path": "3d2431aa3545e2235802c19f33fd0a7ff96763e0398ae19448d016d4307c18c6.jpg" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 90, + 562, + 164, + 622 + ], + "blocks": [ + { + "bbox": [ + 90, + 562, + 164, + 622 + ], + "lines": [ + { + "bbox": [ + 90, + 562, + 164, + 622 + ], + "spans": [ + { + "bbox": [ + 90, + 562, + 164, + 622 + ], + "type": "image", + "image_path": "31f8396aa63da9926854ef470abb9e99765435b8f035951b0f7f713f97359515.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 82, + 625, + 171, + 635 + ], + "lines": [ + { + "bbox": [ + 82, + 625, + 171, + 635 + ], + "spans": [ + { + "bbox": [ + 82, + 625, + 171, + 635 + ], + "type": "text", + "content": "(a) both pixels change " + }, + { + "bbox": [ + 82, + 625, + 171, + 635 + ], + "type": "inline_equation", + "content": "(\\mathrm{A}\\rightarrow \\mathrm{B})" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 187, + 562, + 261, + 622 + ], + "blocks": [ + { + "bbox": [ + 187, + 562, + 261, + 622 + ], + "lines": [ + { + "bbox": [ + 187, + 562, + 261, + 622 + ], + "spans": [ + { + "bbox": [ + 187, + 562, + 261, + 622 + ], + "type": "image", + "image_path": "19a68655c2a8b548079376697f6c0101827163c2921d6edfbdde4a4e268d031e.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 178, + 625, + 272, + 635 + ], + "lines": [ + { + "bbox": [ + 178, + 625, + 272, + 635 + ], + "spans": [ + { + "bbox": [ + 178, + 625, + 272, + 635 + ], + "type": "text", + "content": "(b) only pixel " + }, + { + "bbox": [ + 178, + 625, + 272, + 635 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 178, + 625, + 272, + 635 + ], + "type": "text", + "content": " changes " + }, + { + "bbox": [ + 178, + 625, + 272, + 635 + ], + "type": "inline_equation", + "content": "(\\mathbf{C}\\rightarrow \\mathbf{B})" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 643, + 296, + 710 + ], + "lines": [ + { + "bbox": [ + 55, + 643, + 296, + 710 + ], + "spans": [ + { + "bbox": [ + 55, + 643, + 296, + 710 + ], + "type": "text", + "content": "Figure 1. Examples of \"moves\" for neighboring pixels " + }, + { + "bbox": [ + 55, + 643, + 296, + 710 + ], + "type": "inline_equation", + "content": "\\{i,j\\} \\in \\mathcal{N}" + }, + { + "bbox": [ + 55, + 643, + 296, + 710 + ], + "type": "text", + "content": ". Their (soft) pseudo-labels " + }, + { + "bbox": [ + 55, + 643, + 296, + 710 + ], + "type": "inline_equation", + "content": "y_{i}" + }, + { + "bbox": [ + 55, + 643, + 296, + 710 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 643, + 296, + 710 + ], + "type": "inline_equation", + "content": "y_{j}" + }, + { + "bbox": [ + 55, + 643, + 296, + 710 + ], + "type": "text", + "content": " are illustrated on the probability simplex " + }, + { + "bbox": [ + 55, + 643, + 296, + 710 + ], + "type": "inline_equation", + "content": "\\Delta^K" + }, + { + "bbox": [ + 55, + 643, + 296, + 710 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 55, + 643, + 296, + 710 + ], + "type": "inline_equation", + "content": "K = 3" + }, + { + "bbox": [ + 55, + 643, + 296, + 710 + ], + "type": "text", + "content": ". In (a) both pixels " + }, + { + "bbox": [ + 55, + 643, + 296, + 710 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 643, + 296, + 710 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 643, + 296, + 710 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 55, + 643, + 296, + 710 + ], + "type": "text", + "content": " are inside a region/object changing its label from A to B. In (b) pixels " + }, + { + "bbox": [ + 55, + 643, + 296, + 710 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 643, + 296, + 710 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 643, + 296, + 710 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 55, + 643, + 296, + 710 + ], + "type": "text", + "content": " are on the boundary between two regions/objects; one is fixed to class A and the other changes from class C to B." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 336, + 69, + 535, + 125 + ], + "blocks": [ + { + "bbox": [ + 73, + 136, + 277, + 147 + ], + "lines": [ + { + "bbox": [ + 73, + 136, + 277, + 147 + ], + "spans": [ + { + "bbox": [ + 73, + 136, + 277, + 147 + ], + "type": "text", + "content": "Table 1. Second-order Potts relaxations, see Fig.2(a,b,c)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 336, + 69, + 535, + 125 + ], + "lines": [ + { + "bbox": [ + 336, + 69, + 535, + 125 + ], + "spans": [ + { + "bbox": [ + 336, + 69, + 535, + 125 + ], + "type": "table", + "html": "
collision cross entropy\nPCC(r(p,q) := -ln p^T qlog-quadratic\nP_LQ(p,q) := -ln(1 - ||p-q||^2/2)
collision divergence\nPCD(p,q) := -ln ||p^T q/||p/||q||- ln(1 - 1/2 ||p/||p - q/||q||^2)
", + "image_path": "d4224969ca9d5d4e17edc5da938592645709130742acddba351cd688be4420ae.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 337, + 133, + 531, + 145 + ], + "lines": [ + { + "bbox": [ + 337, + 133, + 531, + 145 + ], + "spans": [ + { + "bbox": [ + 337, + 133, + 531, + 145 + ], + "type": "text", + "content": "Table 2. Log-based Potts relaxations, see Fig.2(d,e,f)" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 153, + 553, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 153, + 553, + 177 + ], + "spans": [ + { + "bbox": [ + 313, + 153, + 553, + 177 + ], + "type": "text", + "content": "while bi-linear relaxation " + }, + { + "bbox": [ + 313, + 153, + 553, + 177 + ], + "type": "inline_equation", + "content": "P_{\\mathrm{BL}}(y_i, y_j) = 1" + }, + { + "bbox": [ + 313, + 153, + 553, + 177 + ], + "type": "text", + "content": " remains constant as " + }, + { + "bbox": [ + 313, + 153, + 553, + 177 + ], + "type": "inline_equation", + "content": "y_j" + }, + { + "bbox": [ + 313, + 153, + 553, + 177 + ], + "type": "text", + "content": " transitions from state C to B." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 178, + 555, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 178, + 555, + 285 + ], + "spans": [ + { + "bbox": [ + 313, + 178, + 555, + 285 + ], + "type": "text", + "content": "We propose a new relaxation, normalized quadratic in Table 1. Normalization leads to equivalence between quadratic and bi-linear formulations combining their benefits. As easy to check, normalized quadratic relaxation " + }, + { + "bbox": [ + 313, + 178, + 555, + 285 + ], + "type": "inline_equation", + "content": "P_{\\mathrm{NQ}}" + }, + { + "bbox": [ + 313, + 178, + 555, + 285 + ], + "type": "text", + "content": " removes the local minima in both examples of Figure 1. Table 2 also proposes \"logarithmic\" versions of the relaxations in Table 1 composing them with function " + }, + { + "bbox": [ + 313, + 178, + 555, + 285 + ], + "type": "inline_equation", + "content": "-\\ln (1 - x)" + }, + { + "bbox": [ + 313, + 178, + 555, + 285 + ], + "type": "text", + "content": ". As illustrated by Figure 2, the logarithmic versions (d-f) address the \"vanishing gradients\" evident from the flat regions in (a-c)." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 293, + 510, + 306 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 293, + 510, + 306 + ], + "spans": [ + { + "bbox": [ + 313, + 293, + 510, + 306 + ], + "type": "text", + "content": "2.2. Cross-entropy and soft pseudo-labels" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 311, + 555, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 311, + 555, + 586 + ], + "spans": [ + { + "bbox": [ + 313, + 311, + 555, + 586 + ], + "type": "text", + "content": "Shannon's cross-entropy " + }, + { + "bbox": [ + 313, + 311, + 555, + 586 + ], + "type": "inline_equation", + "content": "H(y, \\sigma)" + }, + { + "bbox": [ + 313, + 311, + 555, + 586 + ], + "type": "text", + "content": " is the most common loss for training network predictions " + }, + { + "bbox": [ + 313, + 311, + 555, + 586 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 313, + 311, + 555, + 586 + ], + "type": "text", + "content": " from ground truth labels " + }, + { + "bbox": [ + 313, + 311, + 555, + 586 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 313, + 311, + 555, + 586 + ], + "type": "text", + "content": " in the context of classification, semantic segmentation, etc. However, this loss may not be ideal for applications where the targets " + }, + { + "bbox": [ + 313, + 311, + 555, + 586 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 313, + 311, + 555, + 586 + ], + "type": "text", + "content": " are soft categorical distributions representing various forms of class uncertainty. For example, this paper is focused on scribble-based segmentation where the ground truth is not known for most of the pixels, and the network training is done jointly with estimating pseudo-labels " + }, + { + "bbox": [ + 313, + 311, + 555, + 586 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 313, + 311, + 555, + 586 + ], + "type": "text", + "content": " for the unlabeled pixels. In this case, soft labels " + }, + { + "bbox": [ + 313, + 311, + 555, + 586 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 313, + 311, + 555, + 586 + ], + "type": "text", + "content": " are distributions representing class uncertainty. We observe that if such " + }, + { + "bbox": [ + 313, + 311, + 555, + 586 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 313, + 311, + 555, + 586 + ], + "type": "text", + "content": " is used as a target in " + }, + { + "bbox": [ + 313, + 311, + 555, + 586 + ], + "type": "inline_equation", + "content": "H(y, \\sigma)" + }, + { + "bbox": [ + 313, + 311, + 555, + 586 + ], + "type": "text", + "content": ", the network is trained to reproduce the uncertainty, see Figure 3(a). This motivates the discussion of alternative \"cross-entropy\" functions where the quotes indicate an informal interpretation of this information-theoretic concept. Intuitively, such functions should encourage decisiveness, as well as proximity between the predictions and pseudo-labels, but avoid mimicking the uncertainty in both directions: from soft pseudo-labels to predictions and vice-versa. We show that the last property can be achieved in a probabilistically principled manner. The following paragraphs discuss three cross-entropy functions " + }, + { + "bbox": [ + 313, + 311, + 555, + 586 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 313, + 311, + 555, + 586 + ], + "type": "text", + "content": " that we study in the context of our self-labeling loss (8)." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 586, + 555, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 586, + 555, + 622 + ], + "spans": [ + { + "bbox": [ + 313, + 586, + 555, + 622 + ], + "type": "text", + "content": "Standard cross-entropy provides the obvious baseline for evaluating two alternative versions that follow. For completeness, we include its mathematical definition" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 337, + 632, + 555, + 657 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 337, + 632, + 555, + 657 + ], + "spans": [ + { + "bbox": [ + 337, + 632, + 555, + 657 + ], + "type": "interline_equation", + "content": "H _ {\\mathrm {C E}} \\left(y _ {i}, \\sigma_ {i}\\right) = H \\left(y _ {i}, \\sigma_ {i}\\right) \\equiv - \\sum_ {k} y _ {i} ^ {k} \\ln \\sigma_ {i} ^ {k} \\tag {9}", + "image_path": "8d09e03d3d15365c7a0f2630d197e74f91272b36de13ab9fc20698108a28804a.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "type": "text", + "content": "and remind the reader that this loss is primarily used with hard or one-hot labels, in which case it is also equivalent to NLL loss - " + }, + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\ln \\sigma_i^{y_i}" + }, + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "type": "text", + "content": " previously discussed for ground truth labels (3). As mentioned earlier, Figure 3(a) shows that for" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20247" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 70, + 555, + 242 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 555, + 242 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 555, + 242 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 555, + 242 + ], + "type": "image", + "image_path": "46b0c1694453a6d3d12bc007b4636fb591334bcf3f67ea99405c82c8ed82358e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 250, + 555, + 298 + ], + "lines": [ + { + "bbox": [ + 55, + 250, + 555, + 298 + ], + "spans": [ + { + "bbox": [ + 55, + 250, + 555, + 298 + ], + "type": "text", + "content": "Figure 2. Second-order Potts relaxations in Tables 1 and 2: interaction potentials " + }, + { + "bbox": [ + 55, + 250, + 555, + 298 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 55, + 250, + 555, + 298 + ], + "type": "text", + "content": " for pairs of predictions " + }, + { + "bbox": [ + 55, + 250, + 555, + 298 + ], + "type": "inline_equation", + "content": "(\\sigma_i, \\sigma_j)" + }, + { + "bbox": [ + 55, + 250, + 555, + 298 + ], + "type": "text", + "content": " in (4) or pseudo-labels " + }, + { + "bbox": [ + 55, + 250, + 555, + 298 + ], + "type": "inline_equation", + "content": "(y_i, y_j)" + }, + { + "bbox": [ + 55, + 250, + 555, + 298 + ], + "type": "text", + "content": " in (8) are illustrated for " + }, + { + "bbox": [ + 55, + 250, + 555, + 298 + ], + "type": "inline_equation", + "content": "K = 2" + }, + { + "bbox": [ + 55, + 250, + 555, + 298 + ], + "type": "text", + "content": " when each " + }, + { + "bbox": [ + 55, + 250, + 555, + 298 + ], + "type": "inline_equation", + "content": "\\sigma_i" + }, + { + "bbox": [ + 55, + 250, + 555, + 298 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 55, + 250, + 555, + 298 + ], + "type": "inline_equation", + "content": "y_i" + }, + { + "bbox": [ + 55, + 250, + 555, + 298 + ], + "type": "text", + "content": " (binary distributions in " + }, + { + "bbox": [ + 55, + 250, + 555, + 298 + ], + "type": "inline_equation", + "content": "\\Delta^2" + }, + { + "bbox": [ + 55, + 250, + 555, + 298 + ], + "type": "text", + "content": ") can be represented by a single scalar as " + }, + { + "bbox": [ + 55, + 250, + 555, + 298 + ], + "type": "inline_equation", + "content": "(x, 1 - x)" + }, + { + "bbox": [ + 55, + 250, + 555, + 298 + ], + "type": "text", + "content": ". The contour maps are iso-levels of " + }, + { + "bbox": [ + 55, + 250, + 555, + 298 + ], + "type": "inline_equation", + "content": "P((x_i, 1 - x_i), (x_j, 1 - x_j))" + }, + { + "bbox": [ + 55, + 250, + 555, + 298 + ], + "type": "text", + "content": " over domain " + }, + { + "bbox": [ + 55, + 250, + 555, + 298 + ], + "type": "inline_equation", + "content": "(x_i, x_j) \\in [0, 1]^2" + }, + { + "bbox": [ + 55, + 250, + 555, + 298 + ], + "type": "text", + "content": ". The 3D plots above illustrate the potentials " + }, + { + "bbox": [ + 55, + 250, + 555, + 298 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 55, + 250, + 555, + 298 + ], + "type": "text", + "content": " as functions over pairs of \"logits\" " + }, + { + "bbox": [ + 55, + 250, + 555, + 298 + ], + "type": "inline_equation", + "content": "(l_i, l_j) \\in \\mathbb{R}^2" + }, + { + "bbox": [ + 55, + 250, + 555, + 298 + ], + "type": "text", + "content": " where each scalar " + }, + { + "bbox": [ + 55, + 250, + 555, + 298 + ], + "type": "inline_equation", + "content": "l_i" + }, + { + "bbox": [ + 55, + 250, + 555, + 298 + ], + "type": "text", + "content": " defines binary distribution " + }, + { + "bbox": [ + 55, + 250, + 555, + 298 + ], + "type": "inline_equation", + "content": "(x_i, 1 - x_i)" + }, + { + "bbox": [ + 55, + 250, + 555, + 298 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 55, + 250, + 555, + 298 + ], + "type": "inline_equation", + "content": "x_i = \\frac{1}{1 + e^{-2l_i}} \\in [0, 1]" + }, + { + "bbox": [ + 55, + 250, + 555, + 298 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 72, + 305, + 174, + 378 + ], + "blocks": [ + { + "bbox": [ + 72, + 305, + 174, + 378 + ], + "lines": [ + { + "bbox": [ + 72, + 305, + 174, + 378 + ], + "spans": [ + { + "bbox": [ + 72, + 305, + 174, + 378 + ], + "type": "image", + "image_path": "b2db9ae6cb80b535e88f7329a37e10372f6ab064c854737f6e5261af754da1cd.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 78, + 382, + 167, + 394 + ], + "lines": [ + { + "bbox": [ + 78, + 382, + 167, + 394 + ], + "spans": [ + { + "bbox": [ + 78, + 382, + 167, + 394 + ], + "type": "text", + "content": "(a) standard " + }, + { + "bbox": [ + 78, + 382, + 167, + 394 + ], + "type": "inline_equation", + "content": "H_{\\mathrm{CE}}(y,\\sigma)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 177, + 307, + 278, + 378 + ], + "blocks": [ + { + "bbox": [ + 177, + 307, + 278, + 378 + ], + "lines": [ + { + "bbox": [ + 177, + 307, + 278, + 378 + ], + "spans": [ + { + "bbox": [ + 177, + 307, + 278, + 378 + ], + "type": "image", + "image_path": "4d1f589b41658e5832c2fc49bf33b725a1bfef477a4d60b4376fcaa95ef1912e.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 403, + 555, + 470 + ], + "lines": [ + { + "bbox": [ + 55, + 403, + 555, + 470 + ], + "spans": [ + { + "bbox": [ + 55, + 403, + 555, + 470 + ], + "type": "text", + "content": "Figure 3. Illustration of cross-entropy functions: (a) standard (9), (b) reverse (10), and (c) collision (11). (d) shows the empirical comparison on the robustness to label uncertainty. The test uses ResNet-18 architecture on fully-supervised Natural Scene dataset [32] where we corrupted some labels. The horizontal axis shows the percentage " + }, + { + "bbox": [ + 55, + 403, + 555, + 470 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 55, + 403, + 555, + 470 + ], + "type": "text", + "content": " of training images where the correct ground truth labels were replaced by a random label. All losses trained the model using soft target distributions " + }, + { + "bbox": [ + 55, + 403, + 555, + 470 + ], + "type": "inline_equation", + "content": "\\hat{y} = \\eta * u + (1 - \\eta) * y" + }, + { + "bbox": [ + 55, + 403, + 555, + 470 + ], + "type": "text", + "content": " representing the mixture of one-hot distribution " + }, + { + "bbox": [ + 55, + 403, + 555, + 470 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 55, + 403, + 555, + 470 + ], + "type": "text", + "content": " for the observed corrupt label and the uniform distribution " + }, + { + "bbox": [ + 55, + 403, + 555, + 470 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 55, + 403, + 555, + 470 + ], + "type": "text", + "content": ", following [31]. The vertical axis shows the test accuracy. Training with the reverse and collision cross-entropy is robust to much higher levels of label uncertainty." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 282, + 307, + 383, + 378 + ], + "blocks": [ + { + "bbox": [ + 184, + 382, + 272, + 394 + ], + "lines": [ + { + "bbox": [ + 184, + 382, + 272, + 394 + ], + "spans": [ + { + "bbox": [ + 184, + 382, + 272, + 394 + ], + "type": "text", + "content": "(b) reverse " + }, + { + "bbox": [ + 184, + 382, + 272, + 394 + ], + "type": "inline_equation", + "content": "H_{\\mathrm{RCE}}(y,\\sigma)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 282, + 307, + 383, + 378 + ], + "lines": [ + { + "bbox": [ + 282, + 307, + 383, + 378 + ], + "spans": [ + { + "bbox": [ + 282, + 307, + 383, + 378 + ], + "type": "image", + "image_path": "d95d627af0eaa8670a8b91a7672463a1f1a1f281a7923bbff07b04d8e010fe64.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 391, + 307, + 536, + 378 + ], + "blocks": [ + { + "bbox": [ + 286, + 382, + 380, + 394 + ], + "lines": [ + { + "bbox": [ + 286, + 382, + 380, + 394 + ], + "spans": [ + { + "bbox": [ + 286, + 382, + 380, + 394 + ], + "type": "text", + "content": "(c) collision " + }, + { + "bbox": [ + 286, + 382, + 380, + 394 + ], + "type": "inline_equation", + "content": "H_{\\mathrm{CCE}}(y,\\sigma)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 391, + 307, + 536, + 378 + ], + "lines": [ + { + "bbox": [ + 391, + 307, + 536, + 378 + ], + "spans": [ + { + "bbox": [ + 391, + 307, + 536, + 378 + ], + "type": "image", + "image_path": "5d6c5709ee51b1acfe9156ba39ee372ee65e69df03cfdfe2b587762885382f71.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 416, + 382, + 509, + 394 + ], + "lines": [ + { + "bbox": [ + 416, + 382, + 509, + 394 + ], + "spans": [ + { + "bbox": [ + 416, + 382, + 509, + 394 + ], + "type": "text", + "content": "(d) empirical comparison" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 491, + 296, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 491, + 296, + 563 + ], + "spans": [ + { + "bbox": [ + 55, + 491, + 296, + 563 + ], + "type": "text", + "content": "soft pseudo-labels like " + }, + { + "bbox": [ + 55, + 491, + 296, + 563 + ], + "type": "inline_equation", + "content": "y = (0.5, 0.5)" + }, + { + "bbox": [ + 55, + 491, + 296, + 563 + ], + "type": "text", + "content": ", it forces predictions to mimic or replicate the uncertainty " + }, + { + "bbox": [ + 55, + 491, + 296, + 563 + ], + "type": "inline_equation", + "content": "\\sigma \\approx y" + }, + { + "bbox": [ + 55, + 491, + 296, + 563 + ], + "type": "text", + "content": ". In fact, label " + }, + { + "bbox": [ + 55, + 491, + 296, + 563 + ], + "type": "inline_equation", + "content": "y = (0.5, 0.5)" + }, + { + "bbox": [ + 55, + 491, + 296, + 563 + ], + "type": "text", + "content": " just tells that the class is unknown and the network should not be supervised by this point. This problem manifests itself in the poor performance of the standard cross-entropy (9) in our experiment discussed in Figure 3 (d) (red)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 67, + 563, + 268, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 563, + 268, + 575 + ], + "spans": [ + { + "bbox": [ + 67, + 563, + 268, + 575 + ], + "type": "text", + "content": "Reverse cross-entropy is defined in this paper as" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 69, + 584, + 295, + 608 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 584, + 295, + 608 + ], + "spans": [ + { + "bbox": [ + 69, + 584, + 295, + 608 + ], + "type": "interline_equation", + "content": "H _ {\\mathrm {R C E}} \\left(y _ {i}, \\sigma_ {i}\\right) = H \\left(\\sigma_ {i}, y _ {i}\\right) \\equiv - \\sum_ {k} \\sigma_ {i} ^ {k} \\ln y _ {i} ^ {k} \\tag {10}", + "image_path": "1e4f16aebc5c6f47e007f15ee203322ee304582ccc7271b8b83cde9be2c405bd.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 618, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 296, + 714 + ], + "type": "text", + "content": "switching the order of labels and predictions in (9), which is not common. Indeed, Shannon's cross-entropy is not symmetric and the first argument is normally the target distribution and the second is the estimated distribution. However, in our case, both distributions are estimated and there is no reason not to try the reverse order. It is worth noting that our self-labeling formulation (8) suggests that reverse cross-entropy naturally appears when the ADM approach splits the" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 491, + 555, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 491, + 555, + 574 + ], + "spans": [ + { + "bbox": [ + 313, + 491, + 555, + 574 + ], + "type": "text", + "content": "decisiveness and fairness into separate sub-problems. Moreover, as Figure 3(b) shows, in this case, the network does not mimic uncertain pseudo-labels, e.g. the gradient of the blue line is zero. The results for the reverse cross-entropy in Figure 3 (d) (green) are significantly better than for the standard (red). Unfortunately, now pseudo-labels " + }, + { + "bbox": [ + 313, + 491, + 555, + 574 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 313, + 491, + 555, + 574 + ], + "type": "text", + "content": " mimic the uncertainty in predictions " + }, + { + "bbox": [ + 313, + 491, + 555, + 574 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 313, + 491, + 555, + 574 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 326, + 574, + 522, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 574, + 522, + 586 + ], + "spans": [ + { + "bbox": [ + 326, + 574, + 522, + 586 + ], + "type": "text", + "content": "Collision cross-entropy defined in this paper as" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 329, + 593, + 555, + 617 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 329, + 593, + 555, + 617 + ], + "spans": [ + { + "bbox": [ + 329, + 593, + 555, + 617 + ], + "type": "interline_equation", + "content": "H _ {\\mathrm {C C E}} \\left(y _ {i}, \\sigma_ {i}\\right) \\equiv - \\ln \\sum_ {k} \\sigma_ {i} ^ {k} y _ {i} ^ {k} \\equiv - \\ln \\sigma^ {\\top} y \\tag {11}", + "image_path": "7c871e586bfe8b9bc6f9848f72bd4888375a671f2c845726e16a6cc39a229b8e.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 624, + 555, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 624, + 555, + 685 + ], + "spans": [ + { + "bbox": [ + 313, + 624, + 555, + 685 + ], + "type": "text", + "content": "resolves the problem in a principled way. It is symmetric w.r.t. pseudo-labels and predictions. The dot product " + }, + { + "bbox": [ + 313, + 624, + 555, + 685 + ], + "type": "inline_equation", + "content": "\\sigma^{\\top}y" + }, + { + "bbox": [ + 313, + 624, + 555, + 685 + ], + "type": "text", + "content": " can be seen as a probability of equality of two random variables, the predicted class " + }, + { + "bbox": [ + 313, + 624, + 555, + 685 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 313, + 624, + 555, + 685 + ], + "type": "text", + "content": " and unknown true class " + }, + { + "bbox": [ + 313, + 624, + 555, + 685 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 624, + 555, + 685 + ], + "type": "text", + "content": ", which are represented by the distributions " + }, + { + "bbox": [ + 313, + 624, + 555, + 685 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 313, + 624, + 555, + 685 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 624, + 555, + 685 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 313, + 624, + 555, + 685 + ], + "type": "text", + "content": ". Indeed," + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 330, + 690, + 537, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 330, + 690, + 537, + 715 + ], + "spans": [ + { + "bbox": [ + 330, + 690, + 537, + 715 + ], + "type": "interline_equation", + "content": "\\Pr (C = T) = \\sum_ {k} P r (C = k) \\Pr (T = k) = \\sigma^ {\\top} y.", + "image_path": "529b6ce38b8aecc6d006edad6f7a976a74aa6efb90bde7147ae5e547346a1f2f.jpg" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20248" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "text", + "content": "Loss (11) maximizes this \"collision\" probability instead of enforcing the equality of distributions " + }, + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "inline_equation", + "content": "\\sigma = y" + }, + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "text", + "content": ". Figure 3(c) shows no mimicking of uncertainty (blue line). Unlike the reverse cross-entropy, this is also valid when " + }, + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "text", + "content": " is estimated from uncertain predictions " + }, + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "text", + "content": " since (11) is symmetric. This leads to the best performance in Figure 3 (d) (blue). Our extensive experiments are conclusive that collision cross-entropy is the best option for " + }, + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "text", + "content": " in self-labeling loss (8)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 180, + 136, + 193 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 180, + 136, + 193 + ], + "spans": [ + { + "bbox": [ + 55, + 180, + 136, + 193 + ], + "type": "text", + "content": "3. Experiments" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 201, + 297, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 201, + 297, + 357 + ], + "spans": [ + { + "bbox": [ + 54, + 201, + 297, + 357 + ], + "type": "text", + "content": "Our experiments evaluate the components of our loss (8) (cross-entropy, pairwise term, and neighborhood). We compare standard segmentation architectures trained using the \"textbook\" approach, i.e. loss minimization, to the state-of-the-art methods, which often require network modifications and ad-hoc training procedures. Our quantitative tests in Section 3.1 evaluate different Potts relaxations. Some qualitative examples are presented in Figure 4. Section 3.2 compares three cross-entropy terms. Section 3.3 evaluates our soft self-labeling loss on the nearest-neighbor and dense neighborhood systems. We summarized the results in Section 3.4. Section 3.5 shows that our method achieves the SOTA and, in some cases, may outperform the full supervision." + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 80, + 428, + 141, + 487 + ], + "blocks": [ + { + "bbox": [ + 80, + 428, + 141, + 487 + ], + "lines": [ + { + "bbox": [ + 80, + 428, + 141, + 487 + ], + "spans": [ + { + "bbox": [ + 80, + 428, + 141, + 487 + ], + "type": "image", + "image_path": "320ea53741265cd39b5c7e57e714b7d10b28a7711b2a64b6acabd89d1b33d04c.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 81, + 488, + 140, + 496 + ], + "lines": [ + { + "bbox": [ + 81, + 488, + 140, + 496 + ], + "spans": [ + { + "bbox": [ + 81, + 488, + 140, + 496 + ], + "type": "text", + "content": "(a) Image, GT & input" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 506, + 295, + 539 + ], + "lines": [ + { + "bbox": [ + 55, + 506, + 295, + 539 + ], + "spans": [ + { + "bbox": [ + 55, + 506, + 295, + 539 + ], + "type": "text", + "content": "Figure 4. Illustration of the difference among Potts relaxations. The visualization of soft pseudo-labels uses the convex combination of RGB colors for each class weighted by pseudo-label itself." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 164, + 369, + 253, + 486 + ], + "blocks": [ + { + "bbox": [ + 164, + 369, + 253, + 486 + ], + "lines": [ + { + "bbox": [ + 164, + 369, + 253, + 486 + ], + "spans": [ + { + "bbox": [ + 164, + 369, + 253, + 486 + ], + "type": "image", + "image_path": "3fd7e43195ea2d83908a43bfe1cbf33c198ad222dc2286e4f11694458be0b892.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 148, + 488, + 272, + 496 + ], + "lines": [ + { + "bbox": [ + 148, + 488, + 272, + 496 + ], + "spans": [ + { + "bbox": [ + 148, + 488, + 272, + 496 + ], + "type": "text", + "content": "(b) Pseudo-labels using different Potts relaxation" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "content": "Dataset and evaluation We mainly use the standard PASCAL VOC 2012 dataset [18] and scribble annotations for supervision [28]. The dataset contains 21 classes including background. Following the common practice [11, 37, 38], we use the augmented version with 10,582 training images and 1449 images for validation. Our evaluation metric is the standard mean Intersection-over-Union (mIoU) on the validation set. We also test our method on two additional datasets in Section 3.5. One is Cityscapes [14] which is built for urban scenes and consists of 2975 and 500 fine-labeled images for training and validation. There are 19 out of 30 annotated classes for semantic segmentation. The other one" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 334, + 70, + 535, + 167 + ], + "blocks": [ + { + "bbox": [ + 334, + 70, + 535, + 167 + ], + "lines": [ + { + "bbox": [ + 334, + 70, + 535, + 167 + ], + "spans": [ + { + "bbox": [ + 334, + 70, + 535, + 167 + ], + "type": "table", + "html": "
scribble length ratio
00.30.50.81.0
PBL56.4261.7463.8165.7367.24
PNQ59.0165.5367.8070.6371.12
PQ58.9265.3467.8170.4371.05
PCCE56.4061.8263.8165.8167.41
PCD59.0465.5267.8470.9371.22
PLQ59.0365.4467.8170.8071.21
", + "image_path": "15211efd019070799ea0e5a19dd39da347feb5c589fa0d89bac72ace454ff750.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 175, + 554, + 198 + ], + "lines": [ + { + "bbox": [ + 313, + 175, + 554, + 198 + ], + "spans": [ + { + "bbox": [ + 313, + 175, + 554, + 198 + ], + "type": "text", + "content": "Table 3. Comparison of Potts relaxations with self-labeling. mIoUs on the validation set are shown here." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 219, + 555, + 268 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 219, + 555, + 268 + ], + "spans": [ + { + "bbox": [ + 313, + 219, + 555, + 268 + ], + "type": "text", + "content": "is ADE20k [46] which has 150 fine-grained classes. There are 20210 and 2000, images for training and validation. Instead of scribble-based supervision, we followed [27] to use the block-wise annotation as a form of weak supervision." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 281, + 556, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 281, + 556, + 521 + ], + "spans": [ + { + "bbox": [ + 313, + 281, + 556, + 521 + ], + "type": "text", + "content": "Implementation details We use DeepLabv3+ [13] with two backbones: ResNet101 [21] and MobileNetV2 [36]. We use ResNet101 in Section 3.5, and MobileNetV2 in other sections, for efficiency. We also use Vision Transformer backbone [17], specifically vit-base-patch16-224, and linear decoder in Section 3.5. All backbone networks (ResNet101 and MobileNetV2, ViT) are pre-trained onImagenet [16]. Unless stated explicitly, we use batch 12 as the default across all the experiments. To avoid feeding random predictions from the uninitialized classification head into pseudo-labeling sub-problem, we initialize the network using the cross-entropy on scribbles, i.e. the first term in loss (8), similarly to pre-training in [37]. The network parameters are optimized by SGD. The initial learning rate 0.0007 is scheduled by a polynomial decay of power 0.9. All loss variants are optimized for 60 epochs, but all hyperparameters are tuned separately for each variant. For our best result, we use " + }, + { + "bbox": [ + 313, + 281, + 556, + 521 + ], + "type": "inline_equation", + "content": "\\eta = 0.3" + }, + { + "bbox": [ + 313, + 281, + 556, + 521 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 281, + 556, + 521 + ], + "type": "inline_equation", + "content": "\\lambda = 6" + }, + { + "bbox": [ + 313, + 281, + 556, + 521 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 281, + 556, + 521 + ], + "type": "inline_equation", + "content": "H_{\\mathrm{CCE}}" + }, + { + "bbox": [ + 313, + 281, + 556, + 521 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 281, + 556, + 521 + ], + "type": "inline_equation", + "content": "P_{\\mathrm{CD}}" + }, + { + "bbox": [ + 313, + 281, + 556, + 521 + ], + "type": "text", + "content": ". The color-bandwidth in the Potts model is set to 9 across all the experiments on Pascal VOC 2012 and 3 for Cityscapes and ADE20k datasets." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 529, + 487, + 541 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 529, + 487, + 541 + ], + "spans": [ + { + "bbox": [ + 313, + 529, + 487, + 541 + ], + "type": "text", + "content": "3.1. Comparison of Potts relaxations" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 545, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 556, + 713 + ], + "type": "text", + "content": "To compare different Potts relaxations within our self-labeling framework, we select " + }, + { + "bbox": [ + 313, + 545, + 556, + 713 + ], + "type": "inline_equation", + "content": "H_{\\mathrm{CCE}}" + }, + { + "bbox": [ + 313, + 545, + 556, + 713 + ], + "type": "text", + "content": " as the cross-entropy term in (8), as motivated by the results in Section 3.2. The neighborhood system is the nearest neighbors. The quantitative results are in Table 3. First, one can see that the log-based pairwise terms work consistently better. The likely explanation is that the logarithm addresses the gradient vanishing problem that is obvious in the top row of Figure 2. Moreover, the logarithm encourages smoother transitions across the boundaries (see Figure 4). This may benefit network training when the exact ground truth is unknown. Intuitively, higher uncertainty around object boundaries is natural. Second, the normalized relaxations " + }, + { + "bbox": [ + 313, + 545, + 556, + 713 + ], + "type": "inline_equation", + "content": "P_{\\mathrm{NQ}}" + }, + { + "bbox": [ + 313, + 545, + 556, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 545, + 556, + 713 + ], + "type": "inline_equation", + "content": "P_{\\mathrm{CD}}" + }, + { + "bbox": [ + 313, + 545, + 556, + 713 + ], + "type": "text", + "content": " work better, confirming the limitations of " + }, + { + "bbox": [ + 313, + 545, + 556, + 713 + ], + "type": "inline_equation", + "content": "P_{\\mathrm{BL}}" + }, + { + "bbox": [ + 313, + 545, + 556, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 545, + 556, + 713 + ], + "type": "inline_equation", + "content": "P_{\\mathrm{Q}}" + }, + { + "bbox": [ + 313, + 545, + 556, + 713 + ], + "type": "text", + "content": " discussed in Fig. 1." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20249" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 76, + 72, + 273, + 205 + ], + "blocks": [ + { + "bbox": [ + 76, + 72, + 273, + 205 + ], + "lines": [ + { + "bbox": [ + 76, + 72, + 273, + 205 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 273, + 205 + ], + "type": "image", + "image_path": "3122dae951addc3a14b4403a881c4c59407c40048c29dccd68e9857585801043.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 91, + 209, + 259, + 220 + ], + "lines": [ + { + "bbox": [ + 91, + 209, + 259, + 220 + ], + "spans": [ + { + "bbox": [ + 91, + 209, + 259, + 220 + ], + "type": "text", + "content": "Figure 5. Comparison of cross-entropy terms." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 247, + 296, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 247, + 296, + 342 + ], + "spans": [ + { + "bbox": [ + 55, + 247, + 296, + 342 + ], + "type": "text", + "content": "While we focus on a general soft pseudo-label solver (see Suppl.Mat), our results can motivate specialized solvers. For example, quadratic relaxation " + }, + { + "bbox": [ + 55, + 247, + 296, + 342 + ], + "type": "inline_equation", + "content": "P_{\\mathrm{Q}}" + }, + { + "bbox": [ + 55, + 247, + 296, + 342 + ], + "type": "text", + "content": " seems attractive as it defines a convex loss (8). If using \"quadratic cross-entropy\" " + }, + { + "bbox": [ + 55, + 247, + 296, + 342 + ], + "type": "inline_equation", + "content": "H_{\\mathrm{Q}} := \\| y - \\sigma \\|^2" + }, + { + "bbox": [ + 55, + 247, + 296, + 342 + ], + "type": "text", + "content": ", then our loss even has a closed-form solution for " + }, + { + "bbox": [ + 55, + 247, + 296, + 342 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 55, + 247, + 296, + 342 + ], + "type": "text", + "content": " [1, 19]. However, the corresponding pseudolabels tend to be overly soft and additional decidiveness (non-convex entropy) term improves mIoU." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 358, + 244, + 371 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 358, + 244, + 371 + ], + "spans": [ + { + "bbox": [ + 55, + 358, + 244, + 371 + ], + "type": "text", + "content": "3.2. Comparison of cross-entropy terms" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 378, + 296, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 378, + 296, + 474 + ], + "spans": [ + { + "bbox": [ + 55, + 378, + 296, + 474 + ], + "type": "text", + "content": "We compare different cross-entropy terms while fixing the pairwise term to " + }, + { + "bbox": [ + 55, + 378, + 296, + 474 + ], + "type": "inline_equation", + "content": "P_{\\mathrm{Q}}" + }, + { + "bbox": [ + 55, + 378, + 296, + 474 + ], + "type": "text", + "content": " due to its simplicity. We also use the nearest neighbors. The results are in Figure 5. In agreement with Fig.3(d), " + }, + { + "bbox": [ + 55, + 378, + 296, + 474 + ], + "type": "inline_equation", + "content": "H_{\\mathrm{CE}}" + }, + { + "bbox": [ + 55, + 378, + 296, + 474 + ], + "type": "text", + "content": " performs consistently better across different supervision levels, i.e. scribble lengths. Both " + }, + { + "bbox": [ + 55, + 378, + 296, + 474 + ], + "type": "inline_equation", + "content": "H_{\\mathrm{CE}}" + }, + { + "bbox": [ + 55, + 378, + 296, + 474 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 378, + 296, + 474 + ], + "type": "inline_equation", + "content": "H_{\\mathrm{RCE}}" + }, + { + "bbox": [ + 55, + 378, + 296, + 474 + ], + "type": "text", + "content": " have a strong margin over the standard entropy " + }, + { + "bbox": [ + 55, + 378, + 296, + 474 + ], + "type": "inline_equation", + "content": "H_{\\mathrm{CE}}" + }, + { + "bbox": [ + 55, + 378, + 296, + 474 + ], + "type": "text", + "content": " as network training becomes robust to uncertainty and errors in pseudo labels, see Sec.2.2." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 490, + 255, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 490, + 255, + 502 + ], + "spans": [ + { + "bbox": [ + 55, + 490, + 255, + 502 + ], + "type": "text", + "content": "3.3. Comparison of neighborhood systems" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 510, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 296, + 713 + ], + "type": "text", + "content": "So far we used the four nearest neighbors (NN) neighborhood for the pairwise Potts term in (8). Here we compare NN with the dense neighborhoods (DN). We use " + }, + { + "bbox": [ + 55, + 510, + 296, + 713 + ], + "type": "inline_equation", + "content": "H_{\\mathrm{CCE}}" + }, + { + "bbox": [ + 55, + 510, + 296, + 713 + ], + "type": "text", + "content": " as the cross-entropy term. To optimize pseudo-labels over DN, we still use the general gradient descent technique detailed in the Supplementary Material. The gradient computation for DN Potts can employ the efficient bilateral filtering [37], but it applies only to (proper) second-order variants, i.e. " + }, + { + "bbox": [ + 55, + 510, + 296, + 713 + ], + "type": "inline_equation", + "content": "P_{\\mathrm{BL}}" + }, + { + "bbox": [ + 55, + 510, + 296, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 510, + 296, + 713 + ], + "type": "inline_equation", + "content": "P_{\\mathrm{Q}}" + }, + { + "bbox": [ + 55, + 510, + 296, + 713 + ], + "type": "text", + "content": ". We focus on " + }, + { + "bbox": [ + 55, + 510, + 296, + 713 + ], + "type": "inline_equation", + "content": "P_{\\mathrm{Q}}" + }, + { + "bbox": [ + 55, + 510, + 296, + 713 + ], + "type": "text", + "content": " as a preferred option for soft self-labeling. We obtained 71.1% mIoU on nearest neighbors while only getting 67.9% on dense neighborhoods (bandwidth 100). Some qualitative results are shown in Figure 6 indicating that larger neighborhoods induce lower-quality pseudo-labels. A possible explanation is that the Potts model reduces to the cardinality/volume potentials for larger neighborhoods [39]. The nearest neighborhood is better for edge alignment producing better weakly-supervised results." + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 317, + 72, + 359, + 112 + ], + "blocks": [ + { + "bbox": [ + 317, + 72, + 359, + 112 + ], + "lines": [ + { + "bbox": [ + 317, + 72, + 359, + 112 + ], + "spans": [ + { + "bbox": [ + 317, + 72, + 359, + 112 + ], + "type": "image", + "image_path": "c2fc96c343432f316cae67acd322d240a0eaf8957e782a7ae5608eca6dadce14.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 317, + 114, + 358, + 152 + ], + "blocks": [ + { + "bbox": [ + 317, + 114, + 358, + 152 + ], + "lines": [ + { + "bbox": [ + 317, + 114, + 358, + 152 + ], + "spans": [ + { + "bbox": [ + 317, + 114, + 358, + 152 + ], + "type": "image", + "image_path": "6c12f2614429f36f8305135c3d5b0a4640676ce2a9825d65cfdea7c56e107311.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 317, + 153, + 358, + 192 + ], + "blocks": [ + { + "bbox": [ + 317, + 153, + 358, + 192 + ], + "lines": [ + { + "bbox": [ + 317, + 153, + 358, + 192 + ], + "spans": [ + { + "bbox": [ + 317, + 153, + 358, + 192 + ], + "type": "image", + "image_path": "c3aa74185d2a26cb9d267b6338643b0b168ed92dffb2105aa6ec35ba70394505.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 317, + 192, + 357, + 232 + ], + "blocks": [ + { + "bbox": [ + 317, + 192, + 357, + 232 + ], + "lines": [ + { + "bbox": [ + 317, + 192, + 357, + 232 + ], + "spans": [ + { + "bbox": [ + 317, + 192, + 357, + 232 + ], + "type": "image", + "image_path": "de4e540c40c1915846578c05e73d7a31b9034e2d741e940643660f6da029ff74.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 317, + 234, + 360, + 242 + ], + "lines": [ + { + "bbox": [ + 317, + 234, + 360, + 242 + ], + "spans": [ + { + "bbox": [ + 317, + 234, + 360, + 242 + ], + "type": "text", + "content": "(a) Image & scribble" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 365, + 72, + 406, + 111 + ], + "blocks": [ + { + "bbox": [ + 365, + 72, + 406, + 111 + ], + "lines": [ + { + "bbox": [ + 365, + 72, + 406, + 111 + ], + "spans": [ + { + "bbox": [ + 365, + 72, + 406, + 111 + ], + "type": "image", + "image_path": "783c1d36fb18304b47965edbe339884ab6af4f3146fa035f1c6d44d47f665884.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 366, + 113, + 406, + 152 + ], + "blocks": [ + { + "bbox": [ + 366, + 113, + 406, + 152 + ], + "lines": [ + { + "bbox": [ + 366, + 113, + 406, + 152 + ], + "spans": [ + { + "bbox": [ + 366, + 113, + 406, + 152 + ], + "type": "image", + "image_path": "52dca61eca1fb424f26824ea6883747c61bffe9a0c53a2909c060acacf91789a.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 366, + 153, + 405, + 192 + ], + "blocks": [ + { + "bbox": [ + 366, + 153, + 405, + 192 + ], + "lines": [ + { + "bbox": [ + 366, + 153, + 405, + 192 + ], + "spans": [ + { + "bbox": [ + 366, + 153, + 405, + 192 + ], + "type": "image", + "image_path": "ec7576f8952af1dde748ada0014e74f19bc8c9fef8bc6ed9068f9a3fbc54b5dc.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 366, + 193, + 405, + 232 + ], + "blocks": [ + { + "bbox": [ + 366, + 193, + 405, + 232 + ], + "lines": [ + { + "bbox": [ + 366, + 193, + 405, + 232 + ], + "spans": [ + { + "bbox": [ + 366, + 193, + 405, + 232 + ], + "type": "image", + "image_path": "1344c98ce43af0f8d631c5fa7e20b90492e7219f1be7b9ef243b779b6d0a1ad9.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 371, + 234, + 402, + 243 + ], + "lines": [ + { + "bbox": [ + 371, + 234, + 402, + 243 + ], + "spans": [ + { + "bbox": [ + 371, + 234, + 402, + 243 + ], + "type": "text", + "content": "(b) Predictions" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 414, + 72, + 455, + 111 + ], + "blocks": [ + { + "bbox": [ + 414, + 72, + 455, + 111 + ], + "lines": [ + { + "bbox": [ + 414, + 72, + 455, + 111 + ], + "spans": [ + { + "bbox": [ + 414, + 72, + 455, + 111 + ], + "type": "image", + "image_path": "6c2fd150f8362cd43c01e74163f03d5beecae15b6411dc217162fae0a70c3e5c.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 415, + 113, + 455, + 152 + ], + "blocks": [ + { + "bbox": [ + 415, + 113, + 455, + 152 + ], + "lines": [ + { + "bbox": [ + 415, + 113, + 455, + 152 + ], + "spans": [ + { + "bbox": [ + 415, + 113, + 455, + 152 + ], + "type": "image", + "image_path": "78b460e16b3841a2e5edce9ec7d7042c8cae0f926b76845d5388a631d4a9a32b.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 415, + 152, + 455, + 192 + ], + "blocks": [ + { + "bbox": [ + 415, + 152, + 455, + 192 + ], + "lines": [ + { + "bbox": [ + 415, + 152, + 455, + 192 + ], + "spans": [ + { + "bbox": [ + 415, + 152, + 455, + 192 + ], + "type": "image", + "image_path": "978c0339fde0f1cd69429dd2d6b280fb3f870fdd28612c522988713e7b064720.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 415, + 193, + 455, + 232 + ], + "blocks": [ + { + "bbox": [ + 415, + 193, + 455, + 232 + ], + "lines": [ + { + "bbox": [ + 415, + 193, + 455, + 232 + ], + "spans": [ + { + "bbox": [ + 415, + 193, + 455, + 232 + ], + "type": "image", + "image_path": "9bb6d1ace973dbc963681178c47b3eac5869f3d017614fcb4afb4112e8ca4342.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 416, + 233, + 453, + 245 + ], + "lines": [ + { + "bbox": [ + 416, + 233, + 453, + 245 + ], + "spans": [ + { + "bbox": [ + 416, + 233, + 453, + 245 + ], + "type": "text", + "content": "(c) Pseudo-labels for NN" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 464, + 72, + 504, + 111 + ], + "blocks": [ + { + "bbox": [ + 464, + 72, + 504, + 111 + ], + "lines": [ + { + "bbox": [ + 464, + 72, + 504, + 111 + ], + "spans": [ + { + "bbox": [ + 464, + 72, + 504, + 111 + ], + "type": "image", + "image_path": "75b68ae4089eaef4fbb7ffc0b693a307cf1605b1310277fc7e35904208d33cee.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 464, + 113, + 504, + 152 + ], + "blocks": [ + { + "bbox": [ + 464, + 113, + 504, + 152 + ], + "lines": [ + { + "bbox": [ + 464, + 113, + 504, + 152 + ], + "spans": [ + { + "bbox": [ + 464, + 113, + 504, + 152 + ], + "type": "image", + "image_path": "34471a5c14220d5b5dc38f6c58c39ea282751ee6f35897fe7473020a1325c0ce.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 464, + 152, + 504, + 192 + ], + "blocks": [ + { + "bbox": [ + 464, + 152, + 504, + 192 + ], + "lines": [ + { + "bbox": [ + 464, + 152, + 504, + 192 + ], + "spans": [ + { + "bbox": [ + 464, + 152, + 504, + 192 + ], + "type": "image", + "image_path": "f6886a149273b14426a08315df22705302beb983fdff2bfbc531d65588610f27.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 464, + 193, + 504, + 232 + ], + "blocks": [ + { + "bbox": [ + 464, + 193, + 504, + 232 + ], + "lines": [ + { + "bbox": [ + 464, + 193, + 504, + 232 + ], + "spans": [ + { + "bbox": [ + 464, + 193, + 504, + 232 + ], + "type": "image", + "image_path": "608fd117ab593e33fedd0203d633fc1c654639058b3aa991ef6d702219dc9414.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 465, + 233, + 503, + 245 + ], + "lines": [ + { + "bbox": [ + 465, + 233, + 503, + 245 + ], + "spans": [ + { + "bbox": [ + 465, + 233, + 503, + 245 + ], + "type": "text", + "content": "(d) Pseudo-labels for DN (25)" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 255, + 553, + 277 + ], + "lines": [ + { + "bbox": [ + 313, + 255, + 553, + 277 + ], + "spans": [ + { + "bbox": [ + 313, + 255, + 553, + 277 + ], + "type": "text", + "content": "Figure 6. Pseudo-labels generated from given network predictions using different neighborhoods: nearest (NN) and dense (DN)." + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_caption" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 512, + 72, + 552, + 111 + ], + "blocks": [ + { + "bbox": [ + 512, + 72, + 552, + 111 + ], + "lines": [ + { + "bbox": [ + 512, + 72, + 552, + 111 + ], + "spans": [ + { + "bbox": [ + 512, + 72, + 552, + 111 + ], + "type": "image", + "image_path": "76ed42884b0ee07f8d0b0ecb4f44ca3d1cb6f9bd880fe2c0997203184eb68fbd.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 512, + 113, + 552, + 152 + ], + "blocks": [ + { + "bbox": [ + 512, + 113, + 552, + 152 + ], + "lines": [ + { + "bbox": [ + 512, + 113, + 552, + 152 + ], + "spans": [ + { + "bbox": [ + 512, + 113, + 552, + 152 + ], + "type": "image", + "image_path": "d86a57e2d0a898c322e078ca2d59bb4d264831350558c4b62edae94879dc04b2.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 512, + 152, + 552, + 192 + ], + "blocks": [ + { + "bbox": [ + 512, + 152, + 552, + 192 + ], + "lines": [ + { + "bbox": [ + 512, + 152, + 552, + 192 + ], + "spans": [ + { + "bbox": [ + 512, + 152, + 552, + 192 + ], + "type": "image", + "image_path": "8300a86876e651a6d487845539426df4956c283074bdb6c3d3beaa58c18c68c3.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 512, + 193, + 552, + 232 + ], + "blocks": [ + { + "bbox": [ + 512, + 193, + 552, + 232 + ], + "lines": [ + { + "bbox": [ + 512, + 193, + 552, + 232 + ], + "spans": [ + { + "bbox": [ + 512, + 193, + 552, + 232 + ], + "type": "image", + "image_path": "b7f6bd2a92296f335136138e8df2c47706e4626eafd3dcaac07873e39f92b14e.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 513, + 233, + 551, + 245 + ], + "lines": [ + { + "bbox": [ + 513, + 233, + 551, + 245 + ], + "spans": [ + { + "bbox": [ + 513, + 233, + 551, + 245 + ], + "type": "text", + "content": "(e) Pseudo-labels for DN (100)" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + } + ], + "index": 30 + }, + { + "type": "table", + "bbox": [ + 334, + 288, + 535, + 340 + ], + "blocks": [ + { + "bbox": [ + 334, + 288, + 535, + 340 + ], + "lines": [ + { + "bbox": [ + 334, + 288, + 535, + 340 + ], + "spans": [ + { + "bbox": [ + 334, + 288, + 535, + 340 + ], + "type": "table", + "html": "
N
NNDN
GD67.069.5* [38]
SLhard69.6* [29]63.1 [28]
soft71.167.9
", + "image_path": "65de2bad099a03d2e488968bd391a853034cfabaece7a91111b6026639e46ff2.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "table_body" + } + ], + "index": 33 + }, + { + "bbox": [ + 313, + 347, + 555, + 370 + ], + "lines": [ + { + "bbox": [ + 313, + 347, + 555, + 370 + ], + "spans": [ + { + "bbox": [ + 313, + 347, + 555, + 370 + ], + "type": "text", + "content": "Table 4. Summary of Potts-based WSSS (full-scribbles, DeepLab V3+). Reproduced results using public code are marked by \\*." + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 346, + 382, + 524, + 488 + ], + "blocks": [ + { + "bbox": [ + 346, + 382, + 524, + 488 + ], + "lines": [ + { + "bbox": [ + 346, + 382, + 524, + 488 + ], + "spans": [ + { + "bbox": [ + 346, + 382, + 524, + 488 + ], + "type": "image", + "image_path": "ba96d6de4fc604fb80670bc7b8c02a3948d3c7a63ae9d7f9d358a4ca53acf655.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 497, + 555, + 519 + ], + "lines": [ + { + "bbox": [ + 313, + 497, + 555, + 519 + ], + "spans": [ + { + "bbox": [ + 313, + 497, + 555, + 519 + ], + "type": "text", + "content": "Figure 7. Comparison of different methods using Potts relaxations. The architecture is DeeplabV3+ with the backbone MobileNetV2." + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_caption" + } + ], + "index": 35 + }, + { + "bbox": [ + 313, + 540, + 553, + 553 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 540, + 553, + 553 + ], + "spans": [ + { + "bbox": [ + 313, + 540, + 553, + 553 + ], + "type": "text", + "content": "3.4. Soft vs. hard self-labeling vs. gradient descent" + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 313, + 558, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 556, + 713 + ], + "type": "text", + "content": "Here we review our general observations about the Potts-based WSSS using the scribble-based results summarized in Table 4. First, in the simplest approach using direct optimization of the network parameters w.r.t. Potts-regularized losses like (4) and stochastic gradient descent, one benefits from a larger neighborhood size [38]. They use " + }, + { + "bbox": [ + 313, + 558, + 556, + 713 + ], + "type": "inline_equation", + "content": "P_{\\mathrm{BL}}" + }, + { + "bbox": [ + 313, + 558, + 556, + 713 + ], + "type": "text", + "content": " where DN smoothens the Potts model [30] helping SGD to avoid local minima. Higher-order optimization based on auxiliary self-labeling loss (8) significantly benefits from soft pseudo-labels since they can represent uncertainty. In this case (normalized) quadratic Potts relaxations offer two advantages over the tight bilinear relaxation that leads to hard solutions. On the one hand, they do not suppress label soft" + } + ] + } + ], + "index": 38 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20250" + } + ] + } + ], + "index": 39 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 95, + 69, + 517, + 366 + ], + "blocks": [ + { + "bbox": [ + 95, + 69, + 517, + 366 + ], + "lines": [ + { + "bbox": [ + 95, + 69, + 517, + 366 + ], + "spans": [ + { + "bbox": [ + 95, + 69, + 517, + 366 + ], + "type": "table", + "html": "
MethodArchitectureBatchsizeTraining AlgorithmNmIoU
GDSL
hardsoft
Full supervision: standard architectures + full masks training (NLL loss)
Deeplab [12]V212✓---75.6
Deeplab* [13]V3+12✓---76.6
Deeplab* [13]V3+16✓---78.9
Vision Transformer [17]ViT-linear16✓---81.4
Scribble supervision (full-length scribbles)
Modified architectures (△), and/or ad-hoc self-labeling (○)
BPG [41]V2 (△)10✓---73.2
URSS [33]V2 (△)16✓---74.6
SPML [22]V2 (△)16✓---74.2
PSI [44]V3+ (△)---✓(○)-74.9
SEMINAR [10]V3+ (△)12✓---76.2
TEL [27]V3+16--✓(○)-77.1
AGMM [43]ViT-linear (△)16--✓(○)-78.7
Standard architectures + Potts loss optimization (relaxations and/or self-labeling)
ScribbleSup [28]V2 (VGG16)8-✓-DN63.1
DenseCRF loss* [38]V3+12✓--DN75.8
GridCRF loss* [29]V3+12-✓-NN75.6
NonlocalCRF loss* [40]V3+12✓--SN75.7
HCCE + PQV3+12--✓NN77.5
HCCE + PCDV3+12--✓NN77.7
HCCE + PCD (no pretrain)V3+12--✓NN76.7
HCCE + PCD (no pretrain)V3+16--✓NN77.6
HCCE + PCDV3+16--✓NN78.1
HCCE + PCD (no pretrain)V3+16--✓NN77.6
HCCE + PCD (no pretrain)ViT-linear12--✓NN80.8
HCCE + PCD (no pretrain)ViT-linear16--✓NN80.94
", + "image_path": "6512a3766b86c077a3ecbdb0fafb0323372e28bec1e8f999c68539ddf9fac642.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 374, + 555, + 453 + ], + "lines": [ + { + "bbox": [ + 54, + 374, + 555, + 453 + ], + "spans": [ + { + "bbox": [ + 54, + 374, + 555, + 453 + ], + "type": "text", + "content": "Table 5. Comparison of scribble-supervised segmentation methods (without CRF postprocessing). We focus on general solutions applicable to any standard architecture based on mathematically well-founded convergent training procedures optimizing a clearly defined loss (the last block). The second block shows specialized solutions based on architectural modifications and/or ad-hoc training procedures that are often non-convergent and hard to reproduce. The numbers are mIoU on the validation dataset of Pascal VOC 2012. The backbone is ResNet101 unless stated otherwise. V2: deeplabV2. V3+: deeplabV3+. ViT-linear: ViT backbone and linear decoder. " + }, + { + "bbox": [ + 54, + 374, + 555, + 453 + ], + "type": "inline_equation", + "content": "\\mathcal{N}" + }, + { + "bbox": [ + 54, + 374, + 555, + 453 + ], + "type": "text", + "content": ": neighborhood. “*”: reproduced results. GD: gradient descent. SL: self-labeling. “no pretrain” means the segmentation network is not pretrained using cross-entropy on scribbles. In some cases (V3+, batch size 12) soft self-labeling outperforms full supervision (see teal color numbers)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 468, + 294, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 468, + 294, + 576 + ], + "spans": [ + { + "bbox": [ + 54, + 468, + 294, + 576 + ], + "type": "text", + "content": "ness; this softness is even formally related to classification uncertainty in the random walker method [19]. On the other hand, the convexity of quadratic relaxation does not require DN to simplify loss for gradient descent. Then, it is not surprising that NN works better than DN. The latter makes the Potts model behave as a cardinality potential [39], while in the case of NN the Potts model directly regularized the geometric boundary of objects [6]. Figure 7 also compares these approaches across different scribble lengths." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 586, + 179, + 600 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 586, + 179, + 600 + ], + "spans": [ + { + "bbox": [ + 55, + 586, + 179, + 600 + ], + "type": "text", + "content": "3.5. Comparison to SOTA" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 605, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 605, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 54, + 605, + 295, + 713 + ], + "type": "text", + "content": "This section uses DeepLabV3+ with ResNet101 backbone and ViT-linear transformer to evaluate our self-labeling methodology on these standard architectures and to compare with representative SOTA for scribble-based segmentation. Table 5 presents only the results before any post-processing. Our method may outperform the fully-supervised method (teal color numbers for DeepLabV3+ and batch size 12). Our \"text-book\" training technique, i.e. minimization of a well-defined loss, applied to standard architectures out-" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 468, + 555, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 468, + 555, + 492 + ], + "spans": [ + { + "bbox": [ + 313, + 468, + 555, + 492 + ], + "type": "text", + "content": "performs complex WSSS systems requiring architectural modifications and/or ad-hoc multi-stage training procedures." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 502, + 391, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 502, + 391, + 514 + ], + "spans": [ + { + "bbox": [ + 313, + 502, + 391, + 514 + ], + "type": "text", + "content": "4. Conclusions" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 312, + 521, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 521, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 312, + 521, + 556, + 715 + ], + "type": "text", + "content": "This paper proposes a mathematically well-founded self-labeling WSSS framework for standard segmentation networks based on a joint loss (8) w.r.t. network predictions and soft pseudo-labels. The latter was motivated as auxiliary variables simplifying optimization of (4). Conceptual properties and systematic evaluation of the terms in our soft self-labeling loss (8) advocate for the collision cross-entropy and collision divergence. The former replaces the standard cross-entropy " + }, + { + "bbox": [ + 312, + 521, + 556, + 715 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 312, + 521, + 556, + 715 + ], + "type": "text", + "content": ", and the latter is used as a log of a normalized quadratic relaxation of the Potts model " + }, + { + "bbox": [ + 312, + 521, + 556, + 715 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 312, + 521, + 556, + 715 + ], + "type": "text", + "content": " over the NN neighborhood. Our general methodology works with any standard architecture and outperforms complex specialized systems based on architectural modifications and/or ad-hoc multi-stage non-convergent training. In contrast, our ideas are easy to understand, reproduce, and extend to other forms of weakly-supervised segmentation (boxes, class tags, etc.)." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20251" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 297, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 91, + 297, + 134 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 297, + 134 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 297, + 134 + ], + "type": "text", + "content": "[1] Multilabel random walker image segmentation using prior models. In 2005 IEEE computer society conference on computer vision and pattern recognition (CVPR'05), pages 763-770. IEEE, 2005. 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 137, + 296, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 137, + 296, + 190 + ], + "spans": [ + { + "bbox": [ + 61, + 137, + 296, + 190 + ], + "type": "text", + "content": "[2] Jiwoon Ahn and Suha Kwak. Learning pixel-level semantic affinity with image-level supervision for weakly supervised semantic segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4981-4990, 2018. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 194, + 296, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 194, + 296, + 247 + ], + "spans": [ + { + "bbox": [ + 62, + 194, + 296, + 247 + ], + "type": "text", + "content": "[3] Jiwoon Ahn, Sunghyun Cho, and Suha Kwak. Weakly supervised learning of instance segmentation with inter-pixel relations. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2209-2218, 2019. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 250, + 296, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 250, + 296, + 293 + ], + "spans": [ + { + "bbox": [ + 62, + 250, + 296, + 293 + ], + "type": "text", + "content": "[4] Nikita Araslanov and Stefan Roth. Single-stage semantic segmentation from image labels. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4253-4262, 2020. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 295, + 296, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 295, + 296, + 317 + ], + "spans": [ + { + "bbox": [ + 62, + 295, + 296, + 317 + ], + "type": "text", + "content": "[5] Stephen Boyd and Lieven Vandenberghe. Convex optimization. Cambridge university press, 2004. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 319, + 296, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 319, + 296, + 361 + ], + "spans": [ + { + "bbox": [ + 62, + 319, + 296, + 361 + ], + "type": "text", + "content": "[6] Yuri Boykov and Vladimir Kolmogorov. Computing geodesics and minimal surfaces via graph cuts. In International conference on computer vision (ICCV), 2003. 3, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 364, + 296, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 364, + 296, + 407 + ], + "spans": [ + { + "bbox": [ + 62, + 364, + 296, + 407 + ], + "type": "text", + "content": "[7] Yuri Y Boykov and M-P Jolly. Interactive graph cuts for optimal boundary & region segmentation of objects in nd images. In international conference on computer vision (ICCV), pages 105-112, 2001. 2, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 409, + 296, + 452 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 409, + 296, + 452 + ], + "spans": [ + { + "bbox": [ + 62, + 409, + 296, + 452 + ], + "type": "text", + "content": "[8] John Bridle, Anthony Heading, and David MacKay. Unsupervised classifiers, mutual information and phantom targets. Advances in neural information processing systems, 4, 1991. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 455, + 296, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 455, + 296, + 498 + ], + "spans": [ + { + "bbox": [ + 62, + 455, + 296, + 498 + ], + "type": "text", + "content": "[9] Antonin Chambolle and Thomas Pock. A first-order primal-dual algorithm for convex problems with applications to imaging. Journal of mathematical imaging and vision, 40:120-145, 2011. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 501, + 296, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 501, + 296, + 555 + ], + "spans": [ + { + "bbox": [ + 57, + 501, + 296, + 555 + ], + "type": "text", + "content": "[10] Hongjun Chen, Jinbao Wang, Hong Cai Chen, Xiantong Zhen, Feng Zheng, Rongrong Ji, and Ling Shao. Seminar learning for click-level weakly supervised semantic segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6920-6929, 2021. 2, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 557, + 296, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 557, + 296, + 601 + ], + "spans": [ + { + "bbox": [ + 57, + 557, + 296, + 601 + ], + "type": "text", + "content": "[11] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L Yuille. Semantic image segmentation with deep convolutional nets and fully connected crfs. arXiv preprint arXiv:1412.7062, 2014. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 603, + 296, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 603, + 296, + 656 + ], + "spans": [ + { + "bbox": [ + 57, + 603, + 296, + 656 + ], + "type": "text", + "content": "[12] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L Yuille. Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. IEEE transactions on pattern analysis and machine intelligence, 40(4):834-848, 2017. 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 658, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 658, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 658, + 296, + 713 + ], + "type": "text", + "content": "[13] Liang-Chieh Chen, Yukun Zhu, George Papandreou, Florian Schroff, and Hartwig Adam. Encoder-decoder with atrous separable convolution for semantic image segmentation. In Proceedings of the European conference on computer vision (ECCV), pages 801-818, 2018. 6, 8" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 316, + 73, + 555, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 555, + 137 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 555, + 137 + ], + "type": "text", + "content": "[14] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 140, + 555, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 140, + 555, + 183 + ], + "spans": [ + { + "bbox": [ + 317, + 140, + 555, + 183 + ], + "type": "text", + "content": "[15] Camille Couprie, Leo Grady, Laurent Najman, and Hugues Talbot. Power watershed: A unifying graph-based optimization framework. IEEE transactions on pattern analysis and machine intelligence, 33(7):1384-1399, 2010. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 185, + 555, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 185, + 555, + 227 + ], + "spans": [ + { + "bbox": [ + 317, + 185, + 555, + 227 + ], + "type": "text", + "content": "[16] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248–255. IEEE, 2009. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 228, + 555, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 228, + 555, + 259 + ], + "spans": [ + { + "bbox": [ + 317, + 228, + 555, + 259 + ], + "type": "text", + "content": "[17] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 6, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 262, + 555, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 262, + 555, + 304 + ], + "spans": [ + { + "bbox": [ + 317, + 262, + 555, + 304 + ], + "type": "text", + "content": "[18] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. International journal of computer vision, 88:303-308, 2009. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 306, + 554, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 306, + 554, + 337 + ], + "spans": [ + { + "bbox": [ + 316, + 306, + 554, + 337 + ], + "type": "text", + "content": "[19] Leo Grady. Random walks for image segmentation. IEEE transactions on pattern analysis and machine intelligence, 28 (11):1768-1783, 2006. 2, 3, 7, 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 339, + 555, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 339, + 555, + 371 + ], + "spans": [ + { + "bbox": [ + 316, + 339, + 555, + 371 + ], + "type": "text", + "content": "[20] Yves Grandvalet and Yoshua Bengio. Semi-supervised learning by entropy minimization. Advances in neural information processing systems, 17, 2004. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 373, + 555, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 373, + 555, + 415 + ], + "spans": [ + { + "bbox": [ + 316, + 373, + 555, + 415 + ], + "type": "text", + "content": "[21] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 417, + 555, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 417, + 555, + 458 + ], + "spans": [ + { + "bbox": [ + 316, + 417, + 555, + 458 + ], + "type": "text", + "content": "[22] Tsung-Wei Ke, Jyh-Jing Hwang, and Stella X Yu. Universal weakly supervised segmentation by pixel-to-segment contrastive learning. arXiv preprint arXiv:2105.00957, 2021. 2, 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 460, + 555, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 460, + 555, + 525 + ], + "spans": [ + { + "bbox": [ + 317, + 460, + 555, + 525 + ], + "type": "text", + "content": "[23] Alexander Kolesnikov and Christoph H Lampert. Seed, expand and constrain: Three principles for weakly-supervised image segmentation. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part IV 14, pages 695-711. Springer, 2016. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 527, + 555, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 527, + 555, + 567 + ], + "spans": [ + { + "bbox": [ + 317, + 527, + 555, + 567 + ], + "type": "text", + "content": "[24] Philipp Krahenbuhl and Vladlen Koltun. Efficient inference in fully connected CRFs with Gaussian edge potentials. Advances in neural information processing systems, 24, 2011. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 571, + 555, + 603 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 571, + 555, + 603 + ], + "spans": [ + { + "bbox": [ + 317, + 571, + 555, + 603 + ], + "type": "text", + "content": "[25] V. Kulharia, S. Chandra, A. Agrawal, P. Torr, and A. Tyagi. Box2seg: Attention weighted loss and discriminative feature learning for weakly supervised segmentation. In ECCV'20. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 604, + 555, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 604, + 555, + 658 + ], + "spans": [ + { + "bbox": [ + 317, + 604, + 555, + 658 + ], + "type": "text", + "content": "[26] Jungbeom Lee, Eunji Kim, Sungmin Lee, Jangho Lee, and Sungroh Yoon. Ficklenet: Weakly and semi-supervised semantic image segmentation using stochastic inference. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5267-5276, 2019. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 659, + 555, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 659, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 659, + 555, + 713 + ], + "type": "text", + "content": "[27] Zhiyuan Liang, Tiancai Wang, Xiangyu Zhang, Jian Sun, and Jianbing Shen. Tree energy loss: Towards sparsely annotated semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16907-16916, 2022. 2, 3, 6, 8" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "20252" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 297, + 715 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 56, + 73, + 297, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 297, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 297, + 127 + ], + "type": "text", + "content": "[28] Di Lin, Jifeng Dai, Jiaya Jia, Kaiming He, and Jian Sun. Scribblesup: Scribble-supervised convolutional networks for semantic segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3159-3167, 2016. 1, 2, 6, 7, 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 296, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 296, + 173 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 296, + 173 + ], + "type": "text", + "content": "[29] Dmitrii Marin and Yuri Boykov. Robust trust region for weakly supervised segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6608-6618, 2021. 2, 7, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 175, + 296, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 296, + 228 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 296, + 228 + ], + "type": "text", + "content": "[30] Dmitrii Marin, Meng Tang, Ismail Ben Ayed, and Yuri Boykov. Beyond gradient descent for regularized segmentation losses. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10187-10196, 2019. 2, 3, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 230, + 296, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 230, + 296, + 263 + ], + "spans": [ + { + "bbox": [ + 56, + 230, + 296, + 263 + ], + "type": "text", + "content": "[31] Rafael Müller, Simon Kornblith, and Geoffrey E Hinton. When does label smoothing help? Advances in neural information processing systems, 32, 2019. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 265, + 296, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 265, + 296, + 297 + ], + "spans": [ + { + "bbox": [ + 56, + 265, + 296, + 297 + ], + "type": "text", + "content": "[32] NSD. Natural Scenes Dataset [NSD]. https://www.kaggle.com/datasets/nitishabharathi/scene-classification, 2020.5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 298, + 296, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 298, + 296, + 364 + ], + "spans": [ + { + "bbox": [ + 56, + 298, + 296, + 364 + ], + "type": "text", + "content": "[33] Zhiyi Pan, Peng Jiang, Yunhai Wang, Changhe Tu, and Anthony G Cohn. Scribble-supervised semantic segmentation by uncertainty reduction on neural representation and self-supervision on neural eigenspace. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7416-7425, 2021. 2, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 365, + 296, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 365, + 296, + 409 + ], + "spans": [ + { + "bbox": [ + 56, + 365, + 296, + 409 + ], + "type": "text", + "content": "[34] Thomas Pock, Antonin Chambolle, Daniel Cremers, and Horst Bischof. A convex relaxation approach for computing minimal partitions. In IEEE Conference on Computer Vision and Pattern Recognition, pages 810-817, 2009. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 411, + 296, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 411, + 296, + 454 + ], + "spans": [ + { + "bbox": [ + 56, + 411, + 296, + 454 + ], + "type": "text", + "content": "[35] Pradeep Ravikumar and John Lafferty. Quadratic programming relaxations for metric labeling and Markov Random Field MAP estimation. In The 23rd International Conference on Machine Learning, page 737-744, 2006. 2, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 456, + 296, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 456, + 296, + 510 + ], + "spans": [ + { + "bbox": [ + 56, + 456, + 296, + 510 + ], + "type": "text", + "content": "[36] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zhoginov, and Liang-Chieh Chen. *Mobilenetv2: Inverted residuals and linear bottlenecks*. In *Proceedings of the IEEE conference on computer vision and pattern recognition*, pages 4510–4520, 2018. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 511, + 296, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 511, + 296, + 566 + ], + "spans": [ + { + "bbox": [ + 56, + 511, + 296, + 566 + ], + "type": "text", + "content": "[37] Meng Tang, Abdelaziz Djelouah, Federico Perazzi, Yuri Boykov, and Christopher Schroers. Normalized cut loss for weakly-supervised cnn segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1818-1827, 2018. 1, 6, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 567, + 296, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 567, + 296, + 622 + ], + "spans": [ + { + "bbox": [ + 56, + 567, + 296, + 622 + ], + "type": "text", + "content": "[38] Meng Tang, Federico Perazzi, Abdelaziz Djelouah, Ismail Ben Ayed, Christopher Schroers, and Yuri Boykov. On regularized losses for weakly-supervised cnn segmentation. In Proceedings of the European Conference on Computer Vision (ECCV), pages 507-522, 2018. 1, 2, 6, 7, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 624, + 296, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 296, + 656 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 296, + 656 + ], + "type": "text", + "content": "[39] Olga Veksler. Efficient graph cut optimization for full crfs with quantized edges. IEEE transactions on pattern analysis and machine intelligence, 42(4):1005-1012, 2019. 7, 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 658, + 296, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 296, + 690 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 296, + 690 + ], + "type": "text", + "content": "[40] Olga Veksler and Yuri Boykov. Sparse non-local crf. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4493-4503, 2022. 2, 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 692, + 296, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 692, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 56, + 692, + 296, + 715 + ], + "type": "text", + "content": "[41] Bin Wang, Guojun Qi, Sheng Tang, Tianzhu Zhang, Yunchao Wei, Linghui Li, and Yongdong Zhang. Boundary perception" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 353 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 335, + 73, + 555, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 555, + 106 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 555, + 106 + ], + "type": "text", + "content": "guidance: A scribble-supervised semantic segmentation approach. In IJCAI International joint conference on artificial intelligence, 2019. 2, 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 107, + 555, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 555, + 162 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 555, + 162 + ], + "type": "text", + "content": "[42] Changwei Wang, Rongtao Xu, Shibiao Xu, Weiliang Meng, and Xiaopeng Zhang. Treating pseudo-labels generation as image matting for weakly supervised semantic segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 755-765, 2023. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 163, + 555, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 555, + 206 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 555, + 206 + ], + "type": "text", + "content": "[43] Linshan Wu, Zhun Zhong, Jiayi Ma, Yunchao Wei, Hao Chen, Leyuan Fang, and Shutao Li. Modeling the label distributions for weakly-supervised semantic segmentation. arXiv preprint arXiv:2403.13225, 2024. 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 208, + 555, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 208, + 555, + 262 + ], + "spans": [ + { + "bbox": [ + 316, + 208, + 555, + 262 + ], + "type": "text", + "content": "[44] Jingshan Xu, Chuanwei Zhou, Zhen Cui, Chunyan Xu, Yuge Huang, Pengcheng Shen, Shaoxin Li, and Jian Yang. Scribblesupervised semantic segmentation inference. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15354-15363, 2021. 2, 3, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 263, + 555, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 263, + 555, + 307 + ], + "spans": [ + { + "bbox": [ + 316, + 263, + 555, + 307 + ], + "type": "text", + "content": "[45] Christopher Zach, Christian Hane, and Marc Pollefeys. What is optimized in tight convex relaxations for multi-label problems? In IEEE Conference on Computer Vision and Pattern Recognition, pages 1664-1671, 2012. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 308, + 555, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 308, + 555, + 353 + ], + "spans": [ + { + "bbox": [ + 316, + 308, + 555, + 353 + ], + "type": "text", + "content": "[46] Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 633-641, 2017. 6" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20253" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/SoftShadow_ Leveraging Soft Masks for Penumbra-Aware Shadow Removal/1a57ac9e-bdab-4cc5-a82c-bea51caf54d0_content_list.json b/2025/SoftShadow_ Leveraging Soft Masks for Penumbra-Aware Shadow Removal/1a57ac9e-bdab-4cc5-a82c-bea51caf54d0_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..5382fb3491faab18fc267eb8b7c50b2daaf58a6a --- /dev/null +++ b/2025/SoftShadow_ Leveraging Soft Masks for Penumbra-Aware Shadow Removal/1a57ac9e-bdab-4cc5-a82c-bea51caf54d0_content_list.json @@ -0,0 +1,2069 @@ +[ + { + "type": "text", + "text": "SoftShadow: Leveraging Soft Masks for Penumbra-Aware Shadow Removal", + "text_level": 1, + "bbox": [ + 114, + 130, + 883, + 151 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xinrui Wang $^{1}$ , Lanqing Guo $^{2}$ , Xiyu Wang $^{1}$ , Siyu Huang $^{3}$ , Bihan Wen $^{1\\dagger}$ $^{1}$ Nanyang Technological University, Singapore \n $^{2}$ The University of Texas at Austin, USA \n $^{3}$ Clemson University, USA \n{xinrui.wang, bihan.wen}@ntu.edu.sg, xiyu004@e.ntu.edu.sg, lanqing.guo@austin.utexas.edu, siyuh@clemson.edu", + "bbox": [ + 220, + 178, + 781, + 286 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 321, + 326, + 338 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent advancements in deep learning have yielded promising results for the image shadow removal task. However, most existing methods rely on binary pre-generated shadow masks. The binary nature of such masks could potentially lead to artifacts near the boundary between shadow and non-shadow areas. In view of this, inspired by the physical model of shadow formation, we introduce novel soft shadow masks specifically designed for shadow removal. To achieve such soft masks, we propose a SoftShadow framework by leveraging the prior knowledge of pretrained SAM and integrating physical constraints. Specifically, we jointly tune the SAM and the subsequent shadow removal network using penumbra formation constraint loss, mask reconstruction loss, and shadow removal loss. This framework enables accurate predictions of penumbra (partially shaded) and umbra (fully shaded) areas while simultaneously facilitating end-to-end shadow removal. Through extensive experiments on popular datasets, we found that our SoftShadow framework, which generates soft masks, can better restore boundary artifacts, achieve state-of-the-art performance, and demonstrate superior generalizability.", + "bbox": [ + 89, + 354, + 483, + 672 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 700, + 220, + 717 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Shadow removal aims to restore content obscured in shadow regions and correct degraded illumination. Recently, deep learning methods have shown excellent performance in shadow removal tasks relying on large-scale training data. However, one of the main challenges in shadow removal arises in images with soft shadows, where the shadow edges are blurred rather than sharp. This blurred area, known as the penumbra area, occurs due to partial occlusion of light, creating a transition zone where illumination varies signif", + "bbox": [ + 89, + 727, + 482, + 863 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ae3c4a2e397052d05a3208cf1a1029d799c251ea051cade9b83adf59a35af015.jpg", + "image_caption": [ + "Input" + ], + "image_footnote": [], + "bbox": [ + 522, + 323, + 612, + 390 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/67e9c45e2022f38ef0bb891db4b09294360a58ec0f3a780ee648ce2b3e0cca15.jpg", + "image_caption": [ + "Ground Truth" + ], + "image_footnote": [], + "bbox": [ + 617, + 324, + 705, + 390 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6e1e11d3db043007d7f6f8955ec7469e228e8d42d91d556ac2895721a6062941.jpg", + "image_caption": [ + "Hard Mask" + ], + "image_footnote": [], + "bbox": [ + 712, + 323, + 800, + 390 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/19c30de583a65975a35143c95c28edfa7f8f99ba3b237a54341be51825713016.jpg", + "image_caption": [ + "BMNet" + ], + "image_footnote": [], + "bbox": [ + 519, + 401, + 609, + 469 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1b1236cd360fc0784e1a3091948197621d5f59d9324f67d0db23dbe8b69e72e3.jpg", + "image_caption": [ + "SG-ShadowNet" + ], + "image_footnote": [], + "bbox": [ + 617, + 401, + 705, + 469 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5f13ef3994af938eb7618caabf0a0d3c3dedee560971aba38f9c1777587373ef.jpg", + "image_caption": [ + "HomoFormer", + "Figure 1. Illustration of soft shadow removal results using our proposed SoftShadow with soft shadow mask compared to the recent competing methods BMNet [39], SG-ShadowNet [30] and HomoFormer [34] using hard shadow masks. The second rows are sharpened versions of results for better visualization." + ], + "image_footnote": [], + "bbox": [ + 712, + 401, + 800, + 469 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7da857832245867069b99057960d91d50bab59dc1b082f5f780e9c58dd8e5292.jpg", + "image_caption": [ + "Soft Mask" + ], + "image_footnote": [], + "bbox": [ + 808, + 321, + 898, + 390 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d1e43dda1e4a1f971d21d3c2c2eff173642e016567d5a3bc2ee2656992228380.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 807, + 401, + 897, + 469 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "icantly. Removing shadows in these regions is challenging because the gradual light-to-shadow transition often leads to boundary artifacts, compromising the quality of the shadow removal results.", + "bbox": [ + 511, + 595, + 906, + 655 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "A few methods rely on hard shadow masks (binary masks) to indicate the shadow regions [8, 9, 34, 39]. These binary shadow masks are either manually annotated [23, 31], or predicted by off-the-shelf shadow detectors [3]. Obtaining such shadow masks can be costly and complicated, and the choice of shadow detector can significantly impact shadow removal performance. Moreover, hard masks fall short of representing the penumbra area, causing boundary artifacts in removal results, especially when the penumbra areas are highly pronounced. As shown in Figure 1, when the input image contains penumbra regions, the previous state-of-the-art (SOTA) method [34] suffers from boundary artifacts. Other approaches attempt shadow removal without explicitly extracting the shadow mask. Instead, they incorporate semantic parsing modules into shadow removal networks. For instance, some approaches involve predict", + "bbox": [ + 511, + 659, + 908, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution for both authors", + "bbox": [ + 107, + 875, + 303, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author", + "bbox": [ + 111, + 887, + 233, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "23217", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c01c5cf9d5db6c178fb4b2f1ddad8409cea2696767574445619bc2746d04510e.jpg", + "image_caption": [ + "Figure 2. (a) Illustration of the shadow formation geometry that creates the penumbra and umbra [26] regions; the umbra area is where the light is fully occluded, the penumbra area is where light is partially occluded, the lit area is where light is not occluded. (b) Examples of soft shadow images and hard shadow images from commonly used datasets." + ], + "image_footnote": [], + "bbox": [ + 106, + 88, + 264, + 246 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/4ca7a5d78bf1d8ac1fa537ee83766e6540d9684b8b023c6dda47621b57d99b73.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 274, + 88, + 470, + 246 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ing degradation attention [3, 27], while others utilize domain classifiers [16, 17] to better understand and address shadow effects. However, their performance may be limited by the lack of powerful external detectors providing additional shadow location information.", + "bbox": [ + 89, + 372, + 482, + 446 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In view of this, we argue that simply leveraging the binary mask to represent the shadow location is not enough. Differently, we introduce a novel soft shadow mask (grayscale mask) specifically designed for shadow removal as shown in the hard mask v.s. soft mask in Figure 1. The soft masks can precisely locate the inner and outer boundaries of shadow and indicate the penumbra area with proper degradation variance. Moreover, recognizing the effectiveness of employing a powerful pretrained detector for providing semantic information, we utilize the prior knowledge from the pretrained SAM [19]. While existing methods [35] have used the off-the-shelf SAM in a naive way, merely predicting the shadow mask as input to the shadow removal network, we adapt SAM as a soft shadow mask predictor and jointly optimize it with the shadow removal network.", + "bbox": [ + 89, + 448, + 482, + 672 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we first propose a unified shadow removal framework, dubbed SoftShadow, which removes shadow from images in an end-to-end manner without requiring input shadow masks as shown in Figure 3. We integrate the pretrained SAM [19] combined with Low-Rank Adaptation (LoRA) [13] to jointly adapt the SAM model and shadow removal model. The framework enables accurate predictions of soft shadow masks as intermediate results while simultaneously facilitating end-to-end shadow removal. Besides, we introduce a penumbra formation constraint to assist predict soft shadow masks. This constraint regularizes the gradient of the predicted mask in the penumbra area, resulting in more accurate predictions with detailed and spatially varied shadow masks. In the commonly used shadow removal datasets, SRD and LRSS datasets contain", + "bbox": [ + 89, + 674, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "soft shadow images, which makes them more suitable for illustrating the improvements of utilizing soft shadow masks. Experimental results show that SoftShadow achieves superior performance on SRD and LRSS datasets. Our main contributions are summarized as follows:", + "bbox": [ + 511, + 90, + 903, + 165 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a unified shadow removal framework that does not require additional input shadow masks. This framework enables accurate predictions of soft masks as intermediate results, which is specifically designed to capture detailed and spatially varied shadow location information.", + "- We introduce penumbra formation constraint inspired by the physical shadow formation model to further refine the soft mask in the penumbra area. By leveraging the constraint loss and shadow removal loss, we jointly tune the SAM and the subsequent shadow removal network.", + "- Experimental results demonstrate that SoftShadow surpasses state-of-the-art shadow removal methods on the SRD and LRSS datasets, achieving superior performance and even comparable results with previous methods that use ground truth mask inputs." + ], + "bbox": [ + 513, + 171, + 903, + 412 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 428, + 653, + 444 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Shadow Removal", + "text_level": 1, + "bbox": [ + 513, + 454, + 683, + 469 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The degradation of shadows varies in each image, posing a significant challenge for shadow image restoration. Recent deep learning-based approaches have achieved remarkable results in shadow removal. Some methods restore the shadow image with the guidance of shadow masks. For example, SP+M-Net [21] employs two deep networks to predict shadow matte and shadow parameters. Recently, powerful backbones such as transformers [4, 29], and diffusion models [12, 28] have been applied to the shadow removal task. HomoFormer [34] homogenizes the spatial distribution of shadow masks to uniformly recover the entire shadow image, while ShadowDiffusion [9] provides a robust generative prior for producing natural shadow-free images. Some methods aim to eliminate the dependency on shadow mask inputs. DeShadowNet [27] introduces an end-to-end shadow removal method containing a multi-branch fusion module. ST-CGAN [31] connects two GANs [6] in sequence to jointly detect and remove shadows. DC-ShadowNet [16] provides an unsupervised domain-classifier discriminator for guided shadow removal network. More recently, DeS3 [17] implemented a method capable of removing shadows that are cast on the object itself without requiring shadow mask inputs. However, due to the absence of additional shadow location information, their performance may degrade, and the number of parameters used could increase significantly. Different from these existing methods, we introduce a new concept \"soft mask\", designed explicitly for shadow removal.", + "bbox": [ + 511, + 477, + 906, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "23218", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/a7ca1b72b53a7e66f7e79e36318a5bf2df0b9a26991544554e71ec1eb761e1d7.jpg", + "image_caption": [ + "Figure 3. Illustration of the proposed SoftShadow. The left box illustrates the SoftShadow networks, where a shadow image $\\mathbf{y}$ is input into SAM for detecting soft shadow masks. The shadow removal network then processes the soft mask and the shadow image to produce a shadow-free image. The right box shows the three losses we used in SoftShadow. From top to bottom, the shadow removal loss $\\mathcal{L}_{rem}$ is calculated between shadow-free images and shadow-removal images. The mask reconstruction loss $\\mathcal{L}_{mask}$ is calculated between predicted soft masks and ours ground truth soft masks. The penumbra formation constraint loss $\\mathcal{L}_{pen}$ acts as a regularization term. It aims to regularize the gradient of predicted soft masks in the penumbra area. As shown in the soft mask intensity curve, the ideal mask intensity in the penumbra area should not be too large and the gradient orientation should be consistent." + ], + "image_footnote": [], + "bbox": [ + 171, + 85, + 823, + 287 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Shadow Detection", + "text_level": 1, + "bbox": [ + 89, + 422, + 264, + 436 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Shadow detection can be approached as a segmentation task, it often struggles with small shadows and indistinct shadow edges. Various methods have been proposed to address these challenges. BDRAR [37] introduces a bidirectional pyramidal architecture for shadow detection. DSD [36] designs a distraction-aware module to minimize false positives in shadow detection. Chen et al. [2] propose a task-specific semi-supervised learning mechanism to utilize unlabeled shadow images for detection, enhancing the robustness of the model. DHAN [3] is a well-used shadow detection model using dual hierarchical aggregation network training on synthetic shadow and shadow-free image pairs to improve detection accuracy. The SAM-Adapter [1] has been utilized to fine-tune the Segment Anything Model (SAM) [19] specifically for shadow detection tasks. While this approach improves accuracy, it demands a substantial number of model parameters. However, most existing methods detect shadows as hard masks and fail to represent the penumbra area. In contrast, our approach leverages the shadow removal loss and the Penumbra Formation Constraint loss, jointly training the shadow removal network and the shadow detection network.", + "bbox": [ + 91, + 446, + 483, + 779 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methods", + "text_level": 1, + "bbox": [ + 89, + 797, + 189, + 813 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we first explain the motivation behind our newly introduced soft shadow mask, inspired by the physical model of shadow formation. We then propose the unified shadow removal framework SoftShadow, which leverages the powerful segmentation capabilities of SAM to pro", + "bbox": [ + 89, + 825, + 482, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "duce the soft shadow mask and guide the removal network to generate better shadow-free images. Finally, to further mitigate boundary artifacts, we introduce the penumbra formation constraint loss which ensures a smoother transition between the shadow region and the non-shadow region, offering better guidance to the shadow removal network.", + "bbox": [ + 511, + 422, + 906, + 513 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Motivation", + "text_level": 1, + "bbox": [ + 511, + 527, + 635, + 541 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Shadow degradations in the real world exhibit considerable amount of variation and can be classified into two main categories: soft shadows and hard shadows [20]. This classification depends on the light source and the distance between the object and the surface, as shown in Figure 2(a). Soft shadows are characterized by blurry edges and gradual transitions from light to dark, creating relatively large penumbra areas, while hard shadows have sharp edges with few to none penumbra regions.", + "bbox": [ + 511, + 550, + 906, + 686 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Previous methods [5, 8, 34, 39] commonly use hard masks to guide the shadow removal process. The hard masks effectively provide positional guidance for hard masks with sharp edges. However, when the penumbra regions are more pronounced, as shown in Figure 2 (b), the boundary between the umbra and the lit region is hard to determine, making hard masks unsuitable for representing such soft shadows. Additionally, we observed that, in general, the brightness transition within the penumbra area can vary significantly, whereas the brightness of the lit and the umbra areas are relatively uniform.", + "bbox": [ + 511, + 686, + 908, + 852 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To this end, we instead use soft (greyscale) masks to represent the position of the shadows. Formally, we denote the soft mask as s. For a given shadow-free image $\\mathbf{x}$ , the", + "bbox": [ + 511, + 854, + 906, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "23219", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "shadow image $\\mathbf{y}$ can be formulated as such:", + "bbox": [ + 89, + 90, + 382, + 104 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {y} = \\mathbf {a} \\cdot \\mathbf {s} \\cdot \\mathbf {x} + (\\mathbf {1} - \\mathbf {s}) \\cdot \\mathbf {x}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 192, + 119, + 482, + 136 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\cdot$ denotes element-wise multiplication, $s$ is the soft shadow mask in which constant $s = 0$ represents lit area, $s = 1$ represents umbra area, and $s \\in [0,1]$ represents penumbra area. $a \\in [0,1]$ represents the illumination weight in the non-lit area. Based on our experiment, we quantitatively and qualitatively found that the proposed shadow degradation model could lead to fewer boundary artifacts and generally better shadow removal results.", + "bbox": [ + 89, + 148, + 483, + 270 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Architecture of SoftShadow", + "text_level": 1, + "bbox": [ + 89, + 280, + 339, + 294 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "SoftShadow is a unified framework designed to remove object shadows in images in an end-to-end manner. The overall architecture is illustrated in Figure 3. SoftShadow contains a soft shadow detector and a subsequent shadow removal network. The soft shadow detector leverages the strong segmentation capabilities of SAM and further finetuning it to serve as a soft mask detector. This is because the pretrained SAM can fail to accurately identify shadows [15, 19]. Moreover, since SAM has noticeably more parameters compared to conventional shadow detection networks, we use Low-Rank Adaptation (LoRA) [13] to reduce the number of trainable parameters. In practice, we tune all the self-attention blocks of image encoders using LoRA. Since the mask decoder is lightweight, we tune all the parameters of it.", + "bbox": [ + 89, + 301, + 483, + 527 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "SAM is mainly trained for binary predictions, unlike the soft masks we aim to generate. To leverage SAM to generate continuous soft masks, we further introduce the mask reconstruction loss $\\mathcal{L}_{mask}$ and the penumbra formation constraint loss $\\mathcal{L}_{pen}$ to model the characteristics of soft masks (details are described in Section 3.3). They encourage the model to detect the shadow area and produce a continuous soft mask rather than a binary one, thereby capturing the shadow position more accurately. We formulate a mask reconstruction loss $\\mathcal{L}_{mask}$ as follows:", + "bbox": [ + 89, + 529, + 483, + 680 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {m a s k} = \\mathbb {E} _ {n} \\left\\| \\hat {\\mathbf {s}} - \\mathbf {s} _ {g t} \\right\\| _ {F} ^ {2}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 196, + 691, + 482, + 712 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $n$ is the index of shadow masks, $\\hat{\\mathbf{s}}$ is the predicted soft shadow mask, and $\\mathbf{s}_{gt}$ is the ground truth soft shadow mask achieved by dividing the shadow-free image $\\mathbf{y}$ by the shadow image $\\mathbf{x}$ . Specifically, we convert $\\mathbf{y}$ and $\\mathbf{x}$ into YCbCr image space, using the Y channel to illustrate the brightness of these images, denoted as $\\mathbf{y}_Y$ and $\\mathbf{x}_Y$ , respectively. We then divide the $\\mathbf{y}_Y$ by $\\mathbf{x}_Y$ to obtain the difference between them, apply a low-pass filter $f$ to reduce noise, and use a threshold to eliminate outliers as follows:", + "bbox": [ + 89, + 722, + 483, + 857 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {s} _ {g t} = \\max \\left(t, f \\left(\\frac {\\mathbf {x} _ {Y}}{\\mathbf {y} _ {Y}}\\right)\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 191, + 869, + 482, + 905 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/3495dc65292d8cfc40319ed1daacfd1a6ad2258dfba97d292179c530b6891aae.jpg", + "image_caption": [ + "Figure 4. Illustrate the concept of Penumbra Formation Constraint. The $w_{0}$ and $w_{1}$ are points in the inner boundary and the outer boundary of the penumbra area, respectively. In the ideal soft shadow mask, the intensity in the penumbra area decreases progressively from the shadow center to the umbra area. The gradient orientation points from the shadow center towards the non-shadow area." + ], + "image_footnote": [], + "bbox": [ + 578, + 87, + 836, + 250 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $t$ is the threshold used to define the outer boundary between the penumbra area and the lit area.", + "bbox": [ + 511, + 388, + 903, + 419 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Moreover, as illustrate in Figure 3, we incorporet the gradient of the shadow removal loss $\\mathcal{L}_{rem}$ to jointly train SAM and the subsequent shadow removal network:", + "bbox": [ + 511, + 420, + 903, + 465 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {r e m}} = \\mathbb {E} _ {n} \\| \\hat {\\mathbf {x}} - \\mathbf {x} \\| _ {F} ^ {2}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 627, + 476, + 903, + 496 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $n$ indexes all images, $\\hat{\\mathbf{x}}$ is the restored shadow-free image, and $\\mathbf{x}$ is the ground truth shadow-free image. To this end, the intermediate soft shadow masks can be refined using shadow removal results, with the SAM adapter optimized for better location guidance in subsequent shadow removal.", + "bbox": [ + 511, + 507, + 905, + 595 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The overall training objective of SoftShadow is:", + "bbox": [ + 532, + 598, + 849, + 613 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {\\text {m a s k}} + \\lambda_ {1} \\mathcal {L} _ {\\text {p e n}} + \\lambda_ {2} \\mathcal {L} _ {\\text {r e m}}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 594, + 627, + 903, + 643 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\lambda_{1}$ and $\\lambda_{2}$ are weighting coefficients to balance the influence of each term. With the merits of joint training and physical constraints, our SoftShadow framework effectively removes shadows.", + "bbox": [ + 511, + 656, + 905, + 715 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Penumbra Formation Constraint", + "text_level": 1, + "bbox": [ + 511, + 727, + 803, + 741 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We introduce the Penumbra Formation Constraint to enable SAM to predict the soft masks more accurately. The objective of this constraint is to regularize SAM to learn a mask that can reflect the illuminance change of shadow images. To meet the objective, we present two assumptions for our proposed soft mask. For a given shadow image, the soft mask intensity in the penumbra region should gradually transition outward from the shadow center to the outer edge of the shadow. This means that the gradient direction in the penumbra should point from the shadow center toward", + "bbox": [ + 511, + 750, + 906, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "23220", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the boundary of the shadow. Additionally, since the intensity change in this area is gradual, the gradient magnitude should remain relatively small to reflect a smooth transition. Based on these assumptions, we can formulate the penumbra formation constraint. For a predicted soft mask $\\mathbf{s}$ , we define the penumbra area using two thresholds $t_1$ and $t_2$ as", + "bbox": [ + 89, + 90, + 480, + 183 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {W} = \\{(i, j) \\mid t _ {1} \\leq \\mathbf {s} _ {i, j} \\leq t _ {2} \\}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 184, + 195, + 482, + 212 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $i$ and $j$ denote the spatial coordinates of pixels within the soft mask s. As shown in Figure 4, the center of the shadow area is denoted as c. c is defined as the mean of the $x-$ and $y-$ coordinates of the penumbra area. The unit vector representing the direction from c to a point w in the penumbra area $\\mathcal{W}$ can be expressed as:", + "bbox": [ + 89, + 223, + 483, + 315 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nd (w) = \\frac {w - c}{\\| w - c \\|}, \\quad \\text {f o r} w \\in \\mathcal {W}, w \\neq c. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 150, + 327, + 482, + 357 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Following the assumptions we proposed, we can define the penumbra formation constraint loss $\\mathcal{L}_{pen}$ as:", + "bbox": [ + 89, + 368, + 482, + 400 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {p e n}} = \\mathbb {E} _ {n, w} \\left[ R (\\mathbf {d} (w) \\cdot \\nabla M (w)) \\right], \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 166, + 411, + 482, + 429 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $n$ is the index of shadow masks, the desired direction of gradient is $\\mathbf{d}(w)$ , and $\\cdot$ indicate the dot product. $R$ refers to the ReLU function [25]. The ReLU function is applied to filter out gradients conflict with the desired direction. The constraint loss regularizes the magnitude of gradients in alignment with the desired direction, promoting a smooth intensity transition within the penumbra region. By penalizing quick changes in intensity, this constraint encourages the detector to generate softer shadow masks.", + "bbox": [ + 89, + 440, + 483, + 578 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 592, + 223, + 608 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Setups", + "text_level": 1, + "bbox": [ + 89, + 617, + 287, + 633 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation details The shadow removal backbone in our SoftShadow framework can be any shadow removal network. We use ShadowDiffusion [9] as the backbone example for all experiments. We finetune ShadowDiffusion at a resolution of $256 \\times 256$ , followed by previous methods [9]. We empirically set the threshold $t = 0.76$ to get the ground truth soft shadow mask $\\mathbf{s}_{gt}$ . We employ the ViT-H model as the backbone for SAM [19]. A set of LoRA [13] layers with a rank of 8 is added to the self-attention blocks in the image encoder of SAM. We use Adam [18] optimizer. The training batch size is 16. We set $\\lambda_{1} = 0.1$ and $\\lambda_{2} = 1$ . For more training details, please refer to the supplementary.", + "bbox": [ + 89, + 638, + 482, + 835 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Benchmark datasets We work with four benchmark datasets for the various shadow removal experiments. SRD Dataset [27] consists of 2,680 training pairs and 408 testing pairs of shadow and shadow-free images. Notably, the SRD", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "dataset does not provide shadow masks, the previous methods using SRD commonly using masks detected by DHAN [3] methods. LRSS Dataset [7] is a specifically designed soft shadow dataset that includes 137 images. We select 48 paired shadow and shadow-free images as our testing set. UIUC Dataset [11] contains 76 pairs of images for testing. It features a variety of shadow types [17], including soft, hard, and self shadows, which provides a diverse and challenging set of conditions for evaluating shadow removal methods. $\\mathrm{ISTD + }$ Dataset [21] is adjusted ISTD dataset [31], consists of 1330 training pairs and 540 testing pairs of shadow and shadow-free images. $\\mathrm{ISTD + }$ has manually annotated binary masks.", + "bbox": [ + 511, + 90, + 903, + 287 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation metrics Following previous works [3, 9, 34], we employ the Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index (SSIM) [32], and Mean Absolute Error (MAE) [33] as quantitative evaluation metrics. We also calculate all metrics for shadow areas, non-shadow areas, and all pixels between ground truth shadow-free images and generated removal results.", + "bbox": [ + 511, + 291, + 903, + 397 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Comparison with State-of-the-Art", + "text_level": 1, + "bbox": [ + 511, + 406, + 810, + 422 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We compare our proposed method with several state-of-the-art shadow removal methods. We include methods that do not require shadow masks as input, e.g., DCshadowNet [16], DeS3 [17], and SAM-helps-shadow [35]. And methods that require shadow masks, including DHAN [3], BMNet [39], SG-ShadowNet [30], ShadowFormer [8], Inpaint4Shadow [22], ShadowDiffusion [9] and Homoformer [34]. To demonstrate the generalizability of our method, we also compare it with zero-shot methods including BCDiff [10] and G2R-ShadowNet [24] on the LRSS [7] and UIUC [11] datasets.", + "bbox": [ + 511, + 429, + 906, + 595 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Quantitative results Table 1 shows the quantitative results of SRD and $\\mathrm{ISTD + }$ datasets. The SRD dataset contains many soft shadow images, which is more suitable for validating our methods. Specifically, we outperform all competing methods over all metrics, whether they require input shadow masks or not. Compared with the most recent work without requiring mask input, i.e., DeS3 [17], the PSNR is improved from 34.11 dB to 35.57 dB in whole images. When compared with methods that need input masks, we outperform the SOTA method Homoformer [34], which uses DHAN [3] mask as their input. On the $\\mathrm{ISTD + }$ dataset, we significantly outperform DeS3 [17], increasing the PSNR from 31.38 dB to 35.00 dB. For methods that rely on the ground truth masks from the $\\mathrm{ISTD + }$ dataset, we evaluate them using masks detected by the FDRNet [38] followed by previous methods BMNet [39], to ensure a fair comparison. We outperform all competing methods under the condition of without ground truth manually annotated masks. Besides, we even achieve comparable results against some SOTA methods using ground truth masks.", + "bbox": [ + 511, + 598, + 906, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "23221", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/492ea28ba0db74cb42401d5daf9d860fde244bcdd13eacab4c55a6de3678ee9f.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsInput Masksshadownon-shadowall
PSNR↑SSIM↑MAE↓PSNR↑SSIM↑MAE↓PSNR↑SSIM↑MAE↓
SRDDHANDHAN33.670.9788.9434.790.9794.8030.510.9495.67
DC-ShadowNetN/A34.000.9757.7035.530.9813.6531.530.9554.65
BMNetDHAN35.050.9816.6136.020.9823.6131.690.9564.46
SG-ShadowNetDHAN33.730.9797.5336.180.9822.9731.160.9524.23
ShadowFormerDHAN35.550.9826.1436.820.9833.5432.460.9574.28
ShadowDiffusionDHAN38.720.9874.9837.780.9853.4434.730.9703.63
Inpaint4ShadowDHAN36.730.9855.7036.700.9853.2733.270.9673.81
DeS3N/A37.910.9865.2737.450.9843.0334.110.9683.56
HomoformalerDHAN38.810.9874.2539.450.9882.8535.370.9723.33
SAM-helps-shadowN/A33.940.9797.4433.850.9813.7430.720.9524.79
oursN/A39.080.9894.3339.360.9922.5835.570.9753.11
ISTD+BMNetGT37.870.9915.6237.510.9852.4533.980.9722.97
ShadowDiffusionGT39.690.9924.9738.890.9872.2835.670.9752.72
HomoFormerGT39.490.9934.7338.750.9842.2335.350.9752.64
DC-ShadowNetN/A31.060.97612.6227.030.9616.8225.030.9267.77
DeS3N/A36.490.9896.5634.700.9723.4031.380.9583.94
BMNetFDRNet--6.1--2.9--3.5
ShadowDiffusionFDRNet40.120.9925.1536.660.9782.7434.080.9683.12
HomoFormerFDRNet38.840.9915.3134.580.9663.1732.410.9533.51
OursN/A40.360.9934.7837.890.9822.4635.000.9722.85
", + "bbox": [ + 96, + 88, + 898, + 380 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/d42c196bb59f40702c8d58c1b8c56119d275f8c719236595df0d6caf13c10f4d.jpg", + "table_caption": [ + "Table 1. The quantitative results of shadow removal using our SoftShadow and recent methods on SRD and ISTD+ datasets. The “Input Masks” column shows the different types of input masks used by these methods. “N/A” means the method does not require masks as input. “GT” means the method uses manually annotated ground truth mask as input. “DHAN” means using masks generated by DHAN [3] method. “FDRNet” means using masks generated by FDRNet [38] method. The best and the second results are boldfaced and underlined, respectively." + ], + "table_footnote": [], + "table_body": "
MethodsLRSSUIUC
PSNR↑SSIM↑MAE↓PSNR↑SSIM↑MAE↓
DC-ShadowNet20.890.90212.5524.850.8499.51
G2R-ShadowNet20.900.9019.9927.560.8587.43
BCDiff22.130.92210.6826.810.8527.96
Ours23.320.9339.7728.850.9036.48
", + "bbox": [ + 94, + 486, + 477, + 569 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/7c69bd83909777016a60cd580f8a7eee751587af7bd9f22ba982ef391413699b.jpg", + "table_caption": [ + "Table 2. The quantitative results of shadow removal using our SoftShadow and recent methods on the LRSS and UIUC datasets. The best results are boldfaced." + ], + "table_footnote": [], + "table_body": "
MethodsPSNR↑MAE↓
Inpaint4Shadow39.725.27
DeS339.287.07
HomoFormer41.115.91
Ours41.464.40
", + "bbox": [ + 187, + 638, + 387, + 712 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2 shows the generalizability of our methods. We use our pretrained model on the SRD dataset and test it on LRSS and UIUC datasets without further training. The DeS3 [17] has better results because they train their method on the LRSS training set. For comparison, we chose DC-ShadowNet as our baseline since this method does not require masks as input. Our results significantly outperform", + "bbox": [ + 89, + 794, + 483, + 902 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/2a272af29df89caa991f8d90dd30ff1f5ed388ea16b0023b6d28683e234c9efb.jpg", + "table_caption": [ + "Table 3. The quantitative results in the penumbra area. The penumbra area is calculated from our predicted soft masks. The best results are boldfaced." + ], + "table_footnote": [], + "table_body": "
Input MasksShadowDiffusionHomoFormerOurs
Ground Truth mask35.6735.3535.00
Otsu mask35.6233.47
BDRAR mask33.4132.96
FDRNet mask34.0832.41
Pretrained SAM mask28.4528.16
Mean33.4532.47N/A
Std Dev0.9120.993N/A
", + "bbox": [ + 532, + 486, + 887, + 595 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 4. The mask sensitivity evaluation on $\\mathrm{ISTD + }$ dataset. The values in the table are PSNR results. The best results are boldfaced. The results that are higher than ours are underlined.", + "bbox": [ + 511, + 606, + 906, + 648 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "DC-ShadowNet on LRSS and UIUC datasets. Additionally, we compare our methods with two methods, which are designed to have better generalizability. The results demonstrate that our method has the best PSNR/SSIM/MAE among the comparison methods. These results highlight the robustness and adaptability of our model, even on datasets it was not specifically trained on.", + "bbox": [ + 511, + 681, + 906, + 787 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Qualitative evaluation To further demonstrate the effectiveness of our methods compared to other competing methods, Figure 5 illustrates our improvements in removing soft shadows on the SRD dataset compared with methods using hard shadow masks. The soft masks typically appear as soft boundaries of large object shadows or in small shadow regions. Specifically, our samples demonstrate that we ef", + "bbox": [ + 511, + 794, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "23222", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/9a74b8ed5c21e251063c875166a536274446f667ede0b7c8eef6c41d21842dc5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 178, + 87, + 267, + 181 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4ae1f68b9b0e380c1c774c0b52b1804e539b3b51d9bd973ff96c403574c09224.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 178, + 181, + 267, + 272 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e7e2c867b2b114ead37e0962bcb75534115f31d777f39d9843563a9c77bd3cc9.jpg", + "image_caption": [ + "Input", + "(a) BMNet" + ], + "image_footnote": [], + "bbox": [ + 176, + 272, + 267, + 342 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e575ad387457e11efe562115da75df95a81a956b96fe2ac2ca61beeb8dc62cd3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 267, + 88, + 357, + 181 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/fc16a0c3965d03e50257596a09f0d09717d2423ff0ae732fa05396ae35fee6e3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 267, + 181, + 357, + 272 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/57b52f6e05ec73099823e490b0661e20bb67acc979b1c043749f19d7eb1e084e.jpg", + "image_caption": [ + "(b) ShadowDiffusion" + ], + "image_footnote": [], + "bbox": [ + 267, + 272, + 357, + 342 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/529f1391825f1f6d39718cd009e934de4e15188677d0026b4c183f8e5b3c8d6d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 359, + 88, + 450, + 181 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7e911b2b9029b15510d39a1aafa0441c0b91b026604fdc69f3874fd07b6a1918.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 359, + 181, + 450, + 272 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/056c92d0d7e004e1b2290ec47a0e81062cb41c72408d0f90b164db0075c916a3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 359, + 272, + 450, + 342 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/cf590b81a393205a5637b3393852bd9df8622fa72f9d7e0d52a7cce6be987dd2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 452, + 88, + 542, + 181 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6edbe168e397965a43031b6933d332fadb3e51e28fa62ec364754f4324b18165.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 452, + 181, + 542, + 272 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/13fd6f96465afec4915ba9e31f2dd40ce82d1143851380545da3c3b21c44ac9f.jpg", + "image_caption": [ + "(c) Inpaint4Shadow" + ], + "image_footnote": [], + "bbox": [ + 452, + 272, + 542, + 342 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1d40958a37c6a584ee4f263dc0872abd9d2a9658b07265e7960d37671e5a5cdb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 545, + 88, + 635, + 181 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/459b1a8aa28b3366ad5828cefa22127f9c9d05c7812eabd3a032dc913b7420ab.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 545, + 181, + 635, + 272 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e6a4494ad333be3b39be215ae36d0f255335aaff437f65aa1d3849a843ce0ff4.jpg", + "image_caption": [ + "(d) HomoFormer" + ], + "image_footnote": [], + "bbox": [ + 545, + 272, + 635, + 342 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ed245a7606c296b7c73c7525dc5c49e864cba5105d5fa2bed1dd095f463f7295.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 637, + 88, + 725, + 181 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/72440b530bb4b8f65de843da409944f5e9a119ee0f3909f9625632b0bfed7a5f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 637, + 181, + 725, + 272 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9e42a12a120aa18f5222cab3584254c72bd63da378c8e4265f85d036e6772998.jpg", + "image_caption": [ + "(e) Ours" + ], + "image_footnote": [], + "bbox": [ + 637, + 272, + 725, + 342 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5141c770ff3554eddeab19a1f466d9df774639bf548dd0d4e3349e56db5f2f4d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 728, + 88, + 818, + 181 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/08272ac56764c3fb2c046ccf6ed090105dc5a14edb22aab50eb704980ea51dc5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 728, + 181, + 818, + 272 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f9ff74e892c06c8891f0529b75976c0362cf9082e7b5d6afcbd09b3f92369bea.jpg", + "image_caption": [ + "GT" + ], + "image_footnote": [], + "bbox": [ + 728, + 272, + 818, + 342 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3fd1b2e5a7cd38217b90cbf52961015d3b2fe0beb570dedef2d75e338a39b83a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 102, + 434, + 176, + 492 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/09cc3d2a27aede96a00a9e0f41d50d1f344ad667e4f5754468c26307d792180d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 102, + 492, + 176, + 549 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ce55aef3b2fb98669adf97dfe3bb549b4e087e32ab4cc8d2006271596edcc4e5.jpg", + "image_caption": [ + "Input", + "Figure 6. Examples of soft shadow removal results on SRD datasets [27]. The input shadow image, the estimated results of (a) DC-ShadowNet [16], (b) DeS3 [17], and (c) Ours, as well as the ground truth image, respectively." + ], + "image_footnote": [], + "bbox": [ + 102, + 549, + 176, + 604 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d05f4dffa507181f3f2cd5feb3333f44649e4f61e78c0411246376695ec13ee5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 176, + 434, + 251, + 492 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b7a4d945da8b8f191a57f9faf9116331acd8d9572a1d7759b1bf42eaf78ccedc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 176, + 492, + 251, + 549 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4f920aa12fb62068fa8510d35f65b87148c04bffa6112e722d62043b53fd7eeb.jpg", + "image_caption": [ + "(a) DC" + ], + "image_footnote": [], + "bbox": [ + 176, + 549, + 251, + 604 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/cd27d87ca9d83edfde954fd64543008beaefded5ed76be16b0e38ee3a9744f8d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 251, + 434, + 325, + 492 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/183ba488d3db01d337b01eeafc8fe024a3cca07f2047d390b10a3f07980ab6e5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 251, + 492, + 323, + 549 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0ef123f289c40e7ed6e76f1ac9a29d768cdb8e5570cbc42bb85deeabc35ed830.jpg", + "image_caption": [ + "(b) DeS3" + ], + "image_footnote": [], + "bbox": [ + 251, + 549, + 323, + 604 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/83920cde1fa370e4591a94a34f255bf592f47c7f0783d48ee1a4fc7744d8a8ee.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 325, + 434, + 398, + 492 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/97dcd1ae4caad6488054aa445701404dae08ba68fba295bc45a8f9a63bf242f6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 325, + 492, + 398, + 549 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ada482ff5526b4cf5a0797c902111fab036fbaaad21a1a9180c9e76290fa9025.jpg", + "image_caption": [ + "(b) ours" + ], + "image_footnote": [], + "bbox": [ + 325, + 549, + 397, + 604 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/52ebee6b92b1ffe614a77e072359a56d1f3d7cdbeca1c9f8419c2d4b26daa695.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 398, + 434, + 470, + 492 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f2f4a69c00b2cf11296c640b09abeefccdcca229bcfaa23b4ec0eb3b9eb74b03.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 398, + 492, + 470, + 549 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/17ac907c08fbad2e68636dc2fda0d0b684601351bac37a7701cf22c2fd398ac8.jpg", + "image_caption": [ + "GT" + ], + "image_footnote": [], + "bbox": [ + 398, + 549, + 470, + 604 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "fectively remove both types of soft shadows while preserving the illumination transition in penumbra areas, without introducing obvious artifacts. Figure 6 shows our improvement in shadow area accuracy compared to other end-to-end methods. Our approach effectively removes all parts of the shadow shown in the first row of the example. In the second and third rows, we successfully remove shadows in regions with complex textures. Figure 7 highlights our method's generalizability on the LRSS dataset, achieving improved shadow removal performance without additional training on LRSS dataset. These results confirm our method's adaptability and precision in diverse shadowed scenarios.", + "bbox": [ + 89, + 719, + 483, + 900 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/34601ae1269a778ced2e6f4c9335434e5748d9beb04c4d84d603d7f790d0c20f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 526, + 434, + 586, + 481 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/50725bc0220d96e40799bafd318b20e16018b24d968a9e83c9f590ea6d0360cb.jpg", + "image_caption": [ + "Input", + "Figure 7. Examples of shadow removal results on the LRSS dataset [7]. The input shadow image, the estimated results of (a) DC-ShadowNet [16], (b) G2R-ShadowNet [24], (c) Ours, our predicted soft shadow masks and ground truth image, respectively." + ], + "image_footnote": [], + "bbox": [ + 527, + 481, + 586, + 529 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8a6eb088abe55880ad23cf6b0afb2e27a90a5c9c00b2fda374936c742d9dfc14.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 589, + 434, + 648, + 481 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2758aa530fceb31000780249811fc1a0c8b908e0684c2673c5456900336bc1e9.jpg", + "image_caption": [ + "(a) DC" + ], + "image_footnote": [], + "bbox": [ + 589, + 481, + 648, + 529 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6ba136616495a6b8c608e35a9f625c9d09a4694e48adff683a178722cdd2cfdb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 650, + 434, + 710, + 481 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/048caf019344bd26104fc2e1f27624bf59dbdb344305644693552a0f4013e598.jpg", + "image_caption": [ + "(b) G2R" + ], + "image_footnote": [], + "bbox": [ + 650, + 481, + 710, + 529 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/fc5d06e628be907e6ee1552773a3c2cc3a305d3f744ab1d3d77b3651fe84a79f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 710, + 434, + 769, + 481 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/fb86e7d0d2e2e753cfd15411f59ab7b6d0515d8df9c8d2236a1079190104ad31.jpg", + "image_caption": [ + "(c) Ours" + ], + "image_footnote": [], + "bbox": [ + 710, + 481, + 769, + 529 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/db08484af34986ca44063d54c1d2b59b8deb2fab4f55f2f6d0c56d0ab3bf3210.jpg", + "image_caption": [ + "Figure 5. Examples of soft shadow image removal results on the SRD dataset [27]. The input shadow image, the estimated results of (a) BMNet [39], (b) ShadowDiffusion [9], (c) Inpaint4Shadow [22], (d) Homoformer [34], and (e) Ours, as well as the ground truth (GT) image, respectively." + ], + "image_footnote": [], + "bbox": [ + 771, + 434, + 831, + 481 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1a8b5e858811df78da51bd569258bc21a48aca2dad1d9fd943e715645d0d4454.jpg", + "image_caption": [ + "Soft Masks" + ], + "image_footnote": [], + "bbox": [ + 771, + 481, + 831, + 529 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/208f2c9836e272f1594293a4842aae10c45fd0d579a75c5c18ea2eefde562290.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 831, + 434, + 893, + 481 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/fb4ff310f65fe815e4e398de3f66501b786de0cb06f521f84cdbdefb8c964159.jpg", + "image_caption": [ + "GT" + ], + "image_footnote": [], + "bbox": [ + 831, + 481, + 893, + 529 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5. Framework Analysis", + "text_level": 1, + "bbox": [ + 511, + 643, + 712, + 662 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we provide more experiments to evaluate the effectiveness of our proposed SoftShadow framework. We discuss the performance on the penumbra area, the mask sensitive of previous methods, and we discuss the effectiveness of our provided losses in ablation study.", + "bbox": [ + 511, + 672, + 906, + 747 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Penumbra effect evaluation The penumbra region typically occupies only a small portion of the entire image, so improvements within this area have a limited impact on overall image metrics. To further assess the improvements introduced by using soft masks in the penumbra region, we calculate the PSNR and MAE values specifically within this area on the SRD dataset. For a fair comparison, we define the penumbra region following the methods from AEF [5], applying ground truth soft masks $\\mathbf{s}_{gt}$ . The results are shown in Table 3. Compared with the SOTA methods like Homo", + "bbox": [ + 511, + 750, + 908, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "23223", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Former [34], DeS3 [17] and Inpaint4Shadow [22]. We improve PSNR and MAE values in the penumbra area. These results show that on images containing soft shadows, our methods has better restoration performances.", + "bbox": [ + 89, + 90, + 482, + 151 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Mask sensitivity evaluation Previous shadow removal methods that rely on external shadow masks are sensitive to the accuracy of the input masks. Their performance can degrade when lack of ground truth mask. In contrast, end-to-end shadow removal methods mitigate the impact of low-quality masks. To further illustrate the sensitivity of prior SOTA methods to input masks, we evaluate their performance using various mask inputs.", + "bbox": [ + 88, + 155, + 482, + 276 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Specifically, we evaluate ShadowDiffusion [9] and HomoFormer [34] using a range of input masks, including ground truth (GT) masks, Otsu masks [14], BDRAR [37] detected masks, FDRNet [38] detected masks and pretrained SAM [19] detected masks as inputs. The PSNR results on the $\\mathrm{ISTD + }$ [21] datasets are shown in Table 4. The Otsu mask follows the approach in MaskShadowGAN [14], using the difference between shadow-free and shadowed images to generate masks; both Otsu and GT masks include additional shadow-free information. Results indicate that our methods maintain consistent performance across all tests, whereas ShadowDiffusion and HomoFormer have performance declines when using generated masks. Although our method achieves slightly lower PSNR values than those obtained with Otsu or GT masks, this is reasonable given that the $\\mathrm{ISTD + }$ dataset contains a larger proportion of hard shadow images, where our method has less of an advantage. Our approach surpasses the rest results that rely on masks generated by shadow detectors.", + "bbox": [ + 91, + 277, + 482, + 564 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ablation Study We ablate our method, SoftShadow, by adding three losses: the shadow removal loss $\\mathcal{L}_{rem}$ , the mask reconstruction loss $\\mathcal{L}_{mask}$ , and the penumbra formation constraint loss $\\mathcal{L}_{pen}$ , and demonstrate the results on SRD and LRSS datasets. First, we present the shadow removal results using the pretrained SAM and ShadowDiffusion models. Next, we finetune the framework with the shadow removal loss $\\mathcal{L}_{rem}$ . Then, we add the mask reconstruction loss $\\mathcal{L}_{mask}$ to provide SAM with more precise mask position guidance. Finally, we add the penumbra formation constraint loss $\\mathcal{L}_{pen}$ to refine the mask boundaries for smoother transitions in the shadow regions.", + "bbox": [ + 89, + 566, + 482, + 748 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The results in Table 5 are based on training with the SRD training set and evaluating on both the SRD and LRSS datasets. As shown in Table 5, the shadow removal loss $\\mathcal{L}_{rem}$ improves performance on the SRD dataset, primarily by enhancing the shadow removal network. Adding the mask reconstruction loss $\\mathcal{L}_{mask}$ further fine-tunes SAM, producing more accurate shadow masks. This improvement of mask accuracy leads to better shadow removal results. To evaluate the impact of the penumbra formation constraint loss $\\mathcal{L}_{pen}$ , we compare results with and without $\\mathcal{L}_{pen}$ . The", + "bbox": [ + 89, + 750, + 482, + 901 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/aecb285ea9ed94e823eced7f42ed019007d89a0f614bc080b4d40585131947b9.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsLRSSSRD
PSNR↑SSIM↑MAE↓PSNR↑SSIM↑MAE↓
Pretrained Weights21.400.91012.2631.270.9634.41
Lrem21.780.91312.0735.270.9733.32
Lrem + Lmask23.080.9359.9735.440.9743.17
Lrem + Lmask + Lpen23.320.9339.7735.570.9753.11
", + "bbox": [ + 532, + 89, + 887, + 159 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. The ablation studies for the shadow removal loss $\\mathcal{L}_{rem}$ , the mask reconstruction loss $\\mathcal{L}_{mask}$ , and the penumbra formation constraint loss $\\mathcal{L}_{pen}$ in LRSS and SRD datasets. The best results are boldfaced.", + "bbox": [ + 511, + 169, + 906, + 224 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/60082c305a0592f811a8470176c04645a6ab3299647ce5d58ed5652076771e78.jpg", + "image_caption": [ + "Figure 8. Visual comparison of results from our framework with different configurations: using pretrained weights, finetuned with the loss function $\\mathcal{L}_{rem}$ , $\\mathcal{L}_{rem} + \\mathcal{L}_{mask}$ and our full configuration $\\mathcal{L}_{rem} + \\mathcal{L}_{mask} + \\mathcal{L}_{pen}$ (Ours), respectively." + ], + "image_footnote": [], + "bbox": [ + 550, + 241, + 874, + 383 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "results show that $\\mathcal{L}_{pen}$ further aids in shadow removal on the SRD dataset, refining the shadow edges and enhancing the overall shadow removal quality. In the LRSS dataset, both $\\mathcal{L}_{rem}$ and $\\mathcal{L}_{mask}$ significantly improve performance, the $\\mathcal{L}_{pen}$ improves the PSNR and MAE significantly, although it results in a slight decrease in SSIM. As shown in Figure 8, adding the constraint loss $\\mathcal{L}_{pen}$ leads to more accurate and smoother soft mask predictions. The combined effects of $\\mathcal{L}_{mask}$ and $\\mathcal{L}_{pen}$ lead to more accurate and smoother soft mask predictions. Together, these losses provide the removal framework with enhanced soft shadow details, resulting in superior performance in shadow removal.", + "bbox": [ + 511, + 481, + 906, + 662 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 678, + 633, + 694 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we introduce novel soft shadow masks designed specifically for shadow removal. To achieve this, we propose a unified framework, SoftShadow, that eliminates the need for additional shadow mask input. By leveraging a pretrained SAM with LoRA, the framework accurately predicts soft masks as intermediate results, capturing detailed and varied shadow information. Additionally, we introduce a penumbra formation constraint, inspired by the physical shadow formation model, to jointly tune SAM and the shadow removal network, refining the soft mask in the penumbra area and facilitating artifact-free restoration. Extensive experiments demonstrate that our method is superior on various occasions, proving the validity of our method.", + "bbox": [ + 511, + 704, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "23224", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Acknowledgment", + "text_level": 1, + "bbox": [ + 91, + 90, + 263, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 176, + 187, + 193 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Tianrun Chen, Lanyun Zhu, Chaotao Deng, Runlong Cao, Yan Wang, Shangzhan Zhang, Zejian Li, Lingyun Sun, Ying Zang, and Papa Mao. Sam-adapter: Adapting segment anything in underperformed scenes. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3367-3375, 2023. 3", + "[2] Zhihao Chen, Lei Zhu, Liang Wan, Song Wang, Wei Feng, and Pheng-Ann Heng. A multi-task mean teacher for semi-supervised shadow detection. In Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition, pages 5611–5620, 2020. 3", + "[3] Xiaodong Cun, Chi-Man Pun, and Cheng Shi. Towards ghost-free shadow removal via dual hierarchical aggregation network and shadow matting gan. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 10680-10687, 2020. 1, 2, 3, 5, 6", + "[4] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2", + "[5] Lan Fu, Changqing Zhou, Qing Guo, Felix Juefei-Xu, Hongkai Yu, Wei Feng, Yang Liu, and Song Wang. Autoexposure fusion for single-image shadow removal. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10571-10580, 2021. 3, 7", + "[6] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 2", + "[7] Maciej Gryka, Michael Terry, and Gabriel J Brostow. Learning to remove soft shadows. ACM Transactions on Graphics (TOG), 34(5):1-15, 2015. 5, 7", + "[8] Lanqing Guo, Siyu Huang, Ding Liu, Hao Cheng, and Bihan Wen. Shadowformer: Global context helps shadow removal. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 710-718, 2023. 1, 3, 5", + "[9] Lanqing Guo, Chong Wang, Wenhan Yang, Siyu Huang, Yufei Wang, Hanspeter Pfister, and Bihan Wen. Shadowdiffusion: When degradation prior meets diffusion model for shadow removal. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14049-14058, 2023. 1, 2, 5, 7, 8", + "[10] Lanqing Guo, Chong Wang, Wenhan Yang, Yufei Wang, and Bihan Wen. Boundary-aware divide and conquer: A diffusion-based solution for unsupervised shadow removal. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13045-13054, 2023. 5" + ], + "bbox": [ + 93, + 203, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "This work was partially supported by the National Research Foundation Singapore Competitive Research Program (award number CRP29-2022-0003)", + "[11] Ruiqi Guo, Qieyun Dai, and Derek Hoiem. Paired regions for shadow detection and removal. IEEE transactions on pattern analysis and machine intelligence, 35(12):2956-2967, 2012. 5", + "[12] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2", + "[13] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 2, 4, 5", + "[14] Xiaowei Hu, Yitong Jiang, Chi-Wing Fu, and Pheng-Ann Heng. Mask-shadowgan: Learning to remove shadows from unpaired data. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2472–2481, 2019. 8", + "[15] Leiping Jie and Hui Zhang. When sam meets shadow detection. arXiv preprint arXiv:2305.11513, 2023. 4", + "[16] Yeying Jin, Aashish Sharma, and Robby T Tan. Dcshadownet: Single-image hard and soft shadow removal using unsupervised domain-classifier guided network. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5027-5036, 2021. 2, 5, 7", + "[17] Yeying Jin, Wei Ye, Wenhan Yang, Yuan Yuan, and Robby T Tan. Des3: Adaptive attention-driven self and soft shadow removal using vit similarity. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2634-2642, 2024. 2, 5, 6, 7, 8", + "[18] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.5", + "[19] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 2, 3, 4, 5, 8", + "[20] Michael S Langer and Steven W Zucker. What is a light source? In Proceedings of IEEE Computer Society Conference on Computer Vision and Pattern Recognition, pages 172-178. IEEE, 1997. 3", + "[21] Hieu Le and Dimitris Samaras. Shadow removal via shadow image decomposition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8578-8587, 2019. 2, 5, 8", + "[22] Xiaoguang Li, Qing Guo, Rabab Abdelfattah, Di Lin, Wei Feng, Ivor Tsang, and Song Wang. Leveraging inpainting for single-image shadow removal. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13055-13064, 2023. 5, 7, 8", + "[23] Yuhao Liu, Zhanghan Ke, Ke Xu, Fang Liu, Zhenwei Wang, and Rynson WH Lau. Recasting regional lighting for shadow removal. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3810-3818, 2024. 1", + "[24] Zhihao Liu, Hui Yin, Xinyi Wu, Zhenyao Wu, Yang Mi, and Song Wang. From shadow generation to shadow removal. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4927-4936, 2021. 5, 7" + ], + "bbox": [ + 93, + 92, + 905, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "23225", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[25] Vinod Nair and Geoffrey E Hinton. Rectified linear units improve restricted boltzmann machines. In Proceedings of the 27th international conference on machine learning (ICML-10), pages 807–814, 2010. 5", + "[26] Michael Nielsen and Claus B Madsen. Segmentation of soft shadows based on a daylight-and penumbra model. In International Conference on Computer Vision/Computer Graphics Collaboration Techniques and Applications, pages 341-352. Springer, 2007. 2", + "[27] Liangqiong Qu, Jiandong Tian, Shengfeng He, Yandong Tang, and Rynson WH Lau. Deshadownet: A multi-context embedding deep network for shadow removal. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4067-4075, 2017. 2, 5, 7", + "[28] Chitwan Sahara, Jonathan Ho, William Chan, Tim Salimans, David J Fleet, and Mohammad Norouzi. Image superresolution via iterative refinement. IEEE transactions on pattern analysis and machine intelligence, 45(4):4713-4726, 2022. 2", + "[29] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 2", + "[30] Jin Wan, Hui Yin, Zhenyao Wu, Xinyi Wu, Yanting Liu, and Song Wang. Style-guided shadow removal. In European Conference on Computer Vision, pages 361-378. Springer, 2022. 1, 5", + "[31] Jifeng Wang, Xiang Li, and Jian Yang. Stacked conditional generative adversarial networks for jointly learning shadow detection and shadow removal. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1788-1797, 2018. 1, 2, 5", + "[32] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 5", + "[33] Cort J Willmott and Kenji Matsuura. Advantages of the mean absolute error (mae) over the root mean square error (rmse) in assessing average model performance. *Climate research*, 30(1):79–82, 2005. 5", + "[34] Jie Xiao, Xueyang Fu, Yurui Zhu, Dong Li, Jie Huang, Kai Zhu, and Zheng-Jun Zha. Homoformer: Homogenized transformer for image shadow removal. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25617-25626, 2024. 1, 2, 3, 5, 7, 8", + "[35] Xiaofeng Zhang, Chaochen Gu, and Shanying Zhu. Sam-helps-shadow: When segment anything model meet shadow removal. arXiv e-prints, pages arXiv-2306, 2023. 2, 5", + "[36] Quanlong Zheng, Xiaotian Qiao, Ying Cao, and Rynson WH Lau. Distraction-aware shadow detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5167-5176, 2019. 3", + "[37] Lei Zhu, Zijun Deng, Xiaowei Hu, Chi-Wing Fu, Xuemiao Xu, Jing Qin, and Pheng-Ann Heng. Bidirectional feature pyramid network with recurrent attention residual modules for shadow detection. In Proceedings of the European Conference on Computer Vision (ECCV), pages 121–136, 2018." + ], + "bbox": [ + 91, + 92, + 480, + 887 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[38] Lei Zhu, Ke Xu, Zhanghan Ke, and Rynson WH Lau. Mitigating intensity bias in shadow detection via feature decomposition and reweighting. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4702-4711, 2021. 5, 6, 8", + "[39] Yurui Zhu, Jie Huang, Xueyang Fu, Feng Zhao, Qibin Sun, and Zheng-Jun Zha. Bijective mapping network for shadow removal. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5627-5636, 2022. 1, 3, 5, 7" + ], + "bbox": [ + 516, + 92, + 903, + 229 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "23226", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/SoftShadow_ Leveraging Soft Masks for Penumbra-Aware Shadow Removal/1a57ac9e-bdab-4cc5-a82c-bea51caf54d0_model.json b/2025/SoftShadow_ Leveraging Soft Masks for Penumbra-Aware Shadow Removal/1a57ac9e-bdab-4cc5-a82c-bea51caf54d0_model.json new file mode 100644 index 0000000000000000000000000000000000000000..9b5d7f249807fa35373abcef6fbbb39e1d14c909 --- /dev/null +++ b/2025/SoftShadow_ Leveraging Soft Masks for Penumbra-Aware Shadow Removal/1a57ac9e-bdab-4cc5-a82c-bea51caf54d0_model.json @@ -0,0 +1,2695 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.131, + 0.884, + 0.152 + ], + "angle": 0, + "content": "SoftShadow: Leveraging Soft Masks for Penumbra-Aware Shadow Removal" + }, + { + "type": "text", + "bbox": [ + 0.221, + 0.179, + 0.782, + 0.287 + ], + "angle": 0, + "content": "Xinrui Wang\\(^{1}\\), Lanqing Guo\\(^{2}\\), Xiyu Wang\\(^{1}\\), Siyu Huang\\(^{3}\\), Bihan Wen\\(^{1\\dagger}\\) \n\\(^{1}\\)Nanyang Technological University, Singapore \n\\(^{2}\\)The University of Texas at Austin, USA \n\\(^{3}\\)Clemson University, USA \n{xinrui.wang, bihan.wen}@ntu.edu.sg, xiyu004@e.ntu.edu.sg, lanqing.guo@austin.utexas.edu, siyuh@clemson.edu" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.322, + 0.327, + 0.339 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.355, + 0.485, + 0.673 + ], + "angle": 0, + "content": "Recent advancements in deep learning have yielded promising results for the image shadow removal task. However, most existing methods rely on binary pre-generated shadow masks. The binary nature of such masks could potentially lead to artifacts near the boundary between shadow and non-shadow areas. In view of this, inspired by the physical model of shadow formation, we introduce novel soft shadow masks specifically designed for shadow removal. To achieve such soft masks, we propose a SoftShadow framework by leveraging the prior knowledge of pretrained SAM and integrating physical constraints. Specifically, we jointly tune the SAM and the subsequent shadow removal network using penumbra formation constraint loss, mask reconstruction loss, and shadow removal loss. This framework enables accurate predictions of penumbra (partially shaded) and umbra (fully shaded) areas while simultaneously facilitating end-to-end shadow removal. Through extensive experiments on popular datasets, we found that our SoftShadow framework, which generates soft masks, can better restore boundary artifacts, achieve state-of-the-art performance, and demonstrate superior generalizability." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.702, + 0.222, + 0.718 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.728, + 0.483, + 0.864 + ], + "angle": 0, + "content": "Shadow removal aims to restore content obscured in shadow regions and correct degraded illumination. Recently, deep learning methods have shown excellent performance in shadow removal tasks relying on large-scale training data. However, one of the main challenges in shadow removal arises in images with soft shadows, where the shadow edges are blurred rather than sharp. This blurred area, known as the penumbra area, occurs due to partial occlusion of light, creating a transition zone where illumination varies signif" + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.324, + 0.613, + 0.391 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.554, + 0.392, + 0.579, + 0.4 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image", + "bbox": [ + 0.619, + 0.325, + 0.707, + 0.391 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.634, + 0.392, + 0.696, + 0.4 + ], + "angle": 0, + "content": "Ground Truth" + }, + { + "type": "image", + "bbox": [ + 0.714, + 0.324, + 0.802, + 0.391 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.729, + 0.392, + 0.779, + 0.4 + ], + "angle": 0, + "content": "Hard Mask" + }, + { + "type": "image", + "bbox": [ + 0.521, + 0.402, + 0.611, + 0.47 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.551, + 0.471, + 0.585, + 0.479 + ], + "angle": 0, + "content": "BMNet" + }, + { + "type": "image", + "bbox": [ + 0.619, + 0.402, + 0.706, + 0.47 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.627, + 0.471, + 0.697, + 0.479 + ], + "angle": 0, + "content": "SG-ShadowNet" + }, + { + "type": "image", + "bbox": [ + 0.714, + 0.402, + 0.802, + 0.47 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.728, + 0.471, + 0.788, + 0.479 + ], + "angle": 0, + "content": "HomoFormer" + }, + { + "type": "image", + "bbox": [ + 0.809, + 0.323, + 0.899, + 0.391 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.828, + 0.392, + 0.875, + 0.4 + ], + "angle": 0, + "content": "Soft Mask" + }, + { + "type": "image", + "bbox": [ + 0.808, + 0.402, + 0.898, + 0.47 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.842, + 0.471, + 0.865, + 0.479 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.497, + 0.907, + 0.567 + ], + "angle": 0, + "content": "Figure 1. Illustration of soft shadow removal results using our proposed SoftShadow with soft shadow mask compared to the recent competing methods BMNet [39], SG-ShadowNet [30] and HomoFormer [34] using hard shadow masks. The second rows are sharpened versions of results for better visualization." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.597, + 0.907, + 0.656 + ], + "angle": 0, + "content": "icantly. Removing shadows in these regions is challenging because the gradual light-to-shadow transition often leads to boundary artifacts, compromising the quality of the shadow removal results." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.66, + 0.909, + 0.902 + ], + "angle": 0, + "content": "A few methods rely on hard shadow masks (binary masks) to indicate the shadow regions [8, 9, 34, 39]. These binary shadow masks are either manually annotated [23, 31], or predicted by off-the-shelf shadow detectors [3]. Obtaining such shadow masks can be costly and complicated, and the choice of shadow detector can significantly impact shadow removal performance. Moreover, hard masks fall short of representing the penumbra area, causing boundary artifacts in removal results, especially when the penumbra areas are highly pronounced. As shown in Figure 1, when the input image contains penumbra regions, the previous state-of-the-art (SOTA) method [34] suffers from boundary artifacts. Other approaches attempt shadow removal without explicitly extracting the shadow mask. Instead, they incorporate semantic parsing modules into shadow removal networks. For instance, some approaches involve predict" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.875, + 0.304, + 0.888 + ], + "angle": 0, + "content": "*Equal contribution for both authors" + }, + { + "type": "page_footnote", + "bbox": [ + 0.112, + 0.888, + 0.234, + 0.9 + ], + "angle": 0, + "content": "† Corresponding author" + }, + { + "type": "list", + "bbox": [ + 0.109, + 0.875, + 0.304, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23217" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.107, + 0.089, + 0.265, + 0.247 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.276, + 0.089, + 0.472, + 0.247 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.262, + 0.483, + 0.346 + ], + "angle": 0, + "content": "Figure 2. (a) Illustration of the shadow formation geometry that creates the penumbra and umbra [26] regions; the umbra area is where the light is fully occluded, the penumbra area is where light is partially occluded, the lit area is where light is not occluded. (b) Examples of soft shadow images and hard shadow images from commonly used datasets." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.373, + 0.483, + 0.447 + ], + "angle": 0, + "content": "ing degradation attention [3, 27], while others utilize domain classifiers [16, 17] to better understand and address shadow effects. However, their performance may be limited by the lack of powerful external detectors providing additional shadow location information." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.449, + 0.483, + 0.674 + ], + "angle": 0, + "content": "In view of this, we argue that simply leveraging the binary mask to represent the shadow location is not enough. Differently, we introduce a novel soft shadow mask (grayscale mask) specifically designed for shadow removal as shown in the hard mask v.s. soft mask in Figure 1. The soft masks can precisely locate the inner and outer boundaries of shadow and indicate the penumbra area with proper degradation variance. Moreover, recognizing the effectiveness of employing a powerful pretrained detector for providing semantic information, we utilize the prior knowledge from the pretrained SAM [19]. While existing methods [35] have used the off-the-shelf SAM in a naive way, merely predicting the shadow mask as input to the shadow removal network, we adapt SAM as a soft shadow mask predictor and jointly optimize it with the shadow removal network." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.675, + 0.483, + 0.901 + ], + "angle": 0, + "content": "In this paper, we first propose a unified shadow removal framework, dubbed SoftShadow, which removes shadow from images in an end-to-end manner without requiring input shadow masks as shown in Figure 3. We integrate the pretrained SAM [19] combined with Low-Rank Adaptation (LoRA) [13] to jointly adapt the SAM model and shadow removal model. The framework enables accurate predictions of soft shadow masks as intermediate results while simultaneously facilitating end-to-end shadow removal. Besides, we introduce a penumbra formation constraint to assist predict soft shadow masks. This constraint regularizes the gradient of the predicted mask in the penumbra area, resulting in more accurate predictions with detailed and spatially varied shadow masks. In the commonly used shadow removal datasets, SRD and LRSS datasets contain" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.166 + ], + "angle": 0, + "content": "soft shadow images, which makes them more suitable for illustrating the improvements of utilizing soft shadow masks. Experimental results show that SoftShadow achieves superior performance on SRD and LRSS datasets. Our main contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.172, + 0.905, + 0.261 + ], + "angle": 0, + "content": "- We propose a unified shadow removal framework that does not require additional input shadow masks. This framework enables accurate predictions of soft masks as intermediate results, which is specifically designed to capture detailed and spatially varied shadow location information." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.263, + 0.905, + 0.337 + ], + "angle": 0, + "content": "- We introduce penumbra formation constraint inspired by the physical shadow formation model to further refine the soft mask in the penumbra area. By leveraging the constraint loss and shadow removal loss, we jointly tune the SAM and the subsequent shadow removal network." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.338, + 0.905, + 0.413 + ], + "angle": 0, + "content": "- Experimental results demonstrate that SoftShadow surpasses state-of-the-art shadow removal methods on the SRD and LRSS datasets, achieving superior performance and even comparable results with previous methods that use ground truth mask inputs." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.172, + 0.905, + 0.413 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.429, + 0.655, + 0.445 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.455, + 0.684, + 0.47 + ], + "angle": 0, + "content": "2.1. Shadow Removal" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.478, + 0.907, + 0.902 + ], + "angle": 0, + "content": "The degradation of shadows varies in each image, posing a significant challenge for shadow image restoration. Recent deep learning-based approaches have achieved remarkable results in shadow removal. Some methods restore the shadow image with the guidance of shadow masks. For example, SP+M-Net [21] employs two deep networks to predict shadow matte and shadow parameters. Recently, powerful backbones such as transformers [4, 29], and diffusion models [12, 28] have been applied to the shadow removal task. HomoFormer [34] homogenizes the spatial distribution of shadow masks to uniformly recover the entire shadow image, while ShadowDiffusion [9] provides a robust generative prior for producing natural shadow-free images. Some methods aim to eliminate the dependency on shadow mask inputs. DeShadowNet [27] introduces an end-to-end shadow removal method containing a multi-branch fusion module. ST-CGAN [31] connects two GANs [6] in sequence to jointly detect and remove shadows. DC-ShadowNet [16] provides an unsupervised domain-classifier discriminator for guided shadow removal network. More recently, DeS3 [17] implemented a method capable of removing shadows that are cast on the object itself without requiring shadow mask inputs. However, due to the absence of additional shadow location information, their performance may degrade, and the number of parameters used could increase significantly. Different from these existing methods, we introduce a new concept \"soft mask\", designed explicitly for shadow removal." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23218" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.173, + 0.087, + 0.825, + 0.289 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.3, + 0.907, + 0.397 + ], + "angle": 0, + "content": "Figure 3. Illustration of the proposed SoftShadow. The left box illustrates the SoftShadow networks, where a shadow image \\(\\mathbf{y}\\) is input into SAM for detecting soft shadow masks. The shadow removal network then processes the soft mask and the shadow image to produce a shadow-free image. The right box shows the three losses we used in SoftShadow. From top to bottom, the shadow removal loss \\(\\mathcal{L}_{rem}\\) is calculated between shadow-free images and shadow-removal images. The mask reconstruction loss \\(\\mathcal{L}_{mask}\\) is calculated between predicted soft masks and ours ground truth soft masks. The penumbra formation constraint loss \\(\\mathcal{L}_{pen}\\) acts as a regularization term. It aims to regularize the gradient of predicted soft masks in the penumbra area. As shown in the soft mask intensity curve, the ideal mask intensity in the penumbra area should not be too large and the gradient orientation should be consistent." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.423, + 0.266, + 0.437 + ], + "angle": 0, + "content": "2.2. Shadow Detection" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.447, + 0.485, + 0.78 + ], + "angle": 0, + "content": "Shadow detection can be approached as a segmentation task, it often struggles with small shadows and indistinct shadow edges. Various methods have been proposed to address these challenges. BDRAR [37] introduces a bidirectional pyramidal architecture for shadow detection. DSD [36] designs a distraction-aware module to minimize false positives in shadow detection. Chen et al. [2] propose a task-specific semi-supervised learning mechanism to utilize unlabeled shadow images for detection, enhancing the robustness of the model. DHAN [3] is a well-used shadow detection model using dual hierarchical aggregation network training on synthetic shadow and shadow-free image pairs to improve detection accuracy. The SAM-Adapter [1] has been utilized to fine-tune the Segment Anything Model (SAM) [19] specifically for shadow detection tasks. While this approach improves accuracy, it demands a substantial number of model parameters. However, most existing methods detect shadows as hard masks and fail to represent the penumbra area. In contrast, our approach leverages the shadow removal loss and the Penumbra Formation Constraint loss, jointly training the shadow removal network and the shadow detection network." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.799, + 0.191, + 0.814 + ], + "angle": 0, + "content": "3. Methods" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.483, + 0.903 + ], + "angle": 0, + "content": "In this section, we first explain the motivation behind our newly introduced soft shadow mask, inspired by the physical model of shadow formation. We then propose the unified shadow removal framework SoftShadow, which leverages the powerful segmentation capabilities of SAM to pro" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.424, + 0.907, + 0.514 + ], + "angle": 0, + "content": "duce the soft shadow mask and guide the removal network to generate better shadow-free images. Finally, to further mitigate boundary artifacts, we introduce the penumbra formation constraint loss which ensures a smoother transition between the shadow region and the non-shadow region, offering better guidance to the shadow removal network." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.528, + 0.637, + 0.542 + ], + "angle": 0, + "content": "3.1. Motivation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.551, + 0.907, + 0.687 + ], + "angle": 0, + "content": "Shadow degradations in the real world exhibit considerable amount of variation and can be classified into two main categories: soft shadows and hard shadows [20]. This classification depends on the light source and the distance between the object and the surface, as shown in Figure 2(a). Soft shadows are characterized by blurry edges and gradual transitions from light to dark, creating relatively large penumbra areas, while hard shadows have sharp edges with few to none penumbra regions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.688, + 0.909, + 0.853 + ], + "angle": 0, + "content": "Previous methods [5, 8, 34, 39] commonly use hard masks to guide the shadow removal process. The hard masks effectively provide positional guidance for hard masks with sharp edges. However, when the penumbra regions are more pronounced, as shown in Figure 2 (b), the boundary between the umbra and the lit region is hard to determine, making hard masks unsuitable for representing such soft shadows. Additionally, we observed that, in general, the brightness transition within the penumbra area can vary significantly, whereas the brightness of the lit and the umbra areas are relatively uniform." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.908, + 0.901 + ], + "angle": 0, + "content": "To this end, we instead use soft (greyscale) masks to represent the position of the shadows. Formally, we denote the soft mask as s. For a given shadow-free image \\(\\mathbf{x}\\), the" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23219" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.383, + 0.106 + ], + "angle": 0, + "content": "shadow image \\( \\mathbf{y} \\) can be formulated as such:" + }, + { + "type": "equation", + "bbox": [ + 0.194, + 0.121, + 0.483, + 0.137 + ], + "angle": 0, + "content": "\\[\n\\mathbf {y} = \\mathbf {a} \\cdot \\mathbf {s} \\cdot \\mathbf {x} + (\\mathbf {1} - \\mathbf {s}) \\cdot \\mathbf {x}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.149, + 0.484, + 0.271 + ], + "angle": 0, + "content": "where \\(\\cdot\\) denotes element-wise multiplication, \\(s\\) is the soft shadow mask in which constant \\(s = 0\\) represents lit area, \\(s = 1\\) represents umbra area, and \\(s \\in [0,1]\\) represents penumbra area. \\(a \\in [0,1]\\) represents the illumination weight in the non-lit area. Based on our experiment, we quantitatively and qualitatively found that the proposed shadow degradation model could lead to fewer boundary artifacts and generally better shadow removal results." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.281, + 0.341, + 0.295 + ], + "angle": 0, + "content": "3.2. Architecture of SoftShadow" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.303, + 0.484, + 0.528 + ], + "angle": 0, + "content": "SoftShadow is a unified framework designed to remove object shadows in images in an end-to-end manner. The overall architecture is illustrated in Figure 3. SoftShadow contains a soft shadow detector and a subsequent shadow removal network. The soft shadow detector leverages the strong segmentation capabilities of SAM and further finetuning it to serve as a soft mask detector. This is because the pretrained SAM can fail to accurately identify shadows [15, 19]. Moreover, since SAM has noticeably more parameters compared to conventional shadow detection networks, we use Low-Rank Adaptation (LoRA) [13] to reduce the number of trainable parameters. In practice, we tune all the self-attention blocks of image encoders using LoRA. Since the mask decoder is lightweight, we tune all the parameters of it." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.53, + 0.484, + 0.681 + ], + "angle": 0, + "content": "SAM is mainly trained for binary predictions, unlike the soft masks we aim to generate. To leverage SAM to generate continuous soft masks, we further introduce the mask reconstruction loss \\(\\mathcal{L}_{mask}\\) and the penumbra formation constraint loss \\(\\mathcal{L}_{pen}\\) to model the characteristics of soft masks (details are described in Section 3.3). They encourage the model to detect the shadow area and produce a continuous soft mask rather than a binary one, thereby capturing the shadow position more accurately. We formulate a mask reconstruction loss \\(\\mathcal{L}_{mask}\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.197, + 0.692, + 0.483, + 0.713 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {m a s k} = \\mathbb {E} _ {n} \\left\\| \\hat {\\mathbf {s}} - \\mathbf {s} _ {g t} \\right\\| _ {F} ^ {2}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.723, + 0.484, + 0.858 + ], + "angle": 0, + "content": "where \\( n \\) is the index of shadow masks, \\( \\hat{\\mathbf{s}} \\) is the predicted soft shadow mask, and \\( \\mathbf{s}_{gt} \\) is the ground truth soft shadow mask achieved by dividing the shadow-free image \\( \\mathbf{y} \\) by the shadow image \\( \\mathbf{x} \\). Specifically, we convert \\( \\mathbf{y} \\) and \\( \\mathbf{x} \\) into YCbCr image space, using the Y channel to illustrate the brightness of these images, denoted as \\( \\mathbf{y}_Y \\) and \\( \\mathbf{x}_Y \\), respectively. We then divide the \\( \\mathbf{y}_Y \\) by \\( \\mathbf{x}_Y \\) to obtain the difference between them, apply a low-pass filter \\( f \\) to reduce noise, and use a threshold to eliminate outliers as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.192, + 0.871, + 0.483, + 0.906 + ], + "angle": 0, + "content": "\\[\n\\mathbf {s} _ {g t} = \\max \\left(t, f \\left(\\frac {\\mathbf {x} _ {Y}}{\\mathbf {y} _ {Y}}\\right)\\right), \\tag {3}\n\\]" + }, + { + "type": "image", + "bbox": [ + 0.579, + 0.088, + 0.838, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.265, + 0.907, + 0.361 + ], + "angle": 0, + "content": "Figure 4. Illustrate the concept of Penumbra Formation Constraint. The \\( w_{0} \\) and \\( w_{1} \\) are points in the inner boundary and the outer boundary of the penumbra area, respectively. In the ideal soft shadow mask, the intensity in the penumbra area decreases progressively from the shadow center to the umbra area. The gradient orientation points from the shadow center towards the non-shadow area." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.39, + 0.905, + 0.42 + ], + "angle": 0, + "content": "where \\(t\\) is the threshold used to define the outer boundary between the penumbra area and the lit area." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.421, + 0.905, + 0.466 + ], + "angle": 0, + "content": "Moreover, as illustrate in Figure 3, we incorporet the gradient of the shadow removal loss \\(\\mathcal{L}_{rem}\\) to jointly train SAM and the subsequent shadow removal network:" + }, + { + "type": "equation", + "bbox": [ + 0.629, + 0.477, + 0.905, + 0.497 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {r e m}} = \\mathbb {E} _ {n} \\| \\hat {\\mathbf {x}} - \\mathbf {x} \\| _ {F} ^ {2}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.508, + 0.906, + 0.597 + ], + "angle": 0, + "content": "where \\( n \\) indexes all images, \\( \\hat{\\mathbf{x}} \\) is the restored shadow-free image, and \\( \\mathbf{x} \\) is the ground truth shadow-free image. To this end, the intermediate soft shadow masks can be refined using shadow removal results, with the SAM adapter optimized for better location guidance in subsequent shadow removal." + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.599, + 0.851, + 0.614 + ], + "angle": 0, + "content": "The overall training objective of SoftShadow is:" + }, + { + "type": "equation", + "bbox": [ + 0.595, + 0.628, + 0.905, + 0.645 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {\\text {m a s k}} + \\lambda_ {1} \\mathcal {L} _ {\\text {p e n}} + \\lambda_ {2} \\mathcal {L} _ {\\text {r e m}}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.657, + 0.906, + 0.716 + ], + "angle": 0, + "content": "where \\(\\lambda_{1}\\) and \\(\\lambda_{2}\\) are weighting coefficients to balance the influence of each term. With the merits of joint training and physical constraints, our SoftShadow framework effectively removes shadows." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.728, + 0.805, + 0.742 + ], + "angle": 0, + "content": "3.3. Penumbra Formation Constraint" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.907, + 0.901 + ], + "angle": 0, + "content": "We introduce the Penumbra Formation Constraint to enable SAM to predict the soft masks more accurately. The objective of this constraint is to regularize SAM to learn a mask that can reflect the illuminance change of shadow images. To meet the objective, we present two assumptions for our proposed soft mask. For a given shadow image, the soft mask intensity in the penumbra region should gradually transition outward from the shadow center to the outer edge of the shadow. This means that the gradient direction in the penumbra should point from the shadow center toward" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23220" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.184 + ], + "angle": 0, + "content": "the boundary of the shadow. Additionally, since the intensity change in this area is gradual, the gradient magnitude should remain relatively small to reflect a smooth transition. Based on these assumptions, we can formulate the penumbra formation constraint. For a predicted soft mask \\( \\mathbf{s} \\), we define the penumbra area using two thresholds \\( t_1 \\) and \\( t_2 \\) as" + }, + { + "type": "equation", + "bbox": [ + 0.185, + 0.196, + 0.483, + 0.213 + ], + "angle": 0, + "content": "\\[\n\\mathcal {W} = \\{(i, j) \\mid t _ {1} \\leq \\mathbf {s} _ {i, j} \\leq t _ {2} \\}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.224, + 0.484, + 0.316 + ], + "angle": 0, + "content": "where \\(i\\) and \\(j\\) denote the spatial coordinates of pixels within the soft mask s. As shown in Figure 4, the center of the shadow area is denoted as c. c is defined as the mean of the \\(x-\\) and \\(y-\\) coordinates of the penumbra area. The unit vector representing the direction from c to a point w in the penumbra area \\(\\mathcal{W}\\) can be expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.151, + 0.328, + 0.483, + 0.358 + ], + "angle": 0, + "content": "\\[\nd (w) = \\frac {w - c}{\\| w - c \\|}, \\quad \\text {f o r} w \\in \\mathcal {W}, w \\neq c. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.369, + 0.483, + 0.401 + ], + "angle": 0, + "content": "Following the assumptions we proposed, we can define the penumbra formation constraint loss \\(\\mathcal{L}_{pen}\\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.167, + 0.412, + 0.483, + 0.43 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {p e n}} = \\mathbb {E} _ {n, w} \\left[ R (\\mathbf {d} (w) \\cdot \\nabla M (w)) \\right], \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.441, + 0.484, + 0.579 + ], + "angle": 0, + "content": "where \\( n \\) is the index of shadow masks, the desired direction of gradient is \\( \\mathbf{d}(w) \\), and \\( \\cdot \\) indicate the dot product. \\( R \\) refers to the ReLU function [25]. The ReLU function is applied to filter out gradients conflict with the desired direction. The constraint loss regularizes the magnitude of gradients in alignment with the desired direction, promoting a smooth intensity transition within the penumbra region. By penalizing quick changes in intensity, this constraint encourages the detector to generate softer shadow masks." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.593, + 0.224, + 0.609 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.618, + 0.288, + 0.634 + ], + "angle": 0, + "content": "4.1. Experimental Setups" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.64, + 0.483, + 0.837 + ], + "angle": 0, + "content": "Implementation details The shadow removal backbone in our SoftShadow framework can be any shadow removal network. We use ShadowDiffusion [9] as the backbone example for all experiments. We finetune ShadowDiffusion at a resolution of \\(256 \\times 256\\), followed by previous methods [9]. We empirically set the threshold \\(t = 0.76\\) to get the ground truth soft shadow mask \\(\\mathbf{s}_{gt}\\). We employ the ViT-H model as the backbone for SAM [19]. A set of LoRA [13] layers with a rank of 8 is added to the self-attention blocks in the image encoder of SAM. We use Adam [18] optimizer. The training batch size is 16. We set \\(\\lambda_{1} = 0.1\\) and \\(\\lambda_{2} = 1\\). For more training details, please refer to the supplementary." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Benchmark datasets We work with four benchmark datasets for the various shadow removal experiments. SRD Dataset [27] consists of 2,680 training pairs and 408 testing pairs of shadow and shadow-free images. Notably, the SRD" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.288 + ], + "angle": 0, + "content": "dataset does not provide shadow masks, the previous methods using SRD commonly using masks detected by DHAN [3] methods. LRSS Dataset [7] is a specifically designed soft shadow dataset that includes 137 images. We select 48 paired shadow and shadow-free images as our testing set. UIUC Dataset [11] contains 76 pairs of images for testing. It features a variety of shadow types [17], including soft, hard, and self shadows, which provides a diverse and challenging set of conditions for evaluating shadow removal methods. \\(\\mathrm{ISTD + }\\) Dataset [21] is adjusted ISTD dataset [31], consists of 1330 training pairs and 540 testing pairs of shadow and shadow-free images. \\(\\mathrm{ISTD + }\\) has manually annotated binary masks." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.292, + 0.905, + 0.398 + ], + "angle": 0, + "content": "Evaluation metrics Following previous works [3, 9, 34], we employ the Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index (SSIM) [32], and Mean Absolute Error (MAE) [33] as quantitative evaluation metrics. We also calculate all metrics for shadow areas, non-shadow areas, and all pixels between ground truth shadow-free images and generated removal results." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.407, + 0.811, + 0.423 + ], + "angle": 0, + "content": "4.2. Comparison with State-of-the-Art" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.43, + 0.907, + 0.596 + ], + "angle": 0, + "content": "We compare our proposed method with several state-of-the-art shadow removal methods. We include methods that do not require shadow masks as input, e.g., DCshadowNet [16], DeS3 [17], and SAM-helps-shadow [35]. And methods that require shadow masks, including DHAN [3], BMNet [39], SG-ShadowNet [30], ShadowFormer [8], Inpaint4Shadow [22], ShadowDiffusion [9] and Homoformer [34]. To demonstrate the generalizability of our method, we also compare it with zero-shot methods including BCDiff [10] and G2R-ShadowNet [24] on the LRSS [7] and UIUC [11] datasets." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.599, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Quantitative results Table 1 shows the quantitative results of SRD and \\(\\mathrm{ISTD + }\\) datasets. The SRD dataset contains many soft shadow images, which is more suitable for validating our methods. Specifically, we outperform all competing methods over all metrics, whether they require input shadow masks or not. Compared with the most recent work without requiring mask input, i.e., DeS3 [17], the PSNR is improved from 34.11 dB to 35.57 dB in whole images. When compared with methods that need input masks, we outperform the SOTA method Homoformer [34], which uses DHAN [3] mask as their input. On the \\(\\mathrm{ISTD + }\\) dataset, we significantly outperform DeS3 [17], increasing the PSNR from 31.38 dB to 35.00 dB. For methods that rely on the ground truth masks from the \\(\\mathrm{ISTD + }\\) dataset, we evaluate them using masks detected by the FDRNet [38] followed by previous methods BMNet [39], to ensure a fair comparison. We outperform all competing methods under the condition of without ground truth manually annotated masks. Besides, we even achieve comparable results against some SOTA methods using ground truth masks." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "23221" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.098, + 0.089, + 0.9, + 0.381 + ], + "angle": 0, + "content": "
MethodsInput Masksshadownon-shadowall
PSNR↑SSIM↑MAE↓PSNR↑SSIM↑MAE↓PSNR↑SSIM↑MAE↓
SRDDHANDHAN33.670.9788.9434.790.9794.8030.510.9495.67
DC-ShadowNetN/A34.000.9757.7035.530.9813.6531.530.9554.65
BMNetDHAN35.050.9816.6136.020.9823.6131.690.9564.46
SG-ShadowNetDHAN33.730.9797.5336.180.9822.9731.160.9524.23
ShadowFormerDHAN35.550.9826.1436.820.9833.5432.460.9574.28
ShadowDiffusionDHAN38.720.9874.9837.780.9853.4434.730.9703.63
Inpaint4ShadowDHAN36.730.9855.7036.700.9853.2733.270.9673.81
DeS3N/A37.910.9865.2737.450.9843.0334.110.9683.56
HomoformalerDHAN38.810.9874.2539.450.9882.8535.370.9723.33
SAM-helps-shadowN/A33.940.9797.4433.850.9813.7430.720.9524.79
oursN/A39.080.9894.3339.360.9922.5835.570.9753.11
ISTD+BMNetGT37.870.9915.6237.510.9852.4533.980.9722.97
ShadowDiffusionGT39.690.9924.9738.890.9872.2835.670.9752.72
HomoFormerGT39.490.9934.7338.750.9842.2335.350.9752.64
DC-ShadowNetN/A31.060.97612.6227.030.9616.8225.030.9267.77
DeS3N/A36.490.9896.5634.700.9723.4031.380.9583.94
BMNetFDRNet--6.1--2.9--3.5
ShadowDiffusionFDRNet40.120.9925.1536.660.9782.7434.080.9683.12
HomoFormerFDRNet38.840.9915.3134.580.9663.1732.410.9533.51
OursN/A40.360.9934.7837.890.9822.4635.000.9722.85
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.393, + 0.907, + 0.462 + ], + "angle": 0, + "content": "Table 1. The quantitative results of shadow removal using our SoftShadow and recent methods on SRD and ISTD+ datasets. The “Input Masks” column shows the different types of input masks used by these methods. “N/A” means the method does not require masks as input. “GT” means the method uses manually annotated ground truth mask as input. “DHAN” means using masks generated by DHAN [3] method. “FDRNet” means using masks generated by FDRNet [38] method. The best and the second results are boldfaced and underlined, respectively." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.487, + 0.478, + 0.57 + ], + "angle": 0, + "content": "
MethodsLRSSUIUC
PSNR↑SSIM↑MAE↓PSNR↑SSIM↑MAE↓
DC-ShadowNet20.890.90212.5524.850.8499.51
G2R-ShadowNet20.900.9019.9927.560.8587.43
BCDiff22.130.92210.6826.810.8527.96
Ours23.320.9339.7728.850.9036.48
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.581, + 0.483, + 0.622 + ], + "angle": 0, + "content": "Table 2. The quantitative results of shadow removal using our SoftShadow and recent methods on the LRSS and UIUC datasets. The best results are boldfaced." + }, + { + "type": "table", + "bbox": [ + 0.188, + 0.64, + 0.388, + 0.713 + ], + "angle": 0, + "content": "
MethodsPSNR↑MAE↓
Inpaint4Shadow39.725.27
DeS339.287.07
HomoFormer41.115.91
Ours41.464.40
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.724, + 0.483, + 0.765 + ], + "angle": 0, + "content": "Table 3. The quantitative results in the penumbra area. The penumbra area is calculated from our predicted soft masks. The best results are boldfaced." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Table 2 shows the generalizability of our methods. We use our pretrained model on the SRD dataset and test it on LRSS and UIUC datasets without further training. The DeS3 [17] has better results because they train their method on the LRSS training set. For comparison, we chose DC-ShadowNet as our baseline since this method does not require masks as input. Our results significantly outperform" + }, + { + "type": "table", + "bbox": [ + 0.534, + 0.487, + 0.888, + 0.596 + ], + "angle": 0, + "content": "
Input MasksShadowDiffusionHomoFormerOurs
Ground Truth mask35.6735.3535.00
Otsu mask35.6233.47
BDRAR mask33.4132.96
FDRNet mask34.0832.41
Pretrained SAM mask28.4528.16
Mean33.4532.47N/A
Std Dev0.9120.993N/A
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.607, + 0.907, + 0.65 + ], + "angle": 0, + "content": "Table 4. The mask sensitivity evaluation on \\(\\mathrm{ISTD + }\\) dataset. The values in the table are PSNR results. The best results are boldfaced. The results that are higher than ours are underlined." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.682, + 0.907, + 0.789 + ], + "angle": 0, + "content": "DC-ShadowNet on LRSS and UIUC datasets. Additionally, we compare our methods with two methods, which are designed to have better generalizability. The results demonstrate that our method has the best PSNR/SSIM/MAE among the comparison methods. These results highlight the robustness and adaptability of our model, even on datasets it was not specifically trained on." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Qualitative evaluation To further demonstrate the effectiveness of our methods compared to other competing methods, Figure 5 illustrates our improvements in removing soft shadows on the SRD dataset compared with methods using hard shadow masks. The soft masks typically appear as soft boundaries of large object shadows or in small shadow regions. Specifically, our samples demonstrate that we ef" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.946, + 0.521, + 0.957 + ], + "angle": 0, + "content": "23222" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.179, + 0.088, + 0.268, + 0.182 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.179, + 0.182, + 0.268, + 0.273 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.178, + 0.273, + 0.268, + 0.343 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.268, + 0.089, + 0.359, + 0.182 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.268, + 0.182, + 0.359, + 0.273 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.268, + 0.273, + 0.359, + 0.343 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.209, + 0.343, + 0.236, + 0.353 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image_caption", + "bbox": [ + 0.268, + 0.343, + 0.337, + 0.353 + ], + "angle": 0, + "content": "(a) BMNet" + }, + { + "type": "image", + "bbox": [ + 0.36, + 0.089, + 0.451, + 0.182 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.36, + 0.182, + 0.451, + 0.273 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.36, + 0.273, + 0.451, + 0.343 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.358, + 0.343, + 0.451, + 0.353 + ], + "angle": 0, + "content": "(b) ShadowDiffusion" + }, + { + "type": "image", + "bbox": [ + 0.453, + 0.089, + 0.544, + 0.182 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.453, + 0.182, + 0.544, + 0.273 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.453, + 0.273, + 0.543, + 0.343 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.454, + 0.343, + 0.543, + 0.353 + ], + "angle": 0, + "content": "(c) Inpaint4Shadow" + }, + { + "type": "image", + "bbox": [ + 0.546, + 0.089, + 0.636, + 0.182 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.546, + 0.182, + 0.636, + 0.273 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.546, + 0.273, + 0.636, + 0.343 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.553, + 0.343, + 0.632, + 0.353 + ], + "angle": 0, + "content": "(d) HomoFormer" + }, + { + "type": "image", + "bbox": [ + 0.638, + 0.089, + 0.727, + 0.182 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.638, + 0.182, + 0.727, + 0.273 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.638, + 0.273, + 0.727, + 0.343 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.663, + 0.343, + 0.701, + 0.353 + ], + "angle": 0, + "content": "(e) Ours" + }, + { + "type": "image", + "bbox": [ + 0.729, + 0.089, + 0.82, + 0.182 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.729, + 0.182, + 0.82, + 0.273 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.729, + 0.273, + 0.82, + 0.343 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.766, + 0.343, + 0.782, + 0.353 + ], + "angle": 0, + "content": "GT" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.369, + 0.908, + 0.412 + ], + "angle": 0, + "content": "Figure 5. Examples of soft shadow image removal results on the SRD dataset [27]. The input shadow image, the estimated results of (a) BMNet [39], (b) ShadowDiffusion [9], (c) Inpaint4Shadow [22], (d) Homoformer [34], and (e) Ours, as well as the ground truth (GT) image, respectively." + }, + { + "type": "image", + "bbox": [ + 0.104, + 0.435, + 0.177, + 0.493 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.104, + 0.493, + 0.177, + 0.55 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.104, + 0.55, + 0.177, + 0.606 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.132, + 0.607, + 0.157, + 0.616 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image", + "bbox": [ + 0.178, + 0.435, + 0.252, + 0.493 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.178, + 0.493, + 0.252, + 0.55 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.178, + 0.55, + 0.252, + 0.606 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.199, + 0.607, + 0.231, + 0.616 + ], + "angle": 0, + "content": "(a) DC" + }, + { + "type": "image", + "bbox": [ + 0.252, + 0.435, + 0.326, + 0.493 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.252, + 0.493, + 0.325, + 0.55 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.252, + 0.55, + 0.325, + 0.606 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.267, + 0.607, + 0.308, + 0.616 + ], + "angle": 0, + "content": "(b) DeS3" + }, + { + "type": "image", + "bbox": [ + 0.326, + 0.435, + 0.399, + 0.493 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.326, + 0.493, + 0.399, + 0.55 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.326, + 0.55, + 0.398, + 0.606 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.345, + 0.607, + 0.38, + 0.616 + ], + "angle": 0, + "content": "(b) ours" + }, + { + "type": "image", + "bbox": [ + 0.399, + 0.435, + 0.472, + 0.493 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.399, + 0.493, + 0.472, + 0.55 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.399, + 0.55, + 0.472, + 0.606 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.428, + 0.607, + 0.444, + 0.616 + ], + "angle": 0, + "content": "GT" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.633, + 0.486, + 0.689 + ], + "angle": 0, + "content": "Figure 6. Examples of soft shadow removal results on SRD datasets [27]. The input shadow image, the estimated results of (a) DC-ShadowNet [16], (b) DeS3 [17], and (c) Ours, as well as the ground truth image, respectively." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.485, + 0.901 + ], + "angle": 0, + "content": "fectively remove both types of soft shadows while preserving the illumination transition in penumbra areas, without introducing obvious artifacts. Figure 6 shows our improvement in shadow area accuracy compared to other end-to-end methods. Our approach effectively removes all parts of the shadow shown in the first row of the example. In the second and third rows, we successfully remove shadows in regions with complex textures. Figure 7 highlights our method's generalizability on the LRSS dataset, achieving improved shadow removal performance without additional training on LRSS dataset. These results confirm our method's adaptability and precision in diverse shadowed scenarios." + }, + { + "type": "image", + "bbox": [ + 0.527, + 0.435, + 0.588, + 0.482 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.528, + 0.482, + 0.588, + 0.53 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.544, + 0.531, + 0.573, + 0.542 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image", + "bbox": [ + 0.59, + 0.435, + 0.65, + 0.482 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.59, + 0.482, + 0.65, + 0.53 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.601, + 0.531, + 0.636, + 0.542 + ], + "angle": 0, + "content": "(a) DC" + }, + { + "type": "image", + "bbox": [ + 0.651, + 0.435, + 0.711, + 0.482 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.651, + 0.482, + 0.711, + 0.53 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.66, + 0.531, + 0.703, + 0.541 + ], + "angle": 0, + "content": "(b) G2R" + }, + { + "type": "image", + "bbox": [ + 0.712, + 0.435, + 0.771, + 0.482 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.712, + 0.482, + 0.771, + 0.53 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.718, + 0.531, + 0.761, + 0.541 + ], + "angle": 0, + "content": "(c) Ours" + }, + { + "type": "image", + "bbox": [ + 0.772, + 0.435, + 0.832, + 0.482 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.772, + 0.482, + 0.832, + 0.53 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.775, + 0.531, + 0.833, + 0.541 + ], + "angle": 0, + "content": "Soft Masks" + }, + { + "type": "image", + "bbox": [ + 0.833, + 0.435, + 0.894, + 0.482 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.833, + 0.482, + 0.894, + 0.53 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.855, + 0.531, + 0.872, + 0.541 + ], + "angle": 0, + "content": "GT" + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.559, + 0.907, + 0.616 + ], + "angle": 0, + "content": "Figure 7. Examples of shadow removal results on the LRSS dataset [7]. The input shadow image, the estimated results of (a) DC-ShadowNet [16], (b) G2R-ShadowNet [24], (c) Ours, our predicted soft shadow masks and ground truth image, respectively." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.645, + 0.714, + 0.663 + ], + "angle": 0, + "content": "5. Framework Analysis" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.673, + 0.907, + 0.748 + ], + "angle": 0, + "content": "In this section, we provide more experiments to evaluate the effectiveness of our proposed SoftShadow framework. We discuss the performance on the penumbra area, the mask sensitive of previous methods, and we discuss the effectiveness of our provided losses in ablation study." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Penumbra effect evaluation The penumbra region typically occupies only a small portion of the entire image, so improvements within this area have a limited impact on overall image metrics. To further assess the improvements introduced by using soft masks in the penumbra region, we calculate the PSNR and MAE values specifically within this area on the SRD dataset. For a fair comparison, we define the penumbra region following the methods from AEF [5], applying ground truth soft masks \\(\\mathbf{s}_{gt}\\). The results are shown in Table 3. Compared with the SOTA methods like Homo" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23223" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.091, + 0.483, + 0.152 + ], + "angle": 0, + "content": "Former [34], DeS3 [17] and Inpaint4Shadow [22]. We improve PSNR and MAE values in the penumbra area. These results show that on images containing soft shadows, our methods has better restoration performances." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.156, + 0.483, + 0.277 + ], + "angle": 0, + "content": "Mask sensitivity evaluation Previous shadow removal methods that rely on external shadow masks are sensitive to the accuracy of the input masks. Their performance can degrade when lack of ground truth mask. In contrast, end-to-end shadow removal methods mitigate the impact of low-quality masks. To further illustrate the sensitivity of prior SOTA methods to input masks, we evaluate their performance using various mask inputs." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.278, + 0.483, + 0.565 + ], + "angle": 0, + "content": "Specifically, we evaluate ShadowDiffusion [9] and HomoFormer [34] using a range of input masks, including ground truth (GT) masks, Otsu masks [14], BDRAR [37] detected masks, FDRNet [38] detected masks and pretrained SAM [19] detected masks as inputs. The PSNR results on the \\(\\mathrm{ISTD + }\\) [21] datasets are shown in Table 4. The Otsu mask follows the approach in MaskShadowGAN [14], using the difference between shadow-free and shadowed images to generate masks; both Otsu and GT masks include additional shadow-free information. Results indicate that our methods maintain consistent performance across all tests, whereas ShadowDiffusion and HomoFormer have performance declines when using generated masks. Although our method achieves slightly lower PSNR values than those obtained with Otsu or GT masks, this is reasonable given that the \\(\\mathrm{ISTD + }\\) dataset contains a larger proportion of hard shadow images, where our method has less of an advantage. Our approach surpasses the rest results that rely on masks generated by shadow detectors." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.568, + 0.483, + 0.749 + ], + "angle": 0, + "content": "Ablation Study We ablate our method, SoftShadow, by adding three losses: the shadow removal loss \\(\\mathcal{L}_{rem}\\), the mask reconstruction loss \\(\\mathcal{L}_{mask}\\), and the penumbra formation constraint loss \\(\\mathcal{L}_{pen}\\), and demonstrate the results on SRD and LRSS datasets. First, we present the shadow removal results using the pretrained SAM and ShadowDiffusion models. Next, we finetune the framework with the shadow removal loss \\(\\mathcal{L}_{rem}\\). Then, we add the mask reconstruction loss \\(\\mathcal{L}_{mask}\\) to provide SAM with more precise mask position guidance. Finally, we add the penumbra formation constraint loss \\(\\mathcal{L}_{pen}\\) to refine the mask boundaries for smoother transitions in the shadow regions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.902 + ], + "angle": 0, + "content": "The results in Table 5 are based on training with the SRD training set and evaluating on both the SRD and LRSS datasets. As shown in Table 5, the shadow removal loss \\(\\mathcal{L}_{rem}\\) improves performance on the SRD dataset, primarily by enhancing the shadow removal network. Adding the mask reconstruction loss \\(\\mathcal{L}_{mask}\\) further fine-tunes SAM, producing more accurate shadow masks. This improvement of mask accuracy leads to better shadow removal results. To evaluate the impact of the penumbra formation constraint loss \\(\\mathcal{L}_{pen}\\), we compare results with and without \\(\\mathcal{L}_{pen}\\). The" + }, + { + "type": "table", + "bbox": [ + 0.534, + 0.09, + 0.888, + 0.16 + ], + "angle": 0, + "content": "
MethodsLRSSSRD
PSNR↑SSIM↑MAE↓PSNR↑SSIM↑MAE↓
Pretrained Weights21.400.91012.2631.270.9634.41
Lrem21.780.91312.0735.270.9733.32
Lrem + Lmask23.080.9359.9735.440.9743.17
Lrem + Lmask + Lpen23.320.9339.7735.570.9753.11
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.17, + 0.907, + 0.226 + ], + "angle": 0, + "content": "Table 5. The ablation studies for the shadow removal loss \\(\\mathcal{L}_{rem}\\), the mask reconstruction loss \\(\\mathcal{L}_{mask}\\), and the penumbra formation constraint loss \\(\\mathcal{L}_{pen}\\) in LRSS and SRD datasets. The best results are boldfaced." + }, + { + "type": "image", + "bbox": [ + 0.552, + 0.242, + 0.875, + 0.385 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.399, + 0.907, + 0.456 + ], + "angle": 0, + "content": "Figure 8. Visual comparison of results from our framework with different configurations: using pretrained weights, finetuned with the loss function \\(\\mathcal{L}_{rem}\\), \\(\\mathcal{L}_{rem} + \\mathcal{L}_{mask}\\) and our full configuration \\(\\mathcal{L}_{rem} + \\mathcal{L}_{mask} + \\mathcal{L}_{pen}\\) (Ours), respectively." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.482, + 0.907, + 0.664 + ], + "angle": 0, + "content": "results show that \\(\\mathcal{L}_{pen}\\) further aids in shadow removal on the SRD dataset, refining the shadow edges and enhancing the overall shadow removal quality. In the LRSS dataset, both \\(\\mathcal{L}_{rem}\\) and \\(\\mathcal{L}_{mask}\\) significantly improve performance, the \\(\\mathcal{L}_{pen}\\) improves the PSNR and MAE significantly, although it results in a slight decrease in SSIM. As shown in Figure 8, adding the constraint loss \\(\\mathcal{L}_{pen}\\) leads to more accurate and smoother soft mask predictions. The combined effects of \\(\\mathcal{L}_{mask}\\) and \\(\\mathcal{L}_{pen}\\) lead to more accurate and smoother soft mask predictions. Together, these losses provide the removal framework with enhanced soft shadow details, resulting in superior performance in shadow removal." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.679, + 0.634, + 0.695 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In this paper, we introduce novel soft shadow masks designed specifically for shadow removal. To achieve this, we propose a unified framework, SoftShadow, that eliminates the need for additional shadow mask input. By leveraging a pretrained SAM with LoRA, the framework accurately predicts soft masks as intermediate results, capturing detailed and varied shadow information. Additionally, we introduce a penumbra formation constraint, inspired by the physical shadow formation model, to jointly tune SAM and the shadow removal network, refining the soft mask in the penumbra area and facilitating artifact-free restoration. Extensive experiments demonstrate that our method is superior on various occasions, proving the validity of our method." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23224" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.264, + 0.108 + ], + "angle": 0, + "content": "7. Acknowledgment" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.116, + 0.486, + 0.162 + ], + "angle": 0, + "content": "This work was partially supported by the National Research Foundation Singapore Competitive Research Program (award number CRP29-2022-0003)" + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.178, + 0.188, + 0.194 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.204, + 0.484, + 0.287 + ], + "angle": 0, + "content": "[1] Tianrun Chen, Lanyun Zhu, Chaotao Deng, Runlong Cao, Yan Wang, Shangzhan Zhang, Zejian Li, Lingyun Sun, Ying Zang, and Papa Mao. Sam-adapter: Adapting segment anything in underperformed scenes. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3367-3375, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.289, + 0.484, + 0.358 + ], + "angle": 0, + "content": "[2] Zhihao Chen, Lei Zhu, Liang Wan, Song Wang, Wei Feng, and Pheng-Ann Heng. A multi-task mean teacher for semi-supervised shadow detection. In Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition, pages 5611–5620, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.361, + 0.484, + 0.429 + ], + "angle": 0, + "content": "[3] Xiaodong Cun, Chi-Man Pun, and Cheng Shi. Towards ghost-free shadow removal via dual hierarchical aggregation network and shadow matting gan. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 10680-10687, 2020. 1, 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.432, + 0.484, + 0.514 + ], + "angle": 0, + "content": "[4] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.517, + 0.484, + 0.585 + ], + "angle": 0, + "content": "[5] Lan Fu, Changqing Zhou, Qing Guo, Felix Juefei-Xu, Hongkai Yu, Wei Feng, Yang Liu, and Song Wang. Autoexposure fusion for single-image shadow removal. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10571-10580, 2021. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.589, + 0.484, + 0.643 + ], + "angle": 0, + "content": "[6] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.646, + 0.484, + 0.686 + ], + "angle": 0, + "content": "[7] Maciej Gryka, Michael Terry, and Gabriel J Brostow. Learning to remove soft shadows. ACM Transactions on Graphics (TOG), 34(5):1-15, 2015. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.689, + 0.484, + 0.745 + ], + "angle": 0, + "content": "[8] Lanqing Guo, Siyu Huang, Ding Liu, Hao Cheng, and Bihan Wen. Shadowformer: Global context helps shadow removal. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 710-718, 2023. 1, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.748, + 0.484, + 0.829 + ], + "angle": 0, + "content": "[9] Lanqing Guo, Chong Wang, Wenhan Yang, Siyu Huang, Yufei Wang, Hanspeter Pfister, and Bihan Wen. Shadowdiffusion: When degradation prior meets diffusion model for shadow removal. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14049-14058, 2023. 1, 2, 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.832, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[10] Lanqing Guo, Chong Wang, Wenhan Yang, Yufei Wang, and Bihan Wen. Boundary-aware divide and conquer: A diffusion-based solution for unsupervised shadow removal. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13045-13054, 2023. 5" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.204, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.146 + ], + "angle": 0, + "content": "[11] Ruiqi Guo, Qieyun Dai, and Derek Hoiem. Paired regions for shadow detection and removal. IEEE transactions on pattern analysis and machine intelligence, 35(12):2956-2967, 2012. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.149, + 0.906, + 0.19 + ], + "angle": 0, + "content": "[12] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.192, + 0.906, + 0.245 + ], + "angle": 0, + "content": "[13] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.247, + 0.906, + 0.314 + ], + "angle": 0, + "content": "[14] Xiaowei Hu, Yitong Jiang, Chi-Wing Fu, and Pheng-Ann Heng. Mask-shadowgan: Learning to remove shadows from unpaired data. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2472–2481, 2019. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.317, + 0.905, + 0.343 + ], + "angle": 0, + "content": "[15] Leiping Jie and Hui Zhang. When sam meets shadow detection. arXiv preprint arXiv:2305.11513, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.345, + 0.906, + 0.413 + ], + "angle": 0, + "content": "[16] Yeying Jin, Aashish Sharma, and Robby T Tan. Dcshadownet: Single-image hard and soft shadow removal using unsupervised domain-classifier guided network. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5027-5036, 2021. 2, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.415, + 0.906, + 0.482 + ], + "angle": 0, + "content": "[17] Yeying Jin, Wei Ye, Wenhan Yang, Yuan Yuan, and Robby T Tan. Des3: Adaptive attention-driven self and soft shadow removal using vit similarity. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2634-2642, 2024. 2, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.484, + 0.906, + 0.523 + ], + "angle": 0, + "content": "[18] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.526, + 0.906, + 0.607 + ], + "angle": 0, + "content": "[19] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 2, 3, 4, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.609, + 0.906, + 0.663 + ], + "angle": 0, + "content": "[20] Michael S Langer and Steven W Zucker. What is a light source? In Proceedings of IEEE Computer Society Conference on Computer Vision and Pattern Recognition, pages 172-178. IEEE, 1997. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.665, + 0.906, + 0.718 + ], + "angle": 0, + "content": "[21] Hieu Le and Dimitris Samaras. Shadow removal via shadow image decomposition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8578-8587, 2019. 2, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.721, + 0.906, + 0.789 + ], + "angle": 0, + "content": "[22] Xiaoguang Li, Qing Guo, Rabab Abdelfattah, Di Lin, Wei Feng, Ivor Tsang, and Song Wang. Leveraging inpainting for single-image shadow removal. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13055-13064, 2023. 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.791, + 0.906, + 0.845 + ], + "angle": 0, + "content": "[23] Yuhao Liu, Zhanghan Ke, Ke Xu, Fang Liu, Zhenwei Wang, and Rynson WH Lau. Recasting regional lighting for shadow removal. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3810-3818, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.847, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[24] Zhihao Liu, Hui Yin, Xinyi Wu, Zhenyao Wu, Yang Mi, and Song Wang. From shadow generation to shadow removal. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4927-4936, 2021. 5, 7" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.093, + 0.906, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23225" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[25] Vinod Nair and Geoffrey E Hinton. Rectified linear units improve restricted boltzmann machines. In Proceedings of the 27th international conference on machine learning (ICML-10), pages 807–814, 2010. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.218 + ], + "angle": 0, + "content": "[26] Michael Nielsen and Claus B Madsen. Segmentation of soft shadows based on a daylight-and penumbra model. In International Conference on Computer Vision/Computer Graphics Collaboration Techniques and Applications, pages 341-352. Springer, 2007. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.219, + 0.482, + 0.287 + ], + "angle": 0, + "content": "[27] Liangqiong Qu, Jiandong Tian, Shengfeng He, Yandong Tang, and Rynson WH Lau. Deshadownet: A multi-context embedding deep network for shadow removal. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4067-4075, 2017. 2, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.288, + 0.482, + 0.355 + ], + "angle": 0, + "content": "[28] Chitwan Sahara, Jonathan Ho, William Chan, Tim Salimans, David J Fleet, and Mohammad Norouzi. Image superresolution via iterative refinement. IEEE transactions on pattern analysis and machine intelligence, 45(4):4713-4726, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.357, + 0.482, + 0.412 + ], + "angle": 0, + "content": "[29] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.414, + 0.482, + 0.467 + ], + "angle": 0, + "content": "[30] Jin Wan, Hui Yin, Zhenyao Wu, Xinyi Wu, Yanting Liu, and Song Wang. Style-guided shadow removal. In European Conference on Computer Vision, pages 361-378. Springer, 2022. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.469, + 0.482, + 0.537 + ], + "angle": 0, + "content": "[31] Jifeng Wang, Xiang Li, and Jian Yang. Stacked conditional generative adversarial networks for jointly learning shadow detection and shadow removal. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1788-1797, 2018. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.539, + 0.482, + 0.593 + ], + "angle": 0, + "content": "[32] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.595, + 0.482, + 0.649 + ], + "angle": 0, + "content": "[33] Cort J Willmott and Kenji Matsuura. Advantages of the mean absolute error (mae) over the root mean square error (rmse) in assessing average model performance. *Climate research*, 30(1):79–82, 2005. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.651, + 0.482, + 0.718 + ], + "angle": 0, + "content": "[34] Jie Xiao, Xueyang Fu, Yurui Zhu, Dong Li, Jie Huang, Kai Zhu, and Zheng-Jun Zha. Homoformer: Homogenized transformer for image shadow removal. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25617-25626, 2024. 1, 2, 3, 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.72, + 0.482, + 0.761 + ], + "angle": 0, + "content": "[35] Xiaofeng Zhang, Chaochen Gu, and Shanying Zhu. Sam-helps-shadow: When segment anything model meet shadow removal. arXiv e-prints, pages arXiv-2306, 2023. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.763, + 0.482, + 0.817 + ], + "angle": 0, + "content": "[36] Quanlong Zheng, Xiaotian Qiao, Ying Cao, and Rynson WH Lau. Distraction-aware shadow detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5167-5176, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.818, + 0.482, + 0.888 + ], + "angle": 0, + "content": "[37] Lei Zhu, Zijun Deng, Xiaowei Hu, Chi-Wing Fu, Xuemiao Xu, Jing Qin, and Pheng-Ann Heng. Bidirectional feature pyramid network with recurrent attention residual modules for shadow detection. In Proceedings of the European Conference on Computer Vision (ECCV), pages 121–136, 2018." + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.888 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.161 + ], + "angle": 0, + "content": "[38] Lei Zhu, Ke Xu, Zhanghan Ke, and Rynson WH Lau. Mitigating intensity bias in shadow detection via feature decomposition and reweighting. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4702-4711, 2021. 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.164, + 0.905, + 0.231 + ], + "angle": 0, + "content": "[39] Yurui Zhu, Jie Huang, Xueyang Fu, Feng Zhao, Qibin Sun, and Zheng-Jun Zha. Bijective mapping network for shadow removal. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5627-5636, 2022. 1, 3, 5, 7" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.231 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23226" + } + ] +] \ No newline at end of file diff --git a/2025/SoftShadow_ Leveraging Soft Masks for Penumbra-Aware Shadow Removal/1a57ac9e-bdab-4cc5-a82c-bea51caf54d0_origin.pdf b/2025/SoftShadow_ Leveraging Soft Masks for Penumbra-Aware Shadow Removal/1a57ac9e-bdab-4cc5-a82c-bea51caf54d0_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..5551df95cac2d160311066869b5e06bffcd218aa --- /dev/null +++ b/2025/SoftShadow_ Leveraging Soft Masks for Penumbra-Aware Shadow Removal/1a57ac9e-bdab-4cc5-a82c-bea51caf54d0_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:07f93580773b43b5f13e43b3e7b2409dbf348e3fc911fcc7a1e5216e5c6068a6 +size 5615011 diff --git a/2025/SoftShadow_ Leveraging Soft Masks for Penumbra-Aware Shadow Removal/full.md b/2025/SoftShadow_ Leveraging Soft Masks for Penumbra-Aware Shadow Removal/full.md new file mode 100644 index 0000000000000000000000000000000000000000..77dc28055a6a1accd2832f4e5090dbf29c20893b --- /dev/null +++ b/2025/SoftShadow_ Leveraging Soft Masks for Penumbra-Aware Shadow Removal/full.md @@ -0,0 +1,392 @@ +# SoftShadow: Leveraging Soft Masks for Penumbra-Aware Shadow Removal + +Xinrui Wang $^{1}$ , Lanqing Guo $^{2}$ , Xiyu Wang $^{1}$ , Siyu Huang $^{3}$ , Bihan Wen $^{1\dagger}$ $^{1}$ Nanyang Technological University, Singapore + $^{2}$ The University of Texas at Austin, USA + $^{3}$ Clemson University, USA +{xinrui.wang, bihan.wen}@ntu.edu.sg, xiyu004@e.ntu.edu.sg, lanqing.guo@austin.utexas.edu, siyuh@clemson.edu + +# Abstract + +Recent advancements in deep learning have yielded promising results for the image shadow removal task. However, most existing methods rely on binary pre-generated shadow masks. The binary nature of such masks could potentially lead to artifacts near the boundary between shadow and non-shadow areas. In view of this, inspired by the physical model of shadow formation, we introduce novel soft shadow masks specifically designed for shadow removal. To achieve such soft masks, we propose a SoftShadow framework by leveraging the prior knowledge of pretrained SAM and integrating physical constraints. Specifically, we jointly tune the SAM and the subsequent shadow removal network using penumbra formation constraint loss, mask reconstruction loss, and shadow removal loss. This framework enables accurate predictions of penumbra (partially shaded) and umbra (fully shaded) areas while simultaneously facilitating end-to-end shadow removal. Through extensive experiments on popular datasets, we found that our SoftShadow framework, which generates soft masks, can better restore boundary artifacts, achieve state-of-the-art performance, and demonstrate superior generalizability. + +# 1. Introduction + +Shadow removal aims to restore content obscured in shadow regions and correct degraded illumination. Recently, deep learning methods have shown excellent performance in shadow removal tasks relying on large-scale training data. However, one of the main challenges in shadow removal arises in images with soft shadows, where the shadow edges are blurred rather than sharp. This blurred area, known as the penumbra area, occurs due to partial occlusion of light, creating a transition zone where illumination varies signif + +![](images/ae3c4a2e397052d05a3208cf1a1029d799c251ea051cade9b83adf59a35af015.jpg) +Input + +![](images/67e9c45e2022f38ef0bb891db4b09294360a58ec0f3a780ee648ce2b3e0cca15.jpg) +Ground Truth + +![](images/6e1e11d3db043007d7f6f8955ec7469e228e8d42d91d556ac2895721a6062941.jpg) +Hard Mask + +![](images/19c30de583a65975a35143c95c28edfa7f8f99ba3b237a54341be51825713016.jpg) +BMNet + +![](images/1b1236cd360fc0784e1a3091948197621d5f59d9324f67d0db23dbe8b69e72e3.jpg) +SG-ShadowNet + +![](images/5f13ef3994af938eb7618caabf0a0d3c3dedee560971aba38f9c1777587373ef.jpg) +HomoFormer +Figure 1. Illustration of soft shadow removal results using our proposed SoftShadow with soft shadow mask compared to the recent competing methods BMNet [39], SG-ShadowNet [30] and HomoFormer [34] using hard shadow masks. The second rows are sharpened versions of results for better visualization. + +![](images/7da857832245867069b99057960d91d50bab59dc1b082f5f780e9c58dd8e5292.jpg) +Soft Mask + +![](images/d1e43dda1e4a1f971d21d3c2c2eff173642e016567d5a3bc2ee2656992228380.jpg) +Ours + +icantly. Removing shadows in these regions is challenging because the gradual light-to-shadow transition often leads to boundary artifacts, compromising the quality of the shadow removal results. + +A few methods rely on hard shadow masks (binary masks) to indicate the shadow regions [8, 9, 34, 39]. These binary shadow masks are either manually annotated [23, 31], or predicted by off-the-shelf shadow detectors [3]. Obtaining such shadow masks can be costly and complicated, and the choice of shadow detector can significantly impact shadow removal performance. Moreover, hard masks fall short of representing the penumbra area, causing boundary artifacts in removal results, especially when the penumbra areas are highly pronounced. As shown in Figure 1, when the input image contains penumbra regions, the previous state-of-the-art (SOTA) method [34] suffers from boundary artifacts. Other approaches attempt shadow removal without explicitly extracting the shadow mask. Instead, they incorporate semantic parsing modules into shadow removal networks. For instance, some approaches involve predict + +![](images/c01c5cf9d5db6c178fb4b2f1ddad8409cea2696767574445619bc2746d04510e.jpg) +Figure 2. (a) Illustration of the shadow formation geometry that creates the penumbra and umbra [26] regions; the umbra area is where the light is fully occluded, the penumbra area is where light is partially occluded, the lit area is where light is not occluded. (b) Examples of soft shadow images and hard shadow images from commonly used datasets. + +![](images/4ca7a5d78bf1d8ac1fa537ee83766e6540d9684b8b023c6dda47621b57d99b73.jpg) + +ing degradation attention [3, 27], while others utilize domain classifiers [16, 17] to better understand and address shadow effects. However, their performance may be limited by the lack of powerful external detectors providing additional shadow location information. + +In view of this, we argue that simply leveraging the binary mask to represent the shadow location is not enough. Differently, we introduce a novel soft shadow mask (grayscale mask) specifically designed for shadow removal as shown in the hard mask v.s. soft mask in Figure 1. The soft masks can precisely locate the inner and outer boundaries of shadow and indicate the penumbra area with proper degradation variance. Moreover, recognizing the effectiveness of employing a powerful pretrained detector for providing semantic information, we utilize the prior knowledge from the pretrained SAM [19]. While existing methods [35] have used the off-the-shelf SAM in a naive way, merely predicting the shadow mask as input to the shadow removal network, we adapt SAM as a soft shadow mask predictor and jointly optimize it with the shadow removal network. + +In this paper, we first propose a unified shadow removal framework, dubbed SoftShadow, which removes shadow from images in an end-to-end manner without requiring input shadow masks as shown in Figure 3. We integrate the pretrained SAM [19] combined with Low-Rank Adaptation (LoRA) [13] to jointly adapt the SAM model and shadow removal model. The framework enables accurate predictions of soft shadow masks as intermediate results while simultaneously facilitating end-to-end shadow removal. Besides, we introduce a penumbra formation constraint to assist predict soft shadow masks. This constraint regularizes the gradient of the predicted mask in the penumbra area, resulting in more accurate predictions with detailed and spatially varied shadow masks. In the commonly used shadow removal datasets, SRD and LRSS datasets contain + +soft shadow images, which makes them more suitable for illustrating the improvements of utilizing soft shadow masks. Experimental results show that SoftShadow achieves superior performance on SRD and LRSS datasets. Our main contributions are summarized as follows: + +- We propose a unified shadow removal framework that does not require additional input shadow masks. This framework enables accurate predictions of soft masks as intermediate results, which is specifically designed to capture detailed and spatially varied shadow location information. +- We introduce penumbra formation constraint inspired by the physical shadow formation model to further refine the soft mask in the penumbra area. By leveraging the constraint loss and shadow removal loss, we jointly tune the SAM and the subsequent shadow removal network. +- Experimental results demonstrate that SoftShadow surpasses state-of-the-art shadow removal methods on the SRD and LRSS datasets, achieving superior performance and even comparable results with previous methods that use ground truth mask inputs. + +# 2. Related Work + +# 2.1. Shadow Removal + +The degradation of shadows varies in each image, posing a significant challenge for shadow image restoration. Recent deep learning-based approaches have achieved remarkable results in shadow removal. Some methods restore the shadow image with the guidance of shadow masks. For example, SP+M-Net [21] employs two deep networks to predict shadow matte and shadow parameters. Recently, powerful backbones such as transformers [4, 29], and diffusion models [12, 28] have been applied to the shadow removal task. HomoFormer [34] homogenizes the spatial distribution of shadow masks to uniformly recover the entire shadow image, while ShadowDiffusion [9] provides a robust generative prior for producing natural shadow-free images. Some methods aim to eliminate the dependency on shadow mask inputs. DeShadowNet [27] introduces an end-to-end shadow removal method containing a multi-branch fusion module. ST-CGAN [31] connects two GANs [6] in sequence to jointly detect and remove shadows. DC-ShadowNet [16] provides an unsupervised domain-classifier discriminator for guided shadow removal network. More recently, DeS3 [17] implemented a method capable of removing shadows that are cast on the object itself without requiring shadow mask inputs. However, due to the absence of additional shadow location information, their performance may degrade, and the number of parameters used could increase significantly. Different from these existing methods, we introduce a new concept "soft mask", designed explicitly for shadow removal. + +![](images/a7ca1b72b53a7e66f7e79e36318a5bf2df0b9a26991544554e71ec1eb761e1d7.jpg) +Figure 3. Illustration of the proposed SoftShadow. The left box illustrates the SoftShadow networks, where a shadow image $\mathbf{y}$ is input into SAM for detecting soft shadow masks. The shadow removal network then processes the soft mask and the shadow image to produce a shadow-free image. The right box shows the three losses we used in SoftShadow. From top to bottom, the shadow removal loss $\mathcal{L}_{rem}$ is calculated between shadow-free images and shadow-removal images. The mask reconstruction loss $\mathcal{L}_{mask}$ is calculated between predicted soft masks and ours ground truth soft masks. The penumbra formation constraint loss $\mathcal{L}_{pen}$ acts as a regularization term. It aims to regularize the gradient of predicted soft masks in the penumbra area. As shown in the soft mask intensity curve, the ideal mask intensity in the penumbra area should not be too large and the gradient orientation should be consistent. + +# 2.2. Shadow Detection + +Shadow detection can be approached as a segmentation task, it often struggles with small shadows and indistinct shadow edges. Various methods have been proposed to address these challenges. BDRAR [37] introduces a bidirectional pyramidal architecture for shadow detection. DSD [36] designs a distraction-aware module to minimize false positives in shadow detection. Chen et al. [2] propose a task-specific semi-supervised learning mechanism to utilize unlabeled shadow images for detection, enhancing the robustness of the model. DHAN [3] is a well-used shadow detection model using dual hierarchical aggregation network training on synthetic shadow and shadow-free image pairs to improve detection accuracy. The SAM-Adapter [1] has been utilized to fine-tune the Segment Anything Model (SAM) [19] specifically for shadow detection tasks. While this approach improves accuracy, it demands a substantial number of model parameters. However, most existing methods detect shadows as hard masks and fail to represent the penumbra area. In contrast, our approach leverages the shadow removal loss and the Penumbra Formation Constraint loss, jointly training the shadow removal network and the shadow detection network. + +# 3. Methods + +In this section, we first explain the motivation behind our newly introduced soft shadow mask, inspired by the physical model of shadow formation. We then propose the unified shadow removal framework SoftShadow, which leverages the powerful segmentation capabilities of SAM to pro + +duce the soft shadow mask and guide the removal network to generate better shadow-free images. Finally, to further mitigate boundary artifacts, we introduce the penumbra formation constraint loss which ensures a smoother transition between the shadow region and the non-shadow region, offering better guidance to the shadow removal network. + +# 3.1. Motivation + +Shadow degradations in the real world exhibit considerable amount of variation and can be classified into two main categories: soft shadows and hard shadows [20]. This classification depends on the light source and the distance between the object and the surface, as shown in Figure 2(a). Soft shadows are characterized by blurry edges and gradual transitions from light to dark, creating relatively large penumbra areas, while hard shadows have sharp edges with few to none penumbra regions. + +Previous methods [5, 8, 34, 39] commonly use hard masks to guide the shadow removal process. The hard masks effectively provide positional guidance for hard masks with sharp edges. However, when the penumbra regions are more pronounced, as shown in Figure 2 (b), the boundary between the umbra and the lit region is hard to determine, making hard masks unsuitable for representing such soft shadows. Additionally, we observed that, in general, the brightness transition within the penumbra area can vary significantly, whereas the brightness of the lit and the umbra areas are relatively uniform. + +To this end, we instead use soft (greyscale) masks to represent the position of the shadows. Formally, we denote the soft mask as s. For a given shadow-free image $\mathbf{x}$ , the + +shadow image $\mathbf{y}$ can be formulated as such: + +$$ +\mathbf {y} = \mathbf {a} \cdot \mathbf {s} \cdot \mathbf {x} + (\mathbf {1} - \mathbf {s}) \cdot \mathbf {x}, \tag {1} +$$ + +where $\cdot$ denotes element-wise multiplication, $s$ is the soft shadow mask in which constant $s = 0$ represents lit area, $s = 1$ represents umbra area, and $s \in [0,1]$ represents penumbra area. $a \in [0,1]$ represents the illumination weight in the non-lit area. Based on our experiment, we quantitatively and qualitatively found that the proposed shadow degradation model could lead to fewer boundary artifacts and generally better shadow removal results. + +# 3.2. Architecture of SoftShadow + +SoftShadow is a unified framework designed to remove object shadows in images in an end-to-end manner. The overall architecture is illustrated in Figure 3. SoftShadow contains a soft shadow detector and a subsequent shadow removal network. The soft shadow detector leverages the strong segmentation capabilities of SAM and further finetuning it to serve as a soft mask detector. This is because the pretrained SAM can fail to accurately identify shadows [15, 19]. Moreover, since SAM has noticeably more parameters compared to conventional shadow detection networks, we use Low-Rank Adaptation (LoRA) [13] to reduce the number of trainable parameters. In practice, we tune all the self-attention blocks of image encoders using LoRA. Since the mask decoder is lightweight, we tune all the parameters of it. + +SAM is mainly trained for binary predictions, unlike the soft masks we aim to generate. To leverage SAM to generate continuous soft masks, we further introduce the mask reconstruction loss $\mathcal{L}_{mask}$ and the penumbra formation constraint loss $\mathcal{L}_{pen}$ to model the characteristics of soft masks (details are described in Section 3.3). They encourage the model to detect the shadow area and produce a continuous soft mask rather than a binary one, thereby capturing the shadow position more accurately. We formulate a mask reconstruction loss $\mathcal{L}_{mask}$ as follows: + +$$ +\mathcal {L} _ {m a s k} = \mathbb {E} _ {n} \left\| \hat {\mathbf {s}} - \mathbf {s} _ {g t} \right\| _ {F} ^ {2}, \tag {2} +$$ + +where $n$ is the index of shadow masks, $\hat{\mathbf{s}}$ is the predicted soft shadow mask, and $\mathbf{s}_{gt}$ is the ground truth soft shadow mask achieved by dividing the shadow-free image $\mathbf{y}$ by the shadow image $\mathbf{x}$ . Specifically, we convert $\mathbf{y}$ and $\mathbf{x}$ into YCbCr image space, using the Y channel to illustrate the brightness of these images, denoted as $\mathbf{y}_Y$ and $\mathbf{x}_Y$ , respectively. We then divide the $\mathbf{y}_Y$ by $\mathbf{x}_Y$ to obtain the difference between them, apply a low-pass filter $f$ to reduce noise, and use a threshold to eliminate outliers as follows: + +$$ +\mathbf {s} _ {g t} = \max \left(t, f \left(\frac {\mathbf {x} _ {Y}}{\mathbf {y} _ {Y}}\right)\right), \tag {3} +$$ + +![](images/3495dc65292d8cfc40319ed1daacfd1a6ad2258dfba97d292179c530b6891aae.jpg) +Figure 4. Illustrate the concept of Penumbra Formation Constraint. The $w_{0}$ and $w_{1}$ are points in the inner boundary and the outer boundary of the penumbra area, respectively. In the ideal soft shadow mask, the intensity in the penumbra area decreases progressively from the shadow center to the umbra area. The gradient orientation points from the shadow center towards the non-shadow area. + +where $t$ is the threshold used to define the outer boundary between the penumbra area and the lit area. + +Moreover, as illustrate in Figure 3, we incorporet the gradient of the shadow removal loss $\mathcal{L}_{rem}$ to jointly train SAM and the subsequent shadow removal network: + +$$ +\mathcal {L} _ {\text {r e m}} = \mathbb {E} _ {n} \| \hat {\mathbf {x}} - \mathbf {x} \| _ {F} ^ {2}, \tag {4} +$$ + +where $n$ indexes all images, $\hat{\mathbf{x}}$ is the restored shadow-free image, and $\mathbf{x}$ is the ground truth shadow-free image. To this end, the intermediate soft shadow masks can be refined using shadow removal results, with the SAM adapter optimized for better location guidance in subsequent shadow removal. + +The overall training objective of SoftShadow is: + +$$ +\mathcal {L} = \mathcal {L} _ {\text {m a s k}} + \lambda_ {1} \mathcal {L} _ {\text {p e n}} + \lambda_ {2} \mathcal {L} _ {\text {r e m}}, \tag {5} +$$ + +where $\lambda_{1}$ and $\lambda_{2}$ are weighting coefficients to balance the influence of each term. With the merits of joint training and physical constraints, our SoftShadow framework effectively removes shadows. + +# 3.3. Penumbra Formation Constraint + +We introduce the Penumbra Formation Constraint to enable SAM to predict the soft masks more accurately. The objective of this constraint is to regularize SAM to learn a mask that can reflect the illuminance change of shadow images. To meet the objective, we present two assumptions for our proposed soft mask. For a given shadow image, the soft mask intensity in the penumbra region should gradually transition outward from the shadow center to the outer edge of the shadow. This means that the gradient direction in the penumbra should point from the shadow center toward + +the boundary of the shadow. Additionally, since the intensity change in this area is gradual, the gradient magnitude should remain relatively small to reflect a smooth transition. Based on these assumptions, we can formulate the penumbra formation constraint. For a predicted soft mask $\mathbf{s}$ , we define the penumbra area using two thresholds $t_1$ and $t_2$ as + +$$ +\mathcal {W} = \{(i, j) \mid t _ {1} \leq \mathbf {s} _ {i, j} \leq t _ {2} \}, \tag {6} +$$ + +where $i$ and $j$ denote the spatial coordinates of pixels within the soft mask s. As shown in Figure 4, the center of the shadow area is denoted as c. c is defined as the mean of the $x-$ and $y-$ coordinates of the penumbra area. The unit vector representing the direction from c to a point w in the penumbra area $\mathcal{W}$ can be expressed as: + +$$ +d (w) = \frac {w - c}{\| w - c \|}, \quad \text {f o r} w \in \mathcal {W}, w \neq c. \tag {7} +$$ + +Following the assumptions we proposed, we can define the penumbra formation constraint loss $\mathcal{L}_{pen}$ as: + +$$ +\mathcal {L} _ {\text {p e n}} = \mathbb {E} _ {n, w} \left[ R (\mathbf {d} (w) \cdot \nabla M (w)) \right], \tag {8} +$$ + +where $n$ is the index of shadow masks, the desired direction of gradient is $\mathbf{d}(w)$ , and $\cdot$ indicate the dot product. $R$ refers to the ReLU function [25]. The ReLU function is applied to filter out gradients conflict with the desired direction. The constraint loss regularizes the magnitude of gradients in alignment with the desired direction, promoting a smooth intensity transition within the penumbra region. By penalizing quick changes in intensity, this constraint encourages the detector to generate softer shadow masks. + +# 4. Experiments + +# 4.1. Experimental Setups + +Implementation details The shadow removal backbone in our SoftShadow framework can be any shadow removal network. We use ShadowDiffusion [9] as the backbone example for all experiments. We finetune ShadowDiffusion at a resolution of $256 \times 256$ , followed by previous methods [9]. We empirically set the threshold $t = 0.76$ to get the ground truth soft shadow mask $\mathbf{s}_{gt}$ . We employ the ViT-H model as the backbone for SAM [19]. A set of LoRA [13] layers with a rank of 8 is added to the self-attention blocks in the image encoder of SAM. We use Adam [18] optimizer. The training batch size is 16. We set $\lambda_{1} = 0.1$ and $\lambda_{2} = 1$ . For more training details, please refer to the supplementary. + +Benchmark datasets We work with four benchmark datasets for the various shadow removal experiments. SRD Dataset [27] consists of 2,680 training pairs and 408 testing pairs of shadow and shadow-free images. Notably, the SRD + +dataset does not provide shadow masks, the previous methods using SRD commonly using masks detected by DHAN [3] methods. LRSS Dataset [7] is a specifically designed soft shadow dataset that includes 137 images. We select 48 paired shadow and shadow-free images as our testing set. UIUC Dataset [11] contains 76 pairs of images for testing. It features a variety of shadow types [17], including soft, hard, and self shadows, which provides a diverse and challenging set of conditions for evaluating shadow removal methods. $\mathrm{ISTD + }$ Dataset [21] is adjusted ISTD dataset [31], consists of 1330 training pairs and 540 testing pairs of shadow and shadow-free images. $\mathrm{ISTD + }$ has manually annotated binary masks. + +Evaluation metrics Following previous works [3, 9, 34], we employ the Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index (SSIM) [32], and Mean Absolute Error (MAE) [33] as quantitative evaluation metrics. We also calculate all metrics for shadow areas, non-shadow areas, and all pixels between ground truth shadow-free images and generated removal results. + +# 4.2. Comparison with State-of-the-Art + +We compare our proposed method with several state-of-the-art shadow removal methods. We include methods that do not require shadow masks as input, e.g., DCshadowNet [16], DeS3 [17], and SAM-helps-shadow [35]. And methods that require shadow masks, including DHAN [3], BMNet [39], SG-ShadowNet [30], ShadowFormer [8], Inpaint4Shadow [22], ShadowDiffusion [9] and Homoformer [34]. To demonstrate the generalizability of our method, we also compare it with zero-shot methods including BCDiff [10] and G2R-ShadowNet [24] on the LRSS [7] and UIUC [11] datasets. + +Quantitative results Table 1 shows the quantitative results of SRD and $\mathrm{ISTD + }$ datasets. The SRD dataset contains many soft shadow images, which is more suitable for validating our methods. Specifically, we outperform all competing methods over all metrics, whether they require input shadow masks or not. Compared with the most recent work without requiring mask input, i.e., DeS3 [17], the PSNR is improved from 34.11 dB to 35.57 dB in whole images. When compared with methods that need input masks, we outperform the SOTA method Homoformer [34], which uses DHAN [3] mask as their input. On the $\mathrm{ISTD + }$ dataset, we significantly outperform DeS3 [17], increasing the PSNR from 31.38 dB to 35.00 dB. For methods that rely on the ground truth masks from the $\mathrm{ISTD + }$ dataset, we evaluate them using masks detected by the FDRNet [38] followed by previous methods BMNet [39], to ensure a fair comparison. We outperform all competing methods under the condition of without ground truth manually annotated masks. Besides, we even achieve comparable results against some SOTA methods using ground truth masks. + +
MethodsInput Masksshadownon-shadowall
PSNR↑SSIM↑MAE↓PSNR↑SSIM↑MAE↓PSNR↑SSIM↑MAE↓
SRDDHANDHAN33.670.9788.9434.790.9794.8030.510.9495.67
DC-ShadowNetN/A34.000.9757.7035.530.9813.6531.530.9554.65
BMNetDHAN35.050.9816.6136.020.9823.6131.690.9564.46
SG-ShadowNetDHAN33.730.9797.5336.180.9822.9731.160.9524.23
ShadowFormerDHAN35.550.9826.1436.820.9833.5432.460.9574.28
ShadowDiffusionDHAN38.720.9874.9837.780.9853.4434.730.9703.63
Inpaint4ShadowDHAN36.730.9855.7036.700.9853.2733.270.9673.81
DeS3N/A37.910.9865.2737.450.9843.0334.110.9683.56
HomoformalerDHAN38.810.9874.2539.450.9882.8535.370.9723.33
SAM-helps-shadowN/A33.940.9797.4433.850.9813.7430.720.9524.79
oursN/A39.080.9894.3339.360.9922.5835.570.9753.11
ISTD+BMNetGT37.870.9915.6237.510.9852.4533.980.9722.97
ShadowDiffusionGT39.690.9924.9738.890.9872.2835.670.9752.72
HomoFormerGT39.490.9934.7338.750.9842.2335.350.9752.64
DC-ShadowNetN/A31.060.97612.6227.030.9616.8225.030.9267.77
DeS3N/A36.490.9896.5634.700.9723.4031.380.9583.94
BMNetFDRNet--6.1--2.9--3.5
ShadowDiffusionFDRNet40.120.9925.1536.660.9782.7434.080.9683.12
HomoFormerFDRNet38.840.9915.3134.580.9663.1732.410.9533.51
OursN/A40.360.9934.7837.890.9822.4635.000.9722.85
+ +Table 1. The quantitative results of shadow removal using our SoftShadow and recent methods on SRD and ISTD+ datasets. The “Input Masks” column shows the different types of input masks used by these methods. “N/A” means the method does not require masks as input. “GT” means the method uses manually annotated ground truth mask as input. “DHAN” means using masks generated by DHAN [3] method. “FDRNet” means using masks generated by FDRNet [38] method. The best and the second results are boldfaced and underlined, respectively. + +
MethodsLRSSUIUC
PSNR↑SSIM↑MAE↓PSNR↑SSIM↑MAE↓
DC-ShadowNet20.890.90212.5524.850.8499.51
G2R-ShadowNet20.900.9019.9927.560.8587.43
BCDiff22.130.92210.6826.810.8527.96
Ours23.320.9339.7728.850.9036.48
+ +Table 2. The quantitative results of shadow removal using our SoftShadow and recent methods on the LRSS and UIUC datasets. The best results are boldfaced. + +
MethodsPSNR↑MAE↓
Inpaint4Shadow39.725.27
DeS339.287.07
HomoFormer41.115.91
Ours41.464.40
+ +Table 2 shows the generalizability of our methods. We use our pretrained model on the SRD dataset and test it on LRSS and UIUC datasets without further training. The DeS3 [17] has better results because they train their method on the LRSS training set. For comparison, we chose DC-ShadowNet as our baseline since this method does not require masks as input. Our results significantly outperform + +Table 3. The quantitative results in the penumbra area. The penumbra area is calculated from our predicted soft masks. The best results are boldfaced. + +
Input MasksShadowDiffusionHomoFormerOurs
Ground Truth mask35.6735.3535.00
Otsu mask35.6233.47
BDRAR mask33.4132.96
FDRNet mask34.0832.41
Pretrained SAM mask28.4528.16
Mean33.4532.47N/A
Std Dev0.9120.993N/A
+ +Table 4. The mask sensitivity evaluation on $\mathrm{ISTD + }$ dataset. The values in the table are PSNR results. The best results are boldfaced. The results that are higher than ours are underlined. + +DC-ShadowNet on LRSS and UIUC datasets. Additionally, we compare our methods with two methods, which are designed to have better generalizability. The results demonstrate that our method has the best PSNR/SSIM/MAE among the comparison methods. These results highlight the robustness and adaptability of our model, even on datasets it was not specifically trained on. + +Qualitative evaluation To further demonstrate the effectiveness of our methods compared to other competing methods, Figure 5 illustrates our improvements in removing soft shadows on the SRD dataset compared with methods using hard shadow masks. The soft masks typically appear as soft boundaries of large object shadows or in small shadow regions. Specifically, our samples demonstrate that we ef + +![](images/9a74b8ed5c21e251063c875166a536274446f667ede0b7c8eef6c41d21842dc5.jpg) + +![](images/4ae1f68b9b0e380c1c774c0b52b1804e539b3b51d9bd973ff96c403574c09224.jpg) + +![](images/e7e2c867b2b114ead37e0962bcb75534115f31d777f39d9843563a9c77bd3cc9.jpg) +Input +(a) BMNet + +![](images/e575ad387457e11efe562115da75df95a81a956b96fe2ac2ca61beeb8dc62cd3.jpg) + +![](images/fc16a0c3965d03e50257596a09f0d09717d2423ff0ae732fa05396ae35fee6e3.jpg) + +![](images/57b52f6e05ec73099823e490b0661e20bb67acc979b1c043749f19d7eb1e084e.jpg) +(b) ShadowDiffusion + +![](images/529f1391825f1f6d39718cd009e934de4e15188677d0026b4c183f8e5b3c8d6d.jpg) + +![](images/7e911b2b9029b15510d39a1aafa0441c0b91b026604fdc69f3874fd07b6a1918.jpg) + +![](images/056c92d0d7e004e1b2290ec47a0e81062cb41c72408d0f90b164db0075c916a3.jpg) + +![](images/cf590b81a393205a5637b3393852bd9df8622fa72f9d7e0d52a7cce6be987dd2.jpg) + +![](images/6edbe168e397965a43031b6933d332fadb3e51e28fa62ec364754f4324b18165.jpg) + +![](images/13fd6f96465afec4915ba9e31f2dd40ce82d1143851380545da3c3b21c44ac9f.jpg) +(c) Inpaint4Shadow + +![](images/1d40958a37c6a584ee4f263dc0872abd9d2a9658b07265e7960d37671e5a5cdb.jpg) + +![](images/459b1a8aa28b3366ad5828cefa22127f9c9d05c7812eabd3a032dc913b7420ab.jpg) + +![](images/e6a4494ad333be3b39be215ae36d0f255335aaff437f65aa1d3849a843ce0ff4.jpg) +(d) HomoFormer + +![](images/ed245a7606c296b7c73c7525dc5c49e864cba5105d5fa2bed1dd095f463f7295.jpg) + +![](images/72440b530bb4b8f65de843da409944f5e9a119ee0f3909f9625632b0bfed7a5f.jpg) + +![](images/9e42a12a120aa18f5222cab3584254c72bd63da378c8e4265f85d036e6772998.jpg) +(e) Ours + +![](images/5141c770ff3554eddeab19a1f466d9df774639bf548dd0d4e3349e56db5f2f4d.jpg) + +![](images/08272ac56764c3fb2c046ccf6ed090105dc5a14edb22aab50eb704980ea51dc5.jpg) + +![](images/f9ff74e892c06c8891f0529b75976c0362cf9082e7b5d6afcbd09b3f92369bea.jpg) +GT + +![](images/3fd1b2e5a7cd38217b90cbf52961015d3b2fe0beb570dedef2d75e338a39b83a.jpg) + +![](images/09cc3d2a27aede96a00a9e0f41d50d1f344ad667e4f5754468c26307d792180d.jpg) + +![](images/ce55aef3b2fb98669adf97dfe3bb549b4e087e32ab4cc8d2006271596edcc4e5.jpg) +Input +Figure 6. Examples of soft shadow removal results on SRD datasets [27]. The input shadow image, the estimated results of (a) DC-ShadowNet [16], (b) DeS3 [17], and (c) Ours, as well as the ground truth image, respectively. + +![](images/d05f4dffa507181f3f2cd5feb3333f44649e4f61e78c0411246376695ec13ee5.jpg) + +![](images/b7a4d945da8b8f191a57f9faf9116331acd8d9572a1d7759b1bf42eaf78ccedc.jpg) + +![](images/4f920aa12fb62068fa8510d35f65b87148c04bffa6112e722d62043b53fd7eeb.jpg) +(a) DC + +![](images/cd27d87ca9d83edfde954fd64543008beaefded5ed76be16b0e38ee3a9744f8d.jpg) + +![](images/183ba488d3db01d337b01eeafc8fe024a3cca07f2047d390b10a3f07980ab6e5.jpg) + +![](images/0ef123f289c40e7ed6e76f1ac9a29d768cdb8e5570cbc42bb85deeabc35ed830.jpg) +(b) DeS3 + +![](images/83920cde1fa370e4591a94a34f255bf592f47c7f0783d48ee1a4fc7744d8a8ee.jpg) + +![](images/97dcd1ae4caad6488054aa445701404dae08ba68fba295bc45a8f9a63bf242f6.jpg) + +![](images/ada482ff5526b4cf5a0797c902111fab036fbaaad21a1a9180c9e76290fa9025.jpg) +(b) ours + +![](images/52ebee6b92b1ffe614a77e072359a56d1f3d7cdbeca1c9f8419c2d4b26daa695.jpg) + +![](images/f2f4a69c00b2cf11296c640b09abeefccdcca229bcfaa23b4ec0eb3b9eb74b03.jpg) + +![](images/17ac907c08fbad2e68636dc2fda0d0b684601351bac37a7701cf22c2fd398ac8.jpg) +GT + +fectively remove both types of soft shadows while preserving the illumination transition in penumbra areas, without introducing obvious artifacts. Figure 6 shows our improvement in shadow area accuracy compared to other end-to-end methods. Our approach effectively removes all parts of the shadow shown in the first row of the example. In the second and third rows, we successfully remove shadows in regions with complex textures. Figure 7 highlights our method's generalizability on the LRSS dataset, achieving improved shadow removal performance without additional training on LRSS dataset. These results confirm our method's adaptability and precision in diverse shadowed scenarios. + +![](images/34601ae1269a778ced2e6f4c9335434e5748d9beb04c4d84d603d7f790d0c20f.jpg) + +![](images/50725bc0220d96e40799bafd318b20e16018b24d968a9e83c9f590ea6d0360cb.jpg) +Input +Figure 7. Examples of shadow removal results on the LRSS dataset [7]. The input shadow image, the estimated results of (a) DC-ShadowNet [16], (b) G2R-ShadowNet [24], (c) Ours, our predicted soft shadow masks and ground truth image, respectively. + +![](images/8a6eb088abe55880ad23cf6b0afb2e27a90a5c9c00b2fda374936c742d9dfc14.jpg) + +![](images/2758aa530fceb31000780249811fc1a0c8b908e0684c2673c5456900336bc1e9.jpg) +(a) DC + +![](images/6ba136616495a6b8c608e35a9f625c9d09a4694e48adff683a178722cdd2cfdb.jpg) + +![](images/048caf019344bd26104fc2e1f27624bf59dbdb344305644693552a0f4013e598.jpg) +(b) G2R + +![](images/fc5d06e628be907e6ee1552773a3c2cc3a305d3f744ab1d3d77b3651fe84a79f.jpg) + +![](images/fb86e7d0d2e2e753cfd15411f59ab7b6d0515d8df9c8d2236a1079190104ad31.jpg) +(c) Ours + +![](images/db08484af34986ca44063d54c1d2b59b8deb2fab4f55f2f6d0c56d0ab3bf3210.jpg) +Figure 5. Examples of soft shadow image removal results on the SRD dataset [27]. The input shadow image, the estimated results of (a) BMNet [39], (b) ShadowDiffusion [9], (c) Inpaint4Shadow [22], (d) Homoformer [34], and (e) Ours, as well as the ground truth (GT) image, respectively. + +![](images/1a8b5e858811df78da51bd569258bc21a48aca2dad1d9fd943e715645d0d4454.jpg) +Soft Masks + +![](images/208f2c9836e272f1594293a4842aae10c45fd0d579a75c5c18ea2eefde562290.jpg) + +![](images/fb4ff310f65fe815e4e398de3f66501b786de0cb06f521f84cdbdefb8c964159.jpg) +GT + +# 5. Framework Analysis + +In this section, we provide more experiments to evaluate the effectiveness of our proposed SoftShadow framework. We discuss the performance on the penumbra area, the mask sensitive of previous methods, and we discuss the effectiveness of our provided losses in ablation study. + +Penumbra effect evaluation The penumbra region typically occupies only a small portion of the entire image, so improvements within this area have a limited impact on overall image metrics. To further assess the improvements introduced by using soft masks in the penumbra region, we calculate the PSNR and MAE values specifically within this area on the SRD dataset. For a fair comparison, we define the penumbra region following the methods from AEF [5], applying ground truth soft masks $\mathbf{s}_{gt}$ . The results are shown in Table 3. Compared with the SOTA methods like Homo + +Former [34], DeS3 [17] and Inpaint4Shadow [22]. We improve PSNR and MAE values in the penumbra area. These results show that on images containing soft shadows, our methods has better restoration performances. + +Mask sensitivity evaluation Previous shadow removal methods that rely on external shadow masks are sensitive to the accuracy of the input masks. Their performance can degrade when lack of ground truth mask. In contrast, end-to-end shadow removal methods mitigate the impact of low-quality masks. To further illustrate the sensitivity of prior SOTA methods to input masks, we evaluate their performance using various mask inputs. + +Specifically, we evaluate ShadowDiffusion [9] and HomoFormer [34] using a range of input masks, including ground truth (GT) masks, Otsu masks [14], BDRAR [37] detected masks, FDRNet [38] detected masks and pretrained SAM [19] detected masks as inputs. The PSNR results on the $\mathrm{ISTD + }$ [21] datasets are shown in Table 4. The Otsu mask follows the approach in MaskShadowGAN [14], using the difference between shadow-free and shadowed images to generate masks; both Otsu and GT masks include additional shadow-free information. Results indicate that our methods maintain consistent performance across all tests, whereas ShadowDiffusion and HomoFormer have performance declines when using generated masks. Although our method achieves slightly lower PSNR values than those obtained with Otsu or GT masks, this is reasonable given that the $\mathrm{ISTD + }$ dataset contains a larger proportion of hard shadow images, where our method has less of an advantage. Our approach surpasses the rest results that rely on masks generated by shadow detectors. + +Ablation Study We ablate our method, SoftShadow, by adding three losses: the shadow removal loss $\mathcal{L}_{rem}$ , the mask reconstruction loss $\mathcal{L}_{mask}$ , and the penumbra formation constraint loss $\mathcal{L}_{pen}$ , and demonstrate the results on SRD and LRSS datasets. First, we present the shadow removal results using the pretrained SAM and ShadowDiffusion models. Next, we finetune the framework with the shadow removal loss $\mathcal{L}_{rem}$ . Then, we add the mask reconstruction loss $\mathcal{L}_{mask}$ to provide SAM with more precise mask position guidance. Finally, we add the penumbra formation constraint loss $\mathcal{L}_{pen}$ to refine the mask boundaries for smoother transitions in the shadow regions. + +The results in Table 5 are based on training with the SRD training set and evaluating on both the SRD and LRSS datasets. As shown in Table 5, the shadow removal loss $\mathcal{L}_{rem}$ improves performance on the SRD dataset, primarily by enhancing the shadow removal network. Adding the mask reconstruction loss $\mathcal{L}_{mask}$ further fine-tunes SAM, producing more accurate shadow masks. This improvement of mask accuracy leads to better shadow removal results. To evaluate the impact of the penumbra formation constraint loss $\mathcal{L}_{pen}$ , we compare results with and without $\mathcal{L}_{pen}$ . The + +
MethodsLRSSSRD
PSNR↑SSIM↑MAE↓PSNR↑SSIM↑MAE↓
Pretrained Weights21.400.91012.2631.270.9634.41
Lrem21.780.91312.0735.270.9733.32
Lrem + Lmask23.080.9359.9735.440.9743.17
Lrem + Lmask + Lpen23.320.9339.7735.570.9753.11
+ +Table 5. The ablation studies for the shadow removal loss $\mathcal{L}_{rem}$ , the mask reconstruction loss $\mathcal{L}_{mask}$ , and the penumbra formation constraint loss $\mathcal{L}_{pen}$ in LRSS and SRD datasets. The best results are boldfaced. + +![](images/60082c305a0592f811a8470176c04645a6ab3299647ce5d58ed5652076771e78.jpg) +Figure 8. Visual comparison of results from our framework with different configurations: using pretrained weights, finetuned with the loss function $\mathcal{L}_{rem}$ , $\mathcal{L}_{rem} + \mathcal{L}_{mask}$ and our full configuration $\mathcal{L}_{rem} + \mathcal{L}_{mask} + \mathcal{L}_{pen}$ (Ours), respectively. + +results show that $\mathcal{L}_{pen}$ further aids in shadow removal on the SRD dataset, refining the shadow edges and enhancing the overall shadow removal quality. In the LRSS dataset, both $\mathcal{L}_{rem}$ and $\mathcal{L}_{mask}$ significantly improve performance, the $\mathcal{L}_{pen}$ improves the PSNR and MAE significantly, although it results in a slight decrease in SSIM. As shown in Figure 8, adding the constraint loss $\mathcal{L}_{pen}$ leads to more accurate and smoother soft mask predictions. The combined effects of $\mathcal{L}_{mask}$ and $\mathcal{L}_{pen}$ lead to more accurate and smoother soft mask predictions. Together, these losses provide the removal framework with enhanced soft shadow details, resulting in superior performance in shadow removal. + +# 6. Conclusion + +In this paper, we introduce novel soft shadow masks designed specifically for shadow removal. To achieve this, we propose a unified framework, SoftShadow, that eliminates the need for additional shadow mask input. By leveraging a pretrained SAM with LoRA, the framework accurately predicts soft masks as intermediate results, capturing detailed and varied shadow information. Additionally, we introduce a penumbra formation constraint, inspired by the physical shadow formation model, to jointly tune SAM and the shadow removal network, refining the soft mask in the penumbra area and facilitating artifact-free restoration. Extensive experiments demonstrate that our method is superior on various occasions, proving the validity of our method. + +# 7. Acknowledgment + +# References + +[1] Tianrun Chen, Lanyun Zhu, Chaotao Deng, Runlong Cao, Yan Wang, Shangzhan Zhang, Zejian Li, Lingyun Sun, Ying Zang, and Papa Mao. Sam-adapter: Adapting segment anything in underperformed scenes. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3367-3375, 2023. 3 +[2] Zhihao Chen, Lei Zhu, Liang Wan, Song Wang, Wei Feng, and Pheng-Ann Heng. A multi-task mean teacher for semi-supervised shadow detection. In Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition, pages 5611–5620, 2020. 3 +[3] Xiaodong Cun, Chi-Man Pun, and Cheng Shi. Towards ghost-free shadow removal via dual hierarchical aggregation network and shadow matting gan. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 10680-10687, 2020. 1, 2, 3, 5, 6 +[4] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2 +[5] Lan Fu, Changqing Zhou, Qing Guo, Felix Juefei-Xu, Hongkai Yu, Wei Feng, Yang Liu, and Song Wang. Autoexposure fusion for single-image shadow removal. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10571-10580, 2021. 3, 7 +[6] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 2 +[7] Maciej Gryka, Michael Terry, and Gabriel J Brostow. Learning to remove soft shadows. ACM Transactions on Graphics (TOG), 34(5):1-15, 2015. 5, 7 +[8] Lanqing Guo, Siyu Huang, Ding Liu, Hao Cheng, and Bihan Wen. Shadowformer: Global context helps shadow removal. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 710-718, 2023. 1, 3, 5 +[9] Lanqing Guo, Chong Wang, Wenhan Yang, Siyu Huang, Yufei Wang, Hanspeter Pfister, and Bihan Wen. Shadowdiffusion: When degradation prior meets diffusion model for shadow removal. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14049-14058, 2023. 1, 2, 5, 7, 8 +[10] Lanqing Guo, Chong Wang, Wenhan Yang, Yufei Wang, and Bihan Wen. Boundary-aware divide and conquer: A diffusion-based solution for unsupervised shadow removal. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13045-13054, 2023. 5 + +This work was partially supported by the National Research Foundation Singapore Competitive Research Program (award number CRP29-2022-0003) +[11] Ruiqi Guo, Qieyun Dai, and Derek Hoiem. Paired regions for shadow detection and removal. IEEE transactions on pattern analysis and machine intelligence, 35(12):2956-2967, 2012. 5 +[12] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2 +[13] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 2, 4, 5 +[14] Xiaowei Hu, Yitong Jiang, Chi-Wing Fu, and Pheng-Ann Heng. Mask-shadowgan: Learning to remove shadows from unpaired data. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2472–2481, 2019. 8 +[15] Leiping Jie and Hui Zhang. When sam meets shadow detection. arXiv preprint arXiv:2305.11513, 2023. 4 +[16] Yeying Jin, Aashish Sharma, and Robby T Tan. Dcshadownet: Single-image hard and soft shadow removal using unsupervised domain-classifier guided network. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5027-5036, 2021. 2, 5, 7 +[17] Yeying Jin, Wei Ye, Wenhan Yang, Yuan Yuan, and Robby T Tan. Des3: Adaptive attention-driven self and soft shadow removal using vit similarity. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2634-2642, 2024. 2, 5, 6, 7, 8 +[18] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.5 +[19] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 2, 3, 4, 5, 8 +[20] Michael S Langer and Steven W Zucker. What is a light source? In Proceedings of IEEE Computer Society Conference on Computer Vision and Pattern Recognition, pages 172-178. IEEE, 1997. 3 +[21] Hieu Le and Dimitris Samaras. Shadow removal via shadow image decomposition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8578-8587, 2019. 2, 5, 8 +[22] Xiaoguang Li, Qing Guo, Rabab Abdelfattah, Di Lin, Wei Feng, Ivor Tsang, and Song Wang. Leveraging inpainting for single-image shadow removal. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13055-13064, 2023. 5, 7, 8 +[23] Yuhao Liu, Zhanghan Ke, Ke Xu, Fang Liu, Zhenwei Wang, and Rynson WH Lau. Recasting regional lighting for shadow removal. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3810-3818, 2024. 1 +[24] Zhihao Liu, Hui Yin, Xinyi Wu, Zhenyao Wu, Yang Mi, and Song Wang. From shadow generation to shadow removal. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4927-4936, 2021. 5, 7 + +[25] Vinod Nair and Geoffrey E Hinton. Rectified linear units improve restricted boltzmann machines. In Proceedings of the 27th international conference on machine learning (ICML-10), pages 807–814, 2010. 5 +[26] Michael Nielsen and Claus B Madsen. Segmentation of soft shadows based on a daylight-and penumbra model. In International Conference on Computer Vision/Computer Graphics Collaboration Techniques and Applications, pages 341-352. Springer, 2007. 2 +[27] Liangqiong Qu, Jiandong Tian, Shengfeng He, Yandong Tang, and Rynson WH Lau. Deshadownet: A multi-context embedding deep network for shadow removal. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4067-4075, 2017. 2, 5, 7 +[28] Chitwan Sahara, Jonathan Ho, William Chan, Tim Salimans, David J Fleet, and Mohammad Norouzi. Image superresolution via iterative refinement. IEEE transactions on pattern analysis and machine intelligence, 45(4):4713-4726, 2022. 2 +[29] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 2 +[30] Jin Wan, Hui Yin, Zhenyao Wu, Xinyi Wu, Yanting Liu, and Song Wang. Style-guided shadow removal. In European Conference on Computer Vision, pages 361-378. Springer, 2022. 1, 5 +[31] Jifeng Wang, Xiang Li, and Jian Yang. Stacked conditional generative adversarial networks for jointly learning shadow detection and shadow removal. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1788-1797, 2018. 1, 2, 5 +[32] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 5 +[33] Cort J Willmott and Kenji Matsuura. Advantages of the mean absolute error (mae) over the root mean square error (rmse) in assessing average model performance. *Climate research*, 30(1):79–82, 2005. 5 +[34] Jie Xiao, Xueyang Fu, Yurui Zhu, Dong Li, Jie Huang, Kai Zhu, and Zheng-Jun Zha. Homoformer: Homogenized transformer for image shadow removal. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25617-25626, 2024. 1, 2, 3, 5, 7, 8 +[35] Xiaofeng Zhang, Chaochen Gu, and Shanying Zhu. Sam-helps-shadow: When segment anything model meet shadow removal. arXiv e-prints, pages arXiv-2306, 2023. 2, 5 +[36] Quanlong Zheng, Xiaotian Qiao, Ying Cao, and Rynson WH Lau. Distraction-aware shadow detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5167-5176, 2019. 3 +[37] Lei Zhu, Zijun Deng, Xiaowei Hu, Chi-Wing Fu, Xuemiao Xu, Jing Qin, and Pheng-Ann Heng. Bidirectional feature pyramid network with recurrent attention residual modules for shadow detection. In Proceedings of the European Conference on Computer Vision (ECCV), pages 121–136, 2018. + +[38] Lei Zhu, Ke Xu, Zhanghan Ke, and Rynson WH Lau. Mitigating intensity bias in shadow detection via feature decomposition and reweighting. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4702-4711, 2021. 5, 6, 8 +[39] Yurui Zhu, Jie Huang, Xueyang Fu, Feng Zhao, Qibin Sun, and Zheng-Jun Zha. Bijective mapping network for shadow removal. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5627-5636, 2022. 1, 3, 5, 7 \ No newline at end of file diff --git a/2025/SoftShadow_ Leveraging Soft Masks for Penumbra-Aware Shadow Removal/images.zip b/2025/SoftShadow_ Leveraging Soft Masks for Penumbra-Aware Shadow Removal/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..b27ce2f6d4dbd6662bc869faf9bc87998d7416c4 --- /dev/null +++ b/2025/SoftShadow_ Leveraging Soft Masks for Penumbra-Aware Shadow Removal/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d109e0b5bcfc0b6ab20331b02f223e16960c15b4a524e40649f77b2370d44ae1 +size 659203 diff --git a/2025/SoftShadow_ Leveraging Soft Masks for Penumbra-Aware Shadow Removal/layout.json b/2025/SoftShadow_ Leveraging Soft Masks for Penumbra-Aware Shadow Removal/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..88ad6260d94aa9a96cf60192be6e654f5a4e93a4 --- /dev/null +++ b/2025/SoftShadow_ Leveraging Soft Masks for Penumbra-Aware Shadow Removal/layout.json @@ -0,0 +1,10682 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 70, + 103, + 541, + 120 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 103, + 541, + 120 + ], + "spans": [ + { + "bbox": [ + 70, + 103, + 541, + 120 + ], + "type": "text", + "content": "SoftShadow: Leveraging Soft Masks for Penumbra-Aware Shadow Removal" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 135, + 141, + 478, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 141, + 478, + 227 + ], + "spans": [ + { + "bbox": [ + 135, + 141, + 478, + 227 + ], + "type": "text", + "content": "Xinrui Wang" + }, + { + "bbox": [ + 135, + 141, + 478, + 227 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 135, + 141, + 478, + 227 + ], + "type": "text", + "content": ", Lanqing Guo" + }, + { + "bbox": [ + 135, + 141, + 478, + 227 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 135, + 141, + 478, + 227 + ], + "type": "text", + "content": ", Xiyu Wang" + }, + { + "bbox": [ + 135, + 141, + 478, + 227 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 135, + 141, + 478, + 227 + ], + "type": "text", + "content": ", Siyu Huang" + }, + { + "bbox": [ + 135, + 141, + 478, + 227 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 135, + 141, + 478, + 227 + ], + "type": "text", + "content": ", Bihan Wen" + }, + { + "bbox": [ + 135, + 141, + 478, + 227 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 135, + 141, + 478, + 227 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 135, + 141, + 478, + 227 + ], + "type": "text", + "content": "Nanyang Technological University, Singapore \n" + }, + { + "bbox": [ + 135, + 141, + 478, + 227 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 135, + 141, + 478, + 227 + ], + "type": "text", + "content": "The University of Texas at Austin, USA \n" + }, + { + "bbox": [ + 135, + 141, + 478, + 227 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 135, + 141, + 478, + 227 + ], + "type": "text", + "content": "Clemson University, USA \n{xinrui.wang, bihan.wen}@ntu.edu.sg, xiyu004@e.ntu.edu.sg, lanqing.guo@austin.utexas.edu, siyuh@clemson.edu" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 151, + 255, + 200, + 268 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 255, + 200, + 268 + ], + "spans": [ + { + "bbox": [ + 151, + 255, + 200, + 268 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 281, + 296, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 281, + 296, + 533 + ], + "spans": [ + { + "bbox": [ + 55, + 281, + 296, + 533 + ], + "type": "text", + "content": "Recent advancements in deep learning have yielded promising results for the image shadow removal task. However, most existing methods rely on binary pre-generated shadow masks. The binary nature of such masks could potentially lead to artifacts near the boundary between shadow and non-shadow areas. In view of this, inspired by the physical model of shadow formation, we introduce novel soft shadow masks specifically designed for shadow removal. To achieve such soft masks, we propose a SoftShadow framework by leveraging the prior knowledge of pretrained SAM and integrating physical constraints. Specifically, we jointly tune the SAM and the subsequent shadow removal network using penumbra formation constraint loss, mask reconstruction loss, and shadow removal loss. This framework enables accurate predictions of penumbra (partially shaded) and umbra (fully shaded) areas while simultaneously facilitating end-to-end shadow removal. Through extensive experiments on popular datasets, we found that our SoftShadow framework, which generates soft masks, can better restore boundary artifacts, achieve state-of-the-art performance, and demonstrate superior generalizability." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 555, + 135, + 568 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 555, + 135, + 568 + ], + "spans": [ + { + "bbox": [ + 56, + 555, + 135, + 568 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 576, + 295, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 576, + 295, + 684 + ], + "spans": [ + { + "bbox": [ + 55, + 576, + 295, + 684 + ], + "type": "text", + "content": "Shadow removal aims to restore content obscured in shadow regions and correct degraded illumination. Recently, deep learning methods have shown excellent performance in shadow removal tasks relying on large-scale training data. However, one of the main challenges in shadow removal arises in images with soft shadows, where the shadow edges are blurred rather than sharp. This blurred area, known as the penumbra area, occurs due to partial occlusion of light, creating a transition zone where illumination varies signif" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 320, + 256, + 375, + 309 + ], + "blocks": [ + { + "bbox": [ + 320, + 256, + 375, + 309 + ], + "lines": [ + { + "bbox": [ + 320, + 256, + 375, + 309 + ], + "spans": [ + { + "bbox": [ + 320, + 256, + 375, + 309 + ], + "type": "image", + "image_path": "ae3c4a2e397052d05a3208cf1a1029d799c251ea051cade9b83adf59a35af015.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 339, + 310, + 354, + 316 + ], + "lines": [ + { + "bbox": [ + 339, + 310, + 354, + 316 + ], + "spans": [ + { + "bbox": [ + 339, + 310, + 354, + 316 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 378, + 257, + 432, + 309 + ], + "blocks": [ + { + "bbox": [ + 378, + 257, + 432, + 309 + ], + "lines": [ + { + "bbox": [ + 378, + 257, + 432, + 309 + ], + "spans": [ + { + "bbox": [ + 378, + 257, + 432, + 309 + ], + "type": "image", + "image_path": "67e9c45e2022f38ef0bb891db4b09294360a58ec0f3a780ee648ce2b3e0cca15.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 388, + 310, + 425, + 316 + ], + "lines": [ + { + "bbox": [ + 388, + 310, + 425, + 316 + ], + "spans": [ + { + "bbox": [ + 388, + 310, + 425, + 316 + ], + "type": "text", + "content": "Ground Truth" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 436, + 256, + 490, + 309 + ], + "blocks": [ + { + "bbox": [ + 436, + 256, + 490, + 309 + ], + "lines": [ + { + "bbox": [ + 436, + 256, + 490, + 309 + ], + "spans": [ + { + "bbox": [ + 436, + 256, + 490, + 309 + ], + "type": "image", + "image_path": "6e1e11d3db043007d7f6f8955ec7469e228e8d42d91d556ac2895721a6062941.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 446, + 310, + 476, + 316 + ], + "lines": [ + { + "bbox": [ + 446, + 310, + 476, + 316 + ], + "spans": [ + { + "bbox": [ + 446, + 310, + 476, + 316 + ], + "type": "text", + "content": "Hard Mask" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 318, + 318, + 373, + 372 + ], + "blocks": [ + { + "bbox": [ + 318, + 318, + 373, + 372 + ], + "lines": [ + { + "bbox": [ + 318, + 318, + 373, + 372 + ], + "spans": [ + { + "bbox": [ + 318, + 318, + 373, + 372 + ], + "type": "image", + "image_path": "19c30de583a65975a35143c95c28edfa7f8f99ba3b237a54341be51825713016.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 337, + 373, + 358, + 379 + ], + "lines": [ + { + "bbox": [ + 337, + 373, + 358, + 379 + ], + "spans": [ + { + "bbox": [ + 337, + 373, + 358, + 379 + ], + "type": "text", + "content": "BMNet" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 378, + 318, + 432, + 372 + ], + "blocks": [ + { + "bbox": [ + 378, + 318, + 432, + 372 + ], + "lines": [ + { + "bbox": [ + 378, + 318, + 432, + 372 + ], + "spans": [ + { + "bbox": [ + 378, + 318, + 432, + 372 + ], + "type": "image", + "image_path": "1b1236cd360fc0784e1a3091948197621d5f59d9324f67d0db23dbe8b69e72e3.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 383, + 373, + 426, + 379 + ], + "lines": [ + { + "bbox": [ + 383, + 373, + 426, + 379 + ], + "spans": [ + { + "bbox": [ + 383, + 373, + 426, + 379 + ], + "type": "text", + "content": "SG-ShadowNet" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 436, + 318, + 490, + 372 + ], + "blocks": [ + { + "bbox": [ + 436, + 318, + 490, + 372 + ], + "lines": [ + { + "bbox": [ + 436, + 318, + 490, + 372 + ], + "spans": [ + { + "bbox": [ + 436, + 318, + 490, + 372 + ], + "type": "image", + "image_path": "5f13ef3994af938eb7618caabf0a0d3c3dedee560971aba38f9c1777587373ef.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 445, + 373, + 482, + 379 + ], + "lines": [ + { + "bbox": [ + 445, + 373, + 482, + 379 + ], + "spans": [ + { + "bbox": [ + 445, + 373, + 482, + 379 + ], + "type": "text", + "content": "HomoFormer" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 393, + 555, + 449 + ], + "lines": [ + { + "bbox": [ + 313, + 393, + 555, + 449 + ], + "spans": [ + { + "bbox": [ + 313, + 393, + 555, + 449 + ], + "type": "text", + "content": "Figure 1. Illustration of soft shadow removal results using our proposed SoftShadow with soft shadow mask compared to the recent competing methods BMNet [39], SG-ShadowNet [30] and HomoFormer [34] using hard shadow masks. The second rows are sharpened versions of results for better visualization." + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 495, + 255, + 550, + 309 + ], + "blocks": [ + { + "bbox": [ + 495, + 255, + 550, + 309 + ], + "lines": [ + { + "bbox": [ + 495, + 255, + 550, + 309 + ], + "spans": [ + { + "bbox": [ + 495, + 255, + 550, + 309 + ], + "type": "image", + "image_path": "7da857832245867069b99057960d91d50bab59dc1b082f5f780e9c58dd8e5292.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 506, + 310, + 535, + 316 + ], + "lines": [ + { + "bbox": [ + 506, + 310, + 535, + 316 + ], + "spans": [ + { + "bbox": [ + 506, + 310, + 535, + 316 + ], + "type": "text", + "content": "Soft Mask" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 494, + 318, + 549, + 372 + ], + "blocks": [ + { + "bbox": [ + 494, + 318, + 549, + 372 + ], + "lines": [ + { + "bbox": [ + 494, + 318, + 549, + 372 + ], + "spans": [ + { + "bbox": [ + 494, + 318, + 549, + 372 + ], + "type": "image", + "image_path": "d1e43dda1e4a1f971d21d3c2c2eff173642e016567d5a3bc2ee2656992228380.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 515, + 373, + 529, + 379 + ], + "lines": [ + { + "bbox": [ + 515, + 373, + 529, + 379 + ], + "spans": [ + { + "bbox": [ + 515, + 373, + 529, + 379 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 472, + 555, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 472, + 555, + 519 + ], + "spans": [ + { + "bbox": [ + 313, + 472, + 555, + 519 + ], + "type": "text", + "content": "icantly. Removing shadows in these regions is challenging because the gradual light-to-shadow transition often leads to boundary artifacts, compromising the quality of the shadow removal results." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 522, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 522, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 522, + 556, + 714 + ], + "type": "text", + "content": "A few methods rely on hard shadow masks (binary masks) to indicate the shadow regions [8, 9, 34, 39]. These binary shadow masks are either manually annotated [23, 31], or predicted by off-the-shelf shadow detectors [3]. Obtaining such shadow masks can be costly and complicated, and the choice of shadow detector can significantly impact shadow removal performance. Moreover, hard masks fall short of representing the penumbra area, causing boundary artifacts in removal results, especially when the penumbra areas are highly pronounced. As shown in Figure 1, when the input image contains penumbra regions, the previous state-of-the-art (SOTA) method [34] suffers from boundary artifacts. Other approaches attempt shadow removal without explicitly extracting the shadow mask. Instead, they incorporate semantic parsing modules into shadow removal networks. For instance, some approaches involve predict" + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 693, + 186, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 693, + 186, + 703 + ], + "spans": [ + { + "bbox": [ + 66, + 693, + 186, + 703 + ], + "type": "text", + "content": "*Equal contribution for both authors" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 68, + 703, + 143, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 703, + 143, + 712 + ], + "spans": [ + { + "bbox": [ + 68, + 703, + 143, + 712 + ], + "type": "text", + "content": "† Corresponding author" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23217" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 65, + 70, + 162, + 195 + ], + "blocks": [ + { + "bbox": [ + 65, + 70, + 162, + 195 + ], + "lines": [ + { + "bbox": [ + 65, + 70, + 162, + 195 + ], + "spans": [ + { + "bbox": [ + 65, + 70, + 162, + 195 + ], + "type": "image", + "image_path": "c01c5cf9d5db6c178fb4b2f1ddad8409cea2696767574445619bc2746d04510e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 207, + 295, + 274 + ], + "lines": [ + { + "bbox": [ + 55, + 207, + 295, + 274 + ], + "spans": [ + { + "bbox": [ + 55, + 207, + 295, + 274 + ], + "type": "text", + "content": "Figure 2. (a) Illustration of the shadow formation geometry that creates the penumbra and umbra [26] regions; the umbra area is where the light is fully occluded, the penumbra area is where light is partially occluded, the lit area is where light is not occluded. (b) Examples of soft shadow images and hard shadow images from commonly used datasets." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 168, + 70, + 288, + 195 + ], + "blocks": [ + { + "bbox": [ + 168, + 70, + 288, + 195 + ], + "lines": [ + { + "bbox": [ + 168, + 70, + 288, + 195 + ], + "spans": [ + { + "bbox": [ + 168, + 70, + 288, + 195 + ], + "type": "image", + "image_path": "4ca7a5d78bf1d8ac1fa537ee83766e6540d9684b8b023c6dda47621b57d99b73.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 295, + 295, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 295, + 295, + 354 + ], + "spans": [ + { + "bbox": [ + 55, + 295, + 295, + 354 + ], + "type": "text", + "content": "ing degradation attention [3, 27], while others utilize domain classifiers [16, 17] to better understand and address shadow effects. However, their performance may be limited by the lack of powerful external detectors providing additional shadow location information." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 355, + 295, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 355, + 295, + 533 + ], + "spans": [ + { + "bbox": [ + 55, + 355, + 295, + 533 + ], + "type": "text", + "content": "In view of this, we argue that simply leveraging the binary mask to represent the shadow location is not enough. Differently, we introduce a novel soft shadow mask (grayscale mask) specifically designed for shadow removal as shown in the hard mask v.s. soft mask in Figure 1. The soft masks can precisely locate the inner and outer boundaries of shadow and indicate the penumbra area with proper degradation variance. Moreover, recognizing the effectiveness of employing a powerful pretrained detector for providing semantic information, we utilize the prior knowledge from the pretrained SAM [19]. While existing methods [35] have used the off-the-shelf SAM in a naive way, merely predicting the shadow mask as input to the shadow removal network, we adapt SAM as a soft shadow mask predictor and jointly optimize it with the shadow removal network." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "text", + "content": "In this paper, we first propose a unified shadow removal framework, dubbed SoftShadow, which removes shadow from images in an end-to-end manner without requiring input shadow masks as shown in Figure 3. We integrate the pretrained SAM [19] combined with Low-Rank Adaptation (LoRA) [13] to jointly adapt the SAM model and shadow removal model. The framework enables accurate predictions of soft shadow masks as intermediate results while simultaneously facilitating end-to-end shadow removal. Besides, we introduce a penumbra formation constraint to assist predict soft shadow masks. This constraint regularizes the gradient of the predicted mask in the penumbra area, resulting in more accurate predictions with detailed and spatially varied shadow masks. In the commonly used shadow removal datasets, SRD and LRSS datasets contain" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 553, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 131 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 131 + ], + "type": "text", + "content": "soft shadow images, which makes them more suitable for illustrating the improvements of utilizing soft shadow masks. Experimental results show that SoftShadow achieves superior performance on SRD and LRSS datasets. Our main contributions are summarized as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 136, + 553, + 327 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 314, + 136, + 553, + 206 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 136, + 553, + 206 + ], + "spans": [ + { + "bbox": [ + 314, + 136, + 553, + 206 + ], + "type": "text", + "content": "- We propose a unified shadow removal framework that does not require additional input shadow masks. This framework enables accurate predictions of soft masks as intermediate results, which is specifically designed to capture detailed and spatially varied shadow location information." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 208, + 553, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 208, + 553, + 266 + ], + "spans": [ + { + "bbox": [ + 314, + 208, + 553, + 266 + ], + "type": "text", + "content": "- We introduce penumbra formation constraint inspired by the physical shadow formation model to further refine the soft mask in the penumbra area. By leveraging the constraint loss and shadow removal loss, we jointly tune the SAM and the subsequent shadow removal network." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 267, + 553, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 267, + 553, + 327 + ], + "spans": [ + { + "bbox": [ + 314, + 267, + 553, + 327 + ], + "type": "text", + "content": "- Experimental results demonstrate that SoftShadow surpasses state-of-the-art shadow removal methods on the SRD and LRSS datasets, achieving superior performance and even comparable results with previous methods that use ground truth mask inputs." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 339, + 400, + 352 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 339, + 400, + 352 + ], + "spans": [ + { + "bbox": [ + 314, + 339, + 400, + 352 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 360, + 418, + 372 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 360, + 418, + 372 + ], + "spans": [ + { + "bbox": [ + 314, + 360, + 418, + 372 + ], + "type": "text", + "content": "2.1. Shadow Removal" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 378, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 378, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 378, + 555, + 714 + ], + "type": "text", + "content": "The degradation of shadows varies in each image, posing a significant challenge for shadow image restoration. Recent deep learning-based approaches have achieved remarkable results in shadow removal. Some methods restore the shadow image with the guidance of shadow masks. For example, SP+M-Net [21] employs two deep networks to predict shadow matte and shadow parameters. Recently, powerful backbones such as transformers [4, 29], and diffusion models [12, 28] have been applied to the shadow removal task. HomoFormer [34] homogenizes the spatial distribution of shadow masks to uniformly recover the entire shadow image, while ShadowDiffusion [9] provides a robust generative prior for producing natural shadow-free images. Some methods aim to eliminate the dependency on shadow mask inputs. DeShadowNet [27] introduces an end-to-end shadow removal method containing a multi-branch fusion module. ST-CGAN [31] connects two GANs [6] in sequence to jointly detect and remove shadows. DC-ShadowNet [16] provides an unsupervised domain-classifier discriminator for guided shadow removal network. More recently, DeS3 [17] implemented a method capable of removing shadows that are cast on the object itself without requiring shadow mask inputs. However, due to the absence of additional shadow location information, their performance may degrade, and the number of parameters used could increase significantly. Different from these existing methods, we introduce a new concept \"soft mask\", designed explicitly for shadow removal." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23218" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 105, + 68, + 504, + 228 + ], + "blocks": [ + { + "bbox": [ + 105, + 68, + 504, + 228 + ], + "lines": [ + { + "bbox": [ + 105, + 68, + 504, + 228 + ], + "spans": [ + { + "bbox": [ + 105, + 68, + 504, + 228 + ], + "type": "image", + "image_path": "a7ca1b72b53a7e66f7e79e36318a5bf2df0b9a26991544554e71ec1eb761e1d7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 237, + 555, + 314 + ], + "lines": [ + { + "bbox": [ + 54, + 237, + 555, + 314 + ], + "spans": [ + { + "bbox": [ + 54, + 237, + 555, + 314 + ], + "type": "text", + "content": "Figure 3. Illustration of the proposed SoftShadow. The left box illustrates the SoftShadow networks, where a shadow image " + }, + { + "bbox": [ + 54, + 237, + 555, + 314 + ], + "type": "inline_equation", + "content": "\\mathbf{y}" + }, + { + "bbox": [ + 54, + 237, + 555, + 314 + ], + "type": "text", + "content": " is input into SAM for detecting soft shadow masks. The shadow removal network then processes the soft mask and the shadow image to produce a shadow-free image. The right box shows the three losses we used in SoftShadow. From top to bottom, the shadow removal loss " + }, + { + "bbox": [ + 54, + 237, + 555, + 314 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{rem}" + }, + { + "bbox": [ + 54, + 237, + 555, + 314 + ], + "type": "text", + "content": " is calculated between shadow-free images and shadow-removal images. The mask reconstruction loss " + }, + { + "bbox": [ + 54, + 237, + 555, + 314 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{mask}" + }, + { + "bbox": [ + 54, + 237, + 555, + 314 + ], + "type": "text", + "content": " is calculated between predicted soft masks and ours ground truth soft masks. The penumbra formation constraint loss " + }, + { + "bbox": [ + 54, + 237, + 555, + 314 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{pen}" + }, + { + "bbox": [ + 54, + 237, + 555, + 314 + ], + "type": "text", + "content": " acts as a regularization term. It aims to regularize the gradient of predicted soft masks in the penumbra area. As shown in the soft mask intensity curve, the ideal mask intensity in the penumbra area should not be too large and the gradient orientation should be consistent." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 335, + 162, + 346 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 335, + 162, + 346 + ], + "spans": [ + { + "bbox": [ + 55, + 335, + 162, + 346 + ], + "type": "text", + "content": "2.2. Shadow Detection" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 354, + 296, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 354, + 296, + 617 + ], + "spans": [ + { + "bbox": [ + 56, + 354, + 296, + 617 + ], + "type": "text", + "content": "Shadow detection can be approached as a segmentation task, it often struggles with small shadows and indistinct shadow edges. Various methods have been proposed to address these challenges. BDRAR [37] introduces a bidirectional pyramidal architecture for shadow detection. DSD [36] designs a distraction-aware module to minimize false positives in shadow detection. Chen et al. [2] propose a task-specific semi-supervised learning mechanism to utilize unlabeled shadow images for detection, enhancing the robustness of the model. DHAN [3] is a well-used shadow detection model using dual hierarchical aggregation network training on synthetic shadow and shadow-free image pairs to improve detection accuracy. The SAM-Adapter [1] has been utilized to fine-tune the Segment Anything Model (SAM) [19] specifically for shadow detection tasks. While this approach improves accuracy, it demands a substantial number of model parameters. However, most existing methods detect shadows as hard masks and fail to represent the penumbra area. In contrast, our approach leverages the shadow removal loss and the Penumbra Formation Constraint loss, jointly training the shadow removal network and the shadow detection network." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 632, + 116, + 644 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 632, + 116, + 644 + ], + "spans": [ + { + "bbox": [ + 55, + 632, + 116, + 644 + ], + "type": "text", + "content": "3. Methods" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 654, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 295, + 715 + ], + "type": "text", + "content": "In this section, we first explain the motivation behind our newly introduced soft shadow mask, inspired by the physical model of shadow formation. We then propose the unified shadow removal framework SoftShadow, which leverages the powerful segmentation capabilities of SAM to pro" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 335, + 555, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 335, + 555, + 407 + ], + "spans": [ + { + "bbox": [ + 313, + 335, + 555, + 407 + ], + "type": "text", + "content": "duce the soft shadow mask and guide the removal network to generate better shadow-free images. Finally, to further mitigate boundary artifacts, we introduce the penumbra formation constraint loss which ensures a smoother transition between the shadow region and the non-shadow region, offering better guidance to the shadow removal network." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 418, + 389, + 429 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 418, + 389, + 429 + ], + "spans": [ + { + "bbox": [ + 313, + 418, + 389, + 429 + ], + "type": "text", + "content": "3.1. Motivation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 436, + 555, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 436, + 555, + 544 + ], + "spans": [ + { + "bbox": [ + 313, + 436, + 555, + 544 + ], + "type": "text", + "content": "Shadow degradations in the real world exhibit considerable amount of variation and can be classified into two main categories: soft shadows and hard shadows [20]. This classification depends on the light source and the distance between the object and the surface, as shown in Figure 2(a). Soft shadows are characterized by blurry edges and gradual transitions from light to dark, creating relatively large penumbra areas, while hard shadows have sharp edges with few to none penumbra regions." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 544, + 556, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 544, + 556, + 675 + ], + "spans": [ + { + "bbox": [ + 313, + 544, + 556, + 675 + ], + "type": "text", + "content": "Previous methods [5, 8, 34, 39] commonly use hard masks to guide the shadow removal process. The hard masks effectively provide positional guidance for hard masks with sharp edges. However, when the penumbra regions are more pronounced, as shown in Figure 2 (b), the boundary between the umbra and the lit region is hard to determine, making hard masks unsuitable for representing such soft shadows. Additionally, we observed that, in general, the brightness transition within the penumbra area can vary significantly, whereas the brightness of the lit and the umbra areas are relatively uniform." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "content": "To this end, we instead use soft (greyscale) masks to represent the position of the shadows. Formally, we denote the soft mask as s. For a given shadow-free image " + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "content": ", the" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23219" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 234, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 234, + 83 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 234, + 83 + ], + "type": "text", + "content": "shadow image " + }, + { + "bbox": [ + 55, + 72, + 234, + 83 + ], + "type": "inline_equation", + "content": "\\mathbf{y}" + }, + { + "bbox": [ + 55, + 72, + 234, + 83 + ], + "type": "text", + "content": " can be formulated as such:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 118, + 95, + 295, + 108 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 95, + 295, + 108 + ], + "spans": [ + { + "bbox": [ + 118, + 95, + 295, + 108 + ], + "type": "interline_equation", + "content": "\\mathbf {y} = \\mathbf {a} \\cdot \\mathbf {s} \\cdot \\mathbf {x} + (\\mathbf {1} - \\mathbf {s}) \\cdot \\mathbf {x}, \\tag {1}", + "image_path": "c5a7de2e74249582b969d3a0e04a23a7472cdf71f18f38bbed1ddc575e002582.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 118, + 296, + 214 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 118, + 296, + 214 + ], + "spans": [ + { + "bbox": [ + 55, + 118, + 296, + 214 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 118, + 296, + 214 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 55, + 118, + 296, + 214 + ], + "type": "text", + "content": " denotes element-wise multiplication, " + }, + { + "bbox": [ + 55, + 118, + 296, + 214 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 55, + 118, + 296, + 214 + ], + "type": "text", + "content": " is the soft shadow mask in which constant " + }, + { + "bbox": [ + 55, + 118, + 296, + 214 + ], + "type": "inline_equation", + "content": "s = 0" + }, + { + "bbox": [ + 55, + 118, + 296, + 214 + ], + "type": "text", + "content": " represents lit area, " + }, + { + "bbox": [ + 55, + 118, + 296, + 214 + ], + "type": "inline_equation", + "content": "s = 1" + }, + { + "bbox": [ + 55, + 118, + 296, + 214 + ], + "type": "text", + "content": " represents umbra area, and " + }, + { + "bbox": [ + 55, + 118, + 296, + 214 + ], + "type": "inline_equation", + "content": "s \\in [0,1]" + }, + { + "bbox": [ + 55, + 118, + 296, + 214 + ], + "type": "text", + "content": " represents penumbra area. " + }, + { + "bbox": [ + 55, + 118, + 296, + 214 + ], + "type": "inline_equation", + "content": "a \\in [0,1]" + }, + { + "bbox": [ + 55, + 118, + 296, + 214 + ], + "type": "text", + "content": " represents the illumination weight in the non-lit area. Based on our experiment, we quantitatively and qualitatively found that the proposed shadow degradation model could lead to fewer boundary artifacts and generally better shadow removal results." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 222, + 208, + 233 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 222, + 208, + 233 + ], + "spans": [ + { + "bbox": [ + 55, + 222, + 208, + 233 + ], + "type": "text", + "content": "3.2. Architecture of SoftShadow" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 239, + 296, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 239, + 296, + 418 + ], + "spans": [ + { + "bbox": [ + 55, + 239, + 296, + 418 + ], + "type": "text", + "content": "SoftShadow is a unified framework designed to remove object shadows in images in an end-to-end manner. The overall architecture is illustrated in Figure 3. SoftShadow contains a soft shadow detector and a subsequent shadow removal network. The soft shadow detector leverages the strong segmentation capabilities of SAM and further finetuning it to serve as a soft mask detector. This is because the pretrained SAM can fail to accurately identify shadows [15, 19]. Moreover, since SAM has noticeably more parameters compared to conventional shadow detection networks, we use Low-Rank Adaptation (LoRA) [13] to reduce the number of trainable parameters. In practice, we tune all the self-attention blocks of image encoders using LoRA. Since the mask decoder is lightweight, we tune all the parameters of it." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 419, + 296, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 419, + 296, + 539 + ], + "spans": [ + { + "bbox": [ + 55, + 419, + 296, + 539 + ], + "type": "text", + "content": "SAM is mainly trained for binary predictions, unlike the soft masks we aim to generate. To leverage SAM to generate continuous soft masks, we further introduce the mask reconstruction loss " + }, + { + "bbox": [ + 55, + 419, + 296, + 539 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{mask}" + }, + { + "bbox": [ + 55, + 419, + 296, + 539 + ], + "type": "text", + "content": " and the penumbra formation constraint loss " + }, + { + "bbox": [ + 55, + 419, + 296, + 539 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{pen}" + }, + { + "bbox": [ + 55, + 419, + 296, + 539 + ], + "type": "text", + "content": " to model the characteristics of soft masks (details are described in Section 3.3). They encourage the model to detect the shadow area and produce a continuous soft mask rather than a binary one, thereby capturing the shadow position more accurately. We formulate a mask reconstruction loss " + }, + { + "bbox": [ + 55, + 419, + 296, + 539 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{mask}" + }, + { + "bbox": [ + 55, + 419, + 296, + 539 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 120, + 548, + 295, + 564 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 120, + 548, + 295, + 564 + ], + "spans": [ + { + "bbox": [ + 120, + 548, + 295, + 564 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {m a s k} = \\mathbb {E} _ {n} \\left\\| \\hat {\\mathbf {s}} - \\mathbf {s} _ {g t} \\right\\| _ {F} ^ {2}, \\tag {2}", + "image_path": "7e31548bf5beb9e2164753c6240aa267ef4b27fcf1322c9d46ca46edd5969bcf.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 572, + 296, + 679 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 572, + 296, + 679 + ], + "spans": [ + { + "bbox": [ + 55, + 572, + 296, + 679 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 572, + 296, + 679 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 572, + 296, + 679 + ], + "type": "text", + "content": " is the index of shadow masks, " + }, + { + "bbox": [ + 55, + 572, + 296, + 679 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{s}}" + }, + { + "bbox": [ + 55, + 572, + 296, + 679 + ], + "type": "text", + "content": " is the predicted soft shadow mask, and " + }, + { + "bbox": [ + 55, + 572, + 296, + 679 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_{gt}" + }, + { + "bbox": [ + 55, + 572, + 296, + 679 + ], + "type": "text", + "content": " is the ground truth soft shadow mask achieved by dividing the shadow-free image " + }, + { + "bbox": [ + 55, + 572, + 296, + 679 + ], + "type": "inline_equation", + "content": "\\mathbf{y}" + }, + { + "bbox": [ + 55, + 572, + 296, + 679 + ], + "type": "text", + "content": " by the shadow image " + }, + { + "bbox": [ + 55, + 572, + 296, + 679 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 55, + 572, + 296, + 679 + ], + "type": "text", + "content": ". Specifically, we convert " + }, + { + "bbox": [ + 55, + 572, + 296, + 679 + ], + "type": "inline_equation", + "content": "\\mathbf{y}" + }, + { + "bbox": [ + 55, + 572, + 296, + 679 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 572, + 296, + 679 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 55, + 572, + 296, + 679 + ], + "type": "text", + "content": " into YCbCr image space, using the Y channel to illustrate the brightness of these images, denoted as " + }, + { + "bbox": [ + 55, + 572, + 296, + 679 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_Y" + }, + { + "bbox": [ + 55, + 572, + 296, + 679 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 572, + 296, + 679 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_Y" + }, + { + "bbox": [ + 55, + 572, + 296, + 679 + ], + "type": "text", + "content": ", respectively. We then divide the " + }, + { + "bbox": [ + 55, + 572, + 296, + 679 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_Y" + }, + { + "bbox": [ + 55, + 572, + 296, + 679 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 55, + 572, + 296, + 679 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_Y" + }, + { + "bbox": [ + 55, + 572, + 296, + 679 + ], + "type": "text", + "content": " to obtain the difference between them, apply a low-pass filter " + }, + { + "bbox": [ + 55, + 572, + 296, + 679 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 55, + 572, + 296, + 679 + ], + "type": "text", + "content": " to reduce noise, and use a threshold to eliminate outliers as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 117, + 689, + 295, + 717 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 117, + 689, + 295, + 717 + ], + "spans": [ + { + "bbox": [ + 117, + 689, + 295, + 717 + ], + "type": "interline_equation", + "content": "\\mathbf {s} _ {g t} = \\max \\left(t, f \\left(\\frac {\\mathbf {x} _ {Y}}{\\mathbf {y} _ {Y}}\\right)\\right), \\tag {3}", + "image_path": "f33dfc47ecfab024b0955fb7d107be06ccfac7f6938c5e6f0dea725aa629b712.jpg" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 354, + 69, + 512, + 198 + ], + "blocks": [ + { + "bbox": [ + 354, + 69, + 512, + 198 + ], + "lines": [ + { + "bbox": [ + 354, + 69, + 512, + 198 + ], + "spans": [ + { + "bbox": [ + 354, + 69, + 512, + 198 + ], + "type": "image", + "image_path": "3495dc65292d8cfc40319ed1daacfd1a6ad2258dfba97d292179c530b6891aae.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 209, + 555, + 285 + ], + "lines": [ + { + "bbox": [ + 313, + 209, + 555, + 285 + ], + "spans": [ + { + "bbox": [ + 313, + 209, + 555, + 285 + ], + "type": "text", + "content": "Figure 4. Illustrate the concept of Penumbra Formation Constraint. The " + }, + { + "bbox": [ + 313, + 209, + 555, + 285 + ], + "type": "inline_equation", + "content": "w_{0}" + }, + { + "bbox": [ + 313, + 209, + 555, + 285 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 209, + 555, + 285 + ], + "type": "inline_equation", + "content": "w_{1}" + }, + { + "bbox": [ + 313, + 209, + 555, + 285 + ], + "type": "text", + "content": " are points in the inner boundary and the outer boundary of the penumbra area, respectively. In the ideal soft shadow mask, the intensity in the penumbra area decreases progressively from the shadow center to the umbra area. The gradient orientation points from the shadow center towards the non-shadow area." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 308, + 553, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 308, + 553, + 332 + ], + "spans": [ + { + "bbox": [ + 313, + 308, + 553, + 332 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 308, + 553, + 332 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 308, + 553, + 332 + ], + "type": "text", + "content": " is the threshold used to define the outer boundary between the penumbra area and the lit area." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 333, + 553, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 333, + 553, + 369 + ], + "spans": [ + { + "bbox": [ + 313, + 333, + 553, + 369 + ], + "type": "text", + "content": "Moreover, as illustrate in Figure 3, we incorporet the gradient of the shadow removal loss " + }, + { + "bbox": [ + 313, + 333, + 553, + 369 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{rem}" + }, + { + "bbox": [ + 313, + 333, + 553, + 369 + ], + "type": "text", + "content": " to jointly train SAM and the subsequent shadow removal network:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 384, + 377, + 553, + 393 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 384, + 377, + 553, + 393 + ], + "spans": [ + { + "bbox": [ + 384, + 377, + 553, + 393 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {r e m}} = \\mathbb {E} _ {n} \\| \\hat {\\mathbf {x}} - \\mathbf {x} \\| _ {F} ^ {2}, \\tag {4}", + "image_path": "a419bf14aa1d5504317bf1fdc4d7bee0ac4a5479fc058b383d70fe53774e7798.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 402, + 554, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 402, + 554, + 472 + ], + "spans": [ + { + "bbox": [ + 313, + 402, + 554, + 472 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 402, + 554, + 472 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 402, + 554, + 472 + ], + "type": "text", + "content": " indexes all images, " + }, + { + "bbox": [ + 313, + 402, + 554, + 472 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}}" + }, + { + "bbox": [ + 313, + 402, + 554, + 472 + ], + "type": "text", + "content": " is the restored shadow-free image, and " + }, + { + "bbox": [ + 313, + 402, + 554, + 472 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 313, + 402, + 554, + 472 + ], + "type": "text", + "content": " is the ground truth shadow-free image. To this end, the intermediate soft shadow masks can be refined using shadow removal results, with the SAM adapter optimized for better location guidance in subsequent shadow removal." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 326, + 474, + 520, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 474, + 520, + 486 + ], + "spans": [ + { + "bbox": [ + 326, + 474, + 520, + 486 + ], + "type": "text", + "content": "The overall training objective of SoftShadow is:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 364, + 497, + 553, + 510 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 364, + 497, + 553, + 510 + ], + "spans": [ + { + "bbox": [ + 364, + 497, + 553, + 510 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {\\text {m a s k}} + \\lambda_ {1} \\mathcal {L} _ {\\text {p e n}} + \\lambda_ {2} \\mathcal {L} _ {\\text {r e m}}, \\tag {5}", + "image_path": "96a417fe4c711f9f22025835ff5fb490b6ec2bf2f2d3316d3bff577fb2af3d06.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 520, + 554, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 520, + 554, + 567 + ], + "spans": [ + { + "bbox": [ + 313, + 520, + 554, + 567 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 520, + 554, + 567 + ], + "type": "inline_equation", + "content": "\\lambda_{1}" + }, + { + "bbox": [ + 313, + 520, + 554, + 567 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 520, + 554, + 567 + ], + "type": "inline_equation", + "content": "\\lambda_{2}" + }, + { + "bbox": [ + 313, + 520, + 554, + 567 + ], + "type": "text", + "content": " are weighting coefficients to balance the influence of each term. With the merits of joint training and physical constraints, our SoftShadow framework effectively removes shadows." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 576, + 492, + 587 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 576, + 492, + 587 + ], + "spans": [ + { + "bbox": [ + 313, + 576, + 492, + 587 + ], + "type": "text", + "content": "3.3. Penumbra Formation Constraint" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "content": "We introduce the Penumbra Formation Constraint to enable SAM to predict the soft masks more accurately. The objective of this constraint is to regularize SAM to learn a mask that can reflect the illuminance change of shadow images. To meet the objective, we present two assumptions for our proposed soft mask. For a given shadow image, the soft mask intensity in the penumbra region should gradually transition outward from the shadow center to the outer edge of the shadow. This means that the gradient direction in the penumbra should point from the shadow center toward" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23220" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 145 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 145 + ], + "type": "text", + "content": "the boundary of the shadow. Additionally, since the intensity change in this area is gradual, the gradient magnitude should remain relatively small to reflect a smooth transition. Based on these assumptions, we can formulate the penumbra formation constraint. For a predicted soft mask " + }, + { + "bbox": [ + 55, + 72, + 294, + 145 + ], + "type": "inline_equation", + "content": "\\mathbf{s}" + }, + { + "bbox": [ + 55, + 72, + 294, + 145 + ], + "type": "text", + "content": ", we define the penumbra area using two thresholds " + }, + { + "bbox": [ + 55, + 72, + 294, + 145 + ], + "type": "inline_equation", + "content": "t_1" + }, + { + "bbox": [ + 55, + 72, + 294, + 145 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 72, + 294, + 145 + ], + "type": "inline_equation", + "content": "t_2" + }, + { + "bbox": [ + 55, + 72, + 294, + 145 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 113, + 155, + 295, + 168 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 155, + 295, + 168 + ], + "spans": [ + { + "bbox": [ + 113, + 155, + 295, + 168 + ], + "type": "interline_equation", + "content": "\\mathcal {W} = \\{(i, j) \\mid t _ {1} \\leq \\mathbf {s} _ {i, j} \\leq t _ {2} \\}, \\tag {6}", + "image_path": "d8efe06163b58c3d3940a96b3de40a7cfe52eb427a61b896c627f4765dc93eb1.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 177, + 296, + 250 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 177, + 296, + 250 + ], + "spans": [ + { + "bbox": [ + 55, + 177, + 296, + 250 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 177, + 296, + 250 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 177, + 296, + 250 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 177, + 296, + 250 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 55, + 177, + 296, + 250 + ], + "type": "text", + "content": " denote the spatial coordinates of pixels within the soft mask s. As shown in Figure 4, the center of the shadow area is denoted as c. c is defined as the mean of the " + }, + { + "bbox": [ + 55, + 177, + 296, + 250 + ], + "type": "inline_equation", + "content": "x-" + }, + { + "bbox": [ + 55, + 177, + 296, + 250 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 177, + 296, + 250 + ], + "type": "inline_equation", + "content": "y-" + }, + { + "bbox": [ + 55, + 177, + 296, + 250 + ], + "type": "text", + "content": " coordinates of the penumbra area. The unit vector representing the direction from c to a point w in the penumbra area " + }, + { + "bbox": [ + 55, + 177, + 296, + 250 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 55, + 177, + 296, + 250 + ], + "type": "text", + "content": " can be expressed as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 92, + 259, + 295, + 283 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 92, + 259, + 295, + 283 + ], + "spans": [ + { + "bbox": [ + 92, + 259, + 295, + 283 + ], + "type": "interline_equation", + "content": "d (w) = \\frac {w - c}{\\| w - c \\|}, \\quad \\text {f o r} w \\in \\mathcal {W}, w \\neq c. \\tag {7}", + "image_path": "274c35c776999b4e586cb7b10a5d574ba2b81266105d6415ada336641f6bd1ff.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 292, + 295, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 292, + 295, + 317 + ], + "spans": [ + { + "bbox": [ + 55, + 292, + 295, + 317 + ], + "type": "text", + "content": "Following the assumptions we proposed, we can define the penumbra formation constraint loss " + }, + { + "bbox": [ + 55, + 292, + 295, + 317 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{pen}" + }, + { + "bbox": [ + 55, + 292, + 295, + 317 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 102, + 326, + 295, + 340 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 326, + 295, + 340 + ], + "spans": [ + { + "bbox": [ + 102, + 326, + 295, + 340 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {p e n}} = \\mathbb {E} _ {n, w} \\left[ R (\\mathbf {d} (w) \\cdot \\nabla M (w)) \\right], \\tag {8}", + "image_path": "f3afeb9f843535e97aece9c610aa8fd6a70b880c5068abd2d088ff1dbb67427a.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 349, + 296, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 349, + 296, + 458 + ], + "spans": [ + { + "bbox": [ + 55, + 349, + 296, + 458 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 349, + 296, + 458 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 349, + 296, + 458 + ], + "type": "text", + "content": " is the index of shadow masks, the desired direction of gradient is " + }, + { + "bbox": [ + 55, + 349, + 296, + 458 + ], + "type": "inline_equation", + "content": "\\mathbf{d}(w)" + }, + { + "bbox": [ + 55, + 349, + 296, + 458 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 349, + 296, + 458 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 55, + 349, + 296, + 458 + ], + "type": "text", + "content": " indicate the dot product. " + }, + { + "bbox": [ + 55, + 349, + 296, + 458 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 55, + 349, + 296, + 458 + ], + "type": "text", + "content": " refers to the ReLU function [25]. The ReLU function is applied to filter out gradients conflict with the desired direction. The constraint loss regularizes the magnitude of gradients in alignment with the desired direction, promoting a smooth intensity transition within the penumbra region. By penalizing quick changes in intensity, this constraint encourages the detector to generate softer shadow masks." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 469, + 137, + 482 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 469, + 137, + 482 + ], + "spans": [ + { + "bbox": [ + 55, + 469, + 137, + 482 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 489, + 176, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 489, + 176, + 502 + ], + "spans": [ + { + "bbox": [ + 55, + 489, + 176, + 502 + ], + "type": "text", + "content": "4.1. Experimental Setups" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 506, + 295, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 506, + 295, + 662 + ], + "spans": [ + { + "bbox": [ + 55, + 506, + 295, + 662 + ], + "type": "text", + "content": "Implementation details The shadow removal backbone in our SoftShadow framework can be any shadow removal network. We use ShadowDiffusion [9] as the backbone example for all experiments. We finetune ShadowDiffusion at a resolution of " + }, + { + "bbox": [ + 55, + 506, + 295, + 662 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 55, + 506, + 295, + 662 + ], + "type": "text", + "content": ", followed by previous methods [9]. We empirically set the threshold " + }, + { + "bbox": [ + 55, + 506, + 295, + 662 + ], + "type": "inline_equation", + "content": "t = 0.76" + }, + { + "bbox": [ + 55, + 506, + 295, + 662 + ], + "type": "text", + "content": " to get the ground truth soft shadow mask " + }, + { + "bbox": [ + 55, + 506, + 295, + 662 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_{gt}" + }, + { + "bbox": [ + 55, + 506, + 295, + 662 + ], + "type": "text", + "content": ". We employ the ViT-H model as the backbone for SAM [19]. A set of LoRA [13] layers with a rank of 8 is added to the self-attention blocks in the image encoder of SAM. We use Adam [18] optimizer. The training batch size is 16. We set " + }, + { + "bbox": [ + 55, + 506, + 295, + 662 + ], + "type": "inline_equation", + "content": "\\lambda_{1} = 0.1" + }, + { + "bbox": [ + 55, + 506, + 295, + 662 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 506, + 295, + 662 + ], + "type": "inline_equation", + "content": "\\lambda_{2} = 1" + }, + { + "bbox": [ + 55, + 506, + 295, + 662 + ], + "type": "text", + "content": ". For more training details, please refer to the supplementary." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "Benchmark datasets We work with four benchmark datasets for the various shadow removal experiments. SRD Dataset [27] consists of 2,680 training pairs and 408 testing pairs of shadow and shadow-free images. Notably, the SRD" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 72, + 553, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 228 + ], + "type": "text", + "content": "dataset does not provide shadow masks, the previous methods using SRD commonly using masks detected by DHAN [3] methods. LRSS Dataset [7] is a specifically designed soft shadow dataset that includes 137 images. We select 48 paired shadow and shadow-free images as our testing set. UIUC Dataset [11] contains 76 pairs of images for testing. It features a variety of shadow types [17], including soft, hard, and self shadows, which provides a diverse and challenging set of conditions for evaluating shadow removal methods. " + }, + { + "bbox": [ + 313, + 72, + 553, + 228 + ], + "type": "inline_equation", + "content": "\\mathrm{ISTD + }" + }, + { + "bbox": [ + 313, + 72, + 553, + 228 + ], + "type": "text", + "content": " Dataset [21] is adjusted ISTD dataset [31], consists of 1330 training pairs and 540 testing pairs of shadow and shadow-free images. " + }, + { + "bbox": [ + 313, + 72, + 553, + 228 + ], + "type": "inline_equation", + "content": "\\mathrm{ISTD + }" + }, + { + "bbox": [ + 313, + 72, + 553, + 228 + ], + "type": "text", + "content": " has manually annotated binary masks." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 231, + 553, + 315 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 231, + 553, + 315 + ], + "spans": [ + { + "bbox": [ + 313, + 231, + 553, + 315 + ], + "type": "text", + "content": "Evaluation metrics Following previous works [3, 9, 34], we employ the Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index (SSIM) [32], and Mean Absolute Error (MAE) [33] as quantitative evaluation metrics. We also calculate all metrics for shadow areas, non-shadow areas, and all pixels between ground truth shadow-free images and generated removal results." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 322, + 496, + 335 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 322, + 496, + 335 + ], + "spans": [ + { + "bbox": [ + 313, + 322, + 496, + 335 + ], + "type": "text", + "content": "4.2. Comparison with State-of-the-Art" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 340, + 555, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 340, + 555, + 472 + ], + "spans": [ + { + "bbox": [ + 313, + 340, + 555, + 472 + ], + "type": "text", + "content": "We compare our proposed method with several state-of-the-art shadow removal methods. We include methods that do not require shadow masks as input, e.g., DCshadowNet [16], DeS3 [17], and SAM-helps-shadow [35]. And methods that require shadow masks, including DHAN [3], BMNet [39], SG-ShadowNet [30], ShadowFormer [8], Inpaint4Shadow [22], ShadowDiffusion [9] and Homoformer [34]. To demonstrate the generalizability of our method, we also compare it with zero-shot methods including BCDiff [10] and G2R-ShadowNet [24] on the LRSS [7] and UIUC [11] datasets." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 474, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 474, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 474, + 555, + 714 + ], + "type": "text", + "content": "Quantitative results Table 1 shows the quantitative results of SRD and " + }, + { + "bbox": [ + 313, + 474, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{ISTD + }" + }, + { + "bbox": [ + 313, + 474, + 555, + 714 + ], + "type": "text", + "content": " datasets. The SRD dataset contains many soft shadow images, which is more suitable for validating our methods. Specifically, we outperform all competing methods over all metrics, whether they require input shadow masks or not. Compared with the most recent work without requiring mask input, i.e., DeS3 [17], the PSNR is improved from 34.11 dB to 35.57 dB in whole images. When compared with methods that need input masks, we outperform the SOTA method Homoformer [34], which uses DHAN [3] mask as their input. On the " + }, + { + "bbox": [ + 313, + 474, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{ISTD + }" + }, + { + "bbox": [ + 313, + 474, + 555, + 714 + ], + "type": "text", + "content": " dataset, we significantly outperform DeS3 [17], increasing the PSNR from 31.38 dB to 35.00 dB. For methods that rely on the ground truth masks from the " + }, + { + "bbox": [ + 313, + 474, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{ISTD + }" + }, + { + "bbox": [ + 313, + 474, + 555, + 714 + ], + "type": "text", + "content": " dataset, we evaluate them using masks detected by the FDRNet [38] followed by previous methods BMNet [39], to ensure a fair comparison. We outperform all competing methods under the condition of without ground truth manually annotated masks. Besides, we even achieve comparable results against some SOTA methods using ground truth masks." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "23221" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 59, + 70, + 550, + 301 + ], + "blocks": [ + { + "bbox": [ + 59, + 70, + 550, + 301 + ], + "lines": [ + { + "bbox": [ + 59, + 70, + 550, + 301 + ], + "spans": [ + { + "bbox": [ + 59, + 70, + 550, + 301 + ], + "type": "table", + "html": "
MethodsInput Masksshadownon-shadowall
PSNR↑SSIM↑MAE↓PSNR↑SSIM↑MAE↓PSNR↑SSIM↑MAE↓
SRDDHANDHAN33.670.9788.9434.790.9794.8030.510.9495.67
DC-ShadowNetN/A34.000.9757.7035.530.9813.6531.530.9554.65
BMNetDHAN35.050.9816.6136.020.9823.6131.690.9564.46
SG-ShadowNetDHAN33.730.9797.5336.180.9822.9731.160.9524.23
ShadowFormerDHAN35.550.9826.1436.820.9833.5432.460.9574.28
ShadowDiffusionDHAN38.720.9874.9837.780.9853.4434.730.9703.63
Inpaint4ShadowDHAN36.730.9855.7036.700.9853.2733.270.9673.81
DeS3N/A37.910.9865.2737.450.9843.0334.110.9683.56
HomoformalerDHAN38.810.9874.2539.450.9882.8535.370.9723.33
SAM-helps-shadowN/A33.940.9797.4433.850.9813.7430.720.9524.79
oursN/A39.080.9894.3339.360.9922.5835.570.9753.11
ISTD+BMNetGT37.870.9915.6237.510.9852.4533.980.9722.97
ShadowDiffusionGT39.690.9924.9738.890.9872.2835.670.9752.72
HomoFormerGT39.490.9934.7338.750.9842.2335.350.9752.64
DC-ShadowNetN/A31.060.97612.6227.030.9616.8225.030.9267.77
DeS3N/A36.490.9896.5634.700.9723.4031.380.9583.94
BMNetFDRNet--6.1--2.9--3.5
ShadowDiffusionFDRNet40.120.9925.1536.660.9782.7434.080.9683.12
HomoFormerFDRNet38.840.9915.3134.580.9663.1732.410.9533.51
OursN/A40.360.9934.7837.890.9822.4635.000.9722.85
", + "image_path": "492ea28ba0db74cb42401d5daf9d860fde244bcdd13eacab4c55a6de3678ee9f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 58, + 385, + 292, + 451 + ], + "blocks": [ + { + "bbox": [ + 55, + 311, + 555, + 365 + ], + "lines": [ + { + "bbox": [ + 55, + 311, + 555, + 365 + ], + "spans": [ + { + "bbox": [ + 55, + 311, + 555, + 365 + ], + "type": "text", + "content": "Table 1. The quantitative results of shadow removal using our SoftShadow and recent methods on SRD and ISTD+ datasets. The “Input Masks” column shows the different types of input masks used by these methods. “N/A” means the method does not require masks as input. “GT” means the method uses manually annotated ground truth mask as input. “DHAN” means using masks generated by DHAN [3] method. “FDRNet” means using masks generated by FDRNet [38] method. The best and the second results are boldfaced and underlined, respectively." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 385, + 292, + 451 + ], + "lines": [ + { + "bbox": [ + 58, + 385, + 292, + 451 + ], + "spans": [ + { + "bbox": [ + 58, + 385, + 292, + 451 + ], + "type": "table", + "html": "
MethodsLRSSUIUC
PSNR↑SSIM↑MAE↓PSNR↑SSIM↑MAE↓
DC-ShadowNet20.890.90212.5524.850.8499.51
G2R-ShadowNet20.900.9019.9927.560.8587.43
BCDiff22.130.92210.6826.810.8527.96
Ours23.320.9339.7728.850.9036.48
", + "image_path": "d42c196bb59f40702c8d58c1b8c56119d275f8c719236595df0d6caf13c10f4d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 115, + 506, + 237, + 564 + ], + "blocks": [ + { + "bbox": [ + 55, + 460, + 295, + 492 + ], + "lines": [ + { + "bbox": [ + 55, + 460, + 295, + 492 + ], + "spans": [ + { + "bbox": [ + 55, + 460, + 295, + 492 + ], + "type": "text", + "content": "Table 2. The quantitative results of shadow removal using our SoftShadow and recent methods on the LRSS and UIUC datasets. The best results are boldfaced." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 115, + 506, + 237, + 564 + ], + "lines": [ + { + "bbox": [ + 115, + 506, + 237, + 564 + ], + "spans": [ + { + "bbox": [ + 115, + 506, + 237, + 564 + ], + "type": "table", + "html": "
MethodsPSNR↑MAE↓
Inpaint4Shadow39.725.27
DeS339.287.07
HomoFormer41.115.91
Ours41.464.40
", + "image_path": "7c69bd83909777016a60cd580f8a7eee751587af7bd9f22ba982ef391413699b.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "text", + "content": "Table 2 shows the generalizability of our methods. We use our pretrained model on the SRD dataset and test it on LRSS and UIUC datasets without further training. The DeS3 [17] has better results because they train their method on the LRSS training set. For comparison, we chose DC-ShadowNet as our baseline since this method does not require masks as input. Our results significantly outperform" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 326, + 385, + 543, + 472 + ], + "blocks": [ + { + "bbox": [ + 55, + 573, + 295, + 605 + ], + "lines": [ + { + "bbox": [ + 55, + 573, + 295, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 573, + 295, + 605 + ], + "type": "text", + "content": "Table 3. The quantitative results in the penumbra area. The penumbra area is calculated from our predicted soft masks. The best results are boldfaced." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 326, + 385, + 543, + 472 + ], + "lines": [ + { + "bbox": [ + 326, + 385, + 543, + 472 + ], + "spans": [ + { + "bbox": [ + 326, + 385, + 543, + 472 + ], + "type": "table", + "html": "
Input MasksShadowDiffusionHomoFormerOurs
Ground Truth mask35.6735.3535.00
Otsu mask35.6233.47
BDRAR mask33.4132.96
FDRNet mask34.0832.41
Pretrained SAM mask28.4528.16
Mean33.4532.47N/A
Std Dev0.9120.993N/A
", + "image_path": "2a272af29df89caa991f8d90dd30ff1f5ed388ea16b0023b6d28683e234c9efb.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 480, + 555, + 514 + ], + "lines": [ + { + "bbox": [ + 313, + 480, + 555, + 514 + ], + "spans": [ + { + "bbox": [ + 313, + 480, + 555, + 514 + ], + "type": "text", + "content": "Table 4. The mask sensitivity evaluation on " + }, + { + "bbox": [ + 313, + 480, + 555, + 514 + ], + "type": "inline_equation", + "content": "\\mathrm{ISTD + }" + }, + { + "bbox": [ + 313, + 480, + 555, + 514 + ], + "type": "text", + "content": " dataset. The values in the table are PSNR results. The best results are boldfaced. The results that are higher than ours are underlined." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 540, + 555, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 540, + 555, + 624 + ], + "spans": [ + { + "bbox": [ + 313, + 540, + 555, + 624 + ], + "type": "text", + "content": "DC-ShadowNet on LRSS and UIUC datasets. Additionally, we compare our methods with two methods, which are designed to have better generalizability. The results demonstrate that our method has the best PSNR/SSIM/MAE among the comparison methods. These results highlight the robustness and adaptability of our model, even on datasets it was not specifically trained on." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "content": "Qualitative evaluation To further demonstrate the effectiveness of our methods compared to other competing methods, Figure 5 illustrates our improvements in removing soft shadows on the SRD dataset compared with methods using hard shadow masks. The soft masks typically appear as soft boundaries of large object shadows or in small shadow regions. Specifically, our samples demonstrate that we ef" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "23222" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 109, + 69, + 164, + 144 + ], + "blocks": [ + { + "bbox": [ + 109, + 69, + 164, + 144 + ], + "lines": [ + { + "bbox": [ + 109, + 69, + 164, + 144 + ], + "spans": [ + { + "bbox": [ + 109, + 69, + 164, + 144 + ], + "type": "image", + "image_path": "9a74b8ed5c21e251063c875166a536274446f667ede0b7c8eef6c41d21842dc5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 109, + 144, + 164, + 216 + ], + "blocks": [ + { + "bbox": [ + 109, + 144, + 164, + 216 + ], + "lines": [ + { + "bbox": [ + 109, + 144, + 164, + 216 + ], + "spans": [ + { + "bbox": [ + 109, + 144, + 164, + 216 + ], + "type": "image", + "image_path": "4ae1f68b9b0e380c1c774c0b52b1804e539b3b51d9bd973ff96c403574c09224.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 108, + 216, + 164, + 271 + ], + "blocks": [ + { + "bbox": [ + 108, + 216, + 164, + 271 + ], + "lines": [ + { + "bbox": [ + 108, + 216, + 164, + 271 + ], + "spans": [ + { + "bbox": [ + 108, + 216, + 164, + 271 + ], + "type": "image", + "image_path": "e7e2c867b2b114ead37e0962bcb75534115f31d777f39d9843563a9c77bd3cc9.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 127, + 271, + 144, + 279 + ], + "lines": [ + { + "bbox": [ + 127, + 271, + 144, + 279 + ], + "spans": [ + { + "bbox": [ + 127, + 271, + 144, + 279 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 164, + 271, + 206, + 279 + ], + "lines": [ + { + "bbox": [ + 164, + 271, + 206, + 279 + ], + "spans": [ + { + "bbox": [ + 164, + 271, + 206, + 279 + ], + "type": "text", + "content": "(a) BMNet" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 164, + 70, + 219, + 144 + ], + "blocks": [ + { + "bbox": [ + 164, + 70, + 219, + 144 + ], + "lines": [ + { + "bbox": [ + 164, + 70, + 219, + 144 + ], + "spans": [ + { + "bbox": [ + 164, + 70, + 219, + 144 + ], + "type": "image", + "image_path": "e575ad387457e11efe562115da75df95a81a956b96fe2ac2ca61beeb8dc62cd3.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 164, + 144, + 219, + 216 + ], + "blocks": [ + { + "bbox": [ + 164, + 144, + 219, + 216 + ], + "lines": [ + { + "bbox": [ + 164, + 144, + 219, + 216 + ], + "spans": [ + { + "bbox": [ + 164, + 144, + 219, + 216 + ], + "type": "image", + "image_path": "fc16a0c3965d03e50257596a09f0d09717d2423ff0ae732fa05396ae35fee6e3.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 164, + 216, + 219, + 271 + ], + "blocks": [ + { + "bbox": [ + 164, + 216, + 219, + 271 + ], + "lines": [ + { + "bbox": [ + 164, + 216, + 219, + 271 + ], + "spans": [ + { + "bbox": [ + 164, + 216, + 219, + 271 + ], + "type": "image", + "image_path": "57b52f6e05ec73099823e490b0661e20bb67acc979b1c043749f19d7eb1e084e.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 219, + 271, + 276, + 279 + ], + "lines": [ + { + "bbox": [ + 219, + 271, + 276, + 279 + ], + "spans": [ + { + "bbox": [ + 219, + 271, + 276, + 279 + ], + "type": "text", + "content": "(b) ShadowDiffusion" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 220, + 70, + 276, + 144 + ], + "blocks": [ + { + "bbox": [ + 220, + 70, + 276, + 144 + ], + "lines": [ + { + "bbox": [ + 220, + 70, + 276, + 144 + ], + "spans": [ + { + "bbox": [ + 220, + 70, + 276, + 144 + ], + "type": "image", + "image_path": "529f1391825f1f6d39718cd009e934de4e15188677d0026b4c183f8e5b3c8d6d.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 220, + 144, + 276, + 216 + ], + "blocks": [ + { + "bbox": [ + 220, + 144, + 276, + 216 + ], + "lines": [ + { + "bbox": [ + 220, + 144, + 276, + 216 + ], + "spans": [ + { + "bbox": [ + 220, + 144, + 276, + 216 + ], + "type": "image", + "image_path": "7e911b2b9029b15510d39a1aafa0441c0b91b026604fdc69f3874fd07b6a1918.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 220, + 216, + 276, + 271 + ], + "blocks": [ + { + "bbox": [ + 220, + 216, + 276, + 271 + ], + "lines": [ + { + "bbox": [ + 220, + 216, + 276, + 271 + ], + "spans": [ + { + "bbox": [ + 220, + 216, + 276, + 271 + ], + "type": "image", + "image_path": "056c92d0d7e004e1b2290ec47a0e81062cb41c72408d0f90b164db0075c916a3.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 277, + 70, + 332, + 144 + ], + "blocks": [ + { + "bbox": [ + 277, + 70, + 332, + 144 + ], + "lines": [ + { + "bbox": [ + 277, + 70, + 332, + 144 + ], + "spans": [ + { + "bbox": [ + 277, + 70, + 332, + 144 + ], + "type": "image", + "image_path": "cf590b81a393205a5637b3393852bd9df8622fa72f9d7e0d52a7cce6be987dd2.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 277, + 144, + 332, + 216 + ], + "blocks": [ + { + "bbox": [ + 277, + 144, + 332, + 216 + ], + "lines": [ + { + "bbox": [ + 277, + 144, + 332, + 216 + ], + "spans": [ + { + "bbox": [ + 277, + 144, + 332, + 216 + ], + "type": "image", + "image_path": "6edbe168e397965a43031b6933d332fadb3e51e28fa62ec364754f4324b18165.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 277, + 216, + 332, + 271 + ], + "blocks": [ + { + "bbox": [ + 277, + 216, + 332, + 271 + ], + "lines": [ + { + "bbox": [ + 277, + 216, + 332, + 271 + ], + "spans": [ + { + "bbox": [ + 277, + 216, + 332, + 271 + ], + "type": "image", + "image_path": "13fd6f96465afec4915ba9e31f2dd40ce82d1143851380545da3c3b21c44ac9f.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 277, + 271, + 332, + 279 + ], + "lines": [ + { + "bbox": [ + 277, + 271, + 332, + 279 + ], + "spans": [ + { + "bbox": [ + 277, + 271, + 332, + 279 + ], + "type": "text", + "content": "(c) Inpaint4Shadow" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 334, + 70, + 389, + 144 + ], + "blocks": [ + { + "bbox": [ + 334, + 70, + 389, + 144 + ], + "lines": [ + { + "bbox": [ + 334, + 70, + 389, + 144 + ], + "spans": [ + { + "bbox": [ + 334, + 70, + 389, + 144 + ], + "type": "image", + "image_path": "1d40958a37c6a584ee4f263dc0872abd9d2a9658b07265e7960d37671e5a5cdb.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 334, + 144, + 389, + 216 + ], + "blocks": [ + { + "bbox": [ + 334, + 144, + 389, + 216 + ], + "lines": [ + { + "bbox": [ + 334, + 144, + 389, + 216 + ], + "spans": [ + { + "bbox": [ + 334, + 144, + 389, + 216 + ], + "type": "image", + "image_path": "459b1a8aa28b3366ad5828cefa22127f9c9d05c7812eabd3a032dc913b7420ab.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 334, + 216, + 389, + 271 + ], + "blocks": [ + { + "bbox": [ + 334, + 216, + 389, + 271 + ], + "lines": [ + { + "bbox": [ + 334, + 216, + 389, + 271 + ], + "spans": [ + { + "bbox": [ + 334, + 216, + 389, + 271 + ], + "type": "image", + "image_path": "e6a4494ad333be3b39be215ae36d0f255335aaff437f65aa1d3849a843ce0ff4.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 338, + 271, + 386, + 279 + ], + "lines": [ + { + "bbox": [ + 338, + 271, + 386, + 279 + ], + "spans": [ + { + "bbox": [ + 338, + 271, + 386, + 279 + ], + "type": "text", + "content": "(d) HomoFormer" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 390, + 70, + 444, + 144 + ], + "blocks": [ + { + "bbox": [ + 390, + 70, + 444, + 144 + ], + "lines": [ + { + "bbox": [ + 390, + 70, + 444, + 144 + ], + "spans": [ + { + "bbox": [ + 390, + 70, + 444, + 144 + ], + "type": "image", + "image_path": "ed245a7606c296b7c73c7525dc5c49e864cba5105d5fa2bed1dd095f463f7295.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 390, + 144, + 444, + 216 + ], + "blocks": [ + { + "bbox": [ + 390, + 144, + 444, + 216 + ], + "lines": [ + { + "bbox": [ + 390, + 144, + 444, + 216 + ], + "spans": [ + { + "bbox": [ + 390, + 144, + 444, + 216 + ], + "type": "image", + "image_path": "72440b530bb4b8f65de843da409944f5e9a119ee0f3909f9625632b0bfed7a5f.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 390, + 216, + 444, + 271 + ], + "blocks": [ + { + "bbox": [ + 390, + 216, + 444, + 271 + ], + "lines": [ + { + "bbox": [ + 390, + 216, + 444, + 271 + ], + "spans": [ + { + "bbox": [ + 390, + 216, + 444, + 271 + ], + "type": "image", + "image_path": "9e42a12a120aa18f5222cab3584254c72bd63da378c8e4265f85d036e6772998.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 405, + 271, + 429, + 279 + ], + "lines": [ + { + "bbox": [ + 405, + 271, + 429, + 279 + ], + "spans": [ + { + "bbox": [ + 405, + 271, + 429, + 279 + ], + "type": "text", + "content": "(e) Ours" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 446, + 70, + 501, + 144 + ], + "blocks": [ + { + "bbox": [ + 446, + 70, + 501, + 144 + ], + "lines": [ + { + "bbox": [ + 446, + 70, + 501, + 144 + ], + "spans": [ + { + "bbox": [ + 446, + 70, + 501, + 144 + ], + "type": "image", + "image_path": "5141c770ff3554eddeab19a1f466d9df774639bf548dd0d4e3349e56db5f2f4d.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 446, + 144, + 501, + 216 + ], + "blocks": [ + { + "bbox": [ + 446, + 144, + 501, + 216 + ], + "lines": [ + { + "bbox": [ + 446, + 144, + 501, + 216 + ], + "spans": [ + { + "bbox": [ + 446, + 144, + 501, + 216 + ], + "type": "image", + "image_path": "08272ac56764c3fb2c046ccf6ed090105dc5a14edb22aab50eb704980ea51dc5.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 446, + 216, + 501, + 271 + ], + "blocks": [ + { + "bbox": [ + 446, + 216, + 501, + 271 + ], + "lines": [ + { + "bbox": [ + 446, + 216, + 501, + 271 + ], + "spans": [ + { + "bbox": [ + 446, + 216, + 501, + 271 + ], + "type": "image", + "image_path": "f9ff74e892c06c8891f0529b75976c0362cf9082e7b5d6afcbd09b3f92369bea.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 468, + 271, + 478, + 279 + ], + "lines": [ + { + "bbox": [ + 468, + 271, + 478, + 279 + ], + "spans": [ + { + "bbox": [ + 468, + 271, + 478, + 279 + ], + "type": "text", + "content": "GT" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 63, + 344, + 108, + 390 + ], + "blocks": [ + { + "bbox": [ + 63, + 344, + 108, + 390 + ], + "lines": [ + { + "bbox": [ + 63, + 344, + 108, + 390 + ], + "spans": [ + { + "bbox": [ + 63, + 344, + 108, + 390 + ], + "type": "image", + "image_path": "3fd1b2e5a7cd38217b90cbf52961015d3b2fe0beb570dedef2d75e338a39b83a.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 63, + 390, + 108, + 435 + ], + "blocks": [ + { + "bbox": [ + 63, + 390, + 108, + 435 + ], + "lines": [ + { + "bbox": [ + 63, + 390, + 108, + 435 + ], + "spans": [ + { + "bbox": [ + 63, + 390, + 108, + 435 + ], + "type": "image", + "image_path": "09cc3d2a27aede96a00a9e0f41d50d1f344ad667e4f5754468c26307d792180d.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 63, + 435, + 108, + 479 + ], + "blocks": [ + { + "bbox": [ + 63, + 435, + 108, + 479 + ], + "lines": [ + { + "bbox": [ + 63, + 435, + 108, + 479 + ], + "spans": [ + { + "bbox": [ + 63, + 435, + 108, + 479 + ], + "type": "image", + "image_path": "ce55aef3b2fb98669adf97dfe3bb549b4e087e32ab4cc8d2006271596edcc4e5.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 80, + 480, + 96, + 487 + ], + "lines": [ + { + "bbox": [ + 80, + 480, + 96, + 487 + ], + "spans": [ + { + "bbox": [ + 80, + 480, + 96, + 487 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 501, + 297, + 545 + ], + "lines": [ + { + "bbox": [ + 55, + 501, + 297, + 545 + ], + "spans": [ + { + "bbox": [ + 55, + 501, + 297, + 545 + ], + "type": "text", + "content": "Figure 6. Examples of soft shadow removal results on SRD datasets [27]. The input shadow image, the estimated results of (a) DC-ShadowNet [16], (b) DeS3 [17], and (c) Ours, as well as the ground truth image, respectively." + } + ] + } + ], + "index": 49, + "angle": 0, + "type": "image_caption" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 108, + 344, + 154, + 390 + ], + "blocks": [ + { + "bbox": [ + 108, + 344, + 154, + 390 + ], + "lines": [ + { + "bbox": [ + 108, + 344, + 154, + 390 + ], + "spans": [ + { + "bbox": [ + 108, + 344, + 154, + 390 + ], + "type": "image", + "image_path": "d05f4dffa507181f3f2cd5feb3333f44649e4f61e78c0411246376695ec13ee5.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + } + ], + "index": 33 + }, + { + "type": "image", + "bbox": [ + 108, + 390, + 154, + 435 + ], + "blocks": [ + { + "bbox": [ + 108, + 390, + 154, + 435 + ], + "lines": [ + { + "bbox": [ + 108, + 390, + 154, + 435 + ], + "spans": [ + { + "bbox": [ + 108, + 390, + 154, + 435 + ], + "type": "image", + "image_path": "b7a4d945da8b8f191a57f9faf9116331acd8d9572a1d7759b1bf42eaf78ccedc.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 108, + 435, + 154, + 479 + ], + "blocks": [ + { + "bbox": [ + 108, + 435, + 154, + 479 + ], + "lines": [ + { + "bbox": [ + 108, + 435, + 154, + 479 + ], + "spans": [ + { + "bbox": [ + 108, + 435, + 154, + 479 + ], + "type": "image", + "image_path": "4f920aa12fb62068fa8510d35f65b87148c04bffa6112e722d62043b53fd7eeb.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 121, + 480, + 141, + 487 + ], + "lines": [ + { + "bbox": [ + 121, + 480, + 141, + 487 + ], + "spans": [ + { + "bbox": [ + 121, + 480, + 141, + 487 + ], + "type": "text", + "content": "(a) DC" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_caption" + } + ], + "index": 35 + }, + { + "type": "image", + "bbox": [ + 154, + 344, + 199, + 390 + ], + "blocks": [ + { + "bbox": [ + 154, + 344, + 199, + 390 + ], + "lines": [ + { + "bbox": [ + 154, + 344, + 199, + 390 + ], + "spans": [ + { + "bbox": [ + 154, + 344, + 199, + 390 + ], + "type": "image", + "image_path": "cd27d87ca9d83edfde954fd64543008beaefded5ed76be16b0e38ee3a9744f8d.jpg" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_body" + } + ], + "index": 37 + }, + { + "type": "image", + "bbox": [ + 154, + 390, + 198, + 435 + ], + "blocks": [ + { + "bbox": [ + 154, + 390, + 198, + 435 + ], + "lines": [ + { + "bbox": [ + 154, + 390, + 198, + 435 + ], + "spans": [ + { + "bbox": [ + 154, + 390, + 198, + 435 + ], + "type": "image", + "image_path": "183ba488d3db01d337b01eeafc8fe024a3cca07f2047d390b10a3f07980ab6e5.jpg" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_body" + } + ], + "index": 38 + }, + { + "type": "image", + "bbox": [ + 154, + 435, + 198, + 479 + ], + "blocks": [ + { + "bbox": [ + 154, + 435, + 198, + 479 + ], + "lines": [ + { + "bbox": [ + 154, + 435, + 198, + 479 + ], + "spans": [ + { + "bbox": [ + 154, + 435, + 198, + 479 + ], + "type": "image", + "image_path": "0ef123f289c40e7ed6e76f1ac9a29d768cdb8e5570cbc42bb85deeabc35ed830.jpg" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 163, + 480, + 188, + 487 + ], + "lines": [ + { + "bbox": [ + 163, + 480, + 188, + 487 + ], + "spans": [ + { + "bbox": [ + 163, + 480, + 188, + 487 + ], + "type": "text", + "content": "(b) DeS3" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_caption" + } + ], + "index": 39 + }, + { + "type": "image", + "bbox": [ + 199, + 344, + 244, + 390 + ], + "blocks": [ + { + "bbox": [ + 199, + 344, + 244, + 390 + ], + "lines": [ + { + "bbox": [ + 199, + 344, + 244, + 390 + ], + "spans": [ + { + "bbox": [ + 199, + 344, + 244, + 390 + ], + "type": "image", + "image_path": "83920cde1fa370e4591a94a34f255bf592f47c7f0783d48ee1a4fc7744d8a8ee.jpg" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_body" + } + ], + "index": 41 + }, + { + "type": "image", + "bbox": [ + 199, + 390, + 244, + 435 + ], + "blocks": [ + { + "bbox": [ + 199, + 390, + 244, + 435 + ], + "lines": [ + { + "bbox": [ + 199, + 390, + 244, + 435 + ], + "spans": [ + { + "bbox": [ + 199, + 390, + 244, + 435 + ], + "type": "image", + "image_path": "97dcd1ae4caad6488054aa445701404dae08ba68fba295bc45a8f9a63bf242f6.jpg" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_body" + } + ], + "index": 42 + }, + { + "type": "image", + "bbox": [ + 199, + 435, + 243, + 479 + ], + "blocks": [ + { + "bbox": [ + 199, + 435, + 243, + 479 + ], + "lines": [ + { + "bbox": [ + 199, + 435, + 243, + 479 + ], + "spans": [ + { + "bbox": [ + 199, + 435, + 243, + 479 + ], + "type": "image", + "image_path": "ada482ff5526b4cf5a0797c902111fab036fbaaad21a1a9180c9e76290fa9025.jpg" + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 211, + 480, + 232, + 487 + ], + "lines": [ + { + "bbox": [ + 211, + 480, + 232, + 487 + ], + "spans": [ + { + "bbox": [ + 211, + 480, + 232, + 487 + ], + "type": "text", + "content": "(b) ours" + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "image_caption" + } + ], + "index": 43 + }, + { + "type": "image", + "bbox": [ + 244, + 344, + 288, + 390 + ], + "blocks": [ + { + "bbox": [ + 244, + 344, + 288, + 390 + ], + "lines": [ + { + "bbox": [ + 244, + 344, + 288, + 390 + ], + "spans": [ + { + "bbox": [ + 244, + 344, + 288, + 390 + ], + "type": "image", + "image_path": "52ebee6b92b1ffe614a77e072359a56d1f3d7cdbeca1c9f8419c2d4b26daa695.jpg" + } + ] + } + ], + "index": 45, + "angle": 0, + "type": "image_body" + } + ], + "index": 45 + }, + { + "type": "image", + "bbox": [ + 244, + 390, + 288, + 435 + ], + "blocks": [ + { + "bbox": [ + 244, + 390, + 288, + 435 + ], + "lines": [ + { + "bbox": [ + 244, + 390, + 288, + 435 + ], + "spans": [ + { + "bbox": [ + 244, + 390, + 288, + 435 + ], + "type": "image", + "image_path": "f2f4a69c00b2cf11296c640b09abeefccdcca229bcfaa23b4ec0eb3b9eb74b03.jpg" + } + ] + } + ], + "index": 46, + "angle": 0, + "type": "image_body" + } + ], + "index": 46 + }, + { + "type": "image", + "bbox": [ + 244, + 435, + 288, + 479 + ], + "blocks": [ + { + "bbox": [ + 244, + 435, + 288, + 479 + ], + "lines": [ + { + "bbox": [ + 244, + 435, + 288, + 479 + ], + "spans": [ + { + "bbox": [ + 244, + 435, + 288, + 479 + ], + "type": "image", + "image_path": "17ac907c08fbad2e68636dc2fda0d0b684601351bac37a7701cf22c2fd398ac8.jpg" + } + ] + } + ], + "index": 47, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 261, + 480, + 271, + 487 + ], + "lines": [ + { + "bbox": [ + 261, + 480, + 271, + 487 + ], + "spans": [ + { + "bbox": [ + 261, + 480, + 271, + 487 + ], + "type": "text", + "content": "GT" + } + ] + } + ], + "index": 48, + "angle": 0, + "type": "image_caption" + } + ], + "index": 47 + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "content": "fectively remove both types of soft shadows while preserving the illumination transition in penumbra areas, without introducing obvious artifacts. Figure 6 shows our improvement in shadow area accuracy compared to other end-to-end methods. Our approach effectively removes all parts of the shadow shown in the first row of the example. In the second and third rows, we successfully remove shadows in regions with complex textures. Figure 7 highlights our method's generalizability on the LRSS dataset, achieving improved shadow removal performance without additional training on LRSS dataset. These results confirm our method's adaptability and precision in diverse shadowed scenarios." + } + ] + } + ], + "index": 50 + }, + { + "type": "image", + "bbox": [ + 322, + 344, + 359, + 381 + ], + "blocks": [ + { + "bbox": [ + 322, + 344, + 359, + 381 + ], + "lines": [ + { + "bbox": [ + 322, + 344, + 359, + 381 + ], + "spans": [ + { + "bbox": [ + 322, + 344, + 359, + 381 + ], + "type": "image", + "image_path": "34601ae1269a778ced2e6f4c9335434e5748d9beb04c4d84d603d7f790d0c20f.jpg" + } + ] + } + ], + "index": 51, + "angle": 0, + "type": "image_body" + } + ], + "index": 51 + }, + { + "type": "image", + "bbox": [ + 323, + 381, + 359, + 419 + ], + "blocks": [ + { + "bbox": [ + 323, + 381, + 359, + 419 + ], + "lines": [ + { + "bbox": [ + 323, + 381, + 359, + 419 + ], + "spans": [ + { + "bbox": [ + 323, + 381, + 359, + 419 + ], + "type": "image", + "image_path": "50725bc0220d96e40799bafd318b20e16018b24d968a9e83c9f590ea6d0360cb.jpg" + } + ] + } + ], + "index": 52, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 332, + 420, + 350, + 429 + ], + "lines": [ + { + "bbox": [ + 332, + 420, + 350, + 429 + ], + "spans": [ + { + "bbox": [ + 332, + 420, + 350, + 429 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 53, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 442, + 555, + 487 + ], + "lines": [ + { + "bbox": [ + 313, + 442, + 555, + 487 + ], + "spans": [ + { + "bbox": [ + 313, + 442, + 555, + 487 + ], + "type": "text", + "content": "Figure 7. Examples of shadow removal results on the LRSS dataset [7]. The input shadow image, the estimated results of (a) DC-ShadowNet [16], (b) G2R-ShadowNet [24], (c) Ours, our predicted soft shadow masks and ground truth image, respectively." + } + ] + } + ], + "index": 69, + "angle": 0, + "type": "image_caption" + } + ], + "index": 52 + }, + { + "type": "image", + "bbox": [ + 361, + 344, + 397, + 381 + ], + "blocks": [ + { + "bbox": [ + 361, + 344, + 397, + 381 + ], + "lines": [ + { + "bbox": [ + 361, + 344, + 397, + 381 + ], + "spans": [ + { + "bbox": [ + 361, + 344, + 397, + 381 + ], + "type": "image", + "image_path": "8a6eb088abe55880ad23cf6b0afb2e27a90a5c9c00b2fda374936c742d9dfc14.jpg" + } + ] + } + ], + "index": 54, + "angle": 0, + "type": "image_body" + } + ], + "index": 54 + }, + { + "type": "image", + "bbox": [ + 361, + 381, + 397, + 419 + ], + "blocks": [ + { + "bbox": [ + 361, + 381, + 397, + 419 + ], + "lines": [ + { + "bbox": [ + 361, + 381, + 397, + 419 + ], + "spans": [ + { + "bbox": [ + 361, + 381, + 397, + 419 + ], + "type": "image", + "image_path": "2758aa530fceb31000780249811fc1a0c8b908e0684c2673c5456900336bc1e9.jpg" + } + ] + } + ], + "index": 55, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 367, + 420, + 389, + 429 + ], + "lines": [ + { + "bbox": [ + 367, + 420, + 389, + 429 + ], + "spans": [ + { + "bbox": [ + 367, + 420, + 389, + 429 + ], + "type": "text", + "content": "(a) DC" + } + ] + } + ], + "index": 56, + "angle": 0, + "type": "image_caption" + } + ], + "index": 55 + }, + { + "type": "image", + "bbox": [ + 398, + 344, + 435, + 381 + ], + "blocks": [ + { + "bbox": [ + 398, + 344, + 435, + 381 + ], + "lines": [ + { + "bbox": [ + 398, + 344, + 435, + 381 + ], + "spans": [ + { + "bbox": [ + 398, + 344, + 435, + 381 + ], + "type": "image", + "image_path": "6ba136616495a6b8c608e35a9f625c9d09a4694e48adff683a178722cdd2cfdb.jpg" + } + ] + } + ], + "index": 57, + "angle": 0, + "type": "image_body" + } + ], + "index": 57 + }, + { + "type": "image", + "bbox": [ + 398, + 381, + 435, + 419 + ], + "blocks": [ + { + "bbox": [ + 398, + 381, + 435, + 419 + ], + "lines": [ + { + "bbox": [ + 398, + 381, + 435, + 419 + ], + "spans": [ + { + "bbox": [ + 398, + 381, + 435, + 419 + ], + "type": "image", + "image_path": "048caf019344bd26104fc2e1f27624bf59dbdb344305644693552a0f4013e598.jpg" + } + ] + } + ], + "index": 58, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 403, + 420, + 430, + 428 + ], + "lines": [ + { + "bbox": [ + 403, + 420, + 430, + 428 + ], + "spans": [ + { + "bbox": [ + 403, + 420, + 430, + 428 + ], + "type": "text", + "content": "(b) G2R" + } + ] + } + ], + "index": 59, + "angle": 0, + "type": "image_caption" + } + ], + "index": 58 + }, + { + "type": "image", + "bbox": [ + 435, + 344, + 471, + 381 + ], + "blocks": [ + { + "bbox": [ + 435, + 344, + 471, + 381 + ], + "lines": [ + { + "bbox": [ + 435, + 344, + 471, + 381 + ], + "spans": [ + { + "bbox": [ + 435, + 344, + 471, + 381 + ], + "type": "image", + "image_path": "fc5d06e628be907e6ee1552773a3c2cc3a305d3f744ab1d3d77b3651fe84a79f.jpg" + } + ] + } + ], + "index": 60, + "angle": 0, + "type": "image_body" + } + ], + "index": 60 + }, + { + "type": "image", + "bbox": [ + 435, + 381, + 471, + 419 + ], + "blocks": [ + { + "bbox": [ + 435, + 381, + 471, + 419 + ], + "lines": [ + { + "bbox": [ + 435, + 381, + 471, + 419 + ], + "spans": [ + { + "bbox": [ + 435, + 381, + 471, + 419 + ], + "type": "image", + "image_path": "fb86e7d0d2e2e753cfd15411f59ab7b6d0515d8df9c8d2236a1079190104ad31.jpg" + } + ] + } + ], + "index": 61, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 439, + 420, + 465, + 428 + ], + "lines": [ + { + "bbox": [ + 439, + 420, + 465, + 428 + ], + "spans": [ + { + "bbox": [ + 439, + 420, + 465, + 428 + ], + "type": "text", + "content": "(c) Ours" + } + ] + } + ], + "index": 62, + "angle": 0, + "type": "image_caption" + } + ], + "index": 61 + }, + { + "type": "image", + "bbox": [ + 472, + 344, + 509, + 381 + ], + "blocks": [ + { + "bbox": [ + 55, + 292, + 555, + 326 + ], + "lines": [ + { + "bbox": [ + 55, + 292, + 555, + 326 + ], + "spans": [ + { + "bbox": [ + 55, + 292, + 555, + 326 + ], + "type": "text", + "content": "Figure 5. Examples of soft shadow image removal results on the SRD dataset [27]. The input shadow image, the estimated results of (a) BMNet [39], (b) ShadowDiffusion [9], (c) Inpaint4Shadow [22], (d) Homoformer [34], and (e) Ours, as well as the ground truth (GT) image, respectively." + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 472, + 344, + 509, + 381 + ], + "lines": [ + { + "bbox": [ + 472, + 344, + 509, + 381 + ], + "spans": [ + { + "bbox": [ + 472, + 344, + 509, + 381 + ], + "type": "image", + "image_path": "db08484af34986ca44063d54c1d2b59b8deb2fab4f55f2f6d0c56d0ab3bf3210.jpg" + } + ] + } + ], + "index": 63, + "angle": 0, + "type": "image_body" + } + ], + "index": 63 + }, + { + "type": "image", + "bbox": [ + 472, + 381, + 509, + 419 + ], + "blocks": [ + { + "bbox": [ + 472, + 381, + 509, + 419 + ], + "lines": [ + { + "bbox": [ + 472, + 381, + 509, + 419 + ], + "spans": [ + { + "bbox": [ + 472, + 381, + 509, + 419 + ], + "type": "image", + "image_path": "1a8b5e858811df78da51bd569258bc21a48aca2dad1d9fd943e715645d0d4454.jpg" + } + ] + } + ], + "index": 64, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 474, + 420, + 509, + 428 + ], + "lines": [ + { + "bbox": [ + 474, + 420, + 509, + 428 + ], + "spans": [ + { + "bbox": [ + 474, + 420, + 509, + 428 + ], + "type": "text", + "content": "Soft Masks" + } + ] + } + ], + "index": 65, + "angle": 0, + "type": "image_caption" + } + ], + "index": 64 + }, + { + "type": "image", + "bbox": [ + 509, + 344, + 547, + 381 + ], + "blocks": [ + { + "bbox": [ + 509, + 344, + 547, + 381 + ], + "lines": [ + { + "bbox": [ + 509, + 344, + 547, + 381 + ], + "spans": [ + { + "bbox": [ + 509, + 344, + 547, + 381 + ], + "type": "image", + "image_path": "208f2c9836e272f1594293a4842aae10c45fd0d579a75c5c18ea2eefde562290.jpg" + } + ] + } + ], + "index": 66, + "angle": 0, + "type": "image_body" + } + ], + "index": 66 + }, + { + "type": "image", + "bbox": [ + 509, + 381, + 547, + 419 + ], + "blocks": [ + { + "bbox": [ + 509, + 381, + 547, + 419 + ], + "lines": [ + { + "bbox": [ + 509, + 381, + 547, + 419 + ], + "spans": [ + { + "bbox": [ + 509, + 381, + 547, + 419 + ], + "type": "image", + "image_path": "fb4ff310f65fe815e4e398de3f66501b786de0cb06f521f84cdbdefb8c964159.jpg" + } + ] + } + ], + "index": 67, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 523, + 420, + 533, + 428 + ], + "lines": [ + { + "bbox": [ + 523, + 420, + 533, + 428 + ], + "spans": [ + { + "bbox": [ + 523, + 420, + 533, + 428 + ], + "type": "text", + "content": "GT" + } + ] + } + ], + "index": 68, + "angle": 0, + "type": "image_caption" + } + ], + "index": 67 + }, + { + "bbox": [ + 313, + 510, + 436, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 436, + 525 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 436, + 525 + ], + "type": "text", + "content": "5. Framework Analysis" + } + ] + } + ], + "index": 70 + }, + { + "bbox": [ + 313, + 533, + 555, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 555, + 592 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 555, + 592 + ], + "type": "text", + "content": "In this section, we provide more experiments to evaluate the effectiveness of our proposed SoftShadow framework. We discuss the performance on the penumbra area, the mask sensitive of previous methods, and we discuss the effectiveness of our provided losses in ablation study." + } + ] + } + ], + "index": 71 + }, + { + "bbox": [ + 313, + 594, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 556, + 713 + ], + "type": "text", + "content": "Penumbra effect evaluation The penumbra region typically occupies only a small portion of the entire image, so improvements within this area have a limited impact on overall image metrics. To further assess the improvements introduced by using soft masks in the penumbra region, we calculate the PSNR and MAE values specifically within this area on the SRD dataset. For a fair comparison, we define the penumbra region following the methods from AEF [5], applying ground truth soft masks " + }, + { + "bbox": [ + 313, + 594, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_{gt}" + }, + { + "bbox": [ + 313, + 594, + 556, + 713 + ], + "type": "text", + "content": ". The results are shown in Table 3. Compared with the SOTA methods like Homo" + } + ] + } + ], + "index": 72 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23223" + } + ] + } + ], + "index": 73 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 120 + ], + "type": "text", + "content": "Former [34], DeS3 [17] and Inpaint4Shadow [22]. We improve PSNR and MAE values in the penumbra area. These results show that on images containing soft shadows, our methods has better restoration performances." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 123, + 295, + 219 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 123, + 295, + 219 + ], + "spans": [ + { + "bbox": [ + 54, + 123, + 295, + 219 + ], + "type": "text", + "content": "Mask sensitivity evaluation Previous shadow removal methods that rely on external shadow masks are sensitive to the accuracy of the input masks. Their performance can degrade when lack of ground truth mask. In contrast, end-to-end shadow removal methods mitigate the impact of low-quality masks. To further illustrate the sensitivity of prior SOTA methods to input masks, we evaluate their performance using various mask inputs." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 220, + 295, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 220, + 295, + 447 + ], + "spans": [ + { + "bbox": [ + 56, + 220, + 295, + 447 + ], + "type": "text", + "content": "Specifically, we evaluate ShadowDiffusion [9] and HomoFormer [34] using a range of input masks, including ground truth (GT) masks, Otsu masks [14], BDRAR [37] detected masks, FDRNet [38] detected masks and pretrained SAM [19] detected masks as inputs. The PSNR results on the " + }, + { + "bbox": [ + 56, + 220, + 295, + 447 + ], + "type": "inline_equation", + "content": "\\mathrm{ISTD + }" + }, + { + "bbox": [ + 56, + 220, + 295, + 447 + ], + "type": "text", + "content": " [21] datasets are shown in Table 4. The Otsu mask follows the approach in MaskShadowGAN [14], using the difference between shadow-free and shadowed images to generate masks; both Otsu and GT masks include additional shadow-free information. Results indicate that our methods maintain consistent performance across all tests, whereas ShadowDiffusion and HomoFormer have performance declines when using generated masks. Although our method achieves slightly lower PSNR values than those obtained with Otsu or GT masks, this is reasonable given that the " + }, + { + "bbox": [ + 56, + 220, + 295, + 447 + ], + "type": "inline_equation", + "content": "\\mathrm{ISTD + }" + }, + { + "bbox": [ + 56, + 220, + 295, + 447 + ], + "type": "text", + "content": " dataset contains a larger proportion of hard shadow images, where our method has less of an advantage. Our approach surpasses the rest results that rely on masks generated by shadow detectors." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 449, + 295, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 449, + 295, + 593 + ], + "spans": [ + { + "bbox": [ + 55, + 449, + 295, + 593 + ], + "type": "text", + "content": "Ablation Study We ablate our method, SoftShadow, by adding three losses: the shadow removal loss " + }, + { + "bbox": [ + 55, + 449, + 295, + 593 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{rem}" + }, + { + "bbox": [ + 55, + 449, + 295, + 593 + ], + "type": "text", + "content": ", the mask reconstruction loss " + }, + { + "bbox": [ + 55, + 449, + 295, + 593 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{mask}" + }, + { + "bbox": [ + 55, + 449, + 295, + 593 + ], + "type": "text", + "content": ", and the penumbra formation constraint loss " + }, + { + "bbox": [ + 55, + 449, + 295, + 593 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{pen}" + }, + { + "bbox": [ + 55, + 449, + 295, + 593 + ], + "type": "text", + "content": ", and demonstrate the results on SRD and LRSS datasets. First, we present the shadow removal results using the pretrained SAM and ShadowDiffusion models. Next, we finetune the framework with the shadow removal loss " + }, + { + "bbox": [ + 55, + 449, + 295, + 593 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{rem}" + }, + { + "bbox": [ + 55, + 449, + 295, + 593 + ], + "type": "text", + "content": ". Then, we add the mask reconstruction loss " + }, + { + "bbox": [ + 55, + 449, + 295, + 593 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{mask}" + }, + { + "bbox": [ + 55, + 449, + 295, + 593 + ], + "type": "text", + "content": " to provide SAM with more precise mask position guidance. Finally, we add the penumbra formation constraint loss " + }, + { + "bbox": [ + 55, + 449, + 295, + 593 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{pen}" + }, + { + "bbox": [ + 55, + 449, + 295, + 593 + ], + "type": "text", + "content": " to refine the mask boundaries for smoother transitions in the shadow regions." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "content": "The results in Table 5 are based on training with the SRD training set and evaluating on both the SRD and LRSS datasets. As shown in Table 5, the shadow removal loss " + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{rem}" + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "content": " improves performance on the SRD dataset, primarily by enhancing the shadow removal network. Adding the mask reconstruction loss " + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{mask}" + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "content": " further fine-tunes SAM, producing more accurate shadow masks. This improvement of mask accuracy leads to better shadow removal results. To evaluate the impact of the penumbra formation constraint loss " + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{pen}" + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "content": ", we compare results with and without " + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{pen}" + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "content": ". The" + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 326, + 71, + 543, + 126 + ], + "blocks": [ + { + "bbox": [ + 326, + 71, + 543, + 126 + ], + "lines": [ + { + "bbox": [ + 326, + 71, + 543, + 126 + ], + "spans": [ + { + "bbox": [ + 326, + 71, + 543, + 126 + ], + "type": "table", + "html": "
MethodsLRSSSRD
PSNR↑SSIM↑MAE↓PSNR↑SSIM↑MAE↓
Pretrained Weights21.400.91012.2631.270.9634.41
Lrem21.780.91312.0735.270.9733.32
Lrem + Lmask23.080.9359.9735.440.9743.17
Lrem + Lmask + Lpen23.320.9339.7735.570.9753.11
", + "image_path": "aecb285ea9ed94e823eced7f42ed019007d89a0f614bc080b4d40585131947b9.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 134, + 555, + 178 + ], + "lines": [ + { + "bbox": [ + 313, + 134, + 555, + 178 + ], + "spans": [ + { + "bbox": [ + 313, + 134, + 555, + 178 + ], + "type": "text", + "content": "Table 5. The ablation studies for the shadow removal loss " + }, + { + "bbox": [ + 313, + 134, + 555, + 178 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{rem}" + }, + { + "bbox": [ + 313, + 134, + 555, + 178 + ], + "type": "text", + "content": ", the mask reconstruction loss " + }, + { + "bbox": [ + 313, + 134, + 555, + 178 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{mask}" + }, + { + "bbox": [ + 313, + 134, + 555, + 178 + ], + "type": "text", + "content": ", and the penumbra formation constraint loss " + }, + { + "bbox": [ + 313, + 134, + 555, + 178 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{pen}" + }, + { + "bbox": [ + 313, + 134, + 555, + 178 + ], + "type": "text", + "content": " in LRSS and SRD datasets. The best results are boldfaced." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 337, + 191, + 535, + 304 + ], + "blocks": [ + { + "bbox": [ + 337, + 191, + 535, + 304 + ], + "lines": [ + { + "bbox": [ + 337, + 191, + 535, + 304 + ], + "spans": [ + { + "bbox": [ + 337, + 191, + 535, + 304 + ], + "type": "image", + "image_path": "60082c305a0592f811a8470176c04645a6ab3299647ce5d58ed5652076771e78.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 316, + 555, + 361 + ], + "lines": [ + { + "bbox": [ + 313, + 316, + 555, + 361 + ], + "spans": [ + { + "bbox": [ + 313, + 316, + 555, + 361 + ], + "type": "text", + "content": "Figure 8. Visual comparison of results from our framework with different configurations: using pretrained weights, finetuned with the loss function " + }, + { + "bbox": [ + 313, + 316, + 555, + 361 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{rem}" + }, + { + "bbox": [ + 313, + 316, + 555, + 361 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 316, + 555, + 361 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{rem} + \\mathcal{L}_{mask}" + }, + { + "bbox": [ + 313, + 316, + 555, + 361 + ], + "type": "text", + "content": " and our full configuration " + }, + { + "bbox": [ + 313, + 316, + 555, + 361 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{rem} + \\mathcal{L}_{mask} + \\mathcal{L}_{pen}" + }, + { + "bbox": [ + 313, + 316, + 555, + 361 + ], + "type": "text", + "content": " (Ours), respectively." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 381, + 555, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 381, + 555, + 525 + ], + "spans": [ + { + "bbox": [ + 313, + 381, + 555, + 525 + ], + "type": "text", + "content": "results show that " + }, + { + "bbox": [ + 313, + 381, + 555, + 525 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{pen}" + }, + { + "bbox": [ + 313, + 381, + 555, + 525 + ], + "type": "text", + "content": " further aids in shadow removal on the SRD dataset, refining the shadow edges and enhancing the overall shadow removal quality. In the LRSS dataset, both " + }, + { + "bbox": [ + 313, + 381, + 555, + 525 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{rem}" + }, + { + "bbox": [ + 313, + 381, + 555, + 525 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 381, + 555, + 525 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{mask}" + }, + { + "bbox": [ + 313, + 381, + 555, + 525 + ], + "type": "text", + "content": " significantly improve performance, the " + }, + { + "bbox": [ + 313, + 381, + 555, + 525 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{pen}" + }, + { + "bbox": [ + 313, + 381, + 555, + 525 + ], + "type": "text", + "content": " improves the PSNR and MAE significantly, although it results in a slight decrease in SSIM. As shown in Figure 8, adding the constraint loss " + }, + { + "bbox": [ + 313, + 381, + 555, + 525 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{pen}" + }, + { + "bbox": [ + 313, + 381, + 555, + 525 + ], + "type": "text", + "content": " leads to more accurate and smoother soft mask predictions. The combined effects of " + }, + { + "bbox": [ + 313, + 381, + 555, + 525 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{mask}" + }, + { + "bbox": [ + 313, + 381, + 555, + 525 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 381, + 555, + 525 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{pen}" + }, + { + "bbox": [ + 313, + 381, + 555, + 525 + ], + "type": "text", + "content": " lead to more accurate and smoother soft mask predictions. Together, these losses provide the removal framework with enhanced soft shadow details, resulting in superior performance in shadow removal." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 537, + 388, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 537, + 388, + 550 + ], + "spans": [ + { + "bbox": [ + 313, + 537, + 388, + 550 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "text", + "content": "In this paper, we introduce novel soft shadow masks designed specifically for shadow removal. To achieve this, we propose a unified framework, SoftShadow, that eliminates the need for additional shadow mask input. By leveraging a pretrained SAM with LoRA, the framework accurately predicts soft masks as intermediate results, capturing detailed and varied shadow information. Additionally, we introduce a penumbra formation constraint, inspired by the physical shadow formation model, to jointly tune SAM and the shadow removal network, refining the soft mask in the penumbra area and facilitating artifact-free restoration. Extensive experiments demonstrate that our method is superior on various occasions, proving the validity of our method." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "23224" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 161, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 161, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 161, + 85 + ], + "type": "text", + "content": "7. Acknowledgment" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 140, + 115, + 153 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 140, + 115, + 153 + ], + "spans": [ + { + "bbox": [ + 56, + 140, + 115, + 153 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 161, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 161, + 296, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 161, + 296, + 227 + ], + "spans": [ + { + "bbox": [ + 61, + 161, + 296, + 227 + ], + "type": "text", + "content": "[1] Tianrun Chen, Lanyun Zhu, Chaotao Deng, Runlong Cao, Yan Wang, Shangzhan Zhang, Zejian Li, Lingyun Sun, Ying Zang, and Papa Mao. Sam-adapter: Adapting segment anything in underperformed scenes. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3367-3375, 2023. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 228, + 296, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 228, + 296, + 283 + ], + "spans": [ + { + "bbox": [ + 61, + 228, + 296, + 283 + ], + "type": "text", + "content": "[2] Zhihao Chen, Lei Zhu, Liang Wan, Song Wang, Wei Feng, and Pheng-Ann Heng. A multi-task mean teacher for semi-supervised shadow detection. In Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition, pages 5611–5620, 2020. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 285, + 296, + 339 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 285, + 296, + 339 + ], + "spans": [ + { + "bbox": [ + 62, + 285, + 296, + 339 + ], + "type": "text", + "content": "[3] Xiaodong Cun, Chi-Man Pun, and Cheng Shi. Towards ghost-free shadow removal via dual hierarchical aggregation network and shadow matting gan. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 10680-10687, 2020. 1, 2, 3, 5, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 342, + 296, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 342, + 296, + 407 + ], + "spans": [ + { + "bbox": [ + 62, + 342, + 296, + 407 + ], + "type": "text", + "content": "[4] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 409, + 296, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 409, + 296, + 463 + ], + "spans": [ + { + "bbox": [ + 62, + 409, + 296, + 463 + ], + "type": "text", + "content": "[5] Lan Fu, Changqing Zhou, Qing Guo, Felix Juefei-Xu, Hongkai Yu, Wei Feng, Yang Liu, and Song Wang. Autoexposure fusion for single-image shadow removal. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10571-10580, 2021. 3, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 466, + 296, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 466, + 296, + 509 + ], + "spans": [ + { + "bbox": [ + 62, + 466, + 296, + 509 + ], + "type": "text", + "content": "[6] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 511, + 296, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 511, + 296, + 543 + ], + "spans": [ + { + "bbox": [ + 62, + 511, + 296, + 543 + ], + "type": "text", + "content": "[7] Maciej Gryka, Michael Terry, and Gabriel J Brostow. Learning to remove soft shadows. ACM Transactions on Graphics (TOG), 34(5):1-15, 2015. 5, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 545, + 296, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 545, + 296, + 590 + ], + "spans": [ + { + "bbox": [ + 62, + 545, + 296, + 590 + ], + "type": "text", + "content": "[8] Lanqing Guo, Siyu Huang, Ding Liu, Hao Cheng, and Bihan Wen. Shadowformer: Global context helps shadow removal. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 710-718, 2023. 1, 3, 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 592, + 296, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 592, + 296, + 656 + ], + "spans": [ + { + "bbox": [ + 62, + 592, + 296, + 656 + ], + "type": "text", + "content": "[9] Lanqing Guo, Chong Wang, Wenhan Yang, Siyu Huang, Yufei Wang, Hanspeter Pfister, and Bihan Wen. Shadowdiffusion: When degradation prior meets diffusion model for shadow removal. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14049-14058, 2023. 1, 2, 5, 7, 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 658, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 658, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 658, + 296, + 713 + ], + "type": "text", + "content": "[10] Lanqing Guo, Chong Wang, Wenhan Yang, Yufei Wang, and Bihan Wen. Boundary-aware divide and conquer: A diffusion-based solution for unsupervised shadow removal. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13045-13054, 2023. 5" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 57, + 73, + 554, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 55, + 91, + 297, + 128 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 297, + 128 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 297, + 128 + ], + "type": "text", + "content": "This work was partially supported by the National Research Foundation Singapore Competitive Research Program (award number CRP29-2022-0003)" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 316, + 73, + 554, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 115 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 115 + ], + "type": "text", + "content": "[11] Ruiqi Guo, Qieyun Dai, and Derek Hoiem. Paired regions for shadow detection and removal. IEEE transactions on pattern analysis and machine intelligence, 35(12):2956-2967, 2012. 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 118, + 554, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 118, + 554, + 150 + ], + "spans": [ + { + "bbox": [ + 317, + 118, + 554, + 150 + ], + "type": "text", + "content": "[12] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 152, + 554, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 152, + 554, + 194 + ], + "spans": [ + { + "bbox": [ + 317, + 152, + 554, + 194 + ], + "type": "text", + "content": "[13] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 2, 4, 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 195, + 554, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 195, + 554, + 248 + ], + "spans": [ + { + "bbox": [ + 317, + 195, + 554, + 248 + ], + "type": "text", + "content": "[14] Xiaowei Hu, Yitong Jiang, Chi-Wing Fu, and Pheng-Ann Heng. Mask-shadowgan: Learning to remove shadows from unpaired data. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2472–2481, 2019. 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 251, + 553, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 251, + 553, + 271 + ], + "spans": [ + { + "bbox": [ + 316, + 251, + 553, + 271 + ], + "type": "text", + "content": "[15] Leiping Jie and Hui Zhang. When sam meets shadow detection. arXiv preprint arXiv:2305.11513, 2023. 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 273, + 554, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 273, + 554, + 327 + ], + "spans": [ + { + "bbox": [ + 317, + 273, + 554, + 327 + ], + "type": "text", + "content": "[16] Yeying Jin, Aashish Sharma, and Robby T Tan. Dcshadownet: Single-image hard and soft shadow removal using unsupervised domain-classifier guided network. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5027-5036, 2021. 2, 5, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 328, + 554, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 328, + 554, + 381 + ], + "spans": [ + { + "bbox": [ + 317, + 328, + 554, + 381 + ], + "type": "text", + "content": "[17] Yeying Jin, Wei Ye, Wenhan Yang, Yuan Yuan, and Robby T Tan. Des3: Adaptive attention-driven self and soft shadow removal using vit similarity. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2634-2642, 2024. 2, 5, 6, 7, 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 383, + 554, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 383, + 554, + 414 + ], + "spans": [ + { + "bbox": [ + 316, + 383, + 554, + 414 + ], + "type": "text", + "content": "[18] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 416, + 554, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 416, + 554, + 480 + ], + "spans": [ + { + "bbox": [ + 317, + 416, + 554, + 480 + ], + "type": "text", + "content": "[19] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 2, 3, 4, 5, 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 482, + 554, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 482, + 554, + 525 + ], + "spans": [ + { + "bbox": [ + 317, + 482, + 554, + 525 + ], + "type": "text", + "content": "[20] Michael S Langer and Steven W Zucker. What is a light source? In Proceedings of IEEE Computer Society Conference on Computer Vision and Pattern Recognition, pages 172-178. IEEE, 1997. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 526, + 554, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 526, + 554, + 568 + ], + "spans": [ + { + "bbox": [ + 317, + 526, + 554, + 568 + ], + "type": "text", + "content": "[21] Hieu Le and Dimitris Samaras. Shadow removal via shadow image decomposition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8578-8587, 2019. 2, 5, 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 571, + 554, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 571, + 554, + 624 + ], + "spans": [ + { + "bbox": [ + 317, + 571, + 554, + 624 + ], + "type": "text", + "content": "[22] Xiaoguang Li, Qing Guo, Rabab Abdelfattah, Di Lin, Wei Feng, Ivor Tsang, and Song Wang. Leveraging inpainting for single-image shadow removal. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13055-13064, 2023. 5, 7, 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 626, + 554, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 626, + 554, + 669 + ], + "spans": [ + { + "bbox": [ + 317, + 626, + 554, + 669 + ], + "type": "text", + "content": "[23] Yuhao Liu, Zhanghan Ke, Ke Xu, Fang Liu, Zhenwei Wang, and Rynson WH Lau. Recasting regional lighting for shadow removal. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3810-3818, 2024. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 670, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 554, + 713 + ], + "type": "text", + "content": "[24] Zhihao Liu, Hui Yin, Xinyi Wu, Zhenyao Wu, Yang Mi, and Song Wang. From shadow generation to shadow removal. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4927-4936, 2021. 5, 7" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "23225" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 703 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "text", + "content": "[25] Vinod Nair and Geoffrey E Hinton. Rectified linear units improve restricted boltzmann machines. In Proceedings of the 27th international conference on machine learning (ICML-10), pages 807–814, 2010. 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 172 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 172 + ], + "type": "text", + "content": "[26] Michael Nielsen and Claus B Madsen. Segmentation of soft shadows based on a daylight-and penumbra model. In International Conference on Computer Vision/Computer Graphics Collaboration Techniques and Applications, pages 341-352. Springer, 2007. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 173, + 294, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 173, + 294, + 227 + ], + "spans": [ + { + "bbox": [ + 56, + 173, + 294, + 227 + ], + "type": "text", + "content": "[27] Liangqiong Qu, Jiandong Tian, Shengfeng He, Yandong Tang, and Rynson WH Lau. Deshadownet: A multi-context embedding deep network for shadow removal. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4067-4075, 2017. 2, 5, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 228, + 294, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 228, + 294, + 281 + ], + "spans": [ + { + "bbox": [ + 56, + 228, + 294, + 281 + ], + "type": "text", + "content": "[28] Chitwan Sahara, Jonathan Ho, William Chan, Tim Salimans, David J Fleet, and Mohammad Norouzi. Image superresolution via iterative refinement. IEEE transactions on pattern analysis and machine intelligence, 45(4):4713-4726, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 282, + 294, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 282, + 294, + 326 + ], + "spans": [ + { + "bbox": [ + 56, + 282, + 294, + 326 + ], + "type": "text", + "content": "[29] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 327, + 294, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 327, + 294, + 369 + ], + "spans": [ + { + "bbox": [ + 56, + 327, + 294, + 369 + ], + "type": "text", + "content": "[30] Jin Wan, Hui Yin, Zhenyao Wu, Xinyi Wu, Yanting Liu, and Song Wang. Style-guided shadow removal. In European Conference on Computer Vision, pages 361-378. Springer, 2022. 1, 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 371, + 294, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 371, + 294, + 425 + ], + "spans": [ + { + "bbox": [ + 56, + 371, + 294, + 425 + ], + "type": "text", + "content": "[31] Jifeng Wang, Xiang Li, and Jian Yang. Stacked conditional generative adversarial networks for jointly learning shadow detection and shadow removal. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1788-1797, 2018. 1, 2, 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 426, + 294, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 426, + 294, + 469 + ], + "spans": [ + { + "bbox": [ + 56, + 426, + 294, + 469 + ], + "type": "text", + "content": "[32] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 471, + 294, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 471, + 294, + 514 + ], + "spans": [ + { + "bbox": [ + 56, + 471, + 294, + 514 + ], + "type": "text", + "content": "[33] Cort J Willmott and Kenji Matsuura. Advantages of the mean absolute error (mae) over the root mean square error (rmse) in assessing average model performance. *Climate research*, 30(1):79–82, 2005. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 515, + 294, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 515, + 294, + 568 + ], + "spans": [ + { + "bbox": [ + 56, + 515, + 294, + 568 + ], + "type": "text", + "content": "[34] Jie Xiao, Xueyang Fu, Yurui Zhu, Dong Li, Jie Huang, Kai Zhu, and Zheng-Jun Zha. Homoformer: Homogenized transformer for image shadow removal. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25617-25626, 2024. 1, 2, 3, 5, 7, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 570, + 294, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 570, + 294, + 602 + ], + "spans": [ + { + "bbox": [ + 56, + 570, + 294, + 602 + ], + "type": "text", + "content": "[35] Xiaofeng Zhang, Chaochen Gu, and Shanying Zhu. Sam-helps-shadow: When segment anything model meet shadow removal. arXiv e-prints, pages arXiv-2306, 2023. 2, 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 604, + 294, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 604, + 294, + 647 + ], + "spans": [ + { + "bbox": [ + 56, + 604, + 294, + 647 + ], + "type": "text", + "content": "[36] Quanlong Zheng, Xiaotian Qiao, Ying Cao, and Rynson WH Lau. Distraction-aware shadow detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5167-5176, 2019. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 647, + 294, + 703 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 294, + 703 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 294, + 703 + ], + "type": "text", + "content": "[37] Lei Zhu, Zijun Deng, Xiaowei Hu, Chi-Wing Fu, Xuemiao Xu, Jing Qin, and Pheng-Ann Heng. Bidirectional feature pyramid network with recurrent attention residual modules for shadow detection. In Proceedings of the European Conference on Computer Vision (ECCV), pages 121–136, 2018." + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 182 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "type": "text", + "content": "[38] Lei Zhu, Ke Xu, Zhanghan Ke, and Rynson WH Lau. Mitigating intensity bias in shadow detection via feature decomposition and reweighting. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4702-4711, 2021. 5, 6, 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 129, + 553, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 553, + 182 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 553, + 182 + ], + "type": "text", + "content": "[39] Yurui Zhu, Jie Huang, Xueyang Fu, Feng Zhao, Qibin Sun, and Zheng-Jun Zha. Bijective mapping network for shadow removal. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5627-5636, 2022. 1, 3, 5, 7" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "23226" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/SoftVQ-VAE_ Efficient 1-Dimensional Continuous Tokenizer/315fadef-93cf-49d1-961b-c20538f1f4de_content_list.json b/2025/SoftVQ-VAE_ Efficient 1-Dimensional Continuous Tokenizer/315fadef-93cf-49d1-961b-c20538f1f4de_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..710af0e501c7b5654fa6dc07ed6766eff7dcbcb7 --- /dev/null +++ b/2025/SoftVQ-VAE_ Efficient 1-Dimensional Continuous Tokenizer/315fadef-93cf-49d1-961b-c20538f1f4de_content_list.json @@ -0,0 +1,1629 @@ +[ + { + "type": "text", + "text": "SoftVQ-VAE: Efficient 1-Dimensional Continuous Tokenizer", + "text_level": 1, + "bbox": [ + 192, + 130, + 803, + 151 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hao Chen $^{1,2*}$ , Ze Wang $^{2}$ , Xiang Li $^{1}$ , Xineng Sun $^{2}$ , Fangyi Chen $^{1}$ , Jiang Liu $^{2}$ , Jindong Wang $^{3}$ , Bhiksha Raj $^{1,4}$ , Zicheng Liu $^{2}$ , Emad Barsoum $^{2}$", + "bbox": [ + 194, + 179, + 802, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Carnegie Mellon University, $^{2}$ AMD, $^{3}$ William & Mary, $^{4}$ MBZUAI", + "bbox": [ + 295, + 217, + 697, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 268, + 326, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Efficient image tokenization with high compression ratios remains a critical challenge for training generative models. We present SoftVQ-VAE, a continuous image tokenizer that leverages soft categorical posteriors to aggregate multiple codewords into each latent token, substantially increasing the representation capacity of the latent space. When applied to Transformer-based architectures, our approach compresses $256 \\times 256$ and $512 \\times 512$ images using as few as 32 or 64 1-dimensional tokens. Not only does SoftVQ-VAE show consistent and high-quality reconstruction, more importantly, it also achieves state-of-the-art and significantly faster image generation results across different denoising-based generative models. Remarkably, SoftVQ-VAE improves inference throughput by up to $18x$ for generating $256 \\times 256$ images and $55x$ for $512 \\times 512$ images while achieving competitive FID scores of 1.78 and 2.21 for SiT-XL. It also improves the training efficiency of the generative models by reducing the number of training iterations by $2.3x$ while maintaining comparable performance. With its fully-differentiable design and semantic-rich latent space, our experiment demonstrates that SoftVQ-VAE achieves efficient tokenization without compromising generation quality, paving the way for more efficient generative models. Code and model are released1.", + "bbox": [ + 89, + 299, + 485, + 648 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 674, + 220, + 690 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Denoising-based generative modeling has witnessed remarkable progress with recent advances, such as Diffusion Transformers (DiT) [80], Scalable Interpolant Transformers (SiT) [72], and Masked Auto-Regressive models with diffusion loss (MAR) [60], to name a few. Denosing-based generative modeling not only presents impressive generation results on a wide range of modality, including natural language [32, 122], images [16, 25, 69, 83, 88], videos [6, 8, 42], and audios [24, 81], but also shows potential to unify the understanding and generation capabilities across modalities in multi-modal language models [107, 109, 110, 123].", + "bbox": [ + 89, + 700, + 485, + 864 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "A core component of denoising-based generative models is the tokenizer [23, 38, 51, 56, 84, 104, 112, 117], which compresses the raw data of each modality into a set of latent tokens in either a discrete or continuous latent space. The compact latent space therefore allows for more efficient and better generative modeling [88]. Among previous efforts, Kullback-Leibler Variational Auto-Encoders (KL-VAE) [51] and Vector Quantized Variational Auto-Encoders (VQ-VAE) [23, 84, 104] stand out as representatives of tokenizers which introduce continuous and discrete latent spaces, respectively. The former constrains the latent space with a Gaussian distribution using re-parametrization [51] trick, and the latter makes its latent space a categorical discrete distribution with a codebook of finite vocabulary which requires straight-through estimation [7].", + "bbox": [ + 511, + 268, + 908, + 496 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Although both KL-VAE and VQ-VAE (and their variants) have been predominantly adopted in many generative models [10, 80, 88, 98, 123], they still present two major challenges restricting the efficiency and effectiveness of generative modeling: (1) the difficulty of achieving a higher compression ratio [13] and (2) the worse discriminative representations in their latent space than other self-supervised methods [14, 35, 109, 116]. The efficiency of downstream generative models, particularly Transformer-based architectures [105], is fundamentally constrained by their quadratic complexity to the number of latent tokens. Current image tokenizers [23, 88, 96] typically compress $256 \\times 256$ images to at least 256 tokens and $512 \\times 512$ images to at least 1024 tokens, creating a significant computational bottleneck for both training and inference of generative models [60, 72, 80]. Many efforts have been made to reduce the number of latent tokens on the generative model side, such as merging [62, 72, 80], pooling [58, 61, 94], and others [12, 63, 70]. More studies have recently emerged to fundamentally reduce the token number of the tokenizer [11, 30, 62, 115]. For example, TiTok [115] uses 128 tokens, achieving generation results comparable to 256 tokens by adding one extra decoder. DC-AE [11] compresses the initial 1024 tokens of $512 \\times 512$ with 256 tokens. However, further increasing the compression ratio results in a significant degradation in the quality of reconstruction and therefore the generation [11].", + "bbox": [ + 511, + 508, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Work done during internship at AMD. Email: haoc3@andrew.cmu.edu.", + "bbox": [ + 107, + 875, + 482, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "1https://github.com/Hhhhhhao/continuous_tokenizer.", + "bbox": [ + 109, + 887, + 450, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "28358", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/21e994b6b9a87a8fc9dafce3a267dd48eb64d76654861724910fd1a07b80c5fa.jpg", + "image_caption": [ + "Figure 1. ImageNet-1K $256\\times 256$ and $512\\times 512$ generation results of generative models trained on SoftVQ-VAE with 32 and 64 tokens." + ], + "image_footnote": [], + "bbox": [ + 109, + 88, + 883, + 388 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Beyond the compression challenge, the quality of the learned representations of the tokenizer poses another critical limitation. While latent representations are crucial for both understanding [31, 109] and generation tasks [30, 62, 64, 116], current tokenizers usually struggle to capture more discriminative features. Earlier methods attempted to advance the training objectives and recipe to improve the latent space of the tokenizer [14, 19, 23, 34, 124]. Recently, [30, 62, 109, 113, 116, 118, 124] has also explored the representation alignment of latent tokens. However, due to the lossy nature [77, 91] of the smoothness constraints in KL-VAE and the discrete quantization in VQVAE [13, 27, 38, 126], current tokenizers usually fall short in learning the semantics of the latent space [109].", + "bbox": [ + 89, + 412, + 483, + 625 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address the aforementioned challenges, we present SoftVQ-VAE, a simple modification to VQ-VAE that converts it from a discrete tokenizer into a continuous one with a high compression ratio and a semantic-rich latent space. Specifically, we propose using a soft categorical posterior in VAE with a learnable codebook. This straightforward modification introduces a crucial capability: instead of the conventional one-to-one mapping between tokens and codewords in VQ-VAE, our approach enables the adaptive aggregation of multiple codewords into each latent token, substantially increasing the representation capacity of the latent space. We apply SoftVQ-VAE to the Transformer auto-encoder architecture [30, 58, 62, 111, 115], and we succeed in using much fewer 1-dimensional latent tokens (i.e., 32 and 64) for both the reconstruction and subsequent generation tasks. Since SoftVQ-VAE is now fully differentiable, it not only simplifies the learning of the codebook, not requiring the codebook loss or commit loss as in VQ-VAE [23, 104], but", + "bbox": [ + 89, + 628, + 486, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "also enables better representation learning through direct alignment with pre-trained semantic features using a simple cosine similarity objective. Our approach generalizes the K-Means [67] latent space of VQ-VAE to Soft K-Means [22] and further to Gaussian Mixture Models (GMM) [86], while maintaining compatibility with existing VQ-VAE techniques [45, 55, 62, 112]. We comprehensively demonstrate the efficiency and effectiveness of the proposed SoftVQ-VAE in image reconstruction and generation tasks, and show that:", + "bbox": [ + 511, + 412, + 908, + 547 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- SoftVQ-VAE presents more consistent, robust, and high-quality reconstruction and generation results with various latent token lengths ranging from 256 to 32.", + "- SoftVQ-VAE allows diffusion-based DiT [73], flow-based SiT [72], and autoregressive-based MAR [60] to achieve state-of-the-art generation results with only 64 and 32 tokens on both the $256 \\times 256$ , with an FID of 1.78, and $512 \\times 512$ on ImageNet benchmark, with an FID of 2.21.", + "- SoftVQ-VAE improves the inference throughput by 18x and 10x with 32 and 64 tokens to generate $256 \\times 256$ images, and 55x with 64 tokens for generating $512 \\times 512$ images, with significantly lower GFLOPs for all models.", + "- SoftVQ-VAE presents a more discriminative latent space through representation alignment. It also facilitates training efficiency by significantly improving training throughput and reduces training iterations by 2.3x." + ], + "bbox": [ + 514, + 553, + 908, + 794 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Preliminary", + "text_level": 1, + "bbox": [ + 511, + 824, + 638, + 842 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We present an overview of the tokenizers, i.e., KL-VAE [51] and VQ-VAE [104], and denoising-based generative models, i.e., DiT [80], SiT [72], and MAR [60], in this section.", + "bbox": [ + 511, + 854, + 908, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "28359", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Image Tokenizer", + "text_level": 1, + "bbox": [ + 89, + 90, + 256, + 107 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The architecture of image tokenizer generally resembles auto-encoders, consisting of two main components: an encoder $\\mathcal{E}$ and a decoder $\\mathcal{D}$ , parameterized by $\\phi$ and $\\theta$ , respectively. Given an input image $\\mathbf{x} \\in \\mathbb{R}^{H \\times W \\times 3}$ , the encoder $\\mathcal{E}$ maps the high-dimensional $\\mathbf{x}$ to a lower-dimensional latent representation $\\mathbf{z} = \\mathcal{E}(\\mathbf{x}; \\phi)$ . The latent representations can have varying shapes depending on the encoder architecture. We generally view the latent representation as a set of latent tokens $\\mathbf{z} = [\\mathbf{z}^{[0]}, \\mathbf{z}^{[1]}, \\dots, \\mathbf{z}^{[L]}] \\in \\mathbb{R}^{L \\times D}$ , where $L$ is the number of tokens and $D$ is the dimension of the latent representation. For basic auto-encoders (AE), the decoder takes latent tokens as input and reconstructs the original signal $\\hat{\\mathbf{x}} = \\mathcal{D}(\\mathbf{z}; \\theta)$ by minimizing the reconstruction loss with $\\mathbf{x}$ .", + "bbox": [ + 89, + 112, + 485, + 308 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Variational auto-encoders (VAE) [51] extend the basic AE by introducing a probabilistic perspective on the latent space. Specifically, VAE approximates a posterior distribution $p(\\mathbf{z}|\\mathbf{x})$ with a learned distribution $q_{\\phi}(\\mathbf{z}|\\mathbf{x})$ from the encoder, assuming that $\\mathbf{x}$ is generated by the unobserved latent $\\mathbf{z}$ with a prior distribution $p(\\mathbf{z})$ . The decoder instead takes a set of sampled latent tokens $\\mathbf{z} \\sim q_{\\phi}(\\mathbf{z}|\\mathbf{x})$ from the posterior and aims to learn the marginal likelihood of data in a generative process by optimizing the evidence lower bound (ELBO): $\\max_{\\phi,\\theta}\\mathbb{E}_{q_{\\phi}(\\mathbf{z}|\\mathbf{x})}\\log p_{\\theta}(\\mathbf{x}|\\mathbf{z}) - \\mathcal{L}_{\\mathrm{kl}}(q_{\\phi}(\\mathbf{z}|\\mathbf{x})|p(\\mathbf{z}))$ . To make the optimization tractable, assumptions have been made on the prior, leading to different variants of VAE.", + "bbox": [ + 89, + 309, + 483, + 489 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "KL-VAE [38, 51] parametrizes both the prior and posterior as Gaussians. The prior $p(\\mathbf{z})$ is assumed to be the isotropic unit Gaussian $\\mathcal{N}(0,1)$ . The continuous latent code is parameterized with posterior mean $\\mu_{\\phi}$ and variance $\\sigma_{\\phi}^{2}$ predicted by the encoder using the \"re-parametrization\" trick with a noise variable $\\varepsilon$ from standard Gaussian:", + "bbox": [ + 89, + 489, + 483, + 580 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\text {p o s t e r i o r :} q _ {\\phi} (\\mathbf {z} | \\mathbf {x}) = \\mathcal {N} (\\mathbf {z}; \\boldsymbol {\\mu} _ {\\phi} (\\mathbf {x}), \\boldsymbol {\\sigma} _ {\\phi} ^ {2} (\\mathbf {x}))\n$$\n", + "text_format": "latex", + "bbox": [ + 129, + 589, + 380, + 604 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nl a t e n t: \\mathbf {z} = \\boldsymbol {\\mu} _ {\\phi} (\\mathbf {x}) + \\boldsymbol {\\sigma} _ {\\phi} (\\mathbf {x}) \\odot \\varepsilon , \\quad \\epsilon \\sim \\mathcal {N} (0, \\mathbf {I}) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 129, + 607, + 482, + 626 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nk l: \\mathcal {L} _ {\\mathrm {k l}} = - \\frac {1}{2} \\left(1 + \\log \\left(\\boldsymbol {\\sigma} _ {\\phi} ^ {2} (\\mathbf {x})\\right) - \\boldsymbol {\\mu} _ {\\phi} (\\mathbf {x}) ^ {2} - \\boldsymbol {\\sigma} _ {\\phi} ^ {2} (\\mathbf {x})\\right)\n$$\n", + "text_format": "latex", + "bbox": [ + 129, + 626, + 442, + 650 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "While KL-VAE is widely used in diffusion models, it is known to have the \"posterior-collapse\" issue [13]. In addition, the KL loss weight usually imposes a trade-off between the quality of reconstruction and the smoothness of its latent space [100], preventing high compression ratio and the learning of semantics in the latent space of KL-VAE.", + "bbox": [ + 89, + 660, + 483, + 751 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "VQ-VAE [23, 104], in contrast, generates latent code in a discrete space by taking the posterior as a \\(K\\)-way deterministic categorical distribution against a learnable codebook \\(\\mathcal{C} = [\\mathbf{c}^{[0]}, \\dots, \\mathbf{c}^{[K]}\\] \\(\\in \\mathbb{R}^{K \\times D}\\), with \\(\\mathbf{c}\\) as the codeword:", + "bbox": [ + 89, + 752, + 483, + 813 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\text {p o s t e r i o r :} q _ {\\phi} (\\mathbf {z} = k | \\mathbf {x}) = \\left\\{ \\begin{array}{l l} 1 & \\text {i f} k = \\arg \\min _ {j} \\| \\hat {\\mathbf {z}} - \\mathbf {c} ^ {[ j ]} \\| _ {2} \\\\ 0 & \\text {o t h e r w i s e} \\end{array} \\right.\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 823, + 452, + 859 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nl a t e n t: \\mathbf {z} = \\mathbf {c} ^ {[ k ]}, \\quad \\text {w h e r e} k = \\underset {j} {\\arg \\min } \\| \\hat {\\mathbf {z}} - \\mathbf {c} ^ {[ j ]} \\| _ {2} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 101, + 857, + 482, + 883 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nk l: \\mathcal {L} _ {\\mathrm {k l}} = \\log K\n$$\n", + "text_format": "latex", + "bbox": [ + 101, + 888, + 197, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Due to the non-differentiable arg min, a \"straight-through\" trick [7] is adopted to approximate the gradient of the encoder output $\\hat{\\mathbf{z}}$ by directly copying gradient from the decoder input $\\mathbf{z}$ . Since the reconstruction objective does not impose direct gradients on the codebook $\\mathcal{C}$ , VQ-VAE additionally uses a codebook loss as $\\| \\mathrm{sg}[\\hat{\\mathbf{z}}] - \\mathbf{c}\\| _2$ to move codewords toward the encoder output, along with a commit loss $\\| \\hat{\\mathbf{z}} -\\mathrm{sg}[\\mathbf{c}]\\| _2$ to prevent arbitrary growth of codewords, where $\\mathrm{sg}[\\cdot ]$ denotes stop-gradient. The broken gradient hinders high compression ratio and latent space learning [27, 43].", + "bbox": [ + 511, + 90, + 908, + 243 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Denoising-based Generative Models", + "text_level": 1, + "bbox": [ + 511, + 252, + 825, + 268 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Denoising-based generative models synthesize images by progressively transforming Gaussian noise $\\varepsilon \\sim \\mathcal{N}(0,\\mathbf{I})$ into tokenizer latent codes $\\mathbf{z}$ through a forward process [41, 92]:", + "bbox": [ + 511, + 273, + 906, + 320 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {z} _ {t} = \\alpha_ {t} \\mathbf {z} + \\sigma_ {t} \\varepsilon , \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 651, + 333, + 903, + 348 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\alpha_{t}$ and $\\sigma_{t}$ are a decreasing and increasing function of $t$ , and $t \\in [0, T]$ [50]. In this paper, we adopt three approaches to implementing this denoising process as follows.", + "bbox": [ + 511, + 359, + 906, + 406 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "DiT [80] is a diffusion-based model with a transformer architecture. It formulates the process through a forward-time stochastic differential equation (SDE), where $\\mathbf{z}_t$ converges to $\\mathcal{N}(0, \\mathbf{I})$ as $t \\to T$ . Generation occurs via a reverse-time SDE [2], with the model trained to predict noise $\\varepsilon$ at randomly sampled timesteps. Due to the long token length of the tokenizer [96] used in DiT, it has several variants to merge tokens at the input of the transformer backbone.", + "bbox": [ + 511, + 406, + 908, + 526 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "SiT [72] instead uses stochastic interpolants [1] and performs denoising via probability flow ordinary differential equation (PF ODE). The model learns the ODE's velocity field [65] by minimizing mean-squared-error (MSE) loss. This velocity field defines a score function for generation, analogous to DiT's reverse-time SDE.", + "bbox": [ + 511, + 527, + 908, + 617 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "MAR [60] combines diffusion with autoregressive generation. Unlike DiT and SiT's parallel denoising of all latent tokens, MAR adopts an encoder-decoder transformer architecture [35] that progressively denoises tokens in a \"next-set\" autoregressive manner [99], starting from masked tokens, similar to MaskGIT [10] and MAGE [59].", + "bbox": [ + 511, + 617, + 908, + 708 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 511, + 720, + 604, + 737 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we present SoftVQ-VAE, a novel 1D continuous tokenizer with a high compression ratio and a semantic-rich latent space. We first introduce the architecture and the formulation of SoftVQ-VAE. Then, we show that SoftVQ-VAE can learn semantics easily by aligning its latent tokens with pre-trained features via its fully-differentiable property.", + "bbox": [ + 511, + 747, + 908, + 839 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Architecture", + "text_level": 1, + "bbox": [ + 511, + 848, + 648, + 862 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We leverage Vision Transformer (ViT) [21, 111] as the architecture for the encoder $\\mathcal{E}$ and decoder $\\mathcal{D}$ of SoftVQ-VAE.", + "bbox": [ + 511, + 869, + 908, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "28360", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/b757a7fb4120fe2d1d016fc288e2800319d031dd85905ac89c7cb35f0aeebb90.jpg", + "image_caption": [ + "Figure 2. Illustration of SoftVQ-VAE. Left: Transformer encoder-decoder architecture with image tokens, arbitrary length of latent tokens, and mask tokens. Right top: fully-differentiable SoftVQ illustration. Right bottom: latent space representation alignment." + ], + "image_footnote": [], + "bbox": [ + 96, + 88, + 477, + 306 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Similarly to Yu et al. [115], instead of using the image features as latent, we initialize a set of extra 1D learnable tokens and use these tokens for reconstruction and subsequent generation. With the self-attention mechanism, it allows the learnable tokens to adaptively aggregate different image tokens to obtain the latent tokens of SoftVQ, detailed in Sec. 3.2. Specifically, for the encoder transformer with a patch size of $P$ , an image $\\mathbf{x} \\in \\mathbb{R}^{H \\times W \\times 3}$ is patched into a sequence of image tokens $\\mathbf{x}_p \\in \\mathbb{R}^{N \\times D}$ , where $N = HW / P^2$ and $D$ represent the token dimension size. We then concatenate a set of learnable tokens $\\mathbf{z}_l \\in \\mathbb{R}^{L \\times D}$ of length $L$ along with the image tokens as the input to the encoder, and retain only the output corresponding to the learnable latent tokens as the output of our encoder: $\\hat{\\mathbf{z}} = \\mathcal{E}([\\mathbf{x}_p; \\mathbf{z}_l]; \\phi)$ .", + "bbox": [ + 88, + 371, + 483, + 584 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For the decoder, we reconstruct images from a sequence of learnable mask tokens $\\mathbf{m} \\in \\mathbb{R}^{N \\times D}$ [5, 35], concatenated with latent tokens: $\\hat{\\mathbf{x}} = \\mathcal{D}([\\mathbf{m};\\mathbf{z}];\\theta)$ . We use a linear layer at the end of the decoder to regress the pixel values from the mask tokens. For the image tokens at the encoder input and mask tokens at the decoder input, we apply 2D absolute position embeddings and 1D absolute position embedding for the latent tokens. Adopting more advanced position embedding techniques, such as RoPE [36, 97], may lead to better results and is left for future work. This design allows SoftVQ-VAE to use latent codes of arbitrary length, each adaptively associated with image tokens, for reconstruction and subsequent generative modeling, and model images of various resolutions using the same length of latent codes. An overview of the model architecture is shown in Fig. 2.", + "bbox": [ + 89, + 584, + 483, + 809 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. SoftVQ-VAE", + "text_level": 1, + "bbox": [ + 89, + 818, + 227, + 834 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As discussed above, the high compression ratio of both KL-VAE and VQ-VAE usually results in a significant degradation of quality in reconstruction and latent space. To overcome these limitations, we propose SoftVQ-VAE, a simple mod", + "bbox": [ + 89, + 840, + 485, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ification to VQ-VAE, which maintains the advantages of learning codewords to capture the data distribution while bridging it with more representation capacity as a continuous tokenizer. The core idea of SoftVQ-VAE is to allow each latent code to adaptively aggregate multiple codewords from the learnable codebook. Consequently, it imposes a soft categorical distribution on the posterior from the encoder:", + "bbox": [ + 511, + 90, + 906, + 196 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\text {p o s t e r i o r :} q _ {\\phi} (\\mathbf {z} | \\mathbf {x}) = \\operatorname {S o f t m a x} \\left(- \\frac {\\| \\hat {\\mathbf {z}} - \\mathcal {C} \\| _ {2}}{\\tau}\\right) \\\\ \\text {l a t e n t :} \\mathbf {z} = q _ {\\phi} (\\mathbf {z} | \\mathbf {x}) \\mathcal {C} \\end{array} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 558, + 208, + 903, + 257 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nk l \\colon \\mathcal {L} _ {\\mathrm {k l}} = H (q _ {\\phi} (\\mathbf {z} | \\mathbf {x})) - H \\left(\\mathbb {E} _ {\\mathbf {x} \\sim p (\\mathbf {x})} q _ {\\phi} (\\mathbf {z} | \\mathbf {x})\\right)\n$$\n", + "text_format": "latex", + "bbox": [ + 560, + 261, + 852, + 280 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\tau$ is the temperature parameter controlling the sharpness of the softmax probability and is set to 0.07 by ablation. The derivation of $\\mathcal{L}_{kl}$ is shown in Appendix B.1.", + "bbox": [ + 511, + 291, + 906, + 335 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Despite this simple modification, SoftVQ-VAE can significantly reduce the latent token length while maintaining a high quality of the reconstruction and latent space. More importantly, with a soft categorical distribution, SoftVQ-VAE is fully-differentiable, and thus the encoder and codebook can be optimized directly from the reconstruction loss (and other losses). It also allows for easier regularization of various forms on the latent space that drastically improves its quality, and less hyper-parameter tuning without codebook and commit loss [43] in the reduced training objectives.", + "bbox": [ + 511, + 337, + 908, + 487 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Interpretation. The posterior and latent space of VQ-VAE can be interpreted with K-Means [67]. Hence, SoftVQ-VAE can be viewed as soft K-Means [22], with codewords as learnable prototypes in the latent space. This latent space can easily be extended to Gaussian Mixture Models (GMM) [86], which we refer to as GMMVQ-VAE. Instead of using fixed temperature, GMMVQ-VAE predicts data-dependent weights for the codeword prototypes, i.e., Gaussian means, to compute the soft categorical posterior: $q_{\\phi}(\\mathbf{z}|\\mathbf{x}) = \\mathrm{Softmax}(-\\omega (\\hat{\\mathbf{z}})\\| \\hat{\\mathbf{z}} -\\mathcal{C}\\| _2)$ . Although prior work has explored GMM-based VAEs [17, 47], our GMMVQ-VAE differs by learning discrete prototypes rather than continuous latent variables. While providing the additional benefit of interpretable Gaussian components in the latent space, GMMVQ achieves reconstruction quality and downstream generation performance similar to SoftVQ in our experiments, and thus we mainly adopt the simpler SoftVQ throughout this paper. More details are in Appendix B.2.", + "bbox": [ + 511, + 488, + 908, + 760 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Representation Alignment of Latent Space", + "text_level": 1, + "bbox": [ + 511, + 772, + 875, + 787 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "While the reconstruction quality of tokenizer is important, learning a high-quality latent space is more crucial for downstream denoising-based generative modeling [116]. Several recent approaches have explored aligning the latent code with pre-trained features through a contrastive loss [30, 62, 109] and initializing the codebook with pre-trained features [124]. However, imposing effective regularization", + "bbox": [ + 511, + 794, + 906, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "28361", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "on the latent remains a challenge due to the discrete quantization of VQ-VAE and the Gaussian constraints in KL-VAE.", + "bbox": [ + 89, + 90, + 485, + 119 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Thanks to the fully-differentiable property of SoftVQ-VAE, we can now directly impose regularization on the latent space. Inspired by the recent work of REPA [116], we propose aligning the representations of latent codes with pre-trained vision encoders to facilitate learning the latent space. Compared to REPA, which aligns the representation at intermediate layers of downstream generative models, our method of feature alignment in the tokenizer's latent space is equivalent to performing REPA at the input space of generative models. Specifically, to align the latent code with the image tokens from a pre-trained vision encoder, we replicate each latent token by $N / L$ times:", + "bbox": [ + 89, + 121, + 485, + 303 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {z} _ {r} = \\underbrace {\\left[ \\mathbf {z} ^ {[ 0 ]} , \\dots , \\mathbf {z} ^ {[ 0 ]} \\right.} _ {N / L \\text {t i m e s}}, \\underbrace {\\left. \\mathbf {z} ^ {[ 1 ]} , \\dots , \\mathbf {z} ^ {[ 1 ]} \\right.} _ {N / L \\text {t i m e s}}, \\dots , \\underbrace {\\left. \\mathbf {z} ^ {[ L ]} , \\dots , \\mathbf {z} ^ {[ L ]} \\right]} _ {N / L \\text {t i m e s}}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 313, + 483, + 353 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "and encourage the token-wise similarity with image tokens. We employ a projector multilayer perceptron (MLP) on top of the latent tokens to match the pre-trained feature $\\mathbf{y}_{*}$ :", + "bbox": [ + 89, + 364, + 483, + 411 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {a l i g n}} = \\frac {1}{N} \\sum_ {n = 1} ^ {N} \\operatorname {s i m} \\left(\\mathbf {y} _ {*} ^ {[ n ]}, \\mathbf {M L P} \\left(\\mathbf {z} _ {r} ^ {([ n ])}\\right)\\right). \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 140, + 422, + 483, + 465 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This alignment ensures the latent space of SoftVQ-VAE captures semantically discriminative features that are beneficial for training the downstream generative models, even if it does not directly translate into improved reconstruction performance from the tokenizer. However, as we will show in Sec. 4.4, the generation quality of trained generative models often depends more on the semantic structure of the latent space than on the tokenizer's reconstruction capabilities.", + "bbox": [ + 89, + 477, + 483, + 599 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Final Training Objective", + "text_level": 1, + "bbox": [ + 89, + 609, + 316, + 626 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The training objective of SoftVQ-VAE combines the reconstruction loss, perceptual loss [20, 46, 54, 120] and adversarial loss [33, 44] as in VQ [23], and representation alignment:", + "bbox": [ + 89, + 633, + 485, + 679 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {\\text {r e c o n}} + \\lambda_ {1} \\mathcal {L} _ {\\text {p e r c e p}} + \\lambda_ {2} \\mathcal {L} _ {\\text {a d v}} + \\lambda_ {3} \\mathcal {L} _ {\\text {a l i g n}} + \\lambda_ {4} \\mathcal {L} _ {\\mathrm {K L}}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 691, + 482, + 709 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda_1, \\lambda_2, \\lambda_3,$ and $\\lambda_4$ are hyper-parameters. Neither codebook nor commit loss as in VQ is needed. The perceptual loss $\\mathcal{L}_{\\mathrm{percep}}$ helps capture high-level perceptual features, $\\mathcal{L}_{\\mathrm{adv}}$ encourages the decoder to generate realistic images by removing the artifacts from the reconstruction loss only, $\\mathcal{L}_{\\mathrm{align}}$ guides the latent space to align with pre-trained features, and $\\mathcal{L}_{\\mathrm{kl}}$ as the KL divergence term in the ELBO.", + "bbox": [ + 89, + 722, + 485, + 828 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 844, + 223, + 859 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we validate the efficiency and effectiveness of SoftVQ-VAE with extensive experiments.", + "bbox": [ + 89, + 869, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experiments Setup", + "text_level": 1, + "bbox": [ + 513, + 90, + 696, + 107 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation Details of SoftVQ-VAE. We use the LlamaGen codebase [98] to build our SoftVQ-VAE. We instantiate 4 configurations: SoftVQ-S, SoftVQ-B, SoftVQ-BL, SoftVQ-L, with a total of 45M, 173M, 391M, and 608M parameters, respectively. Each configuration has variants of latent codes $L = 64$ and $L = 32$ . For the representation alignment, we choose DINOv2 [79] for the pre-trained features, similarly as Yu et al. [116] and Li et al. [62]. To learn a better semantic latent space, we initialize our encoder from the pre-trained DINOv2 weights. Alignment with other pretrained features is explored in Sec. 4.4. We train the tokenizers on ImageNet [15] of resolution $256 \\times 256$ and $512 \\times 512$ for 250K iterations. Similarly to Tian et al. [99] and Li et al. [62], we adopted the same frozen DINO-S [9, 79] discriminator with a similar architecture to StyleGAN [48, 49]. We use DiffAug [121], consistency regularization [119], and LeCAM regularization [101] for discriminator training as in [99]. For the training objective, we set $\\lambda_1 = 1.0$ , $\\lambda_1 = 0.2$ , $\\lambda_3 = 0.1$ , and $\\lambda_4 = 0.01$ , following previous common practice. More training details are shown in Appendix C.1.", + "bbox": [ + 511, + 112, + 906, + 414 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation Details of Generative Modeling. We select DiT [80], SiT [57], and MAR [60] for downstream denoising-based image generation tasks. For DiT and SiT, we set the patch size of DiT and SiT to 1 and use a 1D absolution position embedding. In our main experiments, we train DiT-XL and SiT-XL of 675M parameters for 3M steps, compared to 4M steps in REPA [116] and 7M steps vanilla version [72, 80]. We strictly follow their original training setup for other settings. For MAR-H, we train for 500 epochs and set the maximum learning rate to $2e-4$ since a higher learning rate leads to NAN issues. For other experiments, we simply train SiT-L of 458M parameters for 400K steps. More experimental details are provided in Appendix C.2.", + "bbox": [ + 511, + 415, + 908, + 609 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation. We use reconstruction Frechet Inception Distance (rFID) [37] on ImageNet validation set to evaluate the tokenizer. To evaluate the performance of generation tasks, we report generation FID (gFID), Inception Score (IS) [90], Precision and Recall [52] (in Appendix D.5), with and without classifier-free guidance (CFG) [40]. We measure the efficiency of the generative models by GLOPs of the model's forward pass on the latent codes of tokenizers, and training and inference throughput for the models using floating point 32 and a batch size of 64 on a single AMD MI250.", + "bbox": [ + 511, + 611, + 908, + 762 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Main Results", + "text_level": 1, + "bbox": [ + 513, + 772, + 651, + 786 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We present the main reconstruction and generation results on the ImageNet benchmark of resolution $256 \\times 256$ and $512 \\times 512$ in Tab. 1 and Tab. 2, respectively. We show that SoftVQ-VAE achieves reconstruction and generation performance (with generative models) comparable to leading systems with 32 and 64 tokens, while presenting a significant improvement in both training and inference efficiency.", + "bbox": [ + 511, + 795, + 906, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "28362", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/6f534cfe17f8f24ba92d11fc7553d01bfe979899be63eafbcb83ef2175368637.jpg", + "table_caption": [ + "Table 1. System-level comparison on ImageNet ${256} \\times {256}$ conditional generation. We compare with both diffusion-based models and auto-regressive models with different types of tokenizers. Grey denotes SoftVQ. † indicates results with DPM-Solver and 50 steps." + ], + "table_footnote": [], + "table_body": "
Gen. ModelTok. Model# Tokens ↓Tok. rFID ↓Gflops ↓Throughput (imgs/sec) ↑w/o CFGw/ CFG
gFID ↓IS ↑gFID ↓IS ↑
Auto-regressive Models
Taming-Trans. [23]VQ2567.94--5.20290.3--
RQ-Trans. [55]RQ2563.20908.917.853.80323.7--
MaskGIT [10]VQ2562.28--6.18182.1--
MAGE [59]VQ256---6.93195.8--
LlamaGen-3B [98]VQ2562.19781.562.90--3.06279.7
TiTok-S-128 [115]VQ1281.6133.036.50--1.97281.8
MAR-H [60]KL2561.22145.080.122.35227.81.55303.7
MAR-HSoftVQ-L320.6167.932.083.83211.22.54273.6
SoftVQ-BL640.6586.550.892.81218.31.93289.4
Diffusion-based Models
LDM-4 [88]KL40960.27157.920.3710.56103.53.60247.7
U-ViT-H/2†[4]128.890.98--2.29263.9
MDTv2-XL/2 [29]125.430.595.06155.61.58314.7
DiT-XL/2 [80]KL10240.6280.730.519.62121.52.27278.2
SiT-XL/2 [72]81.920.548.30131.72.06270.3
+ REPA [116]5.90157.81.42305.7
SoftVQ-B0.898.949.83113.83.91264.2
SoftVQ-BL0.6814.528.819.22115.83.78266.7
SoftVQ-L0.748.749.07117.23.69270.4
DiT-XLSoftVQ-B0.884.706.62129.23.29262.5
SoftVQ-BL0.6528.814.596.53131.93.11268.3
SoftVQ-L0.614.515.83141.32.93268.5
SiT-XLSoftVQ-B0.8910.287.99129.32.51301.3
SoftVQ-BL0.6814.7310.127.67133.92.44308.8
SoftVQ-L0.7410.117.59137.42.44310.6
SiT-XLSoftVQ-B0.885.515.98138.01.78279.0
SoftVQ-BL0.6529.235.425.80143.51.88287.9
SoftVQ-L0.615.395.35151.21.86293.6
", + "bbox": [ + 187, + 125, + 803, + 500 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Consistent high-quality reconstruction. SoftVQ variants achieve remarkable reconstruction quality with a high compression ratio. With only 32 and 64 tokens, our models maintain consistent rFID scores, i.e., 0.61-0.89 on the $256 \\times 256$ and 0.64-0.71 on $512 \\times 512$ benchmark, substantially outperforming VQ tokenizers with 256 tokens used in autoregressive models such as MaskGIT [10] and TiTok [115]. Our results are also comparable to KL tokenizers used in previous denoising-based models, using $32x$ fewer tokens. More discussion of different tokenizers is given in Sec. 4.3.", + "bbox": [ + 88, + 508, + 485, + 657 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Generation performance comparable to SOTA. Notably, with only 32 tokens, the generation performance of DiT-XL and SiT-XL trained on all variants of the proposed SoftVQ significantly outperforms DiT-XL/2 and SiT-XL/2 with KL tokenizers of 1024 tokens, without using CFG. DiT-XL with SoftVQ-L of 32 tokens presents a 0.62 improvement on FID and SiT-XL with SoftVQ-L of 32 tokens shows a 0.71 improvement on FID, without using CFG. The improvement is further strengthened with 64 tokens. SiT-XL with SoftVQ-L of 64 tokens outperforms SiT-XL/2 with REPA [116], without using CFG. Applying CFG, SiT-XL with SoftVQ of 64 tokens achieves a performance comparable to the leading systems with the best FID as 1.78 on 256 and 2.21 on 512 benchmark. Using MAR-H [60], SoftVQ-BL and SoftVQ-L with 64 and 32 tokens present slightly worse results, compared to the KL tokenizer of 256 tokens, possibly due to the", + "bbox": [ + 89, + 659, + 485, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "lowered learning rate. MAR-H and SiT-XL results on 512 benchmark with 32 tokens are included in Appendix D.2.", + "bbox": [ + 511, + 508, + 906, + 537 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Efficiency. The system-level comparison reveals SoftVQ's exceptional efficiency across multiple metrics. When integrated with MAR-H at $256 \\times 256$ resolution, SoftVQ reduces GFLOPs by $40\\%$ compared to standard KL tokenizer, i.e., 86.55 vs 145.08 GFLOPs with 64 tokens and further reducing to 67.93 GFLOPs with 32 tokens. SoftVQ enables DiT-XL and SiT-XL to achieve significantly lower computational costs i.e., 28.81 and 29.23 GFLOPs, respectively with 64 tokens, compared to their KL counterparts (80.73 GFLOPs with 1024 tokens). Importantly, these efficiency improvements come without compromising generation quality. SiT-XL trained on SoftVQ-BL with 64 tokens presents $55\\mathrm{x}$ throughput than the KL counterpart, while showing competitive IS and FID scores in all configurations. Not only do we require fewer training iterations to achieve a comparable performance to leading systems, i.e., 2.5M vs 7M steps of SiT-XL and DiT-XL, the high compression ratio of SoftVQ also improves the training throughput, reducing the time to train SiT-XL for 400K steps from 72 to 20 hours on 8 MI250.", + "bbox": [ + 511, + 537, + 908, + 825 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Comparison of Tokenizers", + "text_level": 1, + "bbox": [ + 511, + 833, + 754, + 849 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We compare the proposed SoftVQ-VAE with the concurrent efficient image tokenizers, i.e., TiTok [115] and DC-AE [11]. To show the superiority of SoftVQ-VAE in achieving a", + "bbox": [ + 511, + 854, + 906, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "28363", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/2e02e28a25eca197a9eff94e7290d8deef87f6a01fadf5c863ee97767a051901.jpg", + "table_caption": [ + "Table 2. System-level comparison on ImageNet ${512} \\times {512}$ conditional generation. We compare with both diffusion-based models and auto-regressive models with different types of tokenizers. Grey denotes SoftVQ. † indicates results with DPM-Solver and 50 steps." + ], + "table_footnote": [], + "table_body": "
Gen. ModelTok. Model# Tokens ↓Tok. rFID ↓Gflops ↓Throughput (imgs/sec) ↑w/o CFGw/ CFG
gFID ↓IS ↑gFID ↓IS ↑
Generative Adversarial Models
BigGAN [10]-------8.43177.9
StyleGAN-XL [48]-------2.41267.7
Auto-regressive Models
MaskGIT [10]VQ10241.97--7.32156.0--
TiTok-B [115]VQ1281.52----2.13261.2
MAR-HSoftVQ-BL640.7186.551.508.21152.93.42261.8
Diffusion-based Models
ADM [16]-----23.2458.063.85221.7
U-ViT-H/4† [4]128.920.58--4.05263.8
DiT-XL/2 [80]KL40960.62373.340.109.62121.53.04240.8
SiT-XL/2 [72]373.320.10--2.62252.2
DiT-XL [11]AE2560.2280.751.029.56-2.84-
UViT-H† [11]128.906.149.83-2.53-
UViT-H† [11]AE640.2232.9915.2312.26-2.66-
UViT-2B† [11]104.189.446.50-2.25-
SiT-XLSoftVQ-BL640.7129.235.387.96133.92.21290.5
", + "bbox": [ + 187, + 122, + 805, + 363 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "high compression ratio, we additionally train VQ-VAE and AE using the small (S) configuration and the same training recipe of SoftVQ. To validate the generation performance, we train a SiT-L for 400K steps and report gFID and IS on 256 ImageNet without using CFG, as shown in Tab. 3.", + "bbox": [ + 89, + 369, + 482, + 445 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Scalable performance. With a much smaller model size, i.e., 46M vs 390M, SoftVQ-S significantly outperforms both TiTok variants at 64 tokens, achieving better reconstruction quality with an rFID of 1.03 compared to 1.25, and better generation performance with a gFID of 11.24 and IS of 89.4 compared to the best TiTok gFID of 19.23 and IS of 61.8. Noteworthy is that the generation results SiT-L trained on SoftVQ with 32 tokens outperform those trained on KL tokenizer with 1024 tokens [72] by a large margin, i.e., 5.9 in gFID. Compared to DC-AE, SoftVQ-S demonstrates competitive scalability across token counts while maintaining significant generation quality even with fewer parameters.", + "bbox": [ + 88, + 446, + 482, + 626 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Less Lossy Property. SoftVQ-S exhibits remarkably consistent performance across different compression ratios. When reducing tokens from 256 to 32, SoftVQ-S maintains a minimal degradation in rFID from 0.80 to 1.24, significantly outperforming both VQ-S, from 1.45 to 10.97, and AE-S, from 1.15 to 2.01. The benefits of the fully-differentiable property of SoftVQ along with representation alignment are particularly evident in generation quality, where SoftVQ-S achieves substantially better gFID of 9.21-12.89, and IS scores of 79.5-93.6, compared to baselines, showing its robustness in preserving information at high compression ratios and a better latent space for generative models.", + "bbox": [ + 89, + 627, + 485, + 808 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Discussions on the Latent Space", + "text_level": 1, + "bbox": [ + 89, + 818, + 372, + 834 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We perform more analysis on the latent of SoftVQ here.", + "bbox": [ + 89, + 839, + 460, + 854 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Alignment with different initialization and target models. As shown in Tab. 4, our results demonstrate the effectiveness of alignment across various pre-trained models. The", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/5fafdd765b3041d841f0695ec4eb87bd1546e4c5578558b67ffaf8c892f6a39b.jpg", + "table_caption": [ + "Table 3. Comparison of tokenizers on class-conditional ${256} \\times {256}$ ImageNet. We report rFID, and gFID and IS of SiT-L without CFG." + ], + "table_footnote": [], + "table_body": "
Tokenizer# Params# TokensDim.rFID ↓SiT-L
gFID ↓IS ↑
KL [96]676M102440.6218.7972.0
TiTok-BL-KL [115]389M64161.2523.3554.7
TiTok-BL-VQ [115]390M64162.0619.2361.8
DC-AE-f32 [11]323M64320.6915.3676.1
DC-AE-f54 [11]677M161280.8124.6649.1
VQ-S46M25681.4519.0258.6
128162.6118.5162.4
64324.0420.0754.5
326410.9721.5747.6
AE-S46M25681.1522.1158.3
128161.3322.3158.5
64321.6425.5354.3
32642.0128.1845.5
SoftVQ-S46M25680.809.2193.6
128160.9210.1285.8
64321.0311.2489.4
32641.2412.8979.5
", + "bbox": [ + 514, + 398, + 903, + 641 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "encoder initialization and alignment with DINOv2-B [79] achieve superior performance with rFID 0.88 and IS 103.4, compared to using either component alone. Further improvements are observed with CLIP-B [82] and EVA-02-B [26] on reconstruction. We reveal that a better rFID does not necessarily translate to a better gFID. Instead, the latent space quality, reflected by linear probing accuracy, is more closely related to the performance of the generative models.", + "bbox": [ + 509, + 643, + 908, + 763 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Linear probing of tokenizer and generative model. Linear probing results in Fig. 3 (details in Appendix C.3) reveal SoftVQ's superior representation quality at both tokenizer and generative model. At the tokenizer level, the linear probing accuracy of SoftVQ variants (B, BL, L) consistently outperforms VQ-S and AE-S across all token counts, with the gaps widening at lower token counts, similarly observed by Yu et al. [115]. This advantage transfers to the generative model with the best linear probing accuracy on the", + "bbox": [ + 511, + 765, + 910, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "28364", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/f1f8bbfa8b3f0d1dbf7e46aee5045af4ff3e9be53c5d870701945fe5a5f5cb30.jpg", + "table_caption": [ + "Table 4. Representation alignment with different encoder initialization and the target alignment models using SoftVQ-B 64 tokens." + ], + "table_footnote": [], + "table_body": "
Enc. Init.Align. TargetrFID ↓L.P. ↑SiT-L
gFID ↓IS ↑
--0.985.4220.3389.4
-DINOv2-B [79]0.9341.0810.96101.9
DINOv2-B [79]-0.7311.8717.2091.5
DINOv2-B [79]DINOv2-B [79]0.8842.4210.13103.4
CLIP-B [82]CLIP-B [82]0.7547.129.89109.3
EVA-02-B [26]EVA-02-B [26]0.8144.2310.34112.7
", + "bbox": [ + 127, + 119, + 442, + 209 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "intermediate features of SiT trained with SoftVQ.", + "bbox": [ + 89, + 213, + 418, + 227 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Latent space visualization. UMAP visualizations latent space in Fig. 4 demonstrate SoftVQ's ability totain structured and discriminative latent representation minimal variance between encoder output $(\\hat{\\mathbf{z}})$ and d input (z). Comparing Fig. 4 (d) and (c), we also observe larger encoder can learn a more discriminative latent showing the superiority of SoftVQ's representation le", + "bbox": [ + 89, + 227, + 439, + 333 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Ablation Studies", + "text_level": 1, + "bbox": [ + 89, + 342, + 254, + 357 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We present a series of ablation studies in Append including SoftVQ variants with product quantization [45], residual quantization (RQ) [55], and GMMVQ book size, latent space size, and the temperature of S The results show the compatibility of SoftVQ with I RQ, with slightly improved performance. GMMVQ p results comparable to SoftVQ. Moreover, SoftVQ shows robustness to other parameters, such as codebook size and softmax temperature. We found that while increasing the dimension of the latent space can result in a significant improvement in reconstruction, it leads to deterioration in generation performance, possibly due to the learning difficulty of generative models with larger dimensions [92].", + "bbox": [ + 89, + 364, + 483, + 561 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Related Work", + "text_level": 1, + "bbox": [ + 89, + 574, + 232, + 589 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Image tokenization has emerged as a fundamental technique to bridge various vision tasks. Traditional approaches using autoencoders [39, 106] laid the groundwork by compressing images into latent representations. This field subsequently diverged into two main branches: generation-focused and understanding-focused approaches. Generation-oriented methods, such as VAE [84, 104] and VQ-GAN [23, 85], emphasized learning latent spaces for detail-preserving compression. These were further refined through variants [55, 73, 112, 124] that enhanced generation quality. Parallel developments in understanding-focused approaches leveraged Large Language Models (LLMs) [105] to create semantic representations for tasks like classification [21] and detection [125]. Recent work [62, 115] has demonstrated the potential of unifying these approaches, with subsequent research [34, 109] exploring the convergence of generation and understanding capabilities using a single tokenizer.", + "bbox": [ + 89, + 599, + 485, + 854 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Image generation have two popular paradigms: autoregressive and diffusion. Autoregressive models start from CNN-based architectures [103] and evolve to the transformer-", + "bbox": [ + 89, + 856, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c62c016c899d097d2b6ba4621e9d2176cf6e114101d22392b9e37a8a01bf5f81.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 535, + 90, + 883, + 215 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/44aa5ef680a0f6a0e40648c3631596a68ea02c174ac34dee31e57d756c0eee2c.jpg", + "image_caption": [ + "Figure 4. Visualization of $\\hat{\\mathbf{z}}$ (top), i.e., encoder output, and $\\mathbf{z}$ (bottom), i.e., decoder input, of (a) VQ-S 64; (b) SoftVQ-S 64; (c) SoftVQ-S 32; (d) SoftVQ-L 32, trained with latent space alignment." + ], + "image_footnote": [], + "bbox": [ + 532, + 271, + 883, + 388 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "based architectures [55, 66, 98, 105, 114] to enhance scalability. Latest works in visual autoregressive modeling, including VAR [99], MAR [60], and ImageFolder [62], have further enhanced generation efficiency. Diffusion models have transformed image generation since their introduction by Sohl-Dickstein et al. [93]. Key advancements by Nichol et al. [75], Dhariwal et al. [16], and Song et al. [95] have significantly improved model performance. A paradigm shift occurred with the adoption of latent space diffusion [88, 102], leveraging pre-trained image encoders as priors [23, 104] to enhance generation efficiency. Architectural innovations, particularly the integration of transformers [80], have further expanded the capabilities of these models. These advances have spawned numerous influential systems, from text-guided generation [18, 76] to high-fidelity synthesis [28, 89]. Commercial successes [74, 78, 83, 87] have also demonstrated the practical impact of diffusion models.", + "bbox": [ + 511, + 439, + 908, + 696 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 709, + 633, + 724 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We presented SoftVQ-VAE, a continuous image tokenizer leveraging soft categorical posteriors to aggregate multiple codewords into each latent token. Our approach compresses images to just 32 or 64 tokens while maintaining high reconstruction quality and enabling state-of-the-art generation results with DiT, SiT, and MAR. The substantial improvements in both inference throughput (up to $55\\mathrm{x}$ with 64 tokens) and training efficiency, combined with the fully-differentiable design's ability to learn semantic-rich representations, demonstrate SoftVQ-VAE's potential as a foundation for efficient generative and vision-language models.", + "bbox": [ + 511, + 734, + 908, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "28365", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Michael S Albergo, Nicholas M Boffi, and Eric Vanden-Eijnden. Stochastic interpolants: A unifying framework for flows and diffusions. arXiv preprint arXiv:2303.08797, 2023. 3", + "[2] Brian DO Anderson. Reverse-time diffusion equation models. Stochastic Processes and their Applications, 12(3): 313-326, 1982. 3", + "[3] Alexei Baevski, Yuhao Zhou, Abdelrahman Mohamed, and Michael Auli. wav2vec 2.0: A framework for self-supervised learning of speech representations. Advances in neural information processing systems, 33:12449-12460, 2020. 3", + "[4] Fan Bao, Shen Nie, Kaiwen Xue, Yue Cao, Chongxuan Li, Hang Su, and Jun Zhu. All are worth words: A vit backbone for diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 22669-22679, 2023. 6, 7, 5", + "[5] Hangbo Bao, Li Dong, Songhao Piao, and Furu Wei. Beit: Bert pre-training of image transformers. arXiv preprint arXiv:2106.08254, 2021. 4", + "[6] Omer Bar-Tal, Hila Chefer, Omer Tov, Charles Herrmann, Roni Paiss, Shiran Zada, Ariel Ephrat, Junhwa Hur, Guanghui Liu, Amit Raj, et al. Lumiere: A spacetime diffusion model for video generation. arXiv preprint arXiv:2401.12945, 2024. 1", + "[7] Yoshua Bengio, Nicholas Léonard, and Aaron Courville. Estimating or propagating gradients through stochastic neurons for conditional computation. arXiv preprint arXiv:1308.3432, 2013. 1, 3", + "[8] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 1", + "[9] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 5, 2", + "[10] Huiwen Chang, Han Zhang, Lu Jiang, Ce Liu, and William T. Freeman. Maskgit: Masked generative image transformer, 2022. 1, 3, 6, 7, 5", + "[11] Junyu Chen, Han Cai, Junsong Chen, Enze Xie, Shang Yang, Haotian Tang, Muyang Li, Yao Lu, and Song Han. Deep compression autoencoder for efficient high-resolution diffusion models. arXiv preprint arXiv:2410.10733, 2024. 1, 6, 7, 5", + "[12] Liang Chen, Haozhe Zhao, Tianyu Liu, Shuai Bai, Junyang Lin, Chang Zhou, and Baobao Chang. An image is worth 1/2 tokens after layer 2: Plug-and-play inference acceleration for large vision-language models. arXiv preprint arXiv:2403.06764, 2024. 1", + "[13] Xi Chen, Diederik P Kingma, Tim Salimans, Yan Duan, Prafulla Dhariwal, John Schulman, Ilya Sutskever, and Pieter Abbeel. Variational lossy autoencoder. arXiv preprint arXiv:1611.02731, 2016. 1, 2, 3" + ], + "bbox": [ + 101, + 114, + 485, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Xinlei Chen, Zhuang Liu, Saining Xie, and Kaiming He. Deconstructing denoising diffusion models for self-supervised learning. arXiv preprint arXiv:2401.14404, 2024. 1, 2, 3", + "[15] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248–255. IEEE, 2009. 5, 2", + "[16] Prafulla Dhariwal and Alex Nichol. Diffusion models beat gans on image synthesis, 2021. 1, 7, 8, 5", + "[17] Nat Dilokthanakul, Pedro AM Mediano, Marta Garnelo, Matthew CH Lee, Hugh Salimbeni, Kai Arulkumaran, and Murray Shanahan. Deep unsupervised clustering with gaussian mixture variational autoencoders. arXiv preprint arXiv:1611.02648, 2016. 4", + "[18] Ming Ding, Zhuoyi Yang, Wenyi Hong, Wendi Zheng, Chang Zhou, Da Yin, Junyang Lin, Xu Zou, Zhou Shao, Hongxia Yang, et al. Cogview: Mastering text-to-image generation via transformers. Advances in neural information processing systems, 34:19822-19835, 2021. 8", + "[19] Xiaoyi Dong, Jianmin Bao, Ting Zhang, Dongdong Chen, Weiming Zhang, Lu Yuan, Dong Chen, Fang Wen, Nenghai Yu, and Baining Guo. Peco: Perceptual codebook for bert pre-training of vision transformers. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 552-560, 2023. 2", + "[20] Alexey Dosovitskiy and Thomas Brox. Generating images with perceptual similarity metrics based on deep networks. Advances in neural information processing systems, 29, 2016. 5", + "[21] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale, 2021. 3, 8", + "[22] Joseph C Dunn. A fuzzy relative of the isodata process and its use in detecting compact well-separated clusters. 1973. 2, 4", + "[23] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12873-12883, 2021. 1, 2, 3, 5, 6, 8", + "[24] Zach Evans, CJ Carr, Josiah Taylor, Scott H Hawley, and Jordi Pons. Fast timing-conditioned latent audio diffusion. arXiv preprint arXiv:2402.04825, 2024. 1", + "[25] Lijie Fan, Tianhong Li, Siyang Qin, Yuzheng Li, Chen Sun, Michael Rubinstein, Deqing Sun, Kaiming He, and Yonglong Tian. Fluid: Scaling autoregressive text-to-image generative models with continuous tokens. arXiv preprint arXiv:2410.13863, 2024. 1", + "[26] Yuxin Fang, Quan Sun, Xinggang Wang, Tiejun Huang, Xinlong Wang, and Yue Cao. Eva-02: A visual representation for neon genesis. Image and Vision Computing, 149:105171, 2024. 7, 8", + "[27] Christopher Fifty, Ronald G Junkins, Dennis Duan, Aniketh Iger, Jerry W Liu, Ehsan Amid, Sebastian Thrun, and" + ], + "bbox": [ + 522, + 92, + 906, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "28366", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Christopher Ré. Restructuring vector quantization with the rotation trick. arXiv preprint arXiv:2410.06424, 2024. 2, 3", + "[28] Oran Gafni, Adam Polyak, Oron Ashual, Shelly Sheynin, Devi Parikh, and Yaniv Taigman. Make-a-scene: Scene-based text-to-image generation with human priors. In European Conference on Computer Vision, pages 89-106. Springer, 2022. 8", + "[29] Shanghua Gao, Pan Zhou, Ming-Ming Cheng, and Shuicheng Yan. Mdtv2: Masked diffusion transformer is a strong image synthesizer. arXiv preprint arXiv:2303.14389, 2023.6,5", + "[30] Yuying Ge, Yixiao Ge, Ziyun Zeng, Xintao Wang, and Ying Shan. Planting a seed of vision in large language model, 2023. 1, 2, 4", + "[31] Yuying Ge, Sijie Zhao, Ziyun Zeng, Yixiao Ge, Chen Li, Xintao Wang, and Ying Shan. Making llama see and draw with seed tokenizer. arXiv preprint arXiv:2310.01218, 2023. 2", + "[32] Shansan Gong, Mukai Li, Jiangtao Feng, Zhiyong Wu, and LingPeng Kong. Diffuseq: Sequence to sequence text generation with diffusion models. arXiv preprint arXiv:2210.08933, 2022. 1", + "[33] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139-144, 2020. 5", + "[34] Yuchao Gu, Xintao Wang, Yixiao Ge, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Rethinking the objectives of vector-quantized tokenizers for image synthesis, 2023. 2, 8", + "[35] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 1, 3, 4", + "[36] Byeongho Heo, Song Park, Dongyoon Han, and Sangwoo Yun. Rotary position embedding for vision transformer. In European Conference on Computer Vision, pages 289-305. Springer, 2025. 4", + "[37] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in Neural Information Processing Systems, 30, 2017. 5", + "[38] Irina Higgins, Loic Matthew, Arka Pal, Christopher P Burgess, Xavier Glorot, Matthew M Botvinick, Shakir Mohamed, and Alexander Lerchner. beta-vae: Learning basic visual concepts with a constrained variational framework. ICLR, 3, 2017. 1, 2, 3", + "[39] Geoffrey E Hinton and Ruslan R Salakhutdinov. Reducing the dimensionality of data with neural networks. science, 313(5786):504-507, 2006. 8", + "[40] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 5", + "[41] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 3" + ], + "bbox": [ + 99, + 92, + 483, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[42] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. Advances in Neural Information Processing Systems, 35:8633-8646, 2022. 1", + "[43] Minyoung Huh, Brian Cheung, Pulkit Agrawal, and Phillip Isola. Straightening out the straight-through estimator: Overcoming optimization challenges in vector quantized networks. In International Conference on Machine Learning, pages 14096-14113. PMLR, 2023. 3, 4", + "[44] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A. Efros. Image-to-image translation with conditional adversarial networks, 2018. 5", + "[45] Herve Jegou, Matthijs Douze, and Cordelia Schmid. Product quantization for nearest neighbor search. IEEE transactions on pattern analysis and machine intelligence, 33(1):117-128, 2010. 2, 8", + "[46] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14, pages 694-711. Springer, 2016. 5", + "[47] Matthew J Johnson, David K Duvenaud, Alex Wiltschko, Ryan P Adams, and Sandeep R Datta. Composing graphical models with neural networks for structured representations and fast inference. Advances in neural information processing systems, 29, 2016. 4", + "[48] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 5, 7, 2", + "[49] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8110-8119, 2020. 5, 2", + "[50] Diederik Kingma and Ruiqi Gao. Understanding diffusion objectives as the elbo with simple data augmentation. Advances in Neural Information Processing Systems, 36, 2024. 3", + "[51] Diederik P Kingma. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 1, 2, 3", + "[52] Tuomas Kynkänniemi, Tero Karras, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Improved precision and recall metric for assessing generative models. Advances in neural information processing systems, 32, 2019. 5", + "[53] Tuomas Kynkänniemi, Miika Aittala, Tero Karras, Samuli Laine, Timo Aila, and Jaakko Lehtinen. Applying guidance in a limited interval improves sample and distribution quality in diffusion models. arXiv preprint arXiv:2404.07724, 2024. 2", + "[54] Anders Boesen Lindbo Larsen, Søren Kaae Sønderby, Hugo Larochelle, and Ole Winther. Autoencoding beyond pixels using a learned similarity metric. In International conference on machine learning, pages 1558-1566. PMLR, 2016. 5", + "[55] Doyup Lee, Chiheon Kim, Saehoon Kim, Minsu Cho, and Wook-Shin Han. Autoregressive image generation using" + ], + "bbox": [ + 522, + 92, + 906, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "28367", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "residual quantization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11523-11532, 2022. 2, 6, 8, 5", + "[56] Doyup Lee, Chiheon Kim, Saehoon Kim, Minsu Cho, and Wook-Shin Han. Autoregressive image generation using residual quantization, 2022. 1", + "[57] Haopeng Li, Jinyue Yang, Kexin Wang, Xuerui Qiu, Yuhong Chou, Xin Li, and Guoqi Li. Scalable autoregressive image generation with mamba. arXiv preprint arXiv:2408.12245, 2024. 5", + "[58] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730-19742. PMLR, 2023. 1, 2", + "[59] Tianhong Li, Huiwen Chang, Shlok Kumar Mishra, Han Zhang, Dina Katabi, and Dilip Krishnan. Mage: Masked generative encoder to unify representation learning and image synthesis, 2023. 3, 6, 5", + "[60] Tianhong Li, Yonglong Tian, He Li, Mingyang Deng, and Kaiming He. Autoregressive image generation without vector quantization, 2024. 1, 2, 3, 5, 6, 8", + "[61] Wentong Li, Yuqian Yuan, Jian Liu, Dongqi Tang, Song Wang, Jianke Zhu, and Lei Zhang. Tokenpacker: Efficient visual projector for multimodal llm. arXiv preprint arXiv:2407.02392, 2024. 1", + "[62] Xiang Li, Hao Chen, Kai Qiu, Jason Kuen, Jiumiang Gu, Bhiksha Raj, and Zhe Lin. Imagefolder: Autoregressive image generation with folded tokens. arXiv preprint arXiv:2410.01756, 2024. 1, 2, 4, 5, 8", + "[63] Yanwei Li, Yuechen Zhang, Chengyao Wang, Zhisheng Zhong, Yixin Chen, Ruihang Chu, Shaoteng Liu, and Jiaya Jia. Mini-gemini: Mining the potential of multi-modality vision language models. arXiv preprint arXiv:2403.18814, 2024. 1", + "[64] Ziyi Lin, Chris Liu, Renrui Zhang, Peng Gao, Longtian Qiu, Han Xiao, Han Qiu, Chen Lin, Wenqi Shao, Keqin Chen, et al. *Sphinx: The joint mixing of weights, tasks, and visual embeddings for multi-modal large language models.* arXiv preprint arXiv:2311.07575, 2023. 2", + "[65] Yaron Lipman, Ricky TQ Chen, Heli Ben-Hamu, Maximilian Nickel, and Matt Le. Flow matching for generative modeling. arXiv preprint arXiv:2210.02747, 2022. 3", + "[66] Wenze Liu, Le Zhuo, Yi Xin, Sheng Xia, Peng Gao, and Xiangyu Yue. Customize your visual autoregressive recipe with set autoregressive modeling. arXiv preprint arXiv:2410.10511, 2024. 8", + "[67] Stuart Lloyd. Least squares quantization in pmc. IEEE transactions on information theory, 28(2):129-137, 1982. 2, 4", + "[68] I Loshchilov. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 2", + "[69] Cheng Lu and Yang Song. Simplifying, stabilizing and scaling continuous-time consistency models. arXiv preprint arXiv:2410.11081, 2024. 1", + "[70] Gen Luo, Yiyi Zhou, Yuxin Zhang, Xiawu Zheng, Xiaoshuai Sun, and Rongrong Ji. Feast your eyes: Mixture-of" + ], + "bbox": [ + 99, + 92, + 483, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "resolution adaptation for multimodal large language models. arXiv preprint arXiv:2403.03003, 2024.1", + "[71] Zhuoyan Luo, Fengyuan Shi, Yixiao Ge, Yujiu Yang, Limin Wang, and Ying Shan. Open-magtvit2: An open-source project toward democratizing auto-regressive visual generation. arXiv preprint arXiv:2409.04410, 2024. 2", + "[72] Nanye Ma, Mark Goldstein, Michael S Albergo, Nicholas M Boffi, Eric Vanden-Eijnden, and Saining Xie. Sit: Exploring flow and diffusion-based generative models with scalable interpolant transformers. arXiv preprint arXiv:2401.08740, 2024. 1, 2, 3, 5, 6, 7", + "[73] Fabian Mentzer, David Minnen, Eirikur Agustsson, and Michael Tschannen. Finite scalar quantization: Vq-vae made simple, 2023. 2, 8", + "[74] MidJourney Inc. Midjourney. Software available from MidJourney Inc., 2022. 8", + "[75] Alex Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models, 2021. 8", + "[76] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021.8", + "[77] Harry Nyquist. Certain topics in telegraph transmission theory. Transactions of the American Institute of Electrical Engineers, 47(2):617-644, 1928. 2", + "[78] OpenAI. Video generation models as world simulators. OpenAI Blog, 2024. 8", + "[79] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Russell Howes, Po-Yao Huang, Hu Xu, Vasu Sharma, Shang-Wen Li, Wojciech Galuba, Mike Rabbat, Mido Assran, Nicolas Ballas, Gabriel Synnaeve, Ishan Misra, Herve Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. Dinov2: Learning robust visual features without supervision, 2023. 5, 7, 8, 2", + "[80] William Peebles and Saining Xie. Scalable diffusion models with transformers, 2023. 1, 2, 3, 5, 6, 7, 8", + "[81] Vadim Popov, Ivan Vovk, Vladimir Gogoryan, Tasnima Sadekova, and Mikhail Kudinov. Grad-tts: A diffusion probabilistic model for text-to-speech. In International Conference on Machine Learning, pages 8599-8608. PMLR, 2021. 1", + "[82] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 7, 8", + "[83] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International conference on machine learning, pages 8821-8831. Pmlr, 2021. 1, 8", + "[84] Ali Razavi, Aaron Van den Oord, and Oriol Vinyals. Generating diverse high-fidelity images with vq-vae-2. Advances in neural information processing systems, 32, 2019. 1, 8" + ], + "bbox": [ + 522, + 92, + 906, + 900 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "28368", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[85] Ali Razavi, Aaron van den Oord, and Oriol Vinyals. Generating diverse high-fidelity images with vq-vae-2, 2019. 8", + "[86] Douglas A Reynolds et al. Gaussian mixture models. Encyclopedia of biometrics, 741(659-663), 2009. 2, 4", + "[87] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 8", + "[88] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models, 2022. 1, 6, 8, 5", + "[89] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in neural information processing systems, 35:36479-36494, 2022. 8", + "[90] Tim Salimans, Ian Goodfellow, Wojciech Zaremba, Vicki Cheung, Alec Radford, and Xi Chen. Improved techniques for training gans. Advances in Neural Information Processing Systems, 29, 2016. 5", + "[91] Claude Elwood Shannon. Communication in the presence of noise. Proceedings of the IRE, 37(1):10-21, 1949. 2", + "[92] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International conference on machine learning, pages 2256-2265. PMLR, 2015. 3, 8", + "[93] Jascha Sohl-Dickstein, Eric A. Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics, 2015. 8", + "[94] Dingjie Song, Wenjun Wang, Shunian Chen, Xidong Wang, Michael Guan, and Benyou Wang. Less is more: A simple yet effective token reduction method for efficient multimodal llms. arXiv preprint arXiv:2409.10994, 2024. 1", + "[95] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models, 2022. 8", + "[96] stabilityai. Sd vae ft ema, 2023. Accessed: 2023. 1, 3, 7", + "[97] Jianlin Su, Murtadha Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu. Roformer: Enhanced transformer with rotary position embedding. Neurocomputing, 568:127063, 2024. 4", + "[98] Peize Sun, Yi Jiang, Shoufa Chen, Shilong Zhang, Bingyue Peng, Ping Luo, and Zehuan Yuan. Autoregressive model beats diffusion: Llama for scalable image generation. arXiv preprint arXiv:2406.06525, 2024. 1, 5, 6, 8", + "[99] Keyu Tian, Yi Jiang, Zehuan Yuan, Bingyue Peng, and Liwei Wang. Visual autoregressive modeling: Scalable image generation via next-scale prediction, 2024. 3, 5, 8, 2", + "[100] Michael Tschannen, Cian Eastwood, and Fabian Mentzer. Givt: Generative infinite-vocabulary transformers. In European Conference on Computer Vision, pages 292-309. Springer, 2025. 3", + "[101] Hung-Yu Tseng, Lu Jiang, Ce Liu, Ming-Hsuan Yang, and Weilong Yang. Regularizing generative adversarial networks under limited data, 2021. 5, 2" + ], + "bbox": [ + 93, + 92, + 485, + 898 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[102] Arash Vahdat, Karsten Kreis, and Jan Kautz. Score-based generative modeling in latent space, 2021. 8", + "[103] Aaron Van den Oord, Nal Kalchbrenner, Lasse Espeholt, Oriol Vinyals, Alex Graves, et al. Conditional image generation with pixelCNN decoders. Advances in neural information processing systems, 29, 2016. 8", + "[104] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017. 1, 2, 3, 8", + "[105] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need, 2023. 1, 8", + "[106] Pascal Vincent, Hugo Larochelle, Yoshua Bengio, and Pierre-Antoine Manzagol. Extracting and composing robust features with denoising autoencoders. In Proceedings of the 25th international conference on Machine learning, pages 1096-1103, 2008. 8", + "[107] Xinlong Wang, Xiaosong Zhang, Zhengxiong Luo, Quan Sun, Yufeng Cui, Jinsheng Wang, Fan Zhang, Yueze Wang, Zhen Li, Qiying Yu, et al. Emu3: Next-token prediction is all you need. arXiv preprint arXiv:2409.18869, 2024. 1", + "[108] Mark Weber, Lijun Yu, Qihang Yu, Xueqing Deng, Xiaohui Shen, Daniel Cremers, and Liang-Chieh Chen. Maskbit: Embedding-free image generation via bit tokens. arXiv preprint arXiv:2409.16211, 2024. 2", + "[109] Yecheng Wu, Zhuoyang Zhang, Junyu Chen, Haotian Tang, Dacheng Li, Yunhao Fang, Ligeng Zhu, Enze Xie, Hongxu Yin, Li Yi, et al. Vila-u: a unified foundation model integrating visual understanding and generation. arXiv preprint arXiv:2409.04429, 2024. 1, 2, 4, 8", + "[110] Jinheng Xie, Weijia Mao, Zechen Bai, David Junhao Zhang, Weihao Wang, Kevin Qinghong Lin, Yuchao Gu, Zhijie Chen, Zhenheng Yang, and Mike Zheng Shou. Show-o: One single transformer to unify multimodal understanding and generation. arXiv preprint arXiv:2408.12528, 2024. 1", + "[111] Jiahui Yu, Xin Li, Jing Yu Koh, Han Zhang, Ruoming Pang, James Qin, Alexander Ku, Yuanzhong Xu, Jason Baldridge, and Yonghui Wu. Vector-quantized image modeling with improved vqgan. arXiv preprint arXiv:2110.04627, 2021. 2, 3", + "[112] Lijun Yu, José Lezama, Nitesh B Gundavarapu, Luca Versari, Kihyuk Sohn, David Minnen, Yong Cheng, Vighnesh Birodkar, Agrim Gupta, Xiuye Gu, et al. Language model beats diffusion-tokenizer is key to visual generation. arXiv preprint arXiv:2310.05737, 2023. 1, 2, 8", + "[113] Lijun Yu, Yong Cheng, Zhiruo Wang, Vivek Kumar, Wolfgang Macherey, Yanping Huang, David Ross, Irfan Essa, Yonatan Bisk, Ming-Hsuan Yang, et al. Spae: Semantic pyramid autoencoder for multimodal generation with frozen llms. Advances in Neural Information Processing Systems, 36, 2024. 2", + "[114] Qihang Yu, Ju He, Xueqing Deng, Xiaohui Shen, and Liang-Chieh Chen. Randomized autoregressive visual generation. arXiv preprint arXiv:2411.00776, 2024. 8", + "[115] Qihang Yu, Mark Weber, Xueqing Deng, Xiaohui Shen, Daniel Cremers, and Liang-Chieh Chen. An image is worth 32 tokens for reconstruction and generation. arxiv: 2406.07550, 2024. 1, 2, 4, 6, 7, 8, 5" + ], + "bbox": [ + 93, + 92, + 906, + 898 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "28369", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[116] Sihyun Yu, Sangkyung Kwak, Huiwon Jang, Jongheon Jeong, Jonathan Huang, Jinwoo Shin, and Saining Xie. Representation alignment for generation: Training diffusion transformers is easier than you think. arXiv preprint arXiv:2410.06940, 2024. 1, 2, 4, 5, 6", + "[117] Neil Zeghidour, Alejandro Luebs, Ahmed Omran, Jan Skoglund, and Marco Tagliasacchi. Soundstream: An end-to-end neural audio codec. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 30:495-507, 2021. 1", + "[118] Baoquan Zhang, Huaibin Wang, Chuyao Luo, Xutao Li, Guotao Liang, Yunming Ye, Xiaochen Qi, and Yao He. Codebook transfer with part-of-speech for vector-quantized image modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7757–7766, 2024. 2", + "[119] Han Zhang, Zizhao Zhang, Augustus Odena, and Honglak Lee. Consistency regularization for generative adversarial networks. arXiv preprint arXiv:1910.12027, 2019. 5, 2", + "[120] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric, 2018. 5", + "[121] Shengyu Zhao, Zhijian Liu, Ji Lin, Jun-Yan Zhu, and Song Han. Differentiable augmentation for data-efficient gan training. Advances in neural information processing systems, 33:7559-7570, 2020. 5, 2", + "[122] Lin Zheng, Jianbo Yuan, Lei Yu, and Lingpeng Kong. A reparameterized discrete diffusion model for text generation. arXiv preprint arXiv:2302.05737, 2023. 1", + "[123] Chunting Zhou, Lili Yu, Arun Babu, Kushal Tirumala, Michihiro Yasunaga, Leonid Shamis, Jacob Kahn, Xuezhe Ma, Luke Zettlemoyer, and Omer Levy. Transfusion: Predict the next token and diffuse images with one multi-modal model. arXiv preprint arXiv:2408.11039, 2024. 1", + "[124] Lei Zhu, Fangyun Wei, Yanye Lu, and Dong Chen. Scaling the codebook size of vqgan to 100,000 with a utilization rate of $99\\%$ . arXiv preprint arXiv:2406.11837, 2024. 2, 4, 8", + "[125] X Zhu, W Su, L Lu, B Li, X Wang, and J Dai. Deformable detr: Deformable transformers for end-to-end object detection. arxiv 2020. arXiv preprint arXiv:2010.04159, 2010.8", + "[126] Yongxin Zhu, Bocheng Li, Yifei Xin, and Linli Xu. Addressing representation collapse in vector quantized models with one linear layer. arXiv preprint arXiv:2411.02038, 2024. 2" + ], + "bbox": [ + 91, + 90, + 485, + 699 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "28370", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 12 + } +] \ No newline at end of file diff --git a/2025/SoftVQ-VAE_ Efficient 1-Dimensional Continuous Tokenizer/315fadef-93cf-49d1-961b-c20538f1f4de_model.json b/2025/SoftVQ-VAE_ Efficient 1-Dimensional Continuous Tokenizer/315fadef-93cf-49d1-961b-c20538f1f4de_model.json new file mode 100644 index 0000000000000000000000000000000000000000..7877b9b261a514843dcbd1d891f6ca2e6a1a34bb --- /dev/null +++ b/2025/SoftVQ-VAE_ Efficient 1-Dimensional Continuous Tokenizer/315fadef-93cf-49d1-961b-c20538f1f4de_model.json @@ -0,0 +1,2976 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.194, + 0.131, + 0.804, + 0.152 + ], + "angle": 0, + "content": "SoftVQ-VAE: Efficient 1-Dimensional Continuous Tokenizer" + }, + { + "type": "text", + "bbox": [ + 0.195, + 0.18, + 0.803, + 0.217 + ], + "angle": 0, + "content": "Hao Chen\\(^{1,2*}\\), Ze Wang\\(^{2}\\), Xiang Li\\(^{1}\\), Xineng Sun\\(^{2}\\), Fangyi Chen\\(^{1}\\), Jiang Liu\\(^{2}\\), Jindong Wang\\(^{3}\\), Bhiksha Raj\\(^{1,4}\\), Zicheng Liu\\(^{2}\\), Emad Barsoum\\(^{2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.297, + 0.218, + 0.698, + 0.234 + ], + "angle": 0, + "content": "\\(^{1}\\)Carnegie Mellon University, \\(^{2}\\)AMD, \\(^{3}\\)William & Mary, \\(^{4}\\)MBZUAI" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.269, + 0.327, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.3, + 0.486, + 0.649 + ], + "angle": 0, + "content": "Efficient image tokenization with high compression ratios remains a critical challenge for training generative models. We present SoftVQ-VAE, a continuous image tokenizer that leverages soft categorical posteriors to aggregate multiple codewords into each latent token, substantially increasing the representation capacity of the latent space. When applied to Transformer-based architectures, our approach compresses \\(256 \\times 256\\) and \\(512 \\times 512\\) images using as few as 32 or 64 1-dimensional tokens. Not only does SoftVQ-VAE show consistent and high-quality reconstruction, more importantly, it also achieves state-of-the-art and significantly faster image generation results across different denoising-based generative models. Remarkably, SoftVQ-VAE improves inference throughput by up to \\(18x\\) for generating \\(256 \\times 256\\) images and \\(55x\\) for \\(512 \\times 512\\) images while achieving competitive FID scores of 1.78 and 2.21 for SiT-XL. It also improves the training efficiency of the generative models by reducing the number of training iterations by \\(2.3x\\) while maintaining comparable performance. With its fully-differentiable design and semantic-rich latent space, our experiment demonstrates that SoftVQ-VAE achieves efficient tokenization without compromising generation quality, paving the way for more efficient generative models. Code and model are released1." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.675, + 0.222, + 0.691 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.701, + 0.486, + 0.866 + ], + "angle": 0, + "content": "Denoising-based generative modeling has witnessed remarkable progress with recent advances, such as Diffusion Transformers (DiT) [80], Scalable Interpolant Transformers (SiT) [72], and Masked Auto-Regressive models with diffusion loss (MAR) [60], to name a few. Denosing-based generative modeling not only presents impressive generation results on a wide range of modality, including natural language [32, 122], images [16, 25, 69, 83, 88], videos [6, 8, 42], and audios [24, 81], but also shows potential to unify the understanding and generation capabilities across modalities in multi-modal language models [107, 109, 110, 123]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.27, + 0.909, + 0.497 + ], + "angle": 0, + "content": "A core component of denoising-based generative models is the tokenizer [23, 38, 51, 56, 84, 104, 112, 117], which compresses the raw data of each modality into a set of latent tokens in either a discrete or continuous latent space. The compact latent space therefore allows for more efficient and better generative modeling [88]. Among previous efforts, Kullback-Leibler Variational Auto-Encoders (KL-VAE) [51] and Vector Quantized Variational Auto-Encoders (VQ-VAE) [23, 84, 104] stand out as representatives of tokenizers which introduce continuous and discrete latent spaces, respectively. The former constrains the latent space with a Gaussian distribution using re-parametrization [51] trick, and the latter makes its latent space a categorical discrete distribution with a codebook of finite vocabulary which requires straight-through estimation [7]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.509, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Although both KL-VAE and VQ-VAE (and their variants) have been predominantly adopted in many generative models [10, 80, 88, 98, 123], they still present two major challenges restricting the efficiency and effectiveness of generative modeling: (1) the difficulty of achieving a higher compression ratio [13] and (2) the worse discriminative representations in their latent space than other self-supervised methods [14, 35, 109, 116]. The efficiency of downstream generative models, particularly Transformer-based architectures [105], is fundamentally constrained by their quadratic complexity to the number of latent tokens. Current image tokenizers [23, 88, 96] typically compress \\(256 \\times 256\\) images to at least 256 tokens and \\(512 \\times 512\\) images to at least 1024 tokens, creating a significant computational bottleneck for both training and inference of generative models [60, 72, 80]. Many efforts have been made to reduce the number of latent tokens on the generative model side, such as merging [62, 72, 80], pooling [58, 61, 94], and others [12, 63, 70]. More studies have recently emerged to fundamentally reduce the token number of the tokenizer [11, 30, 62, 115]. For example, TiTok [115] uses 128 tokens, achieving generation results comparable to 256 tokens by adding one extra decoder. DC-AE [11] compresses the initial 1024 tokens of \\(512 \\times 512\\) with 256 tokens. However, further increasing the compression ratio results in a significant degradation in the quality of reconstruction and therefore the generation [11]." + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.875, + 0.483, + 0.888 + ], + "angle": 0, + "content": "*Work done during internship at AMD. Email: haoc3@andrew.cmu.edu." + }, + { + "type": "page_footnote", + "bbox": [ + 0.111, + 0.888, + 0.451, + 0.9 + ], + "angle": 0, + "content": "1https://github.com/Hhhhhhao/continuous_tokenizer." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "28358" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.111, + 0.089, + 0.885, + 0.389 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.097, + 0.39, + 0.9, + 0.405 + ], + "angle": 0, + "content": "Figure 1. ImageNet-1K \\(256\\times 256\\) and \\(512\\times 512\\) generation results of generative models trained on SoftVQ-VAE with 32 and 64 tokens." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.413, + 0.484, + 0.625 + ], + "angle": 0, + "content": "Beyond the compression challenge, the quality of the learned representations of the tokenizer poses another critical limitation. While latent representations are crucial for both understanding [31, 109] and generation tasks [30, 62, 64, 116], current tokenizers usually struggle to capture more discriminative features. Earlier methods attempted to advance the training objectives and recipe to improve the latent space of the tokenizer [14, 19, 23, 34, 124]. Recently, [30, 62, 109, 113, 116, 118, 124] has also explored the representation alignment of latent tokens. However, due to the lossy nature [77, 91] of the smoothness constraints in KL-VAE and the discrete quantization in VQVAE [13, 27, 38, 126], current tokenizers usually fall short in learning the semantics of the latent space [109]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.63, + 0.487, + 0.902 + ], + "angle": 0, + "content": "To address the aforementioned challenges, we present SoftVQ-VAE, a simple modification to VQ-VAE that converts it from a discrete tokenizer into a continuous one with a high compression ratio and a semantic-rich latent space. Specifically, we propose using a soft categorical posterior in VAE with a learnable codebook. This straightforward modification introduces a crucial capability: instead of the conventional one-to-one mapping between tokens and codewords in VQ-VAE, our approach enables the adaptive aggregation of multiple codewords into each latent token, substantially increasing the representation capacity of the latent space. We apply SoftVQ-VAE to the Transformer auto-encoder architecture [30, 58, 62, 111, 115], and we succeed in using much fewer 1-dimensional latent tokens (i.e., 32 and 64) for both the reconstruction and subsequent generation tasks. Since SoftVQ-VAE is now fully differentiable, it not only simplifies the learning of the codebook, not requiring the codebook loss or commit loss as in VQ-VAE [23, 104], but" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.413, + 0.909, + 0.549 + ], + "angle": 0, + "content": "also enables better representation learning through direct alignment with pre-trained semantic features using a simple cosine similarity objective. Our approach generalizes the K-Means [67] latent space of VQ-VAE to Soft K-Means [22] and further to Gaussian Mixture Models (GMM) [86], while maintaining compatibility with existing VQ-VAE techniques [45, 55, 62, 112]. We comprehensively demonstrate the efficiency and effectiveness of the proposed SoftVQ-VAE in image reconstruction and generation tasks, and show that:" + }, + { + "type": "text", + "bbox": [ + 0.516, + 0.554, + 0.907, + 0.599 + ], + "angle": 0, + "content": "- SoftVQ-VAE presents more consistent, robust, and high-quality reconstruction and generation results with various latent token lengths ranging from 256 to 32." + }, + { + "type": "text", + "bbox": [ + 0.516, + 0.6, + 0.907, + 0.674 + ], + "angle": 0, + "content": "- SoftVQ-VAE allows diffusion-based DiT [73], flow-based SiT [72], and autoregressive-based MAR [60] to achieve state-of-the-art generation results with only 64 and 32 tokens on both the \\(256 \\times 256\\), with an FID of 1.78, and \\(512 \\times 512\\) on ImageNet benchmark, with an FID of 2.21." + }, + { + "type": "text", + "bbox": [ + 0.516, + 0.675, + 0.909, + 0.735 + ], + "angle": 0, + "content": "- SoftVQ-VAE improves the inference throughput by 18x and 10x with 32 and 64 tokens to generate \\(256 \\times 256\\) images, and 55x with 64 tokens for generating \\(512 \\times 512\\) images, with significantly lower GFLOPs for all models." + }, + { + "type": "text", + "bbox": [ + 0.516, + 0.735, + 0.909, + 0.795 + ], + "angle": 0, + "content": "- SoftVQ-VAE presents a more discriminative latent space through representation alignment. It also facilitates training efficiency by significantly improving training throughput and reduces training iterations by 2.3x." + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.554, + 0.909, + 0.795 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.825, + 0.64, + 0.843 + ], + "angle": 0, + "content": "2. Preliminary" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.91, + 0.903 + ], + "angle": 0, + "content": "We present an overview of the tokenizers, i.e., KL-VAE [51] and VQ-VAE [104], and denoising-based generative models, i.e., DiT [80], SiT [72], and MAR [60], in this section." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "28359" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.258, + 0.108 + ], + "angle": 0, + "content": "2.1. Image Tokenizer" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.113, + 0.486, + 0.309 + ], + "angle": 0, + "content": "The architecture of image tokenizer generally resembles auto-encoders, consisting of two main components: an encoder \\(\\mathcal{E}\\) and a decoder \\(\\mathcal{D}\\), parameterized by \\(\\phi\\) and \\(\\theta\\), respectively. Given an input image \\(\\mathbf{x} \\in \\mathbb{R}^{H \\times W \\times 3}\\), the encoder \\(\\mathcal{E}\\) maps the high-dimensional \\(\\mathbf{x}\\) to a lower-dimensional latent representation \\(\\mathbf{z} = \\mathcal{E}(\\mathbf{x}; \\phi)\\). The latent representations can have varying shapes depending on the encoder architecture. We generally view the latent representation as a set of latent tokens \\(\\mathbf{z} = [\\mathbf{z}^{[0]}, \\mathbf{z}^{[1]}, \\dots, \\mathbf{z}^{[L]}] \\in \\mathbb{R}^{L \\times D}\\), where \\(L\\) is the number of tokens and \\(D\\) is the dimension of the latent representation. For basic auto-encoders (AE), the decoder takes latent tokens as input and reconstructs the original signal \\(\\hat{\\mathbf{x}} = \\mathcal{D}(\\mathbf{z}; \\theta)\\) by minimizing the reconstruction loss with \\(\\mathbf{x}\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.31, + 0.485, + 0.49 + ], + "angle": 0, + "content": "Variational auto-encoders (VAE) [51] extend the basic AE by introducing a probabilistic perspective on the latent space. Specifically, VAE approximates a posterior distribution \\( p(\\mathbf{z}|\\mathbf{x}) \\) with a learned distribution \\( q_{\\phi}(\\mathbf{z}|\\mathbf{x}) \\) from the encoder, assuming that \\( \\mathbf{x} \\) is generated by the unobserved latent \\( \\mathbf{z} \\) with a prior distribution \\( p(\\mathbf{z}) \\). The decoder instead takes a set of sampled latent tokens \\( \\mathbf{z} \\sim q_{\\phi}(\\mathbf{z}|\\mathbf{x}) \\) from the posterior and aims to learn the marginal likelihood of data in a generative process by optimizing the evidence lower bound (ELBO): \\( \\max_{\\phi,\\theta}\\mathbb{E}_{q_{\\phi}(\\mathbf{z}|\\mathbf{x})}\\log p_{\\theta}(\\mathbf{x}|\\mathbf{z}) - \\mathcal{L}_{\\mathrm{kl}}(q_{\\phi}(\\mathbf{z}|\\mathbf{x})|p(\\mathbf{z})) \\). To make the optimization tractable, assumptions have been made on the prior, leading to different variants of VAE." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.491, + 0.485, + 0.581 + ], + "angle": 0, + "content": "KL-VAE [38, 51] parametrizes both the prior and posterior as Gaussians. The prior \\( p(\\mathbf{z}) \\) is assumed to be the isotropic unit Gaussian \\( \\mathcal{N}(0,1) \\). The continuous latent code is parameterized with posterior mean \\( \\mu_{\\phi} \\) and variance \\( \\sigma_{\\phi}^{2} \\) predicted by the encoder using the \"re-parametrization\" trick with a noise variable \\( \\varepsilon \\) from standard Gaussian:" + }, + { + "type": "equation", + "bbox": [ + 0.13, + 0.59, + 0.381, + 0.606 + ], + "angle": 0, + "content": "\\[\n\\text {p o s t e r i o r :} q _ {\\phi} (\\mathbf {z} | \\mathbf {x}) = \\mathcal {N} (\\mathbf {z}; \\boldsymbol {\\mu} _ {\\phi} (\\mathbf {x}), \\boldsymbol {\\sigma} _ {\\phi} ^ {2} (\\mathbf {x}))\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.13, + 0.608, + 0.483, + 0.627 + ], + "angle": 0, + "content": "\\[\nl a t e n t: \\mathbf {z} = \\boldsymbol {\\mu} _ {\\phi} (\\mathbf {x}) + \\boldsymbol {\\sigma} _ {\\phi} (\\mathbf {x}) \\odot \\varepsilon , \\quad \\epsilon \\sim \\mathcal {N} (0, \\mathbf {I}) \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.13, + 0.627, + 0.443, + 0.651 + ], + "angle": 0, + "content": "\\[\nk l: \\mathcal {L} _ {\\mathrm {k l}} = - \\frac {1}{2} \\left(1 + \\log \\left(\\boldsymbol {\\sigma} _ {\\phi} ^ {2} (\\mathbf {x})\\right) - \\boldsymbol {\\mu} _ {\\phi} (\\mathbf {x}) ^ {2} - \\boldsymbol {\\sigma} _ {\\phi} ^ {2} (\\mathbf {x})\\right)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.661, + 0.485, + 0.752 + ], + "angle": 0, + "content": "While KL-VAE is widely used in diffusion models, it is known to have the \"posterior-collapse\" issue [13]. In addition, the KL loss weight usually imposes a trade-off between the quality of reconstruction and the smoothness of its latent space [100], preventing high compression ratio and the learning of semantics in the latent space of KL-VAE." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.753, + 0.485, + 0.814 + ], + "angle": 0, + "content": "VQ-VAE [23, 104], in contrast, generates latent code in a discrete space by taking the posterior as a \\(K\\)-way deterministic categorical distribution against a learnable codebook \\(\\mathcal{C} = [\\mathbf{c}^{[0]}, \\dots, \\mathbf{c}^{[K]}\\] \\(\\in \\mathbb{R}^{K \\times D}\\), with \\(\\mathbf{c}\\) as the codeword:" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.824, + 0.453, + 0.86 + ], + "angle": 0, + "content": "\\[\n\\text {p o s t e r i o r :} q _ {\\phi} (\\mathbf {z} = k | \\mathbf {x}) = \\left\\{ \\begin{array}{l l} 1 & \\text {i f} k = \\arg \\min _ {j} \\| \\hat {\\mathbf {z}} - \\mathbf {c} ^ {[ j ]} \\| _ {2} \\\\ 0 & \\text {o t h e r w i s e} \\end{array} \\right.\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.102, + 0.858, + 0.483, + 0.885 + ], + "angle": 0, + "content": "\\[\nl a t e n t: \\mathbf {z} = \\mathbf {c} ^ {[ k ]}, \\quad \\text {w h e r e} k = \\underset {j} {\\arg \\min } \\| \\hat {\\mathbf {z}} - \\mathbf {c} ^ {[ j ]} \\| _ {2} \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.102, + 0.889, + 0.199, + 0.904 + ], + "angle": 0, + "content": "\\[\nk l: \\mathcal {L} _ {\\mathrm {k l}} = \\log K\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.909, + 0.244 + ], + "angle": 0, + "content": "Due to the non-differentiable arg min, a \"straight-through\" trick [7] is adopted to approximate the gradient of the encoder output \\(\\hat{\\mathbf{z}}\\) by directly copying gradient from the decoder input \\(\\mathbf{z}\\). Since the reconstruction objective does not impose direct gradients on the codebook \\(\\mathcal{C}\\), VQ-VAE additionally uses a codebook loss as \\(\\| \\mathrm{sg}[\\hat{\\mathbf{z}}] - \\mathbf{c}\\| _2\\) to move codewords toward the encoder output, along with a commit loss \\(\\| \\hat{\\mathbf{z}} -\\mathrm{sg}[\\mathbf{c}]\\| _2\\) to prevent arbitrary growth of codewords, where \\(\\mathrm{sg}[\\cdot ]\\) denotes stop-gradient. The broken gradient hinders high compression ratio and latent space learning [27, 43]." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.253, + 0.826, + 0.269 + ], + "angle": 0, + "content": "2.2. Denoising-based Generative Models" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.275, + 0.908, + 0.321 + ], + "angle": 0, + "content": "Denoising-based generative models synthesize images by progressively transforming Gaussian noise \\(\\varepsilon \\sim \\mathcal{N}(0,\\mathbf{I})\\) into tokenizer latent codes \\(\\mathbf{z}\\) through a forward process [41, 92]:" + }, + { + "type": "equation", + "bbox": [ + 0.652, + 0.334, + 0.905, + 0.349 + ], + "angle": 0, + "content": "\\[\n\\mathbf {z} _ {t} = \\alpha_ {t} \\mathbf {z} + \\sigma_ {t} \\varepsilon , \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.361, + 0.908, + 0.407 + ], + "angle": 0, + "content": "where \\(\\alpha_{t}\\) and \\(\\sigma_{t}\\) are a decreasing and increasing function of \\(t\\), and \\(t \\in [0, T]\\) [50]. In this paper, we adopt three approaches to implementing this denoising process as follows." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.407, + 0.909, + 0.527 + ], + "angle": 0, + "content": "DiT [80] is a diffusion-based model with a transformer architecture. It formulates the process through a forward-time stochastic differential equation (SDE), where \\(\\mathbf{z}_t\\) converges to \\(\\mathcal{N}(0, \\mathbf{I})\\) as \\(t \\to T\\). Generation occurs via a reverse-time SDE [2], with the model trained to predict noise \\(\\varepsilon\\) at randomly sampled timesteps. Due to the long token length of the tokenizer [96] used in DiT, it has several variants to merge tokens at the input of the transformer backbone." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.528, + 0.909, + 0.618 + ], + "angle": 0, + "content": "SiT [72] instead uses stochastic interpolants [1] and performs denoising via probability flow ordinary differential equation (PF ODE). The model learns the ODE's velocity field [65] by minimizing mean-squared-error (MSE) loss. This velocity field defines a score function for generation, analogous to DiT's reverse-time SDE." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.618, + 0.909, + 0.709 + ], + "angle": 0, + "content": "MAR [60] combines diffusion with autoregressive generation. Unlike DiT and SiT's parallel denoising of all latent tokens, MAR adopts an encoder-decoder transformer architecture [35] that progressively denoises tokens in a \"next-set\" autoregressive manner [99], starting from masked tokens, similar to MaskGIT [10] and MAGE [59]." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.722, + 0.605, + 0.738 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.748, + 0.909, + 0.84 + ], + "angle": 0, + "content": "In this section, we present SoftVQ-VAE, a novel 1D continuous tokenizer with a high compression ratio and a semantic-rich latent space. We first introduce the architecture and the formulation of SoftVQ-VAE. Then, we show that SoftVQ-VAE can learn semantics easily by aligning its latent tokens with pre-trained features via its fully-differentiable property." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.849, + 0.649, + 0.863 + ], + "angle": 0, + "content": "3.1. Architecture" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.909, + 0.901 + ], + "angle": 0, + "content": "We leverage Vision Transformer (ViT) [21, 111] as the architecture for the encoder \\(\\mathcal{E}\\) and decoder \\(\\mathcal{D}\\) of SoftVQ-VAE." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "28360" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.098, + 0.089, + 0.478, + 0.307 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.31, + 0.484, + 0.366 + ], + "angle": 0, + "content": "Figure 2. Illustration of SoftVQ-VAE. Left: Transformer encoder-decoder architecture with image tokens, arbitrary length of latent tokens, and mask tokens. Right top: fully-differentiable SoftVQ illustration. Right bottom: latent space representation alignment." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.372, + 0.484, + 0.585 + ], + "angle": 0, + "content": "Similarly to Yu et al. [115], instead of using the image features as latent, we initialize a set of extra 1D learnable tokens and use these tokens for reconstruction and subsequent generation. With the self-attention mechanism, it allows the learnable tokens to adaptively aggregate different image tokens to obtain the latent tokens of SoftVQ, detailed in Sec. 3.2. Specifically, for the encoder transformer with a patch size of \\( P \\), an image \\( \\mathbf{x} \\in \\mathbb{R}^{H \\times W \\times 3} \\) is patched into a sequence of image tokens \\( \\mathbf{x}_p \\in \\mathbb{R}^{N \\times D} \\), where \\( N = HW / P^2 \\) and \\( D \\) represent the token dimension size. We then concatenate a set of learnable tokens \\( \\mathbf{z}_l \\in \\mathbb{R}^{L \\times D} \\) of length \\( L \\) along with the image tokens as the input to the encoder, and retain only the output corresponding to the learnable latent tokens as the output of our encoder: \\( \\hat{\\mathbf{z}} = \\mathcal{E}([\\mathbf{x}_p; \\mathbf{z}_l]; \\phi) \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.585, + 0.484, + 0.81 + ], + "angle": 0, + "content": "For the decoder, we reconstruct images from a sequence of learnable mask tokens \\(\\mathbf{m} \\in \\mathbb{R}^{N \\times D}\\) [5, 35], concatenated with latent tokens: \\(\\hat{\\mathbf{x}} = \\mathcal{D}([\\mathbf{m};\\mathbf{z}];\\theta)\\). We use a linear layer at the end of the decoder to regress the pixel values from the mask tokens. For the image tokens at the encoder input and mask tokens at the decoder input, we apply 2D absolute position embeddings and 1D absolute position embedding for the latent tokens. Adopting more advanced position embedding techniques, such as RoPE [36, 97], may lead to better results and is left for future work. This design allows SoftVQ-VAE to use latent codes of arbitrary length, each adaptively associated with image tokens, for reconstruction and subsequent generative modeling, and model images of various resolutions using the same length of latent codes. An overview of the model architecture is shown in Fig. 2." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.819, + 0.228, + 0.835 + ], + "angle": 0, + "content": "3.2. SoftVQ-VAE" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.841, + 0.486, + 0.902 + ], + "angle": 0, + "content": "As discussed above, the high compression ratio of both KL-VAE and VQ-VAE usually results in a significant degradation of quality in reconstruction and latent space. To overcome these limitations, we propose SoftVQ-VAE, a simple mod" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.908, + 0.198 + ], + "angle": 0, + "content": "ification to VQ-VAE, which maintains the advantages of learning codewords to capture the data distribution while bridging it with more representation capacity as a continuous tokenizer. The core idea of SoftVQ-VAE is to allow each latent code to adaptively aggregate multiple codewords from the learnable codebook. Consequently, it imposes a soft categorical distribution on the posterior from the encoder:" + }, + { + "type": "equation", + "bbox": [ + 0.56, + 0.209, + 0.905, + 0.258 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\text {p o s t e r i o r :} q _ {\\phi} (\\mathbf {z} | \\mathbf {x}) = \\operatorname {S o f t m a x} \\left(- \\frac {\\| \\hat {\\mathbf {z}} - \\mathcal {C} \\| _ {2}}{\\tau}\\right) \\\\ \\text {l a t e n t :} \\mathbf {z} = q _ {\\phi} (\\mathbf {z} | \\mathbf {x}) \\mathcal {C} \\end{array} \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.562, + 0.262, + 0.853, + 0.281 + ], + "angle": 0, + "content": "\\[\nk l \\colon \\mathcal {L} _ {\\mathrm {k l}} = H (q _ {\\phi} (\\mathbf {z} | \\mathbf {x})) - H \\left(\\mathbb {E} _ {\\mathbf {x} \\sim p (\\mathbf {x})} q _ {\\phi} (\\mathbf {z} | \\mathbf {x})\\right)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.292, + 0.908, + 0.336 + ], + "angle": 0, + "content": "where \\(\\tau\\) is the temperature parameter controlling the sharpness of the softmax probability and is set to 0.07 by ablation. The derivation of \\(\\mathcal{L}_{kl}\\) is shown in Appendix B.1." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.338, + 0.909, + 0.488 + ], + "angle": 0, + "content": "Despite this simple modification, SoftVQ-VAE can significantly reduce the latent token length while maintaining a high quality of the reconstruction and latent space. More importantly, with a soft categorical distribution, SoftVQ-VAE is fully-differentiable, and thus the encoder and codebook can be optimized directly from the reconstruction loss (and other losses). It also allows for easier regularization of various forms on the latent space that drastically improves its quality, and less hyper-parameter tuning without codebook and commit loss [43] in the reduced training objectives." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.489, + 0.909, + 0.761 + ], + "angle": 0, + "content": "Interpretation. The posterior and latent space of VQ-VAE can be interpreted with K-Means [67]. Hence, SoftVQ-VAE can be viewed as soft K-Means [22], with codewords as learnable prototypes in the latent space. This latent space can easily be extended to Gaussian Mixture Models (GMM) [86], which we refer to as GMMVQ-VAE. Instead of using fixed temperature, GMMVQ-VAE predicts data-dependent weights for the codeword prototypes, i.e., Gaussian means, to compute the soft categorical posterior: \\( q_{\\phi}(\\mathbf{z}|\\mathbf{x}) = \\mathrm{Softmax}(-\\omega (\\hat{\\mathbf{z}})\\| \\hat{\\mathbf{z}} -\\mathcal{C}\\| _2) \\). Although prior work has explored GMM-based VAEs [17, 47], our GMMVQ-VAE differs by learning discrete prototypes rather than continuous latent variables. While providing the additional benefit of interpretable Gaussian components in the latent space, GMMVQ achieves reconstruction quality and downstream generation performance similar to SoftVQ in our experiments, and thus we mainly adopt the simpler SoftVQ throughout this paper. More details are in Appendix B.2." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.773, + 0.877, + 0.789 + ], + "angle": 0, + "content": "3.3. Representation Alignment of Latent Space" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.908, + 0.901 + ], + "angle": 0, + "content": "While the reconstruction quality of tokenizer is important, learning a high-quality latent space is more crucial for downstream denoising-based generative modeling [116]. Several recent approaches have explored aligning the latent code with pre-trained features through a contrastive loss [30, 62, 109] and initializing the codebook with pre-trained features [124]. However, imposing effective regularization" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "28361" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.121 + ], + "angle": 0, + "content": "on the latent remains a challenge due to the discrete quantization of VQ-VAE and the Gaussian constraints in KL-VAE." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.122, + 0.486, + 0.304 + ], + "angle": 0, + "content": "Thanks to the fully-differentiable property of SoftVQ-VAE, we can now directly impose regularization on the latent space. Inspired by the recent work of REPA [116], we propose aligning the representations of latent codes with pre-trained vision encoders to facilitate learning the latent space. Compared to REPA, which aligns the representation at intermediate layers of downstream generative models, our method of feature alignment in the tokenizer's latent space is equivalent to performing REPA at the input space of generative models. Specifically, to align the latent code with the image tokens from a pre-trained vision encoder, we replicate each latent token by \\( N / L \\) times:" + }, + { + "type": "equation", + "bbox": [ + 0.101, + 0.314, + 0.484, + 0.354 + ], + "angle": 0, + "content": "\\[\n\\mathbf {z} _ {r} = \\underbrace {\\left[ \\mathbf {z} ^ {[ 0 ]} , \\dots , \\mathbf {z} ^ {[ 0 ]} \\right.} _ {N / L \\text {t i m e s}}, \\underbrace {\\left. \\mathbf {z} ^ {[ 1 ]} , \\dots , \\mathbf {z} ^ {[ 1 ]} \\right.} _ {N / L \\text {t i m e s}}, \\dots , \\underbrace {\\left. \\mathbf {z} ^ {[ L ]} , \\dots , \\mathbf {z} ^ {[ L ]} \\right]} _ {N / L \\text {t i m e s}}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.366, + 0.485, + 0.412 + ], + "angle": 0, + "content": "and encourage the token-wise similarity with image tokens. We employ a projector multilayer perceptron (MLP) on top of the latent tokens to match the pre-trained feature \\(\\mathbf{y}_{*}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.142, + 0.424, + 0.484, + 0.466 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {a l i g n}} = \\frac {1}{N} \\sum_ {n = 1} ^ {N} \\operatorname {s i m} \\left(\\mathbf {y} _ {*} ^ {[ n ]}, \\mathbf {M L P} \\left(\\mathbf {z} _ {r} ^ {([ n ])}\\right)\\right). \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.478, + 0.484, + 0.6 + ], + "angle": 0, + "content": "This alignment ensures the latent space of SoftVQ-VAE captures semantically discriminative features that are beneficial for training the downstream generative models, even if it does not directly translate into improved reconstruction performance from the tokenizer. However, as we will show in Sec. 4.4, the generation quality of trained generative models often depends more on the semantic structure of the latent space than on the tokenizer's reconstruction capabilities." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.611, + 0.317, + 0.627 + ], + "angle": 0, + "content": "3.4. Final Training Objective" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.634, + 0.486, + 0.68 + ], + "angle": 0, + "content": "The training objective of SoftVQ-VAE combines the reconstruction loss, perceptual loss [20, 46, 54, 120] and adversarial loss [33, 44] as in VQ [23], and representation alignment:" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.693, + 0.483, + 0.71 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {\\text {r e c o n}} + \\lambda_ {1} \\mathcal {L} _ {\\text {p e r c e p}} + \\lambda_ {2} \\mathcal {L} _ {\\text {a d v}} + \\lambda_ {3} \\mathcal {L} _ {\\text {a l i g n}} + \\lambda_ {4} \\mathcal {L} _ {\\mathrm {K L}}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.723, + 0.486, + 0.829 + ], + "angle": 0, + "content": "where \\(\\lambda_1, \\lambda_2, \\lambda_3,\\) and \\(\\lambda_4\\) are hyper-parameters. Neither codebook nor commit loss as in VQ is needed. The perceptual loss \\(\\mathcal{L}_{\\mathrm{percep}}\\) helps capture high-level perceptual features, \\(\\mathcal{L}_{\\mathrm{adv}}\\) encourages the decoder to generate realistic images by removing the artifacts from the reconstruction loss only, \\(\\mathcal{L}_{\\mathrm{align}}\\) guides the latent space to align with pre-trained features, and \\(\\mathcal{L}_{\\mathrm{kl}}\\) as the KL divergence term in the ELBO." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.845, + 0.224, + 0.861 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.901 + ], + "angle": 0, + "content": "In this section, we validate the efficiency and effectiveness of SoftVQ-VAE with extensive experiments." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.092, + 0.697, + 0.108 + ], + "angle": 0, + "content": "4.1. Experiments Setup" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.113, + 0.908, + 0.415 + ], + "angle": 0, + "content": "Implementation Details of SoftVQ-VAE. We use the LlamaGen codebase [98] to build our SoftVQ-VAE. We instantiate 4 configurations: SoftVQ-S, SoftVQ-B, SoftVQ-BL, SoftVQ-L, with a total of 45M, 173M, 391M, and 608M parameters, respectively. Each configuration has variants of latent codes \\( L = 64 \\) and \\( L = 32 \\). For the representation alignment, we choose DINOv2 [79] for the pre-trained features, similarly as Yu et al. [116] and Li et al. [62]. To learn a better semantic latent space, we initialize our encoder from the pre-trained DINOv2 weights. Alignment with other pretrained features is explored in Sec. 4.4. We train the tokenizers on ImageNet [15] of resolution \\( 256 \\times 256 \\) and \\( 512 \\times 512 \\) for 250K iterations. Similarly to Tian et al. [99] and Li et al. [62], we adopted the same frozen DINO-S [9, 79] discriminator with a similar architecture to StyleGAN [48, 49]. We use DiffAug [121], consistency regularization [119], and LeCAM regularization [101] for discriminator training as in [99]. For the training objective, we set \\( \\lambda_1 = 1.0 \\), \\( \\lambda_1 = 0.2 \\), \\( \\lambda_3 = 0.1 \\), and \\( \\lambda_4 = 0.01 \\), following previous common practice. More training details are shown in Appendix C.1." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.416, + 0.909, + 0.611 + ], + "angle": 0, + "content": "Implementation Details of Generative Modeling. We select DiT [80], SiT [57], and MAR [60] for downstream denoising-based image generation tasks. For DiT and SiT, we set the patch size of DiT and SiT to 1 and use a 1D absolution position embedding. In our main experiments, we train DiT-XL and SiT-XL of 675M parameters for 3M steps, compared to 4M steps in REPA [116] and 7M steps vanilla version [72, 80]. We strictly follow their original training setup for other settings. For MAR-H, we train for 500 epochs and set the maximum learning rate to \\(2e-4\\) since a higher learning rate leads to NAN issues. For other experiments, we simply train SiT-L of 458M parameters for 400K steps. More experimental details are provided in Appendix C.2." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.612, + 0.909, + 0.763 + ], + "angle": 0, + "content": "Evaluation. We use reconstruction Frechet Inception Distance (rFID) [37] on ImageNet validation set to evaluate the tokenizer. To evaluate the performance of generation tasks, we report generation FID (gFID), Inception Score (IS) [90], Precision and Recall [52] (in Appendix D.5), with and without classifier-free guidance (CFG) [40]. We measure the efficiency of the generative models by GLOPs of the model's forward pass on the latent codes of tokenizers, and training and inference throughput for the models using floating point 32 and a batch size of 64 on a single AMD MI250." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.773, + 0.653, + 0.787 + ], + "angle": 0, + "content": "4.2. Main Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.908, + 0.901 + ], + "angle": 0, + "content": "We present the main reconstruction and generation results on the ImageNet benchmark of resolution \\(256 \\times 256\\) and \\(512 \\times 512\\) in Tab. 1 and Tab. 2, respectively. We show that SoftVQ-VAE achieves reconstruction and generation performance (with generative models) comparable to leading systems with 32 and 64 tokens, while presenting a significant improvement in both training and inference efficiency." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "28362" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.908, + 0.12 + ], + "angle": 0, + "content": "Table 1. System-level comparison on ImageNet \\( {256} \\times {256} \\) conditional generation. We compare with both diffusion-based models and auto-regressive models with different types of tokenizers. Grey denotes SoftVQ. † indicates results with DPM-Solver and 50 steps." + }, + { + "type": "table", + "bbox": [ + 0.189, + 0.125, + 0.805, + 0.501 + ], + "angle": 0, + "content": "
Gen. ModelTok. Model# Tokens ↓Tok. rFID ↓Gflops ↓Throughput (imgs/sec) ↑w/o CFGw/ CFG
gFID ↓IS ↑gFID ↓IS ↑
Auto-regressive Models
Taming-Trans. [23]VQ2567.94--5.20290.3--
RQ-Trans. [55]RQ2563.20908.917.853.80323.7--
MaskGIT [10]VQ2562.28--6.18182.1--
MAGE [59]VQ256---6.93195.8--
LlamaGen-3B [98]VQ2562.19781.562.90--3.06279.7
TiTok-S-128 [115]VQ1281.6133.036.50--1.97281.8
MAR-H [60]KL2561.22145.080.122.35227.81.55303.7
MAR-HSoftVQ-L320.6167.932.083.83211.22.54273.6
SoftVQ-BL640.6586.550.892.81218.31.93289.4
Diffusion-based Models
LDM-4 [88]KL40960.27157.920.3710.56103.53.60247.7
U-ViT-H/2†[4]128.890.98--2.29263.9
MDTv2-XL/2 [29]125.430.595.06155.61.58314.7
DiT-XL/2 [80]KL10240.6280.730.519.62121.52.27278.2
SiT-XL/2 [72]81.920.548.30131.72.06270.3
+ REPA [116]5.90157.81.42305.7
SoftVQ-B0.898.949.83113.83.91264.2
SoftVQ-BL0.6814.528.819.22115.83.78266.7
SoftVQ-L0.748.749.07117.23.69270.4
DiT-XLSoftVQ-B0.884.706.62129.23.29262.5
SoftVQ-BL0.6528.814.596.53131.93.11268.3
SoftVQ-L0.614.515.83141.32.93268.5
SiT-XLSoftVQ-B0.8910.287.99129.32.51301.3
SoftVQ-BL0.6814.7310.127.67133.92.44308.8
SoftVQ-L0.7410.117.59137.42.44310.6
SiT-XLSoftVQ-B0.885.515.98138.01.78279.0
SoftVQ-BL0.6529.235.425.80143.51.88287.9
SoftVQ-L0.615.395.35151.21.86293.6
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.509, + 0.486, + 0.659 + ], + "angle": 0, + "content": "Consistent high-quality reconstruction. SoftVQ variants achieve remarkable reconstruction quality with a high compression ratio. With only 32 and 64 tokens, our models maintain consistent rFID scores, i.e., 0.61-0.89 on the \\(256 \\times 256\\) and 0.64-0.71 on \\(512 \\times 512\\) benchmark, substantially outperforming VQ tokenizers with 256 tokens used in autoregressive models such as MaskGIT [10] and TiTok [115]. Our results are also comparable to KL tokenizers used in previous denoising-based models, using \\(32x\\) fewer tokens. More discussion of different tokenizers is given in Sec. 4.3." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.66, + 0.486, + 0.901 + ], + "angle": 0, + "content": "Generation performance comparable to SOTA. Notably, with only 32 tokens, the generation performance of DiT-XL and SiT-XL trained on all variants of the proposed SoftVQ significantly outperforms DiT-XL/2 and SiT-XL/2 with KL tokenizers of 1024 tokens, without using CFG. DiT-XL with SoftVQ-L of 32 tokens presents a 0.62 improvement on FID and SiT-XL with SoftVQ-L of 32 tokens shows a 0.71 improvement on FID, without using CFG. The improvement is further strengthened with 64 tokens. SiT-XL with SoftVQ-L of 64 tokens outperforms SiT-XL/2 with REPA [116], without using CFG. Applying CFG, SiT-XL with SoftVQ of 64 tokens achieves a performance comparable to the leading systems with the best FID as 1.78 on 256 and 2.21 on 512 benchmark. Using MAR-H [60], SoftVQ-BL and SoftVQ-L with 64 and 32 tokens present slightly worse results, compared to the KL tokenizer of 256 tokens, possibly due to the" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.509, + 0.908, + 0.539 + ], + "angle": 0, + "content": "lowered learning rate. MAR-H and SiT-XL results on 512 benchmark with 32 tokens are included in Appendix D.2." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.539, + 0.909, + 0.826 + ], + "angle": 0, + "content": "Efficiency. The system-level comparison reveals SoftVQ's exceptional efficiency across multiple metrics. When integrated with MAR-H at \\(256 \\times 256\\) resolution, SoftVQ reduces GFLOPs by \\(40\\%\\) compared to standard KL tokenizer, i.e., 86.55 vs 145.08 GFLOPs with 64 tokens and further reducing to 67.93 GFLOPs with 32 tokens. SoftVQ enables DiT-XL and SiT-XL to achieve significantly lower computational costs i.e., 28.81 and 29.23 GFLOPs, respectively with 64 tokens, compared to their KL counterparts (80.73 GFLOPs with 1024 tokens). Importantly, these efficiency improvements come without compromising generation quality. SiT-XL trained on SoftVQ-BL with 64 tokens presents \\(55\\mathrm{x}\\) throughput than the KL counterpart, while showing competitive IS and FID scores in all configurations. Not only do we require fewer training iterations to achieve a comparable performance to leading systems, i.e., 2.5M vs 7M steps of SiT-XL and DiT-XL, the high compression ratio of SoftVQ also improves the training throughput, reducing the time to train SiT-XL for 400K steps from 72 to 20 hours on 8 MI250." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.834, + 0.755, + 0.851 + ], + "angle": 0, + "content": "4.3. Comparison of Tokenizers" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.908, + 0.903 + ], + "angle": 0, + "content": "We compare the proposed SoftVQ-VAE with the concurrent efficient image tokenizers, i.e., TiTok [115] and DC-AE [11]. To show the superiority of SoftVQ-VAE in achieving a" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "28363" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.908, + 0.12 + ], + "angle": 0, + "content": "Table 2. System-level comparison on ImageNet \\( {512} \\times {512} \\) conditional generation. We compare with both diffusion-based models and auto-regressive models with different types of tokenizers. Grey denotes SoftVQ. † indicates results with DPM-Solver and 50 steps." + }, + { + "type": "table", + "bbox": [ + 0.189, + 0.123, + 0.807, + 0.364 + ], + "angle": 0, + "content": "
Gen. ModelTok. Model# Tokens ↓Tok. rFID ↓Gflops ↓Throughput (imgs/sec) ↑w/o CFGw/ CFG
gFID ↓IS ↑gFID ↓IS ↑
Generative Adversarial Models
BigGAN [10]-------8.43177.9
StyleGAN-XL [48]-------2.41267.7
Auto-regressive Models
MaskGIT [10]VQ10241.97--7.32156.0--
TiTok-B [115]VQ1281.52----2.13261.2
MAR-HSoftVQ-BL640.7186.551.508.21152.93.42261.8
Diffusion-based Models
ADM [16]-----23.2458.063.85221.7
U-ViT-H/4† [4]128.920.58--4.05263.8
DiT-XL/2 [80]KL40960.62373.340.109.62121.53.04240.8
SiT-XL/2 [72]373.320.10--2.62252.2
DiT-XL [11]AE2560.2280.751.029.56-2.84-
UViT-H† [11]128.906.149.83-2.53-
UViT-H† [11]AE640.2232.9915.2312.26-2.66-
UViT-2B† [11]104.189.446.50-2.25-
SiT-XLSoftVQ-BL640.7129.235.387.96133.92.21290.5
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.371, + 0.483, + 0.446 + ], + "angle": 0, + "content": "high compression ratio, we additionally train VQ-VAE and AE using the small (S) configuration and the same training recipe of SoftVQ. To validate the generation performance, we train a SiT-L for 400K steps and report gFID and IS on 256 ImageNet without using CFG, as shown in Tab. 3." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.447, + 0.483, + 0.627 + ], + "angle": 0, + "content": "Scalable performance. With a much smaller model size, i.e., 46M vs 390M, SoftVQ-S significantly outperforms both TiTok variants at 64 tokens, achieving better reconstruction quality with an rFID of 1.03 compared to 1.25, and better generation performance with a gFID of 11.24 and IS of 89.4 compared to the best TiTok gFID of 19.23 and IS of 61.8. Noteworthy is that the generation results SiT-L trained on SoftVQ with 32 tokens outperform those trained on KL tokenizer with 1024 tokens [72] by a large margin, i.e., 5.9 in gFID. Compared to DC-AE, SoftVQ-S demonstrates competitive scalability across token counts while maintaining significant generation quality even with fewer parameters." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.628, + 0.486, + 0.809 + ], + "angle": 0, + "content": "Less Lossy Property. SoftVQ-S exhibits remarkably consistent performance across different compression ratios. When reducing tokens from 256 to 32, SoftVQ-S maintains a minimal degradation in rFID from 0.80 to 1.24, significantly outperforming both VQ-S, from 1.45 to 10.97, and AE-S, from 1.15 to 2.01. The benefits of the fully-differentiable property of SoftVQ along with representation alignment are particularly evident in generation quality, where SoftVQ-S achieves substantially better gFID of 9.21-12.89, and IS scores of 79.5-93.6, compared to baselines, showing its robustness in preserving information at high compression ratios and a better latent space for generative models." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.819, + 0.373, + 0.835 + ], + "angle": 0, + "content": "4.4. Discussions on the Latent Space" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.461, + 0.855 + ], + "angle": 0, + "content": "We perform more analysis on the latent of SoftVQ here." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.485, + 0.901 + ], + "angle": 0, + "content": "Alignment with different initialization and target models. As shown in Tab. 4, our results demonstrate the effectiveness of alignment across various pre-trained models. The" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.369, + 0.907, + 0.397 + ], + "angle": 0, + "content": "Table 3. Comparison of tokenizers on class-conditional \\( {256} \\times {256} \\) ImageNet. We report rFID, and gFID and IS of SiT-L without CFG." + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.4, + 0.905, + 0.642 + ], + "angle": 0, + "content": "
Tokenizer# Params# TokensDim.rFID ↓SiT-L
gFID ↓IS ↑
KL [96]676M102440.6218.7972.0
TiTok-BL-KL [115]389M64161.2523.3554.7
TiTok-BL-VQ [115]390M64162.0619.2361.8
DC-AE-f32 [11]323M64320.6915.3676.1
DC-AE-f54 [11]677M161280.8124.6649.1
VQ-S46M25681.4519.0258.6
128162.6118.5162.4
64324.0420.0754.5
326410.9721.5747.6
AE-S46M25681.1522.1158.3
128161.3322.3158.5
64321.6425.5354.3
32642.0128.1845.5
SoftVQ-S46M25680.809.2193.6
128160.9210.1285.8
64321.0311.2489.4
32641.2412.8979.5
" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.644, + 0.909, + 0.765 + ], + "angle": 0, + "content": "encoder initialization and alignment with DINOv2-B [79] achieve superior performance with rFID 0.88 and IS 103.4, compared to using either component alone. Further improvements are observed with CLIP-B [82] and EVA-02-B [26] on reconstruction. We reveal that a better rFID does not necessarily translate to a better gFID. Instead, the latent space quality, reflected by linear probing accuracy, is more closely related to the performance of the generative models." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.766, + 0.911, + 0.902 + ], + "angle": 0, + "content": "Linear probing of tokenizer and generative model. Linear probing results in Fig. 3 (details in Appendix C.3) reveal SoftVQ's superior representation quality at both tokenizer and generative model. At the tokenizer level, the linear probing accuracy of SoftVQ variants (B, BL, L) consistently outperforms VQ-S and AE-S across all token counts, with the gaps widening at lower token counts, similarly observed by Yu et al. [115]. This advantage transfers to the generative model with the best linear probing accuracy on the" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "28364" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.09, + 0.486, + 0.118 + ], + "angle": 0, + "content": "Table 4. Representation alignment with different encoder initialization and the target alignment models using SoftVQ-B 64 tokens." + }, + { + "type": "table", + "bbox": [ + 0.128, + 0.121, + 0.444, + 0.21 + ], + "angle": 0, + "content": "
Enc. Init.Align. TargetrFID ↓L.P. ↑SiT-L
gFID ↓IS ↑
--0.985.4220.3389.4
-DINOv2-B [79]0.9341.0810.96101.9
DINOv2-B [79]-0.7311.8717.2091.5
DINOv2-B [79]DINOv2-B [79]0.8842.4210.13103.4
CLIP-B [82]CLIP-B [82]0.7547.129.89109.3
EVA-02-B [26]EVA-02-B [26]0.8144.2310.34112.7
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.214, + 0.419, + 0.228 + ], + "angle": 0, + "content": "intermediate features of SiT trained with SoftVQ." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.228, + 0.44, + 0.334 + ], + "angle": 0, + "content": "Latent space visualization. UMAP visualizations latent space in Fig. 4 demonstrate SoftVQ's ability totain structured and discriminative latent representation minimal variance between encoder output \\((\\hat{\\mathbf{z}})\\) and d input (z). Comparing Fig. 4 (d) and (c), we also observe larger encoder can learn a more discriminative latent showing the superiority of SoftVQ's representation le" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.343, + 0.255, + 0.358 + ], + "angle": 0, + "content": "4.5. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.365, + 0.485, + 0.562 + ], + "angle": 0, + "content": "We present a series of ablation studies in Append including SoftVQ variants with product quantization [45], residual quantization (RQ) [55], and GMMVQ book size, latent space size, and the temperature of S The results show the compatibility of SoftVQ with I RQ, with slightly improved performance. GMMVQ p results comparable to SoftVQ. Moreover, SoftVQ shows robustness to other parameters, such as codebook size and softmax temperature. We found that while increasing the dimension of the latent space can result in a significant improvement in reconstruction, it leads to deterioration in generation performance, possibly due to the learning difficulty of generative models with larger dimensions [92]." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.575, + 0.233, + 0.59 + ], + "angle": 0, + "content": "5. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.6, + 0.486, + 0.856 + ], + "angle": 0, + "content": "Image tokenization has emerged as a fundamental technique to bridge various vision tasks. Traditional approaches using autoencoders [39, 106] laid the groundwork by compressing images into latent representations. This field subsequently diverged into two main branches: generation-focused and understanding-focused approaches. Generation-oriented methods, such as VAE [84, 104] and VQ-GAN [23, 85], emphasized learning latent spaces for detail-preserving compression. These were further refined through variants [55, 73, 112, 124] that enhanced generation quality. Parallel developments in understanding-focused approaches leveraged Large Language Models (LLMs) [105] to create semantic representations for tasks like classification [21] and detection [125]. Recent work [62, 115] has demonstrated the potential of unifying these approaches, with subsequent research [34, 109] exploring the convergence of generation and understanding capabilities using a single tokenizer." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.857, + 0.485, + 0.901 + ], + "angle": 0, + "content": "Image generation have two popular paradigms: autoregressive and diffusion. Autoregressive models start from CNN-based architectures [103] and evolve to the transformer-" + }, + { + "type": "image", + "bbox": [ + 0.536, + 0.092, + 0.885, + 0.217 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.533, + 0.272, + 0.885, + 0.39 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.391, + 0.907, + 0.433 + ], + "angle": 0, + "content": "Figure 4. Visualization of \\(\\hat{\\mathbf{z}}\\) (top), i.e., encoder output, and \\(\\mathbf{z}\\) (bottom), i.e., decoder input, of (a) VQ-S 64; (b) SoftVQ-S 64; (c) SoftVQ-S 32; (d) SoftVQ-L 32, trained with latent space alignment." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.44, + 0.909, + 0.697 + ], + "angle": 0, + "content": "based architectures [55, 66, 98, 105, 114] to enhance scalability. Latest works in visual autoregressive modeling, including VAR [99], MAR [60], and ImageFolder [62], have further enhanced generation efficiency. Diffusion models have transformed image generation since their introduction by Sohl-Dickstein et al. [93]. Key advancements by Nichol et al. [75], Dhariwal et al. [16], and Song et al. [95] have significantly improved model performance. A paradigm shift occurred with the adoption of latent space diffusion [88, 102], leveraging pre-trained image encoders as priors [23, 104] to enhance generation efficiency. Architectural innovations, particularly the integration of transformers [80], have further expanded the capabilities of these models. These advances have spawned numerous influential systems, from text-guided generation [18, 76] to high-fidelity synthesis [28, 89]. Commercial successes [74, 78, 83, 87] have also demonstrated the practical impact of diffusion models." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.71, + 0.634, + 0.726 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.909, + 0.903 + ], + "angle": 0, + "content": "We presented SoftVQ-VAE, a continuous image tokenizer leveraging soft categorical posteriors to aggregate multiple codewords into each latent token. Our approach compresses images to just 32 or 64 tokens while maintaining high reconstruction quality and enabling state-of-the-art generation results with DiT, SiT, and MAR. The substantial improvements in both inference throughput (up to \\(55\\mathrm{x}\\) with 64 tokens) and training efficiency, combined with the fully-differentiable design's ability to learn semantic-rich representations, demonstrate SoftVQ-VAE's potential as a foundation for efficient generative and vision-language models." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "28365" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.115, + 0.486, + 0.169 + ], + "angle": 0, + "content": "[1] Michael S Albergo, Nicholas M Boffi, and Eric Vanden-Eijnden. Stochastic interpolants: A unifying framework for flows and diffusions. arXiv preprint arXiv:2303.08797, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.172, + 0.484, + 0.211 + ], + "angle": 0, + "content": "[2] Brian DO Anderson. Reverse-time diffusion equation models. Stochastic Processes and their Applications, 12(3): 313-326, 1982. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.215, + 0.484, + 0.282 + ], + "angle": 0, + "content": "[3] Alexei Baevski, Yuhao Zhou, Abdelrahman Mohamed, and Michael Auli. wav2vec 2.0: A framework for self-supervised learning of speech representations. Advances in neural information processing systems, 33:12449-12460, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.284, + 0.484, + 0.352 + ], + "angle": 0, + "content": "[4] Fan Bao, Shen Nie, Kaiwen Xue, Yue Cao, Chongxuan Li, Hang Su, and Jun Zhu. All are worth words: A vit backbone for diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 22669-22679, 2023. 6, 7, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.354, + 0.484, + 0.395 + ], + "angle": 0, + "content": "[5] Hangbo Bao, Li Dong, Songhao Piao, and Furu Wei. Beit: Bert pre-training of image transformers. arXiv preprint arXiv:2106.08254, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.397, + 0.484, + 0.464 + ], + "angle": 0, + "content": "[6] Omer Bar-Tal, Hila Chefer, Omer Tov, Charles Herrmann, Roni Paiss, Shiran Zada, Ariel Ephrat, Junhwa Hur, Guanghui Liu, Amit Raj, et al. Lumiere: A spacetime diffusion model for video generation. arXiv preprint arXiv:2401.12945, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.467, + 0.484, + 0.521 + ], + "angle": 0, + "content": "[7] Yoshua Bengio, Nicholas Léonard, and Aaron Courville. Estimating or propagating gradients through stochastic neurons for conditional computation. arXiv preprint arXiv:1308.3432, 2013. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.524, + 0.484, + 0.592 + ], + "angle": 0, + "content": "[8] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.594, + 0.484, + 0.661 + ], + "angle": 0, + "content": "[9] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 5, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.664, + 0.484, + 0.703 + ], + "angle": 0, + "content": "[10] Huiwen Chang, Han Zhang, Lu Jiang, Ce Liu, and William T. Freeman. Maskgit: Masked generative image transformer, 2022. 1, 3, 6, 7, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.706, + 0.484, + 0.773 + ], + "angle": 0, + "content": "[11] Junyu Chen, Han Cai, Junsong Chen, Enze Xie, Shang Yang, Haotian Tang, Muyang Li, Yao Lu, and Song Han. Deep compression autoencoder for efficient high-resolution diffusion models. arXiv preprint arXiv:2410.10733, 2024. 1, 6, 7, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.776, + 0.484, + 0.843 + ], + "angle": 0, + "content": "[12] Liang Chen, Haozhe Zhao, Tianyu Liu, Shuai Bai, Junyang Lin, Chang Zhou, and Baobao Chang. An image is worth 1/2 tokens after layer 2: Plug-and-play inference acceleration for large vision-language models. arXiv preprint arXiv:2403.06764, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.846, + 0.484, + 0.9 + ], + "angle": 0, + "content": "[13] Xi Chen, Diederik P Kingma, Tim Salimans, Yan Duan, Prafulla Dhariwal, John Schulman, Ilya Sutskever, and Pieter Abbeel. Variational lossy autoencoder. arXiv preprint arXiv:1611.02731, 2016. 1, 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.102, + 0.115, + 0.486, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.093, + 0.908, + 0.134 + ], + "angle": 0, + "content": "[14] Xinlei Chen, Zhuang Liu, Saining Xie, and Kaiming He. Deconstructing denoising diffusion models for self-supervised learning. arXiv preprint arXiv:2401.14404, 2024. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.136, + 0.908, + 0.191 + ], + "angle": 0, + "content": "[15] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248–255. IEEE, 2009. 5, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.193, + 0.907, + 0.221 + ], + "angle": 0, + "content": "[16] Prafulla Dhariwal and Alex Nichol. Diffusion models beat gans on image synthesis, 2021. 1, 7, 8, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.223, + 0.907, + 0.29 + ], + "angle": 0, + "content": "[17] Nat Dilokthanakul, Pedro AM Mediano, Marta Garnelo, Matthew CH Lee, Hugh Salimbeni, Kai Arulkumaran, and Murray Shanahan. Deep unsupervised clustering with gaussian mixture variational autoencoders. arXiv preprint arXiv:1611.02648, 2016. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.293, + 0.907, + 0.361 + ], + "angle": 0, + "content": "[18] Ming Ding, Zhuoyi Yang, Wenyi Hong, Wendi Zheng, Chang Zhou, Da Yin, Junyang Lin, Xu Zou, Zhou Shao, Hongxia Yang, et al. Cogview: Mastering text-to-image generation via transformers. Advances in neural information processing systems, 34:19822-19835, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.364, + 0.907, + 0.445 + ], + "angle": 0, + "content": "[19] Xiaoyi Dong, Jianmin Bao, Ting Zhang, Dongdong Chen, Weiming Zhang, Lu Yuan, Dong Chen, Fang Wen, Nenghai Yu, and Baining Guo. Peco: Perceptual codebook for bert pre-training of vision transformers. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 552-560, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.448, + 0.908, + 0.502 + ], + "angle": 0, + "content": "[20] Alexey Dosovitskiy and Thomas Brox. Generating images with perceptual similarity metrics based on deep networks. Advances in neural information processing systems, 29, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.504, + 0.908, + 0.587 + ], + "angle": 0, + "content": "[21] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale, 2021. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.589, + 0.908, + 0.629 + ], + "angle": 0, + "content": "[22] Joseph C Dunn. A fuzzy relative of the isodata process and its use in detecting compact well-separated clusters. 1973. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.632, + 0.908, + 0.7 + ], + "angle": 0, + "content": "[23] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12873-12883, 2021. 1, 2, 3, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.703, + 0.908, + 0.744 + ], + "angle": 0, + "content": "[24] Zach Evans, CJ Carr, Josiah Taylor, Scott H Hawley, and Jordi Pons. Fast timing-conditioned latent audio diffusion. arXiv preprint arXiv:2402.04825, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.746, + 0.907, + 0.814 + ], + "angle": 0, + "content": "[25] Lijie Fan, Tianhong Li, Siyang Qin, Yuzheng Li, Chen Sun, Michael Rubinstein, Deqing Sun, Kaiming He, and Yonglong Tian. Fluid: Scaling autoregressive text-to-image generative models with continuous tokens. arXiv preprint arXiv:2410.13863, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.817, + 0.908, + 0.87 + ], + "angle": 0, + "content": "[26] Yuxin Fang, Quan Sun, Xinggang Wang, Tiejun Huang, Xinlong Wang, and Yue Cao. Eva-02: A visual representation for neon genesis. Image and Vision Computing, 149:105171, 2024. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.873, + 0.907, + 0.902 + ], + "angle": 0, + "content": "[27] Christopher Fifty, Ronald G Junkins, Dennis Duan, Aniketh Iger, Jerry W Liu, Ehsan Amid, Sebastian Thrun, and" + }, + { + "type": "list", + "bbox": [ + 0.523, + 0.093, + 0.908, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "28366" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.133, + 0.093, + 0.483, + 0.12 + ], + "angle": 0, + "content": "Christopher Ré. Restructuring vector quantization with the rotation trick. arXiv preprint arXiv:2410.06424, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.123, + 0.484, + 0.191 + ], + "angle": 0, + "content": "[28] Oran Gafni, Adam Polyak, Oron Ashual, Shelly Sheynin, Devi Parikh, and Yaniv Taigman. Make-a-scene: Scene-based text-to-image generation with human priors. In European Conference on Computer Vision, pages 89-106. Springer, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.194, + 0.484, + 0.248 + ], + "angle": 0, + "content": "[29] Shanghua Gao, Pan Zhou, Ming-Ming Cheng, and Shuicheng Yan. Mdtv2: Masked diffusion transformer is a strong image synthesizer. arXiv preprint arXiv:2303.14389, 2023.6,5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.252, + 0.484, + 0.293 + ], + "angle": 0, + "content": "[30] Yuying Ge, Yixiao Ge, Ziyun Zeng, Xintao Wang, and Ying Shan. Planting a seed of vision in large language model, 2023. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.296, + 0.484, + 0.349 + ], + "angle": 0, + "content": "[31] Yuying Ge, Sijie Zhao, Ziyun Zeng, Yixiao Ge, Chen Li, Xintao Wang, and Ying Shan. Making llama see and draw with seed tokenizer. arXiv preprint arXiv:2310.01218, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.354, + 0.483, + 0.408 + ], + "angle": 0, + "content": "[32] Shansan Gong, Mukai Li, Jiangtao Feng, Zhiyong Wu, and LingPeng Kong. Diffuseq: Sequence to sequence text generation with diffusion models. arXiv preprint arXiv:2210.08933, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.412, + 0.484, + 0.466 + ], + "angle": 0, + "content": "[33] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139-144, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.47, + 0.484, + 0.511 + ], + "angle": 0, + "content": "[34] Yuchao Gu, Xintao Wang, Yixiao Ge, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Rethinking the objectives of vector-quantized tokenizers for image synthesis, 2023. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.513, + 0.484, + 0.581 + ], + "angle": 0, + "content": "[35] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 1, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.585, + 0.484, + 0.64 + ], + "angle": 0, + "content": "[36] Byeongho Heo, Song Park, Dongyoon Han, and Sangwoo Yun. Rotary position embedding for vision transformer. In European Conference on Computer Vision, pages 289-305. Springer, 2025. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.643, + 0.484, + 0.71 + ], + "angle": 0, + "content": "[37] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in Neural Information Processing Systems, 30, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.714, + 0.484, + 0.782 + ], + "angle": 0, + "content": "[38] Irina Higgins, Loic Matthew, Arka Pal, Christopher P Burgess, Xavier Glorot, Matthew M Botvinick, Shakir Mohamed, and Alexander Lerchner. beta-vae: Learning basic visual concepts with a constrained variational framework. ICLR, 3, 2017. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.786, + 0.484, + 0.826 + ], + "angle": 0, + "content": "[39] Geoffrey E Hinton and Ruslan R Salakhutdinov. Reducing the dimensionality of data with neural networks. science, 313(5786):504-507, 2006. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.829, + 0.483, + 0.857 + ], + "angle": 0, + "content": "[40] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.86, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[41] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 3" + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.093, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.093, + 0.907, + 0.147 + ], + "angle": 0, + "content": "[42] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. Advances in Neural Information Processing Systems, 35:8633-8646, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.15, + 0.907, + 0.219 + ], + "angle": 0, + "content": "[43] Minyoung Huh, Brian Cheung, Pulkit Agrawal, and Phillip Isola. Straightening out the straight-through estimator: Overcoming optimization challenges in vector quantized networks. In International Conference on Machine Learning, pages 14096-14113. PMLR, 2023. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.221, + 0.907, + 0.261 + ], + "angle": 0, + "content": "[44] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A. Efros. Image-to-image translation with conditional adversarial networks, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.264, + 0.907, + 0.317 + ], + "angle": 0, + "content": "[45] Herve Jegou, Matthijs Douze, and Cordelia Schmid. Product quantization for nearest neighbor search. IEEE transactions on pattern analysis and machine intelligence, 33(1):117-128, 2010. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.321, + 0.907, + 0.389 + ], + "angle": 0, + "content": "[46] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14, pages 694-711. Springer, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.391, + 0.907, + 0.459 + ], + "angle": 0, + "content": "[47] Matthew J Johnson, David K Duvenaud, Alex Wiltschko, Ryan P Adams, and Sandeep R Datta. Composing graphical models with neural networks for structured representations and fast inference. Advances in neural information processing systems, 29, 2016. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.462, + 0.907, + 0.529 + ], + "angle": 0, + "content": "[48] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 5, 7, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.533, + 0.907, + 0.601 + ], + "angle": 0, + "content": "[49] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8110-8119, 2020. 5, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.604, + 0.907, + 0.656 + ], + "angle": 0, + "content": "[50] Diederik Kingma and Ruiqi Gao. Understanding diffusion objectives as the elbo with simple data augmentation. Advances in Neural Information Processing Systems, 36, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.66, + 0.906, + 0.687 + ], + "angle": 0, + "content": "[51] Diederik P Kingma. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.689, + 0.906, + 0.744 + ], + "angle": 0, + "content": "[52] Tuomas Kynkänniemi, Tero Karras, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Improved precision and recall metric for assessing generative models. Advances in neural information processing systems, 32, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.747, + 0.907, + 0.813 + ], + "angle": 0, + "content": "[53] Tuomas Kynkänniemi, Miika Aittala, Tero Karras, Samuli Laine, Timo Aila, and Jaakko Lehtinen. Applying guidance in a limited interval improves sample and distribution quality in diffusion models. arXiv preprint arXiv:2404.07724, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.817, + 0.906, + 0.871 + ], + "angle": 0, + "content": "[54] Anders Boesen Lindbo Larsen, Søren Kaae Sønderby, Hugo Larochelle, and Ole Winther. Autoencoding beyond pixels using a learned similarity metric. In International conference on machine learning, pages 1558-1566. PMLR, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.874, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[55] Doyup Lee, Chiheon Kim, Saehoon Kim, Minsu Cho, and Wook-Shin Han. Autoregressive image generation using" + }, + { + "type": "list", + "bbox": [ + 0.523, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28367" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.133, + 0.093, + 0.484, + 0.134 + ], + "angle": 0, + "content": "residual quantization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11523-11532, 2022. 2, 6, 8, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.136, + 0.483, + 0.177 + ], + "angle": 0, + "content": "[56] Doyup Lee, Chiheon Kim, Saehoon Kim, Minsu Cho, and Wook-Shin Han. Autoregressive image generation using residual quantization, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.179, + 0.483, + 0.233 + ], + "angle": 0, + "content": "[57] Haopeng Li, Jinyue Yang, Kexin Wang, Xuerui Qiu, Yuhong Chou, Xin Li, and Guoqi Li. Scalable autoregressive image generation with mamba. arXiv preprint arXiv:2408.12245, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.235, + 0.484, + 0.303 + ], + "angle": 0, + "content": "[58] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730-19742. PMLR, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.305, + 0.483, + 0.36 + ], + "angle": 0, + "content": "[59] Tianhong Li, Huiwen Chang, Shlok Kumar Mishra, Han Zhang, Dina Katabi, and Dilip Krishnan. Mage: Masked generative encoder to unify representation learning and image synthesis, 2023. 3, 6, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.362, + 0.483, + 0.403 + ], + "angle": 0, + "content": "[60] Tianhong Li, Yonglong Tian, He Li, Mingyang Deng, and Kaiming He. Autoregressive image generation without vector quantization, 2024. 1, 2, 3, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.405, + 0.483, + 0.459 + ], + "angle": 0, + "content": "[61] Wentong Li, Yuqian Yuan, Jian Liu, Dongqi Tang, Song Wang, Jianke Zhu, and Lei Zhang. Tokenpacker: Efficient visual projector for multimodal llm. arXiv preprint arXiv:2407.02392, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.461, + 0.483, + 0.516 + ], + "angle": 0, + "content": "[62] Xiang Li, Hao Chen, Kai Qiu, Jason Kuen, Jiumiang Gu, Bhiksha Raj, and Zhe Lin. Imagefolder: Autoregressive image generation with folded tokens. arXiv preprint arXiv:2410.01756, 2024. 1, 2, 4, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.518, + 0.483, + 0.585 + ], + "angle": 0, + "content": "[63] Yanwei Li, Yuechen Zhang, Chengyao Wang, Zhisheng Zhong, Yixin Chen, Ruihang Chu, Shaoteng Liu, and Jiaya Jia. Mini-gemini: Mining the potential of multi-modality vision language models. arXiv preprint arXiv:2403.18814, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.588, + 0.483, + 0.657 + ], + "angle": 0, + "content": "[64] Ziyi Lin, Chris Liu, Renrui Zhang, Peng Gao, Longtian Qiu, Han Xiao, Han Qiu, Chen Lin, Wenqi Shao, Keqin Chen, et al. *Sphinx: The joint mixing of weights, tasks, and visual embeddings for multi-modal large language models.* arXiv preprint arXiv:2311.07575, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.659, + 0.483, + 0.7 + ], + "angle": 0, + "content": "[65] Yaron Lipman, Ricky TQ Chen, Heli Ben-Hamu, Maximilian Nickel, and Matt Le. Flow matching for generative modeling. arXiv preprint arXiv:2210.02747, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.702, + 0.483, + 0.757 + ], + "angle": 0, + "content": "[66] Wenze Liu, Le Zhuo, Yi Xin, Sheng Xia, Peng Gao, and Xiangyu Yue. Customize your visual autoregressive recipe with set autoregressive modeling. arXiv preprint arXiv:2410.10511, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.758, + 0.483, + 0.799 + ], + "angle": 0, + "content": "[67] Stuart Lloyd. Least squares quantization in pmc. IEEE transactions on information theory, 28(2):129-137, 1982. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.802, + 0.483, + 0.829 + ], + "angle": 0, + "content": "[68] I Loshchilov. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.831, + 0.483, + 0.871 + ], + "angle": 0, + "content": "[69] Cheng Lu and Yang Song. Simplifying, stabilizing and scaling continuous-time consistency models. arXiv preprint arXiv:2410.11081, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.873, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[70] Gen Luo, Yiyi Zhou, Yuxin Zhang, Xiawu Zheng, Xiaoshuai Sun, and Rongrong Ji. Feast your eyes: Mixture-of" + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.093, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.554, + 0.093, + 0.907, + 0.12 + ], + "angle": 0, + "content": "resolution adaptation for multimodal large language models. arXiv preprint arXiv:2403.03003, 2024.1" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.122, + 0.907, + 0.176 + ], + "angle": 0, + "content": "[71] Zhuoyan Luo, Fengyuan Shi, Yixiao Ge, Yujiu Yang, Limin Wang, and Ying Shan. Open-magtvit2: An open-source project toward democratizing auto-regressive visual generation. arXiv preprint arXiv:2409.04410, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.177, + 0.906, + 0.245 + ], + "angle": 0, + "content": "[72] Nanye Ma, Mark Goldstein, Michael S Albergo, Nicholas M Boffi, Eric Vanden-Eijnden, and Saining Xie. Sit: Exploring flow and diffusion-based generative models with scalable interpolant transformers. arXiv preprint arXiv:2401.08740, 2024. 1, 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.247, + 0.906, + 0.287 + ], + "angle": 0, + "content": "[73] Fabian Mentzer, David Minnen, Eirikur Agustsson, and Michael Tschannen. Finite scalar quantization: Vq-vae made simple, 2023. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.288, + 0.906, + 0.316 + ], + "angle": 0, + "content": "[74] MidJourney Inc. Midjourney. Software available from MidJourney Inc., 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.317, + 0.906, + 0.344 + ], + "angle": 0, + "content": "[75] Alex Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.345, + 0.907, + 0.413 + ], + "angle": 0, + "content": "[76] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021.8" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.414, + 0.906, + 0.455 + ], + "angle": 0, + "content": "[77] Harry Nyquist. Certain topics in telegraph transmission theory. Transactions of the American Institute of Electrical Engineers, 47(2):617-644, 1928. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.457, + 0.907, + 0.483 + ], + "angle": 0, + "content": "[78] OpenAI. Video generation models as world simulators. OpenAI Blog, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.484, + 0.907, + 0.608 + ], + "angle": 0, + "content": "[79] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Russell Howes, Po-Yao Huang, Hu Xu, Vasu Sharma, Shang-Wen Li, Wojciech Galuba, Mike Rabbat, Mido Assran, Nicolas Ballas, Gabriel Synnaeve, Ishan Misra, Herve Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. Dinov2: Learning robust visual features without supervision, 2023. 5, 7, 8, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.609, + 0.906, + 0.637 + ], + "angle": 0, + "content": "[80] William Peebles and Saining Xie. Scalable diffusion models with transformers, 2023. 1, 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.638, + 0.907, + 0.704 + ], + "angle": 0, + "content": "[81] Vadim Popov, Ivan Vovk, Vladimir Gogoryan, Tasnima Sadekova, and Mikhail Kudinov. Grad-tts: A diffusion probabilistic model for text-to-speech. In International Conference on Machine Learning, pages 8599-8608. PMLR, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.706, + 0.907, + 0.789 + ], + "angle": 0, + "content": "[82] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.79, + 0.907, + 0.857 + ], + "angle": 0, + "content": "[83] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International conference on machine learning, pages 8821-8831. Pmlr, 2021. 1, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.859, + 0.907, + 0.901 + ], + "angle": 0, + "content": "[84] Ali Razavi, Aaron Van den Oord, and Oriol Vinyals. Generating diverse high-fidelity images with vq-vae-2. Advances in neural information processing systems, 32, 2019. 1, 8" + }, + { + "type": "list", + "bbox": [ + 0.524, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "28368" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.093, + 0.486, + 0.133 + ], + "angle": 0, + "content": "[85] Ali Razavi, Aaron van den Oord, and Oriol Vinyals. Generating diverse high-fidelity images with vq-vae-2, 2019. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.134, + 0.485, + 0.163 + ], + "angle": 0, + "content": "[86] Douglas A Reynolds et al. Gaussian mixture models. Encyclopedia of biometrics, 741(659-663), 2009. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.164, + 0.484, + 0.232 + ], + "angle": 0, + "content": "[87] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.233, + 0.484, + 0.273 + ], + "angle": 0, + "content": "[88] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models, 2022. 1, 6, 8, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.274, + 0.484, + 0.357 + ], + "angle": 0, + "content": "[89] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in neural information processing systems, 35:36479-36494, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.358, + 0.484, + 0.412 + ], + "angle": 0, + "content": "[90] Tim Salimans, Ian Goodfellow, Wojciech Zaremba, Vicki Cheung, Alec Radford, and Xi Chen. Improved techniques for training gans. Advances in Neural Information Processing Systems, 29, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.413, + 0.482, + 0.441 + ], + "angle": 0, + "content": "[91] Claude Elwood Shannon. Communication in the presence of noise. Proceedings of the IRE, 37(1):10-21, 1949. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.442, + 0.484, + 0.509 + ], + "angle": 0, + "content": "[92] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International conference on machine learning, pages 2256-2265. PMLR, 2015. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.511, + 0.484, + 0.552 + ], + "angle": 0, + "content": "[93] Jascha Sohl-Dickstein, Eric A. Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics, 2015. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.553, + 0.484, + 0.607 + ], + "angle": 0, + "content": "[94] Dingjie Song, Wenjun Wang, Shunian Chen, Xidong Wang, Michael Guan, and Benyou Wang. Less is more: A simple yet effective token reduction method for efficient multimodal llms. arXiv preprint arXiv:2409.10994, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.608, + 0.484, + 0.636 + ], + "angle": 0, + "content": "[95] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.637, + 0.468, + 0.65 + ], + "angle": 0, + "content": "[96] stabilityai. Sd vae ft ema, 2023. Accessed: 2023. 1, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.651, + 0.484, + 0.704 + ], + "angle": 0, + "content": "[97] Jianlin Su, Murtadha Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu. Roformer: Enhanced transformer with rotary position embedding. Neurocomputing, 568:127063, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.706, + 0.484, + 0.761 + ], + "angle": 0, + "content": "[98] Peize Sun, Yi Jiang, Shoufa Chen, Shilong Zhang, Bingyue Peng, Ping Luo, and Zehuan Yuan. Autoregressive model beats diffusion: Llama for scalable image generation. arXiv preprint arXiv:2406.06525, 2024. 1, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.762, + 0.484, + 0.803 + ], + "angle": 0, + "content": "[99] Keyu Tian, Yi Jiang, Zehuan Yuan, Bingyue Peng, and Liwei Wang. Visual autoregressive modeling: Scalable image generation via next-scale prediction, 2024. 3, 5, 8, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.805, + 0.485, + 0.859 + ], + "angle": 0, + "content": "[100] Michael Tschannen, Cian Eastwood, and Fabian Mentzer. Givt: Generative infinite-vocabulary transformers. In European Conference on Computer Vision, pages 292-309. Springer, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.86, + 0.484, + 0.9 + ], + "angle": 0, + "content": "[101] Hung-Yu Tseng, Lu Jiang, Ce Liu, Ming-Hsuan Yang, and Weilong Yang. Regularizing generative adversarial networks under limited data, 2021. 5, 2" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.093, + 0.486, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.121 + ], + "angle": 0, + "content": "[102] Arash Vahdat, Karsten Kreis, and Jan Kautz. Score-based generative modeling in latent space, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.122, + 0.908, + 0.176 + ], + "angle": 0, + "content": "[103] Aaron Van den Oord, Nal Kalchbrenner, Lasse Espeholt, Oriol Vinyals, Alex Graves, et al. Conditional image generation with pixelCNN decoders. Advances in neural information processing systems, 29, 2016. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.177, + 0.906, + 0.218 + ], + "angle": 0, + "content": "[104] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017. 1, 2, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.219, + 0.908, + 0.26 + ], + "angle": 0, + "content": "[105] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need, 2023. 1, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.261, + 0.908, + 0.33 + ], + "angle": 0, + "content": "[106] Pascal Vincent, Hugo Larochelle, Yoshua Bengio, and Pierre-Antoine Manzagol. Extracting and composing robust features with denoising autoencoders. In Proceedings of the 25th international conference on Machine learning, pages 1096-1103, 2008. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.331, + 0.906, + 0.385 + ], + "angle": 0, + "content": "[107] Xinlong Wang, Xiaosong Zhang, Zhengxiong Luo, Quan Sun, Yufeng Cui, Jinsheng Wang, Fan Zhang, Yueze Wang, Zhen Li, Qiying Yu, et al. Emu3: Next-token prediction is all you need. arXiv preprint arXiv:2409.18869, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.386, + 0.906, + 0.442 + ], + "angle": 0, + "content": "[108] Mark Weber, Lijun Yu, Qihang Yu, Xueqing Deng, Xiaohui Shen, Daniel Cremers, and Liang-Chieh Chen. Maskbit: Embedding-free image generation via bit tokens. arXiv preprint arXiv:2409.16211, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.443, + 0.906, + 0.511 + ], + "angle": 0, + "content": "[109] Yecheng Wu, Zhuoyang Zhang, Junyu Chen, Haotian Tang, Dacheng Li, Yunhao Fang, Ligeng Zhu, Enze Xie, Hongxu Yin, Li Yi, et al. Vila-u: a unified foundation model integrating visual understanding and generation. arXiv preprint arXiv:2409.04429, 2024. 1, 2, 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.511, + 0.906, + 0.58 + ], + "angle": 0, + "content": "[110] Jinheng Xie, Weijia Mao, Zechen Bai, David Junhao Zhang, Weihao Wang, Kevin Qinghong Lin, Yuchao Gu, Zhijie Chen, Zhenheng Yang, and Mike Zheng Shou. Show-o: One single transformer to unify multimodal understanding and generation. arXiv preprint arXiv:2408.12528, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.581, + 0.908, + 0.649 + ], + "angle": 0, + "content": "[111] Jiahui Yu, Xin Li, Jing Yu Koh, Han Zhang, Ruoming Pang, James Qin, Alexander Ku, Yuanzhong Xu, Jason Baldridge, and Yonghui Wu. Vector-quantized image modeling with improved vqgan. arXiv preprint arXiv:2110.04627, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.65, + 0.908, + 0.72 + ], + "angle": 0, + "content": "[112] Lijun Yu, José Lezama, Nitesh B Gundavarapu, Luca Versari, Kihyuk Sohn, David Minnen, Yong Cheng, Vighnesh Birodkar, Agrim Gupta, Xiuye Gu, et al. Language model beats diffusion-tokenizer is key to visual generation. arXiv preprint arXiv:2310.05737, 2023. 1, 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.721, + 0.908, + 0.802 + ], + "angle": 0, + "content": "[113] Lijun Yu, Yong Cheng, Zhiruo Wang, Vivek Kumar, Wolfgang Macherey, Yanping Huang, David Ross, Irfan Essa, Yonatan Bisk, Ming-Hsuan Yang, et al. Spae: Semantic pyramid autoencoder for multimodal generation with frozen llms. Advances in Neural Information Processing Systems, 36, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.804, + 0.908, + 0.845 + ], + "angle": 0, + "content": "[114] Qihang Yu, Ju He, Xueqing Deng, Xiaohui Shen, and Liang-Chieh Chen. Randomized autoregressive visual generation. arXiv preprint arXiv:2411.00776, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.908, + 0.9 + ], + "angle": 0, + "content": "[115] Qihang Yu, Mark Weber, Xueqing Deng, Xiaohui Shen, Daniel Cremers, and Liang-Chieh Chen. An image is worth 32 tokens for reconstruction and generation. arxiv: 2406.07550, 2024. 1, 2, 4, 6, 7, 8, 5" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.093, + 0.908, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "28369" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.161 + ], + "angle": 0, + "content": "[116] Sihyun Yu, Sangkyung Kwak, Huiwon Jang, Jongheon Jeong, Jonathan Huang, Jinwoo Shin, and Saining Xie. Representation alignment for generation: Training diffusion transformers is easier than you think. arXiv preprint arXiv:2410.06940, 2024. 1, 2, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.162, + 0.485, + 0.219 + ], + "angle": 0, + "content": "[117] Neil Zeghidour, Alejandro Luebs, Ahmed Omran, Jan Skoglund, and Marco Tagliasacchi. Soundstream: An end-to-end neural audio codec. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 30:495-507, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.22, + 0.485, + 0.302 + ], + "angle": 0, + "content": "[118] Baoquan Zhang, Huaibin Wang, Chuyao Luo, Xutao Li, Guotao Liang, Yunming Ye, Xiaochen Qi, and Yao He. Codebook transfer with part-of-speech for vector-quantized image modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7757–7766, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.303, + 0.484, + 0.346 + ], + "angle": 0, + "content": "[119] Han Zhang, Zizhao Zhang, Augustus Odena, and Honglak Lee. Consistency regularization for generative adversarial networks. arXiv preprint arXiv:1910.12027, 2019. 5, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.346, + 0.484, + 0.389 + ], + "angle": 0, + "content": "[120] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.389, + 0.484, + 0.444 + ], + "angle": 0, + "content": "[121] Shengyu Zhao, Zhijian Liu, Ji Lin, Jun-Yan Zhu, and Song Han. Differentiable augmentation for data-efficient gan training. Advances in neural information processing systems, 33:7559-7570, 2020. 5, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.446, + 0.484, + 0.488 + ], + "angle": 0, + "content": "[122] Lin Zheng, Jianbo Yuan, Lei Yu, and Lingpeng Kong. A reparameterized discrete diffusion model for text generation. arXiv preprint arXiv:2302.05737, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.489, + 0.484, + 0.557 + ], + "angle": 0, + "content": "[123] Chunting Zhou, Lili Yu, Arun Babu, Kushal Tirumala, Michihiro Yasunaga, Leonid Shamis, Jacob Kahn, Xuezhe Ma, Luke Zettlemoyer, and Omer Levy. Transfusion: Predict the next token and diffuse images with one multi-modal model. arXiv preprint arXiv:2408.11039, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.559, + 0.484, + 0.601 + ], + "angle": 0, + "content": "[124] Lei Zhu, Fangyun Wei, Yanye Lu, and Dong Chen. Scaling the codebook size of vqgan to 100,000 with a utilization rate of \\(99\\%\\). arXiv preprint arXiv:2406.11837, 2024. 2, 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.601, + 0.484, + 0.655 + ], + "angle": 0, + "content": "[125] X Zhu, W Su, L Lu, B Li, X Wang, and J Dai. Deformable detr: Deformable transformers for end-to-end object detection. arxiv 2020. arXiv preprint arXiv:2010.04159, 2010.8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.657, + 0.484, + 0.7 + ], + "angle": 0, + "content": "[126] Yongxin Zhu, Bocheng Li, Yifei Xin, and Linli Xu. Addressing representation collapse in vector quantized models with one linear layer. arXiv preprint arXiv:2411.02038, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.7 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "28370" + } + ] +] \ No newline at end of file diff --git a/2025/SoftVQ-VAE_ Efficient 1-Dimensional Continuous Tokenizer/315fadef-93cf-49d1-961b-c20538f1f4de_origin.pdf b/2025/SoftVQ-VAE_ Efficient 1-Dimensional Continuous Tokenizer/315fadef-93cf-49d1-961b-c20538f1f4de_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..1cb53fb252db07461a88ccdff32900470e254b8a --- /dev/null +++ b/2025/SoftVQ-VAE_ Efficient 1-Dimensional Continuous Tokenizer/315fadef-93cf-49d1-961b-c20538f1f4de_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:25c943d7b841645efc4dca760980e96408d0ee63c83db559e645b8cbc741b7d1 +size 356697 diff --git a/2025/SoftVQ-VAE_ Efficient 1-Dimensional Continuous Tokenizer/full.md b/2025/SoftVQ-VAE_ Efficient 1-Dimensional Continuous Tokenizer/full.md new file mode 100644 index 0000000000000000000000000000000000000000..2163540f053374d1fd65b7bd2b9087dbf78bd83c --- /dev/null +++ b/2025/SoftVQ-VAE_ Efficient 1-Dimensional Continuous Tokenizer/full.md @@ -0,0 +1,379 @@ +# SoftVQ-VAE: Efficient 1-Dimensional Continuous Tokenizer + +Hao Chen $^{1,2*}$ , Ze Wang $^{2}$ , Xiang Li $^{1}$ , Xineng Sun $^{2}$ , Fangyi Chen $^{1}$ , Jiang Liu $^{2}$ , Jindong Wang $^{3}$ , Bhiksha Raj $^{1,4}$ , Zicheng Liu $^{2}$ , Emad Barsoum $^{2}$ + +$^{1}$ Carnegie Mellon University, $^{2}$ AMD, $^{3}$ William & Mary, $^{4}$ MBZUAI + +# Abstract + +Efficient image tokenization with high compression ratios remains a critical challenge for training generative models. We present SoftVQ-VAE, a continuous image tokenizer that leverages soft categorical posteriors to aggregate multiple codewords into each latent token, substantially increasing the representation capacity of the latent space. When applied to Transformer-based architectures, our approach compresses $256 \times 256$ and $512 \times 512$ images using as few as 32 or 64 1-dimensional tokens. Not only does SoftVQ-VAE show consistent and high-quality reconstruction, more importantly, it also achieves state-of-the-art and significantly faster image generation results across different denoising-based generative models. Remarkably, SoftVQ-VAE improves inference throughput by up to $18x$ for generating $256 \times 256$ images and $55x$ for $512 \times 512$ images while achieving competitive FID scores of 1.78 and 2.21 for SiT-XL. It also improves the training efficiency of the generative models by reducing the number of training iterations by $2.3x$ while maintaining comparable performance. With its fully-differentiable design and semantic-rich latent space, our experiment demonstrates that SoftVQ-VAE achieves efficient tokenization without compromising generation quality, paving the way for more efficient generative models. Code and model are released1. + +# 1. Introduction + +Denoising-based generative modeling has witnessed remarkable progress with recent advances, such as Diffusion Transformers (DiT) [80], Scalable Interpolant Transformers (SiT) [72], and Masked Auto-Regressive models with diffusion loss (MAR) [60], to name a few. Denosing-based generative modeling not only presents impressive generation results on a wide range of modality, including natural language [32, 122], images [16, 25, 69, 83, 88], videos [6, 8, 42], and audios [24, 81], but also shows potential to unify the understanding and generation capabilities across modalities in multi-modal language models [107, 109, 110, 123]. + +A core component of denoising-based generative models is the tokenizer [23, 38, 51, 56, 84, 104, 112, 117], which compresses the raw data of each modality into a set of latent tokens in either a discrete or continuous latent space. The compact latent space therefore allows for more efficient and better generative modeling [88]. Among previous efforts, Kullback-Leibler Variational Auto-Encoders (KL-VAE) [51] and Vector Quantized Variational Auto-Encoders (VQ-VAE) [23, 84, 104] stand out as representatives of tokenizers which introduce continuous and discrete latent spaces, respectively. The former constrains the latent space with a Gaussian distribution using re-parametrization [51] trick, and the latter makes its latent space a categorical discrete distribution with a codebook of finite vocabulary which requires straight-through estimation [7]. + +Although both KL-VAE and VQ-VAE (and their variants) have been predominantly adopted in many generative models [10, 80, 88, 98, 123], they still present two major challenges restricting the efficiency and effectiveness of generative modeling: (1) the difficulty of achieving a higher compression ratio [13] and (2) the worse discriminative representations in their latent space than other self-supervised methods [14, 35, 109, 116]. The efficiency of downstream generative models, particularly Transformer-based architectures [105], is fundamentally constrained by their quadratic complexity to the number of latent tokens. Current image tokenizers [23, 88, 96] typically compress $256 \times 256$ images to at least 256 tokens and $512 \times 512$ images to at least 1024 tokens, creating a significant computational bottleneck for both training and inference of generative models [60, 72, 80]. Many efforts have been made to reduce the number of latent tokens on the generative model side, such as merging [62, 72, 80], pooling [58, 61, 94], and others [12, 63, 70]. More studies have recently emerged to fundamentally reduce the token number of the tokenizer [11, 30, 62, 115]. For example, TiTok [115] uses 128 tokens, achieving generation results comparable to 256 tokens by adding one extra decoder. DC-AE [11] compresses the initial 1024 tokens of $512 \times 512$ with 256 tokens. However, further increasing the compression ratio results in a significant degradation in the quality of reconstruction and therefore the generation [11]. + +![](images/21e994b6b9a87a8fc9dafce3a267dd48eb64d76654861724910fd1a07b80c5fa.jpg) +Figure 1. ImageNet-1K $256\times 256$ and $512\times 512$ generation results of generative models trained on SoftVQ-VAE with 32 and 64 tokens. + +Beyond the compression challenge, the quality of the learned representations of the tokenizer poses another critical limitation. While latent representations are crucial for both understanding [31, 109] and generation tasks [30, 62, 64, 116], current tokenizers usually struggle to capture more discriminative features. Earlier methods attempted to advance the training objectives and recipe to improve the latent space of the tokenizer [14, 19, 23, 34, 124]. Recently, [30, 62, 109, 113, 116, 118, 124] has also explored the representation alignment of latent tokens. However, due to the lossy nature [77, 91] of the smoothness constraints in KL-VAE and the discrete quantization in VQVAE [13, 27, 38, 126], current tokenizers usually fall short in learning the semantics of the latent space [109]. + +To address the aforementioned challenges, we present SoftVQ-VAE, a simple modification to VQ-VAE that converts it from a discrete tokenizer into a continuous one with a high compression ratio and a semantic-rich latent space. Specifically, we propose using a soft categorical posterior in VAE with a learnable codebook. This straightforward modification introduces a crucial capability: instead of the conventional one-to-one mapping between tokens and codewords in VQ-VAE, our approach enables the adaptive aggregation of multiple codewords into each latent token, substantially increasing the representation capacity of the latent space. We apply SoftVQ-VAE to the Transformer auto-encoder architecture [30, 58, 62, 111, 115], and we succeed in using much fewer 1-dimensional latent tokens (i.e., 32 and 64) for both the reconstruction and subsequent generation tasks. Since SoftVQ-VAE is now fully differentiable, it not only simplifies the learning of the codebook, not requiring the codebook loss or commit loss as in VQ-VAE [23, 104], but + +also enables better representation learning through direct alignment with pre-trained semantic features using a simple cosine similarity objective. Our approach generalizes the K-Means [67] latent space of VQ-VAE to Soft K-Means [22] and further to Gaussian Mixture Models (GMM) [86], while maintaining compatibility with existing VQ-VAE techniques [45, 55, 62, 112]. We comprehensively demonstrate the efficiency and effectiveness of the proposed SoftVQ-VAE in image reconstruction and generation tasks, and show that: + +- SoftVQ-VAE presents more consistent, robust, and high-quality reconstruction and generation results with various latent token lengths ranging from 256 to 32. +- SoftVQ-VAE allows diffusion-based DiT [73], flow-based SiT [72], and autoregressive-based MAR [60] to achieve state-of-the-art generation results with only 64 and 32 tokens on both the $256 \times 256$ , with an FID of 1.78, and $512 \times 512$ on ImageNet benchmark, with an FID of 2.21. +- SoftVQ-VAE improves the inference throughput by 18x and 10x with 32 and 64 tokens to generate $256 \times 256$ images, and 55x with 64 tokens for generating $512 \times 512$ images, with significantly lower GFLOPs for all models. +- SoftVQ-VAE presents a more discriminative latent space through representation alignment. It also facilitates training efficiency by significantly improving training throughput and reduces training iterations by 2.3x. + +# 2. Preliminary + +We present an overview of the tokenizers, i.e., KL-VAE [51] and VQ-VAE [104], and denoising-based generative models, i.e., DiT [80], SiT [72], and MAR [60], in this section. + +# 2.1. Image Tokenizer + +The architecture of image tokenizer generally resembles auto-encoders, consisting of two main components: an encoder $\mathcal{E}$ and a decoder $\mathcal{D}$ , parameterized by $\phi$ and $\theta$ , respectively. Given an input image $\mathbf{x} \in \mathbb{R}^{H \times W \times 3}$ , the encoder $\mathcal{E}$ maps the high-dimensional $\mathbf{x}$ to a lower-dimensional latent representation $\mathbf{z} = \mathcal{E}(\mathbf{x}; \phi)$ . The latent representations can have varying shapes depending on the encoder architecture. We generally view the latent representation as a set of latent tokens $\mathbf{z} = [\mathbf{z}^{[0]}, \mathbf{z}^{[1]}, \dots, \mathbf{z}^{[L]}] \in \mathbb{R}^{L \times D}$ , where $L$ is the number of tokens and $D$ is the dimension of the latent representation. For basic auto-encoders (AE), the decoder takes latent tokens as input and reconstructs the original signal $\hat{\mathbf{x}} = \mathcal{D}(\mathbf{z}; \theta)$ by minimizing the reconstruction loss with $\mathbf{x}$ . + +Variational auto-encoders (VAE) [51] extend the basic AE by introducing a probabilistic perspective on the latent space. Specifically, VAE approximates a posterior distribution $p(\mathbf{z}|\mathbf{x})$ with a learned distribution $q_{\phi}(\mathbf{z}|\mathbf{x})$ from the encoder, assuming that $\mathbf{x}$ is generated by the unobserved latent $\mathbf{z}$ with a prior distribution $p(\mathbf{z})$ . The decoder instead takes a set of sampled latent tokens $\mathbf{z} \sim q_{\phi}(\mathbf{z}|\mathbf{x})$ from the posterior and aims to learn the marginal likelihood of data in a generative process by optimizing the evidence lower bound (ELBO): $\max_{\phi,\theta}\mathbb{E}_{q_{\phi}(\mathbf{z}|\mathbf{x})}\log p_{\theta}(\mathbf{x}|\mathbf{z}) - \mathcal{L}_{\mathrm{kl}}(q_{\phi}(\mathbf{z}|\mathbf{x})|p(\mathbf{z}))$ . To make the optimization tractable, assumptions have been made on the prior, leading to different variants of VAE. + +KL-VAE [38, 51] parametrizes both the prior and posterior as Gaussians. The prior $p(\mathbf{z})$ is assumed to be the isotropic unit Gaussian $\mathcal{N}(0,1)$ . The continuous latent code is parameterized with posterior mean $\mu_{\phi}$ and variance $\sigma_{\phi}^{2}$ predicted by the encoder using the "re-parametrization" trick with a noise variable $\varepsilon$ from standard Gaussian: + +$$ +\text {p o s t e r i o r :} q _ {\phi} (\mathbf {z} | \mathbf {x}) = \mathcal {N} (\mathbf {z}; \boldsymbol {\mu} _ {\phi} (\mathbf {x}), \boldsymbol {\sigma} _ {\phi} ^ {2} (\mathbf {x})) +$$ + +$$ +l a t e n t: \mathbf {z} = \boldsymbol {\mu} _ {\phi} (\mathbf {x}) + \boldsymbol {\sigma} _ {\phi} (\mathbf {x}) \odot \varepsilon , \quad \epsilon \sim \mathcal {N} (0, \mathbf {I}) \tag {1} +$$ + +$$ +k l: \mathcal {L} _ {\mathrm {k l}} = - \frac {1}{2} \left(1 + \log \left(\boldsymbol {\sigma} _ {\phi} ^ {2} (\mathbf {x})\right) - \boldsymbol {\mu} _ {\phi} (\mathbf {x}) ^ {2} - \boldsymbol {\sigma} _ {\phi} ^ {2} (\mathbf {x})\right) +$$ + +While KL-VAE is widely used in diffusion models, it is known to have the "posterior-collapse" issue [13]. In addition, the KL loss weight usually imposes a trade-off between the quality of reconstruction and the smoothness of its latent space [100], preventing high compression ratio and the learning of semantics in the latent space of KL-VAE. + +VQ-VAE [23, 104], in contrast, generates latent code in a discrete space by taking the posterior as a \(K\)-way deterministic categorical distribution against a learnable codebook \(\mathcal{C} = [\mathbf{c}^{[0]}, \dots, \mathbf{c}^{[K]}\] \(\in \mathbb{R}^{K \times D}\), with \(\mathbf{c}\) as the codeword: + +$$ +\text {p o s t e r i o r :} q _ {\phi} (\mathbf {z} = k | \mathbf {x}) = \left\{ \begin{array}{l l} 1 & \text {i f} k = \arg \min _ {j} \| \hat {\mathbf {z}} - \mathbf {c} ^ {[ j ]} \| _ {2} \\ 0 & \text {o t h e r w i s e} \end{array} \right. +$$ + +$$ +l a t e n t: \mathbf {z} = \mathbf {c} ^ {[ k ]}, \quad \text {w h e r e} k = \underset {j} {\arg \min } \| \hat {\mathbf {z}} - \mathbf {c} ^ {[ j ]} \| _ {2} \tag {2} +$$ + +$$ +k l: \mathcal {L} _ {\mathrm {k l}} = \log K +$$ + +Due to the non-differentiable arg min, a "straight-through" trick [7] is adopted to approximate the gradient of the encoder output $\hat{\mathbf{z}}$ by directly copying gradient from the decoder input $\mathbf{z}$ . Since the reconstruction objective does not impose direct gradients on the codebook $\mathcal{C}$ , VQ-VAE additionally uses a codebook loss as $\| \mathrm{sg}[\hat{\mathbf{z}}] - \mathbf{c}\| _2$ to move codewords toward the encoder output, along with a commit loss $\| \hat{\mathbf{z}} -\mathrm{sg}[\mathbf{c}]\| _2$ to prevent arbitrary growth of codewords, where $\mathrm{sg}[\cdot ]$ denotes stop-gradient. The broken gradient hinders high compression ratio and latent space learning [27, 43]. + +# 2.2. Denoising-based Generative Models + +Denoising-based generative models synthesize images by progressively transforming Gaussian noise $\varepsilon \sim \mathcal{N}(0,\mathbf{I})$ into tokenizer latent codes $\mathbf{z}$ through a forward process [41, 92]: + +$$ +\mathbf {z} _ {t} = \alpha_ {t} \mathbf {z} + \sigma_ {t} \varepsilon , \tag {3} +$$ + +where $\alpha_{t}$ and $\sigma_{t}$ are a decreasing and increasing function of $t$ , and $t \in [0, T]$ [50]. In this paper, we adopt three approaches to implementing this denoising process as follows. + +DiT [80] is a diffusion-based model with a transformer architecture. It formulates the process through a forward-time stochastic differential equation (SDE), where $\mathbf{z}_t$ converges to $\mathcal{N}(0, \mathbf{I})$ as $t \to T$ . Generation occurs via a reverse-time SDE [2], with the model trained to predict noise $\varepsilon$ at randomly sampled timesteps. Due to the long token length of the tokenizer [96] used in DiT, it has several variants to merge tokens at the input of the transformer backbone. + +SiT [72] instead uses stochastic interpolants [1] and performs denoising via probability flow ordinary differential equation (PF ODE). The model learns the ODE's velocity field [65] by minimizing mean-squared-error (MSE) loss. This velocity field defines a score function for generation, analogous to DiT's reverse-time SDE. + +MAR [60] combines diffusion with autoregressive generation. Unlike DiT and SiT's parallel denoising of all latent tokens, MAR adopts an encoder-decoder transformer architecture [35] that progressively denoises tokens in a "next-set" autoregressive manner [99], starting from masked tokens, similar to MaskGIT [10] and MAGE [59]. + +# 3. Method + +In this section, we present SoftVQ-VAE, a novel 1D continuous tokenizer with a high compression ratio and a semantic-rich latent space. We first introduce the architecture and the formulation of SoftVQ-VAE. Then, we show that SoftVQ-VAE can learn semantics easily by aligning its latent tokens with pre-trained features via its fully-differentiable property. + +# 3.1. Architecture + +We leverage Vision Transformer (ViT) [21, 111] as the architecture for the encoder $\mathcal{E}$ and decoder $\mathcal{D}$ of SoftVQ-VAE. + +![](images/b757a7fb4120fe2d1d016fc288e2800319d031dd85905ac89c7cb35f0aeebb90.jpg) +Figure 2. Illustration of SoftVQ-VAE. Left: Transformer encoder-decoder architecture with image tokens, arbitrary length of latent tokens, and mask tokens. Right top: fully-differentiable SoftVQ illustration. Right bottom: latent space representation alignment. + +Similarly to Yu et al. [115], instead of using the image features as latent, we initialize a set of extra 1D learnable tokens and use these tokens for reconstruction and subsequent generation. With the self-attention mechanism, it allows the learnable tokens to adaptively aggregate different image tokens to obtain the latent tokens of SoftVQ, detailed in Sec. 3.2. Specifically, for the encoder transformer with a patch size of $P$ , an image $\mathbf{x} \in \mathbb{R}^{H \times W \times 3}$ is patched into a sequence of image tokens $\mathbf{x}_p \in \mathbb{R}^{N \times D}$ , where $N = HW / P^2$ and $D$ represent the token dimension size. We then concatenate a set of learnable tokens $\mathbf{z}_l \in \mathbb{R}^{L \times D}$ of length $L$ along with the image tokens as the input to the encoder, and retain only the output corresponding to the learnable latent tokens as the output of our encoder: $\hat{\mathbf{z}} = \mathcal{E}([\mathbf{x}_p; \mathbf{z}_l]; \phi)$ . + +For the decoder, we reconstruct images from a sequence of learnable mask tokens $\mathbf{m} \in \mathbb{R}^{N \times D}$ [5, 35], concatenated with latent tokens: $\hat{\mathbf{x}} = \mathcal{D}([\mathbf{m};\mathbf{z}];\theta)$ . We use a linear layer at the end of the decoder to regress the pixel values from the mask tokens. For the image tokens at the encoder input and mask tokens at the decoder input, we apply 2D absolute position embeddings and 1D absolute position embedding for the latent tokens. Adopting more advanced position embedding techniques, such as RoPE [36, 97], may lead to better results and is left for future work. This design allows SoftVQ-VAE to use latent codes of arbitrary length, each adaptively associated with image tokens, for reconstruction and subsequent generative modeling, and model images of various resolutions using the same length of latent codes. An overview of the model architecture is shown in Fig. 2. + +# 3.2. SoftVQ-VAE + +As discussed above, the high compression ratio of both KL-VAE and VQ-VAE usually results in a significant degradation of quality in reconstruction and latent space. To overcome these limitations, we propose SoftVQ-VAE, a simple mod + +ification to VQ-VAE, which maintains the advantages of learning codewords to capture the data distribution while bridging it with more representation capacity as a continuous tokenizer. The core idea of SoftVQ-VAE is to allow each latent code to adaptively aggregate multiple codewords from the learnable codebook. Consequently, it imposes a soft categorical distribution on the posterior from the encoder: + +$$ +\begin{array}{l} \text {p o s t e r i o r :} q _ {\phi} (\mathbf {z} | \mathbf {x}) = \operatorname {S o f t m a x} \left(- \frac {\| \hat {\mathbf {z}} - \mathcal {C} \| _ {2}}{\tau}\right) \\ \text {l a t e n t :} \mathbf {z} = q _ {\phi} (\mathbf {z} | \mathbf {x}) \mathcal {C} \end{array} \tag {4} +$$ + +$$ +k l \colon \mathcal {L} _ {\mathrm {k l}} = H (q _ {\phi} (\mathbf {z} | \mathbf {x})) - H \left(\mathbb {E} _ {\mathbf {x} \sim p (\mathbf {x})} q _ {\phi} (\mathbf {z} | \mathbf {x})\right) +$$ + +where $\tau$ is the temperature parameter controlling the sharpness of the softmax probability and is set to 0.07 by ablation. The derivation of $\mathcal{L}_{kl}$ is shown in Appendix B.1. + +Despite this simple modification, SoftVQ-VAE can significantly reduce the latent token length while maintaining a high quality of the reconstruction and latent space. More importantly, with a soft categorical distribution, SoftVQ-VAE is fully-differentiable, and thus the encoder and codebook can be optimized directly from the reconstruction loss (and other losses). It also allows for easier regularization of various forms on the latent space that drastically improves its quality, and less hyper-parameter tuning without codebook and commit loss [43] in the reduced training objectives. + +Interpretation. The posterior and latent space of VQ-VAE can be interpreted with K-Means [67]. Hence, SoftVQ-VAE can be viewed as soft K-Means [22], with codewords as learnable prototypes in the latent space. This latent space can easily be extended to Gaussian Mixture Models (GMM) [86], which we refer to as GMMVQ-VAE. Instead of using fixed temperature, GMMVQ-VAE predicts data-dependent weights for the codeword prototypes, i.e., Gaussian means, to compute the soft categorical posterior: $q_{\phi}(\mathbf{z}|\mathbf{x}) = \mathrm{Softmax}(-\omega (\hat{\mathbf{z}})\| \hat{\mathbf{z}} -\mathcal{C}\| _2)$ . Although prior work has explored GMM-based VAEs [17, 47], our GMMVQ-VAE differs by learning discrete prototypes rather than continuous latent variables. While providing the additional benefit of interpretable Gaussian components in the latent space, GMMVQ achieves reconstruction quality and downstream generation performance similar to SoftVQ in our experiments, and thus we mainly adopt the simpler SoftVQ throughout this paper. More details are in Appendix B.2. + +# 3.3. Representation Alignment of Latent Space + +While the reconstruction quality of tokenizer is important, learning a high-quality latent space is more crucial for downstream denoising-based generative modeling [116]. Several recent approaches have explored aligning the latent code with pre-trained features through a contrastive loss [30, 62, 109] and initializing the codebook with pre-trained features [124]. However, imposing effective regularization + +on the latent remains a challenge due to the discrete quantization of VQ-VAE and the Gaussian constraints in KL-VAE. + +Thanks to the fully-differentiable property of SoftVQ-VAE, we can now directly impose regularization on the latent space. Inspired by the recent work of REPA [116], we propose aligning the representations of latent codes with pre-trained vision encoders to facilitate learning the latent space. Compared to REPA, which aligns the representation at intermediate layers of downstream generative models, our method of feature alignment in the tokenizer's latent space is equivalent to performing REPA at the input space of generative models. Specifically, to align the latent code with the image tokens from a pre-trained vision encoder, we replicate each latent token by $N / L$ times: + +$$ +\mathbf {z} _ {r} = \underbrace {\left[ \mathbf {z} ^ {[ 0 ]} , \dots , \mathbf {z} ^ {[ 0 ]} \right.} _ {N / L \text {t i m e s}}, \underbrace {\left. \mathbf {z} ^ {[ 1 ]} , \dots , \mathbf {z} ^ {[ 1 ]} \right.} _ {N / L \text {t i m e s}}, \dots , \underbrace {\left. \mathbf {z} ^ {[ L ]} , \dots , \mathbf {z} ^ {[ L ]} \right]} _ {N / L \text {t i m e s}}, \tag {5} +$$ + +and encourage the token-wise similarity with image tokens. We employ a projector multilayer perceptron (MLP) on top of the latent tokens to match the pre-trained feature $\mathbf{y}_{*}$ : + +$$ +\mathcal {L} _ {\text {a l i g n}} = \frac {1}{N} \sum_ {n = 1} ^ {N} \operatorname {s i m} \left(\mathbf {y} _ {*} ^ {[ n ]}, \mathbf {M L P} \left(\mathbf {z} _ {r} ^ {([ n ])}\right)\right). \tag {6} +$$ + +This alignment ensures the latent space of SoftVQ-VAE captures semantically discriminative features that are beneficial for training the downstream generative models, even if it does not directly translate into improved reconstruction performance from the tokenizer. However, as we will show in Sec. 4.4, the generation quality of trained generative models often depends more on the semantic structure of the latent space than on the tokenizer's reconstruction capabilities. + +# 3.4. Final Training Objective + +The training objective of SoftVQ-VAE combines the reconstruction loss, perceptual loss [20, 46, 54, 120] and adversarial loss [33, 44] as in VQ [23], and representation alignment: + +$$ +\mathcal {L} = \mathcal {L} _ {\text {r e c o n}} + \lambda_ {1} \mathcal {L} _ {\text {p e r c e p}} + \lambda_ {2} \mathcal {L} _ {\text {a d v}} + \lambda_ {3} \mathcal {L} _ {\text {a l i g n}} + \lambda_ {4} \mathcal {L} _ {\mathrm {K L}}, \tag {7} +$$ + +where $\lambda_1, \lambda_2, \lambda_3,$ and $\lambda_4$ are hyper-parameters. Neither codebook nor commit loss as in VQ is needed. The perceptual loss $\mathcal{L}_{\mathrm{percep}}$ helps capture high-level perceptual features, $\mathcal{L}_{\mathrm{adv}}$ encourages the decoder to generate realistic images by removing the artifacts from the reconstruction loss only, $\mathcal{L}_{\mathrm{align}}$ guides the latent space to align with pre-trained features, and $\mathcal{L}_{\mathrm{kl}}$ as the KL divergence term in the ELBO. + +# 4. Experiments + +In this section, we validate the efficiency and effectiveness of SoftVQ-VAE with extensive experiments. + +# 4.1. Experiments Setup + +Implementation Details of SoftVQ-VAE. We use the LlamaGen codebase [98] to build our SoftVQ-VAE. We instantiate 4 configurations: SoftVQ-S, SoftVQ-B, SoftVQ-BL, SoftVQ-L, with a total of 45M, 173M, 391M, and 608M parameters, respectively. Each configuration has variants of latent codes $L = 64$ and $L = 32$ . For the representation alignment, we choose DINOv2 [79] for the pre-trained features, similarly as Yu et al. [116] and Li et al. [62]. To learn a better semantic latent space, we initialize our encoder from the pre-trained DINOv2 weights. Alignment with other pretrained features is explored in Sec. 4.4. We train the tokenizers on ImageNet [15] of resolution $256 \times 256$ and $512 \times 512$ for 250K iterations. Similarly to Tian et al. [99] and Li et al. [62], we adopted the same frozen DINO-S [9, 79] discriminator with a similar architecture to StyleGAN [48, 49]. We use DiffAug [121], consistency regularization [119], and LeCAM regularization [101] for discriminator training as in [99]. For the training objective, we set $\lambda_1 = 1.0$ , $\lambda_1 = 0.2$ , $\lambda_3 = 0.1$ , and $\lambda_4 = 0.01$ , following previous common practice. More training details are shown in Appendix C.1. + +Implementation Details of Generative Modeling. We select DiT [80], SiT [57], and MAR [60] for downstream denoising-based image generation tasks. For DiT and SiT, we set the patch size of DiT and SiT to 1 and use a 1D absolution position embedding. In our main experiments, we train DiT-XL and SiT-XL of 675M parameters for 3M steps, compared to 4M steps in REPA [116] and 7M steps vanilla version [72, 80]. We strictly follow their original training setup for other settings. For MAR-H, we train for 500 epochs and set the maximum learning rate to $2e-4$ since a higher learning rate leads to NAN issues. For other experiments, we simply train SiT-L of 458M parameters for 400K steps. More experimental details are provided in Appendix C.2. + +Evaluation. We use reconstruction Frechet Inception Distance (rFID) [37] on ImageNet validation set to evaluate the tokenizer. To evaluate the performance of generation tasks, we report generation FID (gFID), Inception Score (IS) [90], Precision and Recall [52] (in Appendix D.5), with and without classifier-free guidance (CFG) [40]. We measure the efficiency of the generative models by GLOPs of the model's forward pass on the latent codes of tokenizers, and training and inference throughput for the models using floating point 32 and a batch size of 64 on a single AMD MI250. + +# 4.2. Main Results + +We present the main reconstruction and generation results on the ImageNet benchmark of resolution $256 \times 256$ and $512 \times 512$ in Tab. 1 and Tab. 2, respectively. We show that SoftVQ-VAE achieves reconstruction and generation performance (with generative models) comparable to leading systems with 32 and 64 tokens, while presenting a significant improvement in both training and inference efficiency. + +Table 1. System-level comparison on ImageNet ${256} \times {256}$ conditional generation. We compare with both diffusion-based models and auto-regressive models with different types of tokenizers. Grey denotes SoftVQ. † indicates results with DPM-Solver and 50 steps. + +
Gen. ModelTok. Model# Tokens ↓Tok. rFID ↓Gflops ↓Throughput (imgs/sec) ↑w/o CFGw/ CFG
gFID ↓IS ↑gFID ↓IS ↑
Auto-regressive Models
Taming-Trans. [23]VQ2567.94--5.20290.3--
RQ-Trans. [55]RQ2563.20908.917.853.80323.7--
MaskGIT [10]VQ2562.28--6.18182.1--
MAGE [59]VQ256---6.93195.8--
LlamaGen-3B [98]VQ2562.19781.562.90--3.06279.7
TiTok-S-128 [115]VQ1281.6133.036.50--1.97281.8
MAR-H [60]KL2561.22145.080.122.35227.81.55303.7
MAR-HSoftVQ-L320.6167.932.083.83211.22.54273.6
SoftVQ-BL640.6586.550.892.81218.31.93289.4
Diffusion-based Models
LDM-4 [88]KL40960.27157.920.3710.56103.53.60247.7
U-ViT-H/2†[4]128.890.98--2.29263.9
MDTv2-XL/2 [29]125.430.595.06155.61.58314.7
DiT-XL/2 [80]KL10240.6280.730.519.62121.52.27278.2
SiT-XL/2 [72]81.920.548.30131.72.06270.3
+ REPA [116]5.90157.81.42305.7
SoftVQ-B0.898.949.83113.83.91264.2
SoftVQ-BL0.6814.528.819.22115.83.78266.7
SoftVQ-L0.748.749.07117.23.69270.4
DiT-XLSoftVQ-B0.884.706.62129.23.29262.5
SoftVQ-BL0.6528.814.596.53131.93.11268.3
SoftVQ-L0.614.515.83141.32.93268.5
SiT-XLSoftVQ-B0.8910.287.99129.32.51301.3
SoftVQ-BL0.6814.7310.127.67133.92.44308.8
SoftVQ-L0.7410.117.59137.42.44310.6
SiT-XLSoftVQ-B0.885.515.98138.01.78279.0
SoftVQ-BL0.6529.235.425.80143.51.88287.9
SoftVQ-L0.615.395.35151.21.86293.6
+ +Consistent high-quality reconstruction. SoftVQ variants achieve remarkable reconstruction quality with a high compression ratio. With only 32 and 64 tokens, our models maintain consistent rFID scores, i.e., 0.61-0.89 on the $256 \times 256$ and 0.64-0.71 on $512 \times 512$ benchmark, substantially outperforming VQ tokenizers with 256 tokens used in autoregressive models such as MaskGIT [10] and TiTok [115]. Our results are also comparable to KL tokenizers used in previous denoising-based models, using $32x$ fewer tokens. More discussion of different tokenizers is given in Sec. 4.3. + +Generation performance comparable to SOTA. Notably, with only 32 tokens, the generation performance of DiT-XL and SiT-XL trained on all variants of the proposed SoftVQ significantly outperforms DiT-XL/2 and SiT-XL/2 with KL tokenizers of 1024 tokens, without using CFG. DiT-XL with SoftVQ-L of 32 tokens presents a 0.62 improvement on FID and SiT-XL with SoftVQ-L of 32 tokens shows a 0.71 improvement on FID, without using CFG. The improvement is further strengthened with 64 tokens. SiT-XL with SoftVQ-L of 64 tokens outperforms SiT-XL/2 with REPA [116], without using CFG. Applying CFG, SiT-XL with SoftVQ of 64 tokens achieves a performance comparable to the leading systems with the best FID as 1.78 on 256 and 2.21 on 512 benchmark. Using MAR-H [60], SoftVQ-BL and SoftVQ-L with 64 and 32 tokens present slightly worse results, compared to the KL tokenizer of 256 tokens, possibly due to the + +lowered learning rate. MAR-H and SiT-XL results on 512 benchmark with 32 tokens are included in Appendix D.2. + +Efficiency. The system-level comparison reveals SoftVQ's exceptional efficiency across multiple metrics. When integrated with MAR-H at $256 \times 256$ resolution, SoftVQ reduces GFLOPs by $40\%$ compared to standard KL tokenizer, i.e., 86.55 vs 145.08 GFLOPs with 64 tokens and further reducing to 67.93 GFLOPs with 32 tokens. SoftVQ enables DiT-XL and SiT-XL to achieve significantly lower computational costs i.e., 28.81 and 29.23 GFLOPs, respectively with 64 tokens, compared to their KL counterparts (80.73 GFLOPs with 1024 tokens). Importantly, these efficiency improvements come without compromising generation quality. SiT-XL trained on SoftVQ-BL with 64 tokens presents $55\mathrm{x}$ throughput than the KL counterpart, while showing competitive IS and FID scores in all configurations. Not only do we require fewer training iterations to achieve a comparable performance to leading systems, i.e., 2.5M vs 7M steps of SiT-XL and DiT-XL, the high compression ratio of SoftVQ also improves the training throughput, reducing the time to train SiT-XL for 400K steps from 72 to 20 hours on 8 MI250. + +# 4.3. Comparison of Tokenizers + +We compare the proposed SoftVQ-VAE with the concurrent efficient image tokenizers, i.e., TiTok [115] and DC-AE [11]. To show the superiority of SoftVQ-VAE in achieving a + +Table 2. System-level comparison on ImageNet ${512} \times {512}$ conditional generation. We compare with both diffusion-based models and auto-regressive models with different types of tokenizers. Grey denotes SoftVQ. † indicates results with DPM-Solver and 50 steps. + +
Gen. ModelTok. Model# Tokens ↓Tok. rFID ↓Gflops ↓Throughput (imgs/sec) ↑w/o CFGw/ CFG
gFID ↓IS ↑gFID ↓IS ↑
Generative Adversarial Models
BigGAN [10]-------8.43177.9
StyleGAN-XL [48]-------2.41267.7
Auto-regressive Models
MaskGIT [10]VQ10241.97--7.32156.0--
TiTok-B [115]VQ1281.52----2.13261.2
MAR-HSoftVQ-BL640.7186.551.508.21152.93.42261.8
Diffusion-based Models
ADM [16]-----23.2458.063.85221.7
U-ViT-H/4† [4]128.920.58--4.05263.8
DiT-XL/2 [80]KL40960.62373.340.109.62121.53.04240.8
SiT-XL/2 [72]373.320.10--2.62252.2
DiT-XL [11]AE2560.2280.751.029.56-2.84-
UViT-H† [11]128.906.149.83-2.53-
UViT-H† [11]AE640.2232.9915.2312.26-2.66-
UViT-2B† [11]104.189.446.50-2.25-
SiT-XLSoftVQ-BL640.7129.235.387.96133.92.21290.5
+ +high compression ratio, we additionally train VQ-VAE and AE using the small (S) configuration and the same training recipe of SoftVQ. To validate the generation performance, we train a SiT-L for 400K steps and report gFID and IS on 256 ImageNet without using CFG, as shown in Tab. 3. + +Scalable performance. With a much smaller model size, i.e., 46M vs 390M, SoftVQ-S significantly outperforms both TiTok variants at 64 tokens, achieving better reconstruction quality with an rFID of 1.03 compared to 1.25, and better generation performance with a gFID of 11.24 and IS of 89.4 compared to the best TiTok gFID of 19.23 and IS of 61.8. Noteworthy is that the generation results SiT-L trained on SoftVQ with 32 tokens outperform those trained on KL tokenizer with 1024 tokens [72] by a large margin, i.e., 5.9 in gFID. Compared to DC-AE, SoftVQ-S demonstrates competitive scalability across token counts while maintaining significant generation quality even with fewer parameters. + +Less Lossy Property. SoftVQ-S exhibits remarkably consistent performance across different compression ratios. When reducing tokens from 256 to 32, SoftVQ-S maintains a minimal degradation in rFID from 0.80 to 1.24, significantly outperforming both VQ-S, from 1.45 to 10.97, and AE-S, from 1.15 to 2.01. The benefits of the fully-differentiable property of SoftVQ along with representation alignment are particularly evident in generation quality, where SoftVQ-S achieves substantially better gFID of 9.21-12.89, and IS scores of 79.5-93.6, compared to baselines, showing its robustness in preserving information at high compression ratios and a better latent space for generative models. + +# 4.4. Discussions on the Latent Space + +We perform more analysis on the latent of SoftVQ here. + +Alignment with different initialization and target models. As shown in Tab. 4, our results demonstrate the effectiveness of alignment across various pre-trained models. The + +Table 3. Comparison of tokenizers on class-conditional ${256} \times {256}$ ImageNet. We report rFID, and gFID and IS of SiT-L without CFG. + +
Tokenizer# Params# TokensDim.rFID ↓SiT-L
gFID ↓IS ↑
KL [96]676M102440.6218.7972.0
TiTok-BL-KL [115]389M64161.2523.3554.7
TiTok-BL-VQ [115]390M64162.0619.2361.8
DC-AE-f32 [11]323M64320.6915.3676.1
DC-AE-f54 [11]677M161280.8124.6649.1
VQ-S46M25681.4519.0258.6
128162.6118.5162.4
64324.0420.0754.5
326410.9721.5747.6
AE-S46M25681.1522.1158.3
128161.3322.3158.5
64321.6425.5354.3
32642.0128.1845.5
SoftVQ-S46M25680.809.2193.6
128160.9210.1285.8
64321.0311.2489.4
32641.2412.8979.5
+ +encoder initialization and alignment with DINOv2-B [79] achieve superior performance with rFID 0.88 and IS 103.4, compared to using either component alone. Further improvements are observed with CLIP-B [82] and EVA-02-B [26] on reconstruction. We reveal that a better rFID does not necessarily translate to a better gFID. Instead, the latent space quality, reflected by linear probing accuracy, is more closely related to the performance of the generative models. + +Linear probing of tokenizer and generative model. Linear probing results in Fig. 3 (details in Appendix C.3) reveal SoftVQ's superior representation quality at both tokenizer and generative model. At the tokenizer level, the linear probing accuracy of SoftVQ variants (B, BL, L) consistently outperforms VQ-S and AE-S across all token counts, with the gaps widening at lower token counts, similarly observed by Yu et al. [115]. This advantage transfers to the generative model with the best linear probing accuracy on the + +Table 4. Representation alignment with different encoder initialization and the target alignment models using SoftVQ-B 64 tokens. + +
Enc. Init.Align. TargetrFID ↓L.P. ↑SiT-L
gFID ↓IS ↑
--0.985.4220.3389.4
-DINOv2-B [79]0.9341.0810.96101.9
DINOv2-B [79]-0.7311.8717.2091.5
DINOv2-B [79]DINOv2-B [79]0.8842.4210.13103.4
CLIP-B [82]CLIP-B [82]0.7547.129.89109.3
EVA-02-B [26]EVA-02-B [26]0.8144.2310.34112.7
+ +intermediate features of SiT trained with SoftVQ. + +Latent space visualization. UMAP visualizations latent space in Fig. 4 demonstrate SoftVQ's ability totain structured and discriminative latent representation minimal variance between encoder output $(\hat{\mathbf{z}})$ and d input (z). Comparing Fig. 4 (d) and (c), we also observe larger encoder can learn a more discriminative latent showing the superiority of SoftVQ's representation le + +# 4.5. Ablation Studies + +We present a series of ablation studies in Append including SoftVQ variants with product quantization [45], residual quantization (RQ) [55], and GMMVQ book size, latent space size, and the temperature of S The results show the compatibility of SoftVQ with I RQ, with slightly improved performance. GMMVQ p results comparable to SoftVQ. Moreover, SoftVQ shows robustness to other parameters, such as codebook size and softmax temperature. We found that while increasing the dimension of the latent space can result in a significant improvement in reconstruction, it leads to deterioration in generation performance, possibly due to the learning difficulty of generative models with larger dimensions [92]. + +# 5. Related Work + +Image tokenization has emerged as a fundamental technique to bridge various vision tasks. Traditional approaches using autoencoders [39, 106] laid the groundwork by compressing images into latent representations. This field subsequently diverged into two main branches: generation-focused and understanding-focused approaches. Generation-oriented methods, such as VAE [84, 104] and VQ-GAN [23, 85], emphasized learning latent spaces for detail-preserving compression. These were further refined through variants [55, 73, 112, 124] that enhanced generation quality. Parallel developments in understanding-focused approaches leveraged Large Language Models (LLMs) [105] to create semantic representations for tasks like classification [21] and detection [125]. Recent work [62, 115] has demonstrated the potential of unifying these approaches, with subsequent research [34, 109] exploring the convergence of generation and understanding capabilities using a single tokenizer. + +Image generation have two popular paradigms: autoregressive and diffusion. Autoregressive models start from CNN-based architectures [103] and evolve to the transformer- + +![](images/c62c016c899d097d2b6ba4621e9d2176cf6e114101d22392b9e37a8a01bf5f81.jpg) + +![](images/44aa5ef680a0f6a0e40648c3631596a68ea02c174ac34dee31e57d756c0eee2c.jpg) +Figure 4. Visualization of $\hat{\mathbf{z}}$ (top), i.e., encoder output, and $\mathbf{z}$ (bottom), i.e., decoder input, of (a) VQ-S 64; (b) SoftVQ-S 64; (c) SoftVQ-S 32; (d) SoftVQ-L 32, trained with latent space alignment. + +based architectures [55, 66, 98, 105, 114] to enhance scalability. Latest works in visual autoregressive modeling, including VAR [99], MAR [60], and ImageFolder [62], have further enhanced generation efficiency. Diffusion models have transformed image generation since their introduction by Sohl-Dickstein et al. [93]. Key advancements by Nichol et al. [75], Dhariwal et al. [16], and Song et al. [95] have significantly improved model performance. A paradigm shift occurred with the adoption of latent space diffusion [88, 102], leveraging pre-trained image encoders as priors [23, 104] to enhance generation efficiency. Architectural innovations, particularly the integration of transformers [80], have further expanded the capabilities of these models. These advances have spawned numerous influential systems, from text-guided generation [18, 76] to high-fidelity synthesis [28, 89]. Commercial successes [74, 78, 83, 87] have also demonstrated the practical impact of diffusion models. + +# 6. Conclusion + +We presented SoftVQ-VAE, a continuous image tokenizer leveraging soft categorical posteriors to aggregate multiple codewords into each latent token. Our approach compresses images to just 32 or 64 tokens while maintaining high reconstruction quality and enabling state-of-the-art generation results with DiT, SiT, and MAR. The substantial improvements in both inference throughput (up to $55\mathrm{x}$ with 64 tokens) and training efficiency, combined with the fully-differentiable design's ability to learn semantic-rich representations, demonstrate SoftVQ-VAE's potential as a foundation for efficient generative and vision-language models. + +# References + +[1] Michael S Albergo, Nicholas M Boffi, and Eric Vanden-Eijnden. Stochastic interpolants: A unifying framework for flows and diffusions. arXiv preprint arXiv:2303.08797, 2023. 3 +[2] Brian DO Anderson. Reverse-time diffusion equation models. Stochastic Processes and their Applications, 12(3): 313-326, 1982. 3 +[3] Alexei Baevski, Yuhao Zhou, Abdelrahman Mohamed, and Michael Auli. wav2vec 2.0: A framework for self-supervised learning of speech representations. Advances in neural information processing systems, 33:12449-12460, 2020. 3 +[4] Fan Bao, Shen Nie, Kaiwen Xue, Yue Cao, Chongxuan Li, Hang Su, and Jun Zhu. All are worth words: A vit backbone for diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 22669-22679, 2023. 6, 7, 5 +[5] Hangbo Bao, Li Dong, Songhao Piao, and Furu Wei. Beit: Bert pre-training of image transformers. arXiv preprint arXiv:2106.08254, 2021. 4 +[6] Omer Bar-Tal, Hila Chefer, Omer Tov, Charles Herrmann, Roni Paiss, Shiran Zada, Ariel Ephrat, Junhwa Hur, Guanghui Liu, Amit Raj, et al. Lumiere: A spacetime diffusion model for video generation. arXiv preprint arXiv:2401.12945, 2024. 1 +[7] Yoshua Bengio, Nicholas Léonard, and Aaron Courville. Estimating or propagating gradients through stochastic neurons for conditional computation. arXiv preprint arXiv:1308.3432, 2013. 1, 3 +[8] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 1 +[9] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 5, 2 +[10] Huiwen Chang, Han Zhang, Lu Jiang, Ce Liu, and William T. Freeman. Maskgit: Masked generative image transformer, 2022. 1, 3, 6, 7, 5 +[11] Junyu Chen, Han Cai, Junsong Chen, Enze Xie, Shang Yang, Haotian Tang, Muyang Li, Yao Lu, and Song Han. Deep compression autoencoder for efficient high-resolution diffusion models. arXiv preprint arXiv:2410.10733, 2024. 1, 6, 7, 5 +[12] Liang Chen, Haozhe Zhao, Tianyu Liu, Shuai Bai, Junyang Lin, Chang Zhou, and Baobao Chang. An image is worth 1/2 tokens after layer 2: Plug-and-play inference acceleration for large vision-language models. arXiv preprint arXiv:2403.06764, 2024. 1 +[13] Xi Chen, Diederik P Kingma, Tim Salimans, Yan Duan, Prafulla Dhariwal, John Schulman, Ilya Sutskever, and Pieter Abbeel. Variational lossy autoencoder. arXiv preprint arXiv:1611.02731, 2016. 1, 2, 3 + +[14] Xinlei Chen, Zhuang Liu, Saining Xie, and Kaiming He. Deconstructing denoising diffusion models for self-supervised learning. arXiv preprint arXiv:2401.14404, 2024. 1, 2, 3 +[15] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248–255. IEEE, 2009. 5, 2 +[16] Prafulla Dhariwal and Alex Nichol. Diffusion models beat gans on image synthesis, 2021. 1, 7, 8, 5 +[17] Nat Dilokthanakul, Pedro AM Mediano, Marta Garnelo, Matthew CH Lee, Hugh Salimbeni, Kai Arulkumaran, and Murray Shanahan. Deep unsupervised clustering with gaussian mixture variational autoencoders. arXiv preprint arXiv:1611.02648, 2016. 4 +[18] Ming Ding, Zhuoyi Yang, Wenyi Hong, Wendi Zheng, Chang Zhou, Da Yin, Junyang Lin, Xu Zou, Zhou Shao, Hongxia Yang, et al. Cogview: Mastering text-to-image generation via transformers. Advances in neural information processing systems, 34:19822-19835, 2021. 8 +[19] Xiaoyi Dong, Jianmin Bao, Ting Zhang, Dongdong Chen, Weiming Zhang, Lu Yuan, Dong Chen, Fang Wen, Nenghai Yu, and Baining Guo. Peco: Perceptual codebook for bert pre-training of vision transformers. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 552-560, 2023. 2 +[20] Alexey Dosovitskiy and Thomas Brox. Generating images with perceptual similarity metrics based on deep networks. Advances in neural information processing systems, 29, 2016. 5 +[21] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale, 2021. 3, 8 +[22] Joseph C Dunn. A fuzzy relative of the isodata process and its use in detecting compact well-separated clusters. 1973. 2, 4 +[23] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12873-12883, 2021. 1, 2, 3, 5, 6, 8 +[24] Zach Evans, CJ Carr, Josiah Taylor, Scott H Hawley, and Jordi Pons. Fast timing-conditioned latent audio diffusion. arXiv preprint arXiv:2402.04825, 2024. 1 +[25] Lijie Fan, Tianhong Li, Siyang Qin, Yuzheng Li, Chen Sun, Michael Rubinstein, Deqing Sun, Kaiming He, and Yonglong Tian. Fluid: Scaling autoregressive text-to-image generative models with continuous tokens. arXiv preprint arXiv:2410.13863, 2024. 1 +[26] Yuxin Fang, Quan Sun, Xinggang Wang, Tiejun Huang, Xinlong Wang, and Yue Cao. Eva-02: A visual representation for neon genesis. Image and Vision Computing, 149:105171, 2024. 7, 8 +[27] Christopher Fifty, Ronald G Junkins, Dennis Duan, Aniketh Iger, Jerry W Liu, Ehsan Amid, Sebastian Thrun, and + +Christopher Ré. Restructuring vector quantization with the rotation trick. arXiv preprint arXiv:2410.06424, 2024. 2, 3 +[28] Oran Gafni, Adam Polyak, Oron Ashual, Shelly Sheynin, Devi Parikh, and Yaniv Taigman. Make-a-scene: Scene-based text-to-image generation with human priors. In European Conference on Computer Vision, pages 89-106. Springer, 2022. 8 +[29] Shanghua Gao, Pan Zhou, Ming-Ming Cheng, and Shuicheng Yan. Mdtv2: Masked diffusion transformer is a strong image synthesizer. arXiv preprint arXiv:2303.14389, 2023.6,5 +[30] Yuying Ge, Yixiao Ge, Ziyun Zeng, Xintao Wang, and Ying Shan. Planting a seed of vision in large language model, 2023. 1, 2, 4 +[31] Yuying Ge, Sijie Zhao, Ziyun Zeng, Yixiao Ge, Chen Li, Xintao Wang, and Ying Shan. Making llama see and draw with seed tokenizer. arXiv preprint arXiv:2310.01218, 2023. 2 +[32] Shansan Gong, Mukai Li, Jiangtao Feng, Zhiyong Wu, and LingPeng Kong. Diffuseq: Sequence to sequence text generation with diffusion models. arXiv preprint arXiv:2210.08933, 2022. 1 +[33] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139-144, 2020. 5 +[34] Yuchao Gu, Xintao Wang, Yixiao Ge, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Rethinking the objectives of vector-quantized tokenizers for image synthesis, 2023. 2, 8 +[35] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 1, 3, 4 +[36] Byeongho Heo, Song Park, Dongyoon Han, and Sangwoo Yun. Rotary position embedding for vision transformer. In European Conference on Computer Vision, pages 289-305. Springer, 2025. 4 +[37] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in Neural Information Processing Systems, 30, 2017. 5 +[38] Irina Higgins, Loic Matthew, Arka Pal, Christopher P Burgess, Xavier Glorot, Matthew M Botvinick, Shakir Mohamed, and Alexander Lerchner. beta-vae: Learning basic visual concepts with a constrained variational framework. ICLR, 3, 2017. 1, 2, 3 +[39] Geoffrey E Hinton and Ruslan R Salakhutdinov. Reducing the dimensionality of data with neural networks. science, 313(5786):504-507, 2006. 8 +[40] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 5 +[41] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 3 + +[42] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. Advances in Neural Information Processing Systems, 35:8633-8646, 2022. 1 +[43] Minyoung Huh, Brian Cheung, Pulkit Agrawal, and Phillip Isola. Straightening out the straight-through estimator: Overcoming optimization challenges in vector quantized networks. In International Conference on Machine Learning, pages 14096-14113. PMLR, 2023. 3, 4 +[44] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A. Efros. Image-to-image translation with conditional adversarial networks, 2018. 5 +[45] Herve Jegou, Matthijs Douze, and Cordelia Schmid. Product quantization for nearest neighbor search. IEEE transactions on pattern analysis and machine intelligence, 33(1):117-128, 2010. 2, 8 +[46] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14, pages 694-711. Springer, 2016. 5 +[47] Matthew J Johnson, David K Duvenaud, Alex Wiltschko, Ryan P Adams, and Sandeep R Datta. Composing graphical models with neural networks for structured representations and fast inference. Advances in neural information processing systems, 29, 2016. 4 +[48] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 5, 7, 2 +[49] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8110-8119, 2020. 5, 2 +[50] Diederik Kingma and Ruiqi Gao. Understanding diffusion objectives as the elbo with simple data augmentation. Advances in Neural Information Processing Systems, 36, 2024. 3 +[51] Diederik P Kingma. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 1, 2, 3 +[52] Tuomas Kynkänniemi, Tero Karras, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Improved precision and recall metric for assessing generative models. Advances in neural information processing systems, 32, 2019. 5 +[53] Tuomas Kynkänniemi, Miika Aittala, Tero Karras, Samuli Laine, Timo Aila, and Jaakko Lehtinen. Applying guidance in a limited interval improves sample and distribution quality in diffusion models. arXiv preprint arXiv:2404.07724, 2024. 2 +[54] Anders Boesen Lindbo Larsen, Søren Kaae Sønderby, Hugo Larochelle, and Ole Winther. Autoencoding beyond pixels using a learned similarity metric. In International conference on machine learning, pages 1558-1566. PMLR, 2016. 5 +[55] Doyup Lee, Chiheon Kim, Saehoon Kim, Minsu Cho, and Wook-Shin Han. Autoregressive image generation using + +residual quantization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11523-11532, 2022. 2, 6, 8, 5 +[56] Doyup Lee, Chiheon Kim, Saehoon Kim, Minsu Cho, and Wook-Shin Han. Autoregressive image generation using residual quantization, 2022. 1 +[57] Haopeng Li, Jinyue Yang, Kexin Wang, Xuerui Qiu, Yuhong Chou, Xin Li, and Guoqi Li. Scalable autoregressive image generation with mamba. arXiv preprint arXiv:2408.12245, 2024. 5 +[58] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730-19742. PMLR, 2023. 1, 2 +[59] Tianhong Li, Huiwen Chang, Shlok Kumar Mishra, Han Zhang, Dina Katabi, and Dilip Krishnan. Mage: Masked generative encoder to unify representation learning and image synthesis, 2023. 3, 6, 5 +[60] Tianhong Li, Yonglong Tian, He Li, Mingyang Deng, and Kaiming He. Autoregressive image generation without vector quantization, 2024. 1, 2, 3, 5, 6, 8 +[61] Wentong Li, Yuqian Yuan, Jian Liu, Dongqi Tang, Song Wang, Jianke Zhu, and Lei Zhang. Tokenpacker: Efficient visual projector for multimodal llm. arXiv preprint arXiv:2407.02392, 2024. 1 +[62] Xiang Li, Hao Chen, Kai Qiu, Jason Kuen, Jiumiang Gu, Bhiksha Raj, and Zhe Lin. Imagefolder: Autoregressive image generation with folded tokens. arXiv preprint arXiv:2410.01756, 2024. 1, 2, 4, 5, 8 +[63] Yanwei Li, Yuechen Zhang, Chengyao Wang, Zhisheng Zhong, Yixin Chen, Ruihang Chu, Shaoteng Liu, and Jiaya Jia. Mini-gemini: Mining the potential of multi-modality vision language models. arXiv preprint arXiv:2403.18814, 2024. 1 +[64] Ziyi Lin, Chris Liu, Renrui Zhang, Peng Gao, Longtian Qiu, Han Xiao, Han Qiu, Chen Lin, Wenqi Shao, Keqin Chen, et al. *Sphinx: The joint mixing of weights, tasks, and visual embeddings for multi-modal large language models.* arXiv preprint arXiv:2311.07575, 2023. 2 +[65] Yaron Lipman, Ricky TQ Chen, Heli Ben-Hamu, Maximilian Nickel, and Matt Le. Flow matching for generative modeling. arXiv preprint arXiv:2210.02747, 2022. 3 +[66] Wenze Liu, Le Zhuo, Yi Xin, Sheng Xia, Peng Gao, and Xiangyu Yue. Customize your visual autoregressive recipe with set autoregressive modeling. arXiv preprint arXiv:2410.10511, 2024. 8 +[67] Stuart Lloyd. Least squares quantization in pmc. IEEE transactions on information theory, 28(2):129-137, 1982. 2, 4 +[68] I Loshchilov. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 2 +[69] Cheng Lu and Yang Song. Simplifying, stabilizing and scaling continuous-time consistency models. arXiv preprint arXiv:2410.11081, 2024. 1 +[70] Gen Luo, Yiyi Zhou, Yuxin Zhang, Xiawu Zheng, Xiaoshuai Sun, and Rongrong Ji. Feast your eyes: Mixture-of + +resolution adaptation for multimodal large language models. arXiv preprint arXiv:2403.03003, 2024.1 +[71] Zhuoyan Luo, Fengyuan Shi, Yixiao Ge, Yujiu Yang, Limin Wang, and Ying Shan. Open-magtvit2: An open-source project toward democratizing auto-regressive visual generation. arXiv preprint arXiv:2409.04410, 2024. 2 +[72] Nanye Ma, Mark Goldstein, Michael S Albergo, Nicholas M Boffi, Eric Vanden-Eijnden, and Saining Xie. Sit: Exploring flow and diffusion-based generative models with scalable interpolant transformers. arXiv preprint arXiv:2401.08740, 2024. 1, 2, 3, 5, 6, 7 +[73] Fabian Mentzer, David Minnen, Eirikur Agustsson, and Michael Tschannen. Finite scalar quantization: Vq-vae made simple, 2023. 2, 8 +[74] MidJourney Inc. Midjourney. Software available from MidJourney Inc., 2022. 8 +[75] Alex Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models, 2021. 8 +[76] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021.8 +[77] Harry Nyquist. Certain topics in telegraph transmission theory. Transactions of the American Institute of Electrical Engineers, 47(2):617-644, 1928. 2 +[78] OpenAI. Video generation models as world simulators. OpenAI Blog, 2024. 8 +[79] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Russell Howes, Po-Yao Huang, Hu Xu, Vasu Sharma, Shang-Wen Li, Wojciech Galuba, Mike Rabbat, Mido Assran, Nicolas Ballas, Gabriel Synnaeve, Ishan Misra, Herve Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. Dinov2: Learning robust visual features without supervision, 2023. 5, 7, 8, 2 +[80] William Peebles and Saining Xie. Scalable diffusion models with transformers, 2023. 1, 2, 3, 5, 6, 7, 8 +[81] Vadim Popov, Ivan Vovk, Vladimir Gogoryan, Tasnima Sadekova, and Mikhail Kudinov. Grad-tts: A diffusion probabilistic model for text-to-speech. In International Conference on Machine Learning, pages 8599-8608. PMLR, 2021. 1 +[82] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 7, 8 +[83] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International conference on machine learning, pages 8821-8831. Pmlr, 2021. 1, 8 +[84] Ali Razavi, Aaron Van den Oord, and Oriol Vinyals. Generating diverse high-fidelity images with vq-vae-2. Advances in neural information processing systems, 32, 2019. 1, 8 + +[85] Ali Razavi, Aaron van den Oord, and Oriol Vinyals. Generating diverse high-fidelity images with vq-vae-2, 2019. 8 +[86] Douglas A Reynolds et al. Gaussian mixture models. Encyclopedia of biometrics, 741(659-663), 2009. 2, 4 +[87] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 8 +[88] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models, 2022. 1, 6, 8, 5 +[89] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in neural information processing systems, 35:36479-36494, 2022. 8 +[90] Tim Salimans, Ian Goodfellow, Wojciech Zaremba, Vicki Cheung, Alec Radford, and Xi Chen. Improved techniques for training gans. Advances in Neural Information Processing Systems, 29, 2016. 5 +[91] Claude Elwood Shannon. Communication in the presence of noise. Proceedings of the IRE, 37(1):10-21, 1949. 2 +[92] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International conference on machine learning, pages 2256-2265. PMLR, 2015. 3, 8 +[93] Jascha Sohl-Dickstein, Eric A. Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics, 2015. 8 +[94] Dingjie Song, Wenjun Wang, Shunian Chen, Xidong Wang, Michael Guan, and Benyou Wang. Less is more: A simple yet effective token reduction method for efficient multimodal llms. arXiv preprint arXiv:2409.10994, 2024. 1 +[95] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models, 2022. 8 +[96] stabilityai. Sd vae ft ema, 2023. Accessed: 2023. 1, 3, 7 +[97] Jianlin Su, Murtadha Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu. Roformer: Enhanced transformer with rotary position embedding. Neurocomputing, 568:127063, 2024. 4 +[98] Peize Sun, Yi Jiang, Shoufa Chen, Shilong Zhang, Bingyue Peng, Ping Luo, and Zehuan Yuan. Autoregressive model beats diffusion: Llama for scalable image generation. arXiv preprint arXiv:2406.06525, 2024. 1, 5, 6, 8 +[99] Keyu Tian, Yi Jiang, Zehuan Yuan, Bingyue Peng, and Liwei Wang. Visual autoregressive modeling: Scalable image generation via next-scale prediction, 2024. 3, 5, 8, 2 +[100] Michael Tschannen, Cian Eastwood, and Fabian Mentzer. Givt: Generative infinite-vocabulary transformers. In European Conference on Computer Vision, pages 292-309. Springer, 2025. 3 +[101] Hung-Yu Tseng, Lu Jiang, Ce Liu, Ming-Hsuan Yang, and Weilong Yang. Regularizing generative adversarial networks under limited data, 2021. 5, 2 + +[102] Arash Vahdat, Karsten Kreis, and Jan Kautz. Score-based generative modeling in latent space, 2021. 8 +[103] Aaron Van den Oord, Nal Kalchbrenner, Lasse Espeholt, Oriol Vinyals, Alex Graves, et al. Conditional image generation with pixelCNN decoders. Advances in neural information processing systems, 29, 2016. 8 +[104] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017. 1, 2, 3, 8 +[105] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need, 2023. 1, 8 +[106] Pascal Vincent, Hugo Larochelle, Yoshua Bengio, and Pierre-Antoine Manzagol. Extracting and composing robust features with denoising autoencoders. In Proceedings of the 25th international conference on Machine learning, pages 1096-1103, 2008. 8 +[107] Xinlong Wang, Xiaosong Zhang, Zhengxiong Luo, Quan Sun, Yufeng Cui, Jinsheng Wang, Fan Zhang, Yueze Wang, Zhen Li, Qiying Yu, et al. Emu3: Next-token prediction is all you need. arXiv preprint arXiv:2409.18869, 2024. 1 +[108] Mark Weber, Lijun Yu, Qihang Yu, Xueqing Deng, Xiaohui Shen, Daniel Cremers, and Liang-Chieh Chen. Maskbit: Embedding-free image generation via bit tokens. arXiv preprint arXiv:2409.16211, 2024. 2 +[109] Yecheng Wu, Zhuoyang Zhang, Junyu Chen, Haotian Tang, Dacheng Li, Yunhao Fang, Ligeng Zhu, Enze Xie, Hongxu Yin, Li Yi, et al. Vila-u: a unified foundation model integrating visual understanding and generation. arXiv preprint arXiv:2409.04429, 2024. 1, 2, 4, 8 +[110] Jinheng Xie, Weijia Mao, Zechen Bai, David Junhao Zhang, Weihao Wang, Kevin Qinghong Lin, Yuchao Gu, Zhijie Chen, Zhenheng Yang, and Mike Zheng Shou. Show-o: One single transformer to unify multimodal understanding and generation. arXiv preprint arXiv:2408.12528, 2024. 1 +[111] Jiahui Yu, Xin Li, Jing Yu Koh, Han Zhang, Ruoming Pang, James Qin, Alexander Ku, Yuanzhong Xu, Jason Baldridge, and Yonghui Wu. Vector-quantized image modeling with improved vqgan. arXiv preprint arXiv:2110.04627, 2021. 2, 3 +[112] Lijun Yu, José Lezama, Nitesh B Gundavarapu, Luca Versari, Kihyuk Sohn, David Minnen, Yong Cheng, Vighnesh Birodkar, Agrim Gupta, Xiuye Gu, et al. Language model beats diffusion-tokenizer is key to visual generation. arXiv preprint arXiv:2310.05737, 2023. 1, 2, 8 +[113] Lijun Yu, Yong Cheng, Zhiruo Wang, Vivek Kumar, Wolfgang Macherey, Yanping Huang, David Ross, Irfan Essa, Yonatan Bisk, Ming-Hsuan Yang, et al. Spae: Semantic pyramid autoencoder for multimodal generation with frozen llms. Advances in Neural Information Processing Systems, 36, 2024. 2 +[114] Qihang Yu, Ju He, Xueqing Deng, Xiaohui Shen, and Liang-Chieh Chen. Randomized autoregressive visual generation. arXiv preprint arXiv:2411.00776, 2024. 8 +[115] Qihang Yu, Mark Weber, Xueqing Deng, Xiaohui Shen, Daniel Cremers, and Liang-Chieh Chen. An image is worth 32 tokens for reconstruction and generation. arxiv: 2406.07550, 2024. 1, 2, 4, 6, 7, 8, 5 + +[116] Sihyun Yu, Sangkyung Kwak, Huiwon Jang, Jongheon Jeong, Jonathan Huang, Jinwoo Shin, and Saining Xie. Representation alignment for generation: Training diffusion transformers is easier than you think. arXiv preprint arXiv:2410.06940, 2024. 1, 2, 4, 5, 6 +[117] Neil Zeghidour, Alejandro Luebs, Ahmed Omran, Jan Skoglund, and Marco Tagliasacchi. Soundstream: An end-to-end neural audio codec. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 30:495-507, 2021. 1 +[118] Baoquan Zhang, Huaibin Wang, Chuyao Luo, Xutao Li, Guotao Liang, Yunming Ye, Xiaochen Qi, and Yao He. Codebook transfer with part-of-speech for vector-quantized image modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7757–7766, 2024. 2 +[119] Han Zhang, Zizhao Zhang, Augustus Odena, and Honglak Lee. Consistency regularization for generative adversarial networks. arXiv preprint arXiv:1910.12027, 2019. 5, 2 +[120] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric, 2018. 5 +[121] Shengyu Zhao, Zhijian Liu, Ji Lin, Jun-Yan Zhu, and Song Han. Differentiable augmentation for data-efficient gan training. Advances in neural information processing systems, 33:7559-7570, 2020. 5, 2 +[122] Lin Zheng, Jianbo Yuan, Lei Yu, and Lingpeng Kong. A reparameterized discrete diffusion model for text generation. arXiv preprint arXiv:2302.05737, 2023. 1 +[123] Chunting Zhou, Lili Yu, Arun Babu, Kushal Tirumala, Michihiro Yasunaga, Leonid Shamis, Jacob Kahn, Xuezhe Ma, Luke Zettlemoyer, and Omer Levy. Transfusion: Predict the next token and diffuse images with one multi-modal model. arXiv preprint arXiv:2408.11039, 2024. 1 +[124] Lei Zhu, Fangyun Wei, Yanye Lu, and Dong Chen. Scaling the codebook size of vqgan to 100,000 with a utilization rate of $99\%$ . arXiv preprint arXiv:2406.11837, 2024. 2, 4, 8 +[125] X Zhu, W Su, L Lu, B Li, X Wang, and J Dai. Deformable detr: Deformable transformers for end-to-end object detection. arxiv 2020. arXiv preprint arXiv:2010.04159, 2010.8 +[126] Yongxin Zhu, Bocheng Li, Yifei Xin, and Linli Xu. Addressing representation collapse in vector quantized models with one linear layer. arXiv preprint arXiv:2411.02038, 2024. 2 \ No newline at end of file diff --git a/2025/SoftVQ-VAE_ Efficient 1-Dimensional Continuous Tokenizer/images.zip b/2025/SoftVQ-VAE_ Efficient 1-Dimensional Continuous Tokenizer/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..c9ad99488cde4f3c52d36b191ca5b2a490f7cc17 --- /dev/null +++ b/2025/SoftVQ-VAE_ Efficient 1-Dimensional Continuous Tokenizer/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8e0d07af631d3b3e0232f6064423074472205916d94169405fcbdafb7deccaf6 +size 571952 diff --git a/2025/SoftVQ-VAE_ Efficient 1-Dimensional Continuous Tokenizer/layout.json b/2025/SoftVQ-VAE_ Efficient 1-Dimensional Continuous Tokenizer/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..a19c2167c5a5d3cc33b1c4fcea6160e49b271579 --- /dev/null +++ b/2025/SoftVQ-VAE_ Efficient 1-Dimensional Continuous Tokenizer/layout.json @@ -0,0 +1,11275 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 118, + 103, + 492, + 120 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 103, + 492, + 120 + ], + "spans": [ + { + "bbox": [ + 118, + 103, + 492, + 120 + ], + "type": "text", + "content": "SoftVQ-VAE: Efficient 1-Dimensional Continuous Tokenizer" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 119, + 142, + 491, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 119, + 142, + 491, + 171 + ], + "spans": [ + { + "bbox": [ + 119, + 142, + 491, + 171 + ], + "type": "text", + "content": "Hao Chen" + }, + { + "bbox": [ + 119, + 142, + 491, + 171 + ], + "type": "inline_equation", + "content": "^{1,2*}" + }, + { + "bbox": [ + 119, + 142, + 491, + 171 + ], + "type": "text", + "content": ", Ze Wang" + }, + { + "bbox": [ + 119, + 142, + 491, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 119, + 142, + 491, + 171 + ], + "type": "text", + "content": ", Xiang Li" + }, + { + "bbox": [ + 119, + 142, + 491, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 119, + 142, + 491, + 171 + ], + "type": "text", + "content": ", Xineng Sun" + }, + { + "bbox": [ + 119, + 142, + 491, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 119, + 142, + 491, + 171 + ], + "type": "text", + "content": ", Fangyi Chen" + }, + { + "bbox": [ + 119, + 142, + 491, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 119, + 142, + 491, + 171 + ], + "type": "text", + "content": ", Jiang Liu" + }, + { + "bbox": [ + 119, + 142, + 491, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 119, + 142, + 491, + 171 + ], + "type": "text", + "content": ", Jindong Wang" + }, + { + "bbox": [ + 119, + 142, + 491, + 171 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 119, + 142, + 491, + 171 + ], + "type": "text", + "content": ", Bhiksha Raj" + }, + { + "bbox": [ + 119, + 142, + 491, + 171 + ], + "type": "inline_equation", + "content": "^{1,4}" + }, + { + "bbox": [ + 119, + 142, + 491, + 171 + ], + "type": "text", + "content": ", Zicheng Liu" + }, + { + "bbox": [ + 119, + 142, + 491, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 119, + 142, + 491, + 171 + ], + "type": "text", + "content": ", Emad Barsoum" + }, + { + "bbox": [ + 119, + 142, + 491, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 181, + 172, + 427, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 181, + 172, + 427, + 185 + ], + "spans": [ + { + "bbox": [ + 181, + 172, + 427, + 185 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 181, + 172, + 427, + 185 + ], + "type": "text", + "content": "Carnegie Mellon University, " + }, + { + "bbox": [ + 181, + 172, + 427, + 185 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 181, + 172, + 427, + 185 + ], + "type": "text", + "content": "AMD, " + }, + { + "bbox": [ + 181, + 172, + 427, + 185 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 181, + 172, + 427, + 185 + ], + "type": "text", + "content": "William & Mary, " + }, + { + "bbox": [ + 181, + 172, + 427, + 185 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 181, + 172, + 427, + 185 + ], + "type": "text", + "content": "MBZUAI" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 213, + 200, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 213, + 200, + 225 + ], + "spans": [ + { + "bbox": [ + 151, + 213, + 200, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 237, + 297, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 237, + 297, + 514 + ], + "spans": [ + { + "bbox": [ + 55, + 237, + 297, + 514 + ], + "type": "text", + "content": "Efficient image tokenization with high compression ratios remains a critical challenge for training generative models. We present SoftVQ-VAE, a continuous image tokenizer that leverages soft categorical posteriors to aggregate multiple codewords into each latent token, substantially increasing the representation capacity of the latent space. When applied to Transformer-based architectures, our approach compresses " + }, + { + "bbox": [ + 55, + 237, + 297, + 514 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 55, + 237, + 297, + 514 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 237, + 297, + 514 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 55, + 237, + 297, + 514 + ], + "type": "text", + "content": " images using as few as 32 or 64 1-dimensional tokens. Not only does SoftVQ-VAE show consistent and high-quality reconstruction, more importantly, it also achieves state-of-the-art and significantly faster image generation results across different denoising-based generative models. Remarkably, SoftVQ-VAE improves inference throughput by up to " + }, + { + "bbox": [ + 55, + 237, + 297, + 514 + ], + "type": "inline_equation", + "content": "18x" + }, + { + "bbox": [ + 55, + 237, + 297, + 514 + ], + "type": "text", + "content": " for generating " + }, + { + "bbox": [ + 55, + 237, + 297, + 514 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 55, + 237, + 297, + 514 + ], + "type": "text", + "content": " images and " + }, + { + "bbox": [ + 55, + 237, + 297, + 514 + ], + "type": "inline_equation", + "content": "55x" + }, + { + "bbox": [ + 55, + 237, + 297, + 514 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 55, + 237, + 297, + 514 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 55, + 237, + 297, + 514 + ], + "type": "text", + "content": " images while achieving competitive FID scores of 1.78 and 2.21 for SiT-XL. It also improves the training efficiency of the generative models by reducing the number of training iterations by " + }, + { + "bbox": [ + 55, + 237, + 297, + 514 + ], + "type": "inline_equation", + "content": "2.3x" + }, + { + "bbox": [ + 55, + 237, + 297, + 514 + ], + "type": "text", + "content": " while maintaining comparable performance. With its fully-differentiable design and semantic-rich latent space, our experiment demonstrates that SoftVQ-VAE achieves efficient tokenization without compromising generation quality, paving the way for more efficient generative models. Code and model are released1." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 534, + 135, + 547 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 534, + 135, + 547 + ], + "spans": [ + { + "bbox": [ + 56, + 534, + 135, + 547 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 555, + 297, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 555, + 297, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 555, + 297, + 685 + ], + "type": "text", + "content": "Denoising-based generative modeling has witnessed remarkable progress with recent advances, such as Diffusion Transformers (DiT) [80], Scalable Interpolant Transformers (SiT) [72], and Masked Auto-Regressive models with diffusion loss (MAR) [60], to name a few. Denosing-based generative modeling not only presents impressive generation results on a wide range of modality, including natural language [32, 122], images [16, 25, 69, 83, 88], videos [6, 8, 42], and audios [24, 81], but also shows potential to unify the understanding and generation capabilities across modalities in multi-modal language models [107, 109, 110, 123]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 213, + 556, + 393 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 213, + 556, + 393 + ], + "spans": [ + { + "bbox": [ + 313, + 213, + 556, + 393 + ], + "type": "text", + "content": "A core component of denoising-based generative models is the tokenizer [23, 38, 51, 56, 84, 104, 112, 117], which compresses the raw data of each modality into a set of latent tokens in either a discrete or continuous latent space. The compact latent space therefore allows for more efficient and better generative modeling [88]. Among previous efforts, Kullback-Leibler Variational Auto-Encoders (KL-VAE) [51] and Vector Quantized Variational Auto-Encoders (VQ-VAE) [23, 84, 104] stand out as representatives of tokenizers which introduce continuous and discrete latent spaces, respectively. The former constrains the latent space with a Gaussian distribution using re-parametrization [51] trick, and the latter makes its latent space a categorical discrete distribution with a codebook of finite vocabulary which requires straight-through estimation [7]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 403, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 403, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 403, + 556, + 715 + ], + "type": "text", + "content": "Although both KL-VAE and VQ-VAE (and their variants) have been predominantly adopted in many generative models [10, 80, 88, 98, 123], they still present two major challenges restricting the efficiency and effectiveness of generative modeling: (1) the difficulty of achieving a higher compression ratio [13] and (2) the worse discriminative representations in their latent space than other self-supervised methods [14, 35, 109, 116]. The efficiency of downstream generative models, particularly Transformer-based architectures [105], is fundamentally constrained by their quadratic complexity to the number of latent tokens. Current image tokenizers [23, 88, 96] typically compress " + }, + { + "bbox": [ + 313, + 403, + 556, + 715 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 313, + 403, + 556, + 715 + ], + "type": "text", + "content": " images to at least 256 tokens and " + }, + { + "bbox": [ + 313, + 403, + 556, + 715 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 313, + 403, + 556, + 715 + ], + "type": "text", + "content": " images to at least 1024 tokens, creating a significant computational bottleneck for both training and inference of generative models [60, 72, 80]. Many efforts have been made to reduce the number of latent tokens on the generative model side, such as merging [62, 72, 80], pooling [58, 61, 94], and others [12, 63, 70]. More studies have recently emerged to fundamentally reduce the token number of the tokenizer [11, 30, 62, 115]. For example, TiTok [115] uses 128 tokens, achieving generation results comparable to 256 tokens by adding one extra decoder. DC-AE [11] compresses the initial 1024 tokens of " + }, + { + "bbox": [ + 313, + 403, + 556, + 715 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 313, + 403, + 556, + 715 + ], + "type": "text", + "content": " with 256 tokens. However, further increasing the compression ratio results in a significant degradation in the quality of reconstruction and therefore the generation [11]." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 693, + 295, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 693, + 295, + 703 + ], + "spans": [ + { + "bbox": [ + 66, + 693, + 295, + 703 + ], + "type": "text", + "content": "*Work done during internship at AMD. Email: haoc3@andrew.cmu.edu." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 67, + 703, + 276, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 276, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 276, + 712 + ], + "type": "text", + "content": "1https://github.com/Hhhhhhao/continuous_tokenizer." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "28358" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 67, + 70, + 541, + 308 + ], + "blocks": [ + { + "bbox": [ + 67, + 70, + 541, + 308 + ], + "lines": [ + { + "bbox": [ + 67, + 70, + 541, + 308 + ], + "spans": [ + { + "bbox": [ + 67, + 70, + 541, + 308 + ], + "type": "image", + "image_path": "21e994b6b9a87a8fc9dafce3a267dd48eb64d76654861724910fd1a07b80c5fa.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 59, + 308, + 550, + 320 + ], + "lines": [ + { + "bbox": [ + 59, + 308, + 550, + 320 + ], + "spans": [ + { + "bbox": [ + 59, + 308, + 550, + 320 + ], + "type": "text", + "content": "Figure 1. ImageNet-1K " + }, + { + "bbox": [ + 59, + 308, + 550, + 320 + ], + "type": "inline_equation", + "content": "256\\times 256" + }, + { + "bbox": [ + 59, + 308, + 550, + 320 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 59, + 308, + 550, + 320 + ], + "type": "inline_equation", + "content": "512\\times 512" + }, + { + "bbox": [ + 59, + 308, + 550, + 320 + ], + "type": "text", + "content": " generation results of generative models trained on SoftVQ-VAE with 32 and 64 tokens." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 327, + 296, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 327, + 296, + 495 + ], + "spans": [ + { + "bbox": [ + 55, + 327, + 296, + 495 + ], + "type": "text", + "content": "Beyond the compression challenge, the quality of the learned representations of the tokenizer poses another critical limitation. While latent representations are crucial for both understanding [31, 109] and generation tasks [30, 62, 64, 116], current tokenizers usually struggle to capture more discriminative features. Earlier methods attempted to advance the training objectives and recipe to improve the latent space of the tokenizer [14, 19, 23, 34, 124]. Recently, [30, 62, 109, 113, 116, 118, 124] has also explored the representation alignment of latent tokens. However, due to the lossy nature [77, 91] of the smoothness constraints in KL-VAE and the discrete quantization in VQVAE [13, 27, 38, 126], current tokenizers usually fall short in learning the semantics of the latent space [109]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 498, + 298, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 298, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 298, + 714 + ], + "type": "text", + "content": "To address the aforementioned challenges, we present SoftVQ-VAE, a simple modification to VQ-VAE that converts it from a discrete tokenizer into a continuous one with a high compression ratio and a semantic-rich latent space. Specifically, we propose using a soft categorical posterior in VAE with a learnable codebook. This straightforward modification introduces a crucial capability: instead of the conventional one-to-one mapping between tokens and codewords in VQ-VAE, our approach enables the adaptive aggregation of multiple codewords into each latent token, substantially increasing the representation capacity of the latent space. We apply SoftVQ-VAE to the Transformer auto-encoder architecture [30, 58, 62, 111, 115], and we succeed in using much fewer 1-dimensional latent tokens (i.e., 32 and 64) for both the reconstruction and subsequent generation tasks. Since SoftVQ-VAE is now fully differentiable, it not only simplifies the learning of the codebook, not requiring the codebook loss or commit loss as in VQ-VAE [23, 104], but" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 327, + 556, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 327, + 556, + 434 + ], + "spans": [ + { + "bbox": [ + 313, + 327, + 556, + 434 + ], + "type": "text", + "content": "also enables better representation learning through direct alignment with pre-trained semantic features using a simple cosine similarity objective. Our approach generalizes the K-Means [67] latent space of VQ-VAE to Soft K-Means [22] and further to Gaussian Mixture Models (GMM) [86], while maintaining compatibility with existing VQ-VAE techniques [45, 55, 62, 112]. We comprehensively demonstrate the efficiency and effectiveness of the proposed SoftVQ-VAE in image reconstruction and generation tasks, and show that:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 315, + 438, + 556, + 629 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 315, + 438, + 555, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 438, + 555, + 474 + ], + "spans": [ + { + "bbox": [ + 315, + 438, + 555, + 474 + ], + "type": "text", + "content": "- SoftVQ-VAE presents more consistent, robust, and high-quality reconstruction and generation results with various latent token lengths ranging from 256 to 32." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 315, + 475, + 555, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 475, + 555, + 533 + ], + "spans": [ + { + "bbox": [ + 315, + 475, + 555, + 533 + ], + "type": "text", + "content": "- SoftVQ-VAE allows diffusion-based DiT [73], flow-based SiT [72], and autoregressive-based MAR [60] to achieve state-of-the-art generation results with only 64 and 32 tokens on both the " + }, + { + "bbox": [ + 315, + 475, + 555, + 533 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 315, + 475, + 555, + 533 + ], + "type": "text", + "content": ", with an FID of 1.78, and " + }, + { + "bbox": [ + 315, + 475, + 555, + 533 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 315, + 475, + 555, + 533 + ], + "type": "text", + "content": " on ImageNet benchmark, with an FID of 2.21." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 315, + 534, + 556, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 534, + 556, + 582 + ], + "spans": [ + { + "bbox": [ + 315, + 534, + 556, + 582 + ], + "type": "text", + "content": "- SoftVQ-VAE improves the inference throughput by 18x and 10x with 32 and 64 tokens to generate " + }, + { + "bbox": [ + 315, + 534, + 556, + 582 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 315, + 534, + 556, + 582 + ], + "type": "text", + "content": " images, and 55x with 64 tokens for generating " + }, + { + "bbox": [ + 315, + 534, + 556, + 582 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 315, + 534, + 556, + 582 + ], + "type": "text", + "content": " images, with significantly lower GFLOPs for all models." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 315, + 582, + 556, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 582, + 556, + 629 + ], + "spans": [ + { + "bbox": [ + 315, + 582, + 556, + 629 + ], + "type": "text", + "content": "- SoftVQ-VAE presents a more discriminative latent space through representation alignment. It also facilitates training efficiency by significantly improving training throughput and reduces training iterations by 2.3x." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 653, + 391, + 667 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 391, + 667 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 391, + 667 + ], + "type": "text", + "content": "2. Preliminary" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "text", + "content": "We present an overview of the tokenizers, i.e., KL-VAE [51] and VQ-VAE [104], and denoising-based generative models, i.e., DiT [80], SiT [72], and MAR [60], in this section." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "28359" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 157, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 157, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 157, + 85 + ], + "type": "text", + "content": "2.1. Image Tokenizer" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 89, + 297, + 244 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 89, + 297, + 244 + ], + "spans": [ + { + "bbox": [ + 55, + 89, + 297, + 244 + ], + "type": "text", + "content": "The architecture of image tokenizer generally resembles auto-encoders, consisting of two main components: an encoder " + }, + { + "bbox": [ + 55, + 89, + 297, + 244 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 55, + 89, + 297, + 244 + ], + "type": "text", + "content": " and a decoder " + }, + { + "bbox": [ + 55, + 89, + 297, + 244 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 55, + 89, + 297, + 244 + ], + "type": "text", + "content": ", parameterized by " + }, + { + "bbox": [ + 55, + 89, + 297, + 244 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 55, + 89, + 297, + 244 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 89, + 297, + 244 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 55, + 89, + 297, + 244 + ], + "type": "text", + "content": ", respectively. Given an input image " + }, + { + "bbox": [ + 55, + 89, + 297, + 244 + ], + "type": "inline_equation", + "content": "\\mathbf{x} \\in \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 55, + 89, + 297, + 244 + ], + "type": "text", + "content": ", the encoder " + }, + { + "bbox": [ + 55, + 89, + 297, + 244 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 55, + 89, + 297, + 244 + ], + "type": "text", + "content": " maps the high-dimensional " + }, + { + "bbox": [ + 55, + 89, + 297, + 244 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 55, + 89, + 297, + 244 + ], + "type": "text", + "content": " to a lower-dimensional latent representation " + }, + { + "bbox": [ + 55, + 89, + 297, + 244 + ], + "type": "inline_equation", + "content": "\\mathbf{z} = \\mathcal{E}(\\mathbf{x}; \\phi)" + }, + { + "bbox": [ + 55, + 89, + 297, + 244 + ], + "type": "text", + "content": ". The latent representations can have varying shapes depending on the encoder architecture. We generally view the latent representation as a set of latent tokens " + }, + { + "bbox": [ + 55, + 89, + 297, + 244 + ], + "type": "inline_equation", + "content": "\\mathbf{z} = [\\mathbf{z}^{[0]}, \\mathbf{z}^{[1]}, \\dots, \\mathbf{z}^{[L]}] \\in \\mathbb{R}^{L \\times D}" + }, + { + "bbox": [ + 55, + 89, + 297, + 244 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 89, + 297, + 244 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 55, + 89, + 297, + 244 + ], + "type": "text", + "content": " is the number of tokens and " + }, + { + "bbox": [ + 55, + 89, + 297, + 244 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 55, + 89, + 297, + 244 + ], + "type": "text", + "content": " is the dimension of the latent representation. For basic auto-encoders (AE), the decoder takes latent tokens as input and reconstructs the original signal " + }, + { + "bbox": [ + 55, + 89, + 297, + 244 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}} = \\mathcal{D}(\\mathbf{z}; \\theta)" + }, + { + "bbox": [ + 55, + 89, + 297, + 244 + ], + "type": "text", + "content": " by minimizing the reconstruction loss with " + }, + { + "bbox": [ + 55, + 89, + 297, + 244 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 55, + 89, + 297, + 244 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 245, + 296, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 245, + 296, + 388 + ], + "spans": [ + { + "bbox": [ + 55, + 245, + 296, + 388 + ], + "type": "text", + "content": "Variational auto-encoders (VAE) [51] extend the basic AE by introducing a probabilistic perspective on the latent space. Specifically, VAE approximates a posterior distribution " + }, + { + "bbox": [ + 55, + 245, + 296, + 388 + ], + "type": "inline_equation", + "content": "p(\\mathbf{z}|\\mathbf{x})" + }, + { + "bbox": [ + 55, + 245, + 296, + 388 + ], + "type": "text", + "content": " with a learned distribution " + }, + { + "bbox": [ + 55, + 245, + 296, + 388 + ], + "type": "inline_equation", + "content": "q_{\\phi}(\\mathbf{z}|\\mathbf{x})" + }, + { + "bbox": [ + 55, + 245, + 296, + 388 + ], + "type": "text", + "content": " from the encoder, assuming that " + }, + { + "bbox": [ + 55, + 245, + 296, + 388 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 55, + 245, + 296, + 388 + ], + "type": "text", + "content": " is generated by the unobserved latent " + }, + { + "bbox": [ + 55, + 245, + 296, + 388 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 55, + 245, + 296, + 388 + ], + "type": "text", + "content": " with a prior distribution " + }, + { + "bbox": [ + 55, + 245, + 296, + 388 + ], + "type": "inline_equation", + "content": "p(\\mathbf{z})" + }, + { + "bbox": [ + 55, + 245, + 296, + 388 + ], + "type": "text", + "content": ". The decoder instead takes a set of sampled latent tokens " + }, + { + "bbox": [ + 55, + 245, + 296, + 388 + ], + "type": "inline_equation", + "content": "\\mathbf{z} \\sim q_{\\phi}(\\mathbf{z}|\\mathbf{x})" + }, + { + "bbox": [ + 55, + 245, + 296, + 388 + ], + "type": "text", + "content": " from the posterior and aims to learn the marginal likelihood of data in a generative process by optimizing the evidence lower bound (ELBO): " + }, + { + "bbox": [ + 55, + 245, + 296, + 388 + ], + "type": "inline_equation", + "content": "\\max_{\\phi,\\theta}\\mathbb{E}_{q_{\\phi}(\\mathbf{z}|\\mathbf{x})}\\log p_{\\theta}(\\mathbf{x}|\\mathbf{z}) - \\mathcal{L}_{\\mathrm{kl}}(q_{\\phi}(\\mathbf{z}|\\mathbf{x})|p(\\mathbf{z}))" + }, + { + "bbox": [ + 55, + 245, + 296, + 388 + ], + "type": "text", + "content": ". To make the optimization tractable, assumptions have been made on the prior, leading to different variants of VAE." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 388, + 296, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 388, + 296, + 460 + ], + "spans": [ + { + "bbox": [ + 55, + 388, + 296, + 460 + ], + "type": "text", + "content": "KL-VAE [38, 51] parametrizes both the prior and posterior as Gaussians. The prior " + }, + { + "bbox": [ + 55, + 388, + 296, + 460 + ], + "type": "inline_equation", + "content": "p(\\mathbf{z})" + }, + { + "bbox": [ + 55, + 388, + 296, + 460 + ], + "type": "text", + "content": " is assumed to be the isotropic unit Gaussian " + }, + { + "bbox": [ + 55, + 388, + 296, + 460 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(0,1)" + }, + { + "bbox": [ + 55, + 388, + 296, + 460 + ], + "type": "text", + "content": ". The continuous latent code is parameterized with posterior mean " + }, + { + "bbox": [ + 55, + 388, + 296, + 460 + ], + "type": "inline_equation", + "content": "\\mu_{\\phi}" + }, + { + "bbox": [ + 55, + 388, + 296, + 460 + ], + "type": "text", + "content": " and variance " + }, + { + "bbox": [ + 55, + 388, + 296, + 460 + ], + "type": "inline_equation", + "content": "\\sigma_{\\phi}^{2}" + }, + { + "bbox": [ + 55, + 388, + 296, + 460 + ], + "type": "text", + "content": " predicted by the encoder using the \"re-parametrization\" trick with a noise variable " + }, + { + "bbox": [ + 55, + 388, + 296, + 460 + ], + "type": "inline_equation", + "content": "\\varepsilon" + }, + { + "bbox": [ + 55, + 388, + 296, + 460 + ], + "type": "text", + "content": " from standard Gaussian:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 79, + 467, + 233, + 479 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 467, + 233, + 479 + ], + "spans": [ + { + "bbox": [ + 79, + 467, + 233, + 479 + ], + "type": "interline_equation", + "content": "\\text {p o s t e r i o r :} q _ {\\phi} (\\mathbf {z} | \\mathbf {x}) = \\mathcal {N} (\\mathbf {z}; \\boldsymbol {\\mu} _ {\\phi} (\\mathbf {x}), \\boldsymbol {\\sigma} _ {\\phi} ^ {2} (\\mathbf {x}))", + "image_path": "fd8f211e242dc912ced3e5d31eeaad5423663a396edf0c383951a2be28614fdf.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 79, + 481, + 295, + 496 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 481, + 295, + 496 + ], + "spans": [ + { + "bbox": [ + 79, + 481, + 295, + 496 + ], + "type": "interline_equation", + "content": "l a t e n t: \\mathbf {z} = \\boldsymbol {\\mu} _ {\\phi} (\\mathbf {x}) + \\boldsymbol {\\sigma} _ {\\phi} (\\mathbf {x}) \\odot \\varepsilon , \\quad \\epsilon \\sim \\mathcal {N} (0, \\mathbf {I}) \\tag {1}", + "image_path": "601d86cfd5e9d222d6c67926b4bc8fea00f2b95cd6ec0f7fad8b7736a5125dd9.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 79, + 496, + 271, + 515 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 496, + 271, + 515 + ], + "spans": [ + { + "bbox": [ + 79, + 496, + 271, + 515 + ], + "type": "interline_equation", + "content": "k l: \\mathcal {L} _ {\\mathrm {k l}} = - \\frac {1}{2} \\left(1 + \\log \\left(\\boldsymbol {\\sigma} _ {\\phi} ^ {2} (\\mathbf {x})\\right) - \\boldsymbol {\\mu} _ {\\phi} (\\mathbf {x}) ^ {2} - \\boldsymbol {\\sigma} _ {\\phi} ^ {2} (\\mathbf {x})\\right)", + "image_path": "0c715b626fe1e5996501e4f01fcf1302950776d7786323949f34fcce07f38220.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 523, + 296, + 595 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 523, + 296, + 595 + ], + "spans": [ + { + "bbox": [ + 55, + 523, + 296, + 595 + ], + "type": "text", + "content": "While KL-VAE is widely used in diffusion models, it is known to have the \"posterior-collapse\" issue [13]. In addition, the KL loss weight usually imposes a trade-off between the quality of reconstruction and the smoothness of its latent space [100], preventing high compression ratio and the learning of semantics in the latent space of KL-VAE." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 596, + 296, + 644 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 596, + 296, + 644 + ], + "spans": [ + { + "bbox": [ + 55, + 596, + 296, + 644 + ], + "type": "text", + "content": "VQ-VAE [23, 104], in contrast, generates latent code in a discrete space by taking the posterior as a \\(K\\)-way deterministic categorical distribution against a learnable codebook \\(\\mathcal{C} = [\\mathbf{c}^{[0]}, \\dots, \\mathbf{c}^{[K]}\\] \\(\\in \\mathbb{R}^{K \\times D}\\), with \\(\\mathbf{c}\\) as the codeword:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 61, + 652, + 277, + 681 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 652, + 277, + 681 + ], + "spans": [ + { + "bbox": [ + 61, + 652, + 277, + 681 + ], + "type": "interline_equation", + "content": "\\text {p o s t e r i o r :} q _ {\\phi} (\\mathbf {z} = k | \\mathbf {x}) = \\left\\{ \\begin{array}{l l} 1 & \\text {i f} k = \\arg \\min _ {j} \\| \\hat {\\mathbf {z}} - \\mathbf {c} ^ {[ j ]} \\| _ {2} \\\\ 0 & \\text {o t h e r w i s e} \\end{array} \\right.", + "image_path": "63961cb8560d9f28d6c436aaab925294e3ec1a053a9a89943d1481721b036e46.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 679, + 295, + 700 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 679, + 295, + 700 + ], + "spans": [ + { + "bbox": [ + 62, + 679, + 295, + 700 + ], + "type": "interline_equation", + "content": "l a t e n t: \\mathbf {z} = \\mathbf {c} ^ {[ k ]}, \\quad \\text {w h e r e} k = \\underset {j} {\\arg \\min } \\| \\hat {\\mathbf {z}} - \\mathbf {c} ^ {[ j ]} \\| _ {2} \\tag {2}", + "image_path": "ef74abb4bd76f9e0a0f948bfc4aebce838268b60ada89d6fe10a7f97d19f4616.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 704, + 121, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 704, + 121, + 715 + ], + "spans": [ + { + "bbox": [ + 62, + 704, + 121, + 715 + ], + "type": "interline_equation", + "content": "k l: \\mathcal {L} _ {\\mathrm {k l}} = \\log K", + "image_path": "a9778271dfd89411c4db9aaa79393fe447bf0d843691b6a0d97ef87121312358.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 72, + 556, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 556, + 193 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 556, + 193 + ], + "type": "text", + "content": "Due to the non-differentiable arg min, a \"straight-through\" trick [7] is adopted to approximate the gradient of the encoder output " + }, + { + "bbox": [ + 313, + 72, + 556, + 193 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{z}}" + }, + { + "bbox": [ + 313, + 72, + 556, + 193 + ], + "type": "text", + "content": " by directly copying gradient from the decoder input " + }, + { + "bbox": [ + 313, + 72, + 556, + 193 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 313, + 72, + 556, + 193 + ], + "type": "text", + "content": ". Since the reconstruction objective does not impose direct gradients on the codebook " + }, + { + "bbox": [ + 313, + 72, + 556, + 193 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 313, + 72, + 556, + 193 + ], + "type": "text", + "content": ", VQ-VAE additionally uses a codebook loss as " + }, + { + "bbox": [ + 313, + 72, + 556, + 193 + ], + "type": "inline_equation", + "content": "\\| \\mathrm{sg}[\\hat{\\mathbf{z}}] - \\mathbf{c}\\| _2" + }, + { + "bbox": [ + 313, + 72, + 556, + 193 + ], + "type": "text", + "content": " to move codewords toward the encoder output, along with a commit loss " + }, + { + "bbox": [ + 313, + 72, + 556, + 193 + ], + "type": "inline_equation", + "content": "\\| \\hat{\\mathbf{z}} -\\mathrm{sg}[\\mathbf{c}]\\| _2" + }, + { + "bbox": [ + 313, + 72, + 556, + 193 + ], + "type": "text", + "content": " to prevent arbitrary growth of codewords, where " + }, + { + "bbox": [ + 313, + 72, + 556, + 193 + ], + "type": "inline_equation", + "content": "\\mathrm{sg}[\\cdot ]" + }, + { + "bbox": [ + 313, + 72, + 556, + 193 + ], + "type": "text", + "content": " denotes stop-gradient. The broken gradient hinders high compression ratio and latent space learning [27, 43]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 200, + 505, + 213 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 200, + 505, + 213 + ], + "spans": [ + { + "bbox": [ + 313, + 200, + 505, + 213 + ], + "type": "text", + "content": "2.2. Denoising-based Generative Models" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 217, + 555, + 254 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 217, + 555, + 254 + ], + "spans": [ + { + "bbox": [ + 313, + 217, + 555, + 254 + ], + "type": "text", + "content": "Denoising-based generative models synthesize images by progressively transforming Gaussian noise " + }, + { + "bbox": [ + 313, + 217, + 555, + 254 + ], + "type": "inline_equation", + "content": "\\varepsilon \\sim \\mathcal{N}(0,\\mathbf{I})" + }, + { + "bbox": [ + 313, + 217, + 555, + 254 + ], + "type": "text", + "content": " into tokenizer latent codes " + }, + { + "bbox": [ + 313, + 217, + 555, + 254 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 313, + 217, + 555, + 254 + ], + "type": "text", + "content": " through a forward process [41, 92]:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 399, + 264, + 553, + 276 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 399, + 264, + 553, + 276 + ], + "spans": [ + { + "bbox": [ + 399, + 264, + 553, + 276 + ], + "type": "interline_equation", + "content": "\\mathbf {z} _ {t} = \\alpha_ {t} \\mathbf {z} + \\sigma_ {t} \\varepsilon , \\tag {3}", + "image_path": "ae353456ef04cb75cdd4de07e2e27e04111e4d4ab7211c7c91e7e1d6adf6e087.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 285, + 555, + 322 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 285, + 555, + 322 + ], + "spans": [ + { + "bbox": [ + 313, + 285, + 555, + 322 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 285, + 555, + 322 + ], + "type": "inline_equation", + "content": "\\alpha_{t}" + }, + { + "bbox": [ + 313, + 285, + 555, + 322 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 285, + 555, + 322 + ], + "type": "inline_equation", + "content": "\\sigma_{t}" + }, + { + "bbox": [ + 313, + 285, + 555, + 322 + ], + "type": "text", + "content": " are a decreasing and increasing function of " + }, + { + "bbox": [ + 313, + 285, + 555, + 322 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 285, + 555, + 322 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 285, + 555, + 322 + ], + "type": "inline_equation", + "content": "t \\in [0, T]" + }, + { + "bbox": [ + 313, + 285, + 555, + 322 + ], + "type": "text", + "content": " [50]. In this paper, we adopt three approaches to implementing this denoising process as follows." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 322, + 556, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 322, + 556, + 417 + ], + "spans": [ + { + "bbox": [ + 313, + 322, + 556, + 417 + ], + "type": "text", + "content": "DiT [80] is a diffusion-based model with a transformer architecture. It formulates the process through a forward-time stochastic differential equation (SDE), where " + }, + { + "bbox": [ + 313, + 322, + 556, + 417 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_t" + }, + { + "bbox": [ + 313, + 322, + 556, + 417 + ], + "type": "text", + "content": " converges to " + }, + { + "bbox": [ + 313, + 322, + 556, + 417 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(0, \\mathbf{I})" + }, + { + "bbox": [ + 313, + 322, + 556, + 417 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 313, + 322, + 556, + 417 + ], + "type": "inline_equation", + "content": "t \\to T" + }, + { + "bbox": [ + 313, + 322, + 556, + 417 + ], + "type": "text", + "content": ". Generation occurs via a reverse-time SDE [2], with the model trained to predict noise " + }, + { + "bbox": [ + 313, + 322, + 556, + 417 + ], + "type": "inline_equation", + "content": "\\varepsilon" + }, + { + "bbox": [ + 313, + 322, + 556, + 417 + ], + "type": "text", + "content": " at randomly sampled timesteps. Due to the long token length of the tokenizer [96] used in DiT, it has several variants to merge tokens at the input of the transformer backbone." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 418, + 556, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 418, + 556, + 489 + ], + "spans": [ + { + "bbox": [ + 313, + 418, + 556, + 489 + ], + "type": "text", + "content": "SiT [72] instead uses stochastic interpolants [1] and performs denoising via probability flow ordinary differential equation (PF ODE). The model learns the ODE's velocity field [65] by minimizing mean-squared-error (MSE) loss. This velocity field defines a score function for generation, analogous to DiT's reverse-time SDE." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 489, + 556, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 489, + 556, + 561 + ], + "spans": [ + { + "bbox": [ + 313, + 489, + 556, + 561 + ], + "type": "text", + "content": "MAR [60] combines diffusion with autoregressive generation. Unlike DiT and SiT's parallel denoising of all latent tokens, MAR adopts an encoder-decoder transformer architecture [35] that progressively denoises tokens in a \"next-set\" autoregressive manner [99], starting from masked tokens, similar to MaskGIT [10] and MAGE [59]." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 571, + 370, + 584 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 571, + 370, + 584 + ], + "spans": [ + { + "bbox": [ + 313, + 571, + 370, + 584 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 592, + 556, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 592, + 556, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 592, + 556, + 665 + ], + "type": "text", + "content": "In this section, we present SoftVQ-VAE, a novel 1D continuous tokenizer with a high compression ratio and a semantic-rich latent space. We first introduce the architecture and the formulation of SoftVQ-VAE. Then, we show that SoftVQ-VAE can learn semantics easily by aligning its latent tokens with pre-trained features via its fully-differentiable property." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 672, + 397, + 683 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 672, + 397, + 683 + ], + "spans": [ + { + "bbox": [ + 313, + 672, + 397, + 683 + ], + "type": "text", + "content": "3.1. Architecture" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 689, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 556, + 713 + ], + "type": "text", + "content": "We leverage Vision Transformer (ViT) [21, 111] as the architecture for the encoder " + }, + { + "bbox": [ + 313, + 689, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 313, + 689, + 556, + 713 + ], + "type": "text", + "content": " and decoder " + }, + { + "bbox": [ + 313, + 689, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 313, + 689, + 556, + 713 + ], + "type": "text", + "content": " of SoftVQ-VAE." + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "28360" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 70, + 292, + 243 + ], + "blocks": [ + { + "bbox": [ + 59, + 70, + 292, + 243 + ], + "lines": [ + { + "bbox": [ + 59, + 70, + 292, + 243 + ], + "spans": [ + { + "bbox": [ + 59, + 70, + 292, + 243 + ], + "type": "image", + "image_path": "b757a7fb4120fe2d1d016fc288e2800319d031dd85905ac89c7cb35f0aeebb90.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 245, + 296, + 289 + ], + "lines": [ + { + "bbox": [ + 55, + 245, + 296, + 289 + ], + "spans": [ + { + "bbox": [ + 55, + 245, + 296, + 289 + ], + "type": "text", + "content": "Figure 2. Illustration of SoftVQ-VAE. Left: Transformer encoder-decoder architecture with image tokens, arbitrary length of latent tokens, and mask tokens. Right top: fully-differentiable SoftVQ illustration. Right bottom: latent space representation alignment." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 294, + 296, + 463 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 294, + 296, + 463 + ], + "spans": [ + { + "bbox": [ + 54, + 294, + 296, + 463 + ], + "type": "text", + "content": "Similarly to Yu et al. [115], instead of using the image features as latent, we initialize a set of extra 1D learnable tokens and use these tokens for reconstruction and subsequent generation. With the self-attention mechanism, it allows the learnable tokens to adaptively aggregate different image tokens to obtain the latent tokens of SoftVQ, detailed in Sec. 3.2. Specifically, for the encoder transformer with a patch size of " + }, + { + "bbox": [ + 54, + 294, + 296, + 463 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 54, + 294, + 296, + 463 + ], + "type": "text", + "content": ", an image " + }, + { + "bbox": [ + 54, + 294, + 296, + 463 + ], + "type": "inline_equation", + "content": "\\mathbf{x} \\in \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 54, + 294, + 296, + 463 + ], + "type": "text", + "content": " is patched into a sequence of image tokens " + }, + { + "bbox": [ + 54, + 294, + 296, + 463 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_p \\in \\mathbb{R}^{N \\times D}" + }, + { + "bbox": [ + 54, + 294, + 296, + 463 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 54, + 294, + 296, + 463 + ], + "type": "inline_equation", + "content": "N = HW / P^2" + }, + { + "bbox": [ + 54, + 294, + 296, + 463 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 294, + 296, + 463 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 54, + 294, + 296, + 463 + ], + "type": "text", + "content": " represent the token dimension size. We then concatenate a set of learnable tokens " + }, + { + "bbox": [ + 54, + 294, + 296, + 463 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_l \\in \\mathbb{R}^{L \\times D}" + }, + { + "bbox": [ + 54, + 294, + 296, + 463 + ], + "type": "text", + "content": " of length " + }, + { + "bbox": [ + 54, + 294, + 296, + 463 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 54, + 294, + 296, + 463 + ], + "type": "text", + "content": " along with the image tokens as the input to the encoder, and retain only the output corresponding to the learnable latent tokens as the output of our encoder: " + }, + { + "bbox": [ + 54, + 294, + 296, + 463 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{z}} = \\mathcal{E}([\\mathbf{x}_p; \\mathbf{z}_l]; \\phi)" + }, + { + "bbox": [ + 54, + 294, + 296, + 463 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 463, + 296, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 463, + 296, + 641 + ], + "spans": [ + { + "bbox": [ + 55, + 463, + 296, + 641 + ], + "type": "text", + "content": "For the decoder, we reconstruct images from a sequence of learnable mask tokens " + }, + { + "bbox": [ + 55, + 463, + 296, + 641 + ], + "type": "inline_equation", + "content": "\\mathbf{m} \\in \\mathbb{R}^{N \\times D}" + }, + { + "bbox": [ + 55, + 463, + 296, + 641 + ], + "type": "text", + "content": " [5, 35], concatenated with latent tokens: " + }, + { + "bbox": [ + 55, + 463, + 296, + 641 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{x}} = \\mathcal{D}([\\mathbf{m};\\mathbf{z}];\\theta)" + }, + { + "bbox": [ + 55, + 463, + 296, + 641 + ], + "type": "text", + "content": ". We use a linear layer at the end of the decoder to regress the pixel values from the mask tokens. For the image tokens at the encoder input and mask tokens at the decoder input, we apply 2D absolute position embeddings and 1D absolute position embedding for the latent tokens. Adopting more advanced position embedding techniques, such as RoPE [36, 97], may lead to better results and is left for future work. This design allows SoftVQ-VAE to use latent codes of arbitrary length, each adaptively associated with image tokens, for reconstruction and subsequent generative modeling, and model images of various resolutions using the same length of latent codes. An overview of the model architecture is shown in Fig. 2." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 648, + 139, + 661 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 648, + 139, + 661 + ], + "spans": [ + { + "bbox": [ + 55, + 648, + 139, + 661 + ], + "type": "text", + "content": "3.2. SoftVQ-VAE" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 666, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 666, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 666, + 297, + 714 + ], + "type": "text", + "content": "As discussed above, the high compression ratio of both KL-VAE and VQ-VAE usually results in a significant degradation of quality in reconstruction and latent space. To overcome these limitations, we propose SoftVQ-VAE, a simple mod" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "type": "text", + "content": "ification to VQ-VAE, which maintains the advantages of learning codewords to capture the data distribution while bridging it with more representation capacity as a continuous tokenizer. The core idea of SoftVQ-VAE is to allow each latent code to adaptively aggregate multiple codewords from the learnable codebook. Consequently, it imposes a soft categorical distribution on the posterior from the encoder:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 342, + 165, + 553, + 204 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 342, + 165, + 553, + 204 + ], + "spans": [ + { + "bbox": [ + 342, + 165, + 553, + 204 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\text {p o s t e r i o r :} q _ {\\phi} (\\mathbf {z} | \\mathbf {x}) = \\operatorname {S o f t m a x} \\left(- \\frac {\\| \\hat {\\mathbf {z}} - \\mathcal {C} \\| _ {2}}{\\tau}\\right) \\\\ \\text {l a t e n t :} \\mathbf {z} = q _ {\\phi} (\\mathbf {z} | \\mathbf {x}) \\mathcal {C} \\end{array} \\tag {4}", + "image_path": "13bc2892c8f7e25d3e0d9e8136b6615800d19b17fc17a9065f1dcac4b695b893.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 343, + 207, + 522, + 222 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 343, + 207, + 522, + 222 + ], + "spans": [ + { + "bbox": [ + 343, + 207, + 522, + 222 + ], + "type": "interline_equation", + "content": "k l \\colon \\mathcal {L} _ {\\mathrm {k l}} = H (q _ {\\phi} (\\mathbf {z} | \\mathbf {x})) - H \\left(\\mathbb {E} _ {\\mathbf {x} \\sim p (\\mathbf {x})} q _ {\\phi} (\\mathbf {z} | \\mathbf {x})\\right)", + "image_path": "ffb1800de19d0848b838eaaba7fb6cc2b527299a096b31ae7bcfb05be9cdb83d.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 231, + 555, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 231, + 555, + 266 + ], + "spans": [ + { + "bbox": [ + 313, + 231, + 555, + 266 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 231, + 555, + 266 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 313, + 231, + 555, + 266 + ], + "type": "text", + "content": " is the temperature parameter controlling the sharpness of the softmax probability and is set to 0.07 by ablation. The derivation of " + }, + { + "bbox": [ + 313, + 231, + 555, + 266 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{kl}" + }, + { + "bbox": [ + 313, + 231, + 555, + 266 + ], + "type": "text", + "content": " is shown in Appendix B.1." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 267, + 556, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 267, + 556, + 386 + ], + "spans": [ + { + "bbox": [ + 313, + 267, + 556, + 386 + ], + "type": "text", + "content": "Despite this simple modification, SoftVQ-VAE can significantly reduce the latent token length while maintaining a high quality of the reconstruction and latent space. More importantly, with a soft categorical distribution, SoftVQ-VAE is fully-differentiable, and thus the encoder and codebook can be optimized directly from the reconstruction loss (and other losses). It also allows for easier regularization of various forms on the latent space that drastically improves its quality, and less hyper-parameter tuning without codebook and commit loss [43] in the reduced training objectives." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 387, + 556, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 387, + 556, + 602 + ], + "spans": [ + { + "bbox": [ + 313, + 387, + 556, + 602 + ], + "type": "text", + "content": "Interpretation. The posterior and latent space of VQ-VAE can be interpreted with K-Means [67]. Hence, SoftVQ-VAE can be viewed as soft K-Means [22], with codewords as learnable prototypes in the latent space. This latent space can easily be extended to Gaussian Mixture Models (GMM) [86], which we refer to as GMMVQ-VAE. Instead of using fixed temperature, GMMVQ-VAE predicts data-dependent weights for the codeword prototypes, i.e., Gaussian means, to compute the soft categorical posterior: " + }, + { + "bbox": [ + 313, + 387, + 556, + 602 + ], + "type": "inline_equation", + "content": "q_{\\phi}(\\mathbf{z}|\\mathbf{x}) = \\mathrm{Softmax}(-\\omega (\\hat{\\mathbf{z}})\\| \\hat{\\mathbf{z}} -\\mathcal{C}\\| _2)" + }, + { + "bbox": [ + 313, + 387, + 556, + 602 + ], + "type": "text", + "content": ". Although prior work has explored GMM-based VAEs [17, 47], our GMMVQ-VAE differs by learning discrete prototypes rather than continuous latent variables. While providing the additional benefit of interpretable Gaussian components in the latent space, GMMVQ achieves reconstruction quality and downstream generation performance similar to SoftVQ in our experiments, and thus we mainly adopt the simpler SoftVQ throughout this paper. More details are in Appendix B.2." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 612, + 536, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 612, + 536, + 624 + ], + "spans": [ + { + "bbox": [ + 313, + 612, + 536, + 624 + ], + "type": "text", + "content": "3.3. Representation Alignment of Latent Space" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 629, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 555, + 713 + ], + "type": "text", + "content": "While the reconstruction quality of tokenizer is important, learning a high-quality latent space is more crucial for downstream denoising-based generative modeling [116]. Several recent approaches have explored aligning the latent code with pre-trained features through a contrastive loss [30, 62, 109] and initializing the codebook with pre-trained features [124]. However, imposing effective regularization" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28361" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 95 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 95 + ], + "type": "text", + "content": "on the latent remains a challenge due to the discrete quantization of VQ-VAE and the Gaussian constraints in KL-VAE." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 96, + 297, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 96, + 297, + 240 + ], + "spans": [ + { + "bbox": [ + 55, + 96, + 297, + 240 + ], + "type": "text", + "content": "Thanks to the fully-differentiable property of SoftVQ-VAE, we can now directly impose regularization on the latent space. Inspired by the recent work of REPA [116], we propose aligning the representations of latent codes with pre-trained vision encoders to facilitate learning the latent space. Compared to REPA, which aligns the representation at intermediate layers of downstream generative models, our method of feature alignment in the tokenizer's latent space is equivalent to performing REPA at the input space of generative models. Specifically, to align the latent code with the image tokens from a pre-trained vision encoder, we replicate each latent token by " + }, + { + "bbox": [ + 55, + 96, + 297, + 240 + ], + "type": "inline_equation", + "content": "N / L" + }, + { + "bbox": [ + 55, + 96, + 297, + 240 + ], + "type": "text", + "content": " times:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 248, + 296, + 280 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 248, + 296, + 280 + ], + "spans": [ + { + "bbox": [ + 61, + 248, + 296, + 280 + ], + "type": "interline_equation", + "content": "\\mathbf {z} _ {r} = \\underbrace {\\left[ \\mathbf {z} ^ {[ 0 ]} , \\dots , \\mathbf {z} ^ {[ 0 ]} \\right.} _ {N / L \\text {t i m e s}}, \\underbrace {\\left. \\mathbf {z} ^ {[ 1 ]} , \\dots , \\mathbf {z} ^ {[ 1 ]} \\right.} _ {N / L \\text {t i m e s}}, \\dots , \\underbrace {\\left. \\mathbf {z} ^ {[ L ]} , \\dots , \\mathbf {z} ^ {[ L ]} \\right]} _ {N / L \\text {t i m e s}}, \\tag {5}", + "image_path": "7435d8a50ffc3346447ed2c7f524225e8ac59c996afe83bc8c8abc657f0eda6b.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 289, + 296, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 289, + 296, + 326 + ], + "spans": [ + { + "bbox": [ + 55, + 289, + 296, + 326 + ], + "type": "text", + "content": "and encourage the token-wise similarity with image tokens. We employ a projector multilayer perceptron (MLP) on top of the latent tokens to match the pre-trained feature " + }, + { + "bbox": [ + 55, + 289, + 296, + 326 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_{*}" + }, + { + "bbox": [ + 55, + 289, + 296, + 326 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 86, + 335, + 296, + 369 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 335, + 296, + 369 + ], + "spans": [ + { + "bbox": [ + 86, + 335, + 296, + 369 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {a l i g n}} = \\frac {1}{N} \\sum_ {n = 1} ^ {N} \\operatorname {s i m} \\left(\\mathbf {y} _ {*} ^ {[ n ]}, \\mathbf {M L P} \\left(\\mathbf {z} _ {r} ^ {([ n ])}\\right)\\right). \\tag {6}", + "image_path": "65efae21bf1f73f89415510a7810b3c07e7176454012dd26356e91513913f052.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 378, + 296, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 378, + 296, + 475 + ], + "spans": [ + { + "bbox": [ + 55, + 378, + 296, + 475 + ], + "type": "text", + "content": "This alignment ensures the latent space of SoftVQ-VAE captures semantically discriminative features that are beneficial for training the downstream generative models, even if it does not directly translate into improved reconstruction performance from the tokenizer. However, as we will show in Sec. 4.4, the generation quality of trained generative models often depends more on the semantic structure of the latent space than on the tokenizer's reconstruction capabilities." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 483, + 194, + 496 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 483, + 194, + 496 + ], + "spans": [ + { + "bbox": [ + 55, + 483, + 194, + 496 + ], + "type": "text", + "content": "3.4. Final Training Objective" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 502, + 297, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 502, + 297, + 538 + ], + "spans": [ + { + "bbox": [ + 55, + 502, + 297, + 538 + ], + "type": "text", + "content": "The training objective of SoftVQ-VAE combines the reconstruction loss, perceptual loss [20, 46, 54, 120] and adversarial loss [33, 44] as in VQ [23], and representation alignment:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 61, + 548, + 295, + 562 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 548, + 295, + 562 + ], + "spans": [ + { + "bbox": [ + 61, + 548, + 295, + 562 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {\\text {r e c o n}} + \\lambda_ {1} \\mathcal {L} _ {\\text {p e r c e p}} + \\lambda_ {2} \\mathcal {L} _ {\\text {a d v}} + \\lambda_ {3} \\mathcal {L} _ {\\text {a l i g n}} + \\lambda_ {4} \\mathcal {L} _ {\\mathrm {K L}}, \\tag {7}", + "image_path": "d0daa8a82fa4dae58becb89a0b9ecb258ee688a5633aa87b145bb3bc6a4f2d16.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 572, + 297, + 656 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 572, + 297, + 656 + ], + "spans": [ + { + "bbox": [ + 55, + 572, + 297, + 656 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 572, + 297, + 656 + ], + "type": "inline_equation", + "content": "\\lambda_1, \\lambda_2, \\lambda_3," + }, + { + "bbox": [ + 55, + 572, + 297, + 656 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 572, + 297, + 656 + ], + "type": "inline_equation", + "content": "\\lambda_4" + }, + { + "bbox": [ + 55, + 572, + 297, + 656 + ], + "type": "text", + "content": " are hyper-parameters. Neither codebook nor commit loss as in VQ is needed. The perceptual loss " + }, + { + "bbox": [ + 55, + 572, + 297, + 656 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{percep}}" + }, + { + "bbox": [ + 55, + 572, + 297, + 656 + ], + "type": "text", + "content": " helps capture high-level perceptual features, " + }, + { + "bbox": [ + 55, + 572, + 297, + 656 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{adv}}" + }, + { + "bbox": [ + 55, + 572, + 297, + 656 + ], + "type": "text", + "content": " encourages the decoder to generate realistic images by removing the artifacts from the reconstruction loss only, " + }, + { + "bbox": [ + 55, + 572, + 297, + 656 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{align}}" + }, + { + "bbox": [ + 55, + 572, + 297, + 656 + ], + "type": "text", + "content": " guides the latent space to align with pre-trained features, and " + }, + { + "bbox": [ + 55, + 572, + 297, + 656 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{kl}}" + }, + { + "bbox": [ + 55, + 572, + 297, + 656 + ], + "type": "text", + "content": " as the KL divergence term in the ELBO." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 669, + 137, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 669, + 137, + 681 + ], + "spans": [ + { + "bbox": [ + 55, + 669, + 137, + 681 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": "In this section, we validate the efficiency and effectiveness of SoftVQ-VAE with extensive experiments." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 72, + 426, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 426, + 85 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 426, + 85 + ], + "type": "text", + "content": "4.1. Experiments Setup" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 89, + 555, + 328 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 89, + 555, + 328 + ], + "spans": [ + { + "bbox": [ + 313, + 89, + 555, + 328 + ], + "type": "text", + "content": "Implementation Details of SoftVQ-VAE. We use the LlamaGen codebase [98] to build our SoftVQ-VAE. We instantiate 4 configurations: SoftVQ-S, SoftVQ-B, SoftVQ-BL, SoftVQ-L, with a total of 45M, 173M, 391M, and 608M parameters, respectively. Each configuration has variants of latent codes " + }, + { + "bbox": [ + 313, + 89, + 555, + 328 + ], + "type": "inline_equation", + "content": "L = 64" + }, + { + "bbox": [ + 313, + 89, + 555, + 328 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 89, + 555, + 328 + ], + "type": "inline_equation", + "content": "L = 32" + }, + { + "bbox": [ + 313, + 89, + 555, + 328 + ], + "type": "text", + "content": ". For the representation alignment, we choose DINOv2 [79] for the pre-trained features, similarly as Yu et al. [116] and Li et al. [62]. To learn a better semantic latent space, we initialize our encoder from the pre-trained DINOv2 weights. Alignment with other pretrained features is explored in Sec. 4.4. We train the tokenizers on ImageNet [15] of resolution " + }, + { + "bbox": [ + 313, + 89, + 555, + 328 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 313, + 89, + 555, + 328 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 89, + 555, + 328 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 313, + 89, + 555, + 328 + ], + "type": "text", + "content": " for 250K iterations. Similarly to Tian et al. [99] and Li et al. [62], we adopted the same frozen DINO-S [9, 79] discriminator with a similar architecture to StyleGAN [48, 49]. We use DiffAug [121], consistency regularization [119], and LeCAM regularization [101] for discriminator training as in [99]. For the training objective, we set " + }, + { + "bbox": [ + 313, + 89, + 555, + 328 + ], + "type": "inline_equation", + "content": "\\lambda_1 = 1.0" + }, + { + "bbox": [ + 313, + 89, + 555, + 328 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 89, + 555, + 328 + ], + "type": "inline_equation", + "content": "\\lambda_1 = 0.2" + }, + { + "bbox": [ + 313, + 89, + 555, + 328 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 89, + 555, + 328 + ], + "type": "inline_equation", + "content": "\\lambda_3 = 0.1" + }, + { + "bbox": [ + 313, + 89, + 555, + 328 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 89, + 555, + 328 + ], + "type": "inline_equation", + "content": "\\lambda_4 = 0.01" + }, + { + "bbox": [ + 313, + 89, + 555, + 328 + ], + "type": "text", + "content": ", following previous common practice. More training details are shown in Appendix C.1." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 329, + 556, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 329, + 556, + 483 + ], + "spans": [ + { + "bbox": [ + 313, + 329, + 556, + 483 + ], + "type": "text", + "content": "Implementation Details of Generative Modeling. We select DiT [80], SiT [57], and MAR [60] for downstream denoising-based image generation tasks. For DiT and SiT, we set the patch size of DiT and SiT to 1 and use a 1D absolution position embedding. In our main experiments, we train DiT-XL and SiT-XL of 675M parameters for 3M steps, compared to 4M steps in REPA [116] and 7M steps vanilla version [72, 80]. We strictly follow their original training setup for other settings. For MAR-H, we train for 500 epochs and set the maximum learning rate to " + }, + { + "bbox": [ + 313, + 329, + 556, + 483 + ], + "type": "inline_equation", + "content": "2e-4" + }, + { + "bbox": [ + 313, + 329, + 556, + 483 + ], + "type": "text", + "content": " since a higher learning rate leads to NAN issues. For other experiments, we simply train SiT-L of 458M parameters for 400K steps. More experimental details are provided in Appendix C.2." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 484, + 556, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 484, + 556, + 604 + ], + "spans": [ + { + "bbox": [ + 313, + 484, + 556, + 604 + ], + "type": "text", + "content": "Evaluation. We use reconstruction Frechet Inception Distance (rFID) [37] on ImageNet validation set to evaluate the tokenizer. To evaluate the performance of generation tasks, we report generation FID (gFID), Inception Score (IS) [90], Precision and Recall [52] (in Appendix D.5), with and without classifier-free guidance (CFG) [40]. We measure the efficiency of the generative models by GLOPs of the model's forward pass on the latent codes of tokenizers, and training and inference throughput for the models using floating point 32 and a batch size of 64 on a single AMD MI250." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 314, + 612, + 399, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 612, + 399, + 623 + ], + "spans": [ + { + "bbox": [ + 314, + 612, + 399, + 623 + ], + "type": "text", + "content": "4.2. Main Results" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 630, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 555, + 713 + ], + "type": "text", + "content": "We present the main reconstruction and generation results on the ImageNet benchmark of resolution " + }, + { + "bbox": [ + 313, + 630, + 555, + 713 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 313, + 630, + 555, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 630, + 555, + 713 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 313, + 630, + 555, + 713 + ], + "type": "text", + "content": " in Tab. 1 and Tab. 2, respectively. We show that SoftVQ-VAE achieves reconstruction and generation performance (with generative models) comparable to leading systems with 32 and 64 tokens, while presenting a significant improvement in both training and inference efficiency." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "28362" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 115, + 99, + 492, + 396 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 95 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 95 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 95 + ], + "type": "text", + "content": "Table 1. System-level comparison on ImageNet " + }, + { + "bbox": [ + 55, + 70, + 555, + 95 + ], + "type": "inline_equation", + "content": "{256} \\times {256}" + }, + { + "bbox": [ + 55, + 70, + 555, + 95 + ], + "type": "text", + "content": " conditional generation. We compare with both diffusion-based models and auto-regressive models with different types of tokenizers. Grey denotes SoftVQ. † indicates results with DPM-Solver and 50 steps." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 115, + 99, + 492, + 396 + ], + "lines": [ + { + "bbox": [ + 115, + 99, + 492, + 396 + ], + "spans": [ + { + "bbox": [ + 115, + 99, + 492, + 396 + ], + "type": "table", + "html": "
Gen. ModelTok. Model# Tokens ↓Tok. rFID ↓Gflops ↓Throughput (imgs/sec) ↑w/o CFGw/ CFG
gFID ↓IS ↑gFID ↓IS ↑
Auto-regressive Models
Taming-Trans. [23]VQ2567.94--5.20290.3--
RQ-Trans. [55]RQ2563.20908.917.853.80323.7--
MaskGIT [10]VQ2562.28--6.18182.1--
MAGE [59]VQ256---6.93195.8--
LlamaGen-3B [98]VQ2562.19781.562.90--3.06279.7
TiTok-S-128 [115]VQ1281.6133.036.50--1.97281.8
MAR-H [60]KL2561.22145.080.122.35227.81.55303.7
MAR-HSoftVQ-L320.6167.932.083.83211.22.54273.6
SoftVQ-BL640.6586.550.892.81218.31.93289.4
Diffusion-based Models
LDM-4 [88]KL40960.27157.920.3710.56103.53.60247.7
U-ViT-H/2†[4]128.890.98--2.29263.9
MDTv2-XL/2 [29]125.430.595.06155.61.58314.7
DiT-XL/2 [80]KL10240.6280.730.519.62121.52.27278.2
SiT-XL/2 [72]81.920.548.30131.72.06270.3
+ REPA [116]5.90157.81.42305.7
SoftVQ-B0.898.949.83113.83.91264.2
SoftVQ-BL0.6814.528.819.22115.83.78266.7
SoftVQ-L0.748.749.07117.23.69270.4
DiT-XLSoftVQ-B0.884.706.62129.23.29262.5
SoftVQ-BL0.6528.814.596.53131.93.11268.3
SoftVQ-L0.614.515.83141.32.93268.5
SiT-XLSoftVQ-B0.8910.287.99129.32.51301.3
SoftVQ-BL0.6814.7310.127.67133.92.44308.8
SoftVQ-L0.7410.117.59137.42.44310.6
SiT-XLSoftVQ-B0.885.515.98138.01.78279.0
SoftVQ-BL0.6529.235.425.80143.51.88287.9
SoftVQ-L0.615.395.35151.21.86293.6
", + "image_path": "6f534cfe17f8f24ba92d11fc7553d01bfe979899be63eafbcb83ef2175368637.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 403, + 297, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 403, + 297, + 521 + ], + "spans": [ + { + "bbox": [ + 54, + 403, + 297, + 521 + ], + "type": "text", + "content": "Consistent high-quality reconstruction. SoftVQ variants achieve remarkable reconstruction quality with a high compression ratio. With only 32 and 64 tokens, our models maintain consistent rFID scores, i.e., 0.61-0.89 on the " + }, + { + "bbox": [ + 54, + 403, + 297, + 521 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 54, + 403, + 297, + 521 + ], + "type": "text", + "content": " and 0.64-0.71 on " + }, + { + "bbox": [ + 54, + 403, + 297, + 521 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 54, + 403, + 297, + 521 + ], + "type": "text", + "content": " benchmark, substantially outperforming VQ tokenizers with 256 tokens used in autoregressive models such as MaskGIT [10] and TiTok [115]. Our results are also comparable to KL tokenizers used in previous denoising-based models, using " + }, + { + "bbox": [ + 54, + 403, + 297, + 521 + ], + "type": "inline_equation", + "content": "32x" + }, + { + "bbox": [ + 54, + 403, + 297, + 521 + ], + "type": "text", + "content": " fewer tokens. More discussion of different tokenizers is given in Sec. 4.3." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 522, + 297, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 522, + 297, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 522, + 297, + 713 + ], + "type": "text", + "content": "Generation performance comparable to SOTA. Notably, with only 32 tokens, the generation performance of DiT-XL and SiT-XL trained on all variants of the proposed SoftVQ significantly outperforms DiT-XL/2 and SiT-XL/2 with KL tokenizers of 1024 tokens, without using CFG. DiT-XL with SoftVQ-L of 32 tokens presents a 0.62 improvement on FID and SiT-XL with SoftVQ-L of 32 tokens shows a 0.71 improvement on FID, without using CFG. The improvement is further strengthened with 64 tokens. SiT-XL with SoftVQ-L of 64 tokens outperforms SiT-XL/2 with REPA [116], without using CFG. Applying CFG, SiT-XL with SoftVQ of 64 tokens achieves a performance comparable to the leading systems with the best FID as 1.78 on 256 and 2.21 on 512 benchmark. Using MAR-H [60], SoftVQ-BL and SoftVQ-L with 64 and 32 tokens present slightly worse results, compared to the KL tokenizer of 256 tokens, possibly due to the" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 403, + 555, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 403, + 555, + 426 + ], + "spans": [ + { + "bbox": [ + 313, + 403, + 555, + 426 + ], + "type": "text", + "content": "lowered learning rate. MAR-H and SiT-XL results on 512 benchmark with 32 tokens are included in Appendix D.2." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 426, + 556, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 556, + 654 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 556, + 654 + ], + "type": "text", + "content": "Efficiency. The system-level comparison reveals SoftVQ's exceptional efficiency across multiple metrics. When integrated with MAR-H at " + }, + { + "bbox": [ + 313, + 426, + 556, + 654 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 313, + 426, + 556, + 654 + ], + "type": "text", + "content": " resolution, SoftVQ reduces GFLOPs by " + }, + { + "bbox": [ + 313, + 426, + 556, + 654 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 313, + 426, + 556, + 654 + ], + "type": "text", + "content": " compared to standard KL tokenizer, i.e., 86.55 vs 145.08 GFLOPs with 64 tokens and further reducing to 67.93 GFLOPs with 32 tokens. SoftVQ enables DiT-XL and SiT-XL to achieve significantly lower computational costs i.e., 28.81 and 29.23 GFLOPs, respectively with 64 tokens, compared to their KL counterparts (80.73 GFLOPs with 1024 tokens). Importantly, these efficiency improvements come without compromising generation quality. SiT-XL trained on SoftVQ-BL with 64 tokens presents " + }, + { + "bbox": [ + 313, + 426, + 556, + 654 + ], + "type": "inline_equation", + "content": "55\\mathrm{x}" + }, + { + "bbox": [ + 313, + 426, + 556, + 654 + ], + "type": "text", + "content": " throughput than the KL counterpart, while showing competitive IS and FID scores in all configurations. Not only do we require fewer training iterations to achieve a comparable performance to leading systems, i.e., 2.5M vs 7M steps of SiT-XL and DiT-XL, the high compression ratio of SoftVQ also improves the training throughput, reducing the time to train SiT-XL for 400K steps from 72 to 20 hours on 8 MI250." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 660, + 462, + 673 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 660, + 462, + 673 + ], + "spans": [ + { + "bbox": [ + 313, + 660, + 462, + 673 + ], + "type": "text", + "content": "4.3. Comparison of Tokenizers" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "text", + "content": "We compare the proposed SoftVQ-VAE with the concurrent efficient image tokenizers, i.e., TiTok [115] and DC-AE [11]. To show the superiority of SoftVQ-VAE in achieving a" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "28363" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 115, + 97, + 493, + 288 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 95 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 95 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 95 + ], + "type": "text", + "content": "Table 2. System-level comparison on ImageNet " + }, + { + "bbox": [ + 55, + 70, + 555, + 95 + ], + "type": "inline_equation", + "content": "{512} \\times {512}" + }, + { + "bbox": [ + 55, + 70, + 555, + 95 + ], + "type": "text", + "content": " conditional generation. We compare with both diffusion-based models and auto-regressive models with different types of tokenizers. Grey denotes SoftVQ. † indicates results with DPM-Solver and 50 steps." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 115, + 97, + 493, + 288 + ], + "lines": [ + { + "bbox": [ + 115, + 97, + 493, + 288 + ], + "spans": [ + { + "bbox": [ + 115, + 97, + 493, + 288 + ], + "type": "table", + "html": "
Gen. ModelTok. Model# Tokens ↓Tok. rFID ↓Gflops ↓Throughput (imgs/sec) ↑w/o CFGw/ CFG
gFID ↓IS ↑gFID ↓IS ↑
Generative Adversarial Models
BigGAN [10]-------8.43177.9
StyleGAN-XL [48]-------2.41267.7
Auto-regressive Models
MaskGIT [10]VQ10241.97--7.32156.0--
TiTok-B [115]VQ1281.52----2.13261.2
MAR-HSoftVQ-BL640.7186.551.508.21152.93.42261.8
Diffusion-based Models
ADM [16]-----23.2458.063.85221.7
U-ViT-H/4† [4]128.920.58--4.05263.8
DiT-XL/2 [80]KL40960.62373.340.109.62121.53.04240.8
SiT-XL/2 [72]373.320.10--2.62252.2
DiT-XL [11]AE2560.2280.751.029.56-2.84-
UViT-H† [11]128.906.149.83-2.53-
UViT-H† [11]AE640.2232.9915.2312.26-2.66-
UViT-2B† [11]104.189.446.50-2.25-
SiT-XLSoftVQ-BL640.7129.235.387.96133.92.21290.5
", + "image_path": "2e02e28a25eca197a9eff94e7290d8deef87f6a01fadf5c863ee97767a051901.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 293, + 295, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 293, + 295, + 353 + ], + "spans": [ + { + "bbox": [ + 55, + 293, + 295, + 353 + ], + "type": "text", + "content": "high compression ratio, we additionally train VQ-VAE and AE using the small (S) configuration and the same training recipe of SoftVQ. To validate the generation performance, we train a SiT-L for 400K steps and report gFID and IS on 256 ImageNet without using CFG, as shown in Tab. 3." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 354, + 295, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 354, + 295, + 496 + ], + "spans": [ + { + "bbox": [ + 54, + 354, + 295, + 496 + ], + "type": "text", + "content": "Scalable performance. With a much smaller model size, i.e., 46M vs 390M, SoftVQ-S significantly outperforms both TiTok variants at 64 tokens, achieving better reconstruction quality with an rFID of 1.03 compared to 1.25, and better generation performance with a gFID of 11.24 and IS of 89.4 compared to the best TiTok gFID of 19.23 and IS of 61.8. Noteworthy is that the generation results SiT-L trained on SoftVQ with 32 tokens outperform those trained on KL tokenizer with 1024 tokens [72] by a large margin, i.e., 5.9 in gFID. Compared to DC-AE, SoftVQ-S demonstrates competitive scalability across token counts while maintaining significant generation quality even with fewer parameters." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 497, + 297, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 497, + 297, + 640 + ], + "spans": [ + { + "bbox": [ + 55, + 497, + 297, + 640 + ], + "type": "text", + "content": "Less Lossy Property. SoftVQ-S exhibits remarkably consistent performance across different compression ratios. When reducing tokens from 256 to 32, SoftVQ-S maintains a minimal degradation in rFID from 0.80 to 1.24, significantly outperforming both VQ-S, from 1.45 to 10.97, and AE-S, from 1.15 to 2.01. The benefits of the fully-differentiable property of SoftVQ along with representation alignment are particularly evident in generation quality, where SoftVQ-S achieves substantially better gFID of 9.21-12.89, and IS scores of 79.5-93.6, compared to baselines, showing its robustness in preserving information at high compression ratios and a better latent space for generative models." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 648, + 228, + 661 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 648, + 228, + 661 + ], + "spans": [ + { + "bbox": [ + 55, + 648, + 228, + 661 + ], + "type": "text", + "content": "4.4. Discussions on the Latent Space" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 665, + 282, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 282, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 282, + 677 + ], + "type": "text", + "content": "We perform more analysis on the latent of SoftVQ here." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "Alignment with different initialization and target models. As shown in Tab. 4, our results demonstrate the effectiveness of alignment across various pre-trained models. The" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 315, + 316, + 553, + 508 + ], + "blocks": [ + { + "bbox": [ + 313, + 292, + 555, + 314 + ], + "lines": [ + { + "bbox": [ + 313, + 292, + 555, + 314 + ], + "spans": [ + { + "bbox": [ + 313, + 292, + 555, + 314 + ], + "type": "text", + "content": "Table 3. Comparison of tokenizers on class-conditional " + }, + { + "bbox": [ + 313, + 292, + 555, + 314 + ], + "type": "inline_equation", + "content": "{256} \\times {256}" + }, + { + "bbox": [ + 313, + 292, + 555, + 314 + ], + "type": "text", + "content": " ImageNet. We report rFID, and gFID and IS of SiT-L without CFG." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 316, + 553, + 508 + ], + "lines": [ + { + "bbox": [ + 315, + 316, + 553, + 508 + ], + "spans": [ + { + "bbox": [ + 315, + 316, + 553, + 508 + ], + "type": "table", + "html": "
Tokenizer# Params# TokensDim.rFID ↓SiT-L
gFID ↓IS ↑
KL [96]676M102440.6218.7972.0
TiTok-BL-KL [115]389M64161.2523.3554.7
TiTok-BL-VQ [115]390M64162.0619.2361.8
DC-AE-f32 [11]323M64320.6915.3676.1
DC-AE-f54 [11]677M161280.8124.6649.1
VQ-S46M25681.4519.0258.6
128162.6118.5162.4
64324.0420.0754.5
326410.9721.5747.6
AE-S46M25681.1522.1158.3
128161.3322.3158.5
64321.6425.5354.3
32642.0128.1845.5
SoftVQ-S46M25680.809.2193.6
128160.9210.1285.8
64321.0311.2489.4
32641.2412.8979.5
", + "image_path": "5fafdd765b3041d841f0695ec4eb87bd1546e4c5578558b67ffaf8c892f6a39b.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 312, + 510, + 556, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 510, + 556, + 605 + ], + "spans": [ + { + "bbox": [ + 312, + 510, + 556, + 605 + ], + "type": "text", + "content": "encoder initialization and alignment with DINOv2-B [79] achieve superior performance with rFID 0.88 and IS 103.4, compared to using either component alone. Further improvements are observed with CLIP-B [82] and EVA-02-B [26] on reconstruction. We reveal that a better rFID does not necessarily translate to a better gFID. Instead, the latent space quality, reflected by linear probing accuracy, is more closely related to the performance of the generative models." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 606, + 557, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 606, + 557, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 606, + 557, + 714 + ], + "type": "text", + "content": "Linear probing of tokenizer and generative model. Linear probing results in Fig. 3 (details in Appendix C.3) reveal SoftVQ's superior representation quality at both tokenizer and generative model. At the tokenizer level, the linear probing accuracy of SoftVQ variants (B, BL, L) consistently outperforms VQ-S and AE-S across all token counts, with the gaps widening at lower token counts, similarly observed by Yu et al. [115]. This advantage transfers to the generative model with the best linear probing accuracy on the" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "28364" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 78, + 95, + 271, + 166 + ], + "blocks": [ + { + "bbox": [ + 55, + 71, + 297, + 93 + ], + "lines": [ + { + "bbox": [ + 55, + 71, + 297, + 93 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 297, + 93 + ], + "type": "text", + "content": "Table 4. Representation alignment with different encoder initialization and the target alignment models using SoftVQ-B 64 tokens." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 78, + 95, + 271, + 166 + ], + "lines": [ + { + "bbox": [ + 78, + 95, + 271, + 166 + ], + "spans": [ + { + "bbox": [ + 78, + 95, + 271, + 166 + ], + "type": "table", + "html": "
Enc. Init.Align. TargetrFID ↓L.P. ↑SiT-L
gFID ↓IS ↑
--0.985.4220.3389.4
-DINOv2-B [79]0.9341.0810.96101.9
DINOv2-B [79]-0.7311.8717.2091.5
DINOv2-B [79]DINOv2-B [79]0.8842.4210.13103.4
CLIP-B [82]CLIP-B [82]0.7547.129.89109.3
EVA-02-B [26]EVA-02-B [26]0.8144.2310.34112.7
", + "image_path": "f1f8bbfa8b3f0d1dbf7e46aee5045af4ff3e9be53c5d870701945fe5a5f5cb30.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 169, + 256, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 169, + 256, + 180 + ], + "spans": [ + { + "bbox": [ + 55, + 169, + 256, + 180 + ], + "type": "text", + "content": "intermediate features of SiT trained with SoftVQ." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 180, + 269, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 180, + 269, + 264 + ], + "spans": [ + { + "bbox": [ + 55, + 180, + 269, + 264 + ], + "type": "text", + "content": "Latent space visualization. UMAP visualizations latent space in Fig. 4 demonstrate SoftVQ's ability totain structured and discriminative latent representation minimal variance between encoder output " + }, + { + "bbox": [ + 55, + 180, + 269, + 264 + ], + "type": "inline_equation", + "content": "(\\hat{\\mathbf{z}})" + }, + { + "bbox": [ + 55, + 180, + 269, + 264 + ], + "type": "text", + "content": " and d input (z). Comparing Fig. 4 (d) and (c), we also observe larger encoder can learn a more discriminative latent showing the superiority of SoftVQ's representation le" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 271, + 156, + 283 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 271, + 156, + 283 + ], + "spans": [ + { + "bbox": [ + 55, + 271, + 156, + 283 + ], + "type": "text", + "content": "4.5. Ablation Studies" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 289, + 296, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 289, + 296, + 445 + ], + "spans": [ + { + "bbox": [ + 55, + 289, + 296, + 445 + ], + "type": "text", + "content": "We present a series of ablation studies in Append including SoftVQ variants with product quantization [45], residual quantization (RQ) [55], and GMMVQ book size, latent space size, and the temperature of S The results show the compatibility of SoftVQ with I RQ, with slightly improved performance. GMMVQ p results comparable to SoftVQ. Moreover, SoftVQ shows robustness to other parameters, such as codebook size and softmax temperature. We found that while increasing the dimension of the latent space can result in a significant improvement in reconstruction, it leads to deterioration in generation performance, possibly due to the learning difficulty of generative models with larger dimensions [92]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 455, + 142, + 467 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 455, + 142, + 467 + ], + "spans": [ + { + "bbox": [ + 55, + 455, + 142, + 467 + ], + "type": "text", + "content": "5. Related Work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 475, + 297, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 475, + 297, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 475, + 297, + 677 + ], + "type": "text", + "content": "Image tokenization has emerged as a fundamental technique to bridge various vision tasks. Traditional approaches using autoencoders [39, 106] laid the groundwork by compressing images into latent representations. This field subsequently diverged into two main branches: generation-focused and understanding-focused approaches. Generation-oriented methods, such as VAE [84, 104] and VQ-GAN [23, 85], emphasized learning latent spaces for detail-preserving compression. These were further refined through variants [55, 73, 112, 124] that enhanced generation quality. Parallel developments in understanding-focused approaches leveraged Large Language Models (LLMs) [105] to create semantic representations for tasks like classification [21] and detection [125]. Recent work [62, 115] has demonstrated the potential of unifying these approaches, with subsequent research [34, 109] exploring the convergence of generation and understanding capabilities using a single tokenizer." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 678, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 678, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 678, + 296, + 713 + ], + "type": "text", + "content": "Image generation have two popular paradigms: autoregressive and diffusion. Autoregressive models start from CNN-based architectures [103] and evolve to the transformer-" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 328, + 72, + 541, + 171 + ], + "blocks": [ + { + "bbox": [ + 328, + 72, + 541, + 171 + ], + "lines": [ + { + "bbox": [ + 328, + 72, + 541, + 171 + ], + "spans": [ + { + "bbox": [ + 328, + 72, + 541, + 171 + ], + "type": "image", + "image_path": "c62c016c899d097d2b6ba4621e9d2176cf6e114101d22392b9e37a8a01bf5f81.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 326, + 215, + 541, + 308 + ], + "blocks": [ + { + "bbox": [ + 326, + 215, + 541, + 308 + ], + "lines": [ + { + "bbox": [ + 326, + 215, + 541, + 308 + ], + "spans": [ + { + "bbox": [ + 326, + 215, + 541, + 308 + ], + "type": "image", + "image_path": "44aa5ef680a0f6a0e40648c3631596a68ea02c174ac34dee31e57d756c0eee2c.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 309, + 555, + 342 + ], + "lines": [ + { + "bbox": [ + 313, + 309, + 555, + 342 + ], + "spans": [ + { + "bbox": [ + 313, + 309, + 555, + 342 + ], + "type": "text", + "content": "Figure 4. Visualization of " + }, + { + "bbox": [ + 313, + 309, + 555, + 342 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{z}}" + }, + { + "bbox": [ + 313, + 309, + 555, + 342 + ], + "type": "text", + "content": " (top), i.e., encoder output, and " + }, + { + "bbox": [ + 313, + 309, + 555, + 342 + ], + "type": "inline_equation", + "content": "\\mathbf{z}" + }, + { + "bbox": [ + 313, + 309, + 555, + 342 + ], + "type": "text", + "content": " (bottom), i.e., decoder input, of (a) VQ-S 64; (b) SoftVQ-S 64; (c) SoftVQ-S 32; (d) SoftVQ-L 32, trained with latent space alignment." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 348, + 556, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 348, + 556, + 552 + ], + "spans": [ + { + "bbox": [ + 313, + 348, + 556, + 552 + ], + "type": "text", + "content": "based architectures [55, 66, 98, 105, 114] to enhance scalability. Latest works in visual autoregressive modeling, including VAR [99], MAR [60], and ImageFolder [62], have further enhanced generation efficiency. Diffusion models have transformed image generation since their introduction by Sohl-Dickstein et al. [93]. Key advancements by Nichol et al. [75], Dhariwal et al. [16], and Song et al. [95] have significantly improved model performance. A paradigm shift occurred with the adoption of latent space diffusion [88, 102], leveraging pre-trained image encoders as priors [23, 104] to enhance generation efficiency. Architectural innovations, particularly the integration of transformers [80], have further expanded the capabilities of these models. These advances have spawned numerous influential systems, from text-guided generation [18, 76] to high-fidelity synthesis [28, 89]. Commercial successes [74, 78, 83, 87] have also demonstrated the practical impact of diffusion models." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 562, + 388, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 562, + 388, + 574 + ], + "spans": [ + { + "bbox": [ + 314, + 562, + 388, + 574 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "text", + "content": "We presented SoftVQ-VAE, a continuous image tokenizer leveraging soft categorical posteriors to aggregate multiple codewords into each latent token. Our approach compresses images to just 32 or 64 tokens while maintaining high reconstruction quality and enabling state-of-the-art generation results with DiT, SiT, and MAR. The substantial improvements in both inference throughput (up to " + }, + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "inline_equation", + "content": "55\\mathrm{x}" + }, + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "text", + "content": " with 64 tokens) and training efficiency, combined with the fully-differentiable design's ability to learn semantic-rich representations, demonstrate SoftVQ-VAE's potential as a foundation for efficient generative and vision-language models." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28365" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 62, + 91, + 297, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 66, + 91, + 297, + 133 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 91, + 297, + 133 + ], + "spans": [ + { + "bbox": [ + 66, + 91, + 297, + 133 + ], + "type": "text", + "content": "[1] Michael S Albergo, Nicholas M Boffi, and Eric Vanden-Eijnden. Stochastic interpolants: A unifying framework for flows and diffusions. arXiv preprint arXiv:2303.08797, 2023. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 136, + 296, + 167 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 136, + 296, + 167 + ], + "spans": [ + { + "bbox": [ + 66, + 136, + 296, + 167 + ], + "type": "text", + "content": "[2] Brian DO Anderson. Reverse-time diffusion equation models. Stochastic Processes and their Applications, 12(3): 313-326, 1982. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 66, + 170, + 296, + 223 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 170, + 296, + 223 + ], + "spans": [ + { + "bbox": [ + 66, + 170, + 296, + 223 + ], + "type": "text", + "content": "[3] Alexei Baevski, Yuhao Zhou, Abdelrahman Mohamed, and Michael Auli. wav2vec 2.0: A framework for self-supervised learning of speech representations. Advances in neural information processing systems, 33:12449-12460, 2020. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 66, + 224, + 296, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 224, + 296, + 278 + ], + "spans": [ + { + "bbox": [ + 66, + 224, + 296, + 278 + ], + "type": "text", + "content": "[4] Fan Bao, Shen Nie, Kaiwen Xue, Yue Cao, Chongxuan Li, Hang Su, and Jun Zhu. All are worth words: A vit backbone for diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 22669-22679, 2023. 6, 7, 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 66, + 280, + 296, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 280, + 296, + 312 + ], + "spans": [ + { + "bbox": [ + 66, + 280, + 296, + 312 + ], + "type": "text", + "content": "[5] Hangbo Bao, Li Dong, Songhao Piao, and Furu Wei. Beit: Bert pre-training of image transformers. arXiv preprint arXiv:2106.08254, 2021. 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 66, + 314, + 296, + 367 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 314, + 296, + 367 + ], + "spans": [ + { + "bbox": [ + 66, + 314, + 296, + 367 + ], + "type": "text", + "content": "[6] Omer Bar-Tal, Hila Chefer, Omer Tov, Charles Herrmann, Roni Paiss, Shiran Zada, Ariel Ephrat, Junhwa Hur, Guanghui Liu, Amit Raj, et al. Lumiere: A spacetime diffusion model for video generation. arXiv preprint arXiv:2401.12945, 2024. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 66, + 369, + 296, + 412 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 369, + 296, + 412 + ], + "spans": [ + { + "bbox": [ + 66, + 369, + 296, + 412 + ], + "type": "text", + "content": "[7] Yoshua Bengio, Nicholas Léonard, and Aaron Courville. Estimating or propagating gradients through stochastic neurons for conditional computation. arXiv preprint arXiv:1308.3432, 2013. 1, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 66, + 415, + 296, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 415, + 296, + 468 + ], + "spans": [ + { + "bbox": [ + 66, + 415, + 296, + 468 + ], + "type": "text", + "content": "[8] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 66, + 470, + 296, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 470, + 296, + 523 + ], + "spans": [ + { + "bbox": [ + 66, + 470, + 296, + 523 + ], + "type": "text", + "content": "[9] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 5, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 525, + 296, + 556 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 525, + 296, + 556 + ], + "spans": [ + { + "bbox": [ + 62, + 525, + 296, + 556 + ], + "type": "text", + "content": "[10] Huiwen Chang, Han Zhang, Lu Jiang, Ce Liu, and William T. Freeman. Maskgit: Masked generative image transformer, 2022. 1, 3, 6, 7, 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 559, + 296, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 559, + 296, + 612 + ], + "spans": [ + { + "bbox": [ + 62, + 559, + 296, + 612 + ], + "type": "text", + "content": "[11] Junyu Chen, Han Cai, Junsong Chen, Enze Xie, Shang Yang, Haotian Tang, Muyang Li, Yao Lu, and Song Han. Deep compression autoencoder for efficient high-resolution diffusion models. arXiv preprint arXiv:2410.10733, 2024. 1, 6, 7, 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 62, + 614, + 296, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 614, + 296, + 667 + ], + "spans": [ + { + "bbox": [ + 62, + 614, + 296, + 667 + ], + "type": "text", + "content": "[12] Liang Chen, Haozhe Zhao, Tianyu Liu, Shuai Bai, Junyang Lin, Chang Zhou, and Baobao Chang. An image is worth 1/2 tokens after layer 2: Plug-and-play inference acceleration for large vision-language models. arXiv preprint arXiv:2403.06764, 2024. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 62, + 670, + 296, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 670, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 62, + 670, + 296, + 712 + ], + "type": "text", + "content": "[13] Xi Chen, Diederik P Kingma, Tim Salimans, Yan Duan, Prafulla Dhariwal, John Schulman, Ilya Sutskever, and Pieter Abbeel. Variational lossy autoencoder. arXiv preprint arXiv:1611.02731, 2016. 1, 2, 3" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 320, + 73, + 555, + 714 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 320, + 73, + 555, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 73, + 555, + 106 + ], + "spans": [ + { + "bbox": [ + 320, + 73, + 555, + 106 + ], + "type": "text", + "content": "[14] Xinlei Chen, Zhuang Liu, Saining Xie, and Kaiming He. Deconstructing denoising diffusion models for self-supervised learning. arXiv preprint arXiv:2401.14404, 2024. 1, 2, 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 320, + 107, + 555, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 107, + 555, + 151 + ], + "spans": [ + { + "bbox": [ + 320, + 107, + 555, + 151 + ], + "type": "text", + "content": "[15] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248–255. IEEE, 2009. 5, 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 320, + 152, + 555, + 175 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 152, + 555, + 175 + ], + "spans": [ + { + "bbox": [ + 320, + 152, + 555, + 175 + ], + "type": "text", + "content": "[16] Prafulla Dhariwal and Alex Nichol. Diffusion models beat gans on image synthesis, 2021. 1, 7, 8, 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 320, + 176, + 555, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 176, + 555, + 229 + ], + "spans": [ + { + "bbox": [ + 320, + 176, + 555, + 229 + ], + "type": "text", + "content": "[17] Nat Dilokthanakul, Pedro AM Mediano, Marta Garnelo, Matthew CH Lee, Hugh Salimbeni, Kai Arulkumaran, and Murray Shanahan. Deep unsupervised clustering with gaussian mixture variational autoencoders. arXiv preprint arXiv:1611.02648, 2016. 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 320, + 232, + 555, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 232, + 555, + 285 + ], + "spans": [ + { + "bbox": [ + 320, + 232, + 555, + 285 + ], + "type": "text", + "content": "[18] Ming Ding, Zhuoyi Yang, Wenyi Hong, Wendi Zheng, Chang Zhou, Da Yin, Junyang Lin, Xu Zou, Zhou Shao, Hongxia Yang, et al. Cogview: Mastering text-to-image generation via transformers. Advances in neural information processing systems, 34:19822-19835, 2021. 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 320, + 288, + 555, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 288, + 555, + 352 + ], + "spans": [ + { + "bbox": [ + 320, + 288, + 555, + 352 + ], + "type": "text", + "content": "[19] Xiaoyi Dong, Jianmin Bao, Ting Zhang, Dongdong Chen, Weiming Zhang, Lu Yuan, Dong Chen, Fang Wen, Nenghai Yu, and Baining Guo. Peco: Perceptual codebook for bert pre-training of vision transformers. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 552-560, 2023. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 320, + 354, + 555, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 354, + 555, + 397 + ], + "spans": [ + { + "bbox": [ + 320, + 354, + 555, + 397 + ], + "type": "text", + "content": "[20] Alexey Dosovitskiy and Thomas Brox. Generating images with perceptual similarity metrics based on deep networks. Advances in neural information processing systems, 29, 2016. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 320, + 399, + 555, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 399, + 555, + 464 + ], + "spans": [ + { + "bbox": [ + 320, + 399, + 555, + 464 + ], + "type": "text", + "content": "[21] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale, 2021. 3, 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 320, + 466, + 555, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 466, + 555, + 498 + ], + "spans": [ + { + "bbox": [ + 320, + 466, + 555, + 498 + ], + "type": "text", + "content": "[22] Joseph C Dunn. A fuzzy relative of the isodata process and its use in detecting compact well-separated clusters. 1973. 2, 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 320, + 500, + 555, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 500, + 555, + 554 + ], + "spans": [ + { + "bbox": [ + 320, + 500, + 555, + 554 + ], + "type": "text", + "content": "[23] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12873-12883, 2021. 1, 2, 3, 5, 6, 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 320, + 556, + 555, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 556, + 555, + 589 + ], + "spans": [ + { + "bbox": [ + 320, + 556, + 555, + 589 + ], + "type": "text", + "content": "[24] Zach Evans, CJ Carr, Josiah Taylor, Scott H Hawley, and Jordi Pons. Fast timing-conditioned latent audio diffusion. arXiv preprint arXiv:2402.04825, 2024. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 320, + 590, + 555, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 590, + 555, + 644 + ], + "spans": [ + { + "bbox": [ + 320, + 590, + 555, + 644 + ], + "type": "text", + "content": "[25] Lijie Fan, Tianhong Li, Siyang Qin, Yuzheng Li, Chen Sun, Michael Rubinstein, Deqing Sun, Kaiming He, and Yonglong Tian. Fluid: Scaling autoregressive text-to-image generative models with continuous tokens. arXiv preprint arXiv:2410.13863, 2024. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 320, + 647, + 555, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 647, + 555, + 689 + ], + "spans": [ + { + "bbox": [ + 320, + 647, + 555, + 689 + ], + "type": "text", + "content": "[26] Yuxin Fang, Quan Sun, Xinggang Wang, Tiejun Huang, Xinlong Wang, and Yue Cao. Eva-02: A visual representation for neon genesis. Image and Vision Computing, 149:105171, 2024. 7, 8" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 320, + 691, + 555, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 691, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 320, + 691, + 555, + 714 + ], + "type": "text", + "content": "[27] Christopher Fifty, Ronald G Junkins, Dennis Duan, Aniketh Iger, Jerry W Liu, Ehsan Amid, Sebastian Thrun, and" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "28366" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 61, + 73, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 81, + 73, + 295, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 73, + 295, + 95 + ], + "spans": [ + { + "bbox": [ + 81, + 73, + 295, + 95 + ], + "type": "text", + "content": "Christopher Ré. Restructuring vector quantization with the rotation trick. arXiv preprint arXiv:2410.06424, 2024. 2, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 97, + 296, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 97, + 296, + 151 + ], + "spans": [ + { + "bbox": [ + 61, + 97, + 296, + 151 + ], + "type": "text", + "content": "[28] Oran Gafni, Adam Polyak, Oron Ashual, Shelly Sheynin, Devi Parikh, and Yaniv Taigman. Make-a-scene: Scene-based text-to-image generation with human priors. In European Conference on Computer Vision, pages 89-106. Springer, 2022. 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 153, + 296, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 153, + 296, + 196 + ], + "spans": [ + { + "bbox": [ + 61, + 153, + 296, + 196 + ], + "type": "text", + "content": "[29] Shanghua Gao, Pan Zhou, Ming-Ming Cheng, and Shuicheng Yan. Mdtv2: Masked diffusion transformer is a strong image synthesizer. arXiv preprint arXiv:2303.14389, 2023.6,5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 199, + 296, + 232 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 199, + 296, + 232 + ], + "spans": [ + { + "bbox": [ + 62, + 199, + 296, + 232 + ], + "type": "text", + "content": "[30] Yuying Ge, Yixiao Ge, Ziyun Zeng, Xintao Wang, and Ying Shan. Planting a seed of vision in large language model, 2023. 1, 2, 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 234, + 296, + 276 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 234, + 296, + 276 + ], + "spans": [ + { + "bbox": [ + 62, + 234, + 296, + 276 + ], + "type": "text", + "content": "[31] Yuying Ge, Sijie Zhao, Ziyun Zeng, Yixiao Ge, Chen Li, Xintao Wang, and Ying Shan. Making llama see and draw with seed tokenizer. arXiv preprint arXiv:2310.01218, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 280, + 295, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 280, + 295, + 323 + ], + "spans": [ + { + "bbox": [ + 62, + 280, + 295, + 323 + ], + "type": "text", + "content": "[32] Shansan Gong, Mukai Li, Jiangtao Feng, Zhiyong Wu, and LingPeng Kong. Diffuseq: Sequence to sequence text generation with diffusion models. arXiv preprint arXiv:2210.08933, 2022. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 326, + 296, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 326, + 296, + 369 + ], + "spans": [ + { + "bbox": [ + 62, + 326, + 296, + 369 + ], + "type": "text", + "content": "[33] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139-144, 2020. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 372, + 296, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 372, + 296, + 404 + ], + "spans": [ + { + "bbox": [ + 62, + 372, + 296, + 404 + ], + "type": "text", + "content": "[34] Yuchao Gu, Xintao Wang, Yixiao Ge, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Rethinking the objectives of vector-quantized tokenizers for image synthesis, 2023. 2, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 406, + 296, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 406, + 296, + 460 + ], + "spans": [ + { + "bbox": [ + 62, + 406, + 296, + 460 + ], + "type": "text", + "content": "[35] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 1, 3, 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 463, + 296, + 506 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 463, + 296, + 506 + ], + "spans": [ + { + "bbox": [ + 62, + 463, + 296, + 506 + ], + "type": "text", + "content": "[36] Byeongho Heo, Song Park, Dongyoon Han, and Sangwoo Yun. Rotary position embedding for vision transformer. In European Conference on Computer Vision, pages 289-305. Springer, 2025. 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 509, + 296, + 562 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 509, + 296, + 562 + ], + "spans": [ + { + "bbox": [ + 62, + 509, + 296, + 562 + ], + "type": "text", + "content": "[37] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in Neural Information Processing Systems, 30, 2017. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 565, + 296, + 619 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 565, + 296, + 619 + ], + "spans": [ + { + "bbox": [ + 62, + 565, + 296, + 619 + ], + "type": "text", + "content": "[38] Irina Higgins, Loic Matthew, Arka Pal, Christopher P Burgess, Xavier Glorot, Matthew M Botvinick, Shakir Mohamed, and Alexander Lerchner. beta-vae: Learning basic visual concepts with a constrained variational framework. ICLR, 3, 2017. 1, 2, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 62, + 622, + 296, + 654 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 622, + 296, + 654 + ], + "spans": [ + { + "bbox": [ + 62, + 622, + 296, + 654 + ], + "type": "text", + "content": "[39] Geoffrey E Hinton and Ruslan R Salakhutdinov. Reducing the dimensionality of data with neural networks. science, 313(5786):504-507, 2006. 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 62, + 656, + 295, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 656, + 295, + 678 + ], + "spans": [ + { + "bbox": [ + 62, + 656, + 295, + 678 + ], + "type": "text", + "content": "[40] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 62, + 681, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 681, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 62, + 681, + 295, + 713 + ], + "type": "text", + "content": "[41] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 3" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 320, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 320, + 73, + 555, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 73, + 555, + 116 + ], + "spans": [ + { + "bbox": [ + 320, + 73, + 555, + 116 + ], + "type": "text", + "content": "[42] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. Advances in Neural Information Processing Systems, 35:8633-8646, 2022. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 320, + 118, + 555, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 118, + 555, + 173 + ], + "spans": [ + { + "bbox": [ + 320, + 118, + 555, + 173 + ], + "type": "text", + "content": "[43] Minyoung Huh, Brian Cheung, Pulkit Agrawal, and Phillip Isola. Straightening out the straight-through estimator: Overcoming optimization challenges in vector quantized networks. In International Conference on Machine Learning, pages 14096-14113. PMLR, 2023. 3, 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 320, + 175, + 555, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 175, + 555, + 206 + ], + "spans": [ + { + "bbox": [ + 320, + 175, + 555, + 206 + ], + "type": "text", + "content": "[44] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A. Efros. Image-to-image translation with conditional adversarial networks, 2018. 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 320, + 209, + 555, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 209, + 555, + 251 + ], + "spans": [ + { + "bbox": [ + 320, + 209, + 555, + 251 + ], + "type": "text", + "content": "[45] Herve Jegou, Matthijs Douze, and Cordelia Schmid. Product quantization for nearest neighbor search. IEEE transactions on pattern analysis and machine intelligence, 33(1):117-128, 2010. 2, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 320, + 254, + 555, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 254, + 555, + 308 + ], + "spans": [ + { + "bbox": [ + 320, + 254, + 555, + 308 + ], + "type": "text", + "content": "[46] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14, pages 694-711. Springer, 2016. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 320, + 309, + 555, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 309, + 555, + 363 + ], + "spans": [ + { + "bbox": [ + 320, + 309, + 555, + 363 + ], + "type": "text", + "content": "[47] Matthew J Johnson, David K Duvenaud, Alex Wiltschko, Ryan P Adams, and Sandeep R Datta. Composing graphical models with neural networks for structured representations and fast inference. Advances in neural information processing systems, 29, 2016. 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 320, + 365, + 555, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 365, + 555, + 418 + ], + "spans": [ + { + "bbox": [ + 320, + 365, + 555, + 418 + ], + "type": "text", + "content": "[48] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4401-4410, 2019. 5, 7, 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 320, + 422, + 555, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 422, + 555, + 475 + ], + "spans": [ + { + "bbox": [ + 320, + 422, + 555, + 475 + ], + "type": "text", + "content": "[49] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8110-8119, 2020. 5, 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 320, + 478, + 555, + 519 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 478, + 555, + 519 + ], + "spans": [ + { + "bbox": [ + 320, + 478, + 555, + 519 + ], + "type": "text", + "content": "[50] Diederik Kingma and Ruiqi Gao. Understanding diffusion objectives as the elbo with simple data augmentation. Advances in Neural Information Processing Systems, 36, 2024. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 320, + 522, + 554, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 522, + 554, + 544 + ], + "spans": [ + { + "bbox": [ + 320, + 522, + 554, + 544 + ], + "type": "text", + "content": "[51] Diederik P Kingma. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 1, 2, 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 320, + 545, + 554, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 545, + 554, + 589 + ], + "spans": [ + { + "bbox": [ + 320, + 545, + 554, + 589 + ], + "type": "text", + "content": "[52] Tuomas Kynkänniemi, Tero Karras, Samuli Laine, Jaakko Lehtinen, and Timo Aila. Improved precision and recall metric for assessing generative models. Advances in neural information processing systems, 32, 2019. 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 320, + 591, + 555, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 591, + 555, + 643 + ], + "spans": [ + { + "bbox": [ + 320, + 591, + 555, + 643 + ], + "type": "text", + "content": "[53] Tuomas Kynkänniemi, Miika Aittala, Tero Karras, Samuli Laine, Timo Aila, and Jaakko Lehtinen. Applying guidance in a limited interval improves sample and distribution quality in diffusion models. arXiv preprint arXiv:2404.07724, 2024. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 320, + 647, + 554, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 647, + 554, + 689 + ], + "spans": [ + { + "bbox": [ + 320, + 647, + 554, + 689 + ], + "type": "text", + "content": "[54] Anders Boesen Lindbo Larsen, Søren Kaae Sønderby, Hugo Larochelle, and Ole Winther. Autoencoding beyond pixels using a learned similarity metric. In International conference on machine learning, pages 1558-1566. PMLR, 2016. 5" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 320, + 692, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 692, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 320, + 692, + 554, + 713 + ], + "type": "text", + "content": "[55] Doyup Lee, Chiheon Kim, Saehoon Kim, Minsu Cho, and Wook-Shin Han. Autoregressive image generation using" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28367" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 61, + 73, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 81, + 73, + 296, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 73, + 296, + 106 + ], + "spans": [ + { + "bbox": [ + 81, + 73, + 296, + 106 + ], + "type": "text", + "content": "residual quantization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11523-11532, 2022. 2, 6, 8, 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 107, + 295, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 107, + 295, + 140 + ], + "spans": [ + { + "bbox": [ + 61, + 107, + 295, + 140 + ], + "type": "text", + "content": "[56] Doyup Lee, Chiheon Kim, Saehoon Kim, Minsu Cho, and Wook-Shin Han. Autoregressive image generation using residual quantization, 2022. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 141, + 295, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 141, + 295, + 184 + ], + "spans": [ + { + "bbox": [ + 61, + 141, + 295, + 184 + ], + "type": "text", + "content": "[57] Haopeng Li, Jinyue Yang, Kexin Wang, Xuerui Qiu, Yuhong Chou, Xin Li, and Guoqi Li. Scalable autoregressive image generation with mamba. arXiv preprint arXiv:2408.12245, 2024. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 186, + 296, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 186, + 296, + 239 + ], + "spans": [ + { + "bbox": [ + 62, + 186, + 296, + 239 + ], + "type": "text", + "content": "[58] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730-19742. PMLR, 2023. 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 241, + 295, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 241, + 295, + 285 + ], + "spans": [ + { + "bbox": [ + 62, + 241, + 295, + 285 + ], + "type": "text", + "content": "[59] Tianhong Li, Huiwen Chang, Shlok Kumar Mishra, Han Zhang, Dina Katabi, and Dilip Krishnan. Mage: Masked generative encoder to unify representation learning and image synthesis, 2023. 3, 6, 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 286, + 295, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 286, + 295, + 319 + ], + "spans": [ + { + "bbox": [ + 62, + 286, + 295, + 319 + ], + "type": "text", + "content": "[60] Tianhong Li, Yonglong Tian, He Li, Mingyang Deng, and Kaiming He. Autoregressive image generation without vector quantization, 2024. 1, 2, 3, 5, 6, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 320, + 295, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 320, + 295, + 363 + ], + "spans": [ + { + "bbox": [ + 62, + 320, + 295, + 363 + ], + "type": "text", + "content": "[61] Wentong Li, Yuqian Yuan, Jian Liu, Dongqi Tang, Song Wang, Jianke Zhu, and Lei Zhang. Tokenpacker: Efficient visual projector for multimodal llm. arXiv preprint arXiv:2407.02392, 2024. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 365, + 295, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 365, + 295, + 408 + ], + "spans": [ + { + "bbox": [ + 62, + 365, + 295, + 408 + ], + "type": "text", + "content": "[62] Xiang Li, Hao Chen, Kai Qiu, Jason Kuen, Jiumiang Gu, Bhiksha Raj, and Zhe Lin. Imagefolder: Autoregressive image generation with folded tokens. arXiv preprint arXiv:2410.01756, 2024. 1, 2, 4, 5, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 410, + 295, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 410, + 295, + 463 + ], + "spans": [ + { + "bbox": [ + 62, + 410, + 295, + 463 + ], + "type": "text", + "content": "[63] Yanwei Li, Yuechen Zhang, Chengyao Wang, Zhisheng Zhong, Yixin Chen, Ruihang Chu, Shaoteng Liu, and Jiaya Jia. Mini-gemini: Mining the potential of multi-modality vision language models. arXiv preprint arXiv:2403.18814, 2024. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 465, + 295, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 465, + 295, + 520 + ], + "spans": [ + { + "bbox": [ + 62, + 465, + 295, + 520 + ], + "type": "text", + "content": "[64] Ziyi Lin, Chris Liu, Renrui Zhang, Peng Gao, Longtian Qiu, Han Xiao, Han Qiu, Chen Lin, Wenqi Shao, Keqin Chen, et al. *Sphinx: The joint mixing of weights, tasks, and visual embeddings for multi-modal large language models.* arXiv preprint arXiv:2311.07575, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 521, + 295, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 521, + 295, + 554 + ], + "spans": [ + { + "bbox": [ + 62, + 521, + 295, + 554 + ], + "type": "text", + "content": "[65] Yaron Lipman, Ricky TQ Chen, Heli Ben-Hamu, Maximilian Nickel, and Matt Le. Flow matching for generative modeling. arXiv preprint arXiv:2210.02747, 2022. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 555, + 295, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 555, + 295, + 599 + ], + "spans": [ + { + "bbox": [ + 62, + 555, + 295, + 599 + ], + "type": "text", + "content": "[66] Wenze Liu, Le Zhuo, Yi Xin, Sheng Xia, Peng Gao, and Xiangyu Yue. Customize your visual autoregressive recipe with set autoregressive modeling. arXiv preprint arXiv:2410.10511, 2024. 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 62, + 600, + 295, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 600, + 295, + 632 + ], + "spans": [ + { + "bbox": [ + 62, + 600, + 295, + 632 + ], + "type": "text", + "content": "[67] Stuart Lloyd. Least squares quantization in pmc. IEEE transactions on information theory, 28(2):129-137, 1982. 2, 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 62, + 635, + 295, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 635, + 295, + 656 + ], + "spans": [ + { + "bbox": [ + 62, + 635, + 295, + 656 + ], + "type": "text", + "content": "[68] I Loshchilov. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 62, + 658, + 295, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 658, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 62, + 658, + 295, + 689 + ], + "type": "text", + "content": "[69] Cheng Lu and Yang Song. Simplifying, stabilizing and scaling continuous-time consistency models. arXiv preprint arXiv:2410.11081, 2024. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 62, + 691, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 691, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 62, + 691, + 295, + 713 + ], + "type": "text", + "content": "[70] Gen Luo, Yiyi Zhou, Yuxin Zhang, Xiawu Zheng, Xiaoshuai Sun, and Rongrong Ji. Feast your eyes: Mixture-of" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 320, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 339, + 73, + 555, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 73, + 555, + 95 + ], + "spans": [ + { + "bbox": [ + 339, + 73, + 555, + 95 + ], + "type": "text", + "content": "resolution adaptation for multimodal large language models. arXiv preprint arXiv:2403.03003, 2024.1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 321, + 96, + 555, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 96, + 555, + 139 + ], + "spans": [ + { + "bbox": [ + 321, + 96, + 555, + 139 + ], + "type": "text", + "content": "[71] Zhuoyan Luo, Fengyuan Shi, Yixiao Ge, Yujiu Yang, Limin Wang, and Ying Shan. Open-magtvit2: An open-source project toward democratizing auto-regressive visual generation. arXiv preprint arXiv:2409.04410, 2024. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 320, + 140, + 554, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 140, + 554, + 194 + ], + "spans": [ + { + "bbox": [ + 320, + 140, + 554, + 194 + ], + "type": "text", + "content": "[72] Nanye Ma, Mark Goldstein, Michael S Albergo, Nicholas M Boffi, Eric Vanden-Eijnden, and Saining Xie. Sit: Exploring flow and diffusion-based generative models with scalable interpolant transformers. arXiv preprint arXiv:2401.08740, 2024. 1, 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 320, + 195, + 554, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 195, + 554, + 227 + ], + "spans": [ + { + "bbox": [ + 320, + 195, + 554, + 227 + ], + "type": "text", + "content": "[73] Fabian Mentzer, David Minnen, Eirikur Agustsson, and Michael Tschannen. Finite scalar quantization: Vq-vae made simple, 2023. 2, 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 320, + 228, + 554, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 228, + 554, + 250 + ], + "spans": [ + { + "bbox": [ + 320, + 228, + 554, + 250 + ], + "type": "text", + "content": "[74] MidJourney Inc. Midjourney. Software available from MidJourney Inc., 2022. 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 320, + 251, + 554, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 251, + 554, + 272 + ], + "spans": [ + { + "bbox": [ + 320, + 251, + 554, + 272 + ], + "type": "text", + "content": "[75] Alex Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models, 2021. 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 320, + 273, + 555, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 273, + 555, + 327 + ], + "spans": [ + { + "bbox": [ + 320, + 273, + 555, + 327 + ], + "type": "text", + "content": "[76] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021.8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 320, + 327, + 554, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 327, + 554, + 360 + ], + "spans": [ + { + "bbox": [ + 320, + 327, + 554, + 360 + ], + "type": "text", + "content": "[77] Harry Nyquist. Certain topics in telegraph transmission theory. Transactions of the American Institute of Electrical Engineers, 47(2):617-644, 1928. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 320, + 361, + 555, + 382 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 361, + 555, + 382 + ], + "spans": [ + { + "bbox": [ + 320, + 361, + 555, + 382 + ], + "type": "text", + "content": "[78] OpenAI. Video generation models as world simulators. OpenAI Blog, 2024. 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 320, + 383, + 555, + 481 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 383, + 555, + 481 + ], + "spans": [ + { + "bbox": [ + 320, + 383, + 555, + 481 + ], + "type": "text", + "content": "[79] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Russell Howes, Po-Yao Huang, Hu Xu, Vasu Sharma, Shang-Wen Li, Wojciech Galuba, Mike Rabbat, Mido Assran, Nicolas Ballas, Gabriel Synnaeve, Ishan Misra, Herve Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. Dinov2: Learning robust visual features without supervision, 2023. 5, 7, 8, 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 320, + 482, + 554, + 504 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 482, + 554, + 504 + ], + "spans": [ + { + "bbox": [ + 320, + 482, + 554, + 504 + ], + "type": "text", + "content": "[80] William Peebles and Saining Xie. Scalable diffusion models with transformers, 2023. 1, 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 320, + 505, + 555, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 505, + 555, + 557 + ], + "spans": [ + { + "bbox": [ + 320, + 505, + 555, + 557 + ], + "type": "text", + "content": "[81] Vadim Popov, Ivan Vovk, Vladimir Gogoryan, Tasnima Sadekova, and Mikhail Kudinov. Grad-tts: A diffusion probabilistic model for text-to-speech. In International Conference on Machine Learning, pages 8599-8608. PMLR, 2021. 1" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 320, + 559, + 555, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 559, + 555, + 624 + ], + "spans": [ + { + "bbox": [ + 320, + 559, + 555, + 624 + ], + "type": "text", + "content": "[82] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 7, 8" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 320, + 625, + 555, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 625, + 555, + 678 + ], + "spans": [ + { + "bbox": [ + 320, + 625, + 555, + 678 + ], + "type": "text", + "content": "[83] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International conference on machine learning, pages 8821-8831. Pmlr, 2021. 1, 8" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 320, + 680, + 555, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 680, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 320, + 680, + 555, + 713 + ], + "type": "text", + "content": "[84] Ali Razavi, Aaron Van den Oord, and Oriol Vinyals. Generating diverse high-fidelity images with vq-vae-2. Advances in neural information processing systems, 32, 2019. 1, 8" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28368" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 57, + 73, + 297, + 712 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 61, + 73, + 297, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 73, + 297, + 105 + ], + "spans": [ + { + "bbox": [ + 61, + 73, + 297, + 105 + ], + "type": "text", + "content": "[85] Ali Razavi, Aaron van den Oord, and Oriol Vinyals. Generating diverse high-fidelity images with vq-vae-2, 2019. 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 106, + 296, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 106, + 296, + 129 + ], + "spans": [ + { + "bbox": [ + 61, + 106, + 296, + 129 + ], + "type": "text", + "content": "[86] Douglas A Reynolds et al. Gaussian mixture models. Encyclopedia of biometrics, 741(659-663), 2009. 2, 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 129, + 296, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 129, + 296, + 183 + ], + "spans": [ + { + "bbox": [ + 62, + 129, + 296, + 183 + ], + "type": "text", + "content": "[87] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 184, + 296, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 184, + 296, + 216 + ], + "spans": [ + { + "bbox": [ + 62, + 184, + 296, + 216 + ], + "type": "text", + "content": "[88] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models, 2022. 1, 6, 8, 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 217, + 296, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 217, + 296, + 282 + ], + "spans": [ + { + "bbox": [ + 62, + 217, + 296, + 282 + ], + "type": "text", + "content": "[89] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in neural information processing systems, 35:36479-36494, 2022. 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 283, + 296, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 283, + 296, + 326 + ], + "spans": [ + { + "bbox": [ + 62, + 283, + 296, + 326 + ], + "type": "text", + "content": "[90] Tim Salimans, Ian Goodfellow, Wojciech Zaremba, Vicki Cheung, Alec Radford, and Xi Chen. Improved techniques for training gans. Advances in Neural Information Processing Systems, 29, 2016. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 327, + 294, + 349 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 327, + 294, + 349 + ], + "spans": [ + { + "bbox": [ + 62, + 327, + 294, + 349 + ], + "type": "text", + "content": "[91] Claude Elwood Shannon. Communication in the presence of noise. Proceedings of the IRE, 37(1):10-21, 1949. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 350, + 296, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 350, + 296, + 403 + ], + "spans": [ + { + "bbox": [ + 62, + 350, + 296, + 403 + ], + "type": "text", + "content": "[92] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International conference on machine learning, pages 2256-2265. PMLR, 2015. 3, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 404, + 296, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 404, + 296, + 437 + ], + "spans": [ + { + "bbox": [ + 62, + 404, + 296, + 437 + ], + "type": "text", + "content": "[93] Jascha Sohl-Dickstein, Eric A. Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics, 2015. 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 437, + 296, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 437, + 296, + 480 + ], + "spans": [ + { + "bbox": [ + 62, + 437, + 296, + 480 + ], + "type": "text", + "content": "[94] Dingjie Song, Wenjun Wang, Shunian Chen, Xidong Wang, Michael Guan, and Benyou Wang. Less is more: A simple yet effective token reduction method for efficient multimodal llms. arXiv preprint arXiv:2409.10994, 2024. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 481, + 296, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 481, + 296, + 503 + ], + "spans": [ + { + "bbox": [ + 62, + 481, + 296, + 503 + ], + "type": "text", + "content": "[95] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models, 2022. 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 504, + 286, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 504, + 286, + 514 + ], + "spans": [ + { + "bbox": [ + 62, + 504, + 286, + 514 + ], + "type": "text", + "content": "[96] stabilityai. Sd vae ft ema, 2023. Accessed: 2023. 1, 3, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 62, + 515, + 296, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 515, + 296, + 557 + ], + "spans": [ + { + "bbox": [ + 62, + 515, + 296, + 557 + ], + "type": "text", + "content": "[97] Jianlin Su, Murtadha Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu. Roformer: Enhanced transformer with rotary position embedding. Neurocomputing, 568:127063, 2024. 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 62, + 559, + 296, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 559, + 296, + 602 + ], + "spans": [ + { + "bbox": [ + 62, + 559, + 296, + 602 + ], + "type": "text", + "content": "[98] Peize Sun, Yi Jiang, Shoufa Chen, Shilong Zhang, Bingyue Peng, Ping Luo, and Zehuan Yuan. Autoregressive model beats diffusion: Llama for scalable image generation. arXiv preprint arXiv:2406.06525, 2024. 1, 5, 6, 8" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 62, + 603, + 296, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 603, + 296, + 635 + ], + "spans": [ + { + "bbox": [ + 62, + 603, + 296, + 635 + ], + "type": "text", + "content": "[99] Keyu Tian, Yi Jiang, Zehuan Yuan, Bingyue Peng, and Liwei Wang. Visual autoregressive modeling: Scalable image generation via next-scale prediction, 2024. 3, 5, 8, 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 637, + 296, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 637, + 296, + 680 + ], + "spans": [ + { + "bbox": [ + 57, + 637, + 296, + 680 + ], + "type": "text", + "content": "[100] Michael Tschannen, Cian Eastwood, and Fabian Mentzer. Givt: Generative infinite-vocabulary transformers. In European Conference on Computer Vision, pages 292-309. Springer, 2025. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 57, + 681, + 296, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 681, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 681, + 296, + 712 + ], + "type": "text", + "content": "[101] Hung-Yu Tseng, Lu Jiang, Ce Liu, Ming-Hsuan Yang, and Weilong Yang. Regularizing generative adversarial networks under limited data, 2021. 5, 2" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 57, + 73, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 316, + 73, + 554, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 95 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 95 + ], + "type": "text", + "content": "[102] Arash Vahdat, Karsten Kreis, and Jan Kautz. Score-based generative modeling in latent space, 2021. 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 96, + 555, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 96, + 555, + 139 + ], + "spans": [ + { + "bbox": [ + 317, + 96, + 555, + 139 + ], + "type": "text", + "content": "[103] Aaron Van den Oord, Nal Kalchbrenner, Lasse Espeholt, Oriol Vinyals, Alex Graves, et al. Conditional image generation with pixelCNN decoders. Advances in neural information processing systems, 29, 2016. 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 140, + 554, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 140, + 554, + 172 + ], + "spans": [ + { + "bbox": [ + 317, + 140, + 554, + 172 + ], + "type": "text", + "content": "[104] Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017. 1, 2, 3, 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 173, + 555, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 173, + 555, + 205 + ], + "spans": [ + { + "bbox": [ + 317, + 173, + 555, + 205 + ], + "type": "text", + "content": "[105] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need, 2023. 1, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 206, + 555, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 206, + 555, + 261 + ], + "spans": [ + { + "bbox": [ + 317, + 206, + 555, + 261 + ], + "type": "text", + "content": "[106] Pascal Vincent, Hugo Larochelle, Yoshua Bengio, and Pierre-Antoine Manzagol. Extracting and composing robust features with denoising autoencoders. In Proceedings of the 25th international conference on Machine learning, pages 1096-1103, 2008. 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 262, + 554, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 262, + 554, + 304 + ], + "spans": [ + { + "bbox": [ + 316, + 262, + 554, + 304 + ], + "type": "text", + "content": "[107] Xinlong Wang, Xiaosong Zhang, Zhengxiong Luo, Quan Sun, Yufeng Cui, Jinsheng Wang, Fan Zhang, Yueze Wang, Zhen Li, Qiying Yu, et al. Emu3: Next-token prediction is all you need. arXiv preprint arXiv:2409.18869, 2024. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 305, + 554, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 305, + 554, + 350 + ], + "spans": [ + { + "bbox": [ + 316, + 305, + 554, + 350 + ], + "type": "text", + "content": "[108] Mark Weber, Lijun Yu, Qihang Yu, Xueqing Deng, Xiaohui Shen, Daniel Cremers, and Liang-Chieh Chen. Maskbit: Embedding-free image generation via bit tokens. arXiv preprint arXiv:2409.16211, 2024. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 350, + 554, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 350, + 554, + 404 + ], + "spans": [ + { + "bbox": [ + 316, + 350, + 554, + 404 + ], + "type": "text", + "content": "[109] Yecheng Wu, Zhuoyang Zhang, Junyu Chen, Haotian Tang, Dacheng Li, Yunhao Fang, Ligeng Zhu, Enze Xie, Hongxu Yin, Li Yi, et al. Vila-u: a unified foundation model integrating visual understanding and generation. arXiv preprint arXiv:2409.04429, 2024. 1, 2, 4, 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 404, + 554, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 404, + 554, + 459 + ], + "spans": [ + { + "bbox": [ + 316, + 404, + 554, + 459 + ], + "type": "text", + "content": "[110] Jinheng Xie, Weijia Mao, Zechen Bai, David Junhao Zhang, Weihao Wang, Kevin Qinghong Lin, Yuchao Gu, Zhijie Chen, Zhenheng Yang, and Mike Zheng Shou. Show-o: One single transformer to unify multimodal understanding and generation. arXiv preprint arXiv:2408.12528, 2024. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 460, + 555, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 460, + 555, + 514 + ], + "spans": [ + { + "bbox": [ + 317, + 460, + 555, + 514 + ], + "type": "text", + "content": "[111] Jiahui Yu, Xin Li, Jing Yu Koh, Han Zhang, Ruoming Pang, James Qin, Alexander Ku, Yuanzhong Xu, Jason Baldridge, and Yonghui Wu. Vector-quantized image modeling with improved vqgan. arXiv preprint arXiv:2110.04627, 2021. 2, 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 514, + 555, + 570 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 514, + 555, + 570 + ], + "spans": [ + { + "bbox": [ + 317, + 514, + 555, + 570 + ], + "type": "text", + "content": "[112] Lijun Yu, José Lezama, Nitesh B Gundavarapu, Luca Versari, Kihyuk Sohn, David Minnen, Yong Cheng, Vighnesh Birodkar, Agrim Gupta, Xiuye Gu, et al. Language model beats diffusion-tokenizer is key to visual generation. arXiv preprint arXiv:2310.05737, 2023. 1, 2, 8" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 317, + 571, + 555, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 571, + 555, + 635 + ], + "spans": [ + { + "bbox": [ + 317, + 571, + 555, + 635 + ], + "type": "text", + "content": "[113] Lijun Yu, Yong Cheng, Zhiruo Wang, Vivek Kumar, Wolfgang Macherey, Yanping Huang, David Ross, Irfan Essa, Yonatan Bisk, Ming-Hsuan Yang, et al. Spae: Semantic pyramid autoencoder for multimodal generation with frozen llms. Advances in Neural Information Processing Systems, 36, 2024. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 636, + 555, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 636, + 555, + 669 + ], + "spans": [ + { + "bbox": [ + 316, + 636, + 555, + 669 + ], + "type": "text", + "content": "[114] Qihang Yu, Ju He, Xueqing Deng, Xiaohui Shen, and Liang-Chieh Chen. Randomized autoregressive visual generation. arXiv preprint arXiv:2411.00776, 2024. 8" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 670, + 555, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 555, + 712 + ], + "type": "text", + "content": "[115] Qihang Yu, Mark Weber, Xueqing Deng, Xiaohui Shen, Daniel Cremers, and Liang-Chieh Chen. An image is worth 32 tokens for reconstruction and generation. arxiv: 2406.07550, 2024. 1, 2, 4, 6, 7, 8, 5" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "28369" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 554 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 297, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 297, + 127 + ], + "type": "text", + "content": "[116] Sihyun Yu, Sangkyung Kwak, Huiwon Jang, Jongheon Jeong, Jonathan Huang, Jinwoo Shin, and Saining Xie. Representation alignment for generation: Training diffusion transformers is easier than you think. arXiv preprint arXiv:2410.06940, 2024. 1, 2, 4, 5, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 128, + 296, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 128, + 296, + 173 + ], + "spans": [ + { + "bbox": [ + 56, + 128, + 296, + 173 + ], + "type": "text", + "content": "[117] Neil Zeghidour, Alejandro Luebs, Ahmed Omran, Jan Skoglund, and Marco Tagliasacchi. Soundstream: An end-to-end neural audio codec. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 30:495-507, 2021. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 174, + 296, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 174, + 296, + 239 + ], + "spans": [ + { + "bbox": [ + 56, + 174, + 296, + 239 + ], + "type": "text", + "content": "[118] Baoquan Zhang, Huaibin Wang, Chuyao Luo, Xutao Li, Guotao Liang, Yunming Ye, Xiaochen Qi, and Yao He. Codebook transfer with part-of-speech for vector-quantized image modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7757–7766, 2024. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 239, + 296, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 239, + 296, + 274 + ], + "spans": [ + { + "bbox": [ + 56, + 239, + 296, + 274 + ], + "type": "text", + "content": "[119] Han Zhang, Zizhao Zhang, Augustus Odena, and Honglak Lee. Consistency regularization for generative adversarial networks. arXiv preprint arXiv:1910.12027, 2019. 5, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 274, + 296, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 274, + 296, + 308 + ], + "spans": [ + { + "bbox": [ + 56, + 274, + 296, + 308 + ], + "type": "text", + "content": "[120] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric, 2018. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 308, + 296, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 308, + 296, + 351 + ], + "spans": [ + { + "bbox": [ + 56, + 308, + 296, + 351 + ], + "type": "text", + "content": "[121] Shengyu Zhao, Zhijian Liu, Ji Lin, Jun-Yan Zhu, and Song Han. Differentiable augmentation for data-efficient gan training. Advances in neural information processing systems, 33:7559-7570, 2020. 5, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 353, + 296, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 353, + 296, + 386 + ], + "spans": [ + { + "bbox": [ + 56, + 353, + 296, + 386 + ], + "type": "text", + "content": "[122] Lin Zheng, Jianbo Yuan, Lei Yu, and Lingpeng Kong. A reparameterized discrete diffusion model for text generation. arXiv preprint arXiv:2302.05737, 2023. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 387, + 296, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 387, + 296, + 441 + ], + "spans": [ + { + "bbox": [ + 56, + 387, + 296, + 441 + ], + "type": "text", + "content": "[123] Chunting Zhou, Lili Yu, Arun Babu, Kushal Tirumala, Michihiro Yasunaga, Leonid Shamis, Jacob Kahn, Xuezhe Ma, Luke Zettlemoyer, and Omer Levy. Transfusion: Predict the next token and diffuse images with one multi-modal model. arXiv preprint arXiv:2408.11039, 2024. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 442, + 296, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 442, + 296, + 475 + ], + "spans": [ + { + "bbox": [ + 56, + 442, + 296, + 475 + ], + "type": "text", + "content": "[124] Lei Zhu, Fangyun Wei, Yanye Lu, and Dong Chen. Scaling the codebook size of vqgan to 100,000 with a utilization rate of " + }, + { + "bbox": [ + 56, + 442, + 296, + 475 + ], + "type": "inline_equation", + "content": "99\\%" + }, + { + "bbox": [ + 56, + 442, + 296, + 475 + ], + "type": "text", + "content": ". arXiv preprint arXiv:2406.11837, 2024. 2, 4, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 475, + 296, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 475, + 296, + 518 + ], + "spans": [ + { + "bbox": [ + 56, + 475, + 296, + 518 + ], + "type": "text", + "content": "[125] X Zhu, W Su, L Lu, B Li, X Wang, and J Dai. Deformable detr: Deformable transformers for end-to-end object detection. arxiv 2020. arXiv preprint arXiv:2010.04159, 2010.8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 520, + 296, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 520, + 296, + 554 + ], + "spans": [ + { + "bbox": [ + 56, + 520, + 296, + 554 + ], + "type": "text", + "content": "[126] Yongxin Zhu, Bocheng Li, Yifei Xin, and Linli Xu. Addressing representation collapse in vector quantized models with one linear layer. arXiv preprint arXiv:2411.02038, 2024. 2" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "28370" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Solving Instance Detection from an Open-World Perspective/ab6d1961-27f9-4a61-af19-064b80505b12_content_list.json b/2025/Solving Instance Detection from an Open-World Perspective/ab6d1961-27f9-4a61-af19-064b80505b12_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..c429e5800b337e0fef3333c626af9c0a2b5c5a85 --- /dev/null +++ b/2025/Solving Instance Detection from an Open-World Perspective/ab6d1961-27f9-4a61-af19-064b80505b12_content_list.json @@ -0,0 +1,1796 @@ +[ + { + "type": "text", + "text": "Solving Instance Detection from an Open-World Perspective", + "text_level": 1, + "bbox": [ + 194, + 130, + 803, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Qianqian Shen $^{1,*}$ Yunhan Zhao $^{2,*}$ Nahyun Kwon $^{3}$ Jeeun Kim $^{3}$ Yanan Li $^{4,\\dagger}$ Shu Kong $^{5,6,\\dagger}$", + "bbox": [ + 111, + 178, + 883, + 200 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1Zhejiang University, 2UC Irvine, 3Texas A&M University, 4Zhejiang Lab", + "bbox": [ + 276, + 200, + 720, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "5University of Macau, 6Institute of Collaborative Innovation", + "bbox": [ + 316, + 218, + 674, + 232 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "website and code: https://shenqq377.github.io/IDOW", + "bbox": [ + 297, + 237, + 694, + 250 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 286, + 326, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Instance detection (InsDet) aims to localize specific object instances within a novel scene imagery based on given visual references. Technically, it requires proposal detection to identify all possible object instances, followed by instance-level matching to pinpoint the ones of interest. Its open-world nature supports its broad applications from robotics to AR/VR but also presents significant challenges: methods must generalize to unknown testing data distributions because (1) the testing scene imagery is unseen during training, and (2) there are domain gaps between visual references and detected proposals. Existing methods tackle these challenges by synthesizing diverse training examples or utilizing off-the-shelf foundation models (FMs). However, they only partially capitalize the available open-world information. In contrast, we approach InsDet from an Open-World perspective, introducing our method IDOW. We find that, while pretrained FMs yield high recall in instance detection, they are not specifically optimized for instance-level feature matching. Therefore, we adapt pretrained FMs for improved instance-level matching using open-world data. Our approach incorporates metric learning along with novel data augmentations, which sample distractors as negative examples and synthesize novel-view instances to enrich the visual references. Extensive experiments demonstrate that our method significantly outperforms prior works, achieving $>10$ AP over previous results on two recently released challenging benchmark datasets in both conventional and novel instance detection settings.", + "bbox": [ + 86, + 319, + 488, + 743 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 772, + 220, + 787 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Instance detection (InsDet) aims to localize object instances of interest in novel scene imagery (Fig. 1), where these objects are specified by some visual references (as known as support templates) [3, 26, 34, 48, 53]. Its open-world nature supports its wide-ranging applications", + "bbox": [ + 89, + 799, + 485, + 876 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "visual references", + "bbox": [ + 529, + 285, + 632, + 297 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4dbdf2596cf884036f25bbb68bc704f0e6a57fb09760fcef5952eb6840d1de6c.jpg", + "image_caption": [ + "Figure 1. Instance Detection (InsDet) aims to localize specific object instances in novel scene imagery based on given visual references (aka support templates). It is a challenging problem due to its open-world nature: the testing scene images are unseen during training and thus are unknown to InsDet models, and visual references and detected proposals have domain gaps (e.g., due to occlusions and arbitrary lighting conditions in the latter)." + ], + "image_footnote": [], + "bbox": [ + 517, + 299, + 643, + 411 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "novel testing scene imagery", + "bbox": [ + 661, + 285, + 828, + 297 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e13db6646b9f7799e8a6d6bedf3d1c112968821e5e898db2ea44acdb8be3f68c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 648, + 299, + 844, + 411 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/939f085b0a352c97ec1ca4bcaeece617fe2eb30f6cba3f06694e04dc3c815409.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 844, + 297, + 897, + 410 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "in robotics and AR/VR. For example, a robot can be commanded to search for a customer's suitcase at an airport [26], or my_keys in my cluttered bedroom [62]. Its open-world nature also presents significant challenges: during training, one has no knowledge of testing data distribution, which can be instantiated by the unknown scene imagery [9], novel object instances encountered only in testing [26], and domain gaps between visual references and detected proposals [48].", + "bbox": [ + 509, + 521, + 908, + 643 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Status quo. The literature of InsDet has two settings and existing InsDet methods attempt to address the open world in different aspects (Fig. 2). The conventional instance detection (CID) setting provides visual references of object instances for training. Yet, the testing scene is still unknown in training time. Therefore, early method Cut Paste Learn (CPL) collects random background images to model diverse testing scenes (Fig. 2a), cuts instances from the visual references, and pastes them on these background images to construct a synthetic training set [9, 11]. It is worth noting that background image sampling is utilizing open-world information in a way of data collection in the open world. On the other hand, the novel instance detection (NID) setting, introduced recently [26], studies detecting novel object instances that are specified only in testing time where a trained model is not allowed to be tuned further. To approach NID, VoxDet [26] uses large-scale synthetic data (Fig. 2b) to learn", + "bbox": [ + 511, + 643, + 910, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contributions; †Corresponding authors.", + "bbox": [ + 107, + 886, + 357, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "9901", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/fea24deb0d5969838533254bf5891ec288951e87511a7e67116d10d516cd3aaf.jpg", + "image_caption": [ + "(a) background imagery [9, 11]", + "Figure 2. Existing InsDet methods leverage the open-world information in different aspects, such as (a) background image sampling (from the open world) to synthesize training data [9, 11], (b) object image sampling (from the open world) to learn feature representations [26], and (c) foundation model utilization (pretrained in the open world) for proposal detection and instance-level feature matching [48]. (d) As FMs are not specifically designed for instance-level feature matching required by InsDet, we propose to adapt them by leveraging rich data sampled from the open world. We gather data from multiple sources: (1) any available visual references of instances in the CID setting, (2) abundant multi-view object images sampled in the open world similar to (b), (3) synthetic data by training NeRF [1, 36] to generate novel-view images based on the given instances, (4) distractors by running FMs (esp. SAM [22]) on random open-world imagery to generate random object-like proposals (Fig. 5). We use the data above to adapt FM through metric learning. The technical novelty of our work lies in (3) and (4), as well as the design choice of metric learning to adapt FMs for InsDet." + ], + "image_footnote": [], + "bbox": [ + 98, + 98, + 243, + 200 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/a82e0ac888d6ea996040eb2e268f525d5594eed2795a249564d9a7eff4ddf182.jpg", + "image_caption": [ + "(b) object images [26]", + "(c) Foundation Models [48]" + ], + "image_footnote": [], + "bbox": [ + 251, + 98, + 370, + 200 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/05b8064b4eee9ec5218df50a58fa07d35e8e2b29b97a29926a4cbfdc025b5ffb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 383, + 98, + 491, + 200 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/b8e2450a827beaff13567c6ac5add694cbb0ca36efe16ea321df09abb650aa07.jpg", + "image_caption": [ + "(d) Solving Instance Detection in the Open World (ours)" + ], + "image_footnote": [], + "bbox": [ + 506, + 98, + 893, + 200 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3D voxel representations from visual references and a matching function between 2D proposals and 3D voxels. Recent work OTS-FM [48] uses off-the-shelf pretrained foundation models (Fig. 2c) for InsDet. Specifically, it uses SAM [22] to detect object proposals in scene imagery, and DINOv2 [38] to represent visual references and proposals for instance-level feature matching. OTS-FM is a non-learned method and hence is applicable in both CID and NID settings. Importantly, it achieves the state-of-the-art performance on the recently released benchmark HR-InsDet [48].", + "bbox": [ + 88, + 340, + 483, + 491 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The open-world perspective. As illustrated in Fig. 1, the challenges of InsDet lie in its open-world nature: the unknown testing data distribution, and domain gaps between visual references and detected proposals. To address the open-world challenges, existing methods take different strategies to exploit the open-world information (Fig. 2a-c). For example, to address the unknown testing scene imagery, some methods sample diverse background images [9, 11], demonstrating a practice of sampling data in the open world. To mitigate the domain gap between detected proposals and visual references, [26] attempts to learn features for instance-level matching by utilizing external datasets (ShapeNet [5] and ABO [8]). This data utilization can be considered as data collection in the open world. For better proposal detection and more discriminative feature representations, [48] adopts off-the-shelf pretrained foundation models (FMs) SAM [22] and DINOv2 [38], respectively. One can think of FMs as being pretrained in the open world. Different from existing methods which have not fully capitalized the open-world information, we take the open-world perspective as the first principle and solve InsDet by exploiting the Open World (Fig. 2d). We call our approach IDOW.", + "bbox": [ + 91, + 492, + 483, + 824 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Insights and novelties. Our insights are rooted in the understanding of InsDet's open-world nature and the practice of existing methods. Existing methods address open-world issues to some extent (Fig. 2), particularly by sampling data in the open world [9, 23, 26] and adopting foundation mod", + "bbox": [ + 89, + 825, + 483, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "els (FMs) pretrained in the open world [48]. These motivate us to develop InsDet methods in the open world (IDOW). We find that open-world detectors yield high recall on object instances (see precision-recall curves in Fig. 9 of the Appendix) but FMs are not tailored to instance-level matching between proposals and visual references – directly using an FM for InsDet is sub-optimal. Therefore, we adapt FMs to instance-level feature representation by metric learning. We propose novel and simple techniques for data augmentation to enhance FM adaptation and feature learning: distractor sampling, and multi-view synthesis. The former helps learn more discriminative features for instance comparison, and the latter enhances visual references particularly when they are few. With these techniques, our IDOW achieves significant boosts over previously reported results on recently released challenging benchmarks (Table 1 and 2).", + "bbox": [ + 511, + 340, + 906, + 582 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Contributions. We make three major contributions:", + "text_level": 1, + "bbox": [ + 532, + 583, + 877, + 597 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. We elaborate on the open-world challenge of InsDet and motivate the necessity of developing methods in the open world. We solve InsDet by exploiting diverse open data available in the open world and FMs pretrained therein.", + "2. We introduce simple and effective techniques to adapt FMs to InsDet: a metric learning loss, and data augmentation with distractor sampling and novel-view synthesis.", + "3. We extensively compare our approach with existing ones on two recently released datasets, demonstrating that our approach outperforms them by $>10$ AP in both conventional and novel instance detection settings." + ], + "bbox": [ + 513, + 599, + 906, + 765 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 782, + 653, + 797 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Instance detection. Early methods use local features such as SIFT [31] and SURF [2] to match visual references [43] and image regions to localize instances [16]. Recent methods train deep neural networks for instance-agnostic proposal detection and proposal-instance matching, achieving significant improvements [26, 34, 48]. In the con", + "bbox": [ + 511, + 810, + 908, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "9902", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ventional instance detection (CID) setting, visual references of object instances are given during training. To approach CID, prevalent method synthesize training data in a cut-paste-learn (CPL) strategy [9, 11]: cutting instances from visual references and pasting them on random background images, then learning a detector on such synthetic data. Other works strive to obtain more training samples by rendering realistic instance examples [17, 19], using data augmentation [9] and synthesizing better training images [9, 12, 25]. Moreover, recent literature introduces the novel instance detection (NID) setting [26], which requires detecting novel instances specified only in testing and does not allow for further model finetuning during testing. VoxDet [26] utilizes a pretrained open-world detector [20] for proposal detection. Nowadays, the open-world perspective advances proposal detection in a way of learning more generalizable detectors and carrying out open-vocabulary recognition [29]. While a foundational open-vocabulary detector can attach a textual description (e.g., a short phrase) to each proposal, simply using it is insufficient for instance-level matching. [48] realizes this issue although foundational detectors can yield nearly perfect recall. As a result, the recent method OTS-FM uses another FM DINOv2 [38] for proposal-instance matching [48]. Differently, we solve InsDet from the open-world perspective, exploiting open FMs and open data to learn more discriminative instance-level feature representations.", + "bbox": [ + 93, + 90, + 483, + 483 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Foundation models (FMs) are pre-trained on web-scale data, which can be thought of as being sampled in the open world. Different FMs are trained on different types of data and can shine in different downstream tasks related to visual perception, natural language processing, or both. As this work concerns about visual perception, we briefly review Vision-Language Models (VLMs) and Visual Foundation Models (VFMs). VLMs are pretrained on web-scale image-text paired data [18, 28, 29, 44], demonstrating impressive results in high-level tasks such as visual grounding and image captioning. VFMs are pretrained primarily on visual data [4, 7, 14, 22, 38, 55, 58, 63] and can yield impressive visual perception results such as proposal or open-set object detection [29, 59]. In InsDet, the mainstream framework requires proposal detection and visual feature matching. In this framework, open-world detectors appear to achieve high recall on the instances of interest, owing to that the openworld pretrained detectors generalize quite well in the open world [29, 46, 59]. In contrast, feature matching still remains a challenge in improving InsDet. Therefore, we focus on improving features for InsDet by adapting VFMs using data sampled in the open world. [48] shows that using an off-the-shelf FM DINOv2 [38] for feature matching greatly enhances InsDet detection; yet, FM adaptation has been still under-explored in InsDet, although it is a well studied in other areas such as few-shot perception [30, 32, 51] and zero-shot perception [13, 40, 41]. Our work introduces simple", + "bbox": [ + 93, + 493, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "and novel techniques to adapt FMs to improve InsDet.", + "bbox": [ + 516, + 92, + 870, + 106 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The open-world data is publicly available for pretraining FMs. Coping with such data is also often challenging. For example, data between training and testing has distribution shifts or domain gaps [30, 49], as clearly instantiated in InsDet by the unknown testing scene imagery, in addition to the domain gaps between (clean) visual references and (occluded) proposals (Fig. 1). Speaking of distribution shifts, some existing methods sample open-world data, demonstrating a need to develop robust models in the open world [15, 23]. Existing InsDet methods also sample open-world data such as diverse background images and instance examples (Fig. 2) [9, 11]. Even so, while open-world data is abundant, most of it is not directly useful for training models for specific downstream tasks. In contrast, data augmentation aims to synthesize more tailored data to enhance models' robustness and generalization, including geometry-aware augmentation (e.g., random cropping and flipping) [7, 24, 61], photometric-aware augmentation (e.g., color jittering) [21, 56, 57], and robustness-aware augmentation (e.g., adversarial perturbation) [60]. It is worth noting that Neural Radiance Field (NeRF) [35], a recent rendering technique, can be used to synthesize novel-view images. In this work, we particularly adopt NeRF to synthesize novel-view visual references of object instances and use them to adapt FMs towards more tailored features to InsDet. To the best of our knowledge, using NeRF is under-explored in the literature of InsDet, but we find that doing so remarkably improves InsDet performance.", + "bbox": [ + 514, + 109, + 906, + 529 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Instance Detection: Protocols and Methods", + "text_level": 1, + "bbox": [ + 516, + 549, + 895, + 564 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We first introduce the protocols of InsDet, including the problem definition, evaluation, and two settings. We then present our techniques to enhance foundation model adaptation to InsDet.", + "bbox": [ + 514, + 575, + 906, + 633 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Protocols", + "text_level": 1, + "bbox": [ + 516, + 650, + 619, + 664 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Problem definition. Instance Detection (InsDet) aims to detect specific object instances from a 2D scene image (Fig. 1), where the objects of interest are defined by a \"support set\" of some visual references [9, 26, 48]. Previous works capture the visual references from multiple camera views, with QR code being pasted to help estimate camera poses (see visual references in Fig. 1).", + "bbox": [ + 514, + 672, + 906, + 777 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Challenges. InsDet is a challenging problem due to its open-world nature: one has zero knowledge about the testing data distribution which is instantiated by unknown testing scenes, never-before-seen matters, arbitrary occlusions, background clutters, etc. In particular, object instances of interest can be novel and are only defined during testing, as emphasized by the novel instance detection setting below. Nevertheless, InsDet models must be capable of detecting", + "bbox": [ + 514, + 780, + 908, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "9903", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/9acc9f19f1e8fb0dd2d830ffbfda8222b4ad67ec9bb3fecef2c70d7870232a39.jpg", + "image_caption": [ + "Figure 3. Visual comparison of InsDet results by different methods in the CID setting on HR-InsDet [48]. The testing scene image contains sparse placement of small instances with challenging illumination. We mark the ground-truth and predictions using green and blue boxes, respectively. We attach instance IDs to them to highlight whether the instance recognition is correct compared to the visual references (i.e., the leftmost references). Compared with Cut-Paste-Learn and OTS-FM, our IDOW detects more instances (see orange arrows) with better accuracy (see red arrows). Comparison between OTS-FM and IDOW suggests that adapting the FM DINOv2 yields better features for IDOW to perform much better in front of domain shifts, i.e., challenging illumination conditions in this testing scene image." + ], + "image_footnote": [], + "bbox": [ + 89, + 87, + 320, + 261 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/18b7a33130cc1bee7ed8b0187c08c3b38d9123636647441d95516d430266c1bf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 330, + 85, + 513, + 261 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/d735d4105590ba8f0f98a663f6eceb51a9b7c9771272d9a40e666cae9c994d2b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 519, + 85, + 704, + 261 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/2aa413d1b18d69f4f0f452a819509693153e0ed990c2dfc2599396608187c941.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 714, + 85, + 897, + 261 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "any object instances of interest and robustly comparing them with visual references.", + "bbox": [ + 89, + 363, + 482, + 391 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Two settings. There are two settings of InsDet, which differ in whether object instances are pre-defined during training or defined on the fly in testing time.", + "bbox": [ + 89, + 393, + 483, + 438 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Conventional Instance Detection (CID). In this setting, all object instances of interest are pre-defined during training [9]. That said, one can use their visual references as training data to train detectors. Yet, only visual references are available and testing scene data is still unknown. Hence, InsDet methods in CID often choose to sample diverse background images in the open world [9, 12]. This setting simulates application scenarios in AR/VR. For instance, robots should memorize the items concerned by the customer so as to provide better customized services.", + "- Novel Instance Detection (NID). This setting requires InsDet models to detect novel object instances defined only in testing time and does not allow the models to be finetuned further during testing [26]. In this setting, the de facto practice to train InsDet models turns to external data, which can be either synthetic data or those sampled in the open world. This setting simulates some real-world scenarios, for instance, robots must search for a never-before-seen luggage of a customer at an airport." + ], + "bbox": [ + 89, + 439, + 485, + 724 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Learning Instance Detector in Open World", + "text_level": 1, + "bbox": [ + 89, + 733, + 457, + 750 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We present our solution of instance detection from an open-world perspective (IDOW), focusing on foundation model adaptation and data augmentation.", + "bbox": [ + 89, + 755, + 482, + 800 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.1. Foundation Model Adaptation", + "text_level": 1, + "bbox": [ + 89, + 806, + 349, + 821 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As an open-world detector yields quite high recall [26, 48], we focus on adapting a foundation model to enhance feature representations for InsDet. Although a foundational feature model (e.g., DINOv2 [38]) offers feature representations applicable to various downstream tasks, it is not tai-", + "bbox": [ + 89, + 825, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "lored to InsDet. To improve feature representation, one can finetune this FM over relevant data, such as those sampled in the open world [26]. Intuitively, the finetuned feature is expected to match proposals and visual references better if they are from the same object instances; otherwise, distinguish them. Therefore, we propose to use a metric learning loss for FM adaptation. Moreover, we sample distractors from imagery in the open world. The distractors serve as negatives in metric learning.", + "bbox": [ + 511, + 362, + 906, + 498 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We denote an FM as $f_{\\theta}$ , parameterized by $\\theta$ . This FM serves as the feature extractor $f_{\\theta}(\\mathbf{I}) : \\mathbf{I} \\to \\mathbb{R}^q$ to produce $q$ -dimensional features for any given visual reference image or proposal $\\mathbf{I}$ . We finetune $f_{\\theta}(\\cdot)$ to better serve InsDet such that features of reference images from the same instance should present high similarity, otherwise low similarity. We achieve this by finetuning $f_{\\theta}$ using a simple metric learning loss $\\ell$ . Specifically, we construct triplets of training data $(\\mathbf{I}_a, \\mathbf{I}_p, \\mathbf{I}_n)$ , where $\\mathbf{I}_a, \\mathbf{I}_p$ and $\\mathbf{I}_n$ represent an anchor visual reference, positive and negative samples, respectively. The training data contains examples sampled or synthesized in the open world [26], as well as visual references available in the CID setting. Negative examples are cross-instance visual examples and distractors (detailed later). Below is the loss:", + "bbox": [ + 511, + 500, + 908, + 712 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\ell = \\left[ d \\left(f _ {\\theta} \\left(\\mathbf {I} _ {a}\\right), f _ {\\theta} \\left(\\mathbf {I} _ {p}\\right)\\right) - d \\left(f _ {\\theta} \\left(\\mathbf {I} _ {a}\\right), f _ {\\theta} \\left(\\mathbf {I} _ {n}\\right)\\right) + \\alpha \\right] _ {+}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 527, + 727, + 906, + 752 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\alpha$ is a hyper-parameter determining the margin between an instance and other negative data. $d(\\cdot)$ measures the distance between two examples, e.g., inverse cosine similarity used in our work. After training, we use the learned features to represent both proposals and visual references. We follow the instance-proposal matching pipeline [48] to produce the final InsDet results: computing pairwise similarities between each proposal and each visual reference, running stable matching algorithm [10, 33], and returning", + "bbox": [ + 511, + 763, + 908, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "9904", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/c52f8bbd473a77799bf9b4d1de247f976a8f32707f718a714dfe22158720ae25.jpg", + "image_caption": [ + "Support" + ], + "image_footnote": [], + "bbox": [ + 99, + 102, + 186, + 236 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/2f118a273c6ba9e6288fa5a7f90048931795a1486c728c68eaaa8688d9b2040a.jpg", + "image_caption": [ + "Ground Truth" + ], + "image_footnote": [], + "bbox": [ + 187, + 102, + 359, + 236 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/2564597b30890f9ddc99bac46796810c4cfe428baeae9b3880b177c5eacf3e3c.jpg", + "image_caption": [ + "VoxDet [26]", + "Figure 4. Visual comparison of InsDet results by different methods in the NID setting on RoboTools [26]. On the cluttered testing scene image, we mark the ground-truth and predictions using green and blue boxes, respectively. We attach instance IDs to them to highlight whether the instance recognition is correct compared to the visual references (i.e., the leftmost references). Compared with VoxDet and OTS-FM, our IDOW detects more instances (see orange arrows) with better accuracy (see red arrows). Comparing OTS-FM and our IDOW suggests that, although having not seen the visual references during adaption in this NID setting, features by our IDOW are a better representation for InsDet than the off-the-shelf foundation model (i.e., DINOv2) used in OTS-FM." + ], + "image_footnote": [], + "bbox": [ + 367, + 102, + 537, + 236 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/bffecc67dbe4d09cf7afd173469cbe8045cbc77f2aabe89b5312145f9a2c9b2c.jpg", + "image_caption": [ + "OTS-FM [48]" + ], + "image_footnote": [], + "bbox": [ + 545, + 102, + 717, + 236 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/b0ca2c049ef5a9441f7abc088463327a050887f5eed635b016e7553aaa67ecc8.jpg", + "image_caption": [ + "IDOW (ours)" + ], + "image_footnote": [], + "bbox": [ + 725, + 102, + 895, + 236 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "the matched pairs if they have similarity scores greater than a predefined threshold (0.4 as used in [48]).", + "bbox": [ + 89, + 339, + 482, + 368 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Hard example sampling. To construct training batches of triplets, it is important to effectively sample negative examples, easy negative examples do not provide sufficient gradients during training [50]. In this work, we adopt a batch-level hard negative sampling strategy. In each training batch, we first sample a reference image as an anchor $\\mathbf{I}_a$ . We then sample a random example from the same instance as the positive one $\\mathbf{I}_p$ . We finally sample a hard negative example from the union of the reference images of all the other instances and the distractors $S_{dt}$ (explained in the next subsection). Mathematically, we find hard negative patch $\\mathbf{I}_{hn}$ with the objective: $\\arg \\min_{\\mathbf{I}_n} d(f_\\theta(\\mathbf{I}_a), f_\\theta(\\mathbf{I}_n))$ , $\\mathbf{I}_n \\in \\{C(\\mathbf{I}_i) \\neq C(\\mathbf{I}_a)\\}$ , for $i = 1 \\dots N\\} \\cup S_{dt}$ , where $C(\\mathbf{I}_i)$ represents the instance ID of the data $\\mathbf{I}_i$ , and $N$ indicates the total number of visual reference images.", + "bbox": [ + 89, + 369, + 483, + 597 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2.2. The Proposed Data Augmentation", + "text_level": 1, + "bbox": [ + 89, + 608, + 372, + 622 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To better adapt FM for instance-level feature matching, we introduce two data augmentation techniques below.", + "bbox": [ + 89, + 627, + 483, + 657 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Distractor sampling aims to sample patches of random background images as universal negative data to all object instances. It is a data augmentation technique, expected to define the open space and help features better characterize meaningful object instances. Distractor sampling has been adopted in the literature. For example, open-set recognition samples distractors as a separate class, known as \"the-other-class\" or \"background-class\", to define the open space. In InsDet, the cut-paste-learn strategy [9] samples random background photos to define the background content. In this work, we run SAM [22] on random background photos and use the segments as distractors. One might worry about the possibility of sampled distractors being exactly some instances of interest and including them as negatives may destabilize model adaptation. Yet, given a large number of distractors, this rarely becomes an issue as demonstrated by the success", + "bbox": [ + 89, + 659, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/210ab2b74669701f38b34cded9c4adb572b83785ffad4cddd52c715be54d89dd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 516, + 335, + 578, + 386 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/b672221f98d026a64c1900ff47d14121e0810d0dc3354bed182a875306a50e9a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 581, + 335, + 642, + 386 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/cb5287d79ea04599ea5a1ac389160c0e4ee9e34a83bc5b731dc7b0b53db188ad.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 642, + 335, + 705, + 386 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/3de5432c7f80dc5c58ba739f8ab8165307d9e350d64aad10822dea7445f807c0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 705, + 335, + 769, + 386 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/7118f2b44624c755ec6c59899245246ea3aec875a0529b81a7d1fe3512e01956.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 769, + 335, + 834, + 386 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/6c119278d1c3f1d98d9fbef146fc06ae216513b82ae2bbb1e8000e9001163d30.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 834, + 335, + 900, + 386 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/a81afc1f65cb57a2a4b85f6ea9718430dd8d5801eabbd7a001ecea5df7140387.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 517, + 388, + 573, + 436 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/35c1a7d54a5ef6af0836c5d0d601919bc1c7cc9c92900870df6ad555181b8a14.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 584, + 388, + 633, + 436 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/e444383cb9a8750cbc2a30e1af7c81d9f9eb2c6efddca6df33ae10e431d2d83b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 635, + 388, + 687, + 436 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/57134a4754730b4882d93742904b59ee5869e3d6fed9d70cfd46237e3db1f0e1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 689, + 388, + 761, + 436 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/dad18aa7f2248a0407ec92fec14ea9d085e12f1a4c08875ed2ba1d9be5804932.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 761, + 388, + 828, + 436 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/49f5260a048b6661ba0b6e12321bd67aec00f7250beb1e51277a9f0e2ab9ffd1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 828, + 388, + 890, + 436 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/829d828f813648f8ae7a0f65200a4dddf3533c0ca52c977dc877b030c2e7fbd1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 517, + 440, + 573, + 486 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/a5fcc0e5a9a3ad6d0b3065ebda2fbbc7cb33e688df95ebddfd9130d596833f7d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 583, + 440, + 633, + 486 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/34d842752f58c2d6a72de7c6c01d99629fc1d36963f3c1a1742e5ef018262c22.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 635, + 440, + 687, + 486 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/de1e1669113de871305e5f24ef148c0b789ac126d238d24f700c9c489450d3b2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 689, + 440, + 761, + 486 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/f40580b085c30e268001f17eb472427d5c06bd829dd5f4e655d0f8cdb7f1845c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 761, + 440, + 828, + 486 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/b6ba97081f556ea8867104d01590efb35130fa1e4a13a14f1807e797d6744b6c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 828, + 440, + 880, + 486 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/621451e5081d7528f487226a6ab565d338947f7b43248db89a92ea51f39f0959.jpg", + "image_caption": [ + "Figure 5. We use NeRF to synthesize novel-view object images to augment the limited given visual references. In this work, we train per-instance Zip-NeRF [1]. We visualize synthesized images at different angles (in row 2-4) together with the raw visual references (in the first row). The visual references are from HR-InsDet dataset [48], where the QR code is used for camera pose estimation. Overall, we find synthesized images show high visual quality." + ], + "image_footnote": [], + "bbox": [ + 517, + 489, + 573, + 536 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/bb4f3f82f28371d4d03f0f8d1b9e6c8dbcea35d0537c1895015eaa76de69bed7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 583, + 488, + 642, + 536 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/2227a2aea790a14e09cae9f7b8b6d0bddc810cc2fdbbf176ac0d4683fba10118.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 642, + 488, + 687, + 536 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/1552f025dba22887fa9fddef0fc9920007df0f52bd5d2806d6a65ff4b5b199cc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 689, + 488, + 761, + 536 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/50ae24f9bd6d764b65a57adfc26143be676ae3158a864bf0e338e4ecddc04e6a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 761, + 488, + 834, + 536 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/847c9b59247efcca4dd8995311cba223229c42966137d36b6d609e760de94cb5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 834, + 488, + 883, + 536 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "of various self-supervised learning methods [38] which may sample examples serving as both positive and negative data at the same time.", + "bbox": [ + 511, + 654, + 906, + 698 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Novel-view synthesis. Each object instance might be represented by only a few visual examples, especially in the CID setting which provides limited real visual references for instances of interest. To augment available visual references for better adapting FM, we learn a NeRF [35] for each instance and use NeRF to synthesize novel-view images. In this process, we estimate relative camera poses (required by NeRF) using COLMAP [47], as commonly done in NeRF methods [1, 35] and InsDet approaches [26, 48]. Fig. 5 displays some random NeRF-generated novel-view images with good quality in visuals. It is worth noting that a NeRF needs to be trained only once on the visual references of an object instance. Once it is trained, we use it to generate", + "bbox": [ + 511, + 704, + 908, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "9905", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/2b8da109456542c5f7b6efb99356fe161983071d1397ef3dfb459886d3a32a2d.jpg", + "table_caption": [ + "Table 1. Benchmarking results in the CID setting on the HR-InsDet dataset. We compare our IDOW with state-of-the-arts and make three salient conclusions. First, IDOW significantly outperforms previous methods, e.g., $\\mathrm{IDOW}_{\\mathrm{GroundingDINO}}$ (57.01 AP) > OTS-FM_{GroundingDINO} (51.68 AP) > CPL_{DINO} (27.99 AP). This confirms the importance of addressing InsDet from the open-world perspective. Second, adapting FMs by our IDOW further boosts the performance by 5-7 AP, e.g., $\\mathrm{IDOW}_{\\mathrm{SAM}}$ (48.75 AP) > OTS-FMSAM (41.61 AP). Third, IDOW and OTS-FM are applicable to different pretrained FMs and adopting stronger FMs achieves better performance, e.g., using GroundingDINO yields $>8$ AP than SAM in IDOW." + ], + "table_footnote": [], + "table_body": "
MethodVenue & YearAP\\( \\mathbf{A}\\mathbf{P}_{50} \\)\\( \\mathbf{A}\\mathbf{P}_{75} \\)
avghardeasysmallmediumlarge
\\( CPL_{FasterRCNN} \\)[9, 45]NeurIPS 201519.5410.2623.755.0322.2037.9729.2123.26
\\( CPL_{RetinaNet} \\)[9, 27]ICCV 201722.2214.9226.495.4825.8042.7131.1924.98
\\( CPL_{CenterNet} \\)[9, 64]CVPR 201921.1211.8525.705.9024.1540.3832.7223.60
\\( CPL_{FCOS} \\)[9, 54]ICCV 201922.4013.2228.686.1726.4638.1332.8025.47
\\( CPL_{DINO} \\)[9, 59]ICLR 202327.9917.8932.6511.5131.6048.3539.6232.19
OTS-FMSAM[22, 48]NeurIPS 202341.6128.0347.5714.5845.8369.1449.1045.95
OTS-FMGroundingDINO[29, 48]NeurIPS 202351.6837.2358.7228.7958.5569.2262.5056.78
\\( IDOW_{SAM} \\)48.7532.0956.5020.7555.2673.4357.5954.06
\\( IDOW_{GroundingDINO} \\)57.0140.7464.3635.2562.9873.6469.3362.84
", + "bbox": [ + 91, + 174, + 906, + 354 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "more visual references not only for training but for testing. In particular, such NeRF-synthesized novel-view images can be stored together with the original visual references in testing for proposal-instance matching. Doing so significantly improves InsDet performance (Table 3 and Fig. 7). To the best of our knowledge, our work is the first that exploits NeRF to synthesize data for InsDet.", + "bbox": [ + 89, + 366, + 482, + 470 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 489, + 222, + 507 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we first introduce the experimental setup, then evaluate our IDOW by comparing existing methods in both CID and NID settings, and finally conduct ablation studies to analyze our approach.", + "bbox": [ + 89, + 515, + 483, + 575 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 89, + 588, + 279, + 604 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. We use two recently published datasets in our experiments. HR-InsDet [48] is developed for the CID setting, containing 100 daily object instances, 160 high-resolution testing images from 14 indoor scenarios (see Fig. 3 for an example). RoboTools [26] is developed for the NID setting, containing 20 robotic tool instances and 1,581 testing images from 24 indoor scenarios (see Fig. 4 for an example). For a fair comparison, we use the OWID dataset released by [26] to adapt FM in the NID setting.", + "bbox": [ + 89, + 611, + 482, + 747 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Metrics. We use metrics adopted in both datasets mentioned above, including average precision (AP) at IoU thresholds from 0.5 to 0.95 with the step size 0.05 as the primary metric and also report $\\mathrm{AP}_{50}$ and $\\mathrm{AP}_{75}$ averaged over all instances with IoU threshold as 0.5 and 0.75, respectively. The HR-InsDet dataset contains tags for hard/easy testing scenes and small/medium/large object instances. We follow [48] to break down results on these tags.", + "bbox": [ + 89, + 748, + 482, + 868 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Compared methods. We compare previous methods including CPL, OTS-FM, and VoxDet. These methods have", + "bbox": [ + 89, + 869, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "options to use different components (so do ours). For example, CPL can train a detector using one of the popular architectures Faster RCNN [45], RetinaNet [27], CenterNet [64], FCOS [54], and DINO [59]. OTS-FM can use SAM or GroundingDINO for proposal detection. We implement these methods by using different backbones to improve their performance (Table 1). We use subscripts to mark the backbones used in each method. Moreover, we compare more methods in the NID setting, including OS2D [39],DTOID [34], and $\\mathrm{OLN}_{\\mathrm{Corr}}$ [20].", + "bbox": [ + 511, + 366, + 906, + 517 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation details. We carry out all experiments on a single NVIDIA 3090 GPU and implement methods with PyTorch. Unless otherwise specified, our IDOW uses GroundingDINO [29] as the proposal detector and DINOv2 [38] as the feature foundation model to finetune. We use the pretrained [CLS] token to extract features and class embedding to compute similarity scores for matching. When finetuning, we set hyper-parameter $\\alpha = 0.5$ in the metric learning loss. We use Adam optimizer with a learning rate 1e-3 and weight decay 0.5. We set the batch size as 100 and finetune the model for 10 epochs.", + "bbox": [ + 511, + 518, + 908, + 684 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Benchmarking Results", + "text_level": 1, + "bbox": [ + 511, + 696, + 723, + 710 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Quantitative results. We evaluate our method against prior approaches on the InsDet and RoboTools dataset. Table 1 and 2 list detailed results in the CID and NID settings, respectively. We summarize three salient observations:", + "bbox": [ + 511, + 718, + 906, + 777 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "First, approaching InsDet from the open-world perspective significantly outperforms previous methods. From both CID and NID settings, we find that, despite not being explicitly trained on the instances of interests, OTS-FM already achieves $>10$ AP than traditional detector-based methods which particularly train on such instances. Adapting FMs by our IDOW boosts performance by 3-5 AP.", + "bbox": [ + 511, + 779, + 908, + 883 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Second, IDOW significantly outperforms state-of-the-art", + "bbox": [ + 531, + 885, + 906, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "9906", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/6cf9eb521679c594896c30332d81f55203f31da2c04693e8d76effc39daf572d.jpg", + "image_caption": [ + "Figure 6. t-SNE visualization of features learned by different methods. We compare our finetuned DINOv2 against a finetuned ImageNet-pretrained ResNet101 model and the baseline instance detector CPL with a FasterRCNN architecture. We color feature points w.r.t instance IDs and overlay star-marked visual references (zoom-in to see better). Visually, the finetuned DINOv2 extracts more discriminative features." + ], + "image_footnote": [], + "bbox": [ + 99, + 90, + 356, + 287 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/029a4bf98c259efb71d3b9a3329752f557e396db67ca6526b7b80f9bcaa44a4d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 367, + 90, + 625, + 287 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/50bd78c6da05e1478ae947e40c67079c5050e9fbd893202cc4caea089b925398.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 638, + 90, + 895, + 287 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/09bdb5564a148b0df38bb63d3c5f6b234b7abb47b61fae056a12557379718581.jpg", + "table_caption": [ + "Table 2. Benchmarking results in the NID setting on RoboTools. We compare our IDOW with state-of-the-arts and make three salient conclusions. First, IDOW significantly outperforms previous methods, e.g., $\\mathrm{IDOW}_{\\mathrm{GroundingDINO}}$ (59.4 AP) > OTS-FM_{GroundingDINO} (56.7 AP) > VoxDet (18.7 AP). This demonstrates the effectiveness of approaching InsDet from the open-world perspective. Second, adapting FMs by our IDOW further boosts the performance by 3-5 AP, e.g., $\\mathrm{IDOW}_{\\mathrm{SAM}}$ (51.9 AP) > OTS-FMSAM (46.5 AP). Third, adopting stronger FMs achieves better performance, e.g., $\\mathrm{IDOW}_{\\mathrm{GroundingDINO}}$ (59.4 AP) > $\\mathrm{IDOW}_{\\mathrm{SAM}}$ (51.9 AP) in IDOW." + ], + "table_footnote": [], + "table_body": "
MethodVenue & YearAPAP50AP75
OS2D [39]ECCV 20202.96.52.0
DTOID [34]WACV 20213.69.02.0
OLNCorr. [20]RA-L 202214.418.115.7
VoxDet [26]NeurIPS 202318.723.620.5
OTS-FMSAM [48]NeurIPS 202346.555.950.4
OTS-FMGroundingDINO [48]NeurIPS 202356.764.859.0
IDOWSAM51.963.856.5
IDOWGroundingDINO59.467.861.8
", + "bbox": [ + 91, + 492, + 483, + 640 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "InsDet methods, achieving over 5 AP higher on average than methods like OTS-FM in both CID and NID settings, demonstrating the effectiveness of our techniques.", + "bbox": [ + 89, + 655, + 483, + 700 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Third, IDOW is versatile and applicable to different open-world pretrained FMs. By comparing $\\mathrm{IDOW}_{\\mathrm{SAM}}$ and $\\mathrm{IDOW}_{\\mathrm{GroundingDINO}}$ , we find that adopting stronger FMs (GroundingDINO vs. SAM) achieves better performance.", + "bbox": [ + 89, + 702, + 483, + 762 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Qualitative Results. Fig. 3 and 4 compares InsDet results by different methods in the CID and NID settings, respectively. We can see that our IDOW has higher precision and detects more wanted instances, which are small and under low lighting conditions. Fig. 6 uses tSNE to visualize features of visual references and detected proposals (in the HR-InsDet dataset) computed by our finetuned DINOv2 (ft-DINOv2) in the left panel, and other non-FM methods including a finetuned ImageNet-pretrained ResNet101 back", + "bbox": [ + 88, + 763, + 485, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "bone (ft-ImageNet-Res101) and CPL_FasterRCNN in the middle and right panels, respectively. Clearly, ft-DINOv2 matches proposals and references much better than other models.", + "bbox": [ + 511, + 349, + 906, + 396 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Study and Further Analysis", + "text_level": 1, + "bbox": [ + 511, + 402, + 834, + 420 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In Table 3, we ablate the proposed techniques of NeRF-synthesis and distractor sampling. Results show that adapting the FM DINOv2 greatly boosts InsDet performance (from 51.68 AP by the baseline to 53.94 AP). Importantly, including NeRF-synthesized images in both training and testing significantly improves the performance. Yet, using them in testing brings more performance gains than training (refer to the next study in Fig. 7). Eventually, applying our techniques boosts AP to 57.01. The supplemental material contains more quantitative and qualitative results.", + "bbox": [ + 511, + 425, + 908, + 575 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Following the above, we study the affects of using different number of NeRF-synthesized images in training and test. Fig. 7 depicts the results. (1) Adding NeRF-generated images helps foundation model adaptation for InsDet. Specifically, adding $<24$ generated images boosts the performance, especially when 0 synthesized images are used in testing. However, adding $>24$ generated images for training induces diminishing performance gains, likely because that the adapted FM overfits to artifacts in NeRF generated images. (2) Adopting more synthetic images in testing consistently improves performance. Yet, the performance by adding more (e.g., $>36$ ) in testing saturates. The peak performance occurs when 36 NeRF-synthesized images are used in testing.", + "bbox": [ + 511, + 577, + 910, + 773 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5. Discussions", + "text_level": 1, + "bbox": [ + 511, + 784, + 635, + 800 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Broader Impacts. Approaching InsDet in the open-world has a significant impact on various downstream applications, such as robotics and AR/VR. Detecting object instances in RGB images is considered the first step of contemporary perception algorithms. A stronger InsDet model provides richer information for the following steps, such", + "bbox": [ + 511, + 810, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "9907", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/25caaab3137560315849c22265fc94c72df8d46ad39e8900454754ea7ac7e8f0.jpg", + "table_caption": [ + "Table 3. Ablation study of each strategy involved in our IDOW. We carry out the study in the CID setting on the HR-InsDet dataset. We use OTS-FM $_{\\text{GroundingDINO}}$ as a baseline, over which we incrementally add each strategy. \"Train\" means foundation model adaptation through finetuning on the available data. DA denotes Data Augmentation with NeRF-generated novel-views; DS denotes Distractor Sampling. Results clearly demonstrate that all the four strategies help achieve better InsDet performance. Finetuning FMs on the given visual references enhances detection performance, cf. Train (53.94 AP) > baseline (51.68 AP). Moreover, NeRF-based data augmentation improves the final detection performance, particularly when used in testing, cf. Train+DA@Test (56.44 AP) > Train+DA@Train (54.48 AP) > baseline (51.68 AP). Lastly, applying distractor sampling (DS) improves the final performance further, cf. Train+DS (54.10 AP) > Train (53.94 AP)." + ], + "table_footnote": [], + "table_body": "
StrategiesAP\\( \\mathbf{A}\\mathbf{P}_{50} \\)\\( \\mathbf{A}\\mathbf{P}_{75} \\)
DA@TestTrainDA@TrainDSavghardeasysmallmediumlarge
baseline: OTS-FMGroundingDINO51.6837.2358.7228.7958.5569.2262.5056.78
✓53.9437.5461.5230.1860.9871.6065.1859.36
✓✓54.4838.3161.7131.1861.1072.5566.0659.98
✓✓54.1037.7361.6530.2761.1971.2765.3759.57
✓✓✓54.9238.2862.4132.0061.6174.2166.5860.50
✓55.2439.6462.8134.0061.2072.3866.9960.88
✓✓56.4439.7164.2434.7263.1374.4268.5062.31
✓✓✓56.9239.8564.8534.6662.7475.4868.8962.80
✓✓✓56.5139.7064.3634.3461.6575.8968.7162.38
✓✓✓✓57.0140.7464.3635.2562.9873.6469.3362.84
", + "bbox": [ + 91, + 188, + 906, + 388 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7460d61347fe8fa2529db2ae5da64ac653743cf68825e265c42b724202a0ce72.jpg", + "image_caption": [ + "Figure 7. Analysis of the number of added NeRF-synthesized images in training and testing w.r.t AP improvement over the \"Train\" method (Table 3), which by default finetunes DINOv2 using the available visual references (24 per instance). We carry out this analysis in the CID setting on the HR-InsDet dataset. On the one hand, adding NeRF-synthesized images helps finetune FM for InsDet, but adding more (e.g., $>24$ ) hurts the performance, likely due to overfitting to the artifacts of NeRF-synthesized images. On the other hand, adding more NeRF images in testing consistently improves performance." + ], + "image_footnote": [], + "bbox": [ + 101, + 406, + 465, + 585 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "as grasping or trajectory forecasting in robotics. Additionally, our work shows the promising performance of adapting FMs to InsDet, which will foster future research in studying different properties of FMs, such as adapting FMs to other tasks. Nonetheless, our research utilizes FMs that are potentially pretrained on private data. At this point, it is unclear how such FMs might have bias or ethical issues. As a result, we have no safeguard measure to ensure that the finetuned/adapted FMs are fair, even though we find no visible issues in the image-only dataset used for finetuning.", + "bbox": [ + 89, + 750, + 485, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Further, adopting SAM may result in missed detections of critical instances relevant to specific downstream applications, such as medical devices or other rare objects. Such potential issues are not addressed in this work.", + "bbox": [ + 511, + 407, + 906, + 468 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations and future work. Experiments demonstrate that our proposed techniques effectively leverage FMs to address InsDet across various settings. Nonetheless, we note several limitations for future improvements. First, the current pipeline relies on the proposals generated from GroundingDINO. Despite its high recall, it is not yet ready for real-time applications. Future work should consider faster proposal detection methods. Second, the NeRF augmentation step is also slow since it requires training a NeRF for every instance, which takes 1 hour and 500MB space using one NVIDIA 3090 GPU. While developing efficient NeRF methods is beyond our scope, we point out that efficient NeRF methods can be readily used in our pipeline, such as [37] which trains a NeRF in seconds and [6] which can train a single NeRF for multiple objects.", + "bbox": [ + 511, + 470, + 908, + 698 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 720, + 633, + 736 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Instance detection is highly applicable in various scenarios such as in robotics and AR/VR. We elaborate on its open-world nature, necessitating the development of InsDet models in the open world. We solve InsDet from the open world perspective, embracing foundation models (FMs) and data sampled in the open world. In particular, we propose simple and effective techniques such as metric learning, distractor sampling, and novel-view synthesis. Extensive experiments in different settings validate that our techniques in adapting FMs significantly boost InsDet performance.", + "bbox": [ + 511, + 750, + 908, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "9908", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 91, + 90, + 250, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported by FDCT (0067/2024/ITP2), the University of Macau (SRG2023-00044-FST), the Institute of Collaborative Innovation and NSFC (No.62206256).", + "bbox": [ + 89, + 116, + 485, + 161 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 179, + 187, + 194 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Zip-nerf: Anti-aliased grid-based neural radiance fields. In ICCV, 2023. 2, 5", + "[2] Herbert Bay, Tinne Tuytelaars, and Luc Van Gool. Surf: Speeded up robust features. In ECCV, 2006. 2", + "[3] Richard Bormann, Xinjie Wang, Markus Volk, Kilian Kleeberger, and Jochen Lindermayr. Real-time instance detection with fast incremental learning. In ICRA. IEEE, 2021. 1", + "[4] Mathilde Caron, Hugo Touvron, Ishan Misra, Herve Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, 2021. 3", + "[5] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv:1512.03012, 2015. 2", + "[6] Hansheng Chen, Jiatao Gu, Anpei Chen, Wei Tian, Zhuowen Tu, Lingjie Liu, and Hao Su. Single-stage diffusion nerf: A unified approach to 3d generation and reconstruction. In ICCV, 2023. 8", + "[7] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In ICML, 2020. 3, 1", + "[8] Jasmine Collins, Shubham Goel, Kenan Deng, Achleshwar Luthra, Leon Xu, Erhan Gundogdu, Xi Zhang, Tomas F Yago Vicente, Thomas Dideriksen, Himanshu Arora, et al. Abo: Dataset and benchmarks for real-world 3d object understanding. In CVPR, 2022. 2", + "[9] Debidatta Dwibedi, Ishan Misra, and Martial Hebert. Cut, paste and learn: Surprisingly easy synthesis for instance detection. In ICCV, 2017. 1, 2, 3, 4, 5, 6", + "[10] David Gale and Lloyd S Shapley. College admissions and the stability of marriage. The American Mathematical Monthly, 69(1):9-15, 1962. 4", + "[11] Georgios Georgakis, Arsalan Mousavian, Alexander C. Berg, and Jana Kosecka. Synthesizing training data for object detection in indoor scenes. In Proceedings of Robotics: Science and Systems, 2017. 1, 2, 3", + "[12] Georgios Georgakis, Arsalan Mousavian, Alexander C Berg, and Jana Košecká. Synthesizing training data for object detection in indoor scenes. Robotics: Science and Systems (RSS), 2017. 3, 4", + "[13] Sachin Goyal, Ananya Kumar, Sankalp Garg, Zico Kolter, and Aditi Raghunathan. Finetune like you pretrain: Improved finetuning of zero-shot vision models. In CVPR, 2023. 3", + "[14] Sunan He, Taian Guo, Tao Dai, Ruizhi Qiao, Chen Wu, Xiujun Shu, and Bo Ren. Vlmae: Vision-language masked autoencoder. arXiv:2208.09374, 2022. 3" + ], + "bbox": [ + 93, + 204, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] Dan Hendrycks, Mantas Mazeika, and Thomas Dietterich. Deep anomaly detection with outlier exposure. In ICLR, 2019. 3", + "[16] Stefan Hinterstoisser, Cedric Cagniart, Slobodan Ilic, Peter Sturm, Nassir Navab, Pascal Fua, and Vincent Lepetit. Gradient response maps for real-time detection of textureless objects. IEEE transactions on pattern analysis and machine intelligence, 34(5):876-888, 2011. 2", + "[17] Tomás Hodań, Vibhav Vineet, Ran Gal, Emanuel Shalev, Jon Hanzelka, Treb Connell, Pedro Urbina, Sudipta N Sinha, and Brian Guenter. Photorealistic image synthesis for object instance detection. In IEEE international conference on image processing (ICIP), 2019. 3", + "[18] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc V. Le, Yunhsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In ICML, 2021. 3", + "[19] Wadim Kehl, Fabian Manhardt, Federico Tombari, Slobodan Ilic, and Nassir Navab. Ssd-6d: Making rgb-based 3d detection and 6d pose estimation great again. In ICCV, 2017. 3", + "[20] Dahun Kim, Tsung-Yi Lin, Anelia Angelova, In So Kweon, and Weicheng Kuo. Learning open-world object proposals without learning to classify. IEEE Robotics and Automation Letters, 7(2):5453-5460, 2022. 3, 6, 7", + "[21] Eun Kyeong Kim, Hansoo Lee, Jin Yong Kim, and Sungshin Kim. Data augmentation method by applying color perturbation of inverse psnr and geometric transformations for object recognition based on deep learning. Applied Sciences, 10(11): 3755, 2020. 3", + "[22] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In ICCV, 2023. 2, 3, 5, 6", + "[23] Shu Kong and Deva Ramanan. Opengan: Open-set recognition via open data generation. In ICCV, 2021. 2, 3", + "[24] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. ImageNet classification with deep convolutional neural networks. NeurIPS, 25, 2012. 3", + "[25] Kevin Lai, Liefeng Bo, and Dieter Fox. Unsupervised feature learning for 3d scene labeling. ICRA, 2014. 3", + "[26] Bowen Li, Jiashun Wang, Yaoyu Hu, Chen Wang, and Sebastian Scherer. Voxdet: Voxel learning for novel instance detection. NeurIPS, 2024. 1, 2, 3, 4, 5, 6, 7", + "[27] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. In ICCV, 2017. 6, 3", + "[28] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. NeurIPS, 2023. 3", + "[29] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. ECCV, 2024. 3, 6", + "[30] Tian Liu, Huixin Zhang, Shubham Parashar, and Shu Kong. Few-shot recognition via stage-wise retrieval-augmented finetuning. In CVPR, 2025. 3", + "[31] David G Lowe. Distinctive image features from scale-invariant keypoints. *IJCV*, 60:91-110, 2004. 2" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9909", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[32] Anish Madan, Neehar Peri, Shu Kong, and Deva Ramanan. Revisiting few-shot object detection with vision-language models. In NeurIPS Datasets & Benchmark Track, 2024. 3", + "[33] David G McVitie and Leslie B Wilson. The stable marriage problem. Communications of the ACM, 14(7):486-490, 1971. 4", + "[34] Jean-Philippe Mercier, Mathieu Garon, Philippe Giguere, and Jean-Francois Lalonde. Deep template-based object instance detection. In WACV, 2021. 1, 2, 6, 7", + "[35] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 3, 5", + "[36] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 2", + "[37] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM transactions on graphics (TOG), 41(4):1-15, 2022. 8", + "[38] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv:2304.07193, 2023. 2, 3, 4, 5, 6", + "[39] Anton Osokin, Denis Sumin, and Vasily Lomakin. Os2d: One-stage one-shot object detection by matching anchor features. In ECCV, 2020. 6, 7", + "[40] Shubham Parashar, Zhiqiu Lin, Yanan Li, and Shu Kong. Prompting scientific names for zero-shot species recognition. In EMNLP, 2023. 3", + "[41] Shubham Parashar, Zhiqiu Lin, Tian Liu, Xiangjue Dong, Yanan Li, Deva Ramanan, James Caverlee, and Shu Kong. The neglected tails in vision-language models. In CVPR, 2024. 3", + "[42] Samira Pouyanfar, Saad Sadiq, Yilin Yan, Haiman Tian, Yudong Tao, Maria Presa Reyes, Mei-Ling Shyu, Shu-Ching Chen, and Sundaraja S Iyengar. A survey on deep learning: Algorithms, techniques, and applications. ACM computing surveys (CSUR), 51(5):1-36, 2018. 1", + "[43] A Quadros, James Patrick Underwood, and Bertrand Douillard. An occlusion-aware feature for range images. In ICRA, 2012. 2", + "[44] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021. 3", + "[45] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. In NeurIPS, 2015. 6, 3", + "[46] Tianhe Ren, Shilong Liu, Ailing Zeng, Jing Lin, Kunchang Li, He Cao, Jiayu Chen, Xinyu Huang, Yukang Chen, Feng Yan, et al. Grounded sam: Assembling open-world models for diverse visual tasks. arXiv:2401.14159, 2024. 3" + ], + "bbox": [ + 91, + 92, + 483, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[47] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-Motion Revisited. In CVPR, 2016. 5", + "[48] Qianqian Shen, Yunhan Zhao, Nahyun Kwon, Jeeun Kim, Yanan Li, and Shu Kong. A high-resolution dataset for instance detection with multi-view object capture. In NeurIPS Datasets & Benchmark Track, 2023. 1, 2, 3, 4, 5, 6, 7", + "[49] Jia Shi, Gautam Gare, Jinjin Tian, Siqi Chai, Zhiqiu Lin, Arun Vasudevan, Di Feng, Francesco Ferroni, and Shu Kong. Lcaon-the-line: Benchmarking out-of-distribution generalization with class taxonomies. In ICML, 2024. 3", + "[50] Abhinav Shrivastava, Abhinav Gupta, and Ross Girshick. Training region-based object detectors with online hard example mining. In CVPR, 2016. 5", + "[51] Julio Silva-Rodriguez, Sina Hajimiri, Ismail Ben Ayed, and Jose Dolz. A closer look at the few-shot adaptation of large vision-language models. In CVPR, 2024. 3", + "[52] Afrina Tabassum, Muntasir Wahed, Hoda Eldardiry, and Ismini Lourentzou. Hard negative sampling strategies for contrastive representation learning. arXiv:2206.01197, 2022. 1", + "[53] Peng Tang, Xinggang Wang, Xiang Bai, and Wenyu Liu. Multiple instance detection network with online instance classifier refinement. In CVPR, 2017. 1", + "[54] Zhi Tian, Chunhua Shen, Hao Chen, and Tong He. Fcos: Fully convolutional one-stage object detection. In ICCV, 2019. 6, 3", + "[55] Hugo Touvron, Matthieu Cord, and Hervé Jégou. Deit iii: Revenge of the vit. In ECCV, 2022. 3", + "[56] Riccardo Volpi, Pietro Morierio, Silvio Savarese, and Vittorio Murino. Adversarial feature augmentation for unsupervised domain adaptation. In CVPR, 2018. 3", + "[57] Riccardo Volpi, Hongseok Namkoong, Ozan Sener, John C Duchi, Vittorio Murino, and Silvio Savarese. Generalizing to unseen domains via adversarial data augmentation. NeurIPS, 2018. 3", + "[58] XuDong Wang, Jingfeng Yang, and Trevor Darrell. Segment anything without supervision. In NeurIPS, 2024. 3", + "[59] Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel Ni, and Harry Shum. Dino: Detr with improved denoising anchor boxes for end-to-end object detection. In ICLR, 2022. 3, 6", + "[60] Xinyu Zhang, Qiang Wang, Jian Zhang, and Zhao Zhong. Adversarial autoaugment. In ICLR, 2020. 3", + "[61] Yunhan Zhao, Shu Kong, and Charless Fowlkes. Camera pose matters: Improving depth prediction by mitigating pose distribution bias. In CVPR, 2021. 3", + "[62] Yunhan Zhao, Haoyu Ma, Shu Kong, and Charless Fowlkes. Instance tracking in 3d scenes from egocentric videos. In CVPR, 2024. 1", + "[63] Jinghao Zhou, Chen Wei, Huiyu Wang, Wei Shen, Cihang Xie, Alan Yuille, and Tao Kong. Image bert pre-training with online tokenizer. In ICLR, 2022. 3", + "[64] Xingyi Zhou, Dequan Wang, and Philipp Krahenbuhl. Objects as points. In arXiv:1904.07850, 2019. 6, 3" + ], + "bbox": [ + 516, + 92, + 906, + 833 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "9910", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Solving Instance Detection from an Open-World Perspective/ab6d1961-27f9-4a61-af19-064b80505b12_model.json b/2025/Solving Instance Detection from an Open-World Perspective/ab6d1961-27f9-4a61-af19-064b80505b12_model.json new file mode 100644 index 0000000000000000000000000000000000000000..f5edb76a5d68afd45ded1702ade0da2b0dbe8f35 --- /dev/null +++ b/2025/Solving Instance Detection from an Open-World Perspective/ab6d1961-27f9-4a61-af19-064b80505b12_model.json @@ -0,0 +1,2552 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.195, + 0.131, + 0.804, + 0.154 + ], + "angle": 0, + "content": "Solving Instance Detection from an Open-World Perspective" + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.179, + 0.884, + 0.201 + ], + "angle": 0, + "content": "Qianqian Shen\\(^{1,*}\\) Yunhan Zhao\\(^{2,*}\\) Nahyun Kwon\\(^{3}\\) Jeeun Kim\\(^{3}\\) Yanan Li\\(^{4,\\dagger}\\) Shu Kong\\(^{5,6,\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.277, + 0.201, + 0.721, + 0.217 + ], + "angle": 0, + "content": "1Zhejiang University, 2UC Irvine, 3Texas A&M University, 4Zhejiang Lab" + }, + { + "type": "text", + "bbox": [ + 0.318, + 0.219, + 0.676, + 0.233 + ], + "angle": 0, + "content": "5University of Macau, 6Institute of Collaborative Innovation" + }, + { + "type": "text", + "bbox": [ + 0.299, + 0.238, + 0.695, + 0.251 + ], + "angle": 0, + "content": "website and code: https://shenqq377.github.io/IDOW" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.287, + 0.327, + 0.303 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.32, + 0.49, + 0.744 + ], + "angle": 0, + "content": "Instance detection (InsDet) aims to localize specific object instances within a novel scene imagery based on given visual references. Technically, it requires proposal detection to identify all possible object instances, followed by instance-level matching to pinpoint the ones of interest. Its open-world nature supports its broad applications from robotics to AR/VR but also presents significant challenges: methods must generalize to unknown testing data distributions because (1) the testing scene imagery is unseen during training, and (2) there are domain gaps between visual references and detected proposals. Existing methods tackle these challenges by synthesizing diverse training examples or utilizing off-the-shelf foundation models (FMs). However, they only partially capitalize the available open-world information. In contrast, we approach InsDet from an Open-World perspective, introducing our method IDOW. We find that, while pretrained FMs yield high recall in instance detection, they are not specifically optimized for instance-level feature matching. Therefore, we adapt pretrained FMs for improved instance-level matching using open-world data. Our approach incorporates metric learning along with novel data augmentations, which sample distractors as negative examples and synthesize novel-view instances to enrich the visual references. Extensive experiments demonstrate that our method significantly outperforms prior works, achieving \\(>10\\) AP over previous results on two recently released challenging benchmark datasets in both conventional and novel instance detection settings." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.773, + 0.222, + 0.789 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.8, + 0.486, + 0.877 + ], + "angle": 0, + "content": "Instance detection (InsDet) aims to localize object instances of interest in novel scene imagery (Fig. 1), where these objects are specified by some visual references (as known as support templates) [3, 26, 34, 48, 53]. Its open-world nature supports its wide-ranging applications" + }, + { + "type": "text", + "bbox": [ + 0.53, + 0.286, + 0.633, + 0.298 + ], + "angle": 0, + "content": "visual references" + }, + { + "type": "image", + "bbox": [ + 0.519, + 0.3, + 0.645, + 0.412 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.662, + 0.286, + 0.83, + 0.299 + ], + "angle": 0, + "content": "novel testing scene imagery" + }, + { + "type": "image", + "bbox": [ + 0.65, + 0.3, + 0.845, + 0.412 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.846, + 0.299, + 0.898, + 0.411 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.413, + 0.907, + 0.51 + ], + "angle": 0, + "content": "Figure 1. Instance Detection (InsDet) aims to localize specific object instances in novel scene imagery based on given visual references (aka support templates). It is a challenging problem due to its open-world nature: the testing scene images are unseen during training and thus are unknown to InsDet models, and visual references and detected proposals have domain gaps (e.g., due to occlusions and arbitrary lighting conditions in the latter)." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.522, + 0.909, + 0.644 + ], + "angle": 0, + "content": "in robotics and AR/VR. For example, a robot can be commanded to search for a customer's suitcase at an airport [26], or my_keys in my cluttered bedroom [62]. Its open-world nature also presents significant challenges: during training, one has no knowledge of testing data distribution, which can be instantiated by the unknown scene imagery [9], novel object instances encountered only in testing [26], and domain gaps between visual references and detected proposals [48]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.645, + 0.911, + 0.902 + ], + "angle": 0, + "content": "Status quo. The literature of InsDet has two settings and existing InsDet methods attempt to address the open world in different aspects (Fig. 2). The conventional instance detection (CID) setting provides visual references of object instances for training. Yet, the testing scene is still unknown in training time. Therefore, early method Cut Paste Learn (CPL) collects random background images to model diverse testing scenes (Fig. 2a), cuts instances from the visual references, and pastes them on these background images to construct a synthetic training set [9, 11]. It is worth noting that background image sampling is utilizing open-world information in a way of data collection in the open world. On the other hand, the novel instance detection (NID) setting, introduced recently [26], studies detecting novel object instances that are specified only in testing time where a trained model is not allowed to be tuned further. To approach NID, VoxDet [26] uses large-scale synthetic data (Fig. 2b) to learn" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.887, + 0.358, + 0.901 + ], + "angle": 0, + "content": "*Equal contributions; †Corresponding authors." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "9901" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.095, + 0.088, + 0.24, + 0.099 + ], + "angle": 0, + "content": "(a) background imagery [9, 11]" + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.099, + 0.245, + 0.201 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.254, + 0.088, + 0.359, + 0.099 + ], + "angle": 0, + "content": "(b) object images [26]" + }, + { + "type": "image", + "bbox": [ + 0.252, + 0.099, + 0.372, + 0.201 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.37, + 0.088, + 0.496, + 0.099 + ], + "angle": 0, + "content": "(c) Foundation Models [48]" + }, + { + "type": "image", + "bbox": [ + 0.384, + 0.099, + 0.492, + 0.201 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.553, + 0.088, + 0.809, + 0.099 + ], + "angle": 0, + "content": "(d) Solving Instance Detection in the Open World (ours)" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.099, + 0.895, + 0.202 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.204, + 0.907, + 0.328 + ], + "angle": 0, + "content": "Figure 2. Existing InsDet methods leverage the open-world information in different aspects, such as (a) background image sampling (from the open world) to synthesize training data [9, 11], (b) object image sampling (from the open world) to learn feature representations [26], and (c) foundation model utilization (pretrained in the open world) for proposal detection and instance-level feature matching [48]. (d) As FMs are not specifically designed for instance-level feature matching required by InsDet, we propose to adapt them by leveraging rich data sampled from the open world. We gather data from multiple sources: (1) any available visual references of instances in the CID setting, (2) abundant multi-view object images sampled in the open world similar to (b), (3) synthetic data by training NeRF [1, 36] to generate novel-view images based on the given instances, (4) distractors by running FMs (esp. SAM [22]) on random open-world imagery to generate random object-like proposals (Fig. 5). We use the data above to adapt FM through metric learning. The technical novelty of our work lies in (3) and (4), as well as the design choice of metric learning to adapt FMs for InsDet." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.341, + 0.484, + 0.492 + ], + "angle": 0, + "content": "3D voxel representations from visual references and a matching function between 2D proposals and 3D voxels. Recent work OTS-FM [48] uses off-the-shelf pretrained foundation models (Fig. 2c) for InsDet. Specifically, it uses SAM [22] to detect object proposals in scene imagery, and DINOv2 [38] to represent visual references and proposals for instance-level feature matching. OTS-FM is a non-learned method and hence is applicable in both CID and NID settings. Importantly, it achieves the state-of-the-art performance on the recently released benchmark HR-InsDet [48]." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.493, + 0.485, + 0.825 + ], + "angle": 0, + "content": "The open-world perspective. As illustrated in Fig. 1, the challenges of InsDet lie in its open-world nature: the unknown testing data distribution, and domain gaps between visual references and detected proposals. To address the open-world challenges, existing methods take different strategies to exploit the open-world information (Fig. 2a-c). For example, to address the unknown testing scene imagery, some methods sample diverse background images [9, 11], demonstrating a practice of sampling data in the open world. To mitigate the domain gap between detected proposals and visual references, [26] attempts to learn features for instance-level matching by utilizing external datasets (ShapeNet [5] and ABO [8]). This data utilization can be considered as data collection in the open world. For better proposal detection and more discriminative feature representations, [48] adopts off-the-shelf pretrained foundation models (FMs) SAM [22] and DINOv2 [38], respectively. One can think of FMs as being pretrained in the open world. Different from existing methods which have not fully capitalized the open-world information, we take the open-world perspective as the first principle and solve InsDet by exploiting the Open World (Fig. 2d). We call our approach IDOW." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.485, + 0.902 + ], + "angle": 0, + "content": "Insights and novelties. Our insights are rooted in the understanding of InsDet's open-world nature and the practice of existing methods. Existing methods address open-world issues to some extent (Fig. 2), particularly by sampling data in the open world [9, 23, 26] and adopting foundation mod" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.341, + 0.908, + 0.583 + ], + "angle": 0, + "content": "els (FMs) pretrained in the open world [48]. These motivate us to develop InsDet methods in the open world (IDOW). We find that open-world detectors yield high recall on object instances (see precision-recall curves in Fig. 9 of the Appendix) but FMs are not tailored to instance-level matching between proposals and visual references – directly using an FM for InsDet is sub-optimal. Therefore, we adapt FMs to instance-level feature representation by metric learning. We propose novel and simple techniques for data augmentation to enhance FM adaptation and feature learning: distractor sampling, and multi-view synthesis. The former helps learn more discriminative features for instance comparison, and the latter enhances visual references particularly when they are few. With these techniques, our IDOW achieves significant boosts over previously reported results on recently released challenging benchmarks (Table 1 and 2)." + }, + { + "type": "title", + "bbox": [ + 0.534, + 0.584, + 0.879, + 0.598 + ], + "angle": 0, + "content": "Contributions. We make three major contributions:" + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.6, + 0.906, + 0.66 + ], + "angle": 0, + "content": "1. We elaborate on the open-world challenge of InsDet and motivate the necessity of developing methods in the open world. We solve InsDet by exploiting diverse open data available in the open world and FMs pretrained therein." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.661, + 0.907, + 0.706 + ], + "angle": 0, + "content": "2. We introduce simple and effective techniques to adapt FMs to InsDet: a metric learning loss, and data augmentation with distractor sampling and novel-view synthesis." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.707, + 0.907, + 0.766 + ], + "angle": 0, + "content": "3. We extensively compare our approach with existing ones on two recently released datasets, demonstrating that our approach outperforms them by \\(>10\\) AP in both conventional and novel instance detection settings." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.6, + 0.907, + 0.766 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.784, + 0.655, + 0.799 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Instance detection. Early methods use local features such as SIFT [31] and SURF [2] to match visual references [43] and image regions to localize instances [16]. Recent methods train deep neural networks for instance-agnostic proposal detection and proposal-instance matching, achieving significant improvements [26, 34, 48]. In the con" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9902" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.094, + 0.092, + 0.484, + 0.484 + ], + "angle": 0, + "content": "ventional instance detection (CID) setting, visual references of object instances are given during training. To approach CID, prevalent method synthesize training data in a cut-paste-learn (CPL) strategy [9, 11]: cutting instances from visual references and pasting them on random background images, then learning a detector on such synthetic data. Other works strive to obtain more training samples by rendering realistic instance examples [17, 19], using data augmentation [9] and synthesizing better training images [9, 12, 25]. Moreover, recent literature introduces the novel instance detection (NID) setting [26], which requires detecting novel instances specified only in testing and does not allow for further model finetuning during testing. VoxDet [26] utilizes a pretrained open-world detector [20] for proposal detection. Nowadays, the open-world perspective advances proposal detection in a way of learning more generalizable detectors and carrying out open-vocabulary recognition [29]. While a foundational open-vocabulary detector can attach a textual description (e.g., a short phrase) to each proposal, simply using it is insufficient for instance-level matching. [48] realizes this issue although foundational detectors can yield nearly perfect recall. As a result, the recent method OTS-FM uses another FM DINOv2 [38] for proposal-instance matching [48]. Differently, we solve InsDet from the open-world perspective, exploiting open FMs and open data to learn more discriminative instance-level feature representations." + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.494, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Foundation models (FMs) are pre-trained on web-scale data, which can be thought of as being sampled in the open world. Different FMs are trained on different types of data and can shine in different downstream tasks related to visual perception, natural language processing, or both. As this work concerns about visual perception, we briefly review Vision-Language Models (VLMs) and Visual Foundation Models (VFMs). VLMs are pretrained on web-scale image-text paired data [18, 28, 29, 44], demonstrating impressive results in high-level tasks such as visual grounding and image captioning. VFMs are pretrained primarily on visual data [4, 7, 14, 22, 38, 55, 58, 63] and can yield impressive visual perception results such as proposal or open-set object detection [29, 59]. In InsDet, the mainstream framework requires proposal detection and visual feature matching. In this framework, open-world detectors appear to achieve high recall on the instances of interest, owing to that the openworld pretrained detectors generalize quite well in the open world [29, 46, 59]. In contrast, feature matching still remains a challenge in improving InsDet. Therefore, we focus on improving features for InsDet by adapting VFMs using data sampled in the open world. [48] shows that using an off-the-shelf FM DINOv2 [38] for feature matching greatly enhances InsDet detection; yet, FM adaptation has been still under-explored in InsDet, although it is a well studied in other areas such as few-shot perception [30, 32, 51] and zero-shot perception [13, 40, 41]. Our work introduces simple" + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.093, + 0.871, + 0.107 + ], + "angle": 0, + "content": "and novel techniques to adapt FMs to improve InsDet." + }, + { + "type": "text", + "bbox": [ + 0.516, + 0.11, + 0.907, + 0.53 + ], + "angle": 0, + "content": "The open-world data is publicly available for pretraining FMs. Coping with such data is also often challenging. For example, data between training and testing has distribution shifts or domain gaps [30, 49], as clearly instantiated in InsDet by the unknown testing scene imagery, in addition to the domain gaps between (clean) visual references and (occluded) proposals (Fig. 1). Speaking of distribution shifts, some existing methods sample open-world data, demonstrating a need to develop robust models in the open world [15, 23]. Existing InsDet methods also sample open-world data such as diverse background images and instance examples (Fig. 2) [9, 11]. Even so, while open-world data is abundant, most of it is not directly useful for training models for specific downstream tasks. In contrast, data augmentation aims to synthesize more tailored data to enhance models' robustness and generalization, including geometry-aware augmentation (e.g., random cropping and flipping) [7, 24, 61], photometric-aware augmentation (e.g., color jittering) [21, 56, 57], and robustness-aware augmentation (e.g., adversarial perturbation) [60]. It is worth noting that Neural Radiance Field (NeRF) [35], a recent rendering technique, can be used to synthesize novel-view images. In this work, we particularly adopt NeRF to synthesize novel-view visual references of object instances and use them to adapt FMs towards more tailored features to InsDet. To the best of our knowledge, using NeRF is under-explored in the literature of InsDet, but we find that doing so remarkably improves InsDet performance." + }, + { + "type": "title", + "bbox": [ + 0.517, + 0.55, + 0.897, + 0.565 + ], + "angle": 0, + "content": "3. Instance Detection: Protocols and Methods" + }, + { + "type": "text", + "bbox": [ + 0.516, + 0.576, + 0.907, + 0.635 + ], + "angle": 0, + "content": "We first introduce the protocols of InsDet, including the problem definition, evaluation, and two settings. We then present our techniques to enhance foundation model adaptation to InsDet." + }, + { + "type": "title", + "bbox": [ + 0.517, + 0.651, + 0.62, + 0.665 + ], + "angle": 0, + "content": "3.1. Protocols" + }, + { + "type": "text", + "bbox": [ + 0.516, + 0.673, + 0.907, + 0.779 + ], + "angle": 0, + "content": "Problem definition. Instance Detection (InsDet) aims to detect specific object instances from a 2D scene image (Fig. 1), where the objects of interest are defined by a \"support set\" of some visual references [9, 26, 48]. Previous works capture the visual references from multiple camera views, with QR code being pasted to help estimate camera poses (see visual references in Fig. 1)." + }, + { + "type": "text", + "bbox": [ + 0.516, + 0.781, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Challenges. InsDet is a challenging problem due to its open-world nature: one has zero knowledge about the testing data distribution which is instantiated by unknown testing scenes, never-before-seen matters, arbitrary occlusions, background clutters, etc. In particular, object instances of interest can be novel and are only defined during testing, as emphasized by the novel instance detection setting below. Nevertheless, InsDet models must be capable of detecting" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.515, + 0.957 + ], + "angle": 0, + "content": "9903" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.091, + 0.088, + 0.321, + 0.262 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.331, + 0.087, + 0.514, + 0.262 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.087, + 0.705, + 0.262 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.715, + 0.087, + 0.898, + 0.262 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.264, + 0.907, + 0.349 + ], + "angle": 0, + "content": "Figure 3. Visual comparison of InsDet results by different methods in the CID setting on HR-InsDet [48]. The testing scene image contains sparse placement of small instances with challenging illumination. We mark the ground-truth and predictions using green and blue boxes, respectively. We attach instance IDs to them to highlight whether the instance recognition is correct compared to the visual references (i.e., the leftmost references). Compared with Cut-Paste-Learn and OTS-FM, our IDOW detects more instances (see orange arrows) with better accuracy (see red arrows). Comparison between OTS-FM and IDOW suggests that adapting the FM DINOv2 yields better features for IDOW to perform much better in front of domain shifts, i.e., challenging illumination conditions in this testing scene image." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.364, + 0.483, + 0.392 + ], + "angle": 0, + "content": "any object instances of interest and robustly comparing them with visual references." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.394, + 0.484, + 0.439 + ], + "angle": 0, + "content": "Two settings. There are two settings of InsDet, which differ in whether object instances are pre-defined during training or defined on the fly in testing time." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.44, + 0.485, + 0.589 + ], + "angle": 0, + "content": "- Conventional Instance Detection (CID). In this setting, all object instances of interest are pre-defined during training [9]. That said, one can use their visual references as training data to train detectors. Yet, only visual references are available and testing scene data is still unknown. Hence, InsDet methods in CID often choose to sample diverse background images in the open world [9, 12]. This setting simulates application scenarios in AR/VR. For instance, robots should memorize the items concerned by the customer so as to provide better customized services." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.59, + 0.486, + 0.726 + ], + "angle": 0, + "content": "- Novel Instance Detection (NID). This setting requires InsDet models to detect novel object instances defined only in testing time and does not allow the models to be finetuned further during testing [26]. In this setting, the de facto practice to train InsDet models turns to external data, which can be either synthetic data or those sampled in the open world. This setting simulates some real-world scenarios, for instance, robots must search for a never-before-seen luggage of a customer at an airport." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.44, + 0.486, + 0.726 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.734, + 0.458, + 0.75 + ], + "angle": 0, + "content": "3.2. Learning Instance Detector in Open World" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.756, + 0.483, + 0.801 + ], + "angle": 0, + "content": "We present our solution of instance detection from an open-world perspective (IDOW), focusing on foundation model adaptation and data augmentation." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.807, + 0.35, + 0.822 + ], + "angle": 0, + "content": "3.2.1. Foundation Model Adaptation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.485, + 0.901 + ], + "angle": 0, + "content": "As an open-world detector yields quite high recall [26, 48], we focus on adapting a foundation model to enhance feature representations for InsDet. Although a foundational feature model (e.g., DINOv2 [38]) offers feature representations applicable to various downstream tasks, it is not tai-" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.363, + 0.907, + 0.499 + ], + "angle": 0, + "content": "lored to InsDet. To improve feature representation, one can finetune this FM over relevant data, such as those sampled in the open world [26]. Intuitively, the finetuned feature is expected to match proposals and visual references better if they are from the same object instances; otherwise, distinguish them. Therefore, we propose to use a metric learning loss for FM adaptation. Moreover, we sample distractors from imagery in the open world. The distractors serve as negatives in metric learning." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.501, + 0.909, + 0.713 + ], + "angle": 0, + "content": "We denote an FM as \\( f_{\\theta} \\), parameterized by \\( \\theta \\). This FM serves as the feature extractor \\( f_{\\theta}(\\mathbf{I}) : \\mathbf{I} \\to \\mathbb{R}^q \\) to produce \\( q \\)-dimensional features for any given visual reference image or proposal \\( \\mathbf{I} \\). We finetune \\( f_{\\theta}(\\cdot) \\) to better serve InsDet such that features of reference images from the same instance should present high similarity, otherwise low similarity. We achieve this by finetuning \\( f_{\\theta} \\) using a simple metric learning loss \\( \\ell \\). Specifically, we construct triplets of training data \\( (\\mathbf{I}_a, \\mathbf{I}_p, \\mathbf{I}_n) \\), where \\( \\mathbf{I}_a, \\mathbf{I}_p \\) and \\( \\mathbf{I}_n \\) represent an anchor visual reference, positive and negative samples, respectively. The training data contains examples sampled or synthesized in the open world [26], as well as visual references available in the CID setting. Negative examples are cross-instance visual examples and distractors (detailed later). Below is the loss:" + }, + { + "type": "equation", + "bbox": [ + 0.528, + 0.728, + 0.907, + 0.753 + ], + "angle": 0, + "content": "\\[\n\\ell = \\left[ d \\left(f _ {\\theta} \\left(\\mathbf {I} _ {a}\\right), f _ {\\theta} \\left(\\mathbf {I} _ {p}\\right)\\right) - d \\left(f _ {\\theta} \\left(\\mathbf {I} _ {a}\\right), f _ {\\theta} \\left(\\mathbf {I} _ {n}\\right)\\right) + \\alpha \\right] _ {+}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.909, + 0.903 + ], + "angle": 0, + "content": "where \\(\\alpha\\) is a hyper-parameter determining the margin between an instance and other negative data. \\(d(\\cdot)\\) measures the distance between two examples, e.g., inverse cosine similarity used in our work. After training, we use the learned features to represent both proposals and visual references. We follow the instance-proposal matching pipeline [48] to produce the final InsDet results: computing pairwise similarities between each proposal and each visual reference, running stable matching algorithm [10, 33], and returning" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9904" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.126, + 0.089, + 0.175, + 0.101 + ], + "angle": 0, + "content": "Support" + }, + { + "type": "image", + "bbox": [ + 0.1, + 0.103, + 0.187, + 0.237 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.249, + 0.089, + 0.332, + 0.1 + ], + "angle": 0, + "content": "Ground Truth" + }, + { + "type": "image", + "bbox": [ + 0.189, + 0.103, + 0.361, + 0.237 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.425, + 0.089, + 0.5, + 0.102 + ], + "angle": 0, + "content": "VoxDet [26]" + }, + { + "type": "image", + "bbox": [ + 0.369, + 0.103, + 0.539, + 0.237 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.593, + 0.089, + 0.676, + 0.102 + ], + "angle": 0, + "content": "OTS-FM [48]" + }, + { + "type": "image", + "bbox": [ + 0.547, + 0.103, + 0.718, + 0.237 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.765, + 0.089, + 0.85, + 0.101 + ], + "angle": 0, + "content": "IDOW (ours)" + }, + { + "type": "image", + "bbox": [ + 0.726, + 0.103, + 0.897, + 0.237 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.24, + 0.907, + 0.323 + ], + "angle": 0, + "content": "Figure 4. Visual comparison of InsDet results by different methods in the NID setting on RoboTools [26]. On the cluttered testing scene image, we mark the ground-truth and predictions using green and blue boxes, respectively. We attach instance IDs to them to highlight whether the instance recognition is correct compared to the visual references (i.e., the leftmost references). Compared with VoxDet and OTS-FM, our IDOW detects more instances (see orange arrows) with better accuracy (see red arrows). Comparing OTS-FM and our IDOW suggests that, although having not seen the visual references during adaption in this NID setting, features by our IDOW are a better representation for InsDet than the off-the-shelf foundation model (i.e., DINOv2) used in OTS-FM." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.34, + 0.483, + 0.369 + ], + "angle": 0, + "content": "the matched pairs if they have similarity scores greater than a predefined threshold (0.4 as used in [48])." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.371, + 0.484, + 0.598 + ], + "angle": 0, + "content": "Hard example sampling. To construct training batches of triplets, it is important to effectively sample negative examples, easy negative examples do not provide sufficient gradients during training [50]. In this work, we adopt a batch-level hard negative sampling strategy. In each training batch, we first sample a reference image as an anchor \\(\\mathbf{I}_a\\). We then sample a random example from the same instance as the positive one \\(\\mathbf{I}_p\\). We finally sample a hard negative example from the union of the reference images of all the other instances and the distractors \\(S_{dt}\\) (explained in the next subsection). Mathematically, we find hard negative patch \\(\\mathbf{I}_{hn}\\) with the objective: \\(\\arg \\min_{\\mathbf{I}_n} d(f_\\theta(\\mathbf{I}_a), f_\\theta(\\mathbf{I}_n))\\), \\(\\mathbf{I}_n \\in \\{C(\\mathbf{I}_i) \\neq C(\\mathbf{I}_a)\\}\\), for \\(i = 1 \\dots N\\} \\cup S_{dt}\\), where \\(C(\\mathbf{I}_i)\\) represents the instance ID of the data \\(\\mathbf{I}_i\\), and \\(N\\) indicates the total number of visual reference images." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.609, + 0.373, + 0.623 + ], + "angle": 0, + "content": "3.2.2. The Proposed Data Augmentation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.628, + 0.484, + 0.658 + ], + "angle": 0, + "content": "To better adapt FM for instance-level feature matching, we introduce two data augmentation techniques below." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.66, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Distractor sampling aims to sample patches of random background images as universal negative data to all object instances. It is a data augmentation technique, expected to define the open space and help features better characterize meaningful object instances. Distractor sampling has been adopted in the literature. For example, open-set recognition samples distractors as a separate class, known as \"the-other-class\" or \"background-class\", to define the open space. In InsDet, the cut-paste-learn strategy [9] samples random background photos to define the background content. In this work, we run SAM [22] on random background photos and use the segments as distractors. One might worry about the possibility of sampled distractors being exactly some instances of interest and including them as negatives may destabilize model adaptation. Yet, given a large number of distractors, this rarely becomes an issue as demonstrated by the success" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.337, + 0.58, + 0.387 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.582, + 0.337, + 0.643, + 0.387 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.643, + 0.337, + 0.707, + 0.387 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.707, + 0.337, + 0.771, + 0.387 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.771, + 0.337, + 0.836, + 0.387 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.836, + 0.337, + 0.901, + 0.387 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.39, + 0.574, + 0.437 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.585, + 0.39, + 0.635, + 0.437 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.636, + 0.39, + 0.688, + 0.437 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.69, + 0.39, + 0.763, + 0.437 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.763, + 0.39, + 0.829, + 0.438 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.83, + 0.39, + 0.891, + 0.438 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.441, + 0.575, + 0.487 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.584, + 0.441, + 0.634, + 0.487 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.636, + 0.441, + 0.688, + 0.487 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.69, + 0.441, + 0.763, + 0.487 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.763, + 0.441, + 0.829, + 0.487 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.83, + 0.441, + 0.882, + 0.487 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.49, + 0.574, + 0.537 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.584, + 0.489, + 0.643, + 0.537 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.643, + 0.489, + 0.688, + 0.537 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.69, + 0.489, + 0.763, + 0.537 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.763, + 0.489, + 0.836, + 0.537 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.836, + 0.489, + 0.885, + 0.537 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.54, + 0.907, + 0.636 + ], + "angle": 0, + "content": "Figure 5. We use NeRF to synthesize novel-view object images to augment the limited given visual references. In this work, we train per-instance Zip-NeRF [1]. We visualize synthesized images at different angles (in row 2-4) together with the raw visual references (in the first row). The visual references are from HR-InsDet dataset [48], where the QR code is used for camera pose estimation. Overall, we find synthesized images show high visual quality." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.655, + 0.907, + 0.699 + ], + "angle": 0, + "content": "of various self-supervised learning methods [38] which may sample examples serving as both positive and negative data at the same time." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Novel-view synthesis. Each object instance might be represented by only a few visual examples, especially in the CID setting which provides limited real visual references for instances of interest. To augment available visual references for better adapting FM, we learn a NeRF [35] for each instance and use NeRF to synthesize novel-view images. In this process, we estimate relative camera poses (required by NeRF) using COLMAP [47], as commonly done in NeRF methods [1, 35] and InsDet approaches [26, 48]. Fig. 5 displays some random NeRF-generated novel-view images with good quality in visuals. It is worth noting that a NeRF needs to be trained only once on the visual references of an object instance. Once it is trained, we use it to generate" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9905" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.907, + 0.173 + ], + "angle": 0, + "content": "Table 1. Benchmarking results in the CID setting on the HR-InsDet dataset. We compare our IDOW with state-of-the-arts and make three salient conclusions. First, IDOW significantly outperforms previous methods, e.g., \\(\\mathrm{IDOW}_{\\mathrm{GroundingDINO}}\\) (57.01 AP) > OTS-FM_{GroundingDINO} (51.68 AP) > CPL_{DINO} (27.99 AP). This confirms the importance of addressing InsDet from the open-world perspective. Second, adapting FMs by our IDOW further boosts the performance by 5-7 AP, e.g., \\(\\mathrm{IDOW}_{\\mathrm{SAM}}\\) (48.75 AP) > OTS-FMSAM (41.61 AP). Third, IDOW and OTS-FM are applicable to different pretrained FMs and adopting stronger FMs achieves better performance, e.g., using GroundingDINO yields \\(>8\\) AP than SAM in IDOW." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.175, + 0.907, + 0.356 + ], + "angle": 0, + "content": "
MethodVenue & YearAP\\( \\mathbf{A}\\mathbf{P}_{50} \\)\\( \\mathbf{A}\\mathbf{P}_{75} \\)
avghardeasysmallmediumlarge
\\( CPL_{FasterRCNN} \\)[9, 45]NeurIPS 201519.5410.2623.755.0322.2037.9729.2123.26
\\( CPL_{RetinaNet} \\)[9, 27]ICCV 201722.2214.9226.495.4825.8042.7131.1924.98
\\( CPL_{CenterNet} \\)[9, 64]CVPR 201921.1211.8525.705.9024.1540.3832.7223.60
\\( CPL_{FCOS} \\)[9, 54]ICCV 201922.4013.2228.686.1726.4638.1332.8025.47
\\( CPL_{DINO} \\)[9, 59]ICLR 202327.9917.8932.6511.5131.6048.3539.6232.19
OTS-FMSAM[22, 48]NeurIPS 202341.6128.0347.5714.5845.8369.1449.1045.95
OTS-FMGroundingDINO[29, 48]NeurIPS 202351.6837.2358.7228.7958.5569.2262.5056.78
\\( IDOW_{SAM} \\)48.7532.0956.5020.7555.2673.4357.5954.06
\\( IDOW_{GroundingDINO} \\)57.0140.7464.3635.2562.9873.6469.3362.84
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.367, + 0.483, + 0.472 + ], + "angle": 0, + "content": "more visual references not only for training but for testing. In particular, such NeRF-synthesized novel-view images can be stored together with the original visual references in testing for proposal-instance matching. Doing so significantly improves InsDet performance (Table 3 and Fig. 7). To the best of our knowledge, our work is the first that exploits NeRF to synthesize data for InsDet." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.49, + 0.223, + 0.508 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.516, + 0.484, + 0.577 + ], + "angle": 0, + "content": "In this section, we first introduce the experimental setup, then evaluate our IDOW by comparing existing methods in both CID and NID settings, and finally conduct ablation studies to analyze our approach." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.589, + 0.281, + 0.606 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.612, + 0.483, + 0.748 + ], + "angle": 0, + "content": "Datasets. We use two recently published datasets in our experiments. HR-InsDet [48] is developed for the CID setting, containing 100 daily object instances, 160 high-resolution testing images from 14 indoor scenarios (see Fig. 3 for an example). RoboTools [26] is developed for the NID setting, containing 20 robotic tool instances and 1,581 testing images from 24 indoor scenarios (see Fig. 4 for an example). For a fair comparison, we use the OWID dataset released by [26] to adapt FM in the NID setting." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.749, + 0.483, + 0.869 + ], + "angle": 0, + "content": "Metrics. We use metrics adopted in both datasets mentioned above, including average precision (AP) at IoU thresholds from 0.5 to 0.95 with the step size 0.05 as the primary metric and also report \\(\\mathrm{AP}_{50}\\) and \\(\\mathrm{AP}_{75}\\) averaged over all instances with IoU threshold as 0.5 and 0.75, respectively. The HR-InsDet dataset contains tags for hard/easy testing scenes and small/medium/large object instances. We follow [48] to break down results on these tags." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Compared methods. We compare previous methods including CPL, OTS-FM, and VoxDet. These methods have" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.367, + 0.907, + 0.518 + ], + "angle": 0, + "content": "options to use different components (so do ours). For example, CPL can train a detector using one of the popular architectures Faster RCNN [45], RetinaNet [27], CenterNet [64], FCOS [54], and DINO [59]. OTS-FM can use SAM or GroundingDINO for proposal detection. We implement these methods by using different backbones to improve their performance (Table 1). We use subscripts to mark the backbones used in each method. Moreover, we compare more methods in the NID setting, including OS2D [39],DTOID [34], and \\(\\mathrm{OLN}_{\\mathrm{Corr}}\\) [20]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.519, + 0.909, + 0.685 + ], + "angle": 0, + "content": "Implementation details. We carry out all experiments on a single NVIDIA 3090 GPU and implement methods with PyTorch. Unless otherwise specified, our IDOW uses GroundingDINO [29] as the proposal detector and DINOv2 [38] as the feature foundation model to finetune. We use the pretrained [CLS] token to extract features and class embedding to compute similarity scores for matching. When finetuning, we set hyper-parameter \\(\\alpha = 0.5\\) in the metric learning loss. We use Adam optimizer with a learning rate 1e-3 and weight decay 0.5. We set the batch size as 100 and finetune the model for 10 epochs." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.697, + 0.725, + 0.712 + ], + "angle": 0, + "content": "4.2. Benchmarking Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.719, + 0.907, + 0.779 + ], + "angle": 0, + "content": "Quantitative results. We evaluate our method against prior approaches on the InsDet and RoboTools dataset. Table 1 and 2 list detailed results in the CID and NID settings, respectively. We summarize three salient observations:" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.909, + 0.885 + ], + "angle": 0, + "content": "First, approaching InsDet from the open-world perspective significantly outperforms previous methods. From both CID and NID settings, we find that, despite not being explicitly trained on the instances of interests, OTS-FM already achieves \\(>10\\) AP than traditional detector-based methods which particularly train on such instances. Adapting FMs by our IDOW boosts performance by 3-5 AP." + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.886, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Second, IDOW significantly outperforms state-of-the-art" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9906" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.1, + 0.092, + 0.357, + 0.288 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.368, + 0.091, + 0.627, + 0.289 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.639, + 0.092, + 0.897, + 0.289 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.29, + 0.908, + 0.334 + ], + "angle": 0, + "content": "Figure 6. t-SNE visualization of features learned by different methods. We compare our finetuned DINOv2 against a finetuned ImageNet-pretrained ResNet101 model and the baseline instance detector CPL with a FasterRCNN architecture. We color feature points w.r.t instance IDs and overlay star-marked visual references (zoom-in to see better). Visually, the finetuned DINOv2 extracts more discriminative features." + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.349, + 0.486, + 0.489 + ], + "angle": 0, + "content": "Table 2. Benchmarking results in the NID setting on RoboTools. We compare our IDOW with state-of-the-arts and make three salient conclusions. First, IDOW significantly outperforms previous methods, e.g., \\(\\mathrm{IDOW}_{\\mathrm{GroundingDINO}}\\) (59.4 AP) > OTS-FM_{GroundingDINO} (56.7 AP) > VoxDet (18.7 AP). This demonstrates the effectiveness of approaching InsDet from the open-world perspective. Second, adapting FMs by our IDOW further boosts the performance by 3-5 AP, e.g., \\(\\mathrm{IDOW}_{\\mathrm{SAM}}\\) (51.9 AP) > OTS-FMSAM (46.5 AP). Third, adopting stronger FMs achieves better performance, e.g., \\(\\mathrm{IDOW}_{\\mathrm{GroundingDINO}}\\) (59.4 AP) > \\(\\mathrm{IDOW}_{\\mathrm{SAM}}\\) (51.9 AP) in IDOW." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.493, + 0.485, + 0.641 + ], + "angle": 0, + "content": "
MethodVenue & YearAPAP50AP75
OS2D [39]ECCV 20202.96.52.0
DTOID [34]WACV 20213.69.02.0
OLNCorr. [20]RA-L 202214.418.115.7
VoxDet [26]NeurIPS 202318.723.620.5
OTS-FMSAM [48]NeurIPS 202346.555.950.4
OTS-FMGroundingDINO [48]NeurIPS 202356.764.859.0
IDOWSAM51.963.856.5
IDOWGroundingDINO59.467.861.8
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.656, + 0.484, + 0.701 + ], + "angle": 0, + "content": "InsDet methods, achieving over 5 AP higher on average than methods like OTS-FM in both CID and NID settings, demonstrating the effectiveness of our techniques." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.703, + 0.484, + 0.763 + ], + "angle": 0, + "content": "Third, IDOW is versatile and applicable to different open-world pretrained FMs. By comparing \\(\\mathrm{IDOW}_{\\mathrm{SAM}}\\) and \\(\\mathrm{IDOW}_{\\mathrm{GroundingDINO}}\\), we find that adopting stronger FMs (GroundingDINO vs. SAM) achieves better performance." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.765, + 0.486, + 0.902 + ], + "angle": 0, + "content": "Qualitative Results. Fig. 3 and 4 compares InsDet results by different methods in the CID and NID settings, respectively. We can see that our IDOW has higher precision and detects more wanted instances, which are small and under low lighting conditions. Fig. 6 uses tSNE to visualize features of visual references and detected proposals (in the HR-InsDet dataset) computed by our finetuned DINOv2 (ft-DINOv2) in the left panel, and other non-FM methods including a finetuned ImageNet-pretrained ResNet101 back" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.351, + 0.907, + 0.397 + ], + "angle": 0, + "content": "bone (ft-ImageNet-Res101) and CPL_FasterRCNN in the middle and right panels, respectively. Clearly, ft-DINOv2 matches proposals and references much better than other models." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.404, + 0.836, + 0.421 + ], + "angle": 0, + "content": "4.3. Ablation Study and Further Analysis" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.426, + 0.909, + 0.577 + ], + "angle": 0, + "content": "In Table 3, we ablate the proposed techniques of NeRF-synthesis and distractor sampling. Results show that adapting the FM DINOv2 greatly boosts InsDet performance (from 51.68 AP by the baseline to 53.94 AP). Importantly, including NeRF-synthesized images in both training and testing significantly improves the performance. Yet, using them in testing brings more performance gains than training (refer to the next study in Fig. 7). Eventually, applying our techniques boosts AP to 57.01. The supplemental material contains more quantitative and qualitative results." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.578, + 0.911, + 0.775 + ], + "angle": 0, + "content": "Following the above, we study the affects of using different number of NeRF-synthesized images in training and test. Fig. 7 depicts the results. (1) Adding NeRF-generated images helps foundation model adaptation for InsDet. Specifically, adding \\(<24\\) generated images boosts the performance, especially when 0 synthesized images are used in testing. However, adding \\(>24\\) generated images for training induces diminishing performance gains, likely because that the adapted FM overfits to artifacts in NeRF generated images. (2) Adopting more synthetic images in testing consistently improves performance. Yet, the performance by adding more (e.g., \\(>36\\)) in testing saturates. The peak performance occurs when 36 NeRF-synthesized images are used in testing." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.785, + 0.636, + 0.801 + ], + "angle": 0, + "content": "5. Discussions" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Broader Impacts. Approaching InsDet in the open-world has a significant impact on various downstream applications, such as robotics and AR/VR. Detecting object instances in RGB images is considered the first step of contemporary perception algorithms. A stronger InsDet model provides richer information for the following steps, such" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9907" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.907, + 0.187 + ], + "angle": 0, + "content": "Table 3. Ablation study of each strategy involved in our IDOW. We carry out the study in the CID setting on the HR-InsDet dataset. We use OTS-FM\\(_{\\text{GroundingDINO}}\\) as a baseline, over which we incrementally add each strategy. \"Train\" means foundation model adaptation through finetuning on the available data. DA denotes Data Augmentation with NeRF-generated novel-views; DS denotes Distractor Sampling. Results clearly demonstrate that all the four strategies help achieve better InsDet performance. Finetuning FMs on the given visual references enhances detection performance, cf. Train (53.94 AP) > baseline (51.68 AP). Moreover, NeRF-based data augmentation improves the final detection performance, particularly when used in testing, cf. Train+DA@Test (56.44 AP) > Train+DA@Train (54.48 AP) > baseline (51.68 AP). Lastly, applying distractor sampling (DS) improves the final performance further, cf. Train+DS (54.10 AP) > Train (53.94 AP)." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.189, + 0.907, + 0.389 + ], + "angle": 0, + "content": "
StrategiesAP\\( \\mathbf{A}\\mathbf{P}_{50} \\)\\( \\mathbf{A}\\mathbf{P}_{75} \\)
DA@TestTrainDA@TrainDSavghardeasysmallmediumlarge
baseline: OTS-FMGroundingDINO51.6837.2358.7228.7958.5569.2262.5056.78
✓53.9437.5461.5230.1860.9871.6065.1859.36
✓✓54.4838.3161.7131.1861.1072.5566.0659.98
✓✓54.1037.7361.6530.2761.1971.2765.3759.57
✓✓✓54.9238.2862.4132.0061.6174.2166.5860.50
✓55.2439.6462.8134.0061.2072.3866.9960.88
✓✓56.4439.7164.2434.7263.1374.4268.5062.31
✓✓✓56.9239.8564.8534.6662.7475.4868.8962.80
✓✓✓56.5139.7064.3634.3461.6575.8968.7162.38
✓✓✓✓57.0140.7464.3635.2562.9873.6469.3362.84
" + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.407, + 0.466, + 0.587 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.594, + 0.486, + 0.734 + ], + "angle": 0, + "content": "Figure 7. Analysis of the number of added NeRF-synthesized images in training and testing w.r.t AP improvement over the \"Train\" method (Table 3), which by default finetunes DINOv2 using the available visual references (24 per instance). We carry out this analysis in the CID setting on the HR-InsDet dataset. On the one hand, adding NeRF-synthesized images helps finetune FM for InsDet, but adding more (e.g., \\(>24\\)) hurts the performance, likely due to overfitting to the artifacts of NeRF-synthesized images. On the other hand, adding more NeRF images in testing consistently improves performance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.486, + 0.902 + ], + "angle": 0, + "content": "as grasping or trajectory forecasting in robotics. Additionally, our work shows the promising performance of adapting FMs to InsDet, which will foster future research in studying different properties of FMs, such as adapting FMs to other tasks. Nonetheless, our research utilizes FMs that are potentially pretrained on private data. At this point, it is unclear how such FMs might have bias or ethical issues. As a result, we have no safeguard measure to ensure that the finetuned/adapted FMs are fair, even though we find no visible issues in the image-only dataset used for finetuning." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.408, + 0.907, + 0.469 + ], + "angle": 0, + "content": "Further, adopting SAM may result in missed detections of critical instances relevant to specific downstream applications, such as medical devices or other rare objects. Such potential issues are not addressed in this work." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.472, + 0.909, + 0.699 + ], + "angle": 0, + "content": "Limitations and future work. Experiments demonstrate that our proposed techniques effectively leverage FMs to address InsDet across various settings. Nonetheless, we note several limitations for future improvements. First, the current pipeline relies on the proposals generated from GroundingDINO. Despite its high recall, it is not yet ready for real-time applications. Future work should consider faster proposal detection methods. Second, the NeRF augmentation step is also slow since it requires training a NeRF for every instance, which takes 1 hour and 500MB space using one NVIDIA 3090 GPU. While developing efficient NeRF methods is beyond our scope, we point out that efficient NeRF methods can be readily used in our pipeline, such as [37] which trains a NeRF in seconds and [6] which can train a single NeRF for multiple objects." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.722, + 0.634, + 0.737 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Instance detection is highly applicable in various scenarios such as in robotics and AR/VR. We elaborate on its open-world nature, necessitating the development of InsDet models in the open world. We solve InsDet from the open world perspective, embracing foundation models (FMs) and data sampled in the open world. In particular, we propose simple and effective techniques such as metric learning, distractor sampling, and novel-view synthesis. Extensive experiments in different settings validate that our techniques in adapting FMs significantly boost InsDet performance." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9908" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.091, + 0.251, + 0.108 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.117, + 0.486, + 0.162 + ], + "angle": 0, + "content": "This work was supported by FDCT (0067/2024/ITP2), the University of Macau (SRG2023-00044-FST), the Institute of Collaborative Innovation and NSFC (No.62206256)." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.18, + 0.188, + 0.195 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.205, + 0.484, + 0.245 + ], + "angle": 0, + "content": "[1] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Zip-nerf: Anti-aliased grid-based neural radiance fields. In ICCV, 2023. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.249, + 0.484, + 0.275 + ], + "angle": 0, + "content": "[2] Herbert Bay, Tinne Tuytelaars, and Luc Van Gool. Surf: Speeded up robust features. In ECCV, 2006. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.279, + 0.484, + 0.32 + ], + "angle": 0, + "content": "[3] Richard Bormann, Xinjie Wang, Markus Volk, Kilian Kleeberger, and Jochen Lindermayr. Real-time instance detection with fast incremental learning. In ICRA. IEEE, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.323, + 0.484, + 0.376 + ], + "angle": 0, + "content": "[4] Mathilde Caron, Hugo Touvron, Ishan Misra, Herve Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.381, + 0.484, + 0.435 + ], + "angle": 0, + "content": "[5] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv:1512.03012, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.439, + 0.484, + 0.492 + ], + "angle": 0, + "content": "[6] Hansheng Chen, Jiatao Gu, Anpei Chen, Wei Tian, Zhuowen Tu, Lingjie Liu, and Hao Su. Single-stage diffusion nerf: A unified approach to 3d generation and reconstruction. In ICCV, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.497, + 0.484, + 0.537 + ], + "angle": 0, + "content": "[7] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In ICML, 2020. 3, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.541, + 0.484, + 0.609 + ], + "angle": 0, + "content": "[8] Jasmine Collins, Shubham Goel, Kenan Deng, Achleshwar Luthra, Leon Xu, Erhan Gundogdu, Xi Zhang, Tomas F Yago Vicente, Thomas Dideriksen, Himanshu Arora, et al. Abo: Dataset and benchmarks for real-world 3d object understanding. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.612, + 0.484, + 0.653 + ], + "angle": 0, + "content": "[9] Debidatta Dwibedi, Ishan Misra, and Martial Hebert. Cut, paste and learn: Surprisingly easy synthesis for instance detection. In ICCV, 2017. 1, 2, 3, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.657, + 0.484, + 0.696 + ], + "angle": 0, + "content": "[10] David Gale and Lloyd S Shapley. College admissions and the stability of marriage. The American Mathematical Monthly, 69(1):9-15, 1962. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.7, + 0.484, + 0.754 + ], + "angle": 0, + "content": "[11] Georgios Georgakis, Arsalan Mousavian, Alexander C. Berg, and Jana Kosecka. Synthesizing training data for object detection in indoor scenes. In Proceedings of Robotics: Science and Systems, 2017. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.758, + 0.484, + 0.811 + ], + "angle": 0, + "content": "[12] Georgios Georgakis, Arsalan Mousavian, Alexander C Berg, and Jana Košecká. Synthesizing training data for object detection in indoor scenes. Robotics: Science and Systems (RSS), 2017. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.816, + 0.484, + 0.856 + ], + "angle": 0, + "content": "[13] Sachin Goyal, Ananya Kumar, Sankalp Garg, Zico Kolter, and Aditi Raghunathan. Finetune like you pretrain: Improved finetuning of zero-shot vision models. In CVPR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.86, + 0.484, + 0.9 + ], + "angle": 0, + "content": "[14] Sunan He, Taian Guo, Tao Dai, Ruizhi Qiao, Chen Wu, Xiujun Shu, and Bo Ren. Vlmae: Vision-language masked autoencoder. arXiv:2208.09374, 2022. 3" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.205, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.133 + ], + "angle": 0, + "content": "[15] Dan Hendrycks, Mantas Mazeika, and Thomas Dietterich. Deep anomaly detection with outlier exposure. In ICLR, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.136, + 0.908, + 0.204 + ], + "angle": 0, + "content": "[16] Stefan Hinterstoisser, Cedric Cagniart, Slobodan Ilic, Peter Sturm, Nassir Navab, Pascal Fua, and Vincent Lepetit. Gradient response maps for real-time detection of textureless objects. IEEE transactions on pattern analysis and machine intelligence, 34(5):876-888, 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.205, + 0.907, + 0.273 + ], + "angle": 0, + "content": "[17] Tomás Hodań, Vibhav Vineet, Ran Gal, Emanuel Shalev, Jon Hanzelka, Treb Connell, Pedro Urbina, Sudipta N Sinha, and Brian Guenter. Photorealistic image synthesis for object instance detection. In IEEE international conference on image processing (ICIP), 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.275, + 0.906, + 0.328 + ], + "angle": 0, + "content": "[18] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc V. Le, Yunhsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In ICML, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.33, + 0.908, + 0.382 + ], + "angle": 0, + "content": "[19] Wadim Kehl, Fabian Manhardt, Federico Tombari, Slobodan Ilic, and Nassir Navab. Ssd-6d: Making rgb-based 3d detection and 6d pose estimation great again. In ICCV, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.386, + 0.906, + 0.439 + ], + "angle": 0, + "content": "[20] Dahun Kim, Tsung-Yi Lin, Anelia Angelova, In So Kweon, and Weicheng Kuo. Learning open-world object proposals without learning to classify. IEEE Robotics and Automation Letters, 7(2):5453-5460, 2022. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.442, + 0.907, + 0.508 + ], + "angle": 0, + "content": "[21] Eun Kyeong Kim, Hansoo Lee, Jin Yong Kim, and Sungshin Kim. Data augmentation method by applying color perturbation of inverse psnr and geometric transformations for object recognition based on deep learning. Applied Sciences, 10(11): 3755, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.511, + 0.908, + 0.565 + ], + "angle": 0, + "content": "[22] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In ICCV, 2023. 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.568, + 0.907, + 0.593 + ], + "angle": 0, + "content": "[23] Shu Kong and Deva Ramanan. Opengan: Open-set recognition via open data generation. In ICCV, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.596, + 0.908, + 0.635 + ], + "angle": 0, + "content": "[24] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. ImageNet classification with deep convolutional neural networks. NeurIPS, 25, 2012. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.637, + 0.906, + 0.663 + ], + "angle": 0, + "content": "[25] Kevin Lai, Liefeng Bo, and Dieter Fox. Unsupervised feature learning for 3d scene labeling. ICRA, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.665, + 0.907, + 0.704 + ], + "angle": 0, + "content": "[26] Bowen Li, Jiashun Wang, Yaoyu Hu, Chen Wang, and Sebastian Scherer. Voxdet: Voxel learning for novel instance detection. NeurIPS, 2024. 1, 2, 3, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.707, + 0.908, + 0.747 + ], + "angle": 0, + "content": "[27] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. In ICCV, 2017. 6, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.749, + 0.907, + 0.775 + ], + "angle": 0, + "content": "[28] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. NeurIPS, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.777, + 0.906, + 0.831 + ], + "angle": 0, + "content": "[29] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. ECCV, 2024. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.833, + 0.907, + 0.872 + ], + "angle": 0, + "content": "[30] Tian Liu, Huixin Zhang, Shubham Parashar, and Shu Kong. Few-shot recognition via stage-wise retrieval-augmented finetuning. In CVPR, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.907, + 0.901 + ], + "angle": 0, + "content": "[31] David G Lowe. Distinctive image features from scale-invariant keypoints. *IJCV*, 60:91-110, 2004. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9909" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.484, + 0.134 + ], + "angle": 0, + "content": "[32] Anish Madan, Neehar Peri, Shu Kong, and Deva Ramanan. Revisiting few-shot object detection with vision-language models. In NeurIPS Datasets & Benchmark Track, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.484, + 0.176 + ], + "angle": 0, + "content": "[33] David G McVitie and Leslie B Wilson. The stable marriage problem. Communications of the ACM, 14(7):486-490, 1971. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.179, + 0.482, + 0.22 + ], + "angle": 0, + "content": "[34] Jean-Philippe Mercier, Mathieu Garon, Philippe Giguere, and Jean-Francois Lalonde. Deep template-based object instance detection. In WACV, 2021. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.222, + 0.484, + 0.276 + ], + "angle": 0, + "content": "[35] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.279, + 0.484, + 0.345 + ], + "angle": 0, + "content": "[36] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.349, + 0.484, + 0.403 + ], + "angle": 0, + "content": "[37] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM transactions on graphics (TOG), 41(4):1-15, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.406, + 0.484, + 0.473 + ], + "angle": 0, + "content": "[38] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv:2304.07193, 2023. 2, 3, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.476, + 0.484, + 0.517 + ], + "angle": 0, + "content": "[39] Anton Osokin, Denis Sumin, and Vasily Lomakin. Os2d: One-stage one-shot object detection by matching anchor features. In ECCV, 2020. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.519, + 0.484, + 0.559 + ], + "angle": 0, + "content": "[40] Shubham Parashar, Zhiqiu Lin, Yanan Li, and Shu Kong. Prompting scientific names for zero-shot species recognition. In EMNLP, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.562, + 0.484, + 0.616 + ], + "angle": 0, + "content": "[41] Shubham Parashar, Zhiqiu Lin, Tian Liu, Xiangjue Dong, Yanan Li, Deva Ramanan, James Caverlee, and Shu Kong. The neglected tails in vision-language models. In CVPR, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.619, + 0.484, + 0.687 + ], + "angle": 0, + "content": "[42] Samira Pouyanfar, Saad Sadiq, Yilin Yan, Haiman Tian, Yudong Tao, Maria Presa Reyes, Mei-Ling Shyu, Shu-Ching Chen, and Sundaraja S Iyengar. A survey on deep learning: Algorithms, techniques, and applications. ACM computing surveys (CSUR), 51(5):1-36, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.689, + 0.484, + 0.729 + ], + "angle": 0, + "content": "[43] A Quadros, James Patrick Underwood, and Bertrand Douillard. An occlusion-aware feature for range images. In ICRA, 2012. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.732, + 0.484, + 0.801 + ], + "angle": 0, + "content": "[44] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.803, + 0.484, + 0.844 + ], + "angle": 0, + "content": "[45] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. In NeurIPS, 2015. 6, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.484, + 0.9 + ], + "angle": 0, + "content": "[46] Tianhe Ren, Shilong Liu, Ailing Zeng, Jing Lin, Kunchang Li, He Cao, Jiayu Chen, Xinyu Huang, Yukang Chen, Feng Yan, et al. Grounded sam: Assembling open-world models for diverse visual tasks. arXiv:2401.14159, 2024. 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.12 + ], + "angle": 0, + "content": "[47] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-Motion Revisited. In CVPR, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.908, + 0.176 + ], + "angle": 0, + "content": "[48] Qianqian Shen, Yunhan Zhao, Nahyun Kwon, Jeeun Kim, Yanan Li, and Shu Kong. A high-resolution dataset for instance detection with multi-view object capture. In NeurIPS Datasets & Benchmark Track, 2023. 1, 2, 3, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.908, + 0.233 + ], + "angle": 0, + "content": "[49] Jia Shi, Gautam Gare, Jinjin Tian, Siqi Chai, Zhiqiu Lin, Arun Vasudevan, Di Feng, Francesco Ferroni, and Shu Kong. Lcaon-the-line: Benchmarking out-of-distribution generalization with class taxonomies. In ICML, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.235, + 0.908, + 0.276 + ], + "angle": 0, + "content": "[50] Abhinav Shrivastava, Abhinav Gupta, and Ross Girshick. Training region-based object detectors with online hard example mining. In CVPR, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.278, + 0.907, + 0.319 + ], + "angle": 0, + "content": "[51] Julio Silva-Rodriguez, Sina Hajimiri, Ismail Ben Ayed, and Jose Dolz. A closer look at the few-shot adaptation of large vision-language models. In CVPR, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.321, + 0.908, + 0.361 + ], + "angle": 0, + "content": "[52] Afrina Tabassum, Muntasir Wahed, Hoda Eldardiry, and Ismini Lourentzou. Hard negative sampling strategies for contrastive representation learning. arXiv:2206.01197, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.364, + 0.907, + 0.404 + ], + "angle": 0, + "content": "[53] Peng Tang, Xinggang Wang, Xiang Bai, and Wenyu Liu. Multiple instance detection network with online instance classifier refinement. In CVPR, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.406, + 0.907, + 0.433 + ], + "angle": 0, + "content": "[54] Zhi Tian, Chunhua Shen, Hao Chen, and Tong He. Fcos: Fully convolutional one-stage object detection. In ICCV, 2019. 6, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.435, + 0.907, + 0.462 + ], + "angle": 0, + "content": "[55] Hugo Touvron, Matthieu Cord, and Hervé Jégou. Deit iii: Revenge of the vit. In ECCV, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.464, + 0.906, + 0.505 + ], + "angle": 0, + "content": "[56] Riccardo Volpi, Pietro Morierio, Silvio Savarese, and Vittorio Murino. Adversarial feature augmentation for unsupervised domain adaptation. In CVPR, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.507, + 0.908, + 0.56 + ], + "angle": 0, + "content": "[57] Riccardo Volpi, Hongseok Namkoong, Ozan Sener, John C Duchi, Vittorio Murino, and Silvio Savarese. Generalizing to unseen domains via adversarial data augmentation. NeurIPS, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.563, + 0.907, + 0.591 + ], + "angle": 0, + "content": "[58] XuDong Wang, Jingfeng Yang, and Trevor Darrell. Segment anything without supervision. In NeurIPS, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.593, + 0.907, + 0.647 + ], + "angle": 0, + "content": "[59] Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel Ni, and Harry Shum. Dino: Detr with improved denoising anchor boxes for end-to-end object detection. In ICLR, 2022. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.649, + 0.907, + 0.676 + ], + "angle": 0, + "content": "[60] Xinyu Zhang, Qiang Wang, Jian Zhang, and Zhao Zhong. Adversarial autoaugment. In ICLR, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.678, + 0.907, + 0.718 + ], + "angle": 0, + "content": "[61] Yunhan Zhao, Shu Kong, and Charless Fowlkes. Camera pose matters: Improving depth prediction by mitigating pose distribution bias. In CVPR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.72, + 0.907, + 0.761 + ], + "angle": 0, + "content": "[62] Yunhan Zhao, Haoyu Ma, Shu Kong, and Charless Fowlkes. Instance tracking in 3d scenes from egocentric videos. In CVPR, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.763, + 0.906, + 0.804 + ], + "angle": 0, + "content": "[63] Jinghao Zhou, Chen Wei, Huiyu Wang, Wei Shen, Cihang Xie, Alan Yuille, and Tao Kong. Image bert pre-training with online tokenizer. In ICLR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.806, + 0.906, + 0.834 + ], + "angle": 0, + "content": "[64] Xingyi Zhou, Dequan Wang, and Philipp Krahenbuhl. Objects as points. In arXiv:1904.07850, 2019. 6, 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.834 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "9910" + } + ] +] \ No newline at end of file diff --git a/2025/Solving Instance Detection from an Open-World Perspective/ab6d1961-27f9-4a61-af19-064b80505b12_origin.pdf b/2025/Solving Instance Detection from an Open-World Perspective/ab6d1961-27f9-4a61-af19-064b80505b12_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..1d41de7a1d9d7ceb8461361f685cb88c04a0ff3a --- /dev/null +++ b/2025/Solving Instance Detection from an Open-World Perspective/ab6d1961-27f9-4a61-af19-064b80505b12_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0853b23cf2923a855841122be3ef867031a8ac544ac43f0082eb91efc38fdfa3 +size 2182926 diff --git a/2025/Solving Instance Detection from an Open-World Perspective/full.md b/2025/Solving Instance Detection from an Open-World Perspective/full.md new file mode 100644 index 0000000000000000000000000000000000000000..22a63a93cde920086566d6a3a4a29ae7a314fd2c --- /dev/null +++ b/2025/Solving Instance Detection from an Open-World Perspective/full.md @@ -0,0 +1,344 @@ +# Solving Instance Detection from an Open-World Perspective + +Qianqian Shen $^{1,*}$ Yunhan Zhao $^{2,*}$ Nahyun Kwon $^{3}$ Jeeun Kim $^{3}$ Yanan Li $^{4,\dagger}$ Shu Kong $^{5,6,\dagger}$ + +1Zhejiang University, 2UC Irvine, 3Texas A&M University, 4Zhejiang Lab + +5University of Macau, 6Institute of Collaborative Innovation + +website and code: https://shenqq377.github.io/IDOW + +# Abstract + +Instance detection (InsDet) aims to localize specific object instances within a novel scene imagery based on given visual references. Technically, it requires proposal detection to identify all possible object instances, followed by instance-level matching to pinpoint the ones of interest. Its open-world nature supports its broad applications from robotics to AR/VR but also presents significant challenges: methods must generalize to unknown testing data distributions because (1) the testing scene imagery is unseen during training, and (2) there are domain gaps between visual references and detected proposals. Existing methods tackle these challenges by synthesizing diverse training examples or utilizing off-the-shelf foundation models (FMs). However, they only partially capitalize the available open-world information. In contrast, we approach InsDet from an Open-World perspective, introducing our method IDOW. We find that, while pretrained FMs yield high recall in instance detection, they are not specifically optimized for instance-level feature matching. Therefore, we adapt pretrained FMs for improved instance-level matching using open-world data. Our approach incorporates metric learning along with novel data augmentations, which sample distractors as negative examples and synthesize novel-view instances to enrich the visual references. Extensive experiments demonstrate that our method significantly outperforms prior works, achieving $>10$ AP over previous results on two recently released challenging benchmark datasets in both conventional and novel instance detection settings. + +# 1. Introduction + +Instance detection (InsDet) aims to localize object instances of interest in novel scene imagery (Fig. 1), where these objects are specified by some visual references (as known as support templates) [3, 26, 34, 48, 53]. Its open-world nature supports its wide-ranging applications + +visual references + +![](images/4dbdf2596cf884036f25bbb68bc704f0e6a57fb09760fcef5952eb6840d1de6c.jpg) +Figure 1. Instance Detection (InsDet) aims to localize specific object instances in novel scene imagery based on given visual references (aka support templates). It is a challenging problem due to its open-world nature: the testing scene images are unseen during training and thus are unknown to InsDet models, and visual references and detected proposals have domain gaps (e.g., due to occlusions and arbitrary lighting conditions in the latter). + +novel testing scene imagery + +![](images/e13db6646b9f7799e8a6d6bedf3d1c112968821e5e898db2ea44acdb8be3f68c.jpg) + +![](images/939f085b0a352c97ec1ca4bcaeece617fe2eb30f6cba3f06694e04dc3c815409.jpg) + +in robotics and AR/VR. For example, a robot can be commanded to search for a customer's suitcase at an airport [26], or my_keys in my cluttered bedroom [62]. Its open-world nature also presents significant challenges: during training, one has no knowledge of testing data distribution, which can be instantiated by the unknown scene imagery [9], novel object instances encountered only in testing [26], and domain gaps between visual references and detected proposals [48]. + +Status quo. The literature of InsDet has two settings and existing InsDet methods attempt to address the open world in different aspects (Fig. 2). The conventional instance detection (CID) setting provides visual references of object instances for training. Yet, the testing scene is still unknown in training time. Therefore, early method Cut Paste Learn (CPL) collects random background images to model diverse testing scenes (Fig. 2a), cuts instances from the visual references, and pastes them on these background images to construct a synthetic training set [9, 11]. It is worth noting that background image sampling is utilizing open-world information in a way of data collection in the open world. On the other hand, the novel instance detection (NID) setting, introduced recently [26], studies detecting novel object instances that are specified only in testing time where a trained model is not allowed to be tuned further. To approach NID, VoxDet [26] uses large-scale synthetic data (Fig. 2b) to learn + +![](images/fea24deb0d5969838533254bf5891ec288951e87511a7e67116d10d516cd3aaf.jpg) +(a) background imagery [9, 11] +Figure 2. Existing InsDet methods leverage the open-world information in different aspects, such as (a) background image sampling (from the open world) to synthesize training data [9, 11], (b) object image sampling (from the open world) to learn feature representations [26], and (c) foundation model utilization (pretrained in the open world) for proposal detection and instance-level feature matching [48]. (d) As FMs are not specifically designed for instance-level feature matching required by InsDet, we propose to adapt them by leveraging rich data sampled from the open world. We gather data from multiple sources: (1) any available visual references of instances in the CID setting, (2) abundant multi-view object images sampled in the open world similar to (b), (3) synthetic data by training NeRF [1, 36] to generate novel-view images based on the given instances, (4) distractors by running FMs (esp. SAM [22]) on random open-world imagery to generate random object-like proposals (Fig. 5). We use the data above to adapt FM through metric learning. The technical novelty of our work lies in (3) and (4), as well as the design choice of metric learning to adapt FMs for InsDet. + +![](images/a82e0ac888d6ea996040eb2e268f525d5594eed2795a249564d9a7eff4ddf182.jpg) +(b) object images [26] +(c) Foundation Models [48] + +![](images/05b8064b4eee9ec5218df50a58fa07d35e8e2b29b97a29926a4cbfdc025b5ffb.jpg) + +![](images/b8e2450a827beaff13567c6ac5add694cbb0ca36efe16ea321df09abb650aa07.jpg) +(d) Solving Instance Detection in the Open World (ours) + +3D voxel representations from visual references and a matching function between 2D proposals and 3D voxels. Recent work OTS-FM [48] uses off-the-shelf pretrained foundation models (Fig. 2c) for InsDet. Specifically, it uses SAM [22] to detect object proposals in scene imagery, and DINOv2 [38] to represent visual references and proposals for instance-level feature matching. OTS-FM is a non-learned method and hence is applicable in both CID and NID settings. Importantly, it achieves the state-of-the-art performance on the recently released benchmark HR-InsDet [48]. + +The open-world perspective. As illustrated in Fig. 1, the challenges of InsDet lie in its open-world nature: the unknown testing data distribution, and domain gaps between visual references and detected proposals. To address the open-world challenges, existing methods take different strategies to exploit the open-world information (Fig. 2a-c). For example, to address the unknown testing scene imagery, some methods sample diverse background images [9, 11], demonstrating a practice of sampling data in the open world. To mitigate the domain gap between detected proposals and visual references, [26] attempts to learn features for instance-level matching by utilizing external datasets (ShapeNet [5] and ABO [8]). This data utilization can be considered as data collection in the open world. For better proposal detection and more discriminative feature representations, [48] adopts off-the-shelf pretrained foundation models (FMs) SAM [22] and DINOv2 [38], respectively. One can think of FMs as being pretrained in the open world. Different from existing methods which have not fully capitalized the open-world information, we take the open-world perspective as the first principle and solve InsDet by exploiting the Open World (Fig. 2d). We call our approach IDOW. + +Insights and novelties. Our insights are rooted in the understanding of InsDet's open-world nature and the practice of existing methods. Existing methods address open-world issues to some extent (Fig. 2), particularly by sampling data in the open world [9, 23, 26] and adopting foundation mod + +els (FMs) pretrained in the open world [48]. These motivate us to develop InsDet methods in the open world (IDOW). We find that open-world detectors yield high recall on object instances (see precision-recall curves in Fig. 9 of the Appendix) but FMs are not tailored to instance-level matching between proposals and visual references – directly using an FM for InsDet is sub-optimal. Therefore, we adapt FMs to instance-level feature representation by metric learning. We propose novel and simple techniques for data augmentation to enhance FM adaptation and feature learning: distractor sampling, and multi-view synthesis. The former helps learn more discriminative features for instance comparison, and the latter enhances visual references particularly when they are few. With these techniques, our IDOW achieves significant boosts over previously reported results on recently released challenging benchmarks (Table 1 and 2). + +# Contributions. We make three major contributions: + +1. We elaborate on the open-world challenge of InsDet and motivate the necessity of developing methods in the open world. We solve InsDet by exploiting diverse open data available in the open world and FMs pretrained therein. +2. We introduce simple and effective techniques to adapt FMs to InsDet: a metric learning loss, and data augmentation with distractor sampling and novel-view synthesis. +3. We extensively compare our approach with existing ones on two recently released datasets, demonstrating that our approach outperforms them by $>10$ AP in both conventional and novel instance detection settings. + +# 2. Related Work + +Instance detection. Early methods use local features such as SIFT [31] and SURF [2] to match visual references [43] and image regions to localize instances [16]. Recent methods train deep neural networks for instance-agnostic proposal detection and proposal-instance matching, achieving significant improvements [26, 34, 48]. In the con + +ventional instance detection (CID) setting, visual references of object instances are given during training. To approach CID, prevalent method synthesize training data in a cut-paste-learn (CPL) strategy [9, 11]: cutting instances from visual references and pasting them on random background images, then learning a detector on such synthetic data. Other works strive to obtain more training samples by rendering realistic instance examples [17, 19], using data augmentation [9] and synthesizing better training images [9, 12, 25]. Moreover, recent literature introduces the novel instance detection (NID) setting [26], which requires detecting novel instances specified only in testing and does not allow for further model finetuning during testing. VoxDet [26] utilizes a pretrained open-world detector [20] for proposal detection. Nowadays, the open-world perspective advances proposal detection in a way of learning more generalizable detectors and carrying out open-vocabulary recognition [29]. While a foundational open-vocabulary detector can attach a textual description (e.g., a short phrase) to each proposal, simply using it is insufficient for instance-level matching. [48] realizes this issue although foundational detectors can yield nearly perfect recall. As a result, the recent method OTS-FM uses another FM DINOv2 [38] for proposal-instance matching [48]. Differently, we solve InsDet from the open-world perspective, exploiting open FMs and open data to learn more discriminative instance-level feature representations. + +Foundation models (FMs) are pre-trained on web-scale data, which can be thought of as being sampled in the open world. Different FMs are trained on different types of data and can shine in different downstream tasks related to visual perception, natural language processing, or both. As this work concerns about visual perception, we briefly review Vision-Language Models (VLMs) and Visual Foundation Models (VFMs). VLMs are pretrained on web-scale image-text paired data [18, 28, 29, 44], demonstrating impressive results in high-level tasks such as visual grounding and image captioning. VFMs are pretrained primarily on visual data [4, 7, 14, 22, 38, 55, 58, 63] and can yield impressive visual perception results such as proposal or open-set object detection [29, 59]. In InsDet, the mainstream framework requires proposal detection and visual feature matching. In this framework, open-world detectors appear to achieve high recall on the instances of interest, owing to that the openworld pretrained detectors generalize quite well in the open world [29, 46, 59]. In contrast, feature matching still remains a challenge in improving InsDet. Therefore, we focus on improving features for InsDet by adapting VFMs using data sampled in the open world. [48] shows that using an off-the-shelf FM DINOv2 [38] for feature matching greatly enhances InsDet detection; yet, FM adaptation has been still under-explored in InsDet, although it is a well studied in other areas such as few-shot perception [30, 32, 51] and zero-shot perception [13, 40, 41]. Our work introduces simple + +and novel techniques to adapt FMs to improve InsDet. + +The open-world data is publicly available for pretraining FMs. Coping with such data is also often challenging. For example, data between training and testing has distribution shifts or domain gaps [30, 49], as clearly instantiated in InsDet by the unknown testing scene imagery, in addition to the domain gaps between (clean) visual references and (occluded) proposals (Fig. 1). Speaking of distribution shifts, some existing methods sample open-world data, demonstrating a need to develop robust models in the open world [15, 23]. Existing InsDet methods also sample open-world data such as diverse background images and instance examples (Fig. 2) [9, 11]. Even so, while open-world data is abundant, most of it is not directly useful for training models for specific downstream tasks. In contrast, data augmentation aims to synthesize more tailored data to enhance models' robustness and generalization, including geometry-aware augmentation (e.g., random cropping and flipping) [7, 24, 61], photometric-aware augmentation (e.g., color jittering) [21, 56, 57], and robustness-aware augmentation (e.g., adversarial perturbation) [60]. It is worth noting that Neural Radiance Field (NeRF) [35], a recent rendering technique, can be used to synthesize novel-view images. In this work, we particularly adopt NeRF to synthesize novel-view visual references of object instances and use them to adapt FMs towards more tailored features to InsDet. To the best of our knowledge, using NeRF is under-explored in the literature of InsDet, but we find that doing so remarkably improves InsDet performance. + +# 3. Instance Detection: Protocols and Methods + +We first introduce the protocols of InsDet, including the problem definition, evaluation, and two settings. We then present our techniques to enhance foundation model adaptation to InsDet. + +# 3.1. Protocols + +Problem definition. Instance Detection (InsDet) aims to detect specific object instances from a 2D scene image (Fig. 1), where the objects of interest are defined by a "support set" of some visual references [9, 26, 48]. Previous works capture the visual references from multiple camera views, with QR code being pasted to help estimate camera poses (see visual references in Fig. 1). + +Challenges. InsDet is a challenging problem due to its open-world nature: one has zero knowledge about the testing data distribution which is instantiated by unknown testing scenes, never-before-seen matters, arbitrary occlusions, background clutters, etc. In particular, object instances of interest can be novel and are only defined during testing, as emphasized by the novel instance detection setting below. Nevertheless, InsDet models must be capable of detecting + +![](images/9acc9f19f1e8fb0dd2d830ffbfda8222b4ad67ec9bb3fecef2c70d7870232a39.jpg) +Figure 3. Visual comparison of InsDet results by different methods in the CID setting on HR-InsDet [48]. The testing scene image contains sparse placement of small instances with challenging illumination. We mark the ground-truth and predictions using green and blue boxes, respectively. We attach instance IDs to them to highlight whether the instance recognition is correct compared to the visual references (i.e., the leftmost references). Compared with Cut-Paste-Learn and OTS-FM, our IDOW detects more instances (see orange arrows) with better accuracy (see red arrows). Comparison between OTS-FM and IDOW suggests that adapting the FM DINOv2 yields better features for IDOW to perform much better in front of domain shifts, i.e., challenging illumination conditions in this testing scene image. + +![](images/18b7a33130cc1bee7ed8b0187c08c3b38d9123636647441d95516d430266c1bf.jpg) + +![](images/d735d4105590ba8f0f98a663f6eceb51a9b7c9771272d9a40e666cae9c994d2b.jpg) + +![](images/2aa413d1b18d69f4f0f452a819509693153e0ed990c2dfc2599396608187c941.jpg) + +any object instances of interest and robustly comparing them with visual references. + +Two settings. There are two settings of InsDet, which differ in whether object instances are pre-defined during training or defined on the fly in testing time. + +- Conventional Instance Detection (CID). In this setting, all object instances of interest are pre-defined during training [9]. That said, one can use their visual references as training data to train detectors. Yet, only visual references are available and testing scene data is still unknown. Hence, InsDet methods in CID often choose to sample diverse background images in the open world [9, 12]. This setting simulates application scenarios in AR/VR. For instance, robots should memorize the items concerned by the customer so as to provide better customized services. +- Novel Instance Detection (NID). This setting requires InsDet models to detect novel object instances defined only in testing time and does not allow the models to be finetuned further during testing [26]. In this setting, the de facto practice to train InsDet models turns to external data, which can be either synthetic data or those sampled in the open world. This setting simulates some real-world scenarios, for instance, robots must search for a never-before-seen luggage of a customer at an airport. + +# 3.2. Learning Instance Detector in Open World + +We present our solution of instance detection from an open-world perspective (IDOW), focusing on foundation model adaptation and data augmentation. + +# 3.2.1. Foundation Model Adaptation + +As an open-world detector yields quite high recall [26, 48], we focus on adapting a foundation model to enhance feature representations for InsDet. Although a foundational feature model (e.g., DINOv2 [38]) offers feature representations applicable to various downstream tasks, it is not tai- + +lored to InsDet. To improve feature representation, one can finetune this FM over relevant data, such as those sampled in the open world [26]. Intuitively, the finetuned feature is expected to match proposals and visual references better if they are from the same object instances; otherwise, distinguish them. Therefore, we propose to use a metric learning loss for FM adaptation. Moreover, we sample distractors from imagery in the open world. The distractors serve as negatives in metric learning. + +We denote an FM as $f_{\theta}$ , parameterized by $\theta$ . This FM serves as the feature extractor $f_{\theta}(\mathbf{I}) : \mathbf{I} \to \mathbb{R}^q$ to produce $q$ -dimensional features for any given visual reference image or proposal $\mathbf{I}$ . We finetune $f_{\theta}(\cdot)$ to better serve InsDet such that features of reference images from the same instance should present high similarity, otherwise low similarity. We achieve this by finetuning $f_{\theta}$ using a simple metric learning loss $\ell$ . Specifically, we construct triplets of training data $(\mathbf{I}_a, \mathbf{I}_p, \mathbf{I}_n)$ , where $\mathbf{I}_a, \mathbf{I}_p$ and $\mathbf{I}_n$ represent an anchor visual reference, positive and negative samples, respectively. The training data contains examples sampled or synthesized in the open world [26], as well as visual references available in the CID setting. Negative examples are cross-instance visual examples and distractors (detailed later). Below is the loss: + +$$ +\ell = \left[ d \left(f _ {\theta} \left(\mathbf {I} _ {a}\right), f _ {\theta} \left(\mathbf {I} _ {p}\right)\right) - d \left(f _ {\theta} \left(\mathbf {I} _ {a}\right), f _ {\theta} \left(\mathbf {I} _ {n}\right)\right) + \alpha \right] _ {+}, \tag {1} +$$ + +where $\alpha$ is a hyper-parameter determining the margin between an instance and other negative data. $d(\cdot)$ measures the distance between two examples, e.g., inverse cosine similarity used in our work. After training, we use the learned features to represent both proposals and visual references. We follow the instance-proposal matching pipeline [48] to produce the final InsDet results: computing pairwise similarities between each proposal and each visual reference, running stable matching algorithm [10, 33], and returning + +![](images/c52f8bbd473a77799bf9b4d1de247f976a8f32707f718a714dfe22158720ae25.jpg) +Support + +![](images/2f118a273c6ba9e6288fa5a7f90048931795a1486c728c68eaaa8688d9b2040a.jpg) +Ground Truth + +![](images/2564597b30890f9ddc99bac46796810c4cfe428baeae9b3880b177c5eacf3e3c.jpg) +VoxDet [26] +Figure 4. Visual comparison of InsDet results by different methods in the NID setting on RoboTools [26]. On the cluttered testing scene image, we mark the ground-truth and predictions using green and blue boxes, respectively. We attach instance IDs to them to highlight whether the instance recognition is correct compared to the visual references (i.e., the leftmost references). Compared with VoxDet and OTS-FM, our IDOW detects more instances (see orange arrows) with better accuracy (see red arrows). Comparing OTS-FM and our IDOW suggests that, although having not seen the visual references during adaption in this NID setting, features by our IDOW are a better representation for InsDet than the off-the-shelf foundation model (i.e., DINOv2) used in OTS-FM. + +![](images/bffecc67dbe4d09cf7afd173469cbe8045cbc77f2aabe89b5312145f9a2c9b2c.jpg) +OTS-FM [48] + +![](images/b0ca2c049ef5a9441f7abc088463327a050887f5eed635b016e7553aaa67ecc8.jpg) +IDOW (ours) + +the matched pairs if they have similarity scores greater than a predefined threshold (0.4 as used in [48]). + +Hard example sampling. To construct training batches of triplets, it is important to effectively sample negative examples, easy negative examples do not provide sufficient gradients during training [50]. In this work, we adopt a batch-level hard negative sampling strategy. In each training batch, we first sample a reference image as an anchor $\mathbf{I}_a$ . We then sample a random example from the same instance as the positive one $\mathbf{I}_p$ . We finally sample a hard negative example from the union of the reference images of all the other instances and the distractors $S_{dt}$ (explained in the next subsection). Mathematically, we find hard negative patch $\mathbf{I}_{hn}$ with the objective: $\arg \min_{\mathbf{I}_n} d(f_\theta(\mathbf{I}_a), f_\theta(\mathbf{I}_n))$ , $\mathbf{I}_n \in \{C(\mathbf{I}_i) \neq C(\mathbf{I}_a)\}$ , for $i = 1 \dots N\} \cup S_{dt}$ , where $C(\mathbf{I}_i)$ represents the instance ID of the data $\mathbf{I}_i$ , and $N$ indicates the total number of visual reference images. + +# 3.2.2. The Proposed Data Augmentation + +To better adapt FM for instance-level feature matching, we introduce two data augmentation techniques below. + +Distractor sampling aims to sample patches of random background images as universal negative data to all object instances. It is a data augmentation technique, expected to define the open space and help features better characterize meaningful object instances. Distractor sampling has been adopted in the literature. For example, open-set recognition samples distractors as a separate class, known as "the-other-class" or "background-class", to define the open space. In InsDet, the cut-paste-learn strategy [9] samples random background photos to define the background content. In this work, we run SAM [22] on random background photos and use the segments as distractors. One might worry about the possibility of sampled distractors being exactly some instances of interest and including them as negatives may destabilize model adaptation. Yet, given a large number of distractors, this rarely becomes an issue as demonstrated by the success + +![](images/210ab2b74669701f38b34cded9c4adb572b83785ffad4cddd52c715be54d89dd.jpg) + +![](images/b672221f98d026a64c1900ff47d14121e0810d0dc3354bed182a875306a50e9a.jpg) + +![](images/cb5287d79ea04599ea5a1ac389160c0e4ee9e34a83bc5b731dc7b0b53db188ad.jpg) + +![](images/3de5432c7f80dc5c58ba739f8ab8165307d9e350d64aad10822dea7445f807c0.jpg) + +![](images/7118f2b44624c755ec6c59899245246ea3aec875a0529b81a7d1fe3512e01956.jpg) + +![](images/6c119278d1c3f1d98d9fbef146fc06ae216513b82ae2bbb1e8000e9001163d30.jpg) + +![](images/a81afc1f65cb57a2a4b85f6ea9718430dd8d5801eabbd7a001ecea5df7140387.jpg) + +![](images/35c1a7d54a5ef6af0836c5d0d601919bc1c7cc9c92900870df6ad555181b8a14.jpg) + +![](images/e444383cb9a8750cbc2a30e1af7c81d9f9eb2c6efddca6df33ae10e431d2d83b.jpg) + +![](images/57134a4754730b4882d93742904b59ee5869e3d6fed9d70cfd46237e3db1f0e1.jpg) + +![](images/dad18aa7f2248a0407ec92fec14ea9d085e12f1a4c08875ed2ba1d9be5804932.jpg) + +![](images/49f5260a048b6661ba0b6e12321bd67aec00f7250beb1e51277a9f0e2ab9ffd1.jpg) + +![](images/829d828f813648f8ae7a0f65200a4dddf3533c0ca52c977dc877b030c2e7fbd1.jpg) + +![](images/a5fcc0e5a9a3ad6d0b3065ebda2fbbc7cb33e688df95ebddfd9130d596833f7d.jpg) + +![](images/34d842752f58c2d6a72de7c6c01d99629fc1d36963f3c1a1742e5ef018262c22.jpg) + +![](images/de1e1669113de871305e5f24ef148c0b789ac126d238d24f700c9c489450d3b2.jpg) + +![](images/f40580b085c30e268001f17eb472427d5c06bd829dd5f4e655d0f8cdb7f1845c.jpg) + +![](images/b6ba97081f556ea8867104d01590efb35130fa1e4a13a14f1807e797d6744b6c.jpg) + +![](images/621451e5081d7528f487226a6ab565d338947f7b43248db89a92ea51f39f0959.jpg) +Figure 5. We use NeRF to synthesize novel-view object images to augment the limited given visual references. In this work, we train per-instance Zip-NeRF [1]. We visualize synthesized images at different angles (in row 2-4) together with the raw visual references (in the first row). The visual references are from HR-InsDet dataset [48], where the QR code is used for camera pose estimation. Overall, we find synthesized images show high visual quality. + +![](images/bb4f3f82f28371d4d03f0f8d1b9e6c8dbcea35d0537c1895015eaa76de69bed7.jpg) + +![](images/2227a2aea790a14e09cae9f7b8b6d0bddc810cc2fdbbf176ac0d4683fba10118.jpg) + +![](images/1552f025dba22887fa9fddef0fc9920007df0f52bd5d2806d6a65ff4b5b199cc.jpg) + +![](images/50ae24f9bd6d764b65a57adfc26143be676ae3158a864bf0e338e4ecddc04e6a.jpg) + +![](images/847c9b59247efcca4dd8995311cba223229c42966137d36b6d609e760de94cb5.jpg) + +of various self-supervised learning methods [38] which may sample examples serving as both positive and negative data at the same time. + +Novel-view synthesis. Each object instance might be represented by only a few visual examples, especially in the CID setting which provides limited real visual references for instances of interest. To augment available visual references for better adapting FM, we learn a NeRF [35] for each instance and use NeRF to synthesize novel-view images. In this process, we estimate relative camera poses (required by NeRF) using COLMAP [47], as commonly done in NeRF methods [1, 35] and InsDet approaches [26, 48]. Fig. 5 displays some random NeRF-generated novel-view images with good quality in visuals. It is worth noting that a NeRF needs to be trained only once on the visual references of an object instance. Once it is trained, we use it to generate + +Table 1. Benchmarking results in the CID setting on the HR-InsDet dataset. We compare our IDOW with state-of-the-arts and make three salient conclusions. First, IDOW significantly outperforms previous methods, e.g., $\mathrm{IDOW}_{\mathrm{GroundingDINO}}$ (57.01 AP) > OTS-FM_{GroundingDINO} (51.68 AP) > CPL_{DINO} (27.99 AP). This confirms the importance of addressing InsDet from the open-world perspective. Second, adapting FMs by our IDOW further boosts the performance by 5-7 AP, e.g., $\mathrm{IDOW}_{\mathrm{SAM}}$ (48.75 AP) > OTS-FMSAM (41.61 AP). Third, IDOW and OTS-FM are applicable to different pretrained FMs and adopting stronger FMs achieves better performance, e.g., using GroundingDINO yields $>8$ AP than SAM in IDOW. + +
MethodVenue & YearAP\( \mathbf{A}\mathbf{P}_{50} \)\( \mathbf{A}\mathbf{P}_{75} \)
avghardeasysmallmediumlarge
\( CPL_{FasterRCNN} \)[9, 45]NeurIPS 201519.5410.2623.755.0322.2037.9729.2123.26
\( CPL_{RetinaNet} \)[9, 27]ICCV 201722.2214.9226.495.4825.8042.7131.1924.98
\( CPL_{CenterNet} \)[9, 64]CVPR 201921.1211.8525.705.9024.1540.3832.7223.60
\( CPL_{FCOS} \)[9, 54]ICCV 201922.4013.2228.686.1726.4638.1332.8025.47
\( CPL_{DINO} \)[9, 59]ICLR 202327.9917.8932.6511.5131.6048.3539.6232.19
OTS-FMSAM[22, 48]NeurIPS 202341.6128.0347.5714.5845.8369.1449.1045.95
OTS-FMGroundingDINO[29, 48]NeurIPS 202351.6837.2358.7228.7958.5569.2262.5056.78
\( IDOW_{SAM} \)48.7532.0956.5020.7555.2673.4357.5954.06
\( IDOW_{GroundingDINO} \)57.0140.7464.3635.2562.9873.6469.3362.84
+ +more visual references not only for training but for testing. In particular, such NeRF-synthesized novel-view images can be stored together with the original visual references in testing for proposal-instance matching. Doing so significantly improves InsDet performance (Table 3 and Fig. 7). To the best of our knowledge, our work is the first that exploits NeRF to synthesize data for InsDet. + +# 4. Experiments + +In this section, we first introduce the experimental setup, then evaluate our IDOW by comparing existing methods in both CID and NID settings, and finally conduct ablation studies to analyze our approach. + +# 4.1. Experimental Setup + +Datasets. We use two recently published datasets in our experiments. HR-InsDet [48] is developed for the CID setting, containing 100 daily object instances, 160 high-resolution testing images from 14 indoor scenarios (see Fig. 3 for an example). RoboTools [26] is developed for the NID setting, containing 20 robotic tool instances and 1,581 testing images from 24 indoor scenarios (see Fig. 4 for an example). For a fair comparison, we use the OWID dataset released by [26] to adapt FM in the NID setting. + +Metrics. We use metrics adopted in both datasets mentioned above, including average precision (AP) at IoU thresholds from 0.5 to 0.95 with the step size 0.05 as the primary metric and also report $\mathrm{AP}_{50}$ and $\mathrm{AP}_{75}$ averaged over all instances with IoU threshold as 0.5 and 0.75, respectively. The HR-InsDet dataset contains tags for hard/easy testing scenes and small/medium/large object instances. We follow [48] to break down results on these tags. + +Compared methods. We compare previous methods including CPL, OTS-FM, and VoxDet. These methods have + +options to use different components (so do ours). For example, CPL can train a detector using one of the popular architectures Faster RCNN [45], RetinaNet [27], CenterNet [64], FCOS [54], and DINO [59]. OTS-FM can use SAM or GroundingDINO for proposal detection. We implement these methods by using different backbones to improve their performance (Table 1). We use subscripts to mark the backbones used in each method. Moreover, we compare more methods in the NID setting, including OS2D [39],DTOID [34], and $\mathrm{OLN}_{\mathrm{Corr}}$ [20]. + +Implementation details. We carry out all experiments on a single NVIDIA 3090 GPU and implement methods with PyTorch. Unless otherwise specified, our IDOW uses GroundingDINO [29] as the proposal detector and DINOv2 [38] as the feature foundation model to finetune. We use the pretrained [CLS] token to extract features and class embedding to compute similarity scores for matching. When finetuning, we set hyper-parameter $\alpha = 0.5$ in the metric learning loss. We use Adam optimizer with a learning rate 1e-3 and weight decay 0.5. We set the batch size as 100 and finetune the model for 10 epochs. + +# 4.2. Benchmarking Results + +Quantitative results. We evaluate our method against prior approaches on the InsDet and RoboTools dataset. Table 1 and 2 list detailed results in the CID and NID settings, respectively. We summarize three salient observations: + +First, approaching InsDet from the open-world perspective significantly outperforms previous methods. From both CID and NID settings, we find that, despite not being explicitly trained on the instances of interests, OTS-FM already achieves $>10$ AP than traditional detector-based methods which particularly train on such instances. Adapting FMs by our IDOW boosts performance by 3-5 AP. + +Second, IDOW significantly outperforms state-of-the-art + +![](images/6cf9eb521679c594896c30332d81f55203f31da2c04693e8d76effc39daf572d.jpg) +Figure 6. t-SNE visualization of features learned by different methods. We compare our finetuned DINOv2 against a finetuned ImageNet-pretrained ResNet101 model and the baseline instance detector CPL with a FasterRCNN architecture. We color feature points w.r.t instance IDs and overlay star-marked visual references (zoom-in to see better). Visually, the finetuned DINOv2 extracts more discriminative features. + +![](images/029a4bf98c259efb71d3b9a3329752f557e396db67ca6526b7b80f9bcaa44a4d.jpg) + +![](images/50bd78c6da05e1478ae947e40c67079c5050e9fbd893202cc4caea089b925398.jpg) + +Table 2. Benchmarking results in the NID setting on RoboTools. We compare our IDOW with state-of-the-arts and make three salient conclusions. First, IDOW significantly outperforms previous methods, e.g., $\mathrm{IDOW}_{\mathrm{GroundingDINO}}$ (59.4 AP) > OTS-FM_{GroundingDINO} (56.7 AP) > VoxDet (18.7 AP). This demonstrates the effectiveness of approaching InsDet from the open-world perspective. Second, adapting FMs by our IDOW further boosts the performance by 3-5 AP, e.g., $\mathrm{IDOW}_{\mathrm{SAM}}$ (51.9 AP) > OTS-FMSAM (46.5 AP). Third, adopting stronger FMs achieves better performance, e.g., $\mathrm{IDOW}_{\mathrm{GroundingDINO}}$ (59.4 AP) > $\mathrm{IDOW}_{\mathrm{SAM}}$ (51.9 AP) in IDOW. + +
MethodVenue & YearAPAP50AP75
OS2D [39]ECCV 20202.96.52.0
DTOID [34]WACV 20213.69.02.0
OLNCorr. [20]RA-L 202214.418.115.7
VoxDet [26]NeurIPS 202318.723.620.5
OTS-FMSAM [48]NeurIPS 202346.555.950.4
OTS-FMGroundingDINO [48]NeurIPS 202356.764.859.0
IDOWSAM51.963.856.5
IDOWGroundingDINO59.467.861.8
+ +InsDet methods, achieving over 5 AP higher on average than methods like OTS-FM in both CID and NID settings, demonstrating the effectiveness of our techniques. + +Third, IDOW is versatile and applicable to different open-world pretrained FMs. By comparing $\mathrm{IDOW}_{\mathrm{SAM}}$ and $\mathrm{IDOW}_{\mathrm{GroundingDINO}}$ , we find that adopting stronger FMs (GroundingDINO vs. SAM) achieves better performance. + +Qualitative Results. Fig. 3 and 4 compares InsDet results by different methods in the CID and NID settings, respectively. We can see that our IDOW has higher precision and detects more wanted instances, which are small and under low lighting conditions. Fig. 6 uses tSNE to visualize features of visual references and detected proposals (in the HR-InsDet dataset) computed by our finetuned DINOv2 (ft-DINOv2) in the left panel, and other non-FM methods including a finetuned ImageNet-pretrained ResNet101 back + +bone (ft-ImageNet-Res101) and CPL_FasterRCNN in the middle and right panels, respectively. Clearly, ft-DINOv2 matches proposals and references much better than other models. + +# 4.3. Ablation Study and Further Analysis + +In Table 3, we ablate the proposed techniques of NeRF-synthesis and distractor sampling. Results show that adapting the FM DINOv2 greatly boosts InsDet performance (from 51.68 AP by the baseline to 53.94 AP). Importantly, including NeRF-synthesized images in both training and testing significantly improves the performance. Yet, using them in testing brings more performance gains than training (refer to the next study in Fig. 7). Eventually, applying our techniques boosts AP to 57.01. The supplemental material contains more quantitative and qualitative results. + +Following the above, we study the affects of using different number of NeRF-synthesized images in training and test. Fig. 7 depicts the results. (1) Adding NeRF-generated images helps foundation model adaptation for InsDet. Specifically, adding $<24$ generated images boosts the performance, especially when 0 synthesized images are used in testing. However, adding $>24$ generated images for training induces diminishing performance gains, likely because that the adapted FM overfits to artifacts in NeRF generated images. (2) Adopting more synthetic images in testing consistently improves performance. Yet, the performance by adding more (e.g., $>36$ ) in testing saturates. The peak performance occurs when 36 NeRF-synthesized images are used in testing. + +# 5. Discussions + +Broader Impacts. Approaching InsDet in the open-world has a significant impact on various downstream applications, such as robotics and AR/VR. Detecting object instances in RGB images is considered the first step of contemporary perception algorithms. A stronger InsDet model provides richer information for the following steps, such + +Table 3. Ablation study of each strategy involved in our IDOW. We carry out the study in the CID setting on the HR-InsDet dataset. We use OTS-FM $_{\text{GroundingDINO}}$ as a baseline, over which we incrementally add each strategy. "Train" means foundation model adaptation through finetuning on the available data. DA denotes Data Augmentation with NeRF-generated novel-views; DS denotes Distractor Sampling. Results clearly demonstrate that all the four strategies help achieve better InsDet performance. Finetuning FMs on the given visual references enhances detection performance, cf. Train (53.94 AP) > baseline (51.68 AP). Moreover, NeRF-based data augmentation improves the final detection performance, particularly when used in testing, cf. Train+DA@Test (56.44 AP) > Train+DA@Train (54.48 AP) > baseline (51.68 AP). Lastly, applying distractor sampling (DS) improves the final performance further, cf. Train+DS (54.10 AP) > Train (53.94 AP). + +
StrategiesAP\( \mathbf{A}\mathbf{P}_{50} \)\( \mathbf{A}\mathbf{P}_{75} \)
DA@TestTrainDA@TrainDSavghardeasysmallmediumlarge
baseline: OTS-FMGroundingDINO51.6837.2358.7228.7958.5569.2262.5056.78
✓53.9437.5461.5230.1860.9871.6065.1859.36
✓✓54.4838.3161.7131.1861.1072.5566.0659.98
✓✓54.1037.7361.6530.2761.1971.2765.3759.57
✓✓✓54.9238.2862.4132.0061.6174.2166.5860.50
✓55.2439.6462.8134.0061.2072.3866.9960.88
✓✓56.4439.7164.2434.7263.1374.4268.5062.31
✓✓✓56.9239.8564.8534.6662.7475.4868.8962.80
✓✓✓56.5139.7064.3634.3461.6575.8968.7162.38
✓✓✓✓57.0140.7464.3635.2562.9873.6469.3362.84
+ +![](images/7460d61347fe8fa2529db2ae5da64ac653743cf68825e265c42b724202a0ce72.jpg) +Figure 7. Analysis of the number of added NeRF-synthesized images in training and testing w.r.t AP improvement over the "Train" method (Table 3), which by default finetunes DINOv2 using the available visual references (24 per instance). We carry out this analysis in the CID setting on the HR-InsDet dataset. On the one hand, adding NeRF-synthesized images helps finetune FM for InsDet, but adding more (e.g., $>24$ ) hurts the performance, likely due to overfitting to the artifacts of NeRF-synthesized images. On the other hand, adding more NeRF images in testing consistently improves performance. + +as grasping or trajectory forecasting in robotics. Additionally, our work shows the promising performance of adapting FMs to InsDet, which will foster future research in studying different properties of FMs, such as adapting FMs to other tasks. Nonetheless, our research utilizes FMs that are potentially pretrained on private data. At this point, it is unclear how such FMs might have bias or ethical issues. As a result, we have no safeguard measure to ensure that the finetuned/adapted FMs are fair, even though we find no visible issues in the image-only dataset used for finetuning. + +Further, adopting SAM may result in missed detections of critical instances relevant to specific downstream applications, such as medical devices or other rare objects. Such potential issues are not addressed in this work. + +Limitations and future work. Experiments demonstrate that our proposed techniques effectively leverage FMs to address InsDet across various settings. Nonetheless, we note several limitations for future improvements. First, the current pipeline relies on the proposals generated from GroundingDINO. Despite its high recall, it is not yet ready for real-time applications. Future work should consider faster proposal detection methods. Second, the NeRF augmentation step is also slow since it requires training a NeRF for every instance, which takes 1 hour and 500MB space using one NVIDIA 3090 GPU. While developing efficient NeRF methods is beyond our scope, we point out that efficient NeRF methods can be readily used in our pipeline, such as [37] which trains a NeRF in seconds and [6] which can train a single NeRF for multiple objects. + +# 6. Conclusion + +Instance detection is highly applicable in various scenarios such as in robotics and AR/VR. We elaborate on its open-world nature, necessitating the development of InsDet models in the open world. We solve InsDet from the open world perspective, embracing foundation models (FMs) and data sampled in the open world. In particular, we propose simple and effective techniques such as metric learning, distractor sampling, and novel-view synthesis. Extensive experiments in different settings validate that our techniques in adapting FMs significantly boost InsDet performance. + +# Acknowledgments + +This work was supported by FDCT (0067/2024/ITP2), the University of Macau (SRG2023-00044-FST), the Institute of Collaborative Innovation and NSFC (No.62206256). + +# References + +[1] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Zip-nerf: Anti-aliased grid-based neural radiance fields. In ICCV, 2023. 2, 5 +[2] Herbert Bay, Tinne Tuytelaars, and Luc Van Gool. Surf: Speeded up robust features. In ECCV, 2006. 2 +[3] Richard Bormann, Xinjie Wang, Markus Volk, Kilian Kleeberger, and Jochen Lindermayr. Real-time instance detection with fast incremental learning. In ICRA. IEEE, 2021. 1 +[4] Mathilde Caron, Hugo Touvron, Ishan Misra, Herve Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, 2021. 3 +[5] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv:1512.03012, 2015. 2 +[6] Hansheng Chen, Jiatao Gu, Anpei Chen, Wei Tian, Zhuowen Tu, Lingjie Liu, and Hao Su. Single-stage diffusion nerf: A unified approach to 3d generation and reconstruction. In ICCV, 2023. 8 +[7] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In ICML, 2020. 3, 1 +[8] Jasmine Collins, Shubham Goel, Kenan Deng, Achleshwar Luthra, Leon Xu, Erhan Gundogdu, Xi Zhang, Tomas F Yago Vicente, Thomas Dideriksen, Himanshu Arora, et al. Abo: Dataset and benchmarks for real-world 3d object understanding. In CVPR, 2022. 2 +[9] Debidatta Dwibedi, Ishan Misra, and Martial Hebert. Cut, paste and learn: Surprisingly easy synthesis for instance detection. In ICCV, 2017. 1, 2, 3, 4, 5, 6 +[10] David Gale and Lloyd S Shapley. College admissions and the stability of marriage. The American Mathematical Monthly, 69(1):9-15, 1962. 4 +[11] Georgios Georgakis, Arsalan Mousavian, Alexander C. Berg, and Jana Kosecka. Synthesizing training data for object detection in indoor scenes. In Proceedings of Robotics: Science and Systems, 2017. 1, 2, 3 +[12] Georgios Georgakis, Arsalan Mousavian, Alexander C Berg, and Jana Košecká. Synthesizing training data for object detection in indoor scenes. Robotics: Science and Systems (RSS), 2017. 3, 4 +[13] Sachin Goyal, Ananya Kumar, Sankalp Garg, Zico Kolter, and Aditi Raghunathan. Finetune like you pretrain: Improved finetuning of zero-shot vision models. In CVPR, 2023. 3 +[14] Sunan He, Taian Guo, Tao Dai, Ruizhi Qiao, Chen Wu, Xiujun Shu, and Bo Ren. Vlmae: Vision-language masked autoencoder. arXiv:2208.09374, 2022. 3 + +[15] Dan Hendrycks, Mantas Mazeika, and Thomas Dietterich. Deep anomaly detection with outlier exposure. In ICLR, 2019. 3 +[16] Stefan Hinterstoisser, Cedric Cagniart, Slobodan Ilic, Peter Sturm, Nassir Navab, Pascal Fua, and Vincent Lepetit. Gradient response maps for real-time detection of textureless objects. IEEE transactions on pattern analysis and machine intelligence, 34(5):876-888, 2011. 2 +[17] Tomás Hodań, Vibhav Vineet, Ran Gal, Emanuel Shalev, Jon Hanzelka, Treb Connell, Pedro Urbina, Sudipta N Sinha, and Brian Guenter. Photorealistic image synthesis for object instance detection. In IEEE international conference on image processing (ICIP), 2019. 3 +[18] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc V. Le, Yunhsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In ICML, 2021. 3 +[19] Wadim Kehl, Fabian Manhardt, Federico Tombari, Slobodan Ilic, and Nassir Navab. Ssd-6d: Making rgb-based 3d detection and 6d pose estimation great again. In ICCV, 2017. 3 +[20] Dahun Kim, Tsung-Yi Lin, Anelia Angelova, In So Kweon, and Weicheng Kuo. Learning open-world object proposals without learning to classify. IEEE Robotics and Automation Letters, 7(2):5453-5460, 2022. 3, 6, 7 +[21] Eun Kyeong Kim, Hansoo Lee, Jin Yong Kim, and Sungshin Kim. Data augmentation method by applying color perturbation of inverse psnr and geometric transformations for object recognition based on deep learning. Applied Sciences, 10(11): 3755, 2020. 3 +[22] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In ICCV, 2023. 2, 3, 5, 6 +[23] Shu Kong and Deva Ramanan. Opengan: Open-set recognition via open data generation. In ICCV, 2021. 2, 3 +[24] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. ImageNet classification with deep convolutional neural networks. NeurIPS, 25, 2012. 3 +[25] Kevin Lai, Liefeng Bo, and Dieter Fox. Unsupervised feature learning for 3d scene labeling. ICRA, 2014. 3 +[26] Bowen Li, Jiashun Wang, Yaoyu Hu, Chen Wang, and Sebastian Scherer. Voxdet: Voxel learning for novel instance detection. NeurIPS, 2024. 1, 2, 3, 4, 5, 6, 7 +[27] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. In ICCV, 2017. 6, 3 +[28] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. NeurIPS, 2023. 3 +[29] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. ECCV, 2024. 3, 6 +[30] Tian Liu, Huixin Zhang, Shubham Parashar, and Shu Kong. Few-shot recognition via stage-wise retrieval-augmented finetuning. In CVPR, 2025. 3 +[31] David G Lowe. Distinctive image features from scale-invariant keypoints. *IJCV*, 60:91-110, 2004. 2 + +[32] Anish Madan, Neehar Peri, Shu Kong, and Deva Ramanan. Revisiting few-shot object detection with vision-language models. In NeurIPS Datasets & Benchmark Track, 2024. 3 +[33] David G McVitie and Leslie B Wilson. The stable marriage problem. Communications of the ACM, 14(7):486-490, 1971. 4 +[34] Jean-Philippe Mercier, Mathieu Garon, Philippe Giguere, and Jean-Francois Lalonde. Deep template-based object instance detection. In WACV, 2021. 1, 2, 6, 7 +[35] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 3, 5 +[36] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 2 +[37] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM transactions on graphics (TOG), 41(4):1-15, 2022. 8 +[38] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv:2304.07193, 2023. 2, 3, 4, 5, 6 +[39] Anton Osokin, Denis Sumin, and Vasily Lomakin. Os2d: One-stage one-shot object detection by matching anchor features. In ECCV, 2020. 6, 7 +[40] Shubham Parashar, Zhiqiu Lin, Yanan Li, and Shu Kong. Prompting scientific names for zero-shot species recognition. In EMNLP, 2023. 3 +[41] Shubham Parashar, Zhiqiu Lin, Tian Liu, Xiangjue Dong, Yanan Li, Deva Ramanan, James Caverlee, and Shu Kong. The neglected tails in vision-language models. In CVPR, 2024. 3 +[42] Samira Pouyanfar, Saad Sadiq, Yilin Yan, Haiman Tian, Yudong Tao, Maria Presa Reyes, Mei-Ling Shyu, Shu-Ching Chen, and Sundaraja S Iyengar. A survey on deep learning: Algorithms, techniques, and applications. ACM computing surveys (CSUR), 51(5):1-36, 2018. 1 +[43] A Quadros, James Patrick Underwood, and Bertrand Douillard. An occlusion-aware feature for range images. In ICRA, 2012. 2 +[44] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021. 3 +[45] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. In NeurIPS, 2015. 6, 3 +[46] Tianhe Ren, Shilong Liu, Ailing Zeng, Jing Lin, Kunchang Li, He Cao, Jiayu Chen, Xinyu Huang, Yukang Chen, Feng Yan, et al. Grounded sam: Assembling open-world models for diverse visual tasks. arXiv:2401.14159, 2024. 3 + +[47] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-Motion Revisited. In CVPR, 2016. 5 +[48] Qianqian Shen, Yunhan Zhao, Nahyun Kwon, Jeeun Kim, Yanan Li, and Shu Kong. A high-resolution dataset for instance detection with multi-view object capture. In NeurIPS Datasets & Benchmark Track, 2023. 1, 2, 3, 4, 5, 6, 7 +[49] Jia Shi, Gautam Gare, Jinjin Tian, Siqi Chai, Zhiqiu Lin, Arun Vasudevan, Di Feng, Francesco Ferroni, and Shu Kong. Lcaon-the-line: Benchmarking out-of-distribution generalization with class taxonomies. In ICML, 2024. 3 +[50] Abhinav Shrivastava, Abhinav Gupta, and Ross Girshick. Training region-based object detectors with online hard example mining. In CVPR, 2016. 5 +[51] Julio Silva-Rodriguez, Sina Hajimiri, Ismail Ben Ayed, and Jose Dolz. A closer look at the few-shot adaptation of large vision-language models. In CVPR, 2024. 3 +[52] Afrina Tabassum, Muntasir Wahed, Hoda Eldardiry, and Ismini Lourentzou. Hard negative sampling strategies for contrastive representation learning. arXiv:2206.01197, 2022. 1 +[53] Peng Tang, Xinggang Wang, Xiang Bai, and Wenyu Liu. Multiple instance detection network with online instance classifier refinement. In CVPR, 2017. 1 +[54] Zhi Tian, Chunhua Shen, Hao Chen, and Tong He. Fcos: Fully convolutional one-stage object detection. In ICCV, 2019. 6, 3 +[55] Hugo Touvron, Matthieu Cord, and Hervé Jégou. Deit iii: Revenge of the vit. In ECCV, 2022. 3 +[56] Riccardo Volpi, Pietro Morierio, Silvio Savarese, and Vittorio Murino. Adversarial feature augmentation for unsupervised domain adaptation. In CVPR, 2018. 3 +[57] Riccardo Volpi, Hongseok Namkoong, Ozan Sener, John C Duchi, Vittorio Murino, and Silvio Savarese. Generalizing to unseen domains via adversarial data augmentation. NeurIPS, 2018. 3 +[58] XuDong Wang, Jingfeng Yang, and Trevor Darrell. Segment anything without supervision. In NeurIPS, 2024. 3 +[59] Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel Ni, and Harry Shum. Dino: Detr with improved denoising anchor boxes for end-to-end object detection. In ICLR, 2022. 3, 6 +[60] Xinyu Zhang, Qiang Wang, Jian Zhang, and Zhao Zhong. Adversarial autoaugment. In ICLR, 2020. 3 +[61] Yunhan Zhao, Shu Kong, and Charless Fowlkes. Camera pose matters: Improving depth prediction by mitigating pose distribution bias. In CVPR, 2021. 3 +[62] Yunhan Zhao, Haoyu Ma, Shu Kong, and Charless Fowlkes. Instance tracking in 3d scenes from egocentric videos. In CVPR, 2024. 1 +[63] Jinghao Zhou, Chen Wei, Huiyu Wang, Wei Shen, Cihang Xie, Alan Yuille, and Tao Kong. Image bert pre-training with online tokenizer. In ICLR, 2022. 3 +[64] Xingyi Zhou, Dequan Wang, and Philipp Krahenbuhl. Objects as points. In arXiv:1904.07850, 2019. 6, 3 \ No newline at end of file diff --git a/2025/Solving Instance Detection from an Open-World Perspective/images.zip b/2025/Solving Instance Detection from an Open-World Perspective/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..7b40c79c0d85d42cc1cc1ac518a943f201e6f014 --- /dev/null +++ b/2025/Solving Instance Detection from an Open-World Perspective/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1e8e2ea360a69076791f833782137853f551940dadd4852fa34af6daa7097d5a +size 662185 diff --git a/2025/Solving Instance Detection from an Open-World Perspective/layout.json b/2025/Solving Instance Detection from an Open-World Perspective/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..1c73943bb2570c6fb36864fb64d217fa9fc6a55c --- /dev/null +++ b/2025/Solving Instance Detection from an Open-World Perspective/layout.json @@ -0,0 +1,9150 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 119, + 103, + 492, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 119, + 103, + 492, + 121 + ], + "spans": [ + { + "bbox": [ + 119, + 103, + 492, + 121 + ], + "type": "text", + "content": "Solving Instance Detection from an Open-World Perspective" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 68, + 141, + 541, + 159 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 141, + 541, + 159 + ], + "spans": [ + { + "bbox": [ + 68, + 141, + 541, + 159 + ], + "type": "text", + "content": "Qianqian Shen" + }, + { + "bbox": [ + 68, + 141, + 541, + 159 + ], + "type": "inline_equation", + "content": "^{1,*}" + }, + { + "bbox": [ + 68, + 141, + 541, + 159 + ], + "type": "text", + "content": " Yunhan Zhao" + }, + { + "bbox": [ + 68, + 141, + 541, + 159 + ], + "type": "inline_equation", + "content": "^{2,*}" + }, + { + "bbox": [ + 68, + 141, + 541, + 159 + ], + "type": "text", + "content": " Nahyun Kwon" + }, + { + "bbox": [ + 68, + 141, + 541, + 159 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 68, + 141, + 541, + 159 + ], + "type": "text", + "content": " Jeeun Kim" + }, + { + "bbox": [ + 68, + 141, + 541, + 159 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 68, + 141, + 541, + 159 + ], + "type": "text", + "content": " Yanan Li" + }, + { + "bbox": [ + 68, + 141, + 541, + 159 + ], + "type": "inline_equation", + "content": "^{4,\\dagger}" + }, + { + "bbox": [ + 68, + 141, + 541, + 159 + ], + "type": "text", + "content": " Shu Kong" + }, + { + "bbox": [ + 68, + 141, + 541, + 159 + ], + "type": "inline_equation", + "content": "^{5,6,\\dagger}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 169, + 159, + 441, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 169, + 159, + 441, + 171 + ], + "spans": [ + { + "bbox": [ + 169, + 159, + 441, + 171 + ], + "type": "text", + "content": "1Zhejiang University, 2UC Irvine, 3Texas A&M University, 4Zhejiang Lab" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 194, + 173, + 413, + 184 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 194, + 173, + 413, + 184 + ], + "spans": [ + { + "bbox": [ + 194, + 173, + 413, + 184 + ], + "type": "text", + "content": "5University of Macau, 6Institute of Collaborative Innovation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 182, + 188, + 425, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 182, + 188, + 425, + 198 + ], + "spans": [ + { + "bbox": [ + 182, + 188, + 425, + 198 + ], + "type": "text", + "content": "website and code: https://shenqq377.github.io/IDOW" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "spans": [ + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 253, + 299, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 253, + 299, + 589 + ], + "spans": [ + { + "bbox": [ + 53, + 253, + 299, + 589 + ], + "type": "text", + "content": "Instance detection (InsDet) aims to localize specific object instances within a novel scene imagery based on given visual references. Technically, it requires proposal detection to identify all possible object instances, followed by instance-level matching to pinpoint the ones of interest. Its open-world nature supports its broad applications from robotics to AR/VR but also presents significant challenges: methods must generalize to unknown testing data distributions because (1) the testing scene imagery is unseen during training, and (2) there are domain gaps between visual references and detected proposals. Existing methods tackle these challenges by synthesizing diverse training examples or utilizing off-the-shelf foundation models (FMs). However, they only partially capitalize the available open-world information. In contrast, we approach InsDet from an Open-World perspective, introducing our method IDOW. We find that, while pretrained FMs yield high recall in instance detection, they are not specifically optimized for instance-level feature matching. Therefore, we adapt pretrained FMs for improved instance-level matching using open-world data. Our approach incorporates metric learning along with novel data augmentations, which sample distractors as negative examples and synthesize novel-view instances to enrich the visual references. Extensive experiments demonstrate that our method significantly outperforms prior works, achieving " + }, + { + "bbox": [ + 53, + 253, + 299, + 589 + ], + "type": "inline_equation", + "content": ">10" + }, + { + "bbox": [ + 53, + 253, + 299, + 589 + ], + "type": "text", + "content": " AP over previous results on two recently released challenging benchmark datasets in both conventional and novel instance detection settings." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 612, + 135, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 612, + 135, + 624 + ], + "spans": [ + { + "bbox": [ + 56, + 612, + 135, + 624 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 633, + 297, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 633, + 297, + 694 + ], + "spans": [ + { + "bbox": [ + 55, + 633, + 297, + 694 + ], + "type": "text", + "content": "Instance detection (InsDet) aims to localize object instances of interest in novel scene imagery (Fig. 1), where these objects are specified by some visual references (as known as support templates) [3, 26, 34, 48, 53]. Its open-world nature supports its wide-ranging applications" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 324, + 226, + 387, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 226, + 387, + 236 + ], + "spans": [ + { + "bbox": [ + 324, + 226, + 387, + 236 + ], + "type": "text", + "content": "visual references" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 317, + 237, + 394, + 326 + ], + "blocks": [ + { + "bbox": [ + 317, + 237, + 394, + 326 + ], + "lines": [ + { + "bbox": [ + 317, + 237, + 394, + 326 + ], + "spans": [ + { + "bbox": [ + 317, + 237, + 394, + 326 + ], + "type": "image", + "image_path": "4dbdf2596cf884036f25bbb68bc704f0e6a57fb09760fcef5952eb6840d1de6c.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 327, + 555, + 403 + ], + "lines": [ + { + "bbox": [ + 313, + 327, + 555, + 403 + ], + "spans": [ + { + "bbox": [ + 313, + 327, + 555, + 403 + ], + "type": "text", + "content": "Figure 1. Instance Detection (InsDet) aims to localize specific object instances in novel scene imagery based on given visual references (aka support templates). It is a challenging problem due to its open-world nature: the testing scene images are unseen during training and thus are unknown to InsDet models, and visual references and detected proposals have domain gaps (e.g., due to occlusions and arbitrary lighting conditions in the latter)." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 405, + 226, + 507, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 405, + 226, + 507, + 236 + ], + "spans": [ + { + "bbox": [ + 405, + 226, + 507, + 236 + ], + "type": "text", + "content": "novel testing scene imagery" + } + ] + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 397, + 237, + 517, + 326 + ], + "blocks": [ + { + "bbox": [ + 397, + 237, + 517, + 326 + ], + "lines": [ + { + "bbox": [ + 397, + 237, + 517, + 326 + ], + "spans": [ + { + "bbox": [ + 397, + 237, + 517, + 326 + ], + "type": "image", + "image_path": "e13db6646b9f7799e8a6d6bedf3d1c112968821e5e898db2ea44acdb8be3f68c.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 517, + 236, + 549, + 325 + ], + "blocks": [ + { + "bbox": [ + 517, + 236, + 549, + 325 + ], + "lines": [ + { + "bbox": [ + 517, + 236, + 549, + 325 + ], + "spans": [ + { + "bbox": [ + 517, + 236, + 549, + 325 + ], + "type": "image", + "image_path": "939f085b0a352c97ec1ca4bcaeece617fe2eb30f6cba3f06694e04dc3c815409.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "bbox": [ + 312, + 413, + 556, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 413, + 556, + 510 + ], + "spans": [ + { + "bbox": [ + 312, + 413, + 556, + 510 + ], + "type": "text", + "content": "in robotics and AR/VR. For example, a robot can be commanded to search for a customer's suitcase at an airport [26], or my_keys in my cluttered bedroom [62]. Its open-world nature also presents significant challenges: during training, one has no knowledge of testing data distribution, which can be instantiated by the unknown scene imagery [9], novel object instances encountered only in testing [26], and domain gaps between visual references and detected proposals [48]." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 510, + 557, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 557, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 557, + 714 + ], + "type": "text", + "content": "Status quo. The literature of InsDet has two settings and existing InsDet methods attempt to address the open world in different aspects (Fig. 2). The conventional instance detection (CID) setting provides visual references of object instances for training. Yet, the testing scene is still unknown in training time. Therefore, early method Cut Paste Learn (CPL) collects random background images to model diverse testing scenes (Fig. 2a), cuts instances from the visual references, and pastes them on these background images to construct a synthetic training set [9, 11]. It is worth noting that background image sampling is utilizing open-world information in a way of data collection in the open world. On the other hand, the novel instance detection (NID) setting, introduced recently [26], studies detecting novel object instances that are specified only in testing time where a trained model is not allowed to be tuned further. To approach NID, VoxDet [26] uses large-scale synthetic data (Fig. 2b) to learn" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 702, + 219, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 702, + 219, + 713 + ], + "spans": [ + { + "bbox": [ + 66, + 702, + 219, + 713 + ], + "type": "text", + "content": "*Equal contributions; †Corresponding authors." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "9901" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 78, + 149, + 159 + ], + "blocks": [ + { + "bbox": [ + 58, + 69, + 146, + 78 + ], + "lines": [ + { + "bbox": [ + 58, + 69, + 146, + 78 + ], + "spans": [ + { + "bbox": [ + 58, + 69, + 146, + 78 + ], + "type": "text", + "content": "(a) background imagery [9, 11]" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 60, + 78, + 149, + 159 + ], + "lines": [ + { + "bbox": [ + 60, + 78, + 149, + 159 + ], + "spans": [ + { + "bbox": [ + 60, + 78, + 149, + 159 + ], + "type": "image", + "image_path": "fea24deb0d5969838533254bf5891ec288951e87511a7e67116d10d516cd3aaf.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 161, + 555, + 259 + ], + "lines": [ + { + "bbox": [ + 55, + 161, + 555, + 259 + ], + "spans": [ + { + "bbox": [ + 55, + 161, + 555, + 259 + ], + "type": "text", + "content": "Figure 2. Existing InsDet methods leverage the open-world information in different aspects, such as (a) background image sampling (from the open world) to synthesize training data [9, 11], (b) object image sampling (from the open world) to learn feature representations [26], and (c) foundation model utilization (pretrained in the open world) for proposal detection and instance-level feature matching [48]. (d) As FMs are not specifically designed for instance-level feature matching required by InsDet, we propose to adapt them by leveraging rich data sampled from the open world. We gather data from multiple sources: (1) any available visual references of instances in the CID setting, (2) abundant multi-view object images sampled in the open world similar to (b), (3) synthetic data by training NeRF [1, 36] to generate novel-view images based on the given instances, (4) distractors by running FMs (esp. SAM [22]) on random open-world imagery to generate random object-like proposals (Fig. 5). We use the data above to adapt FM through metric learning. The technical novelty of our work lies in (3) and (4), as well as the design choice of metric learning to adapt FMs for InsDet." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 154, + 78, + 227, + 159 + ], + "blocks": [ + { + "bbox": [ + 155, + 69, + 219, + 78 + ], + "lines": [ + { + "bbox": [ + 155, + 69, + 219, + 78 + ], + "spans": [ + { + "bbox": [ + 155, + 69, + 219, + 78 + ], + "type": "text", + "content": "(b) object images [26]" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 154, + 78, + 227, + 159 + ], + "lines": [ + { + "bbox": [ + 154, + 78, + 227, + 159 + ], + "spans": [ + { + "bbox": [ + 154, + 78, + 227, + 159 + ], + "type": "image", + "image_path": "a82e0ac888d6ea996040eb2e268f525d5594eed2795a249564d9a7eff4ddf182.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 226, + 69, + 303, + 78 + ], + "lines": [ + { + "bbox": [ + 226, + 69, + 303, + 78 + ], + "spans": [ + { + "bbox": [ + 226, + 69, + 303, + 78 + ], + "type": "text", + "content": "(c) Foundation Models [48]" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 235, + 78, + 301, + 159 + ], + "blocks": [ + { + "bbox": [ + 235, + 78, + 301, + 159 + ], + "lines": [ + { + "bbox": [ + 235, + 78, + 301, + 159 + ], + "spans": [ + { + "bbox": [ + 235, + 78, + 301, + 159 + ], + "type": "image", + "image_path": "05b8064b4eee9ec5218df50a58fa07d35e8e2b29b97a29926a4cbfdc025b5ffb.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 310, + 78, + 547, + 159 + ], + "blocks": [ + { + "bbox": [ + 338, + 69, + 495, + 78 + ], + "lines": [ + { + "bbox": [ + 338, + 69, + 495, + 78 + ], + "spans": [ + { + "bbox": [ + 338, + 69, + 495, + 78 + ], + "type": "text", + "content": "(d) Solving Instance Detection in the Open World (ours)" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 310, + 78, + 547, + 159 + ], + "lines": [ + { + "bbox": [ + 310, + 78, + 547, + 159 + ], + "spans": [ + { + "bbox": [ + 310, + 78, + 547, + 159 + ], + "type": "image", + "image_path": "b8e2450a827beaff13567c6ac5add694cbb0ca36efe16ea321df09abb650aa07.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 270, + 296, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 270, + 296, + 389 + ], + "spans": [ + { + "bbox": [ + 54, + 270, + 296, + 389 + ], + "type": "text", + "content": "3D voxel representations from visual references and a matching function between 2D proposals and 3D voxels. Recent work OTS-FM [48] uses off-the-shelf pretrained foundation models (Fig. 2c) for InsDet. Specifically, it uses SAM [22] to detect object proposals in scene imagery, and DINOv2 [38] to represent visual references and proposals for instance-level feature matching. OTS-FM is a non-learned method and hence is applicable in both CID and NID settings. Importantly, it achieves the state-of-the-art performance on the recently released benchmark HR-InsDet [48]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 390, + 296, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 390, + 296, + 653 + ], + "spans": [ + { + "bbox": [ + 56, + 390, + 296, + 653 + ], + "type": "text", + "content": "The open-world perspective. As illustrated in Fig. 1, the challenges of InsDet lie in its open-world nature: the unknown testing data distribution, and domain gaps between visual references and detected proposals. To address the open-world challenges, existing methods take different strategies to exploit the open-world information (Fig. 2a-c). For example, to address the unknown testing scene imagery, some methods sample diverse background images [9, 11], demonstrating a practice of sampling data in the open world. To mitigate the domain gap between detected proposals and visual references, [26] attempts to learn features for instance-level matching by utilizing external datasets (ShapeNet [5] and ABO [8]). This data utilization can be considered as data collection in the open world. For better proposal detection and more discriminative feature representations, [48] adopts off-the-shelf pretrained foundation models (FMs) SAM [22] and DINOv2 [38], respectively. One can think of FMs as being pretrained in the open world. Different from existing methods which have not fully capitalized the open-world information, we take the open-world perspective as the first principle and solve InsDet by exploiting the Open World (Fig. 2d). We call our approach IDOW." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": "Insights and novelties. Our insights are rooted in the understanding of InsDet's open-world nature and the practice of existing methods. Existing methods address open-world issues to some extent (Fig. 2), particularly by sampling data in the open world [9, 23, 26] and adopting foundation mod" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 270, + 555, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 270, + 555, + 461 + ], + "spans": [ + { + "bbox": [ + 313, + 270, + 555, + 461 + ], + "type": "text", + "content": "els (FMs) pretrained in the open world [48]. These motivate us to develop InsDet methods in the open world (IDOW). We find that open-world detectors yield high recall on object instances (see precision-recall curves in Fig. 9 of the Appendix) but FMs are not tailored to instance-level matching between proposals and visual references – directly using an FM for InsDet is sub-optimal. Therefore, we adapt FMs to instance-level feature representation by metric learning. We propose novel and simple techniques for data augmentation to enhance FM adaptation and feature learning: distractor sampling, and multi-view synthesis. The former helps learn more discriminative features for instance comparison, and the latter enhances visual references particularly when they are few. With these techniques, our IDOW achieves significant boosts over previously reported results on recently released challenging benchmarks (Table 1 and 2)." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 326, + 462, + 537, + 473 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 462, + 537, + 473 + ], + "spans": [ + { + "bbox": [ + 326, + 462, + 537, + 473 + ], + "type": "text", + "content": "Contributions. We make three major contributions:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 314, + 475, + 555, + 606 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 315, + 475, + 554, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 475, + 554, + 522 + ], + "spans": [ + { + "bbox": [ + 315, + 475, + 554, + 522 + ], + "type": "text", + "content": "1. We elaborate on the open-world challenge of InsDet and motivate the necessity of developing methods in the open world. We solve InsDet by exploiting diverse open data available in the open world and FMs pretrained therein." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 314, + 523, + 555, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 523, + 555, + 559 + ], + "spans": [ + { + "bbox": [ + 314, + 523, + 555, + 559 + ], + "type": "text", + "content": "2. We introduce simple and effective techniques to adapt FMs to InsDet: a metric learning loss, and data augmentation with distractor sampling and novel-view synthesis." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 314, + 559, + 555, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 559, + 555, + 606 + ], + "spans": [ + { + "bbox": [ + 314, + 559, + 555, + 606 + ], + "type": "text", + "content": "3. We extensively compare our approach with existing ones on two recently released datasets, demonstrating that our approach outperforms them by " + }, + { + "bbox": [ + 314, + 559, + 555, + 606 + ], + "type": "inline_equation", + "content": ">10" + }, + { + "bbox": [ + 314, + 559, + 555, + 606 + ], + "type": "text", + "content": " AP in both conventional and novel instance detection settings." + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 620, + 400, + 632 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 620, + 400, + 632 + ], + "spans": [ + { + "bbox": [ + 314, + 620, + 400, + 632 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "content": "Instance detection. Early methods use local features such as SIFT [31] and SURF [2] to match visual references [43] and image regions to localize instances [16]. Recent methods train deep neural networks for instance-agnostic proposal detection and proposal-instance matching, achieving significant improvements [26, 34, 48]. In the con" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9902" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 57, + 72, + 296, + 383 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 72, + 296, + 383 + ], + "spans": [ + { + "bbox": [ + 57, + 72, + 296, + 383 + ], + "type": "text", + "content": "ventional instance detection (CID) setting, visual references of object instances are given during training. To approach CID, prevalent method synthesize training data in a cut-paste-learn (CPL) strategy [9, 11]: cutting instances from visual references and pasting them on random background images, then learning a detector on such synthetic data. Other works strive to obtain more training samples by rendering realistic instance examples [17, 19], using data augmentation [9] and synthesizing better training images [9, 12, 25]. Moreover, recent literature introduces the novel instance detection (NID) setting [26], which requires detecting novel instances specified only in testing and does not allow for further model finetuning during testing. VoxDet [26] utilizes a pretrained open-world detector [20] for proposal detection. Nowadays, the open-world perspective advances proposal detection in a way of learning more generalizable detectors and carrying out open-vocabulary recognition [29]. While a foundational open-vocabulary detector can attach a textual description (e.g., a short phrase) to each proposal, simply using it is insufficient for instance-level matching. [48] realizes this issue although foundational detectors can yield nearly perfect recall. As a result, the recent method OTS-FM uses another FM DINOv2 [38] for proposal-instance matching [48]. Differently, we solve InsDet from the open-world perspective, exploiting open FMs and open data to learn more discriminative instance-level feature representations." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 391, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 391, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 391, + 296, + 713 + ], + "type": "text", + "content": "Foundation models (FMs) are pre-trained on web-scale data, which can be thought of as being sampled in the open world. Different FMs are trained on different types of data and can shine in different downstream tasks related to visual perception, natural language processing, or both. As this work concerns about visual perception, we briefly review Vision-Language Models (VLMs) and Visual Foundation Models (VFMs). VLMs are pretrained on web-scale image-text paired data [18, 28, 29, 44], demonstrating impressive results in high-level tasks such as visual grounding and image captioning. VFMs are pretrained primarily on visual data [4, 7, 14, 22, 38, 55, 58, 63] and can yield impressive visual perception results such as proposal or open-set object detection [29, 59]. In InsDet, the mainstream framework requires proposal detection and visual feature matching. In this framework, open-world detectors appear to achieve high recall on the instances of interest, owing to that the openworld pretrained detectors generalize quite well in the open world [29, 46, 59]. In contrast, feature matching still remains a challenge in improving InsDet. Therefore, we focus on improving features for InsDet by adapting VFMs using data sampled in the open world. [48] shows that using an off-the-shelf FM DINOv2 [38] for feature matching greatly enhances InsDet detection; yet, FM adaptation has been still under-explored in InsDet, although it is a well studied in other areas such as few-shot perception [30, 32, 51] and zero-shot perception [13, 40, 41]. Our work introduces simple" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 316, + 73, + 533, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 533, + 84 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 533, + 84 + ], + "type": "text", + "content": "and novel techniques to adapt FMs to improve InsDet." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 315, + 87, + 555, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 87, + 555, + 419 + ], + "spans": [ + { + "bbox": [ + 315, + 87, + 555, + 419 + ], + "type": "text", + "content": "The open-world data is publicly available for pretraining FMs. Coping with such data is also often challenging. For example, data between training and testing has distribution shifts or domain gaps [30, 49], as clearly instantiated in InsDet by the unknown testing scene imagery, in addition to the domain gaps between (clean) visual references and (occluded) proposals (Fig. 1). Speaking of distribution shifts, some existing methods sample open-world data, demonstrating a need to develop robust models in the open world [15, 23]. Existing InsDet methods also sample open-world data such as diverse background images and instance examples (Fig. 2) [9, 11]. Even so, while open-world data is abundant, most of it is not directly useful for training models for specific downstream tasks. In contrast, data augmentation aims to synthesize more tailored data to enhance models' robustness and generalization, including geometry-aware augmentation (e.g., random cropping and flipping) [7, 24, 61], photometric-aware augmentation (e.g., color jittering) [21, 56, 57], and robustness-aware augmentation (e.g., adversarial perturbation) [60]. It is worth noting that Neural Radiance Field (NeRF) [35], a recent rendering technique, can be used to synthesize novel-view images. In this work, we particularly adopt NeRF to synthesize novel-view visual references of object instances and use them to adapt FMs towards more tailored features to InsDet. To the best of our knowledge, using NeRF is under-explored in the literature of InsDet, but we find that doing so remarkably improves InsDet performance." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 316, + 435, + 548, + 447 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 435, + 548, + 447 + ], + "spans": [ + { + "bbox": [ + 316, + 435, + 548, + 447 + ], + "type": "text", + "content": "3. Instance Detection: Protocols and Methods" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 315, + 456, + 555, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 456, + 555, + 502 + ], + "spans": [ + { + "bbox": [ + 315, + 456, + 555, + 502 + ], + "type": "text", + "content": "We first introduce the protocols of InsDet, including the problem definition, evaluation, and two settings. We then present our techniques to enhance foundation model adaptation to InsDet." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 316, + 515, + 379, + 526 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 515, + 379, + 526 + ], + "spans": [ + { + "bbox": [ + 316, + 515, + 379, + 526 + ], + "type": "text", + "content": "3.1. Protocols" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 315, + 533, + 555, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 533, + 555, + 616 + ], + "spans": [ + { + "bbox": [ + 315, + 533, + 555, + 616 + ], + "type": "text", + "content": "Problem definition. Instance Detection (InsDet) aims to detect specific object instances from a 2D scene image (Fig. 1), where the objects of interest are defined by a \"support set\" of some visual references [9, 26, 48]. Previous works capture the visual references from multiple camera views, with QR code being pasted to help estimate camera poses (see visual references in Fig. 1)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 315, + 618, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 618, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 315, + 618, + 556, + 713 + ], + "type": "text", + "content": "Challenges. InsDet is a challenging problem due to its open-world nature: one has zero knowledge about the testing data distribution which is instantiated by unknown testing scenes, never-before-seen matters, arbitrary occlusions, background clutters, etc. In particular, object instances of interest can be novel and are only defined during testing, as emphasized by the novel instance detection setting below. Nevertheless, InsDet models must be capable of detecting" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "9903" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 55, + 69, + 196, + 207 + ], + "blocks": [ + { + "bbox": [ + 55, + 69, + 196, + 207 + ], + "lines": [ + { + "bbox": [ + 55, + 69, + 196, + 207 + ], + "spans": [ + { + "bbox": [ + 55, + 69, + 196, + 207 + ], + "type": "image", + "image_path": "9acc9f19f1e8fb0dd2d830ffbfda8222b4ad67ec9bb3fecef2c70d7870232a39.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 209, + 555, + 276 + ], + "lines": [ + { + "bbox": [ + 55, + 209, + 555, + 276 + ], + "spans": [ + { + "bbox": [ + 55, + 209, + 555, + 276 + ], + "type": "text", + "content": "Figure 3. Visual comparison of InsDet results by different methods in the CID setting on HR-InsDet [48]. The testing scene image contains sparse placement of small instances with challenging illumination. We mark the ground-truth and predictions using green and blue boxes, respectively. We attach instance IDs to them to highlight whether the instance recognition is correct compared to the visual references (i.e., the leftmost references). Compared with Cut-Paste-Learn and OTS-FM, our IDOW detects more instances (see orange arrows) with better accuracy (see red arrows). Comparison between OTS-FM and IDOW suggests that adapting the FM DINOv2 yields better features for IDOW to perform much better in front of domain shifts, i.e., challenging illumination conditions in this testing scene image." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 202, + 68, + 314, + 207 + ], + "blocks": [ + { + "bbox": [ + 202, + 68, + 314, + 207 + ], + "lines": [ + { + "bbox": [ + 202, + 68, + 314, + 207 + ], + "spans": [ + { + "bbox": [ + 202, + 68, + 314, + 207 + ], + "type": "image", + "image_path": "18b7a33130cc1bee7ed8b0187c08c3b38d9123636647441d95516d430266c1bf.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 318, + 68, + 431, + 207 + ], + "blocks": [ + { + "bbox": [ + 318, + 68, + 431, + 207 + ], + "lines": [ + { + "bbox": [ + 318, + 68, + 431, + 207 + ], + "spans": [ + { + "bbox": [ + 318, + 68, + 431, + 207 + ], + "type": "image", + "image_path": "d735d4105590ba8f0f98a663f6eceb51a9b7c9771272d9a40e666cae9c994d2b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 437, + 68, + 549, + 207 + ], + "blocks": [ + { + "bbox": [ + 437, + 68, + 549, + 207 + ], + "lines": [ + { + "bbox": [ + 437, + 68, + 549, + 207 + ], + "spans": [ + { + "bbox": [ + 437, + 68, + 549, + 207 + ], + "type": "image", + "image_path": "2aa413d1b18d69f4f0f452a819509693153e0ed990c2dfc2599396608187c941.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 288, + 295, + 310 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 288, + 295, + 310 + ], + "spans": [ + { + "bbox": [ + 55, + 288, + 295, + 310 + ], + "type": "text", + "content": "any object instances of interest and robustly comparing them with visual references." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 312, + 296, + 347 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 312, + 296, + 347 + ], + "spans": [ + { + "bbox": [ + 55, + 312, + 296, + 347 + ], + "type": "text", + "content": "Two settings. There are two settings of InsDet, which differ in whether object instances are pre-defined during training or defined on the fly in testing time." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 348, + 297, + 574 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 55, + 348, + 296, + 466 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 348, + 296, + 466 + ], + "spans": [ + { + "bbox": [ + 55, + 348, + 296, + 466 + ], + "type": "text", + "content": "- Conventional Instance Detection (CID). In this setting, all object instances of interest are pre-defined during training [9]. That said, one can use their visual references as training data to train detectors. Yet, only visual references are available and testing scene data is still unknown. Hence, InsDet methods in CID often choose to sample diverse background images in the open world [9, 12]. This setting simulates application scenarios in AR/VR. For instance, robots should memorize the items concerned by the customer so as to provide better customized services." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 467, + 297, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 467, + 297, + 574 + ], + "spans": [ + { + "bbox": [ + 55, + 467, + 297, + 574 + ], + "type": "text", + "content": "- Novel Instance Detection (NID). This setting requires InsDet models to detect novel object instances defined only in testing time and does not allow the models to be finetuned further during testing [26]. In this setting, the de facto practice to train InsDet models turns to external data, which can be either synthetic data or those sampled in the open world. This setting simulates some real-world scenarios, for instance, robots must search for a never-before-seen luggage of a customer at an airport." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 581, + 280, + 594 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 581, + 280, + 594 + ], + "spans": [ + { + "bbox": [ + 55, + 581, + 280, + 594 + ], + "type": "text", + "content": "3.2. Learning Instance Detector in Open World" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 598, + 295, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 598, + 295, + 634 + ], + "spans": [ + { + "bbox": [ + 55, + 598, + 295, + 634 + ], + "type": "text", + "content": "We present our solution of instance detection from an open-world perspective (IDOW), focusing on foundation model adaptation and data augmentation." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 639, + 214, + 651 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 639, + 214, + 651 + ], + "spans": [ + { + "bbox": [ + 55, + 639, + 214, + 651 + ], + "type": "text", + "content": "3.2.1. Foundation Model Adaptation" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "content": "As an open-world detector yields quite high recall [26, 48], we focus on adapting a foundation model to enhance feature representations for InsDet. Although a foundational feature model (e.g., DINOv2 [38]) offers feature representations applicable to various downstream tasks, it is not tai-" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 287, + 555, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 287, + 555, + 395 + ], + "spans": [ + { + "bbox": [ + 313, + 287, + 555, + 395 + ], + "type": "text", + "content": "lored to InsDet. To improve feature representation, one can finetune this FM over relevant data, such as those sampled in the open world [26]. Intuitively, the finetuned feature is expected to match proposals and visual references better if they are from the same object instances; otherwise, distinguish them. Therefore, we propose to use a metric learning loss for FM adaptation. Moreover, we sample distractors from imagery in the open world. The distractors serve as negatives in metric learning." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 396, + 556, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 396, + 556, + 564 + ], + "spans": [ + { + "bbox": [ + 313, + 396, + 556, + 564 + ], + "type": "text", + "content": "We denote an FM as " + }, + { + "bbox": [ + 313, + 396, + 556, + 564 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 313, + 396, + 556, + 564 + ], + "type": "text", + "content": ", parameterized by " + }, + { + "bbox": [ + 313, + 396, + 556, + 564 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 396, + 556, + 564 + ], + "type": "text", + "content": ". This FM serves as the feature extractor " + }, + { + "bbox": [ + 313, + 396, + 556, + 564 + ], + "type": "inline_equation", + "content": "f_{\\theta}(\\mathbf{I}) : \\mathbf{I} \\to \\mathbb{R}^q" + }, + { + "bbox": [ + 313, + 396, + 556, + 564 + ], + "type": "text", + "content": " to produce " + }, + { + "bbox": [ + 313, + 396, + 556, + 564 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 313, + 396, + 556, + 564 + ], + "type": "text", + "content": "-dimensional features for any given visual reference image or proposal " + }, + { + "bbox": [ + 313, + 396, + 556, + 564 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 313, + 396, + 556, + 564 + ], + "type": "text", + "content": ". We finetune " + }, + { + "bbox": [ + 313, + 396, + 556, + 564 + ], + "type": "inline_equation", + "content": "f_{\\theta}(\\cdot)" + }, + { + "bbox": [ + 313, + 396, + 556, + 564 + ], + "type": "text", + "content": " to better serve InsDet such that features of reference images from the same instance should present high similarity, otherwise low similarity. We achieve this by finetuning " + }, + { + "bbox": [ + 313, + 396, + 556, + 564 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 313, + 396, + 556, + 564 + ], + "type": "text", + "content": " using a simple metric learning loss " + }, + { + "bbox": [ + 313, + 396, + 556, + 564 + ], + "type": "inline_equation", + "content": "\\ell" + }, + { + "bbox": [ + 313, + 396, + 556, + 564 + ], + "type": "text", + "content": ". Specifically, we construct triplets of training data " + }, + { + "bbox": [ + 313, + 396, + 556, + 564 + ], + "type": "inline_equation", + "content": "(\\mathbf{I}_a, \\mathbf{I}_p, \\mathbf{I}_n)" + }, + { + "bbox": [ + 313, + 396, + 556, + 564 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 396, + 556, + 564 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_a, \\mathbf{I}_p" + }, + { + "bbox": [ + 313, + 396, + 556, + 564 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 396, + 556, + 564 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_n" + }, + { + "bbox": [ + 313, + 396, + 556, + 564 + ], + "type": "text", + "content": " represent an anchor visual reference, positive and negative samples, respectively. The training data contains examples sampled or synthesized in the open world [26], as well as visual references available in the CID setting. Negative examples are cross-instance visual examples and distractors (detailed later). Below is the loss:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 323, + 576, + 555, + 596 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 576, + 555, + 596 + ], + "spans": [ + { + "bbox": [ + 323, + 576, + 555, + 596 + ], + "type": "interline_equation", + "content": "\\ell = \\left[ d \\left(f _ {\\theta} \\left(\\mathbf {I} _ {a}\\right), f _ {\\theta} \\left(\\mathbf {I} _ {p}\\right)\\right) - d \\left(f _ {\\theta} \\left(\\mathbf {I} _ {a}\\right), f _ {\\theta} \\left(\\mathbf {I} _ {n}\\right)\\right) + \\alpha \\right] _ {+}, \\tag {1}", + "image_path": "5463912de7769213310facc7026ef24f5e78f060832e0b1c9b0868e77e45be43.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "content": " is a hyper-parameter determining the margin between an instance and other negative data. " + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "inline_equation", + "content": "d(\\cdot)" + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "content": " measures the distance between two examples, e.g., inverse cosine similarity used in our work. After training, we use the learned features to represent both proposals and visual references. We follow the instance-proposal matching pipeline [48] to produce the final InsDet results: computing pairwise similarities between each proposal and each visual reference, running stable matching algorithm [10, 33], and returning" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9904" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 81, + 114, + 187 + ], + "blocks": [ + { + "bbox": [ + 77, + 70, + 107, + 79 + ], + "lines": [ + { + "bbox": [ + 77, + 70, + 107, + 79 + ], + "spans": [ + { + "bbox": [ + 77, + 70, + 107, + 79 + ], + "type": "text", + "content": "Support" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 61, + 81, + 114, + 187 + ], + "lines": [ + { + "bbox": [ + 61, + 81, + 114, + 187 + ], + "spans": [ + { + "bbox": [ + 61, + 81, + 114, + 187 + ], + "type": "image", + "image_path": "c52f8bbd473a77799bf9b4d1de247f976a8f32707f718a714dfe22158720ae25.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 115, + 81, + 220, + 187 + ], + "blocks": [ + { + "bbox": [ + 152, + 70, + 203, + 79 + ], + "lines": [ + { + "bbox": [ + 152, + 70, + 203, + 79 + ], + "spans": [ + { + "bbox": [ + 152, + 70, + 203, + 79 + ], + "type": "text", + "content": "Ground Truth" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 115, + 81, + 220, + 187 + ], + "lines": [ + { + "bbox": [ + 115, + 81, + 220, + 187 + ], + "spans": [ + { + "bbox": [ + 115, + 81, + 220, + 187 + ], + "type": "image", + "image_path": "2f118a273c6ba9e6288fa5a7f90048931795a1486c728c68eaaa8688d9b2040a.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 225, + 81, + 329, + 187 + ], + "blocks": [ + { + "bbox": [ + 260, + 70, + 306, + 80 + ], + "lines": [ + { + "bbox": [ + 260, + 70, + 306, + 80 + ], + "spans": [ + { + "bbox": [ + 260, + 70, + 306, + 80 + ], + "type": "text", + "content": "VoxDet [26]" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 225, + 81, + 329, + 187 + ], + "lines": [ + { + "bbox": [ + 225, + 81, + 329, + 187 + ], + "spans": [ + { + "bbox": [ + 225, + 81, + 329, + 187 + ], + "type": "image", + "image_path": "2564597b30890f9ddc99bac46796810c4cfe428baeae9b3880b177c5eacf3e3c.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 190, + 555, + 255 + ], + "lines": [ + { + "bbox": [ + 55, + 190, + 555, + 255 + ], + "spans": [ + { + "bbox": [ + 55, + 190, + 555, + 255 + ], + "type": "text", + "content": "Figure 4. Visual comparison of InsDet results by different methods in the NID setting on RoboTools [26]. On the cluttered testing scene image, we mark the ground-truth and predictions using green and blue boxes, respectively. We attach instance IDs to them to highlight whether the instance recognition is correct compared to the visual references (i.e., the leftmost references). Compared with VoxDet and OTS-FM, our IDOW detects more instances (see orange arrows) with better accuracy (see red arrows). Comparing OTS-FM and our IDOW suggests that, although having not seen the visual references during adaption in this NID setting, features by our IDOW are a better representation for InsDet than the off-the-shelf foundation model (i.e., DINOv2) used in OTS-FM." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 334, + 81, + 439, + 187 + ], + "blocks": [ + { + "bbox": [ + 362, + 70, + 413, + 80 + ], + "lines": [ + { + "bbox": [ + 362, + 70, + 413, + 80 + ], + "spans": [ + { + "bbox": [ + 362, + 70, + 413, + 80 + ], + "type": "text", + "content": "OTS-FM [48]" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 334, + 81, + 439, + 187 + ], + "lines": [ + { + "bbox": [ + 334, + 81, + 439, + 187 + ], + "spans": [ + { + "bbox": [ + 334, + 81, + 439, + 187 + ], + "type": "image", + "image_path": "bffecc67dbe4d09cf7afd173469cbe8045cbc77f2aabe89b5312145f9a2c9b2c.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 444, + 81, + 548, + 187 + ], + "blocks": [ + { + "bbox": [ + 468, + 70, + 520, + 79 + ], + "lines": [ + { + "bbox": [ + 468, + 70, + 520, + 79 + ], + "spans": [ + { + "bbox": [ + 468, + 70, + 520, + 79 + ], + "type": "text", + "content": "IDOW (ours)" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 444, + 81, + 548, + 187 + ], + "lines": [ + { + "bbox": [ + 444, + 81, + 548, + 187 + ], + "spans": [ + { + "bbox": [ + 444, + 81, + 548, + 187 + ], + "type": "image", + "image_path": "b0ca2c049ef5a9441f7abc088463327a050887f5eed635b016e7553aaa67ecc8.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 269, + 295, + 292 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 269, + 295, + 292 + ], + "spans": [ + { + "bbox": [ + 55, + 269, + 295, + 292 + ], + "type": "text", + "content": "the matched pairs if they have similarity scores greater than a predefined threshold (0.4 as used in [48])." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 293, + 296, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 293, + 296, + 473 + ], + "spans": [ + { + "bbox": [ + 55, + 293, + 296, + 473 + ], + "type": "text", + "content": "Hard example sampling. To construct training batches of triplets, it is important to effectively sample negative examples, easy negative examples do not provide sufficient gradients during training [50]. In this work, we adopt a batch-level hard negative sampling strategy. In each training batch, we first sample a reference image as an anchor " + }, + { + "bbox": [ + 55, + 293, + 296, + 473 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_a" + }, + { + "bbox": [ + 55, + 293, + 296, + 473 + ], + "type": "text", + "content": ". We then sample a random example from the same instance as the positive one " + }, + { + "bbox": [ + 55, + 293, + 296, + 473 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_p" + }, + { + "bbox": [ + 55, + 293, + 296, + 473 + ], + "type": "text", + "content": ". We finally sample a hard negative example from the union of the reference images of all the other instances and the distractors " + }, + { + "bbox": [ + 55, + 293, + 296, + 473 + ], + "type": "inline_equation", + "content": "S_{dt}" + }, + { + "bbox": [ + 55, + 293, + 296, + 473 + ], + "type": "text", + "content": " (explained in the next subsection). Mathematically, we find hard negative patch " + }, + { + "bbox": [ + 55, + 293, + 296, + 473 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{hn}" + }, + { + "bbox": [ + 55, + 293, + 296, + 473 + ], + "type": "text", + "content": " with the objective: " + }, + { + "bbox": [ + 55, + 293, + 296, + 473 + ], + "type": "inline_equation", + "content": "\\arg \\min_{\\mathbf{I}_n} d(f_\\theta(\\mathbf{I}_a), f_\\theta(\\mathbf{I}_n))" + }, + { + "bbox": [ + 55, + 293, + 296, + 473 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 293, + 296, + 473 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_n \\in \\{C(\\mathbf{I}_i) \\neq C(\\mathbf{I}_a)\\}" + }, + { + "bbox": [ + 55, + 293, + 296, + 473 + ], + "type": "text", + "content": ", for " + }, + { + "bbox": [ + 55, + 293, + 296, + 473 + ], + "type": "inline_equation", + "content": "i = 1 \\dots N\\} \\cup S_{dt}" + }, + { + "bbox": [ + 55, + 293, + 296, + 473 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 293, + 296, + 473 + ], + "type": "inline_equation", + "content": "C(\\mathbf{I}_i)" + }, + { + "bbox": [ + 55, + 293, + 296, + 473 + ], + "type": "text", + "content": " represents the instance ID of the data " + }, + { + "bbox": [ + 55, + 293, + 296, + 473 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_i" + }, + { + "bbox": [ + 55, + 293, + 296, + 473 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 293, + 296, + 473 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 293, + 296, + 473 + ], + "type": "text", + "content": " indicates the total number of visual reference images." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 482, + 228, + 493 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 482, + 228, + 493 + ], + "spans": [ + { + "bbox": [ + 55, + 482, + 228, + 493 + ], + "type": "text", + "content": "3.2.2. The Proposed Data Augmentation" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 497, + 296, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 497, + 296, + 521 + ], + "spans": [ + { + "bbox": [ + 55, + 497, + 296, + 521 + ], + "type": "text", + "content": "To better adapt FM for instance-level feature matching, we introduce two data augmentation techniques below." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 522, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 522, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 522, + 296, + 713 + ], + "type": "text", + "content": "Distractor sampling aims to sample patches of random background images as universal negative data to all object instances. It is a data augmentation technique, expected to define the open space and help features better characterize meaningful object instances. Distractor sampling has been adopted in the literature. For example, open-set recognition samples distractors as a separate class, known as \"the-other-class\" or \"background-class\", to define the open space. In InsDet, the cut-paste-learn strategy [9] samples random background photos to define the background content. In this work, we run SAM [22] on random background photos and use the segments as distractors. One might worry about the possibility of sampled distractors being exactly some instances of interest and including them as negatives may destabilize model adaptation. Yet, given a large number of distractors, this rarely becomes an issue as demonstrated by the success" + } + ] + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 316, + 266, + 354, + 306 + ], + "blocks": [ + { + "bbox": [ + 316, + 266, + 354, + 306 + ], + "lines": [ + { + "bbox": [ + 316, + 266, + 354, + 306 + ], + "spans": [ + { + "bbox": [ + 316, + 266, + 354, + 306 + ], + "type": "image", + "image_path": "210ab2b74669701f38b34cded9c4adb572b83785ffad4cddd52c715be54d89dd.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 356, + 266, + 393, + 306 + ], + "blocks": [ + { + "bbox": [ + 356, + 266, + 393, + 306 + ], + "lines": [ + { + "bbox": [ + 356, + 266, + 393, + 306 + ], + "spans": [ + { + "bbox": [ + 356, + 266, + 393, + 306 + ], + "type": "image", + "image_path": "b672221f98d026a64c1900ff47d14121e0810d0dc3354bed182a875306a50e9a.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 393, + 266, + 432, + 306 + ], + "blocks": [ + { + "bbox": [ + 393, + 266, + 432, + 306 + ], + "lines": [ + { + "bbox": [ + 393, + 266, + 432, + 306 + ], + "spans": [ + { + "bbox": [ + 393, + 266, + 432, + 306 + ], + "type": "image", + "image_path": "cb5287d79ea04599ea5a1ac389160c0e4ee9e34a83bc5b731dc7b0b53db188ad.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 432, + 266, + 471, + 306 + ], + "blocks": [ + { + "bbox": [ + 432, + 266, + 471, + 306 + ], + "lines": [ + { + "bbox": [ + 432, + 266, + 471, + 306 + ], + "spans": [ + { + "bbox": [ + 432, + 266, + 471, + 306 + ], + "type": "image", + "image_path": "3de5432c7f80dc5c58ba739f8ab8165307d9e350d64aad10822dea7445f807c0.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 471, + 266, + 511, + 306 + ], + "blocks": [ + { + "bbox": [ + 471, + 266, + 511, + 306 + ], + "lines": [ + { + "bbox": [ + 471, + 266, + 511, + 306 + ], + "spans": [ + { + "bbox": [ + 471, + 266, + 511, + 306 + ], + "type": "image", + "image_path": "7118f2b44624c755ec6c59899245246ea3aec875a0529b81a7d1fe3512e01956.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 511, + 266, + 551, + 306 + ], + "blocks": [ + { + "bbox": [ + 511, + 266, + 551, + 306 + ], + "lines": [ + { + "bbox": [ + 511, + 266, + 551, + 306 + ], + "spans": [ + { + "bbox": [ + 511, + 266, + 551, + 306 + ], + "type": "image", + "image_path": "6c119278d1c3f1d98d9fbef146fc06ae216513b82ae2bbb1e8000e9001163d30.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 317, + 308, + 351, + 346 + ], + "blocks": [ + { + "bbox": [ + 317, + 308, + 351, + 346 + ], + "lines": [ + { + "bbox": [ + 317, + 308, + 351, + 346 + ], + "spans": [ + { + "bbox": [ + 317, + 308, + 351, + 346 + ], + "type": "image", + "image_path": "a81afc1f65cb57a2a4b85f6ea9718430dd8d5801eabbd7a001ecea5df7140387.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 358, + 308, + 388, + 346 + ], + "blocks": [ + { + "bbox": [ + 358, + 308, + 388, + 346 + ], + "lines": [ + { + "bbox": [ + 358, + 308, + 388, + 346 + ], + "spans": [ + { + "bbox": [ + 358, + 308, + 388, + 346 + ], + "type": "image", + "image_path": "35c1a7d54a5ef6af0836c5d0d601919bc1c7cc9c92900870df6ad555181b8a14.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 389, + 308, + 421, + 346 + ], + "blocks": [ + { + "bbox": [ + 389, + 308, + 421, + 346 + ], + "lines": [ + { + "bbox": [ + 389, + 308, + 421, + 346 + ], + "spans": [ + { + "bbox": [ + 389, + 308, + 421, + 346 + ], + "type": "image", + "image_path": "e444383cb9a8750cbc2a30e1af7c81d9f9eb2c6efddca6df33ae10e431d2d83b.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 422, + 308, + 466, + 346 + ], + "blocks": [ + { + "bbox": [ + 422, + 308, + 466, + 346 + ], + "lines": [ + { + "bbox": [ + 422, + 308, + 466, + 346 + ], + "spans": [ + { + "bbox": [ + 422, + 308, + 466, + 346 + ], + "type": "image", + "image_path": "57134a4754730b4882d93742904b59ee5869e3d6fed9d70cfd46237e3db1f0e1.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 466, + 308, + 507, + 346 + ], + "blocks": [ + { + "bbox": [ + 466, + 308, + 507, + 346 + ], + "lines": [ + { + "bbox": [ + 466, + 308, + 507, + 346 + ], + "spans": [ + { + "bbox": [ + 466, + 308, + 507, + 346 + ], + "type": "image", + "image_path": "dad18aa7f2248a0407ec92fec14ea9d085e12f1a4c08875ed2ba1d9be5804932.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 507, + 308, + 545, + 346 + ], + "blocks": [ + { + "bbox": [ + 507, + 308, + 545, + 346 + ], + "lines": [ + { + "bbox": [ + 507, + 308, + 545, + 346 + ], + "spans": [ + { + "bbox": [ + 507, + 308, + 545, + 346 + ], + "type": "image", + "image_path": "49f5260a048b6661ba0b6e12321bd67aec00f7250beb1e51277a9f0e2ab9ffd1.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 317, + 349, + 351, + 385 + ], + "blocks": [ + { + "bbox": [ + 317, + 349, + 351, + 385 + ], + "lines": [ + { + "bbox": [ + 317, + 349, + 351, + 385 + ], + "spans": [ + { + "bbox": [ + 317, + 349, + 351, + 385 + ], + "type": "image", + "image_path": "829d828f813648f8ae7a0f65200a4dddf3533c0ca52c977dc877b030c2e7fbd1.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 357, + 349, + 388, + 385 + ], + "blocks": [ + { + "bbox": [ + 357, + 349, + 388, + 385 + ], + "lines": [ + { + "bbox": [ + 357, + 349, + 388, + 385 + ], + "spans": [ + { + "bbox": [ + 357, + 349, + 388, + 385 + ], + "type": "image", + "image_path": "a5fcc0e5a9a3ad6d0b3065ebda2fbbc7cb33e688df95ebddfd9130d596833f7d.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 389, + 349, + 421, + 385 + ], + "blocks": [ + { + "bbox": [ + 389, + 349, + 421, + 385 + ], + "lines": [ + { + "bbox": [ + 389, + 349, + 421, + 385 + ], + "spans": [ + { + "bbox": [ + 389, + 349, + 421, + 385 + ], + "type": "image", + "image_path": "34d842752f58c2d6a72de7c6c01d99629fc1d36963f3c1a1742e5ef018262c22.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 422, + 349, + 466, + 385 + ], + "blocks": [ + { + "bbox": [ + 422, + 349, + 466, + 385 + ], + "lines": [ + { + "bbox": [ + 422, + 349, + 466, + 385 + ], + "spans": [ + { + "bbox": [ + 422, + 349, + 466, + 385 + ], + "type": "image", + "image_path": "de1e1669113de871305e5f24ef148c0b789ac126d238d24f700c9c489450d3b2.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 466, + 349, + 507, + 385 + ], + "blocks": [ + { + "bbox": [ + 466, + 349, + 507, + 385 + ], + "lines": [ + { + "bbox": [ + 466, + 349, + 507, + 385 + ], + "spans": [ + { + "bbox": [ + 466, + 349, + 507, + 385 + ], + "type": "image", + "image_path": "f40580b085c30e268001f17eb472427d5c06bd829dd5f4e655d0f8cdb7f1845c.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 507, + 349, + 539, + 385 + ], + "blocks": [ + { + "bbox": [ + 507, + 349, + 539, + 385 + ], + "lines": [ + { + "bbox": [ + 507, + 349, + 539, + 385 + ], + "spans": [ + { + "bbox": [ + 507, + 349, + 539, + 385 + ], + "type": "image", + "image_path": "b6ba97081f556ea8867104d01590efb35130fa1e4a13a14f1807e797d6744b6c.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + } + ], + "index": 33 + }, + { + "type": "image", + "bbox": [ + 317, + 388, + 351, + 425 + ], + "blocks": [ + { + "bbox": [ + 317, + 388, + 351, + 425 + ], + "lines": [ + { + "bbox": [ + 317, + 388, + 351, + 425 + ], + "spans": [ + { + "bbox": [ + 317, + 388, + 351, + 425 + ], + "type": "image", + "image_path": "621451e5081d7528f487226a6ab565d338947f7b43248db89a92ea51f39f0959.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 427, + 555, + 503 + ], + "lines": [ + { + "bbox": [ + 313, + 427, + 555, + 503 + ], + "spans": [ + { + "bbox": [ + 313, + 427, + 555, + 503 + ], + "type": "text", + "content": "Figure 5. We use NeRF to synthesize novel-view object images to augment the limited given visual references. In this work, we train per-instance Zip-NeRF [1]. We visualize synthesized images at different angles (in row 2-4) together with the raw visual references (in the first row). The visual references are from HR-InsDet dataset [48], where the QR code is used for camera pose estimation. Overall, we find synthesized images show high visual quality." + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_caption" + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 357, + 387, + 393, + 425 + ], + "blocks": [ + { + "bbox": [ + 357, + 387, + 393, + 425 + ], + "lines": [ + { + "bbox": [ + 357, + 387, + 393, + 425 + ], + "spans": [ + { + "bbox": [ + 357, + 387, + 393, + 425 + ], + "type": "image", + "image_path": "bb4f3f82f28371d4d03f0f8d1b9e6c8dbcea35d0537c1895015eaa76de69bed7.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + } + ], + "index": 35 + }, + { + "type": "image", + "bbox": [ + 393, + 387, + 421, + 425 + ], + "blocks": [ + { + "bbox": [ + 393, + 387, + 421, + 425 + ], + "lines": [ + { + "bbox": [ + 393, + 387, + 421, + 425 + ], + "spans": [ + { + "bbox": [ + 393, + 387, + 421, + 425 + ], + "type": "image", + "image_path": "2227a2aea790a14e09cae9f7b8b6d0bddc810cc2fdbbf176ac0d4683fba10118.jpg" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_body" + } + ], + "index": 36 + }, + { + "type": "image", + "bbox": [ + 422, + 387, + 466, + 425 + ], + "blocks": [ + { + "bbox": [ + 422, + 387, + 466, + 425 + ], + "lines": [ + { + "bbox": [ + 422, + 387, + 466, + 425 + ], + "spans": [ + { + "bbox": [ + 422, + 387, + 466, + 425 + ], + "type": "image", + "image_path": "1552f025dba22887fa9fddef0fc9920007df0f52bd5d2806d6a65ff4b5b199cc.jpg" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_body" + } + ], + "index": 37 + }, + { + "type": "image", + "bbox": [ + 466, + 387, + 511, + 425 + ], + "blocks": [ + { + "bbox": [ + 466, + 387, + 511, + 425 + ], + "lines": [ + { + "bbox": [ + 466, + 387, + 511, + 425 + ], + "spans": [ + { + "bbox": [ + 466, + 387, + 511, + 425 + ], + "type": "image", + "image_path": "50ae24f9bd6d764b65a57adfc26143be676ae3158a864bf0e338e4ecddc04e6a.jpg" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_body" + } + ], + "index": 38 + }, + { + "type": "image", + "bbox": [ + 511, + 387, + 541, + 425 + ], + "blocks": [ + { + "bbox": [ + 511, + 387, + 541, + 425 + ], + "lines": [ + { + "bbox": [ + 511, + 387, + 541, + 425 + ], + "spans": [ + { + "bbox": [ + 511, + 387, + 541, + 425 + ], + "type": "image", + "image_path": "847c9b59247efcca4dd8995311cba223229c42966137d36b6d609e760de94cb5.jpg" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_body" + } + ], + "index": 39 + }, + { + "bbox": [ + 313, + 518, + 555, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 518, + 555, + 553 + ], + "spans": [ + { + "bbox": [ + 313, + 518, + 555, + 553 + ], + "type": "text", + "content": "of various self-supervised learning methods [38] which may sample examples serving as both positive and negative data at the same time." + } + ] + } + ], + "index": 41 + }, + { + "bbox": [ + 313, + 558, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 556, + 713 + ], + "type": "text", + "content": "Novel-view synthesis. Each object instance might be represented by only a few visual examples, especially in the CID setting which provides limited real visual references for instances of interest. To augment available visual references for better adapting FM, we learn a NeRF [35] for each instance and use NeRF to synthesize novel-view images. In this process, we estimate relative camera poses (required by NeRF) using COLMAP [47], as commonly done in NeRF methods [1, 35] and InsDet approaches [26, 48]. Fig. 5 displays some random NeRF-generated novel-view images with good quality in visuals. It is worth noting that a NeRF needs to be trained only once on the visual references of an object instance. Once it is trained, we use it to generate" + } + ] + } + ], + "index": 42 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "9905" + } + ] + } + ], + "index": 43 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 138, + 555, + 281 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 137 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 137 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 137 + ], + "type": "text", + "content": "Table 1. Benchmarking results in the CID setting on the HR-InsDet dataset. We compare our IDOW with state-of-the-arts and make three salient conclusions. First, IDOW significantly outperforms previous methods, e.g., " + }, + { + "bbox": [ + 55, + 70, + 555, + 137 + ], + "type": "inline_equation", + "content": "\\mathrm{IDOW}_{\\mathrm{GroundingDINO}}" + }, + { + "bbox": [ + 55, + 70, + 555, + 137 + ], + "type": "text", + "content": " (57.01 AP) > OTS-FM_{GroundingDINO} (51.68 AP) > CPL_{DINO} (27.99 AP). This confirms the importance of addressing InsDet from the open-world perspective. Second, adapting FMs by our IDOW further boosts the performance by 5-7 AP, e.g., " + }, + { + "bbox": [ + 55, + 70, + 555, + 137 + ], + "type": "inline_equation", + "content": "\\mathrm{IDOW}_{\\mathrm{SAM}}" + }, + { + "bbox": [ + 55, + 70, + 555, + 137 + ], + "type": "text", + "content": " (48.75 AP) > OTS-FMSAM (41.61 AP). Third, IDOW and OTS-FM are applicable to different pretrained FMs and adopting stronger FMs achieves better performance, e.g., using GroundingDINO yields " + }, + { + "bbox": [ + 55, + 70, + 555, + 137 + ], + "type": "inline_equation", + "content": ">8" + }, + { + "bbox": [ + 55, + 70, + 555, + 137 + ], + "type": "text", + "content": " AP than SAM in IDOW." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 138, + 555, + 281 + ], + "lines": [ + { + "bbox": [ + 56, + 138, + 555, + 281 + ], + "spans": [ + { + "bbox": [ + 56, + 138, + 555, + 281 + ], + "type": "table", + "html": "
MethodVenue & YearAP\\( \\mathbf{A}\\mathbf{P}_{50} \\)\\( \\mathbf{A}\\mathbf{P}_{75} \\)
avghardeasysmallmediumlarge
\\( CPL_{FasterRCNN} \\)[9, 45]NeurIPS 201519.5410.2623.755.0322.2037.9729.2123.26
\\( CPL_{RetinaNet} \\)[9, 27]ICCV 201722.2214.9226.495.4825.8042.7131.1924.98
\\( CPL_{CenterNet} \\)[9, 64]CVPR 201921.1211.8525.705.9024.1540.3832.7223.60
\\( CPL_{FCOS} \\)[9, 54]ICCV 201922.4013.2228.686.1726.4638.1332.8025.47
\\( CPL_{DINO} \\)[9, 59]ICLR 202327.9917.8932.6511.5131.6048.3539.6232.19
OTS-FMSAM[22, 48]NeurIPS 202341.6128.0347.5714.5845.8369.1449.1045.95
OTS-FMGroundingDINO[29, 48]NeurIPS 202351.6837.2358.7228.7958.5569.2262.5056.78
\\( IDOW_{SAM} \\)48.7532.0956.5020.7555.2673.4357.5954.06
\\( IDOW_{GroundingDINO} \\)57.0140.7464.3635.2562.9873.6469.3362.84
", + "image_path": "2b8da109456542c5f7b6efb99356fe161983071d1397ef3dfb459886d3a32a2d.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 290, + 295, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 290, + 295, + 373 + ], + "spans": [ + { + "bbox": [ + 55, + 290, + 295, + 373 + ], + "type": "text", + "content": "more visual references not only for training but for testing. In particular, such NeRF-synthesized novel-view images can be stored together with the original visual references in testing for proposal-instance matching. Doing so significantly improves InsDet performance (Table 3 and Fig. 7). To the best of our knowledge, our work is the first that exploits NeRF to synthesize data for InsDet." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 388, + 136, + 402 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 388, + 136, + 402 + ], + "spans": [ + { + "bbox": [ + 55, + 388, + 136, + 402 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 408, + 296, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 408, + 296, + 456 + ], + "spans": [ + { + "bbox": [ + 55, + 408, + 296, + 456 + ], + "type": "text", + "content": "In this section, we first introduce the experimental setup, then evaluate our IDOW by comparing existing methods in both CID and NID settings, and finally conduct ablation studies to analyze our approach." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 466, + 171, + 479 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 466, + 171, + 479 + ], + "spans": [ + { + "bbox": [ + 55, + 466, + 171, + 479 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 484, + 295, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 484, + 295, + 592 + ], + "spans": [ + { + "bbox": [ + 55, + 484, + 295, + 592 + ], + "type": "text", + "content": "Datasets. We use two recently published datasets in our experiments. HR-InsDet [48] is developed for the CID setting, containing 100 daily object instances, 160 high-resolution testing images from 14 indoor scenarios (see Fig. 3 for an example). RoboTools [26] is developed for the NID setting, containing 20 robotic tool instances and 1,581 testing images from 24 indoor scenarios (see Fig. 4 for an example). For a fair comparison, we use the OWID dataset released by [26] to adapt FM in the NID setting." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 593, + 295, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 593, + 295, + 688 + ], + "spans": [ + { + "bbox": [ + 55, + 593, + 295, + 688 + ], + "type": "text", + "content": "Metrics. We use metrics adopted in both datasets mentioned above, including average precision (AP) at IoU thresholds from 0.5 to 0.95 with the step size 0.05 as the primary metric and also report " + }, + { + "bbox": [ + 55, + 593, + 295, + 688 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_{50}" + }, + { + "bbox": [ + 55, + 593, + 295, + 688 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 593, + 295, + 688 + ], + "type": "inline_equation", + "content": "\\mathrm{AP}_{75}" + }, + { + "bbox": [ + 55, + 593, + 295, + 688 + ], + "type": "text", + "content": " averaged over all instances with IoU threshold as 0.5 and 0.75, respectively. The HR-InsDet dataset contains tags for hard/easy testing scenes and small/medium/large object instances. We follow [48] to break down results on these tags." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": "Compared methods. We compare previous methods including CPL, OTS-FM, and VoxDet. These methods have" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 290, + 555, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 290, + 555, + 410 + ], + "spans": [ + { + "bbox": [ + 313, + 290, + 555, + 410 + ], + "type": "text", + "content": "options to use different components (so do ours). For example, CPL can train a detector using one of the popular architectures Faster RCNN [45], RetinaNet [27], CenterNet [64], FCOS [54], and DINO [59]. OTS-FM can use SAM or GroundingDINO for proposal detection. We implement these methods by using different backbones to improve their performance (Table 1). We use subscripts to mark the backbones used in each method. Moreover, we compare more methods in the NID setting, including OS2D [39],DTOID [34], and " + }, + { + "bbox": [ + 313, + 290, + 555, + 410 + ], + "type": "inline_equation", + "content": "\\mathrm{OLN}_{\\mathrm{Corr}}" + }, + { + "bbox": [ + 313, + 290, + 555, + 410 + ], + "type": "text", + "content": " [20]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 411, + 556, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 411, + 556, + 542 + ], + "spans": [ + { + "bbox": [ + 313, + 411, + 556, + 542 + ], + "type": "text", + "content": "Implementation details. We carry out all experiments on a single NVIDIA 3090 GPU and implement methods with PyTorch. Unless otherwise specified, our IDOW uses GroundingDINO [29] as the proposal detector and DINOv2 [38] as the feature foundation model to finetune. We use the pretrained [CLS] token to extract features and class embedding to compute similarity scores for matching. When finetuning, we set hyper-parameter " + }, + { + "bbox": [ + 313, + 411, + 556, + 542 + ], + "type": "inline_equation", + "content": "\\alpha = 0.5" + }, + { + "bbox": [ + 313, + 411, + 556, + 542 + ], + "type": "text", + "content": " in the metric learning loss. We use Adam optimizer with a learning rate 1e-3 and weight decay 0.5. We set the batch size as 100 and finetune the model for 10 epochs." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 552, + 443, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 552, + 443, + 563 + ], + "spans": [ + { + "bbox": [ + 313, + 552, + 443, + 563 + ], + "type": "text", + "content": "4.2. Benchmarking Results" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 569, + 555, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 569, + 555, + 616 + ], + "spans": [ + { + "bbox": [ + 313, + 569, + 555, + 616 + ], + "type": "text", + "content": "Quantitative results. We evaluate our method against prior approaches on the InsDet and RoboTools dataset. Table 1 and 2 list detailed results in the CID and NID settings, respectively. We summarize three salient observations:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 617, + 556, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 556, + 700 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 556, + 700 + ], + "type": "text", + "content": "First, approaching InsDet from the open-world perspective significantly outperforms previous methods. From both CID and NID settings, we find that, despite not being explicitly trained on the instances of interests, OTS-FM already achieves " + }, + { + "bbox": [ + 313, + 617, + 556, + 700 + ], + "type": "inline_equation", + "content": ">10" + }, + { + "bbox": [ + 313, + 617, + 556, + 700 + ], + "type": "text", + "content": " AP than traditional detector-based methods which particularly train on such instances. Adapting FMs by our IDOW boosts performance by 3-5 AP." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 325, + 701, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 701, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 325, + 701, + 555, + 714 + ], + "type": "text", + "content": "Second, IDOW significantly outperforms state-of-the-art" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "9906" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 72, + 218, + 228 + ], + "blocks": [ + { + "bbox": [ + 61, + 72, + 218, + 228 + ], + "lines": [ + { + "bbox": [ + 61, + 72, + 218, + 228 + ], + "spans": [ + { + "bbox": [ + 61, + 72, + 218, + 228 + ], + "type": "image", + "image_path": "6cf9eb521679c594896c30332d81f55203f31da2c04693e8d76effc39daf572d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 229, + 555, + 264 + ], + "lines": [ + { + "bbox": [ + 55, + 229, + 555, + 264 + ], + "spans": [ + { + "bbox": [ + 55, + 229, + 555, + 264 + ], + "type": "text", + "content": "Figure 6. t-SNE visualization of features learned by different methods. We compare our finetuned DINOv2 against a finetuned ImageNet-pretrained ResNet101 model and the baseline instance detector CPL with a FasterRCNN architecture. We color feature points w.r.t instance IDs and overlay star-marked visual references (zoom-in to see better). Visually, the finetuned DINOv2 extracts more discriminative features." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 225, + 72, + 383, + 228 + ], + "blocks": [ + { + "bbox": [ + 225, + 72, + 383, + 228 + ], + "lines": [ + { + "bbox": [ + 225, + 72, + 383, + 228 + ], + "spans": [ + { + "bbox": [ + 225, + 72, + 383, + 228 + ], + "type": "image", + "image_path": "029a4bf98c259efb71d3b9a3329752f557e396db67ca6526b7b80f9bcaa44a4d.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 391, + 72, + 548, + 228 + ], + "blocks": [ + { + "bbox": [ + 391, + 72, + 548, + 228 + ], + "lines": [ + { + "bbox": [ + 391, + 72, + 548, + 228 + ], + "spans": [ + { + "bbox": [ + 391, + 72, + 548, + 228 + ], + "type": "image", + "image_path": "50bd78c6da05e1478ae947e40c67079c5050e9fbd893202cc4caea089b925398.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 56, + 390, + 296, + 507 + ], + "blocks": [ + { + "bbox": [ + 55, + 276, + 297, + 387 + ], + "lines": [ + { + "bbox": [ + 55, + 276, + 297, + 387 + ], + "spans": [ + { + "bbox": [ + 55, + 276, + 297, + 387 + ], + "type": "text", + "content": "Table 2. Benchmarking results in the NID setting on RoboTools. We compare our IDOW with state-of-the-arts and make three salient conclusions. First, IDOW significantly outperforms previous methods, e.g., " + }, + { + "bbox": [ + 55, + 276, + 297, + 387 + ], + "type": "inline_equation", + "content": "\\mathrm{IDOW}_{\\mathrm{GroundingDINO}}" + }, + { + "bbox": [ + 55, + 276, + 297, + 387 + ], + "type": "text", + "content": " (59.4 AP) > OTS-FM_{GroundingDINO} (56.7 AP) > VoxDet (18.7 AP). This demonstrates the effectiveness of approaching InsDet from the open-world perspective. Second, adapting FMs by our IDOW further boosts the performance by 3-5 AP, e.g., " + }, + { + "bbox": [ + 55, + 276, + 297, + 387 + ], + "type": "inline_equation", + "content": "\\mathrm{IDOW}_{\\mathrm{SAM}}" + }, + { + "bbox": [ + 55, + 276, + 297, + 387 + ], + "type": "text", + "content": " (51.9 AP) > OTS-FMSAM (46.5 AP). Third, adopting stronger FMs achieves better performance, e.g., " + }, + { + "bbox": [ + 55, + 276, + 297, + 387 + ], + "type": "inline_equation", + "content": "\\mathrm{IDOW}_{\\mathrm{GroundingDINO}}" + }, + { + "bbox": [ + 55, + 276, + 297, + 387 + ], + "type": "text", + "content": " (59.4 AP) > " + }, + { + "bbox": [ + 55, + 276, + 297, + 387 + ], + "type": "inline_equation", + "content": "\\mathrm{IDOW}_{\\mathrm{SAM}}" + }, + { + "bbox": [ + 55, + 276, + 297, + 387 + ], + "type": "text", + "content": " (51.9 AP) in IDOW." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 390, + 296, + 507 + ], + "lines": [ + { + "bbox": [ + 56, + 390, + 296, + 507 + ], + "spans": [ + { + "bbox": [ + 56, + 390, + 296, + 507 + ], + "type": "table", + "html": "
MethodVenue & YearAPAP50AP75
OS2D [39]ECCV 20202.96.52.0
DTOID [34]WACV 20213.69.02.0
OLNCorr. [20]RA-L 202214.418.115.7
VoxDet [26]NeurIPS 202318.723.620.5
OTS-FMSAM [48]NeurIPS 202346.555.950.4
OTS-FMGroundingDINO [48]NeurIPS 202356.764.859.0
IDOWSAM51.963.856.5
IDOWGroundingDINO59.467.861.8
", + "image_path": "09bdb5564a148b0df38bb63d3c5f6b234b7abb47b61fae056a12557379718581.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 519, + 296, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 519, + 296, + 555 + ], + "spans": [ + { + "bbox": [ + 55, + 519, + 296, + 555 + ], + "type": "text", + "content": "InsDet methods, achieving over 5 AP higher on average than methods like OTS-FM in both CID and NID settings, demonstrating the effectiveness of our techniques." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 556, + 296, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 556, + 296, + 604 + ], + "spans": [ + { + "bbox": [ + 55, + 556, + 296, + 604 + ], + "type": "text", + "content": "Third, IDOW is versatile and applicable to different open-world pretrained FMs. By comparing " + }, + { + "bbox": [ + 55, + 556, + 296, + 604 + ], + "type": "inline_equation", + "content": "\\mathrm{IDOW}_{\\mathrm{SAM}}" + }, + { + "bbox": [ + 55, + 556, + 296, + 604 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 556, + 296, + 604 + ], + "type": "inline_equation", + "content": "\\mathrm{IDOW}_{\\mathrm{GroundingDINO}}" + }, + { + "bbox": [ + 55, + 556, + 296, + 604 + ], + "type": "text", + "content": ", we find that adopting stronger FMs (GroundingDINO vs. SAM) achieves better performance." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 605, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 605, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 605, + 297, + 714 + ], + "type": "text", + "content": "Qualitative Results. Fig. 3 and 4 compares InsDet results by different methods in the CID and NID settings, respectively. We can see that our IDOW has higher precision and detects more wanted instances, which are small and under low lighting conditions. Fig. 6 uses tSNE to visualize features of visual references and detected proposals (in the HR-InsDet dataset) computed by our finetuned DINOv2 (ft-DINOv2) in the left panel, and other non-FM methods including a finetuned ImageNet-pretrained ResNet101 back" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 277, + 555, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 277, + 555, + 314 + ], + "spans": [ + { + "bbox": [ + 313, + 277, + 555, + 314 + ], + "type": "text", + "content": "bone (ft-ImageNet-Res101) and CPL_FasterRCNN in the middle and right panels, respectively. Clearly, ft-DINOv2 matches proposals and references much better than other models." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 319, + 511, + 333 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 319, + 511, + 333 + ], + "spans": [ + { + "bbox": [ + 313, + 319, + 511, + 333 + ], + "type": "text", + "content": "4.3. Ablation Study and Further Analysis" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 337, + 556, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 337, + 556, + 456 + ], + "spans": [ + { + "bbox": [ + 313, + 337, + 556, + 456 + ], + "type": "text", + "content": "In Table 3, we ablate the proposed techniques of NeRF-synthesis and distractor sampling. Results show that adapting the FM DINOv2 greatly boosts InsDet performance (from 51.68 AP by the baseline to 53.94 AP). Importantly, including NeRF-synthesized images in both training and testing significantly improves the performance. Yet, using them in testing brings more performance gains than training (refer to the next study in Fig. 7). Eventually, applying our techniques boosts AP to 57.01. The supplemental material contains more quantitative and qualitative results." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 457, + 557, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 457, + 557, + 613 + ], + "spans": [ + { + "bbox": [ + 313, + 457, + 557, + 613 + ], + "type": "text", + "content": "Following the above, we study the affects of using different number of NeRF-synthesized images in training and test. Fig. 7 depicts the results. (1) Adding NeRF-generated images helps foundation model adaptation for InsDet. Specifically, adding " + }, + { + "bbox": [ + 313, + 457, + 557, + 613 + ], + "type": "inline_equation", + "content": "<24" + }, + { + "bbox": [ + 313, + 457, + 557, + 613 + ], + "type": "text", + "content": " generated images boosts the performance, especially when 0 synthesized images are used in testing. However, adding " + }, + { + "bbox": [ + 313, + 457, + 557, + 613 + ], + "type": "inline_equation", + "content": ">24" + }, + { + "bbox": [ + 313, + 457, + 557, + 613 + ], + "type": "text", + "content": " generated images for training induces diminishing performance gains, likely because that the adapted FM overfits to artifacts in NeRF generated images. (2) Adopting more synthetic images in testing consistently improves performance. Yet, the performance by adding more (e.g., " + }, + { + "bbox": [ + 313, + 457, + 557, + 613 + ], + "type": "inline_equation", + "content": ">36" + }, + { + "bbox": [ + 313, + 457, + 557, + 613 + ], + "type": "text", + "content": ") in testing saturates. The peak performance occurs when 36 NeRF-synthesized images are used in testing." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 621, + 389, + 634 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 621, + 389, + 634 + ], + "spans": [ + { + "bbox": [ + 313, + 621, + 389, + 634 + ], + "type": "text", + "content": "5. Discussions" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "text", + "content": "Broader Impacts. Approaching InsDet in the open-world has a significant impact on various downstream applications, such as robotics and AR/VR. Detecting object instances in RGB images is considered the first step of contemporary perception algorithms. A stronger InsDet model provides richer information for the following steps, such" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9907" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 149, + 555, + 308 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 148 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 148 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 148 + ], + "type": "text", + "content": "Table 3. Ablation study of each strategy involved in our IDOW. We carry out the study in the CID setting on the HR-InsDet dataset. We use OTS-FM" + }, + { + "bbox": [ + 55, + 70, + 555, + 148 + ], + "type": "inline_equation", + "content": "_{\\text{GroundingDINO}}" + }, + { + "bbox": [ + 55, + 70, + 555, + 148 + ], + "type": "text", + "content": " as a baseline, over which we incrementally add each strategy. \"Train\" means foundation model adaptation through finetuning on the available data. DA denotes Data Augmentation with NeRF-generated novel-views; DS denotes Distractor Sampling. Results clearly demonstrate that all the four strategies help achieve better InsDet performance. Finetuning FMs on the given visual references enhances detection performance, cf. Train (53.94 AP) > baseline (51.68 AP). Moreover, NeRF-based data augmentation improves the final detection performance, particularly when used in testing, cf. Train+DA@Test (56.44 AP) > Train+DA@Train (54.48 AP) > baseline (51.68 AP). Lastly, applying distractor sampling (DS) improves the final performance further, cf. Train+DS (54.10 AP) > Train (53.94 AP)." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 149, + 555, + 308 + ], + "lines": [ + { + "bbox": [ + 56, + 149, + 555, + 308 + ], + "spans": [ + { + "bbox": [ + 56, + 149, + 555, + 308 + ], + "type": "table", + "html": "
StrategiesAP\\( \\mathbf{A}\\mathbf{P}_{50} \\)\\( \\mathbf{A}\\mathbf{P}_{75} \\)
DA@TestTrainDA@TrainDSavghardeasysmallmediumlarge
baseline: OTS-FMGroundingDINO51.6837.2358.7228.7958.5569.2262.5056.78
✓53.9437.5461.5230.1860.9871.6065.1859.36
✓✓54.4838.3161.7131.1861.1072.5566.0659.98
✓✓54.1037.7361.6530.2761.1971.2765.3759.57
✓✓✓54.9238.2862.4132.0061.6174.2166.5860.50
✓55.2439.6462.8134.0061.2072.3866.9960.88
✓✓56.4439.7164.2434.7263.1374.4268.5062.31
✓✓✓56.9239.8564.8534.6662.7475.4868.8962.80
✓✓✓56.5139.7064.3634.3461.6575.8968.7162.38
✓✓✓✓57.0140.7464.3635.2562.9873.6469.3362.84
", + "image_path": "25caaab3137560315849c22265fc94c72df8d46ad39e8900454754ea7ac7e8f0.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 62, + 322, + 285, + 464 + ], + "blocks": [ + { + "bbox": [ + 62, + 322, + 285, + 464 + ], + "lines": [ + { + "bbox": [ + 62, + 322, + 285, + 464 + ], + "spans": [ + { + "bbox": [ + 62, + 322, + 285, + 464 + ], + "type": "image", + "image_path": "7460d61347fe8fa2529db2ae5da64ac653743cf68825e265c42b724202a0ce72.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 470, + 297, + 581 + ], + "lines": [ + { + "bbox": [ + 55, + 470, + 297, + 581 + ], + "spans": [ + { + "bbox": [ + 55, + 470, + 297, + 581 + ], + "type": "text", + "content": "Figure 7. Analysis of the number of added NeRF-synthesized images in training and testing w.r.t AP improvement over the \"Train\" method (Table 3), which by default finetunes DINOv2 using the available visual references (24 per instance). We carry out this analysis in the CID setting on the HR-InsDet dataset. On the one hand, adding NeRF-synthesized images helps finetune FM for InsDet, but adding more (e.g., " + }, + { + "bbox": [ + 55, + 470, + 297, + 581 + ], + "type": "inline_equation", + "content": ">24" + }, + { + "bbox": [ + 55, + 470, + 297, + 581 + ], + "type": "text", + "content": ") hurts the performance, likely due to overfitting to the artifacts of NeRF-synthesized images. On the other hand, adding more NeRF images in testing consistently improves performance." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 594, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 297, + 714 + ], + "type": "text", + "content": "as grasping or trajectory forecasting in robotics. Additionally, our work shows the promising performance of adapting FMs to InsDet, which will foster future research in studying different properties of FMs, such as adapting FMs to other tasks. Nonetheless, our research utilizes FMs that are potentially pretrained on private data. At this point, it is unclear how such FMs might have bias or ethical issues. As a result, we have no safeguard measure to ensure that the finetuned/adapted FMs are fair, even though we find no visible issues in the image-only dataset used for finetuning." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 323, + 555, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 323, + 555, + 371 + ], + "spans": [ + { + "bbox": [ + 313, + 323, + 555, + 371 + ], + "type": "text", + "content": "Further, adopting SAM may result in missed detections of critical instances relevant to specific downstream applications, such as medical devices or other rare objects. Such potential issues are not addressed in this work." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 373, + 556, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 373, + 556, + 553 + ], + "spans": [ + { + "bbox": [ + 313, + 373, + 556, + 553 + ], + "type": "text", + "content": "Limitations and future work. Experiments demonstrate that our proposed techniques effectively leverage FMs to address InsDet across various settings. Nonetheless, we note several limitations for future improvements. First, the current pipeline relies on the proposals generated from GroundingDINO. Despite its high recall, it is not yet ready for real-time applications. Future work should consider faster proposal detection methods. Second, the NeRF augmentation step is also slow since it requires training a NeRF for every instance, which takes 1 hour and 500MB space using one NVIDIA 3090 GPU. While developing efficient NeRF methods is beyond our scope, we point out that efficient NeRF methods can be readily used in our pipeline, such as [37] which trains a NeRF in seconds and [6] which can train a single NeRF for multiple objects." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 571, + 388, + 583 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 571, + 388, + 583 + ], + "spans": [ + { + "bbox": [ + 314, + 571, + 388, + 583 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "content": "Instance detection is highly applicable in various scenarios such as in robotics and AR/VR. We elaborate on its open-world nature, necessitating the development of InsDet models in the open world. We solve InsDet from the open world perspective, embracing foundation models (FMs) and data sampled in the open world. In particular, we propose simple and effective techniques such as metric learning, distractor sampling, and novel-view synthesis. Extensive experiments in different settings validate that our techniques in adapting FMs significantly boost InsDet performance." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "9908" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 92, + 297, + 128 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 92, + 297, + 128 + ], + "spans": [ + { + "bbox": [ + 55, + 92, + 297, + 128 + ], + "type": "text", + "content": "This work was supported by FDCT (0067/2024/ITP2), the University of Macau (SRG2023-00044-FST), the Institute of Collaborative Innovation and NSFC (No.62206256)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 142, + 115, + 154 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 142, + 115, + 154 + ], + "spans": [ + { + "bbox": [ + 56, + 142, + 115, + 154 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 162, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 61, + 162, + 296, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 162, + 296, + 194 + ], + "spans": [ + { + "bbox": [ + 61, + 162, + 296, + 194 + ], + "type": "text", + "content": "[1] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Zip-nerf: Anti-aliased grid-based neural radiance fields. In ICCV, 2023. 2, 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 197, + 296, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 197, + 296, + 217 + ], + "spans": [ + { + "bbox": [ + 62, + 197, + 296, + 217 + ], + "type": "text", + "content": "[2] Herbert Bay, Tinne Tuytelaars, and Luc Van Gool. Surf: Speeded up robust features. In ECCV, 2006. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 220, + 296, + 253 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 220, + 296, + 253 + ], + "spans": [ + { + "bbox": [ + 62, + 220, + 296, + 253 + ], + "type": "text", + "content": "[3] Richard Bormann, Xinjie Wang, Markus Volk, Kilian Kleeberger, and Jochen Lindermayr. Real-time instance detection with fast incremental learning. In ICRA. IEEE, 2021. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 255, + 296, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 255, + 296, + 297 + ], + "spans": [ + { + "bbox": [ + 62, + 255, + 296, + 297 + ], + "type": "text", + "content": "[4] Mathilde Caron, Hugo Touvron, Ishan Misra, Herve Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, 2021. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 301, + 296, + 344 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 301, + 296, + 344 + ], + "spans": [ + { + "bbox": [ + 62, + 301, + 296, + 344 + ], + "type": "text", + "content": "[5] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv:1512.03012, 2015. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 347, + 296, + 389 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 347, + 296, + 389 + ], + "spans": [ + { + "bbox": [ + 62, + 347, + 296, + 389 + ], + "type": "text", + "content": "[6] Hansheng Chen, Jiatao Gu, Anpei Chen, Wei Tian, Zhuowen Tu, Lingjie Liu, and Hao Su. Single-stage diffusion nerf: A unified approach to 3d generation and reconstruction. In ICCV, 2023. 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 393, + 296, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 393, + 296, + 425 + ], + "spans": [ + { + "bbox": [ + 62, + 393, + 296, + 425 + ], + "type": "text", + "content": "[7] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In ICML, 2020. 3, 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 428, + 296, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 428, + 296, + 482 + ], + "spans": [ + { + "bbox": [ + 62, + 428, + 296, + 482 + ], + "type": "text", + "content": "[8] Jasmine Collins, Shubham Goel, Kenan Deng, Achleshwar Luthra, Leon Xu, Erhan Gundogdu, Xi Zhang, Tomas F Yago Vicente, Thomas Dideriksen, Himanshu Arora, et al. Abo: Dataset and benchmarks for real-world 3d object understanding. In CVPR, 2022. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 484, + 296, + 517 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 484, + 296, + 517 + ], + "spans": [ + { + "bbox": [ + 62, + 484, + 296, + 517 + ], + "type": "text", + "content": "[9] Debidatta Dwibedi, Ishan Misra, and Martial Hebert. Cut, paste and learn: Surprisingly easy synthesis for instance detection. In ICCV, 2017. 1, 2, 3, 4, 5, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 520, + 296, + 551 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 520, + 296, + 551 + ], + "spans": [ + { + "bbox": [ + 57, + 520, + 296, + 551 + ], + "type": "text", + "content": "[10] David Gale and Lloyd S Shapley. College admissions and the stability of marriage. The American Mathematical Monthly, 69(1):9-15, 1962. 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 554, + 296, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 554, + 296, + 597 + ], + "spans": [ + { + "bbox": [ + 57, + 554, + 296, + 597 + ], + "type": "text", + "content": "[11] Georgios Georgakis, Arsalan Mousavian, Alexander C. Berg, and Jana Kosecka. Synthesizing training data for object detection in indoor scenes. In Proceedings of Robotics: Science and Systems, 2017. 1, 2, 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 600, + 296, + 642 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 600, + 296, + 642 + ], + "spans": [ + { + "bbox": [ + 57, + 600, + 296, + 642 + ], + "type": "text", + "content": "[12] Georgios Georgakis, Arsalan Mousavian, Alexander C Berg, and Jana Košecká. Synthesizing training data for object detection in indoor scenes. Robotics: Science and Systems (RSS), 2017. 3, 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 646, + 296, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 646, + 296, + 677 + ], + "spans": [ + { + "bbox": [ + 57, + 646, + 296, + 677 + ], + "type": "text", + "content": "[13] Sachin Goyal, Ananya Kumar, Sankalp Garg, Zico Kolter, and Aditi Raghunathan. Finetune like you pretrain: Improved finetuning of zero-shot vision models. In CVPR, 2023. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 57, + 681, + 296, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 681, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 681, + 296, + 712 + ], + "type": "text", + "content": "[14] Sunan He, Taian Guo, Tao Dai, Ruizhi Qiao, Chen Wu, Xiujun Shu, and Bo Ren. Vlmae: Vision-language masked autoencoder. arXiv:2208.09374, 2022. 3" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 35, + "blocks": [ + { + "bbox": [ + 316, + 73, + 555, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 555, + 105 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 555, + 105 + ], + "type": "text", + "content": "[15] Dan Hendrycks, Mantas Mazeika, and Thomas Dietterich. Deep anomaly detection with outlier exposure. In ICLR, 2019. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 107, + 555, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 107, + 555, + 161 + ], + "spans": [ + { + "bbox": [ + 317, + 107, + 555, + 161 + ], + "type": "text", + "content": "[16] Stefan Hinterstoisser, Cedric Cagniart, Slobodan Ilic, Peter Sturm, Nassir Navab, Pascal Fua, and Vincent Lepetit. Gradient response maps for real-time detection of textureless objects. IEEE transactions on pattern analysis and machine intelligence, 34(5):876-888, 2011. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 162, + 555, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 162, + 555, + 216 + ], + "spans": [ + { + "bbox": [ + 317, + 162, + 555, + 216 + ], + "type": "text", + "content": "[17] Tomás Hodań, Vibhav Vineet, Ran Gal, Emanuel Shalev, Jon Hanzelka, Treb Connell, Pedro Urbina, Sudipta N Sinha, and Brian Guenter. Photorealistic image synthesis for object instance detection. In IEEE international conference on image processing (ICIP), 2019. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 217, + 554, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 217, + 554, + 259 + ], + "spans": [ + { + "bbox": [ + 317, + 217, + 554, + 259 + ], + "type": "text", + "content": "[18] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc V. Le, Yunhsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In ICML, 2021. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 261, + 555, + 302 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 261, + 555, + 302 + ], + "spans": [ + { + "bbox": [ + 317, + 261, + 555, + 302 + ], + "type": "text", + "content": "[19] Wadim Kehl, Fabian Manhardt, Federico Tombari, Slobodan Ilic, and Nassir Navab. Ssd-6d: Making rgb-based 3d detection and 6d pose estimation great again. In ICCV, 2017. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 305, + 554, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 305, + 554, + 347 + ], + "spans": [ + { + "bbox": [ + 316, + 305, + 554, + 347 + ], + "type": "text", + "content": "[20] Dahun Kim, Tsung-Yi Lin, Anelia Angelova, In So Kweon, and Weicheng Kuo. Learning open-world object proposals without learning to classify. IEEE Robotics and Automation Letters, 7(2):5453-5460, 2022. 3, 6, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 350, + 555, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 350, + 555, + 402 + ], + "spans": [ + { + "bbox": [ + 316, + 350, + 555, + 402 + ], + "type": "text", + "content": "[21] Eun Kyeong Kim, Hansoo Lee, Jin Yong Kim, and Sungshin Kim. Data augmentation method by applying color perturbation of inverse psnr and geometric transformations for object recognition based on deep learning. Applied Sciences, 10(11): 3755, 2020. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 404, + 555, + 447 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 404, + 555, + 447 + ], + "spans": [ + { + "bbox": [ + 316, + 404, + 555, + 447 + ], + "type": "text", + "content": "[22] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In ICCV, 2023. 2, 3, 5, 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 449, + 555, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 449, + 555, + 469 + ], + "spans": [ + { + "bbox": [ + 316, + 449, + 555, + 469 + ], + "type": "text", + "content": "[23] Shu Kong and Deva Ramanan. Opengan: Open-set recognition via open data generation. In ICCV, 2021. 2, 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 472, + 555, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 472, + 555, + 502 + ], + "spans": [ + { + "bbox": [ + 317, + 472, + 555, + 502 + ], + "type": "text", + "content": "[24] Alex Krizhevsky, Ilya Sutskever, and Geoffrey E Hinton. ImageNet classification with deep convolutional neural networks. NeurIPS, 25, 2012. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 504, + 554, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 504, + 554, + 525 + ], + "spans": [ + { + "bbox": [ + 317, + 504, + 554, + 525 + ], + "type": "text", + "content": "[25] Kevin Lai, Liefeng Bo, and Dieter Fox. Unsupervised feature learning for 3d scene labeling. ICRA, 2014. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 317, + 526, + 555, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 526, + 555, + 557 + ], + "spans": [ + { + "bbox": [ + 317, + 526, + 555, + 557 + ], + "type": "text", + "content": "[26] Bowen Li, Jiashun Wang, Yaoyu Hu, Chen Wang, and Sebastian Scherer. Voxdet: Voxel learning for novel instance detection. NeurIPS, 2024. 1, 2, 3, 4, 5, 6, 7" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 317, + 559, + 555, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 559, + 555, + 591 + ], + "spans": [ + { + "bbox": [ + 317, + 559, + 555, + 591 + ], + "type": "text", + "content": "[27] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. In ICCV, 2017. 6, 3" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 317, + 593, + 555, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 593, + 555, + 613 + ], + "spans": [ + { + "bbox": [ + 317, + 593, + 555, + 613 + ], + "type": "text", + "content": "[28] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. NeurIPS, 2023. 3" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 317, + 615, + 554, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 615, + 554, + 658 + ], + "spans": [ + { + "bbox": [ + 317, + 615, + 554, + 658 + ], + "type": "text", + "content": "[29] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. ECCV, 2024. 3, 6" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 317, + 659, + 555, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 659, + 555, + 690 + ], + "spans": [ + { + "bbox": [ + 317, + 659, + 555, + 690 + ], + "type": "text", + "content": "[30] Tian Liu, Huixin Zhang, Shubham Parashar, and Shu Kong. Few-shot recognition via stage-wise retrieval-augmented finetuning. In CVPR, 2025. 3" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 316, + 692, + 555, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 555, + 713 + ], + "type": "text", + "content": "[31] David G Lowe. Distinctive image features from scale-invariant keypoints. *IJCV*, 60:91-110, 2004. 2" + } + ] + } + ], + "index": 34 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "9909" + } + ] + } + ], + "index": 36 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 56, + 73, + 296, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 296, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 296, + 106 + ], + "type": "text", + "content": "[32] Anish Madan, Neehar Peri, Shu Kong, and Deva Ramanan. Revisiting few-shot object detection with vision-language models. In NeurIPS Datasets & Benchmark Track, 2024. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 296, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 296, + 139 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 296, + 139 + ], + "type": "text", + "content": "[33] David G McVitie and Leslie B Wilson. The stable marriage problem. Communications of the ACM, 14(7):486-490, 1971. 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 141, + 294, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 141, + 294, + 174 + ], + "spans": [ + { + "bbox": [ + 56, + 141, + 294, + 174 + ], + "type": "text", + "content": "[34] Jean-Philippe Mercier, Mathieu Garon, Philippe Giguere, and Jean-Francois Lalonde. Deep template-based object instance detection. In WACV, 2021. 1, 2, 6, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 175, + 296, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 296, + 218 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 296, + 218 + ], + "type": "text", + "content": "[35] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 3, 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 220, + 296, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 220, + 296, + 273 + ], + "spans": [ + { + "bbox": [ + 56, + 220, + 296, + 273 + ], + "type": "text", + "content": "[36] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 276, + 296, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 276, + 296, + 319 + ], + "spans": [ + { + "bbox": [ + 56, + 276, + 296, + 319 + ], + "type": "text", + "content": "[37] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM transactions on graphics (TOG), 41(4):1-15, 2022. 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 321, + 296, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 321, + 296, + 374 + ], + "spans": [ + { + "bbox": [ + 56, + 321, + 296, + 374 + ], + "type": "text", + "content": "[38] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv:2304.07193, 2023. 2, 3, 4, 5, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 376, + 296, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 376, + 296, + 409 + ], + "spans": [ + { + "bbox": [ + 56, + 376, + 296, + 409 + ], + "type": "text", + "content": "[39] Anton Osokin, Denis Sumin, and Vasily Lomakin. Os2d: One-stage one-shot object detection by matching anchor features. In ECCV, 2020. 6, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 411, + 296, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 411, + 296, + 442 + ], + "spans": [ + { + "bbox": [ + 56, + 411, + 296, + 442 + ], + "type": "text", + "content": "[40] Shubham Parashar, Zhiqiu Lin, Yanan Li, and Shu Kong. Prompting scientific names for zero-shot species recognition. In EMNLP, 2023. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 445, + 296, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 445, + 296, + 487 + ], + "spans": [ + { + "bbox": [ + 56, + 445, + 296, + 487 + ], + "type": "text", + "content": "[41] Shubham Parashar, Zhiqiu Lin, Tian Liu, Xiangjue Dong, Yanan Li, Deva Ramanan, James Caverlee, and Shu Kong. The neglected tails in vision-language models. In CVPR, 2024. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 490, + 296, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 490, + 296, + 544 + ], + "spans": [ + { + "bbox": [ + 56, + 490, + 296, + 544 + ], + "type": "text", + "content": "[42] Samira Pouyanfar, Saad Sadiq, Yilin Yan, Haiman Tian, Yudong Tao, Maria Presa Reyes, Mei-Ling Shyu, Shu-Ching Chen, and Sundaraja S Iyengar. A survey on deep learning: Algorithms, techniques, and applications. ACM computing surveys (CSUR), 51(5):1-36, 2018. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 545, + 296, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 545, + 296, + 577 + ], + "spans": [ + { + "bbox": [ + 56, + 545, + 296, + 577 + ], + "type": "text", + "content": "[43] A Quadros, James Patrick Underwood, and Bertrand Douillard. An occlusion-aware feature for range images. In ICRA, 2012. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 579, + 296, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 579, + 296, + 634 + ], + "spans": [ + { + "bbox": [ + 56, + 579, + 296, + 634 + ], + "type": "text", + "content": "[44] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In ICML, 2021. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 635, + 296, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 296, + 668 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 296, + 668 + ], + "type": "text", + "content": "[45] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. In NeurIPS, 2015. 6, 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 670, + 296, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 296, + 712 + ], + "type": "text", + "content": "[46] Tianhe Ren, Shilong Liu, Ailing Zeng, Jing Lin, Kunchang Li, He Cao, Jiayu Chen, Xinyu Huang, Yukang Chen, Feng Yan, et al. Grounded sam: Assembling open-world models for diverse visual tasks. arXiv:2401.14159, 2024. 3" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 660 + ], + "type": "list", + "angle": 0, + "index": 34, + "blocks": [ + { + "bbox": [ + 316, + 73, + 555, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 555, + 95 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 555, + 95 + ], + "type": "text", + "content": "[47] Johannes Lutz Schonberger and Jan-Michael Frahm. Structure-from-Motion Revisited. In CVPR, 2016. 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 96, + 555, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 555, + 139 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 555, + 139 + ], + "type": "text", + "content": "[48] Qianqian Shen, Yunhan Zhao, Nahyun Kwon, Jeeun Kim, Yanan Li, and Shu Kong. A high-resolution dataset for instance detection with multi-view object capture. In NeurIPS Datasets & Benchmark Track, 2023. 1, 2, 3, 4, 5, 6, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 141, + 555, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 555, + 184 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 555, + 184 + ], + "type": "text", + "content": "[49] Jia Shi, Gautam Gare, Jinjin Tian, Siqi Chai, Zhiqiu Lin, Arun Vasudevan, Di Feng, Francesco Ferroni, and Shu Kong. Lcaon-the-line: Benchmarking out-of-distribution generalization with class taxonomies. In ICML, 2024. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 186, + 555, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 186, + 555, + 218 + ], + "spans": [ + { + "bbox": [ + 316, + 186, + 555, + 218 + ], + "type": "text", + "content": "[50] Abhinav Shrivastava, Abhinav Gupta, and Ross Girshick. Training region-based object detectors with online hard example mining. In CVPR, 2016. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 220, + 555, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 220, + 555, + 252 + ], + "spans": [ + { + "bbox": [ + 316, + 220, + 555, + 252 + ], + "type": "text", + "content": "[51] Julio Silva-Rodriguez, Sina Hajimiri, Ismail Ben Ayed, and Jose Dolz. A closer look at the few-shot adaptation of large vision-language models. In CVPR, 2024. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 254, + 555, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 254, + 555, + 285 + ], + "spans": [ + { + "bbox": [ + 316, + 254, + 555, + 285 + ], + "type": "text", + "content": "[52] Afrina Tabassum, Muntasir Wahed, Hoda Eldardiry, and Ismini Lourentzou. Hard negative sampling strategies for contrastive representation learning. arXiv:2206.01197, 2022. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 288, + 555, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 288, + 555, + 319 + ], + "spans": [ + { + "bbox": [ + 316, + 288, + 555, + 319 + ], + "type": "text", + "content": "[53] Peng Tang, Xinggang Wang, Xiang Bai, and Wenyu Liu. Multiple instance detection network with online instance classifier refinement. In CVPR, 2017. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 321, + 555, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 321, + 555, + 342 + ], + "spans": [ + { + "bbox": [ + 316, + 321, + 555, + 342 + ], + "type": "text", + "content": "[54] Zhi Tian, Chunhua Shen, Hao Chen, and Tong He. Fcos: Fully convolutional one-stage object detection. In ICCV, 2019. 6, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 344, + 555, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 344, + 555, + 365 + ], + "spans": [ + { + "bbox": [ + 316, + 344, + 555, + 365 + ], + "type": "text", + "content": "[55] Hugo Touvron, Matthieu Cord, and Hervé Jégou. Deit iii: Revenge of the vit. In ECCV, 2022. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 367, + 554, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 367, + 554, + 399 + ], + "spans": [ + { + "bbox": [ + 316, + 367, + 554, + 399 + ], + "type": "text", + "content": "[56] Riccardo Volpi, Pietro Morierio, Silvio Savarese, and Vittorio Murino. Adversarial feature augmentation for unsupervised domain adaptation. In CVPR, 2018. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 401, + 555, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 401, + 555, + 443 + ], + "spans": [ + { + "bbox": [ + 316, + 401, + 555, + 443 + ], + "type": "text", + "content": "[57] Riccardo Volpi, Hongseok Namkoong, Ozan Sener, John C Duchi, Vittorio Murino, and Silvio Savarese. Generalizing to unseen domains via adversarial data augmentation. NeurIPS, 2018. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 445, + 555, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 445, + 555, + 468 + ], + "spans": [ + { + "bbox": [ + 316, + 445, + 555, + 468 + ], + "type": "text", + "content": "[58] XuDong Wang, Jingfeng Yang, and Trevor Darrell. Segment anything without supervision. In NeurIPS, 2024. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 469, + 555, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 469, + 555, + 512 + ], + "spans": [ + { + "bbox": [ + 316, + 469, + 555, + 512 + ], + "type": "text", + "content": "[59] Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel Ni, and Harry Shum. Dino: Detr with improved denoising anchor boxes for end-to-end object detection. In ICLR, 2022. 3, 6" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 514, + 555, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 514, + 555, + 535 + ], + "spans": [ + { + "bbox": [ + 316, + 514, + 555, + 535 + ], + "type": "text", + "content": "[60] Xinyu Zhang, Qiang Wang, Jian Zhang, and Zhao Zhong. Adversarial autoaugment. In ICLR, 2020. 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 536, + 555, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 536, + 555, + 568 + ], + "spans": [ + { + "bbox": [ + 316, + 536, + 555, + 568 + ], + "type": "text", + "content": "[61] Yunhan Zhao, Shu Kong, and Charless Fowlkes. Camera pose matters: Improving depth prediction by mitigating pose distribution bias. In CVPR, 2021. 3" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 570, + 555, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 570, + 555, + 602 + ], + "spans": [ + { + "bbox": [ + 316, + 570, + 555, + 602 + ], + "type": "text", + "content": "[62] Yunhan Zhao, Haoyu Ma, Shu Kong, and Charless Fowlkes. Instance tracking in 3d scenes from egocentric videos. In CVPR, 2024. 1" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 316, + 604, + 554, + 636 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 604, + 554, + 636 + ], + "spans": [ + { + "bbox": [ + 316, + 604, + 554, + 636 + ], + "type": "text", + "content": "[63] Jinghao Zhou, Chen Wei, Huiyu Wang, Wei Shen, Cihang Xie, Alan Yuille, and Tao Kong. Image bert pre-training with online tokenizer. In ICLR, 2022. 3" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 316, + 638, + 554, + 660 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 638, + 554, + 660 + ], + "spans": [ + { + "bbox": [ + 316, + 638, + 554, + 660 + ], + "type": "text", + "content": "[64] Xingyi Zhou, Dequan Wang, and Philipp Krahenbuhl. Objects as points. In arXiv:1904.07850, 2019. 6, 3" + } + ] + } + ], + "index": 33 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "9910" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Sonata_ Self-Supervised Learning of Reliable Point Representations/3e06b74b-3b4f-4789-9e14-04119c218bb3_content_list.json b/2025/Sonata_ Self-Supervised Learning of Reliable Point Representations/3e06b74b-3b4f-4789-9e14-04119c218bb3_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..357e3fd4426f4b83a0553e990474a5d70bdc8cb9 --- /dev/null +++ b/2025/Sonata_ Self-Supervised Learning of Reliable Point Representations/3e06b74b-3b4f-4789-9e14-04119c218bb3_content_list.json @@ -0,0 +1,1616 @@ +[ + { + "type": "text", + "text": "Sonata: Self-Supervised Learning of Reliable Point Representations", + "text_level": 1, + "bbox": [ + 156, + 69, + 841, + 92 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xiaoyang Wu $^{1,2}$ Daniel DeTone $^{2*}$ Duncan Frost $^{2*}$ Tianwei Shen $^{2*}$ Chris Xie $^{2*}$ Nan Yang $^{2*}$ Jakob Engel $^{2}$ Richard Newcombe $^{2}$ Hengshuang Zhao $^{1}$ Julian Straub $^{2}$", + "bbox": [ + 114, + 118, + 880, + 157 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1The University of Hong Kong 2Meta Reality Labs Research", + "bbox": [ + 246, + 172, + 750, + 191 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "https://github.com/facebookresearch/sonata", + "bbox": [ + 307, + 193, + 683, + 208 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/589d4300f43854af0829f5157255ce003401f1421bbb5571ff40b45a650e5aba.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 99, + 215, + 377, + 421 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/836a6f483902be083a4e0b93f8367b5b8d30e5b14d1b990fba0f77f1190fb148.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 383, + 215, + 616, + 420 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/56c11ac45288e311d5e6dbd4e365b9f89130bdd1783474991acb5e5bfaa77ffe.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 619, + 215, + 898, + 420 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6e714a08721434cff40b5c0910e86ef7bb6c7bb036bc13df01cef90c14096176.jpg", + "image_caption": [ + "Figure 1. Main properties. Sonata leads to reliable 3D self-supervised pretraining with the following superior and emerging properties: 1. Perception. Sonata advances state-of-the-art results across 3D indoor and outdoor perception tasks; 2. Linear probing. With less than $0.2\\%$ learnable parameters, Sonata achieves strong and usable linear probing performance which is $3.3\\times$ better than previous SOTA; 3. Decoder-free. Sonata moves beyond the inflexible U-Net structure, offering multi-scale representations that unchain future 3D research from previous architectural constraints. 4. Semantic awareness. Sonata reveals semantic structure in PCA and K-means visualizations. 5. Spatial reasoning. Sonata allows spatial correspondence even under strong augmentations as visualized via feature similarity." + ], + "image_footnote": [], + "bbox": [ + 99, + 422, + 493, + 603 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/989600f4ef288c1b4c60baa764c2c9036547f67de46eb78da322b2e565fdf5a7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 498, + 422, + 898, + 603 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 700, + 326, + 715 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we question whether we have a reliable self-supervised point cloud model that can be used for diverse 3D tasks via simple linear probing, even with limited data and minimal computation. We find that existing 3D self-supervised learning approaches fall short when evaluated on representation quality through linear probing. We hypothesize that this is due to what we term the \"geometric shortcut\", which causes representations to collapse to low-level spatial features. This challenge is unique to 3D and arises from the sparse nature of point cloud data.", + "bbox": [ + 89, + 720, + 485, + 875 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We address it through two key strategies: obscuring spatial information and enhancing the reliance on input features, ultimately composing a Sonata of 140k point clouds through self-distillation. Sonata is simple and intuitive, yet its learned representations are strong and reliable: zero-shot visualizations demonstrate semantic grouping, alongside strong spatial reasoning through nearest-neighbor relationships. Sonata demonstrates exceptional parameter and data efficiency, tripling linear probing accuracy (from $21.8\\%$ to $72.5\\%$ ) on ScanNet and nearly doubling performance with only $1\\%$ of the data compared to previous approaches. Full fine-tuning further advances SOTA across both 3D indoor and outdoor perception tasks.", + "bbox": [ + 511, + 700, + 908, + 898 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation.", + "bbox": [ + 236, + 0, + 807, + 18 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "Except for this watermark, it is identical to the accepted version;", + "bbox": [ + 323, + 16, + 722, + 30 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 292, + 31, + 753, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution in alphabetic order.", + "bbox": [ + 107, + 887, + 320, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "22193", + "bbox": [ + 478, + 944, + 517, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 89, + 222, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Self-supervised learning (SSL) with images [2, 13, 34, 94, 104] has seen a continuous increase in model simplicity, capacity, and capability over the past decade [25]. Tuning a single linear layer can achieve performance close to full fine-tuning [31, 35, 104], fostering growing trust in its reliability. This trust has been further strengthened by witnessing the semantic meaning of learned image representations through direct visualization [60, 110]. Consequently, these reliable self-supervised models have become the foundation for emerging approaches [14, 84, 98, 99] across various fields involving images.", + "bbox": [ + 89, + 111, + 483, + 277 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In contrast to the image domain, self-supervised learning with point clouds [18, 61, 93, 102, 106] is still in its early stages. Despite the broad reliance on 3D applications in autonomous driving [9, 74], robotic learning [32], mixed reality [5, 43] and egocentric perception [27, 73], the latest self-supervised point cloud models are seldom included in their pipelines. This gap prompts us to consider a simple yet critical question: do we have a reliable point self-supervised learning approach that provides strong representations, usable with simple linear probing across these applications? Not yet. Previous SOTAs [38, 88, 93] fall short on this higher-level criterion, achieving only a maximum of $21.8\\%$ mIoU on ScanNet [23] with linear probing, especially given that the performance from scratch is $77.6\\%$ .", + "bbox": [ + 89, + 279, + 483, + 489 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We identify the geometric shortcut as the primary issue hindering these prior point cloud SSL approaches from learning reliable representations. This shortcut refers to the tendency of the model to collapse to easily accessible, low-level geometric cues, such as normal direction or point height, as demonstrated in prior works and visualized in Fig. 2. This spatial information is inevitably introduced into point cloud operators along with point coordinates rather than through input features, making it difficult to obscure and nearly impossible to mask effectively.", + "bbox": [ + 89, + 491, + 483, + 642 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, the model collapse caused by geometric shortcuts can be mitigated through two key strategies: obscuring spatial information and emphasizing input features. Specifically, we address this issue by applying SSL losses at coarser spatial scales, disturbing the spatial information of masked points without features, and progressively increasing task difficulty to reduce reliance on accessible geometric cues. Coupled with a point self-distillation framework and scaling techniques inspired by recent advances in image SSL [11, 12, 34, 60, 110], we ultimately compose a Sonata using 140k point cloud scenes [1, 3, 5, 23, 68, 101, 109].", + "bbox": [ + 89, + 643, + 482, + 809 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Sonata demonstrates strong zero-shot capabilities, with PCA-colored visualizations of point clouds, k-means clustering of features, and nearest-neighbor matching between point clouds (see Fig. 1). Sonata also proves highly data-efficient, raising semantic segmentation performance from $25.8\\%$ to $45.3\\%$ under extremely limited data conditions", + "bbox": [ + 89, + 810, + 483, + 900 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/b05cacfdb88fae41dddb3a576b8a396fcf586d24cec5ed0e7025e057ff9a51d6.jpg", + "image_caption": [ + "Figure 2. Geometric shortcut. We select a point on the sofa arm and compute pairwise similarity with other points. The similarity heatmap reveals that CSC [38] collapses to surface normals, and MSC [88] overfits to point height. In contrast, our Sonata can extract higher-level concepts, as can be seen by the high similarity between all sofa arms highlighted in red." + ], + "image_footnote": [], + "bbox": [ + 513, + 88, + 903, + 306 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "(1% of ScanNet). Additionally, Sonata significantly boosts linear probing accuracy on ScanNet semantic segmentation, increasing it by over $3.3 \\times$ from $21.8\\%$ to $72.5\\%$ and surpassing the accuracy of DINOv2 features aggregated onto the point cloud (63.1%). Moreover, combining Sonata features with DINOv2 features further enhances accuracy (76.4%), underscoring that Sonata captures unique 3D information beyond what is visible in images alone. Finally, Sonata achieves state-of-the-art results across various indoor and outdoor perception tasks with full fine-tuning.", + "bbox": [ + 511, + 397, + 906, + 547 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 511, + 560, + 653, + 575 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Image self-supervised learning. Over the past decade, remarkable advancements [25, 26, 36, 59, 91] have been made in image self-supervised learning, and our research is largely inspired by two pivotal moments in this field. First, linear probing, a method that assesses representation quality by optimizing a minimal linear transformation, has become a standard in 2D image SSL [10, 13, 34, 91]. In some cases, such as when the distribution shift is large, linear probing surpasses full fine-tuning [105]. Second, the ability to directly perceive the semantic meaning of learned representations through zero-shot visualization like PCA or attention [12, 60] has further strengthened trust in reliability.", + "bbox": [ + 511, + 582, + 905, + 763 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Point self-supervised learning. Sonata follows the research path initiated by PointContrast [38, 93], emphasizing self-supervised learning with scene-level data [23]. While previous efforts do implement strategies to prevent collapse from geometric shortcuts, they remain limited. For example, Masked Scene Contrast (MSC) [88] encourages learning beyond naive geometric cues by predicting color or normal vectors but still partially anchors representations to predefined tasks. GroupContrast (GC) [80] employs graph-", + "bbox": [ + 511, + 765, + 905, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "22194", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/c61e24b74f3f68c8dd8dbf6f8b7bfde203cf8897155a2cb77a0785bd964a9ac1.jpg", + "image_caption": [ + "Figure 3. The geometric shortcut is unique to 3D. When comparing the information contained in 2D image and 3D point cloud data after removing the input feature (indicated by color), it is evident that in images all information is within the input feature. Whereas point clouds retain geometric information in point positions, which is directly utilized by operators. This characteristic leads to what we term geometric shortcuts in 3D SSL." + ], + "image_footnote": [], + "bbox": [ + 93, + 80, + 480, + 251 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "based segment guidance [28], though it is constrained by its reliance on human-designed algorithms. Building on MSC, Sonata directly addresses the geometric shortcut and scales up training [90] to establish a more reliable approach to point cloud self-supervised learning.", + "bbox": [ + 89, + 354, + 483, + 430 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Point cloud backbones. This area [17, 51, 56, 64, 67, 77] has significantly benefited from the U-Net structure [70], particularly with its hierarchical decoding and skip connections, as first introduced by PointNet++ [66]. However, the tight coupling between the encoder and decoder restricts flexibility and generalization capacity [92]. Sonata addresses this limitation by focusing self-supervised learning exclusively on the encoder, thus removing the hierarchical decoder. Moreover, unlike previous SOTAs [38, 88, 93] in point SSL, which primarily use SparseUNet [17, 20, 29] to balance efficiency and accuracy, we leverage Point Transformer V3 (PTv3) [89], an efficient, accurate, and scalable transformer backbone. This shift alone yields a $7.7\\%$ improvement in linear probing performance over MSC [88].", + "bbox": [ + 89, + 430, + 483, + 642 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Pilot Study and Design Principle", + "text_level": 1, + "bbox": [ + 89, + 652, + 387, + 670 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we qualitatively study the problems of current point self-supervised approaches and point cloud backbones to inform the Sonata approach.", + "bbox": [ + 89, + 674, + 482, + 718 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Uncovering the geometric shortcut in point SSL. The history of image self-supervised learning [34, 60, 104, 112] can be summarized as a continuous battle against shortcuts, where models often exploit trivial solutions (mode collapse) rather than understanding deeper semantics. Each advancement has involved identifying, understanding, and overcoming these shortcuts, refining pretext tasks to push models to \"struggle\" in learning stronger representations.", + "bbox": [ + 89, + 719, + 482, + 839 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "However in 3D point SSL, despite various attempts [38, 88, 93] to increase the difficulty of pretext tasks, a curious phenomenon persists—the loss consistently reduces rapidly to an ideal range in the early stages of training. We hypo-", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/666811aba4d3585c2cf6b2ea3f73df98e6e959b8fcc6969307f1f17a20722f56.jpg", + "image_caption": [ + "Figure 4. What is learned by the hierarchical backbone? We visualize PCA embeddings from different stages of a hierarchical encoder and decoder, trained for semantic segmentation. The encoder captures diverse and dispersed feature patterns, indicating a broad range of information. Notably, as the point cloud becomes coarser, accessible geometric information within point coordinates becomes increasingly global. In contrast, the decoder's representations are more uniform and structured, suggesting a focus on refining features for task-specific outputs." + ], + "image_footnote": [], + "bbox": [ + 516, + 80, + 903, + 166 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "esize that this lack of \"struggle\" in learning indicates that a shortcut exists that leads to a collapse of the representation to trivial solutions. Indeed as Fig. 2 illustrates, previous 3D SSL approaches seem to learn representations that are sensitive to local surface normals or point height. The collapse to such naive solutions even for well-designed SSL pretext tasks is what we term the geometric shortcut.", + "bbox": [ + 511, + 301, + 905, + 407 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Intuitively, representations affected by this geometric shortcut have not learned sufficient semantics. To quantify this problem, we leverage the standard criteria from 2D SSL of linear probing: we linearly probe the semantic class of each point from the learned representation. As illustrated in Fig. 1, with linear probing, previous 3D SSL methods PC [93] and MSC [88] indeed achieve only $5.6\\%$ mIoU and $21.8\\%$ mIoU respectively on ScanNet semantic segmentation. When compared to the $63.1\\%$ mIoU achieved by 3D-aggregated image representations from DINOv2 [60], it becomes obvious that the current 3D SSL methods do not learn semantic information.", + "bbox": [ + 511, + 410, + 905, + 590 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We hypothesize that the geometric shortcut stems from the sparse nature of point cloud data. Every operator, whether for point clouds or images, inherently relies on point (pixel) coordinates to define the kernel. However, unlike images with regularly spaced dense pixels, the sparsity of point cloud data necessitates introducing point coordinate information into point cloud operators rather than just through input features (see Fig. 3). This fact makes it difficult to obscure and nearly impossible to effectively mask out, ultimately resulting in the geometric shortcut.", + "bbox": [ + 511, + 594, + 905, + 744 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "These observations motivate the training methodology of Sonata aimed to prevent the dominance of naive spatial information inherent in point cloud organization.", + "bbox": [ + 511, + 747, + 905, + 792 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Focusing self-supervised learning on encoder only. Following our hypothesis that the geometric arrangement leaks information into the learned representation via the geometric shortcut, we examine the point cloud model. The typical U-Net structure [66, 70] is effective at handling large point clouds in a coarse-to-fine way, but the tight coupling of the encoder and decoder via skip connections restricts", + "bbox": [ + 511, + 795, + 905, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "22195", + "bbox": [ + 478, + 944, + 517, + 957 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "flexibility. Specifically, the decoder enforces per-point features at the original high-resolution scale, with shallow feature channels. This constraint limits the capacity to provide richer representations, which is crucial for SSL as evidenced by large channel dimensions in state-of-the-art 2D approaches [12]. Most importantly, decoding point clouds at the original scale unavoidably introduces local geometric cues into operators facilitating the geometric shortcut.", + "bbox": [ + 89, + 90, + 480, + 210 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In Fig. 4, we visualize PCA embeddings from different stages of PTv3 encoder and decoder, trained in a supervised manner for semantic segmentation. We observe that the encoder learns diverse features capturing spatial features at different scales in contrast to the decoder which produces task-specific higher-level representations. Additionally, as the spatial resolution of the point cloud decreases via the max-pooling stages, feature representations become less local. This resolution reduction fundamentally limits reliance on fine spatial information tied to point coordinates.", + "bbox": [ + 89, + 212, + 482, + 362 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "These observations motivate us to remove the decoder during self-supervised learning for two main reasons. First and foremost, training directly with features at coarser point resolutions inherently restricts access to fine-grained spatial information, reducing the possibility of geometric shortcuts. Second, task-specific features can be probed or fine-tuned on top of a more expressive multi-scale point representation.", + "bbox": [ + 89, + 363, + 482, + 470 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Point Self-distillation with Sonata", + "text_level": 1, + "bbox": [ + 89, + 479, + 395, + 494 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This section details the methodology of Sonata, the point self-distillation framework designed to address geometric shortcuts, remove structural constraints, and deliver strong linear probing results as discussed in Sec. 3. A roadmap of incremental ablation is illustrated in Fig. 6.", + "bbox": [ + 89, + 502, + 482, + 578 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Macro Framework", + "text_level": 1, + "bbox": [ + 89, + 582, + 272, + 595 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Before diving into the specific micro designs, we begin with a point self-distillation framework derived from insights gained through previous efforts in point [88, 93] and image [11, 34, 60, 104, 110] self-supervised learning. This macro framework provides a solid foundation for pretext task design, allowing us to focus on addressing the geometric shortcut without additional concerns.", + "bbox": [ + 89, + 597, + 482, + 702 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In essence, (point) self-supervised learning aims to make things (points) that should be the same, the same (identical in representation). This forms the basic recipe of point SSL: generating two views of a given point cloud with random spatial (e.g., crop, rotate, distort) and photometric (e.g., jitter) augmentations, then matching and aligning the feature embeddings of points that are close in the original space.", + "bbox": [ + 89, + 704, + 482, + 809 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "However, the superior robustness of self-supervised representation is rooted in a core principle: continuously increasing the difficulty of pretext tasks as long as the model continues to converge. This principle encourages enhancing the basic recipe with local-global view alignment and mask-unmask view alignment. Specifically, it involves aligning", + "bbox": [ + 89, + 810, + 482, + 901 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/6cc7c5098377002cf7d37801254da98aad83255b8febd0aa13ed7539d6847ec5.jpg", + "image_caption": [ + "Figure 5. Self-distillation framework of Sonata. (1) Local views (bottom left) and global views (right) are generated with dedicated spatial and photometric augmentations, while masked views are created by randomly masking out grid-based patches from the global views (top left). (2) Embeddings from local and masked views are extracted by the student, with global views processed by the teacher (top). (3) Points from local and masked views are matched with corresponding points in the global views based on their original spatial distance, allowing for the distillation of embeddings from global views to local and masked views (bottom)." + ], + "image_footnote": [], + "bbox": [ + 517, + 82, + 898, + 268 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "neighboring points from local views generated by small cropping ratios and masked views created with large masking ratios (see Fig. 5 left), with unmasked global views that contain relatively richer information (see Fig. 5 right). The difficulty of pretext tasks can be scaled by adjusting the cropping and masking ratios.", + "bbox": [ + 511, + 415, + 903, + 506 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The more challenging the pretext tasks become, the higher the risk of model collapse. This instability calls for an asymmetric encoding approach, specifically exponential moving average (EMA) [34]: rather than encoding all views with a shared-weight model, the approach encodes challenging local and masked views with an actively learned student model, while using a stable teacher model, updated with a moving average of the student parameters, to encode global views (see Fig. 5 top). With the teacher preventing the student from being misled, the student is less likely to get lost in \"mission impossible\" and has a greater chance to discover treasure within \"impossible\" (i.e. extreme $5\\%$ crop ratio for local views and $70\\%$ mask ratio for masked views).", + "bbox": [ + 511, + 507, + 905, + 702 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In terms of SSL criteria, we move away from prior approaches that rely on contrastive and generative learning [88]. Contrastive learning, limited by the number of point pairs in pairwise similarity computations, restricts scalability. Generative learning, meanwhile, partially anchors representations to predefined cues, limiting the model's capacity to capture more generalizable features. Following DINov2, we adopt a self-distillation approach driven by Sinkhorn-Knopp centering [11], KoLeo regularization [72], and clustering assignments [11]. This adaptation initially exacerbates collapse into geometric shortcuts but holds greater potential for robust representation once this challenge is addressed.", + "bbox": [ + 511, + 704, + 905, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "22196", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Micro Design", + "text_level": 1, + "bbox": [ + 89, + 90, + 230, + 107 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We now discuss micro designs aimed at addressing geometric shortcuts. Since the problematic spatial information is inherently tied to point coordinates and directly used by operators, it is nearly impossible to mask out. This constraint defines the key strategies of our micro designs: obscuring spatial information and emphasizing input features.", + "bbox": [ + 89, + 108, + 483, + 199 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Decoder removal. Previous approaches adhere to the original U-Net-style backbone for feature extraction; however, our observations in Sec. 3 motivate us to remove the complex hierarchical decoder and perform self-distillation directly using the encoder's output. This simple adjustment is key in our fight against geometric shortcuts, as it: (a) increases the feature channels participating in self-distillation (from 96 to 512), (b) streamlines the pipeline by involving fewer points in the pretext task after hierarchical pooling, and, most importantly, (c) introduces a natural way to obscure geometric cues: the positional information of points becomes naturally disturbed during hierarchical encoding and pooling. In fact, this removal proves to be even more beneficial than expected, boosting the linear probing result from $20.7\\%$ to $60.4\\%$ .", + "bbox": [ + 89, + 200, + 483, + 426 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Feature up-casting. While the removal effectively prevents over-reliance on naive geometric cues, it also introduces certain limitations, particularly in leveraging multiscale contexts. In the original U-Net structure, the decoder plays a key role in progressively aggregating features across scales to reconstruct semantic details. Without this process, self-distillation struggles to capture multi-scale spatial information and broader contextual relationships. To retain multi-scale features, we introduce a parameter-free feature up-casting process similar to hypercolumns in image segmentation [33]: progressively up-casting features back to the scale of the previous encoding stage, with the mapping relationships preserved through pooling layers and concatenation with features from the prior encoding stage. This approach provides richer, coarse-to-fine features from multiscale encoding. While it does increase the risk of the model falling into geometric shortcuts, finding the right balance is key. Our ablation study shows that up-casting features twice achieve the best performance.", + "bbox": [ + 89, + 429, + 483, + 715 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Masked points jitter. A slight random Gaussian jitter $(\\sigma = 0.005)$ is applied to point coordinates as part of data augmentation. However, specifically for points to be masked, we additionally apply a stronger Gaussian jitter $(\\sigma = 0.01)$ to further disrupt their spatial relationships. We pay special attention to these masked points because models are more likely to collapse to naive geometric cues, especially when point features are masked, making it difficult to derive solutions from neighboring unmasked points.", + "bbox": [ + 89, + 717, + 483, + 853 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Progressive parameter scheduler. Geometric shortcuts are like traps on the convergence path of point cloud self-supervised learning. Similarly, we can set our own \"trap\"", + "bbox": [ + 89, + 854, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/8a03d2c187c4b70f274a3c5da3bebb8c98825d9400e2e0661479788c0b34228e.jpg", + "image_caption": [ + "Figure 6. The roadmap. We evolve Mask Scene Contrast [88] into our Sonata by modernizing self-supervised learning with self-distillation, addressing the geometric shortcut, and scaling up training. Our designs are validated through progressive ablation with linear and decoder probing on ScanNet semantic segmentation [23]. Starting with $23\\mathrm{k}$ training data (a combination of ScanNet and Structured3D [109]) and a $39\\mathrm{M}$ PTv3 model [89], we ultimately scale up to $140\\mathrm{k}$ assets (Tab. 1) and a $108\\mathrm{M}$ PTv3 model." + ], + "image_footnote": [], + "bbox": [ + 521, + 84, + 903, + 450 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "during training. For example, rather than starting with a challenging large mask size and mask ratio, we begin with a relatively small mask size (10 cm) and mask ratio (30%), gradually increasing them to 40 cm (0.1 $\\Rightarrow$ 0.4) and 70% (0.3 $\\Rightarrow$ 0.7) over the first 5% of the training process. This strategy encourages the model to rely more on input features to solve the pre-defined pretext tasks, preventing it from shifting its reliance to point coordinates as training difficulty increases. This approach aligns with curriculum learning [7], progressively challenging the model as it adapts. Similarly, in addition to the common learning rate scheduler, we also implement custom progressive schedulers for teacher temperature (0.04 $\\Rightarrow$ 0.07) and weight decay (0.04 $\\Rightarrow$ 0.2). We found that this design pushes these parameters to more extreme levels previously unexplored, further enhancing model performance.", + "bbox": [ + 511, + 578, + 906, + 820 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Implementation and Evaluation Protocols", + "text_level": 1, + "bbox": [ + 511, + 823, + 870, + 838 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we introduce the implementation details and the evaluation protocols for our experiments.", + "bbox": [ + 511, + 839, + 903, + 869 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Backbone. We build our Sonata with Point Transformer V3 (PTv3) [89] and refer to Pointcept [19] for details of imple", + "bbox": [ + 511, + 869, + 903, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "22197", + "bbox": [ + 478, + 944, + 517, + 957 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/3835880d60da44d365e2818248dcdca8d0a3030064241bfd58cd9ec893d644a8.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetSourceTrainValTestTotal
ScanNet [23]real1,2013121001,613
ScanNet++ [101]real7121781261,016
S3DIS [1]real204680272
ArkitScenes [5]real4,49854905,047
HM3D [68]real8,8811,119010,000
Structured3D [109]sim.18,3481,7761,69721,821
ASE [3]sim.90,00010,0000100,000
Sonata (ours)mixed123,84414,0021,923139,769
", + "bbox": [ + 93, + 88, + 480, + 227 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/fd73843727f38e6b0dfe20421a52580b86b1ad863350b8a6b3dcbf5ac93c1e76.jpg", + "table_caption": [ + "Table 1. Data source collection." + ], + "table_footnote": [], + "table_body": "
MethodRealSimTotalMultiplier
PC [93]1,61301,613×1
MSC [88]6,66006,660×4.1
PPT [90]1,88521,82123,706×14.7
Sonata (ours)17,948121,821139,768×86.7
", + "bbox": [ + 96, + 246, + 472, + 327 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Data scale comparison.", + "bbox": [ + 187, + 329, + 385, + 342 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "mentation. Building on this, we made an additional adjustment to enhance scalability: replacing all Batch Normalization (BN) [41] layers with Layer Normalization (LN) [4]. Although this replacement results in some initial accuracy degradation, it enhances domain adaptation by eliminating the need for additional domain-specific adjustments when scaling up with multi-dataset joint training [90]. Along with the scaling up of data, we also scale up the encoder block depths from $[2, 2, 2, 6, 2]$ to $[3, 3, 3, 12, 3]$ and widths from $[32, 64, 128, 256, 512]$ to $[48, 96, 192, 384, 512]$ . This PTv3 model has $108\\mathrm{M}$ parameters.", + "bbox": [ + 89, + 354, + 482, + 520 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Data. We extend the multi-dataset joint training approach introduced by PPT [90], further expanding the data scale by removing the constraint of human labeling through unsupervised learning. This results in a collection of 140k scene-level point clouds from both real-world and simulated environments (outlined in Tab. 1), making it $86.7 \\times$ larger than the data scale of PointContrast [93] and $5.9 \\times$ larger than the data collection of PPT, as detailed in Tab. 2.", + "bbox": [ + 89, + 521, + 482, + 642 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Training. We train Sonata on the 140k data collection for 200 epochs, using the AdamW optimizer [54] with a batch size of 96, distributed across 32 GPUs. The learning rate linearly warms up over the first 10 epochs to a base value of 0.004, then decays following a cosine schedule [55]. Additionally, a layer-wise learning rate decay of 0.9 is applied to model parameters [111]. Weight decay is also controlled by a cosine schedule, progressively increasing from 0.04 to 0.2. For EMA, the student temperature is set to 0.1, while the teacher temperature gradually rises from 0.04 to 0.07 during the first 10 epochs [60]. The momentum starts at 0.994 and increases to 1 by the final iteration. For data augmentation and view generation, we follow the augmentation pipeline designed by MSC [88]. We generate 2 global views (sampling $40\\%$ to $100\\%$ of scene points) and 4 local views (sampling $5\\%$ to $40\\%$ of scene points) for training, with 2 masked views generated based on the global views.", + "bbox": [ + 89, + 643, + 482, + 901 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/06aa10040be016adeacb37fd5767052dffa88a788de5bc8d35dfdbf31a2a7d1a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 517, + 88, + 612, + 160 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/71d19533408976e6644598cdf508d18a7c5e44b67d330f9724f037577f5ccdcb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 614, + 88, + 903, + 161 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/dcb26bf0cf2ce76820349c2c9376b5eed8bb8c499453a344a54c3494c4ab1d93.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 517, + 161, + 612, + 233 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/82222571f32c270eb9e8c30df905551ea5b2080d6d6f0e6479638834f53e2866.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 614, + 162, + 903, + 234 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/44a957cddc13afa6e27b5da41df947c9bf978364d7f07cc157fac6052703723a.jpg", + "image_caption": [ + "Figure 7. Zero-shot comparison with DINOv2. We compare the PCA visualizations of DINOv2, Sonata, and their combined feature representation. DINOv2 excels at capturing photometric details, while Sonata better distinguishes spatial information. The combined model demonstrates improved coherence and detail, showcasing the complementary strengths of both models." + ], + "image_footnote": [], + "bbox": [ + 517, + 234, + 612, + 306 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/aa552af8ddd0526f922499b890339fd7daffb2482750d8c8e5ac2f10a3aa2d92.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 614, + 234, + 903, + 306 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation. We evaluate the quality of the learned representation using the following three protocols after initializing the encoder with Sonata:", + "bbox": [ + 511, + 407, + 903, + 452 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- In linear probing, we keep the encoder frozen and upcast the features to their original scale. A single linear layer, comprising less than $0.2\\%$ of the total parameters, is then used to adapt these features to downstream tasks.", + "- In decoder probing, we take a step back and reintroduce a lightweight hierarchical decoder, which accounts for $13\\%$ of the total parameters. We then freeze the encoder, allowing only the decoder to actively learn.", + "- In full fine-tuning, we follow the traditional approach by unfreezing the entire PTv3 U-Net-style backbone, tuning the learned representation to downstream tasks." + ], + "bbox": [ + 521, + 453, + 903, + 619 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We advocate linear probing as the primary evaluation criterion for point SSL, considering other methods as intermediate steps. We look forward to the day when a linear-probed self-supervised model outperforms a fully fine-tuned one.", + "bbox": [ + 511, + 619, + 905, + 679 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Main Results", + "text_level": 1, + "bbox": [ + 511, + 686, + 648, + 700 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We validate the reliability of the Sonata representation using the evaluation protocols discussed in Sec. 4.3 and analyze the main properties based on these results.", + "bbox": [ + 511, + 704, + 905, + 748 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Comparison with image self-supervised model. In Tab. 3, we compare the linear probing and decoder probing results on ScanNet and ScanNet200 semantic segmentation with the linear probing accuracy of representations transferred from the image self-supervised models DINOv2 [60] and DINOv2.5 [24]. Specifically, we aggregate unprojected pixel embeddings using ground truth camera poses and depth calculated by ray intersection with a reconstructed mesh, which provides more accuracy than sensor per-frame depth. Our results indicate that while the DINOs' represent", + "bbox": [ + 511, + 750, + 906, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "22198", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/3c9033ea0b3d84da3d0cd5cd522a5addb74d470841676098228016ea58820fd8.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
2D × 3DScanNet Val [23]ScanNet200 Val [23]
MethodsmIoUmAccallAccmIoUmAccallAcc
• DINOv2 (lin.) [60]63.0975.5082.4227.4237.5972.80
• DINOv2.5 (lin.) [24]63.3675.9482.3027.7539.2372.53
• Sonata (lin.)72.5283.1189.7429.2541.6181.15
+DINOv2 (lin.)75.9185.3691.2536.6746.9882.85
+DINOv2.5 (lin.)76.4485.6891.3336.9648.2382.77
• Sonata (dec.)79.0786.5792.6833.5444.4884.07
+DINOv2 (dec.)79.1287.2392.4737.7349.3883.31
+DINOv2.5 (dec.)79.1986.6692.5038.2748.5783.77
", + "bbox": [ + 91, + 88, + 457, + 244 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/74276ce4432bdf579eebed4474ab5b041af07d66bcc013f40763b9bafac876fe.jpg", + "table_caption": [ + "Table 3. Numerical comparison with DINO series." + ], + "table_footnote": [], + "table_body": "
Data EfficiencyLimited Scenes (Pct.)Limited Annotation (Pts.)
Methods1%5%10%20%Full2050100200Full
○ SparseUNet [17]26.047.856.762.972.241.953.962.265.572.2
• CSC [38]28.949.859.464.673.855.560.565.968.273.8
• MSC [88]29.250.761.064.975.461.065.668.969.675.4
○ PTv2 [87]24.848.159.866.375.458.466.170.371.275.4
○ PTv3 [89]25.848.961.067.077.260.167.971.472.777.2
• PPT [90] (sup.)31.152.663.368.278.262.469.174.375.578.2
• Sonata (lin.)43.662.568.669.872.569.070.571.171.572.5
• Sonata (dec.)44.564.169.872.579.169.873.175.076.379.1
• Sonata (full)45.365.772.472.879.470.573.676.077.079.4
", + "bbox": [ + 465, + 90, + 903, + 244 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/673a21537d1f42efbee2d20b90e3cb6de7f05fe926829185f23d551e69643274.jpg", + "table_caption": [ + "Table 4. Data efficiency." + ], + "table_footnote": [], + "table_body": "
Param. Efficiency MethodsParamsScanNet Val [23]ScanNet200 Val [71]ScanNet++ Val [101]S3DIS Area 5 [1]S3DIS 6-fold [1]
Learn.Pct.mIoUmAccallAccmIoUmAccallAccmIoUmAccallAccmIoUmAccallAccmIoUmAccallAcc
○ SparseUNet [17]39.2M100%72.380.290.025.032.980.428.838.480.166.372.589.872.480.989.9
• PC [93] (lin.)<0.2M<0.1%5.69.750.00.50.940.31.83.146.411.418.652.311.719.051.2
• CSC [38] (lin.)<0.2M<0.1%12.618.164.21.32.153.02.84.553.624.432.066.424.932.566.9
• MSC [88] (lin.)<0.2M<0.1%14.120.362.91.52.553.64.56.661.327.935.571.129.937.971.3
○PTv3 [89]124.8M100%77.685.092.035.346.083.442.153.485.673.478.991.777.785.391.5
• MSC [88] (lin.)<0.2M<0.2%21.832.265.53.35.557.58.111.964.732.142.470.934.646.071.3
• Sonata (lin.)<0.2M<0.2%72.583.189.729.341.681.237.350.984.372.381.290.976.587.490.8
• Sonata (dec.)16.3M13%79.186.692.733.544.584.140.952.686.374.580.492.681.588.893.0
", + "bbox": [ + 91, + 266, + 898, + 422 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 5. Parameter efficiency.", + "bbox": [ + 395, + 425, + 583, + 439 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "tation demonstrates impressive robustness, Sonata offers a more suitable representation for 3D tasks, achieving $+9.2\\%$ on ScanNet and $+1.5\\%$ on ScanNet200 for semantic segmentation. Furthermore, combining Sonata with the DINOs yields higher accuracy than any single data modality alone $(+3.9\\%$ and $+7.7\\%$ , respectively), underscoring the promising potential of cross-modal self-distillation. Additional zero-shot comparisons (see Fig. 7) through PCA visualizations further corroborate these numerical findings.", + "bbox": [ + 88, + 448, + 480, + 585 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Data efficiency. In Tab. 4, we present the semantic segmentation performance of Sonata when probed or fine-tuned on the ScanNet dataset with limited scenes and annotations [38]. The results demonstrate the exceptional data efficiency of Sonata, with improvements of $19.5\\%$ in extreme data scarcity (1% of scenes) and $10.4\\%$ in limited annotation scenarios (20 points per scene) compared to training from scratch. Notably, even linear probing surpasses previous SOTA by a substantial margin (12.5% with 1% scenes), highlighting Sonata's reliability in low-data scenarios.", + "bbox": [ + 88, + 590, + 482, + 743 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Parameter efficiency. In Tab. 5, we demonstrate parameter efficiency using both linear and decoder probing across various indoor semantic segmentation benchmarks, including ScanNet [23], ScanNet200 [71], ScanNet++ [101], S3DIS Area5[1], and S3DIS 6-fold cross-validation [65]. Semantic segmentation is emphasized as it provides a direct measure of point cloud representation quality. Results show that a single linear layer with a negligible number of parameters (<0.02% of total parameters) is sufficient for Sonata to achieve strong performance on these benchmarks (e.g.,", + "bbox": [ + 88, + 750, + 482, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "72.5% on ScanNet and 73.4% on S3DIS Area5). Furthermore, probing a decoder with only 13% of the model's parameters yields even higher accuracy (e.g., 79.1% on ScanNet and 81.5% on S3DIS 6-fold cross-validation). However, while decoder probing achieves SOTA results on ScanNet (20 classes) and S3DIS (13 classes), performance on ScanNet200 (200 classes) and ScanNet++ (100 classes) remains limited. This shows a limitation of the learned representation in distinguishing a large number of classes.", + "bbox": [ + 511, + 448, + 903, + 584 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Indoor semantic segmentation. In Tab. 6, we further enhance Sonata's semantic segmentation accuracy through full fine-tuning, consistently pushing SOTA results to new heights across the five widely recognized benchmarks, e.g., achieving $79.4\\%$ on ScanNet and $82.3\\%$ on S3DIS 6-fold cross-validation. However, we view full fine-tuning as an intermediate step toward a future where linear probing surpasses it. At present, full fine-tuning remains essential to achieve the highest performance on these benchmarks and close the remaining gap of $7.1\\%$ and $5.8\\%$ respectively.", + "bbox": [ + 511, + 590, + 905, + 742 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Indoor instance segmentation. In Tab. 7, We also validate the robustness of Sonata representation on indoor instance segmentation benchmarks, including ScanNet [23], ScanNet200 [71], ScanNet++ [101], and S3DIS [1]. Consistent with our findings in semantic segmentation, Sonata demonstrates strong parameter efficiency, achieving significant improvements with linear probing (10× mAP50 on ScanNet and 21× on ScanNet200) and decoder probing (12× mAP50 on ScanNet and 33× on ScanNet200). Full fine-tuning further boosts these results, achieving SOTA", + "bbox": [ + 511, + 750, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "22199", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/0a34a86f99f1624747ff060519ccc3e421dfeb2b64d76c2e1d805ff2a8146b9f.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Indoor Sem. SegParamsScanNet Val [23]ScanNet200 Val [71]ScanNet++ Val [101]S3DIS Area 5 [1]S3DIS 6-fold [1]
MethodsLearn.Pct.mIoUmAccallAccmIoUmAccallAccmIoUmAccallAccmIoUmAccallAccmIoUmAccallAcc
○ SparseUNet [17]39.2M100%72.380.290.025.032.980.428.838.480.166.372.589.872.480.989.9
• PC [93]39.2M100%72.380.990.126.233.079.929.239.782.768.173.590.074.783.390.6
• CSC [38]39.2M100%72.881.090.726.933.780.632.541.183.770.776.490.875.584.090.9
• MSC [88]39.2M100%75.783.491.332.041.682.339.449.684.970.776.191.077.485.391.5
○PTv3 [89]124.8M100%77.685.092.035.346.083.442.153.485.673.478.991.777.785.391.5
• MSC [88]124.8M100%78.285.392.233.443.783.442.453.685.969.974.991.277.484.791.7
• PPT [90] (sup.)124.8M100%78.685.992.336.046.283.843.355.786.474.380.192.080.887.792.6
• Sonata124.8M100%79.486.192.536.846.584.443.755.886.676.081.693.082.389.993.3
", + "bbox": [ + 91, + 88, + 901, + 244 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/725e6fcaf9b132f2531998223fa93a4dd29f8794ba8eba66c8f5d2f77abbddb6.jpg", + "table_caption": [ + "Table 6. Indoor semantic segmentation." + ], + "table_footnote": [], + "table_body": "
Indoor Ins. SegParamsScanNet Val [23]ScanNet200 Val [71]ScanNet++ Val [101]S3DIS Area 5 [1]
MethodsLearn.Pct.mAP25mAP50mAPmAP25mAP50mAPmAP25mAP50mAPmAP25mAP50mAP
○PointGroup [45]124.8M100%77.561.740.940.133.223.136.730.720.955.749.437.8
•MSC (lin.)<0.2M<0.2%13.35.32.32.31.00.44.82.61.319.013.09.7
•Sonata (lin.)<0.2M<0.2%72.653.930.730.921.310.931.622.412.245.836.626.1
•Sonata (dec.)16.3M13%76.862.840.840.833.322.838.129.118.863.757.145.1
•MSC (full)124.8M100%78.462.941.140.533.823.438.930.921.756.350.538.1
•PPT [90] (sup.)124.8M100%78.963.542.140.834.124.039.332.821.957.551.239.7
•Sonata (full)124.8M100%79.263.942.442.135.625.440.333.622.363.857.445.5
", + "bbox": [ + 91, + 265, + 901, + 411 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/6ec5eca4dfc1955ebcbdb48b12537d4a84c76555532556618daddc271253a44b.jpg", + "table_caption": [ + "Table 7. Indoor instance segmentation." + ], + "table_footnote": [], + "table_body": "
Outdoor Sem. Seg.nuScenes Val [9]Waymo Val [74]Sem.KITTI Val [6]
MethodsmIoUmAccallAccmIoUmAccallAccmIoUmAcc allAcc
o PTv3 [89]80.487.294.771.380.594.769.176.192.6
• Sonata (lin.)66.177.292.460.572.592.562.072.591.0
• Sonata (dec.)77.385.994.270.878.894.368.476.592.3
• PPT [90] (sup.)81.287.594.872.181.394.872.377.593.4
• Sonata (full)81.787.995.072.981.994.972.677.993.4
", + "bbox": [ + 91, + 438, + 480, + 553 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 8. Outdoor semantic segmentation.", + "bbox": [ + 156, + 555, + 418, + 569 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "pre-training performance across benchmarks. This clearly demonstrates that, unlike previous SSL approaches, the Sonata representation encodes instance-level information.", + "bbox": [ + 89, + 574, + 482, + 618 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Outdoor semantic segmentation. In Tab. 8, we adapt pre-training paradigm of Sonata to outdoor LiDAR scenarios through joint training on nuScenes [9], Waymo [74], and SemanticKITTI [6] and evaluate semantic segmentation performance using our evaluation protocols. With linear probing, Sonata sets a robust parameter efficiency baseline. Decoder probing achieves significant gains. In full fine-tuning, Sonata surpasses the supervised PPT [90], establishing new SOTA mIoU scores of 81.7, 72.9, and 72.6 across these benchmarks, underscoring the effectiveness of Sonata in outdoor perception tasks. Note that most of that performance can be recovered by more efficient decoder-only probing with a $95\\%$ , $97\\%$ , and $94\\%$ of full fine-tuning.", + "bbox": [ + 88, + 619, + 482, + 815 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion and Discussion", + "text_level": 1, + "bbox": [ + 89, + 821, + 339, + 837 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we make progress towards self-supervised learning of a strong and reliable 3D point representation, that can zero-shot correspond semantically similar 3D points, down to the instance level, even under the presence", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "of spatial and visual perturbations. We demonstrate that such a representation can serve as the foundation for 3D tasks in semantic and instance-level grouping. We find existing 3D self-supervised learning approaches lacking, and hypothesize that this is due to what we term the geometric shortcut, a problem unique to 3D, which causes representations to collapse to low-level spatial features. We demonstrate the deficiency of these collapsed representations through linear probing. Beginning with a point self-distillation framework, we tackle the geometric shortcut by attaching SSL losses at coarser spatial scales, disturbing the spatial information of masked points with no features, and progressively increasing task difficulty to prevent overreliance on accessible geometric cues. This change enables effective scaling up, ultimately composing a Sonata from 140k point clouds. Sonata demonstrates semantically meaningful zero-shot visualization, as well as exceptional parameter and data efficiency. Full fine-tuning further advances SOTA across 3D indoor and outdoor perception tasks.", + "bbox": [ + 511, + 436, + 906, + 724 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We summarize our limitations and future works and encourage readers to refer to the supplementary for details: 1. Sonata representations could be further semantically enriched through augmentation with 1M object-level point clouds [53]; 2. Pursuing unified training across indoor and outdoor scenarios as a feasible and promising direction; 3. Scaling up training with pixel-aligned [8] and SLAM-generated [27] point clouds from video data; 4. Leveraging cross-modal distillation to allow Sonata and DINOV2 to complement each other. We hope our insights will contribute to future research in 3D representation learning.", + "bbox": [ + 511, + 734, + 908, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "22200", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 91, + 90, + 258, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We extend our gratitude to Maxime Oquab and Piotr Bojanowski for their guidance on the DINOv2 training details, to Saining Xie for insightful discussions on the vision of 3D representation learning, and to Paul Mcvay for his thoughts on the JPEA framework with sparse point cloud data.", + "bbox": [ + 89, + 114, + 483, + 191 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 203, + 187, + 218 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Iro Armeni, Ozan Sener, Amir R. Zamir, Helen Jiang, Ioannis Brilakis, Martin Fischer, and Silvio Savarese. 3d semantic parsing of large-scale indoor spaces. In CVPR, 2016. 2, 6, 7, 8, 16", + "[2] Mahmoud Assran, Quentin Duval, Ishan Misra, Piotr Bojanowski, Pascal Vincent, Michael Rabbat, Yann LeCun, and Nicolas Ballas. Self-supervised learning from images with a joint-embedding predictive architecture. In CVPR, 2023. 2", + "[3] Armen Avetisyan, Christopher Xie, Henry Howard-Jenkins, Tsun-Yi Yang, Samir Aroudj, Suvam Patra, Fuyang Zhang, Duncan Frost, Luke Holland, Campbell Orme, et al. Scenescript: Reconstructing scenes with an autoregressive structured language model. In ECCV, 2024. 2, 6, 16", + "[4] Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. Stat, 2016. 6", + "[5] Gilad Baruch, Zhuoyuan Chen, Afshin Dehghan, Tal Dimry, Yuri Feigin, Peter Fu, Thomas Gebauer, Brandon Joffe, Daniel Kurz, Arik Schwartz, and Elad Shulman. ARKitscenes - a diverse real-world dataset for 3d indoor scene understanding using mobile RGB-d data. In NeurIPSW, 2021. 2, 6", + "[6] Jens Behley, Martin Garbade, Andres Milioto, Jan Quenzel, Sven Behnke, Cyril Stachniss, and Jurgen Gall. Semantickitti: A dataset for semantic scene understanding of lidar sequences. In ICCV, 2019. 8", + "[7] Yoshua Bengio, Jérôme Louradour, Ronan Collobert, and Jason Weston. Curriculum learning. In ICML, 2009. 5", + "[8] Aleksei Bochkovskii, Amael Delaunoy, Hugo Germain, Marcel Santos, Yichao Zhou, Stephan R Richter, and Vladlen Koltun. Depth pro: Sharp monocular metric depth in less than a second. arXiv preprint arXiv:2410.02073, 2024. 8, 13", + "[9] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In CVPR, 2020. 2, 8", + "[10] Mathilde Caron, Piotr Bojanowski, Armand Joulin, and Matthijs Douze. Deep clustering for unsupervised learning of visual features. In ECCV, 2018. 2", + "[11] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. In NeurIPS, 2020. 2, 4", + "[12] Mathilde Caron, Hugo Touvron, Ishan Misra, Herve Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin." + ], + "bbox": [ + 99, + 227, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Emerging properties in self-supervised vision transformers. In CVPR, 2021. 2, 4", + "[13] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In ICML, 2020. 2", + "[14] Xi Chen, Lianghua Huang, Yu Liu, Yujun Shen, Deli Zhao, and Hengshuang Zhao. Anydoor: Zero-shot object-level image customization. In CVPR, 2024. 2", + "[15] Yukang Chen, Jianhui Liu, Xiangyu Zhang, Xiaojuan Qi, and Jiaya Jia. Largekernel3d: Scaling up kernels in 3d sparse cnns. In CVPR, 2023. 16", + "[16] Hung-Yueh Chiang, Yen-Liang Lin, Yueh-Cheng Liu, and Winston H Hsu. A unified point-based framework for 3d segmentation. In 3DV, 2019. 16", + "[17] Christopher Choy, JunYoung Gwak, and Silvio Savarese. 4d spatio-temporal convnets: Minkowski convolutional neural networks. In CVPR, 2019. 3, 7, 8, 16, 17", + "[18] Christopher Choy, Jaesik Park, and Vladlen Koltun. Fully convolutional geometric features. In ICCV, 2019. 2", + "[19] Pointcept Contributors. Pointcept: A codebase for point cloud perception research. https://github.com/Pointcept/Pointcept, 2023.5", + "[20] Spconv Contributors. Spconv: Spatially sparse convolution library. https://github.com/traveller59/spconv, 2022.3", + "[21] Brian Curless and Marc Levoy. A volumetric method for building complex models from range images. In Proceedings of the 23rd annual conference on Computer graphics and interactive techniques, pages 303-312, 1996. 16", + "[22] Angela Dai and Matthias Nießner. 3dmv: Joint 3d-multiview prediction for 3d semantic scene segmentation. In ECCV, 2018. 16", + "[23] Angela Dai, Angel X. Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In CVPR, 2017. 2, 5, 6, 7, 8, 16", + "[24] Timothee Darcet, Maxime Oquab, Julien Mairal, and Piotr Bojanowski. Vision transformers need registers. arXiv:2309.16588, 2023. 6, 7", + "[25] Carl Doersch, Abhinav Gupta, and Alexei A Efros. Unsupervised visual representation learning by context prediction. In ICCV, 2015. 2", + "[26] Alexey Dosovitskiy, Philipp Fischer, Jost Tobias Springenberg, Martin Riedmiller, and Thomas Brox. Discriminative unsupervised feature learning with exemplar convolutional neural networks. TPAMI, 2015. 2", + "[27] Jakob Engel, Kiran Somasundaram, Michael Goesele, Albert Sun, Alexander Gamino, Andrew Turner, Arjang Talattof, Arnie Yuan, Bilal Souti, Brighid Meredith, et al. Project aria: A new tool for egocentric multi-modal ai research. arXiv:2308.13561, 2023. 2, 8, 13", + "[28] Pedro F Felzenszwalb and Daniel P Huttenlocher. Efficient graph-based image segmentation. IJCV, 2004. 3", + "[29] Benjamin Graham and Laurens van der Maaten. Submanifold sparse convolutional networks. arXiv:1706.01307, 2017. 3" + ], + "bbox": [ + 522, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "22201", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[30] Benjamin Graham, Martin Engelcke, and Laurens van der Maaten. 3d semantic segmentation with submanifold sparse convolutional networks. In CVPR, 2018. 16", + "[31] Jean-Bastien Grill, Florian Strub, Florent Alché, Coretin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. In NerIPS, 2020. 2", + "[32] Jiayuan Gu, Fanbo Xiang, Xuanlin Li, Zhan Ling, Xiqiang Liu, Tongzhou Mu, Yihe Tang, Stone Tao, Xinyue Wei, Yunchao Yao, Xiaodi Yuan, Pengwei Xie, Zhiao Huang, Rui Chen, and Hao Su. Maniskill2: A unified benchmark for generalizable manipulation skills. In ICLR, 2023. 2", + "[33] Bharath Hariharan, Pablo Arbelaez, Ross Girshick, and Jitendra Malik. Hypercolumns for object segmentation and fine-grained localization. In CVPR, 2015. 5", + "[34] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In CVPR, 2020. 2, 3, 4", + "[35] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In CVPR, 2022. 2", + "[36] R. Devon Hjelm, Alex Fedorov, Samuel Lavoie-Marchildon, Karan Grewal, Phil Bachman, Adam Trischler, and Yoshua Bengio. Learning deep representations by mutual information estimation and maximization. In ICLR, 2018. 2", + "[37] Yicong Hong, Kai Zhang, Jiuming Gu, Sai Bi, Yang Zhou, Difan Liu, Feng Liu, Kalyan Sunkavalli, Trung Bui, and Hao Tan. Lrm: Large reconstruction model for single image to 3d. arXiv preprint arXiv:2311.04400, 2023. 16", + "[38] Ji Hou, Benjamin Graham, Matthias Nießner, and Saining Xie. Exploring data-efficient 3d scene understanding with contrastive scene contexts. In CVPR, 2021. 2, 3, 7, 8, 16, 17", + "[39] Qingyong Hu, Bo Yang, Linhai Xie, Stefano Rosa, Yulan Guo, Zhihua Wang, Niki Trigoni, and Andrew Markham. Randla-net: Efficient semantic segmentation of large-scale point clouds. In CVPR, 2020. 16", + "[40] Zeyu Hu, Mingmin Zhen, Xuyang Bai, Hongbo Fu, and Chiew-lan Tai. Jsenet: Joint semantic segmentation and edge detection network for 3d point clouds. In ECCV, 2020. 16", + "[41] Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In ICML, 2015. 6", + "[42] Ayush Jain, Pushkal Katara, Nikolaos Gkanatsios, Adam W. Harley, Gabriel Sarch, Kriti Aggarwal, Vishrav Chaudhary, and Katerina Fragkiadaki. Odin: A single model for 2d and 3d perception. In CVPR, 2024. 16", + "[43] Stephen James, Zicong Ma, David Rovick Arrojo, and Andrew J. Davison. Rlbench: The robot learning benchmark & learning environment. RAL, 2020. 2", + "[44] Li Jiang, Hengshuang Zhao, Shu Liu, Xiaoyong Shen, Chi-Wing Fu, and Jiaya Jia. Hierarchical point-edge interaction network for point cloud semantic segmentation. In ICCV, 2019. 17" + ], + "bbox": [ + 99, + 92, + 482, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[45] Li Jiang, Hengshuang Zhao, Shaoshuai Shi, Shu Liu, Chi-Wing Fu, and Jiaya Jia. Pointgroup: Dual-set point grouping for 3d instance segmentation. CVPR, 2020. 8", + "[46] Maxim Kolodiazhnyi, Anna Vorontsova, Anton Konushin, and Danila Rukhovich. Oneformer3d: One transformer for unified point cloud segmentation. In CVPR, 2024. 16", + "[47] Xin Lai, Jianhui Liu, Li Jiang, Liwei Wang, Hengshuang Zhao, Shu Liu, Xiaojuan Qi, and Jiaya Jia. Stratified transformer for 3d point cloud segmentation. In CVPR, 2022. 16, 17", + "[48] Loic Landrieu and Martin Simonovsky. Large-scale point cloud semantic segmentation with superpoint graphs. In CVPR, 2018. 17", + "[49] Huan Lei, Naveed Akhtar, and Ajmal Mian. Seggen: Efficient 3d point cloud segmentation with fuzzy spherical kernel. In CVPR, 2020. 16, 17", + "[50] Yangyan Li, Rui Bu, Mingchao Sun, Wei Wu, Xinhan Di, and Baoquan Chen. Pointconn: Convolution on x-transformed points. NeurIPS, 2018. 16, 17", + "[51] Dingkang Liang, Xin Zhou, Wei Xu, Xingkui Zhu, Zhikang Zou, Xiaqing Ye, Xiao Tan, and Xiang Bai. Pointmamba: A simple state space model for point cloud analysis. In Advances in Neural Information Processing Systems, 2024. 3", + "[52] Haojia Lin, Xiawu Zheng, Lijiang Li, Fei Chao, Shanshan Wang, Yan Wang, Yonghong Tian, and Rongrong Ji. Meta architecture for point cloud analysis. In CVPR, pages 17682-17691, 2023. 16, 17", + "[53] Minghua Liu, Ruoxi Shi, Kaiming Kuang, Yinhao Zhu, Xu-anlin Li, Shizhong Han, Hong Cai, Fatih Porikli, and Hao Su. Openshape: Scaling up 3d shape representation towards open-world understanding. In NeurIPS, 2023. 8, 13", + "[54] I Loshchilov. Decoupled weight decay regularization. In ICLR, 2019. 6", + "[55] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. In *ICLR*, 2017. 6", + "[56] Xu Ma, Can Qin, Haoxuan You, Haoxi Ran, and Yun Fu. Rethinking network design and local geometry in point cloud: A simple residual mlp framework. *ICLR*, 2022. 3", + "[57] Gaku Narita, Takashi Seno, Tomoya Ishikawa, and Yohsuke Kaji. Panopticfusion: Online volumetric semantic mapping at the level of stuff and things. In IROS, 2019. 16", + "[58] Richard A Newcombe, Shahram Izadi, Otmar Hilliges, David Molyneaux, David Kim, Andrew J Davison, Pushmeet Kohi, Jamie Shotton, Steve Hodges, and Andrew Fitzgibbon. Kinectfusion: Real-time dense surface mapping and tracking. In 2011 10th IEEE international symposium on mixed and augmented reality, pages 127-136. IEEE, 2011. 16", + "[59] Mehdi Noroozi and Paolo Favaro. Unsupervised learning of visual representations by solving jigsaw puzzles. In ECCV, 2016. 2", + "[60] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Russell Howes, Po-Yao Huang, Hu Xu, Vasu Sharma, Shang-Wen Li, Wojciech Galuba, Mike Rabbat, Mido Assan, Nicolas Ballas, Gabriel Synnaeve, Ishan Misra, Herve" + ], + "bbox": [ + 524, + 92, + 903, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "22202", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. Dinov2: Learning robust visual features without supervision. TMLR, 2024. 2, 3, 4, 6, 7, 13", + "[61] Yatian Pang, Wenxiao Wang, Francis EH Tay, Wei Liu, Yonghong Tian, and Li Yuan. Masked autoencoders for point cloud self-supervised learning. In ECCV, 2022. 2", + "[62] Chunghyun Park, Yoonwoo Jeong, Minsu Cho, and Jaesik Park. Fast point transformer. In CVPR, 2022. 16", + "[63] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. NeurIPS, 2019. 16", + "[64] Bohao Peng, Xiaoyang Wu, Li Jiang, Yukang Chen, Hengshuang Zhao, Zhuotao Tian, and Jiaya Jia. Oa-cnns: Omni-adaptive sparse cnns for 3d semantic segmentation. In CVPR, 2024. 3", + "[65] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In CVPR, 2017. 7, 17", + "[66] Charles R Qi, Li Yi, Hao Su, and Leonidas J Guibas. Point-net++: Deep hierarchical feature learning on point sets in a metric space. In NeurIPS, 2017. 3, 16, 17", + "[67] Guocheng Qian, Yuchen Li, Houwen Peng, Jinjie Mai, Hasan Hammoud, Mohamed Elhoseiny, and Bernard Ghanem. Pointnext: Revisiting pointnet++ with improved training and scaling strategies. In NeurIPS, 2022. 3, 16, 17", + "[68] Santhosh Kumar Ramakrishnan, Aaron Gokaslan, Erik Wijmans, Oleksandr Maksymets, Alexander Clegg, John M Turner, Eric Undersander, Wojciech Galuba, Andrew Westbury, Angel X Chang, Manolis Savva, Yili Zhao, and Dhruv Batra. Habitat-matterport 3d dataset (HM3d): 1000 large-scale 3d environments for embodied AI. In NeurIPS, 2021. 2, 6, 14, 15", + "[69] Damien Robert, Hugo Raguet, and Loic Landrieu. Efficient 3d semantic segmentation with superpoint transformer. In ICCV, 2023. 17", + "[70] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In MICCAI, 2015. 3", + "[71] David Rozenberszki, Or Litany, and Angela Dai. Language-grounded indoor 3d semantic segmentation in the wild. In ECCV, 2022. 7, 8", + "[72] Alexandre Sablayrolles, Matthijs Douze, Cordelia Schmid, and Hervé Jégou. Spreading vectors for similarity search. In ICLR, 2019. 4", + "[73] Julian Straub, Daniel DeTone, Tianwei Shen, Nan Yang, Chris Sweeney, and Richard Newcombe. Efm3d: A benchmark for measuring progress towards 3d egocentric foundation models. arXiv:2406.10224, 2024. 2, 14, 15, 16", + "[74] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, et al. Scalability in perception for autonomous driving: Waymo open dataset. In CVPR, 2020. 2, 8", + "[75] Maxim Tatarchenko, Jaesik Park, Vladlen Koltun, and Qian-Yi Zhou. Tangent convolutions for dense prediction in 3d. In CVPR, 2018. 17" + ], + "bbox": [ + 99, + 90, + 482, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[76] Lyne Tchapmi, Christopher Choy, Iro Armeni, JunYoung Gwak, and Silvio Savarese. Segcloud: Semantic segmentation of 3d point clouds. In 3DV, 2017. 17", + "[77] Hugues Thomas, Charles R Qi, Jean-Emmanuel Deschaud, Beatrix Marcotegui, François Goulette, and Leonidas J Guibas. Kpconv: Flexible and deformable convolution for point clouds. In ICCV, 2019. 3, 16, 17", + "[78] Hugues Thomas, Yao-Hung Hubert Tsai, Timothy D Barfoot, and Jian Zhang. Kpconvx: Modernizing kernel point convolution with kernel attention. In CVPR, 2024. 16", + "[79] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 16", + "[80] Chengyao Wang, Li Jiang, Xiaoyang Wu, Zhuotao Tian, Bohao Peng, Hengshuang Zhao, and Jiaya Jia. Groupcontrast: Semantic-aware self-supervised representation learning for 3d understanding. In CVPR, 2024. 2", + "[81] Lei Wang, Yuchun Huang, Yaolin Hou, Shenman Zhang, and Jie Shan. Graph attention convolution for point cloud semantic segmentation. In CVPR, 2019. 17", + "[82] Peng-Shuai Wang. Octformer: Octree-based transformers for 3D point clouds. SIGGRAPH, 2023. 16", + "[83] Shenlong Wang, Simon Suo, Wei-Chiu Ma, Andrei Pokrovsky, and Raquel Urtasun. Deep parametric continuous convolutional neural networks. In CVPR, 2018. 17", + "[84] Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. Dust3r: Geometric 3d vision made easy. In CVPR, 2024. 2, 13", + "[85] Wenxuan Wu, Zhongang Qi, and Li Fuxin. Pointconv: Deep convolutional networks on 3d point clouds. In CVPR, 2019. 16", + "[86] Wenxuan Wu, Li Fuxin, and Qi Shan. Pointconvformer: Revenge of the point-based convolution. In CVPR, 2023. 16", + "[87] Xiaoyang Wu, Yixing Lao, Li Jiang, Xihui Liu, and Hengshuang Zhao. Point transformer v2: Grouped vector attention and partition-based pooling. In NeurIPS, 2022. 7, 16, 17", + "[88] Xiaoyang Wu, Xin Wen, Xihui Liu, and Hengshuang Zhao. Masked scene contrast: A scalable framework for unsupervised 3d representation learning. In CVPR, 2023. 2, 3, 4, 5, 6, 7, 8, 14, 16, 17", + "[89] Xiaoyang Wu, Li Jiang, Peng-Shuai Wang, Zhijian Liu, Xihui Liu, Yu Qiao, Wanli Ouyang, Tong He, and Hengshuang Zhao. Point transformer v3: Simpler, faster, stronger. In CVPR, 2024. 3, 5, 7, 8, 16, 17", + "[90] Xiaoyang Wu, Zhuotao Tian, Xin Wen, Bohao Peng, Xihui Liu, Kaicheng Yu, and Hengshuang Zhao. Towards large-scale 3d representation learning with multi-dataset point prompt training. In CVPR, 2024. 3, 6, 7, 8, 16, 17", + "[91] Zhirong Wu, Yuanjun Xiong, Stella X. Yu, and Dahua Lin. Unsupervised feature learning via non-parametric instance discrimination. In CVPR, 2018. 2", + "[92] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. In NeurIPS, 2021. 3" + ], + "bbox": [ + 522, + 92, + 903, + 898 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "22203", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[93] Saining Xie, Jiatao Gu, Demi Guo, Charles R Qi, Leonidas Guibas, and Or Litany. Pointcontrast: Unsupervised pretraining for 3d point cloud understanding. In ECCV, 2020. 2, 3, 4, 6, 7, 8, 16, 17", + "[94] Zhenda Xie, Zheng Zhang, Yue Cao, Yutong Lin, Jianmin Bao, Zhuliang Yao, Qi Dai, and Han Hu. Simmim: A simple framework for masked image modeling. In CVPR, 2022. 2", + "[95] Mutian Xu, Runyu Ding, Hengshuang Zhao, and Xiaojuan Qi. Paconv: Position adaptive convolution with dynamic kernel assembling on point clouds. In CVPR, 2021. 17", + "[96] Xu Yan, Chaoda Zheng, Zhen Li, Sheng Wang, and Shuguang Cui. Pointasnl: Robust point clouds processing using nonlocal neural networks with adaptive sampling. In CVPR, 2020. 16", + "[97] Jiancheng Yang, Qiang Zhang, Bingbing Ni, Linguuo Li, Jinxian Liu, Mengdie Zhou, and Qi Tian. Modeling point clouds with self-attention and gumbel subset sampling. In CVPR, 2019. 17", + "[98] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Ji-ashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. In CVPR, 2024. 2", + "[99] Lihe Yang, Bingyi Kang, Zilong Huang, Zhen Zhao, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything v2. In NeurIPS, 2024. 2", + "[100] Yu-Qi Yang, Yu-Xiao Guo, Jian-Yu Xiong, Yang Liu, Hao Pan, Peng-Shuai Wang, Xin Tong, and Baining Guo. Swin3d: A pretrained transformer backbone for 3d indoor scene understanding. arXiv:2304.06906, 2023. 16, 17", + "[101] Chandan Yeshwanth, Yueh-Cheng Liu, Matthias Nießner, and Angela Dai. Scannet++: A high-fidelity dataset of 3d indoor scenes. In ICCV, 2023. 2, 6, 7, 8", + "[102] Xumin Yu, Lulu Tang, Yongming Rao, Tiejun Huang, Jie Zhou, and Jiwen Lu. Point-bert: Pre-training 3d point cloud transformers with masked point modeling. In CVPR, 2022. 2", + "[103] Feihu Zhang, Jin Fang, Benjamin Wah, and Philip Torr. Deep fusionnet for point cloud semantic segmentation. In ECCV, 2020. 16", + "[104] Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel M. Ni, and Heung-Yeung Shum. Dino: Detr with improved denoising anchor boxes for end-to-end object detection. arXiv:2203.03605, 2022. 2, 3, 4", + "[105] Michael Zhang, Aditi Raghunathan Wang, Shiori Sagawa, Tatsunori B Hashimoto, and Percy Liang. Fine-tuning can distort pretrained features and underperform out-of-distribution. In ICLR, 2022. 2", + "[106] Renrui Zhang, Ziyu Guo, Peng Gao, Rongyao Fang, Bin Zhao, Dong Wang, Yu Qiao, and Hongsheng Li. Pointm2ae: Multi-scale masked autoencoders for hierarchical point cloud pre-training. In NeurIPS, 2022. 2", + "[107] Hengshuang Zhao, Li Jiang, Chi-Wing Fu, and Jiaya Jia. Pointweb: Enhancing local neighborhood features for point cloud processing. In CVPR, 2019. 17", + "[108] Hengshuang Zhao, Li Jiang, Jiaya Jia, Philip Torr, and Vladlen Koltun. Point transformer. In ICCV, 2021. 16, 17" + ], + "bbox": [ + 93, + 92, + 480, + 898 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[109] Jia Zheng, Junfei Zhang, Jing Li, Rui Tang, Shenghua Gao, and Zihan Zhou. Structured3d: A large photo-realistic dataset for structured 3d modeling. In ECCV, 2020. 2, 5, 6", + "[110] Jinghao Zhou, Chen Wei, Huiyu Wang, Wei Shen, Cihang Xie, Alan Yuille, and Tao Kong. ibot: Image bert pretraining with online tokenizer. In ICLR, 2022. 2, 4", + "[111] Jinghao Zhou, Chen Wei, Huiyu Wang, Wei Shen, Cihang Xie, Alan Yuille, and Tao Kong. Image BERT pre-training with online tokenizer. In ICLR, 2022. 6", + "[112] Xingyi Zhou, Vladlen Koltun, and Philipp Krahenbuhl. Simple multi-dataset detection. In CVPR, 2022. 3" + ], + "bbox": [ + 516, + 92, + 903, + 262 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "22204", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2025/Sonata_ Self-Supervised Learning of Reliable Point Representations/3e06b74b-3b4f-4789-9e14-04119c218bb3_model.json b/2025/Sonata_ Self-Supervised Learning of Reliable Point Representations/3e06b74b-3b4f-4789-9e14-04119c218bb3_model.json new file mode 100644 index 0000000000000000000000000000000000000000..23f4e247938a79c89e8edc73f17f7e7c41ff5c10 --- /dev/null +++ b/2025/Sonata_ Self-Supervised Learning of Reliable Point Representations/3e06b74b-3b4f-4789-9e14-04119c218bb3_model.json @@ -0,0 +1,2831 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.02 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation." + }, + { + "type": "header", + "bbox": [ + 0.325, + 0.017, + 0.723, + 0.031 + ], + "angle": 0, + "content": "Except for this watermark, it is identical to the accepted version;" + }, + { + "type": "header", + "bbox": [ + 0.294, + 0.032, + 0.754, + 0.047 + ], + "angle": 0, + "content": "the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.157, + 0.07, + 0.842, + 0.093 + ], + "angle": 0, + "content": "Sonata: Self-Supervised Learning of Reliable Point Representations" + }, + { + "type": "text", + "bbox": [ + 0.115, + 0.119, + 0.882, + 0.158 + ], + "angle": 0, + "content": "Xiaoyang Wu\\(^{1,2}\\) Daniel DeTone\\(^{2*}\\) Duncan Frost\\(^{2*}\\) Tianwei Shen\\(^{2*}\\) Chris Xie\\(^{2*}\\) Nan Yang\\(^{2*}\\) Jakob Engel\\(^{2}\\) Richard Newcombe\\(^{2}\\) Hengshuang Zhao\\(^{1}\\) Julian Straub\\(^{2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.247, + 0.173, + 0.751, + 0.192 + ], + "angle": 0, + "content": "1The University of Hong Kong 2Meta Reality Labs Research" + }, + { + "type": "text", + "bbox": [ + 0.308, + 0.194, + 0.684, + 0.209 + ], + "angle": 0, + "content": "https://github.com/facebookresearch/sonata" + }, + { + "type": "image", + "bbox": [ + 0.1, + 0.217, + 0.379, + 0.422 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.384, + 0.217, + 0.617, + 0.421 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.62, + 0.217, + 0.9, + 0.421 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.1, + 0.424, + 0.495, + 0.604 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.499, + 0.424, + 0.899, + 0.604 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.612, + 0.907, + 0.696 + ], + "angle": 0, + "content": "Figure 1. Main properties. Sonata leads to reliable 3D self-supervised pretraining with the following superior and emerging properties: 1. Perception. Sonata advances state-of-the-art results across 3D indoor and outdoor perception tasks; 2. Linear probing. With less than \\(0.2\\%\\) learnable parameters, Sonata achieves strong and usable linear probing performance which is \\(3.3\\times\\) better than previous SOTA; 3. Decoder-free. Sonata moves beyond the inflexible U-Net structure, offering multi-scale representations that unchain future 3D research from previous architectural constraints. 4. Semantic awareness. Sonata reveals semantic structure in PCA and K-means visualizations. 5. Spatial reasoning. Sonata allows spatial correspondence even under strong augmentations as visualized via feature similarity." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.701, + 0.327, + 0.716 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.722, + 0.486, + 0.875 + ], + "angle": 0, + "content": "In this paper, we question whether we have a reliable self-supervised point cloud model that can be used for diverse 3D tasks via simple linear probing, even with limited data and minimal computation. We find that existing 3D self-supervised learning approaches fall short when evaluated on representation quality through linear probing. We hypothesize that this is due to what we term the \"geometric shortcut\", which causes representations to collapse to low-level spatial features. This challenge is unique to 3D and arises from the sparse nature of point cloud data." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.702, + 0.909, + 0.9 + ], + "angle": 0, + "content": "We address it through two key strategies: obscuring spatial information and enhancing the reliance on input features, ultimately composing a Sonata of 140k point clouds through self-distillation. Sonata is simple and intuitive, yet its learned representations are strong and reliable: zero-shot visualizations demonstrate semantic grouping, alongside strong spatial reasoning through nearest-neighbor relationships. Sonata demonstrates exceptional parameter and data efficiency, tripling linear probing accuracy (from \\(21.8\\%\\) to \\(72.5\\%\\)) on ScanNet and nearly doubling performance with only \\(1\\%\\) of the data compared to previous approaches. Full fine-tuning further advances SOTA across both 3D indoor and outdoor perception tasks." + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.888, + 0.321, + 0.901 + ], + "angle": 0, + "content": "*Equal contribution in alphabetic order." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.958 + ], + "angle": 0, + "content": "22193" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.09, + 0.223, + 0.105 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.112, + 0.484, + 0.279 + ], + "angle": 0, + "content": "Self-supervised learning (SSL) with images [2, 13, 34, 94, 104] has seen a continuous increase in model simplicity, capacity, and capability over the past decade [25]. Tuning a single linear layer can achieve performance close to full fine-tuning [31, 35, 104], fostering growing trust in its reliability. This trust has been further strengthened by witnessing the semantic meaning of learned image representations through direct visualization [60, 110]. Consequently, these reliable self-supervised models have become the foundation for emerging approaches [14, 84, 98, 99] across various fields involving images." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.28, + 0.484, + 0.49 + ], + "angle": 0, + "content": "In contrast to the image domain, self-supervised learning with point clouds [18, 61, 93, 102, 106] is still in its early stages. Despite the broad reliance on 3D applications in autonomous driving [9, 74], robotic learning [32], mixed reality [5, 43] and egocentric perception [27, 73], the latest self-supervised point cloud models are seldom included in their pipelines. This gap prompts us to consider a simple yet critical question: do we have a reliable point self-supervised learning approach that provides strong representations, usable with simple linear probing across these applications? Not yet. Previous SOTAs [38, 88, 93] fall short on this higher-level criterion, achieving only a maximum of \\(21.8\\%\\) mIoU on ScanNet [23] with linear probing, especially given that the performance from scratch is \\(77.6\\%\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.492, + 0.484, + 0.643 + ], + "angle": 0, + "content": "We identify the geometric shortcut as the primary issue hindering these prior point cloud SSL approaches from learning reliable representations. This shortcut refers to the tendency of the model to collapse to easily accessible, low-level geometric cues, such as normal direction or point height, as demonstrated in prior works and visualized in Fig. 2. This spatial information is inevitably introduced into point cloud operators along with point coordinates rather than through input features, making it difficult to obscure and nearly impossible to mask effectively." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.644, + 0.483, + 0.81 + ], + "angle": 0, + "content": "However, the model collapse caused by geometric shortcuts can be mitigated through two key strategies: obscuring spatial information and emphasizing input features. Specifically, we address this issue by applying SSL losses at coarser spatial scales, disturbing the spatial information of masked points without features, and progressively increasing task difficulty to reduce reliance on accessible geometric cues. Coupled with a point self-distillation framework and scaling techniques inspired by recent advances in image SSL [11, 12, 34, 60, 110], we ultimately compose a Sonata using 140k point cloud scenes [1, 3, 5, 23, 68, 101, 109]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Sonata demonstrates strong zero-shot capabilities, with PCA-colored visualizations of point clouds, k-means clustering of features, and nearest-neighbor matching between point clouds (see Fig. 1). Sonata also proves highly data-efficient, raising semantic segmentation performance from \\(25.8\\%\\) to \\(45.3\\%\\) under extremely limited data conditions" + }, + { + "type": "image", + "bbox": [ + 0.514, + 0.089, + 0.905, + 0.308 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.309, + 0.905, + 0.392 + ], + "angle": 0, + "content": "Figure 2. Geometric shortcut. We select a point on the sofa arm and compute pairwise similarity with other points. The similarity heatmap reveals that CSC [38] collapses to surface normals, and MSC [88] overfits to point height. In contrast, our Sonata can extract higher-level concepts, as can be seen by the high similarity between all sofa arms highlighted in red." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.398, + 0.907, + 0.549 + ], + "angle": 0, + "content": "(1% of ScanNet). Additionally, Sonata significantly boosts linear probing accuracy on ScanNet semantic segmentation, increasing it by over \\(3.3 \\times\\) from \\(21.8\\%\\) to \\(72.5\\%\\) and surpassing the accuracy of DINOv2 features aggregated onto the point cloud (63.1%). Moreover, combining Sonata features with DINOv2 features further enhances accuracy (76.4%), underscoring that Sonata captures unique 3D information beyond what is visible in images alone. Finally, Sonata achieves state-of-the-art results across various indoor and outdoor perception tasks with full fine-tuning." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.561, + 0.655, + 0.576 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.583, + 0.906, + 0.765 + ], + "angle": 0, + "content": "Image self-supervised learning. Over the past decade, remarkable advancements [25, 26, 36, 59, 91] have been made in image self-supervised learning, and our research is largely inspired by two pivotal moments in this field. First, linear probing, a method that assesses representation quality by optimizing a minimal linear transformation, has become a standard in 2D image SSL [10, 13, 34, 91]. In some cases, such as when the distribution shift is large, linear probing surpasses full fine-tuning [105]. Second, the ability to directly perceive the semantic meaning of learned representations through zero-shot visualization like PCA or attention [12, 60] has further strengthened trust in reliability." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.766, + 0.906, + 0.902 + ], + "angle": 0, + "content": "Point self-supervised learning. Sonata follows the research path initiated by PointContrast [38, 93], emphasizing self-supervised learning with scene-level data [23]. While previous efforts do implement strategies to prevent collapse from geometric shortcuts, they remain limited. For example, Masked Scene Contrast (MSC) [88] encourages learning beyond naive geometric cues by predicting color or normal vectors but still partially anchors representations to predefined tasks. GroupContrast (GC) [80] employs graph-" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22194" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.082, + 0.482, + 0.252 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.254, + 0.483, + 0.351 + ], + "angle": 0, + "content": "Figure 3. The geometric shortcut is unique to 3D. When comparing the information contained in 2D image and 3D point cloud data after removing the input feature (indicated by color), it is evident that in images all information is within the input feature. Whereas point clouds retain geometric information in point positions, which is directly utilized by operators. This characteristic leads to what we term geometric shortcuts in 3D SSL." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.356, + 0.484, + 0.431 + ], + "angle": 0, + "content": "based segment guidance [28], though it is constrained by its reliance on human-designed algorithms. Building on MSC, Sonata directly addresses the geometric shortcut and scales up training [90] to establish a more reliable approach to point cloud self-supervised learning." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.431, + 0.484, + 0.643 + ], + "angle": 0, + "content": "Point cloud backbones. This area [17, 51, 56, 64, 67, 77] has significantly benefited from the U-Net structure [70], particularly with its hierarchical decoding and skip connections, as first introduced by PointNet++ [66]. However, the tight coupling between the encoder and decoder restricts flexibility and generalization capacity [92]. Sonata addresses this limitation by focusing self-supervised learning exclusively on the encoder, thus removing the hierarchical decoder. Moreover, unlike previous SOTAs [38, 88, 93] in point SSL, which primarily use SparseUNet [17, 20, 29] to balance efficiency and accuracy, we leverage Point Transformer V3 (PTv3) [89], an efficient, accurate, and scalable transformer backbone. This shift alone yields a \\(7.7\\%\\) improvement in linear probing performance over MSC [88]." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.653, + 0.388, + 0.671 + ], + "angle": 0, + "content": "3. Pilot Study and Design Principle" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.675, + 0.483, + 0.719 + ], + "angle": 0, + "content": "In this section, we qualitatively study the problems of current point self-supervised approaches and point cloud backbones to inform the Sonata approach." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.483, + 0.84 + ], + "angle": 0, + "content": "Uncovering the geometric shortcut in point SSL. The history of image self-supervised learning [34, 60, 104, 112] can be summarized as a continuous battle against shortcuts, where models often exploit trivial solutions (mode collapse) rather than understanding deeper semantics. Each advancement has involved identifying, understanding, and overcoming these shortcuts, refining pretext tasks to push models to \"struggle\" in learning stronger representations." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "However in 3D point SSL, despite various attempts [38, 88, 93] to increase the difficulty of pretext tasks, a curious phenomenon persists—the loss consistently reduces rapidly to an ideal range in the early stages of training. We hypo-" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.082, + 0.905, + 0.167 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.169, + 0.907, + 0.294 + ], + "angle": 0, + "content": "Figure 4. What is learned by the hierarchical backbone? We visualize PCA embeddings from different stages of a hierarchical encoder and decoder, trained for semantic segmentation. The encoder captures diverse and dispersed feature patterns, indicating a broad range of information. Notably, as the point cloud becomes coarser, accessible geometric information within point coordinates becomes increasingly global. In contrast, the decoder's representations are more uniform and structured, suggesting a focus on refining features for task-specific outputs." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.303, + 0.906, + 0.409 + ], + "angle": 0, + "content": "esize that this lack of \"struggle\" in learning indicates that a shortcut exists that leads to a collapse of the representation to trivial solutions. Indeed as Fig. 2 illustrates, previous 3D SSL approaches seem to learn representations that are sensitive to local surface normals or point height. The collapse to such naive solutions even for well-designed SSL pretext tasks is what we term the geometric shortcut." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.411, + 0.906, + 0.591 + ], + "angle": 0, + "content": "Intuitively, representations affected by this geometric shortcut have not learned sufficient semantics. To quantify this problem, we leverage the standard criteria from 2D SSL of linear probing: we linearly probe the semantic class of each point from the learned representation. As illustrated in Fig. 1, with linear probing, previous 3D SSL methods PC [93] and MSC [88] indeed achieve only \\(5.6\\%\\) mIoU and \\(21.8\\%\\) mIoU respectively on ScanNet semantic segmentation. When compared to the \\(63.1\\%\\) mIoU achieved by 3D-aggregated image representations from DINOv2 [60], it becomes obvious that the current 3D SSL methods do not learn semantic information." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.595, + 0.906, + 0.746 + ], + "angle": 0, + "content": "We hypothesize that the geometric shortcut stems from the sparse nature of point cloud data. Every operator, whether for point clouds or images, inherently relies on point (pixel) coordinates to define the kernel. However, unlike images with regularly spaced dense pixels, the sparsity of point cloud data necessitates introducing point coordinate information into point cloud operators rather than just through input features (see Fig. 3). This fact makes it difficult to obscure and nearly impossible to effectively mask out, ultimately resulting in the geometric shortcut." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.748, + 0.906, + 0.793 + ], + "angle": 0, + "content": "These observations motivate the training methodology of Sonata aimed to prevent the dominance of naive spatial information inherent in point cloud organization." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Focusing self-supervised learning on encoder only. Following our hypothesis that the geometric arrangement leaks information into the learned representation via the geometric shortcut, we examine the point cloud model. The typical U-Net structure [66, 70] is effective at handling large point clouds in a coarse-to-fine way, but the tight coupling of the encoder and decoder via skip connections restricts" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.958 + ], + "angle": 0, + "content": "22195" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.212 + ], + "angle": 0, + "content": "flexibility. Specifically, the decoder enforces per-point features at the original high-resolution scale, with shallow feature channels. This constraint limits the capacity to provide richer representations, which is crucial for SSL as evidenced by large channel dimensions in state-of-the-art 2D approaches [12]. Most importantly, decoding point clouds at the original scale unavoidably introduces local geometric cues into operators facilitating the geometric shortcut." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.213, + 0.483, + 0.363 + ], + "angle": 0, + "content": "In Fig. 4, we visualize PCA embeddings from different stages of PTv3 encoder and decoder, trained in a supervised manner for semantic segmentation. We observe that the encoder learns diverse features capturing spatial features at different scales in contrast to the decoder which produces task-specific higher-level representations. Additionally, as the spatial resolution of the point cloud decreases via the max-pooling stages, feature representations become less local. This resolution reduction fundamentally limits reliance on fine spatial information tied to point coordinates." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.364, + 0.483, + 0.471 + ], + "angle": 0, + "content": "These observations motivate us to remove the decoder during self-supervised learning for two main reasons. First and foremost, training directly with features at coarser point resolutions inherently restricts access to fine-grained spatial information, reducing the possibility of geometric shortcuts. Second, task-specific features can be probed or fine-tuned on top of a more expressive multi-scale point representation." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.481, + 0.396, + 0.496 + ], + "angle": 0, + "content": "4. Point Self-distillation with Sonata" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.503, + 0.483, + 0.579 + ], + "angle": 0, + "content": "This section details the methodology of Sonata, the point self-distillation framework designed to address geometric shortcuts, remove structural constraints, and deliver strong linear probing results as discussed in Sec. 3. A roadmap of incremental ablation is illustrated in Fig. 6." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.583, + 0.274, + 0.596 + ], + "angle": 0, + "content": "4.1. Macro Framework" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.598, + 0.483, + 0.703 + ], + "angle": 0, + "content": "Before diving into the specific micro designs, we begin with a point self-distillation framework derived from insights gained through previous efforts in point [88, 93] and image [11, 34, 60, 104, 110] self-supervised learning. This macro framework provides a solid foundation for pretext task design, allowing us to focus on addressing the geometric shortcut without additional concerns." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.483, + 0.81 + ], + "angle": 0, + "content": "In essence, (point) self-supervised learning aims to make things (points) that should be the same, the same (identical in representation). This forms the basic recipe of point SSL: generating two views of a given point cloud with random spatial (e.g., crop, rotate, distort) and photometric (e.g., jitter) augmentations, then matching and aligning the feature embeddings of points that are close in the original space." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.483, + 0.902 + ], + "angle": 0, + "content": "However, the superior robustness of self-supervised representation is rooted in a core principle: continuously increasing the difficulty of pretext tasks as long as the model continues to converge. This principle encourages enhancing the basic recipe with local-global view alignment and mask-unmask view alignment. Specifically, it involves aligning" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.083, + 0.9, + 0.269 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.271, + 0.905, + 0.41 + ], + "angle": 0, + "content": "Figure 5. Self-distillation framework of Sonata. (1) Local views (bottom left) and global views (right) are generated with dedicated spatial and photometric augmentations, while masked views are created by randomly masking out grid-based patches from the global views (top left). (2) Embeddings from local and masked views are extracted by the student, with global views processed by the teacher (top). (3) Points from local and masked views are matched with corresponding points in the global views based on their original spatial distance, allowing for the distillation of embeddings from global views to local and masked views (bottom)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.416, + 0.905, + 0.507 + ], + "angle": 0, + "content": "neighboring points from local views generated by small cropping ratios and masked views created with large masking ratios (see Fig. 5 left), with unmasked global views that contain relatively richer information (see Fig. 5 right). The difficulty of pretext tasks can be scaled by adjusting the cropping and masking ratios." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.508, + 0.906, + 0.703 + ], + "angle": 0, + "content": "The more challenging the pretext tasks become, the higher the risk of model collapse. This instability calls for an asymmetric encoding approach, specifically exponential moving average (EMA) [34]: rather than encoding all views with a shared-weight model, the approach encodes challenging local and masked views with an actively learned student model, while using a stable teacher model, updated with a moving average of the student parameters, to encode global views (see Fig. 5 top). With the teacher preventing the student from being misled, the student is less likely to get lost in \"mission impossible\" and has a greater chance to discover treasure within \"impossible\" (i.e. extreme \\(5\\%\\) crop ratio for local views and \\(70\\%\\) mask ratio for masked views)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.906, + 0.901 + ], + "angle": 0, + "content": "In terms of SSL criteria, we move away from prior approaches that rely on contrastive and generative learning [88]. Contrastive learning, limited by the number of point pairs in pairwise similarity computations, restricts scalability. Generative learning, meanwhile, partially anchors representations to predefined cues, limiting the model's capacity to capture more generalizable features. Following DINov2, we adopt a self-distillation approach driven by Sinkhorn-Knopp centering [11], KoLeo regularization [72], and clustering assignments [11]. This adaptation initially exacerbates collapse into geometric shortcuts but holds greater potential for robust representation once this challenge is addressed." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22196" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.232, + 0.108 + ], + "angle": 0, + "content": "4.2. Micro Design" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.109, + 0.484, + 0.2 + ], + "angle": 0, + "content": "We now discuss micro designs aimed at addressing geometric shortcuts. Since the problematic spatial information is inherently tied to point coordinates and directly used by operators, it is nearly impossible to mask out. This constraint defines the key strategies of our micro designs: obscuring spatial information and emphasizing input features." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.201, + 0.484, + 0.427 + ], + "angle": 0, + "content": "Decoder removal. Previous approaches adhere to the original U-Net-style backbone for feature extraction; however, our observations in Sec. 3 motivate us to remove the complex hierarchical decoder and perform self-distillation directly using the encoder's output. This simple adjustment is key in our fight against geometric shortcuts, as it: (a) increases the feature channels participating in self-distillation (from 96 to 512), (b) streamlines the pipeline by involving fewer points in the pretext task after hierarchical pooling, and, most importantly, (c) introduces a natural way to obscure geometric cues: the positional information of points becomes naturally disturbed during hierarchical encoding and pooling. In fact, this removal proves to be even more beneficial than expected, boosting the linear probing result from \\(20.7\\%\\) to \\(60.4\\%\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.43, + 0.484, + 0.717 + ], + "angle": 0, + "content": "Feature up-casting. While the removal effectively prevents over-reliance on naive geometric cues, it also introduces certain limitations, particularly in leveraging multiscale contexts. In the original U-Net structure, the decoder plays a key role in progressively aggregating features across scales to reconstruct semantic details. Without this process, self-distillation struggles to capture multi-scale spatial information and broader contextual relationships. To retain multi-scale features, we introduce a parameter-free feature up-casting process similar to hypercolumns in image segmentation [33]: progressively up-casting features back to the scale of the previous encoding stage, with the mapping relationships preserved through pooling layers and concatenation with features from the prior encoding stage. This approach provides richer, coarse-to-fine features from multiscale encoding. While it does increase the risk of the model falling into geometric shortcuts, finding the right balance is key. Our ablation study shows that up-casting features twice achieve the best performance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.718, + 0.484, + 0.854 + ], + "angle": 0, + "content": "Masked points jitter. A slight random Gaussian jitter \\((\\sigma = 0.005)\\) is applied to point coordinates as part of data augmentation. However, specifically for points to be masked, we additionally apply a stronger Gaussian jitter \\((\\sigma = 0.01)\\) to further disrupt their spatial relationships. We pay special attention to these masked points because models are more likely to collapse to naive geometric cues, especially when point features are masked, making it difficult to derive solutions from neighboring unmasked points." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Progressive parameter scheduler. Geometric shortcuts are like traps on the convergence path of point cloud self-supervised learning. Similarly, we can set our own \"trap\"" + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.085, + 0.905, + 0.452 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.454, + 0.907, + 0.567 + ], + "angle": 0, + "content": "Figure 6. The roadmap. We evolve Mask Scene Contrast [88] into our Sonata by modernizing self-supervised learning with self-distillation, addressing the geometric shortcut, and scaling up training. Our designs are validated through progressive ablation with linear and decoder probing on ScanNet semantic segmentation [23]. Starting with \\(23\\mathrm{k}\\) training data (a combination of ScanNet and Structured3D [109]) and a \\(39\\mathrm{M}\\) PTv3 model [89], we ultimately scale up to \\(140\\mathrm{k}\\) assets (Tab. 1) and a \\(108\\mathrm{M}\\) PTv3 model." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.579, + 0.907, + 0.821 + ], + "angle": 0, + "content": "during training. For example, rather than starting with a challenging large mask size and mask ratio, we begin with a relatively small mask size (10 cm) and mask ratio (30%), gradually increasing them to 40 cm (0.1 \\(\\Rightarrow\\) 0.4) and 70% (0.3 \\(\\Rightarrow\\) 0.7) over the first 5% of the training process. This strategy encourages the model to rely more on input features to solve the pre-defined pretext tasks, preventing it from shifting its reliance to point coordinates as training difficulty increases. This approach aligns with curriculum learning [7], progressively challenging the model as it adapts. Similarly, in addition to the common learning rate scheduler, we also implement custom progressive schedulers for teacher temperature (0.04 \\(\\Rightarrow\\) 0.07) and weight decay (0.04 \\(\\Rightarrow\\) 0.2). We found that this design pushes these parameters to more extreme levels previously unexplored, further enhancing model performance." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.824, + 0.871, + 0.839 + ], + "angle": 0, + "content": "4.3. Implementation and Evaluation Protocols" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.84, + 0.905, + 0.871 + ], + "angle": 0, + "content": "In this section, we introduce the implementation details and the evaluation protocols for our experiments." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.905, + 0.902 + ], + "angle": 0, + "content": "Backbone. We build our Sonata with Point Transformer V3 (PTv3) [89] and refer to Pointcept [19] for details of imple" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.958 + ], + "angle": 0, + "content": "22197" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.094, + 0.089, + 0.482, + 0.228 + ], + "angle": 0, + "content": "
DatasetSourceTrainValTestTotal
ScanNet [23]real1,2013121001,613
ScanNet++ [101]real7121781261,016
S3DIS [1]real204680272
ArkitScenes [5]real4,49854905,047
HM3D [68]real8,8811,119010,000
Structured3D [109]sim.18,3481,7761,69721,821
ASE [3]sim.90,00010,0000100,000
Sonata (ours)mixed123,84414,0021,923139,769
" + }, + { + "type": "table_caption", + "bbox": [ + 0.189, + 0.23, + 0.386, + 0.242 + ], + "angle": 0, + "content": "Table 1. Data source collection." + }, + { + "type": "table", + "bbox": [ + 0.097, + 0.247, + 0.473, + 0.328 + ], + "angle": 0, + "content": "
MethodRealSimTotalMultiplier
PC [93]1,61301,613×1
MSC [88]6,66006,660×4.1
PPT [90]1,88521,82123,706×14.7
Sonata (ours)17,948121,821139,768×86.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.188, + 0.33, + 0.387, + 0.343 + ], + "angle": 0, + "content": "Table 2. Data scale comparison." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.355, + 0.483, + 0.521 + ], + "angle": 0, + "content": "mentation. Building on this, we made an additional adjustment to enhance scalability: replacing all Batch Normalization (BN) [41] layers with Layer Normalization (LN) [4]. Although this replacement results in some initial accuracy degradation, it enhances domain adaptation by eliminating the need for additional domain-specific adjustments when scaling up with multi-dataset joint training [90]. Along with the scaling up of data, we also scale up the encoder block depths from \\([2, 2, 2, 6, 2]\\) to \\([3, 3, 3, 12, 3]\\) and widths from \\([32, 64, 128, 256, 512]\\) to \\([48, 96, 192, 384, 512]\\). This PTv3 model has \\(108\\mathrm{M}\\) parameters." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.522, + 0.483, + 0.643 + ], + "angle": 0, + "content": "Data. We extend the multi-dataset joint training approach introduced by PPT [90], further expanding the data scale by removing the constraint of human labeling through unsupervised learning. This results in a collection of 140k scene-level point clouds from both real-world and simulated environments (outlined in Tab. 1), making it \\(86.7 \\times\\) larger than the data scale of PointContrast [93] and \\(5.9 \\times\\) larger than the data collection of PPT, as detailed in Tab. 2." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.644, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Training. We train Sonata on the 140k data collection for 200 epochs, using the AdamW optimizer [54] with a batch size of 96, distributed across 32 GPUs. The learning rate linearly warms up over the first 10 epochs to a base value of 0.004, then decays following a cosine schedule [55]. Additionally, a layer-wise learning rate decay of 0.9 is applied to model parameters [111]. Weight decay is also controlled by a cosine schedule, progressively increasing from 0.04 to 0.2. For EMA, the student temperature is set to 0.1, while the teacher temperature gradually rises from 0.04 to 0.07 during the first 10 epochs [60]. The momentum starts at 0.994 and increases to 1 by the final iteration. For data augmentation and view generation, we follow the augmentation pipeline designed by MSC [88]. We generate 2 global views (sampling \\(40\\%\\) to \\(100\\%\\) of scene points) and 4 local views (sampling \\(5\\%\\) to \\(40\\%\\) of scene points) for training, with 2 masked views generated based on the global views." + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.089, + 0.613, + 0.161 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.616, + 0.089, + 0.905, + 0.162 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.162, + 0.613, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.616, + 0.163, + 0.904, + 0.235 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.236, + 0.613, + 0.308 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.616, + 0.236, + 0.904, + 0.308 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.313, + 0.905, + 0.397 + ], + "angle": 0, + "content": "Figure 7. Zero-shot comparison with DINOv2. We compare the PCA visualizations of DINOv2, Sonata, and their combined feature representation. DINOv2 excels at capturing photometric details, while Sonata better distinguishes spatial information. The combined model demonstrates improved coherence and detail, showcasing the complementary strengths of both models." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.409, + 0.905, + 0.453 + ], + "angle": 0, + "content": "Evaluation. We evaluate the quality of the learned representation using the following three protocols after initializing the encoder with Sonata:" + }, + { + "type": "text", + "bbox": [ + 0.522, + 0.454, + 0.905, + 0.514 + ], + "angle": 0, + "content": "- In linear probing, we keep the encoder frozen and upcast the features to their original scale. A single linear layer, comprising less than \\(0.2\\%\\) of the total parameters, is then used to adapt these features to downstream tasks." + }, + { + "type": "text", + "bbox": [ + 0.522, + 0.515, + 0.905, + 0.575 + ], + "angle": 0, + "content": "- In decoder probing, we take a step back and reintroduce a lightweight hierarchical decoder, which accounts for \\(13\\%\\) of the total parameters. We then freeze the encoder, allowing only the decoder to actively learn." + }, + { + "type": "text", + "bbox": [ + 0.522, + 0.575, + 0.905, + 0.62 + ], + "angle": 0, + "content": "- In full fine-tuning, we follow the traditional approach by unfreezing the entire PTv3 U-Net-style backbone, tuning the learned representation to downstream tasks." + }, + { + "type": "list", + "bbox": [ + 0.522, + 0.454, + 0.905, + 0.62 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.621, + 0.906, + 0.68 + ], + "angle": 0, + "content": "We advocate linear probing as the primary evaluation criterion for point SSL, considering other methods as intermediate steps. We look forward to the day when a linear-probed self-supervised model outperforms a fully fine-tuned one." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.687, + 0.65, + 0.702 + ], + "angle": 0, + "content": "5. Main Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.906, + 0.749 + ], + "angle": 0, + "content": "We validate the reliability of the Sonata representation using the evaluation protocols discussed in Sec. 4.3 and analyze the main properties based on these results." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Comparison with image self-supervised model. In Tab. 3, we compare the linear probing and decoder probing results on ScanNet and ScanNet200 semantic segmentation with the linear probing accuracy of representations transferred from the image self-supervised models DINOv2 [60] and DINOv2.5 [24]. Specifically, we aggregate unprojected pixel embeddings using ground truth camera poses and depth calculated by ray intersection with a reconstructed mesh, which provides more accuracy than sensor per-frame depth. Our results indicate that while the DINOs' represent" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22198" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.093, + 0.089, + 0.458, + 0.246 + ], + "angle": 0, + "content": "
2D × 3DScanNet Val [23]ScanNet200 Val [23]
MethodsmIoUmAccallAccmIoUmAccallAcc
• DINOv2 (lin.) [60]63.0975.5082.4227.4237.5972.80
• DINOv2.5 (lin.) [24]63.3675.9482.3027.7539.2372.53
• Sonata (lin.)72.5283.1189.7429.2541.6181.15
+DINOv2 (lin.)75.9185.3691.2536.6746.9882.85
+DINOv2.5 (lin.)76.4485.6891.3336.9648.2382.77
• Sonata (dec.)79.0786.5792.6833.5444.4884.07
+DINOv2 (dec.)79.1287.2392.4737.7349.3883.31
+DINOv2.5 (dec.)79.1986.6692.5038.2748.5783.77
" + }, + { + "type": "table_caption", + "bbox": [ + 0.115, + 0.249, + 0.429, + 0.262 + ], + "angle": 0, + "content": "Table 3. Numerical comparison with DINO series." + }, + { + "type": "table", + "bbox": [ + 0.467, + 0.091, + 0.904, + 0.246 + ], + "angle": 0, + "content": "
Data EfficiencyLimited Scenes (Pct.)Limited Annotation (Pts.)
Methods1%5%10%20%Full2050100200Full
○ SparseUNet [17]26.047.856.762.972.241.953.962.265.572.2
• CSC [38]28.949.859.464.673.855.560.565.968.273.8
• MSC [88]29.250.761.064.975.461.065.668.969.675.4
○ PTv2 [87]24.848.159.866.375.458.466.170.371.275.4
○ PTv3 [89]25.848.961.067.077.260.167.971.472.777.2
• PPT [90] (sup.)31.152.663.368.278.262.469.174.375.578.2
• Sonata (lin.)43.662.568.669.872.569.070.571.171.572.5
• Sonata (dec.)44.564.169.872.579.169.873.175.076.379.1
• Sonata (full)45.365.772.472.879.470.573.676.077.079.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.609, + 0.249, + 0.76, + 0.262 + ], + "angle": 0, + "content": "Table 4. Data efficiency." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.267, + 0.9, + 0.424 + ], + "angle": 0, + "content": "
Param. Efficiency MethodsParamsScanNet Val [23]ScanNet200 Val [71]ScanNet++ Val [101]S3DIS Area 5 [1]S3DIS 6-fold [1]
Learn.Pct.mIoUmAccallAccmIoUmAccallAccmIoUmAccallAccmIoUmAccallAccmIoUmAccallAcc
○ SparseUNet [17]39.2M100%72.380.290.025.032.980.428.838.480.166.372.589.872.480.989.9
• PC [93] (lin.)<0.2M<0.1%5.69.750.00.50.940.31.83.146.411.418.652.311.719.051.2
• CSC [38] (lin.)<0.2M<0.1%12.618.164.21.32.153.02.84.553.624.432.066.424.932.566.9
• MSC [88] (lin.)<0.2M<0.1%14.120.362.91.52.553.64.56.661.327.935.571.129.937.971.3
○PTv3 [89]124.8M100%77.685.092.035.346.083.442.153.485.673.478.991.777.785.391.5
• MSC [88] (lin.)<0.2M<0.2%21.832.265.53.35.557.58.111.964.732.142.470.934.646.071.3
• Sonata (lin.)<0.2M<0.2%72.583.189.729.341.681.237.350.984.372.381.290.976.587.490.8
• Sonata (dec.)16.3M13%79.186.692.733.544.584.140.952.686.374.580.492.681.588.893.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.397, + 0.426, + 0.584, + 0.44 + ], + "angle": 0, + "content": "Table 5. Parameter efficiency." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.449, + 0.482, + 0.586 + ], + "angle": 0, + "content": "tation demonstrates impressive robustness, Sonata offers a more suitable representation for 3D tasks, achieving \\(+9.2\\%\\) on ScanNet and \\(+1.5\\%\\) on ScanNet200 for semantic segmentation. Furthermore, combining Sonata with the DINOs yields higher accuracy than any single data modality alone \\((+3.9\\%\\) and \\(+7.7\\%\\), respectively), underscoring the promising potential of cross-modal self-distillation. Additional zero-shot comparisons (see Fig. 7) through PCA visualizations further corroborate these numerical findings." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.592, + 0.483, + 0.744 + ], + "angle": 0, + "content": "Data efficiency. In Tab. 4, we present the semantic segmentation performance of Sonata when probed or fine-tuned on the ScanNet dataset with limited scenes and annotations [38]. The results demonstrate the exceptional data efficiency of Sonata, with improvements of \\(19.5\\%\\) in extreme data scarcity (1% of scenes) and \\(10.4\\%\\) in limited annotation scenarios (20 points per scene) compared to training from scratch. Notably, even linear probing surpasses previous SOTA by a substantial margin (12.5% with 1% scenes), highlighting Sonata's reliability in low-data scenarios." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.75, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Parameter efficiency. In Tab. 5, we demonstrate parameter efficiency using both linear and decoder probing across various indoor semantic segmentation benchmarks, including ScanNet [23], ScanNet200 [71], ScanNet++ [101], S3DIS Area5[1], and S3DIS 6-fold cross-validation [65]. Semantic segmentation is emphasized as it provides a direct measure of point cloud representation quality. Results show that a single linear layer with a negligible number of parameters (<0.02% of total parameters) is sufficient for Sonata to achieve strong performance on these benchmarks (e.g.," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.449, + 0.905, + 0.585 + ], + "angle": 0, + "content": "72.5% on ScanNet and 73.4% on S3DIS Area5). Furthermore, probing a decoder with only 13% of the model's parameters yields even higher accuracy (e.g., 79.1% on ScanNet and 81.5% on S3DIS 6-fold cross-validation). However, while decoder probing achieves SOTA results on ScanNet (20 classes) and S3DIS (13 classes), performance on ScanNet200 (200 classes) and ScanNet++ (100 classes) remains limited. This shows a limitation of the learned representation in distinguishing a large number of classes." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.592, + 0.906, + 0.743 + ], + "angle": 0, + "content": "Indoor semantic segmentation. In Tab. 6, we further enhance Sonata's semantic segmentation accuracy through full fine-tuning, consistently pushing SOTA results to new heights across the five widely recognized benchmarks, e.g., achieving \\(79.4\\%\\) on ScanNet and \\(82.3\\%\\) on S3DIS 6-fold cross-validation. However, we view full fine-tuning as an intermediate step toward a future where linear probing surpasses it. At present, full fine-tuning remains essential to achieve the highest performance on these benchmarks and close the remaining gap of \\(7.1\\%\\) and \\(5.8\\%\\) respectively." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Indoor instance segmentation. In Tab. 7, We also validate the robustness of Sonata representation on indoor instance segmentation benchmarks, including ScanNet [23], ScanNet200 [71], ScanNet++ [101], and S3DIS [1]. Consistent with our findings in semantic segmentation, Sonata demonstrates strong parameter efficiency, achieving significant improvements with linear probing (10× mAP50 on ScanNet and 21× on ScanNet200) and decoder probing (12× mAP50 on ScanNet and 33× on ScanNet200). Full fine-tuning further boosts these results, achieving SOTA" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22199" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.093, + 0.089, + 0.902, + 0.246 + ], + "angle": 0, + "content": "
Indoor Sem. SegParamsScanNet Val [23]ScanNet200 Val [71]ScanNet++ Val [101]S3DIS Area 5 [1]S3DIS 6-fold [1]
MethodsLearn.Pct.mIoUmAccallAccmIoUmAccallAccmIoUmAccallAccmIoUmAccallAccmIoUmAccallAcc
○ SparseUNet [17]39.2M100%72.380.290.025.032.980.428.838.480.166.372.589.872.480.989.9
• PC [93]39.2M100%72.380.990.126.233.079.929.239.782.768.173.590.074.783.390.6
• CSC [38]39.2M100%72.881.090.726.933.780.632.541.183.770.776.490.875.584.090.9
• MSC [88]39.2M100%75.783.491.332.041.682.339.449.684.970.776.191.077.485.391.5
○PTv3 [89]124.8M100%77.685.092.035.346.083.442.153.485.673.478.991.777.785.391.5
• MSC [88]124.8M100%78.285.392.233.443.783.442.453.685.969.974.991.277.484.791.7
• PPT [90] (sup.)124.8M100%78.685.992.336.046.283.843.355.786.474.380.192.080.887.792.6
• Sonata124.8M100%79.486.192.536.846.584.443.755.886.676.081.693.082.389.993.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.365, + 0.249, + 0.615, + 0.262 + ], + "angle": 0, + "content": "Table 6. Indoor semantic segmentation." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.266, + 0.902, + 0.412 + ], + "angle": 0, + "content": "
Indoor Ins. SegParamsScanNet Val [23]ScanNet200 Val [71]ScanNet++ Val [101]S3DIS Area 5 [1]
MethodsLearn.Pct.mAP25mAP50mAPmAP25mAP50mAPmAP25mAP50mAPmAP25mAP50mAP
○PointGroup [45]124.8M100%77.561.740.940.133.223.136.730.720.955.749.437.8
•MSC (lin.)<0.2M<0.2%13.35.32.32.31.00.44.82.61.319.013.09.7
•Sonata (lin.)<0.2M<0.2%72.653.930.730.921.310.931.622.412.245.836.626.1
•Sonata (dec.)16.3M13%76.862.840.840.833.322.838.129.118.863.757.145.1
•MSC (full)124.8M100%78.462.941.140.533.823.438.930.921.756.350.538.1
•PPT [90] (sup.)124.8M100%78.963.542.140.834.124.039.332.821.957.551.239.7
•Sonata (full)124.8M100%79.263.942.442.135.625.440.333.622.363.857.445.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.368, + 0.413, + 0.612, + 0.427 + ], + "angle": 0, + "content": "Table 7. Indoor instance segmentation." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.439, + 0.482, + 0.554 + ], + "angle": 0, + "content": "
Outdoor Sem. Seg.nuScenes Val [9]Waymo Val [74]Sem.KITTI Val [6]
MethodsmIoUmAccallAccmIoUmAccallAccmIoUmAcc allAcc
o PTv3 [89]80.487.294.771.380.594.769.176.192.6
• Sonata (lin.)66.177.292.460.572.592.562.072.591.0
• Sonata (dec.)77.385.994.270.878.894.368.476.592.3
• PPT [90] (sup.)81.287.594.872.181.394.872.377.593.4
• Sonata (full)81.787.995.072.981.994.972.677.993.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.158, + 0.556, + 0.419, + 0.57 + ], + "angle": 0, + "content": "Table 8. Outdoor semantic segmentation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.575, + 0.483, + 0.619 + ], + "angle": 0, + "content": "pre-training performance across benchmarks. This clearly demonstrates that, unlike previous SSL approaches, the Sonata representation encodes instance-level information." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.62, + 0.483, + 0.816 + ], + "angle": 0, + "content": "Outdoor semantic segmentation. In Tab. 8, we adapt pre-training paradigm of Sonata to outdoor LiDAR scenarios through joint training on nuScenes [9], Waymo [74], and SemanticKITTI [6] and evaluate semantic segmentation performance using our evaluation protocols. With linear probing, Sonata sets a robust parameter efficiency baseline. Decoder probing achieves significant gains. In full fine-tuning, Sonata surpasses the supervised PPT [90], establishing new SOTA mIoU scores of 81.7, 72.9, and 72.6 across these benchmarks, underscoring the effectiveness of Sonata in outdoor perception tasks. Note that most of that performance can be recovered by more efficient decoder-only probing with a \\(95\\%\\), \\(97\\%\\), and \\(94\\%\\) of full fine-tuning." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.823, + 0.34, + 0.838 + ], + "angle": 0, + "content": "6. Conclusion and Discussion" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "In this work, we make progress towards self-supervised learning of a strong and reliable 3D point representation, that can zero-shot correspond semantically similar 3D points, down to the instance level, even under the presence" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.437, + 0.907, + 0.726 + ], + "angle": 0, + "content": "of spatial and visual perturbations. We demonstrate that such a representation can serve as the foundation for 3D tasks in semantic and instance-level grouping. We find existing 3D self-supervised learning approaches lacking, and hypothesize that this is due to what we term the geometric shortcut, a problem unique to 3D, which causes representations to collapse to low-level spatial features. We demonstrate the deficiency of these collapsed representations through linear probing. Beginning with a point self-distillation framework, we tackle the geometric shortcut by attaching SSL losses at coarser spatial scales, disturbing the spatial information of masked points with no features, and progressively increasing task difficulty to prevent overreliance on accessible geometric cues. This change enables effective scaling up, ultimately composing a Sonata from 140k point clouds. Sonata demonstrates semantically meaningful zero-shot visualization, as well as exceptional parameter and data efficiency. Full fine-tuning further advances SOTA across 3D indoor and outdoor perception tasks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.909, + 0.903 + ], + "angle": 0, + "content": "We summarize our limitations and future works and encourage readers to refer to the supplementary for details: 1. Sonata representations could be further semantically enriched through augmentation with 1M object-level point clouds [53]; 2. Pursuing unified training across indoor and outdoor scenarios as a feasible and promising direction; 3. Scaling up training with pixel-aligned [8] and SLAM-generated [27] point clouds from video data; 4. Leveraging cross-modal distillation to allow Sonata and DINOV2 to complement each other. We hope our insights will contribute to future research in 3D representation learning." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22200" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.259, + 0.108 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.484, + 0.192 + ], + "angle": 0, + "content": "We extend our gratitude to Maxime Oquab and Piotr Bojanowski for their guidance on the DINOv2 training details, to Saining Xie for insightful discussions on the vision of 3D representation learning, and to Paul Mcvay for his thoughts on the JPEA framework with sparse point cloud data." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.204, + 0.188, + 0.219 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.228, + 0.484, + 0.284 + ], + "angle": 0, + "content": "[1] Iro Armeni, Ozan Sener, Amir R. Zamir, Helen Jiang, Ioannis Brilakis, Martin Fischer, and Silvio Savarese. 3d semantic parsing of large-scale indoor spaces. In CVPR, 2016. 2, 6, 7, 8, 16" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.285, + 0.484, + 0.352 + ], + "angle": 0, + "content": "[2] Mahmoud Assran, Quentin Duval, Ishan Misra, Piotr Bojanowski, Pascal Vincent, Michael Rabbat, Yann LeCun, and Nicolas Ballas. Self-supervised learning from images with a joint-embedding predictive architecture. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.354, + 0.483, + 0.436 + ], + "angle": 0, + "content": "[3] Armen Avetisyan, Christopher Xie, Henry Howard-Jenkins, Tsun-Yi Yang, Samir Aroudj, Suvam Patra, Fuyang Zhang, Duncan Frost, Luke Holland, Campbell Orme, et al. Scenescript: Reconstructing scenes with an autoregressive structured language model. In ECCV, 2024. 2, 6, 16" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.438, + 0.482, + 0.466 + ], + "angle": 0, + "content": "[4] Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. Stat, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.467, + 0.484, + 0.549 + ], + "angle": 0, + "content": "[5] Gilad Baruch, Zhuoyuan Chen, Afshin Dehghan, Tal Dimry, Yuri Feigin, Peter Fu, Thomas Gebauer, Brandon Joffe, Daniel Kurz, Arik Schwartz, and Elad Shulman. ARKitscenes - a diverse real-world dataset for 3d indoor scene understanding using mobile RGB-d data. In NeurIPSW, 2021. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.551, + 0.484, + 0.606 + ], + "angle": 0, + "content": "[6] Jens Behley, Martin Garbade, Andres Milioto, Jan Quenzel, Sven Behnke, Cyril Stachniss, and Jurgen Gall. Semantickitti: A dataset for semantic scene understanding of lidar sequences. In ICCV, 2019. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.607, + 0.482, + 0.636 + ], + "angle": 0, + "content": "[7] Yoshua Bengio, Jérôme Louradour, Ronan Collobert, and Jason Weston. Curriculum learning. In ICML, 2009. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.636, + 0.482, + 0.704 + ], + "angle": 0, + "content": "[8] Aleksei Bochkovskii, Amael Delaunoy, Hugo Germain, Marcel Santos, Yichao Zhou, Stephan R Richter, and Vladlen Koltun. Depth pro: Sharp monocular metric depth in less than a second. arXiv preprint arXiv:2410.02073, 2024. 8, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.705, + 0.482, + 0.773 + ], + "angle": 0, + "content": "[9] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In CVPR, 2020. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.774, + 0.482, + 0.816 + ], + "angle": 0, + "content": "[10] Mathilde Caron, Piotr Bojanowski, Armand Joulin, and Matthijs Douze. Deep clustering for unsupervised learning of visual features. In ECCV, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.817, + 0.482, + 0.872 + ], + "angle": 0, + "content": "[11] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. In NeurIPS, 2020. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.873, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[12] Mathilde Caron, Hugo Touvron, Ishan Misra, Herve Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin." + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.228, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.554, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "Emerging properties in self-supervised vision transformers. In CVPR, 2021. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.122, + 0.905, + 0.164 + ], + "angle": 0, + "content": "[13] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In ICML, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.166, + 0.905, + 0.207 + ], + "angle": 0, + "content": "[14] Xi Chen, Lianghua Huang, Yu Liu, Yujun Shen, Deli Zhao, and Hengshuang Zhao. Anydoor: Zero-shot object-level image customization. In CVPR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.209, + 0.905, + 0.251 + ], + "angle": 0, + "content": "[15] Yukang Chen, Jianhui Liu, Xiangyu Zhang, Xiaojuan Qi, and Jiaya Jia. Largekernel3d: Scaling up kernels in 3d sparse cnns. In CVPR, 2023. 16" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.253, + 0.905, + 0.295 + ], + "angle": 0, + "content": "[16] Hung-Yueh Chiang, Yen-Liang Lin, Yueh-Cheng Liu, and Winston H Hsu. A unified point-based framework for 3d segmentation. In 3DV, 2019. 16" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.296, + 0.905, + 0.338 + ], + "angle": 0, + "content": "[17] Christopher Choy, JunYoung Gwak, and Silvio Savarese. 4d spatio-temporal convnets: Minkowski convolutional neural networks. In CVPR, 2019. 3, 7, 8, 16, 17" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.34, + 0.905, + 0.368 + ], + "angle": 0, + "content": "[18] Christopher Choy, Jaesik Park, and Vladlen Koltun. Fully convolutional geometric features. In ICCV, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.37, + 0.905, + 0.411 + ], + "angle": 0, + "content": "[19] Pointcept Contributors. Pointcept: A codebase for point cloud perception research. https://github.com/Pointcept/Pointcept, 2023.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.413, + 0.905, + 0.455 + ], + "angle": 0, + "content": "[20] Spconv Contributors. Spconv: Spatially sparse convolution library. https://github.com/traveller59/spconv, 2022.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.456, + 0.905, + 0.512 + ], + "angle": 0, + "content": "[21] Brian Curless and Marc Levoy. A volumetric method for building complex models from range images. In Proceedings of the 23rd annual conference on Computer graphics and interactive techniques, pages 303-312, 1996. 16" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.514, + 0.905, + 0.554 + ], + "angle": 0, + "content": "[22] Angela Dai and Matthias Nießner. 3dmv: Joint 3d-multiview prediction for 3d semantic scene segmentation. In ECCV, 2018. 16" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.557, + 0.905, + 0.612 + ], + "angle": 0, + "content": "[23] Angela Dai, Angel X. Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In CVPR, 2017. 2, 5, 6, 7, 8, 16" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.614, + 0.905, + 0.655 + ], + "angle": 0, + "content": "[24] Timothee Darcet, Maxime Oquab, Julien Mairal, and Piotr Bojanowski. Vision transformers need registers. arXiv:2309.16588, 2023. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.657, + 0.905, + 0.699 + ], + "angle": 0, + "content": "[25] Carl Doersch, Abhinav Gupta, and Alexei A Efros. Unsupervised visual representation learning by context prediction. In ICCV, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.701, + 0.905, + 0.757 + ], + "angle": 0, + "content": "[26] Alexey Dosovitskiy, Philipp Fischer, Jost Tobias Springenberg, Martin Riedmiller, and Thomas Brox. Discriminative unsupervised feature learning with exemplar convolutional neural networks. TPAMI, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.758, + 0.905, + 0.828 + ], + "angle": 0, + "content": "[27] Jakob Engel, Kiran Somasundaram, Michael Goesele, Albert Sun, Alexander Gamino, Andrew Turner, Arjang Talattof, Arnie Yuan, Bilal Souti, Brighid Meredith, et al. Project aria: A new tool for egocentric multi-modal ai research. arXiv:2308.13561, 2023. 2, 8, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.829, + 0.905, + 0.858 + ], + "angle": 0, + "content": "[28] Pedro F Felzenszwalb and Daniel P Huttenlocher. Efficient graph-based image segmentation. IJCV, 2004. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.859, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[29] Benjamin Graham and Laurens van der Maaten. Submanifold sparse convolutional networks. arXiv:1706.01307, 2017. 3" + }, + { + "type": "list", + "bbox": [ + 0.524, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "22201" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.093, + 0.483, + 0.134 + ], + "angle": 0, + "content": "[30] Benjamin Graham, Martin Engelcke, and Laurens van der Maaten. 3d semantic segmentation with submanifold sparse convolutional networks. In CVPR, 2018. 16" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.136, + 0.483, + 0.205 + ], + "angle": 0, + "content": "[31] Jean-Bastien Grill, Florian Strub, Florent Alché, Coretin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. In NerIPS, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.207, + 0.483, + 0.275 + ], + "angle": 0, + "content": "[32] Jiayuan Gu, Fanbo Xiang, Xuanlin Li, Zhan Ling, Xiqiang Liu, Tongzhou Mu, Yihe Tang, Stone Tao, Xinyue Wei, Yunchao Yao, Xiaodi Yuan, Pengwei Xie, Zhiao Huang, Rui Chen, and Hao Su. Maniskill2: A unified benchmark for generalizable manipulation skills. In ICLR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.277, + 0.482, + 0.318 + ], + "angle": 0, + "content": "[33] Bharath Hariharan, Pablo Arbelaez, Ross Girshick, and Jitendra Malik. Hypercolumns for object segmentation and fine-grained localization. In CVPR, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.32, + 0.482, + 0.361 + ], + "angle": 0, + "content": "[34] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In CVPR, 2020. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.362, + 0.482, + 0.403 + ], + "angle": 0, + "content": "[35] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.406, + 0.482, + 0.473 + ], + "angle": 0, + "content": "[36] R. Devon Hjelm, Alex Fedorov, Samuel Lavoie-Marchildon, Karan Grewal, Phil Bachman, Adam Trischler, and Yoshua Bengio. Learning deep representations by mutual information estimation and maximization. In ICLR, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.476, + 0.482, + 0.532 + ], + "angle": 0, + "content": "[37] Yicong Hong, Kai Zhang, Jiuming Gu, Sai Bi, Yang Zhou, Difan Liu, Feng Liu, Kalyan Sunkavalli, Trung Bui, and Hao Tan. Lrm: Large reconstruction model for single image to 3d. arXiv preprint arXiv:2311.04400, 2023. 16" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.533, + 0.482, + 0.587 + ], + "angle": 0, + "content": "[38] Ji Hou, Benjamin Graham, Matthias Nießner, and Saining Xie. Exploring data-efficient 3d scene understanding with contrastive scene contexts. In CVPR, 2021. 2, 3, 7, 8, 16, 17" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.59, + 0.482, + 0.645 + ], + "angle": 0, + "content": "[39] Qingyong Hu, Bo Yang, Linhai Xie, Stefano Rosa, Yulan Guo, Zhihua Wang, Niki Trigoni, and Andrew Markham. Randla-net: Efficient semantic segmentation of large-scale point clouds. In CVPR, 2020. 16" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.647, + 0.482, + 0.7 + ], + "angle": 0, + "content": "[40] Zeyu Hu, Mingmin Zhen, Xuyang Bai, Hongbo Fu, and Chiew-lan Tai. Jsenet: Joint semantic segmentation and edge detection network for 3d point clouds. In ECCV, 2020. 16" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.703, + 0.482, + 0.744 + ], + "angle": 0, + "content": "[41] Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In ICML, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.746, + 0.482, + 0.802 + ], + "angle": 0, + "content": "[42] Ayush Jain, Pushkal Katara, Nikolaos Gkanatsios, Adam W. Harley, Gabriel Sarch, Kriti Aggarwal, Vishrav Chaudhary, and Katerina Fragkiadaki. Odin: A single model for 2d and 3d perception. In CVPR, 2024. 16" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.803, + 0.482, + 0.844 + ], + "angle": 0, + "content": "[43] Stephen James, Zicong Ma, David Rovick Arrojo, and Andrew J. Davison. Rlbench: The robot learning benchmark & learning environment. RAL, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.846, + 0.482, + 0.899 + ], + "angle": 0, + "content": "[44] Li Jiang, Hengshuang Zhao, Shu Liu, Xiaoyong Shen, Chi-Wing Fu, and Jiaya Jia. Hierarchical point-edge interaction network for point cloud semantic segmentation. In ICCV, 2019. 17" + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.093, + 0.483, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "[45] Li Jiang, Hengshuang Zhao, Shaoshuai Shi, Shu Liu, Chi-Wing Fu, and Jiaya Jia. Pointgroup: Dual-set point grouping for 3d instance segmentation. CVPR, 2020. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.136, + 0.905, + 0.176 + ], + "angle": 0, + "content": "[46] Maxim Kolodiazhnyi, Anna Vorontsova, Anton Konushin, and Danila Rukhovich. Oneformer3d: One transformer for unified point cloud segmentation. In CVPR, 2024. 16" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.178, + 0.905, + 0.23 + ], + "angle": 0, + "content": "[47] Xin Lai, Jianhui Liu, Li Jiang, Liwei Wang, Hengshuang Zhao, Shu Liu, Xiaojuan Qi, and Jiaya Jia. Stratified transformer for 3d point cloud segmentation. In CVPR, 2022. 16, 17" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.233, + 0.905, + 0.272 + ], + "angle": 0, + "content": "[48] Loic Landrieu and Martin Simonovsky. Large-scale point cloud semantic segmentation with superpoint graphs. In CVPR, 2018. 17" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.274, + 0.905, + 0.315 + ], + "angle": 0, + "content": "[49] Huan Lei, Naveed Akhtar, and Ajmal Mian. Seggen: Efficient 3d point cloud segmentation with fuzzy spherical kernel. In CVPR, 2020. 16, 17" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.317, + 0.905, + 0.357 + ], + "angle": 0, + "content": "[50] Yangyan Li, Rui Bu, Mingchao Sun, Wei Wu, Xinhan Di, and Baoquan Chen. Pointconn: Convolution on x-transformed points. NeurIPS, 2018. 16, 17" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.359, + 0.905, + 0.425 + ], + "angle": 0, + "content": "[51] Dingkang Liang, Xin Zhou, Wei Xu, Xingkui Zhu, Zhikang Zou, Xiaqing Ye, Xiao Tan, and Xiang Bai. Pointmamba: A simple state space model for point cloud analysis. In Advances in Neural Information Processing Systems, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.428, + 0.905, + 0.482 + ], + "angle": 0, + "content": "[52] Haojia Lin, Xiawu Zheng, Lijiang Li, Fei Chao, Shanshan Wang, Yan Wang, Yonghong Tian, and Rongrong Ji. Meta architecture for point cloud analysis. In CVPR, pages 17682-17691, 2023. 16, 17" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.484, + 0.905, + 0.538 + ], + "angle": 0, + "content": "[53] Minghua Liu, Ruoxi Shi, Kaiming Kuang, Yinhao Zhu, Xu-anlin Li, Shizhong Han, Hong Cai, Fatih Porikli, and Hao Su. Openshape: Scaling up 3d shape representation towards open-world understanding. In NeurIPS, 2023. 8, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.54, + 0.905, + 0.566 + ], + "angle": 0, + "content": "[54] I Loshchilov. Decoupled weight decay regularization. In ICLR, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.568, + 0.905, + 0.593 + ], + "angle": 0, + "content": "[55] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. In *ICLR*, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.596, + 0.905, + 0.636 + ], + "angle": 0, + "content": "[56] Xu Ma, Can Qin, Haoxuan You, Haoxi Ran, and Yun Fu. Rethinking network design and local geometry in point cloud: A simple residual mlp framework. *ICLR*, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.638, + 0.905, + 0.678 + ], + "angle": 0, + "content": "[57] Gaku Narita, Takashi Seno, Tomoya Ishikawa, and Yohsuke Kaji. Panopticfusion: Online volumetric semantic mapping at the level of stuff and things. In IROS, 2019. 16" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.68, + 0.905, + 0.774 + ], + "angle": 0, + "content": "[58] Richard A Newcombe, Shahram Izadi, Otmar Hilliges, David Molyneaux, David Kim, Andrew J Davison, Pushmeet Kohi, Jamie Shotton, Steve Hodges, and Andrew Fitzgibbon. Kinectfusion: Real-time dense surface mapping and tracking. In 2011 10th IEEE international symposium on mixed and augmented reality, pages 127-136. IEEE, 2011. 16" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.776, + 0.905, + 0.816 + ], + "angle": 0, + "content": "[59] Mehdi Noroozi and Paolo Favaro. Unsupervised learning of visual representations by solving jigsaw puzzles. In ECCV, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.818, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[60] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Russell Howes, Po-Yao Huang, Hu Xu, Vasu Sharma, Shang-Wen Li, Wojciech Galuba, Mike Rabbat, Mido Assan, Nicolas Ballas, Gabriel Synnaeve, Ishan Misra, Herve" + }, + { + "type": "list", + "bbox": [ + 0.525, + 0.093, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22202" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.131, + 0.092, + 0.482, + 0.134 + ], + "angle": 0, + "content": "Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. Dinov2: Learning robust visual features without supervision. TMLR, 2024. 2, 3, 4, 6, 7, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.135, + 0.483, + 0.176 + ], + "angle": 0, + "content": "[61] Yatian Pang, Wenxiao Wang, Francis EH Tay, Wei Liu, Yonghong Tian, and Li Yuan. Masked autoencoders for point cloud self-supervised learning. In ECCV, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.177, + 0.482, + 0.204 + ], + "angle": 0, + "content": "[62] Chunghyun Park, Yoonwoo Jeong, Minsu Cho, and Jaesik Park. Fast point transformer. In CVPR, 2022. 16" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.205, + 0.483, + 0.272 + ], + "angle": 0, + "content": "[63] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. NeurIPS, 2019. 16" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.274, + 0.482, + 0.328 + ], + "angle": 0, + "content": "[64] Bohao Peng, Xiaoyang Wu, Li Jiang, Yukang Chen, Hengshuang Zhao, Zhuotao Tian, and Jiaya Jia. Oa-cnns: Omni-adaptive sparse cnns for 3d semantic segmentation. In CVPR, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.33, + 0.482, + 0.371 + ], + "angle": 0, + "content": "[65] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In CVPR, 2017. 7, 17" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.372, + 0.482, + 0.412 + ], + "angle": 0, + "content": "[66] Charles R Qi, Li Yi, Hao Su, and Leonidas J Guibas. Point-net++: Deep hierarchical feature learning on point sets in a metric space. In NeurIPS, 2017. 3, 16, 17" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.414, + 0.482, + 0.469 + ], + "angle": 0, + "content": "[67] Guocheng Qian, Yuchen Li, Houwen Peng, Jinjie Mai, Hasan Hammoud, Mohamed Elhoseiny, and Bernard Ghanem. Pointnext: Revisiting pointnet++ with improved training and scaling strategies. In NeurIPS, 2022. 3, 16, 17" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.47, + 0.482, + 0.565 + ], + "angle": 0, + "content": "[68] Santhosh Kumar Ramakrishnan, Aaron Gokaslan, Erik Wijmans, Oleksandr Maksymets, Alexander Clegg, John M Turner, Eric Undersander, Wojciech Galuba, Andrew Westbury, Angel X Chang, Manolis Savva, Yili Zhao, and Dhruv Batra. Habitat-matterport 3d dataset (HM3d): 1000 large-scale 3d environments for embodied AI. In NeurIPS, 2021. 2, 6, 14, 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.567, + 0.482, + 0.607 + ], + "angle": 0, + "content": "[69] Damien Robert, Hugo Raguet, and Loic Landrieu. Efficient 3d semantic segmentation with superpoint transformer. In ICCV, 2023. 17" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.609, + 0.482, + 0.649 + ], + "angle": 0, + "content": "[70] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In MICCAI, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.65, + 0.482, + 0.691 + ], + "angle": 0, + "content": "[71] David Rozenberszki, Or Litany, and Angela Dai. Language-grounded indoor 3d semantic segmentation in the wild. In ECCV, 2022. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.692, + 0.482, + 0.733 + ], + "angle": 0, + "content": "[72] Alexandre Sablayrolles, Matthijs Douze, Cordelia Schmid, and Hervé Jégou. Spreading vectors for similarity search. In ICLR, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.735, + 0.482, + 0.789 + ], + "angle": 0, + "content": "[73] Julian Straub, Daniel DeTone, Tianwei Shen, Nan Yang, Chris Sweeney, and Richard Newcombe. Efm3d: A benchmark for measuring progress towards 3d egocentric foundation models. arXiv:2406.10224, 2024. 2, 14, 15, 16" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.79, + 0.482, + 0.858 + ], + "angle": 0, + "content": "[74] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, et al. Scalability in perception for autonomous driving: Waymo open dataset. In CVPR, 2020. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.86, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[75] Maxim Tatarchenko, Jaesik Park, Vladlen Koltun, and Qian-Yi Zhou. Tangent convolutions for dense prediction in 3d. In CVPR, 2018. 17" + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.092, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "[76] Lyne Tchapmi, Christopher Choy, Iro Armeni, JunYoung Gwak, and Silvio Savarese. Segcloud: Semantic segmentation of 3d point clouds. In 3DV, 2017. 17" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.136, + 0.905, + 0.189 + ], + "angle": 0, + "content": "[77] Hugues Thomas, Charles R Qi, Jean-Emmanuel Deschaud, Beatrix Marcotegui, François Goulette, and Leonidas J Guibas. Kpconv: Flexible and deformable convolution for point clouds. In ICCV, 2019. 3, 16, 17" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.191, + 0.905, + 0.231 + ], + "angle": 0, + "content": "[78] Hugues Thomas, Yao-Hung Hubert Tsai, Timothy D Barfoot, and Jian Zhang. Kpconvx: Modernizing kernel point convolution with kernel attention. In CVPR, 2024. 16" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.232, + 0.905, + 0.286 + ], + "angle": 0, + "content": "[79] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 16" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.288, + 0.905, + 0.343 + ], + "angle": 0, + "content": "[80] Chengyao Wang, Li Jiang, Xiaoyang Wu, Zhuotao Tian, Bohao Peng, Hengshuang Zhao, and Jiaya Jia. Groupcontrast: Semantic-aware self-supervised representation learning for 3d understanding. In CVPR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.344, + 0.905, + 0.385 + ], + "angle": 0, + "content": "[81] Lei Wang, Yuchun Huang, Yaolin Hou, Shenman Zhang, and Jie Shan. Graph attention convolution for point cloud semantic segmentation. In CVPR, 2019. 17" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.386, + 0.904, + 0.412 + ], + "angle": 0, + "content": "[82] Peng-Shuai Wang. Octformer: Octree-based transformers for 3D point clouds. SIGGRAPH, 2023. 16" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.414, + 0.904, + 0.454 + ], + "angle": 0, + "content": "[83] Shenlong Wang, Simon Suo, Wei-Chiu Ma, Andrei Pokrovsky, and Raquel Urtasun. Deep parametric continuous convolutional neural networks. In CVPR, 2018. 17" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.455, + 0.905, + 0.496 + ], + "angle": 0, + "content": "[84] Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. Dust3r: Geometric 3d vision made easy. In CVPR, 2024. 2, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.497, + 0.905, + 0.537 + ], + "angle": 0, + "content": "[85] Wenxuan Wu, Zhongang Qi, and Li Fuxin. Pointconv: Deep convolutional networks on 3d point clouds. In CVPR, 2019. 16" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.539, + 0.905, + 0.578 + ], + "angle": 0, + "content": "[86] Wenxuan Wu, Li Fuxin, and Qi Shan. Pointconvformer: Revenge of the point-based convolution. In CVPR, 2023. 16" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.581, + 0.905, + 0.634 + ], + "angle": 0, + "content": "[87] Xiaoyang Wu, Yixing Lao, Li Jiang, Xihui Liu, and Hengshuang Zhao. Point transformer v2: Grouped vector attention and partition-based pooling. In NeurIPS, 2022. 7, 16, 17" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.636, + 0.905, + 0.691 + ], + "angle": 0, + "content": "[88] Xiaoyang Wu, Xin Wen, Xihui Liu, and Hengshuang Zhao. Masked scene contrast: A scalable framework for unsupervised 3d representation learning. In CVPR, 2023. 2, 3, 4, 5, 6, 7, 8, 14, 16, 17" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.692, + 0.905, + 0.748 + ], + "angle": 0, + "content": "[89] Xiaoyang Wu, Li Jiang, Peng-Shuai Wang, Zhijian Liu, Xihui Liu, Yu Qiao, Wanli Ouyang, Tong He, and Hengshuang Zhao. Point transformer v3: Simpler, faster, stronger. In CVPR, 2024. 3, 5, 7, 8, 16, 17" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.749, + 0.905, + 0.803 + ], + "angle": 0, + "content": "[90] Xiaoyang Wu, Zhuotao Tian, Xin Wen, Bohao Peng, Xihui Liu, Kaicheng Yu, and Hengshuang Zhao. Towards large-scale 3d representation learning with multi-dataset point prompt training. In CVPR, 2024. 3, 6, 7, 8, 16, 17" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.804, + 0.905, + 0.844 + ], + "angle": 0, + "content": "[91] Zhirong Wu, Yuanjun Xiong, Stella X. Yu, and Dahua Lin. Unsupervised feature learning via non-parametric instance discrimination. In CVPR, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.845, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[92] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. In NeurIPS, 2021. 3" + }, + { + "type": "list", + "bbox": [ + 0.524, + 0.093, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22203" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.093, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[93] Saining Xie, Jiatao Gu, Demi Guo, Charles R Qi, Leonidas Guibas, and Or Litany. Pointcontrast: Unsupervised pretraining for 3d point cloud understanding. In ECCV, 2020. 2, 3, 4, 6, 7, 8, 16, 17" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.149, + 0.482, + 0.203 + ], + "angle": 0, + "content": "[94] Zhenda Xie, Zheng Zhang, Yue Cao, Yutong Lin, Jianmin Bao, Zhuliang Yao, Qi Dai, and Han Hu. Simmim: A simple framework for masked image modeling. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.205, + 0.482, + 0.247 + ], + "angle": 0, + "content": "[95] Mutian Xu, Runyu Ding, Hengshuang Zhao, and Xiaojuan Qi. Paconv: Position adaptive convolution with dynamic kernel assembling on point clouds. In CVPR, 2021. 17" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.247, + 0.482, + 0.3 + ], + "angle": 0, + "content": "[96] Xu Yan, Chaoda Zheng, Zhen Li, Sheng Wang, and Shuguang Cui. Pointasnl: Robust point clouds processing using nonlocal neural networks with adaptive sampling. In CVPR, 2020. 16" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.302, + 0.482, + 0.357 + ], + "angle": 0, + "content": "[97] Jiancheng Yang, Qiang Zhang, Bingbing Ni, Linguuo Li, Jinxian Liu, Mengdie Zhou, and Qi Tian. Modeling point clouds with self-attention and gumbel subset sampling. In CVPR, 2019. 17" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.359, + 0.482, + 0.412 + ], + "angle": 0, + "content": "[98] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Ji-ashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. In CVPR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.414, + 0.482, + 0.455 + ], + "angle": 0, + "content": "[99] Lihe Yang, Bingyi Kang, Zilong Huang, Zhen Zhao, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything v2. In NeurIPS, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.457, + 0.482, + 0.51 + ], + "angle": 0, + "content": "[100] Yu-Qi Yang, Yu-Xiao Guo, Jian-Yu Xiong, Yang Liu, Hao Pan, Peng-Shuai Wang, Xin Tong, and Baining Guo. Swin3d: A pretrained transformer backbone for 3d indoor scene understanding. arXiv:2304.06906, 2023. 16, 17" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.511, + 0.482, + 0.552 + ], + "angle": 0, + "content": "[101] Chandan Yeshwanth, Yueh-Cheng Liu, Matthias Nießner, and Angela Dai. Scannet++: A high-fidelity dataset of 3d indoor scenes. In ICCV, 2023. 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.553, + 0.482, + 0.607 + ], + "angle": 0, + "content": "[102] Xumin Yu, Lulu Tang, Yongming Rao, Tiejun Huang, Jie Zhou, and Jiwen Lu. Point-bert: Pre-training 3d point cloud transformers with masked point modeling. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.609, + 0.482, + 0.65 + ], + "angle": 0, + "content": "[103] Feihu Zhang, Jin Fang, Benjamin Wah, and Philip Torr. Deep fusionnet for point cloud semantic segmentation. In ECCV, 2020. 16" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.651, + 0.482, + 0.705 + ], + "angle": 0, + "content": "[104] Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel M. Ni, and Heung-Yeung Shum. Dino: Detr with improved denoising anchor boxes for end-to-end object detection. arXiv:2203.03605, 2022. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.707, + 0.482, + 0.761 + ], + "angle": 0, + "content": "[105] Michael Zhang, Aditi Raghunathan Wang, Shiori Sagawa, Tatsunori B Hashimoto, and Percy Liang. Fine-tuning can distort pretrained features and underperform out-of-distribution. In ICLR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.762, + 0.482, + 0.817 + ], + "angle": 0, + "content": "[106] Renrui Zhang, Ziyu Guo, Peng Gao, Rongyao Fang, Bin Zhao, Dong Wang, Yu Qiao, and Hongsheng Li. Pointm2ae: Multi-scale masked autoencoders for hierarchical point cloud pre-training. In NeurIPS, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.818, + 0.482, + 0.859 + ], + "angle": 0, + "content": "[107] Hengshuang Zhao, Li Jiang, Chi-Wing Fu, and Jiaya Jia. Pointweb: Enhancing local neighborhood features for point cloud processing. In CVPR, 2019. 17" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.86, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[108] Hengshuang Zhao, Li Jiang, Jiaya Jia, Philip Torr, and Vladlen Koltun. Point transformer. In ICCV, 2021. 16, 17" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.093, + 0.482, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[109] Jia Zheng, Junfei Zhang, Jing Li, Rui Tang, Shenghua Gao, and Zihan Zhou. Structured3d: A large photo-realistic dataset for structured 3d modeling. In ECCV, 2020. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.149, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[110] Jinghao Zhou, Chen Wei, Huiyu Wang, Wei Shen, Cihang Xie, Alan Yuille, and Tao Kong. ibot: Image bert pretraining with online tokenizer. In ICLR, 2022. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.192, + 0.905, + 0.233 + ], + "angle": 0, + "content": "[111] Jinghao Zhou, Chen Wei, Huiyu Wang, Wei Shen, Cihang Xie, Alan Yuille, and Tao Kong. Image BERT pre-training with online tokenizer. In ICLR, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.235, + 0.905, + 0.263 + ], + "angle": 0, + "content": "[112] Xingyi Zhou, Vladlen Koltun, and Philipp Krahenbuhl. Simple multi-dataset detection. In CVPR, 2022. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22204" + } + ] +] \ No newline at end of file diff --git a/2025/Sonata_ Self-Supervised Learning of Reliable Point Representations/3e06b74b-3b4f-4789-9e14-04119c218bb3_origin.pdf b/2025/Sonata_ Self-Supervised Learning of Reliable Point Representations/3e06b74b-3b4f-4789-9e14-04119c218bb3_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..bef8bcd7ad981db45c93a01d4fa28a67f48a3a85 --- /dev/null +++ b/2025/Sonata_ Self-Supervised Learning of Reliable Point Representations/3e06b74b-3b4f-4789-9e14-04119c218bb3_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ae9de0cb85849b78fc5fc8e7213ddce91de6d8771188b533840f1e87bc284923 +size 9682215 diff --git a/2025/Sonata_ Self-Supervised Learning of Reliable Point Representations/full.md b/2025/Sonata_ Self-Supervised Learning of Reliable Point Representations/full.md new file mode 100644 index 0000000000000000000000000000000000000000..18509a41c7b8097ccaa69c6d1ad29ecaa792d19e --- /dev/null +++ b/2025/Sonata_ Self-Supervised Learning of Reliable Point Representations/full.md @@ -0,0 +1,342 @@ +# Sonata: Self-Supervised Learning of Reliable Point Representations + +Xiaoyang Wu $^{1,2}$ Daniel DeTone $^{2*}$ Duncan Frost $^{2*}$ Tianwei Shen $^{2*}$ Chris Xie $^{2*}$ Nan Yang $^{2*}$ Jakob Engel $^{2}$ Richard Newcombe $^{2}$ Hengshuang Zhao $^{1}$ Julian Straub $^{2}$ + +1The University of Hong Kong 2Meta Reality Labs Research + +https://github.com/facebookresearch/sonata + +![](images/589d4300f43854af0829f5157255ce003401f1421bbb5571ff40b45a650e5aba.jpg) + +![](images/836a6f483902be083a4e0b93f8367b5b8d30e5b14d1b990fba0f77f1190fb148.jpg) + +![](images/56c11ac45288e311d5e6dbd4e365b9f89130bdd1783474991acb5e5bfaa77ffe.jpg) + +![](images/6e714a08721434cff40b5c0910e86ef7bb6c7bb036bc13df01cef90c14096176.jpg) +Figure 1. Main properties. Sonata leads to reliable 3D self-supervised pretraining with the following superior and emerging properties: 1. Perception. Sonata advances state-of-the-art results across 3D indoor and outdoor perception tasks; 2. Linear probing. With less than $0.2\%$ learnable parameters, Sonata achieves strong and usable linear probing performance which is $3.3\times$ better than previous SOTA; 3. Decoder-free. Sonata moves beyond the inflexible U-Net structure, offering multi-scale representations that unchain future 3D research from previous architectural constraints. 4. Semantic awareness. Sonata reveals semantic structure in PCA and K-means visualizations. 5. Spatial reasoning. Sonata allows spatial correspondence even under strong augmentations as visualized via feature similarity. + +![](images/989600f4ef288c1b4c60baa764c2c9036547f67de46eb78da322b2e565fdf5a7.jpg) + +# Abstract + +In this paper, we question whether we have a reliable self-supervised point cloud model that can be used for diverse 3D tasks via simple linear probing, even with limited data and minimal computation. We find that existing 3D self-supervised learning approaches fall short when evaluated on representation quality through linear probing. We hypothesize that this is due to what we term the "geometric shortcut", which causes representations to collapse to low-level spatial features. This challenge is unique to 3D and arises from the sparse nature of point cloud data. + +We address it through two key strategies: obscuring spatial information and enhancing the reliance on input features, ultimately composing a Sonata of 140k point clouds through self-distillation. Sonata is simple and intuitive, yet its learned representations are strong and reliable: zero-shot visualizations demonstrate semantic grouping, alongside strong spatial reasoning through nearest-neighbor relationships. Sonata demonstrates exceptional parameter and data efficiency, tripling linear probing accuracy (from $21.8\%$ to $72.5\%$ ) on ScanNet and nearly doubling performance with only $1\%$ of the data compared to previous approaches. Full fine-tuning further advances SOTA across both 3D indoor and outdoor perception tasks. + +# 1. Introduction + +Self-supervised learning (SSL) with images [2, 13, 34, 94, 104] has seen a continuous increase in model simplicity, capacity, and capability over the past decade [25]. Tuning a single linear layer can achieve performance close to full fine-tuning [31, 35, 104], fostering growing trust in its reliability. This trust has been further strengthened by witnessing the semantic meaning of learned image representations through direct visualization [60, 110]. Consequently, these reliable self-supervised models have become the foundation for emerging approaches [14, 84, 98, 99] across various fields involving images. + +In contrast to the image domain, self-supervised learning with point clouds [18, 61, 93, 102, 106] is still in its early stages. Despite the broad reliance on 3D applications in autonomous driving [9, 74], robotic learning [32], mixed reality [5, 43] and egocentric perception [27, 73], the latest self-supervised point cloud models are seldom included in their pipelines. This gap prompts us to consider a simple yet critical question: do we have a reliable point self-supervised learning approach that provides strong representations, usable with simple linear probing across these applications? Not yet. Previous SOTAs [38, 88, 93] fall short on this higher-level criterion, achieving only a maximum of $21.8\%$ mIoU on ScanNet [23] with linear probing, especially given that the performance from scratch is $77.6\%$ . + +We identify the geometric shortcut as the primary issue hindering these prior point cloud SSL approaches from learning reliable representations. This shortcut refers to the tendency of the model to collapse to easily accessible, low-level geometric cues, such as normal direction or point height, as demonstrated in prior works and visualized in Fig. 2. This spatial information is inevitably introduced into point cloud operators along with point coordinates rather than through input features, making it difficult to obscure and nearly impossible to mask effectively. + +However, the model collapse caused by geometric shortcuts can be mitigated through two key strategies: obscuring spatial information and emphasizing input features. Specifically, we address this issue by applying SSL losses at coarser spatial scales, disturbing the spatial information of masked points without features, and progressively increasing task difficulty to reduce reliance on accessible geometric cues. Coupled with a point self-distillation framework and scaling techniques inspired by recent advances in image SSL [11, 12, 34, 60, 110], we ultimately compose a Sonata using 140k point cloud scenes [1, 3, 5, 23, 68, 101, 109]. + +Sonata demonstrates strong zero-shot capabilities, with PCA-colored visualizations of point clouds, k-means clustering of features, and nearest-neighbor matching between point clouds (see Fig. 1). Sonata also proves highly data-efficient, raising semantic segmentation performance from $25.8\%$ to $45.3\%$ under extremely limited data conditions + +![](images/b05cacfdb88fae41dddb3a576b8a396fcf586d24cec5ed0e7025e057ff9a51d6.jpg) +Figure 2. Geometric shortcut. We select a point on the sofa arm and compute pairwise similarity with other points. The similarity heatmap reveals that CSC [38] collapses to surface normals, and MSC [88] overfits to point height. In contrast, our Sonata can extract higher-level concepts, as can be seen by the high similarity between all sofa arms highlighted in red. + +(1% of ScanNet). Additionally, Sonata significantly boosts linear probing accuracy on ScanNet semantic segmentation, increasing it by over $3.3 \times$ from $21.8\%$ to $72.5\%$ and surpassing the accuracy of DINOv2 features aggregated onto the point cloud (63.1%). Moreover, combining Sonata features with DINOv2 features further enhances accuracy (76.4%), underscoring that Sonata captures unique 3D information beyond what is visible in images alone. Finally, Sonata achieves state-of-the-art results across various indoor and outdoor perception tasks with full fine-tuning. + +# 2. Related Work + +Image self-supervised learning. Over the past decade, remarkable advancements [25, 26, 36, 59, 91] have been made in image self-supervised learning, and our research is largely inspired by two pivotal moments in this field. First, linear probing, a method that assesses representation quality by optimizing a minimal linear transformation, has become a standard in 2D image SSL [10, 13, 34, 91]. In some cases, such as when the distribution shift is large, linear probing surpasses full fine-tuning [105]. Second, the ability to directly perceive the semantic meaning of learned representations through zero-shot visualization like PCA or attention [12, 60] has further strengthened trust in reliability. + +Point self-supervised learning. Sonata follows the research path initiated by PointContrast [38, 93], emphasizing self-supervised learning with scene-level data [23]. While previous efforts do implement strategies to prevent collapse from geometric shortcuts, they remain limited. For example, Masked Scene Contrast (MSC) [88] encourages learning beyond naive geometric cues by predicting color or normal vectors but still partially anchors representations to predefined tasks. GroupContrast (GC) [80] employs graph- + +![](images/c61e24b74f3f68c8dd8dbf6f8b7bfde203cf8897155a2cb77a0785bd964a9ac1.jpg) +Figure 3. The geometric shortcut is unique to 3D. When comparing the information contained in 2D image and 3D point cloud data after removing the input feature (indicated by color), it is evident that in images all information is within the input feature. Whereas point clouds retain geometric information in point positions, which is directly utilized by operators. This characteristic leads to what we term geometric shortcuts in 3D SSL. + +based segment guidance [28], though it is constrained by its reliance on human-designed algorithms. Building on MSC, Sonata directly addresses the geometric shortcut and scales up training [90] to establish a more reliable approach to point cloud self-supervised learning. + +Point cloud backbones. This area [17, 51, 56, 64, 67, 77] has significantly benefited from the U-Net structure [70], particularly with its hierarchical decoding and skip connections, as first introduced by PointNet++ [66]. However, the tight coupling between the encoder and decoder restricts flexibility and generalization capacity [92]. Sonata addresses this limitation by focusing self-supervised learning exclusively on the encoder, thus removing the hierarchical decoder. Moreover, unlike previous SOTAs [38, 88, 93] in point SSL, which primarily use SparseUNet [17, 20, 29] to balance efficiency and accuracy, we leverage Point Transformer V3 (PTv3) [89], an efficient, accurate, and scalable transformer backbone. This shift alone yields a $7.7\%$ improvement in linear probing performance over MSC [88]. + +# 3. Pilot Study and Design Principle + +In this section, we qualitatively study the problems of current point self-supervised approaches and point cloud backbones to inform the Sonata approach. + +Uncovering the geometric shortcut in point SSL. The history of image self-supervised learning [34, 60, 104, 112] can be summarized as a continuous battle against shortcuts, where models often exploit trivial solutions (mode collapse) rather than understanding deeper semantics. Each advancement has involved identifying, understanding, and overcoming these shortcuts, refining pretext tasks to push models to "struggle" in learning stronger representations. + +However in 3D point SSL, despite various attempts [38, 88, 93] to increase the difficulty of pretext tasks, a curious phenomenon persists—the loss consistently reduces rapidly to an ideal range in the early stages of training. We hypo- + +![](images/666811aba4d3585c2cf6b2ea3f73df98e6e959b8fcc6969307f1f17a20722f56.jpg) +Figure 4. What is learned by the hierarchical backbone? We visualize PCA embeddings from different stages of a hierarchical encoder and decoder, trained for semantic segmentation. The encoder captures diverse and dispersed feature patterns, indicating a broad range of information. Notably, as the point cloud becomes coarser, accessible geometric information within point coordinates becomes increasingly global. In contrast, the decoder's representations are more uniform and structured, suggesting a focus on refining features for task-specific outputs. + +esize that this lack of "struggle" in learning indicates that a shortcut exists that leads to a collapse of the representation to trivial solutions. Indeed as Fig. 2 illustrates, previous 3D SSL approaches seem to learn representations that are sensitive to local surface normals or point height. The collapse to such naive solutions even for well-designed SSL pretext tasks is what we term the geometric shortcut. + +Intuitively, representations affected by this geometric shortcut have not learned sufficient semantics. To quantify this problem, we leverage the standard criteria from 2D SSL of linear probing: we linearly probe the semantic class of each point from the learned representation. As illustrated in Fig. 1, with linear probing, previous 3D SSL methods PC [93] and MSC [88] indeed achieve only $5.6\%$ mIoU and $21.8\%$ mIoU respectively on ScanNet semantic segmentation. When compared to the $63.1\%$ mIoU achieved by 3D-aggregated image representations from DINOv2 [60], it becomes obvious that the current 3D SSL methods do not learn semantic information. + +We hypothesize that the geometric shortcut stems from the sparse nature of point cloud data. Every operator, whether for point clouds or images, inherently relies on point (pixel) coordinates to define the kernel. However, unlike images with regularly spaced dense pixels, the sparsity of point cloud data necessitates introducing point coordinate information into point cloud operators rather than just through input features (see Fig. 3). This fact makes it difficult to obscure and nearly impossible to effectively mask out, ultimately resulting in the geometric shortcut. + +These observations motivate the training methodology of Sonata aimed to prevent the dominance of naive spatial information inherent in point cloud organization. + +Focusing self-supervised learning on encoder only. Following our hypothesis that the geometric arrangement leaks information into the learned representation via the geometric shortcut, we examine the point cloud model. The typical U-Net structure [66, 70] is effective at handling large point clouds in a coarse-to-fine way, but the tight coupling of the encoder and decoder via skip connections restricts + +flexibility. Specifically, the decoder enforces per-point features at the original high-resolution scale, with shallow feature channels. This constraint limits the capacity to provide richer representations, which is crucial for SSL as evidenced by large channel dimensions in state-of-the-art 2D approaches [12]. Most importantly, decoding point clouds at the original scale unavoidably introduces local geometric cues into operators facilitating the geometric shortcut. + +In Fig. 4, we visualize PCA embeddings from different stages of PTv3 encoder and decoder, trained in a supervised manner for semantic segmentation. We observe that the encoder learns diverse features capturing spatial features at different scales in contrast to the decoder which produces task-specific higher-level representations. Additionally, as the spatial resolution of the point cloud decreases via the max-pooling stages, feature representations become less local. This resolution reduction fundamentally limits reliance on fine spatial information tied to point coordinates. + +These observations motivate us to remove the decoder during self-supervised learning for two main reasons. First and foremost, training directly with features at coarser point resolutions inherently restricts access to fine-grained spatial information, reducing the possibility of geometric shortcuts. Second, task-specific features can be probed or fine-tuned on top of a more expressive multi-scale point representation. + +# 4. Point Self-distillation with Sonata + +This section details the methodology of Sonata, the point self-distillation framework designed to address geometric shortcuts, remove structural constraints, and deliver strong linear probing results as discussed in Sec. 3. A roadmap of incremental ablation is illustrated in Fig. 6. + +# 4.1. Macro Framework + +Before diving into the specific micro designs, we begin with a point self-distillation framework derived from insights gained through previous efforts in point [88, 93] and image [11, 34, 60, 104, 110] self-supervised learning. This macro framework provides a solid foundation for pretext task design, allowing us to focus on addressing the geometric shortcut without additional concerns. + +In essence, (point) self-supervised learning aims to make things (points) that should be the same, the same (identical in representation). This forms the basic recipe of point SSL: generating two views of a given point cloud with random spatial (e.g., crop, rotate, distort) and photometric (e.g., jitter) augmentations, then matching and aligning the feature embeddings of points that are close in the original space. + +However, the superior robustness of self-supervised representation is rooted in a core principle: continuously increasing the difficulty of pretext tasks as long as the model continues to converge. This principle encourages enhancing the basic recipe with local-global view alignment and mask-unmask view alignment. Specifically, it involves aligning + +![](images/6cc7c5098377002cf7d37801254da98aad83255b8febd0aa13ed7539d6847ec5.jpg) +Figure 5. Self-distillation framework of Sonata. (1) Local views (bottom left) and global views (right) are generated with dedicated spatial and photometric augmentations, while masked views are created by randomly masking out grid-based patches from the global views (top left). (2) Embeddings from local and masked views are extracted by the student, with global views processed by the teacher (top). (3) Points from local and masked views are matched with corresponding points in the global views based on their original spatial distance, allowing for the distillation of embeddings from global views to local and masked views (bottom). + +neighboring points from local views generated by small cropping ratios and masked views created with large masking ratios (see Fig. 5 left), with unmasked global views that contain relatively richer information (see Fig. 5 right). The difficulty of pretext tasks can be scaled by adjusting the cropping and masking ratios. + +The more challenging the pretext tasks become, the higher the risk of model collapse. This instability calls for an asymmetric encoding approach, specifically exponential moving average (EMA) [34]: rather than encoding all views with a shared-weight model, the approach encodes challenging local and masked views with an actively learned student model, while using a stable teacher model, updated with a moving average of the student parameters, to encode global views (see Fig. 5 top). With the teacher preventing the student from being misled, the student is less likely to get lost in "mission impossible" and has a greater chance to discover treasure within "impossible" (i.e. extreme $5\%$ crop ratio for local views and $70\%$ mask ratio for masked views). + +In terms of SSL criteria, we move away from prior approaches that rely on contrastive and generative learning [88]. Contrastive learning, limited by the number of point pairs in pairwise similarity computations, restricts scalability. Generative learning, meanwhile, partially anchors representations to predefined cues, limiting the model's capacity to capture more generalizable features. Following DINov2, we adopt a self-distillation approach driven by Sinkhorn-Knopp centering [11], KoLeo regularization [72], and clustering assignments [11]. This adaptation initially exacerbates collapse into geometric shortcuts but holds greater potential for robust representation once this challenge is addressed. + +# 4.2. Micro Design + +We now discuss micro designs aimed at addressing geometric shortcuts. Since the problematic spatial information is inherently tied to point coordinates and directly used by operators, it is nearly impossible to mask out. This constraint defines the key strategies of our micro designs: obscuring spatial information and emphasizing input features. + +Decoder removal. Previous approaches adhere to the original U-Net-style backbone for feature extraction; however, our observations in Sec. 3 motivate us to remove the complex hierarchical decoder and perform self-distillation directly using the encoder's output. This simple adjustment is key in our fight against geometric shortcuts, as it: (a) increases the feature channels participating in self-distillation (from 96 to 512), (b) streamlines the pipeline by involving fewer points in the pretext task after hierarchical pooling, and, most importantly, (c) introduces a natural way to obscure geometric cues: the positional information of points becomes naturally disturbed during hierarchical encoding and pooling. In fact, this removal proves to be even more beneficial than expected, boosting the linear probing result from $20.7\%$ to $60.4\%$ . + +Feature up-casting. While the removal effectively prevents over-reliance on naive geometric cues, it also introduces certain limitations, particularly in leveraging multiscale contexts. In the original U-Net structure, the decoder plays a key role in progressively aggregating features across scales to reconstruct semantic details. Without this process, self-distillation struggles to capture multi-scale spatial information and broader contextual relationships. To retain multi-scale features, we introduce a parameter-free feature up-casting process similar to hypercolumns in image segmentation [33]: progressively up-casting features back to the scale of the previous encoding stage, with the mapping relationships preserved through pooling layers and concatenation with features from the prior encoding stage. This approach provides richer, coarse-to-fine features from multiscale encoding. While it does increase the risk of the model falling into geometric shortcuts, finding the right balance is key. Our ablation study shows that up-casting features twice achieve the best performance. + +Masked points jitter. A slight random Gaussian jitter $(\sigma = 0.005)$ is applied to point coordinates as part of data augmentation. However, specifically for points to be masked, we additionally apply a stronger Gaussian jitter $(\sigma = 0.01)$ to further disrupt their spatial relationships. We pay special attention to these masked points because models are more likely to collapse to naive geometric cues, especially when point features are masked, making it difficult to derive solutions from neighboring unmasked points. + +Progressive parameter scheduler. Geometric shortcuts are like traps on the convergence path of point cloud self-supervised learning. Similarly, we can set our own "trap" + +![](images/8a03d2c187c4b70f274a3c5da3bebb8c98825d9400e2e0661479788c0b34228e.jpg) +Figure 6. The roadmap. We evolve Mask Scene Contrast [88] into our Sonata by modernizing self-supervised learning with self-distillation, addressing the geometric shortcut, and scaling up training. Our designs are validated through progressive ablation with linear and decoder probing on ScanNet semantic segmentation [23]. Starting with $23\mathrm{k}$ training data (a combination of ScanNet and Structured3D [109]) and a $39\mathrm{M}$ PTv3 model [89], we ultimately scale up to $140\mathrm{k}$ assets (Tab. 1) and a $108\mathrm{M}$ PTv3 model. + +during training. For example, rather than starting with a challenging large mask size and mask ratio, we begin with a relatively small mask size (10 cm) and mask ratio (30%), gradually increasing them to 40 cm (0.1 $\Rightarrow$ 0.4) and 70% (0.3 $\Rightarrow$ 0.7) over the first 5% of the training process. This strategy encourages the model to rely more on input features to solve the pre-defined pretext tasks, preventing it from shifting its reliance to point coordinates as training difficulty increases. This approach aligns with curriculum learning [7], progressively challenging the model as it adapts. Similarly, in addition to the common learning rate scheduler, we also implement custom progressive schedulers for teacher temperature (0.04 $\Rightarrow$ 0.07) and weight decay (0.04 $\Rightarrow$ 0.2). We found that this design pushes these parameters to more extreme levels previously unexplored, further enhancing model performance. + +# 4.3. Implementation and Evaluation Protocols + +In this section, we introduce the implementation details and the evaluation protocols for our experiments. + +Backbone. We build our Sonata with Point Transformer V3 (PTv3) [89] and refer to Pointcept [19] for details of imple + +
DatasetSourceTrainValTestTotal
ScanNet [23]real1,2013121001,613
ScanNet++ [101]real7121781261,016
S3DIS [1]real204680272
ArkitScenes [5]real4,49854905,047
HM3D [68]real8,8811,119010,000
Structured3D [109]sim.18,3481,7761,69721,821
ASE [3]sim.90,00010,0000100,000
Sonata (ours)mixed123,84414,0021,923139,769
+ +Table 1. Data source collection. + +
MethodRealSimTotalMultiplier
PC [93]1,61301,613×1
MSC [88]6,66006,660×4.1
PPT [90]1,88521,82123,706×14.7
Sonata (ours)17,948121,821139,768×86.7
+ +Table 2. Data scale comparison. + +mentation. Building on this, we made an additional adjustment to enhance scalability: replacing all Batch Normalization (BN) [41] layers with Layer Normalization (LN) [4]. Although this replacement results in some initial accuracy degradation, it enhances domain adaptation by eliminating the need for additional domain-specific adjustments when scaling up with multi-dataset joint training [90]. Along with the scaling up of data, we also scale up the encoder block depths from $[2, 2, 2, 6, 2]$ to $[3, 3, 3, 12, 3]$ and widths from $[32, 64, 128, 256, 512]$ to $[48, 96, 192, 384, 512]$ . This PTv3 model has $108\mathrm{M}$ parameters. + +Data. We extend the multi-dataset joint training approach introduced by PPT [90], further expanding the data scale by removing the constraint of human labeling through unsupervised learning. This results in a collection of 140k scene-level point clouds from both real-world and simulated environments (outlined in Tab. 1), making it $86.7 \times$ larger than the data scale of PointContrast [93] and $5.9 \times$ larger than the data collection of PPT, as detailed in Tab. 2. + +Training. We train Sonata on the 140k data collection for 200 epochs, using the AdamW optimizer [54] with a batch size of 96, distributed across 32 GPUs. The learning rate linearly warms up over the first 10 epochs to a base value of 0.004, then decays following a cosine schedule [55]. Additionally, a layer-wise learning rate decay of 0.9 is applied to model parameters [111]. Weight decay is also controlled by a cosine schedule, progressively increasing from 0.04 to 0.2. For EMA, the student temperature is set to 0.1, while the teacher temperature gradually rises from 0.04 to 0.07 during the first 10 epochs [60]. The momentum starts at 0.994 and increases to 1 by the final iteration. For data augmentation and view generation, we follow the augmentation pipeline designed by MSC [88]. We generate 2 global views (sampling $40\%$ to $100\%$ of scene points) and 4 local views (sampling $5\%$ to $40\%$ of scene points) for training, with 2 masked views generated based on the global views. + +![](images/06aa10040be016adeacb37fd5767052dffa88a788de5bc8d35dfdbf31a2a7d1a.jpg) + +![](images/71d19533408976e6644598cdf508d18a7c5e44b67d330f9724f037577f5ccdcb.jpg) + +![](images/dcb26bf0cf2ce76820349c2c9376b5eed8bb8c499453a344a54c3494c4ab1d93.jpg) + +![](images/82222571f32c270eb9e8c30df905551ea5b2080d6d6f0e6479638834f53e2866.jpg) + +![](images/44a957cddc13afa6e27b5da41df947c9bf978364d7f07cc157fac6052703723a.jpg) +Figure 7. Zero-shot comparison with DINOv2. We compare the PCA visualizations of DINOv2, Sonata, and their combined feature representation. DINOv2 excels at capturing photometric details, while Sonata better distinguishes spatial information. The combined model demonstrates improved coherence and detail, showcasing the complementary strengths of both models. + +![](images/aa552af8ddd0526f922499b890339fd7daffb2482750d8c8e5ac2f10a3aa2d92.jpg) + +Evaluation. We evaluate the quality of the learned representation using the following three protocols after initializing the encoder with Sonata: + +- In linear probing, we keep the encoder frozen and upcast the features to their original scale. A single linear layer, comprising less than $0.2\%$ of the total parameters, is then used to adapt these features to downstream tasks. +- In decoder probing, we take a step back and reintroduce a lightweight hierarchical decoder, which accounts for $13\%$ of the total parameters. We then freeze the encoder, allowing only the decoder to actively learn. +- In full fine-tuning, we follow the traditional approach by unfreezing the entire PTv3 U-Net-style backbone, tuning the learned representation to downstream tasks. + +We advocate linear probing as the primary evaluation criterion for point SSL, considering other methods as intermediate steps. We look forward to the day when a linear-probed self-supervised model outperforms a fully fine-tuned one. + +# 5. Main Results + +We validate the reliability of the Sonata representation using the evaluation protocols discussed in Sec. 4.3 and analyze the main properties based on these results. + +Comparison with image self-supervised model. In Tab. 3, we compare the linear probing and decoder probing results on ScanNet and ScanNet200 semantic segmentation with the linear probing accuracy of representations transferred from the image self-supervised models DINOv2 [60] and DINOv2.5 [24]. Specifically, we aggregate unprojected pixel embeddings using ground truth camera poses and depth calculated by ray intersection with a reconstructed mesh, which provides more accuracy than sensor per-frame depth. Our results indicate that while the DINOs' represent + +
2D × 3DScanNet Val [23]ScanNet200 Val [23]
MethodsmIoUmAccallAccmIoUmAccallAcc
• DINOv2 (lin.) [60]63.0975.5082.4227.4237.5972.80
• DINOv2.5 (lin.) [24]63.3675.9482.3027.7539.2372.53
• Sonata (lin.)72.5283.1189.7429.2541.6181.15
+DINOv2 (lin.)75.9185.3691.2536.6746.9882.85
+DINOv2.5 (lin.)76.4485.6891.3336.9648.2382.77
• Sonata (dec.)79.0786.5792.6833.5444.4884.07
+DINOv2 (dec.)79.1287.2392.4737.7349.3883.31
+DINOv2.5 (dec.)79.1986.6692.5038.2748.5783.77
+ +Table 3. Numerical comparison with DINO series. + +
Data EfficiencyLimited Scenes (Pct.)Limited Annotation (Pts.)
Methods1%5%10%20%Full2050100200Full
○ SparseUNet [17]26.047.856.762.972.241.953.962.265.572.2
• CSC [38]28.949.859.464.673.855.560.565.968.273.8
• MSC [88]29.250.761.064.975.461.065.668.969.675.4
○ PTv2 [87]24.848.159.866.375.458.466.170.371.275.4
○ PTv3 [89]25.848.961.067.077.260.167.971.472.777.2
• PPT [90] (sup.)31.152.663.368.278.262.469.174.375.578.2
• Sonata (lin.)43.662.568.669.872.569.070.571.171.572.5
• Sonata (dec.)44.564.169.872.579.169.873.175.076.379.1
• Sonata (full)45.365.772.472.879.470.573.676.077.079.4
+ +Table 4. Data efficiency. + +
Param. Efficiency MethodsParamsScanNet Val [23]ScanNet200 Val [71]ScanNet++ Val [101]S3DIS Area 5 [1]S3DIS 6-fold [1]
Learn.Pct.mIoUmAccallAccmIoUmAccallAccmIoUmAccallAccmIoUmAccallAccmIoUmAccallAcc
○ SparseUNet [17]39.2M100%72.380.290.025.032.980.428.838.480.166.372.589.872.480.989.9
• PC [93] (lin.)<0.2M<0.1%5.69.750.00.50.940.31.83.146.411.418.652.311.719.051.2
• CSC [38] (lin.)<0.2M<0.1%12.618.164.21.32.153.02.84.553.624.432.066.424.932.566.9
• MSC [88] (lin.)<0.2M<0.1%14.120.362.91.52.553.64.56.661.327.935.571.129.937.971.3
○PTv3 [89]124.8M100%77.685.092.035.346.083.442.153.485.673.478.991.777.785.391.5
• MSC [88] (lin.)<0.2M<0.2%21.832.265.53.35.557.58.111.964.732.142.470.934.646.071.3
• Sonata (lin.)<0.2M<0.2%72.583.189.729.341.681.237.350.984.372.381.290.976.587.490.8
• Sonata (dec.)16.3M13%79.186.692.733.544.584.140.952.686.374.580.492.681.588.893.0
+ +Table 5. Parameter efficiency. + +tation demonstrates impressive robustness, Sonata offers a more suitable representation for 3D tasks, achieving $+9.2\%$ on ScanNet and $+1.5\%$ on ScanNet200 for semantic segmentation. Furthermore, combining Sonata with the DINOs yields higher accuracy than any single data modality alone $(+3.9\%$ and $+7.7\%$ , respectively), underscoring the promising potential of cross-modal self-distillation. Additional zero-shot comparisons (see Fig. 7) through PCA visualizations further corroborate these numerical findings. + +Data efficiency. In Tab. 4, we present the semantic segmentation performance of Sonata when probed or fine-tuned on the ScanNet dataset with limited scenes and annotations [38]. The results demonstrate the exceptional data efficiency of Sonata, with improvements of $19.5\%$ in extreme data scarcity (1% of scenes) and $10.4\%$ in limited annotation scenarios (20 points per scene) compared to training from scratch. Notably, even linear probing surpasses previous SOTA by a substantial margin (12.5% with 1% scenes), highlighting Sonata's reliability in low-data scenarios. + +Parameter efficiency. In Tab. 5, we demonstrate parameter efficiency using both linear and decoder probing across various indoor semantic segmentation benchmarks, including ScanNet [23], ScanNet200 [71], ScanNet++ [101], S3DIS Area5[1], and S3DIS 6-fold cross-validation [65]. Semantic segmentation is emphasized as it provides a direct measure of point cloud representation quality. Results show that a single linear layer with a negligible number of parameters (<0.02% of total parameters) is sufficient for Sonata to achieve strong performance on these benchmarks (e.g., + +72.5% on ScanNet and 73.4% on S3DIS Area5). Furthermore, probing a decoder with only 13% of the model's parameters yields even higher accuracy (e.g., 79.1% on ScanNet and 81.5% on S3DIS 6-fold cross-validation). However, while decoder probing achieves SOTA results on ScanNet (20 classes) and S3DIS (13 classes), performance on ScanNet200 (200 classes) and ScanNet++ (100 classes) remains limited. This shows a limitation of the learned representation in distinguishing a large number of classes. + +Indoor semantic segmentation. In Tab. 6, we further enhance Sonata's semantic segmentation accuracy through full fine-tuning, consistently pushing SOTA results to new heights across the five widely recognized benchmarks, e.g., achieving $79.4\%$ on ScanNet and $82.3\%$ on S3DIS 6-fold cross-validation. However, we view full fine-tuning as an intermediate step toward a future where linear probing surpasses it. At present, full fine-tuning remains essential to achieve the highest performance on these benchmarks and close the remaining gap of $7.1\%$ and $5.8\%$ respectively. + +Indoor instance segmentation. In Tab. 7, We also validate the robustness of Sonata representation on indoor instance segmentation benchmarks, including ScanNet [23], ScanNet200 [71], ScanNet++ [101], and S3DIS [1]. Consistent with our findings in semantic segmentation, Sonata demonstrates strong parameter efficiency, achieving significant improvements with linear probing (10× mAP50 on ScanNet and 21× on ScanNet200) and decoder probing (12× mAP50 on ScanNet and 33× on ScanNet200). Full fine-tuning further boosts these results, achieving SOTA + +
Indoor Sem. SegParamsScanNet Val [23]ScanNet200 Val [71]ScanNet++ Val [101]S3DIS Area 5 [1]S3DIS 6-fold [1]
MethodsLearn.Pct.mIoUmAccallAccmIoUmAccallAccmIoUmAccallAccmIoUmAccallAccmIoUmAccallAcc
○ SparseUNet [17]39.2M100%72.380.290.025.032.980.428.838.480.166.372.589.872.480.989.9
• PC [93]39.2M100%72.380.990.126.233.079.929.239.782.768.173.590.074.783.390.6
• CSC [38]39.2M100%72.881.090.726.933.780.632.541.183.770.776.490.875.584.090.9
• MSC [88]39.2M100%75.783.491.332.041.682.339.449.684.970.776.191.077.485.391.5
○PTv3 [89]124.8M100%77.685.092.035.346.083.442.153.485.673.478.991.777.785.391.5
• MSC [88]124.8M100%78.285.392.233.443.783.442.453.685.969.974.991.277.484.791.7
• PPT [90] (sup.)124.8M100%78.685.992.336.046.283.843.355.786.474.380.192.080.887.792.6
• Sonata124.8M100%79.486.192.536.846.584.443.755.886.676.081.693.082.389.993.3
+ +Table 6. Indoor semantic segmentation. + +
Indoor Ins. SegParamsScanNet Val [23]ScanNet200 Val [71]ScanNet++ Val [101]S3DIS Area 5 [1]
MethodsLearn.Pct.mAP25mAP50mAPmAP25mAP50mAPmAP25mAP50mAPmAP25mAP50mAP
○PointGroup [45]124.8M100%77.561.740.940.133.223.136.730.720.955.749.437.8
•MSC (lin.)<0.2M<0.2%13.35.32.32.31.00.44.82.61.319.013.09.7
•Sonata (lin.)<0.2M<0.2%72.653.930.730.921.310.931.622.412.245.836.626.1
•Sonata (dec.)16.3M13%76.862.840.840.833.322.838.129.118.863.757.145.1
•MSC (full)124.8M100%78.462.941.140.533.823.438.930.921.756.350.538.1
•PPT [90] (sup.)124.8M100%78.963.542.140.834.124.039.332.821.957.551.239.7
•Sonata (full)124.8M100%79.263.942.442.135.625.440.333.622.363.857.445.5
+ +Table 7. Indoor instance segmentation. + +
Outdoor Sem. Seg.nuScenes Val [9]Waymo Val [74]Sem.KITTI Val [6]
MethodsmIoUmAccallAccmIoUmAccallAccmIoUmAcc allAcc
o PTv3 [89]80.487.294.771.380.594.769.176.192.6
• Sonata (lin.)66.177.292.460.572.592.562.072.591.0
• Sonata (dec.)77.385.994.270.878.894.368.476.592.3
• PPT [90] (sup.)81.287.594.872.181.394.872.377.593.4
• Sonata (full)81.787.995.072.981.994.972.677.993.4
+ +Table 8. Outdoor semantic segmentation. + +pre-training performance across benchmarks. This clearly demonstrates that, unlike previous SSL approaches, the Sonata representation encodes instance-level information. + +Outdoor semantic segmentation. In Tab. 8, we adapt pre-training paradigm of Sonata to outdoor LiDAR scenarios through joint training on nuScenes [9], Waymo [74], and SemanticKITTI [6] and evaluate semantic segmentation performance using our evaluation protocols. With linear probing, Sonata sets a robust parameter efficiency baseline. Decoder probing achieves significant gains. In full fine-tuning, Sonata surpasses the supervised PPT [90], establishing new SOTA mIoU scores of 81.7, 72.9, and 72.6 across these benchmarks, underscoring the effectiveness of Sonata in outdoor perception tasks. Note that most of that performance can be recovered by more efficient decoder-only probing with a $95\%$ , $97\%$ , and $94\%$ of full fine-tuning. + +# 6. Conclusion and Discussion + +In this work, we make progress towards self-supervised learning of a strong and reliable 3D point representation, that can zero-shot correspond semantically similar 3D points, down to the instance level, even under the presence + +of spatial and visual perturbations. We demonstrate that such a representation can serve as the foundation for 3D tasks in semantic and instance-level grouping. We find existing 3D self-supervised learning approaches lacking, and hypothesize that this is due to what we term the geometric shortcut, a problem unique to 3D, which causes representations to collapse to low-level spatial features. We demonstrate the deficiency of these collapsed representations through linear probing. Beginning with a point self-distillation framework, we tackle the geometric shortcut by attaching SSL losses at coarser spatial scales, disturbing the spatial information of masked points with no features, and progressively increasing task difficulty to prevent overreliance on accessible geometric cues. This change enables effective scaling up, ultimately composing a Sonata from 140k point clouds. Sonata demonstrates semantically meaningful zero-shot visualization, as well as exceptional parameter and data efficiency. Full fine-tuning further advances SOTA across 3D indoor and outdoor perception tasks. + +We summarize our limitations and future works and encourage readers to refer to the supplementary for details: 1. Sonata representations could be further semantically enriched through augmentation with 1M object-level point clouds [53]; 2. Pursuing unified training across indoor and outdoor scenarios as a feasible and promising direction; 3. Scaling up training with pixel-aligned [8] and SLAM-generated [27] point clouds from video data; 4. Leveraging cross-modal distillation to allow Sonata and DINOV2 to complement each other. We hope our insights will contribute to future research in 3D representation learning. + +# Acknowledgements + +We extend our gratitude to Maxime Oquab and Piotr Bojanowski for their guidance on the DINOv2 training details, to Saining Xie for insightful discussions on the vision of 3D representation learning, and to Paul Mcvay for his thoughts on the JPEA framework with sparse point cloud data. + +# References + +[1] Iro Armeni, Ozan Sener, Amir R. Zamir, Helen Jiang, Ioannis Brilakis, Martin Fischer, and Silvio Savarese. 3d semantic parsing of large-scale indoor spaces. In CVPR, 2016. 2, 6, 7, 8, 16 +[2] Mahmoud Assran, Quentin Duval, Ishan Misra, Piotr Bojanowski, Pascal Vincent, Michael Rabbat, Yann LeCun, and Nicolas Ballas. Self-supervised learning from images with a joint-embedding predictive architecture. In CVPR, 2023. 2 +[3] Armen Avetisyan, Christopher Xie, Henry Howard-Jenkins, Tsun-Yi Yang, Samir Aroudj, Suvam Patra, Fuyang Zhang, Duncan Frost, Luke Holland, Campbell Orme, et al. Scenescript: Reconstructing scenes with an autoregressive structured language model. In ECCV, 2024. 2, 6, 16 +[4] Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. Stat, 2016. 6 +[5] Gilad Baruch, Zhuoyuan Chen, Afshin Dehghan, Tal Dimry, Yuri Feigin, Peter Fu, Thomas Gebauer, Brandon Joffe, Daniel Kurz, Arik Schwartz, and Elad Shulman. ARKitscenes - a diverse real-world dataset for 3d indoor scene understanding using mobile RGB-d data. In NeurIPSW, 2021. 2, 6 +[6] Jens Behley, Martin Garbade, Andres Milioto, Jan Quenzel, Sven Behnke, Cyril Stachniss, and Jurgen Gall. Semantickitti: A dataset for semantic scene understanding of lidar sequences. In ICCV, 2019. 8 +[7] Yoshua Bengio, Jérôme Louradour, Ronan Collobert, and Jason Weston. Curriculum learning. In ICML, 2009. 5 +[8] Aleksei Bochkovskii, Amael Delaunoy, Hugo Germain, Marcel Santos, Yichao Zhou, Stephan R Richter, and Vladlen Koltun. Depth pro: Sharp monocular metric depth in less than a second. arXiv preprint arXiv:2410.02073, 2024. 8, 13 +[9] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In CVPR, 2020. 2, 8 +[10] Mathilde Caron, Piotr Bojanowski, Armand Joulin, and Matthijs Douze. Deep clustering for unsupervised learning of visual features. In ECCV, 2018. 2 +[11] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. In NeurIPS, 2020. 2, 4 +[12] Mathilde Caron, Hugo Touvron, Ishan Misra, Herve Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. + +Emerging properties in self-supervised vision transformers. In CVPR, 2021. 2, 4 +[13] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In ICML, 2020. 2 +[14] Xi Chen, Lianghua Huang, Yu Liu, Yujun Shen, Deli Zhao, and Hengshuang Zhao. Anydoor: Zero-shot object-level image customization. In CVPR, 2024. 2 +[15] Yukang Chen, Jianhui Liu, Xiangyu Zhang, Xiaojuan Qi, and Jiaya Jia. Largekernel3d: Scaling up kernels in 3d sparse cnns. In CVPR, 2023. 16 +[16] Hung-Yueh Chiang, Yen-Liang Lin, Yueh-Cheng Liu, and Winston H Hsu. A unified point-based framework for 3d segmentation. In 3DV, 2019. 16 +[17] Christopher Choy, JunYoung Gwak, and Silvio Savarese. 4d spatio-temporal convnets: Minkowski convolutional neural networks. In CVPR, 2019. 3, 7, 8, 16, 17 +[18] Christopher Choy, Jaesik Park, and Vladlen Koltun. Fully convolutional geometric features. In ICCV, 2019. 2 +[19] Pointcept Contributors. Pointcept: A codebase for point cloud perception research. https://github.com/Pointcept/Pointcept, 2023.5 +[20] Spconv Contributors. Spconv: Spatially sparse convolution library. https://github.com/traveller59/spconv, 2022.3 +[21] Brian Curless and Marc Levoy. A volumetric method for building complex models from range images. In Proceedings of the 23rd annual conference on Computer graphics and interactive techniques, pages 303-312, 1996. 16 +[22] Angela Dai and Matthias Nießner. 3dmv: Joint 3d-multiview prediction for 3d semantic scene segmentation. In ECCV, 2018. 16 +[23] Angela Dai, Angel X. Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In CVPR, 2017. 2, 5, 6, 7, 8, 16 +[24] Timothee Darcet, Maxime Oquab, Julien Mairal, and Piotr Bojanowski. Vision transformers need registers. arXiv:2309.16588, 2023. 6, 7 +[25] Carl Doersch, Abhinav Gupta, and Alexei A Efros. Unsupervised visual representation learning by context prediction. In ICCV, 2015. 2 +[26] Alexey Dosovitskiy, Philipp Fischer, Jost Tobias Springenberg, Martin Riedmiller, and Thomas Brox. Discriminative unsupervised feature learning with exemplar convolutional neural networks. TPAMI, 2015. 2 +[27] Jakob Engel, Kiran Somasundaram, Michael Goesele, Albert Sun, Alexander Gamino, Andrew Turner, Arjang Talattof, Arnie Yuan, Bilal Souti, Brighid Meredith, et al. Project aria: A new tool for egocentric multi-modal ai research. arXiv:2308.13561, 2023. 2, 8, 13 +[28] Pedro F Felzenszwalb and Daniel P Huttenlocher. Efficient graph-based image segmentation. IJCV, 2004. 3 +[29] Benjamin Graham and Laurens van der Maaten. Submanifold sparse convolutional networks. arXiv:1706.01307, 2017. 3 + +[30] Benjamin Graham, Martin Engelcke, and Laurens van der Maaten. 3d semantic segmentation with submanifold sparse convolutional networks. In CVPR, 2018. 16 +[31] Jean-Bastien Grill, Florian Strub, Florent Alché, Coretin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. In NerIPS, 2020. 2 +[32] Jiayuan Gu, Fanbo Xiang, Xuanlin Li, Zhan Ling, Xiqiang Liu, Tongzhou Mu, Yihe Tang, Stone Tao, Xinyue Wei, Yunchao Yao, Xiaodi Yuan, Pengwei Xie, Zhiao Huang, Rui Chen, and Hao Su. Maniskill2: A unified benchmark for generalizable manipulation skills. In ICLR, 2023. 2 +[33] Bharath Hariharan, Pablo Arbelaez, Ross Girshick, and Jitendra Malik. Hypercolumns for object segmentation and fine-grained localization. In CVPR, 2015. 5 +[34] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In CVPR, 2020. 2, 3, 4 +[35] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In CVPR, 2022. 2 +[36] R. Devon Hjelm, Alex Fedorov, Samuel Lavoie-Marchildon, Karan Grewal, Phil Bachman, Adam Trischler, and Yoshua Bengio. Learning deep representations by mutual information estimation and maximization. In ICLR, 2018. 2 +[37] Yicong Hong, Kai Zhang, Jiuming Gu, Sai Bi, Yang Zhou, Difan Liu, Feng Liu, Kalyan Sunkavalli, Trung Bui, and Hao Tan. Lrm: Large reconstruction model for single image to 3d. arXiv preprint arXiv:2311.04400, 2023. 16 +[38] Ji Hou, Benjamin Graham, Matthias Nießner, and Saining Xie. Exploring data-efficient 3d scene understanding with contrastive scene contexts. In CVPR, 2021. 2, 3, 7, 8, 16, 17 +[39] Qingyong Hu, Bo Yang, Linhai Xie, Stefano Rosa, Yulan Guo, Zhihua Wang, Niki Trigoni, and Andrew Markham. Randla-net: Efficient semantic segmentation of large-scale point clouds. In CVPR, 2020. 16 +[40] Zeyu Hu, Mingmin Zhen, Xuyang Bai, Hongbo Fu, and Chiew-lan Tai. Jsenet: Joint semantic segmentation and edge detection network for 3d point clouds. In ECCV, 2020. 16 +[41] Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In ICML, 2015. 6 +[42] Ayush Jain, Pushkal Katara, Nikolaos Gkanatsios, Adam W. Harley, Gabriel Sarch, Kriti Aggarwal, Vishrav Chaudhary, and Katerina Fragkiadaki. Odin: A single model for 2d and 3d perception. In CVPR, 2024. 16 +[43] Stephen James, Zicong Ma, David Rovick Arrojo, and Andrew J. Davison. Rlbench: The robot learning benchmark & learning environment. RAL, 2020. 2 +[44] Li Jiang, Hengshuang Zhao, Shu Liu, Xiaoyong Shen, Chi-Wing Fu, and Jiaya Jia. Hierarchical point-edge interaction network for point cloud semantic segmentation. In ICCV, 2019. 17 + +[45] Li Jiang, Hengshuang Zhao, Shaoshuai Shi, Shu Liu, Chi-Wing Fu, and Jiaya Jia. Pointgroup: Dual-set point grouping for 3d instance segmentation. CVPR, 2020. 8 +[46] Maxim Kolodiazhnyi, Anna Vorontsova, Anton Konushin, and Danila Rukhovich. Oneformer3d: One transformer for unified point cloud segmentation. In CVPR, 2024. 16 +[47] Xin Lai, Jianhui Liu, Li Jiang, Liwei Wang, Hengshuang Zhao, Shu Liu, Xiaojuan Qi, and Jiaya Jia. Stratified transformer for 3d point cloud segmentation. In CVPR, 2022. 16, 17 +[48] Loic Landrieu and Martin Simonovsky. Large-scale point cloud semantic segmentation with superpoint graphs. In CVPR, 2018. 17 +[49] Huan Lei, Naveed Akhtar, and Ajmal Mian. Seggen: Efficient 3d point cloud segmentation with fuzzy spherical kernel. In CVPR, 2020. 16, 17 +[50] Yangyan Li, Rui Bu, Mingchao Sun, Wei Wu, Xinhan Di, and Baoquan Chen. Pointconn: Convolution on x-transformed points. NeurIPS, 2018. 16, 17 +[51] Dingkang Liang, Xin Zhou, Wei Xu, Xingkui Zhu, Zhikang Zou, Xiaqing Ye, Xiao Tan, and Xiang Bai. Pointmamba: A simple state space model for point cloud analysis. In Advances in Neural Information Processing Systems, 2024. 3 +[52] Haojia Lin, Xiawu Zheng, Lijiang Li, Fei Chao, Shanshan Wang, Yan Wang, Yonghong Tian, and Rongrong Ji. Meta architecture for point cloud analysis. In CVPR, pages 17682-17691, 2023. 16, 17 +[53] Minghua Liu, Ruoxi Shi, Kaiming Kuang, Yinhao Zhu, Xu-anlin Li, Shizhong Han, Hong Cai, Fatih Porikli, and Hao Su. Openshape: Scaling up 3d shape representation towards open-world understanding. In NeurIPS, 2023. 8, 13 +[54] I Loshchilov. Decoupled weight decay regularization. In ICLR, 2019. 6 +[55] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. In *ICLR*, 2017. 6 +[56] Xu Ma, Can Qin, Haoxuan You, Haoxi Ran, and Yun Fu. Rethinking network design and local geometry in point cloud: A simple residual mlp framework. *ICLR*, 2022. 3 +[57] Gaku Narita, Takashi Seno, Tomoya Ishikawa, and Yohsuke Kaji. Panopticfusion: Online volumetric semantic mapping at the level of stuff and things. In IROS, 2019. 16 +[58] Richard A Newcombe, Shahram Izadi, Otmar Hilliges, David Molyneaux, David Kim, Andrew J Davison, Pushmeet Kohi, Jamie Shotton, Steve Hodges, and Andrew Fitzgibbon. Kinectfusion: Real-time dense surface mapping and tracking. In 2011 10th IEEE international symposium on mixed and augmented reality, pages 127-136. IEEE, 2011. 16 +[59] Mehdi Noroozi and Paolo Favaro. Unsupervised learning of visual representations by solving jigsaw puzzles. In ECCV, 2016. 2 +[60] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Russell Howes, Po-Yao Huang, Hu Xu, Vasu Sharma, Shang-Wen Li, Wojciech Galuba, Mike Rabbat, Mido Assan, Nicolas Ballas, Gabriel Synnaeve, Ishan Misra, Herve + +Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. Dinov2: Learning robust visual features without supervision. TMLR, 2024. 2, 3, 4, 6, 7, 13 +[61] Yatian Pang, Wenxiao Wang, Francis EH Tay, Wei Liu, Yonghong Tian, and Li Yuan. Masked autoencoders for point cloud self-supervised learning. In ECCV, 2022. 2 +[62] Chunghyun Park, Yoonwoo Jeong, Minsu Cho, and Jaesik Park. Fast point transformer. In CVPR, 2022. 16 +[63] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. NeurIPS, 2019. 16 +[64] Bohao Peng, Xiaoyang Wu, Li Jiang, Yukang Chen, Hengshuang Zhao, Zhuotao Tian, and Jiaya Jia. Oa-cnns: Omni-adaptive sparse cnns for 3d semantic segmentation. In CVPR, 2024. 3 +[65] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In CVPR, 2017. 7, 17 +[66] Charles R Qi, Li Yi, Hao Su, and Leonidas J Guibas. Point-net++: Deep hierarchical feature learning on point sets in a metric space. In NeurIPS, 2017. 3, 16, 17 +[67] Guocheng Qian, Yuchen Li, Houwen Peng, Jinjie Mai, Hasan Hammoud, Mohamed Elhoseiny, and Bernard Ghanem. Pointnext: Revisiting pointnet++ with improved training and scaling strategies. In NeurIPS, 2022. 3, 16, 17 +[68] Santhosh Kumar Ramakrishnan, Aaron Gokaslan, Erik Wijmans, Oleksandr Maksymets, Alexander Clegg, John M Turner, Eric Undersander, Wojciech Galuba, Andrew Westbury, Angel X Chang, Manolis Savva, Yili Zhao, and Dhruv Batra. Habitat-matterport 3d dataset (HM3d): 1000 large-scale 3d environments for embodied AI. In NeurIPS, 2021. 2, 6, 14, 15 +[69] Damien Robert, Hugo Raguet, and Loic Landrieu. Efficient 3d semantic segmentation with superpoint transformer. In ICCV, 2023. 17 +[70] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In MICCAI, 2015. 3 +[71] David Rozenberszki, Or Litany, and Angela Dai. Language-grounded indoor 3d semantic segmentation in the wild. In ECCV, 2022. 7, 8 +[72] Alexandre Sablayrolles, Matthijs Douze, Cordelia Schmid, and Hervé Jégou. Spreading vectors for similarity search. In ICLR, 2019. 4 +[73] Julian Straub, Daniel DeTone, Tianwei Shen, Nan Yang, Chris Sweeney, and Richard Newcombe. Efm3d: A benchmark for measuring progress towards 3d egocentric foundation models. arXiv:2406.10224, 2024. 2, 14, 15, 16 +[74] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, et al. Scalability in perception for autonomous driving: Waymo open dataset. In CVPR, 2020. 2, 8 +[75] Maxim Tatarchenko, Jaesik Park, Vladlen Koltun, and Qian-Yi Zhou. Tangent convolutions for dense prediction in 3d. In CVPR, 2018. 17 + +[76] Lyne Tchapmi, Christopher Choy, Iro Armeni, JunYoung Gwak, and Silvio Savarese. Segcloud: Semantic segmentation of 3d point clouds. In 3DV, 2017. 17 +[77] Hugues Thomas, Charles R Qi, Jean-Emmanuel Deschaud, Beatrix Marcotegui, François Goulette, and Leonidas J Guibas. Kpconv: Flexible and deformable convolution for point clouds. In ICCV, 2019. 3, 16, 17 +[78] Hugues Thomas, Yao-Hung Hubert Tsai, Timothy D Barfoot, and Jian Zhang. Kpconvx: Modernizing kernel point convolution with kernel attention. In CVPR, 2024. 16 +[79] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 16 +[80] Chengyao Wang, Li Jiang, Xiaoyang Wu, Zhuotao Tian, Bohao Peng, Hengshuang Zhao, and Jiaya Jia. Groupcontrast: Semantic-aware self-supervised representation learning for 3d understanding. In CVPR, 2024. 2 +[81] Lei Wang, Yuchun Huang, Yaolin Hou, Shenman Zhang, and Jie Shan. Graph attention convolution for point cloud semantic segmentation. In CVPR, 2019. 17 +[82] Peng-Shuai Wang. Octformer: Octree-based transformers for 3D point clouds. SIGGRAPH, 2023. 16 +[83] Shenlong Wang, Simon Suo, Wei-Chiu Ma, Andrei Pokrovsky, and Raquel Urtasun. Deep parametric continuous convolutional neural networks. In CVPR, 2018. 17 +[84] Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. Dust3r: Geometric 3d vision made easy. In CVPR, 2024. 2, 13 +[85] Wenxuan Wu, Zhongang Qi, and Li Fuxin. Pointconv: Deep convolutional networks on 3d point clouds. In CVPR, 2019. 16 +[86] Wenxuan Wu, Li Fuxin, and Qi Shan. Pointconvformer: Revenge of the point-based convolution. In CVPR, 2023. 16 +[87] Xiaoyang Wu, Yixing Lao, Li Jiang, Xihui Liu, and Hengshuang Zhao. Point transformer v2: Grouped vector attention and partition-based pooling. In NeurIPS, 2022. 7, 16, 17 +[88] Xiaoyang Wu, Xin Wen, Xihui Liu, and Hengshuang Zhao. Masked scene contrast: A scalable framework for unsupervised 3d representation learning. In CVPR, 2023. 2, 3, 4, 5, 6, 7, 8, 14, 16, 17 +[89] Xiaoyang Wu, Li Jiang, Peng-Shuai Wang, Zhijian Liu, Xihui Liu, Yu Qiao, Wanli Ouyang, Tong He, and Hengshuang Zhao. Point transformer v3: Simpler, faster, stronger. In CVPR, 2024. 3, 5, 7, 8, 16, 17 +[90] Xiaoyang Wu, Zhuotao Tian, Xin Wen, Bohao Peng, Xihui Liu, Kaicheng Yu, and Hengshuang Zhao. Towards large-scale 3d representation learning with multi-dataset point prompt training. In CVPR, 2024. 3, 6, 7, 8, 16, 17 +[91] Zhirong Wu, Yuanjun Xiong, Stella X. Yu, and Dahua Lin. Unsupervised feature learning via non-parametric instance discrimination. In CVPR, 2018. 2 +[92] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. In NeurIPS, 2021. 3 + +[93] Saining Xie, Jiatao Gu, Demi Guo, Charles R Qi, Leonidas Guibas, and Or Litany. Pointcontrast: Unsupervised pretraining for 3d point cloud understanding. In ECCV, 2020. 2, 3, 4, 6, 7, 8, 16, 17 +[94] Zhenda Xie, Zheng Zhang, Yue Cao, Yutong Lin, Jianmin Bao, Zhuliang Yao, Qi Dai, and Han Hu. Simmim: A simple framework for masked image modeling. In CVPR, 2022. 2 +[95] Mutian Xu, Runyu Ding, Hengshuang Zhao, and Xiaojuan Qi. Paconv: Position adaptive convolution with dynamic kernel assembling on point clouds. In CVPR, 2021. 17 +[96] Xu Yan, Chaoda Zheng, Zhen Li, Sheng Wang, and Shuguang Cui. Pointasnl: Robust point clouds processing using nonlocal neural networks with adaptive sampling. In CVPR, 2020. 16 +[97] Jiancheng Yang, Qiang Zhang, Bingbing Ni, Linguuo Li, Jinxian Liu, Mengdie Zhou, and Qi Tian. Modeling point clouds with self-attention and gumbel subset sampling. In CVPR, 2019. 17 +[98] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Ji-ashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. In CVPR, 2024. 2 +[99] Lihe Yang, Bingyi Kang, Zilong Huang, Zhen Zhao, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything v2. In NeurIPS, 2024. 2 +[100] Yu-Qi Yang, Yu-Xiao Guo, Jian-Yu Xiong, Yang Liu, Hao Pan, Peng-Shuai Wang, Xin Tong, and Baining Guo. Swin3d: A pretrained transformer backbone for 3d indoor scene understanding. arXiv:2304.06906, 2023. 16, 17 +[101] Chandan Yeshwanth, Yueh-Cheng Liu, Matthias Nießner, and Angela Dai. Scannet++: A high-fidelity dataset of 3d indoor scenes. In ICCV, 2023. 2, 6, 7, 8 +[102] Xumin Yu, Lulu Tang, Yongming Rao, Tiejun Huang, Jie Zhou, and Jiwen Lu. Point-bert: Pre-training 3d point cloud transformers with masked point modeling. In CVPR, 2022. 2 +[103] Feihu Zhang, Jin Fang, Benjamin Wah, and Philip Torr. Deep fusionnet for point cloud semantic segmentation. In ECCV, 2020. 16 +[104] Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel M. Ni, and Heung-Yeung Shum. Dino: Detr with improved denoising anchor boxes for end-to-end object detection. arXiv:2203.03605, 2022. 2, 3, 4 +[105] Michael Zhang, Aditi Raghunathan Wang, Shiori Sagawa, Tatsunori B Hashimoto, and Percy Liang. Fine-tuning can distort pretrained features and underperform out-of-distribution. In ICLR, 2022. 2 +[106] Renrui Zhang, Ziyu Guo, Peng Gao, Rongyao Fang, Bin Zhao, Dong Wang, Yu Qiao, and Hongsheng Li. Pointm2ae: Multi-scale masked autoencoders for hierarchical point cloud pre-training. In NeurIPS, 2022. 2 +[107] Hengshuang Zhao, Li Jiang, Chi-Wing Fu, and Jiaya Jia. Pointweb: Enhancing local neighborhood features for point cloud processing. In CVPR, 2019. 17 +[108] Hengshuang Zhao, Li Jiang, Jiaya Jia, Philip Torr, and Vladlen Koltun. Point transformer. In ICCV, 2021. 16, 17 + +[109] Jia Zheng, Junfei Zhang, Jing Li, Rui Tang, Shenghua Gao, and Zihan Zhou. Structured3d: A large photo-realistic dataset for structured 3d modeling. In ECCV, 2020. 2, 5, 6 +[110] Jinghao Zhou, Chen Wei, Huiyu Wang, Wei Shen, Cihang Xie, Alan Yuille, and Tao Kong. ibot: Image bert pretraining with online tokenizer. In ICLR, 2022. 2, 4 +[111] Jinghao Zhou, Chen Wei, Huiyu Wang, Wei Shen, Cihang Xie, Alan Yuille, and Tao Kong. Image BERT pre-training with online tokenizer. In ICLR, 2022. 6 +[112] Xingyi Zhou, Vladlen Koltun, and Philipp Krahenbuhl. Simple multi-dataset detection. In CVPR, 2022. 3 \ No newline at end of file diff --git a/2025/Sonata_ Self-Supervised Learning of Reliable Point Representations/images.zip b/2025/Sonata_ Self-Supervised Learning of Reliable Point Representations/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..10effd340b34844e5d751562c8ec758831cc95dd --- /dev/null +++ b/2025/Sonata_ Self-Supervised Learning of Reliable Point Representations/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0b174d259247b8cf2353cef72462fa6511cf47115c06d194bd8c4b2827a9d1c7 +size 986587 diff --git a/2025/Sonata_ Self-Supervised Learning of Reliable Point Representations/layout.json b/2025/Sonata_ Self-Supervised Learning of Reliable Point Representations/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..d39a3f4ea313a732784f6b1358033ea1f3636a98 --- /dev/null +++ b/2025/Sonata_ Self-Supervised Learning of Reliable Point Representations/layout.json @@ -0,0 +1,9988 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 96, + 55, + 515, + 73 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 55, + 515, + 73 + ], + "spans": [ + { + "bbox": [ + 96, + 55, + 515, + 73 + ], + "type": "text", + "content": "Sonata: Self-Supervised Learning of Reliable Point Representations" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 70, + 94, + 539, + 125 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 94, + 539, + 125 + ], + "spans": [ + { + "bbox": [ + 70, + 94, + 539, + 125 + ], + "type": "text", + "content": "Xiaoyang Wu" + }, + { + "bbox": [ + 70, + 94, + 539, + 125 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 70, + 94, + 539, + 125 + ], + "type": "text", + "content": " Daniel DeTone" + }, + { + "bbox": [ + 70, + 94, + 539, + 125 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 70, + 94, + 539, + 125 + ], + "type": "text", + "content": " Duncan Frost" + }, + { + "bbox": [ + 70, + 94, + 539, + 125 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 70, + 94, + 539, + 125 + ], + "type": "text", + "content": " Tianwei Shen" + }, + { + "bbox": [ + 70, + 94, + 539, + 125 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 70, + 94, + 539, + 125 + ], + "type": "text", + "content": " Chris Xie" + }, + { + "bbox": [ + 70, + 94, + 539, + 125 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 70, + 94, + 539, + 125 + ], + "type": "text", + "content": " Nan Yang" + }, + { + "bbox": [ + 70, + 94, + 539, + 125 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 70, + 94, + 539, + 125 + ], + "type": "text", + "content": " Jakob Engel" + }, + { + "bbox": [ + 70, + 94, + 539, + 125 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 70, + 94, + 539, + 125 + ], + "type": "text", + "content": " Richard Newcombe" + }, + { + "bbox": [ + 70, + 94, + 539, + 125 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 70, + 94, + 539, + 125 + ], + "type": "text", + "content": " Hengshuang Zhao" + }, + { + "bbox": [ + 70, + 94, + 539, + 125 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 70, + 94, + 539, + 125 + ], + "type": "text", + "content": " Julian Straub" + }, + { + "bbox": [ + 70, + 94, + 539, + 125 + ], + "type": "inline_equation", + "content": "^{2}" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 151, + 137, + 459, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 137, + 459, + 152 + ], + "spans": [ + { + "bbox": [ + 151, + 137, + 459, + 152 + ], + "type": "text", + "content": "1The University of Hong Kong 2Meta Reality Labs Research" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 188, + 153, + 418, + 165 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 188, + 153, + 418, + 165 + ], + "spans": [ + { + "bbox": [ + 188, + 153, + 418, + 165 + ], + "type": "text", + "content": "https://github.com/facebookresearch/sonata" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 61, + 171, + 231, + 334 + ], + "blocks": [ + { + "bbox": [ + 61, + 171, + 231, + 334 + ], + "lines": [ + { + "bbox": [ + 61, + 171, + 231, + 334 + ], + "spans": [ + { + "bbox": [ + 61, + 171, + 231, + 334 + ], + "type": "image", + "image_path": "589d4300f43854af0829f5157255ce003401f1421bbb5571ff40b45a650e5aba.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 235, + 171, + 377, + 333 + ], + "blocks": [ + { + "bbox": [ + 235, + 171, + 377, + 333 + ], + "lines": [ + { + "bbox": [ + 235, + 171, + 377, + 333 + ], + "spans": [ + { + "bbox": [ + 235, + 171, + 377, + 333 + ], + "type": "image", + "image_path": "836a6f483902be083a4e0b93f8367b5b8d30e5b14d1b990fba0f77f1190fb148.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 379, + 171, + 550, + 333 + ], + "blocks": [ + { + "bbox": [ + 379, + 171, + 550, + 333 + ], + "lines": [ + { + "bbox": [ + 379, + 171, + 550, + 333 + ], + "spans": [ + { + "bbox": [ + 379, + 171, + 550, + 333 + ], + "type": "image", + "image_path": "56c11ac45288e311d5e6dbd4e365b9f89130bdd1783474991acb5e5bfaa77ffe.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 61, + 335, + 302, + 478 + ], + "blocks": [ + { + "bbox": [ + 61, + 335, + 302, + 478 + ], + "lines": [ + { + "bbox": [ + 61, + 335, + 302, + 478 + ], + "spans": [ + { + "bbox": [ + 61, + 335, + 302, + 478 + ], + "type": "image", + "image_path": "6e714a08721434cff40b5c0910e86ef7bb6c7bb036bc13df01cef90c14096176.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 484, + 555, + 551 + ], + "lines": [ + { + "bbox": [ + 55, + 484, + 555, + 551 + ], + "spans": [ + { + "bbox": [ + 55, + 484, + 555, + 551 + ], + "type": "text", + "content": "Figure 1. Main properties. Sonata leads to reliable 3D self-supervised pretraining with the following superior and emerging properties: 1. Perception. Sonata advances state-of-the-art results across 3D indoor and outdoor perception tasks; 2. Linear probing. With less than " + }, + { + "bbox": [ + 55, + 484, + 555, + 551 + ], + "type": "inline_equation", + "content": "0.2\\%" + }, + { + "bbox": [ + 55, + 484, + 555, + 551 + ], + "type": "text", + "content": " learnable parameters, Sonata achieves strong and usable linear probing performance which is " + }, + { + "bbox": [ + 55, + 484, + 555, + 551 + ], + "type": "inline_equation", + "content": "3.3\\times" + }, + { + "bbox": [ + 55, + 484, + 555, + 551 + ], + "type": "text", + "content": " better than previous SOTA; 3. Decoder-free. Sonata moves beyond the inflexible U-Net structure, offering multi-scale representations that unchain future 3D research from previous architectural constraints. 4. Semantic awareness. Sonata reveals semantic structure in PCA and K-means visualizations. 5. Spatial reasoning. Sonata allows spatial correspondence even under strong augmentations as visualized via feature similarity." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 305, + 335, + 550, + 478 + ], + "blocks": [ + { + "bbox": [ + 305, + 335, + 550, + 478 + ], + "lines": [ + { + "bbox": [ + 305, + 335, + 550, + 478 + ], + "spans": [ + { + "bbox": [ + 305, + 335, + 550, + 478 + ], + "type": "image", + "image_path": "989600f4ef288c1b4c60baa764c2c9036547f67de46eb78da322b2e565fdf5a7.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 151, + 555, + 200, + 567 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 555, + 200, + 567 + ], + "spans": [ + { + "bbox": [ + 151, + 555, + 200, + 567 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 571, + 297, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 571, + 297, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 571, + 297, + 693 + ], + "type": "text", + "content": "In this paper, we question whether we have a reliable self-supervised point cloud model that can be used for diverse 3D tasks via simple linear probing, even with limited data and minimal computation. We find that existing 3D self-supervised learning approaches fall short when evaluated on representation quality through linear probing. We hypothesize that this is due to what we term the \"geometric shortcut\", which causes representations to collapse to low-level spatial features. This challenge is unique to 3D and arises from the sparse nature of point cloud data." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 555, + 556, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 555, + 556, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 555, + 556, + 712 + ], + "type": "text", + "content": "We address it through two key strategies: obscuring spatial information and enhancing the reliance on input features, ultimately composing a Sonata of 140k point clouds through self-distillation. Sonata is simple and intuitive, yet its learned representations are strong and reliable: zero-shot visualizations demonstrate semantic grouping, alongside strong spatial reasoning through nearest-neighbor relationships. Sonata demonstrates exceptional parameter and data efficiency, tripling linear probing accuracy (from " + }, + { + "bbox": [ + 313, + 555, + 556, + 712 + ], + "type": "inline_equation", + "content": "21.8\\%" + }, + { + "bbox": [ + 313, + 555, + 556, + 712 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 555, + 556, + 712 + ], + "type": "inline_equation", + "content": "72.5\\%" + }, + { + "bbox": [ + 313, + 555, + 556, + 712 + ], + "type": "text", + "content": ") on ScanNet and nearly doubling performance with only " + }, + { + "bbox": [ + 313, + 555, + 556, + 712 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 313, + 555, + 556, + 712 + ], + "type": "text", + "content": " of the data compared to previous approaches. Full fine-tuning further advances SOTA across both 3D indoor and outdoor perception tasks." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 15 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 15 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 15 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 198, + 13, + 442, + 24 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 13, + 442, + 24 + ], + "spans": [ + { + "bbox": [ + 198, + 13, + 442, + 24 + ], + "type": "text", + "content": "Except for this watermark, it is identical to the accepted version;" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 179, + 25, + 461, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 25, + 461, + 37 + ], + "spans": [ + { + "bbox": [ + 179, + 25, + 461, + 37 + ], + "type": "text", + "content": "the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 66, + 703, + 196, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 703, + 196, + 713 + ], + "spans": [ + { + "bbox": [ + 66, + 703, + 196, + 713 + ], + "type": "text", + "content": "*Equal contribution in alphabetic order." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "text", + "content": "22193" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 136, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 136, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 136, + 83 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 88, + 296, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 88, + 296, + 220 + ], + "spans": [ + { + "bbox": [ + 55, + 88, + 296, + 220 + ], + "type": "text", + "content": "Self-supervised learning (SSL) with images [2, 13, 34, 94, 104] has seen a continuous increase in model simplicity, capacity, and capability over the past decade [25]. Tuning a single linear layer can achieve performance close to full fine-tuning [31, 35, 104], fostering growing trust in its reliability. This trust has been further strengthened by witnessing the semantic meaning of learned image representations through direct visualization [60, 110]. Consequently, these reliable self-supervised models have become the foundation for emerging approaches [14, 84, 98, 99] across various fields involving images." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 221, + 296, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 221, + 296, + 388 + ], + "spans": [ + { + "bbox": [ + 55, + 221, + 296, + 388 + ], + "type": "text", + "content": "In contrast to the image domain, self-supervised learning with point clouds [18, 61, 93, 102, 106] is still in its early stages. Despite the broad reliance on 3D applications in autonomous driving [9, 74], robotic learning [32], mixed reality [5, 43] and egocentric perception [27, 73], the latest self-supervised point cloud models are seldom included in their pipelines. This gap prompts us to consider a simple yet critical question: do we have a reliable point self-supervised learning approach that provides strong representations, usable with simple linear probing across these applications? Not yet. Previous SOTAs [38, 88, 93] fall short on this higher-level criterion, achieving only a maximum of " + }, + { + "bbox": [ + 55, + 221, + 296, + 388 + ], + "type": "inline_equation", + "content": "21.8\\%" + }, + { + "bbox": [ + 55, + 221, + 296, + 388 + ], + "type": "text", + "content": " mIoU on ScanNet [23] with linear probing, especially given that the performance from scratch is " + }, + { + "bbox": [ + 55, + 221, + 296, + 388 + ], + "type": "inline_equation", + "content": "77.6\\%" + }, + { + "bbox": [ + 55, + 221, + 296, + 388 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 389, + 296, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 389, + 296, + 509 + ], + "spans": [ + { + "bbox": [ + 55, + 389, + 296, + 509 + ], + "type": "text", + "content": "We identify the geometric shortcut as the primary issue hindering these prior point cloud SSL approaches from learning reliable representations. This shortcut refers to the tendency of the model to collapse to easily accessible, low-level geometric cues, such as normal direction or point height, as demonstrated in prior works and visualized in Fig. 2. This spatial information is inevitably introduced into point cloud operators along with point coordinates rather than through input features, making it difficult to obscure and nearly impossible to mask effectively." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 510, + 295, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 295, + 641 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 295, + 641 + ], + "type": "text", + "content": "However, the model collapse caused by geometric shortcuts can be mitigated through two key strategies: obscuring spatial information and emphasizing input features. Specifically, we address this issue by applying SSL losses at coarser spatial scales, disturbing the spatial information of masked points without features, and progressively increasing task difficulty to reduce reliance on accessible geometric cues. Coupled with a point self-distillation framework and scaling techniques inspired by recent advances in image SSL [11, 12, 34, 60, 110], we ultimately compose a Sonata using 140k point cloud scenes [1, 3, 5, 23, 68, 101, 109]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 642, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 713 + ], + "type": "text", + "content": "Sonata demonstrates strong zero-shot capabilities, with PCA-colored visualizations of point clouds, k-means clustering of features, and nearest-neighbor matching between point clouds (see Fig. 1). Sonata also proves highly data-efficient, raising semantic segmentation performance from " + }, + { + "bbox": [ + 55, + 642, + 296, + 713 + ], + "type": "inline_equation", + "content": "25.8\\%" + }, + { + "bbox": [ + 55, + 642, + 296, + 713 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 642, + 296, + 713 + ], + "type": "inline_equation", + "content": "45.3\\%" + }, + { + "bbox": [ + 55, + 642, + 296, + 713 + ], + "type": "text", + "content": " under extremely limited data conditions" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 314, + 70, + 553, + 243 + ], + "blocks": [ + { + "bbox": [ + 314, + 70, + 553, + 243 + ], + "lines": [ + { + "bbox": [ + 314, + 70, + 553, + 243 + ], + "spans": [ + { + "bbox": [ + 314, + 70, + 553, + 243 + ], + "type": "image", + "image_path": "b05cacfdb88fae41dddb3a576b8a396fcf586d24cec5ed0e7025e057ff9a51d6.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 244, + 553, + 310 + ], + "lines": [ + { + "bbox": [ + 313, + 244, + 553, + 310 + ], + "spans": [ + { + "bbox": [ + 313, + 244, + 553, + 310 + ], + "type": "text", + "content": "Figure 2. Geometric shortcut. We select a point on the sofa arm and compute pairwise similarity with other points. The similarity heatmap reveals that CSC [38] collapses to surface normals, and MSC [88] overfits to point height. In contrast, our Sonata can extract higher-level concepts, as can be seen by the high similarity between all sofa arms highlighted in red." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 315, + 555, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 315, + 555, + 434 + ], + "spans": [ + { + "bbox": [ + 313, + 315, + 555, + 434 + ], + "type": "text", + "content": "(1% of ScanNet). Additionally, Sonata significantly boosts linear probing accuracy on ScanNet semantic segmentation, increasing it by over " + }, + { + "bbox": [ + 313, + 315, + 555, + 434 + ], + "type": "inline_equation", + "content": "3.3 \\times" + }, + { + "bbox": [ + 313, + 315, + 555, + 434 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 313, + 315, + 555, + 434 + ], + "type": "inline_equation", + "content": "21.8\\%" + }, + { + "bbox": [ + 313, + 315, + 555, + 434 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 315, + 555, + 434 + ], + "type": "inline_equation", + "content": "72.5\\%" + }, + { + "bbox": [ + 313, + 315, + 555, + 434 + ], + "type": "text", + "content": " and surpassing the accuracy of DINOv2 features aggregated onto the point cloud (63.1%). Moreover, combining Sonata features with DINOv2 features further enhances accuracy (76.4%), underscoring that Sonata captures unique 3D information beyond what is visible in images alone. Finally, Sonata achieves state-of-the-art results across various indoor and outdoor perception tasks with full fine-tuning." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 444, + 400, + 456 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 444, + 400, + 456 + ], + "spans": [ + { + "bbox": [ + 313, + 444, + 400, + 456 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 461, + 554, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 461, + 554, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 461, + 554, + 605 + ], + "type": "text", + "content": "Image self-supervised learning. Over the past decade, remarkable advancements [25, 26, 36, 59, 91] have been made in image self-supervised learning, and our research is largely inspired by two pivotal moments in this field. First, linear probing, a method that assesses representation quality by optimizing a minimal linear transformation, has become a standard in 2D image SSL [10, 13, 34, 91]. In some cases, such as when the distribution shift is large, linear probing surpasses full fine-tuning [105]. Second, the ability to directly perceive the semantic meaning of learned representations through zero-shot visualization like PCA or attention [12, 60] has further strengthened trust in reliability." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 606, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 606, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 606, + 554, + 714 + ], + "type": "text", + "content": "Point self-supervised learning. Sonata follows the research path initiated by PointContrast [38, 93], emphasizing self-supervised learning with scene-level data [23]. While previous efforts do implement strategies to prevent collapse from geometric shortcuts, they remain limited. For example, Masked Scene Contrast (MSC) [88] encourages learning beyond naive geometric cues by predicting color or normal vectors but still partially anchors representations to predefined tasks. GroupContrast (GC) [80] employs graph-" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22194" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 64, + 294, + 199 + ], + "blocks": [ + { + "bbox": [ + 57, + 64, + 294, + 199 + ], + "lines": [ + { + "bbox": [ + 57, + 64, + 294, + 199 + ], + "spans": [ + { + "bbox": [ + 57, + 64, + 294, + 199 + ], + "type": "image", + "image_path": "c61e24b74f3f68c8dd8dbf6f8b7bfde203cf8897155a2cb77a0785bd964a9ac1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 201, + 295, + 277 + ], + "lines": [ + { + "bbox": [ + 55, + 201, + 295, + 277 + ], + "spans": [ + { + "bbox": [ + 55, + 201, + 295, + 277 + ], + "type": "text", + "content": "Figure 3. The geometric shortcut is unique to 3D. When comparing the information contained in 2D image and 3D point cloud data after removing the input feature (indicated by color), it is evident that in images all information is within the input feature. Whereas point clouds retain geometric information in point positions, which is directly utilized by operators. This characteristic leads to what we term geometric shortcuts in 3D SSL." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 281, + 296, + 341 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 281, + 296, + 341 + ], + "spans": [ + { + "bbox": [ + 55, + 281, + 296, + 341 + ], + "type": "text", + "content": "based segment guidance [28], though it is constrained by its reliance on human-designed algorithms. Building on MSC, Sonata directly addresses the geometric shortcut and scales up training [90] to establish a more reliable approach to point cloud self-supervised learning." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 341, + 296, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 341, + 296, + 509 + ], + "spans": [ + { + "bbox": [ + 55, + 341, + 296, + 509 + ], + "type": "text", + "content": "Point cloud backbones. This area [17, 51, 56, 64, 67, 77] has significantly benefited from the U-Net structure [70], particularly with its hierarchical decoding and skip connections, as first introduced by PointNet++ [66]. However, the tight coupling between the encoder and decoder restricts flexibility and generalization capacity [92]. Sonata addresses this limitation by focusing self-supervised learning exclusively on the encoder, thus removing the hierarchical decoder. Moreover, unlike previous SOTAs [38, 88, 93] in point SSL, which primarily use SparseUNet [17, 20, 29] to balance efficiency and accuracy, we leverage Point Transformer V3 (PTv3) [89], an efficient, accurate, and scalable transformer backbone. This shift alone yields a " + }, + { + "bbox": [ + 55, + 341, + 296, + 509 + ], + "type": "inline_equation", + "content": "7.7\\%" + }, + { + "bbox": [ + 55, + 341, + 296, + 509 + ], + "type": "text", + "content": " improvement in linear probing performance over MSC [88]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 517, + 237, + 531 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 517, + 237, + 531 + ], + "spans": [ + { + "bbox": [ + 55, + 517, + 237, + 531 + ], + "type": "text", + "content": "3. Pilot Study and Design Principle" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 534, + 295, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 295, + 569 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 295, + 569 + ], + "type": "text", + "content": "In this section, we qualitatively study the problems of current point self-supervised approaches and point cloud backbones to inform the Sonata approach." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 570, + 295, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 295, + 665 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 295, + 665 + ], + "type": "text", + "content": "Uncovering the geometric shortcut in point SSL. The history of image self-supervised learning [34, 60, 104, 112] can be summarized as a continuous battle against shortcuts, where models often exploit trivial solutions (mode collapse) rather than understanding deeper semantics. Each advancement has involved identifying, understanding, and overcoming these shortcuts, refining pretext tasks to push models to \"struggle\" in learning stronger representations." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "However in 3D point SSL, despite various attempts [38, 88, 93] to increase the difficulty of pretext tasks, a curious phenomenon persists—the loss consistently reduces rapidly to an ideal range in the early stages of training. We hypo-" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 316, + 64, + 553, + 132 + ], + "blocks": [ + { + "bbox": [ + 316, + 64, + 553, + 132 + ], + "lines": [ + { + "bbox": [ + 316, + 64, + 553, + 132 + ], + "spans": [ + { + "bbox": [ + 316, + 64, + 553, + 132 + ], + "type": "image", + "image_path": "666811aba4d3585c2cf6b2ea3f73df98e6e959b8fcc6969307f1f17a20722f56.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 133, + 555, + 232 + ], + "lines": [ + { + "bbox": [ + 313, + 133, + 555, + 232 + ], + "spans": [ + { + "bbox": [ + 313, + 133, + 555, + 232 + ], + "type": "text", + "content": "Figure 4. What is learned by the hierarchical backbone? We visualize PCA embeddings from different stages of a hierarchical encoder and decoder, trained for semantic segmentation. The encoder captures diverse and dispersed feature patterns, indicating a broad range of information. Notably, as the point cloud becomes coarser, accessible geometric information within point coordinates becomes increasingly global. In contrast, the decoder's representations are more uniform and structured, suggesting a focus on refining features for task-specific outputs." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 239, + 554, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 239, + 554, + 323 + ], + "spans": [ + { + "bbox": [ + 313, + 239, + 554, + 323 + ], + "type": "text", + "content": "esize that this lack of \"struggle\" in learning indicates that a shortcut exists that leads to a collapse of the representation to trivial solutions. Indeed as Fig. 2 illustrates, previous 3D SSL approaches seem to learn representations that are sensitive to local surface normals or point height. The collapse to such naive solutions even for well-designed SSL pretext tasks is what we term the geometric shortcut." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 325, + 554, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 325, + 554, + 468 + ], + "spans": [ + { + "bbox": [ + 313, + 325, + 554, + 468 + ], + "type": "text", + "content": "Intuitively, representations affected by this geometric shortcut have not learned sufficient semantics. To quantify this problem, we leverage the standard criteria from 2D SSL of linear probing: we linearly probe the semantic class of each point from the learned representation. As illustrated in Fig. 1, with linear probing, previous 3D SSL methods PC [93] and MSC [88] indeed achieve only " + }, + { + "bbox": [ + 313, + 325, + 554, + 468 + ], + "type": "inline_equation", + "content": "5.6\\%" + }, + { + "bbox": [ + 313, + 325, + 554, + 468 + ], + "type": "text", + "content": " mIoU and " + }, + { + "bbox": [ + 313, + 325, + 554, + 468 + ], + "type": "inline_equation", + "content": "21.8\\%" + }, + { + "bbox": [ + 313, + 325, + 554, + 468 + ], + "type": "text", + "content": " mIoU respectively on ScanNet semantic segmentation. When compared to the " + }, + { + "bbox": [ + 313, + 325, + 554, + 468 + ], + "type": "inline_equation", + "content": "63.1\\%" + }, + { + "bbox": [ + 313, + 325, + 554, + 468 + ], + "type": "text", + "content": " mIoU achieved by 3D-aggregated image representations from DINOv2 [60], it becomes obvious that the current 3D SSL methods do not learn semantic information." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 471, + 554, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 471, + 554, + 590 + ], + "spans": [ + { + "bbox": [ + 313, + 471, + 554, + 590 + ], + "type": "text", + "content": "We hypothesize that the geometric shortcut stems from the sparse nature of point cloud data. Every operator, whether for point clouds or images, inherently relies on point (pixel) coordinates to define the kernel. However, unlike images with regularly spaced dense pixels, the sparsity of point cloud data necessitates introducing point coordinate information into point cloud operators rather than just through input features (see Fig. 3). This fact makes it difficult to obscure and nearly impossible to effectively mask out, ultimately resulting in the geometric shortcut." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 592, + 554, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 592, + 554, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 592, + 554, + 628 + ], + "type": "text", + "content": "These observations motivate the training methodology of Sonata aimed to prevent the dominance of naive spatial information inherent in point cloud organization." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "text", + "content": "Focusing self-supervised learning on encoder only. Following our hypothesis that the geometric arrangement leaks information into the learned representation via the geometric shortcut, we examine the point cloud model. The typical U-Net structure [66, 70] is effective at handling large point clouds in a coarse-to-fine way, but the tight coupling of the encoder and decoder via skip connections restricts" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "text", + "content": "22195" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "content": "flexibility. Specifically, the decoder enforces per-point features at the original high-resolution scale, with shallow feature channels. This constraint limits the capacity to provide richer representations, which is crucial for SSL as evidenced by large channel dimensions in state-of-the-art 2D approaches [12]. Most importantly, decoding point clouds at the original scale unavoidably introduces local geometric cues into operators facilitating the geometric shortcut." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 168, + 295, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 168, + 295, + 287 + ], + "spans": [ + { + "bbox": [ + 55, + 168, + 295, + 287 + ], + "type": "text", + "content": "In Fig. 4, we visualize PCA embeddings from different stages of PTv3 encoder and decoder, trained in a supervised manner for semantic segmentation. We observe that the encoder learns diverse features capturing spatial features at different scales in contrast to the decoder which produces task-specific higher-level representations. Additionally, as the spatial resolution of the point cloud decreases via the max-pooling stages, feature representations become less local. This resolution reduction fundamentally limits reliance on fine spatial information tied to point coordinates." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 288, + 295, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 288, + 295, + 373 + ], + "spans": [ + { + "bbox": [ + 55, + 288, + 295, + 373 + ], + "type": "text", + "content": "These observations motivate us to remove the decoder during self-supervised learning for two main reasons. First and foremost, training directly with features at coarser point resolutions inherently restricts access to fine-grained spatial information, reducing the possibility of geometric shortcuts. Second, task-specific features can be probed or fine-tuned on top of a more expressive multi-scale point representation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 380, + 242, + 392 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 380, + 242, + 392 + ], + "spans": [ + { + "bbox": [ + 55, + 380, + 242, + 392 + ], + "type": "text", + "content": "4. Point Self-distillation with Sonata" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 398, + 295, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 398, + 295, + 458 + ], + "spans": [ + { + "bbox": [ + 55, + 398, + 295, + 458 + ], + "type": "text", + "content": "This section details the methodology of Sonata, the point self-distillation framework designed to address geometric shortcuts, remove structural constraints, and deliver strong linear probing results as discussed in Sec. 3. A roadmap of incremental ablation is illustrated in Fig. 6." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 461, + 167, + 472 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 461, + 167, + 472 + ], + "spans": [ + { + "bbox": [ + 55, + 461, + 167, + 472 + ], + "type": "text", + "content": "4.1. Macro Framework" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 473, + 295, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 473, + 295, + 556 + ], + "spans": [ + { + "bbox": [ + 55, + 473, + 295, + 556 + ], + "type": "text", + "content": "Before diving into the specific micro designs, we begin with a point self-distillation framework derived from insights gained through previous efforts in point [88, 93] and image [11, 34, 60, 104, 110] self-supervised learning. This macro framework provides a solid foundation for pretext task design, allowing us to focus on addressing the geometric shortcut without additional concerns." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 558, + 295, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 295, + 641 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 295, + 641 + ], + "type": "text", + "content": "In essence, (point) self-supervised learning aims to make things (points) that should be the same, the same (identical in representation). This forms the basic recipe of point SSL: generating two views of a given point cloud with random spatial (e.g., crop, rotate, distort) and photometric (e.g., jitter) augmentations, then matching and aligning the feature embeddings of points that are close in the original space." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "type": "text", + "content": "However, the superior robustness of self-supervised representation is rooted in a core principle: continuously increasing the difficulty of pretext tasks as long as the model continues to converge. This principle encourages enhancing the basic recipe with local-global view alignment and mask-unmask view alignment. Specifically, it involves aligning" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 317, + 65, + 550, + 213 + ], + "blocks": [ + { + "bbox": [ + 317, + 65, + 550, + 213 + ], + "lines": [ + { + "bbox": [ + 317, + 65, + 550, + 213 + ], + "spans": [ + { + "bbox": [ + 317, + 65, + 550, + 213 + ], + "type": "image", + "image_path": "6cc7c5098377002cf7d37801254da98aad83255b8febd0aa13ed7539d6847ec5.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 214, + 553, + 324 + ], + "lines": [ + { + "bbox": [ + 313, + 214, + 553, + 324 + ], + "spans": [ + { + "bbox": [ + 313, + 214, + 553, + 324 + ], + "type": "text", + "content": "Figure 5. Self-distillation framework of Sonata. (1) Local views (bottom left) and global views (right) are generated with dedicated spatial and photometric augmentations, while masked views are created by randomly masking out grid-based patches from the global views (top left). (2) Embeddings from local and masked views are extracted by the student, with global views processed by the teacher (top). (3) Points from local and masked views are matched with corresponding points in the global views based on their original spatial distance, allowing for the distillation of embeddings from global views to local and masked views (bottom)." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 329, + 553, + 401 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 329, + 553, + 401 + ], + "spans": [ + { + "bbox": [ + 313, + 329, + 553, + 401 + ], + "type": "text", + "content": "neighboring points from local views generated by small cropping ratios and masked views created with large masking ratios (see Fig. 5 left), with unmasked global views that contain relatively richer information (see Fig. 5 right). The difficulty of pretext tasks can be scaled by adjusting the cropping and masking ratios." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 402, + 554, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 402, + 554, + 556 + ], + "spans": [ + { + "bbox": [ + 313, + 402, + 554, + 556 + ], + "type": "text", + "content": "The more challenging the pretext tasks become, the higher the risk of model collapse. This instability calls for an asymmetric encoding approach, specifically exponential moving average (EMA) [34]: rather than encoding all views with a shared-weight model, the approach encodes challenging local and masked views with an actively learned student model, while using a stable teacher model, updated with a moving average of the student parameters, to encode global views (see Fig. 5 top). With the teacher preventing the student from being misled, the student is less likely to get lost in \"mission impossible\" and has a greater chance to discover treasure within \"impossible\" (i.e. extreme " + }, + { + "bbox": [ + 313, + 402, + 554, + 556 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 313, + 402, + 554, + 556 + ], + "type": "text", + "content": " crop ratio for local views and " + }, + { + "bbox": [ + 313, + 402, + 554, + 556 + ], + "type": "inline_equation", + "content": "70\\%" + }, + { + "bbox": [ + 313, + 402, + 554, + 556 + ], + "type": "text", + "content": " mask ratio for masked views)." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 558, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 554, + 713 + ], + "type": "text", + "content": "In terms of SSL criteria, we move away from prior approaches that rely on contrastive and generative learning [88]. Contrastive learning, limited by the number of point pairs in pairwise similarity computations, restricts scalability. Generative learning, meanwhile, partially anchors representations to predefined cues, limiting the model's capacity to capture more generalizable features. Following DINov2, we adopt a self-distillation approach driven by Sinkhorn-Knopp centering [11], KoLeo regularization [72], and clustering assignments [11]. This adaptation initially exacerbates collapse into geometric shortcuts but holds greater potential for robust representation once this challenge is addressed." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22196" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 141, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 141, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 141, + 85 + ], + "type": "text", + "content": "4.2. Micro Design" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 86, + 296, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 86, + 296, + 158 + ], + "spans": [ + { + "bbox": [ + 55, + 86, + 296, + 158 + ], + "type": "text", + "content": "We now discuss micro designs aimed at addressing geometric shortcuts. Since the problematic spatial information is inherently tied to point coordinates and directly used by operators, it is nearly impossible to mask out. This constraint defines the key strategies of our micro designs: obscuring spatial information and emphasizing input features." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 159, + 296, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 159, + 296, + 338 + ], + "spans": [ + { + "bbox": [ + 55, + 159, + 296, + 338 + ], + "type": "text", + "content": "Decoder removal. Previous approaches adhere to the original U-Net-style backbone for feature extraction; however, our observations in Sec. 3 motivate us to remove the complex hierarchical decoder and perform self-distillation directly using the encoder's output. This simple adjustment is key in our fight against geometric shortcuts, as it: (a) increases the feature channels participating in self-distillation (from 96 to 512), (b) streamlines the pipeline by involving fewer points in the pretext task after hierarchical pooling, and, most importantly, (c) introduces a natural way to obscure geometric cues: the positional information of points becomes naturally disturbed during hierarchical encoding and pooling. In fact, this removal proves to be even more beneficial than expected, boosting the linear probing result from " + }, + { + "bbox": [ + 55, + 159, + 296, + 338 + ], + "type": "inline_equation", + "content": "20.7\\%" + }, + { + "bbox": [ + 55, + 159, + 296, + 338 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 159, + 296, + 338 + ], + "type": "inline_equation", + "content": "60.4\\%" + }, + { + "bbox": [ + 55, + 159, + 296, + 338 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 340, + 296, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 340, + 296, + 567 + ], + "spans": [ + { + "bbox": [ + 55, + 340, + 296, + 567 + ], + "type": "text", + "content": "Feature up-casting. While the removal effectively prevents over-reliance on naive geometric cues, it also introduces certain limitations, particularly in leveraging multiscale contexts. In the original U-Net structure, the decoder plays a key role in progressively aggregating features across scales to reconstruct semantic details. Without this process, self-distillation struggles to capture multi-scale spatial information and broader contextual relationships. To retain multi-scale features, we introduce a parameter-free feature up-casting process similar to hypercolumns in image segmentation [33]: progressively up-casting features back to the scale of the previous encoding stage, with the mapping relationships preserved through pooling layers and concatenation with features from the prior encoding stage. This approach provides richer, coarse-to-fine features from multiscale encoding. While it does increase the risk of the model falling into geometric shortcuts, finding the right balance is key. Our ablation study shows that up-casting features twice achieve the best performance." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 568, + 296, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 568, + 296, + 676 + ], + "spans": [ + { + "bbox": [ + 55, + 568, + 296, + 676 + ], + "type": "text", + "content": "Masked points jitter. A slight random Gaussian jitter " + }, + { + "bbox": [ + 55, + 568, + 296, + 676 + ], + "type": "inline_equation", + "content": "(\\sigma = 0.005)" + }, + { + "bbox": [ + 55, + 568, + 296, + 676 + ], + "type": "text", + "content": " is applied to point coordinates as part of data augmentation. However, specifically for points to be masked, we additionally apply a stronger Gaussian jitter " + }, + { + "bbox": [ + 55, + 568, + 296, + 676 + ], + "type": "inline_equation", + "content": "(\\sigma = 0.01)" + }, + { + "bbox": [ + 55, + 568, + 296, + 676 + ], + "type": "text", + "content": " to further disrupt their spatial relationships. We pay special attention to these masked points because models are more likely to collapse to naive geometric cues, especially when point features are masked, making it difficult to derive solutions from neighboring unmasked points." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": "Progressive parameter scheduler. Geometric shortcuts are like traps on the convergence path of point cloud self-supervised learning. Similarly, we can set our own \"trap\"" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 319, + 67, + 553, + 357 + ], + "blocks": [ + { + "bbox": [ + 319, + 67, + 553, + 357 + ], + "lines": [ + { + "bbox": [ + 319, + 67, + 553, + 357 + ], + "spans": [ + { + "bbox": [ + 319, + 67, + 553, + 357 + ], + "type": "image", + "image_path": "8a03d2c187c4b70f274a3c5da3bebb8c98825d9400e2e0661479788c0b34228e.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 359, + 555, + 449 + ], + "lines": [ + { + "bbox": [ + 313, + 359, + 555, + 449 + ], + "spans": [ + { + "bbox": [ + 313, + 359, + 555, + 449 + ], + "type": "text", + "content": "Figure 6. The roadmap. We evolve Mask Scene Contrast [88] into our Sonata by modernizing self-supervised learning with self-distillation, addressing the geometric shortcut, and scaling up training. Our designs are validated through progressive ablation with linear and decoder probing on ScanNet semantic segmentation [23]. Starting with " + }, + { + "bbox": [ + 313, + 359, + 555, + 449 + ], + "type": "inline_equation", + "content": "23\\mathrm{k}" + }, + { + "bbox": [ + 313, + 359, + 555, + 449 + ], + "type": "text", + "content": " training data (a combination of ScanNet and Structured3D [109]) and a " + }, + { + "bbox": [ + 313, + 359, + 555, + 449 + ], + "type": "inline_equation", + "content": "39\\mathrm{M}" + }, + { + "bbox": [ + 313, + 359, + 555, + 449 + ], + "type": "text", + "content": " PTv3 model [89], we ultimately scale up to " + }, + { + "bbox": [ + 313, + 359, + 555, + 449 + ], + "type": "inline_equation", + "content": "140\\mathrm{k}" + }, + { + "bbox": [ + 313, + 359, + 555, + 449 + ], + "type": "text", + "content": " assets (Tab. 1) and a " + }, + { + "bbox": [ + 313, + 359, + 555, + 449 + ], + "type": "inline_equation", + "content": "108\\mathrm{M}" + }, + { + "bbox": [ + 313, + 359, + 555, + 449 + ], + "type": "text", + "content": " PTv3 model." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 458, + 555, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 458, + 555, + 650 + ], + "spans": [ + { + "bbox": [ + 313, + 458, + 555, + 650 + ], + "type": "text", + "content": "during training. For example, rather than starting with a challenging large mask size and mask ratio, we begin with a relatively small mask size (10 cm) and mask ratio (30%), gradually increasing them to 40 cm (0.1 " + }, + { + "bbox": [ + 313, + 458, + 555, + 650 + ], + "type": "inline_equation", + "content": "\\Rightarrow" + }, + { + "bbox": [ + 313, + 458, + 555, + 650 + ], + "type": "text", + "content": " 0.4) and 70% (0.3 " + }, + { + "bbox": [ + 313, + 458, + 555, + 650 + ], + "type": "inline_equation", + "content": "\\Rightarrow" + }, + { + "bbox": [ + 313, + 458, + 555, + 650 + ], + "type": "text", + "content": " 0.7) over the first 5% of the training process. This strategy encourages the model to rely more on input features to solve the pre-defined pretext tasks, preventing it from shifting its reliance to point coordinates as training difficulty increases. This approach aligns with curriculum learning [7], progressively challenging the model as it adapts. Similarly, in addition to the common learning rate scheduler, we also implement custom progressive schedulers for teacher temperature (0.04 " + }, + { + "bbox": [ + 313, + 458, + 555, + 650 + ], + "type": "inline_equation", + "content": "\\Rightarrow" + }, + { + "bbox": [ + 313, + 458, + 555, + 650 + ], + "type": "text", + "content": " 0.07) and weight decay (0.04 " + }, + { + "bbox": [ + 313, + 458, + 555, + 650 + ], + "type": "inline_equation", + "content": "\\Rightarrow" + }, + { + "bbox": [ + 313, + 458, + 555, + 650 + ], + "type": "text", + "content": " 0.2). We found that this design pushes these parameters to more extreme levels previously unexplored, further enhancing model performance." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 652, + 533, + 664 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 652, + 533, + 664 + ], + "spans": [ + { + "bbox": [ + 313, + 652, + 533, + 664 + ], + "type": "text", + "content": "4.3. Implementation and Evaluation Protocols" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 665, + 553, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 553, + 689 + ], + "type": "text", + "content": "In this section, we introduce the implementation details and the evaluation protocols for our experiments." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "type": "text", + "content": "Backbone. We build our Sonata with Point Transformer V3 (PTv3) [89] and refer to Pointcept [19] for details of imple" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "text", + "content": "22197" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 57, + 70, + 294, + 180 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 294, + 180 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 294, + 180 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 294, + 180 + ], + "type": "table", + "html": "
DatasetSourceTrainValTestTotal
ScanNet [23]real1,2013121001,613
ScanNet++ [101]real7121781261,016
S3DIS [1]real204680272
ArkitScenes [5]real4,49854905,047
HM3D [68]real8,8811,119010,000
Structured3D [109]sim.18,3481,7761,69721,821
ASE [3]sim.90,00010,0000100,000
Sonata (ours)mixed123,84414,0021,923139,769
", + "image_path": "3835880d60da44d365e2818248dcdca8d0a3030064241bfd58cd9ec893d644a8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 59, + 195, + 289, + 259 + ], + "blocks": [ + { + "bbox": [ + 115, + 182, + 236, + 191 + ], + "lines": [ + { + "bbox": [ + 115, + 182, + 236, + 191 + ], + "spans": [ + { + "bbox": [ + 115, + 182, + 236, + 191 + ], + "type": "text", + "content": "Table 1. Data source collection." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 59, + 195, + 289, + 259 + ], + "lines": [ + { + "bbox": [ + 59, + 195, + 289, + 259 + ], + "spans": [ + { + "bbox": [ + 59, + 195, + 289, + 259 + ], + "type": "table", + "html": "
MethodRealSimTotalMultiplier
PC [93]1,61301,613×1
MSC [88]6,66006,660×4.1
PPT [90]1,88521,82123,706×14.7
Sonata (ours)17,948121,821139,768×86.7
", + "image_path": "fd73843727f38e6b0dfe20421a52580b86b1ad863350b8a6b3dcbf5ac93c1e76.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 115, + 261, + 236, + 271 + ], + "lines": [ + { + "bbox": [ + 115, + 261, + 236, + 271 + ], + "spans": [ + { + "bbox": [ + 115, + 261, + 236, + 271 + ], + "type": "text", + "content": "Table 2. Data scale comparison." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 281, + 295, + 412 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 281, + 295, + 412 + ], + "spans": [ + { + "bbox": [ + 55, + 281, + 295, + 412 + ], + "type": "text", + "content": "mentation. Building on this, we made an additional adjustment to enhance scalability: replacing all Batch Normalization (BN) [41] layers with Layer Normalization (LN) [4]. Although this replacement results in some initial accuracy degradation, it enhances domain adaptation by eliminating the need for additional domain-specific adjustments when scaling up with multi-dataset joint training [90]. Along with the scaling up of data, we also scale up the encoder block depths from " + }, + { + "bbox": [ + 55, + 281, + 295, + 412 + ], + "type": "inline_equation", + "content": "[2, 2, 2, 6, 2]" + }, + { + "bbox": [ + 55, + 281, + 295, + 412 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 281, + 295, + 412 + ], + "type": "inline_equation", + "content": "[3, 3, 3, 12, 3]" + }, + { + "bbox": [ + 55, + 281, + 295, + 412 + ], + "type": "text", + "content": " and widths from " + }, + { + "bbox": [ + 55, + 281, + 295, + 412 + ], + "type": "inline_equation", + "content": "[32, 64, 128, 256, 512]" + }, + { + "bbox": [ + 55, + 281, + 295, + 412 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 281, + 295, + 412 + ], + "type": "inline_equation", + "content": "[48, 96, 192, 384, 512]" + }, + { + "bbox": [ + 55, + 281, + 295, + 412 + ], + "type": "text", + "content": ". This PTv3 model has " + }, + { + "bbox": [ + 55, + 281, + 295, + 412 + ], + "type": "inline_equation", + "content": "108\\mathrm{M}" + }, + { + "bbox": [ + 55, + 281, + 295, + 412 + ], + "type": "text", + "content": " parameters." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 413, + 295, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 413, + 295, + 509 + ], + "spans": [ + { + "bbox": [ + 55, + 413, + 295, + 509 + ], + "type": "text", + "content": "Data. We extend the multi-dataset joint training approach introduced by PPT [90], further expanding the data scale by removing the constraint of human labeling through unsupervised learning. This results in a collection of 140k scene-level point clouds from both real-world and simulated environments (outlined in Tab. 1), making it " + }, + { + "bbox": [ + 55, + 413, + 295, + 509 + ], + "type": "inline_equation", + "content": "86.7 \\times" + }, + { + "bbox": [ + 55, + 413, + 295, + 509 + ], + "type": "text", + "content": " larger than the data scale of PointContrast [93] and " + }, + { + "bbox": [ + 55, + 413, + 295, + 509 + ], + "type": "inline_equation", + "content": "5.9 \\times" + }, + { + "bbox": [ + 55, + 413, + 295, + 509 + ], + "type": "text", + "content": " larger than the data collection of PPT, as detailed in Tab. 2." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 510, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 295, + 714 + ], + "type": "text", + "content": "Training. We train Sonata on the 140k data collection for 200 epochs, using the AdamW optimizer [54] with a batch size of 96, distributed across 32 GPUs. The learning rate linearly warms up over the first 10 epochs to a base value of 0.004, then decays following a cosine schedule [55]. Additionally, a layer-wise learning rate decay of 0.9 is applied to model parameters [111]. Weight decay is also controlled by a cosine schedule, progressively increasing from 0.04 to 0.2. For EMA, the student temperature is set to 0.1, while the teacher temperature gradually rises from 0.04 to 0.07 during the first 10 epochs [60]. The momentum starts at 0.994 and increases to 1 by the final iteration. For data augmentation and view generation, we follow the augmentation pipeline designed by MSC [88]. We generate 2 global views (sampling " + }, + { + "bbox": [ + 55, + 510, + 295, + 714 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 55, + 510, + 295, + 714 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 510, + 295, + 714 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 55, + 510, + 295, + 714 + ], + "type": "text", + "content": " of scene points) and 4 local views (sampling " + }, + { + "bbox": [ + 55, + 510, + 295, + 714 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 55, + 510, + 295, + 714 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 510, + 295, + 714 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 55, + 510, + 295, + 714 + ], + "type": "text", + "content": " of scene points) for training, with 2 masked views generated based on the global views." + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 317, + 70, + 375, + 127 + ], + "blocks": [ + { + "bbox": [ + 317, + 70, + 375, + 127 + ], + "lines": [ + { + "bbox": [ + 317, + 70, + 375, + 127 + ], + "spans": [ + { + "bbox": [ + 317, + 70, + 375, + 127 + ], + "type": "image", + "image_path": "06aa10040be016adeacb37fd5767052dffa88a788de5bc8d35dfdbf31a2a7d1a.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 376, + 70, + 553, + 128 + ], + "blocks": [ + { + "bbox": [ + 376, + 70, + 553, + 128 + ], + "lines": [ + { + "bbox": [ + 376, + 70, + 553, + 128 + ], + "spans": [ + { + "bbox": [ + 376, + 70, + 553, + 128 + ], + "type": "image", + "image_path": "71d19533408976e6644598cdf508d18a7c5e44b67d330f9724f037577f5ccdcb.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 317, + 128, + 375, + 185 + ], + "blocks": [ + { + "bbox": [ + 317, + 128, + 375, + 185 + ], + "lines": [ + { + "bbox": [ + 317, + 128, + 375, + 185 + ], + "spans": [ + { + "bbox": [ + 317, + 128, + 375, + 185 + ], + "type": "image", + "image_path": "dcb26bf0cf2ce76820349c2c9376b5eed8bb8c499453a344a54c3494c4ab1d93.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 376, + 129, + 553, + 186 + ], + "blocks": [ + { + "bbox": [ + 376, + 129, + 553, + 186 + ], + "lines": [ + { + "bbox": [ + 376, + 129, + 553, + 186 + ], + "spans": [ + { + "bbox": [ + 376, + 129, + 553, + 186 + ], + "type": "image", + "image_path": "82222571f32c270eb9e8c30df905551ea5b2080d6d6f0e6479638834f53e2866.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 317, + 186, + 375, + 243 + ], + "blocks": [ + { + "bbox": [ + 317, + 186, + 375, + 243 + ], + "lines": [ + { + "bbox": [ + 317, + 186, + 375, + 243 + ], + "spans": [ + { + "bbox": [ + 317, + 186, + 375, + 243 + ], + "type": "image", + "image_path": "44a957cddc13afa6e27b5da41df947c9bf978364d7f07cc157fac6052703723a.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 247, + 553, + 314 + ], + "lines": [ + { + "bbox": [ + 313, + 247, + 553, + 314 + ], + "spans": [ + { + "bbox": [ + 313, + 247, + 553, + 314 + ], + "type": "text", + "content": "Figure 7. Zero-shot comparison with DINOv2. We compare the PCA visualizations of DINOv2, Sonata, and their combined feature representation. DINOv2 excels at capturing photometric details, while Sonata better distinguishes spatial information. The combined model demonstrates improved coherence and detail, showcasing the complementary strengths of both models." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 376, + 186, + 553, + 243 + ], + "blocks": [ + { + "bbox": [ + 376, + 186, + 553, + 243 + ], + "lines": [ + { + "bbox": [ + 376, + 186, + 553, + 243 + ], + "spans": [ + { + "bbox": [ + 376, + 186, + 553, + 243 + ], + "type": "image", + "image_path": "aa552af8ddd0526f922499b890339fd7daffb2482750d8c8e5ac2f10a3aa2d92.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 323, + 553, + 358 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 323, + 553, + 358 + ], + "spans": [ + { + "bbox": [ + 313, + 323, + 553, + 358 + ], + "type": "text", + "content": "Evaluation. We evaluate the quality of the learned representation using the following three protocols after initializing the encoder with Sonata:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 319, + 359, + 553, + 491 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 319, + 359, + 553, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 359, + 553, + 407 + ], + "spans": [ + { + "bbox": [ + 319, + 359, + 553, + 407 + ], + "type": "text", + "content": "- In linear probing, we keep the encoder frozen and upcast the features to their original scale. A single linear layer, comprising less than " + }, + { + "bbox": [ + 319, + 359, + 553, + 407 + ], + "type": "inline_equation", + "content": "0.2\\%" + }, + { + "bbox": [ + 319, + 359, + 553, + 407 + ], + "type": "text", + "content": " of the total parameters, is then used to adapt these features to downstream tasks." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 319, + 407, + 553, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 407, + 553, + 455 + ], + "spans": [ + { + "bbox": [ + 319, + 407, + 553, + 455 + ], + "type": "text", + "content": "- In decoder probing, we take a step back and reintroduce a lightweight hierarchical decoder, which accounts for " + }, + { + "bbox": [ + 319, + 407, + 553, + 455 + ], + "type": "inline_equation", + "content": "13\\%" + }, + { + "bbox": [ + 319, + 407, + 553, + 455 + ], + "type": "text", + "content": " of the total parameters. We then freeze the encoder, allowing only the decoder to actively learn." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 319, + 455, + 553, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 455, + 553, + 491 + ], + "spans": [ + { + "bbox": [ + 319, + 455, + 553, + 491 + ], + "type": "text", + "content": "- In full fine-tuning, we follow the traditional approach by unfreezing the entire PTv3 U-Net-style backbone, tuning the learned representation to downstream tasks." + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 491, + 554, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 491, + 554, + 538 + ], + "spans": [ + { + "bbox": [ + 313, + 491, + 554, + 538 + ], + "type": "text", + "content": "We advocate linear probing as the primary evaluation criterion for point SSL, considering other methods as intermediate steps. We look forward to the day when a linear-probed self-supervised model outperforms a fully fine-tuned one." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 544, + 397, + 555 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 544, + 397, + 555 + ], + "spans": [ + { + "bbox": [ + 313, + 544, + 397, + 555 + ], + "type": "text", + "content": "5. Main Results" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 558, + 554, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 554, + 593 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 554, + 593 + ], + "type": "text", + "content": "We validate the reliability of the Sonata representation using the evaluation protocols discussed in Sec. 4.3 and analyze the main properties based on these results." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "content": "Comparison with image self-supervised model. In Tab. 3, we compare the linear probing and decoder probing results on ScanNet and ScanNet200 semantic segmentation with the linear probing accuracy of representations transferred from the image self-supervised models DINOv2 [60] and DINOv2.5 [24]. Specifically, we aggregate unprojected pixel embeddings using ground truth camera poses and depth calculated by ray intersection with a reconstructed mesh, which provides more accuracy than sensor per-frame depth. Our results indicate that while the DINOs' represent" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22198" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 70, + 280, + 194 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 280, + 194 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 280, + 194 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 280, + 194 + ], + "type": "table", + "html": "
2D × 3DScanNet Val [23]ScanNet200 Val [23]
MethodsmIoUmAccallAccmIoUmAccallAcc
• DINOv2 (lin.) [60]63.0975.5082.4227.4237.5972.80
• DINOv2.5 (lin.) [24]63.3675.9482.3027.7539.2372.53
• Sonata (lin.)72.5283.1189.7429.2541.6181.15
+DINOv2 (lin.)75.9185.3691.2536.6746.9882.85
+DINOv2.5 (lin.)76.4485.6891.3336.9648.2382.77
• Sonata (dec.)79.0786.5792.6833.5444.4884.07
+DINOv2 (dec.)79.1287.2392.4737.7349.3883.31
+DINOv2.5 (dec.)79.1986.6692.5038.2748.5783.77
", + "image_path": "3c9033ea0b3d84da3d0cd5cd522a5addb74d470841676098228016ea58820fd8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 285, + 72, + 553, + 194 + ], + "blocks": [ + { + "bbox": [ + 70, + 197, + 262, + 207 + ], + "lines": [ + { + "bbox": [ + 70, + 197, + 262, + 207 + ], + "spans": [ + { + "bbox": [ + 70, + 197, + 262, + 207 + ], + "type": "text", + "content": "Table 3. Numerical comparison with DINO series." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 285, + 72, + 553, + 194 + ], + "lines": [ + { + "bbox": [ + 285, + 72, + 553, + 194 + ], + "spans": [ + { + "bbox": [ + 285, + 72, + 553, + 194 + ], + "type": "table", + "html": "
Data EfficiencyLimited Scenes (Pct.)Limited Annotation (Pts.)
Methods1%5%10%20%Full2050100200Full
○ SparseUNet [17]26.047.856.762.972.241.953.962.265.572.2
• CSC [38]28.949.859.464.673.855.560.565.968.273.8
• MSC [88]29.250.761.064.975.461.065.668.969.675.4
○ PTv2 [87]24.848.159.866.375.458.466.170.371.275.4
○ PTv3 [89]25.848.961.067.077.260.167.971.472.777.2
• PPT [90] (sup.)31.152.663.368.278.262.469.174.375.578.2
• Sonata (lin.)43.662.568.669.872.569.070.571.171.572.5
• Sonata (dec.)44.564.169.872.579.169.873.175.076.379.1
• Sonata (full)45.365.772.472.879.470.573.676.077.079.4
", + "image_path": "74276ce4432bdf579eebed4474ab5b041af07d66bcc013f40763b9bafac876fe.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 56, + 211, + 550, + 335 + ], + "blocks": [ + { + "bbox": [ + 372, + 197, + 465, + 207 + ], + "lines": [ + { + "bbox": [ + 372, + 197, + 465, + 207 + ], + "spans": [ + { + "bbox": [ + 372, + 197, + 465, + 207 + ], + "type": "text", + "content": "Table 4. Data efficiency." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 211, + 550, + 335 + ], + "lines": [ + { + "bbox": [ + 56, + 211, + 550, + 335 + ], + "spans": [ + { + "bbox": [ + 56, + 211, + 550, + 335 + ], + "type": "table", + "html": "
Param. Efficiency MethodsParamsScanNet Val [23]ScanNet200 Val [71]ScanNet++ Val [101]S3DIS Area 5 [1]S3DIS 6-fold [1]
Learn.Pct.mIoUmAccallAccmIoUmAccallAccmIoUmAccallAccmIoUmAccallAccmIoUmAccallAcc
○ SparseUNet [17]39.2M100%72.380.290.025.032.980.428.838.480.166.372.589.872.480.989.9
• PC [93] (lin.)<0.2M<0.1%5.69.750.00.50.940.31.83.146.411.418.652.311.719.051.2
• CSC [38] (lin.)<0.2M<0.1%12.618.164.21.32.153.02.84.553.624.432.066.424.932.566.9
• MSC [88] (lin.)<0.2M<0.1%14.120.362.91.52.553.64.56.661.327.935.571.129.937.971.3
○PTv3 [89]124.8M100%77.685.092.035.346.083.442.153.485.673.478.991.777.785.391.5
• MSC [88] (lin.)<0.2M<0.2%21.832.265.53.35.557.58.111.964.732.142.470.934.646.071.3
• Sonata (lin.)<0.2M<0.2%72.583.189.729.341.681.237.350.984.372.381.290.976.587.490.8
• Sonata (dec.)16.3M13%79.186.692.733.544.584.140.952.686.374.580.492.681.588.893.0
", + "image_path": "673a21537d1f42efbee2d20b90e3cb6de7f05fe926829185f23d551e69643274.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 242, + 337, + 357, + 348 + ], + "lines": [ + { + "bbox": [ + 242, + 337, + 357, + 348 + ], + "spans": [ + { + "bbox": [ + 242, + 337, + 357, + 348 + ], + "type": "text", + "content": "Table 5. Parameter efficiency." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 355, + 294, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 355, + 294, + 464 + ], + "spans": [ + { + "bbox": [ + 54, + 355, + 294, + 464 + ], + "type": "text", + "content": "tation demonstrates impressive robustness, Sonata offers a more suitable representation for 3D tasks, achieving " + }, + { + "bbox": [ + 54, + 355, + 294, + 464 + ], + "type": "inline_equation", + "content": "+9.2\\%" + }, + { + "bbox": [ + 54, + 355, + 294, + 464 + ], + "type": "text", + "content": " on ScanNet and " + }, + { + "bbox": [ + 54, + 355, + 294, + 464 + ], + "type": "inline_equation", + "content": "+1.5\\%" + }, + { + "bbox": [ + 54, + 355, + 294, + 464 + ], + "type": "text", + "content": " on ScanNet200 for semantic segmentation. Furthermore, combining Sonata with the DINOs yields higher accuracy than any single data modality alone " + }, + { + "bbox": [ + 54, + 355, + 294, + 464 + ], + "type": "inline_equation", + "content": "(+3.9\\%" + }, + { + "bbox": [ + 54, + 355, + 294, + 464 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 355, + 294, + 464 + ], + "type": "inline_equation", + "content": "+7.7\\%" + }, + { + "bbox": [ + 54, + 355, + 294, + 464 + ], + "type": "text", + "content": ", respectively), underscoring the promising potential of cross-modal self-distillation. Additional zero-shot comparisons (see Fig. 7) through PCA visualizations further corroborate these numerical findings." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 468, + 295, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 468, + 295, + 589 + ], + "spans": [ + { + "bbox": [ + 54, + 468, + 295, + 589 + ], + "type": "text", + "content": "Data efficiency. In Tab. 4, we present the semantic segmentation performance of Sonata when probed or fine-tuned on the ScanNet dataset with limited scenes and annotations [38]. The results demonstrate the exceptional data efficiency of Sonata, with improvements of " + }, + { + "bbox": [ + 54, + 468, + 295, + 589 + ], + "type": "inline_equation", + "content": "19.5\\%" + }, + { + "bbox": [ + 54, + 468, + 295, + 589 + ], + "type": "text", + "content": " in extreme data scarcity (1% of scenes) and " + }, + { + "bbox": [ + 54, + 468, + 295, + 589 + ], + "type": "inline_equation", + "content": "10.4\\%" + }, + { + "bbox": [ + 54, + 468, + 295, + 589 + ], + "type": "text", + "content": " in limited annotation scenarios (20 points per scene) compared to training from scratch. Notably, even linear probing surpasses previous SOTA by a substantial margin (12.5% with 1% scenes), highlighting Sonata's reliability in low-data scenarios." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 594, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 594, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 594, + 295, + 714 + ], + "type": "text", + "content": "Parameter efficiency. In Tab. 5, we demonstrate parameter efficiency using both linear and decoder probing across various indoor semantic segmentation benchmarks, including ScanNet [23], ScanNet200 [71], ScanNet++ [101], S3DIS Area5[1], and S3DIS 6-fold cross-validation [65]. Semantic segmentation is emphasized as it provides a direct measure of point cloud representation quality. Results show that a single linear layer with a negligible number of parameters (<0.02% of total parameters) is sufficient for Sonata to achieve strong performance on these benchmarks (e.g.," + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 355, + 553, + 463 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 355, + 553, + 463 + ], + "spans": [ + { + "bbox": [ + 313, + 355, + 553, + 463 + ], + "type": "text", + "content": "72.5% on ScanNet and 73.4% on S3DIS Area5). Furthermore, probing a decoder with only 13% of the model's parameters yields even higher accuracy (e.g., 79.1% on ScanNet and 81.5% on S3DIS 6-fold cross-validation). However, while decoder probing achieves SOTA results on ScanNet (20 classes) and S3DIS (13 classes), performance on ScanNet200 (200 classes) and ScanNet++ (100 classes) remains limited. This shows a limitation of the learned representation in distinguishing a large number of classes." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 468, + 554, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 468, + 554, + 588 + ], + "spans": [ + { + "bbox": [ + 313, + 468, + 554, + 588 + ], + "type": "text", + "content": "Indoor semantic segmentation. In Tab. 6, we further enhance Sonata's semantic segmentation accuracy through full fine-tuning, consistently pushing SOTA results to new heights across the five widely recognized benchmarks, e.g., achieving " + }, + { + "bbox": [ + 313, + 468, + 554, + 588 + ], + "type": "inline_equation", + "content": "79.4\\%" + }, + { + "bbox": [ + 313, + 468, + 554, + 588 + ], + "type": "text", + "content": " on ScanNet and " + }, + { + "bbox": [ + 313, + 468, + 554, + 588 + ], + "type": "inline_equation", + "content": "82.3\\%" + }, + { + "bbox": [ + 313, + 468, + 554, + 588 + ], + "type": "text", + "content": " on S3DIS 6-fold cross-validation. However, we view full fine-tuning as an intermediate step toward a future where linear probing surpasses it. At present, full fine-tuning remains essential to achieve the highest performance on these benchmarks and close the remaining gap of " + }, + { + "bbox": [ + 313, + 468, + 554, + 588 + ], + "type": "inline_equation", + "content": "7.1\\%" + }, + { + "bbox": [ + 313, + 468, + 554, + 588 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 468, + 554, + 588 + ], + "type": "inline_equation", + "content": "5.8\\%" + }, + { + "bbox": [ + 313, + 468, + 554, + 588 + ], + "type": "text", + "content": " respectively." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "content": "Indoor instance segmentation. In Tab. 7, We also validate the robustness of Sonata representation on indoor instance segmentation benchmarks, including ScanNet [23], ScanNet200 [71], ScanNet++ [101], and S3DIS [1]. Consistent with our findings in semantic segmentation, Sonata demonstrates strong parameter efficiency, achieving significant improvements with linear probing (10× mAP50 on ScanNet and 21× on ScanNet200) and decoder probing (12× mAP50 on ScanNet and 33× on ScanNet200). Full fine-tuning further boosts these results, achieving SOTA" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22199" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 70, + 552, + 194 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 552, + 194 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 552, + 194 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 552, + 194 + ], + "type": "table", + "html": "
Indoor Sem. SegParamsScanNet Val [23]ScanNet200 Val [71]ScanNet++ Val [101]S3DIS Area 5 [1]S3DIS 6-fold [1]
MethodsLearn.Pct.mIoUmAccallAccmIoUmAccallAccmIoUmAccallAccmIoUmAccallAccmIoUmAccallAcc
○ SparseUNet [17]39.2M100%72.380.290.025.032.980.428.838.480.166.372.589.872.480.989.9
• PC [93]39.2M100%72.380.990.126.233.079.929.239.782.768.173.590.074.783.390.6
• CSC [38]39.2M100%72.881.090.726.933.780.632.541.183.770.776.490.875.584.090.9
• MSC [88]39.2M100%75.783.491.332.041.682.339.449.684.970.776.191.077.485.391.5
○PTv3 [89]124.8M100%77.685.092.035.346.083.442.153.485.673.478.991.777.785.391.5
• MSC [88]124.8M100%78.285.392.233.443.783.442.453.685.969.974.991.277.484.791.7
• PPT [90] (sup.)124.8M100%78.685.992.336.046.283.843.355.786.474.380.192.080.887.792.6
• Sonata124.8M100%79.486.192.536.846.584.443.755.886.676.081.693.082.389.993.3
", + "image_path": "0a34a86f99f1624747ff060519ccc3e421dfeb2b64d76c2e1d805ff2a8146b9f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 56, + 210, + 552, + 326 + ], + "blocks": [ + { + "bbox": [ + 223, + 197, + 376, + 207 + ], + "lines": [ + { + "bbox": [ + 223, + 197, + 376, + 207 + ], + "spans": [ + { + "bbox": [ + 223, + 197, + 376, + 207 + ], + "type": "text", + "content": "Table 6. Indoor semantic segmentation." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 210, + 552, + 326 + ], + "lines": [ + { + "bbox": [ + 56, + 210, + 552, + 326 + ], + "spans": [ + { + "bbox": [ + 56, + 210, + 552, + 326 + ], + "type": "table", + "html": "
Indoor Ins. SegParamsScanNet Val [23]ScanNet200 Val [71]ScanNet++ Val [101]S3DIS Area 5 [1]
MethodsLearn.Pct.mAP25mAP50mAPmAP25mAP50mAPmAP25mAP50mAPmAP25mAP50mAP
○PointGroup [45]124.8M100%77.561.740.940.133.223.136.730.720.955.749.437.8
•MSC (lin.)<0.2M<0.2%13.35.32.32.31.00.44.82.61.319.013.09.7
•Sonata (lin.)<0.2M<0.2%72.653.930.730.921.310.931.622.412.245.836.626.1
•Sonata (dec.)16.3M13%76.862.840.840.833.322.838.129.118.863.757.145.1
•MSC (full)124.8M100%78.462.941.140.533.823.438.930.921.756.350.538.1
•PPT [90] (sup.)124.8M100%78.963.542.140.834.124.039.332.821.957.551.239.7
•Sonata (full)124.8M100%79.263.942.442.135.625.440.333.622.363.857.445.5
", + "image_path": "725e6fcaf9b132f2531998223fa93a4dd29f8794ba8eba66c8f5d2f77abbddb6.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 56, + 347, + 294, + 438 + ], + "blocks": [ + { + "bbox": [ + 225, + 327, + 374, + 338 + ], + "lines": [ + { + "bbox": [ + 225, + 327, + 374, + 338 + ], + "spans": [ + { + "bbox": [ + 225, + 327, + 374, + 338 + ], + "type": "text", + "content": "Table 7. Indoor instance segmentation." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 347, + 294, + 438 + ], + "lines": [ + { + "bbox": [ + 56, + 347, + 294, + 438 + ], + "spans": [ + { + "bbox": [ + 56, + 347, + 294, + 438 + ], + "type": "table", + "html": "
Outdoor Sem. Seg.nuScenes Val [9]Waymo Val [74]Sem.KITTI Val [6]
MethodsmIoUmAccallAccmIoUmAccallAccmIoUmAcc allAcc
o PTv3 [89]80.487.294.771.380.594.769.176.192.6
• Sonata (lin.)66.177.292.460.572.592.562.072.591.0
• Sonata (dec.)77.385.994.270.878.894.368.476.592.3
• PPT [90] (sup.)81.287.594.872.181.394.872.377.593.4
• Sonata (full)81.787.995.072.981.994.972.677.993.4
", + "image_path": "6ec5eca4dfc1955ebcbdb48b12537d4a84c76555532556618daddc271253a44b.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 96, + 440, + 256, + 451 + ], + "lines": [ + { + "bbox": [ + 96, + 440, + 256, + 451 + ], + "spans": [ + { + "bbox": [ + 96, + 440, + 256, + 451 + ], + "type": "text", + "content": "Table 8. Outdoor semantic segmentation." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 455, + 295, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 455, + 295, + 490 + ], + "spans": [ + { + "bbox": [ + 55, + 455, + 295, + 490 + ], + "type": "text", + "content": "pre-training performance across benchmarks. This clearly demonstrates that, unlike previous SSL approaches, the Sonata representation encodes instance-level information." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 491, + 295, + 646 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 491, + 295, + 646 + ], + "spans": [ + { + "bbox": [ + 54, + 491, + 295, + 646 + ], + "type": "text", + "content": "Outdoor semantic segmentation. In Tab. 8, we adapt pre-training paradigm of Sonata to outdoor LiDAR scenarios through joint training on nuScenes [9], Waymo [74], and SemanticKITTI [6] and evaluate semantic segmentation performance using our evaluation protocols. With linear probing, Sonata sets a robust parameter efficiency baseline. Decoder probing achieves significant gains. In full fine-tuning, Sonata surpasses the supervised PPT [90], establishing new SOTA mIoU scores of 81.7, 72.9, and 72.6 across these benchmarks, underscoring the effectiveness of Sonata in outdoor perception tasks. Note that most of that performance can be recovered by more efficient decoder-only probing with a " + }, + { + "bbox": [ + 54, + 491, + 295, + 646 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 54, + 491, + 295, + 646 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 54, + 491, + 295, + 646 + ], + "type": "inline_equation", + "content": "97\\%" + }, + { + "bbox": [ + 54, + 491, + 295, + 646 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 54, + 491, + 295, + 646 + ], + "type": "inline_equation", + "content": "94\\%" + }, + { + "bbox": [ + 54, + 491, + 295, + 646 + ], + "type": "text", + "content": " of full fine-tuning." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 651, + 208, + 663 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 651, + 208, + 663 + ], + "spans": [ + { + "bbox": [ + 55, + 651, + 208, + 663 + ], + "type": "text", + "content": "6. Conclusion and Discussion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "In this work, we make progress towards self-supervised learning of a strong and reliable 3D point representation, that can zero-shot correspond semantically similar 3D points, down to the instance level, even under the presence" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 346, + 555, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 346, + 555, + 574 + ], + "spans": [ + { + "bbox": [ + 313, + 346, + 555, + 574 + ], + "type": "text", + "content": "of spatial and visual perturbations. We demonstrate that such a representation can serve as the foundation for 3D tasks in semantic and instance-level grouping. We find existing 3D self-supervised learning approaches lacking, and hypothesize that this is due to what we term the geometric shortcut, a problem unique to 3D, which causes representations to collapse to low-level spatial features. We demonstrate the deficiency of these collapsed representations through linear probing. Beginning with a point self-distillation framework, we tackle the geometric shortcut by attaching SSL losses at coarser spatial scales, disturbing the spatial information of masked points with no features, and progressively increasing task difficulty to prevent overreliance on accessible geometric cues. This change enables effective scaling up, ultimately composing a Sonata from 140k point clouds. Sonata demonstrates semantically meaningful zero-shot visualization, as well as exceptional parameter and data efficiency. Full fine-tuning further advances SOTA across 3D indoor and outdoor perception tasks." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 556, + 715 + ], + "type": "text", + "content": "We summarize our limitations and future works and encourage readers to refer to the supplementary for details: 1. Sonata representations could be further semantically enriched through augmentation with 1M object-level point clouds [53]; 2. Pursuing unified training across indoor and outdoor scenarios as a feasible and promising direction; 3. Scaling up training with pixel-aligned [8] and SLAM-generated [27] point clouds from video data; 4. Leveraging cross-modal distillation to allow Sonata and DINOV2 to complement each other. We hope our insights will contribute to future research in 3D representation learning." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22200" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 152 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 152 + ], + "type": "text", + "content": "We extend our gratitude to Maxime Oquab and Piotr Bojanowski for their guidance on the DINOv2 training details, to Saining Xie for insightful discussions on the vision of 3D representation learning, and to Paul Mcvay for his thoughts on the JPEA framework with sparse point cloud data." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 161, + 115, + 173 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 161, + 115, + 173 + ], + "spans": [ + { + "bbox": [ + 56, + 161, + 115, + 173 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 180, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 66, + 180, + 296, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 180, + 296, + 224 + ], + "spans": [ + { + "bbox": [ + 66, + 180, + 296, + 224 + ], + "type": "text", + "content": "[1] Iro Armeni, Ozan Sener, Amir R. Zamir, Helen Jiang, Ioannis Brilakis, Martin Fischer, and Silvio Savarese. 3d semantic parsing of large-scale indoor spaces. In CVPR, 2016. 2, 6, 7, 8, 16" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 66, + 225, + 296, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 225, + 296, + 278 + ], + "spans": [ + { + "bbox": [ + 66, + 225, + 296, + 278 + ], + "type": "text", + "content": "[2] Mahmoud Assran, Quentin Duval, Ishan Misra, Piotr Bojanowski, Pascal Vincent, Michael Rabbat, Yann LeCun, and Nicolas Ballas. Self-supervised learning from images with a joint-embedding predictive architecture. In CVPR, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 66, + 280, + 295, + 345 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 280, + 295, + 345 + ], + "spans": [ + { + "bbox": [ + 66, + 280, + 295, + 345 + ], + "type": "text", + "content": "[3] Armen Avetisyan, Christopher Xie, Henry Howard-Jenkins, Tsun-Yi Yang, Samir Aroudj, Suvam Patra, Fuyang Zhang, Duncan Frost, Luke Holland, Campbell Orme, et al. Scenescript: Reconstructing scenes with an autoregressive structured language model. In ECCV, 2024. 2, 6, 16" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 66, + 346, + 294, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 346, + 294, + 369 + ], + "spans": [ + { + "bbox": [ + 66, + 346, + 294, + 369 + ], + "type": "text", + "content": "[4] Jimmy Lei Ba, Jamie Ryan Kiros, and Geoffrey E Hinton. Layer normalization. Stat, 2016. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 66, + 369, + 296, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 369, + 296, + 434 + ], + "spans": [ + { + "bbox": [ + 66, + 369, + 296, + 434 + ], + "type": "text", + "content": "[5] Gilad Baruch, Zhuoyuan Chen, Afshin Dehghan, Tal Dimry, Yuri Feigin, Peter Fu, Thomas Gebauer, Brandon Joffe, Daniel Kurz, Arik Schwartz, and Elad Shulman. ARKitscenes - a diverse real-world dataset for 3d indoor scene understanding using mobile RGB-d data. In NeurIPSW, 2021. 2, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 66, + 436, + 296, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 436, + 296, + 479 + ], + "spans": [ + { + "bbox": [ + 66, + 436, + 296, + 479 + ], + "type": "text", + "content": "[6] Jens Behley, Martin Garbade, Andres Milioto, Jan Quenzel, Sven Behnke, Cyril Stachniss, and Jurgen Gall. Semantickitti: A dataset for semantic scene understanding of lidar sequences. In ICCV, 2019. 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 66, + 480, + 294, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 480, + 294, + 503 + ], + "spans": [ + { + "bbox": [ + 66, + 480, + 294, + 503 + ], + "type": "text", + "content": "[7] Yoshua Bengio, Jérôme Louradour, Ronan Collobert, and Jason Weston. Curriculum learning. In ICML, 2009. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 66, + 503, + 294, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 503, + 294, + 557 + ], + "spans": [ + { + "bbox": [ + 66, + 503, + 294, + 557 + ], + "type": "text", + "content": "[8] Aleksei Bochkovskii, Amael Delaunoy, Hugo Germain, Marcel Santos, Yichao Zhou, Stephan R Richter, and Vladlen Koltun. Depth pro: Sharp monocular metric depth in less than a second. arXiv preprint arXiv:2410.02073, 2024. 8, 13" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 66, + 558, + 294, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 558, + 294, + 612 + ], + "spans": [ + { + "bbox": [ + 66, + 558, + 294, + 612 + ], + "type": "text", + "content": "[9] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In CVPR, 2020. 2, 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 61, + 613, + 294, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 613, + 294, + 646 + ], + "spans": [ + { + "bbox": [ + 61, + 613, + 294, + 646 + ], + "type": "text", + "content": "[10] Mathilde Caron, Piotr Bojanowski, Armand Joulin, and Matthijs Douze. Deep clustering for unsupervised learning of visual features. In ECCV, 2018. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 61, + 647, + 294, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 647, + 294, + 690 + ], + "spans": [ + { + "bbox": [ + 61, + 647, + 294, + 690 + ], + "type": "text", + "content": "[11] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. In NeurIPS, 2020. 2, 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 61, + 691, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 691, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 61, + 691, + 294, + 713 + ], + "type": "text", + "content": "[12] Mathilde Caron, Hugo Touvron, Ishan Misra, Herve Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin." + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 320, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 34, + "blocks": [ + { + "bbox": [ + 339, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 339, + 73, + 553, + 95 + ], + "type": "text", + "content": "Emerging properties in self-supervised vision transformers. In CVPR, 2021. 2, 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 321, + 96, + 553, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 96, + 553, + 129 + ], + "spans": [ + { + "bbox": [ + 321, + 96, + 553, + 129 + ], + "type": "text", + "content": "[13] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In ICML, 2020. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 320, + 131, + 553, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 131, + 553, + 163 + ], + "spans": [ + { + "bbox": [ + 320, + 131, + 553, + 163 + ], + "type": "text", + "content": "[14] Xi Chen, Lianghua Huang, Yu Liu, Yujun Shen, Deli Zhao, and Hengshuang Zhao. Anydoor: Zero-shot object-level image customization. In CVPR, 2024. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 320, + 165, + 553, + 198 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 165, + 553, + 198 + ], + "spans": [ + { + "bbox": [ + 320, + 165, + 553, + 198 + ], + "type": "text", + "content": "[15] Yukang Chen, Jianhui Liu, Xiangyu Zhang, Xiaojuan Qi, and Jiaya Jia. Largekernel3d: Scaling up kernels in 3d sparse cnns. In CVPR, 2023. 16" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 320, + 200, + 553, + 233 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 200, + 553, + 233 + ], + "spans": [ + { + "bbox": [ + 320, + 200, + 553, + 233 + ], + "type": "text", + "content": "[16] Hung-Yueh Chiang, Yen-Liang Lin, Yueh-Cheng Liu, and Winston H Hsu. A unified point-based framework for 3d segmentation. In 3DV, 2019. 16" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 320, + 234, + 553, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 234, + 553, + 267 + ], + "spans": [ + { + "bbox": [ + 320, + 234, + 553, + 267 + ], + "type": "text", + "content": "[17] Christopher Choy, JunYoung Gwak, and Silvio Savarese. 4d spatio-temporal convnets: Minkowski convolutional neural networks. In CVPR, 2019. 3, 7, 8, 16, 17" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 320, + 269, + 553, + 291 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 269, + 553, + 291 + ], + "spans": [ + { + "bbox": [ + 320, + 269, + 553, + 291 + ], + "type": "text", + "content": "[18] Christopher Choy, Jaesik Park, and Vladlen Koltun. Fully convolutional geometric features. In ICCV, 2019. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 320, + 293, + 553, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 293, + 553, + 325 + ], + "spans": [ + { + "bbox": [ + 320, + 293, + 553, + 325 + ], + "type": "text", + "content": "[19] Pointcept Contributors. Pointcept: A codebase for point cloud perception research. https://github.com/Pointcept/Pointcept, 2023.5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 320, + 327, + 553, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 327, + 553, + 360 + ], + "spans": [ + { + "bbox": [ + 320, + 327, + 553, + 360 + ], + "type": "text", + "content": "[20] Spconv Contributors. Spconv: Spatially sparse convolution library. https://github.com/traveller59/spconv, 2022.3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 320, + 361, + 553, + 405 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 361, + 553, + 405 + ], + "spans": [ + { + "bbox": [ + 320, + 361, + 553, + 405 + ], + "type": "text", + "content": "[21] Brian Curless and Marc Levoy. A volumetric method for building complex models from range images. In Proceedings of the 23rd annual conference on Computer graphics and interactive techniques, pages 303-312, 1996. 16" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 320, + 407, + 553, + 438 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 407, + 553, + 438 + ], + "spans": [ + { + "bbox": [ + 320, + 407, + 553, + 438 + ], + "type": "text", + "content": "[22] Angela Dai and Matthias Nießner. 3dmv: Joint 3d-multiview prediction for 3d semantic scene segmentation. In ECCV, 2018. 16" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 320, + 441, + 553, + 484 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 441, + 553, + 484 + ], + "spans": [ + { + "bbox": [ + 320, + 441, + 553, + 484 + ], + "type": "text", + "content": "[23] Angela Dai, Angel X. Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In CVPR, 2017. 2, 5, 6, 7, 8, 16" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 320, + 486, + 553, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 486, + 553, + 518 + ], + "spans": [ + { + "bbox": [ + 320, + 486, + 553, + 518 + ], + "type": "text", + "content": "[24] Timothee Darcet, Maxime Oquab, Julien Mairal, and Piotr Bojanowski. Vision transformers need registers. arXiv:2309.16588, 2023. 6, 7" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 320, + 520, + 553, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 520, + 553, + 553 + ], + "spans": [ + { + "bbox": [ + 320, + 520, + 553, + 553 + ], + "type": "text", + "content": "[25] Carl Doersch, Abhinav Gupta, and Alexei A Efros. Unsupervised visual representation learning by context prediction. In ICCV, 2015. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 320, + 555, + 553, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 555, + 553, + 599 + ], + "spans": [ + { + "bbox": [ + 320, + 555, + 553, + 599 + ], + "type": "text", + "content": "[26] Alexey Dosovitskiy, Philipp Fischer, Jost Tobias Springenberg, Martin Riedmiller, and Thomas Brox. Discriminative unsupervised feature learning with exemplar convolutional neural networks. TPAMI, 2015. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 320, + 600, + 553, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 600, + 553, + 655 + ], + "spans": [ + { + "bbox": [ + 320, + 600, + 553, + 655 + ], + "type": "text", + "content": "[27] Jakob Engel, Kiran Somasundaram, Michael Goesele, Albert Sun, Alexander Gamino, Andrew Turner, Arjang Talattof, Arnie Yuan, Bilal Souti, Brighid Meredith, et al. Project aria: A new tool for egocentric multi-modal ai research. arXiv:2308.13561, 2023. 2, 8, 13" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 320, + 656, + 553, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 656, + 553, + 679 + ], + "spans": [ + { + "bbox": [ + 320, + 656, + 553, + 679 + ], + "type": "text", + "content": "[28] Pedro F Felzenszwalb and Daniel P Huttenlocher. Efficient graph-based image segmentation. IJCV, 2004. 3" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 320, + 680, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 680, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 320, + 680, + 553, + 713 + ], + "type": "text", + "content": "[29] Benjamin Graham and Laurens van der Maaten. Submanifold sparse convolutional networks. arXiv:1706.01307, 2017. 3" + } + ] + } + ], + "index": 33 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22201" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 61, + 73, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 73, + 295, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 73, + 295, + 106 + ], + "spans": [ + { + "bbox": [ + 61, + 73, + 295, + 106 + ], + "type": "text", + "content": "[30] Benjamin Graham, Martin Engelcke, and Laurens van der Maaten. 3d semantic segmentation with submanifold sparse convolutional networks. In CVPR, 2018. 16" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 107, + 295, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 107, + 295, + 162 + ], + "spans": [ + { + "bbox": [ + 61, + 107, + 295, + 162 + ], + "type": "text", + "content": "[31] Jean-Bastien Grill, Florian Strub, Florent Alché, Coretin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. In NerIPS, 2020. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 163, + 295, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 163, + 295, + 217 + ], + "spans": [ + { + "bbox": [ + 62, + 163, + 295, + 217 + ], + "type": "text", + "content": "[32] Jiayuan Gu, Fanbo Xiang, Xuanlin Li, Zhan Ling, Xiqiang Liu, Tongzhou Mu, Yihe Tang, Stone Tao, Xinyue Wei, Yunchao Yao, Xiaodi Yuan, Pengwei Xie, Zhiao Huang, Rui Chen, and Hao Su. Maniskill2: A unified benchmark for generalizable manipulation skills. In ICLR, 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 219, + 294, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 219, + 294, + 251 + ], + "spans": [ + { + "bbox": [ + 62, + 219, + 294, + 251 + ], + "type": "text", + "content": "[33] Bharath Hariharan, Pablo Arbelaez, Ross Girshick, and Jitendra Malik. Hypercolumns for object segmentation and fine-grained localization. In CVPR, 2015. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 253, + 294, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 253, + 294, + 285 + ], + "spans": [ + { + "bbox": [ + 62, + 253, + 294, + 285 + ], + "type": "text", + "content": "[34] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In CVPR, 2020. 2, 3, 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 286, + 294, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 286, + 294, + 319 + ], + "spans": [ + { + "bbox": [ + 62, + 286, + 294, + 319 + ], + "type": "text", + "content": "[35] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In CVPR, 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 321, + 294, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 321, + 294, + 374 + ], + "spans": [ + { + "bbox": [ + 62, + 321, + 294, + 374 + ], + "type": "text", + "content": "[36] R. Devon Hjelm, Alex Fedorov, Samuel Lavoie-Marchildon, Karan Grewal, Phil Bachman, Adam Trischler, and Yoshua Bengio. Learning deep representations by mutual information estimation and maximization. In ICLR, 2018. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 376, + 294, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 376, + 294, + 421 + ], + "spans": [ + { + "bbox": [ + 62, + 376, + 294, + 421 + ], + "type": "text", + "content": "[37] Yicong Hong, Kai Zhang, Jiuming Gu, Sai Bi, Yang Zhou, Difan Liu, Feng Liu, Kalyan Sunkavalli, Trung Bui, and Hao Tan. Lrm: Large reconstruction model for single image to 3d. arXiv preprint arXiv:2311.04400, 2023. 16" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 422, + 294, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 422, + 294, + 464 + ], + "spans": [ + { + "bbox": [ + 62, + 422, + 294, + 464 + ], + "type": "text", + "content": "[38] Ji Hou, Benjamin Graham, Matthias Nießner, and Saining Xie. Exploring data-efficient 3d scene understanding with contrastive scene contexts. In CVPR, 2021. 2, 3, 7, 8, 16, 17" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 467, + 294, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 467, + 294, + 510 + ], + "spans": [ + { + "bbox": [ + 62, + 467, + 294, + 510 + ], + "type": "text", + "content": "[39] Qingyong Hu, Bo Yang, Linhai Xie, Stefano Rosa, Yulan Guo, Zhihua Wang, Niki Trigoni, and Andrew Markham. Randla-net: Efficient semantic segmentation of large-scale point clouds. In CVPR, 2020. 16" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 512, + 294, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 512, + 294, + 554 + ], + "spans": [ + { + "bbox": [ + 62, + 512, + 294, + 554 + ], + "type": "text", + "content": "[40] Zeyu Hu, Mingmin Zhen, Xuyang Bai, Hongbo Fu, and Chiew-lan Tai. Jsenet: Joint semantic segmentation and edge detection network for 3d point clouds. In ECCV, 2020. 16" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 556, + 294, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 556, + 294, + 589 + ], + "spans": [ + { + "bbox": [ + 62, + 556, + 294, + 589 + ], + "type": "text", + "content": "[41] Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In ICML, 2015. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 62, + 590, + 294, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 590, + 294, + 635 + ], + "spans": [ + { + "bbox": [ + 62, + 590, + 294, + 635 + ], + "type": "text", + "content": "[42] Ayush Jain, Pushkal Katara, Nikolaos Gkanatsios, Adam W. Harley, Gabriel Sarch, Kriti Aggarwal, Vishrav Chaudhary, and Katerina Fragkiadaki. Odin: A single model for 2d and 3d perception. In CVPR, 2024. 16" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 62, + 635, + 294, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 635, + 294, + 668 + ], + "spans": [ + { + "bbox": [ + 62, + 635, + 294, + 668 + ], + "type": "text", + "content": "[43] Stephen James, Zicong Ma, David Rovick Arrojo, and Andrew J. Davison. Rlbench: The robot learning benchmark & learning environment. RAL, 2020. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 62, + 670, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 670, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 62, + 670, + 294, + 712 + ], + "type": "text", + "content": "[44] Li Jiang, Hengshuang Zhao, Shu Liu, Xiaoyong Shen, Chi-Wing Fu, and Jiaya Jia. Hierarchical point-edge interaction network for point cloud semantic segmentation. In ICCV, 2019. 17" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 321, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 321, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 321, + 73, + 553, + 106 + ], + "type": "text", + "content": "[45] Li Jiang, Hengshuang Zhao, Shaoshuai Shi, Shu Liu, Chi-Wing Fu, and Jiaya Jia. Pointgroup: Dual-set point grouping for 3d instance segmentation. CVPR, 2020. 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 321, + 107, + 553, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 107, + 553, + 139 + ], + "spans": [ + { + "bbox": [ + 321, + 107, + 553, + 139 + ], + "type": "text", + "content": "[46] Maxim Kolodiazhnyi, Anna Vorontsova, Anton Konushin, and Danila Rukhovich. Oneformer3d: One transformer for unified point cloud segmentation. In CVPR, 2024. 16" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 321, + 140, + 553, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 140, + 553, + 182 + ], + "spans": [ + { + "bbox": [ + 321, + 140, + 553, + 182 + ], + "type": "text", + "content": "[47] Xin Lai, Jianhui Liu, Li Jiang, Liwei Wang, Hengshuang Zhao, Shu Liu, Xiaojuan Qi, and Jiaya Jia. Stratified transformer for 3d point cloud segmentation. In CVPR, 2022. 16, 17" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 321, + 184, + 553, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 184, + 553, + 215 + ], + "spans": [ + { + "bbox": [ + 321, + 184, + 553, + 215 + ], + "type": "text", + "content": "[48] Loic Landrieu and Martin Simonovsky. Large-scale point cloud semantic segmentation with superpoint graphs. In CVPR, 2018. 17" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 321, + 217, + 553, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 217, + 553, + 249 + ], + "spans": [ + { + "bbox": [ + 321, + 217, + 553, + 249 + ], + "type": "text", + "content": "[49] Huan Lei, Naveed Akhtar, and Ajmal Mian. Seggen: Efficient 3d point cloud segmentation with fuzzy spherical kernel. In CVPR, 2020. 16, 17" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 321, + 251, + 553, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 251, + 553, + 282 + ], + "spans": [ + { + "bbox": [ + 321, + 251, + 553, + 282 + ], + "type": "text", + "content": "[50] Yangyan Li, Rui Bu, Mingchao Sun, Wei Wu, Xinhan Di, and Baoquan Chen. Pointconn: Convolution on x-transformed points. NeurIPS, 2018. 16, 17" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 321, + 284, + 553, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 284, + 553, + 336 + ], + "spans": [ + { + "bbox": [ + 321, + 284, + 553, + 336 + ], + "type": "text", + "content": "[51] Dingkang Liang, Xin Zhou, Wei Xu, Xingkui Zhu, Zhikang Zou, Xiaqing Ye, Xiao Tan, and Xiang Bai. Pointmamba: A simple state space model for point cloud analysis. In Advances in Neural Information Processing Systems, 2024. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 321, + 338, + 553, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 338, + 553, + 381 + ], + "spans": [ + { + "bbox": [ + 321, + 338, + 553, + 381 + ], + "type": "text", + "content": "[52] Haojia Lin, Xiawu Zheng, Lijiang Li, Fei Chao, Shanshan Wang, Yan Wang, Yonghong Tian, and Rongrong Ji. Meta architecture for point cloud analysis. In CVPR, pages 17682-17691, 2023. 16, 17" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 321, + 383, + 553, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 383, + 553, + 426 + ], + "spans": [ + { + "bbox": [ + 321, + 383, + 553, + 426 + ], + "type": "text", + "content": "[53] Minghua Liu, Ruoxi Shi, Kaiming Kuang, Yinhao Zhu, Xu-anlin Li, Shizhong Han, Hong Cai, Fatih Porikli, and Hao Su. Openshape: Scaling up 3d shape representation towards open-world understanding. In NeurIPS, 2023. 8, 13" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 321, + 427, + 553, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 427, + 553, + 448 + ], + "spans": [ + { + "bbox": [ + 321, + 427, + 553, + 448 + ], + "type": "text", + "content": "[54] I Loshchilov. Decoupled weight decay regularization. In ICLR, 2019. 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 321, + 449, + 553, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 449, + 553, + 469 + ], + "spans": [ + { + "bbox": [ + 321, + 449, + 553, + 469 + ], + "type": "text", + "content": "[55] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. In *ICLR*, 2017. 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 321, + 472, + 553, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 472, + 553, + 503 + ], + "spans": [ + { + "bbox": [ + 321, + 472, + 553, + 503 + ], + "type": "text", + "content": "[56] Xu Ma, Can Qin, Haoxuan You, Haoxi Ran, and Yun Fu. Rethinking network design and local geometry in point cloud: A simple residual mlp framework. *ICLR*, 2022. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 321, + 505, + 553, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 505, + 553, + 536 + ], + "spans": [ + { + "bbox": [ + 321, + 505, + 553, + 536 + ], + "type": "text", + "content": "[57] Gaku Narita, Takashi Seno, Tomoya Ishikawa, and Yohsuke Kaji. Panopticfusion: Online volumetric semantic mapping at the level of stuff and things. In IROS, 2019. 16" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 321, + 538, + 553, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 538, + 553, + 613 + ], + "spans": [ + { + "bbox": [ + 321, + 538, + 553, + 613 + ], + "type": "text", + "content": "[58] Richard A Newcombe, Shahram Izadi, Otmar Hilliges, David Molyneaux, David Kim, Andrew J Davison, Pushmeet Kohi, Jamie Shotton, Steve Hodges, and Andrew Fitzgibbon. Kinectfusion: Real-time dense surface mapping and tracking. In 2011 10th IEEE international symposium on mixed and augmented reality, pages 127-136. IEEE, 2011. 16" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 321, + 614, + 553, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 614, + 553, + 646 + ], + "spans": [ + { + "bbox": [ + 321, + 614, + 553, + 646 + ], + "type": "text", + "content": "[59] Mehdi Noroozi and Paolo Favaro. Unsupervised learning of visual representations by solving jigsaw puzzles. In ECCV, 2016. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 321, + 647, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 647, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 321, + 647, + 553, + 712 + ], + "type": "text", + "content": "[60] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Russell Howes, Po-Yao Huang, Hu Xu, Vasu Sharma, Shang-Wen Li, Wojciech Galuba, Mike Rabbat, Mido Assan, Nicolas Ballas, Gabriel Synnaeve, Ishan Misra, Herve" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "22202" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 61, + 72, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 80, + 72, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 72, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 80, + 72, + 294, + 106 + ], + "type": "text", + "content": "Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. Dinov2: Learning robust visual features without supervision. TMLR, 2024. 2, 3, 4, 6, 7, 13" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 106, + 295, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 106, + 295, + 139 + ], + "spans": [ + { + "bbox": [ + 61, + 106, + 295, + 139 + ], + "type": "text", + "content": "[61] Yatian Pang, Wenxiao Wang, Francis EH Tay, Wei Liu, Yonghong Tian, and Li Yuan. Masked autoencoders for point cloud self-supervised learning. In ECCV, 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 140, + 294, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 140, + 294, + 161 + ], + "spans": [ + { + "bbox": [ + 61, + 140, + 294, + 161 + ], + "type": "text", + "content": "[62] Chunghyun Park, Yoonwoo Jeong, Minsu Cho, and Jaesik Park. Fast point transformer. In CVPR, 2022. 16" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 162, + 295, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 162, + 295, + 215 + ], + "spans": [ + { + "bbox": [ + 62, + 162, + 295, + 215 + ], + "type": "text", + "content": "[63] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. NeurIPS, 2019. 16" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 217, + 294, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 217, + 294, + 259 + ], + "spans": [ + { + "bbox": [ + 62, + 217, + 294, + 259 + ], + "type": "text", + "content": "[64] Bohao Peng, Xiaoyang Wu, Li Jiang, Yukang Chen, Hengshuang Zhao, Zhuotao Tian, and Jiaya Jia. Oa-cnns: Omni-adaptive sparse cnns for 3d semantic segmentation. In CVPR, 2024. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 261, + 294, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 261, + 294, + 293 + ], + "spans": [ + { + "bbox": [ + 62, + 261, + 294, + 293 + ], + "type": "text", + "content": "[65] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In CVPR, 2017. 7, 17" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 294, + 294, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 294, + 294, + 326 + ], + "spans": [ + { + "bbox": [ + 62, + 294, + 294, + 326 + ], + "type": "text", + "content": "[66] Charles R Qi, Li Yi, Hao Su, and Leonidas J Guibas. Point-net++: Deep hierarchical feature learning on point sets in a metric space. In NeurIPS, 2017. 3, 16, 17" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 327, + 294, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 327, + 294, + 371 + ], + "spans": [ + { + "bbox": [ + 62, + 327, + 294, + 371 + ], + "type": "text", + "content": "[67] Guocheng Qian, Yuchen Li, Houwen Peng, Jinjie Mai, Hasan Hammoud, Mohamed Elhoseiny, and Bernard Ghanem. Pointnext: Revisiting pointnet++ with improved training and scaling strategies. In NeurIPS, 2022. 3, 16, 17" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 372, + 294, + 447 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 372, + 294, + 447 + ], + "spans": [ + { + "bbox": [ + 62, + 372, + 294, + 447 + ], + "type": "text", + "content": "[68] Santhosh Kumar Ramakrishnan, Aaron Gokaslan, Erik Wijmans, Oleksandr Maksymets, Alexander Clegg, John M Turner, Eric Undersander, Wojciech Galuba, Andrew Westbury, Angel X Chang, Manolis Savva, Yili Zhao, and Dhruv Batra. Habitat-matterport 3d dataset (HM3d): 1000 large-scale 3d environments for embodied AI. In NeurIPS, 2021. 2, 6, 14, 15" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 449, + 294, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 449, + 294, + 480 + ], + "spans": [ + { + "bbox": [ + 62, + 449, + 294, + 480 + ], + "type": "text", + "content": "[69] Damien Robert, Hugo Raguet, and Loic Landrieu. Efficient 3d semantic segmentation with superpoint transformer. In ICCV, 2023. 17" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 482, + 294, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 482, + 294, + 514 + ], + "spans": [ + { + "bbox": [ + 62, + 482, + 294, + 514 + ], + "type": "text", + "content": "[70] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In MICCAI, 2015. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 514, + 294, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 514, + 294, + 547 + ], + "spans": [ + { + "bbox": [ + 62, + 514, + 294, + 547 + ], + "type": "text", + "content": "[71] David Rozenberszki, Or Litany, and Angela Dai. Language-grounded indoor 3d semantic segmentation in the wild. In ECCV, 2022. 7, 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 62, + 548, + 294, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 548, + 294, + 580 + ], + "spans": [ + { + "bbox": [ + 62, + 548, + 294, + 580 + ], + "type": "text", + "content": "[72] Alexandre Sablayrolles, Matthijs Douze, Cordelia Schmid, and Hervé Jégou. Spreading vectors for similarity search. In ICLR, 2019. 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 62, + 582, + 294, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 582, + 294, + 624 + ], + "spans": [ + { + "bbox": [ + 62, + 582, + 294, + 624 + ], + "type": "text", + "content": "[73] Julian Straub, Daniel DeTone, Tianwei Shen, Nan Yang, Chris Sweeney, and Richard Newcombe. Efm3d: A benchmark for measuring progress towards 3d egocentric foundation models. arXiv:2406.10224, 2024. 2, 14, 15, 16" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 62, + 625, + 294, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 625, + 294, + 679 + ], + "spans": [ + { + "bbox": [ + 62, + 625, + 294, + 679 + ], + "type": "text", + "content": "[74] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, et al. Scalability in perception for autonomous driving: Waymo open dataset. In CVPR, 2020. 2, 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 62, + 681, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 681, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 62, + 681, + 294, + 712 + ], + "type": "text", + "content": "[75] Maxim Tatarchenko, Jaesik Park, Vladlen Koltun, and Qian-Yi Zhou. Tangent convolutions for dense prediction in 3d. In CVPR, 2018. 17" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 320, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 34, + "blocks": [ + { + "bbox": [ + 320, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 320, + 73, + 553, + 106 + ], + "type": "text", + "content": "[76] Lyne Tchapmi, Christopher Choy, Iro Armeni, JunYoung Gwak, and Silvio Savarese. Segcloud: Semantic segmentation of 3d point clouds. In 3DV, 2017. 17" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 320, + 107, + 553, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 107, + 553, + 149 + ], + "spans": [ + { + "bbox": [ + 320, + 107, + 553, + 149 + ], + "type": "text", + "content": "[77] Hugues Thomas, Charles R Qi, Jean-Emmanuel Deschaud, Beatrix Marcotegui, François Goulette, and Leonidas J Guibas. Kpconv: Flexible and deformable convolution for point clouds. In ICCV, 2019. 3, 16, 17" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 320, + 151, + 553, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 151, + 553, + 182 + ], + "spans": [ + { + "bbox": [ + 320, + 151, + 553, + 182 + ], + "type": "text", + "content": "[78] Hugues Thomas, Yao-Hung Hubert Tsai, Timothy D Barfoot, and Jian Zhang. Kpconvx: Modernizing kernel point convolution with kernel attention. In CVPR, 2024. 16" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 320, + 183, + 553, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 183, + 553, + 226 + ], + "spans": [ + { + "bbox": [ + 320, + 183, + 553, + 226 + ], + "type": "text", + "content": "[79] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 16" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 320, + 228, + 553, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 228, + 553, + 271 + ], + "spans": [ + { + "bbox": [ + 320, + 228, + 553, + 271 + ], + "type": "text", + "content": "[80] Chengyao Wang, Li Jiang, Xiaoyang Wu, Zhuotao Tian, Bohao Peng, Hengshuang Zhao, and Jiaya Jia. Groupcontrast: Semantic-aware self-supervised representation learning for 3d understanding. In CVPR, 2024. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 320, + 272, + 553, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 272, + 553, + 304 + ], + "spans": [ + { + "bbox": [ + 320, + 272, + 553, + 304 + ], + "type": "text", + "content": "[81] Lei Wang, Yuchun Huang, Yaolin Hou, Shenman Zhang, and Jie Shan. Graph attention convolution for point cloud semantic segmentation. In CVPR, 2019. 17" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 320, + 305, + 553, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 305, + 553, + 326 + ], + "spans": [ + { + "bbox": [ + 320, + 305, + 553, + 326 + ], + "type": "text", + "content": "[82] Peng-Shuai Wang. Octformer: Octree-based transformers for 3D point clouds. SIGGRAPH, 2023. 16" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 320, + 327, + 553, + 359 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 327, + 553, + 359 + ], + "spans": [ + { + "bbox": [ + 320, + 327, + 553, + 359 + ], + "type": "text", + "content": "[83] Shenlong Wang, Simon Suo, Wei-Chiu Ma, Andrei Pokrovsky, and Raquel Urtasun. Deep parametric continuous convolutional neural networks. In CVPR, 2018. 17" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 320, + 360, + 553, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 360, + 553, + 392 + ], + "spans": [ + { + "bbox": [ + 320, + 360, + 553, + 392 + ], + "type": "text", + "content": "[84] Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. Dust3r: Geometric 3d vision made easy. In CVPR, 2024. 2, 13" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 320, + 393, + 553, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 393, + 553, + 425 + ], + "spans": [ + { + "bbox": [ + 320, + 393, + 553, + 425 + ], + "type": "text", + "content": "[85] Wenxuan Wu, Zhongang Qi, and Li Fuxin. Pointconv: Deep convolutional networks on 3d point clouds. In CVPR, 2019. 16" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 320, + 426, + 553, + 457 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 426, + 553, + 457 + ], + "spans": [ + { + "bbox": [ + 320, + 426, + 553, + 457 + ], + "type": "text", + "content": "[86] Wenxuan Wu, Li Fuxin, and Qi Shan. Pointconvformer: Revenge of the point-based convolution. In CVPR, 2023. 16" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 320, + 460, + 553, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 460, + 553, + 502 + ], + "spans": [ + { + "bbox": [ + 320, + 460, + 553, + 502 + ], + "type": "text", + "content": "[87] Xiaoyang Wu, Yixing Lao, Li Jiang, Xihui Liu, and Hengshuang Zhao. Point transformer v2: Grouped vector attention and partition-based pooling. In NeurIPS, 2022. 7, 16, 17" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 320, + 503, + 553, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 503, + 553, + 547 + ], + "spans": [ + { + "bbox": [ + 320, + 503, + 553, + 547 + ], + "type": "text", + "content": "[88] Xiaoyang Wu, Xin Wen, Xihui Liu, and Hengshuang Zhao. Masked scene contrast: A scalable framework for unsupervised 3d representation learning. In CVPR, 2023. 2, 3, 4, 5, 6, 7, 8, 14, 16, 17" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 320, + 548, + 553, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 548, + 553, + 592 + ], + "spans": [ + { + "bbox": [ + 320, + 548, + 553, + 592 + ], + "type": "text", + "content": "[89] Xiaoyang Wu, Li Jiang, Peng-Shuai Wang, Zhijian Liu, Xihui Liu, Yu Qiao, Wanli Ouyang, Tong He, and Hengshuang Zhao. Point transformer v3: Simpler, faster, stronger. In CVPR, 2024. 3, 5, 7, 8, 16, 17" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 320, + 593, + 553, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 593, + 553, + 635 + ], + "spans": [ + { + "bbox": [ + 320, + 593, + 553, + 635 + ], + "type": "text", + "content": "[90] Xiaoyang Wu, Zhuotao Tian, Xin Wen, Bohao Peng, Xihui Liu, Kaicheng Yu, and Hengshuang Zhao. Towards large-scale 3d representation learning with multi-dataset point prompt training. In CVPR, 2024. 3, 6, 7, 8, 16, 17" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 320, + 636, + 553, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 636, + 553, + 668 + ], + "spans": [ + { + "bbox": [ + 320, + 636, + 553, + 668 + ], + "type": "text", + "content": "[91] Zhirong Wu, Yuanjun Xiong, Stella X. Yu, and Dahua Lin. Unsupervised feature learning via non-parametric instance discrimination. In CVPR, 2018. 2" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 320, + 669, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 669, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 320, + 669, + 553, + 712 + ], + "type": "text", + "content": "[92] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. In NeurIPS, 2021. 3" + } + ] + } + ], + "index": 33 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22203" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 57, + 73, + 294, + 712 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 61, + 73, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 73, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 61, + 73, + 294, + 116 + ], + "type": "text", + "content": "[93] Saining Xie, Jiatao Gu, Demi Guo, Charles R Qi, Leonidas Guibas, and Or Litany. Pointcontrast: Unsupervised pretraining for 3d point cloud understanding. In ECCV, 2020. 2, 3, 4, 6, 7, 8, 16, 17" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 118, + 294, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 118, + 294, + 160 + ], + "spans": [ + { + "bbox": [ + 61, + 118, + 294, + 160 + ], + "type": "text", + "content": "[94] Zhenda Xie, Zheng Zhang, Yue Cao, Yutong Lin, Jianmin Bao, Zhuliang Yao, Qi Dai, and Han Hu. Simmim: A simple framework for masked image modeling. In CVPR, 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 162, + 294, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 162, + 294, + 195 + ], + "spans": [ + { + "bbox": [ + 62, + 162, + 294, + 195 + ], + "type": "text", + "content": "[95] Mutian Xu, Runyu Ding, Hengshuang Zhao, and Xiaojuan Qi. Paconv: Position adaptive convolution with dynamic kernel assembling on point clouds. In CVPR, 2021. 17" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 195, + 294, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 195, + 294, + 237 + ], + "spans": [ + { + "bbox": [ + 62, + 195, + 294, + 237 + ], + "type": "text", + "content": "[96] Xu Yan, Chaoda Zheng, Zhen Li, Sheng Wang, and Shuguang Cui. Pointasnl: Robust point clouds processing using nonlocal neural networks with adaptive sampling. In CVPR, 2020. 16" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 239, + 294, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 239, + 294, + 282 + ], + "spans": [ + { + "bbox": [ + 62, + 239, + 294, + 282 + ], + "type": "text", + "content": "[97] Jiancheng Yang, Qiang Zhang, Bingbing Ni, Linguuo Li, Jinxian Liu, Mengdie Zhou, and Qi Tian. Modeling point clouds with self-attention and gumbel subset sampling. In CVPR, 2019. 17" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 284, + 294, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 284, + 294, + 326 + ], + "spans": [ + { + "bbox": [ + 62, + 284, + 294, + 326 + ], + "type": "text", + "content": "[98] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Ji-ashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. In CVPR, 2024. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 327, + 294, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 327, + 294, + 360 + ], + "spans": [ + { + "bbox": [ + 62, + 327, + 294, + 360 + ], + "type": "text", + "content": "[99] Lihe Yang, Bingyi Kang, Zilong Huang, Zhen Zhao, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything v2. In NeurIPS, 2024. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 57, + 361, + 294, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 361, + 294, + 403 + ], + "spans": [ + { + "bbox": [ + 57, + 361, + 294, + 403 + ], + "type": "text", + "content": "[100] Yu-Qi Yang, Yu-Xiao Guo, Jian-Yu Xiong, Yang Liu, Hao Pan, Peng-Shuai Wang, Xin Tong, and Baining Guo. Swin3d: A pretrained transformer backbone for 3d indoor scene understanding. arXiv:2304.06906, 2023. 16, 17" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 57, + 404, + 294, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 404, + 294, + 437 + ], + "spans": [ + { + "bbox": [ + 57, + 404, + 294, + 437 + ], + "type": "text", + "content": "[101] Chandan Yeshwanth, Yueh-Cheng Liu, Matthias Nießner, and Angela Dai. Scannet++: A high-fidelity dataset of 3d indoor scenes. In ICCV, 2023. 2, 6, 7, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 57, + 437, + 294, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 437, + 294, + 480 + ], + "spans": [ + { + "bbox": [ + 57, + 437, + 294, + 480 + ], + "type": "text", + "content": "[102] Xumin Yu, Lulu Tang, Yongming Rao, Tiejun Huang, Jie Zhou, and Jiwen Lu. Point-bert: Pre-training 3d point cloud transformers with masked point modeling. In CVPR, 2022. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 482, + 294, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 482, + 294, + 514 + ], + "spans": [ + { + "bbox": [ + 57, + 482, + 294, + 514 + ], + "type": "text", + "content": "[103] Feihu Zhang, Jin Fang, Benjamin Wah, and Philip Torr. Deep fusionnet for point cloud semantic segmentation. In ECCV, 2020. 16" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 515, + 294, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 515, + 294, + 558 + ], + "spans": [ + { + "bbox": [ + 57, + 515, + 294, + 558 + ], + "type": "text", + "content": "[104] Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel M. Ni, and Heung-Yeung Shum. Dino: Detr with improved denoising anchor boxes for end-to-end object detection. arXiv:2203.03605, 2022. 2, 3, 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 559, + 294, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 559, + 294, + 602 + ], + "spans": [ + { + "bbox": [ + 57, + 559, + 294, + 602 + ], + "type": "text", + "content": "[105] Michael Zhang, Aditi Raghunathan Wang, Shiori Sagawa, Tatsunori B Hashimoto, and Percy Liang. Fine-tuning can distort pretrained features and underperform out-of-distribution. In ICLR, 2022. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 603, + 294, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 603, + 294, + 647 + ], + "spans": [ + { + "bbox": [ + 57, + 603, + 294, + 647 + ], + "type": "text", + "content": "[106] Renrui Zhang, Ziyu Guo, Peng Gao, Rongyao Fang, Bin Zhao, Dong Wang, Yu Qiao, and Hongsheng Li. Pointm2ae: Multi-scale masked autoencoders for hierarchical point cloud pre-training. In NeurIPS, 2022. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 647, + 294, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 647, + 294, + 680 + ], + "spans": [ + { + "bbox": [ + 57, + 647, + 294, + 680 + ], + "type": "text", + "content": "[107] Hengshuang Zhao, Li Jiang, Chi-Wing Fu, and Jiaya Jia. Pointweb: Enhancing local neighborhood features for point cloud processing. In CVPR, 2019. 17" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 681, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 681, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 681, + 294, + 712 + ], + "type": "text", + "content": "[108] Hengshuang Zhao, Li Jiang, Jiaya Jia, Philip Torr, and Vladlen Koltun. Point transformer. In ICCV, 2021. 16, 17" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 208 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "text", + "content": "[109] Jia Zheng, Junfei Zhang, Jing Li, Rui Tang, Shenghua Gao, and Zihan Zhou. Structured3d: A large photo-realistic dataset for structured 3d modeling. In ECCV, 2020. 2, 5, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 118, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 553, + 150 + ], + "type": "text", + "content": "[110] Jinghao Zhou, Chen Wei, Huiyu Wang, Wei Shen, Cihang Xie, Alan Yuille, and Tao Kong. ibot: Image bert pretraining with online tokenizer. In ICLR, 2022. 2, 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 152, + 553, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 553, + 184 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 553, + 184 + ], + "type": "text", + "content": "[111] Jinghao Zhou, Chen Wei, Huiyu Wang, Wei Shen, Cihang Xie, Alan Yuille, and Tao Kong. Image BERT pre-training with online tokenizer. In ICLR, 2022. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 186, + 553, + 208 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 186, + 553, + 208 + ], + "spans": [ + { + "bbox": [ + 316, + 186, + 553, + 208 + ], + "type": "text", + "content": "[112] Xingyi Zhou, Vladlen Koltun, and Philipp Krahenbuhl. Simple multi-dataset detection. In CVPR, 2022. 3" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "22204" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Sonic_ Shifting Focus to Global Audio Perception in Portrait Animation/108dd369-6bdf-4c82-9e76-1c8b1e3be7c1_content_list.json b/2025/Sonic_ Shifting Focus to Global Audio Perception in Portrait Animation/108dd369-6bdf-4c82-9e76-1c8b1e3be7c1_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..c6f68f8ff64be9f02f6903f8588e0b7b29dbd434 --- /dev/null +++ b/2025/Sonic_ Shifting Focus to Global Audio Perception in Portrait Animation/108dd369-6bdf-4c82-9e76-1c8b1e3be7c1_content_list.json @@ -0,0 +1,1295 @@ +[ + { + "type": "text", + "text": "Sonic: Shifting Focus to Global Audio Perception in Portrait Animation", + "text_level": 1, + "bbox": [ + 137, + 130, + 859, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xiaozhong Ji $^{1*}$ Xiaobin Hu $^{1*}$ Zhihong Xu $^{2}$ Junwei Zhu $^{1\\dagger}$ Chuming Lin $^{1}$ Qingdong He $^{1}$ Jiangning Zhang $^{1}$ Donghao Luo $^{1}$ Yi Chen $^{1}$ Qin Lin $^{1}$ Qinglin Lu $^{1}$ Chengjie Wang $^{1}$ $^{1}$ Tencent $^{2}$ Zhejiang University", + "bbox": [ + 119, + 179, + 875, + 236 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4cdac659236a27f42818a2d5d96f7d65ea4e189ed7cb8f635c69cd635001d723.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 138, + 257, + 439, + 351 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/258097e14bae0c985142d9d7ac64a532dbf2f780721f6225b9a92e1b8ecba330.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 460, + 257, + 854, + 352 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1457ec753d8173dbf7a05a442559e8dc0dbe82f94280f86c25afafe79bdae14f.jpg", + "image_caption": [ + "Figure 1. Sonic excels in producing vivid portrait animation videos given a reference image and an audio clip. Beyond its fundamental lip-syncing capabilities, Sonic demonstrates proficiency in creating a diverse spectrum of facial expressions and adaptable head movements. Notably, when dealing with extended videos, Sonic can yield stable and seamless outcomes in a parallel fashion, all while maintaining an unified paradigm focusing on global audio perception." + ], + "image_footnote": [], + "bbox": [ + 137, + 354, + 429, + 477 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b5e2e39a45c721c1e21dd6850c86200acd2cfbcb49d4f9634c16dc26d769eea5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 444, + 354, + 849, + 477 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 547, + 326, + 561 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The study of talking face generation mainly explores the intricacies of synchronizing facial movements and crafting visually appealing, temporally-coherent animations. However, due to the limited exploration of global audio perception, current approaches predominantly employ auxiliary visual and spatial knowledge to stabilize the movements, which often results in the deterioration of the naturalness and temporal inconsistencies. Considering the essence of audio-driven animation, the audio signal serves as the ideal and unique priors to adjust facial expressions and lip movements, without resorting to interference of any visual signals. Based on this motivation, we propose a novel paradigm, dubbed as Sonic, to shift focus on the exploration of global audio perception. To effectively leverage global audio knowledge, we disentangle it into intra- and inter-clip audio perception and collaborate with both aspects to enhance overall perception. For the intra-clip audio perception, 1). Context-enhanced audio learning, in", + "bbox": [ + 88, + 580, + 485, + 853 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "which long-range intra-clip temporal audio knowledge is extracted to provide facial expression and lip motion priors implicitly expressed as the tone and speed of speech. 2). Motion-decoupled controller, in which the motion of the head and expression movement are disentangled and independently controlled by intra-audio clips. Most importantly, for inter-clip audio perception, as a bridge to connect the intra-clips to achieve the global perception, Time-aware position shift fusion, in which the global inter-clip audio information is considered and fused for long-audio inference via through consecutively time-aware shifted windows. Extensive experiments demonstrate that the novel audio-driven paradigm outperform existing SOTA methodologies in terms of video quality, temporally consistency, lip synchronization precision, and motion diversity.", + "bbox": [ + 511, + 549, + 908, + 777 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 811, + 645, + 828 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Talking Face Animation is a technique aimed at creating a lifelike representation of a speaking person by animating a still image in sync with speech audio. This technology holds great potential in enhancing the realism of virtual re", + "bbox": [ + 511, + 839, + 906, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Equal Contribution, † Project Leader \nCorresponding Author \nProject page: https://jixiaozhong.github.io/Sonic/.", + "bbox": [ + 89, + 862, + 452, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "193", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ality [5], improving the efficiency of film and gaming making [25], facilitating accessibility for individuals with communication impairments [24], and offering therapeutic support and social interaction in healthcare [31]. Prominent progress in this domain is demonstrated by notable works based on Stable Diffusion [32], which project the parametric or implicit representations of lip movements [29, 50], facial expressions [16, 28, 33], and head movement [35] in the latent space [11, 27, 37, 45] to generate high-quality video with a high level of realism and liveliness.", + "bbox": [ + 89, + 90, + 480, + 242 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, precise audio control and temporal coherence have not yet been effectively achieved. Current approaches [8, 23, 37, 40, 42, 44, 49, 50] predominantly addressed these two issues in a completely segregated manner, overlooking the holistic coordination between audio and vision. For audio control, audio features were segmented according to timestamps and matched to each visual frame. By applying cross-attention within a spatial latent space, each frame was restricted to the nearby audio information, which hindered the transformation into an optimal motion representation. To achieve temporal coherence, self-attention along the temporal dimension was applied to smooth visual features within a clip, while overlapping or motion frames were used to enhance coherence between different clips. However, these stability strategies have limited temporal receptive fields, which can reduce the richness of motion. Additionally, they do not take audio information into account, potentially disrupting the synchronization between audio and visual elements.", + "bbox": [ + 91, + 247, + 483, + 532 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address these issues, we propose a novel paradigm, dubbed as Sonic, to extent the range of receptive field to the global level guided by the audio priors. To achieve global audio perception, we decompose the talking head video generation task into two aspects: intra-clip (i.e., lip sync, expression alignment, and head movement) and inter-clip (i.e., motion rhythm and smoothness, long video stability). Sonic aims to exploit global audio perception rather than the motion frames and other visual motion (i.e., referring to the face locator and landmark prior adopted by existing methods, which hinders the ability to perceive information directly from audio). Performing as long-range and global signal, audio shows tone, rhythm and speed of speech which implicitly provides expression and head movement motion priors. However, we have to face such a tricky challenge that, as a weak correlation cross-modality signal without the guidance of any visual motion (e.g., motion frames), current audio learning paradigm leads to the temporal jitter and unsatisfactory pattern and fails to well-align the audio and motion. To this end, we first customize efficient global audio perception learning mechanism. Specifically, context-enhanced audio learning is proposed as long-range audio learning module to extract the intra-clip audio temporal knowledge from the input audio clip. Then such intra", + "bbox": [ + 91, + 539, + 483, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "clip audio temporal information is mapped as a temporal-embeddings for the latter fusion by audio-temporal cross-attention. The temporal-embeddings projection operation can effectively reduce the computational burden and capture temporal priors within current audio clip. Although context-enhanced audio learning directs the temporal movements correlated with audio, such as lip movements synchronized with speech and facial expressions related to the content, there still exists the absence of habitual head movements. Secondly, we design motion-decoupled controller to disentangle the habitual head and expression movement and support the independent control via two explicit parameters learned from the current intra-clip audio. To enrich the functionality and playability of animation, we also add an plug to allow users to define some exaggerated movements. Most importantly, for inter-clip audio perception, time-aware position shift fusion enlarges the intra-clip audio perception to the global inter-clip audio perception via time-aware shifted windows consecutively bridging the preceding clip, providing the global inter-clip connection that significantly enhance modeling temporal power, as demonstrated in Figure 1. It is worth noting that our time-aware position shift fusion requires no extra training cost compared with the motion frames in most recent works, such as EMO [37], Loopy [23] and also no additional inference time caused by overlapping frames. Our contributions are summarized as:", + "bbox": [ + 511, + 90, + 906, + 496 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a novel unified paradigm without the guidance of visual motion, dubbed as Sonic, to focus on the exploration of global audio perception.", + "- For intra-clip audio perception, context-enhanced audio learning and motion-decoupled controller excavate the intra-clip audio temporal knowledge to direct the temporal movements.", + "- For inter-clip audio perception, we propose time-aware position shift fusion to enhance the global inter-clip audio perception by extending the intra-clip audio perception through consecutively time-aware shifted windows.", + "- Extensive experiments demonstrate that the novel unified global audio perception paradigm outperforms existing SOTA methods in terms of video quality, temporal consistency, lip synchronization, and motion diversity." + ], + "bbox": [ + 513, + 500, + 903, + 726 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 513, + 744, + 661, + 758 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Diffusion-based Video Generation", + "text_level": 1, + "bbox": [ + 513, + 770, + 808, + 784 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The remarkable achievements in diffusion models in image generation have sparked widespread research interest for video generation. Early endeavors in video generation [14, 19, 20, 26, 34] have primarily concentrated on devising image-based diffusion models with temporal module to effectively capture the dynamic characteristics inherent in video sequences. To leverage the powerful abil", + "bbox": [ + 511, + 795, + 906, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "194", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ity of pretrained image diffusion models, Video LDM [4] incorporates temporal layers that learn to align images in a temporally consistent manner. AnimateDiff [15] provides a motion-specific module by multi-stage progressive training manner, and the motion module can be seamlessly integrated into existing text-to-image architectures without any further modifications. VideoComposer [41] and VideoCrafter [6] further explore the synthesis of image-to-video generation via textual and visual features combinations. Despite these notable advancements, the field still struggles to overcome the challenge of generating temporal coherent and high-quality videos. As a powerful motion representation trained on a large well-curated high-quality dataset, Stable Video Diffusion [3] shows the remarkable temporal-consistency and the high-quality visual characters for downstream tasks. There exists a large gap between the audio-driven portrait animation and general video generation due to the fact that audio-driven task aims to use the weak cross-modality audio signal to precisely control motion movements (e.g., lip and expression movement).", + "bbox": [ + 93, + 90, + 480, + 392 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Audio-driven Talking Face Generation", + "text_level": 1, + "bbox": [ + 93, + 419, + 419, + 434 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The generation of talking face videos from audio inputs has been a longstanding task. In the early stages, the emphasis was primarily on synthesizing lip movements alone, accomplished by directly mapping audio signals to lip movements while keeping other facial attributes unchanged [7, 22, 29, 36, 46]. In more recent efforts, the scope has been broadened to encompass a wider range of facial expressions and head movements, which are derived from audio inputs and a single reference image. These methods can be broadly classified into with or without intermediate representation. For instance, by employing 3D coefficients representation (e.g., Sadtalker [50], Dreamtalk [28], Ani-Portrait [42], VASA-1 [45]), the video generation process can be informed through the direct or indirect visible control signal in the aspect of head movement and expression. Nevertheless, a recurring challenge faced by these techniques is the limited capacity of the 3D mesh to capture intricate details, consequently constraining the overall dynamism and realism. In contrast, the methods that omit intermediate representation exhibit superior naturalness and consistent preservation of identity with the original image. EMO [37], Hallo [44], and Loopy [23] employ a direct audio-to-video synthesis approach to ensure a high level of realism and naturalness, eliminating the need for intermediate 3D representation. However, these approaches primarily rely on additional conditions associated with spatial motion or motion frames to maintain temporal consistency. In this regard, they overlook the crucial role of audio in audio-driven animation, consequently diminishing both expressiveness and audio-motion synchronization.", + "bbox": [ + 93, + 448, + 480, + 898 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Long Video Inference", + "text_level": 1, + "bbox": [ + 517, + 90, + 714, + 106 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Significant research efforts have been devoted to extending the duration of generated videos, aiming to expand their practical applications. Autoregressively predicting successive frames [17, 39] and hierarchical coarse-to-fine methods [47] have emerged to generate long videos. Then Lumiere [2] divides the video into overlapping temporal segments, denoises each segment independently, and finally fuse these segments in an optimization algorithm. However there is no available high-efficiency method for audio-condition global perception. As one step towards achieving stronger capabilities, our Sonic proposes time-aware position shift fusion strategy progressively denoise intra-clip audio to establish a global inter-clip connection through the use of time-aware shifted windows that bridge the preceding clip along the timesteps axis. This inter-clip fusion with global receptive field needs no additional training cost and introduce no extra inference time without overlapping frames.", + "bbox": [ + 517, + 114, + 903, + 369 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 517, + 387, + 645, + 402 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Overall framework. Given a single portrait reference image and an input audio, our one-stage framework can generate a portrait video well-synchronized with the driven audio. Our proposed Sonic framework aims to focus on the exploration of global audio perception, primarily consisting of context-enhanced audio learning, motion-decoupled controller, and time-aware position shift fusion module, as illustrated in Figure 2.", + "bbox": [ + 517, + 412, + 903, + 532 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Technical contribution. To achieve global audio perception, we decompose the talking head video generation task into two aspects: intra-clip (i.e., lip sync, expression alignment, and head movement) and inter-clip (i.e., motion rhythm and smoothness, long video stability). For intra-clip, we enhance the visual realism that matches audio signal through the context-enhanced audio learning and the motion controller. For inter-clip, we improve the stability using the time-aware position shift fusion strategy.", + "bbox": [ + 517, + 534, + 903, + 667 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Context-enhanced Audio Learning", + "text_level": 1, + "bbox": [ + 517, + 681, + 813, + 695 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For audio-talking face animation, the audio is the ideally unique and global signal to adjust the face expression and movement without interference from other visual signals. Thus, the motivation behind context-enhanced audio learning is to guide the generation of spatial and temporal frames solely based on audio signals, rather than motion frames or other visual signals, to achieve the high-realism generation. Specifically, we adopt Whisper-Tiny [30] as the pretrained audio feature extraction model, which is more lightweight than Wav2Vec [1] used by most existing methods [8, 37, 44]. For each audio frame, the features from the last layers of five stages are concatenated together for multiscale understanding. For each video frame, a duration of", + "bbox": [ + 517, + 704, + 903, + 898 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "195", + "bbox": [ + 488, + 946, + 509, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/a1775a9d25b07070e02f41b8fc2df232ce1ecec833567882c7ae742a6b6389ed.jpg", + "image_caption": [ + "Figure 2. Framework of our approach. Sonic processes each clip of the long audio in parallel, shifting to a new context at each time step to progressively fuse inter-clip latent features across global audio perception. In Sonic, we enhance intra-clip temporal audio context learning and decouple motion to improve dynamics." + ], + "image_footnote": [], + "bbox": [ + 135, + 90, + 352, + 340 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/5339a31b068a57e4d1d9409ccc5d5a80d511b2d1e6bb19c7de2f4d2a66916b02.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 367, + 90, + 500, + 338 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/f484da2b775929682cd4a9d4bf8ec8589faca09ae2f7863e73281d2d607cac78.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 509, + 95, + 859, + 340 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "0.2s audio feature is used frame-wise to give rich context. We use three linear layers to project audio feature to match the dimension of cross attention. The transformed audio embedding is denoted as $c_{a} \\in \\mathcal{R}^{b \\times f \\times d \\times c}$ , where $b$ denotes batch size, $f$ is audio length, $d$ is num of context tokens, and $c$ is hidden size of the cross-attention. To focus on talking face area, the core latent features is restricted with a mask $M$ determined by a joint bounding boxes of faces. The audio signal can act as the spatial knowledge provider being injected into the spatial cross-attention layers as follows:", + "bbox": [ + 89, + 398, + 483, + 551 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nz _ {s} ^ {\\prime} = z _ {s} + \\operatorname {C r o s s A t t n} \\left(Q \\left(z _ {s}\\right), K \\left(c _ {a}\\right), V \\left(c _ {a}\\right)\\right) \\cdot M, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 114, + 556, + 482, + 575 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $z_{s}$ is spatial latent features of self-attention layer, and $z_{s}^{\\prime}$ is the adjusted spatial features guided by audio signals in spatial-aware level.", + "bbox": [ + 89, + 583, + 482, + 628 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Furthermore, audio is a long-range and global signal that provides expression and head movement motion priors implicitly expressed through the tone and speed of speech. To excavate the temporal information from the long-range audio knowledge, we customize a temporal-audio module for precise motion control. The temporal module in existing methods [37, 50] followsAnimateDiff, which performs only self-attention across the temporal dimension without audio aids. Some other talking-face [23, 44] works extract the motion priors from the visual knowledge (e.g., motion frames), which deviates from the essence of audio-driven task that relies solely on audio to drive movement. Thus we propose a temporal audio cross-attention module, introducing audio embedding to guide temporal alignment.", + "bbox": [ + 89, + 628, + 482, + 840 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The preceding latent from temporal self-attention layers can be reshaped as $z_{t} \\in \\mathcal{R}^{(b \\times h \\times w) \\times f \\times c}$ . To reduce the computation burden, the audio feature $c_{a}$ is average-pooled along temporal dimension as the shape $\\mathcal{R}^{(b \\times 1) \\times d \\times c}$ , then", + "bbox": [ + 89, + 840, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "repeated and reshaped as $c_{ta} \\in \\mathcal{R}^{(b \\times h \\times w) \\times d \\times c}$ . The projected audio temporal-embeddings are injected into the denoising U-Net via a temporal cross-attention blocks as:", + "bbox": [ + 511, + 398, + 906, + 445 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nz _ {t} ^ {\\prime} = z _ {t} + \\operatorname {C r o s s A t t n} (Q (z _ {t}), K (c _ {t a}), V (c _ {t a})). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 555, + 450, + 906, + 470 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The spatial and temporal audio attention layers are cascaded in multiple down-stages and up-stages to enable natural appearance and smooth temporal movement.", + "bbox": [ + 511, + 474, + 906, + 521 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Motion-decoupled Controller", + "text_level": 1, + "bbox": [ + 511, + 529, + 774, + 545 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Context-enhanced Audio Learning governs the temporal movements strongly correlated with audio, such as lip movements synchronized with speech and facial expressions related to the content. On the other hand, our motion-decoupled controller performs a complementary role by directing the temporal movements weakly associated with audio, such as habitual head movements and random changes in perspective. This controller can disentangle motion into explicit head and expression movement amplitudes, thereby enhancing both functionality, playability, and interactivity. Thus, we introduce independent augmented motion-bucket parameters which effectively influence motion patterns of head movements and expression strength. During the training phase, the translation motion-bucket $m_t$ is computed as the variance of bounding boxes of video clip, while the expression motion-bucket $m_e$ is variance of relative landmarks. Both buckets are integers in the range of [0, 128]. Through position encoding and linear projection, they are added into ResNet blocks as follows:", + "bbox": [ + 511, + 551, + 906, + 839 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ne m b = W \\left[ \\operatorname {P e} \\left(m _ {t}\\right), \\operatorname {P e} \\left(m _ {e}\\right) \\right], \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 617, + 847, + 903, + 864 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $emb$ is the embedding, $\\mathrm{Pe}$ is the position encoding, and $W$ represents the linear projection weights.", + "bbox": [ + 511, + 869, + 906, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "196", + "bbox": [ + 486, + 945, + 511, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For practice application, motion-decoupled controller also supports the automatically adjustable parameters learning from the audio and reference image, and deeply explore the correlation between audio and portrait reference without requiring the parameter setting from users. To simplify multiple parameters and maintain adjustability, the predicted motion buckets are multiplying by a scale $\\beta$ (0.5 for mild dynamic, 1.0 for moderate dynamic, and 2.0 for intense dynamic). Specifically, we utilize a audio-to-bucket manner to adaptively predict the motion-buckets relying on the audio and reference image. In inference phase, the motion buckets are predicted and rescaled as:", + "bbox": [ + 89, + 90, + 483, + 272 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nm _ {t} ^ {\\prime}, m _ {e} ^ {\\prime} = \\beta \\times \\mathrm {E} _ {\\mathrm {b}} \\left(c _ {a}, R _ {i m g}\\right). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 192, + 277, + 482, + 297 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $m_t'$ , $m_e'$ are the estimated translation and expression motion-bucket achieved by the audio $c_a$ and single reference image CLIP embedding $R_{img}$ , and $E_b$ is the audio-to-bucket projection consisting of there linear layers with the ReLU as activation function.", + "bbox": [ + 89, + 305, + 483, + 380 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Time-aware Position Shift Fusion", + "text_level": 1, + "bbox": [ + 89, + 388, + 382, + 402 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/4967c461a53e2a8190cbf3f9b40249587298b2997e31cc5da103aab701f3a0b3.jpg", + "image_caption": [ + "Figure 3. Illustration of the proposed time-aware position-shift fusion. The model processes each clip non-overlapping. In next timestep, the model starts from a new position determined by the offset, thereby integrating long-range context. Specifically, the tail latents are filled cyclically from the head." + ], + "image_footnote": [], + "bbox": [ + 98, + 426, + 464, + 537 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Recent methods [23, 37, 44] use motion frames or overlapping latent during coherent denoising, which suffer from a limited receptive field and increased the training or inference computational complexity for audio-conditioned video generation problem. In contrast, we propose a bridge, the time-aware position shift fusion, to extent the intra-clip audio perception to the global inter-clip audio-perception via consecutively time-aware shifted windows. The simple yet effective fusion strategy, without increasing extra inference or training time costs, can effectively address jitter and abrupt transition issues. Audio serves as the input condition for the talking face, where longer audio prompts the generation of longer videos. Based on this principle, we tailor time-aware position-shift fusion mechanism along the timestep axis where the latent sliding window is progressively shifted by a certain length at different timesteps while also preserving the frame correspondence relationship between audio and video signal.", + "bbox": [ + 89, + 628, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Specifically, as described in Figure 3 and Algorithm 1, our time-aware position-shift fusion consists of two nested loops. The outer loop is inverse diffusion process, while the inner loop is a sliding window process that model predicts for the full-length audio conditions. At each timestep, the model takes a clip of audio input to predict the corresponding latent with a start offset to smooth the segment from the last timestep. The video diffusion model naturally bridges the context, generating continuous movement following the audio prompts. In detail, we set a cumulative shift offset $\\alpha_{\\Sigma}$ , accumulating $\\alpha$ at each timestep to ensure that the model starts sliding window from different positions. The offset $\\alpha$ is experimentally set to a small integer, such as 3, 7; we will report analysis on this parameters in experiments part. Note that the relative positioning between the latents and the audio prompts has not been altered. The only modification occurs in their recombination during the input phase to the model. In a special case, at the end of the long sequence, the shifted position index may exceed the sequence length. To address this issue, we adopt a circular padding strategy, padding the start audio prompt and latents to the end.", + "bbox": [ + 511, + 90, + 906, + 422 + ], + "page_idx": 4 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1 Time-aware Position-shift Fusion" + ], + "code_body": "Require: Audio embedding $c_{a}^{[0,l]}$ with length $l$ , denoising steps $T$ , initial noisy latent $z_{T}^{[0,l]}$ , pretrained Sonic model $\\mathrm{ST}(\\cdot)$ for sequence length $f$ , position-shift offset $\\alpha < f < l$ . \nEnsure: Denoised latent $z_{0}^{[0,l]}$ . \n1: Initialize accumulated shift offset $\\alpha_{\\Sigma} = 0$ . \n2: for $t = T, \\dots, 1$ do \n3: // Denoising loop \n4: Initialize start point $s = \\alpha_{\\Sigma}$ , end $e = s + f$ , processed length $n = 0$ . // start from new position for each timestep. \n5: while $n < l$ do \n6: // Sequence loop \n7: $z_{t-1}^{[s,e]} = \\mathrm{ST}(z_{t}^{[s,e]}, c_{a}^{[s,e]}, t)$ \n8: $s \\gets s + f, e \\gets e + f, n \\gets n + f$ . // Move to next clip non-overlapping \n9: if $s > l$ or $e > l$ then \n10: $s \\gets s\\% l, e \\gets e\\% l$ . // Padding circularly \n11: end if \n12: end while \n13: $\\alpha_{\\Sigma} \\gets \\alpha_{\\Sigma} + \\alpha$ . // Accumulate shift offset \n14: end for \n15: return Denoised latent $z_{0}^{[0,l]}$ .", + "bbox": [ + 514, + 453, + 906, + 752 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Computation cost. We analyze the computation efficiency of our time-aware position shift fusion compared to motion frames and overlapping. Denote the FLOPs of a single forward pass of the video diffusion model as $\\Omega$ , the number of frames to generate as $n \\times f$ , and the denoising steps as $T$ . The accumulated FLOPs of Sonic is: $\\Omega(Shift) = \\Omega \\times T \\times n$ . For overlapping, the overlapped latents (overlap $o$ ) are calculated twice, $\\Omega(Overlap) = \\Omega \\times T \\times \\left(n + \\frac{o}{f}\\right)$ . For motion frames, the FLOPs of a single forward of Ref-", + "bbox": [ + 511, + 763, + 906, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "197", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/26b26caeb9ca3b21b294c9ab03ed7d14a80bd06669eeac0e8c5a16c1cffad260.jpg", + "table_caption": [ + "Table 1. Quantitative comparisons with the state-of-the-arts on the HDTF and CelebV-HQ test set. * indicates GAN-based methods and others are Diffusion-based methods. The best results are bold, and the second are underlined." + ], + "table_footnote": [], + "table_body": "
DatasetMethodFID↓FVD↓Sync-C↑Sync-D↓E-FID↓F-SIM↑Smooth↑Runtime↓
HDTFSadTalker* (2023) [50]61.672397.1141.75510.6952.4820.92870.99613.75
Aniportrait (2024) [42]36.965471.4521.09512.4613.1610.95080.992144.03
V-Express (2024) [40]47.396758.0231.25612.3942.6340.90930.996839.04
Hallo (2024) [44]30.176347.3584.0609.5511.7920.95550.994174.65
Hallo2 (2024) [12]38.673328.5404.1369.4652.2030.96060.994245.75
EchoMimic (2024) [8]33.207384.3042.51410.7431.4860.95270.99345.45
Sonic (Ours)29.104301.1734.1979.3711.7450.95950.997017.04
CelebV-HQSadTalker* (2024) [50]57.574841.9621.97810.9152.2520.94340.99593.75
Aniportrait (2024) [42]53.746590.3730.99612.0843.2960.95220.991144.03
V-Express (2024) [40]65.400889.9850.80913.2552.7130.90190.996539.04
Hallo (2024) [44]47.403488.4992.68010.2922.2730.96070.994274.65
Hallo2 (2024) [12]52.396481.3362.63810.3432.8190.95870.994245.75
EchoMimic (2024) [8]48.267596.8701.94910.7542.1360.96120.99325.45
Sonic (Ours)43.137483.1082.68910.1941.7830.96240.997217.04
", + "bbox": [ + 135, + 121, + 859, + 305 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/07353588234a301c7bb4c2ad734f4be157930330510d866ef8192f839e350115.jpg", + "image_caption": [ + "Figure 4. Qualitative comparisons with State-of-the-Art talking head generation methods. Due to image does not reflect important sync, naturalness and stability, the full video comparison will be included in supplementary materials as well as comparison with demos from other non-open source works." + ], + "image_footnote": [], + "bbox": [ + 94, + 320, + 470, + 621 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/67196a5f3134c1acae4602dea6d68ca40d7b1b00322cb8a9c55662c610341211.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 475, + 320, + 903, + 621 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "erence Net is $\\omega_{r}$ , the number of motion frames is $o$ and the Flops of motion modules computing $f$ frames is $\\omega_{m}$ , $\\Omega (MF) = \\Omega \\times T\\times n + \\omega_{r}\\times o\\times n + \\omega_{m}\\times T\\times n\\times \\frac{2of + o^{2}}{f^{2}}$ . Overall, the proposed time-aware position shift fusion significantly reduces the computation complexity.", + "bbox": [ + 89, + 689, + 482, + 768 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 787, + 223, + 805 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experiments Setup", + "text_level": 1, + "bbox": [ + 89, + 814, + 272, + 830 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details. The Sonic training is in a single stage, with initialization of the spatial module and temporal modules from stable-video-diffusion-xt-1-1[3]. To separately enable different conditions in training, we manipulate", + "bbox": [ + 89, + 839, + 483, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "the data such that $5\\%$ of it drops audio, another $5\\%$ drops image, and $5\\%$ drops both condition. In inference phase, multi-condition CFG is performed. The CFG of the reference image $r_i$ is set to 2.0, and the CFG of the audio $r_a$ is set to 7.5. In our full model, the dynamic scale $\\beta$ is set to 1.0 and the shift offset $\\alpha$ is set to 7.", + "bbox": [ + 511, + 689, + 906, + 780 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets and Evaluation Metrics. For the training set, we created a diverse dataset by aggregating data from open-source datasets, including VFHQ [43], CelebV-Text [48], VoxCeleb2 [10]. For the test set, we conducted experiments on two widely used datasets, HDTF [52] and CelebV-HQ [51]. We randomly selected 50 four-second clips from CelebV-HQ and 100 from HDTF. We utilize seven met", + "bbox": [ + 511, + 794, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "198", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/760a4446925b36607c2b04ee26f8c7dbc5ecac4b05fc2dfcdcc9805be274a672.jpg", + "image_caption": [ + "Figure 5. Qualitative results under different styles of portrait images and various types of audio inputs. The images and audios were collected from recent works and the Internet. The upper section presents results for audio input ranging from 20 seconds to 2 minutes in duration, while the lower section shows results for longer audio inputs, up to 10 minutes. Our Sonic demonstrates versatility across various portrait styles and maintains vividness over extended durations." + ], + "image_footnote": [], + "bbox": [ + 132, + 88, + 859, + 361 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "rics to evaluate. The Fréchet Inception Distance (FID) [18] and Fréchet Video Distance (FVD) [38] metrics measure the quality of the generated data. Expression-FID (E-FID) [37] uses face reconstruction method [13] to extract features and calculate the FID between generated and ground truth frames, while F-SIM measures facial similarity. SyncNet (Sync-C&Sync-D) [9] and VBench's smooth metric [21] assess lip synchronization and motion fluidity. In ablation experiments, we employ VBench's dynamic degree metric to evaluate amplitude of head movements.", + "bbox": [ + 89, + 421, + 483, + 571 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Results and Analysis", + "text_level": 1, + "bbox": [ + 89, + 587, + 285, + 603 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Quantitative Comparison. In Table 1, our Sonic are significantly superior than other recent 6 methods in terms of FID, Sync-confidence and Smoothness evaluated on the CelebV-HQ and HDTF datasets. On the HDTF dataset, our Sonic achieve the lowest FID and FVD scores, $8\\%$ lower than the second-best, indicating superior generation quality and alignment with real data. For the CelebV-HQ dataset, which features more complex characters and backgrounds, our method maintains superior performance across most metrics and competitive FVD. We notably excel in E-FID, $16.5\\%$ lower than the second-best, which demonstrates our method has superior expression diversity. EMO [37] as a representative work that tackles talking head generation task, remains closed source, and not available for fair quantitative comparison. We provide video results using input from their online demos for a qualitative comparison.", + "bbox": [ + 89, + 611, + 482, + 853 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Qualitative Comparison. Figure 4 provides a visualization comparison on open datasets. Analyzing the results, Previous approaches relied on motion frames, limiting the rich", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ness of generated expressions and the dynamism of head movements. In contrast, our proposed framework incorporates global audio perception, enhancing the integration of audio and temporal variables while providing a deeper understanding of the audio input. This enables our method to generate a broader range of expressions that align with the audio and facilitate more natural head movements.", + "bbox": [ + 511, + 421, + 906, + 526 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Visualization Results in Complex Scenarios. We further explore the generation performance of our method under different input conditions. In terms of images, we employ three styles: real human, animation, and 3D. Each character portrait is paired with various types of audio relevant to its theme, such as singing, speech, and rap. As shown in Figure 5, our method successfully generates a wide array of expressions and diverse movements across various categories of input conditions, demonstrating strong robustness.", + "bbox": [ + 511, + 527, + 908, + 666 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/c5ca55a57343619a0c23fb2ff996c3338117959f7fc90872a3079ce20028f460.jpg", + "table_caption": [ + "Table 2. User study comparison on open dataset." + ], + "table_footnote": [], + "table_body": "
Method / MetricLip syncMotion diversityID consistencyVideo Smoothness
Aniportrait1.421.623.112.09
SadTalker1.982.342.952.95
Echomimic2.772.653.482.71
Hallo23.152.373.342.94
Sonic(Ours)4.58 (45%↑)4.55 (72%↑)4.29 (23%↑)4.66 (58%↑)
", + "bbox": [ + 516, + 696, + 897, + 757 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "User Studies We conducted subjective evaluation on open dataset to assess the comparative methods along four key dimensions: Lip sync, motion diversity, identity similarity, and video smoothness. Total 40 participants gives scores from 1 to 5 for results of five comparative methods. In Table 2, our Sonic outperforms the others in all four dimensions, especially showing a significant $72\\%$ improvement in motion diversity. The video results for user study will be included in the supplementary materials.", + "bbox": [ + 511, + 763, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "199", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9f65a05c834be73952ef0238eac9444ab11f66f211836806959bc12ea4be9bc7.jpg", + "image_caption": [ + "Figure 6. Visual results for ablation on temporal audio attention and motion controller." + ], + "image_footnote": [], + "bbox": [ + 109, + 89, + 460, + 262 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8887c8f1533892c2fd1904150b799428ccb6c80fc1d53ec83dde211bc79b5735.jpg", + "image_caption": [ + "Figure 7. Demonstration on effect of time-aware position shift fusion." + ], + "image_footnote": [], + "bbox": [ + 106, + 290, + 467, + 455 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/da278e5b21cc667bb7fe03becd4931751acbb0c07a87146bda4b95be82e4554b.jpg", + "table_caption": [ + "Table 3. Quantitative results of ablations on CelebV-HQ along with our results obtained using different dynamic scales. In the Sonic model, the default setting for $\\beta$ is 1.0, i.e. Ours-Moderate." + ], + "table_footnote": [], + "table_body": "
Method / MetricSync-C↑Sync-D↓E-FID↓Smooth↑Dynamic↑
w/o temporal-audio2.61010.3102.3110.99690.62
w/o motion-controller2.19310.7313.2310.99690.22
w/o shift-fusion2.50910.3792.2020.99690.76
Ours-Mild (β = 0.5)2.65510.2422.0600.99700.78
Ours-Intense (β = 2.0)2.80310.1252.0380.99690.98
Ours-Moderate (β = 1.0)2.68910.1941.7830.99720.86
", + "bbox": [ + 93, + 534, + 478, + 618 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/640c08b665c9969263c73cdf1d4484256725346904e553721d31c73ef26a618d.jpg", + "table_caption": [ + "Table 4. Quantitative comparison on motion frames, overlap and the proposed time-aware shift fusion." + ], + "table_footnote": [], + "table_body": "
Method / MetricSync-C↑Sync-D↓Smooth↑Runtime↓
Motion frames(8)2.46510.3230.996919.70
Overlap(8)2.54410.2650.996922.49
Shift-fusion2.68910.1940.997217.04
", + "bbox": [ + 101, + 654, + 465, + 715 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Ablation Studies", + "text_level": 1, + "bbox": [ + 91, + 724, + 254, + 739 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Temporal Control Inspires Intra-clip Diversity. In Table 3, the absence of temporal audio attention results in a decrease in both lip-sync and E-FID. We observe that the motion controller plays a crucial role in CelebV-HQ given that its videos exhibit more movement that weakly correlates with audio. We also investigate the effect of different dynamic scale. Increasing $\\beta$ can further stimulate the lip-audio synchronization and the diversity of movements, while there is a slight decrease in stability and expression similarity. The instances in Figure 6 indicate", + "bbox": [ + 89, + 750, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b5ca64c6b30ab4751ecc4a740e92de9fc5491979d2dfd064d22c163a405fab67.jpg", + "image_caption": [ + "Figure 8. Visual results compared with motion frames and overlapping." + ], + "image_footnote": [], + "bbox": [ + 519, + 97, + 903, + 255 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "temporal audio attention helps learn effective control of micro-expressions, while the motion controller is critical for agility and expressiveness.", + "bbox": [ + 511, + 305, + 906, + 351 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Larger Audio Perception Bridges Inter-clip Stability. Initially, we ablate the time-aware position shift fusion (briefly termed Shift) and generate clips independently as illustrated in Figure 7. The fusion process is performed at each timestep, gradually inverting the random noise to continuous and smooth long range videos. Subsequently, we explore different long video inference strategy: motion frames, overlap, our shift. The number of motion frames and overlap frames was set to 8. As indicated in Table 4, our shift approach attains superior lip-sync and smoothness without necessitating any auxiliary computation. From the qualitative comparison demonstrated in Figure 8, our method successfully handle long term consistency, circumventing identity similarity decrease in motion frames and mixed blurry texture in overlapping results.", + "bbox": [ + 511, + 354, + 906, + 582 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 609, + 633, + 625 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In summary, our work present Sonic, an audio-driven portrait animation that focuses on the exploration of global audio perception to enhance the efficient generation of realistic lip synchronization, naturalness and temporal consistencies. The technical cornerstone mainly lies on context-enhanced audio learning and time-aware position shift fusion that work together to extend the intra-clip audio temporal to the global inter-clip audio perception. It significantly outperforms existing SOTA methods in video quality, motion diversity and naturalness.", + "bbox": [ + 511, + 638, + 906, + 789 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements We thank Dongshan Li, Zixiang Zhou, Zhentao Yu, Hongmei Wang, Joeyu Wang, Chengfei Cai, Shiyu Tang, Tianxiang Zheng, Junshu Tang, Shuai Shao, Yubin Zeng, Junxin Huang, Yue Zhang, Chao Zhan, Zhengyan Tong, Zhaokang Chen, Bin Wu, Wenjiang Zhou, Yang Zhang, Junqi Cheng, Jiaxi Gu, and Fangyuan Zou for helpful discussion.", + "bbox": [ + 511, + 795, + 908, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "200", + "bbox": [ + 485, + 945, + 511, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Alexei Baevski, Yuhao Zhou, Abdelrahman Mohamed, and Michael Auli. wav2vec 2.0: A framework for self-supervised learning of speech representations. Advances in neural information processing systems, 33:12449-12460, 2020. 3", + "[2] Omer Bar-Tal, Hila Chefer, Omer Tov, Charles Herrmann, Roni Paiss, Shiran Zada, Ariel Ephrat, Junhwa Hur, Yuanzhen Li, Tomer Michaeli, et al. Lumiere: A spacetime diffusion model for video generation. arXiv preprint arXiv:2401.12945, 2024. 3", + "[3] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 3, 6", + "[4] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22563-22575, 2023. 3", + "[5] Aras Bozkurt, Xiao Junhong, Sarah Lambert, Angelica Pazurek, Helen Crompton, Suzan Koseoglu, Robert Farrow, Melissa Bond, Chrissi Nerantzi, Sarah Honeychurch, et al. Speculative futures on chatgpt and generative artificial intelligence (ai): A collective reflection from the educational landscape. Asian Journal of Distance Education, 18(1):53-130, 2023. 2", + "[6] Haoxin Chen, Yong Zhang, Xiaodong Cun, Menghan Xia, Xintao Wang, Chao Weng, and Ying Shan. Videocrafter2: Overcoming data limitations for high-quality video diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7310-7320, 2024. 3", + "[7] Lele Chen, Zhiheng Li, Ross K Maddox, Zhiyao Duan, and Chenliang Xu. Lip movements generation at a glance. In Proceedings of the European conference on computer vision (ECCV), pages 520-535, 2018. 3", + "[8] Zhiyuan Chen, Jiajiong Cao, Zhiquan Chen, Yuming Li, and Chenguang Ma. Echomimic: Lifelike audio-driven portrait animations through editable landmark conditions. arXiv preprint arXiv:2407.08136, 2024. 2, 3, 6", + "[9] Joon Son Chung and Andrew Zisserman. Out of time: automated lip sync in the wild. In Computer Vision-ACCV 2016 Workshops: ACCV 2016 International Workshops, Taipei, Taiwan, November 20-24, 2016, Revised Selected Papers, Part II 13, pages 251-263. Springer, 2017. 7", + "[10] Joon Son Chung, Arsha Nagrani, and Andrew Zisserman. Voxceleb2: Deep speaker recognition. arXiv preprint arXiv:1806.05622, 2018. 6", + "[11] Enric Corona, Andrei Zanfir, Eduard Gabriel Bazavan, Nikos Kolotouros, Thiemo Alldieck, and Cristian Sminchisescu. Vlogger: Multimodal diffusion for embodied avatar synthesis. arXiv preprint arXiv:2403.08764, 2024. 2", + "[12] Jiahao Cui, Hui Li, Yao Yao, Hao Zhu, Hanlin Shang, Kaihui Cheng, Hang Zhou, Siyu Zhu, and Jingdong Wang. Hallo2:" + ], + "bbox": [ + 93, + 114, + 483, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Long-duration and high-resolution audio-driven portrait image animation. arXiv preprint arXiv:2410.07718, 2024. 6", + "[13] Yu Deng, Jiaolong Yang, Sicheng Xu, Dong Chen, Yunde Jia, and Xin Tong. Accurate 3d face reconstruction with weakly-supervised learning: From single image to image set. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 0–0, 2019. 7", + "[14] Patrick Esser, Johnathan Chiu, Parmida Atighehchian, Jonathan Granskog, and Anastasis Germanidis. Structure and content-guided video synthesis with diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7346-7356, 2023. 2", + "[15] Yuwei Guo, Ceyuan Yang, Anyi Rao, Zhengyang Liang, Yaohui Wang, Yu Qiao, Maneesh Agrawala, Dahua Lin, and Bo Dai. Animatediff: Animate your personalized text-to-image diffusion models without specific tuning. International Conference on Learning Representations, 2024. 3", + "[16] Tianyu He, Junliang Guo, Runyi Yu, Yuchi Wang, Jialiang Zhu, Kaikai An, Leyi Li, Xu Tan, Chunyu Wang, Han Hu, et al. Gaia: Zero-shot talking avatar generation. arXiv preprint arXiv:2311.15230, 2023. 2", + "[17] Yingqing He, Tianyu Yang, Yong Zhang, Ying Shan, and Qifeng Chen. Latent video diffusion models for high-fidelity long video generation. arXiv preprint arXiv:2211.13221, 2022.3", + "[18] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 7", + "[19] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. Advances in Neural Information Processing Systems, 35:8633-8646, 2022. 2", + "[20] Wenyi Hong, Ming Ding, Wendi Zheng, Xinghan Liu, and Jie Tang. Cogvideo: Large-scale pretraining for text-to-video generation via transformers. arXiv preprint arXiv:2205.15868, 2022. 2", + "[21] Ziqi Huang, Yinan He, Jiashuo Yu, Fan Zhang, Chenyang Si, Yuming Jiang, Yuanhan Zhang, Tianxing Wu, Qingyang Jin, Nattapol Chanpaisit, et al. Vbench: Comprehensive benchmark suite for video generative models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21807-21818, 2024. 7", + "[22] Xiaozhong Ji, Chuming Lin, Zhonggan Ding, Ying Tai, Jian Yang, Junwei Zhu, Xiaobin Hu, Jiangning Zhang, Donghao Luo, and Chengjie Wang. Realtalk: Real-time and realistic audio-driven face generation with 3d facial prior-guided identity alignment network. arXiv preprint arXiv:2406.18284, 2024. 3", + "[23] Jianwen Jiang, Chao Liang, Jiaqi Yang, Gaojie Lin, Tianyun Zhong, and Yanbo Zheng. Loopy: Taming audio-driven portrait avatar with long-term motion dependency. arXiv preprint arXiv:2409.02634, 2024. 2, 3, 4, 5", + "[24] Esperanza Johnson, Ramón Hervás, Carlos Gutiérrez López de la Franca, Tania Mondéjar, Sergio F Ochoa, and Jesús" + ], + "bbox": [ + 516, + 92, + 903, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "201", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Favela. Assessing empathy and managing emotions through interactions with an affective avatar. Health informatics journal, 24(2):182-193, 2018. 2", + "[25] Oytun Kal and Yavuz Samur. Educational virtual reality game design for film and animation. Encyclopedia of Computer Graphics and Games, pages 621-636, 2024. 2", + "[26] Levon Khachatryan, Andranik Movsisyan, Vahram Tadevosyan, Roberto Henschel, Zhangyang Wang, Shant Navasardyan, and Humphrey Shi. Text2video-zero: Text-to-image diffusion models are zero-shot video generators. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15954-15964, 2023. 2", + "[27] Tao Liu, Feilong Chen, Shuai Fan, Chenpeng Du, Qi Chen, Xie Chen, and Kai Yu. Anitalker: Animate vivid and diverse talking faces through identity-decoupled facial motion encoding. arXiv preprint arXiv:2405.03121, 2024. 2", + "[28] Yifeng Ma, Shiwei Zhang, Jiayu Wang, Xiang Wang, Yingya Zhang, and Zhidong Deng. Dreamtalk: When expressive talking head generation meets diffusion probabilistic models. arXiv preprint arXiv:2312.09767, 2023. 2, 3", + "[29] KR Prajwal, Rudrabha Mukhopadhyay, Vinay P Namboodiri, and CV Jawahar. A lip sync expert is all you need for speech to lip generation in the wild. In Proceedings of the 28th ACM international conference on multimedia, pages 484-492, 2020. 2, 3", + "[30] Alec Radford, Jong Wook Kim, Tao Xu, Greg Brockman, Christine McLeavey, and Ilya Sutskever. Robust speech recognition via large-scale weak supervision. In International conference on machine learning, pages 28492-28518. PMLR, 2023. 3", + "[31] Imogen C Rehm, Emily Foenander, Klaire Wallace, JoAnne M Abbott, Michael Kyrios, and Neil Thomas. What role can avatars play in e-mental health interventions? exploring new models of client-therapist interaction. Frontiers in Psychiatry, 7:186, 2016. 2", + "[32] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2", + "[33] Shuai Shen, Wenliang Zhao, Zibin Meng, Wanhua Li, Zheng Zhu, Jie Zhou, and Jiwen Lu. Difftalk: Crafting diffusion models for generalized audio-driven portraits animation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1982-1991, 2023. 2", + "[34] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792, 2022. 2", + "[35] Xusen Sun, Longhao Zhang, Hao Zhu, Peng Zhang, Bang Zhang, Xinya Ji, Kangneng Zhou, Daiheng Gao, Liefeng Bo, and Xun Cao. Vividtalk: One-shot audio-driven talking head generation based on 3d hybrid prior. arXiv preprint arXiv:2312.01841, 2023. 2", + "[36] Supasorn Suwajanakorn, Steven M Seitz, and Ira Kemelmacher-Shlizerman. Synthesizing obama: learn" + ], + "bbox": [ + 91, + 92, + 480, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ing lip sync from audio. ACM Transactions on Graphics (ToG), 36(4):1-13, 2017. 3", + "[37] Linrui Tian, Qi Wang, Bang Zhang, and Liefeng Bo. Emo: Emote portrait alive-generating expressive portrait videos with audio2video diffusion model under weak conditions. arXiv preprint arXiv:2402.17485, 2024. 2, 3, 4, 5, 7", + "[38] Thomas Unterthiner, Sjoerd van Steenkiste, Karol Kurach, Raphaël Marinier, Marcin Michalski, and Sylvain Gelly. Fvd: A new metric for video generation. 2019. 7", + "[39] Vikram Voleti, Alexia Jolicoeur-Martineau, and Chris Pal. Mcvd-masked conditional video diffusion for prediction, generation, and interpolation. Advances in neural information processing systems, 35:23371-23385, 2022. 3", + "[40] Cong Wang, Kuan Tian, Jun Zhang, Yonghang Guan, Feng Luo, Fei Shen, Zhiwei Jiang, Qing Gu, Xiao Han, and Wei Yang. V-express: Conditional dropout for progressive training of portrait video generation. arXiv preprint arXiv:2406.02511, 2024. 2, 6", + "[41] Xiang Wang, Hangjie Yuan, Shiwei Zhang, Dayou Chen, Jiuniu Wang, Yingya Zhang, Yujun Shen, Deli Zhao, and Jingren Zhou. Videocomposer: Compositional video synthesis with motion controllability. Advances in Neural Information Processing Systems, 36, 2024. 3", + "[42] Huawei Wei, Zejun Yang, and Zhisheng Wang. Aniportrait: Audio-driven synthesis of photorealistic portrait animation. arXiv preprint arXiv:2403.17694, 2024. 2, 3, 6", + "[43] Liangbin Xie, Xintao Wang, Honglun Zhang, Chao Dong, and Ying Shan. Vfhq: A high-quality dataset and benchmark for video face super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 657-666, 2022. 6", + "[44] Mingwang Xu, Hui Li, Qingkun Su, Hanlin Shang, Liwei Zhang, Ce Liu, Jingdong Wang, Luc Van Gool, Yao Yao, and Siyu Zhu. Hallo: Hierarchical audio-driven visual synthesis for portrait image animation. arXiv preprint arXiv:2406.08801, 2024. 2, 3, 4, 5, 6", + "[45] Sicheng Xu, Guojun Chen, Yu-Xiao Guo, Jiaolong Yang, Chong Li, Zhenyu Zang, Yizhong Zhang, Xin Tong, and Baining Guo. Vasa-1: Lifelike audio-driven talking faces generated in real time. arXiv preprint arXiv:2404.10667, 2024. 2, 3", + "[46] Fei Yin, Yong Zhang, Xiaodong Cun, Mingdeng Cao, Yanbo Fan, Xuan Wang, Qingyan Bai, Baoyuan Wu, Jue Wang, and Yujiu Yang. Styleheat: One-shot high-resolution editable talking face generation via pre-trained stylegan. In European conference on computer vision, pages 85-101. Springer, 2022. 3", + "[47] Shengming Yin, Chenfei Wu, Huan Yang, Jianfeng Wang, Xiaodong Wang, Minheng Ni, Zhengyuan Yang, Linjie Li, Shuguang Liu, Fan Yang, et al. Nuwa-xl: Diffusion over diffusion for extremely long video generation. arXiv preprint arXiv:2303.12346, 2023. 3", + "[48] Jianhui Yu, Hao Zhu, Liming Jiang, Chen Change Loy, Weidong Cai, and Wayne Wu. Celebv-text: A large-scale facial text-video dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14805-14814, 2023. 6" + ], + "bbox": [ + 516, + 92, + 906, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "202", + "bbox": [ + 485, + 945, + 511, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[49] Rui Zhang, Yixiao Fang, Zhengnan Lu, Pei Cheng, Zebiao Huang, and Bin Fu. Linguualinker: Audio-driven portraits animation with implicit facial control enhancement. arXiv preprint arXiv:2407.18595, 2024. 2", + "[50] Wenxuan Zhang, Xiaodong Cun, Xuan Wang, Yong Zhang, Xi Shen, Yu Guo, Ying Shan, and Fei Wang. Sadtalker: Learning realistic 3d motion coefficients for stylized audio-driven single image talking face animation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8652-8661, 2023. 2, 3, 4, 6", + "[51] Zhimeng Zhang, Lincheng Li, Yu Ding, and Changjie Fan. Flow-guided one-shot talking face generation with a high-resolution audio-visual dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3661–3670, 2021. 6", + "[52] Hao Zhu, Wayne Wu, Wentao Zhu, Liming Jiang, Siwei Tang, Li Zhang, Ziwei Liu, and Chen Change Loy. Celebvyhq: A large-scale video facial attributes dataset. In European conference on computer vision, pages 650-667. Springer, 2022. 6" + ], + "bbox": [ + 89, + 90, + 483, + 371 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "203", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Sonic_ Shifting Focus to Global Audio Perception in Portrait Animation/108dd369-6bdf-4c82-9e76-1c8b1e3be7c1_model.json b/2025/Sonic_ Shifting Focus to Global Audio Perception in Portrait Animation/108dd369-6bdf-4c82-9e76-1c8b1e3be7c1_model.json new file mode 100644 index 0000000000000000000000000000000000000000..d723375fa60eff492a8d9f39420103954ad8d731 --- /dev/null +++ b/2025/Sonic_ Shifting Focus to Global Audio Perception in Portrait Animation/108dd369-6bdf-4c82-9e76-1c8b1e3be7c1_model.json @@ -0,0 +1,1949 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.138, + 0.131, + 0.86, + 0.154 + ], + "angle": 0, + "content": "Sonic: Shifting Focus to Global Audio Perception in Portrait Animation" + }, + { + "type": "text", + "bbox": [ + 0.12, + 0.18, + 0.877, + 0.237 + ], + "angle": 0, + "content": "Xiaozhong Ji\\(^{1*}\\) Xiaobin Hu\\(^{1*}\\) Zhihong Xu\\(^{2}\\) Junwei Zhu\\(^{1\\dagger}\\) Chuming Lin\\(^{1}\\) Qingdong He\\(^{1}\\) Jiangning Zhang\\(^{1}\\) Donghao Luo\\(^{1}\\) Yi Chen\\(^{1}\\) Qin Lin\\(^{1}\\) Qinglin Lu\\(^{1}\\) Chengjie Wang\\(^{1}\\) \n\\(^{1}\\)Tencent \\(^{2}\\)Zhejiang University" + }, + { + "type": "image", + "bbox": [ + 0.139, + 0.258, + 0.44, + 0.352 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.462, + 0.258, + 0.856, + 0.353 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.138, + 0.355, + 0.431, + 0.478 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.445, + 0.356, + 0.851, + 0.478 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.48, + 0.907, + 0.537 + ], + "angle": 0, + "content": "Figure 1. Sonic excels in producing vivid portrait animation videos given a reference image and an audio clip. Beyond its fundamental lip-syncing capabilities, Sonic demonstrates proficiency in creating a diverse spectrum of facial expressions and adaptable head movements. Notably, when dealing with extended videos, Sonic can yield stable and seamless outcomes in a parallel fashion, all while maintaining an unified paradigm focusing on global audio perception." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.549, + 0.327, + 0.563 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.581, + 0.486, + 0.854 + ], + "angle": 0, + "content": "The study of talking face generation mainly explores the intricacies of synchronizing facial movements and crafting visually appealing, temporally-coherent animations. However, due to the limited exploration of global audio perception, current approaches predominantly employ auxiliary visual and spatial knowledge to stabilize the movements, which often results in the deterioration of the naturalness and temporal inconsistencies. Considering the essence of audio-driven animation, the audio signal serves as the ideal and unique priors to adjust facial expressions and lip movements, without resorting to interference of any visual signals. Based on this motivation, we propose a novel paradigm, dubbed as Sonic, to shift focus on the exploration of global audio perception. To effectively leverage global audio knowledge, we disentangle it into intra- and inter-clip audio perception and collaborate with both aspects to enhance overall perception. For the intra-clip audio perception, 1). Context-enhanced audio learning, in" + }, + { + "type": "page_footnote", + "bbox": [ + 0.09, + 0.863, + 0.454, + 0.901 + ], + "angle": 0, + "content": "* Equal Contribution, † Project Leader \nCorresponding Author \nProject page: https://jixiaozhong.github.io/Sonic/." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.55, + 0.909, + 0.779 + ], + "angle": 0, + "content": "which long-range intra-clip temporal audio knowledge is extracted to provide facial expression and lip motion priors implicitly expressed as the tone and speed of speech. 2). Motion-decoupled controller, in which the motion of the head and expression movement are disentangled and independently controlled by intra-audio clips. Most importantly, for inter-clip audio perception, as a bridge to connect the intra-clips to achieve the global perception, Time-aware position shift fusion, in which the global inter-clip audio information is considered and fused for long-audio inference via through consecutively time-aware shifted windows. Extensive experiments demonstrate that the novel audio-driven paradigm outperform existing SOTA methodologies in terms of video quality, temporally consistency, lip synchronization precision, and motion diversity." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.813, + 0.646, + 0.829 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.908, + 0.903 + ], + "angle": 0, + "content": "Talking Face Animation is a technique aimed at creating a lifelike representation of a speaking person by animating a still image in sync with speech audio. This technology holds great potential in enhancing the realism of virtual re" + }, + { + "type": "page_number", + "bbox": [ + 0.488, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "193" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.243 + ], + "angle": 0, + "content": "ality [5], improving the efficiency of film and gaming making [25], facilitating accessibility for individuals with communication impairments [24], and offering therapeutic support and social interaction in healthcare [31]. Prominent progress in this domain is demonstrated by notable works based on Stable Diffusion [32], which project the parametric or implicit representations of lip movements [29, 50], facial expressions [16, 28, 33], and head movement [35] in the latent space [11, 27, 37, 45] to generate high-quality video with a high level of realism and liveliness." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.248, + 0.485, + 0.534 + ], + "angle": 0, + "content": "However, precise audio control and temporal coherence have not yet been effectively achieved. Current approaches [8, 23, 37, 40, 42, 44, 49, 50] predominantly addressed these two issues in a completely segregated manner, overlooking the holistic coordination between audio and vision. For audio control, audio features were segmented according to timestamps and matched to each visual frame. By applying cross-attention within a spatial latent space, each frame was restricted to the nearby audio information, which hindered the transformation into an optimal motion representation. To achieve temporal coherence, self-attention along the temporal dimension was applied to smooth visual features within a clip, while overlapping or motion frames were used to enhance coherence between different clips. However, these stability strategies have limited temporal receptive fields, which can reduce the richness of motion. Additionally, they do not take audio information into account, potentially disrupting the synchronization between audio and visual elements." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.54, + 0.484, + 0.902 + ], + "angle": 0, + "content": "To address these issues, we propose a novel paradigm, dubbed as Sonic, to extent the range of receptive field to the global level guided by the audio priors. To achieve global audio perception, we decompose the talking head video generation task into two aspects: intra-clip (i.e., lip sync, expression alignment, and head movement) and inter-clip (i.e., motion rhythm and smoothness, long video stability). Sonic aims to exploit global audio perception rather than the motion frames and other visual motion (i.e., referring to the face locator and landmark prior adopted by existing methods, which hinders the ability to perceive information directly from audio). Performing as long-range and global signal, audio shows tone, rhythm and speed of speech which implicitly provides expression and head movement motion priors. However, we have to face such a tricky challenge that, as a weak correlation cross-modality signal without the guidance of any visual motion (e.g., motion frames), current audio learning paradigm leads to the temporal jitter and unsatisfactory pattern and fails to well-align the audio and motion. To this end, we first customize efficient global audio perception learning mechanism. Specifically, context-enhanced audio learning is proposed as long-range audio learning module to extract the intra-clip audio temporal knowledge from the input audio clip. Then such intra" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.908, + 0.497 + ], + "angle": 0, + "content": "clip audio temporal information is mapped as a temporal-embeddings for the latter fusion by audio-temporal cross-attention. The temporal-embeddings projection operation can effectively reduce the computational burden and capture temporal priors within current audio clip. Although context-enhanced audio learning directs the temporal movements correlated with audio, such as lip movements synchronized with speech and facial expressions related to the content, there still exists the absence of habitual head movements. Secondly, we design motion-decoupled controller to disentangle the habitual head and expression movement and support the independent control via two explicit parameters learned from the current intra-clip audio. To enrich the functionality and playability of animation, we also add an plug to allow users to define some exaggerated movements. Most importantly, for inter-clip audio perception, time-aware position shift fusion enlarges the intra-clip audio perception to the global inter-clip audio perception via time-aware shifted windows consecutively bridging the preceding clip, providing the global inter-clip connection that significantly enhance modeling temporal power, as demonstrated in Figure 1. It is worth noting that our time-aware position shift fusion requires no extra training cost compared with the motion frames in most recent works, such as EMO [37], Loopy [23] and also no additional inference time caused by overlapping frames. Our contributions are summarized as:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.501, + 0.905, + 0.546 + ], + "angle": 0, + "content": "- We propose a novel unified paradigm without the guidance of visual motion, dubbed as Sonic, to focus on the exploration of global audio perception." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.547, + 0.905, + 0.605 + ], + "angle": 0, + "content": "- For intra-clip audio perception, context-enhanced audio learning and motion-decoupled controller excavate the intra-clip audio temporal knowledge to direct the temporal movements." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.607, + 0.905, + 0.667 + ], + "angle": 0, + "content": "- For inter-clip audio perception, we propose time-aware position shift fusion to enhance the global inter-clip audio perception by extending the intra-clip audio perception through consecutively time-aware shifted windows." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.668, + 0.905, + 0.727 + ], + "angle": 0, + "content": "- Extensive experiments demonstrate that the novel unified global audio perception paradigm outperforms existing SOTA methods in terms of video quality, temporal consistency, lip synchronization, and motion diversity." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.501, + 0.905, + 0.727 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.745, + 0.663, + 0.76 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.771, + 0.81, + 0.785 + ], + "angle": 0, + "content": "2.1. Diffusion-based Video Generation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.907, + 0.903 + ], + "angle": 0, + "content": "The remarkable achievements in diffusion models in image generation have sparked widespread research interest for video generation. Early endeavors in video generation [14, 19, 20, 26, 34] have primarily concentrated on devising image-based diffusion models with temporal module to effectively capture the dynamic characteristics inherent in video sequences. To leverage the powerful abil" + }, + { + "type": "page_number", + "bbox": [ + 0.488, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "194" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.094, + 0.092, + 0.482, + 0.393 + ], + "angle": 0, + "content": "ity of pretrained image diffusion models, Video LDM [4] incorporates temporal layers that learn to align images in a temporally consistent manner. AnimateDiff [15] provides a motion-specific module by multi-stage progressive training manner, and the motion module can be seamlessly integrated into existing text-to-image architectures without any further modifications. VideoComposer [41] and VideoCrafter [6] further explore the synthesis of image-to-video generation via textual and visual features combinations. Despite these notable advancements, the field still struggles to overcome the challenge of generating temporal coherent and high-quality videos. As a powerful motion representation trained on a large well-curated high-quality dataset, Stable Video Diffusion [3] shows the remarkable temporal-consistency and the high-quality visual characters for downstream tasks. There exists a large gap between the audio-driven portrait animation and general video generation due to the fact that audio-driven task aims to use the weak cross-modality audio signal to precisely control motion movements (e.g., lip and expression movement)." + }, + { + "type": "title", + "bbox": [ + 0.094, + 0.42, + 0.421, + 0.435 + ], + "angle": 0, + "content": "2.2. Audio-driven Talking Face Generation" + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.449, + 0.482, + 0.9 + ], + "angle": 0, + "content": "The generation of talking face videos from audio inputs has been a longstanding task. In the early stages, the emphasis was primarily on synthesizing lip movements alone, accomplished by directly mapping audio signals to lip movements while keeping other facial attributes unchanged [7, 22, 29, 36, 46]. In more recent efforts, the scope has been broadened to encompass a wider range of facial expressions and head movements, which are derived from audio inputs and a single reference image. These methods can be broadly classified into with or without intermediate representation. For instance, by employing 3D coefficients representation (e.g., Sadtalker [50], Dreamtalk [28], Ani-Portrait [42], VASA-1 [45]), the video generation process can be informed through the direct or indirect visible control signal in the aspect of head movement and expression. Nevertheless, a recurring challenge faced by these techniques is the limited capacity of the 3D mesh to capture intricate details, consequently constraining the overall dynamism and realism. In contrast, the methods that omit intermediate representation exhibit superior naturalness and consistent preservation of identity with the original image. EMO [37], Hallo [44], and Loopy [23] employ a direct audio-to-video synthesis approach to ensure a high level of realism and naturalness, eliminating the need for intermediate 3D representation. However, these approaches primarily rely on additional conditions associated with spatial motion or motion frames to maintain temporal consistency. In this regard, they overlook the crucial role of audio in audio-driven animation, consequently diminishing both expressiveness and audio-motion synchronization." + }, + { + "type": "title", + "bbox": [ + 0.518, + 0.092, + 0.715, + 0.107 + ], + "angle": 0, + "content": "2.3. Long Video Inference" + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.115, + 0.904, + 0.371 + ], + "angle": 0, + "content": "Significant research efforts have been devoted to extending the duration of generated videos, aiming to expand their practical applications. Autoregressively predicting successive frames [17, 39] and hierarchical coarse-to-fine methods [47] have emerged to generate long videos. Then Lumiere [2] divides the video into overlapping temporal segments, denoises each segment independently, and finally fuse these segments in an optimization algorithm. However there is no available high-efficiency method for audio-condition global perception. As one step towards achieving stronger capabilities, our Sonic proposes time-aware position shift fusion strategy progressively denoise intra-clip audio to establish a global inter-clip connection through the use of time-aware shifted windows that bridge the preceding clip along the timesteps axis. This inter-clip fusion with global receptive field needs no additional training cost and introduce no extra inference time without overlapping frames." + }, + { + "type": "title", + "bbox": [ + 0.518, + 0.388, + 0.646, + 0.403 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.414, + 0.904, + 0.533 + ], + "angle": 0, + "content": "Overall framework. Given a single portrait reference image and an input audio, our one-stage framework can generate a portrait video well-synchronized with the driven audio. Our proposed Sonic framework aims to focus on the exploration of global audio perception, primarily consisting of context-enhanced audio learning, motion-decoupled controller, and time-aware position shift fusion module, as illustrated in Figure 2." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.535, + 0.904, + 0.669 + ], + "angle": 0, + "content": "Technical contribution. To achieve global audio perception, we decompose the talking head video generation task into two aspects: intra-clip (i.e., lip sync, expression alignment, and head movement) and inter-clip (i.e., motion rhythm and smoothness, long video stability). For intra-clip, we enhance the visual realism that matches audio signal through the context-enhanced audio learning and the motion controller. For inter-clip, we improve the stability using the time-aware position shift fusion strategy." + }, + { + "type": "title", + "bbox": [ + 0.518, + 0.682, + 0.815, + 0.696 + ], + "angle": 0, + "content": "3.1. Context-enhanced Audio Learning" + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.705, + 0.904, + 0.899 + ], + "angle": 0, + "content": "For audio-talking face animation, the audio is the ideally unique and global signal to adjust the face expression and movement without interference from other visual signals. Thus, the motivation behind context-enhanced audio learning is to guide the generation of spatial and temporal frames solely based on audio signals, rather than motion frames or other visual signals, to achieve the high-realism generation. Specifically, we adopt Whisper-Tiny [30] as the pretrained audio feature extraction model, which is more lightweight than Wav2Vec [1] used by most existing methods [8, 37, 44]. For each audio frame, the features from the last layers of five stages are concatenated together for multiscale understanding. For each video frame, a duration of" + }, + { + "type": "page_number", + "bbox": [ + 0.489, + 0.947, + 0.51, + 0.957 + ], + "angle": 0, + "content": "195" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.137, + 0.092, + 0.353, + 0.342 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.368, + 0.092, + 0.501, + 0.339 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.096, + 0.861, + 0.342 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.343, + 0.908, + 0.385 + ], + "angle": 0, + "content": "Figure 2. Framework of our approach. Sonic processes each clip of the long audio in parallel, shifting to a new context at each time step to progressively fuse inter-clip latent features across global audio perception. In Sonic, we enhance intra-clip temporal audio context learning and decouple motion to improve dynamics." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.4, + 0.484, + 0.552 + ], + "angle": 0, + "content": "0.2s audio feature is used frame-wise to give rich context. We use three linear layers to project audio feature to match the dimension of cross attention. The transformed audio embedding is denoted as \\( c_{a} \\in \\mathcal{R}^{b \\times f \\times d \\times c} \\), where \\( b \\) denotes batch size, \\( f \\) is audio length, \\( d \\) is num of context tokens, and \\( c \\) is hidden size of the cross-attention. To focus on talking face area, the core latent features is restricted with a mask \\( M \\) determined by a joint bounding boxes of faces. The audio signal can act as the spatial knowledge provider being injected into the spatial cross-attention layers as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.116, + 0.558, + 0.483, + 0.577 + ], + "angle": 0, + "content": "\\[\nz _ {s} ^ {\\prime} = z _ {s} + \\operatorname {C r o s s A t t n} \\left(Q \\left(z _ {s}\\right), K \\left(c _ {a}\\right), V \\left(c _ {a}\\right)\\right) \\cdot M, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.584, + 0.483, + 0.629 + ], + "angle": 0, + "content": "where \\( z_{s} \\) is spatial latent features of self-attention layer, and \\( z_{s}^{\\prime} \\) is the adjusted spatial features guided by audio signals in spatial-aware level." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.63, + 0.483, + 0.841 + ], + "angle": 0, + "content": "Furthermore, audio is a long-range and global signal that provides expression and head movement motion priors implicitly expressed through the tone and speed of speech. To excavate the temporal information from the long-range audio knowledge, we customize a temporal-audio module for precise motion control. The temporal module in existing methods [37, 50] followsAnimateDiff, which performs only self-attention across the temporal dimension without audio aids. Some other talking-face [23, 44] works extract the motion priors from the visual knowledge (e.g., motion frames), which deviates from the essence of audio-driven task that relies solely on audio to drive movement. Thus we propose a temporal audio cross-attention module, introducing audio embedding to guide temporal alignment." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.841, + 0.484, + 0.902 + ], + "angle": 0, + "content": "The preceding latent from temporal self-attention layers can be reshaped as \\( z_{t} \\in \\mathcal{R}^{(b \\times h \\times w) \\times f \\times c} \\). To reduce the computation burden, the audio feature \\( c_{a} \\) is average-pooled along temporal dimension as the shape \\( \\mathcal{R}^{(b \\times 1) \\times d \\times c} \\), then" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.4, + 0.907, + 0.446 + ], + "angle": 0, + "content": "repeated and reshaped as \\( c_{ta} \\in \\mathcal{R}^{(b \\times h \\times w) \\times d \\times c} \\). The projected audio temporal-embeddings are injected into the denoising U-Net via a temporal cross-attention blocks as:" + }, + { + "type": "equation", + "bbox": [ + 0.557, + 0.452, + 0.907, + 0.471 + ], + "angle": 0, + "content": "\\[\nz _ {t} ^ {\\prime} = z _ {t} + \\operatorname {C r o s s A t t n} (Q (z _ {t}), K (c _ {t a}), V (c _ {t a})). \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.476, + 0.907, + 0.522 + ], + "angle": 0, + "content": "The spatial and temporal audio attention layers are cascaded in multiple down-stages and up-stages to enable natural appearance and smooth temporal movement." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.53, + 0.776, + 0.546 + ], + "angle": 0, + "content": "3.2. Motion-decoupled Controller" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.553, + 0.907, + 0.84 + ], + "angle": 0, + "content": "Context-enhanced Audio Learning governs the temporal movements strongly correlated with audio, such as lip movements synchronized with speech and facial expressions related to the content. On the other hand, our motion-decoupled controller performs a complementary role by directing the temporal movements weakly associated with audio, such as habitual head movements and random changes in perspective. This controller can disentangle motion into explicit head and expression movement amplitudes, thereby enhancing both functionality, playability, and interactivity. Thus, we introduce independent augmented motion-bucket parameters which effectively influence motion patterns of head movements and expression strength. During the training phase, the translation motion-bucket \\( m_t \\) is computed as the variance of bounding boxes of video clip, while the expression motion-bucket \\( m_e \\) is variance of relative landmarks. Both buckets are integers in the range of [0, 128]. Through position encoding and linear projection, they are added into ResNet blocks as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.618, + 0.848, + 0.905, + 0.865 + ], + "angle": 0, + "content": "\\[\ne m b = W \\left[ \\operatorname {P e} \\left(m _ {t}\\right), \\operatorname {P e} \\left(m _ {e}\\right) \\right], \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.907, + 0.902 + ], + "angle": 0, + "content": "where \\(emb\\) is the embedding, \\(\\mathrm{Pe}\\) is the position encoding, and \\(W\\) represents the linear projection weights." + }, + { + "type": "page_number", + "bbox": [ + 0.488, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "196" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.273 + ], + "angle": 0, + "content": "For practice application, motion-decoupled controller also supports the automatically adjustable parameters learning from the audio and reference image, and deeply explore the correlation between audio and portrait reference without requiring the parameter setting from users. To simplify multiple parameters and maintain adjustability, the predicted motion buckets are multiplying by a scale \\(\\beta\\) (0.5 for mild dynamic, 1.0 for moderate dynamic, and 2.0 for intense dynamic). Specifically, we utilize a audio-to-bucket manner to adaptively predict the motion-buckets relying on the audio and reference image. In inference phase, the motion buckets are predicted and rescaled as:" + }, + { + "type": "equation", + "bbox": [ + 0.193, + 0.278, + 0.483, + 0.298 + ], + "angle": 0, + "content": "\\[\nm _ {t} ^ {\\prime}, m _ {e} ^ {\\prime} = \\beta \\times \\mathrm {E} _ {\\mathrm {b}} \\left(c _ {a}, R _ {i m g}\\right). \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.306, + 0.484, + 0.381 + ], + "angle": 0, + "content": "where \\( m_t' \\), \\( m_e' \\) are the estimated translation and expression motion-bucket achieved by the audio \\( c_a \\) and single reference image CLIP embedding \\( R_{img} \\), and \\( E_b \\) is the audio-to-bucket projection consisting of there linear layers with the ReLU as activation function." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.39, + 0.383, + 0.404 + ], + "angle": 0, + "content": "3.3. Time-aware Position Shift Fusion" + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.427, + 0.465, + 0.539 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.542, + 0.484, + 0.611 + ], + "angle": 0, + "content": "Figure 3. Illustration of the proposed time-aware position-shift fusion. The model processes each clip non-overlapping. In next timestep, the model starts from a new position determined by the offset, thereby integrating long-range context. Specifically, the tail latents are filled cyclically from the head." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.629, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Recent methods [23, 37, 44] use motion frames or overlapping latent during coherent denoising, which suffer from a limited receptive field and increased the training or inference computational complexity for audio-conditioned video generation problem. In contrast, we propose a bridge, the time-aware position shift fusion, to extent the intra-clip audio perception to the global inter-clip audio-perception via consecutively time-aware shifted windows. The simple yet effective fusion strategy, without increasing extra inference or training time costs, can effectively address jitter and abrupt transition issues. Audio serves as the input condition for the talking face, where longer audio prompts the generation of longer videos. Based on this principle, we tailor time-aware position-shift fusion mechanism along the timestep axis where the latent sliding window is progressively shifted by a certain length at different timesteps while also preserving the frame correspondence relationship between audio and video signal." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.424 + ], + "angle": 0, + "content": "Specifically, as described in Figure 3 and Algorithm 1, our time-aware position-shift fusion consists of two nested loops. The outer loop is inverse diffusion process, while the inner loop is a sliding window process that model predicts for the full-length audio conditions. At each timestep, the model takes a clip of audio input to predict the corresponding latent with a start offset to smooth the segment from the last timestep. The video diffusion model naturally bridges the context, generating continuous movement following the audio prompts. In detail, we set a cumulative shift offset \\(\\alpha_{\\Sigma}\\), accumulating \\(\\alpha\\) at each timestep to ensure that the model starts sliding window from different positions. The offset \\(\\alpha\\) is experimentally set to a small integer, such as 3, 7; we will report analysis on this parameters in experiments part. Note that the relative positioning between the latents and the audio prompts has not been altered. The only modification occurs in their recombination during the input phase to the model. In a special case, at the end of the long sequence, the shifted position index may exceed the sequence length. To address this issue, we adopt a circular padding strategy, padding the start audio prompt and latents to the end." + }, + { + "type": "code_caption", + "bbox": [ + 0.516, + 0.434, + 0.826, + 0.448 + ], + "angle": 0, + "content": "Algorithm 1 Time-aware Position-shift Fusion" + }, + { + "type": "algorithm", + "bbox": [ + 0.516, + 0.454, + 0.907, + 0.753 + ], + "angle": 0, + "content": "Require: Audio embedding \\(c_{a}^{[0,l]}\\) with length \\(l\\), denoising steps \\(T\\), initial noisy latent \\(z_{T}^{[0,l]}\\), pretrained Sonic model \\(\\mathrm{ST}(\\cdot)\\) for sequence length \\(f\\), position-shift offset \\(\\alpha < f < l\\). \nEnsure: Denoised latent \\(z_{0}^{[0,l]}\\). \n1: Initialize accumulated shift offset \\(\\alpha_{\\Sigma} = 0\\). \n2: for \\(t = T, \\dots, 1\\) do \n3: // Denoising loop \n4: Initialize start point \\(s = \\alpha_{\\Sigma}\\), end \\(e = s + f\\), processed length \\(n = 0\\). // start from new position for each timestep. \n5: while \\(n < l\\) do \n6: // Sequence loop \n7: \\(z_{t-1}^{[s,e]} = \\mathrm{ST}(z_{t}^{[s,e]}, c_{a}^{[s,e]}, t)\\) \n8: \\(s \\gets s + f, e \\gets e + f, n \\gets n + f\\). // Move to next clip non-overlapping \n9: if \\(s > l\\) or \\(e > l\\) then \n10: \\(s \\gets s\\% l, e \\gets e\\% l\\). // Padding circularly \n11: end if \n12: end while \n13: \\(\\alpha_{\\Sigma} \\gets \\alpha_{\\Sigma} + \\alpha\\). // Accumulate shift offset \n14: end for \n15: return Denoised latent \\(z_{0}^{[0,l]}\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Computation cost. We analyze the computation efficiency of our time-aware position shift fusion compared to motion frames and overlapping. Denote the FLOPs of a single forward pass of the video diffusion model as \\(\\Omega\\), the number of frames to generate as \\(n \\times f\\), and the denoising steps as \\(T\\). The accumulated FLOPs of Sonic is: \\(\\Omega(Shift) = \\Omega \\times T \\times n\\). For overlapping, the overlapped latents (overlap \\(o\\)) are calculated twice, \\(\\Omega(Overlap) = \\Omega \\times T \\times \\left(n + \\frac{o}{f}\\right)\\). For motion frames, the FLOPs of a single forward of Ref-" + }, + { + "type": "page_number", + "bbox": [ + 0.488, + 0.946, + 0.511, + 0.956 + ], + "angle": 0, + "content": "197" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.908, + 0.118 + ], + "angle": 0, + "content": "Table 1. Quantitative comparisons with the state-of-the-arts on the HDTF and CelebV-HQ test set. * indicates GAN-based methods and others are Diffusion-based methods. The best results are bold, and the second are underlined." + }, + { + "type": "table", + "bbox": [ + 0.136, + 0.122, + 0.86, + 0.306 + ], + "angle": 0, + "content": "
DatasetMethodFID↓FVD↓Sync-C↑Sync-D↓E-FID↓F-SIM↑Smooth↑Runtime↓
HDTFSadTalker* (2023) [50]61.672397.1141.75510.6952.4820.92870.99613.75
Aniportrait (2024) [42]36.965471.4521.09512.4613.1610.95080.992144.03
V-Express (2024) [40]47.396758.0231.25612.3942.6340.90930.996839.04
Hallo (2024) [44]30.176347.3584.0609.5511.7920.95550.994174.65
Hallo2 (2024) [12]38.673328.5404.1369.4652.2030.96060.994245.75
EchoMimic (2024) [8]33.207384.3042.51410.7431.4860.95270.99345.45
Sonic (Ours)29.104301.1734.1979.3711.7450.95950.997017.04
CelebV-HQSadTalker* (2024) [50]57.574841.9621.97810.9152.2520.94340.99593.75
Aniportrait (2024) [42]53.746590.3730.99612.0843.2960.95220.991144.03
V-Express (2024) [40]65.400889.9850.80913.2552.7130.90190.996539.04
Hallo (2024) [44]47.403488.4992.68010.2922.2730.96070.994274.65
Hallo2 (2024) [12]52.396481.3362.63810.3432.8190.95870.994245.75
EchoMimic (2024) [8]48.267596.8701.94910.7542.1360.96120.99325.45
Sonic (Ours)43.137483.1082.68910.1941.7830.96240.997217.04
" + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.321, + 0.472, + 0.622 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.476, + 0.321, + 0.905, + 0.622 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.632, + 0.908, + 0.677 + ], + "angle": 0, + "content": "Figure 4. Qualitative comparisons with State-of-the-Art talking head generation methods. Due to image does not reflect important sync, naturalness and stability, the full video comparison will be included in supplementary materials as well as comparison with demos from other non-open source works." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.69, + 0.483, + 0.77 + ], + "angle": 0, + "content": "erence Net is \\(\\omega_{r}\\), the number of motion frames is \\(o\\) and the Flops of motion modules computing \\(f\\) frames is \\(\\omega_{m}\\), \\(\\Omega (MF) = \\Omega \\times T\\times n + \\omega_{r}\\times o\\times n + \\omega_{m}\\times T\\times n\\times \\frac{2of + o^{2}}{f^{2}}\\). Overall, the proposed time-aware position shift fusion significantly reduces the computation complexity." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.788, + 0.224, + 0.806 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.815, + 0.274, + 0.832 + ], + "angle": 0, + "content": "4.1. Experiments Setup" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Implementation Details. The Sonic training is in a single stage, with initialization of the spatial module and temporal modules from stable-video-diffusion-xt-1-1[3]. To separately enable different conditions in training, we manipulate" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.69, + 0.907, + 0.781 + ], + "angle": 0, + "content": "the data such that \\(5\\%\\) of it drops audio, another \\(5\\%\\) drops image, and \\(5\\%\\) drops both condition. In inference phase, multi-condition CFG is performed. The CFG of the reference image \\(r_i\\) is set to 2.0, and the CFG of the audio \\(r_a\\) is set to 7.5. In our full model, the dynamic scale \\(\\beta\\) is set to 1.0 and the shift offset \\(\\alpha\\) is set to 7." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Datasets and Evaluation Metrics. For the training set, we created a diverse dataset by aggregating data from open-source datasets, including VFHQ [43], CelebV-Text [48], VoxCeleb2 [10]. For the test set, we conducted experiments on two widely used datasets, HDTF [52] and CelebV-HQ [51]. We randomly selected 50 four-second clips from CelebV-HQ and 100 from HDTF. We utilize seven met" + }, + { + "type": "page_number", + "bbox": [ + 0.488, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "198" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.133, + 0.089, + 0.861, + 0.362 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.365, + 0.907, + 0.42 + ], + "angle": 0, + "content": "Figure 5. Qualitative results under different styles of portrait images and various types of audio inputs. The images and audios were collected from recent works and the Internet. The upper section presents results for audio input ranging from 20 seconds to 2 minutes in duration, while the lower section shows results for longer audio inputs, up to 10 minutes. Our Sonic demonstrates versatility across various portrait styles and maintains vividness over extended durations." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.422, + 0.484, + 0.573 + ], + "angle": 0, + "content": "rics to evaluate. The Fréchet Inception Distance (FID) [18] and Fréchet Video Distance (FVD) [38] metrics measure the quality of the generated data. Expression-FID (E-FID) [37] uses face reconstruction method [13] to extract features and calculate the FID between generated and ground truth frames, while F-SIM measures facial similarity. SyncNet (Sync-C&Sync-D) [9] and VBench's smooth metric [21] assess lip synchronization and motion fluidity. In ablation experiments, we employ VBench's dynamic degree metric to evaluate amplitude of head movements." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.588, + 0.286, + 0.604 + ], + "angle": 0, + "content": "4.2. Results and Analysis" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.612, + 0.483, + 0.854 + ], + "angle": 0, + "content": "Quantitative Comparison. In Table 1, our Sonic are significantly superior than other recent 6 methods in terms of FID, Sync-confidence and Smoothness evaluated on the CelebV-HQ and HDTF datasets. On the HDTF dataset, our Sonic achieve the lowest FID and FVD scores, \\(8\\%\\) lower than the second-best, indicating superior generation quality and alignment with real data. For the CelebV-HQ dataset, which features more complex characters and backgrounds, our method maintains superior performance across most metrics and competitive FVD. We notably excel in E-FID, \\(16.5\\%\\) lower than the second-best, which demonstrates our method has superior expression diversity. EMO [37] as a representative work that tackles talking head generation task, remains closed source, and not available for fair quantitative comparison. We provide video results using input from their online demos for a qualitative comparison." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Qualitative Comparison. Figure 4 provides a visualization comparison on open datasets. Analyzing the results, Previous approaches relied on motion frames, limiting the rich" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.422, + 0.907, + 0.527 + ], + "angle": 0, + "content": "ness of generated expressions and the dynamism of head movements. In contrast, our proposed framework incorporates global audio perception, enhancing the integration of audio and temporal variables while providing a deeper understanding of the audio input. This enables our method to generate a broader range of expressions that align with the audio and facilitate more natural head movements." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.529, + 0.909, + 0.667 + ], + "angle": 0, + "content": "Visualization Results in Complex Scenarios. We further explore the generation performance of our method under different input conditions. In terms of images, we employ three styles: real human, animation, and 3D. Each character portrait is paired with various types of audio relevant to its theme, such as singing, speech, and rap. As shown in Figure 5, our method successfully generates a wide array of expressions and diverse movements across various categories of input conditions, demonstrating strong robustness." + }, + { + "type": "table_caption", + "bbox": [ + 0.564, + 0.682, + 0.854, + 0.696 + ], + "angle": 0, + "content": "Table 2. User study comparison on open dataset." + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.697, + 0.898, + 0.758 + ], + "angle": 0, + "content": "
Method / MetricLip syncMotion diversityID consistencyVideo Smoothness
Aniportrait1.421.623.112.09
SadTalker1.982.342.952.95
Echomimic2.772.653.482.71
Hallo23.152.373.342.94
Sonic(Ours)4.58 (45%↑)4.55 (72%↑)4.29 (23%↑)4.66 (58%↑)
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.907, + 0.901 + ], + "angle": 0, + "content": "User Studies We conducted subjective evaluation on open dataset to assess the comparative methods along four key dimensions: Lip sync, motion diversity, identity similarity, and video smoothness. Total 40 participants gives scores from 1 to 5 for results of five comparative methods. In Table 2, our Sonic outperforms the others in all four dimensions, especially showing a significant \\(72\\%\\) improvement in motion diversity. The video results for user study will be included in the supplementary materials." + }, + { + "type": "page_number", + "bbox": [ + 0.488, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "199" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.111, + 0.09, + 0.462, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.092, + 0.264, + 0.483, + 0.29 + ], + "angle": 0, + "content": "Figure 6. Visual results for ablation on temporal audio attention and motion controller." + }, + { + "type": "image", + "bbox": [ + 0.107, + 0.291, + 0.468, + 0.457 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.092, + 0.459, + 0.483, + 0.485 + ], + "angle": 0, + "content": "Figure 7. Demonstration on effect of time-aware position shift fusion." + }, + { + "type": "table_caption", + "bbox": [ + 0.092, + 0.493, + 0.483, + 0.534 + ], + "angle": 0, + "content": "Table 3. Quantitative results of ablations on CelebV-HQ along with our results obtained using different dynamic scales. In the Sonic model, the default setting for \\(\\beta\\) is 1.0, i.e. Ours-Moderate." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.535, + 0.479, + 0.619 + ], + "angle": 0, + "content": "
Method / MetricSync-C↑Sync-D↓E-FID↓Smooth↑Dynamic↑
w/o temporal-audio2.61010.3102.3110.99690.62
w/o motion-controller2.19310.7313.2310.99690.22
w/o shift-fusion2.50910.3792.2020.99690.76
Ours-Mild (β = 0.5)2.65510.2422.0600.99700.78
Ours-Intense (β = 2.0)2.80310.1252.0380.99690.98
Ours-Moderate (β = 1.0)2.68910.1941.7830.99720.86
" + }, + { + "type": "table_caption", + "bbox": [ + 0.092, + 0.625, + 0.483, + 0.654 + ], + "angle": 0, + "content": "Table 4. Quantitative comparison on motion frames, overlap and the proposed time-aware shift fusion." + }, + { + "type": "table", + "bbox": [ + 0.102, + 0.655, + 0.466, + 0.717 + ], + "angle": 0, + "content": "
Method / MetricSync-C↑Sync-D↓Smooth↑Runtime↓
Motion frames(8)2.46510.3230.996919.70
Overlap(8)2.54410.2650.996922.49
Shift-fusion2.68910.1940.997217.04
" + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.726, + 0.255, + 0.74 + ], + "angle": 0, + "content": "4.3. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Temporal Control Inspires Intra-clip Diversity. In Table 3, the absence of temporal audio attention results in a decrease in both lip-sync and E-FID. We observe that the motion controller plays a crucial role in CelebV-HQ given that its videos exhibit more movement that weakly correlates with audio. We also investigate the effect of different dynamic scale. Increasing \\(\\beta\\) can further stimulate the lip-audio synchronization and the diversity of movements, while there is a slight decrease in stability and expression similarity. The instances in Figure 6 indicate" + }, + { + "type": "image", + "bbox": [ + 0.521, + 0.098, + 0.905, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.26, + 0.905, + 0.289 + ], + "angle": 0, + "content": "Figure 8. Visual results compared with motion frames and overlapping." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.306, + 0.907, + 0.352 + ], + "angle": 0, + "content": "temporal audio attention helps learn effective control of micro-expressions, while the motion controller is critical for agility and expressiveness." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.356, + 0.907, + 0.583 + ], + "angle": 0, + "content": "Larger Audio Perception Bridges Inter-clip Stability. Initially, we ablate the time-aware position shift fusion (briefly termed Shift) and generate clips independently as illustrated in Figure 7. The fusion process is performed at each timestep, gradually inverting the random noise to continuous and smooth long range videos. Subsequently, we explore different long video inference strategy: motion frames, overlap, our shift. The number of motion frames and overlap frames was set to 8. As indicated in Table 4, our shift approach attains superior lip-sync and smoothness without necessitating any auxiliary computation. From the qualitative comparison demonstrated in Figure 8, our method successfully handle long term consistency, circumventing identity similarity decrease in motion frames and mixed blurry texture in overlapping results." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.61, + 0.634, + 0.626 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.64, + 0.907, + 0.79 + ], + "angle": 0, + "content": "In summary, our work present Sonic, an audio-driven portrait animation that focuses on the exploration of global audio perception to enhance the efficient generation of realistic lip synchronization, naturalness and temporal consistencies. The technical cornerstone mainly lies on context-enhanced audio learning and time-aware position shift fusion that work together to extend the intra-clip audio temporal to the global inter-clip audio perception. It significantly outperforms existing SOTA methods in video quality, motion diversity and naturalness." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Acknowledgements We thank Dongshan Li, Zixiang Zhou, Zhentao Yu, Hongmei Wang, Joeyu Wang, Chengfei Cai, Shiyu Tang, Tianxiang Zheng, Junshu Tang, Shuai Shao, Yubin Zeng, Junxin Huang, Yue Zhang, Chao Zhan, Zhengyan Tong, Zhaokang Chen, Bin Wu, Wenjiang Zhou, Yang Zhang, Junqi Cheng, Jiaxi Gu, and Fangyuan Zou for helpful discussion." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "200" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.171 + ], + "angle": 0, + "content": "[1] Alexei Baevski, Yuhao Zhou, Abdelrahman Mohamed, and Michael Auli. wav2vec 2.0: A framework for self-supervised learning of speech representations. Advances in neural information processing systems, 33:12449-12460, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.173, + 0.484, + 0.243 + ], + "angle": 0, + "content": "[2] Omer Bar-Tal, Hila Chefer, Omer Tov, Charles Herrmann, Roni Paiss, Shiran Zada, Ariel Ephrat, Junhwa Hur, Yuanzhen Li, Tomer Michaeli, et al. Lumiere: A spacetime diffusion model for video generation. arXiv preprint arXiv:2401.12945, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.245, + 0.483, + 0.314 + ], + "angle": 0, + "content": "[3] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.316, + 0.483, + 0.398 + ], + "angle": 0, + "content": "[4] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22563-22575, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.402, + 0.483, + 0.496 + ], + "angle": 0, + "content": "[5] Aras Bozkurt, Xiao Junhong, Sarah Lambert, Angelica Pazurek, Helen Crompton, Suzan Koseoglu, Robert Farrow, Melissa Bond, Chrissi Nerantzi, Sarah Honeychurch, et al. Speculative futures on chatgpt and generative artificial intelligence (ai): A collective reflection from the educational landscape. Asian Journal of Distance Education, 18(1):53-130, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.5, + 0.483, + 0.582 + ], + "angle": 0, + "content": "[6] Haoxin Chen, Yong Zhang, Xiaodong Cun, Menghan Xia, Xintao Wang, Chao Weng, and Ying Shan. Videocrafter2: Overcoming data limitations for high-quality video diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7310-7320, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.585, + 0.483, + 0.641 + ], + "angle": 0, + "content": "[7] Lele Chen, Zhiheng Li, Ross K Maddox, Zhiyao Duan, and Chenliang Xu. Lip movements generation at a glance. In Proceedings of the European conference on computer vision (ECCV), pages 520-535, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.643, + 0.483, + 0.698 + ], + "angle": 0, + "content": "[8] Zhiyuan Chen, Jiajiong Cao, Zhiquan Chen, Yuming Li, and Chenguang Ma. Echomimic: Lifelike audio-driven portrait animations through editable landmark conditions. arXiv preprint arXiv:2407.08136, 2024. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.701, + 0.483, + 0.77 + ], + "angle": 0, + "content": "[9] Joon Son Chung and Andrew Zisserman. Out of time: automated lip sync in the wild. In Computer Vision-ACCV 2016 Workshops: ACCV 2016 International Workshops, Taipei, Taiwan, November 20-24, 2016, Revised Selected Papers, Part II 13, pages 251-263. Springer, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.772, + 0.483, + 0.813 + ], + "angle": 0, + "content": "[10] Joon Son Chung, Arsha Nagrani, and Andrew Zisserman. Voxceleb2: Deep speaker recognition. arXiv preprint arXiv:1806.05622, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.816, + 0.483, + 0.871 + ], + "angle": 0, + "content": "[11] Enric Corona, Andrei Zanfir, Eduard Gabriel Bazavan, Nikos Kolotouros, Thiemo Alldieck, and Cristian Sminchisescu. Vlogger: Multimodal diffusion for embodied avatar synthesis. arXiv preprint arXiv:2403.08764, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.483, + 0.902 + ], + "angle": 0, + "content": "[12] Jiahao Cui, Hui Li, Yao Yao, Hao Zhu, Hanlin Shang, Kaihui Cheng, Hang Zhou, Siyu Zhu, and Jingdong Wang. Hallo2:" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.121 + ], + "angle": 0, + "content": "Long-duration and high-resolution audio-driven portrait image animation. arXiv preprint arXiv:2410.07718, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.123, + 0.905, + 0.205 + ], + "angle": 0, + "content": "[13] Yu Deng, Jiaolong Yang, Sicheng Xu, Dong Chen, Yunde Jia, and Xin Tong. Accurate 3d face reconstruction with weakly-supervised learning: From single image to image set. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 0–0, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.207, + 0.905, + 0.276 + ], + "angle": 0, + "content": "[14] Patrick Esser, Johnathan Chiu, Parmida Atighehchian, Jonathan Granskog, and Anastasis Germanidis. Structure and content-guided video synthesis with diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7346-7356, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.278, + 0.905, + 0.348 + ], + "angle": 0, + "content": "[15] Yuwei Guo, Ceyuan Yang, Anyi Rao, Zhengyang Liang, Yaohui Wang, Yu Qiao, Maneesh Agrawala, Dahua Lin, and Bo Dai. Animatediff: Animate your personalized text-to-image diffusion models without specific tuning. International Conference on Learning Representations, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.349, + 0.905, + 0.404 + ], + "angle": 0, + "content": "[16] Tianyu He, Junliang Guo, Runyi Yu, Yuchi Wang, Jialiang Zhu, Kaikai An, Leyi Li, Xu Tan, Chunyu Wang, Han Hu, et al. Gaia: Zero-shot talking avatar generation. arXiv preprint arXiv:2311.15230, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.406, + 0.905, + 0.46 + ], + "angle": 0, + "content": "[17] Yingqing He, Tianyu Yang, Yong Zhang, Ying Shan, and Qifeng Chen. Latent video diffusion models for high-fidelity long video generation. arXiv preprint arXiv:2211.13221, 2022.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.462, + 0.905, + 0.531 + ], + "angle": 0, + "content": "[18] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.533, + 0.905, + 0.588 + ], + "angle": 0, + "content": "[19] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. Advances in Neural Information Processing Systems, 35:8633-8646, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.59, + 0.905, + 0.645 + ], + "angle": 0, + "content": "[20] Wenyi Hong, Ming Ding, Wendi Zheng, Xinghan Liu, and Jie Tang. Cogvideo: Large-scale pretraining for text-to-video generation via transformers. arXiv preprint arXiv:2205.15868, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.647, + 0.905, + 0.731 + ], + "angle": 0, + "content": "[21] Ziqi Huang, Yinan He, Jiashuo Yu, Fan Zhang, Chenyang Si, Yuming Jiang, Yuanhan Zhang, Tianxing Wu, Qingyang Jin, Nattapol Chanpaisit, et al. Vbench: Comprehensive benchmark suite for video generative models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21807-21818, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.732, + 0.905, + 0.813 + ], + "angle": 0, + "content": "[22] Xiaozhong Ji, Chuming Lin, Zhonggan Ding, Ying Tai, Jian Yang, Junwei Zhu, Xiaobin Hu, Jiangning Zhang, Donghao Luo, and Chengjie Wang. Realtalk: Real-time and realistic audio-driven face generation with 3d facial prior-guided identity alignment network. arXiv preprint arXiv:2406.18284, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.816, + 0.905, + 0.872 + ], + "angle": 0, + "content": "[23] Jianwen Jiang, Chao Liang, Jiaqi Yang, Gaojie Lin, Tianyun Zhong, and Yanbo Zheng. Loopy: Taming audio-driven portrait avatar with long-term motion dependency. arXiv preprint arXiv:2409.02634, 2024. 2, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.902 + ], + "angle": 0, + "content": "[24] Esperanza Johnson, Ramón Hervás, Carlos Gutiérrez López de la Franca, Tania Mondéjar, Sergio F Ochoa, and Jesús" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "201" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.093, + 0.482, + 0.134 + ], + "angle": 0, + "content": "Favela. Assessing empathy and managing emotions through interactions with an affective avatar. Health informatics journal, 24(2):182-193, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.482, + 0.178 + ], + "angle": 0, + "content": "[25] Oytun Kal and Yavuz Samur. Educational virtual reality game design for film and animation. Encyclopedia of Computer Graphics and Games, pages 621-636, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.179, + 0.482, + 0.262 + ], + "angle": 0, + "content": "[26] Levon Khachatryan, Andranik Movsisyan, Vahram Tadevosyan, Roberto Henschel, Zhangyang Wang, Shant Navasardyan, and Humphrey Shi. Text2video-zero: Text-to-image diffusion models are zero-shot video generators. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15954-15964, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.264, + 0.482, + 0.319 + ], + "angle": 0, + "content": "[27] Tao Liu, Feilong Chen, Shuai Fan, Chenpeng Du, Qi Chen, Xie Chen, and Kai Yu. Anitalker: Animate vivid and diverse talking faces through identity-decoupled facial motion encoding. arXiv preprint arXiv:2405.03121, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.321, + 0.482, + 0.376 + ], + "angle": 0, + "content": "[28] Yifeng Ma, Shiwei Zhang, Jiayu Wang, Xiang Wang, Yingya Zhang, and Zhidong Deng. Dreamtalk: When expressive talking head generation meets diffusion probabilistic models. arXiv preprint arXiv:2312.09767, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.378, + 0.482, + 0.446 + ], + "angle": 0, + "content": "[29] KR Prajwal, Rudrabha Mukhopadhyay, Vinay P Namboodiri, and CV Jawahar. A lip sync expert is all you need for speech to lip generation in the wild. In Proceedings of the 28th ACM international conference on multimedia, pages 484-492, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.449, + 0.482, + 0.517 + ], + "angle": 0, + "content": "[30] Alec Radford, Jong Wook Kim, Tao Xu, Greg Brockman, Christine McLeavey, and Ilya Sutskever. Robust speech recognition via large-scale weak supervision. In International conference on machine learning, pages 28492-28518. PMLR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.519, + 0.482, + 0.588 + ], + "angle": 0, + "content": "[31] Imogen C Rehm, Emily Foenander, Klaire Wallace, JoAnne M Abbott, Michael Kyrios, and Neil Thomas. What role can avatars play in e-mental health interventions? exploring new models of client-therapist interaction. Frontiers in Psychiatry, 7:186, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.59, + 0.482, + 0.659 + ], + "angle": 0, + "content": "[32] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.661, + 0.482, + 0.731 + ], + "angle": 0, + "content": "[33] Shuai Shen, Wenliang Zhao, Zibin Meng, Wanhua Li, Zheng Zhu, Jie Zhou, and Jiwen Lu. Difftalk: Crafting diffusion models for generalized audio-driven portraits animation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1982-1991, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.732, + 0.482, + 0.8 + ], + "angle": 0, + "content": "[34] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.803, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[35] Xusen Sun, Longhao Zhang, Hao Zhu, Peng Zhang, Bang Zhang, Xinya Ji, Kangneng Zhou, Daiheng Gao, Liefeng Bo, and Xun Cao. Vividtalk: One-shot audio-driven talking head generation based on 3d hybrid prior. arXiv preprint arXiv:2312.01841, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[36] Supasorn Suwajanakorn, Steven M Seitz, and Ira Kemelmacher-Shlizerman. Synthesizing obama: learn" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "ing lip sync from audio. ACM Transactions on Graphics (ToG), 36(4):1-13, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.178 + ], + "angle": 0, + "content": "[37] Linrui Tian, Qi Wang, Bang Zhang, and Liefeng Bo. Emo: Emote portrait alive-generating expressive portrait videos with audio2video diffusion model under weak conditions. arXiv preprint arXiv:2402.17485, 2024. 2, 3, 4, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.905, + 0.22 + ], + "angle": 0, + "content": "[38] Thomas Unterthiner, Sjoerd van Steenkiste, Karol Kurach, Raphaël Marinier, Marcin Michalski, and Sylvain Gelly. Fvd: A new metric for video generation. 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.222, + 0.905, + 0.277 + ], + "angle": 0, + "content": "[39] Vikram Voleti, Alexia Jolicoeur-Martineau, and Chris Pal. Mcvd-masked conditional video diffusion for prediction, generation, and interpolation. Advances in neural information processing systems, 35:23371-23385, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.279, + 0.905, + 0.348 + ], + "angle": 0, + "content": "[40] Cong Wang, Kuan Tian, Jun Zhang, Yonghang Guan, Feng Luo, Fei Shen, Zhiwei Jiang, Qing Gu, Xiao Han, and Wei Yang. V-express: Conditional dropout for progressive training of portrait video generation. arXiv preprint arXiv:2406.02511, 2024. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.35, + 0.905, + 0.419 + ], + "angle": 0, + "content": "[41] Xiang Wang, Hangjie Yuan, Shiwei Zhang, Dayou Chen, Jiuniu Wang, Yingya Zhang, Yujun Shen, Deli Zhao, and Jingren Zhou. Videocomposer: Compositional video synthesis with motion controllability. Advances in Neural Information Processing Systems, 36, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.421, + 0.905, + 0.462 + ], + "angle": 0, + "content": "[42] Huawei Wei, Zejun Yang, and Zhisheng Wang. Aniportrait: Audio-driven synthesis of photorealistic portrait animation. arXiv preprint arXiv:2403.17694, 2024. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.464, + 0.907, + 0.533 + ], + "angle": 0, + "content": "[43] Liangbin Xie, Xintao Wang, Honglun Zhang, Chao Dong, and Ying Shan. Vfhq: A high-quality dataset and benchmark for video face super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 657-666, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.535, + 0.905, + 0.604 + ], + "angle": 0, + "content": "[44] Mingwang Xu, Hui Li, Qingkun Su, Hanlin Shang, Liwei Zhang, Ce Liu, Jingdong Wang, Luc Van Gool, Yao Yao, and Siyu Zhu. Hallo: Hierarchical audio-driven visual synthesis for portrait image animation. arXiv preprint arXiv:2406.08801, 2024. 2, 3, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.606, + 0.905, + 0.673 + ], + "angle": 0, + "content": "[45] Sicheng Xu, Guojun Chen, Yu-Xiao Guo, Jiaolong Yang, Chong Li, Zhenyu Zang, Yizhong Zhang, Xin Tong, and Baining Guo. Vasa-1: Lifelike audio-driven talking faces generated in real time. arXiv preprint arXiv:2404.10667, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.676, + 0.905, + 0.759 + ], + "angle": 0, + "content": "[46] Fei Yin, Yong Zhang, Xiaodong Cun, Mingdeng Cao, Yanbo Fan, Xuan Wang, Qingyan Bai, Baoyuan Wu, Jue Wang, and Yujiu Yang. Styleheat: One-shot high-resolution editable talking face generation via pre-trained stylegan. In European conference on computer vision, pages 85-101. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.761, + 0.905, + 0.829 + ], + "angle": 0, + "content": "[47] Shengming Yin, Chenfei Wu, Huan Yang, Jianfeng Wang, Xiaodong Wang, Minheng Ni, Zhengyuan Yang, Linjie Li, Shuguang Liu, Fan Yang, et al. Nuwa-xl: Diffusion over diffusion for extremely long video generation. arXiv preprint arXiv:2303.12346, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.832, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[48] Jianhui Yu, Hao Zhu, Liming Jiang, Chen Change Loy, Weidong Cai, and Wayne Wu. Celebv-text: A large-scale facial text-video dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14805-14814, 2023. 6" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "202" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.092, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[49] Rui Zhang, Yixiao Fang, Zhengnan Lu, Pei Cheng, Zebiao Huang, and Bin Fu. Linguualinker: Audio-driven portraits animation with implicit facial control enhancement. arXiv preprint arXiv:2407.18595, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.091, + 0.149, + 0.483, + 0.232 + ], + "angle": 0, + "content": "[50] Wenxuan Zhang, Xiaodong Cun, Xuan Wang, Yong Zhang, Xi Shen, Yu Guo, Ying Shan, and Fei Wang. Sadtalker: Learning realistic 3d motion coefficients for stylized audio-driven single image talking face animation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8652-8661, 2023. 2, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.234, + 0.484, + 0.302 + ], + "angle": 0, + "content": "[51] Zhimeng Zhang, Lincheng Li, Yu Ding, and Changjie Fan. Flow-guided one-shot talking face generation with a high-resolution audio-visual dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3661–3670, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.304, + 0.483, + 0.372 + ], + "angle": 0, + "content": "[52] Hao Zhu, Wayne Wu, Wentao Zhu, Liming Jiang, Siwei Tang, Li Zhang, Ziwei Liu, and Chen Change Loy. Celebvyhq: A large-scale video facial attributes dataset. In European conference on computer vision, pages 650-667. Springer, 2022. 6" + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.092, + 0.484, + 0.372 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.511, + 0.956 + ], + "angle": 0, + "content": "203" + } + ] +] \ No newline at end of file diff --git a/2025/Sonic_ Shifting Focus to Global Audio Perception in Portrait Animation/108dd369-6bdf-4c82-9e76-1c8b1e3be7c1_origin.pdf b/2025/Sonic_ Shifting Focus to Global Audio Perception in Portrait Animation/108dd369-6bdf-4c82-9e76-1c8b1e3be7c1_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..eaf380bfe55d84f6ee11313f1fb67dc16e621346 --- /dev/null +++ b/2025/Sonic_ Shifting Focus to Global Audio Perception in Portrait Animation/108dd369-6bdf-4c82-9e76-1c8b1e3be7c1_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4a0e9efc2315cea253bf45df9cde92d618c4e92a1605e2bd36227efa9cf3a3f8 +size 7930494 diff --git a/2025/Sonic_ Shifting Focus to Global Audio Perception in Portrait Animation/full.md b/2025/Sonic_ Shifting Focus to Global Audio Perception in Portrait Animation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..c85e32e3b6972c6c89e797239063d18c2ed6dce4 --- /dev/null +++ b/2025/Sonic_ Shifting Focus to Global Audio Perception in Portrait Animation/full.md @@ -0,0 +1,271 @@ +# Sonic: Shifting Focus to Global Audio Perception in Portrait Animation + +Xiaozhong Ji $^{1*}$ Xiaobin Hu $^{1*}$ Zhihong Xu $^{2}$ Junwei Zhu $^{1\dagger}$ Chuming Lin $^{1}$ Qingdong He $^{1}$ Jiangning Zhang $^{1}$ Donghao Luo $^{1}$ Yi Chen $^{1}$ Qin Lin $^{1}$ Qinglin Lu $^{1}$ Chengjie Wang $^{1}$ $^{1}$ Tencent $^{2}$ Zhejiang University + +![](images/4cdac659236a27f42818a2d5d96f7d65ea4e189ed7cb8f635c69cd635001d723.jpg) + +![](images/258097e14bae0c985142d9d7ac64a532dbf2f780721f6225b9a92e1b8ecba330.jpg) + +![](images/1457ec753d8173dbf7a05a442559e8dc0dbe82f94280f86c25afafe79bdae14f.jpg) +Figure 1. Sonic excels in producing vivid portrait animation videos given a reference image and an audio clip. Beyond its fundamental lip-syncing capabilities, Sonic demonstrates proficiency in creating a diverse spectrum of facial expressions and adaptable head movements. Notably, when dealing with extended videos, Sonic can yield stable and seamless outcomes in a parallel fashion, all while maintaining an unified paradigm focusing on global audio perception. + +![](images/b5e2e39a45c721c1e21dd6850c86200acd2cfbcb49d4f9634c16dc26d769eea5.jpg) + +# Abstract + +The study of talking face generation mainly explores the intricacies of synchronizing facial movements and crafting visually appealing, temporally-coherent animations. However, due to the limited exploration of global audio perception, current approaches predominantly employ auxiliary visual and spatial knowledge to stabilize the movements, which often results in the deterioration of the naturalness and temporal inconsistencies. Considering the essence of audio-driven animation, the audio signal serves as the ideal and unique priors to adjust facial expressions and lip movements, without resorting to interference of any visual signals. Based on this motivation, we propose a novel paradigm, dubbed as Sonic, to shift focus on the exploration of global audio perception. To effectively leverage global audio knowledge, we disentangle it into intra- and inter-clip audio perception and collaborate with both aspects to enhance overall perception. For the intra-clip audio perception, 1). Context-enhanced audio learning, in + +which long-range intra-clip temporal audio knowledge is extracted to provide facial expression and lip motion priors implicitly expressed as the tone and speed of speech. 2). Motion-decoupled controller, in which the motion of the head and expression movement are disentangled and independently controlled by intra-audio clips. Most importantly, for inter-clip audio perception, as a bridge to connect the intra-clips to achieve the global perception, Time-aware position shift fusion, in which the global inter-clip audio information is considered and fused for long-audio inference via through consecutively time-aware shifted windows. Extensive experiments demonstrate that the novel audio-driven paradigm outperform existing SOTA methodologies in terms of video quality, temporally consistency, lip synchronization precision, and motion diversity. + +# 1. Introduction + +Talking Face Animation is a technique aimed at creating a lifelike representation of a speaking person by animating a still image in sync with speech audio. This technology holds great potential in enhancing the realism of virtual re + +ality [5], improving the efficiency of film and gaming making [25], facilitating accessibility for individuals with communication impairments [24], and offering therapeutic support and social interaction in healthcare [31]. Prominent progress in this domain is demonstrated by notable works based on Stable Diffusion [32], which project the parametric or implicit representations of lip movements [29, 50], facial expressions [16, 28, 33], and head movement [35] in the latent space [11, 27, 37, 45] to generate high-quality video with a high level of realism and liveliness. + +However, precise audio control and temporal coherence have not yet been effectively achieved. Current approaches [8, 23, 37, 40, 42, 44, 49, 50] predominantly addressed these two issues in a completely segregated manner, overlooking the holistic coordination between audio and vision. For audio control, audio features were segmented according to timestamps and matched to each visual frame. By applying cross-attention within a spatial latent space, each frame was restricted to the nearby audio information, which hindered the transformation into an optimal motion representation. To achieve temporal coherence, self-attention along the temporal dimension was applied to smooth visual features within a clip, while overlapping or motion frames were used to enhance coherence between different clips. However, these stability strategies have limited temporal receptive fields, which can reduce the richness of motion. Additionally, they do not take audio information into account, potentially disrupting the synchronization between audio and visual elements. + +To address these issues, we propose a novel paradigm, dubbed as Sonic, to extent the range of receptive field to the global level guided by the audio priors. To achieve global audio perception, we decompose the talking head video generation task into two aspects: intra-clip (i.e., lip sync, expression alignment, and head movement) and inter-clip (i.e., motion rhythm and smoothness, long video stability). Sonic aims to exploit global audio perception rather than the motion frames and other visual motion (i.e., referring to the face locator and landmark prior adopted by existing methods, which hinders the ability to perceive information directly from audio). Performing as long-range and global signal, audio shows tone, rhythm and speed of speech which implicitly provides expression and head movement motion priors. However, we have to face such a tricky challenge that, as a weak correlation cross-modality signal without the guidance of any visual motion (e.g., motion frames), current audio learning paradigm leads to the temporal jitter and unsatisfactory pattern and fails to well-align the audio and motion. To this end, we first customize efficient global audio perception learning mechanism. Specifically, context-enhanced audio learning is proposed as long-range audio learning module to extract the intra-clip audio temporal knowledge from the input audio clip. Then such intra + +clip audio temporal information is mapped as a temporal-embeddings for the latter fusion by audio-temporal cross-attention. The temporal-embeddings projection operation can effectively reduce the computational burden and capture temporal priors within current audio clip. Although context-enhanced audio learning directs the temporal movements correlated with audio, such as lip movements synchronized with speech and facial expressions related to the content, there still exists the absence of habitual head movements. Secondly, we design motion-decoupled controller to disentangle the habitual head and expression movement and support the independent control via two explicit parameters learned from the current intra-clip audio. To enrich the functionality and playability of animation, we also add an plug to allow users to define some exaggerated movements. Most importantly, for inter-clip audio perception, time-aware position shift fusion enlarges the intra-clip audio perception to the global inter-clip audio perception via time-aware shifted windows consecutively bridging the preceding clip, providing the global inter-clip connection that significantly enhance modeling temporal power, as demonstrated in Figure 1. It is worth noting that our time-aware position shift fusion requires no extra training cost compared with the motion frames in most recent works, such as EMO [37], Loopy [23] and also no additional inference time caused by overlapping frames. Our contributions are summarized as: + +- We propose a novel unified paradigm without the guidance of visual motion, dubbed as Sonic, to focus on the exploration of global audio perception. +- For intra-clip audio perception, context-enhanced audio learning and motion-decoupled controller excavate the intra-clip audio temporal knowledge to direct the temporal movements. +- For inter-clip audio perception, we propose time-aware position shift fusion to enhance the global inter-clip audio perception by extending the intra-clip audio perception through consecutively time-aware shifted windows. +- Extensive experiments demonstrate that the novel unified global audio perception paradigm outperforms existing SOTA methods in terms of video quality, temporal consistency, lip synchronization, and motion diversity. + +# 2. Related Works + +# 2.1. Diffusion-based Video Generation + +The remarkable achievements in diffusion models in image generation have sparked widespread research interest for video generation. Early endeavors in video generation [14, 19, 20, 26, 34] have primarily concentrated on devising image-based diffusion models with temporal module to effectively capture the dynamic characteristics inherent in video sequences. To leverage the powerful abil + +ity of pretrained image diffusion models, Video LDM [4] incorporates temporal layers that learn to align images in a temporally consistent manner. AnimateDiff [15] provides a motion-specific module by multi-stage progressive training manner, and the motion module can be seamlessly integrated into existing text-to-image architectures without any further modifications. VideoComposer [41] and VideoCrafter [6] further explore the synthesis of image-to-video generation via textual and visual features combinations. Despite these notable advancements, the field still struggles to overcome the challenge of generating temporal coherent and high-quality videos. As a powerful motion representation trained on a large well-curated high-quality dataset, Stable Video Diffusion [3] shows the remarkable temporal-consistency and the high-quality visual characters for downstream tasks. There exists a large gap between the audio-driven portrait animation and general video generation due to the fact that audio-driven task aims to use the weak cross-modality audio signal to precisely control motion movements (e.g., lip and expression movement). + +# 2.2. Audio-driven Talking Face Generation + +The generation of talking face videos from audio inputs has been a longstanding task. In the early stages, the emphasis was primarily on synthesizing lip movements alone, accomplished by directly mapping audio signals to lip movements while keeping other facial attributes unchanged [7, 22, 29, 36, 46]. In more recent efforts, the scope has been broadened to encompass a wider range of facial expressions and head movements, which are derived from audio inputs and a single reference image. These methods can be broadly classified into with or without intermediate representation. For instance, by employing 3D coefficients representation (e.g., Sadtalker [50], Dreamtalk [28], Ani-Portrait [42], VASA-1 [45]), the video generation process can be informed through the direct or indirect visible control signal in the aspect of head movement and expression. Nevertheless, a recurring challenge faced by these techniques is the limited capacity of the 3D mesh to capture intricate details, consequently constraining the overall dynamism and realism. In contrast, the methods that omit intermediate representation exhibit superior naturalness and consistent preservation of identity with the original image. EMO [37], Hallo [44], and Loopy [23] employ a direct audio-to-video synthesis approach to ensure a high level of realism and naturalness, eliminating the need for intermediate 3D representation. However, these approaches primarily rely on additional conditions associated with spatial motion or motion frames to maintain temporal consistency. In this regard, they overlook the crucial role of audio in audio-driven animation, consequently diminishing both expressiveness and audio-motion synchronization. + +# 2.3. Long Video Inference + +Significant research efforts have been devoted to extending the duration of generated videos, aiming to expand their practical applications. Autoregressively predicting successive frames [17, 39] and hierarchical coarse-to-fine methods [47] have emerged to generate long videos. Then Lumiere [2] divides the video into overlapping temporal segments, denoises each segment independently, and finally fuse these segments in an optimization algorithm. However there is no available high-efficiency method for audio-condition global perception. As one step towards achieving stronger capabilities, our Sonic proposes time-aware position shift fusion strategy progressively denoise intra-clip audio to establish a global inter-clip connection through the use of time-aware shifted windows that bridge the preceding clip along the timesteps axis. This inter-clip fusion with global receptive field needs no additional training cost and introduce no extra inference time without overlapping frames. + +# 3. Methodology + +Overall framework. Given a single portrait reference image and an input audio, our one-stage framework can generate a portrait video well-synchronized with the driven audio. Our proposed Sonic framework aims to focus on the exploration of global audio perception, primarily consisting of context-enhanced audio learning, motion-decoupled controller, and time-aware position shift fusion module, as illustrated in Figure 2. + +Technical contribution. To achieve global audio perception, we decompose the talking head video generation task into two aspects: intra-clip (i.e., lip sync, expression alignment, and head movement) and inter-clip (i.e., motion rhythm and smoothness, long video stability). For intra-clip, we enhance the visual realism that matches audio signal through the context-enhanced audio learning and the motion controller. For inter-clip, we improve the stability using the time-aware position shift fusion strategy. + +# 3.1. Context-enhanced Audio Learning + +For audio-talking face animation, the audio is the ideally unique and global signal to adjust the face expression and movement without interference from other visual signals. Thus, the motivation behind context-enhanced audio learning is to guide the generation of spatial and temporal frames solely based on audio signals, rather than motion frames or other visual signals, to achieve the high-realism generation. Specifically, we adopt Whisper-Tiny [30] as the pretrained audio feature extraction model, which is more lightweight than Wav2Vec [1] used by most existing methods [8, 37, 44]. For each audio frame, the features from the last layers of five stages are concatenated together for multiscale understanding. For each video frame, a duration of + +![](images/a1775a9d25b07070e02f41b8fc2df232ce1ecec833567882c7ae742a6b6389ed.jpg) +Figure 2. Framework of our approach. Sonic processes each clip of the long audio in parallel, shifting to a new context at each time step to progressively fuse inter-clip latent features across global audio perception. In Sonic, we enhance intra-clip temporal audio context learning and decouple motion to improve dynamics. + +![](images/5339a31b068a57e4d1d9409ccc5d5a80d511b2d1e6bb19c7de2f4d2a66916b02.jpg) + +![](images/f484da2b775929682cd4a9d4bf8ec8589faca09ae2f7863e73281d2d607cac78.jpg) + +0.2s audio feature is used frame-wise to give rich context. We use three linear layers to project audio feature to match the dimension of cross attention. The transformed audio embedding is denoted as $c_{a} \in \mathcal{R}^{b \times f \times d \times c}$ , where $b$ denotes batch size, $f$ is audio length, $d$ is num of context tokens, and $c$ is hidden size of the cross-attention. To focus on talking face area, the core latent features is restricted with a mask $M$ determined by a joint bounding boxes of faces. The audio signal can act as the spatial knowledge provider being injected into the spatial cross-attention layers as follows: + +$$ +z _ {s} ^ {\prime} = z _ {s} + \operatorname {C r o s s A t t n} \left(Q \left(z _ {s}\right), K \left(c _ {a}\right), V \left(c _ {a}\right)\right) \cdot M, \tag {1} +$$ + +where $z_{s}$ is spatial latent features of self-attention layer, and $z_{s}^{\prime}$ is the adjusted spatial features guided by audio signals in spatial-aware level. + +Furthermore, audio is a long-range and global signal that provides expression and head movement motion priors implicitly expressed through the tone and speed of speech. To excavate the temporal information from the long-range audio knowledge, we customize a temporal-audio module for precise motion control. The temporal module in existing methods [37, 50] followsAnimateDiff, which performs only self-attention across the temporal dimension without audio aids. Some other talking-face [23, 44] works extract the motion priors from the visual knowledge (e.g., motion frames), which deviates from the essence of audio-driven task that relies solely on audio to drive movement. Thus we propose a temporal audio cross-attention module, introducing audio embedding to guide temporal alignment. + +The preceding latent from temporal self-attention layers can be reshaped as $z_{t} \in \mathcal{R}^{(b \times h \times w) \times f \times c}$ . To reduce the computation burden, the audio feature $c_{a}$ is average-pooled along temporal dimension as the shape $\mathcal{R}^{(b \times 1) \times d \times c}$ , then + +repeated and reshaped as $c_{ta} \in \mathcal{R}^{(b \times h \times w) \times d \times c}$ . The projected audio temporal-embeddings are injected into the denoising U-Net via a temporal cross-attention blocks as: + +$$ +z _ {t} ^ {\prime} = z _ {t} + \operatorname {C r o s s A t t n} (Q (z _ {t}), K (c _ {t a}), V (c _ {t a})). \tag {2} +$$ + +The spatial and temporal audio attention layers are cascaded in multiple down-stages and up-stages to enable natural appearance and smooth temporal movement. + +# 3.2. Motion-decoupled Controller + +Context-enhanced Audio Learning governs the temporal movements strongly correlated with audio, such as lip movements synchronized with speech and facial expressions related to the content. On the other hand, our motion-decoupled controller performs a complementary role by directing the temporal movements weakly associated with audio, such as habitual head movements and random changes in perspective. This controller can disentangle motion into explicit head and expression movement amplitudes, thereby enhancing both functionality, playability, and interactivity. Thus, we introduce independent augmented motion-bucket parameters which effectively influence motion patterns of head movements and expression strength. During the training phase, the translation motion-bucket $m_t$ is computed as the variance of bounding boxes of video clip, while the expression motion-bucket $m_e$ is variance of relative landmarks. Both buckets are integers in the range of [0, 128]. Through position encoding and linear projection, they are added into ResNet blocks as follows: + +$$ +e m b = W \left[ \operatorname {P e} \left(m _ {t}\right), \operatorname {P e} \left(m _ {e}\right) \right], \tag {3} +$$ + +where $emb$ is the embedding, $\mathrm{Pe}$ is the position encoding, and $W$ represents the linear projection weights. + +For practice application, motion-decoupled controller also supports the automatically adjustable parameters learning from the audio and reference image, and deeply explore the correlation between audio and portrait reference without requiring the parameter setting from users. To simplify multiple parameters and maintain adjustability, the predicted motion buckets are multiplying by a scale $\beta$ (0.5 for mild dynamic, 1.0 for moderate dynamic, and 2.0 for intense dynamic). Specifically, we utilize a audio-to-bucket manner to adaptively predict the motion-buckets relying on the audio and reference image. In inference phase, the motion buckets are predicted and rescaled as: + +$$ +m _ {t} ^ {\prime}, m _ {e} ^ {\prime} = \beta \times \mathrm {E} _ {\mathrm {b}} \left(c _ {a}, R _ {i m g}\right). \tag {4} +$$ + +where $m_t'$ , $m_e'$ are the estimated translation and expression motion-bucket achieved by the audio $c_a$ and single reference image CLIP embedding $R_{img}$ , and $E_b$ is the audio-to-bucket projection consisting of there linear layers with the ReLU as activation function. + +# 3.3. Time-aware Position Shift Fusion + +![](images/4967c461a53e2a8190cbf3f9b40249587298b2997e31cc5da103aab701f3a0b3.jpg) +Figure 3. Illustration of the proposed time-aware position-shift fusion. The model processes each clip non-overlapping. In next timestep, the model starts from a new position determined by the offset, thereby integrating long-range context. Specifically, the tail latents are filled cyclically from the head. + +Recent methods [23, 37, 44] use motion frames or overlapping latent during coherent denoising, which suffer from a limited receptive field and increased the training or inference computational complexity for audio-conditioned video generation problem. In contrast, we propose a bridge, the time-aware position shift fusion, to extent the intra-clip audio perception to the global inter-clip audio-perception via consecutively time-aware shifted windows. The simple yet effective fusion strategy, without increasing extra inference or training time costs, can effectively address jitter and abrupt transition issues. Audio serves as the input condition for the talking face, where longer audio prompts the generation of longer videos. Based on this principle, we tailor time-aware position-shift fusion mechanism along the timestep axis where the latent sliding window is progressively shifted by a certain length at different timesteps while also preserving the frame correspondence relationship between audio and video signal. + +Specifically, as described in Figure 3 and Algorithm 1, our time-aware position-shift fusion consists of two nested loops. The outer loop is inverse diffusion process, while the inner loop is a sliding window process that model predicts for the full-length audio conditions. At each timestep, the model takes a clip of audio input to predict the corresponding latent with a start offset to smooth the segment from the last timestep. The video diffusion model naturally bridges the context, generating continuous movement following the audio prompts. In detail, we set a cumulative shift offset $\alpha_{\Sigma}$ , accumulating $\alpha$ at each timestep to ensure that the model starts sliding window from different positions. The offset $\alpha$ is experimentally set to a small integer, such as 3, 7; we will report analysis on this parameters in experiments part. Note that the relative positioning between the latents and the audio prompts has not been altered. The only modification occurs in their recombination during the input phase to the model. In a special case, at the end of the long sequence, the shifted position index may exceed the sequence length. To address this issue, we adopt a circular padding strategy, padding the start audio prompt and latents to the end. + +Algorithm 1 Time-aware Position-shift Fusion +Require: Audio embedding $c_{a}^{[0,l]}$ with length $l$ , denoising steps $T$ , initial noisy latent $z_{T}^{[0,l]}$ , pretrained Sonic model $\mathrm{ST}(\cdot)$ for sequence length $f$ , position-shift offset $\alpha < f < l$ . +Ensure: Denoised latent $z_{0}^{[0,l]}$ . +1: Initialize accumulated shift offset $\alpha_{\Sigma} = 0$ . +2: for $t = T, \dots, 1$ do +3: // Denoising loop +4: Initialize start point $s = \alpha_{\Sigma}$ , end $e = s + f$ , processed length $n = 0$ . // start from new position for each timestep. +5: while $n < l$ do +6: // Sequence loop +7: $z_{t-1}^{[s,e]} = \mathrm{ST}(z_{t}^{[s,e]}, c_{a}^{[s,e]}, t)$ +8: $s \gets s + f, e \gets e + f, n \gets n + f$ . // Move to next clip non-overlapping +9: if $s > l$ or $e > l$ then +10: $s \gets s\% l, e \gets e\% l$ . // Padding circularly +11: end if +12: end while +13: $\alpha_{\Sigma} \gets \alpha_{\Sigma} + \alpha$ . // Accumulate shift offset +14: end for +15: return Denoised latent $z_{0}^{[0,l]}$ . + +Computation cost. We analyze the computation efficiency of our time-aware position shift fusion compared to motion frames and overlapping. Denote the FLOPs of a single forward pass of the video diffusion model as $\Omega$ , the number of frames to generate as $n \times f$ , and the denoising steps as $T$ . The accumulated FLOPs of Sonic is: $\Omega(Shift) = \Omega \times T \times n$ . For overlapping, the overlapped latents (overlap $o$ ) are calculated twice, $\Omega(Overlap) = \Omega \times T \times \left(n + \frac{o}{f}\right)$ . For motion frames, the FLOPs of a single forward of Ref- + +Table 1. Quantitative comparisons with the state-of-the-arts on the HDTF and CelebV-HQ test set. * indicates GAN-based methods and others are Diffusion-based methods. The best results are bold, and the second are underlined. + +
DatasetMethodFID↓FVD↓Sync-C↑Sync-D↓E-FID↓F-SIM↑Smooth↑Runtime↓
HDTFSadTalker* (2023) [50]61.672397.1141.75510.6952.4820.92870.99613.75
Aniportrait (2024) [42]36.965471.4521.09512.4613.1610.95080.992144.03
V-Express (2024) [40]47.396758.0231.25612.3942.6340.90930.996839.04
Hallo (2024) [44]30.176347.3584.0609.5511.7920.95550.994174.65
Hallo2 (2024) [12]38.673328.5404.1369.4652.2030.96060.994245.75
EchoMimic (2024) [8]33.207384.3042.51410.7431.4860.95270.99345.45
Sonic (Ours)29.104301.1734.1979.3711.7450.95950.997017.04
CelebV-HQSadTalker* (2024) [50]57.574841.9621.97810.9152.2520.94340.99593.75
Aniportrait (2024) [42]53.746590.3730.99612.0843.2960.95220.991144.03
V-Express (2024) [40]65.400889.9850.80913.2552.7130.90190.996539.04
Hallo (2024) [44]47.403488.4992.68010.2922.2730.96070.994274.65
Hallo2 (2024) [12]52.396481.3362.63810.3432.8190.95870.994245.75
EchoMimic (2024) [8]48.267596.8701.94910.7542.1360.96120.99325.45
Sonic (Ours)43.137483.1082.68910.1941.7830.96240.997217.04
+ +![](images/07353588234a301c7bb4c2ad734f4be157930330510d866ef8192f839e350115.jpg) +Figure 4. Qualitative comparisons with State-of-the-Art talking head generation methods. Due to image does not reflect important sync, naturalness and stability, the full video comparison will be included in supplementary materials as well as comparison with demos from other non-open source works. + +![](images/67196a5f3134c1acae4602dea6d68ca40d7b1b00322cb8a9c55662c610341211.jpg) + +erence Net is $\omega_{r}$ , the number of motion frames is $o$ and the Flops of motion modules computing $f$ frames is $\omega_{m}$ , $\Omega (MF) = \Omega \times T\times n + \omega_{r}\times o\times n + \omega_{m}\times T\times n\times \frac{2of + o^{2}}{f^{2}}$ . Overall, the proposed time-aware position shift fusion significantly reduces the computation complexity. + +# 4. Experiments + +# 4.1. Experiments Setup + +Implementation Details. The Sonic training is in a single stage, with initialization of the spatial module and temporal modules from stable-video-diffusion-xt-1-1[3]. To separately enable different conditions in training, we manipulate + +the data such that $5\%$ of it drops audio, another $5\%$ drops image, and $5\%$ drops both condition. In inference phase, multi-condition CFG is performed. The CFG of the reference image $r_i$ is set to 2.0, and the CFG of the audio $r_a$ is set to 7.5. In our full model, the dynamic scale $\beta$ is set to 1.0 and the shift offset $\alpha$ is set to 7. + +Datasets and Evaluation Metrics. For the training set, we created a diverse dataset by aggregating data from open-source datasets, including VFHQ [43], CelebV-Text [48], VoxCeleb2 [10]. For the test set, we conducted experiments on two widely used datasets, HDTF [52] and CelebV-HQ [51]. We randomly selected 50 four-second clips from CelebV-HQ and 100 from HDTF. We utilize seven met + +![](images/760a4446925b36607c2b04ee26f8c7dbc5ecac4b05fc2dfcdcc9805be274a672.jpg) +Figure 5. Qualitative results under different styles of portrait images and various types of audio inputs. The images and audios were collected from recent works and the Internet. The upper section presents results for audio input ranging from 20 seconds to 2 minutes in duration, while the lower section shows results for longer audio inputs, up to 10 minutes. Our Sonic demonstrates versatility across various portrait styles and maintains vividness over extended durations. + +rics to evaluate. The Fréchet Inception Distance (FID) [18] and Fréchet Video Distance (FVD) [38] metrics measure the quality of the generated data. Expression-FID (E-FID) [37] uses face reconstruction method [13] to extract features and calculate the FID between generated and ground truth frames, while F-SIM measures facial similarity. SyncNet (Sync-C&Sync-D) [9] and VBench's smooth metric [21] assess lip synchronization and motion fluidity. In ablation experiments, we employ VBench's dynamic degree metric to evaluate amplitude of head movements. + +# 4.2. Results and Analysis + +Quantitative Comparison. In Table 1, our Sonic are significantly superior than other recent 6 methods in terms of FID, Sync-confidence and Smoothness evaluated on the CelebV-HQ and HDTF datasets. On the HDTF dataset, our Sonic achieve the lowest FID and FVD scores, $8\%$ lower than the second-best, indicating superior generation quality and alignment with real data. For the CelebV-HQ dataset, which features more complex characters and backgrounds, our method maintains superior performance across most metrics and competitive FVD. We notably excel in E-FID, $16.5\%$ lower than the second-best, which demonstrates our method has superior expression diversity. EMO [37] as a representative work that tackles talking head generation task, remains closed source, and not available for fair quantitative comparison. We provide video results using input from their online demos for a qualitative comparison. + +Qualitative Comparison. Figure 4 provides a visualization comparison on open datasets. Analyzing the results, Previous approaches relied on motion frames, limiting the rich + +ness of generated expressions and the dynamism of head movements. In contrast, our proposed framework incorporates global audio perception, enhancing the integration of audio and temporal variables while providing a deeper understanding of the audio input. This enables our method to generate a broader range of expressions that align with the audio and facilitate more natural head movements. + +Visualization Results in Complex Scenarios. We further explore the generation performance of our method under different input conditions. In terms of images, we employ three styles: real human, animation, and 3D. Each character portrait is paired with various types of audio relevant to its theme, such as singing, speech, and rap. As shown in Figure 5, our method successfully generates a wide array of expressions and diverse movements across various categories of input conditions, demonstrating strong robustness. + +Table 2. User study comparison on open dataset. + +
Method / MetricLip syncMotion diversityID consistencyVideo Smoothness
Aniportrait1.421.623.112.09
SadTalker1.982.342.952.95
Echomimic2.772.653.482.71
Hallo23.152.373.342.94
Sonic(Ours)4.58 (45%↑)4.55 (72%↑)4.29 (23%↑)4.66 (58%↑)
+ +User Studies We conducted subjective evaluation on open dataset to assess the comparative methods along four key dimensions: Lip sync, motion diversity, identity similarity, and video smoothness. Total 40 participants gives scores from 1 to 5 for results of five comparative methods. In Table 2, our Sonic outperforms the others in all four dimensions, especially showing a significant $72\%$ improvement in motion diversity. The video results for user study will be included in the supplementary materials. + +![](images/9f65a05c834be73952ef0238eac9444ab11f66f211836806959bc12ea4be9bc7.jpg) +Figure 6. Visual results for ablation on temporal audio attention and motion controller. + +![](images/8887c8f1533892c2fd1904150b799428ccb6c80fc1d53ec83dde211bc79b5735.jpg) +Figure 7. Demonstration on effect of time-aware position shift fusion. + +Table 3. Quantitative results of ablations on CelebV-HQ along with our results obtained using different dynamic scales. In the Sonic model, the default setting for $\beta$ is 1.0, i.e. Ours-Moderate. + +
Method / MetricSync-C↑Sync-D↓E-FID↓Smooth↑Dynamic↑
w/o temporal-audio2.61010.3102.3110.99690.62
w/o motion-controller2.19310.7313.2310.99690.22
w/o shift-fusion2.50910.3792.2020.99690.76
Ours-Mild (β = 0.5)2.65510.2422.0600.99700.78
Ours-Intense (β = 2.0)2.80310.1252.0380.99690.98
Ours-Moderate (β = 1.0)2.68910.1941.7830.99720.86
+ +Table 4. Quantitative comparison on motion frames, overlap and the proposed time-aware shift fusion. + +
Method / MetricSync-C↑Sync-D↓Smooth↑Runtime↓
Motion frames(8)2.46510.3230.996919.70
Overlap(8)2.54410.2650.996922.49
Shift-fusion2.68910.1940.997217.04
+ +# 4.3. Ablation Studies + +Temporal Control Inspires Intra-clip Diversity. In Table 3, the absence of temporal audio attention results in a decrease in both lip-sync and E-FID. We observe that the motion controller plays a crucial role in CelebV-HQ given that its videos exhibit more movement that weakly correlates with audio. We also investigate the effect of different dynamic scale. Increasing $\beta$ can further stimulate the lip-audio synchronization and the diversity of movements, while there is a slight decrease in stability and expression similarity. The instances in Figure 6 indicate + +![](images/b5ca64c6b30ab4751ecc4a740e92de9fc5491979d2dfd064d22c163a405fab67.jpg) +Figure 8. Visual results compared with motion frames and overlapping. + +temporal audio attention helps learn effective control of micro-expressions, while the motion controller is critical for agility and expressiveness. + +Larger Audio Perception Bridges Inter-clip Stability. Initially, we ablate the time-aware position shift fusion (briefly termed Shift) and generate clips independently as illustrated in Figure 7. The fusion process is performed at each timestep, gradually inverting the random noise to continuous and smooth long range videos. Subsequently, we explore different long video inference strategy: motion frames, overlap, our shift. The number of motion frames and overlap frames was set to 8. As indicated in Table 4, our shift approach attains superior lip-sync and smoothness without necessitating any auxiliary computation. From the qualitative comparison demonstrated in Figure 8, our method successfully handle long term consistency, circumventing identity similarity decrease in motion frames and mixed blurry texture in overlapping results. + +# 5. Conclusion + +In summary, our work present Sonic, an audio-driven portrait animation that focuses on the exploration of global audio perception to enhance the efficient generation of realistic lip synchronization, naturalness and temporal consistencies. The technical cornerstone mainly lies on context-enhanced audio learning and time-aware position shift fusion that work together to extend the intra-clip audio temporal to the global inter-clip audio perception. It significantly outperforms existing SOTA methods in video quality, motion diversity and naturalness. + +Acknowledgements We thank Dongshan Li, Zixiang Zhou, Zhentao Yu, Hongmei Wang, Joeyu Wang, Chengfei Cai, Shiyu Tang, Tianxiang Zheng, Junshu Tang, Shuai Shao, Yubin Zeng, Junxin Huang, Yue Zhang, Chao Zhan, Zhengyan Tong, Zhaokang Chen, Bin Wu, Wenjiang Zhou, Yang Zhang, Junqi Cheng, Jiaxi Gu, and Fangyuan Zou for helpful discussion. + +# References + +[1] Alexei Baevski, Yuhao Zhou, Abdelrahman Mohamed, and Michael Auli. wav2vec 2.0: A framework for self-supervised learning of speech representations. Advances in neural information processing systems, 33:12449-12460, 2020. 3 +[2] Omer Bar-Tal, Hila Chefer, Omer Tov, Charles Herrmann, Roni Paiss, Shiran Zada, Ariel Ephrat, Junhwa Hur, Yuanzhen Li, Tomer Michaeli, et al. Lumiere: A spacetime diffusion model for video generation. arXiv preprint arXiv:2401.12945, 2024. 3 +[3] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 3, 6 +[4] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22563-22575, 2023. 3 +[5] Aras Bozkurt, Xiao Junhong, Sarah Lambert, Angelica Pazurek, Helen Crompton, Suzan Koseoglu, Robert Farrow, Melissa Bond, Chrissi Nerantzi, Sarah Honeychurch, et al. Speculative futures on chatgpt and generative artificial intelligence (ai): A collective reflection from the educational landscape. Asian Journal of Distance Education, 18(1):53-130, 2023. 2 +[6] Haoxin Chen, Yong Zhang, Xiaodong Cun, Menghan Xia, Xintao Wang, Chao Weng, and Ying Shan. Videocrafter2: Overcoming data limitations for high-quality video diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7310-7320, 2024. 3 +[7] Lele Chen, Zhiheng Li, Ross K Maddox, Zhiyao Duan, and Chenliang Xu. Lip movements generation at a glance. In Proceedings of the European conference on computer vision (ECCV), pages 520-535, 2018. 3 +[8] Zhiyuan Chen, Jiajiong Cao, Zhiquan Chen, Yuming Li, and Chenguang Ma. Echomimic: Lifelike audio-driven portrait animations through editable landmark conditions. arXiv preprint arXiv:2407.08136, 2024. 2, 3, 6 +[9] Joon Son Chung and Andrew Zisserman. Out of time: automated lip sync in the wild. In Computer Vision-ACCV 2016 Workshops: ACCV 2016 International Workshops, Taipei, Taiwan, November 20-24, 2016, Revised Selected Papers, Part II 13, pages 251-263. Springer, 2017. 7 +[10] Joon Son Chung, Arsha Nagrani, and Andrew Zisserman. Voxceleb2: Deep speaker recognition. arXiv preprint arXiv:1806.05622, 2018. 6 +[11] Enric Corona, Andrei Zanfir, Eduard Gabriel Bazavan, Nikos Kolotouros, Thiemo Alldieck, and Cristian Sminchisescu. Vlogger: Multimodal diffusion for embodied avatar synthesis. arXiv preprint arXiv:2403.08764, 2024. 2 +[12] Jiahao Cui, Hui Li, Yao Yao, Hao Zhu, Hanlin Shang, Kaihui Cheng, Hang Zhou, Siyu Zhu, and Jingdong Wang. Hallo2: + +Long-duration and high-resolution audio-driven portrait image animation. arXiv preprint arXiv:2410.07718, 2024. 6 +[13] Yu Deng, Jiaolong Yang, Sicheng Xu, Dong Chen, Yunde Jia, and Xin Tong. Accurate 3d face reconstruction with weakly-supervised learning: From single image to image set. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 0–0, 2019. 7 +[14] Patrick Esser, Johnathan Chiu, Parmida Atighehchian, Jonathan Granskog, and Anastasis Germanidis. Structure and content-guided video synthesis with diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7346-7356, 2023. 2 +[15] Yuwei Guo, Ceyuan Yang, Anyi Rao, Zhengyang Liang, Yaohui Wang, Yu Qiao, Maneesh Agrawala, Dahua Lin, and Bo Dai. Animatediff: Animate your personalized text-to-image diffusion models without specific tuning. International Conference on Learning Representations, 2024. 3 +[16] Tianyu He, Junliang Guo, Runyi Yu, Yuchi Wang, Jialiang Zhu, Kaikai An, Leyi Li, Xu Tan, Chunyu Wang, Han Hu, et al. Gaia: Zero-shot talking avatar generation. arXiv preprint arXiv:2311.15230, 2023. 2 +[17] Yingqing He, Tianyu Yang, Yong Zhang, Ying Shan, and Qifeng Chen. Latent video diffusion models for high-fidelity long video generation. arXiv preprint arXiv:2211.13221, 2022.3 +[18] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 7 +[19] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. Advances in Neural Information Processing Systems, 35:8633-8646, 2022. 2 +[20] Wenyi Hong, Ming Ding, Wendi Zheng, Xinghan Liu, and Jie Tang. Cogvideo: Large-scale pretraining for text-to-video generation via transformers. arXiv preprint arXiv:2205.15868, 2022. 2 +[21] Ziqi Huang, Yinan He, Jiashuo Yu, Fan Zhang, Chenyang Si, Yuming Jiang, Yuanhan Zhang, Tianxing Wu, Qingyang Jin, Nattapol Chanpaisit, et al. Vbench: Comprehensive benchmark suite for video generative models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21807-21818, 2024. 7 +[22] Xiaozhong Ji, Chuming Lin, Zhonggan Ding, Ying Tai, Jian Yang, Junwei Zhu, Xiaobin Hu, Jiangning Zhang, Donghao Luo, and Chengjie Wang. Realtalk: Real-time and realistic audio-driven face generation with 3d facial prior-guided identity alignment network. arXiv preprint arXiv:2406.18284, 2024. 3 +[23] Jianwen Jiang, Chao Liang, Jiaqi Yang, Gaojie Lin, Tianyun Zhong, and Yanbo Zheng. Loopy: Taming audio-driven portrait avatar with long-term motion dependency. arXiv preprint arXiv:2409.02634, 2024. 2, 3, 4, 5 +[24] Esperanza Johnson, Ramón Hervás, Carlos Gutiérrez López de la Franca, Tania Mondéjar, Sergio F Ochoa, and Jesús + +Favela. Assessing empathy and managing emotions through interactions with an affective avatar. Health informatics journal, 24(2):182-193, 2018. 2 +[25] Oytun Kal and Yavuz Samur. Educational virtual reality game design for film and animation. Encyclopedia of Computer Graphics and Games, pages 621-636, 2024. 2 +[26] Levon Khachatryan, Andranik Movsisyan, Vahram Tadevosyan, Roberto Henschel, Zhangyang Wang, Shant Navasardyan, and Humphrey Shi. Text2video-zero: Text-to-image diffusion models are zero-shot video generators. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15954-15964, 2023. 2 +[27] Tao Liu, Feilong Chen, Shuai Fan, Chenpeng Du, Qi Chen, Xie Chen, and Kai Yu. Anitalker: Animate vivid and diverse talking faces through identity-decoupled facial motion encoding. arXiv preprint arXiv:2405.03121, 2024. 2 +[28] Yifeng Ma, Shiwei Zhang, Jiayu Wang, Xiang Wang, Yingya Zhang, and Zhidong Deng. Dreamtalk: When expressive talking head generation meets diffusion probabilistic models. arXiv preprint arXiv:2312.09767, 2023. 2, 3 +[29] KR Prajwal, Rudrabha Mukhopadhyay, Vinay P Namboodiri, and CV Jawahar. A lip sync expert is all you need for speech to lip generation in the wild. In Proceedings of the 28th ACM international conference on multimedia, pages 484-492, 2020. 2, 3 +[30] Alec Radford, Jong Wook Kim, Tao Xu, Greg Brockman, Christine McLeavey, and Ilya Sutskever. Robust speech recognition via large-scale weak supervision. In International conference on machine learning, pages 28492-28518. PMLR, 2023. 3 +[31] Imogen C Rehm, Emily Foenander, Klaire Wallace, JoAnne M Abbott, Michael Kyrios, and Neil Thomas. What role can avatars play in e-mental health interventions? exploring new models of client-therapist interaction. Frontiers in Psychiatry, 7:186, 2016. 2 +[32] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2 +[33] Shuai Shen, Wenliang Zhao, Zibin Meng, Wanhua Li, Zheng Zhu, Jie Zhou, and Jiwen Lu. Difftalk: Crafting diffusion models for generalized audio-driven portraits animation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1982-1991, 2023. 2 +[34] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792, 2022. 2 +[35] Xusen Sun, Longhao Zhang, Hao Zhu, Peng Zhang, Bang Zhang, Xinya Ji, Kangneng Zhou, Daiheng Gao, Liefeng Bo, and Xun Cao. Vividtalk: One-shot audio-driven talking head generation based on 3d hybrid prior. arXiv preprint arXiv:2312.01841, 2023. 2 +[36] Supasorn Suwajanakorn, Steven M Seitz, and Ira Kemelmacher-Shlizerman. Synthesizing obama: learn + +ing lip sync from audio. ACM Transactions on Graphics (ToG), 36(4):1-13, 2017. 3 +[37] Linrui Tian, Qi Wang, Bang Zhang, and Liefeng Bo. Emo: Emote portrait alive-generating expressive portrait videos with audio2video diffusion model under weak conditions. arXiv preprint arXiv:2402.17485, 2024. 2, 3, 4, 5, 7 +[38] Thomas Unterthiner, Sjoerd van Steenkiste, Karol Kurach, Raphaël Marinier, Marcin Michalski, and Sylvain Gelly. Fvd: A new metric for video generation. 2019. 7 +[39] Vikram Voleti, Alexia Jolicoeur-Martineau, and Chris Pal. Mcvd-masked conditional video diffusion for prediction, generation, and interpolation. Advances in neural information processing systems, 35:23371-23385, 2022. 3 +[40] Cong Wang, Kuan Tian, Jun Zhang, Yonghang Guan, Feng Luo, Fei Shen, Zhiwei Jiang, Qing Gu, Xiao Han, and Wei Yang. V-express: Conditional dropout for progressive training of portrait video generation. arXiv preprint arXiv:2406.02511, 2024. 2, 6 +[41] Xiang Wang, Hangjie Yuan, Shiwei Zhang, Dayou Chen, Jiuniu Wang, Yingya Zhang, Yujun Shen, Deli Zhao, and Jingren Zhou. Videocomposer: Compositional video synthesis with motion controllability. Advances in Neural Information Processing Systems, 36, 2024. 3 +[42] Huawei Wei, Zejun Yang, and Zhisheng Wang. Aniportrait: Audio-driven synthesis of photorealistic portrait animation. arXiv preprint arXiv:2403.17694, 2024. 2, 3, 6 +[43] Liangbin Xie, Xintao Wang, Honglun Zhang, Chao Dong, and Ying Shan. Vfhq: A high-quality dataset and benchmark for video face super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 657-666, 2022. 6 +[44] Mingwang Xu, Hui Li, Qingkun Su, Hanlin Shang, Liwei Zhang, Ce Liu, Jingdong Wang, Luc Van Gool, Yao Yao, and Siyu Zhu. Hallo: Hierarchical audio-driven visual synthesis for portrait image animation. arXiv preprint arXiv:2406.08801, 2024. 2, 3, 4, 5, 6 +[45] Sicheng Xu, Guojun Chen, Yu-Xiao Guo, Jiaolong Yang, Chong Li, Zhenyu Zang, Yizhong Zhang, Xin Tong, and Baining Guo. Vasa-1: Lifelike audio-driven talking faces generated in real time. arXiv preprint arXiv:2404.10667, 2024. 2, 3 +[46] Fei Yin, Yong Zhang, Xiaodong Cun, Mingdeng Cao, Yanbo Fan, Xuan Wang, Qingyan Bai, Baoyuan Wu, Jue Wang, and Yujiu Yang. Styleheat: One-shot high-resolution editable talking face generation via pre-trained stylegan. In European conference on computer vision, pages 85-101. Springer, 2022. 3 +[47] Shengming Yin, Chenfei Wu, Huan Yang, Jianfeng Wang, Xiaodong Wang, Minheng Ni, Zhengyuan Yang, Linjie Li, Shuguang Liu, Fan Yang, et al. Nuwa-xl: Diffusion over diffusion for extremely long video generation. arXiv preprint arXiv:2303.12346, 2023. 3 +[48] Jianhui Yu, Hao Zhu, Liming Jiang, Chen Change Loy, Weidong Cai, and Wayne Wu. Celebv-text: A large-scale facial text-video dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14805-14814, 2023. 6 + +[49] Rui Zhang, Yixiao Fang, Zhengnan Lu, Pei Cheng, Zebiao Huang, and Bin Fu. Linguualinker: Audio-driven portraits animation with implicit facial control enhancement. arXiv preprint arXiv:2407.18595, 2024. 2 +[50] Wenxuan Zhang, Xiaodong Cun, Xuan Wang, Yong Zhang, Xi Shen, Yu Guo, Ying Shan, and Fei Wang. Sadtalker: Learning realistic 3d motion coefficients for stylized audio-driven single image talking face animation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8652-8661, 2023. 2, 3, 4, 6 +[51] Zhimeng Zhang, Lincheng Li, Yu Ding, and Changjie Fan. Flow-guided one-shot talking face generation with a high-resolution audio-visual dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3661–3670, 2021. 6 +[52] Hao Zhu, Wayne Wu, Wentao Zhu, Liming Jiang, Siwei Tang, Li Zhang, Ziwei Liu, and Chen Change Loy. Celebvyhq: A large-scale video facial attributes dataset. In European conference on computer vision, pages 650-667. Springer, 2022. 6 \ No newline at end of file diff --git a/2025/Sonic_ Shifting Focus to Global Audio Perception in Portrait Animation/images.zip b/2025/Sonic_ Shifting Focus to Global Audio Perception in Portrait Animation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..11e3e68005e663428a3a4046394c1a6ba2632dab --- /dev/null +++ b/2025/Sonic_ Shifting Focus to Global Audio Perception in Portrait Animation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:846f0a141c02070708a1228f33321f0f55ebf1ccf26dc8076ae12b406c40e067 +size 954598 diff --git a/2025/Sonic_ Shifting Focus to Global Audio Perception in Portrait Animation/layout.json b/2025/Sonic_ Shifting Focus to Global Audio Perception in Portrait Animation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..68c9309e92e42e44d901c710d1f0de155ba43742 --- /dev/null +++ b/2025/Sonic_ Shifting Focus to Global Audio Perception in Portrait Animation/layout.json @@ -0,0 +1,7695 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 84, + 103, + 526, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 84, + 103, + 526, + 121 + ], + "spans": [ + { + "bbox": [ + 84, + 103, + 526, + 121 + ], + "type": "text", + "content": "Sonic: Shifting Focus to Global Audio Perception in Portrait Animation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 73, + 142, + 536, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 142, + 536, + 187 + ], + "spans": [ + { + "bbox": [ + 73, + 142, + 536, + 187 + ], + "type": "text", + "content": "Xiaozhong Ji" + }, + { + "bbox": [ + 73, + 142, + 536, + 187 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 73, + 142, + 536, + 187 + ], + "type": "text", + "content": " Xiaobin Hu" + }, + { + "bbox": [ + 73, + 142, + 536, + 187 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 73, + 142, + 536, + 187 + ], + "type": "text", + "content": " Zhihong Xu" + }, + { + "bbox": [ + 73, + 142, + 536, + 187 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 73, + 142, + 536, + 187 + ], + "type": "text", + "content": " Junwei Zhu" + }, + { + "bbox": [ + 73, + 142, + 536, + 187 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 73, + 142, + 536, + 187 + ], + "type": "text", + "content": " Chuming Lin" + }, + { + "bbox": [ + 73, + 142, + 536, + 187 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 73, + 142, + 536, + 187 + ], + "type": "text", + "content": " Qingdong He" + }, + { + "bbox": [ + 73, + 142, + 536, + 187 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 73, + 142, + 536, + 187 + ], + "type": "text", + "content": " Jiangning Zhang" + }, + { + "bbox": [ + 73, + 142, + 536, + 187 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 73, + 142, + 536, + 187 + ], + "type": "text", + "content": " Donghao Luo" + }, + { + "bbox": [ + 73, + 142, + 536, + 187 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 73, + 142, + 536, + 187 + ], + "type": "text", + "content": " Yi Chen" + }, + { + "bbox": [ + 73, + 142, + 536, + 187 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 73, + 142, + 536, + 187 + ], + "type": "text", + "content": " Qin Lin" + }, + { + "bbox": [ + 73, + 142, + 536, + 187 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 73, + 142, + 536, + 187 + ], + "type": "text", + "content": " Qinglin Lu" + }, + { + "bbox": [ + 73, + 142, + 536, + 187 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 73, + 142, + 536, + 187 + ], + "type": "text", + "content": " Chengjie Wang" + }, + { + "bbox": [ + 73, + 142, + 536, + 187 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 73, + 142, + 536, + 187 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 73, + 142, + 536, + 187 + ], + "type": "text", + "content": "Tencent " + }, + { + "bbox": [ + 73, + 142, + 536, + 187 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 73, + 142, + 536, + 187 + ], + "type": "text", + "content": "Zhejiang University" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 85, + 204, + 269, + 278 + ], + "blocks": [ + { + "bbox": [ + 85, + 204, + 269, + 278 + ], + "lines": [ + { + "bbox": [ + 85, + 204, + 269, + 278 + ], + "spans": [ + { + "bbox": [ + 85, + 204, + 269, + 278 + ], + "type": "image", + "image_path": "4cdac659236a27f42818a2d5d96f7d65ea4e189ed7cb8f635c69cd635001d723.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 282, + 204, + 523, + 279 + ], + "blocks": [ + { + "bbox": [ + 282, + 204, + 523, + 279 + ], + "lines": [ + { + "bbox": [ + 282, + 204, + 523, + 279 + ], + "spans": [ + { + "bbox": [ + 282, + 204, + 523, + 279 + ], + "type": "image", + "image_path": "258097e14bae0c985142d9d7ac64a532dbf2f780721f6225b9a92e1b8ecba330.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 84, + 281, + 263, + 378 + ], + "blocks": [ + { + "bbox": [ + 84, + 281, + 263, + 378 + ], + "lines": [ + { + "bbox": [ + 84, + 281, + 263, + 378 + ], + "spans": [ + { + "bbox": [ + 84, + 281, + 263, + 378 + ], + "type": "image", + "image_path": "1457ec753d8173dbf7a05a442559e8dc0dbe82f94280f86c25afafe79bdae14f.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 380, + 555, + 425 + ], + "lines": [ + { + "bbox": [ + 55, + 380, + 555, + 425 + ], + "spans": [ + { + "bbox": [ + 55, + 380, + 555, + 425 + ], + "type": "text", + "content": "Figure 1. Sonic excels in producing vivid portrait animation videos given a reference image and an audio clip. Beyond its fundamental lip-syncing capabilities, Sonic demonstrates proficiency in creating a diverse spectrum of facial expressions and adaptable head movements. Notably, when dealing with extended videos, Sonic can yield stable and seamless outcomes in a parallel fashion, all while maintaining an unified paradigm focusing on global audio perception." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 272, + 281, + 520, + 378 + ], + "blocks": [ + { + "bbox": [ + 272, + 281, + 520, + 378 + ], + "lines": [ + { + "bbox": [ + 272, + 281, + 520, + 378 + ], + "spans": [ + { + "bbox": [ + 272, + 281, + 520, + 378 + ], + "type": "image", + "image_path": "b5e2e39a45c721c1e21dd6850c86200acd2cfbcb49d4f9634c16dc26d769eea5.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 151, + 434, + 200, + 445 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 434, + 200, + 445 + ], + "spans": [ + { + "bbox": [ + 151, + 434, + 200, + 445 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 54, + 460, + 297, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 460, + 297, + 676 + ], + "spans": [ + { + "bbox": [ + 54, + 460, + 297, + 676 + ], + "type": "text", + "content": "The study of talking face generation mainly explores the intricacies of synchronizing facial movements and crafting visually appealing, temporally-coherent animations. However, due to the limited exploration of global audio perception, current approaches predominantly employ auxiliary visual and spatial knowledge to stabilize the movements, which often results in the deterioration of the naturalness and temporal inconsistencies. Considering the essence of audio-driven animation, the audio signal serves as the ideal and unique priors to adjust facial expressions and lip movements, without resorting to interference of any visual signals. Based on this motivation, we propose a novel paradigm, dubbed as Sonic, to shift focus on the exploration of global audio perception. To effectively leverage global audio knowledge, we disentangle it into intra- and inter-clip audio perception and collaborate with both aspects to enhance overall perception. For the intra-clip audio perception, 1). Context-enhanced audio learning, in" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 435, + 556, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 435, + 556, + 616 + ], + "spans": [ + { + "bbox": [ + 313, + 435, + 556, + 616 + ], + "type": "text", + "content": "which long-range intra-clip temporal audio knowledge is extracted to provide facial expression and lip motion priors implicitly expressed as the tone and speed of speech. 2). Motion-decoupled controller, in which the motion of the head and expression movement are disentangled and independently controlled by intra-audio clips. Most importantly, for inter-clip audio perception, as a bridge to connect the intra-clips to achieve the global perception, Time-aware position shift fusion, in which the global inter-clip audio information is considered and fused for long-audio inference via through consecutively time-aware shifted windows. Extensive experiments demonstrate that the novel audio-driven paradigm outperform existing SOTA methodologies in terms of video quality, temporally consistency, lip synchronization precision, and motion diversity." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 643, + 395, + 656 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 643, + 395, + 656 + ], + "spans": [ + { + "bbox": [ + 314, + 643, + 395, + 656 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 665, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 555, + 715 + ], + "type": "text", + "content": "Talking Face Animation is a technique aimed at creating a lifelike representation of a speaking person by animating a still image in sync with speech audio. This technology holds great potential in enhancing the realism of virtual re" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 683, + 277, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 683, + 277, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 683, + 277, + 713 + ], + "type": "text", + "content": "* Equal Contribution, † Project Leader \nCorresponding Author \nProject page: https://jixiaozhong.github.io/Sonic/." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "193" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 192 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 192 + ], + "type": "text", + "content": "ality [5], improving the efficiency of film and gaming making [25], facilitating accessibility for individuals with communication impairments [24], and offering therapeutic support and social interaction in healthcare [31]. Prominent progress in this domain is demonstrated by notable works based on Stable Diffusion [32], which project the parametric or implicit representations of lip movements [29, 50], facial expressions [16, 28, 33], and head movement [35] in the latent space [11, 27, 37, 45] to generate high-quality video with a high level of realism and liveliness." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 196, + 296, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 196, + 296, + 422 + ], + "spans": [ + { + "bbox": [ + 56, + 196, + 296, + 422 + ], + "type": "text", + "content": "However, precise audio control and temporal coherence have not yet been effectively achieved. Current approaches [8, 23, 37, 40, 42, 44, 49, 50] predominantly addressed these two issues in a completely segregated manner, overlooking the holistic coordination between audio and vision. For audio control, audio features were segmented according to timestamps and matched to each visual frame. By applying cross-attention within a spatial latent space, each frame was restricted to the nearby audio information, which hindered the transformation into an optimal motion representation. To achieve temporal coherence, self-attention along the temporal dimension was applied to smooth visual features within a clip, while overlapping or motion frames were used to enhance coherence between different clips. However, these stability strategies have limited temporal receptive fields, which can reduce the richness of motion. Additionally, they do not take audio information into account, potentially disrupting the synchronization between audio and visual elements." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 427, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 427, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 56, + 427, + 296, + 714 + ], + "type": "text", + "content": "To address these issues, we propose a novel paradigm, dubbed as Sonic, to extent the range of receptive field to the global level guided by the audio priors. To achieve global audio perception, we decompose the talking head video generation task into two aspects: intra-clip (i.e., lip sync, expression alignment, and head movement) and inter-clip (i.e., motion rhythm and smoothness, long video stability). Sonic aims to exploit global audio perception rather than the motion frames and other visual motion (i.e., referring to the face locator and landmark prior adopted by existing methods, which hinders the ability to perceive information directly from audio). Performing as long-range and global signal, audio shows tone, rhythm and speed of speech which implicitly provides expression and head movement motion priors. However, we have to face such a tricky challenge that, as a weak correlation cross-modality signal without the guidance of any visual motion (e.g., motion frames), current audio learning paradigm leads to the temporal jitter and unsatisfactory pattern and fails to well-align the audio and motion. To this end, we first customize efficient global audio perception learning mechanism. Specifically, context-enhanced audio learning is proposed as long-range audio learning module to extract the intra-clip audio temporal knowledge from the input audio clip. Then such intra" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 313, + 72, + 555, + 393 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 393 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 393 + ], + "type": "text", + "content": "clip audio temporal information is mapped as a temporal-embeddings for the latter fusion by audio-temporal cross-attention. The temporal-embeddings projection operation can effectively reduce the computational burden and capture temporal priors within current audio clip. Although context-enhanced audio learning directs the temporal movements correlated with audio, such as lip movements synchronized with speech and facial expressions related to the content, there still exists the absence of habitual head movements. Secondly, we design motion-decoupled controller to disentangle the habitual head and expression movement and support the independent control via two explicit parameters learned from the current intra-clip audio. To enrich the functionality and playability of animation, we also add an plug to allow users to define some exaggerated movements. Most importantly, for inter-clip audio perception, time-aware position shift fusion enlarges the intra-clip audio perception to the global inter-clip audio perception via time-aware shifted windows consecutively bridging the preceding clip, providing the global inter-clip connection that significantly enhance modeling temporal power, as demonstrated in Figure 1. It is worth noting that our time-aware position shift fusion requires no extra training cost compared with the motion frames in most recent works, such as EMO [37], Loopy [23] and also no additional inference time caused by overlapping frames. Our contributions are summarized as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 314, + 396, + 553, + 575 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 314, + 396, + 553, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 396, + 553, + 432 + ], + "spans": [ + { + "bbox": [ + 314, + 396, + 553, + 432 + ], + "type": "text", + "content": "- We propose a novel unified paradigm without the guidance of visual motion, dubbed as Sonic, to focus on the exploration of global audio perception." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 314, + 433, + 553, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 433, + 553, + 479 + ], + "spans": [ + { + "bbox": [ + 314, + 433, + 553, + 479 + ], + "type": "text", + "content": "- For intra-clip audio perception, context-enhanced audio learning and motion-decoupled controller excavate the intra-clip audio temporal knowledge to direct the temporal movements." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 480, + 553, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 480, + 553, + 528 + ], + "spans": [ + { + "bbox": [ + 314, + 480, + 553, + 528 + ], + "type": "text", + "content": "- For inter-clip audio perception, we propose time-aware position shift fusion to enhance the global inter-clip audio perception by extending the intra-clip audio perception through consecutively time-aware shifted windows." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 529, + 553, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 529, + 553, + 575 + ], + "spans": [ + { + "bbox": [ + 314, + 529, + 553, + 575 + ], + "type": "text", + "content": "- Extensive experiments demonstrate that the novel unified global audio perception paradigm outperforms existing SOTA methods in terms of video quality, temporal consistency, lip synchronization, and motion diversity." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 590, + 405, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 590, + 405, + 601 + ], + "spans": [ + { + "bbox": [ + 314, + 590, + 405, + 601 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 610, + 495, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 610, + 495, + 621 + ], + "spans": [ + { + "bbox": [ + 314, + 610, + 495, + 621 + ], + "type": "text", + "content": "2.1. Diffusion-based Video Generation" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 630, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 555, + 715 + ], + "type": "text", + "content": "The remarkable achievements in diffusion models in image generation have sparked widespread research interest for video generation. Early endeavors in video generation [14, 19, 20, 26, 34] have primarily concentrated on devising image-based diffusion models with temporal module to effectively capture the dynamic characteristics inherent in video sequences. To leverage the powerful abil" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "194" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 57, + 72, + 294, + 311 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 72, + 294, + 311 + ], + "spans": [ + { + "bbox": [ + 57, + 72, + 294, + 311 + ], + "type": "text", + "content": "ity of pretrained image diffusion models, Video LDM [4] incorporates temporal layers that learn to align images in a temporally consistent manner. AnimateDiff [15] provides a motion-specific module by multi-stage progressive training manner, and the motion module can be seamlessly integrated into existing text-to-image architectures without any further modifications. VideoComposer [41] and VideoCrafter [6] further explore the synthesis of image-to-video generation via textual and visual features combinations. Despite these notable advancements, the field still struggles to overcome the challenge of generating temporal coherent and high-quality videos. As a powerful motion representation trained on a large well-curated high-quality dataset, Stable Video Diffusion [3] shows the remarkable temporal-consistency and the high-quality visual characters for downstream tasks. There exists a large gap between the audio-driven portrait animation and general video generation due to the fact that audio-driven task aims to use the weak cross-modality audio signal to precisely control motion movements (e.g., lip and expression movement)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 332, + 257, + 344 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 332, + 257, + 344 + ], + "spans": [ + { + "bbox": [ + 57, + 332, + 257, + 344 + ], + "type": "text", + "content": "2.2. Audio-driven Talking Face Generation" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 355, + 294, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 355, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 355, + 294, + 712 + ], + "type": "text", + "content": "The generation of talking face videos from audio inputs has been a longstanding task. In the early stages, the emphasis was primarily on synthesizing lip movements alone, accomplished by directly mapping audio signals to lip movements while keeping other facial attributes unchanged [7, 22, 29, 36, 46]. In more recent efforts, the scope has been broadened to encompass a wider range of facial expressions and head movements, which are derived from audio inputs and a single reference image. These methods can be broadly classified into with or without intermediate representation. For instance, by employing 3D coefficients representation (e.g., Sadtalker [50], Dreamtalk [28], Ani-Portrait [42], VASA-1 [45]), the video generation process can be informed through the direct or indirect visible control signal in the aspect of head movement and expression. Nevertheless, a recurring challenge faced by these techniques is the limited capacity of the 3D mesh to capture intricate details, consequently constraining the overall dynamism and realism. In contrast, the methods that omit intermediate representation exhibit superior naturalness and consistent preservation of identity with the original image. EMO [37], Hallo [44], and Loopy [23] employ a direct audio-to-video synthesis approach to ensure a high level of realism and naturalness, eliminating the need for intermediate 3D representation. However, these approaches primarily rely on additional conditions associated with spatial motion or motion frames to maintain temporal consistency. In this regard, they overlook the crucial role of audio in audio-driven animation, consequently diminishing both expressiveness and audio-motion synchronization." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 317, + 72, + 437, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 72, + 437, + 84 + ], + "spans": [ + { + "bbox": [ + 317, + 72, + 437, + 84 + ], + "type": "text", + "content": "2.3. Long Video Inference" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 317, + 91, + 553, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 91, + 553, + 293 + ], + "spans": [ + { + "bbox": [ + 317, + 91, + 553, + 293 + ], + "type": "text", + "content": "Significant research efforts have been devoted to extending the duration of generated videos, aiming to expand their practical applications. Autoregressively predicting successive frames [17, 39] and hierarchical coarse-to-fine methods [47] have emerged to generate long videos. Then Lumiere [2] divides the video into overlapping temporal segments, denoises each segment independently, and finally fuse these segments in an optimization algorithm. However there is no available high-efficiency method for audio-condition global perception. As one step towards achieving stronger capabilities, our Sonic proposes time-aware position shift fusion strategy progressively denoise intra-clip audio to establish a global inter-clip connection through the use of time-aware shifted windows that bridge the preceding clip along the timesteps axis. This inter-clip fusion with global receptive field needs no additional training cost and introduce no extra inference time without overlapping frames." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 317, + 307, + 395, + 319 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 307, + 395, + 319 + ], + "spans": [ + { + "bbox": [ + 317, + 307, + 395, + 319 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 317, + 327, + 553, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 327, + 553, + 422 + ], + "spans": [ + { + "bbox": [ + 317, + 327, + 553, + 422 + ], + "type": "text", + "content": "Overall framework. Given a single portrait reference image and an input audio, our one-stage framework can generate a portrait video well-synchronized with the driven audio. Our proposed Sonic framework aims to focus on the exploration of global audio perception, primarily consisting of context-enhanced audio learning, motion-decoupled controller, and time-aware position shift fusion module, as illustrated in Figure 2." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 317, + 423, + 553, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 423, + 553, + 529 + ], + "spans": [ + { + "bbox": [ + 317, + 423, + 553, + 529 + ], + "type": "text", + "content": "Technical contribution. To achieve global audio perception, we decompose the talking head video generation task into two aspects: intra-clip (i.e., lip sync, expression alignment, and head movement) and inter-clip (i.e., motion rhythm and smoothness, long video stability). For intra-clip, we enhance the visual realism that matches audio signal through the context-enhanced audio learning and the motion controller. For inter-clip, we improve the stability using the time-aware position shift fusion strategy." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 317, + 540, + 498, + 551 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 540, + 498, + 551 + ], + "spans": [ + { + "bbox": [ + 317, + 540, + 498, + 551 + ], + "type": "text", + "content": "3.1. Context-enhanced Audio Learning" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 317, + 558, + 553, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 558, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 317, + 558, + 553, + 712 + ], + "type": "text", + "content": "For audio-talking face animation, the audio is the ideally unique and global signal to adjust the face expression and movement without interference from other visual signals. Thus, the motivation behind context-enhanced audio learning is to guide the generation of spatial and temporal frames solely based on audio signals, rather than motion frames or other visual signals, to achieve the high-realism generation. Specifically, we adopt Whisper-Tiny [30] as the pretrained audio feature extraction model, which is more lightweight than Wav2Vec [1] used by most existing methods [8, 37, 44]. For each audio frame, the features from the last layers of five stages are concatenated together for multiscale understanding. For each video frame, a duration of" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 299, + 750, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 299, + 750, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 299, + 750, + 312, + 757 + ], + "type": "text", + "content": "195" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 83, + 72, + 216, + 270 + ], + "blocks": [ + { + "bbox": [ + 83, + 72, + 216, + 270 + ], + "lines": [ + { + "bbox": [ + 83, + 72, + 216, + 270 + ], + "spans": [ + { + "bbox": [ + 83, + 72, + 216, + 270 + ], + "type": "image", + "image_path": "a1775a9d25b07070e02f41b8fc2df232ce1ecec833567882c7ae742a6b6389ed.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 271, + 555, + 304 + ], + "lines": [ + { + "bbox": [ + 55, + 271, + 555, + 304 + ], + "spans": [ + { + "bbox": [ + 55, + 271, + 555, + 304 + ], + "type": "text", + "content": "Figure 2. Framework of our approach. Sonic processes each clip of the long audio in parallel, shifting to a new context at each time step to progressively fuse inter-clip latent features across global audio perception. In Sonic, we enhance intra-clip temporal audio context learning and decouple motion to improve dynamics." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 225, + 72, + 306, + 268 + ], + "blocks": [ + { + "bbox": [ + 225, + 72, + 306, + 268 + ], + "lines": [ + { + "bbox": [ + 225, + 72, + 306, + 268 + ], + "spans": [ + { + "bbox": [ + 225, + 72, + 306, + 268 + ], + "type": "image", + "image_path": "5339a31b068a57e4d1d9409ccc5d5a80d511b2d1e6bb19c7de2f4d2a66916b02.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 312, + 76, + 526, + 270 + ], + "blocks": [ + { + "bbox": [ + 312, + 76, + 526, + 270 + ], + "lines": [ + { + "bbox": [ + 312, + 76, + 526, + 270 + ], + "spans": [ + { + "bbox": [ + 312, + 76, + 526, + 270 + ], + "type": "image", + "image_path": "f484da2b775929682cd4a9d4bf8ec8589faca09ae2f7863e73281d2d607cac78.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 316, + 296, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 316, + 296, + 437 + ], + "spans": [ + { + "bbox": [ + 55, + 316, + 296, + 437 + ], + "type": "text", + "content": "0.2s audio feature is used frame-wise to give rich context. We use three linear layers to project audio feature to match the dimension of cross attention. The transformed audio embedding is denoted as " + }, + { + "bbox": [ + 55, + 316, + 296, + 437 + ], + "type": "inline_equation", + "content": "c_{a} \\in \\mathcal{R}^{b \\times f \\times d \\times c}" + }, + { + "bbox": [ + 55, + 316, + 296, + 437 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 316, + 296, + 437 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 55, + 316, + 296, + 437 + ], + "type": "text", + "content": " denotes batch size, " + }, + { + "bbox": [ + 55, + 316, + 296, + 437 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 55, + 316, + 296, + 437 + ], + "type": "text", + "content": " is audio length, " + }, + { + "bbox": [ + 55, + 316, + 296, + 437 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 55, + 316, + 296, + 437 + ], + "type": "text", + "content": " is num of context tokens, and " + }, + { + "bbox": [ + 55, + 316, + 296, + 437 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 316, + 296, + 437 + ], + "type": "text", + "content": " is hidden size of the cross-attention. To focus on talking face area, the core latent features is restricted with a mask " + }, + { + "bbox": [ + 55, + 316, + 296, + 437 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 316, + 296, + 437 + ], + "type": "text", + "content": " determined by a joint bounding boxes of faces. The audio signal can act as the spatial knowledge provider being injected into the spatial cross-attention layers as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 70, + 441, + 295, + 456 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 441, + 295, + 456 + ], + "spans": [ + { + "bbox": [ + 70, + 441, + 295, + 456 + ], + "type": "interline_equation", + "content": "z _ {s} ^ {\\prime} = z _ {s} + \\operatorname {C r o s s A t t n} \\left(Q \\left(z _ {s}\\right), K \\left(c _ {a}\\right), V \\left(c _ {a}\\right)\\right) \\cdot M, \\tag {1}", + "image_path": "67873c8c5b3c1dc1aab0d21ad1b040f238b0ded0999eec8ae273e2562e3c242a.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 462, + 295, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 462, + 295, + 498 + ], + "spans": [ + { + "bbox": [ + 55, + 462, + 295, + 498 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 462, + 295, + 498 + ], + "type": "inline_equation", + "content": "z_{s}" + }, + { + "bbox": [ + 55, + 462, + 295, + 498 + ], + "type": "text", + "content": " is spatial latent features of self-attention layer, and " + }, + { + "bbox": [ + 55, + 462, + 295, + 498 + ], + "type": "inline_equation", + "content": "z_{s}^{\\prime}" + }, + { + "bbox": [ + 55, + 462, + 295, + 498 + ], + "type": "text", + "content": " is the adjusted spatial features guided by audio signals in spatial-aware level." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 498, + 295, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 295, + 666 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 295, + 666 + ], + "type": "text", + "content": "Furthermore, audio is a long-range and global signal that provides expression and head movement motion priors implicitly expressed through the tone and speed of speech. To excavate the temporal information from the long-range audio knowledge, we customize a temporal-audio module for precise motion control. The temporal module in existing methods [37, 50] followsAnimateDiff, which performs only self-attention across the temporal dimension without audio aids. Some other talking-face [23, 44] works extract the motion priors from the visual knowledge (e.g., motion frames), which deviates from the essence of audio-driven task that relies solely on audio to drive movement. Thus we propose a temporal audio cross-attention module, introducing audio embedding to guide temporal alignment." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 666, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 666, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 666, + 296, + 714 + ], + "type": "text", + "content": "The preceding latent from temporal self-attention layers can be reshaped as " + }, + { + "bbox": [ + 55, + 666, + 296, + 714 + ], + "type": "inline_equation", + "content": "z_{t} \\in \\mathcal{R}^{(b \\times h \\times w) \\times f \\times c}" + }, + { + "bbox": [ + 55, + 666, + 296, + 714 + ], + "type": "text", + "content": ". To reduce the computation burden, the audio feature " + }, + { + "bbox": [ + 55, + 666, + 296, + 714 + ], + "type": "inline_equation", + "content": "c_{a}" + }, + { + "bbox": [ + 55, + 666, + 296, + 714 + ], + "type": "text", + "content": " is average-pooled along temporal dimension as the shape " + }, + { + "bbox": [ + 55, + 666, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{R}^{(b \\times 1) \\times d \\times c}" + }, + { + "bbox": [ + 55, + 666, + 296, + 714 + ], + "type": "text", + "content": ", then" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 316, + 555, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 316, + 555, + 353 + ], + "spans": [ + { + "bbox": [ + 313, + 316, + 555, + 353 + ], + "type": "text", + "content": "repeated and reshaped as " + }, + { + "bbox": [ + 313, + 316, + 555, + 353 + ], + "type": "inline_equation", + "content": "c_{ta} \\in \\mathcal{R}^{(b \\times h \\times w) \\times d \\times c}" + }, + { + "bbox": [ + 313, + 316, + 555, + 353 + ], + "type": "text", + "content": ". The projected audio temporal-embeddings are injected into the denoising U-Net via a temporal cross-attention blocks as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 340, + 357, + 555, + 373 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 340, + 357, + 555, + 373 + ], + "spans": [ + { + "bbox": [ + 340, + 357, + 555, + 373 + ], + "type": "interline_equation", + "content": "z _ {t} ^ {\\prime} = z _ {t} + \\operatorname {C r o s s A t t n} (Q (z _ {t}), K (c _ {t a}), V (c _ {t a})). \\tag {2}", + "image_path": "258bb6bc96ae773c108820d7d37d0a3a478023f38fd9267c53b295b53b640ea8.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 376, + 555, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 376, + 555, + 413 + ], + "spans": [ + { + "bbox": [ + 313, + 376, + 555, + 413 + ], + "type": "text", + "content": "The spatial and temporal audio attention layers are cascaded in multiple down-stages and up-stages to enable natural appearance and smooth temporal movement." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 419, + 474, + 432 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 419, + 474, + 432 + ], + "spans": [ + { + "bbox": [ + 313, + 419, + 474, + 432 + ], + "type": "text", + "content": "3.2. Motion-decoupled Controller" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 437, + 555, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 555, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 555, + 665 + ], + "type": "text", + "content": "Context-enhanced Audio Learning governs the temporal movements strongly correlated with audio, such as lip movements synchronized with speech and facial expressions related to the content. On the other hand, our motion-decoupled controller performs a complementary role by directing the temporal movements weakly associated with audio, such as habitual head movements and random changes in perspective. This controller can disentangle motion into explicit head and expression movement amplitudes, thereby enhancing both functionality, playability, and interactivity. Thus, we introduce independent augmented motion-bucket parameters which effectively influence motion patterns of head movements and expression strength. During the training phase, the translation motion-bucket " + }, + { + "bbox": [ + 313, + 437, + 555, + 665 + ], + "type": "inline_equation", + "content": "m_t" + }, + { + "bbox": [ + 313, + 437, + 555, + 665 + ], + "type": "text", + "content": " is computed as the variance of bounding boxes of video clip, while the expression motion-bucket " + }, + { + "bbox": [ + 313, + 437, + 555, + 665 + ], + "type": "inline_equation", + "content": "m_e" + }, + { + "bbox": [ + 313, + 437, + 555, + 665 + ], + "type": "text", + "content": " is variance of relative landmarks. Both buckets are integers in the range of [0, 128]. Through position encoding and linear projection, they are added into ResNet blocks as follows:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 378, + 671, + 553, + 685 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 378, + 671, + 553, + 685 + ], + "spans": [ + { + "bbox": [ + 378, + 671, + 553, + 685 + ], + "type": "interline_equation", + "content": "e m b = W \\left[ \\operatorname {P e} \\left(m _ {t}\\right), \\operatorname {P e} \\left(m _ {e}\\right) \\right], \\tag {3}", + "image_path": "95f7aca5fdd9bc7d5c02ee115182ad29c54de8fcdc26a105fc0145caa1fecae4.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "inline_equation", + "content": "emb" + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": " is the embedding, " + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{Pe}" + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": " is the position encoding, and " + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": " represents the linear projection weights." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "text", + "content": "196" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 216 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 216 + ], + "type": "text", + "content": "For practice application, motion-decoupled controller also supports the automatically adjustable parameters learning from the audio and reference image, and deeply explore the correlation between audio and portrait reference without requiring the parameter setting from users. To simplify multiple parameters and maintain adjustability, the predicted motion buckets are multiplying by a scale " + }, + { + "bbox": [ + 55, + 72, + 296, + 216 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 55, + 72, + 296, + 216 + ], + "type": "text", + "content": " (0.5 for mild dynamic, 1.0 for moderate dynamic, and 2.0 for intense dynamic). Specifically, we utilize a audio-to-bucket manner to adaptively predict the motion-buckets relying on the audio and reference image. In inference phase, the motion buckets are predicted and rescaled as:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 118, + 220, + 295, + 236 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 220, + 295, + 236 + ], + "spans": [ + { + "bbox": [ + 118, + 220, + 295, + 236 + ], + "type": "interline_equation", + "content": "m _ {t} ^ {\\prime}, m _ {e} ^ {\\prime} = \\beta \\times \\mathrm {E} _ {\\mathrm {b}} \\left(c _ {a}, R _ {i m g}\\right). \\tag {4}", + "image_path": "b32912c5ba671dbbf244e807275500592833dcc631b83cf57a48e03c5037c4d2.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 242, + 296, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 242, + 296, + 301 + ], + "spans": [ + { + "bbox": [ + 55, + 242, + 296, + 301 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 242, + 296, + 301 + ], + "type": "inline_equation", + "content": "m_t'" + }, + { + "bbox": [ + 55, + 242, + 296, + 301 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 242, + 296, + 301 + ], + "type": "inline_equation", + "content": "m_e'" + }, + { + "bbox": [ + 55, + 242, + 296, + 301 + ], + "type": "text", + "content": " are the estimated translation and expression motion-bucket achieved by the audio " + }, + { + "bbox": [ + 55, + 242, + 296, + 301 + ], + "type": "inline_equation", + "content": "c_a" + }, + { + "bbox": [ + 55, + 242, + 296, + 301 + ], + "type": "text", + "content": " and single reference image CLIP embedding " + }, + { + "bbox": [ + 55, + 242, + 296, + 301 + ], + "type": "inline_equation", + "content": "R_{img}" + }, + { + "bbox": [ + 55, + 242, + 296, + 301 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 242, + 296, + 301 + ], + "type": "inline_equation", + "content": "E_b" + }, + { + "bbox": [ + 55, + 242, + 296, + 301 + ], + "type": "text", + "content": " is the audio-to-bucket projection consisting of there linear layers with the ReLU as activation function." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 308, + 234, + 319 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 308, + 234, + 319 + ], + "spans": [ + { + "bbox": [ + 55, + 308, + 234, + 319 + ], + "type": "text", + "content": "3.3. Time-aware Position Shift Fusion" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 60, + 338, + 284, + 426 + ], + "blocks": [ + { + "bbox": [ + 60, + 338, + 284, + 426 + ], + "lines": [ + { + "bbox": [ + 60, + 338, + 284, + 426 + ], + "spans": [ + { + "bbox": [ + 60, + 338, + 284, + 426 + ], + "type": "image", + "image_path": "4967c461a53e2a8190cbf3f9b40249587298b2997e31cc5da103aab701f3a0b3.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 429, + 296, + 483 + ], + "lines": [ + { + "bbox": [ + 55, + 429, + 296, + 483 + ], + "spans": [ + { + "bbox": [ + 55, + 429, + 296, + 483 + ], + "type": "text", + "content": "Figure 3. Illustration of the proposed time-aware position-shift fusion. The model processes each clip non-overlapping. In next timestep, the model starts from a new position determined by the offset, thereby integrating long-range context. Specifically, the tail latents are filled cyclically from the head." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 498, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 296, + 713 + ], + "type": "text", + "content": "Recent methods [23, 37, 44] use motion frames or overlapping latent during coherent denoising, which suffer from a limited receptive field and increased the training or inference computational complexity for audio-conditioned video generation problem. In contrast, we propose a bridge, the time-aware position shift fusion, to extent the intra-clip audio perception to the global inter-clip audio-perception via consecutively time-aware shifted windows. The simple yet effective fusion strategy, without increasing extra inference or training time costs, can effectively address jitter and abrupt transition issues. Audio serves as the input condition for the talking face, where longer audio prompts the generation of longer videos. Based on this principle, we tailor time-aware position-shift fusion mechanism along the timestep axis where the latent sliding window is progressively shifted by a certain length at different timesteps while also preserving the frame correspondence relationship between audio and video signal." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 555, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 335 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 335 + ], + "type": "text", + "content": "Specifically, as described in Figure 3 and Algorithm 1, our time-aware position-shift fusion consists of two nested loops. The outer loop is inverse diffusion process, while the inner loop is a sliding window process that model predicts for the full-length audio conditions. At each timestep, the model takes a clip of audio input to predict the corresponding latent with a start offset to smooth the segment from the last timestep. The video diffusion model naturally bridges the context, generating continuous movement following the audio prompts. In detail, we set a cumulative shift offset " + }, + { + "bbox": [ + 313, + 72, + 555, + 335 + ], + "type": "inline_equation", + "content": "\\alpha_{\\Sigma}" + }, + { + "bbox": [ + 313, + 72, + 555, + 335 + ], + "type": "text", + "content": ", accumulating " + }, + { + "bbox": [ + 313, + 72, + 555, + 335 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 72, + 555, + 335 + ], + "type": "text", + "content": " at each timestep to ensure that the model starts sliding window from different positions. The offset " + }, + { + "bbox": [ + 313, + 72, + 555, + 335 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 72, + 555, + 335 + ], + "type": "text", + "content": " is experimentally set to a small integer, such as 3, 7; we will report analysis on this parameters in experiments part. Note that the relative positioning between the latents and the audio prompts has not been altered. The only modification occurs in their recombination during the input phase to the model. In a special case, at the end of the long sequence, the shifted position index may exceed the sequence length. To address this issue, we adopt a circular padding strategy, padding the start audio prompt and latents to the end." + } + ] + } + ], + "index": 7 + }, + { + "type": "code", + "bbox": [ + 315, + 359, + 555, + 596 + ], + "blocks": [ + { + "bbox": [ + 315, + 343, + 505, + 354 + ], + "lines": [ + { + "bbox": [ + 315, + 343, + 505, + 354 + ], + "spans": [ + { + "bbox": [ + 315, + 343, + 505, + 354 + ], + "type": "text", + "content": "Algorithm 1 Time-aware Position-shift Fusion" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 315, + 359, + 555, + 596 + ], + "lines": [ + { + "bbox": [ + 315, + 359, + 555, + 596 + ], + "spans": [ + { + "bbox": [ + 315, + 359, + 555, + 596 + ], + "type": "text", + "content": "Require: Audio embedding " + }, + { + "bbox": [ + 315, + 359, + 555, + 596 + ], + "type": "inline_equation", + "content": "c_{a}^{[0,l]}" + }, + { + "bbox": [ + 315, + 359, + 555, + 596 + ], + "type": "text", + "content": " with length " + }, + { + "bbox": [ + 315, + 359, + 555, + 596 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 315, + 359, + 555, + 596 + ], + "type": "text", + "content": ", denoising steps " + }, + { + "bbox": [ + 315, + 359, + 555, + 596 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 315, + 359, + 555, + 596 + ], + "type": "text", + "content": ", initial noisy latent " + }, + { + "bbox": [ + 315, + 359, + 555, + 596 + ], + "type": "inline_equation", + "content": "z_{T}^{[0,l]}" + }, + { + "bbox": [ + 315, + 359, + 555, + 596 + ], + "type": "text", + "content": ", pretrained Sonic model " + }, + { + "bbox": [ + 315, + 359, + 555, + 596 + ], + "type": "inline_equation", + "content": "\\mathrm{ST}(\\cdot)" + }, + { + "bbox": [ + 315, + 359, + 555, + 596 + ], + "type": "text", + "content": " for sequence length " + }, + { + "bbox": [ + 315, + 359, + 555, + 596 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 315, + 359, + 555, + 596 + ], + "type": "text", + "content": ", position-shift offset " + }, + { + "bbox": [ + 315, + 359, + 555, + 596 + ], + "type": "inline_equation", + "content": "\\alpha < f < l" + }, + { + "bbox": [ + 315, + 359, + 555, + 596 + ], + "type": "text", + "content": ". \nEnsure: Denoised latent " + }, + { + "bbox": [ + 315, + 359, + 555, + 596 + ], + "type": "inline_equation", + "content": "z_{0}^{[0,l]}" + }, + { + "bbox": [ + 315, + 359, + 555, + 596 + ], + "type": "text", + "content": ". \n1: Initialize accumulated shift offset " + }, + { + "bbox": [ + 315, + 359, + 555, + 596 + ], + "type": "inline_equation", + "content": "\\alpha_{\\Sigma} = 0" + }, + { + "bbox": [ + 315, + 359, + 555, + 596 + ], + "type": "text", + "content": ". \n2: for " + }, + { + "bbox": [ + 315, + 359, + 555, + 596 + ], + "type": "inline_equation", + "content": "t = T, \\dots, 1" + }, + { + "bbox": [ + 315, + 359, + 555, + 596 + ], + "type": "text", + "content": " do \n3: // Denoising loop \n4: Initialize start point " + }, + { + "bbox": [ + 315, + 359, + 555, + 596 + ], + "type": "inline_equation", + "content": "s = \\alpha_{\\Sigma}" + }, + { + "bbox": [ + 315, + 359, + 555, + 596 + ], + "type": "text", + "content": ", end " + }, + { + "bbox": [ + 315, + 359, + 555, + 596 + ], + "type": "inline_equation", + "content": "e = s + f" + }, + { + "bbox": [ + 315, + 359, + 555, + 596 + ], + "type": "text", + "content": ", processed length " + }, + { + "bbox": [ + 315, + 359, + 555, + 596 + ], + "type": "inline_equation", + "content": "n = 0" + }, + { + "bbox": [ + 315, + 359, + 555, + 596 + ], + "type": "text", + "content": ". // start from new position for each timestep. \n5: while " + }, + { + "bbox": [ + 315, + 359, + 555, + 596 + ], + "type": "inline_equation", + "content": "n < l" + }, + { + "bbox": [ + 315, + 359, + 555, + 596 + ], + "type": "text", + "content": " do \n6: // Sequence loop \n7: " + }, + { + "bbox": [ + 315, + 359, + 555, + 596 + ], + "type": "inline_equation", + "content": "z_{t-1}^{[s,e]} = \\mathrm{ST}(z_{t}^{[s,e]}, c_{a}^{[s,e]}, t)" + }, + { + "bbox": [ + 315, + 359, + 555, + 596 + ], + "type": "text", + "content": " \n8: " + }, + { + "bbox": [ + 315, + 359, + 555, + 596 + ], + "type": "inline_equation", + "content": "s \\gets s + f, e \\gets e + f, n \\gets n + f" + }, + { + "bbox": [ + 315, + 359, + 555, + 596 + ], + "type": "text", + "content": ". // Move to next clip non-overlapping \n9: if " + }, + { + "bbox": [ + 315, + 359, + 555, + 596 + ], + "type": "inline_equation", + "content": "s > l" + }, + { + "bbox": [ + 315, + 359, + 555, + 596 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 315, + 359, + 555, + 596 + ], + "type": "inline_equation", + "content": "e > l" + }, + { + "bbox": [ + 315, + 359, + 555, + 596 + ], + "type": "text", + "content": " then \n10: " + }, + { + "bbox": [ + 315, + 359, + 555, + 596 + ], + "type": "inline_equation", + "content": "s \\gets s\\% l, e \\gets e\\% l" + }, + { + "bbox": [ + 315, + 359, + 555, + 596 + ], + "type": "text", + "content": ". // Padding circularly \n11: end if \n12: end while \n13: " + }, + { + "bbox": [ + 315, + 359, + 555, + 596 + ], + "type": "inline_equation", + "content": "\\alpha_{\\Sigma} \\gets \\alpha_{\\Sigma} + \\alpha" + }, + { + "bbox": [ + 315, + 359, + 555, + 596 + ], + "type": "text", + "content": ". // Accumulate shift offset \n14: end for \n15: return Denoised latent " + }, + { + "bbox": [ + 315, + 359, + 555, + 596 + ], + "type": "inline_equation", + "content": "z_{0}^{[0,l]}" + }, + { + "bbox": [ + 315, + 359, + 555, + 596 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "code_body" + } + ], + "index": 9, + "sub_type": "algorithm" + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "content": "Computation cost. We analyze the computation efficiency of our time-aware position shift fusion compared to motion frames and overlapping. Denote the FLOPs of a single forward pass of the video diffusion model as " + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\Omega" + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "content": ", the number of frames to generate as " + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "inline_equation", + "content": "n \\times f" + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "content": ", and the denoising steps as " + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "content": ". The accumulated FLOPs of Sonic is: " + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\Omega(Shift) = \\Omega \\times T \\times n" + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "content": ". For overlapping, the overlapped latents (overlap " + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "inline_equation", + "content": "o" + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "content": ") are calculated twice, " + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\Omega(Overlap) = \\Omega \\times T \\times \\left(n + \\frac{o}{f}\\right)" + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "content": ". For motion frames, the FLOPs of a single forward of Ref-" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "197" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 83, + 96, + 526, + 242 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 93 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 93 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 93 + ], + "type": "text", + "content": "Table 1. Quantitative comparisons with the state-of-the-arts on the HDTF and CelebV-HQ test set. * indicates GAN-based methods and others are Diffusion-based methods. The best results are bold, and the second are underlined." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 83, + 96, + 526, + 242 + ], + "lines": [ + { + "bbox": [ + 83, + 96, + 526, + 242 + ], + "spans": [ + { + "bbox": [ + 83, + 96, + 526, + 242 + ], + "type": "table", + "html": "
DatasetMethodFID↓FVD↓Sync-C↑Sync-D↓E-FID↓F-SIM↑Smooth↑Runtime↓
HDTFSadTalker* (2023) [50]61.672397.1141.75510.6952.4820.92870.99613.75
Aniportrait (2024) [42]36.965471.4521.09512.4613.1610.95080.992144.03
V-Express (2024) [40]47.396758.0231.25612.3942.6340.90930.996839.04
Hallo (2024) [44]30.176347.3584.0609.5511.7920.95550.994174.65
Hallo2 (2024) [12]38.673328.5404.1369.4652.2030.96060.994245.75
EchoMimic (2024) [8]33.207384.3042.51410.7431.4860.95270.99345.45
Sonic (Ours)29.104301.1734.1979.3711.7450.95950.997017.04
CelebV-HQSadTalker* (2024) [50]57.574841.9621.97810.9152.2520.94340.99593.75
Aniportrait (2024) [42]53.746590.3730.99612.0843.2960.95220.991144.03
V-Express (2024) [40]65.400889.9850.80913.2552.7130.90190.996539.04
Hallo (2024) [44]47.403488.4992.68010.2922.2730.96070.994274.65
Hallo2 (2024) [12]52.396481.3362.63810.3432.8190.95870.994245.75
EchoMimic (2024) [8]48.267596.8701.94910.7542.1360.96120.99325.45
Sonic (Ours)43.137483.1082.68910.1941.7830.96240.997217.04
", + "image_path": "26b26caeb9ca3b21b294c9ab03ed7d14a80bd06669eeac0e8c5a16c1cffad260.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 58, + 254, + 288, + 492 + ], + "blocks": [ + { + "bbox": [ + 58, + 254, + 288, + 492 + ], + "lines": [ + { + "bbox": [ + 58, + 254, + 288, + 492 + ], + "spans": [ + { + "bbox": [ + 58, + 254, + 288, + 492 + ], + "type": "image", + "image_path": "07353588234a301c7bb4c2ad734f4be157930330510d866ef8192f839e350115.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 500, + 555, + 536 + ], + "lines": [ + { + "bbox": [ + 55, + 500, + 555, + 536 + ], + "spans": [ + { + "bbox": [ + 55, + 500, + 555, + 536 + ], + "type": "text", + "content": "Figure 4. Qualitative comparisons with State-of-the-Art talking head generation methods. Due to image does not reflect important sync, naturalness and stability, the full video comparison will be included in supplementary materials as well as comparison with demos from other non-open source works." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 291, + 254, + 553, + 492 + ], + "blocks": [ + { + "bbox": [ + 291, + 254, + 553, + 492 + ], + "lines": [ + { + "bbox": [ + 291, + 254, + 553, + 492 + ], + "spans": [ + { + "bbox": [ + 291, + 254, + 553, + 492 + ], + "type": "image", + "image_path": "67196a5f3134c1acae4602dea6d68ca40d7b1b00322cb8a9c55662c610341211.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 546, + 295, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 546, + 295, + 609 + ], + "spans": [ + { + "bbox": [ + 55, + 546, + 295, + 609 + ], + "type": "text", + "content": "erence Net is " + }, + { + "bbox": [ + 55, + 546, + 295, + 609 + ], + "type": "inline_equation", + "content": "\\omega_{r}" + }, + { + "bbox": [ + 55, + 546, + 295, + 609 + ], + "type": "text", + "content": ", the number of motion frames is " + }, + { + "bbox": [ + 55, + 546, + 295, + 609 + ], + "type": "inline_equation", + "content": "o" + }, + { + "bbox": [ + 55, + 546, + 295, + 609 + ], + "type": "text", + "content": " and the Flops of motion modules computing " + }, + { + "bbox": [ + 55, + 546, + 295, + 609 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 55, + 546, + 295, + 609 + ], + "type": "text", + "content": " frames is " + }, + { + "bbox": [ + 55, + 546, + 295, + 609 + ], + "type": "inline_equation", + "content": "\\omega_{m}" + }, + { + "bbox": [ + 55, + 546, + 295, + 609 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 546, + 295, + 609 + ], + "type": "inline_equation", + "content": "\\Omega (MF) = \\Omega \\times T\\times n + \\omega_{r}\\times o\\times n + \\omega_{m}\\times T\\times n\\times \\frac{2of + o^{2}}{f^{2}}" + }, + { + "bbox": [ + 55, + 546, + 295, + 609 + ], + "type": "text", + "content": ". Overall, the proposed time-aware position shift fusion significantly reduces the computation complexity." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 624, + 137, + 638 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 624, + 137, + 638 + ], + "spans": [ + { + "bbox": [ + 55, + 624, + 137, + 638 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 645, + 167, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 645, + 167, + 658 + ], + "spans": [ + { + "bbox": [ + 55, + 645, + 167, + 658 + ], + "type": "text", + "content": "4.1. Experiments Setup" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "content": "Implementation Details. The Sonic training is in a single stage, with initialization of the spatial module and temporal modules from stable-video-diffusion-xt-1-1[3]. To separately enable different conditions in training, we manipulate" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 546, + 555, + 618 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 546, + 555, + 618 + ], + "spans": [ + { + "bbox": [ + 313, + 546, + 555, + 618 + ], + "type": "text", + "content": "the data such that " + }, + { + "bbox": [ + 313, + 546, + 555, + 618 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 313, + 546, + 555, + 618 + ], + "type": "text", + "content": " of it drops audio, another " + }, + { + "bbox": [ + 313, + 546, + 555, + 618 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 313, + 546, + 555, + 618 + ], + "type": "text", + "content": " drops image, and " + }, + { + "bbox": [ + 313, + 546, + 555, + 618 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 313, + 546, + 555, + 618 + ], + "type": "text", + "content": " drops both condition. In inference phase, multi-condition CFG is performed. The CFG of the reference image " + }, + { + "bbox": [ + 313, + 546, + 555, + 618 + ], + "type": "inline_equation", + "content": "r_i" + }, + { + "bbox": [ + 313, + 546, + 555, + 618 + ], + "type": "text", + "content": " is set to 2.0, and the CFG of the audio " + }, + { + "bbox": [ + 313, + 546, + 555, + 618 + ], + "type": "inline_equation", + "content": "r_a" + }, + { + "bbox": [ + 313, + 546, + 555, + 618 + ], + "type": "text", + "content": " is set to 7.5. In our full model, the dynamic scale " + }, + { + "bbox": [ + 313, + 546, + 555, + 618 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 313, + 546, + 555, + 618 + ], + "type": "text", + "content": " is set to 1.0 and the shift offset " + }, + { + "bbox": [ + 313, + 546, + 555, + 618 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 546, + 555, + 618 + ], + "type": "text", + "content": " is set to 7." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "content": "Datasets and Evaluation Metrics. For the training set, we created a diverse dataset by aggregating data from open-source datasets, including VFHQ [43], CelebV-Text [48], VoxCeleb2 [10]. For the test set, we conducted experiments on two widely used datasets, HDTF [52] and CelebV-HQ [51]. We randomly selected 50 four-second clips from CelebV-HQ and 100 from HDTF. We utilize seven met" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "198" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 81, + 70, + 526, + 286 + ], + "blocks": [ + { + "bbox": [ + 81, + 70, + 526, + 286 + ], + "lines": [ + { + "bbox": [ + 81, + 70, + 526, + 286 + ], + "spans": [ + { + "bbox": [ + 81, + 70, + 526, + 286 + ], + "type": "image", + "image_path": "760a4446925b36607c2b04ee26f8c7dbc5ecac4b05fc2dfcdcc9805be274a672.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 289, + 555, + 332 + ], + "lines": [ + { + "bbox": [ + 55, + 289, + 555, + 332 + ], + "spans": [ + { + "bbox": [ + 55, + 289, + 555, + 332 + ], + "type": "text", + "content": "Figure 5. Qualitative results under different styles of portrait images and various types of audio inputs. The images and audios were collected from recent works and the Internet. The upper section presents results for audio input ranging from 20 seconds to 2 minutes in duration, while the lower section shows results for longer audio inputs, up to 10 minutes. Our Sonic demonstrates versatility across various portrait styles and maintains vividness over extended durations." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 334, + 296, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 334, + 296, + 453 + ], + "spans": [ + { + "bbox": [ + 55, + 334, + 296, + 453 + ], + "type": "text", + "content": "rics to evaluate. The Fréchet Inception Distance (FID) [18] and Fréchet Video Distance (FVD) [38] metrics measure the quality of the generated data. Expression-FID (E-FID) [37] uses face reconstruction method [13] to extract features and calculate the FID between generated and ground truth frames, while F-SIM measures facial similarity. SyncNet (Sync-C&Sync-D) [9] and VBench's smooth metric [21] assess lip synchronization and motion fluidity. In ablation experiments, we employ VBench's dynamic degree metric to evaluate amplitude of head movements." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 465, + 175, + 478 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 465, + 175, + 478 + ], + "spans": [ + { + "bbox": [ + 55, + 465, + 175, + 478 + ], + "type": "text", + "content": "4.2. Results and Analysis" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 484, + 295, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 484, + 295, + 676 + ], + "spans": [ + { + "bbox": [ + 55, + 484, + 295, + 676 + ], + "type": "text", + "content": "Quantitative Comparison. In Table 1, our Sonic are significantly superior than other recent 6 methods in terms of FID, Sync-confidence and Smoothness evaluated on the CelebV-HQ and HDTF datasets. On the HDTF dataset, our Sonic achieve the lowest FID and FVD scores, " + }, + { + "bbox": [ + 55, + 484, + 295, + 676 + ], + "type": "inline_equation", + "content": "8\\%" + }, + { + "bbox": [ + 55, + 484, + 295, + 676 + ], + "type": "text", + "content": " lower than the second-best, indicating superior generation quality and alignment with real data. For the CelebV-HQ dataset, which features more complex characters and backgrounds, our method maintains superior performance across most metrics and competitive FVD. We notably excel in E-FID, " + }, + { + "bbox": [ + 55, + 484, + 295, + 676 + ], + "type": "inline_equation", + "content": "16.5\\%" + }, + { + "bbox": [ + 55, + 484, + 295, + 676 + ], + "type": "text", + "content": " lower than the second-best, which demonstrates our method has superior expression diversity. EMO [37] as a representative work that tackles talking head generation task, remains closed source, and not available for fair quantitative comparison. We provide video results using input from their online demos for a qualitative comparison." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "Qualitative Comparison. Figure 4 provides a visualization comparison on open datasets. Analyzing the results, Previous approaches relied on motion frames, limiting the rich" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 334, + 555, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 334, + 555, + 417 + ], + "spans": [ + { + "bbox": [ + 313, + 334, + 555, + 417 + ], + "type": "text", + "content": "ness of generated expressions and the dynamism of head movements. In contrast, our proposed framework incorporates global audio perception, enhancing the integration of audio and temporal variables while providing a deeper understanding of the audio input. This enables our method to generate a broader range of expressions that align with the audio and facilitate more natural head movements." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 418, + 556, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 418, + 556, + 528 + ], + "spans": [ + { + "bbox": [ + 313, + 418, + 556, + 528 + ], + "type": "text", + "content": "Visualization Results in Complex Scenarios. We further explore the generation performance of our method under different input conditions. In terms of images, we employ three styles: real human, animation, and 3D. Each character portrait is paired with various types of audio relevant to its theme, such as singing, speech, and rap. As shown in Figure 5, our method successfully generates a wide array of expressions and diverse movements across various categories of input conditions, demonstrating strong robustness." + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 316, + 552, + 549, + 600 + ], + "blocks": [ + { + "bbox": [ + 345, + 540, + 522, + 551 + ], + "lines": [ + { + "bbox": [ + 345, + 540, + 522, + 551 + ], + "spans": [ + { + "bbox": [ + 345, + 540, + 522, + 551 + ], + "type": "text", + "content": "Table 2. User study comparison on open dataset." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 316, + 552, + 549, + 600 + ], + "lines": [ + { + "bbox": [ + 316, + 552, + 549, + 600 + ], + "spans": [ + { + "bbox": [ + 316, + 552, + 549, + 600 + ], + "type": "table", + "html": "
Method / MetricLip syncMotion diversityID consistencyVideo Smoothness
Aniportrait1.421.623.112.09
SadTalker1.982.342.952.95
Echomimic2.772.653.482.71
Hallo23.152.373.342.94
Sonic(Ours)4.58 (45%↑)4.55 (72%↑)4.29 (23%↑)4.66 (58%↑)
", + "image_path": "c5ca55a57343619a0c23fb2ff996c3338117959f7fc90872a3079ce20028f460.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "content": "User Studies We conducted subjective evaluation on open dataset to assess the comparative methods along four key dimensions: Lip sync, motion diversity, identity similarity, and video smoothness. Total 40 participants gives scores from 1 to 5 for results of five comparative methods. In Table 2, our Sonic outperforms the others in all four dimensions, especially showing a significant " + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "inline_equation", + "content": "72\\%" + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "content": " improvement in motion diversity. The video results for user study will be included in the supplementary materials." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "199" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 67, + 71, + 282, + 208 + ], + "blocks": [ + { + "bbox": [ + 67, + 71, + 282, + 208 + ], + "lines": [ + { + "bbox": [ + 67, + 71, + 282, + 208 + ], + "spans": [ + { + "bbox": [ + 67, + 71, + 282, + 208 + ], + "type": "image", + "image_path": "9f65a05c834be73952ef0238eac9444ab11f66f211836806959bc12ea4be9bc7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 56, + 209, + 295, + 229 + ], + "lines": [ + { + "bbox": [ + 56, + 209, + 295, + 229 + ], + "spans": [ + { + "bbox": [ + 56, + 209, + 295, + 229 + ], + "type": "text", + "content": "Figure 6. Visual results for ablation on temporal audio attention and motion controller." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 65, + 230, + 286, + 361 + ], + "blocks": [ + { + "bbox": [ + 65, + 230, + 286, + 361 + ], + "lines": [ + { + "bbox": [ + 65, + 230, + 286, + 361 + ], + "spans": [ + { + "bbox": [ + 65, + 230, + 286, + 361 + ], + "type": "image", + "image_path": "8887c8f1533892c2fd1904150b799428ccb6c80fc1d53ec83dde211bc79b5735.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 56, + 363, + 295, + 384 + ], + "lines": [ + { + "bbox": [ + 56, + 363, + 295, + 384 + ], + "spans": [ + { + "bbox": [ + 56, + 363, + 295, + 384 + ], + "type": "text", + "content": "Figure 7. Demonstration on effect of time-aware position shift fusion." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 57, + 423, + 293, + 490 + ], + "blocks": [ + { + "bbox": [ + 56, + 390, + 295, + 422 + ], + "lines": [ + { + "bbox": [ + 56, + 390, + 295, + 422 + ], + "spans": [ + { + "bbox": [ + 56, + 390, + 295, + 422 + ], + "type": "text", + "content": "Table 3. Quantitative results of ablations on CelebV-HQ along with our results obtained using different dynamic scales. In the Sonic model, the default setting for " + }, + { + "bbox": [ + 56, + 390, + 295, + 422 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 56, + 390, + 295, + 422 + ], + "type": "text", + "content": " is 1.0, i.e. Ours-Moderate." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 423, + 293, + 490 + ], + "lines": [ + { + "bbox": [ + 57, + 423, + 293, + 490 + ], + "spans": [ + { + "bbox": [ + 57, + 423, + 293, + 490 + ], + "type": "table", + "html": "
Method / MetricSync-C↑Sync-D↓E-FID↓Smooth↑Dynamic↑
w/o temporal-audio2.61010.3102.3110.99690.62
w/o motion-controller2.19310.7313.2310.99690.22
w/o shift-fusion2.50910.3792.2020.99690.76
Ours-Mild (β = 0.5)2.65510.2422.0600.99700.78
Ours-Intense (β = 2.0)2.80310.1252.0380.99690.98
Ours-Moderate (β = 1.0)2.68910.1941.7830.99720.86
", + "image_path": "da278e5b21cc667bb7fe03becd4931751acbb0c07a87146bda4b95be82e4554b.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 62, + 518, + 285, + 567 + ], + "blocks": [ + { + "bbox": [ + 56, + 495, + 295, + 517 + ], + "lines": [ + { + "bbox": [ + 56, + 495, + 295, + 517 + ], + "spans": [ + { + "bbox": [ + 56, + 495, + 295, + 517 + ], + "type": "text", + "content": "Table 4. Quantitative comparison on motion frames, overlap and the proposed time-aware shift fusion." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 62, + 518, + 285, + 567 + ], + "lines": [ + { + "bbox": [ + 62, + 518, + 285, + 567 + ], + "spans": [ + { + "bbox": [ + 62, + 518, + 285, + 567 + ], + "type": "table", + "html": "
Method / MetricSync-C↑Sync-D↓Smooth↑Runtime↓
Motion frames(8)2.46510.3230.996919.70
Overlap(8)2.54410.2650.996922.49
Shift-fusion2.68910.1940.997217.04
", + "image_path": "640c08b665c9969263c73cdf1d4484256725346904e553721d31c73ef26a618d.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 574, + 156, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 574, + 156, + 586 + ], + "spans": [ + { + "bbox": [ + 56, + 574, + 156, + 586 + ], + "type": "text", + "content": "4.3. Ablation Studies" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "text", + "content": "Temporal Control Inspires Intra-clip Diversity. In Table 3, the absence of temporal audio attention results in a decrease in both lip-sync and E-FID. We observe that the motion controller plays a crucial role in CelebV-HQ given that its videos exhibit more movement that weakly correlates with audio. We also investigate the effect of different dynamic scale. Increasing " + }, + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "text", + "content": " can further stimulate the lip-audio synchronization and the diversity of movements, while there is a slight decrease in stability and expression similarity. The instances in Figure 6 indicate" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 318, + 77, + 553, + 202 + ], + "blocks": [ + { + "bbox": [ + 318, + 77, + 553, + 202 + ], + "lines": [ + { + "bbox": [ + 318, + 77, + 553, + 202 + ], + "spans": [ + { + "bbox": [ + 318, + 77, + 553, + 202 + ], + "type": "image", + "image_path": "b5ca64c6b30ab4751ecc4a740e92de9fc5491979d2dfd064d22c163a405fab67.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 205, + 553, + 228 + ], + "lines": [ + { + "bbox": [ + 313, + 205, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 313, + 205, + 553, + 228 + ], + "type": "text", + "content": "Figure 8. Visual results compared with motion frames and overlapping." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 242, + 555, + 278 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 242, + 555, + 278 + ], + "spans": [ + { + "bbox": [ + 313, + 242, + 555, + 278 + ], + "type": "text", + "content": "temporal audio attention helps learn effective control of micro-expressions, while the motion controller is critical for agility and expressiveness." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 281, + 555, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 281, + 555, + 461 + ], + "spans": [ + { + "bbox": [ + 313, + 281, + 555, + 461 + ], + "type": "text", + "content": "Larger Audio Perception Bridges Inter-clip Stability. Initially, we ablate the time-aware position shift fusion (briefly termed Shift) and generate clips independently as illustrated in Figure 7. The fusion process is performed at each timestep, gradually inverting the random noise to continuous and smooth long range videos. Subsequently, we explore different long video inference strategy: motion frames, overlap, our shift. The number of motion frames and overlap frames was set to 8. As indicated in Table 4, our shift approach attains superior lip-sync and smoothness without necessitating any auxiliary computation. From the qualitative comparison demonstrated in Figure 8, our method successfully handle long term consistency, circumventing identity similarity decrease in motion frames and mixed blurry texture in overlapping results." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 314, + 483, + 388, + 495 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 483, + 388, + 495 + ], + "spans": [ + { + "bbox": [ + 314, + 483, + 388, + 495 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 506, + 555, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 506, + 555, + 625 + ], + "spans": [ + { + "bbox": [ + 313, + 506, + 555, + 625 + ], + "type": "text", + "content": "In summary, our work present Sonic, an audio-driven portrait animation that focuses on the exploration of global audio perception to enhance the efficient generation of realistic lip synchronization, naturalness and temporal consistencies. The technical cornerstone mainly lies on context-enhanced audio learning and time-aware position shift fusion that work together to extend the intra-clip audio temporal to the global inter-clip audio perception. It significantly outperforms existing SOTA methods in video quality, motion diversity and naturalness." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 630, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 556, + 713 + ], + "type": "text", + "content": "Acknowledgements We thank Dongshan Li, Zixiang Zhou, Zhentao Yu, Hongmei Wang, Joeyu Wang, Chengfei Cai, Shiyu Tang, Tianxiang Zheng, Junshu Tang, Shuai Shao, Yubin Zeng, Junxin Huang, Yue Zhang, Chao Zhan, Zhengyan Tong, Zhaokang Chen, Bin Wu, Wenjiang Zhou, Yang Zhang, Junqi Cheng, Jiaxi Gu, and Fangyuan Zou for helpful discussion." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "text", + "content": "200" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "type": "text", + "content": "[1] Alexei Baevski, Yuhao Zhou, Abdelrahman Mohamed, and Michael Auli. wav2vec 2.0: A framework for self-supervised learning of speech representations. Advances in neural information processing systems, 33:12449-12460, 2020. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 137, + 296, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 137, + 296, + 192 + ], + "spans": [ + { + "bbox": [ + 61, + 137, + 296, + 192 + ], + "type": "text", + "content": "[2] Omer Bar-Tal, Hila Chefer, Omer Tov, Charles Herrmann, Roni Paiss, Shiran Zada, Ariel Ephrat, Junhwa Hur, Yuanzhen Li, Tomer Michaeli, et al. Lumiere: A spacetime diffusion model for video generation. arXiv preprint arXiv:2401.12945, 2024. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 194, + 295, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 194, + 295, + 248 + ], + "spans": [ + { + "bbox": [ + 62, + 194, + 295, + 248 + ], + "type": "text", + "content": "[3] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 3, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 250, + 295, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 250, + 295, + 315 + ], + "spans": [ + { + "bbox": [ + 62, + 250, + 295, + 315 + ], + "type": "text", + "content": "[4] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22563-22575, 2023. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 318, + 295, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 318, + 295, + 392 + ], + "spans": [ + { + "bbox": [ + 62, + 318, + 295, + 392 + ], + "type": "text", + "content": "[5] Aras Bozkurt, Xiao Junhong, Sarah Lambert, Angelica Pazurek, Helen Crompton, Suzan Koseoglu, Robert Farrow, Melissa Bond, Chrissi Nerantzi, Sarah Honeychurch, et al. Speculative futures on chatgpt and generative artificial intelligence (ai): A collective reflection from the educational landscape. Asian Journal of Distance Education, 18(1):53-130, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 396, + 295, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 396, + 295, + 460 + ], + "spans": [ + { + "bbox": [ + 62, + 396, + 295, + 460 + ], + "type": "text", + "content": "[6] Haoxin Chen, Yong Zhang, Xiaodong Cun, Menghan Xia, Xintao Wang, Chao Weng, and Ying Shan. Videocrafter2: Overcoming data limitations for high-quality video diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7310-7320, 2024. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 463, + 295, + 507 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 463, + 295, + 507 + ], + "spans": [ + { + "bbox": [ + 62, + 463, + 295, + 507 + ], + "type": "text", + "content": "[7] Lele Chen, Zhiheng Li, Ross K Maddox, Zhiyao Duan, and Chenliang Xu. Lip movements generation at a glance. In Proceedings of the European conference on computer vision (ECCV), pages 520-535, 2018. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 509, + 295, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 509, + 295, + 552 + ], + "spans": [ + { + "bbox": [ + 62, + 509, + 295, + 552 + ], + "type": "text", + "content": "[8] Zhiyuan Chen, Jiajiong Cao, Zhiquan Chen, Yuming Li, and Chenguang Ma. Echomimic: Lifelike audio-driven portrait animations through editable landmark conditions. arXiv preprint arXiv:2407.08136, 2024. 2, 3, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 555, + 295, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 555, + 295, + 609 + ], + "spans": [ + { + "bbox": [ + 62, + 555, + 295, + 609 + ], + "type": "text", + "content": "[9] Joon Son Chung and Andrew Zisserman. Out of time: automated lip sync in the wild. In Computer Vision-ACCV 2016 Workshops: ACCV 2016 International Workshops, Taipei, Taiwan, November 20-24, 2016, Revised Selected Papers, Part II 13, pages 251-263. Springer, 2017. 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 611, + 295, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 611, + 295, + 643 + ], + "spans": [ + { + "bbox": [ + 57, + 611, + 295, + 643 + ], + "type": "text", + "content": "[10] Joon Son Chung, Arsha Nagrani, and Andrew Zisserman. Voxceleb2: Deep speaker recognition. arXiv preprint arXiv:1806.05622, 2018. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 646, + 295, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 646, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 57, + 646, + 295, + 689 + ], + "type": "text", + "content": "[11] Enric Corona, Andrei Zanfir, Eduard Gabriel Bazavan, Nikos Kolotouros, Thiemo Alldieck, and Cristian Sminchisescu. Vlogger: Multimodal diffusion for embodied avatar synthesis. arXiv preprint arXiv:2403.08764, 2024. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 691, + 295, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 295, + 714 + ], + "type": "text", + "content": "[12] Jiahao Cui, Hui Li, Yao Yao, Hao Zhu, Hanlin Shang, Kaihui Cheng, Hang Zhou, Siyu Zhu, and Jingdong Wang. Hallo2:" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 714 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "Long-duration and high-resolution audio-driven portrait image animation. arXiv preprint arXiv:2410.07718, 2024. 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 97, + 553, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 97, + 553, + 162 + ], + "spans": [ + { + "bbox": [ + 316, + 97, + 553, + 162 + ], + "type": "text", + "content": "[13] Yu Deng, Jiaolong Yang, Sicheng Xu, Dong Chen, Yunde Jia, and Xin Tong. Accurate 3d face reconstruction with weakly-supervised learning: From single image to image set. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 0–0, 2019. 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 163, + 553, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 553, + 218 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 553, + 218 + ], + "type": "text", + "content": "[14] Patrick Esser, Johnathan Chiu, Parmida Atighehchian, Jonathan Granskog, and Anastasis Germanidis. Structure and content-guided video synthesis with diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7346-7356, 2023. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 220, + 553, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 220, + 553, + 275 + ], + "spans": [ + { + "bbox": [ + 316, + 220, + 553, + 275 + ], + "type": "text", + "content": "[15] Yuwei Guo, Ceyuan Yang, Anyi Rao, Zhengyang Liang, Yaohui Wang, Yu Qiao, Maneesh Agrawala, Dahua Lin, and Bo Dai. Animatediff: Animate your personalized text-to-image diffusion models without specific tuning. International Conference on Learning Representations, 2024. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 276, + 553, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 276, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 316, + 276, + 553, + 319 + ], + "type": "text", + "content": "[16] Tianyu He, Junliang Guo, Runyi Yu, Yuchi Wang, Jialiang Zhu, Kaikai An, Leyi Li, Xu Tan, Chunyu Wang, Han Hu, et al. Gaia: Zero-shot talking avatar generation. arXiv preprint arXiv:2311.15230, 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 321, + 553, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 321, + 553, + 364 + ], + "spans": [ + { + "bbox": [ + 316, + 321, + 553, + 364 + ], + "type": "text", + "content": "[17] Yingqing He, Tianyu Yang, Yong Zhang, Ying Shan, and Qifeng Chen. Latent video diffusion models for high-fidelity long video generation. arXiv preprint arXiv:2211.13221, 2022.3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 365, + 553, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 365, + 553, + 420 + ], + "spans": [ + { + "bbox": [ + 316, + 365, + 553, + 420 + ], + "type": "text", + "content": "[18] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 422, + 553, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 422, + 553, + 465 + ], + "spans": [ + { + "bbox": [ + 316, + 422, + 553, + 465 + ], + "type": "text", + "content": "[19] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. Advances in Neural Information Processing Systems, 35:8633-8646, 2022. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 467, + 553, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 467, + 553, + 510 + ], + "spans": [ + { + "bbox": [ + 316, + 467, + 553, + 510 + ], + "type": "text", + "content": "[20] Wenyi Hong, Ming Ding, Wendi Zheng, Xinghan Liu, and Jie Tang. Cogvideo: Large-scale pretraining for text-to-video generation via transformers. arXiv preprint arXiv:2205.15868, 2022. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 512, + 553, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 512, + 553, + 578 + ], + "spans": [ + { + "bbox": [ + 316, + 512, + 553, + 578 + ], + "type": "text", + "content": "[21] Ziqi Huang, Yinan He, Jiashuo Yu, Fan Zhang, Chenyang Si, Yuming Jiang, Yuanhan Zhang, Tianxing Wu, Qingyang Jin, Nattapol Chanpaisit, et al. Vbench: Comprehensive benchmark suite for video generative models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21807-21818, 2024. 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 579, + 553, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 579, + 553, + 643 + ], + "spans": [ + { + "bbox": [ + 316, + 579, + 553, + 643 + ], + "type": "text", + "content": "[22] Xiaozhong Ji, Chuming Lin, Zhonggan Ding, Ying Tai, Jian Yang, Junwei Zhu, Xiaobin Hu, Jiangning Zhang, Donghao Luo, and Chengjie Wang. Realtalk: Real-time and realistic audio-driven face generation with 3d facial prior-guided identity alignment network. arXiv preprint arXiv:2406.18284, 2024. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 646, + 553, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 646, + 553, + 690 + ], + "spans": [ + { + "bbox": [ + 316, + 646, + 553, + 690 + ], + "type": "text", + "content": "[23] Jianwen Jiang, Chao Liang, Jiaqi Yang, Gaojie Lin, Tianyun Zhong, and Yanbo Zheng. Loopy: Taming audio-driven portrait avatar with long-term motion dependency. arXiv preprint arXiv:2409.02634, 2024. 2, 3, 4, 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 691, + 553, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 714 + ], + "type": "text", + "content": "[24] Esperanza Johnson, Ramón Hervás, Carlos Gutiérrez López de la Franca, Tania Mondéjar, Sergio F Ochoa, and Jesús" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "201" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 76, + 73, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 73, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 76, + 73, + 294, + 106 + ], + "type": "text", + "content": "Favela. Assessing empathy and managing emotions through interactions with an affective avatar. Health informatics journal, 24(2):182-193, 2018. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 294, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 294, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 294, + 140 + ], + "type": "text", + "content": "[25] Oytun Kal and Yavuz Samur. Educational virtual reality game design for film and animation. Encyclopedia of Computer Graphics and Games, pages 621-636, 2024. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 141, + 294, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 141, + 294, + 207 + ], + "spans": [ + { + "bbox": [ + 56, + 141, + 294, + 207 + ], + "type": "text", + "content": "[26] Levon Khachatryan, Andranik Movsisyan, Vahram Tadevosyan, Roberto Henschel, Zhangyang Wang, Shant Navasardyan, and Humphrey Shi. Text2video-zero: Text-to-image diffusion models are zero-shot video generators. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15954-15964, 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 209, + 294, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 209, + 294, + 252 + ], + "spans": [ + { + "bbox": [ + 56, + 209, + 294, + 252 + ], + "type": "text", + "content": "[27] Tao Liu, Feilong Chen, Shuai Fan, Chenpeng Du, Qi Chen, Xie Chen, and Kai Yu. Anitalker: Animate vivid and diverse talking faces through identity-decoupled facial motion encoding. arXiv preprint arXiv:2405.03121, 2024. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 254, + 294, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 254, + 294, + 297 + ], + "spans": [ + { + "bbox": [ + 56, + 254, + 294, + 297 + ], + "type": "text", + "content": "[28] Yifeng Ma, Shiwei Zhang, Jiayu Wang, Xiang Wang, Yingya Zhang, and Zhidong Deng. Dreamtalk: When expressive talking head generation meets diffusion probabilistic models. arXiv preprint arXiv:2312.09767, 2023. 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 299, + 294, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 299, + 294, + 353 + ], + "spans": [ + { + "bbox": [ + 56, + 299, + 294, + 353 + ], + "type": "text", + "content": "[29] KR Prajwal, Rudrabha Mukhopadhyay, Vinay P Namboodiri, and CV Jawahar. A lip sync expert is all you need for speech to lip generation in the wild. In Proceedings of the 28th ACM international conference on multimedia, pages 484-492, 2020. 2, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 355, + 294, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 355, + 294, + 409 + ], + "spans": [ + { + "bbox": [ + 56, + 355, + 294, + 409 + ], + "type": "text", + "content": "[30] Alec Radford, Jong Wook Kim, Tao Xu, Greg Brockman, Christine McLeavey, and Ilya Sutskever. Robust speech recognition via large-scale weak supervision. In International conference on machine learning, pages 28492-28518. PMLR, 2023. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 411, + 294, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 411, + 294, + 465 + ], + "spans": [ + { + "bbox": [ + 56, + 411, + 294, + 465 + ], + "type": "text", + "content": "[31] Imogen C Rehm, Emily Foenander, Klaire Wallace, JoAnne M Abbott, Michael Kyrios, and Neil Thomas. What role can avatars play in e-mental health interventions? exploring new models of client-therapist interaction. Frontiers in Psychiatry, 7:186, 2016. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 467, + 294, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 467, + 294, + 521 + ], + "spans": [ + { + "bbox": [ + 56, + 467, + 294, + 521 + ], + "type": "text", + "content": "[32] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 523, + 294, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 523, + 294, + 578 + ], + "spans": [ + { + "bbox": [ + 56, + 523, + 294, + 578 + ], + "type": "text", + "content": "[33] Shuai Shen, Wenliang Zhao, Zibin Meng, Wanhua Li, Zheng Zhu, Jie Zhou, and Jiwen Lu. Difftalk: Crafting diffusion models for generalized audio-driven portraits animation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1982-1991, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 579, + 294, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 579, + 294, + 633 + ], + "spans": [ + { + "bbox": [ + 56, + 579, + 294, + 633 + ], + "type": "text", + "content": "[34] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792, 2022. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 635, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 294, + 689 + ], + "type": "text", + "content": "[35] Xusen Sun, Longhao Zhang, Hao Zhu, Peng Zhang, Bang Zhang, Xinya Ji, Kangneng Zhou, Daiheng Gao, Liefeng Bo, and Xun Cao. Vividtalk: One-shot audio-driven talking head generation based on 3d hybrid prior. arXiv preprint arXiv:2312.01841, 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "text", + "content": "[36] Supasorn Suwajanakorn, Steven M Seitz, and Ira Kemelmacher-Shlizerman. Synthesizing obama: learn" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "ing lip sync from audio. ACM Transactions on Graphics (ToG), 36(4):1-13, 2017. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "type": "text", + "content": "[37] Linrui Tian, Qi Wang, Bang Zhang, and Liefeng Bo. Emo: Emote portrait alive-generating expressive portrait videos with audio2video diffusion model under weak conditions. arXiv preprint arXiv:2402.17485, 2024. 2, 3, 4, 5, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 141, + 553, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 553, + 174 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 553, + 174 + ], + "type": "text", + "content": "[38] Thomas Unterthiner, Sjoerd van Steenkiste, Karol Kurach, Raphaël Marinier, Marcin Michalski, and Sylvain Gelly. Fvd: A new metric for video generation. 2019. 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 175, + 553, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 553, + 219 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 553, + 219 + ], + "type": "text", + "content": "[39] Vikram Voleti, Alexia Jolicoeur-Martineau, and Chris Pal. Mcvd-masked conditional video diffusion for prediction, generation, and interpolation. Advances in neural information processing systems, 35:23371-23385, 2022. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 220, + 553, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 220, + 553, + 275 + ], + "spans": [ + { + "bbox": [ + 316, + 220, + 553, + 275 + ], + "type": "text", + "content": "[40] Cong Wang, Kuan Tian, Jun Zhang, Yonghang Guan, Feng Luo, Fei Shen, Zhiwei Jiang, Qing Gu, Xiao Han, and Wei Yang. V-express: Conditional dropout for progressive training of portrait video generation. arXiv preprint arXiv:2406.02511, 2024. 2, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 277, + 553, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 277, + 553, + 331 + ], + "spans": [ + { + "bbox": [ + 316, + 277, + 553, + 331 + ], + "type": "text", + "content": "[41] Xiang Wang, Hangjie Yuan, Shiwei Zhang, Dayou Chen, Jiuniu Wang, Yingya Zhang, Yujun Shen, Deli Zhao, and Jingren Zhou. Videocomposer: Compositional video synthesis with motion controllability. Advances in Neural Information Processing Systems, 36, 2024. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 333, + 553, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 333, + 553, + 365 + ], + "spans": [ + { + "bbox": [ + 316, + 333, + 553, + 365 + ], + "type": "text", + "content": "[42] Huawei Wei, Zejun Yang, and Zhisheng Wang. Aniportrait: Audio-driven synthesis of photorealistic portrait animation. arXiv preprint arXiv:2403.17694, 2024. 2, 3, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 367, + 555, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 367, + 555, + 422 + ], + "spans": [ + { + "bbox": [ + 316, + 367, + 555, + 422 + ], + "type": "text", + "content": "[43] Liangbin Xie, Xintao Wang, Honglun Zhang, Chao Dong, and Ying Shan. Vfhq: A high-quality dataset and benchmark for video face super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 657-666, 2022. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 423, + 553, + 478 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 423, + 553, + 478 + ], + "spans": [ + { + "bbox": [ + 316, + 423, + 553, + 478 + ], + "type": "text", + "content": "[44] Mingwang Xu, Hui Li, Qingkun Su, Hanlin Shang, Liwei Zhang, Ce Liu, Jingdong Wang, Luc Van Gool, Yao Yao, and Siyu Zhu. Hallo: Hierarchical audio-driven visual synthesis for portrait image animation. arXiv preprint arXiv:2406.08801, 2024. 2, 3, 4, 5, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 479, + 553, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 479, + 553, + 533 + ], + "spans": [ + { + "bbox": [ + 316, + 479, + 553, + 533 + ], + "type": "text", + "content": "[45] Sicheng Xu, Guojun Chen, Yu-Xiao Guo, Jiaolong Yang, Chong Li, Zhenyu Zang, Yizhong Zhang, Xin Tong, and Baining Guo. Vasa-1: Lifelike audio-driven talking faces generated in real time. arXiv preprint arXiv:2404.10667, 2024. 2, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 535, + 553, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 535, + 553, + 601 + ], + "spans": [ + { + "bbox": [ + 316, + 535, + 553, + 601 + ], + "type": "text", + "content": "[46] Fei Yin, Yong Zhang, Xiaodong Cun, Mingdeng Cao, Yanbo Fan, Xuan Wang, Qingyan Bai, Baoyuan Wu, Jue Wang, and Yujiu Yang. Styleheat: One-shot high-resolution editable talking face generation via pre-trained stylegan. In European conference on computer vision, pages 85-101. Springer, 2022. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 602, + 553, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 602, + 553, + 656 + ], + "spans": [ + { + "bbox": [ + 316, + 602, + 553, + 656 + ], + "type": "text", + "content": "[47] Shengming Yin, Chenfei Wu, Huan Yang, Jianfeng Wang, Xiaodong Wang, Minheng Ni, Zhengyuan Yang, Linjie Li, Shuguang Liu, Fan Yang, et al. Nuwa-xl: Diffusion over diffusion for extremely long video generation. arXiv preprint arXiv:2303.12346, 2023. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 658, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 553, + 712 + ], + "type": "text", + "content": "[48] Jianhui Yu, Hao Zhu, Liming Jiang, Chen Change Loy, Weidong Cai, and Wayne Wu. Celebv-text: A large-scale facial text-video dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14805-14814, 2023. 6" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "text", + "content": "202" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 294 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "type": "text", + "content": "[49] Rui Zhang, Yixiao Fang, Zhengnan Lu, Pei Cheng, Zebiao Huang, and Bin Fu. Linguualinker: Audio-driven portraits animation with implicit facial control enhancement. arXiv preprint arXiv:2407.18595, 2024. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 118, + 295, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 118, + 295, + 183 + ], + "spans": [ + { + "bbox": [ + 55, + 118, + 295, + 183 + ], + "type": "text", + "content": "[50] Wenxuan Zhang, Xiaodong Cun, Xuan Wang, Yong Zhang, Xi Shen, Yu Guo, Ying Shan, and Fei Wang. Sadtalker: Learning realistic 3d motion coefficients for stylized audio-driven single image talking face animation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8652-8661, 2023. 2, 3, 4, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 185, + 296, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 185, + 296, + 239 + ], + "spans": [ + { + "bbox": [ + 57, + 185, + 296, + 239 + ], + "type": "text", + "content": "[51] Zhimeng Zhang, Lincheng Li, Yu Ding, and Changjie Fan. Flow-guided one-shot talking face generation with a high-resolution audio-visual dataset. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3661–3670, 2021. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 240, + 295, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 240, + 295, + 294 + ], + "spans": [ + { + "bbox": [ + 57, + 240, + 295, + 294 + ], + "type": "text", + "content": "[52] Hao Zhu, Wayne Wu, Wentao Zhu, Liming Jiang, Siwei Tang, Li Zhang, Ziwei Liu, and Chen Change Loy. Celebvyhq: A large-scale video facial attributes dataset. In European conference on computer vision, pages 650-667. Springer, 2022. 6" + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "203" + } + ] + } + ], + "index": 5 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Sound Bridge_ Associating Egocentric and Exocentric Videos via Audio Cues/37c295df-9911-4189-b46a-8795fdcf6a6f_content_list.json b/2025/Sound Bridge_ Associating Egocentric and Exocentric Videos via Audio Cues/37c295df-9911-4189-b46a-8795fdcf6a6f_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..085280ed14a4dd9ae218fb96a271588ff10e658a --- /dev/null +++ b/2025/Sound Bridge_ Associating Egocentric and Exocentric Videos via Audio Cues/37c295df-9911-4189-b46a-8795fdcf6a6f_content_list.json @@ -0,0 +1,1534 @@ +[ + { + "type": "text", + "text": "Sound Bridge: Associating Egocentric and Exocentric Videos via Audio Cues", + "text_level": 1, + "bbox": [ + 166, + 184, + 828, + 203 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Sihong Huang $^{1,5,*}$ , Jiaxin Wu $^{2,*}$ , Xiaoyong Wei $^{3,2,5,\\dagger}$ , Yi Cai $^{1,\\dagger}$ , Dongmei Jiang $^{5}$ , Yaowei Wang $^{4,5}$", + "bbox": [ + 171, + 224, + 821, + 241 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1South China University of Technology, 2The Hong Kong Polytechnic University", + "bbox": [ + 227, + 239, + 769, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{3}$ Sichuan University, $^{4}$ Harbin Institute of Technology, Shenzhen, $^{5}$ Peng Cheng Laboratory", + "bbox": [ + 197, + 256, + 799, + 272 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{fthuangsh, ycai}@mail.scut.edu.cn, jiaxwu@polyu.edu.hk, cswei@scu.edu.cn", + "bbox": [ + 223, + 272, + 772, + 286 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "jiangdm@pcl.ac.cn, wangyaowei@hit.edu.cn", + "bbox": [ + 344, + 287, + 647, + 299 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 285, + 340, + 351, + 354 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Understanding human behavior and environmental information in egocentric videos is very challenging due to the invisibility of some actions (e.g., laughing and sneezing) and the local nature of the first-person view. Leveraging the corresponding exocentric video to provide global context has shown promising results. However, existing visual-to-visual and visual-to-textual Ego-Exo video alignment methods struggle with the issue that some activities may have non-visual overlap. To address this, we propose using sound as a bridge, as audio is often consistent across Ego-Exo videos. However, direct audio-to-audio alignment lacks context. Thus, we introduce two context-aware sound modules: one aligns audio with vision via a visual-audio cross-attention module, and another aligns text with sound closed caption generated by LLM. Experimental results on two Ego-Exo video association benchmarks show that each of the proposed modules enhances the state-of-the-art methods. Moreover, the proposed sound-aware egocentric or exocentric representation boosts the performance of downstream tasks, such as action recognition of exocentric videos and scene recognition of egocentric videos. The code and models can be accessed at https://github.com/shhuangcoder/SoundBridge.", + "bbox": [ + 148, + 367, + 486, + 650 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 151, + 674, + 264, + 686 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Egocentric video analysis has been receiving extensive attention recently [10, 12, 34] due to its potential applications in robotic vision analysis, embedded AI, etc. Compared to the large-scale and in-depth studies on exocentric video analysis tasks, research on egocentric videos is still underexplored. Moreover, there are significant differences between egocentric and exocentric videos; thus, many findings and methodologies derived from an exocentric perspective are not directly applicable to egocentric video analysis.", + "bbox": [ + 148, + 695, + 486, + 819 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f55b5b9896177bd7c98e9d7433355a7627dc23fbc84ec69561270c1c2a024903.jpg", + "image_caption": [ + "Figure 1. Given a video shooting script, egocentric and exocentric views can exhibit significant visual variances for the same activity, while their audio signals remain relatively consistent." + ], + "image_footnote": [], + "bbox": [ + 511, + 328, + 844, + 470 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The main difference between egocentric and exocentric videos is that they capture events from different perspectives. First-person view videos focus on local and detailed information about the current activity, whereas third-person view videos emphasize global and environmental context. An example is shown in Figure 1. Although egocentric and exocentric videos follow the same video shooting script, they exhibit significant visual variations. For example, in the case of the same activity, such as eating a snack, the egocentric view primarily captures hand-object interactions, such as a hand picking up a raisin, while the location remains unknown. In contrast, the exocentric view captures the overall scene of the event but lacks finer details—for example, it may show a person lying on a bed eating food, but the specific type of food remains unclear. Moreover, certain activities, such as laughing and sneezing, do not produce distinct visual cues in egocentric videos apart from camera shaking. Consequently, when attempting to align visual representations from exocentric to egocentric videos, there is often no overlapping visual content to match.", + "bbox": [ + 508, + 534, + 844, + 779 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "However, recent studies [12, 34] have shown that exocentric videos can supplement the understanding of their semantically aligned egocentric counterparts. In this paper, we also focus on constructing an Ego-Exo alignment space to facilitate egocentric video analysis. Previous studies on", + "bbox": [ + 508, + 779, + 844, + 840 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation.", + "bbox": [ + 236, + 0, + 807, + 18 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "Except for this watermark, it is identical to the accepted version;", + "bbox": [ + 323, + 16, + 722, + 30 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 292, + 31, + 753, + 45 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Equal Contribution. † Corresponding Authors.", + "bbox": [ + 171, + 829, + 390, + 840 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "28942", + "bbox": [ + 478, + 914, + 519, + 926 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "egocentric and exocentric video alignment have primarily focused on matching visual content [35] or aligning both video types to an anchor (i.e., their semantic descriptions or a shared textual representation) [12, 34]. However, these alignment strategies may be insufficient due to the substantial visual variations between views. In contrast to the significant differences in view-dependent visual content, we observe that audio signals for the same event remain consistent across views. For example, in the case of laughing, the laughter sound—whether captured in an egocentric or exocentric video—can infer that someone is laughing.", + "bbox": [ + 148, + 152, + 485, + 286 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we propose using audio signals as a bridge to model the egocentric-exocentric video alignment space. However, directly linking audio signals across videos without context may be suboptimal, as audio information alone is not sufficiently discriminative [2]. For example, the sound of a click associated with the action \"switch off the light\" in Figure 1 could also occur in other activities, such as \"flipping a power switch on a household appliance\" or \"turning off a desk lamp\". To address this limitation, we propose constructing contextual audio features for alignment by incorporating visual and textual context. An overview of our proposed method is presented in Figure 2. Specifically, we introduce sound-text and sound-vision modules to bridge Ego-Exo videos. The sound-vision module follows existing methods [13, 18] to build contextual audiovisual representations through cross-attention. Meanwhile, the sound-text module utilizes a large language model (LLM) to generate contextual audio descriptions and construct sound-aware textual representations. In particular, we provide the LLM with semantic text describing the video scenario and prompt it to generate an audio description of the actions occurring within the scene in a closed-caption format [15]. Additionally, to ensure that the generated sound captions emphasize unique information within the paired cross-view videos rather than commonly occurring content in the dataset, we filter out frequently appearing and less informative concepts (e.g., person) from the generated descriptions. As a result, the output sound caption for the example in Figure 1 is \"[Eating the snack] (munching), [dishes on the floor] (crashing and scraping), [Laughing] (giggles), [leaving the room] (Footsteps), [turns off the light] (click)\". The generated sound caption and the semantic text are subsequently processed through a cross-attention block to learn a sound-aware textual representation. Finally, we treat the textual representation as an anchor and align the audiovisual representations of egocentric and exocentric videos to it, thereby learning an Ego-Exo semantic feature space. We conduct extensive experiments on two Ego-Exo video retrieval benchmarks, namely EgoExoLearn [12] and CharadesEgo [1]. The results demonstrate that the learned sound-aware features outperform state-of-the-art approaches in both egocentric-exocentric video retrieval and text-to-egocentric/exocentric video retrieval. Moreover, the learned Ego- and Exo- features enhance the performance of downstream tasks, such as exocentric video action", + "bbox": [ + 150, + 286, + 485, + 837 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "recognition and egocentric video scene recognition.", + "bbox": [ + 509, + 152, + 800, + 164 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 509, + 174, + 633, + 188 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Egocentric Video Analysis", + "text_level": 1, + "bbox": [ + 509, + 195, + 712, + 209 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The number of research studies on egocentric video analysis is growing rapidly, as this task is a fundamental component of many applications [10, 12], such as robotic vision analysis and embedded AI. However, egocentric videos are fundamentally different from traditional exocentric videos. Egocentric videos capture a unique immersive viewpoint from a first-person perspective, directly reflecting an individual's interaction with the current activity, whereas exocentric videos provide an external view of human interactions within an environment. This distinct perspective in egocentric videos introduces a range of challenges for human activity analysis, including hand detection [5, 9, 26], action recognition [8, 22, 28], human-computer interaction [16, 21, 33], and applications in virtual and augmented reality [14, 19, 29]. Despite these differences, recent studies [12, 34] have demonstrated that egocentric video analysis can benefit from the contextual information provided by corresponding exocentric videos. For example, Xu et al. [34] show that related exocentric videos can enhance egocentric video captioning. Similarly, Huang et al. [12] demonstrate that the exocentric perspective provides complementary information that aids robots in performing complex tasks. In this paper, we also focus on developing effective Ego-Exo video alignment techniques to enhance egocentric video analysis.", + "bbox": [ + 509, + 214, + 844, + 521 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Ego-Exo Video Representation Learning", + "text_level": 1, + "bbox": [ + 509, + 527, + 807, + 541 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The goal of Ego-Exo video representation learning is to construct a semantic space in which egocentric and exocentric videos with similar semantics are positioned closely, while dissimilar Ego-Exo videos are placed farther apart [12]. Early studies on Ego-Exo representation learning align cross-view videos using visual information. For example, AE2 [35] extracts visual regions of hands and active objects and performs temporal alignment on these extracted regions across views, assuming that the visual content of a fine-grained action remains consistent across perspectives. AE2 constructs four action-specific cross-view datasets for evaluation: break egg, pour milk, pour liquid, and tennis forehand. Recent approaches aim to learn a semantic Ego-Exo representation space by aligning videos to either a shared semantic text or separate textual descriptions. For example, EgoExo [12] employs implicit Ego-Exo alignment by associating each video with a view-dependent text, based on the assumption that semantically similar texts will be positioned closely in the learned space. Beyond implicit alignment, EgoExoNCE [34] takes this further by explicitly aligning Ego-Exo videos to a mutual text anchor. This shared text is generated by extracting key nouns and verbs from both the egocentric and exocentric textual descriptions. In this paper, we follow the pipeline of previous", + "bbox": [ + 509, + 547, + 844, + 843 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "28943", + "bbox": [ + 478, + 914, + 517, + 926 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "work by aligning Ego-Exo videos to a mutual text anchor. However, rather than directly matching entire videos with full-text descriptions without attention, we propose using audio as a bridge to address the challenge of visual variance. Our approach incorporates sound-vision and sound-text modules to bridge the gap between egocentric and exocentric perspectives. To the best of our knowledge, we are the first to integrate sound into Ego-Exo video alignment.", + "bbox": [ + 148, + 151, + 486, + 251 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Sound-aware Representation Learning", + "text_level": 1, + "bbox": [ + 151, + 271, + 434, + 286 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Numerous studies have demonstrated that audio signals complement visual signals in video analysis tasks [18]. For example, the Eclipse method has shown that audio signals can serve as an alternative to visual signals in long-range video retrieval, significantly accelerating the search process [18]. Consequently, many approaches have sought to develop sound-aware representations. Several works have incorporated visual information as context to create audiovisual features for exocentric video analysis, applying them to tasks such as video retrieval [13, 18], action recognition [20], and video summarization [11]. More recently, audio-based methods have also been explored in egocentric video analysis, such as active speaker spatial detection [25]. Additionally, Ibrahimi et al. [13] integrate textual context to construct audio-textual representations. Common approaches to incorporating audio signals into video analysis include appending audio features to visual or textual features [20, 25] or learning audio-textual/visual cross-attention representations [13, 18]. Another method involves converting audio signals into text using Automatic Speech Recognition (ASR) and subsequently appending the ASR-derived text features to visual features [11]. Our proposed sound-vision model follows prior work by learning contextual audiovisual features through cross-attention. Additionally, we introduce a novel approach to learning audiotextual representations in our sound-text module, which leverages a large language model (LLM) to generate sound captions.", + "bbox": [ + 148, + 290, + 486, + 633 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Sound-aligned Ego-Exo Representation Learning", + "text_level": 1, + "bbox": [ + 151, + 657, + 485, + 686 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The architecture of our proposed method is illustrated in Figure 2. The main objective is to construct a sound-aligned Ego-Exo representation space that integrates local (Ego) and global (Exo) visual features along with their corresponding semantic concepts. Specifically, to address the challenge of aligning semantic text with the significant visual variations present in Ego-Exo videos, we leverage sound as a unifying element. This is achieved by incorporating sound-vision and sound-text modules to generate sound-aware representations. Subsequently, we align the sound-aware video features with the sound-aware textual features to learn a unified representation space.", + "bbox": [ + 148, + 693, + 486, + 842 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Sound-Vision Module", + "text_level": 1, + "bbox": [ + 509, + 151, + 684, + 162 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given a semantically aligned egocentric and exocentric (ego-exo) video pair $(v^{ego}, v^{exo})$ and their corresponding audio signals $(a^{ego}, a^{exo})$ , we propose a sound-vision cross-attention module to learn their sound-aware visual representations. The sound-vision module consists of a video encoder $E_v(x)$ , an audio encoder $E_a(x)$ , and a multi-head cross-attention block Multihead $(Q, K, V)$ [27]. The module inputs a video $v$ and its corresponding audio signal $a$ and outputs sound-aware visual feature $f_{sv}$ as:", + "bbox": [ + 508, + 170, + 846, + 282 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nf _ {v} = E _ {v} (v), f _ {a} = E _ {a} (a), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 602, + 291, + 844, + 306 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nC r o s s A B (Q, K, V) = S o f t m a x \\left(\\frac {Q \\times K ^ {T}}{\\sqrt {d _ {k}}}\\right) \\cdot V, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 527, + 313, + 844, + 344 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {M u l t i H e a d} (Q, K, V) = \\operatorname {C o n c a t} \\left(\\text {h e a d} _ {1}, \\dots , \\text {h e a d} _ {h}\\right) W ^ {O} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 349, + 844, + 375 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nf _ {s v} = \\operatorname {M u l t i H e a d} \\left(Q \\left(f _ {a}\\right), K \\left(f _ {v}\\right), V \\left(f _ {v}\\right)\\right). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 553, + 377, + 844, + 392 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where the $Q(x)$ , $K(x)$ , $V(x)$ map the input feature $x$ to a hidden space though linear projections and $h$ represents the number of head. The egocentric and exocentric sound-vision modules in Figure 2 share the same architecture, as illustrated in Eqs. (1)-(4), but with view-dependent weights. Specifically, we feed egocentric videos and exocentric videos and their audio signals to the sound-vision module to obtain sound-aware representations as:", + "bbox": [ + 509, + 395, + 846, + 494 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nf _ {v} ^ {e g o} = E _ {v} ^ {e g o} \\left(v ^ {e g o}\\right), f _ {v} ^ {e x o} = E _ {v} ^ {e x o} \\left(v ^ {e x o}\\right). \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 562, + 502, + 844, + 517 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nf _ {a} ^ {e g o} = E _ {a} \\left(a ^ {e g o}\\right), f _ {a} ^ {e x o} = E _ {a} \\left(a ^ {e x o}\\right). \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 573, + 521, + 844, + 536 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nf _ {s v} ^ {e g o} = \\text {M u l t i H e a d} \\left(Q \\left(f _ {a} ^ {e g o}\\right), K \\left(f _ {v} ^ {e g o}\\right), V \\left(f _ {v} ^ {e g o}\\right)\\right) \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 526, + 545, + 844, + 561 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nf _ {s v} ^ {e x o} = \\text {M u l t i H e a d} \\left(Q \\left(f _ {a} ^ {e x o}\\right), K \\left(f _ {v} ^ {e x o}\\right), V \\left(f _ {v} ^ {e x o}\\right)\\right) \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 574, + 844, + 590 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Sound-Text Module", + "text_level": 1, + "bbox": [ + 509, + 597, + 671, + 609 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given a text $t$ that captures the semantics of Ego-Exo videos, we use an LLM to generate a sound caption $c$ and propose a sound-text cross-attention module to learn a sound-aware text representation $f_{st}$ . The sound-text module shares the same architecture as the sound-vision module but uses different encoders:", + "bbox": [ + 509, + 616, + 846, + 689 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nf _ {t} = E _ {t} (t), f _ {c} = E _ {t} (c), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 607, + 699, + 844, + 714 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nf _ {s t} = \\text {M u l t i H e a d} \\left(Q \\left(f _ {c}\\right), K \\left(f _ {t}\\right), V \\left(f _ {t}\\right)\\right). \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 557, + 723, + 844, + 738 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Specifically, the sound caption $c$ is generated from the given text $t$ by an LLM. With $t$ providing the video context, we prompt the LLM to generate sound descriptions for all the actions in the videos as: $c = LLM(t, prompt_{Txt2Sound})$ . The sound description needs to contain the source of the sound and the sound effect, formatted in Closed Caption (CC) style [15]. Closed captions (CC) are a specialized type of caption designed for people who are deaf or hard of", + "bbox": [ + 508, + 742, + 846, + 842 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "28944", + "bbox": [ + 478, + 914, + 519, + 926 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/c29c5ed9232595f4f2ea2bdeaaaabf1ee848e54d437d4c748636ad6c99eabe7e.jpg", + "image_caption": [ + "Figure 2. Overview of the proposed method. We leverage sound as a bridge to connect Ego-Exo videos with their corresponding semantic text to build a sound-aligned representation space." + ], + "image_footnote": [], + "bbox": [ + 155, + 151, + 836, + 372 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "hearing. They additionally include an audio description in parentheses and indicate the source of the sound in brackets. For example, for a given text $t =$ \"a person turning off the light\", the generated sound caption is $c =$ \"[a person turning off the light] (click)\". Furthermore, to ensure that the sound caption focuses on the unique information of the paired videos, we filter out frequently appearing concepts (e.g., \"a person\") in CC. To achieve this, we use the LLM to refine $c$ to $\\tilde{c}$ by providing the frequently appeared concepts, i.e., $\\tilde{c} = \\text{LLM}(c, \\text{frequentConcepts}, \\text{prompt}_{\\text{soundRefine}})$ . Due to space limitations, details of all prompts are provided in the supplementary material. We input the filtered and refined closed sound caption $\\tilde{c}$ along with the text $t$ into the sound-text module to obtain the sound-aware text representation: $f_{st} = \\text{MultiHead}(Q(f_{\\tilde{c}}), K(f_t), V(f_t))$ .", + "bbox": [ + 148, + 426, + 485, + 611 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Sound-aware Representation Space Learning", + "text_level": 1, + "bbox": [ + 150, + 616, + 478, + 630 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given $m$ triplets of sound-aware representations for the Ego-Exo video pairs and their corresponding texts: $\\{(f_{sv}^{ego(1)}, f_{sv}^{exo(1)}, f_{st}^{(1)}), \\dots, (f_{sv}^{ego(m)}, f_{sv}^{exo(m)}, f_{st}^{(m)})\\}$ , the visual representations are trained to be closer to the text representation within the same triplet while remaining distant from other text representations using two contrastive losses.", + "bbox": [ + 150, + 635, + 485, + 712 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\text {l o s s} ^ {\\text {e g o 2 t e x t}} = - \\frac {1}{m} \\sum_ {i = 1} ^ {m} \\log \\frac {\\exp \\left(\\sin \\left(f _ {s v} ^ {e g o (i)} , f _ {s t} ^ {(i)}\\right) / \\tau\\right)}{\\sum_ {k \\neq i} \\exp \\left(\\sin \\left(f _ {s v} ^ {e g o (i)} , f _ {s t} ^ {(k)}\\right) / \\tau\\right)} \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 158, + 722, + 485, + 763 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\text {l o s s} ^ {\\text {e x o 2 t e x t}} = - \\frac {1}{m} \\sum_ {i = 1} ^ {m} \\log \\frac {\\exp \\left(\\sin \\left(f _ {s v} ^ {e x o (i)} , f _ {s t} ^ {(i)}\\right) / \\tau\\right)}{\\sum_ {k \\neq i} \\exp \\left(\\sin \\left(f _ {s v} ^ {e x o (i)} , f _ {s t} ^ {(k)}\\right) / \\tau\\right)} \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 158, + 773, + 485, + 815 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\tau$ is the temperature parameter. We use the sum of these two losses to train the sound-aligned Ego-Exo repre", + "bbox": [ + 150, + 816, + 485, + 840 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "sentation space.", + "bbox": [ + 509, + 428, + 602, + 439 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For the Ego-Exo video retrieval and text-video retrieval tasks, we compute the similarity between $(f_{sv}^{ego}, f_{sv}^{exo})$ and $(f_{sv}, f_{st})$ to obtain ranked lists.", + "bbox": [ + 509, + 439, + 844, + 477 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Sound-aligned Space for Downstream Tasks", + "text_level": 1, + "bbox": [ + 509, + 486, + 843, + 501 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To verify the effectiveness of the learned sound-aligned Ego-Exo representation space, we apply the learned sound-aware visual representation to two tasks: scene recognition on egocentric and action recognition on exocentric video. Scene recognition is particularly challenging in an egocentric view because the ego view typically captures local information, such as hand-object interactions. Here, we investigate whether the proposed egocentric sound-vision representation can learn global and environmental information from an exocentric view, thereby enhancing scene recognition performance. We predict the scene classes as follows:", + "bbox": [ + 509, + 507, + 844, + 642 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {y} _ {\\text {s c e n e}} = \\operatorname {s o f t m a x} \\left(F C \\left(f _ {s v} ^ {e g o}\\right)\\right). \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 584, + 652, + 843, + 667 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $FC$ is a fully connected layer that maps the sound-aligned feature to scene classes.", + "bbox": [ + 509, + 670, + 843, + 694 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Similarly, we predict the action label using the exocentric sound-aware visual feature as:", + "bbox": [ + 509, + 695, + 843, + 719 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {y} _ {\\text {a c t i o n}} = \\operatorname {s o f t m a x} \\left(F C \\left(f _ {s v} ^ {e x o}\\right)\\right). \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 581, + 726, + 843, + 742 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $FC$ is a fully connected layer that maps the sound-aligned feature to action classes. We train the FC layers using cross-entropy loss with the ground truth labels.", + "bbox": [ + 509, + 747, + 844, + 785 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 509, + 794, + 624, + 809 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we first evaluate the effectiveness of the sound-aware representation space on Ego-Exo video re", + "bbox": [ + 509, + 815, + 844, + 840 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "28945", + "bbox": [ + 478, + 914, + 517, + 926 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "trieval and text-video retrieval tasks. Next, we apply the sound-aware visual representations to two downstream tasks: scene recognition in egocentric videos and action recognition in exocentric videos. Finally, we conduct ablation studies to verify the effectiveness of our proposed modules.", + "bbox": [ + 148, + 151, + 485, + 224 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. Experimental Settings", + "text_level": 1, + "bbox": [ + 151, + 234, + 328, + 250 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1.1 Datasets and Evaluation Metric", + "text_level": 1, + "bbox": [ + 151, + 253, + 385, + 266 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We conduct the retrieval tasks on two Ego-Exo benchmarks. One benchmark consists of strictly paired Ego-Exo videos (i.e., CharadesEgo [1]), which are recorded using the same shooting script. The other benchmark consists of loosely paired Ego-Exo videos (i.e., EgoExoLearn [12]), where most video pairs share only the same verbs and nouns in their descriptions. CharadesEgo [1] is a large-scale EgoExo dataset featuring videos of daily activities. It comprises 7,985 video clips with a total duration of approximately 80 hours. The dataset is divided into training, validation, and test subsets, containing 5,338, 1,334, and 1,313 video pairs, respectively, along with corresponding shooting script descriptions. EgoExoLearn [12] is another large-scale EgoExo dataset, consisting of 747 video clips with a total duration of 120 hours. Following [34], we construct 36,372 Ego-Exo pairs by aligning nouns and verbs in the video text descriptions. We use the official data split, which includes 36,373 video pairs for training, 800 for validation, and 2,200 for testing. The downstream tasks are also conducted on CharadesEgo [1] and EgoExoLearn [12]. CharadesEgo contains 16 scene categories (e.g., kitchen and bedroom), while EgoExoLearn covers 19 action categories from both daily life and laboratory scenes (e.g., pick-up and washing). We use CharadesEgo for egocentric video scene recognition and EgoExoLearn for exocentric video action recognition. Following previous works [12, 34], we report top-1 accuracy for both retrieval and recognition tasks.", + "bbox": [ + 148, + 273, + 485, + 604 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1.2 Implementation Details", + "text_level": 1, + "bbox": [ + 151, + 625, + 336, + 636 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We use the video transformer from [3] with a ViT-B/16 backbone initialized with CLIP weights [23] as the video encoder $E_{v}(x)$ . For the sound-vision module, we apply the audio encoder from [6] as $E_{a}(x)$ , keeping it frozen during training. In the sound-text module, we use the pretrained CLIP textual encoder as $E_{t}(x)$ to encode both semantic text and sound captions. The output dimension of the sound-aligned Ego-Exo representation space is 256. We use Llama3-8B as the LLM to generate sound captions. In the multi-head cross-attention module, we set the number of heads to $h = 4$ . The temperature parameter $\\tau$ is set to 0.07. We utilize 8 V100 GPUs to train the sound-aware representation space with a learning rate of 3e-5, using AdamW as the optimizer for improved convergence and stability.", + "bbox": [ + 148, + 643, + 485, + 815 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "It is worth noting that for video pairs with high synchronization, applying the Dynamic Time Warping (DTW) algo", + "bbox": [ + 150, + 815, + 485, + 842 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/3e2b74335314434726a3ec48af61d1455bc84b8ed88a201f2d418b2844017e41.jpg", + "table_caption": [ + "Table 1. Retrieval result comparison (top-1 accuracy) of Exo-Ego video alignment task on two benchmarks." + ], + "table_footnote": [], + "table_body": "
MethodEgoExoLearnCharadesEgo
Eg2ExEx2EgAvgEg2ExEx2EgAvg
FrozenInTime [3]14.1013.4013.756.734.965.85
LaViLa [36]28.7025.7027.2058.2446.9352.59
EgoVLP [17]32.1028.930.5058.2760.7659.52
InternVideo [32]30.6021.7026.1553.5561.2357.39
EgoExo [12]49.0045.3047.1568.3262.0665.19
EgoInstructor [34]47.1045.8246.4662.0656.3859.22
SoundBridge50.2750.7350.5077.6671.6374.65
", + "bbox": [ + 511, + 181, + 846, + 303 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "rhythm [24] to align the two audio sequences before inputting the audio signals into the cross-attention module can help mitigate the impact of varying sound speeds. We use this setting by default for videos in CharadesEgo and provide an ablation study on DTW in the supplementary material.", + "bbox": [ + 509, + 313, + 844, + 377 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1.3 Baseline Models", + "text_level": 1, + "bbox": [ + 509, + 390, + 653, + 402 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We include several types of models as baselines. These include recent text-video representation models pre-trained on large-scale exocentric datasets, such as FrozenInTime [3] and InternVideo [32]. Additionally, we include pretrained models on large-scale egocentric datasets, namely LaViLa [36] and EgoVLP [17]. Furthermore, we compare our method with two recent approaches designed for EgoExo video alignment, namely EgoExo [12] and EgoInstructor [34]. For the downstream tasks, we also include task-specific models, namely VideoMAE [30] for exocentric action recognition and InternImage [31] for egocentric scene recognition.", + "bbox": [ + 509, + 409, + 844, + 556 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.2. Results on Ego-Exo Video Alignment", + "text_level": 1, + "bbox": [ + 509, + 564, + 784, + 579 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1 compares the results of Ego-to-Exo and Exo-to-Ego video retrieval with other baselines. Specifically, we report the zero-shot retrieval performance of the FrozenIn-Time model and the fine-tuned results of other methods on the EgoExoLearn and CharadesEgo datasets. Our proposed method, SoundBridge, consistently outperforms other baselines, including pre-trained egocentric models, pre-trained exocentric models, and recent Ego-Exo models, on both Ego-to-Exo and Exo-to-Ego video retrieval tasks across the two benchmarks. We are the first model to achieve a top-1 accuracy exceeding $50\\%$ on the EgoExoLearn dataset. Specifically, SoundBridge achieves higher retrieval accuracy than EgoExo [12] and EgoInstructor [34] by $8.13\\%$ and $15.93\\%$ on Ego-to-Exo video retrieval, and by $13.71\\%$ and $18.88\\%$ on Exo-to-Ego video retrieval, respectively. The performance boost primarily comes from the incorporation of sound-aware modules in the alignment process. For example, given an egocentric video query recording picking up the garlic with the left hand, the EgoExo model ranks an exocentric video of put the ginger down with both hand on the iron bowl in the top-1 position. In contrast, our", + "bbox": [ + 509, + 583, + 846, + 842 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "28946", + "bbox": [ + 478, + 914, + 519, + 926 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/a483729a140a77a2666eb0ddcfb786f9c61e9f26ff1360a6e3d17caf0ef9ae4f.jpg", + "image_caption": [ + "Figure 3. Visualization of an Ego-to-Exo video retrieval example, compared with EgoExo [12] and EgoInstructor [34], on an egocentric video query looking for A person is working on a vacuum in the man cave, then sneezes and reaches for some medicine." + ], + "image_footnote": [], + "bbox": [ + 166, + 148, + 830, + 300 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1cfded1faa5ed2bdf84ee725960a128edc2f1bfa2bd5295f33969693c527ba44.jpg", + "image_caption": [ + "Figure 4. Visualization of an Exo-to-Ego video retrieval example, compared with EgoExo [12] and EgoInstructor [34], on an exocentric video query looking for a person throws a box of groceries onto the table. The person opens the box and takes out an item." + ], + "image_footnote": [], + "bbox": [ + 166, + 347, + 828, + 500 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "method differentiates between the actions \"pick-up\" and \"put down\" using sound cues. The pick-up action produces a soft rustling sound, whereas the put-down action generates a clunk sound when interacting with the iron bowl. Our proposed sound-aware modules effectively encode these audio features and their semantic descriptions (e.g., [pick up ginger] (soft rustling)). As a result, our method correctly ranks the ground-truth exocentric video in the top-1 position.", + "bbox": [ + 148, + 556, + 485, + 655 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Figure 3 visualizes an Ego-to-Exo video retrieval example, comparing our method with EgoExo [12] and EgoInstructor [34]. The egocentric video depicts a person working on a vacuum, sneezing, and picking up a medicine bottle, while the visual cue for the sneeze is barely visible in the egocentric video. All methods successfully rank an egocentric video of a person with a vacuum in the top-1 position. However, EgoExo and EgoInstructor fail to capture either the sneezing action or the action of reaching for the medicine, leading to incorrect top-1 results. In contrast, our method effectively detects the sneeze and the action of picking up the pill bottle using sound cues, thereby successfully retrieving the ground-truth exocentric video. Similarly, Figure 4 visualizes an Exo-to-Ego retrieval example with a query for: a person throws a box of groceries onto", + "bbox": [ + 148, + 656, + 486, + 839 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "the table. The person opens the box and takes out an item. All methods successfully retrieve videos that capture the actions of opening the box and taking out items. However, only our SoundBridge method correctly identifies the action of throwing the box onto the table, leveraging the thud sound as an essential cue.", + "bbox": [ + 509, + 556, + 844, + 630 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.3. Results on Text-Video Retrieval", + "text_level": 1, + "bbox": [ + 509, + 637, + 750, + 651 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The proposed sound-aware Ego-Exo representation space also enables text-video retrieval. Therefore, we compare its performance with baselines on both ego-(exo-)video-to-text and text-to-ego-(exo-)video retrieval tasks. Table 2 demonstrates that our method significantly enhances text-video retrieval performance on Ego-Exo datasets. Specifically, our method substantially improves the performance of the pre-trained egocentric model LaViLa [36], and the pre-trained exocentric model InternVideo [32]. Furthermore, our method outperforms recent Ego-Exo models, namely EgoExo [12] and EgoInstructor [34], by approximately $30.50\\%$ and $23.47\\%$ in overall performance across both datasets. The main improvement over recent Ego-Exo models is observed in queries with multiple nearest-neighbor (NN) answers. By incorporating sound aware", + "bbox": [ + 508, + 657, + 846, + 842 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "28947", + "bbox": [ + 478, + 914, + 517, + 926 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/07afd1a626762d99c078aac83d271451b12ce167020c90b460f0ad5f134c7bc2.jpg", + "table_caption": [ + "Table 2. Performance comparison on text-video retrieval task." + ], + "table_footnote": [], + "table_body": "
MethodEgoExoLearnCharadesEgo
T2VV2TAvgT2VV2TAvg
FrozenInTime [3]6.034.055.049.405.627.51
LaViLa [36]34.8740.1837.5343.7441.6742.71
EgoVLP [17]19.0926.1022.6045.5143.2144.36
InternVideo [32]18.5926.5522.5747.3446.6346.99
EgoExo [12]58.6458.7358.6948.0045.2146.61
EgoInstructor [34]66.3766.6866.5346.9345.9846.46
SoundBridge71.4173.1472.2864.9063.6064.25
", + "bbox": [ + 151, + 170, + 486, + 297 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ness, our method effectively ranks the ground-truth videos higher among similar nearest-neighbor results. For example, given a text query searching for exocentric videos of a person washing dishes in the kitchen, EgoExo retrieves videos of a person chopping potatoes, while EgoInstruc-tor retrieves videos of a person stir-frying. In contrast, our method distinguishes the washing action among these nearest neighbors by leveraging water-running sounds, allowing the ground-truth video to be ranked higher than videos of chopping and stir-frying.", + "bbox": [ + 148, + 306, + 486, + 431 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.4. Results on Downstream Tasks", + "text_level": 1, + "bbox": [ + 151, + 441, + 377, + 455 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We also evaluate the effectiveness of the learned egocentric visual feature (i.e., $f_{sv}^{ego}$ ) for scene recognition and the learned exocentric visual feature (i.e., $f_{sv}^{exo}$ ) for action recognition. We evaluate our model under two settings: (1) fine-tuning, where the training and test sets come from the same dataset, and (2) transfer learning, where the model is trained on one dataset and tested on another. The report performances of other models are all with the fine-tuning setting.", + "bbox": [ + 148, + 460, + 485, + 571 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3 compares the top-1 recognition accuracy with baseline methods. Specifically, for the egocentric scene recognition task, our method improves upon the egocentric model LaViLa [36] and the Ego-Exo model EgoInstructor [34] by $23.42\\%$ and $9.78\\%$ , respectively. Figure 5 illustrates an example of scene recognition. Given an egocentric video query where only a hand wiping a table with a towel is visible, it is difficult to infer that the scene is a kitchen. Other methods fail to predict the correct scene label. For example, FrozenInTime [3] misclassifies the video as a living room, while LaViLa [36] predicts it as a bathroom. In contrast, our method correctly identifies the scene by successfully aligning the query video with the corresponding exocentric video using sound cues from multiple actions (e.g., walking to the table and leaving the room). The egocentric visual feature (i.e., $f_{sv}^{ego}$ ) learns environmental and global visual cues from the exocentric video, which contains a range hood, an oven, and a refrigerator, allowing it to infer that the scene is a kitchen. For exocentric action recognition, our method achieves an average accuracy of $56.25\\%$ , which is lower than VideoMAE [30], as the model is trained on a large-scale action dataset. Despite this, our", + "bbox": [ + 148, + 571, + 486, + 840 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f8f7697f2b8942fb7b259c5c2c74f92bfe93fecb5f1205f8579975ae616a260d.jpg", + "image_caption": [ + "Figure 5. Visualization of the egocentric scene recognition results compared with FrozenInTime [3] and LaViLa [36]." + ], + "image_footnote": [], + "bbox": [ + 513, + 150, + 841, + 218 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4e538bb7800667e0b8c86a8c7206f5e9042021b611e147d188f064d69ac48f20.jpg", + "image_caption": [ + "Figure 6. Visualization of the exocentric action recognition results compared with InternVideo [32] and EgoInstructor [34]." + ], + "image_footnote": [], + "bbox": [ + 517, + 263, + 841, + 330 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/f2038d25561f482732a244e344f12a3ba6fba96adabbafec7ef050a075d8f2e4.jpg", + "table_caption": [ + "Table 3. Result comparison on CharadesEgo scene recognition and EgoExoLearn action recognition. An asterisk (*) indicates the transfer learning setting, where a different dataset is used for training." + ], + "table_footnote": [], + "table_body": "
MethodEgo-SceneExo-Action
FrozenInTime [3]8.256.45
LaViLa [36]41.3726.25
EgoVLP [17]40.3138.75
InternVideo [32]43.0045.00
EgoExo [12]42.5546.25
EgoInstructor [34]46.5147.50
VideoMAEv2 [30]-62.50
InterImage [31]39.60-
SoundBridge*48.7652.50
SoundBridge51.0656.25
", + "bbox": [ + 511, + 431, + 844, + 574 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "approach benefits from the sound-aware Exo-Ego representation space, as shown in Figure 6. Predicting a person's action from an exocentric view is challenging due to occlusions caused by the human body and the person being far from the camera. InternVideo [32] misclassifies the query video as the action \"put\", while EgoInstructor [34] assigns a high probability to the action \"pick-up\". In contrast, our method aligns the query video with its corresponding egocentric view using the sound of washing dishes (e.g., water splashing) and correctly classifies the action as \"washing\" with the highest probability.", + "bbox": [ + 508, + 584, + 846, + 720 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.5. Ablation study", + "text_level": 1, + "bbox": [ + 509, + 727, + 638, + 741 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.5.1 Results on Different LLMs", + "text_level": 1, + "bbox": [ + 509, + 746, + 717, + 758 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We evaluated the impact of audio descriptions generated by different Large Language Models (LLMs) on the performance of cross-view video association tasks. Specifically, we used GPT-3.5 [4] and LLaMA3 [7] to generate audio descriptions using the same prompt. As shown in Table 5, the results indicate that while the cross-view video", + "bbox": [ + 508, + 766, + 844, + 840 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "28948", + "bbox": [ + 478, + 914, + 519, + 926 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/935c1ea4eb1ae08b61690a50fe549bb57a02200f1a94e2ed6a7a4df24904ad65.jpg", + "table_caption": [ + "Table 4. Ablation study on the proposed sound-vision and sound-text modules." + ], + "table_footnote": [], + "table_body": "
MethodEgoExoLearnCharadesEgo
Eg2ExEx2EgT2VV2TEg2ExEx2EgT2VV2T
Baseline49.0045.3058.6458.7368.3262.0648.0045.21
Baseline + Sound-Text Module48.8250.1869.2371.9672.8165.2563.5561.35
Baseline + Sound-Vision Module48.4647.7369.1073.0974.3566.5563.5961.11
Baseline + Sound-Text&Vision Modules50.2750.7371.4173.1477.6671.6364.9063.60
", + "bbox": [ + 153, + 171, + 844, + 253 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/32fdd16a18cc10b8f999769cac3cf2e87b7ab54dadf1aa391983970f187cf56b.jpg", + "table_caption": [ + "Table 5. Results with different LLMs." + ], + "table_footnote": [], + "table_body": "
LLMEgoExoLearnCharadesEgo
Eg2ExEx2EgAvgEg2ExEx2EgAvg
GPT-3.549.2748.3648.8273.7671.5172.64
LLaMA350.2750.7350.5077.6671.6374.65
", + "bbox": [ + 153, + 292, + 485, + 353 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/59ccdc396d19220324bd66f4d9a641d325e73180f2ada8bcb4a7eaa442124e22.jpg", + "table_caption": [ + "Table 6. Performance comparison of different anchors." + ], + "table_footnote": [], + "table_body": "
AnchorEgoExoLearnCharadesEgo
Eg2ExEx2EgEg2ExEx2Eg
Origin Caption48.6448.7374.3566.55
Sound Caption47.0949.9156.9751.42
Sound Feature36.6726.6756.2748.38
Sound-Text Module50.2750.7377.6671.63
", + "bbox": [ + 153, + 385, + 485, + 460 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "association performance with GPT-3.5 is lower than that achieved with LLaMA3, our method still surpasses existing approaches. This demonstrates the effectiveness of our proposed method.", + "bbox": [ + 150, + 469, + 485, + 518 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.5.2 Impact of the Sound-aware Modules", + "text_level": 1, + "bbox": [ + 151, + 537, + 413, + 549 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We conduct an ablation study on the proposed sound-vision and sound-text modules, using the state-of-the-art Ego-Exo model EgoExo [12] as the baseline. The performance of Ego-Exo video alignment and the overall accuracy of text-video retrieval are presented in Table 4. When using only the sound-vision or sound-text module, our method demonstrates competitive performance in the Ego-to-Exo video retrieval task. However, when both modules are jointly utilized, our method achieves significantly higher performance across the evaluated retrieval tasks.", + "bbox": [ + 150, + 556, + 485, + 679 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.5.3 Results of Different Anchors", + "text_level": 1, + "bbox": [ + 151, + 696, + 366, + 709 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We also investigate the effectiveness of using different features as anchors to align sound-aware visual representations. Specifically, we replace the $f_{st}$ in Eq. (11) and Eq. (12) with $f_{t}, f_{c}$ or $\\text{sum}(f_{a}^{ego}, f_{a}^{exo})$ . As shown in Table 6, using audio alone as an anchor is not ideal, especially for the short video dataset EgoExoLearn, where events heavily overlap and exhibit only slight variations. Upon analyzing the erroneous cases, we found that in many samples, the audio anchor alone fails to provide sufficient information to distinguish between videos from different views. As il", + "bbox": [ + 150, + 717, + 485, + 840 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5605f5e34f91f4b7e706fce05920b5b7331a927c86abbc3b344732b3f512eb60.jpg", + "image_caption": [ + "Figure 7. An example of Exo-Ego video retrieval using different anchors in the SoundBridge model." + ], + "image_footnote": [], + "bbox": [ + 514, + 272, + 839, + 364 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "lustrated in Figure 7, a person is slicing garlic in the query video, and the text describes the semantic information of \"a person slicing garlic\". However, the sound features only convey the action \"slice\". During training, different-view videos were aligned solely based on the \"slice\" action, leading to incorrect retrieval of a video showing \"slicing ginger\" instead of \"slicing garlic\". Similarly, the sound caption lacks full contextual information and misinterprets the one-hand slicing action as two-hand slicing.", + "bbox": [ + 509, + 411, + 844, + 521 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 509, + 537, + 612, + 550 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we proposed two sound-aware modules as a bridge to address the large visual variance challenge in egocentric and exocentric video alignment. Both modules demonstrated significant improvements over state-of-the-art methods in Ego-Exo video retrieval and text-video retrieval tasks. Moreover, we verified that both egocentric and exocentric representations benefit from the sound-aware learning paradigm, achieving substantial improvements in downstream recognition tasks by leveraging complementary information from a correctly associated view. Furthermore, we found that our sound-vision and sound-text modules provide effective cues to distinguish the target video from its nearest neighbors based on sound.", + "bbox": [ + 509, + 558, + 844, + 717 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 509, + 733, + 653, + 748 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This research was supported by the National Natural Science Foundation of China (Grant No.: 62372314, 62476097); the Guangdong Provincial Fund for Basic and Applied Basic Research—Regional Joint Fund Project (Key Project) (2023B1515120078); the Guangdong Provincial Natural Science Foundation for Outstanding Youth Team Project (2024B1515040010).", + "bbox": [ + 509, + 753, + 844, + 840 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "28949", + "bbox": [ + 478, + 914, + 517, + 926 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 153, + 150, + 233, + 162 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Karteek Alahari. Actor and observer: Joint modeling of first and third-person videos. In Proceedings of the 1st Workshop and Challenge on Comprehensive Video Understanding in the Wild, page 3, 2018. 2, 5", + "[2] Relja Arandjelovic and Andrew Zisserman. Look, listen and learn. In Proceedings of the IEEE international conference on computer vision, pages 609-617, 2017. 2", + "[3] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1728-1738, 2021. 5, 7", + "[4] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 7", + "[5] Minjie Cai, Kris M Kitani, and Yoichi Sato. Understanding hand-object manipulation with grasp types and object attributes. In Robotics: Science and Systems, 2016. 2", + "[6] Sanyuan Chen, Yu Wu, Chengyi Wang, Shujie Liu, Daniel Tompkins, Zhuo Chen, and Furu Wei. Beats: Audio pre-training with acoustic tokenizers. arXiv preprint arXiv:2212.09058, 2022. 5", + "[7] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 7", + "[8] Antonino Furnari and Giovanni Maria Farinella. What would you expect? anticipating egocentric actions with rolling-unrolling lstms and modality attention. In Proceedings of the IEEE/CVF International conference on computer vision, pages 6252-6261, 2019. 2", + "[9] Mohit Goyal, Sahil Modi, Rishabh Goyal, and Saurabh Gupta. Human hands as probes for interactive object understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3293-3303, 2022. 2", + "[10] Kristen Grauman, Andrew Westbury, Eugene Byrne, Zachary Chavis, Antonino Furnari, Rohit Girdhar, Jackson Hamburger, Hao Jiang, Miao Liu, Xingyu Liu, Miguel Martin, Tushar Nagarajan, Iija Radosavovic, Santhosh Kumar Ramakrishnan, Fiona Ryan, Jayant Sharma, Michael Wray, Mengmeng Xu, Eric Zhongcong Xu, Chen Zhao, Siddhant Bansal, Dhruv Batra, Vincent Cartillier, Sean Crane, Tien Do, Morrie Doulaty, Akshay Erapalli, Christoph Feichtenhofer, Adriano Fragomeni, Qichen Fu, Abrham Gebreselasie, Cristina Gonzalez, James Hillis, Xuhua Huang, Yifei Huang, Wenqi Jia, Weslie Khoo, Jachym" + ], + "bbox": [ + 153, + 171, + 486, + 840 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Kolář, Satwik Kottur, Anurag Kumar, Federico Landini, Chao Li, Yanghao Li, Zhenqiang Li, Karttikeya Mangalam, Raghava Modhugu, Jonathan Munro, Tullie Murrell, Takumi Nishiyasu, Will Price, Paola Ruiz, Merey Ramazanova, Leda Sari, Kiran Somasundaram, Audrey Southerland, Yusuke Sugano, Ruijie Tao, Minh Vo, Yuchen Wang, Xindi Wu, Takuma Yagi, Ziwei Zhao, Yunyi Zhu, Pablo Arbeláez, David Crandall, Dima Damen, Giovanni Maria Farinella, Christian Fuegen, Bernard Ghanem, Vamsi Krishna Ithapu, C. V. Jawahar, Hanbyul Joo, Kris Kitani, Haizhou Li, Richard Newcombe, Aude Oliva, Hyun Soo Park, James M. Rehg, Yoichi Sato, Jianbo Shi, Mike Zheng Shou, Antonio Torralba, Lorenzo Torresani, Mingfei Yan, and Jitendra Malik. Ego4d: Around the world in 3,000 hours of egocentric video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18995-19012, 2022. 1, 2", + "[11] Bo He, Jun Wang, Jielin Qiu, Trung Bui, Abhinav Shrivastava, and Zhaowen Wang. Align and attend: Multimodal summarization with dual contrastive losses. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14867-14878, 2023. 3", + "[12] Yifei Huang, Guo Chen, Jilan Xu, Mingfang Zhang, Lijin Yang, Baoqi Pei, Hongjie Zhang, Lu Dong, Yali Wang, Limin Wang, et al. Egoxolearn: A dataset for bridging asynchronous ego-and exo-centric view of procedural activities in real world. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22072-22086, 2024. 1, 2, 5, 6, 7, 8", + "[13] Sarah Ibrahimi, Xiaohang Sun, Pichao Wang, Amanmeet Garg, Ashutosh Sanan, and Mohamed Omar. Audio-enhanced text-to-video retrieval using text-conditioned feature alignment. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12054-12064, 2023. 2, 3", + "[14] Hao Jiang and Vamsi Krishna Ithapu. Egocentric pose estimation from human vision span. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 10986-10994. IEEE, 2021. 2", + "[15] Patricia S Koskinen, Robert M Wilson, and Carl J Jensema. Using closed-captioned television in the teaching of reading to deaf students. American Annals of the Deaf, 131(1):43-46, 1986. 2, 3", + "[16] Haoxin Li, Yijun Cai, and Wei-Shi Zheng. Deep dual relation modeling for egocentric interaction recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7932-7941, 2019. 2", + "[17] Kevin Qinghong Lin, Jinpeng Wang, Mattia Soldan, Michael Wray, Rui Yan, Eric Z Xu, Difei Gao, Rong-Cheng Tu, Wenzhe Zhao, Weijie Kong, et al. Ego-centric video-language pretraining. Advances in Neu" + ], + "bbox": [ + 513, + 152, + 844, + 840 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "28950", + "bbox": [ + 478, + 914, + 517, + 926 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ral Information Processing Systems, 35:7575-7586, 2022. 5, 7", + "[18] Yan-Bo Lin, Jie Lei, Mohit Bansal, and Gedas Bertasius. Eclipse: Efficient long-range video retrieval using sight and sound. In European Conference on Computer Vision, pages 413–430. Springer, 2022. 2, 3", + "[19] Shaowei Liu, Subarna Tripathi, Somdeb Majumdar, and Xiaolong Wang. Joint hand motion and interaction hotspots prediction from egocentric videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3282-3292, 2022. 2", + "[20] Sagnik Majumder, Ziad Al-Halah, and Kristen Grauman. Learning spatial features from audio-visual correspondence in egocentric videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 27058-27068, 2024. 3", + "[21] Tushar Nagarajan, Yanghao Li, Christoph Feichtenhofer, and Kristen Grauman. Ego-topo: Environment affordances from egocentric video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 163-172, 2020. 2", + "[22] Chiara Plizzari, Mirco Planamente, Gabriele Goletto, Marco Cannici, Emanuele Gusso, Matteo Matteucci, and Barbara Caputo. E2 (go) motion: Motion augmented event stream for egocentric action recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 19935-19947, 2022. 2", + "[23] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 5", + "[24] Hiroaki Sakoe and Seibi Chiba. Dynamic programming algorithm optimization for spoken word recognition. IEEE transactions on acoustics, speech, and signal processing, 26(1):43-49, 1978. 5", + "[25] Muhammad Bilal Shaikh, Douglas Chai, Syed Mohammed Shamsul Islam, and Naveed Akhtar. Multimodal fusion for audio-image and video action recognition. Neural Computing and Applications, 36(10): 5499-5513, 2024. 3", + "[26] Dandan Shan, Jiaqi Geng, Michelle Shu, and David F Fouhey. Understanding human hands in contact at internet scale. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9869-9878, 2020. 2", + "[27] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 3", + "[28] Huiyu Wang, Mitesh Kumar Singh, and Lorenzo Torresani. Ego-only: Egocentric action detection with" + ], + "bbox": [ + 153, + 152, + 486, + 840 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "out exocentric transferring. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5250-5261, 2023. 2", + "[29] Jian Wang, Diogo Luvizon, Weipeng Xu, Lingjie Liu, Kripasindhu Sarkar, and Christian Theobalt. Scene-aware egocentric 3d human pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13031-13040, 2023. 2", + "[30] Limin Wang, Bingkun Huang, Zhiyu Zhao, Zhan Tong, Yinan He, Yi Wang, Yali Wang, and Yu Qiao. Videomae v2: Scaling video masked autoencoders with dual masking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14549-14560, 2023. 5, 7", + "[31] Wenhai Wang, Jifeng Dai, Zhe Chen, Zhenhang Huang, Zhiqi Li, Xizhou Zhu, Xiaowei Hu, Tong Lu, Lewei Lu, Hongsheng Li, et al. Internimage: Exploring large-scale vision foundation models with deformable convolutions. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14408-14419, 2023. 5, 7", + "[32] Yi Wang, Yinan He, Yizhuo Li, Kunchang Li, Jiashuo Yu, Xin Ma, Xinhao Li, Guo Chen, Xinyuan Chen, Yaohui Wang, Ping Luo, Ziwei Liu, Yali Wang, Limin Wang, and Yu Qiao. Internvid: A large-scale videotext dataset for multimodal understanding and generation. In The Twelfth International Conference on Learning Representations, 2024. 5, 6, 7", + "[33] Jianjia Xin, Lichun Wang, Kai Xu, Chao Yang, and Baocai Yin. Learning interaction regions and motion trajectories simultaneously from egocentric demonstration videos. IEEE Robotics and Automation Letters, 2023. 2", + "[34] Jilan Xu, Yifei Huang, Junlin Hou, Guo Chen, Yuejie Zhang, Rui Feng, and Weidi Xie. Retrieval-augmented egocentric video captioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13525-13536, 2024. 1, 2, 5, 6, 7", + "[35] Zihui Sherry Xue and Kristen Grauman. Learning fine-grained view-invariant representations from unpaired ego-exo videos via temporal alignment. Advances in Neural Information Processing Systems, 36: 53688-53710, 2023. 2", + "[36] Yue Zhao, Ishan Misra, Philipp Krahenbuhl, and Rohit Girdhar. Learning video representations from large language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6586-6597, 2023. 5, 6, 7" + ], + "bbox": [ + 513, + 152, + 844, + 771 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "28951", + "bbox": [ + 478, + 914, + 517, + 926 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Sound Bridge_ Associating Egocentric and Exocentric Videos via Audio Cues/37c295df-9911-4189-b46a-8795fdcf6a6f_model.json b/2025/Sound Bridge_ Associating Egocentric and Exocentric Videos via Audio Cues/37c295df-9911-4189-b46a-8795fdcf6a6f_model.json new file mode 100644 index 0000000000000000000000000000000000000000..a7e26b926a61fdd771ba16d70caf1ea4fc1367a6 --- /dev/null +++ b/2025/Sound Bridge_ Associating Egocentric and Exocentric Videos via Audio Cues/37c295df-9911-4189-b46a-8795fdcf6a6f_model.json @@ -0,0 +1,1980 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.02 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation." + }, + { + "type": "header", + "bbox": [ + 0.325, + 0.017, + 0.723, + 0.031 + ], + "angle": 0, + "content": "Except for this watermark, it is identical to the accepted version;" + }, + { + "type": "header", + "bbox": [ + 0.294, + 0.032, + 0.754, + 0.046 + ], + "angle": 0, + "content": "the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.168, + 0.185, + 0.83, + 0.204 + ], + "angle": 0, + "content": "Sound Bridge: Associating Egocentric and Exocentric Videos via Audio Cues" + }, + { + "type": "text", + "bbox": [ + 0.173, + 0.226, + 0.823, + 0.242 + ], + "angle": 0, + "content": "Sihong Huang\\(^{1,5,*}\\), Jiaxin Wu\\(^{2,*}\\), Xiaoyong Wei\\(^{3,2,5,\\dagger}\\), Yi Cai\\(^{1,\\dagger}\\), Dongmei Jiang\\(^{5}\\), Yaowei Wang\\(^{4,5}\\)" + }, + { + "type": "text", + "bbox": [ + 0.228, + 0.241, + 0.771, + 0.258 + ], + "angle": 0, + "content": "1South China University of Technology, 2The Hong Kong Polytechnic University" + }, + { + "type": "text", + "bbox": [ + 0.199, + 0.257, + 0.8, + 0.273 + ], + "angle": 0, + "content": "\\(^{3}\\)Sichuan University, \\(^{4}\\)Harbin Institute of Technology, Shenzhen, \\(^{5}\\)Peng Cheng Laboratory" + }, + { + "type": "text", + "bbox": [ + 0.225, + 0.273, + 0.773, + 0.287 + ], + "angle": 0, + "content": "{fthuangsh, ycai}@mail.scut.edu.cn, jiaxwu@polyu.edu.hk, cswei@scu.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.346, + 0.288, + 0.648, + 0.3 + ], + "angle": 0, + "content": "jiangdm@pcl.ac.cn, wangyaowei@hit.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.286, + 0.342, + 0.352, + 0.356 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.15, + 0.368, + 0.487, + 0.651 + ], + "angle": 0, + "content": "Understanding human behavior and environmental information in egocentric videos is very challenging due to the invisibility of some actions (e.g., laughing and sneezing) and the local nature of the first-person view. Leveraging the corresponding exocentric video to provide global context has shown promising results. However, existing visual-to-visual and visual-to-textual Ego-Exo video alignment methods struggle with the issue that some activities may have non-visual overlap. To address this, we propose using sound as a bridge, as audio is often consistent across Ego-Exo videos. However, direct audio-to-audio alignment lacks context. Thus, we introduce two context-aware sound modules: one aligns audio with vision via a visual-audio cross-attention module, and another aligns text with sound closed caption generated by LLM. Experimental results on two Ego-Exo video association benchmarks show that each of the proposed modules enhances the state-of-the-art methods. Moreover, the proposed sound-aware egocentric or exocentric representation boosts the performance of downstream tasks, such as action recognition of exocentric videos and scene recognition of egocentric videos. The code and models can be accessed at https://github.com/shhuangcoder/SoundBridge." + }, + { + "type": "title", + "bbox": [ + 0.152, + 0.675, + 0.265, + 0.688 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.15, + 0.696, + 0.487, + 0.82 + ], + "angle": 0, + "content": "Egocentric video analysis has been receiving extensive attention recently [10, 12, 34] due to its potential applications in robotic vision analysis, embedded AI, etc. Compared to the large-scale and in-depth studies on exocentric video analysis tasks, research on egocentric videos is still underexplored. Moreover, there are significant differences between egocentric and exocentric videos; thus, many findings and methodologies derived from an exocentric perspective are not directly applicable to egocentric video analysis." + }, + { + "type": "image", + "bbox": [ + 0.513, + 0.329, + 0.845, + 0.472 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.51, + 0.481, + 0.846, + 0.517 + ], + "angle": 0, + "content": "Figure 1. Given a video shooting script, egocentric and exocentric views can exhibit significant visual variances for the same activity, while their audio signals remain relatively consistent." + }, + { + "type": "text", + "bbox": [ + 0.509, + 0.535, + 0.846, + 0.78 + ], + "angle": 0, + "content": "The main difference between egocentric and exocentric videos is that they capture events from different perspectives. First-person view videos focus on local and detailed information about the current activity, whereas third-person view videos emphasize global and environmental context. An example is shown in Figure 1. Although egocentric and exocentric videos follow the same video shooting script, they exhibit significant visual variations. For example, in the case of the same activity, such as eating a snack, the egocentric view primarily captures hand-object interactions, such as a hand picking up a raisin, while the location remains unknown. In contrast, the exocentric view captures the overall scene of the event but lacks finer details—for example, it may show a person lying on a bed eating food, but the specific type of food remains unclear. Moreover, certain activities, such as laughing and sneezing, do not produce distinct visual cues in egocentric videos apart from camera shaking. Consequently, when attempting to align visual representations from exocentric to egocentric videos, there is often no overlapping visual content to match." + }, + { + "type": "text", + "bbox": [ + 0.509, + 0.78, + 0.846, + 0.842 + ], + "angle": 0, + "content": "However, recent studies [12, 34] have shown that exocentric videos can supplement the understanding of their semantically aligned egocentric counterparts. In this paper, we also focus on constructing an Ego-Exo alignment space to facilitate egocentric video analysis. Previous studies on" + }, + { + "type": "page_footnote", + "bbox": [ + 0.172, + 0.83, + 0.392, + 0.842 + ], + "angle": 0, + "content": "* Equal Contribution. † Corresponding Authors." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.915, + 0.521, + 0.927 + ], + "angle": 0, + "content": "28942" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.15, + 0.153, + 0.486, + 0.287 + ], + "angle": 0, + "content": "egocentric and exocentric video alignment have primarily focused on matching visual content [35] or aligning both video types to an anchor (i.e., their semantic descriptions or a shared textual representation) [12, 34]. However, these alignment strategies may be insufficient due to the substantial visual variations between views. In contrast to the significant differences in view-dependent visual content, we observe that audio signals for the same event remain consistent across views. For example, in the case of laughing, the laughter sound—whether captured in an egocentric or exocentric video—can infer that someone is laughing." + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.287, + 0.486, + 0.838 + ], + "angle": 0, + "content": "In this paper, we propose using audio signals as a bridge to model the egocentric-exocentric video alignment space. However, directly linking audio signals across videos without context may be suboptimal, as audio information alone is not sufficiently discriminative [2]. For example, the sound of a click associated with the action \"switch off the light\" in Figure 1 could also occur in other activities, such as \"flipping a power switch on a household appliance\" or \"turning off a desk lamp\". To address this limitation, we propose constructing contextual audio features for alignment by incorporating visual and textual context. An overview of our proposed method is presented in Figure 2. Specifically, we introduce sound-text and sound-vision modules to bridge Ego-Exo videos. The sound-vision module follows existing methods [13, 18] to build contextual audiovisual representations through cross-attention. Meanwhile, the sound-text module utilizes a large language model (LLM) to generate contextual audio descriptions and construct sound-aware textual representations. In particular, we provide the LLM with semantic text describing the video scenario and prompt it to generate an audio description of the actions occurring within the scene in a closed-caption format [15]. Additionally, to ensure that the generated sound captions emphasize unique information within the paired cross-view videos rather than commonly occurring content in the dataset, we filter out frequently appearing and less informative concepts (e.g., person) from the generated descriptions. As a result, the output sound caption for the example in Figure 1 is \"[Eating the snack] (munching), [dishes on the floor] (crashing and scraping), [Laughing] (giggles), [leaving the room] (Footsteps), [turns off the light] (click)\". The generated sound caption and the semantic text are subsequently processed through a cross-attention block to learn a sound-aware textual representation. Finally, we treat the textual representation as an anchor and align the audiovisual representations of egocentric and exocentric videos to it, thereby learning an Ego-Exo semantic feature space. We conduct extensive experiments on two Ego-Exo video retrieval benchmarks, namely EgoExoLearn [12] and CharadesEgo [1]. The results demonstrate that the learned sound-aware features outperform state-of-the-art approaches in both egocentric-exocentric video retrieval and text-to-egocentric/exocentric video retrieval. Moreover, the learned Ego- and Exo- features enhance the performance of downstream tasks, such as exocentric video action" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.153, + 0.802, + 0.165 + ], + "angle": 0, + "content": "recognition and egocentric video scene recognition." + }, + { + "type": "title", + "bbox": [ + 0.511, + 0.175, + 0.634, + 0.189 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.511, + 0.196, + 0.714, + 0.21 + ], + "angle": 0, + "content": "2.1. Egocentric Video Analysis" + }, + { + "type": "text", + "bbox": [ + 0.51, + 0.215, + 0.846, + 0.522 + ], + "angle": 0, + "content": "The number of research studies on egocentric video analysis is growing rapidly, as this task is a fundamental component of many applications [10, 12], such as robotic vision analysis and embedded AI. However, egocentric videos are fundamentally different from traditional exocentric videos. Egocentric videos capture a unique immersive viewpoint from a first-person perspective, directly reflecting an individual's interaction with the current activity, whereas exocentric videos provide an external view of human interactions within an environment. This distinct perspective in egocentric videos introduces a range of challenges for human activity analysis, including hand detection [5, 9, 26], action recognition [8, 22, 28], human-computer interaction [16, 21, 33], and applications in virtual and augmented reality [14, 19, 29]. Despite these differences, recent studies [12, 34] have demonstrated that egocentric video analysis can benefit from the contextual information provided by corresponding exocentric videos. For example, Xu et al. [34] show that related exocentric videos can enhance egocentric video captioning. Similarly, Huang et al. [12] demonstrate that the exocentric perspective provides complementary information that aids robots in performing complex tasks. In this paper, we also focus on developing effective Ego-Exo video alignment techniques to enhance egocentric video analysis." + }, + { + "type": "title", + "bbox": [ + 0.511, + 0.528, + 0.808, + 0.542 + ], + "angle": 0, + "content": "2.2. Ego-Exo Video Representation Learning" + }, + { + "type": "text", + "bbox": [ + 0.51, + 0.548, + 0.846, + 0.844 + ], + "angle": 0, + "content": "The goal of Ego-Exo video representation learning is to construct a semantic space in which egocentric and exocentric videos with similar semantics are positioned closely, while dissimilar Ego-Exo videos are placed farther apart [12]. Early studies on Ego-Exo representation learning align cross-view videos using visual information. For example, AE2 [35] extracts visual regions of hands and active objects and performs temporal alignment on these extracted regions across views, assuming that the visual content of a fine-grained action remains consistent across perspectives. AE2 constructs four action-specific cross-view datasets for evaluation: break egg, pour milk, pour liquid, and tennis forehand. Recent approaches aim to learn a semantic Ego-Exo representation space by aligning videos to either a shared semantic text or separate textual descriptions. For example, EgoExo [12] employs implicit Ego-Exo alignment by associating each video with a view-dependent text, based on the assumption that semantically similar texts will be positioned closely in the learned space. Beyond implicit alignment, EgoExoNCE [34] takes this further by explicitly aligning Ego-Exo videos to a mutual text anchor. This shared text is generated by extracting key nouns and verbs from both the egocentric and exocentric textual descriptions. In this paper, we follow the pipeline of previous" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.915, + 0.519, + 0.927 + ], + "angle": 0, + "content": "28943" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.15, + 0.152, + 0.487, + 0.252 + ], + "angle": 0, + "content": "work by aligning Ego-Exo videos to a mutual text anchor. However, rather than directly matching entire videos with full-text descriptions without attention, we propose using audio as a bridge to address the challenge of visual variance. Our approach incorporates sound-vision and sound-text modules to bridge the gap between egocentric and exocentric perspectives. To the best of our knowledge, we are the first to integrate sound into Ego-Exo video alignment." + }, + { + "type": "title", + "bbox": [ + 0.152, + 0.272, + 0.436, + 0.287 + ], + "angle": 0, + "content": "2.3. Sound-aware Representation Learning" + }, + { + "type": "text", + "bbox": [ + 0.15, + 0.291, + 0.487, + 0.635 + ], + "angle": 0, + "content": "Numerous studies have demonstrated that audio signals complement visual signals in video analysis tasks [18]. For example, the Eclipse method has shown that audio signals can serve as an alternative to visual signals in long-range video retrieval, significantly accelerating the search process [18]. Consequently, many approaches have sought to develop sound-aware representations. Several works have incorporated visual information as context to create audiovisual features for exocentric video analysis, applying them to tasks such as video retrieval [13, 18], action recognition [20], and video summarization [11]. More recently, audio-based methods have also been explored in egocentric video analysis, such as active speaker spatial detection [25]. Additionally, Ibrahimi et al. [13] integrate textual context to construct audio-textual representations. Common approaches to incorporating audio signals into video analysis include appending audio features to visual or textual features [20, 25] or learning audio-textual/visual cross-attention representations [13, 18]. Another method involves converting audio signals into text using Automatic Speech Recognition (ASR) and subsequently appending the ASR-derived text features to visual features [11]. Our proposed sound-vision model follows prior work by learning contextual audiovisual features through cross-attention. Additionally, we introduce a novel approach to learning audiotextual representations in our sound-text module, which leverages a large language model (LLM) to generate sound captions." + }, + { + "type": "title", + "bbox": [ + 0.152, + 0.659, + 0.486, + 0.688 + ], + "angle": 0, + "content": "3. Sound-aligned Ego-Exo Representation Learning" + }, + { + "type": "text", + "bbox": [ + 0.15, + 0.694, + 0.487, + 0.843 + ], + "angle": 0, + "content": "The architecture of our proposed method is illustrated in Figure 2. The main objective is to construct a sound-aligned Ego-Exo representation space that integrates local (Ego) and global (Exo) visual features along with their corresponding semantic concepts. Specifically, to address the challenge of aligning semantic text with the significant visual variations present in Ego-Exo videos, we leverage sound as a unifying element. This is achieved by incorporating sound-vision and sound-text modules to generate sound-aware representations. Subsequently, we align the sound-aware video features with the sound-aware textual features to learn a unified representation space." + }, + { + "type": "title", + "bbox": [ + 0.511, + 0.152, + 0.686, + 0.164 + ], + "angle": 0, + "content": "3.1. Sound-Vision Module" + }, + { + "type": "text", + "bbox": [ + 0.509, + 0.171, + 0.847, + 0.283 + ], + "angle": 0, + "content": "Given a semantically aligned egocentric and exocentric (ego-exo) video pair \\((v^{ego}, v^{exo})\\) and their corresponding audio signals \\((a^{ego}, a^{exo})\\), we propose a sound-vision cross-attention module to learn their sound-aware visual representations. The sound-vision module consists of a video encoder \\(E_v(x)\\), an audio encoder \\(E_a(x)\\), and a multi-head cross-attention block Multihead \\((Q, K, V)\\) [27]. The module inputs a video \\(v\\) and its corresponding audio signal \\(a\\) and outputs sound-aware visual feature \\(f_{sv}\\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.604, + 0.292, + 0.846, + 0.308 + ], + "angle": 0, + "content": "\\[\nf _ {v} = E _ {v} (v), f _ {a} = E _ {a} (a), \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.529, + 0.314, + 0.846, + 0.345 + ], + "angle": 0, + "content": "\\[\nC r o s s A B (Q, K, V) = S o f t m a x \\left(\\frac {Q \\times K ^ {T}}{\\sqrt {d _ {k}}}\\right) \\cdot V, \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.515, + 0.35, + 0.845, + 0.376 + ], + "angle": 0, + "content": "\\[\n\\operatorname {M u l t i H e a d} (Q, K, V) = \\operatorname {C o n c a t} \\left(\\text {h e a d} _ {1}, \\dots , \\text {h e a d} _ {h}\\right) W ^ {O} \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.555, + 0.378, + 0.845, + 0.393 + ], + "angle": 0, + "content": "\\[\nf _ {s v} = \\operatorname {M u l t i H e a d} \\left(Q \\left(f _ {a}\\right), K \\left(f _ {v}\\right), V \\left(f _ {v}\\right)\\right). \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.51, + 0.396, + 0.847, + 0.495 + ], + "angle": 0, + "content": "where the \\( Q(x) \\), \\( K(x) \\), \\( V(x) \\) map the input feature \\( x \\) to a hidden space though linear projections and \\( h \\) represents the number of head. The egocentric and exocentric sound-vision modules in Figure 2 share the same architecture, as illustrated in Eqs. (1)-(4), but with view-dependent weights. Specifically, we feed egocentric videos and exocentric videos and their audio signals to the sound-vision module to obtain sound-aware representations as:" + }, + { + "type": "equation", + "bbox": [ + 0.563, + 0.503, + 0.846, + 0.518 + ], + "angle": 0, + "content": "\\[\nf _ {v} ^ {e g o} = E _ {v} ^ {e g o} \\left(v ^ {e g o}\\right), f _ {v} ^ {e x o} = E _ {v} ^ {e x o} \\left(v ^ {e x o}\\right). \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.575, + 0.522, + 0.845, + 0.537 + ], + "angle": 0, + "content": "\\[\nf _ {a} ^ {e g o} = E _ {a} \\left(a ^ {e g o}\\right), f _ {a} ^ {e x o} = E _ {a} \\left(a ^ {e x o}\\right). \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.527, + 0.546, + 0.845, + 0.562 + ], + "angle": 0, + "content": "\\[\nf _ {s v} ^ {e g o} = \\text {M u l t i H e a d} \\left(Q \\left(f _ {a} ^ {e g o}\\right), K \\left(f _ {v} ^ {e g o}\\right), V \\left(f _ {v} ^ {e g o}\\right)\\right) \\tag {7}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.526, + 0.575, + 0.845, + 0.591 + ], + "angle": 0, + "content": "\\[\nf _ {s v} ^ {e x o} = \\text {M u l t i H e a d} \\left(Q \\left(f _ {a} ^ {e x o}\\right), K \\left(f _ {v} ^ {e x o}\\right), V \\left(f _ {v} ^ {e x o}\\right)\\right) \\tag {8}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.511, + 0.598, + 0.673, + 0.61 + ], + "angle": 0, + "content": "3.2. Sound-Text Module" + }, + { + "type": "text", + "bbox": [ + 0.51, + 0.617, + 0.847, + 0.69 + ], + "angle": 0, + "content": "Given a text \\(t\\) that captures the semantics of Ego-Exo videos, we use an LLM to generate a sound caption \\(c\\) and propose a sound-text cross-attention module to learn a sound-aware text representation \\(f_{st}\\). The sound-text module shares the same architecture as the sound-vision module but uses different encoders:" + }, + { + "type": "equation", + "bbox": [ + 0.608, + 0.7, + 0.846, + 0.715 + ], + "angle": 0, + "content": "\\[\nf _ {t} = E _ {t} (t), f _ {c} = E _ {t} (c), \\tag {9}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.558, + 0.724, + 0.845, + 0.739 + ], + "angle": 0, + "content": "\\[\nf _ {s t} = \\text {M u l t i H e a d} \\left(Q \\left(f _ {c}\\right), K \\left(f _ {t}\\right), V \\left(f _ {t}\\right)\\right). \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.509, + 0.743, + 0.847, + 0.843 + ], + "angle": 0, + "content": "Specifically, the sound caption \\( c \\) is generated from the given text \\( t \\) by an LLM. With \\( t \\) providing the video context, we prompt the LLM to generate sound descriptions for all the actions in the videos as: \\( c = LLM(t, prompt_{Txt2Sound}) \\). The sound description needs to contain the source of the sound and the sound effect, formatted in Closed Caption (CC) style [15]. Closed captions (CC) are a specialized type of caption designed for people who are deaf or hard of" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.915, + 0.521, + 0.927 + ], + "angle": 0, + "content": "28944" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.156, + 0.152, + 0.837, + 0.373 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.15, + 0.382, + 0.846, + 0.406 + ], + "angle": 0, + "content": "Figure 2. Overview of the proposed method. We leverage sound as a bridge to connect Ego-Exo videos with their corresponding semantic text to build a sound-aligned representation space." + }, + { + "type": "text", + "bbox": [ + 0.15, + 0.428, + 0.486, + 0.612 + ], + "angle": 0, + "content": "hearing. They additionally include an audio description in parentheses and indicate the source of the sound in brackets. For example, for a given text \\( t = \\) \"a person turning off the light\", the generated sound caption is \\( c = \\) \"[a person turning off the light] (click)\". Furthermore, to ensure that the sound caption focuses on the unique information of the paired videos, we filter out frequently appearing concepts (e.g., \"a person\") in CC. To achieve this, we use the LLM to refine \\( c \\) to \\( \\tilde{c} \\) by providing the frequently appeared concepts, i.e., \\( \\tilde{c} = \\text{LLM}(c, \\text{frequentConcepts}, \\text{prompt}_{\\text{soundRefine}}) \\). Due to space limitations, details of all prompts are provided in the supplementary material. We input the filtered and refined closed sound caption \\( \\tilde{c} \\) along with the text \\( t \\) into the sound-text module to obtain the sound-aware text representation: \\( f_{st} = \\text{MultiHead}(Q(f_{\\tilde{c}}), K(f_t), V(f_t)) \\)." + }, + { + "type": "title", + "bbox": [ + 0.151, + 0.617, + 0.48, + 0.631 + ], + "angle": 0, + "content": "3.3. Sound-aware Representation Space Learning" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.636, + 0.486, + 0.713 + ], + "angle": 0, + "content": "Given \\(m\\) triplets of sound-aware representations for the Ego-Exo video pairs and their corresponding texts: \\(\\{(f_{sv}^{ego(1)}, f_{sv}^{exo(1)}, f_{st}^{(1)}), \\dots, (f_{sv}^{ego(m)}, f_{sv}^{exo(m)}, f_{st}^{(m)})\\}\\), the visual representations are trained to be closer to the text representation within the same triplet while remaining distant from other text representations using two contrastive losses." + }, + { + "type": "equation", + "bbox": [ + 0.16, + 0.723, + 0.486, + 0.765 + ], + "angle": 0, + "content": "\\[\n\\text {l o s s} ^ {\\text {e g o 2 t e x t}} = - \\frac {1}{m} \\sum_ {i = 1} ^ {m} \\log \\frac {\\exp \\left(\\sin \\left(f _ {s v} ^ {e g o (i)} , f _ {s t} ^ {(i)}\\right) / \\tau\\right)}{\\sum_ {k \\neq i} \\exp \\left(\\sin \\left(f _ {s v} ^ {e g o (i)} , f _ {s t} ^ {(k)}\\right) / \\tau\\right)} \\tag {11}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.16, + 0.775, + 0.486, + 0.816 + ], + "angle": 0, + "content": "\\[\n\\text {l o s s} ^ {\\text {e x o 2 t e x t}} = - \\frac {1}{m} \\sum_ {i = 1} ^ {m} \\log \\frac {\\exp \\left(\\sin \\left(f _ {s v} ^ {e x o (i)} , f _ {s t} ^ {(i)}\\right) / \\tau\\right)}{\\sum_ {k \\neq i} \\exp \\left(\\sin \\left(f _ {s v} ^ {e x o (i)} , f _ {s t} ^ {(k)}\\right) / \\tau\\right)} \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.817, + 0.486, + 0.842 + ], + "angle": 0, + "content": "where \\(\\tau\\) is the temperature parameter. We use the sum of these two losses to train the sound-aligned Ego-Exo repre" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.429, + 0.603, + 0.44 + ], + "angle": 0, + "content": "sentation space." + }, + { + "type": "text", + "bbox": [ + 0.51, + 0.44, + 0.845, + 0.478 + ], + "angle": 0, + "content": "For the Ego-Exo video retrieval and text-video retrieval tasks, we compute the similarity between \\((f_{sv}^{ego}, f_{sv}^{exo})\\) and \\((f_{sv}, f_{st})\\) to obtain ranked lists." + }, + { + "type": "title", + "bbox": [ + 0.51, + 0.487, + 0.844, + 0.502 + ], + "angle": 0, + "content": "4. Sound-aligned Space for Downstream Tasks" + }, + { + "type": "text", + "bbox": [ + 0.51, + 0.508, + 0.845, + 0.643 + ], + "angle": 0, + "content": "To verify the effectiveness of the learned sound-aligned Ego-Exo representation space, we apply the learned sound-aware visual representation to two tasks: scene recognition on egocentric and action recognition on exocentric video. Scene recognition is particularly challenging in an egocentric view because the ego view typically captures local information, such as hand-object interactions. Here, we investigate whether the proposed egocentric sound-vision representation can learn global and environmental information from an exocentric view, thereby enhancing scene recognition performance. We predict the scene classes as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.585, + 0.654, + 0.844, + 0.668 + ], + "angle": 0, + "content": "\\[\n\\hat {y} _ {\\text {s c e n e}} = \\operatorname {s o f t m a x} \\left(F C \\left(f _ {s v} ^ {e g o}\\right)\\right). \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.51, + 0.671, + 0.844, + 0.695 + ], + "angle": 0, + "content": "where \\(FC\\) is a fully connected layer that maps the sound-aligned feature to scene classes." + }, + { + "type": "text", + "bbox": [ + 0.51, + 0.696, + 0.844, + 0.72 + ], + "angle": 0, + "content": "Similarly, we predict the action label using the exocentric sound-aware visual feature as:" + }, + { + "type": "equation", + "bbox": [ + 0.583, + 0.727, + 0.844, + 0.743 + ], + "angle": 0, + "content": "\\[\n\\hat {y} _ {\\text {a c t i o n}} = \\operatorname {s o f t m a x} \\left(F C \\left(f _ {s v} ^ {e x o}\\right)\\right). \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.51, + 0.748, + 0.845, + 0.786 + ], + "angle": 0, + "content": "where \\(FC\\) is a fully connected layer that maps the sound-aligned feature to action classes. We train the FC layers using cross-entropy loss with the ground truth labels." + }, + { + "type": "title", + "bbox": [ + 0.511, + 0.795, + 0.625, + 0.81 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.51, + 0.816, + 0.845, + 0.842 + ], + "angle": 0, + "content": "In this section, we first evaluate the effectiveness of the sound-aware representation space on Ego-Exo video re" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.915, + 0.518, + 0.927 + ], + "angle": 0, + "content": "28945" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.15, + 0.152, + 0.486, + 0.226 + ], + "angle": 0, + "content": "trieval and text-video retrieval tasks. Next, we apply the sound-aware visual representations to two downstream tasks: scene recognition in egocentric videos and action recognition in exocentric videos. Finally, we conduct ablation studies to verify the effectiveness of our proposed modules." + }, + { + "type": "title", + "bbox": [ + 0.152, + 0.235, + 0.33, + 0.25 + ], + "angle": 0, + "content": "5.1. Experimental Settings" + }, + { + "type": "title", + "bbox": [ + 0.152, + 0.255, + 0.386, + 0.267 + ], + "angle": 0, + "content": "5.1.1 Datasets and Evaluation Metric" + }, + { + "type": "text", + "bbox": [ + 0.15, + 0.275, + 0.486, + 0.606 + ], + "angle": 0, + "content": "We conduct the retrieval tasks on two Ego-Exo benchmarks. One benchmark consists of strictly paired Ego-Exo videos (i.e., CharadesEgo [1]), which are recorded using the same shooting script. The other benchmark consists of loosely paired Ego-Exo videos (i.e., EgoExoLearn [12]), where most video pairs share only the same verbs and nouns in their descriptions. CharadesEgo [1] is a large-scale EgoExo dataset featuring videos of daily activities. It comprises 7,985 video clips with a total duration of approximately 80 hours. The dataset is divided into training, validation, and test subsets, containing 5,338, 1,334, and 1,313 video pairs, respectively, along with corresponding shooting script descriptions. EgoExoLearn [12] is another large-scale EgoExo dataset, consisting of 747 video clips with a total duration of 120 hours. Following [34], we construct 36,372 Ego-Exo pairs by aligning nouns and verbs in the video text descriptions. We use the official data split, which includes 36,373 video pairs for training, 800 for validation, and 2,200 for testing. The downstream tasks are also conducted on CharadesEgo [1] and EgoExoLearn [12]. CharadesEgo contains 16 scene categories (e.g., kitchen and bedroom), while EgoExoLearn covers 19 action categories from both daily life and laboratory scenes (e.g., pick-up and washing). We use CharadesEgo for egocentric video scene recognition and EgoExoLearn for exocentric video action recognition. Following previous works [12, 34], we report top-1 accuracy for both retrieval and recognition tasks." + }, + { + "type": "title", + "bbox": [ + 0.152, + 0.625, + 0.338, + 0.637 + ], + "angle": 0, + "content": "5.1.2 Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.15, + 0.645, + 0.486, + 0.816 + ], + "angle": 0, + "content": "We use the video transformer from [3] with a ViT-B/16 backbone initialized with CLIP weights [23] as the video encoder \\( E_{v}(x) \\). For the sound-vision module, we apply the audio encoder from [6] as \\( E_{a}(x) \\), keeping it frozen during training. In the sound-text module, we use the pretrained CLIP textual encoder as \\( E_{t}(x) \\) to encode both semantic text and sound captions. The output dimension of the sound-aligned Ego-Exo representation space is 256. We use Llama3-8B as the LLM to generate sound captions. In the multi-head cross-attention module, we set the number of heads to \\( h = 4 \\). The temperature parameter \\( \\tau \\) is set to 0.07. We utilize 8 V100 GPUs to train the sound-aware representation space with a learning rate of 3e-5, using AdamW as the optimizer for improved convergence and stability." + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.816, + 0.486, + 0.843 + ], + "angle": 0, + "content": "It is worth noting that for video pairs with high synchronization, applying the Dynamic Time Warping (DTW) algo" + }, + { + "type": "table_caption", + "bbox": [ + 0.51, + 0.15, + 0.846, + 0.174 + ], + "angle": 0, + "content": "Table 1. Retrieval result comparison (top-1 accuracy) of Exo-Ego video alignment task on two benchmarks." + }, + { + "type": "table", + "bbox": [ + 0.513, + 0.183, + 0.847, + 0.304 + ], + "angle": 0, + "content": "
MethodEgoExoLearnCharadesEgo
Eg2ExEx2EgAvgEg2ExEx2EgAvg
FrozenInTime [3]14.1013.4013.756.734.965.85
LaViLa [36]28.7025.7027.2058.2446.9352.59
EgoVLP [17]32.1028.930.5058.2760.7659.52
InternVideo [32]30.6021.7026.1553.5561.2357.39
EgoExo [12]49.0045.3047.1568.3262.0665.19
EgoInstructor [34]47.1045.8246.4662.0656.3859.22
SoundBridge50.2750.7350.5077.6671.6374.65
" + }, + { + "type": "text", + "bbox": [ + 0.51, + 0.314, + 0.846, + 0.378 + ], + "angle": 0, + "content": "rhythm [24] to align the two audio sequences before inputting the audio signals into the cross-attention module can help mitigate the impact of varying sound speeds. We use this setting by default for videos in CharadesEgo and provide an ablation study on DTW in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.511, + 0.391, + 0.655, + 0.403 + ], + "angle": 0, + "content": "5.1.3 Baseline Models" + }, + { + "type": "text", + "bbox": [ + 0.51, + 0.41, + 0.846, + 0.558 + ], + "angle": 0, + "content": "We include several types of models as baselines. These include recent text-video representation models pre-trained on large-scale exocentric datasets, such as FrozenInTime [3] and InternVideo [32]. Additionally, we include pretrained models on large-scale egocentric datasets, namely LaViLa [36] and EgoVLP [17]. Furthermore, we compare our method with two recent approaches designed for EgoExo video alignment, namely EgoExo [12] and EgoInstructor [34]. For the downstream tasks, we also include task-specific models, namely VideoMAE [30] for exocentric action recognition and InternImage [31] for egocentric scene recognition." + }, + { + "type": "title", + "bbox": [ + 0.511, + 0.565, + 0.785, + 0.58 + ], + "angle": 0, + "content": "5.2. Results on Ego-Exo Video Alignment" + }, + { + "type": "text", + "bbox": [ + 0.51, + 0.584, + 0.847, + 0.843 + ], + "angle": 0, + "content": "Table 1 compares the results of Ego-to-Exo and Exo-to-Ego video retrieval with other baselines. Specifically, we report the zero-shot retrieval performance of the FrozenIn-Time model and the fine-tuned results of other methods on the EgoExoLearn and CharadesEgo datasets. Our proposed method, SoundBridge, consistently outperforms other baselines, including pre-trained egocentric models, pre-trained exocentric models, and recent Ego-Exo models, on both Ego-to-Exo and Exo-to-Ego video retrieval tasks across the two benchmarks. We are the first model to achieve a top-1 accuracy exceeding \\(50\\%\\) on the EgoExoLearn dataset. Specifically, SoundBridge achieves higher retrieval accuracy than EgoExo [12] and EgoInstructor [34] by \\(8.13\\%\\) and \\(15.93\\%\\) on Ego-to-Exo video retrieval, and by \\(13.71\\%\\) and \\(18.88\\%\\) on Exo-to-Ego video retrieval, respectively. The performance boost primarily comes from the incorporation of sound-aware modules in the alignment process. For example, given an egocentric video query recording picking up the garlic with the left hand, the EgoExo model ranks an exocentric video of put the ginger down with both hand on the iron bowl in the top-1 position. In contrast, our" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.915, + 0.521, + 0.927 + ], + "angle": 0, + "content": "28946" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.167, + 0.149, + 0.831, + 0.301 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.15, + 0.313, + 0.846, + 0.337 + ], + "angle": 0, + "content": "Figure 3. Visualization of an Ego-to-Exo video retrieval example, compared with EgoExo [12] and EgoInstructor [34], on an egocentric video query looking for A person is working on a vacuum in the man cave, then sneezes and reaches for some medicine." + }, + { + "type": "image", + "bbox": [ + 0.168, + 0.348, + 0.83, + 0.5 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.15, + 0.512, + 0.846, + 0.536 + ], + "angle": 0, + "content": "Figure 4. Visualization of an Exo-to-Ego video retrieval example, compared with EgoExo [12] and EgoInstructor [34], on an exocentric video query looking for a person throws a box of groceries onto the table. The person opens the box and takes out an item." + }, + { + "type": "text", + "bbox": [ + 0.15, + 0.558, + 0.486, + 0.656 + ], + "angle": 0, + "content": "method differentiates between the actions \"pick-up\" and \"put down\" using sound cues. The pick-up action produces a soft rustling sound, whereas the put-down action generates a clunk sound when interacting with the iron bowl. Our proposed sound-aware modules effectively encode these audio features and their semantic descriptions (e.g., [pick up ginger] (soft rustling)). As a result, our method correctly ranks the ground-truth exocentric video in the top-1 position." + }, + { + "type": "text", + "bbox": [ + 0.15, + 0.657, + 0.487, + 0.84 + ], + "angle": 0, + "content": "Figure 3 visualizes an Ego-to-Exo video retrieval example, comparing our method with EgoExo [12] and EgoInstructor [34]. The egocentric video depicts a person working on a vacuum, sneezing, and picking up a medicine bottle, while the visual cue for the sneeze is barely visible in the egocentric video. All methods successfully rank an egocentric video of a person with a vacuum in the top-1 position. However, EgoExo and EgoInstructor fail to capture either the sneezing action or the action of reaching for the medicine, leading to incorrect top-1 results. In contrast, our method effectively detects the sneeze and the action of picking up the pill bottle using sound cues, thereby successfully retrieving the ground-truth exocentric video. Similarly, Figure 4 visualizes an Exo-to-Ego retrieval example with a query for: a person throws a box of groceries onto" + }, + { + "type": "text", + "bbox": [ + 0.51, + 0.558, + 0.845, + 0.631 + ], + "angle": 0, + "content": "the table. The person opens the box and takes out an item. All methods successfully retrieve videos that capture the actions of opening the box and taking out items. However, only our SoundBridge method correctly identifies the action of throwing the box onto the table, leveraging the thud sound as an essential cue." + }, + { + "type": "title", + "bbox": [ + 0.511, + 0.638, + 0.75, + 0.652 + ], + "angle": 0, + "content": "5.3. Results on Text-Video Retrieval" + }, + { + "type": "text", + "bbox": [ + 0.509, + 0.658, + 0.847, + 0.843 + ], + "angle": 0, + "content": "The proposed sound-aware Ego-Exo representation space also enables text-video retrieval. Therefore, we compare its performance with baselines on both ego-(exo-)video-to-text and text-to-ego-(exo-)video retrieval tasks. Table 2 demonstrates that our method significantly enhances text-video retrieval performance on Ego-Exo datasets. Specifically, our method substantially improves the performance of the pre-trained egocentric model LaViLa [36], and the pre-trained exocentric model InternVideo [32]. Furthermore, our method outperforms recent Ego-Exo models, namely EgoExo [12] and EgoInstructor [34], by approximately \\(30.50\\%\\) and \\(23.47\\%\\) in overall performance across both datasets. The main improvement over recent Ego-Exo models is observed in queries with multiple nearest-neighbor (NN) answers. By incorporating sound aware" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.915, + 0.519, + 0.927 + ], + "angle": 0, + "content": "28947" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.162, + 0.15, + 0.474, + 0.162 + ], + "angle": 0, + "content": "Table 2. Performance comparison on text-video retrieval task." + }, + { + "type": "table", + "bbox": [ + 0.153, + 0.171, + 0.487, + 0.298 + ], + "angle": 0, + "content": "
MethodEgoExoLearnCharadesEgo
T2VV2TAvgT2VV2TAvg
FrozenInTime [3]6.034.055.049.405.627.51
LaViLa [36]34.8740.1837.5343.7441.6742.71
EgoVLP [17]19.0926.1022.6045.5143.2144.36
InternVideo [32]18.5926.5522.5747.3446.6346.99
EgoExo [12]58.6458.7358.6948.0045.2146.61
EgoInstructor [34]66.3766.6866.5346.9345.9846.46
SoundBridge71.4173.1472.2864.9063.6064.25
" + }, + { + "type": "text", + "bbox": [ + 0.15, + 0.308, + 0.487, + 0.432 + ], + "angle": 0, + "content": "ness, our method effectively ranks the ground-truth videos higher among similar nearest-neighbor results. For example, given a text query searching for exocentric videos of a person washing dishes in the kitchen, EgoExo retrieves videos of a person chopping potatoes, while EgoInstruc-tor retrieves videos of a person stir-frying. In contrast, our method distinguishes the washing action among these nearest neighbors by leveraging water-running sounds, allowing the ground-truth video to be ranked higher than videos of chopping and stir-frying." + }, + { + "type": "title", + "bbox": [ + 0.152, + 0.443, + 0.379, + 0.456 + ], + "angle": 0, + "content": "5.4. Results on Downstream Tasks" + }, + { + "type": "text", + "bbox": [ + 0.15, + 0.462, + 0.486, + 0.572 + ], + "angle": 0, + "content": "We also evaluate the effectiveness of the learned egocentric visual feature (i.e., \\( f_{sv}^{ego} \\)) for scene recognition and the learned exocentric visual feature (i.e., \\( f_{sv}^{exo} \\)) for action recognition. We evaluate our model under two settings: (1) fine-tuning, where the training and test sets come from the same dataset, and (2) transfer learning, where the model is trained on one dataset and tested on another. The report performances of other models are all with the fine-tuning setting." + }, + { + "type": "text", + "bbox": [ + 0.15, + 0.572, + 0.487, + 0.842 + ], + "angle": 0, + "content": "Table 3 compares the top-1 recognition accuracy with baseline methods. Specifically, for the egocentric scene recognition task, our method improves upon the egocentric model LaViLa [36] and the Ego-Exo model EgoInstructor [34] by \\(23.42\\%\\) and \\(9.78\\%\\), respectively. Figure 5 illustrates an example of scene recognition. Given an egocentric video query where only a hand wiping a table with a towel is visible, it is difficult to infer that the scene is a kitchen. Other methods fail to predict the correct scene label. For example, FrozenInTime [3] misclassifies the video as a living room, while LaViLa [36] predicts it as a bathroom. In contrast, our method correctly identifies the scene by successfully aligning the query video with the corresponding exocentric video using sound cues from multiple actions (e.g., walking to the table and leaving the room). The egocentric visual feature (i.e., \\(f_{sv}^{ego}\\)) learns environmental and global visual cues from the exocentric video, which contains a range hood, an oven, and a refrigerator, allowing it to infer that the scene is a kitchen. For exocentric action recognition, our method achieves an average accuracy of \\(56.25\\%\\), which is lower than VideoMAE [30], as the model is trained on a large-scale action dataset. Despite this, our" + }, + { + "type": "image", + "bbox": [ + 0.514, + 0.151, + 0.843, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.51, + 0.227, + 0.846, + 0.251 + ], + "angle": 0, + "content": "Figure 5. Visualization of the egocentric scene recognition results compared with FrozenInTime [3] and LaViLa [36]." + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.264, + 0.842, + 0.331 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.51, + 0.34, + 0.846, + 0.365 + ], + "angle": 0, + "content": "Figure 6. Visualization of the exocentric action recognition results compared with InternVideo [32] and EgoInstructor [34]." + }, + { + "type": "table_caption", + "bbox": [ + 0.51, + 0.376, + 0.846, + 0.423 + ], + "angle": 0, + "content": "Table 3. Result comparison on CharadesEgo scene recognition and EgoExoLearn action recognition. An asterisk (*) indicates the transfer learning setting, where a different dataset is used for training." + }, + { + "type": "table", + "bbox": [ + 0.513, + 0.432, + 0.846, + 0.575 + ], + "angle": 0, + "content": "
MethodEgo-SceneExo-Action
FrozenInTime [3]8.256.45
LaViLa [36]41.3726.25
EgoVLP [17]40.3138.75
InternVideo [32]43.0045.00
EgoExo [12]42.5546.25
EgoInstructor [34]46.5147.50
VideoMAEv2 [30]-62.50
InterImage [31]39.60-
SoundBridge*48.7652.50
SoundBridge51.0656.25
" + }, + { + "type": "text", + "bbox": [ + 0.509, + 0.585, + 0.847, + 0.721 + ], + "angle": 0, + "content": "approach benefits from the sound-aware Exo-Ego representation space, as shown in Figure 6. Predicting a person's action from an exocentric view is challenging due to occlusions caused by the human body and the person being far from the camera. InternVideo [32] misclassifies the query video as the action \"put\", while EgoInstructor [34] assigns a high probability to the action \"pick-up\". In contrast, our method aligns the query video with its corresponding egocentric view using the sound of washing dishes (e.g., water splashing) and correctly classifies the action as \"washing\" with the highest probability." + }, + { + "type": "title", + "bbox": [ + 0.511, + 0.728, + 0.64, + 0.742 + ], + "angle": 0, + "content": "5.5. Ablation study" + }, + { + "type": "title", + "bbox": [ + 0.511, + 0.747, + 0.718, + 0.76 + ], + "angle": 0, + "content": "5.5.1 Results on Different LLMs" + }, + { + "type": "text", + "bbox": [ + 0.509, + 0.767, + 0.845, + 0.842 + ], + "angle": 0, + "content": "We evaluated the impact of audio descriptions generated by different Large Language Models (LLMs) on the performance of cross-view video association tasks. Specifically, we used GPT-3.5 [4] and LLaMA3 [7] to generate audio descriptions using the same prompt. As shown in Table 5, the results indicate that while the cross-view video" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.915, + 0.521, + 0.927 + ], + "angle": 0, + "content": "28948" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.3, + 0.15, + 0.697, + 0.163 + ], + "angle": 0, + "content": "Table 4. Ablation study on the proposed sound-vision and sound-text modules." + }, + { + "type": "table", + "bbox": [ + 0.155, + 0.172, + 0.846, + 0.254 + ], + "angle": 0, + "content": "
MethodEgoExoLearnCharadesEgo
Eg2ExEx2EgT2VV2TEg2ExEx2EgT2VV2T
Baseline49.0045.3058.6458.7368.3262.0648.0045.21
Baseline + Sound-Text Module48.8250.1869.2371.9672.8165.2563.5561.35
Baseline + Sound-Vision Module48.4647.7369.1073.0974.3566.5563.5961.11
Baseline + Sound-Text&Vision Modules50.2750.7371.4173.1477.6671.6364.9063.60
" + }, + { + "type": "table_caption", + "bbox": [ + 0.222, + 0.273, + 0.414, + 0.284 + ], + "angle": 0, + "content": "Table 5. Results with different LLMs." + }, + { + "type": "table", + "bbox": [ + 0.154, + 0.294, + 0.486, + 0.354 + ], + "angle": 0, + "content": "
LLMEgoExoLearnCharadesEgo
Eg2ExEx2EgAvgEg2ExEx2EgAvg
GPT-3.549.2748.3648.8273.7671.5172.64
LLaMA350.2750.7350.5077.6671.6374.65
" + }, + { + "type": "table_caption", + "bbox": [ + 0.18, + 0.365, + 0.457, + 0.377 + ], + "angle": 0, + "content": "Table 6. Performance comparison of different anchors." + }, + { + "type": "table", + "bbox": [ + 0.154, + 0.386, + 0.486, + 0.461 + ], + "angle": 0, + "content": "
AnchorEgoExoLearnCharadesEgo
Eg2ExEx2EgEg2ExEx2Eg
Origin Caption48.6448.7374.3566.55
Sound Caption47.0949.9156.9751.42
Sound Feature36.6726.6756.2748.38
Sound-Text Module50.2750.7377.6671.63
" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.47, + 0.486, + 0.52 + ], + "angle": 0, + "content": "association performance with GPT-3.5 is lower than that achieved with LLaMA3, our method still surpasses existing approaches. This demonstrates the effectiveness of our proposed method." + }, + { + "type": "title", + "bbox": [ + 0.152, + 0.538, + 0.414, + 0.55 + ], + "angle": 0, + "content": "5.5.2 Impact of the Sound-aware Modules" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.558, + 0.486, + 0.68 + ], + "angle": 0, + "content": "We conduct an ablation study on the proposed sound-vision and sound-text modules, using the state-of-the-art Ego-Exo model EgoExo [12] as the baseline. The performance of Ego-Exo video alignment and the overall accuracy of text-video retrieval are presented in Table 4. When using only the sound-vision or sound-text module, our method demonstrates competitive performance in the Ego-to-Exo video retrieval task. However, when both modules are jointly utilized, our method achieves significantly higher performance across the evaluated retrieval tasks." + }, + { + "type": "title", + "bbox": [ + 0.152, + 0.698, + 0.367, + 0.71 + ], + "angle": 0, + "content": "5.5.3 Results of Different Anchors" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.718, + 0.486, + 0.842 + ], + "angle": 0, + "content": "We also investigate the effectiveness of using different features as anchors to align sound-aware visual representations. Specifically, we replace the \\(f_{st}\\) in Eq. (11) and Eq. (12) with \\(f_{t}, f_{c}\\) or \\(\\text{sum}(f_{a}^{ego}, f_{a}^{exo})\\). As shown in Table 6, using audio alone as an anchor is not ideal, especially for the short video dataset EgoExoLearn, where events heavily overlap and exhibit only slight variations. Upon analyzing the erroneous cases, we found that in many samples, the audio anchor alone fails to provide sufficient information to distinguish between videos from different views. As il" + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.273, + 0.84, + 0.366 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.51, + 0.376, + 0.846, + 0.4 + ], + "angle": 0, + "content": "Figure 7. An example of Exo-Ego video retrieval using different anchors in the SoundBridge model." + }, + { + "type": "text", + "bbox": [ + 0.51, + 0.412, + 0.845, + 0.522 + ], + "angle": 0, + "content": "lustrated in Figure 7, a person is slicing garlic in the query video, and the text describes the semantic information of \"a person slicing garlic\". However, the sound features only convey the action \"slice\". During training, different-view videos were aligned solely based on the \"slice\" action, leading to incorrect retrieval of a video showing \"slicing ginger\" instead of \"slicing garlic\". Similarly, the sound caption lacks full contextual information and misinterprets the one-hand slicing action as two-hand slicing." + }, + { + "type": "title", + "bbox": [ + 0.511, + 0.538, + 0.613, + 0.551 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.51, + 0.559, + 0.846, + 0.718 + ], + "angle": 0, + "content": "In this paper, we proposed two sound-aware modules as a bridge to address the large visual variance challenge in egocentric and exocentric video alignment. Both modules demonstrated significant improvements over state-of-the-art methods in Ego-Exo video retrieval and text-video retrieval tasks. Moreover, we verified that both egocentric and exocentric representations benefit from the sound-aware learning paradigm, achieving substantial improvements in downstream recognition tasks by leveraging complementary information from a correctly associated view. Furthermore, we found that our sound-vision and sound-text modules provide effective cues to distinguish the target video from its nearest neighbors based on sound." + }, + { + "type": "title", + "bbox": [ + 0.511, + 0.734, + 0.655, + 0.749 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.51, + 0.755, + 0.846, + 0.842 + ], + "angle": 0, + "content": "This research was supported by the National Natural Science Foundation of China (Grant No.: 62372314, 62476097); the Guangdong Provincial Fund for Basic and Applied Basic Research—Regional Joint Fund Project (Key Project) (2023B1515120078); the Guangdong Provincial Natural Science Foundation for Outstanding Youth Team Project (2024B1515040010)." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.915, + 0.519, + 0.927 + ], + "angle": 0, + "content": "28949" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.154, + 0.151, + 0.234, + 0.164 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.161, + 0.172, + 0.486, + 0.222 + ], + "angle": 0, + "content": "[1] Karteek Alahari. Actor and observer: Joint modeling of first and third-person videos. In Proceedings of the 1st Workshop and Challenge on Comprehensive Video Understanding in the Wild, page 3, 2018. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.161, + 0.223, + 0.486, + 0.271 + ], + "angle": 0, + "content": "[2] Relja Arandjelovic and Andrew Zisserman. Look, listen and learn. In Proceedings of the IEEE international conference on computer vision, pages 609-617, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.161, + 0.272, + 0.487, + 0.333 + ], + "angle": 0, + "content": "[3] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1728-1738, 2021. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.161, + 0.334, + 0.486, + 0.407 + ], + "angle": 0, + "content": "[4] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.161, + 0.408, + 0.486, + 0.457 + ], + "angle": 0, + "content": "[5] Minjie Cai, Kris M Kitani, and Yoichi Sato. Understanding hand-object manipulation with grasp types and object attributes. In Robotics: Science and Systems, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.161, + 0.458, + 0.486, + 0.508 + ], + "angle": 0, + "content": "[6] Sanyuan Chen, Yu Wu, Chengyi Wang, Shujie Liu, Daniel Tompkins, Zhuo Chen, and Furu Wei. Beats: Audio pre-training with acoustic tokenizers. arXiv preprint arXiv:2212.09058, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.161, + 0.509, + 0.486, + 0.569 + ], + "angle": 0, + "content": "[7] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.161, + 0.57, + 0.486, + 0.632 + ], + "angle": 0, + "content": "[8] Antonino Furnari and Giovanni Maria Farinella. What would you expect? anticipating egocentric actions with rolling-unrolling lstms and modality attention. In Proceedings of the IEEE/CVF International conference on computer vision, pages 6252-6261, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.161, + 0.633, + 0.486, + 0.693 + ], + "angle": 0, + "content": "[9] Mohit Goyal, Sahil Modi, Rishabh Goyal, and Saurabh Gupta. Human hands as probes for interactive object understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3293-3303, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.154, + 0.694, + 0.487, + 0.842 + ], + "angle": 0, + "content": "[10] Kristen Grauman, Andrew Westbury, Eugene Byrne, Zachary Chavis, Antonino Furnari, Rohit Girdhar, Jackson Hamburger, Hao Jiang, Miao Liu, Xingyu Liu, Miguel Martin, Tushar Nagarajan, Iija Radosavovic, Santhosh Kumar Ramakrishnan, Fiona Ryan, Jayant Sharma, Michael Wray, Mengmeng Xu, Eric Zhongcong Xu, Chen Zhao, Siddhant Bansal, Dhruv Batra, Vincent Cartillier, Sean Crane, Tien Do, Morrie Doulaty, Akshay Erapalli, Christoph Feichtenhofer, Adriano Fragomeni, Qichen Fu, Abrham Gebreselasie, Cristina Gonzalez, James Hillis, Xuhua Huang, Yifei Huang, Wenqi Jia, Weslie Khoo, Jachym" + }, + { + "type": "list", + "bbox": [ + 0.154, + 0.172, + 0.487, + 0.842 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.543, + 0.153, + 0.845, + 0.383 + ], + "angle": 0, + "content": "Kolář, Satwik Kottur, Anurag Kumar, Federico Landini, Chao Li, Yanghao Li, Zhenqiang Li, Karttikeya Mangalam, Raghava Modhugu, Jonathan Munro, Tullie Murrell, Takumi Nishiyasu, Will Price, Paola Ruiz, Merey Ramazanova, Leda Sari, Kiran Somasundaram, Audrey Southerland, Yusuke Sugano, Ruijie Tao, Minh Vo, Yuchen Wang, Xindi Wu, Takuma Yagi, Ziwei Zhao, Yunyi Zhu, Pablo Arbeláez, David Crandall, Dima Damen, Giovanni Maria Farinella, Christian Fuegen, Bernard Ghanem, Vamsi Krishna Ithapu, C. V. Jawahar, Hanbyul Joo, Kris Kitani, Haizhou Li, Richard Newcombe, Aude Oliva, Hyun Soo Park, James M. Rehg, Yoichi Sato, Jianbo Shi, Mike Zheng Shou, Antonio Torralba, Lorenzo Torresani, Mingfei Yan, and Jitendra Malik. Ego4d: Around the world in 3,000 hours of egocentric video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18995-19012, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.514, + 0.385, + 0.845, + 0.459 + ], + "angle": 0, + "content": "[11] Bo He, Jun Wang, Jielin Qiu, Trung Bui, Abhinav Shrivastava, and Zhaowen Wang. Align and attend: Multimodal summarization with dual contrastive losses. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14867-14878, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.514, + 0.46, + 0.845, + 0.557 + ], + "angle": 0, + "content": "[12] Yifei Huang, Guo Chen, Jilan Xu, Mingfang Zhang, Lijin Yang, Baoqi Pei, Hongjie Zhang, Lu Dong, Yali Wang, Limin Wang, et al. Egoxolearn: A dataset for bridging asynchronous ego-and exo-centric view of procedural activities in real world. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22072-22086, 2024. 1, 2, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.514, + 0.558, + 0.844, + 0.631 + ], + "angle": 0, + "content": "[13] Sarah Ibrahimi, Xiaohang Sun, Pichao Wang, Amanmeet Garg, Ashutosh Sanan, and Mohamed Omar. Audio-enhanced text-to-video retrieval using text-conditioned feature alignment. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12054-12064, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.514, + 0.632, + 0.844, + 0.68 + ], + "angle": 0, + "content": "[14] Hao Jiang and Vamsi Krishna Ithapu. Egocentric pose estimation from human vision span. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 10986-10994. IEEE, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.514, + 0.681, + 0.844, + 0.73 + ], + "angle": 0, + "content": "[15] Patricia S Koskinen, Robert M Wilson, and Carl J Jensema. Using closed-captioned television in the teaching of reading to deaf students. American Annals of the Deaf, 131(1):43-46, 1986. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.514, + 0.731, + 0.845, + 0.791 + ], + "angle": 0, + "content": "[16] Haoxin Li, Yijun Cai, and Wei-Shi Zheng. Deep dual relation modeling for egocentric interaction recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7932-7941, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.514, + 0.792, + 0.845, + 0.842 + ], + "angle": 0, + "content": "[17] Kevin Qinghong Lin, Jinpeng Wang, Mattia Soldan, Michael Wray, Rui Yan, Eric Z Xu, Difei Gao, Rong-Cheng Tu, Wenzhe Zhao, Weijie Kong, et al. Ego-centric video-language pretraining. Advances in Neu" + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.153, + 0.845, + 0.842 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.915, + 0.519, + 0.927 + ], + "angle": 0, + "content": "28950" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.184, + 0.153, + 0.485, + 0.176 + ], + "angle": 0, + "content": "ral Information Processing Systems, 35:7575-7586, 2022. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.156, + 0.178, + 0.484, + 0.226 + ], + "angle": 0, + "content": "[18] Yan-Bo Lin, Jie Lei, Mohit Bansal, and Gedas Bertasius. Eclipse: Efficient long-range video retrieval using sight and sound. In European Conference on Computer Vision, pages 413–430. Springer, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.155, + 0.227, + 0.485, + 0.299 + ], + "angle": 0, + "content": "[19] Shaowei Liu, Subarna Tripathi, Somdeb Majumdar, and Xiaolong Wang. Joint hand motion and interaction hotspots prediction from egocentric videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3282-3292, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.155, + 0.301, + 0.487, + 0.361 + ], + "angle": 0, + "content": "[20] Sagnik Majumder, Ziad Al-Halah, and Kristen Grauman. Learning spatial features from audio-visual correspondence in egocentric videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 27058-27068, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.155, + 0.362, + 0.485, + 0.422 + ], + "angle": 0, + "content": "[21] Tushar Nagarajan, Yanghao Li, Christoph Feichtenhofer, and Kristen Grauman. Ego-topo: Environment affordances from egocentric video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 163-172, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.155, + 0.423, + 0.485, + 0.508 + ], + "angle": 0, + "content": "[22] Chiara Plizzari, Mirco Planamente, Gabriele Goletto, Marco Cannici, Emanuele Gusso, Matteo Matteucci, and Barbara Caputo. E2 (go) motion: Motion augmented event stream for egocentric action recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 19935-19947, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.155, + 0.51, + 0.485, + 0.582 + ], + "angle": 0, + "content": "[23] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.155, + 0.583, + 0.485, + 0.631 + ], + "angle": 0, + "content": "[24] Hiroaki Sakoe and Seibi Chiba. Dynamic programming algorithm optimization for spoken word recognition. IEEE transactions on acoustics, speech, and signal processing, 26(1):43-49, 1978. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.155, + 0.632, + 0.485, + 0.692 + ], + "angle": 0, + "content": "[25] Muhammad Bilal Shaikh, Douglas Chai, Syed Mohammed Shamsul Islam, and Naveed Akhtar. Multimodal fusion for audio-image and video action recognition. Neural Computing and Applications, 36(10): 5499-5513, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.155, + 0.693, + 0.485, + 0.755 + ], + "angle": 0, + "content": "[26] Dandan Shan, Jiaqi Geng, Michelle Shu, and David F Fouhey. Understanding human hands in contact at internet scale. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9869-9878, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.155, + 0.756, + 0.485, + 0.815 + ], + "angle": 0, + "content": "[27] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.155, + 0.817, + 0.485, + 0.841 + ], + "angle": 0, + "content": "[28] Huiyu Wang, Mitesh Kumar Singh, and Lorenzo Torresani. Ego-only: Egocentric action detection with" + }, + { + "type": "list", + "bbox": [ + 0.155, + 0.153, + 0.487, + 0.841 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.543, + 0.154, + 0.845, + 0.189 + ], + "angle": 0, + "content": "out exocentric transferring. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5250-5261, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.515, + 0.191, + 0.845, + 0.263 + ], + "angle": 0, + "content": "[29] Jian Wang, Diogo Luvizon, Weipeng Xu, Lingjie Liu, Kripasindhu Sarkar, and Christian Theobalt. Scene-aware egocentric 3d human pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13031-13040, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.514, + 0.265, + 0.844, + 0.338 + ], + "angle": 0, + "content": "[30] Limin Wang, Bingkun Huang, Zhiyu Zhao, Zhan Tong, Yinan He, Yi Wang, Yali Wang, and Yu Qiao. Videomae v2: Scaling video masked autoencoders with dual masking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14549-14560, 2023. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.514, + 0.339, + 0.844, + 0.424 + ], + "angle": 0, + "content": "[31] Wenhai Wang, Jifeng Dai, Zhe Chen, Zhenhang Huang, Zhiqi Li, Xizhou Zhu, Xiaowei Hu, Tong Lu, Lewei Lu, Hongsheng Li, et al. Internimage: Exploring large-scale vision foundation models with deformable convolutions. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14408-14419, 2023. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.514, + 0.426, + 0.844, + 0.511 + ], + "angle": 0, + "content": "[32] Yi Wang, Yinan He, Yizhuo Li, Kunchang Li, Jiashuo Yu, Xin Ma, Xinhao Li, Guo Chen, Xinyuan Chen, Yaohui Wang, Ping Luo, Ziwei Liu, Yali Wang, Limin Wang, and Yu Qiao. Internvid: A large-scale videotext dataset for multimodal understanding and generation. In The Twelfth International Conference on Learning Representations, 2024. 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.514, + 0.512, + 0.844, + 0.572 + ], + "angle": 0, + "content": "[33] Jianjia Xin, Lichun Wang, Kai Xu, Chao Yang, and Baocai Yin. Learning interaction regions and motion trajectories simultaneously from egocentric demonstration videos. IEEE Robotics and Automation Letters, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.514, + 0.574, + 0.844, + 0.647 + ], + "angle": 0, + "content": "[34] Jilan Xu, Yifei Huang, Junlin Hou, Guo Chen, Yuejie Zhang, Rui Feng, and Weidi Xie. Retrieval-augmented egocentric video captioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13525-13536, 2024. 1, 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.514, + 0.649, + 0.844, + 0.708 + ], + "angle": 0, + "content": "[35] Zihui Sherry Xue and Kristen Grauman. Learning fine-grained view-invariant representations from unpaired ego-exo videos via temporal alignment. Advances in Neural Information Processing Systems, 36: 53688-53710, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.514, + 0.71, + 0.844, + 0.772 + ], + "angle": 0, + "content": "[36] Yue Zhao, Ishan Misra, Philipp Krahenbuhl, and Rohit Girdhar. Learning video representations from large language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6586-6597, 2023. 5, 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.154, + 0.845, + 0.772 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.915, + 0.518, + 0.927 + ], + "angle": 0, + "content": "28951" + } + ] +] \ No newline at end of file diff --git a/2025/Sound Bridge_ Associating Egocentric and Exocentric Videos via Audio Cues/37c295df-9911-4189-b46a-8795fdcf6a6f_origin.pdf b/2025/Sound Bridge_ Associating Egocentric and Exocentric Videos via Audio Cues/37c295df-9911-4189-b46a-8795fdcf6a6f_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..b769e1ec703b448573540e538ef625f58f407694 --- /dev/null +++ b/2025/Sound Bridge_ Associating Egocentric and Exocentric Videos via Audio Cues/37c295df-9911-4189-b46a-8795fdcf6a6f_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7ca3cb327008b19def1b95f42f043655d7b888804ecc1e980ff787382dc3ba27 +size 952489 diff --git a/2025/Sound Bridge_ Associating Egocentric and Exocentric Videos via Audio Cues/full.md b/2025/Sound Bridge_ Associating Egocentric and Exocentric Videos via Audio Cues/full.md new file mode 100644 index 0000000000000000000000000000000000000000..7e6902c8698d90dfd8ecbcbeb855208a4139c3e3 --- /dev/null +++ b/2025/Sound Bridge_ Associating Egocentric and Exocentric Videos via Audio Cues/full.md @@ -0,0 +1,303 @@ +# Sound Bridge: Associating Egocentric and Exocentric Videos via Audio Cues + +Sihong Huang $^{1,5,*}$ , Jiaxin Wu $^{2,*}$ , Xiaoyong Wei $^{3,2,5,\dagger}$ , Yi Cai $^{1,\dagger}$ , Dongmei Jiang $^{5}$ , Yaowei Wang $^{4,5}$ + +1South China University of Technology, 2The Hong Kong Polytechnic University + +$^{3}$ Sichuan University, $^{4}$ Harbin Institute of Technology, Shenzhen, $^{5}$ Peng Cheng Laboratory + +{fthuangsh, ycai}@mail.scut.edu.cn, jiaxwu@polyu.edu.hk, cswei@scu.edu.cn + +jiangdm@pcl.ac.cn, wangyaowei@hit.edu.cn + +# Abstract + +Understanding human behavior and environmental information in egocentric videos is very challenging due to the invisibility of some actions (e.g., laughing and sneezing) and the local nature of the first-person view. Leveraging the corresponding exocentric video to provide global context has shown promising results. However, existing visual-to-visual and visual-to-textual Ego-Exo video alignment methods struggle with the issue that some activities may have non-visual overlap. To address this, we propose using sound as a bridge, as audio is often consistent across Ego-Exo videos. However, direct audio-to-audio alignment lacks context. Thus, we introduce two context-aware sound modules: one aligns audio with vision via a visual-audio cross-attention module, and another aligns text with sound closed caption generated by LLM. Experimental results on two Ego-Exo video association benchmarks show that each of the proposed modules enhances the state-of-the-art methods. Moreover, the proposed sound-aware egocentric or exocentric representation boosts the performance of downstream tasks, such as action recognition of exocentric videos and scene recognition of egocentric videos. The code and models can be accessed at https://github.com/shhuangcoder/SoundBridge. + +# 1. Introduction + +Egocentric video analysis has been receiving extensive attention recently [10, 12, 34] due to its potential applications in robotic vision analysis, embedded AI, etc. Compared to the large-scale and in-depth studies on exocentric video analysis tasks, research on egocentric videos is still underexplored. Moreover, there are significant differences between egocentric and exocentric videos; thus, many findings and methodologies derived from an exocentric perspective are not directly applicable to egocentric video analysis. + +![](images/f55b5b9896177bd7c98e9d7433355a7627dc23fbc84ec69561270c1c2a024903.jpg) +Figure 1. Given a video shooting script, egocentric and exocentric views can exhibit significant visual variances for the same activity, while their audio signals remain relatively consistent. + +The main difference between egocentric and exocentric videos is that they capture events from different perspectives. First-person view videos focus on local and detailed information about the current activity, whereas third-person view videos emphasize global and environmental context. An example is shown in Figure 1. Although egocentric and exocentric videos follow the same video shooting script, they exhibit significant visual variations. For example, in the case of the same activity, such as eating a snack, the egocentric view primarily captures hand-object interactions, such as a hand picking up a raisin, while the location remains unknown. In contrast, the exocentric view captures the overall scene of the event but lacks finer details—for example, it may show a person lying on a bed eating food, but the specific type of food remains unclear. Moreover, certain activities, such as laughing and sneezing, do not produce distinct visual cues in egocentric videos apart from camera shaking. Consequently, when attempting to align visual representations from exocentric to egocentric videos, there is often no overlapping visual content to match. + +However, recent studies [12, 34] have shown that exocentric videos can supplement the understanding of their semantically aligned egocentric counterparts. In this paper, we also focus on constructing an Ego-Exo alignment space to facilitate egocentric video analysis. Previous studies on + +egocentric and exocentric video alignment have primarily focused on matching visual content [35] or aligning both video types to an anchor (i.e., their semantic descriptions or a shared textual representation) [12, 34]. However, these alignment strategies may be insufficient due to the substantial visual variations between views. In contrast to the significant differences in view-dependent visual content, we observe that audio signals for the same event remain consistent across views. For example, in the case of laughing, the laughter sound—whether captured in an egocentric or exocentric video—can infer that someone is laughing. + +In this paper, we propose using audio signals as a bridge to model the egocentric-exocentric video alignment space. However, directly linking audio signals across videos without context may be suboptimal, as audio information alone is not sufficiently discriminative [2]. For example, the sound of a click associated with the action "switch off the light" in Figure 1 could also occur in other activities, such as "flipping a power switch on a household appliance" or "turning off a desk lamp". To address this limitation, we propose constructing contextual audio features for alignment by incorporating visual and textual context. An overview of our proposed method is presented in Figure 2. Specifically, we introduce sound-text and sound-vision modules to bridge Ego-Exo videos. The sound-vision module follows existing methods [13, 18] to build contextual audiovisual representations through cross-attention. Meanwhile, the sound-text module utilizes a large language model (LLM) to generate contextual audio descriptions and construct sound-aware textual representations. In particular, we provide the LLM with semantic text describing the video scenario and prompt it to generate an audio description of the actions occurring within the scene in a closed-caption format [15]. Additionally, to ensure that the generated sound captions emphasize unique information within the paired cross-view videos rather than commonly occurring content in the dataset, we filter out frequently appearing and less informative concepts (e.g., person) from the generated descriptions. As a result, the output sound caption for the example in Figure 1 is "[Eating the snack] (munching), [dishes on the floor] (crashing and scraping), [Laughing] (giggles), [leaving the room] (Footsteps), [turns off the light] (click)". The generated sound caption and the semantic text are subsequently processed through a cross-attention block to learn a sound-aware textual representation. Finally, we treat the textual representation as an anchor and align the audiovisual representations of egocentric and exocentric videos to it, thereby learning an Ego-Exo semantic feature space. We conduct extensive experiments on two Ego-Exo video retrieval benchmarks, namely EgoExoLearn [12] and CharadesEgo [1]. The results demonstrate that the learned sound-aware features outperform state-of-the-art approaches in both egocentric-exocentric video retrieval and text-to-egocentric/exocentric video retrieval. Moreover, the learned Ego- and Exo- features enhance the performance of downstream tasks, such as exocentric video action + +recognition and egocentric video scene recognition. + +# 2. Related Work + +# 2.1. Egocentric Video Analysis + +The number of research studies on egocentric video analysis is growing rapidly, as this task is a fundamental component of many applications [10, 12], such as robotic vision analysis and embedded AI. However, egocentric videos are fundamentally different from traditional exocentric videos. Egocentric videos capture a unique immersive viewpoint from a first-person perspective, directly reflecting an individual's interaction with the current activity, whereas exocentric videos provide an external view of human interactions within an environment. This distinct perspective in egocentric videos introduces a range of challenges for human activity analysis, including hand detection [5, 9, 26], action recognition [8, 22, 28], human-computer interaction [16, 21, 33], and applications in virtual and augmented reality [14, 19, 29]. Despite these differences, recent studies [12, 34] have demonstrated that egocentric video analysis can benefit from the contextual information provided by corresponding exocentric videos. For example, Xu et al. [34] show that related exocentric videos can enhance egocentric video captioning. Similarly, Huang et al. [12] demonstrate that the exocentric perspective provides complementary information that aids robots in performing complex tasks. In this paper, we also focus on developing effective Ego-Exo video alignment techniques to enhance egocentric video analysis. + +# 2.2. Ego-Exo Video Representation Learning + +The goal of Ego-Exo video representation learning is to construct a semantic space in which egocentric and exocentric videos with similar semantics are positioned closely, while dissimilar Ego-Exo videos are placed farther apart [12]. Early studies on Ego-Exo representation learning align cross-view videos using visual information. For example, AE2 [35] extracts visual regions of hands and active objects and performs temporal alignment on these extracted regions across views, assuming that the visual content of a fine-grained action remains consistent across perspectives. AE2 constructs four action-specific cross-view datasets for evaluation: break egg, pour milk, pour liquid, and tennis forehand. Recent approaches aim to learn a semantic Ego-Exo representation space by aligning videos to either a shared semantic text or separate textual descriptions. For example, EgoExo [12] employs implicit Ego-Exo alignment by associating each video with a view-dependent text, based on the assumption that semantically similar texts will be positioned closely in the learned space. Beyond implicit alignment, EgoExoNCE [34] takes this further by explicitly aligning Ego-Exo videos to a mutual text anchor. This shared text is generated by extracting key nouns and verbs from both the egocentric and exocentric textual descriptions. In this paper, we follow the pipeline of previous + +work by aligning Ego-Exo videos to a mutual text anchor. However, rather than directly matching entire videos with full-text descriptions without attention, we propose using audio as a bridge to address the challenge of visual variance. Our approach incorporates sound-vision and sound-text modules to bridge the gap between egocentric and exocentric perspectives. To the best of our knowledge, we are the first to integrate sound into Ego-Exo video alignment. + +# 2.3. Sound-aware Representation Learning + +Numerous studies have demonstrated that audio signals complement visual signals in video analysis tasks [18]. For example, the Eclipse method has shown that audio signals can serve as an alternative to visual signals in long-range video retrieval, significantly accelerating the search process [18]. Consequently, many approaches have sought to develop sound-aware representations. Several works have incorporated visual information as context to create audiovisual features for exocentric video analysis, applying them to tasks such as video retrieval [13, 18], action recognition [20], and video summarization [11]. More recently, audio-based methods have also been explored in egocentric video analysis, such as active speaker spatial detection [25]. Additionally, Ibrahimi et al. [13] integrate textual context to construct audio-textual representations. Common approaches to incorporating audio signals into video analysis include appending audio features to visual or textual features [20, 25] or learning audio-textual/visual cross-attention representations [13, 18]. Another method involves converting audio signals into text using Automatic Speech Recognition (ASR) and subsequently appending the ASR-derived text features to visual features [11]. Our proposed sound-vision model follows prior work by learning contextual audiovisual features through cross-attention. Additionally, we introduce a novel approach to learning audiotextual representations in our sound-text module, which leverages a large language model (LLM) to generate sound captions. + +# 3. Sound-aligned Ego-Exo Representation Learning + +The architecture of our proposed method is illustrated in Figure 2. The main objective is to construct a sound-aligned Ego-Exo representation space that integrates local (Ego) and global (Exo) visual features along with their corresponding semantic concepts. Specifically, to address the challenge of aligning semantic text with the significant visual variations present in Ego-Exo videos, we leverage sound as a unifying element. This is achieved by incorporating sound-vision and sound-text modules to generate sound-aware representations. Subsequently, we align the sound-aware video features with the sound-aware textual features to learn a unified representation space. + +# 3.1. Sound-Vision Module + +Given a semantically aligned egocentric and exocentric (ego-exo) video pair $(v^{ego}, v^{exo})$ and their corresponding audio signals $(a^{ego}, a^{exo})$ , we propose a sound-vision cross-attention module to learn their sound-aware visual representations. The sound-vision module consists of a video encoder $E_v(x)$ , an audio encoder $E_a(x)$ , and a multi-head cross-attention block Multihead $(Q, K, V)$ [27]. The module inputs a video $v$ and its corresponding audio signal $a$ and outputs sound-aware visual feature $f_{sv}$ as: + +$$ +f _ {v} = E _ {v} (v), f _ {a} = E _ {a} (a), \tag {1} +$$ + +$$ +C r o s s A B (Q, K, V) = S o f t m a x \left(\frac {Q \times K ^ {T}}{\sqrt {d _ {k}}}\right) \cdot V, \tag {2} +$$ + +$$ +\operatorname {M u l t i H e a d} (Q, K, V) = \operatorname {C o n c a t} \left(\text {h e a d} _ {1}, \dots , \text {h e a d} _ {h}\right) W ^ {O} \tag {3} +$$ + +$$ +f _ {s v} = \operatorname {M u l t i H e a d} \left(Q \left(f _ {a}\right), K \left(f _ {v}\right), V \left(f _ {v}\right)\right). \tag {4} +$$ + +where the $Q(x)$ , $K(x)$ , $V(x)$ map the input feature $x$ to a hidden space though linear projections and $h$ represents the number of head. The egocentric and exocentric sound-vision modules in Figure 2 share the same architecture, as illustrated in Eqs. (1)-(4), but with view-dependent weights. Specifically, we feed egocentric videos and exocentric videos and their audio signals to the sound-vision module to obtain sound-aware representations as: + +$$ +f _ {v} ^ {e g o} = E _ {v} ^ {e g o} \left(v ^ {e g o}\right), f _ {v} ^ {e x o} = E _ {v} ^ {e x o} \left(v ^ {e x o}\right). \tag {5} +$$ + +$$ +f _ {a} ^ {e g o} = E _ {a} \left(a ^ {e g o}\right), f _ {a} ^ {e x o} = E _ {a} \left(a ^ {e x o}\right). \tag {6} +$$ + +$$ +f _ {s v} ^ {e g o} = \text {M u l t i H e a d} \left(Q \left(f _ {a} ^ {e g o}\right), K \left(f _ {v} ^ {e g o}\right), V \left(f _ {v} ^ {e g o}\right)\right) \tag {7} +$$ + +$$ +f _ {s v} ^ {e x o} = \text {M u l t i H e a d} \left(Q \left(f _ {a} ^ {e x o}\right), K \left(f _ {v} ^ {e x o}\right), V \left(f _ {v} ^ {e x o}\right)\right) \tag {8} +$$ + +# 3.2. Sound-Text Module + +Given a text $t$ that captures the semantics of Ego-Exo videos, we use an LLM to generate a sound caption $c$ and propose a sound-text cross-attention module to learn a sound-aware text representation $f_{st}$ . The sound-text module shares the same architecture as the sound-vision module but uses different encoders: + +$$ +f _ {t} = E _ {t} (t), f _ {c} = E _ {t} (c), \tag {9} +$$ + +$$ +f _ {s t} = \text {M u l t i H e a d} \left(Q \left(f _ {c}\right), K \left(f _ {t}\right), V \left(f _ {t}\right)\right). \tag {10} +$$ + +Specifically, the sound caption $c$ is generated from the given text $t$ by an LLM. With $t$ providing the video context, we prompt the LLM to generate sound descriptions for all the actions in the videos as: $c = LLM(t, prompt_{Txt2Sound})$ . The sound description needs to contain the source of the sound and the sound effect, formatted in Closed Caption (CC) style [15]. Closed captions (CC) are a specialized type of caption designed for people who are deaf or hard of + +![](images/c29c5ed9232595f4f2ea2bdeaaaabf1ee848e54d437d4c748636ad6c99eabe7e.jpg) +Figure 2. Overview of the proposed method. We leverage sound as a bridge to connect Ego-Exo videos with their corresponding semantic text to build a sound-aligned representation space. + +hearing. They additionally include an audio description in parentheses and indicate the source of the sound in brackets. For example, for a given text $t =$ "a person turning off the light", the generated sound caption is $c =$ "[a person turning off the light] (click)". Furthermore, to ensure that the sound caption focuses on the unique information of the paired videos, we filter out frequently appearing concepts (e.g., "a person") in CC. To achieve this, we use the LLM to refine $c$ to $\tilde{c}$ by providing the frequently appeared concepts, i.e., $\tilde{c} = \text{LLM}(c, \text{frequentConcepts}, \text{prompt}_{\text{soundRefine}})$ . Due to space limitations, details of all prompts are provided in the supplementary material. We input the filtered and refined closed sound caption $\tilde{c}$ along with the text $t$ into the sound-text module to obtain the sound-aware text representation: $f_{st} = \text{MultiHead}(Q(f_{\tilde{c}}), K(f_t), V(f_t))$ . + +# 3.3. Sound-aware Representation Space Learning + +Given $m$ triplets of sound-aware representations for the Ego-Exo video pairs and their corresponding texts: $\{(f_{sv}^{ego(1)}, f_{sv}^{exo(1)}, f_{st}^{(1)}), \dots, (f_{sv}^{ego(m)}, f_{sv}^{exo(m)}, f_{st}^{(m)})\}$ , the visual representations are trained to be closer to the text representation within the same triplet while remaining distant from other text representations using two contrastive losses. + +$$ +\text {l o s s} ^ {\text {e g o 2 t e x t}} = - \frac {1}{m} \sum_ {i = 1} ^ {m} \log \frac {\exp \left(\sin \left(f _ {s v} ^ {e g o (i)} , f _ {s t} ^ {(i)}\right) / \tau\right)}{\sum_ {k \neq i} \exp \left(\sin \left(f _ {s v} ^ {e g o (i)} , f _ {s t} ^ {(k)}\right) / \tau\right)} \tag {11} +$$ + +$$ +\text {l o s s} ^ {\text {e x o 2 t e x t}} = - \frac {1}{m} \sum_ {i = 1} ^ {m} \log \frac {\exp \left(\sin \left(f _ {s v} ^ {e x o (i)} , f _ {s t} ^ {(i)}\right) / \tau\right)}{\sum_ {k \neq i} \exp \left(\sin \left(f _ {s v} ^ {e x o (i)} , f _ {s t} ^ {(k)}\right) / \tau\right)} \tag {12} +$$ + +where $\tau$ is the temperature parameter. We use the sum of these two losses to train the sound-aligned Ego-Exo repre + +sentation space. + +For the Ego-Exo video retrieval and text-video retrieval tasks, we compute the similarity between $(f_{sv}^{ego}, f_{sv}^{exo})$ and $(f_{sv}, f_{st})$ to obtain ranked lists. + +# 4. Sound-aligned Space for Downstream Tasks + +To verify the effectiveness of the learned sound-aligned Ego-Exo representation space, we apply the learned sound-aware visual representation to two tasks: scene recognition on egocentric and action recognition on exocentric video. Scene recognition is particularly challenging in an egocentric view because the ego view typically captures local information, such as hand-object interactions. Here, we investigate whether the proposed egocentric sound-vision representation can learn global and environmental information from an exocentric view, thereby enhancing scene recognition performance. We predict the scene classes as follows: + +$$ +\hat {y} _ {\text {s c e n e}} = \operatorname {s o f t m a x} \left(F C \left(f _ {s v} ^ {e g o}\right)\right). \tag {13} +$$ + +where $FC$ is a fully connected layer that maps the sound-aligned feature to scene classes. + +Similarly, we predict the action label using the exocentric sound-aware visual feature as: + +$$ +\hat {y} _ {\text {a c t i o n}} = \operatorname {s o f t m a x} \left(F C \left(f _ {s v} ^ {e x o}\right)\right). \tag {14} +$$ + +where $FC$ is a fully connected layer that maps the sound-aligned feature to action classes. We train the FC layers using cross-entropy loss with the ground truth labels. + +# 5. Experiments + +In this section, we first evaluate the effectiveness of the sound-aware representation space on Ego-Exo video re + +trieval and text-video retrieval tasks. Next, we apply the sound-aware visual representations to two downstream tasks: scene recognition in egocentric videos and action recognition in exocentric videos. Finally, we conduct ablation studies to verify the effectiveness of our proposed modules. + +# 5.1. Experimental Settings + +# 5.1.1 Datasets and Evaluation Metric + +We conduct the retrieval tasks on two Ego-Exo benchmarks. One benchmark consists of strictly paired Ego-Exo videos (i.e., CharadesEgo [1]), which are recorded using the same shooting script. The other benchmark consists of loosely paired Ego-Exo videos (i.e., EgoExoLearn [12]), where most video pairs share only the same verbs and nouns in their descriptions. CharadesEgo [1] is a large-scale EgoExo dataset featuring videos of daily activities. It comprises 7,985 video clips with a total duration of approximately 80 hours. The dataset is divided into training, validation, and test subsets, containing 5,338, 1,334, and 1,313 video pairs, respectively, along with corresponding shooting script descriptions. EgoExoLearn [12] is another large-scale EgoExo dataset, consisting of 747 video clips with a total duration of 120 hours. Following [34], we construct 36,372 Ego-Exo pairs by aligning nouns and verbs in the video text descriptions. We use the official data split, which includes 36,373 video pairs for training, 800 for validation, and 2,200 for testing. The downstream tasks are also conducted on CharadesEgo [1] and EgoExoLearn [12]. CharadesEgo contains 16 scene categories (e.g., kitchen and bedroom), while EgoExoLearn covers 19 action categories from both daily life and laboratory scenes (e.g., pick-up and washing). We use CharadesEgo for egocentric video scene recognition and EgoExoLearn for exocentric video action recognition. Following previous works [12, 34], we report top-1 accuracy for both retrieval and recognition tasks. + +# 5.1.2 Implementation Details + +We use the video transformer from [3] with a ViT-B/16 backbone initialized with CLIP weights [23] as the video encoder $E_{v}(x)$ . For the sound-vision module, we apply the audio encoder from [6] as $E_{a}(x)$ , keeping it frozen during training. In the sound-text module, we use the pretrained CLIP textual encoder as $E_{t}(x)$ to encode both semantic text and sound captions. The output dimension of the sound-aligned Ego-Exo representation space is 256. We use Llama3-8B as the LLM to generate sound captions. In the multi-head cross-attention module, we set the number of heads to $h = 4$ . The temperature parameter $\tau$ is set to 0.07. We utilize 8 V100 GPUs to train the sound-aware representation space with a learning rate of 3e-5, using AdamW as the optimizer for improved convergence and stability. + +It is worth noting that for video pairs with high synchronization, applying the Dynamic Time Warping (DTW) algo + +Table 1. Retrieval result comparison (top-1 accuracy) of Exo-Ego video alignment task on two benchmarks. + +
MethodEgoExoLearnCharadesEgo
Eg2ExEx2EgAvgEg2ExEx2EgAvg
FrozenInTime [3]14.1013.4013.756.734.965.85
LaViLa [36]28.7025.7027.2058.2446.9352.59
EgoVLP [17]32.1028.930.5058.2760.7659.52
InternVideo [32]30.6021.7026.1553.5561.2357.39
EgoExo [12]49.0045.3047.1568.3262.0665.19
EgoInstructor [34]47.1045.8246.4662.0656.3859.22
SoundBridge50.2750.7350.5077.6671.6374.65
+ +rhythm [24] to align the two audio sequences before inputting the audio signals into the cross-attention module can help mitigate the impact of varying sound speeds. We use this setting by default for videos in CharadesEgo and provide an ablation study on DTW in the supplementary material. + +# 5.1.3 Baseline Models + +We include several types of models as baselines. These include recent text-video representation models pre-trained on large-scale exocentric datasets, such as FrozenInTime [3] and InternVideo [32]. Additionally, we include pretrained models on large-scale egocentric datasets, namely LaViLa [36] and EgoVLP [17]. Furthermore, we compare our method with two recent approaches designed for EgoExo video alignment, namely EgoExo [12] and EgoInstructor [34]. For the downstream tasks, we also include task-specific models, namely VideoMAE [30] for exocentric action recognition and InternImage [31] for egocentric scene recognition. + +# 5.2. Results on Ego-Exo Video Alignment + +Table 1 compares the results of Ego-to-Exo and Exo-to-Ego video retrieval with other baselines. Specifically, we report the zero-shot retrieval performance of the FrozenIn-Time model and the fine-tuned results of other methods on the EgoExoLearn and CharadesEgo datasets. Our proposed method, SoundBridge, consistently outperforms other baselines, including pre-trained egocentric models, pre-trained exocentric models, and recent Ego-Exo models, on both Ego-to-Exo and Exo-to-Ego video retrieval tasks across the two benchmarks. We are the first model to achieve a top-1 accuracy exceeding $50\%$ on the EgoExoLearn dataset. Specifically, SoundBridge achieves higher retrieval accuracy than EgoExo [12] and EgoInstructor [34] by $8.13\%$ and $15.93\%$ on Ego-to-Exo video retrieval, and by $13.71\%$ and $18.88\%$ on Exo-to-Ego video retrieval, respectively. The performance boost primarily comes from the incorporation of sound-aware modules in the alignment process. For example, given an egocentric video query recording picking up the garlic with the left hand, the EgoExo model ranks an exocentric video of put the ginger down with both hand on the iron bowl in the top-1 position. In contrast, our + +![](images/a483729a140a77a2666eb0ddcfb786f9c61e9f26ff1360a6e3d17caf0ef9ae4f.jpg) +Figure 3. Visualization of an Ego-to-Exo video retrieval example, compared with EgoExo [12] and EgoInstructor [34], on an egocentric video query looking for A person is working on a vacuum in the man cave, then sneezes and reaches for some medicine. + +![](images/1cfded1faa5ed2bdf84ee725960a128edc2f1bfa2bd5295f33969693c527ba44.jpg) +Figure 4. Visualization of an Exo-to-Ego video retrieval example, compared with EgoExo [12] and EgoInstructor [34], on an exocentric video query looking for a person throws a box of groceries onto the table. The person opens the box and takes out an item. + +method differentiates between the actions "pick-up" and "put down" using sound cues. The pick-up action produces a soft rustling sound, whereas the put-down action generates a clunk sound when interacting with the iron bowl. Our proposed sound-aware modules effectively encode these audio features and their semantic descriptions (e.g., [pick up ginger] (soft rustling)). As a result, our method correctly ranks the ground-truth exocentric video in the top-1 position. + +Figure 3 visualizes an Ego-to-Exo video retrieval example, comparing our method with EgoExo [12] and EgoInstructor [34]. The egocentric video depicts a person working on a vacuum, sneezing, and picking up a medicine bottle, while the visual cue for the sneeze is barely visible in the egocentric video. All methods successfully rank an egocentric video of a person with a vacuum in the top-1 position. However, EgoExo and EgoInstructor fail to capture either the sneezing action or the action of reaching for the medicine, leading to incorrect top-1 results. In contrast, our method effectively detects the sneeze and the action of picking up the pill bottle using sound cues, thereby successfully retrieving the ground-truth exocentric video. Similarly, Figure 4 visualizes an Exo-to-Ego retrieval example with a query for: a person throws a box of groceries onto + +the table. The person opens the box and takes out an item. All methods successfully retrieve videos that capture the actions of opening the box and taking out items. However, only our SoundBridge method correctly identifies the action of throwing the box onto the table, leveraging the thud sound as an essential cue. + +# 5.3. Results on Text-Video Retrieval + +The proposed sound-aware Ego-Exo representation space also enables text-video retrieval. Therefore, we compare its performance with baselines on both ego-(exo-)video-to-text and text-to-ego-(exo-)video retrieval tasks. Table 2 demonstrates that our method significantly enhances text-video retrieval performance on Ego-Exo datasets. Specifically, our method substantially improves the performance of the pre-trained egocentric model LaViLa [36], and the pre-trained exocentric model InternVideo [32]. Furthermore, our method outperforms recent Ego-Exo models, namely EgoExo [12] and EgoInstructor [34], by approximately $30.50\%$ and $23.47\%$ in overall performance across both datasets. The main improvement over recent Ego-Exo models is observed in queries with multiple nearest-neighbor (NN) answers. By incorporating sound aware + +Table 2. Performance comparison on text-video retrieval task. + +
MethodEgoExoLearnCharadesEgo
T2VV2TAvgT2VV2TAvg
FrozenInTime [3]6.034.055.049.405.627.51
LaViLa [36]34.8740.1837.5343.7441.6742.71
EgoVLP [17]19.0926.1022.6045.5143.2144.36
InternVideo [32]18.5926.5522.5747.3446.6346.99
EgoExo [12]58.6458.7358.6948.0045.2146.61
EgoInstructor [34]66.3766.6866.5346.9345.9846.46
SoundBridge71.4173.1472.2864.9063.6064.25
+ +ness, our method effectively ranks the ground-truth videos higher among similar nearest-neighbor results. For example, given a text query searching for exocentric videos of a person washing dishes in the kitchen, EgoExo retrieves videos of a person chopping potatoes, while EgoInstruc-tor retrieves videos of a person stir-frying. In contrast, our method distinguishes the washing action among these nearest neighbors by leveraging water-running sounds, allowing the ground-truth video to be ranked higher than videos of chopping and stir-frying. + +# 5.4. Results on Downstream Tasks + +We also evaluate the effectiveness of the learned egocentric visual feature (i.e., $f_{sv}^{ego}$ ) for scene recognition and the learned exocentric visual feature (i.e., $f_{sv}^{exo}$ ) for action recognition. We evaluate our model under two settings: (1) fine-tuning, where the training and test sets come from the same dataset, and (2) transfer learning, where the model is trained on one dataset and tested on another. The report performances of other models are all with the fine-tuning setting. + +Table 3 compares the top-1 recognition accuracy with baseline methods. Specifically, for the egocentric scene recognition task, our method improves upon the egocentric model LaViLa [36] and the Ego-Exo model EgoInstructor [34] by $23.42\%$ and $9.78\%$ , respectively. Figure 5 illustrates an example of scene recognition. Given an egocentric video query where only a hand wiping a table with a towel is visible, it is difficult to infer that the scene is a kitchen. Other methods fail to predict the correct scene label. For example, FrozenInTime [3] misclassifies the video as a living room, while LaViLa [36] predicts it as a bathroom. In contrast, our method correctly identifies the scene by successfully aligning the query video with the corresponding exocentric video using sound cues from multiple actions (e.g., walking to the table and leaving the room). The egocentric visual feature (i.e., $f_{sv}^{ego}$ ) learns environmental and global visual cues from the exocentric video, which contains a range hood, an oven, and a refrigerator, allowing it to infer that the scene is a kitchen. For exocentric action recognition, our method achieves an average accuracy of $56.25\%$ , which is lower than VideoMAE [30], as the model is trained on a large-scale action dataset. Despite this, our + +![](images/f8f7697f2b8942fb7b259c5c2c74f92bfe93fecb5f1205f8579975ae616a260d.jpg) +Figure 5. Visualization of the egocentric scene recognition results compared with FrozenInTime [3] and LaViLa [36]. + +![](images/4e538bb7800667e0b8c86a8c7206f5e9042021b611e147d188f064d69ac48f20.jpg) +Figure 6. Visualization of the exocentric action recognition results compared with InternVideo [32] and EgoInstructor [34]. + +Table 3. Result comparison on CharadesEgo scene recognition and EgoExoLearn action recognition. An asterisk (*) indicates the transfer learning setting, where a different dataset is used for training. + +
MethodEgo-SceneExo-Action
FrozenInTime [3]8.256.45
LaViLa [36]41.3726.25
EgoVLP [17]40.3138.75
InternVideo [32]43.0045.00
EgoExo [12]42.5546.25
EgoInstructor [34]46.5147.50
VideoMAEv2 [30]-62.50
InterImage [31]39.60-
SoundBridge*48.7652.50
SoundBridge51.0656.25
+ +approach benefits from the sound-aware Exo-Ego representation space, as shown in Figure 6. Predicting a person's action from an exocentric view is challenging due to occlusions caused by the human body and the person being far from the camera. InternVideo [32] misclassifies the query video as the action "put", while EgoInstructor [34] assigns a high probability to the action "pick-up". In contrast, our method aligns the query video with its corresponding egocentric view using the sound of washing dishes (e.g., water splashing) and correctly classifies the action as "washing" with the highest probability. + +# 5.5. Ablation study + +# 5.5.1 Results on Different LLMs + +We evaluated the impact of audio descriptions generated by different Large Language Models (LLMs) on the performance of cross-view video association tasks. Specifically, we used GPT-3.5 [4] and LLaMA3 [7] to generate audio descriptions using the same prompt. As shown in Table 5, the results indicate that while the cross-view video + +Table 4. Ablation study on the proposed sound-vision and sound-text modules. + +
MethodEgoExoLearnCharadesEgo
Eg2ExEx2EgT2VV2TEg2ExEx2EgT2VV2T
Baseline49.0045.3058.6458.7368.3262.0648.0045.21
Baseline + Sound-Text Module48.8250.1869.2371.9672.8165.2563.5561.35
Baseline + Sound-Vision Module48.4647.7369.1073.0974.3566.5563.5961.11
Baseline + Sound-Text&Vision Modules50.2750.7371.4173.1477.6671.6364.9063.60
+ +Table 5. Results with different LLMs. + +
LLMEgoExoLearnCharadesEgo
Eg2ExEx2EgAvgEg2ExEx2EgAvg
GPT-3.549.2748.3648.8273.7671.5172.64
LLaMA350.2750.7350.5077.6671.6374.65
+ +Table 6. Performance comparison of different anchors. + +
AnchorEgoExoLearnCharadesEgo
Eg2ExEx2EgEg2ExEx2Eg
Origin Caption48.6448.7374.3566.55
Sound Caption47.0949.9156.9751.42
Sound Feature36.6726.6756.2748.38
Sound-Text Module50.2750.7377.6671.63
+ +association performance with GPT-3.5 is lower than that achieved with LLaMA3, our method still surpasses existing approaches. This demonstrates the effectiveness of our proposed method. + +# 5.5.2 Impact of the Sound-aware Modules + +We conduct an ablation study on the proposed sound-vision and sound-text modules, using the state-of-the-art Ego-Exo model EgoExo [12] as the baseline. The performance of Ego-Exo video alignment and the overall accuracy of text-video retrieval are presented in Table 4. When using only the sound-vision or sound-text module, our method demonstrates competitive performance in the Ego-to-Exo video retrieval task. However, when both modules are jointly utilized, our method achieves significantly higher performance across the evaluated retrieval tasks. + +# 5.5.3 Results of Different Anchors + +We also investigate the effectiveness of using different features as anchors to align sound-aware visual representations. Specifically, we replace the $f_{st}$ in Eq. (11) and Eq. (12) with $f_{t}, f_{c}$ or $\text{sum}(f_{a}^{ego}, f_{a}^{exo})$ . As shown in Table 6, using audio alone as an anchor is not ideal, especially for the short video dataset EgoExoLearn, where events heavily overlap and exhibit only slight variations. Upon analyzing the erroneous cases, we found that in many samples, the audio anchor alone fails to provide sufficient information to distinguish between videos from different views. As il + +![](images/5605f5e34f91f4b7e706fce05920b5b7331a927c86abbc3b344732b3f512eb60.jpg) +Figure 7. An example of Exo-Ego video retrieval using different anchors in the SoundBridge model. + +lustrated in Figure 7, a person is slicing garlic in the query video, and the text describes the semantic information of "a person slicing garlic". However, the sound features only convey the action "slice". During training, different-view videos were aligned solely based on the "slice" action, leading to incorrect retrieval of a video showing "slicing ginger" instead of "slicing garlic". Similarly, the sound caption lacks full contextual information and misinterprets the one-hand slicing action as two-hand slicing. + +# 6. Conclusion + +In this paper, we proposed two sound-aware modules as a bridge to address the large visual variance challenge in egocentric and exocentric video alignment. Both modules demonstrated significant improvements over state-of-the-art methods in Ego-Exo video retrieval and text-video retrieval tasks. Moreover, we verified that both egocentric and exocentric representations benefit from the sound-aware learning paradigm, achieving substantial improvements in downstream recognition tasks by leveraging complementary information from a correctly associated view. Furthermore, we found that our sound-vision and sound-text modules provide effective cues to distinguish the target video from its nearest neighbors based on sound. + +# Acknowledgements + +This research was supported by the National Natural Science Foundation of China (Grant No.: 62372314, 62476097); the Guangdong Provincial Fund for Basic and Applied Basic Research—Regional Joint Fund Project (Key Project) (2023B1515120078); the Guangdong Provincial Natural Science Foundation for Outstanding Youth Team Project (2024B1515040010). + +# References + +[1] Karteek Alahari. Actor and observer: Joint modeling of first and third-person videos. In Proceedings of the 1st Workshop and Challenge on Comprehensive Video Understanding in the Wild, page 3, 2018. 2, 5 +[2] Relja Arandjelovic and Andrew Zisserman. Look, listen and learn. In Proceedings of the IEEE international conference on computer vision, pages 609-617, 2017. 2 +[3] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1728-1738, 2021. 5, 7 +[4] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 7 +[5] Minjie Cai, Kris M Kitani, and Yoichi Sato. Understanding hand-object manipulation with grasp types and object attributes. In Robotics: Science and Systems, 2016. 2 +[6] Sanyuan Chen, Yu Wu, Chengyi Wang, Shujie Liu, Daniel Tompkins, Zhuo Chen, and Furu Wei. Beats: Audio pre-training with acoustic tokenizers. arXiv preprint arXiv:2212.09058, 2022. 5 +[7] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 7 +[8] Antonino Furnari and Giovanni Maria Farinella. What would you expect? anticipating egocentric actions with rolling-unrolling lstms and modality attention. In Proceedings of the IEEE/CVF International conference on computer vision, pages 6252-6261, 2019. 2 +[9] Mohit Goyal, Sahil Modi, Rishabh Goyal, and Saurabh Gupta. Human hands as probes for interactive object understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3293-3303, 2022. 2 +[10] Kristen Grauman, Andrew Westbury, Eugene Byrne, Zachary Chavis, Antonino Furnari, Rohit Girdhar, Jackson Hamburger, Hao Jiang, Miao Liu, Xingyu Liu, Miguel Martin, Tushar Nagarajan, Iija Radosavovic, Santhosh Kumar Ramakrishnan, Fiona Ryan, Jayant Sharma, Michael Wray, Mengmeng Xu, Eric Zhongcong Xu, Chen Zhao, Siddhant Bansal, Dhruv Batra, Vincent Cartillier, Sean Crane, Tien Do, Morrie Doulaty, Akshay Erapalli, Christoph Feichtenhofer, Adriano Fragomeni, Qichen Fu, Abrham Gebreselasie, Cristina Gonzalez, James Hillis, Xuhua Huang, Yifei Huang, Wenqi Jia, Weslie Khoo, Jachym + +Kolář, Satwik Kottur, Anurag Kumar, Federico Landini, Chao Li, Yanghao Li, Zhenqiang Li, Karttikeya Mangalam, Raghava Modhugu, Jonathan Munro, Tullie Murrell, Takumi Nishiyasu, Will Price, Paola Ruiz, Merey Ramazanova, Leda Sari, Kiran Somasundaram, Audrey Southerland, Yusuke Sugano, Ruijie Tao, Minh Vo, Yuchen Wang, Xindi Wu, Takuma Yagi, Ziwei Zhao, Yunyi Zhu, Pablo Arbeláez, David Crandall, Dima Damen, Giovanni Maria Farinella, Christian Fuegen, Bernard Ghanem, Vamsi Krishna Ithapu, C. V. Jawahar, Hanbyul Joo, Kris Kitani, Haizhou Li, Richard Newcombe, Aude Oliva, Hyun Soo Park, James M. Rehg, Yoichi Sato, Jianbo Shi, Mike Zheng Shou, Antonio Torralba, Lorenzo Torresani, Mingfei Yan, and Jitendra Malik. Ego4d: Around the world in 3,000 hours of egocentric video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18995-19012, 2022. 1, 2 +[11] Bo He, Jun Wang, Jielin Qiu, Trung Bui, Abhinav Shrivastava, and Zhaowen Wang. Align and attend: Multimodal summarization with dual contrastive losses. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14867-14878, 2023. 3 +[12] Yifei Huang, Guo Chen, Jilan Xu, Mingfang Zhang, Lijin Yang, Baoqi Pei, Hongjie Zhang, Lu Dong, Yali Wang, Limin Wang, et al. Egoxolearn: A dataset for bridging asynchronous ego-and exo-centric view of procedural activities in real world. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22072-22086, 2024. 1, 2, 5, 6, 7, 8 +[13] Sarah Ibrahimi, Xiaohang Sun, Pichao Wang, Amanmeet Garg, Ashutosh Sanan, and Mohamed Omar. Audio-enhanced text-to-video retrieval using text-conditioned feature alignment. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12054-12064, 2023. 2, 3 +[14] Hao Jiang and Vamsi Krishna Ithapu. Egocentric pose estimation from human vision span. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 10986-10994. IEEE, 2021. 2 +[15] Patricia S Koskinen, Robert M Wilson, and Carl J Jensema. Using closed-captioned television in the teaching of reading to deaf students. American Annals of the Deaf, 131(1):43-46, 1986. 2, 3 +[16] Haoxin Li, Yijun Cai, and Wei-Shi Zheng. Deep dual relation modeling for egocentric interaction recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7932-7941, 2019. 2 +[17] Kevin Qinghong Lin, Jinpeng Wang, Mattia Soldan, Michael Wray, Rui Yan, Eric Z Xu, Difei Gao, Rong-Cheng Tu, Wenzhe Zhao, Weijie Kong, et al. Ego-centric video-language pretraining. Advances in Neu + +ral Information Processing Systems, 35:7575-7586, 2022. 5, 7 +[18] Yan-Bo Lin, Jie Lei, Mohit Bansal, and Gedas Bertasius. Eclipse: Efficient long-range video retrieval using sight and sound. In European Conference on Computer Vision, pages 413–430. Springer, 2022. 2, 3 +[19] Shaowei Liu, Subarna Tripathi, Somdeb Majumdar, and Xiaolong Wang. Joint hand motion and interaction hotspots prediction from egocentric videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3282-3292, 2022. 2 +[20] Sagnik Majumder, Ziad Al-Halah, and Kristen Grauman. Learning spatial features from audio-visual correspondence in egocentric videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 27058-27068, 2024. 3 +[21] Tushar Nagarajan, Yanghao Li, Christoph Feichtenhofer, and Kristen Grauman. Ego-topo: Environment affordances from egocentric video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 163-172, 2020. 2 +[22] Chiara Plizzari, Mirco Planamente, Gabriele Goletto, Marco Cannici, Emanuele Gusso, Matteo Matteucci, and Barbara Caputo. E2 (go) motion: Motion augmented event stream for egocentric action recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 19935-19947, 2022. 2 +[23] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 5 +[24] Hiroaki Sakoe and Seibi Chiba. Dynamic programming algorithm optimization for spoken word recognition. IEEE transactions on acoustics, speech, and signal processing, 26(1):43-49, 1978. 5 +[25] Muhammad Bilal Shaikh, Douglas Chai, Syed Mohammed Shamsul Islam, and Naveed Akhtar. Multimodal fusion for audio-image and video action recognition. Neural Computing and Applications, 36(10): 5499-5513, 2024. 3 +[26] Dandan Shan, Jiaqi Geng, Michelle Shu, and David F Fouhey. Understanding human hands in contact at internet scale. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9869-9878, 2020. 2 +[27] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 3 +[28] Huiyu Wang, Mitesh Kumar Singh, and Lorenzo Torresani. Ego-only: Egocentric action detection with + +out exocentric transferring. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5250-5261, 2023. 2 +[29] Jian Wang, Diogo Luvizon, Weipeng Xu, Lingjie Liu, Kripasindhu Sarkar, and Christian Theobalt. Scene-aware egocentric 3d human pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13031-13040, 2023. 2 +[30] Limin Wang, Bingkun Huang, Zhiyu Zhao, Zhan Tong, Yinan He, Yi Wang, Yali Wang, and Yu Qiao. Videomae v2: Scaling video masked autoencoders with dual masking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14549-14560, 2023. 5, 7 +[31] Wenhai Wang, Jifeng Dai, Zhe Chen, Zhenhang Huang, Zhiqi Li, Xizhou Zhu, Xiaowei Hu, Tong Lu, Lewei Lu, Hongsheng Li, et al. Internimage: Exploring large-scale vision foundation models with deformable convolutions. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14408-14419, 2023. 5, 7 +[32] Yi Wang, Yinan He, Yizhuo Li, Kunchang Li, Jiashuo Yu, Xin Ma, Xinhao Li, Guo Chen, Xinyuan Chen, Yaohui Wang, Ping Luo, Ziwei Liu, Yali Wang, Limin Wang, and Yu Qiao. Internvid: A large-scale videotext dataset for multimodal understanding and generation. In The Twelfth International Conference on Learning Representations, 2024. 5, 6, 7 +[33] Jianjia Xin, Lichun Wang, Kai Xu, Chao Yang, and Baocai Yin. Learning interaction regions and motion trajectories simultaneously from egocentric demonstration videos. IEEE Robotics and Automation Letters, 2023. 2 +[34] Jilan Xu, Yifei Huang, Junlin Hou, Guo Chen, Yuejie Zhang, Rui Feng, and Weidi Xie. Retrieval-augmented egocentric video captioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13525-13536, 2024. 1, 2, 5, 6, 7 +[35] Zihui Sherry Xue and Kristen Grauman. Learning fine-grained view-invariant representations from unpaired ego-exo videos via temporal alignment. Advances in Neural Information Processing Systems, 36: 53688-53710, 2023. 2 +[36] Yue Zhao, Ishan Misra, Philipp Krahenbuhl, and Rohit Girdhar. Learning video representations from large language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6586-6597, 2023. 5, 6, 7 \ No newline at end of file diff --git a/2025/Sound Bridge_ Associating Egocentric and Exocentric Videos via Audio Cues/images.zip b/2025/Sound Bridge_ Associating Egocentric and Exocentric Videos via Audio Cues/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..fc879bbebb90d8e0c8cfc6896010578ce6a0f448 --- /dev/null +++ b/2025/Sound Bridge_ Associating Egocentric and Exocentric Videos via Audio Cues/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2faeb282a1a44f7eef96467bda4ba9d48a38127649acb1b86ffb4b99b52e62e5 +size 451536 diff --git a/2025/Sound Bridge_ Associating Egocentric and Exocentric Videos via Audio Cues/layout.json b/2025/Sound Bridge_ Associating Egocentric and Exocentric Videos via Audio Cues/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..0156058f9f7deea111a886fd6748468f18547da4 --- /dev/null +++ b/2025/Sound Bridge_ Associating Egocentric and Exocentric Videos via Audio Cues/layout.json @@ -0,0 +1,7370 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 102, + 146, + 507, + 161 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 146, + 507, + 161 + ], + "spans": [ + { + "bbox": [ + 102, + 146, + 507, + 161 + ], + "type": "text", + "content": "Sound Bridge: Associating Egocentric and Exocentric Videos via Audio Cues" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 178, + 503, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 178, + 503, + 191 + ], + "spans": [ + { + "bbox": [ + 105, + 178, + 503, + 191 + ], + "type": "text", + "content": "Sihong Huang" + }, + { + "bbox": [ + 105, + 178, + 503, + 191 + ], + "type": "inline_equation", + "content": "^{1,5,*}" + }, + { + "bbox": [ + 105, + 178, + 503, + 191 + ], + "type": "text", + "content": ", Jiaxin Wu" + }, + { + "bbox": [ + 105, + 178, + 503, + 191 + ], + "type": "inline_equation", + "content": "^{2,*}" + }, + { + "bbox": [ + 105, + 178, + 503, + 191 + ], + "type": "text", + "content": ", Xiaoyong Wei" + }, + { + "bbox": [ + 105, + 178, + 503, + 191 + ], + "type": "inline_equation", + "content": "^{3,2,5,\\dagger}" + }, + { + "bbox": [ + 105, + 178, + 503, + 191 + ], + "type": "text", + "content": ", Yi Cai" + }, + { + "bbox": [ + 105, + 178, + 503, + 191 + ], + "type": "inline_equation", + "content": "^{1,\\dagger}" + }, + { + "bbox": [ + 105, + 178, + 503, + 191 + ], + "type": "text", + "content": ", Dongmei Jiang" + }, + { + "bbox": [ + 105, + 178, + 503, + 191 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 105, + 178, + 503, + 191 + ], + "type": "text", + "content": ", Yaowei Wang" + }, + { + "bbox": [ + 105, + 178, + 503, + 191 + ], + "type": "inline_equation", + "content": "^{4,5}" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 139, + 190, + 471, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 190, + 471, + 204 + ], + "spans": [ + { + "bbox": [ + 139, + 190, + 471, + 204 + ], + "type": "text", + "content": "1South China University of Technology, 2The Hong Kong Polytechnic University" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 121, + 203, + 489, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 203, + 489, + 216 + ], + "spans": [ + { + "bbox": [ + 121, + 203, + 489, + 216 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 121, + 203, + 489, + 216 + ], + "type": "text", + "content": "Sichuan University, " + }, + { + "bbox": [ + 121, + 203, + 489, + 216 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 121, + 203, + 489, + 216 + ], + "type": "text", + "content": "Harbin Institute of Technology, Shenzhen, " + }, + { + "bbox": [ + 121, + 203, + 489, + 216 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 121, + 203, + 489, + 216 + ], + "type": "text", + "content": "Peng Cheng Laboratory" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 137, + 216, + 473, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 216, + 473, + 227 + ], + "spans": [ + { + "bbox": [ + 137, + 216, + 473, + 227 + ], + "type": "text", + "content": "{fthuangsh, ycai}@mail.scut.edu.cn, jiaxwu@polyu.edu.hk, cswei@scu.edu.cn" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 211, + 228, + 396, + 237 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 211, + 228, + 396, + 237 + ], + "spans": [ + { + "bbox": [ + 211, + 228, + 396, + 237 + ], + "type": "text", + "content": "jiangdm@pcl.ac.cn, wangyaowei@hit.edu.cn" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 175, + 270, + 215, + 281 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 175, + 270, + 215, + 281 + ], + "spans": [ + { + "bbox": [ + 175, + 270, + 215, + 281 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 91, + 291, + 298, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 291, + 298, + 515 + ], + "spans": [ + { + "bbox": [ + 91, + 291, + 298, + 515 + ], + "type": "text", + "content": "Understanding human behavior and environmental information in egocentric videos is very challenging due to the invisibility of some actions (e.g., laughing and sneezing) and the local nature of the first-person view. Leveraging the corresponding exocentric video to provide global context has shown promising results. However, existing visual-to-visual and visual-to-textual Ego-Exo video alignment methods struggle with the issue that some activities may have non-visual overlap. To address this, we propose using sound as a bridge, as audio is often consistent across Ego-Exo videos. However, direct audio-to-audio alignment lacks context. Thus, we introduce two context-aware sound modules: one aligns audio with vision via a visual-audio cross-attention module, and another aligns text with sound closed caption generated by LLM. Experimental results on two Ego-Exo video association benchmarks show that each of the proposed modules enhances the state-of-the-art methods. Moreover, the proposed sound-aware egocentric or exocentric representation boosts the performance of downstream tasks, such as action recognition of exocentric videos and scene recognition of egocentric videos. The code and models can be accessed at https://github.com/shhuangcoder/SoundBridge." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 93, + 534, + 162, + 544 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 534, + 162, + 544 + ], + "spans": [ + { + "bbox": [ + 93, + 534, + 162, + 544 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 91, + 551, + 298, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 551, + 298, + 649 + ], + "spans": [ + { + "bbox": [ + 91, + 551, + 298, + 649 + ], + "type": "text", + "content": "Egocentric video analysis has been receiving extensive attention recently [10, 12, 34] due to its potential applications in robotic vision analysis, embedded AI, etc. Compared to the large-scale and in-depth studies on exocentric video analysis tasks, research on egocentric videos is still underexplored. Moreover, there are significant differences between egocentric and exocentric videos; thus, many findings and methodologies derived from an exocentric perspective are not directly applicable to egocentric video analysis." + } + ] + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 313, + 260, + 517, + 373 + ], + "blocks": [ + { + "bbox": [ + 313, + 260, + 517, + 373 + ], + "lines": [ + { + "bbox": [ + 313, + 260, + 517, + 373 + ], + "spans": [ + { + "bbox": [ + 313, + 260, + 517, + 373 + ], + "type": "image", + "image_path": "f55b5b9896177bd7c98e9d7433355a7627dc23fbc84ec69561270c1c2a024903.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 312, + 380, + 517, + 409 + ], + "lines": [ + { + "bbox": [ + 312, + 380, + 517, + 409 + ], + "spans": [ + { + "bbox": [ + 312, + 380, + 517, + 409 + ], + "type": "text", + "content": "Figure 1. Given a video shooting script, egocentric and exocentric views can exhibit significant visual variances for the same activity, while their audio signals remain relatively consistent." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 311, + 423, + 517, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 423, + 517, + 617 + ], + "spans": [ + { + "bbox": [ + 311, + 423, + 517, + 617 + ], + "type": "text", + "content": "The main difference between egocentric and exocentric videos is that they capture events from different perspectives. First-person view videos focus on local and detailed information about the current activity, whereas third-person view videos emphasize global and environmental context. An example is shown in Figure 1. Although egocentric and exocentric videos follow the same video shooting script, they exhibit significant visual variations. For example, in the case of the same activity, such as eating a snack, the egocentric view primarily captures hand-object interactions, such as a hand picking up a raisin, while the location remains unknown. In contrast, the exocentric view captures the overall scene of the event but lacks finer details—for example, it may show a person lying on a bed eating food, but the specific type of food remains unclear. Moreover, certain activities, such as laughing and sneezing, do not produce distinct visual cues in egocentric videos apart from camera shaking. Consequently, when attempting to align visual representations from exocentric to egocentric videos, there is often no overlapping visual content to match." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 311, + 617, + 517, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 617, + 517, + 666 + ], + "spans": [ + { + "bbox": [ + 311, + 617, + 517, + 666 + ], + "type": "text", + "content": "However, recent studies [12, 34] have shown that exocentric videos can supplement the understanding of their semantically aligned egocentric counterparts. In this paper, we also focus on constructing an Ego-Exo alignment space to facilitate egocentric video analysis. Previous studies on" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 15 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 15 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 15 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 198, + 13, + 442, + 24 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 13, + 442, + 24 + ], + "spans": [ + { + "bbox": [ + 198, + 13, + 442, + 24 + ], + "type": "text", + "content": "Except for this watermark, it is identical to the accepted version;" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 179, + 25, + 461, + 36 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 25, + 461, + 36 + ], + "spans": [ + { + "bbox": [ + 179, + 25, + 461, + 36 + ], + "type": "text", + "content": "the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 657, + 239, + 666 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 657, + 239, + 666 + ], + "spans": [ + { + "bbox": [ + 105, + 657, + 239, + 666 + ], + "type": "text", + "content": "* Equal Contribution. † Corresponding Authors." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 293, + 724, + 318, + 734 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 724, + 318, + 734 + ], + "spans": [ + { + "bbox": [ + 293, + 724, + 318, + 734 + ], + "type": "text", + "content": "28942" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 91, + 121, + 297, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 121, + 297, + 227 + ], + "spans": [ + { + "bbox": [ + 91, + 121, + 297, + 227 + ], + "type": "text", + "content": "egocentric and exocentric video alignment have primarily focused on matching visual content [35] or aligning both video types to an anchor (i.e., their semantic descriptions or a shared textual representation) [12, 34]. However, these alignment strategies may be insufficient due to the substantial visual variations between views. In contrast to the significant differences in view-dependent visual content, we observe that audio signals for the same event remain consistent across views. For example, in the case of laughing, the laughter sound—whether captured in an egocentric or exocentric video—can infer that someone is laughing." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 92, + 227, + 297, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 92, + 227, + 297, + 663 + ], + "spans": [ + { + "bbox": [ + 92, + 227, + 297, + 663 + ], + "type": "text", + "content": "In this paper, we propose using audio signals as a bridge to model the egocentric-exocentric video alignment space. However, directly linking audio signals across videos without context may be suboptimal, as audio information alone is not sufficiently discriminative [2]. For example, the sound of a click associated with the action \"switch off the light\" in Figure 1 could also occur in other activities, such as \"flipping a power switch on a household appliance\" or \"turning off a desk lamp\". To address this limitation, we propose constructing contextual audio features for alignment by incorporating visual and textual context. An overview of our proposed method is presented in Figure 2. Specifically, we introduce sound-text and sound-vision modules to bridge Ego-Exo videos. The sound-vision module follows existing methods [13, 18] to build contextual audiovisual representations through cross-attention. Meanwhile, the sound-text module utilizes a large language model (LLM) to generate contextual audio descriptions and construct sound-aware textual representations. In particular, we provide the LLM with semantic text describing the video scenario and prompt it to generate an audio description of the actions occurring within the scene in a closed-caption format [15]. Additionally, to ensure that the generated sound captions emphasize unique information within the paired cross-view videos rather than commonly occurring content in the dataset, we filter out frequently appearing and less informative concepts (e.g., person) from the generated descriptions. As a result, the output sound caption for the example in Figure 1 is \"[Eating the snack] (munching), [dishes on the floor] (crashing and scraping), [Laughing] (giggles), [leaving the room] (Footsteps), [turns off the light] (click)\". The generated sound caption and the semantic text are subsequently processed through a cross-attention block to learn a sound-aware textual representation. Finally, we treat the textual representation as an anchor and align the audiovisual representations of egocentric and exocentric videos to it, thereby learning an Ego-Exo semantic feature space. We conduct extensive experiments on two Ego-Exo video retrieval benchmarks, namely EgoExoLearn [12] and CharadesEgo [1]. The results demonstrate that the learned sound-aware features outperform state-of-the-art approaches in both egocentric-exocentric video retrieval and text-to-egocentric/exocentric video retrieval. Moreover, the learned Ego- and Exo- features enhance the performance of downstream tasks, such as exocentric video action" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 312, + 121, + 490, + 130 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 121, + 490, + 130 + ], + "spans": [ + { + "bbox": [ + 312, + 121, + 490, + 130 + ], + "type": "text", + "content": "recognition and egocentric video scene recognition." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 312, + 138, + 388, + 149 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 138, + 388, + 149 + ], + "spans": [ + { + "bbox": [ + 312, + 138, + 388, + 149 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 312, + 155, + 436, + 166 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 155, + 436, + 166 + ], + "spans": [ + { + "bbox": [ + 312, + 155, + 436, + 166 + ], + "type": "text", + "content": "2.1. Egocentric Video Analysis" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 312, + 170, + 517, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 170, + 517, + 413 + ], + "spans": [ + { + "bbox": [ + 312, + 170, + 517, + 413 + ], + "type": "text", + "content": "The number of research studies on egocentric video analysis is growing rapidly, as this task is a fundamental component of many applications [10, 12], such as robotic vision analysis and embedded AI. However, egocentric videos are fundamentally different from traditional exocentric videos. Egocentric videos capture a unique immersive viewpoint from a first-person perspective, directly reflecting an individual's interaction with the current activity, whereas exocentric videos provide an external view of human interactions within an environment. This distinct perspective in egocentric videos introduces a range of challenges for human activity analysis, including hand detection [5, 9, 26], action recognition [8, 22, 28], human-computer interaction [16, 21, 33], and applications in virtual and augmented reality [14, 19, 29]. Despite these differences, recent studies [12, 34] have demonstrated that egocentric video analysis can benefit from the contextual information provided by corresponding exocentric videos. For example, Xu et al. [34] show that related exocentric videos can enhance egocentric video captioning. Similarly, Huang et al. [12] demonstrate that the exocentric perspective provides complementary information that aids robots in performing complex tasks. In this paper, we also focus on developing effective Ego-Exo video alignment techniques to enhance egocentric video analysis." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 312, + 418, + 494, + 429 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 418, + 494, + 429 + ], + "spans": [ + { + "bbox": [ + 312, + 418, + 494, + 429 + ], + "type": "text", + "content": "2.2. Ego-Exo Video Representation Learning" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 312, + 434, + 517, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 434, + 517, + 668 + ], + "spans": [ + { + "bbox": [ + 312, + 434, + 517, + 668 + ], + "type": "text", + "content": "The goal of Ego-Exo video representation learning is to construct a semantic space in which egocentric and exocentric videos with similar semantics are positioned closely, while dissimilar Ego-Exo videos are placed farther apart [12]. Early studies on Ego-Exo representation learning align cross-view videos using visual information. For example, AE2 [35] extracts visual regions of hands and active objects and performs temporal alignment on these extracted regions across views, assuming that the visual content of a fine-grained action remains consistent across perspectives. AE2 constructs four action-specific cross-view datasets for evaluation: break egg, pour milk, pour liquid, and tennis forehand. Recent approaches aim to learn a semantic Ego-Exo representation space by aligning videos to either a shared semantic text or separate textual descriptions. For example, EgoExo [12] employs implicit Ego-Exo alignment by associating each video with a view-dependent text, based on the assumption that semantically similar texts will be positioned closely in the learned space. Beyond implicit alignment, EgoExoNCE [34] takes this further by explicitly aligning Ego-Exo videos to a mutual text anchor. This shared text is generated by extracting key nouns and verbs from both the egocentric and exocentric textual descriptions. In this paper, we follow the pipeline of previous" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 724, + 317, + 734 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 724, + 317, + 734 + ], + "spans": [ + { + "bbox": [ + 293, + 724, + 317, + 734 + ], + "type": "text", + "content": "28943" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 91, + 120, + 298, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 120, + 298, + 199 + ], + "spans": [ + { + "bbox": [ + 91, + 120, + 298, + 199 + ], + "type": "text", + "content": "work by aligning Ego-Exo videos to a mutual text anchor. However, rather than directly matching entire videos with full-text descriptions without attention, we propose using audio as a bridge to address the challenge of visual variance. Our approach incorporates sound-vision and sound-text modules to bridge the gap between egocentric and exocentric perspectives. To the best of our knowledge, we are the first to integrate sound into Ego-Exo video alignment." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 93, + 215, + 266, + 227 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 215, + 266, + 227 + ], + "spans": [ + { + "bbox": [ + 93, + 215, + 266, + 227 + ], + "type": "text", + "content": "2.3. Sound-aware Representation Learning" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 91, + 230, + 298, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 230, + 298, + 502 + ], + "spans": [ + { + "bbox": [ + 91, + 230, + 298, + 502 + ], + "type": "text", + "content": "Numerous studies have demonstrated that audio signals complement visual signals in video analysis tasks [18]. For example, the Eclipse method has shown that audio signals can serve as an alternative to visual signals in long-range video retrieval, significantly accelerating the search process [18]. Consequently, many approaches have sought to develop sound-aware representations. Several works have incorporated visual information as context to create audiovisual features for exocentric video analysis, applying them to tasks such as video retrieval [13, 18], action recognition [20], and video summarization [11]. More recently, audio-based methods have also been explored in egocentric video analysis, such as active speaker spatial detection [25]. Additionally, Ibrahimi et al. [13] integrate textual context to construct audio-textual representations. Common approaches to incorporating audio signals into video analysis include appending audio features to visual or textual features [20, 25] or learning audio-textual/visual cross-attention representations [13, 18]. Another method involves converting audio signals into text using Automatic Speech Recognition (ASR) and subsequently appending the ASR-derived text features to visual features [11]. Our proposed sound-vision model follows prior work by learning contextual audiovisual features through cross-attention. Additionally, we introduce a novel approach to learning audiotextual representations in our sound-text module, which leverages a large language model (LLM) to generate sound captions." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 93, + 521, + 297, + 544 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 521, + 297, + 544 + ], + "spans": [ + { + "bbox": [ + 93, + 521, + 297, + 544 + ], + "type": "text", + "content": "3. Sound-aligned Ego-Exo Representation Learning" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 91, + 549, + 298, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 549, + 298, + 667 + ], + "spans": [ + { + "bbox": [ + 91, + 549, + 298, + 667 + ], + "type": "text", + "content": "The architecture of our proposed method is illustrated in Figure 2. The main objective is to construct a sound-aligned Ego-Exo representation space that integrates local (Ego) and global (Exo) visual features along with their corresponding semantic concepts. Specifically, to address the challenge of aligning semantic text with the significant visual variations present in Ego-Exo videos, we leverage sound as a unifying element. This is achieved by incorporating sound-vision and sound-text modules to generate sound-aware representations. Subsequently, we align the sound-aware video features with the sound-aware textual features to learn a unified representation space." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 312, + 120, + 419, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 120, + 419, + 129 + ], + "spans": [ + { + "bbox": [ + 312, + 120, + 419, + 129 + ], + "type": "text", + "content": "3.1. Sound-Vision Module" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 311, + 135, + 518, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 135, + 518, + 224 + ], + "spans": [ + { + "bbox": [ + 311, + 135, + 518, + 224 + ], + "type": "text", + "content": "Given a semantically aligned egocentric and exocentric (ego-exo) video pair " + }, + { + "bbox": [ + 311, + 135, + 518, + 224 + ], + "type": "inline_equation", + "content": "(v^{ego}, v^{exo})" + }, + { + "bbox": [ + 311, + 135, + 518, + 224 + ], + "type": "text", + "content": " and their corresponding audio signals " + }, + { + "bbox": [ + 311, + 135, + 518, + 224 + ], + "type": "inline_equation", + "content": "(a^{ego}, a^{exo})" + }, + { + "bbox": [ + 311, + 135, + 518, + 224 + ], + "type": "text", + "content": ", we propose a sound-vision cross-attention module to learn their sound-aware visual representations. The sound-vision module consists of a video encoder " + }, + { + "bbox": [ + 311, + 135, + 518, + 224 + ], + "type": "inline_equation", + "content": "E_v(x)" + }, + { + "bbox": [ + 311, + 135, + 518, + 224 + ], + "type": "text", + "content": ", an audio encoder " + }, + { + "bbox": [ + 311, + 135, + 518, + 224 + ], + "type": "inline_equation", + "content": "E_a(x)" + }, + { + "bbox": [ + 311, + 135, + 518, + 224 + ], + "type": "text", + "content": ", and a multi-head cross-attention block Multihead " + }, + { + "bbox": [ + 311, + 135, + 518, + 224 + ], + "type": "inline_equation", + "content": "(Q, K, V)" + }, + { + "bbox": [ + 311, + 135, + 518, + 224 + ], + "type": "text", + "content": " [27]. The module inputs a video " + }, + { + "bbox": [ + 311, + 135, + 518, + 224 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 311, + 135, + 518, + 224 + ], + "type": "text", + "content": " and its corresponding audio signal " + }, + { + "bbox": [ + 311, + 135, + 518, + 224 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 311, + 135, + 518, + 224 + ], + "type": "text", + "content": " and outputs sound-aware visual feature " + }, + { + "bbox": [ + 311, + 135, + 518, + 224 + ], + "type": "inline_equation", + "content": "f_{sv}" + }, + { + "bbox": [ + 311, + 135, + 518, + 224 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 369, + 231, + 517, + 243 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 369, + 231, + 517, + 243 + ], + "spans": [ + { + "bbox": [ + 369, + 231, + 517, + 243 + ], + "type": "interline_equation", + "content": "f _ {v} = E _ {v} (v), f _ {a} = E _ {a} (a), \\tag {1}", + "image_path": "95df07d4005bb287f904145122429ac0fd9b1b5b203c656737853706550bb9dd.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 323, + 248, + 517, + 273 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 248, + 517, + 273 + ], + "spans": [ + { + "bbox": [ + 323, + 248, + 517, + 273 + ], + "type": "interline_equation", + "content": "C r o s s A B (Q, K, V) = S o f t m a x \\left(\\frac {Q \\times K ^ {T}}{\\sqrt {d _ {k}}}\\right) \\cdot V, \\tag {2}", + "image_path": "d787cab96d8d11386412ac9ae642e91bd3590384168b902e7a67cda68ef872cf.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 315, + 277, + 517, + 297 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 277, + 517, + 297 + ], + "spans": [ + { + "bbox": [ + 315, + 277, + 517, + 297 + ], + "type": "interline_equation", + "content": "\\operatorname {M u l t i H e a d} (Q, K, V) = \\operatorname {C o n c a t} \\left(\\text {h e a d} _ {1}, \\dots , \\text {h e a d} _ {h}\\right) W ^ {O} \\tag {3}", + "image_path": "2f34f72f07ee8e8de4ef358cd19610ab5c6e2a543a9af2a786345fdf992fa1c7.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 339, + 299, + 517, + 311 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 299, + 517, + 311 + ], + "spans": [ + { + "bbox": [ + 339, + 299, + 517, + 311 + ], + "type": "interline_equation", + "content": "f _ {s v} = \\operatorname {M u l t i H e a d} \\left(Q \\left(f _ {a}\\right), K \\left(f _ {v}\\right), V \\left(f _ {v}\\right)\\right). \\tag {4}", + "image_path": "e3f5ba89af8b5e04e5826d503940ab325f61f4ca060ff9d6079188aad3bb746f.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 312, + 313, + 518, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 313, + 518, + 392 + ], + "spans": [ + { + "bbox": [ + 312, + 313, + 518, + 392 + ], + "type": "text", + "content": "where the " + }, + { + "bbox": [ + 312, + 313, + 518, + 392 + ], + "type": "inline_equation", + "content": "Q(x)" + }, + { + "bbox": [ + 312, + 313, + 518, + 392 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 312, + 313, + 518, + 392 + ], + "type": "inline_equation", + "content": "K(x)" + }, + { + "bbox": [ + 312, + 313, + 518, + 392 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 312, + 313, + 518, + 392 + ], + "type": "inline_equation", + "content": "V(x)" + }, + { + "bbox": [ + 312, + 313, + 518, + 392 + ], + "type": "text", + "content": " map the input feature " + }, + { + "bbox": [ + 312, + 313, + 518, + 392 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 312, + 313, + 518, + 392 + ], + "type": "text", + "content": " to a hidden space though linear projections and " + }, + { + "bbox": [ + 312, + 313, + 518, + 392 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 312, + 313, + 518, + 392 + ], + "type": "text", + "content": " represents the number of head. The egocentric and exocentric sound-vision modules in Figure 2 share the same architecture, as illustrated in Eqs. (1)-(4), but with view-dependent weights. Specifically, we feed egocentric videos and exocentric videos and their audio signals to the sound-vision module to obtain sound-aware representations as:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 344, + 398, + 517, + 410 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 344, + 398, + 517, + 410 + ], + "spans": [ + { + "bbox": [ + 344, + 398, + 517, + 410 + ], + "type": "interline_equation", + "content": "f _ {v} ^ {e g o} = E _ {v} ^ {e g o} \\left(v ^ {e g o}\\right), f _ {v} ^ {e x o} = E _ {v} ^ {e x o} \\left(v ^ {e x o}\\right). \\tag {5}", + "image_path": "7752145fbcfe3eac595058c7fc0061cd6b69432d98d58d193df07623258b1ef3.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 351, + 413, + 517, + 425 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 413, + 517, + 425 + ], + "spans": [ + { + "bbox": [ + 351, + 413, + 517, + 425 + ], + "type": "interline_equation", + "content": "f _ {a} ^ {e g o} = E _ {a} \\left(a ^ {e g o}\\right), f _ {a} ^ {e x o} = E _ {a} \\left(a ^ {e x o}\\right). \\tag {6}", + "image_path": "ea09a9b79fab9565409ea29b73297e4b1c91d25e5ee2689fde286c7efcbb1e32.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 322, + 432, + 517, + 445 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 432, + 517, + 445 + ], + "spans": [ + { + "bbox": [ + 322, + 432, + 517, + 445 + ], + "type": "interline_equation", + "content": "f _ {s v} ^ {e g o} = \\text {M u l t i H e a d} \\left(Q \\left(f _ {a} ^ {e g o}\\right), K \\left(f _ {v} ^ {e g o}\\right), V \\left(f _ {v} ^ {e g o}\\right)\\right) \\tag {7}", + "image_path": "7f1569bdd148dff96e4b9f6cbdb4bab70e7e93bc5c90345538cb17eb6bea835f.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 321, + 455, + 517, + 468 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 455, + 517, + 468 + ], + "spans": [ + { + "bbox": [ + 321, + 455, + 517, + 468 + ], + "type": "interline_equation", + "content": "f _ {s v} ^ {e x o} = \\text {M u l t i H e a d} \\left(Q \\left(f _ {a} ^ {e x o}\\right), K \\left(f _ {v} ^ {e x o}\\right), V \\left(f _ {v} ^ {e x o}\\right)\\right) \\tag {8}", + "image_path": "fc5f545e30f2cf42e9a4427307f3bd5faa43a2a451105afe754b6404c51d32aa.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 312, + 473, + 411, + 483 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 473, + 411, + 483 + ], + "spans": [ + { + "bbox": [ + 312, + 473, + 411, + 483 + ], + "type": "text", + "content": "3.2. Sound-Text Module" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 312, + 488, + 518, + 546 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 488, + 518, + 546 + ], + "spans": [ + { + "bbox": [ + 312, + 488, + 518, + 546 + ], + "type": "text", + "content": "Given a text " + }, + { + "bbox": [ + 312, + 488, + 518, + 546 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 312, + 488, + 518, + 546 + ], + "type": "text", + "content": " that captures the semantics of Ego-Exo videos, we use an LLM to generate a sound caption " + }, + { + "bbox": [ + 312, + 488, + 518, + 546 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 312, + 488, + 518, + 546 + ], + "type": "text", + "content": " and propose a sound-text cross-attention module to learn a sound-aware text representation " + }, + { + "bbox": [ + 312, + 488, + 518, + 546 + ], + "type": "inline_equation", + "content": "f_{st}" + }, + { + "bbox": [ + 312, + 488, + 518, + 546 + ], + "type": "text", + "content": ". The sound-text module shares the same architecture as the sound-vision module but uses different encoders:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 372, + 554, + 517, + 566 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 554, + 517, + 566 + ], + "spans": [ + { + "bbox": [ + 372, + 554, + 517, + 566 + ], + "type": "interline_equation", + "content": "f _ {t} = E _ {t} (t), f _ {c} = E _ {t} (c), \\tag {9}", + "image_path": "6652cf9fbe0e51c15a9828cd7572d7d449e12bc7adaf25325edfef56c48102d9.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 341, + 573, + 517, + 585 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 341, + 573, + 517, + 585 + ], + "spans": [ + { + "bbox": [ + 341, + 573, + 517, + 585 + ], + "type": "interline_equation", + "content": "f _ {s t} = \\text {M u l t i H e a d} \\left(Q \\left(f _ {c}\\right), K \\left(f _ {t}\\right), V \\left(f _ {t}\\right)\\right). \\tag {10}", + "image_path": "14cd0dd6bf116fa435080814ef6c4ec64db50d8c58c5e1280cdbae1c1f80b9f5.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 311, + 588, + 518, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 588, + 518, + 667 + ], + "spans": [ + { + "bbox": [ + 311, + 588, + 518, + 667 + ], + "type": "text", + "content": "Specifically, the sound caption " + }, + { + "bbox": [ + 311, + 588, + 518, + 667 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 311, + 588, + 518, + 667 + ], + "type": "text", + "content": " is generated from the given text " + }, + { + "bbox": [ + 311, + 588, + 518, + 667 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 311, + 588, + 518, + 667 + ], + "type": "text", + "content": " by an LLM. With " + }, + { + "bbox": [ + 311, + 588, + 518, + 667 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 311, + 588, + 518, + 667 + ], + "type": "text", + "content": " providing the video context, we prompt the LLM to generate sound descriptions for all the actions in the videos as: " + }, + { + "bbox": [ + 311, + 588, + 518, + 667 + ], + "type": "inline_equation", + "content": "c = LLM(t, prompt_{Txt2Sound})" + }, + { + "bbox": [ + 311, + 588, + 518, + 667 + ], + "type": "text", + "content": ". The sound description needs to contain the source of the sound and the sound effect, formatted in Closed Caption (CC) style [15]. Closed captions (CC) are a specialized type of caption designed for people who are deaf or hard of" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 724, + 318, + 734 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 724, + 318, + 734 + ], + "spans": [ + { + "bbox": [ + 293, + 724, + 318, + 734 + ], + "type": "text", + "content": "28944" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 95, + 120, + 512, + 295 + ], + "blocks": [ + { + "bbox": [ + 95, + 120, + 512, + 295 + ], + "lines": [ + { + "bbox": [ + 95, + 120, + 512, + 295 + ], + "spans": [ + { + "bbox": [ + 95, + 120, + 512, + 295 + ], + "type": "image", + "image_path": "c29c5ed9232595f4f2ea2bdeaaaabf1ee848e54d437d4c748636ad6c99eabe7e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 91, + 302, + 517, + 321 + ], + "lines": [ + { + "bbox": [ + 91, + 302, + 517, + 321 + ], + "spans": [ + { + "bbox": [ + 91, + 302, + 517, + 321 + ], + "type": "text", + "content": "Figure 2. Overview of the proposed method. We leverage sound as a bridge to connect Ego-Exo videos with their corresponding semantic text to build a sound-aligned representation space." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 91, + 338, + 297, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 338, + 297, + 484 + ], + "spans": [ + { + "bbox": [ + 91, + 338, + 297, + 484 + ], + "type": "text", + "content": "hearing. They additionally include an audio description in parentheses and indicate the source of the sound in brackets. For example, for a given text " + }, + { + "bbox": [ + 91, + 338, + 297, + 484 + ], + "type": "inline_equation", + "content": "t =" + }, + { + "bbox": [ + 91, + 338, + 297, + 484 + ], + "type": "text", + "content": " \"a person turning off the light\", the generated sound caption is " + }, + { + "bbox": [ + 91, + 338, + 297, + 484 + ], + "type": "inline_equation", + "content": "c =" + }, + { + "bbox": [ + 91, + 338, + 297, + 484 + ], + "type": "text", + "content": " \"[a person turning off the light] (click)\". Furthermore, to ensure that the sound caption focuses on the unique information of the paired videos, we filter out frequently appearing concepts (e.g., \"a person\") in CC. To achieve this, we use the LLM to refine " + }, + { + "bbox": [ + 91, + 338, + 297, + 484 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 91, + 338, + 297, + 484 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 91, + 338, + 297, + 484 + ], + "type": "inline_equation", + "content": "\\tilde{c}" + }, + { + "bbox": [ + 91, + 338, + 297, + 484 + ], + "type": "text", + "content": " by providing the frequently appeared concepts, i.e., " + }, + { + "bbox": [ + 91, + 338, + 297, + 484 + ], + "type": "inline_equation", + "content": "\\tilde{c} = \\text{LLM}(c, \\text{frequentConcepts}, \\text{prompt}_{\\text{soundRefine}})" + }, + { + "bbox": [ + 91, + 338, + 297, + 484 + ], + "type": "text", + "content": ". Due to space limitations, details of all prompts are provided in the supplementary material. We input the filtered and refined closed sound caption " + }, + { + "bbox": [ + 91, + 338, + 297, + 484 + ], + "type": "inline_equation", + "content": "\\tilde{c}" + }, + { + "bbox": [ + 91, + 338, + 297, + 484 + ], + "type": "text", + "content": " along with the text " + }, + { + "bbox": [ + 91, + 338, + 297, + 484 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 91, + 338, + 297, + 484 + ], + "type": "text", + "content": " into the sound-text module to obtain the sound-aware text representation: " + }, + { + "bbox": [ + 91, + 338, + 297, + 484 + ], + "type": "inline_equation", + "content": "f_{st} = \\text{MultiHead}(Q(f_{\\tilde{c}}), K(f_t), V(f_t))" + }, + { + "bbox": [ + 91, + 338, + 297, + 484 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 92, + 488, + 293, + 499 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 92, + 488, + 293, + 499 + ], + "spans": [ + { + "bbox": [ + 92, + 488, + 293, + 499 + ], + "type": "text", + "content": "3.3. Sound-aware Representation Space Learning" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 92, + 503, + 297, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 92, + 503, + 297, + 564 + ], + "spans": [ + { + "bbox": [ + 92, + 503, + 297, + 564 + ], + "type": "text", + "content": "Given " + }, + { + "bbox": [ + 92, + 503, + 297, + 564 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 92, + 503, + 297, + 564 + ], + "type": "text", + "content": " triplets of sound-aware representations for the Ego-Exo video pairs and their corresponding texts: " + }, + { + "bbox": [ + 92, + 503, + 297, + 564 + ], + "type": "inline_equation", + "content": "\\{(f_{sv}^{ego(1)}, f_{sv}^{exo(1)}, f_{st}^{(1)}), \\dots, (f_{sv}^{ego(m)}, f_{sv}^{exo(m)}, f_{st}^{(m)})\\}" + }, + { + "bbox": [ + 92, + 503, + 297, + 564 + ], + "type": "text", + "content": ", the visual representations are trained to be closer to the text representation within the same triplet while remaining distant from other text representations using two contrastive losses." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 97, + 572, + 297, + 605 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 572, + 297, + 605 + ], + "spans": [ + { + "bbox": [ + 97, + 572, + 297, + 605 + ], + "type": "interline_equation", + "content": "\\text {l o s s} ^ {\\text {e g o 2 t e x t}} = - \\frac {1}{m} \\sum_ {i = 1} ^ {m} \\log \\frac {\\exp \\left(\\sin \\left(f _ {s v} ^ {e g o (i)} , f _ {s t} ^ {(i)}\\right) / \\tau\\right)}{\\sum_ {k \\neq i} \\exp \\left(\\sin \\left(f _ {s v} ^ {e g o (i)} , f _ {s t} ^ {(k)}\\right) / \\tau\\right)} \\tag {11}", + "image_path": "b9b3d561de59a2a659cec68b29991c742e2a3db5175b747331cdecdea26f15ed.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 97, + 613, + 297, + 646 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 613, + 297, + 646 + ], + "spans": [ + { + "bbox": [ + 97, + 613, + 297, + 646 + ], + "type": "interline_equation", + "content": "\\text {l o s s} ^ {\\text {e x o 2 t e x t}} = - \\frac {1}{m} \\sum_ {i = 1} ^ {m} \\log \\frac {\\exp \\left(\\sin \\left(f _ {s v} ^ {e x o (i)} , f _ {s t} ^ {(i)}\\right) / \\tau\\right)}{\\sum_ {k \\neq i} \\exp \\left(\\sin \\left(f _ {s v} ^ {e x o (i)} , f _ {s t} ^ {(k)}\\right) / \\tau\\right)} \\tag {12}", + "image_path": "ac531421400daec274516c7df60e794b6bf35589b1fe4b5e960ea96f069c72b4.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 92, + 647, + 297, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 92, + 647, + 297, + 666 + ], + "spans": [ + { + "bbox": [ + 92, + 647, + 297, + 666 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 92, + 647, + 297, + 666 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 92, + 647, + 297, + 666 + ], + "type": "text", + "content": " is the temperature parameter. We use the sum of these two losses to train the sound-aligned Ego-Exo repre" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 312, + 339, + 369, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 339, + 369, + 348 + ], + "spans": [ + { + "bbox": [ + 312, + 339, + 369, + 348 + ], + "type": "text", + "content": "sentation space." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 312, + 348, + 517, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 348, + 517, + 378 + ], + "spans": [ + { + "bbox": [ + 312, + 348, + 517, + 378 + ], + "type": "text", + "content": "For the Ego-Exo video retrieval and text-video retrieval tasks, we compute the similarity between " + }, + { + "bbox": [ + 312, + 348, + 517, + 378 + ], + "type": "inline_equation", + "content": "(f_{sv}^{ego}, f_{sv}^{exo})" + }, + { + "bbox": [ + 312, + 348, + 517, + 378 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 312, + 348, + 517, + 378 + ], + "type": "inline_equation", + "content": "(f_{sv}, f_{st})" + }, + { + "bbox": [ + 312, + 348, + 517, + 378 + ], + "type": "text", + "content": " to obtain ranked lists." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 312, + 385, + 516, + 397 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 385, + 516, + 397 + ], + "spans": [ + { + "bbox": [ + 312, + 385, + 516, + 397 + ], + "type": "text", + "content": "4. Sound-aligned Space for Downstream Tasks" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 312, + 402, + 517, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 402, + 517, + 509 + ], + "spans": [ + { + "bbox": [ + 312, + 402, + 517, + 509 + ], + "type": "text", + "content": "To verify the effectiveness of the learned sound-aligned Ego-Exo representation space, we apply the learned sound-aware visual representation to two tasks: scene recognition on egocentric and action recognition on exocentric video. Scene recognition is particularly challenging in an egocentric view because the ego view typically captures local information, such as hand-object interactions. Here, we investigate whether the proposed egocentric sound-vision representation can learn global and environmental information from an exocentric view, thereby enhancing scene recognition performance. We predict the scene classes as follows:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 358, + 517, + 516, + 529 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 517, + 516, + 529 + ], + "spans": [ + { + "bbox": [ + 358, + 517, + 516, + 529 + ], + "type": "interline_equation", + "content": "\\hat {y} _ {\\text {s c e n e}} = \\operatorname {s o f t m a x} \\left(F C \\left(f _ {s v} ^ {e g o}\\right)\\right). \\tag {13}", + "image_path": "489c24a19c0cf0ca37e9506ec7cb93f0a746fd347ad5960d2b2ff540895e8d1b.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 312, + 531, + 516, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 531, + 516, + 550 + ], + "spans": [ + { + "bbox": [ + 312, + 531, + 516, + 550 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 312, + 531, + 516, + 550 + ], + "type": "inline_equation", + "content": "FC" + }, + { + "bbox": [ + 312, + 531, + 516, + 550 + ], + "type": "text", + "content": " is a fully connected layer that maps the sound-aligned feature to scene classes." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 312, + 551, + 516, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 551, + 516, + 570 + ], + "spans": [ + { + "bbox": [ + 312, + 551, + 516, + 570 + ], + "type": "text", + "content": "Similarly, we predict the action label using the exocentric sound-aware visual feature as:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 356, + 575, + 516, + 588 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 356, + 575, + 516, + 588 + ], + "spans": [ + { + "bbox": [ + 356, + 575, + 516, + 588 + ], + "type": "interline_equation", + "content": "\\hat {y} _ {\\text {a c t i o n}} = \\operatorname {s o f t m a x} \\left(F C \\left(f _ {s v} ^ {e x o}\\right)\\right). \\tag {14}", + "image_path": "7e0c09228086faf74407e268801e6581d016bfb24e4d7b1f9f5224093f765df4.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 312, + 592, + 517, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 592, + 517, + 622 + ], + "spans": [ + { + "bbox": [ + 312, + 592, + 517, + 622 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 312, + 592, + 517, + 622 + ], + "type": "inline_equation", + "content": "FC" + }, + { + "bbox": [ + 312, + 592, + 517, + 622 + ], + "type": "text", + "content": " is a fully connected layer that maps the sound-aligned feature to action classes. We train the FC layers using cross-entropy loss with the ground truth labels." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 312, + 629, + 382, + 641 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 629, + 382, + 641 + ], + "spans": [ + { + "bbox": [ + 312, + 629, + 382, + 641 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 312, + 646, + 517, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 646, + 517, + 666 + ], + "spans": [ + { + "bbox": [ + 312, + 646, + 517, + 666 + ], + "type": "text", + "content": "In this section, we first evaluate the effectiveness of the sound-aware representation space on Ego-Exo video re" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 724, + 317, + 734 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 724, + 317, + 734 + ], + "spans": [ + { + "bbox": [ + 293, + 724, + 317, + 734 + ], + "type": "text", + "content": "28945" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 91, + 120, + 297, + 178 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 120, + 297, + 178 + ], + "spans": [ + { + "bbox": [ + 91, + 120, + 297, + 178 + ], + "type": "text", + "content": "trieval and text-video retrieval tasks. Next, we apply the sound-aware visual representations to two downstream tasks: scene recognition in egocentric videos and action recognition in exocentric videos. Finally, we conduct ablation studies to verify the effectiveness of our proposed modules." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 93, + 186, + 201, + 198 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 186, + 201, + 198 + ], + "spans": [ + { + "bbox": [ + 93, + 186, + 201, + 198 + ], + "type": "text", + "content": "5.1. Experimental Settings" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 93, + 201, + 236, + 211 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 201, + 236, + 211 + ], + "spans": [ + { + "bbox": [ + 93, + 201, + 236, + 211 + ], + "type": "text", + "content": "5.1.1 Datasets and Evaluation Metric" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 91, + 217, + 297, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 217, + 297, + 479 + ], + "spans": [ + { + "bbox": [ + 91, + 217, + 297, + 479 + ], + "type": "text", + "content": "We conduct the retrieval tasks on two Ego-Exo benchmarks. One benchmark consists of strictly paired Ego-Exo videos (i.e., CharadesEgo [1]), which are recorded using the same shooting script. The other benchmark consists of loosely paired Ego-Exo videos (i.e., EgoExoLearn [12]), where most video pairs share only the same verbs and nouns in their descriptions. CharadesEgo [1] is a large-scale EgoExo dataset featuring videos of daily activities. It comprises 7,985 video clips with a total duration of approximately 80 hours. The dataset is divided into training, validation, and test subsets, containing 5,338, 1,334, and 1,313 video pairs, respectively, along with corresponding shooting script descriptions. EgoExoLearn [12] is another large-scale EgoExo dataset, consisting of 747 video clips with a total duration of 120 hours. Following [34], we construct 36,372 Ego-Exo pairs by aligning nouns and verbs in the video text descriptions. We use the official data split, which includes 36,373 video pairs for training, 800 for validation, and 2,200 for testing. The downstream tasks are also conducted on CharadesEgo [1] and EgoExoLearn [12]. CharadesEgo contains 16 scene categories (e.g., kitchen and bedroom), while EgoExoLearn covers 19 action categories from both daily life and laboratory scenes (e.g., pick-up and washing). We use CharadesEgo for egocentric video scene recognition and EgoExoLearn for exocentric video action recognition. Following previous works [12, 34], we report top-1 accuracy for both retrieval and recognition tasks." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 93, + 495, + 206, + 504 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 495, + 206, + 504 + ], + "spans": [ + { + "bbox": [ + 93, + 495, + 206, + 504 + ], + "type": "text", + "content": "5.1.2 Implementation Details" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 91, + 510, + 297, + 646 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 510, + 297, + 646 + ], + "spans": [ + { + "bbox": [ + 91, + 510, + 297, + 646 + ], + "type": "text", + "content": "We use the video transformer from [3] with a ViT-B/16 backbone initialized with CLIP weights [23] as the video encoder " + }, + { + "bbox": [ + 91, + 510, + 297, + 646 + ], + "type": "inline_equation", + "content": "E_{v}(x)" + }, + { + "bbox": [ + 91, + 510, + 297, + 646 + ], + "type": "text", + "content": ". For the sound-vision module, we apply the audio encoder from [6] as " + }, + { + "bbox": [ + 91, + 510, + 297, + 646 + ], + "type": "inline_equation", + "content": "E_{a}(x)" + }, + { + "bbox": [ + 91, + 510, + 297, + 646 + ], + "type": "text", + "content": ", keeping it frozen during training. In the sound-text module, we use the pretrained CLIP textual encoder as " + }, + { + "bbox": [ + 91, + 510, + 297, + 646 + ], + "type": "inline_equation", + "content": "E_{t}(x)" + }, + { + "bbox": [ + 91, + 510, + 297, + 646 + ], + "type": "text", + "content": " to encode both semantic text and sound captions. The output dimension of the sound-aligned Ego-Exo representation space is 256. We use Llama3-8B as the LLM to generate sound captions. In the multi-head cross-attention module, we set the number of heads to " + }, + { + "bbox": [ + 91, + 510, + 297, + 646 + ], + "type": "inline_equation", + "content": "h = 4" + }, + { + "bbox": [ + 91, + 510, + 297, + 646 + ], + "type": "text", + "content": ". The temperature parameter " + }, + { + "bbox": [ + 91, + 510, + 297, + 646 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 91, + 510, + 297, + 646 + ], + "type": "text", + "content": " is set to 0.07. We utilize 8 V100 GPUs to train the sound-aware representation space with a learning rate of 3e-5, using AdamW as the optimizer for improved convergence and stability." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 92, + 646, + 297, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 92, + 646, + 297, + 667 + ], + "spans": [ + { + "bbox": [ + 92, + 646, + 297, + 667 + ], + "type": "text", + "content": "It is worth noting that for video pairs with high synchronization, applying the Dynamic Time Warping (DTW) algo" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 313, + 144, + 518, + 240 + ], + "blocks": [ + { + "bbox": [ + 312, + 118, + 517, + 137 + ], + "lines": [ + { + "bbox": [ + 312, + 118, + 517, + 137 + ], + "spans": [ + { + "bbox": [ + 312, + 118, + 517, + 137 + ], + "type": "text", + "content": "Table 1. Retrieval result comparison (top-1 accuracy) of Exo-Ego video alignment task on two benchmarks." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 313, + 144, + 518, + 240 + ], + "lines": [ + { + "bbox": [ + 313, + 144, + 518, + 240 + ], + "spans": [ + { + "bbox": [ + 313, + 144, + 518, + 240 + ], + "type": "table", + "html": "
MethodEgoExoLearnCharadesEgo
Eg2ExEx2EgAvgEg2ExEx2EgAvg
FrozenInTime [3]14.1013.4013.756.734.965.85
LaViLa [36]28.7025.7027.2058.2446.9352.59
EgoVLP [17]32.1028.930.5058.2760.7659.52
InternVideo [32]30.6021.7026.1553.5561.2357.39
EgoExo [12]49.0045.3047.1568.3262.0665.19
EgoInstructor [34]47.1045.8246.4662.0656.3859.22
SoundBridge50.2750.7350.5077.6671.6374.65
", + "image_path": "3e2b74335314434726a3ec48af61d1455bc84b8ed88a201f2d418b2844017e41.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 312, + 248, + 517, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 248, + 517, + 299 + ], + "spans": [ + { + "bbox": [ + 312, + 248, + 517, + 299 + ], + "type": "text", + "content": "rhythm [24] to align the two audio sequences before inputting the audio signals into the cross-attention module can help mitigate the impact of varying sound speeds. We use this setting by default for videos in CharadesEgo and provide an ablation study on DTW in the supplementary material." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 312, + 309, + 400, + 319 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 309, + 400, + 319 + ], + "spans": [ + { + "bbox": [ + 312, + 309, + 400, + 319 + ], + "type": "text", + "content": "5.1.3 Baseline Models" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 312, + 324, + 517, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 324, + 517, + 441 + ], + "spans": [ + { + "bbox": [ + 312, + 324, + 517, + 441 + ], + "type": "text", + "content": "We include several types of models as baselines. These include recent text-video representation models pre-trained on large-scale exocentric datasets, such as FrozenInTime [3] and InternVideo [32]. Additionally, we include pretrained models on large-scale egocentric datasets, namely LaViLa [36] and EgoVLP [17]. Furthermore, we compare our method with two recent approaches designed for EgoExo video alignment, namely EgoExo [12] and EgoInstructor [34]. For the downstream tasks, we also include task-specific models, namely VideoMAE [30] for exocentric action recognition and InternImage [31] for egocentric scene recognition." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 312, + 447, + 480, + 459 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 447, + 480, + 459 + ], + "spans": [ + { + "bbox": [ + 312, + 447, + 480, + 459 + ], + "type": "text", + "content": "5.2. Results on Ego-Exo Video Alignment" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 312, + 462, + 518, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 462, + 518, + 667 + ], + "spans": [ + { + "bbox": [ + 312, + 462, + 518, + 667 + ], + "type": "text", + "content": "Table 1 compares the results of Ego-to-Exo and Exo-to-Ego video retrieval with other baselines. Specifically, we report the zero-shot retrieval performance of the FrozenIn-Time model and the fine-tuned results of other methods on the EgoExoLearn and CharadesEgo datasets. Our proposed method, SoundBridge, consistently outperforms other baselines, including pre-trained egocentric models, pre-trained exocentric models, and recent Ego-Exo models, on both Ego-to-Exo and Exo-to-Ego video retrieval tasks across the two benchmarks. We are the first model to achieve a top-1 accuracy exceeding " + }, + { + "bbox": [ + 312, + 462, + 518, + 667 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 312, + 462, + 518, + 667 + ], + "type": "text", + "content": " on the EgoExoLearn dataset. Specifically, SoundBridge achieves higher retrieval accuracy than EgoExo [12] and EgoInstructor [34] by " + }, + { + "bbox": [ + 312, + 462, + 518, + 667 + ], + "type": "inline_equation", + "content": "8.13\\%" + }, + { + "bbox": [ + 312, + 462, + 518, + 667 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 312, + 462, + 518, + 667 + ], + "type": "inline_equation", + "content": "15.93\\%" + }, + { + "bbox": [ + 312, + 462, + 518, + 667 + ], + "type": "text", + "content": " on Ego-to-Exo video retrieval, and by " + }, + { + "bbox": [ + 312, + 462, + 518, + 667 + ], + "type": "inline_equation", + "content": "13.71\\%" + }, + { + "bbox": [ + 312, + 462, + 518, + 667 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 312, + 462, + 518, + 667 + ], + "type": "inline_equation", + "content": "18.88\\%" + }, + { + "bbox": [ + 312, + 462, + 518, + 667 + ], + "type": "text", + "content": " on Exo-to-Ego video retrieval, respectively. The performance boost primarily comes from the incorporation of sound-aware modules in the alignment process. For example, given an egocentric video query recording picking up the garlic with the left hand, the EgoExo model ranks an exocentric video of put the ginger down with both hand on the iron bowl in the top-1 position. In contrast, our" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 724, + 318, + 734 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 724, + 318, + 734 + ], + "spans": [ + { + "bbox": [ + 293, + 724, + 318, + 734 + ], + "type": "text", + "content": "28946" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 102, + 118, + 508, + 238 + ], + "blocks": [ + { + "bbox": [ + 102, + 118, + 508, + 238 + ], + "lines": [ + { + "bbox": [ + 102, + 118, + 508, + 238 + ], + "spans": [ + { + "bbox": [ + 102, + 118, + 508, + 238 + ], + "type": "image", + "image_path": "a483729a140a77a2666eb0ddcfb786f9c61e9f26ff1360a6e3d17caf0ef9ae4f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 91, + 247, + 517, + 266 + ], + "lines": [ + { + "bbox": [ + 91, + 247, + 517, + 266 + ], + "spans": [ + { + "bbox": [ + 91, + 247, + 517, + 266 + ], + "type": "text", + "content": "Figure 3. Visualization of an Ego-to-Exo video retrieval example, compared with EgoExo [12] and EgoInstructor [34], on an egocentric video query looking for A person is working on a vacuum in the man cave, then sneezes and reaches for some medicine." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 102, + 275, + 507, + 396 + ], + "blocks": [ + { + "bbox": [ + 102, + 275, + 507, + 396 + ], + "lines": [ + { + "bbox": [ + 102, + 275, + 507, + 396 + ], + "spans": [ + { + "bbox": [ + 102, + 275, + 507, + 396 + ], + "type": "image", + "image_path": "1cfded1faa5ed2bdf84ee725960a128edc2f1bfa2bd5295f33969693c527ba44.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 91, + 405, + 517, + 424 + ], + "lines": [ + { + "bbox": [ + 91, + 405, + 517, + 424 + ], + "spans": [ + { + "bbox": [ + 91, + 405, + 517, + 424 + ], + "type": "text", + "content": "Figure 4. Visualization of an Exo-to-Ego video retrieval example, compared with EgoExo [12] and EgoInstructor [34], on an exocentric video query looking for a person throws a box of groceries onto the table. The person opens the box and takes out an item." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 91, + 441, + 297, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 441, + 297, + 519 + ], + "spans": [ + { + "bbox": [ + 91, + 441, + 297, + 519 + ], + "type": "text", + "content": "method differentiates between the actions \"pick-up\" and \"put down\" using sound cues. The pick-up action produces a soft rustling sound, whereas the put-down action generates a clunk sound when interacting with the iron bowl. Our proposed sound-aware modules effectively encode these audio features and their semantic descriptions (e.g., [pick up ginger] (soft rustling)). As a result, our method correctly ranks the ground-truth exocentric video in the top-1 position." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 91, + 520, + 298, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 520, + 298, + 665 + ], + "spans": [ + { + "bbox": [ + 91, + 520, + 298, + 665 + ], + "type": "text", + "content": "Figure 3 visualizes an Ego-to-Exo video retrieval example, comparing our method with EgoExo [12] and EgoInstructor [34]. The egocentric video depicts a person working on a vacuum, sneezing, and picking up a medicine bottle, while the visual cue for the sneeze is barely visible in the egocentric video. All methods successfully rank an egocentric video of a person with a vacuum in the top-1 position. However, EgoExo and EgoInstructor fail to capture either the sneezing action or the action of reaching for the medicine, leading to incorrect top-1 results. In contrast, our method effectively detects the sneeze and the action of picking up the pill bottle using sound cues, thereby successfully retrieving the ground-truth exocentric video. Similarly, Figure 4 visualizes an Exo-to-Ego retrieval example with a query for: a person throws a box of groceries onto" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 312, + 441, + 517, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 441, + 517, + 499 + ], + "spans": [ + { + "bbox": [ + 312, + 441, + 517, + 499 + ], + "type": "text", + "content": "the table. The person opens the box and takes out an item. All methods successfully retrieve videos that capture the actions of opening the box and taking out items. However, only our SoundBridge method correctly identifies the action of throwing the box onto the table, leveraging the thud sound as an essential cue." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 312, + 505, + 459, + 516 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 505, + 459, + 516 + ], + "spans": [ + { + "bbox": [ + 312, + 505, + 459, + 516 + ], + "type": "text", + "content": "5.3. Results on Text-Video Retrieval" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 311, + 521, + 518, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 521, + 518, + 667 + ], + "spans": [ + { + "bbox": [ + 311, + 521, + 518, + 667 + ], + "type": "text", + "content": "The proposed sound-aware Ego-Exo representation space also enables text-video retrieval. Therefore, we compare its performance with baselines on both ego-(exo-)video-to-text and text-to-ego-(exo-)video retrieval tasks. Table 2 demonstrates that our method significantly enhances text-video retrieval performance on Ego-Exo datasets. Specifically, our method substantially improves the performance of the pre-trained egocentric model LaViLa [36], and the pre-trained exocentric model InternVideo [32]. Furthermore, our method outperforms recent Ego-Exo models, namely EgoExo [12] and EgoInstructor [34], by approximately " + }, + { + "bbox": [ + 311, + 521, + 518, + 667 + ], + "type": "inline_equation", + "content": "30.50\\%" + }, + { + "bbox": [ + 311, + 521, + 518, + 667 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 311, + 521, + 518, + 667 + ], + "type": "inline_equation", + "content": "23.47\\%" + }, + { + "bbox": [ + 311, + 521, + 518, + 667 + ], + "type": "text", + "content": " in overall performance across both datasets. The main improvement over recent Ego-Exo models is observed in queries with multiple nearest-neighbor (NN) answers. By incorporating sound aware" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 724, + 317, + 734 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 724, + 317, + 734 + ], + "spans": [ + { + "bbox": [ + 293, + 724, + 317, + 734 + ], + "type": "text", + "content": "28947" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 93, + 135, + 298, + 236 + ], + "blocks": [ + { + "bbox": [ + 99, + 118, + 290, + 128 + ], + "lines": [ + { + "bbox": [ + 99, + 118, + 290, + 128 + ], + "spans": [ + { + "bbox": [ + 99, + 118, + 290, + 128 + ], + "type": "text", + "content": "Table 2. Performance comparison on text-video retrieval task." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 93, + 135, + 298, + 236 + ], + "lines": [ + { + "bbox": [ + 93, + 135, + 298, + 236 + ], + "spans": [ + { + "bbox": [ + 93, + 135, + 298, + 236 + ], + "type": "table", + "html": "
MethodEgoExoLearnCharadesEgo
T2VV2TAvgT2VV2TAvg
FrozenInTime [3]6.034.055.049.405.627.51
LaViLa [36]34.8740.1837.5343.7441.6742.71
EgoVLP [17]19.0926.1022.6045.5143.2144.36
InternVideo [32]18.5926.5522.5747.3446.6346.99
EgoExo [12]58.6458.7358.6948.0045.2146.61
EgoInstructor [34]66.3766.6866.5346.9345.9846.46
SoundBridge71.4173.1472.2864.9063.6064.25
", + "image_path": "07afd1a626762d99c078aac83d271451b12ce167020c90b460f0ad5f134c7bc2.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 91, + 243, + 298, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 243, + 298, + 342 + ], + "spans": [ + { + "bbox": [ + 91, + 243, + 298, + 342 + ], + "type": "text", + "content": "ness, our method effectively ranks the ground-truth videos higher among similar nearest-neighbor results. For example, given a text query searching for exocentric videos of a person washing dishes in the kitchen, EgoExo retrieves videos of a person chopping potatoes, while EgoInstruc-tor retrieves videos of a person stir-frying. In contrast, our method distinguishes the washing action among these nearest neighbors by leveraging water-running sounds, allowing the ground-truth video to be ranked higher than videos of chopping and stir-frying." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 93, + 350, + 231, + 361 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 350, + 231, + 361 + ], + "spans": [ + { + "bbox": [ + 93, + 350, + 231, + 361 + ], + "type": "text", + "content": "5.4. Results on Downstream Tasks" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 91, + 365, + 297, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 365, + 297, + 453 + ], + "spans": [ + { + "bbox": [ + 91, + 365, + 297, + 453 + ], + "type": "text", + "content": "We also evaluate the effectiveness of the learned egocentric visual feature (i.e., " + }, + { + "bbox": [ + 91, + 365, + 297, + 453 + ], + "type": "inline_equation", + "content": "f_{sv}^{ego}" + }, + { + "bbox": [ + 91, + 365, + 297, + 453 + ], + "type": "text", + "content": ") for scene recognition and the learned exocentric visual feature (i.e., " + }, + { + "bbox": [ + 91, + 365, + 297, + 453 + ], + "type": "inline_equation", + "content": "f_{sv}^{exo}" + }, + { + "bbox": [ + 91, + 365, + 297, + 453 + ], + "type": "text", + "content": ") for action recognition. We evaluate our model under two settings: (1) fine-tuning, where the training and test sets come from the same dataset, and (2) transfer learning, where the model is trained on one dataset and tested on another. The report performances of other models are all with the fine-tuning setting." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 91, + 453, + 298, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 453, + 298, + 666 + ], + "spans": [ + { + "bbox": [ + 91, + 453, + 298, + 666 + ], + "type": "text", + "content": "Table 3 compares the top-1 recognition accuracy with baseline methods. Specifically, for the egocentric scene recognition task, our method improves upon the egocentric model LaViLa [36] and the Ego-Exo model EgoInstructor [34] by " + }, + { + "bbox": [ + 91, + 453, + 298, + 666 + ], + "type": "inline_equation", + "content": "23.42\\%" + }, + { + "bbox": [ + 91, + 453, + 298, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 91, + 453, + 298, + 666 + ], + "type": "inline_equation", + "content": "9.78\\%" + }, + { + "bbox": [ + 91, + 453, + 298, + 666 + ], + "type": "text", + "content": ", respectively. Figure 5 illustrates an example of scene recognition. Given an egocentric video query where only a hand wiping a table with a towel is visible, it is difficult to infer that the scene is a kitchen. Other methods fail to predict the correct scene label. For example, FrozenInTime [3] misclassifies the video as a living room, while LaViLa [36] predicts it as a bathroom. In contrast, our method correctly identifies the scene by successfully aligning the query video with the corresponding exocentric video using sound cues from multiple actions (e.g., walking to the table and leaving the room). The egocentric visual feature (i.e., " + }, + { + "bbox": [ + 91, + 453, + 298, + 666 + ], + "type": "inline_equation", + "content": "f_{sv}^{ego}" + }, + { + "bbox": [ + 91, + 453, + 298, + 666 + ], + "type": "text", + "content": ") learns environmental and global visual cues from the exocentric video, which contains a range hood, an oven, and a refrigerator, allowing it to infer that the scene is a kitchen. For exocentric action recognition, our method achieves an average accuracy of " + }, + { + "bbox": [ + 91, + 453, + 298, + 666 + ], + "type": "inline_equation", + "content": "56.25\\%" + }, + { + "bbox": [ + 91, + 453, + 298, + 666 + ], + "type": "text", + "content": ", which is lower than VideoMAE [30], as the model is trained on a large-scale action dataset. Despite this, our" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 314, + 119, + 515, + 173 + ], + "blocks": [ + { + "bbox": [ + 314, + 119, + 515, + 173 + ], + "lines": [ + { + "bbox": [ + 314, + 119, + 515, + 173 + ], + "spans": [ + { + "bbox": [ + 314, + 119, + 515, + 173 + ], + "type": "image", + "image_path": "f8f7697f2b8942fb7b259c5c2c74f92bfe93fecb5f1205f8579975ae616a260d.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 312, + 179, + 517, + 198 + ], + "lines": [ + { + "bbox": [ + 312, + 179, + 517, + 198 + ], + "spans": [ + { + "bbox": [ + 312, + 179, + 517, + 198 + ], + "type": "text", + "content": "Figure 5. Visualization of the egocentric scene recognition results compared with FrozenInTime [3] and LaViLa [36]." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 317, + 209, + 515, + 262 + ], + "blocks": [ + { + "bbox": [ + 317, + 209, + 515, + 262 + ], + "lines": [ + { + "bbox": [ + 317, + 209, + 515, + 262 + ], + "spans": [ + { + "bbox": [ + 317, + 209, + 515, + 262 + ], + "type": "image", + "image_path": "4e538bb7800667e0b8c86a8c7206f5e9042021b611e147d188f064d69ac48f20.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 312, + 269, + 517, + 289 + ], + "lines": [ + { + "bbox": [ + 312, + 269, + 517, + 289 + ], + "spans": [ + { + "bbox": [ + 312, + 269, + 517, + 289 + ], + "type": "text", + "content": "Figure 6. Visualization of the exocentric action recognition results compared with InternVideo [32] and EgoInstructor [34]." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 313, + 342, + 517, + 455 + ], + "blocks": [ + { + "bbox": [ + 312, + 297, + 517, + 335 + ], + "lines": [ + { + "bbox": [ + 312, + 297, + 517, + 335 + ], + "spans": [ + { + "bbox": [ + 312, + 297, + 517, + 335 + ], + "type": "text", + "content": "Table 3. Result comparison on CharadesEgo scene recognition and EgoExoLearn action recognition. An asterisk (*) indicates the transfer learning setting, where a different dataset is used for training." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 313, + 342, + 517, + 455 + ], + "lines": [ + { + "bbox": [ + 313, + 342, + 517, + 455 + ], + "spans": [ + { + "bbox": [ + 313, + 342, + 517, + 455 + ], + "type": "table", + "html": "
MethodEgo-SceneExo-Action
FrozenInTime [3]8.256.45
LaViLa [36]41.3726.25
EgoVLP [17]40.3138.75
InternVideo [32]43.0045.00
EgoExo [12]42.5546.25
EgoInstructor [34]46.5147.50
VideoMAEv2 [30]-62.50
InterImage [31]39.60-
SoundBridge*48.7652.50
SoundBridge51.0656.25
", + "image_path": "f2038d25561f482732a244e344f12a3ba6fba96adabbafec7ef050a075d8f2e4.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 311, + 463, + 518, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 463, + 518, + 571 + ], + "spans": [ + { + "bbox": [ + 311, + 463, + 518, + 571 + ], + "type": "text", + "content": "approach benefits from the sound-aware Exo-Ego representation space, as shown in Figure 6. Predicting a person's action from an exocentric view is challenging due to occlusions caused by the human body and the person being far from the camera. InternVideo [32] misclassifies the query video as the action \"put\", while EgoInstructor [34] assigns a high probability to the action \"pick-up\". In contrast, our method aligns the query video with its corresponding egocentric view using the sound of washing dishes (e.g., water splashing) and correctly classifies the action as \"washing\" with the highest probability." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 312, + 576, + 391, + 587 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 576, + 391, + 587 + ], + "spans": [ + { + "bbox": [ + 312, + 576, + 391, + 587 + ], + "type": "text", + "content": "5.5. Ablation study" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 312, + 591, + 439, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 591, + 439, + 601 + ], + "spans": [ + { + "bbox": [ + 312, + 591, + 439, + 601 + ], + "type": "text", + "content": "5.5.1 Results on Different LLMs" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 311, + 607, + 517, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 607, + 517, + 666 + ], + "spans": [ + { + "bbox": [ + 311, + 607, + 517, + 666 + ], + "type": "text", + "content": "We evaluated the impact of audio descriptions generated by different Large Language Models (LLMs) on the performance of cross-view video association tasks. Specifically, we used GPT-3.5 [4] and LLaMA3 [7] to generate audio descriptions using the same prompt. As shown in Table 5, the results indicate that while the cross-view video" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 724, + 318, + 734 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 724, + 318, + 734 + ], + "spans": [ + { + "bbox": [ + 293, + 724, + 318, + 734 + ], + "type": "text", + "content": "28948" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 94, + 136, + 517, + 201 + ], + "blocks": [ + { + "bbox": [ + 183, + 118, + 426, + 129 + ], + "lines": [ + { + "bbox": [ + 183, + 118, + 426, + 129 + ], + "spans": [ + { + "bbox": [ + 183, + 118, + 426, + 129 + ], + "type": "text", + "content": "Table 4. Ablation study on the proposed sound-vision and sound-text modules." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 94, + 136, + 517, + 201 + ], + "lines": [ + { + "bbox": [ + 94, + 136, + 517, + 201 + ], + "spans": [ + { + "bbox": [ + 94, + 136, + 517, + 201 + ], + "type": "table", + "html": "
MethodEgoExoLearnCharadesEgo
Eg2ExEx2EgT2VV2TEg2ExEx2EgT2VV2T
Baseline49.0045.3058.6458.7368.3262.0648.0045.21
Baseline + Sound-Text Module48.8250.1869.2371.9672.8165.2563.5561.35
Baseline + Sound-Vision Module48.4647.7369.1073.0974.3566.5563.5961.11
Baseline + Sound-Text&Vision Modules50.2750.7371.4173.1477.6671.6364.9063.60
", + "image_path": "935c1ea4eb1ae08b61690a50fe549bb57a02200f1a94e2ed6a7a4df24904ad65.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 94, + 232, + 297, + 280 + ], + "blocks": [ + { + "bbox": [ + 135, + 216, + 253, + 224 + ], + "lines": [ + { + "bbox": [ + 135, + 216, + 253, + 224 + ], + "spans": [ + { + "bbox": [ + 135, + 216, + 253, + 224 + ], + "type": "text", + "content": "Table 5. Results with different LLMs." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 94, + 232, + 297, + 280 + ], + "lines": [ + { + "bbox": [ + 94, + 232, + 297, + 280 + ], + "spans": [ + { + "bbox": [ + 94, + 232, + 297, + 280 + ], + "type": "table", + "html": "
LLMEgoExoLearnCharadesEgo
Eg2ExEx2EgAvgEg2ExEx2EgAvg
GPT-3.549.2748.3648.8273.7671.5172.64
LLaMA350.2750.7350.5077.6671.6374.65
", + "image_path": "32fdd16a18cc10b8f999769cac3cf2e87b7ab54dadf1aa391983970f187cf56b.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 94, + 305, + 297, + 365 + ], + "blocks": [ + { + "bbox": [ + 110, + 289, + 279, + 298 + ], + "lines": [ + { + "bbox": [ + 110, + 289, + 279, + 298 + ], + "spans": [ + { + "bbox": [ + 110, + 289, + 279, + 298 + ], + "type": "text", + "content": "Table 6. Performance comparison of different anchors." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 94, + 305, + 297, + 365 + ], + "lines": [ + { + "bbox": [ + 94, + 305, + 297, + 365 + ], + "spans": [ + { + "bbox": [ + 94, + 305, + 297, + 365 + ], + "type": "table", + "html": "
AnchorEgoExoLearnCharadesEgo
Eg2ExEx2EgEg2ExEx2Eg
Origin Caption48.6448.7374.3566.55
Sound Caption47.0949.9156.9751.42
Sound Feature36.6726.6756.2748.38
Sound-Text Module50.2750.7377.6671.63
", + "image_path": "59ccdc396d19220324bd66f4d9a641d325e73180f2ada8bcb4a7eaa442124e22.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 92, + 372, + 297, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 92, + 372, + 297, + 411 + ], + "spans": [ + { + "bbox": [ + 92, + 372, + 297, + 411 + ], + "type": "text", + "content": "association performance with GPT-3.5 is lower than that achieved with LLaMA3, our method still surpasses existing approaches. This demonstrates the effectiveness of our proposed method." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 93, + 426, + 253, + 435 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 426, + 253, + 435 + ], + "spans": [ + { + "bbox": [ + 93, + 426, + 253, + 435 + ], + "type": "text", + "content": "5.5.2 Impact of the Sound-aware Modules" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 92, + 441, + 297, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 92, + 441, + 297, + 538 + ], + "spans": [ + { + "bbox": [ + 92, + 441, + 297, + 538 + ], + "type": "text", + "content": "We conduct an ablation study on the proposed sound-vision and sound-text modules, using the state-of-the-art Ego-Exo model EgoExo [12] as the baseline. The performance of Ego-Exo video alignment and the overall accuracy of text-video retrieval are presented in Table 4. When using only the sound-vision or sound-text module, our method demonstrates competitive performance in the Ego-to-Exo video retrieval task. However, when both modules are jointly utilized, our method achieves significantly higher performance across the evaluated retrieval tasks." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 93, + 552, + 224, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 552, + 224, + 562 + ], + "spans": [ + { + "bbox": [ + 93, + 552, + 224, + 562 + ], + "type": "text", + "content": "5.5.3 Results of Different Anchors" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 92, + 568, + 297, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 92, + 568, + 297, + 666 + ], + "spans": [ + { + "bbox": [ + 92, + 568, + 297, + 666 + ], + "type": "text", + "content": "We also investigate the effectiveness of using different features as anchors to align sound-aware visual representations. Specifically, we replace the " + }, + { + "bbox": [ + 92, + 568, + 297, + 666 + ], + "type": "inline_equation", + "content": "f_{st}" + }, + { + "bbox": [ + 92, + 568, + 297, + 666 + ], + "type": "text", + "content": " in Eq. (11) and Eq. (12) with " + }, + { + "bbox": [ + 92, + 568, + 297, + 666 + ], + "type": "inline_equation", + "content": "f_{t}, f_{c}" + }, + { + "bbox": [ + 92, + 568, + 297, + 666 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 92, + 568, + 297, + 666 + ], + "type": "inline_equation", + "content": "\\text{sum}(f_{a}^{ego}, f_{a}^{exo})" + }, + { + "bbox": [ + 92, + 568, + 297, + 666 + ], + "type": "text", + "content": ". As shown in Table 6, using audio alone as an anchor is not ideal, especially for the short video dataset EgoExoLearn, where events heavily overlap and exhibit only slight variations. Upon analyzing the erroneous cases, we found that in many samples, the audio anchor alone fails to provide sufficient information to distinguish between videos from different views. As il" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 315, + 216, + 514, + 289 + ], + "blocks": [ + { + "bbox": [ + 315, + 216, + 514, + 289 + ], + "lines": [ + { + "bbox": [ + 315, + 216, + 514, + 289 + ], + "spans": [ + { + "bbox": [ + 315, + 216, + 514, + 289 + ], + "type": "image", + "image_path": "5605f5e34f91f4b7e706fce05920b5b7331a927c86abbc3b344732b3f512eb60.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 312, + 297, + 517, + 316 + ], + "lines": [ + { + "bbox": [ + 312, + 297, + 517, + 316 + ], + "spans": [ + { + "bbox": [ + 312, + 297, + 517, + 316 + ], + "type": "text", + "content": "Figure 7. An example of Exo-Ego video retrieval using different anchors in the SoundBridge model." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 312, + 326, + 517, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 326, + 517, + 413 + ], + "spans": [ + { + "bbox": [ + 312, + 326, + 517, + 413 + ], + "type": "text", + "content": "lustrated in Figure 7, a person is slicing garlic in the query video, and the text describes the semantic information of \"a person slicing garlic\". However, the sound features only convey the action \"slice\". During training, different-view videos were aligned solely based on the \"slice\" action, leading to incorrect retrieval of a video showing \"slicing ginger\" instead of \"slicing garlic\". Similarly, the sound caption lacks full contextual information and misinterprets the one-hand slicing action as two-hand slicing." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 312, + 426, + 375, + 436 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 426, + 375, + 436 + ], + "spans": [ + { + "bbox": [ + 312, + 426, + 375, + 436 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 312, + 442, + 517, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 442, + 517, + 568 + ], + "spans": [ + { + "bbox": [ + 312, + 442, + 517, + 568 + ], + "type": "text", + "content": "In this paper, we proposed two sound-aware modules as a bridge to address the large visual variance challenge in egocentric and exocentric video alignment. Both modules demonstrated significant improvements over state-of-the-art methods in Ego-Exo video retrieval and text-video retrieval tasks. Moreover, we verified that both egocentric and exocentric representations benefit from the sound-aware learning paradigm, achieving substantial improvements in downstream recognition tasks by leveraging complementary information from a correctly associated view. Furthermore, we found that our sound-vision and sound-text modules provide effective cues to distinguish the target video from its nearest neighbors based on sound." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 312, + 581, + 400, + 593 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 581, + 400, + 593 + ], + "spans": [ + { + "bbox": [ + 312, + 581, + 400, + 593 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 312, + 597, + 517, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 597, + 517, + 666 + ], + "spans": [ + { + "bbox": [ + 312, + 597, + 517, + 666 + ], + "type": "text", + "content": "This research was supported by the National Natural Science Foundation of China (Grant No.: 62372314, 62476097); the Guangdong Provincial Fund for Basic and Applied Basic Research—Regional Joint Fund Project (Key Project) (2023B1515120078); the Guangdong Provincial Natural Science Foundation for Outstanding Youth Team Project (2024B1515040010)." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 724, + 317, + 734 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 724, + 317, + 734 + ], + "spans": [ + { + "bbox": [ + 293, + 724, + 317, + 734 + ], + "type": "text", + "content": "28949" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 94, + 119, + 143, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 119, + 143, + 129 + ], + "spans": [ + { + "bbox": [ + 94, + 119, + 143, + 129 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 94, + 136, + 298, + 666 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 98, + 136, + 297, + 175 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 98, + 136, + 297, + 175 + ], + "spans": [ + { + "bbox": [ + 98, + 136, + 297, + 175 + ], + "type": "text", + "content": "[1] Karteek Alahari. Actor and observer: Joint modeling of first and third-person videos. In Proceedings of the 1st Workshop and Challenge on Comprehensive Video Understanding in the Wild, page 3, 2018. 2, 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 98, + 176, + 297, + 214 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 98, + 176, + 297, + 214 + ], + "spans": [ + { + "bbox": [ + 98, + 176, + 297, + 214 + ], + "type": "text", + "content": "[2] Relja Arandjelovic and Andrew Zisserman. Look, listen and learn. In Proceedings of the IEEE international conference on computer vision, pages 609-617, 2017. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 98, + 215, + 298, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 98, + 215, + 298, + 263 + ], + "spans": [ + { + "bbox": [ + 98, + 215, + 298, + 263 + ], + "type": "text", + "content": "[3] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1728-1738, 2021. 5, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 98, + 264, + 297, + 322 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 98, + 264, + 297, + 322 + ], + "spans": [ + { + "bbox": [ + 98, + 264, + 297, + 322 + ], + "type": "text", + "content": "[4] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 98, + 323, + 297, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 98, + 323, + 297, + 361 + ], + "spans": [ + { + "bbox": [ + 98, + 323, + 297, + 361 + ], + "type": "text", + "content": "[5] Minjie Cai, Kris M Kitani, and Yoichi Sato. Understanding hand-object manipulation with grasp types and object attributes. In Robotics: Science and Systems, 2016. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 98, + 362, + 297, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 98, + 362, + 297, + 402 + ], + "spans": [ + { + "bbox": [ + 98, + 362, + 297, + 402 + ], + "type": "text", + "content": "[6] Sanyuan Chen, Yu Wu, Chengyi Wang, Shujie Liu, Daniel Tompkins, Zhuo Chen, and Furu Wei. Beats: Audio pre-training with acoustic tokenizers. arXiv preprint arXiv:2212.09058, 2022. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 98, + 403, + 297, + 450 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 98, + 403, + 297, + 450 + ], + "spans": [ + { + "bbox": [ + 98, + 403, + 297, + 450 + ], + "type": "text", + "content": "[7] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 98, + 451, + 297, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 98, + 451, + 297, + 500 + ], + "spans": [ + { + "bbox": [ + 98, + 451, + 297, + 500 + ], + "type": "text", + "content": "[8] Antonino Furnari and Giovanni Maria Farinella. What would you expect? anticipating egocentric actions with rolling-unrolling lstms and modality attention. In Proceedings of the IEEE/CVF International conference on computer vision, pages 6252-6261, 2019. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 98, + 501, + 297, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 98, + 501, + 297, + 548 + ], + "spans": [ + { + "bbox": [ + 98, + 501, + 297, + 548 + ], + "type": "text", + "content": "[9] Mohit Goyal, Sahil Modi, Rishabh Goyal, and Saurabh Gupta. Human hands as probes for interactive object understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3293-3303, 2022. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 94, + 549, + 298, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 549, + 298, + 666 + ], + "spans": [ + { + "bbox": [ + 94, + 549, + 298, + 666 + ], + "type": "text", + "content": "[10] Kristen Grauman, Andrew Westbury, Eugene Byrne, Zachary Chavis, Antonino Furnari, Rohit Girdhar, Jackson Hamburger, Hao Jiang, Miao Liu, Xingyu Liu, Miguel Martin, Tushar Nagarajan, Iija Radosavovic, Santhosh Kumar Ramakrishnan, Fiona Ryan, Jayant Sharma, Michael Wray, Mengmeng Xu, Eric Zhongcong Xu, Chen Zhao, Siddhant Bansal, Dhruv Batra, Vincent Cartillier, Sean Crane, Tien Do, Morrie Doulaty, Akshay Erapalli, Christoph Feichtenhofer, Adriano Fragomeni, Qichen Fu, Abrham Gebreselasie, Cristina Gonzalez, James Hillis, Xuhua Huang, Yifei Huang, Wenqi Jia, Weslie Khoo, Jachym" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 314, + 121, + 517, + 666 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 332, + 121, + 517, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 332, + 121, + 517, + 303 + ], + "spans": [ + { + "bbox": [ + 332, + 121, + 517, + 303 + ], + "type": "text", + "content": "Kolář, Satwik Kottur, Anurag Kumar, Federico Landini, Chao Li, Yanghao Li, Zhenqiang Li, Karttikeya Mangalam, Raghava Modhugu, Jonathan Munro, Tullie Murrell, Takumi Nishiyasu, Will Price, Paola Ruiz, Merey Ramazanova, Leda Sari, Kiran Somasundaram, Audrey Southerland, Yusuke Sugano, Ruijie Tao, Minh Vo, Yuchen Wang, Xindi Wu, Takuma Yagi, Ziwei Zhao, Yunyi Zhu, Pablo Arbeláez, David Crandall, Dima Damen, Giovanni Maria Farinella, Christian Fuegen, Bernard Ghanem, Vamsi Krishna Ithapu, C. V. Jawahar, Hanbyul Joo, Kris Kitani, Haizhou Li, Richard Newcombe, Aude Oliva, Hyun Soo Park, James M. Rehg, Yoichi Sato, Jianbo Shi, Mike Zheng Shou, Antonio Torralba, Lorenzo Torresani, Mingfei Yan, and Jitendra Malik. Ego4d: Around the world in 3,000 hours of egocentric video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18995-19012, 2022. 1, 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 304, + 517, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 304, + 517, + 363 + ], + "spans": [ + { + "bbox": [ + 314, + 304, + 517, + 363 + ], + "type": "text", + "content": "[11] Bo He, Jun Wang, Jielin Qiu, Trung Bui, Abhinav Shrivastava, and Zhaowen Wang. Align and attend: Multimodal summarization with dual contrastive losses. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14867-14878, 2023. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 314, + 364, + 517, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 364, + 517, + 441 + ], + "spans": [ + { + "bbox": [ + 314, + 364, + 517, + 441 + ], + "type": "text", + "content": "[12] Yifei Huang, Guo Chen, Jilan Xu, Mingfang Zhang, Lijin Yang, Baoqi Pei, Hongjie Zhang, Lu Dong, Yali Wang, Limin Wang, et al. Egoxolearn: A dataset for bridging asynchronous ego-and exo-centric view of procedural activities in real world. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22072-22086, 2024. 1, 2, 5, 6, 7, 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 314, + 441, + 516, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 441, + 516, + 499 + ], + "spans": [ + { + "bbox": [ + 314, + 441, + 516, + 499 + ], + "type": "text", + "content": "[13] Sarah Ibrahimi, Xiaohang Sun, Pichao Wang, Amanmeet Garg, Ashutosh Sanan, and Mohamed Omar. Audio-enhanced text-to-video retrieval using text-conditioned feature alignment. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12054-12064, 2023. 2, 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 314, + 500, + 516, + 538 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 500, + 516, + 538 + ], + "spans": [ + { + "bbox": [ + 314, + 500, + 516, + 538 + ], + "type": "text", + "content": "[14] Hao Jiang and Vamsi Krishna Ithapu. Egocentric pose estimation from human vision span. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 10986-10994. IEEE, 2021. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 314, + 539, + 516, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 539, + 516, + 578 + ], + "spans": [ + { + "bbox": [ + 314, + 539, + 516, + 578 + ], + "type": "text", + "content": "[15] Patricia S Koskinen, Robert M Wilson, and Carl J Jensema. Using closed-captioned television in the teaching of reading to deaf students. American Annals of the Deaf, 131(1):43-46, 1986. 2, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 314, + 578, + 517, + 626 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 578, + 517, + 626 + ], + "spans": [ + { + "bbox": [ + 314, + 578, + 517, + 626 + ], + "type": "text", + "content": "[16] Haoxin Li, Yijun Cai, and Wei-Shi Zheng. Deep dual relation modeling for egocentric interaction recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7932-7941, 2019. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 314, + 627, + 517, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 627, + 517, + 666 + ], + "spans": [ + { + "bbox": [ + 314, + 627, + 517, + 666 + ], + "type": "text", + "content": "[17] Kevin Qinghong Lin, Jinpeng Wang, Mattia Soldan, Michael Wray, Rui Yan, Eric Z Xu, Difei Gao, Rong-Cheng Tu, Wenzhe Zhao, Weijie Kong, et al. Ego-centric video-language pretraining. Advances in Neu" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 724, + 317, + 734 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 724, + 317, + 734 + ], + "spans": [ + { + "bbox": [ + 293, + 724, + 317, + 734 + ], + "type": "text", + "content": "28950" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 94, + 121, + 298, + 666 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 112, + 121, + 296, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 112, + 121, + 296, + 139 + ], + "spans": [ + { + "bbox": [ + 112, + 121, + 296, + 139 + ], + "type": "text", + "content": "ral Information Processing Systems, 35:7575-7586, 2022. 5, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 95, + 140, + 296, + 178 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 95, + 140, + 296, + 178 + ], + "spans": [ + { + "bbox": [ + 95, + 140, + 296, + 178 + ], + "type": "text", + "content": "[18] Yan-Bo Lin, Jie Lei, Mohit Bansal, and Gedas Bertasius. Eclipse: Efficient long-range video retrieval using sight and sound. In European Conference on Computer Vision, pages 413–430. Springer, 2022. 2, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 94, + 179, + 296, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 179, + 296, + 236 + ], + "spans": [ + { + "bbox": [ + 94, + 179, + 296, + 236 + ], + "type": "text", + "content": "[19] Shaowei Liu, Subarna Tripathi, Somdeb Majumdar, and Xiaolong Wang. Joint hand motion and interaction hotspots prediction from egocentric videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3282-3292, 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 94, + 238, + 298, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 238, + 298, + 285 + ], + "spans": [ + { + "bbox": [ + 94, + 238, + 298, + 285 + ], + "type": "text", + "content": "[20] Sagnik Majumder, Ziad Al-Halah, and Kristen Grauman. Learning spatial features from audio-visual correspondence in egocentric videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 27058-27068, 2024. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 94, + 286, + 296, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 286, + 296, + 334 + ], + "spans": [ + { + "bbox": [ + 94, + 286, + 296, + 334 + ], + "type": "text", + "content": "[21] Tushar Nagarajan, Yanghao Li, Christoph Feichtenhofer, and Kristen Grauman. Ego-topo: Environment affordances from egocentric video. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 163-172, 2020. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 94, + 335, + 296, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 335, + 296, + 402 + ], + "spans": [ + { + "bbox": [ + 94, + 335, + 296, + 402 + ], + "type": "text", + "content": "[22] Chiara Plizzari, Mirco Planamente, Gabriele Goletto, Marco Cannici, Emanuele Gusso, Matteo Matteucci, and Barbara Caputo. E2 (go) motion: Motion augmented event stream for egocentric action recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 19935-19947, 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 94, + 403, + 296, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 403, + 296, + 460 + ], + "spans": [ + { + "bbox": [ + 94, + 403, + 296, + 460 + ], + "type": "text", + "content": "[23] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 94, + 461, + 296, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 461, + 296, + 499 + ], + "spans": [ + { + "bbox": [ + 94, + 461, + 296, + 499 + ], + "type": "text", + "content": "[24] Hiroaki Sakoe and Seibi Chiba. Dynamic programming algorithm optimization for spoken word recognition. IEEE transactions on acoustics, speech, and signal processing, 26(1):43-49, 1978. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 94, + 500, + 296, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 500, + 296, + 548 + ], + "spans": [ + { + "bbox": [ + 94, + 500, + 296, + 548 + ], + "type": "text", + "content": "[25] Muhammad Bilal Shaikh, Douglas Chai, Syed Mohammed Shamsul Islam, and Naveed Akhtar. Multimodal fusion for audio-image and video action recognition. Neural Computing and Applications, 36(10): 5499-5513, 2024. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 94, + 548, + 296, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 548, + 296, + 597 + ], + "spans": [ + { + "bbox": [ + 94, + 548, + 296, + 597 + ], + "type": "text", + "content": "[26] Dandan Shan, Jiaqi Geng, Michelle Shu, and David F Fouhey. Understanding human hands in contact at internet scale. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9869-9878, 2020. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 94, + 598, + 296, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 598, + 296, + 645 + ], + "spans": [ + { + "bbox": [ + 94, + 598, + 296, + 645 + ], + "type": "text", + "content": "[27] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 94, + 647, + 296, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 647, + 296, + 666 + ], + "spans": [ + { + "bbox": [ + 94, + 647, + 296, + 666 + ], + "type": "text", + "content": "[28] Huiyu Wang, Mitesh Kumar Singh, and Lorenzo Torresani. Ego-only: Egocentric action detection with" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 314, + 121, + 517, + 611 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 332, + 121, + 517, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 332, + 121, + 517, + 149 + ], + "spans": [ + { + "bbox": [ + 332, + 121, + 517, + 149 + ], + "type": "text", + "content": "out exocentric transferring. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5250-5261, 2023. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 315, + 151, + 517, + 208 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 151, + 517, + 208 + ], + "spans": [ + { + "bbox": [ + 315, + 151, + 517, + 208 + ], + "type": "text", + "content": "[29] Jian Wang, Diogo Luvizon, Weipeng Xu, Lingjie Liu, Kripasindhu Sarkar, and Christian Theobalt. Scene-aware egocentric 3d human pose estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13031-13040, 2023. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 314, + 209, + 516, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 209, + 516, + 267 + ], + "spans": [ + { + "bbox": [ + 314, + 209, + 516, + 267 + ], + "type": "text", + "content": "[30] Limin Wang, Bingkun Huang, Zhiyu Zhao, Zhan Tong, Yinan He, Yi Wang, Yali Wang, and Yu Qiao. Videomae v2: Scaling video masked autoencoders with dual masking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14549-14560, 2023. 5, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 314, + 268, + 516, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 268, + 516, + 335 + ], + "spans": [ + { + "bbox": [ + 314, + 268, + 516, + 335 + ], + "type": "text", + "content": "[31] Wenhai Wang, Jifeng Dai, Zhe Chen, Zhenhang Huang, Zhiqi Li, Xizhou Zhu, Xiaowei Hu, Tong Lu, Lewei Lu, Hongsheng Li, et al. Internimage: Exploring large-scale vision foundation models with deformable convolutions. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14408-14419, 2023. 5, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 314, + 337, + 516, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 337, + 516, + 404 + ], + "spans": [ + { + "bbox": [ + 314, + 337, + 516, + 404 + ], + "type": "text", + "content": "[32] Yi Wang, Yinan He, Yizhuo Li, Kunchang Li, Jiashuo Yu, Xin Ma, Xinhao Li, Guo Chen, Xinyuan Chen, Yaohui Wang, Ping Luo, Ziwei Liu, Yali Wang, Limin Wang, and Yu Qiao. Internvid: A large-scale videotext dataset for multimodal understanding and generation. In The Twelfth International Conference on Learning Representations, 2024. 5, 6, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 314, + 405, + 516, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 405, + 516, + 453 + ], + "spans": [ + { + "bbox": [ + 314, + 405, + 516, + 453 + ], + "type": "text", + "content": "[33] Jianjia Xin, Lichun Wang, Kai Xu, Chao Yang, and Baocai Yin. Learning interaction regions and motion trajectories simultaneously from egocentric demonstration videos. IEEE Robotics and Automation Letters, 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 314, + 454, + 516, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 454, + 516, + 512 + ], + "spans": [ + { + "bbox": [ + 314, + 454, + 516, + 512 + ], + "type": "text", + "content": "[34] Jilan Xu, Yifei Huang, Junlin Hou, Guo Chen, Yuejie Zhang, Rui Feng, and Weidi Xie. Retrieval-augmented egocentric video captioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13525-13536, 2024. 1, 2, 5, 6, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 314, + 514, + 516, + 560 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 514, + 516, + 560 + ], + "spans": [ + { + "bbox": [ + 314, + 514, + 516, + 560 + ], + "type": "text", + "content": "[35] Zihui Sherry Xue and Kristen Grauman. Learning fine-grained view-invariant representations from unpaired ego-exo videos via temporal alignment. Advances in Neural Information Processing Systems, 36: 53688-53710, 2023. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 314, + 562, + 516, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 562, + 516, + 611 + ], + "spans": [ + { + "bbox": [ + 314, + 562, + 516, + 611 + ], + "type": "text", + "content": "[36] Yue Zhao, Ishan Misra, Philipp Krahenbuhl, and Rohit Girdhar. Learning video representations from large language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6586-6597, 2023. 5, 6, 7" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 724, + 317, + 734 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 724, + 317, + 734 + ], + "spans": [ + { + "bbox": [ + 293, + 724, + 317, + 734 + ], + "type": "text", + "content": "28951" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/SoundVista_ Novel-View Ambient Sound Synthesis via Visual-Acoustic Binding/9c309271-ab78-4453-a939-59e8aa02fe41_content_list.json b/2025/SoundVista_ Novel-View Ambient Sound Synthesis via Visual-Acoustic Binding/9c309271-ab78-4453-a939-59e8aa02fe41_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..a20a40a956bd67a180f30b96c3220c603b58e53e --- /dev/null +++ b/2025/SoundVista_ Novel-View Ambient Sound Synthesis via Visual-Acoustic Binding/9c309271-ab78-4453-a939-59e8aa02fe41_content_list.json @@ -0,0 +1,1469 @@ +[ + { + "type": "text", + "text": "SoundVista: Novel-View Ambient Sound Synthesis via Visual-Acoustic Binding", + "text_level": 1, + "bbox": [ + 99, + 130, + 897, + 154 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Mingfei Chen $^{1,2*}$ Israel D. Gebru $^{2}$ Ishwarya Ananthabhotla $^{3}$ Christian Richardt $^{2}$ \nDejan Markovic $^{2}$ Jake Sandakly $^{2}$ Steven Krenn $^{2}$ Todd Keebler $^{2}$ Eli Shlizerman $^{1}$ Alexander Richard $^{2}$", + "bbox": [ + 125, + 178, + 866, + 243 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1University of Washington 2Codec Avatars Lab, Pittsburgh, Meta 3Reality Labs Research, Meta", + "bbox": [ + 102, + 250, + 888, + 271 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8b2ef0b4b30c3b3170c3f44021a83871628edb3d04c44413ad1bccf479f2500f.jpg", + "image_caption": [ + "Figure 1. SoundVista: a novel method that synthesizes binaural ambient sound for arbitrary scenes from novel viewpoints. Our method leverages pre-acquired audio recordings and visual data captured from sparsely distributed reference points and synthesizes binaural audio consistent with the target 3D position and pose." + ], + "image_footnote": [], + "bbox": [ + 119, + 304, + 877, + 474 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 541, + 326, + 556 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We introduce SoundVista, a method to generate the ambient sound of an arbitrary scene at novel viewpoints. Given a pre-acquired recording of the scene from sparsely distributed microphones, SoundVista can synthesize the sound of that scene from an unseen target viewpoint. The method learns the underlying acoustic transfer function that relates the signals acquired at the distributed microphones to the signal at the target viewpoint, using a limited number of known recordings. Unlike existing works, our method does not require constraints or prior knowledge of sound source details. Moreover, our method efficiently adapts to diverse room layouts, reference microphone configurations and unseen environments. To enable this, we introduce a visual-acoustic binding module that learns visual embeddings linked with local acoustic properties from panoramic RGB and depth data. We first leverage these embeddings to optimize the placement of reference microphones in any given scene. During synthesis, we leverage multiple embeddings extracted from reference locations to get adaptive weights for their contribution, conditioned on target viewpoint. We benchmark the task", + "bbox": [ + 88, + 573, + 485, + 876 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "on both publicly available data and real-world settings. We demonstrate significant improvements over existing methods.", + "bbox": [ + 511, + 542, + 906, + 573 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 585, + 645, + 599 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent advances in 3D reconstruction and Novel-View Synthesis (NVS) have significantly enhanced our ability to create photorealistic visual models of real-world scenes [17, 34, 48, 49]. These developments have paved the way for applications in 3D virtual tours, experience recreation, and spatial media. However, the audio counterpart – Novel-View Acoustic Synthesis (NVAS) [1, 6] – has been under-explored compared to its visual counterpart and has not received the rigorous attention required to generate acoustic virtual scenes that match their visual surroundings.", + "bbox": [ + 509, + 611, + 908, + 762 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To address this gap, we introduce SoundVista: a novel method that creates a truly immersive acoustic experience. SoundVista can generate realistic and spatially accurate binaural audio at novel viewpoints in arbitrary scenes given a sparse set of sample recordings from different viewpoint.", + "bbox": [ + 511, + 763, + 908, + 839 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Unlike novel-view synthesis for visual 3D scenes, which are mostly static, NVAS faces significant challenges due to the dynamic nature of real-world acoustic environments. Ambient sound – the overall acoustic state describing all sounds", + "bbox": [ + 511, + 839, + 908, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Work done during an internship at Meta.", + "bbox": [ + 112, + 886, + 339, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "8331", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "in a scene [55] – can encompass multiple time-varying, nonstationary sound sources without limitations on the type and number of sound sources. In an ideal scenario, with complete information about the ambient sound (i.e. clean signals of individual sound sources and their locations over time), and a fully reconstructed 3D visual model, one could synthesize the ambient sound at any location using standard acoustic renderers with room impulse responses (RIRs) [1, 4, 9, 19, 43]. In real-world scenes, however, we lack detailed information about the sound sources that comprise the ambient sound. Accurately determining their clean source content and emitter positions is a challenging problem.", + "bbox": [ + 89, + 90, + 485, + 272 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "NVAS also faces challenges in balancing data sampling costs and synthesis quality. Current methods in the acoustic community use grids of microphones, and techniques ranging from interpolation [28, 29] to complex signal processing [2, 50] to synthesize audio at novel locations. However, these approaches do not scale well to large spaces, and efficient sampling and data utilization remain open questions.", + "bbox": [ + 89, + 273, + 483, + 381 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recent deep learning based NVAS methods [5, 6, 8, 21, 25] utilize multimodal data, including visual inputs, to transfer sparse reference sounds to binaural sound at target viewpoints. These methods often simplify the task by focusing on re-synthesizing primary sounds such as speech and music, with a limited number of sound sources (typically 1-2). They often ignore other sounds that contribute to the natural characteristics of the ambient sound scene. While feasible with few references, they struggle to adapt to diverse, large scenes with complex layout and acoustic environment. Furthermore, the lack of an optimal reference location sampling strategy hinders adaptation, leading to a loss of acoustic details and making it difficult to synthesize high-quality sound that has accurate binaural effects for target viewpoints.", + "bbox": [ + 89, + 382, + 485, + 594 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we propose a novel approach that avoids the challenge of obtaining granular information about individual sound sources and addresses some of the shortcomings of existing NVAS techniques. Our method relies on sparsely distributed \"reference microphones\" to capture acoustic snapshots, taking their recorded audio signals as \"reference recordings\". This allows us to get a holistic representation of the acoustic environment from references recordings. Given many examples of reference recordings, and sounds recorded at known locations within the scene, we formulate NVAS as learning the transformation from reference recordings to sound recorded at target viewpoints. At inference time, given any pre-acquired reference recordings of the scene as input and a query for an arbitrary target position, we expect the model to output binaural audio that is acoustically consistent with the query viewpoint and the content from reference recordings.", + "bbox": [ + 89, + 595, + 485, + 852 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address the limitations of existing NVAS techniques in reference microphone placement, we develop a multi-modal approach for optimal reference location sampling. Our ap-", + "bbox": [ + 89, + 854, + 485, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "proach can adapt to diverse, large complex scenes using a Visual-Acoustic Binding (VAB) module. VAB learns acoustically relevant features (VAB embeddings) by pretraining an encoder to align visual features from panoramic RGB-D captures with acoustic features from echo responses. Using VAB embeddings from numerous candidate reference locations, we employ a sampler that automatically identifies representative spots by finding spatial regions with similar embeddings. These spatial regions represent areas with similar acoustic properties and are usually free from obstacles that significantly hinder sound propagation. This approach optimizes reference microphone placement within a limited reference budget which enhances the overall performance.", + "bbox": [ + 511, + 90, + 906, + 287 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Furthermore, to manage varying numbers of references – which typically depend on scene size and budget – and their unequal contributions to the final audio, we introduce an adaptive reference integration module. This module models sequences via a transformer, using VAB features to reweight reference inputs and viewpoints. These reweighted inputs then serve as conditions for the final binaural audio renderer. The resulting conditions are both scene-adapted and content-invariant, enabling effective handling of various reference microphone configurations and audio content.", + "bbox": [ + 511, + 287, + 908, + 436 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To demonstrate the effectiveness of our approach, we benchmark the proposed model in both real-world setting and in a challenging simulated dataset derived from Matterport3D [3] scenes with SoundSpaces [4]. SoundVista outperforms existing methods [1, 5, 6, 8, 21] on scenes with multiple varieties of sound sources, handling varying numbers up to ten sound sources.", + "bbox": [ + 511, + 438, + 908, + 544 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 554, + 656, + 569 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Acoustic Scene Synthesis. Traditional methods focus on estimating the room impulse response (RIR) to recreate spatial audio. This is achieved by convolving the sound from each emitter with the RIR corresponding to the emitter-listener pair and summing the results [9, 20, 23, 25, 35-37, 39, 43]. These RIR-based techniques often require detailed source information, such as a clean signal of each source and its precise location, which are typically unknown in real-world scenarios, making them challenging to implement. Alternative approaches use scene-descriptor images for direct spatial sound synthesis based on reference sound input, such as visual acoustic matching [5, 7] or visual-guided audio spatialization [12, 13, 51, 57]. However, these methods may be inaccurate when viewpoints change continuously.", + "bbox": [ + 511, + 580, + 908, + 791 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Novel-view Acoustic Synthesis. Novel-view acoustic synthesis techniques address limitations in acoustic scene synthesis by learning transfer functions from reference sounds to target viewpoints [6, 8, 21, 38, 46]. Methods such as Few-shotRIR [25] and BEE [8] use multiple references but lack optimal reference sampling strategies, relying on heuristic settings such as a single close reference [6, 21, 38, 46],", + "bbox": [ + 511, + 795, + 910, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "8332", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/616cb5b354904c43c02e740986cc445de139992565e0cdfa10327101b660d2f5.jpg", + "image_caption": [ + "Figure 2. Details of the SoundVista Pipeline: (a) The reference location sampler selects optimal reference locations leveraging embeddings from visual-acoustic binding (VAB). (b) The reference integration transformer uses VAB embeddings to derive contribution weights for each reference. (c) Reweighting by contribution weights adjusts and integrates reference recording channels and pose conditioning for precise sound synthesis. (d) The spatial audio renderer converts reweighted channels and conditions to binaural sound at the target viewpoint." + ], + "image_footnote": [], + "bbox": [ + 101, + 90, + 885, + 343 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "random sampling [25], or fixed references [8], which limit adaptation to diverse scene layouts. In contrast, our method optimizes reference location sampling via visual-acoustic binding and derives adaptive reference contribution weights, enhancing sound synthesis accuracy and generalization.", + "bbox": [ + 88, + 415, + 483, + 492 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Acoustic-Visual Learning. Techniques such as CLIP [11, 22, 24, 31, 52] have successfully aligned visual and linguistic inputs to learn matched deep semantic representations. These advances led to recent works to leverage complementary nature of audio and visual data for acoustic related tasks. For example, AV-RIR [37] binds RIR with panoramic images to improve the late reverberation modeling of RIR. Advances in multi-modal visual understanding have facilitated sound source localization [15, 18, 27, 44], audio-visual speech enhancement and source separation [10, 26, 30, 45, 54, 56], as well as acoustic scene reconstruction [5, 6, 8, 25, 37, 41]. These works underscore the strong link between visual and acoustic modalities. Inspired by these approaches, we developed a visual encoder that aligns features from RGB-D visual data with acoustic features from echo responses. The alignment enables us to infer acoustic properties using visual data only. As a result, our method does not require knowing acoustic parameters or RIRs during inference. It is also easily adaptable to various scene and acoustic environments.", + "bbox": [ + 91, + 494, + 483, + 782 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 792, + 181, + 808 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Problem Formulation", + "text_level": 1, + "bbox": [ + 89, + 816, + 294, + 832 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given $N$ pairs of co-located microphones and cameras, we aim to strategically place them in the acoustic scene. Let $\\mathbf{L}_i = (x_i,y_i,z_i)\\in \\mathbb{R}^3$ denote the location of the devices and $\\theta_{i}\\in \\mathbb{R}^{2}$ denote their orientations. The microphones cap", + "bbox": [ + 89, + 839, + 485, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ture reference audio recordings $\\mathbf{A}_i$ , and the cameras capture panoramic RGB-D views $\\mathbf{V}_i$ . These \"reference captures\" encapsulate the ambient sound and visual properties, providing the basis for reconstructing binaural ambient sound $\\tilde{\\mathbf{A}}_k$ for a target listener at arbitrary location $\\mathbf{L}_k = (x_k, y_k, z_k) \\in \\mathbb{R}^3$ with orientation $\\theta_k \\in \\mathbb{R}^2$ . The reference recording can be of any type, but we use ambisonic microphones as they are ideal for capturing sound fields from multiple directions and simplify binaural signal extraction. In contrast to existing methods [1], we avoid enumerating sound sources. Instead, we learn a transfer function $\\mathcal{F}$ from reference captures to the target as:", + "bbox": [ + 511, + 416, + 906, + 597 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {F}: \\left(\\left\\{\\mathbf {A} _ {i}, g _ {i}, r _ {k i}, \\boldsymbol {\\theta} _ {i} \\right\\} _ {i = 1} ^ {N}, g _ {k}, \\boldsymbol {\\theta} _ {k}\\right) \\mapsto \\tilde {\\mathbf {A}} _ {k}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 552, + 601, + 906, + 627 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\ng _ {i} = \\phi (\\mathbf {V} _ {i}), g _ {k} = \\phi (\\mathbf {V} _ {k}), \\text {a n d} r _ {k i} = \\mathbf {L} _ {k} - \\mathbf {L} _ {i},\n$$\n", + "text_format": "latex", + "bbox": [ + 552, + 630, + 866, + 646 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $g_{i}$ and $g_{k}$ are Visual-Acoustic Binding (VAB) embeddings extracted with a pre-trained visual encoder $\\phi (\\cdot)$ , designed to align visual with acoustic representations. $r_{ki}$ is relative vector introduced to avoid overfitting on absolute locations and to handle diverse scene coordinate systems.", + "bbox": [ + 511, + 657, + 906, + 734 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We parameterize $\\mathcal{F}$ using neural networks that incorporate a transformer network and spatial audio renderer module. The transformer generates an adaptive mask to weight the contribution of each reference recording based on the target viewpoint and the VAB embeddings associated with reference viewpoints. We call this the \"Reference Integration Transformer\". The spatial audio renderer then processes the reweighted audio channels along with conditional information from the target view features to produce the final binaural sound. We provide an overview of these modules in Figure 2, with details explained in the following sections.", + "bbox": [ + 511, + 734, + 908, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "8333", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Visual-Acoustic Binding (VAB)", + "text_level": 1, + "bbox": [ + 89, + 90, + 364, + 107 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Obtaining acoustic properties of real-world scenes is challenging [9]. We propose to leverage RGB-D data, which can provide rich information linked to acoustic properties of a scene. For example, depth data provides information about obstacles and room geometry, while pixel-level textures reveal material differences and detailed obstacle information.", + "bbox": [ + 89, + 113, + 485, + 204 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The goal is to infer the acoustic properties of a scene using only the visual data. To achieve this, we first collected extensive paired panoramic RGB-D and acoustic echo responses data by navigating walkable areas in the SoundSpaces simulator [4]. We then trained a neural network, which we refer as \"Visual-Acoustic Binding (VAB) module\", to predict the acoustic representations from the visual features.", + "bbox": [ + 89, + 205, + 485, + 311 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Acoustic Representation. In acoustics, an impulse response is a function of time and the positions of the emitter and the listener [47]. It describes how sound propagates through a medium and interacts with the environment [32, 33]. To simplify data collection and infer local acoustic characteristic, we focus on echo impulse responses [25], where the emitter and target locations are the same. We extracted reverberation time (aka $\\mathrm{RT}_{60}$ parameter) from echo response to use as acoustic representation. $\\mathrm{RT}_{60}$ measures how long it takes for the acoustic energy to decay by $60\\mathrm{dB}$ and can reveal information about room geometry, obstacles and reflection. Figure 2(a) shows an example $\\mathrm{RT}_{60}$ map of a scene. Significant value changes in $\\mathrm{RT}_{60}$ map can indicate major obstacles or surface variations, highlighting key acoustic regions.", + "bbox": [ + 89, + 316, + 485, + 527 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Visual Representation. Following prior works [6, 37], we use panoramic RGB-D captures as inputs and extract visual features from ResNet-18 [14]. We train the VAB module to predict $\\mathrm{RT}_{60}$ . Therefore, effective binding visual representation with acoustic representation.", + "bbox": [ + 89, + 531, + 483, + 608 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Reference Location Sampler", + "text_level": 1, + "bbox": [ + 89, + 619, + 346, + 637 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "While having more microphones and cameras is ideal for this task, in practice resources are often limited. To maximize performance and make the best use of available resources, we propose a strategic approach for placement of reference microphones throughout a given scene.", + "bbox": [ + 89, + 642, + 483, + 718 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We argue that ideal microphone placements align with the acoustic partitions of the scene-areas with unique acoustic properties and free from obstacles such as walls. To identify these partitions, we capture panoramic RGB-D images from all candidate locations. One can use novel-view synthesis [49] to render these images; without needing actual photographs at each spot. We then extract VAB embeddings from each capture. VAB embeddings which correspond to scene acoustic parameters serve as strong cues for identifying acoustic partitions, such as distinguishing regions separated by obstacles. To enhance reliability, we combine the extracted VAB embeddings with positional information,", + "bbox": [ + 89, + 719, + 485, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "allowing them to complement each other. Using these embeddings, we perform data clustering of the candidate locations and take the center of each cluster as a reference location. This is illustrated in Figure 2(a).", + "bbox": [ + 511, + 90, + 908, + 152 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Reference Integration Transformer", + "text_level": 1, + "bbox": [ + 511, + 157, + 821, + 172 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We want our model to work effectively across diverse scenes. This requires the model to adapt to varying numbers of reference audio and visual inputs. Logically, larger scenes would benefit from more microphones and cameras, while smaller spaces can operate efficiently with fewer resources. Moreover, since the task (Section 3.1) is to transfer reference recordings to target sounds at specific viewpoints, distant microphones, or those in non-adjacent acoustic partitions usually contribute significantly less to the process than closer ones. However, weighting based solely on distance is insufficient due to potential obstacles such as walls and objects.", + "bbox": [ + 511, + 179, + 908, + 345 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To address these issues, we propose a transformer network. We treat each reference input as part of a sequence, allowing us to manage varying numbers of reference inputs. To derive the weights for their unequal contributions to the final audio, we exploit VAB embeddings extracted from visual captures at the reference and target viewpoints.", + "bbox": [ + 511, + 345, + 908, + 436 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We formed the queries input by combining the VAB embedding at the target location $g_{k}$ , with a learnable latent query embedding $\\mathbf{e}$ . The query is initialized from a normal distribution and optimized during training, to adjust the contribution of reference microphones based on their relative location vector to the target location. Let $g_{k}^{e} = [g_{k}||\\mathbf{e}] \\in \\mathbb{R}^{\\mathrm{C}}$ to denote the queries. The keys and values inputs are created by combining the VAB embedding of each reference with its relative vector $r_{ki}$ . Let $g_{i}^{r} = [g_{i}||r_{ki}] \\in \\mathbb{R}^{\\mathrm{C}}$ to denote the keys and values of $i^{\\text{th}}$ reference for the attention operation. The attention weight $a_{ki}$ can be calculated as follows:", + "bbox": [ + 511, + 436, + 908, + 602 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\na _ {k i} = \\frac {g _ {k} ^ {e} \\cdot g _ {i} ^ {r} \\top}{\\sqrt {\\mathsf {C}}}, \\quad i = \\{1, 2, \\dots , N \\}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 581, + 609, + 906, + 643 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The attention weight $a_{ki}$ indicates contribution of the reference $i$ to the target $k$ normalized by the Softmax function over $N$ references. Higher weights indicate greater influence.", + "bbox": [ + 511, + 650, + 908, + 696 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.5. Spatial Audio Renderer", + "text_level": 1, + "bbox": [ + 511, + 704, + 732, + 720 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The Spatial Audio Renderer takes the reference recordings and conditional information to generate binaural sound at the target viewpoint.", + "bbox": [ + 511, + 726, + 906, + 772 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Reweighting Reference Recordings. We apply the Short-Time Fourier Transform (STFT) to the reference recordings to extract their spectrograms. Then, we reweight the channels of each reference recording using the attention weights $a_{ki}$ from Section 3.4 to integrate them, encoding the result into a latent space as the input for the renderer.", + "bbox": [ + 511, + 775, + 908, + 867 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Reweighting Condition Inputs. The binaural sound is derived based on the target's position and pose. To capture the", + "bbox": [ + 511, + 869, + 908, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "8334", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "different aspects of the binaural effects, we decouple the conditioning into global and local components. This accounts for the cause of spatialization effects at a given target point, which may primarily vary from the distance to the sound source or remain invariant to head orientation. The global condition $c_{g}$ determines how the target viewpoint's relative position to reference locations influences the binaural sound. The local condition $c_{l}$ accounts more for the effects of head orientation. They are defined as follows:", + "bbox": [ + 89, + 90, + 485, + 227 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nc _ {g} = \\sum_ {i = 1} ^ {N} a _ {k i} \\cdot \\psi_ {1} \\left(g _ {i} \\| r _ {k i}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 158, + 233, + 483, + 273 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nc _ {l} = \\left(\\sum_ {i = 1} ^ {N} a _ {k i} \\cdot \\psi_ {2} \\left(r _ {k i} \\| q _ {k i}\\right)\\right) + \\sigma \\left(\\theta_ {k}\\right). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 158, + 277, + 483, + 319 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Here $q_{ki}$ is the relative orientation quaternion for target $k$ relative to reference $i$ ; $\\psi_1$ and $\\psi_2$ are MLP layers projecting concatenated inputs to a latent condition embedding; and $\\sigma(\\theta_k)$ denotes the target's rotational features from $\\theta_k$ .", + "bbox": [ + 89, + 324, + 483, + 385 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "**Renderer.** The Audio Renderer network is designed as a stacked U-Net [40, 42], depicted in Figure 2(d). It features down-sampling and up-sampling blocks that incorporate input conditions at each layer. To preserve quality and content details from the reference input, skip connections are employed at multiple resolutions. After processing the integrated reference recordings, we apply the inverse STFT to convert the output spectrogram into a binaural waveform.", + "bbox": [ + 89, + 388, + 483, + 508 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training from scratch in complex environments with challenging audio content can hinder the renderer to adapt to head orientation while producing high-quality audio. To address this, we pretrain the binauralization capability of our renderer by fixing the target location and varying head rotations and audio sources. After pre-training, the renderer effectively understands binauralization across different orientations, enhancing the accuracy of spatial effect modeling.", + "bbox": [ + 89, + 508, + 485, + 630 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.6. Loss", + "text_level": 1, + "bbox": [ + 89, + 637, + 163, + 651 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We design the loss function as a weighted combination of three components, each regulating different aspects of the model: 1) Waveform Loss measures the mean squared error between the predicted and target waveforms, ensuring precision of the predicted waveform amplitude and phases. 2) Binaural Interaural Level Difference (ILD) Loss [16] focuses on the energy difference between binaural audio channels to ensure accurate spatial effect prediction. 3) Multi-resolution Spectrogram Magnitude Loss [53] ensures that the predicted audio matches the target spectrogram magnitude across multiple resolutions. It includes two terms: the first compares log-scaled magnitudes, and the second, the Scaled Spectrogram Magnitude Loss, calculates magnitude distance over the target scale, addressing high variance in audio magnitude. We found that resolution with large FFT and small hop sizes can benefit modeling ambient noise, such as air vibrations.", + "bbox": [ + 89, + 659, + 485, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 513, + 90, + 645, + 107 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Benchmarks and Metrics", + "text_level": 1, + "bbox": [ + 513, + 116, + 743, + 131 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "N2S Benchmark. Captured in a real room of 8.5 by 6 meters dimensions with two internal rooms, the N2S benchmark utilizes microphone arrays with 32 microphones for 32-channel recordings. Visual captures are rendered by the NeRF-based 3D NVS model, VR-NeRF [49], after it has been trained on visual data collected by navigating the room. 11 static microphone arrays are uniformly distributed to provide reference ambient recordings, while 6 mobile arrays cover 557 locations with 15 seconds of recording per location. Motion capture system, OptiTrack, records 6DoF tracking data for both reference and target microphones. $35\\%$ of these locations with 20 orientations are selected for training.", + "bbox": [ + 511, + 142, + 906, + 325 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Soundspace-Ambient Matterport3D. We build this benchmark based on Matterport3D scenes, comprising 39 complex training scenes and 23 unseen scenes. It is challenging as each scene can include 2 to 10 sound sources, selected from 128 different sounds, from fan noise to speech. Reference locations are spaced every 5 meters on average, with simulated second-order ambisonics RIRs to render reference sounds by Soundspaces [4]. $85\\%$ of walkable locations are used for training with azimuth angles of 0, 90, 180, and 270.", + "bbox": [ + 511, + 329, + 908, + 465 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Metrics. We evaluate sound synthesis performance using four key metrics (the lower, the better): 1) L1 distance of STFT spectrogram (STFT) [13] for left and right channels; 2) Magnitude Spectrogram Distance (MAG) [13], measuring the closeness of reconstructed audio to the groundtruth; 3) Energy Envelope Error (ENV) [13], assessing the Euclidean distance between energy envelopes of groundtruth and predicted audio channels; 4) Left-Right Energy Ratio Error (LRE) [6], evaluating binaural effect accuracy by calculating energy ratio difference between left and right channels.", + "bbox": [ + 511, + 470, + 908, + 623 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Comparison with Baselines", + "text_level": 1, + "bbox": [ + 511, + 635, + 759, + 651 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We compare our method with the following baseline approaches: (1) Nearest GT: Binaural sound from the nearest reference microphone aligned with the target orientation. (2) Interp GT: Employs binaural sound at the target orientation from the four nearest reference microphones, interpolating based on distance to the target location. (3) AV-NeRF [21]: A NeRF-based system that synthesizes binaural audio from a given camera pose and RGB-D renderings. We adapted it to use recordings, poses, and visual context from references while preserving its core model components, enabling fair comparisons in dynamic scenes. (4) Few-shotRIR [25]: A transformer-based method that infers RIRs from a sparse set of observed images and echoes. Adapted to replace the reference and target impulse responses with the ambient sounds at the corresponding viewpoints. (5) VAM [5]: Matches the acoustics of input audio with a target image. (6) BEE [8]:", + "bbox": [ + 511, + 657, + 908, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "8335", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/a2a4394d4e3363eb3b3d7b115b7bdefa4b913f7a9b2b77ee6c3b79174a5a7f54.jpg", + "table_caption": [ + "Table 1. Results Comparison on Soundspace-Ambient Benchmark: Average metrics for 10,189 samples in 39 seen scenes with novel target locations and sources, and 6,534 samples in 23 unseen scenes. Sampling strategies: location only, vis+location (our sampler), and w/o VAB (ours without VAB pretraining). SoundVista Ref Num $k$ : only references with top- $k$ contribution weights are used for fair comparison." + ], + "table_footnote": [], + "table_body": "
MethodSamplingRef NumSeen ScenesUnseen Scenes
STFT ↓MAG ↓ENV ↓LRE ↓STFT ↓MAG ↓ENV ↓LRE ↓
nearest GTlocation14.4480.3510.1541.5964.0340.3530.1551.617
nearest GTw/o VAB14.4140.3410.1511.5764.6800.3470.1531.537
nearest GTvis+location13.9160.3360.1461.5723.8350.3440.1511.557
interp GTlocation43.9220.3260.1441.5843.4100.3270.1451.587
interp GTw/o VAB43.6600.3190.1421.5703.7660.3200.1421.531
interp GTvis+location43.1790.3130.1371.5593.4150.3210.1411.531
AV-NeRF [21]vis+location19.4240.4260.1951.9229.3210.4280.1961.979
VAM [5]vis+location15.2240.4200.1781.9024.9360.4360.1821.977
ViGAS [6]vis+location13.7400.3610.1542.0403.4380.3710.1572.051
SoundVista (Ours)vis+location12.5260.2910.1321.4082.6760.3090.1401.386
BEE [8]vis+location44.0980.3650.1622.0835.6350.3960.1782.131
SoundVista (Ours)vis+location42.4440.2890.1301.3902.5170.3050.1371.371
Few-shotRIR [25]vis+locationall5.9370.4590.2131.8925.4570.4710.2151.960
SoundVista w/o vislocationall3.2280.3060.1411.4252.8900.3120.1421.439
SoundVista (Ours)vis+locationall2.4420.2890.1301.3902.5140.3050.1371.372
", + "bbox": [ + 94, + 132, + 898, + 371 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/5f423369912946fdb394c5f6a262dac1a2f4662332065dc7ab65a2840dbf7b73.jpg", + "table_caption": [ + "Table 2. Testing Results Comparison on N2S Benchmark. In this real-world scene, SoundVista with visual modality largely boosts the performance accuracy, especially on the binaural effect (LRE)." + ], + "table_footnote": [], + "table_body": "
MethodSTFT ↓MAG ↓ENV ↓LRE ↓
Nearest DSP2.4200.2120.1361.447
Interp DSP1.6590.2030.1421.383
AV-NeRF [21]2.1940.1870.1190.840
Few-shotRIR [25]1.7650.1990.1340.909
VAM [5]1.9720.1900.1190.916
BEE [8]1.4710.2000.1410.995
ViGAS [6]1.2010.1850.1190.873
SoundVista w/o vis1.2420.1850.1180.894
SoundVista (Ours)1.0730.1770.1130.776
", + "bbox": [ + 93, + 426, + 480, + 568 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "A generalizable rendering pipeline that reconstructs the binaural audio at an arbitrary listener location using inputs from sparse reference audio-visual samples in the scene. (7) ViGAS [6]: Transforms sound to the target viewpoint given the observed audio and visual captures at the reference viewpoint. Specifically, for AV-NeRF, VAM, and ViGAS, which need a single reference, we use the nearest reference microphone. BEE is adapted to use the nearest 4 microphones. Deep-learning baselines requiring visual inputs utilize panoramic RGB-D images and the same visual encoder as our model for consistency.", + "bbox": [ + 88, + 580, + 482, + 747 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our results in Table 1 and Table 2 demonstrate that our method, SoundVista, consistently surpasses baselines across diverse novel scenes and real scenarios. Despite the challenge of obtaining binaural groundtruth (GT) for arbitrary target orientations, SoundVista significantly reduces errors across all metrics compared with nearest GT and interp GT. On the challenging Soundspace-Ambient benchmark, with diverse sound sources and complex layouts, most deep-learning baselines underperform compared to non-learning methods due to the need for robust conditioning models.", + "bbox": [ + 88, + 750, + 482, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For a fair comparison, we evaluate SoundVista using the top 1 and 4 reference microphones by contribution weight, respectively. Compared to learning-based methods with 1 reference, SoundVista outperforms the best baseline ViGAS by $32.5\\%$ (STFT), $19.4\\%$ (MAG), $14.3\\%$ (ENV), and $31\\%$ (LRE), demonstrating the strength of our conditioned audio renderer. The nearest microphone may miss critical audio content, while using the top 4 references improves performance, matching that of all microphones since most references are not required due to their sparse distribution. Since contribution weights are independent of audio content, fewer than 4 references can be selected per target location by precomputing contribution weights once per scene.", + "bbox": [ + 511, + 386, + 906, + 582 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "While $Few-shotRIR$ and $BEE$ use audio content to integrate references, the diverse sound distribution makes this challenging. During testing, content often falls outside the training distribution, degrading performance. Thus, more references aren't always better; for example, $Few-shotRIR$ did not gain an increase in synthesis accuracy and it was even reduced compared to baselines using only one reference.", + "bbox": [ + 511, + 582, + 908, + 686 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The deep-learning baselines perform better on the N2S benchmark, which features a single static scene. Compared to these deep-learning baselines, our SoundVista still significantly outperforms on accuracy of energy (ENV) and binaural effect (LRE) by $5\\%$ and $7.6\\%$ , respectively.", + "bbox": [ + 511, + 688, + 908, + 763 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Qualitative Results Comparison", + "text_level": 1, + "bbox": [ + 511, + 772, + 795, + 789 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Figure 3 visualizes our results for qualitative comparison with several competitive sound synthesis baselines: ViGAS, BEE, and Few-shotRIR. The first row displays loudness maps, while the second row shows generated binaural waveforms with corresponding LRE error displayed in the third row. ViGAS relies on the nearest microphone, leading to discrete loudness maps in complex layouts and unreliable results", + "bbox": [ + 511, + 795, + 908, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "8336", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/75fa673d8d6d7d5b358f14cb724594368076c405ae0d9bd1071cfe2b986128b8.jpg", + "image_caption": [ + "Figure 3. Comparison of Qualitative Results: First row: Loudness Map (black triangle: sources; purple triangle: target viewpoint). Second row: Reconstructed binaural waveform at target viewpoint (red: prediction; green: GT). Third row: LRE (lower for better binaural effect)." + ], + "image_footnote": [], + "bbox": [ + 98, + 90, + 511, + 280 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/198ee6a5e5efbd9a81f3aa7240b2cd7592cf067f6ff7f8aab5dbdced42457704.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 514, + 90, + 900, + 280 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "when obstacles are present, as shown in the unseen results. Additionally, ViGAS struggles with accurate binaural effects shown from the LRE results. BEE produces discontinuous loudness maps, deviating from the ground truth. Both BEE and Few-shotRIR have difficulty generating high-fidelity waveforms close to the groundtruth. Their reliance on incorporating audio content into the renderer's conditioning can bias the renderer leading to confusion and reduced quality when encountering content outside the training distribution. This is particularly common in our complex task setting with diverse audio sources. In contrast, SoundVista, closely matches the GT in loudness maps and binaural waveforms for both seen and unseen scenes, excelling in high-quality novel-view ambient sound synthesis.", + "bbox": [ + 88, + 325, + 485, + 537 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 550, + 243, + 566 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We conducted ablation studies on the Soundspace-Ambient benchmark to verify the contribution of key components.", + "bbox": [ + 89, + 573, + 483, + 604 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Visual-Acoustic Binding (VAB). In the design of VAB, we optimized the integration of visual features that correlate with acoustic parameters. Table 3 shows the accuracy of $\\mathrm{RT}_{60}$ value predictions in novel scenes without finetuning (w/o finetune) and with few-shot finetuning at reference locations (w/finetune). We tested various modality inputs: location only, panoramic rgb, depth, and rgb+depth. Depth alone significantly enhances binding, reducing error by over $50\\%$ compared to use location only. With finetuning, rgb+depth achieves superior results. Given the ability to capture references in scenes, we select rgb+depth as our visual input.", + "bbox": [ + 89, + 608, + 483, + 775 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Reference Sampler via VAB. The Reference Sampler via Visual-Acoustic Binding (VAB) (vis+location) significantly enhances the accuracy of non-learning baselines such as nearest GT and interp GT, as shown in Table 1. This method improves predictions of sound magnitude (MAG, ENV) and spectrogram phases (STFT). Without pretrained VAB (w/o VAB), these benefits diminish, especially in seen scenes. Results in Table 1, 2, and 4 further highlight the critical role of", + "bbox": [ + 89, + 780, + 483, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/ce8b73efac2edd617cd034a9d39636ae9b6f563c917eba30fd51014f1cc9c1a7.jpg", + "table_caption": [ + "Table 3. $\\mathbf{RT}_{60}$ Prediction Results on Matterport3D. w/ finetune involves few-shot finetuning on unseen scenes. err: distance between predicted and GT $\\mathrm{RT}_{60}$ , error ratio: scaled err over GT." + ], + "table_footnote": [], + "table_body": "
Modalityw/o finetunew/ finetune
err ↓err ratio ↓err ↓err ratio ↓
location0.1700.4110.0670.199
rgb0.1170.2360.0440.126
depth0.0820.1730.0380.109
rgb+depth0.0870.1850.0360.103
", + "bbox": [ + 535, + 366, + 880, + 453 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/1969225b58fb26b43f1d2d510b309f88220fbf27e1648d1a0efb1431759f262f.jpg", + "table_caption": [ + "Table 4. Ablations on the SoundSpace-Ambient benchmark." + ], + "table_footnote": [], + "table_body": "
MethodSTFT ↓MAG ↓ENV ↓LRE ↓
full model2.4420.2890.1301.390
w/o vis3.2280.3060.1411.425
w/o reweight2.9080.2950.1341.391
w/o render_pretrain2.5820.3100.1351.877
w/o decoup_cond3.0170.3000.1381.433
full loss2.4420.2890.1301.390
w/o ILD_loss2.4620.2920.1311.420
w/o mag_scale2.4240.3010.1361.422
w/o wav_loss3.4220.2910.1431.369
", + "bbox": [ + 516, + 470, + 903, + 614 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "visual input, as its absence increases errors across all metrics, underscoring the importance of VAB with visual information for adapting to diverse scene layouts.", + "bbox": [ + 511, + 625, + 906, + 669 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Figure 4 (left) illustrates the STFT error curve relative to reference density. Our model is trained with a reference density of 1.0, and tested on novel Soundspace-Ambient scenes with varying densities. Compared to Ours - w/o vis and interp - w/o vis (interp GT with location-only sampler), Ours - w/ vis outperforms others, with the performance gap widening as density decreases, demonstrating robustness of ours equipped with VAB vs. reference density variance.", + "bbox": [ + 511, + 670, + 908, + 790 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Reference Integration. To evaluate the contribution of the adaptive reference integration, we conducted experiments with the w/o reweight variant, which uses only the nearest reference microphone along with its corresponding visual and pose information to calculate the condition for the audio renderer. As shown in Table 4, the results of w/o reweight compared to full SoundVista model reveal a noticeable degra", + "bbox": [ + 511, + 795, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "8337", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c703f44ed5855cd266012abd0f0191e9c6812998ff18178544cffdee41b39ab5.jpg", + "image_caption": [ + "Figure 4. STFT Error Curve for different reference densities (left) and training location data ratios (right)." + ], + "image_footnote": [], + "bbox": [ + 91, + 87, + 282, + 218 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a28dfea29a199b2eb5e4f80c31a1aebe037947c3d4de2d0fcd1b0bd769b06a1e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 282, + 88, + 482, + 217 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "dation in performance, particularly in the STFT error, which increases by $19\\%$ . This indicates that the adaptive Reference Integration Transformer and reweighting modules effectively integrate references to enhance sound synthesis accuracy, as the nearest microphone may miss critical content required for the target sound synthesis.", + "bbox": [ + 89, + 276, + 483, + 367 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Conditioned Spatial Audio Renderer. In the w/o decoup_cond variant, we replaced the decoupled global and local conditions from reweighting in Section 3.5 with only the global condition plus target rotational features. This absence of local conditions significantly increases STFT error by $23.5\\%$ , MAG error by $3.8\\%$ , and ENV error by $6.2\\%$ .", + "bbox": [ + 89, + 376, + 483, + 467 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As detailed in Section 3.5, we pretrain the renderer with binauralization capability to help the model learn the complex task of orientation-sensitive novel view sound synthesis. The w/o render_pretrain variant shows a $35\\%$ increase in LRE error, underscoring the importance of renderer pretraining for achieving a more accurate binaural effect.", + "bbox": [ + 89, + 473, + 483, + 565 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Loss. The ablations in Table 4 highlight the contributions of different loss components. The Waveform Loss (wav_loss) constrains the energy envelope and regulates phase learning. Removing it significantly reduces STFT and ENV accuracy, increasing the STFT error by $29\\%$ . The Scaled Spectrogram Magnitude Loss (mag_scale) calculates magnitude loss over the target magnitude scale, addressing high variance in audio content and mitigating scale differences. Removing mag_scale will increase MAG error largely.", + "bbox": [ + 89, + 573, + 483, + 710 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Robustness w.r.t. Training Data Ratio. To assess the impact of training location data ratio on performance, we trained two variants, Ours - w/o vis and Ours - w/ vis, using different proportions of training locations on the Soundspace-Ambient benchmark. Figure 4 (right) shows that STFT error increases as the data ratio decreases. When the data ratio is below $10\\%$ , Ours - w/o vis underperforms compared to the non-learning baseline interp - w/o vis, while Ours - w/ vis maintains superior performance. Additionally, Ours - w/ vis benefits more from increased training locations, with STFT error decreasing rapidly as the ratio nears $100\\%$ , widening the performance gap with Ours - w/o vis and interp - w/o vis.", + "bbox": [ + 89, + 719, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7e8414d5fee898c7db8d2922422fbe0a8e9439733164ceac4b62769d684d1a1a.jpg", + "image_caption": [ + "Figure 5. Visualization Analysis. First three columns: Clustering results for loc only, Vis - w/o VAB (ours without VAB pretraining), and Vis - w/ VAB (ours). Colors indicate cluster regions. VAB helps clustering to align better with room partitions. Last two columns: Attention weights visualization for references, with colored stars (size proportional to weights) and a purple triangle for the target." + ], + "image_footnote": [], + "bbox": [ + 521, + 90, + 893, + 253 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Visualization Analysis", + "text_level": 1, + "bbox": [ + 511, + 356, + 720, + 372 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Figure 5 visualizes the clustering results of Reference Sampler via VAB, comparing loc only, Vis - w/o VAB, and Vis - w/ VAB. Different colors represent cluster regions. The Reference Sampler Vis - w/ VAB accurately segments clusters aligning with actual room partitions, proving to be more reliable. In contrast, other methods may incorrectly cluster non-adjacent rooms together which cannot share the same acoustic environment, highlighted by red-circles.", + "bbox": [ + 511, + 378, + 906, + 498 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "After identifying reference locations, we visualize the attention weights from the Reference Integration Transformer in Figure 5 (column 4-5 from left). Colored stars denote sampled references, sized by contribution weight, with the target marked by a purple triangle. Unlike Distance-only weights, which may incorrectly prioritize closer references despite obstacles, our model, Ref Attn - w/ VAB, trained with the adaptive Reference Integration Transformer, effectively highlights reliable references in similar acoustic environments.", + "bbox": [ + 511, + 500, + 908, + 635 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 648, + 633, + 664 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We introduce SoundVista, a novel system designed to synthesize ambient sound from arbitrary scenes at novel viewpoints. Our approach introduces a visual-acoustic binding module to effectively learn visual embeddings linked with local acoustic properties. This enables our system to optimize the placement of reference microphones and derive adaptive weights for each microphone's contribution, conditioned on the target viewpoint and visual captures, thereby, facilitating the final conditioned sound synthesis. SoundVista adapts to diverse room layouts, microphone configurations, and unseen environments, rendering high-quality binaural ambient sound without requiring prior constraints or detailed knowledge of sound sources. Our results, validated on both publicly available data and real-world settings, demonstrate state-of-the-art sound synthesis accuracy and generalization.", + "bbox": [ + 511, + 674, + 908, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8338", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Byeongjoo Ahn, Karren Yang, Brian Hamilton, Jonathan Sheaffer, Anurag Ranjan, Miguel Sarabia, Oncel Tuzel, and Jen-Hao Rick Chang. Novel-view acoustic synthesis from 3d reconstructed rooms. arXiv preprint arXiv:2310.15130, 2023. 1, 2, 3", + "[2] Federico Borra, Israel Dejene Gebru, and Dejan Markovic. Soundfield reconstruction in reverberant environments using higher-order microphones and impulse response measurements. In ICASSP 2019-2019 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 281-285. IEEE, 2019. 2", + "[3] Angel Chang, Angela Dai, Thomas Funkhouser, Maciej Halber, Matthias Niessner, Manolis Savva, Shuran Song, Andy Zeng, and Yinda Zhang. Matterport3d: Learning from rgb-d data in indoor environments. International Conference on 3D Vision (3DV), 2017. 2", + "[4] Changan Chen, Unnat Jain, Carl Schissler, Sebastia Vicenc Amengual Gari, Ziad Al-Halah, Vamsi Krishna Ithapu, Philip Robinson, and Kristen Grauman. Soundspaces: Audiovisual navigation in 3d environments. In European Conference on Computer Vision, pages 17-36. Springer, 2020. 2, 4, 5", + "[5] Changan Chen, Ruohan Gao, Paul Calamia, and Kristen Grauman. Visual acoustic matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18858-18868, 2022. 2, 3, 5, 6", + "[6] Changan Chen, Alexander Richard, Roman Shapovalov, Vamsi Krishna Ithapu, Natalia Neverova, Kristen Grauman, and Andrea Vedaldi. Novel-view acoustic synthesis. arXiv preprint arXiv:2301.08730, 2023. 1, 2, 3, 4, 5, 6", + "[7] Changan Chen, Wei Sun, David Harwath, and Kristen Grauman. Learning audio-visual dereverberation. In ICASSP, 2023. 2", + "[8] Mingfei Chen, Kun Su, and Eli Shlizerman. Be everywhere—hear everything (bee): Audio scene reconstruction by sparse audio-visual samples. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7853–7862, 2023. 2, 3, 5, 6", + "[9] Ziyang Chen, Israel D Gebru, Christian Richardt, Anurag Kumar, William Laney, Andrew Owens, and Alexander Richard. Real acoustic fields: An audio-visual room acoustics dataset and benchmark. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21886-21896, 2024. 2, 4", + "[10] Ariel Ephrat, Inbar Mosseri, Oran Lang, Tali Dekel, Kevin Wilson, Avinatan Hassidim, William T Freeman, and Michael Rubinstein. Looking to listen at the cocktail party: A speaker-independent audio-visual model for speech separation. arXiv preprint arXiv:1804.03619, 2018. 3", + "[11] Han Fang, Pengfei Xiong, Luhui Xu, and Yu Chen. Clip2video: Mastering video-text retrieval via image clip. arXiv preprint arXiv:2106.11097, 2021. 3", + "[12] Ruohan Gao and Kristen Grauman. 2.5d visual sound. In CVPR, 2019. 2", + "[13] Ruohan Gao and Kristen Grauman. 2.5 d visual sound. In Pro" + ], + "bbox": [ + 93, + 114, + 485, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 324-333, 2019. 2, 5", + "[14] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 4", + "[15] Di Hu, Rui Qian, Minyue Jiang, Xiao Tan, Shilei Wen, Errui Ding, Weiyao Lin, and Dejing Dou. Discriminative sounding objects localization via self-supervised audiovisual matching. Advances in Neural Information Processing Systems, 33: 10077-10087, 2020. 3", + "[16] Christopher A. Ick, Gordon Wichern, Yoshiki Masuyama, François Germain, and Jonathan Le Roux. Spatially-aware losses for enhanced neural acoustic fields. In Audio Imagination: NeurIPS 2024 Workshop AI-Driven Speech, Music, and Sound Generation, 2024. 5", + "[17] Hyeonjoong Jang, Andreas Meuleman, Dahiyun Kang, Donggun Kim, Christian Richardt, and Min H Kim. Egocentric scene reconstruction from an omnidirectional video. ACM Transactions on Graphics (TOG), 41(4):1-12, 2022. 1", + "[18] Hao Jiang, Calvin Murdock, and Vamsi Krishna Ithapu. Ego-centric deep multi-channel audio-visual active speaker localization. arXiv preprint arXiv:2201.01928, 2022. 3", + "[19] Shoichi Koyama, Tomoya Nishida, Keisuke Kimura, Takumi Abe, Natsuki Ueno, and Jesper Brunnström. Meshrir: A dataset of room impulse responses on meshed grid points for evaluating sound field analysis and synthesis methods. In 2021 IEEE Workshop on Applications of Signal Processing to Audio and Acoustics (WASPAA), pages 1-5, 2021. 2", + "[20] Susan Liang, Chao Huang, Yapeng Tian, Anurag Kumar, and Chenliang Xu. Neural acoustic context field: Rendering realistic room impulse response with neural fields. ArXiv, 2023. 2", + "[21] Susan Liang, Chao Huang, Yapeng Tian, Anurag Kumar, and Chenliang Xu. Av-nerf: Learning neural fields for real-world audio-visual scene synthesis. In Conference on Neural Information Processing Systems (NeurIPS), 2023. 2, 5, 6", + "[22] Ziyi Lin, Shijie Geng, Renrui Zhang, Peng Gao, Gerard De Melo, Xiaogang Wang, Jifeng Dai, Yu Qiao, and Hongsheng Li. Frozen clip models are efficient video learners. In European Conference on Computer Vision, pages 388-404. Springer, 2022. 3", + "[23] Andrew Luo, Yilun Du, Michael J Tarr, Joshua B Tenenbaum, Antonio Torralba, and Chuang Gan. Learning neural acoustic fields. arXiv preprint arXiv., 2021. 2", + "[24] Huaishao Luo, Lei Ji, Ming Zhong, Yang Chen, Wen Lei, Nan Duan, and Tianrui Li. CLIP4Clip: An empirical study of clip for end to end video clip retrieval. arXiv preprint arXiv:2104.08860, 2021. 3", + "[25] Sagnik Majumder, Changan Chen, Ziad Al-Halah, and Kristen Grauman. Few-shot audio-visual learning of environment acoustics, 2022. 2, 3, 4, 5, 6", + "[26] Daniel Michelsanti, Zheng-Hua Tan, Shi-Xiong Zhang, Yong Xu, Meng Yu, Dong Yu, and Jesper Jensen. An overview of deep-learning-based audio-visual speech enhancement and separation. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 2021. 3" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "8339", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[27] Shentong Mo and Yapeng Tian. Audio-visual grouping network for sound localization from mixtures. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10565–10574, 2023. 3", + "[28] Bartlomiej Mróz, Marek Kabaciński, Tomasz Ciotucha, Andrzej Ruminski, and Tomasz Žernicki. Production of six-degrees-of-freedom (6dof) navigable audio using 30 ambisonic microphones. In 2021 Immersive and 3D Audio: from Architecture to Automotive (I3DA), pages 1-5. IEEE, 2021. 2", + "[29] Orhun Olgun, Ege Erdem, and Hüseyin Hachhabiboglu. Sound field interpolation via sparse plane wave decomposition for 6 dof immersive audio. In 2023 Immersive and 3D Audio: from Architecture to Automotive (I3DA), pages 1-10. IEEE, 2023. 2", + "[30] Andrew Owens and Alexei A Efros. Audio-visual scene analysis with self-supervised multisensory features. In Proceedings of the European Conference on Computer Vision (ECCV), pages 631–648, 2018. 3", + "[31] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 3", + "[32] Nikunj Raghuvanshi and John Snyder. Parametric wave field coding for precomputed sound propagation. ACM Transactions on Graphics (TOG), 33(4):1-11, 2014. 4", + "[33] Nikunj Raghuvanshi and John Snyder. Parametric directional coding for precomputed sound propagation. ACM Transactions on Graphics (TOG), 37(4):1-14, 2018. 4", + "[34] Santhosh K Ramakrishnan, Aaron Gokaslan, Erik Wijmans, Oleksandr Maksymets, Alex Clegg, John Turner, Eric Undersander, Wojciech Galuba, Andrew Westbury, Angel X Chang, et al. Habitat-matterport 3d dataset (hm3d): 1000 large-scale 3d environments for embodied ai. arXiv preprint arXiv:2109.08238, 2021. 1", + "[35] Anton Ratnarajah, Zhenyu Tang, and Dinesh Manocha. *Irgan: Room impulse response generator for far-field speech recognition.* arXiv preprint arXiv:2010.13219, 2020. 2", + "[36] Anton Ratnarajah, Shi-Xiong Zhang, Meng Yu, Zhenyu Tang, Dinesh Manocha, and Dong Yu. Fast-rir: Fast neural diffuse room impulse response generator. arXiv preprint arXiv:2110.04057, 2021.", + "[37] Anton Ratnarajah, Sreyan Ghosh, Sonal Kumar, Purva Chiniya, and Dinesh Manocha. Av-rir: Audio-visual room impulse response estimation. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 27154–27165, 2023. 2, 3, 4", + "[38] Alexander Richard, Dejan Markovic, Israel D. Gebru, Steven Krenn, Gladstone Alexander Butler, Fernando Torre, and Yaser Sheikh. Neural synthesis of binaural speech from mono audio. In International Conference on Learning Representations, 2021. 2", + "[39] Alexander Richard, Peter Dodds, and Vamsi Krishna Ithapu. Deep impulse responses: Estimating and parameterizing filters with deep networks. arXiv preprint arXiv:2202.03416, 2022. 2" + ], + "bbox": [ + 91, + 92, + 485, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[40] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention-MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, pages 234-241. Springer, 2015. 5", + "[41] Nikhil Singh, Jeff Mentch, Jerry Ng, Matthew Beveridge, and Iddo Drori. Image2reverb: Cross-modal reverb impulse response synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 286-295, 2021. 3", + "[42] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In International Conference on Learning Representations, 2021. 5", + "[43] Kun Su, Mingfei Chen, and Eli Shlizerman. Inras: Implicit neural representation for audio scenes. Advances in Neural Information Processing Systems, 35:8144-8158, 2022. 2", + "[44] Ruijie Tao, Zexu Pan, Rohan Kumar Das, Xinyuan Qian, Mike Zheng Shou, and Haizhou Li. Is someone speaking? exploring long-term temporal features for audio-visual active speaker detection. In Proceedings of the 29th ACM international conference on multimedia, pages 3927-3935, 2021. 3", + "[45] Yapeng Tian, Di Hu, and Chenliang Xu. Cyclic co-learning of sounding object visual grounding and sound separation. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2744-2753, 2021. 3", + "[46] Aäron van den Oord, Sander Dieleman, Heiga Zen, Karen Simonyan, Oriol Vinyals, Alexander Graves, Nal Kalchbrenner, Andrew Senior, and Koray Kavukcuoglu. Wavenet: A generative model for raw audio. In Arxiv, 2016. 2", + "[47] Tor Erik Vigran. Building acoustics. CRC Press, 2014. 4", + "[48] Xiuchao Wu, Jiamin Xu, Zihan Zhu, Hujun Bao, Qixing Huang, James Tompkin, and Weiwei Xu. Scalable neural indoor scene rendering. ACM transactions on graphics, 41 (4), 2022. 1", + "[49] Linning Xu, Vasu Agrawal, William Laney, Tony Garcia, Aayush Bansal, Changil Kim, Samuel Rota Bulò, Lorenzo Porzi, Peter Kontschieder, Aljaž Božić, Dahua Lin, Michael Zollhöfer, and Christian Richardt. VR-NeRF: High-fidelity virtualized walkable spaces. In SIGGRAPH Asia Conference Proceedings, 2023. 1, 4, 5", + "[50] Shaoheng Xu, Jihui Aimee Zhang, Thushara D. Abhayapala, Amy Bastine, Wei-Ting Lai, and Prasanga N. Samarasinghe. Sparse sound field representation using complex orthogonal matching pursuit. In ICASSP 2024 - 2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1336-1340, 2024. 2", + "[51] Xudong Xu, Hang Zhou, Ziwei Liu, Bo Dai, Xiaogang Wang, and Dahua Lin. Visually informed binaural audio generation without binaural audiios. In CVPR, 2021. 2", + "[52] Hongwei Xue, Yuchong Sun, Bei Liu, Jianlong Fu, Ruihua Song, Houqiang Li, and Jiebo Luo. Clip-vip: Adapting pretrained image-text model to video-language representation alignment. arXiv preprint arXiv:2209.06430, 2022. 3" + ], + "bbox": [ + 516, + 92, + 906, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "8340", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[53] Ryuichi Yamamoto, Eunwoo Song, and Jae-Min Kim. Parallel wavegan: A fast waveform generation model based on generative adversarial networks with multi-resolution spectrogram. In ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 6199-6203. IEEE, 2020. 5", + "[54] Yuxin Ye, Wenming Yang, and Yapeng Tian. Lavss: Location-guided audio-visual spatial audio separation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 5508–5519, 2024. 3", + "[55] Zechen Zhang, Nikunj Raghuvanshi, John Snyder, and Steve Marschner. Ambient sound propagation. ACM Transactions on Graphics (TOG), 37(6):1-10, 2018. 2", + "[56] Hang Zhou, Ziwei Liu, Xudong Xu, Ping Luo, and Xiaogang Wang. Vision-infused deep audio inpainting. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 283-292, 2019. 3", + "[57] Hang Zhou, Xudong Xu, Dahua Lin, Xiaogang Wang, and Ziwei Liu. Sep-stereo: Visually guided stereophonic audio generation by associating source separation. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XII 16, pages 52-69. Springer, 2020. 2" + ], + "bbox": [ + 91, + 92, + 485, + 415 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "8341", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/SoundVista_ Novel-View Ambient Sound Synthesis via Visual-Acoustic Binding/9c309271-ab78-4453-a939-59e8aa02fe41_model.json b/2025/SoundVista_ Novel-View Ambient Sound Synthesis via Visual-Acoustic Binding/9c309271-ab78-4453-a939-59e8aa02fe41_model.json new file mode 100644 index 0000000000000000000000000000000000000000..525d161b132e4b1d690d8e12b81a575e95326fa5 --- /dev/null +++ b/2025/SoundVista_ Novel-View Ambient Sound Synthesis via Visual-Acoustic Binding/9c309271-ab78-4453-a939-59e8aa02fe41_model.json @@ -0,0 +1,2092 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.101, + 0.131, + 0.898, + 0.155 + ], + "angle": 0, + "content": "SoundVista: Novel-View Ambient Sound Synthesis via Visual-Acoustic Binding" + }, + { + "type": "text", + "bbox": [ + 0.126, + 0.179, + 0.867, + 0.244 + ], + "angle": 0, + "content": "Mingfei Chen\\(^{1,2*}\\) Israel D. Gebru\\(^{2}\\) Ishwarya Ananthabhotla\\(^{3}\\) Christian Richardt\\(^{2}\\) \nDejan Markovic\\(^{2}\\) Jake Sandakly\\(^{2}\\) Steven Krenn\\(^{2}\\) Todd Keebler\\(^{2}\\) Eli Shlizerman\\(^{1}\\) Alexander Richard\\(^{2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.104, + 0.251, + 0.89, + 0.272 + ], + "angle": 0, + "content": "1University of Washington 2Codec Avatars Lab, Pittsburgh, Meta 3Reality Labs Research, Meta" + }, + { + "type": "image", + "bbox": [ + 0.12, + 0.305, + 0.879, + 0.476 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.487, + 0.908, + 0.53 + ], + "angle": 0, + "content": "Figure 1. SoundVista: a novel method that synthesizes binaural ambient sound for arbitrary scenes from novel viewpoints. Our method leverages pre-acquired audio recordings and visual data captured from sparsely distributed reference points and synthesizes binaural audio consistent with the target 3D position and pose." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.542, + 0.327, + 0.558 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.574, + 0.486, + 0.877 + ], + "angle": 0, + "content": "We introduce SoundVista, a method to generate the ambient sound of an arbitrary scene at novel viewpoints. Given a pre-acquired recording of the scene from sparsely distributed microphones, SoundVista can synthesize the sound of that scene from an unseen target viewpoint. The method learns the underlying acoustic transfer function that relates the signals acquired at the distributed microphones to the signal at the target viewpoint, using a limited number of known recordings. Unlike existing works, our method does not require constraints or prior knowledge of sound source details. Moreover, our method efficiently adapts to diverse room layouts, reference microphone configurations and unseen environments. To enable this, we introduce a visual-acoustic binding module that learns visual embeddings linked with local acoustic properties from panoramic RGB and depth data. We first leverage these embeddings to optimize the placement of reference microphones in any given scene. During synthesis, we leverage multiple embeddings extracted from reference locations to get adaptive weights for their contribution, conditioned on target viewpoint. We benchmark the task" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.543, + 0.908, + 0.574 + ], + "angle": 0, + "content": "on both publicly available data and real-world settings. We demonstrate significant improvements over existing methods." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.586, + 0.646, + 0.601 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.612, + 0.909, + 0.763 + ], + "angle": 0, + "content": "Recent advances in 3D reconstruction and Novel-View Synthesis (NVS) have significantly enhanced our ability to create photorealistic visual models of real-world scenes [17, 34, 48, 49]. These developments have paved the way for applications in 3D virtual tours, experience recreation, and spatial media. However, the audio counterpart – Novel-View Acoustic Synthesis (NVAS) [1, 6] – has been under-explored compared to its visual counterpart and has not received the rigorous attention required to generate acoustic virtual scenes that match their visual surroundings." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.764, + 0.909, + 0.84 + ], + "angle": 0, + "content": "To address this gap, we introduce SoundVista: a novel method that creates a truly immersive acoustic experience. SoundVista can generate realistic and spatially accurate binaural audio at novel viewpoints in arbitrary scenes given a sparse set of sample recordings from different viewpoint." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Unlike novel-view synthesis for visual 3D scenes, which are mostly static, NVAS faces significant challenges due to the dynamic nature of real-world acoustic environments. Ambient sound – the overall acoustic state describing all sounds" + }, + { + "type": "page_footnote", + "bbox": [ + 0.113, + 0.887, + 0.341, + 0.901 + ], + "angle": 0, + "content": "* Work done during an internship at Meta." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "8331" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.273 + ], + "angle": 0, + "content": "in a scene [55] – can encompass multiple time-varying, nonstationary sound sources without limitations on the type and number of sound sources. In an ideal scenario, with complete information about the ambient sound (i.e. clean signals of individual sound sources and their locations over time), and a fully reconstructed 3D visual model, one could synthesize the ambient sound at any location using standard acoustic renderers with room impulse responses (RIRs) [1, 4, 9, 19, 43]. In real-world scenes, however, we lack detailed information about the sound sources that comprise the ambient sound. Accurately determining their clean source content and emitter positions is a challenging problem." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.275, + 0.485, + 0.382 + ], + "angle": 0, + "content": "NVAS also faces challenges in balancing data sampling costs and synthesis quality. Current methods in the acoustic community use grids of microphones, and techniques ranging from interpolation [28, 29] to complex signal processing [2, 50] to synthesize audio at novel locations. However, these approaches do not scale well to large spaces, and efficient sampling and data utilization remain open questions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.383, + 0.486, + 0.595 + ], + "angle": 0, + "content": "Recent deep learning based NVAS methods [5, 6, 8, 21, 25] utilize multimodal data, including visual inputs, to transfer sparse reference sounds to binaural sound at target viewpoints. These methods often simplify the task by focusing on re-synthesizing primary sounds such as speech and music, with a limited number of sound sources (typically 1-2). They often ignore other sounds that contribute to the natural characteristics of the ambient sound scene. While feasible with few references, they struggle to adapt to diverse, large scenes with complex layout and acoustic environment. Furthermore, the lack of an optimal reference location sampling strategy hinders adaptation, leading to a loss of acoustic details and making it difficult to synthesize high-quality sound that has accurate binaural effects for target viewpoints." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.597, + 0.486, + 0.853 + ], + "angle": 0, + "content": "In this paper, we propose a novel approach that avoids the challenge of obtaining granular information about individual sound sources and addresses some of the shortcomings of existing NVAS techniques. Our method relies on sparsely distributed \"reference microphones\" to capture acoustic snapshots, taking their recorded audio signals as \"reference recordings\". This allows us to get a holistic representation of the acoustic environment from references recordings. Given many examples of reference recordings, and sounds recorded at known locations within the scene, we formulate NVAS as learning the transformation from reference recordings to sound recorded at target viewpoints. At inference time, given any pre-acquired reference recordings of the scene as input and a query for an arbitrary target position, we expect the model to output binaural audio that is acoustically consistent with the query viewpoint and the content from reference recordings." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.486, + 0.903 + ], + "angle": 0, + "content": "To address the limitations of existing NVAS techniques in reference microphone placement, we develop a multi-modal approach for optimal reference location sampling. Our ap-" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.908, + 0.288 + ], + "angle": 0, + "content": "proach can adapt to diverse, large complex scenes using a Visual-Acoustic Binding (VAB) module. VAB learns acoustically relevant features (VAB embeddings) by pretraining an encoder to align visual features from panoramic RGB-D captures with acoustic features from echo responses. Using VAB embeddings from numerous candidate reference locations, we employ a sampler that automatically identifies representative spots by finding spatial regions with similar embeddings. These spatial regions represent areas with similar acoustic properties and are usually free from obstacles that significantly hinder sound propagation. This approach optimizes reference microphone placement within a limited reference budget which enhances the overall performance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.288, + 0.909, + 0.438 + ], + "angle": 0, + "content": "Furthermore, to manage varying numbers of references – which typically depend on scene size and budget – and their unequal contributions to the final audio, we introduce an adaptive reference integration module. This module models sequences via a transformer, using VAB features to reweight reference inputs and viewpoints. These reweighted inputs then serve as conditions for the final binaural audio renderer. The resulting conditions are both scene-adapted and content-invariant, enabling effective handling of various reference microphone configurations and audio content." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.439, + 0.909, + 0.545 + ], + "angle": 0, + "content": "To demonstrate the effectiveness of our approach, we benchmark the proposed model in both real-world setting and in a challenging simulated dataset derived from Matterport3D [3] scenes with SoundSpaces [4]. SoundVista outperforms existing methods [1, 5, 6, 8, 21] on scenes with multiple varieties of sound sources, handling varying numbers up to ten sound sources." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.555, + 0.657, + 0.57 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.581, + 0.909, + 0.792 + ], + "angle": 0, + "content": "Acoustic Scene Synthesis. Traditional methods focus on estimating the room impulse response (RIR) to recreate spatial audio. This is achieved by convolving the sound from each emitter with the RIR corresponding to the emitter-listener pair and summing the results [9, 20, 23, 25, 35-37, 39, 43]. These RIR-based techniques often require detailed source information, such as a clean signal of each source and its precise location, which are typically unknown in real-world scenarios, making them challenging to implement. Alternative approaches use scene-descriptor images for direct spatial sound synthesis based on reference sound input, such as visual acoustic matching [5, 7] or visual-guided audio spatialization [12, 13, 51, 57]. However, these methods may be inaccurate when viewpoints change continuously." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.911, + 0.903 + ], + "angle": 0, + "content": "Novel-view Acoustic Synthesis. Novel-view acoustic synthesis techniques address limitations in acoustic scene synthesis by learning transfer functions from reference sounds to target viewpoints [6, 8, 21, 38, 46]. Methods such as Few-shotRIR [25] and BEE [8] use multiple references but lack optimal reference sampling strategies, relying on heuristic settings such as a single close reference [6, 21, 38, 46]," + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "8332" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.102, + 0.092, + 0.887, + 0.344 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.346, + 0.907, + 0.403 + ], + "angle": 0, + "content": "Figure 2. Details of the SoundVista Pipeline: (a) The reference location sampler selects optimal reference locations leveraging embeddings from visual-acoustic binding (VAB). (b) The reference integration transformer uses VAB embeddings to derive contribution weights for each reference. (c) Reweighting by contribution weights adjusts and integrates reference recording channels and pose conditioning for precise sound synthesis. (d) The spatial audio renderer converts reweighted channels and conditions to binaural sound at the target viewpoint." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.416, + 0.484, + 0.493 + ], + "angle": 0, + "content": "random sampling [25], or fixed references [8], which limit adaptation to diverse scene layouts. In contrast, our method optimizes reference location sampling via visual-acoustic binding and derives adaptive reference contribution weights, enhancing sound synthesis accuracy and generalization." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.496, + 0.485, + 0.784 + ], + "angle": 0, + "content": "Acoustic-Visual Learning. Techniques such as CLIP [11, 22, 24, 31, 52] have successfully aligned visual and linguistic inputs to learn matched deep semantic representations. These advances led to recent works to leverage complementary nature of audio and visual data for acoustic related tasks. For example, AV-RIR [37] binds RIR with panoramic images to improve the late reverberation modeling of RIR. Advances in multi-modal visual understanding have facilitated sound source localization [15, 18, 27, 44], audio-visual speech enhancement and source separation [10, 26, 30, 45, 54, 56], as well as acoustic scene reconstruction [5, 6, 8, 25, 37, 41]. These works underscore the strong link between visual and acoustic modalities. Inspired by these approaches, we developed a visual encoder that aligns features from RGB-D visual data with acoustic features from echo responses. The alignment enables us to infer acoustic properties using visual data only. As a result, our method does not require knowing acoustic parameters or RIRs during inference. It is also easily adaptable to various scene and acoustic environments." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.794, + 0.182, + 0.809 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.818, + 0.295, + 0.833 + ], + "angle": 0, + "content": "3.1. Problem Formulation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.486, + 0.903 + ], + "angle": 0, + "content": "Given \\(N\\) pairs of co-located microphones and cameras, we aim to strategically place them in the acoustic scene. Let \\(\\mathbf{L}_i = (x_i,y_i,z_i)\\in \\mathbb{R}^3\\) denote the location of the devices and \\(\\theta_{i}\\in \\mathbb{R}^{2}\\) denote their orientations. The microphones cap" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.417, + 0.908, + 0.598 + ], + "angle": 0, + "content": "ture reference audio recordings \\(\\mathbf{A}_i\\), and the cameras capture panoramic RGB-D views \\(\\mathbf{V}_i\\). These \"reference captures\" encapsulate the ambient sound and visual properties, providing the basis for reconstructing binaural ambient sound \\(\\tilde{\\mathbf{A}}_k\\) for a target listener at arbitrary location \\(\\mathbf{L}_k = (x_k, y_k, z_k) \\in \\mathbb{R}^3\\) with orientation \\(\\theta_k \\in \\mathbb{R}^2\\). The reference recording can be of any type, but we use ambisonic microphones as they are ideal for capturing sound fields from multiple directions and simplify binaural signal extraction. In contrast to existing methods [1], we avoid enumerating sound sources. Instead, we learn a transfer function \\(\\mathcal{F}\\) from reference captures to the target as:" + }, + { + "type": "equation", + "bbox": [ + 0.553, + 0.602, + 0.907, + 0.628 + ], + "angle": 0, + "content": "\\[\n\\mathcal {F}: \\left(\\left\\{\\mathbf {A} _ {i}, g _ {i}, r _ {k i}, \\boldsymbol {\\theta} _ {i} \\right\\} _ {i = 1} ^ {N}, g _ {k}, \\boldsymbol {\\theta} _ {k}\\right) \\mapsto \\tilde {\\mathbf {A}} _ {k}, \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.553, + 0.631, + 0.867, + 0.647 + ], + "angle": 0, + "content": "\\[\ng _ {i} = \\phi (\\mathbf {V} _ {i}), g _ {k} = \\phi (\\mathbf {V} _ {k}), \\text {a n d} r _ {k i} = \\mathbf {L} _ {k} - \\mathbf {L} _ {i},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.659, + 0.907, + 0.735 + ], + "angle": 0, + "content": "where \\(g_{i}\\) and \\(g_{k}\\) are Visual-Acoustic Binding (VAB) embeddings extracted with a pre-trained visual encoder \\(\\phi (\\cdot)\\), designed to align visual with acoustic representations. \\(r_{ki}\\) is relative vector introduced to avoid overfitting on absolute locations and to handle diverse scene coordinate systems." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.909, + 0.902 + ], + "angle": 0, + "content": "We parameterize \\(\\mathcal{F}\\) using neural networks that incorporate a transformer network and spatial audio renderer module. The transformer generates an adaptive mask to weight the contribution of each reference recording based on the target viewpoint and the VAB embeddings associated with reference viewpoints. We call this the \"Reference Integration Transformer\". The spatial audio renderer then processes the reweighted audio channels along with conditional information from the target view features to produce the final binaural sound. We provide an overview of these modules in Figure 2, with details explained in the following sections." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8333" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.366, + 0.108 + ], + "angle": 0, + "content": "3.2. Visual-Acoustic Binding (VAB)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.114, + 0.486, + 0.205 + ], + "angle": 0, + "content": "Obtaining acoustic properties of real-world scenes is challenging [9]. We propose to leverage RGB-D data, which can provide rich information linked to acoustic properties of a scene. For example, depth data provides information about obstacles and room geometry, while pixel-level textures reveal material differences and detailed obstacle information." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.206, + 0.486, + 0.312 + ], + "angle": 0, + "content": "The goal is to infer the acoustic properties of a scene using only the visual data. To achieve this, we first collected extensive paired panoramic RGB-D and acoustic echo responses data by navigating walkable areas in the SoundSpaces simulator [4]. We then trained a neural network, which we refer as \"Visual-Acoustic Binding (VAB) module\", to predict the acoustic representations from the visual features." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.317, + 0.486, + 0.529 + ], + "angle": 0, + "content": "Acoustic Representation. In acoustics, an impulse response is a function of time and the positions of the emitter and the listener [47]. It describes how sound propagates through a medium and interacts with the environment [32, 33]. To simplify data collection and infer local acoustic characteristic, we focus on echo impulse responses [25], where the emitter and target locations are the same. We extracted reverberation time (aka \\(\\mathrm{RT}_{60}\\) parameter) from echo response to use as acoustic representation. \\(\\mathrm{RT}_{60}\\) measures how long it takes for the acoustic energy to decay by \\(60\\mathrm{dB}\\) and can reveal information about room geometry, obstacles and reflection. Figure 2(a) shows an example \\(\\mathrm{RT}_{60}\\) map of a scene. Significant value changes in \\(\\mathrm{RT}_{60}\\) map can indicate major obstacles or surface variations, highlighting key acoustic regions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.532, + 0.485, + 0.609 + ], + "angle": 0, + "content": "Visual Representation. Following prior works [6, 37], we use panoramic RGB-D captures as inputs and extract visual features from ResNet-18 [14]. We train the VAB module to predict \\(\\mathrm{RT}_{60}\\). Therefore, effective binding visual representation with acoustic representation." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.62, + 0.347, + 0.638 + ], + "angle": 0, + "content": "3.3. Reference Location Sampler" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.643, + 0.485, + 0.719 + ], + "angle": 0, + "content": "While having more microphones and cameras is ideal for this task, in practice resources are often limited. To maximize performance and make the best use of available resources, we propose a strategic approach for placement of reference microphones throughout a given scene." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.486, + 0.902 + ], + "angle": 0, + "content": "We argue that ideal microphone placements align with the acoustic partitions of the scene-areas with unique acoustic properties and free from obstacles such as walls. To identify these partitions, we capture panoramic RGB-D images from all candidate locations. One can use novel-view synthesis [49] to render these images; without needing actual photographs at each spot. We then extract VAB embeddings from each capture. VAB embeddings which correspond to scene acoustic parameters serve as strong cues for identifying acoustic partitions, such as distinguishing regions separated by obstacles. To enhance reliability, we combine the extracted VAB embeddings with positional information," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.909, + 0.153 + ], + "angle": 0, + "content": "allowing them to complement each other. Using these embeddings, we perform data clustering of the candidate locations and take the center of each cluster as a reference location. This is illustrated in Figure 2(a)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.158, + 0.822, + 0.173 + ], + "angle": 0, + "content": "3.4. Reference Integration Transformer" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.18, + 0.909, + 0.346 + ], + "angle": 0, + "content": "We want our model to work effectively across diverse scenes. This requires the model to adapt to varying numbers of reference audio and visual inputs. Logically, larger scenes would benefit from more microphones and cameras, while smaller spaces can operate efficiently with fewer resources. Moreover, since the task (Section 3.1) is to transfer reference recordings to target sounds at specific viewpoints, distant microphones, or those in non-adjacent acoustic partitions usually contribute significantly less to the process than closer ones. However, weighting based solely on distance is insufficient due to potential obstacles such as walls and objects." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.346, + 0.909, + 0.437 + ], + "angle": 0, + "content": "To address these issues, we propose a transformer network. We treat each reference input as part of a sequence, allowing us to manage varying numbers of reference inputs. To derive the weights for their unequal contributions to the final audio, we exploit VAB embeddings extracted from visual captures at the reference and target viewpoints." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.437, + 0.909, + 0.603 + ], + "angle": 0, + "content": "We formed the queries input by combining the VAB embedding at the target location \\( g_{k} \\), with a learnable latent query embedding \\( \\mathbf{e} \\). The query is initialized from a normal distribution and optimized during training, to adjust the contribution of reference microphones based on their relative location vector to the target location. Let \\( g_{k}^{e} = [g_{k}||\\mathbf{e}] \\in \\mathbb{R}^{\\mathrm{C}} \\) to denote the queries. The keys and values inputs are created by combining the VAB embedding of each reference with its relative vector \\( r_{ki} \\). Let \\( g_{i}^{r} = [g_{i}||r_{ki}] \\in \\mathbb{R}^{\\mathrm{C}} \\) to denote the keys and values of \\( i^{\\text{th}} \\) reference for the attention operation. The attention weight \\( a_{ki} \\) can be calculated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.582, + 0.611, + 0.907, + 0.645 + ], + "angle": 0, + "content": "\\[\na _ {k i} = \\frac {g _ {k} ^ {e} \\cdot g _ {i} ^ {r} \\top}{\\sqrt {\\mathsf {C}}}, \\quad i = \\{1, 2, \\dots , N \\}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.651, + 0.91, + 0.698 + ], + "angle": 0, + "content": "The attention weight \\( a_{ki} \\) indicates contribution of the reference \\( i \\) to the target \\( k \\) normalized by the Softmax function over \\( N \\) references. Higher weights indicate greater influence." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.705, + 0.733, + 0.722 + ], + "angle": 0, + "content": "3.5. Spatial Audio Renderer" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.727, + 0.907, + 0.773 + ], + "angle": 0, + "content": "The Spatial Audio Renderer takes the reference recordings and conditional information to generate binaural sound at the target viewpoint." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.776, + 0.909, + 0.868 + ], + "angle": 0, + "content": "Reweighting Reference Recordings. We apply the Short-Time Fourier Transform (STFT) to the reference recordings to extract their spectrograms. Then, we reweight the channels of each reference recording using the attention weights \\( a_{ki} \\) from Section 3.4 to integrate them, encoding the result into a latent space as the input for the renderer." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Reweighting Condition Inputs. The binaural sound is derived based on the target's position and pose. To capture the" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8334" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.228 + ], + "angle": 0, + "content": "different aspects of the binaural effects, we decouple the conditioning into global and local components. This accounts for the cause of spatialization effects at a given target point, which may primarily vary from the distance to the sound source or remain invariant to head orientation. The global condition \\( c_{g} \\) determines how the target viewpoint's relative position to reference locations influences the binaural sound. The local condition \\( c_{l} \\) accounts more for the effects of head orientation. They are defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.16, + 0.234, + 0.484, + 0.275 + ], + "angle": 0, + "content": "\\[\nc _ {g} = \\sum_ {i = 1} ^ {N} a _ {k i} \\cdot \\psi_ {1} \\left(g _ {i} \\| r _ {k i}\\right), \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.16, + 0.279, + 0.484, + 0.32 + ], + "angle": 0, + "content": "\\[\nc _ {l} = \\left(\\sum_ {i = 1} ^ {N} a _ {k i} \\cdot \\psi_ {2} \\left(r _ {k i} \\| q _ {k i}\\right)\\right) + \\sigma \\left(\\theta_ {k}\\right). \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.325, + 0.484, + 0.386 + ], + "angle": 0, + "content": "Here \\( q_{ki} \\) is the relative orientation quaternion for target \\( k \\) relative to reference \\( i \\); \\( \\psi_1 \\) and \\( \\psi_2 \\) are MLP layers projecting concatenated inputs to a latent condition embedding; and \\( \\sigma(\\theta_k) \\) denotes the target's rotational features from \\( \\theta_k \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.389, + 0.484, + 0.509 + ], + "angle": 0, + "content": "**Renderer.** The Audio Renderer network is designed as a stacked U-Net [40, 42], depicted in Figure 2(d). It features down-sampling and up-sampling blocks that incorporate input conditions at each layer. To preserve quality and content details from the reference input, skip connections are employed at multiple resolutions. After processing the integrated reference recordings, we apply the inverse STFT to convert the output spectrogram into a binaural waveform." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.51, + 0.486, + 0.631 + ], + "angle": 0, + "content": "Training from scratch in complex environments with challenging audio content can hinder the renderer to adapt to head orientation while producing high-quality audio. To address this, we pretrain the binauralization capability of our renderer by fixing the target location and varying head rotations and audio sources. After pre-training, the renderer effectively understands binauralization across different orientations, enhancing the accuracy of spatial effect modeling." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.638, + 0.165, + 0.652 + ], + "angle": 0, + "content": "3.6. Loss" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.66, + 0.486, + 0.901 + ], + "angle": 0, + "content": "We design the loss function as a weighted combination of three components, each regulating different aspects of the model: 1) Waveform Loss measures the mean squared error between the predicted and target waveforms, ensuring precision of the predicted waveform amplitude and phases. 2) Binaural Interaural Level Difference (ILD) Loss [16] focuses on the energy difference between binaural audio channels to ensure accurate spatial effect prediction. 3) Multi-resolution Spectrogram Magnitude Loss [53] ensures that the predicted audio matches the target spectrogram magnitude across multiple resolutions. It includes two terms: the first compares log-scaled magnitudes, and the second, the Scaled Spectrogram Magnitude Loss, calculates magnitude distance over the target scale, addressing high variance in audio magnitude. We found that resolution with large FFT and small hop sizes can benefit modeling ambient noise, such as air vibrations." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.091, + 0.646, + 0.108 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.117, + 0.744, + 0.132 + ], + "angle": 0, + "content": "4.1. Benchmarks and Metrics" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.143, + 0.908, + 0.326 + ], + "angle": 0, + "content": "N2S Benchmark. Captured in a real room of 8.5 by 6 meters dimensions with two internal rooms, the N2S benchmark utilizes microphone arrays with 32 microphones for 32-channel recordings. Visual captures are rendered by the NeRF-based 3D NVS model, VR-NeRF [49], after it has been trained on visual data collected by navigating the room. 11 static microphone arrays are uniformly distributed to provide reference ambient recordings, while 6 mobile arrays cover 557 locations with 15 seconds of recording per location. Motion capture system, OptiTrack, records 6DoF tracking data for both reference and target microphones. \\(35\\%\\) of these locations with 20 orientations are selected for training." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.33, + 0.909, + 0.466 + ], + "angle": 0, + "content": "Soundspace-Ambient Matterport3D. We build this benchmark based on Matterport3D scenes, comprising 39 complex training scenes and 23 unseen scenes. It is challenging as each scene can include 2 to 10 sound sources, selected from 128 different sounds, from fan noise to speech. Reference locations are spaced every 5 meters on average, with simulated second-order ambisonics RIRs to render reference sounds by Soundspaces [4]. \\(85\\%\\) of walkable locations are used for training with azimuth angles of 0, 90, 180, and 270." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.471, + 0.909, + 0.624 + ], + "angle": 0, + "content": "Metrics. We evaluate sound synthesis performance using four key metrics (the lower, the better): 1) L1 distance of STFT spectrogram (STFT) [13] for left and right channels; 2) Magnitude Spectrogram Distance (MAG) [13], measuring the closeness of reconstructed audio to the groundtruth; 3) Energy Envelope Error (ENV) [13], assessing the Euclidean distance between energy envelopes of groundtruth and predicted audio channels; 4) Left-Right Energy Ratio Error (LRE) [6], evaluating binaural effect accuracy by calculating energy ratio difference between left and right channels." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.636, + 0.76, + 0.652 + ], + "angle": 0, + "content": "4.2. Comparison with Baselines" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.659, + 0.909, + 0.903 + ], + "angle": 0, + "content": "We compare our method with the following baseline approaches: (1) Nearest GT: Binaural sound from the nearest reference microphone aligned with the target orientation. (2) Interp GT: Employs binaural sound at the target orientation from the four nearest reference microphones, interpolating based on distance to the target location. (3) AV-NeRF [21]: A NeRF-based system that synthesizes binaural audio from a given camera pose and RGB-D renderings. We adapted it to use recordings, poses, and visual context from references while preserving its core model components, enabling fair comparisons in dynamic scenes. (4) Few-shotRIR [25]: A transformer-based method that infers RIRs from a sparse set of observed images and echoes. Adapted to replace the reference and target impulse responses with the ambient sounds at the corresponding viewpoints. (5) VAM [5]: Matches the acoustics of input audio with a target image. (6) BEE [8]:" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8335" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.907, + 0.131 + ], + "angle": 0, + "content": "Table 1. Results Comparison on Soundspace-Ambient Benchmark: Average metrics for 10,189 samples in 39 seen scenes with novel target locations and sources, and 6,534 samples in 23 unseen scenes. Sampling strategies: location only, vis+location (our sampler), and w/o VAB (ours without VAB pretraining). SoundVista Ref Num \\( k \\) : only references with top- \\( k \\) contribution weights are used for fair comparison." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.133, + 0.899, + 0.372 + ], + "angle": 0, + "content": "
MethodSamplingRef NumSeen ScenesUnseen Scenes
STFT ↓MAG ↓ENV ↓LRE ↓STFT ↓MAG ↓ENV ↓LRE ↓
nearest GTlocation14.4480.3510.1541.5964.0340.3530.1551.617
nearest GTw/o VAB14.4140.3410.1511.5764.6800.3470.1531.537
nearest GTvis+location13.9160.3360.1461.5723.8350.3440.1511.557
interp GTlocation43.9220.3260.1441.5843.4100.3270.1451.587
interp GTw/o VAB43.6600.3190.1421.5703.7660.3200.1421.531
interp GTvis+location43.1790.3130.1371.5593.4150.3210.1411.531
AV-NeRF [21]vis+location19.4240.4260.1951.9229.3210.4280.1961.979
VAM [5]vis+location15.2240.4200.1781.9024.9360.4360.1821.977
ViGAS [6]vis+location13.7400.3610.1542.0403.4380.3710.1572.051
SoundVista (Ours)vis+location12.5260.2910.1321.4082.6760.3090.1401.386
BEE [8]vis+location44.0980.3650.1622.0835.6350.3960.1782.131
SoundVista (Ours)vis+location42.4440.2890.1301.3902.5170.3050.1371.371
Few-shotRIR [25]vis+locationall5.9370.4590.2131.8925.4570.4710.2151.960
SoundVista w/o vislocationall3.2280.3060.1411.4252.8900.3120.1421.439
SoundVista (Ours)vis+locationall2.4420.2890.1301.3902.5140.3050.1371.372
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.385, + 0.483, + 0.426 + ], + "angle": 0, + "content": "Table 2. Testing Results Comparison on N2S Benchmark. In this real-world scene, SoundVista with visual modality largely boosts the performance accuracy, especially on the binaural effect (LRE)." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.427, + 0.481, + 0.569 + ], + "angle": 0, + "content": "
MethodSTFT ↓MAG ↓ENV ↓LRE ↓
Nearest DSP2.4200.2120.1361.447
Interp DSP1.6590.2030.1421.383
AV-NeRF [21]2.1940.1870.1190.840
Few-shotRIR [25]1.7650.1990.1340.909
VAM [5]1.9720.1900.1190.916
BEE [8]1.4710.2000.1410.995
ViGAS [6]1.2010.1850.1190.873
SoundVista w/o vis1.2420.1850.1180.894
SoundVista (Ours)1.0730.1770.1130.776
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.582, + 0.483, + 0.748 + ], + "angle": 0, + "content": "A generalizable rendering pipeline that reconstructs the binaural audio at an arbitrary listener location using inputs from sparse reference audio-visual samples in the scene. (7) ViGAS [6]: Transforms sound to the target viewpoint given the observed audio and visual captures at the reference viewpoint. Specifically, for AV-NeRF, VAM, and ViGAS, which need a single reference, we use the nearest reference microphone. BEE is adapted to use the nearest 4 microphones. Deep-learning baselines requiring visual inputs utilize panoramic RGB-D images and the same visual encoder as our model for consistency." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.75, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Our results in Table 1 and Table 2 demonstrate that our method, SoundVista, consistently surpasses baselines across diverse novel scenes and real scenarios. Despite the challenge of obtaining binaural groundtruth (GT) for arbitrary target orientations, SoundVista significantly reduces errors across all metrics compared with nearest GT and interp GT. On the challenging Soundspace-Ambient benchmark, with diverse sound sources and complex layouts, most deep-learning baselines underperform compared to non-learning methods due to the need for robust conditioning models." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.387, + 0.908, + 0.583 + ], + "angle": 0, + "content": "For a fair comparison, we evaluate SoundVista using the top 1 and 4 reference microphones by contribution weight, respectively. Compared to learning-based methods with 1 reference, SoundVista outperforms the best baseline ViGAS by \\(32.5\\%\\) (STFT), \\(19.4\\%\\) (MAG), \\(14.3\\%\\) (ENV), and \\(31\\%\\) (LRE), demonstrating the strength of our conditioned audio renderer. The nearest microphone may miss critical audio content, while using the top 4 references improves performance, matching that of all microphones since most references are not required due to their sparse distribution. Since contribution weights are independent of audio content, fewer than 4 references can be selected per target location by precomputing contribution weights once per scene." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.583, + 0.909, + 0.688 + ], + "angle": 0, + "content": "While \\(Few-shotRIR\\) and \\(BEE\\) use audio content to integrate references, the diverse sound distribution makes this challenging. During testing, content often falls outside the training distribution, degrading performance. Thus, more references aren't always better; for example, \\(Few-shotRIR\\) did not gain an increase in synthesis accuracy and it was even reduced compared to baselines using only one reference." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.909, + 0.765 + ], + "angle": 0, + "content": "The deep-learning baselines perform better on the N2S benchmark, which features a single static scene. Compared to these deep-learning baselines, our SoundVista still significantly outperforms on accuracy of energy (ENV) and binaural effect (LRE) by \\(5\\%\\) and \\(7.6\\%\\), respectively." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.773, + 0.796, + 0.79 + ], + "angle": 0, + "content": "4.3. Qualitative Results Comparison" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Figure 3 visualizes our results for qualitative comparison with several competitive sound synthesis baselines: ViGAS, BEE, and Few-shotRIR. The first row displays loudness maps, while the second row shows generated binaural waveforms with corresponding LRE error displayed in the third row. ViGAS relies on the nearest microphone, leading to discrete loudness maps in complex layouts and unreliable results" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8336" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.092, + 0.512, + 0.281 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.092, + 0.901, + 0.281 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.282, + 0.908, + 0.311 + ], + "angle": 0, + "content": "Figure 3. Comparison of Qualitative Results: First row: Loudness Map (black triangle: sources; purple triangle: target viewpoint). Second row: Reconstructed binaural waveform at target viewpoint (red: prediction; green: GT). Third row: LRE (lower for better binaural effect)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.326, + 0.486, + 0.538 + ], + "angle": 0, + "content": "when obstacles are present, as shown in the unseen results. Additionally, ViGAS struggles with accurate binaural effects shown from the LRE results. BEE produces discontinuous loudness maps, deviating from the ground truth. Both BEE and Few-shotRIR have difficulty generating high-fidelity waveforms close to the groundtruth. Their reliance on incorporating audio content into the renderer's conditioning can bias the renderer leading to confusion and reduced quality when encountering content outside the training distribution. This is particularly common in our complex task setting with diverse audio sources. In contrast, SoundVista, closely matches the GT in loudness maps and binaural waveforms for both seen and unseen scenes, excelling in high-quality novel-view ambient sound synthesis." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.551, + 0.245, + 0.567 + ], + "angle": 0, + "content": "4.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.574, + 0.484, + 0.606 + ], + "angle": 0, + "content": "We conducted ablation studies on the Soundspace-Ambient benchmark to verify the contribution of key components." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.609, + 0.484, + 0.776 + ], + "angle": 0, + "content": "Visual-Acoustic Binding (VAB). In the design of VAB, we optimized the integration of visual features that correlate with acoustic parameters. Table 3 shows the accuracy of \\(\\mathrm{RT}_{60}\\) value predictions in novel scenes without finetuning (w/o finetune) and with few-shot finetuning at reference locations (w/finetune). We tested various modality inputs: location only, panoramic rgb, depth, and rgb+depth. Depth alone significantly enhances binding, reducing error by over \\(50\\%\\) compared to use location only. With finetuning, rgb+depth achieves superior results. Given the ability to capture references in scenes, we select rgb+depth as our visual input." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Reference Sampler via VAB. The Reference Sampler via Visual-Acoustic Binding (VAB) (vis+location) significantly enhances the accuracy of non-learning baselines such as nearest GT and interp GT, as shown in Table 1. This method improves predictions of sound magnitude (MAG, ENV) and spectrogram phases (STFT). Without pretrained VAB (w/o VAB), these benefits diminish, especially in seen scenes. Results in Table 1, 2, and 4 further highlight the critical role of" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.324, + 0.908, + 0.366 + ], + "angle": 0, + "content": "Table 3. \\(\\mathbf{RT}_{60}\\) Prediction Results on Matterport3D. w/ finetune involves few-shot finetuning on unseen scenes. err: distance between predicted and GT \\(\\mathrm{RT}_{60}\\), error ratio: scaled err over GT." + }, + { + "type": "table", + "bbox": [ + 0.536, + 0.367, + 0.882, + 0.454 + ], + "angle": 0, + "content": "
Modalityw/o finetunew/ finetune
err ↓err ratio ↓err ↓err ratio ↓
location0.1700.4110.0670.199
rgb0.1170.2360.0440.126
depth0.0820.1730.0380.109
rgb+depth0.0870.1850.0360.103
" + }, + { + "type": "table_caption", + "bbox": [ + 0.521, + 0.457, + 0.898, + 0.47 + ], + "angle": 0, + "content": "Table 4. Ablations on the SoundSpace-Ambient benchmark." + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.472, + 0.905, + 0.616 + ], + "angle": 0, + "content": "
MethodSTFT ↓MAG ↓ENV ↓LRE ↓
full model2.4420.2890.1301.390
w/o vis3.2280.3060.1411.425
w/o reweight2.9080.2950.1341.391
w/o render_pretrain2.5820.3100.1351.877
w/o decoup_cond3.0170.3000.1381.433
full loss2.4420.2890.1301.390
w/o ILD_loss2.4620.2920.1311.420
w/o mag_scale2.4240.3010.1361.422
w/o wav_loss3.4220.2910.1431.369
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.625, + 0.907, + 0.67 + ], + "angle": 0, + "content": "visual input, as its absence increases errors across all metrics, underscoring the importance of VAB with visual information for adapting to diverse scene layouts." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.671, + 0.909, + 0.791 + ], + "angle": 0, + "content": "Figure 4 (left) illustrates the STFT error curve relative to reference density. Our model is trained with a reference density of 1.0, and tested on novel Soundspace-Ambient scenes with varying densities. Compared to Ours - w/o vis and interp - w/o vis (interp GT with location-only sampler), Ours - w/ vis outperforms others, with the performance gap widening as density decreases, demonstrating robustness of ours equipped with VAB vs. reference density variance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Reference Integration. To evaluate the contribution of the adaptive reference integration, we conducted experiments with the w/o reweight variant, which uses only the nearest reference microphone along with its corresponding visual and pose information to calculate the condition for the audio renderer. As shown in Table 4, the results of w/o reweight compared to full SoundVista model reveal a noticeable degra" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8337" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.088, + 0.283, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.283, + 0.089, + 0.483, + 0.218 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.091, + 0.228, + 0.483, + 0.259 + ], + "angle": 0, + "content": "Figure 4. STFT Error Curve for different reference densities (left) and training location data ratios (right)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.277, + 0.484, + 0.368 + ], + "angle": 0, + "content": "dation in performance, particularly in the STFT error, which increases by \\(19\\%\\). This indicates that the adaptive Reference Integration Transformer and reweighting modules effectively integrate references to enhance sound synthesis accuracy, as the nearest microphone may miss critical content required for the target sound synthesis." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.377, + 0.484, + 0.468 + ], + "angle": 0, + "content": "Conditioned Spatial Audio Renderer. In the w/o decoup_cond variant, we replaced the decoupled global and local conditions from reweighting in Section 3.5 with only the global condition plus target rotational features. This absence of local conditions significantly increases STFT error by \\(23.5\\%\\), MAG error by \\(3.8\\%\\), and ENV error by \\(6.2\\%\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.474, + 0.484, + 0.566 + ], + "angle": 0, + "content": "As detailed in Section 3.5, we pretrain the renderer with binauralization capability to help the model learn the complex task of orientation-sensitive novel view sound synthesis. The w/o render_pretrain variant shows a \\(35\\%\\) increase in LRE error, underscoring the importance of renderer pretraining for achieving a more accurate binaural effect." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.574, + 0.484, + 0.711 + ], + "angle": 0, + "content": "Loss. The ablations in Table 4 highlight the contributions of different loss components. The Waveform Loss (wav_loss) constrains the energy envelope and regulates phase learning. Removing it significantly reduces STFT and ENV accuracy, increasing the STFT error by \\(29\\%\\). The Scaled Spectrogram Magnitude Loss (mag_scale) calculates magnitude loss over the target magnitude scale, addressing high variance in audio content and mitigating scale differences. Removing mag_scale will increase MAG error largely." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Robustness w.r.t. Training Data Ratio. To assess the impact of training location data ratio on performance, we trained two variants, Ours - w/o vis and Ours - w/ vis, using different proportions of training locations on the Soundspace-Ambient benchmark. Figure 4 (right) shows that STFT error increases as the data ratio decreases. When the data ratio is below \\(10\\%\\), Ours - w/o vis underperforms compared to the non-learning baseline interp - w/o vis, while Ours - w/ vis maintains superior performance. Additionally, Ours - w/ vis benefits more from increased training locations, with STFT error decreasing rapidly as the ratio nears \\(100\\%\\), widening the performance gap with Ours - w/o vis and interp - w/o vis." + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.091, + 0.895, + 0.254 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.257, + 0.907, + 0.342 + ], + "angle": 0, + "content": "Figure 5. Visualization Analysis. First three columns: Clustering results for loc only, Vis - w/o VAB (ours without VAB pretraining), and Vis - w/ VAB (ours). Colors indicate cluster regions. VAB helps clustering to align better with room partitions. Last two columns: Attention weights visualization for references, with colored stars (size proportional to weights) and a purple triangle for the target." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.357, + 0.722, + 0.373 + ], + "angle": 0, + "content": "4.5. Visualization Analysis" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.379, + 0.908, + 0.499 + ], + "angle": 0, + "content": "Figure 5 visualizes the clustering results of Reference Sampler via VAB, comparing loc only, Vis - w/o VAB, and Vis - w/ VAB. Different colors represent cluster regions. The Reference Sampler Vis - w/ VAB accurately segments clusters aligning with actual room partitions, proving to be more reliable. In contrast, other methods may incorrectly cluster non-adjacent rooms together which cannot share the same acoustic environment, highlighted by red-circles." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.501, + 0.909, + 0.636 + ], + "angle": 0, + "content": "After identifying reference locations, we visualize the attention weights from the Reference Integration Transformer in Figure 5 (column 4-5 from left). Colored stars denote sampled references, sized by contribution weight, with the target marked by a purple triangle. Unlike Distance-only weights, which may incorrectly prioritize closer references despite obstacles, our model, Ref Attn - w/ VAB, trained with the adaptive Reference Integration Transformer, effectively highlights reliable references in similar acoustic environments." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.649, + 0.634, + 0.665 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.909, + 0.901 + ], + "angle": 0, + "content": "We introduce SoundVista, a novel system designed to synthesize ambient sound from arbitrary scenes at novel viewpoints. Our approach introduces a visual-acoustic binding module to effectively learn visual embeddings linked with local acoustic properties. This enables our system to optimize the placement of reference microphones and derive adaptive weights for each microphone's contribution, conditioned on the target viewpoint and visual captures, thereby, facilitating the final conditioned sound synthesis. SoundVista adapts to diverse room layouts, microphone configurations, and unseen environments, rendering high-quality binaural ambient sound without requiring prior constraints or detailed knowledge of sound sources. Our results, validated on both publicly available data and real-world settings, demonstrate state-of-the-art sound synthesis accuracy and generalization." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8338" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.486, + 0.184 + ], + "angle": 0, + "content": "[1] Byeongjoo Ahn, Karren Yang, Brian Hamilton, Jonathan Sheaffer, Anurag Ranjan, Miguel Sarabia, Oncel Tuzel, and Jen-Hao Rick Chang. Novel-view acoustic synthesis from 3d reconstructed rooms. arXiv preprint arXiv:2310.15130, 2023. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.187, + 0.485, + 0.269 + ], + "angle": 0, + "content": "[2] Federico Borra, Israel Dejene Gebru, and Dejan Markovic. Soundfield reconstruction in reverberant environments using higher-order microphones and impulse response measurements. In ICASSP 2019-2019 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 281-285. IEEE, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.272, + 0.484, + 0.342 + ], + "angle": 0, + "content": "[3] Angel Chang, Angela Dai, Thomas Funkhouser, Maciej Halber, Matthias Niessner, Manolis Savva, Shuran Song, Andy Zeng, and Yinda Zhang. Matterport3d: Learning from rgb-d data in indoor environments. International Conference on 3D Vision (3DV), 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.343, + 0.484, + 0.424 + ], + "angle": 0, + "content": "[4] Changan Chen, Unnat Jain, Carl Schissler, Sebastia Vicenc Amengual Gari, Ziad Al-Halah, Vamsi Krishna Ithapu, Philip Robinson, and Kristen Grauman. Soundspaces: Audiovisual navigation in 3d environments. In European Conference on Computer Vision, pages 17-36. Springer, 2020. 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.428, + 0.485, + 0.483 + ], + "angle": 0, + "content": "[5] Changan Chen, Ruohan Gao, Paul Calamia, and Kristen Grauman. Visual acoustic matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18858-18868, 2022. 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.486, + 0.484, + 0.541 + ], + "angle": 0, + "content": "[6] Changan Chen, Alexander Richard, Roman Shapovalov, Vamsi Krishna Ithapu, Natalia Neverova, Kristen Grauman, and Andrea Vedaldi. Novel-view acoustic synthesis. arXiv preprint arXiv:2301.08730, 2023. 1, 2, 3, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.543, + 0.484, + 0.583 + ], + "angle": 0, + "content": "[7] Changan Chen, Wei Sun, David Harwath, and Kristen Grauman. Learning audio-visual dereverberation. In ICASSP, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.586, + 0.484, + 0.655 + ], + "angle": 0, + "content": "[8] Mingfei Chen, Kun Su, and Eli Shlizerman. Be everywhere—hear everything (bee): Audio scene reconstruction by sparse audio-visual samples. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7853–7862, 2023. 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.657, + 0.484, + 0.74 + ], + "angle": 0, + "content": "[9] Ziyang Chen, Israel D Gebru, Christian Richardt, Anurag Kumar, William Laney, Andrew Owens, and Alexander Richard. Real acoustic fields: An audio-visual room acoustics dataset and benchmark. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21886-21896, 2024. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.743, + 0.484, + 0.812 + ], + "angle": 0, + "content": "[10] Ariel Ephrat, Inbar Mosseri, Oran Lang, Tali Dekel, Kevin Wilson, Avinatan Hassidim, William T Freeman, and Michael Rubinstein. Looking to listen at the cocktail party: A speaker-independent audio-visual model for speech separation. arXiv preprint arXiv:1804.03619, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.814, + 0.484, + 0.855 + ], + "angle": 0, + "content": "[11] Han Fang, Pengfei Xiong, Luhui Xu, and Yu Chen. Clip2video: Mastering video-text retrieval via image clip. arXiv preprint arXiv:2106.11097, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.858, + 0.484, + 0.884 + ], + "angle": 0, + "content": "[12] Ruohan Gao and Kristen Grauman. 2.5d visual sound. In CVPR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.887, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[13] Ruohan Gao and Kristen Grauman. 2.5 d visual sound. In Pro" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.486, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.548, + 0.093, + 0.905, + 0.121 + ], + "angle": 0, + "content": "ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 324-333, 2019. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.123, + 0.907, + 0.178 + ], + "angle": 0, + "content": "[14] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.179, + 0.907, + 0.247 + ], + "angle": 0, + "content": "[15] Di Hu, Rui Qian, Minyue Jiang, Xiao Tan, Shilei Wen, Errui Ding, Weiyao Lin, and Dejing Dou. Discriminative sounding objects localization via self-supervised audiovisual matching. Advances in Neural Information Processing Systems, 33: 10077-10087, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.25, + 0.907, + 0.318 + ], + "angle": 0, + "content": "[16] Christopher A. Ick, Gordon Wichern, Yoshiki Masuyama, François Germain, and Jonathan Le Roux. Spatially-aware losses for enhanced neural acoustic fields. In Audio Imagination: NeurIPS 2024 Workshop AI-Driven Speech, Music, and Sound Generation, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.321, + 0.907, + 0.375 + ], + "angle": 0, + "content": "[17] Hyeonjoong Jang, Andreas Meuleman, Dahiyun Kang, Donggun Kim, Christian Richardt, and Min H Kim. Egocentric scene reconstruction from an omnidirectional video. ACM Transactions on Graphics (TOG), 41(4):1-12, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.378, + 0.907, + 0.418 + ], + "angle": 0, + "content": "[18] Hao Jiang, Calvin Murdock, and Vamsi Krishna Ithapu. Ego-centric deep multi-channel audio-visual active speaker localization. arXiv preprint arXiv:2201.01928, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.421, + 0.906, + 0.503 + ], + "angle": 0, + "content": "[19] Shoichi Koyama, Tomoya Nishida, Keisuke Kimura, Takumi Abe, Natsuki Ueno, and Jesper Brunnström. Meshrir: A dataset of room impulse responses on meshed grid points for evaluating sound field analysis and synthesis methods. In 2021 IEEE Workshop on Applications of Signal Processing to Audio and Acoustics (WASPAA), pages 1-5, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.505, + 0.907, + 0.559 + ], + "angle": 0, + "content": "[20] Susan Liang, Chao Huang, Yapeng Tian, Anurag Kumar, and Chenliang Xu. Neural acoustic context field: Rendering realistic room impulse response with neural fields. ArXiv, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.562, + 0.907, + 0.617 + ], + "angle": 0, + "content": "[21] Susan Liang, Chao Huang, Yapeng Tian, Anurag Kumar, and Chenliang Xu. Av-nerf: Learning neural fields for real-world audio-visual scene synthesis. In Conference on Neural Information Processing Systems (NeurIPS), 2023. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.619, + 0.907, + 0.687 + ], + "angle": 0, + "content": "[22] Ziyi Lin, Shijie Geng, Renrui Zhang, Peng Gao, Gerard De Melo, Xiaogang Wang, Jifeng Dai, Yu Qiao, and Hongsheng Li. Frozen clip models are efficient video learners. In European Conference on Computer Vision, pages 388-404. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.689, + 0.907, + 0.731 + ], + "angle": 0, + "content": "[23] Andrew Luo, Yilun Du, Michael J Tarr, Joshua B Tenenbaum, Antonio Torralba, and Chuang Gan. Learning neural acoustic fields. arXiv preprint arXiv., 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.733, + 0.906, + 0.786 + ], + "angle": 0, + "content": "[24] Huaishao Luo, Lei Ji, Ming Zhong, Yang Chen, Wen Lei, Nan Duan, and Tianrui Li. CLIP4Clip: An empirical study of clip for end to end video clip retrieval. arXiv preprint arXiv:2104.08860, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.789, + 0.906, + 0.83 + ], + "angle": 0, + "content": "[25] Sagnik Majumder, Changan Chen, Ziad Al-Halah, and Kristen Grauman. Few-shot audio-visual learning of environment acoustics, 2022. 2, 3, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.832, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[26] Daniel Michelsanti, Zheng-Hua Tan, Shi-Xiong Zhang, Yong Xu, Meng Yu, Dong Yu, and Jesper Jensen. An overview of deep-learning-based audio-visual speech enhancement and separation. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 2021. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8339" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.486, + 0.147 + ], + "angle": 0, + "content": "[27] Shentong Mo and Yapeng Tian. Audio-visual grouping network for sound localization from mixtures. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10565–10574, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.486, + 0.231 + ], + "angle": 0, + "content": "[28] Bartlomiej Mróz, Marek Kabaciński, Tomasz Ciotucha, Andrzej Ruminski, and Tomasz Žernicki. Production of six-degrees-of-freedom (6dof) navigable audio using 30 ambisonic microphones. In 2021 Immersive and 3D Audio: from Architecture to Automotive (I3DA), pages 1-5. IEEE, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.233, + 0.486, + 0.3 + ], + "angle": 0, + "content": "[29] Orhun Olgun, Ege Erdem, and Hüseyin Hachhabiboglu. Sound field interpolation via sparse plane wave decomposition for 6 dof immersive audio. In 2023 Immersive and 3D Audio: from Architecture to Automotive (I3DA), pages 1-10. IEEE, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.302, + 0.486, + 0.357 + ], + "angle": 0, + "content": "[30] Andrew Owens and Alexei A Efros. Audio-visual scene analysis with self-supervised multisensory features. In Proceedings of the European Conference on Computer Vision (ECCV), pages 631–648, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.359, + 0.486, + 0.44 + ], + "angle": 0, + "content": "[31] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.442, + 0.486, + 0.482 + ], + "angle": 0, + "content": "[32] Nikunj Raghuvanshi and John Snyder. Parametric wave field coding for precomputed sound propagation. ACM Transactions on Graphics (TOG), 33(4):1-11, 2014. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.484, + 0.486, + 0.524 + ], + "angle": 0, + "content": "[33] Nikunj Raghuvanshi and John Snyder. Parametric directional coding for precomputed sound propagation. ACM Transactions on Graphics (TOG), 37(4):1-14, 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.526, + 0.486, + 0.607 + ], + "angle": 0, + "content": "[34] Santhosh K Ramakrishnan, Aaron Gokaslan, Erik Wijmans, Oleksandr Maksymets, Alex Clegg, John Turner, Eric Undersander, Wojciech Galuba, Andrew Westbury, Angel X Chang, et al. Habitat-matterport 3d dataset (hm3d): 1000 large-scale 3d environments for embodied ai. arXiv preprint arXiv:2109.08238, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.609, + 0.486, + 0.65 + ], + "angle": 0, + "content": "[35] Anton Ratnarajah, Zhenyu Tang, and Dinesh Manocha. *Irgan: Room impulse response generator for far-field speech recognition.* arXiv preprint arXiv:2010.13219, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.651, + 0.486, + 0.704 + ], + "angle": 0, + "content": "[36] Anton Ratnarajah, Shi-Xiong Zhang, Meng Yu, Zhenyu Tang, Dinesh Manocha, and Dong Yu. Fast-rir: Fast neural diffuse room impulse response generator. arXiv preprint arXiv:2110.04057, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.707, + 0.486, + 0.774 + ], + "angle": 0, + "content": "[37] Anton Ratnarajah, Sreyan Ghosh, Sonal Kumar, Purva Chiniya, and Dinesh Manocha. Av-rir: Audio-visual room impulse response estimation. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 27154–27165, 2023. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.776, + 0.486, + 0.844 + ], + "angle": 0, + "content": "[38] Alexander Richard, Dejan Markovic, Israel D. Gebru, Steven Krenn, Gladstone Alexander Butler, Fernando Torre, and Yaser Sheikh. Neural synthesis of binaural speech from mono audio. In International Conference on Learning Representations, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.486, + 0.9 + ], + "angle": 0, + "content": "[39] Alexander Richard, Peter Dodds, and Vamsi Krishna Ithapu. Deep impulse responses: Estimating and parameterizing filters with deep networks. arXiv preprint arXiv:2202.03416, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.486, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.176 + ], + "angle": 0, + "content": "[40] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention-MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, pages 234-241. Springer, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.908, + 0.245 + ], + "angle": 0, + "content": "[41] Nikhil Singh, Jeff Mentch, Jerry Ng, Matthew Beveridge, and Iddo Drori. Image2reverb: Cross-modal reverb impulse response synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 286-295, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.249, + 0.908, + 0.316 + ], + "angle": 0, + "content": "[42] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In International Conference on Learning Representations, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.319, + 0.907, + 0.36 + ], + "angle": 0, + "content": "[43] Kun Su, Mingfei Chen, and Eli Shlizerman. Inras: Implicit neural representation for audio scenes. Advances in Neural Information Processing Systems, 35:8144-8158, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.362, + 0.908, + 0.442 + ], + "angle": 0, + "content": "[44] Ruijie Tao, Zexu Pan, Rohan Kumar Das, Xinyuan Qian, Mike Zheng Shou, and Haizhou Li. Is someone speaking? exploring long-term temporal features for audio-visual active speaker detection. In Proceedings of the 29th ACM international conference on multimedia, pages 3927-3935, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.447, + 0.908, + 0.503 + ], + "angle": 0, + "content": "[45] Yapeng Tian, Di Hu, and Chenliang Xu. Cyclic co-learning of sounding object visual grounding and sound separation. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2744-2753, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.504, + 0.908, + 0.559 + ], + "angle": 0, + "content": "[46] Aäron van den Oord, Sander Dieleman, Heiga Zen, Karen Simonyan, Oriol Vinyals, Alexander Graves, Nal Kalchbrenner, Andrew Senior, and Koray Kavukcuoglu. Wavenet: A generative model for raw audio. In Arxiv, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.562, + 0.888, + 0.574 + ], + "angle": 0, + "content": "[47] Tor Erik Vigran. Building acoustics. CRC Press, 2014. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.577, + 0.908, + 0.631 + ], + "angle": 0, + "content": "[48] Xiuchao Wu, Jiamin Xu, Zihan Zhu, Hujun Bao, Qixing Huang, James Tompkin, and Weiwei Xu. Scalable neural indoor scene rendering. ACM transactions on graphics, 41 (4), 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.633, + 0.908, + 0.716 + ], + "angle": 0, + "content": "[49] Linning Xu, Vasu Agrawal, William Laney, Tony Garcia, Aayush Bansal, Changil Kim, Samuel Rota Bulò, Lorenzo Porzi, Peter Kontschieder, Aljaž Božić, Dahua Lin, Michael Zollhöfer, and Christian Richardt. VR-NeRF: High-fidelity virtualized walkable spaces. In SIGGRAPH Asia Conference Proceedings, 2023. 1, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.718, + 0.908, + 0.8 + ], + "angle": 0, + "content": "[50] Shaoheng Xu, Jihui Aimee Zhang, Thushara D. Abhayapala, Amy Bastine, Wei-Ting Lai, and Prasanga N. Samarasinghe. Sparse sound field representation using complex orthogonal matching pursuit. In ICASSP 2024 - 2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1336-1340, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.802, + 0.908, + 0.844 + ], + "angle": 0, + "content": "[51] Xudong Xu, Hang Zhou, Ziwei Liu, Bo Dai, Xiaogang Wang, and Dahua Lin. Visually informed binaural audio generation without binaural audiios. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.908, + 0.9 + ], + "angle": 0, + "content": "[52] Hongwei Xue, Yuchong Sun, Bei Liu, Jianlong Fu, Ruihua Song, Houqiang Li, and Jiebo Luo. Clip-vip: Adapting pretrained image-text model to video-language representation alignment. arXiv preprint arXiv:2209.06430, 2022. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "8340" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.486, + 0.174 + ], + "angle": 0, + "content": "[53] Ryuichi Yamamoto, Eunwoo Song, and Jae-Min Kim. Parallel wavegan: A fast waveform generation model based on generative adversarial networks with multi-resolution spectrogram. In ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 6199-6203. IEEE, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.178, + 0.485, + 0.232 + ], + "angle": 0, + "content": "[54] Yuxin Ye, Wenming Yang, and Yapeng Tian. Lavss: Location-guided audio-visual spatial audio separation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 5508–5519, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.234, + 0.484, + 0.275 + ], + "angle": 0, + "content": "[55] Zechen Zhang, Nikunj Raghuvanshi, John Snyder, and Steve Marschner. Ambient sound propagation. ACM Transactions on Graphics (TOG), 37(6):1-10, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.277, + 0.484, + 0.331 + ], + "angle": 0, + "content": "[56] Hang Zhou, Ziwei Liu, Xudong Xu, Ping Luo, and Xiaogang Wang. Vision-infused deep audio inpainting. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 283-292, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.333, + 0.484, + 0.416 + ], + "angle": 0, + "content": "[57] Hang Zhou, Xudong Xu, Dahua Lin, Xiaogang Wang, and Ziwei Liu. Sep-stereo: Visually guided stereophonic audio generation by associating source separation. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XII 16, pages 52-69. Springer, 2020. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.486, + 0.416 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.515, + 0.957 + ], + "angle": 0, + "content": "8341" + } + ] +] \ No newline at end of file diff --git a/2025/SoundVista_ Novel-View Ambient Sound Synthesis via Visual-Acoustic Binding/9c309271-ab78-4453-a939-59e8aa02fe41_origin.pdf b/2025/SoundVista_ Novel-View Ambient Sound Synthesis via Visual-Acoustic Binding/9c309271-ab78-4453-a939-59e8aa02fe41_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..1900d68480b9c27181504838e18d6bd353bfeca8 --- /dev/null +++ b/2025/SoundVista_ Novel-View Ambient Sound Synthesis via Visual-Acoustic Binding/9c309271-ab78-4453-a939-59e8aa02fe41_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:233f7d8fa3ecf9918087bc3ade7817a6b95aec7ab4d61c3ea87ba0ff0102b450 +size 1099927 diff --git a/2025/SoundVista_ Novel-View Ambient Sound Synthesis via Visual-Acoustic Binding/full.md b/2025/SoundVista_ Novel-View Ambient Sound Synthesis via Visual-Acoustic Binding/full.md new file mode 100644 index 0000000000000000000000000000000000000000..c25c633d6fefb81eb371b5011559a9c0ff456a1d --- /dev/null +++ b/2025/SoundVista_ Novel-View Ambient Sound Synthesis via Visual-Acoustic Binding/full.md @@ -0,0 +1,290 @@ +# SoundVista: Novel-View Ambient Sound Synthesis via Visual-Acoustic Binding + +Mingfei Chen $^{1,2*}$ Israel D. Gebru $^{2}$ Ishwarya Ananthabhotla $^{3}$ Christian Richardt $^{2}$ +Dejan Markovic $^{2}$ Jake Sandakly $^{2}$ Steven Krenn $^{2}$ Todd Keebler $^{2}$ Eli Shlizerman $^{1}$ Alexander Richard $^{2}$ + +1University of Washington 2Codec Avatars Lab, Pittsburgh, Meta 3Reality Labs Research, Meta + +![](images/8b2ef0b4b30c3b3170c3f44021a83871628edb3d04c44413ad1bccf479f2500f.jpg) +Figure 1. SoundVista: a novel method that synthesizes binaural ambient sound for arbitrary scenes from novel viewpoints. Our method leverages pre-acquired audio recordings and visual data captured from sparsely distributed reference points and synthesizes binaural audio consistent with the target 3D position and pose. + +# Abstract + +We introduce SoundVista, a method to generate the ambient sound of an arbitrary scene at novel viewpoints. Given a pre-acquired recording of the scene from sparsely distributed microphones, SoundVista can synthesize the sound of that scene from an unseen target viewpoint. The method learns the underlying acoustic transfer function that relates the signals acquired at the distributed microphones to the signal at the target viewpoint, using a limited number of known recordings. Unlike existing works, our method does not require constraints or prior knowledge of sound source details. Moreover, our method efficiently adapts to diverse room layouts, reference microphone configurations and unseen environments. To enable this, we introduce a visual-acoustic binding module that learns visual embeddings linked with local acoustic properties from panoramic RGB and depth data. We first leverage these embeddings to optimize the placement of reference microphones in any given scene. During synthesis, we leverage multiple embeddings extracted from reference locations to get adaptive weights for their contribution, conditioned on target viewpoint. We benchmark the task + +on both publicly available data and real-world settings. We demonstrate significant improvements over existing methods. + +# 1. Introduction + +Recent advances in 3D reconstruction and Novel-View Synthesis (NVS) have significantly enhanced our ability to create photorealistic visual models of real-world scenes [17, 34, 48, 49]. These developments have paved the way for applications in 3D virtual tours, experience recreation, and spatial media. However, the audio counterpart – Novel-View Acoustic Synthesis (NVAS) [1, 6] – has been under-explored compared to its visual counterpart and has not received the rigorous attention required to generate acoustic virtual scenes that match their visual surroundings. + +To address this gap, we introduce SoundVista: a novel method that creates a truly immersive acoustic experience. SoundVista can generate realistic and spatially accurate binaural audio at novel viewpoints in arbitrary scenes given a sparse set of sample recordings from different viewpoint. + +Unlike novel-view synthesis for visual 3D scenes, which are mostly static, NVAS faces significant challenges due to the dynamic nature of real-world acoustic environments. Ambient sound – the overall acoustic state describing all sounds + +in a scene [55] – can encompass multiple time-varying, nonstationary sound sources without limitations on the type and number of sound sources. In an ideal scenario, with complete information about the ambient sound (i.e. clean signals of individual sound sources and their locations over time), and a fully reconstructed 3D visual model, one could synthesize the ambient sound at any location using standard acoustic renderers with room impulse responses (RIRs) [1, 4, 9, 19, 43]. In real-world scenes, however, we lack detailed information about the sound sources that comprise the ambient sound. Accurately determining their clean source content and emitter positions is a challenging problem. + +NVAS also faces challenges in balancing data sampling costs and synthesis quality. Current methods in the acoustic community use grids of microphones, and techniques ranging from interpolation [28, 29] to complex signal processing [2, 50] to synthesize audio at novel locations. However, these approaches do not scale well to large spaces, and efficient sampling and data utilization remain open questions. + +Recent deep learning based NVAS methods [5, 6, 8, 21, 25] utilize multimodal data, including visual inputs, to transfer sparse reference sounds to binaural sound at target viewpoints. These methods often simplify the task by focusing on re-synthesizing primary sounds such as speech and music, with a limited number of sound sources (typically 1-2). They often ignore other sounds that contribute to the natural characteristics of the ambient sound scene. While feasible with few references, they struggle to adapt to diverse, large scenes with complex layout and acoustic environment. Furthermore, the lack of an optimal reference location sampling strategy hinders adaptation, leading to a loss of acoustic details and making it difficult to synthesize high-quality sound that has accurate binaural effects for target viewpoints. + +In this paper, we propose a novel approach that avoids the challenge of obtaining granular information about individual sound sources and addresses some of the shortcomings of existing NVAS techniques. Our method relies on sparsely distributed "reference microphones" to capture acoustic snapshots, taking their recorded audio signals as "reference recordings". This allows us to get a holistic representation of the acoustic environment from references recordings. Given many examples of reference recordings, and sounds recorded at known locations within the scene, we formulate NVAS as learning the transformation from reference recordings to sound recorded at target viewpoints. At inference time, given any pre-acquired reference recordings of the scene as input and a query for an arbitrary target position, we expect the model to output binaural audio that is acoustically consistent with the query viewpoint and the content from reference recordings. + +To address the limitations of existing NVAS techniques in reference microphone placement, we develop a multi-modal approach for optimal reference location sampling. Our ap- + +proach can adapt to diverse, large complex scenes using a Visual-Acoustic Binding (VAB) module. VAB learns acoustically relevant features (VAB embeddings) by pretraining an encoder to align visual features from panoramic RGB-D captures with acoustic features from echo responses. Using VAB embeddings from numerous candidate reference locations, we employ a sampler that automatically identifies representative spots by finding spatial regions with similar embeddings. These spatial regions represent areas with similar acoustic properties and are usually free from obstacles that significantly hinder sound propagation. This approach optimizes reference microphone placement within a limited reference budget which enhances the overall performance. + +Furthermore, to manage varying numbers of references – which typically depend on scene size and budget – and their unequal contributions to the final audio, we introduce an adaptive reference integration module. This module models sequences via a transformer, using VAB features to reweight reference inputs and viewpoints. These reweighted inputs then serve as conditions for the final binaural audio renderer. The resulting conditions are both scene-adapted and content-invariant, enabling effective handling of various reference microphone configurations and audio content. + +To demonstrate the effectiveness of our approach, we benchmark the proposed model in both real-world setting and in a challenging simulated dataset derived from Matterport3D [3] scenes with SoundSpaces [4]. SoundVista outperforms existing methods [1, 5, 6, 8, 21] on scenes with multiple varieties of sound sources, handling varying numbers up to ten sound sources. + +# 2. Related Work + +Acoustic Scene Synthesis. Traditional methods focus on estimating the room impulse response (RIR) to recreate spatial audio. This is achieved by convolving the sound from each emitter with the RIR corresponding to the emitter-listener pair and summing the results [9, 20, 23, 25, 35-37, 39, 43]. These RIR-based techniques often require detailed source information, such as a clean signal of each source and its precise location, which are typically unknown in real-world scenarios, making them challenging to implement. Alternative approaches use scene-descriptor images for direct spatial sound synthesis based on reference sound input, such as visual acoustic matching [5, 7] or visual-guided audio spatialization [12, 13, 51, 57]. However, these methods may be inaccurate when viewpoints change continuously. + +Novel-view Acoustic Synthesis. Novel-view acoustic synthesis techniques address limitations in acoustic scene synthesis by learning transfer functions from reference sounds to target viewpoints [6, 8, 21, 38, 46]. Methods such as Few-shotRIR [25] and BEE [8] use multiple references but lack optimal reference sampling strategies, relying on heuristic settings such as a single close reference [6, 21, 38, 46], + +![](images/616cb5b354904c43c02e740986cc445de139992565e0cdfa10327101b660d2f5.jpg) +Figure 2. Details of the SoundVista Pipeline: (a) The reference location sampler selects optimal reference locations leveraging embeddings from visual-acoustic binding (VAB). (b) The reference integration transformer uses VAB embeddings to derive contribution weights for each reference. (c) Reweighting by contribution weights adjusts and integrates reference recording channels and pose conditioning for precise sound synthesis. (d) The spatial audio renderer converts reweighted channels and conditions to binaural sound at the target viewpoint. + +random sampling [25], or fixed references [8], which limit adaptation to diverse scene layouts. In contrast, our method optimizes reference location sampling via visual-acoustic binding and derives adaptive reference contribution weights, enhancing sound synthesis accuracy and generalization. + +Acoustic-Visual Learning. Techniques such as CLIP [11, 22, 24, 31, 52] have successfully aligned visual and linguistic inputs to learn matched deep semantic representations. These advances led to recent works to leverage complementary nature of audio and visual data for acoustic related tasks. For example, AV-RIR [37] binds RIR with panoramic images to improve the late reverberation modeling of RIR. Advances in multi-modal visual understanding have facilitated sound source localization [15, 18, 27, 44], audio-visual speech enhancement and source separation [10, 26, 30, 45, 54, 56], as well as acoustic scene reconstruction [5, 6, 8, 25, 37, 41]. These works underscore the strong link between visual and acoustic modalities. Inspired by these approaches, we developed a visual encoder that aligns features from RGB-D visual data with acoustic features from echo responses. The alignment enables us to infer acoustic properties using visual data only. As a result, our method does not require knowing acoustic parameters or RIRs during inference. It is also easily adaptable to various scene and acoustic environments. + +# 3. Method + +# 3.1. Problem Formulation + +Given $N$ pairs of co-located microphones and cameras, we aim to strategically place them in the acoustic scene. Let $\mathbf{L}_i = (x_i,y_i,z_i)\in \mathbb{R}^3$ denote the location of the devices and $\theta_{i}\in \mathbb{R}^{2}$ denote their orientations. The microphones cap + +ture reference audio recordings $\mathbf{A}_i$ , and the cameras capture panoramic RGB-D views $\mathbf{V}_i$ . These "reference captures" encapsulate the ambient sound and visual properties, providing the basis for reconstructing binaural ambient sound $\tilde{\mathbf{A}}_k$ for a target listener at arbitrary location $\mathbf{L}_k = (x_k, y_k, z_k) \in \mathbb{R}^3$ with orientation $\theta_k \in \mathbb{R}^2$ . The reference recording can be of any type, but we use ambisonic microphones as they are ideal for capturing sound fields from multiple directions and simplify binaural signal extraction. In contrast to existing methods [1], we avoid enumerating sound sources. Instead, we learn a transfer function $\mathcal{F}$ from reference captures to the target as: + +$$ +\mathcal {F}: \left(\left\{\mathbf {A} _ {i}, g _ {i}, r _ {k i}, \boldsymbol {\theta} _ {i} \right\} _ {i = 1} ^ {N}, g _ {k}, \boldsymbol {\theta} _ {k}\right) \mapsto \tilde {\mathbf {A}} _ {k}, \tag {1} +$$ + +$$ +g _ {i} = \phi (\mathbf {V} _ {i}), g _ {k} = \phi (\mathbf {V} _ {k}), \text {a n d} r _ {k i} = \mathbf {L} _ {k} - \mathbf {L} _ {i}, +$$ + +where $g_{i}$ and $g_{k}$ are Visual-Acoustic Binding (VAB) embeddings extracted with a pre-trained visual encoder $\phi (\cdot)$ , designed to align visual with acoustic representations. $r_{ki}$ is relative vector introduced to avoid overfitting on absolute locations and to handle diverse scene coordinate systems. + +We parameterize $\mathcal{F}$ using neural networks that incorporate a transformer network and spatial audio renderer module. The transformer generates an adaptive mask to weight the contribution of each reference recording based on the target viewpoint and the VAB embeddings associated with reference viewpoints. We call this the "Reference Integration Transformer". The spatial audio renderer then processes the reweighted audio channels along with conditional information from the target view features to produce the final binaural sound. We provide an overview of these modules in Figure 2, with details explained in the following sections. + +# 3.2. Visual-Acoustic Binding (VAB) + +Obtaining acoustic properties of real-world scenes is challenging [9]. We propose to leverage RGB-D data, which can provide rich information linked to acoustic properties of a scene. For example, depth data provides information about obstacles and room geometry, while pixel-level textures reveal material differences and detailed obstacle information. + +The goal is to infer the acoustic properties of a scene using only the visual data. To achieve this, we first collected extensive paired panoramic RGB-D and acoustic echo responses data by navigating walkable areas in the SoundSpaces simulator [4]. We then trained a neural network, which we refer as "Visual-Acoustic Binding (VAB) module", to predict the acoustic representations from the visual features. + +Acoustic Representation. In acoustics, an impulse response is a function of time and the positions of the emitter and the listener [47]. It describes how sound propagates through a medium and interacts with the environment [32, 33]. To simplify data collection and infer local acoustic characteristic, we focus on echo impulse responses [25], where the emitter and target locations are the same. We extracted reverberation time (aka $\mathrm{RT}_{60}$ parameter) from echo response to use as acoustic representation. $\mathrm{RT}_{60}$ measures how long it takes for the acoustic energy to decay by $60\mathrm{dB}$ and can reveal information about room geometry, obstacles and reflection. Figure 2(a) shows an example $\mathrm{RT}_{60}$ map of a scene. Significant value changes in $\mathrm{RT}_{60}$ map can indicate major obstacles or surface variations, highlighting key acoustic regions. + +Visual Representation. Following prior works [6, 37], we use panoramic RGB-D captures as inputs and extract visual features from ResNet-18 [14]. We train the VAB module to predict $\mathrm{RT}_{60}$ . Therefore, effective binding visual representation with acoustic representation. + +# 3.3. Reference Location Sampler + +While having more microphones and cameras is ideal for this task, in practice resources are often limited. To maximize performance and make the best use of available resources, we propose a strategic approach for placement of reference microphones throughout a given scene. + +We argue that ideal microphone placements align with the acoustic partitions of the scene-areas with unique acoustic properties and free from obstacles such as walls. To identify these partitions, we capture panoramic RGB-D images from all candidate locations. One can use novel-view synthesis [49] to render these images; without needing actual photographs at each spot. We then extract VAB embeddings from each capture. VAB embeddings which correspond to scene acoustic parameters serve as strong cues for identifying acoustic partitions, such as distinguishing regions separated by obstacles. To enhance reliability, we combine the extracted VAB embeddings with positional information, + +allowing them to complement each other. Using these embeddings, we perform data clustering of the candidate locations and take the center of each cluster as a reference location. This is illustrated in Figure 2(a). + +# 3.4. Reference Integration Transformer + +We want our model to work effectively across diverse scenes. This requires the model to adapt to varying numbers of reference audio and visual inputs. Logically, larger scenes would benefit from more microphones and cameras, while smaller spaces can operate efficiently with fewer resources. Moreover, since the task (Section 3.1) is to transfer reference recordings to target sounds at specific viewpoints, distant microphones, or those in non-adjacent acoustic partitions usually contribute significantly less to the process than closer ones. However, weighting based solely on distance is insufficient due to potential obstacles such as walls and objects. + +To address these issues, we propose a transformer network. We treat each reference input as part of a sequence, allowing us to manage varying numbers of reference inputs. To derive the weights for their unequal contributions to the final audio, we exploit VAB embeddings extracted from visual captures at the reference and target viewpoints. + +We formed the queries input by combining the VAB embedding at the target location $g_{k}$ , with a learnable latent query embedding $\mathbf{e}$ . The query is initialized from a normal distribution and optimized during training, to adjust the contribution of reference microphones based on their relative location vector to the target location. Let $g_{k}^{e} = [g_{k}||\mathbf{e}] \in \mathbb{R}^{\mathrm{C}}$ to denote the queries. The keys and values inputs are created by combining the VAB embedding of each reference with its relative vector $r_{ki}$ . Let $g_{i}^{r} = [g_{i}||r_{ki}] \in \mathbb{R}^{\mathrm{C}}$ to denote the keys and values of $i^{\text{th}}$ reference for the attention operation. The attention weight $a_{ki}$ can be calculated as follows: + +$$ +a _ {k i} = \frac {g _ {k} ^ {e} \cdot g _ {i} ^ {r} \top}{\sqrt {\mathsf {C}}}, \quad i = \{1, 2, \dots , N \}. \tag {2} +$$ + +The attention weight $a_{ki}$ indicates contribution of the reference $i$ to the target $k$ normalized by the Softmax function over $N$ references. Higher weights indicate greater influence. + +# 3.5. Spatial Audio Renderer + +The Spatial Audio Renderer takes the reference recordings and conditional information to generate binaural sound at the target viewpoint. + +Reweighting Reference Recordings. We apply the Short-Time Fourier Transform (STFT) to the reference recordings to extract their spectrograms. Then, we reweight the channels of each reference recording using the attention weights $a_{ki}$ from Section 3.4 to integrate them, encoding the result into a latent space as the input for the renderer. + +Reweighting Condition Inputs. The binaural sound is derived based on the target's position and pose. To capture the + +different aspects of the binaural effects, we decouple the conditioning into global and local components. This accounts for the cause of spatialization effects at a given target point, which may primarily vary from the distance to the sound source or remain invariant to head orientation. The global condition $c_{g}$ determines how the target viewpoint's relative position to reference locations influences the binaural sound. The local condition $c_{l}$ accounts more for the effects of head orientation. They are defined as follows: + +$$ +c _ {g} = \sum_ {i = 1} ^ {N} a _ {k i} \cdot \psi_ {1} \left(g _ {i} \| r _ {k i}\right), \tag {3} +$$ + +$$ +c _ {l} = \left(\sum_ {i = 1} ^ {N} a _ {k i} \cdot \psi_ {2} \left(r _ {k i} \| q _ {k i}\right)\right) + \sigma \left(\theta_ {k}\right). \tag {4} +$$ + +Here $q_{ki}$ is the relative orientation quaternion for target $k$ relative to reference $i$ ; $\psi_1$ and $\psi_2$ are MLP layers projecting concatenated inputs to a latent condition embedding; and $\sigma(\theta_k)$ denotes the target's rotational features from $\theta_k$ . + +**Renderer.** The Audio Renderer network is designed as a stacked U-Net [40, 42], depicted in Figure 2(d). It features down-sampling and up-sampling blocks that incorporate input conditions at each layer. To preserve quality and content details from the reference input, skip connections are employed at multiple resolutions. After processing the integrated reference recordings, we apply the inverse STFT to convert the output spectrogram into a binaural waveform. + +Training from scratch in complex environments with challenging audio content can hinder the renderer to adapt to head orientation while producing high-quality audio. To address this, we pretrain the binauralization capability of our renderer by fixing the target location and varying head rotations and audio sources. After pre-training, the renderer effectively understands binauralization across different orientations, enhancing the accuracy of spatial effect modeling. + +# 3.6. Loss + +We design the loss function as a weighted combination of three components, each regulating different aspects of the model: 1) Waveform Loss measures the mean squared error between the predicted and target waveforms, ensuring precision of the predicted waveform amplitude and phases. 2) Binaural Interaural Level Difference (ILD) Loss [16] focuses on the energy difference between binaural audio channels to ensure accurate spatial effect prediction. 3) Multi-resolution Spectrogram Magnitude Loss [53] ensures that the predicted audio matches the target spectrogram magnitude across multiple resolutions. It includes two terms: the first compares log-scaled magnitudes, and the second, the Scaled Spectrogram Magnitude Loss, calculates magnitude distance over the target scale, addressing high variance in audio magnitude. We found that resolution with large FFT and small hop sizes can benefit modeling ambient noise, such as air vibrations. + +# 4. Experiments + +# 4.1. Benchmarks and Metrics + +N2S Benchmark. Captured in a real room of 8.5 by 6 meters dimensions with two internal rooms, the N2S benchmark utilizes microphone arrays with 32 microphones for 32-channel recordings. Visual captures are rendered by the NeRF-based 3D NVS model, VR-NeRF [49], after it has been trained on visual data collected by navigating the room. 11 static microphone arrays are uniformly distributed to provide reference ambient recordings, while 6 mobile arrays cover 557 locations with 15 seconds of recording per location. Motion capture system, OptiTrack, records 6DoF tracking data for both reference and target microphones. $35\%$ of these locations with 20 orientations are selected for training. + +Soundspace-Ambient Matterport3D. We build this benchmark based on Matterport3D scenes, comprising 39 complex training scenes and 23 unseen scenes. It is challenging as each scene can include 2 to 10 sound sources, selected from 128 different sounds, from fan noise to speech. Reference locations are spaced every 5 meters on average, with simulated second-order ambisonics RIRs to render reference sounds by Soundspaces [4]. $85\%$ of walkable locations are used for training with azimuth angles of 0, 90, 180, and 270. + +Metrics. We evaluate sound synthesis performance using four key metrics (the lower, the better): 1) L1 distance of STFT spectrogram (STFT) [13] for left and right channels; 2) Magnitude Spectrogram Distance (MAG) [13], measuring the closeness of reconstructed audio to the groundtruth; 3) Energy Envelope Error (ENV) [13], assessing the Euclidean distance between energy envelopes of groundtruth and predicted audio channels; 4) Left-Right Energy Ratio Error (LRE) [6], evaluating binaural effect accuracy by calculating energy ratio difference between left and right channels. + +# 4.2. Comparison with Baselines + +We compare our method with the following baseline approaches: (1) Nearest GT: Binaural sound from the nearest reference microphone aligned with the target orientation. (2) Interp GT: Employs binaural sound at the target orientation from the four nearest reference microphones, interpolating based on distance to the target location. (3) AV-NeRF [21]: A NeRF-based system that synthesizes binaural audio from a given camera pose and RGB-D renderings. We adapted it to use recordings, poses, and visual context from references while preserving its core model components, enabling fair comparisons in dynamic scenes. (4) Few-shotRIR [25]: A transformer-based method that infers RIRs from a sparse set of observed images and echoes. Adapted to replace the reference and target impulse responses with the ambient sounds at the corresponding viewpoints. (5) VAM [5]: Matches the acoustics of input audio with a target image. (6) BEE [8]: + +Table 1. Results Comparison on Soundspace-Ambient Benchmark: Average metrics for 10,189 samples in 39 seen scenes with novel target locations and sources, and 6,534 samples in 23 unseen scenes. Sampling strategies: location only, vis+location (our sampler), and w/o VAB (ours without VAB pretraining). SoundVista Ref Num $k$ : only references with top- $k$ contribution weights are used for fair comparison. + +
MethodSamplingRef NumSeen ScenesUnseen Scenes
STFT ↓MAG ↓ENV ↓LRE ↓STFT ↓MAG ↓ENV ↓LRE ↓
nearest GTlocation14.4480.3510.1541.5964.0340.3530.1551.617
nearest GTw/o VAB14.4140.3410.1511.5764.6800.3470.1531.537
nearest GTvis+location13.9160.3360.1461.5723.8350.3440.1511.557
interp GTlocation43.9220.3260.1441.5843.4100.3270.1451.587
interp GTw/o VAB43.6600.3190.1421.5703.7660.3200.1421.531
interp GTvis+location43.1790.3130.1371.5593.4150.3210.1411.531
AV-NeRF [21]vis+location19.4240.4260.1951.9229.3210.4280.1961.979
VAM [5]vis+location15.2240.4200.1781.9024.9360.4360.1821.977
ViGAS [6]vis+location13.7400.3610.1542.0403.4380.3710.1572.051
SoundVista (Ours)vis+location12.5260.2910.1321.4082.6760.3090.1401.386
BEE [8]vis+location44.0980.3650.1622.0835.6350.3960.1782.131
SoundVista (Ours)vis+location42.4440.2890.1301.3902.5170.3050.1371.371
Few-shotRIR [25]vis+locationall5.9370.4590.2131.8925.4570.4710.2151.960
SoundVista w/o vislocationall3.2280.3060.1411.4252.8900.3120.1421.439
SoundVista (Ours)vis+locationall2.4420.2890.1301.3902.5140.3050.1371.372
+ +Table 2. Testing Results Comparison on N2S Benchmark. In this real-world scene, SoundVista with visual modality largely boosts the performance accuracy, especially on the binaural effect (LRE). + +
MethodSTFT ↓MAG ↓ENV ↓LRE ↓
Nearest DSP2.4200.2120.1361.447
Interp DSP1.6590.2030.1421.383
AV-NeRF [21]2.1940.1870.1190.840
Few-shotRIR [25]1.7650.1990.1340.909
VAM [5]1.9720.1900.1190.916
BEE [8]1.4710.2000.1410.995
ViGAS [6]1.2010.1850.1190.873
SoundVista w/o vis1.2420.1850.1180.894
SoundVista (Ours)1.0730.1770.1130.776
+ +A generalizable rendering pipeline that reconstructs the binaural audio at an arbitrary listener location using inputs from sparse reference audio-visual samples in the scene. (7) ViGAS [6]: Transforms sound to the target viewpoint given the observed audio and visual captures at the reference viewpoint. Specifically, for AV-NeRF, VAM, and ViGAS, which need a single reference, we use the nearest reference microphone. BEE is adapted to use the nearest 4 microphones. Deep-learning baselines requiring visual inputs utilize panoramic RGB-D images and the same visual encoder as our model for consistency. + +Our results in Table 1 and Table 2 demonstrate that our method, SoundVista, consistently surpasses baselines across diverse novel scenes and real scenarios. Despite the challenge of obtaining binaural groundtruth (GT) for arbitrary target orientations, SoundVista significantly reduces errors across all metrics compared with nearest GT and interp GT. On the challenging Soundspace-Ambient benchmark, with diverse sound sources and complex layouts, most deep-learning baselines underperform compared to non-learning methods due to the need for robust conditioning models. + +For a fair comparison, we evaluate SoundVista using the top 1 and 4 reference microphones by contribution weight, respectively. Compared to learning-based methods with 1 reference, SoundVista outperforms the best baseline ViGAS by $32.5\%$ (STFT), $19.4\%$ (MAG), $14.3\%$ (ENV), and $31\%$ (LRE), demonstrating the strength of our conditioned audio renderer. The nearest microphone may miss critical audio content, while using the top 4 references improves performance, matching that of all microphones since most references are not required due to their sparse distribution. Since contribution weights are independent of audio content, fewer than 4 references can be selected per target location by precomputing contribution weights once per scene. + +While $Few-shotRIR$ and $BEE$ use audio content to integrate references, the diverse sound distribution makes this challenging. During testing, content often falls outside the training distribution, degrading performance. Thus, more references aren't always better; for example, $Few-shotRIR$ did not gain an increase in synthesis accuracy and it was even reduced compared to baselines using only one reference. + +The deep-learning baselines perform better on the N2S benchmark, which features a single static scene. Compared to these deep-learning baselines, our SoundVista still significantly outperforms on accuracy of energy (ENV) and binaural effect (LRE) by $5\%$ and $7.6\%$ , respectively. + +# 4.3. Qualitative Results Comparison + +Figure 3 visualizes our results for qualitative comparison with several competitive sound synthesis baselines: ViGAS, BEE, and Few-shotRIR. The first row displays loudness maps, while the second row shows generated binaural waveforms with corresponding LRE error displayed in the third row. ViGAS relies on the nearest microphone, leading to discrete loudness maps in complex layouts and unreliable results + +![](images/75fa673d8d6d7d5b358f14cb724594368076c405ae0d9bd1071cfe2b986128b8.jpg) +Figure 3. Comparison of Qualitative Results: First row: Loudness Map (black triangle: sources; purple triangle: target viewpoint). Second row: Reconstructed binaural waveform at target viewpoint (red: prediction; green: GT). Third row: LRE (lower for better binaural effect). + +![](images/198ee6a5e5efbd9a81f3aa7240b2cd7592cf067f6ff7f8aab5dbdced42457704.jpg) + +when obstacles are present, as shown in the unseen results. Additionally, ViGAS struggles with accurate binaural effects shown from the LRE results. BEE produces discontinuous loudness maps, deviating from the ground truth. Both BEE and Few-shotRIR have difficulty generating high-fidelity waveforms close to the groundtruth. Their reliance on incorporating audio content into the renderer's conditioning can bias the renderer leading to confusion and reduced quality when encountering content outside the training distribution. This is particularly common in our complex task setting with diverse audio sources. In contrast, SoundVista, closely matches the GT in loudness maps and binaural waveforms for both seen and unseen scenes, excelling in high-quality novel-view ambient sound synthesis. + +# 4.4. Ablation Study + +We conducted ablation studies on the Soundspace-Ambient benchmark to verify the contribution of key components. + +Visual-Acoustic Binding (VAB). In the design of VAB, we optimized the integration of visual features that correlate with acoustic parameters. Table 3 shows the accuracy of $\mathrm{RT}_{60}$ value predictions in novel scenes without finetuning (w/o finetune) and with few-shot finetuning at reference locations (w/finetune). We tested various modality inputs: location only, panoramic rgb, depth, and rgb+depth. Depth alone significantly enhances binding, reducing error by over $50\%$ compared to use location only. With finetuning, rgb+depth achieves superior results. Given the ability to capture references in scenes, we select rgb+depth as our visual input. + +Reference Sampler via VAB. The Reference Sampler via Visual-Acoustic Binding (VAB) (vis+location) significantly enhances the accuracy of non-learning baselines such as nearest GT and interp GT, as shown in Table 1. This method improves predictions of sound magnitude (MAG, ENV) and spectrogram phases (STFT). Without pretrained VAB (w/o VAB), these benefits diminish, especially in seen scenes. Results in Table 1, 2, and 4 further highlight the critical role of + +Table 3. $\mathbf{RT}_{60}$ Prediction Results on Matterport3D. w/ finetune involves few-shot finetuning on unseen scenes. err: distance between predicted and GT $\mathrm{RT}_{60}$ , error ratio: scaled err over GT. + +
Modalityw/o finetunew/ finetune
err ↓err ratio ↓err ↓err ratio ↓
location0.1700.4110.0670.199
rgb0.1170.2360.0440.126
depth0.0820.1730.0380.109
rgb+depth0.0870.1850.0360.103
+ +Table 4. Ablations on the SoundSpace-Ambient benchmark. + +
MethodSTFT ↓MAG ↓ENV ↓LRE ↓
full model2.4420.2890.1301.390
w/o vis3.2280.3060.1411.425
w/o reweight2.9080.2950.1341.391
w/o render_pretrain2.5820.3100.1351.877
w/o decoup_cond3.0170.3000.1381.433
full loss2.4420.2890.1301.390
w/o ILD_loss2.4620.2920.1311.420
w/o mag_scale2.4240.3010.1361.422
w/o wav_loss3.4220.2910.1431.369
+ +visual input, as its absence increases errors across all metrics, underscoring the importance of VAB with visual information for adapting to diverse scene layouts. + +Figure 4 (left) illustrates the STFT error curve relative to reference density. Our model is trained with a reference density of 1.0, and tested on novel Soundspace-Ambient scenes with varying densities. Compared to Ours - w/o vis and interp - w/o vis (interp GT with location-only sampler), Ours - w/ vis outperforms others, with the performance gap widening as density decreases, demonstrating robustness of ours equipped with VAB vs. reference density variance. + +Reference Integration. To evaluate the contribution of the adaptive reference integration, we conducted experiments with the w/o reweight variant, which uses only the nearest reference microphone along with its corresponding visual and pose information to calculate the condition for the audio renderer. As shown in Table 4, the results of w/o reweight compared to full SoundVista model reveal a noticeable degra + +![](images/c703f44ed5855cd266012abd0f0191e9c6812998ff18178544cffdee41b39ab5.jpg) +Figure 4. STFT Error Curve for different reference densities (left) and training location data ratios (right). + +![](images/a28dfea29a199b2eb5e4f80c31a1aebe037947c3d4de2d0fcd1b0bd769b06a1e.jpg) + +dation in performance, particularly in the STFT error, which increases by $19\%$ . This indicates that the adaptive Reference Integration Transformer and reweighting modules effectively integrate references to enhance sound synthesis accuracy, as the nearest microphone may miss critical content required for the target sound synthesis. + +Conditioned Spatial Audio Renderer. In the w/o decoup_cond variant, we replaced the decoupled global and local conditions from reweighting in Section 3.5 with only the global condition plus target rotational features. This absence of local conditions significantly increases STFT error by $23.5\%$ , MAG error by $3.8\%$ , and ENV error by $6.2\%$ . + +As detailed in Section 3.5, we pretrain the renderer with binauralization capability to help the model learn the complex task of orientation-sensitive novel view sound synthesis. The w/o render_pretrain variant shows a $35\%$ increase in LRE error, underscoring the importance of renderer pretraining for achieving a more accurate binaural effect. + +Loss. The ablations in Table 4 highlight the contributions of different loss components. The Waveform Loss (wav_loss) constrains the energy envelope and regulates phase learning. Removing it significantly reduces STFT and ENV accuracy, increasing the STFT error by $29\%$ . The Scaled Spectrogram Magnitude Loss (mag_scale) calculates magnitude loss over the target magnitude scale, addressing high variance in audio content and mitigating scale differences. Removing mag_scale will increase MAG error largely. + +Robustness w.r.t. Training Data Ratio. To assess the impact of training location data ratio on performance, we trained two variants, Ours - w/o vis and Ours - w/ vis, using different proportions of training locations on the Soundspace-Ambient benchmark. Figure 4 (right) shows that STFT error increases as the data ratio decreases. When the data ratio is below $10\%$ , Ours - w/o vis underperforms compared to the non-learning baseline interp - w/o vis, while Ours - w/ vis maintains superior performance. Additionally, Ours - w/ vis benefits more from increased training locations, with STFT error decreasing rapidly as the ratio nears $100\%$ , widening the performance gap with Ours - w/o vis and interp - w/o vis. + +![](images/7e8414d5fee898c7db8d2922422fbe0a8e9439733164ceac4b62769d684d1a1a.jpg) +Figure 5. Visualization Analysis. First three columns: Clustering results for loc only, Vis - w/o VAB (ours without VAB pretraining), and Vis - w/ VAB (ours). Colors indicate cluster regions. VAB helps clustering to align better with room partitions. Last two columns: Attention weights visualization for references, with colored stars (size proportional to weights) and a purple triangle for the target. + +# 4.5. Visualization Analysis + +Figure 5 visualizes the clustering results of Reference Sampler via VAB, comparing loc only, Vis - w/o VAB, and Vis - w/ VAB. Different colors represent cluster regions. The Reference Sampler Vis - w/ VAB accurately segments clusters aligning with actual room partitions, proving to be more reliable. In contrast, other methods may incorrectly cluster non-adjacent rooms together which cannot share the same acoustic environment, highlighted by red-circles. + +After identifying reference locations, we visualize the attention weights from the Reference Integration Transformer in Figure 5 (column 4-5 from left). Colored stars denote sampled references, sized by contribution weight, with the target marked by a purple triangle. Unlike Distance-only weights, which may incorrectly prioritize closer references despite obstacles, our model, Ref Attn - w/ VAB, trained with the adaptive Reference Integration Transformer, effectively highlights reliable references in similar acoustic environments. + +# 5. Conclusion + +We introduce SoundVista, a novel system designed to synthesize ambient sound from arbitrary scenes at novel viewpoints. Our approach introduces a visual-acoustic binding module to effectively learn visual embeddings linked with local acoustic properties. This enables our system to optimize the placement of reference microphones and derive adaptive weights for each microphone's contribution, conditioned on the target viewpoint and visual captures, thereby, facilitating the final conditioned sound synthesis. SoundVista adapts to diverse room layouts, microphone configurations, and unseen environments, rendering high-quality binaural ambient sound without requiring prior constraints or detailed knowledge of sound sources. Our results, validated on both publicly available data and real-world settings, demonstrate state-of-the-art sound synthesis accuracy and generalization. + +# References + +[1] Byeongjoo Ahn, Karren Yang, Brian Hamilton, Jonathan Sheaffer, Anurag Ranjan, Miguel Sarabia, Oncel Tuzel, and Jen-Hao Rick Chang. Novel-view acoustic synthesis from 3d reconstructed rooms. arXiv preprint arXiv:2310.15130, 2023. 1, 2, 3 +[2] Federico Borra, Israel Dejene Gebru, and Dejan Markovic. Soundfield reconstruction in reverberant environments using higher-order microphones and impulse response measurements. In ICASSP 2019-2019 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 281-285. IEEE, 2019. 2 +[3] Angel Chang, Angela Dai, Thomas Funkhouser, Maciej Halber, Matthias Niessner, Manolis Savva, Shuran Song, Andy Zeng, and Yinda Zhang. Matterport3d: Learning from rgb-d data in indoor environments. International Conference on 3D Vision (3DV), 2017. 2 +[4] Changan Chen, Unnat Jain, Carl Schissler, Sebastia Vicenc Amengual Gari, Ziad Al-Halah, Vamsi Krishna Ithapu, Philip Robinson, and Kristen Grauman. Soundspaces: Audiovisual navigation in 3d environments. In European Conference on Computer Vision, pages 17-36. Springer, 2020. 2, 4, 5 +[5] Changan Chen, Ruohan Gao, Paul Calamia, and Kristen Grauman. Visual acoustic matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18858-18868, 2022. 2, 3, 5, 6 +[6] Changan Chen, Alexander Richard, Roman Shapovalov, Vamsi Krishna Ithapu, Natalia Neverova, Kristen Grauman, and Andrea Vedaldi. Novel-view acoustic synthesis. arXiv preprint arXiv:2301.08730, 2023. 1, 2, 3, 4, 5, 6 +[7] Changan Chen, Wei Sun, David Harwath, and Kristen Grauman. Learning audio-visual dereverberation. In ICASSP, 2023. 2 +[8] Mingfei Chen, Kun Su, and Eli Shlizerman. Be everywhere—hear everything (bee): Audio scene reconstruction by sparse audio-visual samples. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7853–7862, 2023. 2, 3, 5, 6 +[9] Ziyang Chen, Israel D Gebru, Christian Richardt, Anurag Kumar, William Laney, Andrew Owens, and Alexander Richard. Real acoustic fields: An audio-visual room acoustics dataset and benchmark. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21886-21896, 2024. 2, 4 +[10] Ariel Ephrat, Inbar Mosseri, Oran Lang, Tali Dekel, Kevin Wilson, Avinatan Hassidim, William T Freeman, and Michael Rubinstein. Looking to listen at the cocktail party: A speaker-independent audio-visual model for speech separation. arXiv preprint arXiv:1804.03619, 2018. 3 +[11] Han Fang, Pengfei Xiong, Luhui Xu, and Yu Chen. Clip2video: Mastering video-text retrieval via image clip. arXiv preprint arXiv:2106.11097, 2021. 3 +[12] Ruohan Gao and Kristen Grauman. 2.5d visual sound. In CVPR, 2019. 2 +[13] Ruohan Gao and Kristen Grauman. 2.5 d visual sound. In Pro + +ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 324-333, 2019. 2, 5 +[14] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 4 +[15] Di Hu, Rui Qian, Minyue Jiang, Xiao Tan, Shilei Wen, Errui Ding, Weiyao Lin, and Dejing Dou. Discriminative sounding objects localization via self-supervised audiovisual matching. Advances in Neural Information Processing Systems, 33: 10077-10087, 2020. 3 +[16] Christopher A. Ick, Gordon Wichern, Yoshiki Masuyama, François Germain, and Jonathan Le Roux. Spatially-aware losses for enhanced neural acoustic fields. In Audio Imagination: NeurIPS 2024 Workshop AI-Driven Speech, Music, and Sound Generation, 2024. 5 +[17] Hyeonjoong Jang, Andreas Meuleman, Dahiyun Kang, Donggun Kim, Christian Richardt, and Min H Kim. Egocentric scene reconstruction from an omnidirectional video. ACM Transactions on Graphics (TOG), 41(4):1-12, 2022. 1 +[18] Hao Jiang, Calvin Murdock, and Vamsi Krishna Ithapu. Ego-centric deep multi-channel audio-visual active speaker localization. arXiv preprint arXiv:2201.01928, 2022. 3 +[19] Shoichi Koyama, Tomoya Nishida, Keisuke Kimura, Takumi Abe, Natsuki Ueno, and Jesper Brunnström. Meshrir: A dataset of room impulse responses on meshed grid points for evaluating sound field analysis and synthesis methods. In 2021 IEEE Workshop on Applications of Signal Processing to Audio and Acoustics (WASPAA), pages 1-5, 2021. 2 +[20] Susan Liang, Chao Huang, Yapeng Tian, Anurag Kumar, and Chenliang Xu. Neural acoustic context field: Rendering realistic room impulse response with neural fields. ArXiv, 2023. 2 +[21] Susan Liang, Chao Huang, Yapeng Tian, Anurag Kumar, and Chenliang Xu. Av-nerf: Learning neural fields for real-world audio-visual scene synthesis. In Conference on Neural Information Processing Systems (NeurIPS), 2023. 2, 5, 6 +[22] Ziyi Lin, Shijie Geng, Renrui Zhang, Peng Gao, Gerard De Melo, Xiaogang Wang, Jifeng Dai, Yu Qiao, and Hongsheng Li. Frozen clip models are efficient video learners. In European Conference on Computer Vision, pages 388-404. Springer, 2022. 3 +[23] Andrew Luo, Yilun Du, Michael J Tarr, Joshua B Tenenbaum, Antonio Torralba, and Chuang Gan. Learning neural acoustic fields. arXiv preprint arXiv., 2021. 2 +[24] Huaishao Luo, Lei Ji, Ming Zhong, Yang Chen, Wen Lei, Nan Duan, and Tianrui Li. CLIP4Clip: An empirical study of clip for end to end video clip retrieval. arXiv preprint arXiv:2104.08860, 2021. 3 +[25] Sagnik Majumder, Changan Chen, Ziad Al-Halah, and Kristen Grauman. Few-shot audio-visual learning of environment acoustics, 2022. 2, 3, 4, 5, 6 +[26] Daniel Michelsanti, Zheng-Hua Tan, Shi-Xiong Zhang, Yong Xu, Meng Yu, Dong Yu, and Jesper Jensen. An overview of deep-learning-based audio-visual speech enhancement and separation. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 2021. 3 + +[27] Shentong Mo and Yapeng Tian. Audio-visual grouping network for sound localization from mixtures. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10565–10574, 2023. 3 +[28] Bartlomiej Mróz, Marek Kabaciński, Tomasz Ciotucha, Andrzej Ruminski, and Tomasz Žernicki. Production of six-degrees-of-freedom (6dof) navigable audio using 30 ambisonic microphones. In 2021 Immersive and 3D Audio: from Architecture to Automotive (I3DA), pages 1-5. IEEE, 2021. 2 +[29] Orhun Olgun, Ege Erdem, and Hüseyin Hachhabiboglu. Sound field interpolation via sparse plane wave decomposition for 6 dof immersive audio. In 2023 Immersive and 3D Audio: from Architecture to Automotive (I3DA), pages 1-10. IEEE, 2023. 2 +[30] Andrew Owens and Alexei A Efros. Audio-visual scene analysis with self-supervised multisensory features. In Proceedings of the European Conference on Computer Vision (ECCV), pages 631–648, 2018. 3 +[31] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 3 +[32] Nikunj Raghuvanshi and John Snyder. Parametric wave field coding for precomputed sound propagation. ACM Transactions on Graphics (TOG), 33(4):1-11, 2014. 4 +[33] Nikunj Raghuvanshi and John Snyder. Parametric directional coding for precomputed sound propagation. ACM Transactions on Graphics (TOG), 37(4):1-14, 2018. 4 +[34] Santhosh K Ramakrishnan, Aaron Gokaslan, Erik Wijmans, Oleksandr Maksymets, Alex Clegg, John Turner, Eric Undersander, Wojciech Galuba, Andrew Westbury, Angel X Chang, et al. Habitat-matterport 3d dataset (hm3d): 1000 large-scale 3d environments for embodied ai. arXiv preprint arXiv:2109.08238, 2021. 1 +[35] Anton Ratnarajah, Zhenyu Tang, and Dinesh Manocha. *Irgan: Room impulse response generator for far-field speech recognition.* arXiv preprint arXiv:2010.13219, 2020. 2 +[36] Anton Ratnarajah, Shi-Xiong Zhang, Meng Yu, Zhenyu Tang, Dinesh Manocha, and Dong Yu. Fast-rir: Fast neural diffuse room impulse response generator. arXiv preprint arXiv:2110.04057, 2021. +[37] Anton Ratnarajah, Sreyan Ghosh, Sonal Kumar, Purva Chiniya, and Dinesh Manocha. Av-rir: Audio-visual room impulse response estimation. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 27154–27165, 2023. 2, 3, 4 +[38] Alexander Richard, Dejan Markovic, Israel D. Gebru, Steven Krenn, Gladstone Alexander Butler, Fernando Torre, and Yaser Sheikh. Neural synthesis of binaural speech from mono audio. In International Conference on Learning Representations, 2021. 2 +[39] Alexander Richard, Peter Dodds, and Vamsi Krishna Ithapu. Deep impulse responses: Estimating and parameterizing filters with deep networks. arXiv preprint arXiv:2202.03416, 2022. 2 + +[40] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention-MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, pages 234-241. Springer, 2015. 5 +[41] Nikhil Singh, Jeff Mentch, Jerry Ng, Matthew Beveridge, and Iddo Drori. Image2reverb: Cross-modal reverb impulse response synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 286-295, 2021. 3 +[42] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In International Conference on Learning Representations, 2021. 5 +[43] Kun Su, Mingfei Chen, and Eli Shlizerman. Inras: Implicit neural representation for audio scenes. Advances in Neural Information Processing Systems, 35:8144-8158, 2022. 2 +[44] Ruijie Tao, Zexu Pan, Rohan Kumar Das, Xinyuan Qian, Mike Zheng Shou, and Haizhou Li. Is someone speaking? exploring long-term temporal features for audio-visual active speaker detection. In Proceedings of the 29th ACM international conference on multimedia, pages 3927-3935, 2021. 3 +[45] Yapeng Tian, Di Hu, and Chenliang Xu. Cyclic co-learning of sounding object visual grounding and sound separation. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2744-2753, 2021. 3 +[46] Aäron van den Oord, Sander Dieleman, Heiga Zen, Karen Simonyan, Oriol Vinyals, Alexander Graves, Nal Kalchbrenner, Andrew Senior, and Koray Kavukcuoglu. Wavenet: A generative model for raw audio. In Arxiv, 2016. 2 +[47] Tor Erik Vigran. Building acoustics. CRC Press, 2014. 4 +[48] Xiuchao Wu, Jiamin Xu, Zihan Zhu, Hujun Bao, Qixing Huang, James Tompkin, and Weiwei Xu. Scalable neural indoor scene rendering. ACM transactions on graphics, 41 (4), 2022. 1 +[49] Linning Xu, Vasu Agrawal, William Laney, Tony Garcia, Aayush Bansal, Changil Kim, Samuel Rota Bulò, Lorenzo Porzi, Peter Kontschieder, Aljaž Božić, Dahua Lin, Michael Zollhöfer, and Christian Richardt. VR-NeRF: High-fidelity virtualized walkable spaces. In SIGGRAPH Asia Conference Proceedings, 2023. 1, 4, 5 +[50] Shaoheng Xu, Jihui Aimee Zhang, Thushara D. Abhayapala, Amy Bastine, Wei-Ting Lai, and Prasanga N. Samarasinghe. Sparse sound field representation using complex orthogonal matching pursuit. In ICASSP 2024 - 2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1336-1340, 2024. 2 +[51] Xudong Xu, Hang Zhou, Ziwei Liu, Bo Dai, Xiaogang Wang, and Dahua Lin. Visually informed binaural audio generation without binaural audiios. In CVPR, 2021. 2 +[52] Hongwei Xue, Yuchong Sun, Bei Liu, Jianlong Fu, Ruihua Song, Houqiang Li, and Jiebo Luo. Clip-vip: Adapting pretrained image-text model to video-language representation alignment. arXiv preprint arXiv:2209.06430, 2022. 3 + +[53] Ryuichi Yamamoto, Eunwoo Song, and Jae-Min Kim. Parallel wavegan: A fast waveform generation model based on generative adversarial networks with multi-resolution spectrogram. In ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 6199-6203. IEEE, 2020. 5 +[54] Yuxin Ye, Wenming Yang, and Yapeng Tian. Lavss: Location-guided audio-visual spatial audio separation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 5508–5519, 2024. 3 +[55] Zechen Zhang, Nikunj Raghuvanshi, John Snyder, and Steve Marschner. Ambient sound propagation. ACM Transactions on Graphics (TOG), 37(6):1-10, 2018. 2 +[56] Hang Zhou, Ziwei Liu, Xudong Xu, Ping Luo, and Xiaogang Wang. Vision-infused deep audio inpainting. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 283-292, 2019. 3 +[57] Hang Zhou, Xudong Xu, Dahua Lin, Xiaogang Wang, and Ziwei Liu. Sep-stereo: Visually guided stereophonic audio generation by associating source separation. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XII 16, pages 52-69. Springer, 2020. 2 \ No newline at end of file diff --git a/2025/SoundVista_ Novel-View Ambient Sound Synthesis via Visual-Acoustic Binding/images.zip b/2025/SoundVista_ Novel-View Ambient Sound Synthesis via Visual-Acoustic Binding/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..6ddfdd8a8c01006d6252515c8db6624c91577389 --- /dev/null +++ b/2025/SoundVista_ Novel-View Ambient Sound Synthesis via Visual-Acoustic Binding/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9002f346d31127dd50368ad7f572a1703ad38f8addc26894296f989b6e322740 +size 592398 diff --git a/2025/SoundVista_ Novel-View Ambient Sound Synthesis via Visual-Acoustic Binding/layout.json b/2025/SoundVista_ Novel-View Ambient Sound Synthesis via Visual-Acoustic Binding/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..fb604c457b4a26cefa0965c936a8d5b47620979b --- /dev/null +++ b/2025/SoundVista_ Novel-View Ambient Sound Synthesis via Visual-Acoustic Binding/layout.json @@ -0,0 +1,7887 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 61, + 103, + 549, + 122 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 103, + 549, + 122 + ], + "spans": [ + { + "bbox": [ + 61, + 103, + 549, + 122 + ], + "type": "text", + "content": "SoundVista: Novel-View Ambient Sound Synthesis via Visual-Acoustic Binding" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 77, + 141, + 530, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 141, + 530, + 193 + ], + "spans": [ + { + "bbox": [ + 77, + 141, + 530, + 193 + ], + "type": "text", + "content": "Mingfei Chen" + }, + { + "bbox": [ + 77, + 141, + 530, + 193 + ], + "type": "inline_equation", + "content": "^{1,2*}" + }, + { + "bbox": [ + 77, + 141, + 530, + 193 + ], + "type": "text", + "content": " Israel D. Gebru" + }, + { + "bbox": [ + 77, + 141, + 530, + 193 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 77, + 141, + 530, + 193 + ], + "type": "text", + "content": " Ishwarya Ananthabhotla" + }, + { + "bbox": [ + 77, + 141, + 530, + 193 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 77, + 141, + 530, + 193 + ], + "type": "text", + "content": " Christian Richardt" + }, + { + "bbox": [ + 77, + 141, + 530, + 193 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 77, + 141, + 530, + 193 + ], + "type": "text", + "content": " \nDejan Markovic" + }, + { + "bbox": [ + 77, + 141, + 530, + 193 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 77, + 141, + 530, + 193 + ], + "type": "text", + "content": " Jake Sandakly" + }, + { + "bbox": [ + 77, + 141, + 530, + 193 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 77, + 141, + 530, + 193 + ], + "type": "text", + "content": " Steven Krenn" + }, + { + "bbox": [ + 77, + 141, + 530, + 193 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 77, + 141, + 530, + 193 + ], + "type": "text", + "content": " Todd Keebler" + }, + { + "bbox": [ + 77, + 141, + 530, + 193 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 77, + 141, + 530, + 193 + ], + "type": "text", + "content": " Eli Shlizerman" + }, + { + "bbox": [ + 77, + 141, + 530, + 193 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 77, + 141, + 530, + 193 + ], + "type": "text", + "content": " Alexander Richard" + }, + { + "bbox": [ + 77, + 141, + 530, + 193 + ], + "type": "inline_equation", + "content": "^{2}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 63, + 198, + 544, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 198, + 544, + 215 + ], + "spans": [ + { + "bbox": [ + 63, + 198, + 544, + 215 + ], + "type": "text", + "content": "1University of Washington 2Codec Avatars Lab, Pittsburgh, Meta 3Reality Labs Research, Meta" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 73, + 241, + 537, + 376 + ], + "blocks": [ + { + "bbox": [ + 73, + 241, + 537, + 376 + ], + "lines": [ + { + "bbox": [ + 73, + 241, + 537, + 376 + ], + "spans": [ + { + "bbox": [ + 73, + 241, + 537, + 376 + ], + "type": "image", + "image_path": "8b2ef0b4b30c3b3170c3f44021a83871628edb3d04c44413ad1bccf479f2500f.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 385, + 555, + 419 + ], + "lines": [ + { + "bbox": [ + 54, + 385, + 555, + 419 + ], + "spans": [ + { + "bbox": [ + 54, + 385, + 555, + 419 + ], + "type": "text", + "content": "Figure 1. SoundVista: a novel method that synthesizes binaural ambient sound for arbitrary scenes from novel viewpoints. Our method leverages pre-acquired audio recordings and visual data captured from sparsely distributed reference points and synthesizes binaural audio consistent with the target 3D position and pose." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 151, + 429, + 200, + 441 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 429, + 200, + 441 + ], + "spans": [ + { + "bbox": [ + 151, + 429, + 200, + 441 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 454, + 297, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 454, + 297, + 694 + ], + "spans": [ + { + "bbox": [ + 54, + 454, + 297, + 694 + ], + "type": "text", + "content": "We introduce SoundVista, a method to generate the ambient sound of an arbitrary scene at novel viewpoints. Given a pre-acquired recording of the scene from sparsely distributed microphones, SoundVista can synthesize the sound of that scene from an unseen target viewpoint. The method learns the underlying acoustic transfer function that relates the signals acquired at the distributed microphones to the signal at the target viewpoint, using a limited number of known recordings. Unlike existing works, our method does not require constraints or prior knowledge of sound source details. Moreover, our method efficiently adapts to diverse room layouts, reference microphone configurations and unseen environments. To enable this, we introduce a visual-acoustic binding module that learns visual embeddings linked with local acoustic properties from panoramic RGB and depth data. We first leverage these embeddings to optimize the placement of reference microphones in any given scene. During synthesis, we leverage multiple embeddings extracted from reference locations to get adaptive weights for their contribution, conditioned on target viewpoint. We benchmark the task" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 430, + 555, + 454 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 430, + 555, + 454 + ], + "spans": [ + { + "bbox": [ + 313, + 430, + 555, + 454 + ], + "type": "text", + "content": "on both publicly available data and real-world settings. We demonstrate significant improvements over existing methods." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 464, + 395, + 475 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 464, + 395, + 475 + ], + "spans": [ + { + "bbox": [ + 314, + 464, + 395, + 475 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 312, + 484, + 556, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 484, + 556, + 604 + ], + "spans": [ + { + "bbox": [ + 312, + 484, + 556, + 604 + ], + "type": "text", + "content": "Recent advances in 3D reconstruction and Novel-View Synthesis (NVS) have significantly enhanced our ability to create photorealistic visual models of real-world scenes [17, 34, 48, 49]. These developments have paved the way for applications in 3D virtual tours, experience recreation, and spatial media. However, the audio counterpart – Novel-View Acoustic Synthesis (NVAS) [1, 6] – has been under-explored compared to its visual counterpart and has not received the rigorous attention required to generate acoustic virtual scenes that match their visual surroundings." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 605, + 556, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 556, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 556, + 665 + ], + "type": "text", + "content": "To address this gap, we introduce SoundVista: a novel method that creates a truly immersive acoustic experience. SoundVista can generate realistic and spatially accurate binaural audio at novel viewpoints in arbitrary scenes given a sparse set of sample recordings from different viewpoint." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 665, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 556, + 713 + ], + "type": "text", + "content": "Unlike novel-view synthesis for visual 3D scenes, which are mostly static, NVAS faces significant challenges due to the dynamic nature of real-world acoustic environments. Ambient sound – the overall acoustic state describing all sounds" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 69, + 702, + 208, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 702, + 208, + 713 + ], + "spans": [ + { + "bbox": [ + 69, + 702, + 208, + 713 + ], + "type": "text", + "content": "* Work done during an internship at Meta." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "8331" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 216 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 216 + ], + "type": "text", + "content": "in a scene [55] – can encompass multiple time-varying, nonstationary sound sources without limitations on the type and number of sound sources. In an ideal scenario, with complete information about the ambient sound (i.e. clean signals of individual sound sources and their locations over time), and a fully reconstructed 3D visual model, one could synthesize the ambient sound at any location using standard acoustic renderers with room impulse responses (RIRs) [1, 4, 9, 19, 43]. In real-world scenes, however, we lack detailed information about the sound sources that comprise the ambient sound. Accurately determining their clean source content and emitter positions is a challenging problem." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 217, + 296, + 302 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 217, + 296, + 302 + ], + "spans": [ + { + "bbox": [ + 55, + 217, + 296, + 302 + ], + "type": "text", + "content": "NVAS also faces challenges in balancing data sampling costs and synthesis quality. Current methods in the acoustic community use grids of microphones, and techniques ranging from interpolation [28, 29] to complex signal processing [2, 50] to synthesize audio at novel locations. However, these approaches do not scale well to large spaces, and efficient sampling and data utilization remain open questions." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 303, + 297, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 303, + 297, + 471 + ], + "spans": [ + { + "bbox": [ + 55, + 303, + 297, + 471 + ], + "type": "text", + "content": "Recent deep learning based NVAS methods [5, 6, 8, 21, 25] utilize multimodal data, including visual inputs, to transfer sparse reference sounds to binaural sound at target viewpoints. These methods often simplify the task by focusing on re-synthesizing primary sounds such as speech and music, with a limited number of sound sources (typically 1-2). They often ignore other sounds that contribute to the natural characteristics of the ambient sound scene. While feasible with few references, they struggle to adapt to diverse, large scenes with complex layout and acoustic environment. Furthermore, the lack of an optimal reference location sampling strategy hinders adaptation, leading to a loss of acoustic details and making it difficult to synthesize high-quality sound that has accurate binaural effects for target viewpoints." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 472, + 297, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 472, + 297, + 675 + ], + "spans": [ + { + "bbox": [ + 55, + 472, + 297, + 675 + ], + "type": "text", + "content": "In this paper, we propose a novel approach that avoids the challenge of obtaining granular information about individual sound sources and addresses some of the shortcomings of existing NVAS techniques. Our method relies on sparsely distributed \"reference microphones\" to capture acoustic snapshots, taking their recorded audio signals as \"reference recordings\". This allows us to get a holistic representation of the acoustic environment from references recordings. Given many examples of reference recordings, and sounds recorded at known locations within the scene, we formulate NVAS as learning the transformation from reference recordings to sound recorded at target viewpoints. At inference time, given any pre-acquired reference recordings of the scene as input and a query for an arbitrary target position, we expect the model to output binaural audio that is acoustically consistent with the query viewpoint and the content from reference recordings." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 677, + 297, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 297, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 297, + 715 + ], + "type": "text", + "content": "To address the limitations of existing NVAS techniques in reference microphone placement, we develop a multi-modal approach for optimal reference location sampling. Our ap-" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 555, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 228 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 228 + ], + "type": "text", + "content": "proach can adapt to diverse, large complex scenes using a Visual-Acoustic Binding (VAB) module. VAB learns acoustically relevant features (VAB embeddings) by pretraining an encoder to align visual features from panoramic RGB-D captures with acoustic features from echo responses. Using VAB embeddings from numerous candidate reference locations, we employ a sampler that automatically identifies representative spots by finding spatial regions with similar embeddings. These spatial regions represent areas with similar acoustic properties and are usually free from obstacles that significantly hinder sound propagation. This approach optimizes reference microphone placement within a limited reference budget which enhances the overall performance." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 228, + 556, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 228, + 556, + 346 + ], + "spans": [ + { + "bbox": [ + 313, + 228, + 556, + 346 + ], + "type": "text", + "content": "Furthermore, to manage varying numbers of references – which typically depend on scene size and budget – and their unequal contributions to the final audio, we introduce an adaptive reference integration module. This module models sequences via a transformer, using VAB features to reweight reference inputs and viewpoints. These reweighted inputs then serve as conditions for the final binaural audio renderer. The resulting conditions are both scene-adapted and content-invariant, enabling effective handling of various reference microphone configurations and audio content." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 347, + 556, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 347, + 556, + 431 + ], + "spans": [ + { + "bbox": [ + 313, + 347, + 556, + 431 + ], + "type": "text", + "content": "To demonstrate the effectiveness of our approach, we benchmark the proposed model in both real-world setting and in a challenging simulated dataset derived from Matterport3D [3] scenes with SoundSpaces [4]. SoundVista outperforms existing methods [1, 5, 6, 8, 21] on scenes with multiple varieties of sound sources, handling varying numbers up to ten sound sources." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 439, + 402, + 451 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 439, + 402, + 451 + ], + "spans": [ + { + "bbox": [ + 314, + 439, + 402, + 451 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 460, + 556, + 627 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 460, + 556, + 627 + ], + "spans": [ + { + "bbox": [ + 313, + 460, + 556, + 627 + ], + "type": "text", + "content": "Acoustic Scene Synthesis. Traditional methods focus on estimating the room impulse response (RIR) to recreate spatial audio. This is achieved by convolving the sound from each emitter with the RIR corresponding to the emitter-listener pair and summing the results [9, 20, 23, 25, 35-37, 39, 43]. These RIR-based techniques often require detailed source information, such as a clean signal of each source and its precise location, which are typically unknown in real-world scenarios, making them challenging to implement. Alternative approaches use scene-descriptor images for direct spatial sound synthesis based on reference sound input, such as visual acoustic matching [5, 7] or visual-guided audio spatialization [12, 13, 51, 57]. However, these methods may be inaccurate when viewpoints change continuously." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 630, + 557, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 557, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 557, + 715 + ], + "type": "text", + "content": "Novel-view Acoustic Synthesis. Novel-view acoustic synthesis techniques address limitations in acoustic scene synthesis by learning transfer functions from reference sounds to target viewpoints [6, 8, 21, 38, 46]. Methods such as Few-shotRIR [25] and BEE [8] use multiple references but lack optimal reference sampling strategies, relying on heuristic settings such as a single close reference [6, 21, 38, 46]," + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "8332" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 62, + 72, + 542, + 272 + ], + "blocks": [ + { + "bbox": [ + 62, + 72, + 542, + 272 + ], + "lines": [ + { + "bbox": [ + 62, + 72, + 542, + 272 + ], + "spans": [ + { + "bbox": [ + 62, + 72, + 542, + 272 + ], + "type": "image", + "image_path": "616cb5b354904c43c02e740986cc445de139992565e0cdfa10327101b660d2f5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 274, + 555, + 319 + ], + "lines": [ + { + "bbox": [ + 55, + 274, + 555, + 319 + ], + "spans": [ + { + "bbox": [ + 55, + 274, + 555, + 319 + ], + "type": "text", + "content": "Figure 2. Details of the SoundVista Pipeline: (a) The reference location sampler selects optimal reference locations leveraging embeddings from visual-acoustic binding (VAB). (b) The reference integration transformer uses VAB embeddings to derive contribution weights for each reference. (c) Reweighting by contribution weights adjusts and integrates reference recording channels and pose conditioning for precise sound synthesis. (d) The spatial audio renderer converts reweighted channels and conditions to binaural sound at the target viewpoint." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 329, + 296, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 329, + 296, + 390 + ], + "spans": [ + { + "bbox": [ + 54, + 329, + 296, + 390 + ], + "type": "text", + "content": "random sampling [25], or fixed references [8], which limit adaptation to diverse scene layouts. In contrast, our method optimizes reference location sampling via visual-acoustic binding and derives adaptive reference contribution weights, enhancing sound synthesis accuracy and generalization." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 392, + 296, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 392, + 296, + 620 + ], + "spans": [ + { + "bbox": [ + 56, + 392, + 296, + 620 + ], + "type": "text", + "content": "Acoustic-Visual Learning. Techniques such as CLIP [11, 22, 24, 31, 52] have successfully aligned visual and linguistic inputs to learn matched deep semantic representations. These advances led to recent works to leverage complementary nature of audio and visual data for acoustic related tasks. For example, AV-RIR [37] binds RIR with panoramic images to improve the late reverberation modeling of RIR. Advances in multi-modal visual understanding have facilitated sound source localization [15, 18, 27, 44], audio-visual speech enhancement and source separation [10, 26, 30, 45, 54, 56], as well as acoustic scene reconstruction [5, 6, 8, 25, 37, 41]. These works underscore the strong link between visual and acoustic modalities. Inspired by these approaches, we developed a visual encoder that aligns features from RGB-D visual data with acoustic features from echo responses. The alignment enables us to infer acoustic properties using visual data only. As a result, our method does not require knowing acoustic parameters or RIRs during inference. It is also easily adaptable to various scene and acoustic environments." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 628, + 111, + 640 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 628, + 111, + 640 + ], + "spans": [ + { + "bbox": [ + 55, + 628, + 111, + 640 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 647, + 180, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 647, + 180, + 659 + ], + "spans": [ + { + "bbox": [ + 55, + 647, + 180, + 659 + ], + "type": "text", + "content": "3.1. Problem Formulation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 665, + 297, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 297, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 297, + 715 + ], + "type": "text", + "content": "Given " + }, + { + "bbox": [ + 55, + 665, + 297, + 715 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 665, + 297, + 715 + ], + "type": "text", + "content": " pairs of co-located microphones and cameras, we aim to strategically place them in the acoustic scene. Let " + }, + { + "bbox": [ + 55, + 665, + 297, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{L}_i = (x_i,y_i,z_i)\\in \\mathbb{R}^3" + }, + { + "bbox": [ + 55, + 665, + 297, + 715 + ], + "type": "text", + "content": " denote the location of the devices and " + }, + { + "bbox": [ + 55, + 665, + 297, + 715 + ], + "type": "inline_equation", + "content": "\\theta_{i}\\in \\mathbb{R}^{2}" + }, + { + "bbox": [ + 55, + 665, + 297, + 715 + ], + "type": "text", + "content": " denote their orientations. The microphones cap" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 330, + 555, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 330, + 555, + 473 + ], + "spans": [ + { + "bbox": [ + 313, + 330, + 555, + 473 + ], + "type": "text", + "content": "ture reference audio recordings " + }, + { + "bbox": [ + 313, + 330, + 555, + 473 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_i" + }, + { + "bbox": [ + 313, + 330, + 555, + 473 + ], + "type": "text", + "content": ", and the cameras capture panoramic RGB-D views " + }, + { + "bbox": [ + 313, + 330, + 555, + 473 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_i" + }, + { + "bbox": [ + 313, + 330, + 555, + 473 + ], + "type": "text", + "content": ". These \"reference captures\" encapsulate the ambient sound and visual properties, providing the basis for reconstructing binaural ambient sound " + }, + { + "bbox": [ + 313, + 330, + 555, + 473 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{A}}_k" + }, + { + "bbox": [ + 313, + 330, + 555, + 473 + ], + "type": "text", + "content": " for a target listener at arbitrary location " + }, + { + "bbox": [ + 313, + 330, + 555, + 473 + ], + "type": "inline_equation", + "content": "\\mathbf{L}_k = (x_k, y_k, z_k) \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 313, + 330, + 555, + 473 + ], + "type": "text", + "content": " with orientation " + }, + { + "bbox": [ + 313, + 330, + 555, + 473 + ], + "type": "inline_equation", + "content": "\\theta_k \\in \\mathbb{R}^2" + }, + { + "bbox": [ + 313, + 330, + 555, + 473 + ], + "type": "text", + "content": ". The reference recording can be of any type, but we use ambisonic microphones as they are ideal for capturing sound fields from multiple directions and simplify binaural signal extraction. In contrast to existing methods [1], we avoid enumerating sound sources. Instead, we learn a transfer function " + }, + { + "bbox": [ + 313, + 330, + 555, + 473 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 313, + 330, + 555, + 473 + ], + "type": "text", + "content": " from reference captures to the target as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 338, + 476, + 555, + 497 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 338, + 476, + 555, + 497 + ], + "spans": [ + { + "bbox": [ + 338, + 476, + 555, + 497 + ], + "type": "interline_equation", + "content": "\\mathcal {F}: \\left(\\left\\{\\mathbf {A} _ {i}, g _ {i}, r _ {k i}, \\boldsymbol {\\theta} _ {i} \\right\\} _ {i = 1} ^ {N}, g _ {k}, \\boldsymbol {\\theta} _ {k}\\right) \\mapsto \\tilde {\\mathbf {A}} _ {k}, \\tag {1}", + "image_path": "0728d75adbd2e08ca8623c969f79e0a18e34fb78eeca42e844ebc862e47b9d61.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 338, + 499, + 530, + 512 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 338, + 499, + 530, + 512 + ], + "spans": [ + { + "bbox": [ + 338, + 499, + 530, + 512 + ], + "type": "interline_equation", + "content": "g _ {i} = \\phi (\\mathbf {V} _ {i}), g _ {k} = \\phi (\\mathbf {V} _ {k}), \\text {a n d} r _ {k i} = \\mathbf {L} _ {k} - \\mathbf {L} _ {i},", + "image_path": "206877fb7349ea5a5f08925ca85ce6460a68ce4473895fb5883d7286bebdb273.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 521, + 555, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 555, + 582 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 555, + 582 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 521, + 555, + 582 + ], + "type": "inline_equation", + "content": "g_{i}" + }, + { + "bbox": [ + 313, + 521, + 555, + 582 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 521, + 555, + 582 + ], + "type": "inline_equation", + "content": "g_{k}" + }, + { + "bbox": [ + 313, + 521, + 555, + 582 + ], + "type": "text", + "content": " are Visual-Acoustic Binding (VAB) embeddings extracted with a pre-trained visual encoder " + }, + { + "bbox": [ + 313, + 521, + 555, + 582 + ], + "type": "inline_equation", + "content": "\\phi (\\cdot)" + }, + { + "bbox": [ + 313, + 521, + 555, + 582 + ], + "type": "text", + "content": ", designed to align visual with acoustic representations. " + }, + { + "bbox": [ + 313, + 521, + 555, + 582 + ], + "type": "inline_equation", + "content": "r_{ki}" + }, + { + "bbox": [ + 313, + 521, + 555, + 582 + ], + "type": "text", + "content": " is relative vector introduced to avoid overfitting on absolute locations and to handle diverse scene coordinate systems." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 582, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 556, + 714 + ], + "type": "text", + "content": "We parameterize " + }, + { + "bbox": [ + 313, + 582, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 313, + 582, + 556, + 714 + ], + "type": "text", + "content": " using neural networks that incorporate a transformer network and spatial audio renderer module. The transformer generates an adaptive mask to weight the contribution of each reference recording based on the target viewpoint and the VAB embeddings associated with reference viewpoints. We call this the \"Reference Integration Transformer\". The spatial audio renderer then processes the reweighted audio channels along with conditional information from the target view features to produce the final binaural sound. We provide an overview of these modules in Figure 2, with details explained in the following sections." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8333" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 223, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 223, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 223, + 85 + ], + "type": "text", + "content": "3.2. Visual-Acoustic Binding (VAB)" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 90, + 297, + 162 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 90, + 297, + 162 + ], + "spans": [ + { + "bbox": [ + 55, + 90, + 297, + 162 + ], + "type": "text", + "content": "Obtaining acoustic properties of real-world scenes is challenging [9]. We propose to leverage RGB-D data, which can provide rich information linked to acoustic properties of a scene. For example, depth data provides information about obstacles and room geometry, while pixel-level textures reveal material differences and detailed obstacle information." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 163, + 297, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 163, + 297, + 247 + ], + "spans": [ + { + "bbox": [ + 55, + 163, + 297, + 247 + ], + "type": "text", + "content": "The goal is to infer the acoustic properties of a scene using only the visual data. To achieve this, we first collected extensive paired panoramic RGB-D and acoustic echo responses data by navigating walkable areas in the SoundSpaces simulator [4]. We then trained a neural network, which we refer as \"Visual-Acoustic Binding (VAB) module\", to predict the acoustic representations from the visual features." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 251, + 297, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 251, + 297, + 418 + ], + "spans": [ + { + "bbox": [ + 55, + 251, + 297, + 418 + ], + "type": "text", + "content": "Acoustic Representation. In acoustics, an impulse response is a function of time and the positions of the emitter and the listener [47]. It describes how sound propagates through a medium and interacts with the environment [32, 33]. To simplify data collection and infer local acoustic characteristic, we focus on echo impulse responses [25], where the emitter and target locations are the same. We extracted reverberation time (aka " + }, + { + "bbox": [ + 55, + 251, + 297, + 418 + ], + "type": "inline_equation", + "content": "\\mathrm{RT}_{60}" + }, + { + "bbox": [ + 55, + 251, + 297, + 418 + ], + "type": "text", + "content": " parameter) from echo response to use as acoustic representation. " + }, + { + "bbox": [ + 55, + 251, + 297, + 418 + ], + "type": "inline_equation", + "content": "\\mathrm{RT}_{60}" + }, + { + "bbox": [ + 55, + 251, + 297, + 418 + ], + "type": "text", + "content": " measures how long it takes for the acoustic energy to decay by " + }, + { + "bbox": [ + 55, + 251, + 297, + 418 + ], + "type": "inline_equation", + "content": "60\\mathrm{dB}" + }, + { + "bbox": [ + 55, + 251, + 297, + 418 + ], + "type": "text", + "content": " and can reveal information about room geometry, obstacles and reflection. Figure 2(a) shows an example " + }, + { + "bbox": [ + 55, + 251, + 297, + 418 + ], + "type": "inline_equation", + "content": "\\mathrm{RT}_{60}" + }, + { + "bbox": [ + 55, + 251, + 297, + 418 + ], + "type": "text", + "content": " map of a scene. Significant value changes in " + }, + { + "bbox": [ + 55, + 251, + 297, + 418 + ], + "type": "inline_equation", + "content": "\\mathrm{RT}_{60}" + }, + { + "bbox": [ + 55, + 251, + 297, + 418 + ], + "type": "text", + "content": " map can indicate major obstacles or surface variations, highlighting key acoustic regions." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 421, + 296, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 421, + 296, + 482 + ], + "spans": [ + { + "bbox": [ + 55, + 421, + 296, + 482 + ], + "type": "text", + "content": "Visual Representation. Following prior works [6, 37], we use panoramic RGB-D captures as inputs and extract visual features from ResNet-18 [14]. We train the VAB module to predict " + }, + { + "bbox": [ + 55, + 421, + 296, + 482 + ], + "type": "inline_equation", + "content": "\\mathrm{RT}_{60}" + }, + { + "bbox": [ + 55, + 421, + 296, + 482 + ], + "type": "text", + "content": ". Therefore, effective binding visual representation with acoustic representation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 491, + 212, + 505 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 491, + 212, + 505 + ], + "spans": [ + { + "bbox": [ + 55, + 491, + 212, + 505 + ], + "type": "text", + "content": "3.3. Reference Location Sampler" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 509, + 296, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 509, + 296, + 569 + ], + "spans": [ + { + "bbox": [ + 55, + 509, + 296, + 569 + ], + "type": "text", + "content": "While having more microphones and cameras is ideal for this task, in practice resources are often limited. To maximize performance and make the best use of available resources, we propose a strategic approach for placement of reference microphones throughout a given scene." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 570, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 297, + 714 + ], + "type": "text", + "content": "We argue that ideal microphone placements align with the acoustic partitions of the scene-areas with unique acoustic properties and free from obstacles such as walls. To identify these partitions, we capture panoramic RGB-D images from all candidate locations. One can use novel-view synthesis [49] to render these images; without needing actual photographs at each spot. We then extract VAB embeddings from each capture. VAB embeddings which correspond to scene acoustic parameters serve as strong cues for identifying acoustic partitions, such as distinguishing regions separated by obstacles. To enhance reliability, we combine the extracted VAB embeddings with positional information," + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 556, + 121 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 556, + 121 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 556, + 121 + ], + "type": "text", + "content": "allowing them to complement each other. Using these embeddings, we perform data clustering of the candidate locations and take the center of each cluster as a reference location. This is illustrated in Figure 2(a)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 125, + 503, + 137 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 125, + 503, + 137 + ], + "spans": [ + { + "bbox": [ + 313, + 125, + 503, + 137 + ], + "type": "text", + "content": "3.4. Reference Integration Transformer" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 142, + 556, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 142, + 556, + 274 + ], + "spans": [ + { + "bbox": [ + 313, + 142, + 556, + 274 + ], + "type": "text", + "content": "We want our model to work effectively across diverse scenes. This requires the model to adapt to varying numbers of reference audio and visual inputs. Logically, larger scenes would benefit from more microphones and cameras, while smaller spaces can operate efficiently with fewer resources. Moreover, since the task (Section 3.1) is to transfer reference recordings to target sounds at specific viewpoints, distant microphones, or those in non-adjacent acoustic partitions usually contribute significantly less to the process than closer ones. However, weighting based solely on distance is insufficient due to potential obstacles such as walls and objects." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 274, + 556, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 274, + 556, + 346 + ], + "spans": [ + { + "bbox": [ + 313, + 274, + 556, + 346 + ], + "type": "text", + "content": "To address these issues, we propose a transformer network. We treat each reference input as part of a sequence, allowing us to manage varying numbers of reference inputs. To derive the weights for their unequal contributions to the final audio, we exploit VAB embeddings extracted from visual captures at the reference and target viewpoints." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 346, + 556, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 346, + 556, + 477 + ], + "spans": [ + { + "bbox": [ + 313, + 346, + 556, + 477 + ], + "type": "text", + "content": "We formed the queries input by combining the VAB embedding at the target location " + }, + { + "bbox": [ + 313, + 346, + 556, + 477 + ], + "type": "inline_equation", + "content": "g_{k}" + }, + { + "bbox": [ + 313, + 346, + 556, + 477 + ], + "type": "text", + "content": ", with a learnable latent query embedding " + }, + { + "bbox": [ + 313, + 346, + 556, + 477 + ], + "type": "inline_equation", + "content": "\\mathbf{e}" + }, + { + "bbox": [ + 313, + 346, + 556, + 477 + ], + "type": "text", + "content": ". The query is initialized from a normal distribution and optimized during training, to adjust the contribution of reference microphones based on their relative location vector to the target location. Let " + }, + { + "bbox": [ + 313, + 346, + 556, + 477 + ], + "type": "inline_equation", + "content": "g_{k}^{e} = [g_{k}||\\mathbf{e}] \\in \\mathbb{R}^{\\mathrm{C}}" + }, + { + "bbox": [ + 313, + 346, + 556, + 477 + ], + "type": "text", + "content": " to denote the queries. The keys and values inputs are created by combining the VAB embedding of each reference with its relative vector " + }, + { + "bbox": [ + 313, + 346, + 556, + 477 + ], + "type": "inline_equation", + "content": "r_{ki}" + }, + { + "bbox": [ + 313, + 346, + 556, + 477 + ], + "type": "text", + "content": ". Let " + }, + { + "bbox": [ + 313, + 346, + 556, + 477 + ], + "type": "inline_equation", + "content": "g_{i}^{r} = [g_{i}||r_{ki}] \\in \\mathbb{R}^{\\mathrm{C}}" + }, + { + "bbox": [ + 313, + 346, + 556, + 477 + ], + "type": "text", + "content": " to denote the keys and values of " + }, + { + "bbox": [ + 313, + 346, + 556, + 477 + ], + "type": "inline_equation", + "content": "i^{\\text{th}}" + }, + { + "bbox": [ + 313, + 346, + 556, + 477 + ], + "type": "text", + "content": " reference for the attention operation. The attention weight " + }, + { + "bbox": [ + 313, + 346, + 556, + 477 + ], + "type": "inline_equation", + "content": "a_{ki}" + }, + { + "bbox": [ + 313, + 346, + 556, + 477 + ], + "type": "text", + "content": " can be calculated as follows:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 356, + 483, + 555, + 510 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 356, + 483, + 555, + 510 + ], + "spans": [ + { + "bbox": [ + 356, + 483, + 555, + 510 + ], + "type": "interline_equation", + "content": "a _ {k i} = \\frac {g _ {k} ^ {e} \\cdot g _ {i} ^ {r} \\top}{\\sqrt {\\mathsf {C}}}, \\quad i = \\{1, 2, \\dots , N \\}. \\tag {2}", + "image_path": "1f2c08668e0ed99861afae3b4ec8bda6024ebcdd04993fb0fbdc304851595b22.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 515, + 556, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 515, + 556, + 552 + ], + "spans": [ + { + "bbox": [ + 313, + 515, + 556, + 552 + ], + "type": "text", + "content": "The attention weight " + }, + { + "bbox": [ + 313, + 515, + 556, + 552 + ], + "type": "inline_equation", + "content": "a_{ki}" + }, + { + "bbox": [ + 313, + 515, + 556, + 552 + ], + "type": "text", + "content": " indicates contribution of the reference " + }, + { + "bbox": [ + 313, + 515, + 556, + 552 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 515, + 556, + 552 + ], + "type": "text", + "content": " to the target " + }, + { + "bbox": [ + 313, + 515, + 556, + 552 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 515, + 556, + 552 + ], + "type": "text", + "content": " normalized by the Softmax function over " + }, + { + "bbox": [ + 313, + 515, + 556, + 552 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 515, + 556, + 552 + ], + "type": "text", + "content": " references. Higher weights indicate greater influence." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 558, + 448, + 571 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 448, + 571 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 448, + 571 + ], + "type": "text", + "content": "3.5. Spatial Audio Renderer" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 575, + 555, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 575, + 555, + 612 + ], + "spans": [ + { + "bbox": [ + 313, + 575, + 555, + 612 + ], + "type": "text", + "content": "The Spatial Audio Renderer takes the reference recordings and conditional information to generate binaural sound at the target viewpoint." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 614, + 556, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 614, + 556, + 687 + ], + "spans": [ + { + "bbox": [ + 313, + 614, + 556, + 687 + ], + "type": "text", + "content": "Reweighting Reference Recordings. We apply the Short-Time Fourier Transform (STFT) to the reference recordings to extract their spectrograms. Then, we reweight the channels of each reference recording using the attention weights " + }, + { + "bbox": [ + 313, + 614, + 556, + 687 + ], + "type": "inline_equation", + "content": "a_{ki}" + }, + { + "bbox": [ + 313, + 614, + 556, + 687 + ], + "type": "text", + "content": " from Section 3.4 to integrate them, encoding the result into a latent space as the input for the renderer." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 689, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 556, + 715 + ], + "type": "text", + "content": "Reweighting Condition Inputs. The binaural sound is derived based on the target's position and pose. To capture the" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "8334" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 180 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 180 + ], + "type": "text", + "content": "different aspects of the binaural effects, we decouple the conditioning into global and local components. This accounts for the cause of spatialization effects at a given target point, which may primarily vary from the distance to the sound source or remain invariant to head orientation. The global condition " + }, + { + "bbox": [ + 55, + 72, + 297, + 180 + ], + "type": "inline_equation", + "content": "c_{g}" + }, + { + "bbox": [ + 55, + 72, + 297, + 180 + ], + "type": "text", + "content": " determines how the target viewpoint's relative position to reference locations influences the binaural sound. The local condition " + }, + { + "bbox": [ + 55, + 72, + 297, + 180 + ], + "type": "inline_equation", + "content": "c_{l}" + }, + { + "bbox": [ + 55, + 72, + 297, + 180 + ], + "type": "text", + "content": " accounts more for the effects of head orientation. They are defined as follows:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 97, + 185, + 296, + 217 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 185, + 296, + 217 + ], + "spans": [ + { + "bbox": [ + 97, + 185, + 296, + 217 + ], + "type": "interline_equation", + "content": "c _ {g} = \\sum_ {i = 1} ^ {N} a _ {k i} \\cdot \\psi_ {1} \\left(g _ {i} \\| r _ {k i}\\right), \\tag {3}", + "image_path": "d2f8193006f4ccc64c6e6e77c7155a59bb4b6ba19b7e4c9c2a6ebf03f1f09e1f.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 97, + 220, + 296, + 253 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 220, + 296, + 253 + ], + "spans": [ + { + "bbox": [ + 97, + 220, + 296, + 253 + ], + "type": "interline_equation", + "content": "c _ {l} = \\left(\\sum_ {i = 1} ^ {N} a _ {k i} \\cdot \\psi_ {2} \\left(r _ {k i} \\| q _ {k i}\\right)\\right) + \\sigma \\left(\\theta_ {k}\\right). \\tag {4}", + "image_path": "4828052b58a1c568c852eb00c7d1402a3451484d190bcb981dd16fe8c17bd7a9.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 257, + 296, + 305 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 257, + 296, + 305 + ], + "spans": [ + { + "bbox": [ + 55, + 257, + 296, + 305 + ], + "type": "text", + "content": "Here " + }, + { + "bbox": [ + 55, + 257, + 296, + 305 + ], + "type": "inline_equation", + "content": "q_{ki}" + }, + { + "bbox": [ + 55, + 257, + 296, + 305 + ], + "type": "text", + "content": " is the relative orientation quaternion for target " + }, + { + "bbox": [ + 55, + 257, + 296, + 305 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 257, + 296, + 305 + ], + "type": "text", + "content": " relative to reference " + }, + { + "bbox": [ + 55, + 257, + 296, + 305 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 257, + 296, + 305 + ], + "type": "text", + "content": "; " + }, + { + "bbox": [ + 55, + 257, + 296, + 305 + ], + "type": "inline_equation", + "content": "\\psi_1" + }, + { + "bbox": [ + 55, + 257, + 296, + 305 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 257, + 296, + 305 + ], + "type": "inline_equation", + "content": "\\psi_2" + }, + { + "bbox": [ + 55, + 257, + 296, + 305 + ], + "type": "text", + "content": " are MLP layers projecting concatenated inputs to a latent condition embedding; and " + }, + { + "bbox": [ + 55, + 257, + 296, + 305 + ], + "type": "inline_equation", + "content": "\\sigma(\\theta_k)" + }, + { + "bbox": [ + 55, + 257, + 296, + 305 + ], + "type": "text", + "content": " denotes the target's rotational features from " + }, + { + "bbox": [ + 55, + 257, + 296, + 305 + ], + "type": "inline_equation", + "content": "\\theta_k" + }, + { + "bbox": [ + 55, + 257, + 296, + 305 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 308, + 296, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 308, + 296, + 403 + ], + "spans": [ + { + "bbox": [ + 55, + 308, + 296, + 403 + ], + "type": "text", + "content": "**Renderer.** The Audio Renderer network is designed as a stacked U-Net [40, 42], depicted in Figure 2(d). It features down-sampling and up-sampling blocks that incorporate input conditions at each layer. To preserve quality and content details from the reference input, skip connections are employed at multiple resolutions. After processing the integrated reference recordings, we apply the inverse STFT to convert the output spectrogram into a binaural waveform." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 403, + 297, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 403, + 297, + 499 + ], + "spans": [ + { + "bbox": [ + 55, + 403, + 297, + 499 + ], + "type": "text", + "content": "Training from scratch in complex environments with challenging audio content can hinder the renderer to adapt to head orientation while producing high-quality audio. To address this, we pretrain the binauralization capability of our renderer by fixing the target location and varying head rotations and audio sources. After pre-training, the renderer effectively understands binauralization across different orientations, enhancing the accuracy of spatial effect modeling." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 505, + 100, + 516 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 505, + 100, + 516 + ], + "spans": [ + { + "bbox": [ + 55, + 505, + 100, + 516 + ], + "type": "text", + "content": "3.6. Loss" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 522, + 297, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 522, + 297, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 522, + 297, + 713 + ], + "type": "text", + "content": "We design the loss function as a weighted combination of three components, each regulating different aspects of the model: 1) Waveform Loss measures the mean squared error between the predicted and target waveforms, ensuring precision of the predicted waveform amplitude and phases. 2) Binaural Interaural Level Difference (ILD) Loss [16] focuses on the energy difference between binaural audio channels to ensure accurate spatial effect prediction. 3) Multi-resolution Spectrogram Magnitude Loss [53] ensures that the predicted audio matches the target spectrogram magnitude across multiple resolutions. It includes two terms: the first compares log-scaled magnitudes, and the second, the Scaled Spectrogram Magnitude Loss, calculates magnitude distance over the target scale, addressing high variance in audio magnitude. We found that resolution with large FFT and small hop sizes can benefit modeling ambient noise, such as air vibrations." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 72, + 395, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 395, + 85 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 395, + 85 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 92, + 455, + 104 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 92, + 455, + 104 + ], + "spans": [ + { + "bbox": [ + 314, + 92, + 455, + 104 + ], + "type": "text", + "content": "4.1. Benchmarks and Metrics" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 113, + 555, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 113, + 555, + 258 + ], + "spans": [ + { + "bbox": [ + 313, + 113, + 555, + 258 + ], + "type": "text", + "content": "N2S Benchmark. Captured in a real room of 8.5 by 6 meters dimensions with two internal rooms, the N2S benchmark utilizes microphone arrays with 32 microphones for 32-channel recordings. Visual captures are rendered by the NeRF-based 3D NVS model, VR-NeRF [49], after it has been trained on visual data collected by navigating the room. 11 static microphone arrays are uniformly distributed to provide reference ambient recordings, while 6 mobile arrays cover 557 locations with 15 seconds of recording per location. Motion capture system, OptiTrack, records 6DoF tracking data for both reference and target microphones. " + }, + { + "bbox": [ + 313, + 113, + 555, + 258 + ], + "type": "inline_equation", + "content": "35\\%" + }, + { + "bbox": [ + 313, + 113, + 555, + 258 + ], + "type": "text", + "content": " of these locations with 20 orientations are selected for training." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 261, + 556, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 261, + 556, + 369 + ], + "spans": [ + { + "bbox": [ + 313, + 261, + 556, + 369 + ], + "type": "text", + "content": "Soundspace-Ambient Matterport3D. We build this benchmark based on Matterport3D scenes, comprising 39 complex training scenes and 23 unseen scenes. It is challenging as each scene can include 2 to 10 sound sources, selected from 128 different sounds, from fan noise to speech. Reference locations are spaced every 5 meters on average, with simulated second-order ambisonics RIRs to render reference sounds by Soundspaces [4]. " + }, + { + "bbox": [ + 313, + 261, + 556, + 369 + ], + "type": "inline_equation", + "content": "85\\%" + }, + { + "bbox": [ + 313, + 261, + 556, + 369 + ], + "type": "text", + "content": " of walkable locations are used for training with azimuth angles of 0, 90, 180, and 270." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 373, + 556, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 373, + 556, + 494 + ], + "spans": [ + { + "bbox": [ + 313, + 373, + 556, + 494 + ], + "type": "text", + "content": "Metrics. We evaluate sound synthesis performance using four key metrics (the lower, the better): 1) L1 distance of STFT spectrogram (STFT) [13] for left and right channels; 2) Magnitude Spectrogram Distance (MAG) [13], measuring the closeness of reconstructed audio to the groundtruth; 3) Energy Envelope Error (ENV) [13], assessing the Euclidean distance between energy envelopes of groundtruth and predicted audio channels; 4) Left-Right Energy Ratio Error (LRE) [6], evaluating binaural effect accuracy by calculating energy ratio difference between left and right channels." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 503, + 465, + 516 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 503, + 465, + 516 + ], + "spans": [ + { + "bbox": [ + 313, + 503, + 465, + 516 + ], + "type": "text", + "content": "4.2. Comparison with Baselines" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 521, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 556, + 715 + ], + "type": "text", + "content": "We compare our method with the following baseline approaches: (1) Nearest GT: Binaural sound from the nearest reference microphone aligned with the target orientation. (2) Interp GT: Employs binaural sound at the target orientation from the four nearest reference microphones, interpolating based on distance to the target location. (3) AV-NeRF [21]: A NeRF-based system that synthesizes binaural audio from a given camera pose and RGB-D renderings. We adapted it to use recordings, poses, and visual context from references while preserving its core model components, enabling fair comparisons in dynamic scenes. (4) Few-shotRIR [25]: A transformer-based method that infers RIRs from a sparse set of observed images and echoes. Adapted to replace the reference and target impulse responses with the ambient sounds at the corresponding viewpoints. (5) VAM [5]: Matches the acoustics of input audio with a target image. (6) BEE [8]:" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8335" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 105, + 550, + 294 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 103 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 103 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 103 + ], + "type": "text", + "content": "Table 1. Results Comparison on Soundspace-Ambient Benchmark: Average metrics for 10,189 samples in 39 seen scenes with novel target locations and sources, and 6,534 samples in 23 unseen scenes. Sampling strategies: location only, vis+location (our sampler), and w/o VAB (ours without VAB pretraining). SoundVista Ref Num " + }, + { + "bbox": [ + 55, + 70, + 555, + 103 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 70, + 555, + 103 + ], + "type": "text", + "content": " : only references with top- " + }, + { + "bbox": [ + 55, + 70, + 555, + 103 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 70, + 555, + 103 + ], + "type": "text", + "content": " contribution weights are used for fair comparison." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 105, + 550, + 294 + ], + "lines": [ + { + "bbox": [ + 58, + 105, + 550, + 294 + ], + "spans": [ + { + "bbox": [ + 58, + 105, + 550, + 294 + ], + "type": "table", + "html": "
MethodSamplingRef NumSeen ScenesUnseen Scenes
STFT ↓MAG ↓ENV ↓LRE ↓STFT ↓MAG ↓ENV ↓LRE ↓
nearest GTlocation14.4480.3510.1541.5964.0340.3530.1551.617
nearest GTw/o VAB14.4140.3410.1511.5764.6800.3470.1531.537
nearest GTvis+location13.9160.3360.1461.5723.8350.3440.1511.557
interp GTlocation43.9220.3260.1441.5843.4100.3270.1451.587
interp GTw/o VAB43.6600.3190.1421.5703.7660.3200.1421.531
interp GTvis+location43.1790.3130.1371.5593.4150.3210.1411.531
AV-NeRF [21]vis+location19.4240.4260.1951.9229.3210.4280.1961.979
VAM [5]vis+location15.2240.4200.1781.9024.9360.4360.1821.977
ViGAS [6]vis+location13.7400.3610.1542.0403.4380.3710.1572.051
SoundVista (Ours)vis+location12.5260.2910.1321.4082.6760.3090.1401.386
BEE [8]vis+location44.0980.3650.1622.0835.6350.3960.1782.131
SoundVista (Ours)vis+location42.4440.2890.1301.3902.5170.3050.1371.371
Few-shotRIR [25]vis+locationall5.9370.4590.2131.8925.4570.4710.2151.960
SoundVista w/o vislocationall3.2280.3060.1411.4252.8900.3120.1421.439
SoundVista (Ours)vis+locationall2.4420.2890.1301.3902.5140.3050.1371.372
", + "image_path": "a2a4394d4e3363eb3b3d7b115b7bdefa4b913f7a9b2b77ee6c3b79174a5a7f54.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 57, + 338, + 294, + 450 + ], + "blocks": [ + { + "bbox": [ + 55, + 304, + 295, + 337 + ], + "lines": [ + { + "bbox": [ + 55, + 304, + 295, + 337 + ], + "spans": [ + { + "bbox": [ + 55, + 304, + 295, + 337 + ], + "type": "text", + "content": "Table 2. Testing Results Comparison on N2S Benchmark. In this real-world scene, SoundVista with visual modality largely boosts the performance accuracy, especially on the binaural effect (LRE)." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 338, + 294, + 450 + ], + "lines": [ + { + "bbox": [ + 57, + 338, + 294, + 450 + ], + "spans": [ + { + "bbox": [ + 57, + 338, + 294, + 450 + ], + "type": "table", + "html": "
MethodSTFT ↓MAG ↓ENV ↓LRE ↓
Nearest DSP2.4200.2120.1361.447
Interp DSP1.6590.2030.1421.383
AV-NeRF [21]2.1940.1870.1190.840
Few-shotRIR [25]1.7650.1990.1340.909
VAM [5]1.9720.1900.1190.916
BEE [8]1.4710.2000.1410.995
ViGAS [6]1.2010.1850.1190.873
SoundVista w/o vis1.2420.1850.1180.894
SoundVista (Ours)1.0730.1770.1130.776
", + "image_path": "5f423369912946fdb394c5f6a262dac1a2f4662332065dc7ab65a2840dbf7b73.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 460, + 295, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 460, + 295, + 592 + ], + "spans": [ + { + "bbox": [ + 54, + 460, + 295, + 592 + ], + "type": "text", + "content": "A generalizable rendering pipeline that reconstructs the binaural audio at an arbitrary listener location using inputs from sparse reference audio-visual samples in the scene. (7) ViGAS [6]: Transforms sound to the target viewpoint given the observed audio and visual captures at the reference viewpoint. Specifically, for AV-NeRF, VAM, and ViGAS, which need a single reference, we use the nearest reference microphone. BEE is adapted to use the nearest 4 microphones. Deep-learning baselines requiring visual inputs utilize panoramic RGB-D images and the same visual encoder as our model for consistency." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 594, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 594, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 594, + 295, + 714 + ], + "type": "text", + "content": "Our results in Table 1 and Table 2 demonstrate that our method, SoundVista, consistently surpasses baselines across diverse novel scenes and real scenarios. Despite the challenge of obtaining binaural groundtruth (GT) for arbitrary target orientations, SoundVista significantly reduces errors across all metrics compared with nearest GT and interp GT. On the challenging Soundspace-Ambient benchmark, with diverse sound sources and complex layouts, most deep-learning baselines underperform compared to non-learning methods due to the need for robust conditioning models." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 306, + 555, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 306, + 555, + 461 + ], + "spans": [ + { + "bbox": [ + 313, + 306, + 555, + 461 + ], + "type": "text", + "content": "For a fair comparison, we evaluate SoundVista using the top 1 and 4 reference microphones by contribution weight, respectively. Compared to learning-based methods with 1 reference, SoundVista outperforms the best baseline ViGAS by " + }, + { + "bbox": [ + 313, + 306, + 555, + 461 + ], + "type": "inline_equation", + "content": "32.5\\%" + }, + { + "bbox": [ + 313, + 306, + 555, + 461 + ], + "type": "text", + "content": " (STFT), " + }, + { + "bbox": [ + 313, + 306, + 555, + 461 + ], + "type": "inline_equation", + "content": "19.4\\%" + }, + { + "bbox": [ + 313, + 306, + 555, + 461 + ], + "type": "text", + "content": " (MAG), " + }, + { + "bbox": [ + 313, + 306, + 555, + 461 + ], + "type": "inline_equation", + "content": "14.3\\%" + }, + { + "bbox": [ + 313, + 306, + 555, + 461 + ], + "type": "text", + "content": " (ENV), and " + }, + { + "bbox": [ + 313, + 306, + 555, + 461 + ], + "type": "inline_equation", + "content": "31\\%" + }, + { + "bbox": [ + 313, + 306, + 555, + 461 + ], + "type": "text", + "content": " (LRE), demonstrating the strength of our conditioned audio renderer. The nearest microphone may miss critical audio content, while using the top 4 references improves performance, matching that of all microphones since most references are not required due to their sparse distribution. Since contribution weights are independent of audio content, fewer than 4 references can be selected per target location by precomputing contribution weights once per scene." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 461, + 556, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 461, + 556, + 544 + ], + "spans": [ + { + "bbox": [ + 313, + 461, + 556, + 544 + ], + "type": "text", + "content": "While " + }, + { + "bbox": [ + 313, + 461, + 556, + 544 + ], + "type": "inline_equation", + "content": "Few-shotRIR" + }, + { + "bbox": [ + 313, + 461, + 556, + 544 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 461, + 556, + 544 + ], + "type": "inline_equation", + "content": "BEE" + }, + { + "bbox": [ + 313, + 461, + 556, + 544 + ], + "type": "text", + "content": " use audio content to integrate references, the diverse sound distribution makes this challenging. During testing, content often falls outside the training distribution, degrading performance. Thus, more references aren't always better; for example, " + }, + { + "bbox": [ + 313, + 461, + 556, + 544 + ], + "type": "inline_equation", + "content": "Few-shotRIR" + }, + { + "bbox": [ + 313, + 461, + 556, + 544 + ], + "type": "text", + "content": " did not gain an increase in synthesis accuracy and it was even reduced compared to baselines using only one reference." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 545, + 556, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 556, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 556, + 605 + ], + "type": "text", + "content": "The deep-learning baselines perform better on the N2S benchmark, which features a single static scene. Compared to these deep-learning baselines, our SoundVista still significantly outperforms on accuracy of energy (ENV) and binaural effect (LRE) by " + }, + { + "bbox": [ + 313, + 545, + 556, + 605 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 313, + 545, + 556, + 605 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 545, + 556, + 605 + ], + "type": "inline_equation", + "content": "7.6\\%" + }, + { + "bbox": [ + 313, + 545, + 556, + 605 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 612, + 487, + 625 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 612, + 487, + 625 + ], + "spans": [ + { + "bbox": [ + 313, + 612, + 487, + 625 + ], + "type": "text", + "content": "4.3. Qualitative Results Comparison" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 630, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 556, + 713 + ], + "type": "text", + "content": "Figure 3 visualizes our results for qualitative comparison with several competitive sound synthesis baselines: ViGAS, BEE, and Few-shotRIR. The first row displays loudness maps, while the second row shows generated binaural waveforms with corresponding LRE error displayed in the third row. ViGAS relies on the nearest microphone, leading to discrete loudness maps in complex layouts and unreliable results" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "8336" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 72, + 313, + 222 + ], + "blocks": [ + { + "bbox": [ + 60, + 72, + 313, + 222 + ], + "lines": [ + { + "bbox": [ + 60, + 72, + 313, + 222 + ], + "spans": [ + { + "bbox": [ + 60, + 72, + 313, + 222 + ], + "type": "image", + "image_path": "75fa673d8d6d7d5b358f14cb724594368076c405ae0d9bd1071cfe2b986128b8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 223, + 555, + 246 + ], + "lines": [ + { + "bbox": [ + 55, + 223, + 555, + 246 + ], + "spans": [ + { + "bbox": [ + 55, + 223, + 555, + 246 + ], + "type": "text", + "content": "Figure 3. Comparison of Qualitative Results: First row: Loudness Map (black triangle: sources; purple triangle: target viewpoint). Second row: Reconstructed binaural waveform at target viewpoint (red: prediction; green: GT). Third row: LRE (lower for better binaural effect)." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 315, + 72, + 551, + 222 + ], + "blocks": [ + { + "bbox": [ + 315, + 72, + 551, + 222 + ], + "lines": [ + { + "bbox": [ + 315, + 72, + 551, + 222 + ], + "spans": [ + { + "bbox": [ + 315, + 72, + 551, + 222 + ], + "type": "image", + "image_path": "198ee6a5e5efbd9a81f3aa7240b2cd7592cf067f6ff7f8aab5dbdced42457704.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 258, + 297, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 258, + 297, + 426 + ], + "spans": [ + { + "bbox": [ + 54, + 258, + 297, + 426 + ], + "type": "text", + "content": "when obstacles are present, as shown in the unseen results. Additionally, ViGAS struggles with accurate binaural effects shown from the LRE results. BEE produces discontinuous loudness maps, deviating from the ground truth. Both BEE and Few-shotRIR have difficulty generating high-fidelity waveforms close to the groundtruth. Their reliance on incorporating audio content into the renderer's conditioning can bias the renderer leading to confusion and reduced quality when encountering content outside the training distribution. This is particularly common in our complex task setting with diverse audio sources. In contrast, SoundVista, closely matches the GT in loudness maps and binaural waveforms for both seen and unseen scenes, excelling in high-quality novel-view ambient sound synthesis." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 436, + 149, + 449 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 436, + 149, + 449 + ], + "spans": [ + { + "bbox": [ + 55, + 436, + 149, + 449 + ], + "type": "text", + "content": "4.4. Ablation Study" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 454, + 296, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 454, + 296, + 479 + ], + "spans": [ + { + "bbox": [ + 55, + 454, + 296, + 479 + ], + "type": "text", + "content": "We conducted ablation studies on the Soundspace-Ambient benchmark to verify the contribution of key components." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 482, + 296, + 614 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 482, + 296, + 614 + ], + "spans": [ + { + "bbox": [ + 55, + 482, + 296, + 614 + ], + "type": "text", + "content": "Visual-Acoustic Binding (VAB). In the design of VAB, we optimized the integration of visual features that correlate with acoustic parameters. Table 3 shows the accuracy of " + }, + { + "bbox": [ + 55, + 482, + 296, + 614 + ], + "type": "inline_equation", + "content": "\\mathrm{RT}_{60}" + }, + { + "bbox": [ + 55, + 482, + 296, + 614 + ], + "type": "text", + "content": " value predictions in novel scenes without finetuning (w/o finetune) and with few-shot finetuning at reference locations (w/finetune). We tested various modality inputs: location only, panoramic rgb, depth, and rgb+depth. Depth alone significantly enhances binding, reducing error by over " + }, + { + "bbox": [ + 55, + 482, + 296, + 614 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 55, + 482, + 296, + 614 + ], + "type": "text", + "content": " compared to use location only. With finetuning, rgb+depth achieves superior results. Given the ability to capture references in scenes, we select rgb+depth as our visual input." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 618, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 296, + 713 + ], + "type": "text", + "content": "Reference Sampler via VAB. The Reference Sampler via Visual-Acoustic Binding (VAB) (vis+location) significantly enhances the accuracy of non-learning baselines such as nearest GT and interp GT, as shown in Table 1. This method improves predictions of sound magnitude (MAG, ENV) and spectrogram phases (STFT). Without pretrained VAB (w/o VAB), these benefits diminish, especially in seen scenes. Results in Table 1, 2, and 4 further highlight the critical role of" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 328, + 290, + 539, + 359 + ], + "blocks": [ + { + "bbox": [ + 313, + 256, + 555, + 289 + ], + "lines": [ + { + "bbox": [ + 313, + 256, + 555, + 289 + ], + "spans": [ + { + "bbox": [ + 313, + 256, + 555, + 289 + ], + "type": "text", + "content": "Table 3. " + }, + { + "bbox": [ + 313, + 256, + 555, + 289 + ], + "type": "inline_equation", + "content": "\\mathbf{RT}_{60}" + }, + { + "bbox": [ + 313, + 256, + 555, + 289 + ], + "type": "text", + "content": " Prediction Results on Matterport3D. w/ finetune involves few-shot finetuning on unseen scenes. err: distance between predicted and GT " + }, + { + "bbox": [ + 313, + 256, + 555, + 289 + ], + "type": "inline_equation", + "content": "\\mathrm{RT}_{60}" + }, + { + "bbox": [ + 313, + 256, + 555, + 289 + ], + "type": "text", + "content": ", error ratio: scaled err over GT." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 328, + 290, + 539, + 359 + ], + "lines": [ + { + "bbox": [ + 328, + 290, + 539, + 359 + ], + "spans": [ + { + "bbox": [ + 328, + 290, + 539, + 359 + ], + "type": "table", + "html": "
Modalityw/o finetunew/ finetune
err ↓err ratio ↓err ↓err ratio ↓
location0.1700.4110.0670.199
rgb0.1170.2360.0440.126
depth0.0820.1730.0380.109
rgb+depth0.0870.1850.0360.103
", + "image_path": "ce8b73efac2edd617cd034a9d39636ae9b6f563c917eba30fd51014f1cc9c1a7.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 316, + 373, + 553, + 487 + ], + "blocks": [ + { + "bbox": [ + 318, + 361, + 549, + 372 + ], + "lines": [ + { + "bbox": [ + 318, + 361, + 549, + 372 + ], + "spans": [ + { + "bbox": [ + 318, + 361, + 549, + 372 + ], + "type": "text", + "content": "Table 4. Ablations on the SoundSpace-Ambient benchmark." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 316, + 373, + 553, + 487 + ], + "lines": [ + { + "bbox": [ + 316, + 373, + 553, + 487 + ], + "spans": [ + { + "bbox": [ + 316, + 373, + 553, + 487 + ], + "type": "table", + "html": "
MethodSTFT ↓MAG ↓ENV ↓LRE ↓
full model2.4420.2890.1301.390
w/o vis3.2280.3060.1411.425
w/o reweight2.9080.2950.1341.391
w/o render_pretrain2.5820.3100.1351.877
w/o decoup_cond3.0170.3000.1381.433
full loss2.4420.2890.1301.390
w/o ILD_loss2.4620.2920.1311.420
w/o mag_scale2.4240.3010.1361.422
w/o wav_loss3.4220.2910.1431.369
", + "image_path": "1969225b58fb26b43f1d2d510b309f88220fbf27e1648d1a0efb1431759f262f.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 495, + 555, + 530 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 495, + 555, + 530 + ], + "spans": [ + { + "bbox": [ + 313, + 495, + 555, + 530 + ], + "type": "text", + "content": "visual input, as its absence increases errors across all metrics, underscoring the importance of VAB with visual information for adapting to diverse scene layouts." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 531, + 556, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 531, + 556, + 626 + ], + "spans": [ + { + "bbox": [ + 313, + 531, + 556, + 626 + ], + "type": "text", + "content": "Figure 4 (left) illustrates the STFT error curve relative to reference density. Our model is trained with a reference density of 1.0, and tested on novel Soundspace-Ambient scenes with varying densities. Compared to Ours - w/o vis and interp - w/o vis (interp GT with location-only sampler), Ours - w/ vis outperforms others, with the performance gap widening as density decreases, demonstrating robustness of ours equipped with VAB vs. reference density variance." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "content": "Reference Integration. To evaluate the contribution of the adaptive reference integration, we conducted experiments with the w/o reweight variant, which uses only the nearest reference microphone along with its corresponding visual and pose information to calculate the condition for the audio renderer. As shown in Table 4, the results of w/o reweight compared to full SoundVista model reveal a noticeable degra" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8337" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 69, + 173, + 173 + ], + "blocks": [ + { + "bbox": [ + 56, + 69, + 173, + 173 + ], + "lines": [ + { + "bbox": [ + 56, + 69, + 173, + 173 + ], + "spans": [ + { + "bbox": [ + 56, + 69, + 173, + 173 + ], + "type": "image", + "image_path": "c703f44ed5855cd266012abd0f0191e9c6812998ff18178544cffdee41b39ab5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 180, + 295, + 205 + ], + "lines": [ + { + "bbox": [ + 55, + 180, + 295, + 205 + ], + "spans": [ + { + "bbox": [ + 55, + 180, + 295, + 205 + ], + "type": "text", + "content": "Figure 4. STFT Error Curve for different reference densities (left) and training location data ratios (right)." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 173, + 70, + 295, + 172 + ], + "blocks": [ + { + "bbox": [ + 173, + 70, + 295, + 172 + ], + "lines": [ + { + "bbox": [ + 173, + 70, + 295, + 172 + ], + "spans": [ + { + "bbox": [ + 173, + 70, + 295, + 172 + ], + "type": "image", + "image_path": "a28dfea29a199b2eb5e4f80c31a1aebe037947c3d4de2d0fcd1b0bd769b06a1e.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 219, + 296, + 291 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 219, + 296, + 291 + ], + "spans": [ + { + "bbox": [ + 55, + 219, + 296, + 291 + ], + "type": "text", + "content": "dation in performance, particularly in the STFT error, which increases by " + }, + { + "bbox": [ + 55, + 219, + 296, + 291 + ], + "type": "inline_equation", + "content": "19\\%" + }, + { + "bbox": [ + 55, + 219, + 296, + 291 + ], + "type": "text", + "content": ". This indicates that the adaptive Reference Integration Transformer and reweighting modules effectively integrate references to enhance sound synthesis accuracy, as the nearest microphone may miss critical content required for the target sound synthesis." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 298, + 296, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 298, + 296, + 370 + ], + "spans": [ + { + "bbox": [ + 55, + 298, + 296, + 370 + ], + "type": "text", + "content": "Conditioned Spatial Audio Renderer. In the w/o decoup_cond variant, we replaced the decoupled global and local conditions from reweighting in Section 3.5 with only the global condition plus target rotational features. This absence of local conditions significantly increases STFT error by " + }, + { + "bbox": [ + 55, + 298, + 296, + 370 + ], + "type": "inline_equation", + "content": "23.5\\%" + }, + { + "bbox": [ + 55, + 298, + 296, + 370 + ], + "type": "text", + "content": ", MAG error by " + }, + { + "bbox": [ + 55, + 298, + 296, + 370 + ], + "type": "inline_equation", + "content": "3.8\\%" + }, + { + "bbox": [ + 55, + 298, + 296, + 370 + ], + "type": "text", + "content": ", and ENV error by " + }, + { + "bbox": [ + 55, + 298, + 296, + 370 + ], + "type": "inline_equation", + "content": "6.2\\%" + }, + { + "bbox": [ + 55, + 298, + 296, + 370 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 375, + 296, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 375, + 296, + 448 + ], + "spans": [ + { + "bbox": [ + 55, + 375, + 296, + 448 + ], + "type": "text", + "content": "As detailed in Section 3.5, we pretrain the renderer with binauralization capability to help the model learn the complex task of orientation-sensitive novel view sound synthesis. The w/o render_pretrain variant shows a " + }, + { + "bbox": [ + 55, + 375, + 296, + 448 + ], + "type": "inline_equation", + "content": "35\\%" + }, + { + "bbox": [ + 55, + 375, + 296, + 448 + ], + "type": "text", + "content": " increase in LRE error, underscoring the importance of renderer pretraining for achieving a more accurate binaural effect." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 454, + 296, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 454, + 296, + 563 + ], + "spans": [ + { + "bbox": [ + 55, + 454, + 296, + 563 + ], + "type": "text", + "content": "Loss. The ablations in Table 4 highlight the contributions of different loss components. The Waveform Loss (wav_loss) constrains the energy envelope and regulates phase learning. Removing it significantly reduces STFT and ENV accuracy, increasing the STFT error by " + }, + { + "bbox": [ + 55, + 454, + 296, + 563 + ], + "type": "inline_equation", + "content": "29\\%" + }, + { + "bbox": [ + 55, + 454, + 296, + 563 + ], + "type": "text", + "content": ". The Scaled Spectrogram Magnitude Loss (mag_scale) calculates magnitude loss over the target magnitude scale, addressing high variance in audio content and mitigating scale differences. Removing mag_scale will increase MAG error largely." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "content": "Robustness w.r.t. Training Data Ratio. To assess the impact of training location data ratio on performance, we trained two variants, Ours - w/o vis and Ours - w/ vis, using different proportions of training locations on the Soundspace-Ambient benchmark. Figure 4 (right) shows that STFT error increases as the data ratio decreases. When the data ratio is below " + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "content": ", Ours - w/o vis underperforms compared to the non-learning baseline interp - w/o vis, while Ours - w/ vis maintains superior performance. Additionally, Ours - w/ vis benefits more from increased training locations, with STFT error decreasing rapidly as the ratio nears " + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "content": ", widening the performance gap with Ours - w/o vis and interp - w/o vis." + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 319, + 72, + 547, + 201 + ], + "blocks": [ + { + "bbox": [ + 319, + 72, + 547, + 201 + ], + "lines": [ + { + "bbox": [ + 319, + 72, + 547, + 201 + ], + "spans": [ + { + "bbox": [ + 319, + 72, + 547, + 201 + ], + "type": "image", + "image_path": "7e8414d5fee898c7db8d2922422fbe0a8e9439733164ceac4b62769d684d1a1a.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 203, + 555, + 270 + ], + "lines": [ + { + "bbox": [ + 313, + 203, + 555, + 270 + ], + "spans": [ + { + "bbox": [ + 313, + 203, + 555, + 270 + ], + "type": "text", + "content": "Figure 5. Visualization Analysis. First three columns: Clustering results for loc only, Vis - w/o VAB (ours without VAB pretraining), and Vis - w/ VAB (ours). Colors indicate cluster regions. VAB helps clustering to align better with room partitions. Last two columns: Attention weights visualization for references, with colored stars (size proportional to weights) and a purple triangle for the target." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 282, + 441, + 295 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 282, + 441, + 295 + ], + "spans": [ + { + "bbox": [ + 313, + 282, + 441, + 295 + ], + "type": "text", + "content": "4.5. Visualization Analysis" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 300, + 555, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 300, + 555, + 395 + ], + "spans": [ + { + "bbox": [ + 313, + 300, + 555, + 395 + ], + "type": "text", + "content": "Figure 5 visualizes the clustering results of Reference Sampler via VAB, comparing loc only, Vis - w/o VAB, and Vis - w/ VAB. Different colors represent cluster regions. The Reference Sampler Vis - w/ VAB accurately segments clusters aligning with actual room partitions, proving to be more reliable. In contrast, other methods may incorrectly cluster non-adjacent rooms together which cannot share the same acoustic environment, highlighted by red-circles." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 396, + 556, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 396, + 556, + 503 + ], + "spans": [ + { + "bbox": [ + 313, + 396, + 556, + 503 + ], + "type": "text", + "content": "After identifying reference locations, we visualize the attention weights from the Reference Integration Transformer in Figure 5 (column 4-5 from left). Colored stars denote sampled references, sized by contribution weight, with the target marked by a purple triangle. Unlike Distance-only weights, which may incorrectly prioritize closer references despite obstacles, our model, Ref Attn - w/ VAB, trained with the adaptive Reference Integration Transformer, effectively highlights reliable references in similar acoustic environments." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 514, + 388, + 526 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 514, + 388, + 526 + ], + "spans": [ + { + "bbox": [ + 313, + 514, + 388, + 526 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 534, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 556, + 713 + ], + "type": "text", + "content": "We introduce SoundVista, a novel system designed to synthesize ambient sound from arbitrary scenes at novel viewpoints. Our approach introduces a visual-acoustic binding module to effectively learn visual embeddings linked with local acoustic properties. This enables our system to optimize the placement of reference microphones and derive adaptive weights for each microphone's contribution, conditioned on the target viewpoint and visual captures, thereby, facilitating the final conditioned sound synthesis. SoundVista adapts to diverse room layouts, microphone configurations, and unseen environments, rendering high-quality binaural ambient sound without requiring prior constraints or detailed knowledge of sound sources. Our results, validated on both publicly available data and real-world settings, demonstrate state-of-the-art sound synthesis accuracy and generalization." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8338" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 297, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 91, + 297, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 297, + 145 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 297, + 145 + ], + "type": "text", + "content": "[1] Byeongjoo Ahn, Karren Yang, Brian Hamilton, Jonathan Sheaffer, Anurag Ranjan, Miguel Sarabia, Oncel Tuzel, and Jen-Hao Rick Chang. Novel-view acoustic synthesis from 3d reconstructed rooms. arXiv preprint arXiv:2310.15130, 2023. 1, 2, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 148, + 296, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 148, + 296, + 213 + ], + "spans": [ + { + "bbox": [ + 61, + 148, + 296, + 213 + ], + "type": "text", + "content": "[2] Federico Borra, Israel Dejene Gebru, and Dejan Markovic. Soundfield reconstruction in reverberant environments using higher-order microphones and impulse response measurements. In ICASSP 2019-2019 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 281-285. IEEE, 2019. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 215, + 296, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 215, + 296, + 270 + ], + "spans": [ + { + "bbox": [ + 62, + 215, + 296, + 270 + ], + "type": "text", + "content": "[3] Angel Chang, Angela Dai, Thomas Funkhouser, Maciej Halber, Matthias Niessner, Manolis Savva, Shuran Song, Andy Zeng, and Yinda Zhang. Matterport3d: Learning from rgb-d data in indoor environments. International Conference on 3D Vision (3DV), 2017. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 271, + 296, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 271, + 296, + 335 + ], + "spans": [ + { + "bbox": [ + 62, + 271, + 296, + 335 + ], + "type": "text", + "content": "[4] Changan Chen, Unnat Jain, Carl Schissler, Sebastia Vicenc Amengual Gari, Ziad Al-Halah, Vamsi Krishna Ithapu, Philip Robinson, and Kristen Grauman. Soundspaces: Audiovisual navigation in 3d environments. In European Conference on Computer Vision, pages 17-36. Springer, 2020. 2, 4, 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 338, + 296, + 382 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 338, + 296, + 382 + ], + "spans": [ + { + "bbox": [ + 62, + 338, + 296, + 382 + ], + "type": "text", + "content": "[5] Changan Chen, Ruohan Gao, Paul Calamia, and Kristen Grauman. Visual acoustic matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18858-18868, 2022. 2, 3, 5, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 384, + 296, + 428 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 384, + 296, + 428 + ], + "spans": [ + { + "bbox": [ + 62, + 384, + 296, + 428 + ], + "type": "text", + "content": "[6] Changan Chen, Alexander Richard, Roman Shapovalov, Vamsi Krishna Ithapu, Natalia Neverova, Kristen Grauman, and Andrea Vedaldi. Novel-view acoustic synthesis. arXiv preprint arXiv:2301.08730, 2023. 1, 2, 3, 4, 5, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 430, + 296, + 461 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 430, + 296, + 461 + ], + "spans": [ + { + "bbox": [ + 62, + 430, + 296, + 461 + ], + "type": "text", + "content": "[7] Changan Chen, Wei Sun, David Harwath, and Kristen Grauman. Learning audio-visual dereverberation. In ICASSP, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 464, + 296, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 464, + 296, + 518 + ], + "spans": [ + { + "bbox": [ + 62, + 464, + 296, + 518 + ], + "type": "text", + "content": "[8] Mingfei Chen, Kun Su, and Eli Shlizerman. Be everywhere—hear everything (bee): Audio scene reconstruction by sparse audio-visual samples. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7853–7862, 2023. 2, 3, 5, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 520, + 296, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 520, + 296, + 586 + ], + "spans": [ + { + "bbox": [ + 62, + 520, + 296, + 586 + ], + "type": "text", + "content": "[9] Ziyang Chen, Israel D Gebru, Christian Richardt, Anurag Kumar, William Laney, Andrew Owens, and Alexander Richard. Real acoustic fields: An audio-visual room acoustics dataset and benchmark. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21886-21896, 2024. 2, 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 588, + 296, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 588, + 296, + 643 + ], + "spans": [ + { + "bbox": [ + 57, + 588, + 296, + 643 + ], + "type": "text", + "content": "[10] Ariel Ephrat, Inbar Mosseri, Oran Lang, Tali Dekel, Kevin Wilson, Avinatan Hassidim, William T Freeman, and Michael Rubinstein. Looking to listen at the cocktail party: A speaker-independent audio-visual model for speech separation. arXiv preprint arXiv:1804.03619, 2018. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 644, + 296, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 644, + 296, + 677 + ], + "spans": [ + { + "bbox": [ + 57, + 644, + 296, + 677 + ], + "type": "text", + "content": "[11] Han Fang, Pengfei Xiong, Luhui Xu, and Yu Chen. Clip2video: Mastering video-text retrieval via image clip. arXiv preprint arXiv:2106.11097, 2021. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 679, + 296, + 700 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 679, + 296, + 700 + ], + "spans": [ + { + "bbox": [ + 57, + 679, + 296, + 700 + ], + "type": "text", + "content": "[12] Ruohan Gao and Kristen Grauman. 2.5d visual sound. In CVPR, 2019. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 702, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 702, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 702, + 296, + 713 + ], + "type": "text", + "content": "[13] Ruohan Gao and Kristen Grauman. 2.5 d visual sound. In Pro" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "text", + "content": "ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 324-333, 2019. 2, 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 97, + 555, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 97, + 555, + 140 + ], + "spans": [ + { + "bbox": [ + 317, + 97, + 555, + 140 + ], + "type": "text", + "content": "[14] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 141, + 555, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 141, + 555, + 195 + ], + "spans": [ + { + "bbox": [ + 317, + 141, + 555, + 195 + ], + "type": "text", + "content": "[15] Di Hu, Rui Qian, Minyue Jiang, Xiao Tan, Shilei Wen, Errui Ding, Weiyao Lin, and Dejing Dou. Discriminative sounding objects localization via self-supervised audiovisual matching. Advances in Neural Information Processing Systems, 33: 10077-10087, 2020. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 198, + 555, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 198, + 555, + 251 + ], + "spans": [ + { + "bbox": [ + 316, + 198, + 555, + 251 + ], + "type": "text", + "content": "[16] Christopher A. Ick, Gordon Wichern, Yoshiki Masuyama, François Germain, and Jonathan Le Roux. Spatially-aware losses for enhanced neural acoustic fields. In Audio Imagination: NeurIPS 2024 Workshop AI-Driven Speech, Music, and Sound Generation, 2024. 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 254, + 555, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 254, + 555, + 297 + ], + "spans": [ + { + "bbox": [ + 316, + 254, + 555, + 297 + ], + "type": "text", + "content": "[17] Hyeonjoong Jang, Andreas Meuleman, Dahiyun Kang, Donggun Kim, Christian Richardt, and Min H Kim. Egocentric scene reconstruction from an omnidirectional video. ACM Transactions on Graphics (TOG), 41(4):1-12, 2022. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 299, + 555, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 299, + 555, + 331 + ], + "spans": [ + { + "bbox": [ + 316, + 299, + 555, + 331 + ], + "type": "text", + "content": "[18] Hao Jiang, Calvin Murdock, and Vamsi Krishna Ithapu. Ego-centric deep multi-channel audio-visual active speaker localization. arXiv preprint arXiv:2201.01928, 2022. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 333, + 554, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 333, + 554, + 398 + ], + "spans": [ + { + "bbox": [ + 317, + 333, + 554, + 398 + ], + "type": "text", + "content": "[19] Shoichi Koyama, Tomoya Nishida, Keisuke Kimura, Takumi Abe, Natsuki Ueno, and Jesper Brunnström. Meshrir: A dataset of room impulse responses on meshed grid points for evaluating sound field analysis and synthesis methods. In 2021 IEEE Workshop on Applications of Signal Processing to Audio and Acoustics (WASPAA), pages 1-5, 2021. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 399, + 555, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 399, + 555, + 442 + ], + "spans": [ + { + "bbox": [ + 317, + 399, + 555, + 442 + ], + "type": "text", + "content": "[20] Susan Liang, Chao Huang, Yapeng Tian, Anurag Kumar, and Chenliang Xu. Neural acoustic context field: Rendering realistic room impulse response with neural fields. ArXiv, 2023. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 445, + 555, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 445, + 555, + 488 + ], + "spans": [ + { + "bbox": [ + 317, + 445, + 555, + 488 + ], + "type": "text", + "content": "[21] Susan Liang, Chao Huang, Yapeng Tian, Anurag Kumar, and Chenliang Xu. Av-nerf: Learning neural fields for real-world audio-visual scene synthesis. In Conference on Neural Information Processing Systems (NeurIPS), 2023. 2, 5, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 490, + 555, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 490, + 555, + 544 + ], + "spans": [ + { + "bbox": [ + 317, + 490, + 555, + 544 + ], + "type": "text", + "content": "[22] Ziyi Lin, Shijie Geng, Renrui Zhang, Peng Gao, Gerard De Melo, Xiaogang Wang, Jifeng Dai, Yu Qiao, and Hongsheng Li. Frozen clip models are efficient video learners. In European Conference on Computer Vision, pages 388-404. Springer, 2022. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 545, + 555, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 545, + 555, + 578 + ], + "spans": [ + { + "bbox": [ + 317, + 545, + 555, + 578 + ], + "type": "text", + "content": "[23] Andrew Luo, Yilun Du, Michael J Tarr, Joshua B Tenenbaum, Antonio Torralba, and Chuang Gan. Learning neural acoustic fields. arXiv preprint arXiv., 2021. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 580, + 554, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 580, + 554, + 622 + ], + "spans": [ + { + "bbox": [ + 317, + 580, + 554, + 622 + ], + "type": "text", + "content": "[24] Huaishao Luo, Lei Ji, Ming Zhong, Yang Chen, Wen Lei, Nan Duan, and Tianrui Li. CLIP4Clip: An empirical study of clip for end to end video clip retrieval. arXiv preprint arXiv:2104.08860, 2021. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 624, + 554, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 554, + 657 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 554, + 657 + ], + "type": "text", + "content": "[25] Sagnik Majumder, Changan Chen, Ziad Al-Halah, and Kristen Grauman. Few-shot audio-visual learning of environment acoustics, 2022. 2, 3, 4, 5, 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 658, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 658, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 658, + 554, + 713 + ], + "type": "text", + "content": "[26] Daniel Michelsanti, Zheng-Hua Tan, Shi-Xiong Zhang, Yong Xu, Meng Yu, Dong Yu, and Jesper Jensen. An overview of deep-learning-based audio-visual speech enhancement and separation. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 2021. 3" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8339" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 297, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 73, + 297, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 297, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 297, + 116 + ], + "type": "text", + "content": "[27] Shentong Mo and Yapeng Tian. Audio-visual grouping network for sound localization from mixtures. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10565–10574, 2023. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 297, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 297, + 182 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 297, + 182 + ], + "type": "text", + "content": "[28] Bartlomiej Mróz, Marek Kabaciński, Tomasz Ciotucha, Andrzej Ruminski, and Tomasz Žernicki. Production of six-degrees-of-freedom (6dof) navigable audio using 30 ambisonic microphones. In 2021 Immersive and 3D Audio: from Architecture to Automotive (I3DA), pages 1-5. IEEE, 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 184, + 297, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 184, + 297, + 237 + ], + "spans": [ + { + "bbox": [ + 56, + 184, + 297, + 237 + ], + "type": "text", + "content": "[29] Orhun Olgun, Ege Erdem, and Hüseyin Hachhabiboglu. Sound field interpolation via sparse plane wave decomposition for 6 dof immersive audio. In 2023 Immersive and 3D Audio: from Architecture to Automotive (I3DA), pages 1-10. IEEE, 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 239, + 297, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 239, + 297, + 282 + ], + "spans": [ + { + "bbox": [ + 56, + 239, + 297, + 282 + ], + "type": "text", + "content": "[30] Andrew Owens and Alexei A Efros. Audio-visual scene analysis with self-supervised multisensory features. In Proceedings of the European Conference on Computer Vision (ECCV), pages 631–648, 2018. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 284, + 297, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 284, + 297, + 348 + ], + "spans": [ + { + "bbox": [ + 56, + 284, + 297, + 348 + ], + "type": "text", + "content": "[31] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 350, + 297, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 350, + 297, + 381 + ], + "spans": [ + { + "bbox": [ + 56, + 350, + 297, + 381 + ], + "type": "text", + "content": "[32] Nikunj Raghuvanshi and John Snyder. Parametric wave field coding for precomputed sound propagation. ACM Transactions on Graphics (TOG), 33(4):1-11, 2014. 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 383, + 297, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 383, + 297, + 415 + ], + "spans": [ + { + "bbox": [ + 56, + 383, + 297, + 415 + ], + "type": "text", + "content": "[33] Nikunj Raghuvanshi and John Snyder. Parametric directional coding for precomputed sound propagation. ACM Transactions on Graphics (TOG), 37(4):1-14, 2018. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 416, + 297, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 416, + 297, + 480 + ], + "spans": [ + { + "bbox": [ + 56, + 416, + 297, + 480 + ], + "type": "text", + "content": "[34] Santhosh K Ramakrishnan, Aaron Gokaslan, Erik Wijmans, Oleksandr Maksymets, Alex Clegg, John Turner, Eric Undersander, Wojciech Galuba, Andrew Westbury, Angel X Chang, et al. Habitat-matterport 3d dataset (hm3d): 1000 large-scale 3d environments for embodied ai. arXiv preprint arXiv:2109.08238, 2021. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 482, + 297, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 482, + 297, + 514 + ], + "spans": [ + { + "bbox": [ + 56, + 482, + 297, + 514 + ], + "type": "text", + "content": "[35] Anton Ratnarajah, Zhenyu Tang, and Dinesh Manocha. *Irgan: Room impulse response generator for far-field speech recognition.* arXiv preprint arXiv:2010.13219, 2020. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 515, + 297, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 515, + 297, + 557 + ], + "spans": [ + { + "bbox": [ + 56, + 515, + 297, + 557 + ], + "type": "text", + "content": "[36] Anton Ratnarajah, Shi-Xiong Zhang, Meng Yu, Zhenyu Tang, Dinesh Manocha, and Dong Yu. Fast-rir: Fast neural diffuse room impulse response generator. arXiv preprint arXiv:2110.04057, 2021." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 559, + 297, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 559, + 297, + 613 + ], + "spans": [ + { + "bbox": [ + 56, + 559, + 297, + 613 + ], + "type": "text", + "content": "[37] Anton Ratnarajah, Sreyan Ghosh, Sonal Kumar, Purva Chiniya, and Dinesh Manocha. Av-rir: Audio-visual room impulse response estimation. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 27154–27165, 2023. 2, 3, 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 614, + 297, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 614, + 297, + 668 + ], + "spans": [ + { + "bbox": [ + 56, + 614, + 297, + 668 + ], + "type": "text", + "content": "[38] Alexander Richard, Dejan Markovic, Israel D. Gebru, Steven Krenn, Gladstone Alexander Butler, Fernando Torre, and Yaser Sheikh. Neural synthesis of binaural speech from mono audio. In International Conference on Learning Representations, 2021. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 670, + 297, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 297, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 297, + 712 + ], + "type": "text", + "content": "[39] Alexander Richard, Peter Dodds, and Vamsi Krishna Ithapu. Deep impulse responses: Estimating and parameterizing filters with deep networks. arXiv preprint arXiv:2202.03416, 2022. 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 316, + 73, + 555, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 555, + 139 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 555, + 139 + ], + "type": "text", + "content": "[40] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention-MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, pages 234-241. Springer, 2015. 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 141, + 555, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 555, + 194 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 555, + 194 + ], + "type": "text", + "content": "[41] Nikhil Singh, Jeff Mentch, Jerry Ng, Matthew Beveridge, and Iddo Drori. Image2reverb: Cross-modal reverb impulse response synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 286-295, 2021. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 197, + 555, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 197, + 555, + 250 + ], + "spans": [ + { + "bbox": [ + 316, + 197, + 555, + 250 + ], + "type": "text", + "content": "[42] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In International Conference on Learning Representations, 2021. 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 252, + 555, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 252, + 555, + 285 + ], + "spans": [ + { + "bbox": [ + 316, + 252, + 555, + 285 + ], + "type": "text", + "content": "[43] Kun Su, Mingfei Chen, and Eli Shlizerman. Inras: Implicit neural representation for audio scenes. Advances in Neural Information Processing Systems, 35:8144-8158, 2022. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 286, + 555, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 286, + 555, + 350 + ], + "spans": [ + { + "bbox": [ + 316, + 286, + 555, + 350 + ], + "type": "text", + "content": "[44] Ruijie Tao, Zexu Pan, Rohan Kumar Das, Xinyuan Qian, Mike Zheng Shou, and Haizhou Li. Is someone speaking? exploring long-term temporal features for audio-visual active speaker detection. In Proceedings of the 29th ACM international conference on multimedia, pages 3927-3935, 2021. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 354, + 555, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 354, + 555, + 398 + ], + "spans": [ + { + "bbox": [ + 316, + 354, + 555, + 398 + ], + "type": "text", + "content": "[45] Yapeng Tian, Di Hu, and Chenliang Xu. Cyclic co-learning of sounding object visual grounding and sound separation. 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2744-2753, 2021. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 399, + 555, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 399, + 555, + 442 + ], + "spans": [ + { + "bbox": [ + 316, + 399, + 555, + 442 + ], + "type": "text", + "content": "[46] Aäron van den Oord, Sander Dieleman, Heiga Zen, Karen Simonyan, Oriol Vinyals, Alexander Graves, Nal Kalchbrenner, Andrew Senior, and Koray Kavukcuoglu. Wavenet: A generative model for raw audio. In Arxiv, 2016. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 445, + 543, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 445, + 543, + 454 + ], + "spans": [ + { + "bbox": [ + 316, + 445, + 543, + 454 + ], + "type": "text", + "content": "[47] Tor Erik Vigran. Building acoustics. CRC Press, 2014. 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 456, + 555, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 456, + 555, + 499 + ], + "spans": [ + { + "bbox": [ + 316, + 456, + 555, + 499 + ], + "type": "text", + "content": "[48] Xiuchao Wu, Jiamin Xu, Zihan Zhu, Hujun Bao, Qixing Huang, James Tompkin, and Weiwei Xu. Scalable neural indoor scene rendering. ACM transactions on graphics, 41 (4), 2022. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 501, + 555, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 501, + 555, + 567 + ], + "spans": [ + { + "bbox": [ + 316, + 501, + 555, + 567 + ], + "type": "text", + "content": "[49] Linning Xu, Vasu Agrawal, William Laney, Tony Garcia, Aayush Bansal, Changil Kim, Samuel Rota Bulò, Lorenzo Porzi, Peter Kontschieder, Aljaž Božić, Dahua Lin, Michael Zollhöfer, and Christian Richardt. VR-NeRF: High-fidelity virtualized walkable spaces. In SIGGRAPH Asia Conference Proceedings, 2023. 1, 4, 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 568, + 555, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 568, + 555, + 633 + ], + "spans": [ + { + "bbox": [ + 316, + 568, + 555, + 633 + ], + "type": "text", + "content": "[50] Shaoheng Xu, Jihui Aimee Zhang, Thushara D. Abhayapala, Amy Bastine, Wei-Ting Lai, and Prasanga N. Samarasinghe. Sparse sound field representation using complex orthogonal matching pursuit. In ICASSP 2024 - 2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1336-1340, 2024. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 635, + 555, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 555, + 668 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 555, + 668 + ], + "type": "text", + "content": "[51] Xudong Xu, Hang Zhou, Ziwei Liu, Bo Dai, Xiaogang Wang, and Dahua Lin. Visually informed binaural audio generation without binaural audiios. In CVPR, 2021. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 670, + 555, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 555, + 712 + ], + "type": "text", + "content": "[52] Hongwei Xue, Yuchong Sun, Bei Liu, Jianlong Fu, Ruihua Song, Houqiang Li, and Jiebo Luo. Clip-vip: Adapting pretrained image-text model to video-language representation alignment. arXiv preprint arXiv:2209.06430, 2022. 3" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "8340" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 297, + 329 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 56, + 73, + 297, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 297, + 137 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 297, + 137 + ], + "type": "text", + "content": "[53] Ryuichi Yamamoto, Eunwoo Song, and Jae-Min Kim. Parallel wavegan: A fast waveform generation model based on generative adversarial networks with multi-resolution spectrogram. In ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 6199-6203. IEEE, 2020. 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 140, + 296, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 140, + 296, + 183 + ], + "spans": [ + { + "bbox": [ + 56, + 140, + 296, + 183 + ], + "type": "text", + "content": "[54] Yuxin Ye, Wenming Yang, and Yapeng Tian. Lavss: Location-guided audio-visual spatial audio separation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 5508–5519, 2024. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 185, + 296, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 185, + 296, + 217 + ], + "spans": [ + { + "bbox": [ + 56, + 185, + 296, + 217 + ], + "type": "text", + "content": "[55] Zechen Zhang, Nikunj Raghuvanshi, John Snyder, and Steve Marschner. Ambient sound propagation. ACM Transactions on Graphics (TOG), 37(6):1-10, 2018. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 219, + 296, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 219, + 296, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 219, + 296, + 262 + ], + "type": "text", + "content": "[56] Hang Zhou, Ziwei Liu, Xudong Xu, Ping Luo, and Xiaogang Wang. Vision-infused deep audio inpainting. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 283-292, 2019. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 263, + 296, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 263, + 296, + 329 + ], + "spans": [ + { + "bbox": [ + 56, + 263, + 296, + 329 + ], + "type": "text", + "content": "[57] Hang Zhou, Xudong Xu, Dahua Lin, Xiaogang Wang, and Ziwei Liu. Sep-stereo: Visually guided stereophonic audio generation by associating source separation. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XII 16, pages 52-69. Springer, 2020. 2" + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "8341" + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Sparse Point Cloud Patches Rendering via Splitting 2D Gaussians/dd278d9d-da45-4b1e-bcc2-33b0924a1fa1_content_list.json b/2025/Sparse Point Cloud Patches Rendering via Splitting 2D Gaussians/dd278d9d-da45-4b1e-bcc2-33b0924a1fa1_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..280834fe8d52296adcef79c7eb262aa419c0ee56 --- /dev/null +++ b/2025/Sparse Point Cloud Patches Rendering via Splitting 2D Gaussians/dd278d9d-da45-4b1e-bcc2-33b0924a1fa1_content_list.json @@ -0,0 +1,1555 @@ +[ + { + "type": "text", + "text": "Sparse Point Cloud Patches Rendering via Splitting 2D Gaussians", + "text_level": 1, + "bbox": [ + 166, + 130, + 828, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Changfeng Ma $^{1}$ , Ran Bi $^{1}$ , Jie Guo $^{1}$ , Chongjun Wang $^{1}$ , Yanwen Guo $^{12*}$ $^{1}$ Nanjing University, Nanjing, China \n $^{2}$ School of Software, North University of China", + "bbox": [ + 163, + 179, + 833, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{changfengma, 211250233}@smail.nju.edu.cn", + "bbox": [ + 318, + 219, + 679, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{guojie, chjwang, ywguo}@nju.edu.cn", + "bbox": [ + 351, + 236, + 640, + 251 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9d4d6b033eabffbf029eaf37ab8a1448429297366823f9b8d6fa2a01ceac05ee.jpg", + "image_caption": [ + "Car", + "Point Cloud" + ], + "image_footnote": [], + "bbox": [ + 138, + 273, + 235, + 343 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/bf5f8ecb853a56f2c13091a2a43ee81d4ae0a6e3012240cb93d9c4696f7d0689.jpg", + "image_caption": [ + "NPBG+±" + ], + "image_footnote": [], + "bbox": [ + 251, + 272, + 339, + 335 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/591b037e67a826ba4bf302c9828820d4f1bbd5eeb832e4072bb62dba7c4aedbb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 243, + 335, + 344, + 362 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/bc7518ba8c933acdda4fa80540ae016e91b2a7ba9ab93f2571a60109a6c64ff9.jpg", + "image_caption": [ + "TriVol" + ], + "image_footnote": [], + "bbox": [ + 359, + 272, + 447, + 335 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ad480242986f91afb37657047c5966ade456c4b558d3075eb78c6436e8733a09.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 352, + 335, + 454, + 362 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/15a6c7afc42cd95fa20fe573507e1d61feb5c739cdfd5d4d48dcf993d02d1e98.jpg", + "image_caption": [ + "PFGS" + ], + "image_footnote": [], + "bbox": [ + 143, + 375, + 232, + 438 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b1220996de86ded1acbef8a69f08538b905c0c6ba1e8c6a2734f1207b379fbcf.jpg", + "image_caption": [ + "Figure 1. Our method excels in rendering intricate, photo-realistic images from point clouds of different categories. In this instance, our model is trained on the Car category, utilizing 20K points as input." + ], + "image_footnote": [], + "bbox": [ + 135, + 439, + 235, + 465 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7659f8d1a9478ac4cafb70a5ec795e9e447c8ef8389ec92774c739b32e421efa.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 251, + 375, + 339, + 438 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f25f999b074e6f5d71444861da5a33b76655a61a669a79cdcb6be4623646d986.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 243, + 439, + 344, + 465 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d33e9d13af948c7082cbd821c671c6e9f444f1c88395d45f44dfab564002ac9f.jpg", + "image_caption": [ + "GT" + ], + "image_footnote": [], + "bbox": [ + 361, + 372, + 447, + 436 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/996bbb729c98e5b0318897b6919590734b0358864ed5a774ca336f04f6b58a56.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 351, + 439, + 454, + 465 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c4e48cfab5721918aad94808b29ad87164af7eb18d9f79af61d8c94f7accefe7.jpg", + "image_caption": [ + "Generalize to Other Categories" + ], + "image_footnote": [], + "bbox": [ + 488, + 273, + 625, + 377 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d294a263598f72cd8895f81e57fd6f224b2b8b15f677352d7ee3768d05cc5b86.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 486, + 378, + 625, + 465 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2ea980b1f830657c635adfc199d2f0b7234b3375a833be1d0b788853f0dfddc5.jpg", + "image_caption": [ + "Sculpture" + ], + "image_footnote": [], + "bbox": [ + 627, + 273, + 756, + 392 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/756cc0419e4f854567bf136d5e3b7cd287982c77e453943d2447eabbc11b71ec.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 629, + 402, + 705, + 464 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/56228bb22cd5c45271e6adc5f71f364b3a7a8dde46090ef763073123f11c4286.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 761, + 273, + 859, + 329 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2bcc0458d0620247aaefe164414bb61082bc343a159bd2957a268515dbb7c507.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 763, + 330, + 859, + 392 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b9669c9491e4f05faec33dcb8eb7c6ed71fd289b25d545fa2b11802cc558ed0e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 710, + 393, + 800, + 464 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/dc52fd9874f362f2def01a2be70f3b28299b1a60fdc3c862d6c37bf01ae6279b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 803, + 393, + 861, + 428 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2bd0f885f9ed91620037d0f37d4edbb7eb87f9e3edaa5f727e6e7ff75a2b8924.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 803, + 429, + 861, + 464 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 527, + 326, + 544 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Current learning-based methods predict NeRF or 3D Gaussians from point clouds to achieve photo-realistic rendering but still depend on categorical priors, dense point clouds, or additional refinements. Hence, we introduce a novel point cloud rendering method by predicting 2D Gaussians from point clouds. Our method incorporates two identical modules with an entire-patch architecture enabling the network to be generalized to multiple datasets. The module normalizes and initializes the Gaussians utilizing the point cloud information including normals, colors and distances. Then, splitting decoders are employed to refine the initial Gaussians by duplicating them and predicting more accurate results, making our methodology effectively accommodate sparse point clouds as well. Once trained, our approach exhibits direct generalization to point clouds across different categories. The predicted Gaussians are employed directly for rendering without additional refinement on the rendered images, retaining the benefits of 2D Gaussians. We conduct extensive experiments on various datasets, and the results demonstrate the superiority and generalization of our method, which achieves SOTA performance. The code is", + "bbox": [ + 88, + 560, + 485, + 878 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "available at https://github.com/murcherful/GauPCRender.", + "bbox": [ + 513, + 530, + 898, + 545 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 573, + 643, + 588 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Point clouds can be easily acquired using various types of 3D scanners, making them a recent research hotspot. Photorealistic rendering of point clouds is significant for many applications such as visualization, virtual reality and automatic driving. However, the sparsity and discrete nature of point clouds continue to pose challenges for achieving photo-realistic rendering.", + "bbox": [ + 509, + 598, + 906, + 704 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Traditional graphics rendering methods [2, 22] merely project points as planes or spheres onto images. Consequently, their outcomes lack photo-realism and are riddled with numerous holes. Previous deep learning-based works [1, 8, 23] necessitate point clouds and corresponding images from several views as input, which are difficult to acquire in practice. These methods also require training or fine-tuning for each scene or object. Recent works [13, 14] have focused on predicting NeRF [20] from given point clouds to render images from arbitrary views. PFGS [28] predicts the 3D Gaussians [17] with features for point clouds. After rendering, PFGS refines the rendered images based on these features. Although these methods can render point clouds", + "bbox": [ + 509, + 704, + 908, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author.", + "bbox": [ + 114, + 887, + 243, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "27285", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "table", + "img_path": "images/a57b6cd5ae8ccd58bd3dd30d48d68c03884fb676e453c9c514625135e3b57b0f.jpg", + "table_caption": [ + "Table 1. The comparison of different methods across several dimensions. Here, \"S. S/O.\", \"S. C.\" and \"M. C.\" represent \"single scene or object\", \"single category\" and \"multiple categories\", respectively. $\\mathcal{P}$ and $\\mathcal{I}$ indicate point clouds and images, separately. $\\dagger$ : Traditional graphics rendering approach is unable to render photo-realistic images. $\\ddagger$ : PFGS requires further refinement on images rendered from 3D gaussians." + ], + "table_footnote": [], + "table_body": "
MethodGeneralization abilityInputOutputPre- Fine- trainFine- tunePoint number
\\( GR^† \\)AnyPI✘✘Any
NPBG [1]S. S/O.\\( P+I \\)I✓✓100K
NPCR [8]S. S/O.\\( P+I \\)I✘✓100K
NPBG++ [23]S. S/O.\\( P+I \\)I✓✘100K
TriVol [13]S. C.PNeRF[20]✓✘100K
Point2Pix [14]S. C.PNeRF[20]✓✘100K
PFGS [28]S. C.P3DGS[17]‡✓✘80K
OursM. C.P2DGS[15]✓✘2K-100K
", + "bbox": [ + 91, + 191, + 483, + 364 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "without training or fine-tuning on each data, they still have limitations, such as the need for dense point clouds, poor generalization, slow rendering speeds, and blurry results.", + "bbox": [ + 89, + 385, + 482, + 430 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2D Gaussian Splatting [15] has recently been proposed, which includes explicit normals of Gaussians. These normals facilitate easy initialization, learning, and prediction by the network from point clouds. Motivated by this, we introduce a novel method to predict the 2D Gaussians of the given point cloud and rasterize the predicted 2D Gaussians into images for photo-realistic point cloud rendering.", + "bbox": [ + 89, + 431, + 482, + 537 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Due to taking entire point clouds as input, previous methods are hard to generalize to other categories or adapt to different point distributions. Processing only point cloud patches with the network can enhance its generalization capability and reduce dependence on category priors [3, 12]. However, rendering Gaussians predicted from patch points results in incomplete images that are unsupervisible. Therefore, we employ an entire-patch architecture that includes two identical 2D Gaussian prediction modules to handle entire and patch point clouds, separately. We use the rendering results of the entire point cloud as a \"background\" to obtain a complete image for proper supervision. The module first normalizes and initializes the 2D Gaussians based on the given point clouds, encompassing the estimated normals, scales, and colors, which are crucial for our network to converge and predict more accurate results. Subsequently, a point cloud encoder and multiple splitting decoders are employed to extract local features and predict the parameters of the duplicated Gaussians. Equipped with the splitting decoder, our network can handle sparse point clouds and produce high-quality details with dense Gaussians. Thanks to the proposed architecture, our method also exhibits strong generalization across scenes and objects. Without further refinement on rendered images, our", + "bbox": [ + 89, + 537, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "method retains the advantages of 2D Gaussian Splatting in rendering, such as rapid rendering speed.", + "bbox": [ + 511, + 90, + 903, + 121 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We conduct extensive comparative and ablation experiments, and the results demonstrate the superiority of our method, which can render point clouds into high-quality detailed images with excellent generalization. We will make our dataset and codes public in the future. Our main contributions are as follows:", + "bbox": [ + 511, + 122, + 903, + 212 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce a method that directly predicts 2D Gaussians from point clouds to render photo-realistic images without any refinement, while retaining the advantages of 2D Gaussian Splatting.", + "- Equipped with an entire-patch architecture, our method can handle point cloud patches, leading to significant generalization across scenes and objects.", + "- We introduce the splitting decoder, which splits Gaussians into denser outcomes, enabling our method to render detailed images from even sparse point clouds." + ], + "bbox": [ + 513, + 214, + 903, + 364 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 381, + 653, + 398 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Point Cloud Rendering", + "text_level": 1, + "bbox": [ + 513, + 407, + 728, + 424 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Traditional Point Cloud Rendering. Point cloud rendering has consistently been a significant research topic [18]. Conventional point cloud rendering techniques [6, 24, 27, 32] directly rasterize points onto 2D screens for rendering, yet they grapple with the issue of holes. Subsequent studies [2, 22, 25, 33] have proposed splatting points with elliptic discs, ellipsoids, or surfels for point cloud rendering to mitigate the holes observed during the rendering process. Although these graphic rendering methods can render any point cloud, the resulting images often contain holes and exhibit blurry details, lacking photo-realism.", + "bbox": [ + 511, + 431, + 903, + 597 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Point-based novel view synthesis. For photo-realistic point cloud rendering, several recently proposed methods utilize corresponding images of point clouds as additional information. NPBG [1] signs descriptors to point clouds and rasterizes into multi-resolution raw images with features and employs a U-Net to render the final results. Dai et al. [8] aggregate points with features onto multiple planes based on a given view and point cloud. Following multi-plane-based voxelization, they employ a 3D U-Net to predict the rendered image. ADOP [26] is introduced for rendering HD images from point clouds, which employs several global parameters, a neural renderer and a differentiable physically-based tonemapper for optimization. These methods necessitate training for each point cloud with images from several views to render images of novel views, which is a time-consuming process. To reduce training time, NPBG++ [23] is trained on a collection of point clouds using a network that extracts features from corresponding images. This network aligns the features of input images to point clouds without the need for fine-tuning. However, this", + "bbox": [ + 511, + 599, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "27286", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "method still requires images as input, which can be challenging to obtain in practice.", + "bbox": [ + 89, + 90, + 480, + 119 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Learning-based Point Cloud Rendering. Several works [13, 14, 28] are proposed for rendering point clouds without the need for training or fine-tuning on each point cloud and corresponding images. TriVol [13] utilizes triple-plane-based grouping to query point features for predicting a NeRF from a given point cloud. Hu et al. [14] propose multi-scale radiance fields to predict NeRFs from point clouds. They also introduce a sampling strategy to accelerate fine-tuning and rendering speed. PFGS [28], a two-stage approach, for the first time employs feature splatting to predict 3D Gaussians from point clouds and uses a recurrent decoder to refine the rendered images. These methods are capable of rendering point clouds without fine-tuning or the need for input images, yet they still depend on categorical priors, dense points, or further refinements. Table 1 summarizes the distinctions between our method and the previous methods across several dimensions, highlighting the innovation and superiority of our proposed approach.", + "bbox": [ + 89, + 121, + 482, + 393 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Gaussian Splitting", + "text_level": 1, + "bbox": [ + 89, + 402, + 272, + 417 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3D Gaussian Splitting [17] is proposed with remarkable performance in novel view synthesis and further explores the rendering capability of the point clouds. 3DGS utilizes the differentiable splatting rendering to optimize 3D Gaussian primitives initialized from point clouds. Scenes are explicitly represented with 3D Gaussians that can be rendered with remarkable speed. Meshes can also be reconstructed from 3D Gaussians[10]. However, the normals of 3D Gaussians are so vague that the resulting geometries lack precision. To generate more accurate geometries, Hua et al. [15] propose 2DGS that employs 2D Gaussian disks containing explicit normals as representation primitives. With precise normals, meshes can accurately reconstruct the geometry of objects. Given that the normals of 2D Gaussians can be readily determined, 2D Gaussians can also be initialized using normals. Luckily, the normals of points can be effortlessly estimated [11] facilitating the initialization and forecasting of 2D Gaussians. Motivated by this, we opt to predict 2D Gaussians in place of 3D Gaussians from the provided point clouds.", + "bbox": [ + 89, + 422, + 482, + 724 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 739, + 181, + 753 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given a point cloud $\\mathcal{P}$ and a camera $\\mathcal{C}$ , point cloud rendering aims to render a photo-realistic image $\\mathcal{I}$ . Here, point cloud $\\mathcal{P}$ contains of coordinates of 3D points $\\mathbf{P}_{\\mathbf{x}} \\in \\mathbb{R}_{N \\times 3}$ and colors of points $\\mathbf{P}_{\\mathbf{c}} \\in [0,1]_{N \\times 3}$ , where $N$ indicates the point number of $\\mathcal{P}$ . Camera $\\mathcal{C}$ consists of intrinsic matrix $\\mathbf{K} \\in \\mathbb{R}_{3 \\times 3}$ and extrinsic matrix $\\mathbf{P} \\in \\mathbb{R}_{4 \\times 4}$ .", + "bbox": [ + 89, + 763, + 480, + 854 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For photo-realistic point cloud rendering, we propose a network $\\mathcal{N}$ to predict 2D Gaussians $\\mathcal{G} = \\mathcal{N}(\\mathcal{P})$ given point cloud $\\mathcal{P}$ , and employ differentiable splatting rendering [15]", + "bbox": [ + 89, + 856, + 482, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "$f_{R}$ to synthesize images $\\mathcal{I} = f_{R}(\\mathcal{G},\\mathcal{C})$ . The parameters of a 2D Gaussian $\\mathbf{g} = \\{\\mathbf{x},\\mathbf{s},o,\\mathbf{c},\\mathbf{n},\\alpha \\} \\in \\mathcal{G}$ include: (1) position $\\mathbf{x}\\in \\mathbb{R}_3$ , (2) scale $\\mathbf{s}\\in \\mathbb{R}_2$ , (3) opacity $o\\in [0,1]$ , (4) spherical harmonic of color $\\mathbf{c}\\in \\mathbb{R}_d$ , (5) normal $\\mathbf{n}\\in \\mathbb{R}_3$ and (6) rotation angle along normal $\\alpha \\in [0,2\\pi ]$ , where $d$ represents the parameter number of spherical harmonic for color. As shown in Figure 2, our network contains two identical 2D Gaussian prediction modules (3.1) with an entire-patch architecture (3.2).", + "bbox": [ + 511, + 90, + 903, + 226 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1.2D Gaussian Prediction Module", + "text_level": 1, + "bbox": [ + 511, + 238, + 794, + 252 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1.1. Initialization of 2D Gaussians", + "text_level": 1, + "bbox": [ + 511, + 261, + 764, + 273 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We first introduce 2D Gaussian prediction module to predict the 2D Gaussians $\\mathcal{G}$ given a point cloud $\\mathcal{P}$ as input. The positions of the point cloud $\\mathcal{P}$ are normalized to the range $[-1,1]$ based on the geometric center $\\mathbf{c} \\in \\mathbb{R}_3$ and scale $s \\in R$ of the input point cloud to facilitate easier feature extraction and prediction of results, where the scale is defined as the maximum distance between any point and $\\mathbf{c}$ . Our module then initializes one 2D Gaussian $\\mathbf{g}$ for each normalized point $\\mathbf{p} = \\{\\mathbf{p}_{\\mathbf{x}}, \\mathbf{p}_{\\mathbf{c}}\\} \\in \\mathcal{P}$ , where $\\mathbf{p}_{\\mathbf{x}} \\in \\mathcal{R}_3$ and $\\mathbf{p}_{\\mathbf{c}} \\in \\mathcal{R}_3$ indicate the position and color of the point respectively. The position $\\mathbf{x}$ of the Gaussian is initialized with $\\mathbf{p}_{\\mathbf{x}}$ . The spherical harmonic $\\mathbf{c}$ and scale $\\mathbf{s}$ are initialized as the same as 2DGS [15] according to $\\mathbf{p}_{\\mathbf{c}}$ and the minimum distance between points, separately. The opacity $o$ is set to 1 to ensure that all Gaussians are visible. As we mentioned before, the geometry of given point cloud $\\mathcal{P}$ provides important clues for the rotation of Gaussians. Gaussian Splatting works [15, 17] utilize quaternions to represent the rotation of Gaussians and only randomly initialize them as Figure 3(b) shows, which are difficult for networks to predict intuitively. Fortunately, the normals of given point cloud can be easily estimated. Therefore, we represent the rotation of the Gaussian in terms of the normal $\\mathbf{n}$ and the rotation angle $\\alpha$ around that normal, and initialize the normal of the Gaussian using the estimated normal of the point cloud. Figure 3 (a) and (c) illustrate this initialization process. Specifically, we estimate the normals for $\\mathcal{P}$ [11, 32], where the normal $\\hat{\\mathbf{n}}$ for each point $\\mathbf{p}$ is utilized to initialize the normal $\\mathbf{n}$ of Gaussian. And the rotation angle $\\alpha$ of Gaussian is set to 0. The estimated normal may not be precise, but it provides crucial information and ensures that each Gaussian is visible, thereby guaranteeing that the network can converge correctly. After initializing for all the points, the initialized Gaussians $\\mathcal{G}^i$ composing of the initialized positions, normals, scales, spherical harmonics, opacities and rotation angles ( $\\mathbf{X}^i, \\mathbf{N}^i \\in \\mathbb{R}_{N\\times 3}, \\mathbf{S}^i \\in \\mathbb{R}_{N\\times 2}, \\mathbf{C}^i \\in \\mathbb{R}_{N\\times 3}, O^i, A^i \\in \\mathbb{R}_N$ ) is passed to our network as input and predicting basis.", + "bbox": [ + 511, + 279, + 903, + 840 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1.2. Splitting Decoder", + "text_level": 1, + "bbox": [ + 511, + 842, + 681, + 856 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As shown in Figure 2, local features are required before splitting and predicting 2D Gaussians. To extract features from the initialized 2D Gaussians, we employ", + "bbox": [ + 511, + 856, + 903, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "27287", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/554cdefc4dec6134350b048c8b7bec91da332ec181726e8f215df849700885ea.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 98, + 90, + 637, + 242 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/311ad8e7a092496b911ba656a3b4c686079d96882aeceedeebe70fbe13718379.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 643, + 90, + 898, + 242 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/bdcbd1a965c81961da135a0fcf2765fee40118f781745bb22881f183f1d6201d.jpg", + "image_caption": [ + "Figure 2. Overview of our proposed method. Our method predicts 2D Gaussians for point cloud rendering, employing an entire-patch architecture (bottom) and the 2D Gaussian Prediction Module (top left) with splitting decoders (top right)." + ], + "image_footnote": [], + "bbox": [ + 96, + 244, + 898, + 335 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/5677a16993197090d21fa6b6705109dc5505b68d8823bbf81cbd9e3c2d709160.jpg", + "image_caption": [ + "Figure 3. The illustration of our initialization approach, where the left side of each picture is a 2D schematic diagram and the right side is the rendered images of Gaussians. (a) The estimated normals of the point cloud. (b) Randomly Initialization. (c) Our Initialization. (d) Predicted 2D Gaussians." + ], + "image_footnote": [], + "bbox": [ + 93, + 392, + 480, + 523 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "PointMLP[19] as the encoder $E$ , which is a simple residual MLP framework for point cloud learning with great performance on many tasks. The local feature $\\mathbf{F}_l$ is extracted by:", + "bbox": [ + 89, + 608, + 482, + 667 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {F} _ {l} = E (\\mathbf {X} ^ {i}, \\mathbf {C} ^ {i}, \\mathbf {N} ^ {i}, \\mathbf {S} ^ {i}).\n$$\n", + "text_format": "latex", + "bbox": [ + 200, + 667, + 370, + 685 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Subsequently, $\\mathbf{F}_l$ is augmented with $\\mathbf{X}^i$ , $\\mathbf{C}^i$ , $\\mathbf{N}^i$ , $\\mathbf{S}^i$ as the input features for the splitting decoders. Our splitting decoders $D$ also utilize $\\mathcal{G}^i$ as the basis for prediction and forecast 2D Gaussians $\\mathcal{G}^p$ for rendering purposes. There are six splitting decoders $D_x, D_s, D_c, D_n, D_\\alpha, D_o$ responsible for predicting the six parameters $X^p, S^p, C^p, N^p, A^p, O^p$ of 2D Gaussians. Each decoder contains a weight-shared MLP [5] as shown in Figure 2. The splitting and predicting process for the 6 parameters is basically the same, except for slight differences in shapes. Take the process of predicting the position $X^p$ as an example. The decoder $D_x$ predicts $K$ shifts $\\Delta^x$ based on the position $\\mathbf{X}^i$ :", + "bbox": [ + 89, + 691, + 482, + 872 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\Delta} _ {1} ^ {x}, \\boldsymbol {\\Delta} _ {2} ^ {x},..., \\boldsymbol {\\Delta} _ {K} ^ {x} = D _ {x} (\\mathbf {F} _ {l}, \\mathbf {X} ^ {i}, \\mathbf {C} ^ {i}, \\mathbf {N} ^ {i}, \\mathbf {S} ^ {i}),\n$$\n", + "text_format": "latex", + "bbox": [ + 137, + 883, + 434, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $K$ indicates the number of splits. The $K$ shifts are added to the same $\\mathbf{X}^i$ that is \"split\" the initial position $\\mathbf{X}^i$ to $K$ new positions. The predicted position $\\mathbf{X}^p\\in \\mathbb{R}_{K\\cdot N\\times 3}$ is the combination of these $K$ new positions:", + "bbox": [ + 511, + 393, + 903, + 455 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {X} ^ {p} = \\bigcup_ {j = 1} ^ {K} \\left(\\boldsymbol {\\Delta} _ {j} ^ {x} + \\mathbf {X} ^ {i}\\right).\n$$\n", + "text_format": "latex", + "bbox": [ + 632, + 465, + 782, + 508 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "After applying all the splitting decoders, we can obtain the parameters of $K \\cdot N$ Gaussians $\\mathcal{G}^p$ . With the splitting decoders, our method can manage sparse point clouds by augmenting the number of Gaussians, serving as a form of point cloud upsampling. A greater number of Gaussians can also capture more complex features, enhancing the details of rendered images. The predicted 2D Gaussians are based on normalized points, hence they need to be de-normalized to achieve accurate rendering results. The de-normalization is applied to the predicted position $X^p$ and scale $S^p$ .", + "bbox": [ + 511, + 521, + 906, + 672 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Entire-Patch Architecture", + "text_level": 1, + "bbox": [ + 511, + 681, + 750, + 696 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "However, the 2D Gaussian prediction module can only be trained on complete point clouds, which leads to poor generalization across different categories with varying point distributions, as we only have ground truth images of entire objects for supervision. Numerous methods for other point cloud tasks, including surface reconstruction[3] and semantic segmentation[12], utilize point cloud patches as processing units to enhance the generalization capabilities of their networks. However, the corresponding ground truth images of point cloud patches are unavailable in practice, which makes training on patches impossible. To facilitate the application of our method to point cloud patches, we utilize an entire-patch architecture, as illustrated at the bottom of Fig-", + "bbox": [ + 511, + 704, + 906, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "27288", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ure 2. Our approach encompasses two 2D Gaussian prediction modules, $\\mathcal{N}_e$ and $\\mathcal{N}_p$ , which operate on the entire point cloud and the point cloud patch, respectively. Given an entire point cloud $\\mathcal{P}_e$ for rendering, our method randomly selects a point as the center point of a point cloud patch. The point cloud patch $\\mathcal{P}_p$ is obtained by gathering $N_p$ nearest points around the center point. The 2D Gaussians of point cloud $\\mathcal{P}_e$ and patch $\\mathcal{P}_p$ are predicted by two prediction modules, where $\\mathcal{G}_e = \\mathcal{N}_e(\\mathcal{P}_e)$ and $\\mathcal{G}_p = \\mathcal{N}_p(\\mathcal{P}_p)$ . The non-patch part $\\mathcal{G}_e'$ of $\\mathcal{G}_e$ are the 2D Gaussians corresponding to the non-patch points $\\mathcal{P}_e' = \\mathcal{P}_e - \\mathcal{P}_p$ . The final 2D Gaussians $\\mathcal{G} = \\mathcal{G}_e' \\cup \\mathcal{G}_p$ are used to render images for supervision during training. Here, $\\mathcal{G}_e'$ serves as the \"background\" ensuring that the rendered image is complete for calculating the loss function and $\\mathcal{N}_p$ can be properly trained. In practice, we initially train $\\mathcal{N}_e$ separately and then freeze $\\mathcal{N}_e$ while training $\\mathcal{N}_p$ for our method. Both $\\mathcal{N}_e$ and $\\mathcal{N}_p$ can be utilized for inference. The inference of $\\mathcal{N}_e$ directly takes $\\mathcal{P}_e$ as input and outputs the entire 2D Gaussians for rendering. For $\\mathcal{N}_p$ , our method divides the point cloud into several patches. To ensure that all points of the input point cloud are covered, we repeatedly perform the process of randomly selecting a central point and extracting surrounding points as the patch point cloud, continuing until all points are included in at least one patch. The 2D Gaussian of the entire point cloud is the combination of the predictions from all patches.", + "bbox": [ + 89, + 90, + 483, + 484 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Training Details", + "text_level": 1, + "bbox": [ + 89, + 493, + 251, + 508 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given $N_{c}$ cameras $\\mathcal{C}_1,\\ldots ,\\mathcal{C}_{N_c}$ , the rendered images $\\mathcal{I}_1^p,\\dots ,\\mathcal{I}_{N_c}^p$ is obtained by $f_{R}(\\mathcal{G},\\mathcal{C}_{1}),\\ldots ,f_{R}(\\mathcal{G},\\mathcal{C}_{N_{c}})$ . Similar to [15], we employ MSE and SSIM between rendered images and the corresponding ground truth images $\\mathcal{I}_1^g,\\dots ,\\mathcal{I}_{N_c}^g$ as loss functions for supervision:", + "bbox": [ + 89, + 515, + 483, + 594 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\frac {1}{N _ {c}} \\sum_ {i = 1} ^ {N _ {c}} \\left(\\beta \\mathcal {L} _ {M S E} (\\mathcal {I} _ {i} ^ {p}, \\mathcal {I} _ {i} ^ {g}) + (1 - \\beta) \\mathcal {L} _ {S S I M} (\\mathcal {I} _ {i} ^ {p}, \\mathcal {I} _ {i} ^ {g})\\right).\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 604, + 483, + 647 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We implement our method with PyTorch [21]. The split number $K$ is set to 4. The point number $N_{p}$ of a patch is 2048. $\\beta$ is set to 0.8 for the loss function. The number of rendered images $N_{c}$ is 8. The resolutions of images are 256x256, 640x512 and 512x512 for objects, scenes and human bodies. We employ Adam optimizer whose learning rate is $1.0 \\times 10^{-4}$ . The batch size is set to 8, and the maximum epoch of training is 480. Training our method takes about 30 hours with a RTX 3090 GPU.", + "bbox": [ + 89, + 657, + 483, + 792 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 806, + 223, + 824 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Settings", + "text_level": 1, + "bbox": [ + 89, + 832, + 189, + 848 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Dataset. We compare our method with previous works across 5 datasets that include scenes, objects, and human bodies, encompassing the following datasets. ShapeNet [4]", + "bbox": [ + 89, + 854, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/f6c7fbe43e16219a6e30f2505829e02f287298609e8cba653ab4d046d12f84e5.jpg", + "image_caption": [ + "Figure 4. The evaluation of our method, TriVol and PFGS trained with different point numbers on Car category. The legend in the lower right corner indicates different methods." + ], + "image_footnote": [], + "bbox": [ + 514, + 87, + 905, + 195 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "is a large-scale dataset that contains CAD models across various categories. GSO [9] is a high-quality dataset comprising 3D scanned household items. ScanNet [7] is a dataset consisting of real-scanned indoor scenes. DTU [16] is a multi-view stereo dataset featuring high-quality and high-density scenes. THuman2.0 [30] is a dataset comprising high-quality 3D models of human bodies. We select the chairs and tables from ShapeNet and shoes from GSO for object-level comparison. All settings are consistent with those of TriVol [13] and PFGS [28].", + "bbox": [ + 511, + 277, + 906, + 428 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Metrics. We evaluate the rendering results using three widely accepted metrics: PSNR, SSIM[29], and LIPIS[31]. Baselines. We employ traditional graphics rendering (GR), point-based novel view synthesis method NPBG++[23] and learning-based method TriVol[13] and PFGS[28] for comparison on different datasets.", + "bbox": [ + 511, + 429, + 906, + 518 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Evaluation of Rendering", + "text_level": 1, + "bbox": [ + 511, + 529, + 740, + 545 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The evaluation results of our method and baselines are presented in Table 2. The point numbers $N$ of input point clouds are 100K for objects and scenes, and 80K for human bodies. The results demonstrate that our method outperforms nearly all other methods across various metrics and categories. Our approach outperforms existing methods, showing average improvements of $2.60\\%$ , $1.05\\%$ , and $6.06\\%$ in PSNR, SSIM, and LIPIs across all datasets, respectively. For qualitative comparison, we represent the rendering results of different methods across several datasets in Figure 5. Our rendering results not only exhibit superior overall quality, but also display crisp and detailed elements, such as patterns, wheels, and human faces. Quantitative and qualitative comparisons demonstrate that our method achieves SOTA performance in point cloud rendering.", + "bbox": [ + 511, + 551, + 906, + 792 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Evaluation on Different Point Number", + "text_level": 1, + "bbox": [ + 511, + 803, + 844, + 816 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As shown in Table 2, we also display the evaluation results of our method trained with different input point numbers, including 20K and 40K. Despite utilizing fewer points, our method still achieves the best performance on object categories such as chairs, tables, and shoes. Note that our", + "bbox": [ + 511, + 825, + 905, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "27289", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/74ae7f8c770fface135dc89a53a4068b99441dea058f7192ddb8562d56aa23a9.jpg", + "image_caption": [ + "Figure 5. The rendering results of our method and previous methods on different categories. From top to bottom: scene, car, chair, shoe and human body." + ], + "image_footnote": [], + "bbox": [ + 91, + 88, + 908, + 676 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "method only utilizes $20\\%$ of the points as input, indicating that our method also yields good results for sparse point clouds. We also showcase the evaluation results for TriVol and PFGS when trained with varying point counts (2K, 10K, 20K, 40K, 100K) on the Car category to illustrate their performance with sparse point clouds. The comparisons are depicted in Figure 4. All methods exhibit improved performance as the number of points used during training increases, with our method outperforming all others across all input point numbers. The comparisons highlight the superiority of our method in rendering sparse point clouds.", + "bbox": [ + 89, + 723, + 485, + 891 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Evaluation of Generalization Capability", + "text_level": 1, + "bbox": [ + 511, + 722, + 856, + 739 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To evaluate the generalization capability of our $\\mathcal{N}_p$ on different categories with different input point numbers, we directly test our network on DTU dataset, which is trained on the Car category using 20K points. The setting here is the same as that in PFGS [28]. The results are presented in Table 3, where the methods are assessed using 20K, 40K, 100K and 1M points on the DTU dataset. We also present the evaluation results of PFGS when trained on DTU with 1M points and on the Car category of ShapeNet with 20K points, in order to compare the generalization ca", + "bbox": [ + 511, + 750, + 906, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "27290", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/8979cf7d439c699d3f47a9804385ee9e7e7b6d1f8fde6808496f46266875c906.jpg", + "table_caption": [ + "Table 2. The comparison of our method against previous methods on different datasets including scenes, objects and human bodies. The top three results are highlighted in red, orange, and yellow, respectively. †:The input point number for THuman2.0 is 80K for all methods." + ], + "table_footnote": [], + "table_body": "
MethodPoint NumberScanNet[7]Car (ShapeNet[4])Chair (ShapeNet[4])Shoe (GSO[9])THuman2.0[30]
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
GR100K†13.620.5280.77919.240.8140.18217.780.7790.20123.140.8290.15320.260.9050.337
NPBG++[23]100K16.810.6710.58525.320.8740.12025.780.9160.10129.420.9290.08126.810.9520.062
TriVol[13]100K18.560.7340.47327.220.9270.08428.850.9600.07831.240.9610.04525.970.9350.059
PFGS[28]100K19.860.7580.45227.340.9420.07727.520.9560.07829.530.9570.05834.740.9830.009
Ours20K18.570.7240.60027.880.9490.06828.890.9620.07730.910.9680.05133.700.9790.014
40K19.430.7430.55228.570.9570.06229.520.9670.07031.490.9740.04734.290.9810.012
100K20.240.7590.49028.730.9600.06029.100.9650.07532.080.9780.04235.430.9870.009
", + "bbox": [ + 93, + 128, + 903, + 291 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/d31a0a04768613b72f445825f2b030e9312109bec421b9437a0d2af2693659a4.jpg", + "table_caption": [ + "Table 3. The generalization capability of our method, traditional graphics rendering, and PFGS on the DTU dataset. Here, our method is trained on the Car category of ShapeNet using 20K points. Two models of PFGS are trained on the DTU dataset with 1M points and on the Car category of ShapeNet with 20K points, respectively. The first column records the input point numbers for evaluation. The top two results are highlighted in red and orange." + ], + "table_footnote": [], + "table_body": "
Point NumberPFGS DTU 1M pointsPFGS Car (ShapeNet) 20K points
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
20K10.1770.3130.6478.0650.2970.792
40K12.0260.3580.5508.1600.3020.765
100K17.9780.5540.3128.6470.3220.673
1M23.0010.7700.11213.6760.5490.401
Point NumberGROurs \nCar (ShapeNet) 20K points
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
20K9.1050.3160.62516.7130.5900.361
40K9.9850.3280.64116.6230.5820.315
100K12.0350.3630.66616.5440.5730.263
1M18.0030.5940.39616.6430.5940.216
", + "bbox": [ + 96, + 422, + 478, + 657 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "pability of PFGS. Like previous methods, PFGS requires entire point clouds as input for predicting Gaussians, hence these methods struggle to generalize to other datasets or varying point densities. Being trained with point cloud patches, our method does not depend on global priors and possesses superior generalization capabilities, resulting in enhanced performance on other datasets and with varying point densities. Figure 6 shows the results of methods on DTU dataset with different point numbers, where our rendering results preserve clear details. Note that previous methods are unable to function on sparse point clouds or on scene categories that are significantly different from object categories. Both quantitative and qualitative evaluations reveal the exceptional generalization capability of our", + "bbox": [ + 88, + 688, + 485, + 902 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/45c2d67f60af888019e01bfedba89a09eb09af02f582eec2d279709365838d48.jpg", + "image_caption": [ + "Figure 6. The evaluation results of our method and previous methods on the DTU dataset, utilizing 20K, 40K, 100K and 1M points, with all methods trained on the Car category with 20K points." + ], + "image_footnote": [], + "bbox": [ + 516, + 311, + 903, + 494 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "method.", + "bbox": [ + 513, + 575, + 571, + 589 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Visualization of Gaussians", + "text_level": 1, + "bbox": [ + 511, + 601, + 754, + 617 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Figure 7(a)(b) illustrates the predicted Gaussians and corresponding rendered images of our method and PFGS[28]. As mentioned in Section 2.1, PFGS first predicts Gaussians and then refines the rendered images. We visualize the Gaussians of its first stage. Different from our predicted Gaussians, which are evenly distributed with fewer gaps, the predicted Gaussians of PFGS are chaos. Its rendered image is not photo-realistic requiring refinement in the second stage, and thus cannot be considered a final rendering result.", + "bbox": [ + 511, + 625, + 906, + 762 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.6. Evaluation of Rendering Speed", + "text_level": 1, + "bbox": [ + 511, + 771, + 790, + 787 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4 represents the rendering speeds of different methods. Our method directly predicts Gaussians for rendering images, eliminating the need for further refinement once the prediction is complete. The rendering speed of our method is the same as 2DGS [15], thus our method achieves real-time rendering with the fastest rendering speed among previous methods.", + "bbox": [ + 511, + 794, + 908, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "27291", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f68cc7b16a78b6c1c92221e4a68ad3b33b9f24d1d91ad4ac465b4c6c2a10bd98.jpg", + "image_caption": [ + "Figure 7. The visualization of Gaussians and their rendered images predicted by our method, our ablated methods and PFGS. (a): PFGS trained on Car (100K). (b) and (c): Our method trained on Car (100K and 20K). (d), (e) and (f): Our ablated methods trained on Car (20K)." + ], + "image_footnote": [], + "bbox": [ + 93, + 88, + 480, + 309 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/517cd0d228b1bee68fcc2501f3949d7aeb6b33bb414dbb2ba9974d8cecc1ff0b.jpg", + "table_caption": [ + "Table 4. The rendering speeds of different methods, where the rendering frames per second (FPS) are presented. All methods are evaluated using an RTX 3090." + ], + "table_footnote": [], + "table_body": "
MethodNPBG++[23]TriVol[13]PFGS[28]Ours
Speed (FPS)37.451.623.80142.86
", + "bbox": [ + 94, + 465, + 478, + 505 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.7. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 530, + 243, + 546 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Split Number. We conduct experiments with different split numbers $K$ (1, 2, 4, 8) and different input point cloud numbers $N$ (2K, 10K, 20K, 40K) on the Car category of ShapeNet. The results are shown in Figure 8. Performance is enhanced with a growing number of splits, underscoring the efficacy of our split decoders, especially when it comes to sparse point clouds.", + "bbox": [ + 89, + 551, + 483, + 657 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Component Ablation. We first evaluate the effectiveness of our initialization in the 2D Gaussian prediction module. The visualization of initialized Gaussians is shown in Figure 3 (c), which indicates a coarse rendering result providing a great foundation for optimization. Without initialization, the performance of our method sharply drops as Table 5 shows. Randomly initialized Gaussians might not be visible in certain views, resulting in non-convergence, as shown in Figure 7 (d). The ablation studies of our entire-patch architecture contain two experiments: (1) removing the entire module $\\mathcal{N}_e$ and utilizing the predicted incomplete images rendered from point patches for supervision, denoted as “No $\\mathcal{N}_e$ ” in Table 5; (2) splitting the entire point cloud to multiple patches as same as the inference process of $\\mathcal{N}_p$ to render complete images during training, denoted as “ $\\mathcal{N}_p$ only” in Table 5. the “ $\\mathcal{N}_p$ only” method will consume a sub", + "bbox": [ + 89, + 659, + 483, + 901 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/b06d724f0d2fa884f7cf32187a6860ac3098355f003e9bdfa989ef9888191bd4.jpg", + "table_caption": [ + "Table 5. The ablation studies of our method. The ablated methods are trained on Car with 20K points." + ], + "table_footnote": [], + "table_body": "
PSNR ↑SSIM ↑LPIPS ↓
No Initialization9.620.6850.655
No Nc12.940.6760.441
Np only23.230.8290.126
Ours27.880.9490.068
", + "bbox": [ + 547, + 123, + 874, + 215 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7bb36793d49020206ecd95505b0cb2dd2232f2e8ed64e162bbfbd6442670419f.jpg", + "image_caption": [ + "Figure 8. The ablation study of split number $K$ with different input point numbers. The legend in the lower right corner indicates the input point number for training and evaluating." + ], + "image_footnote": [], + "bbox": [ + 514, + 238, + 903, + 348 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "stantial amount of resources for training and inference. The evaluation results of these two ablated methods are worse than our method, where their visualization results are shown in Figure 7(e)(f)(c).", + "bbox": [ + 511, + 416, + 905, + 477 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The results of ablation studies validate the effectiveness and necessity of the components within our method.", + "bbox": [ + 511, + 482, + 905, + 513 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion and Discussion", + "text_level": 1, + "bbox": [ + 511, + 537, + 763, + 553 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we introduce a novel approach featuring an entire-patch architecture and the 2D Gaussian prediction module for point cloud rendering. Our prediction module directly forecasts the 2D Gaussians from the provided point clouds, beginning with initialization derived from normals. The splitting decoders produce multiple Gaussians from a point, thereby better accommodating sparse point clouds. To enhance the generalization capability, the entire-patch architecture employs a module for \"background\" area prediction and another module for point cloud patch prediction, ensuring proper supervision. Comprehensive experiments and comparisons demonstrate the superiority and generalization capability of our method, which can render point clouds into photo-realistic images with clear details and achieves SOTA performance. A limitation of our method may arise when a portion of a point cloud is absent, as our method cannot render a complete image without the direct support of the points. Our future work aims to address this limitation by incorporating point cloud completion techniques and also attempting to render point clouds with noise, such as those directly scanned from wild and outdoor scenes.", + "bbox": [ + 511, + 568, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "27292", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 91, + 90, + 250, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 161, + 187, + 176 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Kara-Ali Aliev, Artem Sevastopolsky, Maria Kolos, Dmitry Ulyanov, and Victor Lempitsky. Neural point-based graphics. In Computer Vision - ECCV 2020, pages 696-712, Cham, 2020. Springer International Publishing. 1, 2", + "[2] M. Botsch, A. Hornung, M. Zwicker, and L. Kobbelt. High-quality surface splatting on today's gpus. In Proceedings Eurographics/IEEE VGTC Symposium Point-Based Graphics, 2005., pages 17-141, 2005. 1, 2", + "[3] Alexandre Boulch and Renaud Marlet. Poco: Point convolution for surface reconstruction. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6292-6304, 2022. 2, 4", + "[4] Angel X. Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, Jianxiong Xiao, Li Yi, and Fisher Yu. Shapenet: An information-rich 3d model repository, 2015. 5, 7", + "[5] Kaichun Mo Leonidas J. Guibas Charles R. Qi, Hao Su. Pointnet: Deep learning on point sets for 3d classification and segmentation. In 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2017. 4", + "[6] Paolo Cignoni, Marco Callieri, Massimiliano Corsini, Matteo Dellepiane, Fabio Ganovelli, and Guido Ranzuglia. MeshLab: an Open-Source Mesh Processing Tool. In Eurographics Italian Chapter Conference. The Eurographics Association, 2008. 2", + "[7] Angela Dai, Angel X. Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proc. Computer Vision and Pattern Recognition (CVPR), IEEE, 2017. 5, 7", + "[8] Peng Dai, Yinda Zhang, Zhuwen Li, Shuaicheng Liu, and Bing Zeng. Neural point cloud rendering via multi-plane projection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7830-7839, 2020. 1, 2", + "[9] Laura Downs, Anthony Francis, Nate Koenig, Brandon Kinman, Ryan Michael Hickman, Krista Reymann, Thomas Barlow McHugh, and Vincent Vanhoucke. Google scanned objects: A high-quality dataset of 3d scanned household items. 2022 International Conference on Robotics and Automation (ICRA), pages 2553-2560, 2022. 5, 7", + "[10] Antoine Guédon and Vincent Lepetit. Sugar: Surface-aligned gaussian splatting for efficient 3d mesh reconstruction and high-quality mesh rendering. CVPR, 2024. 3", + "[11] Hugues Hoppe, Tony DeRose, Tom Duchamp, John Alan McDonald, and Werner Stuetzle. Surface reconstruction from unorganized points. Proceedings of the 19th annual conference on Computer graphics and interactive techniques, 1992. 3" + ], + "bbox": [ + 93, + 186, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "This work was supported by the National Natural Science Foundation of China under Grant 62032011.", + "[12] Qingyong Hu, Bo Yang, Linhai Xie, Stefano Rosa, Yulan Guo, Zhihua Wang, Niki Trigoni, and Andrew Markham. Randla-net: Efficient semantic segmentation of large-scale point clouds. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2020. 2, 4", + "[13] Tao Hu, Xiaogang Xu, Ruihang Chu, and Jiaya Jia. Trivol: Point cloud rendering via triple volumes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20732-20741, 2023. 1, 2, 3, 5, 7, 8", + "[14] T. Hu, Xiaogang Xu, Shu Liu, and Jiaya Jia. Point2pix: Photo-realistic point cloud rendering via neural radiance fields. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8349-8358, 2023. 1, 2, 3", + "[15] Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2d gaussian splatting for geometrically accurate radiance fields. In SIGGRAPH 2024 Conference Papers. Association for Computing Machinery, 2024. 2, 3, 5, 7", + "[16] Rasmus Jensen, Anders Dahl, George Vogiatzis, Engil Tola, and Henrik Aanæs. Large scale multi-view stereopsis evaluation. In 2014 IEEE Conference on Computer Vision and Pattern Recognition, pages 406-413, 2014. 5", + "[17] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42 (4), 2023. 1, 2, 3", + "[18] Leif Kobbelt and Mario Botsch. A survey of point-based techniques in computer graphics. Comput. Graph., 28(6): 801-814, 2004. 2", + "[19] Xu Ma, Can Qin, Haoxuan You, Haoxi Ran, and Yun Fu. Rethinking network design and local geometry in point cloud: A simple residual mlp framework. *ICLR*, 2022. 4", + "[20] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 1, 2", + "[21] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Kopf, Edward Yang, Zachary DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. Pytorch: An imperative style, high-performance deep learning library. In Advances in Neural Information Processing Systems 32, pages 8024-8035. Curran Associates, Inc., 2019. 5", + "[22] Hanspeter Pfister, Matthias Zwicker, Jeroen van Baar, and Markus H. Gross. Surfels: surface elements as rendering primitives. Proceedings of the 27th annual conference on Computer graphics and interactive techniques, 2000. 1, 2", + "[23] Ruslan Rakhimov, Andrei-Timotei Ardelean, Victor Lempitsky, and Evgeny Burnaev. Npbg++: Accelerating neural point-based graphics. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15969-15979, 2022. 1, 2, 5, 7, 8", + "[24] Nikhila Ravi, Jeremy Reizenstein, David Novotny, Taylor Gordon, Wan-Yen Lo, Justin Johnson, and Georgia" + ], + "bbox": [ + 93, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "27293", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Gkioxari. Accelerating 3d deep learning with pytorch3d. arXiv:2007.08501, 2020. 2", + "[25] Liu Ren, Hanspeter Pfister, and Matthias Zwicker. Object space ewa surface splatting: A hardware accelerated approach to high quality point rendering. Computer Graphics Forum, 21(3):461-470, 2002. 2", + "[26] Darius Rückert, Linus Franke, and Marc Stamminger. Adop: Approximate differentiable one-pixel point rendering. ACM Transactions on Graphics (ToG), 41(4):1-14, 2022. 2", + "[27] Bane Sullivan and Alexander Kaszynski. PyVista: 3D plotting and mesh analysis through a streamlined interface for the Visualization Toolkit (VTK). Journal of Open Source Software, 4(37):1450, 2019. 2", + "[28] Jiaxu Wang, Ziyi Zhang, Junhao He, and Renjing Xu. Pfgs: High fidelity point cloud rendering via feature splatting. ECCV 2024, 2024. 1, 2, 3, 5, 6, 7, 8", + "[29] Zhou Wang, A.C. Bovik, H.R. Sheikh, and E.P. Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing, 13(4): 600-612, 2004. 5", + "[30] Tao Yu, Zerong Zheng, Kaiwen Guo, Pengpeng Liu, Qionghai Dai, and Yebin Liu. Function4d: Real-time human volumetric capture from very sparse consumer rgbd sensors. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR2021), 2021. 5, 7", + "[31] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 5", + "[32] Qian-Yi Zhou, Jaesik Park, and Vladlen Koltun. Open3D: A modern library for 3D data processing. arXiv:1801.09847, 2018. 2, 3", + "[33] Matthias Zwicker, Hanspeter Pfister, Jeroen Baar, and Markus Gross. Surface splattering. Proceedings of the ACM SIGGRAPH Conference on Computer Graphics, 2001, 2001. 2" + ], + "bbox": [ + 91, + 90, + 482, + 585 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "27294", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Sparse Point Cloud Patches Rendering via Splitting 2D Gaussians/dd278d9d-da45-4b1e-bcc2-33b0924a1fa1_model.json b/2025/Sparse Point Cloud Patches Rendering via Splitting 2D Gaussians/dd278d9d-da45-4b1e-bcc2-33b0924a1fa1_model.json new file mode 100644 index 0000000000000000000000000000000000000000..9d393b4733eb1e941e567159157894895802767f --- /dev/null +++ b/2025/Sparse Point Cloud Patches Rendering via Splitting 2D Gaussians/dd278d9d-da45-4b1e-bcc2-33b0924a1fa1_model.json @@ -0,0 +1,2046 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.167, + 0.131, + 0.83, + 0.154 + ], + "angle": 0, + "content": "Sparse Point Cloud Patches Rendering via Splitting 2D Gaussians" + }, + { + "type": "text", + "bbox": [ + 0.165, + 0.18, + 0.834, + 0.218 + ], + "angle": 0, + "content": "Changfeng Ma\\(^{1}\\), Ran Bi\\(^{1}\\), Jie Guo\\(^{1}\\), Chongjun Wang\\(^{1}\\), Yanwen Guo\\(^{12*}\\) \n\\(^{1}\\)Nanjing University, Nanjing, China \n\\(^{2}\\)School of Software, North University of China" + }, + { + "type": "text", + "bbox": [ + 0.319, + 0.22, + 0.68, + 0.234 + ], + "angle": 0, + "content": "{changfengma, 211250233}@smail.nju.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.352, + 0.237, + 0.642, + 0.252 + ], + "angle": 0, + "content": "{guojie, chjwang, ywguo}@nju.edu.cn" + }, + { + "type": "image_caption", + "bbox": [ + 0.145, + 0.264, + 0.169, + 0.275 + ], + "angle": 0, + "content": "Car" + }, + { + "type": "image", + "bbox": [ + 0.139, + 0.275, + 0.236, + 0.344 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.153, + 0.347, + 0.217, + 0.358 + ], + "angle": 0, + "content": "Point Cloud" + }, + { + "type": "image_caption", + "bbox": [ + 0.248, + 0.264, + 0.3, + 0.274 + ], + "angle": 0, + "content": "NPBG+±" + }, + { + "type": "image", + "bbox": [ + 0.252, + 0.273, + 0.341, + 0.336 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.245, + 0.337, + 0.345, + 0.363 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.356, + 0.264, + 0.394, + 0.275 + ], + "angle": 0, + "content": "TriVol" + }, + { + "type": "image", + "bbox": [ + 0.361, + 0.273, + 0.449, + 0.336 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.354, + 0.337, + 0.455, + 0.363 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.139, + 0.367, + 0.173, + 0.377 + ], + "angle": 0, + "content": "PFGS" + }, + { + "type": "image", + "bbox": [ + 0.145, + 0.376, + 0.233, + 0.439 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.136, + 0.44, + 0.236, + 0.466 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.248, + 0.368, + 0.276, + 0.377 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.253, + 0.376, + 0.341, + 0.439 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.245, + 0.44, + 0.346, + 0.466 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.357, + 0.368, + 0.377, + 0.377 + ], + "angle": 0, + "content": "GT" + }, + { + "type": "image", + "bbox": [ + 0.362, + 0.373, + 0.449, + 0.438 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.352, + 0.44, + 0.455, + 0.466 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.584, + 0.263, + 0.774, + 0.275 + ], + "angle": 0, + "content": "Generalize to Other Categories" + }, + { + "type": "image", + "bbox": [ + 0.49, + 0.275, + 0.627, + 0.378 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.487, + 0.379, + 0.627, + 0.466 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.629, + 0.275, + 0.758, + 0.393 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.631, + 0.393, + 0.707, + 0.405 + ], + "angle": 0, + "content": "Sculpture" + }, + { + "type": "image", + "bbox": [ + 0.63, + 0.403, + 0.707, + 0.465 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.762, + 0.275, + 0.861, + 0.33 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.764, + 0.332, + 0.861, + 0.393 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.711, + 0.394, + 0.801, + 0.465 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.804, + 0.394, + 0.862, + 0.429 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.805, + 0.43, + 0.862, + 0.465 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.478, + 0.908, + 0.507 + ], + "angle": 0, + "content": "Figure 1. Our method excels in rendering intricate, photo-realistic images from point clouds of different categories. In this instance, our model is trained on the Car category, utilizing 20K points as input." + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.529, + 0.327, + 0.545 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.561, + 0.486, + 0.879 + ], + "angle": 0, + "content": "Current learning-based methods predict NeRF or 3D Gaussians from point clouds to achieve photo-realistic rendering but still depend on categorical priors, dense point clouds, or additional refinements. Hence, we introduce a novel point cloud rendering method by predicting 2D Gaussians from point clouds. Our method incorporates two identical modules with an entire-patch architecture enabling the network to be generalized to multiple datasets. The module normalizes and initializes the Gaussians utilizing the point cloud information including normals, colors and distances. Then, splitting decoders are employed to refine the initial Gaussians by duplicating them and predicting more accurate results, making our methodology effectively accommodate sparse point clouds as well. Once trained, our approach exhibits direct generalization to point clouds across different categories. The predicted Gaussians are employed directly for rendering without additional refinement on the rendered images, retaining the benefits of 2D Gaussians. We conduct extensive experiments on various datasets, and the results demonstrate the superiority and generalization of our method, which achieves SOTA performance. The code is" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.531, + 0.899, + 0.546 + ], + "angle": 0, + "content": "available at https://github.com/murcherful/GauPCRender." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.574, + 0.645, + 0.589 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.599, + 0.907, + 0.705 + ], + "angle": 0, + "content": "Point clouds can be easily acquired using various types of 3D scanners, making them a recent research hotspot. Photorealistic rendering of point clouds is significant for many applications such as visualization, virtual reality and automatic driving. However, the sparsity and discrete nature of point clouds continue to pose challenges for achieving photo-realistic rendering." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.705, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Traditional graphics rendering methods [2, 22] merely project points as planes or spheres onto images. Consequently, their outcomes lack photo-realism and are riddled with numerous holes. Previous deep learning-based works [1, 8, 23] necessitate point clouds and corresponding images from several views as input, which are difficult to acquire in practice. These methods also require training or fine-tuning for each scene or object. Recent works [13, 14] have focused on predicting NeRF [20] from given point clouds to render images from arbitrary views. PFGS [28] predicts the 3D Gaussians [17] with features for point clouds. After rendering, PFGS refines the rendered images based on these features. Although these methods can render point clouds" + }, + { + "type": "page_footnote", + "bbox": [ + 0.116, + 0.888, + 0.244, + 0.901 + ], + "angle": 0, + "content": "*Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "27285" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.482, + 0.187 + ], + "angle": 0, + "content": "Table 1. The comparison of different methods across several dimensions. Here, \"S. S/O.\", \"S. C.\" and \"M. C.\" represent \"single scene or object\", \"single category\" and \"multiple categories\", respectively. \\(\\mathcal{P}\\) and \\(\\mathcal{I}\\) indicate point clouds and images, separately. \\(\\dagger\\): Traditional graphics rendering approach is unable to render photo-realistic images. \\(\\ddagger\\): PFGS requires further refinement on images rendered from 3D gaussians." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.192, + 0.485, + 0.365 + ], + "angle": 0, + "content": "
MethodGeneralization abilityInputOutputPre- Fine- trainFine- tunePoint number
\\( GR^† \\)AnyPI✘✘Any
NPBG [1]S. S/O.\\( P+I \\)I✓✓100K
NPCR [8]S. S/O.\\( P+I \\)I✘✓100K
NPBG++ [23]S. S/O.\\( P+I \\)I✓✘100K
TriVol [13]S. C.PNeRF[20]✓✘100K
Point2Pix [14]S. C.PNeRF[20]✓✘100K
PFGS [28]S. C.P3DGS[17]‡✓✘80K
OursM. C.P2DGS[15]✓✘2K-100K
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.386, + 0.483, + 0.431 + ], + "angle": 0, + "content": "without training or fine-tuning on each data, they still have limitations, such as the need for dense point clouds, poor generalization, slow rendering speeds, and blurry results." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.432, + 0.483, + 0.538 + ], + "angle": 0, + "content": "2D Gaussian Splatting [15] has recently been proposed, which includes explicit normals of Gaussians. These normals facilitate easy initialization, learning, and prediction by the network from point clouds. Motivated by this, we introduce a novel method to predict the 2D Gaussians of the given point cloud and rasterize the predicted 2D Gaussians into images for photo-realistic point cloud rendering." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.539, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Due to taking entire point clouds as input, previous methods are hard to generalize to other categories or adapt to different point distributions. Processing only point cloud patches with the network can enhance its generalization capability and reduce dependence on category priors [3, 12]. However, rendering Gaussians predicted from patch points results in incomplete images that are unsupervisible. Therefore, we employ an entire-patch architecture that includes two identical 2D Gaussian prediction modules to handle entire and patch point clouds, separately. We use the rendering results of the entire point cloud as a \"background\" to obtain a complete image for proper supervision. The module first normalizes and initializes the 2D Gaussians based on the given point clouds, encompassing the estimated normals, scales, and colors, which are crucial for our network to converge and predict more accurate results. Subsequently, a point cloud encoder and multiple splitting decoders are employed to extract local features and predict the parameters of the duplicated Gaussians. Equipped with the splitting decoder, our network can handle sparse point clouds and produce high-quality details with dense Gaussians. Thanks to the proposed architecture, our method also exhibits strong generalization across scenes and objects. Without further refinement on rendered images, our" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.122 + ], + "angle": 0, + "content": "method retains the advantages of 2D Gaussian Splatting in rendering, such as rapid rendering speed." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.123, + 0.905, + 0.213 + ], + "angle": 0, + "content": "We conduct extensive comparative and ablation experiments, and the results demonstrate the superiority of our method, which can render point clouds into high-quality detailed images with excellent generalization. We will make our dataset and codes public in the future. Our main contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.215, + 0.905, + 0.275 + ], + "angle": 0, + "content": "- We introduce a method that directly predicts 2D Gaussians from point clouds to render photo-realistic images without any refinement, while retaining the advantages of 2D Gaussian Splatting." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.276, + 0.905, + 0.32 + ], + "angle": 0, + "content": "- Equipped with an entire-patch architecture, our method can handle point cloud patches, leading to significant generalization across scenes and objects." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.321, + 0.905, + 0.366 + ], + "angle": 0, + "content": "- We introduce the splitting decoder, which splits Gaussians into denser outcomes, enabling our method to render detailed images from even sparse point clouds." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.215, + 0.905, + 0.366 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.382, + 0.655, + 0.399 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.409, + 0.729, + 0.425 + ], + "angle": 0, + "content": "2.1. Point Cloud Rendering" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.432, + 0.905, + 0.598 + ], + "angle": 0, + "content": "Traditional Point Cloud Rendering. Point cloud rendering has consistently been a significant research topic [18]. Conventional point cloud rendering techniques [6, 24, 27, 32] directly rasterize points onto 2D screens for rendering, yet they grapple with the issue of holes. Subsequent studies [2, 22, 25, 33] have proposed splatting points with elliptic discs, ellipsoids, or surfels for point cloud rendering to mitigate the holes observed during the rendering process. Although these graphic rendering methods can render any point cloud, the resulting images often contain holes and exhibit blurry details, lacking photo-realism." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.6, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Point-based novel view synthesis. For photo-realistic point cloud rendering, several recently proposed methods utilize corresponding images of point clouds as additional information. NPBG [1] signs descriptors to point clouds and rasterizes into multi-resolution raw images with features and employs a U-Net to render the final results. Dai et al. [8] aggregate points with features onto multiple planes based on a given view and point cloud. Following multi-plane-based voxelization, they employ a 3D U-Net to predict the rendered image. ADOP [26] is introduced for rendering HD images from point clouds, which employs several global parameters, a neural renderer and a differentiable physically-based tonemapper for optimization. These methods necessitate training for each point cloud with images from several views to render images of novel views, which is a time-consuming process. To reduce training time, NPBG++ [23] is trained on a collection of point clouds using a network that extracts features from corresponding images. This network aligns the features of input images to point clouds without the need for fine-tuning. However, this" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.521, + 0.957 + ], + "angle": 0, + "content": "27286" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.121 + ], + "angle": 0, + "content": "method still requires images as input, which can be challenging to obtain in practice." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.122, + 0.483, + 0.394 + ], + "angle": 0, + "content": "Learning-based Point Cloud Rendering. Several works [13, 14, 28] are proposed for rendering point clouds without the need for training or fine-tuning on each point cloud and corresponding images. TriVol [13] utilizes triple-plane-based grouping to query point features for predicting a NeRF from a given point cloud. Hu et al. [14] propose multi-scale radiance fields to predict NeRFs from point clouds. They also introduce a sampling strategy to accelerate fine-tuning and rendering speed. PFGS [28], a two-stage approach, for the first time employs feature splatting to predict 3D Gaussians from point clouds and uses a recurrent decoder to refine the rendered images. These methods are capable of rendering point clouds without fine-tuning or the need for input images, yet they still depend on categorical priors, dense points, or further refinements. Table 1 summarizes the distinctions between our method and the previous methods across several dimensions, highlighting the innovation and superiority of our proposed approach." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.403, + 0.273, + 0.419 + ], + "angle": 0, + "content": "2.2. Gaussian Splitting" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.424, + 0.483, + 0.726 + ], + "angle": 0, + "content": "3D Gaussian Splitting [17] is proposed with remarkable performance in novel view synthesis and further explores the rendering capability of the point clouds. 3DGS utilizes the differentiable splatting rendering to optimize 3D Gaussian primitives initialized from point clouds. Scenes are explicitly represented with 3D Gaussians that can be rendered with remarkable speed. Meshes can also be reconstructed from 3D Gaussians[10]. However, the normals of 3D Gaussians are so vague that the resulting geometries lack precision. To generate more accurate geometries, Hua et al. [15] propose 2DGS that employs 2D Gaussian disks containing explicit normals as representation primitives. With precise normals, meshes can accurately reconstruct the geometry of objects. Given that the normals of 2D Gaussians can be readily determined, 2D Gaussians can also be initialized using normals. Luckily, the normals of points can be effortlessly estimated [11] facilitating the initialization and forecasting of 2D Gaussians. Motivated by this, we opt to predict 2D Gaussians in place of 3D Gaussians from the provided point clouds." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.74, + 0.182, + 0.755 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.482, + 0.856 + ], + "angle": 0, + "content": "Given a point cloud \\(\\mathcal{P}\\) and a camera \\(\\mathcal{C}\\), point cloud rendering aims to render a photo-realistic image \\(\\mathcal{I}\\). Here, point cloud \\(\\mathcal{P}\\) contains of coordinates of 3D points \\(\\mathbf{P}_{\\mathbf{x}} \\in \\mathbb{R}_{N \\times 3}\\) and colors of points \\(\\mathbf{P}_{\\mathbf{c}} \\in [0,1]_{N \\times 3}\\), where \\(N\\) indicates the point number of \\(\\mathcal{P}\\). Camera \\(\\mathcal{C}\\) consists of intrinsic matrix \\(\\mathbf{K} \\in \\mathbb{R}_{3 \\times 3}\\) and extrinsic matrix \\(\\mathbf{P} \\in \\mathbb{R}_{4 \\times 4}\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.857, + 0.483, + 0.901 + ], + "angle": 0, + "content": "For photo-realistic point cloud rendering, we propose a network \\(\\mathcal{N}\\) to predict 2D Gaussians \\(\\mathcal{G} = \\mathcal{N}(\\mathcal{P})\\) given point cloud \\(\\mathcal{P}\\), and employ differentiable splatting rendering [15]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.227 + ], + "angle": 0, + "content": "\\(f_{R}\\) to synthesize images \\(\\mathcal{I} = f_{R}(\\mathcal{G},\\mathcal{C})\\). The parameters of a 2D Gaussian \\(\\mathbf{g} = \\{\\mathbf{x},\\mathbf{s},o,\\mathbf{c},\\mathbf{n},\\alpha \\} \\in \\mathcal{G}\\) include: (1) position \\(\\mathbf{x}\\in \\mathbb{R}_3\\), (2) scale \\(\\mathbf{s}\\in \\mathbb{R}_2\\), (3) opacity \\(o\\in [0,1]\\), (4) spherical harmonic of color \\(\\mathbf{c}\\in \\mathbb{R}_d\\), (5) normal \\(\\mathbf{n}\\in \\mathbb{R}_3\\) and (6) rotation angle along normal \\(\\alpha \\in [0,2\\pi ]\\), where \\(d\\) represents the parameter number of spherical harmonic for color. As shown in Figure 2, our network contains two identical 2D Gaussian prediction modules (3.1) with an entire-patch architecture (3.2)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.239, + 0.795, + 0.253 + ], + "angle": 0, + "content": "3.1.2D Gaussian Prediction Module" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.262, + 0.766, + 0.275 + ], + "angle": 0, + "content": "3.1.1. Initialization of 2D Gaussians" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.28, + 0.905, + 0.841 + ], + "angle": 0, + "content": "We first introduce 2D Gaussian prediction module to predict the 2D Gaussians \\(\\mathcal{G}\\) given a point cloud \\(\\mathcal{P}\\) as input. The positions of the point cloud \\(\\mathcal{P}\\) are normalized to the range \\([-1,1]\\) based on the geometric center \\(\\mathbf{c} \\in \\mathbb{R}_3\\) and scale \\(s \\in R\\) of the input point cloud to facilitate easier feature extraction and prediction of results, where the scale is defined as the maximum distance between any point and \\(\\mathbf{c}\\). Our module then initializes one 2D Gaussian \\(\\mathbf{g}\\) for each normalized point \\(\\mathbf{p} = \\{\\mathbf{p}_{\\mathbf{x}}, \\mathbf{p}_{\\mathbf{c}}\\} \\in \\mathcal{P}\\), where \\(\\mathbf{p}_{\\mathbf{x}} \\in \\mathcal{R}_3\\) and \\(\\mathbf{p}_{\\mathbf{c}} \\in \\mathcal{R}_3\\) indicate the position and color of the point respectively. The position \\(\\mathbf{x}\\) of the Gaussian is initialized with \\(\\mathbf{p}_{\\mathbf{x}}\\). The spherical harmonic \\(\\mathbf{c}\\) and scale \\(\\mathbf{s}\\) are initialized as the same as 2DGS [15] according to \\(\\mathbf{p}_{\\mathbf{c}}\\) and the minimum distance between points, separately. The opacity \\(o\\) is set to 1 to ensure that all Gaussians are visible. As we mentioned before, the geometry of given point cloud \\(\\mathcal{P}\\) provides important clues for the rotation of Gaussians. Gaussian Splatting works [15, 17] utilize quaternions to represent the rotation of Gaussians and only randomly initialize them as Figure 3(b) shows, which are difficult for networks to predict intuitively. Fortunately, the normals of given point cloud can be easily estimated. Therefore, we represent the rotation of the Gaussian in terms of the normal \\(\\mathbf{n}\\) and the rotation angle \\(\\alpha\\) around that normal, and initialize the normal of the Gaussian using the estimated normal of the point cloud. Figure 3 (a) and (c) illustrate this initialization process. Specifically, we estimate the normals for \\(\\mathcal{P}\\) [11, 32], where the normal \\(\\hat{\\mathbf{n}}\\) for each point \\(\\mathbf{p}\\) is utilized to initialize the normal \\(\\mathbf{n}\\) of Gaussian. And the rotation angle \\(\\alpha\\) of Gaussian is set to 0. The estimated normal may not be precise, but it provides crucial information and ensures that each Gaussian is visible, thereby guaranteeing that the network can converge correctly. After initializing for all the points, the initialized Gaussians \\(\\mathcal{G}^i\\) composing of the initialized positions, normals, scales, spherical harmonics, opacities and rotation angles (\\(\\mathbf{X}^i, \\mathbf{N}^i \\in \\mathbb{R}_{N\\times 3}, \\mathbf{S}^i \\in \\mathbb{R}_{N\\times 2}, \\mathbf{C}^i \\in \\mathbb{R}_{N\\times 3}, O^i, A^i \\in \\mathbb{R}_N\\)) is passed to our network as input and predicting basis." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.843, + 0.682, + 0.857 + ], + "angle": 0, + "content": "3.1.2. Splitting Decoder" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.857, + 0.905, + 0.901 + ], + "angle": 0, + "content": "As shown in Figure 2, local features are required before splitting and predicting 2D Gaussians. To extract features from the initialized 2D Gaussians, we employ" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "27287" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.092, + 0.638, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.645, + 0.092, + 0.9, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.245, + 0.9, + 0.337 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.353, + 0.905, + 0.381 + ], + "angle": 0, + "content": "Figure 2. Overview of our proposed method. Our method predicts 2D Gaussians for point cloud rendering, employing an entire-patch architecture (bottom) and the 2D Gaussian Prediction Module (top left) with splitting decoders (top right)." + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.393, + 0.481, + 0.525 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.533, + 0.483, + 0.601 + ], + "angle": 0, + "content": "Figure 3. The illustration of our initialization approach, where the left side of each picture is a 2D schematic diagram and the right side is the rendered images of Gaussians. (a) The estimated normals of the point cloud. (b) Randomly Initialization. (c) Our Initialization. (d) Predicted 2D Gaussians." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.609, + 0.483, + 0.669 + ], + "angle": 0, + "content": "PointMLP[19] as the encoder \\(E\\), which is a simple residual MLP framework for point cloud learning with great performance on many tasks. The local feature \\(\\mathbf{F}_l\\) is extracted by:" + }, + { + "type": "equation", + "bbox": [ + 0.202, + 0.669, + 0.371, + 0.686 + ], + "angle": 0, + "content": "\\[\n\\mathbf {F} _ {l} = E (\\mathbf {X} ^ {i}, \\mathbf {C} ^ {i}, \\mathbf {N} ^ {i}, \\mathbf {S} ^ {i}).\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.692, + 0.483, + 0.873 + ], + "angle": 0, + "content": "Subsequently, \\(\\mathbf{F}_l\\) is augmented with \\(\\mathbf{X}^i\\), \\(\\mathbf{C}^i\\), \\(\\mathbf{N}^i\\), \\(\\mathbf{S}^i\\) as the input features for the splitting decoders. Our splitting decoders \\(D\\) also utilize \\(\\mathcal{G}^i\\) as the basis for prediction and forecast 2D Gaussians \\(\\mathcal{G}^p\\) for rendering purposes. There are six splitting decoders \\(D_x, D_s, D_c, D_n, D_\\alpha, D_o\\) responsible for predicting the six parameters \\(X^p, S^p, C^p, N^p, A^p, O^p\\) of 2D Gaussians. Each decoder contains a weight-shared MLP [5] as shown in Figure 2. The splitting and predicting process for the 6 parameters is basically the same, except for slight differences in shapes. Take the process of predicting the position \\(X^p\\) as an example. The decoder \\(D_x\\) predicts \\(K\\) shifts \\(\\Delta^x\\) based on the position \\(\\mathbf{X}^i\\):" + }, + { + "type": "equation", + "bbox": [ + 0.138, + 0.884, + 0.435, + 0.902 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\Delta} _ {1} ^ {x}, \\boldsymbol {\\Delta} _ {2} ^ {x},..., \\boldsymbol {\\Delta} _ {K} ^ {x} = D _ {x} (\\mathbf {F} _ {l}, \\mathbf {X} ^ {i}, \\mathbf {C} ^ {i}, \\mathbf {N} ^ {i}, \\mathbf {S} ^ {i}),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.395, + 0.905, + 0.456 + ], + "angle": 0, + "content": "where \\(K\\) indicates the number of splits. The \\(K\\) shifts are added to the same \\(\\mathbf{X}^i\\) that is \"split\" the initial position \\(\\mathbf{X}^i\\) to \\(K\\) new positions. The predicted position \\(\\mathbf{X}^p\\in \\mathbb{R}_{K\\cdot N\\times 3}\\) is the combination of these \\(K\\) new positions:" + }, + { + "type": "equation", + "bbox": [ + 0.633, + 0.467, + 0.784, + 0.51 + ], + "angle": 0, + "content": "\\[\n\\mathbf {X} ^ {p} = \\bigcup_ {j = 1} ^ {K} \\left(\\boldsymbol {\\Delta} _ {j} ^ {x} + \\mathbf {X} ^ {i}\\right).\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.522, + 0.907, + 0.673 + ], + "angle": 0, + "content": "After applying all the splitting decoders, we can obtain the parameters of \\( K \\cdot N \\) Gaussians \\( \\mathcal{G}^p \\). With the splitting decoders, our method can manage sparse point clouds by augmenting the number of Gaussians, serving as a form of point cloud upsampling. A greater number of Gaussians can also capture more complex features, enhancing the details of rendered images. The predicted 2D Gaussians are based on normalized points, hence they need to be de-normalized to achieve accurate rendering results. The de-normalization is applied to the predicted position \\( X^p \\) and scale \\( S^p \\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.683, + 0.75, + 0.697 + ], + "angle": 0, + "content": "3.2. Entire-Patch Architecture" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.907, + 0.901 + ], + "angle": 0, + "content": "However, the 2D Gaussian prediction module can only be trained on complete point clouds, which leads to poor generalization across different categories with varying point distributions, as we only have ground truth images of entire objects for supervision. Numerous methods for other point cloud tasks, including surface reconstruction[3] and semantic segmentation[12], utilize point cloud patches as processing units to enhance the generalization capabilities of their networks. However, the corresponding ground truth images of point cloud patches are unavailable in practice, which makes training on patches impossible. To facilitate the application of our method to point cloud patches, we utilize an entire-patch architecture, as illustrated at the bottom of Fig-" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "27288" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.485 + ], + "angle": 0, + "content": "ure 2. Our approach encompasses two 2D Gaussian prediction modules, \\(\\mathcal{N}_e\\) and \\(\\mathcal{N}_p\\), which operate on the entire point cloud and the point cloud patch, respectively. Given an entire point cloud \\(\\mathcal{P}_e\\) for rendering, our method randomly selects a point as the center point of a point cloud patch. The point cloud patch \\(\\mathcal{P}_p\\) is obtained by gathering \\(N_p\\) nearest points around the center point. The 2D Gaussians of point cloud \\(\\mathcal{P}_e\\) and patch \\(\\mathcal{P}_p\\) are predicted by two prediction modules, where \\(\\mathcal{G}_e = \\mathcal{N}_e(\\mathcal{P}_e)\\) and \\(\\mathcal{G}_p = \\mathcal{N}_p(\\mathcal{P}_p)\\). The non-patch part \\(\\mathcal{G}_e'\\) of \\(\\mathcal{G}_e\\) are the 2D Gaussians corresponding to the non-patch points \\(\\mathcal{P}_e' = \\mathcal{P}_e - \\mathcal{P}_p\\). The final 2D Gaussians \\(\\mathcal{G} = \\mathcal{G}_e' \\cup \\mathcal{G}_p\\) are used to render images for supervision during training. Here, \\(\\mathcal{G}_e'\\) serves as the \"background\" ensuring that the rendered image is complete for calculating the loss function and \\(\\mathcal{N}_p\\) can be properly trained. In practice, we initially train \\(\\mathcal{N}_e\\) separately and then freeze \\(\\mathcal{N}_e\\) while training \\(\\mathcal{N}_p\\) for our method. Both \\(\\mathcal{N}_e\\) and \\(\\mathcal{N}_p\\) can be utilized for inference. The inference of \\(\\mathcal{N}_e\\) directly takes \\(\\mathcal{P}_e\\) as input and outputs the entire 2D Gaussians for rendering. For \\(\\mathcal{N}_p\\), our method divides the point cloud into several patches. To ensure that all points of the input point cloud are covered, we repeatedly perform the process of randomly selecting a central point and extracting surrounding points as the patch point cloud, continuing until all points are included in at least one patch. The 2D Gaussian of the entire point cloud is the combination of the predictions from all patches." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.494, + 0.253, + 0.51 + ], + "angle": 0, + "content": "3.3. Training Details" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.516, + 0.484, + 0.595 + ], + "angle": 0, + "content": "Given \\(N_{c}\\) cameras \\(\\mathcal{C}_1,\\ldots ,\\mathcal{C}_{N_c}\\), the rendered images \\(\\mathcal{I}_1^p,\\dots ,\\mathcal{I}_{N_c}^p\\) is obtained by \\(f_{R}(\\mathcal{G},\\mathcal{C}_{1}),\\ldots ,f_{R}(\\mathcal{G},\\mathcal{C}_{N_{c}})\\). Similar to [15], we employ MSE and SSIM between rendered images and the corresponding ground truth images \\(\\mathcal{I}_1^g,\\dots ,\\mathcal{I}_{N_c}^g\\) as loss functions for supervision:" + }, + { + "type": "equation", + "bbox": [ + 0.09, + 0.606, + 0.484, + 0.648 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\frac {1}{N _ {c}} \\sum_ {i = 1} ^ {N _ {c}} \\left(\\beta \\mathcal {L} _ {M S E} (\\mathcal {I} _ {i} ^ {p}, \\mathcal {I} _ {i} ^ {g}) + (1 - \\beta) \\mathcal {L} _ {S S I M} (\\mathcal {I} _ {i} ^ {p}, \\mathcal {I} _ {i} ^ {g})\\right).\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.658, + 0.484, + 0.794 + ], + "angle": 0, + "content": "We implement our method with PyTorch [21]. The split number \\( K \\) is set to 4. The point number \\( N_{p} \\) of a patch is 2048. \\( \\beta \\) is set to 0.8 for the loss function. The number of rendered images \\( N_{c} \\) is 8. The resolutions of images are 256x256, 640x512 and 512x512 for objects, scenes and human bodies. We employ Adam optimizer whose learning rate is \\( 1.0 \\times 10^{-4} \\). The batch size is set to 8, and the maximum epoch of training is 480. Training our method takes about 30 hours with a RTX 3090 GPU." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.808, + 0.224, + 0.825 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.833, + 0.19, + 0.849 + ], + "angle": 0, + "content": "4.1. Settings" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Dataset. We compare our method with previous works across 5 datasets that include scenes, objects, and human bodies, encompassing the following datasets. ShapeNet [4]" + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.088, + 0.906, + 0.196 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.209, + 0.907, + 0.251 + ], + "angle": 0, + "content": "Figure 4. The evaluation of our method, TriVol and PFGS trained with different point numbers on Car category. The legend in the lower right corner indicates different methods." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.278, + 0.907, + 0.429 + ], + "angle": 0, + "content": "is a large-scale dataset that contains CAD models across various categories. GSO [9] is a high-quality dataset comprising 3D scanned household items. ScanNet [7] is a dataset consisting of real-scanned indoor scenes. DTU [16] is a multi-view stereo dataset featuring high-quality and high-density scenes. THuman2.0 [30] is a dataset comprising high-quality 3D models of human bodies. We select the chairs and tables from ShapeNet and shoes from GSO for object-level comparison. All settings are consistent with those of TriVol [13] and PFGS [28]." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.43, + 0.907, + 0.52 + ], + "angle": 0, + "content": "Metrics. We evaluate the rendering results using three widely accepted metrics: PSNR, SSIM[29], and LIPIS[31]. Baselines. We employ traditional graphics rendering (GR), point-based novel view synthesis method NPBG++[23] and learning-based method TriVol[13] and PFGS[28] for comparison on different datasets." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.53, + 0.741, + 0.546 + ], + "angle": 0, + "content": "4.2. Evaluation of Rendering" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.552, + 0.907, + 0.793 + ], + "angle": 0, + "content": "The evaluation results of our method and baselines are presented in Table 2. The point numbers \\(N\\) of input point clouds are 100K for objects and scenes, and 80K for human bodies. The results demonstrate that our method outperforms nearly all other methods across various metrics and categories. Our approach outperforms existing methods, showing average improvements of \\(2.60\\%\\), \\(1.05\\%\\), and \\(6.06\\%\\) in PSNR, SSIM, and LIPIs across all datasets, respectively. For qualitative comparison, we represent the rendering results of different methods across several datasets in Figure 5. Our rendering results not only exhibit superior overall quality, but also display crisp and detailed elements, such as patterns, wheels, and human faces. Quantitative and qualitative comparisons demonstrate that our method achieves SOTA performance in point cloud rendering." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.804, + 0.845, + 0.818 + ], + "angle": 0, + "content": "4.3. Evaluation on Different Point Number" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.906, + 0.901 + ], + "angle": 0, + "content": "As shown in Table 2, we also display the evaluation results of our method trained with different input point numbers, including 20K and 40K. Despite utilizing fewer points, our method still achieves the best performance on object categories such as chairs, tables, and shoes. Note that our" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "27289" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.089, + 0.909, + 0.677 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.688, + 0.908, + 0.717 + ], + "angle": 0, + "content": "Figure 5. The rendering results of our method and previous methods on different categories. From top to bottom: scene, car, chair, shoe and human body." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.724, + 0.486, + 0.892 + ], + "angle": 0, + "content": "method only utilizes \\(20\\%\\) of the points as input, indicating that our method also yields good results for sparse point clouds. We also showcase the evaluation results for TriVol and PFGS when trained with varying point counts (2K, 10K, 20K, 40K, 100K) on the Car category to illustrate their performance with sparse point clouds. The comparisons are depicted in Figure 4. All methods exhibit improved performance as the number of points used during training increases, with our method outperforming all others across all input point numbers. The comparisons highlight the superiority of our method in rendering sparse point clouds." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.723, + 0.857, + 0.74 + ], + "angle": 0, + "content": "4.4. Evaluation of Generalization Capability" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.907, + 0.903 + ], + "angle": 0, + "content": "To evaluate the generalization capability of our \\(\\mathcal{N}_p\\) on different categories with different input point numbers, we directly test our network on DTU dataset, which is trained on the Car category using 20K points. The setting here is the same as that in PFGS [28]. The results are presented in Table 3, where the methods are assessed using 20K, 40K, 100K and 1M points on the DTU dataset. We also present the evaluation results of PFGS when trained on DTU with 1M points and on the Car category of ShapeNet with 20K points, in order to compare the generalization ca" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "27290" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.908, + 0.119 + ], + "angle": 0, + "content": "Table 2. The comparison of our method against previous methods on different datasets including scenes, objects and human bodies. The top three results are highlighted in red, orange, and yellow, respectively. †:The input point number for THuman2.0 is 80K for all methods." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.13, + 0.905, + 0.292 + ], + "angle": 0, + "content": "
MethodPoint NumberScanNet[7]Car (ShapeNet[4])Chair (ShapeNet[4])Shoe (GSO[9])THuman2.0[30]
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
GR100K†13.620.5280.77919.240.8140.18217.780.7790.20123.140.8290.15320.260.9050.337
NPBG++[23]100K16.810.6710.58525.320.8740.12025.780.9160.10129.420.9290.08126.810.9520.062
TriVol[13]100K18.560.7340.47327.220.9270.08428.850.9600.07831.240.9610.04525.970.9350.059
PFGS[28]100K19.860.7580.45227.340.9420.07727.520.9560.07829.530.9570.05834.740.9830.009
Ours20K18.570.7240.60027.880.9490.06828.890.9620.07730.910.9680.05133.700.9790.014
40K19.430.7430.55228.570.9570.06229.520.9670.07031.490.9740.04734.290.9810.012
100K20.240.7590.49028.730.9600.06029.100.9650.07532.080.9780.04235.430.9870.009
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.314, + 0.486, + 0.412 + ], + "angle": 0, + "content": "Table 3. The generalization capability of our method, traditional graphics rendering, and PFGS on the DTU dataset. Here, our method is trained on the Car category of ShapeNet using 20K points. Two models of PFGS are trained on the DTU dataset with 1M points and on the Car category of ShapeNet with 20K points, respectively. The first column records the input point numbers for evaluation. The top two results are highlighted in red and orange." + }, + { + "type": "table", + "bbox": [ + 0.098, + 0.423, + 0.48, + 0.659 + ], + "angle": 0, + "content": "
Point NumberPFGS DTU 1M pointsPFGS Car (ShapeNet) 20K points
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
20K10.1770.3130.6478.0650.2970.792
40K12.0260.3580.5508.1600.3020.765
100K17.9780.5540.3128.6470.3220.673
1M23.0010.7700.11213.6760.5490.401
Point NumberGROurs \nCar (ShapeNet) 20K points
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
20K9.1050.3160.62516.7130.5900.361
40K9.9850.3280.64116.6230.5820.315
100K12.0350.3630.66616.5440.5730.263
1M18.0030.5940.39616.6430.5940.216
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.689, + 0.486, + 0.903 + ], + "angle": 0, + "content": "pability of PFGS. Like previous methods, PFGS requires entire point clouds as input for predicting Gaussians, hence these methods struggle to generalize to other datasets or varying point densities. Being trained with point cloud patches, our method does not depend on global priors and possesses superior generalization capabilities, resulting in enhanced performance on other datasets and with varying point densities. Figure 6 shows the results of methods on DTU dataset with different point numbers, where our rendering results preserve clear details. Note that previous methods are unable to function on sparse point clouds or on scene categories that are significantly different from object categories. Both quantitative and qualitative evaluations reveal the exceptional generalization capability of our" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.313, + 0.905, + 0.496 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.506, + 0.908, + 0.549 + ], + "angle": 0, + "content": "Figure 6. The evaluation results of our method and previous methods on the DTU dataset, utilizing 20K, 40K, 100K and 1M points, with all methods trained on the Car category with 20K points." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.577, + 0.573, + 0.59 + ], + "angle": 0, + "content": "method." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.602, + 0.755, + 0.618 + ], + "angle": 0, + "content": "4.5. Visualization of Gaussians" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.625, + 0.907, + 0.763 + ], + "angle": 0, + "content": "Figure 7(a)(b) illustrates the predicted Gaussians and corresponding rendered images of our method and PFGS[28]. As mentioned in Section 2.1, PFGS first predicts Gaussians and then refines the rendered images. We visualize the Gaussians of its first stage. Different from our predicted Gaussians, which are evenly distributed with fewer gaps, the predicted Gaussians of PFGS are chaos. Its rendered image is not photo-realistic requiring refinement in the second stage, and thus cannot be considered a final rendering result." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.772, + 0.791, + 0.789 + ], + "angle": 0, + "content": "4.6. Evaluation of Rendering Speed" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Table 4 represents the rendering speeds of different methods. Our method directly predicts Gaussians for rendering images, eliminating the need for further refinement once the prediction is complete. The rendering speed of our method is the same as 2DGS [15], thus our method achieves real-time rendering with the fastest rendering speed among previous methods." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "27291" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.089, + 0.482, + 0.31 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.328, + 0.485, + 0.399 + ], + "angle": 0, + "content": "Figure 7. The visualization of Gaussians and their rendered images predicted by our method, our ablated methods and PFGS. (a): PFGS trained on Car (100K). (b) and (c): Our method trained on Car (100K and 20K). (d), (e) and (f): Our ablated methods trained on Car (20K)." + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.413, + 0.484, + 0.456 + ], + "angle": 0, + "content": "Table 4. The rendering speeds of different methods, where the rendering frames per second (FPS) are presented. All methods are evaluated using an RTX 3090." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.467, + 0.48, + 0.506 + ], + "angle": 0, + "content": "
MethodNPBG++[23]TriVol[13]PFGS[28]Ours
Speed (FPS)37.451.623.80142.86
" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.531, + 0.245, + 0.547 + ], + "angle": 0, + "content": "4.7. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.553, + 0.484, + 0.659 + ], + "angle": 0, + "content": "Split Number. We conduct experiments with different split numbers \\( K \\) (1, 2, 4, 8) and different input point cloud numbers \\( N \\) (2K, 10K, 20K, 40K) on the Car category of ShapeNet. The results are shown in Figure 8. Performance is enhanced with a growing number of splits, underscoring the efficacy of our split decoders, especially when it comes to sparse point clouds." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.66, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Component Ablation. We first evaluate the effectiveness of our initialization in the 2D Gaussian prediction module. The visualization of initialized Gaussians is shown in Figure 3 (c), which indicates a coarse rendering result providing a great foundation for optimization. Without initialization, the performance of our method sharply drops as Table 5 shows. Randomly initialized Gaussians might not be visible in certain views, resulting in non-convergence, as shown in Figure 7 (d). The ablation studies of our entire-patch architecture contain two experiments: (1) removing the entire module \\(\\mathcal{N}_e\\) and utilizing the predicted incomplete images rendered from point patches for supervision, denoted as “No \\(\\mathcal{N}_e\\)” in Table 5; (2) splitting the entire point cloud to multiple patches as same as the inference process of \\(\\mathcal{N}_p\\) to render complete images during training, denoted as “\\(\\mathcal{N}_p\\) only” in Table 5. the “\\(\\mathcal{N}_p\\) only” method will consume a sub" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.09, + 0.907, + 0.118 + ], + "angle": 0, + "content": "Table 5. The ablation studies of our method. The ablated methods are trained on Car with 20K points." + }, + { + "type": "table", + "bbox": [ + 0.549, + 0.124, + 0.875, + 0.217 + ], + "angle": 0, + "content": "
PSNR ↑SSIM ↑LPIPS ↓
No Initialization9.620.6850.655
No Nc12.940.6760.441
Np only23.230.8290.126
Ours27.880.9490.068
" + }, + { + "type": "image", + "bbox": [ + 0.515, + 0.239, + 0.905, + 0.349 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.356, + 0.907, + 0.399 + ], + "angle": 0, + "content": "Figure 8. The ablation study of split number \\( K \\) with different input point numbers. The legend in the lower right corner indicates the input point number for training and evaluating." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.417, + 0.906, + 0.478 + ], + "angle": 0, + "content": "stantial amount of resources for training and inference. The evaluation results of these two ablated methods are worse than our method, where their visualization results are shown in Figure 7(e)(f)(c)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.483, + 0.906, + 0.514 + ], + "angle": 0, + "content": "The results of ablation studies validate the effectiveness and necessity of the components within our method." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.538, + 0.764, + 0.554 + ], + "angle": 0, + "content": "5. Conclusion and Discussion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.569, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In this paper, we introduce a novel approach featuring an entire-patch architecture and the 2D Gaussian prediction module for point cloud rendering. Our prediction module directly forecasts the 2D Gaussians from the provided point clouds, beginning with initialization derived from normals. The splitting decoders produce multiple Gaussians from a point, thereby better accommodating sparse point clouds. To enhance the generalization capability, the entire-patch architecture employs a module for \"background\" area prediction and another module for point cloud patch prediction, ensuring proper supervision. Comprehensive experiments and comparisons demonstrate the superiority and generalization capability of our method, which can render point clouds into photo-realistic images with clear details and achieves SOTA performance. A limitation of our method may arise when a portion of a point cloud is absent, as our method cannot render a complete image without the direct support of the points. Our future work aims to address this limitation by incorporating point cloud completion techniques and also attempting to render point clouds with noise, such as those directly scanned from wild and outdoor scenes." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "27292" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.251, + 0.108 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.116, + 0.484, + 0.147 + ], + "angle": 0, + "content": "This work was supported by the National Natural Science Foundation of China under Grant 62032011." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.162, + 0.188, + 0.178 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.188, + 0.484, + 0.243 + ], + "angle": 0, + "content": "[1] Kara-Ali Aliev, Artem Sevastopolsky, Maria Kolos, Dmitry Ulyanov, and Victor Lempitsky. Neural point-based graphics. In Computer Vision - ECCV 2020, pages 696-712, Cham, 2020. Springer International Publishing. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.246, + 0.484, + 0.302 + ], + "angle": 0, + "content": "[2] M. Botsch, A. Hornung, M. Zwicker, and L. Kobbelt. High-quality surface splatting on today's gpus. In Proceedings Eurographics/IEEE VGTC Symposium Point-Based Graphics, 2005., pages 17-141, 2005. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.304, + 0.484, + 0.358 + ], + "angle": 0, + "content": "[3] Alexandre Boulch and Renaud Marlet. Poco: Point convolution for surface reconstruction. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6292-6304, 2022. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.361, + 0.484, + 0.429 + ], + "angle": 0, + "content": "[4] Angel X. Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, Jianxiong Xiao, Li Yi, and Fisher Yu. Shapenet: An information-rich 3d model repository, 2015. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.432, + 0.484, + 0.487 + ], + "angle": 0, + "content": "[5] Kaichun Mo Leonidas J. Guibas Charles R. Qi, Hao Su. Pointnet: Deep learning on point sets for 3d classification and segmentation. In 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.49, + 0.484, + 0.558 + ], + "angle": 0, + "content": "[6] Paolo Cignoni, Marco Callieri, Massimiliano Corsini, Matteo Dellepiane, Fabio Ganovelli, and Guido Ranzuglia. MeshLab: an Open-Source Mesh Processing Tool. In Eurographics Italian Chapter Conference. The Eurographics Association, 2008. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.561, + 0.484, + 0.629 + ], + "angle": 0, + "content": "[7] Angela Dai, Angel X. Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proc. Computer Vision and Pattern Recognition (CVPR), IEEE, 2017. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.632, + 0.484, + 0.701 + ], + "angle": 0, + "content": "[8] Peng Dai, Yinda Zhang, Zhuwen Li, Shuaicheng Liu, and Bing Zeng. Neural point cloud rendering via multi-plane projection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7830-7839, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.704, + 0.484, + 0.786 + ], + "angle": 0, + "content": "[9] Laura Downs, Anthony Francis, Nate Koenig, Brandon Kinman, Ryan Michael Hickman, Krista Reymann, Thomas Barlow McHugh, and Vincent Vanhoucke. Google scanned objects: A high-quality dataset of 3d scanned household items. 2022 International Conference on Robotics and Automation (ICRA), pages 2553-2560, 2022. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.789, + 0.483, + 0.83 + ], + "angle": 0, + "content": "[10] Antoine Guédon and Vincent Lepetit. Sugar: Surface-aligned gaussian splatting for efficient 3d mesh reconstruction and high-quality mesh rendering. CVPR, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.832, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[11] Hugues Hoppe, Tony DeRose, Tom Duchamp, John Alan McDonald, and Werner Stuetzle. Surface reconstruction from unorganized points. Proceedings of the 19th annual conference on Computer graphics and interactive techniques, 1992. 3" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.188, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.162 + ], + "angle": 0, + "content": "[12] Qingyong Hu, Bo Yang, Linhai Xie, Stefano Rosa, Yulan Guo, Zhihua Wang, Niki Trigoni, and Andrew Markham. Randla-net: Efficient semantic segmentation of large-scale point clouds. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2020. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.164, + 0.908, + 0.233 + ], + "angle": 0, + "content": "[13] Tao Hu, Xiaogang Xu, Ruihang Chu, and Jiaya Jia. Trivol: Point cloud rendering via triple volumes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20732-20741, 2023. 1, 2, 3, 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.235, + 0.907, + 0.302 + ], + "angle": 0, + "content": "[14] T. Hu, Xiaogang Xu, Shu Liu, and Jiaya Jia. Point2pix: Photo-realistic point cloud rendering via neural radiance fields. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8349-8358, 2023. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.304, + 0.906, + 0.36 + ], + "angle": 0, + "content": "[15] Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2d gaussian splatting for geometrically accurate radiance fields. In SIGGRAPH 2024 Conference Papers. Association for Computing Machinery, 2024. 2, 3, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.362, + 0.906, + 0.417 + ], + "angle": 0, + "content": "[16] Rasmus Jensen, Anders Dahl, George Vogiatzis, Engil Tola, and Henrik Aanæs. Large scale multi-view stereopsis evaluation. In 2014 IEEE Conference on Computer Vision and Pattern Recognition, pages 406-413, 2014. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.42, + 0.906, + 0.474 + ], + "angle": 0, + "content": "[17] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42 (4), 2023. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.476, + 0.906, + 0.517 + ], + "angle": 0, + "content": "[18] Leif Kobbelt and Mario Botsch. A survey of point-based techniques in computer graphics. Comput. Graph., 28(6): 801-814, 2004. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.519, + 0.906, + 0.561 + ], + "angle": 0, + "content": "[19] Xu Ma, Can Qin, Haoxuan You, Haoxi Ran, and Yun Fu. Rethinking network design and local geometry in point cloud: A simple residual mlp framework. *ICLR*, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.563, + 0.906, + 0.617 + ], + "angle": 0, + "content": "[20] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.62, + 0.906, + 0.744 + ], + "angle": 0, + "content": "[21] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Kopf, Edward Yang, Zachary DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. Pytorch: An imperative style, high-performance deep learning library. In Advances in Neural Information Processing Systems 32, pages 8024-8035. Curran Associates, Inc., 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.746, + 0.906, + 0.801 + ], + "angle": 0, + "content": "[22] Hanspeter Pfister, Matthias Zwicker, Jeroen van Baar, and Markus H. Gross. Surfels: surface elements as rendering primitives. Proceedings of the 27th annual conference on Computer graphics and interactive techniques, 2000. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.803, + 0.906, + 0.872 + ], + "angle": 0, + "content": "[23] Ruslan Rakhimov, Andrei-Timotei Ardelean, Victor Lempitsky, and Evgeny Burnaev. Npbg++: Accelerating neural point-based graphics. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15969-15979, 2022. 1, 2, 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[24] Nikhila Ravi, Jeremy Reizenstein, David Novotny, Taylor Gordon, Wan-Yen Lo, Justin Johnson, and Georgia" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.093, + 0.908, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "27293" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.12 + ], + "angle": 0, + "content": "Gkioxari. Accelerating 3d deep learning with pytorch3d. arXiv:2007.08501, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.177 + ], + "angle": 0, + "content": "[25] Liu Ren, Hanspeter Pfister, and Matthias Zwicker. Object space ewa surface splatting: A hardware accelerated approach to high quality point rendering. Computer Graphics Forum, 21(3):461-470, 2002. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.179, + 0.482, + 0.22 + ], + "angle": 0, + "content": "[26] Darius Rückert, Linus Franke, and Marc Stamminger. Adop: Approximate differentiable one-pixel point rendering. ACM Transactions on Graphics (ToG), 41(4):1-14, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.222, + 0.482, + 0.276 + ], + "angle": 0, + "content": "[27] Bane Sullivan and Alexander Kaszynski. PyVista: 3D plotting and mesh analysis through a streamlined interface for the Visualization Toolkit (VTK). Journal of Open Source Software, 4(37):1450, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.278, + 0.482, + 0.318 + ], + "angle": 0, + "content": "[28] Jiaxu Wang, Ziyi Zhang, Junhao He, and Renjing Xu. Pfgs: High fidelity point cloud rendering via feature splatting. ECCV 2024, 2024. 1, 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.321, + 0.482, + 0.375 + ], + "angle": 0, + "content": "[29] Zhou Wang, A.C. Bovik, H.R. Sheikh, and E.P. Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing, 13(4): 600-612, 2004. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.378, + 0.482, + 0.445 + ], + "angle": 0, + "content": "[30] Tao Yu, Zerong Zheng, Kaiwen Guo, Pengpeng Liu, Qionghai Dai, and Yebin Liu. Function4d: Real-time human volumetric capture from very sparse consumer rgbd sensors. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR2021), 2021. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.448, + 0.482, + 0.489 + ], + "angle": 0, + "content": "[31] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.491, + 0.482, + 0.531 + ], + "angle": 0, + "content": "[32] Qian-Yi Zhou, Jaesik Park, and Vladlen Koltun. Open3D: A modern library for 3D data processing. arXiv:1801.09847, 2018. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.533, + 0.482, + 0.586 + ], + "angle": 0, + "content": "[33] Matthias Zwicker, Hanspeter Pfister, Jeroen Baar, and Markus Gross. Surface splattering. Proceedings of the ACM SIGGRAPH Conference on Computer Graphics, 2001, 2001. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.586 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "27294" + } + ] +] \ No newline at end of file diff --git a/2025/Sparse Point Cloud Patches Rendering via Splitting 2D Gaussians/dd278d9d-da45-4b1e-bcc2-33b0924a1fa1_origin.pdf b/2025/Sparse Point Cloud Patches Rendering via Splitting 2D Gaussians/dd278d9d-da45-4b1e-bcc2-33b0924a1fa1_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..d14a19671353bf99dfdb01b4718e7c180fd2a3ba --- /dev/null +++ b/2025/Sparse Point Cloud Patches Rendering via Splitting 2D Gaussians/dd278d9d-da45-4b1e-bcc2-33b0924a1fa1_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5bc95e5700e6e0f25cc509c49f2f7c6bf989b5e027f83fea770f3ba9d26feb63 +size 9362914 diff --git a/2025/Sparse Point Cloud Patches Rendering via Splitting 2D Gaussians/full.md b/2025/Sparse Point Cloud Patches Rendering via Splitting 2D Gaussians/full.md new file mode 100644 index 0000000000000000000000000000000000000000..a51df90f6f937ae8ae0cd80577af48f7523d34ee --- /dev/null +++ b/2025/Sparse Point Cloud Patches Rendering via Splitting 2D Gaussians/full.md @@ -0,0 +1,292 @@ +# Sparse Point Cloud Patches Rendering via Splitting 2D Gaussians + +Changfeng Ma $^{1}$ , Ran Bi $^{1}$ , Jie Guo $^{1}$ , Chongjun Wang $^{1}$ , Yanwen Guo $^{12*}$ $^{1}$ Nanjing University, Nanjing, China + $^{2}$ School of Software, North University of China + +{changfengma, 211250233}@smail.nju.edu.cn + +{guojie, chjwang, ywguo}@nju.edu.cn + +![](images/9d4d6b033eabffbf029eaf37ab8a1448429297366823f9b8d6fa2a01ceac05ee.jpg) +Car +Point Cloud + +![](images/bf5f8ecb853a56f2c13091a2a43ee81d4ae0a6e3012240cb93d9c4696f7d0689.jpg) +NPBG+± + +![](images/591b037e67a826ba4bf302c9828820d4f1bbd5eeb832e4072bb62dba7c4aedbb.jpg) + +![](images/bc7518ba8c933acdda4fa80540ae016e91b2a7ba9ab93f2571a60109a6c64ff9.jpg) +TriVol + +![](images/ad480242986f91afb37657047c5966ade456c4b558d3075eb78c6436e8733a09.jpg) + +![](images/15a6c7afc42cd95fa20fe573507e1d61feb5c739cdfd5d4d48dcf993d02d1e98.jpg) +PFGS + +![](images/b1220996de86ded1acbef8a69f08538b905c0c6ba1e8c6a2734f1207b379fbcf.jpg) +Figure 1. Our method excels in rendering intricate, photo-realistic images from point clouds of different categories. In this instance, our model is trained on the Car category, utilizing 20K points as input. + +![](images/7659f8d1a9478ac4cafb70a5ec795e9e447c8ef8389ec92774c739b32e421efa.jpg) +Ours + +![](images/f25f999b074e6f5d71444861da5a33b76655a61a669a79cdcb6be4623646d986.jpg) + +![](images/d33e9d13af948c7082cbd821c671c6e9f444f1c88395d45f44dfab564002ac9f.jpg) +GT + +![](images/996bbb729c98e5b0318897b6919590734b0358864ed5a774ca336f04f6b58a56.jpg) + +![](images/c4e48cfab5721918aad94808b29ad87164af7eb18d9f79af61d8c94f7accefe7.jpg) +Generalize to Other Categories + +![](images/d294a263598f72cd8895f81e57fd6f224b2b8b15f677352d7ee3768d05cc5b86.jpg) + +![](images/2ea980b1f830657c635adfc199d2f0b7234b3375a833be1d0b788853f0dfddc5.jpg) +Sculpture + +![](images/756cc0419e4f854567bf136d5e3b7cd287982c77e453943d2447eabbc11b71ec.jpg) + +![](images/56228bb22cd5c45271e6adc5f71f364b3a7a8dde46090ef763073123f11c4286.jpg) + +![](images/2bcc0458d0620247aaefe164414bb61082bc343a159bd2957a268515dbb7c507.jpg) + +![](images/b9669c9491e4f05faec33dcb8eb7c6ed71fd289b25d545fa2b11802cc558ed0e.jpg) + +![](images/dc52fd9874f362f2def01a2be70f3b28299b1a60fdc3c862d6c37bf01ae6279b.jpg) + +![](images/2bd0f885f9ed91620037d0f37d4edbb7eb87f9e3edaa5f727e6e7ff75a2b8924.jpg) + +# Abstract + +Current learning-based methods predict NeRF or 3D Gaussians from point clouds to achieve photo-realistic rendering but still depend on categorical priors, dense point clouds, or additional refinements. Hence, we introduce a novel point cloud rendering method by predicting 2D Gaussians from point clouds. Our method incorporates two identical modules with an entire-patch architecture enabling the network to be generalized to multiple datasets. The module normalizes and initializes the Gaussians utilizing the point cloud information including normals, colors and distances. Then, splitting decoders are employed to refine the initial Gaussians by duplicating them and predicting more accurate results, making our methodology effectively accommodate sparse point clouds as well. Once trained, our approach exhibits direct generalization to point clouds across different categories. The predicted Gaussians are employed directly for rendering without additional refinement on the rendered images, retaining the benefits of 2D Gaussians. We conduct extensive experiments on various datasets, and the results demonstrate the superiority and generalization of our method, which achieves SOTA performance. The code is + +available at https://github.com/murcherful/GauPCRender. + +# 1. Introduction + +Point clouds can be easily acquired using various types of 3D scanners, making them a recent research hotspot. Photorealistic rendering of point clouds is significant for many applications such as visualization, virtual reality and automatic driving. However, the sparsity and discrete nature of point clouds continue to pose challenges for achieving photo-realistic rendering. + +Traditional graphics rendering methods [2, 22] merely project points as planes or spheres onto images. Consequently, their outcomes lack photo-realism and are riddled with numerous holes. Previous deep learning-based works [1, 8, 23] necessitate point clouds and corresponding images from several views as input, which are difficult to acquire in practice. These methods also require training or fine-tuning for each scene or object. Recent works [13, 14] have focused on predicting NeRF [20] from given point clouds to render images from arbitrary views. PFGS [28] predicts the 3D Gaussians [17] with features for point clouds. After rendering, PFGS refines the rendered images based on these features. Although these methods can render point clouds + +Table 1. The comparison of different methods across several dimensions. Here, "S. S/O.", "S. C." and "M. C." represent "single scene or object", "single category" and "multiple categories", respectively. $\mathcal{P}$ and $\mathcal{I}$ indicate point clouds and images, separately. $\dagger$ : Traditional graphics rendering approach is unable to render photo-realistic images. $\ddagger$ : PFGS requires further refinement on images rendered from 3D gaussians. + +
MethodGeneralization abilityInputOutputPre- Fine- trainFine- tunePoint number
\( GR^† \)AnyPI✘✘Any
NPBG [1]S. S/O.\( P+I \)I✓✓100K
NPCR [8]S. S/O.\( P+I \)I✘✓100K
NPBG++ [23]S. S/O.\( P+I \)I✓✘100K
TriVol [13]S. C.PNeRF[20]✓✘100K
Point2Pix [14]S. C.PNeRF[20]✓✘100K
PFGS [28]S. C.P3DGS[17]‡✓✘80K
OursM. C.P2DGS[15]✓✘2K-100K
+ +without training or fine-tuning on each data, they still have limitations, such as the need for dense point clouds, poor generalization, slow rendering speeds, and blurry results. + +2D Gaussian Splatting [15] has recently been proposed, which includes explicit normals of Gaussians. These normals facilitate easy initialization, learning, and prediction by the network from point clouds. Motivated by this, we introduce a novel method to predict the 2D Gaussians of the given point cloud and rasterize the predicted 2D Gaussians into images for photo-realistic point cloud rendering. + +Due to taking entire point clouds as input, previous methods are hard to generalize to other categories or adapt to different point distributions. Processing only point cloud patches with the network can enhance its generalization capability and reduce dependence on category priors [3, 12]. However, rendering Gaussians predicted from patch points results in incomplete images that are unsupervisible. Therefore, we employ an entire-patch architecture that includes two identical 2D Gaussian prediction modules to handle entire and patch point clouds, separately. We use the rendering results of the entire point cloud as a "background" to obtain a complete image for proper supervision. The module first normalizes and initializes the 2D Gaussians based on the given point clouds, encompassing the estimated normals, scales, and colors, which are crucial for our network to converge and predict more accurate results. Subsequently, a point cloud encoder and multiple splitting decoders are employed to extract local features and predict the parameters of the duplicated Gaussians. Equipped with the splitting decoder, our network can handle sparse point clouds and produce high-quality details with dense Gaussians. Thanks to the proposed architecture, our method also exhibits strong generalization across scenes and objects. Without further refinement on rendered images, our + +method retains the advantages of 2D Gaussian Splatting in rendering, such as rapid rendering speed. + +We conduct extensive comparative and ablation experiments, and the results demonstrate the superiority of our method, which can render point clouds into high-quality detailed images with excellent generalization. We will make our dataset and codes public in the future. Our main contributions are as follows: + +- We introduce a method that directly predicts 2D Gaussians from point clouds to render photo-realistic images without any refinement, while retaining the advantages of 2D Gaussian Splatting. +- Equipped with an entire-patch architecture, our method can handle point cloud patches, leading to significant generalization across scenes and objects. +- We introduce the splitting decoder, which splits Gaussians into denser outcomes, enabling our method to render detailed images from even sparse point clouds. + +# 2. Related Work + +# 2.1. Point Cloud Rendering + +Traditional Point Cloud Rendering. Point cloud rendering has consistently been a significant research topic [18]. Conventional point cloud rendering techniques [6, 24, 27, 32] directly rasterize points onto 2D screens for rendering, yet they grapple with the issue of holes. Subsequent studies [2, 22, 25, 33] have proposed splatting points with elliptic discs, ellipsoids, or surfels for point cloud rendering to mitigate the holes observed during the rendering process. Although these graphic rendering methods can render any point cloud, the resulting images often contain holes and exhibit blurry details, lacking photo-realism. + +Point-based novel view synthesis. For photo-realistic point cloud rendering, several recently proposed methods utilize corresponding images of point clouds as additional information. NPBG [1] signs descriptors to point clouds and rasterizes into multi-resolution raw images with features and employs a U-Net to render the final results. Dai et al. [8] aggregate points with features onto multiple planes based on a given view and point cloud. Following multi-plane-based voxelization, they employ a 3D U-Net to predict the rendered image. ADOP [26] is introduced for rendering HD images from point clouds, which employs several global parameters, a neural renderer and a differentiable physically-based tonemapper for optimization. These methods necessitate training for each point cloud with images from several views to render images of novel views, which is a time-consuming process. To reduce training time, NPBG++ [23] is trained on a collection of point clouds using a network that extracts features from corresponding images. This network aligns the features of input images to point clouds without the need for fine-tuning. However, this + +method still requires images as input, which can be challenging to obtain in practice. + +Learning-based Point Cloud Rendering. Several works [13, 14, 28] are proposed for rendering point clouds without the need for training or fine-tuning on each point cloud and corresponding images. TriVol [13] utilizes triple-plane-based grouping to query point features for predicting a NeRF from a given point cloud. Hu et al. [14] propose multi-scale radiance fields to predict NeRFs from point clouds. They also introduce a sampling strategy to accelerate fine-tuning and rendering speed. PFGS [28], a two-stage approach, for the first time employs feature splatting to predict 3D Gaussians from point clouds and uses a recurrent decoder to refine the rendered images. These methods are capable of rendering point clouds without fine-tuning or the need for input images, yet they still depend on categorical priors, dense points, or further refinements. Table 1 summarizes the distinctions between our method and the previous methods across several dimensions, highlighting the innovation and superiority of our proposed approach. + +# 2.2. Gaussian Splitting + +3D Gaussian Splitting [17] is proposed with remarkable performance in novel view synthesis and further explores the rendering capability of the point clouds. 3DGS utilizes the differentiable splatting rendering to optimize 3D Gaussian primitives initialized from point clouds. Scenes are explicitly represented with 3D Gaussians that can be rendered with remarkable speed. Meshes can also be reconstructed from 3D Gaussians[10]. However, the normals of 3D Gaussians are so vague that the resulting geometries lack precision. To generate more accurate geometries, Hua et al. [15] propose 2DGS that employs 2D Gaussian disks containing explicit normals as representation primitives. With precise normals, meshes can accurately reconstruct the geometry of objects. Given that the normals of 2D Gaussians can be readily determined, 2D Gaussians can also be initialized using normals. Luckily, the normals of points can be effortlessly estimated [11] facilitating the initialization and forecasting of 2D Gaussians. Motivated by this, we opt to predict 2D Gaussians in place of 3D Gaussians from the provided point clouds. + +# 3. Method + +Given a point cloud $\mathcal{P}$ and a camera $\mathcal{C}$ , point cloud rendering aims to render a photo-realistic image $\mathcal{I}$ . Here, point cloud $\mathcal{P}$ contains of coordinates of 3D points $\mathbf{P}_{\mathbf{x}} \in \mathbb{R}_{N \times 3}$ and colors of points $\mathbf{P}_{\mathbf{c}} \in [0,1]_{N \times 3}$ , where $N$ indicates the point number of $\mathcal{P}$ . Camera $\mathcal{C}$ consists of intrinsic matrix $\mathbf{K} \in \mathbb{R}_{3 \times 3}$ and extrinsic matrix $\mathbf{P} \in \mathbb{R}_{4 \times 4}$ . + +For photo-realistic point cloud rendering, we propose a network $\mathcal{N}$ to predict 2D Gaussians $\mathcal{G} = \mathcal{N}(\mathcal{P})$ given point cloud $\mathcal{P}$ , and employ differentiable splatting rendering [15] + +$f_{R}$ to synthesize images $\mathcal{I} = f_{R}(\mathcal{G},\mathcal{C})$ . The parameters of a 2D Gaussian $\mathbf{g} = \{\mathbf{x},\mathbf{s},o,\mathbf{c},\mathbf{n},\alpha \} \in \mathcal{G}$ include: (1) position $\mathbf{x}\in \mathbb{R}_3$ , (2) scale $\mathbf{s}\in \mathbb{R}_2$ , (3) opacity $o\in [0,1]$ , (4) spherical harmonic of color $\mathbf{c}\in \mathbb{R}_d$ , (5) normal $\mathbf{n}\in \mathbb{R}_3$ and (6) rotation angle along normal $\alpha \in [0,2\pi ]$ , where $d$ represents the parameter number of spherical harmonic for color. As shown in Figure 2, our network contains two identical 2D Gaussian prediction modules (3.1) with an entire-patch architecture (3.2). + +# 3.1.2D Gaussian Prediction Module + +# 3.1.1. Initialization of 2D Gaussians + +We first introduce 2D Gaussian prediction module to predict the 2D Gaussians $\mathcal{G}$ given a point cloud $\mathcal{P}$ as input. The positions of the point cloud $\mathcal{P}$ are normalized to the range $[-1,1]$ based on the geometric center $\mathbf{c} \in \mathbb{R}_3$ and scale $s \in R$ of the input point cloud to facilitate easier feature extraction and prediction of results, where the scale is defined as the maximum distance between any point and $\mathbf{c}$ . Our module then initializes one 2D Gaussian $\mathbf{g}$ for each normalized point $\mathbf{p} = \{\mathbf{p}_{\mathbf{x}}, \mathbf{p}_{\mathbf{c}}\} \in \mathcal{P}$ , where $\mathbf{p}_{\mathbf{x}} \in \mathcal{R}_3$ and $\mathbf{p}_{\mathbf{c}} \in \mathcal{R}_3$ indicate the position and color of the point respectively. The position $\mathbf{x}$ of the Gaussian is initialized with $\mathbf{p}_{\mathbf{x}}$ . The spherical harmonic $\mathbf{c}$ and scale $\mathbf{s}$ are initialized as the same as 2DGS [15] according to $\mathbf{p}_{\mathbf{c}}$ and the minimum distance between points, separately. The opacity $o$ is set to 1 to ensure that all Gaussians are visible. As we mentioned before, the geometry of given point cloud $\mathcal{P}$ provides important clues for the rotation of Gaussians. Gaussian Splatting works [15, 17] utilize quaternions to represent the rotation of Gaussians and only randomly initialize them as Figure 3(b) shows, which are difficult for networks to predict intuitively. Fortunately, the normals of given point cloud can be easily estimated. Therefore, we represent the rotation of the Gaussian in terms of the normal $\mathbf{n}$ and the rotation angle $\alpha$ around that normal, and initialize the normal of the Gaussian using the estimated normal of the point cloud. Figure 3 (a) and (c) illustrate this initialization process. Specifically, we estimate the normals for $\mathcal{P}$ [11, 32], where the normal $\hat{\mathbf{n}}$ for each point $\mathbf{p}$ is utilized to initialize the normal $\mathbf{n}$ of Gaussian. And the rotation angle $\alpha$ of Gaussian is set to 0. The estimated normal may not be precise, but it provides crucial information and ensures that each Gaussian is visible, thereby guaranteeing that the network can converge correctly. After initializing for all the points, the initialized Gaussians $\mathcal{G}^i$ composing of the initialized positions, normals, scales, spherical harmonics, opacities and rotation angles ( $\mathbf{X}^i, \mathbf{N}^i \in \mathbb{R}_{N\times 3}, \mathbf{S}^i \in \mathbb{R}_{N\times 2}, \mathbf{C}^i \in \mathbb{R}_{N\times 3}, O^i, A^i \in \mathbb{R}_N$ ) is passed to our network as input and predicting basis. + +# 3.1.2. Splitting Decoder + +As shown in Figure 2, local features are required before splitting and predicting 2D Gaussians. To extract features from the initialized 2D Gaussians, we employ + +![](images/554cdefc4dec6134350b048c8b7bec91da332ec181726e8f215df849700885ea.jpg) + +![](images/311ad8e7a092496b911ba656a3b4c686079d96882aeceedeebe70fbe13718379.jpg) + +![](images/bdcbd1a965c81961da135a0fcf2765fee40118f781745bb22881f183f1d6201d.jpg) +Figure 2. Overview of our proposed method. Our method predicts 2D Gaussians for point cloud rendering, employing an entire-patch architecture (bottom) and the 2D Gaussian Prediction Module (top left) with splitting decoders (top right). + +![](images/5677a16993197090d21fa6b6705109dc5505b68d8823bbf81cbd9e3c2d709160.jpg) +Figure 3. The illustration of our initialization approach, where the left side of each picture is a 2D schematic diagram and the right side is the rendered images of Gaussians. (a) The estimated normals of the point cloud. (b) Randomly Initialization. (c) Our Initialization. (d) Predicted 2D Gaussians. + +PointMLP[19] as the encoder $E$ , which is a simple residual MLP framework for point cloud learning with great performance on many tasks. The local feature $\mathbf{F}_l$ is extracted by: + +$$ +\mathbf {F} _ {l} = E (\mathbf {X} ^ {i}, \mathbf {C} ^ {i}, \mathbf {N} ^ {i}, \mathbf {S} ^ {i}). +$$ + +Subsequently, $\mathbf{F}_l$ is augmented with $\mathbf{X}^i$ , $\mathbf{C}^i$ , $\mathbf{N}^i$ , $\mathbf{S}^i$ as the input features for the splitting decoders. Our splitting decoders $D$ also utilize $\mathcal{G}^i$ as the basis for prediction and forecast 2D Gaussians $\mathcal{G}^p$ for rendering purposes. There are six splitting decoders $D_x, D_s, D_c, D_n, D_\alpha, D_o$ responsible for predicting the six parameters $X^p, S^p, C^p, N^p, A^p, O^p$ of 2D Gaussians. Each decoder contains a weight-shared MLP [5] as shown in Figure 2. The splitting and predicting process for the 6 parameters is basically the same, except for slight differences in shapes. Take the process of predicting the position $X^p$ as an example. The decoder $D_x$ predicts $K$ shifts $\Delta^x$ based on the position $\mathbf{X}^i$ : + +$$ +\boldsymbol {\Delta} _ {1} ^ {x}, \boldsymbol {\Delta} _ {2} ^ {x},..., \boldsymbol {\Delta} _ {K} ^ {x} = D _ {x} (\mathbf {F} _ {l}, \mathbf {X} ^ {i}, \mathbf {C} ^ {i}, \mathbf {N} ^ {i}, \mathbf {S} ^ {i}), +$$ + +where $K$ indicates the number of splits. The $K$ shifts are added to the same $\mathbf{X}^i$ that is "split" the initial position $\mathbf{X}^i$ to $K$ new positions. The predicted position $\mathbf{X}^p\in \mathbb{R}_{K\cdot N\times 3}$ is the combination of these $K$ new positions: + +$$ +\mathbf {X} ^ {p} = \bigcup_ {j = 1} ^ {K} \left(\boldsymbol {\Delta} _ {j} ^ {x} + \mathbf {X} ^ {i}\right). +$$ + +After applying all the splitting decoders, we can obtain the parameters of $K \cdot N$ Gaussians $\mathcal{G}^p$ . With the splitting decoders, our method can manage sparse point clouds by augmenting the number of Gaussians, serving as a form of point cloud upsampling. A greater number of Gaussians can also capture more complex features, enhancing the details of rendered images. The predicted 2D Gaussians are based on normalized points, hence they need to be de-normalized to achieve accurate rendering results. The de-normalization is applied to the predicted position $X^p$ and scale $S^p$ . + +# 3.2. Entire-Patch Architecture + +However, the 2D Gaussian prediction module can only be trained on complete point clouds, which leads to poor generalization across different categories with varying point distributions, as we only have ground truth images of entire objects for supervision. Numerous methods for other point cloud tasks, including surface reconstruction[3] and semantic segmentation[12], utilize point cloud patches as processing units to enhance the generalization capabilities of their networks. However, the corresponding ground truth images of point cloud patches are unavailable in practice, which makes training on patches impossible. To facilitate the application of our method to point cloud patches, we utilize an entire-patch architecture, as illustrated at the bottom of Fig- + +ure 2. Our approach encompasses two 2D Gaussian prediction modules, $\mathcal{N}_e$ and $\mathcal{N}_p$ , which operate on the entire point cloud and the point cloud patch, respectively. Given an entire point cloud $\mathcal{P}_e$ for rendering, our method randomly selects a point as the center point of a point cloud patch. The point cloud patch $\mathcal{P}_p$ is obtained by gathering $N_p$ nearest points around the center point. The 2D Gaussians of point cloud $\mathcal{P}_e$ and patch $\mathcal{P}_p$ are predicted by two prediction modules, where $\mathcal{G}_e = \mathcal{N}_e(\mathcal{P}_e)$ and $\mathcal{G}_p = \mathcal{N}_p(\mathcal{P}_p)$ . The non-patch part $\mathcal{G}_e'$ of $\mathcal{G}_e$ are the 2D Gaussians corresponding to the non-patch points $\mathcal{P}_e' = \mathcal{P}_e - \mathcal{P}_p$ . The final 2D Gaussians $\mathcal{G} = \mathcal{G}_e' \cup \mathcal{G}_p$ are used to render images for supervision during training. Here, $\mathcal{G}_e'$ serves as the "background" ensuring that the rendered image is complete for calculating the loss function and $\mathcal{N}_p$ can be properly trained. In practice, we initially train $\mathcal{N}_e$ separately and then freeze $\mathcal{N}_e$ while training $\mathcal{N}_p$ for our method. Both $\mathcal{N}_e$ and $\mathcal{N}_p$ can be utilized for inference. The inference of $\mathcal{N}_e$ directly takes $\mathcal{P}_e$ as input and outputs the entire 2D Gaussians for rendering. For $\mathcal{N}_p$ , our method divides the point cloud into several patches. To ensure that all points of the input point cloud are covered, we repeatedly perform the process of randomly selecting a central point and extracting surrounding points as the patch point cloud, continuing until all points are included in at least one patch. The 2D Gaussian of the entire point cloud is the combination of the predictions from all patches. + +# 3.3. Training Details + +Given $N_{c}$ cameras $\mathcal{C}_1,\ldots ,\mathcal{C}_{N_c}$ , the rendered images $\mathcal{I}_1^p,\dots ,\mathcal{I}_{N_c}^p$ is obtained by $f_{R}(\mathcal{G},\mathcal{C}_{1}),\ldots ,f_{R}(\mathcal{G},\mathcal{C}_{N_{c}})$ . Similar to [15], we employ MSE and SSIM between rendered images and the corresponding ground truth images $\mathcal{I}_1^g,\dots ,\mathcal{I}_{N_c}^g$ as loss functions for supervision: + +$$ +\mathcal {L} = \frac {1}{N _ {c}} \sum_ {i = 1} ^ {N _ {c}} \left(\beta \mathcal {L} _ {M S E} (\mathcal {I} _ {i} ^ {p}, \mathcal {I} _ {i} ^ {g}) + (1 - \beta) \mathcal {L} _ {S S I M} (\mathcal {I} _ {i} ^ {p}, \mathcal {I} _ {i} ^ {g})\right). +$$ + +We implement our method with PyTorch [21]. The split number $K$ is set to 4. The point number $N_{p}$ of a patch is 2048. $\beta$ is set to 0.8 for the loss function. The number of rendered images $N_{c}$ is 8. The resolutions of images are 256x256, 640x512 and 512x512 for objects, scenes and human bodies. We employ Adam optimizer whose learning rate is $1.0 \times 10^{-4}$ . The batch size is set to 8, and the maximum epoch of training is 480. Training our method takes about 30 hours with a RTX 3090 GPU. + +# 4. Experiments + +# 4.1. Settings + +Dataset. We compare our method with previous works across 5 datasets that include scenes, objects, and human bodies, encompassing the following datasets. ShapeNet [4] + +![](images/f6c7fbe43e16219a6e30f2505829e02f287298609e8cba653ab4d046d12f84e5.jpg) +Figure 4. The evaluation of our method, TriVol and PFGS trained with different point numbers on Car category. The legend in the lower right corner indicates different methods. + +is a large-scale dataset that contains CAD models across various categories. GSO [9] is a high-quality dataset comprising 3D scanned household items. ScanNet [7] is a dataset consisting of real-scanned indoor scenes. DTU [16] is a multi-view stereo dataset featuring high-quality and high-density scenes. THuman2.0 [30] is a dataset comprising high-quality 3D models of human bodies. We select the chairs and tables from ShapeNet and shoes from GSO for object-level comparison. All settings are consistent with those of TriVol [13] and PFGS [28]. + +Metrics. We evaluate the rendering results using three widely accepted metrics: PSNR, SSIM[29], and LIPIS[31]. Baselines. We employ traditional graphics rendering (GR), point-based novel view synthesis method NPBG++[23] and learning-based method TriVol[13] and PFGS[28] for comparison on different datasets. + +# 4.2. Evaluation of Rendering + +The evaluation results of our method and baselines are presented in Table 2. The point numbers $N$ of input point clouds are 100K for objects and scenes, and 80K for human bodies. The results demonstrate that our method outperforms nearly all other methods across various metrics and categories. Our approach outperforms existing methods, showing average improvements of $2.60\%$ , $1.05\%$ , and $6.06\%$ in PSNR, SSIM, and LIPIs across all datasets, respectively. For qualitative comparison, we represent the rendering results of different methods across several datasets in Figure 5. Our rendering results not only exhibit superior overall quality, but also display crisp and detailed elements, such as patterns, wheels, and human faces. Quantitative and qualitative comparisons demonstrate that our method achieves SOTA performance in point cloud rendering. + +# 4.3. Evaluation on Different Point Number + +As shown in Table 2, we also display the evaluation results of our method trained with different input point numbers, including 20K and 40K. Despite utilizing fewer points, our method still achieves the best performance on object categories such as chairs, tables, and shoes. Note that our + +![](images/74ae7f8c770fface135dc89a53a4068b99441dea058f7192ddb8562d56aa23a9.jpg) +Figure 5. The rendering results of our method and previous methods on different categories. From top to bottom: scene, car, chair, shoe and human body. + +method only utilizes $20\%$ of the points as input, indicating that our method also yields good results for sparse point clouds. We also showcase the evaluation results for TriVol and PFGS when trained with varying point counts (2K, 10K, 20K, 40K, 100K) on the Car category to illustrate their performance with sparse point clouds. The comparisons are depicted in Figure 4. All methods exhibit improved performance as the number of points used during training increases, with our method outperforming all others across all input point numbers. The comparisons highlight the superiority of our method in rendering sparse point clouds. + +# 4.4. Evaluation of Generalization Capability + +To evaluate the generalization capability of our $\mathcal{N}_p$ on different categories with different input point numbers, we directly test our network on DTU dataset, which is trained on the Car category using 20K points. The setting here is the same as that in PFGS [28]. The results are presented in Table 3, where the methods are assessed using 20K, 40K, 100K and 1M points on the DTU dataset. We also present the evaluation results of PFGS when trained on DTU with 1M points and on the Car category of ShapeNet with 20K points, in order to compare the generalization ca + +Table 2. The comparison of our method against previous methods on different datasets including scenes, objects and human bodies. The top three results are highlighted in red, orange, and yellow, respectively. †:The input point number for THuman2.0 is 80K for all methods. + +
MethodPoint NumberScanNet[7]Car (ShapeNet[4])Chair (ShapeNet[4])Shoe (GSO[9])THuman2.0[30]
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
GR100K†13.620.5280.77919.240.8140.18217.780.7790.20123.140.8290.15320.260.9050.337
NPBG++[23]100K16.810.6710.58525.320.8740.12025.780.9160.10129.420.9290.08126.810.9520.062
TriVol[13]100K18.560.7340.47327.220.9270.08428.850.9600.07831.240.9610.04525.970.9350.059
PFGS[28]100K19.860.7580.45227.340.9420.07727.520.9560.07829.530.9570.05834.740.9830.009
Ours20K18.570.7240.60027.880.9490.06828.890.9620.07730.910.9680.05133.700.9790.014
40K19.430.7430.55228.570.9570.06229.520.9670.07031.490.9740.04734.290.9810.012
100K20.240.7590.49028.730.9600.06029.100.9650.07532.080.9780.04235.430.9870.009
+ +Table 3. The generalization capability of our method, traditional graphics rendering, and PFGS on the DTU dataset. Here, our method is trained on the Car category of ShapeNet using 20K points. Two models of PFGS are trained on the DTU dataset with 1M points and on the Car category of ShapeNet with 20K points, respectively. The first column records the input point numbers for evaluation. The top two results are highlighted in red and orange. + +
Point NumberPFGS DTU 1M pointsPFGS Car (ShapeNet) 20K points
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
20K10.1770.3130.6478.0650.2970.792
40K12.0260.3580.5508.1600.3020.765
100K17.9780.5540.3128.6470.3220.673
1M23.0010.7700.11213.6760.5490.401
Point NumberGROurs +Car (ShapeNet) 20K points
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
20K9.1050.3160.62516.7130.5900.361
40K9.9850.3280.64116.6230.5820.315
100K12.0350.3630.66616.5440.5730.263
1M18.0030.5940.39616.6430.5940.216
+ +pability of PFGS. Like previous methods, PFGS requires entire point clouds as input for predicting Gaussians, hence these methods struggle to generalize to other datasets or varying point densities. Being trained with point cloud patches, our method does not depend on global priors and possesses superior generalization capabilities, resulting in enhanced performance on other datasets and with varying point densities. Figure 6 shows the results of methods on DTU dataset with different point numbers, where our rendering results preserve clear details. Note that previous methods are unable to function on sparse point clouds or on scene categories that are significantly different from object categories. Both quantitative and qualitative evaluations reveal the exceptional generalization capability of our + +![](images/45c2d67f60af888019e01bfedba89a09eb09af02f582eec2d279709365838d48.jpg) +Figure 6. The evaluation results of our method and previous methods on the DTU dataset, utilizing 20K, 40K, 100K and 1M points, with all methods trained on the Car category with 20K points. + +method. + +# 4.5. Visualization of Gaussians + +Figure 7(a)(b) illustrates the predicted Gaussians and corresponding rendered images of our method and PFGS[28]. As mentioned in Section 2.1, PFGS first predicts Gaussians and then refines the rendered images. We visualize the Gaussians of its first stage. Different from our predicted Gaussians, which are evenly distributed with fewer gaps, the predicted Gaussians of PFGS are chaos. Its rendered image is not photo-realistic requiring refinement in the second stage, and thus cannot be considered a final rendering result. + +# 4.6. Evaluation of Rendering Speed + +Table 4 represents the rendering speeds of different methods. Our method directly predicts Gaussians for rendering images, eliminating the need for further refinement once the prediction is complete. The rendering speed of our method is the same as 2DGS [15], thus our method achieves real-time rendering with the fastest rendering speed among previous methods. + +![](images/f68cc7b16a78b6c1c92221e4a68ad3b33b9f24d1d91ad4ac465b4c6c2a10bd98.jpg) +Figure 7. The visualization of Gaussians and their rendered images predicted by our method, our ablated methods and PFGS. (a): PFGS trained on Car (100K). (b) and (c): Our method trained on Car (100K and 20K). (d), (e) and (f): Our ablated methods trained on Car (20K). + +Table 4. The rendering speeds of different methods, where the rendering frames per second (FPS) are presented. All methods are evaluated using an RTX 3090. + +
MethodNPBG++[23]TriVol[13]PFGS[28]Ours
Speed (FPS)37.451.623.80142.86
+ +# 4.7. Ablation Study + +Split Number. We conduct experiments with different split numbers $K$ (1, 2, 4, 8) and different input point cloud numbers $N$ (2K, 10K, 20K, 40K) on the Car category of ShapeNet. The results are shown in Figure 8. Performance is enhanced with a growing number of splits, underscoring the efficacy of our split decoders, especially when it comes to sparse point clouds. + +Component Ablation. We first evaluate the effectiveness of our initialization in the 2D Gaussian prediction module. The visualization of initialized Gaussians is shown in Figure 3 (c), which indicates a coarse rendering result providing a great foundation for optimization. Without initialization, the performance of our method sharply drops as Table 5 shows. Randomly initialized Gaussians might not be visible in certain views, resulting in non-convergence, as shown in Figure 7 (d). The ablation studies of our entire-patch architecture contain two experiments: (1) removing the entire module $\mathcal{N}_e$ and utilizing the predicted incomplete images rendered from point patches for supervision, denoted as “No $\mathcal{N}_e$ ” in Table 5; (2) splitting the entire point cloud to multiple patches as same as the inference process of $\mathcal{N}_p$ to render complete images during training, denoted as “ $\mathcal{N}_p$ only” in Table 5. the “ $\mathcal{N}_p$ only” method will consume a sub + +Table 5. The ablation studies of our method. The ablated methods are trained on Car with 20K points. + +
PSNR ↑SSIM ↑LPIPS ↓
No Initialization9.620.6850.655
No Nc12.940.6760.441
Np only23.230.8290.126
Ours27.880.9490.068
+ +![](images/7bb36793d49020206ecd95505b0cb2dd2232f2e8ed64e162bbfbd6442670419f.jpg) +Figure 8. The ablation study of split number $K$ with different input point numbers. The legend in the lower right corner indicates the input point number for training and evaluating. + +stantial amount of resources for training and inference. The evaluation results of these two ablated methods are worse than our method, where their visualization results are shown in Figure 7(e)(f)(c). + +The results of ablation studies validate the effectiveness and necessity of the components within our method. + +# 5. Conclusion and Discussion + +In this paper, we introduce a novel approach featuring an entire-patch architecture and the 2D Gaussian prediction module for point cloud rendering. Our prediction module directly forecasts the 2D Gaussians from the provided point clouds, beginning with initialization derived from normals. The splitting decoders produce multiple Gaussians from a point, thereby better accommodating sparse point clouds. To enhance the generalization capability, the entire-patch architecture employs a module for "background" area prediction and another module for point cloud patch prediction, ensuring proper supervision. Comprehensive experiments and comparisons demonstrate the superiority and generalization capability of our method, which can render point clouds into photo-realistic images with clear details and achieves SOTA performance. A limitation of our method may arise when a portion of a point cloud is absent, as our method cannot render a complete image without the direct support of the points. Our future work aims to address this limitation by incorporating point cloud completion techniques and also attempting to render point clouds with noise, such as those directly scanned from wild and outdoor scenes. + +# Acknowledgments + +# References + +[1] Kara-Ali Aliev, Artem Sevastopolsky, Maria Kolos, Dmitry Ulyanov, and Victor Lempitsky. Neural point-based graphics. In Computer Vision - ECCV 2020, pages 696-712, Cham, 2020. Springer International Publishing. 1, 2 +[2] M. Botsch, A. Hornung, M. Zwicker, and L. Kobbelt. High-quality surface splatting on today's gpus. In Proceedings Eurographics/IEEE VGTC Symposium Point-Based Graphics, 2005., pages 17-141, 2005. 1, 2 +[3] Alexandre Boulch and Renaud Marlet. Poco: Point convolution for surface reconstruction. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6292-6304, 2022. 2, 4 +[4] Angel X. Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, Jianxiong Xiao, Li Yi, and Fisher Yu. Shapenet: An information-rich 3d model repository, 2015. 5, 7 +[5] Kaichun Mo Leonidas J. Guibas Charles R. Qi, Hao Su. Pointnet: Deep learning on point sets for 3d classification and segmentation. In 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2017. 4 +[6] Paolo Cignoni, Marco Callieri, Massimiliano Corsini, Matteo Dellepiane, Fabio Ganovelli, and Guido Ranzuglia. MeshLab: an Open-Source Mesh Processing Tool. In Eurographics Italian Chapter Conference. The Eurographics Association, 2008. 2 +[7] Angela Dai, Angel X. Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proc. Computer Vision and Pattern Recognition (CVPR), IEEE, 2017. 5, 7 +[8] Peng Dai, Yinda Zhang, Zhuwen Li, Shuaicheng Liu, and Bing Zeng. Neural point cloud rendering via multi-plane projection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7830-7839, 2020. 1, 2 +[9] Laura Downs, Anthony Francis, Nate Koenig, Brandon Kinman, Ryan Michael Hickman, Krista Reymann, Thomas Barlow McHugh, and Vincent Vanhoucke. Google scanned objects: A high-quality dataset of 3d scanned household items. 2022 International Conference on Robotics and Automation (ICRA), pages 2553-2560, 2022. 5, 7 +[10] Antoine Guédon and Vincent Lepetit. Sugar: Surface-aligned gaussian splatting for efficient 3d mesh reconstruction and high-quality mesh rendering. CVPR, 2024. 3 +[11] Hugues Hoppe, Tony DeRose, Tom Duchamp, John Alan McDonald, and Werner Stuetzle. Surface reconstruction from unorganized points. Proceedings of the 19th annual conference on Computer graphics and interactive techniques, 1992. 3 + +This work was supported by the National Natural Science Foundation of China under Grant 62032011. +[12] Qingyong Hu, Bo Yang, Linhai Xie, Stefano Rosa, Yulan Guo, Zhihua Wang, Niki Trigoni, and Andrew Markham. Randla-net: Efficient semantic segmentation of large-scale point clouds. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2020. 2, 4 +[13] Tao Hu, Xiaogang Xu, Ruihang Chu, and Jiaya Jia. Trivol: Point cloud rendering via triple volumes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20732-20741, 2023. 1, 2, 3, 5, 7, 8 +[14] T. Hu, Xiaogang Xu, Shu Liu, and Jiaya Jia. Point2pix: Photo-realistic point cloud rendering via neural radiance fields. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8349-8358, 2023. 1, 2, 3 +[15] Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2d gaussian splatting for geometrically accurate radiance fields. In SIGGRAPH 2024 Conference Papers. Association for Computing Machinery, 2024. 2, 3, 5, 7 +[16] Rasmus Jensen, Anders Dahl, George Vogiatzis, Engil Tola, and Henrik Aanæs. Large scale multi-view stereopsis evaluation. In 2014 IEEE Conference on Computer Vision and Pattern Recognition, pages 406-413, 2014. 5 +[17] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42 (4), 2023. 1, 2, 3 +[18] Leif Kobbelt and Mario Botsch. A survey of point-based techniques in computer graphics. Comput. Graph., 28(6): 801-814, 2004. 2 +[19] Xu Ma, Can Qin, Haoxuan You, Haoxi Ran, and Yun Fu. Rethinking network design and local geometry in point cloud: A simple residual mlp framework. *ICLR*, 2022. 4 +[20] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 1, 2 +[21] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Kopf, Edward Yang, Zachary DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. Pytorch: An imperative style, high-performance deep learning library. In Advances in Neural Information Processing Systems 32, pages 8024-8035. Curran Associates, Inc., 2019. 5 +[22] Hanspeter Pfister, Matthias Zwicker, Jeroen van Baar, and Markus H. Gross. Surfels: surface elements as rendering primitives. Proceedings of the 27th annual conference on Computer graphics and interactive techniques, 2000. 1, 2 +[23] Ruslan Rakhimov, Andrei-Timotei Ardelean, Victor Lempitsky, and Evgeny Burnaev. Npbg++: Accelerating neural point-based graphics. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15969-15979, 2022. 1, 2, 5, 7, 8 +[24] Nikhila Ravi, Jeremy Reizenstein, David Novotny, Taylor Gordon, Wan-Yen Lo, Justin Johnson, and Georgia + +Gkioxari. Accelerating 3d deep learning with pytorch3d. arXiv:2007.08501, 2020. 2 +[25] Liu Ren, Hanspeter Pfister, and Matthias Zwicker. Object space ewa surface splatting: A hardware accelerated approach to high quality point rendering. Computer Graphics Forum, 21(3):461-470, 2002. 2 +[26] Darius Rückert, Linus Franke, and Marc Stamminger. Adop: Approximate differentiable one-pixel point rendering. ACM Transactions on Graphics (ToG), 41(4):1-14, 2022. 2 +[27] Bane Sullivan and Alexander Kaszynski. PyVista: 3D plotting and mesh analysis through a streamlined interface for the Visualization Toolkit (VTK). Journal of Open Source Software, 4(37):1450, 2019. 2 +[28] Jiaxu Wang, Ziyi Zhang, Junhao He, and Renjing Xu. Pfgs: High fidelity point cloud rendering via feature splatting. ECCV 2024, 2024. 1, 2, 3, 5, 6, 7, 8 +[29] Zhou Wang, A.C. Bovik, H.R. Sheikh, and E.P. Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing, 13(4): 600-612, 2004. 5 +[30] Tao Yu, Zerong Zheng, Kaiwen Guo, Pengpeng Liu, Qionghai Dai, and Yebin Liu. Function4d: Real-time human volumetric capture from very sparse consumer rgbd sensors. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR2021), 2021. 5, 7 +[31] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 5 +[32] Qian-Yi Zhou, Jaesik Park, and Vladlen Koltun. Open3D: A modern library for 3D data processing. arXiv:1801.09847, 2018. 2, 3 +[33] Matthias Zwicker, Hanspeter Pfister, Jeroen Baar, and Markus Gross. Surface splattering. Proceedings of the ACM SIGGRAPH Conference on Computer Graphics, 2001, 2001. 2 \ No newline at end of file diff --git a/2025/Sparse Point Cloud Patches Rendering via Splitting 2D Gaussians/images.zip b/2025/Sparse Point Cloud Patches Rendering via Splitting 2D Gaussians/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..6c1b0debfe71f9c3a7c09d5599fba631c461db98 --- /dev/null +++ b/2025/Sparse Point Cloud Patches Rendering via Splitting 2D Gaussians/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6d91c9b9552cd38b161e355fc77bec00f8b52f813af910310f2b73b2bf2ac954 +size 953931 diff --git a/2025/Sparse Point Cloud Patches Rendering via Splitting 2D Gaussians/layout.json b/2025/Sparse Point Cloud Patches Rendering via Splitting 2D Gaussians/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..b2fae57d3a8014d7875ecd179b1f27c8db9ec277 --- /dev/null +++ b/2025/Sparse Point Cloud Patches Rendering via Splitting 2D Gaussians/layout.json @@ -0,0 +1,9249 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 102, + 103, + 507, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 103, + 507, + 121 + ], + "spans": [ + { + "bbox": [ + 102, + 103, + 507, + 121 + ], + "type": "text", + "content": "Sparse Point Cloud Patches Rendering via Splitting 2D Gaussians" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 100, + 142, + 510, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 142, + 510, + 172 + ], + "spans": [ + { + "bbox": [ + 100, + 142, + 510, + 172 + ], + "type": "text", + "content": "Changfeng Ma" + }, + { + "bbox": [ + 100, + 142, + 510, + 172 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 100, + 142, + 510, + 172 + ], + "type": "text", + "content": ", Ran Bi" + }, + { + "bbox": [ + 100, + 142, + 510, + 172 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 100, + 142, + 510, + 172 + ], + "type": "text", + "content": ", Jie Guo" + }, + { + "bbox": [ + 100, + 142, + 510, + 172 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 100, + 142, + 510, + 172 + ], + "type": "text", + "content": ", Chongjun Wang" + }, + { + "bbox": [ + 100, + 142, + 510, + 172 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 100, + 142, + 510, + 172 + ], + "type": "text", + "content": ", Yanwen Guo" + }, + { + "bbox": [ + 100, + 142, + 510, + 172 + ], + "type": "inline_equation", + "content": "^{12*}" + }, + { + "bbox": [ + 100, + 142, + 510, + 172 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 100, + 142, + 510, + 172 + ], + "type": "text", + "content": "Nanjing University, Nanjing, China \n" + }, + { + "bbox": [ + 100, + 142, + 510, + 172 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 100, + 142, + 510, + 172 + ], + "type": "text", + "content": "School of Software, North University of China" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 195, + 174, + 416, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 195, + 174, + 416, + 185 + ], + "spans": [ + { + "bbox": [ + 195, + 174, + 416, + 185 + ], + "type": "text", + "content": "{changfengma, 211250233}@smail.nju.edu.cn" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 215, + 187, + 392, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 215, + 187, + 392, + 199 + ], + "spans": [ + { + "bbox": [ + 215, + 187, + 392, + 199 + ], + "type": "text", + "content": "{guojie, chjwang, ywguo}@nju.edu.cn" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 85, + 217, + 144, + 272 + ], + "blocks": [ + { + "bbox": [ + 88, + 209, + 103, + 217 + ], + "lines": [ + { + "bbox": [ + 88, + 209, + 103, + 217 + ], + "spans": [ + { + "bbox": [ + 88, + 209, + 103, + 217 + ], + "type": "text", + "content": "Car" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 85, + 217, + 144, + 272 + ], + "lines": [ + { + "bbox": [ + 85, + 217, + 144, + 272 + ], + "spans": [ + { + "bbox": [ + 85, + 217, + 144, + 272 + ], + "type": "image", + "image_path": "9d4d6b033eabffbf029eaf37ab8a1448429297366823f9b8d6fa2a01ceac05ee.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 93, + 274, + 132, + 283 + ], + "lines": [ + { + "bbox": [ + 93, + 274, + 132, + 283 + ], + "spans": [ + { + "bbox": [ + 93, + 274, + 132, + 283 + ], + "type": "text", + "content": "Point Cloud" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 154, + 216, + 208, + 266 + ], + "blocks": [ + { + "bbox": [ + 151, + 209, + 183, + 217 + ], + "lines": [ + { + "bbox": [ + 151, + 209, + 183, + 217 + ], + "spans": [ + { + "bbox": [ + 151, + 209, + 183, + 217 + ], + "type": "text", + "content": "NPBG+±" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 154, + 216, + 208, + 266 + ], + "lines": [ + { + "bbox": [ + 154, + 216, + 208, + 266 + ], + "spans": [ + { + "bbox": [ + 154, + 216, + 208, + 266 + ], + "type": "image", + "image_path": "bf5f8ecb853a56f2c13091a2a43ee81d4ae0a6e3012240cb93d9c4696f7d0689.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 149, + 266, + 211, + 287 + ], + "blocks": [ + { + "bbox": [ + 149, + 266, + 211, + 287 + ], + "lines": [ + { + "bbox": [ + 149, + 266, + 211, + 287 + ], + "spans": [ + { + "bbox": [ + 149, + 266, + 211, + 287 + ], + "type": "image", + "image_path": "591b037e67a826ba4bf302c9828820d4f1bbd5eeb832e4072bb62dba7c4aedbb.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 220, + 216, + 274, + 266 + ], + "blocks": [ + { + "bbox": [ + 217, + 209, + 241, + 217 + ], + "lines": [ + { + "bbox": [ + 217, + 209, + 241, + 217 + ], + "spans": [ + { + "bbox": [ + 217, + 209, + 241, + 217 + ], + "type": "text", + "content": "TriVol" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 220, + 216, + 274, + 266 + ], + "lines": [ + { + "bbox": [ + 220, + 216, + 274, + 266 + ], + "spans": [ + { + "bbox": [ + 220, + 216, + 274, + 266 + ], + "type": "image", + "image_path": "bc7518ba8c933acdda4fa80540ae016e91b2a7ba9ab93f2571a60109a6c64ff9.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 216, + 266, + 278, + 287 + ], + "blocks": [ + { + "bbox": [ + 216, + 266, + 278, + 287 + ], + "lines": [ + { + "bbox": [ + 216, + 266, + 278, + 287 + ], + "spans": [ + { + "bbox": [ + 216, + 266, + 278, + 287 + ], + "type": "image", + "image_path": "ad480242986f91afb37657047c5966ade456c4b558d3075eb78c6436e8733a09.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 88, + 297, + 142, + 347 + ], + "blocks": [ + { + "bbox": [ + 85, + 290, + 105, + 298 + ], + "lines": [ + { + "bbox": [ + 85, + 290, + 105, + 298 + ], + "spans": [ + { + "bbox": [ + 85, + 290, + 105, + 298 + ], + "type": "text", + "content": "PFGS" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 88, + 297, + 142, + 347 + ], + "lines": [ + { + "bbox": [ + 88, + 297, + 142, + 347 + ], + "spans": [ + { + "bbox": [ + 88, + 297, + 142, + 347 + ], + "type": "image", + "image_path": "15a6c7afc42cd95fa20fe573507e1d61feb5c739cdfd5d4d48dcf993d02d1e98.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 83, + 348, + 144, + 369 + ], + "blocks": [ + { + "bbox": [ + 83, + 348, + 144, + 369 + ], + "lines": [ + { + "bbox": [ + 83, + 348, + 144, + 369 + ], + "spans": [ + { + "bbox": [ + 83, + 348, + 144, + 369 + ], + "type": "image", + "image_path": "b1220996de86ded1acbef8a69f08538b905c0c6ba1e8c6a2734f1207b379fbcf.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 378, + 555, + 401 + ], + "lines": [ + { + "bbox": [ + 55, + 378, + 555, + 401 + ], + "spans": [ + { + "bbox": [ + 55, + 378, + 555, + 401 + ], + "type": "text", + "content": "Figure 1. Our method excels in rendering intricate, photo-realistic images from point clouds of different categories. In this instance, our model is trained on the Car category, utilizing 20K points as input." + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 154, + 297, + 208, + 347 + ], + "blocks": [ + { + "bbox": [ + 151, + 291, + 168, + 298 + ], + "lines": [ + { + "bbox": [ + 151, + 291, + 168, + 298 + ], + "spans": [ + { + "bbox": [ + 151, + 291, + 168, + 298 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 154, + 297, + 208, + 347 + ], + "lines": [ + { + "bbox": [ + 154, + 297, + 208, + 347 + ], + "spans": [ + { + "bbox": [ + 154, + 297, + 208, + 347 + ], + "type": "image", + "image_path": "7659f8d1a9478ac4cafb70a5ec795e9e447c8ef8389ec92774c739b32e421efa.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 149, + 348, + 211, + 369 + ], + "blocks": [ + { + "bbox": [ + 149, + 348, + 211, + 369 + ], + "lines": [ + { + "bbox": [ + 149, + 348, + 211, + 369 + ], + "spans": [ + { + "bbox": [ + 149, + 348, + 211, + 369 + ], + "type": "image", + "image_path": "f25f999b074e6f5d71444861da5a33b76655a61a669a79cdcb6be4623646d986.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 221, + 295, + 274, + 346 + ], + "blocks": [ + { + "bbox": [ + 218, + 291, + 230, + 298 + ], + "lines": [ + { + "bbox": [ + 218, + 291, + 230, + 298 + ], + "spans": [ + { + "bbox": [ + 218, + 291, + 230, + 298 + ], + "type": "text", + "content": "GT" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 221, + 295, + 274, + 346 + ], + "lines": [ + { + "bbox": [ + 221, + 295, + 274, + 346 + ], + "spans": [ + { + "bbox": [ + 221, + 295, + 274, + 346 + ], + "type": "image", + "image_path": "d33e9d13af948c7082cbd821c671c6e9f444f1c88395d45f44dfab564002ac9f.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 215, + 348, + 278, + 369 + ], + "blocks": [ + { + "bbox": [ + 215, + 348, + 278, + 369 + ], + "lines": [ + { + "bbox": [ + 215, + 348, + 278, + 369 + ], + "spans": [ + { + "bbox": [ + 215, + 348, + 278, + 369 + ], + "type": "image", + "image_path": "996bbb729c98e5b0318897b6919590734b0358864ed5a774ca336f04f6b58a56.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 299, + 217, + 383, + 299 + ], + "blocks": [ + { + "bbox": [ + 357, + 208, + 473, + 217 + ], + "lines": [ + { + "bbox": [ + 357, + 208, + 473, + 217 + ], + "spans": [ + { + "bbox": [ + 357, + 208, + 473, + 217 + ], + "type": "text", + "content": "Generalize to Other Categories" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 299, + 217, + 383, + 299 + ], + "lines": [ + { + "bbox": [ + 299, + 217, + 383, + 299 + ], + "spans": [ + { + "bbox": [ + 299, + 217, + 383, + 299 + ], + "type": "image", + "image_path": "c4e48cfab5721918aad94808b29ad87164af7eb18d9f79af61d8c94f7accefe7.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 298, + 300, + 383, + 369 + ], + "blocks": [ + { + "bbox": [ + 298, + 300, + 383, + 369 + ], + "lines": [ + { + "bbox": [ + 298, + 300, + 383, + 369 + ], + "spans": [ + { + "bbox": [ + 298, + 300, + 383, + 369 + ], + "type": "image", + "image_path": "d294a263598f72cd8895f81e57fd6f224b2b8b15f677352d7ee3768d05cc5b86.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 384, + 217, + 463, + 311 + ], + "blocks": [ + { + "bbox": [ + 384, + 217, + 463, + 311 + ], + "lines": [ + { + "bbox": [ + 384, + 217, + 463, + 311 + ], + "spans": [ + { + "bbox": [ + 384, + 217, + 463, + 311 + ], + "type": "image", + "image_path": "2ea980b1f830657c635adfc199d2f0b7234b3375a833be1d0b788853f0dfddc5.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 386, + 311, + 432, + 320 + ], + "lines": [ + { + "bbox": [ + 386, + 311, + 432, + 320 + ], + "spans": [ + { + "bbox": [ + 386, + 311, + 432, + 320 + ], + "type": "text", + "content": "Sculpture" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 385, + 319, + 432, + 368 + ], + "blocks": [ + { + "bbox": [ + 385, + 319, + 432, + 368 + ], + "lines": [ + { + "bbox": [ + 385, + 319, + 432, + 368 + ], + "spans": [ + { + "bbox": [ + 385, + 319, + 432, + 368 + ], + "type": "image", + "image_path": "756cc0419e4f854567bf136d5e3b7cd287982c77e453943d2447eabbc11b71ec.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 466, + 217, + 526, + 261 + ], + "blocks": [ + { + "bbox": [ + 466, + 217, + 526, + 261 + ], + "lines": [ + { + "bbox": [ + 466, + 217, + 526, + 261 + ], + "spans": [ + { + "bbox": [ + 466, + 217, + 526, + 261 + ], + "type": "image", + "image_path": "56228bb22cd5c45271e6adc5f71f364b3a7a8dde46090ef763073123f11c4286.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 467, + 262, + 526, + 311 + ], + "blocks": [ + { + "bbox": [ + 467, + 262, + 526, + 311 + ], + "lines": [ + { + "bbox": [ + 467, + 262, + 526, + 311 + ], + "spans": [ + { + "bbox": [ + 467, + 262, + 526, + 311 + ], + "type": "image", + "image_path": "2bcc0458d0620247aaefe164414bb61082bc343a159bd2957a268515dbb7c507.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 435, + 312, + 490, + 368 + ], + "blocks": [ + { + "bbox": [ + 435, + 312, + 490, + 368 + ], + "lines": [ + { + "bbox": [ + 435, + 312, + 490, + 368 + ], + "spans": [ + { + "bbox": [ + 435, + 312, + 490, + 368 + ], + "type": "image", + "image_path": "b9669c9491e4f05faec33dcb8eb7c6ed71fd289b25d545fa2b11802cc558ed0e.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 492, + 312, + 527, + 339 + ], + "blocks": [ + { + "bbox": [ + 492, + 312, + 527, + 339 + ], + "lines": [ + { + "bbox": [ + 492, + 312, + 527, + 339 + ], + "spans": [ + { + "bbox": [ + 492, + 312, + 527, + 339 + ], + "type": "image", + "image_path": "dc52fd9874f362f2def01a2be70f3b28299b1a60fdc3c862d6c37bf01ae6279b.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + } + ], + "index": 33 + }, + { + "type": "image", + "bbox": [ + 492, + 340, + 527, + 368 + ], + "blocks": [ + { + "bbox": [ + 492, + 340, + 527, + 368 + ], + "lines": [ + { + "bbox": [ + 492, + 340, + 527, + 368 + ], + "spans": [ + { + "bbox": [ + 492, + 340, + 527, + 368 + ], + "type": "image", + "image_path": "2bd0f885f9ed91620037d0f37d4edbb7eb87f9e3edaa5f727e6e7ff75a2b8924.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + } + ], + "index": 34 + }, + { + "bbox": [ + 152, + 418, + 200, + 431 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 418, + 200, + 431 + ], + "spans": [ + { + "bbox": [ + 152, + 418, + 200, + 431 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 54, + 444, + 297, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 444, + 297, + 696 + ], + "spans": [ + { + "bbox": [ + 54, + 444, + 297, + 696 + ], + "type": "text", + "content": "Current learning-based methods predict NeRF or 3D Gaussians from point clouds to achieve photo-realistic rendering but still depend on categorical priors, dense point clouds, or additional refinements. Hence, we introduce a novel point cloud rendering method by predicting 2D Gaussians from point clouds. Our method incorporates two identical modules with an entire-patch architecture enabling the network to be generalized to multiple datasets. The module normalizes and initializes the Gaussians utilizing the point cloud information including normals, colors and distances. Then, splitting decoders are employed to refine the initial Gaussians by duplicating them and predicting more accurate results, making our methodology effectively accommodate sparse point clouds as well. Once trained, our approach exhibits direct generalization to point clouds across different categories. The predicted Gaussians are employed directly for rendering without additional refinement on the rendered images, retaining the benefits of 2D Gaussians. We conduct extensive experiments on various datasets, and the results demonstrate the superiority and generalization of our method, which achieves SOTA performance. The code is" + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 314, + 420, + 550, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 420, + 550, + 432 + ], + "spans": [ + { + "bbox": [ + 314, + 420, + 550, + 432 + ], + "type": "text", + "content": "available at https://github.com/murcherful/GauPCRender." + } + ] + } + ], + "index": 38 + }, + { + "bbox": [ + 314, + 454, + 394, + 466 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 454, + 394, + 466 + ], + "spans": [ + { + "bbox": [ + 314, + 454, + 394, + 466 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 39 + }, + { + "bbox": [ + 312, + 474, + 555, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 474, + 555, + 558 + ], + "spans": [ + { + "bbox": [ + 312, + 474, + 555, + 558 + ], + "type": "text", + "content": "Point clouds can be easily acquired using various types of 3D scanners, making them a recent research hotspot. Photorealistic rendering of point clouds is significant for many applications such as visualization, virtual reality and automatic driving. However, the sparsity and discrete nature of point clouds continue to pose challenges for achieving photo-realistic rendering." + } + ] + } + ], + "index": 40 + }, + { + "bbox": [ + 312, + 558, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 558, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 312, + 558, + 556, + 714 + ], + "type": "text", + "content": "Traditional graphics rendering methods [2, 22] merely project points as planes or spheres onto images. Consequently, their outcomes lack photo-realism and are riddled with numerous holes. Previous deep learning-based works [1, 8, 23] necessitate point clouds and corresponding images from several views as input, which are difficult to acquire in practice. These methods also require training or fine-tuning for each scene or object. Recent works [13, 14] have focused on predicting NeRF [20] from given point clouds to render images from arbitrary views. PFGS [28] predicts the 3D Gaussians [17] with features for point clouds. After rendering, PFGS refines the rendered images based on these features. Although these methods can render point clouds" + } + ] + } + ], + "index": 41 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 70, + 703, + 149, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 703, + 149, + 713 + ], + "spans": [ + { + "bbox": [ + 70, + 703, + 149, + 713 + ], + "type": "text", + "content": "*Corresponding author." + } + ] + } + ], + "index": 42 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27285" + } + ] + } + ], + "index": 43 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 152, + 296, + 289 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 294, + 148 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 294, + 148 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 294, + 148 + ], + "type": "text", + "content": "Table 1. The comparison of different methods across several dimensions. Here, \"S. S/O.\", \"S. C.\" and \"M. C.\" represent \"single scene or object\", \"single category\" and \"multiple categories\", respectively. " + }, + { + "bbox": [ + 55, + 70, + 294, + 148 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 55, + 70, + 294, + 148 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 70, + 294, + 148 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 55, + 70, + 294, + 148 + ], + "type": "text", + "content": " indicate point clouds and images, separately. " + }, + { + "bbox": [ + 55, + 70, + 294, + 148 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 55, + 70, + 294, + 148 + ], + "type": "text", + "content": ": Traditional graphics rendering approach is unable to render photo-realistic images. " + }, + { + "bbox": [ + 55, + 70, + 294, + 148 + ], + "type": "inline_equation", + "content": "\\ddagger" + }, + { + "bbox": [ + 55, + 70, + 294, + 148 + ], + "type": "text", + "content": ": PFGS requires further refinement on images rendered from 3D gaussians." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 152, + 296, + 289 + ], + "lines": [ + { + "bbox": [ + 56, + 152, + 296, + 289 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 296, + 289 + ], + "type": "table", + "html": "
MethodGeneralization abilityInputOutputPre- Fine- trainFine- tunePoint number
\\( GR^† \\)AnyPI✘✘Any
NPBG [1]S. S/O.\\( P+I \\)I✓✓100K
NPCR [8]S. S/O.\\( P+I \\)I✘✓100K
NPBG++ [23]S. S/O.\\( P+I \\)I✓✘100K
TriVol [13]S. C.PNeRF[20]✓✘100K
Point2Pix [14]S. C.PNeRF[20]✓✘100K
PFGS [28]S. C.P3DGS[17]‡✓✘80K
OursM. C.P2DGS[15]✓✘2K-100K
", + "image_path": "a57b6cd5ae8ccd58bd3dd30d48d68c03884fb676e453c9c514625135e3b57b0f.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 305, + 295, + 341 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 305, + 295, + 341 + ], + "spans": [ + { + "bbox": [ + 55, + 305, + 295, + 341 + ], + "type": "text", + "content": "without training or fine-tuning on each data, they still have limitations, such as the need for dense point clouds, poor generalization, slow rendering speeds, and blurry results." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 342, + 295, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 342, + 295, + 426 + ], + "spans": [ + { + "bbox": [ + 55, + 342, + 295, + 426 + ], + "type": "text", + "content": "2D Gaussian Splatting [15] has recently been proposed, which includes explicit normals of Gaussians. These normals facilitate easy initialization, learning, and prediction by the network from point clouds. Motivated by this, we introduce a novel method to predict the 2D Gaussians of the given point cloud and rasterize the predicted 2D Gaussians into images for photo-realistic point cloud rendering." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 426, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 426, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 426, + 295, + 714 + ], + "type": "text", + "content": "Due to taking entire point clouds as input, previous methods are hard to generalize to other categories or adapt to different point distributions. Processing only point cloud patches with the network can enhance its generalization capability and reduce dependence on category priors [3, 12]. However, rendering Gaussians predicted from patch points results in incomplete images that are unsupervisible. Therefore, we employ an entire-patch architecture that includes two identical 2D Gaussian prediction modules to handle entire and patch point clouds, separately. We use the rendering results of the entire point cloud as a \"background\" to obtain a complete image for proper supervision. The module first normalizes and initializes the 2D Gaussians based on the given point clouds, encompassing the estimated normals, scales, and colors, which are crucial for our network to converge and predict more accurate results. Subsequently, a point cloud encoder and multiple splitting decoders are employed to extract local features and predict the parameters of the duplicated Gaussians. Equipped with the splitting decoder, our network can handle sparse point clouds and produce high-quality details with dense Gaussians. Thanks to the proposed architecture, our method also exhibits strong generalization across scenes and objects. Without further refinement on rendered images, our" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "content": "method retains the advantages of 2D Gaussian Splatting in rendering, such as rapid rendering speed." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 97, + 553, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 97, + 553, + 168 + ], + "spans": [ + { + "bbox": [ + 313, + 97, + 553, + 168 + ], + "type": "text", + "content": "We conduct extensive comparative and ablation experiments, and the results demonstrate the superiority of our method, which can render point clouds into high-quality detailed images with excellent generalization. We will make our dataset and codes public in the future. Our main contributions are as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 170, + 553, + 289 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 314, + 170, + 553, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 170, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 314, + 170, + 553, + 217 + ], + "type": "text", + "content": "- We introduce a method that directly predicts 2D Gaussians from point clouds to render photo-realistic images without any refinement, while retaining the advantages of 2D Gaussian Splatting." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 218, + 553, + 253 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 218, + 553, + 253 + ], + "spans": [ + { + "bbox": [ + 314, + 218, + 553, + 253 + ], + "type": "text", + "content": "- Equipped with an entire-patch architecture, our method can handle point cloud patches, leading to significant generalization across scenes and objects." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 254, + 553, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 254, + 553, + 289 + ], + "spans": [ + { + "bbox": [ + 314, + 254, + 553, + 289 + ], + "type": "text", + "content": "- We introduce the splitting decoder, which splits Gaussians into denser outcomes, enabling our method to render detailed images from even sparse point clouds." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 302, + 400, + 316 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 302, + 400, + 316 + ], + "spans": [ + { + "bbox": [ + 314, + 302, + 400, + 316 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 323, + 446, + 336 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 323, + 446, + 336 + ], + "spans": [ + { + "bbox": [ + 314, + 323, + 446, + 336 + ], + "type": "text", + "content": "2.1. Point Cloud Rendering" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 342, + 553, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 342, + 553, + 473 + ], + "spans": [ + { + "bbox": [ + 313, + 342, + 553, + 473 + ], + "type": "text", + "content": "Traditional Point Cloud Rendering. Point cloud rendering has consistently been a significant research topic [18]. Conventional point cloud rendering techniques [6, 24, 27, 32] directly rasterize points onto 2D screens for rendering, yet they grapple with the issue of holes. Subsequent studies [2, 22, 25, 33] have proposed splatting points with elliptic discs, ellipsoids, or surfels for point cloud rendering to mitigate the holes observed during the rendering process. Although these graphic rendering methods can render any point cloud, the resulting images often contain holes and exhibit blurry details, lacking photo-realism." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 475, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 475, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 475, + 553, + 713 + ], + "type": "text", + "content": "Point-based novel view synthesis. For photo-realistic point cloud rendering, several recently proposed methods utilize corresponding images of point clouds as additional information. NPBG [1] signs descriptors to point clouds and rasterizes into multi-resolution raw images with features and employs a U-Net to render the final results. Dai et al. [8] aggregate points with features onto multiple planes based on a given view and point cloud. Following multi-plane-based voxelization, they employ a 3D U-Net to predict the rendered image. ADOP [26] is introduced for rendering HD images from point clouds, which employs several global parameters, a neural renderer and a differentiable physically-based tonemapper for optimization. These methods necessitate training for each point cloud with images from several views to render images of novel views, which is a time-consuming process. To reduce training time, NPBG++ [23] is trained on a collection of point clouds using a network that extracts features from corresponding images. This network aligns the features of input images to point clouds without the need for fine-tuning. However, this" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "27286" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 95 + ], + "type": "text", + "content": "method still requires images as input, which can be challenging to obtain in practice." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 96, + 295, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 96, + 295, + 312 + ], + "spans": [ + { + "bbox": [ + 55, + 96, + 295, + 312 + ], + "type": "text", + "content": "Learning-based Point Cloud Rendering. Several works [13, 14, 28] are proposed for rendering point clouds without the need for training or fine-tuning on each point cloud and corresponding images. TriVol [13] utilizes triple-plane-based grouping to query point features for predicting a NeRF from a given point cloud. Hu et al. [14] propose multi-scale radiance fields to predict NeRFs from point clouds. They also introduce a sampling strategy to accelerate fine-tuning and rendering speed. PFGS [28], a two-stage approach, for the first time employs feature splatting to predict 3D Gaussians from point clouds and uses a recurrent decoder to refine the rendered images. These methods are capable of rendering point clouds without fine-tuning or the need for input images, yet they still depend on categorical priors, dense points, or further refinements. Table 1 summarizes the distinctions between our method and the previous methods across several dimensions, highlighting the innovation and superiority of our proposed approach." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 319, + 167, + 331 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 319, + 167, + 331 + ], + "spans": [ + { + "bbox": [ + 55, + 319, + 167, + 331 + ], + "type": "text", + "content": "2.2. Gaussian Splitting" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 335, + 295, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 335, + 295, + 574 + ], + "spans": [ + { + "bbox": [ + 55, + 335, + 295, + 574 + ], + "type": "text", + "content": "3D Gaussian Splitting [17] is proposed with remarkable performance in novel view synthesis and further explores the rendering capability of the point clouds. 3DGS utilizes the differentiable splatting rendering to optimize 3D Gaussian primitives initialized from point clouds. Scenes are explicitly represented with 3D Gaussians that can be rendered with remarkable speed. Meshes can also be reconstructed from 3D Gaussians[10]. However, the normals of 3D Gaussians are so vague that the resulting geometries lack precision. To generate more accurate geometries, Hua et al. [15] propose 2DGS that employs 2D Gaussian disks containing explicit normals as representation primitives. With precise normals, meshes can accurately reconstruct the geometry of objects. Given that the normals of 2D Gaussians can be readily determined, 2D Gaussians can also be initialized using normals. Luckily, the normals of points can be effortlessly estimated [11] facilitating the initialization and forecasting of 2D Gaussians. Motivated by this, we opt to predict 2D Gaussians in place of 3D Gaussians from the provided point clouds." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 586, + 111, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 586, + 111, + 597 + ], + "spans": [ + { + "bbox": [ + 55, + 586, + 111, + 597 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 605, + 294, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 294, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 294, + 677 + ], + "type": "text", + "content": "Given a point cloud " + }, + { + "bbox": [ + 55, + 605, + 294, + 677 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 55, + 605, + 294, + 677 + ], + "type": "text", + "content": " and a camera " + }, + { + "bbox": [ + 55, + 605, + 294, + 677 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 55, + 605, + 294, + 677 + ], + "type": "text", + "content": ", point cloud rendering aims to render a photo-realistic image " + }, + { + "bbox": [ + 55, + 605, + 294, + 677 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 55, + 605, + 294, + 677 + ], + "type": "text", + "content": ". Here, point cloud " + }, + { + "bbox": [ + 55, + 605, + 294, + 677 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 55, + 605, + 294, + 677 + ], + "type": "text", + "content": " contains of coordinates of 3D points " + }, + { + "bbox": [ + 55, + 605, + 294, + 677 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{\\mathbf{x}} \\in \\mathbb{R}_{N \\times 3}" + }, + { + "bbox": [ + 55, + 605, + 294, + 677 + ], + "type": "text", + "content": " and colors of points " + }, + { + "bbox": [ + 55, + 605, + 294, + 677 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{\\mathbf{c}} \\in [0,1]_{N \\times 3}" + }, + { + "bbox": [ + 55, + 605, + 294, + 677 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 605, + 294, + 677 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 605, + 294, + 677 + ], + "type": "text", + "content": " indicates the point number of " + }, + { + "bbox": [ + 55, + 605, + 294, + 677 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 55, + 605, + 294, + 677 + ], + "type": "text", + "content": ". Camera " + }, + { + "bbox": [ + 55, + 605, + 294, + 677 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 55, + 605, + 294, + 677 + ], + "type": "text", + "content": " consists of intrinsic matrix " + }, + { + "bbox": [ + 55, + 605, + 294, + 677 + ], + "type": "inline_equation", + "content": "\\mathbf{K} \\in \\mathbb{R}_{3 \\times 3}" + }, + { + "bbox": [ + 55, + 605, + 294, + 677 + ], + "type": "text", + "content": " and extrinsic matrix " + }, + { + "bbox": [ + 55, + 605, + 294, + 677 + ], + "type": "inline_equation", + "content": "\\mathbf{P} \\in \\mathbb{R}_{4 \\times 4}" + }, + { + "bbox": [ + 55, + 605, + 294, + 677 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 678, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 678, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 678, + 295, + 713 + ], + "type": "text", + "content": "For photo-realistic point cloud rendering, we propose a network " + }, + { + "bbox": [ + 55, + 678, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{N}" + }, + { + "bbox": [ + 55, + 678, + 295, + 713 + ], + "type": "text", + "content": " to predict 2D Gaussians " + }, + { + "bbox": [ + 55, + 678, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{G} = \\mathcal{N}(\\mathcal{P})" + }, + { + "bbox": [ + 55, + 678, + 295, + 713 + ], + "type": "text", + "content": " given point cloud " + }, + { + "bbox": [ + 55, + 678, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 55, + 678, + 295, + 713 + ], + "type": "text", + "content": ", and employ differentiable splatting rendering [15]" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "inline_equation", + "content": "f_{R}" + }, + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "text", + "content": " to synthesize images " + }, + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "inline_equation", + "content": "\\mathcal{I} = f_{R}(\\mathcal{G},\\mathcal{C})" + }, + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "text", + "content": ". The parameters of a 2D Gaussian " + }, + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "inline_equation", + "content": "\\mathbf{g} = \\{\\mathbf{x},\\mathbf{s},o,\\mathbf{c},\\mathbf{n},\\alpha \\} \\in \\mathcal{G}" + }, + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "text", + "content": " include: (1) position " + }, + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "inline_equation", + "content": "\\mathbf{x}\\in \\mathbb{R}_3" + }, + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "text", + "content": ", (2) scale " + }, + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "inline_equation", + "content": "\\mathbf{s}\\in \\mathbb{R}_2" + }, + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "text", + "content": ", (3) opacity " + }, + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "inline_equation", + "content": "o\\in [0,1]" + }, + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "text", + "content": ", (4) spherical harmonic of color " + }, + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "inline_equation", + "content": "\\mathbf{c}\\in \\mathbb{R}_d" + }, + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "text", + "content": ", (5) normal " + }, + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "inline_equation", + "content": "\\mathbf{n}\\in \\mathbb{R}_3" + }, + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "text", + "content": " and (6) rotation angle along normal " + }, + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "inline_equation", + "content": "\\alpha \\in [0,2\\pi ]" + }, + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "text", + "content": " represents the parameter number of spherical harmonic for color. As shown in Figure 2, our network contains two identical 2D Gaussian prediction modules (3.1) with an entire-patch architecture (3.2)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 189, + 486, + 200 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 189, + 486, + 200 + ], + "spans": [ + { + "bbox": [ + 313, + 189, + 486, + 200 + ], + "type": "text", + "content": "3.1.2D Gaussian Prediction Module" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 207, + 468, + 217 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 207, + 468, + 217 + ], + "spans": [ + { + "bbox": [ + 313, + 207, + 468, + 217 + ], + "type": "text", + "content": "3.1.1. Initialization of 2D Gaussians" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 221, + 553, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 221, + 553, + 666 + ], + "spans": [ + { + "bbox": [ + 313, + 221, + 553, + 666 + ], + "type": "text", + "content": "We first introduce 2D Gaussian prediction module to predict the 2D Gaussians " + }, + { + "bbox": [ + 313, + 221, + 553, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 313, + 221, + 553, + 666 + ], + "type": "text", + "content": " given a point cloud " + }, + { + "bbox": [ + 313, + 221, + 553, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 313, + 221, + 553, + 666 + ], + "type": "text", + "content": " as input. The positions of the point cloud " + }, + { + "bbox": [ + 313, + 221, + 553, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 313, + 221, + 553, + 666 + ], + "type": "text", + "content": " are normalized to the range " + }, + { + "bbox": [ + 313, + 221, + 553, + 666 + ], + "type": "inline_equation", + "content": "[-1,1]" + }, + { + "bbox": [ + 313, + 221, + 553, + 666 + ], + "type": "text", + "content": " based on the geometric center " + }, + { + "bbox": [ + 313, + 221, + 553, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{c} \\in \\mathbb{R}_3" + }, + { + "bbox": [ + 313, + 221, + 553, + 666 + ], + "type": "text", + "content": " and scale " + }, + { + "bbox": [ + 313, + 221, + 553, + 666 + ], + "type": "inline_equation", + "content": "s \\in R" + }, + { + "bbox": [ + 313, + 221, + 553, + 666 + ], + "type": "text", + "content": " of the input point cloud to facilitate easier feature extraction and prediction of results, where the scale is defined as the maximum distance between any point and " + }, + { + "bbox": [ + 313, + 221, + 553, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 313, + 221, + 553, + 666 + ], + "type": "text", + "content": ". Our module then initializes one 2D Gaussian " + }, + { + "bbox": [ + 313, + 221, + 553, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{g}" + }, + { + "bbox": [ + 313, + 221, + 553, + 666 + ], + "type": "text", + "content": " for each normalized point " + }, + { + "bbox": [ + 313, + 221, + 553, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{p} = \\{\\mathbf{p}_{\\mathbf{x}}, \\mathbf{p}_{\\mathbf{c}}\\} \\in \\mathcal{P}" + }, + { + "bbox": [ + 313, + 221, + 553, + 666 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 221, + 553, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_{\\mathbf{x}} \\in \\mathcal{R}_3" + }, + { + "bbox": [ + 313, + 221, + 553, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 221, + 553, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_{\\mathbf{c}} \\in \\mathcal{R}_3" + }, + { + "bbox": [ + 313, + 221, + 553, + 666 + ], + "type": "text", + "content": " indicate the position and color of the point respectively. The position " + }, + { + "bbox": [ + 313, + 221, + 553, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 313, + 221, + 553, + 666 + ], + "type": "text", + "content": " of the Gaussian is initialized with " + }, + { + "bbox": [ + 313, + 221, + 553, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_{\\mathbf{x}}" + }, + { + "bbox": [ + 313, + 221, + 553, + 666 + ], + "type": "text", + "content": ". The spherical harmonic " + }, + { + "bbox": [ + 313, + 221, + 553, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 313, + 221, + 553, + 666 + ], + "type": "text", + "content": " and scale " + }, + { + "bbox": [ + 313, + 221, + 553, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{s}" + }, + { + "bbox": [ + 313, + 221, + 553, + 666 + ], + "type": "text", + "content": " are initialized as the same as 2DGS [15] according to " + }, + { + "bbox": [ + 313, + 221, + 553, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_{\\mathbf{c}}" + }, + { + "bbox": [ + 313, + 221, + 553, + 666 + ], + "type": "text", + "content": " and the minimum distance between points, separately. The opacity " + }, + { + "bbox": [ + 313, + 221, + 553, + 666 + ], + "type": "inline_equation", + "content": "o" + }, + { + "bbox": [ + 313, + 221, + 553, + 666 + ], + "type": "text", + "content": " is set to 1 to ensure that all Gaussians are visible. As we mentioned before, the geometry of given point cloud " + }, + { + "bbox": [ + 313, + 221, + 553, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 313, + 221, + 553, + 666 + ], + "type": "text", + "content": " provides important clues for the rotation of Gaussians. Gaussian Splatting works [15, 17] utilize quaternions to represent the rotation of Gaussians and only randomly initialize them as Figure 3(b) shows, which are difficult for networks to predict intuitively. Fortunately, the normals of given point cloud can be easily estimated. Therefore, we represent the rotation of the Gaussian in terms of the normal " + }, + { + "bbox": [ + 313, + 221, + 553, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{n}" + }, + { + "bbox": [ + 313, + 221, + 553, + 666 + ], + "type": "text", + "content": " and the rotation angle " + }, + { + "bbox": [ + 313, + 221, + 553, + 666 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 221, + 553, + 666 + ], + "type": "text", + "content": " around that normal, and initialize the normal of the Gaussian using the estimated normal of the point cloud. Figure 3 (a) and (c) illustrate this initialization process. Specifically, we estimate the normals for " + }, + { + "bbox": [ + 313, + 221, + 553, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 313, + 221, + 553, + 666 + ], + "type": "text", + "content": " [11, 32], where the normal " + }, + { + "bbox": [ + 313, + 221, + 553, + 666 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{n}}" + }, + { + "bbox": [ + 313, + 221, + 553, + 666 + ], + "type": "text", + "content": " for each point " + }, + { + "bbox": [ + 313, + 221, + 553, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 313, + 221, + 553, + 666 + ], + "type": "text", + "content": " is utilized to initialize the normal " + }, + { + "bbox": [ + 313, + 221, + 553, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{n}" + }, + { + "bbox": [ + 313, + 221, + 553, + 666 + ], + "type": "text", + "content": " of Gaussian. And the rotation angle " + }, + { + "bbox": [ + 313, + 221, + 553, + 666 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 221, + 553, + 666 + ], + "type": "text", + "content": " of Gaussian is set to 0. The estimated normal may not be precise, but it provides crucial information and ensures that each Gaussian is visible, thereby guaranteeing that the network can converge correctly. After initializing for all the points, the initialized Gaussians " + }, + { + "bbox": [ + 313, + 221, + 553, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{G}^i" + }, + { + "bbox": [ + 313, + 221, + 553, + 666 + ], + "type": "text", + "content": " composing of the initialized positions, normals, scales, spherical harmonics, opacities and rotation angles (" + }, + { + "bbox": [ + 313, + 221, + 553, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{X}^i, \\mathbf{N}^i \\in \\mathbb{R}_{N\\times 3}, \\mathbf{S}^i \\in \\mathbb{R}_{N\\times 2}, \\mathbf{C}^i \\in \\mathbb{R}_{N\\times 3}, O^i, A^i \\in \\mathbb{R}_N" + }, + { + "bbox": [ + 313, + 221, + 553, + 666 + ], + "type": "text", + "content": ") is passed to our network as input and predicting basis." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 667, + 417, + 678 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 667, + 417, + 678 + ], + "spans": [ + { + "bbox": [ + 313, + 667, + 417, + 678 + ], + "type": "text", + "content": "3.1.2. Splitting Decoder" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 678, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 678, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 678, + 553, + 713 + ], + "type": "text", + "content": "As shown in Figure 2, local features are required before splitting and predicting 2D Gaussians. To extract features from the initialized 2D Gaussians, we employ" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27287" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 72, + 390, + 192 + ], + "blocks": [ + { + "bbox": [ + 60, + 72, + 390, + 192 + ], + "lines": [ + { + "bbox": [ + 60, + 72, + 390, + 192 + ], + "spans": [ + { + "bbox": [ + 60, + 72, + 390, + 192 + ], + "type": "image", + "image_path": "554cdefc4dec6134350b048c8b7bec91da332ec181726e8f215df849700885ea.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 394, + 72, + 550, + 192 + ], + "blocks": [ + { + "bbox": [ + 394, + 72, + 550, + 192 + ], + "lines": [ + { + "bbox": [ + 394, + 72, + 550, + 192 + ], + "spans": [ + { + "bbox": [ + 394, + 72, + 550, + 192 + ], + "type": "image", + "image_path": "311ad8e7a092496b911ba656a3b4c686079d96882aeceedeebe70fbe13718379.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 59, + 194, + 550, + 266 + ], + "blocks": [ + { + "bbox": [ + 59, + 194, + 550, + 266 + ], + "lines": [ + { + "bbox": [ + 59, + 194, + 550, + 266 + ], + "spans": [ + { + "bbox": [ + 59, + 194, + 550, + 266 + ], + "type": "image", + "image_path": "bdcbd1a965c81961da135a0fcf2765fee40118f781745bb22881f183f1d6201d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 279, + 553, + 301 + ], + "lines": [ + { + "bbox": [ + 55, + 279, + 553, + 301 + ], + "spans": [ + { + "bbox": [ + 55, + 279, + 553, + 301 + ], + "type": "text", + "content": "Figure 2. Overview of our proposed method. Our method predicts 2D Gaussians for point cloud rendering, employing an entire-patch architecture (bottom) and the 2D Gaussian Prediction Module (top left) with splitting decoders (top right)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 57, + 311, + 294, + 415 + ], + "blocks": [ + { + "bbox": [ + 57, + 311, + 294, + 415 + ], + "lines": [ + { + "bbox": [ + 57, + 311, + 294, + 415 + ], + "spans": [ + { + "bbox": [ + 57, + 311, + 294, + 415 + ], + "type": "image", + "image_path": "5677a16993197090d21fa6b6705109dc5505b68d8823bbf81cbd9e3c2d709160.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 422, + 295, + 475 + ], + "lines": [ + { + "bbox": [ + 55, + 422, + 295, + 475 + ], + "spans": [ + { + "bbox": [ + 55, + 422, + 295, + 475 + ], + "type": "text", + "content": "Figure 3. The illustration of our initialization approach, where the left side of each picture is a 2D schematic diagram and the right side is the rendered images of Gaussians. (a) The estimated normals of the point cloud. (b) Randomly Initialization. (c) Our Initialization. (d) Predicted 2D Gaussians." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 482, + 295, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 482, + 295, + 529 + ], + "spans": [ + { + "bbox": [ + 55, + 482, + 295, + 529 + ], + "type": "text", + "content": "PointMLP[19] as the encoder " + }, + { + "bbox": [ + 55, + 482, + 295, + 529 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 55, + 482, + 295, + 529 + ], + "type": "text", + "content": ", which is a simple residual MLP framework for point cloud learning with great performance on many tasks. The local feature " + }, + { + "bbox": [ + 55, + 482, + 295, + 529 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_l" + }, + { + "bbox": [ + 55, + 482, + 295, + 529 + ], + "type": "text", + "content": " is extracted by:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 123, + 529, + 227, + 543 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 123, + 529, + 227, + 543 + ], + "spans": [ + { + "bbox": [ + 123, + 529, + 227, + 543 + ], + "type": "interline_equation", + "content": "\\mathbf {F} _ {l} = E (\\mathbf {X} ^ {i}, \\mathbf {C} ^ {i}, \\mathbf {N} ^ {i}, \\mathbf {S} ^ {i}).", + "image_path": "70b47b8a9fc3d63c9daec59d8e9028b1ffb0ba66fd6d6808c11b9d01f02e5fb6.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 548, + 295, + 691 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 548, + 295, + 691 + ], + "spans": [ + { + "bbox": [ + 55, + 548, + 295, + 691 + ], + "type": "text", + "content": "Subsequently, " + }, + { + "bbox": [ + 55, + 548, + 295, + 691 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_l" + }, + { + "bbox": [ + 55, + 548, + 295, + 691 + ], + "type": "text", + "content": " is augmented with " + }, + { + "bbox": [ + 55, + 548, + 295, + 691 + ], + "type": "inline_equation", + "content": "\\mathbf{X}^i" + }, + { + "bbox": [ + 55, + 548, + 295, + 691 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 548, + 295, + 691 + ], + "type": "inline_equation", + "content": "\\mathbf{C}^i" + }, + { + "bbox": [ + 55, + 548, + 295, + 691 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 548, + 295, + 691 + ], + "type": "inline_equation", + "content": "\\mathbf{N}^i" + }, + { + "bbox": [ + 55, + 548, + 295, + 691 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 548, + 295, + 691 + ], + "type": "inline_equation", + "content": "\\mathbf{S}^i" + }, + { + "bbox": [ + 55, + 548, + 295, + 691 + ], + "type": "text", + "content": " as the input features for the splitting decoders. Our splitting decoders " + }, + { + "bbox": [ + 55, + 548, + 295, + 691 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 55, + 548, + 295, + 691 + ], + "type": "text", + "content": " also utilize " + }, + { + "bbox": [ + 55, + 548, + 295, + 691 + ], + "type": "inline_equation", + "content": "\\mathcal{G}^i" + }, + { + "bbox": [ + 55, + 548, + 295, + 691 + ], + "type": "text", + "content": " as the basis for prediction and forecast 2D Gaussians " + }, + { + "bbox": [ + 55, + 548, + 295, + 691 + ], + "type": "inline_equation", + "content": "\\mathcal{G}^p" + }, + { + "bbox": [ + 55, + 548, + 295, + 691 + ], + "type": "text", + "content": " for rendering purposes. There are six splitting decoders " + }, + { + "bbox": [ + 55, + 548, + 295, + 691 + ], + "type": "inline_equation", + "content": "D_x, D_s, D_c, D_n, D_\\alpha, D_o" + }, + { + "bbox": [ + 55, + 548, + 295, + 691 + ], + "type": "text", + "content": " responsible for predicting the six parameters " + }, + { + "bbox": [ + 55, + 548, + 295, + 691 + ], + "type": "inline_equation", + "content": "X^p, S^p, C^p, N^p, A^p, O^p" + }, + { + "bbox": [ + 55, + 548, + 295, + 691 + ], + "type": "text", + "content": " of 2D Gaussians. Each decoder contains a weight-shared MLP [5] as shown in Figure 2. The splitting and predicting process for the 6 parameters is basically the same, except for slight differences in shapes. Take the process of predicting the position " + }, + { + "bbox": [ + 55, + 548, + 295, + 691 + ], + "type": "inline_equation", + "content": "X^p" + }, + { + "bbox": [ + 55, + 548, + 295, + 691 + ], + "type": "text", + "content": " as an example. The decoder " + }, + { + "bbox": [ + 55, + 548, + 295, + 691 + ], + "type": "inline_equation", + "content": "D_x" + }, + { + "bbox": [ + 55, + 548, + 295, + 691 + ], + "type": "text", + "content": " predicts " + }, + { + "bbox": [ + 55, + 548, + 295, + 691 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 548, + 295, + 691 + ], + "type": "text", + "content": " shifts " + }, + { + "bbox": [ + 55, + 548, + 295, + 691 + ], + "type": "inline_equation", + "content": "\\Delta^x" + }, + { + "bbox": [ + 55, + 548, + 295, + 691 + ], + "type": "text", + "content": " based on the position " + }, + { + "bbox": [ + 55, + 548, + 295, + 691 + ], + "type": "inline_equation", + "content": "\\mathbf{X}^i" + }, + { + "bbox": [ + 55, + 548, + 295, + 691 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 84, + 700, + 266, + 714 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 84, + 700, + 266, + 714 + ], + "spans": [ + { + "bbox": [ + 84, + 700, + 266, + 714 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\Delta} _ {1} ^ {x}, \\boldsymbol {\\Delta} _ {2} ^ {x},..., \\boldsymbol {\\Delta} _ {K} ^ {x} = D _ {x} (\\mathbf {F} _ {l}, \\mathbf {X} ^ {i}, \\mathbf {C} ^ {i}, \\mathbf {N} ^ {i}, \\mathbf {S} ^ {i}),", + "image_path": "ecb96895282834b25d331749f839b1183e4961c970198c44c6f0e2bb114f86e3.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 312, + 553, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 312, + 553, + 361 + ], + "spans": [ + { + "bbox": [ + 313, + 312, + 553, + 361 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 312, + 553, + 361 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 312, + 553, + 361 + ], + "type": "text", + "content": " indicates the number of splits. The " + }, + { + "bbox": [ + 313, + 312, + 553, + 361 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 312, + 553, + 361 + ], + "type": "text", + "content": " shifts are added to the same " + }, + { + "bbox": [ + 313, + 312, + 553, + 361 + ], + "type": "inline_equation", + "content": "\\mathbf{X}^i" + }, + { + "bbox": [ + 313, + 312, + 553, + 361 + ], + "type": "text", + "content": " that is \"split\" the initial position " + }, + { + "bbox": [ + 313, + 312, + 553, + 361 + ], + "type": "inline_equation", + "content": "\\mathbf{X}^i" + }, + { + "bbox": [ + 313, + 312, + 553, + 361 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 312, + 553, + 361 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 312, + 553, + 361 + ], + "type": "text", + "content": " new positions. The predicted position " + }, + { + "bbox": [ + 313, + 312, + 553, + 361 + ], + "type": "inline_equation", + "content": "\\mathbf{X}^p\\in \\mathbb{R}_{K\\cdot N\\times 3}" + }, + { + "bbox": [ + 313, + 312, + 553, + 361 + ], + "type": "text", + "content": " is the combination of these " + }, + { + "bbox": [ + 313, + 312, + 553, + 361 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 312, + 553, + 361 + ], + "type": "text", + "content": " new positions:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 387, + 369, + 479, + 403 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 387, + 369, + 479, + 403 + ], + "spans": [ + { + "bbox": [ + 387, + 369, + 479, + 403 + ], + "type": "interline_equation", + "content": "\\mathbf {X} ^ {p} = \\bigcup_ {j = 1} ^ {K} \\left(\\boldsymbol {\\Delta} _ {j} ^ {x} + \\mathbf {X} ^ {i}\\right).", + "image_path": "ec4cbbd2acc3810518a6f70c925341b91aff65fc0b4b2505baf09c2889892a44.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 413, + 555, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 413, + 555, + 533 + ], + "spans": [ + { + "bbox": [ + 313, + 413, + 555, + 533 + ], + "type": "text", + "content": "After applying all the splitting decoders, we can obtain the parameters of " + }, + { + "bbox": [ + 313, + 413, + 555, + 533 + ], + "type": "inline_equation", + "content": "K \\cdot N" + }, + { + "bbox": [ + 313, + 413, + 555, + 533 + ], + "type": "text", + "content": " Gaussians " + }, + { + "bbox": [ + 313, + 413, + 555, + 533 + ], + "type": "inline_equation", + "content": "\\mathcal{G}^p" + }, + { + "bbox": [ + 313, + 413, + 555, + 533 + ], + "type": "text", + "content": ". With the splitting decoders, our method can manage sparse point clouds by augmenting the number of Gaussians, serving as a form of point cloud upsampling. A greater number of Gaussians can also capture more complex features, enhancing the details of rendered images. The predicted 2D Gaussians are based on normalized points, hence they need to be de-normalized to achieve accurate rendering results. The de-normalization is applied to the predicted position " + }, + { + "bbox": [ + 313, + 413, + 555, + 533 + ], + "type": "inline_equation", + "content": "X^p" + }, + { + "bbox": [ + 313, + 413, + 555, + 533 + ], + "type": "text", + "content": " and scale " + }, + { + "bbox": [ + 313, + 413, + 555, + 533 + ], + "type": "inline_equation", + "content": "S^p" + }, + { + "bbox": [ + 313, + 413, + 555, + 533 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 540, + 459, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 540, + 459, + 552 + ], + "spans": [ + { + "bbox": [ + 313, + 540, + 459, + 552 + ], + "type": "text", + "content": "3.2. Entire-Patch Architecture" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "text", + "content": "However, the 2D Gaussian prediction module can only be trained on complete point clouds, which leads to poor generalization across different categories with varying point distributions, as we only have ground truth images of entire objects for supervision. Numerous methods for other point cloud tasks, including surface reconstruction[3] and semantic segmentation[12], utilize point cloud patches as processing units to enhance the generalization capabilities of their networks. However, the corresponding ground truth images of point cloud patches are unavailable in practice, which makes training on patches impossible. To facilitate the application of our method to point cloud patches, we utilize an entire-patch architecture, as illustrated at the bottom of Fig-" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27288" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 384 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 384 + ], + "type": "text", + "content": "ure 2. Our approach encompasses two 2D Gaussian prediction modules, " + }, + { + "bbox": [ + 55, + 72, + 296, + 384 + ], + "type": "inline_equation", + "content": "\\mathcal{N}_e" + }, + { + "bbox": [ + 55, + 72, + 296, + 384 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 72, + 296, + 384 + ], + "type": "inline_equation", + "content": "\\mathcal{N}_p" + }, + { + "bbox": [ + 55, + 72, + 296, + 384 + ], + "type": "text", + "content": ", which operate on the entire point cloud and the point cloud patch, respectively. Given an entire point cloud " + }, + { + "bbox": [ + 55, + 72, + 296, + 384 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_e" + }, + { + "bbox": [ + 55, + 72, + 296, + 384 + ], + "type": "text", + "content": " for rendering, our method randomly selects a point as the center point of a point cloud patch. The point cloud patch " + }, + { + "bbox": [ + 55, + 72, + 296, + 384 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_p" + }, + { + "bbox": [ + 55, + 72, + 296, + 384 + ], + "type": "text", + "content": " is obtained by gathering " + }, + { + "bbox": [ + 55, + 72, + 296, + 384 + ], + "type": "inline_equation", + "content": "N_p" + }, + { + "bbox": [ + 55, + 72, + 296, + 384 + ], + "type": "text", + "content": " nearest points around the center point. The 2D Gaussians of point cloud " + }, + { + "bbox": [ + 55, + 72, + 296, + 384 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_e" + }, + { + "bbox": [ + 55, + 72, + 296, + 384 + ], + "type": "text", + "content": " and patch " + }, + { + "bbox": [ + 55, + 72, + 296, + 384 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_p" + }, + { + "bbox": [ + 55, + 72, + 296, + 384 + ], + "type": "text", + "content": " are predicted by two prediction modules, where " + }, + { + "bbox": [ + 55, + 72, + 296, + 384 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_e = \\mathcal{N}_e(\\mathcal{P}_e)" + }, + { + "bbox": [ + 55, + 72, + 296, + 384 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 72, + 296, + 384 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_p = \\mathcal{N}_p(\\mathcal{P}_p)" + }, + { + "bbox": [ + 55, + 72, + 296, + 384 + ], + "type": "text", + "content": ". The non-patch part " + }, + { + "bbox": [ + 55, + 72, + 296, + 384 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_e'" + }, + { + "bbox": [ + 55, + 72, + 296, + 384 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 55, + 72, + 296, + 384 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_e" + }, + { + "bbox": [ + 55, + 72, + 296, + 384 + ], + "type": "text", + "content": " are the 2D Gaussians corresponding to the non-patch points " + }, + { + "bbox": [ + 55, + 72, + 296, + 384 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_e' = \\mathcal{P}_e - \\mathcal{P}_p" + }, + { + "bbox": [ + 55, + 72, + 296, + 384 + ], + "type": "text", + "content": ". The final 2D Gaussians " + }, + { + "bbox": [ + 55, + 72, + 296, + 384 + ], + "type": "inline_equation", + "content": "\\mathcal{G} = \\mathcal{G}_e' \\cup \\mathcal{G}_p" + }, + { + "bbox": [ + 55, + 72, + 296, + 384 + ], + "type": "text", + "content": " are used to render images for supervision during training. Here, " + }, + { + "bbox": [ + 55, + 72, + 296, + 384 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_e'" + }, + { + "bbox": [ + 55, + 72, + 296, + 384 + ], + "type": "text", + "content": " serves as the \"background\" ensuring that the rendered image is complete for calculating the loss function and " + }, + { + "bbox": [ + 55, + 72, + 296, + 384 + ], + "type": "inline_equation", + "content": "\\mathcal{N}_p" + }, + { + "bbox": [ + 55, + 72, + 296, + 384 + ], + "type": "text", + "content": " can be properly trained. In practice, we initially train " + }, + { + "bbox": [ + 55, + 72, + 296, + 384 + ], + "type": "inline_equation", + "content": "\\mathcal{N}_e" + }, + { + "bbox": [ + 55, + 72, + 296, + 384 + ], + "type": "text", + "content": " separately and then freeze " + }, + { + "bbox": [ + 55, + 72, + 296, + 384 + ], + "type": "inline_equation", + "content": "\\mathcal{N}_e" + }, + { + "bbox": [ + 55, + 72, + 296, + 384 + ], + "type": "text", + "content": " while training " + }, + { + "bbox": [ + 55, + 72, + 296, + 384 + ], + "type": "inline_equation", + "content": "\\mathcal{N}_p" + }, + { + "bbox": [ + 55, + 72, + 296, + 384 + ], + "type": "text", + "content": " for our method. Both " + }, + { + "bbox": [ + 55, + 72, + 296, + 384 + ], + "type": "inline_equation", + "content": "\\mathcal{N}_e" + }, + { + "bbox": [ + 55, + 72, + 296, + 384 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 72, + 296, + 384 + ], + "type": "inline_equation", + "content": "\\mathcal{N}_p" + }, + { + "bbox": [ + 55, + 72, + 296, + 384 + ], + "type": "text", + "content": " can be utilized for inference. The inference of " + }, + { + "bbox": [ + 55, + 72, + 296, + 384 + ], + "type": "inline_equation", + "content": "\\mathcal{N}_e" + }, + { + "bbox": [ + 55, + 72, + 296, + 384 + ], + "type": "text", + "content": " directly takes " + }, + { + "bbox": [ + 55, + 72, + 296, + 384 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_e" + }, + { + "bbox": [ + 55, + 72, + 296, + 384 + ], + "type": "text", + "content": " as input and outputs the entire 2D Gaussians for rendering. For " + }, + { + "bbox": [ + 55, + 72, + 296, + 384 + ], + "type": "inline_equation", + "content": "\\mathcal{N}_p" + }, + { + "bbox": [ + 55, + 72, + 296, + 384 + ], + "type": "text", + "content": ", our method divides the point cloud into several patches. To ensure that all points of the input point cloud are covered, we repeatedly perform the process of randomly selecting a central point and extracting surrounding points as the patch point cloud, continuing until all points are included in at least one patch. The 2D Gaussian of the entire point cloud is the combination of the predictions from all patches." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 391, + 154, + 403 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 391, + 154, + 403 + ], + "spans": [ + { + "bbox": [ + 55, + 391, + 154, + 403 + ], + "type": "text", + "content": "3.3. Training Details" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 408, + 296, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 408, + 296, + 471 + ], + "spans": [ + { + "bbox": [ + 55, + 408, + 296, + 471 + ], + "type": "text", + "content": "Given " + }, + { + "bbox": [ + 55, + 408, + 296, + 471 + ], + "type": "inline_equation", + "content": "N_{c}" + }, + { + "bbox": [ + 55, + 408, + 296, + 471 + ], + "type": "text", + "content": " cameras " + }, + { + "bbox": [ + 55, + 408, + 296, + 471 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_1,\\ldots ,\\mathcal{C}_{N_c}" + }, + { + "bbox": [ + 55, + 408, + 296, + 471 + ], + "type": "text", + "content": ", the rendered images " + }, + { + "bbox": [ + 55, + 408, + 296, + 471 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_1^p,\\dots ,\\mathcal{I}_{N_c}^p" + }, + { + "bbox": [ + 55, + 408, + 296, + 471 + ], + "type": "text", + "content": " is obtained by " + }, + { + "bbox": [ + 55, + 408, + 296, + 471 + ], + "type": "inline_equation", + "content": "f_{R}(\\mathcal{G},\\mathcal{C}_{1}),\\ldots ,f_{R}(\\mathcal{G},\\mathcal{C}_{N_{c}})" + }, + { + "bbox": [ + 55, + 408, + 296, + 471 + ], + "type": "text", + "content": ". Similar to [15], we employ MSE and SSIM between rendered images and the corresponding ground truth images " + }, + { + "bbox": [ + 55, + 408, + 296, + 471 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_1^g,\\dots ,\\mathcal{I}_{N_c}^g" + }, + { + "bbox": [ + 55, + 408, + 296, + 471 + ], + "type": "text", + "content": " as loss functions for supervision:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 479, + 296, + 513 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 479, + 296, + 513 + ], + "spans": [ + { + "bbox": [ + 55, + 479, + 296, + 513 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\frac {1}{N _ {c}} \\sum_ {i = 1} ^ {N _ {c}} \\left(\\beta \\mathcal {L} _ {M S E} (\\mathcal {I} _ {i} ^ {p}, \\mathcal {I} _ {i} ^ {g}) + (1 - \\beta) \\mathcal {L} _ {S S I M} (\\mathcal {I} _ {i} ^ {p}, \\mathcal {I} _ {i} ^ {g})\\right).", + "image_path": "baf8a591c483efaeb043b234f213021ab1ab0cc0206394a08d45f4781e344368.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 521, + 296, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 296, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 296, + 628 + ], + "type": "text", + "content": "We implement our method with PyTorch [21]. The split number " + }, + { + "bbox": [ + 55, + 521, + 296, + 628 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 521, + 296, + 628 + ], + "type": "text", + "content": " is set to 4. The point number " + }, + { + "bbox": [ + 55, + 521, + 296, + 628 + ], + "type": "inline_equation", + "content": "N_{p}" + }, + { + "bbox": [ + 55, + 521, + 296, + 628 + ], + "type": "text", + "content": " of a patch is 2048. " + }, + { + "bbox": [ + 55, + 521, + 296, + 628 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 55, + 521, + 296, + 628 + ], + "type": "text", + "content": " is set to 0.8 for the loss function. The number of rendered images " + }, + { + "bbox": [ + 55, + 521, + 296, + 628 + ], + "type": "inline_equation", + "content": "N_{c}" + }, + { + "bbox": [ + 55, + 521, + 296, + 628 + ], + "type": "text", + "content": " is 8. The resolutions of images are 256x256, 640x512 and 512x512 for objects, scenes and human bodies. We employ Adam optimizer whose learning rate is " + }, + { + "bbox": [ + 55, + 521, + 296, + 628 + ], + "type": "inline_equation", + "content": "1.0 \\times 10^{-4}" + }, + { + "bbox": [ + 55, + 521, + 296, + 628 + ], + "type": "text", + "content": ". The batch size is set to 8, and the maximum epoch of training is 480. Training our method takes about 30 hours with a RTX 3090 GPU." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 639, + 137, + 653 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 639, + 137, + 653 + ], + "spans": [ + { + "bbox": [ + 55, + 639, + 137, + 653 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 659, + 116, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 659, + 116, + 672 + ], + "spans": [ + { + "bbox": [ + 55, + 659, + 116, + 672 + ], + "type": "text", + "content": "4.1. Settings" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": "Dataset. We compare our method with previous works across 5 datasets that include scenes, objects, and human bodies, encompassing the following datasets. ShapeNet [4]" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 315, + 69, + 554, + 155 + ], + "blocks": [ + { + "bbox": [ + 315, + 69, + 554, + 155 + ], + "lines": [ + { + "bbox": [ + 315, + 69, + 554, + 155 + ], + "spans": [ + { + "bbox": [ + 315, + 69, + 554, + 155 + ], + "type": "image", + "image_path": "f6c7fbe43e16219a6e30f2505829e02f287298609e8cba653ab4d046d12f84e5.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 165, + 555, + 198 + ], + "lines": [ + { + "bbox": [ + 313, + 165, + 555, + 198 + ], + "spans": [ + { + "bbox": [ + 313, + 165, + 555, + 198 + ], + "type": "text", + "content": "Figure 4. The evaluation of our method, TriVol and PFGS trained with different point numbers on Car category. The legend in the lower right corner indicates different methods." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 220, + 555, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 220, + 555, + 339 + ], + "spans": [ + { + "bbox": [ + 313, + 220, + 555, + 339 + ], + "type": "text", + "content": "is a large-scale dataset that contains CAD models across various categories. GSO [9] is a high-quality dataset comprising 3D scanned household items. ScanNet [7] is a dataset consisting of real-scanned indoor scenes. DTU [16] is a multi-view stereo dataset featuring high-quality and high-density scenes. THuman2.0 [30] is a dataset comprising high-quality 3D models of human bodies. We select the chairs and tables from ShapeNet and shoes from GSO for object-level comparison. All settings are consistent with those of TriVol [13] and PFGS [28]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 340, + 555, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 340, + 555, + 411 + ], + "spans": [ + { + "bbox": [ + 313, + 340, + 555, + 411 + ], + "type": "text", + "content": "Metrics. We evaluate the rendering results using three widely accepted metrics: PSNR, SSIM[29], and LIPIS[31]. Baselines. We employ traditional graphics rendering (GR), point-based novel view synthesis method NPBG++[23] and learning-based method TriVol[13] and PFGS[28] for comparison on different datasets." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 419, + 453, + 432 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 419, + 453, + 432 + ], + "spans": [ + { + "bbox": [ + 313, + 419, + 453, + 432 + ], + "type": "text", + "content": "4.2. Evaluation of Rendering" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 437, + 555, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 555, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 555, + 628 + ], + "type": "text", + "content": "The evaluation results of our method and baselines are presented in Table 2. The point numbers " + }, + { + "bbox": [ + 313, + 437, + 555, + 628 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 437, + 555, + 628 + ], + "type": "text", + "content": " of input point clouds are 100K for objects and scenes, and 80K for human bodies. The results demonstrate that our method outperforms nearly all other methods across various metrics and categories. Our approach outperforms existing methods, showing average improvements of " + }, + { + "bbox": [ + 313, + 437, + 555, + 628 + ], + "type": "inline_equation", + "content": "2.60\\%" + }, + { + "bbox": [ + 313, + 437, + 555, + 628 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 437, + 555, + 628 + ], + "type": "inline_equation", + "content": "1.05\\%" + }, + { + "bbox": [ + 313, + 437, + 555, + 628 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 437, + 555, + 628 + ], + "type": "inline_equation", + "content": "6.06\\%" + }, + { + "bbox": [ + 313, + 437, + 555, + 628 + ], + "type": "text", + "content": " in PSNR, SSIM, and LIPIs across all datasets, respectively. For qualitative comparison, we represent the rendering results of different methods across several datasets in Figure 5. Our rendering results not only exhibit superior overall quality, but also display crisp and detailed elements, such as patterns, wheels, and human faces. Quantitative and qualitative comparisons demonstrate that our method achieves SOTA performance in point cloud rendering." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 636, + 517, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 636, + 517, + 647 + ], + "spans": [ + { + "bbox": [ + 313, + 636, + 517, + 647 + ], + "type": "text", + "content": "4.3. Evaluation on Different Point Number" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "text", + "content": "As shown in Table 2, we also display the evaluation results of our method trained with different input point numbers, including 20K and 40K. Despite utilizing fewer points, our method still achieves the best performance on object categories such as chairs, tables, and shoes. Note that our" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27289" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 70, + 556, + 536 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 556, + 536 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 556, + 536 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 556, + 536 + ], + "type": "image", + "image_path": "74ae7f8c770fface135dc89a53a4068b99441dea058f7192ddb8562d56aa23a9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 544, + 555, + 567 + ], + "lines": [ + { + "bbox": [ + 55, + 544, + 555, + 567 + ], + "spans": [ + { + "bbox": [ + 55, + 544, + 555, + 567 + ], + "type": "text", + "content": "Figure 5. The rendering results of our method and previous methods on different categories. From top to bottom: scene, car, chair, shoe and human body." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 573, + 297, + 706 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 573, + 297, + 706 + ], + "spans": [ + { + "bbox": [ + 55, + 573, + 297, + 706 + ], + "type": "text", + "content": "method only utilizes " + }, + { + "bbox": [ + 55, + 573, + 297, + 706 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 55, + 573, + 297, + 706 + ], + "type": "text", + "content": " of the points as input, indicating that our method also yields good results for sparse point clouds. We also showcase the evaluation results for TriVol and PFGS when trained with varying point counts (2K, 10K, 20K, 40K, 100K) on the Car category to illustrate their performance with sparse point clouds. The comparisons are depicted in Figure 4. All methods exhibit improved performance as the number of points used during training increases, with our method outperforming all others across all input point numbers. The comparisons highlight the superiority of our method in rendering sparse point clouds." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 313, + 572, + 524, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 572, + 524, + 586 + ], + "spans": [ + { + "bbox": [ + 313, + 572, + 524, + 586 + ], + "type": "text", + "content": "4.4. Evaluation of Generalization Capability" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "text", + "content": "To evaluate the generalization capability of our " + }, + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{N}_p" + }, + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "text", + "content": " on different categories with different input point numbers, we directly test our network on DTU dataset, which is trained on the Car category using 20K points. The setting here is the same as that in PFGS [28]. The results are presented in Table 3, where the methods are assessed using 20K, 40K, 100K and 1M points on the DTU dataset. We also present the evaluation results of PFGS when trained on DTU with 1M points and on the Car category of ShapeNet with 20K points, in order to compare the generalization ca" + } + ] + } + ], + "index": 4 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "27290" + } + ] + } + ], + "index": 5 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 57, + 102, + 553, + 231 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "text", + "content": "Table 2. The comparison of our method against previous methods on different datasets including scenes, objects and human bodies. The top three results are highlighted in red, orange, and yellow, respectively. †:The input point number for THuman2.0 is 80K for all methods." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 102, + 553, + 231 + ], + "lines": [ + { + "bbox": [ + 57, + 102, + 553, + 231 + ], + "spans": [ + { + "bbox": [ + 57, + 102, + 553, + 231 + ], + "type": "table", + "html": "
MethodPoint NumberScanNet[7]Car (ShapeNet[4])Chair (ShapeNet[4])Shoe (GSO[9])THuman2.0[30]
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
GR100K†13.620.5280.77919.240.8140.18217.780.7790.20123.140.8290.15320.260.9050.337
NPBG++[23]100K16.810.6710.58525.320.8740.12025.780.9160.10129.420.9290.08126.810.9520.062
TriVol[13]100K18.560.7340.47327.220.9270.08428.850.9600.07831.240.9610.04525.970.9350.059
PFGS[28]100K19.860.7580.45227.340.9420.07727.520.9560.07829.530.9570.05834.740.9830.009
Ours20K18.570.7240.60027.880.9490.06828.890.9620.07730.910.9680.05133.700.9790.014
40K19.430.7430.55228.570.9570.06229.520.9670.07031.490.9740.04734.290.9810.012
100K20.240.7590.49028.730.9600.06029.100.9650.07532.080.9780.04235.430.9870.009
", + "image_path": "8979cf7d439c699d3f47a9804385ee9e7e7b6d1f8fde6808496f46266875c906.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 59, + 335, + 293, + 521 + ], + "blocks": [ + { + "bbox": [ + 55, + 248, + 297, + 326 + ], + "lines": [ + { + "bbox": [ + 55, + 248, + 297, + 326 + ], + "spans": [ + { + "bbox": [ + 55, + 248, + 297, + 326 + ], + "type": "text", + "content": "Table 3. The generalization capability of our method, traditional graphics rendering, and PFGS on the DTU dataset. Here, our method is trained on the Car category of ShapeNet using 20K points. Two models of PFGS are trained on the DTU dataset with 1M points and on the Car category of ShapeNet with 20K points, respectively. The first column records the input point numbers for evaluation. The top two results are highlighted in red and orange." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 59, + 335, + 293, + 521 + ], + "lines": [ + { + "bbox": [ + 59, + 335, + 293, + 521 + ], + "spans": [ + { + "bbox": [ + 59, + 335, + 293, + 521 + ], + "type": "table", + "html": "
Point NumberPFGS DTU 1M pointsPFGS Car (ShapeNet) 20K points
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
20K10.1770.3130.6478.0650.2970.792
40K12.0260.3580.5508.1600.3020.765
100K17.9780.5540.3128.6470.3220.673
1M23.0010.7700.11213.6760.5490.401
Point NumberGROurs \nCar (ShapeNet) 20K points
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
20K9.1050.3160.62516.7130.5900.361
40K9.9850.3280.64116.6230.5820.315
100K12.0350.3630.66616.5440.5730.263
1M18.0030.5940.39616.6430.5940.216
", + "image_path": "d31a0a04768613b72f445825f2b030e9312109bec421b9437a0d2af2693659a4.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 545, + 297, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 545, + 297, + 715 + ], + "spans": [ + { + "bbox": [ + 54, + 545, + 297, + 715 + ], + "type": "text", + "content": "pability of PFGS. Like previous methods, PFGS requires entire point clouds as input for predicting Gaussians, hence these methods struggle to generalize to other datasets or varying point densities. Being trained with point cloud patches, our method does not depend on global priors and possesses superior generalization capabilities, resulting in enhanced performance on other datasets and with varying point densities. Figure 6 shows the results of methods on DTU dataset with different point numbers, where our rendering results preserve clear details. Note that previous methods are unable to function on sparse point clouds or on scene categories that are significantly different from object categories. Both quantitative and qualitative evaluations reveal the exceptional generalization capability of our" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 316, + 247, + 553, + 392 + ], + "blocks": [ + { + "bbox": [ + 316, + 247, + 553, + 392 + ], + "lines": [ + { + "bbox": [ + 316, + 247, + 553, + 392 + ], + "spans": [ + { + "bbox": [ + 316, + 247, + 553, + 392 + ], + "type": "image", + "image_path": "45c2d67f60af888019e01bfedba89a09eb09af02f582eec2d279709365838d48.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 400, + 555, + 434 + ], + "lines": [ + { + "bbox": [ + 313, + 400, + 555, + 434 + ], + "spans": [ + { + "bbox": [ + 313, + 400, + 555, + 434 + ], + "type": "text", + "content": "Figure 6. The evaluation results of our method and previous methods on the DTU dataset, utilizing 20K, 40K, 100K and 1M points, with all methods trained on the Car category with 20K points." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 456, + 350, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 456, + 350, + 467 + ], + "spans": [ + { + "bbox": [ + 314, + 456, + 350, + 467 + ], + "type": "text", + "content": "method." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 476, + 462, + 489 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 476, + 462, + 489 + ], + "spans": [ + { + "bbox": [ + 313, + 476, + 462, + 489 + ], + "type": "text", + "content": "4.5. Visualization of Gaussians" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 495, + 555, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 495, + 555, + 604 + ], + "spans": [ + { + "bbox": [ + 313, + 495, + 555, + 604 + ], + "type": "text", + "content": "Figure 7(a)(b) illustrates the predicted Gaussians and corresponding rendered images of our method and PFGS[28]. As mentioned in Section 2.1, PFGS first predicts Gaussians and then refines the rendered images. We visualize the Gaussians of its first stage. Different from our predicted Gaussians, which are evenly distributed with fewer gaps, the predicted Gaussians of PFGS are chaos. Its rendered image is not photo-realistic requiring refinement in the second stage, and thus cannot be considered a final rendering result." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 611, + 484, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 611, + 484, + 624 + ], + "spans": [ + { + "bbox": [ + 313, + 611, + 484, + 624 + ], + "type": "text", + "content": "4.6. Evaluation of Rendering Speed" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 629, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 556, + 713 + ], + "type": "text", + "content": "Table 4 represents the rendering speeds of different methods. Our method directly predicts Gaussians for rendering images, eliminating the need for further refinement once the prediction is complete. The rendering speed of our method is the same as 2DGS [15], thus our method achieves real-time rendering with the fastest rendering speed among previous methods." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27291" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 70, + 294, + 245 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 294, + 245 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 294, + 245 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 294, + 245 + ], + "type": "image", + "image_path": "f68cc7b16a78b6c1c92221e4a68ad3b33b9f24d1d91ad4ac465b4c6c2a10bd98.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 259, + 296, + 316 + ], + "lines": [ + { + "bbox": [ + 55, + 259, + 296, + 316 + ], + "spans": [ + { + "bbox": [ + 55, + 259, + 296, + 316 + ], + "type": "text", + "content": "Figure 7. The visualization of Gaussians and their rendered images predicted by our method, our ablated methods and PFGS. (a): PFGS trained on Car (100K). (b) and (c): Our method trained on Car (100K and 20K). (d), (e) and (f): Our ablated methods trained on Car (20K)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 58, + 369, + 293, + 400 + ], + "blocks": [ + { + "bbox": [ + 55, + 327, + 296, + 361 + ], + "lines": [ + { + "bbox": [ + 55, + 327, + 296, + 361 + ], + "spans": [ + { + "bbox": [ + 55, + 327, + 296, + 361 + ], + "type": "text", + "content": "Table 4. The rendering speeds of different methods, where the rendering frames per second (FPS) are presented. All methods are evaluated using an RTX 3090." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 369, + 293, + 400 + ], + "lines": [ + { + "bbox": [ + 58, + 369, + 293, + 400 + ], + "spans": [ + { + "bbox": [ + 58, + 369, + 293, + 400 + ], + "type": "table", + "html": "
MethodNPBG++[23]TriVol[13]PFGS[28]Ours
Speed (FPS)37.451.623.80142.86
", + "image_path": "517cd0d228b1bee68fcc2501f3949d7aeb6b33bb414dbb2ba9974d8cecc1ff0b.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 420, + 149, + 433 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 420, + 149, + 433 + ], + "spans": [ + { + "bbox": [ + 55, + 420, + 149, + 433 + ], + "type": "text", + "content": "4.7. Ablation Study" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 437, + 296, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 437, + 296, + 521 + ], + "spans": [ + { + "bbox": [ + 55, + 437, + 296, + 521 + ], + "type": "text", + "content": "Split Number. We conduct experiments with different split numbers " + }, + { + "bbox": [ + 55, + 437, + 296, + 521 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 437, + 296, + 521 + ], + "type": "text", + "content": " (1, 2, 4, 8) and different input point cloud numbers " + }, + { + "bbox": [ + 55, + 437, + 296, + 521 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 437, + 296, + 521 + ], + "type": "text", + "content": " (2K, 10K, 20K, 40K) on the Car category of ShapeNet. The results are shown in Figure 8. Performance is enhanced with a growing number of splits, underscoring the efficacy of our split decoders, especially when it comes to sparse point clouds." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 522, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 522, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 522, + 296, + 714 + ], + "type": "text", + "content": "Component Ablation. We first evaluate the effectiveness of our initialization in the 2D Gaussian prediction module. The visualization of initialized Gaussians is shown in Figure 3 (c), which indicates a coarse rendering result providing a great foundation for optimization. Without initialization, the performance of our method sharply drops as Table 5 shows. Randomly initialized Gaussians might not be visible in certain views, resulting in non-convergence, as shown in Figure 7 (d). The ablation studies of our entire-patch architecture contain two experiments: (1) removing the entire module " + }, + { + "bbox": [ + 55, + 522, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{N}_e" + }, + { + "bbox": [ + 55, + 522, + 296, + 714 + ], + "type": "text", + "content": " and utilizing the predicted incomplete images rendered from point patches for supervision, denoted as “No " + }, + { + "bbox": [ + 55, + 522, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{N}_e" + }, + { + "bbox": [ + 55, + 522, + 296, + 714 + ], + "type": "text", + "content": "” in Table 5; (2) splitting the entire point cloud to multiple patches as same as the inference process of " + }, + { + "bbox": [ + 55, + 522, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{N}_p" + }, + { + "bbox": [ + 55, + 522, + 296, + 714 + ], + "type": "text", + "content": " to render complete images during training, denoted as “" + }, + { + "bbox": [ + 55, + 522, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{N}_p" + }, + { + "bbox": [ + 55, + 522, + 296, + 714 + ], + "type": "text", + "content": " only” in Table 5. the “" + }, + { + "bbox": [ + 55, + 522, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{N}_p" + }, + { + "bbox": [ + 55, + 522, + 296, + 714 + ], + "type": "text", + "content": " only” method will consume a sub" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 335, + 98, + 535, + 171 + ], + "blocks": [ + { + "bbox": [ + 313, + 71, + 555, + 93 + ], + "lines": [ + { + "bbox": [ + 313, + 71, + 555, + 93 + ], + "spans": [ + { + "bbox": [ + 313, + 71, + 555, + 93 + ], + "type": "text", + "content": "Table 5. The ablation studies of our method. The ablated methods are trained on Car with 20K points." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 335, + 98, + 535, + 171 + ], + "lines": [ + { + "bbox": [ + 335, + 98, + 535, + 171 + ], + "spans": [ + { + "bbox": [ + 335, + 98, + 535, + 171 + ], + "type": "table", + "html": "
PSNR ↑SSIM ↑LPIPS ↓
No Initialization9.620.6850.655
No Nc12.940.6760.441
Np only23.230.8290.126
Ours27.880.9490.068
", + "image_path": "b06d724f0d2fa884f7cf32187a6860ac3098355f003e9bdfa989ef9888191bd4.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 315, + 189, + 553, + 276 + ], + "blocks": [ + { + "bbox": [ + 315, + 189, + 553, + 276 + ], + "lines": [ + { + "bbox": [ + 315, + 189, + 553, + 276 + ], + "spans": [ + { + "bbox": [ + 315, + 189, + 553, + 276 + ], + "type": "image", + "image_path": "7bb36793d49020206ecd95505b0cb2dd2232f2e8ed64e162bbfbd6442670419f.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 281, + 555, + 316 + ], + "lines": [ + { + "bbox": [ + 313, + 281, + 555, + 316 + ], + "spans": [ + { + "bbox": [ + 313, + 281, + 555, + 316 + ], + "type": "text", + "content": "Figure 8. The ablation study of split number " + }, + { + "bbox": [ + 313, + 281, + 555, + 316 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 281, + 555, + 316 + ], + "type": "text", + "content": " with different input point numbers. The legend in the lower right corner indicates the input point number for training and evaluating." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 330, + 554, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 330, + 554, + 378 + ], + "spans": [ + { + "bbox": [ + 313, + 330, + 554, + 378 + ], + "type": "text", + "content": "stantial amount of resources for training and inference. The evaluation results of these two ablated methods are worse than our method, where their visualization results are shown in Figure 7(e)(f)(c)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 382, + 554, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 382, + 554, + 407 + ], + "spans": [ + { + "bbox": [ + 313, + 382, + 554, + 407 + ], + "type": "text", + "content": "The results of ablation studies validate the effectiveness and necessity of the components within our method." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 426, + 467, + 438 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 467, + 438 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 467, + 438 + ], + "type": "text", + "content": "5. Conclusion and Discussion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 450, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 450, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 450, + 555, + 713 + ], + "type": "text", + "content": "In this paper, we introduce a novel approach featuring an entire-patch architecture and the 2D Gaussian prediction module for point cloud rendering. Our prediction module directly forecasts the 2D Gaussians from the provided point clouds, beginning with initialization derived from normals. The splitting decoders produce multiple Gaussians from a point, thereby better accommodating sparse point clouds. To enhance the generalization capability, the entire-patch architecture employs a module for \"background\" area prediction and another module for point cloud patch prediction, ensuring proper supervision. Comprehensive experiments and comparisons demonstrate the superiority and generalization capability of our method, which can render point clouds into photo-realistic images with clear details and achieves SOTA performance. A limitation of our method may arise when a portion of a point cloud is absent, as our method cannot render a complete image without the direct support of the points. Our future work aims to address this limitation by incorporating point cloud completion techniques and also attempting to render point clouds with noise, such as those directly scanned from wild and outdoor scenes." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "27292" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 128, + 115, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 128, + 115, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 128, + 115, + 140 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 148, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 62, + 148, + 296, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 148, + 296, + 192 + ], + "spans": [ + { + "bbox": [ + 62, + 148, + 296, + 192 + ], + "type": "text", + "content": "[1] Kara-Ali Aliev, Artem Sevastopolsky, Maria Kolos, Dmitry Ulyanov, and Victor Lempitsky. Neural point-based graphics. In Computer Vision - ECCV 2020, pages 696-712, Cham, 2020. Springer International Publishing. 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 194, + 296, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 194, + 296, + 239 + ], + "spans": [ + { + "bbox": [ + 61, + 194, + 296, + 239 + ], + "type": "text", + "content": "[2] M. Botsch, A. Hornung, M. Zwicker, and L. Kobbelt. High-quality surface splatting on today's gpus. In Proceedings Eurographics/IEEE VGTC Symposium Point-Based Graphics, 2005., pages 17-141, 2005. 1, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 240, + 296, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 240, + 296, + 283 + ], + "spans": [ + { + "bbox": [ + 62, + 240, + 296, + 283 + ], + "type": "text", + "content": "[3] Alexandre Boulch and Renaud Marlet. Poco: Point convolution for surface reconstruction. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6292-6304, 2022. 2, 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 285, + 296, + 339 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 285, + 296, + 339 + ], + "spans": [ + { + "bbox": [ + 62, + 285, + 296, + 339 + ], + "type": "text", + "content": "[4] Angel X. Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, Jianxiong Xiao, Li Yi, and Fisher Yu. Shapenet: An information-rich 3d model repository, 2015. 5, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 342, + 296, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 342, + 296, + 385 + ], + "spans": [ + { + "bbox": [ + 62, + 342, + 296, + 385 + ], + "type": "text", + "content": "[5] Kaichun Mo Leonidas J. Guibas Charles R. Qi, Hao Su. Pointnet: Deep learning on point sets for 3d classification and segmentation. In 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2017. 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 388, + 296, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 388, + 296, + 441 + ], + "spans": [ + { + "bbox": [ + 62, + 388, + 296, + 441 + ], + "type": "text", + "content": "[6] Paolo Cignoni, Marco Callieri, Massimiliano Corsini, Matteo Dellepiane, Fabio Ganovelli, and Guido Ranzuglia. MeshLab: an Open-Source Mesh Processing Tool. In Eurographics Italian Chapter Conference. The Eurographics Association, 2008. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 444, + 296, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 444, + 296, + 498 + ], + "spans": [ + { + "bbox": [ + 62, + 444, + 296, + 498 + ], + "type": "text", + "content": "[7] Angela Dai, Angel X. Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proc. Computer Vision and Pattern Recognition (CVPR), IEEE, 2017. 5, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 500, + 296, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 500, + 296, + 555 + ], + "spans": [ + { + "bbox": [ + 62, + 500, + 296, + 555 + ], + "type": "text", + "content": "[8] Peng Dai, Yinda Zhang, Zhuwen Li, Shuaicheng Liu, and Bing Zeng. Neural point cloud rendering via multi-plane projection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7830-7839, 2020. 1, 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 557, + 296, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 557, + 296, + 622 + ], + "spans": [ + { + "bbox": [ + 62, + 557, + 296, + 622 + ], + "type": "text", + "content": "[9] Laura Downs, Anthony Francis, Nate Koenig, Brandon Kinman, Ryan Michael Hickman, Krista Reymann, Thomas Barlow McHugh, and Vincent Vanhoucke. Google scanned objects: A high-quality dataset of 3d scanned household items. 2022 International Conference on Robotics and Automation (ICRA), pages 2553-2560, 2022. 5, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 624, + 295, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 624, + 295, + 657 + ], + "spans": [ + { + "bbox": [ + 57, + 624, + 295, + 657 + ], + "type": "text", + "content": "[10] Antoine Guédon and Vincent Lepetit. Sugar: Surface-aligned gaussian splatting for efficient 3d mesh reconstruction and high-quality mesh rendering. CVPR, 2024. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 658, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 658, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 658, + 296, + 713 + ], + "type": "text", + "content": "[11] Hugues Hoppe, Tony DeRose, Tom Duchamp, John Alan McDonald, and Werner Stuetzle. Surface reconstruction from unorganized points. Proceedings of the 19th annual conference on Computer graphics and interactive techniques, 1992. 3" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 57, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 55, + 91, + 296, + 116 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 116 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 116 + ], + "type": "text", + "content": "This work was supported by the National Natural Science Foundation of China under Grant 62032011." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 316, + 73, + 555, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 555, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 555, + 128 + ], + "type": "text", + "content": "[12] Qingyong Hu, Bo Yang, Linhai Xie, Stefano Rosa, Yulan Guo, Zhihua Wang, Niki Trigoni, and Andrew Markham. Randla-net: Efficient semantic segmentation of large-scale point clouds. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2020. 2, 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 129, + 555, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 129, + 555, + 184 + ], + "spans": [ + { + "bbox": [ + 317, + 129, + 555, + 184 + ], + "type": "text", + "content": "[13] Tao Hu, Xiaogang Xu, Ruihang Chu, and Jiaya Jia. Trivol: Point cloud rendering via triple volumes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20732-20741, 2023. 1, 2, 3, 5, 7, 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 186, + 555, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 186, + 555, + 239 + ], + "spans": [ + { + "bbox": [ + 317, + 186, + 555, + 239 + ], + "type": "text", + "content": "[14] T. Hu, Xiaogang Xu, Shu Liu, and Jiaya Jia. Point2pix: Photo-realistic point cloud rendering via neural radiance fields. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8349-8358, 2023. 1, 2, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 240, + 554, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 240, + 554, + 285 + ], + "spans": [ + { + "bbox": [ + 316, + 240, + 554, + 285 + ], + "type": "text", + "content": "[15] Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2d gaussian splatting for geometrically accurate radiance fields. In SIGGRAPH 2024 Conference Papers. Association for Computing Machinery, 2024. 2, 3, 5, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 286, + 554, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 286, + 554, + 330 + ], + "spans": [ + { + "bbox": [ + 316, + 286, + 554, + 330 + ], + "type": "text", + "content": "[16] Rasmus Jensen, Anders Dahl, George Vogiatzis, Engil Tola, and Henrik Aanæs. Large scale multi-view stereopsis evaluation. In 2014 IEEE Conference on Computer Vision and Pattern Recognition, pages 406-413, 2014. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 332, + 554, + 375 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 332, + 554, + 375 + ], + "spans": [ + { + "bbox": [ + 316, + 332, + 554, + 375 + ], + "type": "text", + "content": "[17] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42 (4), 2023. 1, 2, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 376, + 554, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 376, + 554, + 409 + ], + "spans": [ + { + "bbox": [ + 316, + 376, + 554, + 409 + ], + "type": "text", + "content": "[18] Leif Kobbelt and Mario Botsch. A survey of point-based techniques in computer graphics. Comput. Graph., 28(6): 801-814, 2004. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 411, + 554, + 444 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 411, + 554, + 444 + ], + "spans": [ + { + "bbox": [ + 316, + 411, + 554, + 444 + ], + "type": "text", + "content": "[19] Xu Ma, Can Qin, Haoxuan You, Haoxi Ran, and Yun Fu. Rethinking network design and local geometry in point cloud: A simple residual mlp framework. *ICLR*, 2022. 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 445, + 554, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 445, + 554, + 488 + ], + "spans": [ + { + "bbox": [ + 316, + 445, + 554, + 488 + ], + "type": "text", + "content": "[20] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 1, 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 491, + 554, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 491, + 554, + 589 + ], + "spans": [ + { + "bbox": [ + 316, + 491, + 554, + 589 + ], + "type": "text", + "content": "[21] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Kopf, Edward Yang, Zachary DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. Pytorch: An imperative style, high-performance deep learning library. In Advances in Neural Information Processing Systems 32, pages 8024-8035. Curran Associates, Inc., 2019. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 590, + 554, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 590, + 554, + 634 + ], + "spans": [ + { + "bbox": [ + 316, + 590, + 554, + 634 + ], + "type": "text", + "content": "[22] Hanspeter Pfister, Matthias Zwicker, Jeroen van Baar, and Markus H. Gross. Surfels: surface elements as rendering primitives. Proceedings of the 27th annual conference on Computer graphics and interactive techniques, 2000. 1, 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 635, + 554, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 554, + 690 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 554, + 690 + ], + "type": "text", + "content": "[23] Ruslan Rakhimov, Andrei-Timotei Ardelean, Victor Lempitsky, and Evgeny Burnaev. Npbg++: Accelerating neural point-based graphics. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 15969-15979, 2022. 1, 2, 5, 7, 8" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 692, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 554, + 713 + ], + "type": "text", + "content": "[24] Nikhila Ravi, Jeremy Reizenstein, David Novotny, Taylor Gordon, Wan-Yen Lo, Justin Johnson, and Georgia" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27293" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 464 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "text", + "content": "Gkioxari. Accelerating 3d deep learning with pytorch3d. arXiv:2007.08501, 2020. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 140 + ], + "type": "text", + "content": "[25] Liu Ren, Hanspeter Pfister, and Matthias Zwicker. Object space ewa surface splatting: A hardware accelerated approach to high quality point rendering. Computer Graphics Forum, 21(3):461-470, 2002. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 141, + 294, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 141, + 294, + 174 + ], + "spans": [ + { + "bbox": [ + 56, + 141, + 294, + 174 + ], + "type": "text", + "content": "[26] Darius Rückert, Linus Franke, and Marc Stamminger. Adop: Approximate differentiable one-pixel point rendering. ACM Transactions on Graphics (ToG), 41(4):1-14, 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 175, + 294, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 294, + 218 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 294, + 218 + ], + "type": "text", + "content": "[27] Bane Sullivan and Alexander Kaszynski. PyVista: 3D plotting and mesh analysis through a streamlined interface for the Visualization Toolkit (VTK). Journal of Open Source Software, 4(37):1450, 2019. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 220, + 294, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 220, + 294, + 251 + ], + "spans": [ + { + "bbox": [ + 56, + 220, + 294, + 251 + ], + "type": "text", + "content": "[28] Jiaxu Wang, Ziyi Zhang, Junhao He, and Renjing Xu. Pfgs: High fidelity point cloud rendering via feature splatting. ECCV 2024, 2024. 1, 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 254, + 294, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 254, + 294, + 297 + ], + "spans": [ + { + "bbox": [ + 56, + 254, + 294, + 297 + ], + "type": "text", + "content": "[29] Zhou Wang, A.C. Bovik, H.R. Sheikh, and E.P. Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing, 13(4): 600-612, 2004. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 299, + 294, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 299, + 294, + 352 + ], + "spans": [ + { + "bbox": [ + 56, + 299, + 294, + 352 + ], + "type": "text", + "content": "[30] Tao Yu, Zerong Zheng, Kaiwen Guo, Pengpeng Liu, Qionghai Dai, and Yebin Liu. Function4d: Real-time human volumetric capture from very sparse consumer rgbd sensors. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR2021), 2021. 5, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 354, + 294, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 354, + 294, + 387 + ], + "spans": [ + { + "bbox": [ + 56, + 354, + 294, + 387 + ], + "type": "text", + "content": "[31] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 388, + 294, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 388, + 294, + 420 + ], + "spans": [ + { + "bbox": [ + 56, + 388, + 294, + 420 + ], + "type": "text", + "content": "[32] Qian-Yi Zhou, Jaesik Park, and Vladlen Koltun. Open3D: A modern library for 3D data processing. arXiv:1801.09847, 2018. 2, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 422, + 294, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 422, + 294, + 464 + ], + "spans": [ + { + "bbox": [ + 56, + 422, + 294, + 464 + ], + "type": "text", + "content": "[33] Matthias Zwicker, Hanspeter Pfister, Jeroen Baar, and Markus Gross. Surface splattering. Proceedings of the ACM SIGGRAPH Conference on Computer Graphics, 2001, 2001. 2" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27294" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Sparse Voxels Rasterization_ Real-time High-fidelity Radiance Field Rendering/a6bd5a3d-4547-4194-a227-b9c1f40e864c_content_list.json b/2025/Sparse Voxels Rasterization_ Real-time High-fidelity Radiance Field Rendering/a6bd5a3d-4547-4194-a227-b9c1f40e864c_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..1ffe09798f6f566540730b25da3a52f27d3f0e2a --- /dev/null +++ b/2025/Sparse Voxels Rasterization_ Real-time High-fidelity Radiance Field Rendering/a6bd5a3d-4547-4194-a227-b9c1f40e864c_content_list.json @@ -0,0 +1,2031 @@ +[ + { + "type": "text", + "text": "Sparse Voxels Rasterization: Real-time High-fidelity Radiance Field Rendering", + "text_level": 1, + "bbox": [ + 99, + 130, + 897, + 154 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Cheng Sun1", + "bbox": [ + 104, + 179, + 202, + 199 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jaesung Choe1 \n1NVIDIA", + "bbox": [ + 233, + 180, + 354, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Charles Loop1 \n2Cornell University", + "bbox": [ + 382, + 180, + 501, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Wei-Chiu Ma $^{2}$ $^{3}$ National Taiwan University", + "bbox": [ + 519, + 180, + 692, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yu-Chiang Frank Wang $^{1,3}$ $^{1}$", + "bbox": [ + 684, + 180, + 888, + 215 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d0938896d4565a5e05a7fe016e57bbf7bc58a98e6675b90538972db605606e47.jpg", + "image_caption": [ + "(a) Volume rendering by rasterizing sparse voxels." + ], + "image_footnote": [], + "bbox": [ + 94, + 224, + 385, + 402 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6d7b51f99847323ed14a071228798e98352c46ddac178ac7489b5a7a73cb0130.jpg", + "image_caption": [ + "(b) Novel-view rendering on Mip-NeRF360 scenes.", + "Figure 1. We propose SVRaster, a novel framework for multi-view reconstruction and novel view synthesis. (a) Sparse voxel representation effectively captures the volume density and radiance field of the scene, without the need for neural networks, 3D Gaussians, and sparse-points prior. (b) Using our customized sparse voxel rasterizer, we can learn the underlying 3D scene efficiently and achieve state-of-the-art performance in both rendering quality and speed. (c) Notably, lifting 2D modal to the trained sparse voxels is simple and efficient by integrating the classic Volume Fusion [7, 8, 34]. We show examples of vision foundation model feature field from RADIO [38], semantic field from Segformer [52], and signed distance field from rendered depth, making it flexible and suitable for a wide range of applications." + ], + "image_footnote": [], + "bbox": [ + 385, + 226, + 679, + 417 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a75af022143a16049ee7fb478ceb92d03cb17d8d73599b3744176bdd2326eb2c.jpg", + "image_caption": [ + "(c) 2D-to-3D made easy." + ], + "image_footnote": [], + "bbox": [ + 686, + 224, + 900, + 415 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 546, + 326, + 561 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We propose an efficient radiance field rendering algorithm that incorporates a rasterization process on adaptive sparse voxels without neural networks or 3D Gaussians. There are two key contributions coupled with the proposed system. The first is to adaptively and explicitly allocate sparse voxels to different levels of detail within scenes, faithfully reproducing scene details with $65536^{3}$ grid resolution while achieving high rendering frame rates. Second, we customize a rasterizer for efficient adaptive sparse voxels rendering. We render voxels in the correct depth order by using ray direction-dependent Morton ordering, which avoids the well-known popping artifact found in Gaussian splatting. Our method improves the previous neural-free voxel model by over 4db PSNR and more than 10x FPS speedup, achieving state-of-the-art comparable novel-view synthesis results. Additionally, our voxel representation is seamlessly compatible with grid-based 3D processing techniques such as Volume Fusion, Voxel Pooling, and Marching Cubes, enabling a wide range of future extensions and applications.", + "bbox": [ + 89, + 580, + 483, + 868 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Code: github.com/NVlabs/svraster", + "bbox": [ + 89, + 868, + 323, + 882 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 514, + 546, + 643, + 561 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Gaussian splatting [20] has emerged as one of the most promising solutions for novel view synthesis. It has drawn wide attention across multiple communities due to its exceptional rendering speed and ability to capture the nuanced details of a scene. Nevertheless, Gaussian splatting in its base form has two key limitations: first, sorting Gaussians based on their centers does not guarantee proper depth ordering. It may result in popping artifacts [37] (i.e., sudden color changes for consistent geometry) when changing views. Second, the volume density of a 3D point is ill-defined when covered by multiple Gaussians. This ambiguity makes surface reconstruction non-trivial.", + "bbox": [ + 509, + 575, + 906, + 756 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "On the other hand, grid representations inherently avoid these issues—both ordering and volume are well-defined. However, due to ray casting, the rendering speed of these methods [4, 5, 24, 43, 44, 51] is slower than that of Gaussian Splatting. This raises the question: is it possible to take the best of both worlds? Can we combine the efficiency of Gaussian splatting with the well-defined grid properties?", + "bbox": [ + 511, + 760, + 908, + 867 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Fortunately, the answer is yes. Our key insight is to revisit voxels—a well-established primitive with decades of", + "bbox": [ + 511, + 869, + 908, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "16187", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "history. Voxel representations are inherently compatible with modern graphics engines and can be rasterized efficiently. Additionally, previous work has demonstrated their ability to model scene volume densities, despite through volume ray casting. This positions voxels as the perfect bridge between rasterization and volumetric representation. However, naively adopting voxels does not work well [11]. Since a scene may consist of different levels of detail.", + "bbox": [ + 89, + 90, + 480, + 210 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "With these observations in mind, we present SVRaster, a novel framework that combines the efficiency of rasterization in 3DGS with the structured volumetric approach of grid-based representations. SVRaster leverages (1) multilevel sparse voxels to model 3D scenes and (2) implements a direction-dependent Morton order encoding that facilitates the rasterization rendering from our adaptive-sized sparse voxel representation. Our rendering is free from popping artifacts because the 3D space is partitioned into disjoint voxels, and our sorting ensures the correct rendering order. Moreover, thanks to the volumetric nature and neural-free representation of SVRaster, our sparse voxel grid can be easily and seamlessly integrated with classical grid-based 3D processing algorithms.", + "bbox": [ + 89, + 212, + 482, + 422 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We show that our method is training fast, rendering fast, and achieves novel-view synthesis quality comparable to the state-of-the-art. We also integrate Volume Fusion, Voxel Pooling, and Marching Cubes operations into our adaptive sparse voxels, which showcases promising results of mesh extraction and 2D foundation feature fusion.", + "bbox": [ + 89, + 422, + 482, + 513 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 527, + 232, + 544 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Differentiable volume rendering has made significant strides in 3D scene reconstruction and novel-view synthesis tasks. Neural Radiance Fields (NeRF) [30] laid the foundation for this progress by optimizing a volumetric function, parameterized by multi-layer perceptrons (MLPs) to encode both geometry and appearance through differential rendering. Subsequent studies focus on accelerating speed by decomposing large MLPs into grid-based representations, with a shallow MLP typically still being employed. Several grid representations have been explored—dense grids [43], factorized grids [5], tri-planes [4], and hash grids [32].", + "bbox": [ + 89, + 553, + 482, + 718 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In particular, our method is closely related to sparse voxel grid representations. Sparse Voxel Octrees [22] and VDB trees [33] are commonly used to manage sparse voxels and facilitate rendering. In contrast, our sparse voxels are stored in a 1D array without advanced data structures. Our rasterizer with the proposed direction-dependent Morton order ensures correct rendering order. To model volumes in sparse leaf nodes, different strategies exist, such as low-resolution 3D grid [33, 50] or implicit neural field [25, 45]. Our leaf node is a single voxel with explicit density and color parameters, like Plenoxels [11]. Regarding voxel levels, previous methods allocate voxels to a tar", + "bbox": [ + 89, + 720, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "get level [11, 25, 45, 55] or use a shallow tree [33, 50]. Our voxels can adaptively fit into different levels across the entire tree depth, maximizing flexibility and scalability.", + "bbox": [ + 511, + 90, + 903, + 136 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Another scalability issue of the previous grid-based methods is that they still use some sort of dense 3D grid in the field of novel-view synthesis. For instance, dense occupancy grids [5, 23, 32, 43] for free-space skipping or dense pointer grid [11] to support sparse voxel lookup. We do not use any 3D dense data structures.", + "bbox": [ + 511, + 138, + 903, + 228 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3D Gaussian Splatting (3DGS) [20] takes a different approach by representing scenes with 3D Gaussian primitives and using rasterization for rendering, achieving state-of-the-art trade-off for quality and speed. Our work is inspired by the efficiency of using a rasterizer by 3DGS. However, 3DGS exhibits view-inconsistent popping artifacts due to inaccurate rendering order and primitive overlapping. Some recent works mitigate this artifact [31, 37], but completely resolving it significantly harms rendering speed [28]. Our method does not suffer from the popping artifact. Typically, 3DGS [20] initializes Gaussians from the triangulated points via COLMAP [41], which is especially critical for unbounded scenes. The sparse points geometry is also employed to guide the training of volumetric representations and has shown improvement in some NeRF-based methods [9, 12]. In this work, we do not use the sparse points.", + "bbox": [ + 511, + 229, + 903, + 472 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Mesh reconstruction is an important extended topic for both volumetric-based and GS-based methods. Recent 3DGS variants [14, 16] with proper regularization have shown a good speed-accuracy trade-off. However, directly extracting meshes from 3DGS remains challenging, so these methods still rely on volumetric-based post-processing algorithms to extract meshes. NeRF variants [24, 44, 47, 48, 51, 53] improve surface quality by rendering a signed distance field (SDF) instead of density field. The well-defined volume makes it straightforward to extract a mesh from the isosurface. Our sparse voxels are also a volumetric representation. Though our voxel also follows volumetric representation focusing on the density field, we still can achieve promising accuracy with minimal time.", + "bbox": [ + 511, + 474, + 903, + 685 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "For efficient rendering, another line of research is to convert (or distill) the geometry and appearance of a high-quality but slow model into other efficient kind of representations such as smaller MLPs [39, 46], meshes [40, 49, 54], grids [10], or the recent Gaussians [35]. Typically, the powerful Zip-NeRF [3] is employed as the teacher model which requires hours of training per scene. Our volumetric representation could enable training-free conversion from a large implicit field, which is however out of our current scope.", + "bbox": [ + 511, + 686, + 903, + 823 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Finally, the sparse voxel is also a widely-used representation for 3D processing [7, 8, 19, 26, 34, 36] and understanding [6, 13, 50]. We mainly focus on rendering in this work while we believe that adapting our method to these techniques is a promising future direction.", + "bbox": [ + 511, + 825, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "16188", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/c555599a72e63f8845ff350e9ad912262f6383be220dfe60687eb61aba27006d.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 93, + 88, + 282, + 233 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/00d9280cf0e4d38f6e0fd48a2ca55da6578dc7b213db74d1c52d927959d0bf5c.jpg", + "image_caption": [ + "(b)", + "Figure 2. Sparse voxels scene representation. (Left) We allocate voxel under an Octree layout. Each voxels has its own Spherical Harmonic coefficient for view-dependent appearance. The color field is approximated as a constant inside a voxel when rendering a view for efficiency. The density field is trilaterally varied inside a voxel and is modeled by the density values on the corner grid points (i.e., the black dots $\\bullet$ ) of each voxel. The grid points densities are shared between adjacent voxels. (Right) We evenly sample $K$ points inside the segment of ray-voxel intersection to compute volume integration for its alpha value contributing to the pixel ray. See Sec. 3.1.1 for details." + ], + "image_footnote": [], + "bbox": [ + 282, + 87, + 478, + 234 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Approach", + "text_level": 1, + "bbox": [ + 89, + 429, + 199, + 446 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We present our approach as follows. First in Sec. 3.1, we introduce our sparse voxels scene representation and our rasterizer for rendering sparse voxels into pixels. Later in Sec. 3.2, we describe the progressive scene optimization strategy, which is designed for faithfully reconstructing the scene from multi-view images using our sparse voxels.", + "bbox": [ + 89, + 454, + 483, + 546 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Sparse Voxels Rasterization", + "text_level": 1, + "bbox": [ + 89, + 556, + 341, + 571 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Recent neural radiance field rendering approaches, such as NeRF [30] and 3DGS [20] variants, use the following alpha composition equation to render a pixel's color C:", + "bbox": [ + 89, + 579, + 483, + 625 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {C} = \\sum_ {i = 1} ^ {N} T _ {i} \\cdot \\alpha_ {i} \\cdot c _ {i}, \\quad T _ {i} = \\prod_ {j = 1} ^ {i - 1} \\left(1 - \\alpha_ {j}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 109, + 627, + 482, + 656 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\alpha_{i}\\in \\mathbb{R}_{\\in [0,1]}$ and $c_{i}\\in \\mathbb{R}_{\\in [0,1]}^{3}$ are alpha and view-dependent color at the $i$ -th sampled point or primitive on the pixel-ray. The quantity $T_{i}$ is known as the transmittance. At a high-level, the difference between each method comes down to: $i$ ) how they find the $N$ points or primitives to composite a pixel-ray and $ii$ ) how the rgb and alpha values are determined from their scene representations.", + "bbox": [ + 89, + 660, + 482, + 765 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our sparse-voxel method follows the same principle. In Sec. 3.1.1, we provide details of our sparse voxels scene representation and how the $\\alpha_{i}$ and $c_{i}$ in Eq. (1) are computed from a voxel. In Sec. 3.1.2, we present our rasterizer, that gathers the $N$ voxels to be composited for each pixel.", + "bbox": [ + 89, + 767, + 483, + 843 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1.1. Scene Representation", + "text_level": 1, + "bbox": [ + 89, + 851, + 287, + 866 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We first describe the grid layout for allocating our sparse voxels and then derive the alpha value, view-dependent", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "color, and other geometric properties needed for the composite rendering of a voxel.", + "bbox": [ + 511, + 90, + 903, + 121 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Sparse voxel grid. Our SVRaster constructs 3D scenes using a sparse voxel representation. We allocate voxels following an Octree space partition rule (i.e., Octree layout as illustrated in Fig. 2a) for two reasons necessary for achieving high-quality results. First, it facilitates the correct rendering order of voxels with various sizes (Sec. 3.1.2). Second, we can adaptively fit the sparse voxels to different scene level-of-details (Sec. 3.2). Note that our representation does not replicate a traditional Octree data structure with parent-child pointers or linear Octree. Specifically, we only keep voxels at the Octree leaf nodes without any ancestor nodes. Our sorting-based rasterizer will project voxels to image space and guarantee all voxels are in the correct order when rendering. In sum, we store individual voxels in arbitrary order without the need to maintain a more complex data structure, thanks to the flexibility provided by our rasterizer.", + "bbox": [ + 511, + 130, + 906, + 385 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We choose a maximum level of detail $L$ ( $= 16$ in this work) that defines a maximum grid resolution at $(2^L)^3$ . Let $\\mathbf{w}_{\\mathrm{s}} \\in \\mathbb{R}$ be the Octree size and $\\mathbf{w}_{\\mathrm{c}} \\in \\mathbb{R}^3$ be the Octree center in the world space. The voxel index $v = \\{i,j,k\\} \\in [0,\\dots,2^{L} - 1]^{3}$ together with a Octree level $l \\in [1,L]$ ( $l = 0$ represent root node and is not used) define voxel size $\\mathbf{v}_{\\mathrm{s}}$ and voxel center $\\mathbf{v}_{\\mathrm{c}}$ as:", + "bbox": [ + 511, + 388, + 905, + 494 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {v} _ {\\mathrm {s}} = \\mathbf {w} _ {\\mathrm {s}} \\cdot 2 ^ {- l}, \\quad \\mathbf {v} _ {\\mathrm {c}} = \\mathbf {w} _ {\\mathrm {c}} - 0. 5 \\cdot \\mathbf {w} _ {\\mathrm {s}} + \\mathbf {v} _ {\\mathrm {s}} \\cdot v. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 535, + 500, + 903, + 516 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Internally, we map the grid index to its Morton code using a well-known bit interleaving operation in the low-level CUDA implementation. Please see supplementary materials for more details.", + "bbox": [ + 511, + 521, + 905, + 580 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Voxel alpha from density field. Next, we present details for the geometry and appearance modeling of each voxel primitive. For scene geometry, we use eight parameters corresponding to the voxel corners to model a trilinear density field inside each voxel, denoted as $\\mathbf{v}_{\\mathrm{geo}} \\in \\mathbb{R}^{2 \\times 2 \\times 2}$ . Sharing corners among adjacent voxels results in a continuous density field.", + "bbox": [ + 511, + 590, + 905, + 696 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We also need an activation function to ensure a nonnegative density value for the raw density from $\\mathbf{v}_{\\mathrm{geo}}$ . For this purpose, we use exponential-linear activation:", + "bbox": [ + 511, + 699, + 905, + 744 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {e x p l i n} (x) = \\left\\{ \\begin{array}{l l} x & \\text {i f} x > 1. 1 \\\\ \\exp \\left(\\frac {x}{1 . 1} - 1 + \\ln 1. 1\\right) & \\text {o t h e r w i s e} \\end{array} , \\right. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 748, + 903, + 790 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "which approximates softplus but is more efficient to compute. For a sharp density field inside a voxel, we apply the non-linear activation after trilinear interpolation [18, 43].", + "bbox": [ + 511, + 792, + 903, + 837 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To derive the alpha value of a voxel contributing to the alpha composition formulation in Eq. (1), we evenly sample $K$ points in the ray segment of ray-voxel intersection as depicted in Fig. 2b. The equation follows the numerical", + "bbox": [ + 511, + 839, + 905, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "16189", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/e8373eed911a159dc8296b70ba2ae245e42822207ef5b1bb96d018aa9efd4d19.jpg", + "image_caption": [ + "Figure 3. Rasterization procedure. Refer to Sec. 3.1.2 for details." + ], + "image_footnote": [], + "bbox": [ + 93, + 88, + 480, + 157 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "integration for volume rendering as in NeRF [29, 30]:", + "bbox": [ + 89, + 215, + 447, + 231 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\alpha = 1 - \\exp \\left(- \\frac {l}{K} \\sum_ {k = 1} ^ {K} \\operatorname {e x p l i n} \\left(\\operatorname {i n t e r p} \\left(\\mathbf {v} _ {\\text {g e o}}, \\mathbf {q} _ {k}\\right)\\right)\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 96, + 234, + 482, + 276 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $l$ is the ray segment length, $\\mathbf{q}_k$ is the local voxel coordinate of the $k$ -th sample point, and $\\mathrm{interp}(\\cdot)$ indicates trilinear interpolation.", + "bbox": [ + 89, + 280, + 482, + 325 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Voxel view-dependent color from SHs. To model view-dependent scene appearance, we use $N_{\\mathrm{shd}}$ degree SH. For increased efficiency, we assume the SH coefficients stay constant inside a voxel, denoted as $\\mathbf{v}_{\\mathrm{sh}} \\in \\mathbb{R}^{(N_{\\mathrm{shd}} + 1)^2 \\times 3}$ . We approximate voxel colors as a function of the direction from the camera position $\\mathbf{r}_0$ to the voxel center $\\mathbf{v}_{\\mathrm{c}}$ instead of individual ray direction $\\mathbf{r}_{\\mathrm{d}}$ for the sake of efficiency following 3DGS:", + "bbox": [ + 89, + 338, + 483, + 460 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nc = \\max (0, \\operatorname {s h} _ {-} \\operatorname {e v a l} (\\mathbf {v} _ {\\mathrm {s h}}, \\text {n o r m a l i z e} (\\mathbf {v} _ {\\mathrm {c}} - \\mathbf {r} _ {\\mathrm {o}}))), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 112, + 465, + 482, + 482 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "which is the view-dependent color intensity of the voxel contributing to the pixel composition Eq. (1). Due to the approximation, the resulting SH color of a voxel can be shared by all covered pixels in the image rather than evaluating the SH for each intersecting ray.", + "bbox": [ + 89, + 486, + 483, + 561 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Voxel normal. The rendering of other features or properties is similar to rendering a color image by replacing the color term $c$ in Eq. (1) with the target modality like the normal of a voxel density field. For rendering efficiency, we assume the normal stays constant inside a voxel, which is represented by the analytical gradient of the density field at the voxel center:", + "bbox": [ + 89, + 574, + 483, + 679 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\vec {\\mathbf {n}} = \\text {n o r m a l i z e} \\left(\\nabla_ {\\mathbf {q}} \\operatorname {i n t e r p} \\left(\\mathbf {v} _ {\\mathrm {g e o}}, \\mathbf {q} _ {\\mathrm {c}}\\right)\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 150, + 685, + 482, + 703 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{q}_{\\mathrm{c}} = (0.5, 0.5, 0.5)$ and the closed-form equations for forward and backward passes are in the supplementary material. Similar to the SH colors, the differentiable voxel normals are computed once in pre-processing and are shared by all the intersecting rays in the image.", + "bbox": [ + 89, + 705, + 483, + 782 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Voxel depth. Unlike colors and normals, the point depths to composite is efficient to compute so we do the same $K$ points sampling as in the voxel alpha value in Eq. (4) for a more precise depth rendering. We manually expand and simplify the forward and backward computation for small number of $K$ in our CUDA implementation. Please refer to supplementary materials for details.", + "bbox": [ + 89, + 794, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/2f1341aea32eb6c2d516926feb1d92f0e0388fb88b66f5c291f65973c22bfb46.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 516, + 89, + 584, + 143 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/6de3319d7697aa716df85624bd8d2fdc2cf5a1bdea4db5821848e2bd7dc11632.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 516, + 143, + 584, + 208 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/9eccfdd93cd764748253f609b75e6d7abeca7798ced325c90a2de9445dd45b15.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 594, + 92, + 656, + 143 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/e243c7d69faf023ea5177c92999eaac74d2e0e9df41eb1d7b6ab0972582ecaaa.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 596, + 145, + 656, + 205 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/5de485304965ce85f7df2cbbe759db978491b2216adb77fe66d611bd27bf9007.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 658, + 93, + 743, + 143 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/1058bcc0c9c7bd80855967ef1de2ccc3e62c7a8b1e920cb607ab5c7b7d735964.jpg", + "image_caption": [ + "Figure 4. Illustration of the rendering order. (a) In both cases, the smaller voxels should be rendered first but they will arranged behind the larger voxels if using voxel centers or the nearest corners as the sorting order. (b) We show the four types of Morton order under the 2D world. The voxel rendering order under an Octree node is depend on which world quadrant the ray direction is pointing to. (c) An toy example of sorting the Morton order encoding. All the ray directions going toward the up-right quadrant can use the sorted voxels for a correct rendering order. See Sec. 3.1.2 for details." + ], + "image_footnote": [], + "bbox": [ + 658, + 145, + 743, + 205 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/2583482f33309c08029722bffad55bc8cd414c4476a363bdae81a4358d68e317.jpg", + "image_caption": [ + "(c)" + ], + "image_footnote": [], + "bbox": [ + 756, + 94, + 903, + 208 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1.2. Rasterization Algorithm", + "text_level": 1, + "bbox": [ + 511, + 395, + 728, + 410 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The overview of our sparse voxel rasterization procedure is depicted in Fig. 3. We build our sorting-based rasterizer based on the highly efficient CUDA implementation of 3DGS [20]. The procedure is detailed in the following.", + "bbox": [ + 511, + 414, + 906, + 474 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Projection to image space. The first step of rasterization is projecting sparse voxels onto image space and assigning the voxels to the tiles (i.e., image patches) they cover. In practice, we project the eight corner points of each voxel. The voxel is assigned to all tiles overlapped with the axis-aligned bounding box formed by the projected eight points.", + "bbox": [ + 511, + 479, + 905, + 571 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Pre-processing voxels. For active voxels assigned to tiles of the target view, we gather their densities $\\mathbf{v}_{\\mathrm{geo}}$ from the grid points, compute view-dependent colors from their spherical harmonic coefficients with Eq. (5), and derive voxel normals by Eq. (6). The pre-processed voxel properties are shared among all pixels during rendering.", + "bbox": [ + 511, + 577, + 905, + 667 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Sorting voxels. For accurate rasterization, primitive rendering order is important. Similar to the challenge in 3DGS [20], using primitive centers or their closest distance to the camera can produce incorrect ordering, producing popping artifacts [37]. We show two incorrect ordering results using naive sorting criteria in Fig. 4a. Thanks to the Octree layout (Sec. 3.1.1), we can sort by Morton order using our sparse voxels representation. As illustrated in Fig. 4b, we can follow certain types of Morton order to render the voxels under an Octree node for correct ordering. The type of Morton order to follow is solely dependent on the positive/negative signs of the ray direction (the ray origin doesn't matter). That is to say, we have eight permutations of Morton order for different ray directions in the 3D space. Finally, the generalization to multi-level voxels can", + "bbox": [ + 511, + 674, + 906, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "16190", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "be proved by induction. An ordering example in 2D with three levels is depicted in Fig. 4c.", + "bbox": [ + 89, + 90, + 482, + 119 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The sorting is applied for each image tile. In case all the pixels in a tile share the same ray direction signs, we can simply sort the assigned voxels by their type of Morton order. We handle the corner case when multiple Morton orders are required in supplementary materials.", + "bbox": [ + 89, + 121, + 483, + 196 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Rendering pixels. Finally, we proceed with alpha composition, Eq. (1), to render pixels. In our case, the $N$ primitives blend of a pixel-ray depends on the number of sparse voxels assigned to the tile that the pixel-ray belongs to. The computation of the alpha, color, and other geometric properties from our sparse voxels are described in Sec. 3.1.1. When rendering sparse voxels for a pixel-ray, we compute ray-aabb intersection to determine the ray segment to sample (for voxel alpha in Eq. (4)) and skip some non-intersected sparse voxels. We do early termination of the alpha composition if the transmittance of a sparse voxel is below a threshold $T_{i} < h_{\\mathrm{T}}$ .", + "bbox": [ + 89, + 202, + 483, + 383 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Anti-aliasing. To mitigate aliasing artifacts, we render in $h_{\\mathrm{ss}}$ times higher resolution and then apply image downsampling to the target resolution with an anti-aliasing filter.", + "bbox": [ + 89, + 390, + 483, + 436 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2. Progressive Sparse Voxels Optimization", + "text_level": 1, + "bbox": [ + 89, + 444, + 431, + 460 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we describe the procedure to optimize a 3D scene from the input frames with known camera parameters using our SVRaster presented in Sec. 3.1.", + "bbox": [ + 89, + 465, + 483, + 511 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Voxel max sampling rate. We first define the maximum sampling rate $\\mathbf{v}_{\\mathrm{rate}}$ of each voxel on the training images, which reflects the image region a voxel can cover. A smaller $\\mathbf{v}_{\\mathrm{rate}}$ indicates that the voxel is more prone to overfitting due to less observation. We use $\\mathbf{v}_{\\mathrm{rate}}$ in our voxel initialization and subdivision process described later. Given $N_{\\mathrm{cam}}$ training cameras, we estimate the maximum sampling rate of a voxel as follows with visualization in Fig. 5a:", + "bbox": [ + 89, + 517, + 483, + 638 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {v} _ {\\text {r a t e}} = \\max _ {i} ^ {N _ {\\text {c a m}}} \\frac {\\mathbf {v} _ {\\mathrm {s}}}{\\mathbf {v} _ {\\text {i n t e r v a l}} ^ {(i)}}, \\tag {7a}\n$$\n", + "text_format": "latex", + "bbox": [ + 133, + 641, + 482, + 675 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {v} _ {\\text {i n t e r v a l}} ^ {(i)} = \\underbrace {\\left(\\mathbf {v} _ {\\mathrm {c}} - \\mathbf {r} _ {\\mathrm {o}} ^ {(i)}\\right) ^ {\\intercal} \\ell^ {(i)}} _ {\\text {V o x e l z - d i s t a n c e}} \\cdot \\underbrace {\\frac {\\tan \\left(0 . 5 \\cdot \\theta_ {\\text {f o v - x}} ^ {(i)}\\right)}{0 . 5 \\cdot W ^ {(i)}}} _ {\\text {U n i t - d i s t a n c e p i x e l s i z e}}, \\tag {7b}\n$$\n", + "text_format": "latex", + "bbox": [ + 114, + 676, + 482, + 736 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\ell$ is camera lookat vector, $\\theta_{\\mathrm{fov - x}}$ is camera horizontal field of view, and $W$ is image width. The sampling rate indicates the estimated number of rays along the image's horizontal axis direction that may hit the voxel.", + "bbox": [ + 89, + 739, + 483, + 800 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2.1. Scene Initialization", + "text_level": 1, + "bbox": [ + 89, + 806, + 272, + 820 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Without employing an additional prior, we initialize all the parameters to constant. We start with volume density approaching zero by setting voxel raw density to a negative number $h_{\\mathrm{geo}}$ so that the initial activated density $\\mathrm{expl}(h_{\\mathrm{geo}}) \\approx 0$ . We set the SH coefficients to zero for", + "bbox": [ + 89, + 824, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/d27e95ea3cd38fe76c3b5b7b47305570c9c3cd3a9d55ca912138c2e92baed9d5.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 517, + 93, + 637, + 196 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/ee5739594973dabe5c4f9c288e4650a6b06d8665c3264b832b8191c9de6539c7.jpg", + "image_caption": [ + "(b)", + "Figure 5. Visualization of voxel sampling rate and grid layout initialization. (a) We visualize the voxel sampling rate defined in Eq. (7). (b) We depict the foreground main region and the background region under different shell levels. In unbounded scenes, we apply different grid layout initialization strategies for foreground and background regions. See Sec. 3.2.1 for details." + ], + "image_footnote": [], + "bbox": [ + 705, + 95, + 898, + 196 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "non-zero degrees and set the view-independent zero-degree component to yield gray color (i.e., intensity equal 0.5). We detail the Octree grid layout initialization in the following.", + "bbox": [ + 511, + 318, + 905, + 364 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Bounded scenes. In case the scenes or the objects to reconstruct are enclosed in a known bounded region, we simply initialize the layout as a dense grid with $h_{1\\mathrm{v}}$ levels and remove voxels unobserved by any training images. The number of voxels is $\\leq (2^{h_{1\\mathrm{v}}})^3$ after initialization.", + "bbox": [ + 511, + 371, + 905, + 446 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Unbounded scenes. For the unbounded scenes, we first split the space into the main and the unbounded background regions, depicted in Fig. 5b, each with a different heuristic. We use the training camera positions to determine a cuboid for the main region. The cuboid center is set to the average camera positions and the radius is set to the median distance between the cuboid center and the cameras. The same as the bounded scenes, we initialize a dense grid with $h_{\\mathrm{lv}}$ levels for the main region. For the background region, we allocate $h_{\\mathrm{out}}$ level of background shells enclosing the main region, which means that the radius of the entire scene is $2^{h_{\\mathrm{out}}}$ of the main region. In each background shell level, we start with the coarsest voxel size, i.e., $4^3 - 2^3 = 56$ voxels in each shell level. We then iteratively subdivide shell voxels with the highest sampling rate and remove voxels unobserved by any training cameras. The process repeats until the ratio of the number of voxels in the background and the main regions is $h_{\\mathrm{ratio}}$ . The number of voxels is $\\leq (1 + h_{\\mathrm{ratio}})(2^{h_{\\mathrm{lv}}})^3$ after initialization.", + "bbox": [ + 511, + 454, + 906, + 742 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2.2. Adaptive Pruning and Subdivision", + "text_level": 1, + "bbox": [ + 511, + 752, + 799, + 767 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The initialized grid layout only coarsely covers the entire scene that should be adaptively aligned to different levels-of-detail for the scene during the training progress. We apply the following two operations every $h_{\\text{every}}$ training iterations to achieve this purpose.", + "bbox": [ + 511, + 772, + 905, + 848 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Pruning. We compute the maximum blending weight $(T_{i}\\alpha_{i})$ from Eq. (1) of each voxel using all the training cameras. We remove voxels with maximum blending weight", + "bbox": [ + 511, + 854, + 906, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "16191", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "lower than $h_{\\mathrm{prune}}$", + "bbox": [ + 89, + 90, + 215, + 106 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Subdivision. Our heuristic is that a voxel with a larger training loss gradient indicates that the voxel region requires finer voxels to model. Specifically, we accumulate the subdivision priority as the following:", + "bbox": [ + 89, + 112, + 483, + 172 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {v} _ {\\text {p r i o r i t y}} = \\sum_ {\\mathbf {r} \\in R} \\left\\| \\alpha (\\mathbf {r}) \\cdot \\frac {\\partial \\mathcal {L} (\\mathbf {r})}{\\partial \\alpha (\\mathbf {r})} \\right\\|, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 160, + 176, + 482, + 210 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $R$ is the set of all training pixel rays throughout the $h_{\\text{every}}$ iterations and $\\mathcal{L}(\\mathbf{r})$ is the training loss of the ray. The gradient is weighted by alpha values contributed from the voxel to the ray. Higher $\\mathbf{v}_{\\text{priority}}$ indicates higher subdivision priority. To prevent voxels from overfitting few pixels, we set the priority to zero for voxels with maximum sampling rate lower than a sampling rate threshold $\\mathbf{v}_{\\text{rate}} < 2h_{\\text{rate}}$ . Finally, we select the voxels with priority above the top $h_{\\text{percent}}$ percent to subdivide, i.e., the total number of voxels is increased by $(h_{\\text{percent}} \\cdot (8 - 1))$ percent. Note that we only keep the leaf nodes in the Octree layout so we remove the source voxels once they are subdivided.", + "bbox": [ + 89, + 213, + 483, + 393 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "When voxels are pruned and subdivided, the voxel Spherical Harmonic (SH) coefficients and the grid point densities need to be updated accordingly. The SH coefficients are simply pruned together with voxels and duplicated to the subdivided children voxels. Grid point densities are slightly more complex as the eight voxel corner grid points are shared between adjacent voxels (Fig. 2). We remove a grid point only when it does not belong to any voxel corners. When subdividing, we use trilinear interpolation to compute the densities of the new grid points. The duplicated grid points are merged and their densities are averaged.", + "bbox": [ + 89, + 393, + 483, + 561 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2.3. Optimization objectives", + "text_level": 1, + "bbox": [ + 89, + 566, + 302, + 580 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We use MSE and SSIM as the photometric loss between the rendered and the ground truth images. The overall training objective is summarized as:", + "bbox": [ + 89, + 584, + 483, + 630 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} = \\mathcal {L} _ {\\mathrm {m s e}} + \\lambda_ {\\mathrm {s s i m}} \\mathcal {L} _ {\\mathrm {s s i m}} \\\\ + \\lambda_ {\\mathrm {T}} \\mathcal {L} _ {\\mathrm {T}} + \\lambda_ {\\text {d i s t}} \\mathcal {L} _ {\\text {d i s t}} + \\lambda_ {\\mathrm {R}} \\mathcal {L} _ {\\mathrm {R}} + \\lambda_ {\\mathrm {t v}} \\mathcal {L} _ {\\mathrm {t v}}, \\tag {9} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 106, + 638, + 482, + 674 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\lambda$ are the loss weights, $\\mathcal{L}_{\\mathrm{T}}$ encourages the final ray transmittances to be either zero or one, $\\mathcal{L}_{\\mathrm{dist}}$ is the distortion loss [2], $\\mathcal{L}_{\\mathrm{R}}$ is the per-point rgb loss [43], and $\\mathcal{L}_{\\mathrm{tv}}$ is the total variation loss on the sparse density grid. In mesh extraction task, we also add the depth-normal consistency loss from 2DGS [16]:", + "bbox": [ + 89, + 679, + 483, + 770 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {m e s h}} = \\lambda_ {\\mathrm {n - d m e a n}} \\mathcal {L} _ {\\mathrm {n - d m e a n}} + \\lambda_ {\\mathrm {n - d m e d}} \\mathcal {L} _ {\\mathrm {n - d m e d}}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 107, + 779, + 482, + 796 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where both losses encourage alignment between rendered normals and depth-derived normals from mean and median depth. More details are in the supplementary materials.", + "bbox": [ + 89, + 800, + 483, + 845 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.2.4. Sparse-voxel TSDF Fusion and Marching Cubes", + "text_level": 1, + "bbox": [ + 89, + 851, + 473, + 867 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our sparse voxels can be seamlessly integrated with grid-based algorithms. To extract a mesh, we implement March-", + "bbox": [ + 89, + 869, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e8d58f37de38ce27a4313b2c32656c80caa7f7355eac5dd81e3e7120f3af224a.jpg", + "image_caption": [ + "Figure 6. A qualitative comparison with 3DGS [20]. Our result here corresponding to the base version in Tab. 1. We achieve similar visual quality comparing to 3DGS. Note that 3DGS use the coarse geometry from SfM while we do not rely on this prior." + ], + "image_footnote": [], + "bbox": [ + 516, + 89, + 609, + 195 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f2cdbeb90dd4abee73ab8b8c63a6013e6e04bbe3fc68c6c5845dbcd3b53d9d7f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 611, + 90, + 702, + 196 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/851623337ba4d253a469d16b9ea9ae8fe9f3869f701e1b6b534e627b8293da51.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 704, + 90, + 795, + 196 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/caa7d463f4514c280dbd6282de13a96454543cbace5d77917e5c84f7ec808758.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 797, + 90, + 831, + 196 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/0ec9cc628271902213e2cb573f03effa40ef87426d03421724d02ccbd8e96e5c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 831, + 90, + 864, + 196 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e64aaead89acfcf17a5539b96d4c9ff713b8c2a4ae898c578e48a0290765ace3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 867, + 90, + 901, + 196 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a7c38bd876137a04723797d560bb11eca4c0b1c529af8ee98bd7f065f00553e6.jpg", + "image_caption": [ + "Figure 7. Visualization of the reconstructed surface. We show the rendering images, normal maps, and the final meshes on Tanks&Temples and DTU datasets. Note that we only model the scene with density field and do not use the coarse geometry prior from SfM sparse points in this work." + ], + "image_footnote": [], + "bbox": [ + 516, + 273, + 645, + 383 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f06a010990d5fb1e31ed80fe85bfdc08b7cd83bb1863c8f34e06f5872d098a1f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 645, + 275, + 774, + 383 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/6823587515660a834d4214405ce47f47d4325b980720fb75ff8393ff3084ee6f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 774, + 275, + 903, + 383 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ing Cubes [27] to extract the triangles of an isosurface over density from the sparse voxels. The duplicated vertices from adjacent voxels are merged to produce a unique set of vertices. When the adjacent voxels belong to different Octree levels, the extracted triangle may not be connected as the density field is not continuous for voxels in different levels. Such discontinuities can be removed by simply subdividing all voxels to the finest level.", + "bbox": [ + 511, + 489, + 905, + 609 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Deciding the target level set for extracting the isosurface can be tricky for the density field. Instead, we implement sparse-voxel TSDF-Fusion [7, 8, 36] to compute the truncated signed distance values of the sparse grid points. We can then directly extract the surface of the zero-level set using the former sparse-voxel Marching Cubes. Future extensions of our method could directly model signed distance fields following NeuS [47] and VolSDF [53]. The sparse-voxel TSDF-Fusion can still be beneficial to directly initialize our sparse voxel representation from sensor depth.", + "bbox": [ + 511, + 611, + 905, + 763 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 777, + 645, + 794 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Implementation Details", + "text_level": 1, + "bbox": [ + 511, + 801, + 730, + 818 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We use the following implementation details except stated otherwise. We start the optimization from empty space with raw density set to $h_{\\mathrm{geo}} = -10$ . The initial Octree level is $h_{\\mathrm{lv}} = 6$ (i.e., $64^{3}$ voxels) for the bounded scenes and the foreground main region of the unbounded scenes. To model", + "bbox": [ + 511, + 825, + 905, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "16192", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/24ab81e445e67f37047e83605b3b962d07e05809cc85260a06ddcdb14bc03292.jpg", + "table_caption": [], + "table_footnote": [ + "† Re-evaluate on our machine.", + "Table 1. Novel-view synthesis results comparison on MipNeRF360 dataset [2]. The results are averaged from 4 indoor scenes and 5 outdoor scenes. 3DGS uses the sparse points prior from COLMAP [41], whereas the other methods and ours do not." + ], + "table_body": "
MethodMip-NeRF360 dataset
FPS↑Tr. Time↓LPIPS↓PSNR↑SSIM↑
NeRF [30]<1~day0.45123.850.605
M-NeRF [1]<1~day0.44124.040.616
M-NeRF360 [2]<1~day0.23727.690.792
Zip-NeRF [3]†<1~hrs0.18728.550.828
Plenoxels [11]<10~30m0.46323.080.626
INGP [32]~10~5m0.30225.680.705
3DGS [20]†13124m0.21627.450.815
Ours fast-rend2589m0.21026.870.804
Ours fast-train1314.5m0.19927.080.816
Ours12115m0.18527.330.822
", + "bbox": [ + 94, + 88, + 480, + 282 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/2cbe86ae046fa0762e44379bbaf66191b7dc7e45dbb6c1ac4a49375766a695ff.jpg", + "table_caption": [], + "table_footnote": [ + "† Re-evaluate on our machine.", + "Table 2. Comparison on Tanks&Temples [21] and Deep Blending [15] datasets. We follow 3DGS [20] to use two outdoor scenes from Tanks&Temples and two indoor scenes from Deep Blending." + ], + "table_body": "
MethodTanks&TemplesDeep Blending
FPS↑LPIPS↓PSNR↑FPS↑LPIPS↓PSNR↑
Plenoxels [11]130.37921.08110.51023.06
INGP [32]140.30521.9230.39024.96
M-NeRF360 [2]<10.25722.22<10.24529.40
3DGS [20]†1800.17623.751400.24429.60
Ours1250.14423.043660.22829.84
", + "bbox": [ + 93, + 368, + 478, + 476 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5e9d8977242621a3b6547198a855d2e243fb3c60f0ed9a2bf96b41490f0b22d1.jpg", + "image_caption": [ + "GT", + "Figure 8. Failure case. On scenes with severe exposure variation of training views, our method struggles and produces clear boundary of different brightnesses and allocates many floaters. 3DGS on the other hand is less sensitive to photometric variation of GT. This explains our worse PSNR and FPS on Tanks&Temples in Tab. 2." + ], + "image_footnote": [], + "bbox": [ + 93, + 540, + 222, + 652 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/222bcc7c2e59533a270851ebb9c140da2e79d6cf75f99240d4d09ad3a0c54e15.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 222, + 540, + 349, + 652 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/130f561e738b9d0c977e30ff2cc1e61da18ccba322897372a54334905d12ba96.jpg", + "image_caption": [ + "3DGS" + ], + "image_footnote": [], + "bbox": [ + 351, + 540, + 480, + 652 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "unbounded scenes, we use $h_{\\mathrm{out}} = 5$ background shell levels with $h_{\\mathrm{ratio}} = 2$ times the number of foreground voxels (Sec. 3.2.1 and Fig. 5). The early ray stopping threshold is set to $h_{\\mathrm{T}} = 1\\mathrm{e} - 4$ and the supersampling scale is set to $h_{\\mathrm{ss}} = 1.5$ . Inside each voxel, we sample $K = 1$ point for novel-view synthesis and $K = 3$ points for mesh reconstruction task. We train our model for 20K iterations. The voxels are pruned and subdivided every $h_{\\mathrm{every}} = 1,000$ iterations. The pruning threshold $h_{\\mathrm{prune}}$ is linearly scale from 0.0001 to 0.05 and the subdivision percentage is $h_{\\mathrm{percent}} = 5$ . More", + "bbox": [ + 88, + 750, + 482, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/9b69265384f9ed7079cbc0cfc04742f36d73ff8c343f93bc14c0695337105c0f.jpg", + "table_caption": [], + "table_footnote": [ + "Table 3. Memory, model size, and high-res FPS. The results are averaged on Mip-NeRF360 dataset. The standard 1x evaluation resolution have about 1–1.6M pixels per frame." + ], + "table_body": "
MethodPeak GPU mem.↓Model size↓FPS at higher res.↑
1x2x3x
3DGS1.8GB0.7GB1316939
Ours3.9GB1.8GB12110369
", + "bbox": [ + 516, + 88, + 903, + 159 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "details are provided in the supplement.", + "bbox": [ + 511, + 220, + 769, + 234 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We use the standard LPIPS [56], SSIM, and PSNR metrics to evaluate novel-view quality. We directly employ the LPIPS implementation from prior work [20], which uses [0, 1] image intensity range with VGG [42] network. For mesh accuracy, we follow the benchmarks to use F-score and chamfer distance. We use the test-set images to measure the FPS on a desktop computer with a 3090 Ti GPU.", + "bbox": [ + 511, + 236, + 903, + 340 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Novel-view Synthesis", + "text_level": 1, + "bbox": [ + 511, + 364, + 710, + 380 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In Tab. 1, we show the quantitative comparison on the MipNeRF-360 dataset. We also provide a fast-rendering variant with $>2\\mathrm{x}$ FPS by setting $h_{\\mathrm{ss}} = 1.5$ , $h_{\\mathrm{prune}} = 0.15$ , and a fast-training variant with $>3\\mathrm{x}$ training speedup by scaling all iteration related hyperparameter by 0.3. Both fast variants only reduce quality moderately.", + "bbox": [ + 511, + 387, + 903, + 477 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our rendering speed is comparable to 3DGS on average. However, FPS varies significantly across scenes (see supplementary for detailed per-scene results). Several differences impact speed between our method and 3DGS. For instance, 3DGS sorts 32-bit floats, while we sort 48-bit Morton codes (16 levels, each with 3 bits) for primitive ordering. We also avoid the overhead of computing inverse covariance matrices from quaternions and scaling parameters. We directly intersect rays with voxels in 3D space, whereas 3DGS approximates the projection of 3D Gaussians to 2D via a linear affine transform. Finally, Gaussian distributions decay gradually to zero, while our post-activation [18, 43] voxel density field can be arbitrarily sharp. As a result, the average numbers of primitives contributing to a pixel intensity is 63 for 3DGS and 27 for us. The influence of these factors on rendering speed is scene-dependent.", + "bbox": [ + 511, + 478, + 905, + 718 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Regarding rendering quality metrics, our method achieves significantly better LPIPS than 3DGS and even surpasses Zip-NeRF. This can be explained by the qualitative comparison in Fig. 6 where our method recover finer details. The SSIM comparison results with 3DGS is more scene-dependent. SSIM comparisons with 3DGS are more scene-dependent, while PSNR tends to favor 3DGS, as it prefers smoother rendering in uncertain regions. More visual comparisons are in the supplementary.", + "bbox": [ + 511, + 719, + 905, + 854 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our method uses much more voxel primitives than Gaussian, resulting in larger model size and requires more GPU memory as shown in Tab. 3. Interestingly, our high FPS is", + "bbox": [ + 511, + 854, + 903, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "16193", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/cb96906182bd6ece656ed5f686f26c632e6a59bdabf19cb50d95484101a52747.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Resolution of main2563512310243adaptivePlenoxels 6403
LPIPS↓0.4440.3260.2000.452
PSNR↑23.9825.37OOM28.0123.29
FPS↑457190171<10
", + "bbox": [ + 93, + 88, + 480, + 151 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "more scalable to higher resolution rendering perhaps due to the fewer contributing primitives per pixel.", + "bbox": [ + 89, + 250, + 482, + 281 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We compare results on two more datasets in Tab. 2. On Deep Blending [15] dataset, we achieve much better speed and quality than 3DGS. On Tanks&Temples dataset [21], despite having better LPIPS, our PSNR and FPS are worse than 3DGS. The failure results are shown in Fig. 8 where the training views have large exposure variation. As a result, our method produces a clear brightness boundary in the scene with many floaters that slow down our rendering.", + "bbox": [ + 89, + 281, + 483, + 402 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Ablation Studies", + "text_level": 1, + "bbox": [ + 89, + 425, + 254, + 439 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The rasterization process improves the rendering speed of sparse voxels, while the key to achieve high-quality results is the adaptive voxel size for different levels of detail. We show an ablation experiment in Tab. 4. Instead of subdividing voxels adaptively, the uniform voxel size variant subdivides all the voxels at certain training iterations until the grid resolution of the foreground main region reaches $256^{3}$ , $512^{3}$ , or $1024^{3}$ . We also align the background and the pruning setup where the details are deferred to the supplementary. As shown in the table, the quality of uniform voxel size in $512^{3}$ resolution is much worse than the adaptive voxel size under similar FPS. Our machine with 24GB GPU memory fails to reach $1024^{3}$ grid resolution despite the voxels being pruned and sparse. Plenoxels [11] renders sparse voxels by ray casting and sampling following NeRF. Their FPS is significantly lower than our rasterization-based sparse voxel rendering. As they use a dense 3D pointer grid to support point sampling on rays, the model scalability for the unbounded background region is thus limited. To workaround, they apply a multi-sphere image with 64 layers to model the background. Conversely, we set the scene size as $32\\mathrm{x}$ larger ( $h_{\\mathrm{out}} = 5$ ) than the foreground cuboid and direct model the entire scene by our sparse voxels. As a result, our uniform size variant with $512^{3}$ foreground resolution still outperforms Plenoxels with $640^{3}$ grid.", + "bbox": [ + 89, + 446, + 483, + 824 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In our rasterizer, we use direction-dependent Morton order to ensure the sorting of voxels from different levels is correct. However, like the case in 3DGS, the popping artifact by the incorrect order is not a major factor in the numerical results. Instead, we provide more visualization in", + "bbox": [ + 89, + 825, + 483, + 901 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/8885cb64f54161fd086768fff6adaffd1d225ab6ae826bd24a9be6aadb45f0bd.jpg", + "table_caption": [ + "Table 4. Ablation experiments of adaptive and uniform voxel sizes. The results are evaluated on the indoor bonsai and the outdoor bicycle scenes from MipNeRF-360 dataset. The resolutions at the first row indicate the final grid resolution of the main foreground cuboid. Plenxoels is the previous fully-explicit voxel grid approach. Please refer to Sec. 4.3 for more discussion." + ], + "table_footnote": [], + "table_body": "
MethodGeom.Init.Tansk&TemplesDTU
F-score↑Tr. time↓Cf.↓Tr. time↓
NeRF [30]DensityRandom--1.49hrs
NeuS [47]SDFSphere0.38hrs0.84hrs
Voxurf [51]SDFSphere--0.7615m
Neuralangelo [24]SDFSphere0.50hrs0.61hrs
3DGS [20]GaussiansSfM pts.0.1914m1.9611m
SuGaR [14]GaussiansSfM pts.0.09~1h1.33~1h
2DGS [16]GaussiansSfM pts.0.3216m0.8011m
OursDensityConstant0.4011m0.765m
", + "bbox": [ + 516, + 88, + 903, + 218 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. Mesh results comparison on the Tanks&Temples [21] and DTU [17] datasets. Our method with volume density and constant field initialization achieves good accuracy-time trade-off on the unbounded-level and object-level datasets.", + "bbox": [ + 511, + 219, + 906, + 273 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "the supplementary videos to show its effect. See more ablation studies in the supplementary materials.", + "bbox": [ + 511, + 287, + 903, + 318 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Mesh Reconstruction", + "text_level": 1, + "bbox": [ + 511, + 325, + 714, + 340 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We extract mesh by adapting TSDF-Fusion [34] and Marching Cubes [27] into our sparse voxels. The results comparison on the large-scale Tanks&Temples and the object-scale DTU datasets is provided in Tab. 5, where our method achieves a good balance of accuracy and training time on both datasets. See qualitative examples in Fig. 7 and the supplementary materials. We find that our current method tends to produce unnecessary geometric bumps for texture details. Future work could improve surface smoothness by adapting our density field for direct signed distance field modeling or incorporating SfM sparse points prior.", + "bbox": [ + 511, + 348, + 906, + 515 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. 2D Feature Fusion", + "text_level": 1, + "bbox": [ + 511, + 523, + 692, + 537 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We showcase some results in Fig. 1, where we fuse 2D semantic segmentation and high-dimensional foundation features into our sparse voxels using Voxel Pooling and Volume Fusion. The multi-view ensemble smoothes the inconsistent 2D predictions, while our detailed 3D geometry enables higher-resolution rendering of 2D features. See more results in the released code.", + "bbox": [ + 511, + 545, + 905, + 650 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 664, + 633, + 679 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work presents a novel differentiable radiance field rendering system that integrates an efficient rasterizer with a multi-level sparse voxel scene representation. We reconstruct a scene from the multi-view input images by adaptively fitting the sparse voxels into different levels-of-detail. The results reveal that fully explicit voxel models, without neural networks or Gaussians, can achieve state-of-the-art comparable novel-view rendering speed and quality. The key breakthrough is overcoming the scalability constraint of voxel grids with adaptive sparse voxels and improving rendering speed through rasterization. We believe that incorporating our method with classical 3D processing algorithm [7, 8, 19, 26, 36] and 3D neural network [6, 13, 50] are promising future directions.", + "bbox": [ + 511, + 688, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "16194", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jonathan T. Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P. Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In ICCV, 2021. 7", + "[2] Jonathan T. Barron, Ben Mildenhall, Dor Verbin, Pratul P. Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In CVPR, 2022. 6, 7", + "[3] Jonathan T. Barron, Ben Mildenhall, Dor Verbin, Pratul P. Srinivasan, and Peter Hedman. Zip-nerf: Anti-aliased grid-based neural radiance fields. In ICCV, 2023. 2, 7", + "[4] Eric R. Chan, Connor Z. Lin, Matthew A. Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J. Guibas, Jonathan Tremblay, Sameh Khamis, Tero Karras, and Gordon Wetzstein. Efficient geometry-aware 3d generative adversarial networks. In CVPR, 2022. 1, 2", + "[5] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In ECCV, 2022. 1, 2", + "[6] Christopher Choy, JunYoung Gwak, and Silvio Savarese. 4d spatio-temporal convnets: Minkowski convolutional neural networks. In CVPR, 2019. 2, 8", + "[7] Brian Curless and Marc Levoy. A volumetric method for building complex models from range images. ACM TOG, 1996. 1, 2, 6, 8", + "[8] Angela Dai, Matthias Nießner, Michael Zollöfer, Shahram Izadi, and Christian Theobalt. Bundlefusion: Real-time globally consistent 3d reconstruction using on-the-fly surface re-integration. ACM TOG, 2017. 1, 2, 6, 8", + "[9] Kangle Deng, Andrew Liu, Jun-Yan Zhu, and Deva Ramanan. Depth-supervised nerf: Fewer views and faster training for free. In CVPR, 2022. 2", + "[10] Daniel Duckworth, Peter Hedman, Christian Reiser, Peter Zhizhin, Jean-François Thibert, Mario Lucic, Richard Szeliski, and Jonathan T. Barron. SMERF: streamable memory efficient radiance fields for real-time large-scene exploration. ACM TOG, 2024. 2", + "[11] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In CVPR, 2022. 2, 7, 8", + "[12] Qiancheng Fu, Qingshan Xu, Yew Soon Ong, and Wenbing Tao. Geo-neus: Geometry-consistent neural implicit surfaces learning for multi-view reconstruction. In _NeurIPS_, 2022. 2", + "[13] Benjamin Graham and Laurens Van der Maaten. Submanifold sparse convolutional networks. arXiv preprint arXiv:1706.01307, 2017. 2, 8", + "[14] Antoine Guédon and Vincent Lepetit. Sugar: Surface-aligned gaussian splatting for efficient 3d mesh reconstruction and high-quality mesh rendering. In CVPR, 2024. 2, 8", + "[15] Peter Hedman, Julien Philip, True Price, Jan-Michael Frahm, George Drettakis, and Gabriel J. Brostow. Deep blending for free-viewpoint image-based rendering. ACM TOG, 2018. 7, 8" + ], + "bbox": [ + 93, + 114, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[16] Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2d gaussian splatting for geometrically accurate radiance fields. ACM TOG, 2024. 2, 6, 8", + "[17] Rasmus Ramsbøl Jensen, Anders Lindbjerg Dahl, George Vogiatzis, Engin Tola, and Henrik Aanæs. Large scale multiview stereopsis evaluation. In CVPR, 2014. 8", + "[18] Animesh Karnewar, Tobias Ritschel, Oliver Wang, and Niloy J. Mitra. Relu fields: The little non-linearity that could. In ACM TOG, 2022. 3, 7", + "[19] Michael Kazhdan, Matthew Bolitho, and Hugues Hoppe. Poisson surface reconstruction. In SGP, 2006. 2, 8", + "[20] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM TOG, 2023. 1, 2, 3, 4, 6, 7, 8", + "[21] Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and temples: benchmarking large-scale scene reconstruction. ACM TOG, 2017. 7, 8", + "[22] Samuli Laine and Tero Karras. Efficient sparse voxel octrees. In ACM I3D, 2010. 2", + "[23] Ruilong Li, Hang Gao, Matthew Tancik, and Angjoo Kanazawa. Nerfacc: Efficient sampling accelerates nerfs. In ICCV, 2023. 2", + "[24] Zhaoshuo Li, Thomas Müller, Alex Evans, Russell H. Taylor, Mathias Unberath, Ming-Yu Liu, and Chen-Hsuan Lin. Neuralangelo: High-fidelity neural surface reconstruction. In CVPR, 2023. 1, 2, 8", + "[25] Lingjie Liu, Jiatao Gu, Kyaw Zaw Lin, Tat-Seng Chua, and Christian Theobalt. Neural sparse voxel fields. In NeurIPS, 2020. 2", + "[26] Charles Loop, Qin Cai, Sergio Orts-Escolano, and Philip A Chou. A closed-form bayesian fusion equation using occupancy probabilities. In 3DV, 2016. 2, 8", + "[27] William E. Lorensen and Harvey E. Cline. Marching cubes: A high resolution 3d surface construction algorithm. In ACM TOG, 1987. 6, 8", + "[28] Alexander Mai, Peter Hedman, George Kopanas, Dor Verbin, David Futschik, Qiangeng Xu, Falko Kuester, Jonathan T. Barron, and Yinda Zhang. Ever: Exact volumetric ellipsoid rendering for real-time view synthesis. arXiv preprint arXiv:2410.01804, 2024. 2", + "[29] Nelson L. Max. Optical models for direct volume rendering. IEEE TVCG, 1995. 4", + "[30] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 2, 3, 4, 7, 8", + "[31] Nicolas Moenne-Loccoz, Ashkan Mirzaei, Or Perel, Ricardo de Lutio, Janick Martinez Esturo, Gavriel State, Sanja Fidler, Nicholas Sharp, and Zan Gojcic. 3d gaussian ray tracing: Fast tracing of particle scenes. ACM TOG, 2024. 2", + "[32] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM TOG, 2022. 2, 7", + "[33] Ken Museth. Vdb: High-resolution sparse volumes with dynamic topology. ACM TOG, 2013. 2", + "[34] Richard A Newcombe, Shahram Izadi, Otmar Hilliges, David Molyneaux, David Kim, Andrew J Davison, Pushmeet" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "16195", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Kohi, Jamie Shotton, Steve Hodges, and Andrew Fitzgibbon. Kinectfusion: Real-time dense surface mapping and tracking. In IEEE ISMAR, 2011. 1, 2, 8", + "[35] Michael Niemeyer, Fabian Manhardt, Marie-Julie Rakoto-saona, Michael Oechsle, Daniel Duckworth, Rama Gosula, Keisuke Tateno, John Bates, Dominik Kaeser, and Federico Tombari. Radsplat: Radiance field-informed gaussian splat-ting for robust real-time rendering with $900+$ FPS. arXiv preprint arXiv:2403.13806, 2024. 2", + "[36] Matthias Nießner, Michael Zollhöfer, Shahram Izadi, and Marc Stamminger. Real-time 3d reconstruction at scale using voxel hashing. ACM TOG, 2013. 2, 6, 8", + "[37] Lukas Radl, Michael Steiner, Mathias Parger, Alexander Weinrauch, Bernhard Kerbl, and Markus Steinberger. Stophepop: Sorted gaussian splatting for view-consistent real-time rendering. ACM TOG, 2024. 1, 2, 4", + "[38] Mike Ranzinger, Greg Heinrich, Jan Kautz, and Pavlo Molchanov. Am-radio: Agglomerative vision foundation model reduce all domains into one. In CVPR, 2024. 1", + "[39] Christian Reiser, Songyou Peng, Yiyi Liao, and Andreas Geiger. Kilonerf: Speeding up neural radiance fields with thousands of tiny mllps. In ICCV, 2021. 2", + "[40] Christian Reiser, Stephan J. Garbin, Pratul P. Srinivasan, Dor Verbin, Richard Szeliski, Ben Mildenhall, Jonathan T. Barron, Peter Hedman, and Andreas Geiger. Binary opacity grids: Capturing fine geometric detail for mesh-based view synthesis. ACM TOG, 2024. 2", + "[41] Johannes L. Schonberger, Enliang Zheng, Jan-Michael Frahm, and Marc Pollefeys. Pixelwise view selection for unstructured multi-view stereo. In ECCV, 2016. 2, 7", + "[42] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. In ICLR, 2015. 7", + "[43] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In CVPR, 2022. 1, 2, 3, 6, 7", + "[44] Cheng Sun, Guangyan Cai, Zhengqin Li, Kai Yan, Cheng Zhang, Carl S. Marshall, Jia-Bin Huang, Shuang Zhao, and Zhao Dong. Neural-pbir reconstruction of shape, material, and illumination. In ICCV, 2023. 1, 2", + "[45] Towaki Takikawa, Joey Litalien, Kangxue Yin, Karsten Kreis, Charles T. Loop, Derek Nowrouzezahrai, Alec Jacobson, Morgan McGuire, and Sanja Fidler. Neural geometric level of detail: Real-time rendering with implicit 3d shapes. In CVPR, 2021. 2", + "[46] Haithem Turki, Vasu Agrawal, Samuel Rota Bulò, Lorenzo Porzi, Peter Kontschieder, Deva Ramanan, Michael Zollhöfer, and Christian Richardt. Hybridnerf: Efficient neural rendering via adaptive volumetric surfaces. In CVPR, 2024. 2", + "[47] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. In NeurIPS, 2021. 2, 6, 8", + "[48] Yiming Wang, Qin Han, Marc Habermann, Kostas Dani-ilidis, Christian Theobalt, and Lingjie Liu. Neus2: Fast learning of neural implicit surfaces for multi-view reconstruction. In ICCV, 2023. 2" + ], + "bbox": [ + 91, + 90, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[49] Zian Wang, Tianchang Shen, Merlin Nimier-David, Nicholas Sharp, Jun Gao, Alexander Keller, Sanja Fidler, Thomas Müller, and Zan Gojcic. Adaptive shells for efficient neural radiance field rendering. ACM TOG, 2023. 2", + "[50] Francis Williams, Jiahui Huang, Jonathan Swartz, Gergely Klár, Vijay Thakkar, Matthew Cong, Xuanchi Ren, Ruilong Li, Clement Fuji-Tsang, Sanja Fidler, Eftychios Sifakis, and Ken Museth. fvdb: A deep-learning framework for sparse, large scale, and high performance spatial intelligence. ACM TOG, 2024. 2, 8", + "[51] Tong Wu, Jiaqi Wang, Xingang Pan, Xudong Xu, Christian Theobalt, Ziwei Liu, and Dahua Lin. Voxurf: Voxel-based efficient and accurate neural surface reconstruction. In ICLR, 2023. 1, 2, 8", + "[52] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, José M. Álvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. In NeurIPS, 2021. 1", + "[53] Lior Yariv, Jiatao Gu, Yoni Kasten, and Yaron Lipman. Volume rendering of neural implicit surfaces. In NeurIPS, 2021. 2, 6", + "[54] Lior Yariv, Peter Hedman, Christian Reiser, Dor Verbin, Pratul P. Srinivasan, Richard Szeliski, Jonathan T. Barron, and Ben Mildenhall. Bakedsdf: Meshing neural sdfs for real-time view synthesis. In ACM TOG, 2023. 2", + "[55] Alex Yu, Ruilong Li, Matthew Tancik, Hao Li, Ren Ng, and Angjoo Kanazawa. Plenoctrees for real-time rendering of neural radiance fields. In ICCV, 2021. 2", + "[56] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 7" + ], + "bbox": [ + 516, + 92, + 903, + 530 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "16196", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Sparse Voxels Rasterization_ Real-time High-fidelity Radiance Field Rendering/a6bd5a3d-4547-4194-a227-b9c1f40e864c_model.json b/2025/Sparse Voxels Rasterization_ Real-time High-fidelity Radiance Field Rendering/a6bd5a3d-4547-4194-a227-b9c1f40e864c_model.json new file mode 100644 index 0000000000000000000000000000000000000000..e9105633684156b1d3c84f13c28b3f5fb60ad2a5 --- /dev/null +++ b/2025/Sparse Voxels Rasterization_ Real-time High-fidelity Radiance Field Rendering/a6bd5a3d-4547-4194-a227-b9c1f40e864c_model.json @@ -0,0 +1,2783 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.1, + 0.131, + 0.898, + 0.155 + ], + "angle": 0, + "content": "Sparse Voxels Rasterization: Real-time High-fidelity Radiance Field Rendering" + }, + { + "type": "text", + "bbox": [ + 0.106, + 0.18, + 0.204, + 0.2 + ], + "angle": 0, + "content": "Cheng Sun1" + }, + { + "type": "text", + "bbox": [ + 0.235, + 0.181, + 0.355, + 0.217 + ], + "angle": 0, + "content": "Jaesung Choe1 \n1NVIDIA" + }, + { + "type": "text", + "bbox": [ + 0.383, + 0.181, + 0.503, + 0.217 + ], + "angle": 0, + "content": "Charles Loop1 \n2Cornell University" + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.181, + 0.694, + 0.217 + ], + "angle": 0, + "content": "Wei-Chiu Ma\\(^{2}\\) \n\\(^{3}\\)National Taiwan University" + }, + { + "type": "text", + "bbox": [ + 0.685, + 0.181, + 0.89, + 0.217 + ], + "angle": 0, + "content": "Yu-Chiang Frank Wang\\(^{1,3}\\) \n\\(^{1}\\)" + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.226, + 0.387, + 0.404 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.107, + 0.42, + 0.376, + 0.432 + ], + "angle": 0, + "content": "(a) Volume rendering by rasterizing sparse voxels." + }, + { + "type": "image", + "bbox": [ + 0.387, + 0.227, + 0.681, + 0.419 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.413, + 0.42, + 0.68, + 0.432 + ], + "angle": 0, + "content": "(b) Novel-view rendering on Mip-NeRF360 scenes." + }, + { + "type": "image", + "bbox": [ + 0.687, + 0.226, + 0.901, + 0.416 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.727, + 0.42, + 0.856, + 0.432 + ], + "angle": 0, + "content": "(c) 2D-to-3D made easy." + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.439, + 0.907, + 0.522 + ], + "angle": 0, + "content": "Figure 1. We propose SVRaster, a novel framework for multi-view reconstruction and novel view synthesis. (a) Sparse voxel representation effectively captures the volume density and radiance field of the scene, without the need for neural networks, 3D Gaussians, and sparse-points prior. (b) Using our customized sparse voxel rasterizer, we can learn the underlying 3D scene efficiently and achieve state-of-the-art performance in both rendering quality and speed. (c) Notably, lifting 2D modal to the trained sparse voxels is simple and efficient by integrating the classic Volume Fusion [7, 8, 34]. We show examples of vision foundation model feature field from RADIO [38], semantic field from Segformer [52], and signed distance field from rendered depth, making it flexible and suitable for a wide range of applications." + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.547, + 0.327, + 0.563 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.581, + 0.485, + 0.869 + ], + "angle": 0, + "content": "We propose an efficient radiance field rendering algorithm that incorporates a rasterization process on adaptive sparse voxels without neural networks or 3D Gaussians. There are two key contributions coupled with the proposed system. The first is to adaptively and explicitly allocate sparse voxels to different levels of detail within scenes, faithfully reproducing scene details with \\(65536^{3}\\) grid resolution while achieving high rendering frame rates. Second, we customize a rasterizer for efficient adaptive sparse voxels rendering. We render voxels in the correct depth order by using ray direction-dependent Morton ordering, which avoids the well-known popping artifact found in Gaussian splatting. Our method improves the previous neural-free voxel model by over 4db PSNR and more than 10x FPS speedup, achieving state-of-the-art comparable novel-view synthesis results. Additionally, our voxel representation is seamlessly compatible with grid-based 3D processing techniques such as Volume Fusion, Voxel Pooling, and Marching Cubes, enabling a wide range of future extensions and applications." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.869, + 0.324, + 0.883 + ], + "angle": 0, + "content": "Code: github.com/NVlabs/svraster" + }, + { + "type": "title", + "bbox": [ + 0.515, + 0.547, + 0.645, + 0.563 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.576, + 0.907, + 0.757 + ], + "angle": 0, + "content": "Gaussian splatting [20] has emerged as one of the most promising solutions for novel view synthesis. It has drawn wide attention across multiple communities due to its exceptional rendering speed and ability to capture the nuanced details of a scene. Nevertheless, Gaussian splatting in its base form has two key limitations: first, sorting Gaussians based on their centers does not guarantee proper depth ordering. It may result in popping artifacts [37] (i.e., sudden color changes for consistent geometry) when changing views. Second, the volume density of a 3D point is ill-defined when covered by multiple Gaussians. This ambiguity makes surface reconstruction non-trivial." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.761, + 0.909, + 0.868 + ], + "angle": 0, + "content": "On the other hand, grid representations inherently avoid these issues—both ordering and volume are well-defined. However, due to ray casting, the rendering speed of these methods [4, 5, 24, 43, 44, 51] is slower than that of Gaussian Splatting. This raises the question: is it possible to take the best of both worlds? Can we combine the efficiency of Gaussian splatting with the well-defined grid properties?" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Fortunately, the answer is yes. Our key insight is to revisit voxels—a well-established primitive with decades of" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16187" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.212 + ], + "angle": 0, + "content": "history. Voxel representations are inherently compatible with modern graphics engines and can be rasterized efficiently. Additionally, previous work has demonstrated their ability to model scene volume densities, despite through volume ray casting. This positions voxels as the perfect bridge between rasterization and volumetric representation. However, naively adopting voxels does not work well [11]. Since a scene may consist of different levels of detail." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.213, + 0.483, + 0.424 + ], + "angle": 0, + "content": "With these observations in mind, we present SVRaster, a novel framework that combines the efficiency of rasterization in 3DGS with the structured volumetric approach of grid-based representations. SVRaster leverages (1) multilevel sparse voxels to model 3D scenes and (2) implements a direction-dependent Morton order encoding that facilitates the rasterization rendering from our adaptive-sized sparse voxel representation. Our rendering is free from popping artifacts because the 3D space is partitioned into disjoint voxels, and our sorting ensures the correct rendering order. Moreover, thanks to the volumetric nature and neural-free representation of SVRaster, our sparse voxel grid can be easily and seamlessly integrated with classical grid-based 3D processing algorithms." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.424, + 0.483, + 0.514 + ], + "angle": 0, + "content": "We show that our method is training fast, rendering fast, and achieves novel-view synthesis quality comparable to the state-of-the-art. We also integrate Volume Fusion, Voxel Pooling, and Marching Cubes operations into our adaptive sparse voxels, which showcases promising results of mesh extraction and 2D foundation feature fusion." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.529, + 0.233, + 0.545 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.554, + 0.483, + 0.719 + ], + "angle": 0, + "content": "Differentiable volume rendering has made significant strides in 3D scene reconstruction and novel-view synthesis tasks. Neural Radiance Fields (NeRF) [30] laid the foundation for this progress by optimizing a volumetric function, parameterized by multi-layer perceptrons (MLPs) to encode both geometry and appearance through differential rendering. Subsequent studies focus on accelerating speed by decomposing large MLPs into grid-based representations, with a shallow MLP typically still being employed. Several grid representations have been explored—dense grids [43], factorized grids [5], tri-planes [4], and hash grids [32]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.721, + 0.483, + 0.901 + ], + "angle": 0, + "content": "In particular, our method is closely related to sparse voxel grid representations. Sparse Voxel Octrees [22] and VDB trees [33] are commonly used to manage sparse voxels and facilitate rendering. In contrast, our sparse voxels are stored in a 1D array without advanced data structures. Our rasterizer with the proposed direction-dependent Morton order ensures correct rendering order. To model volumes in sparse leaf nodes, different strategies exist, such as low-resolution 3D grid [33, 50] or implicit neural field [25, 45]. Our leaf node is a single voxel with explicit density and color parameters, like Plenoxels [11]. Regarding voxel levels, previous methods allocate voxels to a tar" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.137 + ], + "angle": 0, + "content": "get level [11, 25, 45, 55] or use a shallow tree [33, 50]. Our voxels can adaptively fit into different levels across the entire tree depth, maximizing flexibility and scalability." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.139, + 0.905, + 0.229 + ], + "angle": 0, + "content": "Another scalability issue of the previous grid-based methods is that they still use some sort of dense 3D grid in the field of novel-view synthesis. For instance, dense occupancy grids [5, 23, 32, 43] for free-space skipping or dense pointer grid [11] to support sparse voxel lookup. We do not use any 3D dense data structures." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.231, + 0.905, + 0.473 + ], + "angle": 0, + "content": "3D Gaussian Splatting (3DGS) [20] takes a different approach by representing scenes with 3D Gaussian primitives and using rasterization for rendering, achieving state-of-the-art trade-off for quality and speed. Our work is inspired by the efficiency of using a rasterizer by 3DGS. However, 3DGS exhibits view-inconsistent popping artifacts due to inaccurate rendering order and primitive overlapping. Some recent works mitigate this artifact [31, 37], but completely resolving it significantly harms rendering speed [28]. Our method does not suffer from the popping artifact. Typically, 3DGS [20] initializes Gaussians from the triangulated points via COLMAP [41], which is especially critical for unbounded scenes. The sparse points geometry is also employed to guide the training of volumetric representations and has shown improvement in some NeRF-based methods [9, 12]. In this work, we do not use the sparse points." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.475, + 0.905, + 0.686 + ], + "angle": 0, + "content": "Mesh reconstruction is an important extended topic for both volumetric-based and GS-based methods. Recent 3DGS variants [14, 16] with proper regularization have shown a good speed-accuracy trade-off. However, directly extracting meshes from 3DGS remains challenging, so these methods still rely on volumetric-based post-processing algorithms to extract meshes. NeRF variants [24, 44, 47, 48, 51, 53] improve surface quality by rendering a signed distance field (SDF) instead of density field. The well-defined volume makes it straightforward to extract a mesh from the isosurface. Our sparse voxels are also a volumetric representation. Though our voxel also follows volumetric representation focusing on the density field, we still can achieve promising accuracy with minimal time." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.688, + 0.905, + 0.824 + ], + "angle": 0, + "content": "For efficient rendering, another line of research is to convert (or distill) the geometry and appearance of a high-quality but slow model into other efficient kind of representations such as smaller MLPs [39, 46], meshes [40, 49, 54], grids [10], or the recent Gaussians [35]. Typically, the powerful Zip-NeRF [3] is employed as the teacher model which requires hours of training per scene. Our volumetric representation could enable training-free conversion from a large implicit field, which is however out of our current scope." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Finally, the sparse voxel is also a widely-used representation for 3D processing [7, 8, 19, 26, 34, 36] and understanding [6, 13, 50]. We mainly focus on rendering in this work while we believe that adapting our method to these techniques is a promising future direction." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "16188" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.089, + 0.283, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.179, + 0.24, + 0.196, + 0.251 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.284, + 0.088, + 0.48, + 0.236 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.38, + 0.239, + 0.398, + 0.251 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.263, + 0.485, + 0.417 + ], + "angle": 0, + "content": "Figure 2. Sparse voxels scene representation. (Left) We allocate voxel under an Octree layout. Each voxels has its own Spherical Harmonic coefficient for view-dependent appearance. The color field is approximated as a constant inside a voxel when rendering a view for efficiency. The density field is trilaterally varied inside a voxel and is modeled by the density values on the corner grid points (i.e., the black dots \\(\\bullet\\)) of each voxel. The grid points densities are shared between adjacent voxels. (Right) We evenly sample \\(K\\) points inside the segment of ray-voxel intersection to compute volume integration for its alpha value contributing to the pixel ray. See Sec. 3.1.1 for details." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.43, + 0.2, + 0.447 + ], + "angle": 0, + "content": "3. Approach" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.455, + 0.484, + 0.547 + ], + "angle": 0, + "content": "We present our approach as follows. First in Sec. 3.1, we introduce our sparse voxels scene representation and our rasterizer for rendering sparse voxels into pixels. Later in Sec. 3.2, we describe the progressive scene optimization strategy, which is designed for faithfully reconstructing the scene from multi-view images using our sparse voxels." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.557, + 0.342, + 0.573 + ], + "angle": 0, + "content": "3.1. Sparse Voxels Rasterization" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.58, + 0.484, + 0.625 + ], + "angle": 0, + "content": "Recent neural radiance field rendering approaches, such as NeRF [30] and 3DGS [20] variants, use the following alpha composition equation to render a pixel's color C:" + }, + { + "type": "equation", + "bbox": [ + 0.11, + 0.628, + 0.483, + 0.657 + ], + "angle": 0, + "content": "\\[\n\\mathrm {C} = \\sum_ {i = 1} ^ {N} T _ {i} \\cdot \\alpha_ {i} \\cdot c _ {i}, \\quad T _ {i} = \\prod_ {j = 1} ^ {i - 1} \\left(1 - \\alpha_ {j}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.661, + 0.483, + 0.766 + ], + "angle": 0, + "content": "where \\(\\alpha_{i}\\in \\mathbb{R}_{\\in [0,1]}\\) and \\(c_{i}\\in \\mathbb{R}_{\\in [0,1]}^{3}\\) are alpha and view-dependent color at the \\(i\\)-th sampled point or primitive on the pixel-ray. The quantity \\(T_{i}\\) is known as the transmittance. At a high-level, the difference between each method comes down to: \\(i\\) ) how they find the \\(N\\) points or primitives to composite a pixel-ray and \\(ii\\) ) how the rgb and alpha values are determined from their scene representations." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.768, + 0.484, + 0.844 + ], + "angle": 0, + "content": "Our sparse-voxel method follows the same principle. In Sec. 3.1.1, we provide details of our sparse voxels scene representation and how the \\(\\alpha_{i}\\) and \\(c_{i}\\) in Eq. (1) are computed from a voxel. In Sec. 3.1.2, we present our rasterizer, that gathers the \\(N\\) voxels to be composited for each pixel." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.852, + 0.288, + 0.867 + ], + "angle": 0, + "content": "3.1.1. Scene Representation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "We first describe the grid layout for allocating our sparse voxels and then derive the alpha value, view-dependent" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.122 + ], + "angle": 0, + "content": "color, and other geometric properties needed for the composite rendering of a voxel." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.131, + 0.907, + 0.386 + ], + "angle": 0, + "content": "Sparse voxel grid. Our SVRaster constructs 3D scenes using a sparse voxel representation. We allocate voxels following an Octree space partition rule (i.e., Octree layout as illustrated in Fig. 2a) for two reasons necessary for achieving high-quality results. First, it facilitates the correct rendering order of voxels with various sizes (Sec. 3.1.2). Second, we can adaptively fit the sparse voxels to different scene level-of-details (Sec. 3.2). Note that our representation does not replicate a traditional Octree data structure with parent-child pointers or linear Octree. Specifically, we only keep voxels at the Octree leaf nodes without any ancestor nodes. Our sorting-based rasterizer will project voxels to image space and guarantee all voxels are in the correct order when rendering. In sum, we store individual voxels in arbitrary order without the need to maintain a more complex data structure, thanks to the flexibility provided by our rasterizer." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.39, + 0.906, + 0.495 + ], + "angle": 0, + "content": "We choose a maximum level of detail \\( L \\) (\\( = 16 \\) in this work) that defines a maximum grid resolution at \\( (2^L)^3 \\). Let \\( \\mathbf{w}_{\\mathrm{s}} \\in \\mathbb{R} \\) be the Octree size and \\( \\mathbf{w}_{\\mathrm{c}} \\in \\mathbb{R}^3 \\) be the Octree center in the world space. The voxel index \\( v = \\{i,j,k\\} \\in [0,\\dots,2^{L} - 1]^{3} \\) together with a Octree level \\( l \\in [1,L] \\) (\\( l = 0 \\) represent root node and is not used) define voxel size \\( \\mathbf{v}_{\\mathrm{s}} \\) and voxel center \\( \\mathbf{v}_{\\mathrm{c}} \\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.537, + 0.5, + 0.905, + 0.517 + ], + "angle": 0, + "content": "\\[\n\\mathbf {v} _ {\\mathrm {s}} = \\mathbf {w} _ {\\mathrm {s}} \\cdot 2 ^ {- l}, \\quad \\mathbf {v} _ {\\mathrm {c}} = \\mathbf {w} _ {\\mathrm {c}} - 0. 5 \\cdot \\mathbf {w} _ {\\mathrm {s}} + \\mathbf {v} _ {\\mathrm {s}} \\cdot v. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.522, + 0.906, + 0.582 + ], + "angle": 0, + "content": "Internally, we map the grid index to its Morton code using a well-known bit interleaving operation in the low-level CUDA implementation. Please see supplementary materials for more details." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.591, + 0.906, + 0.697 + ], + "angle": 0, + "content": "Voxel alpha from density field. Next, we present details for the geometry and appearance modeling of each voxel primitive. For scene geometry, we use eight parameters corresponding to the voxel corners to model a trilinear density field inside each voxel, denoted as \\(\\mathbf{v}_{\\mathrm{geo}} \\in \\mathbb{R}^{2 \\times 2 \\times 2}\\). Sharing corners among adjacent voxels results in a continuous density field." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.7, + 0.906, + 0.745 + ], + "angle": 0, + "content": "We also need an activation function to ensure a nonnegative density value for the raw density from \\(\\mathbf{v}_{\\mathrm{geo}}\\). For this purpose, we use exponential-linear activation:" + }, + { + "type": "equation", + "bbox": [ + 0.525, + 0.749, + 0.905, + 0.791 + ], + "angle": 0, + "content": "\\[\n\\operatorname {e x p l i n} (x) = \\left\\{ \\begin{array}{l l} x & \\text {i f} x > 1. 1 \\\\ \\exp \\left(\\frac {x}{1 . 1} - 1 + \\ln 1. 1\\right) & \\text {o t h e r w i s e} \\end{array} , \\right. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.793, + 0.905, + 0.838 + ], + "angle": 0, + "content": "which approximates softplus but is more efficient to compute. For a sharp density field inside a voxel, we apply the non-linear activation after trilinear interpolation [18, 43]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.906, + 0.901 + ], + "angle": 0, + "content": "To derive the alpha value of a voxel contributing to the alpha composition formulation in Eq. (1), we evenly sample \\(K\\) points in the ray segment of ray-voxel intersection as depicted in Fig. 2b. The equation follows the numerical" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16189" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.089, + 0.482, + 0.158 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.091, + 0.169, + 0.483, + 0.184 + ], + "angle": 0, + "content": "Figure 3. Rasterization procedure. Refer to Sec. 3.1.2 for details." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.216, + 0.448, + 0.232 + ], + "angle": 0, + "content": "integration for volume rendering as in NeRF [29, 30]:" + }, + { + "type": "equation", + "bbox": [ + 0.098, + 0.235, + 0.483, + 0.277 + ], + "angle": 0, + "content": "\\[\n\\alpha = 1 - \\exp \\left(- \\frac {l}{K} \\sum_ {k = 1} ^ {K} \\operatorname {e x p l i n} \\left(\\operatorname {i n t e r p} \\left(\\mathbf {v} _ {\\text {g e o}}, \\mathbf {q} _ {k}\\right)\\right)\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.281, + 0.483, + 0.327 + ], + "angle": 0, + "content": "where \\(l\\) is the ray segment length, \\(\\mathbf{q}_k\\) is the local voxel coordinate of the \\(k\\)-th sample point, and \\(\\mathrm{interp}(\\cdot)\\) indicates trilinear interpolation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.339, + 0.484, + 0.461 + ], + "angle": 0, + "content": "Voxel view-dependent color from SHs. To model view-dependent scene appearance, we use \\(N_{\\mathrm{shd}}\\) degree SH. For increased efficiency, we assume the SH coefficients stay constant inside a voxel, denoted as \\(\\mathbf{v}_{\\mathrm{sh}} \\in \\mathbb{R}^{(N_{\\mathrm{shd}} + 1)^2 \\times 3}\\). We approximate voxel colors as a function of the direction from the camera position \\(\\mathbf{r}_0\\) to the voxel center \\(\\mathbf{v}_{\\mathrm{c}}\\) instead of individual ray direction \\(\\mathbf{r}_{\\mathrm{d}}\\) for the sake of efficiency following 3DGS:" + }, + { + "type": "equation", + "bbox": [ + 0.114, + 0.466, + 0.483, + 0.483 + ], + "angle": 0, + "content": "\\[\nc = \\max (0, \\operatorname {s h} _ {-} \\operatorname {e v a l} (\\mathbf {v} _ {\\mathrm {s h}}, \\text {n o r m a l i z e} (\\mathbf {v} _ {\\mathrm {c}} - \\mathbf {r} _ {\\mathrm {o}}))), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.487, + 0.484, + 0.563 + ], + "angle": 0, + "content": "which is the view-dependent color intensity of the voxel contributing to the pixel composition Eq. (1). Due to the approximation, the resulting SH color of a voxel can be shared by all covered pixels in the image rather than evaluating the SH for each intersecting ray." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.575, + 0.484, + 0.68 + ], + "angle": 0, + "content": "Voxel normal. The rendering of other features or properties is similar to rendering a color image by replacing the color term \\( c \\) in Eq. (1) with the target modality like the normal of a voxel density field. For rendering efficiency, we assume the normal stays constant inside a voxel, which is represented by the analytical gradient of the density field at the voxel center:" + }, + { + "type": "equation", + "bbox": [ + 0.151, + 0.686, + 0.483, + 0.704 + ], + "angle": 0, + "content": "\\[\n\\vec {\\mathbf {n}} = \\text {n o r m a l i z e} \\left(\\nabla_ {\\mathbf {q}} \\operatorname {i n t e r p} \\left(\\mathbf {v} _ {\\mathrm {g e o}}, \\mathbf {q} _ {\\mathrm {c}}\\right)\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.707, + 0.484, + 0.784 + ], + "angle": 0, + "content": "where \\(\\mathbf{q}_{\\mathrm{c}} = (0.5, 0.5, 0.5)\\) and the closed-form equations for forward and backward passes are in the supplementary material. Similar to the SH colors, the differentiable voxel normals are computed once in pre-processing and are shared by all the intersecting rays in the image." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Voxel depth. Unlike colors and normals, the point depths to composite is efficient to compute so we do the same \\(K\\) points sampling as in the voxel alpha value in Eq. (4) for a more precise depth rendering. We manually expand and simplify the forward and backward computation for small number of \\(K\\) in our CUDA implementation. Please refer to supplementary materials for details." + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.09, + 0.586, + 0.144 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.145, + 0.585, + 0.209 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.542, + 0.213, + 0.56, + 0.224 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.596, + 0.093, + 0.657, + 0.145 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.597, + 0.146, + 0.657, + 0.207 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.66, + 0.094, + 0.744, + 0.145 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.66, + 0.146, + 0.744, + 0.207 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.662, + 0.213, + 0.679, + 0.224 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.757, + 0.095, + 0.905, + 0.209 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.822, + 0.213, + 0.839, + 0.223 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.236, + 0.907, + 0.373 + ], + "angle": 0, + "content": "Figure 4. Illustration of the rendering order. (a) In both cases, the smaller voxels should be rendered first but they will arranged behind the larger voxels if using voxel centers or the nearest corners as the sorting order. (b) We show the four types of Morton order under the 2D world. The voxel rendering order under an Octree node is depend on which world quadrant the ray direction is pointing to. (c) An toy example of sorting the Morton order encoding. All the ray directions going toward the up-right quadrant can use the sorted voxels for a correct rendering order. See Sec. 3.1.2 for details." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.396, + 0.73, + 0.411 + ], + "angle": 0, + "content": "3.1.2. Rasterization Algorithm" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.415, + 0.907, + 0.475 + ], + "angle": 0, + "content": "The overview of our sparse voxel rasterization procedure is depicted in Fig. 3. We build our sorting-based rasterizer based on the highly efficient CUDA implementation of 3DGS [20]. The procedure is detailed in the following." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.481, + 0.906, + 0.572 + ], + "angle": 0, + "content": "Projection to image space. The first step of rasterization is projecting sparse voxels onto image space and assigning the voxels to the tiles (i.e., image patches) they cover. In practice, we project the eight corner points of each voxel. The voxel is assigned to all tiles overlapped with the axis-aligned bounding box formed by the projected eight points." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.578, + 0.906, + 0.669 + ], + "angle": 0, + "content": "Pre-processing voxels. For active voxels assigned to tiles of the target view, we gather their densities \\(\\mathbf{v}_{\\mathrm{geo}}\\) from the grid points, compute view-dependent colors from their spherical harmonic coefficients with Eq. (5), and derive voxel normals by Eq. (6). The pre-processed voxel properties are shared among all pixels during rendering." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Sorting voxels. For accurate rasterization, primitive rendering order is important. Similar to the challenge in 3DGS [20], using primitive centers or their closest distance to the camera can produce incorrect ordering, producing popping artifacts [37]. We show two incorrect ordering results using naive sorting criteria in Fig. 4a. Thanks to the Octree layout (Sec. 3.1.1), we can sort by Morton order using our sparse voxels representation. As illustrated in Fig. 4b, we can follow certain types of Morton order to render the voxels under an Octree node for correct ordering. The type of Morton order to follow is solely dependent on the positive/negative signs of the ray direction (the ray origin doesn't matter). That is to say, we have eight permutations of Morton order for different ray directions in the 3D space. Finally, the generalization to multi-level voxels can" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16190" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.121 + ], + "angle": 0, + "content": "be proved by induction. An ordering example in 2D with three levels is depicted in Fig. 4c." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.122, + 0.484, + 0.198 + ], + "angle": 0, + "content": "The sorting is applied for each image tile. In case all the pixels in a tile share the same ray direction signs, we can simply sort the assigned voxels by their type of Morton order. We handle the corner case when multiple Morton orders are required in supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.203, + 0.484, + 0.384 + ], + "angle": 0, + "content": "Rendering pixels. Finally, we proceed with alpha composition, Eq. (1), to render pixels. In our case, the \\(N\\) primitives blend of a pixel-ray depends on the number of sparse voxels assigned to the tile that the pixel-ray belongs to. The computation of the alpha, color, and other geometric properties from our sparse voxels are described in Sec. 3.1.1. When rendering sparse voxels for a pixel-ray, we compute ray-aabb intersection to determine the ray segment to sample (for voxel alpha in Eq. (4)) and skip some non-intersected sparse voxels. We do early termination of the alpha composition if the transmittance of a sparse voxel is below a threshold \\(T_{i} < h_{\\mathrm{T}}\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.391, + 0.484, + 0.437 + ], + "angle": 0, + "content": "Anti-aliasing. To mitigate aliasing artifacts, we render in \\( h_{\\mathrm{ss}} \\) times higher resolution and then apply image downsampling to the target resolution with an anti-aliasing filter." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.445, + 0.433, + 0.461 + ], + "angle": 0, + "content": "3.2. Progressive Sparse Voxels Optimization" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.467, + 0.484, + 0.512 + ], + "angle": 0, + "content": "In this section, we describe the procedure to optimize a 3D scene from the input frames with known camera parameters using our SVRaster presented in Sec. 3.1." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.518, + 0.484, + 0.639 + ], + "angle": 0, + "content": "Voxel max sampling rate. We first define the maximum sampling rate \\(\\mathbf{v}_{\\mathrm{rate}}\\) of each voxel on the training images, which reflects the image region a voxel can cover. A smaller \\(\\mathbf{v}_{\\mathrm{rate}}\\) indicates that the voxel is more prone to overfitting due to less observation. We use \\(\\mathbf{v}_{\\mathrm{rate}}\\) in our voxel initialization and subdivision process described later. Given \\(N_{\\mathrm{cam}}\\) training cameras, we estimate the maximum sampling rate of a voxel as follows with visualization in Fig. 5a:" + }, + { + "type": "equation", + "bbox": [ + 0.135, + 0.642, + 0.483, + 0.676 + ], + "angle": 0, + "content": "\\[\n\\mathbf {v} _ {\\text {r a t e}} = \\max _ {i} ^ {N _ {\\text {c a m}}} \\frac {\\mathbf {v} _ {\\mathrm {s}}}{\\mathbf {v} _ {\\text {i n t e r v a l}} ^ {(i)}}, \\tag {7a}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.115, + 0.678, + 0.483, + 0.737 + ], + "angle": 0, + "content": "\\[\n\\mathbf {v} _ {\\text {i n t e r v a l}} ^ {(i)} = \\underbrace {\\left(\\mathbf {v} _ {\\mathrm {c}} - \\mathbf {r} _ {\\mathrm {o}} ^ {(i)}\\right) ^ {\\intercal} \\ell^ {(i)}} _ {\\text {V o x e l z - d i s t a n c e}} \\cdot \\underbrace {\\frac {\\tan \\left(0 . 5 \\cdot \\theta_ {\\text {f o v - x}} ^ {(i)}\\right)}{0 . 5 \\cdot W ^ {(i)}}} _ {\\text {U n i t - d i s t a n c e p i x e l s i z e}}, \\tag {7b}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.741, + 0.484, + 0.801 + ], + "angle": 0, + "content": "where \\(\\ell\\) is camera lookat vector, \\(\\theta_{\\mathrm{fov - x}}\\) is camera horizontal field of view, and \\(W\\) is image width. The sampling rate indicates the estimated number of rays along the image's horizontal axis direction that may hit the voxel." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.807, + 0.273, + 0.821 + ], + "angle": 0, + "content": "3.2.1. Scene Initialization" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Without employing an additional prior, we initialize all the parameters to constant. We start with volume density approaching zero by setting voxel raw density to a negative number \\( h_{\\mathrm{geo}} \\) so that the initial activated density \\( \\mathrm{expl}(h_{\\mathrm{geo}}) \\approx 0 \\). We set the SH coefficients to zero for" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.094, + 0.638, + 0.197 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.567, + 0.202, + 0.585, + 0.213 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.706, + 0.096, + 0.9, + 0.198 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.795, + 0.201, + 0.812, + 0.213 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.225, + 0.907, + 0.309 + ], + "angle": 0, + "content": "Figure 5. Visualization of voxel sampling rate and grid layout initialization. (a) We visualize the voxel sampling rate defined in Eq. (7). (b) We depict the foreground main region and the background region under different shell levels. In unbounded scenes, we apply different grid layout initialization strategies for foreground and background regions. See Sec. 3.2.1 for details." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.319, + 0.906, + 0.365 + ], + "angle": 0, + "content": "non-zero degrees and set the view-independent zero-degree component to yield gray color (i.e., intensity equal 0.5). We detail the Octree grid layout initialization in the following." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.372, + 0.906, + 0.448 + ], + "angle": 0, + "content": "Bounded scenes. In case the scenes or the objects to reconstruct are enclosed in a known bounded region, we simply initialize the layout as a dense grid with \\( h_{1\\mathrm{v}} \\) levels and remove voxels unobserved by any training images. The number of voxels is \\( \\leq (2^{h_{1\\mathrm{v}}})^3 \\) after initialization." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.455, + 0.907, + 0.743 + ], + "angle": 0, + "content": "Unbounded scenes. For the unbounded scenes, we first split the space into the main and the unbounded background regions, depicted in Fig. 5b, each with a different heuristic. We use the training camera positions to determine a cuboid for the main region. The cuboid center is set to the average camera positions and the radius is set to the median distance between the cuboid center and the cameras. The same as the bounded scenes, we initialize a dense grid with \\( h_{\\mathrm{lv}} \\) levels for the main region. For the background region, we allocate \\( h_{\\mathrm{out}} \\) level of background shells enclosing the main region, which means that the radius of the entire scene is \\( 2^{h_{\\mathrm{out}}} \\) of the main region. In each background shell level, we start with the coarsest voxel size, i.e., \\( 4^3 - 2^3 = 56 \\) voxels in each shell level. We then iteratively subdivide shell voxels with the highest sampling rate and remove voxels unobserved by any training cameras. The process repeats until the ratio of the number of voxels in the background and the main regions is \\( h_{\\mathrm{ratio}} \\). The number of voxels is \\( \\leq (1 + h_{\\mathrm{ratio}})(2^{h_{\\mathrm{lv}}})^3 \\) after initialization." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.753, + 0.8, + 0.768 + ], + "angle": 0, + "content": "3.2.2. Adaptive Pruning and Subdivision" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.773, + 0.906, + 0.849 + ], + "angle": 0, + "content": "The initialized grid layout only coarsely covers the entire scene that should be adaptively aligned to different levels-of-detail for the scene during the training progress. We apply the following two operations every \\( h_{\\text{every}} \\) training iterations to achieve this purpose." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Pruning. We compute the maximum blending weight \\((T_{i}\\alpha_{i})\\) from Eq. (1) of each voxel using all the training cameras. We remove voxels with maximum blending weight" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "16191" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.216, + 0.107 + ], + "angle": 0, + "content": "lower than \\(h_{\\mathrm{prune}}\\)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.113, + 0.485, + 0.174 + ], + "angle": 0, + "content": "Subdivision. Our heuristic is that a voxel with a larger training loss gradient indicates that the voxel region requires finer voxels to model. Specifically, we accumulate the subdivision priority as the following:" + }, + { + "type": "equation", + "bbox": [ + 0.161, + 0.177, + 0.483, + 0.211 + ], + "angle": 0, + "content": "\\[\n\\mathbf {v} _ {\\text {p r i o r i t y}} = \\sum_ {\\mathbf {r} \\in R} \\left\\| \\alpha (\\mathbf {r}) \\cdot \\frac {\\partial \\mathcal {L} (\\mathbf {r})}{\\partial \\alpha (\\mathbf {r})} \\right\\|, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.214, + 0.484, + 0.394 + ], + "angle": 0, + "content": "where \\( R \\) is the set of all training pixel rays throughout the \\( h_{\\text{every}} \\) iterations and \\( \\mathcal{L}(\\mathbf{r}) \\) is the training loss of the ray. The gradient is weighted by alpha values contributed from the voxel to the ray. Higher \\( \\mathbf{v}_{\\text{priority}} \\) indicates higher subdivision priority. To prevent voxels from overfitting few pixels, we set the priority to zero for voxels with maximum sampling rate lower than a sampling rate threshold \\( \\mathbf{v}_{\\text{rate}} < 2h_{\\text{rate}} \\). Finally, we select the voxels with priority above the top \\( h_{\\text{percent}} \\) percent to subdivide, i.e., the total number of voxels is increased by \\( (h_{\\text{percent}} \\cdot (8 - 1)) \\) percent. Note that we only keep the leaf nodes in the Octree layout so we remove the source voxels once they are subdivided." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.395, + 0.484, + 0.562 + ], + "angle": 0, + "content": "When voxels are pruned and subdivided, the voxel Spherical Harmonic (SH) coefficients and the grid point densities need to be updated accordingly. The SH coefficients are simply pruned together with voxels and duplicated to the subdivided children voxels. Grid point densities are slightly more complex as the eight voxel corner grid points are shared between adjacent voxels (Fig. 2). We remove a grid point only when it does not belong to any voxel corners. When subdividing, we use trilinear interpolation to compute the densities of the new grid points. The duplicated grid points are merged and their densities are averaged." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.567, + 0.303, + 0.582 + ], + "angle": 0, + "content": "3.2.3. Optimization objectives" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.585, + 0.484, + 0.631 + ], + "angle": 0, + "content": "We use MSE and SSIM as the photometric loss between the rendered and the ground truth images. The overall training objective is summarized as:" + }, + { + "type": "equation", + "bbox": [ + 0.107, + 0.64, + 0.483, + 0.675 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} = \\mathcal {L} _ {\\mathrm {m s e}} + \\lambda_ {\\mathrm {s s i m}} \\mathcal {L} _ {\\mathrm {s s i m}} \\\\ + \\lambda_ {\\mathrm {T}} \\mathcal {L} _ {\\mathrm {T}} + \\lambda_ {\\text {d i s t}} \\mathcal {L} _ {\\text {d i s t}} + \\lambda_ {\\mathrm {R}} \\mathcal {L} _ {\\mathrm {R}} + \\lambda_ {\\mathrm {t v}} \\mathcal {L} _ {\\mathrm {t v}}, \\tag {9} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.68, + 0.484, + 0.771 + ], + "angle": 0, + "content": "where \\(\\lambda\\) are the loss weights, \\(\\mathcal{L}_{\\mathrm{T}}\\) encourages the final ray transmittances to be either zero or one, \\(\\mathcal{L}_{\\mathrm{dist}}\\) is the distortion loss [2], \\(\\mathcal{L}_{\\mathrm{R}}\\) is the per-point rgb loss [43], and \\(\\mathcal{L}_{\\mathrm{tv}}\\) is the total variation loss on the sparse density grid. In mesh extraction task, we also add the depth-normal consistency loss from 2DGS [16]:" + }, + { + "type": "equation", + "bbox": [ + 0.108, + 0.78, + 0.483, + 0.797 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {m e s h}} = \\lambda_ {\\mathrm {n - d m e a n}} \\mathcal {L} _ {\\mathrm {n - d m e a n}} + \\lambda_ {\\mathrm {n - d m e d}} \\mathcal {L} _ {\\mathrm {n - d m e d}}, \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.801, + 0.484, + 0.847 + ], + "angle": 0, + "content": "where both losses encourage alignment between rendered normals and depth-derived normals from mean and median depth. More details are in the supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.852, + 0.474, + 0.868 + ], + "angle": 0, + "content": "3.2.4. Sparse-voxel TSDF Fusion and Marching Cubes" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Our sparse voxels can be seamlessly integrated with grid-based algorithms. To extract a mesh, we implement March-" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.09, + 0.611, + 0.196 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.612, + 0.091, + 0.704, + 0.197 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.705, + 0.091, + 0.796, + 0.197 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.798, + 0.091, + 0.832, + 0.197 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.833, + 0.091, + 0.865, + 0.197 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.868, + 0.091, + 0.903, + 0.197 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.205, + 0.906, + 0.261 + ], + "angle": 0, + "content": "Figure 6. A qualitative comparison with 3DGS [20]. Our result here corresponding to the base version in Tab. 1. We achieve similar visual quality comparing to 3DGS. Note that 3DGS use the coarse geometry from SfM while we do not rely on this prior." + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.275, + 0.646, + 0.385 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.646, + 0.276, + 0.776, + 0.385 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.776, + 0.276, + 0.905, + 0.385 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.393, + 0.906, + 0.463 + ], + "angle": 0, + "content": "Figure 7. Visualization of the reconstructed surface. We show the rendering images, normal maps, and the final meshes on Tanks&Temples and DTU datasets. Note that we only model the scene with density field and do not use the coarse geometry prior from SfM sparse points in this work." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.49, + 0.906, + 0.61 + ], + "angle": 0, + "content": "ing Cubes [27] to extract the triangles of an isosurface over density from the sparse voxels. The duplicated vertices from adjacent voxels are merged to produce a unique set of vertices. When the adjacent voxels belong to different Octree levels, the extracted triangle may not be connected as the density field is not continuous for voxels in different levels. Such discontinuities can be removed by simply subdividing all voxels to the finest level." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.612, + 0.906, + 0.764 + ], + "angle": 0, + "content": "Deciding the target level set for extracting the isosurface can be tricky for the density field. Instead, we implement sparse-voxel TSDF-Fusion [7, 8, 36] to compute the truncated signed distance values of the sparse grid points. We can then directly extract the surface of the zero-level set using the former sparse-voxel Marching Cubes. Future extensions of our method could directly model signed distance fields following NeuS [47] and VolSDF [53]. The sparse-voxel TSDF-Fusion can still be beneficial to directly initialize our sparse voxel representation from sensor depth." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.778, + 0.646, + 0.795 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.803, + 0.731, + 0.819 + ], + "angle": 0, + "content": "4.1. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.906, + 0.901 + ], + "angle": 0, + "content": "We use the following implementation details except stated otherwise. We start the optimization from empty space with raw density set to \\( h_{\\mathrm{geo}} = -10 \\). The initial Octree level is \\( h_{\\mathrm{lv}} = 6 \\) (i.e., \\( 64^{3} \\) voxels) for the bounded scenes and the foreground main region of the unbounded scenes. To model" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16192" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.089, + 0.482, + 0.284 + ], + "angle": 0, + "content": "
MethodMip-NeRF360 dataset
FPS↑Tr. Time↓LPIPS↓PSNR↑SSIM↑
NeRF [30]<1~day0.45123.850.605
M-NeRF [1]<1~day0.44124.040.616
M-NeRF360 [2]<1~day0.23727.690.792
Zip-NeRF [3]†<1~hrs0.18728.550.828
Plenoxels [11]<10~30m0.46323.080.626
INGP [32]~10~5m0.30225.680.705
3DGS [20]†13124m0.21627.450.815
Ours fast-rend2589m0.21026.870.804
Ours fast-train1314.5m0.19927.080.816
Ours12115m0.18527.330.822
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.094, + 0.284, + 0.242, + 0.295 + ], + "angle": 0, + "content": "† Re-evaluate on our machine." + }, + { + "type": "table_footnote", + "bbox": [ + 0.093, + 0.297, + 0.483, + 0.352 + ], + "angle": 0, + "content": "Table 1. Novel-view synthesis results comparison on MipNeRF360 dataset [2]. The results are averaged from 4 indoor scenes and 5 outdoor scenes. 3DGS uses the sparse points prior from COLMAP [41], whereas the other methods and ours do not." + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.284, + 0.483, + 0.352 + ], + "angle": 0, + "content": null + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.369, + 0.48, + 0.477 + ], + "angle": 0, + "content": "
MethodTanks&TemplesDeep Blending
FPS↑LPIPS↓PSNR↑FPS↑LPIPS↓PSNR↑
Plenoxels [11]130.37921.08110.51023.06
INGP [32]140.30521.9230.39024.96
M-NeRF360 [2]<10.25722.22<10.24529.40
3DGS [20]†1800.17623.751400.24429.60
Ours1250.14423.043660.22829.84
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.097, + 0.478, + 0.223, + 0.487 + ], + "angle": 0, + "content": "† Re-evaluate on our machine." + }, + { + "type": "table_footnote", + "bbox": [ + 0.093, + 0.489, + 0.483, + 0.531 + ], + "angle": 0, + "content": "Table 2. Comparison on Tanks&Temples [21] and Deep Blending [15] datasets. We follow 3DGS [20] to use two outdoor scenes from Tanks&Temples and two indoor scenes from Deep Blending." + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.478, + 0.483, + 0.531 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.541, + 0.223, + 0.653 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.151, + 0.654, + 0.165, + 0.662 + ], + "angle": 0, + "content": "GT" + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.541, + 0.351, + 0.653 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.278, + 0.655, + 0.297, + 0.662 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.352, + 0.541, + 0.481, + 0.653 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.405, + 0.654, + 0.428, + 0.662 + ], + "angle": 0, + "content": "3DGS" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.67, + 0.483, + 0.741 + ], + "angle": 0, + "content": "Figure 8. Failure case. On scenes with severe exposure variation of training views, our method struggles and produces clear boundary of different brightnesses and allocates many floaters. 3DGS on the other hand is less sensitive to photometric variation of GT. This explains our worse PSNR and FPS on Tanks&Temples in Tab. 2." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.75, + 0.483, + 0.902 + ], + "angle": 0, + "content": "unbounded scenes, we use \\( h_{\\mathrm{out}} = 5 \\) background shell levels with \\( h_{\\mathrm{ratio}} = 2 \\) times the number of foreground voxels (Sec. 3.2.1 and Fig. 5). The early ray stopping threshold is set to \\( h_{\\mathrm{T}} = 1\\mathrm{e} - 4 \\) and the supersampling scale is set to \\( h_{\\mathrm{ss}} = 1.5 \\). Inside each voxel, we sample \\( K = 1 \\) point for novel-view synthesis and \\( K = 3 \\) points for mesh reconstruction task. We train our model for 20K iterations. The voxels are pruned and subdivided every \\( h_{\\mathrm{every}} = 1,000 \\) iterations. The pruning threshold \\( h_{\\mathrm{prune}} \\) is linearly scale from 0.0001 to 0.05 and the subdivision percentage is \\( h_{\\mathrm{percent}} = 5 \\). More" + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.089, + 0.905, + 0.16 + ], + "angle": 0, + "content": "
MethodPeak GPU mem.↓Model size↓FPS at higher res.↑
1x2x3x
3DGS1.8GB0.7GB1316939
Ours3.9GB1.8GB12110369
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.513, + 0.161, + 0.905, + 0.202 + ], + "angle": 0, + "content": "Table 3. Memory, model size, and high-res FPS. The results are averaged on Mip-NeRF360 dataset. The standard 1x evaluation resolution have about 1–1.6M pixels per frame." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.222, + 0.771, + 0.236 + ], + "angle": 0, + "content": "details are provided in the supplement." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.237, + 0.905, + 0.342 + ], + "angle": 0, + "content": "We use the standard LPIPS [56], SSIM, and PSNR metrics to evaluate novel-view quality. We directly employ the LPIPS implementation from prior work [20], which uses [0, 1] image intensity range with VGG [42] network. For mesh accuracy, we follow the benchmarks to use F-score and chamfer distance. We use the test-set images to measure the FPS on a desktop computer with a 3090 Ti GPU." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.366, + 0.712, + 0.381 + ], + "angle": 0, + "content": "4.2. Novel-view Synthesis" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.388, + 0.905, + 0.478 + ], + "angle": 0, + "content": "In Tab. 1, we show the quantitative comparison on the MipNeRF-360 dataset. We also provide a fast-rendering variant with \\(>2\\mathrm{x}\\) FPS by setting \\(h_{\\mathrm{ss}} = 1.5\\), \\(h_{\\mathrm{prune}} = 0.15\\), and a fast-training variant with \\(>3\\mathrm{x}\\) training speedup by scaling all iteration related hyperparameter by 0.3. Both fast variants only reduce quality moderately." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.479, + 0.906, + 0.719 + ], + "angle": 0, + "content": "Our rendering speed is comparable to 3DGS on average. However, FPS varies significantly across scenes (see supplementary for detailed per-scene results). Several differences impact speed between our method and 3DGS. For instance, 3DGS sorts 32-bit floats, while we sort 48-bit Morton codes (16 levels, each with 3 bits) for primitive ordering. We also avoid the overhead of computing inverse covariance matrices from quaternions and scaling parameters. We directly intersect rays with voxels in 3D space, whereas 3DGS approximates the projection of 3D Gaussians to 2D via a linear affine transform. Finally, Gaussian distributions decay gradually to zero, while our post-activation [18, 43] voxel density field can be arbitrarily sharp. As a result, the average numbers of primitives contributing to a pixel intensity is 63 for 3DGS and 27 for us. The influence of these factors on rendering speed is scene-dependent." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.906, + 0.856 + ], + "angle": 0, + "content": "Regarding rendering quality metrics, our method achieves significantly better LPIPS than 3DGS and even surpasses Zip-NeRF. This can be explained by the qualitative comparison in Fig. 6 where our method recover finer details. The SSIM comparison results with 3DGS is more scene-dependent. SSIM comparisons with 3DGS are more scene-dependent, while PSNR tends to favor 3DGS, as it prefers smoother rendering in uncertain regions. More visual comparisons are in the supplementary." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Our method uses much more voxel primitives than Gaussian, resulting in larger model size and requires more GPU memory as shown in Tab. 3. Interestingly, our high FPS is" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16193" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.094, + 0.089, + 0.482, + 0.152 + ], + "angle": 0, + "content": "
Resolution of main2563512310243adaptivePlenoxels 6403
LPIPS↓0.4440.3260.2000.452
PSNR↑23.9825.37OOM28.0123.29
FPS↑457190171<10
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.153, + 0.483, + 0.237 + ], + "angle": 0, + "content": "Table 4. Ablation experiments of adaptive and uniform voxel sizes. The results are evaluated on the indoor bonsai and the outdoor bicycle scenes from MipNeRF-360 dataset. The resolutions at the first row indicate the final grid resolution of the main foreground cuboid. Plenxoels is the previous fully-explicit voxel grid approach. Please refer to Sec. 4.3 for more discussion." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.251, + 0.483, + 0.282 + ], + "angle": 0, + "content": "more scalable to higher resolution rendering perhaps due to the fewer contributing primitives per pixel." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.282, + 0.484, + 0.404 + ], + "angle": 0, + "content": "We compare results on two more datasets in Tab. 2. On Deep Blending [15] dataset, we achieve much better speed and quality than 3DGS. On Tanks&Temples dataset [21], despite having better LPIPS, our PSNR and FPS are worse than 3DGS. The failure results are shown in Fig. 8 where the training views have large exposure variation. As a result, our method produces a clear brightness boundary in the scene with many floaters that slow down our rendering." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.426, + 0.255, + 0.44 + ], + "angle": 0, + "content": "4.3. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.448, + 0.484, + 0.825 + ], + "angle": 0, + "content": "The rasterization process improves the rendering speed of sparse voxels, while the key to achieve high-quality results is the adaptive voxel size for different levels of detail. We show an ablation experiment in Tab. 4. Instead of subdividing voxels adaptively, the uniform voxel size variant subdivides all the voxels at certain training iterations until the grid resolution of the foreground main region reaches \\(256^{3}\\), \\(512^{3}\\), or \\(1024^{3}\\). We also align the background and the pruning setup where the details are deferred to the supplementary. As shown in the table, the quality of uniform voxel size in \\(512^{3}\\) resolution is much worse than the adaptive voxel size under similar FPS. Our machine with 24GB GPU memory fails to reach \\(1024^{3}\\) grid resolution despite the voxels being pruned and sparse. Plenoxels [11] renders sparse voxels by ray casting and sampling following NeRF. Their FPS is significantly lower than our rasterization-based sparse voxel rendering. As they use a dense 3D pointer grid to support point sampling on rays, the model scalability for the unbounded background region is thus limited. To workaround, they apply a multi-sphere image with 64 layers to model the background. Conversely, we set the scene size as \\(32\\mathrm{x}\\) larger (\\(h_{\\mathrm{out}} = 5\\)) than the foreground cuboid and direct model the entire scene by our sparse voxels. As a result, our uniform size variant with \\(512^{3}\\) foreground resolution still outperforms Plenoxels with \\(640^{3}\\) grid." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.484, + 0.902 + ], + "angle": 0, + "content": "In our rasterizer, we use direction-dependent Morton order to ensure the sorting of voxels from different levels is correct. However, like the case in 3DGS, the popping artifact by the incorrect order is not a major factor in the numerical results. Instead, we provide more visualization in" + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.089, + 0.905, + 0.219 + ], + "angle": 0, + "content": "
MethodGeom.Init.Tansk&TemplesDTU
F-score↑Tr. time↓Cf.↓Tr. time↓
NeRF [30]DensityRandom--1.49hrs
NeuS [47]SDFSphere0.38hrs0.84hrs
Voxurf [51]SDFSphere--0.7615m
Neuralangelo [24]SDFSphere0.50hrs0.61hrs
3DGS [20]GaussiansSfM pts.0.1914m1.9611m
SuGaR [14]GaussiansSfM pts.0.09~1h1.33~1h
2DGS [16]GaussiansSfM pts.0.3216m0.8011m
OursDensityConstant0.4011m0.765m
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.22, + 0.907, + 0.275 + ], + "angle": 0, + "content": "Table 5. Mesh results comparison on the Tanks&Temples [21] and DTU [17] datasets. Our method with volume density and constant field initialization achieves good accuracy-time trade-off on the unbounded-level and object-level datasets." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.288, + 0.905, + 0.319 + ], + "angle": 0, + "content": "the supplementary videos to show its effect. See more ablation studies in the supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.327, + 0.715, + 0.342 + ], + "angle": 0, + "content": "4.4. Mesh Reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.349, + 0.907, + 0.516 + ], + "angle": 0, + "content": "We extract mesh by adapting TSDF-Fusion [34] and Marching Cubes [27] into our sparse voxels. The results comparison on the large-scale Tanks&Temples and the object-scale DTU datasets is provided in Tab. 5, where our method achieves a good balance of accuracy and training time on both datasets. See qualitative examples in Fig. 7 and the supplementary materials. We find that our current method tends to produce unnecessary geometric bumps for texture details. Future work could improve surface smoothness by adapting our density field for direct signed distance field modeling or incorporating SfM sparse points prior." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.524, + 0.693, + 0.538 + ], + "angle": 0, + "content": "4.5. 2D Feature Fusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.546, + 0.906, + 0.651 + ], + "angle": 0, + "content": "We showcase some results in Fig. 1, where we fuse 2D semantic segmentation and high-dimensional foundation features into our sparse voxels using Voxel Pooling and Volume Fusion. The multi-view ensemble smoothes the inconsistent 2D predictions, while our detailed 3D geometry enables higher-resolution rendering of 2D features. See more results in the released code." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.665, + 0.634, + 0.68 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.907, + 0.901 + ], + "angle": 0, + "content": "This work presents a novel differentiable radiance field rendering system that integrates an efficient rasterizer with a multi-level sparse voxel scene representation. We reconstruct a scene from the multi-view input images by adaptively fitting the sparse voxels into different levels-of-detail. The results reveal that fully explicit voxel models, without neural networks or Gaussians, can achieve state-of-the-art comparable novel-view rendering speed and quality. The key breakthrough is overcoming the scalability constraint of voxel grids with adaptive sparse voxels and improving rendering speed through rasterization. We believe that incorporating our method with classical 3D processing algorithm [7, 8, 19, 26, 36] and 3D neural network [6, 13, 50] are promising future directions." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16194" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.171 + ], + "angle": 0, + "content": "[1] Jonathan T. Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P. Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In ICCV, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.173, + 0.484, + 0.214 + ], + "angle": 0, + "content": "[2] Jonathan T. Barron, Ben Mildenhall, Dor Verbin, Pratul P. Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In CVPR, 2022. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.216, + 0.483, + 0.257 + ], + "angle": 0, + "content": "[3] Jonathan T. Barron, Ben Mildenhall, Dor Verbin, Pratul P. Srinivasan, and Peter Hedman. Zip-nerf: Anti-aliased grid-based neural radiance fields. In ICCV, 2023. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.259, + 0.483, + 0.342 + ], + "angle": 0, + "content": "[4] Eric R. Chan, Connor Z. Lin, Matthew A. Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J. Guibas, Jonathan Tremblay, Sameh Khamis, Tero Karras, and Gordon Wetzstein. Efficient geometry-aware 3d generative adversarial networks. In CVPR, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.344, + 0.483, + 0.384 + ], + "angle": 0, + "content": "[5] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In ECCV, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.387, + 0.483, + 0.428 + ], + "angle": 0, + "content": "[6] Christopher Choy, JunYoung Gwak, and Silvio Savarese. 4d spatio-temporal convnets: Minkowski convolutional neural networks. In CVPR, 2019. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.431, + 0.483, + 0.47 + ], + "angle": 0, + "content": "[7] Brian Curless and Marc Levoy. A volumetric method for building complex models from range images. ACM TOG, 1996. 1, 2, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.474, + 0.483, + 0.529 + ], + "angle": 0, + "content": "[8] Angela Dai, Matthias Nießner, Michael Zollöfer, Shahram Izadi, and Christian Theobalt. Bundlefusion: Real-time globally consistent 3d reconstruction using on-the-fly surface re-integration. ACM TOG, 2017. 1, 2, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.531, + 0.483, + 0.572 + ], + "angle": 0, + "content": "[9] Kangle Deng, Andrew Liu, Jun-Yan Zhu, and Deva Ramanan. Depth-supervised nerf: Fewer views and faster training for free. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.574, + 0.483, + 0.643 + ], + "angle": 0, + "content": "[10] Daniel Duckworth, Peter Hedman, Christian Reiser, Peter Zhizhin, Jean-François Thibert, Mario Lucic, Richard Szeliski, and Jonathan T. Barron. SMERF: streamable memory efficient radiance fields for real-time large-scene exploration. ACM TOG, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.645, + 0.483, + 0.698 + ], + "angle": 0, + "content": "[11] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In CVPR, 2022. 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.702, + 0.483, + 0.744 + ], + "angle": 0, + "content": "[12] Qiancheng Fu, Qingshan Xu, Yew Soon Ong, and Wenbing Tao. Geo-neus: Geometry-consistent neural implicit surfaces learning for multi-view reconstruction. In _NeurIPS_, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.746, + 0.483, + 0.787 + ], + "angle": 0, + "content": "[13] Benjamin Graham and Laurens Van der Maaten. Submanifold sparse convolutional networks. arXiv preprint arXiv:1706.01307, 2017. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.789, + 0.483, + 0.842 + ], + "angle": 0, + "content": "[14] Antoine Guédon and Vincent Lepetit. Sugar: Surface-aligned gaussian splatting for efficient 3d mesh reconstruction and high-quality mesh rendering. In CVPR, 2024. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.483, + 0.899 + ], + "angle": 0, + "content": "[15] Peter Hedman, Julien Philip, True Price, Jan-Michael Frahm, George Drettakis, and Gabriel J. Brostow. Deep blending for free-viewpoint image-based rendering. ACM TOG, 2018. 7, 8" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "[16] Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2d gaussian splatting for geometrically accurate radiance fields. ACM TOG, 2024. 2, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.136, + 0.905, + 0.176 + ], + "angle": 0, + "content": "[17] Rasmus Ramsbøl Jensen, Anders Lindbjerg Dahl, George Vogiatzis, Engin Tola, and Henrik Aanæs. Large scale multiview stereopsis evaluation. In CVPR, 2014. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.179, + 0.905, + 0.219 + ], + "angle": 0, + "content": "[18] Animesh Karnewar, Tobias Ritschel, Oliver Wang, and Niloy J. Mitra. Relu fields: The little non-linearity that could. In ACM TOG, 2022. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.221, + 0.904, + 0.248 + ], + "angle": 0, + "content": "[19] Michael Kazhdan, Matthew Bolitho, and Hugues Hoppe. Poisson surface reconstruction. In SGP, 2006. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.25, + 0.905, + 0.291 + ], + "angle": 0, + "content": "[20] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM TOG, 2023. 1, 2, 3, 4, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.293, + 0.905, + 0.333 + ], + "angle": 0, + "content": "[21] Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and temples: benchmarking large-scale scene reconstruction. ACM TOG, 2017. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.335, + 0.904, + 0.362 + ], + "angle": 0, + "content": "[22] Samuli Laine and Tero Karras. Efficient sparse voxel octrees. In ACM I3D, 2010. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.364, + 0.904, + 0.404 + ], + "angle": 0, + "content": "[23] Ruilong Li, Hang Gao, Matthew Tancik, and Angjoo Kanazawa. Nerfacc: Efficient sampling accelerates nerfs. In ICCV, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.406, + 0.905, + 0.461 + ], + "angle": 0, + "content": "[24] Zhaoshuo Li, Thomas Müller, Alex Evans, Russell H. Taylor, Mathias Unberath, Ming-Yu Liu, and Chen-Hsuan Lin. Neuralangelo: High-fidelity neural surface reconstruction. In CVPR, 2023. 1, 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.463, + 0.905, + 0.503 + ], + "angle": 0, + "content": "[25] Lingjie Liu, Jiatao Gu, Kyaw Zaw Lin, Tat-Seng Chua, and Christian Theobalt. Neural sparse voxel fields. In NeurIPS, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.505, + 0.905, + 0.546 + ], + "angle": 0, + "content": "[26] Charles Loop, Qin Cai, Sergio Orts-Escolano, and Philip A Chou. A closed-form bayesian fusion equation using occupancy probabilities. In 3DV, 2016. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.548, + 0.905, + 0.588 + ], + "angle": 0, + "content": "[27] William E. Lorensen and Harvey E. Cline. Marching cubes: A high resolution 3d surface construction algorithm. In ACM TOG, 1987. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.591, + 0.905, + 0.659 + ], + "angle": 0, + "content": "[28] Alexander Mai, Peter Hedman, George Kopanas, Dor Verbin, David Futschik, Qiangeng Xu, Falko Kuester, Jonathan T. Barron, and Yinda Zhang. Ever: Exact volumetric ellipsoid rendering for real-time view synthesis. arXiv preprint arXiv:2410.01804, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.661, + 0.904, + 0.687 + ], + "angle": 0, + "content": "[29] Nelson L. Max. Optical models for direct volume rendering. IEEE TVCG, 1995. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.689, + 0.905, + 0.744 + ], + "angle": 0, + "content": "[30] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 2, 3, 4, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.746, + 0.905, + 0.801 + ], + "angle": 0, + "content": "[31] Nicolas Moenne-Loccoz, Ashkan Mirzaei, Or Perel, Ricardo de Lutio, Janick Martinez Esturo, Gavriel State, Sanja Fidler, Nicholas Sharp, and Zan Gojcic. 3d gaussian ray tracing: Fast tracing of particle scenes. ACM TOG, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.803, + 0.905, + 0.843 + ], + "angle": 0, + "content": "[32] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM TOG, 2022. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.845, + 0.905, + 0.872 + ], + "angle": 0, + "content": "[33] Ken Museth. Vdb: High-resolution sparse volumes with dynamic topology. ACM TOG, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[34] Richard A Newcombe, Shahram Izadi, Otmar Hilliges, David Molyneaux, David Kim, Andrew J Davison, Pushmeet" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16195" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.134 + ], + "angle": 0, + "content": "Kohi, Jamie Shotton, Steve Hodges, and Andrew Fitzgibbon. Kinectfusion: Real-time dense surface mapping and tracking. In IEEE ISMAR, 2011. 1, 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.135, + 0.483, + 0.218 + ], + "angle": 0, + "content": "[35] Michael Niemeyer, Fabian Manhardt, Marie-Julie Rakoto-saona, Michael Oechsle, Daniel Duckworth, Rama Gosula, Keisuke Tateno, John Bates, Dominik Kaeser, and Federico Tombari. Radsplat: Radiance field-informed gaussian splat-ting for robust real-time rendering with \\(900+\\) FPS. arXiv preprint arXiv:2403.13806, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.219, + 0.482, + 0.26 + ], + "angle": 0, + "content": "[36] Matthias Nießner, Michael Zollhöfer, Shahram Izadi, and Marc Stamminger. Real-time 3d reconstruction at scale using voxel hashing. ACM TOG, 2013. 2, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.261, + 0.482, + 0.315 + ], + "angle": 0, + "content": "[37] Lukas Radl, Michael Steiner, Mathias Parger, Alexander Weinrauch, Bernhard Kerbl, and Markus Steinberger. Stophepop: Sorted gaussian splatting for view-consistent real-time rendering. ACM TOG, 2024. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.316, + 0.482, + 0.357 + ], + "angle": 0, + "content": "[38] Mike Ranzinger, Greg Heinrich, Jan Kautz, and Pavlo Molchanov. Am-radio: Agglomerative vision foundation model reduce all domains into one. In CVPR, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.359, + 0.482, + 0.399 + ], + "angle": 0, + "content": "[39] Christian Reiser, Songyou Peng, Yiyi Liao, and Andreas Geiger. Kilonerf: Speeding up neural radiance fields with thousands of tiny mllps. In ICCV, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.401, + 0.482, + 0.468 + ], + "angle": 0, + "content": "[40] Christian Reiser, Stephan J. Garbin, Pratul P. Srinivasan, Dor Verbin, Richard Szeliski, Ben Mildenhall, Jonathan T. Barron, Peter Hedman, and Andreas Geiger. Binary opacity grids: Capturing fine geometric detail for mesh-based view synthesis. ACM TOG, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.47, + 0.482, + 0.511 + ], + "angle": 0, + "content": "[41] Johannes L. Schonberger, Enliang Zheng, Jan-Michael Frahm, and Marc Pollefeys. Pixelwise view selection for unstructured multi-view stereo. In ECCV, 2016. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.512, + 0.482, + 0.552 + ], + "angle": 0, + "content": "[42] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. In ICLR, 2015. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.554, + 0.482, + 0.594 + ], + "angle": 0, + "content": "[43] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In CVPR, 2022. 1, 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.596, + 0.482, + 0.65 + ], + "angle": 0, + "content": "[44] Cheng Sun, Guangyan Cai, Zhengqin Li, Kai Yan, Cheng Zhang, Carl S. Marshall, Jia-Bin Huang, Shuang Zhao, and Zhao Dong. Neural-pbir reconstruction of shape, material, and illumination. In ICCV, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.652, + 0.482, + 0.719 + ], + "angle": 0, + "content": "[45] Towaki Takikawa, Joey Litalien, Kangxue Yin, Karsten Kreis, Charles T. Loop, Derek Nowrouzezahrai, Alec Jacobson, Morgan McGuire, and Sanja Fidler. Neural geometric level of detail: Real-time rendering with implicit 3d shapes. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.721, + 0.482, + 0.788 + ], + "angle": 0, + "content": "[46] Haithem Turki, Vasu Agrawal, Samuel Rota Bulò, Lorenzo Porzi, Peter Kontschieder, Deva Ramanan, Michael Zollhöfer, and Christian Richardt. Hybridnerf: Efficient neural rendering via adaptive volumetric surfaces. In CVPR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.79, + 0.482, + 0.845 + ], + "angle": 0, + "content": "[47] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. In NeurIPS, 2021. 2, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.847, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[48] Yiming Wang, Qin Han, Marc Habermann, Kostas Dani-ilidis, Christian Theobalt, and Lingjie Liu. Neus2: Fast learning of neural implicit surfaces for multi-view reconstruction. In ICCV, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[49] Zian Wang, Tianchang Shen, Merlin Nimier-David, Nicholas Sharp, Jun Gao, Alexander Keller, Sanja Fidler, Thomas Müller, and Zan Gojcic. Adaptive shells for efficient neural radiance field rendering. ACM TOG, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.15, + 0.905, + 0.232 + ], + "angle": 0, + "content": "[50] Francis Williams, Jiahui Huang, Jonathan Swartz, Gergely Klár, Vijay Thakkar, Matthew Cong, Xuanchi Ren, Ruilong Li, Clement Fuji-Tsang, Sanja Fidler, Eftychios Sifakis, and Ken Museth. fvdb: A deep-learning framework for sparse, large scale, and high performance spatial intelligence. ACM TOG, 2024. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.234, + 0.905, + 0.288 + ], + "angle": 0, + "content": "[51] Tong Wu, Jiaqi Wang, Xingang Pan, Xudong Xu, Christian Theobalt, Ziwei Liu, and Dahua Lin. Voxurf: Voxel-based efficient and accurate neural surface reconstruction. In ICLR, 2023. 1, 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.291, + 0.905, + 0.345 + ], + "angle": 0, + "content": "[52] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, José M. Álvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. In NeurIPS, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.347, + 0.905, + 0.387 + ], + "angle": 0, + "content": "[53] Lior Yariv, Jiatao Gu, Yoni Kasten, and Yaron Lipman. Volume rendering of neural implicit surfaces. In NeurIPS, 2021. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.39, + 0.905, + 0.445 + ], + "angle": 0, + "content": "[54] Lior Yariv, Peter Hedman, Christian Reiser, Dor Verbin, Pratul P. Srinivasan, Richard Szeliski, Jonathan T. Barron, and Ben Mildenhall. Bakedsdf: Meshing neural sdfs for real-time view synthesis. In ACM TOG, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.447, + 0.905, + 0.488 + ], + "angle": 0, + "content": "[55] Alex Yu, Ruilong Li, Matthew Tancik, Hao Li, Ren Ng, and Angjoo Kanazawa. Plenoctrees for real-time rendering of neural radiance fields. In ICCV, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.49, + 0.905, + 0.531 + ], + "angle": 0, + "content": "[56] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 7" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.531 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16196" + } + ] +] \ No newline at end of file diff --git a/2025/Sparse Voxels Rasterization_ Real-time High-fidelity Radiance Field Rendering/a6bd5a3d-4547-4194-a227-b9c1f40e864c_origin.pdf b/2025/Sparse Voxels Rasterization_ Real-time High-fidelity Radiance Field Rendering/a6bd5a3d-4547-4194-a227-b9c1f40e864c_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..90b9d4e0898dd3bb44b332d826f0d0e5ffbcf76c --- /dev/null +++ b/2025/Sparse Voxels Rasterization_ Real-time High-fidelity Radiance Field Rendering/a6bd5a3d-4547-4194-a227-b9c1f40e864c_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:18e506995072a4486db3ab52f5a327322a981d42a9ae2ff14dddb0d888bf4e89 +size 5627470 diff --git a/2025/Sparse Voxels Rasterization_ Real-time High-fidelity Radiance Field Rendering/full.md b/2025/Sparse Voxels Rasterization_ Real-time High-fidelity Radiance Field Rendering/full.md new file mode 100644 index 0000000000000000000000000000000000000000..9a00166f378aa3bcfbc533a840f3ff0db4b96a99 --- /dev/null +++ b/2025/Sparse Voxels Rasterization_ Real-time High-fidelity Radiance Field Rendering/full.md @@ -0,0 +1,413 @@ +# Sparse Voxels Rasterization: Real-time High-fidelity Radiance Field Rendering + +Cheng Sun1 + +Jaesung Choe1 +1NVIDIA + +Charles Loop1 +2Cornell University + +Wei-Chiu Ma $^{2}$ $^{3}$ National Taiwan University + +Yu-Chiang Frank Wang $^{1,3}$ $^{1}$ + +![](images/d0938896d4565a5e05a7fe016e57bbf7bc58a98e6675b90538972db605606e47.jpg) +(a) Volume rendering by rasterizing sparse voxels. + +![](images/6d7b51f99847323ed14a071228798e98352c46ddac178ac7489b5a7a73cb0130.jpg) +(b) Novel-view rendering on Mip-NeRF360 scenes. +Figure 1. We propose SVRaster, a novel framework for multi-view reconstruction and novel view synthesis. (a) Sparse voxel representation effectively captures the volume density and radiance field of the scene, without the need for neural networks, 3D Gaussians, and sparse-points prior. (b) Using our customized sparse voxel rasterizer, we can learn the underlying 3D scene efficiently and achieve state-of-the-art performance in both rendering quality and speed. (c) Notably, lifting 2D modal to the trained sparse voxels is simple and efficient by integrating the classic Volume Fusion [7, 8, 34]. We show examples of vision foundation model feature field from RADIO [38], semantic field from Segformer [52], and signed distance field from rendered depth, making it flexible and suitable for a wide range of applications. + +![](images/a75af022143a16049ee7fb478ceb92d03cb17d8d73599b3744176bdd2326eb2c.jpg) +(c) 2D-to-3D made easy. + +# Abstract + +We propose an efficient radiance field rendering algorithm that incorporates a rasterization process on adaptive sparse voxels without neural networks or 3D Gaussians. There are two key contributions coupled with the proposed system. The first is to adaptively and explicitly allocate sparse voxels to different levels of detail within scenes, faithfully reproducing scene details with $65536^{3}$ grid resolution while achieving high rendering frame rates. Second, we customize a rasterizer for efficient adaptive sparse voxels rendering. We render voxels in the correct depth order by using ray direction-dependent Morton ordering, which avoids the well-known popping artifact found in Gaussian splatting. Our method improves the previous neural-free voxel model by over 4db PSNR and more than 10x FPS speedup, achieving state-of-the-art comparable novel-view synthesis results. Additionally, our voxel representation is seamlessly compatible with grid-based 3D processing techniques such as Volume Fusion, Voxel Pooling, and Marching Cubes, enabling a wide range of future extensions and applications. + +Code: github.com/NVlabs/svraster + +# 1. Introduction + +Gaussian splatting [20] has emerged as one of the most promising solutions for novel view synthesis. It has drawn wide attention across multiple communities due to its exceptional rendering speed and ability to capture the nuanced details of a scene. Nevertheless, Gaussian splatting in its base form has two key limitations: first, sorting Gaussians based on their centers does not guarantee proper depth ordering. It may result in popping artifacts [37] (i.e., sudden color changes for consistent geometry) when changing views. Second, the volume density of a 3D point is ill-defined when covered by multiple Gaussians. This ambiguity makes surface reconstruction non-trivial. + +On the other hand, grid representations inherently avoid these issues—both ordering and volume are well-defined. However, due to ray casting, the rendering speed of these methods [4, 5, 24, 43, 44, 51] is slower than that of Gaussian Splatting. This raises the question: is it possible to take the best of both worlds? Can we combine the efficiency of Gaussian splatting with the well-defined grid properties? + +Fortunately, the answer is yes. Our key insight is to revisit voxels—a well-established primitive with decades of + +history. Voxel representations are inherently compatible with modern graphics engines and can be rasterized efficiently. Additionally, previous work has demonstrated their ability to model scene volume densities, despite through volume ray casting. This positions voxels as the perfect bridge between rasterization and volumetric representation. However, naively adopting voxels does not work well [11]. Since a scene may consist of different levels of detail. + +With these observations in mind, we present SVRaster, a novel framework that combines the efficiency of rasterization in 3DGS with the structured volumetric approach of grid-based representations. SVRaster leverages (1) multilevel sparse voxels to model 3D scenes and (2) implements a direction-dependent Morton order encoding that facilitates the rasterization rendering from our adaptive-sized sparse voxel representation. Our rendering is free from popping artifacts because the 3D space is partitioned into disjoint voxels, and our sorting ensures the correct rendering order. Moreover, thanks to the volumetric nature and neural-free representation of SVRaster, our sparse voxel grid can be easily and seamlessly integrated with classical grid-based 3D processing algorithms. + +We show that our method is training fast, rendering fast, and achieves novel-view synthesis quality comparable to the state-of-the-art. We also integrate Volume Fusion, Voxel Pooling, and Marching Cubes operations into our adaptive sparse voxels, which showcases promising results of mesh extraction and 2D foundation feature fusion. + +# 2. Related Work + +Differentiable volume rendering has made significant strides in 3D scene reconstruction and novel-view synthesis tasks. Neural Radiance Fields (NeRF) [30] laid the foundation for this progress by optimizing a volumetric function, parameterized by multi-layer perceptrons (MLPs) to encode both geometry and appearance through differential rendering. Subsequent studies focus on accelerating speed by decomposing large MLPs into grid-based representations, with a shallow MLP typically still being employed. Several grid representations have been explored—dense grids [43], factorized grids [5], tri-planes [4], and hash grids [32]. + +In particular, our method is closely related to sparse voxel grid representations. Sparse Voxel Octrees [22] and VDB trees [33] are commonly used to manage sparse voxels and facilitate rendering. In contrast, our sparse voxels are stored in a 1D array without advanced data structures. Our rasterizer with the proposed direction-dependent Morton order ensures correct rendering order. To model volumes in sparse leaf nodes, different strategies exist, such as low-resolution 3D grid [33, 50] or implicit neural field [25, 45]. Our leaf node is a single voxel with explicit density and color parameters, like Plenoxels [11]. Regarding voxel levels, previous methods allocate voxels to a tar + +get level [11, 25, 45, 55] or use a shallow tree [33, 50]. Our voxels can adaptively fit into different levels across the entire tree depth, maximizing flexibility and scalability. + +Another scalability issue of the previous grid-based methods is that they still use some sort of dense 3D grid in the field of novel-view synthesis. For instance, dense occupancy grids [5, 23, 32, 43] for free-space skipping or dense pointer grid [11] to support sparse voxel lookup. We do not use any 3D dense data structures. + +3D Gaussian Splatting (3DGS) [20] takes a different approach by representing scenes with 3D Gaussian primitives and using rasterization for rendering, achieving state-of-the-art trade-off for quality and speed. Our work is inspired by the efficiency of using a rasterizer by 3DGS. However, 3DGS exhibits view-inconsistent popping artifacts due to inaccurate rendering order and primitive overlapping. Some recent works mitigate this artifact [31, 37], but completely resolving it significantly harms rendering speed [28]. Our method does not suffer from the popping artifact. Typically, 3DGS [20] initializes Gaussians from the triangulated points via COLMAP [41], which is especially critical for unbounded scenes. The sparse points geometry is also employed to guide the training of volumetric representations and has shown improvement in some NeRF-based methods [9, 12]. In this work, we do not use the sparse points. + +Mesh reconstruction is an important extended topic for both volumetric-based and GS-based methods. Recent 3DGS variants [14, 16] with proper regularization have shown a good speed-accuracy trade-off. However, directly extracting meshes from 3DGS remains challenging, so these methods still rely on volumetric-based post-processing algorithms to extract meshes. NeRF variants [24, 44, 47, 48, 51, 53] improve surface quality by rendering a signed distance field (SDF) instead of density field. The well-defined volume makes it straightforward to extract a mesh from the isosurface. Our sparse voxels are also a volumetric representation. Though our voxel also follows volumetric representation focusing on the density field, we still can achieve promising accuracy with minimal time. + +For efficient rendering, another line of research is to convert (or distill) the geometry and appearance of a high-quality but slow model into other efficient kind of representations such as smaller MLPs [39, 46], meshes [40, 49, 54], grids [10], or the recent Gaussians [35]. Typically, the powerful Zip-NeRF [3] is employed as the teacher model which requires hours of training per scene. Our volumetric representation could enable training-free conversion from a large implicit field, which is however out of our current scope. + +Finally, the sparse voxel is also a widely-used representation for 3D processing [7, 8, 19, 26, 34, 36] and understanding [6, 13, 50]. We mainly focus on rendering in this work while we believe that adapting our method to these techniques is a promising future direction. + +![](images/c555599a72e63f8845ff350e9ad912262f6383be220dfe60687eb61aba27006d.jpg) +(a) + +![](images/00d9280cf0e4d38f6e0fd48a2ca55da6578dc7b213db74d1c52d927959d0bf5c.jpg) +(b) +Figure 2. Sparse voxels scene representation. (Left) We allocate voxel under an Octree layout. Each voxels has its own Spherical Harmonic coefficient for view-dependent appearance. The color field is approximated as a constant inside a voxel when rendering a view for efficiency. The density field is trilaterally varied inside a voxel and is modeled by the density values on the corner grid points (i.e., the black dots $\bullet$ ) of each voxel. The grid points densities are shared between adjacent voxels. (Right) We evenly sample $K$ points inside the segment of ray-voxel intersection to compute volume integration for its alpha value contributing to the pixel ray. See Sec. 3.1.1 for details. + +# 3. Approach + +We present our approach as follows. First in Sec. 3.1, we introduce our sparse voxels scene representation and our rasterizer for rendering sparse voxels into pixels. Later in Sec. 3.2, we describe the progressive scene optimization strategy, which is designed for faithfully reconstructing the scene from multi-view images using our sparse voxels. + +# 3.1. Sparse Voxels Rasterization + +Recent neural radiance field rendering approaches, such as NeRF [30] and 3DGS [20] variants, use the following alpha composition equation to render a pixel's color C: + +$$ +\mathrm {C} = \sum_ {i = 1} ^ {N} T _ {i} \cdot \alpha_ {i} \cdot c _ {i}, \quad T _ {i} = \prod_ {j = 1} ^ {i - 1} \left(1 - \alpha_ {j}\right), \tag {1} +$$ + +where $\alpha_{i}\in \mathbb{R}_{\in [0,1]}$ and $c_{i}\in \mathbb{R}_{\in [0,1]}^{3}$ are alpha and view-dependent color at the $i$ -th sampled point or primitive on the pixel-ray. The quantity $T_{i}$ is known as the transmittance. At a high-level, the difference between each method comes down to: $i$ ) how they find the $N$ points or primitives to composite a pixel-ray and $ii$ ) how the rgb and alpha values are determined from their scene representations. + +Our sparse-voxel method follows the same principle. In Sec. 3.1.1, we provide details of our sparse voxels scene representation and how the $\alpha_{i}$ and $c_{i}$ in Eq. (1) are computed from a voxel. In Sec. 3.1.2, we present our rasterizer, that gathers the $N$ voxels to be composited for each pixel. + +# 3.1.1. Scene Representation + +We first describe the grid layout for allocating our sparse voxels and then derive the alpha value, view-dependent + +color, and other geometric properties needed for the composite rendering of a voxel. + +Sparse voxel grid. Our SVRaster constructs 3D scenes using a sparse voxel representation. We allocate voxels following an Octree space partition rule (i.e., Octree layout as illustrated in Fig. 2a) for two reasons necessary for achieving high-quality results. First, it facilitates the correct rendering order of voxels with various sizes (Sec. 3.1.2). Second, we can adaptively fit the sparse voxels to different scene level-of-details (Sec. 3.2). Note that our representation does not replicate a traditional Octree data structure with parent-child pointers or linear Octree. Specifically, we only keep voxels at the Octree leaf nodes without any ancestor nodes. Our sorting-based rasterizer will project voxels to image space and guarantee all voxels are in the correct order when rendering. In sum, we store individual voxels in arbitrary order without the need to maintain a more complex data structure, thanks to the flexibility provided by our rasterizer. + +We choose a maximum level of detail $L$ ( $= 16$ in this work) that defines a maximum grid resolution at $(2^L)^3$ . Let $\mathbf{w}_{\mathrm{s}} \in \mathbb{R}$ be the Octree size and $\mathbf{w}_{\mathrm{c}} \in \mathbb{R}^3$ be the Octree center in the world space. The voxel index $v = \{i,j,k\} \in [0,\dots,2^{L} - 1]^{3}$ together with a Octree level $l \in [1,L]$ ( $l = 0$ represent root node and is not used) define voxel size $\mathbf{v}_{\mathrm{s}}$ and voxel center $\mathbf{v}_{\mathrm{c}}$ as: + +$$ +\mathbf {v} _ {\mathrm {s}} = \mathbf {w} _ {\mathrm {s}} \cdot 2 ^ {- l}, \quad \mathbf {v} _ {\mathrm {c}} = \mathbf {w} _ {\mathrm {c}} - 0. 5 \cdot \mathbf {w} _ {\mathrm {s}} + \mathbf {v} _ {\mathrm {s}} \cdot v. \tag {2} +$$ + +Internally, we map the grid index to its Morton code using a well-known bit interleaving operation in the low-level CUDA implementation. Please see supplementary materials for more details. + +Voxel alpha from density field. Next, we present details for the geometry and appearance modeling of each voxel primitive. For scene geometry, we use eight parameters corresponding to the voxel corners to model a trilinear density field inside each voxel, denoted as $\mathbf{v}_{\mathrm{geo}} \in \mathbb{R}^{2 \times 2 \times 2}$ . Sharing corners among adjacent voxels results in a continuous density field. + +We also need an activation function to ensure a nonnegative density value for the raw density from $\mathbf{v}_{\mathrm{geo}}$ . For this purpose, we use exponential-linear activation: + +$$ +\operatorname {e x p l i n} (x) = \left\{ \begin{array}{l l} x & \text {i f} x > 1. 1 \\ \exp \left(\frac {x}{1 . 1} - 1 + \ln 1. 1\right) & \text {o t h e r w i s e} \end{array} , \right. \tag {3} +$$ + +which approximates softplus but is more efficient to compute. For a sharp density field inside a voxel, we apply the non-linear activation after trilinear interpolation [18, 43]. + +To derive the alpha value of a voxel contributing to the alpha composition formulation in Eq. (1), we evenly sample $K$ points in the ray segment of ray-voxel intersection as depicted in Fig. 2b. The equation follows the numerical + +![](images/e8373eed911a159dc8296b70ba2ae245e42822207ef5b1bb96d018aa9efd4d19.jpg) +Figure 3. Rasterization procedure. Refer to Sec. 3.1.2 for details. + +integration for volume rendering as in NeRF [29, 30]: + +$$ +\alpha = 1 - \exp \left(- \frac {l}{K} \sum_ {k = 1} ^ {K} \operatorname {e x p l i n} \left(\operatorname {i n t e r p} \left(\mathbf {v} _ {\text {g e o}}, \mathbf {q} _ {k}\right)\right)\right), \tag {4} +$$ + +where $l$ is the ray segment length, $\mathbf{q}_k$ is the local voxel coordinate of the $k$ -th sample point, and $\mathrm{interp}(\cdot)$ indicates trilinear interpolation. + +Voxel view-dependent color from SHs. To model view-dependent scene appearance, we use $N_{\mathrm{shd}}$ degree SH. For increased efficiency, we assume the SH coefficients stay constant inside a voxel, denoted as $\mathbf{v}_{\mathrm{sh}} \in \mathbb{R}^{(N_{\mathrm{shd}} + 1)^2 \times 3}$ . We approximate voxel colors as a function of the direction from the camera position $\mathbf{r}_0$ to the voxel center $\mathbf{v}_{\mathrm{c}}$ instead of individual ray direction $\mathbf{r}_{\mathrm{d}}$ for the sake of efficiency following 3DGS: + +$$ +c = \max (0, \operatorname {s h} _ {-} \operatorname {e v a l} (\mathbf {v} _ {\mathrm {s h}}, \text {n o r m a l i z e} (\mathbf {v} _ {\mathrm {c}} - \mathbf {r} _ {\mathrm {o}}))), \tag {5} +$$ + +which is the view-dependent color intensity of the voxel contributing to the pixel composition Eq. (1). Due to the approximation, the resulting SH color of a voxel can be shared by all covered pixels in the image rather than evaluating the SH for each intersecting ray. + +Voxel normal. The rendering of other features or properties is similar to rendering a color image by replacing the color term $c$ in Eq. (1) with the target modality like the normal of a voxel density field. For rendering efficiency, we assume the normal stays constant inside a voxel, which is represented by the analytical gradient of the density field at the voxel center: + +$$ +\vec {\mathbf {n}} = \text {n o r m a l i z e} \left(\nabla_ {\mathbf {q}} \operatorname {i n t e r p} \left(\mathbf {v} _ {\mathrm {g e o}}, \mathbf {q} _ {\mathrm {c}}\right)\right), \tag {6} +$$ + +where $\mathbf{q}_{\mathrm{c}} = (0.5, 0.5, 0.5)$ and the closed-form equations for forward and backward passes are in the supplementary material. Similar to the SH colors, the differentiable voxel normals are computed once in pre-processing and are shared by all the intersecting rays in the image. + +Voxel depth. Unlike colors and normals, the point depths to composite is efficient to compute so we do the same $K$ points sampling as in the voxel alpha value in Eq. (4) for a more precise depth rendering. We manually expand and simplify the forward and backward computation for small number of $K$ in our CUDA implementation. Please refer to supplementary materials for details. + +![](images/2f1341aea32eb6c2d516926feb1d92f0e0388fb88b66f5c291f65973c22bfb46.jpg) + +![](images/6de3319d7697aa716df85624bd8d2fdc2cf5a1bdea4db5821848e2bd7dc11632.jpg) +(a) + +![](images/9eccfdd93cd764748253f609b75e6d7abeca7798ced325c90a2de9445dd45b15.jpg) + +![](images/e243c7d69faf023ea5177c92999eaac74d2e0e9df41eb1d7b6ab0972582ecaaa.jpg) +(b) + +![](images/5de485304965ce85f7df2cbbe759db978491b2216adb77fe66d611bd27bf9007.jpg) + +![](images/1058bcc0c9c7bd80855967ef1de2ccc3e62c7a8b1e920cb607ab5c7b7d735964.jpg) +Figure 4. Illustration of the rendering order. (a) In both cases, the smaller voxels should be rendered first but they will arranged behind the larger voxels if using voxel centers or the nearest corners as the sorting order. (b) We show the four types of Morton order under the 2D world. The voxel rendering order under an Octree node is depend on which world quadrant the ray direction is pointing to. (c) An toy example of sorting the Morton order encoding. All the ray directions going toward the up-right quadrant can use the sorted voxels for a correct rendering order. See Sec. 3.1.2 for details. + +![](images/2583482f33309c08029722bffad55bc8cd414c4476a363bdae81a4358d68e317.jpg) +(c) + +# 3.1.2. Rasterization Algorithm + +The overview of our sparse voxel rasterization procedure is depicted in Fig. 3. We build our sorting-based rasterizer based on the highly efficient CUDA implementation of 3DGS [20]. The procedure is detailed in the following. + +Projection to image space. The first step of rasterization is projecting sparse voxels onto image space and assigning the voxels to the tiles (i.e., image patches) they cover. In practice, we project the eight corner points of each voxel. The voxel is assigned to all tiles overlapped with the axis-aligned bounding box formed by the projected eight points. + +Pre-processing voxels. For active voxels assigned to tiles of the target view, we gather their densities $\mathbf{v}_{\mathrm{geo}}$ from the grid points, compute view-dependent colors from their spherical harmonic coefficients with Eq. (5), and derive voxel normals by Eq. (6). The pre-processed voxel properties are shared among all pixels during rendering. + +Sorting voxels. For accurate rasterization, primitive rendering order is important. Similar to the challenge in 3DGS [20], using primitive centers or their closest distance to the camera can produce incorrect ordering, producing popping artifacts [37]. We show two incorrect ordering results using naive sorting criteria in Fig. 4a. Thanks to the Octree layout (Sec. 3.1.1), we can sort by Morton order using our sparse voxels representation. As illustrated in Fig. 4b, we can follow certain types of Morton order to render the voxels under an Octree node for correct ordering. The type of Morton order to follow is solely dependent on the positive/negative signs of the ray direction (the ray origin doesn't matter). That is to say, we have eight permutations of Morton order for different ray directions in the 3D space. Finally, the generalization to multi-level voxels can + +be proved by induction. An ordering example in 2D with three levels is depicted in Fig. 4c. + +The sorting is applied for each image tile. In case all the pixels in a tile share the same ray direction signs, we can simply sort the assigned voxels by their type of Morton order. We handle the corner case when multiple Morton orders are required in supplementary materials. + +Rendering pixels. Finally, we proceed with alpha composition, Eq. (1), to render pixels. In our case, the $N$ primitives blend of a pixel-ray depends on the number of sparse voxels assigned to the tile that the pixel-ray belongs to. The computation of the alpha, color, and other geometric properties from our sparse voxels are described in Sec. 3.1.1. When rendering sparse voxels for a pixel-ray, we compute ray-aabb intersection to determine the ray segment to sample (for voxel alpha in Eq. (4)) and skip some non-intersected sparse voxels. We do early termination of the alpha composition if the transmittance of a sparse voxel is below a threshold $T_{i} < h_{\mathrm{T}}$ . + +Anti-aliasing. To mitigate aliasing artifacts, we render in $h_{\mathrm{ss}}$ times higher resolution and then apply image downsampling to the target resolution with an anti-aliasing filter. + +# 3.2. Progressive Sparse Voxels Optimization + +In this section, we describe the procedure to optimize a 3D scene from the input frames with known camera parameters using our SVRaster presented in Sec. 3.1. + +Voxel max sampling rate. We first define the maximum sampling rate $\mathbf{v}_{\mathrm{rate}}$ of each voxel on the training images, which reflects the image region a voxel can cover. A smaller $\mathbf{v}_{\mathrm{rate}}$ indicates that the voxel is more prone to overfitting due to less observation. We use $\mathbf{v}_{\mathrm{rate}}$ in our voxel initialization and subdivision process described later. Given $N_{\mathrm{cam}}$ training cameras, we estimate the maximum sampling rate of a voxel as follows with visualization in Fig. 5a: + +$$ +\mathbf {v} _ {\text {r a t e}} = \max _ {i} ^ {N _ {\text {c a m}}} \frac {\mathbf {v} _ {\mathrm {s}}}{\mathbf {v} _ {\text {i n t e r v a l}} ^ {(i)}}, \tag {7a} +$$ + +$$ +\mathbf {v} _ {\text {i n t e r v a l}} ^ {(i)} = \underbrace {\left(\mathbf {v} _ {\mathrm {c}} - \mathbf {r} _ {\mathrm {o}} ^ {(i)}\right) ^ {\intercal} \ell^ {(i)}} _ {\text {V o x e l z - d i s t a n c e}} \cdot \underbrace {\frac {\tan \left(0 . 5 \cdot \theta_ {\text {f o v - x}} ^ {(i)}\right)}{0 . 5 \cdot W ^ {(i)}}} _ {\text {U n i t - d i s t a n c e p i x e l s i z e}}, \tag {7b} +$$ + +where $\ell$ is camera lookat vector, $\theta_{\mathrm{fov - x}}$ is camera horizontal field of view, and $W$ is image width. The sampling rate indicates the estimated number of rays along the image's horizontal axis direction that may hit the voxel. + +# 3.2.1. Scene Initialization + +Without employing an additional prior, we initialize all the parameters to constant. We start with volume density approaching zero by setting voxel raw density to a negative number $h_{\mathrm{geo}}$ so that the initial activated density $\mathrm{expl}(h_{\mathrm{geo}}) \approx 0$ . We set the SH coefficients to zero for + +![](images/d27e95ea3cd38fe76c3b5b7b47305570c9c3cd3a9d55ca912138c2e92baed9d5.jpg) +(a) + +![](images/ee5739594973dabe5c4f9c288e4650a6b06d8665c3264b832b8191c9de6539c7.jpg) +(b) +Figure 5. Visualization of voxel sampling rate and grid layout initialization. (a) We visualize the voxel sampling rate defined in Eq. (7). (b) We depict the foreground main region and the background region under different shell levels. In unbounded scenes, we apply different grid layout initialization strategies for foreground and background regions. See Sec. 3.2.1 for details. + +non-zero degrees and set the view-independent zero-degree component to yield gray color (i.e., intensity equal 0.5). We detail the Octree grid layout initialization in the following. + +Bounded scenes. In case the scenes or the objects to reconstruct are enclosed in a known bounded region, we simply initialize the layout as a dense grid with $h_{1\mathrm{v}}$ levels and remove voxels unobserved by any training images. The number of voxels is $\leq (2^{h_{1\mathrm{v}}})^3$ after initialization. + +Unbounded scenes. For the unbounded scenes, we first split the space into the main and the unbounded background regions, depicted in Fig. 5b, each with a different heuristic. We use the training camera positions to determine a cuboid for the main region. The cuboid center is set to the average camera positions and the radius is set to the median distance between the cuboid center and the cameras. The same as the bounded scenes, we initialize a dense grid with $h_{\mathrm{lv}}$ levels for the main region. For the background region, we allocate $h_{\mathrm{out}}$ level of background shells enclosing the main region, which means that the radius of the entire scene is $2^{h_{\mathrm{out}}}$ of the main region. In each background shell level, we start with the coarsest voxel size, i.e., $4^3 - 2^3 = 56$ voxels in each shell level. We then iteratively subdivide shell voxels with the highest sampling rate and remove voxels unobserved by any training cameras. The process repeats until the ratio of the number of voxels in the background and the main regions is $h_{\mathrm{ratio}}$ . The number of voxels is $\leq (1 + h_{\mathrm{ratio}})(2^{h_{\mathrm{lv}}})^3$ after initialization. + +# 3.2.2. Adaptive Pruning and Subdivision + +The initialized grid layout only coarsely covers the entire scene that should be adaptively aligned to different levels-of-detail for the scene during the training progress. We apply the following two operations every $h_{\text{every}}$ training iterations to achieve this purpose. + +Pruning. We compute the maximum blending weight $(T_{i}\alpha_{i})$ from Eq. (1) of each voxel using all the training cameras. We remove voxels with maximum blending weight + +lower than $h_{\mathrm{prune}}$ + +Subdivision. Our heuristic is that a voxel with a larger training loss gradient indicates that the voxel region requires finer voxels to model. Specifically, we accumulate the subdivision priority as the following: + +$$ +\mathbf {v} _ {\text {p r i o r i t y}} = \sum_ {\mathbf {r} \in R} \left\| \alpha (\mathbf {r}) \cdot \frac {\partial \mathcal {L} (\mathbf {r})}{\partial \alpha (\mathbf {r})} \right\|, \tag {8} +$$ + +where $R$ is the set of all training pixel rays throughout the $h_{\text{every}}$ iterations and $\mathcal{L}(\mathbf{r})$ is the training loss of the ray. The gradient is weighted by alpha values contributed from the voxel to the ray. Higher $\mathbf{v}_{\text{priority}}$ indicates higher subdivision priority. To prevent voxels from overfitting few pixels, we set the priority to zero for voxels with maximum sampling rate lower than a sampling rate threshold $\mathbf{v}_{\text{rate}} < 2h_{\text{rate}}$ . Finally, we select the voxels with priority above the top $h_{\text{percent}}$ percent to subdivide, i.e., the total number of voxels is increased by $(h_{\text{percent}} \cdot (8 - 1))$ percent. Note that we only keep the leaf nodes in the Octree layout so we remove the source voxels once they are subdivided. + +When voxels are pruned and subdivided, the voxel Spherical Harmonic (SH) coefficients and the grid point densities need to be updated accordingly. The SH coefficients are simply pruned together with voxels and duplicated to the subdivided children voxels. Grid point densities are slightly more complex as the eight voxel corner grid points are shared between adjacent voxels (Fig. 2). We remove a grid point only when it does not belong to any voxel corners. When subdividing, we use trilinear interpolation to compute the densities of the new grid points. The duplicated grid points are merged and their densities are averaged. + +# 3.2.3. Optimization objectives + +We use MSE and SSIM as the photometric loss between the rendered and the ground truth images. The overall training objective is summarized as: + +$$ +\begin{array}{l} \mathcal {L} = \mathcal {L} _ {\mathrm {m s e}} + \lambda_ {\mathrm {s s i m}} \mathcal {L} _ {\mathrm {s s i m}} \\ + \lambda_ {\mathrm {T}} \mathcal {L} _ {\mathrm {T}} + \lambda_ {\text {d i s t}} \mathcal {L} _ {\text {d i s t}} + \lambda_ {\mathrm {R}} \mathcal {L} _ {\mathrm {R}} + \lambda_ {\mathrm {t v}} \mathcal {L} _ {\mathrm {t v}}, \tag {9} \\ \end{array} +$$ + +where $\lambda$ are the loss weights, $\mathcal{L}_{\mathrm{T}}$ encourages the final ray transmittances to be either zero or one, $\mathcal{L}_{\mathrm{dist}}$ is the distortion loss [2], $\mathcal{L}_{\mathrm{R}}$ is the per-point rgb loss [43], and $\mathcal{L}_{\mathrm{tv}}$ is the total variation loss on the sparse density grid. In mesh extraction task, we also add the depth-normal consistency loss from 2DGS [16]: + +$$ +\mathcal {L} _ {\text {m e s h}} = \lambda_ {\mathrm {n - d m e a n}} \mathcal {L} _ {\mathrm {n - d m e a n}} + \lambda_ {\mathrm {n - d m e d}} \mathcal {L} _ {\mathrm {n - d m e d}}, \tag {10} +$$ + +where both losses encourage alignment between rendered normals and depth-derived normals from mean and median depth. More details are in the supplementary materials. + +# 3.2.4. Sparse-voxel TSDF Fusion and Marching Cubes + +Our sparse voxels can be seamlessly integrated with grid-based algorithms. To extract a mesh, we implement March- + +![](images/e8d58f37de38ce27a4313b2c32656c80caa7f7355eac5dd81e3e7120f3af224a.jpg) +Figure 6. A qualitative comparison with 3DGS [20]. Our result here corresponding to the base version in Tab. 1. We achieve similar visual quality comparing to 3DGS. Note that 3DGS use the coarse geometry from SfM while we do not rely on this prior. + +![](images/f2cdbeb90dd4abee73ab8b8c63a6013e6e04bbe3fc68c6c5845dbcd3b53d9d7f.jpg) + +![](images/851623337ba4d253a469d16b9ea9ae8fe9f3869f701e1b6b534e627b8293da51.jpg) + +![](images/caa7d463f4514c280dbd6282de13a96454543cbace5d77917e5c84f7ec808758.jpg) + +![](images/0ec9cc628271902213e2cb573f03effa40ef87426d03421724d02ccbd8e96e5c.jpg) + +![](images/e64aaead89acfcf17a5539b96d4c9ff713b8c2a4ae898c578e48a0290765ace3.jpg) + +![](images/a7c38bd876137a04723797d560bb11eca4c0b1c529af8ee98bd7f065f00553e6.jpg) +Figure 7. Visualization of the reconstructed surface. We show the rendering images, normal maps, and the final meshes on Tanks&Temples and DTU datasets. Note that we only model the scene with density field and do not use the coarse geometry prior from SfM sparse points in this work. + +![](images/f06a010990d5fb1e31ed80fe85bfdc08b7cd83bb1863c8f34e06f5872d098a1f.jpg) + +![](images/6823587515660a834d4214405ce47f47d4325b980720fb75ff8393ff3084ee6f.jpg) + +ing Cubes [27] to extract the triangles of an isosurface over density from the sparse voxels. The duplicated vertices from adjacent voxels are merged to produce a unique set of vertices. When the adjacent voxels belong to different Octree levels, the extracted triangle may not be connected as the density field is not continuous for voxels in different levels. Such discontinuities can be removed by simply subdividing all voxels to the finest level. + +Deciding the target level set for extracting the isosurface can be tricky for the density field. Instead, we implement sparse-voxel TSDF-Fusion [7, 8, 36] to compute the truncated signed distance values of the sparse grid points. We can then directly extract the surface of the zero-level set using the former sparse-voxel Marching Cubes. Future extensions of our method could directly model signed distance fields following NeuS [47] and VolSDF [53]. The sparse-voxel TSDF-Fusion can still be beneficial to directly initialize our sparse voxel representation from sensor depth. + +# 4. Experiments + +# 4.1. Implementation Details + +We use the following implementation details except stated otherwise. We start the optimization from empty space with raw density set to $h_{\mathrm{geo}} = -10$ . The initial Octree level is $h_{\mathrm{lv}} = 6$ (i.e., $64^{3}$ voxels) for the bounded scenes and the foreground main region of the unbounded scenes. To model + +
MethodMip-NeRF360 dataset
FPS↑Tr. Time↓LPIPS↓PSNR↑SSIM↑
NeRF [30]<1~day0.45123.850.605
M-NeRF [1]<1~day0.44124.040.616
M-NeRF360 [2]<1~day0.23727.690.792
Zip-NeRF [3]†<1~hrs0.18728.550.828
Plenoxels [11]<10~30m0.46323.080.626
INGP [32]~10~5m0.30225.680.705
3DGS [20]†13124m0.21627.450.815
Ours fast-rend2589m0.21026.870.804
Ours fast-train1314.5m0.19927.080.816
Ours12115m0.18527.330.822
+ +† Re-evaluate on our machine. +Table 1. Novel-view synthesis results comparison on MipNeRF360 dataset [2]. The results are averaged from 4 indoor scenes and 5 outdoor scenes. 3DGS uses the sparse points prior from COLMAP [41], whereas the other methods and ours do not. + +
MethodTanks&TemplesDeep Blending
FPS↑LPIPS↓PSNR↑FPS↑LPIPS↓PSNR↑
Plenoxels [11]130.37921.08110.51023.06
INGP [32]140.30521.9230.39024.96
M-NeRF360 [2]<10.25722.22<10.24529.40
3DGS [20]†1800.17623.751400.24429.60
Ours1250.14423.043660.22829.84
+ +† Re-evaluate on our machine. +Table 2. Comparison on Tanks&Temples [21] and Deep Blending [15] datasets. We follow 3DGS [20] to use two outdoor scenes from Tanks&Temples and two indoor scenes from Deep Blending. + +![](images/5e9d8977242621a3b6547198a855d2e243fb3c60f0ed9a2bf96b41490f0b22d1.jpg) +GT +Figure 8. Failure case. On scenes with severe exposure variation of training views, our method struggles and produces clear boundary of different brightnesses and allocates many floaters. 3DGS on the other hand is less sensitive to photometric variation of GT. This explains our worse PSNR and FPS on Tanks&Temples in Tab. 2. + +![](images/222bcc7c2e59533a270851ebb9c140da2e79d6cf75f99240d4d09ad3a0c54e15.jpg) +Ours + +![](images/130f561e738b9d0c977e30ff2cc1e61da18ccba322897372a54334905d12ba96.jpg) +3DGS + +unbounded scenes, we use $h_{\mathrm{out}} = 5$ background shell levels with $h_{\mathrm{ratio}} = 2$ times the number of foreground voxels (Sec. 3.2.1 and Fig. 5). The early ray stopping threshold is set to $h_{\mathrm{T}} = 1\mathrm{e} - 4$ and the supersampling scale is set to $h_{\mathrm{ss}} = 1.5$ . Inside each voxel, we sample $K = 1$ point for novel-view synthesis and $K = 3$ points for mesh reconstruction task. We train our model for 20K iterations. The voxels are pruned and subdivided every $h_{\mathrm{every}} = 1,000$ iterations. The pruning threshold $h_{\mathrm{prune}}$ is linearly scale from 0.0001 to 0.05 and the subdivision percentage is $h_{\mathrm{percent}} = 5$ . More + +
MethodPeak GPU mem.↓Model size↓FPS at higher res.↑
1x2x3x
3DGS1.8GB0.7GB1316939
Ours3.9GB1.8GB12110369
+ +Table 3. Memory, model size, and high-res FPS. The results are averaged on Mip-NeRF360 dataset. The standard 1x evaluation resolution have about 1–1.6M pixels per frame. + +details are provided in the supplement. + +We use the standard LPIPS [56], SSIM, and PSNR metrics to evaluate novel-view quality. We directly employ the LPIPS implementation from prior work [20], which uses [0, 1] image intensity range with VGG [42] network. For mesh accuracy, we follow the benchmarks to use F-score and chamfer distance. We use the test-set images to measure the FPS on a desktop computer with a 3090 Ti GPU. + +# 4.2. Novel-view Synthesis + +In Tab. 1, we show the quantitative comparison on the MipNeRF-360 dataset. We also provide a fast-rendering variant with $>2\mathrm{x}$ FPS by setting $h_{\mathrm{ss}} = 1.5$ , $h_{\mathrm{prune}} = 0.15$ , and a fast-training variant with $>3\mathrm{x}$ training speedup by scaling all iteration related hyperparameter by 0.3. Both fast variants only reduce quality moderately. + +Our rendering speed is comparable to 3DGS on average. However, FPS varies significantly across scenes (see supplementary for detailed per-scene results). Several differences impact speed between our method and 3DGS. For instance, 3DGS sorts 32-bit floats, while we sort 48-bit Morton codes (16 levels, each with 3 bits) for primitive ordering. We also avoid the overhead of computing inverse covariance matrices from quaternions and scaling parameters. We directly intersect rays with voxels in 3D space, whereas 3DGS approximates the projection of 3D Gaussians to 2D via a linear affine transform. Finally, Gaussian distributions decay gradually to zero, while our post-activation [18, 43] voxel density field can be arbitrarily sharp. As a result, the average numbers of primitives contributing to a pixel intensity is 63 for 3DGS and 27 for us. The influence of these factors on rendering speed is scene-dependent. + +Regarding rendering quality metrics, our method achieves significantly better LPIPS than 3DGS and even surpasses Zip-NeRF. This can be explained by the qualitative comparison in Fig. 6 where our method recover finer details. The SSIM comparison results with 3DGS is more scene-dependent. SSIM comparisons with 3DGS are more scene-dependent, while PSNR tends to favor 3DGS, as it prefers smoother rendering in uncertain regions. More visual comparisons are in the supplementary. + +Our method uses much more voxel primitives than Gaussian, resulting in larger model size and requires more GPU memory as shown in Tab. 3. Interestingly, our high FPS is + +
Resolution of main2563512310243adaptivePlenoxels 6403
LPIPS↓0.4440.3260.2000.452
PSNR↑23.9825.37OOM28.0123.29
FPS↑457190171<10
+ +more scalable to higher resolution rendering perhaps due to the fewer contributing primitives per pixel. + +We compare results on two more datasets in Tab. 2. On Deep Blending [15] dataset, we achieve much better speed and quality than 3DGS. On Tanks&Temples dataset [21], despite having better LPIPS, our PSNR and FPS are worse than 3DGS. The failure results are shown in Fig. 8 where the training views have large exposure variation. As a result, our method produces a clear brightness boundary in the scene with many floaters that slow down our rendering. + +# 4.3. Ablation Studies + +The rasterization process improves the rendering speed of sparse voxels, while the key to achieve high-quality results is the adaptive voxel size for different levels of detail. We show an ablation experiment in Tab. 4. Instead of subdividing voxels adaptively, the uniform voxel size variant subdivides all the voxels at certain training iterations until the grid resolution of the foreground main region reaches $256^{3}$ , $512^{3}$ , or $1024^{3}$ . We also align the background and the pruning setup where the details are deferred to the supplementary. As shown in the table, the quality of uniform voxel size in $512^{3}$ resolution is much worse than the adaptive voxel size under similar FPS. Our machine with 24GB GPU memory fails to reach $1024^{3}$ grid resolution despite the voxels being pruned and sparse. Plenoxels [11] renders sparse voxels by ray casting and sampling following NeRF. Their FPS is significantly lower than our rasterization-based sparse voxel rendering. As they use a dense 3D pointer grid to support point sampling on rays, the model scalability for the unbounded background region is thus limited. To workaround, they apply a multi-sphere image with 64 layers to model the background. Conversely, we set the scene size as $32\mathrm{x}$ larger ( $h_{\mathrm{out}} = 5$ ) than the foreground cuboid and direct model the entire scene by our sparse voxels. As a result, our uniform size variant with $512^{3}$ foreground resolution still outperforms Plenoxels with $640^{3}$ grid. + +In our rasterizer, we use direction-dependent Morton order to ensure the sorting of voxels from different levels is correct. However, like the case in 3DGS, the popping artifact by the incorrect order is not a major factor in the numerical results. Instead, we provide more visualization in + +Table 4. Ablation experiments of adaptive and uniform voxel sizes. The results are evaluated on the indoor bonsai and the outdoor bicycle scenes from MipNeRF-360 dataset. The resolutions at the first row indicate the final grid resolution of the main foreground cuboid. Plenxoels is the previous fully-explicit voxel grid approach. Please refer to Sec. 4.3 for more discussion. + +
MethodGeom.Init.Tansk&TemplesDTU
F-score↑Tr. time↓Cf.↓Tr. time↓
NeRF [30]DensityRandom--1.49hrs
NeuS [47]SDFSphere0.38hrs0.84hrs
Voxurf [51]SDFSphere--0.7615m
Neuralangelo [24]SDFSphere0.50hrs0.61hrs
3DGS [20]GaussiansSfM pts.0.1914m1.9611m
SuGaR [14]GaussiansSfM pts.0.09~1h1.33~1h
2DGS [16]GaussiansSfM pts.0.3216m0.8011m
OursDensityConstant0.4011m0.765m
+ +Table 5. Mesh results comparison on the Tanks&Temples [21] and DTU [17] datasets. Our method with volume density and constant field initialization achieves good accuracy-time trade-off on the unbounded-level and object-level datasets. + +the supplementary videos to show its effect. See more ablation studies in the supplementary materials. + +# 4.4. Mesh Reconstruction + +We extract mesh by adapting TSDF-Fusion [34] and Marching Cubes [27] into our sparse voxels. The results comparison on the large-scale Tanks&Temples and the object-scale DTU datasets is provided in Tab. 5, where our method achieves a good balance of accuracy and training time on both datasets. See qualitative examples in Fig. 7 and the supplementary materials. We find that our current method tends to produce unnecessary geometric bumps for texture details. Future work could improve surface smoothness by adapting our density field for direct signed distance field modeling or incorporating SfM sparse points prior. + +# 4.5. 2D Feature Fusion + +We showcase some results in Fig. 1, where we fuse 2D semantic segmentation and high-dimensional foundation features into our sparse voxels using Voxel Pooling and Volume Fusion. The multi-view ensemble smoothes the inconsistent 2D predictions, while our detailed 3D geometry enables higher-resolution rendering of 2D features. See more results in the released code. + +# 5. Conclusion + +This work presents a novel differentiable radiance field rendering system that integrates an efficient rasterizer with a multi-level sparse voxel scene representation. We reconstruct a scene from the multi-view input images by adaptively fitting the sparse voxels into different levels-of-detail. The results reveal that fully explicit voxel models, without neural networks or Gaussians, can achieve state-of-the-art comparable novel-view rendering speed and quality. The key breakthrough is overcoming the scalability constraint of voxel grids with adaptive sparse voxels and improving rendering speed through rasterization. We believe that incorporating our method with classical 3D processing algorithm [7, 8, 19, 26, 36] and 3D neural network [6, 13, 50] are promising future directions. + +# References + +[1] Jonathan T. Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P. Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In ICCV, 2021. 7 +[2] Jonathan T. Barron, Ben Mildenhall, Dor Verbin, Pratul P. Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In CVPR, 2022. 6, 7 +[3] Jonathan T. Barron, Ben Mildenhall, Dor Verbin, Pratul P. Srinivasan, and Peter Hedman. Zip-nerf: Anti-aliased grid-based neural radiance fields. In ICCV, 2023. 2, 7 +[4] Eric R. Chan, Connor Z. Lin, Matthew A. Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J. Guibas, Jonathan Tremblay, Sameh Khamis, Tero Karras, and Gordon Wetzstein. Efficient geometry-aware 3d generative adversarial networks. In CVPR, 2022. 1, 2 +[5] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In ECCV, 2022. 1, 2 +[6] Christopher Choy, JunYoung Gwak, and Silvio Savarese. 4d spatio-temporal convnets: Minkowski convolutional neural networks. In CVPR, 2019. 2, 8 +[7] Brian Curless and Marc Levoy. A volumetric method for building complex models from range images. ACM TOG, 1996. 1, 2, 6, 8 +[8] Angela Dai, Matthias Nießner, Michael Zollöfer, Shahram Izadi, and Christian Theobalt. Bundlefusion: Real-time globally consistent 3d reconstruction using on-the-fly surface re-integration. ACM TOG, 2017. 1, 2, 6, 8 +[9] Kangle Deng, Andrew Liu, Jun-Yan Zhu, and Deva Ramanan. Depth-supervised nerf: Fewer views and faster training for free. In CVPR, 2022. 2 +[10] Daniel Duckworth, Peter Hedman, Christian Reiser, Peter Zhizhin, Jean-François Thibert, Mario Lucic, Richard Szeliski, and Jonathan T. Barron. SMERF: streamable memory efficient radiance fields for real-time large-scene exploration. ACM TOG, 2024. 2 +[11] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In CVPR, 2022. 2, 7, 8 +[12] Qiancheng Fu, Qingshan Xu, Yew Soon Ong, and Wenbing Tao. Geo-neus: Geometry-consistent neural implicit surfaces learning for multi-view reconstruction. In _NeurIPS_, 2022. 2 +[13] Benjamin Graham and Laurens Van der Maaten. Submanifold sparse convolutional networks. arXiv preprint arXiv:1706.01307, 2017. 2, 8 +[14] Antoine Guédon and Vincent Lepetit. Sugar: Surface-aligned gaussian splatting for efficient 3d mesh reconstruction and high-quality mesh rendering. In CVPR, 2024. 2, 8 +[15] Peter Hedman, Julien Philip, True Price, Jan-Michael Frahm, George Drettakis, and Gabriel J. Brostow. Deep blending for free-viewpoint image-based rendering. ACM TOG, 2018. 7, 8 + +[16] Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2d gaussian splatting for geometrically accurate radiance fields. ACM TOG, 2024. 2, 6, 8 +[17] Rasmus Ramsbøl Jensen, Anders Lindbjerg Dahl, George Vogiatzis, Engin Tola, and Henrik Aanæs. Large scale multiview stereopsis evaluation. In CVPR, 2014. 8 +[18] Animesh Karnewar, Tobias Ritschel, Oliver Wang, and Niloy J. Mitra. Relu fields: The little non-linearity that could. In ACM TOG, 2022. 3, 7 +[19] Michael Kazhdan, Matthew Bolitho, and Hugues Hoppe. Poisson surface reconstruction. In SGP, 2006. 2, 8 +[20] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM TOG, 2023. 1, 2, 3, 4, 6, 7, 8 +[21] Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and temples: benchmarking large-scale scene reconstruction. ACM TOG, 2017. 7, 8 +[22] Samuli Laine and Tero Karras. Efficient sparse voxel octrees. In ACM I3D, 2010. 2 +[23] Ruilong Li, Hang Gao, Matthew Tancik, and Angjoo Kanazawa. Nerfacc: Efficient sampling accelerates nerfs. In ICCV, 2023. 2 +[24] Zhaoshuo Li, Thomas Müller, Alex Evans, Russell H. Taylor, Mathias Unberath, Ming-Yu Liu, and Chen-Hsuan Lin. Neuralangelo: High-fidelity neural surface reconstruction. In CVPR, 2023. 1, 2, 8 +[25] Lingjie Liu, Jiatao Gu, Kyaw Zaw Lin, Tat-Seng Chua, and Christian Theobalt. Neural sparse voxel fields. In NeurIPS, 2020. 2 +[26] Charles Loop, Qin Cai, Sergio Orts-Escolano, and Philip A Chou. A closed-form bayesian fusion equation using occupancy probabilities. In 3DV, 2016. 2, 8 +[27] William E. Lorensen and Harvey E. Cline. Marching cubes: A high resolution 3d surface construction algorithm. In ACM TOG, 1987. 6, 8 +[28] Alexander Mai, Peter Hedman, George Kopanas, Dor Verbin, David Futschik, Qiangeng Xu, Falko Kuester, Jonathan T. Barron, and Yinda Zhang. Ever: Exact volumetric ellipsoid rendering for real-time view synthesis. arXiv preprint arXiv:2410.01804, 2024. 2 +[29] Nelson L. Max. Optical models for direct volume rendering. IEEE TVCG, 1995. 4 +[30] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 2, 3, 4, 7, 8 +[31] Nicolas Moenne-Loccoz, Ashkan Mirzaei, Or Perel, Ricardo de Lutio, Janick Martinez Esturo, Gavriel State, Sanja Fidler, Nicholas Sharp, and Zan Gojcic. 3d gaussian ray tracing: Fast tracing of particle scenes. ACM TOG, 2024. 2 +[32] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM TOG, 2022. 2, 7 +[33] Ken Museth. Vdb: High-resolution sparse volumes with dynamic topology. ACM TOG, 2013. 2 +[34] Richard A Newcombe, Shahram Izadi, Otmar Hilliges, David Molyneaux, David Kim, Andrew J Davison, Pushmeet + +Kohi, Jamie Shotton, Steve Hodges, and Andrew Fitzgibbon. Kinectfusion: Real-time dense surface mapping and tracking. In IEEE ISMAR, 2011. 1, 2, 8 +[35] Michael Niemeyer, Fabian Manhardt, Marie-Julie Rakoto-saona, Michael Oechsle, Daniel Duckworth, Rama Gosula, Keisuke Tateno, John Bates, Dominik Kaeser, and Federico Tombari. Radsplat: Radiance field-informed gaussian splat-ting for robust real-time rendering with $900+$ FPS. arXiv preprint arXiv:2403.13806, 2024. 2 +[36] Matthias Nießner, Michael Zollhöfer, Shahram Izadi, and Marc Stamminger. Real-time 3d reconstruction at scale using voxel hashing. ACM TOG, 2013. 2, 6, 8 +[37] Lukas Radl, Michael Steiner, Mathias Parger, Alexander Weinrauch, Bernhard Kerbl, and Markus Steinberger. Stophepop: Sorted gaussian splatting for view-consistent real-time rendering. ACM TOG, 2024. 1, 2, 4 +[38] Mike Ranzinger, Greg Heinrich, Jan Kautz, and Pavlo Molchanov. Am-radio: Agglomerative vision foundation model reduce all domains into one. In CVPR, 2024. 1 +[39] Christian Reiser, Songyou Peng, Yiyi Liao, and Andreas Geiger. Kilonerf: Speeding up neural radiance fields with thousands of tiny mllps. In ICCV, 2021. 2 +[40] Christian Reiser, Stephan J. Garbin, Pratul P. Srinivasan, Dor Verbin, Richard Szeliski, Ben Mildenhall, Jonathan T. Barron, Peter Hedman, and Andreas Geiger. Binary opacity grids: Capturing fine geometric detail for mesh-based view synthesis. ACM TOG, 2024. 2 +[41] Johannes L. Schonberger, Enliang Zheng, Jan-Michael Frahm, and Marc Pollefeys. Pixelwise view selection for unstructured multi-view stereo. In ECCV, 2016. 2, 7 +[42] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. In ICLR, 2015. 7 +[43] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In CVPR, 2022. 1, 2, 3, 6, 7 +[44] Cheng Sun, Guangyan Cai, Zhengqin Li, Kai Yan, Cheng Zhang, Carl S. Marshall, Jia-Bin Huang, Shuang Zhao, and Zhao Dong. Neural-pbir reconstruction of shape, material, and illumination. In ICCV, 2023. 1, 2 +[45] Towaki Takikawa, Joey Litalien, Kangxue Yin, Karsten Kreis, Charles T. Loop, Derek Nowrouzezahrai, Alec Jacobson, Morgan McGuire, and Sanja Fidler. Neural geometric level of detail: Real-time rendering with implicit 3d shapes. In CVPR, 2021. 2 +[46] Haithem Turki, Vasu Agrawal, Samuel Rota Bulò, Lorenzo Porzi, Peter Kontschieder, Deva Ramanan, Michael Zollhöfer, and Christian Richardt. Hybridnerf: Efficient neural rendering via adaptive volumetric surfaces. In CVPR, 2024. 2 +[47] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. In NeurIPS, 2021. 2, 6, 8 +[48] Yiming Wang, Qin Han, Marc Habermann, Kostas Dani-ilidis, Christian Theobalt, and Lingjie Liu. Neus2: Fast learning of neural implicit surfaces for multi-view reconstruction. In ICCV, 2023. 2 + +[49] Zian Wang, Tianchang Shen, Merlin Nimier-David, Nicholas Sharp, Jun Gao, Alexander Keller, Sanja Fidler, Thomas Müller, and Zan Gojcic. Adaptive shells for efficient neural radiance field rendering. ACM TOG, 2023. 2 +[50] Francis Williams, Jiahui Huang, Jonathan Swartz, Gergely Klár, Vijay Thakkar, Matthew Cong, Xuanchi Ren, Ruilong Li, Clement Fuji-Tsang, Sanja Fidler, Eftychios Sifakis, and Ken Museth. fvdb: A deep-learning framework for sparse, large scale, and high performance spatial intelligence. ACM TOG, 2024. 2, 8 +[51] Tong Wu, Jiaqi Wang, Xingang Pan, Xudong Xu, Christian Theobalt, Ziwei Liu, and Dahua Lin. Voxurf: Voxel-based efficient and accurate neural surface reconstruction. In ICLR, 2023. 1, 2, 8 +[52] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, José M. Álvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. In NeurIPS, 2021. 1 +[53] Lior Yariv, Jiatao Gu, Yoni Kasten, and Yaron Lipman. Volume rendering of neural implicit surfaces. In NeurIPS, 2021. 2, 6 +[54] Lior Yariv, Peter Hedman, Christian Reiser, Dor Verbin, Pratul P. Srinivasan, Richard Szeliski, Jonathan T. Barron, and Ben Mildenhall. Bakedsdf: Meshing neural sdfs for real-time view synthesis. In ACM TOG, 2023. 2 +[55] Alex Yu, Ruilong Li, Matthew Tancik, Hao Li, Ren Ng, and Angjoo Kanazawa. Plenoctrees for real-time rendering of neural radiance fields. In ICCV, 2021. 2 +[56] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 7 \ No newline at end of file diff --git a/2025/Sparse Voxels Rasterization_ Real-time High-fidelity Radiance Field Rendering/images.zip b/2025/Sparse Voxels Rasterization_ Real-time High-fidelity Radiance Field Rendering/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..b745e9483a9ce597db1501b10923a33d3781f394 --- /dev/null +++ b/2025/Sparse Voxels Rasterization_ Real-time High-fidelity Radiance Field Rendering/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:443b222f084133d2d62e766f8177e9f132398b3719737e50ae24e5b33b6ed89a +size 532579 diff --git a/2025/Sparse Voxels Rasterization_ Real-time High-fidelity Radiance Field Rendering/layout.json b/2025/Sparse Voxels Rasterization_ Real-time High-fidelity Radiance Field Rendering/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..7675a599f47eeda846e5d81bd40b4fd2a9b2ed5f --- /dev/null +++ b/2025/Sparse Voxels Rasterization_ Real-time High-fidelity Radiance Field Rendering/layout.json @@ -0,0 +1,10717 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 61, + 103, + 549, + 122 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 103, + 549, + 122 + ], + "spans": [ + { + "bbox": [ + 61, + 103, + 549, + 122 + ], + "type": "text", + "content": "Sparse Voxels Rasterization: Real-time High-fidelity Radiance Field Rendering" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 64, + 142, + 124, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 142, + 124, + 158 + ], + "spans": [ + { + "bbox": [ + 64, + 142, + 124, + 158 + ], + "type": "text", + "content": "Cheng Sun1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 143, + 217, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 143, + 217, + 171 + ], + "spans": [ + { + "bbox": [ + 143, + 143, + 217, + 171 + ], + "type": "text", + "content": "Jaesung Choe1 \n1NVIDIA" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 234, + 143, + 307, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 234, + 143, + 307, + 171 + ], + "spans": [ + { + "bbox": [ + 234, + 143, + 307, + 171 + ], + "type": "text", + "content": "Charles Loop1 \n2Cornell University" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 318, + 143, + 424, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 143, + 424, + 171 + ], + "spans": [ + { + "bbox": [ + 318, + 143, + 424, + 171 + ], + "type": "text", + "content": "Wei-Chiu Ma" + }, + { + "bbox": [ + 318, + 143, + 424, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 318, + 143, + 424, + 171 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 318, + 143, + 424, + 171 + ], + "type": "text", + "content": "National Taiwan University" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 419, + 143, + 544, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 419, + 143, + 544, + 171 + ], + "spans": [ + { + "bbox": [ + 419, + 143, + 544, + 171 + ], + "type": "text", + "content": "Yu-Chiang Frank Wang" + }, + { + "bbox": [ + 419, + 143, + 544, + 171 + ], + "type": "inline_equation", + "content": "^{1,3}" + }, + { + "bbox": [ + 419, + 143, + 544, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 58, + 178, + 236, + 319 + ], + "blocks": [ + { + "bbox": [ + 58, + 178, + 236, + 319 + ], + "lines": [ + { + "bbox": [ + 58, + 178, + 236, + 319 + ], + "spans": [ + { + "bbox": [ + 58, + 178, + 236, + 319 + ], + "type": "image", + "image_path": "d0938896d4565a5e05a7fe016e57bbf7bc58a98e6675b90538972db605606e47.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 65, + 332, + 230, + 342 + ], + "lines": [ + { + "bbox": [ + 65, + 332, + 230, + 342 + ], + "spans": [ + { + "bbox": [ + 65, + 332, + 230, + 342 + ], + "type": "text", + "content": "(a) Volume rendering by rasterizing sparse voxels." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 236, + 179, + 416, + 331 + ], + "blocks": [ + { + "bbox": [ + 236, + 179, + 416, + 331 + ], + "lines": [ + { + "bbox": [ + 236, + 179, + 416, + 331 + ], + "spans": [ + { + "bbox": [ + 236, + 179, + 416, + 331 + ], + "type": "image", + "image_path": "6d7b51f99847323ed14a071228798e98352c46ddac178ac7489b5a7a73cb0130.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 252, + 332, + 416, + 342 + ], + "lines": [ + { + "bbox": [ + 252, + 332, + 416, + 342 + ], + "spans": [ + { + "bbox": [ + 252, + 332, + 416, + 342 + ], + "type": "text", + "content": "(b) Novel-view rendering on Mip-NeRF360 scenes." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 54, + 347, + 555, + 413 + ], + "lines": [ + { + "bbox": [ + 54, + 347, + 555, + 413 + ], + "spans": [ + { + "bbox": [ + 54, + 347, + 555, + 413 + ], + "type": "text", + "content": "Figure 1. We propose SVRaster, a novel framework for multi-view reconstruction and novel view synthesis. (a) Sparse voxel representation effectively captures the volume density and radiance field of the scene, without the need for neural networks, 3D Gaussians, and sparse-points prior. (b) Using our customized sparse voxel rasterizer, we can learn the underlying 3D scene efficiently and achieve state-of-the-art performance in both rendering quality and speed. (c) Notably, lifting 2D modal to the trained sparse voxels is simple and efficient by integrating the classic Volume Fusion [7, 8, 34]. We show examples of vision foundation model feature field from RADIO [38], semantic field from Segformer [52], and signed distance field from rendered depth, making it flexible and suitable for a wide range of applications." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 420, + 178, + 551, + 329 + ], + "blocks": [ + { + "bbox": [ + 420, + 178, + 551, + 329 + ], + "lines": [ + { + "bbox": [ + 420, + 178, + 551, + 329 + ], + "spans": [ + { + "bbox": [ + 420, + 178, + 551, + 329 + ], + "type": "image", + "image_path": "a75af022143a16049ee7fb478ceb92d03cb17d8d73599b3744176bdd2326eb2c.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 444, + 332, + 523, + 342 + ], + "lines": [ + { + "bbox": [ + 444, + 332, + 523, + 342 + ], + "spans": [ + { + "bbox": [ + 444, + 332, + 523, + 342 + ], + "type": "text", + "content": "(c) 2D-to-3D made easy." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 152, + 433, + 200, + 445 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 433, + 200, + 445 + ], + "spans": [ + { + "bbox": [ + 152, + 433, + 200, + 445 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 55, + 460, + 296, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 460, + 296, + 688 + ], + "spans": [ + { + "bbox": [ + 55, + 460, + 296, + 688 + ], + "type": "text", + "content": "We propose an efficient radiance field rendering algorithm that incorporates a rasterization process on adaptive sparse voxels without neural networks or 3D Gaussians. There are two key contributions coupled with the proposed system. The first is to adaptively and explicitly allocate sparse voxels to different levels of detail within scenes, faithfully reproducing scene details with " + }, + { + "bbox": [ + 55, + 460, + 296, + 688 + ], + "type": "inline_equation", + "content": "65536^{3}" + }, + { + "bbox": [ + 55, + 460, + 296, + 688 + ], + "type": "text", + "content": " grid resolution while achieving high rendering frame rates. Second, we customize a rasterizer for efficient adaptive sparse voxels rendering. We render voxels in the correct depth order by using ray direction-dependent Morton ordering, which avoids the well-known popping artifact found in Gaussian splatting. Our method improves the previous neural-free voxel model by over 4db PSNR and more than 10x FPS speedup, achieving state-of-the-art comparable novel-view synthesis results. Additionally, our voxel representation is seamlessly compatible with grid-based 3D processing techniques such as Volume Fusion, Voxel Pooling, and Marching Cubes, enabling a wide range of future extensions and applications." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 55, + 688, + 198, + 699 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 688, + 198, + 699 + ], + "spans": [ + { + "bbox": [ + 55, + 688, + 198, + 699 + ], + "type": "text", + "content": "Code: github.com/NVlabs/svraster" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 315, + 433, + 394, + 445 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 433, + 394, + 445 + ], + "spans": [ + { + "bbox": [ + 315, + 433, + 394, + 445 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 312, + 456, + 555, + 599 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 456, + 555, + 599 + ], + "spans": [ + { + "bbox": [ + 312, + 456, + 555, + 599 + ], + "type": "text", + "content": "Gaussian splatting [20] has emerged as one of the most promising solutions for novel view synthesis. It has drawn wide attention across multiple communities due to its exceptional rendering speed and ability to capture the nuanced details of a scene. Nevertheless, Gaussian splatting in its base form has two key limitations: first, sorting Gaussians based on their centers does not guarantee proper depth ordering. It may result in popping artifacts [37] (i.e., sudden color changes for consistent geometry) when changing views. Second, the volume density of a 3D point is ill-defined when covered by multiple Gaussians. This ambiguity makes surface reconstruction non-trivial." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 602, + 556, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 602, + 556, + 687 + ], + "spans": [ + { + "bbox": [ + 313, + 602, + 556, + 687 + ], + "type": "text", + "content": "On the other hand, grid representations inherently avoid these issues—both ordering and volume are well-defined. However, due to ray casting, the rendering speed of these methods [4, 5, 24, 43, 44, 51] is slower than that of Gaussian Splatting. This raises the question: is it possible to take the best of both worlds? Can we combine the efficiency of Gaussian splatting with the well-defined grid properties?" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 689, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 556, + 713 + ], + "type": "text", + "content": "Fortunately, the answer is yes. Our key insight is to revisit voxels—a well-established primitive with decades of" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16187" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "content": "history. Voxel representations are inherently compatible with modern graphics engines and can be rasterized efficiently. Additionally, previous work has demonstrated their ability to model scene volume densities, despite through volume ray casting. This positions voxels as the perfect bridge between rasterization and volumetric representation. However, naively adopting voxels does not work well [11]. Since a scene may consist of different levels of detail." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 168, + 295, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 168, + 295, + 335 + ], + "spans": [ + { + "bbox": [ + 55, + 168, + 295, + 335 + ], + "type": "text", + "content": "With these observations in mind, we present SVRaster, a novel framework that combines the efficiency of rasterization in 3DGS with the structured volumetric approach of grid-based representations. SVRaster leverages (1) multilevel sparse voxels to model 3D scenes and (2) implements a direction-dependent Morton order encoding that facilitates the rasterization rendering from our adaptive-sized sparse voxel representation. Our rendering is free from popping artifacts because the 3D space is partitioned into disjoint voxels, and our sorting ensures the correct rendering order. Moreover, thanks to the volumetric nature and neural-free representation of SVRaster, our sparse voxel grid can be easily and seamlessly integrated with classical grid-based 3D processing algorithms." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 335, + 295, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 335, + 295, + 407 + ], + "spans": [ + { + "bbox": [ + 55, + 335, + 295, + 407 + ], + "type": "text", + "content": "We show that our method is training fast, rendering fast, and achieves novel-view synthesis quality comparable to the state-of-the-art. We also integrate Volume Fusion, Voxel Pooling, and Marching Cubes operations into our adaptive sparse voxels, which showcases promising results of mesh extraction and 2D foundation feature fusion." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 418, + 142, + 431 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 418, + 142, + 431 + ], + "spans": [ + { + "bbox": [ + 55, + 418, + 142, + 431 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 438, + 295, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 438, + 295, + 569 + ], + "spans": [ + { + "bbox": [ + 55, + 438, + 295, + 569 + ], + "type": "text", + "content": "Differentiable volume rendering has made significant strides in 3D scene reconstruction and novel-view synthesis tasks. Neural Radiance Fields (NeRF) [30] laid the foundation for this progress by optimizing a volumetric function, parameterized by multi-layer perceptrons (MLPs) to encode both geometry and appearance through differential rendering. Subsequent studies focus on accelerating speed by decomposing large MLPs into grid-based representations, with a shallow MLP typically still being employed. Several grid representations have been explored—dense grids [43], factorized grids [5], tri-planes [4], and hash grids [32]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 571, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 571, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 571, + 295, + 713 + ], + "type": "text", + "content": "In particular, our method is closely related to sparse voxel grid representations. Sparse Voxel Octrees [22] and VDB trees [33] are commonly used to manage sparse voxels and facilitate rendering. In contrast, our sparse voxels are stored in a 1D array without advanced data structures. Our rasterizer with the proposed direction-dependent Morton order ensures correct rendering order. To model volumes in sparse leaf nodes, different strategies exist, such as low-resolution 3D grid [33, 50] or implicit neural field [25, 45]. Our leaf node is a single voxel with explicit density and color parameters, like Plenoxels [11]. Regarding voxel levels, previous methods allocate voxels to a tar" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "text", + "content": "get level [11, 25, 45, 55] or use a shallow tree [33, 50]. Our voxels can adaptively fit into different levels across the entire tree depth, maximizing flexibility and scalability." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 110, + 553, + 181 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 110, + 553, + 181 + ], + "spans": [ + { + "bbox": [ + 313, + 110, + 553, + 181 + ], + "type": "text", + "content": "Another scalability issue of the previous grid-based methods is that they still use some sort of dense 3D grid in the field of novel-view synthesis. For instance, dense occupancy grids [5, 23, 32, 43] for free-space skipping or dense pointer grid [11] to support sparse voxel lookup. We do not use any 3D dense data structures." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 182, + 553, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 182, + 553, + 374 + ], + "spans": [ + { + "bbox": [ + 313, + 182, + 553, + 374 + ], + "type": "text", + "content": "3D Gaussian Splatting (3DGS) [20] takes a different approach by representing scenes with 3D Gaussian primitives and using rasterization for rendering, achieving state-of-the-art trade-off for quality and speed. Our work is inspired by the efficiency of using a rasterizer by 3DGS. However, 3DGS exhibits view-inconsistent popping artifacts due to inaccurate rendering order and primitive overlapping. Some recent works mitigate this artifact [31, 37], but completely resolving it significantly harms rendering speed [28]. Our method does not suffer from the popping artifact. Typically, 3DGS [20] initializes Gaussians from the triangulated points via COLMAP [41], which is especially critical for unbounded scenes. The sparse points geometry is also employed to guide the training of volumetric representations and has shown improvement in some NeRF-based methods [9, 12]. In this work, we do not use the sparse points." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 376, + 553, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 376, + 553, + 543 + ], + "spans": [ + { + "bbox": [ + 313, + 376, + 553, + 543 + ], + "type": "text", + "content": "Mesh reconstruction is an important extended topic for both volumetric-based and GS-based methods. Recent 3DGS variants [14, 16] with proper regularization have shown a good speed-accuracy trade-off. However, directly extracting meshes from 3DGS remains challenging, so these methods still rely on volumetric-based post-processing algorithms to extract meshes. NeRF variants [24, 44, 47, 48, 51, 53] improve surface quality by rendering a signed distance field (SDF) instead of density field. The well-defined volume makes it straightforward to extract a mesh from the isosurface. Our sparse voxels are also a volumetric representation. Though our voxel also follows volumetric representation focusing on the density field, we still can achieve promising accuracy with minimal time." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 544, + 553, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 544, + 553, + 652 + ], + "spans": [ + { + "bbox": [ + 313, + 544, + 553, + 652 + ], + "type": "text", + "content": "For efficient rendering, another line of research is to convert (or distill) the geometry and appearance of a high-quality but slow model into other efficient kind of representations such as smaller MLPs [39, 46], meshes [40, 49, 54], grids [10], or the recent Gaussians [35]. Typically, the powerful Zip-NeRF [3] is employed as the teacher model which requires hours of training per scene. Our volumetric representation could enable training-free conversion from a large implicit field, which is however out of our current scope." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "text", + "content": "Finally, the sparse voxel is also a widely-used representation for 3D processing [7, 8, 19, 26, 34, 36] and understanding [6, 13, 50]. We mainly focus on rendering in this work while we believe that adapting our method to these techniques is a promising future direction." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16188" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 70, + 173, + 185 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 173, + 185 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 173, + 185 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 173, + 185 + ], + "type": "image", + "image_path": "c555599a72e63f8845ff350e9ad912262f6383be220dfe60687eb61aba27006d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 109, + 190, + 119, + 198 + ], + "lines": [ + { + "bbox": [ + 109, + 190, + 119, + 198 + ], + "spans": [ + { + "bbox": [ + 109, + 190, + 119, + 198 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 173, + 69, + 293, + 186 + ], + "blocks": [ + { + "bbox": [ + 173, + 69, + 293, + 186 + ], + "lines": [ + { + "bbox": [ + 173, + 69, + 293, + 186 + ], + "spans": [ + { + "bbox": [ + 173, + 69, + 293, + 186 + ], + "type": "image", + "image_path": "00d9280cf0e4d38f6e0fd48a2ca55da6578dc7b213db74d1c52d927959d0bf5c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 232, + 189, + 243, + 198 + ], + "lines": [ + { + "bbox": [ + 232, + 189, + 243, + 198 + ], + "spans": [ + { + "bbox": [ + 232, + 189, + 243, + 198 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 208, + 296, + 330 + ], + "lines": [ + { + "bbox": [ + 55, + 208, + 296, + 330 + ], + "spans": [ + { + "bbox": [ + 55, + 208, + 296, + 330 + ], + "type": "text", + "content": "Figure 2. Sparse voxels scene representation. (Left) We allocate voxel under an Octree layout. Each voxels has its own Spherical Harmonic coefficient for view-dependent appearance. The color field is approximated as a constant inside a voxel when rendering a view for efficiency. The density field is trilaterally varied inside a voxel and is modeled by the density values on the corner grid points (i.e., the black dots " + }, + { + "bbox": [ + 55, + 208, + 296, + 330 + ], + "type": "inline_equation", + "content": "\\bullet" + }, + { + "bbox": [ + 55, + 208, + 296, + 330 + ], + "type": "text", + "content": ") of each voxel. The grid points densities are shared between adjacent voxels. (Right) We evenly sample " + }, + { + "bbox": [ + 55, + 208, + 296, + 330 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 208, + 296, + 330 + ], + "type": "text", + "content": " points inside the segment of ray-voxel intersection to compute volume integration for its alpha value contributing to the pixel ray. See Sec. 3.1.1 for details." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 340, + 122, + 354 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 340, + 122, + 354 + ], + "spans": [ + { + "bbox": [ + 55, + 340, + 122, + 354 + ], + "type": "text", + "content": "3. Approach" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 360, + 296, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 360, + 296, + 433 + ], + "spans": [ + { + "bbox": [ + 55, + 360, + 296, + 433 + ], + "type": "text", + "content": "We present our approach as follows. First in Sec. 3.1, we introduce our sparse voxels scene representation and our rasterizer for rendering sparse voxels into pixels. Later in Sec. 3.2, we describe the progressive scene optimization strategy, which is designed for faithfully reconstructing the scene from multi-view images using our sparse voxels." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 441, + 209, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 441, + 209, + 453 + ], + "spans": [ + { + "bbox": [ + 55, + 441, + 209, + 453 + ], + "type": "text", + "content": "3.1. Sparse Voxels Rasterization" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 459, + 296, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 459, + 296, + 495 + ], + "spans": [ + { + "bbox": [ + 55, + 459, + 296, + 495 + ], + "type": "text", + "content": "Recent neural radiance field rendering approaches, such as NeRF [30] and 3DGS [20] variants, use the following alpha composition equation to render a pixel's color C:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 67, + 497, + 295, + 520 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 497, + 295, + 520 + ], + "spans": [ + { + "bbox": [ + 67, + 497, + 295, + 520 + ], + "type": "interline_equation", + "content": "\\mathrm {C} = \\sum_ {i = 1} ^ {N} T _ {i} \\cdot \\alpha_ {i} \\cdot c _ {i}, \\quad T _ {i} = \\prod_ {j = 1} ^ {i - 1} \\left(1 - \\alpha_ {j}\\right), \\tag {1}", + "image_path": "b0f26b0374b9ab04e03ddead50bcbbacfbc4c00c4ba6591eaf8e60e5edd0abb3.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 523, + 295, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 523, + 295, + 606 + ], + "spans": [ + { + "bbox": [ + 55, + 523, + 295, + 606 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 523, + 295, + 606 + ], + "type": "inline_equation", + "content": "\\alpha_{i}\\in \\mathbb{R}_{\\in [0,1]}" + }, + { + "bbox": [ + 55, + 523, + 295, + 606 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 523, + 295, + 606 + ], + "type": "inline_equation", + "content": "c_{i}\\in \\mathbb{R}_{\\in [0,1]}^{3}" + }, + { + "bbox": [ + 55, + 523, + 295, + 606 + ], + "type": "text", + "content": " are alpha and view-dependent color at the " + }, + { + "bbox": [ + 55, + 523, + 295, + 606 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 523, + 295, + 606 + ], + "type": "text", + "content": "-th sampled point or primitive on the pixel-ray. The quantity " + }, + { + "bbox": [ + 55, + 523, + 295, + 606 + ], + "type": "inline_equation", + "content": "T_{i}" + }, + { + "bbox": [ + 55, + 523, + 295, + 606 + ], + "type": "text", + "content": " is known as the transmittance. At a high-level, the difference between each method comes down to: " + }, + { + "bbox": [ + 55, + 523, + 295, + 606 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 523, + 295, + 606 + ], + "type": "text", + "content": " ) how they find the " + }, + { + "bbox": [ + 55, + 523, + 295, + 606 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 523, + 295, + 606 + ], + "type": "text", + "content": " points or primitives to composite a pixel-ray and " + }, + { + "bbox": [ + 55, + 523, + 295, + 606 + ], + "type": "inline_equation", + "content": "ii" + }, + { + "bbox": [ + 55, + 523, + 295, + 606 + ], + "type": "text", + "content": " ) how the rgb and alpha values are determined from their scene representations." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 608, + 296, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 608, + 296, + 668 + ], + "spans": [ + { + "bbox": [ + 55, + 608, + 296, + 668 + ], + "type": "text", + "content": "Our sparse-voxel method follows the same principle. In Sec. 3.1.1, we provide details of our sparse voxels scene representation and how the " + }, + { + "bbox": [ + 55, + 608, + 296, + 668 + ], + "type": "inline_equation", + "content": "\\alpha_{i}" + }, + { + "bbox": [ + 55, + 608, + 296, + 668 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 608, + 296, + 668 + ], + "type": "inline_equation", + "content": "c_{i}" + }, + { + "bbox": [ + 55, + 608, + 296, + 668 + ], + "type": "text", + "content": " in Eq. (1) are computed from a voxel. In Sec. 3.1.2, we present our rasterizer, that gathers the " + }, + { + "bbox": [ + 55, + 608, + 296, + 668 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 608, + 296, + 668 + ], + "type": "text", + "content": " voxels to be composited for each pixel." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 674, + 176, + 686 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 674, + 176, + 686 + ], + "spans": [ + { + "bbox": [ + 55, + 674, + 176, + 686 + ], + "type": "text", + "content": "3.1.1. Scene Representation" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "We first describe the grid layout for allocating our sparse voxels and then derive the alpha value, view-dependent" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "content": "color, and other geometric properties needed for the composite rendering of a voxel." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 103, + 555, + 305 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 103, + 555, + 305 + ], + "spans": [ + { + "bbox": [ + 313, + 103, + 555, + 305 + ], + "type": "text", + "content": "Sparse voxel grid. Our SVRaster constructs 3D scenes using a sparse voxel representation. We allocate voxels following an Octree space partition rule (i.e., Octree layout as illustrated in Fig. 2a) for two reasons necessary for achieving high-quality results. First, it facilitates the correct rendering order of voxels with various sizes (Sec. 3.1.2). Second, we can adaptively fit the sparse voxels to different scene level-of-details (Sec. 3.2). Note that our representation does not replicate a traditional Octree data structure with parent-child pointers or linear Octree. Specifically, we only keep voxels at the Octree leaf nodes without any ancestor nodes. Our sorting-based rasterizer will project voxels to image space and guarantee all voxels are in the correct order when rendering. In sum, we store individual voxels in arbitrary order without the need to maintain a more complex data structure, thanks to the flexibility provided by our rasterizer." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 308, + 554, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 308, + 554, + 392 + ], + "spans": [ + { + "bbox": [ + 313, + 308, + 554, + 392 + ], + "type": "text", + "content": "We choose a maximum level of detail " + }, + { + "bbox": [ + 313, + 308, + 554, + 392 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 313, + 308, + 554, + 392 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 313, + 308, + 554, + 392 + ], + "type": "inline_equation", + "content": "= 16" + }, + { + "bbox": [ + 313, + 308, + 554, + 392 + ], + "type": "text", + "content": " in this work) that defines a maximum grid resolution at " + }, + { + "bbox": [ + 313, + 308, + 554, + 392 + ], + "type": "inline_equation", + "content": "(2^L)^3" + }, + { + "bbox": [ + 313, + 308, + 554, + 392 + ], + "type": "text", + "content": ". Let " + }, + { + "bbox": [ + 313, + 308, + 554, + 392 + ], + "type": "inline_equation", + "content": "\\mathbf{w}_{\\mathrm{s}} \\in \\mathbb{R}" + }, + { + "bbox": [ + 313, + 308, + 554, + 392 + ], + "type": "text", + "content": " be the Octree size and " + }, + { + "bbox": [ + 313, + 308, + 554, + 392 + ], + "type": "inline_equation", + "content": "\\mathbf{w}_{\\mathrm{c}} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 313, + 308, + 554, + 392 + ], + "type": "text", + "content": " be the Octree center in the world space. The voxel index " + }, + { + "bbox": [ + 313, + 308, + 554, + 392 + ], + "type": "inline_equation", + "content": "v = \\{i,j,k\\} \\in [0,\\dots,2^{L} - 1]^{3}" + }, + { + "bbox": [ + 313, + 308, + 554, + 392 + ], + "type": "text", + "content": " together with a Octree level " + }, + { + "bbox": [ + 313, + 308, + 554, + 392 + ], + "type": "inline_equation", + "content": "l \\in [1,L]" + }, + { + "bbox": [ + 313, + 308, + 554, + 392 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 313, + 308, + 554, + 392 + ], + "type": "inline_equation", + "content": "l = 0" + }, + { + "bbox": [ + 313, + 308, + 554, + 392 + ], + "type": "text", + "content": " represent root node and is not used) define voxel size " + }, + { + "bbox": [ + 313, + 308, + 554, + 392 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_{\\mathrm{s}}" + }, + { + "bbox": [ + 313, + 308, + 554, + 392 + ], + "type": "text", + "content": " and voxel center " + }, + { + "bbox": [ + 313, + 308, + 554, + 392 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_{\\mathrm{c}}" + }, + { + "bbox": [ + 313, + 308, + 554, + 392 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 328, + 396, + 553, + 409 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 396, + 553, + 409 + ], + "spans": [ + { + "bbox": [ + 328, + 396, + 553, + 409 + ], + "type": "interline_equation", + "content": "\\mathbf {v} _ {\\mathrm {s}} = \\mathbf {w} _ {\\mathrm {s}} \\cdot 2 ^ {- l}, \\quad \\mathbf {v} _ {\\mathrm {c}} = \\mathbf {w} _ {\\mathrm {c}} - 0. 5 \\cdot \\mathbf {w} _ {\\mathrm {s}} + \\mathbf {v} _ {\\mathrm {s}} \\cdot v. \\tag {2}", + "image_path": "5cb732091531bbf9383b4dcf468b94eae2933b732eaa26d4c84691ff96dd768c.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 413, + 554, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 413, + 554, + 460 + ], + "spans": [ + { + "bbox": [ + 313, + 413, + 554, + 460 + ], + "type": "text", + "content": "Internally, we map the grid index to its Morton code using a well-known bit interleaving operation in the low-level CUDA implementation. Please see supplementary materials for more details." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 468, + 554, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 468, + 554, + 552 + ], + "spans": [ + { + "bbox": [ + 313, + 468, + 554, + 552 + ], + "type": "text", + "content": "Voxel alpha from density field. Next, we present details for the geometry and appearance modeling of each voxel primitive. For scene geometry, we use eight parameters corresponding to the voxel corners to model a trilinear density field inside each voxel, denoted as " + }, + { + "bbox": [ + 313, + 468, + 554, + 552 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_{\\mathrm{geo}} \\in \\mathbb{R}^{2 \\times 2 \\times 2}" + }, + { + "bbox": [ + 313, + 468, + 554, + 552 + ], + "type": "text", + "content": ". Sharing corners among adjacent voxels results in a continuous density field." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 554, + 554, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 554, + 554, + 590 + ], + "spans": [ + { + "bbox": [ + 313, + 554, + 554, + 590 + ], + "type": "text", + "content": "We also need an activation function to ensure a nonnegative density value for the raw density from " + }, + { + "bbox": [ + 313, + 554, + 554, + 590 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_{\\mathrm{geo}}" + }, + { + "bbox": [ + 313, + 554, + 554, + 590 + ], + "type": "text", + "content": ". For this purpose, we use exponential-linear activation:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 321, + 593, + 553, + 626 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 593, + 553, + 626 + ], + "spans": [ + { + "bbox": [ + 321, + 593, + 553, + 626 + ], + "type": "interline_equation", + "content": "\\operatorname {e x p l i n} (x) = \\left\\{ \\begin{array}{l l} x & \\text {i f} x > 1. 1 \\\\ \\exp \\left(\\frac {x}{1 . 1} - 1 + \\ln 1. 1\\right) & \\text {o t h e r w i s e} \\end{array} , \\right. \\tag {3}", + "image_path": "5e1bd290f05d92fcea9afedd567440dd984f5a0dd1a66fe4e1a0260e9831ccc3.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 628, + 553, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 628, + 553, + 663 + ], + "spans": [ + { + "bbox": [ + 313, + 628, + 553, + 663 + ], + "type": "text", + "content": "which approximates softplus but is more efficient to compute. For a sharp density field inside a voxel, we apply the non-linear activation after trilinear interpolation [18, 43]." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "text", + "content": "To derive the alpha value of a voxel contributing to the alpha composition formulation in Eq. (1), we evenly sample " + }, + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "text", + "content": " points in the ray segment of ray-voxel intersection as depicted in Fig. 2b. The equation follows the numerical" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16189" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 70, + 294, + 125 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 294, + 125 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 294, + 125 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 294, + 125 + ], + "type": "image", + "image_path": "e8373eed911a159dc8296b70ba2ae245e42822207ef5b1bb96d018aa9efd4d19.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 133, + 295, + 145 + ], + "lines": [ + { + "bbox": [ + 55, + 133, + 295, + 145 + ], + "spans": [ + { + "bbox": [ + 55, + 133, + 295, + 145 + ], + "type": "text", + "content": "Figure 3. Rasterization procedure. Refer to Sec. 3.1.2 for details." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 171, + 274, + 183 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 171, + 274, + 183 + ], + "spans": [ + { + "bbox": [ + 55, + 171, + 274, + 183 + ], + "type": "text", + "content": "integration for volume rendering as in NeRF [29, 30]:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 59, + 186, + 295, + 219 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 186, + 295, + 219 + ], + "spans": [ + { + "bbox": [ + 59, + 186, + 295, + 219 + ], + "type": "interline_equation", + "content": "\\alpha = 1 - \\exp \\left(- \\frac {l}{K} \\sum_ {k = 1} ^ {K} \\operatorname {e x p l i n} \\left(\\operatorname {i n t e r p} \\left(\\mathbf {v} _ {\\text {g e o}}, \\mathbf {q} _ {k}\\right)\\right)\\right), \\tag {4}", + "image_path": "b4880a80812258867c9d16984e82f4dd999d3c97de82fcd272550feb01aac9fd.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 222, + 295, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 222, + 295, + 258 + ], + "spans": [ + { + "bbox": [ + 55, + 222, + 295, + 258 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 222, + 295, + 258 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 55, + 222, + 295, + 258 + ], + "type": "text", + "content": " is the ray segment length, " + }, + { + "bbox": [ + 55, + 222, + 295, + 258 + ], + "type": "inline_equation", + "content": "\\mathbf{q}_k" + }, + { + "bbox": [ + 55, + 222, + 295, + 258 + ], + "type": "text", + "content": " is the local voxel coordinate of the " + }, + { + "bbox": [ + 55, + 222, + 295, + 258 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 222, + 295, + 258 + ], + "type": "text", + "content": "-th sample point, and " + }, + { + "bbox": [ + 55, + 222, + 295, + 258 + ], + "type": "inline_equation", + "content": "\\mathrm{interp}(\\cdot)" + }, + { + "bbox": [ + 55, + 222, + 295, + 258 + ], + "type": "text", + "content": " indicates trilinear interpolation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 268, + 296, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 268, + 296, + 365 + ], + "spans": [ + { + "bbox": [ + 55, + 268, + 296, + 365 + ], + "type": "text", + "content": "Voxel view-dependent color from SHs. To model view-dependent scene appearance, we use " + }, + { + "bbox": [ + 55, + 268, + 296, + 365 + ], + "type": "inline_equation", + "content": "N_{\\mathrm{shd}}" + }, + { + "bbox": [ + 55, + 268, + 296, + 365 + ], + "type": "text", + "content": " degree SH. For increased efficiency, we assume the SH coefficients stay constant inside a voxel, denoted as " + }, + { + "bbox": [ + 55, + 268, + 296, + 365 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_{\\mathrm{sh}} \\in \\mathbb{R}^{(N_{\\mathrm{shd}} + 1)^2 \\times 3}" + }, + { + "bbox": [ + 55, + 268, + 296, + 365 + ], + "type": "text", + "content": ". We approximate voxel colors as a function of the direction from the camera position " + }, + { + "bbox": [ + 55, + 268, + 296, + 365 + ], + "type": "inline_equation", + "content": "\\mathbf{r}_0" + }, + { + "bbox": [ + 55, + 268, + 296, + 365 + ], + "type": "text", + "content": " to the voxel center " + }, + { + "bbox": [ + 55, + 268, + 296, + 365 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_{\\mathrm{c}}" + }, + { + "bbox": [ + 55, + 268, + 296, + 365 + ], + "type": "text", + "content": " instead of individual ray direction " + }, + { + "bbox": [ + 55, + 268, + 296, + 365 + ], + "type": "inline_equation", + "content": "\\mathbf{r}_{\\mathrm{d}}" + }, + { + "bbox": [ + 55, + 268, + 296, + 365 + ], + "type": "text", + "content": " for the sake of efficiency following 3DGS:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 69, + 369, + 295, + 382 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 369, + 295, + 382 + ], + "spans": [ + { + "bbox": [ + 69, + 369, + 295, + 382 + ], + "type": "interline_equation", + "content": "c = \\max (0, \\operatorname {s h} _ {-} \\operatorname {e v a l} (\\mathbf {v} _ {\\mathrm {s h}}, \\text {n o r m a l i z e} (\\mathbf {v} _ {\\mathrm {c}} - \\mathbf {r} _ {\\mathrm {o}}))), \\tag {5}", + "image_path": "070198758a8ed8cd5a820985905bc875a47e3a29c6f8f584c7bcc38f7573f07c.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 385, + 296, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 385, + 296, + 445 + ], + "spans": [ + { + "bbox": [ + 55, + 385, + 296, + 445 + ], + "type": "text", + "content": "which is the view-dependent color intensity of the voxel contributing to the pixel composition Eq. (1). Due to the approximation, the resulting SH color of a voxel can be shared by all covered pixels in the image rather than evaluating the SH for each intersecting ray." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 455, + 296, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 455, + 296, + 538 + ], + "spans": [ + { + "bbox": [ + 55, + 455, + 296, + 538 + ], + "type": "text", + "content": "Voxel normal. The rendering of other features or properties is similar to rendering a color image by replacing the color term " + }, + { + "bbox": [ + 55, + 455, + 296, + 538 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 455, + 296, + 538 + ], + "type": "text", + "content": " in Eq. (1) with the target modality like the normal of a voxel density field. For rendering efficiency, we assume the normal stays constant inside a voxel, which is represented by the analytical gradient of the density field at the voxel center:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 92, + 543, + 295, + 557 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 92, + 543, + 295, + 557 + ], + "spans": [ + { + "bbox": [ + 92, + 543, + 295, + 557 + ], + "type": "interline_equation", + "content": "\\vec {\\mathbf {n}} = \\text {n o r m a l i z e} \\left(\\nabla_ {\\mathbf {q}} \\operatorname {i n t e r p} \\left(\\mathbf {v} _ {\\mathrm {g e o}}, \\mathbf {q} _ {\\mathrm {c}}\\right)\\right), \\tag {6}", + "image_path": "6681f2cebcba2a62d2a78524e6ef1a647429f66e5f263a080841942df2647fc7.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 559, + 296, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 559, + 296, + 620 + ], + "spans": [ + { + "bbox": [ + 55, + 559, + 296, + 620 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 559, + 296, + 620 + ], + "type": "inline_equation", + "content": "\\mathbf{q}_{\\mathrm{c}} = (0.5, 0.5, 0.5)" + }, + { + "bbox": [ + 55, + 559, + 296, + 620 + ], + "type": "text", + "content": " and the closed-form equations for forward and backward passes are in the supplementary material. Similar to the SH colors, the differentiable voxel normals are computed once in pre-processing and are shared by all the intersecting rays in the image." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "content": "Voxel depth. Unlike colors and normals, the point depths to composite is efficient to compute so we do the same " + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "content": " points sampling as in the voxel alpha value in Eq. (4) for a more precise depth rendering. We manually expand and simplify the forward and backward computation for small number of " + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "content": " in our CUDA implementation. Please refer to supplementary materials for details." + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 316, + 71, + 358, + 114 + ], + "blocks": [ + { + "bbox": [ + 316, + 71, + 358, + 114 + ], + "lines": [ + { + "bbox": [ + 316, + 71, + 358, + 114 + ], + "spans": [ + { + "bbox": [ + 316, + 71, + 358, + 114 + ], + "type": "image", + "image_path": "2f1341aea32eb6c2d516926feb1d92f0e0388fb88b66f5c291f65973c22bfb46.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 316, + 114, + 358, + 165 + ], + "blocks": [ + { + "bbox": [ + 316, + 114, + 358, + 165 + ], + "lines": [ + { + "bbox": [ + 316, + 114, + 358, + 165 + ], + "spans": [ + { + "bbox": [ + 316, + 114, + 358, + 165 + ], + "type": "image", + "image_path": "6de3319d7697aa716df85624bd8d2fdc2cf5a1bdea4db5821848e2bd7dc11632.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 331, + 168, + 342, + 177 + ], + "lines": [ + { + "bbox": [ + 331, + 168, + 342, + 177 + ], + "spans": [ + { + "bbox": [ + 331, + 168, + 342, + 177 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 364, + 73, + 402, + 114 + ], + "blocks": [ + { + "bbox": [ + 364, + 73, + 402, + 114 + ], + "lines": [ + { + "bbox": [ + 364, + 73, + 402, + 114 + ], + "spans": [ + { + "bbox": [ + 364, + 73, + 402, + 114 + ], + "type": "image", + "image_path": "9eccfdd93cd764748253f609b75e6d7abeca7798ced325c90a2de9445dd45b15.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 365, + 115, + 402, + 163 + ], + "blocks": [ + { + "bbox": [ + 365, + 115, + 402, + 163 + ], + "lines": [ + { + "bbox": [ + 365, + 115, + 402, + 163 + ], + "spans": [ + { + "bbox": [ + 365, + 115, + 402, + 163 + ], + "type": "image", + "image_path": "e243c7d69faf023ea5177c92999eaac74d2e0e9df41eb1d7b6ab0972582ecaaa.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 405, + 168, + 415, + 177 + ], + "lines": [ + { + "bbox": [ + 405, + 168, + 415, + 177 + ], + "spans": [ + { + "bbox": [ + 405, + 168, + 415, + 177 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 403, + 74, + 455, + 114 + ], + "blocks": [ + { + "bbox": [ + 403, + 74, + 455, + 114 + ], + "lines": [ + { + "bbox": [ + 403, + 74, + 455, + 114 + ], + "spans": [ + { + "bbox": [ + 403, + 74, + 455, + 114 + ], + "type": "image", + "image_path": "5de485304965ce85f7df2cbbe759db978491b2216adb77fe66d611bd27bf9007.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 403, + 115, + 455, + 163 + ], + "blocks": [ + { + "bbox": [ + 403, + 115, + 455, + 163 + ], + "lines": [ + { + "bbox": [ + 403, + 115, + 455, + 163 + ], + "spans": [ + { + "bbox": [ + 403, + 115, + 455, + 163 + ], + "type": "image", + "image_path": "1058bcc0c9c7bd80855967ef1de2ccc3e62c7a8b1e920cb607ab5c7b7d735964.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 186, + 555, + 295 + ], + "lines": [ + { + "bbox": [ + 313, + 186, + 555, + 295 + ], + "spans": [ + { + "bbox": [ + 313, + 186, + 555, + 295 + ], + "type": "text", + "content": "Figure 4. Illustration of the rendering order. (a) In both cases, the smaller voxels should be rendered first but they will arranged behind the larger voxels if using voxel centers or the nearest corners as the sorting order. (b) We show the four types of Morton order under the 2D world. The voxel rendering order under an Octree node is depend on which world quadrant the ray direction is pointing to. (c) An toy example of sorting the Morton order encoding. All the ray directions going toward the up-right quadrant can use the sorted voxels for a correct rendering order. See Sec. 3.1.2 for details." + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 463, + 75, + 553, + 165 + ], + "blocks": [ + { + "bbox": [ + 463, + 75, + 553, + 165 + ], + "lines": [ + { + "bbox": [ + 463, + 75, + 553, + 165 + ], + "spans": [ + { + "bbox": [ + 463, + 75, + 553, + 165 + ], + "type": "image", + "image_path": "2583482f33309c08029722bffad55bc8cd414c4476a363bdae81a4358d68e317.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 503, + 168, + 513, + 176 + ], + "lines": [ + { + "bbox": [ + 503, + 168, + 513, + 176 + ], + "spans": [ + { + "bbox": [ + 503, + 168, + 513, + 176 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 313, + 446, + 325 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 313, + 446, + 325 + ], + "spans": [ + { + "bbox": [ + 313, + 313, + 446, + 325 + ], + "type": "text", + "content": "3.1.2. Rasterization Algorithm" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 328, + 555, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 328, + 555, + 376 + ], + "spans": [ + { + "bbox": [ + 313, + 328, + 555, + 376 + ], + "type": "text", + "content": "The overview of our sparse voxel rasterization procedure is depicted in Fig. 3. We build our sorting-based rasterizer based on the highly efficient CUDA implementation of 3DGS [20]. The procedure is detailed in the following." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 380, + 554, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 380, + 554, + 453 + ], + "spans": [ + { + "bbox": [ + 313, + 380, + 554, + 453 + ], + "type": "text", + "content": "Projection to image space. The first step of rasterization is projecting sparse voxels onto image space and assigning the voxels to the tiles (i.e., image patches) they cover. In practice, we project the eight corner points of each voxel. The voxel is assigned to all tiles overlapped with the axis-aligned bounding box formed by the projected eight points." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 457, + 554, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 457, + 554, + 529 + ], + "spans": [ + { + "bbox": [ + 313, + 457, + 554, + 529 + ], + "type": "text", + "content": "Pre-processing voxels. For active voxels assigned to tiles of the target view, we gather their densities " + }, + { + "bbox": [ + 313, + 457, + 554, + 529 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_{\\mathrm{geo}}" + }, + { + "bbox": [ + 313, + 457, + 554, + 529 + ], + "type": "text", + "content": " from the grid points, compute view-dependent colors from their spherical harmonic coefficients with Eq. (5), and derive voxel normals by Eq. (6). The pre-processed voxel properties are shared among all pixels during rendering." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "text", + "content": "Sorting voxels. For accurate rasterization, primitive rendering order is important. Similar to the challenge in 3DGS [20], using primitive centers or their closest distance to the camera can produce incorrect ordering, producing popping artifacts [37]. We show two incorrect ordering results using naive sorting criteria in Fig. 4a. Thanks to the Octree layout (Sec. 3.1.1), we can sort by Morton order using our sparse voxels representation. As illustrated in Fig. 4b, we can follow certain types of Morton order to render the voxels under an Octree node for correct ordering. The type of Morton order to follow is solely dependent on the positive/negative signs of the ray direction (the ray origin doesn't matter). That is to say, we have eight permutations of Morton order for different ray directions in the 3D space. Finally, the generalization to multi-level voxels can" + } + ] + } + ], + "index": 27 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16190" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "type": "text", + "content": "be proved by induction. An ordering example in 2D with three levels is depicted in Fig. 4c." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 96, + 296, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 96, + 296, + 156 + ], + "spans": [ + { + "bbox": [ + 55, + 96, + 296, + 156 + ], + "type": "text", + "content": "The sorting is applied for each image tile. In case all the pixels in a tile share the same ray direction signs, we can simply sort the assigned voxels by their type of Morton order. We handle the corner case when multiple Morton orders are required in supplementary materials." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 160, + 296, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 160, + 296, + 304 + ], + "spans": [ + { + "bbox": [ + 55, + 160, + 296, + 304 + ], + "type": "text", + "content": "Rendering pixels. Finally, we proceed with alpha composition, Eq. (1), to render pixels. In our case, the " + }, + { + "bbox": [ + 55, + 160, + 296, + 304 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 160, + 296, + 304 + ], + "type": "text", + "content": " primitives blend of a pixel-ray depends on the number of sparse voxels assigned to the tile that the pixel-ray belongs to. The computation of the alpha, color, and other geometric properties from our sparse voxels are described in Sec. 3.1.1. When rendering sparse voxels for a pixel-ray, we compute ray-aabb intersection to determine the ray segment to sample (for voxel alpha in Eq. (4)) and skip some non-intersected sparse voxels. We do early termination of the alpha composition if the transmittance of a sparse voxel is below a threshold " + }, + { + "bbox": [ + 55, + 160, + 296, + 304 + ], + "type": "inline_equation", + "content": "T_{i} < h_{\\mathrm{T}}" + }, + { + "bbox": [ + 55, + 160, + 296, + 304 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 309, + 296, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 309, + 296, + 346 + ], + "spans": [ + { + "bbox": [ + 55, + 309, + 296, + 346 + ], + "type": "text", + "content": "Anti-aliasing. To mitigate aliasing artifacts, we render in " + }, + { + "bbox": [ + 55, + 309, + 296, + 346 + ], + "type": "inline_equation", + "content": "h_{\\mathrm{ss}}" + }, + { + "bbox": [ + 55, + 309, + 296, + 346 + ], + "type": "text", + "content": " times higher resolution and then apply image downsampling to the target resolution with an anti-aliasing filter." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 352, + 264, + 365 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 352, + 264, + 365 + ], + "spans": [ + { + "bbox": [ + 55, + 352, + 264, + 365 + ], + "type": "text", + "content": "3.2. Progressive Sparse Voxels Optimization" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 369, + 296, + 405 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 369, + 296, + 405 + ], + "spans": [ + { + "bbox": [ + 55, + 369, + 296, + 405 + ], + "type": "text", + "content": "In this section, we describe the procedure to optimize a 3D scene from the input frames with known camera parameters using our SVRaster presented in Sec. 3.1." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 410, + 296, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 410, + 296, + 506 + ], + "spans": [ + { + "bbox": [ + 55, + 410, + 296, + 506 + ], + "type": "text", + "content": "Voxel max sampling rate. We first define the maximum sampling rate " + }, + { + "bbox": [ + 55, + 410, + 296, + 506 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_{\\mathrm{rate}}" + }, + { + "bbox": [ + 55, + 410, + 296, + 506 + ], + "type": "text", + "content": " of each voxel on the training images, which reflects the image region a voxel can cover. A smaller " + }, + { + "bbox": [ + 55, + 410, + 296, + 506 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_{\\mathrm{rate}}" + }, + { + "bbox": [ + 55, + 410, + 296, + 506 + ], + "type": "text", + "content": " indicates that the voxel is more prone to overfitting due to less observation. We use " + }, + { + "bbox": [ + 55, + 410, + 296, + 506 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_{\\mathrm{rate}}" + }, + { + "bbox": [ + 55, + 410, + 296, + 506 + ], + "type": "text", + "content": " in our voxel initialization and subdivision process described later. Given " + }, + { + "bbox": [ + 55, + 410, + 296, + 506 + ], + "type": "inline_equation", + "content": "N_{\\mathrm{cam}}" + }, + { + "bbox": [ + 55, + 410, + 296, + 506 + ], + "type": "text", + "content": " training cameras, we estimate the maximum sampling rate of a voxel as follows with visualization in Fig. 5a:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 82, + 508, + 295, + 535 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 508, + 295, + 535 + ], + "spans": [ + { + "bbox": [ + 82, + 508, + 295, + 535 + ], + "type": "interline_equation", + "content": "\\mathbf {v} _ {\\text {r a t e}} = \\max _ {i} ^ {N _ {\\text {c a m}}} \\frac {\\mathbf {v} _ {\\mathrm {s}}}{\\mathbf {v} _ {\\text {i n t e r v a l}} ^ {(i)}}, \\tag {7a}", + "image_path": "fbac5c97aaf22097a29daa07931cf05c7e7deb83e35f838ed32e467c938aa932.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 70, + 536, + 295, + 583 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 536, + 295, + 583 + ], + "spans": [ + { + "bbox": [ + 70, + 536, + 295, + 583 + ], + "type": "interline_equation", + "content": "\\mathbf {v} _ {\\text {i n t e r v a l}} ^ {(i)} = \\underbrace {\\left(\\mathbf {v} _ {\\mathrm {c}} - \\mathbf {r} _ {\\mathrm {o}} ^ {(i)}\\right) ^ {\\intercal} \\ell^ {(i)}} _ {\\text {V o x e l z - d i s t a n c e}} \\cdot \\underbrace {\\frac {\\tan \\left(0 . 5 \\cdot \\theta_ {\\text {f o v - x}} ^ {(i)}\\right)}{0 . 5 \\cdot W ^ {(i)}}} _ {\\text {U n i t - d i s t a n c e p i x e l s i z e}}, \\tag {7b}", + "image_path": "7891be80ff92f06e7e35a3fffd0ebaa1fb9bfa125eeb1cdc9d49d2983789dff9.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 586, + 296, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 586, + 296, + 634 + ], + "spans": [ + { + "bbox": [ + 55, + 586, + 296, + 634 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 586, + 296, + 634 + ], + "type": "inline_equation", + "content": "\\ell" + }, + { + "bbox": [ + 55, + 586, + 296, + 634 + ], + "type": "text", + "content": " is camera lookat vector, " + }, + { + "bbox": [ + 55, + 586, + 296, + 634 + ], + "type": "inline_equation", + "content": "\\theta_{\\mathrm{fov - x}}" + }, + { + "bbox": [ + 55, + 586, + 296, + 634 + ], + "type": "text", + "content": " is camera horizontal field of view, and " + }, + { + "bbox": [ + 55, + 586, + 296, + 634 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 55, + 586, + 296, + 634 + ], + "type": "text", + "content": " is image width. The sampling rate indicates the estimated number of rays along the image's horizontal axis direction that may hit the voxel." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 639, + 167, + 650 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 639, + 167, + 650 + ], + "spans": [ + { + "bbox": [ + 55, + 639, + 167, + 650 + ], + "type": "text", + "content": "3.2.1. Scene Initialization" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "content": "Without employing an additional prior, we initialize all the parameters to constant. We start with volume density approaching zero by setting voxel raw density to a negative number " + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "inline_equation", + "content": "h_{\\mathrm{geo}}" + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "content": " so that the initial activated density " + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{expl}(h_{\\mathrm{geo}}) \\approx 0" + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "content": ". We set the SH coefficients to zero for" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 317, + 74, + 390, + 156 + ], + "blocks": [ + { + "bbox": [ + 317, + 74, + 390, + 156 + ], + "lines": [ + { + "bbox": [ + 317, + 74, + 390, + 156 + ], + "spans": [ + { + "bbox": [ + 317, + 74, + 390, + 156 + ], + "type": "image", + "image_path": "d27e95ea3cd38fe76c3b5b7b47305570c9c3cd3a9d55ca912138c2e92baed9d5.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 347, + 159, + 358, + 168 + ], + "lines": [ + { + "bbox": [ + 347, + 159, + 358, + 168 + ], + "spans": [ + { + "bbox": [ + 347, + 159, + 358, + 168 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 432, + 76, + 550, + 156 + ], + "blocks": [ + { + "bbox": [ + 432, + 76, + 550, + 156 + ], + "lines": [ + { + "bbox": [ + 432, + 76, + 550, + 156 + ], + "spans": [ + { + "bbox": [ + 432, + 76, + 550, + 156 + ], + "type": "image", + "image_path": "ee5739594973dabe5c4f9c288e4650a6b06d8665c3264b832b8191c9de6539c7.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 486, + 159, + 496, + 168 + ], + "lines": [ + { + "bbox": [ + 486, + 159, + 496, + 168 + ], + "spans": [ + { + "bbox": [ + 486, + 159, + 496, + 168 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 178, + 555, + 244 + ], + "lines": [ + { + "bbox": [ + 313, + 178, + 555, + 244 + ], + "spans": [ + { + "bbox": [ + 313, + 178, + 555, + 244 + ], + "type": "text", + "content": "Figure 5. Visualization of voxel sampling rate and grid layout initialization. (a) We visualize the voxel sampling rate defined in Eq. (7). (b) We depict the foreground main region and the background region under different shell levels. In unbounded scenes, we apply different grid layout initialization strategies for foreground and background regions. See Sec. 3.2.1 for details." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 252, + 554, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 252, + 554, + 289 + ], + "spans": [ + { + "bbox": [ + 313, + 252, + 554, + 289 + ], + "type": "text", + "content": "non-zero degrees and set the view-independent zero-degree component to yield gray color (i.e., intensity equal 0.5). We detail the Octree grid layout initialization in the following." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 294, + 554, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 294, + 554, + 354 + ], + "spans": [ + { + "bbox": [ + 313, + 294, + 554, + 354 + ], + "type": "text", + "content": "Bounded scenes. In case the scenes or the objects to reconstruct are enclosed in a known bounded region, we simply initialize the layout as a dense grid with " + }, + { + "bbox": [ + 313, + 294, + 554, + 354 + ], + "type": "inline_equation", + "content": "h_{1\\mathrm{v}}" + }, + { + "bbox": [ + 313, + 294, + 554, + 354 + ], + "type": "text", + "content": " levels and remove voxels unobserved by any training images. The number of voxels is " + }, + { + "bbox": [ + 313, + 294, + 554, + 354 + ], + "type": "inline_equation", + "content": "\\leq (2^{h_{1\\mathrm{v}}})^3" + }, + { + "bbox": [ + 313, + 294, + 554, + 354 + ], + "type": "text", + "content": " after initialization." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 360, + 555, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 360, + 555, + 588 + ], + "spans": [ + { + "bbox": [ + 313, + 360, + 555, + 588 + ], + "type": "text", + "content": "Unbounded scenes. For the unbounded scenes, we first split the space into the main and the unbounded background regions, depicted in Fig. 5b, each with a different heuristic. We use the training camera positions to determine a cuboid for the main region. The cuboid center is set to the average camera positions and the radius is set to the median distance between the cuboid center and the cameras. The same as the bounded scenes, we initialize a dense grid with " + }, + { + "bbox": [ + 313, + 360, + 555, + 588 + ], + "type": "inline_equation", + "content": "h_{\\mathrm{lv}}" + }, + { + "bbox": [ + 313, + 360, + 555, + 588 + ], + "type": "text", + "content": " levels for the main region. For the background region, we allocate " + }, + { + "bbox": [ + 313, + 360, + 555, + 588 + ], + "type": "inline_equation", + "content": "h_{\\mathrm{out}}" + }, + { + "bbox": [ + 313, + 360, + 555, + 588 + ], + "type": "text", + "content": " level of background shells enclosing the main region, which means that the radius of the entire scene is " + }, + { + "bbox": [ + 313, + 360, + 555, + 588 + ], + "type": "inline_equation", + "content": "2^{h_{\\mathrm{out}}}" + }, + { + "bbox": [ + 313, + 360, + 555, + 588 + ], + "type": "text", + "content": " of the main region. In each background shell level, we start with the coarsest voxel size, i.e., " + }, + { + "bbox": [ + 313, + 360, + 555, + 588 + ], + "type": "inline_equation", + "content": "4^3 - 2^3 = 56" + }, + { + "bbox": [ + 313, + 360, + 555, + 588 + ], + "type": "text", + "content": " voxels in each shell level. We then iteratively subdivide shell voxels with the highest sampling rate and remove voxels unobserved by any training cameras. The process repeats until the ratio of the number of voxels in the background and the main regions is " + }, + { + "bbox": [ + 313, + 360, + 555, + 588 + ], + "type": "inline_equation", + "content": "h_{\\mathrm{ratio}}" + }, + { + "bbox": [ + 313, + 360, + 555, + 588 + ], + "type": "text", + "content": ". The number of voxels is " + }, + { + "bbox": [ + 313, + 360, + 555, + 588 + ], + "type": "inline_equation", + "content": "\\leq (1 + h_{\\mathrm{ratio}})(2^{h_{\\mathrm{lv}}})^3" + }, + { + "bbox": [ + 313, + 360, + 555, + 588 + ], + "type": "text", + "content": " after initialization." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 596, + 489, + 608 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 596, + 489, + 608 + ], + "spans": [ + { + "bbox": [ + 313, + 596, + 489, + 608 + ], + "type": "text", + "content": "3.2.2. Adaptive Pruning and Subdivision" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 612, + 554, + 672 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 612, + 554, + 672 + ], + "spans": [ + { + "bbox": [ + 313, + 612, + 554, + 672 + ], + "type": "text", + "content": "The initialized grid layout only coarsely covers the entire scene that should be adaptively aligned to different levels-of-detail for the scene during the training progress. We apply the following two operations every " + }, + { + "bbox": [ + 313, + 612, + 554, + 672 + ], + "type": "inline_equation", + "content": "h_{\\text{every}}" + }, + { + "bbox": [ + 313, + 612, + 554, + 672 + ], + "type": "text", + "content": " training iterations to achieve this purpose." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "type": "text", + "content": "Pruning. We compute the maximum blending weight " + }, + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "type": "inline_equation", + "content": "(T_{i}\\alpha_{i})" + }, + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "type": "text", + "content": " from Eq. (1) of each voxel using all the training cameras. We remove voxels with maximum blending weight" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16191" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 132, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 132, + 84 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 132, + 84 + ], + "type": "text", + "content": "lower than " + }, + { + "bbox": [ + 55, + 72, + 132, + 84 + ], + "type": "inline_equation", + "content": "h_{\\mathrm{prune}}" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 89, + 296, + 137 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 89, + 296, + 137 + ], + "spans": [ + { + "bbox": [ + 55, + 89, + 296, + 137 + ], + "type": "text", + "content": "Subdivision. Our heuristic is that a voxel with a larger training loss gradient indicates that the voxel region requires finer voxels to model. Specifically, we accumulate the subdivision priority as the following:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 98, + 140, + 295, + 167 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 98, + 140, + 295, + 167 + ], + "spans": [ + { + "bbox": [ + 98, + 140, + 295, + 167 + ], + "type": "interline_equation", + "content": "\\mathbf {v} _ {\\text {p r i o r i t y}} = \\sum_ {\\mathbf {r} \\in R} \\left\\| \\alpha (\\mathbf {r}) \\cdot \\frac {\\partial \\mathcal {L} (\\mathbf {r})}{\\partial \\alpha (\\mathbf {r})} \\right\\|, \\tag {8}", + "image_path": "eac252fae811975aadc77d34c146dcc23174842da559e836ea3b62723830243e.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 169, + 296, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 169, + 296, + 312 + ], + "spans": [ + { + "bbox": [ + 55, + 169, + 296, + 312 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 169, + 296, + 312 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 55, + 169, + 296, + 312 + ], + "type": "text", + "content": " is the set of all training pixel rays throughout the " + }, + { + "bbox": [ + 55, + 169, + 296, + 312 + ], + "type": "inline_equation", + "content": "h_{\\text{every}}" + }, + { + "bbox": [ + 55, + 169, + 296, + 312 + ], + "type": "text", + "content": " iterations and " + }, + { + "bbox": [ + 55, + 169, + 296, + 312 + ], + "type": "inline_equation", + "content": "\\mathcal{L}(\\mathbf{r})" + }, + { + "bbox": [ + 55, + 169, + 296, + 312 + ], + "type": "text", + "content": " is the training loss of the ray. The gradient is weighted by alpha values contributed from the voxel to the ray. Higher " + }, + { + "bbox": [ + 55, + 169, + 296, + 312 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_{\\text{priority}}" + }, + { + "bbox": [ + 55, + 169, + 296, + 312 + ], + "type": "text", + "content": " indicates higher subdivision priority. To prevent voxels from overfitting few pixels, we set the priority to zero for voxels with maximum sampling rate lower than a sampling rate threshold " + }, + { + "bbox": [ + 55, + 169, + 296, + 312 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_{\\text{rate}} < 2h_{\\text{rate}}" + }, + { + "bbox": [ + 55, + 169, + 296, + 312 + ], + "type": "text", + "content": ". Finally, we select the voxels with priority above the top " + }, + { + "bbox": [ + 55, + 169, + 296, + 312 + ], + "type": "inline_equation", + "content": "h_{\\text{percent}}" + }, + { + "bbox": [ + 55, + 169, + 296, + 312 + ], + "type": "text", + "content": " percent to subdivide, i.e., the total number of voxels is increased by " + }, + { + "bbox": [ + 55, + 169, + 296, + 312 + ], + "type": "inline_equation", + "content": "(h_{\\text{percent}} \\cdot (8 - 1))" + }, + { + "bbox": [ + 55, + 169, + 296, + 312 + ], + "type": "text", + "content": " percent. Note that we only keep the leaf nodes in the Octree layout so we remove the source voxels once they are subdivided." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 312, + 296, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 312, + 296, + 445 + ], + "spans": [ + { + "bbox": [ + 55, + 312, + 296, + 445 + ], + "type": "text", + "content": "When voxels are pruned and subdivided, the voxel Spherical Harmonic (SH) coefficients and the grid point densities need to be updated accordingly. The SH coefficients are simply pruned together with voxels and duplicated to the subdivided children voxels. Grid point densities are slightly more complex as the eight voxel corner grid points are shared between adjacent voxels (Fig. 2). We remove a grid point only when it does not belong to any voxel corners. When subdividing, we use trilinear interpolation to compute the densities of the new grid points. The duplicated grid points are merged and their densities are averaged." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 449, + 185, + 460 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 449, + 185, + 460 + ], + "spans": [ + { + "bbox": [ + 55, + 449, + 185, + 460 + ], + "type": "text", + "content": "3.2.3. Optimization objectives" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 463, + 296, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 463, + 296, + 499 + ], + "spans": [ + { + "bbox": [ + 55, + 463, + 296, + 499 + ], + "type": "text", + "content": "We use MSE and SSIM as the photometric loss between the rendered and the ground truth images. The overall training objective is summarized as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 65, + 506, + 295, + 534 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 506, + 295, + 534 + ], + "spans": [ + { + "bbox": [ + 65, + 506, + 295, + 534 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} = \\mathcal {L} _ {\\mathrm {m s e}} + \\lambda_ {\\mathrm {s s i m}} \\mathcal {L} _ {\\mathrm {s s i m}} \\\\ + \\lambda_ {\\mathrm {T}} \\mathcal {L} _ {\\mathrm {T}} + \\lambda_ {\\text {d i s t}} \\mathcal {L} _ {\\text {d i s t}} + \\lambda_ {\\mathrm {R}} \\mathcal {L} _ {\\mathrm {R}} + \\lambda_ {\\mathrm {t v}} \\mathcal {L} _ {\\mathrm {t v}}, \\tag {9} \\\\ \\end{array}", + "image_path": "9e0722b13cc6d18d572dc308219a9726a62a245bfe404709deb8d281c0c85125.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 538, + 296, + 610 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 538, + 296, + 610 + ], + "spans": [ + { + "bbox": [ + 55, + 538, + 296, + 610 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 538, + 296, + 610 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 538, + 296, + 610 + ], + "type": "text", + "content": " are the loss weights, " + }, + { + "bbox": [ + 55, + 538, + 296, + 610 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{T}}" + }, + { + "bbox": [ + 55, + 538, + 296, + 610 + ], + "type": "text", + "content": " encourages the final ray transmittances to be either zero or one, " + }, + { + "bbox": [ + 55, + 538, + 296, + 610 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{dist}}" + }, + { + "bbox": [ + 55, + 538, + 296, + 610 + ], + "type": "text", + "content": " is the distortion loss [2], " + }, + { + "bbox": [ + 55, + 538, + 296, + 610 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{R}}" + }, + { + "bbox": [ + 55, + 538, + 296, + 610 + ], + "type": "text", + "content": " is the per-point rgb loss [43], and " + }, + { + "bbox": [ + 55, + 538, + 296, + 610 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{tv}}" + }, + { + "bbox": [ + 55, + 538, + 296, + 610 + ], + "type": "text", + "content": " is the total variation loss on the sparse density grid. In mesh extraction task, we also add the depth-normal consistency loss from 2DGS [16]:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 66, + 617, + 295, + 631 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 617, + 295, + 631 + ], + "spans": [ + { + "bbox": [ + 66, + 617, + 295, + 631 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {m e s h}} = \\lambda_ {\\mathrm {n - d m e a n}} \\mathcal {L} _ {\\mathrm {n - d m e a n}} + \\lambda_ {\\mathrm {n - d m e d}} \\mathcal {L} _ {\\mathrm {n - d m e d}}, \\tag {10}", + "image_path": "5838e2ec5ee3f3ea4fe7705f142f70c9a2cf06661c4c0af72e3ea46a98dec365.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 634, + 296, + 670 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 634, + 296, + 670 + ], + "spans": [ + { + "bbox": [ + 55, + 634, + 296, + 670 + ], + "type": "text", + "content": "where both losses encourage alignment between rendered normals and depth-derived normals from mean and median depth. More details are in the supplementary materials." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 674, + 290, + 687 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 674, + 290, + 687 + ], + "spans": [ + { + "bbox": [ + 55, + 674, + 290, + 687 + ], + "type": "text", + "content": "3.2.4. Sparse-voxel TSDF Fusion and Marching Cubes" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": "Our sparse voxels can be seamlessly integrated with grid-based algorithms. To extract a mesh, we implement March-" + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 316, + 71, + 373, + 155 + ], + "blocks": [ + { + "bbox": [ + 316, + 71, + 373, + 155 + ], + "lines": [ + { + "bbox": [ + 316, + 71, + 373, + 155 + ], + "spans": [ + { + "bbox": [ + 316, + 71, + 373, + 155 + ], + "type": "image", + "image_path": "e8d58f37de38ce27a4313b2c32656c80caa7f7355eac5dd81e3e7120f3af224a.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 162, + 554, + 206 + ], + "lines": [ + { + "bbox": [ + 313, + 162, + 554, + 206 + ], + "spans": [ + { + "bbox": [ + 313, + 162, + 554, + 206 + ], + "type": "text", + "content": "Figure 6. A qualitative comparison with 3DGS [20]. Our result here corresponding to the base version in Tab. 1. We achieve similar visual quality comparing to 3DGS. Note that 3DGS use the coarse geometry from SfM while we do not rely on this prior." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 374, + 72, + 430, + 156 + ], + "blocks": [ + { + "bbox": [ + 374, + 72, + 430, + 156 + ], + "lines": [ + { + "bbox": [ + 374, + 72, + 430, + 156 + ], + "spans": [ + { + "bbox": [ + 374, + 72, + 430, + 156 + ], + "type": "image", + "image_path": "f2cdbeb90dd4abee73ab8b8c63a6013e6e04bbe3fc68c6c5845dbcd3b53d9d7f.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 431, + 72, + 487, + 156 + ], + "blocks": [ + { + "bbox": [ + 431, + 72, + 487, + 156 + ], + "lines": [ + { + "bbox": [ + 431, + 72, + 487, + 156 + ], + "spans": [ + { + "bbox": [ + 431, + 72, + 487, + 156 + ], + "type": "image", + "image_path": "851623337ba4d253a469d16b9ea9ae8fe9f3869f701e1b6b534e627b8293da51.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 488, + 72, + 509, + 156 + ], + "blocks": [ + { + "bbox": [ + 488, + 72, + 509, + 156 + ], + "lines": [ + { + "bbox": [ + 488, + 72, + 509, + 156 + ], + "spans": [ + { + "bbox": [ + 488, + 72, + 509, + 156 + ], + "type": "image", + "image_path": "caa7d463f4514c280dbd6282de13a96454543cbace5d77917e5c84f7ec808758.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 509, + 72, + 529, + 156 + ], + "blocks": [ + { + "bbox": [ + 509, + 72, + 529, + 156 + ], + "lines": [ + { + "bbox": [ + 509, + 72, + 529, + 156 + ], + "spans": [ + { + "bbox": [ + 509, + 72, + 529, + 156 + ], + "type": "image", + "image_path": "0ec9cc628271902213e2cb573f03effa40ef87426d03421724d02ccbd8e96e5c.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 531, + 72, + 552, + 156 + ], + "blocks": [ + { + "bbox": [ + 531, + 72, + 552, + 156 + ], + "lines": [ + { + "bbox": [ + 531, + 72, + 552, + 156 + ], + "spans": [ + { + "bbox": [ + 531, + 72, + 552, + 156 + ], + "type": "image", + "image_path": "e64aaead89acfcf17a5539b96d4c9ff713b8c2a4ae898c578e48a0290765ace3.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 316, + 217, + 395, + 304 + ], + "blocks": [ + { + "bbox": [ + 316, + 217, + 395, + 304 + ], + "lines": [ + { + "bbox": [ + 316, + 217, + 395, + 304 + ], + "spans": [ + { + "bbox": [ + 316, + 217, + 395, + 304 + ], + "type": "image", + "image_path": "a7c38bd876137a04723797d560bb11eca4c0b1c529af8ee98bd7f065f00553e6.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 311, + 554, + 366 + ], + "lines": [ + { + "bbox": [ + 313, + 311, + 554, + 366 + ], + "spans": [ + { + "bbox": [ + 313, + 311, + 554, + 366 + ], + "type": "text", + "content": "Figure 7. Visualization of the reconstructed surface. We show the rendering images, normal maps, and the final meshes on Tanks&Temples and DTU datasets. Note that we only model the scene with density field and do not use the coarse geometry prior from SfM sparse points in this work." + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 395, + 218, + 474, + 304 + ], + "blocks": [ + { + "bbox": [ + 395, + 218, + 474, + 304 + ], + "lines": [ + { + "bbox": [ + 395, + 218, + 474, + 304 + ], + "spans": [ + { + "bbox": [ + 395, + 218, + 474, + 304 + ], + "type": "image", + "image_path": "f06a010990d5fb1e31ed80fe85bfdc08b7cd83bb1863c8f34e06f5872d098a1f.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 474, + 218, + 553, + 304 + ], + "blocks": [ + { + "bbox": [ + 474, + 218, + 553, + 304 + ], + "lines": [ + { + "bbox": [ + 474, + 218, + 553, + 304 + ], + "spans": [ + { + "bbox": [ + 474, + 218, + 553, + 304 + ], + "type": "image", + "image_path": "6823587515660a834d4214405ce47f47d4325b980720fb75ff8393ff3084ee6f.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 388, + 554, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 388, + 554, + 483 + ], + "spans": [ + { + "bbox": [ + 313, + 388, + 554, + 483 + ], + "type": "text", + "content": "ing Cubes [27] to extract the triangles of an isosurface over density from the sparse voxels. The duplicated vertices from adjacent voxels are merged to produce a unique set of vertices. When the adjacent voxels belong to different Octree levels, the extracted triangle may not be connected as the density field is not continuous for voxels in different levels. Such discontinuities can be removed by simply subdividing all voxels to the finest level." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 484, + 554, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 484, + 554, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 484, + 554, + 605 + ], + "type": "text", + "content": "Deciding the target level set for extracting the isosurface can be tricky for the density field. Instead, we implement sparse-voxel TSDF-Fusion [7, 8, 36] to compute the truncated signed distance values of the sparse grid points. We can then directly extract the surface of the zero-level set using the former sparse-voxel Marching Cubes. Future extensions of our method could directly model signed distance fields following NeuS [47] and VolSDF [53]. The sparse-voxel TSDF-Fusion can still be beneficial to directly initialize our sparse voxel representation from sensor depth." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 616, + 395, + 629 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 616, + 395, + 629 + ], + "spans": [ + { + "bbox": [ + 313, + 616, + 395, + 629 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 635, + 447, + 648 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 635, + 447, + 648 + ], + "spans": [ + { + "bbox": [ + 313, + 635, + 447, + 648 + ], + "type": "text", + "content": "4.1. Implementation Details" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "text", + "content": "We use the following implementation details except stated otherwise. We start the optimization from empty space with raw density set to " + }, + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "inline_equation", + "content": "h_{\\mathrm{geo}} = -10" + }, + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "text", + "content": ". The initial Octree level is " + }, + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "inline_equation", + "content": "h_{\\mathrm{lv}} = 6" + }, + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "text", + "content": " (i.e., " + }, + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "inline_equation", + "content": "64^{3}" + }, + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "text", + "content": " voxels) for the bounded scenes and the foreground main region of the unbounded scenes. To model" + } + ] + } + ], + "index": 28 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16192" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 70, + 294, + 224 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 294, + 224 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 294, + 224 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 294, + 224 + ], + "type": "table", + "html": "
MethodMip-NeRF360 dataset
FPS↑Tr. Time↓LPIPS↓PSNR↑SSIM↑
NeRF [30]<1~day0.45123.850.605
M-NeRF [1]<1~day0.44124.040.616
M-NeRF360 [2]<1~day0.23727.690.792
Zip-NeRF [3]†<1~hrs0.18728.550.828
Plenoxels [11]<10~30m0.46323.080.626
INGP [32]~10~5m0.30225.680.705
3DGS [20]†13124m0.21627.450.815
Ours fast-rend2589m0.21026.870.804
Ours fast-train1314.5m0.19927.080.816
Ours12115m0.18527.330.822
", + "image_path": "24ab81e445e67f37047e83605b3b962d07e05809cc85260a06ddcdb14bc03292.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 57, + 224, + 148, + 233 + ], + "lines": [ + { + "bbox": [ + 57, + 224, + 148, + 233 + ], + "spans": [ + { + "bbox": [ + 57, + 224, + 148, + 233 + ], + "type": "text", + "content": "† Re-evaluate on our machine." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_footnote" + }, + { + "bbox": [ + 56, + 235, + 295, + 278 + ], + "lines": [ + { + "bbox": [ + 56, + 235, + 295, + 278 + ], + "spans": [ + { + "bbox": [ + 56, + 235, + 295, + 278 + ], + "type": "text", + "content": "Table 1. Novel-view synthesis results comparison on MipNeRF360 dataset [2]. The results are averaged from 4 indoor scenes and 5 outdoor scenes. 3DGS uses the sparse points prior from COLMAP [41], whereas the other methods and ours do not." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 57, + 292, + 293, + 377 + ], + "blocks": [ + { + "bbox": [ + 57, + 292, + 293, + 377 + ], + "lines": [ + { + "bbox": [ + 57, + 292, + 293, + 377 + ], + "spans": [ + { + "bbox": [ + 57, + 292, + 293, + 377 + ], + "type": "table", + "html": "
MethodTanks&TemplesDeep Blending
FPS↑LPIPS↓PSNR↑FPS↑LPIPS↓PSNR↑
Plenoxels [11]130.37921.08110.51023.06
INGP [32]140.30521.9230.39024.96
M-NeRF360 [2]<10.25722.22<10.24529.40
3DGS [20]†1800.17623.751400.24429.60
Ours1250.14423.043660.22829.84
", + "image_path": "2cbe86ae046fa0762e44379bbaf66191b7dc7e45dbb6c1ac4a49375766a695ff.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 59, + 378, + 136, + 385 + ], + "lines": [ + { + "bbox": [ + 59, + 378, + 136, + 385 + ], + "spans": [ + { + "bbox": [ + 59, + 378, + 136, + 385 + ], + "type": "text", + "content": "† Re-evaluate on our machine." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_footnote" + }, + { + "bbox": [ + 56, + 387, + 295, + 420 + ], + "lines": [ + { + "bbox": [ + 56, + 387, + 295, + 420 + ], + "spans": [ + { + "bbox": [ + 56, + 387, + 295, + 420 + ], + "type": "text", + "content": "Table 2. Comparison on Tanks&Temples [21] and Deep Blending [15] datasets. We follow 3DGS [20] to use two outdoor scenes from Tanks&Temples and two indoor scenes from Deep Blending." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 57, + 428, + 136, + 517 + ], + "blocks": [ + { + "bbox": [ + 57, + 428, + 136, + 517 + ], + "lines": [ + { + "bbox": [ + 57, + 428, + 136, + 517 + ], + "spans": [ + { + "bbox": [ + 57, + 428, + 136, + 517 + ], + "type": "image", + "image_path": "5e9d8977242621a3b6547198a855d2e243fb3c60f0ed9a2bf96b41490f0b22d1.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 92, + 517, + 100, + 524 + ], + "lines": [ + { + "bbox": [ + 92, + 517, + 100, + 524 + ], + "spans": [ + { + "bbox": [ + 92, + 517, + 100, + 524 + ], + "type": "text", + "content": "GT" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 530, + 295, + 586 + ], + "lines": [ + { + "bbox": [ + 55, + 530, + 295, + 586 + ], + "spans": [ + { + "bbox": [ + 55, + 530, + 295, + 586 + ], + "type": "text", + "content": "Figure 8. Failure case. On scenes with severe exposure variation of training views, our method struggles and produces clear boundary of different brightnesses and allocates many floaters. 3DGS on the other hand is less sensitive to photometric variation of GT. This explains our worse PSNR and FPS on Tanks&Temples in Tab. 2." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 136, + 428, + 214, + 517 + ], + "blocks": [ + { + "bbox": [ + 136, + 428, + 214, + 517 + ], + "lines": [ + { + "bbox": [ + 136, + 428, + 214, + 517 + ], + "spans": [ + { + "bbox": [ + 136, + 428, + 214, + 517 + ], + "type": "image", + "image_path": "222bcc7c2e59533a270851ebb9c140da2e79d6cf75f99240d4d09ad3a0c54e15.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 170, + 518, + 181, + 524 + ], + "lines": [ + { + "bbox": [ + 170, + 518, + 181, + 524 + ], + "spans": [ + { + "bbox": [ + 170, + 518, + 181, + 524 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 215, + 428, + 294, + 517 + ], + "blocks": [ + { + "bbox": [ + 215, + 428, + 294, + 517 + ], + "lines": [ + { + "bbox": [ + 215, + 428, + 294, + 517 + ], + "spans": [ + { + "bbox": [ + 215, + 428, + 294, + 517 + ], + "type": "image", + "image_path": "130f561e738b9d0c977e30ff2cc1e61da18ccba322897372a54334905d12ba96.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 247, + 517, + 261, + 524 + ], + "lines": [ + { + "bbox": [ + 247, + 517, + 261, + 524 + ], + "spans": [ + { + "bbox": [ + 247, + 517, + 261, + 524 + ], + "type": "text", + "content": "3DGS" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 54, + 594, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 594, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 594, + 295, + 714 + ], + "type": "text", + "content": "unbounded scenes, we use " + }, + { + "bbox": [ + 54, + 594, + 295, + 714 + ], + "type": "inline_equation", + "content": "h_{\\mathrm{out}} = 5" + }, + { + "bbox": [ + 54, + 594, + 295, + 714 + ], + "type": "text", + "content": " background shell levels with " + }, + { + "bbox": [ + 54, + 594, + 295, + 714 + ], + "type": "inline_equation", + "content": "h_{\\mathrm{ratio}} = 2" + }, + { + "bbox": [ + 54, + 594, + 295, + 714 + ], + "type": "text", + "content": " times the number of foreground voxels (Sec. 3.2.1 and Fig. 5). The early ray stopping threshold is set to " + }, + { + "bbox": [ + 54, + 594, + 295, + 714 + ], + "type": "inline_equation", + "content": "h_{\\mathrm{T}} = 1\\mathrm{e} - 4" + }, + { + "bbox": [ + 54, + 594, + 295, + 714 + ], + "type": "text", + "content": " and the supersampling scale is set to " + }, + { + "bbox": [ + 54, + 594, + 295, + 714 + ], + "type": "inline_equation", + "content": "h_{\\mathrm{ss}} = 1.5" + }, + { + "bbox": [ + 54, + 594, + 295, + 714 + ], + "type": "text", + "content": ". Inside each voxel, we sample " + }, + { + "bbox": [ + 54, + 594, + 295, + 714 + ], + "type": "inline_equation", + "content": "K = 1" + }, + { + "bbox": [ + 54, + 594, + 295, + 714 + ], + "type": "text", + "content": " point for novel-view synthesis and " + }, + { + "bbox": [ + 54, + 594, + 295, + 714 + ], + "type": "inline_equation", + "content": "K = 3" + }, + { + "bbox": [ + 54, + 594, + 295, + 714 + ], + "type": "text", + "content": " points for mesh reconstruction task. We train our model for 20K iterations. The voxels are pruned and subdivided every " + }, + { + "bbox": [ + 54, + 594, + 295, + 714 + ], + "type": "inline_equation", + "content": "h_{\\mathrm{every}} = 1,000" + }, + { + "bbox": [ + 54, + 594, + 295, + 714 + ], + "type": "text", + "content": " iterations. The pruning threshold " + }, + { + "bbox": [ + 54, + 594, + 295, + 714 + ], + "type": "inline_equation", + "content": "h_{\\mathrm{prune}}" + }, + { + "bbox": [ + 54, + 594, + 295, + 714 + ], + "type": "text", + "content": " is linearly scale from 0.0001 to 0.05 and the subdivision percentage is " + }, + { + "bbox": [ + 54, + 594, + 295, + 714 + ], + "type": "inline_equation", + "content": "h_{\\mathrm{percent}} = 5" + }, + { + "bbox": [ + 54, + 594, + 295, + 714 + ], + "type": "text", + "content": ". More" + } + ] + } + ], + "index": 15 + }, + { + "type": "table", + "bbox": [ + 316, + 70, + 553, + 126 + ], + "blocks": [ + { + "bbox": [ + 316, + 70, + 553, + 126 + ], + "lines": [ + { + "bbox": [ + 316, + 70, + 553, + 126 + ], + "spans": [ + { + "bbox": [ + 316, + 70, + 553, + 126 + ], + "type": "table", + "html": "
MethodPeak GPU mem.↓Model size↓FPS at higher res.↑
1x2x3x
3DGS1.8GB0.7GB1316939
Ours3.9GB1.8GB12110369
", + "image_path": "9b69265384f9ed7079cbc0cfc04742f36d73ff8c343f93bc14c0695337105c0f.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 313, + 127, + 553, + 159 + ], + "lines": [ + { + "bbox": [ + 313, + 127, + 553, + 159 + ], + "spans": [ + { + "bbox": [ + 313, + 127, + 553, + 159 + ], + "type": "text", + "content": "Table 3. Memory, model size, and high-res FPS. The results are averaged on Mip-NeRF360 dataset. The standard 1x evaluation resolution have about 1–1.6M pixels per frame." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 175, + 471, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 175, + 471, + 186 + ], + "spans": [ + { + "bbox": [ + 313, + 175, + 471, + 186 + ], + "type": "text", + "content": "details are provided in the supplement." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 187, + 553, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 187, + 553, + 270 + ], + "spans": [ + { + "bbox": [ + 313, + 187, + 553, + 270 + ], + "type": "text", + "content": "We use the standard LPIPS [56], SSIM, and PSNR metrics to evaluate novel-view quality. We directly employ the LPIPS implementation from prior work [20], which uses [0, 1] image intensity range with VGG [42] network. For mesh accuracy, we follow the benchmarks to use F-score and chamfer distance. We use the test-set images to measure the FPS on a desktop computer with a 3090 Ti GPU." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 289, + 435, + 301 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 289, + 435, + 301 + ], + "spans": [ + { + "bbox": [ + 313, + 289, + 435, + 301 + ], + "type": "text", + "content": "4.2. Novel-view Synthesis" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 307, + 553, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 307, + 553, + 378 + ], + "spans": [ + { + "bbox": [ + 313, + 307, + 553, + 378 + ], + "type": "text", + "content": "In Tab. 1, we show the quantitative comparison on the MipNeRF-360 dataset. We also provide a fast-rendering variant with " + }, + { + "bbox": [ + 313, + 307, + 553, + 378 + ], + "type": "inline_equation", + "content": ">2\\mathrm{x}" + }, + { + "bbox": [ + 313, + 307, + 553, + 378 + ], + "type": "text", + "content": " FPS by setting " + }, + { + "bbox": [ + 313, + 307, + 553, + 378 + ], + "type": "inline_equation", + "content": "h_{\\mathrm{ss}} = 1.5" + }, + { + "bbox": [ + 313, + 307, + 553, + 378 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 307, + 553, + 378 + ], + "type": "inline_equation", + "content": "h_{\\mathrm{prune}} = 0.15" + }, + { + "bbox": [ + 313, + 307, + 553, + 378 + ], + "type": "text", + "content": ", and a fast-training variant with " + }, + { + "bbox": [ + 313, + 307, + 553, + 378 + ], + "type": "inline_equation", + "content": ">3\\mathrm{x}" + }, + { + "bbox": [ + 313, + 307, + 553, + 378 + ], + "type": "text", + "content": " training speedup by scaling all iteration related hyperparameter by 0.3. Both fast variants only reduce quality moderately." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 379, + 554, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 379, + 554, + 569 + ], + "spans": [ + { + "bbox": [ + 313, + 379, + 554, + 569 + ], + "type": "text", + "content": "Our rendering speed is comparable to 3DGS on average. However, FPS varies significantly across scenes (see supplementary for detailed per-scene results). Several differences impact speed between our method and 3DGS. For instance, 3DGS sorts 32-bit floats, while we sort 48-bit Morton codes (16 levels, each with 3 bits) for primitive ordering. We also avoid the overhead of computing inverse covariance matrices from quaternions and scaling parameters. We directly intersect rays with voxels in 3D space, whereas 3DGS approximates the projection of 3D Gaussians to 2D via a linear affine transform. Finally, Gaussian distributions decay gradually to zero, while our post-activation [18, 43] voxel density field can be arbitrarily sharp. As a result, the average numbers of primitives contributing to a pixel intensity is 63 for 3DGS and 27 for us. The influence of these factors on rendering speed is scene-dependent." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 570, + 554, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 554, + 677 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 554, + 677 + ], + "type": "text", + "content": "Regarding rendering quality metrics, our method achieves significantly better LPIPS than 3DGS and even surpasses Zip-NeRF. This can be explained by the qualitative comparison in Fig. 6 where our method recover finer details. The SSIM comparison results with 3DGS is more scene-dependent. SSIM comparisons with 3DGS are more scene-dependent, while PSNR tends to favor 3DGS, as it prefers smoother rendering in uncertain regions. More visual comparisons are in the supplementary." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "content": "Our method uses much more voxel primitives than Gaussian, resulting in larger model size and requires more GPU memory as shown in Tab. 3. Interestingly, our high FPS is" + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16193" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 57, + 70, + 294, + 120 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 294, + 120 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 294, + 120 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 294, + 120 + ], + "type": "table", + "html": "
Resolution of main2563512310243adaptivePlenoxels 6403
LPIPS↓0.4440.3260.2000.452
PSNR↑23.9825.37OOM28.0123.29
FPS↑457190171<10
", + "image_path": "cb96906182bd6ece656ed5f686f26c632e6a59bdabf19cb50d95484101a52747.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 198, + 295, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 198, + 295, + 223 + ], + "spans": [ + { + "bbox": [ + 55, + 198, + 295, + 223 + ], + "type": "text", + "content": "more scalable to higher resolution rendering perhaps due to the fewer contributing primitives per pixel." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 223, + 296, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 223, + 296, + 319 + ], + "spans": [ + { + "bbox": [ + 55, + 223, + 296, + 319 + ], + "type": "text", + "content": "We compare results on two more datasets in Tab. 2. On Deep Blending [15] dataset, we achieve much better speed and quality than 3DGS. On Tanks&Temples dataset [21], despite having better LPIPS, our PSNR and FPS are worse than 3DGS. The failure results are shown in Fig. 8 where the training views have large exposure variation. As a result, our method produces a clear brightness boundary in the scene with many floaters that slow down our rendering." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 337, + 156, + 348 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 337, + 156, + 348 + ], + "spans": [ + { + "bbox": [ + 55, + 337, + 156, + 348 + ], + "type": "text", + "content": "4.3. Ablation Studies" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 354, + 296, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 354, + 296, + 653 + ], + "spans": [ + { + "bbox": [ + 55, + 354, + 296, + 653 + ], + "type": "text", + "content": "The rasterization process improves the rendering speed of sparse voxels, while the key to achieve high-quality results is the adaptive voxel size for different levels of detail. We show an ablation experiment in Tab. 4. Instead of subdividing voxels adaptively, the uniform voxel size variant subdivides all the voxels at certain training iterations until the grid resolution of the foreground main region reaches " + }, + { + "bbox": [ + 55, + 354, + 296, + 653 + ], + "type": "inline_equation", + "content": "256^{3}" + }, + { + "bbox": [ + 55, + 354, + 296, + 653 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 354, + 296, + 653 + ], + "type": "inline_equation", + "content": "512^{3}" + }, + { + "bbox": [ + 55, + 354, + 296, + 653 + ], + "type": "text", + "content": ", or " + }, + { + "bbox": [ + 55, + 354, + 296, + 653 + ], + "type": "inline_equation", + "content": "1024^{3}" + }, + { + "bbox": [ + 55, + 354, + 296, + 653 + ], + "type": "text", + "content": ". We also align the background and the pruning setup where the details are deferred to the supplementary. As shown in the table, the quality of uniform voxel size in " + }, + { + "bbox": [ + 55, + 354, + 296, + 653 + ], + "type": "inline_equation", + "content": "512^{3}" + }, + { + "bbox": [ + 55, + 354, + 296, + 653 + ], + "type": "text", + "content": " resolution is much worse than the adaptive voxel size under similar FPS. Our machine with 24GB GPU memory fails to reach " + }, + { + "bbox": [ + 55, + 354, + 296, + 653 + ], + "type": "inline_equation", + "content": "1024^{3}" + }, + { + "bbox": [ + 55, + 354, + 296, + 653 + ], + "type": "text", + "content": " grid resolution despite the voxels being pruned and sparse. Plenoxels [11] renders sparse voxels by ray casting and sampling following NeRF. Their FPS is significantly lower than our rasterization-based sparse voxel rendering. As they use a dense 3D pointer grid to support point sampling on rays, the model scalability for the unbounded background region is thus limited. To workaround, they apply a multi-sphere image with 64 layers to model the background. Conversely, we set the scene size as " + }, + { + "bbox": [ + 55, + 354, + 296, + 653 + ], + "type": "inline_equation", + "content": "32\\mathrm{x}" + }, + { + "bbox": [ + 55, + 354, + 296, + 653 + ], + "type": "text", + "content": " larger (" + }, + { + "bbox": [ + 55, + 354, + 296, + 653 + ], + "type": "inline_equation", + "content": "h_{\\mathrm{out}} = 5" + }, + { + "bbox": [ + 55, + 354, + 296, + 653 + ], + "type": "text", + "content": ") than the foreground cuboid and direct model the entire scene by our sparse voxels. As a result, our uniform size variant with " + }, + { + "bbox": [ + 55, + 354, + 296, + 653 + ], + "type": "inline_equation", + "content": "512^{3}" + }, + { + "bbox": [ + 55, + 354, + 296, + 653 + ], + "type": "text", + "content": " foreground resolution still outperforms Plenoxels with " + }, + { + "bbox": [ + 55, + 354, + 296, + 653 + ], + "type": "inline_equation", + "content": "640^{3}" + }, + { + "bbox": [ + 55, + 354, + 296, + 653 + ], + "type": "text", + "content": " grid." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": "In our rasterizer, we use direction-dependent Morton order to ensure the sorting of voxels from different levels is correct. However, like the case in 3DGS, the popping artifact by the incorrect order is not a major factor in the numerical results. Instead, we provide more visualization in" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 316, + 70, + 553, + 173 + ], + "blocks": [ + { + "bbox": [ + 55, + 121, + 295, + 187 + ], + "lines": [ + { + "bbox": [ + 55, + 121, + 295, + 187 + ], + "spans": [ + { + "bbox": [ + 55, + 121, + 295, + 187 + ], + "type": "text", + "content": "Table 4. Ablation experiments of adaptive and uniform voxel sizes. The results are evaluated on the indoor bonsai and the outdoor bicycle scenes from MipNeRF-360 dataset. The resolutions at the first row indicate the final grid resolution of the main foreground cuboid. Plenxoels is the previous fully-explicit voxel grid approach. Please refer to Sec. 4.3 for more discussion." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 316, + 70, + 553, + 173 + ], + "lines": [ + { + "bbox": [ + 316, + 70, + 553, + 173 + ], + "spans": [ + { + "bbox": [ + 316, + 70, + 553, + 173 + ], + "type": "table", + "html": "
MethodGeom.Init.Tansk&TemplesDTU
F-score↑Tr. time↓Cf.↓Tr. time↓
NeRF [30]DensityRandom--1.49hrs
NeuS [47]SDFSphere0.38hrs0.84hrs
Voxurf [51]SDFSphere--0.7615m
Neuralangelo [24]SDFSphere0.50hrs0.61hrs
3DGS [20]GaussiansSfM pts.0.1914m1.9611m
SuGaR [14]GaussiansSfM pts.0.09~1h1.33~1h
2DGS [16]GaussiansSfM pts.0.3216m0.8011m
OursDensityConstant0.4011m0.765m
", + "image_path": "8885cb64f54161fd086768fff6adaffd1d225ab6ae826bd24a9be6aadb45f0bd.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 174, + 555, + 217 + ], + "lines": [ + { + "bbox": [ + 313, + 174, + 555, + 217 + ], + "spans": [ + { + "bbox": [ + 313, + 174, + 555, + 217 + ], + "type": "text", + "content": "Table 5. Mesh results comparison on the Tanks&Temples [21] and DTU [17] datasets. Our method with volume density and constant field initialization achieves good accuracy-time trade-off on the unbounded-level and object-level datasets." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 228, + 553, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 228, + 553, + 252 + ], + "spans": [ + { + "bbox": [ + 313, + 228, + 553, + 252 + ], + "type": "text", + "content": "the supplementary videos to show its effect. See more ablation studies in the supplementary materials." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 258, + 437, + 270 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 258, + 437, + 270 + ], + "spans": [ + { + "bbox": [ + 313, + 258, + 437, + 270 + ], + "type": "text", + "content": "4.4. Mesh Reconstruction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 276, + 555, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 276, + 555, + 408 + ], + "spans": [ + { + "bbox": [ + 313, + 276, + 555, + 408 + ], + "type": "text", + "content": "We extract mesh by adapting TSDF-Fusion [34] and Marching Cubes [27] into our sparse voxels. The results comparison on the large-scale Tanks&Temples and the object-scale DTU datasets is provided in Tab. 5, where our method achieves a good balance of accuracy and training time on both datasets. See qualitative examples in Fig. 7 and the supplementary materials. We find that our current method tends to produce unnecessary geometric bumps for texture details. Future work could improve surface smoothness by adapting our density field for direct signed distance field modeling or incorporating SfM sparse points prior." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 415, + 424, + 426 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 415, + 424, + 426 + ], + "spans": [ + { + "bbox": [ + 313, + 415, + 424, + 426 + ], + "type": "text", + "content": "4.5. 2D Feature Fusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 432, + 554, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 432, + 554, + 515 + ], + "spans": [ + { + "bbox": [ + 313, + 432, + 554, + 515 + ], + "type": "text", + "content": "We showcase some results in Fig. 1, where we fuse 2D semantic segmentation and high-dimensional foundation features into our sparse voxels using Voxel Pooling and Volume Fusion. The multi-view ensemble smoothes the inconsistent 2D predictions, while our detailed 3D geometry enables higher-resolution rendering of 2D features. See more results in the released code." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 526, + 388, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 526, + 388, + 538 + ], + "spans": [ + { + "bbox": [ + 313, + 526, + 388, + 538 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 545, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 555, + 713 + ], + "type": "text", + "content": "This work presents a novel differentiable radiance field rendering system that integrates an efficient rasterizer with a multi-level sparse voxel scene representation. We reconstruct a scene from the multi-view input images by adaptively fitting the sparse voxels into different levels-of-detail. The results reveal that fully explicit voxel models, without neural networks or Gaussians, can achieve state-of-the-art comparable novel-view rendering speed and quality. The key breakthrough is overcoming the scalability constraint of voxel grids with adaptive sparse voxels and improving rendering speed through rasterization. We believe that incorporating our method with classical 3D processing algorithm [7, 8, 19, 26, 36] and 3D neural network [6, 13, 50] are promising future directions." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16194" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "type": "text", + "content": "[1] Jonathan T. Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P. Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In ICCV, 2021. 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 137, + 296, + 169 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 137, + 296, + 169 + ], + "spans": [ + { + "bbox": [ + 61, + 137, + 296, + 169 + ], + "type": "text", + "content": "[2] Jonathan T. Barron, Ben Mildenhall, Dor Verbin, Pratul P. Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In CVPR, 2022. 6, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 171, + 295, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 171, + 295, + 203 + ], + "spans": [ + { + "bbox": [ + 62, + 171, + 295, + 203 + ], + "type": "text", + "content": "[3] Jonathan T. Barron, Ben Mildenhall, Dor Verbin, Pratul P. Srinivasan, and Peter Hedman. Zip-nerf: Anti-aliased grid-based neural radiance fields. In ICCV, 2023. 2, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 205, + 295, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 205, + 295, + 270 + ], + "spans": [ + { + "bbox": [ + 62, + 205, + 295, + 270 + ], + "type": "text", + "content": "[4] Eric R. Chan, Connor Z. Lin, Matthew A. Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas J. Guibas, Jonathan Tremblay, Sameh Khamis, Tero Karras, and Gordon Wetzstein. Efficient geometry-aware 3d generative adversarial networks. In CVPR, 2022. 1, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 272, + 295, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 272, + 295, + 304 + ], + "spans": [ + { + "bbox": [ + 62, + 272, + 295, + 304 + ], + "type": "text", + "content": "[5] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In ECCV, 2022. 1, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 306, + 295, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 306, + 295, + 338 + ], + "spans": [ + { + "bbox": [ + 62, + 306, + 295, + 338 + ], + "type": "text", + "content": "[6] Christopher Choy, JunYoung Gwak, and Silvio Savarese. 4d spatio-temporal convnets: Minkowski convolutional neural networks. In CVPR, 2019. 2, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 341, + 295, + 372 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 341, + 295, + 372 + ], + "spans": [ + { + "bbox": [ + 62, + 341, + 295, + 372 + ], + "type": "text", + "content": "[7] Brian Curless and Marc Levoy. A volumetric method for building complex models from range images. ACM TOG, 1996. 1, 2, 6, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 375, + 295, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 375, + 295, + 418 + ], + "spans": [ + { + "bbox": [ + 62, + 375, + 295, + 418 + ], + "type": "text", + "content": "[8] Angela Dai, Matthias Nießner, Michael Zollöfer, Shahram Izadi, and Christian Theobalt. Bundlefusion: Real-time globally consistent 3d reconstruction using on-the-fly surface re-integration. ACM TOG, 2017. 1, 2, 6, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 420, + 295, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 420, + 295, + 453 + ], + "spans": [ + { + "bbox": [ + 62, + 420, + 295, + 453 + ], + "type": "text", + "content": "[9] Kangle Deng, Andrew Liu, Jun-Yan Zhu, and Deva Ramanan. Depth-supervised nerf: Fewer views and faster training for free. In CVPR, 2022. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 454, + 295, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 454, + 295, + 509 + ], + "spans": [ + { + "bbox": [ + 57, + 454, + 295, + 509 + ], + "type": "text", + "content": "[10] Daniel Duckworth, Peter Hedman, Christian Reiser, Peter Zhizhin, Jean-François Thibert, Mario Lucic, Richard Szeliski, and Jonathan T. Barron. SMERF: streamable memory efficient radiance fields for real-time large-scene exploration. ACM TOG, 2024. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 510, + 295, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 510, + 295, + 552 + ], + "spans": [ + { + "bbox": [ + 57, + 510, + 295, + 552 + ], + "type": "text", + "content": "[11] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In CVPR, 2022. 2, 7, 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 555, + 295, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 555, + 295, + 589 + ], + "spans": [ + { + "bbox": [ + 57, + 555, + 295, + 589 + ], + "type": "text", + "content": "[12] Qiancheng Fu, Qingshan Xu, Yew Soon Ong, and Wenbing Tao. Geo-neus: Geometry-consistent neural implicit surfaces learning for multi-view reconstruction. In _NeurIPS_, 2022. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 590, + 295, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 590, + 295, + 623 + ], + "spans": [ + { + "bbox": [ + 57, + 590, + 295, + 623 + ], + "type": "text", + "content": "[13] Benjamin Graham and Laurens Van der Maaten. Submanifold sparse convolutional networks. arXiv preprint arXiv:1706.01307, 2017. 2, 8" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 624, + 295, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 624, + 295, + 666 + ], + "spans": [ + { + "bbox": [ + 57, + 624, + 295, + 666 + ], + "type": "text", + "content": "[14] Antoine Guédon and Vincent Lepetit. Sugar: Surface-aligned gaussian splatting for efficient 3d mesh reconstruction and high-quality mesh rendering. In CVPR, 2024. 2, 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 670, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 295, + 712 + ], + "type": "text", + "content": "[15] Peter Hedman, Julien Philip, True Price, Jan-Michael Frahm, George Drettakis, and Gabriel J. Brostow. Deep blending for free-viewpoint image-based rendering. ACM TOG, 2018. 7, 8" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 36, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "text", + "content": "[16] Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2d gaussian splatting for geometrically accurate radiance fields. ACM TOG, 2024. 2, 6, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 107, + 553, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 107, + 553, + 139 + ], + "spans": [ + { + "bbox": [ + 317, + 107, + 553, + 139 + ], + "type": "text", + "content": "[17] Rasmus Ramsbøl Jensen, Anders Lindbjerg Dahl, George Vogiatzis, Engin Tola, and Henrik Aanæs. Large scale multiview stereopsis evaluation. In CVPR, 2014. 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 141, + 553, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 141, + 553, + 173 + ], + "spans": [ + { + "bbox": [ + 317, + 141, + 553, + 173 + ], + "type": "text", + "content": "[18] Animesh Karnewar, Tobias Ritschel, Oliver Wang, and Niloy J. Mitra. Relu fields: The little non-linearity that could. In ACM TOG, 2022. 3, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 175, + 553, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 175, + 553, + 196 + ], + "spans": [ + { + "bbox": [ + 317, + 175, + 553, + 196 + ], + "type": "text", + "content": "[19] Michael Kazhdan, Matthew Bolitho, and Hugues Hoppe. Poisson surface reconstruction. In SGP, 2006. 2, 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 198, + 553, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 198, + 553, + 230 + ], + "spans": [ + { + "bbox": [ + 317, + 198, + 553, + 230 + ], + "type": "text", + "content": "[20] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM TOG, 2023. 1, 2, 3, 4, 6, 7, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 232, + 553, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 232, + 553, + 263 + ], + "spans": [ + { + "bbox": [ + 317, + 232, + 553, + 263 + ], + "type": "text", + "content": "[21] Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and temples: benchmarking large-scale scene reconstruction. ACM TOG, 2017. 7, 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 265, + 553, + 286 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 265, + 553, + 286 + ], + "spans": [ + { + "bbox": [ + 316, + 265, + 553, + 286 + ], + "type": "text", + "content": "[22] Samuli Laine and Tero Karras. Efficient sparse voxel octrees. In ACM I3D, 2010. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 288, + 553, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 288, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 316, + 288, + 553, + 319 + ], + "type": "text", + "content": "[23] Ruilong Li, Hang Gao, Matthew Tancik, and Angjoo Kanazawa. Nerfacc: Efficient sampling accelerates nerfs. In ICCV, 2023. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 321, + 553, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 321, + 553, + 365 + ], + "spans": [ + { + "bbox": [ + 316, + 321, + 553, + 365 + ], + "type": "text", + "content": "[24] Zhaoshuo Li, Thomas Müller, Alex Evans, Russell H. Taylor, Mathias Unberath, Ming-Yu Liu, and Chen-Hsuan Lin. Neuralangelo: High-fidelity neural surface reconstruction. In CVPR, 2023. 1, 2, 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 366, + 553, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 366, + 553, + 398 + ], + "spans": [ + { + "bbox": [ + 316, + 366, + 553, + 398 + ], + "type": "text", + "content": "[25] Lingjie Liu, Jiatao Gu, Kyaw Zaw Lin, Tat-Seng Chua, and Christian Theobalt. Neural sparse voxel fields. In NeurIPS, 2020. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 399, + 553, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 399, + 553, + 432 + ], + "spans": [ + { + "bbox": [ + 316, + 399, + 553, + 432 + ], + "type": "text", + "content": "[26] Charles Loop, Qin Cai, Sergio Orts-Escolano, and Philip A Chou. A closed-form bayesian fusion equation using occupancy probabilities. In 3DV, 2016. 2, 8" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 434, + 553, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 434, + 553, + 465 + ], + "spans": [ + { + "bbox": [ + 316, + 434, + 553, + 465 + ], + "type": "text", + "content": "[27] William E. Lorensen and Harvey E. Cline. Marching cubes: A high resolution 3d surface construction algorithm. In ACM TOG, 1987. 6, 8" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 317, + 468, + 553, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 468, + 553, + 521 + ], + "spans": [ + { + "bbox": [ + 317, + 468, + 553, + 521 + ], + "type": "text", + "content": "[28] Alexander Mai, Peter Hedman, George Kopanas, Dor Verbin, David Futschik, Qiangeng Xu, Falko Kuester, Jonathan T. Barron, and Yinda Zhang. Ever: Exact volumetric ellipsoid rendering for real-time view synthesis. arXiv preprint arXiv:2410.01804, 2024. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 523, + 553, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 523, + 553, + 544 + ], + "spans": [ + { + "bbox": [ + 316, + 523, + 553, + 544 + ], + "type": "text", + "content": "[29] Nelson L. Max. Optical models for direct volume rendering. IEEE TVCG, 1995. 4" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 317, + 545, + 553, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 545, + 553, + 589 + ], + "spans": [ + { + "bbox": [ + 317, + 545, + 553, + 589 + ], + "type": "text", + "content": "[30] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 2, 3, 4, 7, 8" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 317, + 590, + 553, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 590, + 553, + 634 + ], + "spans": [ + { + "bbox": [ + 317, + 590, + 553, + 634 + ], + "type": "text", + "content": "[31] Nicolas Moenne-Loccoz, Ashkan Mirzaei, Or Perel, Ricardo de Lutio, Janick Martinez Esturo, Gavriel State, Sanja Fidler, Nicholas Sharp, and Zan Gojcic. 3d gaussian ray tracing: Fast tracing of particle scenes. ACM TOG, 2024. 2" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 317, + 635, + 553, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 635, + 553, + 667 + ], + "spans": [ + { + "bbox": [ + 317, + 635, + 553, + 667 + ], + "type": "text", + "content": "[32] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM TOG, 2022. 2, 7" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 316, + 669, + 553, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 669, + 553, + 690 + ], + "spans": [ + { + "bbox": [ + 316, + 669, + 553, + 690 + ], + "type": "text", + "content": "[33] Ken Museth. Vdb: High-resolution sparse volumes with dynamic topology. ACM TOG, 2013. 2" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "text", + "content": "[34] Richard A Newcombe, Shahram Izadi, Otmar Hilliges, David Molyneaux, David Kim, Andrew J Davison, Pushmeet" + } + ] + } + ], + "index": 35 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16195" + } + ] + } + ], + "index": 37 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 106 + ], + "type": "text", + "content": "Kohi, Jamie Shotton, Steve Hodges, and Andrew Fitzgibbon. Kinectfusion: Real-time dense surface mapping and tracking. In IEEE ISMAR, 2011. 1, 2, 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 106, + 295, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 106, + 295, + 172 + ], + "spans": [ + { + "bbox": [ + 56, + 106, + 295, + 172 + ], + "type": "text", + "content": "[35] Michael Niemeyer, Fabian Manhardt, Marie-Julie Rakoto-saona, Michael Oechsle, Daniel Duckworth, Rama Gosula, Keisuke Tateno, John Bates, Dominik Kaeser, and Federico Tombari. Radsplat: Radiance field-informed gaussian splat-ting for robust real-time rendering with " + }, + { + "bbox": [ + 56, + 106, + 295, + 172 + ], + "type": "inline_equation", + "content": "900+" + }, + { + "bbox": [ + 56, + 106, + 295, + 172 + ], + "type": "text", + "content": " FPS. arXiv preprint arXiv:2403.13806, 2024. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 173, + 294, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 173, + 294, + 205 + ], + "spans": [ + { + "bbox": [ + 56, + 173, + 294, + 205 + ], + "type": "text", + "content": "[36] Matthias Nießner, Michael Zollhöfer, Shahram Izadi, and Marc Stamminger. Real-time 3d reconstruction at scale using voxel hashing. ACM TOG, 2013. 2, 6, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 206, + 294, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 206, + 294, + 249 + ], + "spans": [ + { + "bbox": [ + 56, + 206, + 294, + 249 + ], + "type": "text", + "content": "[37] Lukas Radl, Michael Steiner, Mathias Parger, Alexander Weinrauch, Bernhard Kerbl, and Markus Steinberger. Stophepop: Sorted gaussian splatting for view-consistent real-time rendering. ACM TOG, 2024. 1, 2, 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 250, + 294, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 250, + 294, + 282 + ], + "spans": [ + { + "bbox": [ + 56, + 250, + 294, + 282 + ], + "type": "text", + "content": "[38] Mike Ranzinger, Greg Heinrich, Jan Kautz, and Pavlo Molchanov. Am-radio: Agglomerative vision foundation model reduce all domains into one. In CVPR, 2024. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 284, + 294, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 284, + 294, + 316 + ], + "spans": [ + { + "bbox": [ + 56, + 284, + 294, + 316 + ], + "type": "text", + "content": "[39] Christian Reiser, Songyou Peng, Yiyi Liao, and Andreas Geiger. Kilonerf: Speeding up neural radiance fields with thousands of tiny mllps. In ICCV, 2021. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 317, + 294, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 317, + 294, + 370 + ], + "spans": [ + { + "bbox": [ + 56, + 317, + 294, + 370 + ], + "type": "text", + "content": "[40] Christian Reiser, Stephan J. Garbin, Pratul P. Srinivasan, Dor Verbin, Richard Szeliski, Ben Mildenhall, Jonathan T. Barron, Peter Hedman, and Andreas Geiger. Binary opacity grids: Capturing fine geometric detail for mesh-based view synthesis. ACM TOG, 2024. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 372, + 294, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 372, + 294, + 404 + ], + "spans": [ + { + "bbox": [ + 56, + 372, + 294, + 404 + ], + "type": "text", + "content": "[41] Johannes L. Schonberger, Enliang Zheng, Jan-Michael Frahm, and Marc Pollefeys. Pixelwise view selection for unstructured multi-view stereo. In ECCV, 2016. 2, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 405, + 294, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 405, + 294, + 437 + ], + "spans": [ + { + "bbox": [ + 56, + 405, + 294, + 437 + ], + "type": "text", + "content": "[42] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. In ICLR, 2015. 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 438, + 294, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 438, + 294, + 470 + ], + "spans": [ + { + "bbox": [ + 56, + 438, + 294, + 470 + ], + "type": "text", + "content": "[43] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In CVPR, 2022. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 472, + 294, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 472, + 294, + 514 + ], + "spans": [ + { + "bbox": [ + 56, + 472, + 294, + 514 + ], + "type": "text", + "content": "[44] Cheng Sun, Guangyan Cai, Zhengqin Li, Kai Yan, Cheng Zhang, Carl S. Marshall, Jia-Bin Huang, Shuang Zhao, and Zhao Dong. Neural-pbir reconstruction of shape, material, and illumination. In ICCV, 2023. 1, 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 516, + 294, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 516, + 294, + 569 + ], + "spans": [ + { + "bbox": [ + 56, + 516, + 294, + 569 + ], + "type": "text", + "content": "[45] Towaki Takikawa, Joey Litalien, Kangxue Yin, Karsten Kreis, Charles T. Loop, Derek Nowrouzezahrai, Alec Jacobson, Morgan McGuire, and Sanja Fidler. Neural geometric level of detail: Real-time rendering with implicit 3d shapes. In CVPR, 2021. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 571, + 294, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 571, + 294, + 624 + ], + "spans": [ + { + "bbox": [ + 56, + 571, + 294, + 624 + ], + "type": "text", + "content": "[46] Haithem Turki, Vasu Agrawal, Samuel Rota Bulò, Lorenzo Porzi, Peter Kontschieder, Deva Ramanan, Michael Zollhöfer, and Christian Richardt. Hybridnerf: Efficient neural rendering via adaptive volumetric surfaces. In CVPR, 2024. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 625, + 294, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 625, + 294, + 669 + ], + "spans": [ + { + "bbox": [ + 56, + 625, + 294, + 669 + ], + "type": "text", + "content": "[47] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. In NeurIPS, 2021. 2, 6, 8" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 670, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 294, + 713 + ], + "type": "text", + "content": "[48] Yiming Wang, Qin Han, Marc Habermann, Kostas Dani-ilidis, Christian Theobalt, and Lingjie Liu. Neus2: Fast learning of neural implicit surfaces for multi-view reconstruction. In ICCV, 2023. 2" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 420 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "text", + "content": "[49] Zian Wang, Tianchang Shen, Merlin Nimier-David, Nicholas Sharp, Jun Gao, Alexander Keller, Sanja Fidler, Thomas Müller, and Zan Gojcic. Adaptive shells for efficient neural radiance field rendering. ACM TOG, 2023. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 118, + 553, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 118, + 553, + 183 + ], + "spans": [ + { + "bbox": [ + 317, + 118, + 553, + 183 + ], + "type": "text", + "content": "[50] Francis Williams, Jiahui Huang, Jonathan Swartz, Gergely Klár, Vijay Thakkar, Matthew Cong, Xuanchi Ren, Ruilong Li, Clement Fuji-Tsang, Sanja Fidler, Eftychios Sifakis, and Ken Museth. fvdb: A deep-learning framework for sparse, large scale, and high performance spatial intelligence. ACM TOG, 2024. 2, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 185, + 553, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 185, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 317, + 185, + 553, + 228 + ], + "type": "text", + "content": "[51] Tong Wu, Jiaqi Wang, Xingang Pan, Xudong Xu, Christian Theobalt, Ziwei Liu, and Dahua Lin. Voxurf: Voxel-based efficient and accurate neural surface reconstruction. In ICLR, 2023. 1, 2, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 230, + 553, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 230, + 553, + 273 + ], + "spans": [ + { + "bbox": [ + 317, + 230, + 553, + 273 + ], + "type": "text", + "content": "[52] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, José M. Álvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. In NeurIPS, 2021. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 274, + 553, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 274, + 553, + 306 + ], + "spans": [ + { + "bbox": [ + 317, + 274, + 553, + 306 + ], + "type": "text", + "content": "[53] Lior Yariv, Jiatao Gu, Yoni Kasten, and Yaron Lipman. Volume rendering of neural implicit surfaces. In NeurIPS, 2021. 2, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 308, + 553, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 308, + 553, + 352 + ], + "spans": [ + { + "bbox": [ + 317, + 308, + 553, + 352 + ], + "type": "text", + "content": "[54] Lior Yariv, Peter Hedman, Christian Reiser, Dor Verbin, Pratul P. Srinivasan, Richard Szeliski, Jonathan T. Barron, and Ben Mildenhall. Bakedsdf: Meshing neural sdfs for real-time view synthesis. In ACM TOG, 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 354, + 553, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 354, + 553, + 386 + ], + "spans": [ + { + "bbox": [ + 316, + 354, + 553, + 386 + ], + "type": "text", + "content": "[55] Alex Yu, Ruilong Li, Matthew Tancik, Hao Li, Ren Ng, and Angjoo Kanazawa. Plenoctrees for real-time rendering of neural radiance fields. In ICCV, 2021. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 388, + 553, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 388, + 553, + 420 + ], + "spans": [ + { + "bbox": [ + 316, + 388, + 553, + 420 + ], + "type": "text", + "content": "[56] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 7" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "16196" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Sparse2DGS_ Geometry-Prioritized Gaussian Splatting for Surface Reconstruction from Sparse Views/3c5dfe86-080c-4ec1-9ca9-65bf2b3453a3_content_list.json b/2025/Sparse2DGS_ Geometry-Prioritized Gaussian Splatting for Surface Reconstruction from Sparse Views/3c5dfe86-080c-4ec1-9ca9-65bf2b3453a3_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..3c1d08aeb82aecb80e7f7f23365c93e3aaf4b324 --- /dev/null +++ b/2025/Sparse2DGS_ Geometry-Prioritized Gaussian Splatting for Surface Reconstruction from Sparse Views/3c5dfe86-080c-4ec1-9ca9-65bf2b3453a3_content_list.json @@ -0,0 +1,1570 @@ +[ + { + "type": "text", + "text": "Sparse2DGS: Geometry-Prioritized Gaussian Splating for Surface Reconstruction from Sparse Views", + "text_level": 1, + "bbox": [ + 202, + 130, + 797, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jiang Wu Rui Li Yu Zhu† Rong Guo Jinqiu Sun Yanning Zhang† Northwestern Polytechnical University", + "bbox": [ + 205, + 202, + 790, + 239 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5de91afae79a2e02772b387629ac4a9a80ef02c5eb2ec8b0f33fbbba5c40b093.jpg", + "image_caption": [ + "Figure 1. Sparse2DGS boosts the strengths of Gaussian Splatting and Multi-view Stereo in sparse-view surface reconstruction, with notable improvement over 2DGS [11], CLMVSNet [31], as well as their plain combination (left). Meanwhile, Sparse2DGS is $2 \\times$ faster than NeRF-based sparse-view method [20] (middle) and leads to complete and accurate reconstruction (right)." + ], + "image_footnote": [], + "bbox": [ + 102, + 280, + 897, + 440 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 506, + 326, + 522 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We present a Gaussian Splatting method for surface reconstruction using sparse input views. Previous methods relying on dense views struggle with extremely sparse Structure-from-Motion points for initialization. While learning-based Multi-view Stereo (MVS) provides dense 3D points, directly combining it with Gaussian Splatting leads to suboptimal results due to the ill-posed nature of sparse-view geometric optimization. We propose Sparse2DGS, an MVS-initialized Gaussian Splatting pipeline for complete and accurate reconstruction. Our key insight is to incorporate the geometric-prioritized enhancement schemes, allowing for direct and robust geometric learning under ill-posed conditions. Sparse2DGS outperforms existing methods by notable margins while being $2 \\times$ faster than the NeRF-based fine-tuning approach. Code is available at https://github.com/Wuuu3511/Sparse2DGS.", + "bbox": [ + 88, + 537, + 485, + 780 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 806, + 220, + 821 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Reconstructing 3D surfaces from multiple views is a fundamental task in computer vision and has been widely used in autonomous driving [9, 18] and virtual reality [21]. While", + "bbox": [ + 89, + 832, + 483, + 878 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "typical methods require abundant input images, reconstructing accurate 3D surfaces from sparse views is particularly challenging and is crucial for scenes that are frequently captured with limited images.", + "bbox": [ + 511, + 508, + 906, + 566 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "NeRF-based surface reconstruction methods [22] have gained popularity by encoding scene geometry into an occupancy field [17, 23] or a signed distance function (SDF) [28, 38]. However, when faced with sparse views, these methods often yield incomplete surfaces with artifacts due to insufficient supervision. Although sparse-view methods [19, 20, 25, 34] use pre-trained feed-forward NeRFs to provide geometric priors, they still exhibit limited generalization to unseen scenes.", + "bbox": [ + 511, + 566, + 908, + 700 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, 3D Gaussian Splatting (3DGS) [14] has leveraged Structure-from-Motion (SfM) point cloud to construct Gaussian primitives for fast and high-quality novel view synthesis (NVS). Subsequent works [4, 5, 7, 11, 41, 43] aim to achieve more efficient 3D surface reconstruction using 3DGS. Notably, 2D Gaussian Splatting (2DGS) [11] proposes a 2D variant of Gaussian primitives, with disk-shaped primitives suitable for geometry representation. However, when dealing with sparse views, existing methods [4, 11, 43] struggle to produce faithful reconstructions due to initially over-sparse SfM points.", + "bbox": [ + 511, + 703, + 908, + 869 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "On the other hand, learning-based multi-view stereo (MVS) [6, 16, 30, 36] predicts dense depth maps from", + "bbox": [ + 511, + 869, + 906, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$^\\dagger$ indicates corresponding authors.", + "bbox": [ + 112, + 886, + 297, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "11307", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "pixel-wise feature matching, significantly improving geometric completeness even with sparse images. However, MVS depth networks focus on generalizable inference instead of test-time geometry enhancement by design. While combining MVS point cloud with Gaussian Splatting is intuitively effective, their plain combination (see Fig. 1, \"2DGS+CLMVS\") yields sub-optimal results due to the ill-posed nature of geometry optimization from sparse views.", + "bbox": [ + 89, + 90, + 480, + 210 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we propose Sparse2DGS, an MVS-initialized Gaussian Splating method for complete and accurate sparse-view 3D surface reconstruction. The key idea is to boost the strengths of two paradigms through geometric-prioritized enhancement against the ill-posed nature of sparse-view reconstruction. Starting from MVS initialization, we first incorporate MVS-derived geometric features into the Gaussian splattering process while keeping feature and color values fixed. This provides geometrically expressive supervision and prevents appearance overfitting, thereby improving geometric reconstruction quality. We further introduce a direct Gaussian primitive regularization, with a reparameterization-based disk sampling strategy that reformulates Gaussian primitives' position, orientation, and scale into point representations. This approach enables a unified optimization of different forms of Gaussian properties within a standard point-based, cross-view consistency loss. Moreover, instead of densifying Gaussian primitives under ill-posed conditions, we propose to selectively update the MVS-initialized Gaussian primitives with rendered geometric cues, allowing for self-guided geometric corrections.", + "bbox": [ + 91, + 213, + 482, + 530 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Sparse2DGS significantly outperforms existing Gaussian Splatting methods in sparse-view surface reconstruction. Meanwhile, it demonstrates notable efficiency advantages over NeRF-based test-time optimization. Our contributions are as follows:", + "bbox": [ + 89, + 531, + 482, + 604 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose Sparse2DGS, a test-time sparse-view 3D surface reconstruction method that boosts the strengths of Gaussian Splitting and MVS, allowing for complete and accurate reconstruction from sparse images.", + "- We focus on geometric-prioritized enhancement by proposing geometrically enhanced supervision, direct Gaussian primitive regularization, and selective Gaussian update throughout the Gaussian Splatting pipeline. Sparse2DGS achieves the lowest error (1.13 CD error) compared to 2DGS (2.81) [11], GOF (2.82) [43], and PGSR (2.08) [4] on DTU dataset with 3 views. Meanwhile, Sparse2DGS is $2 \\times$ faster than fine-tuning NeRF-based SparseNeus [20]." + ], + "bbox": [ + 89, + 607, + 482, + 804 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 89, + 820, + 238, + 835 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. 3D Reconstruction via Novel View Synthesis", + "text_level": 1, + "bbox": [ + 89, + 845, + 464, + 861 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recent progresses [11, 12, 25, 43] in novel view synthesis (NVS) have enabled accurate 3D surface reconstruc", + "bbox": [ + 89, + 869, + 482, + 898 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tion with only image-based supervision. Neural Radiance Fields [22] (NeRFs) have become a significant technique for representing 3D scenes. It has inspired numerous works on directly extracting surfaces by training a parameterized radiance field through differentiable volume rendering. Neus [28] and VolSDF [38] integrate surface learning into volume rendering by re-parameterizing the signed distance function as volume density, reconstructing more accurate surfaces. To reduce geometric deviations and accelerate training, subsequent works [8, 27, 29, 42] have introduced improvements in various areas.", + "bbox": [ + 511, + 90, + 903, + 256 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recently, 3D Gaussian Splatting [14] has shown impressive results in novel view synthesis (NVS), with many recent studies adopting this approach for surface reconstruction. SuGaR [10] aligns 3D Gaussians with scene surfaces by designing a regularization term and extracts an accurate mesh from the Gaussians using Poisson Reconstruction [13]. 2DGS [11] replaces 3D Gaussian ellipsoids with 2D-oriented planar disks, providing view-consistent geometric representations. RaDe-GS [44] further designs a rasterized approach to render depth and normal, enhancing geometric accuracy while ensuring rendering efficiency. GOF [43] extracts the level set from trained Gaussian primitives by constructing a Gaussian opacity field. PGSR [4] improves mesh quality by introducing single-view and multi-view geometric constraints. Additionally, GSDF [41] and NeusSG [5] reconstruct scenes by combining the advantages of 3D Gaussian Splatting and SDF networks. Despite these methods' success, the extracted geometry tends to be noisy and incomplete when the input number of training viewpoints is limited.", + "bbox": [ + 511, + 257, + 903, + 559 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Learning-based Multi-view Stereo", + "text_level": 1, + "bbox": [ + 511, + 570, + 812, + 585 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Deep learning-based Multi-view Stereo (MVS) estimates the reference image's depth from multi-view posed images. MVSNet [36] first introduces the differentiable homography warping to build the cost volume with deep image features and then employs a 3D CNN for cost regularization. Recent follow-up works have introduced improvements in various aspects, including multi-view feature extraction [2, 6], cost aggregation [32, 33, 35, 46], and cost volume regularization [30]. However, these methods require ground-truth depth maps for training, which are challenging to obtain in practice. Unsupervised MVS [3, 15, 31] can be applied in more general, real-world scenarios. MVS methods can infer dense depth maps for each input view and therefore, we select the unsupervised CLMVSNet [31] to initialize the Gaussian point cloud.", + "bbox": [ + 511, + 594, + 903, + 820 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.3. Sparse View Surface Reconstruction", + "text_level": 1, + "bbox": [ + 511, + 830, + 828, + 847 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recent works have explored 3D surface reconstruction from sparse training views. SparseNeuS [20] achieves reasonable mesh reconstruction results by constructing a cascaded", + "bbox": [ + 511, + 854, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "11308", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/6a87a0e2c57120679f272a77cf9a2d5b3d88bc0b0ba4c0d9b0fe890f1f77df64.jpg", + "image_caption": [ + "Figure 2. Overview. Given sparse posed images, we first leverage the MVS points to initialize the Gaussian position $\\mathbf{p}$ (Sec. 3.1). We then leverage the MVS-derived feature for Gaussian Splating, with fixed feature $\\mathbf{f}$ and color $\\mathbf{c}$ values to conduct geometrically expressive supervision and avoid appearance overfitting (Sec. 3.2). We then optimize Gaussian primitive properties by reformulating orientation $\\mathbf{R}$ , scale $\\mathbf{S}$ , and position $\\mathbf{p}$ into sampled points using the reparameterization-based disk sampling, ensuring simultaneously optimizing different forms of Gaussian properties with point-based loss (Sec. 3.3). Finally, we introduce a Selective Gaussian Update strategy that leverages rendered geometry as an alternative to update Gaussian primitive positions, which further improves reconstruction quality (Sec. 3.4)." + ], + "image_footnote": [], + "bbox": [ + 98, + 85, + 898, + 321 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "volume to learn a generalizable radiance field, followed by per-scene fine-tuning. NeuSurf [12] learns dense on-surface priors from sparse points to guide geometric learning. SparseCraft [40] leverages stereopsis cues for geometric regularization based on Taylor expansion. Moreover, VolRecon [25] and ReTR [19] enhance voxel features by designing specialized transformer architectures. C2F2Neus [34] constructs per-view cost frustum for finer geometry estimation. Although these feedforward networks can directly infer scene geometry, their limited generalization to unseen scenes restricts their applicability in real-world scenarios.", + "bbox": [ + 88, + 429, + 483, + 597 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 89, + 603, + 225, + 619 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given sparse view posed images $\\{\\mathbf{I}_i\\}_{i=1}^N$ , where $\\mathbf{I}_i \\in \\mathbb{R}^{H \\times W \\times 3}$ and $N$ denotes the view number, Sparse2DGS aims to reconstruct accurate 3D scene surfaces using Gaussian Splatting. As shown in Fig. 2, we first initialize the Gaussian primitives with the MVS [31] reconstructed point cloud. Then, we sample MVS geometric feature maps into feature splatting with its value and RGB color fixed for supervision. For each Gaussian primitive, we regularize its geometric properties (position, orientation, scale) by a novel reparameterization-based 2D disk sampling strategy that formulates these properties as point-based supervision. Finally, we propose a selective Gaussian update strategy to further enhance Gaussian primitives using rendered geometric cues.", + "bbox": [ + 88, + 628, + 482, + 839 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. 2D Gaussian Splitting with MVS initialization", + "text_level": 1, + "bbox": [ + 89, + 847, + 482, + 864 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2D Gaussian Splitting [11] (2DGS) represents Gaussian primitives as planar disks suitable for representing and op-", + "bbox": [ + 89, + 869, + 483, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "timizing 3D surfaces. However, it relies on sparse SfM [26] 3D points for initialization, which exhibits limited reconstruction quality under sparse-view inputs. To address this, we initialize 2DGS with an MVS point cloud to provide dense geometric representations.", + "bbox": [ + 511, + 429, + 906, + 503 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2D Gaussian Splatting. Each 2D Gaussian primitive is characterized by its central position $\\mathbf{p}$ , two principal tangential vectors $\\mathbf{t_u}$ and $\\mathbf{t_v}$ , and a scaling vector $(s_u, s_v)$ . The rotation matrix of a 2D Gaussian can be represented as $\\mathbf{R} = (\\mathbf{t_u}, \\mathbf{t_v}, \\mathbf{t_z})$ , where $\\mathbf{t_z}$ is defined by two orthogonal tangent vectors $\\mathbf{t_z} = \\mathbf{t_u} \\times \\mathbf{t_v}$ and the scale matrix is $\\mathbf{S} = \\mathrm{diag}(s_u, s_v, 0)$ . A 2D Gaussian is then defined in a local tangent plane in world space, which is parameterized:", + "bbox": [ + 511, + 505, + 908, + 627 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nP (u, v) = \\mathbf {p} + s _ {u} \\mathbf {t} _ {u} u + s _ {v} \\mathbf {t} _ {v} v. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 602, + 638, + 906, + 656 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For the point $\\mathbf{u} = (u,v)$ in $uv$ space, its 2D Gaussian value can then be evaluated by standard Gaussian", + "bbox": [ + 511, + 667, + 905, + 696 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {G} (\\mathbf {u}) = \\exp \\frac {- \\left(u ^ {2} + v ^ {2}\\right)}{2}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 622, + 708, + 905, + 739 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Initializing 2DGS with MVS depth. MVS methods estimate dense depth maps through cross-view matching. In particular, unsupervised MVS [31] reconstructs complete scenes without ground truth for training, offering greater flexibility. Therefore, we use the point cloud reconstructed by the current unsupervised method CLMVSNet [31] to initialize our Gaussian point cloud. Specifically, given sparse views' images $\\{\\mathbf{I}_i\\}_{i=1}^N$ , we use the MVS network for depth maps $\\{\\mathbf{D}_i\\}_{i=1}^N$ , then back-project these depth maps and fuse into a point cloud $\\{\\mathbf{p}_{ij} \\mid 1 \\leq i \\leq N, 1 \\leq j \\leq HW\\}$ as ini-", + "bbox": [ + 511, + 750, + 906, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "11309", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "tialized positions for Gaussian primitives, where $i$ denotes image index and $j$ is the pixel index within each image.", + "bbox": [ + 89, + 90, + 482, + 122 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Geometrically Enhanced Supervision", + "text_level": 1, + "bbox": [ + 89, + 130, + 413, + 147 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3D/2D Gaussian Splatting [11, 14] couples color and geometry optimization through image-based loss. However, this leads to severe degradation under sparse views as color will overfit sparse views, leading to a compromised geometry. Meanwhile, the challenges in image areas (e.g., textureless regions) will be amplified with sparse views. To address this issue, we propose geometrically enhanced supervision by incorporating off-the-shelf expressive MVS features for feature splatting and supervision, with the color and feature fixed to avoid overfitting.", + "bbox": [ + 89, + 154, + 483, + 305 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In addition to providing dense point clouds from input images $\\{\\mathbf{I}_i\\}_{i = 1}^N$ , the MVS network [31] offers off-the-shelf feature maps with geometric expressiveness. We leverage MVS features to enhance geometric supervision by incorporating additional feature attribute $\\mathbf{f}_{ij}$ into Gaussian primitives. Specifically, for each input image $\\mathbf{I}_i$ , we extract its feature map $\\mathbf{F}_i$ from the Feature Pyramid Network (FPN) of CLMVSNet [31]. Since each 3D point corresponds to an image pixel due to MVS initialization, we sample both RGB color and MVS features into each Gaussian primitive.", + "bbox": [ + 89, + 306, + 483, + 457 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {f} _ {i j} = \\mathbf {F} _ {i} (\\pi (\\mathbf {p} _ {i j})), \\mathbf {c} _ {i j} = \\mathbf {I} _ {i} (\\pi (\\mathbf {p} _ {i j})), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 166, + 469, + 482, + 488 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\pi(\\cdot)$ denotes the projection from 3D space onto 2D image plane. We further render the feature into an image plane using the tile-based renderer together with RGB rendering. We show the rendering process given $M$ Gaussian primitives along a ray, with indices $i, j$ omitted for brevity", + "bbox": [ + 89, + 497, + 483, + 574 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\{\\hat {\\mathbf {c}}, \\hat {\\mathbf {f}} \\} = \\sum_ {m = 1} ^ {M} \\left\\{\\mathbf {c} _ {m}, \\mathbf {f} _ {m} \\right\\} \\cdot \\alpha_ {m} \\prod_ {o = 1} ^ {m - 1} (1 - \\alpha_ {m}), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 140, + 597, + 482, + 638 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\alpha_{m}$ is the opacity of the Gaussian. With rendered feature maps that have enhanced geometric expressiveness for the challenging sparse views, we further supervise the rendered features using cosine similarity loss by", + "bbox": [ + 89, + 648, + 483, + 710 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {f e a} = 1 - \\frac {\\mathbf {f} _ {i j} \\cdot \\hat {\\mathbf {f}} _ {i j}}{\\| \\mathbf {f} _ {i j} \\| \\| \\hat {\\mathbf {f}} _ {i j} \\|}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 204, + 720, + 482, + 758 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Note that the simultaneous learning of appearance (i.e., color $\\mathbf{c}_{ij}$ and feature $\\mathbf{f}_{ij}$ ) and geometry lead to geometric degradation under sparse views (see Tab. 4). To address this issue, both $\\mathbf{f}_{ij}$ and $\\mathbf{c}_{ij}$ are fixed during the optimization to enforce the learning specifically for geometry.", + "bbox": [ + 89, + 763, + 482, + 840 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our method essentially differentiates previous feature-based methods [24, 47] in its specific strategy for geometric enhancement. In our method, the expressive MVS features are incorporated, and features of each Gaussian are fixed,", + "bbox": [ + 89, + 840, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/a9ca6fc52cdf00f37bf78b84109e47bd721d5cd4a5c260b55463f271c3ff79df.jpg", + "image_caption": [ + "Figure 3. Direct Gaussian Primitive Regularization. We sample point set $\\mathbf{Z}' \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})$ from the standard Gaussian distribution and leverage the reparameterization technique to transform $\\mathbf{Z}'$ to the point set $\\mathbf{Z}$ on each Gaussian primitive. This process allows for representing Gaussian orientation $\\mathbf{R}$ and scale $\\mathbf{S}$ properties in the form of points easier for geometric supervision. The transformed points $\\mathbf{Z}$ are then projected to the source and target views, supervised by cross-view feature consistency loss." + ], + "image_footnote": [], + "bbox": [ + 532, + 92, + 883, + 280 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "both for geometry enhancement with sparse views. Previous methods leverage splatted features only for segmentation or editing [24, 47] with dense views, while their feature losses do not optimize the scene geometry.", + "bbox": [ + 511, + 422, + 905, + 484 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Direct Gaussian Primitive Regularization", + "text_level": 1, + "bbox": [ + 511, + 496, + 867, + 512 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The disk-shaped 2D Gaussian primitives encode latent scene geometry including local shape orientation $\\mathbf{R}$ , scale $\\mathbf{S}$ , and position $\\mathbf{p}$ , which are essential for accurate surface reconstruction. These geometric properties are typically learned from surrogate image render loss supported by abundant multi-view supervision. When facing sparse-view inputs, the primitives tend to degrade due to insufficient geometric constraints.", + "bbox": [ + 511, + 520, + 905, + 641 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To address this issue, we incorporate direct geometric regularization for 2D Gaussian primitives through multiview consistency guidance. As multi-view consistency approaches only consider point position, we incorporate broad Gaussian primitive properties (i.e., orientation, scale, and position) by correlating their relationship with point position, through a novel reparameterization-based disk sampling strategy. This regularization aligns 2D Gaussian primitives more closely with the latent scene surfaces, leading to improved geometric learning. In the following illustration, we omit the Gaussian primitive indices $i,j$ for brevity.", + "bbox": [ + 511, + 643, + 906, + 808 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Regularizing points with cross-view consistency. Given 3D point sampling positions $\\mathbf{Z} = \\{\\mathbf{z}_k\\}_{k=1}^K$ on each Gaussian, we project them to source and target views to obtain the corresponding 2D positions $\\mathbf{q}_k^{src}$ and $\\mathbf{q}_k^{tar}$ , as well as queried MVS features $\\mathbf{F}_{src}(\\mathbf{q}_k^{src})$ , $\\mathbf{F}_{tar}(\\mathbf{q}_k^{tar})$ . We can then optimize the 3D positions by cross-view feature consistency", + "bbox": [ + 511, + 810, + 908, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "11310", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "loss [45] $\\mathcal{L}_{df}$", + "bbox": [ + 89, + 90, + 184, + 106 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {d f} = \\frac {1}{K} \\sum_ {k = 1} ^ {K} (1 - \\frac {\\mathbf {F} _ {s r c} \\left(\\mathbf {q} _ {k} ^ {s r c}\\right) \\cdot \\mathbf {F} _ {t a r} \\left(\\mathbf {q} _ {k} ^ {t a r}\\right)}{\\| \\mathbf {F} _ {s r c} \\left(\\mathbf {q} _ {k} ^ {s r c}\\right) \\| \\| \\mathbf {F} _ {t a r} \\left(\\mathbf {q} _ {k} ^ {t a r}\\right) \\|}), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 117, + 119, + 480, + 160 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $K$ denotes the number of sampled points on each disk-shaped Gaussian primitive.", + "bbox": [ + 89, + 171, + 482, + 200 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Reparameterization-based disk sampling. The above operation focuses on position regularization only, and thus, it cannot propagate gradients to other Gaussian primitive properties such as orientation $\\mathbf{R}$ and scale $\\mathbf{S}$ . We propose a reparameterization-based sampling strategy to solve this problem. As shown in Fig. 3, for each Gaussian primitive, we first sample points $\\mathbf{Z}'$ from the standard Gaussian distribution $\\mathcal{N}(\\mathbf{0},\\mathbf{I})$ and then transform the distribution to the target Gaussian primitive through its position $\\mathbf{p}$ , orientation $\\mathbf{R}$ and scale $\\mathbf{S}$", + "bbox": [ + 89, + 202, + 483, + 349 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {Z} = \\mathbf {p} + \\mathbf {L} \\mathbf {Z} ^ {\\prime}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 236, + 353, + 482, + 369 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {L} = \\mathbf {R S}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 253, + 378, + 480, + 393 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{Z}' \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})$ . This bridges the gap between Gaussian properties (orientation and scale) and point representation, ensuring regularizing Gaussian geometric properties merely through the surrogate point-based loss. Additionally, since the last scale value of the 2D Gaussian primitive [11] is zero, resulting in difficulties in optimizing the normal vector $\\mathbf{R}[\\dots,2]$ in the last dimension, we apply additional supervision using the rendered normal $\\mathbf{n}$ :", + "bbox": [ + 89, + 401, + 483, + 521 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {d n} = 1 - \\mathbf {R} [ \\dots , 2 ] ^ {\\top} \\mathbf {n}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 202, + 532, + 482, + 551 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The final primitive regularization loss can be defined as:", + "bbox": [ + 89, + 564, + 460, + 579 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {d r} = \\mathcal {L} _ {d f} + \\mathcal {L} _ {d n}. \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 222, + 592, + 480, + 608 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Selective Gaussian Update", + "text_level": 1, + "bbox": [ + 89, + 619, + 331, + 636 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Gaussian Splitting leverages adaptive density control to densify the Gaussian primitives. However, with sparse views, the densified Gaussian primitives are prone to degrade without sufficient multi-view supervision. To address this issue, as shown in Fig. 2, instead of continually densifying the MVS-initialized dense Gaussian primitives with limited supervision, we propose to selectively update them guided by the alternative rendered geometric cues, e.g., depth and normal.", + "bbox": [ + 89, + 643, + 482, + 779 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluating Gaussian primitives via patch warping. Given any Gaussian primitive with its position $\\mathbf{p}$ and orientation $\\mathbf{R}$ , we can warp an image patch from one source view to a target view using its position $\\mathbf{p}$ and orientation $\\mathbf{R}$ together with the pose, following homography warping techniques [3]. Then, we can evaluate the geometric quality of the Gaussian primitive via the patch-based Normalized Cross-Correlation (NCC) [39] metric.", + "bbox": [ + 89, + 780, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {N C C} (x, y) = \\frac {\\sum_ {i = l} ^ {n} \\left(x _ {l} - \\bar {x}\\right) \\left(y _ {l} - \\bar {y}\\right)}{\\sqrt {\\sum_ {l = 1} ^ {n} \\left(x _ {l} - \\bar {x}\\right) ^ {2}} \\sqrt {\\sum_ {l = 1} ^ {n} \\left(y _ {l} - \\bar {y}\\right) ^ {2}}}, \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 109, + 903, + 147 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $x_{l}$ and $y_{l}$ represent the pixels in the patches $x$ and $y$ of the source and target images, respectively.", + "bbox": [ + 511, + 151, + 903, + 181 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Update Gaussian primitive with rendered cues. In addition to using the Gaussian primitive parameter $\\mathbf{p}$ and orientation $\\mathbf{R}$ for patch warping, we can leverage the rendered pixel-wise depth and normals from tiled Gaussian primitives as an alternative for patch warping. We compute patch-wise NCC over two warping strategies, yielding $\\mathrm{NCC}_G$ for Gaussian primitive and $\\mathrm{NCC}_R$ for rendered cues.", + "bbox": [ + 511, + 181, + 903, + 287 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "If $\\mathrm{NCC}_R$ is better than $\\mathrm{NCC}_G$ , we reproject the rendered depth into a 3D point and update the position of the corresponding Gaussian primitive. We update the position of Gaussian primitives every 100 training steps, after which all Gaussian parameters continue to be optimized by gradients. In this way, we improve the initial Gaussian primitive with geometric cues from rendered results, further enhancing the reconstruction quality.", + "bbox": [ + 511, + 287, + 905, + 407 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Training loss", + "text_level": 1, + "bbox": [ + 513, + 415, + 650, + 431 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "During training, in addition to using the disk regularization loss $\\mathcal{L}_{dr}$ and MVS feature splatting loss $\\mathcal{L}_{fea}$ , we retain the original 2DGS [11] image reconstruction loss $\\mathcal{L}_{rgb}$ , depth-normal consistency loss $\\mathcal{L}_n$ , and depth distortion loss $\\mathcal{L}_d$ . The final loss function is defined as:", + "bbox": [ + 511, + 439, + 905, + 513 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {r g b} + \\lambda_ {1} \\mathcal {L} _ {d} + \\lambda_ {2} \\mathcal {L} _ {n}, + \\lambda_ {3} \\mathcal {L} _ {d r} + \\lambda_ {4} \\mathcal {L} _ {f e a}, \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 535, + 521, + 903, + 539 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda_1, \\lambda_2, \\lambda_3, \\lambda_4$ denotes the hyperparameters for the weights of each loss term.", + "bbox": [ + 511, + 545, + 903, + 575 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 587, + 645, + 604 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we first introduce the implementation details and then provide quantitative and qualitative comparisons with current methods on DTU [1] dataset, as well as qualitative results on BlendedMVS [37] dataset. We further analyze the contribution of each module and compare them with different methods through ablation studies.", + "bbox": [ + 511, + 613, + 905, + 704 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Datasets", + "text_level": 1, + "bbox": [ + 511, + 710, + 614, + 724 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The DTU [1] dataset consists of 128 scenes captured in controlled lab environments using structured light scanners containing 49 or 64 images. Similar to SparseNeuS [20], we select 15 scenes and use views 23, 24, and 33 for 3-view reconstruction, with an image resolution of $576 \\times 768$ . We perform the quantitative evaluation on the DTU dataset using the official evaluation code to calculate the Chamfer Distance between the reconstructed mesh and the ground truth point cloud. The BlendedMVS [37] dataset includes several large-scale real-world scenes. We also show the qualitative reconstruction comparisons.", + "bbox": [ + 511, + 734, + 906, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "11311", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/e013745cc6db1209cab451ef33617c107936fdc06ff0e08af75fd77bf969fe4d.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Scan243740556365698397105106110114118122Mean
Colmap [26]0.902.891.631.082.181.941.611.302.341.281.101.420.761.171.141.52
CLMVSNet [31]1.172.521.680.891.511.820.991.291.520.780.881.100.641.041.121.26
Neus [28]4.574.493.974.324.631.954.683.834.152.501.526.471.265.576.114.00
Volsdf [38]4.034.216.120.918.241.732.741.825.143.092.084.810.603.512.183.41
VolRecon [25]1.202.591.561.081.431.921.111.481.421.051.191.380.741.231.271.38
SparseNeus [20]1.292.271.570.881.611.861.061.271.421.070.990.870.541.151.181.27
Retr [19]1.052.311.440.981.181.520.881.351.300.871.070.770.591.051.121.17
C2F2Neus [34]1.122.421.400.751.411.770.851.161.260.760.910.600.460.880.921.11
Sparsecraft [40]1.171.741.800.701.191.530.831.051.420.780.800.560.440.770.841.04
NeuSurf [12]0.782.351.550.751.041.680.601.140.980.700.740.490.390.750.860.99
2DGS [11]3.143.792.311.515.342.221.852.793.621.633.054.111.053.352.442.81
GOF [43]3.275.052.772.043.842.612.232.653.301.652.584.441.162.292.392.82
PGSR [4]3.363.282.751.275.151.840.881.793.491.191.861.110.611.091.522.08
Ours1.052.351.380.831.371.450.841.161.430.740.850.840.570.951.011.13
", + "bbox": [ + 99, + 88, + 898, + 282 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Quantitative results on DTU dataset. Our method achieves the best performance among Gaussian Splitting surface reconstruction approaches and demonstrates competitive performance to previous state-of-the-art NeRF-based methods.", + "bbox": [ + 89, + 287, + 906, + 316 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/9ed3e83273f7eec8f44dfc6e88c43c823b51662c3c3d7bb63b0e62738b1a4a80.jpg", + "image_caption": [ + "Figure 4. DTU surface reconstruction results. Our method achieves more complete reconstructions with finer details." + ], + "image_footnote": [], + "bbox": [ + 120, + 321, + 877, + 683 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Implementation Details", + "text_level": 1, + "bbox": [ + 89, + 710, + 307, + 727 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our method is implemented based on 2DGS [11]. For Gaussian primitive initialization, we use the point cloud back-projected depth maps from CLMVSNet [31]. Instead of using spherical harmonics to represent the color of the Gaussian primitives, we initialize the color using the RGB values from the image. CLMVSNet uses an FPN for feature extraction, and we utilize the full-resolution feature maps with a channel dimension of 8 from the FPN last layer for feature splatting. The remaining parameters are initialized consistently with those in 2DGS. We set the training iter", + "bbox": [ + 89, + 750, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ation to 7000 and replaced the original 2DGS's adaptive density control with our proposed Gaussian update strategy. Similar to 2DGS [11], we used truncated signed distance fusion (TSDF) to extract the final mesh. During TSDF fusion, the voxel size and truncated threshold are set to 0.004 and 0.02 respectively. We set the hyperparameters of the loss function, $\\lambda_{1},\\lambda_{2},\\lambda_{3}$ , and $\\lambda_{4}$ , to 1000, 0.05, 1, and 0.2, respectively.", + "bbox": [ + 511, + 712, + 906, + 833 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Benchmark Comparisons", + "text_level": 1, + "bbox": [ + 511, + 844, + 746, + 859 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We compare the sparse-view reconstruction performance of different NeRF-based and Gaussian-based methods on", + "bbox": [ + 511, + 869, + 906, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "11312", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/444bf781e02a8f027249643b14b295cd4a1c74977a9ce4dca0e4f315b875e77a.jpg", + "image_caption": [ + "Figure 5. BlendedMVS reconstruction results. Our method achieves more complete and detailed reconstructions." + ], + "image_footnote": [], + "bbox": [ + 106, + 93, + 468, + 402 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/7c19cf3d27ff7a49df0c9e2ffa852823eb7047f401e402dc0ab97335c07fcd96.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
2DGSMVS Init.Fixed ColorFeat. Splat.DGPRSGUCD ↓
(a)✓2.813
(b)✓✓1.273
(c)✓✓✓1.212
(d)✓✓✓1.224
(e)✓✓✓✓1.201
(f)✓✓✓✓✓1.142
(g)✓✓✓✓✓1.152
(h)✓✓✓✓✓✓1.125
", + "bbox": [ + 94, + 448, + 480, + 573 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "the DTU dataset. As shown in Tab. 1, our method significantly improves upon Gaussian-based approaches under sparse views, outperforming baseline (2DGS) and previous Gaussian-based SOTA methods (PGSR) by $67\\%$ and $46\\%$ , respectively. Our method also outperforms SparseNeus [20], the first sparse-view NeRF-based method with notable margins. Our method performs competitively with the state-of-the-art NeRF-based surface reconstruction methods. The visualization results are shown in Fig. 4. Our method captures more surface details than implicit reconstruction approaches and outperforms Gaussian-based methods in preserving the complete scene structure. In comparison, Gaussian Splitting methods, i.e., 2DGS, GOF, and PGSR, tend to exhibit noise and missing areas. Additionally, we performed a qualitative comparison on the BlendedMVS dataset. As shown in Fig. 5, our method", + "bbox": [ + 89, + 657, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/4be6d13706e42ef75ce6f84e7688b1ef8320fbf89c8ac71855838743893c8590.jpg", + "table_caption": [ + "Table 2. Ablation on DTU [1] dataset. DGPR and SGU represent proposed Direct Gaussian Primitive Regularization and Selective Gaussian Update respectively. Each module improves upon the baseline, leading to the best final results." + ], + "table_footnote": [], + "table_body": "
Methods2DGS [11]PGSR [4]CLMVS. [31]Ours
3 view SfM Point Cloud2.812.08--
49/64 views SfM Point Cloud1.651.53--
CLMVSNet Point Cloud1.271.381.261.13
", + "bbox": [ + 517, + 88, + 903, + 148 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Comparison of initialization strategies. Initializing with MVS point cloud improves reconstruction performance. However, a plain combination (Chamfer Distance, CD:1.27) between 2DGS and CLMVSNet fails to surpass CLMVSNet results (CD: 1.26). Our method enhances both MVS and 2DGS through geometry-prioritized optimization.", + "bbox": [ + 511, + 155, + 906, + 239 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2af1b2913544c4ea42f028ee04b1610459dff33d8d7f5579ad35977d2f9b96de.jpg", + "image_caption": [ + "Figure 6. Comparison between different initialization strategies of 2DGS and our method. Initializing 2DGS with MVS point cloud improves SfM reconstruction completeness but struggles to capture details. Our method not only enhances completeness but also achieves detailed reconstruction." + ], + "image_footnote": [], + "bbox": [ + 514, + 257, + 888, + 479 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "achieves higher accuracy in reconstructing meshes for real-world scenes.", + "bbox": [ + 511, + 565, + 906, + 594 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Ablations and Analysis", + "text_level": 1, + "bbox": [ + 511, + 604, + 728, + 622 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we perform ablation studies to analyze the effectiveness of different modules. As shown in Tab. 2, using MVS point cloud initialization notably enhances the performance (row a). By fixing the colors and using feature splatting for geometric enhancement supervision, we further improve the accuracy and completeness of the reconstruction (row e). The Direct Gaussian Primitive Regularization (DGPR) enhances the surface modeling capability of each Gaussian primitive (row f). The Selective Gaussian Update (SGU) leverages rendering cues to enhance the precision of the Gaussian point set, further improving surface accuracy (row g).", + "bbox": [ + 509, + 628, + 906, + 809 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparison of different initialization strategies. We compare the reconstruction results using various initialization strategies for Gaussian Splatting methods. We show the results using 3 different initialization strategies: SfM point cloud from 3 viewpoints, SfM point cloud from 49 viewpoints, and the CLMVSNet reconstructed point cloud.", + "bbox": [ + 511, + 810, + 908, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "11313", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/f5eed4349c0e3d755d6fabeb2a149a38dbedbe38b4c9e17afc95f21bceced6ca.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
SettingsAccuracy↓Completion↓Average↓
SH=30.831.711.27
SH=20.841.681.26
SH=10.841.641.24
Fixed Color0.811.611.21
", + "bbox": [ + 109, + 88, + 464, + 159 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/2a8bdd83aae43113c33d60d8860183250fd987a0b53a876652d7693dce8d2298.jpg", + "table_caption": [ + "Table 4. Ablation study on appearance learning. Fixing the color helps in learning better geometry." + ], + "table_footnote": [], + "table_body": "
MethodsAccuracy↓Completeness↓Average↓
Baseline0.7901.6121.201
Adaptive Density Control [14]0.8141.6541.234
Selective Gaussian Update0.7601.5441.152
", + "bbox": [ + 93, + 196, + 486, + 258 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/36f83582a80326fac4c415fb033f74b859b43c5d3033abbba332d76da8f0a613.jpg", + "image_caption": [ + "Figure 7. Visualization of MVS feature splatting. Incorporating MVS feature splatting aids in achieving higher reconstruction accuracy in challenging regions." + ], + "image_footnote": [], + "bbox": [ + 94, + 325, + 468, + 419 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Using MVS point clouds achieves a more complete reconstruction than SfM. Note that as shown in Tab. 3, due to the inherent ill-posedness in sparse-view reconstruction, the plain combination of MVS and 2DGS reconstruction yields suboptimal performance. Even with CLMVSNet point clouds as initialization, the reconstruction quality still falls short of that achieved by CLMVSNet itself. In contrast, our geometric-prioritized enhancement addresses this challenge, with results excelling both CLMVSNet and 2DGS with higher completeness and finer details (Fig. 6).", + "bbox": [ + 88, + 474, + 482, + 626 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Geometrically enhanced supervision. In this paper, we mitigate appearance overfitting by fixing the color and feature parameters during optimization. In the original Gaussian splatting, view-independent color is represented by spherical harmonics (SH) with learnable coefficients. Intuitively, reducing SH degrees can achieve a similar goal in alleviating color overfitting. We compare the fixed color and reduced SH degrees in Tab. 4. Results show that reducing the SH degrees also mitigates the degradation in geometry, with only 1 SH degree outperforming other settings (i.e., 2, 3) in surface reconstruction. However, using learnable color coefficients still performs worse than the adopted choice of fixed color during training. Moreover, we enhance geometric supervision using MVS feature splatting. As shown in Fig. 7, MVS features are more expressive and provide more effective constraints, especially in challenging regions such as repetitive texture areas (e.g., the roof of the building).", + "bbox": [ + 89, + 627, + 482, + 883 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ablation on Direct Gaussian Primitive Regularization.", + "bbox": [ + 89, + 885, + 482, + 901 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/33b552213e1fd606aec1278c91c8866d33c371ae4967a66d4c1cf823a5f18833.jpg", + "table_caption": [ + "Table 5. Ablation on DTU [1] dataset. The proposed Selective Gaussian Update strategy enhances reconstruction accuracy and completeness compared to the generic Adaptive Density Control in sparse views." + ], + "table_footnote": [], + "table_body": "
SettingsAccuracy↓Completion↓Average↓
K=90.7461.5441.145
K=250.7471.5411.144
K=490.7511.5321.142
K=810.7471.5311.139
", + "bbox": [ + 517, + 88, + 905, + 161 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/c5790e3bedd394bda8389b46119c0cbdda3a83444efed989df7598127b698559.jpg", + "table_caption": [ + "Table 6. Comparison of different numbers of sampled points. Increasing the sampling numbers lead to a slight performance improvement." + ], + "table_footnote": [], + "table_body": "
MethodsSparseNeus [20]NeuSurf [12]Ours
Chamfer Distance1.270.991.13
Training Time25mins10 hours10mins
", + "bbox": [ + 526, + 210, + 895, + 265 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 7. Performance and efficiency. Our method achieves comparable performance while requiring the least training time.", + "bbox": [ + 511, + 268, + 903, + 297 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We analyze the impact of the reparameterization-based disk sampling number $K$ on the final results. As shown in Tab. 6, increasing the number of sampled points can slightly enhance the final reconstruction performance.", + "bbox": [ + 511, + 306, + 903, + 366 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Selective Gaussian update strategy. To validate the effectiveness of our selective update strategy, we compared it with the adaptive density control in the original 2DGS. Quantitative results are shown in the Tab. 5. Due to the lack of sufficient multi-view supervision, the densified Gaussian primitives are prone to degradation, leading to declined reconstruction performance (CD:1.23) compared to the baseline (CD: 1.20). The proposed selective Gaussian update strategy enhances the accuracy of the original Gaussian point set by utilizing high-precision points from the rendered cues, which further improves the surface reconstruction quality.", + "bbox": [ + 511, + 366, + 905, + 547 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Performance and efficiency. We compared reconstruction performance and training time with previous state-of-the-art surface reconstruction methods. As shown in Tab. 7, our method achieves better performance and $2 \\times$ faster training speed than SparseNeus [20]. Meanwhile, our method achieves competitive performance while being $60 \\times$ faster than NeuSurf [12].", + "bbox": [ + 511, + 547, + 905, + 654 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 667, + 633, + 684 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we propose Sparse2DGS, addressing the challenge of ill-posed geometry optimization from sparse views through geometry-prioritized enhancement schemes. Our method not only reconstructs accurate and complete scene surfaces but also achieves higher efficiency than NeRF-based methods. Limitation: Despite their effectiveness in dense matching, MVS methods inevitably encounter degradation under occlusions. The error in occluded areas can propagate to our method, leading to potential performance declines.", + "bbox": [ + 511, + 694, + 905, + 843 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements Y. Zhang was supported by NSFC (No.U19B2037) and the Natural Science Basic Research Program of Shaanxi (No.2021JCW-03). Y. Zhu was supported by NSFC (No.61901384).", + "bbox": [ + 511, + 844, + 903, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "11314", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 98, + 187, + 114 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Henrik Aanaes, Rasmus Ramsbøl Jensen, George Vogiatzis, Engin Tola, and Anders Bjorholm Dahl. Large-scale data for multiple-view stereopsis. International Journal of Computer Vision, 120:153-168, 2016. 5, 7, 8", + "[2] Chenjie Cao, Xinlin Ren, and Yanwei Fu. Mvsformer: Learning robust image representations via transformers and temperature-based depth for multi-view stereo. arXiv preprint arXiv:2208.02541, 2022. 2", + "[3] Di Chang, Aljaž Božić, Tong Zhang, Qingsong Yan, Yingcong Chen, Sabine Susstrunk, and Matthias Nießner. Rc-mvsnet: Unsupervised multi-view stereo with neural rendering. In European conference on computer vision, pages 665-680. Springer, 2022. 2, 5", + "[4] Danpeng Chen, Hai Li, Weicai Ye, Yifan Wang, Weijian Xie, Shangjin Zhai, Nan Wang, Haomin Liu, Hujun Bao, and Guofeng Zhang. Pgsr: Planar-based gaussian splatting for efficient and high-fidelity surface reconstruction. arXiv preprint arXiv:2406.06521, 2024. 1, 2, 6, 7", + "[5] Hanlin Chen, Chen Li, and Gim Hee Lee. Neusg: Neural implicit surface reconstruction with 3d gaussian splatting guidance. arXiv preprint arXiv:2312.00846, 2023. 1, 2", + "[6] Yikang Ding, Wentao Yuan, Qingtian Zhu, Haotian Zhang, Xiangyue Liu, Yuanjiang Wang, and Xiao Liu. Transmvsnet: Global context-aware multi-view stereo network with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8585-8594, 2022. 1, 2", + "[7] Lue Fan, Yuxue Yang, Minxing Li, Hongsheng Li, and Zhaoxiang Zhang. Trim 3d gaussian splatting for accurate geometry representation. arXiv preprint arXiv:2406.07499, 2024.1", + "[8] Qiancheng Fu, Qingshan Xu, Yew Soon Ong, and Wenbing Tao. Geo-neus: Geometry-consistent neural implicit surfaces learning for multi-view reconstruction. Advances in Neural Information Processing Systems, 35:3403-3416, 2022. 2", + "[9] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In 2012 IEEE conference on computer vision and pattern recognition, pages 3354-3361. IEEE, 2012. 1", + "[10] Antoine Guédon and Vincent Lepetit. Sugar: Surface-aligned gaussian splatting for efficient 3d mesh reconstruction and high-quality mesh rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5354-5363, 2024. 2", + "[11] Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2d gaussian splatting for geometrically accurate radiance fields. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 1, 2, 3, 4, 5, 6, 7", + "[12] Han Huang, Yulun Wu, Junsheng Zhou, Ge Gao, Ming Gu, and Yu-Shen Liu. Neusurf: On-surface priors for neural surface reconstruction from sparse input views. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2312-2320, 2024. 2, 3, 6, 8", + "[13] Michael Kazhdan, Matthew Bolitho, and Hugues Hoppe. Poisson surface reconstruction. In Proceedings of the fourth Eurographics symposium on Geometry processing, 2006. 2" + ], + "bbox": [ + 93, + 123, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4):139-1, 2023. 1, 2, 4, 8", + "[15] Tejas Khot, Shubham Agrawal, Shubham Tulsiani, Christoph Mertz, Simon Lucey, and Martial Hebert. Learning unsupervised multi-view stereopsis via robust photometric consistency. arXiv preprint arXiv:1905.02706, 2019. 2", + "[16] Rui Li, Dong Gong, Wei Yin, Hao Chen, Yu Zhu, Kaixuan Wang, Xiaozhi Chen, Jinqiu Sun, and Yanning Zhang. Learning to fuse monocular and multi-view cues for multiframe depth estimation in dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21539-21548, 2023. 1", + "[17] Rui Li, Tobias Fischer, Mattia Segu, Marc Pollefeys, Luc Van Gool, and Federico Tombari. Know your neighbors: Improving single-view reconstruction via spatial vision-language reasoning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9848-9858, 2024. 1", + "[18] Yinhao Li, Zheng Ge, Guanyi Yu, Jinrong Yang, Zengran Wang, Yukang Shi, Jianjian Sun, and Zeming Li. Bevdepth: Acquisition of reliable depth for multi-view 3d object detection. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1477-1485, 2023. 1", + "[19] Yixun Liang, Hao He, and Yingcong Chen. Retr: Modeling rendering via transformer for generalizable neural surface reconstruction. Advances in Neural Information Processing Systems, 36, 2024. 1, 3, 6", + "[20] Xiaoxiao Long, Cheng Lin, Peng Wang, Taku Komura, and Wenping Wang. Sparseneus: Fast generalizable neural surface reconstruction from sparse views. In European Conference on Computer Vision, pages 210-227. Springer, 2022. 1, 2, 5, 6, 7, 8", + "[21] Xuan Luo, Jia-Bin Huang, Richard Szeliski, Kevin Matzen, and Johannes Kopf. Consistent video depth estimation. ACM Transactions on Graphics (ToG), 39(4):71-1, 2020. 1", + "[22] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 1, 2", + "[23] Michael Oechsle, Songyou Peng, and Andreas Geiger. Unisurf: Unifying neural implicit surfaces and radiance fields for multi-view reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5589-5599, 2021. 1", + "[24] Ri-Zhao Qiu, Ge Yang, Weijia Zeng, and Xiaolong Wang. Feature splatting: Language-driven physics-based scene synthesis and editing. arXiv preprint arXiv:2404.01223, 2024. 4", + "[25] Yufan Ren, Fangjinhua Wang, Tong Zhang, Marc Pollefeys, and Sabine Susstrunk. Volrecon: Volume rendering of signed ray distance functions for generalizable multi-view reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16685-16695, 2023. 1, 2, 3, 6" + ], + "bbox": [ + 514, + 92, + 906, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "11315", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[26] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4104-4113, 2016. 3, 6", + "[27] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5459-5469, 2022. 2", + "[28] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. arXiv preprint arXiv:2106.10689, 2021. 1, 2, 6", + "[29] Yiqun Wang, Ivan Skorokhodov, and Peter Wonka. Improved surface reconstruction using high-frequency details. 2022. 2", + "[30] Jiang Wu, Rui Li, Haofei Xu, Wenxun Zhao, Yu Zhu, Jinqiu Sun, and Yanning Zhang. Gomvs: Geometrically consistent cost aggregation for multi-view stereo. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20207-20216, 2024. 1, 2", + "[31] Kaiqiang Xiong, Rui Peng, Zhe Zhang, Tianxing Feng, Jianbo Jiao, Feng Gao, and Ronggang Wang. Cl-mvsnet: Unsupervised multi-view stereo with dual-level contrastive learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3769-3780, 2023. 1, 2, 3, 4, 6, 7", + "[32] Jialei Xu, Xianming Liu, Yuanchao Bai, Junjun Jiang, Kaixuan Wang, Xiaozhi Chen, and Xiangyang Ji. Multi-camera collaborative depth prediction via consistent structure estimation. In Proceedings of the 30th ACM international conference on multimedia, pages 2730-2738, 2022. 2", + "[33] Jialei Xu, Wei Yin, Dong Gong, Junjun Jiang, and Xianming Liu. Sdge: Stereo guided depth estimation for $360^{\\circ}$ camera sets. In 2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 11179-11186. IEEE, 2024. 2", + "[34] Luoyuan Xu, Tao Guan, Yuesong Wang, Wenkai Liu, Zhaojie Zeng, Junle Wang, and Wei Yang. C2f2neus: Cascade cost frustum fusion for high fidelity and generalizable neural surface reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 18291-18301, 2023. 1, 3, 6", + "[35] Qingshan Xu, Wanjuan Su, Yuhang Qi, Wenbing Tao, and Marc Pollefeys. Learning inverse depth regression for pixelwise visibility-aware multi-view stereo networks. International Journal of Computer Vision, 130(8):2040-2059, 2022. 2", + "[36] Yao Yao, Zixin Luo, Shiwei Li, Tian Fang, and Long Quan. Mvsnet: Depth inference for unstructured multi-view stereo. In Proceedings of the European conference on computer vision (ECCV), pages 767-783, 2018. 1, 2", + "[37] Yao Yao, Zixin Luo, Shiwei Li, Jingyang Zhang, Yufan Ren, Lei Zhou, Tian Fang, and Long Quan. Blendedmvs: A large-scale dataset for generalized multi-view stereo networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1790-1799, 2020. 5", + "[38] Lior Yariv, Jiatao Gu, Yoni Kasten, and Yaron Lipman. Volume rendering of neural implicit surfaces. Advances in Neu" + ], + "bbox": [ + 91, + 92, + 480, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ral Information Processing Systems, 34:4805-4815, 2021. 1, 2, 6", + "[39] Jae-Chern Yoo and Tae Hee Han. Fast normalized cross-correlation. Circuits, systems and signal processing, 28:819-843, 2009. 5", + "[40] Mae Younes, Amine Ouasfi, and Adnane Boukhayma. Sparsecraft: Few-shot neural reconstruction through stereopsis guided geometric linearization. In European Conference on Computer Vision, pages 37-56. Springer, 2024. 3, 6", + "[41] Mulin Yu, Tao Lu, Linning Xu, Lihan Jiang, Yuanbo Xiangli, and Bo Dai. Gsdf: 3dgs meets sdf for improved rendering and reconstruction. arXiv preprint arXiv:2403.16964, 2024. 1, 2", + "[42] Zehao Yu, Songyou Peng, Michael Niemeyer, Torsten Sattler, and Andreas Geiger. Monosdf: Exploring monocular geometric cues for neural implicit surface reconstruction. Advances in neural information processing systems, 35:25018-25032, 2022. 2", + "[43] Zehao Yu, Torsten Sattler, and Andreas Geiger. Gaussian opacity fields: Efficient and compact surface reconstruction in unbounded scenes. arXiv preprint arXiv:2404.10772, 2024. 1, 2, 6", + "[44] Baowen Zhang, Chuan Fang, Rakesh Shrestha, Yixun Liang, Xiaoxiao Long, and Ping Tan. Rade-gs: Rasterizing depth in gaussian splatting. arXiv preprint arXiv:2406.01467, 2024. 2", + "[45] Jingyang Zhang, Yao Yao, and Long Quan. Learning signed distance field for multi-view surface reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6525-6534, 2021. 5", + "[46] Jingyang Zhang, Shiwei Li, Zixin Luo, Tian Fang, and Yao Yao. Vis-mvsnet: Visibility-aware multi-view stereo network. International Journal of Computer Vision, 131(1): 199-214, 2023. 2", + "[47] Shijie Zhou, Haoran Chang, Sicheng Jiang, Zhiwen Fan, Zehao Zhu, Dejia Xu, Pradyumna Chari, Suya You, Zhangyang Wang, and Achuta Kadambi. Feature 3dgs: Supercharging 3d gaussian splattering to enable distilled feature fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21676-21685, 2024. 4" + ], + "bbox": [ + 516, + 92, + 903, + 657 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "11316", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Sparse2DGS_ Geometry-Prioritized Gaussian Splatting for Surface Reconstruction from Sparse Views/3c5dfe86-080c-4ec1-9ca9-65bf2b3453a3_model.json b/2025/Sparse2DGS_ Geometry-Prioritized Gaussian Splatting for Surface Reconstruction from Sparse Views/3c5dfe86-080c-4ec1-9ca9-65bf2b3453a3_model.json new file mode 100644 index 0000000000000000000000000000000000000000..783229f178a0e69fc4ae970495d967f267148e5a --- /dev/null +++ b/2025/Sparse2DGS_ Geometry-Prioritized Gaussian Splatting for Surface Reconstruction from Sparse Views/3c5dfe86-080c-4ec1-9ca9-65bf2b3453a3_model.json @@ -0,0 +1,2112 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.203, + 0.131, + 0.798, + 0.177 + ], + "angle": 0, + "content": "Sparse2DGS: Geometry-Prioritized Gaussian Splating for Surface Reconstruction from Sparse Views" + }, + { + "type": "text", + "bbox": [ + 0.207, + 0.203, + 0.792, + 0.241 + ], + "angle": 0, + "content": "Jiang Wu Rui Li Yu Zhu† Rong Guo Jinqiu Sun Yanning Zhang† Northwestern Polytechnical University" + }, + { + "type": "image", + "bbox": [ + 0.104, + 0.281, + 0.898, + 0.441 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.452, + 0.908, + 0.496 + ], + "angle": 0, + "content": "Figure 1. Sparse2DGS boosts the strengths of Gaussian Splatting and Multi-view Stereo in sparse-view surface reconstruction, with notable improvement over 2DGS [11], CLMVSNet [31], as well as their plain combination (left). Meanwhile, Sparse2DGS is \\(2 \\times\\) faster than NeRF-based sparse-view method [20] (middle) and leads to complete and accurate reconstruction (right)." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.507, + 0.328, + 0.523 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.539, + 0.486, + 0.781 + ], + "angle": 0, + "content": "We present a Gaussian Splatting method for surface reconstruction using sparse input views. Previous methods relying on dense views struggle with extremely sparse Structure-from-Motion points for initialization. While learning-based Multi-view Stereo (MVS) provides dense 3D points, directly combining it with Gaussian Splatting leads to suboptimal results due to the ill-posed nature of sparse-view geometric optimization. We propose Sparse2DGS, an MVS-initialized Gaussian Splatting pipeline for complete and accurate reconstruction. Our key insight is to incorporate the geometric-prioritized enhancement schemes, allowing for direct and robust geometric learning under ill-posed conditions. Sparse2DGS outperforms existing methods by notable margins while being \\(2 \\times\\) faster than the NeRF-based fine-tuning approach. Code is available at https://github.com/Wuuu3511/Sparse2DGS." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.808, + 0.222, + 0.823 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.833, + 0.485, + 0.88 + ], + "angle": 0, + "content": "Reconstructing 3D surfaces from multiple views is a fundamental task in computer vision and has been widely used in autonomous driving [9, 18] and virtual reality [21]. While" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.509, + 0.907, + 0.568 + ], + "angle": 0, + "content": "typical methods require abundant input images, reconstructing accurate 3D surfaces from sparse views is particularly challenging and is crucial for scenes that are frequently captured with limited images." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.568, + 0.909, + 0.702 + ], + "angle": 0, + "content": "NeRF-based surface reconstruction methods [22] have gained popularity by encoding scene geometry into an occupancy field [17, 23] or a signed distance function (SDF) [28, 38]. However, when faced with sparse views, these methods often yield incomplete surfaces with artifacts due to insufficient supervision. Although sparse-view methods [19, 20, 25, 34] use pre-trained feed-forward NeRFs to provide geometric priors, they still exhibit limited generalization to unseen scenes." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.704, + 0.909, + 0.87 + ], + "angle": 0, + "content": "Recently, 3D Gaussian Splatting (3DGS) [14] has leveraged Structure-from-Motion (SfM) point cloud to construct Gaussian primitives for fast and high-quality novel view synthesis (NVS). Subsequent works [4, 5, 7, 11, 41, 43] aim to achieve more efficient 3D surface reconstruction using 3DGS. Notably, 2D Gaussian Splatting (2DGS) [11] proposes a 2D variant of Gaussian primitives, with disk-shaped primitives suitable for geometry representation. However, when dealing with sparse views, existing methods [4, 11, 43] struggle to produce faithful reconstructions due to initially over-sparse SfM points." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.908, + 0.902 + ], + "angle": 0, + "content": "On the other hand, learning-based multi-view stereo (MVS) [6, 16, 30, 36] predicts dense depth maps from" + }, + { + "type": "page_footnote", + "bbox": [ + 0.114, + 0.887, + 0.298, + 0.901 + ], + "angle": 0, + "content": "\\(^\\dagger\\) indicates corresponding authors." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11307" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.212 + ], + "angle": 0, + "content": "pixel-wise feature matching, significantly improving geometric completeness even with sparse images. However, MVS depth networks focus on generalizable inference instead of test-time geometry enhancement by design. While combining MVS point cloud with Gaussian Splatting is intuitively effective, their plain combination (see Fig. 1, \"2DGS+CLMVS\") yields sub-optimal results due to the ill-posed nature of geometry optimization from sparse views." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.214, + 0.483, + 0.531 + ], + "angle": 0, + "content": "In this paper, we propose Sparse2DGS, an MVS-initialized Gaussian Splating method for complete and accurate sparse-view 3D surface reconstruction. The key idea is to boost the strengths of two paradigms through geometric-prioritized enhancement against the ill-posed nature of sparse-view reconstruction. Starting from MVS initialization, we first incorporate MVS-derived geometric features into the Gaussian splattering process while keeping feature and color values fixed. This provides geometrically expressive supervision and prevents appearance overfitting, thereby improving geometric reconstruction quality. We further introduce a direct Gaussian primitive regularization, with a reparameterization-based disk sampling strategy that reformulates Gaussian primitives' position, orientation, and scale into point representations. This approach enables a unified optimization of different forms of Gaussian properties within a standard point-based, cross-view consistency loss. Moreover, instead of densifying Gaussian primitives under ill-posed conditions, we propose to selectively update the MVS-initialized Gaussian primitives with rendered geometric cues, allowing for self-guided geometric corrections." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.532, + 0.483, + 0.606 + ], + "angle": 0, + "content": "Sparse2DGS significantly outperforms existing Gaussian Splatting methods in sparse-view surface reconstruction. Meanwhile, it demonstrates notable efficiency advantages over NeRF-based test-time optimization. Our contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.608, + 0.483, + 0.669 + ], + "angle": 0, + "content": "- We propose Sparse2DGS, a test-time sparse-view 3D surface reconstruction method that boosts the strengths of Gaussian Splitting and MVS, allowing for complete and accurate reconstruction from sparse images." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.669, + 0.483, + 0.805 + ], + "angle": 0, + "content": "- We focus on geometric-prioritized enhancement by proposing geometrically enhanced supervision, direct Gaussian primitive regularization, and selective Gaussian update throughout the Gaussian Splatting pipeline. Sparse2DGS achieves the lowest error (1.13 CD error) compared to 2DGS (2.81) [11], GOF (2.82) [43], and PGSR (2.08) [4] on DTU dataset with 3 views. Meanwhile, Sparse2DGS is \\(2 \\times\\) faster than fine-tuning NeRF-based SparseNeus [20]." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.608, + 0.483, + 0.805 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.821, + 0.24, + 0.836 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.847, + 0.465, + 0.862 + ], + "angle": 0, + "content": "2.1. 3D Reconstruction via Novel View Synthesis" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.9 + ], + "angle": 0, + "content": "Recent progresses [11, 12, 25, 43] in novel view synthesis (NVS) have enabled accurate 3D surface reconstruc" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.257 + ], + "angle": 0, + "content": "tion with only image-based supervision. Neural Radiance Fields [22] (NeRFs) have become a significant technique for representing 3D scenes. It has inspired numerous works on directly extracting surfaces by training a parameterized radiance field through differentiable volume rendering. Neus [28] and VolSDF [38] integrate surface learning into volume rendering by re-parameterizing the signed distance function as volume density, reconstructing more accurate surfaces. To reduce geometric deviations and accelerate training, subsequent works [8, 27, 29, 42] have introduced improvements in various areas." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.258, + 0.905, + 0.56 + ], + "angle": 0, + "content": "Recently, 3D Gaussian Splatting [14] has shown impressive results in novel view synthesis (NVS), with many recent studies adopting this approach for surface reconstruction. SuGaR [10] aligns 3D Gaussians with scene surfaces by designing a regularization term and extracts an accurate mesh from the Gaussians using Poisson Reconstruction [13]. 2DGS [11] replaces 3D Gaussian ellipsoids with 2D-oriented planar disks, providing view-consistent geometric representations. RaDe-GS [44] further designs a rasterized approach to render depth and normal, enhancing geometric accuracy while ensuring rendering efficiency. GOF [43] extracts the level set from trained Gaussian primitives by constructing a Gaussian opacity field. PGSR [4] improves mesh quality by introducing single-view and multi-view geometric constraints. Additionally, GSDF [41] and NeusSG [5] reconstruct scenes by combining the advantages of 3D Gaussian Splatting and SDF networks. Despite these methods' success, the extracted geometry tends to be noisy and incomplete when the input number of training viewpoints is limited." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.571, + 0.813, + 0.587 + ], + "angle": 0, + "content": "2.2. Learning-based Multi-view Stereo" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.595, + 0.905, + 0.821 + ], + "angle": 0, + "content": "Deep learning-based Multi-view Stereo (MVS) estimates the reference image's depth from multi-view posed images. MVSNet [36] first introduces the differentiable homography warping to build the cost volume with deep image features and then employs a 3D CNN for cost regularization. Recent follow-up works have introduced improvements in various aspects, including multi-view feature extraction [2, 6], cost aggregation [32, 33, 35, 46], and cost volume regularization [30]. However, these methods require ground-truth depth maps for training, which are challenging to obtain in practice. Unsupervised MVS [3, 15, 31] can be applied in more general, real-world scenarios. MVS methods can infer dense depth maps for each input view and therefore, we select the unsupervised CLMVSNet [31] to initialize the Gaussian point cloud." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.832, + 0.829, + 0.848 + ], + "angle": 0, + "content": "2.3. Sparse View Surface Reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Recent works have explored 3D surface reconstruction from sparse training views. SparseNeuS [20] achieves reasonable mesh reconstruction results by constructing a cascaded" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11308" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.087, + 0.9, + 0.322 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.332, + 0.907, + 0.417 + ], + "angle": 0, + "content": "Figure 2. Overview. Given sparse posed images, we first leverage the MVS points to initialize the Gaussian position \\(\\mathbf{p}\\) (Sec. 3.1). We then leverage the MVS-derived feature for Gaussian Splating, with fixed feature \\(\\mathbf{f}\\) and color \\(\\mathbf{c}\\) values to conduct geometrically expressive supervision and avoid appearance overfitting (Sec. 3.2). We then optimize Gaussian primitive properties by reformulating orientation \\(\\mathbf{R}\\), scale \\(\\mathbf{S}\\), and position \\(\\mathbf{p}\\) into sampled points using the reparameterization-based disk sampling, ensuring simultaneously optimizing different forms of Gaussian properties with point-based loss (Sec. 3.3). Finally, we introduce a Selective Gaussian Update strategy that leverages rendered geometry as an alternative to update Gaussian primitive positions, which further improves reconstruction quality (Sec. 3.4)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.43, + 0.484, + 0.598 + ], + "angle": 0, + "content": "volume to learn a generalizable radiance field, followed by per-scene fine-tuning. NeuSurf [12] learns dense on-surface priors from sparse points to guide geometric learning. SparseCraft [40] leverages stereopsis cues for geometric regularization based on Taylor expansion. Moreover, VolRecon [25] and ReTR [19] enhance voxel features by designing specialized transformer architectures. C2F2Neus [34] constructs per-view cost frustum for finer geometry estimation. Although these feedforward networks can directly infer scene geometry, their limited generalization to unseen scenes restricts their applicability in real-world scenarios." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.604, + 0.226, + 0.621 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.629, + 0.483, + 0.84 + ], + "angle": 0, + "content": "Given sparse view posed images \\(\\{\\mathbf{I}_i\\}_{i=1}^N\\), where \\(\\mathbf{I}_i \\in \\mathbb{R}^{H \\times W \\times 3}\\) and \\(N\\) denotes the view number, Sparse2DGS aims to reconstruct accurate 3D scene surfaces using Gaussian Splatting. As shown in Fig. 2, we first initialize the Gaussian primitives with the MVS [31] reconstructed point cloud. Then, we sample MVS geometric feature maps into feature splatting with its value and RGB color fixed for supervision. For each Gaussian primitive, we regularize its geometric properties (position, orientation, scale) by a novel reparameterization-based 2D disk sampling strategy that formulates these properties as point-based supervision. Finally, we propose a selective Gaussian update strategy to further enhance Gaussian primitives using rendered geometric cues." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.848, + 0.483, + 0.865 + ], + "angle": 0, + "content": "3.1. 2D Gaussian Splitting with MVS initialization" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.903 + ], + "angle": 0, + "content": "2D Gaussian Splitting [11] (2DGS) represents Gaussian primitives as planar disks suitable for representing and op-" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.43, + 0.907, + 0.505 + ], + "angle": 0, + "content": "timizing 3D surfaces. However, it relies on sparse SfM [26] 3D points for initialization, which exhibits limited reconstruction quality under sparse-view inputs. To address this, we initialize 2DGS with an MVS point cloud to provide dense geometric representations." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.506, + 0.909, + 0.628 + ], + "angle": 0, + "content": "2D Gaussian Splatting. Each 2D Gaussian primitive is characterized by its central position \\(\\mathbf{p}\\), two principal tangential vectors \\(\\mathbf{t_u}\\) and \\(\\mathbf{t_v}\\), and a scaling vector \\((s_u, s_v)\\). The rotation matrix of a 2D Gaussian can be represented as \\(\\mathbf{R} = (\\mathbf{t_u}, \\mathbf{t_v}, \\mathbf{t_z})\\), where \\(\\mathbf{t_z}\\) is defined by two orthogonal tangent vectors \\(\\mathbf{t_z} = \\mathbf{t_u} \\times \\mathbf{t_v}\\) and the scale matrix is \\(\\mathbf{S} = \\mathrm{diag}(s_u, s_v, 0)\\). A 2D Gaussian is then defined in a local tangent plane in world space, which is parameterized:" + }, + { + "type": "equation", + "bbox": [ + 0.604, + 0.64, + 0.907, + 0.657 + ], + "angle": 0, + "content": "\\[\nP (u, v) = \\mathbf {p} + s _ {u} \\mathbf {t} _ {u} u + s _ {v} \\mathbf {t} _ {v} v. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.668, + 0.906, + 0.698 + ], + "angle": 0, + "content": "For the point \\(\\mathbf{u} = (u,v)\\) in \\(uv\\) space, its 2D Gaussian value can then be evaluated by standard Gaussian" + }, + { + "type": "equation", + "bbox": [ + 0.624, + 0.709, + 0.906, + 0.741 + ], + "angle": 0, + "content": "\\[\n\\mathcal {G} (\\mathbf {u}) = \\exp \\frac {- \\left(u ^ {2} + v ^ {2}\\right)}{2}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Initializing 2DGS with MVS depth. MVS methods estimate dense depth maps through cross-view matching. In particular, unsupervised MVS [31] reconstructs complete scenes without ground truth for training, offering greater flexibility. Therefore, we use the point cloud reconstructed by the current unsupervised method CLMVSNet [31] to initialize our Gaussian point cloud. Specifically, given sparse views' images \\(\\{\\mathbf{I}_i\\}_{i=1}^N\\), we use the MVS network for depth maps \\(\\{\\mathbf{D}_i\\}_{i=1}^N\\), then back-project these depth maps and fuse into a point cloud \\(\\{\\mathbf{p}_{ij} \\mid 1 \\leq i \\leq N, 1 \\leq j \\leq HW\\}\\) as ini-" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "11309" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.123 + ], + "angle": 0, + "content": "tialized positions for Gaussian primitives, where \\(i\\) denotes image index and \\(j\\) is the pixel index within each image." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.131, + 0.414, + 0.148 + ], + "angle": 0, + "content": "3.2. Geometrically Enhanced Supervision" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.155, + 0.484, + 0.306 + ], + "angle": 0, + "content": "3D/2D Gaussian Splatting [11, 14] couples color and geometry optimization through image-based loss. However, this leads to severe degradation under sparse views as color will overfit sparse views, leading to a compromised geometry. Meanwhile, the challenges in image areas (e.g., textureless regions) will be amplified with sparse views. To address this issue, we propose geometrically enhanced supervision by incorporating off-the-shelf expressive MVS features for feature splatting and supervision, with the color and feature fixed to avoid overfitting." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.307, + 0.484, + 0.458 + ], + "angle": 0, + "content": "In addition to providing dense point clouds from input images \\(\\{\\mathbf{I}_i\\}_{i = 1}^N\\), the MVS network [31] offers off-the-shelf feature maps with geometric expressiveness. We leverage MVS features to enhance geometric supervision by incorporating additional feature attribute \\(\\mathbf{f}_{ij}\\) into Gaussian primitives. Specifically, for each input image \\(\\mathbf{I}_i\\), we extract its feature map \\(\\mathbf{F}_i\\) from the Feature Pyramid Network (FPN) of CLMVSNet [31]. Since each 3D point corresponds to an image pixel due to MVS initialization, we sample both RGB color and MVS features into each Gaussian primitive." + }, + { + "type": "equation", + "bbox": [ + 0.168, + 0.47, + 0.483, + 0.489 + ], + "angle": 0, + "content": "\\[\n\\mathbf {f} _ {i j} = \\mathbf {F} _ {i} (\\pi (\\mathbf {p} _ {i j})), \\mathbf {c} _ {i j} = \\mathbf {I} _ {i} (\\pi (\\mathbf {p} _ {i j})), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.498, + 0.484, + 0.575 + ], + "angle": 0, + "content": "where \\(\\pi(\\cdot)\\) denotes the projection from 3D space onto 2D image plane. We further render the feature into an image plane using the tile-based renderer together with RGB rendering. We show the rendering process given \\(M\\) Gaussian primitives along a ray, with indices \\(i, j\\) omitted for brevity" + }, + { + "type": "equation", + "bbox": [ + 0.142, + 0.598, + 0.483, + 0.64 + ], + "angle": 0, + "content": "\\[\n\\{\\hat {\\mathbf {c}}, \\hat {\\mathbf {f}} \\} = \\sum_ {m = 1} ^ {M} \\left\\{\\mathbf {c} _ {m}, \\mathbf {f} _ {m} \\right\\} \\cdot \\alpha_ {m} \\prod_ {o = 1} ^ {m - 1} (1 - \\alpha_ {m}), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.65, + 0.484, + 0.711 + ], + "angle": 0, + "content": "where \\(\\alpha_{m}\\) is the opacity of the Gaussian. With rendered feature maps that have enhanced geometric expressiveness for the challenging sparse views, we further supervise the rendered features using cosine similarity loss by" + }, + { + "type": "equation", + "bbox": [ + 0.205, + 0.722, + 0.483, + 0.76 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {f e a} = 1 - \\frac {\\mathbf {f} _ {i j} \\cdot \\hat {\\mathbf {f}} _ {i j}}{\\| \\mathbf {f} _ {i j} \\| \\| \\hat {\\mathbf {f}} _ {i j} \\|}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.841 + ], + "angle": 0, + "content": "Note that the simultaneous learning of appearance (i.e., color \\(\\mathbf{c}_{ij}\\) and feature \\(\\mathbf{f}_{ij}\\)) and geometry lead to geometric degradation under sparse views (see Tab. 4). To address this issue, both \\(\\mathbf{f}_{ij}\\) and \\(\\mathbf{c}_{ij}\\) are fixed during the optimization to enforce the learning specifically for geometry." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.841, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Our method essentially differentiates previous feature-based methods [24, 47] in its specific strategy for geometric enhancement. In our method, the expressive MVS features are incorporated, and features of each Gaussian are fixed," + }, + { + "type": "image", + "bbox": [ + 0.533, + 0.093, + 0.885, + 0.281 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.296, + 0.907, + 0.407 + ], + "angle": 0, + "content": "Figure 3. Direct Gaussian Primitive Regularization. We sample point set \\(\\mathbf{Z}' \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})\\) from the standard Gaussian distribution and leverage the reparameterization technique to transform \\(\\mathbf{Z}'\\) to the point set \\(\\mathbf{Z}\\) on each Gaussian primitive. This process allows for representing Gaussian orientation \\(\\mathbf{R}\\) and scale \\(\\mathbf{S}\\) properties in the form of points easier for geometric supervision. The transformed points \\(\\mathbf{Z}\\) are then projected to the source and target views, supervised by cross-view feature consistency loss." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.424, + 0.906, + 0.485 + ], + "angle": 0, + "content": "both for geometry enhancement with sparse views. Previous methods leverage splatted features only for segmentation or editing [24, 47] with dense views, while their feature losses do not optimize the scene geometry." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.497, + 0.868, + 0.513 + ], + "angle": 0, + "content": "3.3. Direct Gaussian Primitive Regularization" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.521, + 0.906, + 0.642 + ], + "angle": 0, + "content": "The disk-shaped 2D Gaussian primitives encode latent scene geometry including local shape orientation \\(\\mathbf{R}\\), scale \\(\\mathbf{S}\\), and position \\(\\mathbf{p}\\), which are essential for accurate surface reconstruction. These geometric properties are typically learned from surrogate image render loss supported by abundant multi-view supervision. When facing sparse-view inputs, the primitives tend to degrade due to insufficient geometric constraints." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.644, + 0.907, + 0.809 + ], + "angle": 0, + "content": "To address this issue, we incorporate direct geometric regularization for 2D Gaussian primitives through multiview consistency guidance. As multi-view consistency approaches only consider point position, we incorporate broad Gaussian primitive properties (i.e., orientation, scale, and position) by correlating their relationship with point position, through a novel reparameterization-based disk sampling strategy. This regularization aligns 2D Gaussian primitives more closely with the latent scene surfaces, leading to improved geometric learning. In the following illustration, we omit the Gaussian primitive indices \\( i,j \\) for brevity." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Regularizing points with cross-view consistency. Given 3D point sampling positions \\(\\mathbf{Z} = \\{\\mathbf{z}_k\\}_{k=1}^K\\) on each Gaussian, we project them to source and target views to obtain the corresponding 2D positions \\(\\mathbf{q}_k^{src}\\) and \\(\\mathbf{q}_k^{tar}\\), as well as queried MVS features \\(\\mathbf{F}_{src}(\\mathbf{q}_k^{src})\\), \\(\\mathbf{F}_{tar}(\\mathbf{q}_k^{tar})\\). We can then optimize the 3D positions by cross-view feature consistency" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11310" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.186, + 0.107 + ], + "angle": 0, + "content": "loss [45] \\(\\mathcal{L}_{df}\\)" + }, + { + "type": "equation", + "bbox": [ + 0.118, + 0.12, + 0.482, + 0.161 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {d f} = \\frac {1}{K} \\sum_ {k = 1} ^ {K} (1 - \\frac {\\mathbf {F} _ {s r c} \\left(\\mathbf {q} _ {k} ^ {s r c}\\right) \\cdot \\mathbf {F} _ {t a r} \\left(\\mathbf {q} _ {k} ^ {t a r}\\right)}{\\| \\mathbf {F} _ {s r c} \\left(\\mathbf {q} _ {k} ^ {s r c}\\right) \\| \\| \\mathbf {F} _ {t a r} \\left(\\mathbf {q} _ {k} ^ {t a r}\\right) \\|}), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.172, + 0.483, + 0.202 + ], + "angle": 0, + "content": "where \\(K\\) denotes the number of sampled points on each disk-shaped Gaussian primitive." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.203, + 0.484, + 0.351 + ], + "angle": 0, + "content": "Reparameterization-based disk sampling. The above operation focuses on position regularization only, and thus, it cannot propagate gradients to other Gaussian primitive properties such as orientation \\(\\mathbf{R}\\) and scale \\(\\mathbf{S}\\). We propose a reparameterization-based sampling strategy to solve this problem. As shown in Fig. 3, for each Gaussian primitive, we first sample points \\(\\mathbf{Z}'\\) from the standard Gaussian distribution \\(\\mathcal{N}(\\mathbf{0},\\mathbf{I})\\) and then transform the distribution to the target Gaussian primitive through its position \\(\\mathbf{p}\\), orientation \\(\\mathbf{R}\\) and scale \\(\\mathbf{S}\\)" + }, + { + "type": "equation", + "bbox": [ + 0.237, + 0.354, + 0.483, + 0.37 + ], + "angle": 0, + "content": "\\[\n\\mathbf {Z} = \\mathbf {p} + \\mathbf {L} \\mathbf {Z} ^ {\\prime}, \\tag {7}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.254, + 0.379, + 0.482, + 0.394 + ], + "angle": 0, + "content": "\\[\n\\mathbf {L} = \\mathbf {R S}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.402, + 0.484, + 0.522 + ], + "angle": 0, + "content": "where \\(\\mathbf{Z}' \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})\\). This bridges the gap between Gaussian properties (orientation and scale) and point representation, ensuring regularizing Gaussian geometric properties merely through the surrogate point-based loss. Additionally, since the last scale value of the 2D Gaussian primitive [11] is zero, resulting in difficulties in optimizing the normal vector \\(\\mathbf{R}[\\dots,2]\\) in the last dimension, we apply additional supervision using the rendered normal \\(\\mathbf{n}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.203, + 0.534, + 0.483, + 0.553 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {d n} = 1 - \\mathbf {R} [ \\dots , 2 ] ^ {\\top} \\mathbf {n}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.565, + 0.462, + 0.58 + ], + "angle": 0, + "content": "The final primitive regularization loss can be defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.223, + 0.593, + 0.482, + 0.609 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {d r} = \\mathcal {L} _ {d f} + \\mathcal {L} _ {d n}. \\tag {10}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.621, + 0.332, + 0.637 + ], + "angle": 0, + "content": "3.4. Selective Gaussian Update" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.644, + 0.483, + 0.78 + ], + "angle": 0, + "content": "Gaussian Splitting leverages adaptive density control to densify the Gaussian primitives. However, with sparse views, the densified Gaussian primitives are prone to degrade without sufficient multi-view supervision. To address this issue, as shown in Fig. 2, instead of continually densifying the MVS-initialized dense Gaussian primitives with limited supervision, we propose to selectively update them guided by the alternative rendered geometric cues, e.g., depth and normal." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Evaluating Gaussian primitives via patch warping. Given any Gaussian primitive with its position \\(\\mathbf{p}\\) and orientation \\(\\mathbf{R}\\), we can warp an image patch from one source view to a target view using its position \\(\\mathbf{p}\\) and orientation \\(\\mathbf{R}\\) together with the pose, following homography warping techniques [3]. Then, we can evaluate the geometric quality of the Gaussian primitive via the patch-based Normalized Cross-Correlation (NCC) [39] metric." + }, + { + "type": "equation", + "bbox": [ + 0.525, + 0.11, + 0.905, + 0.148 + ], + "angle": 0, + "content": "\\[\n\\operatorname {N C C} (x, y) = \\frac {\\sum_ {i = l} ^ {n} \\left(x _ {l} - \\bar {x}\\right) \\left(y _ {l} - \\bar {y}\\right)}{\\sqrt {\\sum_ {l = 1} ^ {n} \\left(x _ {l} - \\bar {x}\\right) ^ {2}} \\sqrt {\\sum_ {l = 1} ^ {n} \\left(y _ {l} - \\bar {y}\\right) ^ {2}}}, \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.152, + 0.905, + 0.182 + ], + "angle": 0, + "content": "where \\( x_{l} \\) and \\( y_{l} \\) represent the pixels in the patches \\( x \\) and \\( y \\) of the source and target images, respectively." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.183, + 0.905, + 0.288 + ], + "angle": 0, + "content": "Update Gaussian primitive with rendered cues. In addition to using the Gaussian primitive parameter \\(\\mathbf{p}\\) and orientation \\(\\mathbf{R}\\) for patch warping, we can leverage the rendered pixel-wise depth and normals from tiled Gaussian primitives as an alternative for patch warping. We compute patch-wise NCC over two warping strategies, yielding \\(\\mathrm{NCC}_G\\) for Gaussian primitive and \\(\\mathrm{NCC}_R\\) for rendered cues." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.288, + 0.906, + 0.409 + ], + "angle": 0, + "content": "If \\(\\mathrm{NCC}_R\\) is better than \\(\\mathrm{NCC}_G\\), we reproject the rendered depth into a 3D point and update the position of the corresponding Gaussian primitive. We update the position of Gaussian primitives every 100 training steps, after which all Gaussian parameters continue to be optimized by gradients. In this way, we improve the initial Gaussian primitive with geometric cues from rendered results, further enhancing the reconstruction quality." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.416, + 0.651, + 0.433 + ], + "angle": 0, + "content": "3.5. Training loss" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.44, + 0.906, + 0.515 + ], + "angle": 0, + "content": "During training, in addition to using the disk regularization loss \\(\\mathcal{L}_{dr}\\) and MVS feature splatting loss \\(\\mathcal{L}_{fea}\\), we retain the original 2DGS [11] image reconstruction loss \\(\\mathcal{L}_{rgb}\\), depth-normal consistency loss \\(\\mathcal{L}_n\\), and depth distortion loss \\(\\mathcal{L}_d\\). The final loss function is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.536, + 0.522, + 0.905, + 0.54 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {r g b} + \\lambda_ {1} \\mathcal {L} _ {d} + \\lambda_ {2} \\mathcal {L} _ {n}, + \\lambda_ {3} \\mathcal {L} _ {d r} + \\lambda_ {4} \\mathcal {L} _ {f e a}, \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.546, + 0.905, + 0.577 + ], + "angle": 0, + "content": "where \\(\\lambda_1, \\lambda_2, \\lambda_3, \\lambda_4\\) denotes the hyperparameters for the weights of each loss term." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.588, + 0.646, + 0.605 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.614, + 0.906, + 0.705 + ], + "angle": 0, + "content": "In this section, we first introduce the implementation details and then provide quantitative and qualitative comparisons with current methods on DTU [1] dataset, as well as qualitative results on BlendedMVS [37] dataset. We further analyze the contribution of each module and compare them with different methods through ablation studies." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.712, + 0.616, + 0.726 + ], + "angle": 0, + "content": "4.1. Datasets" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.907, + 0.901 + ], + "angle": 0, + "content": "The DTU [1] dataset consists of 128 scenes captured in controlled lab environments using structured light scanners containing 49 or 64 images. Similar to SparseNeuS [20], we select 15 scenes and use views 23, 24, and 33 for 3-view reconstruction, with an image resolution of \\(576 \\times 768\\). We perform the quantitative evaluation on the DTU dataset using the official evaluation code to calculate the Chamfer Distance between the reconstructed mesh and the ground truth point cloud. The BlendedMVS [37] dataset includes several large-scale real-world scenes. We also show the qualitative reconstruction comparisons." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "11311" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.1, + 0.089, + 0.9, + 0.284 + ], + "angle": 0, + "content": "
Scan243740556365698397105106110114118122Mean
Colmap [26]0.902.891.631.082.181.941.611.302.341.281.101.420.761.171.141.52
CLMVSNet [31]1.172.521.680.891.511.820.991.291.520.780.881.100.641.041.121.26
Neus [28]4.574.493.974.324.631.954.683.834.152.501.526.471.265.576.114.00
Volsdf [38]4.034.216.120.918.241.732.741.825.143.092.084.810.603.512.183.41
VolRecon [25]1.202.591.561.081.431.921.111.481.421.051.191.380.741.231.271.38
SparseNeus [20]1.292.271.570.881.611.861.061.271.421.070.990.870.541.151.181.27
Retr [19]1.052.311.440.981.181.520.881.351.300.871.070.770.591.051.121.17
C2F2Neus [34]1.122.421.400.751.411.770.851.161.260.760.910.600.460.880.921.11
Sparsecraft [40]1.171.741.800.701.191.530.831.051.420.780.800.560.440.770.841.04
NeuSurf [12]0.782.351.550.751.041.680.601.140.980.700.740.490.390.750.860.99
2DGS [11]3.143.792.311.515.342.221.852.793.621.633.054.111.053.352.442.81
GOF [43]3.275.052.772.043.842.612.232.653.301.652.584.441.162.292.392.82
PGSR [4]3.363.282.751.275.151.840.881.793.491.191.861.110.611.091.522.08
Ours1.052.351.380.831.371.450.841.161.430.740.850.840.570.951.011.13
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.289, + 0.907, + 0.318 + ], + "angle": 0, + "content": "Table 1. Quantitative results on DTU dataset. Our method achieves the best performance among Gaussian Splitting surface reconstruction approaches and demonstrates competitive performance to previous state-of-the-art NeRF-based methods." + }, + { + "type": "image", + "bbox": [ + 0.121, + 0.322, + 0.879, + 0.684 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.145, + 0.688, + 0.851, + 0.702 + ], + "angle": 0, + "content": "Figure 4. DTU surface reconstruction results. Our method achieves more complete reconstructions with finer details." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.712, + 0.308, + 0.728 + ], + "angle": 0, + "content": "4.2. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Our method is implemented based on 2DGS [11]. For Gaussian primitive initialization, we use the point cloud back-projected depth maps from CLMVSNet [31]. Instead of using spherical harmonics to represent the color of the Gaussian primitives, we initialize the color using the RGB values from the image. CLMVSNet uses an FPN for feature extraction, and we utilize the full-resolution feature maps with a channel dimension of 8 from the FPN last layer for feature splatting. The remaining parameters are initialized consistently with those in 2DGS. We set the training iter" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.713, + 0.907, + 0.834 + ], + "angle": 0, + "content": "ation to 7000 and replaced the original 2DGS's adaptive density control with our proposed Gaussian update strategy. Similar to 2DGS [11], we used truncated signed distance fusion (TSDF) to extract the final mesh. During TSDF fusion, the voxel size and truncated threshold are set to 0.004 and 0.02 respectively. We set the hyperparameters of the loss function, \\(\\lambda_{1},\\lambda_{2},\\lambda_{3}\\), and \\(\\lambda_{4}\\), to 1000, 0.05, 1, and 0.2, respectively." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.845, + 0.748, + 0.861 + ], + "angle": 0, + "content": "4.3. Benchmark Comparisons" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.907, + 0.901 + ], + "angle": 0, + "content": "We compare the sparse-view reconstruction performance of different NeRF-based and Gaussian-based methods on" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11312" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.107, + 0.094, + 0.47, + 0.404 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.415, + 0.483, + 0.443 + ], + "angle": 0, + "content": "Figure 5. BlendedMVS reconstruction results. Our method achieves more complete and detailed reconstructions." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.449, + 0.481, + 0.574 + ], + "angle": 0, + "content": "
2DGSMVS Init.Fixed ColorFeat. Splat.DGPRSGUCD ↓
(a)✓2.813
(b)✓✓1.273
(c)✓✓✓1.212
(d)✓✓✓1.224
(e)✓✓✓✓1.201
(f)✓✓✓✓✓1.142
(g)✓✓✓✓✓1.152
(h)✓✓✓✓✓✓1.125
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.585, + 0.483, + 0.642 + ], + "angle": 0, + "content": "Table 2. Ablation on DTU [1] dataset. DGPR and SGU represent proposed Direct Gaussian Primitive Regularization and Selective Gaussian Update respectively. Each module improves upon the baseline, leading to the best final results." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.659, + 0.483, + 0.901 + ], + "angle": 0, + "content": "the DTU dataset. As shown in Tab. 1, our method significantly improves upon Gaussian-based approaches under sparse views, outperforming baseline (2DGS) and previous Gaussian-based SOTA methods (PGSR) by \\(67\\%\\) and \\(46\\%\\), respectively. Our method also outperforms SparseNeus [20], the first sparse-view NeRF-based method with notable margins. Our method performs competitively with the state-of-the-art NeRF-based surface reconstruction methods. The visualization results are shown in Fig. 4. Our method captures more surface details than implicit reconstruction approaches and outperforms Gaussian-based methods in preserving the complete scene structure. In comparison, Gaussian Splitting methods, i.e., 2DGS, GOF, and PGSR, tend to exhibit noise and missing areas. Additionally, we performed a qualitative comparison on the BlendedMVS dataset. As shown in Fig. 5, our method" + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.089, + 0.905, + 0.15 + ], + "angle": 0, + "content": "
Methods2DGS [11]PGSR [4]CLMVS. [31]Ours
3 view SfM Point Cloud2.812.08--
49/64 views SfM Point Cloud1.651.53--
CLMVSNet Point Cloud1.271.381.261.13
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.156, + 0.907, + 0.241 + ], + "angle": 0, + "content": "Table 3. Comparison of initialization strategies. Initializing with MVS point cloud improves reconstruction performance. However, a plain combination (Chamfer Distance, CD:1.27) between 2DGS and CLMVSNet fails to surpass CLMVSNet results (CD: 1.26). Our method enhances both MVS and 2DGS through geometry-prioritized optimization." + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.258, + 0.89, + 0.481 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.487, + 0.907, + 0.557 + ], + "angle": 0, + "content": "Figure 6. Comparison between different initialization strategies of 2DGS and our method. Initializing 2DGS with MVS point cloud improves SfM reconstruction completeness but struggles to capture details. Our method not only enhances completeness but also achieves detailed reconstruction." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.566, + 0.907, + 0.595 + ], + "angle": 0, + "content": "achieves higher accuracy in reconstructing meshes for real-world scenes." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.606, + 0.729, + 0.623 + ], + "angle": 0, + "content": "4.4. Ablations and Analysis" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.629, + 0.907, + 0.81 + ], + "angle": 0, + "content": "In this section, we perform ablation studies to analyze the effectiveness of different modules. As shown in Tab. 2, using MVS point cloud initialization notably enhances the performance (row a). By fixing the colors and using feature splatting for geometric enhancement supervision, we further improve the accuracy and completeness of the reconstruction (row e). The Direct Gaussian Primitive Regularization (DGPR) enhances the surface modeling capability of each Gaussian primitive (row f). The Selective Gaussian Update (SGU) leverages rendering cues to enhance the precision of the Gaussian point set, further improving surface accuracy (row g)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Comparison of different initialization strategies. We compare the reconstruction results using various initialization strategies for Gaussian Splatting methods. We show the results using 3 different initialization strategies: SfM point cloud from 3 viewpoints, SfM point cloud from 49 viewpoints, and the CLMVSNet reconstructed point cloud." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11313" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.11, + 0.089, + 0.465, + 0.16 + ], + "angle": 0, + "content": "
SettingsAccuracy↓Completion↓Average↓
SH=30.831.711.27
SH=20.841.681.26
SH=10.841.641.24
Fixed Color0.811.611.21
" + }, + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.162, + 0.483, + 0.191 + ], + "angle": 0, + "content": "Table 4. Ablation study on appearance learning. Fixing the color helps in learning better geometry." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.198, + 0.487, + 0.259 + ], + "angle": 0, + "content": "
MethodsAccuracy↓Completeness↓Average↓
Baseline0.7901.6121.201
Adaptive Density Control [14]0.8141.6541.234
Selective Gaussian Update0.7601.5441.152
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.261, + 0.483, + 0.318 + ], + "angle": 0, + "content": "Table 5. Ablation on DTU [1] dataset. The proposed Selective Gaussian Update strategy enhances reconstruction accuracy and completeness compared to the generic Adaptive Density Control in sparse views." + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.326, + 0.47, + 0.42 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.423, + 0.483, + 0.465 + ], + "angle": 0, + "content": "Figure 7. Visualization of MVS feature splatting. Incorporating MVS feature splatting aids in achieving higher reconstruction accuracy in challenging regions." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.476, + 0.483, + 0.627 + ], + "angle": 0, + "content": "Using MVS point clouds achieves a more complete reconstruction than SfM. Note that as shown in Tab. 3, due to the inherent ill-posedness in sparse-view reconstruction, the plain combination of MVS and 2DGS reconstruction yields suboptimal performance. Even with CLMVSNet point clouds as initialization, the reconstruction quality still falls short of that achieved by CLMVSNet itself. In contrast, our geometric-prioritized enhancement addresses this challenge, with results excelling both CLMVSNet and 2DGS with higher completeness and finer details (Fig. 6)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.628, + 0.483, + 0.885 + ], + "angle": 0, + "content": "Geometrically enhanced supervision. In this paper, we mitigate appearance overfitting by fixing the color and feature parameters during optimization. In the original Gaussian splatting, view-independent color is represented by spherical harmonics (SH) with learnable coefficients. Intuitively, reducing SH degrees can achieve a similar goal in alleviating color overfitting. We compare the fixed color and reduced SH degrees in Tab. 4. Results show that reducing the SH degrees also mitigates the degradation in geometry, with only 1 SH degree outperforming other settings (i.e., 2, 3) in surface reconstruction. However, using learnable color coefficients still performs worse than the adopted choice of fixed color during training. Moreover, we enhance geometric supervision using MVS feature splatting. As shown in Fig. 7, MVS features are more expressive and provide more effective constraints, especially in challenging regions such as repetitive texture areas (e.g., the roof of the building)." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.886, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Ablation on Direct Gaussian Primitive Regularization." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.089, + 0.906, + 0.162 + ], + "angle": 0, + "content": "
SettingsAccuracy↓Completion↓Average↓
K=90.7461.5441.145
K=250.7471.5411.144
K=490.7511.5321.142
K=810.7471.5311.139
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.163, + 0.905, + 0.205 + ], + "angle": 0, + "content": "Table 6. Comparison of different numbers of sampled points. Increasing the sampling numbers lead to a slight performance improvement." + }, + { + "type": "table", + "bbox": [ + 0.527, + 0.211, + 0.897, + 0.266 + ], + "angle": 0, + "content": "
MethodsSparseNeus [20]NeuSurf [12]Ours
Chamfer Distance1.270.991.13
Training Time25mins10 hours10mins
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.27, + 0.905, + 0.299 + ], + "angle": 0, + "content": "Table 7. Performance and efficiency. Our method achieves comparable performance while requiring the least training time." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.307, + 0.905, + 0.367 + ], + "angle": 0, + "content": "We analyze the impact of the reparameterization-based disk sampling number \\( K \\) on the final results. As shown in Tab. 6, increasing the number of sampled points can slightly enhance the final reconstruction performance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.367, + 0.906, + 0.548 + ], + "angle": 0, + "content": "Selective Gaussian update strategy. To validate the effectiveness of our selective update strategy, we compared it with the adaptive density control in the original 2DGS. Quantitative results are shown in the Tab. 5. Due to the lack of sufficient multi-view supervision, the densified Gaussian primitives are prone to degradation, leading to declined reconstruction performance (CD:1.23) compared to the baseline (CD: 1.20). The proposed selective Gaussian update strategy enhances the accuracy of the original Gaussian point set by utilizing high-precision points from the rendered cues, which further improves the surface reconstruction quality." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.549, + 0.906, + 0.655 + ], + "angle": 0, + "content": "Performance and efficiency. We compared reconstruction performance and training time with previous state-of-the-art surface reconstruction methods. As shown in Tab. 7, our method achieves better performance and \\(2 \\times\\) faster training speed than SparseNeus [20]. Meanwhile, our method achieves competitive performance while being \\(60 \\times\\) faster than NeuSurf [12]." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.669, + 0.634, + 0.685 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.695, + 0.906, + 0.844 + ], + "angle": 0, + "content": "In this paper, we propose Sparse2DGS, addressing the challenge of ill-posed geometry optimization from sparse views through geometry-prioritized enhancement schemes. Our method not only reconstructs accurate and complete scene surfaces but also achieves higher efficiency than NeRF-based methods. Limitation: Despite their effectiveness in dense matching, MVS methods inevitably encounter degradation under occlusions. The error in occluded areas can propagate to our method, leading to potential performance declines." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.845, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Acknowledgements Y. Zhang was supported by NSFC (No.U19B2037) and the Natural Science Basic Research Program of Shaanxi (No.2021JCW-03). Y. Zhu was supported by NSFC (No.61901384)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11314" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.099, + 0.188, + 0.115 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.124, + 0.484, + 0.178 + ], + "angle": 0, + "content": "[1] Henrik Aanaes, Rasmus Ramsbøl Jensen, George Vogiatzis, Engin Tola, and Anders Bjorholm Dahl. Large-scale data for multiple-view stereopsis. International Journal of Computer Vision, 120:153-168, 2016. 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.18, + 0.484, + 0.234 + ], + "angle": 0, + "content": "[2] Chenjie Cao, Xinlin Ren, and Yanwei Fu. Mvsformer: Learning robust image representations via transformers and temperature-based depth for multi-view stereo. arXiv preprint arXiv:2208.02541, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.236, + 0.484, + 0.304 + ], + "angle": 0, + "content": "[3] Di Chang, Aljaž Božić, Tong Zhang, Qingsong Yan, Yingcong Chen, Sabine Susstrunk, and Matthias Nießner. Rc-mvsnet: Unsupervised multi-view stereo with neural rendering. In European conference on computer vision, pages 665-680. Springer, 2022. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.305, + 0.483, + 0.374 + ], + "angle": 0, + "content": "[4] Danpeng Chen, Hai Li, Weicai Ye, Yifan Wang, Weijian Xie, Shangjin Zhai, Nan Wang, Haomin Liu, Hujun Bao, and Guofeng Zhang. Pgsr: Planar-based gaussian splatting for efficient and high-fidelity surface reconstruction. arXiv preprint arXiv:2406.06521, 2024. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.375, + 0.483, + 0.415 + ], + "angle": 0, + "content": "[5] Hanlin Chen, Chen Li, and Gim Hee Lee. Neusg: Neural implicit surface reconstruction with 3d gaussian splatting guidance. arXiv preprint arXiv:2312.00846, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.416, + 0.483, + 0.497 + ], + "angle": 0, + "content": "[6] Yikang Ding, Wentao Yuan, Qingtian Zhu, Haotian Zhang, Xiangyue Liu, Yuanjiang Wang, and Xiao Liu. Transmvsnet: Global context-aware multi-view stereo network with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8585-8594, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.499, + 0.483, + 0.553 + ], + "angle": 0, + "content": "[7] Lue Fan, Yuxue Yang, Minxing Li, Hongsheng Li, and Zhaoxiang Zhang. Trim 3d gaussian splatting for accurate geometry representation. arXiv preprint arXiv:2406.07499, 2024.1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.555, + 0.483, + 0.609 + ], + "angle": 0, + "content": "[8] Qiancheng Fu, Qingshan Xu, Yew Soon Ong, and Wenbing Tao. Geo-neus: Geometry-consistent neural implicit surfaces learning for multi-view reconstruction. Advances in Neural Information Processing Systems, 35:3403-3416, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.61, + 0.483, + 0.664 + ], + "angle": 0, + "content": "[9] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In 2012 IEEE conference on computer vision and pattern recognition, pages 3354-3361. IEEE, 2012. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.665, + 0.484, + 0.734 + ], + "angle": 0, + "content": "[10] Antoine Guédon and Vincent Lepetit. Sugar: Surface-aligned gaussian splatting for efficient 3d mesh reconstruction and high-quality mesh rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5354-5363, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.735, + 0.483, + 0.79 + ], + "angle": 0, + "content": "[11] Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2d gaussian splatting for geometrically accurate radiance fields. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 1, 2, 3, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.79, + 0.483, + 0.858 + ], + "angle": 0, + "content": "[12] Han Huang, Yulun Wu, Junsheng Zhou, Ge Gao, Ming Gu, and Yu-Shen Liu. Neusurf: On-surface priors for neural surface reconstruction from sparse input views. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2312-2320, 2024. 2, 3, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.859, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[13] Michael Kazhdan, Matthew Bolitho, and Hugues Hoppe. Poisson surface reconstruction. In Proceedings of the fourth Eurographics symposium on Geometry processing, 2006. 2" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.124, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[14] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4):139-1, 2023. 1, 2, 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.149, + 0.905, + 0.216 + ], + "angle": 0, + "content": "[15] Tejas Khot, Shubham Agrawal, Shubham Tulsiani, Christoph Mertz, Simon Lucey, and Martial Hebert. Learning unsupervised multi-view stereopsis via robust photometric consistency. arXiv preprint arXiv:1905.02706, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.218, + 0.907, + 0.301 + ], + "angle": 0, + "content": "[16] Rui Li, Dong Gong, Wei Yin, Hao Chen, Yu Zhu, Kaixuan Wang, Xiaozhi Chen, Jinqiu Sun, and Yanning Zhang. Learning to fuse monocular and multi-view cues for multiframe depth estimation in dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21539-21548, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.302, + 0.905, + 0.383 + ], + "angle": 0, + "content": "[17] Rui Li, Tobias Fischer, Mattia Segu, Marc Pollefeys, Luc Van Gool, and Federico Tombari. Know your neighbors: Improving single-view reconstruction via spatial vision-language reasoning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9848-9858, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.385, + 0.905, + 0.455 + ], + "angle": 0, + "content": "[18] Yinhao Li, Zheng Ge, Guanyi Yu, Jinrong Yang, Zengran Wang, Yukang Shi, Jianjian Sun, and Zeming Li. Bevdepth: Acquisition of reliable depth for multi-view 3d object detection. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1477-1485, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.456, + 0.905, + 0.51 + ], + "angle": 0, + "content": "[19] Yixun Liang, Hao He, and Yingcong Chen. Retr: Modeling rendering via transformer for generalizable neural surface reconstruction. Advances in Neural Information Processing Systems, 36, 2024. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.511, + 0.905, + 0.579 + ], + "angle": 0, + "content": "[20] Xiaoxiao Long, Cheng Lin, Peng Wang, Taku Komura, and Wenping Wang. Sparseneus: Fast generalizable neural surface reconstruction from sparse views. In European Conference on Computer Vision, pages 210-227. Springer, 2022. 1, 2, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.581, + 0.905, + 0.622 + ], + "angle": 0, + "content": "[21] Xuan Luo, Jia-Bin Huang, Richard Szeliski, Kevin Matzen, and Johannes Kopf. Consistent video depth estimation. ACM Transactions on Graphics (ToG), 39(4):71-1, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.623, + 0.905, + 0.691 + ], + "angle": 0, + "content": "[22] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.692, + 0.905, + 0.761 + ], + "angle": 0, + "content": "[23] Michael Oechsle, Songyou Peng, and Andreas Geiger. Unisurf: Unifying neural implicit surfaces and radiance fields for multi-view reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5589-5599, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.762, + 0.905, + 0.815 + ], + "angle": 0, + "content": "[24] Ri-Zhao Qiu, Ge Yang, Weijia Zeng, and Xiaolong Wang. Feature splatting: Language-driven physics-based scene synthesis and editing. arXiv preprint arXiv:2404.01223, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.817, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[25] Yufan Ren, Fangjinhua Wang, Tong Zhang, Marc Pollefeys, and Sabine Susstrunk. Volrecon: Volume rendering of signed ray distance functions for generalizable multi-view reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16685-16695, 2023. 1, 2, 3, 6" + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.093, + 0.907, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11315" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[26] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4104-4113, 2016. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.217 + ], + "angle": 0, + "content": "[27] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5459-5469, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.218, + 0.482, + 0.273 + ], + "angle": 0, + "content": "[28] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. arXiv preprint arXiv:2106.10689, 2021. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.274, + 0.482, + 0.302 + ], + "angle": 0, + "content": "[29] Yiqun Wang, Ivan Skorokhodov, and Peter Wonka. Improved surface reconstruction using high-frequency details. 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.303, + 0.482, + 0.371 + ], + "angle": 0, + "content": "[30] Jiang Wu, Rui Li, Haofei Xu, Wenxun Zhao, Yu Zhu, Jinqiu Sun, and Yanning Zhang. Gomvs: Geometrically consistent cost aggregation for multi-view stereo. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20207-20216, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.372, + 0.482, + 0.453 + ], + "angle": 0, + "content": "[31] Kaiqiang Xiong, Rui Peng, Zhe Zhang, Tianxing Feng, Jianbo Jiao, Feng Gao, and Ronggang Wang. Cl-mvsnet: Unsupervised multi-view stereo with dual-level contrastive learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3769-3780, 2023. 1, 2, 3, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.455, + 0.482, + 0.524 + ], + "angle": 0, + "content": "[32] Jialei Xu, Xianming Liu, Yuanchao Bai, Junjun Jiang, Kaixuan Wang, Xiaozhi Chen, and Xiangyang Ji. Multi-camera collaborative depth prediction via consistent structure estimation. In Proceedings of the 30th ACM international conference on multimedia, pages 2730-2738, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.525, + 0.482, + 0.593 + ], + "angle": 0, + "content": "[33] Jialei Xu, Wei Yin, Dong Gong, Junjun Jiang, and Xianming Liu. Sdge: Stereo guided depth estimation for \\(360^{\\circ}\\) camera sets. In 2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 11179-11186. IEEE, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.595, + 0.482, + 0.677 + ], + "angle": 0, + "content": "[34] Luoyuan Xu, Tao Guan, Yuesong Wang, Wenkai Liu, Zhaojie Zeng, Junle Wang, and Wei Yang. C2f2neus: Cascade cost frustum fusion for high fidelity and generalizable neural surface reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 18291-18301, 2023. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.679, + 0.482, + 0.746 + ], + "angle": 0, + "content": "[35] Qingshan Xu, Wanjuan Su, Yuhang Qi, Wenbing Tao, and Marc Pollefeys. Learning inverse depth regression for pixelwise visibility-aware multi-view stereo networks. International Journal of Computer Vision, 130(8):2040-2059, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.748, + 0.482, + 0.803 + ], + "angle": 0, + "content": "[36] Yao Yao, Zixin Luo, Shiwei Li, Tian Fang, and Long Quan. Mvsnet: Depth inference for unstructured multi-view stereo. In Proceedings of the European conference on computer vision (ECCV), pages 767-783, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.804, + 0.482, + 0.872 + ], + "angle": 0, + "content": "[37] Yao Yao, Zixin Luo, Shiwei Li, Jingyang Zhang, Yufan Ren, Lei Zhou, Tian Fang, and Long Quan. Blendedmvs: A large-scale dataset for generalized multi-view stereo networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1790-1799, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[38] Lior Yariv, Jiatao Gu, Yoni Kasten, and Yaron Lipman. Volume rendering of neural implicit surfaces. Advances in Neu" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "ral Information Processing Systems, 34:4805-4815, 2021. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.162 + ], + "angle": 0, + "content": "[39] Jae-Chern Yoo and Tae Hee Han. Fast normalized cross-correlation. Circuits, systems and signal processing, 28:819-843, 2009. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.164, + 0.905, + 0.22 + ], + "angle": 0, + "content": "[40] Mae Younes, Amine Ouasfi, and Adnane Boukhayma. Sparsecraft: Few-shot neural reconstruction through stereopsis guided geometric linearization. In European Conference on Computer Vision, pages 37-56. Springer, 2024. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.222, + 0.905, + 0.275 + ], + "angle": 0, + "content": "[41] Mulin Yu, Tao Lu, Linning Xu, Lihan Jiang, Yuanbo Xiangli, and Bo Dai. Gsdf: 3dgs meets sdf for improved rendering and reconstruction. arXiv preprint arXiv:2403.16964, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.278, + 0.905, + 0.346 + ], + "angle": 0, + "content": "[42] Zehao Yu, Songyou Peng, Michael Niemeyer, Torsten Sattler, and Andreas Geiger. Monosdf: Exploring monocular geometric cues for neural implicit surface reconstruction. Advances in neural information processing systems, 35:25018-25032, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.348, + 0.905, + 0.403 + ], + "angle": 0, + "content": "[43] Zehao Yu, Torsten Sattler, and Andreas Geiger. Gaussian opacity fields: Efficient and compact surface reconstruction in unbounded scenes. arXiv preprint arXiv:2404.10772, 2024. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.405, + 0.905, + 0.459 + ], + "angle": 0, + "content": "[44] Baowen Zhang, Chuan Fang, Rakesh Shrestha, Yixun Liang, Xiaoxiao Long, and Ping Tan. Rade-gs: Rasterizing depth in gaussian splatting. arXiv preprint arXiv:2406.01467, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.461, + 0.905, + 0.517 + ], + "angle": 0, + "content": "[45] Jingyang Zhang, Yao Yao, and Long Quan. Learning signed distance field for multi-view surface reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6525-6534, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.518, + 0.905, + 0.572 + ], + "angle": 0, + "content": "[46] Jingyang Zhang, Shiwei Li, Zixin Luo, Tian Fang, and Yao Yao. Vis-mvsnet: Visibility-aware multi-view stereo network. International Journal of Computer Vision, 131(1): 199-214, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.574, + 0.905, + 0.658 + ], + "angle": 0, + "content": "[47] Shijie Zhou, Haoran Chang, Sicheng Jiang, Zhiwen Fan, Zehao Zhu, Dejia Xu, Pradyumna Chari, Suya You, Zhangyang Wang, and Achuta Kadambi. Feature 3dgs: Supercharging 3d gaussian splattering to enable distilled feature fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21676-21685, 2024. 4" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.658 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11316" + } + ] +] \ No newline at end of file diff --git a/2025/Sparse2DGS_ Geometry-Prioritized Gaussian Splatting for Surface Reconstruction from Sparse Views/3c5dfe86-080c-4ec1-9ca9-65bf2b3453a3_origin.pdf b/2025/Sparse2DGS_ Geometry-Prioritized Gaussian Splatting for Surface Reconstruction from Sparse Views/3c5dfe86-080c-4ec1-9ca9-65bf2b3453a3_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..6ac678660e989cdea822af9f40397b29e5641d00 --- /dev/null +++ b/2025/Sparse2DGS_ Geometry-Prioritized Gaussian Splatting for Surface Reconstruction from Sparse Views/3c5dfe86-080c-4ec1-9ca9-65bf2b3453a3_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:085861d13b3734d7642b4dd78240a4d3745bc8737353b1a53b5b749aba6adb41 +size 9315534 diff --git a/2025/Sparse2DGS_ Geometry-Prioritized Gaussian Splatting for Surface Reconstruction from Sparse Views/full.md b/2025/Sparse2DGS_ Geometry-Prioritized Gaussian Splatting for Surface Reconstruction from Sparse Views/full.md new file mode 100644 index 0000000000000000000000000000000000000000..a1dafbe44c6adce1c7158f75666af83572c2fae3 --- /dev/null +++ b/2025/Sparse2DGS_ Geometry-Prioritized Gaussian Splatting for Surface Reconstruction from Sparse Views/full.md @@ -0,0 +1,315 @@ +# Sparse2DGS: Geometry-Prioritized Gaussian Splating for Surface Reconstruction from Sparse Views + +Jiang Wu Rui Li Yu Zhu† Rong Guo Jinqiu Sun Yanning Zhang† Northwestern Polytechnical University + +![](images/5de91afae79a2e02772b387629ac4a9a80ef02c5eb2ec8b0f33fbbba5c40b093.jpg) +Figure 1. Sparse2DGS boosts the strengths of Gaussian Splatting and Multi-view Stereo in sparse-view surface reconstruction, with notable improvement over 2DGS [11], CLMVSNet [31], as well as their plain combination (left). Meanwhile, Sparse2DGS is $2 \times$ faster than NeRF-based sparse-view method [20] (middle) and leads to complete and accurate reconstruction (right). + +# Abstract + +We present a Gaussian Splatting method for surface reconstruction using sparse input views. Previous methods relying on dense views struggle with extremely sparse Structure-from-Motion points for initialization. While learning-based Multi-view Stereo (MVS) provides dense 3D points, directly combining it with Gaussian Splatting leads to suboptimal results due to the ill-posed nature of sparse-view geometric optimization. We propose Sparse2DGS, an MVS-initialized Gaussian Splatting pipeline for complete and accurate reconstruction. Our key insight is to incorporate the geometric-prioritized enhancement schemes, allowing for direct and robust geometric learning under ill-posed conditions. Sparse2DGS outperforms existing methods by notable margins while being $2 \times$ faster than the NeRF-based fine-tuning approach. Code is available at https://github.com/Wuuu3511/Sparse2DGS. + +# 1. Introduction + +Reconstructing 3D surfaces from multiple views is a fundamental task in computer vision and has been widely used in autonomous driving [9, 18] and virtual reality [21]. While + +typical methods require abundant input images, reconstructing accurate 3D surfaces from sparse views is particularly challenging and is crucial for scenes that are frequently captured with limited images. + +NeRF-based surface reconstruction methods [22] have gained popularity by encoding scene geometry into an occupancy field [17, 23] or a signed distance function (SDF) [28, 38]. However, when faced with sparse views, these methods often yield incomplete surfaces with artifacts due to insufficient supervision. Although sparse-view methods [19, 20, 25, 34] use pre-trained feed-forward NeRFs to provide geometric priors, they still exhibit limited generalization to unseen scenes. + +Recently, 3D Gaussian Splatting (3DGS) [14] has leveraged Structure-from-Motion (SfM) point cloud to construct Gaussian primitives for fast and high-quality novel view synthesis (NVS). Subsequent works [4, 5, 7, 11, 41, 43] aim to achieve more efficient 3D surface reconstruction using 3DGS. Notably, 2D Gaussian Splatting (2DGS) [11] proposes a 2D variant of Gaussian primitives, with disk-shaped primitives suitable for geometry representation. However, when dealing with sparse views, existing methods [4, 11, 43] struggle to produce faithful reconstructions due to initially over-sparse SfM points. + +On the other hand, learning-based multi-view stereo (MVS) [6, 16, 30, 36] predicts dense depth maps from + +pixel-wise feature matching, significantly improving geometric completeness even with sparse images. However, MVS depth networks focus on generalizable inference instead of test-time geometry enhancement by design. While combining MVS point cloud with Gaussian Splatting is intuitively effective, their plain combination (see Fig. 1, "2DGS+CLMVS") yields sub-optimal results due to the ill-posed nature of geometry optimization from sparse views. + +In this paper, we propose Sparse2DGS, an MVS-initialized Gaussian Splating method for complete and accurate sparse-view 3D surface reconstruction. The key idea is to boost the strengths of two paradigms through geometric-prioritized enhancement against the ill-posed nature of sparse-view reconstruction. Starting from MVS initialization, we first incorporate MVS-derived geometric features into the Gaussian splattering process while keeping feature and color values fixed. This provides geometrically expressive supervision and prevents appearance overfitting, thereby improving geometric reconstruction quality. We further introduce a direct Gaussian primitive regularization, with a reparameterization-based disk sampling strategy that reformulates Gaussian primitives' position, orientation, and scale into point representations. This approach enables a unified optimization of different forms of Gaussian properties within a standard point-based, cross-view consistency loss. Moreover, instead of densifying Gaussian primitives under ill-posed conditions, we propose to selectively update the MVS-initialized Gaussian primitives with rendered geometric cues, allowing for self-guided geometric corrections. + +Sparse2DGS significantly outperforms existing Gaussian Splatting methods in sparse-view surface reconstruction. Meanwhile, it demonstrates notable efficiency advantages over NeRF-based test-time optimization. Our contributions are as follows: + +- We propose Sparse2DGS, a test-time sparse-view 3D surface reconstruction method that boosts the strengths of Gaussian Splitting and MVS, allowing for complete and accurate reconstruction from sparse images. +- We focus on geometric-prioritized enhancement by proposing geometrically enhanced supervision, direct Gaussian primitive regularization, and selective Gaussian update throughout the Gaussian Splatting pipeline. Sparse2DGS achieves the lowest error (1.13 CD error) compared to 2DGS (2.81) [11], GOF (2.82) [43], and PGSR (2.08) [4] on DTU dataset with 3 views. Meanwhile, Sparse2DGS is $2 \times$ faster than fine-tuning NeRF-based SparseNeus [20]. + +# 2. Related Works + +# 2.1. 3D Reconstruction via Novel View Synthesis + +Recent progresses [11, 12, 25, 43] in novel view synthesis (NVS) have enabled accurate 3D surface reconstruc + +tion with only image-based supervision. Neural Radiance Fields [22] (NeRFs) have become a significant technique for representing 3D scenes. It has inspired numerous works on directly extracting surfaces by training a parameterized radiance field through differentiable volume rendering. Neus [28] and VolSDF [38] integrate surface learning into volume rendering by re-parameterizing the signed distance function as volume density, reconstructing more accurate surfaces. To reduce geometric deviations and accelerate training, subsequent works [8, 27, 29, 42] have introduced improvements in various areas. + +Recently, 3D Gaussian Splatting [14] has shown impressive results in novel view synthesis (NVS), with many recent studies adopting this approach for surface reconstruction. SuGaR [10] aligns 3D Gaussians with scene surfaces by designing a regularization term and extracts an accurate mesh from the Gaussians using Poisson Reconstruction [13]. 2DGS [11] replaces 3D Gaussian ellipsoids with 2D-oriented planar disks, providing view-consistent geometric representations. RaDe-GS [44] further designs a rasterized approach to render depth and normal, enhancing geometric accuracy while ensuring rendering efficiency. GOF [43] extracts the level set from trained Gaussian primitives by constructing a Gaussian opacity field. PGSR [4] improves mesh quality by introducing single-view and multi-view geometric constraints. Additionally, GSDF [41] and NeusSG [5] reconstruct scenes by combining the advantages of 3D Gaussian Splatting and SDF networks. Despite these methods' success, the extracted geometry tends to be noisy and incomplete when the input number of training viewpoints is limited. + +# 2.2. Learning-based Multi-view Stereo + +Deep learning-based Multi-view Stereo (MVS) estimates the reference image's depth from multi-view posed images. MVSNet [36] first introduces the differentiable homography warping to build the cost volume with deep image features and then employs a 3D CNN for cost regularization. Recent follow-up works have introduced improvements in various aspects, including multi-view feature extraction [2, 6], cost aggregation [32, 33, 35, 46], and cost volume regularization [30]. However, these methods require ground-truth depth maps for training, which are challenging to obtain in practice. Unsupervised MVS [3, 15, 31] can be applied in more general, real-world scenarios. MVS methods can infer dense depth maps for each input view and therefore, we select the unsupervised CLMVSNet [31] to initialize the Gaussian point cloud. + +# 2.3. Sparse View Surface Reconstruction + +Recent works have explored 3D surface reconstruction from sparse training views. SparseNeuS [20] achieves reasonable mesh reconstruction results by constructing a cascaded + +![](images/6a87a0e2c57120679f272a77cf9a2d5b3d88bc0b0ba4c0d9b0fe890f1f77df64.jpg) +Figure 2. Overview. Given sparse posed images, we first leverage the MVS points to initialize the Gaussian position $\mathbf{p}$ (Sec. 3.1). We then leverage the MVS-derived feature for Gaussian Splating, with fixed feature $\mathbf{f}$ and color $\mathbf{c}$ values to conduct geometrically expressive supervision and avoid appearance overfitting (Sec. 3.2). We then optimize Gaussian primitive properties by reformulating orientation $\mathbf{R}$ , scale $\mathbf{S}$ , and position $\mathbf{p}$ into sampled points using the reparameterization-based disk sampling, ensuring simultaneously optimizing different forms of Gaussian properties with point-based loss (Sec. 3.3). Finally, we introduce a Selective Gaussian Update strategy that leverages rendered geometry as an alternative to update Gaussian primitive positions, which further improves reconstruction quality (Sec. 3.4). + +volume to learn a generalizable radiance field, followed by per-scene fine-tuning. NeuSurf [12] learns dense on-surface priors from sparse points to guide geometric learning. SparseCraft [40] leverages stereopsis cues for geometric regularization based on Taylor expansion. Moreover, VolRecon [25] and ReTR [19] enhance voxel features by designing specialized transformer architectures. C2F2Neus [34] constructs per-view cost frustum for finer geometry estimation. Although these feedforward networks can directly infer scene geometry, their limited generalization to unseen scenes restricts their applicability in real-world scenarios. + +# 3. Methodology + +Given sparse view posed images $\{\mathbf{I}_i\}_{i=1}^N$ , where $\mathbf{I}_i \in \mathbb{R}^{H \times W \times 3}$ and $N$ denotes the view number, Sparse2DGS aims to reconstruct accurate 3D scene surfaces using Gaussian Splatting. As shown in Fig. 2, we first initialize the Gaussian primitives with the MVS [31] reconstructed point cloud. Then, we sample MVS geometric feature maps into feature splatting with its value and RGB color fixed for supervision. For each Gaussian primitive, we regularize its geometric properties (position, orientation, scale) by a novel reparameterization-based 2D disk sampling strategy that formulates these properties as point-based supervision. Finally, we propose a selective Gaussian update strategy to further enhance Gaussian primitives using rendered geometric cues. + +# 3.1. 2D Gaussian Splitting with MVS initialization + +2D Gaussian Splitting [11] (2DGS) represents Gaussian primitives as planar disks suitable for representing and op- + +timizing 3D surfaces. However, it relies on sparse SfM [26] 3D points for initialization, which exhibits limited reconstruction quality under sparse-view inputs. To address this, we initialize 2DGS with an MVS point cloud to provide dense geometric representations. + +2D Gaussian Splatting. Each 2D Gaussian primitive is characterized by its central position $\mathbf{p}$ , two principal tangential vectors $\mathbf{t_u}$ and $\mathbf{t_v}$ , and a scaling vector $(s_u, s_v)$ . The rotation matrix of a 2D Gaussian can be represented as $\mathbf{R} = (\mathbf{t_u}, \mathbf{t_v}, \mathbf{t_z})$ , where $\mathbf{t_z}$ is defined by two orthogonal tangent vectors $\mathbf{t_z} = \mathbf{t_u} \times \mathbf{t_v}$ and the scale matrix is $\mathbf{S} = \mathrm{diag}(s_u, s_v, 0)$ . A 2D Gaussian is then defined in a local tangent plane in world space, which is parameterized: + +$$ +P (u, v) = \mathbf {p} + s _ {u} \mathbf {t} _ {u} u + s _ {v} \mathbf {t} _ {v} v. \tag {1} +$$ + +For the point $\mathbf{u} = (u,v)$ in $uv$ space, its 2D Gaussian value can then be evaluated by standard Gaussian + +$$ +\mathcal {G} (\mathbf {u}) = \exp \frac {- \left(u ^ {2} + v ^ {2}\right)}{2}. \tag {2} +$$ + +Initializing 2DGS with MVS depth. MVS methods estimate dense depth maps through cross-view matching. In particular, unsupervised MVS [31] reconstructs complete scenes without ground truth for training, offering greater flexibility. Therefore, we use the point cloud reconstructed by the current unsupervised method CLMVSNet [31] to initialize our Gaussian point cloud. Specifically, given sparse views' images $\{\mathbf{I}_i\}_{i=1}^N$ , we use the MVS network for depth maps $\{\mathbf{D}_i\}_{i=1}^N$ , then back-project these depth maps and fuse into a point cloud $\{\mathbf{p}_{ij} \mid 1 \leq i \leq N, 1 \leq j \leq HW\}$ as ini- + +tialized positions for Gaussian primitives, where $i$ denotes image index and $j$ is the pixel index within each image. + +# 3.2. Geometrically Enhanced Supervision + +3D/2D Gaussian Splatting [11, 14] couples color and geometry optimization through image-based loss. However, this leads to severe degradation under sparse views as color will overfit sparse views, leading to a compromised geometry. Meanwhile, the challenges in image areas (e.g., textureless regions) will be amplified with sparse views. To address this issue, we propose geometrically enhanced supervision by incorporating off-the-shelf expressive MVS features for feature splatting and supervision, with the color and feature fixed to avoid overfitting. + +In addition to providing dense point clouds from input images $\{\mathbf{I}_i\}_{i = 1}^N$ , the MVS network [31] offers off-the-shelf feature maps with geometric expressiveness. We leverage MVS features to enhance geometric supervision by incorporating additional feature attribute $\mathbf{f}_{ij}$ into Gaussian primitives. Specifically, for each input image $\mathbf{I}_i$ , we extract its feature map $\mathbf{F}_i$ from the Feature Pyramid Network (FPN) of CLMVSNet [31]. Since each 3D point corresponds to an image pixel due to MVS initialization, we sample both RGB color and MVS features into each Gaussian primitive. + +$$ +\mathbf {f} _ {i j} = \mathbf {F} _ {i} (\pi (\mathbf {p} _ {i j})), \mathbf {c} _ {i j} = \mathbf {I} _ {i} (\pi (\mathbf {p} _ {i j})), \tag {3} +$$ + +where $\pi(\cdot)$ denotes the projection from 3D space onto 2D image plane. We further render the feature into an image plane using the tile-based renderer together with RGB rendering. We show the rendering process given $M$ Gaussian primitives along a ray, with indices $i, j$ omitted for brevity + +$$ +\{\hat {\mathbf {c}}, \hat {\mathbf {f}} \} = \sum_ {m = 1} ^ {M} \left\{\mathbf {c} _ {m}, \mathbf {f} _ {m} \right\} \cdot \alpha_ {m} \prod_ {o = 1} ^ {m - 1} (1 - \alpha_ {m}), \tag {4} +$$ + +where $\alpha_{m}$ is the opacity of the Gaussian. With rendered feature maps that have enhanced geometric expressiveness for the challenging sparse views, we further supervise the rendered features using cosine similarity loss by + +$$ +\mathcal {L} _ {f e a} = 1 - \frac {\mathbf {f} _ {i j} \cdot \hat {\mathbf {f}} _ {i j}}{\| \mathbf {f} _ {i j} \| \| \hat {\mathbf {f}} _ {i j} \|}, \tag {5} +$$ + +Note that the simultaneous learning of appearance (i.e., color $\mathbf{c}_{ij}$ and feature $\mathbf{f}_{ij}$ ) and geometry lead to geometric degradation under sparse views (see Tab. 4). To address this issue, both $\mathbf{f}_{ij}$ and $\mathbf{c}_{ij}$ are fixed during the optimization to enforce the learning specifically for geometry. + +Our method essentially differentiates previous feature-based methods [24, 47] in its specific strategy for geometric enhancement. In our method, the expressive MVS features are incorporated, and features of each Gaussian are fixed, + +![](images/a9ca6fc52cdf00f37bf78b84109e47bd721d5cd4a5c260b55463f271c3ff79df.jpg) +Figure 3. Direct Gaussian Primitive Regularization. We sample point set $\mathbf{Z}' \sim \mathcal{N}(\mathbf{0},\mathbf{I})$ from the standard Gaussian distribution and leverage the reparameterization technique to transform $\mathbf{Z}'$ to the point set $\mathbf{Z}$ on each Gaussian primitive. This process allows for representing Gaussian orientation $\mathbf{R}$ and scale $\mathbf{S}$ properties in the form of points easier for geometric supervision. The transformed points $\mathbf{Z}$ are then projected to the source and target views, supervised by cross-view feature consistency loss. + +both for geometry enhancement with sparse views. Previous methods leverage splatted features only for segmentation or editing [24, 47] with dense views, while their feature losses do not optimize the scene geometry. + +# 3.3. Direct Gaussian Primitive Regularization + +The disk-shaped 2D Gaussian primitives encode latent scene geometry including local shape orientation $\mathbf{R}$ , scale $\mathbf{S}$ , and position $\mathbf{p}$ , which are essential for accurate surface reconstruction. These geometric properties are typically learned from surrogate image render loss supported by abundant multi-view supervision. When facing sparse-view inputs, the primitives tend to degrade due to insufficient geometric constraints. + +To address this issue, we incorporate direct geometric regularization for 2D Gaussian primitives through multiview consistency guidance. As multi-view consistency approaches only consider point position, we incorporate broad Gaussian primitive properties (i.e., orientation, scale, and position) by correlating their relationship with point position, through a novel reparameterization-based disk sampling strategy. This regularization aligns 2D Gaussian primitives more closely with the latent scene surfaces, leading to improved geometric learning. In the following illustration, we omit the Gaussian primitive indices $i,j$ for brevity. + +Regularizing points with cross-view consistency. Given 3D point sampling positions $\mathbf{Z} = \{\mathbf{z}_k\}_{k=1}^K$ on each Gaussian, we project them to source and target views to obtain the corresponding 2D positions $\mathbf{q}_k^{src}$ and $\mathbf{q}_k^{tar}$ , as well as queried MVS features $\mathbf{F}_{src}(\mathbf{q}_k^{src})$ , $\mathbf{F}_{tar}(\mathbf{q}_k^{tar})$ . We can then optimize the 3D positions by cross-view feature consistency + +loss [45] $\mathcal{L}_{df}$ + +$$ +\mathcal {L} _ {d f} = \frac {1}{K} \sum_ {k = 1} ^ {K} (1 - \frac {\mathbf {F} _ {s r c} \left(\mathbf {q} _ {k} ^ {s r c}\right) \cdot \mathbf {F} _ {t a r} \left(\mathbf {q} _ {k} ^ {t a r}\right)}{\| \mathbf {F} _ {s r c} \left(\mathbf {q} _ {k} ^ {s r c}\right) \| \| \mathbf {F} _ {t a r} \left(\mathbf {q} _ {k} ^ {t a r}\right) \|}), \tag {6} +$$ + +where $K$ denotes the number of sampled points on each disk-shaped Gaussian primitive. + +Reparameterization-based disk sampling. The above operation focuses on position regularization only, and thus, it cannot propagate gradients to other Gaussian primitive properties such as orientation $\mathbf{R}$ and scale $\mathbf{S}$ . We propose a reparameterization-based sampling strategy to solve this problem. As shown in Fig. 3, for each Gaussian primitive, we first sample points $\mathbf{Z}'$ from the standard Gaussian distribution $\mathcal{N}(\mathbf{0},\mathbf{I})$ and then transform the distribution to the target Gaussian primitive through its position $\mathbf{p}$ , orientation $\mathbf{R}$ and scale $\mathbf{S}$ + +$$ +\mathbf {Z} = \mathbf {p} + \mathbf {L} \mathbf {Z} ^ {\prime}, \tag {7} +$$ + +$$ +\mathbf {L} = \mathbf {R S}, \tag {8} +$$ + +where $\mathbf{Z}' \sim \mathcal{N}(\mathbf{0},\mathbf{I})$ . This bridges the gap between Gaussian properties (orientation and scale) and point representation, ensuring regularizing Gaussian geometric properties merely through the surrogate point-based loss. Additionally, since the last scale value of the 2D Gaussian primitive [11] is zero, resulting in difficulties in optimizing the normal vector $\mathbf{R}[\dots,2]$ in the last dimension, we apply additional supervision using the rendered normal $\mathbf{n}$ : + +$$ +\mathcal {L} _ {d n} = 1 - \mathbf {R} [ \dots , 2 ] ^ {\top} \mathbf {n}, \tag {9} +$$ + +The final primitive regularization loss can be defined as: + +$$ +\mathcal {L} _ {d r} = \mathcal {L} _ {d f} + \mathcal {L} _ {d n}. \tag {10} +$$ + +# 3.4. Selective Gaussian Update + +Gaussian Splitting leverages adaptive density control to densify the Gaussian primitives. However, with sparse views, the densified Gaussian primitives are prone to degrade without sufficient multi-view supervision. To address this issue, as shown in Fig. 2, instead of continually densifying the MVS-initialized dense Gaussian primitives with limited supervision, we propose to selectively update them guided by the alternative rendered geometric cues, e.g., depth and normal. + +Evaluating Gaussian primitives via patch warping. Given any Gaussian primitive with its position $\mathbf{p}$ and orientation $\mathbf{R}$ , we can warp an image patch from one source view to a target view using its position $\mathbf{p}$ and orientation $\mathbf{R}$ together with the pose, following homography warping techniques [3]. Then, we can evaluate the geometric quality of the Gaussian primitive via the patch-based Normalized Cross-Correlation (NCC) [39] metric. + +$$ +\operatorname {N C C} (x, y) = \frac {\sum_ {i = l} ^ {n} \left(x _ {l} - \bar {x}\right) \left(y _ {l} - \bar {y}\right)}{\sqrt {\sum_ {l = 1} ^ {n} \left(x _ {l} - \bar {x}\right) ^ {2}} \sqrt {\sum_ {l = 1} ^ {n} \left(y _ {l} - \bar {y}\right) ^ {2}}}, \tag {11} +$$ + +where $x_{l}$ and $y_{l}$ represent the pixels in the patches $x$ and $y$ of the source and target images, respectively. + +Update Gaussian primitive with rendered cues. In addition to using the Gaussian primitive parameter $\mathbf{p}$ and orientation $\mathbf{R}$ for patch warping, we can leverage the rendered pixel-wise depth and normals from tiled Gaussian primitives as an alternative for patch warping. We compute patch-wise NCC over two warping strategies, yielding $\mathrm{NCC}_G$ for Gaussian primitive and $\mathrm{NCC}_R$ for rendered cues. + +If $\mathrm{NCC}_R$ is better than $\mathrm{NCC}_G$ , we reproject the rendered depth into a 3D point and update the position of the corresponding Gaussian primitive. We update the position of Gaussian primitives every 100 training steps, after which all Gaussian parameters continue to be optimized by gradients. In this way, we improve the initial Gaussian primitive with geometric cues from rendered results, further enhancing the reconstruction quality. + +# 3.5. Training loss + +During training, in addition to using the disk regularization loss $\mathcal{L}_{dr}$ and MVS feature splatting loss $\mathcal{L}_{fea}$ , we retain the original 2DGS [11] image reconstruction loss $\mathcal{L}_{rgb}$ , depth-normal consistency loss $\mathcal{L}_n$ , and depth distortion loss $\mathcal{L}_d$ . The final loss function is defined as: + +$$ +\mathcal {L} = \mathcal {L} _ {r g b} + \lambda_ {1} \mathcal {L} _ {d} + \lambda_ {2} \mathcal {L} _ {n}, + \lambda_ {3} \mathcal {L} _ {d r} + \lambda_ {4} \mathcal {L} _ {f e a}, \tag {12} +$$ + +where $\lambda_1, \lambda_2, \lambda_3, \lambda_4$ denotes the hyperparameters for the weights of each loss term. + +# 4. Experiments + +In this section, we first introduce the implementation details and then provide quantitative and qualitative comparisons with current methods on DTU [1] dataset, as well as qualitative results on BlendedMVS [37] dataset. We further analyze the contribution of each module and compare them with different methods through ablation studies. + +# 4.1. Datasets + +The DTU [1] dataset consists of 128 scenes captured in controlled lab environments using structured light scanners containing 49 or 64 images. Similar to SparseNeuS [20], we select 15 scenes and use views 23, 24, and 33 for 3-view reconstruction, with an image resolution of $576 \times 768$ . We perform the quantitative evaluation on the DTU dataset using the official evaluation code to calculate the Chamfer Distance between the reconstructed mesh and the ground truth point cloud. The BlendedMVS [37] dataset includes several large-scale real-world scenes. We also show the qualitative reconstruction comparisons. + +
Scan243740556365698397105106110114118122Mean
Colmap [26]0.902.891.631.082.181.941.611.302.341.281.101.420.761.171.141.52
CLMVSNet [31]1.172.521.680.891.511.820.991.291.520.780.881.100.641.041.121.26
Neus [28]4.574.493.974.324.631.954.683.834.152.501.526.471.265.576.114.00
Volsdf [38]4.034.216.120.918.241.732.741.825.143.092.084.810.603.512.183.41
VolRecon [25]1.202.591.561.081.431.921.111.481.421.051.191.380.741.231.271.38
SparseNeus [20]1.292.271.570.881.611.861.061.271.421.070.990.870.541.151.181.27
Retr [19]1.052.311.440.981.181.520.881.351.300.871.070.770.591.051.121.17
C2F2Neus [34]1.122.421.400.751.411.770.851.161.260.760.910.600.460.880.921.11
Sparsecraft [40]1.171.741.800.701.191.530.831.051.420.780.800.560.440.770.841.04
NeuSurf [12]0.782.351.550.751.041.680.601.140.980.700.740.490.390.750.860.99
2DGS [11]3.143.792.311.515.342.221.852.793.621.633.054.111.053.352.442.81
GOF [43]3.275.052.772.043.842.612.232.653.301.652.584.441.162.292.392.82
PGSR [4]3.363.282.751.275.151.840.881.793.491.191.861.110.611.091.522.08
Ours1.052.351.380.831.371.450.841.161.430.740.850.840.570.951.011.13
+ +Table 1. Quantitative results on DTU dataset. Our method achieves the best performance among Gaussian Splitting surface reconstruction approaches and demonstrates competitive performance to previous state-of-the-art NeRF-based methods. + +![](images/9ed3e83273f7eec8f44dfc6e88c43c823b51662c3c3d7bb63b0e62738b1a4a80.jpg) +Figure 4. DTU surface reconstruction results. Our method achieves more complete reconstructions with finer details. + +# 4.2. Implementation Details + +Our method is implemented based on 2DGS [11]. For Gaussian primitive initialization, we use the point cloud back-projected depth maps from CLMVSNet [31]. Instead of using spherical harmonics to represent the color of the Gaussian primitives, we initialize the color using the RGB values from the image. CLMVSNet uses an FPN for feature extraction, and we utilize the full-resolution feature maps with a channel dimension of 8 from the FPN last layer for feature splatting. The remaining parameters are initialized consistently with those in 2DGS. We set the training iter + +ation to 7000 and replaced the original 2DGS's adaptive density control with our proposed Gaussian update strategy. Similar to 2DGS [11], we used truncated signed distance fusion (TSDF) to extract the final mesh. During TSDF fusion, the voxel size and truncated threshold are set to 0.004 and 0.02 respectively. We set the hyperparameters of the loss function, $\lambda_{1},\lambda_{2},\lambda_{3}$ , and $\lambda_{4}$ , to 1000, 0.05, 1, and 0.2, respectively. + +# 4.3. Benchmark Comparisons + +We compare the sparse-view reconstruction performance of different NeRF-based and Gaussian-based methods on + +![](images/444bf781e02a8f027249643b14b295cd4a1c74977a9ce4dca0e4f315b875e77a.jpg) +Figure 5. BlendedMVS reconstruction results. Our method achieves more complete and detailed reconstructions. + +
2DGSMVS Init.Fixed ColorFeat. Splat.DGPRSGUCD ↓
(a)✓2.813
(b)✓✓1.273
(c)✓✓✓1.212
(d)✓✓✓1.224
(e)✓✓✓✓1.201
(f)✓✓✓✓✓1.142
(g)✓✓✓✓✓1.152
(h)✓✓✓✓✓✓1.125
+ +the DTU dataset. As shown in Tab. 1, our method significantly improves upon Gaussian-based approaches under sparse views, outperforming baseline (2DGS) and previous Gaussian-based SOTA methods (PGSR) by $67\%$ and $46\%$ , respectively. Our method also outperforms SparseNeus [20], the first sparse-view NeRF-based method with notable margins. Our method performs competitively with the state-of-the-art NeRF-based surface reconstruction methods. The visualization results are shown in Fig. 4. Our method captures more surface details than implicit reconstruction approaches and outperforms Gaussian-based methods in preserving the complete scene structure. In comparison, Gaussian Splitting methods, i.e., 2DGS, GOF, and PGSR, tend to exhibit noise and missing areas. Additionally, we performed a qualitative comparison on the BlendedMVS dataset. As shown in Fig. 5, our method + +Table 2. Ablation on DTU [1] dataset. DGPR and SGU represent proposed Direct Gaussian Primitive Regularization and Selective Gaussian Update respectively. Each module improves upon the baseline, leading to the best final results. + +
Methods2DGS [11]PGSR [4]CLMVS. [31]Ours
3 view SfM Point Cloud2.812.08--
49/64 views SfM Point Cloud1.651.53--
CLMVSNet Point Cloud1.271.381.261.13
+ +Table 3. Comparison of initialization strategies. Initializing with MVS point cloud improves reconstruction performance. However, a plain combination (Chamfer Distance, CD:1.27) between 2DGS and CLMVSNet fails to surpass CLMVSNet results (CD: 1.26). Our method enhances both MVS and 2DGS through geometry-prioritized optimization. + +![](images/2af1b2913544c4ea42f028ee04b1610459dff33d8d7f5579ad35977d2f9b96de.jpg) +Figure 6. Comparison between different initialization strategies of 2DGS and our method. Initializing 2DGS with MVS point cloud improves SfM reconstruction completeness but struggles to capture details. Our method not only enhances completeness but also achieves detailed reconstruction. + +achieves higher accuracy in reconstructing meshes for real-world scenes. + +# 4.4. Ablations and Analysis + +In this section, we perform ablation studies to analyze the effectiveness of different modules. As shown in Tab. 2, using MVS point cloud initialization notably enhances the performance (row a). By fixing the colors and using feature splatting for geometric enhancement supervision, we further improve the accuracy and completeness of the reconstruction (row e). The Direct Gaussian Primitive Regularization (DGPR) enhances the surface modeling capability of each Gaussian primitive (row f). The Selective Gaussian Update (SGU) leverages rendering cues to enhance the precision of the Gaussian point set, further improving surface accuracy (row g). + +Comparison of different initialization strategies. We compare the reconstruction results using various initialization strategies for Gaussian Splatting methods. We show the results using 3 different initialization strategies: SfM point cloud from 3 viewpoints, SfM point cloud from 49 viewpoints, and the CLMVSNet reconstructed point cloud. + +
SettingsAccuracy↓Completion↓Average↓
SH=30.831.711.27
SH=20.841.681.26
SH=10.841.641.24
Fixed Color0.811.611.21
+ +Table 4. Ablation study on appearance learning. Fixing the color helps in learning better geometry. + +
MethodsAccuracy↓Completeness↓Average↓
Baseline0.7901.6121.201
Adaptive Density Control [14]0.8141.6541.234
Selective Gaussian Update0.7601.5441.152
+ +![](images/36f83582a80326fac4c415fb033f74b859b43c5d3033abbba332d76da8f0a613.jpg) +Figure 7. Visualization of MVS feature splatting. Incorporating MVS feature splatting aids in achieving higher reconstruction accuracy in challenging regions. + +Using MVS point clouds achieves a more complete reconstruction than SfM. Note that as shown in Tab. 3, due to the inherent ill-posedness in sparse-view reconstruction, the plain combination of MVS and 2DGS reconstruction yields suboptimal performance. Even with CLMVSNet point clouds as initialization, the reconstruction quality still falls short of that achieved by CLMVSNet itself. In contrast, our geometric-prioritized enhancement addresses this challenge, with results excelling both CLMVSNet and 2DGS with higher completeness and finer details (Fig. 6). + +Geometrically enhanced supervision. In this paper, we mitigate appearance overfitting by fixing the color and feature parameters during optimization. In the original Gaussian splatting, view-independent color is represented by spherical harmonics (SH) with learnable coefficients. Intuitively, reducing SH degrees can achieve a similar goal in alleviating color overfitting. We compare the fixed color and reduced SH degrees in Tab. 4. Results show that reducing the SH degrees also mitigates the degradation in geometry, with only 1 SH degree outperforming other settings (i.e., 2, 3) in surface reconstruction. However, using learnable color coefficients still performs worse than the adopted choice of fixed color during training. Moreover, we enhance geometric supervision using MVS feature splatting. As shown in Fig. 7, MVS features are more expressive and provide more effective constraints, especially in challenging regions such as repetitive texture areas (e.g., the roof of the building). + +Ablation on Direct Gaussian Primitive Regularization. + +Table 5. Ablation on DTU [1] dataset. The proposed Selective Gaussian Update strategy enhances reconstruction accuracy and completeness compared to the generic Adaptive Density Control in sparse views. + +
SettingsAccuracy↓Completion↓Average↓
K=90.7461.5441.145
K=250.7471.5411.144
K=490.7511.5321.142
K=810.7471.5311.139
+ +Table 6. Comparison of different numbers of sampled points. Increasing the sampling numbers lead to a slight performance improvement. + +
MethodsSparseNeus [20]NeuSurf [12]Ours
Chamfer Distance1.270.991.13
Training Time25mins10 hours10mins
+ +Table 7. Performance and efficiency. Our method achieves comparable performance while requiring the least training time. + +We analyze the impact of the reparameterization-based disk sampling number $K$ on the final results. As shown in Tab. 6, increasing the number of sampled points can slightly enhance the final reconstruction performance. + +Selective Gaussian update strategy. To validate the effectiveness of our selective update strategy, we compared it with the adaptive density control in the original 2DGS. Quantitative results are shown in the Tab. 5. Due to the lack of sufficient multi-view supervision, the densified Gaussian primitives are prone to degradation, leading to declined reconstruction performance (CD:1.23) compared to the baseline (CD: 1.20). The proposed selective Gaussian update strategy enhances the accuracy of the original Gaussian point set by utilizing high-precision points from the rendered cues, which further improves the surface reconstruction quality. + +Performance and efficiency. We compared reconstruction performance and training time with previous state-of-the-art surface reconstruction methods. As shown in Tab. 7, our method achieves better performance and $2 \times$ faster training speed than SparseNeus [20]. Meanwhile, our method achieves competitive performance while being $60 \times$ faster than NeuSurf [12]. + +# 5. Conclusion + +In this paper, we propose Sparse2DGS, addressing the challenge of ill-posed geometry optimization from sparse views through geometry-prioritized enhancement schemes. Our method not only reconstructs accurate and complete scene surfaces but also achieves higher efficiency than NeRF-based methods. Limitation: Despite their effectiveness in dense matching, MVS methods inevitably encounter degradation under occlusions. The error in occluded areas can propagate to our method, leading to potential performance declines. + +Acknowledgements Y. Zhang was supported by NSFC (No.U19B2037) and the Natural Science Basic Research Program of Shaanxi (No.2021JCW-03). Y. Zhu was supported by NSFC (No.61901384). + +# References + +[1] Henrik Aanaes, Rasmus Ramsbøl Jensen, George Vogiatzis, Engin Tola, and Anders Bjorholm Dahl. Large-scale data for multiple-view stereopsis. International Journal of Computer Vision, 120:153-168, 2016. 5, 7, 8 +[2] Chenjie Cao, Xinlin Ren, and Yanwei Fu. Mvsformer: Learning robust image representations via transformers and temperature-based depth for multi-view stereo. arXiv preprint arXiv:2208.02541, 2022. 2 +[3] Di Chang, Aljaž Božić, Tong Zhang, Qingsong Yan, Yingcong Chen, Sabine Susstrunk, and Matthias Nießner. Rc-mvsnet: Unsupervised multi-view stereo with neural rendering. In European conference on computer vision, pages 665-680. Springer, 2022. 2, 5 +[4] Danpeng Chen, Hai Li, Weicai Ye, Yifan Wang, Weijian Xie, Shangjin Zhai, Nan Wang, Haomin Liu, Hujun Bao, and Guofeng Zhang. Pgsr: Planar-based gaussian splatting for efficient and high-fidelity surface reconstruction. arXiv preprint arXiv:2406.06521, 2024. 1, 2, 6, 7 +[5] Hanlin Chen, Chen Li, and Gim Hee Lee. Neusg: Neural implicit surface reconstruction with 3d gaussian splatting guidance. arXiv preprint arXiv:2312.00846, 2023. 1, 2 +[6] Yikang Ding, Wentao Yuan, Qingtian Zhu, Haotian Zhang, Xiangyue Liu, Yuanjiang Wang, and Xiao Liu. Transmvsnet: Global context-aware multi-view stereo network with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8585-8594, 2022. 1, 2 +[7] Lue Fan, Yuxue Yang, Minxing Li, Hongsheng Li, and Zhaoxiang Zhang. Trim 3d gaussian splatting for accurate geometry representation. arXiv preprint arXiv:2406.07499, 2024.1 +[8] Qiancheng Fu, Qingshan Xu, Yew Soon Ong, and Wenbing Tao. Geo-neus: Geometry-consistent neural implicit surfaces learning for multi-view reconstruction. Advances in Neural Information Processing Systems, 35:3403-3416, 2022. 2 +[9] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In 2012 IEEE conference on computer vision and pattern recognition, pages 3354-3361. IEEE, 2012. 1 +[10] Antoine Guédon and Vincent Lepetit. Sugar: Surface-aligned gaussian splatting for efficient 3d mesh reconstruction and high-quality mesh rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5354-5363, 2024. 2 +[11] Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2d gaussian splatting for geometrically accurate radiance fields. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 1, 2, 3, 4, 5, 6, 7 +[12] Han Huang, Yulun Wu, Junsheng Zhou, Ge Gao, Ming Gu, and Yu-Shen Liu. Neusurf: On-surface priors for neural surface reconstruction from sparse input views. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2312-2320, 2024. 2, 3, 6, 8 +[13] Michael Kazhdan, Matthew Bolitho, and Hugues Hoppe. Poisson surface reconstruction. In Proceedings of the fourth Eurographics symposium on Geometry processing, 2006. 2 + +[14] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4):139-1, 2023. 1, 2, 4, 8 +[15] Tejas Khot, Shubham Agrawal, Shubham Tulsiani, Christoph Mertz, Simon Lucey, and Martial Hebert. Learning unsupervised multi-view stereopsis via robust photometric consistency. arXiv preprint arXiv:1905.02706, 2019. 2 +[16] Rui Li, Dong Gong, Wei Yin, Hao Chen, Yu Zhu, Kaixuan Wang, Xiaozhi Chen, Jinqiu Sun, and Yanning Zhang. Learning to fuse monocular and multi-view cues for multiframe depth estimation in dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21539-21548, 2023. 1 +[17] Rui Li, Tobias Fischer, Mattia Segu, Marc Pollefeys, Luc Van Gool, and Federico Tombari. Know your neighbors: Improving single-view reconstruction via spatial vision-language reasoning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9848-9858, 2024. 1 +[18] Yinhao Li, Zheng Ge, Guanyi Yu, Jinrong Yang, Zengran Wang, Yukang Shi, Jianjian Sun, and Zeming Li. Bevdepth: Acquisition of reliable depth for multi-view 3d object detection. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1477-1485, 2023. 1 +[19] Yixun Liang, Hao He, and Yingcong Chen. Retr: Modeling rendering via transformer for generalizable neural surface reconstruction. Advances in Neural Information Processing Systems, 36, 2024. 1, 3, 6 +[20] Xiaoxiao Long, Cheng Lin, Peng Wang, Taku Komura, and Wenping Wang. Sparseneus: Fast generalizable neural surface reconstruction from sparse views. In European Conference on Computer Vision, pages 210-227. Springer, 2022. 1, 2, 5, 6, 7, 8 +[21] Xuan Luo, Jia-Bin Huang, Richard Szeliski, Kevin Matzen, and Johannes Kopf. Consistent video depth estimation. ACM Transactions on Graphics (ToG), 39(4):71-1, 2020. 1 +[22] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 1, 2 +[23] Michael Oechsle, Songyou Peng, and Andreas Geiger. Unisurf: Unifying neural implicit surfaces and radiance fields for multi-view reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5589-5599, 2021. 1 +[24] Ri-Zhao Qiu, Ge Yang, Weijia Zeng, and Xiaolong Wang. Feature splatting: Language-driven physics-based scene synthesis and editing. arXiv preprint arXiv:2404.01223, 2024. 4 +[25] Yufan Ren, Fangjinhua Wang, Tong Zhang, Marc Pollefeys, and Sabine Susstrunk. Volrecon: Volume rendering of signed ray distance functions for generalizable multi-view reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16685-16695, 2023. 1, 2, 3, 6 + +[26] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4104-4113, 2016. 3, 6 +[27] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5459-5469, 2022. 2 +[28] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. arXiv preprint arXiv:2106.10689, 2021. 1, 2, 6 +[29] Yiqun Wang, Ivan Skorokhodov, and Peter Wonka. Improved surface reconstruction using high-frequency details. 2022. 2 +[30] Jiang Wu, Rui Li, Haofei Xu, Wenxun Zhao, Yu Zhu, Jinqiu Sun, and Yanning Zhang. Gomvs: Geometrically consistent cost aggregation for multi-view stereo. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20207-20216, 2024. 1, 2 +[31] Kaiqiang Xiong, Rui Peng, Zhe Zhang, Tianxing Feng, Jianbo Jiao, Feng Gao, and Ronggang Wang. Cl-mvsnet: Unsupervised multi-view stereo with dual-level contrastive learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3769-3780, 2023. 1, 2, 3, 4, 6, 7 +[32] Jialei Xu, Xianming Liu, Yuanchao Bai, Junjun Jiang, Kaixuan Wang, Xiaozhi Chen, and Xiangyang Ji. Multi-camera collaborative depth prediction via consistent structure estimation. In Proceedings of the 30th ACM international conference on multimedia, pages 2730-2738, 2022. 2 +[33] Jialei Xu, Wei Yin, Dong Gong, Junjun Jiang, and Xianming Liu. Sdge: Stereo guided depth estimation for $360^{\circ}$ camera sets. In 2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 11179-11186. IEEE, 2024. 2 +[34] Luoyuan Xu, Tao Guan, Yuesong Wang, Wenkai Liu, Zhaojie Zeng, Junle Wang, and Wei Yang. C2f2neus: Cascade cost frustum fusion for high fidelity and generalizable neural surface reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 18291-18301, 2023. 1, 3, 6 +[35] Qingshan Xu, Wanjuan Su, Yuhang Qi, Wenbing Tao, and Marc Pollefeys. Learning inverse depth regression for pixelwise visibility-aware multi-view stereo networks. International Journal of Computer Vision, 130(8):2040-2059, 2022. 2 +[36] Yao Yao, Zixin Luo, Shiwei Li, Tian Fang, and Long Quan. Mvsnet: Depth inference for unstructured multi-view stereo. In Proceedings of the European conference on computer vision (ECCV), pages 767-783, 2018. 1, 2 +[37] Yao Yao, Zixin Luo, Shiwei Li, Jingyang Zhang, Yufan Ren, Lei Zhou, Tian Fang, and Long Quan. Blendedmvs: A large-scale dataset for generalized multi-view stereo networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1790-1799, 2020. 5 +[38] Lior Yariv, Jiatao Gu, Yoni Kasten, and Yaron Lipman. Volume rendering of neural implicit surfaces. Advances in Neu + +ral Information Processing Systems, 34:4805-4815, 2021. 1, 2, 6 +[39] Jae-Chern Yoo and Tae Hee Han. Fast normalized cross-correlation. Circuits, systems and signal processing, 28:819-843, 2009. 5 +[40] Mae Younes, Amine Ouasfi, and Adnane Boukhayma. Sparsecraft: Few-shot neural reconstruction through stereopsis guided geometric linearization. In European Conference on Computer Vision, pages 37-56. Springer, 2024. 3, 6 +[41] Mulin Yu, Tao Lu, Linning Xu, Lihan Jiang, Yuanbo Xiangli, and Bo Dai. Gsdf: 3dgs meets sdf for improved rendering and reconstruction. arXiv preprint arXiv:2403.16964, 2024. 1, 2 +[42] Zehao Yu, Songyou Peng, Michael Niemeyer, Torsten Sattler, and Andreas Geiger. Monosdf: Exploring monocular geometric cues for neural implicit surface reconstruction. Advances in neural information processing systems, 35:25018-25032, 2022. 2 +[43] Zehao Yu, Torsten Sattler, and Andreas Geiger. Gaussian opacity fields: Efficient and compact surface reconstruction in unbounded scenes. arXiv preprint arXiv:2404.10772, 2024. 1, 2, 6 +[44] Baowen Zhang, Chuan Fang, Rakesh Shrestha, Yixun Liang, Xiaoxiao Long, and Ping Tan. Rade-gs: Rasterizing depth in gaussian splatting. arXiv preprint arXiv:2406.01467, 2024. 2 +[45] Jingyang Zhang, Yao Yao, and Long Quan. Learning signed distance field for multi-view surface reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6525-6534, 2021. 5 +[46] Jingyang Zhang, Shiwei Li, Zixin Luo, Tian Fang, and Yao Yao. Vis-mvsnet: Visibility-aware multi-view stereo network. International Journal of Computer Vision, 131(1): 199-214, 2023. 2 +[47] Shijie Zhou, Haoran Chang, Sicheng Jiang, Zhiwen Fan, Zehao Zhu, Dejia Xu, Pradyumna Chari, Suya You, Zhangyang Wang, and Achuta Kadambi. Feature 3dgs: Supercharging 3d gaussian splattering to enable distilled feature fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21676-21685, 2024. 4 \ No newline at end of file diff --git a/2025/Sparse2DGS_ Geometry-Prioritized Gaussian Splatting for Surface Reconstruction from Sparse Views/images.zip b/2025/Sparse2DGS_ Geometry-Prioritized Gaussian Splatting for Surface Reconstruction from Sparse Views/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..84a7e6f611cdc880797ee63f7c96810f092e14fc --- /dev/null +++ b/2025/Sparse2DGS_ Geometry-Prioritized Gaussian Splatting for Surface Reconstruction from Sparse Views/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5934a2cda8dc364ce4bda020f11ebc94a3145ebb2c28d140e511ea2824adc743 +size 738143 diff --git a/2025/Sparse2DGS_ Geometry-Prioritized Gaussian Splatting for Surface Reconstruction from Sparse Views/layout.json b/2025/Sparse2DGS_ Geometry-Prioritized Gaussian Splatting for Surface Reconstruction from Sparse Views/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..7af6606a6aaa6fedecc872ca0c39ee42e1270b36 --- /dev/null +++ b/2025/Sparse2DGS_ Geometry-Prioritized Gaussian Splatting for Surface Reconstruction from Sparse Views/layout.json @@ -0,0 +1,8368 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 124, + 103, + 488, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 124, + 103, + 488, + 140 + ], + "spans": [ + { + "bbox": [ + 124, + 103, + 488, + 140 + ], + "type": "text", + "content": "Sparse2DGS: Geometry-Prioritized Gaussian Splating for Surface Reconstruction from Sparse Views" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 126, + 160, + 484, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 126, + 160, + 484, + 190 + ], + "spans": [ + { + "bbox": [ + 126, + 160, + 484, + 190 + ], + "type": "text", + "content": "Jiang Wu Rui Li Yu Zhu† Rong Guo Jinqiu Sun Yanning Zhang† Northwestern Polytechnical University" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 63, + 222, + 549, + 349 + ], + "blocks": [ + { + "bbox": [ + 63, + 222, + 549, + 349 + ], + "lines": [ + { + "bbox": [ + 63, + 222, + 549, + 349 + ], + "spans": [ + { + "bbox": [ + 63, + 222, + 549, + 349 + ], + "type": "image", + "image_path": "5de91afae79a2e02772b387629ac4a9a80ef02c5eb2ec8b0f33fbbba5c40b093.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 357, + 555, + 392 + ], + "lines": [ + { + "bbox": [ + 54, + 357, + 555, + 392 + ], + "spans": [ + { + "bbox": [ + 54, + 357, + 555, + 392 + ], + "type": "text", + "content": "Figure 1. Sparse2DGS boosts the strengths of Gaussian Splatting and Multi-view Stereo in sparse-view surface reconstruction, with notable improvement over 2DGS [11], CLMVSNet [31], as well as their plain combination (left). Meanwhile, Sparse2DGS is " + }, + { + "bbox": [ + 54, + 357, + 555, + 392 + ], + "type": "inline_equation", + "content": "2 \\times" + }, + { + "bbox": [ + 54, + 357, + 555, + 392 + ], + "type": "text", + "content": " faster than NeRF-based sparse-view method [20] (middle) and leads to complete and accurate reconstruction (right)." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 401, + 200, + 414 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 401, + 200, + 414 + ], + "spans": [ + { + "bbox": [ + 151, + 401, + 200, + 414 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 426, + 297, + 618 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 426, + 297, + 618 + ], + "spans": [ + { + "bbox": [ + 54, + 426, + 297, + 618 + ], + "type": "text", + "content": "We present a Gaussian Splatting method for surface reconstruction using sparse input views. Previous methods relying on dense views struggle with extremely sparse Structure-from-Motion points for initialization. While learning-based Multi-view Stereo (MVS) provides dense 3D points, directly combining it with Gaussian Splatting leads to suboptimal results due to the ill-posed nature of sparse-view geometric optimization. We propose Sparse2DGS, an MVS-initialized Gaussian Splatting pipeline for complete and accurate reconstruction. Our key insight is to incorporate the geometric-prioritized enhancement schemes, allowing for direct and robust geometric learning under ill-posed conditions. Sparse2DGS outperforms existing methods by notable margins while being " + }, + { + "bbox": [ + 54, + 426, + 297, + 618 + ], + "type": "inline_equation", + "content": "2 \\times" + }, + { + "bbox": [ + 54, + 426, + 297, + 618 + ], + "type": "text", + "content": " faster than the NeRF-based fine-tuning approach. Code is available at https://github.com/Wuuu3511/Sparse2DGS." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 639, + 135, + 651 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 639, + 135, + 651 + ], + "spans": [ + { + "bbox": [ + 55, + 639, + 135, + 651 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 659, + 296, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 659, + 296, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 659, + 296, + 696 + ], + "type": "text", + "content": "Reconstructing 3D surfaces from multiple views is a fundamental task in computer vision and has been widely used in autonomous driving [9, 18] and virtual reality [21]. While" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 403, + 555, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 403, + 555, + 449 + ], + "spans": [ + { + "bbox": [ + 313, + 403, + 555, + 449 + ], + "type": "text", + "content": "typical methods require abundant input images, reconstructing accurate 3D surfaces from sparse views is particularly challenging and is crucial for scenes that are frequently captured with limited images." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 449, + 556, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 449, + 556, + 555 + ], + "spans": [ + { + "bbox": [ + 313, + 449, + 556, + 555 + ], + "type": "text", + "content": "NeRF-based surface reconstruction methods [22] have gained popularity by encoding scene geometry into an occupancy field [17, 23] or a signed distance function (SDF) [28, 38]. However, when faced with sparse views, these methods often yield incomplete surfaces with artifacts due to insufficient supervision. Although sparse-view methods [19, 20, 25, 34] use pre-trained feed-forward NeRFs to provide geometric priors, they still exhibit limited generalization to unseen scenes." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 557, + 556, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 557, + 556, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 557, + 556, + 689 + ], + "type": "text", + "content": "Recently, 3D Gaussian Splatting (3DGS) [14] has leveraged Structure-from-Motion (SfM) point cloud to construct Gaussian primitives for fast and high-quality novel view synthesis (NVS). Subsequent works [4, 5, 7, 11, 41, 43] aim to achieve more efficient 3D surface reconstruction using 3DGS. Notably, 2D Gaussian Splatting (2DGS) [11] proposes a 2D variant of Gaussian primitives, with disk-shaped primitives suitable for geometry representation. However, when dealing with sparse views, existing methods [4, 11, 43] struggle to produce faithful reconstructions due to initially over-sparse SfM points." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": "On the other hand, learning-based multi-view stereo (MVS) [6, 16, 30, 36] predicts dense depth maps from" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 69, + 702, + 182, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 702, + 182, + 713 + ], + "spans": [ + { + "bbox": [ + 69, + 702, + 182, + 713 + ], + "type": "inline_equation", + "content": "^\\dagger" + }, + { + "bbox": [ + 69, + 702, + 182, + 713 + ], + "type": "text", + "content": " indicates corresponding authors." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "11307" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "content": "pixel-wise feature matching, significantly improving geometric completeness even with sparse images. However, MVS depth networks focus on generalizable inference instead of test-time geometry enhancement by design. While combining MVS point cloud with Gaussian Splatting is intuitively effective, their plain combination (see Fig. 1, \"2DGS+CLMVS\") yields sub-optimal results due to the ill-posed nature of geometry optimization from sparse views." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 169, + 295, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 169, + 295, + 420 + ], + "spans": [ + { + "bbox": [ + 56, + 169, + 295, + 420 + ], + "type": "text", + "content": "In this paper, we propose Sparse2DGS, an MVS-initialized Gaussian Splating method for complete and accurate sparse-view 3D surface reconstruction. The key idea is to boost the strengths of two paradigms through geometric-prioritized enhancement against the ill-posed nature of sparse-view reconstruction. Starting from MVS initialization, we first incorporate MVS-derived geometric features into the Gaussian splattering process while keeping feature and color values fixed. This provides geometrically expressive supervision and prevents appearance overfitting, thereby improving geometric reconstruction quality. We further introduce a direct Gaussian primitive regularization, with a reparameterization-based disk sampling strategy that reformulates Gaussian primitives' position, orientation, and scale into point representations. This approach enables a unified optimization of different forms of Gaussian properties within a standard point-based, cross-view consistency loss. Moreover, instead of densifying Gaussian primitives under ill-posed conditions, we propose to selectively update the MVS-initialized Gaussian primitives with rendered geometric cues, allowing for self-guided geometric corrections." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 421, + 295, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 421, + 295, + 479 + ], + "spans": [ + { + "bbox": [ + 55, + 421, + 295, + 479 + ], + "type": "text", + "content": "Sparse2DGS significantly outperforms existing Gaussian Splatting methods in sparse-view surface reconstruction. Meanwhile, it demonstrates notable efficiency advantages over NeRF-based test-time optimization. Our contributions are as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 481, + 295, + 637 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 55, + 481, + 295, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 481, + 295, + 529 + ], + "spans": [ + { + "bbox": [ + 55, + 481, + 295, + 529 + ], + "type": "text", + "content": "- We propose Sparse2DGS, a test-time sparse-view 3D surface reconstruction method that boosts the strengths of Gaussian Splitting and MVS, allowing for complete and accurate reconstruction from sparse images." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 529, + 295, + 637 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 529, + 295, + 637 + ], + "spans": [ + { + "bbox": [ + 55, + 529, + 295, + 637 + ], + "type": "text", + "content": "- We focus on geometric-prioritized enhancement by proposing geometrically enhanced supervision, direct Gaussian primitive regularization, and selective Gaussian update throughout the Gaussian Splatting pipeline. Sparse2DGS achieves the lowest error (1.13 CD error) compared to 2DGS (2.81) [11], GOF (2.82) [43], and PGSR (2.08) [4] on DTU dataset with 3 views. Meanwhile, Sparse2DGS is " + }, + { + "bbox": [ + 55, + 529, + 295, + 637 + ], + "type": "inline_equation", + "content": "2 \\times" + }, + { + "bbox": [ + 55, + 529, + 295, + 637 + ], + "type": "text", + "content": " faster than fine-tuning NeRF-based SparseNeus [20]." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 650, + 146, + 662 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 650, + 146, + 662 + ], + "spans": [ + { + "bbox": [ + 55, + 650, + 146, + 662 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 670, + 284, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 670, + 284, + 682 + ], + "spans": [ + { + "bbox": [ + 55, + 670, + 284, + 682 + ], + "type": "text", + "content": "2.1. 3D Reconstruction via Novel View Synthesis" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 689, + 295, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 712 + ], + "type": "text", + "content": "Recent progresses [11, 12, 25, 43] in novel view synthesis (NVS) have enabled accurate 3D surface reconstruc" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 553, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 203 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 203 + ], + "type": "text", + "content": "tion with only image-based supervision. Neural Radiance Fields [22] (NeRFs) have become a significant technique for representing 3D scenes. It has inspired numerous works on directly extracting surfaces by training a parameterized radiance field through differentiable volume rendering. Neus [28] and VolSDF [38] integrate surface learning into volume rendering by re-parameterizing the signed distance function as volume density, reconstructing more accurate surfaces. To reduce geometric deviations and accelerate training, subsequent works [8, 27, 29, 42] have introduced improvements in various areas." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 204, + 553, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 204, + 553, + 443 + ], + "spans": [ + { + "bbox": [ + 313, + 204, + 553, + 443 + ], + "type": "text", + "content": "Recently, 3D Gaussian Splatting [14] has shown impressive results in novel view synthesis (NVS), with many recent studies adopting this approach for surface reconstruction. SuGaR [10] aligns 3D Gaussians with scene surfaces by designing a regularization term and extracts an accurate mesh from the Gaussians using Poisson Reconstruction [13]. 2DGS [11] replaces 3D Gaussian ellipsoids with 2D-oriented planar disks, providing view-consistent geometric representations. RaDe-GS [44] further designs a rasterized approach to render depth and normal, enhancing geometric accuracy while ensuring rendering efficiency. GOF [43] extracts the level set from trained Gaussian primitives by constructing a Gaussian opacity field. PGSR [4] improves mesh quality by introducing single-view and multi-view geometric constraints. Additionally, GSDF [41] and NeusSG [5] reconstruct scenes by combining the advantages of 3D Gaussian Splatting and SDF networks. Despite these methods' success, the extracted geometry tends to be noisy and incomplete when the input number of training viewpoints is limited." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 452, + 497, + 464 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 452, + 497, + 464 + ], + "spans": [ + { + "bbox": [ + 313, + 452, + 497, + 464 + ], + "type": "text", + "content": "2.2. Learning-based Multi-view Stereo" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 471, + 553, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 471, + 553, + 650 + ], + "spans": [ + { + "bbox": [ + 313, + 471, + 553, + 650 + ], + "type": "text", + "content": "Deep learning-based Multi-view Stereo (MVS) estimates the reference image's depth from multi-view posed images. MVSNet [36] first introduces the differentiable homography warping to build the cost volume with deep image features and then employs a 3D CNN for cost regularization. Recent follow-up works have introduced improvements in various aspects, including multi-view feature extraction [2, 6], cost aggregation [32, 33, 35, 46], and cost volume regularization [30]. However, these methods require ground-truth depth maps for training, which are challenging to obtain in practice. Unsupervised MVS [3, 15, 31] can be applied in more general, real-world scenarios. MVS methods can infer dense depth maps for each input view and therefore, we select the unsupervised CLMVSNet [31] to initialize the Gaussian point cloud." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 658, + 507, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 658, + 507, + 671 + ], + "spans": [ + { + "bbox": [ + 313, + 658, + 507, + 671 + ], + "type": "text", + "content": "2.3. Sparse View Surface Reconstruction" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "content": "Recent works have explored 3D surface reconstruction from sparse training views. SparseNeuS [20] achieves reasonable mesh reconstruction results by constructing a cascaded" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11308" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 68, + 550, + 255 + ], + "blocks": [ + { + "bbox": [ + 60, + 68, + 550, + 255 + ], + "lines": [ + { + "bbox": [ + 60, + 68, + 550, + 255 + ], + "spans": [ + { + "bbox": [ + 60, + 68, + 550, + 255 + ], + "type": "image", + "image_path": "6a87a0e2c57120679f272a77cf9a2d5b3d88bc0b0ba4c0d9b0fe890f1f77df64.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 262, + 555, + 330 + ], + "lines": [ + { + "bbox": [ + 54, + 262, + 555, + 330 + ], + "spans": [ + { + "bbox": [ + 54, + 262, + 555, + 330 + ], + "type": "text", + "content": "Figure 2. Overview. Given sparse posed images, we first leverage the MVS points to initialize the Gaussian position " + }, + { + "bbox": [ + 54, + 262, + 555, + 330 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 54, + 262, + 555, + 330 + ], + "type": "text", + "content": " (Sec. 3.1). We then leverage the MVS-derived feature for Gaussian Splating, with fixed feature " + }, + { + "bbox": [ + 54, + 262, + 555, + 330 + ], + "type": "inline_equation", + "content": "\\mathbf{f}" + }, + { + "bbox": [ + 54, + 262, + 555, + 330 + ], + "type": "text", + "content": " and color " + }, + { + "bbox": [ + 54, + 262, + 555, + 330 + ], + "type": "inline_equation", + "content": "\\mathbf{c}" + }, + { + "bbox": [ + 54, + 262, + 555, + 330 + ], + "type": "text", + "content": " values to conduct geometrically expressive supervision and avoid appearance overfitting (Sec. 3.2). We then optimize Gaussian primitive properties by reformulating orientation " + }, + { + "bbox": [ + 54, + 262, + 555, + 330 + ], + "type": "inline_equation", + "content": "\\mathbf{R}" + }, + { + "bbox": [ + 54, + 262, + 555, + 330 + ], + "type": "text", + "content": ", scale " + }, + { + "bbox": [ + 54, + 262, + 555, + 330 + ], + "type": "inline_equation", + "content": "\\mathbf{S}" + }, + { + "bbox": [ + 54, + 262, + 555, + 330 + ], + "type": "text", + "content": ", and position " + }, + { + "bbox": [ + 54, + 262, + 555, + 330 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 54, + 262, + 555, + 330 + ], + "type": "text", + "content": " into sampled points using the reparameterization-based disk sampling, ensuring simultaneously optimizing different forms of Gaussian properties with point-based loss (Sec. 3.3). Finally, we introduce a Selective Gaussian Update strategy that leverages rendered geometry as an alternative to update Gaussian primitive positions, which further improves reconstruction quality (Sec. 3.4)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 340, + 296, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 340, + 296, + 473 + ], + "spans": [ + { + "bbox": [ + 54, + 340, + 296, + 473 + ], + "type": "text", + "content": "volume to learn a generalizable radiance field, followed by per-scene fine-tuning. NeuSurf [12] learns dense on-surface priors from sparse points to guide geometric learning. SparseCraft [40] leverages stereopsis cues for geometric regularization based on Taylor expansion. Moreover, VolRecon [25] and ReTR [19] enhance voxel features by designing specialized transformer architectures. C2F2Neus [34] constructs per-view cost frustum for finer geometry estimation. Although these feedforward networks can directly infer scene geometry, their limited generalization to unseen scenes restricts their applicability in real-world scenarios." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 478, + 138, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 478, + 138, + 491 + ], + "spans": [ + { + "bbox": [ + 55, + 478, + 138, + 491 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 498, + 295, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 498, + 295, + 665 + ], + "spans": [ + { + "bbox": [ + 54, + 498, + 295, + 665 + ], + "type": "text", + "content": "Given sparse view posed images " + }, + { + "bbox": [ + 54, + 498, + 295, + 665 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{I}_i\\}_{i=1}^N" + }, + { + "bbox": [ + 54, + 498, + 295, + 665 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 54, + 498, + 295, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_i \\in \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 54, + 498, + 295, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 498, + 295, + 665 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 54, + 498, + 295, + 665 + ], + "type": "text", + "content": " denotes the view number, Sparse2DGS aims to reconstruct accurate 3D scene surfaces using Gaussian Splatting. As shown in Fig. 2, we first initialize the Gaussian primitives with the MVS [31] reconstructed point cloud. Then, we sample MVS geometric feature maps into feature splatting with its value and RGB color fixed for supervision. For each Gaussian primitive, we regularize its geometric properties (position, orientation, scale) by a novel reparameterization-based 2D disk sampling strategy that formulates these properties as point-based supervision. Finally, we propose a selective Gaussian update strategy to further enhance Gaussian primitives using rendered geometric cues." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 671, + 295, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 671, + 295, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 671, + 295, + 685 + ], + "type": "text", + "content": "3.1. 2D Gaussian Splitting with MVS initialization" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": "2D Gaussian Splitting [11] (2DGS) represents Gaussian primitives as planar disks suitable for representing and op-" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 340, + 555, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 340, + 555, + 399 + ], + "spans": [ + { + "bbox": [ + 313, + 340, + 555, + 399 + ], + "type": "text", + "content": "timizing 3D surfaces. However, it relies on sparse SfM [26] 3D points for initialization, which exhibits limited reconstruction quality under sparse-view inputs. To address this, we initialize 2DGS with an MVS point cloud to provide dense geometric representations." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 400, + 556, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 400, + 556, + 497 + ], + "spans": [ + { + "bbox": [ + 313, + 400, + 556, + 497 + ], + "type": "text", + "content": "2D Gaussian Splatting. Each 2D Gaussian primitive is characterized by its central position " + }, + { + "bbox": [ + 313, + 400, + 556, + 497 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 313, + 400, + 556, + 497 + ], + "type": "text", + "content": ", two principal tangential vectors " + }, + { + "bbox": [ + 313, + 400, + 556, + 497 + ], + "type": "inline_equation", + "content": "\\mathbf{t_u}" + }, + { + "bbox": [ + 313, + 400, + 556, + 497 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 400, + 556, + 497 + ], + "type": "inline_equation", + "content": "\\mathbf{t_v}" + }, + { + "bbox": [ + 313, + 400, + 556, + 497 + ], + "type": "text", + "content": ", and a scaling vector " + }, + { + "bbox": [ + 313, + 400, + 556, + 497 + ], + "type": "inline_equation", + "content": "(s_u, s_v)" + }, + { + "bbox": [ + 313, + 400, + 556, + 497 + ], + "type": "text", + "content": ". The rotation matrix of a 2D Gaussian can be represented as " + }, + { + "bbox": [ + 313, + 400, + 556, + 497 + ], + "type": "inline_equation", + "content": "\\mathbf{R} = (\\mathbf{t_u}, \\mathbf{t_v}, \\mathbf{t_z})" + }, + { + "bbox": [ + 313, + 400, + 556, + 497 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 400, + 556, + 497 + ], + "type": "inline_equation", + "content": "\\mathbf{t_z}" + }, + { + "bbox": [ + 313, + 400, + 556, + 497 + ], + "type": "text", + "content": " is defined by two orthogonal tangent vectors " + }, + { + "bbox": [ + 313, + 400, + 556, + 497 + ], + "type": "inline_equation", + "content": "\\mathbf{t_z} = \\mathbf{t_u} \\times \\mathbf{t_v}" + }, + { + "bbox": [ + 313, + 400, + 556, + 497 + ], + "type": "text", + "content": " and the scale matrix is " + }, + { + "bbox": [ + 313, + 400, + 556, + 497 + ], + "type": "inline_equation", + "content": "\\mathbf{S} = \\mathrm{diag}(s_u, s_v, 0)" + }, + { + "bbox": [ + 313, + 400, + 556, + 497 + ], + "type": "text", + "content": ". A 2D Gaussian is then defined in a local tangent plane in world space, which is parameterized:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 369, + 506, + 555, + 520 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 369, + 506, + 555, + 520 + ], + "spans": [ + { + "bbox": [ + 369, + 506, + 555, + 520 + ], + "type": "interline_equation", + "content": "P (u, v) = \\mathbf {p} + s _ {u} \\mathbf {t} _ {u} u + s _ {v} \\mathbf {t} _ {v} v. \\tag {1}", + "image_path": "b44272bac82266095d5a7a4a4b5fba1cd22405eaba45f1f6e7bb0a612132acfa.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 529, + 554, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 529, + 554, + 552 + ], + "spans": [ + { + "bbox": [ + 313, + 529, + 554, + 552 + ], + "type": "text", + "content": "For the point " + }, + { + "bbox": [ + 313, + 529, + 554, + 552 + ], + "type": "inline_equation", + "content": "\\mathbf{u} = (u,v)" + }, + { + "bbox": [ + 313, + 529, + 554, + 552 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 313, + 529, + 554, + 552 + ], + "type": "inline_equation", + "content": "uv" + }, + { + "bbox": [ + 313, + 529, + 554, + 552 + ], + "type": "text", + "content": " space, its 2D Gaussian value can then be evaluated by standard Gaussian" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 381, + 561, + 554, + 586 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 381, + 561, + 554, + 586 + ], + "spans": [ + { + "bbox": [ + 381, + 561, + 554, + 586 + ], + "type": "interline_equation", + "content": "\\mathcal {G} (\\mathbf {u}) = \\exp \\frac {- \\left(u ^ {2} + v ^ {2}\\right)}{2}. \\tag {2}", + "image_path": "c45158f8a991b711bcbad07c967e81b5fb35d9485c52575706b4423400828213.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "text", + "content": "Initializing 2DGS with MVS depth. MVS methods estimate dense depth maps through cross-view matching. In particular, unsupervised MVS [31] reconstructs complete scenes without ground truth for training, offering greater flexibility. Therefore, we use the point cloud reconstructed by the current unsupervised method CLMVSNet [31] to initialize our Gaussian point cloud. Specifically, given sparse views' images " + }, + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{I}_i\\}_{i=1}^N" + }, + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "text", + "content": ", we use the MVS network for depth maps " + }, + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{D}_i\\}_{i=1}^N" + }, + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "text", + "content": ", then back-project these depth maps and fuse into a point cloud " + }, + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{p}_{ij} \\mid 1 \\leq i \\leq N, 1 \\leq j \\leq HW\\}" + }, + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "text", + "content": " as ini-" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "11309" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "text", + "content": "tialized positions for Gaussian primitives, where " + }, + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "text", + "content": " denotes image index and " + }, + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "text", + "content": " is the pixel index within each image." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 103, + 253, + 117 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 103, + 253, + 117 + ], + "spans": [ + { + "bbox": [ + 55, + 103, + 253, + 117 + ], + "type": "text", + "content": "3.2. Geometrically Enhanced Supervision" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 122, + 296, + 242 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 122, + 296, + 242 + ], + "spans": [ + { + "bbox": [ + 55, + 122, + 296, + 242 + ], + "type": "text", + "content": "3D/2D Gaussian Splatting [11, 14] couples color and geometry optimization through image-based loss. However, this leads to severe degradation under sparse views as color will overfit sparse views, leading to a compromised geometry. Meanwhile, the challenges in image areas (e.g., textureless regions) will be amplified with sparse views. To address this issue, we propose geometrically enhanced supervision by incorporating off-the-shelf expressive MVS features for feature splatting and supervision, with the color and feature fixed to avoid overfitting." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 243, + 296, + 362 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 243, + 296, + 362 + ], + "spans": [ + { + "bbox": [ + 55, + 243, + 296, + 362 + ], + "type": "text", + "content": "In addition to providing dense point clouds from input images " + }, + { + "bbox": [ + 55, + 243, + 296, + 362 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{I}_i\\}_{i = 1}^N" + }, + { + "bbox": [ + 55, + 243, + 296, + 362 + ], + "type": "text", + "content": ", the MVS network [31] offers off-the-shelf feature maps with geometric expressiveness. We leverage MVS features to enhance geometric supervision by incorporating additional feature attribute " + }, + { + "bbox": [ + 55, + 243, + 296, + 362 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{ij}" + }, + { + "bbox": [ + 55, + 243, + 296, + 362 + ], + "type": "text", + "content": " into Gaussian primitives. Specifically, for each input image " + }, + { + "bbox": [ + 55, + 243, + 296, + 362 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_i" + }, + { + "bbox": [ + 55, + 243, + 296, + 362 + ], + "type": "text", + "content": ", we extract its feature map " + }, + { + "bbox": [ + 55, + 243, + 296, + 362 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_i" + }, + { + "bbox": [ + 55, + 243, + 296, + 362 + ], + "type": "text", + "content": " from the Feature Pyramid Network (FPN) of CLMVSNet [31]. Since each 3D point corresponds to an image pixel due to MVS initialization, we sample both RGB color and MVS features into each Gaussian primitive." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 102, + 372, + 295, + 387 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 372, + 295, + 387 + ], + "spans": [ + { + "bbox": [ + 102, + 372, + 295, + 387 + ], + "type": "interline_equation", + "content": "\\mathbf {f} _ {i j} = \\mathbf {F} _ {i} (\\pi (\\mathbf {p} _ {i j})), \\mathbf {c} _ {i j} = \\mathbf {I} _ {i} (\\pi (\\mathbf {p} _ {i j})), \\tag {3}", + "image_path": "2d0c6bc73978f3d4f3ce341da1328cbc6a1edbb55809ca444f9c89462a8d33aa.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 394, + 296, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 394, + 296, + 455 + ], + "spans": [ + { + "bbox": [ + 55, + 394, + 296, + 455 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 394, + 296, + 455 + ], + "type": "inline_equation", + "content": "\\pi(\\cdot)" + }, + { + "bbox": [ + 55, + 394, + 296, + 455 + ], + "type": "text", + "content": " denotes the projection from 3D space onto 2D image plane. We further render the feature into an image plane using the tile-based renderer together with RGB rendering. We show the rendering process given " + }, + { + "bbox": [ + 55, + 394, + 296, + 455 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 394, + 296, + 455 + ], + "type": "text", + "content": " Gaussian primitives along a ray, with indices " + }, + { + "bbox": [ + 55, + 394, + 296, + 455 + ], + "type": "inline_equation", + "content": "i, j" + }, + { + "bbox": [ + 55, + 394, + 296, + 455 + ], + "type": "text", + "content": " omitted for brevity" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 86, + 473, + 295, + 506 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 473, + 295, + 506 + ], + "spans": [ + { + "bbox": [ + 86, + 473, + 295, + 506 + ], + "type": "interline_equation", + "content": "\\{\\hat {\\mathbf {c}}, \\hat {\\mathbf {f}} \\} = \\sum_ {m = 1} ^ {M} \\left\\{\\mathbf {c} _ {m}, \\mathbf {f} _ {m} \\right\\} \\cdot \\alpha_ {m} \\prod_ {o = 1} ^ {m - 1} (1 - \\alpha_ {m}), \\tag {4}", + "image_path": "e0a5d28c31162cd6a749ed2a5a1dcc339f2adcd8f9fb24286db39c3ba76f8b8e.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 514, + 296, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 514, + 296, + 563 + ], + "spans": [ + { + "bbox": [ + 55, + 514, + 296, + 563 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 514, + 296, + 563 + ], + "type": "inline_equation", + "content": "\\alpha_{m}" + }, + { + "bbox": [ + 55, + 514, + 296, + 563 + ], + "type": "text", + "content": " is the opacity of the Gaussian. With rendered feature maps that have enhanced geometric expressiveness for the challenging sparse views, we further supervise the rendered features using cosine similarity loss by" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 125, + 571, + 295, + 601 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 125, + 571, + 295, + 601 + ], + "spans": [ + { + "bbox": [ + 125, + 571, + 295, + 601 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {f e a} = 1 - \\frac {\\mathbf {f} _ {i j} \\cdot \\hat {\\mathbf {f}} _ {i j}}{\\| \\mathbf {f} _ {i j} \\| \\| \\hat {\\mathbf {f}} _ {i j} \\|}, \\tag {5}", + "image_path": "b9f305df178d3d6fbf7477ae92c232612e4e42a2841bca5748723f2768e9ff4c.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 605, + 295, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 666 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 666 + ], + "type": "text", + "content": "Note that the simultaneous learning of appearance (i.e., color " + }, + { + "bbox": [ + 55, + 605, + 295, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_{ij}" + }, + { + "bbox": [ + 55, + 605, + 295, + 666 + ], + "type": "text", + "content": " and feature " + }, + { + "bbox": [ + 55, + 605, + 295, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{ij}" + }, + { + "bbox": [ + 55, + 605, + 295, + 666 + ], + "type": "text", + "content": ") and geometry lead to geometric degradation under sparse views (see Tab. 4). To address this issue, both " + }, + { + "bbox": [ + 55, + 605, + 295, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{ij}" + }, + { + "bbox": [ + 55, + 605, + 295, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 605, + 295, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_{ij}" + }, + { + "bbox": [ + 55, + 605, + 295, + 666 + ], + "type": "text", + "content": " are fixed during the optimization to enforce the learning specifically for geometry." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 666, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 666, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 666, + 296, + 713 + ], + "type": "text", + "content": "Our method essentially differentiates previous feature-based methods [24, 47] in its specific strategy for geometric enhancement. In our method, the expressive MVS features are incorporated, and features of each Gaussian are fixed," + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 326, + 73, + 541, + 222 + ], + "blocks": [ + { + "bbox": [ + 326, + 73, + 541, + 222 + ], + "lines": [ + { + "bbox": [ + 326, + 73, + 541, + 222 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 541, + 222 + ], + "type": "image", + "image_path": "a9ca6fc52cdf00f37bf78b84109e47bd721d5cd4a5c260b55463f271c3ff79df.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 234, + 555, + 322 + ], + "lines": [ + { + "bbox": [ + 313, + 234, + 555, + 322 + ], + "spans": [ + { + "bbox": [ + 313, + 234, + 555, + 322 + ], + "type": "text", + "content": "Figure 3. Direct Gaussian Primitive Regularization. We sample point set " + }, + { + "bbox": [ + 313, + 234, + 555, + 322 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}' \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})" + }, + { + "bbox": [ + 313, + 234, + 555, + 322 + ], + "type": "text", + "content": " from the standard Gaussian distribution and leverage the reparameterization technique to transform " + }, + { + "bbox": [ + 313, + 234, + 555, + 322 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}'" + }, + { + "bbox": [ + 313, + 234, + 555, + 322 + ], + "type": "text", + "content": " to the point set " + }, + { + "bbox": [ + 313, + 234, + 555, + 322 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}" + }, + { + "bbox": [ + 313, + 234, + 555, + 322 + ], + "type": "text", + "content": " on each Gaussian primitive. This process allows for representing Gaussian orientation " + }, + { + "bbox": [ + 313, + 234, + 555, + 322 + ], + "type": "inline_equation", + "content": "\\mathbf{R}" + }, + { + "bbox": [ + 313, + 234, + 555, + 322 + ], + "type": "text", + "content": " and scale " + }, + { + "bbox": [ + 313, + 234, + 555, + 322 + ], + "type": "inline_equation", + "content": "\\mathbf{S}" + }, + { + "bbox": [ + 313, + 234, + 555, + 322 + ], + "type": "text", + "content": " properties in the form of points easier for geometric supervision. The transformed points " + }, + { + "bbox": [ + 313, + 234, + 555, + 322 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}" + }, + { + "bbox": [ + 313, + 234, + 555, + 322 + ], + "type": "text", + "content": " are then projected to the source and target views, supervised by cross-view feature consistency loss." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 335, + 554, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 335, + 554, + 384 + ], + "spans": [ + { + "bbox": [ + 313, + 335, + 554, + 384 + ], + "type": "text", + "content": "both for geometry enhancement with sparse views. Previous methods leverage splatted features only for segmentation or editing [24, 47] with dense views, while their feature losses do not optimize the scene geometry." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 393, + 531, + 406 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 393, + 531, + 406 + ], + "spans": [ + { + "bbox": [ + 313, + 393, + 531, + 406 + ], + "type": "text", + "content": "3.3. Direct Gaussian Primitive Regularization" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 412, + 554, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 412, + 554, + 508 + ], + "spans": [ + { + "bbox": [ + 313, + 412, + 554, + 508 + ], + "type": "text", + "content": "The disk-shaped 2D Gaussian primitives encode latent scene geometry including local shape orientation " + }, + { + "bbox": [ + 313, + 412, + 554, + 508 + ], + "type": "inline_equation", + "content": "\\mathbf{R}" + }, + { + "bbox": [ + 313, + 412, + 554, + 508 + ], + "type": "text", + "content": ", scale " + }, + { + "bbox": [ + 313, + 412, + 554, + 508 + ], + "type": "inline_equation", + "content": "\\mathbf{S}" + }, + { + "bbox": [ + 313, + 412, + 554, + 508 + ], + "type": "text", + "content": ", and position " + }, + { + "bbox": [ + 313, + 412, + 554, + 508 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 313, + 412, + 554, + 508 + ], + "type": "text", + "content": ", which are essential for accurate surface reconstruction. These geometric properties are typically learned from surrogate image render loss supported by abundant multi-view supervision. When facing sparse-view inputs, the primitives tend to degrade due to insufficient geometric constraints." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 510, + 555, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 555, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 555, + 640 + ], + "type": "text", + "content": "To address this issue, we incorporate direct geometric regularization for 2D Gaussian primitives through multiview consistency guidance. As multi-view consistency approaches only consider point position, we incorporate broad Gaussian primitive properties (i.e., orientation, scale, and position) by correlating their relationship with point position, through a novel reparameterization-based disk sampling strategy. This regularization aligns 2D Gaussian primitives more closely with the latent scene surfaces, leading to improved geometric learning. In the following illustration, we omit the Gaussian primitive indices " + }, + { + "bbox": [ + 313, + 510, + 555, + 640 + ], + "type": "inline_equation", + "content": "i,j" + }, + { + "bbox": [ + 313, + 510, + 555, + 640 + ], + "type": "text", + "content": " for brevity." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "content": "Regularizing points with cross-view consistency. Given 3D point sampling positions " + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{Z} = \\{\\mathbf{z}_k\\}_{k=1}^K" + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "content": " on each Gaussian, we project them to source and target views to obtain the corresponding 2D positions " + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{q}_k^{src}" + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{q}_k^{tar}" + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "content": ", as well as queried MVS features " + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{src}(\\mathbf{q}_k^{src})" + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{tar}(\\mathbf{q}_k^{tar})" + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "content": ". We can then optimize the 3D positions by cross-view feature consistency" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "11310" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 113, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 113, + 84 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 113, + 84 + ], + "type": "text", + "content": "loss [45] " + }, + { + "bbox": [ + 55, + 72, + 113, + 84 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{df}" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 72, + 95, + 294, + 127 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 95, + 294, + 127 + ], + "spans": [ + { + "bbox": [ + 72, + 95, + 294, + 127 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {d f} = \\frac {1}{K} \\sum_ {k = 1} ^ {K} (1 - \\frac {\\mathbf {F} _ {s r c} \\left(\\mathbf {q} _ {k} ^ {s r c}\\right) \\cdot \\mathbf {F} _ {t a r} \\left(\\mathbf {q} _ {k} ^ {t a r}\\right)}{\\| \\mathbf {F} _ {s r c} \\left(\\mathbf {q} _ {k} ^ {s r c}\\right) \\| \\| \\mathbf {F} _ {t a r} \\left(\\mathbf {q} _ {k} ^ {t a r}\\right) \\|}), \\tag {6}", + "image_path": "94b61e49ef862bc75c39d4b1b085d8a513f55cea11fc4aa152e5d5c3a5c55bf6.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 136, + 295, + 159 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 136, + 295, + 159 + ], + "spans": [ + { + "bbox": [ + 55, + 136, + 295, + 159 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 136, + 295, + 159 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 136, + 295, + 159 + ], + "type": "text", + "content": " denotes the number of sampled points on each disk-shaped Gaussian primitive." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 160, + 296, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 160, + 296, + 277 + ], + "spans": [ + { + "bbox": [ + 55, + 160, + 296, + 277 + ], + "type": "text", + "content": "Reparameterization-based disk sampling. The above operation focuses on position regularization only, and thus, it cannot propagate gradients to other Gaussian primitive properties such as orientation " + }, + { + "bbox": [ + 55, + 160, + 296, + 277 + ], + "type": "inline_equation", + "content": "\\mathbf{R}" + }, + { + "bbox": [ + 55, + 160, + 296, + 277 + ], + "type": "text", + "content": " and scale " + }, + { + "bbox": [ + 55, + 160, + 296, + 277 + ], + "type": "inline_equation", + "content": "\\mathbf{S}" + }, + { + "bbox": [ + 55, + 160, + 296, + 277 + ], + "type": "text", + "content": ". We propose a reparameterization-based sampling strategy to solve this problem. As shown in Fig. 3, for each Gaussian primitive, we first sample points " + }, + { + "bbox": [ + 55, + 160, + 296, + 277 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}'" + }, + { + "bbox": [ + 55, + 160, + 296, + 277 + ], + "type": "text", + "content": " from the standard Gaussian distribution " + }, + { + "bbox": [ + 55, + 160, + 296, + 277 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(\\mathbf{0},\\mathbf{I})" + }, + { + "bbox": [ + 55, + 160, + 296, + 277 + ], + "type": "text", + "content": " and then transform the distribution to the target Gaussian primitive through its position " + }, + { + "bbox": [ + 55, + 160, + 296, + 277 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 55, + 160, + 296, + 277 + ], + "type": "text", + "content": ", orientation " + }, + { + "bbox": [ + 55, + 160, + 296, + 277 + ], + "type": "inline_equation", + "content": "\\mathbf{R}" + }, + { + "bbox": [ + 55, + 160, + 296, + 277 + ], + "type": "text", + "content": " and scale " + }, + { + "bbox": [ + 55, + 160, + 296, + 277 + ], + "type": "inline_equation", + "content": "\\mathbf{S}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 145, + 280, + 295, + 293 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 280, + 295, + 293 + ], + "spans": [ + { + "bbox": [ + 145, + 280, + 295, + 293 + ], + "type": "interline_equation", + "content": "\\mathbf {Z} = \\mathbf {p} + \\mathbf {L} \\mathbf {Z} ^ {\\prime}, \\tag {7}", + "image_path": "bff8741e6ac726a26ce11fdcac6909348250a72897dda6f431f61f9750e15193.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 155, + 300, + 294, + 312 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 155, + 300, + 294, + 312 + ], + "spans": [ + { + "bbox": [ + 155, + 300, + 294, + 312 + ], + "type": "interline_equation", + "content": "\\mathbf {L} = \\mathbf {R S}, \\tag {8}", + "image_path": "4590bd42ad453b3e5adfd075265089c5ef2a58149735482731144ed61b4c4aa2.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 318, + 296, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 318, + 296, + 413 + ], + "spans": [ + { + "bbox": [ + 55, + 318, + 296, + 413 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 318, + 296, + 413 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}' \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})" + }, + { + "bbox": [ + 55, + 318, + 296, + 413 + ], + "type": "text", + "content": ". This bridges the gap between Gaussian properties (orientation and scale) and point representation, ensuring regularizing Gaussian geometric properties merely through the surrogate point-based loss. Additionally, since the last scale value of the 2D Gaussian primitive [11] is zero, resulting in difficulties in optimizing the normal vector " + }, + { + "bbox": [ + 55, + 318, + 296, + 413 + ], + "type": "inline_equation", + "content": "\\mathbf{R}[\\dots,2]" + }, + { + "bbox": [ + 55, + 318, + 296, + 413 + ], + "type": "text", + "content": " in the last dimension, we apply additional supervision using the rendered normal " + }, + { + "bbox": [ + 55, + 318, + 296, + 413 + ], + "type": "inline_equation", + "content": "\\mathbf{n}" + }, + { + "bbox": [ + 55, + 318, + 296, + 413 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 124, + 422, + 295, + 437 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 124, + 422, + 295, + 437 + ], + "spans": [ + { + "bbox": [ + 124, + 422, + 295, + 437 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {d n} = 1 - \\mathbf {R} [ \\dots , 2 ] ^ {\\top} \\mathbf {n}, \\tag {9}", + "image_path": "a7ecc46456d429d9312f23a0a2dd82c14f8461569621e0452097dfe2f9b89f44.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 447, + 282, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 447, + 282, + 459 + ], + "spans": [ + { + "bbox": [ + 55, + 447, + 282, + 459 + ], + "type": "text", + "content": "The final primitive regularization loss can be defined as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 136, + 469, + 294, + 482 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 469, + 294, + 482 + ], + "spans": [ + { + "bbox": [ + 136, + 469, + 294, + 482 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {d r} = \\mathcal {L} _ {d f} + \\mathcal {L} _ {d n}. \\tag {10}", + "image_path": "59606971ade1b2cacdc515afad42086f26139d717d0d07cbc0396fcb667aa969.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 491, + 203, + 504 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 491, + 203, + 504 + ], + "spans": [ + { + "bbox": [ + 55, + 491, + 203, + 504 + ], + "type": "text", + "content": "3.4. Selective Gaussian Update" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 510, + 295, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 295, + 617 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 295, + 617 + ], + "type": "text", + "content": "Gaussian Splitting leverages adaptive density control to densify the Gaussian primitives. However, with sparse views, the densified Gaussian primitives are prone to degrade without sufficient multi-view supervision. To address this issue, as shown in Fig. 2, instead of continually densifying the MVS-initialized dense Gaussian primitives with limited supervision, we propose to selectively update them guided by the alternative rendered geometric cues, e.g., depth and normal." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 618, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 296, + 713 + ], + "type": "text", + "content": "Evaluating Gaussian primitives via patch warping. Given any Gaussian primitive with its position " + }, + { + "bbox": [ + 55, + 618, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 55, + 618, + 296, + 713 + ], + "type": "text", + "content": " and orientation " + }, + { + "bbox": [ + 55, + 618, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{R}" + }, + { + "bbox": [ + 55, + 618, + 296, + 713 + ], + "type": "text", + "content": ", we can warp an image patch from one source view to a target view using its position " + }, + { + "bbox": [ + 55, + 618, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 55, + 618, + 296, + 713 + ], + "type": "text", + "content": " and orientation " + }, + { + "bbox": [ + 55, + 618, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{R}" + }, + { + "bbox": [ + 55, + 618, + 296, + 713 + ], + "type": "text", + "content": " together with the pose, following homography warping techniques [3]. Then, we can evaluate the geometric quality of the Gaussian primitive via the patch-based Normalized Cross-Correlation (NCC) [39] metric." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 321, + 87, + 553, + 117 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 87, + 553, + 117 + ], + "spans": [ + { + "bbox": [ + 321, + 87, + 553, + 117 + ], + "type": "interline_equation", + "content": "\\operatorname {N C C} (x, y) = \\frac {\\sum_ {i = l} ^ {n} \\left(x _ {l} - \\bar {x}\\right) \\left(y _ {l} - \\bar {y}\\right)}{\\sqrt {\\sum_ {l = 1} ^ {n} \\left(x _ {l} - \\bar {x}\\right) ^ {2}} \\sqrt {\\sum_ {l = 1} ^ {n} \\left(y _ {l} - \\bar {y}\\right) ^ {2}}}, \\tag {11}", + "image_path": "e9ab009d708e0311906691fe5190cd0c31d4d50c90fdbca412c18f2f391ab7e2.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 120, + 553, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 120, + 553, + 144 + ], + "spans": [ + { + "bbox": [ + 313, + 120, + 553, + 144 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 120, + 553, + 144 + ], + "type": "inline_equation", + "content": "x_{l}" + }, + { + "bbox": [ + 313, + 120, + 553, + 144 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 120, + 553, + 144 + ], + "type": "inline_equation", + "content": "y_{l}" + }, + { + "bbox": [ + 313, + 120, + 553, + 144 + ], + "type": "text", + "content": " represent the pixels in the patches " + }, + { + "bbox": [ + 313, + 120, + 553, + 144 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 120, + 553, + 144 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 120, + 553, + 144 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 313, + 120, + 553, + 144 + ], + "type": "text", + "content": " of the source and target images, respectively." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 144, + 553, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 144, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 313, + 144, + 553, + 228 + ], + "type": "text", + "content": "Update Gaussian primitive with rendered cues. In addition to using the Gaussian primitive parameter " + }, + { + "bbox": [ + 313, + 144, + 553, + 228 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 313, + 144, + 553, + 228 + ], + "type": "text", + "content": " and orientation " + }, + { + "bbox": [ + 313, + 144, + 553, + 228 + ], + "type": "inline_equation", + "content": "\\mathbf{R}" + }, + { + "bbox": [ + 313, + 144, + 553, + 228 + ], + "type": "text", + "content": " for patch warping, we can leverage the rendered pixel-wise depth and normals from tiled Gaussian primitives as an alternative for patch warping. We compute patch-wise NCC over two warping strategies, yielding " + }, + { + "bbox": [ + 313, + 144, + 553, + 228 + ], + "type": "inline_equation", + "content": "\\mathrm{NCC}_G" + }, + { + "bbox": [ + 313, + 144, + 553, + 228 + ], + "type": "text", + "content": " for Gaussian primitive and " + }, + { + "bbox": [ + 313, + 144, + 553, + 228 + ], + "type": "inline_equation", + "content": "\\mathrm{NCC}_R" + }, + { + "bbox": [ + 313, + 144, + 553, + 228 + ], + "type": "text", + "content": " for rendered cues." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 228, + 554, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 228, + 554, + 323 + ], + "spans": [ + { + "bbox": [ + 313, + 228, + 554, + 323 + ], + "type": "text", + "content": "If " + }, + { + "bbox": [ + 313, + 228, + 554, + 323 + ], + "type": "inline_equation", + "content": "\\mathrm{NCC}_R" + }, + { + "bbox": [ + 313, + 228, + 554, + 323 + ], + "type": "text", + "content": " is better than " + }, + { + "bbox": [ + 313, + 228, + 554, + 323 + ], + "type": "inline_equation", + "content": "\\mathrm{NCC}_G" + }, + { + "bbox": [ + 313, + 228, + 554, + 323 + ], + "type": "text", + "content": ", we reproject the rendered depth into a 3D point and update the position of the corresponding Gaussian primitive. We update the position of Gaussian primitives every 100 training steps, after which all Gaussian parameters continue to be optimized by gradients. In this way, we improve the initial Gaussian primitive with geometric cues from rendered results, further enhancing the reconstruction quality." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 314, + 329, + 398, + 342 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 329, + 398, + 342 + ], + "spans": [ + { + "bbox": [ + 314, + 329, + 398, + 342 + ], + "type": "text", + "content": "3.5. Training loss" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 348, + 554, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 348, + 554, + 407 + ], + "spans": [ + { + "bbox": [ + 313, + 348, + 554, + 407 + ], + "type": "text", + "content": "During training, in addition to using the disk regularization loss " + }, + { + "bbox": [ + 313, + 348, + 554, + 407 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{dr}" + }, + { + "bbox": [ + 313, + 348, + 554, + 407 + ], + "type": "text", + "content": " and MVS feature splatting loss " + }, + { + "bbox": [ + 313, + 348, + 554, + 407 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{fea}" + }, + { + "bbox": [ + 313, + 348, + 554, + 407 + ], + "type": "text", + "content": ", we retain the original 2DGS [11] image reconstruction loss " + }, + { + "bbox": [ + 313, + 348, + 554, + 407 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{rgb}" + }, + { + "bbox": [ + 313, + 348, + 554, + 407 + ], + "type": "text", + "content": ", depth-normal consistency loss " + }, + { + "bbox": [ + 313, + 348, + 554, + 407 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_n" + }, + { + "bbox": [ + 313, + 348, + 554, + 407 + ], + "type": "text", + "content": ", and depth distortion loss " + }, + { + "bbox": [ + 313, + 348, + 554, + 407 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_d" + }, + { + "bbox": [ + 313, + 348, + 554, + 407 + ], + "type": "text", + "content": ". The final loss function is defined as:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 328, + 413, + 553, + 427 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 413, + 553, + 427 + ], + "spans": [ + { + "bbox": [ + 328, + 413, + 553, + 427 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {r g b} + \\lambda_ {1} \\mathcal {L} _ {d} + \\lambda_ {2} \\mathcal {L} _ {n}, + \\lambda_ {3} \\mathcal {L} _ {d r} + \\lambda_ {4} \\mathcal {L} _ {f e a}, \\tag {12}", + "image_path": "a79671804e8912fad98b49d6328893bddf733a7c54f513937bd5d19699732832.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 432, + 553, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 432, + 553, + 456 + ], + "spans": [ + { + "bbox": [ + 313, + 432, + 553, + 456 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 432, + 553, + 456 + ], + "type": "inline_equation", + "content": "\\lambda_1, \\lambda_2, \\lambda_3, \\lambda_4" + }, + { + "bbox": [ + 313, + 432, + 553, + 456 + ], + "type": "text", + "content": " denotes the hyperparameters for the weights of each loss term." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 465, + 395, + 479 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 465, + 395, + 479 + ], + "spans": [ + { + "bbox": [ + 313, + 465, + 395, + 479 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 486, + 554, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 486, + 554, + 558 + ], + "spans": [ + { + "bbox": [ + 313, + 486, + 554, + 558 + ], + "type": "text", + "content": "In this section, we first introduce the implementation details and then provide quantitative and qualitative comparisons with current methods on DTU [1] dataset, as well as qualitative results on BlendedMVS [37] dataset. We further analyze the contribution of each module and compare them with different methods through ablation studies." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 563, + 376, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 563, + 376, + 574 + ], + "spans": [ + { + "bbox": [ + 313, + 563, + 376, + 574 + ], + "type": "text", + "content": "4.1. Datasets" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "content": "The DTU [1] dataset consists of 128 scenes captured in controlled lab environments using structured light scanners containing 49 or 64 images. Similar to SparseNeuS [20], we select 15 scenes and use views 23, 24, and 33 for 3-view reconstruction, with an image resolution of " + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "inline_equation", + "content": "576 \\times 768" + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "content": ". We perform the quantitative evaluation on the DTU dataset using the official evaluation code to calculate the Chamfer Distance between the reconstructed mesh and the ground truth point cloud. The BlendedMVS [37] dataset includes several large-scale real-world scenes. We also show the qualitative reconstruction comparisons." + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11311" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 61, + 70, + 550, + 224 + ], + "blocks": [ + { + "bbox": [ + 61, + 70, + 550, + 224 + ], + "lines": [ + { + "bbox": [ + 61, + 70, + 550, + 224 + ], + "spans": [ + { + "bbox": [ + 61, + 70, + 550, + 224 + ], + "type": "table", + "html": "
Scan243740556365698397105106110114118122Mean
Colmap [26]0.902.891.631.082.181.941.611.302.341.281.101.420.761.171.141.52
CLMVSNet [31]1.172.521.680.891.511.820.991.291.520.780.881.100.641.041.121.26
Neus [28]4.574.493.974.324.631.954.683.834.152.501.526.471.265.576.114.00
Volsdf [38]4.034.216.120.918.241.732.741.825.143.092.084.810.603.512.183.41
VolRecon [25]1.202.591.561.081.431.921.111.481.421.051.191.380.741.231.271.38
SparseNeus [20]1.292.271.570.881.611.861.061.271.421.070.990.870.541.151.181.27
Retr [19]1.052.311.440.981.181.520.881.351.300.871.070.770.591.051.121.17
C2F2Neus [34]1.122.421.400.751.411.770.851.161.260.760.910.600.460.880.921.11
Sparsecraft [40]1.171.741.800.701.191.530.831.051.420.780.800.560.440.770.841.04
NeuSurf [12]0.782.351.550.751.041.680.601.140.980.700.740.490.390.750.860.99
2DGS [11]3.143.792.311.515.342.221.852.793.621.633.054.111.053.352.442.81
GOF [43]3.275.052.772.043.842.612.232.653.301.652.584.441.162.292.392.82
PGSR [4]3.363.282.751.275.151.840.881.793.491.191.861.110.611.091.522.08
Ours1.052.351.380.831.371.450.841.161.430.740.850.840.570.951.011.13
", + "image_path": "e013745cc6db1209cab451ef33617c107936fdc06ff0e08af75fd77bf969fe4d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 228, + 555, + 251 + ], + "lines": [ + { + "bbox": [ + 55, + 228, + 555, + 251 + ], + "spans": [ + { + "bbox": [ + 55, + 228, + 555, + 251 + ], + "type": "text", + "content": "Table 1. Quantitative results on DTU dataset. Our method achieves the best performance among Gaussian Splitting surface reconstruction approaches and demonstrates competitive performance to previous state-of-the-art NeRF-based methods." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 74, + 255, + 537, + 541 + ], + "blocks": [ + { + "bbox": [ + 74, + 255, + 537, + 541 + ], + "lines": [ + { + "bbox": [ + 74, + 255, + 537, + 541 + ], + "spans": [ + { + "bbox": [ + 74, + 255, + 537, + 541 + ], + "type": "image", + "image_path": "9ed3e83273f7eec8f44dfc6e88c43c823b51662c3c3d7bb63b0e62738b1a4a80.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 88, + 544, + 520, + 555 + ], + "lines": [ + { + "bbox": [ + 88, + 544, + 520, + 555 + ], + "spans": [ + { + "bbox": [ + 88, + 544, + 520, + 555 + ], + "type": "text", + "content": "Figure 4. DTU surface reconstruction results. Our method achieves more complete reconstructions with finer details." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 563, + 188, + 576 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 563, + 188, + 576 + ], + "spans": [ + { + "bbox": [ + 55, + 563, + 188, + 576 + ], + "type": "text", + "content": "4.2. Implementation Details" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "content": "Our method is implemented based on 2DGS [11]. For Gaussian primitive initialization, we use the point cloud back-projected depth maps from CLMVSNet [31]. Instead of using spherical harmonics to represent the color of the Gaussian primitives, we initialize the color using the RGB values from the image. CLMVSNet uses an FPN for feature extraction, and we utilize the full-resolution feature maps with a channel dimension of 8 from the FPN last layer for feature splatting. The remaining parameters are initialized consistently with those in 2DGS. We set the training iter" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 564, + 555, + 660 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 564, + 555, + 660 + ], + "spans": [ + { + "bbox": [ + 313, + 564, + 555, + 660 + ], + "type": "text", + "content": "ation to 7000 and replaced the original 2DGS's adaptive density control with our proposed Gaussian update strategy. Similar to 2DGS [11], we used truncated signed distance fusion (TSDF) to extract the final mesh. During TSDF fusion, the voxel size and truncated threshold are set to 0.004 and 0.02 respectively. We set the hyperparameters of the loss function, " + }, + { + "bbox": [ + 313, + 564, + 555, + 660 + ], + "type": "inline_equation", + "content": "\\lambda_{1},\\lambda_{2},\\lambda_{3}" + }, + { + "bbox": [ + 313, + 564, + 555, + 660 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 564, + 555, + 660 + ], + "type": "inline_equation", + "content": "\\lambda_{4}" + }, + { + "bbox": [ + 313, + 564, + 555, + 660 + ], + "type": "text", + "content": ", to 1000, 0.05, 1, and 0.2, respectively." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 669, + 457, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 669, + 457, + 681 + ], + "spans": [ + { + "bbox": [ + 313, + 669, + 457, + 681 + ], + "type": "text", + "content": "4.3. Benchmark Comparisons" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "content": "We compare the sparse-view reconstruction performance of different NeRF-based and Gaussian-based methods on" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11312" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 65, + 74, + 287, + 319 + ], + "blocks": [ + { + "bbox": [ + 65, + 74, + 287, + 319 + ], + "lines": [ + { + "bbox": [ + 65, + 74, + 287, + 319 + ], + "spans": [ + { + "bbox": [ + 65, + 74, + 287, + 319 + ], + "type": "image", + "image_path": "444bf781e02a8f027249643b14b295cd4a1c74977a9ce4dca0e4f315b875e77a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 328, + 295, + 350 + ], + "lines": [ + { + "bbox": [ + 55, + 328, + 295, + 350 + ], + "spans": [ + { + "bbox": [ + 55, + 328, + 295, + 350 + ], + "type": "text", + "content": "Figure 5. BlendedMVS reconstruction results. Our method achieves more complete and detailed reconstructions." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 58, + 355, + 294, + 454 + ], + "blocks": [ + { + "bbox": [ + 58, + 355, + 294, + 454 + ], + "lines": [ + { + "bbox": [ + 58, + 355, + 294, + 454 + ], + "spans": [ + { + "bbox": [ + 58, + 355, + 294, + 454 + ], + "type": "table", + "html": "
2DGSMVS Init.Fixed ColorFeat. Splat.DGPRSGUCD ↓
(a)✓2.813
(b)✓✓1.273
(c)✓✓✓1.212
(d)✓✓✓1.224
(e)✓✓✓✓1.201
(f)✓✓✓✓✓1.142
(g)✓✓✓✓✓1.152
(h)✓✓✓✓✓✓1.125
", + "image_path": "7c19cf3d27ff7a49df0c9e2ffa852823eb7047f401e402dc0ab97335c07fcd96.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 521, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 295, + 713 + ], + "type": "text", + "content": "the DTU dataset. As shown in Tab. 1, our method significantly improves upon Gaussian-based approaches under sparse views, outperforming baseline (2DGS) and previous Gaussian-based SOTA methods (PGSR) by " + }, + { + "bbox": [ + 55, + 521, + 295, + 713 + ], + "type": "inline_equation", + "content": "67\\%" + }, + { + "bbox": [ + 55, + 521, + 295, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 521, + 295, + 713 + ], + "type": "inline_equation", + "content": "46\\%" + }, + { + "bbox": [ + 55, + 521, + 295, + 713 + ], + "type": "text", + "content": ", respectively. Our method also outperforms SparseNeus [20], the first sparse-view NeRF-based method with notable margins. Our method performs competitively with the state-of-the-art NeRF-based surface reconstruction methods. The visualization results are shown in Fig. 4. Our method captures more surface details than implicit reconstruction approaches and outperforms Gaussian-based methods in preserving the complete scene structure. In comparison, Gaussian Splitting methods, i.e., 2DGS, GOF, and PGSR, tend to exhibit noise and missing areas. Additionally, we performed a qualitative comparison on the BlendedMVS dataset. As shown in Fig. 5, our method" + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 317, + 70, + 553, + 118 + ], + "blocks": [ + { + "bbox": [ + 55, + 463, + 295, + 508 + ], + "lines": [ + { + "bbox": [ + 55, + 463, + 295, + 508 + ], + "spans": [ + { + "bbox": [ + 55, + 463, + 295, + 508 + ], + "type": "text", + "content": "Table 2. Ablation on DTU [1] dataset. DGPR and SGU represent proposed Direct Gaussian Primitive Regularization and Selective Gaussian Update respectively. Each module improves upon the baseline, leading to the best final results." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 70, + 553, + 118 + ], + "lines": [ + { + "bbox": [ + 317, + 70, + 553, + 118 + ], + "spans": [ + { + "bbox": [ + 317, + 70, + 553, + 118 + ], + "type": "table", + "html": "
Methods2DGS [11]PGSR [4]CLMVS. [31]Ours
3 view SfM Point Cloud2.812.08--
49/64 views SfM Point Cloud1.651.53--
CLMVSNet Point Cloud1.271.381.261.13
", + "image_path": "4be6d13706e42ef75ce6f84e7688b1ef8320fbf89c8ac71855838743893c8590.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 123, + 555, + 190 + ], + "lines": [ + { + "bbox": [ + 313, + 123, + 555, + 190 + ], + "spans": [ + { + "bbox": [ + 313, + 123, + 555, + 190 + ], + "type": "text", + "content": "Table 3. Comparison of initialization strategies. Initializing with MVS point cloud improves reconstruction performance. However, a plain combination (Chamfer Distance, CD:1.27) between 2DGS and CLMVSNet fails to surpass CLMVSNet results (CD: 1.26). Our method enhances both MVS and 2DGS through geometry-prioritized optimization." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 315, + 204, + 544, + 380 + ], + "blocks": [ + { + "bbox": [ + 315, + 204, + 544, + 380 + ], + "lines": [ + { + "bbox": [ + 315, + 204, + 544, + 380 + ], + "spans": [ + { + "bbox": [ + 315, + 204, + 544, + 380 + ], + "type": "image", + "image_path": "2af1b2913544c4ea42f028ee04b1610459dff33d8d7f5579ad35977d2f9b96de.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 385, + 555, + 441 + ], + "lines": [ + { + "bbox": [ + 313, + 385, + 555, + 441 + ], + "spans": [ + { + "bbox": [ + 313, + 385, + 555, + 441 + ], + "type": "text", + "content": "Figure 6. Comparison between different initialization strategies of 2DGS and our method. Initializing 2DGS with MVS point cloud improves SfM reconstruction completeness but struggles to capture details. Our method not only enhances completeness but also achieves detailed reconstruction." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 448, + 555, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 448, + 555, + 471 + ], + "spans": [ + { + "bbox": [ + 313, + 448, + 555, + 471 + ], + "type": "text", + "content": "achieves higher accuracy in reconstructing meshes for real-world scenes." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 479, + 446, + 493 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 479, + 446, + 493 + ], + "spans": [ + { + "bbox": [ + 313, + 479, + 446, + 493 + ], + "type": "text", + "content": "4.4. Ablations and Analysis" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 312, + 498, + 555, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 498, + 555, + 641 + ], + "spans": [ + { + "bbox": [ + 312, + 498, + 555, + 641 + ], + "type": "text", + "content": "In this section, we perform ablation studies to analyze the effectiveness of different modules. As shown in Tab. 2, using MVS point cloud initialization notably enhances the performance (row a). By fixing the colors and using feature splatting for geometric enhancement supervision, we further improve the accuracy and completeness of the reconstruction (row e). The Direct Gaussian Primitive Regularization (DGPR) enhances the surface modeling capability of each Gaussian primitive (row f). The Selective Gaussian Update (SGU) leverages rendering cues to enhance the precision of the Gaussian point set, further improving surface accuracy (row g)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "content": "Comparison of different initialization strategies. We compare the reconstruction results using various initialization strategies for Gaussian Splatting methods. We show the results using 3 different initialization strategies: SfM point cloud from 3 viewpoints, SfM point cloud from 49 viewpoints, and the CLMVSNet reconstructed point cloud." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11313" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 67, + 70, + 284, + 126 + ], + "blocks": [ + { + "bbox": [ + 67, + 70, + 284, + 126 + ], + "lines": [ + { + "bbox": [ + 67, + 70, + 284, + 126 + ], + "spans": [ + { + "bbox": [ + 67, + 70, + 284, + 126 + ], + "type": "table", + "html": "
SettingsAccuracy↓Completion↓Average↓
SH=30.831.711.27
SH=20.841.681.26
SH=10.841.641.24
Fixed Color0.811.611.21
", + "image_path": "f5eed4349c0e3d755d6fabeb2a149a38dbedbe38b4c9e17afc95f21bceced6ca.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 57, + 156, + 298, + 205 + ], + "blocks": [ + { + "bbox": [ + 55, + 128, + 295, + 151 + ], + "lines": [ + { + "bbox": [ + 55, + 128, + 295, + 151 + ], + "spans": [ + { + "bbox": [ + 55, + 128, + 295, + 151 + ], + "type": "text", + "content": "Table 4. Ablation study on appearance learning. Fixing the color helps in learning better geometry." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 156, + 298, + 205 + ], + "lines": [ + { + "bbox": [ + 57, + 156, + 298, + 205 + ], + "spans": [ + { + "bbox": [ + 57, + 156, + 298, + 205 + ], + "type": "table", + "html": "
MethodsAccuracy↓Completeness↓Average↓
Baseline0.7901.6121.201
Adaptive Density Control [14]0.8141.6541.234
Selective Gaussian Update0.7601.5441.152
", + "image_path": "2a8bdd83aae43113c33d60d8860183250fd987a0b53a876652d7693dce8d2298.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 58, + 258, + 287, + 332 + ], + "blocks": [ + { + "bbox": [ + 58, + 258, + 287, + 332 + ], + "lines": [ + { + "bbox": [ + 58, + 258, + 287, + 332 + ], + "spans": [ + { + "bbox": [ + 58, + 258, + 287, + 332 + ], + "type": "image", + "image_path": "36f83582a80326fac4c415fb033f74b859b43c5d3033abbba332d76da8f0a613.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 335, + 295, + 368 + ], + "lines": [ + { + "bbox": [ + 55, + 335, + 295, + 368 + ], + "spans": [ + { + "bbox": [ + 55, + 335, + 295, + 368 + ], + "type": "text", + "content": "Figure 7. Visualization of MVS feature splatting. Incorporating MVS feature splatting aids in achieving higher reconstruction accuracy in challenging regions." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 376, + 295, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 376, + 295, + 496 + ], + "spans": [ + { + "bbox": [ + 54, + 376, + 295, + 496 + ], + "type": "text", + "content": "Using MVS point clouds achieves a more complete reconstruction than SfM. Note that as shown in Tab. 3, due to the inherent ill-posedness in sparse-view reconstruction, the plain combination of MVS and 2DGS reconstruction yields suboptimal performance. Even with CLMVSNet point clouds as initialization, the reconstruction quality still falls short of that achieved by CLMVSNet itself. In contrast, our geometric-prioritized enhancement addresses this challenge, with results excelling both CLMVSNet and 2DGS with higher completeness and finer details (Fig. 6)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 497, + 295, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 497, + 295, + 700 + ], + "spans": [ + { + "bbox": [ + 55, + 497, + 295, + 700 + ], + "type": "text", + "content": "Geometrically enhanced supervision. In this paper, we mitigate appearance overfitting by fixing the color and feature parameters during optimization. In the original Gaussian splatting, view-independent color is represented by spherical harmonics (SH) with learnable coefficients. Intuitively, reducing SH degrees can achieve a similar goal in alleviating color overfitting. We compare the fixed color and reduced SH degrees in Tab. 4. Results show that reducing the SH degrees also mitigates the degradation in geometry, with only 1 SH degree outperforming other settings (i.e., 2, 3) in surface reconstruction. However, using learnable color coefficients still performs worse than the adopted choice of fixed color during training. Moreover, we enhance geometric supervision using MVS feature splatting. As shown in Fig. 7, MVS features are more expressive and provide more effective constraints, especially in challenging regions such as repetitive texture areas (e.g., the roof of the building)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 701, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 701, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 701, + 295, + 714 + ], + "type": "text", + "content": "Ablation on Direct Gaussian Primitive Regularization." + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 317, + 70, + 554, + 128 + ], + "blocks": [ + { + "bbox": [ + 55, + 206, + 295, + 251 + ], + "lines": [ + { + "bbox": [ + 55, + 206, + 295, + 251 + ], + "spans": [ + { + "bbox": [ + 55, + 206, + 295, + 251 + ], + "type": "text", + "content": "Table 5. Ablation on DTU [1] dataset. The proposed Selective Gaussian Update strategy enhances reconstruction accuracy and completeness compared to the generic Adaptive Density Control in sparse views." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 70, + 554, + 128 + ], + "lines": [ + { + "bbox": [ + 317, + 70, + 554, + 128 + ], + "spans": [ + { + "bbox": [ + 317, + 70, + 554, + 128 + ], + "type": "table", + "html": "
SettingsAccuracy↓Completion↓Average↓
K=90.7461.5441.145
K=250.7471.5411.144
K=490.7511.5321.142
K=810.7471.5311.139
", + "image_path": "33b552213e1fd606aec1278c91c8866d33c371ae4967a66d4c1cf823a5f18833.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 322, + 167, + 548, + 210 + ], + "blocks": [ + { + "bbox": [ + 313, + 129, + 553, + 162 + ], + "lines": [ + { + "bbox": [ + 313, + 129, + 553, + 162 + ], + "spans": [ + { + "bbox": [ + 313, + 129, + 553, + 162 + ], + "type": "text", + "content": "Table 6. Comparison of different numbers of sampled points. Increasing the sampling numbers lead to a slight performance improvement." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 322, + 167, + 548, + 210 + ], + "lines": [ + { + "bbox": [ + 322, + 167, + 548, + 210 + ], + "spans": [ + { + "bbox": [ + 322, + 167, + 548, + 210 + ], + "type": "table", + "html": "
MethodsSparseNeus [20]NeuSurf [12]Ours
Chamfer Distance1.270.991.13
Training Time25mins10 hours10mins
", + "image_path": "c5790e3bedd394bda8389b46119c0cbdda3a83444efed989df7598127b698559.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 213, + 553, + 236 + ], + "lines": [ + { + "bbox": [ + 313, + 213, + 553, + 236 + ], + "spans": [ + { + "bbox": [ + 313, + 213, + 553, + 236 + ], + "type": "text", + "content": "Table 7. Performance and efficiency. Our method achieves comparable performance while requiring the least training time." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 243, + 553, + 290 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 243, + 553, + 290 + ], + "spans": [ + { + "bbox": [ + 313, + 243, + 553, + 290 + ], + "type": "text", + "content": "We analyze the impact of the reparameterization-based disk sampling number " + }, + { + "bbox": [ + 313, + 243, + 553, + 290 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 243, + 553, + 290 + ], + "type": "text", + "content": " on the final results. As shown in Tab. 6, increasing the number of sampled points can slightly enhance the final reconstruction performance." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 290, + 554, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 290, + 554, + 434 + ], + "spans": [ + { + "bbox": [ + 313, + 290, + 554, + 434 + ], + "type": "text", + "content": "Selective Gaussian update strategy. To validate the effectiveness of our selective update strategy, we compared it with the adaptive density control in the original 2DGS. Quantitative results are shown in the Tab. 5. Due to the lack of sufficient multi-view supervision, the densified Gaussian primitives are prone to degradation, leading to declined reconstruction performance (CD:1.23) compared to the baseline (CD: 1.20). The proposed selective Gaussian update strategy enhances the accuracy of the original Gaussian point set by utilizing high-precision points from the rendered cues, which further improves the surface reconstruction quality." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 434, + 554, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 434, + 554, + 518 + ], + "spans": [ + { + "bbox": [ + 313, + 434, + 554, + 518 + ], + "type": "text", + "content": "Performance and efficiency. We compared reconstruction performance and training time with previous state-of-the-art surface reconstruction methods. As shown in Tab. 7, our method achieves better performance and " + }, + { + "bbox": [ + 313, + 434, + 554, + 518 + ], + "type": "inline_equation", + "content": "2 \\times" + }, + { + "bbox": [ + 313, + 434, + 554, + 518 + ], + "type": "text", + "content": " faster training speed than SparseNeus [20]. Meanwhile, our method achieves competitive performance while being " + }, + { + "bbox": [ + 313, + 434, + 554, + 518 + ], + "type": "inline_equation", + "content": "60 \\times" + }, + { + "bbox": [ + 313, + 434, + 554, + 518 + ], + "type": "text", + "content": " faster than NeuSurf [12]." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 529, + 388, + 542 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 529, + 388, + 542 + ], + "spans": [ + { + "bbox": [ + 313, + 529, + 388, + 542 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 550, + 554, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 550, + 554, + 668 + ], + "spans": [ + { + "bbox": [ + 313, + 550, + 554, + 668 + ], + "type": "text", + "content": "In this paper, we propose Sparse2DGS, addressing the challenge of ill-posed geometry optimization from sparse views through geometry-prioritized enhancement schemes. Our method not only reconstructs accurate and complete scene surfaces but also achieves higher efficiency than NeRF-based methods. Limitation: Despite their effectiveness in dense matching, MVS methods inevitably encounter degradation under occlusions. The error in occluded areas can propagate to our method, leading to potential performance declines." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 669, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 669, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 669, + 553, + 713 + ], + "type": "text", + "content": "Acknowledgements Y. Zhang was supported by NSFC (No.U19B2037) and the Natural Science Basic Research Program of Shaanxi (No.2021JCW-03). Y. Zhu was supported by NSFC (No.61901384)." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "11314" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 78, + 115, + 91 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 78, + 115, + 91 + ], + "spans": [ + { + "bbox": [ + 56, + 78, + 115, + 91 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 98, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 98, + 296, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 98, + 296, + 140 + ], + "spans": [ + { + "bbox": [ + 61, + 98, + 296, + 140 + ], + "type": "text", + "content": "[1] Henrik Aanaes, Rasmus Ramsbøl Jensen, George Vogiatzis, Engin Tola, and Anders Bjorholm Dahl. Large-scale data for multiple-view stereopsis. International Journal of Computer Vision, 120:153-168, 2016. 5, 7, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 142, + 296, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 142, + 296, + 185 + ], + "spans": [ + { + "bbox": [ + 61, + 142, + 296, + 185 + ], + "type": "text", + "content": "[2] Chenjie Cao, Xinlin Ren, and Yanwei Fu. Mvsformer: Learning robust image representations via transformers and temperature-based depth for multi-view stereo. arXiv preprint arXiv:2208.02541, 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 186, + 296, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 186, + 296, + 240 + ], + "spans": [ + { + "bbox": [ + 62, + 186, + 296, + 240 + ], + "type": "text", + "content": "[3] Di Chang, Aljaž Božić, Tong Zhang, Qingsong Yan, Yingcong Chen, Sabine Susstrunk, and Matthias Nießner. Rc-mvsnet: Unsupervised multi-view stereo with neural rendering. In European conference on computer vision, pages 665-680. Springer, 2022. 2, 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 241, + 295, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 241, + 295, + 296 + ], + "spans": [ + { + "bbox": [ + 62, + 241, + 295, + 296 + ], + "type": "text", + "content": "[4] Danpeng Chen, Hai Li, Weicai Ye, Yifan Wang, Weijian Xie, Shangjin Zhai, Nan Wang, Haomin Liu, Hujun Bao, and Guofeng Zhang. Pgsr: Planar-based gaussian splatting for efficient and high-fidelity surface reconstruction. arXiv preprint arXiv:2406.06521, 2024. 1, 2, 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 297, + 295, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 297, + 295, + 328 + ], + "spans": [ + { + "bbox": [ + 62, + 297, + 295, + 328 + ], + "type": "text", + "content": "[5] Hanlin Chen, Chen Li, and Gim Hee Lee. Neusg: Neural implicit surface reconstruction with 3d gaussian splatting guidance. arXiv preprint arXiv:2312.00846, 2023. 1, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 329, + 295, + 393 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 329, + 295, + 393 + ], + "spans": [ + { + "bbox": [ + 62, + 329, + 295, + 393 + ], + "type": "text", + "content": "[6] Yikang Ding, Wentao Yuan, Qingtian Zhu, Haotian Zhang, Xiangyue Liu, Yuanjiang Wang, and Xiao Liu. Transmvsnet: Global context-aware multi-view stereo network with transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8585-8594, 2022. 1, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 395, + 295, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 395, + 295, + 437 + ], + "spans": [ + { + "bbox": [ + 62, + 395, + 295, + 437 + ], + "type": "text", + "content": "[7] Lue Fan, Yuxue Yang, Minxing Li, Hongsheng Li, and Zhaoxiang Zhang. Trim 3d gaussian splatting for accurate geometry representation. arXiv preprint arXiv:2406.07499, 2024.1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 439, + 295, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 439, + 295, + 482 + ], + "spans": [ + { + "bbox": [ + 62, + 439, + 295, + 482 + ], + "type": "text", + "content": "[8] Qiancheng Fu, Qingshan Xu, Yew Soon Ong, and Wenbing Tao. Geo-neus: Geometry-consistent neural implicit surfaces learning for multi-view reconstruction. Advances in Neural Information Processing Systems, 35:3403-3416, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 483, + 295, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 483, + 295, + 525 + ], + "spans": [ + { + "bbox": [ + 62, + 483, + 295, + 525 + ], + "type": "text", + "content": "[9] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In 2012 IEEE conference on computer vision and pattern recognition, pages 3354-3361. IEEE, 2012. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 526, + 296, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 526, + 296, + 581 + ], + "spans": [ + { + "bbox": [ + 57, + 526, + 296, + 581 + ], + "type": "text", + "content": "[10] Antoine Guédon and Vincent Lepetit. Sugar: Surface-aligned gaussian splatting for efficient 3d mesh reconstruction and high-quality mesh rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5354-5363, 2024. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 582, + 295, + 625 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 582, + 295, + 625 + ], + "spans": [ + { + "bbox": [ + 57, + 582, + 295, + 625 + ], + "type": "text", + "content": "[11] Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2d gaussian splatting for geometrically accurate radiance fields. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 1, 2, 3, 4, 5, 6, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 625, + 295, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 625, + 295, + 679 + ], + "spans": [ + { + "bbox": [ + 57, + 625, + 295, + 679 + ], + "type": "text", + "content": "[12] Han Huang, Yulun Wu, Junsheng Zhou, Ge Gao, Ming Gu, and Yu-Shen Liu. Neusurf: On-surface priors for neural surface reconstruction from sparse input views. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2312-2320, 2024. 2, 3, 6, 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 680, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 680, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 680, + 295, + 713 + ], + "type": "text", + "content": "[13] Michael Kazhdan, Matthew Bolitho, and Hugues Hoppe. Poisson surface reconstruction. In Proceedings of the fourth Eurographics symposium on Geometry processing, 2006. 2" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 73, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "text", + "content": "[14] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4):139-1, 2023. 1, 2, 4, 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 118, + 553, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 118, + 553, + 171 + ], + "spans": [ + { + "bbox": [ + 317, + 118, + 553, + 171 + ], + "type": "text", + "content": "[15] Tejas Khot, Shubham Agrawal, Shubham Tulsiani, Christoph Mertz, Simon Lucey, and Martial Hebert. Learning unsupervised multi-view stereopsis via robust photometric consistency. arXiv preprint arXiv:1905.02706, 2019. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 172, + 555, + 238 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 172, + 555, + 238 + ], + "spans": [ + { + "bbox": [ + 317, + 172, + 555, + 238 + ], + "type": "text", + "content": "[16] Rui Li, Dong Gong, Wei Yin, Hao Chen, Yu Zhu, Kaixuan Wang, Xiaozhi Chen, Jinqiu Sun, and Yanning Zhang. Learning to fuse monocular and multi-view cues for multiframe depth estimation in dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21539-21548, 2023. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 239, + 553, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 239, + 553, + 303 + ], + "spans": [ + { + "bbox": [ + 316, + 239, + 553, + 303 + ], + "type": "text", + "content": "[17] Rui Li, Tobias Fischer, Mattia Segu, Marc Pollefeys, Luc Van Gool, and Federico Tombari. Know your neighbors: Improving single-view reconstruction via spatial vision-language reasoning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9848-9858, 2024. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 304, + 553, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 304, + 553, + 360 + ], + "spans": [ + { + "bbox": [ + 316, + 304, + 553, + 360 + ], + "type": "text", + "content": "[18] Yinhao Li, Zheng Ge, Guanyi Yu, Jinrong Yang, Zengran Wang, Yukang Shi, Jianjian Sun, and Zeming Li. Bevdepth: Acquisition of reliable depth for multi-view 3d object detection. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1477-1485, 2023. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 361, + 553, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 361, + 553, + 403 + ], + "spans": [ + { + "bbox": [ + 316, + 361, + 553, + 403 + ], + "type": "text", + "content": "[19] Yixun Liang, Hao He, and Yingcong Chen. Retr: Modeling rendering via transformer for generalizable neural surface reconstruction. Advances in Neural Information Processing Systems, 36, 2024. 1, 3, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 404, + 553, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 404, + 553, + 458 + ], + "spans": [ + { + "bbox": [ + 316, + 404, + 553, + 458 + ], + "type": "text", + "content": "[20] Xiaoxiao Long, Cheng Lin, Peng Wang, Taku Komura, and Wenping Wang. Sparseneus: Fast generalizable neural surface reconstruction from sparse views. In European Conference on Computer Vision, pages 210-227. Springer, 2022. 1, 2, 5, 6, 7, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 460, + 553, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 460, + 553, + 492 + ], + "spans": [ + { + "bbox": [ + 316, + 460, + 553, + 492 + ], + "type": "text", + "content": "[21] Xuan Luo, Jia-Bin Huang, Richard Szeliski, Kevin Matzen, and Johannes Kopf. Consistent video depth estimation. ACM Transactions on Graphics (ToG), 39(4):71-1, 2020. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 493, + 553, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 493, + 553, + 547 + ], + "spans": [ + { + "bbox": [ + 316, + 493, + 553, + 547 + ], + "type": "text", + "content": "[22] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 1, 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 548, + 553, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 548, + 553, + 602 + ], + "spans": [ + { + "bbox": [ + 316, + 548, + 553, + 602 + ], + "type": "text", + "content": "[23] Michael Oechsle, Songyou Peng, and Andreas Geiger. Unisurf: Unifying neural implicit surfaces and radiance fields for multi-view reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5589-5599, 2021. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 315, + 603, + 553, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 603, + 553, + 645 + ], + "spans": [ + { + "bbox": [ + 315, + 603, + 553, + 645 + ], + "type": "text", + "content": "[24] Ri-Zhao Qiu, Ge Yang, Weijia Zeng, and Xiaolong Wang. Feature splatting: Language-driven physics-based scene synthesis and editing. arXiv preprint arXiv:2404.01223, 2024. 4" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 647, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 712 + ], + "type": "text", + "content": "[25] Yufan Ren, Fangjinhua Wang, Tong Zhang, Marc Pollefeys, and Sabine Susstrunk. Volrecon: Volume rendering of signed ray distance functions for generalizable multi-view reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16685-16695, 2023. 1, 2, 3, 6" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11315" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "text", + "content": "[26] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4104-4113, 2016. 3, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 171 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 171 + ], + "type": "text", + "content": "[27] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Direct voxel grid optimization: Super-fast convergence for radiance fields reconstruction. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5459-5469, 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 172, + 294, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 172, + 294, + 216 + ], + "spans": [ + { + "bbox": [ + 56, + 172, + 294, + 216 + ], + "type": "text", + "content": "[28] Peng Wang, Lingjie Liu, Yuan Liu, Christian Theobalt, Taku Komura, and Wenping Wang. Neus: Learning neural implicit surfaces by volume rendering for multi-view reconstruction. arXiv preprint arXiv:2106.10689, 2021. 1, 2, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 217, + 294, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 217, + 294, + 239 + ], + "spans": [ + { + "bbox": [ + 56, + 217, + 294, + 239 + ], + "type": "text", + "content": "[29] Yiqun Wang, Ivan Skorokhodov, and Peter Wonka. Improved surface reconstruction using high-frequency details. 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 239, + 294, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 239, + 294, + 293 + ], + "spans": [ + { + "bbox": [ + 56, + 239, + 294, + 293 + ], + "type": "text", + "content": "[30] Jiang Wu, Rui Li, Haofei Xu, Wenxun Zhao, Yu Zhu, Jinqiu Sun, and Yanning Zhang. Gomvs: Geometrically consistent cost aggregation for multi-view stereo. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20207-20216, 2024. 1, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 294, + 294, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 294, + 294, + 358 + ], + "spans": [ + { + "bbox": [ + 56, + 294, + 294, + 358 + ], + "type": "text", + "content": "[31] Kaiqiang Xiong, Rui Peng, Zhe Zhang, Tianxing Feng, Jianbo Jiao, Feng Gao, and Ronggang Wang. Cl-mvsnet: Unsupervised multi-view stereo with dual-level contrastive learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3769-3780, 2023. 1, 2, 3, 4, 6, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 360, + 294, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 360, + 294, + 415 + ], + "spans": [ + { + "bbox": [ + 56, + 360, + 294, + 415 + ], + "type": "text", + "content": "[32] Jialei Xu, Xianming Liu, Yuanchao Bai, Junjun Jiang, Kaixuan Wang, Xiaozhi Chen, and Xiangyang Ji. Multi-camera collaborative depth prediction via consistent structure estimation. In Proceedings of the 30th ACM international conference on multimedia, pages 2730-2738, 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 415, + 294, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 415, + 294, + 469 + ], + "spans": [ + { + "bbox": [ + 56, + 415, + 294, + 469 + ], + "type": "text", + "content": "[33] Jialei Xu, Wei Yin, Dong Gong, Junjun Jiang, and Xianming Liu. Sdge: Stereo guided depth estimation for " + }, + { + "bbox": [ + 56, + 415, + 294, + 469 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 56, + 415, + 294, + 469 + ], + "type": "text", + "content": " camera sets. In 2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 11179-11186. IEEE, 2024. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 471, + 294, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 471, + 294, + 536 + ], + "spans": [ + { + "bbox": [ + 56, + 471, + 294, + 536 + ], + "type": "text", + "content": "[34] Luoyuan Xu, Tao Guan, Yuesong Wang, Wenkai Liu, Zhaojie Zeng, Junle Wang, and Wei Yang. C2f2neus: Cascade cost frustum fusion for high fidelity and generalizable neural surface reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 18291-18301, 2023. 1, 3, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 537, + 294, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 537, + 294, + 590 + ], + "spans": [ + { + "bbox": [ + 56, + 537, + 294, + 590 + ], + "type": "text", + "content": "[35] Qingshan Xu, Wanjuan Su, Yuhang Qi, Wenbing Tao, and Marc Pollefeys. Learning inverse depth regression for pixelwise visibility-aware multi-view stereo networks. International Journal of Computer Vision, 130(8):2040-2059, 2022. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 592, + 294, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 592, + 294, + 635 + ], + "spans": [ + { + "bbox": [ + 56, + 592, + 294, + 635 + ], + "type": "text", + "content": "[36] Yao Yao, Zixin Luo, Shiwei Li, Tian Fang, and Long Quan. Mvsnet: Depth inference for unstructured multi-view stereo. In Proceedings of the European conference on computer vision (ECCV), pages 767-783, 2018. 1, 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 636, + 294, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 636, + 294, + 690 + ], + "spans": [ + { + "bbox": [ + 56, + 636, + 294, + 690 + ], + "type": "text", + "content": "[37] Yao Yao, Zixin Luo, Shiwei Li, Jingyang Zhang, Yufan Ren, Lei Zhou, Tian Fang, and Long Quan. Blendedmvs: A large-scale dataset for generalized multi-view stereo networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1790-1799, 2020. 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "text", + "content": "[38] Lior Yariv, Jiatao Gu, Yoni Kasten, and Yaron Lipman. Volume rendering of neural implicit surfaces. Advances in Neu" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 521 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "ral Information Processing Systems, 34:4805-4815, 2021. 1, 2, 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 96, + 553, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 128 + ], + "type": "text", + "content": "[39] Jae-Chern Yoo and Tae Hee Han. Fast normalized cross-correlation. Circuits, systems and signal processing, 28:819-843, 2009. 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 129, + 553, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 553, + 174 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 553, + 174 + ], + "type": "text", + "content": "[40] Mae Younes, Amine Ouasfi, and Adnane Boukhayma. Sparsecraft: Few-shot neural reconstruction through stereopsis guided geometric linearization. In European Conference on Computer Vision, pages 37-56. Springer, 2024. 3, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 175, + 553, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 553, + 217 + ], + "type": "text", + "content": "[41] Mulin Yu, Tao Lu, Linning Xu, Lihan Jiang, Yuanbo Xiangli, and Bo Dai. Gsdf: 3dgs meets sdf for improved rendering and reconstruction. arXiv preprint arXiv:2403.16964, 2024. 1, 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 220, + 553, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 220, + 553, + 274 + ], + "spans": [ + { + "bbox": [ + 316, + 220, + 553, + 274 + ], + "type": "text", + "content": "[42] Zehao Yu, Songyou Peng, Michael Niemeyer, Torsten Sattler, and Andreas Geiger. Monosdf: Exploring monocular geometric cues for neural implicit surface reconstruction. Advances in neural information processing systems, 35:25018-25032, 2022. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 275, + 553, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 275, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 316, + 275, + 553, + 319 + ], + "type": "text", + "content": "[43] Zehao Yu, Torsten Sattler, and Andreas Geiger. Gaussian opacity fields: Efficient and compact surface reconstruction in unbounded scenes. arXiv preprint arXiv:2404.10772, 2024. 1, 2, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 320, + 553, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 320, + 553, + 363 + ], + "spans": [ + { + "bbox": [ + 316, + 320, + 553, + 363 + ], + "type": "text", + "content": "[44] Baowen Zhang, Chuan Fang, Rakesh Shrestha, Yixun Liang, Xiaoxiao Long, and Ping Tan. Rade-gs: Rasterizing depth in gaussian splatting. arXiv preprint arXiv:2406.01467, 2024. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 365, + 553, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 365, + 553, + 409 + ], + "spans": [ + { + "bbox": [ + 316, + 365, + 553, + 409 + ], + "type": "text", + "content": "[45] Jingyang Zhang, Yao Yao, and Long Quan. Learning signed distance field for multi-view surface reconstruction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6525-6534, 2021. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 410, + 553, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 410, + 553, + 453 + ], + "spans": [ + { + "bbox": [ + 316, + 410, + 553, + 453 + ], + "type": "text", + "content": "[46] Jingyang Zhang, Shiwei Li, Zixin Luo, Tian Fang, and Yao Yao. Vis-mvsnet: Visibility-aware multi-view stereo network. International Journal of Computer Vision, 131(1): 199-214, 2023. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 454, + 553, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 454, + 553, + 521 + ], + "spans": [ + { + "bbox": [ + 316, + 454, + 553, + 521 + ], + "type": "text", + "content": "[47] Shijie Zhou, Haoran Chang, Sicheng Jiang, Zhiwen Fan, Zehao Zhu, Dejia Xu, Pradyumna Chari, Suya You, Zhangyang Wang, and Achuta Kadambi. Feature 3dgs: Supercharging 3d gaussian splattering to enable distilled feature fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21676-21685, 2024. 4" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "11316" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/SparseAlign_ a Fully Sparse Framework for Cooperative Object Detection/e9f3f305-e62d-4aad-9255-49e76e43c9cf_content_list.json b/2025/SparseAlign_ a Fully Sparse Framework for Cooperative Object Detection/e9f3f305-e62d-4aad-9255-49e76e43c9cf_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..5b30b298ba9649041fdac910326fd35e07420942 --- /dev/null +++ b/2025/SparseAlign_ a Fully Sparse Framework for Cooperative Object Detection/e9f3f305-e62d-4aad-9255-49e76e43c9cf_content_list.json @@ -0,0 +1,1542 @@ +[ + { + "type": "text", + "text": "SparseAlign: A Fully Sparse Framework for Cooperative Object Detection", + "text_level": 1, + "bbox": [ + 120, + 130, + 875, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yunshuang Yuan $^{1}$ Yan Xia $^{2,3\\dagger}$ Daniel Cremers $^{2,3}$ Monika Sester $^{1}$", + "bbox": [ + 215, + 186, + 779, + 209 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Leibniz University Hannover $^{2}$ Technical University of Munich $^{3}$ Munich Center for Machine Learning (MCML) {yunshuang.yuan, monika.sester}@ikg.uni-hannover.de {yan.xia, cremers}@tum.de", + "bbox": [ + 148, + 210, + 846, + 246 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 280, + 326, + 295 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Cooperative perception can increase the view field and decrease the occlusion of an ego vehicle, hence improving the perception performance and safety of autonomous driving. Despite the success of previous works on cooperative object detection, they mostly operate on dense Bird's Eye View (BEV) feature maps, which are computationally demanding and can hardly be extended to long-range detection problems. More efficient fully sparse frameworks are rarely explored. In this work, we design a fully sparse framework, SparseAlign, with three key features: an enhanced sparse 3D backbone, a query-based temporal context learning module, and a robust detection head specially tailored for sparse features. Extensive experimental results on both OPV2V and DAIR-V2X datasets show that our framework, despite its sparsity, outperforms the state of the art with less communication bandwidth requirements. In addition, experiments on the OPV2Vt and DAIR-V2Xt datasets for time-aligned cooperative object detection also show a significant performance gain compared to the baseline works.", + "bbox": [ + 88, + 311, + 485, + 599 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 626, + 220, + 642 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Cooperative Perception (CP) for autonomous driving has gained significant attention due to its potential to enhance road safety. Sharing Collective Perception Messages (CPMs) among Intelligent (traffic) Agents (IAs) in the vehicular network effectively reduces occlusions and expands the field of view, improving perception performance. In many recent works [1, 16, 24, 29, 35, 41, 42], CP has proven to be effective for different tasks, such as cooperative object detection [1, 24, 29, 41], Bird's Eye View (BEV) semantic segmentation [19, 42], and 3D occupancy semantic segmentation [22]. In this work, we focus on LiDAR-based cooperative vehicle detection.", + "bbox": [ + 89, + 651, + 482, + 830 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Training a cooperative 3D object detector is computationally demanding as it needs to process data from multiple vehicles, especially when the model is built to di", + "bbox": [ + 89, + 832, + 483, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3b92215a35f0c736e791874f42495a26171e0abf1fed4f4b07c69a6b308f2a61.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 514, + 282, + 903, + 359 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1610062c385ae684802eeefeba04c5e130cda9c6944bc712aa6c35442716d363.jpg", + "image_caption": [ + "Figure 1. a,b) Communication strategies for cooperative object detection. c-e) Issues of sparse convolutional backbone networks. c) Center Feature Missing (CFM). d,e) Isolated Convolution Field (ICF) caused by ring disconnectivity and occlusion." + ], + "image_footnote": [], + "bbox": [ + 514, + 364, + 903, + 465 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "gest sequential data from several subsequent frames to learn the temporal context. Following Fcooper [1], recent works [6, 11, 28-30, 33, 35] encode the 3D point cloud into dense BEV feature maps for further processing, sharing, and fusion. However, the computational cost of this pipeline increases quadratically with the perception range. Although cooperation among IAs can improve long-range detection performance, the IAs are normally not in the best positions to optimize the performance gain. The long-range detection performance of the local detector directly influences the overall performance. Besides, sharing the BEV feature maps among IAs requires extensive communication resources. Instead, the sparse operations leverage the sparsity of point clouds and have only linear computational complexity relative to the number of points. In addition, the output sparse features can be more efficient for sharing. It is therefore beneficial to perform cooperative object detection based on sparse operations only.", + "bbox": [ + 511, + 565, + 906, + 838 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "However, building a fully sparse framework to compete with the dense ones is non-trivial. The first unavoidable issue is Center Feature Missing [8] (CFM, shown in Fig. 1 c)). Due to the range view of on-board LiDARs, there are", + "bbox": [ + 511, + 839, + 908, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author.", + "bbox": [ + 109, + 886, + 235, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "22296", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "usually no scan points in the vehicles' center area. However, compared to the edge points, the center points have a better ability to represent the whole object. Lacking center points for learning, the sparse frameworks perform worse than the dense ones that directly learn the features for center points. The second issue is the poor connectivity between points scanned by different laser beams. As shown in Fig. 1 d), this disconnectivity issue in distant areas may lead to isolated voxel blobs; the receptive field only enlarges to the blob scale as the convolutional layers go deeper. The isolated voxel blobs never exchange information with each other, limiting the backbone's capability of capturing global features. We call this issue Isolated Convolution Field (ICF). It also happens in the occlusion areas as shown in Fig. 1 e): although the occluded vehicle has very few scan points (red), it is expected to be detected based on the neighboring vehicles and the clues in the previous frames. However, ICF introduces difficulties for these isolated points to aggregate information from neighbors and the global environment.", + "bbox": [ + 89, + 90, + 480, + 378 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To overcome the above problems, we build a Sparse UNet backbone network SUNet with the Coordinate-Expandable sparse Convolution (CEC) implemented in the MinkowskiEngine [5] library. We apply CEC on $4\\times$ and $8\\times$ downsampled 3D voxels to increase the connectivity between ICFs. On the 2D BEV features, the CECs are used to expand the sparse coordinates, assuring the center features for all scanned vehicles. On top of SUNet, we build TempAlign Module (TAM) to learn the temporal context from the historical frames in a query-based [23, 40] manner, PoseAlign Module (PAM) to correct the relative pose errors between the cooperative and the ego agents, and SpatialAlign Module (SAM) to fuse the shared object queries based on the corrected poses. In addition, we propose the CompassRose encoding for the regression of BBox directions to further improve performance on both Cooperative Object Detection (COOD) and the Time-Aligned COOD [40] (TA-COOD) tasks.", + "bbox": [ + 91, + 380, + 482, + 652 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To summarize, the main contributions of this work are:", + "bbox": [ + 109, + 656, + 472, + 670 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We investigate the issues of building a fully sparse network for LiDAR-based COOD and analyze how these issues impact the detection performance.", + "- We propose a novel fully sparse network, SparseAlign, via addressing the ICF and CFM problems of the sparse 3D backbone SUNet with CECs. Additionally, we introduce the SAM module, specifically tailored to ensure accurate fusion of sparse features.", + "- We assess the communication bandwidth and precision trade-offs of the different approaches, finding that our proposal reduces bandwidth consumption by up to $98\\%$ , compared to previous methods using dense BEV feature maps for LiDAR-based cooperative object detection.", + "We conduct extensive experiments on OPV2V, DAIR-V2X, OPV2Vt, and DAIR-V2Xt benchmarks and show" + ], + "bbox": [ + 91, + 674, + 478, + 896 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "that the proposed SparseAlign has a significant improvement over state-of-the-art methods.", + "bbox": [ + 527, + 90, + 903, + 119 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 138, + 653, + 154 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Cooperative Object Detection (COOD)", + "text_level": 1, + "bbox": [ + 511, + 164, + 846, + 181 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "COOD frameworks are classified into sharing CPMs with raw data, intermediate features, and detections. Recent works [2, 24, 29, 38] have shown that intermediate feature sharing has the best performance due to its adjustable and learnable CPMs. We also use this strategy in our framework. In the perspective of communication, there are two strategies: handshaking and broadcasting as shown in Fig. 1 a) and b). With handshaking communication [11, 28-30, 33, 40, 42], the ego vehicle sends a request message that contains its pose to the cooperative IAs so that they can transform the local raw data to the ego coordinate system and then process their data to obtain the features for sharing. This simplifies the feature fusion process, as the features are directly learned in the goal coordinate system. In contrast, the broadcasting method [6, 17, 24, 37] directly processes the data in the local coordinate system and transmits it to other IAs. In the fusion stage, the ego vehicle must transform the cooperative features into the ego coordinate system and then conduct the fusion process. In comparison, broadcasting is more efficient as it does not require a prior communication agreement with the cooperative agents in highly dynamic driving scenarios where the communication partners change frequently.", + "bbox": [ + 511, + 188, + 903, + 535 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Achieving COOD presents several challenges, including communication latency and bandwidth limitations, inaccurate IA poses, and sensor asynchrony. To reduce the communication bandwidth consumption, [1, 24, 29, 33] compress the learned BEV feature maps to fewer channels via convolution layers, while [13] utilizes the learned confidence map and [42] leverages the uncertainty map to select the most important information for sharing. Instead, [41] employs the two-stage detector PVRCNN [20] and shares the 3D keypoint features from the second stage. To be more compact, [4, 9, 40] share the object query features [23] which are highly flexible to blend all relevant information, such as scanning time and IA pose, about an object into one unique feature vector. In this work, we also leverage the object query features to learn the temporal and spatial context of the object. Pose errors can dramatically deteriorate the detection results [24]. To this end, [39, 41] leverage the predicted Bounding Boxes (BBoxes) and select semantically classified keypoints to align the relative pose between the ego and the cooperative vehicles. Instead, [17] solely uses the BBoxes for pose alignment via the Pose Graph Optimization (PGO). However, these methods require the initial pose errors to be small to find the correct object matching in the first step and then iterate many times, or use the PGO", + "bbox": [ + 511, + 537, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "22297", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/44cc1539675ad7111a932d1986704015c5161b48351f5c39168527d394fa3ad6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 94, + 89, + 570, + 212 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/95d04a62207224d1fbfbf49826dc9e9db10a85618467f71ec704117966537bd3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 576, + 89, + 897, + 212 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/68687900a7229ab0a7b2604115550d9bc2d6419e05d0af4d413294535b8c18f3.jpg", + "image_caption": [ + "SUNet" + ], + "image_footnote": [], + "bbox": [ + 94, + 213, + 509, + 353 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/92b7e57fa75e7da32a2ccc6dd89d886434e1cd007168c160a9c99ed9968218aa.jpg", + "image_caption": [ + "TempAlign Module (TAM)" + ], + "image_footnote": [], + "bbox": [ + 509, + 213, + 700, + 353 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/613cabf80a43f8cd27c2bee3c4e834dcff5a64bb3bbfb4f08604407524ea99f1.jpg", + "image_caption": [ + "PoseAlign Module (PAM)", + "Figure 2. Overview of SparseAlign framework" + ], + "image_footnote": [], + "bbox": [ + 705, + 214, + 897, + 354 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "to refine the result. In this paper, we propose to learn pose-agnostic neighborhood graph features to accurately find the correct matching and then use PGO to refine it. Communication delay can cause that data arriving at the ego agent is outdated. To compensate for this delay, [36, 37] add the first-order derivatives of the BEV feature map to the original feature map to obtain the updated feature map that is aligned to the ego timestamp; [15] employs the LSTM [12] structure on the BEV feature maps to adjust the features to be aligned. Differently, [40] embeds the observation timestamps as well as the temporal context from historical frames into the object queries so that they are time- and motion-aware and capable of predicting the status at any given future timestamp, achieving the time-delay compensation. Different to [15, 36, 37] that compensate the feature maps, [40] leverages the point-wise timestamp of the point clouds to learn the accurate temporal context, hence is also able to cope with sensor asynchrony issues. Therefore, we adopt the query-based strategy for time-related compensation.", + "bbox": [ + 88, + 422, + 485, + 710 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2.3DBackbones", + "text_level": 1, + "bbox": [ + 89, + 724, + 236, + 739 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To extract point cloud features, most previous works [16, 17, 24, 29, 33, 37, 41] for COOD either use dense 3D [25, 44], dense 2D [14, 32] or semi-sparse [21, 26, 31, 43] backbones that first encode the point cloud with 3D sparse convolution and then apply dense 2D convolutions for further processing. In contrast, [40, 42] employ the MinkUnet [5] to extract point cloud features. Despite its efficiency, the MinkUnet backbone struggles with long-range detection, occlusions and large objects, because of CFM and ICF problems. [8] proposed center voting for each scanned ob", + "bbox": [ + 89, + 750, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ject points to solve CFM issue and [3] reveals that further down-sampling the voxel features to lower resolutions can increase the receptive field and refine the large object detection. Instead of shifting the feature locations as [8] to accurately locate object queries or adding additional layers as [3], we build a backbone that solves both CFM and ICF issues by modifying MinkUnet with CEC layers.", + "bbox": [ + 511, + 422, + 906, + 529 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 511, + 547, + 604, + 561 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Task formalization", + "text_level": 1, + "bbox": [ + 511, + 574, + 692, + 588 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this work, we solve the LiDAR-based cooperative object detection task considering unavoidable challenges in reality: communication delay, sensor asynchrony, and localization errors. For this scenario, we assume there are $N$ IAs, including one ego agent $A_{1}$ and $N - 1$ cooperative agents. These agents are sharing CPMs with each other in a broadcasting manner, as shown in Fig. 1b). Specifically, in the $i$ -th frame, $A_{1}$ receives the CPMs $M = \\{(m_j,t_j)|j\\in \\{1,\\dots ,N - 1\\} \\}$ from the cooperative agents, where $m_j$ and $t_j$ are the CPM and the corresponding timestamp from the $j$ -th agent. We aim to optimize the object detection results based on the point clouds captured so far ( $t\\leq t_i$ ) at $A_{1}$ and the received messages $M$ . We model and evaluate both the COOD and TA-COOD tasks from the perspective of an ego agent. The key difference between COOD and TA-COOD lies in the GT BBoxes. As illustrated in Fig. 2, TAM, COOD GT BBoxes (dashed lines) align with the sequentially captured LiDAR points, whereas TA-COD GT BBoxes (solid lines) are referenced to a global timestamp $t_g$ . This distinction requires the model to accurately predict", + "bbox": [ + 511, + 598, + 908, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "22298", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/682b68fa3080748eca50c1f79abb0047e99ec61a838df4a0ff883c2393331b16.jpg", + "image_caption": [ + "Figure 3. Receptive field (RF) after 4 sparse convolutions (convs). a. LiDAR points; b. Spare pixels (white) with a conv center point (red); c. RF coverage of 4 normal sparse convs (red); d. RF coverage of 4 CECs (red+light blue, darkest blue is background)." + ], + "image_footnote": [], + "bbox": [ + 99, + 90, + 192, + 165 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/9e202ffca0ac72622c4111bc1cee2959b3e65db2646a1f716e10f0a6327cdc71.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 196, + 90, + 287, + 162 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/bdbdc5d946e97c88ac9dafc4b3ca949f19f98df3ccb231c8999295f2aa952b5a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 290, + 90, + 382, + 162 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/2c53502a3f07b26b836b892ff66e5e4cd6e3c960e129e07eb3436935a53573ef.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 385, + 90, + 477, + 162 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "BBox locations at $t_g$ . While COOD relies solely on point cloud geometry for location prediction, TA-COOD leverages temporal context to enhance accuracy at $t_g$ and also enables the model to handle location errors due to communication latency.", + "bbox": [ + 89, + 246, + 483, + 321 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Framework overview", + "text_level": 1, + "bbox": [ + 89, + 330, + 290, + 345 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We build an efficient fully sparse framework SparseAlign, as shown in Fig. 2, to process sequential point clouds from multiple agents. All agents share the same network weights. The input at time $t_i$ is the point cloud of the ego agent $PC_{t_i}$ and of the cooperative agents $PC_{t_j}$ . Note that $t_j < t_i$ when communication latency exists. All point clouds are processed by the backbone SUNet to extract the point cloud features. The Region of Interest (RoI) module then selects the most interesting features as the object queries. For simplicity, we define object queries as $Q^{*} = \\{(F_{i}^{*}, x_{i}^{*}, y_{i}^{*}) | i \\in \\{1, \\dots, N^{*}\\}$ , where * represents the query set name, $F$ is the features, and $x, y$ are the coordinates of the corresponding $i$ -th query. With selected queries, the TempAlign Module (TAM) can efficiently align the query features to the globally aligned timestamp $t$ so that errors introduced by communication delay and sensor asynchrony are compensated. The updated query features and the detections from the Local Query Detection head (LQDet) are then shared as CPMs. The PoseAlign Module (PAM) then uses the detected BBoxes to correct the relative poses between the cooperative and the ego agent and the SpatialAlign Module (SAM) fuses the received cooperative features into the ego coordinate system based on the corrected transformation $T_{c}^{e}$ from cooperative to ego coordinates.", + "bbox": [ + 89, + 354, + 483, + 717 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Backbone: Sparse UNet (SUNet)", + "text_level": 1, + "bbox": [ + 89, + 726, + 377, + 742 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Processing data from multiple IAs for cooperative perception is computationally demanding. To improve efficiency, we use a fully sparse 3D UNet, similar to MinkUnet[5], to construct the SUNet (Fig. 2 bottom left). However, the fully sparse convolutions used in such networks suffer from the ICF issue. As shown in Fig. 3, the receptive field (RF) of (c) standard sparse convolutions only covers LiDAR points from a single vehicle. In contrast, the RF of (d) the CEC layers extends to the points of a neighboring vehicle, enabling learning within a larger spatial context. Therefore,", + "bbox": [ + 89, + 750, + 482, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "we use 3D CECs in the sparse convolution (SConv) block 3 and 4. Correspondingly, we contract the coordinates in the transposed sparse convolution (TConv) block 3 and 2. In this way, the output of the TConv blocks $p4$ and $p2$ can match the coordinates in stride $x2$ and $x4$ for the concatenation. In addition, it learns better global features without voxel number soaring in the subsequent layers and maintains efficiency. Against the CFM issue, we use 2D CECs on the 2D BEV sparse features (Fig. 2 SUNet:Coor. Expand.) to expand the coordinates to ensure their coverage over the object centers.", + "bbox": [ + 511, + 90, + 905, + 257 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Alignment modules", + "text_level": 1, + "bbox": [ + 511, + 266, + 699, + 282 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "TempAlign Module (TAM). The TAM aims to learn the temporal context relative to the previous frames to achieve two goals: improve object detection performance by looking for clues in the history and compensate for the object displacement error introduced by asynchronous scanning time (Fig. 2, TAM). We achieve the second goal by predicting the BBoxes at the globally aligned timestamp so that the BBoxes shared for PAM are synchronized. Then PAM only needs to deal with the object displacement caused by localization errors. Technically, the TAM follows [40], for efficiency, by removing the global cross-attention module which learns the global BEV features. We only use a memory queue and a hybrid attention module. The memory queue stores the top $k = 256$ object queries from previous $L = 4$ data frames so that the attention module can allow the interaction between object queries in the current frame and the historical frames. In this way, the object queries learn the temporal context from the sequential data and are able to make predictions for the future state.", + "bbox": [ + 511, + 287, + 906, + 575 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "PoseAlign Module (PAM). Given the detected BBoxes $\\mathbf{B}_i$ and $\\mathbf{B}_j$ , respectively, from the perspective of the agent $A_i$ and $A_j$ in the same scene, humans can easily find the correspondences between $\\mathbf{B}_i$ and $\\mathbf{B}_j$ based on the geometric relationships. Practically, we model the features of these geometric relationships by embedding the neighborhood geometry of each BBox into deep features and then match these features of $\\mathbf{B}_i$ and $\\mathbf{B}_j$ with the Hungarian algorithm. For each BBox $b$ in a BBox set $\\mathbf{B}$ , we select the nearest $k = 8$ neighbors to embed its neighborhood features. As shown in Fig. 2, PAM: 1.Graph Feature, the orientation angles of the BBox are $\\alpha$ and $\\beta$ for $b$ and its neighbor, respectively. We embed the relative orientation with $\\nu_{a} = [\\sin (\\beta -\\alpha);\\cos (\\beta -\\alpha)]^{1}$ . Similarly, the relative edge direction is embedded as $\\epsilon_{a} = [\\sin (\\theta -\\alpha);\\cos (\\theta -\\alpha)]$ . In addition, the dimension $\\nu_{dim} = [l,w,h]$ of the neighbor BBox, and the Euclidean distance between $b$ and its neighbors $\\epsilon_d = d$ , are also embedded into the feature. Note that $\\epsilon_d,\\epsilon_a,\\nu_a$ are all relative features between two BBoxes that are pose-agnostic, leading to robust pose alignment without", + "bbox": [ + 511, + 575, + 908, + 878 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "1[.;.]:concatenation", + "bbox": [ + 531, + 886, + 643, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "22299", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "dependency on the magnitude of initial pose errors. These features are then concatenated and embedded into high-dimensional features by a linear layer with ReLU activation $(\\mathbb{R}^5\\to \\mathbb{R}^{d_{emb}})$ :", + "bbox": [ + 89, + 90, + 483, + 151 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nf _ {n b r} = \\operatorname {l i n e a r} \\left(\\left[ \\epsilon_ {d}; \\epsilon_ {a}; \\nu_ {a}; \\nu_ {\\text {d i m}} \\right]\\right) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 176, + 162, + 482, + 179 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Assembling the features relative to all neighbors, we summarize the neighborhood feature $F_{nbr} \\in \\mathbb{R}^{k \\times d_{emb}}$ by a multi-head self-attention module $attn$ followed by a linear layer operated on the mean and max of the $attn$ outcome, resulting in the feature $F_{nbr} \\in \\mathbb{R}^{d_{emb}}$ :", + "bbox": [ + 89, + 188, + 483, + 263 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nF _ {n b r} = \\operatorname {a t t n} \\left(F _ {n b r}\\right) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 138, + 273, + 482, + 290 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nF _ {n b r} = \\operatorname {l i n e a r} \\left(\\left[ \\operatorname {m e a n} \\left(F _ {n b r}\\right); \\operatorname {m a x} \\left(F _ {n b r}\\right) \\right]\\right) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 140, + 292, + 482, + 310 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "By calculating the Euclidean distance between the learned neighborhood feature $F_{nbr}$ of the BBox in $\\mathbf{B}_i$ and $\\mathbf{B}_j$ , we construct the cost matrix for the linear sum assignment to find the best Match (Fig. 2, PAM: 2-match) between these two BBox sets. However, an object in $\\mathbf{B}_i$ does not always have a projection in $\\mathbf{B}_j$ and vice versa. Thus, we reject the wrong matches that have large distances. Based on the PAM: 2-match result, the relative pose transformation between $A_i$ and $A_j$ can be recovered, completing the PAM: 3Align step. Additionally, we use pose graph optimization to refine the alignment by leveraging loop closure among multiple cooperative agents posing following [17].", + "bbox": [ + 89, + 319, + 483, + 500 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "SpatialAlign Module (SAM). To align the cooperative query features $Q^c$ with the ego query features $Q^e$ , there are two issues to be resolved. First, the features $F^c$ are learned in the cooperative coordinate system and must be adapted to the ego coordinate system. We achieve this using a multilayer perceptron (mlp) as described in Eq. (4) where $F^R \\in \\mathbb{R}^{N \\times 9}$ is the flattened and repeated rotation matrix $R \\in \\mathbb{R}^{3 \\times 3}$ .", + "bbox": [ + 89, + 500, + 483, + 619 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nF ^ {c} = m l p \\left(\\left[ F ^ {c}; F ^ {R} \\right]\\right) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 214, + 628, + 482, + 648 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Second, after rotation with $R$ , the coordinates of $Q^c$ (Fig. 2, SAM, yellow points) do not perfectly align to the discretized grid in the ego coordinate system. Thus, we merge $Q^c$ into the nearest points of this grid. The locations of the fused output object query points $Q^{ts}$ are shown with gray points. For each point in $Q^{ts}$ , we summarize the features of its $k = 8$ nearest neighbors (Fig. 2, SAM: black arrows) in query set $Q^{\\cup} = Q^{c}\\bigcup Q^{e}$ to obtain the fused features $F^{ts}$ . Mathematically, it reads as", + "bbox": [ + 89, + 657, + 483, + 792 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nF _ {i j} = m l p \\left(\\left[ F _ {j} ^ {\\cup}; \\text {l i n e a r} \\left(\\left[ x _ {i} ^ {t s} - x _ {j} ^ {\\cup}, y _ {i} ^ {t s} - y _ {j} ^ {\\cup} \\right]\\right) \\right]\\right) \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 116, + 801, + 482, + 821 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nF _ {i} ^ {t s} = \\left\\{F _ {i j} | j \\in \\{1, \\dots , k \\} \\right\\} \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 116, + 823, + 482, + 840 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nF _ {i} ^ {t s} = \\max _ {j} \\left(F _ {i} ^ {t s}\\right) + \\text {m e a n} _ {j} \\left(F _ {i} ^ {t s}\\right) \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 116, + 843, + 482, + 861 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $i$ is the query index of $Q^{ts}$ and $j$ is the neighbor index of the $i$ -th query.", + "bbox": [ + 89, + 869, + 482, + 901 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/529f1456c1122a5dfe97c8baaff0642621ac6e2ca352e44cfa80719f4f141a2f.jpg", + "image_caption": [ + "Figure 4. CompassRose encoding" + ], + "image_footnote": [], + "bbox": [ + 524, + 90, + 898, + 174 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Detection Heads", + "text_level": 1, + "bbox": [ + 513, + 229, + 674, + 244 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To train SparseAlign, we follow [40] to attach a center-based object detection head to each of the RoI, TAM and SAM to learn query features. These heads are notated as RoIDet, LQDet and GQDet, respectively. They share the same target-encoding method. We use Focal loss to classify fore- and background object queries; all query points inside the ground-truth BBoxes are positive, and those outside are negative. Smooth $L1$ loss is used for regressing $[dx,dy,dz,l,w,h,\\mathcal{B}_{dir},\\mathcal{B}_{scr}]$ , where $l,w,h$ are the dimensions of BBoxes and $dx,dy,dz$ are coordinate offsets between the query points and the center of ground-truth BBoxes. $\\mathcal{B}_{dir},\\mathcal{B}_{scr}$ are the novel CompassRose encoding (Fig. 4) of the ground truth orientation $r_g$ with respect to four anchor angles $\\mathbf{r}_a = [0,0.5\\pi ,\\pi ,1.5\\pi ]$ . It reads as", + "bbox": [ + 511, + 253, + 906, + 465 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {B} _ {\\mathrm {d i r}} = \\left[ \\cos r _ {g} - \\cos \\mathbf {r} _ {a}, \\sin r _ {g} - \\sin \\mathbf {r} _ {a} \\right] (8) \\\\ \\mathcal {B} _ {\\mathrm {s c r}} = \\left[ s _ {0}, s _ {0. 5 \\pi}, s _ {\\pi}, s _ {1. 5 \\pi} \\right] (9) \\\\ = \\mathbf {1} - \\arccos \\left(\\cos \\mathbf {r} _ {a} \\cdot \\cos r _ {g} + \\sin \\mathbf {r} _ {a} \\cdot \\sin r _ {g}\\right) / \\pi \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 529, + 476, + 903, + 531 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Take the encoded sine wave (Fig. 4 left) as an example, the four anchor angles (yellow points) ensure that at least one anchor can reach the target angle monotonically, simplifying the regression process.", + "bbox": [ + 511, + 542, + 905, + 603 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.6. Data augmentation", + "text_level": 1, + "bbox": [ + 513, + 612, + 697, + 628 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In addition to geometric augmentation, which includes randomly flipping the input point cloud along the x- or y-axis and rotating it around the z-axis, we apply Free Space Augmentation (FSA) [42]. FSA enhances the point cloud by adding auxiliary points along LiDAR scanning rays, representing empty spaces where no objects are present. This augmentation helps mitigate the ICF problem in distant regions, where observation points are extremely sparse. For more details, refer to supplementary B.", + "bbox": [ + 511, + 635, + 906, + 772 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiment", + "text_level": 1, + "bbox": [ + 511, + 784, + 637, + 801 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. We evaluate SparseAlign on OPV2V[29] and DAIR-V2X[35] datasets for COOD and their variant datasets OPV2Vt[40] and DAIR-V2Xt[40] for TA-COOD. OPV2V is a synthetic dataset created in different driving scenarios of CARLA [7] simulator. Following [29], the detection range is set to $x \\in [-140, 140]m$ , $y \\in [-40, 40]m$ .", + "bbox": [ + 511, + 809, + 906, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "22300", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/a05d590cc6c4537282991804a69b7a64246ea5a736e4b75d202b9c93e4c82455.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodBW(Mb) ↓AP0.5 ↑AP0.7↑
Fcooper[1]72.080.8870.790
V2VNet[24]72.080.9170.822
FPVRCNN[41]0.240.8730.820
OPV2V[29]126.80.9050.815
CoAlign[17]72.080.9020.833
CoBEVT[19]72.080.9130.861
V2VAM[16]72.080.9160.850
V2VFformer++[34]L+C72.080.9350.895
SparseAlign (ours)< 1.30.9300.892
", + "bbox": [ + 96, + 89, + 478, + 239 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/a567cf47c12a2f46c6f75d2cfe1872891d43eab3a1eadeaf0fb7c9d6dae9eab7.jpg", + "table_caption": [ + "Table 1. COOD Average Precision on OPV2V dataset. ${}^{2}$" + ], + "table_footnote": [], + "table_body": "
MethodOPV2VtDAIR-V2Xt
AP0.5↑AP0.7↑AP0.5↑AP0.7↑
Fcooper[1]+SA0.7630.5530.5970.282
FPVRCNN[43]+SA0.6400.4740.5980.307
OPV2V[29]+SA0.8810.7870.7020.366
StreamLTS[40]0.8530.7210.6400.404
StreamLTS[40]+SA0.8500.7900.7570.497
SparseAlign (full)0.8930.8180.7960.548
", + "bbox": [ + 96, + 291, + 478, + 407 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "DAIR-V2X is a real dataset containing one vehicle- and one infrastructure-side sensor suit. The detection range is set to $x \\in [-100, 100]m, y \\in [-40, 40]m$ . OPV2Vt is generated by interpolating OPV2V data frames into sub-frames and replaying them to mimic the sensor asynchrony and the rolling shutter problem of LiDARs. The GT BBoxes of this dataset are aligned to the global timestamp, the LiDAR scan ending time of each frame. DAIR-V2Xt follows the same idea. The detection range of OPV2Vt and DAIR-V2Xt is set the same as OPV2V and DAIR-V2X.", + "bbox": [ + 88, + 477, + 482, + 628 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation details. All models are trained on a single GTX-4090 and an Intel i7-8700 CPU for 50 epochs (about 35 hours) with the Adam optimizer and batch size two. The learning rate is set to $2e - 4$ with a warm-up stage of 4000 iterations. The random rotation angle is set to the range $[-90^{\\circ}, 90^{\\circ}]$ . For models incorporating the TAM, we load four sequential frames of data during training and calculate losses solely for the last frame. To reduce GPU memory consumption, we enable gradient calculations only for data from a single ego agent and one cooperative agent. For details on efficient gradient scheduling, refer to supplementary $D$ . To ensure that the model generalizes well to the significant time latency introduced by communication, we load data from cooperative vehicles with a random latency ranging from 0 to $200~\\mathrm{ms}$ relative to the ego vehicles.", + "bbox": [ + 89, + 630, + 482, + 858 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/49227effd71dca5d664a5461b91ba5dcf1dec2ad564d1d287bd7d44fa5a3d9e9.jpg", + "table_caption": [ + "Table 3. TA-COOD Average Precision on OPV2Vt and DAIR-V2Xt" + ], + "table_footnote": [], + "table_body": "
MethodBW(Mb) ↓AP0.5 ↑AP0.7↑
Fcooper[1]48.80.7340.559
V2VNet[24]48.80.6640.402
FPVRCNN[41]0.240.6650.505
V2XViT[28]48.80.7040.531
DiscoNet[19]48.80.7360.583
OPV2V[29]97.60.7330.553
CoAlign[17]48.80.7460.604
DI-V2X[27]48.80.7880.662
SparseAlign (ours)< 1.30.8450.685
", + "bbox": [ + 542, + 89, + 901, + 241 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. COOD Average Precision on DAIR-V2X dataset.", + "bbox": [ + 545, + 252, + 893, + 266 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Results", + "text_level": 1, + "bbox": [ + 511, + 292, + 601, + 306 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Comparison to the state-of-the-art", + "text_level": 1, + "bbox": [ + 511, + 316, + 812, + 333 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Tab. 1 presents the COOD results on the OPV2V dataset. All methods use only LiDAR data, except for V2VFormer++ (denoted by $\\mathrm{L + C}$ ), which incorporates both LiDAR and camera data. Compared to the state-of-the-art, our SparseAlign achieves the best performance in LiDAR-based COOD, approaching the results of V2VFormer++, which benefits from LiDAR-Camera fusion. Tab. 2 shows the results on the DAIR-V2X dataset. Here, SparseAlign outperforms other methods by a large margin, achieving a $2.3\\%$ improvement in AP at the IoU threshold of 0.7 and a $5.7\\%$ improvement at the IoU threshold of 0.5. This is mainly due to the enhanced temporal and global reasoning ability of our model for detecting distant and large objects. Compared to other methods that rely on dense BEV feature maps, SparseAlign consumes the least communication bandwidth (BW) of less than 1.3MB (without compression) thanks to the efficient query-based operations.", + "bbox": [ + 511, + 339, + 903, + 595 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For the results of the more challenging TA-COOD task shown in Tab. 3, our SparseAlign outperforms StreamLTS by a great margin. By applying our three alignment modules (StreamLTS + SA), the performance of StreamLTS is also improved. In addition, we compare our framework to Fcooper, FPVRCNN and OPV2V method by attaching the three alignment modules to them to achieve TA-COOD. All show worse performance than the full SparseAlign because of worse 3D backbone or the RoI selection.", + "bbox": [ + 511, + 597, + 905, + 731 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The qualitative results and the additional comparison to methods, such as TransIFF[4], on the 3D AP metric are given in the supplementary A and E, respectively.", + "bbox": [ + 511, + 732, + 905, + 779 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Ablation study", + "text_level": 1, + "bbox": [ + 511, + 786, + 663, + 801 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We conduct an ablation study on the 3D backbone, TAM, and SAM (PAM is discussed separately in Sec. 5.4). Among these components, the spatial fusion module plays a crucial role in multi-agent feature fusion. Without SAM, the module performs the worst, as evidenced by the first row of Tab. 4. When using QUEST [9] spatial fusion, which", + "bbox": [ + 511, + 809, + 905, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "2The results reported in this table use frame-wise local score sorting for AP, same to the official benchmark. Other results use global sorting.", + "bbox": [ + 89, + 875, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "22301", + "bbox": [ + 478, + 944, + 517, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/67b06a0e389cf8f639a5cf911871ca5a9c4bc2ba7f442b83e9f39736a9da98a1.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
CEC 2DCEC 3DTAMSAMOPV2VDAIR-V2X
AP0.5AP0.7AP0.5AP0.7
✓✓✓0.8470.7560.6710.528
✓✓✓QST0.8880.8690.7490.622
✓0.9240.8850.7730.638
✓✓0.9400.9010.7960.662
✓✓✓0.9500.9140.8130.682
✓✓✓✓0.9510.9290.8450.685
", + "bbox": [ + 96, + 89, + 537, + 215 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/5d2885a02827632217ff15559198d9dea5cbc262d3a4f7e9248006670505a394.jpg", + "table_caption": [ + "Table 4. Ablation study on 2D and 3D CEC layers, TAM and SAM. QST: use QUEST [9] spatial fusion instead of SAM." + ], + "table_footnote": [], + "table_body": "
Dir.Enc.OPV2VDAIR-V2X
AP0.5AP0.7AP0.5AP0.7
gt-angle0.8710.7680.6200.427
second0.9080.8740.7430.590
sin_cos0.9150.8810.7480.603
compass0.9240.8850.7730.638
", + "bbox": [ + 580, + 89, + 903, + 188 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/4fbefe9fd2dd11c2d0e99147fe29e9d46620ab6e4b49d9848b1a1e497bc7b792.jpg", + "table_caption": [ + "Table 5. Comparison on BBox angle encoding methods: offset to ground-truth angle (gt-angle), second-style (second), sine-cosine offsets (sin_cos) and compass rose (compass)." + ], + "table_footnote": [], + "table_body": "
DatasetOPV2VtDAIR-V2Xt
Latency0ms100ms200ms0ms100ms200ms
AP threshold0.50.70.50.70.50.70.50.70.50.70.50.7
StreamLTS [40]0.8530.7210.8160.6800.7870.6470.6420.4040.6130.3790.5900.364
SA no TAM0.8900.8020.8560.4710.6310.2090.7200.4570.6720.3930.6350.350
SA no PT0.8350.6150.8310.6250.8150.5810.7600.4690.7370.4590.7220.449
SA no TL0.8850.8430.8620.6190.7230.2790.7930.5340.7490.4740.6980.423
SA no FSA0.8810.8160.8750.7950.8570.7620.7690.5080.7550.5020.7400.496
SparseAlign (SA)0.8930.8180.8850.7950.8670.7650.7960.5480.7860.5430.7720.532
", + "bbox": [ + 116, + 268, + 883, + 406 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 6. AP of TA-COOD with communication latency. PT: Point-wise Timestamps, TL: Train Latency, FSA: Free Space Augmentation.", + "bbox": [ + 91, + 417, + 903, + 431 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "aligns only rotation (Eq. (4)) without correcting spatial displacement (Eqs. (5) to (7)), the AP scores remain significantly lower than those obtained with SAM. Incorporating SAM improves performance, and we further enhance it by gradually replacing standard sparse convolutional layers with CEC layers. Finally, TAM further boosts AP scores by leveraging temporal information from historical frames.", + "bbox": [ + 88, + 459, + 482, + 564 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In Tab. 5, we compare CompassRose encoding with three methods: directly regress the ground-truth angle (gt-angle), second-style [31] encoding (second), and the sine-cosine encoding (sin_cos). For simplicity, we use the configurations in the third row of Tab. 4 for the comparison. The results show that CompassRose performs the best.", + "bbox": [ + 89, + 565, + 482, + 656 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Communication latency and sensor asynchrony", + "text_level": 1, + "bbox": [ + 89, + 665, + 482, + 681 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Tab. 6 shows the results of TA-COOD on the OPV2Vt and the DAIR-V2Xt dataset. Our proposed SparseAlign demonstrates a significant improvement over the baseline framework, StreamLTS [40]. Specifically, the AP at an IoU threshold of 0.7 increased $9.7\\%$ on the OPV2Vt dataset and $14.4\\%$ on the DAIR-V2Xt dataset. As the communication latency increases to 200ms, the AP at IoU 0.5 of our SparseAlign only dropped $2.6\\%$ while the baseline work has lost $6.6\\%$ of its accuracy. This reveals that our framework is more robust against communication latency. To investigate modules or configurations that are making an effect on this robustness, we conducted an ablation study by removing the TAM, Point-wise Timestamps (TP), Training Latency (TL, communication latency during training) and", + "bbox": [ + 89, + 688, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Free Space Augmentation (FSA), respectively.", + "bbox": [ + 511, + 459, + 821, + 473 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Without the TAM, the performance noticeably declines as latency increases, despite achieving comparable results to the full SpareAlign model at $0\\mathrm{ms}$ latency. A similar effect is observed when the full SA model is trained without exposure to data with communication latency (SA without TL). Remarkably, without TL, the model outperforms the full SpareAlign model in terms of AP at an IoU threshold of 0.7 on the OPV2Vt dataset (0.843 against 0.818). This is because the model focuses more on predicting BBoxes within the same frame, where only minimal latency is introduced by asynchronous sensors.", + "bbox": [ + 509, + 474, + 906, + 640 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Without PT, the model performs much worse than the full SpareAlign, showing the importance of fine-grained point-wise timestamps in the point clouds for the model to compensate for the errors introduced by sensor asynchrony and capture the accurate temporal context for prediction. Finally, the ablation study on the FSA module also shows a positive influence on the performance. For example, AP at the IoU of 0.7 increased by $4\\%$ on DAIR-V2Xt by utilizing the FSA (See supplementary D for visual comparison).", + "bbox": [ + 509, + 640, + 908, + 777 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.4. Localization errors", + "text_level": 1, + "bbox": [ + 511, + 787, + 697, + 801 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Fig. 5 shows the performance of SpareAlign against different localization errors. We add random errors to the poses of both ego $(T_e)$ and cooperative $(T_c)$ vehicles. The translation errors $x_{\\epsilon}, y_{\\epsilon}$ along the $x$ - and $y$ -axis and the rotation error $r_{\\epsilon}$ around the $z$ -axis are all assumed to be normally distributed with $\\mathcal{N}(0,1) \\cdot \\epsilon$ , where $\\epsilon \\in [0,1]$ is the error scaling factor.", + "bbox": [ + 511, + 810, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "22302", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/de9edb1b1ab958c861bf4550b46e38a36f72222773356e1ae489e155e178d8c9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 117, + 92, + 455, + 223 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d963ed326dce265d53ea2cafabe9165d8f70ada8d9bc5100897043381cf87a89.jpg", + "image_caption": [ + "Figure 5. AP at IoU threshold of 0.7 with translation errors ranging from $0\\mathrm{m}$ to $1\\mathrm{m}$ along x- and y-axis, and rotation errors from $0^{\\circ}$ to $1.0^{\\circ}$ (horizontal axis) for the different datasets." + ], + "image_footnote": [], + "bbox": [ + 114, + 224, + 455, + 354 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We report the results by gradually increasing $\\epsilon$ with a step size of 0.2. Note that the translation and rotation errors exist in the poses of both the ego and the cooperative vehicle; the resulting relative translation error of the transformation $T_{c}^{e} = inverse(T_{e})\\cdot T_{c}$ from cooperative to ego coordinate system will be amplified as the rotation error increases. Our PAM is specially designed to mitigate the influence of large relative transformation errors.", + "bbox": [ + 89, + 440, + 483, + 559 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our proposed PAM and the baseline method CoAlign [17] both significantly reduce the impact of pose errors on AP performance compared to the configuration lacking any correction module. In addition, our method demonstrates greater robustness against large errors. For instance, on the OPV2V dataset, the AP for CoAlign decreased by approximately $27\\%$ at the error of $1.0m, 1.0m, 1.0^{\\circ}$ , whereas PAM experienced only a $12\\%$ drop. This validates the efficacy of our design, which relies on the pose-agnostic relative neighborhood geometries to match the detected BBoxes of the ego and the cooperative IAs. Note that the DAIR-V2X dataset exhibits smaller changes in AP as localization errors increase, because the pose parameters are not well calibrated. In contrast, [40] refined the poses of DAIR-V2X for generating the DAIR-V2Xt dataset, making the better-calibrated DAIR-V2Xt more sensitive to newly introduced errors.", + "bbox": [ + 89, + 561, + 483, + 819 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.5. CPM sizes", + "text_level": 1, + "bbox": [ + 89, + 830, + 209, + 845 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our framework significantly reduces communication bandwidth without relying on computationally intensive dense feature maps, unlike state-of-the-art methods. This effi", + "bbox": [ + 89, + 854, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c77fbab4063aa88260bffeeeae859f74e50be0d25aeac07ea4597d41fcfb6a98.jpg", + "image_caption": [ + "Figure 6. CPM sizes with different information selection scores. Object query features with detection scores larger than the threshold (x-axis) are shared as CPMs." + ], + "image_footnote": [], + "bbox": [ + 540, + 93, + 709, + 220 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2df0624091bbbf8392d6db71bcd8377fe6a2e7bf9823f149723911361688707e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 710, + 93, + 880, + 220 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ciency is achieved by selecting only the top $K = 1024$ object queries for further processing. In addition to this intrinsic reduction, the size of the CPMs can be further minimized through score-based selection. Specifically, we employ the query-based detection scores from the LQDet head to identify and share only the most critical queries. As Fig. 6 shows, using a CPM score threshold below 0.5 does not lead to a noticeable degradation in performance across all datasets. Remarkably, at a threshold score of 0.5, the average size of the CPMs is reduced to less than $400\\mathrm{KB}$ on all datasets. The final CPM sizes after selection are influenced by the driving scenes. Specifically, scenes with more vehicles tend to have larger CPM sizes. Consequently, the OPV2V and OPV2Vt datasets, which contain more vehicles in their scenes, result in larger CPM sizes compared to the DAIR-V2X and DAIR-V2Xt datasets.", + "bbox": [ + 511, + 309, + 906, + 551 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 571, + 633, + 585 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we proposed SparseAlign, a fully sparse framework for cooperative object detection (COOD). With our novel 3D backbone network SUNet, the Compass-Rose encoding of detection head, the temporal and spatial fusion module TAM and SAM, SparseAlign significantly outperforms state-of-the-art methods on the OPV2V and DAIR-V2X datasets for COOD, as well as on their variants OPV2Vt and DAIR-V2Xt for time-aligned COOD (TA-COOD), achieving this with very low communication bandwidth. Additionally, our proposed pose alignment module PAM demonstrates improved robustness to large localization errors compared to the baseline method CoAlign. Nevertheless, when the pose errors are large, a drop of about $10\\%$ in APs persists even with pose correction. In future work, we therefore plan to explore more robust methods, e.g. leveraging sub-graph detection, to further enhance the pose alignment accuracy. Additionally, more application scenarios, such as roadside cooperative detection benchmarks HoloVIC[18], RCooper[10] and TUMTraf[45], could also be explored.", + "bbox": [ + 511, + 598, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "22303", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Qi Chen. F-cooper: feature based cooperative perception for autonomous vehicle edge computing system using 3d point clouds. Proceedings of the 4th ACM/IEEE Symposium on Edge Computing, 2019. 1, 2, 6", + "[2] Qi Chen, Sihai Tang, Qing Yang, and Song Fu. Cooper: Cooperative perception for connected autonomous vehicles based on 3d point clouds. 2019 IEEE 39th International Conference on Distributed Computing Systems (ICDCS), pages 514-524, 2019. 2", + "[3] Yukang Chen, Jianhui Liu, Xiangyu Zhang, Xiaojuan Qi, and Jiaya Jia. Voxelnext: Fully sparse voxelnet for 3d object detection and tracking. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 21674-21683, 2023. 3", + "[4] Ziming Chen, Yifeng Shi, and Jinrang Jia. TransIFF: An instance-level feature fusion framework for vehicleinfrastructure cooperative 3d detection with transformers. In 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 18159-18168, 2023. 2, 6", + "[5] Christopher Bongsoo Choy, JunYoung Gwak, and Silvio Savarese. 4d spatio-temporal convnets: Minkowski convolutional neural networks. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3070-3079, 2019. 2, 3, 4", + "[6] Jiaxun Cui, Hang Qiu, Dian Chen, Peter Stone, and Yuke Zhu. Coopernaut: End-to-end driving with cooperative perception for networked vehicles. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), page 17252-17262, 2022. 1, 2", + "[7] Alexey Dosovitskiy, German Ros, Felipe Codevilla, Antonio Lopez, and Vladlen Koltun. CARLA: An open urban driving simulator. In Proceedings of the 1st Annual Conference on Robot Learning, pages 1-16, 2017. 5", + "[8] Lue Fan, Feng Wang, Naiyan Wang, and Zhaoxiang Zhang. Fully Sparse 3D Object Detection. In NeurIPS, 2022. 1, 3", + "[9] Siqi Fan, Haibao Yu, Wenxian Yang, Jirui Yuan, and Zaiqing Nie. Quest: Query stream for practical cooperative perception. In IEEE International Conference on Robotics and Automation (ICRA), 2024. 2, 6, 7", + "[10] Ruiyang Hao, Siqi Fan, Yingru Dai, Zhenlin Zhang, Chenxi Li, Yuntian Wang, Haibao Yu, Wenxian Yang, Yuan Jirui, and Zaiqing Nie. Rcooper: A real-world large-scale dataset for roadside cooperative perception. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 22347-22357, 2024. 8", + "[11] Caizhen He, Hai Wang, Long Chen, Tong Luo, and Yingfeng Cai. V2x-ahd:vehicle-to-everything cooperation perception via asymmetric heterogenous distillation network. arXiv, arXiv:2310.06603, 2023. arXiv:2310.06603 [cs]. 1, 2", + "[12] Sepp Hochreiter and Jürgen Schmidhuber. Long short-term memory. Neural Computation, 9(8):1735-1780, 1997. 3", + "[13] Yue Hu, Shaoheng Fang, Zixing Lei, Yiqi Zhong, and Siheng Chen. Where2comm: Communication-efficient collaborative perception via spatial confidence maps. In Thirty-sixth Conference on Neural Information Processing Systems (Neurips), 2022. 2" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Alex H. Lang, Sourabh Vora, Holger Caesar, Lubing Zhou, Jiong Yang, and Oscar Beijbom. Pointpillars: Fast encoders for object detection from point clouds. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12689-12697, 2018. 3", + "[15] Zixing Lei, Shunli Ren, Yue Hu, Wenjun Zhang, and Siheng Chen. Latency-aware collaborative perception. In Computer Vision – ECCV 2022, page 316–332, 2022. 3", + "[16] Jinlong Li, Runsheng Xu, Xinyu Liu, Jin Ma, Zicheng Chi, Jiaqi Ma, and Hongkai Yu. Learning for vehicle-to-vehicle cooperative perception under lossy communication. IEEE Transactions on Intelligent Vehicles, 8(4):2650-2660, 2023. 1, 3, 6", + "[17] Yifan Lu, Quanhao Li, Baoan Liu, Mehrdad Dianati, Chen Feng, Siheng Chen, and Yanfeng Wang. Robust collaborative 3d object detection in presence of pose errors. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 4812-4818. IEEE, 2023. 2, 3, 5, 6, 8", + "[18] Cong Ma, Lei Qiao, Chengkai Zhu, Kai Liu, Zelong Kong, Qing Li, Xueqi Zhou, Yuheng Kan, and Wei Wu. Holovic:large-scale dataset and benchmark for multi-sensor holographic intersection and vehicle-infrastructure cooperative. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 22129-22138, 2024. 8", + "[19] Xu Runsheng, Zhengzhong Tu Tu, Hao Xiang, Wei Shao, Bolei Zhou, and Jiaqi Ma. Cobevt: Cooperative bird's eye view semantic segmentation with sparse transformers. In Conference on Robot Learning (CoRL), 2022. 1, 6", + "[20] Shaoshuai Shi, Chaoxu Guo, Li Jiang, Zhe Wang, Jianping Shi, Xiaogang Wang, and Hongsheng Li. Pv-rcnn: Point-voxel feature set abstraction for 3d object detection. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10526–10535, 2019. 2", + "[21] Shaoshuai Shi, Chaoxu Guo, Li Jiang, Zhe Wang, Jianping Shi, Xiaogang Wang, and Hongsheng Li. PV-RCNN: Point-voxel feature set abstraction for 3d object detection. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10526-10535, 2019. 3", + "[22] Rui Song, Chenwei Liang, Hu Cao, Zhiran Yan, Walter Zimmer, Markus Gross, Andreas Festag, and Alois Knoll. Collaborative semantic occupancy prediction with hybrid feature fusion in connected automated vehicles. In 2024 IEEE/CVF International Conference on Computer Vision and Pattern Recognition (CVPR). IEEE/CVF, 2024. 1", + "[23] Shihao Wang, Ying-Hao Liu, Tiancai Wang, Ying Li, and Xiangyu Zhang. Exploring object-centric temporal modeling for efficient multi-view 3d object detection. 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 3598-3608, 2023. 2", + "[24] Tsun-Hsuan Wang, Sivaban Manivasagam, Ming Liang, Bin Yang, Wenyuan Zeng, and Raquel Urtasun. V2VNet: Vehicle-to-Vehicle Communication for Joint Perception and Prediction, page 605-621. Springer International Publishing, Cham, 2020. 1, 2, 3, 6", + "[25] Yan Xia, Yusheng Xu, Shuang Li, Rui Wang, Juan Du, Daniel Cremers, and Uwe Stilla. Soe-net: A self-attention and orientation encoding network for point cloud based place" + ], + "bbox": [ + 516, + 92, + 905, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "22304", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "recognition. In Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition, pages 11348-11357, 2021. 3", + "[26] Yan Xia, Mariia Gladkova, Rui Wang, Qianyun Li, Uwe Stilla, João F Henriques, and Daniel Cremers. Casspr: Cross attention single scan place recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 8461-8472, 2023. 3", + "[27] Li Xiang, Junbo Yin, Wei Li, Cheng-Zhong Xu, Ruigang Yang, and Jianbing Shen. DI-V2X: Learning domain-invariant representation for vehicle-infrastructure collaborative 3d object detection. AAAI, 2024. 6", + "[28] Runsheng Xu, Hao Xiang, Zhengzhong Tu, Xin Xia, Ming-Hsuan Yang, and Jiaqi Ma. V2x-vit: Vehicle-to-everything cooperative perception with vision transformer. In Computer Vision - ECCV 2022, page 107-124, Cham, 2022. Springer Nature Switzerland. 1, 2, 6", + "[29] Runsheng Xu, Hao Xiang, Xin Xia, Xu Han, Jinlong Li, and Jiaqi Ma. Opv2v: An open benchmark dataset and fusion pipeline for perception with vehicle-to-vehicle communication. 2022 International Conference on Robotics and Automation (ICRA), page 2583–2589, 2022. 1, 2, 3, 5, 6", + "[30] Runsheng Xu, Xin Xia, Jinlong Li, Hanzhao Li, Shuo Zhang, Zhengzhong Tu, Zonglin Meng, Hao Xiang, Xiaoyu Dong, Rui Song, Hongkai Yu, Bolei Zhou, and Jiaqi Ma. V2v4real: A real-world large-scale dataset for vehicle-to-vehicle cooperative perception. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), page 13712–13722, Vancouver, BC, Canada, 2023. IEEE. 1, 2", + "[31] Yan Yan, Yuxing Mao, and Bo Li. Second: Sparsely embedded convolutional detection. Sensors (Basel, Switzerland), 18, 2018. 3, 7", + "[32] Binh Yang, Wenjie Luo, and Raquel Urtasun. Pixor: Realtime 3d object detection from point clouds. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7652-7660, 2018. 3", + "[33] Hongbo Yin, Daxin Tian, Chunmian Lin, Xuting Duan, Jianshan Zhou, Dezong Zhao, and Dongpu Cao. V2VFormer ++: Multi-modal vehicle-to-vehicle cooperative perception via global-local transformer. IEEE Transactions on Intelligent Transportation Systems, page 1-14, 2023. 1, 2, 3", + "[34] Hongbo Yin, Daxin Tian, Chunmian Lin, Xuting Duan, Jianshan Zhou, Dezong Zhao, and Dongpu Cao. V2vformer++: Multi-modal vehicle-to-vehicle cooperative perception via global-local transformer. IEEE Transactions on Intelligent Transportation Systems, 25:2153-2166, 2024. 6", + "[35] Haibao Yu, Yizhen Luo, Mao Shu, Yiyi Huo, Zebang Yang, Yifeng Shi, Zhenglong Guo, Hanyu Li, Xing Hu, Jirui Yuan, and Zaiqing Nie. Dair-v2x: A large-scale dataset for vehicle-infrastructure cooperative 3d object detection. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), page 21329-21338, New Orleans, LA, USA, 2022. IEEE. 1, 5", + "[36] Haibao Yu, Yingjuan Tang, Enze Xie, Jilei Mao, Ping Luo, and Zaiqing Nie. Flow-based feature fusion for vehicleinfrastructure cooperative 3d object detection. In Advances in Neural Information Processing Systems, 2023. 3" + ], + "bbox": [ + 91, + 90, + 483, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[37] Haibao Yu, Wenxian Yang, Hongzhi Ruan, Zhenwei Yang, Yingjuan Tang, Xu Gao, Xin Hao, Yifeng Shi, Yifeng Pan, Ning Sun, Juan Song, Jirui Yuan, Ping Luo, and Zaiqing Nie. V2x-seq: A large-scale sequential dataset for vehicleinfrastructure cooperative perception and forecasting. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), page 5486-5495, Vancouver, BC, Canada, 2023. IEEE. 2, 3", + "[38] Yunshuang Yuan and Monika Sester. COMAP: A synthetic dataset for collective multi-agent perception of autonomous driving. The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, XLIII-B2-2021:255-263, 2021. 2", + "[39] Yunshuang Yuan and Monika Sester. Leveraging dynamic objects for relative localization correction in a connected autonomous vehicle network. ISPRS Annals of the Photogrammetry, Remote Sensing and Spatial Information Sciences, V-1-2022:101-109, 2022. 2", + "[40] Yunshuang Yuan and Monika Sester. StreamLTS: Query-based temporal-spatial lidar fusion for cooperative object detection. In Computer Vision - ECCV 2024 Workshops, 2024. 2, 3, 4, 5, 6, 7, 8", + "[41] Yunshuang Yuan, Hao Cheng, and Monika Sester. Keypoints-based deep feature fusion for cooperative vehicle detection of autonomous driving. IEEE Robotics and Automation Letters, 7(2):3054-3061, 2022. 1, 2, 3, 6", + "[42] Yunshuang Yuan, Hao Cheng, Michael Ying Yang, and Monika Sester. Generating evidential bev maps in continuous driving space. ISPRS Journal of Photogrammetry and Remote Sensing, 204:27-41, 2023. 1, 2, 3, 5", + "[43] Wu Zheng, Weiliang Tang, Sijin Chen, Li Jiang, and Chi-Wing Fu. CIA-SSD: Confident iou-aware single-stage object detector from point cloud. In AAAI, 2021. 3, 6", + "[44] Yin Zhou and Oncel Tuzel. Voxelnet: End-to-end learning for point cloud based 3d object detection. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4490-4499, 2017. 3", + "[45] Walter Zimmer, Christian Creß, Huu Tung Nguyen, and Alois C Knoll. Tumtraf intersection dataset: All you need for urban 3d camera-lidar roadside perception [best student paper award]. In 2023 IEEE Intelligent Transportation Systems ITSC. IEEE, 2023. 8" + ], + "bbox": [ + 516, + 90, + 903, + 681 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "22305", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/SparseAlign_ a Fully Sparse Framework for Cooperative Object Detection/e9f3f305-e62d-4aad-9255-49e76e43c9cf_model.json b/2025/SparseAlign_ a Fully Sparse Framework for Cooperative Object Detection/e9f3f305-e62d-4aad-9255-49e76e43c9cf_model.json new file mode 100644 index 0000000000000000000000000000000000000000..b59861734896abbf75592708b7c3990b0a76dc7b --- /dev/null +++ b/2025/SparseAlign_ a Fully Sparse Framework for Cooperative Object Detection/e9f3f305-e62d-4aad-9255-49e76e43c9cf_model.json @@ -0,0 +1,2090 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.122, + 0.131, + 0.877, + 0.153 + ], + "angle": 0, + "content": "SparseAlign: A Fully Sparse Framework for Cooperative Object Detection" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.187, + 0.78, + 0.21 + ], + "angle": 0, + "content": "Yunshuang Yuan\\(^{1}\\) Yan Xia\\(^{2,3\\dagger}\\) Daniel Cremers\\(^{2,3}\\) Monika Sester\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.149, + 0.212, + 0.848, + 0.247 + ], + "angle": 0, + "content": "\\(^{1}\\)Leibniz University Hannover \\(^{2}\\)Technical University of Munich \\(^{3}\\)Munich Center for Machine Learning (MCML) {yunshuang.yuan, monika.sester}@ikg.uni-hannover.de {yan.xia, cremers}@tum.de" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.281, + 0.327, + 0.296 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.313, + 0.486, + 0.601 + ], + "angle": 0, + "content": "Cooperative perception can increase the view field and decrease the occlusion of an ego vehicle, hence improving the perception performance and safety of autonomous driving. Despite the success of previous works on cooperative object detection, they mostly operate on dense Bird's Eye View (BEV) feature maps, which are computationally demanding and can hardly be extended to long-range detection problems. More efficient fully sparse frameworks are rarely explored. In this work, we design a fully sparse framework, SparseAlign, with three key features: an enhanced sparse 3D backbone, a query-based temporal context learning module, and a robust detection head specially tailored for sparse features. Extensive experimental results on both OPV2V and DAIR-V2X datasets show that our framework, despite its sparsity, outperforms the state of the art with less communication bandwidth requirements. In addition, experiments on the OPV2Vt and DAIR-V2Xt datasets for time-aligned cooperative object detection also show a significant performance gain compared to the baseline works." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.627, + 0.222, + 0.643 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.652, + 0.483, + 0.832 + ], + "angle": 0, + "content": "Cooperative Perception (CP) for autonomous driving has gained significant attention due to its potential to enhance road safety. Sharing Collective Perception Messages (CPMs) among Intelligent (traffic) Agents (IAs) in the vehicular network effectively reduces occlusions and expands the field of view, improving perception performance. In many recent works [1, 16, 24, 29, 35, 41, 42], CP has proven to be effective for different tasks, such as cooperative object detection [1, 24, 29, 41], Bird's Eye View (BEV) semantic segmentation [19, 42], and 3D occupancy semantic segmentation [22]. In this work, we focus on LiDAR-based cooperative vehicle detection." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.833, + 0.484, + 0.879 + ], + "angle": 0, + "content": "Training a cooperative 3D object detector is computationally demanding as it needs to process data from multiple vehicles, especially when the model is built to di" + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.283, + 0.905, + 0.361 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.366, + 0.905, + 0.466 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.479, + 0.908, + 0.536 + ], + "angle": 0, + "content": "Figure 1. a,b) Communication strategies for cooperative object detection. c-e) Issues of sparse convolutional backbone networks. c) Center Feature Missing (CFM). d,e) Isolated Convolution Field (ICF) caused by ring disconnectivity and occlusion." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.566, + 0.907, + 0.839 + ], + "angle": 0, + "content": "gest sequential data from several subsequent frames to learn the temporal context. Following Fcooper [1], recent works [6, 11, 28-30, 33, 35] encode the 3D point cloud into dense BEV feature maps for further processing, sharing, and fusion. However, the computational cost of this pipeline increases quadratically with the perception range. Although cooperation among IAs can improve long-range detection performance, the IAs are normally not in the best positions to optimize the performance gain. The long-range detection performance of the local detector directly influences the overall performance. Besides, sharing the BEV feature maps among IAs requires extensive communication resources. Instead, the sparse operations leverage the sparsity of point clouds and have only linear computational complexity relative to the number of points. In addition, the output sparse features can be more efficient for sharing. It is therefore beneficial to perform cooperative object detection based on sparse operations only." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.909, + 0.901 + ], + "angle": 0, + "content": "However, building a fully sparse framework to compete with the dense ones is non-trivial. The first unavoidable issue is Center Feature Missing [8] (CFM, shown in Fig. 1 c)). Due to the range view of on-board LiDARs, there are" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.887, + 0.236, + 0.9 + ], + "angle": 0, + "content": "† Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.946, + 0.521, + 0.957 + ], + "angle": 0, + "content": "22296" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.379 + ], + "angle": 0, + "content": "usually no scan points in the vehicles' center area. However, compared to the edge points, the center points have a better ability to represent the whole object. Lacking center points for learning, the sparse frameworks perform worse than the dense ones that directly learn the features for center points. The second issue is the poor connectivity between points scanned by different laser beams. As shown in Fig. 1 d), this disconnectivity issue in distant areas may lead to isolated voxel blobs; the receptive field only enlarges to the blob scale as the convolutional layers go deeper. The isolated voxel blobs never exchange information with each other, limiting the backbone's capability of capturing global features. We call this issue Isolated Convolution Field (ICF). It also happens in the occlusion areas as shown in Fig. 1 e): although the occluded vehicle has very few scan points (red), it is expected to be detected based on the neighboring vehicles and the clues in the previous frames. However, ICF introduces difficulties for these isolated points to aggregate information from neighbors and the global environment." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.381, + 0.483, + 0.653 + ], + "angle": 0, + "content": "To overcome the above problems, we build a Sparse UNet backbone network SUNet with the Coordinate-Expandable sparse Convolution (CEC) implemented in the MinkowskiEngine [5] library. We apply CEC on \\(4\\times\\) and \\(8\\times\\) downsampled 3D voxels to increase the connectivity between ICFs. On the 2D BEV features, the CECs are used to expand the sparse coordinates, assuring the center features for all scanned vehicles. On top of SUNet, we build TempAlign Module (TAM) to learn the temporal context from the historical frames in a query-based [23, 40] manner, PoseAlign Module (PAM) to correct the relative pose errors between the cooperative and the ego agents, and SpatialAlign Module (SAM) to fuse the shared object queries based on the corrected poses. In addition, we propose the CompassRose encoding for the regression of BBox directions to further improve performance on both Cooperative Object Detection (COOD) and the Time-Aligned COOD [40] (TA-COOD) tasks." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.657, + 0.473, + 0.671 + ], + "angle": 0, + "content": "To summarize, the main contributions of this work are:" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.675, + 0.48, + 0.715 + ], + "angle": 0, + "content": "- We investigate the issues of building a fully sparse network for LiDAR-based COOD and analyze how these issues impact the detection performance." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.72, + 0.48, + 0.792 + ], + "angle": 0, + "content": "- We propose a novel fully sparse network, SparseAlign, via addressing the ICF and CFM problems of the sparse 3D backbone SUNet with CECs. Additionally, we introduce the SAM module, specifically tailored to ensure accurate fusion of sparse features." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.796, + 0.48, + 0.868 + ], + "angle": 0, + "content": "- We assess the communication bandwidth and precision trade-offs of the different approaches, finding that our proposal reduces bandwidth consumption by up to \\(98\\%\\), compared to previous methods using dense BEV feature maps for LiDAR-based cooperative object detection." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.871, + 0.48, + 0.897 + ], + "angle": 0, + "content": "We conduct extensive experiments on OPV2V, DAIR-V2X, OPV2Vt, and DAIR-V2Xt benchmarks and show" + }, + { + "type": "list", + "bbox": [ + 0.092, + 0.675, + 0.48, + 0.897 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.528, + 0.092, + 0.905, + 0.121 + ], + "angle": 0, + "content": "that the proposed SparseAlign has a significant improvement over state-of-the-art methods." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.139, + 0.655, + 0.155 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.165, + 0.847, + 0.182 + ], + "angle": 0, + "content": "2.1. Cooperative Object Detection (COOD)" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.189, + 0.905, + 0.536 + ], + "angle": 0, + "content": "COOD frameworks are classified into sharing CPMs with raw data, intermediate features, and detections. Recent works [2, 24, 29, 38] have shown that intermediate feature sharing has the best performance due to its adjustable and learnable CPMs. We also use this strategy in our framework. In the perspective of communication, there are two strategies: handshaking and broadcasting as shown in Fig. 1 a) and b). With handshaking communication [11, 28-30, 33, 40, 42], the ego vehicle sends a request message that contains its pose to the cooperative IAs so that they can transform the local raw data to the ego coordinate system and then process their data to obtain the features for sharing. This simplifies the feature fusion process, as the features are directly learned in the goal coordinate system. In contrast, the broadcasting method [6, 17, 24, 37] directly processes the data in the local coordinate system and transmits it to other IAs. In the fusion stage, the ego vehicle must transform the cooperative features into the ego coordinate system and then conduct the fusion process. In comparison, broadcasting is more efficient as it does not require a prior communication agreement with the cooperative agents in highly dynamic driving scenarios where the communication partners change frequently." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.539, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Achieving COOD presents several challenges, including communication latency and bandwidth limitations, inaccurate IA poses, and sensor asynchrony. To reduce the communication bandwidth consumption, [1, 24, 29, 33] compress the learned BEV feature maps to fewer channels via convolution layers, while [13] utilizes the learned confidence map and [42] leverages the uncertainty map to select the most important information for sharing. Instead, [41] employs the two-stage detector PVRCNN [20] and shares the 3D keypoint features from the second stage. To be more compact, [4, 9, 40] share the object query features [23] which are highly flexible to blend all relevant information, such as scanning time and IA pose, about an object into one unique feature vector. In this work, we also leverage the object query features to learn the temporal and spatial context of the object. Pose errors can dramatically deteriorate the detection results [24]. To this end, [39, 41] leverage the predicted Bounding Boxes (BBoxes) and select semantically classified keypoints to align the relative pose between the ego and the cooperative vehicles. Instead, [17] solely uses the BBoxes for pose alignment via the Pose Graph Optimization (PGO). However, these methods require the initial pose errors to be small to find the correct object matching in the first step and then iterate many times, or use the PGO" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "22297" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.09, + 0.571, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.577, + 0.09, + 0.898, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.214, + 0.51, + 0.354 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.104, + 0.355, + 0.135, + 0.363 + ], + "angle": 0, + "content": "SUNet" + }, + { + "type": "image", + "bbox": [ + 0.511, + 0.214, + 0.702, + 0.354 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.515, + 0.356, + 0.622, + 0.366 + ], + "angle": 0, + "content": "TempAlign Module (TAM)" + }, + { + "type": "image", + "bbox": [ + 0.706, + 0.215, + 0.898, + 0.355 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.711, + 0.356, + 0.817, + 0.366 + ], + "angle": 0, + "content": "PoseAlign Module (PAM)" + }, + { + "type": "image_caption", + "bbox": [ + 0.358, + 0.383, + 0.638, + 0.398 + ], + "angle": 0, + "content": "Figure 2. Overview of SparseAlign framework" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.424, + 0.486, + 0.712 + ], + "angle": 0, + "content": "to refine the result. In this paper, we propose to learn pose-agnostic neighborhood graph features to accurately find the correct matching and then use PGO to refine it. Communication delay can cause that data arriving at the ego agent is outdated. To compensate for this delay, [36, 37] add the first-order derivatives of the BEV feature map to the original feature map to obtain the updated feature map that is aligned to the ego timestamp; [15] employs the LSTM [12] structure on the BEV feature maps to adjust the features to be aligned. Differently, [40] embeds the observation timestamps as well as the temporal context from historical frames into the object queries so that they are time- and motion-aware and capable of predicting the status at any given future timestamp, achieving the time-delay compensation. Different to [15, 36, 37] that compensate the feature maps, [40] leverages the point-wise timestamp of the point clouds to learn the accurate temporal context, hence is also able to cope with sensor asynchrony issues. Therefore, we adopt the query-based strategy for time-related compensation." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.725, + 0.238, + 0.74 + ], + "angle": 0, + "content": "2.2.3DBackbones" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.485, + 0.902 + ], + "angle": 0, + "content": "To extract point cloud features, most previous works [16, 17, 24, 29, 33, 37, 41] for COOD either use dense 3D [25, 44], dense 2D [14, 32] or semi-sparse [21, 26, 31, 43] backbones that first encode the point cloud with 3D sparse convolution and then apply dense 2D convolutions for further processing. In contrast, [40, 42] employ the MinkUnet [5] to extract point cloud features. Despite its efficiency, the MinkUnet backbone struggles with long-range detection, occlusions and large objects, because of CFM and ICF problems. [8] proposed center voting for each scanned ob" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.424, + 0.907, + 0.53 + ], + "angle": 0, + "content": "ject points to solve CFM issue and [3] reveals that further down-sampling the voxel features to lower resolutions can increase the receptive field and refine the large object detection. Instead of shifting the feature locations as [8] to accurately locate object queries or adding additional layers as [3], we build a backbone that solves both CFM and ICF issues by modifying MinkUnet with CEC layers." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.548, + 0.605, + 0.563 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.575, + 0.694, + 0.589 + ], + "angle": 0, + "content": "3.1. Task formalization" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.599, + 0.909, + 0.903 + ], + "angle": 0, + "content": "In this work, we solve the LiDAR-based cooperative object detection task considering unavoidable challenges in reality: communication delay, sensor asynchrony, and localization errors. For this scenario, we assume there are \\(N\\) IAs, including one ego agent \\(A_{1}\\) and \\(N - 1\\) cooperative agents. These agents are sharing CPMs with each other in a broadcasting manner, as shown in Fig. 1b). Specifically, in the \\(i\\)-th frame, \\(A_{1}\\) receives the CPMs \\(M = \\{(m_j,t_j)|j\\in \\{1,\\dots ,N - 1\\} \\}\\) from the cooperative agents, where \\(m_j\\) and \\(t_j\\) are the CPM and the corresponding timestamp from the \\(j\\)-th agent. We aim to optimize the object detection results based on the point clouds captured so far (\\(t\\leq t_i\\)) at \\(A_{1}\\) and the received messages \\(M\\). We model and evaluate both the COOD and TA-COOD tasks from the perspective of an ego agent. The key difference between COOD and TA-COOD lies in the GT BBoxes. As illustrated in Fig. 2, TAM, COOD GT BBoxes (dashed lines) align with the sequentially captured LiDAR points, whereas TA-COD GT BBoxes (solid lines) are referenced to a global timestamp \\(t_g\\). This distinction requires the model to accurately predict" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "22298" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.1, + 0.091, + 0.194, + 0.166 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.197, + 0.091, + 0.288, + 0.164 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.291, + 0.092, + 0.383, + 0.163 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.387, + 0.092, + 0.478, + 0.163 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.177, + 0.483, + 0.232 + ], + "angle": 0, + "content": "Figure 3. Receptive field (RF) after 4 sparse convolutions (convs). a. LiDAR points; b. Spare pixels (white) with a conv center point (red); c. RF coverage of 4 normal sparse convs (red); d. RF coverage of 4 CECs (red+light blue, darkest blue is background)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.247, + 0.484, + 0.322 + ], + "angle": 0, + "content": "BBox locations at \\( t_g \\). While COOD relies solely on point cloud geometry for location prediction, TA-COOD leverages temporal context to enhance accuracy at \\( t_g \\) and also enables the model to handle location errors due to communication latency." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.332, + 0.292, + 0.346 + ], + "angle": 0, + "content": "3.2. Framework overview" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.355, + 0.484, + 0.718 + ], + "angle": 0, + "content": "We build an efficient fully sparse framework SparseAlign, as shown in Fig. 2, to process sequential point clouds from multiple agents. All agents share the same network weights. The input at time \\( t_i \\) is the point cloud of the ego agent \\( PC_{t_i} \\) and of the cooperative agents \\( PC_{t_j} \\). Note that \\( t_j < t_i \\) when communication latency exists. All point clouds are processed by the backbone SUNet to extract the point cloud features. The Region of Interest (RoI) module then selects the most interesting features as the object queries. For simplicity, we define object queries as \\( Q^{*} = \\{(F_{i}^{*}, x_{i}^{*}, y_{i}^{*}) | i \\in \\{1, \\dots, N^{*}\\} \\), where * represents the query set name, \\( F \\) is the features, and \\( x, y \\) are the coordinates of the corresponding \\( i \\)-th query. With selected queries, the TempAlign Module (TAM) can efficiently align the query features to the globally aligned timestamp \\( t \\) so that errors introduced by communication delay and sensor asynchrony are compensated. The updated query features and the detections from the Local Query Detection head (LQDet) are then shared as CPMs. The PoseAlign Module (PAM) then uses the detected BBoxes to correct the relative poses between the cooperative and the ego agent and the SpatialAlign Module (SAM) fuses the received cooperative features into the ego coordinate system based on the corrected transformation \\( T_{c}^{e} \\) from cooperative to ego coordinates." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.727, + 0.379, + 0.743 + ], + "angle": 0, + "content": "3.3. Backbone: Sparse UNet (SUNet)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Processing data from multiple IAs for cooperative perception is computationally demanding. To improve efficiency, we use a fully sparse 3D UNet, similar to MinkUnet[5], to construct the SUNet (Fig. 2 bottom left). However, the fully sparse convolutions used in such networks suffer from the ICF issue. As shown in Fig. 3, the receptive field (RF) of (c) standard sparse convolutions only covers LiDAR points from a single vehicle. In contrast, the RF of (d) the CEC layers extends to the points of a neighboring vehicle, enabling learning within a larger spatial context. Therefore," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.258 + ], + "angle": 0, + "content": "we use 3D CECs in the sparse convolution (SConv) block 3 and 4. Correspondingly, we contract the coordinates in the transposed sparse convolution (TConv) block 3 and 2. In this way, the output of the TConv blocks \\( p4 \\) and \\( p2 \\) can match the coordinates in stride \\( x2 \\) and \\( x4 \\) for the concatenation. In addition, it learns better global features without voxel number soaring in the subsequent layers and maintains efficiency. Against the CFM issue, we use 2D CECs on the 2D BEV sparse features (Fig. 2 SUNet:Coor. Expand.) to expand the coordinates to ensure their coverage over the object centers." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.267, + 0.7, + 0.283 + ], + "angle": 0, + "content": "3.4. Alignment modules" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.289, + 0.907, + 0.576 + ], + "angle": 0, + "content": "TempAlign Module (TAM). The TAM aims to learn the temporal context relative to the previous frames to achieve two goals: improve object detection performance by looking for clues in the history and compensate for the object displacement error introduced by asynchronous scanning time (Fig. 2, TAM). We achieve the second goal by predicting the BBoxes at the globally aligned timestamp so that the BBoxes shared for PAM are synchronized. Then PAM only needs to deal with the object displacement caused by localization errors. Technically, the TAM follows [40], for efficiency, by removing the global cross-attention module which learns the global BEV features. We only use a memory queue and a hybrid attention module. The memory queue stores the top \\( k = 256 \\) object queries from previous \\( L = 4 \\) data frames so that the attention module can allow the interaction between object queries in the current frame and the historical frames. In this way, the object queries learn the temporal context from the sequential data and are able to make predictions for the future state." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.577, + 0.909, + 0.88 + ], + "angle": 0, + "content": "PoseAlign Module (PAM). Given the detected BBoxes \\(\\mathbf{B}_i\\) and \\(\\mathbf{B}_j\\), respectively, from the perspective of the agent \\(A_i\\) and \\(A_j\\) in the same scene, humans can easily find the correspondences between \\(\\mathbf{B}_i\\) and \\(\\mathbf{B}_j\\) based on the geometric relationships. Practically, we model the features of these geometric relationships by embedding the neighborhood geometry of each BBox into deep features and then match these features of \\(\\mathbf{B}_i\\) and \\(\\mathbf{B}_j\\) with the Hungarian algorithm. For each BBox \\(b\\) in a BBox set \\(\\mathbf{B}\\), we select the nearest \\(k = 8\\) neighbors to embed its neighborhood features. As shown in Fig. 2, PAM: 1.Graph Feature, the orientation angles of the BBox are \\(\\alpha\\) and \\(\\beta\\) for \\(b\\) and its neighbor, respectively. We embed the relative orientation with \\(\\nu_{a} = [\\sin (\\beta -\\alpha);\\cos (\\beta -\\alpha)]^{1}\\). Similarly, the relative edge direction is embedded as \\(\\epsilon_{a} = [\\sin (\\theta -\\alpha);\\cos (\\theta -\\alpha)]\\). In addition, the dimension \\(\\nu_{dim} = [l,w,h]\\) of the neighbor BBox, and the Euclidean distance between \\(b\\) and its neighbors \\(\\epsilon_d = d\\), are also embedded into the feature. Note that \\(\\epsilon_d,\\epsilon_a,\\nu_a\\) are all relative features between two BBoxes that are pose-agnostic, leading to robust pose alignment without" + }, + { + "type": "page_footnote", + "bbox": [ + 0.532, + 0.887, + 0.645, + 0.901 + ], + "angle": 0, + "content": "1[.;.]:concatenation" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22299" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.152 + ], + "angle": 0, + "content": "dependency on the magnitude of initial pose errors. These features are then concatenated and embedded into high-dimensional features by a linear layer with ReLU activation \\((\\mathbb{R}^5\\to \\mathbb{R}^{d_{emb}})\\):" + }, + { + "type": "equation", + "bbox": [ + 0.177, + 0.163, + 0.483, + 0.18 + ], + "angle": 0, + "content": "\\[\nf _ {n b r} = \\operatorname {l i n e a r} \\left(\\left[ \\epsilon_ {d}; \\epsilon_ {a}; \\nu_ {a}; \\nu_ {\\text {d i m}} \\right]\\right) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.189, + 0.485, + 0.265 + ], + "angle": 0, + "content": "Assembling the features relative to all neighbors, we summarize the neighborhood feature \\( F_{nbr} \\in \\mathbb{R}^{k \\times d_{emb}} \\) by a multi-head self-attention module \\( attn \\) followed by a linear layer operated on the mean and max of the \\( attn \\) outcome, resulting in the feature \\( F_{nbr} \\in \\mathbb{R}^{d_{emb}} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.14, + 0.274, + 0.483, + 0.291 + ], + "angle": 0, + "content": "\\[\nF _ {n b r} = \\operatorname {a t t n} \\left(F _ {n b r}\\right) \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.142, + 0.293, + 0.483, + 0.311 + ], + "angle": 0, + "content": "\\[\nF _ {n b r} = \\operatorname {l i n e a r} \\left(\\left[ \\operatorname {m e a n} \\left(F _ {n b r}\\right); \\operatorname {m a x} \\left(F _ {n b r}\\right) \\right]\\right) \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.32, + 0.484, + 0.501 + ], + "angle": 0, + "content": "By calculating the Euclidean distance between the learned neighborhood feature \\( F_{nbr} \\) of the BBox in \\( \\mathbf{B}_i \\) and \\( \\mathbf{B}_j \\), we construct the cost matrix for the linear sum assignment to find the best Match (Fig. 2, PAM: 2-match) between these two BBox sets. However, an object in \\( \\mathbf{B}_i \\) does not always have a projection in \\( \\mathbf{B}_j \\) and vice versa. Thus, we reject the wrong matches that have large distances. Based on the PAM: 2-match result, the relative pose transformation between \\( A_i \\) and \\( A_j \\) can be recovered, completing the PAM: 3Align step. Additionally, we use pose graph optimization to refine the alignment by leveraging loop closure among multiple cooperative agents posing following [17]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.501, + 0.484, + 0.621 + ], + "angle": 0, + "content": "SpatialAlign Module (SAM). To align the cooperative query features \\( Q^c \\) with the ego query features \\( Q^e \\), there are two issues to be resolved. First, the features \\( F^c \\) are learned in the cooperative coordinate system and must be adapted to the ego coordinate system. We achieve this using a multilayer perceptron (mlp) as described in Eq. (4) where \\( F^R \\in \\mathbb{R}^{N \\times 9} \\) is the flattened and repeated rotation matrix \\( R \\in \\mathbb{R}^{3 \\times 3} \\)." + }, + { + "type": "equation", + "bbox": [ + 0.215, + 0.63, + 0.483, + 0.649 + ], + "angle": 0, + "content": "\\[\nF ^ {c} = m l p \\left(\\left[ F ^ {c}; F ^ {R} \\right]\\right) \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.658, + 0.484, + 0.794 + ], + "angle": 0, + "content": "Second, after rotation with \\( R \\), the coordinates of \\( Q^c \\) (Fig. 2, SAM, yellow points) do not perfectly align to the discretized grid in the ego coordinate system. Thus, we merge \\( Q^c \\) into the nearest points of this grid. The locations of the fused output object query points \\( Q^{ts} \\) are shown with gray points. For each point in \\( Q^{ts} \\), we summarize the features of its \\( k = 8 \\) nearest neighbors (Fig. 2, SAM: black arrows) in query set \\( Q^{\\cup} = Q^{c}\\bigcup Q^{e} \\) to obtain the fused features \\( F^{ts} \\). Mathematically, it reads as" + }, + { + "type": "equation", + "bbox": [ + 0.117, + 0.803, + 0.483, + 0.822 + ], + "angle": 0, + "content": "\\[\nF _ {i j} = m l p \\left(\\left[ F _ {j} ^ {\\cup}; \\text {l i n e a r} \\left(\\left[ x _ {i} ^ {t s} - x _ {j} ^ {\\cup}, y _ {i} ^ {t s} - y _ {j} ^ {\\cup} \\right]\\right) \\right]\\right) \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.117, + 0.824, + 0.483, + 0.842 + ], + "angle": 0, + "content": "\\[\nF _ {i} ^ {t s} = \\left\\{F _ {i j} | j \\in \\{1, \\dots , k \\} \\right\\} \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.117, + 0.844, + 0.483, + 0.862 + ], + "angle": 0, + "content": "\\[\nF _ {i} ^ {t s} = \\max _ {j} \\left(F _ {i} ^ {t s}\\right) + \\text {m e a n} _ {j} \\left(F _ {i} ^ {t s}\\right) \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.902 + ], + "angle": 0, + "content": "where \\(i\\) is the query index of \\(Q^{ts}\\) and \\(j\\) is the neighbor index of the \\(i\\)-th query." + }, + { + "type": "image", + "bbox": [ + 0.526, + 0.091, + 0.9, + 0.175 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.608, + 0.191, + 0.813, + 0.206 + ], + "angle": 0, + "content": "Figure 4. CompassRose encoding" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.231, + 0.676, + 0.246 + ], + "angle": 0, + "content": "3.5. Detection Heads" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.254, + 0.907, + 0.466 + ], + "angle": 0, + "content": "To train SparseAlign, we follow [40] to attach a center-based object detection head to each of the RoI, TAM and SAM to learn query features. These heads are notated as RoIDet, LQDet and GQDet, respectively. They share the same target-encoding method. We use Focal loss to classify fore- and background object queries; all query points inside the ground-truth BBoxes are positive, and those outside are negative. Smooth \\(L1\\) loss is used for regressing \\([dx,dy,dz,l,w,h,\\mathcal{B}_{dir},\\mathcal{B}_{scr}]\\), where \\(l,w,h\\) are the dimensions of BBoxes and \\(dx,dy,dz\\) are coordinate offsets between the query points and the center of ground-truth BBoxes. \\(\\mathcal{B}_{dir},\\mathcal{B}_{scr}\\) are the novel CompassRose encoding (Fig. 4) of the ground truth orientation \\(r_g\\) with respect to four anchor angles \\(\\mathbf{r}_a = [0,0.5\\pi ,\\pi ,1.5\\pi ]\\). It reads as" + }, + { + "type": "equation", + "bbox": [ + 0.531, + 0.477, + 0.905, + 0.532 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {B} _ {\\mathrm {d i r}} = \\left[ \\cos r _ {g} - \\cos \\mathbf {r} _ {a}, \\sin r _ {g} - \\sin \\mathbf {r} _ {a} \\right] (8) \\\\ \\mathcal {B} _ {\\mathrm {s c r}} = \\left[ s _ {0}, s _ {0. 5 \\pi}, s _ {\\pi}, s _ {1. 5 \\pi} \\right] (9) \\\\ = \\mathbf {1} - \\arccos \\left(\\cos \\mathbf {r} _ {a} \\cdot \\cos r _ {g} + \\sin \\mathbf {r} _ {a} \\cdot \\sin r _ {g}\\right) / \\pi \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.543, + 0.906, + 0.604 + ], + "angle": 0, + "content": "Take the encoded sine wave (Fig. 4 left) as an example, the four anchor angles (yellow points) ensure that at least one anchor can reach the target angle monotonically, simplifying the regression process." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.613, + 0.698, + 0.629 + ], + "angle": 0, + "content": "3.6. Data augmentation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.636, + 0.907, + 0.773 + ], + "angle": 0, + "content": "In addition to geometric augmentation, which includes randomly flipping the input point cloud along the x- or y-axis and rotating it around the z-axis, we apply Free Space Augmentation (FSA) [42]. FSA enhances the point cloud by adding auxiliary points along LiDAR scanning rays, representing empty spaces where no objects are present. This augmentation helps mitigate the ICF problem in distant regions, where observation points are extremely sparse. For more details, refer to supplementary B." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.785, + 0.638, + 0.802 + ], + "angle": 0, + "content": "4. Experiment" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Datasets. We evaluate SparseAlign on OPV2V[29] and DAIR-V2X[35] datasets for COOD and their variant datasets OPV2Vt[40] and DAIR-V2Xt[40] for TA-COOD. OPV2V is a synthetic dataset created in different driving scenarios of CARLA [7] simulator. Following [29], the detection range is set to \\( x \\in [-140, 140]m \\), \\( y \\in [-40, 40]m \\)." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22300" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.097, + 0.09, + 0.48, + 0.24 + ], + "angle": 0, + "content": "
MethodBW(Mb) ↓AP0.5 ↑AP0.7↑
Fcooper[1]72.080.8870.790
V2VNet[24]72.080.9170.822
FPVRCNN[41]0.240.8730.820
OPV2V[29]126.80.9050.815
CoAlign[17]72.080.9020.833
CoBEVT[19]72.080.9130.861
V2VAM[16]72.080.9160.850
V2VFformer++[34]L+C72.080.9350.895
SparseAlign (ours)< 1.30.9300.892
" + }, + { + "type": "table_caption", + "bbox": [ + 0.123, + 0.251, + 0.453, + 0.266 + ], + "angle": 0, + "content": "Table 1. COOD Average Precision on OPV2V dataset. \\( {}^{2} \\)" + }, + { + "type": "table", + "bbox": [ + 0.097, + 0.292, + 0.48, + 0.408 + ], + "angle": 0, + "content": "
MethodOPV2VtDAIR-V2Xt
AP0.5↑AP0.7↑AP0.5↑AP0.7↑
Fcooper[1]+SA0.7630.5530.5970.282
FPVRCNN[43]+SA0.6400.4740.5980.307
OPV2V[29]+SA0.8810.7870.7020.366
StreamLTS[40]0.8530.7210.6400.404
StreamLTS[40]+SA0.8500.7900.7570.497
SparseAlign (full)0.8930.8180.7960.548
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.419, + 0.483, + 0.446 + ], + "angle": 0, + "content": "Table 3. TA-COOD Average Precision on OPV2Vt and DAIR-V2Xt" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.478, + 0.483, + 0.629 + ], + "angle": 0, + "content": "DAIR-V2X is a real dataset containing one vehicle- and one infrastructure-side sensor suit. The detection range is set to \\( x \\in [-100, 100]m, y \\in [-40, 40]m \\). OPV2Vt is generated by interpolating OPV2V data frames into sub-frames and replaying them to mimic the sensor asynchrony and the rolling shutter problem of LiDARs. The GT BBoxes of this dataset are aligned to the global timestamp, the LiDAR scan ending time of each frame. DAIR-V2Xt follows the same idea. The detection range of OPV2Vt and DAIR-V2Xt is set the same as OPV2V and DAIR-V2X." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.631, + 0.483, + 0.859 + ], + "angle": 0, + "content": "Implementation details. All models are trained on a single GTX-4090 and an Intel i7-8700 CPU for 50 epochs (about 35 hours) with the Adam optimizer and batch size two. The learning rate is set to \\(2e - 4\\) with a warm-up stage of 4000 iterations. The random rotation angle is set to the range \\([-90^{\\circ}, 90^{\\circ}]\\). For models incorporating the TAM, we load four sequential frames of data during training and calculate losses solely for the last frame. To reduce GPU memory consumption, we enable gradient calculations only for data from a single ego agent and one cooperative agent. For details on efficient gradient scheduling, refer to supplementary \\(D\\). To ensure that the model generalizes well to the significant time latency introduced by communication, we load data from cooperative vehicles with a random latency ranging from 0 to \\(200~\\mathrm{ms}\\) relative to the ego vehicles." + }, + { + "type": "table", + "bbox": [ + 0.543, + 0.09, + 0.902, + 0.242 + ], + "angle": 0, + "content": "
MethodBW(Mb) ↓AP0.5 ↑AP0.7↑
Fcooper[1]48.80.7340.559
V2VNet[24]48.80.6640.402
FPVRCNN[41]0.240.6650.505
V2XViT[28]48.80.7040.531
DiscoNet[19]48.80.7360.583
OPV2V[29]97.60.7330.553
CoAlign[17]48.80.7460.604
DI-V2X[27]48.80.7880.662
SparseAlign (ours)< 1.30.8450.685
" + }, + { + "type": "table_caption", + "bbox": [ + 0.547, + 0.253, + 0.895, + 0.267 + ], + "angle": 0, + "content": "Table 2. COOD Average Precision on DAIR-V2X dataset." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.293, + 0.602, + 0.308 + ], + "angle": 0, + "content": "5. Results" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.317, + 0.813, + 0.334 + ], + "angle": 0, + "content": "5.1. Comparison to the state-of-the-art" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.34, + 0.905, + 0.597 + ], + "angle": 0, + "content": "Tab. 1 presents the COOD results on the OPV2V dataset. All methods use only LiDAR data, except for V2VFormer++ (denoted by \\(\\mathrm{L + C}\\)), which incorporates both LiDAR and camera data. Compared to the state-of-the-art, our SparseAlign achieves the best performance in LiDAR-based COOD, approaching the results of V2VFormer++, which benefits from LiDAR-Camera fusion. Tab. 2 shows the results on the DAIR-V2X dataset. Here, SparseAlign outperforms other methods by a large margin, achieving a \\(2.3\\%\\) improvement in AP at the IoU threshold of 0.7 and a \\(5.7\\%\\) improvement at the IoU threshold of 0.5. This is mainly due to the enhanced temporal and global reasoning ability of our model for detecting distant and large objects. Compared to other methods that rely on dense BEV feature maps, SparseAlign consumes the least communication bandwidth (BW) of less than 1.3MB (without compression) thanks to the efficient query-based operations." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.598, + 0.906, + 0.732 + ], + "angle": 0, + "content": "For the results of the more challenging TA-COOD task shown in Tab. 3, our SparseAlign outperforms StreamLTS by a great margin. By applying our three alignment modules (StreamLTS + SA), the performance of StreamLTS is also improved. In addition, we compare our framework to Fcooper, FPVRCNN and OPV2V method by attaching the three alignment modules to them to achieve TA-COOD. All show worse performance than the full SparseAlign because of worse 3D backbone or the RoI selection." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.733, + 0.906, + 0.78 + ], + "angle": 0, + "content": "The qualitative results and the additional comparison to methods, such as TransIFF[4], on the 3D AP metric are given in the supplementary A and E, respectively." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.787, + 0.665, + 0.803 + ], + "angle": 0, + "content": "5.2. Ablation study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.906, + 0.901 + ], + "angle": 0, + "content": "We conduct an ablation study on the 3D backbone, TAM, and SAM (PAM is discussed separately in Sec. 5.4). Among these components, the spatial fusion module plays a crucial role in multi-agent feature fusion. Without SAM, the module performs the worst, as evidenced by the first row of Tab. 4. When using QUEST [9] spatial fusion, which" + }, + { + "type": "page_footnote", + "bbox": [ + 0.091, + 0.875, + 0.483, + 0.901 + ], + "angle": 0, + "content": "2The results reported in this table use frame-wise local score sorting for AP, same to the official benchmark. Other results use global sorting." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.958 + ], + "angle": 0, + "content": "22301" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.097, + 0.09, + 0.538, + 0.216 + ], + "angle": 0, + "content": "
CEC 2DCEC 3DTAMSAMOPV2VDAIR-V2X
AP0.5AP0.7AP0.5AP0.7
✓✓✓0.8470.7560.6710.528
✓✓✓QST0.8880.8690.7490.622
✓0.9240.8850.7730.638
✓✓0.9400.9010.7960.662
✓✓✓0.9500.9140.8130.682
✓✓✓✓0.9510.9290.8450.685
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.226, + 0.544, + 0.255 + ], + "angle": 0, + "content": "Table 4. Ablation study on 2D and 3D CEC layers, TAM and SAM. QST: use QUEST [9] spatial fusion instead of SAM." + }, + { + "type": "table", + "bbox": [ + 0.581, + 0.09, + 0.905, + 0.189 + ], + "angle": 0, + "content": "
Dir.Enc.OPV2VDAIR-V2X
AP0.5AP0.7AP0.5AP0.7
gt-angle0.8710.7680.6200.427
second0.9080.8740.7430.590
sin_cos0.9150.8810.7480.603
compass0.9240.8850.7730.638
" + }, + { + "type": "table_caption", + "bbox": [ + 0.575, + 0.199, + 0.907, + 0.257 + ], + "angle": 0, + "content": "Table 5. Comparison on BBox angle encoding methods: offset to ground-truth angle (gt-angle), second-style (second), sine-cosine offsets (sin_cos) and compass rose (compass)." + }, + { + "type": "table", + "bbox": [ + 0.117, + 0.27, + 0.885, + 0.407 + ], + "angle": 0, + "content": "
DatasetOPV2VtDAIR-V2Xt
Latency0ms100ms200ms0ms100ms200ms
AP threshold0.50.70.50.70.50.70.50.70.50.70.50.7
StreamLTS [40]0.8530.7210.8160.6800.7870.6470.6420.4040.6130.3790.5900.364
SA no TAM0.8900.8020.8560.4710.6310.2090.7200.4570.6720.3930.6350.350
SA no PT0.8350.6150.8310.6250.8150.5810.7600.4690.7370.4590.7220.449
SA no TL0.8850.8430.8620.6190.7230.2790.7930.5340.7490.4740.6980.423
SA no FSA0.8810.8160.8750.7950.8570.7620.7690.5080.7550.5020.7400.496
SparseAlign (SA)0.8930.8180.8850.7950.8670.7650.7960.5480.7860.5430.7720.532
" + }, + { + "type": "table_caption", + "bbox": [ + 0.092, + 0.419, + 0.904, + 0.433 + ], + "angle": 0, + "content": "Table 6. AP of TA-COOD with communication latency. PT: Point-wise Timestamps, TL: Train Latency, FSA: Free Space Augmentation." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.46, + 0.483, + 0.565 + ], + "angle": 0, + "content": "aligns only rotation (Eq. (4)) without correcting spatial displacement (Eqs. (5) to (7)), the AP scores remain significantly lower than those obtained with SAM. Incorporating SAM improves performance, and we further enhance it by gradually replacing standard sparse convolutional layers with CEC layers. Finally, TAM further boosts AP scores by leveraging temporal information from historical frames." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.566, + 0.483, + 0.657 + ], + "angle": 0, + "content": "In Tab. 5, we compare CompassRose encoding with three methods: directly regress the ground-truth angle (gt-angle), second-style [31] encoding (second), and the sine-cosine encoding (sin_cos). For simplicity, we use the configurations in the third row of Tab. 4 for the comparison. The results show that CompassRose performs the best." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.666, + 0.483, + 0.683 + ], + "angle": 0, + "content": "5.3. Communication latency and sensor asynchrony" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Tab. 6 shows the results of TA-COOD on the OPV2Vt and the DAIR-V2Xt dataset. Our proposed SparseAlign demonstrates a significant improvement over the baseline framework, StreamLTS [40]. Specifically, the AP at an IoU threshold of 0.7 increased \\(9.7\\%\\) on the OPV2Vt dataset and \\(14.4\\%\\) on the DAIR-V2Xt dataset. As the communication latency increases to 200ms, the AP at IoU 0.5 of our SparseAlign only dropped \\(2.6\\%\\) while the baseline work has lost \\(6.6\\%\\) of its accuracy. This reveals that our framework is more robust against communication latency. To investigate modules or configurations that are making an effect on this robustness, we conducted an ablation study by removing the TAM, Point-wise Timestamps (TP), Training Latency (TL, communication latency during training) and" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.46, + 0.822, + 0.474 + ], + "angle": 0, + "content": "Free Space Augmentation (FSA), respectively." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.476, + 0.907, + 0.641 + ], + "angle": 0, + "content": "Without the TAM, the performance noticeably declines as latency increases, despite achieving comparable results to the full SpareAlign model at \\(0\\mathrm{ms}\\) latency. A similar effect is observed when the full SA model is trained without exposure to data with communication latency (SA without TL). Remarkably, without TL, the model outperforms the full SpareAlign model in terms of AP at an IoU threshold of 0.7 on the OPV2Vt dataset (0.843 against 0.818). This is because the model focuses more on predicting BBoxes within the same frame, where only minimal latency is introduced by asynchronous sensors." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.641, + 0.909, + 0.779 + ], + "angle": 0, + "content": "Without PT, the model performs much worse than the full SpareAlign, showing the importance of fine-grained point-wise timestamps in the point clouds for the model to compensate for the errors introduced by sensor asynchrony and capture the accurate temporal context for prediction. Finally, the ablation study on the FSA module also shows a positive influence on the performance. For example, AP at the IoU of 0.7 increased by \\(4\\%\\) on DAIR-V2Xt by utilizing the FSA (See supplementary D for visual comparison)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.788, + 0.698, + 0.802 + ], + "angle": 0, + "content": "5.4. Localization errors" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Fig. 5 shows the performance of SpareAlign against different localization errors. We add random errors to the poses of both ego \\((T_e)\\) and cooperative \\((T_c)\\) vehicles. The translation errors \\(x_{\\epsilon}, y_{\\epsilon}\\) along the \\(x\\)- and \\(y\\)-axis and the rotation error \\(r_{\\epsilon}\\) around the \\(z\\)-axis are all assumed to be normally distributed with \\(\\mathcal{N}(0,1) \\cdot \\epsilon\\), where \\(\\epsilon \\in [0,1]\\) is the error scaling factor." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22302" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.118, + 0.093, + 0.457, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.116, + 0.226, + 0.456, + 0.355 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.371, + 0.483, + 0.414 + ], + "angle": 0, + "content": "Figure 5. AP at IoU threshold of 0.7 with translation errors ranging from \\(0\\mathrm{m}\\) to \\(1\\mathrm{m}\\) along x- and y-axis, and rotation errors from \\(0^{\\circ}\\) to \\(1.0^{\\circ}\\) (horizontal axis) for the different datasets." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.441, + 0.484, + 0.56 + ], + "angle": 0, + "content": "We report the results by gradually increasing \\(\\epsilon\\) with a step size of 0.2. Note that the translation and rotation errors exist in the poses of both the ego and the cooperative vehicle; the resulting relative translation error of the transformation \\(T_{c}^{e} = inverse(T_{e})\\cdot T_{c}\\) from cooperative to ego coordinate system will be amplified as the rotation error increases. Our PAM is specially designed to mitigate the influence of large relative transformation errors." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.563, + 0.484, + 0.82 + ], + "angle": 0, + "content": "Our proposed PAM and the baseline method CoAlign [17] both significantly reduce the impact of pose errors on AP performance compared to the configuration lacking any correction module. In addition, our method demonstrates greater robustness against large errors. For instance, on the OPV2V dataset, the AP for CoAlign decreased by approximately \\(27\\%\\) at the error of \\(1.0m, 1.0m, 1.0^{\\circ}\\), whereas PAM experienced only a \\(12\\%\\) drop. This validates the efficacy of our design, which relies on the pose-agnostic relative neighborhood geometries to match the detected BBoxes of the ego and the cooperative IAs. Note that the DAIR-V2X dataset exhibits smaller changes in AP as localization errors increase, because the pose parameters are not well calibrated. In contrast, [40] refined the poses of DAIR-V2X for generating the DAIR-V2Xt dataset, making the better-calibrated DAIR-V2Xt more sensitive to newly introduced errors." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.832, + 0.21, + 0.847 + ], + "angle": 0, + "content": "5.5. CPM sizes" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Our framework significantly reduces communication bandwidth without relying on computationally intensive dense feature maps, unlike state-of-the-art methods. This effi" + }, + { + "type": "image", + "bbox": [ + 0.541, + 0.094, + 0.71, + 0.221 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.712, + 0.094, + 0.882, + 0.221 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.236, + 0.905, + 0.278 + ], + "angle": 0, + "content": "Figure 6. CPM sizes with different information selection scores. Object query features with detection scores larger than the threshold (x-axis) are shared as CPMs." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.31, + 0.907, + 0.552 + ], + "angle": 0, + "content": "ciency is achieved by selecting only the top \\( K = 1024 \\) object queries for further processing. In addition to this intrinsic reduction, the size of the CPMs can be further minimized through score-based selection. Specifically, we employ the query-based detection scores from the LQDet head to identify and share only the most critical queries. As Fig. 6 shows, using a CPM score threshold below 0.5 does not lead to a noticeable degradation in performance across all datasets. Remarkably, at a threshold score of 0.5, the average size of the CPMs is reduced to less than \\( 400\\mathrm{KB} \\) on all datasets. The final CPM sizes after selection are influenced by the driving scenes. Specifically, scenes with more vehicles tend to have larger CPM sizes. Consequently, the OPV2V and OPV2Vt datasets, which contain more vehicles in their scenes, result in larger CPM sizes compared to the DAIR-V2X and DAIR-V2Xt datasets." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.572, + 0.634, + 0.587 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.599, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In this paper, we proposed SparseAlign, a fully sparse framework for cooperative object detection (COOD). With our novel 3D backbone network SUNet, the Compass-Rose encoding of detection head, the temporal and spatial fusion module TAM and SAM, SparseAlign significantly outperforms state-of-the-art methods on the OPV2V and DAIR-V2X datasets for COOD, as well as on their variants OPV2Vt and DAIR-V2Xt for time-aligned COOD (TA-COOD), achieving this with very low communication bandwidth. Additionally, our proposed pose alignment module PAM demonstrates improved robustness to large localization errors compared to the baseline method CoAlign. Nevertheless, when the pose errors are large, a drop of about \\(10\\%\\) in APs persists even with pose correction. In future work, we therefore plan to explore more robust methods, e.g. leveraging sub-graph detection, to further enhance the pose alignment accuracy. Additionally, more application scenarios, such as roadside cooperative detection benchmarks HoloVIC[18], RCooper[10] and TUMTraf[45], could also be explored." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "22303" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.171 + ], + "angle": 0, + "content": "[1] Qi Chen. F-cooper: feature based cooperative perception for autonomous vehicle edge computing system using 3d point clouds. Proceedings of the 4th ACM/IEEE Symposium on Edge Computing, 2019. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.172, + 0.484, + 0.242 + ], + "angle": 0, + "content": "[2] Qi Chen, Sihai Tang, Qing Yang, and Song Fu. Cooper: Cooperative perception for connected autonomous vehicles based on 3d point clouds. 2019 IEEE 39th International Conference on Distributed Computing Systems (ICDCS), pages 514-524, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.242, + 0.483, + 0.31 + ], + "angle": 0, + "content": "[3] Yukang Chen, Jianhui Liu, Xiangyu Zhang, Xiaojuan Qi, and Jiaya Jia. Voxelnext: Fully sparse voxelnet for 3d object detection and tracking. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 21674-21683, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.311, + 0.483, + 0.381 + ], + "angle": 0, + "content": "[4] Ziming Chen, Yifeng Shi, and Jinrang Jia. TransIFF: An instance-level feature fusion framework for vehicleinfrastructure cooperative 3d detection with transformers. In 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 18159-18168, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.382, + 0.483, + 0.45 + ], + "angle": 0, + "content": "[5] Christopher Bongsoo Choy, JunYoung Gwak, and Silvio Savarese. 4d spatio-temporal convnets: Minkowski convolutional neural networks. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3070-3079, 2019. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.452, + 0.483, + 0.521 + ], + "angle": 0, + "content": "[6] Jiaxun Cui, Hang Qiu, Dian Chen, Peter Stone, and Yuke Zhu. Coopernaut: End-to-end driving with cooperative perception for networked vehicles. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), page 17252-17262, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.522, + 0.483, + 0.578 + ], + "angle": 0, + "content": "[7] Alexey Dosovitskiy, German Ros, Felipe Codevilla, Antonio Lopez, and Vladlen Koltun. CARLA: An open urban driving simulator. In Proceedings of the 1st Annual Conference on Robot Learning, pages 1-16, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.579, + 0.482, + 0.606 + ], + "angle": 0, + "content": "[8] Lue Fan, Feng Wang, Naiyan Wang, and Zhaoxiang Zhang. Fully Sparse 3D Object Detection. In NeurIPS, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.607, + 0.483, + 0.662 + ], + "angle": 0, + "content": "[9] Siqi Fan, Haibao Yu, Wenxian Yang, Jirui Yuan, and Zaiqing Nie. Quest: Query stream for practical cooperative perception. In IEEE International Conference on Robotics and Automation (ICRA), 2024. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.663, + 0.484, + 0.746 + ], + "angle": 0, + "content": "[10] Ruiyang Hao, Siqi Fan, Yingru Dai, Zhenlin Zhang, Chenxi Li, Yuntian Wang, Haibao Yu, Wenxian Yang, Yuan Jirui, and Zaiqing Nie. Rcooper: A real-world large-scale dataset for roadside cooperative perception. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 22347-22357, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.747, + 0.483, + 0.803 + ], + "angle": 0, + "content": "[11] Caizhen He, Hai Wang, Long Chen, Tong Luo, and Yingfeng Cai. V2x-ahd:vehicle-to-everything cooperation perception via asymmetric heterogenous distillation network. arXiv, arXiv:2310.06603, 2023. arXiv:2310.06603 [cs]. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.804, + 0.483, + 0.831 + ], + "angle": 0, + "content": "[12] Sepp Hochreiter and Jürgen Schmidhuber. Long short-term memory. Neural Computation, 9(8):1735-1780, 1997. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.832, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[13] Yue Hu, Shaoheng Fang, Zixing Lei, Yiqi Zhong, and Siheng Chen. Where2comm: Communication-efficient collaborative perception via spatial confidence maps. In Thirty-sixth Conference on Neural Information Processing Systems (Neurips), 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.162 + ], + "angle": 0, + "content": "[14] Alex H. Lang, Sourabh Vora, Holger Caesar, Lubing Zhou, Jiong Yang, and Oscar Beijbom. Pointpillars: Fast encoders for object detection from point clouds. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12689-12697, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.164, + 0.906, + 0.207 + ], + "angle": 0, + "content": "[15] Zixing Lei, Shunli Ren, Yue Hu, Wenjun Zhang, and Siheng Chen. Latency-aware collaborative perception. In Computer Vision – ECCV 2022, page 316–332, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.208, + 0.906, + 0.275 + ], + "angle": 0, + "content": "[16] Jinlong Li, Runsheng Xu, Xinyu Liu, Jin Ma, Zicheng Chi, Jiaqi Ma, and Hongkai Yu. Learning for vehicle-to-vehicle cooperative perception under lossy communication. IEEE Transactions on Intelligent Vehicles, 8(4):2650-2660, 2023. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.278, + 0.906, + 0.348 + ], + "angle": 0, + "content": "[17] Yifan Lu, Quanhao Li, Baoan Liu, Mehrdad Dianati, Chen Feng, Siheng Chen, and Yanfeng Wang. Robust collaborative 3d object detection in presence of pose errors. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 4812-4818. IEEE, 2023. 2, 3, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.349, + 0.906, + 0.432 + ], + "angle": 0, + "content": "[18] Cong Ma, Lei Qiao, Chengkai Zhu, Kai Liu, Zelong Kong, Qing Li, Xueqi Zhou, Yuheng Kan, and Wei Wu. Holovic:large-scale dataset and benchmark for multi-sensor holographic intersection and vehicle-infrastructure cooperative. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 22129-22138, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.434, + 0.906, + 0.49 + ], + "angle": 0, + "content": "[19] Xu Runsheng, Zhengzhong Tu Tu, Hao Xiang, Wei Shao, Bolei Zhou, and Jiaqi Ma. Cobevt: Cooperative bird's eye view semantic segmentation with sparse transformers. In Conference on Robot Learning (CoRL), 2022. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.491, + 0.906, + 0.56 + ], + "angle": 0, + "content": "[20] Shaoshuai Shi, Chaoxu Guo, Li Jiang, Zhe Wang, Jianping Shi, Xiaogang Wang, and Hongsheng Li. Pv-rcnn: Point-voxel feature set abstraction for 3d object detection. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10526–10535, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.561, + 0.906, + 0.631 + ], + "angle": 0, + "content": "[21] Shaoshuai Shi, Chaoxu Guo, Li Jiang, Zhe Wang, Jianping Shi, Xiaogang Wang, and Hongsheng Li. PV-RCNN: Point-voxel feature set abstraction for 3d object detection. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10526-10535, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.632, + 0.906, + 0.716 + ], + "angle": 0, + "content": "[22] Rui Song, Chenwei Liang, Hu Cao, Zhiran Yan, Walter Zimmer, Markus Gross, Andreas Festag, and Alois Knoll. Collaborative semantic occupancy prediction with hybrid feature fusion in connected automated vehicles. In 2024 IEEE/CVF International Conference on Computer Vision and Pattern Recognition (CVPR). IEEE/CVF, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.718, + 0.906, + 0.786 + ], + "angle": 0, + "content": "[23] Shihao Wang, Ying-Hao Liu, Tiancai Wang, Ying Li, and Xiangyu Zhang. Exploring object-centric temporal modeling for efficient multi-view 3d object detection. 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 3598-3608, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.788, + 0.906, + 0.858 + ], + "angle": 0, + "content": "[24] Tsun-Hsuan Wang, Sivaban Manivasagam, Ming Liang, Bin Yang, Wenyuan Zeng, and Raquel Urtasun. V2VNet: Vehicle-to-Vehicle Communication for Joint Perception and Prediction, page 605-621. Springer International Publishing, Cham, 2020. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.859, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[25] Yan Xia, Yusheng Xu, Shuang Li, Rui Wang, Juan Du, Daniel Cremers, and Uwe Stilla. Soe-net: A self-attention and orientation encoding network for point cloud based place" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22304" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.484, + 0.134 + ], + "angle": 0, + "content": "recognition. In Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition, pages 11348-11357, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.484, + 0.207 + ], + "angle": 0, + "content": "[26] Yan Xia, Mariia Gladkova, Rui Wang, Qianyun Li, Uwe Stilla, João F Henriques, and Daniel Cremers. Casspr: Cross attention single scan place recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 8461-8472, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.208, + 0.484, + 0.262 + ], + "angle": 0, + "content": "[27] Li Xiang, Junbo Yin, Wei Li, Cheng-Zhong Xu, Ruigang Yang, and Jianbing Shen. DI-V2X: Learning domain-invariant representation for vehicle-infrastructure collaborative 3d object detection. AAAI, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.264, + 0.483, + 0.332 + ], + "angle": 0, + "content": "[28] Runsheng Xu, Hao Xiang, Zhengzhong Tu, Xin Xia, Ming-Hsuan Yang, and Jiaqi Ma. V2x-vit: Vehicle-to-everything cooperative perception with vision transformer. In Computer Vision - ECCV 2022, page 107-124, Cham, 2022. Springer Nature Switzerland. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.334, + 0.483, + 0.404 + ], + "angle": 0, + "content": "[29] Runsheng Xu, Hao Xiang, Xin Xia, Xu Han, Jinlong Li, and Jiaqi Ma. Opv2v: An open benchmark dataset and fusion pipeline for perception with vehicle-to-vehicle communication. 2022 International Conference on Robotics and Automation (ICRA), page 2583–2589, 2022. 1, 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.406, + 0.483, + 0.502 + ], + "angle": 0, + "content": "[30] Runsheng Xu, Xin Xia, Jinlong Li, Hanzhao Li, Shuo Zhang, Zhengzhong Tu, Zonglin Meng, Hao Xiang, Xiaoyu Dong, Rui Song, Hongkai Yu, Bolei Zhou, and Jiaqi Ma. V2v4real: A real-world large-scale dataset for vehicle-to-vehicle cooperative perception. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), page 13712–13722, Vancouver, BC, Canada, 2023. IEEE. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.504, + 0.483, + 0.546 + ], + "angle": 0, + "content": "[31] Yan Yan, Yuxing Mao, and Bo Li. Second: Sparsely embedded convolutional detection. Sensors (Basel, Switzerland), 18, 2018. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.548, + 0.483, + 0.603 + ], + "angle": 0, + "content": "[32] Binh Yang, Wenjie Luo, and Raquel Urtasun. Pixor: Realtime 3d object detection from point clouds. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7652-7660, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.605, + 0.483, + 0.675 + ], + "angle": 0, + "content": "[33] Hongbo Yin, Daxin Tian, Chunmian Lin, Xuting Duan, Jianshan Zhou, Dezong Zhao, and Dongpu Cao. V2VFormer ++: Multi-modal vehicle-to-vehicle cooperative perception via global-local transformer. IEEE Transactions on Intelligent Transportation Systems, page 1-14, 2023. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.676, + 0.483, + 0.745 + ], + "angle": 0, + "content": "[34] Hongbo Yin, Daxin Tian, Chunmian Lin, Xuting Duan, Jianshan Zhou, Dezong Zhao, and Dongpu Cao. V2vformer++: Multi-modal vehicle-to-vehicle cooperative perception via global-local transformer. IEEE Transactions on Intelligent Transportation Systems, 25:2153-2166, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.747, + 0.483, + 0.843 + ], + "angle": 0, + "content": "[35] Haibao Yu, Yizhen Luo, Mao Shu, Yiyi Huo, Zebang Yang, Yifeng Shi, Zhenglong Guo, Hanyu Li, Xing Hu, Jirui Yuan, and Zaiqing Nie. Dair-v2x: A large-scale dataset for vehicle-infrastructure cooperative 3d object detection. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), page 21329-21338, New Orleans, LA, USA, 2022. IEEE. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.845, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[36] Haibao Yu, Yingjuan Tang, Enze Xie, Jilei Mao, Ping Luo, and Zaiqing Nie. Flow-based feature fusion for vehicleinfrastructure cooperative 3d object detection. In Advances in Neural Information Processing Systems, 2023. 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.203 + ], + "angle": 0, + "content": "[37] Haibao Yu, Wenxian Yang, Hongzhi Ruan, Zhenwei Yang, Yingjuan Tang, Xu Gao, Xin Hao, Yifeng Shi, Yifeng Pan, Ning Sun, Juan Song, Jirui Yuan, Ping Luo, and Zaiqing Nie. V2x-seq: A large-scale sequential dataset for vehicleinfrastructure cooperative perception and forecasting. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), page 5486-5495, Vancouver, BC, Canada, 2023. IEEE. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.205, + 0.905, + 0.273 + ], + "angle": 0, + "content": "[38] Yunshuang Yuan and Monika Sester. COMAP: A synthetic dataset for collective multi-agent perception of autonomous driving. The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, XLIII-B2-2021:255-263, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.275, + 0.905, + 0.343 + ], + "angle": 0, + "content": "[39] Yunshuang Yuan and Monika Sester. Leveraging dynamic objects for relative localization correction in a connected autonomous vehicle network. ISPRS Annals of the Photogrammetry, Remote Sensing and Spatial Information Sciences, V-1-2022:101-109, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.345, + 0.905, + 0.4 + ], + "angle": 0, + "content": "[40] Yunshuang Yuan and Monika Sester. StreamLTS: Query-based temporal-spatial lidar fusion for cooperative object detection. In Computer Vision - ECCV 2024 Workshops, 2024. 2, 3, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.402, + 0.905, + 0.457 + ], + "angle": 0, + "content": "[41] Yunshuang Yuan, Hao Cheng, and Monika Sester. Keypoints-based deep feature fusion for cooperative vehicle detection of autonomous driving. IEEE Robotics and Automation Letters, 7(2):3054-3061, 2022. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.459, + 0.905, + 0.514 + ], + "angle": 0, + "content": "[42] Yunshuang Yuan, Hao Cheng, Michael Ying Yang, and Monika Sester. Generating evidential bev maps in continuous driving space. ISPRS Journal of Photogrammetry and Remote Sensing, 204:27-41, 2023. 1, 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.516, + 0.905, + 0.557 + ], + "angle": 0, + "content": "[43] Wu Zheng, Weiliang Tang, Sijin Chen, Li Jiang, and Chi-Wing Fu. CIA-SSD: Confident iou-aware single-stage object detector from point cloud. In AAAI, 2021. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.558, + 0.905, + 0.613 + ], + "angle": 0, + "content": "[44] Yin Zhou and Oncel Tuzel. Voxelnet: End-to-end learning for point cloud based 3d object detection. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4490-4499, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.615, + 0.905, + 0.683 + ], + "angle": 0, + "content": "[45] Walter Zimmer, Christian Creß, Huu Tung Nguyen, and Alois C Knoll. Tumtraf intersection dataset: All you need for urban 3d camera-lidar roadside perception [best student paper award]. In 2023 IEEE Intelligent Transportation Systems ITSC. IEEE, 2023. 8" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.683 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22305" + } + ] +] \ No newline at end of file diff --git a/2025/SparseAlign_ a Fully Sparse Framework for Cooperative Object Detection/e9f3f305-e62d-4aad-9255-49e76e43c9cf_origin.pdf b/2025/SparseAlign_ a Fully Sparse Framework for Cooperative Object Detection/e9f3f305-e62d-4aad-9255-49e76e43c9cf_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..18988b6a3b81a4be967a68428ef3d3963913b9d7 --- /dev/null +++ b/2025/SparseAlign_ a Fully Sparse Framework for Cooperative Object Detection/e9f3f305-e62d-4aad-9255-49e76e43c9cf_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:98d336cac3588623d979fcaccf3ebb8f23b7189aab7646ee43bee7199eec0ad2 +size 556367 diff --git a/2025/SparseAlign_ a Fully Sparse Framework for Cooperative Object Detection/full.md b/2025/SparseAlign_ a Fully Sparse Framework for Cooperative Object Detection/full.md new file mode 100644 index 0000000000000000000000000000000000000000..19f8fee2e6f6897de48fc3bb5f196a1e24150257 --- /dev/null +++ b/2025/SparseAlign_ a Fully Sparse Framework for Cooperative Object Detection/full.md @@ -0,0 +1,297 @@ +# SparseAlign: A Fully Sparse Framework for Cooperative Object Detection + +Yunshuang Yuan $^{1}$ Yan Xia $^{2,3\dagger}$ Daniel Cremers $^{2,3}$ Monika Sester $^{1}$ + +$^{1}$ Leibniz University Hannover $^{2}$ Technical University of Munich $^{3}$ Munich Center for Machine Learning (MCML) {yunshuang.yuan, monika.sester}@ikg.uni-hannover.de {yan.xia, cremers}@tum.de + +# Abstract + +Cooperative perception can increase the view field and decrease the occlusion of an ego vehicle, hence improving the perception performance and safety of autonomous driving. Despite the success of previous works on cooperative object detection, they mostly operate on dense Bird's Eye View (BEV) feature maps, which are computationally demanding and can hardly be extended to long-range detection problems. More efficient fully sparse frameworks are rarely explored. In this work, we design a fully sparse framework, SparseAlign, with three key features: an enhanced sparse 3D backbone, a query-based temporal context learning module, and a robust detection head specially tailored for sparse features. Extensive experimental results on both OPV2V and DAIR-V2X datasets show that our framework, despite its sparsity, outperforms the state of the art with less communication bandwidth requirements. In addition, experiments on the OPV2Vt and DAIR-V2Xt datasets for time-aligned cooperative object detection also show a significant performance gain compared to the baseline works. + +# 1. Introduction + +Cooperative Perception (CP) for autonomous driving has gained significant attention due to its potential to enhance road safety. Sharing Collective Perception Messages (CPMs) among Intelligent (traffic) Agents (IAs) in the vehicular network effectively reduces occlusions and expands the field of view, improving perception performance. In many recent works [1, 16, 24, 29, 35, 41, 42], CP has proven to be effective for different tasks, such as cooperative object detection [1, 24, 29, 41], Bird's Eye View (BEV) semantic segmentation [19, 42], and 3D occupancy semantic segmentation [22]. In this work, we focus on LiDAR-based cooperative vehicle detection. + +Training a cooperative 3D object detector is computationally demanding as it needs to process data from multiple vehicles, especially when the model is built to di + +![](images/3b92215a35f0c736e791874f42495a26171e0abf1fed4f4b07c69a6b308f2a61.jpg) + +![](images/1610062c385ae684802eeefeba04c5e130cda9c6944bc712aa6c35442716d363.jpg) +Figure 1. a,b) Communication strategies for cooperative object detection. c-e) Issues of sparse convolutional backbone networks. c) Center Feature Missing (CFM). d,e) Isolated Convolution Field (ICF) caused by ring disconnectivity and occlusion. + +gest sequential data from several subsequent frames to learn the temporal context. Following Fcooper [1], recent works [6, 11, 28-30, 33, 35] encode the 3D point cloud into dense BEV feature maps for further processing, sharing, and fusion. However, the computational cost of this pipeline increases quadratically with the perception range. Although cooperation among IAs can improve long-range detection performance, the IAs are normally not in the best positions to optimize the performance gain. The long-range detection performance of the local detector directly influences the overall performance. Besides, sharing the BEV feature maps among IAs requires extensive communication resources. Instead, the sparse operations leverage the sparsity of point clouds and have only linear computational complexity relative to the number of points. In addition, the output sparse features can be more efficient for sharing. It is therefore beneficial to perform cooperative object detection based on sparse operations only. + +However, building a fully sparse framework to compete with the dense ones is non-trivial. The first unavoidable issue is Center Feature Missing [8] (CFM, shown in Fig. 1 c)). Due to the range view of on-board LiDARs, there are + +usually no scan points in the vehicles' center area. However, compared to the edge points, the center points have a better ability to represent the whole object. Lacking center points for learning, the sparse frameworks perform worse than the dense ones that directly learn the features for center points. The second issue is the poor connectivity between points scanned by different laser beams. As shown in Fig. 1 d), this disconnectivity issue in distant areas may lead to isolated voxel blobs; the receptive field only enlarges to the blob scale as the convolutional layers go deeper. The isolated voxel blobs never exchange information with each other, limiting the backbone's capability of capturing global features. We call this issue Isolated Convolution Field (ICF). It also happens in the occlusion areas as shown in Fig. 1 e): although the occluded vehicle has very few scan points (red), it is expected to be detected based on the neighboring vehicles and the clues in the previous frames. However, ICF introduces difficulties for these isolated points to aggregate information from neighbors and the global environment. + +To overcome the above problems, we build a Sparse UNet backbone network SUNet with the Coordinate-Expandable sparse Convolution (CEC) implemented in the MinkowskiEngine [5] library. We apply CEC on $4\times$ and $8\times$ downsampled 3D voxels to increase the connectivity between ICFs. On the 2D BEV features, the CECs are used to expand the sparse coordinates, assuring the center features for all scanned vehicles. On top of SUNet, we build TempAlign Module (TAM) to learn the temporal context from the historical frames in a query-based [23, 40] manner, PoseAlign Module (PAM) to correct the relative pose errors between the cooperative and the ego agents, and SpatialAlign Module (SAM) to fuse the shared object queries based on the corrected poses. In addition, we propose the CompassRose encoding for the regression of BBox directions to further improve performance on both Cooperative Object Detection (COOD) and the Time-Aligned COOD [40] (TA-COOD) tasks. + +To summarize, the main contributions of this work are: + +- We investigate the issues of building a fully sparse network for LiDAR-based COOD and analyze how these issues impact the detection performance. +- We propose a novel fully sparse network, SparseAlign, via addressing the ICF and CFM problems of the sparse 3D backbone SUNet with CECs. Additionally, we introduce the SAM module, specifically tailored to ensure accurate fusion of sparse features. +- We assess the communication bandwidth and precision trade-offs of the different approaches, finding that our proposal reduces bandwidth consumption by up to $98\%$ , compared to previous methods using dense BEV feature maps for LiDAR-based cooperative object detection. +We conduct extensive experiments on OPV2V, DAIR-V2X, OPV2Vt, and DAIR-V2Xt benchmarks and show + +that the proposed SparseAlign has a significant improvement over state-of-the-art methods. + +# 2. Related Work + +# 2.1. Cooperative Object Detection (COOD) + +COOD frameworks are classified into sharing CPMs with raw data, intermediate features, and detections. Recent works [2, 24, 29, 38] have shown that intermediate feature sharing has the best performance due to its adjustable and learnable CPMs. We also use this strategy in our framework. In the perspective of communication, there are two strategies: handshaking and broadcasting as shown in Fig. 1 a) and b). With handshaking communication [11, 28-30, 33, 40, 42], the ego vehicle sends a request message that contains its pose to the cooperative IAs so that they can transform the local raw data to the ego coordinate system and then process their data to obtain the features for sharing. This simplifies the feature fusion process, as the features are directly learned in the goal coordinate system. In contrast, the broadcasting method [6, 17, 24, 37] directly processes the data in the local coordinate system and transmits it to other IAs. In the fusion stage, the ego vehicle must transform the cooperative features into the ego coordinate system and then conduct the fusion process. In comparison, broadcasting is more efficient as it does not require a prior communication agreement with the cooperative agents in highly dynamic driving scenarios where the communication partners change frequently. + +Achieving COOD presents several challenges, including communication latency and bandwidth limitations, inaccurate IA poses, and sensor asynchrony. To reduce the communication bandwidth consumption, [1, 24, 29, 33] compress the learned BEV feature maps to fewer channels via convolution layers, while [13] utilizes the learned confidence map and [42] leverages the uncertainty map to select the most important information for sharing. Instead, [41] employs the two-stage detector PVRCNN [20] and shares the 3D keypoint features from the second stage. To be more compact, [4, 9, 40] share the object query features [23] which are highly flexible to blend all relevant information, such as scanning time and IA pose, about an object into one unique feature vector. In this work, we also leverage the object query features to learn the temporal and spatial context of the object. Pose errors can dramatically deteriorate the detection results [24]. To this end, [39, 41] leverage the predicted Bounding Boxes (BBoxes) and select semantically classified keypoints to align the relative pose between the ego and the cooperative vehicles. Instead, [17] solely uses the BBoxes for pose alignment via the Pose Graph Optimization (PGO). However, these methods require the initial pose errors to be small to find the correct object matching in the first step and then iterate many times, or use the PGO + +![](images/44cc1539675ad7111a932d1986704015c5161b48351f5c39168527d394fa3ad6.jpg) + +![](images/95d04a62207224d1fbfbf49826dc9e9db10a85618467f71ec704117966537bd3.jpg) + +![](images/68687900a7229ab0a7b2604115550d9bc2d6419e05d0af4d413294535b8c18f3.jpg) +SUNet + +![](images/92b7e57fa75e7da32a2ccc6dd89d886434e1cd007168c160a9c99ed9968218aa.jpg) +TempAlign Module (TAM) + +![](images/613cabf80a43f8cd27c2bee3c4e834dcff5a64bb3bbfb4f08604407524ea99f1.jpg) +PoseAlign Module (PAM) +Figure 2. Overview of SparseAlign framework + +to refine the result. In this paper, we propose to learn pose-agnostic neighborhood graph features to accurately find the correct matching and then use PGO to refine it. Communication delay can cause that data arriving at the ego agent is outdated. To compensate for this delay, [36, 37] add the first-order derivatives of the BEV feature map to the original feature map to obtain the updated feature map that is aligned to the ego timestamp; [15] employs the LSTM [12] structure on the BEV feature maps to adjust the features to be aligned. Differently, [40] embeds the observation timestamps as well as the temporal context from historical frames into the object queries so that they are time- and motion-aware and capable of predicting the status at any given future timestamp, achieving the time-delay compensation. Different to [15, 36, 37] that compensate the feature maps, [40] leverages the point-wise timestamp of the point clouds to learn the accurate temporal context, hence is also able to cope with sensor asynchrony issues. Therefore, we adopt the query-based strategy for time-related compensation. + +# 2.2.3DBackbones + +To extract point cloud features, most previous works [16, 17, 24, 29, 33, 37, 41] for COOD either use dense 3D [25, 44], dense 2D [14, 32] or semi-sparse [21, 26, 31, 43] backbones that first encode the point cloud with 3D sparse convolution and then apply dense 2D convolutions for further processing. In contrast, [40, 42] employ the MinkUnet [5] to extract point cloud features. Despite its efficiency, the MinkUnet backbone struggles with long-range detection, occlusions and large objects, because of CFM and ICF problems. [8] proposed center voting for each scanned ob + +ject points to solve CFM issue and [3] reveals that further down-sampling the voxel features to lower resolutions can increase the receptive field and refine the large object detection. Instead of shifting the feature locations as [8] to accurately locate object queries or adding additional layers as [3], we build a backbone that solves both CFM and ICF issues by modifying MinkUnet with CEC layers. + +# 3. Method + +# 3.1. Task formalization + +In this work, we solve the LiDAR-based cooperative object detection task considering unavoidable challenges in reality: communication delay, sensor asynchrony, and localization errors. For this scenario, we assume there are $N$ IAs, including one ego agent $A_{1}$ and $N - 1$ cooperative agents. These agents are sharing CPMs with each other in a broadcasting manner, as shown in Fig. 1b). Specifically, in the $i$ -th frame, $A_{1}$ receives the CPMs $M = \{(m_j,t_j)|j\in \{1,\dots ,N - 1\} \}$ from the cooperative agents, where $m_j$ and $t_j$ are the CPM and the corresponding timestamp from the $j$ -th agent. We aim to optimize the object detection results based on the point clouds captured so far ( $t\leq t_i$ ) at $A_{1}$ and the received messages $M$ . We model and evaluate both the COOD and TA-COOD tasks from the perspective of an ego agent. The key difference between COOD and TA-COOD lies in the GT BBoxes. As illustrated in Fig. 2, TAM, COOD GT BBoxes (dashed lines) align with the sequentially captured LiDAR points, whereas TA-COD GT BBoxes (solid lines) are referenced to a global timestamp $t_g$ . This distinction requires the model to accurately predict + +![](images/682b68fa3080748eca50c1f79abb0047e99ec61a838df4a0ff883c2393331b16.jpg) +Figure 3. Receptive field (RF) after 4 sparse convolutions (convs). a. LiDAR points; b. Spare pixels (white) with a conv center point (red); c. RF coverage of 4 normal sparse convs (red); d. RF coverage of 4 CECs (red+light blue, darkest blue is background). + +![](images/9e202ffca0ac72622c4111bc1cee2959b3e65db2646a1f716e10f0a6327cdc71.jpg) + +![](images/bdbdc5d946e97c88ac9dafc4b3ca949f19f98df3ccb231c8999295f2aa952b5a.jpg) + +![](images/2c53502a3f07b26b836b892ff66e5e4cd6e3c960e129e07eb3436935a53573ef.jpg) + +BBox locations at $t_g$ . While COOD relies solely on point cloud geometry for location prediction, TA-COOD leverages temporal context to enhance accuracy at $t_g$ and also enables the model to handle location errors due to communication latency. + +# 3.2. Framework overview + +We build an efficient fully sparse framework SparseAlign, as shown in Fig. 2, to process sequential point clouds from multiple agents. All agents share the same network weights. The input at time $t_i$ is the point cloud of the ego agent $PC_{t_i}$ and of the cooperative agents $PC_{t_j}$ . Note that $t_j < t_i$ when communication latency exists. All point clouds are processed by the backbone SUNet to extract the point cloud features. The Region of Interest (RoI) module then selects the most interesting features as the object queries. For simplicity, we define object queries as $Q^{*} = \{(F_{i}^{*}, x_{i}^{*}, y_{i}^{*}) | i \in \{1, \dots, N^{*}\}$ , where * represents the query set name, $F$ is the features, and $x, y$ are the coordinates of the corresponding $i$ -th query. With selected queries, the TempAlign Module (TAM) can efficiently align the query features to the globally aligned timestamp $t$ so that errors introduced by communication delay and sensor asynchrony are compensated. The updated query features and the detections from the Local Query Detection head (LQDet) are then shared as CPMs. The PoseAlign Module (PAM) then uses the detected BBoxes to correct the relative poses between the cooperative and the ego agent and the SpatialAlign Module (SAM) fuses the received cooperative features into the ego coordinate system based on the corrected transformation $T_{c}^{e}$ from cooperative to ego coordinates. + +# 3.3. Backbone: Sparse UNet (SUNet) + +Processing data from multiple IAs for cooperative perception is computationally demanding. To improve efficiency, we use a fully sparse 3D UNet, similar to MinkUnet[5], to construct the SUNet (Fig. 2 bottom left). However, the fully sparse convolutions used in such networks suffer from the ICF issue. As shown in Fig. 3, the receptive field (RF) of (c) standard sparse convolutions only covers LiDAR points from a single vehicle. In contrast, the RF of (d) the CEC layers extends to the points of a neighboring vehicle, enabling learning within a larger spatial context. Therefore, + +we use 3D CECs in the sparse convolution (SConv) block 3 and 4. Correspondingly, we contract the coordinates in the transposed sparse convolution (TConv) block 3 and 2. In this way, the output of the TConv blocks $p4$ and $p2$ can match the coordinates in stride $x2$ and $x4$ for the concatenation. In addition, it learns better global features without voxel number soaring in the subsequent layers and maintains efficiency. Against the CFM issue, we use 2D CECs on the 2D BEV sparse features (Fig. 2 SUNet:Coor. Expand.) to expand the coordinates to ensure their coverage over the object centers. + +# 3.4. Alignment modules + +TempAlign Module (TAM). The TAM aims to learn the temporal context relative to the previous frames to achieve two goals: improve object detection performance by looking for clues in the history and compensate for the object displacement error introduced by asynchronous scanning time (Fig. 2, TAM). We achieve the second goal by predicting the BBoxes at the globally aligned timestamp so that the BBoxes shared for PAM are synchronized. Then PAM only needs to deal with the object displacement caused by localization errors. Technically, the TAM follows [40], for efficiency, by removing the global cross-attention module which learns the global BEV features. We only use a memory queue and a hybrid attention module. The memory queue stores the top $k = 256$ object queries from previous $L = 4$ data frames so that the attention module can allow the interaction between object queries in the current frame and the historical frames. In this way, the object queries learn the temporal context from the sequential data and are able to make predictions for the future state. + +PoseAlign Module (PAM). Given the detected BBoxes $\mathbf{B}_i$ and $\mathbf{B}_j$ , respectively, from the perspective of the agent $A_i$ and $A_j$ in the same scene, humans can easily find the correspondences between $\mathbf{B}_i$ and $\mathbf{B}_j$ based on the geometric relationships. Practically, we model the features of these geometric relationships by embedding the neighborhood geometry of each BBox into deep features and then match these features of $\mathbf{B}_i$ and $\mathbf{B}_j$ with the Hungarian algorithm. For each BBox $b$ in a BBox set $\mathbf{B}$ , we select the nearest $k = 8$ neighbors to embed its neighborhood features. As shown in Fig. 2, PAM: 1.Graph Feature, the orientation angles of the BBox are $\alpha$ and $\beta$ for $b$ and its neighbor, respectively. We embed the relative orientation with $\nu_{a} = [\sin (\beta -\alpha);\cos (\beta -\alpha)]^{1}$ . Similarly, the relative edge direction is embedded as $\epsilon_{a} = [\sin (\theta -\alpha);\cos (\theta -\alpha)]$ . In addition, the dimension $\nu_{dim} = [l,w,h]$ of the neighbor BBox, and the Euclidean distance between $b$ and its neighbors $\epsilon_d = d$ , are also embedded into the feature. Note that $\epsilon_d,\epsilon_a,\nu_a$ are all relative features between two BBoxes that are pose-agnostic, leading to robust pose alignment without + +dependency on the magnitude of initial pose errors. These features are then concatenated and embedded into high-dimensional features by a linear layer with ReLU activation $(\mathbb{R}^5\to \mathbb{R}^{d_{emb}})$ : + +$$ +f _ {n b r} = \operatorname {l i n e a r} \left(\left[ \epsilon_ {d}; \epsilon_ {a}; \nu_ {a}; \nu_ {\text {d i m}} \right]\right) \tag {1} +$$ + +Assembling the features relative to all neighbors, we summarize the neighborhood feature $F_{nbr} \in \mathbb{R}^{k \times d_{emb}}$ by a multi-head self-attention module $attn$ followed by a linear layer operated on the mean and max of the $attn$ outcome, resulting in the feature $F_{nbr} \in \mathbb{R}^{d_{emb}}$ : + +$$ +F _ {n b r} = \operatorname {a t t n} \left(F _ {n b r}\right) \tag {2} +$$ + +$$ +F _ {n b r} = \operatorname {l i n e a r} \left(\left[ \operatorname {m e a n} \left(F _ {n b r}\right); \operatorname {m a x} \left(F _ {n b r}\right) \right]\right) \tag {3} +$$ + +By calculating the Euclidean distance between the learned neighborhood feature $F_{nbr}$ of the BBox in $\mathbf{B}_i$ and $\mathbf{B}_j$ , we construct the cost matrix for the linear sum assignment to find the best Match (Fig. 2, PAM: 2-match) between these two BBox sets. However, an object in $\mathbf{B}_i$ does not always have a projection in $\mathbf{B}_j$ and vice versa. Thus, we reject the wrong matches that have large distances. Based on the PAM: 2-match result, the relative pose transformation between $A_i$ and $A_j$ can be recovered, completing the PAM: 3Align step. Additionally, we use pose graph optimization to refine the alignment by leveraging loop closure among multiple cooperative agents posing following [17]. + +SpatialAlign Module (SAM). To align the cooperative query features $Q^c$ with the ego query features $Q^e$ , there are two issues to be resolved. First, the features $F^c$ are learned in the cooperative coordinate system and must be adapted to the ego coordinate system. We achieve this using a multilayer perceptron (mlp) as described in Eq. (4) where $F^R \in \mathbb{R}^{N \times 9}$ is the flattened and repeated rotation matrix $R \in \mathbb{R}^{3 \times 3}$ . + +$$ +F ^ {c} = m l p \left(\left[ F ^ {c}; F ^ {R} \right]\right) \tag {4} +$$ + +Second, after rotation with $R$ , the coordinates of $Q^c$ (Fig. 2, SAM, yellow points) do not perfectly align to the discretized grid in the ego coordinate system. Thus, we merge $Q^c$ into the nearest points of this grid. The locations of the fused output object query points $Q^{ts}$ are shown with gray points. For each point in $Q^{ts}$ , we summarize the features of its $k = 8$ nearest neighbors (Fig. 2, SAM: black arrows) in query set $Q^{\cup} = Q^{c}\bigcup Q^{e}$ to obtain the fused features $F^{ts}$ . Mathematically, it reads as + +$$ +F _ {i j} = m l p \left(\left[ F _ {j} ^ {\cup}; \text {l i n e a r} \left(\left[ x _ {i} ^ {t s} - x _ {j} ^ {\cup}, y _ {i} ^ {t s} - y _ {j} ^ {\cup} \right]\right) \right]\right) \tag {5} +$$ + +$$ +F _ {i} ^ {t s} = \left\{F _ {i j} | j \in \{1, \dots , k \} \right\} \tag {6} +$$ + +$$ +F _ {i} ^ {t s} = \max _ {j} \left(F _ {i} ^ {t s}\right) + \text {m e a n} _ {j} \left(F _ {i} ^ {t s}\right) \tag {7} +$$ + +where $i$ is the query index of $Q^{ts}$ and $j$ is the neighbor index of the $i$ -th query. + +![](images/529f1456c1122a5dfe97c8baaff0642621ac6e2ca352e44cfa80719f4f141a2f.jpg) +Figure 4. CompassRose encoding + +# 3.5. Detection Heads + +To train SparseAlign, we follow [40] to attach a center-based object detection head to each of the RoI, TAM and SAM to learn query features. These heads are notated as RoIDet, LQDet and GQDet, respectively. They share the same target-encoding method. We use Focal loss to classify fore- and background object queries; all query points inside the ground-truth BBoxes are positive, and those outside are negative. Smooth $L1$ loss is used for regressing $[dx,dy,dz,l,w,h,\mathcal{B}_{dir},\mathcal{B}_{scr}]$ , where $l,w,h$ are the dimensions of BBoxes and $dx,dy,dz$ are coordinate offsets between the query points and the center of ground-truth BBoxes. $\mathcal{B}_{dir},\mathcal{B}_{scr}$ are the novel CompassRose encoding (Fig. 4) of the ground truth orientation $r_g$ with respect to four anchor angles $\mathbf{r}_a = [0,0.5\pi ,\pi ,1.5\pi ]$ . It reads as + +$$ +\begin{array}{l} \mathcal {B} _ {\mathrm {d i r}} = \left[ \cos r _ {g} - \cos \mathbf {r} _ {a}, \sin r _ {g} - \sin \mathbf {r} _ {a} \right] (8) \\ \mathcal {B} _ {\mathrm {s c r}} = \left[ s _ {0}, s _ {0. 5 \pi}, s _ {\pi}, s _ {1. 5 \pi} \right] (9) \\ = \mathbf {1} - \arccos \left(\cos \mathbf {r} _ {a} \cdot \cos r _ {g} + \sin \mathbf {r} _ {a} \cdot \sin r _ {g}\right) / \pi \\ \end{array} +$$ + +Take the encoded sine wave (Fig. 4 left) as an example, the four anchor angles (yellow points) ensure that at least one anchor can reach the target angle monotonically, simplifying the regression process. + +# 3.6. Data augmentation + +In addition to geometric augmentation, which includes randomly flipping the input point cloud along the x- or y-axis and rotating it around the z-axis, we apply Free Space Augmentation (FSA) [42]. FSA enhances the point cloud by adding auxiliary points along LiDAR scanning rays, representing empty spaces where no objects are present. This augmentation helps mitigate the ICF problem in distant regions, where observation points are extremely sparse. For more details, refer to supplementary B. + +# 4. Experiment + +Datasets. We evaluate SparseAlign on OPV2V[29] and DAIR-V2X[35] datasets for COOD and their variant datasets OPV2Vt[40] and DAIR-V2Xt[40] for TA-COOD. OPV2V is a synthetic dataset created in different driving scenarios of CARLA [7] simulator. Following [29], the detection range is set to $x \in [-140, 140]m$ , $y \in [-40, 40]m$ . + +
MethodBW(Mb) ↓AP0.5 ↑AP0.7↑
Fcooper[1]72.080.8870.790
V2VNet[24]72.080.9170.822
FPVRCNN[41]0.240.8730.820
OPV2V[29]126.80.9050.815
CoAlign[17]72.080.9020.833
CoBEVT[19]72.080.9130.861
V2VAM[16]72.080.9160.850
V2VFformer++[34]L+C72.080.9350.895
SparseAlign (ours)< 1.30.9300.892
+ +Table 1. COOD Average Precision on OPV2V dataset. ${}^{2}$ + +
MethodOPV2VtDAIR-V2Xt
AP0.5↑AP0.7↑AP0.5↑AP0.7↑
Fcooper[1]+SA0.7630.5530.5970.282
FPVRCNN[43]+SA0.6400.4740.5980.307
OPV2V[29]+SA0.8810.7870.7020.366
StreamLTS[40]0.8530.7210.6400.404
StreamLTS[40]+SA0.8500.7900.7570.497
SparseAlign (full)0.8930.8180.7960.548
+ +DAIR-V2X is a real dataset containing one vehicle- and one infrastructure-side sensor suit. The detection range is set to $x \in [-100, 100]m, y \in [-40, 40]m$ . OPV2Vt is generated by interpolating OPV2V data frames into sub-frames and replaying them to mimic the sensor asynchrony and the rolling shutter problem of LiDARs. The GT BBoxes of this dataset are aligned to the global timestamp, the LiDAR scan ending time of each frame. DAIR-V2Xt follows the same idea. The detection range of OPV2Vt and DAIR-V2Xt is set the same as OPV2V and DAIR-V2X. + +Implementation details. All models are trained on a single GTX-4090 and an Intel i7-8700 CPU for 50 epochs (about 35 hours) with the Adam optimizer and batch size two. The learning rate is set to $2e - 4$ with a warm-up stage of 4000 iterations. The random rotation angle is set to the range $[-90^{\circ}, 90^{\circ}]$ . For models incorporating the TAM, we load four sequential frames of data during training and calculate losses solely for the last frame. To reduce GPU memory consumption, we enable gradient calculations only for data from a single ego agent and one cooperative agent. For details on efficient gradient scheduling, refer to supplementary $D$ . To ensure that the model generalizes well to the significant time latency introduced by communication, we load data from cooperative vehicles with a random latency ranging from 0 to $200~\mathrm{ms}$ relative to the ego vehicles. + +Table 3. TA-COOD Average Precision on OPV2Vt and DAIR-V2Xt + +
MethodBW(Mb) ↓AP0.5 ↑AP0.7↑
Fcooper[1]48.80.7340.559
V2VNet[24]48.80.6640.402
FPVRCNN[41]0.240.6650.505
V2XViT[28]48.80.7040.531
DiscoNet[19]48.80.7360.583
OPV2V[29]97.60.7330.553
CoAlign[17]48.80.7460.604
DI-V2X[27]48.80.7880.662
SparseAlign (ours)< 1.30.8450.685
+ +Table 2. COOD Average Precision on DAIR-V2X dataset. + +# 5. Results + +# 5.1. Comparison to the state-of-the-art + +Tab. 1 presents the COOD results on the OPV2V dataset. All methods use only LiDAR data, except for V2VFormer++ (denoted by $\mathrm{L + C}$ ), which incorporates both LiDAR and camera data. Compared to the state-of-the-art, our SparseAlign achieves the best performance in LiDAR-based COOD, approaching the results of V2VFormer++, which benefits from LiDAR-Camera fusion. Tab. 2 shows the results on the DAIR-V2X dataset. Here, SparseAlign outperforms other methods by a large margin, achieving a $2.3\%$ improvement in AP at the IoU threshold of 0.7 and a $5.7\%$ improvement at the IoU threshold of 0.5. This is mainly due to the enhanced temporal and global reasoning ability of our model for detecting distant and large objects. Compared to other methods that rely on dense BEV feature maps, SparseAlign consumes the least communication bandwidth (BW) of less than 1.3MB (without compression) thanks to the efficient query-based operations. + +For the results of the more challenging TA-COOD task shown in Tab. 3, our SparseAlign outperforms StreamLTS by a great margin. By applying our three alignment modules (StreamLTS + SA), the performance of StreamLTS is also improved. In addition, we compare our framework to Fcooper, FPVRCNN and OPV2V method by attaching the three alignment modules to them to achieve TA-COOD. All show worse performance than the full SparseAlign because of worse 3D backbone or the RoI selection. + +The qualitative results and the additional comparison to methods, such as TransIFF[4], on the 3D AP metric are given in the supplementary A and E, respectively. + +# 5.2. Ablation study + +We conduct an ablation study on the 3D backbone, TAM, and SAM (PAM is discussed separately in Sec. 5.4). Among these components, the spatial fusion module plays a crucial role in multi-agent feature fusion. Without SAM, the module performs the worst, as evidenced by the first row of Tab. 4. When using QUEST [9] spatial fusion, which + +
CEC 2DCEC 3DTAMSAMOPV2VDAIR-V2X
AP0.5AP0.7AP0.5AP0.7
✓✓✓0.8470.7560.6710.528
✓✓✓QST0.8880.8690.7490.622
✓0.9240.8850.7730.638
✓✓0.9400.9010.7960.662
✓✓✓0.9500.9140.8130.682
✓✓✓✓0.9510.9290.8450.685
+ +Table 4. Ablation study on 2D and 3D CEC layers, TAM and SAM. QST: use QUEST [9] spatial fusion instead of SAM. + +
Dir.Enc.OPV2VDAIR-V2X
AP0.5AP0.7AP0.5AP0.7
gt-angle0.8710.7680.6200.427
second0.9080.8740.7430.590
sin_cos0.9150.8810.7480.603
compass0.9240.8850.7730.638
+ +Table 5. Comparison on BBox angle encoding methods: offset to ground-truth angle (gt-angle), second-style (second), sine-cosine offsets (sin_cos) and compass rose (compass). + +
DatasetOPV2VtDAIR-V2Xt
Latency0ms100ms200ms0ms100ms200ms
AP threshold0.50.70.50.70.50.70.50.70.50.70.50.7
StreamLTS [40]0.8530.7210.8160.6800.7870.6470.6420.4040.6130.3790.5900.364
SA no TAM0.8900.8020.8560.4710.6310.2090.7200.4570.6720.3930.6350.350
SA no PT0.8350.6150.8310.6250.8150.5810.7600.4690.7370.4590.7220.449
SA no TL0.8850.8430.8620.6190.7230.2790.7930.5340.7490.4740.6980.423
SA no FSA0.8810.8160.8750.7950.8570.7620.7690.5080.7550.5020.7400.496
SparseAlign (SA)0.8930.8180.8850.7950.8670.7650.7960.5480.7860.5430.7720.532
+ +Table 6. AP of TA-COOD with communication latency. PT: Point-wise Timestamps, TL: Train Latency, FSA: Free Space Augmentation. + +aligns only rotation (Eq. (4)) without correcting spatial displacement (Eqs. (5) to (7)), the AP scores remain significantly lower than those obtained with SAM. Incorporating SAM improves performance, and we further enhance it by gradually replacing standard sparse convolutional layers with CEC layers. Finally, TAM further boosts AP scores by leveraging temporal information from historical frames. + +In Tab. 5, we compare CompassRose encoding with three methods: directly regress the ground-truth angle (gt-angle), second-style [31] encoding (second), and the sine-cosine encoding (sin_cos). For simplicity, we use the configurations in the third row of Tab. 4 for the comparison. The results show that CompassRose performs the best. + +# 5.3. Communication latency and sensor asynchrony + +Tab. 6 shows the results of TA-COOD on the OPV2Vt and the DAIR-V2Xt dataset. Our proposed SparseAlign demonstrates a significant improvement over the baseline framework, StreamLTS [40]. Specifically, the AP at an IoU threshold of 0.7 increased $9.7\%$ on the OPV2Vt dataset and $14.4\%$ on the DAIR-V2Xt dataset. As the communication latency increases to 200ms, the AP at IoU 0.5 of our SparseAlign only dropped $2.6\%$ while the baseline work has lost $6.6\%$ of its accuracy. This reveals that our framework is more robust against communication latency. To investigate modules or configurations that are making an effect on this robustness, we conducted an ablation study by removing the TAM, Point-wise Timestamps (TP), Training Latency (TL, communication latency during training) and + +Free Space Augmentation (FSA), respectively. + +Without the TAM, the performance noticeably declines as latency increases, despite achieving comparable results to the full SpareAlign model at $0\mathrm{ms}$ latency. A similar effect is observed when the full SA model is trained without exposure to data with communication latency (SA without TL). Remarkably, without TL, the model outperforms the full SpareAlign model in terms of AP at an IoU threshold of 0.7 on the OPV2Vt dataset (0.843 against 0.818). This is because the model focuses more on predicting BBoxes within the same frame, where only minimal latency is introduced by asynchronous sensors. + +Without PT, the model performs much worse than the full SpareAlign, showing the importance of fine-grained point-wise timestamps in the point clouds for the model to compensate for the errors introduced by sensor asynchrony and capture the accurate temporal context for prediction. Finally, the ablation study on the FSA module also shows a positive influence on the performance. For example, AP at the IoU of 0.7 increased by $4\%$ on DAIR-V2Xt by utilizing the FSA (See supplementary D for visual comparison). + +# 5.4. Localization errors + +Fig. 5 shows the performance of SpareAlign against different localization errors. We add random errors to the poses of both ego $(T_e)$ and cooperative $(T_c)$ vehicles. The translation errors $x_{\epsilon}, y_{\epsilon}$ along the $x$ - and $y$ -axis and the rotation error $r_{\epsilon}$ around the $z$ -axis are all assumed to be normally distributed with $\mathcal{N}(0,1) \cdot \epsilon$ , where $\epsilon \in [0,1]$ is the error scaling factor. + +![](images/de9edb1b1ab958c861bf4550b46e38a36f72222773356e1ae489e155e178d8c9.jpg) + +![](images/d963ed326dce265d53ea2cafabe9165d8f70ada8d9bc5100897043381cf87a89.jpg) +Figure 5. AP at IoU threshold of 0.7 with translation errors ranging from $0\mathrm{m}$ to $1\mathrm{m}$ along x- and y-axis, and rotation errors from $0^{\circ}$ to $1.0^{\circ}$ (horizontal axis) for the different datasets. + +We report the results by gradually increasing $\epsilon$ with a step size of 0.2. Note that the translation and rotation errors exist in the poses of both the ego and the cooperative vehicle; the resulting relative translation error of the transformation $T_{c}^{e} = inverse(T_{e})\cdot T_{c}$ from cooperative to ego coordinate system will be amplified as the rotation error increases. Our PAM is specially designed to mitigate the influence of large relative transformation errors. + +Our proposed PAM and the baseline method CoAlign [17] both significantly reduce the impact of pose errors on AP performance compared to the configuration lacking any correction module. In addition, our method demonstrates greater robustness against large errors. For instance, on the OPV2V dataset, the AP for CoAlign decreased by approximately $27\%$ at the error of $1.0m, 1.0m, 1.0^{\circ}$ , whereas PAM experienced only a $12\%$ drop. This validates the efficacy of our design, which relies on the pose-agnostic relative neighborhood geometries to match the detected BBoxes of the ego and the cooperative IAs. Note that the DAIR-V2X dataset exhibits smaller changes in AP as localization errors increase, because the pose parameters are not well calibrated. In contrast, [40] refined the poses of DAIR-V2X for generating the DAIR-V2Xt dataset, making the better-calibrated DAIR-V2Xt more sensitive to newly introduced errors. + +# 5.5. CPM sizes + +Our framework significantly reduces communication bandwidth without relying on computationally intensive dense feature maps, unlike state-of-the-art methods. This effi + +![](images/c77fbab4063aa88260bffeeeae859f74e50be0d25aeac07ea4597d41fcfb6a98.jpg) +Figure 6. CPM sizes with different information selection scores. Object query features with detection scores larger than the threshold (x-axis) are shared as CPMs. + +![](images/2df0624091bbbf8392d6db71bcd8377fe6a2e7bf9823f149723911361688707e.jpg) + +ciency is achieved by selecting only the top $K = 1024$ object queries for further processing. In addition to this intrinsic reduction, the size of the CPMs can be further minimized through score-based selection. Specifically, we employ the query-based detection scores from the LQDet head to identify and share only the most critical queries. As Fig. 6 shows, using a CPM score threshold below 0.5 does not lead to a noticeable degradation in performance across all datasets. Remarkably, at a threshold score of 0.5, the average size of the CPMs is reduced to less than $400\mathrm{KB}$ on all datasets. The final CPM sizes after selection are influenced by the driving scenes. Specifically, scenes with more vehicles tend to have larger CPM sizes. Consequently, the OPV2V and OPV2Vt datasets, which contain more vehicles in their scenes, result in larger CPM sizes compared to the DAIR-V2X and DAIR-V2Xt datasets. + +# 6. Conclusion + +In this paper, we proposed SparseAlign, a fully sparse framework for cooperative object detection (COOD). With our novel 3D backbone network SUNet, the Compass-Rose encoding of detection head, the temporal and spatial fusion module TAM and SAM, SparseAlign significantly outperforms state-of-the-art methods on the OPV2V and DAIR-V2X datasets for COOD, as well as on their variants OPV2Vt and DAIR-V2Xt for time-aligned COOD (TA-COOD), achieving this with very low communication bandwidth. Additionally, our proposed pose alignment module PAM demonstrates improved robustness to large localization errors compared to the baseline method CoAlign. Nevertheless, when the pose errors are large, a drop of about $10\%$ in APs persists even with pose correction. In future work, we therefore plan to explore more robust methods, e.g. leveraging sub-graph detection, to further enhance the pose alignment accuracy. Additionally, more application scenarios, such as roadside cooperative detection benchmarks HoloVIC[18], RCooper[10] and TUMTraf[45], could also be explored. + +# References + +[1] Qi Chen. F-cooper: feature based cooperative perception for autonomous vehicle edge computing system using 3d point clouds. Proceedings of the 4th ACM/IEEE Symposium on Edge Computing, 2019. 1, 2, 6 +[2] Qi Chen, Sihai Tang, Qing Yang, and Song Fu. Cooper: Cooperative perception for connected autonomous vehicles based on 3d point clouds. 2019 IEEE 39th International Conference on Distributed Computing Systems (ICDCS), pages 514-524, 2019. 2 +[3] Yukang Chen, Jianhui Liu, Xiangyu Zhang, Xiaojuan Qi, and Jiaya Jia. Voxelnext: Fully sparse voxelnet for 3d object detection and tracking. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 21674-21683, 2023. 3 +[4] Ziming Chen, Yifeng Shi, and Jinrang Jia. TransIFF: An instance-level feature fusion framework for vehicleinfrastructure cooperative 3d detection with transformers. In 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 18159-18168, 2023. 2, 6 +[5] Christopher Bongsoo Choy, JunYoung Gwak, and Silvio Savarese. 4d spatio-temporal convnets: Minkowski convolutional neural networks. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3070-3079, 2019. 2, 3, 4 +[6] Jiaxun Cui, Hang Qiu, Dian Chen, Peter Stone, and Yuke Zhu. Coopernaut: End-to-end driving with cooperative perception for networked vehicles. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), page 17252-17262, 2022. 1, 2 +[7] Alexey Dosovitskiy, German Ros, Felipe Codevilla, Antonio Lopez, and Vladlen Koltun. CARLA: An open urban driving simulator. In Proceedings of the 1st Annual Conference on Robot Learning, pages 1-16, 2017. 5 +[8] Lue Fan, Feng Wang, Naiyan Wang, and Zhaoxiang Zhang. Fully Sparse 3D Object Detection. In NeurIPS, 2022. 1, 3 +[9] Siqi Fan, Haibao Yu, Wenxian Yang, Jirui Yuan, and Zaiqing Nie. Quest: Query stream for practical cooperative perception. In IEEE International Conference on Robotics and Automation (ICRA), 2024. 2, 6, 7 +[10] Ruiyang Hao, Siqi Fan, Yingru Dai, Zhenlin Zhang, Chenxi Li, Yuntian Wang, Haibao Yu, Wenxian Yang, Yuan Jirui, and Zaiqing Nie. Rcooper: A real-world large-scale dataset for roadside cooperative perception. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 22347-22357, 2024. 8 +[11] Caizhen He, Hai Wang, Long Chen, Tong Luo, and Yingfeng Cai. V2x-ahd:vehicle-to-everything cooperation perception via asymmetric heterogenous distillation network. arXiv, arXiv:2310.06603, 2023. arXiv:2310.06603 [cs]. 1, 2 +[12] Sepp Hochreiter and Jürgen Schmidhuber. Long short-term memory. Neural Computation, 9(8):1735-1780, 1997. 3 +[13] Yue Hu, Shaoheng Fang, Zixing Lei, Yiqi Zhong, and Siheng Chen. Where2comm: Communication-efficient collaborative perception via spatial confidence maps. In Thirty-sixth Conference on Neural Information Processing Systems (Neurips), 2022. 2 + +[14] Alex H. Lang, Sourabh Vora, Holger Caesar, Lubing Zhou, Jiong Yang, and Oscar Beijbom. Pointpillars: Fast encoders for object detection from point clouds. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12689-12697, 2018. 3 +[15] Zixing Lei, Shunli Ren, Yue Hu, Wenjun Zhang, and Siheng Chen. Latency-aware collaborative perception. In Computer Vision – ECCV 2022, page 316–332, 2022. 3 +[16] Jinlong Li, Runsheng Xu, Xinyu Liu, Jin Ma, Zicheng Chi, Jiaqi Ma, and Hongkai Yu. Learning for vehicle-to-vehicle cooperative perception under lossy communication. IEEE Transactions on Intelligent Vehicles, 8(4):2650-2660, 2023. 1, 3, 6 +[17] Yifan Lu, Quanhao Li, Baoan Liu, Mehrdad Dianati, Chen Feng, Siheng Chen, and Yanfeng Wang. Robust collaborative 3d object detection in presence of pose errors. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 4812-4818. IEEE, 2023. 2, 3, 5, 6, 8 +[18] Cong Ma, Lei Qiao, Chengkai Zhu, Kai Liu, Zelong Kong, Qing Li, Xueqi Zhou, Yuheng Kan, and Wei Wu. Holovic:large-scale dataset and benchmark for multi-sensor holographic intersection and vehicle-infrastructure cooperative. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 22129-22138, 2024. 8 +[19] Xu Runsheng, Zhengzhong Tu Tu, Hao Xiang, Wei Shao, Bolei Zhou, and Jiaqi Ma. Cobevt: Cooperative bird's eye view semantic segmentation with sparse transformers. In Conference on Robot Learning (CoRL), 2022. 1, 6 +[20] Shaoshuai Shi, Chaoxu Guo, Li Jiang, Zhe Wang, Jianping Shi, Xiaogang Wang, and Hongsheng Li. Pv-rcnn: Point-voxel feature set abstraction for 3d object detection. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10526–10535, 2019. 2 +[21] Shaoshuai Shi, Chaoxu Guo, Li Jiang, Zhe Wang, Jianping Shi, Xiaogang Wang, and Hongsheng Li. PV-RCNN: Point-voxel feature set abstraction for 3d object detection. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10526-10535, 2019. 3 +[22] Rui Song, Chenwei Liang, Hu Cao, Zhiran Yan, Walter Zimmer, Markus Gross, Andreas Festag, and Alois Knoll. Collaborative semantic occupancy prediction with hybrid feature fusion in connected automated vehicles. In 2024 IEEE/CVF International Conference on Computer Vision and Pattern Recognition (CVPR). IEEE/CVF, 2024. 1 +[23] Shihao Wang, Ying-Hao Liu, Tiancai Wang, Ying Li, and Xiangyu Zhang. Exploring object-centric temporal modeling for efficient multi-view 3d object detection. 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 3598-3608, 2023. 2 +[24] Tsun-Hsuan Wang, Sivaban Manivasagam, Ming Liang, Bin Yang, Wenyuan Zeng, and Raquel Urtasun. V2VNet: Vehicle-to-Vehicle Communication for Joint Perception and Prediction, page 605-621. Springer International Publishing, Cham, 2020. 1, 2, 3, 6 +[25] Yan Xia, Yusheng Xu, Shuang Li, Rui Wang, Juan Du, Daniel Cremers, and Uwe Stilla. Soe-net: A self-attention and orientation encoding network for point cloud based place + +recognition. In Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition, pages 11348-11357, 2021. 3 +[26] Yan Xia, Mariia Gladkova, Rui Wang, Qianyun Li, Uwe Stilla, João F Henriques, and Daniel Cremers. Casspr: Cross attention single scan place recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 8461-8472, 2023. 3 +[27] Li Xiang, Junbo Yin, Wei Li, Cheng-Zhong Xu, Ruigang Yang, and Jianbing Shen. DI-V2X: Learning domain-invariant representation for vehicle-infrastructure collaborative 3d object detection. AAAI, 2024. 6 +[28] Runsheng Xu, Hao Xiang, Zhengzhong Tu, Xin Xia, Ming-Hsuan Yang, and Jiaqi Ma. V2x-vit: Vehicle-to-everything cooperative perception with vision transformer. In Computer Vision - ECCV 2022, page 107-124, Cham, 2022. Springer Nature Switzerland. 1, 2, 6 +[29] Runsheng Xu, Hao Xiang, Xin Xia, Xu Han, Jinlong Li, and Jiaqi Ma. Opv2v: An open benchmark dataset and fusion pipeline for perception with vehicle-to-vehicle communication. 2022 International Conference on Robotics and Automation (ICRA), page 2583–2589, 2022. 1, 2, 3, 5, 6 +[30] Runsheng Xu, Xin Xia, Jinlong Li, Hanzhao Li, Shuo Zhang, Zhengzhong Tu, Zonglin Meng, Hao Xiang, Xiaoyu Dong, Rui Song, Hongkai Yu, Bolei Zhou, and Jiaqi Ma. V2v4real: A real-world large-scale dataset for vehicle-to-vehicle cooperative perception. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), page 13712–13722, Vancouver, BC, Canada, 2023. IEEE. 1, 2 +[31] Yan Yan, Yuxing Mao, and Bo Li. Second: Sparsely embedded convolutional detection. Sensors (Basel, Switzerland), 18, 2018. 3, 7 +[32] Binh Yang, Wenjie Luo, and Raquel Urtasun. Pixor: Realtime 3d object detection from point clouds. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7652-7660, 2018. 3 +[33] Hongbo Yin, Daxin Tian, Chunmian Lin, Xuting Duan, Jianshan Zhou, Dezong Zhao, and Dongpu Cao. V2VFormer ++: Multi-modal vehicle-to-vehicle cooperative perception via global-local transformer. IEEE Transactions on Intelligent Transportation Systems, page 1-14, 2023. 1, 2, 3 +[34] Hongbo Yin, Daxin Tian, Chunmian Lin, Xuting Duan, Jianshan Zhou, Dezong Zhao, and Dongpu Cao. V2vformer++: Multi-modal vehicle-to-vehicle cooperative perception via global-local transformer. IEEE Transactions on Intelligent Transportation Systems, 25:2153-2166, 2024. 6 +[35] Haibao Yu, Yizhen Luo, Mao Shu, Yiyi Huo, Zebang Yang, Yifeng Shi, Zhenglong Guo, Hanyu Li, Xing Hu, Jirui Yuan, and Zaiqing Nie. Dair-v2x: A large-scale dataset for vehicle-infrastructure cooperative 3d object detection. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), page 21329-21338, New Orleans, LA, USA, 2022. IEEE. 1, 5 +[36] Haibao Yu, Yingjuan Tang, Enze Xie, Jilei Mao, Ping Luo, and Zaiqing Nie. Flow-based feature fusion for vehicleinfrastructure cooperative 3d object detection. In Advances in Neural Information Processing Systems, 2023. 3 + +[37] Haibao Yu, Wenxian Yang, Hongzhi Ruan, Zhenwei Yang, Yingjuan Tang, Xu Gao, Xin Hao, Yifeng Shi, Yifeng Pan, Ning Sun, Juan Song, Jirui Yuan, Ping Luo, and Zaiqing Nie. V2x-seq: A large-scale sequential dataset for vehicleinfrastructure cooperative perception and forecasting. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), page 5486-5495, Vancouver, BC, Canada, 2023. IEEE. 2, 3 +[38] Yunshuang Yuan and Monika Sester. COMAP: A synthetic dataset for collective multi-agent perception of autonomous driving. The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, XLIII-B2-2021:255-263, 2021. 2 +[39] Yunshuang Yuan and Monika Sester. Leveraging dynamic objects for relative localization correction in a connected autonomous vehicle network. ISPRS Annals of the Photogrammetry, Remote Sensing and Spatial Information Sciences, V-1-2022:101-109, 2022. 2 +[40] Yunshuang Yuan and Monika Sester. StreamLTS: Query-based temporal-spatial lidar fusion for cooperative object detection. In Computer Vision - ECCV 2024 Workshops, 2024. 2, 3, 4, 5, 6, 7, 8 +[41] Yunshuang Yuan, Hao Cheng, and Monika Sester. Keypoints-based deep feature fusion for cooperative vehicle detection of autonomous driving. IEEE Robotics and Automation Letters, 7(2):3054-3061, 2022. 1, 2, 3, 6 +[42] Yunshuang Yuan, Hao Cheng, Michael Ying Yang, and Monika Sester. Generating evidential bev maps in continuous driving space. ISPRS Journal of Photogrammetry and Remote Sensing, 204:27-41, 2023. 1, 2, 3, 5 +[43] Wu Zheng, Weiliang Tang, Sijin Chen, Li Jiang, and Chi-Wing Fu. CIA-SSD: Confident iou-aware single-stage object detector from point cloud. In AAAI, 2021. 3, 6 +[44] Yin Zhou and Oncel Tuzel. Voxelnet: End-to-end learning for point cloud based 3d object detection. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4490-4499, 2017. 3 +[45] Walter Zimmer, Christian Creß, Huu Tung Nguyen, and Alois C Knoll. Tumtraf intersection dataset: All you need for urban 3d camera-lidar roadside perception [best student paper award]. In 2023 IEEE Intelligent Transportation Systems ITSC. IEEE, 2023. 8 \ No newline at end of file diff --git a/2025/SparseAlign_ a Fully Sparse Framework for Cooperative Object Detection/images.zip b/2025/SparseAlign_ a Fully Sparse Framework for Cooperative Object Detection/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..d6ca4252288ab241af08fe5bd3b433ac3baac72f --- /dev/null +++ b/2025/SparseAlign_ a Fully Sparse Framework for Cooperative Object Detection/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:253f1c3374e65f611b8534d84b8814526a30b97b2f31b4164ad89015d55c1f5d +size 526682 diff --git a/2025/SparseAlign_ a Fully Sparse Framework for Cooperative Object Detection/layout.json b/2025/SparseAlign_ a Fully Sparse Framework for Cooperative Object Detection/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..eadd6d668ba6edc9158ade2aa4aa5df3a122364a --- /dev/null +++ b/2025/SparseAlign_ a Fully Sparse Framework for Cooperative Object Detection/layout.json @@ -0,0 +1,9173 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 74, + 103, + 536, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 103, + 536, + 121 + ], + "spans": [ + { + "bbox": [ + 74, + 103, + 536, + 121 + ], + "type": "text", + "content": "SparseAlign: A Fully Sparse Framework for Cooperative Object Detection" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 148, + 477, + 166 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 148, + 477, + 166 + ], + "spans": [ + { + "bbox": [ + 132, + 148, + 477, + 166 + ], + "type": "text", + "content": "Yunshuang Yuan" + }, + { + "bbox": [ + 132, + 148, + 477, + 166 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 132, + 148, + 477, + 166 + ], + "type": "text", + "content": " Yan Xia" + }, + { + "bbox": [ + 132, + 148, + 477, + 166 + ], + "type": "inline_equation", + "content": "^{2,3\\dagger}" + }, + { + "bbox": [ + 132, + 148, + 477, + 166 + ], + "type": "text", + "content": " Daniel Cremers" + }, + { + "bbox": [ + 132, + 148, + 477, + 166 + ], + "type": "inline_equation", + "content": "^{2,3}" + }, + { + "bbox": [ + 132, + 148, + 477, + 166 + ], + "type": "text", + "content": " Monika Sester" + }, + { + "bbox": [ + 132, + 148, + 477, + 166 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 91, + 167, + 518, + 195 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 167, + 518, + 195 + ], + "spans": [ + { + "bbox": [ + 91, + 167, + 518, + 195 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 91, + 167, + 518, + 195 + ], + "type": "text", + "content": "Leibniz University Hannover " + }, + { + "bbox": [ + 91, + 167, + 518, + 195 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 91, + 167, + 518, + 195 + ], + "type": "text", + "content": "Technical University of Munich " + }, + { + "bbox": [ + 91, + 167, + 518, + 195 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 91, + 167, + 518, + 195 + ], + "type": "text", + "content": "Munich Center for Machine Learning (MCML) {yunshuang.yuan, monika.sester}@ikg.uni-hannover.de {yan.xia, cremers}@tum.de" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 222, + 200, + 234 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 222, + 200, + 234 + ], + "spans": [ + { + "bbox": [ + 151, + 222, + 200, + 234 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 247, + 297, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 247, + 297, + 475 + ], + "spans": [ + { + "bbox": [ + 54, + 247, + 297, + 475 + ], + "type": "text", + "content": "Cooperative perception can increase the view field and decrease the occlusion of an ego vehicle, hence improving the perception performance and safety of autonomous driving. Despite the success of previous works on cooperative object detection, they mostly operate on dense Bird's Eye View (BEV) feature maps, which are computationally demanding and can hardly be extended to long-range detection problems. More efficient fully sparse frameworks are rarely explored. In this work, we design a fully sparse framework, SparseAlign, with three key features: an enhanced sparse 3D backbone, a query-based temporal context learning module, and a robust detection head specially tailored for sparse features. Extensive experimental results on both OPV2V and DAIR-V2X datasets show that our framework, despite its sparsity, outperforms the state of the art with less communication bandwidth requirements. In addition, experiments on the OPV2Vt and DAIR-V2Xt datasets for time-aligned cooperative object detection also show a significant performance gain compared to the baseline works." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 496, + 135, + 509 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 496, + 135, + 509 + ], + "spans": [ + { + "bbox": [ + 55, + 496, + 135, + 509 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 516, + 295, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 516, + 295, + 658 + ], + "spans": [ + { + "bbox": [ + 55, + 516, + 295, + 658 + ], + "type": "text", + "content": "Cooperative Perception (CP) for autonomous driving has gained significant attention due to its potential to enhance road safety. Sharing Collective Perception Messages (CPMs) among Intelligent (traffic) Agents (IAs) in the vehicular network effectively reduces occlusions and expands the field of view, improving perception performance. In many recent works [1, 16, 24, 29, 35, 41, 42], CP has proven to be effective for different tasks, such as cooperative object detection [1, 24, 29, 41], Bird's Eye View (BEV) semantic segmentation [19, 42], and 3D occupancy semantic segmentation [22]. In this work, we focus on LiDAR-based cooperative vehicle detection." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 659, + 296, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 659, + 296, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 659, + 296, + 696 + ], + "type": "text", + "content": "Training a cooperative 3D object detector is computationally demanding as it needs to process data from multiple vehicles, especially when the model is built to di" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 315, + 224, + 553, + 285 + ], + "blocks": [ + { + "bbox": [ + 315, + 224, + 553, + 285 + ], + "lines": [ + { + "bbox": [ + 315, + 224, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 315, + 224, + 553, + 285 + ], + "type": "image", + "image_path": "3b92215a35f0c736e791874f42495a26171e0abf1fed4f4b07c69a6b308f2a61.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 315, + 289, + 553, + 369 + ], + "blocks": [ + { + "bbox": [ + 315, + 289, + 553, + 369 + ], + "lines": [ + { + "bbox": [ + 315, + 289, + 553, + 369 + ], + "spans": [ + { + "bbox": [ + 315, + 289, + 553, + 369 + ], + "type": "image", + "image_path": "1610062c385ae684802eeefeba04c5e130cda9c6944bc712aa6c35442716d363.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 379, + 555, + 424 + ], + "lines": [ + { + "bbox": [ + 313, + 379, + 555, + 424 + ], + "spans": [ + { + "bbox": [ + 313, + 379, + 555, + 424 + ], + "type": "text", + "content": "Figure 1. a,b) Communication strategies for cooperative object detection. c-e) Issues of sparse convolutional backbone networks. c) Center Feature Missing (CFM). d,e) Isolated Convolution Field (ICF) caused by ring disconnectivity and occlusion." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 448, + 555, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 448, + 555, + 664 + ], + "spans": [ + { + "bbox": [ + 313, + 448, + 555, + 664 + ], + "type": "text", + "content": "gest sequential data from several subsequent frames to learn the temporal context. Following Fcooper [1], recent works [6, 11, 28-30, 33, 35] encode the 3D point cloud into dense BEV feature maps for further processing, sharing, and fusion. However, the computational cost of this pipeline increases quadratically with the perception range. Although cooperation among IAs can improve long-range detection performance, the IAs are normally not in the best positions to optimize the performance gain. The long-range detection performance of the local detector directly influences the overall performance. Besides, sharing the BEV feature maps among IAs requires extensive communication resources. Instead, the sparse operations leverage the sparsity of point clouds and have only linear computational complexity relative to the number of points. In addition, the output sparse features can be more efficient for sharing. It is therefore beneficial to perform cooperative object detection based on sparse operations only." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 665, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 556, + 713 + ], + "type": "text", + "content": "However, building a fully sparse framework to compete with the dense ones is non-trivial. The first unavoidable issue is Center Feature Missing [8] (CFM, shown in Fig. 1 c)). Due to the range view of on-board LiDARs, there are" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 702, + 144, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 702, + 144, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 702, + 144, + 712 + ], + "type": "text", + "content": "† Corresponding author." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "22296" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 300 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 300 + ], + "type": "text", + "content": "usually no scan points in the vehicles' center area. However, compared to the edge points, the center points have a better ability to represent the whole object. Lacking center points for learning, the sparse frameworks perform worse than the dense ones that directly learn the features for center points. The second issue is the poor connectivity between points scanned by different laser beams. As shown in Fig. 1 d), this disconnectivity issue in distant areas may lead to isolated voxel blobs; the receptive field only enlarges to the blob scale as the convolutional layers go deeper. The isolated voxel blobs never exchange information with each other, limiting the backbone's capability of capturing global features. We call this issue Isolated Convolution Field (ICF). It also happens in the occlusion areas as shown in Fig. 1 e): although the occluded vehicle has very few scan points (red), it is expected to be detected based on the neighboring vehicles and the clues in the previous frames. However, ICF introduces difficulties for these isolated points to aggregate information from neighbors and the global environment." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 301, + 295, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 301, + 295, + 517 + ], + "spans": [ + { + "bbox": [ + 56, + 301, + 295, + 517 + ], + "type": "text", + "content": "To overcome the above problems, we build a Sparse UNet backbone network SUNet with the Coordinate-Expandable sparse Convolution (CEC) implemented in the MinkowskiEngine [5] library. We apply CEC on " + }, + { + "bbox": [ + 56, + 301, + 295, + 517 + ], + "type": "inline_equation", + "content": "4\\times" + }, + { + "bbox": [ + 56, + 301, + 295, + 517 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 56, + 301, + 295, + 517 + ], + "type": "inline_equation", + "content": "8\\times" + }, + { + "bbox": [ + 56, + 301, + 295, + 517 + ], + "type": "text", + "content": " downsampled 3D voxels to increase the connectivity between ICFs. On the 2D BEV features, the CECs are used to expand the sparse coordinates, assuring the center features for all scanned vehicles. On top of SUNet, we build TempAlign Module (TAM) to learn the temporal context from the historical frames in a query-based [23, 40] manner, PoseAlign Module (PAM) to correct the relative pose errors between the cooperative and the ego agents, and SpatialAlign Module (SAM) to fuse the shared object queries based on the corrected poses. In addition, we propose the CompassRose encoding for the regression of BBox directions to further improve performance on both Cooperative Object Detection (COOD) and the Time-Aligned COOD [40] (TA-COOD) tasks." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 520, + 289, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 520, + 289, + 531 + ], + "spans": [ + { + "bbox": [ + 67, + 520, + 289, + 531 + ], + "type": "text", + "content": "To summarize, the main contributions of this work are:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 534, + 293, + 710 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 56, + 534, + 293, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 534, + 293, + 566 + ], + "spans": [ + { + "bbox": [ + 56, + 534, + 293, + 566 + ], + "type": "text", + "content": "- We investigate the issues of building a fully sparse network for LiDAR-based COOD and analyze how these issues impact the detection performance." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 570, + 293, + 627 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 570, + 293, + 627 + ], + "spans": [ + { + "bbox": [ + 56, + 570, + 293, + 627 + ], + "type": "text", + "content": "- We propose a novel fully sparse network, SparseAlign, via addressing the ICF and CFM problems of the sparse 3D backbone SUNet with CECs. Additionally, we introduce the SAM module, specifically tailored to ensure accurate fusion of sparse features." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 630, + 293, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 630, + 293, + 687 + ], + "spans": [ + { + "bbox": [ + 56, + 630, + 293, + 687 + ], + "type": "text", + "content": "- We assess the communication bandwidth and precision trade-offs of the different approaches, finding that our proposal reduces bandwidth consumption by up to " + }, + { + "bbox": [ + 56, + 630, + 293, + 687 + ], + "type": "inline_equation", + "content": "98\\%" + }, + { + "bbox": [ + 56, + 630, + 293, + 687 + ], + "type": "text", + "content": ", compared to previous methods using dense BEV feature maps for LiDAR-based cooperative object detection." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 689, + 293, + 710 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 689, + 293, + 710 + ], + "spans": [ + { + "bbox": [ + 56, + 689, + 293, + 710 + ], + "type": "text", + "content": "We conduct extensive experiments on OPV2V, DAIR-V2X, OPV2Vt, and DAIR-V2Xt benchmarks and show" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 323, + 72, + 553, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 72, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 323, + 72, + 553, + 95 + ], + "type": "text", + "content": "that the proposed SparseAlign has a significant improvement over state-of-the-art methods." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 110, + 400, + 122 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 110, + 400, + 122 + ], + "spans": [ + { + "bbox": [ + 314, + 110, + 400, + 122 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 130, + 518, + 144 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 130, + 518, + 144 + ], + "spans": [ + { + "bbox": [ + 313, + 130, + 518, + 144 + ], + "type": "text", + "content": "2.1. Cooperative Object Detection (COOD)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 149, + 553, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 149, + 553, + 424 + ], + "spans": [ + { + "bbox": [ + 313, + 149, + 553, + 424 + ], + "type": "text", + "content": "COOD frameworks are classified into sharing CPMs with raw data, intermediate features, and detections. Recent works [2, 24, 29, 38] have shown that intermediate feature sharing has the best performance due to its adjustable and learnable CPMs. We also use this strategy in our framework. In the perspective of communication, there are two strategies: handshaking and broadcasting as shown in Fig. 1 a) and b). With handshaking communication [11, 28-30, 33, 40, 42], the ego vehicle sends a request message that contains its pose to the cooperative IAs so that they can transform the local raw data to the ego coordinate system and then process their data to obtain the features for sharing. This simplifies the feature fusion process, as the features are directly learned in the goal coordinate system. In contrast, the broadcasting method [6, 17, 24, 37] directly processes the data in the local coordinate system and transmits it to other IAs. In the fusion stage, the ego vehicle must transform the cooperative features into the ego coordinate system and then conduct the fusion process. In comparison, broadcasting is more efficient as it does not require a prior communication agreement with the cooperative agents in highly dynamic driving scenarios where the communication partners change frequently." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 426, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 553, + 713 + ], + "type": "text", + "content": "Achieving COOD presents several challenges, including communication latency and bandwidth limitations, inaccurate IA poses, and sensor asynchrony. To reduce the communication bandwidth consumption, [1, 24, 29, 33] compress the learned BEV feature maps to fewer channels via convolution layers, while [13] utilizes the learned confidence map and [42] leverages the uncertainty map to select the most important information for sharing. Instead, [41] employs the two-stage detector PVRCNN [20] and shares the 3D keypoint features from the second stage. To be more compact, [4, 9, 40] share the object query features [23] which are highly flexible to blend all relevant information, such as scanning time and IA pose, about an object into one unique feature vector. In this work, we also leverage the object query features to learn the temporal and spatial context of the object. Pose errors can dramatically deteriorate the detection results [24]. To this end, [39, 41] leverage the predicted Bounding Boxes (BBoxes) and select semantically classified keypoints to align the relative pose between the ego and the cooperative vehicles. Instead, [17] solely uses the BBoxes for pose alignment via the Pose Graph Optimization (PGO). However, these methods require the initial pose errors to be small to find the correct object matching in the first step and then iterate many times, or use the PGO" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22297" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 71, + 349, + 168 + ], + "blocks": [ + { + "bbox": [ + 58, + 71, + 349, + 168 + ], + "lines": [ + { + "bbox": [ + 58, + 71, + 349, + 168 + ], + "spans": [ + { + "bbox": [ + 58, + 71, + 349, + 168 + ], + "type": "image", + "image_path": "44cc1539675ad7111a932d1986704015c5161b48351f5c39168527d394fa3ad6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 353, + 71, + 549, + 168 + ], + "blocks": [ + { + "bbox": [ + 353, + 71, + 549, + 168 + ], + "lines": [ + { + "bbox": [ + 353, + 71, + 549, + 168 + ], + "spans": [ + { + "bbox": [ + 353, + 71, + 549, + 168 + ], + "type": "image", + "image_path": "95d04a62207224d1fbfbf49826dc9e9db10a85618467f71ec704117966537bd3.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 58, + 169, + 312, + 280 + ], + "blocks": [ + { + "bbox": [ + 58, + 169, + 312, + 280 + ], + "lines": [ + { + "bbox": [ + 58, + 169, + 312, + 280 + ], + "spans": [ + { + "bbox": [ + 58, + 169, + 312, + 280 + ], + "type": "image", + "image_path": "68687900a7229ab0a7b2604115550d9bc2d6419e05d0af4d413294535b8c18f3.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 63, + 281, + 82, + 287 + ], + "lines": [ + { + "bbox": [ + 63, + 281, + 82, + 287 + ], + "spans": [ + { + "bbox": [ + 63, + 281, + 82, + 287 + ], + "type": "text", + "content": "SUNet" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 312, + 169, + 429, + 280 + ], + "blocks": [ + { + "bbox": [ + 312, + 169, + 429, + 280 + ], + "lines": [ + { + "bbox": [ + 312, + 169, + 429, + 280 + ], + "spans": [ + { + "bbox": [ + 312, + 169, + 429, + 280 + ], + "type": "image", + "image_path": "92b7e57fa75e7da32a2ccc6dd89d886434e1cd007168c160a9c99ed9968218aa.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 315, + 281, + 380, + 289 + ], + "lines": [ + { + "bbox": [ + 315, + 281, + 380, + 289 + ], + "spans": [ + { + "bbox": [ + 315, + 281, + 380, + 289 + ], + "type": "text", + "content": "TempAlign Module (TAM)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 432, + 170, + 549, + 281 + ], + "blocks": [ + { + "bbox": [ + 432, + 170, + 549, + 281 + ], + "lines": [ + { + "bbox": [ + 432, + 170, + 549, + 281 + ], + "spans": [ + { + "bbox": [ + 432, + 170, + 549, + 281 + ], + "type": "image", + "image_path": "613cabf80a43f8cd27c2bee3c4e834dcff5a64bb3bbfb4f08604407524ea99f1.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 435, + 281, + 500, + 289 + ], + "lines": [ + { + "bbox": [ + 435, + 281, + 500, + 289 + ], + "spans": [ + { + "bbox": [ + 435, + 281, + 500, + 289 + ], + "type": "text", + "content": "PoseAlign Module (PAM)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 219, + 303, + 390, + 315 + ], + "lines": [ + { + "bbox": [ + 219, + 303, + 390, + 315 + ], + "spans": [ + { + "bbox": [ + 219, + 303, + 390, + 315 + ], + "type": "text", + "content": "Figure 2. Overview of SparseAlign framework" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 335, + 297, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 335, + 297, + 563 + ], + "spans": [ + { + "bbox": [ + 54, + 335, + 297, + 563 + ], + "type": "text", + "content": "to refine the result. In this paper, we propose to learn pose-agnostic neighborhood graph features to accurately find the correct matching and then use PGO to refine it. Communication delay can cause that data arriving at the ego agent is outdated. To compensate for this delay, [36, 37] add the first-order derivatives of the BEV feature map to the original feature map to obtain the updated feature map that is aligned to the ego timestamp; [15] employs the LSTM [12] structure on the BEV feature maps to adjust the features to be aligned. Differently, [40] embeds the observation timestamps as well as the temporal context from historical frames into the object queries so that they are time- and motion-aware and capable of predicting the status at any given future timestamp, achieving the time-delay compensation. Different to [15, 36, 37] that compensate the feature maps, [40] leverages the point-wise timestamp of the point clouds to learn the accurate temporal context, hence is also able to cope with sensor asynchrony issues. Therefore, we adopt the query-based strategy for time-related compensation." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 574, + 145, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 574, + 145, + 586 + ], + "spans": [ + { + "bbox": [ + 55, + 574, + 145, + 586 + ], + "type": "text", + "content": "2.2.3DBackbones" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "text", + "content": "To extract point cloud features, most previous works [16, 17, 24, 29, 33, 37, 41] for COOD either use dense 3D [25, 44], dense 2D [14, 32] or semi-sparse [21, 26, 31, 43] backbones that first encode the point cloud with 3D sparse convolution and then apply dense 2D convolutions for further processing. In contrast, [40, 42] employ the MinkUnet [5] to extract point cloud features. Despite its efficiency, the MinkUnet backbone struggles with long-range detection, occlusions and large objects, because of CFM and ICF problems. [8] proposed center voting for each scanned ob" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 335, + 555, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 335, + 555, + 419 + ], + "spans": [ + { + "bbox": [ + 313, + 335, + 555, + 419 + ], + "type": "text", + "content": "ject points to solve CFM issue and [3] reveals that further down-sampling the voxel features to lower resolutions can increase the receptive field and refine the large object detection. Instead of shifting the feature locations as [8] to accurately locate object queries or adding additional layers as [3], we build a backbone that solves both CFM and ICF issues by modifying MinkUnet with CEC layers." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 434, + 370, + 445 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 434, + 370, + 445 + ], + "spans": [ + { + "bbox": [ + 313, + 434, + 370, + 445 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 455, + 424, + 466 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 455, + 424, + 466 + ], + "spans": [ + { + "bbox": [ + 313, + 455, + 424, + 466 + ], + "type": "text", + "content": "3.1. Task formalization" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 474, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 474, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 474, + 556, + 715 + ], + "type": "text", + "content": "In this work, we solve the LiDAR-based cooperative object detection task considering unavoidable challenges in reality: communication delay, sensor asynchrony, and localization errors. For this scenario, we assume there are " + }, + { + "bbox": [ + 313, + 474, + 556, + 715 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 474, + 556, + 715 + ], + "type": "text", + "content": " IAs, including one ego agent " + }, + { + "bbox": [ + 313, + 474, + 556, + 715 + ], + "type": "inline_equation", + "content": "A_{1}" + }, + { + "bbox": [ + 313, + 474, + 556, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 474, + 556, + 715 + ], + "type": "inline_equation", + "content": "N - 1" + }, + { + "bbox": [ + 313, + 474, + 556, + 715 + ], + "type": "text", + "content": " cooperative agents. These agents are sharing CPMs with each other in a broadcasting manner, as shown in Fig. 1b). Specifically, in the " + }, + { + "bbox": [ + 313, + 474, + 556, + 715 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 474, + 556, + 715 + ], + "type": "text", + "content": "-th frame, " + }, + { + "bbox": [ + 313, + 474, + 556, + 715 + ], + "type": "inline_equation", + "content": "A_{1}" + }, + { + "bbox": [ + 313, + 474, + 556, + 715 + ], + "type": "text", + "content": " receives the CPMs " + }, + { + "bbox": [ + 313, + 474, + 556, + 715 + ], + "type": "inline_equation", + "content": "M = \\{(m_j,t_j)|j\\in \\{1,\\dots ,N - 1\\} \\}" + }, + { + "bbox": [ + 313, + 474, + 556, + 715 + ], + "type": "text", + "content": " from the cooperative agents, where " + }, + { + "bbox": [ + 313, + 474, + 556, + 715 + ], + "type": "inline_equation", + "content": "m_j" + }, + { + "bbox": [ + 313, + 474, + 556, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 474, + 556, + 715 + ], + "type": "inline_equation", + "content": "t_j" + }, + { + "bbox": [ + 313, + 474, + 556, + 715 + ], + "type": "text", + "content": " are the CPM and the corresponding timestamp from the " + }, + { + "bbox": [ + 313, + 474, + 556, + 715 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 474, + 556, + 715 + ], + "type": "text", + "content": "-th agent. We aim to optimize the object detection results based on the point clouds captured so far (" + }, + { + "bbox": [ + 313, + 474, + 556, + 715 + ], + "type": "inline_equation", + "content": "t\\leq t_i" + }, + { + "bbox": [ + 313, + 474, + 556, + 715 + ], + "type": "text", + "content": ") at " + }, + { + "bbox": [ + 313, + 474, + 556, + 715 + ], + "type": "inline_equation", + "content": "A_{1}" + }, + { + "bbox": [ + 313, + 474, + 556, + 715 + ], + "type": "text", + "content": " and the received messages " + }, + { + "bbox": [ + 313, + 474, + 556, + 715 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 474, + 556, + 715 + ], + "type": "text", + "content": ". We model and evaluate both the COOD and TA-COOD tasks from the perspective of an ego agent. The key difference between COOD and TA-COOD lies in the GT BBoxes. As illustrated in Fig. 2, TAM, COOD GT BBoxes (dashed lines) align with the sequentially captured LiDAR points, whereas TA-COD GT BBoxes (solid lines) are referenced to a global timestamp " + }, + { + "bbox": [ + 313, + 474, + 556, + 715 + ], + "type": "inline_equation", + "content": "t_g" + }, + { + "bbox": [ + 313, + 474, + 556, + 715 + ], + "type": "text", + "content": ". This distinction requires the model to accurately predict" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "22298" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 72, + 118, + 131 + ], + "blocks": [ + { + "bbox": [ + 61, + 72, + 118, + 131 + ], + "lines": [ + { + "bbox": [ + 61, + 72, + 118, + 131 + ], + "spans": [ + { + "bbox": [ + 61, + 72, + 118, + 131 + ], + "type": "image", + "image_path": "682b68fa3080748eca50c1f79abb0047e99ec61a838df4a0ff883c2393331b16.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 140, + 295, + 183 + ], + "lines": [ + { + "bbox": [ + 55, + 140, + 295, + 183 + ], + "spans": [ + { + "bbox": [ + 55, + 140, + 295, + 183 + ], + "type": "text", + "content": "Figure 3. Receptive field (RF) after 4 sparse convolutions (convs). a. LiDAR points; b. Spare pixels (white) with a conv center point (red); c. RF coverage of 4 normal sparse convs (red); d. RF coverage of 4 CECs (red+light blue, darkest blue is background)." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 120, + 72, + 176, + 129 + ], + "blocks": [ + { + "bbox": [ + 120, + 72, + 176, + 129 + ], + "lines": [ + { + "bbox": [ + 120, + 72, + 176, + 129 + ], + "spans": [ + { + "bbox": [ + 120, + 72, + 176, + 129 + ], + "type": "image", + "image_path": "9e202ffca0ac72622c4111bc1cee2959b3e65db2646a1f716e10f0a6327cdc71.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 178, + 72, + 234, + 129 + ], + "blocks": [ + { + "bbox": [ + 178, + 72, + 234, + 129 + ], + "lines": [ + { + "bbox": [ + 178, + 72, + 234, + 129 + ], + "spans": [ + { + "bbox": [ + 178, + 72, + 234, + 129 + ], + "type": "image", + "image_path": "bdbdc5d946e97c88ac9dafc4b3ca949f19f98df3ccb231c8999295f2aa952b5a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 236, + 72, + 292, + 129 + ], + "blocks": [ + { + "bbox": [ + 236, + 72, + 292, + 129 + ], + "lines": [ + { + "bbox": [ + 236, + 72, + 292, + 129 + ], + "spans": [ + { + "bbox": [ + 236, + 72, + 292, + 129 + ], + "type": "image", + "image_path": "2c53502a3f07b26b836b892ff66e5e4cd6e3c960e129e07eb3436935a53573ef.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 195, + 296, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 195, + 296, + 255 + ], + "spans": [ + { + "bbox": [ + 55, + 195, + 296, + 255 + ], + "type": "text", + "content": "BBox locations at " + }, + { + "bbox": [ + 55, + 195, + 296, + 255 + ], + "type": "inline_equation", + "content": "t_g" + }, + { + "bbox": [ + 55, + 195, + 296, + 255 + ], + "type": "text", + "content": ". While COOD relies solely on point cloud geometry for location prediction, TA-COOD leverages temporal context to enhance accuracy at " + }, + { + "bbox": [ + 55, + 195, + 296, + 255 + ], + "type": "inline_equation", + "content": "t_g" + }, + { + "bbox": [ + 55, + 195, + 296, + 255 + ], + "type": "text", + "content": " and also enables the model to handle location errors due to communication latency." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 262, + 178, + 274 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 262, + 178, + 274 + ], + "spans": [ + { + "bbox": [ + 55, + 262, + 178, + 274 + ], + "type": "text", + "content": "3.2. Framework overview" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 281, + 296, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 281, + 296, + 568 + ], + "spans": [ + { + "bbox": [ + 55, + 281, + 296, + 568 + ], + "type": "text", + "content": "We build an efficient fully sparse framework SparseAlign, as shown in Fig. 2, to process sequential point clouds from multiple agents. All agents share the same network weights. The input at time " + }, + { + "bbox": [ + 55, + 281, + 296, + 568 + ], + "type": "inline_equation", + "content": "t_i" + }, + { + "bbox": [ + 55, + 281, + 296, + 568 + ], + "type": "text", + "content": " is the point cloud of the ego agent " + }, + { + "bbox": [ + 55, + 281, + 296, + 568 + ], + "type": "inline_equation", + "content": "PC_{t_i}" + }, + { + "bbox": [ + 55, + 281, + 296, + 568 + ], + "type": "text", + "content": " and of the cooperative agents " + }, + { + "bbox": [ + 55, + 281, + 296, + 568 + ], + "type": "inline_equation", + "content": "PC_{t_j}" + }, + { + "bbox": [ + 55, + 281, + 296, + 568 + ], + "type": "text", + "content": ". Note that " + }, + { + "bbox": [ + 55, + 281, + 296, + 568 + ], + "type": "inline_equation", + "content": "t_j < t_i" + }, + { + "bbox": [ + 55, + 281, + 296, + 568 + ], + "type": "text", + "content": " when communication latency exists. All point clouds are processed by the backbone SUNet to extract the point cloud features. The Region of Interest (RoI) module then selects the most interesting features as the object queries. For simplicity, we define object queries as " + }, + { + "bbox": [ + 55, + 281, + 296, + 568 + ], + "type": "inline_equation", + "content": "Q^{*} = \\{(F_{i}^{*}, x_{i}^{*}, y_{i}^{*}) | i \\in \\{1, \\dots, N^{*}\\}" + }, + { + "bbox": [ + 55, + 281, + 296, + 568 + ], + "type": "text", + "content": ", where * represents the query set name, " + }, + { + "bbox": [ + 55, + 281, + 296, + 568 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 55, + 281, + 296, + 568 + ], + "type": "text", + "content": " is the features, and " + }, + { + "bbox": [ + 55, + 281, + 296, + 568 + ], + "type": "inline_equation", + "content": "x, y" + }, + { + "bbox": [ + 55, + 281, + 296, + 568 + ], + "type": "text", + "content": " are the coordinates of the corresponding " + }, + { + "bbox": [ + 55, + 281, + 296, + 568 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 281, + 296, + 568 + ], + "type": "text", + "content": "-th query. With selected queries, the TempAlign Module (TAM) can efficiently align the query features to the globally aligned timestamp " + }, + { + "bbox": [ + 55, + 281, + 296, + 568 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 281, + 296, + 568 + ], + "type": "text", + "content": " so that errors introduced by communication delay and sensor asynchrony are compensated. The updated query features and the detections from the Local Query Detection head (LQDet) are then shared as CPMs. The PoseAlign Module (PAM) then uses the detected BBoxes to correct the relative poses between the cooperative and the ego agent and the SpatialAlign Module (SAM) fuses the received cooperative features into the ego coordinate system based on the corrected transformation " + }, + { + "bbox": [ + 55, + 281, + 296, + 568 + ], + "type": "inline_equation", + "content": "T_{c}^{e}" + }, + { + "bbox": [ + 55, + 281, + 296, + 568 + ], + "type": "text", + "content": " from cooperative to ego coordinates." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 575, + 231, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 575, + 231, + 588 + ], + "spans": [ + { + "bbox": [ + 55, + 575, + 231, + 588 + ], + "type": "text", + "content": "3.3. Backbone: Sparse UNet (SUNet)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "text", + "content": "Processing data from multiple IAs for cooperative perception is computationally demanding. To improve efficiency, we use a fully sparse 3D UNet, similar to MinkUnet[5], to construct the SUNet (Fig. 2 bottom left). However, the fully sparse convolutions used in such networks suffer from the ICF issue. As shown in Fig. 3, the receptive field (RF) of (c) standard sparse convolutions only covers LiDAR points from a single vehicle. In contrast, the RF of (d) the CEC layers extends to the points of a neighboring vehicle, enabling learning within a larger spatial context. Therefore," + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 554, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 204 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 204 + ], + "type": "text", + "content": "we use 3D CECs in the sparse convolution (SConv) block 3 and 4. Correspondingly, we contract the coordinates in the transposed sparse convolution (TConv) block 3 and 2. In this way, the output of the TConv blocks " + }, + { + "bbox": [ + 313, + 72, + 554, + 204 + ], + "type": "inline_equation", + "content": "p4" + }, + { + "bbox": [ + 313, + 72, + 554, + 204 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 72, + 554, + 204 + ], + "type": "inline_equation", + "content": "p2" + }, + { + "bbox": [ + 313, + 72, + 554, + 204 + ], + "type": "text", + "content": " can match the coordinates in stride " + }, + { + "bbox": [ + 313, + 72, + 554, + 204 + ], + "type": "inline_equation", + "content": "x2" + }, + { + "bbox": [ + 313, + 72, + 554, + 204 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 72, + 554, + 204 + ], + "type": "inline_equation", + "content": "x4" + }, + { + "bbox": [ + 313, + 72, + 554, + 204 + ], + "type": "text", + "content": " for the concatenation. In addition, it learns better global features without voxel number soaring in the subsequent layers and maintains efficiency. Against the CFM issue, we use 2D CECs on the 2D BEV sparse features (Fig. 2 SUNet:Coor. Expand.) to expand the coordinates to ensure their coverage over the object centers." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 211, + 428, + 224 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 211, + 428, + 224 + ], + "spans": [ + { + "bbox": [ + 313, + 211, + 428, + 224 + ], + "type": "text", + "content": "3.4. Alignment modules" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 228, + 555, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 228, + 555, + 456 + ], + "spans": [ + { + "bbox": [ + 313, + 228, + 555, + 456 + ], + "type": "text", + "content": "TempAlign Module (TAM). The TAM aims to learn the temporal context relative to the previous frames to achieve two goals: improve object detection performance by looking for clues in the history and compensate for the object displacement error introduced by asynchronous scanning time (Fig. 2, TAM). We achieve the second goal by predicting the BBoxes at the globally aligned timestamp so that the BBoxes shared for PAM are synchronized. Then PAM only needs to deal with the object displacement caused by localization errors. Technically, the TAM follows [40], for efficiency, by removing the global cross-attention module which learns the global BEV features. We only use a memory queue and a hybrid attention module. The memory queue stores the top " + }, + { + "bbox": [ + 313, + 228, + 555, + 456 + ], + "type": "inline_equation", + "content": "k = 256" + }, + { + "bbox": [ + 313, + 228, + 555, + 456 + ], + "type": "text", + "content": " object queries from previous " + }, + { + "bbox": [ + 313, + 228, + 555, + 456 + ], + "type": "inline_equation", + "content": "L = 4" + }, + { + "bbox": [ + 313, + 228, + 555, + 456 + ], + "type": "text", + "content": " data frames so that the attention module can allow the interaction between object queries in the current frame and the historical frames. In this way, the object queries learn the temporal context from the sequential data and are able to make predictions for the future state." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 456, + 556, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 456, + 556, + 696 + ], + "spans": [ + { + "bbox": [ + 313, + 456, + 556, + 696 + ], + "type": "text", + "content": "PoseAlign Module (PAM). Given the detected BBoxes " + }, + { + "bbox": [ + 313, + 456, + 556, + 696 + ], + "type": "inline_equation", + "content": "\\mathbf{B}_i" + }, + { + "bbox": [ + 313, + 456, + 556, + 696 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 456, + 556, + 696 + ], + "type": "inline_equation", + "content": "\\mathbf{B}_j" + }, + { + "bbox": [ + 313, + 456, + 556, + 696 + ], + "type": "text", + "content": ", respectively, from the perspective of the agent " + }, + { + "bbox": [ + 313, + 456, + 556, + 696 + ], + "type": "inline_equation", + "content": "A_i" + }, + { + "bbox": [ + 313, + 456, + 556, + 696 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 456, + 556, + 696 + ], + "type": "inline_equation", + "content": "A_j" + }, + { + "bbox": [ + 313, + 456, + 556, + 696 + ], + "type": "text", + "content": " in the same scene, humans can easily find the correspondences between " + }, + { + "bbox": [ + 313, + 456, + 556, + 696 + ], + "type": "inline_equation", + "content": "\\mathbf{B}_i" + }, + { + "bbox": [ + 313, + 456, + 556, + 696 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 456, + 556, + 696 + ], + "type": "inline_equation", + "content": "\\mathbf{B}_j" + }, + { + "bbox": [ + 313, + 456, + 556, + 696 + ], + "type": "text", + "content": " based on the geometric relationships. Practically, we model the features of these geometric relationships by embedding the neighborhood geometry of each BBox into deep features and then match these features of " + }, + { + "bbox": [ + 313, + 456, + 556, + 696 + ], + "type": "inline_equation", + "content": "\\mathbf{B}_i" + }, + { + "bbox": [ + 313, + 456, + 556, + 696 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 456, + 556, + 696 + ], + "type": "inline_equation", + "content": "\\mathbf{B}_j" + }, + { + "bbox": [ + 313, + 456, + 556, + 696 + ], + "type": "text", + "content": " with the Hungarian algorithm. For each BBox " + }, + { + "bbox": [ + 313, + 456, + 556, + 696 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 313, + 456, + 556, + 696 + ], + "type": "text", + "content": " in a BBox set " + }, + { + "bbox": [ + 313, + 456, + 556, + 696 + ], + "type": "inline_equation", + "content": "\\mathbf{B}" + }, + { + "bbox": [ + 313, + 456, + 556, + 696 + ], + "type": "text", + "content": ", we select the nearest " + }, + { + "bbox": [ + 313, + 456, + 556, + 696 + ], + "type": "inline_equation", + "content": "k = 8" + }, + { + "bbox": [ + 313, + 456, + 556, + 696 + ], + "type": "text", + "content": " neighbors to embed its neighborhood features. As shown in Fig. 2, PAM: 1.Graph Feature, the orientation angles of the BBox are " + }, + { + "bbox": [ + 313, + 456, + 556, + 696 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 456, + 556, + 696 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 456, + 556, + 696 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 313, + 456, + 556, + 696 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 313, + 456, + 556, + 696 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 313, + 456, + 556, + 696 + ], + "type": "text", + "content": " and its neighbor, respectively. We embed the relative orientation with " + }, + { + "bbox": [ + 313, + 456, + 556, + 696 + ], + "type": "inline_equation", + "content": "\\nu_{a} = [\\sin (\\beta -\\alpha);\\cos (\\beta -\\alpha)]^{1}" + }, + { + "bbox": [ + 313, + 456, + 556, + 696 + ], + "type": "text", + "content": ". Similarly, the relative edge direction is embedded as " + }, + { + "bbox": [ + 313, + 456, + 556, + 696 + ], + "type": "inline_equation", + "content": "\\epsilon_{a} = [\\sin (\\theta -\\alpha);\\cos (\\theta -\\alpha)]" + }, + { + "bbox": [ + 313, + 456, + 556, + 696 + ], + "type": "text", + "content": ". In addition, the dimension " + }, + { + "bbox": [ + 313, + 456, + 556, + 696 + ], + "type": "inline_equation", + "content": "\\nu_{dim} = [l,w,h]" + }, + { + "bbox": [ + 313, + 456, + 556, + 696 + ], + "type": "text", + "content": " of the neighbor BBox, and the Euclidean distance between " + }, + { + "bbox": [ + 313, + 456, + 556, + 696 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 313, + 456, + 556, + 696 + ], + "type": "text", + "content": " and its neighbors " + }, + { + "bbox": [ + 313, + 456, + 556, + 696 + ], + "type": "inline_equation", + "content": "\\epsilon_d = d" + }, + { + "bbox": [ + 313, + 456, + 556, + 696 + ], + "type": "text", + "content": ", are also embedded into the feature. Note that " + }, + { + "bbox": [ + 313, + 456, + 556, + 696 + ], + "type": "inline_equation", + "content": "\\epsilon_d,\\epsilon_a,\\nu_a" + }, + { + "bbox": [ + 313, + 456, + 556, + 696 + ], + "type": "text", + "content": " are all relative features between two BBoxes that are pose-agnostic, leading to robust pose alignment without" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 325, + 702, + 394, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 702, + 394, + 713 + ], + "spans": [ + { + "bbox": [ + 325, + 702, + 394, + 713 + ], + "type": "text", + "content": "1[.;.]:concatenation" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "22299" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "text", + "content": "dependency on the magnitude of initial pose errors. These features are then concatenated and embedded into high-dimensional features by a linear layer with ReLU activation " + }, + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "inline_equation", + "content": "(\\mathbb{R}^5\\to \\mathbb{R}^{d_{emb}})" + }, + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 108, + 129, + 295, + 142 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 129, + 295, + 142 + ], + "spans": [ + { + "bbox": [ + 108, + 129, + 295, + 142 + ], + "type": "interline_equation", + "content": "f _ {n b r} = \\operatorname {l i n e a r} \\left(\\left[ \\epsilon_ {d}; \\epsilon_ {a}; \\nu_ {a}; \\nu_ {\\text {d i m}} \\right]\\right) \\tag {1}", + "image_path": "9b6015ad206075dded46bc509e51f72dff4b4c848007fb8c973d65a3d980122e.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 149, + 296, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 149, + 296, + 209 + ], + "spans": [ + { + "bbox": [ + 55, + 149, + 296, + 209 + ], + "type": "text", + "content": "Assembling the features relative to all neighbors, we summarize the neighborhood feature " + }, + { + "bbox": [ + 55, + 149, + 296, + 209 + ], + "type": "inline_equation", + "content": "F_{nbr} \\in \\mathbb{R}^{k \\times d_{emb}}" + }, + { + "bbox": [ + 55, + 149, + 296, + 209 + ], + "type": "text", + "content": " by a multi-head self-attention module " + }, + { + "bbox": [ + 55, + 149, + 296, + 209 + ], + "type": "inline_equation", + "content": "attn" + }, + { + "bbox": [ + 55, + 149, + 296, + 209 + ], + "type": "text", + "content": " followed by a linear layer operated on the mean and max of the " + }, + { + "bbox": [ + 55, + 149, + 296, + 209 + ], + "type": "inline_equation", + "content": "attn" + }, + { + "bbox": [ + 55, + 149, + 296, + 209 + ], + "type": "text", + "content": " outcome, resulting in the feature " + }, + { + "bbox": [ + 55, + 149, + 296, + 209 + ], + "type": "inline_equation", + "content": "F_{nbr} \\in \\mathbb{R}^{d_{emb}}" + }, + { + "bbox": [ + 55, + 149, + 296, + 209 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 85, + 217, + 295, + 230 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 217, + 295, + 230 + ], + "spans": [ + { + "bbox": [ + 85, + 217, + 295, + 230 + ], + "type": "interline_equation", + "content": "F _ {n b r} = \\operatorname {a t t n} \\left(F _ {n b r}\\right) \\tag {2}", + "image_path": "bd429ad2935ab1ca1664a74b55e017cdeded5ee9f317785927b993f20e484221.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 86, + 232, + 295, + 246 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 232, + 295, + 246 + ], + "spans": [ + { + "bbox": [ + 86, + 232, + 295, + 246 + ], + "type": "interline_equation", + "content": "F _ {n b r} = \\operatorname {l i n e a r} \\left(\\left[ \\operatorname {m e a n} \\left(F _ {n b r}\\right); \\operatorname {m a x} \\left(F _ {n b r}\\right) \\right]\\right) \\tag {3}", + "image_path": "f1b88de14d0f1632e02f8e9256574909ba5eb40670d67feda371b2adc701e1f2.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 253, + 296, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 253, + 296, + 396 + ], + "spans": [ + { + "bbox": [ + 55, + 253, + 296, + 396 + ], + "type": "text", + "content": "By calculating the Euclidean distance between the learned neighborhood feature " + }, + { + "bbox": [ + 55, + 253, + 296, + 396 + ], + "type": "inline_equation", + "content": "F_{nbr}" + }, + { + "bbox": [ + 55, + 253, + 296, + 396 + ], + "type": "text", + "content": " of the BBox in " + }, + { + "bbox": [ + 55, + 253, + 296, + 396 + ], + "type": "inline_equation", + "content": "\\mathbf{B}_i" + }, + { + "bbox": [ + 55, + 253, + 296, + 396 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 253, + 296, + 396 + ], + "type": "inline_equation", + "content": "\\mathbf{B}_j" + }, + { + "bbox": [ + 55, + 253, + 296, + 396 + ], + "type": "text", + "content": ", we construct the cost matrix for the linear sum assignment to find the best Match (Fig. 2, PAM: 2-match) between these two BBox sets. However, an object in " + }, + { + "bbox": [ + 55, + 253, + 296, + 396 + ], + "type": "inline_equation", + "content": "\\mathbf{B}_i" + }, + { + "bbox": [ + 55, + 253, + 296, + 396 + ], + "type": "text", + "content": " does not always have a projection in " + }, + { + "bbox": [ + 55, + 253, + 296, + 396 + ], + "type": "inline_equation", + "content": "\\mathbf{B}_j" + }, + { + "bbox": [ + 55, + 253, + 296, + 396 + ], + "type": "text", + "content": " and vice versa. Thus, we reject the wrong matches that have large distances. Based on the PAM: 2-match result, the relative pose transformation between " + }, + { + "bbox": [ + 55, + 253, + 296, + 396 + ], + "type": "inline_equation", + "content": "A_i" + }, + { + "bbox": [ + 55, + 253, + 296, + 396 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 253, + 296, + 396 + ], + "type": "inline_equation", + "content": "A_j" + }, + { + "bbox": [ + 55, + 253, + 296, + 396 + ], + "type": "text", + "content": " can be recovered, completing the PAM: 3Align step. Additionally, we use pose graph optimization to refine the alignment by leveraging loop closure among multiple cooperative agents posing following [17]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 396, + 296, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 396, + 296, + 491 + ], + "spans": [ + { + "bbox": [ + 55, + 396, + 296, + 491 + ], + "type": "text", + "content": "SpatialAlign Module (SAM). To align the cooperative query features " + }, + { + "bbox": [ + 55, + 396, + 296, + 491 + ], + "type": "inline_equation", + "content": "Q^c" + }, + { + "bbox": [ + 55, + 396, + 296, + 491 + ], + "type": "text", + "content": " with the ego query features " + }, + { + "bbox": [ + 55, + 396, + 296, + 491 + ], + "type": "inline_equation", + "content": "Q^e" + }, + { + "bbox": [ + 55, + 396, + 296, + 491 + ], + "type": "text", + "content": ", there are two issues to be resolved. First, the features " + }, + { + "bbox": [ + 55, + 396, + 296, + 491 + ], + "type": "inline_equation", + "content": "F^c" + }, + { + "bbox": [ + 55, + 396, + 296, + 491 + ], + "type": "text", + "content": " are learned in the cooperative coordinate system and must be adapted to the ego coordinate system. We achieve this using a multilayer perceptron (mlp) as described in Eq. (4) where " + }, + { + "bbox": [ + 55, + 396, + 296, + 491 + ], + "type": "inline_equation", + "content": "F^R \\in \\mathbb{R}^{N \\times 9}" + }, + { + "bbox": [ + 55, + 396, + 296, + 491 + ], + "type": "text", + "content": " is the flattened and repeated rotation matrix " + }, + { + "bbox": [ + 55, + 396, + 296, + 491 + ], + "type": "inline_equation", + "content": "R \\in \\mathbb{R}^{3 \\times 3}" + }, + { + "bbox": [ + 55, + 396, + 296, + 491 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 498, + 295, + 514 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 498, + 295, + 514 + ], + "spans": [ + { + "bbox": [ + 131, + 498, + 295, + 514 + ], + "type": "interline_equation", + "content": "F ^ {c} = m l p \\left(\\left[ F ^ {c}; F ^ {R} \\right]\\right) \\tag {4}", + "image_path": "930f7751e42e8d4f1446df6048b150738b5c49cd32de1c0882cb00f08989e08a.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 521, + 296, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 296, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 296, + 628 + ], + "type": "text", + "content": "Second, after rotation with " + }, + { + "bbox": [ + 55, + 521, + 296, + 628 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 55, + 521, + 296, + 628 + ], + "type": "text", + "content": ", the coordinates of " + }, + { + "bbox": [ + 55, + 521, + 296, + 628 + ], + "type": "inline_equation", + "content": "Q^c" + }, + { + "bbox": [ + 55, + 521, + 296, + 628 + ], + "type": "text", + "content": " (Fig. 2, SAM, yellow points) do not perfectly align to the discretized grid in the ego coordinate system. Thus, we merge " + }, + { + "bbox": [ + 55, + 521, + 296, + 628 + ], + "type": "inline_equation", + "content": "Q^c" + }, + { + "bbox": [ + 55, + 521, + 296, + 628 + ], + "type": "text", + "content": " into the nearest points of this grid. The locations of the fused output object query points " + }, + { + "bbox": [ + 55, + 521, + 296, + 628 + ], + "type": "inline_equation", + "content": "Q^{ts}" + }, + { + "bbox": [ + 55, + 521, + 296, + 628 + ], + "type": "text", + "content": " are shown with gray points. For each point in " + }, + { + "bbox": [ + 55, + 521, + 296, + 628 + ], + "type": "inline_equation", + "content": "Q^{ts}" + }, + { + "bbox": [ + 55, + 521, + 296, + 628 + ], + "type": "text", + "content": ", we summarize the features of its " + }, + { + "bbox": [ + 55, + 521, + 296, + 628 + ], + "type": "inline_equation", + "content": "k = 8" + }, + { + "bbox": [ + 55, + 521, + 296, + 628 + ], + "type": "text", + "content": " nearest neighbors (Fig. 2, SAM: black arrows) in query set " + }, + { + "bbox": [ + 55, + 521, + 296, + 628 + ], + "type": "inline_equation", + "content": "Q^{\\cup} = Q^{c}\\bigcup Q^{e}" + }, + { + "bbox": [ + 55, + 521, + 296, + 628 + ], + "type": "text", + "content": " to obtain the fused features " + }, + { + "bbox": [ + 55, + 521, + 296, + 628 + ], + "type": "inline_equation", + "content": "F^{ts}" + }, + { + "bbox": [ + 55, + 521, + 296, + 628 + ], + "type": "text", + "content": ". Mathematically, it reads as" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 71, + 635, + 295, + 651 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 635, + 295, + 651 + ], + "spans": [ + { + "bbox": [ + 71, + 635, + 295, + 651 + ], + "type": "interline_equation", + "content": "F _ {i j} = m l p \\left(\\left[ F _ {j} ^ {\\cup}; \\text {l i n e a r} \\left(\\left[ x _ {i} ^ {t s} - x _ {j} ^ {\\cup}, y _ {i} ^ {t s} - y _ {j} ^ {\\cup} \\right]\\right) \\right]\\right) \\tag {5}", + "image_path": "a8b1cd92dc3398d84729345d6ce1d3d8ef188d468d03037646e9f8125618caa9.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 71, + 652, + 295, + 666 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 652, + 295, + 666 + ], + "spans": [ + { + "bbox": [ + 71, + 652, + 295, + 666 + ], + "type": "interline_equation", + "content": "F _ {i} ^ {t s} = \\left\\{F _ {i j} | j \\in \\{1, \\dots , k \\} \\right\\} \\tag {6}", + "image_path": "894d8ac5c53b967e9a4a9684786bd7eea61405c431a3eac16247cb4f95bc63ea.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 71, + 668, + 295, + 682 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 668, + 295, + 682 + ], + "spans": [ + { + "bbox": [ + 71, + 668, + 295, + 682 + ], + "type": "interline_equation", + "content": "F _ {i} ^ {t s} = \\max _ {j} \\left(F _ {i} ^ {t s}\\right) + \\text {m e a n} _ {j} \\left(F _ {i} ^ {t s}\\right) \\tag {7}", + "image_path": "82ccdc0f5be3374324950bd89058caa55e7124f01ad0b70e89b296627e246abd.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": " is the query index of " + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "inline_equation", + "content": "Q^{ts}" + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": " is the neighbor index of the " + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": "-th query." + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 321, + 72, + 550, + 138 + ], + "blocks": [ + { + "bbox": [ + 321, + 72, + 550, + 138 + ], + "lines": [ + { + "bbox": [ + 321, + 72, + 550, + 138 + ], + "spans": [ + { + "bbox": [ + 321, + 72, + 550, + 138 + ], + "type": "image", + "image_path": "529f1456c1122a5dfe97c8baaff0642621ac6e2ca352e44cfa80719f4f141a2f.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 372, + 151, + 497, + 163 + ], + "lines": [ + { + "bbox": [ + 372, + 151, + 497, + 163 + ], + "spans": [ + { + "bbox": [ + 372, + 151, + 497, + 163 + ], + "type": "text", + "content": "Figure 4. CompassRose encoding" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 314, + 182, + 413, + 194 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 182, + 413, + 194 + ], + "spans": [ + { + "bbox": [ + 314, + 182, + 413, + 194 + ], + "type": "text", + "content": "3.5. Detection Heads" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 201, + 555, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 201, + 555, + 369 + ], + "spans": [ + { + "bbox": [ + 313, + 201, + 555, + 369 + ], + "type": "text", + "content": "To train SparseAlign, we follow [40] to attach a center-based object detection head to each of the RoI, TAM and SAM to learn query features. These heads are notated as RoIDet, LQDet and GQDet, respectively. They share the same target-encoding method. We use Focal loss to classify fore- and background object queries; all query points inside the ground-truth BBoxes are positive, and those outside are negative. Smooth " + }, + { + "bbox": [ + 313, + 201, + 555, + 369 + ], + "type": "inline_equation", + "content": "L1" + }, + { + "bbox": [ + 313, + 201, + 555, + 369 + ], + "type": "text", + "content": " loss is used for regressing " + }, + { + "bbox": [ + 313, + 201, + 555, + 369 + ], + "type": "inline_equation", + "content": "[dx,dy,dz,l,w,h,\\mathcal{B}_{dir},\\mathcal{B}_{scr}]" + }, + { + "bbox": [ + 313, + 201, + 555, + 369 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 201, + 555, + 369 + ], + "type": "inline_equation", + "content": "l,w,h" + }, + { + "bbox": [ + 313, + 201, + 555, + 369 + ], + "type": "text", + "content": " are the dimensions of BBoxes and " + }, + { + "bbox": [ + 313, + 201, + 555, + 369 + ], + "type": "inline_equation", + "content": "dx,dy,dz" + }, + { + "bbox": [ + 313, + 201, + 555, + 369 + ], + "type": "text", + "content": " are coordinate offsets between the query points and the center of ground-truth BBoxes. " + }, + { + "bbox": [ + 313, + 201, + 555, + 369 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_{dir},\\mathcal{B}_{scr}" + }, + { + "bbox": [ + 313, + 201, + 555, + 369 + ], + "type": "text", + "content": " are the novel CompassRose encoding (Fig. 4) of the ground truth orientation " + }, + { + "bbox": [ + 313, + 201, + 555, + 369 + ], + "type": "inline_equation", + "content": "r_g" + }, + { + "bbox": [ + 313, + 201, + 555, + 369 + ], + "type": "text", + "content": " with respect to four anchor angles " + }, + { + "bbox": [ + 313, + 201, + 555, + 369 + ], + "type": "inline_equation", + "content": "\\mathbf{r}_a = [0,0.5\\pi ,\\pi ,1.5\\pi ]" + }, + { + "bbox": [ + 313, + 201, + 555, + 369 + ], + "type": "text", + "content": ". It reads as" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 324, + 377, + 553, + 421 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 377, + 553, + 421 + ], + "spans": [ + { + "bbox": [ + 324, + 377, + 553, + 421 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {B} _ {\\mathrm {d i r}} = \\left[ \\cos r _ {g} - \\cos \\mathbf {r} _ {a}, \\sin r _ {g} - \\sin \\mathbf {r} _ {a} \\right] (8) \\\\ \\mathcal {B} _ {\\mathrm {s c r}} = \\left[ s _ {0}, s _ {0. 5 \\pi}, s _ {\\pi}, s _ {1. 5 \\pi} \\right] (9) \\\\ = \\mathbf {1} - \\arccos \\left(\\cos \\mathbf {r} _ {a} \\cdot \\cos r _ {g} + \\sin \\mathbf {r} _ {a} \\cdot \\sin r _ {g}\\right) / \\pi \\\\ \\end{array}", + "image_path": "c8c0a60b5a0939a3074f60eac0cdd403aeab141f96d00f66c310b71647f11dad.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 430, + 554, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 430, + 554, + 478 + ], + "spans": [ + { + "bbox": [ + 313, + 430, + 554, + 478 + ], + "type": "text", + "content": "Take the encoded sine wave (Fig. 4 left) as an example, the four anchor angles (yellow points) ensure that at least one anchor can reach the target angle monotonically, simplifying the regression process." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 314, + 485, + 427, + 498 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 485, + 427, + 498 + ], + "spans": [ + { + "bbox": [ + 314, + 485, + 427, + 498 + ], + "type": "text", + "content": "3.6. Data augmentation" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 503, + 555, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 503, + 555, + 612 + ], + "spans": [ + { + "bbox": [ + 313, + 503, + 555, + 612 + ], + "type": "text", + "content": "In addition to geometric augmentation, which includes randomly flipping the input point cloud along the x- or y-axis and rotating it around the z-axis, we apply Free Space Augmentation (FSA) [42]. FSA enhances the point cloud by adding auxiliary points along LiDAR scanning rays, representing empty spaces where no objects are present. This augmentation helps mitigate the ICF problem in distant regions, where observation points are extremely sparse. For more details, refer to supplementary B." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 621, + 390, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 621, + 390, + 635 + ], + "spans": [ + { + "bbox": [ + 313, + 621, + 390, + 635 + ], + "type": "text", + "content": "4. Experiment" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 641, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 555, + 715 + ], + "type": "text", + "content": "Datasets. We evaluate SparseAlign on OPV2V[29] and DAIR-V2X[35] datasets for COOD and their variant datasets OPV2Vt[40] and DAIR-V2Xt[40] for TA-COOD. OPV2V is a synthetic dataset created in different driving scenarios of CARLA [7] simulator. Following [29], the detection range is set to " + }, + { + "bbox": [ + 313, + 641, + 555, + 715 + ], + "type": "inline_equation", + "content": "x \\in [-140, 140]m" + }, + { + "bbox": [ + 313, + 641, + 555, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 641, + 555, + 715 + ], + "type": "inline_equation", + "content": "y \\in [-40, 40]m" + }, + { + "bbox": [ + 313, + 641, + 555, + 715 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22300" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 59, + 71, + 293, + 190 + ], + "blocks": [ + { + "bbox": [ + 59, + 71, + 293, + 190 + ], + "lines": [ + { + "bbox": [ + 59, + 71, + 293, + 190 + ], + "spans": [ + { + "bbox": [ + 59, + 71, + 293, + 190 + ], + "type": "table", + "html": "
MethodBW(Mb) ↓AP0.5 ↑AP0.7↑
Fcooper[1]72.080.8870.790
V2VNet[24]72.080.9170.822
FPVRCNN[41]0.240.8730.820
OPV2V[29]126.80.9050.815
CoAlign[17]72.080.9020.833
CoBEVT[19]72.080.9130.861
V2VAM[16]72.080.9160.850
V2VFformer++[34]L+C72.080.9350.895
SparseAlign (ours)< 1.30.9300.892
", + "image_path": "a05d590cc6c4537282991804a69b7a64246ea5a736e4b75d202b9c93e4c82455.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 59, + 231, + 293, + 323 + ], + "blocks": [ + { + "bbox": [ + 75, + 198, + 277, + 210 + ], + "lines": [ + { + "bbox": [ + 75, + 198, + 277, + 210 + ], + "spans": [ + { + "bbox": [ + 75, + 198, + 277, + 210 + ], + "type": "text", + "content": "Table 1. COOD Average Precision on OPV2V dataset. " + }, + { + "bbox": [ + 75, + 198, + 277, + 210 + ], + "type": "inline_equation", + "content": "{}^{2}" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 59, + 231, + 293, + 323 + ], + "lines": [ + { + "bbox": [ + 59, + 231, + 293, + 323 + ], + "spans": [ + { + "bbox": [ + 59, + 231, + 293, + 323 + ], + "type": "table", + "html": "
MethodOPV2VtDAIR-V2Xt
AP0.5↑AP0.7↑AP0.5↑AP0.7↑
Fcooper[1]+SA0.7630.5530.5970.282
FPVRCNN[43]+SA0.6400.4740.5980.307
OPV2V[29]+SA0.8810.7870.7020.366
StreamLTS[40]0.8530.7210.6400.404
StreamLTS[40]+SA0.8500.7900.7570.497
SparseAlign (full)0.8930.8180.7960.548
", + "image_path": "a567cf47c12a2f46c6f75d2cfe1872891d43eab3a1eadeaf0fb7c9d6dae9eab7.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 378, + 295, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 378, + 295, + 498 + ], + "spans": [ + { + "bbox": [ + 54, + 378, + 295, + 498 + ], + "type": "text", + "content": "DAIR-V2X is a real dataset containing one vehicle- and one infrastructure-side sensor suit. The detection range is set to " + }, + { + "bbox": [ + 54, + 378, + 295, + 498 + ], + "type": "inline_equation", + "content": "x \\in [-100, 100]m, y \\in [-40, 40]m" + }, + { + "bbox": [ + 54, + 378, + 295, + 498 + ], + "type": "text", + "content": ". OPV2Vt is generated by interpolating OPV2V data frames into sub-frames and replaying them to mimic the sensor asynchrony and the rolling shutter problem of LiDARs. The GT BBoxes of this dataset are aligned to the global timestamp, the LiDAR scan ending time of each frame. DAIR-V2Xt follows the same idea. The detection range of OPV2Vt and DAIR-V2Xt is set the same as OPV2V and DAIR-V2X." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 499, + 295, + 680 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 499, + 295, + 680 + ], + "spans": [ + { + "bbox": [ + 55, + 499, + 295, + 680 + ], + "type": "text", + "content": "Implementation details. All models are trained on a single GTX-4090 and an Intel i7-8700 CPU for 50 epochs (about 35 hours) with the Adam optimizer and batch size two. The learning rate is set to " + }, + { + "bbox": [ + 55, + 499, + 295, + 680 + ], + "type": "inline_equation", + "content": "2e - 4" + }, + { + "bbox": [ + 55, + 499, + 295, + 680 + ], + "type": "text", + "content": " with a warm-up stage of 4000 iterations. The random rotation angle is set to the range " + }, + { + "bbox": [ + 55, + 499, + 295, + 680 + ], + "type": "inline_equation", + "content": "[-90^{\\circ}, 90^{\\circ}]" + }, + { + "bbox": [ + 55, + 499, + 295, + 680 + ], + "type": "text", + "content": ". For models incorporating the TAM, we load four sequential frames of data during training and calculate losses solely for the last frame. To reduce GPU memory consumption, we enable gradient calculations only for data from a single ego agent and one cooperative agent. For details on efficient gradient scheduling, refer to supplementary " + }, + { + "bbox": [ + 55, + 499, + 295, + 680 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 55, + 499, + 295, + 680 + ], + "type": "text", + "content": ". To ensure that the model generalizes well to the significant time latency introduced by communication, we load data from cooperative vehicles with a random latency ranging from 0 to " + }, + { + "bbox": [ + 55, + 499, + 295, + 680 + ], + "type": "inline_equation", + "content": "200~\\mathrm{ms}" + }, + { + "bbox": [ + 55, + 499, + 295, + 680 + ], + "type": "text", + "content": " relative to the ego vehicles." + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 332, + 71, + 552, + 191 + ], + "blocks": [ + { + "bbox": [ + 55, + 331, + 295, + 353 + ], + "lines": [ + { + "bbox": [ + 55, + 331, + 295, + 353 + ], + "spans": [ + { + "bbox": [ + 55, + 331, + 295, + 353 + ], + "type": "text", + "content": "Table 3. TA-COOD Average Precision on OPV2Vt and DAIR-V2Xt" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 332, + 71, + 552, + 191 + ], + "lines": [ + { + "bbox": [ + 332, + 71, + 552, + 191 + ], + "spans": [ + { + "bbox": [ + 332, + 71, + 552, + 191 + ], + "type": "table", + "html": "
MethodBW(Mb) ↓AP0.5 ↑AP0.7↑
Fcooper[1]48.80.7340.559
V2VNet[24]48.80.6640.402
FPVRCNN[41]0.240.6650.505
V2XViT[28]48.80.7040.531
DiscoNet[19]48.80.7360.583
OPV2V[29]97.60.7330.553
CoAlign[17]48.80.7460.604
DI-V2X[27]48.80.7880.662
SparseAlign (ours)< 1.30.8450.685
", + "image_path": "49227effd71dca5d664a5461b91ba5dcf1dec2ad564d1d287bd7d44fa5a3d9e9.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 334, + 200, + 547, + 211 + ], + "lines": [ + { + "bbox": [ + 334, + 200, + 547, + 211 + ], + "spans": [ + { + "bbox": [ + 334, + 200, + 547, + 211 + ], + "type": "text", + "content": "Table 2. COOD Average Precision on DAIR-V2X dataset." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 232, + 368, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 232, + 368, + 243 + ], + "spans": [ + { + "bbox": [ + 313, + 232, + 368, + 243 + ], + "type": "text", + "content": "5. Results" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 251, + 497, + 264 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 251, + 497, + 264 + ], + "spans": [ + { + "bbox": [ + 313, + 251, + 497, + 264 + ], + "type": "text", + "content": "5.1. Comparison to the state-of-the-art" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 269, + 553, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 269, + 553, + 472 + ], + "spans": [ + { + "bbox": [ + 313, + 269, + 553, + 472 + ], + "type": "text", + "content": "Tab. 1 presents the COOD results on the OPV2V dataset. All methods use only LiDAR data, except for V2VFormer++ (denoted by " + }, + { + "bbox": [ + 313, + 269, + 553, + 472 + ], + "type": "inline_equation", + "content": "\\mathrm{L + C}" + }, + { + "bbox": [ + 313, + 269, + 553, + 472 + ], + "type": "text", + "content": "), which incorporates both LiDAR and camera data. Compared to the state-of-the-art, our SparseAlign achieves the best performance in LiDAR-based COOD, approaching the results of V2VFormer++, which benefits from LiDAR-Camera fusion. Tab. 2 shows the results on the DAIR-V2X dataset. Here, SparseAlign outperforms other methods by a large margin, achieving a " + }, + { + "bbox": [ + 313, + 269, + 553, + 472 + ], + "type": "inline_equation", + "content": "2.3\\%" + }, + { + "bbox": [ + 313, + 269, + 553, + 472 + ], + "type": "text", + "content": " improvement in AP at the IoU threshold of 0.7 and a " + }, + { + "bbox": [ + 313, + 269, + 553, + 472 + ], + "type": "inline_equation", + "content": "5.7\\%" + }, + { + "bbox": [ + 313, + 269, + 553, + 472 + ], + "type": "text", + "content": " improvement at the IoU threshold of 0.5. This is mainly due to the enhanced temporal and global reasoning ability of our model for detecting distant and large objects. Compared to other methods that rely on dense BEV feature maps, SparseAlign consumes the least communication bandwidth (BW) of less than 1.3MB (without compression) thanks to the efficient query-based operations." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 473, + 554, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 473, + 554, + 579 + ], + "spans": [ + { + "bbox": [ + 313, + 473, + 554, + 579 + ], + "type": "text", + "content": "For the results of the more challenging TA-COOD task shown in Tab. 3, our SparseAlign outperforms StreamLTS by a great margin. By applying our three alignment modules (StreamLTS + SA), the performance of StreamLTS is also improved. In addition, we compare our framework to Fcooper, FPVRCNN and OPV2V method by attaching the three alignment modules to them to achieve TA-COOD. All show worse performance than the full SparseAlign because of worse 3D backbone or the RoI selection." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 580, + 554, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 580, + 554, + 617 + ], + "spans": [ + { + "bbox": [ + 313, + 580, + 554, + 617 + ], + "type": "text", + "content": "The qualitative results and the additional comparison to methods, such as TransIFF[4], on the 3D AP metric are given in the supplementary A and E, respectively." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 623, + 406, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 623, + 406, + 635 + ], + "spans": [ + { + "bbox": [ + 313, + 623, + 406, + 635 + ], + "type": "text", + "content": "5.2. Ablation study" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 641, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 554, + 713 + ], + "type": "text", + "content": "We conduct an ablation study on the 3D backbone, TAM, and SAM (PAM is discussed separately in Sec. 5.4). Among these components, the spatial fusion module plays a crucial role in multi-agent feature fusion. Without SAM, the module performs the worst, as evidenced by the first row of Tab. 4. When using QUEST [9] spatial fusion, which" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "type": "text", + "content": "2The results reported in this table use frame-wise local score sorting for AP, same to the official benchmark. Other results use global sorting." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "text", + "content": "22301" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 59, + 71, + 329, + 171 + ], + "blocks": [ + { + "bbox": [ + 59, + 71, + 329, + 171 + ], + "lines": [ + { + "bbox": [ + 59, + 71, + 329, + 171 + ], + "spans": [ + { + "bbox": [ + 59, + 71, + 329, + 171 + ], + "type": "table", + "html": "
CEC 2DCEC 3DTAMSAMOPV2VDAIR-V2X
AP0.5AP0.7AP0.5AP0.7
✓✓✓0.8470.7560.6710.528
✓✓✓QST0.8880.8690.7490.622
✓0.9240.8850.7730.638
✓✓0.9400.9010.7960.662
✓✓✓0.9500.9140.8130.682
✓✓✓✓0.9510.9290.8450.685
", + "image_path": "67b06a0e389cf8f639a5cf911871ca5a9c4bc2ba7f442b83e9f39736a9da98a1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 355, + 71, + 553, + 149 + ], + "blocks": [ + { + "bbox": [ + 55, + 178, + 332, + 201 + ], + "lines": [ + { + "bbox": [ + 55, + 178, + 332, + 201 + ], + "spans": [ + { + "bbox": [ + 55, + 178, + 332, + 201 + ], + "type": "text", + "content": "Table 4. Ablation study on 2D and 3D CEC layers, TAM and SAM. QST: use QUEST [9] spatial fusion instead of SAM." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 355, + 71, + 553, + 149 + ], + "lines": [ + { + "bbox": [ + 355, + 71, + 553, + 149 + ], + "spans": [ + { + "bbox": [ + 355, + 71, + 553, + 149 + ], + "type": "table", + "html": "
Dir.Enc.OPV2VDAIR-V2X
AP0.5AP0.7AP0.5AP0.7
gt-angle0.8710.7680.6200.427
second0.9080.8740.7430.590
sin_cos0.9150.8810.7480.603
compass0.9240.8850.7730.638
", + "image_path": "5d2885a02827632217ff15559198d9dea5cbc262d3a4f7e9248006670505a394.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 71, + 213, + 541, + 322 + ], + "blocks": [ + { + "bbox": [ + 351, + 157, + 555, + 203 + ], + "lines": [ + { + "bbox": [ + 351, + 157, + 555, + 203 + ], + "spans": [ + { + "bbox": [ + 351, + 157, + 555, + 203 + ], + "type": "text", + "content": "Table 5. Comparison on BBox angle encoding methods: offset to ground-truth angle (gt-angle), second-style (second), sine-cosine offsets (sin_cos) and compass rose (compass)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 71, + 213, + 541, + 322 + ], + "lines": [ + { + "bbox": [ + 71, + 213, + 541, + 322 + ], + "spans": [ + { + "bbox": [ + 71, + 213, + 541, + 322 + ], + "type": "table", + "html": "
DatasetOPV2VtDAIR-V2Xt
Latency0ms100ms200ms0ms100ms200ms
AP threshold0.50.70.50.70.50.70.50.70.50.70.50.7
StreamLTS [40]0.8530.7210.8160.6800.7870.6470.6420.4040.6130.3790.5900.364
SA no TAM0.8900.8020.8560.4710.6310.2090.7200.4570.6720.3930.6350.350
SA no PT0.8350.6150.8310.6250.8150.5810.7600.4690.7370.4590.7220.449
SA no TL0.8850.8430.8620.6190.7230.2790.7930.5340.7490.4740.6980.423
SA no FSA0.8810.8160.8750.7950.8570.7620.7690.5080.7550.5020.7400.496
SparseAlign (SA)0.8930.8180.8850.7950.8670.7650.7960.5480.7860.5430.7720.532
", + "image_path": "4fbefe9fd2dd11c2d0e99147fe29e9d46620ab6e4b49d9848b1a1e497bc7b792.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 331, + 553, + 342 + ], + "lines": [ + { + "bbox": [ + 56, + 331, + 553, + 342 + ], + "spans": [ + { + "bbox": [ + 56, + 331, + 553, + 342 + ], + "type": "text", + "content": "Table 6. AP of TA-COOD with communication latency. PT: Point-wise Timestamps, TL: Train Latency, FSA: Free Space Augmentation." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 364, + 295, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 364, + 295, + 447 + ], + "spans": [ + { + "bbox": [ + 54, + 364, + 295, + 447 + ], + "type": "text", + "content": "aligns only rotation (Eq. (4)) without correcting spatial displacement (Eqs. (5) to (7)), the AP scores remain significantly lower than those obtained with SAM. Incorporating SAM improves performance, and we further enhance it by gradually replacing standard sparse convolutional layers with CEC layers. Finally, TAM further boosts AP scores by leveraging temporal information from historical frames." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 448, + 295, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 448, + 295, + 520 + ], + "spans": [ + { + "bbox": [ + 55, + 448, + 295, + 520 + ], + "type": "text", + "content": "In Tab. 5, we compare CompassRose encoding with three methods: directly regress the ground-truth angle (gt-angle), second-style [31] encoding (second), and the sine-cosine encoding (sin_cos). For simplicity, we use the configurations in the third row of Tab. 4 for the comparison. The results show that CompassRose performs the best." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 527, + 295, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 527, + 295, + 540 + ], + "spans": [ + { + "bbox": [ + 55, + 527, + 295, + 540 + ], + "type": "text", + "content": "5.3. Communication latency and sensor asynchrony" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "text", + "content": "Tab. 6 shows the results of TA-COOD on the OPV2Vt and the DAIR-V2Xt dataset. Our proposed SparseAlign demonstrates a significant improvement over the baseline framework, StreamLTS [40]. Specifically, the AP at an IoU threshold of 0.7 increased " + }, + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "inline_equation", + "content": "9.7\\%" + }, + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "text", + "content": " on the OPV2Vt dataset and " + }, + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "inline_equation", + "content": "14.4\\%" + }, + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "text", + "content": " on the DAIR-V2Xt dataset. As the communication latency increases to 200ms, the AP at IoU 0.5 of our SparseAlign only dropped " + }, + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "inline_equation", + "content": "2.6\\%" + }, + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "text", + "content": " while the baseline work has lost " + }, + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "inline_equation", + "content": "6.6\\%" + }, + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "text", + "content": " of its accuracy. This reveals that our framework is more robust against communication latency. To investigate modules or configurations that are making an effect on this robustness, we conducted an ablation study by removing the TAM, Point-wise Timestamps (TP), Training Latency (TL, communication latency during training) and" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 364, + 503, + 375 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 364, + 503, + 375 + ], + "spans": [ + { + "bbox": [ + 313, + 364, + 503, + 375 + ], + "type": "text", + "content": "Free Space Augmentation (FSA), respectively." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 312, + 376, + 555, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 376, + 555, + 507 + ], + "spans": [ + { + "bbox": [ + 312, + 376, + 555, + 507 + ], + "type": "text", + "content": "Without the TAM, the performance noticeably declines as latency increases, despite achieving comparable results to the full SpareAlign model at " + }, + { + "bbox": [ + 312, + 376, + 555, + 507 + ], + "type": "inline_equation", + "content": "0\\mathrm{ms}" + }, + { + "bbox": [ + 312, + 376, + 555, + 507 + ], + "type": "text", + "content": " latency. A similar effect is observed when the full SA model is trained without exposure to data with communication latency (SA without TL). Remarkably, without TL, the model outperforms the full SpareAlign model in terms of AP at an IoU threshold of 0.7 on the OPV2Vt dataset (0.843 against 0.818). This is because the model focuses more on predicting BBoxes within the same frame, where only minimal latency is introduced by asynchronous sensors." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 312, + 507, + 556, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 507, + 556, + 616 + ], + "spans": [ + { + "bbox": [ + 312, + 507, + 556, + 616 + ], + "type": "text", + "content": "Without PT, the model performs much worse than the full SpareAlign, showing the importance of fine-grained point-wise timestamps in the point clouds for the model to compensate for the errors introduced by sensor asynchrony and capture the accurate temporal context for prediction. Finally, the ablation study on the FSA module also shows a positive influence on the performance. For example, AP at the IoU of 0.7 increased by " + }, + { + "bbox": [ + 312, + 507, + 556, + 616 + ], + "type": "inline_equation", + "content": "4\\%" + }, + { + "bbox": [ + 312, + 507, + 556, + 616 + ], + "type": "text", + "content": " on DAIR-V2Xt by utilizing the FSA (See supplementary D for visual comparison)." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 624, + 427, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 624, + 427, + 635 + ], + "spans": [ + { + "bbox": [ + 313, + 624, + 427, + 635 + ], + "type": "text", + "content": "5.4. Localization errors" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "text", + "content": "Fig. 5 shows the performance of SpareAlign against different localization errors. We add random errors to the poses of both ego " + }, + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "inline_equation", + "content": "(T_e)" + }, + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "text", + "content": " and cooperative " + }, + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "inline_equation", + "content": "(T_c)" + }, + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "text", + "content": " vehicles. The translation errors " + }, + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "inline_equation", + "content": "x_{\\epsilon}, y_{\\epsilon}" + }, + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "text", + "content": " along the " + }, + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "text", + "content": "- and " + }, + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "text", + "content": "-axis and the rotation error " + }, + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "inline_equation", + "content": "r_{\\epsilon}" + }, + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "text", + "content": " around the " + }, + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "text", + "content": "-axis are all assumed to be normally distributed with " + }, + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(0,1) \\cdot \\epsilon" + }, + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\epsilon \\in [0,1]" + }, + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "text", + "content": " is the error scaling factor." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22302" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 72, + 73, + 279, + 177 + ], + "blocks": [ + { + "bbox": [ + 72, + 73, + 279, + 177 + ], + "lines": [ + { + "bbox": [ + 72, + 73, + 279, + 177 + ], + "spans": [ + { + "bbox": [ + 72, + 73, + 279, + 177 + ], + "type": "image", + "image_path": "de9edb1b1ab958c861bf4550b46e38a36f72222773356e1ae489e155e178d8c9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 70, + 178, + 279, + 281 + ], + "blocks": [ + { + "bbox": [ + 70, + 178, + 279, + 281 + ], + "lines": [ + { + "bbox": [ + 70, + 178, + 279, + 281 + ], + "spans": [ + { + "bbox": [ + 70, + 178, + 279, + 281 + ], + "type": "image", + "image_path": "d963ed326dce265d53ea2cafabe9165d8f70ada8d9bc5100897043381cf87a89.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 293, + 295, + 327 + ], + "lines": [ + { + "bbox": [ + 55, + 293, + 295, + 327 + ], + "spans": [ + { + "bbox": [ + 55, + 293, + 295, + 327 + ], + "type": "text", + "content": "Figure 5. AP at IoU threshold of 0.7 with translation errors ranging from " + }, + { + "bbox": [ + 55, + 293, + 295, + 327 + ], + "type": "inline_equation", + "content": "0\\mathrm{m}" + }, + { + "bbox": [ + 55, + 293, + 295, + 327 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 293, + 295, + 327 + ], + "type": "inline_equation", + "content": "1\\mathrm{m}" + }, + { + "bbox": [ + 55, + 293, + 295, + 327 + ], + "type": "text", + "content": " along x- and y-axis, and rotation errors from " + }, + { + "bbox": [ + 55, + 293, + 295, + 327 + ], + "type": "inline_equation", + "content": "0^{\\circ}" + }, + { + "bbox": [ + 55, + 293, + 295, + 327 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 293, + 295, + 327 + ], + "type": "inline_equation", + "content": "1.0^{\\circ}" + }, + { + "bbox": [ + 55, + 293, + 295, + 327 + ], + "type": "text", + "content": " (horizontal axis) for the different datasets." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 349, + 296, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 349, + 296, + 443 + ], + "spans": [ + { + "bbox": [ + 55, + 349, + 296, + 443 + ], + "type": "text", + "content": "We report the results by gradually increasing " + }, + { + "bbox": [ + 55, + 349, + 296, + 443 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 55, + 349, + 296, + 443 + ], + "type": "text", + "content": " with a step size of 0.2. Note that the translation and rotation errors exist in the poses of both the ego and the cooperative vehicle; the resulting relative translation error of the transformation " + }, + { + "bbox": [ + 55, + 349, + 296, + 443 + ], + "type": "inline_equation", + "content": "T_{c}^{e} = inverse(T_{e})\\cdot T_{c}" + }, + { + "bbox": [ + 55, + 349, + 296, + 443 + ], + "type": "text", + "content": " from cooperative to ego coordinate system will be amplified as the rotation error increases. Our PAM is specially designed to mitigate the influence of large relative transformation errors." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 445, + 296, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 445, + 296, + 649 + ], + "spans": [ + { + "bbox": [ + 55, + 445, + 296, + 649 + ], + "type": "text", + "content": "Our proposed PAM and the baseline method CoAlign [17] both significantly reduce the impact of pose errors on AP performance compared to the configuration lacking any correction module. In addition, our method demonstrates greater robustness against large errors. For instance, on the OPV2V dataset, the AP for CoAlign decreased by approximately " + }, + { + "bbox": [ + 55, + 445, + 296, + 649 + ], + "type": "inline_equation", + "content": "27\\%" + }, + { + "bbox": [ + 55, + 445, + 296, + 649 + ], + "type": "text", + "content": " at the error of " + }, + { + "bbox": [ + 55, + 445, + 296, + 649 + ], + "type": "inline_equation", + "content": "1.0m, 1.0m, 1.0^{\\circ}" + }, + { + "bbox": [ + 55, + 445, + 296, + 649 + ], + "type": "text", + "content": ", whereas PAM experienced only a " + }, + { + "bbox": [ + 55, + 445, + 296, + 649 + ], + "type": "inline_equation", + "content": "12\\%" + }, + { + "bbox": [ + 55, + 445, + 296, + 649 + ], + "type": "text", + "content": " drop. This validates the efficacy of our design, which relies on the pose-agnostic relative neighborhood geometries to match the detected BBoxes of the ego and the cooperative IAs. Note that the DAIR-V2X dataset exhibits smaller changes in AP as localization errors increase, because the pose parameters are not well calibrated. In contrast, [40] refined the poses of DAIR-V2X for generating the DAIR-V2Xt dataset, making the better-calibrated DAIR-V2Xt more sensitive to newly introduced errors." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 658, + 128, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 658, + 128, + 670 + ], + "spans": [ + { + "bbox": [ + 55, + 658, + 128, + 670 + ], + "type": "text", + "content": "5.5. CPM sizes" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "content": "Our framework significantly reduces communication bandwidth without relying on computationally intensive dense feature maps, unlike state-of-the-art methods. This effi" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 331, + 74, + 434, + 175 + ], + "blocks": [ + { + "bbox": [ + 331, + 74, + 434, + 175 + ], + "lines": [ + { + "bbox": [ + 331, + 74, + 434, + 175 + ], + "spans": [ + { + "bbox": [ + 331, + 74, + 434, + 175 + ], + "type": "image", + "image_path": "c77fbab4063aa88260bffeeeae859f74e50be0d25aeac07ea4597d41fcfb6a98.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 186, + 553, + 220 + ], + "lines": [ + { + "bbox": [ + 313, + 186, + 553, + 220 + ], + "spans": [ + { + "bbox": [ + 313, + 186, + 553, + 220 + ], + "type": "text", + "content": "Figure 6. CPM sizes with different information selection scores. Object query features with detection scores larger than the threshold (x-axis) are shared as CPMs." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 435, + 74, + 539, + 175 + ], + "blocks": [ + { + "bbox": [ + 435, + 74, + 539, + 175 + ], + "lines": [ + { + "bbox": [ + 435, + 74, + 539, + 175 + ], + "spans": [ + { + "bbox": [ + 435, + 74, + 539, + 175 + ], + "type": "image", + "image_path": "2df0624091bbbf8392d6db71bcd8377fe6a2e7bf9823f149723911361688707e.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 245, + 555, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 245, + 555, + 437 + ], + "spans": [ + { + "bbox": [ + 313, + 245, + 555, + 437 + ], + "type": "text", + "content": "ciency is achieved by selecting only the top " + }, + { + "bbox": [ + 313, + 245, + 555, + 437 + ], + "type": "inline_equation", + "content": "K = 1024" + }, + { + "bbox": [ + 313, + 245, + 555, + 437 + ], + "type": "text", + "content": " object queries for further processing. In addition to this intrinsic reduction, the size of the CPMs can be further minimized through score-based selection. Specifically, we employ the query-based detection scores from the LQDet head to identify and share only the most critical queries. As Fig. 6 shows, using a CPM score threshold below 0.5 does not lead to a noticeable degradation in performance across all datasets. Remarkably, at a threshold score of 0.5, the average size of the CPMs is reduced to less than " + }, + { + "bbox": [ + 313, + 245, + 555, + 437 + ], + "type": "inline_equation", + "content": "400\\mathrm{KB}" + }, + { + "bbox": [ + 313, + 245, + 555, + 437 + ], + "type": "text", + "content": " on all datasets. The final CPM sizes after selection are influenced by the driving scenes. Specifically, scenes with more vehicles tend to have larger CPM sizes. Consequently, the OPV2V and OPV2Vt datasets, which contain more vehicles in their scenes, result in larger CPM sizes compared to the DAIR-V2X and DAIR-V2Xt datasets." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 453, + 388, + 464 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 453, + 388, + 464 + ], + "spans": [ + { + "bbox": [ + 314, + 453, + 388, + 464 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 474, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 474, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 474, + 555, + 713 + ], + "type": "text", + "content": "In this paper, we proposed SparseAlign, a fully sparse framework for cooperative object detection (COOD). With our novel 3D backbone network SUNet, the Compass-Rose encoding of detection head, the temporal and spatial fusion module TAM and SAM, SparseAlign significantly outperforms state-of-the-art methods on the OPV2V and DAIR-V2X datasets for COOD, as well as on their variants OPV2Vt and DAIR-V2Xt for time-aligned COOD (TA-COOD), achieving this with very low communication bandwidth. Additionally, our proposed pose alignment module PAM demonstrates improved robustness to large localization errors compared to the baseline method CoAlign. Nevertheless, when the pose errors are large, a drop of about " + }, + { + "bbox": [ + 313, + 474, + 555, + 713 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 313, + 474, + 555, + 713 + ], + "type": "text", + "content": " in APs persists even with pose correction. In future work, we therefore plan to explore more robust methods, e.g. leveraging sub-graph detection, to further enhance the pose alignment accuracy. Additionally, more application scenarios, such as roadside cooperative detection benchmarks HoloVIC[18], RCooper[10] and TUMTraf[45], could also be explored." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22303" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "type": "text", + "content": "[1] Qi Chen. F-cooper: feature based cooperative perception for autonomous vehicle edge computing system using 3d point clouds. Proceedings of the 4th ACM/IEEE Symposium on Edge Computing, 2019. 1, 2, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 136, + 296, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 136, + 296, + 191 + ], + "spans": [ + { + "bbox": [ + 61, + 136, + 296, + 191 + ], + "type": "text", + "content": "[2] Qi Chen, Sihai Tang, Qing Yang, and Song Fu. Cooper: Cooperative perception for connected autonomous vehicles based on 3d point clouds. 2019 IEEE 39th International Conference on Distributed Computing Systems (ICDCS), pages 514-524, 2019. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 191, + 295, + 245 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 191, + 295, + 245 + ], + "spans": [ + { + "bbox": [ + 62, + 191, + 295, + 245 + ], + "type": "text", + "content": "[3] Yukang Chen, Jianhui Liu, Xiangyu Zhang, Xiaojuan Qi, and Jiaya Jia. Voxelnext: Fully sparse voxelnet for 3d object detection and tracking. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 21674-21683, 2023. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 246, + 295, + 301 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 246, + 295, + 301 + ], + "spans": [ + { + "bbox": [ + 62, + 246, + 295, + 301 + ], + "type": "text", + "content": "[4] Ziming Chen, Yifeng Shi, and Jinrang Jia. TransIFF: An instance-level feature fusion framework for vehicleinfrastructure cooperative 3d detection with transformers. In 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 18159-18168, 2023. 2, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 302, + 295, + 356 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 302, + 295, + 356 + ], + "spans": [ + { + "bbox": [ + 62, + 302, + 295, + 356 + ], + "type": "text", + "content": "[5] Christopher Bongsoo Choy, JunYoung Gwak, and Silvio Savarese. 4d spatio-temporal convnets: Minkowski convolutional neural networks. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3070-3079, 2019. 2, 3, 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 357, + 295, + 412 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 357, + 295, + 412 + ], + "spans": [ + { + "bbox": [ + 62, + 357, + 295, + 412 + ], + "type": "text", + "content": "[6] Jiaxun Cui, Hang Qiu, Dian Chen, Peter Stone, and Yuke Zhu. Coopernaut: End-to-end driving with cooperative perception for networked vehicles. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), page 17252-17262, 2022. 1, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 413, + 295, + 457 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 413, + 295, + 457 + ], + "spans": [ + { + "bbox": [ + 62, + 413, + 295, + 457 + ], + "type": "text", + "content": "[7] Alexey Dosovitskiy, German Ros, Felipe Codevilla, Antonio Lopez, and Vladlen Koltun. CARLA: An open urban driving simulator. In Proceedings of the 1st Annual Conference on Robot Learning, pages 1-16, 2017. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 458, + 294, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 458, + 294, + 479 + ], + "spans": [ + { + "bbox": [ + 62, + 458, + 294, + 479 + ], + "type": "text", + "content": "[8] Lue Fan, Feng Wang, Naiyan Wang, and Zhaoxiang Zhang. Fully Sparse 3D Object Detection. In NeurIPS, 2022. 1, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 480, + 295, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 480, + 295, + 524 + ], + "spans": [ + { + "bbox": [ + 62, + 480, + 295, + 524 + ], + "type": "text", + "content": "[9] Siqi Fan, Haibao Yu, Wenxian Yang, Jirui Yuan, and Zaiqing Nie. Quest: Query stream for practical cooperative perception. In IEEE International Conference on Robotics and Automation (ICRA), 2024. 2, 6, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 525, + 296, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 525, + 296, + 590 + ], + "spans": [ + { + "bbox": [ + 57, + 525, + 296, + 590 + ], + "type": "text", + "content": "[10] Ruiyang Hao, Siqi Fan, Yingru Dai, Zhenlin Zhang, Chenxi Li, Yuntian Wang, Haibao Yu, Wenxian Yang, Yuan Jirui, and Zaiqing Nie. Rcooper: A real-world large-scale dataset for roadside cooperative perception. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 22347-22357, 2024. 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 591, + 295, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 591, + 295, + 635 + ], + "spans": [ + { + "bbox": [ + 57, + 591, + 295, + 635 + ], + "type": "text", + "content": "[11] Caizhen He, Hai Wang, Long Chen, Tong Luo, and Yingfeng Cai. V2x-ahd:vehicle-to-everything cooperation perception via asymmetric heterogenous distillation network. arXiv, arXiv:2310.06603, 2023. arXiv:2310.06603 [cs]. 1, 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 636, + 295, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 636, + 295, + 658 + ], + "spans": [ + { + "bbox": [ + 57, + 636, + 295, + 658 + ], + "type": "text", + "content": "[12] Sepp Hochreiter and Jürgen Schmidhuber. Long short-term memory. Neural Computation, 9(8):1735-1780, 1997. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 658, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 658, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 658, + 295, + 713 + ], + "type": "text", + "content": "[13] Yue Hu, Shaoheng Fang, Zixing Lei, Yiqi Zhong, and Siheng Chen. Where2comm: Communication-efficient collaborative perception via spatial confidence maps. In Thirty-sixth Conference on Neural Information Processing Systems (Neurips), 2022. 2" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 554, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 316, + 73, + 554, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 128 + ], + "type": "text", + "content": "[14] Alex H. Lang, Sourabh Vora, Holger Caesar, Lubing Zhou, Jiong Yang, and Oscar Beijbom. Pointpillars: Fast encoders for object detection from point clouds. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12689-12697, 2018. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 129, + 554, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 554, + 163 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 554, + 163 + ], + "type": "text", + "content": "[15] Zixing Lei, Shunli Ren, Yue Hu, Wenjun Zhang, and Siheng Chen. Latency-aware collaborative perception. In Computer Vision – ECCV 2022, page 316–332, 2022. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 164, + 554, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 164, + 554, + 217 + ], + "spans": [ + { + "bbox": [ + 317, + 164, + 554, + 217 + ], + "type": "text", + "content": "[16] Jinlong Li, Runsheng Xu, Xinyu Liu, Jin Ma, Zicheng Chi, Jiaqi Ma, and Hongkai Yu. Learning for vehicle-to-vehicle cooperative perception under lossy communication. IEEE Transactions on Intelligent Vehicles, 8(4):2650-2660, 2023. 1, 3, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 220, + 554, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 220, + 554, + 275 + ], + "spans": [ + { + "bbox": [ + 316, + 220, + 554, + 275 + ], + "type": "text", + "content": "[17] Yifan Lu, Quanhao Li, Baoan Liu, Mehrdad Dianati, Chen Feng, Siheng Chen, and Yanfeng Wang. Robust collaborative 3d object detection in presence of pose errors. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 4812-4818. IEEE, 2023. 2, 3, 5, 6, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 276, + 554, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 276, + 554, + 342 + ], + "spans": [ + { + "bbox": [ + 316, + 276, + 554, + 342 + ], + "type": "text", + "content": "[18] Cong Ma, Lei Qiao, Chengkai Zhu, Kai Liu, Zelong Kong, Qing Li, Xueqi Zhou, Yuheng Kan, and Wei Wu. Holovic:large-scale dataset and benchmark for multi-sensor holographic intersection and vehicle-infrastructure cooperative. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 22129-22138, 2024. 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 343, + 554, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 343, + 554, + 388 + ], + "spans": [ + { + "bbox": [ + 316, + 343, + 554, + 388 + ], + "type": "text", + "content": "[19] Xu Runsheng, Zhengzhong Tu Tu, Hao Xiang, Wei Shao, Bolei Zhou, and Jiaqi Ma. Cobevt: Cooperative bird's eye view semantic segmentation with sparse transformers. In Conference on Robot Learning (CoRL), 2022. 1, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 388, + 554, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 388, + 554, + 443 + ], + "spans": [ + { + "bbox": [ + 316, + 388, + 554, + 443 + ], + "type": "text", + "content": "[20] Shaoshuai Shi, Chaoxu Guo, Li Jiang, Zhe Wang, Jianping Shi, Xiaogang Wang, and Hongsheng Li. Pv-rcnn: Point-voxel feature set abstraction for 3d object detection. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10526–10535, 2019. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 444, + 554, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 444, + 554, + 499 + ], + "spans": [ + { + "bbox": [ + 317, + 444, + 554, + 499 + ], + "type": "text", + "content": "[21] Shaoshuai Shi, Chaoxu Guo, Li Jiang, Zhe Wang, Jianping Shi, Xiaogang Wang, and Hongsheng Li. PV-RCNN: Point-voxel feature set abstraction for 3d object detection. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10526-10535, 2019. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 500, + 554, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 500, + 554, + 567 + ], + "spans": [ + { + "bbox": [ + 317, + 500, + 554, + 567 + ], + "type": "text", + "content": "[22] Rui Song, Chenwei Liang, Hu Cao, Zhiran Yan, Walter Zimmer, Markus Gross, Andreas Festag, and Alois Knoll. Collaborative semantic occupancy prediction with hybrid feature fusion in connected automated vehicles. In 2024 IEEE/CVF International Conference on Computer Vision and Pattern Recognition (CVPR). IEEE/CVF, 2024. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 568, + 554, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 568, + 554, + 622 + ], + "spans": [ + { + "bbox": [ + 317, + 568, + 554, + 622 + ], + "type": "text", + "content": "[23] Shihao Wang, Ying-Hao Liu, Tiancai Wang, Ying Li, and Xiangyu Zhang. Exploring object-centric temporal modeling for efficient multi-view 3d object detection. 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 3598-3608, 2023. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 624, + 554, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 554, + 679 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 554, + 679 + ], + "type": "text", + "content": "[24] Tsun-Hsuan Wang, Sivaban Manivasagam, Ming Liang, Bin Yang, Wenyuan Zeng, and Raquel Urtasun. V2VNet: Vehicle-to-Vehicle Communication for Joint Perception and Prediction, page 605-621. Springer International Publishing, Cham, 2020. 1, 2, 3, 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 680, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 680, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 680, + 554, + 713 + ], + "type": "text", + "content": "[25] Yan Xia, Yusheng Xu, Shuang Li, Rui Wang, Juan Du, Daniel Cremers, and Uwe Stilla. Soe-net: A self-attention and orientation encoding network for point cloud based place" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22304" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 76, + 72, + 296, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 296, + 106 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 296, + 106 + ], + "type": "text", + "content": "recognition. In Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition, pages 11348-11357, 2021. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 296, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 296, + 163 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 296, + 163 + ], + "type": "text", + "content": "[26] Yan Xia, Mariia Gladkova, Rui Wang, Qianyun Li, Uwe Stilla, João F Henriques, and Daniel Cremers. Casspr: Cross attention single scan place recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 8461-8472, 2023. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 164, + 296, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 164, + 296, + 207 + ], + "spans": [ + { + "bbox": [ + 56, + 164, + 296, + 207 + ], + "type": "text", + "content": "[27] Li Xiang, Junbo Yin, Wei Li, Cheng-Zhong Xu, Ruigang Yang, and Jianbing Shen. DI-V2X: Learning domain-invariant representation for vehicle-infrastructure collaborative 3d object detection. AAAI, 2024. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 209, + 295, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 209, + 295, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 209, + 295, + 262 + ], + "type": "text", + "content": "[28] Runsheng Xu, Hao Xiang, Zhengzhong Tu, Xin Xia, Ming-Hsuan Yang, and Jiaqi Ma. V2x-vit: Vehicle-to-everything cooperative perception with vision transformer. In Computer Vision - ECCV 2022, page 107-124, Cham, 2022. Springer Nature Switzerland. 1, 2, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 264, + 295, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 264, + 295, + 319 + ], + "spans": [ + { + "bbox": [ + 56, + 264, + 295, + 319 + ], + "type": "text", + "content": "[29] Runsheng Xu, Hao Xiang, Xin Xia, Xu Han, Jinlong Li, and Jiaqi Ma. Opv2v: An open benchmark dataset and fusion pipeline for perception with vehicle-to-vehicle communication. 2022 International Conference on Robotics and Automation (ICRA), page 2583–2589, 2022. 1, 2, 3, 5, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 321, + 295, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 321, + 295, + 397 + ], + "spans": [ + { + "bbox": [ + 56, + 321, + 295, + 397 + ], + "type": "text", + "content": "[30] Runsheng Xu, Xin Xia, Jinlong Li, Hanzhao Li, Shuo Zhang, Zhengzhong Tu, Zonglin Meng, Hao Xiang, Xiaoyu Dong, Rui Song, Hongkai Yu, Bolei Zhou, and Jiaqi Ma. V2v4real: A real-world large-scale dataset for vehicle-to-vehicle cooperative perception. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), page 13712–13722, Vancouver, BC, Canada, 2023. IEEE. 1, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 399, + 295, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 399, + 295, + 432 + ], + "spans": [ + { + "bbox": [ + 56, + 399, + 295, + 432 + ], + "type": "text", + "content": "[31] Yan Yan, Yuxing Mao, and Bo Li. Second: Sparsely embedded convolutional detection. Sensors (Basel, Switzerland), 18, 2018. 3, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 434, + 295, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 434, + 295, + 477 + ], + "spans": [ + { + "bbox": [ + 56, + 434, + 295, + 477 + ], + "type": "text", + "content": "[32] Binh Yang, Wenjie Luo, and Raquel Urtasun. Pixor: Realtime 3d object detection from point clouds. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7652-7660, 2018. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 479, + 295, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 479, + 295, + 534 + ], + "spans": [ + { + "bbox": [ + 56, + 479, + 295, + 534 + ], + "type": "text", + "content": "[33] Hongbo Yin, Daxin Tian, Chunmian Lin, Xuting Duan, Jianshan Zhou, Dezong Zhao, and Dongpu Cao. V2VFormer ++: Multi-modal vehicle-to-vehicle cooperative perception via global-local transformer. IEEE Transactions on Intelligent Transportation Systems, page 1-14, 2023. 1, 2, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 535, + 295, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 535, + 295, + 590 + ], + "spans": [ + { + "bbox": [ + 56, + 535, + 295, + 590 + ], + "type": "text", + "content": "[34] Hongbo Yin, Daxin Tian, Chunmian Lin, Xuting Duan, Jianshan Zhou, Dezong Zhao, and Dongpu Cao. V2vformer++: Multi-modal vehicle-to-vehicle cooperative perception via global-local transformer. IEEE Transactions on Intelligent Transportation Systems, 25:2153-2166, 2024. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 591, + 295, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 591, + 295, + 667 + ], + "spans": [ + { + "bbox": [ + 56, + 591, + 295, + 667 + ], + "type": "text", + "content": "[35] Haibao Yu, Yizhen Luo, Mao Shu, Yiyi Huo, Zebang Yang, Yifeng Shi, Zhenglong Guo, Hanyu Li, Xing Hu, Jirui Yuan, and Zaiqing Nie. Dair-v2x: A large-scale dataset for vehicle-infrastructure cooperative 3d object detection. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), page 21329-21338, New Orleans, LA, USA, 2022. IEEE. 1, 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 669, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 669, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 669, + 295, + 713 + ], + "type": "text", + "content": "[36] Haibao Yu, Yingjuan Tang, Enze Xie, Jilei Mao, Ping Luo, and Zaiqing Nie. Flow-based feature fusion for vehicleinfrastructure cooperative 3d object detection. In Advances in Neural Information Processing Systems, 2023. 3" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 540 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 160 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 160 + ], + "type": "text", + "content": "[37] Haibao Yu, Wenxian Yang, Hongzhi Ruan, Zhenwei Yang, Yingjuan Tang, Xu Gao, Xin Hao, Yifeng Shi, Yifeng Pan, Ning Sun, Juan Song, Jirui Yuan, Ping Luo, and Zaiqing Nie. V2x-seq: A large-scale sequential dataset for vehicleinfrastructure cooperative perception and forecasting. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), page 5486-5495, Vancouver, BC, Canada, 2023. IEEE. 2, 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 162, + 553, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 162, + 553, + 216 + ], + "spans": [ + { + "bbox": [ + 316, + 162, + 553, + 216 + ], + "type": "text", + "content": "[38] Yunshuang Yuan and Monika Sester. COMAP: A synthetic dataset for collective multi-agent perception of autonomous driving. The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, XLIII-B2-2021:255-263, 2021. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 217, + 553, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 217, + 553, + 271 + ], + "spans": [ + { + "bbox": [ + 316, + 217, + 553, + 271 + ], + "type": "text", + "content": "[39] Yunshuang Yuan and Monika Sester. Leveraging dynamic objects for relative localization correction in a connected autonomous vehicle network. ISPRS Annals of the Photogrammetry, Remote Sensing and Spatial Information Sciences, V-1-2022:101-109, 2022. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 273, + 553, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 273, + 553, + 316 + ], + "spans": [ + { + "bbox": [ + 316, + 273, + 553, + 316 + ], + "type": "text", + "content": "[40] Yunshuang Yuan and Monika Sester. StreamLTS: Query-based temporal-spatial lidar fusion for cooperative object detection. In Computer Vision - ECCV 2024 Workshops, 2024. 2, 3, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 318, + 553, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 318, + 553, + 361 + ], + "spans": [ + { + "bbox": [ + 316, + 318, + 553, + 361 + ], + "type": "text", + "content": "[41] Yunshuang Yuan, Hao Cheng, and Monika Sester. Keypoints-based deep feature fusion for cooperative vehicle detection of autonomous driving. IEEE Robotics and Automation Letters, 7(2):3054-3061, 2022. 1, 2, 3, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 363, + 553, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 363, + 553, + 407 + ], + "spans": [ + { + "bbox": [ + 316, + 363, + 553, + 407 + ], + "type": "text", + "content": "[42] Yunshuang Yuan, Hao Cheng, Michael Ying Yang, and Monika Sester. Generating evidential bev maps in continuous driving space. ISPRS Journal of Photogrammetry and Remote Sensing, 204:27-41, 2023. 1, 2, 3, 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 408, + 553, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 408, + 553, + 441 + ], + "spans": [ + { + "bbox": [ + 316, + 408, + 553, + 441 + ], + "type": "text", + "content": "[43] Wu Zheng, Weiliang Tang, Sijin Chen, Li Jiang, and Chi-Wing Fu. CIA-SSD: Confident iou-aware single-stage object detector from point cloud. In AAAI, 2021. 3, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 441, + 553, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 441, + 553, + 485 + ], + "spans": [ + { + "bbox": [ + 316, + 441, + 553, + 485 + ], + "type": "text", + "content": "[44] Yin Zhou and Oncel Tuzel. Voxelnet: End-to-end learning for point cloud based 3d object detection. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4490-4499, 2017. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 487, + 553, + 540 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 487, + 553, + 540 + ], + "spans": [ + { + "bbox": [ + 316, + 487, + 553, + 540 + ], + "type": "text", + "content": "[45] Walter Zimmer, Christian Creß, Huu Tung Nguyen, and Alois C Knoll. Tumtraf intersection dataset: All you need for urban 3d camera-lidar roadside perception [best student paper award]. In 2023 IEEE Intelligent Transportation Systems ITSC. IEEE, 2023. 8" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22305" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Spatial Transport Optimization by Repositioning Attention Map for Training-Free Text-to-Image Synthesis/68174923-d215-4833-8530-d036bb11b7fd_content_list.json b/2025/Spatial Transport Optimization by Repositioning Attention Map for Training-Free Text-to-Image Synthesis/68174923-d215-4833-8530-d036bb11b7fd_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..441f006c582276a2b578a719e06477ef7b45ee8d --- /dev/null +++ b/2025/Spatial Transport Optimization by Repositioning Attention Map for Training-Free Text-to-Image Synthesis/68174923-d215-4833-8530-d036bb11b7fd_content_list.json @@ -0,0 +1,1407 @@ +[ + { + "type": "text", + "text": "Spatial Transport Optimization by Repositioning Attention Map for Training-Free Text-to-Image Synthesis", + "text_level": 1, + "bbox": [ + 173, + 130, + 823, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Woojung Han Yeonkyung Lee Chanyoung Kim Kwanghyun Park Seong Jae Hwang* Yonsei University", + "bbox": [ + 138, + 203, + 864, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{dnwjDDL, yeonkyung.lee, chanyoung, kwanghyun.park, seongjae}@yonsei.ac.kr", + "bbox": [ + 169, + 241, + 823, + 257 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4ac46047851eb13618cbfeaeacfa209b6ec31355d9bf81b5dd48867cccff1c19.jpg", + "image_caption": [ + "Figure 1. Three main challenges in training-free text-to-image (T2I) generation: (1) missing objects, (2) mismatched attributes, and (3) mislocated objects. While existing approaches address missing objects and mismatched attributes, effectively controlling object positioning remains an open problem. Our proposed model, STORM, introduces a dynamic approach to aligning relative object positions throughout the denoising process, enabling precise spatial control without additional spatial templates. Red underline highlights errors made by SD." + ], + "image_footnote": [], + "bbox": [ + 93, + 297, + 903, + 523 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 595, + 326, + 611 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Diffusion-based text-to-image (T2I) models have recently excelled in high-quality image generation, particularly in a training-free manner, enabling cost-effective adaptability and generalization across diverse tasks. However, while the existing methods have been continuously focusing on several challenges such as \"missing objects\" and \"mismatched attributes,\" another critical issue of \"mislocated objects\" remains where generated spatial positions fail to align with text prompts. Surprisingly, ensuring such seemingly basic functionality remains challenging in popular T2I models due to the inherent difficulty of imposing explicit spatial guidance via text forms. To address this, we propose STORM (Spatial Transport Optimization by Repositioning Attention Map), a novel training-free approach for spatially coherent T2I synthesis. STORM employs Spatial Transport Optimization (STO), rooted in optimal transport theory, to", + "bbox": [ + 88, + 628, + 485, + 872 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "dynamically adjust object attention maps for precise spatial adherence, supported by a Spatial Transport (ST) Cost function that enhances spatial understanding. Our analysis shows that integrating spatial awareness is most effective in the early denoising stages, while later phases refine details. Extensive experiments demonstrate that STORM surpasses existing methods, effectively mitigating mislocated objects while improving missing and mismatched attributes, setting a new benchmark for spatial alignment in T2I synthesis.", + "bbox": [ + 511, + 597, + 908, + 733 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 746, + 643, + 760 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Diffusion-based text-to-image (T2I) models exhibit remarkable versatility across various applications, generating high-quality images from textual prompts [13, 22, 24, 32, 41]. Notably, training-free methods offer efficiency, real-time adaptability, lower costs, and broad task generalization [5, 6, 9]. These training-free methods leverage their advantages while addressing three key challenges in fully capturing textual details. In particular, as shown in Fig. 1, these challenges include (1) missing objects, (2) mismatched at", + "bbox": [ + 511, + 763, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "Project Page: https://micv-yonsei.github.io/storm2025/", + "bbox": [ + 112, + 887, + 457, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "18401", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tributes, and (3) mislocated objects. These limitations may lead to images that poorly represent the intended scene or convey misleading information, reducing their reliability.", + "bbox": [ + 89, + 90, + 482, + 136 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "While there has been substantial progress in addressing issues like missing objects and mismatched attributes in diffusion-based image synthesis [5, 12, 21, 23], the challenge of mislocated objects remains relatively unexplored despite its critical importance. For example, in Fig. 1, the fifth column shows that the backpack should be below a donut, but Stable Diffusion (SD) [32] generates donut to the left of a backpack. Furthermore, SD consistently generates similar images regardless of spatial cues in the text (e.g., same seed with only positional changes for teddy bear and vase in Fig. 2), indicating limited spatial understanding. This limitation highlights the importance of addressing the problem of mislocated objects and demonstrates the need for more advanced methods of interpreting and aligning with spatial cues. Overcoming these challenges is key to improving T2I model reliability and supporting adaptable applications across diverse settings.", + "bbox": [ + 88, + 138, + 482, + 396 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Existing solutions for mislocated objects often rely on fixed spatial templates or predefined layouts [6, 7, 38], which offer some control but impose rigid constraints and require additional inputs. For instance, these templates may fix an object's position in the image (e.g., far-left corner) instead of offering flexible guidance like placing one object \"to the left\" of another. While effective in structured fields like automated design, these approaches present limitations in more dynamic areas, such as digital art or interactive media, by restricting spatial flexibility. Therefore, instead of relying on restrictive templates or requiring additional inputs, we aim to push the boundaries of T2I models by enabling precise alignment with textual guidance, which is essential to unlocking their full potential across applications.", + "bbox": [ + 89, + 398, + 482, + 609 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To this end, we propose Spatial Transport Optimization by Repositioning Attention Map, STORM, a training-free approach that adjusts relative object positions dynamically, without relying on fixed spatial templates. While SD disregards spatial cues (see Fig. 2), STORM achieves various object positioning with the same model weights, ensuring spatial alignment. To implement STORM, we introduce the Spatial Transport Optimization (STO) framework, which interprets each text token's attention map as a distribution, and leverages Optimal Transport (OT) [37] to efficiently reposition distributions. At an early diffusion timestep, STO adjusts the placement of objects based on their relative positioning (see Fig.3 for attention map changes across the denoising process). While OT considers both distribution and distance to minimize transformation costs, we propose the Spatial Transport Cost (ST Cost) to align with our objectives and guide distribution positioning. This function updates and optimizes the latent vector through backward guidance, following methods from prior work [5, 6].", + "bbox": [ + 89, + 613, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/516592f1799a6376dd7c3f34af248e0d527c1e134b0644c6d03cbd802ad9e1e9.jpg", + "image_caption": [ + "Figure 2. Comparison of Spatial Awareness. $\\{\\text{position}^*\\}$ in each prompt denotes the spatial relationship in each column (e.g., \"to the left of\", \"to the right of\", \"above, and below\"). While Stable Diffusion (SD) shows limited spatial awareness by generating similar images regardless of spatial prompts, our model accurately reflects specified positions. (Same seed for all synthesis)." + ], + "image_footnote": [], + "bbox": [ + 516, + 87, + 898, + 253 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our analysis suggests that integrating spatial awareness yields optimal results in the early denoising stages, whereas the later stages are better suited for focusing on fine details. Furthermore, we found that missing objects can be resolved concurrently as mislocated objects are corrected in the early steps. Extensive experiments show that our approach outperforms existing methods, setting a new benchmark for spatial alignment in diffusion models.", + "bbox": [ + 511, + 353, + 906, + 474 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our main contributions are as follows:", + "bbox": [ + 527, + 478, + 785, + 491 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose STORM, a training-free method that dynamically adjusts relative object positions, achieving accurate spatial alignment without relying on fixed templates.", + "- We introduce the Spatial Transport Optimization (STO) framework based on Optimal Transport (OT), which treats attention maps as distributions and repositions objects relative to each other across diffusion timesteps.", + "- We develop ST Cost function that penalizes overlaps and progressively enhances spatial accuracy, demonstrating its effectiveness through extensive experiments that set a new benchmark for spatial alignment in T2I models." + ], + "bbox": [ + 513, + 496, + 903, + 662 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 681, + 653, + 696 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Text-to-Image Synthesis. Text-to-image (T2I) synthesis aims to generate visually realistic images that align with given text prompts. Early studies focused on GANs [18, 19, 27] and autoregressive models [29, 39], but diffusion models [15] have taken the lead, demonstrating remarkable performance in T2I synthesis [22, 32]. Despite advances driven by VLMs [32, 34, 39], aligning generated images with text prompts remains a challenge. To address this, some approaches [2, 30, 34] have expanded network architectures, though these methods require retraining and are difficult to adapt to existing frameworks. Alternatively, other research [5, 9, 21, 23] focused on training-free improvement strategies, which aligns closely with our approach.", + "bbox": [ + 511, + 703, + 906, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "18402", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/90b79f3547110a014b5a218fc9829f909e771d7e8a89a6371c54e946fc654ddf.jpg", + "image_caption": [ + "Figure 3. Comparison of Attention Map Progression During Denoising. Visualization of the attention maps for \"a red bird to the right of a green plant\" throughout the denoising process for both Stable Diffusion (a) and our model (b). While Stable Diffusion struggles to distinctly capture the spatial relationship between the bird and the plant, our model effectively aligns the objects according to the specified spatial cue (\"to the right of\"). The resulting image from our model demonstrates improved spatial accuracy compared to Stable Diffusion." + ], + "image_footnote": [], + "bbox": [ + 94, + 88, + 478, + 303 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Training-free T2I Models. Previous studies [1, 5, 9, 12, 21, 22, 31] have explored various approaches to enhance how text prompts are aligned within text-to-image diffusion models. Composable Diffusion [22] and Structure Diffusion [9] aim to improve compositionality through multi-model or cross-attention methods, yet they still face issues like subject mixing and attribute binding. To address these, Attend-and-Excite [5] dynamically adjusts attention during inference to capture all subjects, while Divide-and-Bind [21] improves attribute binding and compositional accuracy with targeted loss functions. CONFORM [23] uses a contrastive method to separate objects in attention maps while keeping related attributes together, and INITNO [12] further tackles spatial misalignment by improving control over object placement to reduce overlap. Although these studies focus on the missing object issue by reducing object overlap, they often overlook the crucial task of capturing accurate relative positioning and directional relationships between objects.", + "bbox": [ + 91, + 441, + 483, + 729 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Training-free T2I Models for Mislocated Objects. The advantages of attention maps for controlling spatial positioning have led to several studies using fixed templates, such as layouts, to improve object placement [6, 7, 11, 25, 35, 38]. Wu et al. [38] use layout predictors to assign pixel regions, while ZestGuide [7] and Chen et al. [6] leverage cross-attention layers for alignment with segmentations or user-defined layouts. However, these methods often require additional layout inputs and lack flexibility for relative positioning. As a result, there is a need for methods that enhance object placement without relying on predefined layouts.", + "bbox": [ + 89, + 734, + 483, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Preliminaries", + "text_level": 1, + "bbox": [ + 513, + 89, + 651, + 104 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Text Conditioned via Cross-Attention. Our approach uses Stable Diffusion (SD) [32], which generates images from text by integrating text conditioning through cross-attention. A pre-trained CLIP encoder [28] converts the text prompt into a conditioning vector $c$ , fed into the UNet model [33] via cross-attention layers at different resolutions. At each timestep $t$ , the attention map $A_{t} \\in \\mathbb{R}^{P \\times P \\times N}$ assigns probabilities for transferring token information to spatial positions, where $P$ is the spatial dimension of the feature map and $N$ the number of tokens in the text prompt. In our work, we use $16 \\times 16$ attention maps as they contain the most semantically meaningful information [14].", + "bbox": [ + 511, + 107, + 906, + 289 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Optimal Transport with Sinkhorn Algorithm. Optimal Transport (OT) [37] provides a framework for transforming one probability distribution into another while preserving spatial relationships. Given two discrete distributions $\\mu = \\{\\mu_u\\}_{u=1}^n$ and $\\nu = \\{\\nu_v\\}_{v=1}^n$ , each with $n$ discrete locations, and a cost matrix $\\mathbf{C} \\in \\mathbb{R}^{n \\times n}$ , where $\\mathbf{C}_{uv}$ is the cost of moving distribution from point $u$ to $v$ , OT seeks a transport plan $\\mathbf{P} \\in \\mathbb{R}^{n \\times n}$ that minimizes the total cost, $\\min_{P \\geq 0} \\sum_{u=1}^{n} \\sum_{v=1}^{n} \\mathbf{C}_{uv} \\mathbf{P}_{uv}$ . For large $n$ , the Sinkhorn algorithm [8] applies entropic regularization for a faster, scalable, and differentiable solution:", + "bbox": [ + 511, + 291, + 908, + 445 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {P \\geq 0} \\quad \\sum_ {u = 1} ^ {n} \\sum_ {v = 1} ^ {n} \\mathbf {C} _ {u v} \\mathbf {P} _ {u v} - \\varepsilon \\sum_ {u = 1} ^ {n} \\sum_ {v = 1} ^ {n} \\mathbf {P} _ {u v} (\\log \\mathbf {P} _ {u v} - 1), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 527, + 450, + 903, + 484 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\varepsilon > 0$ balances transport cost and entropy. We further integrate spatial alignment objectives into this OT-based approach in our framework.", + "bbox": [ + 511, + 483, + 906, + 527 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Method", + "text_level": 1, + "bbox": [ + 513, + 532, + 604, + 547 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We describe the full pipeline of our model, STORM, as shown in Fig. 4. To support the understanding of our Spatial Transport Optimization (STO) framework, we first provide background information in Sec. 4.1. The full details of STO in Sec. 4.2, while Sec. 4.3 covers optimizations for non-spatial text prompts, and Sec. 4.4 details the update process.", + "bbox": [ + 511, + 551, + 906, + 643 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1. Background of STO", + "text_level": 1, + "bbox": [ + 511, + 645, + 705, + 661 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Since objects are generated in regions of high attention within the attention map (see Fig. 3), our goal is to reposition the attention map based on the positional information from the text prompt, using relative objects as a reference. To achieve this, our approach involves three types of attention maps: the attention map of the object to be moved, called the source distribution (e.g., blue area in Fig. 4); the attention map of the relative object, which acts as a reference distribution (e.g., red area in Fig. 4); and the target distribution which represents the intended position for the source distribution to move (e.g., gray area in Fig. 4). The target distribution is an arbitrary distribution defined based on the reference point, which is the centroid of the reference distribution, $(j_{\\mathrm{ref}}, i_{\\mathrm{ref}})$ . It primarily serves to guide the intended position of the source distribution using reference points without exerting significant direct influence.", + "bbox": [ + 511, + 662, + 908, + 904 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "18403", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/8096962e3fc7166fc54cbbc04e1742ffb70c4eac9855ddc3a881a154e038120b.jpg", + "image_caption": [ + "Figure 4. Overview pipeline of STORM. Our method leverages Optimal Transport in a training-free manner, allowing the model to accurately reflect relative object positions at each step without additional inputs. Given the prompt \"A car to the left of an elephant\", our method dynamically adjusts the attention maps to induce the specified spatial relationship. The process starts with initial attention maps for the \"car\" and \"elephant\" at time step $z_{t}$ . Using the centroids of these attention maps, the Spatial Transport Optimization (STO) computes the losses to correct positional relationships (e.g., ensuring the car is to the left of the elephant). The updated attention map is then used to refine the latent representation $z_{t}$ , leading to a final image that adheres to the desired spatial arrangement. The comparison of attention maps (before and after STO) shows improved alignment, effectively placing the car to the left of the elephant as instructed in the prompt." + ], + "image_footnote": [], + "bbox": [ + 94, + 89, + 897, + 327 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Spatial Transport Optimization Framework", + "text_level": 1, + "bbox": [ + 89, + 440, + 464, + 455 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we detail our Spatial Transport Optimization (STO) framework, which dynamically repositions attention maps for accurate object alignment without modifying model weights. This framework consists of two main components: a cost function and a transport plan. The cost function directs the distribution shift from the source in the intended direction, allowing the attention map to adjust positionally according to spatial relationships. Although we build on standard cost functions that measure distance between locations, our Spatial Transport Cost (ST Cost) incorporates additional objectives adapted to our specific spatial alignment goals. The transport plan, which optimally moves the distribution between points, is implemented using the Sinkhorn algorithm. Beginning with an initial attention map, STO iteratively updates the latent vector by applying STO-derived losses, gradually guiding the distribution into the target alignment.", + "bbox": [ + 88, + 458, + 482, + 715 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2.1 Construction of Target Distributions", + "text_level": 1, + "bbox": [ + 89, + 719, + 398, + 734 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our approach focuses on relative positioning adjustments at each timestep, using an arbitrary target distribution (e.g., gray area in Fig. 4) to guide the source distribution based on the reference. Depending on the text-specified direction (e.g., horizontal or vertical), only the relevant reference point of $(j_{\\mathrm{ref}},i_{\\mathrm{ref}})$ is considered: for 'left' or 'right' positioning, only the horizontal dimension of the reference point $j_{\\mathrm{ref}}$ is used, while for 'above' or 'below' positioning, only the vertical dimension $i_{\\mathrm{ref}}$ is considered, allowing the flexibility in the opposite dimension. For example, in Fig. 4a, with the car attention map as a source distribution and the", + "bbox": [ + 89, + 734, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "elephant attention map as a reference distribution, the target distribution is placed to the left of a elephant's distribution. This approach similarly applies to another source. Supplementary Materials provide details on computing the target distribution and reference point.", + "bbox": [ + 511, + 441, + 906, + 517 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2.2 Spatial Transport Cost Function", + "text_level": 1, + "bbox": [ + 511, + 523, + 792, + 537 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the STO framework, we aim to find an optimal transport plan that minimizes the cost of transferring the attention map. The cost function measures the expense of moving distribution mass by computing distances between points, where a higher value indicates less desirable locations. While our overall cost function merges the standard OT loss with an additional term (detailed in the supplementary material), here we focus on our newly introduced ST Cost, which is built on two core principles:", + "bbox": [ + 511, + 537, + 906, + 672 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "(I) Positional Cost: We developed a cost function to guide the distribution to its intended position relative to the reference point. In simple terms, when the source distribution is intended to be on the left side of the reference point, it has a low cost when positioned on the left and a high cost when positioned on the right. To quantify this directional positioning, we define a set of positional $\\delta$ values, representing the distances between the current point and the reference point in each of the four directions. Specifically, the $\\delta$ values are defined as follows: $\\delta_{ij}^{\\leftarrow} = j_{\\mathrm{ref}} - j,\\delta_{ij}^{\\rightarrow} = j - j_{\\mathrm{ref}},\\delta_{ij}^{\\uparrow} = i_{\\mathrm{ref}} - i,\\delta_{ij}^{\\downarrow} = i - i_{\\mathrm{ref}}$ with $j$ and $i$ representing the current position values, and the arrows indicate the desired and restricted direction for positioning the source attention map relative to the reference point. When constructing the overall cost function that incorporates all directional $\\delta$ values, each direction is categorized as either de", + "bbox": [ + 511, + 672, + 908, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "18404", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "sired or restricted. The desired direction $(\\delta^{\\mathrm{des}})$ corresponds to the intended movement of the source distribution, while the restricted direction $(\\delta^{\\mathrm{res}})$ represents the opposite or undesired direction. For instance, if the source distribution should move to the left, $\\delta_{ij}^{\\leftarrow}$ is marked as the desired direction $(\\delta^{\\mathrm{des}})$ , while $\\delta_{ij}^{\\rightarrow}$ is treated as the restricted direction $(\\delta^{\\mathrm{res}})$ . The combined cost function, accounting for both desired and restricted positions, can be expressed as:", + "bbox": [ + 89, + 90, + 480, + 205 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\Delta_ {i j} \\left(\\delta_ {i j} ^ {\\mathrm {d e s}}, \\delta_ {i j} ^ {\\mathrm {r e s}}\\right) = \\frac {1}{\\omega \\left(\\delta_ {i j} ^ {\\mathrm {d e s}} + \\epsilon\\right)} \\mathbb {1} _ {+} \\left[ \\delta_ {i j} ^ {\\mathrm {d e s}} \\right] + \\omega \\left(\\delta_ {i j} ^ {\\mathrm {r e s}} + \\epsilon\\right) \\mathbb {1} _ {+} \\left[ \\delta_ {i j} ^ {\\mathrm {r e s}} \\right], \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 209, + 482, + 239 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\omega (\\cdot) > 1$ is a progressive adaptive weight that controls the alignment importance across timesteps, and $\\epsilon$ is a stabilizing factor. $\\mathbb{1}_{+}$ is an indicator function that equals 1 when the input is positive and 0 otherwise, allowing selective application of computation based on positivity. As the source distribution aligns more closely with the desired direction, the $\\Delta$ value approaches zero, indicating minimal cost for correctly aligned positioning (see Example below).", + "bbox": [ + 89, + 247, + 483, + 368 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "(II) Non-overlap Cost: In this process, it is crucial to prevent the source distribution from overlapping with the reference distribution during its movement. This is one of the primary issues in addressing the missing object problem in SD, where avoiding overlap is essential for ensuring that each object is distinctly represented. To achieve this, we simply embed the reference distribution directly into the cost function, guiding objects to occupy separate locations. Specifically, with $A_{ij}$ representing the attention weight of the reference distribution, we incorporated this distribution into the cost function, assigning high costs to regions occupied by the reference distribution and low costs elsewhere.", + "bbox": [ + 89, + 373, + 483, + 553 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Combining this principle with our core ideas, our cost function can be defined as follows:", + "bbox": [ + 89, + 554, + 482, + 580 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nC _ {i j} = A _ {i j} \\Delta_ {i j} \\left(\\delta_ {i j} ^ {\\mathrm {d e s}}, \\delta_ {i j} ^ {\\mathrm {r e s}}\\right). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 207, + 592, + 480, + 609 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Example. In Fig. 4, the car should be left of the elephant, and the elephant to the right of the car. The cost functions for each of them are represented by $C_{ij}^{\\leftarrow} = A_{ij}\\Delta_{ij}(\\delta_{ij}^{\\leftarrow},\\delta_{ij}^{\\rightarrow})$ , $C_{ij}^{\\rightarrow} = A_{ij}\\Delta_{ij}(\\delta_{ij}^{\\rightarrow},\\delta_{ij}^{\\leftarrow})$ , respectively. Taking the car as an example, it needs to be positioned to the left of the elephant relative to the elephant's centroid point. In more detail, if the current point is positioned to the left of the centroid, the $\\delta_{ij}^{\\mathrm{des}}$ is activated; if positioned to the right of the centroid, the $\\delta_{ij}^{\\mathrm{res}}$ is activated. Additionally, as the car moves to the left of the elephant, it should not overlap the elephant it passes along the way. The elephant's attention map is used in the cost function, guiding the car to avoid overlapping with the elephant's area. The resulting $C_{ij}$ is utilized as the cost function for STO.", + "bbox": [ + 89, + 617, + 483, + 829 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2.3 Sinkhorn Algorithm-based Transport Plan", + "text_level": 1, + "bbox": [ + 89, + 840, + 442, + 854 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After combining the standard OT approach with ST components (see Supplementary Material for the complete formulation), we compute the transport plan using the Sinkhorn", + "bbox": [ + 89, + 854, + 482, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "algorithm [8], which applies entropic regularization for computational efficiency and stability, especially for high-dimensional attention maps. The regularized OT problem is formulated as in Eq. (1) using target distribution calculated in Sec. 4.2.1. The Sinkhorn algorithm iteratively updates transport plan $\\mathbf{P}$ to meet the marginal constraints of source and target distributions, applying the cost function bidirectionally for both objects.", + "bbox": [ + 511, + 90, + 903, + 210 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Optimization for Objects and Attributes", + "text_level": 1, + "bbox": [ + 511, + 215, + 857, + 231 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our method also extends to cases where the text lacks spatial information. (i) Object-level Optimization: We handle cases where the text prompt specifies objects without explicit spatial guidance by randomly placing the target distribution of each object. This flexible positioning encourages diverse object placement when spatial relationships are not provided. Therefore, the cost function is simplified to focus on solely on preventing overlap, expressed as $C_{ij} = A_{ij} \\otimes \\mathbf{1}_N$ , where $A$ is the reference attention map. (ii) Attribute Matching Optimization: The attention map for an object and its corresponding attribute should be identical, reflecting the same attention focus. Therefore, we used the standard OT cost function, calculating Euclidean distance between corresponding positions across the object, and attribute attention maps to quantify differences.", + "bbox": [ + 511, + 232, + 903, + 458 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.4. Optimization and Update Process", + "text_level": 1, + "bbox": [ + 511, + 464, + 808, + 479 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We update the latent vector using STO-based losses to guide the spatial alignment of the attention maps in a differentiable manner. At each timestep, we compute the OT loss $\\mathcal{L}$ and update the latent vector $z_{t}$ as $z_{t}^{\\prime}\\gets z_{t} - \\alpha_{t}\\cdot \\nabla z_{t}\\mathcal{L}$ where $\\alpha_{t}$ is the step size. After the update, a forward pass computes $z_{t - 1}$ for the next denoising step.", + "bbox": [ + 511, + 482, + 903, + 571 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Remark. As shown in Fig.8, addressing the mislocated object problem is most effective in early timesteps, which, according to related work[26], establish the overall image structure while later timesteps refine finer details. Based on this, STORM establishes object positioning and spatial configurations early in the denoising process, minimizing the risk of misplacements and overlaps. Furthermore, as shown in the bottom row of Fig. 8, solving the mislocated object problem before addressing missing objects can naturally resolve the latter, highlighting the importance of prioritizing mislocated objects as a critical step in the synthesis process.", + "bbox": [ + 511, + 575, + 905, + 742 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 511, + 758, + 643, + 773 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation Details. Following methods established in prior research [5, 21, 23], we conducted a comprehensive evaluation of our model. We used the official Stable Diffusion models [32] v1.4 and v2.1, configuring the denoising timestep to 50, and the guidance scale to 7.5 in line with common practices in the field. To smooth the cross-attention map, a Gaussian filter with a kernel size of 3 and a standard deviation of 0.5 was applied. Optimization", + "bbox": [ + 511, + 779, + 905, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "18405", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/855e16195c9c9137b7cf42571bd23f7c7ccc4dcc00637fcc9f96d36e00520bf4.jpg", + "table_caption": [ + "Table 1. Performance comparison between different models on VISOR (\\%) and Object Accuracy (OA) (\\%) metrics, based on Stable Diffusion 1.4 and Stable Diffusion 2.1 [32]. For entries marked with $\\dagger$ , official values were unavailable; thus, we directly calculated these values. Bold values indicate the best scores, while underlined values indicate the second-best scores among the fair comparison. Notes on VISOR. VISOR cond: spatial accuracy using the results in which both objects are sufficiently generated; VISOR uncond: spatial accuracy using all results, including results with failed object generations; VISOR $1/2/3/4$ : ratios of one, two, three, or four spatially accurate results out of four seed results, respectively. (e.g., $\\mathrm{VISOR}_1$ : % of results with exactly one spatially correct result out of four generated results)" + ], + "table_footnote": [], + "table_body": "
ModelVenueExtra TrainingExtra InputOA (%)VISOR (%)
uncondcond1234
Stable Diffusion 1.4 Based
SD 1.4 [32]-X29.8618.8162.9846.6020.116.891.63
SD 1.4 + CDM [22]ECCV22✓X23.2714.9964.4139.4414.564.841.12
GLIDE [24]ICML22✓X3.361.9859.066.721.020.170.03
GLIDE + CDM [22]ECCV22✓X10.176.4363.2120.074.690.830.11
Control-GPT [42]arXiv23✓✓48.3344.1765.9769.8051.2035.6720.48
Layout Guidance [6]WACV24X✓40.0138.8095.95----
REVISION [3]ECCV24X✓53.9652.7197.6977.7961.0244.9027.15
Structure Diffusion [9]ICLR23XX28.6517.8762.3644.7018.736.571.46
Attend-and-Excite [5]SIGGRAPH23XX42.0725.7561.2149.2919.334.560.08
Divide-and-Bind†[21]BMVC24XX46.0331.6268.7064.7237.8218.645.30
INITNO†[12]CVPR24XX60.4035.1858.2471.2042.7120.096.72
CONFORM†[23]CVPR24XX60.7338.4862.3373.0145.8225.579.52
Ours (SD 1.4)XX61.0157.5894.3985.9369.7149.0125.70
Stable Diffusion 2.1 Based
SD 2.1 [32]-X47.8330.2563.2464.4235.7416.134.70
SPRIGHT [4]ECCV24✓X60.6842.2371.2471.7851.8833.0916.15
REVISION [3]ECCV24X✓48.2647.1197.6176.0755.7537.1019.53
Ours (SD 2.1)XX62.5559.3594.8888.3471.7552.0325.42
", + "bbox": [ + 94, + 172, + 901, + 455 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "was performed iteratively with forward updates at specified timesteps ( $t \\in 5, 10, 15, 20$ ), stopping at $t = 25$ to enhance quality. All experiments use NVIDIA RTX A6000 Ada*.", + "bbox": [ + 89, + 469, + 482, + 518 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Evaluation Results", + "text_level": 1, + "bbox": [ + 89, + 530, + 272, + 544 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We evaluated our method against existing training-free T2I models using both qualitative and quantitative measures, focusing on two key aspects. (i) Spatial Accuracy: To measure alignment with spatial configurations specified in prompts, we used the VISOR [10] and the T2I-CompBench [17], both provide a comprehensive assessment of spatial relationships captured by the model. (ii) Object and Attribute Consistency: We assessed object and attribute representation using T2I-CompBench in combination with CLIP-space distance [28] and TIFA score [16], following established methods [5, 21, 23]. We provide detailed descriptions of each evaluation metric and datasets in the Supplementary Materials for additional clarity.", + "bbox": [ + 88, + 547, + 482, + 744 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Quantitative Evaluation: VISOR. To evaluate spatial accuracy, we use the VISOR [10] to measure the ability of a model to position objects based on spatial cues (e.g., above) in text prompts. VISOR benchmark contains over 25K prompts describing spatial relationships between objects. Notably, as seen in Table 1, the increase in $\\mathrm{VISOR}_4$ demonstrates the model's reliability in generating spatially accurate images across prompts and seeds.", + "bbox": [ + 88, + 750, + 482, + 869 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/72ea1ed6d83bee0e516d508fa5acedd7097248e3268d1b20eeb4fb7535b4a2da.jpg", + "table_caption": [ + "Table 2. Comparison of methods on T2I-CompBench, with attribute binding and spatial relationships calculated using models based on Stable Diffusion 1.4 and 2.1." + ], + "table_footnote": [], + "table_body": "
MethodAttribute Binding (Color ↑)Attribute Binding (Shape ↑)Attribute Binding (Texture ↑)Object Relationship (Spatial ↑)
Stable Diffusion 1.4 Based
Stable Diffusion-v1.4 [32]0.37650.35760.41560.1246
Ours (SD 1.4)0.64580.59830.75390.1613
Stable Diffusion 2.1 Based
Stable Diffusion-v2.1 [32]0.50650.42210.49220.1342
Composable Diffusion [22]0.40630.32990.36450.0800
Structured Diffusion [9]0.49900.42180.49000.1386
Attend-and-Excite [5]0.64000.45170.59630.1455
Ours (SD 2.1)0.67770.62260.78840.1981
", + "bbox": [ + 516, + 508, + 903, + 646 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Quantitative Evaluation: T2I-CompBench. The T2I-CompBench [17] evaluates object presence, spatial relationships, and attribute alignment from text prompts. We used the attribute and spatial configuration datasets provided by T2I-CompBench, averaging results over more than 10 randomly selected seeds for each dataset. Attribute Binding is evaluated with the BLIP-VQA metric [20], which poses separate questions for each attribute-object pair in the prompt. UniDet-based metric [43] is used for spatial relationships. In Table 2, our score shows that the generated images position the specified objects correctly and accurately reflect the given attributes (e.g., color, shape, texture).", + "bbox": [ + 511, + 654, + 906, + 835 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Quantitative Evaluation: CLIP Score and TIFA Score. Following previous work [5, 12, 23], we evaluate our method using CLIP Image-Text and Text-Text Similarity, generating 64 images per prompt and averaging CLIP co", + "bbox": [ + 511, + 839, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "* Advanced Database System Infrastructure(NFEC-2024-11-300458)", + "bbox": [ + 114, + 887, + 434, + 898 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "18406", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/6f6563efb61becaf78ec9b87df5e987ca8f097923a40e1b4f8c2da7eda98e414.jpg", + "image_caption": [ + "Figure 5. Qualitative comparison across the custom prompt, which involves attribute and positional information in text, evaluating previous state-of-the-art training-free T2I methods, Attend&Excite [5], Divide&Bind [21], INITNO [12], CONFORM [23], and ours." + ], + "image_footnote": [], + "bbox": [ + 94, + 88, + 898, + 464 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/587f27950643b7e7cf1ae84dcc5a10e4ec2b0cf6ab7dd4b47e10ae1605e716b0.jpg", + "table_caption": [ + "Table 3. User study evaluating model performance on object synthesis, attribute matching, spatial correctness, and overall fidelity." + ], + "table_footnote": [], + "table_body": "
MethodObject Accuracy (%)Attribute Matching (%)Spatial Correctness (%)Overall Fidelity (%)
Stable Diffusion [32]14.6814.3111.8115.02
Attend-and-Excite [5]16.3416.5013.2016.04
Divide-and-Bind [21]16.3915.6913.0116.52
INITNO [12]15.9816.0313.1515.63
CONFORM [23]15.6416.7912.9114.14
Ours20.9720.6835.9222.65
", + "bbox": [ + 93, + 531, + 480, + 612 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "sine similarity scores across methods. Fig. 7 reports two CLIP Image-Text Similarity metrics: full prompt similarity (cosine similarity between the entire prompt and generated image) and minimum object similarity (minimum similarity between the generated image and each subject prompt) to see the accuracy of each object. We generate captions for the images with BLIP [20] and compute cosine similarity with the original prompt. Our method achieves comparable performance to other leading methods. However, as the reliability of the CLIP score is limited by its \"bag-of-words\" approach [40] (as noted by [5]), we also use the TIFA metric [16] for improved evaluation of complex prompts. TIFA [16] evaluates image faithfulness to the input prompt by generating question-answer pairs using the LLaMA 2 [36] model. Our method performs well compared to other state-of-the-art models.", + "bbox": [ + 88, + 625, + 482, + 864 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Quantitative Evaluation: User Studies. We create 10 custom text prompts with detailed objects, attributes, and spa", + "bbox": [ + 89, + 869, + 482, + 902 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2bb58b2f12a24e2ab41514c50056e828d9ee092f414d32d83dd9037591a9e270.jpg", + "image_caption": [ + "Figure 6. Generated images with complex positioning details in the prompt (Additional results in Supplementary Materials)." + ], + "image_footnote": [], + "bbox": [ + 522, + 507, + 890, + 702 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "tial information, then generate images using random seeds. With 30 participants, our user study evaluates (1) object accuracy, (2) attribute matching, (3) spatial correctness, and (4) overall fidelity. In Table 3, our model shows surpassing results across the existing models of all categories.", + "bbox": [ + 511, + 744, + 906, + 821 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Qualitative Analysis. Fig. 5 provides a qualitative comparison with recent state-of-the-art models [12, 23]. To compare attribute matching and object localization accuracy, we create custom text prompts to evaluate the qualitative performance of our model. Existing methods often misalign", + "bbox": [ + 511, + 824, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "18407", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/aa791de998ec864ce81dec2e93acea1b5a7b3be5e8b45f5e07858bccb7380be6.jpg", + "image_caption": [ + "Figure 7. Comparison of various models on CLIP Image-Text, CLIP Text-Text, and TIFA scores across prompts involving Animal-Animal, Animal-Object, and Object-Object pairs. (The seed is randomly selected.)" + ], + "image_footnote": [], + "bbox": [ + 91, + 88, + 521, + 191 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9e207de35df137f1d0c9f77e81141e5f34d45b2f925b244536d31e132b1062d3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 527, + 89, + 898, + 191 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/74600838616d46d450a3a651c4ebed241cdabee538f6c15666c23f3d848a38fe.jpg", + "image_caption": [ + "Exp.#A0 (SD)" + ], + "image_footnote": [], + "bbox": [ + 94, + 239, + 171, + 287 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6872d1bd4154ec196101ba94c8f7256b1c13f9735dc134ab3d36c0d12397652d.jpg", + "image_caption": [ + "\"A cake to the left of a tie\"", + "Exp.#A1" + ], + "image_footnote": [], + "bbox": [ + 173, + 239, + 250, + 287 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b09aef1acabae910e6c694cd23ee3d9f1b474fd6538fbf633845ecc9fffa204d.jpg", + "image_caption": [ + "Exp.#A2" + ], + "image_footnote": [], + "bbox": [ + 251, + 239, + 326, + 287 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c9d5d8e87e46b948fdfed57cbe639efd94cea39461f3f4c541458e0f8bfba06b.jpg", + "image_caption": [ + "Exp.#A3" + ], + "image_footnote": [], + "bbox": [ + 326, + 239, + 401, + 287 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4fec3d3107de492ac659bc0c9d9ffeaaa20e19b00d450510d890f5e0a20b4508.jpg", + "image_caption": [ + "Exp.#A4(Ours)" + ], + "image_footnote": [], + "bbox": [ + 401, + 239, + 478, + 287 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2e4b2b3ed6e933b26ffa99e29d43124a879e0c7ef07a3a09d245dd144fd10e1e.jpg", + "image_caption": [ + "Exp.#A0 (SD)", + "Figure 8. Comparison of results when applying STO at different timesteps. Experiments are organized as follows: no STO, STO applied from timesteps 19-24, 13-24, 7-24, and 1-24 (Ours). As seen in the images, earlier STO application improves object positioning and reduces overlap, resulting in more accurately positioned objects. (See Table 4 for quantitative evaluation.)" + ], + "image_footnote": [], + "bbox": [ + 94, + 311, + 171, + 369 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/21b28c35622515d03f64baec19d1cf117751fb078e0b62d59c28dca30788160d.jpg", + "image_caption": [ + "\"An orange above a suitcase\"", + "Exp.#A1" + ], + "image_footnote": [], + "bbox": [ + 171, + 311, + 250, + 369 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7571c6ab5d0366785ebe52fc87797ddda207e51302110d75305fe97f5246e394.jpg", + "image_caption": [ + "Exp.#A2" + ], + "image_footnote": [], + "bbox": [ + 250, + 311, + 325, + 369 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c1bf3e168887c321ef9c5e16dfe12bad45b47318d9b3efe286cb4299f805a121.jpg", + "image_caption": [ + "Exp.#A3" + ], + "image_footnote": [], + "bbox": [ + 326, + 311, + 401, + 369 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3c0a3cf3070b282ad151499bb0a32a2681ebcd07be4c75197d9fc5268a077866.jpg", + "image_caption": [ + "Exp.#A4 (Ours)" + ], + "image_footnote": [], + "bbox": [ + 401, + 311, + 478, + 369 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "spatial details, leading to errors such as a rabbit overlapping a backpack instead of being positioned to its right. In contrast, our method, which is focused on precise positioning, demonstrates a strong understanding of spatial information and generates objects accurately. Our method also can generate images that accurately reflect complex spatial information, even when prompts contain intricate positioning details (refer Fig. 6).", + "bbox": [ + 89, + 474, + 483, + 597 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.2. Ablation Studies", + "text_level": 1, + "bbox": [ + 89, + 604, + 254, + 618 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effects of Optimizing through Timestep. In Table 4, we compare STO applied at different timestep ranges (Exp.#A1: 19-24, Exp.#A2: 13-24, Exp.#A3: 7-24, Exp.#A4: 1-24). These experiments show that setting spatial configurations in the early stages is essential not only for positional accuracy but also for achieving higher OA. This improvement increases as spatial relationships are established earlier in the process. Notably, all configurations with STO outperform the baseline (Exp.#A0, without STO), demonstrating a substantial improvement in both object positioning and accuracy. In Fig. 8, tie placement is optimal when spatial cues are applied early, while delayed guidance leads to misplacements or suboptimal synthesis.", + "bbox": [ + 89, + 622, + 483, + 819 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effects of Cost Function. Table 5 presents an ablation study on methods for computing movement costs between a source and target distribution. Exp.#B0 uses Manhattan distance (only horizontal and vertical movement), while Exp.#B1 considers the mass distribution of the attention", + "bbox": [ + 89, + 824, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/f0e6c7f7a885726381c498896adab9009a56d58c1b7bca2f1cca5e0defbbe8c4.jpg", + "table_caption": [ + "Table 4. Ablation study on the impact of applying STO at different timesteps. Exp.#A0 shows baseline results from SD without STO. From Exp.#A1 to Exp.#A4, STO is progressively applied across increasing timestep ranges (19-24, 13-24, 7-24, and 1-24)." + ], + "table_footnote": [], + "table_body": "
# Exp.OA (%)VISOR
uncondcond1234
A029.8618.8162.9846.6020.116.891.63
A146.6237.4880.3874.1145.7322.307.82
A255.6547.9086.0781.6360.3035.8913.81
A359.8553.6289.6083.6167.5344.1219.31
A4 (Ours)61.0157.5894.3985.9369.7149.0125.70
", + "bbox": [ + 517, + 287, + 903, + 377 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/2057f576cb755ae6570a25ff681ddad80e6982e0a43bd3bccc35704ce03a015d.jpg", + "table_caption": [ + "Table 5. Ablation study on computing cost function, highlighting the impact of ST Cost components on OA and VISOR performance. (B0: Manhattan distance; B1: Euclidean distance; B2: B0 + B1; B3: ST cost without overlap prevention.)" + ], + "table_footnote": [], + "table_body": "
# Exp.OA (%)VISOR
uncondcond1234
B027.3217.9165.5142.8419.257.602.00
B147.4830.3563.9365.2037.7114.034.54
B246.6029.2062.9363.9233.9114.434.60
B331.3827.7888.5361.3032.6113.443.82
B4 (Ours)61.0157.5894.3985.9369.7149.0125.70
", + "bbox": [ + 517, + 441, + 901, + 532 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "map using Euclidean distance. Exp.#B2 combines Exp.#B0 and Exp.#B1. Exp.#B3 excludes the overlap prevention component from ST Cost to examine its impact. Our results (Exp.#B4) show that ST Cost outperforms standard OT functions, with overlap prevention proving critical for maintaining OA, despite minimal impact on VISOR.", + "bbox": [ + 511, + 545, + 906, + 638 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 646, + 633, + 661 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we introduced STORM, a framework that dynamically addresses spatial misalignment issues in training-free T2I synthesis. By leveraging the STO framework, which combines OT and an ST Cost function, STORM not only resolves misplaced objects but also tackles missing objects and mismatched attributes. Extensive experiments show that STORM significantly improves spatial alignment and object accuracy, surpassing existing state-of-the-art methods. This advancement establishes STORM as a new benchmark for spatial control in T2I generation, pushing the boundaries of adaptability and efficiency in training-free approaches.", + "bbox": [ + 511, + 667, + 906, + 849 + ], + "page_idx": 7 + }, + { + "type": "page_footnote", + "text": "Acknowledgement. This work was supported in part by IITP RS-2024-00457882 (AI Research Hub Project), IITP 2020-0-01361, NRF RS-2024-00345806, NRF RS-2023-00219019, and RS-2024-00403860 (Korea Basic Science Institute, National Research Facilities and Equipment Center).", + "bbox": [ + 511, + 857, + 905, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "18408", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Aishwarya Agarwal, Srikrishna Karanam, KJ Joseph, Apoorv Saxena, Koustava Goswami, and Balaji Vasan Srinivasan. A-star: Test-time attention segregation and retention for text-to-image synthesis. In ICCV, 2023. 3", + "[2] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Ji-aming Song, Qinsheng Zhang, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, et al. ediff-i: Text-to-image diffusion models with an ensemble of expert denoisers. arXiv preprint arXiv:2211.01324, 2022. 2", + "[3] Agneet Chatterjee, Yiran Luo, Tejas Gokhale, Yezhou Yang, and Chitta Baral. Revision: Rendering tools enable spatial fidelity in vision-language models. In ECCV, 2025. 6", + "[4] Agneet Chatterjee, Gabriela Ben Melech Stan, Estelle Aflalo, Sayak Paul, Dhruba Ghosh, Tejas Gokhale, Ludwig Schmidt, Hannaneh Hajishirzi, Vasudev Lal, Chitta Baral, et al. Getting it right: Improving spatial consistency in text-to-image models. In ECCV, 2025. 6", + "[5] Hila Chefer, Yuval Alaluf, Yael Vinker, Lior Wolf, and Daniel Cohen-Or. Attend-and-excite: Attention-based semantic guidance for text-to-image diffusion models. SIGGRAPH, 2023. 1, 2, 3, 5, 6, 7", + "[6] Minghao Chen, Iro Laina, and Andrea Vedaldi. Training-free layout control with cross-attention guidance. In WACV, 2024. 1, 2, 3, 6", + "[7] Guillaume Couairon, Marlene Careil, Matthieu Cord, Stéphane Lathuiliere, and Jakob Verbeek. Zero-shot spatial layout conditioning for text-to-image diffusion models. In ICCV, 2023. 2, 3", + "[8] Marco Cuturi. Sinkhorn distances: Lightspeed computation of optimal transport. NeurIPS, 2013. 3, 5", + "[9] Weixi Feng, Xuehai He, Tsu-Jui Fu, Varun Jampani, Arjun Akula, Pradyumna Narayana, Sugato Basu, Xin Eric Wang, and William Yang Wang. Training-free structured diffusion guidance for compositional text-to-image synthesis. arXiv preprint arXiv:2212.05032, 2022. 1, 2, 3, 6", + "[10] Tejas Gokhale, Hamid Palangi, Besmira Nushi, Vibhav Vineet, Eric Horvitz, Ece Kamar, Chitta Baral, and Yezhou Yang. Benchmarking spatial relationships in text-to-image generation. arXiv preprint arXiv:2212.10015, 2022. 6", + "[11] Biao Gong, Siteng Huang, Yutong Feng, Shiwei Zhang, Yuyuan Li, and Yu Liu. Check locate rectify: A training-free layout calibration system for text-to-image generation. In CVPR, 2024. 3", + "[12] Xiefan Guo, Jinlin Liu, Miaomiao Cui, Jiankai Li, Hongyu Yang, and Di Huang. Initno: Boosting text-to-image diffusion models via initial noise optimization. In CVPR, 2024, 2, 3, 6, 7", + "[13] Woojung Han, Chanyoung Kim, Dayun Ju, Yumin Shim, and Seong Jae Hwang. Advancing text-driven chest x-ray generation with policy-based reinforcement learning. MICCAI, 2024. 1", + "[14] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 3" + ], + "bbox": [ + 93, + 114, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. NeurIPS, 2020. 2", + "[16] Yushi Hu, Benlin Liu, Jungo Kasai, Yizhong Wang, Mari Ostendorf, Ranjay Krishna, and Noah A Smith. Tifa: Accurate and interpretable text-to-image faithfulness evaluation with question answering. In ICCV, 2023. 6, 7", + "[17] Kaiyi Huang, Kaiyue Sun, Enze Xie, Zhenguo Li, and Xihui Liu. T2i-compbench: A comprehensive benchmark for open-world compositional text-to-image generation. NeurIPS, 2023. 6", + "[18] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. In CVPR, 2017. 2", + "[19] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In CVPR, 2019. 2", + "[20] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, 2022. 6, 7", + "[21] Jiachen Li, Jitesh Jain, and Humphrey Shi. Matting anything. In CVPR, 2024. 2, 3, 5, 6, 7", + "[22] Nan Liu, Shuang Li, Yilun Du, Antonio Torralba, and Joshua B Tenenbaum. Compositional visual generation with composable diffusion models. In ECCV, 2022. 1, 2, 3, 6", + "[23] Tuna Han Salih Meral, Enis Simsar, Federico Tombari, and Pinar Yanardag. Conform: Contrast is all you need for high-fidelity text-to-image diffusion models. In CVPR, 2024. 2, 3, 5, 6, 7", + "[24] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 1, 6", + "[25] Weili Nie, Sifei Liu, Morteza Mardani, Chao Liu, Benjamin Eckart, and Arash Vahdat. Compositional text-to-image generation with dense blob representations. In ICML, 2024. 3", + "[26] Ji-Hoon Park, Yeong-Joon Ju, and Seong-Whan Lee. Explaining generative diffusion models via visual analysis for interpretable decision-making process. Expert Systems with Applications, 2024. 5", + "[27] Taesung Park, Ming-Yu Liu, Ting-Chun Wang, and Jun-Yan Zhu. Semantic image synthesis with spatially-adaptive normalization. In CVPR, 2019. 2", + "[28] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 3, 6", + "[29] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In ICML, 2021. 2", + "[30] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 2" + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "18409", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[31] Royi Rassin, Eran Hirsch, Daniel Glickman, Shauli Ravfogel, Yoav Goldberg, and Gal Chechik. Linguistic binding in diffusion models: Enhancing attribute correspondence through attention map alignment. NeurIPS, 2024. 3", + "[32] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 1, 2, 3, 5, 6, 7", + "[33] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In MICCAI, 2015. 3", + "[34] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. NeurIPS, 2022. 2", + "[35] Wenqiang Sun, Teng Li, Zehong Lin, and Jun Zhang. Spatial-aware latent initialization for controllable image generation. arXiv preprint arXiv:2401.16157, 2024. 3", + "[36] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 7", + "[37] Cédric Villani et al. Optimal transport: old and new. Springer, 338, 2009. 2, 3", + "[38] Qiucheng Wu, Yujuan Liu, Handong Zhao, Trung Bui, Zhe Lin, Yang Zhang, and Shiyu Chang. Harnessing the spatial-temporal attention of diffusion models for high-fidelity text-to-image synthesis. In ICCV, 2023. 2, 3", + "[39] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, et al. Scaling autoregressive models for content-rich text-to-image generation. arXiv preprint arXiv:2206.10789, 2022. 2", + "[40] Mert Yuksekgonul, Federico Bianchi, Pratyusha Kalluri, Dan Jurafsky, and James Zou. When and why vision-language models behave like bags-of-words, and what to do about it? arXiv preprint arXiv:2210.01936, 2022. 7", + "[41] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, 2023. 1", + "[42] Tianjun Zhang, Yi Zhang, Vibhav Vineet, Neel Joshi, and Xin Wang. Controllable text-to-image generation with gpt-4. arXiv preprint arXiv:2305.18583, 2023. 6", + "[43] Xingyi Zhou, Vladlen Koltun, and Philipp Krahenbuhl. Simple multi-dataset detection. In CVPR, 2022. 6" + ], + "bbox": [ + 91, + 92, + 480, + 757 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "18410", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Spatial Transport Optimization by Repositioning Attention Map for Training-Free Text-to-Image Synthesis/68174923-d215-4833-8530-d036bb11b7fd_model.json b/2025/Spatial Transport Optimization by Repositioning Attention Map for Training-Free Text-to-Image Synthesis/68174923-d215-4833-8530-d036bb11b7fd_model.json new file mode 100644 index 0000000000000000000000000000000000000000..c337929c5b8bd9c0a6ce8cbfd662d52eea7ba22b --- /dev/null +++ b/2025/Spatial Transport Optimization by Repositioning Attention Map for Training-Free Text-to-Image Synthesis/68174923-d215-4833-8530-d036bb11b7fd_model.json @@ -0,0 +1,2035 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.174, + 0.131, + 0.825, + 0.177 + ], + "angle": 0, + "content": "Spatial Transport Optimization by Repositioning Attention Map for Training-Free Text-to-Image Synthesis" + }, + { + "type": "text", + "bbox": [ + 0.139, + 0.204, + 0.866, + 0.24 + ], + "angle": 0, + "content": "Woojung Han Yeonkyung Lee Chanyoung Kim Kwanghyun Park Seong Jae Hwang* Yonsei University" + }, + { + "type": "text", + "bbox": [ + 0.17, + 0.242, + 0.825, + 0.258 + ], + "angle": 0, + "content": "{dnwjDDL, yeonkyung.lee, chanyoung, kwanghyun.park, seongjae}@yonsei.ac.kr" + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.298, + 0.905, + 0.525 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.527, + 0.908, + 0.586 + ], + "angle": 0, + "content": "Figure 1. Three main challenges in training-free text-to-image (T2I) generation: (1) missing objects, (2) mismatched attributes, and (3) mislocated objects. While existing approaches address missing objects and mismatched attributes, effectively controlling object positioning remains an open problem. Our proposed model, STORM, introduces a dynamic approach to aligning relative object positions throughout the denoising process, enabling precise spatial control without additional spatial templates. Red underline highlights errors made by SD." + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.596, + 0.327, + 0.612 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.629, + 0.486, + 0.873 + ], + "angle": 0, + "content": "Diffusion-based text-to-image (T2I) models have recently excelled in high-quality image generation, particularly in a training-free manner, enabling cost-effective adaptability and generalization across diverse tasks. However, while the existing methods have been continuously focusing on several challenges such as \"missing objects\" and \"mismatched attributes,\" another critical issue of \"mislocated objects\" remains where generated spatial positions fail to align with text prompts. Surprisingly, ensuring such seemingly basic functionality remains challenging in popular T2I models due to the inherent difficulty of imposing explicit spatial guidance via text forms. To address this, we propose STORM (Spatial Transport Optimization by Repositioning Attention Map), a novel training-free approach for spatially coherent T2I synthesis. STORM employs Spatial Transport Optimization (STO), rooted in optimal transport theory, to" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.598, + 0.909, + 0.734 + ], + "angle": 0, + "content": "dynamically adjust object attention maps for precise spatial adherence, supported by a Spatial Transport (ST) Cost function that enhances spatial understanding. Our analysis shows that integrating spatial awareness is most effective in the early denoising stages, while later phases refine details. Extensive experiments demonstrate that STORM surpasses existing methods, effectively mitigating mislocated objects while improving missing and mismatched attributes, setting a new benchmark for spatial alignment in T2I synthesis." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.747, + 0.645, + 0.761 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Diffusion-based text-to-image (T2I) models exhibit remarkable versatility across various applications, generating high-quality images from textual prompts [13, 22, 24, 32, 41]. Notably, training-free methods offer efficiency, real-time adaptability, lower costs, and broad task generalization [5, 6, 9]. These training-free methods leverage their advantages while addressing three key challenges in fully capturing textual details. In particular, as shown in Fig. 1, these challenges include (1) missing objects, (2) mismatched at" + }, + { + "type": "page_footnote", + "bbox": [ + 0.114, + 0.888, + 0.458, + 0.901 + ], + "angle": 0, + "content": "Project Page: https://micv-yonsei.github.io/storm2025/" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "18401" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.137 + ], + "angle": 0, + "content": "tributes, and (3) mislocated objects. These limitations may lead to images that poorly represent the intended scene or convey misleading information, reducing their reliability." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.14, + 0.483, + 0.397 + ], + "angle": 0, + "content": "While there has been substantial progress in addressing issues like missing objects and mismatched attributes in diffusion-based image synthesis [5, 12, 21, 23], the challenge of mislocated objects remains relatively unexplored despite its critical importance. For example, in Fig. 1, the fifth column shows that the backpack should be below a donut, but Stable Diffusion (SD) [32] generates donut to the left of a backpack. Furthermore, SD consistently generates similar images regardless of spatial cues in the text (e.g., same seed with only positional changes for teddy bear and vase in Fig. 2), indicating limited spatial understanding. This limitation highlights the importance of addressing the problem of mislocated objects and demonstrates the need for more advanced methods of interpreting and aligning with spatial cues. Overcoming these challenges is key to improving T2I model reliability and supporting adaptable applications across diverse settings." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.4, + 0.483, + 0.611 + ], + "angle": 0, + "content": "Existing solutions for mislocated objects often rely on fixed spatial templates or predefined layouts [6, 7, 38], which offer some control but impose rigid constraints and require additional inputs. For instance, these templates may fix an object's position in the image (e.g., far-left corner) instead of offering flexible guidance like placing one object \"to the left\" of another. While effective in structured fields like automated design, these approaches present limitations in more dynamic areas, such as digital art or interactive media, by restricting spatial flexibility. Therefore, instead of relying on restrictive templates or requiring additional inputs, we aim to push the boundaries of T2I models by enabling precise alignment with textual guidance, which is essential to unlocking their full potential across applications." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.614, + 0.483, + 0.902 + ], + "angle": 0, + "content": "To this end, we propose Spatial Transport Optimization by Repositioning Attention Map, STORM, a training-free approach that adjusts relative object positions dynamically, without relying on fixed spatial templates. While SD disregards spatial cues (see Fig. 2), STORM achieves various object positioning with the same model weights, ensuring spatial alignment. To implement STORM, we introduce the Spatial Transport Optimization (STO) framework, which interprets each text token's attention map as a distribution, and leverages Optimal Transport (OT) [37] to efficiently reposition distributions. At an early diffusion timestep, STO adjusts the placement of objects based on their relative positioning (see Fig.3 for attention map changes across the denoising process). While OT considers both distribution and distance to minimize transformation costs, we propose the Spatial Transport Cost (ST Cost) to align with our objectives and guide distribution positioning. This function updates and optimizes the latent vector through backward guidance, following methods from prior work [5, 6]." + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.088, + 0.9, + 0.254 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.255, + 0.907, + 0.34 + ], + "angle": 0, + "content": "Figure 2. Comparison of Spatial Awareness. \\(\\{\\text{position}^*\\}\\) in each prompt denotes the spatial relationship in each column (e.g., \"to the left of\", \"to the right of\", \"above, and below\"). While Stable Diffusion (SD) shows limited spatial awareness by generating similar images regardless of spatial prompts, our model accurately reflects specified positions. (Same seed for all synthesis)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.354, + 0.907, + 0.475 + ], + "angle": 0, + "content": "Our analysis suggests that integrating spatial awareness yields optimal results in the early denoising stages, whereas the later stages are better suited for focusing on fine details. Furthermore, we found that missing objects can be resolved concurrently as mislocated objects are corrected in the early steps. Extensive experiments show that our approach outperforms existing methods, setting a new benchmark for spatial alignment in diffusion models." + }, + { + "type": "text", + "bbox": [ + 0.529, + 0.479, + 0.786, + 0.492 + ], + "angle": 0, + "content": "Our main contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.497, + 0.905, + 0.542 + ], + "angle": 0, + "content": "- We propose STORM, a training-free method that dynamically adjusts relative object positions, achieving accurate spatial alignment without relying on fixed templates." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.543, + 0.905, + 0.603 + ], + "angle": 0, + "content": "- We introduce the Spatial Transport Optimization (STO) framework based on Optimal Transport (OT), which treats attention maps as distributions and repositions objects relative to each other across diffusion timesteps." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.603, + 0.905, + 0.663 + ], + "angle": 0, + "content": "- We develop ST Cost function that penalizes overlaps and progressively enhances spatial accuracy, demonstrating its effectiveness through extensive experiments that set a new benchmark for spatial alignment in T2I models." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.497, + 0.905, + 0.663 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.682, + 0.655, + 0.698 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.704, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Text-to-Image Synthesis. Text-to-image (T2I) synthesis aims to generate visually realistic images that align with given text prompts. Early studies focused on GANs [18, 19, 27] and autoregressive models [29, 39], but diffusion models [15] have taken the lead, demonstrating remarkable performance in T2I synthesis [22, 32]. Despite advances driven by VLMs [32, 34, 39], aligning generated images with text prompts remains a challenge. To address this, some approaches [2, 30, 34] have expanded network architectures, though these methods require retraining and are difficult to adapt to existing frameworks. Alternatively, other research [5, 9, 21, 23] focused on training-free improvement strategies, which aligns closely with our approach." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18402" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.089, + 0.48, + 0.304 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.309, + 0.485, + 0.435 + ], + "angle": 0, + "content": "Figure 3. Comparison of Attention Map Progression During Denoising. Visualization of the attention maps for \"a red bird to the right of a green plant\" throughout the denoising process for both Stable Diffusion (a) and our model (b). While Stable Diffusion struggles to distinctly capture the spatial relationship between the bird and the plant, our model effectively aligns the objects according to the specified spatial cue (\"to the right of\"). The resulting image from our model demonstrates improved spatial accuracy compared to Stable Diffusion." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.442, + 0.484, + 0.73 + ], + "angle": 0, + "content": "Training-free T2I Models. Previous studies [1, 5, 9, 12, 21, 22, 31] have explored various approaches to enhance how text prompts are aligned within text-to-image diffusion models. Composable Diffusion [22] and Structure Diffusion [9] aim to improve compositionality through multi-model or cross-attention methods, yet they still face issues like subject mixing and attribute binding. To address these, Attend-and-Excite [5] dynamically adjusts attention during inference to capture all subjects, while Divide-and-Bind [21] improves attribute binding and compositional accuracy with targeted loss functions. CONFORM [23] uses a contrastive method to separate objects in attention maps while keeping related attributes together, and INITNO [12] further tackles spatial misalignment by improving control over object placement to reduce overlap. Although these studies focus on the missing object issue by reducing object overlap, they often overlook the crucial task of capturing accurate relative positioning and directional relationships between objects." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Training-free T2I Models for Mislocated Objects. The advantages of attention maps for controlling spatial positioning have led to several studies using fixed templates, such as layouts, to improve object placement [6, 7, 11, 25, 35, 38]. Wu et al. [38] use layout predictors to assign pixel regions, while ZestGuide [7] and Chen et al. [6] leverage cross-attention layers for alignment with segmentations or user-defined layouts. However, these methods often require additional layout inputs and lack flexibility for relative positioning. As a result, there is a need for methods that enhance object placement without relying on predefined layouts." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.09, + 0.652, + 0.105 + ], + "angle": 0, + "content": "3. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.108, + 0.907, + 0.29 + ], + "angle": 0, + "content": "Text Conditioned via Cross-Attention. Our approach uses Stable Diffusion (SD) [32], which generates images from text by integrating text conditioning through cross-attention. A pre-trained CLIP encoder [28] converts the text prompt into a conditioning vector \\( c \\), fed into the UNet model [33] via cross-attention layers at different resolutions. At each timestep \\( t \\), the attention map \\( A_{t} \\in \\mathbb{R}^{P \\times P \\times N} \\) assigns probabilities for transferring token information to spatial positions, where \\( P \\) is the spatial dimension of the feature map and \\( N \\) the number of tokens in the text prompt. In our work, we use \\( 16 \\times 16 \\) attention maps as they contain the most semantically meaningful information [14]." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.292, + 0.909, + 0.446 + ], + "angle": 0, + "content": "Optimal Transport with Sinkhorn Algorithm. Optimal Transport (OT) [37] provides a framework for transforming one probability distribution into another while preserving spatial relationships. Given two discrete distributions \\(\\mu = \\{\\mu_u\\}_{u=1}^n\\) and \\(\\nu = \\{\\nu_v\\}_{v=1}^n\\), each with \\(n\\) discrete locations, and a cost matrix \\(\\mathbf{C} \\in \\mathbb{R}^{n \\times n}\\), where \\(\\mathbf{C}_{uv}\\) is the cost of moving distribution from point \\(u\\) to \\(v\\), OT seeks a transport plan \\(\\mathbf{P} \\in \\mathbb{R}^{n \\times n}\\) that minimizes the total cost, \\(\\min_{P \\geq 0} \\sum_{u=1}^{n} \\sum_{v=1}^{n} \\mathbf{C}_{uv} \\mathbf{P}_{uv}\\). For large \\(n\\), the Sinkhorn algorithm [8] applies entropic regularization for a faster, scalable, and differentiable solution:" + }, + { + "type": "equation", + "bbox": [ + 0.529, + 0.451, + 0.905, + 0.486 + ], + "angle": 0, + "content": "\\[\n\\min _ {P \\geq 0} \\quad \\sum_ {u = 1} ^ {n} \\sum_ {v = 1} ^ {n} \\mathbf {C} _ {u v} \\mathbf {P} _ {u v} - \\varepsilon \\sum_ {u = 1} ^ {n} \\sum_ {v = 1} ^ {n} \\mathbf {P} _ {u v} (\\log \\mathbf {P} _ {u v} - 1), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.484, + 0.907, + 0.529 + ], + "angle": 0, + "content": "where \\(\\varepsilon > 0\\) balances transport cost and entropy. We further integrate spatial alignment objectives into this OT-based approach in our framework." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.534, + 0.605, + 0.549 + ], + "angle": 0, + "content": "4. Method" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.552, + 0.907, + 0.644 + ], + "angle": 0, + "content": "We describe the full pipeline of our model, STORM, as shown in Fig. 4. To support the understanding of our Spatial Transport Optimization (STO) framework, we first provide background information in Sec. 4.1. The full details of STO in Sec. 4.2, while Sec. 4.3 covers optimizations for non-spatial text prompts, and Sec. 4.4 details the update process." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.646, + 0.706, + 0.662 + ], + "angle": 0, + "content": "4.1. Background of STO" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.663, + 0.909, + 0.905 + ], + "angle": 0, + "content": "Since objects are generated in regions of high attention within the attention map (see Fig. 3), our goal is to reposition the attention map based on the positional information from the text prompt, using relative objects as a reference. To achieve this, our approach involves three types of attention maps: the attention map of the object to be moved, called the source distribution (e.g., blue area in Fig. 4); the attention map of the relative object, which acts as a reference distribution (e.g., red area in Fig. 4); and the target distribution which represents the intended position for the source distribution to move (e.g., gray area in Fig. 4). The target distribution is an arbitrary distribution defined based on the reference point, which is the centroid of the reference distribution, \\((j_{\\mathrm{ref}}, i_{\\mathrm{ref}})\\). It primarily serves to guide the intended position of the source distribution using reference points without exerting significant direct influence." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18403" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.09, + 0.898, + 0.328 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.33, + 0.908, + 0.43 + ], + "angle": 0, + "content": "Figure 4. Overview pipeline of STORM. Our method leverages Optimal Transport in a training-free manner, allowing the model to accurately reflect relative object positions at each step without additional inputs. Given the prompt \"A car to the left of an elephant\", our method dynamically adjusts the attention maps to induce the specified spatial relationship. The process starts with initial attention maps for the \"car\" and \"elephant\" at time step \\( z_{t} \\). Using the centroids of these attention maps, the Spatial Transport Optimization (STO) computes the losses to correct positional relationships (e.g., ensuring the car is to the left of the elephant). The updated attention map is then used to refine the latent representation \\( z_{t} \\), leading to a final image that adheres to the desired spatial arrangement. The comparison of attention maps (before and after STO) shows improved alignment, effectively placing the car to the left of the elephant as instructed in the prompt." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.441, + 0.465, + 0.457 + ], + "angle": 0, + "content": "4.2. Spatial Transport Optimization Framework" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.459, + 0.483, + 0.716 + ], + "angle": 0, + "content": "In this section, we detail our Spatial Transport Optimization (STO) framework, which dynamically repositions attention maps for accurate object alignment without modifying model weights. This framework consists of two main components: a cost function and a transport plan. The cost function directs the distribution shift from the source in the intended direction, allowing the attention map to adjust positionally according to spatial relationships. Although we build on standard cost functions that measure distance between locations, our Spatial Transport Cost (ST Cost) incorporates additional objectives adapted to our specific spatial alignment goals. The transport plan, which optimally moves the distribution between points, is implemented using the Sinkhorn algorithm. Beginning with an initial attention map, STO iteratively updates the latent vector by applying STO-derived losses, gradually guiding the distribution into the target alignment." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.72, + 0.4, + 0.735 + ], + "angle": 0, + "content": "4.2.1 Construction of Target Distributions" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Our approach focuses on relative positioning adjustments at each timestep, using an arbitrary target distribution (e.g., gray area in Fig. 4) to guide the source distribution based on the reference. Depending on the text-specified direction (e.g., horizontal or vertical), only the relevant reference point of \\((j_{\\mathrm{ref}},i_{\\mathrm{ref}})\\) is considered: for 'left' or 'right' positioning, only the horizontal dimension of the reference point \\(j_{\\mathrm{ref}}\\) is used, while for 'above' or 'below' positioning, only the vertical dimension \\(i_{\\mathrm{ref}}\\) is considered, allowing the flexibility in the opposite dimension. For example, in Fig. 4a, with the car attention map as a source distribution and the" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.442, + 0.907, + 0.518 + ], + "angle": 0, + "content": "elephant attention map as a reference distribution, the target distribution is placed to the left of a elephant's distribution. This approach similarly applies to another source. Supplementary Materials provide details on computing the target distribution and reference point." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.524, + 0.794, + 0.538 + ], + "angle": 0, + "content": "4.2.2 Spatial Transport Cost Function" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.539, + 0.907, + 0.674 + ], + "angle": 0, + "content": "In the STO framework, we aim to find an optimal transport plan that minimizes the cost of transferring the attention map. The cost function measures the expense of moving distribution mass by computing distances between points, where a higher value indicates less desirable locations. While our overall cost function merges the standard OT loss with an additional term (detailed in the supplementary material), here we focus on our newly introduced ST Cost, which is built on two core principles:" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.674, + 0.909, + 0.901 + ], + "angle": 0, + "content": "(I) Positional Cost: We developed a cost function to guide the distribution to its intended position relative to the reference point. In simple terms, when the source distribution is intended to be on the left side of the reference point, it has a low cost when positioned on the left and a high cost when positioned on the right. To quantify this directional positioning, we define a set of positional \\(\\delta\\) values, representing the distances between the current point and the reference point in each of the four directions. Specifically, the \\(\\delta\\) values are defined as follows: \\(\\delta_{ij}^{\\leftarrow} = j_{\\mathrm{ref}} - j,\\delta_{ij}^{\\rightarrow} = j - j_{\\mathrm{ref}},\\delta_{ij}^{\\uparrow} = i_{\\mathrm{ref}} - i,\\delta_{ij}^{\\downarrow} = i - i_{\\mathrm{ref}}\\) with \\(j\\) and \\(i\\) representing the current position values, and the arrows indicate the desired and restricted direction for positioning the source attention map relative to the reference point. When constructing the overall cost function that incorporates all directional \\(\\delta\\) values, each direction is categorized as either de" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18404" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.091, + 0.482, + 0.206 + ], + "angle": 0, + "content": "sired or restricted. The desired direction \\((\\delta^{\\mathrm{des}})\\) corresponds to the intended movement of the source distribution, while the restricted direction \\((\\delta^{\\mathrm{res}})\\) represents the opposite or undesired direction. For instance, if the source distribution should move to the left, \\(\\delta_{ij}^{\\leftarrow}\\) is marked as the desired direction \\((\\delta^{\\mathrm{des}})\\), while \\(\\delta_{ij}^{\\rightarrow}\\) is treated as the restricted direction \\((\\delta^{\\mathrm{res}})\\). The combined cost function, accounting for both desired and restricted positions, can be expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.101, + 0.21, + 0.483, + 0.241 + ], + "angle": 0, + "content": "\\[\n\\Delta_ {i j} \\left(\\delta_ {i j} ^ {\\mathrm {d e s}}, \\delta_ {i j} ^ {\\mathrm {r e s}}\\right) = \\frac {1}{\\omega \\left(\\delta_ {i j} ^ {\\mathrm {d e s}} + \\epsilon\\right)} \\mathbb {1} _ {+} \\left[ \\delta_ {i j} ^ {\\mathrm {d e s}} \\right] + \\omega \\left(\\delta_ {i j} ^ {\\mathrm {r e s}} + \\epsilon\\right) \\mathbb {1} _ {+} \\left[ \\delta_ {i j} ^ {\\mathrm {r e s}} \\right], \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.248, + 0.484, + 0.369 + ], + "angle": 0, + "content": "where \\(\\omega (\\cdot) > 1\\) is a progressive adaptive weight that controls the alignment importance across timesteps, and \\(\\epsilon\\) is a stabilizing factor. \\(\\mathbb{1}_{+}\\) is an indicator function that equals 1 when the input is positive and 0 otherwise, allowing selective application of computation based on positivity. As the source distribution aligns more closely with the desired direction, the \\(\\Delta\\) value approaches zero, indicating minimal cost for correctly aligned positioning (see Example below)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.374, + 0.484, + 0.554 + ], + "angle": 0, + "content": "(II) Non-overlap Cost: In this process, it is crucial to prevent the source distribution from overlapping with the reference distribution during its movement. This is one of the primary issues in addressing the missing object problem in SD, where avoiding overlap is essential for ensuring that each object is distinctly represented. To achieve this, we simply embed the reference distribution directly into the cost function, guiding objects to occupy separate locations. Specifically, with \\( A_{ij} \\) representing the attention weight of the reference distribution, we incorporated this distribution into the cost function, assigning high costs to regions occupied by the reference distribution and low costs elsewhere." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.555, + 0.483, + 0.582 + ], + "angle": 0, + "content": "Combining this principle with our core ideas, our cost function can be defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.209, + 0.593, + 0.482, + 0.611 + ], + "angle": 0, + "content": "\\[\nC _ {i j} = A _ {i j} \\Delta_ {i j} \\left(\\delta_ {i j} ^ {\\mathrm {d e s}}, \\delta_ {i j} ^ {\\mathrm {r e s}}\\right). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.618, + 0.484, + 0.83 + ], + "angle": 0, + "content": "Example. In Fig. 4, the car should be left of the elephant, and the elephant to the right of the car. The cost functions for each of them are represented by \\( C_{ij}^{\\leftarrow} = A_{ij}\\Delta_{ij}(\\delta_{ij}^{\\leftarrow},\\delta_{ij}^{\\rightarrow}) \\), \\( C_{ij}^{\\rightarrow} = A_{ij}\\Delta_{ij}(\\delta_{ij}^{\\rightarrow},\\delta_{ij}^{\\leftarrow}) \\), respectively. Taking the car as an example, it needs to be positioned to the left of the elephant relative to the elephant's centroid point. In more detail, if the current point is positioned to the left of the centroid, the \\( \\delta_{ij}^{\\mathrm{des}} \\) is activated; if positioned to the right of the centroid, the \\( \\delta_{ij}^{\\mathrm{res}} \\) is activated. Additionally, as the car moves to the left of the elephant, it should not overlap the elephant it passes along the way. The elephant's attention map is used in the cost function, guiding the car to avoid overlapping with the elephant's area. The resulting \\( C_{ij} \\) is utilized as the cost function for STO." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.841, + 0.443, + 0.855 + ], + "angle": 0, + "content": "4.2.3 Sinkhorn Algorithm-based Transport Plan" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.483, + 0.901 + ], + "angle": 0, + "content": "After combining the standard OT approach with ST components (see Supplementary Material for the complete formulation), we compute the transport plan using the Sinkhorn" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.212 + ], + "angle": 0, + "content": "algorithm [8], which applies entropic regularization for computational efficiency and stability, especially for high-dimensional attention maps. The regularized OT problem is formulated as in Eq. (1) using target distribution calculated in Sec. 4.2.1. The Sinkhorn algorithm iteratively updates transport plan \\(\\mathbf{P}\\) to meet the marginal constraints of source and target distributions, applying the cost function bidirectionally for both objects." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.217, + 0.859, + 0.232 + ], + "angle": 0, + "content": "4.3. Optimization for Objects and Attributes" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.233, + 0.905, + 0.459 + ], + "angle": 0, + "content": "Our method also extends to cases where the text lacks spatial information. (i) Object-level Optimization: We handle cases where the text prompt specifies objects without explicit spatial guidance by randomly placing the target distribution of each object. This flexible positioning encourages diverse object placement when spatial relationships are not provided. Therefore, the cost function is simplified to focus on solely on preventing overlap, expressed as \\( C_{ij} = A_{ij} \\otimes \\mathbf{1}_N \\), where \\( A \\) is the reference attention map. (ii) Attribute Matching Optimization: The attention map for an object and its corresponding attribute should be identical, reflecting the same attention focus. Therefore, we used the standard OT cost function, calculating Euclidean distance between corresponding positions across the object, and attribute attention maps to quantify differences." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.465, + 0.809, + 0.481 + ], + "angle": 0, + "content": "4.4. Optimization and Update Process" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.483, + 0.905, + 0.573 + ], + "angle": 0, + "content": "We update the latent vector using STO-based losses to guide the spatial alignment of the attention maps in a differentiable manner. At each timestep, we compute the OT loss \\(\\mathcal{L}\\) and update the latent vector \\(z_{t}\\) as \\(z_{t}^{\\prime}\\gets z_{t} - \\alpha_{t}\\cdot \\nabla z_{t}\\mathcal{L}\\) where \\(\\alpha_{t}\\) is the step size. After the update, a forward pass computes \\(z_{t - 1}\\) for the next denoising step." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.576, + 0.906, + 0.743 + ], + "angle": 0, + "content": "Remark. As shown in Fig.8, addressing the mislocated object problem is most effective in early timesteps, which, according to related work[26], establish the overall image structure while later timesteps refine finer details. Based on this, STORM establishes object positioning and spatial configurations early in the denoising process, minimizing the risk of misplacements and overlaps. Furthermore, as shown in the bottom row of Fig. 8, solving the mislocated object problem before addressing missing objects can naturally resolve the latter, highlighting the importance of prioritizing mislocated objects as a critical step in the synthesis process." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.759, + 0.645, + 0.775 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Implementation Details. Following methods established in prior research [5, 21, 23], we conducted a comprehensive evaluation of our model. We used the official Stable Diffusion models [32] v1.4 and v2.1, configuring the denoising timestep to 50, and the guidance scale to 7.5 in line with common practices in the field. To smooth the cross-attention map, a Gaussian filter with a kernel size of 3 and a standard deviation of 0.5 was applied. Optimization" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "18405" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.907, + 0.173 + ], + "angle": 0, + "content": "Table 1. Performance comparison between different models on VISOR (\\%) and Object Accuracy (OA) (\\%) metrics, based on Stable Diffusion 1.4 and Stable Diffusion 2.1 [32]. For entries marked with \\(\\dagger\\), official values were unavailable; thus, we directly calculated these values. Bold values indicate the best scores, while underlined values indicate the second-best scores among the fair comparison. Notes on VISOR. VISOR cond: spatial accuracy using the results in which both objects are sufficiently generated; VISOR uncond: spatial accuracy using all results, including results with failed object generations; VISOR \\(1/2/3/4\\): ratios of one, two, three, or four spatially accurate results out of four seed results, respectively. (e.g., \\(\\mathrm{VISOR}_1\\): % of results with exactly one spatially correct result out of four generated results)" + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.173, + 0.903, + 0.457 + ], + "angle": 0, + "content": "
ModelVenueExtra TrainingExtra InputOA (%)VISOR (%)
uncondcond1234
Stable Diffusion 1.4 Based
SD 1.4 [32]-X29.8618.8162.9846.6020.116.891.63
SD 1.4 + CDM [22]ECCV22✓X23.2714.9964.4139.4414.564.841.12
GLIDE [24]ICML22✓X3.361.9859.066.721.020.170.03
GLIDE + CDM [22]ECCV22✓X10.176.4363.2120.074.690.830.11
Control-GPT [42]arXiv23✓✓48.3344.1765.9769.8051.2035.6720.48
Layout Guidance [6]WACV24X✓40.0138.8095.95----
REVISION [3]ECCV24X✓53.9652.7197.6977.7961.0244.9027.15
Structure Diffusion [9]ICLR23XX28.6517.8762.3644.7018.736.571.46
Attend-and-Excite [5]SIGGRAPH23XX42.0725.7561.2149.2919.334.560.08
Divide-and-Bind†[21]BMVC24XX46.0331.6268.7064.7237.8218.645.30
INITNO†[12]CVPR24XX60.4035.1858.2471.2042.7120.096.72
CONFORM†[23]CVPR24XX60.7338.4862.3373.0145.8225.579.52
Ours (SD 1.4)XX61.0157.5894.3985.9369.7149.0125.70
Stable Diffusion 2.1 Based
SD 2.1 [32]-X47.8330.2563.2464.4235.7416.134.70
SPRIGHT [4]ECCV24✓X60.6842.2371.2471.7851.8833.0916.15
REVISION [3]ECCV24X✓48.2647.1197.6176.0755.7537.1019.53
Ours (SD 2.1)XX62.5559.3594.8888.3471.7552.0325.42
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.47, + 0.483, + 0.519 + ], + "angle": 0, + "content": "was performed iteratively with forward updates at specified timesteps (\\(t \\in 5, 10, 15, 20\\)), stopping at \\(t = 25\\) to enhance quality. All experiments use NVIDIA RTX A6000 Ada*." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.531, + 0.273, + 0.545 + ], + "angle": 0, + "content": "5.1. Evaluation Results" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.549, + 0.483, + 0.746 + ], + "angle": 0, + "content": "We evaluated our method against existing training-free T2I models using both qualitative and quantitative measures, focusing on two key aspects. (i) Spatial Accuracy: To measure alignment with spatial configurations specified in prompts, we used the VISOR [10] and the T2I-CompBench [17], both provide a comprehensive assessment of spatial relationships captured by the model. (ii) Object and Attribute Consistency: We assessed object and attribute representation using T2I-CompBench in combination with CLIP-space distance [28] and TIFA score [16], following established methods [5, 21, 23]. We provide detailed descriptions of each evaluation metric and datasets in the Supplementary Materials for additional clarity." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.75, + 0.483, + 0.871 + ], + "angle": 0, + "content": "Quantitative Evaluation: VISOR. To evaluate spatial accuracy, we use the VISOR [10] to measure the ability of a model to position objects based on spatial cues (e.g., above) in text prompts. VISOR benchmark contains over 25K prompts describing spatial relationships between objects. Notably, as seen in Table 1, the increase in \\(\\mathrm{VISOR}_4\\) demonstrates the model's reliability in generating spatially accurate images across prompts and seeds." + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.469, + 0.907, + 0.509 + ], + "angle": 0, + "content": "Table 2. Comparison of methods on T2I-CompBench, with attribute binding and spatial relationships calculated using models based on Stable Diffusion 1.4 and 2.1." + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.51, + 0.905, + 0.647 + ], + "angle": 0, + "content": "
MethodAttribute Binding (Color ↑)Attribute Binding (Shape ↑)Attribute Binding (Texture ↑)Object Relationship (Spatial ↑)
Stable Diffusion 1.4 Based
Stable Diffusion-v1.4 [32]0.37650.35760.41560.1246
Ours (SD 1.4)0.64580.59830.75390.1613
Stable Diffusion 2.1 Based
Stable Diffusion-v2.1 [32]0.50650.42210.49220.1342
Composable Diffusion [22]0.40630.32990.36450.0800
Structured Diffusion [9]0.49900.42180.49000.1386
Attend-and-Excite [5]0.64000.45170.59630.1455
Ours (SD 2.1)0.67770.62260.78840.1981
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.655, + 0.907, + 0.837 + ], + "angle": 0, + "content": "Quantitative Evaluation: T2I-CompBench. The T2I-CompBench [17] evaluates object presence, spatial relationships, and attribute alignment from text prompts. We used the attribute and spatial configuration datasets provided by T2I-CompBench, averaging results over more than 10 randomly selected seeds for each dataset. Attribute Binding is evaluated with the BLIP-VQA metric [20], which poses separate questions for each attribute-object pair in the prompt. UniDet-based metric [43] is used for spatial relationships. In Table 2, our score shows that the generated images position the specified objects correctly and accurately reflect the given attributes (e.g., color, shape, texture)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Quantitative Evaluation: CLIP Score and TIFA Score. Following previous work [5, 12, 23], we evaluate our method using CLIP Image-Text and Text-Text Similarity, generating 64 images per prompt and averaging CLIP co" + }, + { + "type": "page_footnote", + "bbox": [ + 0.115, + 0.888, + 0.436, + 0.9 + ], + "angle": 0, + "content": "* Advanced Database System Infrastructure(NFEC-2024-11-300458)" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18406" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.089, + 0.9, + 0.465 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.467, + 0.907, + 0.496 + ], + "angle": 0, + "content": "Figure 5. Qualitative comparison across the custom prompt, which involves attribute and positional information in text, evaluating previous state-of-the-art training-free T2I methods, Attend&Excite [5], Divide&Bind [21], INITNO [12], CONFORM [23], and ours." + }, + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.503, + 0.483, + 0.531 + ], + "angle": 0, + "content": "Table 3. User study evaluating model performance on object synthesis, attribute matching, spatial correctness, and overall fidelity." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.532, + 0.481, + 0.613 + ], + "angle": 0, + "content": "
MethodObject Accuracy (%)Attribute Matching (%)Spatial Correctness (%)Overall Fidelity (%)
Stable Diffusion [32]14.6814.3111.8115.02
Attend-and-Excite [5]16.3416.5013.2016.04
Divide-and-Bind [21]16.3915.6913.0116.52
INITNO [12]15.9816.0313.1515.63
CONFORM [23]15.6416.7912.9114.14
Ours20.9720.6835.9222.65
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.625, + 0.483, + 0.866 + ], + "angle": 0, + "content": "sine similarity scores across methods. Fig. 7 reports two CLIP Image-Text Similarity metrics: full prompt similarity (cosine similarity between the entire prompt and generated image) and minimum object similarity (minimum similarity between the generated image and each subject prompt) to see the accuracy of each object. We generate captions for the images with BLIP [20] and compute cosine similarity with the original prompt. Our method achieves comparable performance to other leading methods. However, as the reliability of the CLIP score is limited by its \"bag-of-words\" approach [40] (as noted by [5]), we also use the TIFA metric [16] for improved evaluation of complex prompts. TIFA [16] evaluates image faithfulness to the input prompt by generating question-answer pairs using the LLaMA 2 [36] model. Our method performs well compared to other state-of-the-art models." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.903 + ], + "angle": 0, + "content": "Quantitative Evaluation: User Studies. We create 10 custom text prompts with detailed objects, attributes, and spa" + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.508, + 0.892, + 0.703 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.705, + 0.907, + 0.734 + ], + "angle": 0, + "content": "Figure 6. Generated images with complex positioning details in the prompt (Additional results in Supplementary Materials)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.745, + 0.907, + 0.822 + ], + "angle": 0, + "content": "tial information, then generate images using random seeds. With 30 participants, our user study evaluates (1) object accuracy, (2) attribute matching, (3) spatial correctness, and (4) overall fidelity. In Table 3, our model shows surpassing results across the existing models of all categories." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Qualitative Analysis. Fig. 5 provides a qualitative comparison with recent state-of-the-art models [12, 23]. To compare attribute matching and object localization accuracy, we create custom text prompts to evaluate the qualitative performance of our model. Existing methods often misalign" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18407" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.089, + 0.522, + 0.193 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.528, + 0.09, + 0.9, + 0.192 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.195, + 0.908, + 0.223 + ], + "angle": 0, + "content": "Figure 7. Comparison of various models on CLIP Image-Text, CLIP Text-Text, and TIFA scores across prompts involving Animal-Animal, Animal-Object, and Object-Object pairs. (The seed is randomly selected.)" + }, + { + "type": "image_caption", + "bbox": [ + 0.186, + 0.23, + 0.388, + 0.24 + ], + "angle": 0, + "content": "\"A cake to the left of a tie\"" + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.24, + 0.172, + 0.289 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.098, + 0.289, + 0.167, + 0.299 + ], + "angle": 0, + "content": "Exp.#A0 (SD)" + }, + { + "type": "image", + "bbox": [ + 0.174, + 0.24, + 0.25, + 0.289 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.188, + 0.289, + 0.232, + 0.299 + ], + "angle": 0, + "content": "Exp.#A1" + }, + { + "type": "image", + "bbox": [ + 0.252, + 0.24, + 0.327, + 0.289 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.265, + 0.289, + 0.31, + 0.299 + ], + "angle": 0, + "content": "Exp.#A2" + }, + { + "type": "image", + "bbox": [ + 0.328, + 0.24, + 0.402, + 0.289 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.342, + 0.289, + 0.386, + 0.299 + ], + "angle": 0, + "content": "Exp.#A3" + }, + { + "type": "image", + "bbox": [ + 0.403, + 0.24, + 0.48, + 0.289 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.404, + 0.289, + 0.479, + 0.299 + ], + "angle": 0, + "content": "Exp.#A4(Ours)" + }, + { + "type": "image_caption", + "bbox": [ + 0.189, + 0.301, + 0.385, + 0.313 + ], + "angle": 0, + "content": "\"An orange above a suitcase\"" + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.313, + 0.172, + 0.37 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.098, + 0.37, + 0.168, + 0.38 + ], + "angle": 0, + "content": "Exp.#A0 (SD)" + }, + { + "type": "image", + "bbox": [ + 0.173, + 0.313, + 0.25, + 0.37 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.188, + 0.37, + 0.232, + 0.38 + ], + "angle": 0, + "content": "Exp.#A1" + }, + { + "type": "image", + "bbox": [ + 0.251, + 0.313, + 0.326, + 0.37 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.266, + 0.37, + 0.31, + 0.38 + ], + "angle": 0, + "content": "Exp.#A2" + }, + { + "type": "image", + "bbox": [ + 0.327, + 0.313, + 0.402, + 0.37 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.342, + 0.37, + 0.386, + 0.38 + ], + "angle": 0, + "content": "Exp.#A3" + }, + { + "type": "image", + "bbox": [ + 0.403, + 0.313, + 0.48, + 0.37 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.404, + 0.37, + 0.479, + 0.38 + ], + "angle": 0, + "content": "Exp.#A4 (Ours)" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.383, + 0.484, + 0.466 + ], + "angle": 0, + "content": "Figure 8. Comparison of results when applying STO at different timesteps. Experiments are organized as follows: no STO, STO applied from timesteps 19-24, 13-24, 7-24, and 1-24 (Ours). As seen in the images, earlier STO application improves object positioning and reduces overlap, resulting in more accurately positioned objects. (See Table 4 for quantitative evaluation.)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.476, + 0.484, + 0.598 + ], + "angle": 0, + "content": "spatial details, leading to errors such as a rabbit overlapping a backpack instead of being positioned to its right. In contrast, our method, which is focused on precise positioning, demonstrates a strong understanding of spatial information and generates objects accurately. Our method also can generate images that accurately reflect complex spatial information, even when prompts contain intricate positioning details (refer Fig. 6)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.605, + 0.255, + 0.619 + ], + "angle": 0, + "content": "5.2. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.623, + 0.484, + 0.82 + ], + "angle": 0, + "content": "Effects of Optimizing through Timestep. In Table 4, we compare STO applied at different timestep ranges (Exp.#A1: 19-24, Exp.#A2: 13-24, Exp.#A3: 7-24, Exp.#A4: 1-24). These experiments show that setting spatial configurations in the early stages is essential not only for positional accuracy but also for achieving higher OA. This improvement increases as spatial relationships are established earlier in the process. Notably, all configurations with STO outperform the baseline (Exp.#A0, without STO), demonstrating a substantial improvement in both object positioning and accuracy. In Fig. 8, tie placement is optimal when spatial cues are applied early, while delayed guidance leads to misplacements or suboptimal synthesis." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Effects of Cost Function. Table 5 presents an ablation study on methods for computing movement costs between a source and target distribution. Exp.#B0 uses Manhattan distance (only horizontal and vertical movement), while Exp.#B1 considers the mass distribution of the attention" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.231, + 0.906, + 0.287 + ], + "angle": 0, + "content": "Table 4. Ablation study on the impact of applying STO at different timesteps. Exp.#A0 shows baseline results from SD without STO. From Exp.#A1 to Exp.#A4, STO is progressively applied across increasing timestep ranges (19-24, 13-24, 7-24, and 1-24)." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.289, + 0.905, + 0.378 + ], + "angle": 0, + "content": "
# Exp.OA (%)VISOR
uncondcond1234
A029.8618.8162.9846.6020.116.891.63
A146.6237.4880.3874.1145.7322.307.82
A255.6547.9086.0781.6360.3035.8913.81
A359.8553.6289.6083.6167.5344.1219.31
A4 (Ours)61.0157.5894.3985.9369.7149.0125.70
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.386, + 0.906, + 0.442 + ], + "angle": 0, + "content": "Table 5. Ablation study on computing cost function, highlighting the impact of ST Cost components on OA and VISOR performance. (B0: Manhattan distance; B1: Euclidean distance; B2: B0 + B1; B3: ST cost without overlap prevention.)" + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.443, + 0.903, + 0.533 + ], + "angle": 0, + "content": "
# Exp.OA (%)VISOR
uncondcond1234
B027.3217.9165.5142.8419.257.602.00
B147.4830.3563.9365.2037.7114.034.54
B246.6029.2062.9363.9233.9114.434.60
B331.3827.7888.5361.3032.6113.443.82
B4 (Ours)61.0157.5894.3985.9369.7149.0125.70
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.546, + 0.907, + 0.639 + ], + "angle": 0, + "content": "map using Euclidean distance. Exp.#B2 combines Exp.#B0 and Exp.#B1. Exp.#B3 excludes the overlap prevention component from ST Cost to examine its impact. Our results (Exp.#B4) show that ST Cost outperforms standard OT functions, with overlap prevention proving critical for maintaining OA, despite minimal impact on VISOR." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.647, + 0.634, + 0.662 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.668, + 0.907, + 0.85 + ], + "angle": 0, + "content": "In this work, we introduced STORM, a framework that dynamically addresses spatial misalignment issues in training-free T2I synthesis. By leveraging the STO framework, which combines OT and an ST Cost function, STORM not only resolves misplaced objects but also tackles missing objects and mismatched attributes. Extensive experiments show that STORM significantly improves spatial alignment and object accuracy, surpassing existing state-of-the-art methods. This advancement establishes STORM as a new benchmark for spatial control in T2I generation, pushing the boundaries of adaptability and efficiency in training-free approaches." + }, + { + "type": "page_footnote", + "bbox": [ + 0.512, + 0.858, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Acknowledgement. This work was supported in part by IITP RS-2024-00457882 (AI Research Hub Project), IITP 2020-0-01361, NRF RS-2024-00345806, NRF RS-2023-00219019, and RS-2024-00403860 (Korea Basic Science Institute, National Research Facilities and Equipment Center)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18408" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.171 + ], + "angle": 0, + "content": "[1] Aishwarya Agarwal, Srikrishna Karanam, KJ Joseph, Apoorv Saxena, Koustava Goswami, and Balaji Vasan Srinivasan. A-star: Test-time attention segregation and retention for text-to-image synthesis. In ICCV, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.173, + 0.484, + 0.243 + ], + "angle": 0, + "content": "[2] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Ji-aming Song, Qinsheng Zhang, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, et al. ediff-i: Text-to-image diffusion models with an ensemble of expert denoisers. arXiv preprint arXiv:2211.01324, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.245, + 0.483, + 0.285 + ], + "angle": 0, + "content": "[3] Agneet Chatterjee, Yiran Luo, Tejas Gokhale, Yezhou Yang, and Chitta Baral. Revision: Rendering tools enable spatial fidelity in vision-language models. In ECCV, 2025. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.287, + 0.483, + 0.356 + ], + "angle": 0, + "content": "[4] Agneet Chatterjee, Gabriela Ben Melech Stan, Estelle Aflalo, Sayak Paul, Dhruba Ghosh, Tejas Gokhale, Ludwig Schmidt, Hannaneh Hajishirzi, Vasudev Lal, Chitta Baral, et al. Getting it right: Improving spatial consistency in text-to-image models. In ECCV, 2025. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.359, + 0.483, + 0.413 + ], + "angle": 0, + "content": "[5] Hila Chefer, Yuval Alaluf, Yael Vinker, Lior Wolf, and Daniel Cohen-Or. Attend-and-excite: Attention-based semantic guidance for text-to-image diffusion models. SIGGRAPH, 2023. 1, 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.415, + 0.483, + 0.455 + ], + "angle": 0, + "content": "[6] Minghao Chen, Iro Laina, and Andrea Vedaldi. Training-free layout control with cross-attention guidance. In WACV, 2024. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.459, + 0.483, + 0.514 + ], + "angle": 0, + "content": "[7] Guillaume Couairon, Marlene Careil, Matthieu Cord, Stéphane Lathuiliere, and Jakob Verbeek. Zero-shot spatial layout conditioning for text-to-image diffusion models. In ICCV, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.517, + 0.483, + 0.544 + ], + "angle": 0, + "content": "[8] Marco Cuturi. Sinkhorn distances: Lightspeed computation of optimal transport. NeurIPS, 2013. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.546, + 0.483, + 0.614 + ], + "angle": 0, + "content": "[9] Weixi Feng, Xuehai He, Tsu-Jui Fu, Varun Jampani, Arjun Akula, Pradyumna Narayana, Sugato Basu, Xin Eric Wang, and William Yang Wang. Training-free structured diffusion guidance for compositional text-to-image synthesis. arXiv preprint arXiv:2212.05032, 2022. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.617, + 0.483, + 0.672 + ], + "angle": 0, + "content": "[10] Tejas Gokhale, Hamid Palangi, Besmira Nushi, Vibhav Vineet, Eric Horvitz, Ece Kamar, Chitta Baral, and Yezhou Yang. Benchmarking spatial relationships in text-to-image generation. arXiv preprint arXiv:2212.10015, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.674, + 0.483, + 0.728 + ], + "angle": 0, + "content": "[11] Biao Gong, Siteng Huang, Yutong Feng, Shiwei Zhang, Yuyuan Li, and Yu Liu. Check locate rectify: A training-free layout calibration system for text-to-image generation. In CVPR, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.731, + 0.483, + 0.786 + ], + "angle": 0, + "content": "[12] Xiefan Guo, Jinlin Liu, Miaomiao Cui, Jiankai Li, Hongyu Yang, and Di Huang. Initno: Boosting text-to-image diffusion models via initial noise optimization. In CVPR, 2024, 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.789, + 0.483, + 0.842 + ], + "angle": 0, + "content": "[13] Woojung Han, Chanyoung Kim, Dayun Ju, Yumin Shim, and Seong Jae Hwang. Advancing text-driven chest x-ray generation with policy-based reinforcement learning. MICCAI, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.483, + 0.9 + ], + "angle": 0, + "content": "[14] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 3" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "[15] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. NeurIPS, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.123, + 0.905, + 0.178 + ], + "angle": 0, + "content": "[16] Yushi Hu, Benlin Liu, Jungo Kasai, Yizhong Wang, Mari Ostendorf, Ranjay Krishna, and Noah A Smith. Tifa: Accurate and interpretable text-to-image faithfulness evaluation with question answering. In ICCV, 2023. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.181, + 0.905, + 0.234 + ], + "angle": 0, + "content": "[17] Kaiyi Huang, Kaiyue Sun, Enze Xie, Zhenguo Li, and Xihui Liu. T2i-compbench: A comprehensive benchmark for open-world compositional text-to-image generation. NeurIPS, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.238, + 0.905, + 0.279 + ], + "angle": 0, + "content": "[18] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. In CVPR, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.282, + 0.905, + 0.322 + ], + "angle": 0, + "content": "[19] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In CVPR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.326, + 0.905, + 0.38 + ], + "angle": 0, + "content": "[20] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, 2022. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.383, + 0.904, + 0.41 + ], + "angle": 0, + "content": "[21] Jiachen Li, Jitesh Jain, and Humphrey Shi. Matting anything. In CVPR, 2024. 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.413, + 0.905, + 0.453 + ], + "angle": 0, + "content": "[22] Nan Liu, Shuang Li, Yilun Du, Antonio Torralba, and Joshua B Tenenbaum. Compositional visual generation with composable diffusion models. In ECCV, 2022. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.457, + 0.905, + 0.511 + ], + "angle": 0, + "content": "[23] Tuna Han Salih Meral, Enis Simsar, Federico Tombari, and Pinar Yanardag. Conform: Contrast is all you need for high-fidelity text-to-image diffusion models. In CVPR, 2024. 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.514, + 0.905, + 0.582 + ], + "angle": 0, + "content": "[24] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.585, + 0.905, + 0.627 + ], + "angle": 0, + "content": "[25] Weili Nie, Sifei Liu, Morteza Mardani, Chao Liu, Benjamin Eckart, and Arash Vahdat. Compositional text-to-image generation with dense blob representations. In ICML, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.63, + 0.905, + 0.685 + ], + "angle": 0, + "content": "[26] Ji-Hoon Park, Yeong-Joon Ju, and Seong-Whan Lee. Explaining generative diffusion models via visual analysis for interpretable decision-making process. Expert Systems with Applications, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.687, + 0.905, + 0.728 + ], + "angle": 0, + "content": "[27] Taesung Park, Ming-Yu Liu, Ting-Chun Wang, and Jun-Yan Zhu. Semantic image synthesis with spatially-adaptive normalization. In CVPR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.731, + 0.905, + 0.799 + ], + "angle": 0, + "content": "[28] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.802, + 0.905, + 0.843 + ], + "angle": 0, + "content": "[29] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In ICML, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.905, + 0.899 + ], + "angle": 0, + "content": "[30] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18409" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[31] Royi Rassin, Eran Hirsch, Daniel Glickman, Shauli Ravfogel, Yoav Goldberg, and Gal Chechik. Linguistic binding in diffusion models: Enhancing attribute correspondence through attention map alignment. NeurIPS, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.204 + ], + "angle": 0, + "content": "[32] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 1, 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.206, + 0.482, + 0.247 + ], + "angle": 0, + "content": "[33] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In MICCAI, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.249, + 0.482, + 0.318 + ], + "angle": 0, + "content": "[34] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. NeurIPS, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.32, + 0.482, + 0.36 + ], + "angle": 0, + "content": "[35] Wenqiang Sun, Teng Li, Zehong Lin, and Jun Zhang. Spatial-aware latent initialization for controllable image generation. arXiv preprint arXiv:2401.16157, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.362, + 0.482, + 0.431 + ], + "angle": 0, + "content": "[36] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.433, + 0.482, + 0.459 + ], + "angle": 0, + "content": "[37] Cédric Villani et al. Optimal transport: old and new. Springer, 338, 2009. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.462, + 0.482, + 0.516 + ], + "angle": 0, + "content": "[38] Qiucheng Wu, Yujuan Liu, Handong Zhao, Trung Bui, Zhe Lin, Yang Zhang, and Shiyu Chang. Harnessing the spatial-temporal attention of diffusion models for high-fidelity text-to-image synthesis. In ICCV, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.518, + 0.482, + 0.587 + ], + "angle": 0, + "content": "[39] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, et al. Scaling autoregressive models for content-rich text-to-image generation. arXiv preprint arXiv:2206.10789, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.589, + 0.482, + 0.643 + ], + "angle": 0, + "content": "[40] Mert Yuksekgonul, Federico Bianchi, Pratyusha Kalluri, Dan Jurafsky, and James Zou. When and why vision-language models behave like bags-of-words, and what to do about it? arXiv preprint arXiv:2210.01936, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.645, + 0.482, + 0.685 + ], + "angle": 0, + "content": "[41] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.688, + 0.482, + 0.729 + ], + "angle": 0, + "content": "[42] Tianjun Zhang, Yi Zhang, Vibhav Vineet, Neel Joshi, and Xin Wang. Controllable text-to-image generation with gpt-4. arXiv preprint arXiv:2305.18583, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.731, + 0.482, + 0.758 + ], + "angle": 0, + "content": "[43] Xingyi Zhou, Vladlen Koltun, and Philipp Krahenbuhl. Simple multi-dataset detection. In CVPR, 2022. 6" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.758 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18410" + } + ] +] \ No newline at end of file diff --git a/2025/Spatial Transport Optimization by Repositioning Attention Map for Training-Free Text-to-Image Synthesis/68174923-d215-4833-8530-d036bb11b7fd_origin.pdf b/2025/Spatial Transport Optimization by Repositioning Attention Map for Training-Free Text-to-Image Synthesis/68174923-d215-4833-8530-d036bb11b7fd_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..d3ba6f37b3a97ae02200795427bff2a8c8e325e7 --- /dev/null +++ b/2025/Spatial Transport Optimization by Repositioning Attention Map for Training-Free Text-to-Image Synthesis/68174923-d215-4833-8530-d036bb11b7fd_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:39c04124f05e4e168e4beed735333f7acdb56e92a8d4f6cc9f59f019f90f747e +size 4203492 diff --git a/2025/Spatial Transport Optimization by Repositioning Attention Map for Training-Free Text-to-Image Synthesis/full.md b/2025/Spatial Transport Optimization by Repositioning Attention Map for Training-Free Text-to-Image Synthesis/full.md new file mode 100644 index 0000000000000000000000000000000000000000..c41d0b6d0bcc291044d85d93c226b7b017cc9a18 --- /dev/null +++ b/2025/Spatial Transport Optimization by Repositioning Attention Map for Training-Free Text-to-Image Synthesis/full.md @@ -0,0 +1,271 @@ +# Spatial Transport Optimization by Repositioning Attention Map for Training-Free Text-to-Image Synthesis + +Woojung Han Yeonkyung Lee Chanyoung Kim Kwanghyun Park Seong Jae Hwang* Yonsei University + +{dnwjDDL, yeonkyung.lee, chanyoung, kwanghyun.park, seongjae}@yonsei.ac.kr + +![](images/4ac46047851eb13618cbfeaeacfa209b6ec31355d9bf81b5dd48867cccff1c19.jpg) +Figure 1. Three main challenges in training-free text-to-image (T2I) generation: (1) missing objects, (2) mismatched attributes, and (3) mislocated objects. While existing approaches address missing objects and mismatched attributes, effectively controlling object positioning remains an open problem. Our proposed model, STORM, introduces a dynamic approach to aligning relative object positions throughout the denoising process, enabling precise spatial control without additional spatial templates. Red underline highlights errors made by SD. + +# Abstract + +Diffusion-based text-to-image (T2I) models have recently excelled in high-quality image generation, particularly in a training-free manner, enabling cost-effective adaptability and generalization across diverse tasks. However, while the existing methods have been continuously focusing on several challenges such as "missing objects" and "mismatched attributes," another critical issue of "mislocated objects" remains where generated spatial positions fail to align with text prompts. Surprisingly, ensuring such seemingly basic functionality remains challenging in popular T2I models due to the inherent difficulty of imposing explicit spatial guidance via text forms. To address this, we propose STORM (Spatial Transport Optimization by Repositioning Attention Map), a novel training-free approach for spatially coherent T2I synthesis. STORM employs Spatial Transport Optimization (STO), rooted in optimal transport theory, to + +dynamically adjust object attention maps for precise spatial adherence, supported by a Spatial Transport (ST) Cost function that enhances spatial understanding. Our analysis shows that integrating spatial awareness is most effective in the early denoising stages, while later phases refine details. Extensive experiments demonstrate that STORM surpasses existing methods, effectively mitigating mislocated objects while improving missing and mismatched attributes, setting a new benchmark for spatial alignment in T2I synthesis. + +# 1. Introduction + +Diffusion-based text-to-image (T2I) models exhibit remarkable versatility across various applications, generating high-quality images from textual prompts [13, 22, 24, 32, 41]. Notably, training-free methods offer efficiency, real-time adaptability, lower costs, and broad task generalization [5, 6, 9]. These training-free methods leverage their advantages while addressing three key challenges in fully capturing textual details. In particular, as shown in Fig. 1, these challenges include (1) missing objects, (2) mismatched at + +tributes, and (3) mislocated objects. These limitations may lead to images that poorly represent the intended scene or convey misleading information, reducing their reliability. + +While there has been substantial progress in addressing issues like missing objects and mismatched attributes in diffusion-based image synthesis [5, 12, 21, 23], the challenge of mislocated objects remains relatively unexplored despite its critical importance. For example, in Fig. 1, the fifth column shows that the backpack should be below a donut, but Stable Diffusion (SD) [32] generates donut to the left of a backpack. Furthermore, SD consistently generates similar images regardless of spatial cues in the text (e.g., same seed with only positional changes for teddy bear and vase in Fig. 2), indicating limited spatial understanding. This limitation highlights the importance of addressing the problem of mislocated objects and demonstrates the need for more advanced methods of interpreting and aligning with spatial cues. Overcoming these challenges is key to improving T2I model reliability and supporting adaptable applications across diverse settings. + +Existing solutions for mislocated objects often rely on fixed spatial templates or predefined layouts [6, 7, 38], which offer some control but impose rigid constraints and require additional inputs. For instance, these templates may fix an object's position in the image (e.g., far-left corner) instead of offering flexible guidance like placing one object "to the left" of another. While effective in structured fields like automated design, these approaches present limitations in more dynamic areas, such as digital art or interactive media, by restricting spatial flexibility. Therefore, instead of relying on restrictive templates or requiring additional inputs, we aim to push the boundaries of T2I models by enabling precise alignment with textual guidance, which is essential to unlocking their full potential across applications. + +To this end, we propose Spatial Transport Optimization by Repositioning Attention Map, STORM, a training-free approach that adjusts relative object positions dynamically, without relying on fixed spatial templates. While SD disregards spatial cues (see Fig. 2), STORM achieves various object positioning with the same model weights, ensuring spatial alignment. To implement STORM, we introduce the Spatial Transport Optimization (STO) framework, which interprets each text token's attention map as a distribution, and leverages Optimal Transport (OT) [37] to efficiently reposition distributions. At an early diffusion timestep, STO adjusts the placement of objects based on their relative positioning (see Fig.3 for attention map changes across the denoising process). While OT considers both distribution and distance to minimize transformation costs, we propose the Spatial Transport Cost (ST Cost) to align with our objectives and guide distribution positioning. This function updates and optimizes the latent vector through backward guidance, following methods from prior work [5, 6]. + +![](images/516592f1799a6376dd7c3f34af248e0d527c1e134b0644c6d03cbd802ad9e1e9.jpg) +Figure 2. Comparison of Spatial Awareness. $\{\text{position}^*\}$ in each prompt denotes the spatial relationship in each column (e.g., "to the left of", "to the right of", "above, and below"). While Stable Diffusion (SD) shows limited spatial awareness by generating similar images regardless of spatial prompts, our model accurately reflects specified positions. (Same seed for all synthesis). + +Our analysis suggests that integrating spatial awareness yields optimal results in the early denoising stages, whereas the later stages are better suited for focusing on fine details. Furthermore, we found that missing objects can be resolved concurrently as mislocated objects are corrected in the early steps. Extensive experiments show that our approach outperforms existing methods, setting a new benchmark for spatial alignment in diffusion models. + +Our main contributions are as follows: + +- We propose STORM, a training-free method that dynamically adjusts relative object positions, achieving accurate spatial alignment without relying on fixed templates. +- We introduce the Spatial Transport Optimization (STO) framework based on Optimal Transport (OT), which treats attention maps as distributions and repositions objects relative to each other across diffusion timesteps. +- We develop ST Cost function that penalizes overlaps and progressively enhances spatial accuracy, demonstrating its effectiveness through extensive experiments that set a new benchmark for spatial alignment in T2I models. + +# 2. Related Work + +Text-to-Image Synthesis. Text-to-image (T2I) synthesis aims to generate visually realistic images that align with given text prompts. Early studies focused on GANs [18, 19, 27] and autoregressive models [29, 39], but diffusion models [15] have taken the lead, demonstrating remarkable performance in T2I synthesis [22, 32]. Despite advances driven by VLMs [32, 34, 39], aligning generated images with text prompts remains a challenge. To address this, some approaches [2, 30, 34] have expanded network architectures, though these methods require retraining and are difficult to adapt to existing frameworks. Alternatively, other research [5, 9, 21, 23] focused on training-free improvement strategies, which aligns closely with our approach. + +![](images/90b79f3547110a014b5a218fc9829f909e771d7e8a89a6371c54e946fc654ddf.jpg) +Figure 3. Comparison of Attention Map Progression During Denoising. Visualization of the attention maps for "a red bird to the right of a green plant" throughout the denoising process for both Stable Diffusion (a) and our model (b). While Stable Diffusion struggles to distinctly capture the spatial relationship between the bird and the plant, our model effectively aligns the objects according to the specified spatial cue ("to the right of"). The resulting image from our model demonstrates improved spatial accuracy compared to Stable Diffusion. + +Training-free T2I Models. Previous studies [1, 5, 9, 12, 21, 22, 31] have explored various approaches to enhance how text prompts are aligned within text-to-image diffusion models. Composable Diffusion [22] and Structure Diffusion [9] aim to improve compositionality through multi-model or cross-attention methods, yet they still face issues like subject mixing and attribute binding. To address these, Attend-and-Excite [5] dynamically adjusts attention during inference to capture all subjects, while Divide-and-Bind [21] improves attribute binding and compositional accuracy with targeted loss functions. CONFORM [23] uses a contrastive method to separate objects in attention maps while keeping related attributes together, and INITNO [12] further tackles spatial misalignment by improving control over object placement to reduce overlap. Although these studies focus on the missing object issue by reducing object overlap, they often overlook the crucial task of capturing accurate relative positioning and directional relationships between objects. + +Training-free T2I Models for Mislocated Objects. The advantages of attention maps for controlling spatial positioning have led to several studies using fixed templates, such as layouts, to improve object placement [6, 7, 11, 25, 35, 38]. Wu et al. [38] use layout predictors to assign pixel regions, while ZestGuide [7] and Chen et al. [6] leverage cross-attention layers for alignment with segmentations or user-defined layouts. However, these methods often require additional layout inputs and lack flexibility for relative positioning. As a result, there is a need for methods that enhance object placement without relying on predefined layouts. + +# 3. Preliminaries + +Text Conditioned via Cross-Attention. Our approach uses Stable Diffusion (SD) [32], which generates images from text by integrating text conditioning through cross-attention. A pre-trained CLIP encoder [28] converts the text prompt into a conditioning vector $c$ , fed into the UNet model [33] via cross-attention layers at different resolutions. At each timestep $t$ , the attention map $A_{t} \in \mathbb{R}^{P \times P \times N}$ assigns probabilities for transferring token information to spatial positions, where $P$ is the spatial dimension of the feature map and $N$ the number of tokens in the text prompt. In our work, we use $16 \times 16$ attention maps as they contain the most semantically meaningful information [14]. + +Optimal Transport with Sinkhorn Algorithm. Optimal Transport (OT) [37] provides a framework for transforming one probability distribution into another while preserving spatial relationships. Given two discrete distributions $\mu = \{\mu_u\}_{u=1}^n$ and $\nu = \{\nu_v\}_{v=1}^n$ , each with $n$ discrete locations, and a cost matrix $\mathbf{C} \in \mathbb{R}^{n \times n}$ , where $\mathbf{C}_{uv}$ is the cost of moving distribution from point $u$ to $v$ , OT seeks a transport plan $\mathbf{P} \in \mathbb{R}^{n \times n}$ that minimizes the total cost, $\min_{P \geq 0} \sum_{u=1}^{n} \sum_{v=1}^{n} \mathbf{C}_{uv} \mathbf{P}_{uv}$ . For large $n$ , the Sinkhorn algorithm [8] applies entropic regularization for a faster, scalable, and differentiable solution: + +$$ +\min _ {P \geq 0} \quad \sum_ {u = 1} ^ {n} \sum_ {v = 1} ^ {n} \mathbf {C} _ {u v} \mathbf {P} _ {u v} - \varepsilon \sum_ {u = 1} ^ {n} \sum_ {v = 1} ^ {n} \mathbf {P} _ {u v} (\log \mathbf {P} _ {u v} - 1), \tag {1} +$$ + +where $\varepsilon > 0$ balances transport cost and entropy. We further integrate spatial alignment objectives into this OT-based approach in our framework. + +# 4. Method + +We describe the full pipeline of our model, STORM, as shown in Fig. 4. To support the understanding of our Spatial Transport Optimization (STO) framework, we first provide background information in Sec. 4.1. The full details of STO in Sec. 4.2, while Sec. 4.3 covers optimizations for non-spatial text prompts, and Sec. 4.4 details the update process. + +# 4.1. Background of STO + +Since objects are generated in regions of high attention within the attention map (see Fig. 3), our goal is to reposition the attention map based on the positional information from the text prompt, using relative objects as a reference. To achieve this, our approach involves three types of attention maps: the attention map of the object to be moved, called the source distribution (e.g., blue area in Fig. 4); the attention map of the relative object, which acts as a reference distribution (e.g., red area in Fig. 4); and the target distribution which represents the intended position for the source distribution to move (e.g., gray area in Fig. 4). The target distribution is an arbitrary distribution defined based on the reference point, which is the centroid of the reference distribution, $(j_{\mathrm{ref}}, i_{\mathrm{ref}})$ . It primarily serves to guide the intended position of the source distribution using reference points without exerting significant direct influence. + +![](images/8096962e3fc7166fc54cbbc04e1742ffb70c4eac9855ddc3a881a154e038120b.jpg) +Figure 4. Overview pipeline of STORM. Our method leverages Optimal Transport in a training-free manner, allowing the model to accurately reflect relative object positions at each step without additional inputs. Given the prompt "A car to the left of an elephant", our method dynamically adjusts the attention maps to induce the specified spatial relationship. The process starts with initial attention maps for the "car" and "elephant" at time step $z_{t}$ . Using the centroids of these attention maps, the Spatial Transport Optimization (STO) computes the losses to correct positional relationships (e.g., ensuring the car is to the left of the elephant). The updated attention map is then used to refine the latent representation $z_{t}$ , leading to a final image that adheres to the desired spatial arrangement. The comparison of attention maps (before and after STO) shows improved alignment, effectively placing the car to the left of the elephant as instructed in the prompt. + +# 4.2. Spatial Transport Optimization Framework + +In this section, we detail our Spatial Transport Optimization (STO) framework, which dynamically repositions attention maps for accurate object alignment without modifying model weights. This framework consists of two main components: a cost function and a transport plan. The cost function directs the distribution shift from the source in the intended direction, allowing the attention map to adjust positionally according to spatial relationships. Although we build on standard cost functions that measure distance between locations, our Spatial Transport Cost (ST Cost) incorporates additional objectives adapted to our specific spatial alignment goals. The transport plan, which optimally moves the distribution between points, is implemented using the Sinkhorn algorithm. Beginning with an initial attention map, STO iteratively updates the latent vector by applying STO-derived losses, gradually guiding the distribution into the target alignment. + +# 4.2.1 Construction of Target Distributions + +Our approach focuses on relative positioning adjustments at each timestep, using an arbitrary target distribution (e.g., gray area in Fig. 4) to guide the source distribution based on the reference. Depending on the text-specified direction (e.g., horizontal or vertical), only the relevant reference point of $(j_{\mathrm{ref}},i_{\mathrm{ref}})$ is considered: for 'left' or 'right' positioning, only the horizontal dimension of the reference point $j_{\mathrm{ref}}$ is used, while for 'above' or 'below' positioning, only the vertical dimension $i_{\mathrm{ref}}$ is considered, allowing the flexibility in the opposite dimension. For example, in Fig. 4a, with the car attention map as a source distribution and the + +elephant attention map as a reference distribution, the target distribution is placed to the left of a elephant's distribution. This approach similarly applies to another source. Supplementary Materials provide details on computing the target distribution and reference point. + +# 4.2.2 Spatial Transport Cost Function + +In the STO framework, we aim to find an optimal transport plan that minimizes the cost of transferring the attention map. The cost function measures the expense of moving distribution mass by computing distances between points, where a higher value indicates less desirable locations. While our overall cost function merges the standard OT loss with an additional term (detailed in the supplementary material), here we focus on our newly introduced ST Cost, which is built on two core principles: + +(I) Positional Cost: We developed a cost function to guide the distribution to its intended position relative to the reference point. In simple terms, when the source distribution is intended to be on the left side of the reference point, it has a low cost when positioned on the left and a high cost when positioned on the right. To quantify this directional positioning, we define a set of positional $\delta$ values, representing the distances between the current point and the reference point in each of the four directions. Specifically, the $\delta$ values are defined as follows: $\delta_{ij}^{\leftarrow} = j_{\mathrm{ref}} - j,\delta_{ij}^{\rightarrow} = j - j_{\mathrm{ref}},\delta_{ij}^{\uparrow} = i_{\mathrm{ref}} - i,\delta_{ij}^{\downarrow} = i - i_{\mathrm{ref}}$ with $j$ and $i$ representing the current position values, and the arrows indicate the desired and restricted direction for positioning the source attention map relative to the reference point. When constructing the overall cost function that incorporates all directional $\delta$ values, each direction is categorized as either de + +sired or restricted. The desired direction $(\delta^{\mathrm{des}})$ corresponds to the intended movement of the source distribution, while the restricted direction $(\delta^{\mathrm{res}})$ represents the opposite or undesired direction. For instance, if the source distribution should move to the left, $\delta_{ij}^{\leftarrow}$ is marked as the desired direction $(\delta^{\mathrm{des}})$ , while $\delta_{ij}^{\rightarrow}$ is treated as the restricted direction $(\delta^{\mathrm{res}})$ . The combined cost function, accounting for both desired and restricted positions, can be expressed as: + +$$ +\Delta_ {i j} \left(\delta_ {i j} ^ {\mathrm {d e s}}, \delta_ {i j} ^ {\mathrm {r e s}}\right) = \frac {1}{\omega \left(\delta_ {i j} ^ {\mathrm {d e s}} + \epsilon\right)} \mathbb {1} _ {+} \left[ \delta_ {i j} ^ {\mathrm {d e s}} \right] + \omega \left(\delta_ {i j} ^ {\mathrm {r e s}} + \epsilon\right) \mathbb {1} _ {+} \left[ \delta_ {i j} ^ {\mathrm {r e s}} \right], \tag {2} +$$ + +where $\omega (\cdot) > 1$ is a progressive adaptive weight that controls the alignment importance across timesteps, and $\epsilon$ is a stabilizing factor. $\mathbb{1}_{+}$ is an indicator function that equals 1 when the input is positive and 0 otherwise, allowing selective application of computation based on positivity. As the source distribution aligns more closely with the desired direction, the $\Delta$ value approaches zero, indicating minimal cost for correctly aligned positioning (see Example below). + +(II) Non-overlap Cost: In this process, it is crucial to prevent the source distribution from overlapping with the reference distribution during its movement. This is one of the primary issues in addressing the missing object problem in SD, where avoiding overlap is essential for ensuring that each object is distinctly represented. To achieve this, we simply embed the reference distribution directly into the cost function, guiding objects to occupy separate locations. Specifically, with $A_{ij}$ representing the attention weight of the reference distribution, we incorporated this distribution into the cost function, assigning high costs to regions occupied by the reference distribution and low costs elsewhere. + +Combining this principle with our core ideas, our cost function can be defined as follows: + +$$ +C _ {i j} = A _ {i j} \Delta_ {i j} \left(\delta_ {i j} ^ {\mathrm {d e s}}, \delta_ {i j} ^ {\mathrm {r e s}}\right). \tag {3} +$$ + +Example. In Fig. 4, the car should be left of the elephant, and the elephant to the right of the car. The cost functions for each of them are represented by $C_{ij}^{\leftarrow} = A_{ij}\Delta_{ij}(\delta_{ij}^{\leftarrow},\delta_{ij}^{\rightarrow})$ , $C_{ij}^{\rightarrow} = A_{ij}\Delta_{ij}(\delta_{ij}^{\rightarrow},\delta_{ij}^{\leftarrow})$ , respectively. Taking the car as an example, it needs to be positioned to the left of the elephant relative to the elephant's centroid point. In more detail, if the current point is positioned to the left of the centroid, the $\delta_{ij}^{\mathrm{des}}$ is activated; if positioned to the right of the centroid, the $\delta_{ij}^{\mathrm{res}}$ is activated. Additionally, as the car moves to the left of the elephant, it should not overlap the elephant it passes along the way. The elephant's attention map is used in the cost function, guiding the car to avoid overlapping with the elephant's area. The resulting $C_{ij}$ is utilized as the cost function for STO. + +# 4.2.3 Sinkhorn Algorithm-based Transport Plan + +After combining the standard OT approach with ST components (see Supplementary Material for the complete formulation), we compute the transport plan using the Sinkhorn + +algorithm [8], which applies entropic regularization for computational efficiency and stability, especially for high-dimensional attention maps. The regularized OT problem is formulated as in Eq. (1) using target distribution calculated in Sec. 4.2.1. The Sinkhorn algorithm iteratively updates transport plan $\mathbf{P}$ to meet the marginal constraints of source and target distributions, applying the cost function bidirectionally for both objects. + +# 4.3. Optimization for Objects and Attributes + +Our method also extends to cases where the text lacks spatial information. (i) Object-level Optimization: We handle cases where the text prompt specifies objects without explicit spatial guidance by randomly placing the target distribution of each object. This flexible positioning encourages diverse object placement when spatial relationships are not provided. Therefore, the cost function is simplified to focus on solely on preventing overlap, expressed as $C_{ij} = A_{ij} \otimes \mathbf{1}_N$ , where $A$ is the reference attention map. (ii) Attribute Matching Optimization: The attention map for an object and its corresponding attribute should be identical, reflecting the same attention focus. Therefore, we used the standard OT cost function, calculating Euclidean distance between corresponding positions across the object, and attribute attention maps to quantify differences. + +# 4.4. Optimization and Update Process + +We update the latent vector using STO-based losses to guide the spatial alignment of the attention maps in a differentiable manner. At each timestep, we compute the OT loss $\mathcal{L}$ and update the latent vector $z_{t}$ as $z_{t}^{\prime}\gets z_{t} - \alpha_{t}\cdot \nabla z_{t}\mathcal{L}$ where $\alpha_{t}$ is the step size. After the update, a forward pass computes $z_{t - 1}$ for the next denoising step. + +Remark. As shown in Fig.8, addressing the mislocated object problem is most effective in early timesteps, which, according to related work[26], establish the overall image structure while later timesteps refine finer details. Based on this, STORM establishes object positioning and spatial configurations early in the denoising process, minimizing the risk of misplacements and overlaps. Furthermore, as shown in the bottom row of Fig. 8, solving the mislocated object problem before addressing missing objects can naturally resolve the latter, highlighting the importance of prioritizing mislocated objects as a critical step in the synthesis process. + +# 5. Experiments + +Implementation Details. Following methods established in prior research [5, 21, 23], we conducted a comprehensive evaluation of our model. We used the official Stable Diffusion models [32] v1.4 and v2.1, configuring the denoising timestep to 50, and the guidance scale to 7.5 in line with common practices in the field. To smooth the cross-attention map, a Gaussian filter with a kernel size of 3 and a standard deviation of 0.5 was applied. Optimization + +Table 1. Performance comparison between different models on VISOR (\%) and Object Accuracy (OA) (\%) metrics, based on Stable Diffusion 1.4 and Stable Diffusion 2.1 [32]. For entries marked with $\dagger$ , official values were unavailable; thus, we directly calculated these values. Bold values indicate the best scores, while underlined values indicate the second-best scores among the fair comparison. Notes on VISOR. VISOR cond: spatial accuracy using the results in which both objects are sufficiently generated; VISOR uncond: spatial accuracy using all results, including results with failed object generations; VISOR $1/2/3/4$ : ratios of one, two, three, or four spatially accurate results out of four seed results, respectively. (e.g., $\mathrm{VISOR}_1$ : % of results with exactly one spatially correct result out of four generated results) + +
ModelVenueExtra TrainingExtra InputOA (%)VISOR (%)
uncondcond1234
Stable Diffusion 1.4 Based
SD 1.4 [32]-X29.8618.8162.9846.6020.116.891.63
SD 1.4 + CDM [22]ECCV22✓X23.2714.9964.4139.4414.564.841.12
GLIDE [24]ICML22✓X3.361.9859.066.721.020.170.03
GLIDE + CDM [22]ECCV22✓X10.176.4363.2120.074.690.830.11
Control-GPT [42]arXiv23✓✓48.3344.1765.9769.8051.2035.6720.48
Layout Guidance [6]WACV24X✓40.0138.8095.95----
REVISION [3]ECCV24X✓53.9652.7197.6977.7961.0244.9027.15
Structure Diffusion [9]ICLR23XX28.6517.8762.3644.7018.736.571.46
Attend-and-Excite [5]SIGGRAPH23XX42.0725.7561.2149.2919.334.560.08
Divide-and-Bind†[21]BMVC24XX46.0331.6268.7064.7237.8218.645.30
INITNO†[12]CVPR24XX60.4035.1858.2471.2042.7120.096.72
CONFORM†[23]CVPR24XX60.7338.4862.3373.0145.8225.579.52
Ours (SD 1.4)XX61.0157.5894.3985.9369.7149.0125.70
Stable Diffusion 2.1 Based
SD 2.1 [32]-X47.8330.2563.2464.4235.7416.134.70
SPRIGHT [4]ECCV24✓X60.6842.2371.2471.7851.8833.0916.15
REVISION [3]ECCV24X✓48.2647.1197.6176.0755.7537.1019.53
Ours (SD 2.1)XX62.5559.3594.8888.3471.7552.0325.42
+ +was performed iteratively with forward updates at specified timesteps ( $t \in 5, 10, 15, 20$ ), stopping at $t = 25$ to enhance quality. All experiments use NVIDIA RTX A6000 Ada*. + +# 5.1. Evaluation Results + +We evaluated our method against existing training-free T2I models using both qualitative and quantitative measures, focusing on two key aspects. (i) Spatial Accuracy: To measure alignment with spatial configurations specified in prompts, we used the VISOR [10] and the T2I-CompBench [17], both provide a comprehensive assessment of spatial relationships captured by the model. (ii) Object and Attribute Consistency: We assessed object and attribute representation using T2I-CompBench in combination with CLIP-space distance [28] and TIFA score [16], following established methods [5, 21, 23]. We provide detailed descriptions of each evaluation metric and datasets in the Supplementary Materials for additional clarity. + +Quantitative Evaluation: VISOR. To evaluate spatial accuracy, we use the VISOR [10] to measure the ability of a model to position objects based on spatial cues (e.g., above) in text prompts. VISOR benchmark contains over 25K prompts describing spatial relationships between objects. Notably, as seen in Table 1, the increase in $\mathrm{VISOR}_4$ demonstrates the model's reliability in generating spatially accurate images across prompts and seeds. + +Table 2. Comparison of methods on T2I-CompBench, with attribute binding and spatial relationships calculated using models based on Stable Diffusion 1.4 and 2.1. + +
MethodAttribute Binding (Color ↑)Attribute Binding (Shape ↑)Attribute Binding (Texture ↑)Object Relationship (Spatial ↑)
Stable Diffusion 1.4 Based
Stable Diffusion-v1.4 [32]0.37650.35760.41560.1246
Ours (SD 1.4)0.64580.59830.75390.1613
Stable Diffusion 2.1 Based
Stable Diffusion-v2.1 [32]0.50650.42210.49220.1342
Composable Diffusion [22]0.40630.32990.36450.0800
Structured Diffusion [9]0.49900.42180.49000.1386
Attend-and-Excite [5]0.64000.45170.59630.1455
Ours (SD 2.1)0.67770.62260.78840.1981
+ +Quantitative Evaluation: T2I-CompBench. The T2I-CompBench [17] evaluates object presence, spatial relationships, and attribute alignment from text prompts. We used the attribute and spatial configuration datasets provided by T2I-CompBench, averaging results over more than 10 randomly selected seeds for each dataset. Attribute Binding is evaluated with the BLIP-VQA metric [20], which poses separate questions for each attribute-object pair in the prompt. UniDet-based metric [43] is used for spatial relationships. In Table 2, our score shows that the generated images position the specified objects correctly and accurately reflect the given attributes (e.g., color, shape, texture). + +Quantitative Evaluation: CLIP Score and TIFA Score. Following previous work [5, 12, 23], we evaluate our method using CLIP Image-Text and Text-Text Similarity, generating 64 images per prompt and averaging CLIP co + +![](images/6f6563efb61becaf78ec9b87df5e987ca8f097923a40e1b4f8c2da7eda98e414.jpg) +Figure 5. Qualitative comparison across the custom prompt, which involves attribute and positional information in text, evaluating previous state-of-the-art training-free T2I methods, Attend&Excite [5], Divide&Bind [21], INITNO [12], CONFORM [23], and ours. + +Table 3. User study evaluating model performance on object synthesis, attribute matching, spatial correctness, and overall fidelity. + +
MethodObject Accuracy (%)Attribute Matching (%)Spatial Correctness (%)Overall Fidelity (%)
Stable Diffusion [32]14.6814.3111.8115.02
Attend-and-Excite [5]16.3416.5013.2016.04
Divide-and-Bind [21]16.3915.6913.0116.52
INITNO [12]15.9816.0313.1515.63
CONFORM [23]15.6416.7912.9114.14
Ours20.9720.6835.9222.65
+ +sine similarity scores across methods. Fig. 7 reports two CLIP Image-Text Similarity metrics: full prompt similarity (cosine similarity between the entire prompt and generated image) and minimum object similarity (minimum similarity between the generated image and each subject prompt) to see the accuracy of each object. We generate captions for the images with BLIP [20] and compute cosine similarity with the original prompt. Our method achieves comparable performance to other leading methods. However, as the reliability of the CLIP score is limited by its "bag-of-words" approach [40] (as noted by [5]), we also use the TIFA metric [16] for improved evaluation of complex prompts. TIFA [16] evaluates image faithfulness to the input prompt by generating question-answer pairs using the LLaMA 2 [36] model. Our method performs well compared to other state-of-the-art models. + +Quantitative Evaluation: User Studies. We create 10 custom text prompts with detailed objects, attributes, and spa + +![](images/2bb58b2f12a24e2ab41514c50056e828d9ee092f414d32d83dd9037591a9e270.jpg) +Figure 6. Generated images with complex positioning details in the prompt (Additional results in Supplementary Materials). + +tial information, then generate images using random seeds. With 30 participants, our user study evaluates (1) object accuracy, (2) attribute matching, (3) spatial correctness, and (4) overall fidelity. In Table 3, our model shows surpassing results across the existing models of all categories. + +Qualitative Analysis. Fig. 5 provides a qualitative comparison with recent state-of-the-art models [12, 23]. To compare attribute matching and object localization accuracy, we create custom text prompts to evaluate the qualitative performance of our model. Existing methods often misalign + +![](images/aa791de998ec864ce81dec2e93acea1b5a7b3be5e8b45f5e07858bccb7380be6.jpg) +Figure 7. Comparison of various models on CLIP Image-Text, CLIP Text-Text, and TIFA scores across prompts involving Animal-Animal, Animal-Object, and Object-Object pairs. (The seed is randomly selected.) + +![](images/9e207de35df137f1d0c9f77e81141e5f34d45b2f925b244536d31e132b1062d3.jpg) + +![](images/74600838616d46d450a3a651c4ebed241cdabee538f6c15666c23f3d848a38fe.jpg) +Exp.#A0 (SD) + +![](images/6872d1bd4154ec196101ba94c8f7256b1c13f9735dc134ab3d36c0d12397652d.jpg) +"A cake to the left of a tie" +Exp.#A1 + +![](images/b09aef1acabae910e6c694cd23ee3d9f1b474fd6538fbf633845ecc9fffa204d.jpg) +Exp.#A2 + +![](images/c9d5d8e87e46b948fdfed57cbe639efd94cea39461f3f4c541458e0f8bfba06b.jpg) +Exp.#A3 + +![](images/4fec3d3107de492ac659bc0c9d9ffeaaa20e19b00d450510d890f5e0a20b4508.jpg) +Exp.#A4(Ours) + +![](images/2e4b2b3ed6e933b26ffa99e29d43124a879e0c7ef07a3a09d245dd144fd10e1e.jpg) +Exp.#A0 (SD) +Figure 8. Comparison of results when applying STO at different timesteps. Experiments are organized as follows: no STO, STO applied from timesteps 19-24, 13-24, 7-24, and 1-24 (Ours). As seen in the images, earlier STO application improves object positioning and reduces overlap, resulting in more accurately positioned objects. (See Table 4 for quantitative evaluation.) + +![](images/21b28c35622515d03f64baec19d1cf117751fb078e0b62d59c28dca30788160d.jpg) +"An orange above a suitcase" +Exp.#A1 + +![](images/7571c6ab5d0366785ebe52fc87797ddda207e51302110d75305fe97f5246e394.jpg) +Exp.#A2 + +![](images/c1bf3e168887c321ef9c5e16dfe12bad45b47318d9b3efe286cb4299f805a121.jpg) +Exp.#A3 + +![](images/3c0a3cf3070b282ad151499bb0a32a2681ebcd07be4c75197d9fc5268a077866.jpg) +Exp.#A4 (Ours) + +spatial details, leading to errors such as a rabbit overlapping a backpack instead of being positioned to its right. In contrast, our method, which is focused on precise positioning, demonstrates a strong understanding of spatial information and generates objects accurately. Our method also can generate images that accurately reflect complex spatial information, even when prompts contain intricate positioning details (refer Fig. 6). + +# 5.2. Ablation Studies + +Effects of Optimizing through Timestep. In Table 4, we compare STO applied at different timestep ranges (Exp.#A1: 19-24, Exp.#A2: 13-24, Exp.#A3: 7-24, Exp.#A4: 1-24). These experiments show that setting spatial configurations in the early stages is essential not only for positional accuracy but also for achieving higher OA. This improvement increases as spatial relationships are established earlier in the process. Notably, all configurations with STO outperform the baseline (Exp.#A0, without STO), demonstrating a substantial improvement in both object positioning and accuracy. In Fig. 8, tie placement is optimal when spatial cues are applied early, while delayed guidance leads to misplacements or suboptimal synthesis. + +Effects of Cost Function. Table 5 presents an ablation study on methods for computing movement costs between a source and target distribution. Exp.#B0 uses Manhattan distance (only horizontal and vertical movement), while Exp.#B1 considers the mass distribution of the attention + +Table 4. Ablation study on the impact of applying STO at different timesteps. Exp.#A0 shows baseline results from SD without STO. From Exp.#A1 to Exp.#A4, STO is progressively applied across increasing timestep ranges (19-24, 13-24, 7-24, and 1-24). + +
# Exp.OA (%)VISOR
uncondcond1234
A029.8618.8162.9846.6020.116.891.63
A146.6237.4880.3874.1145.7322.307.82
A255.6547.9086.0781.6360.3035.8913.81
A359.8553.6289.6083.6167.5344.1219.31
A4 (Ours)61.0157.5894.3985.9369.7149.0125.70
+ +Table 5. Ablation study on computing cost function, highlighting the impact of ST Cost components on OA and VISOR performance. (B0: Manhattan distance; B1: Euclidean distance; B2: B0 + B1; B3: ST cost without overlap prevention.) + +
# Exp.OA (%)VISOR
uncondcond1234
B027.3217.9165.5142.8419.257.602.00
B147.4830.3563.9365.2037.7114.034.54
B246.6029.2062.9363.9233.9114.434.60
B331.3827.7888.5361.3032.6113.443.82
B4 (Ours)61.0157.5894.3985.9369.7149.0125.70
+ +map using Euclidean distance. Exp.#B2 combines Exp.#B0 and Exp.#B1. Exp.#B3 excludes the overlap prevention component from ST Cost to examine its impact. Our results (Exp.#B4) show that ST Cost outperforms standard OT functions, with overlap prevention proving critical for maintaining OA, despite minimal impact on VISOR. + +# 6. Conclusion + +In this work, we introduced STORM, a framework that dynamically addresses spatial misalignment issues in training-free T2I synthesis. By leveraging the STO framework, which combines OT and an ST Cost function, STORM not only resolves misplaced objects but also tackles missing objects and mismatched attributes. Extensive experiments show that STORM significantly improves spatial alignment and object accuracy, surpassing existing state-of-the-art methods. This advancement establishes STORM as a new benchmark for spatial control in T2I generation, pushing the boundaries of adaptability and efficiency in training-free approaches. + +# References + +[1] Aishwarya Agarwal, Srikrishna Karanam, KJ Joseph, Apoorv Saxena, Koustava Goswami, and Balaji Vasan Srinivasan. A-star: Test-time attention segregation and retention for text-to-image synthesis. In ICCV, 2023. 3 +[2] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Ji-aming Song, Qinsheng Zhang, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, et al. ediff-i: Text-to-image diffusion models with an ensemble of expert denoisers. arXiv preprint arXiv:2211.01324, 2022. 2 +[3] Agneet Chatterjee, Yiran Luo, Tejas Gokhale, Yezhou Yang, and Chitta Baral. Revision: Rendering tools enable spatial fidelity in vision-language models. In ECCV, 2025. 6 +[4] Agneet Chatterjee, Gabriela Ben Melech Stan, Estelle Aflalo, Sayak Paul, Dhruba Ghosh, Tejas Gokhale, Ludwig Schmidt, Hannaneh Hajishirzi, Vasudev Lal, Chitta Baral, et al. Getting it right: Improving spatial consistency in text-to-image models. In ECCV, 2025. 6 +[5] Hila Chefer, Yuval Alaluf, Yael Vinker, Lior Wolf, and Daniel Cohen-Or. Attend-and-excite: Attention-based semantic guidance for text-to-image diffusion models. SIGGRAPH, 2023. 1, 2, 3, 5, 6, 7 +[6] Minghao Chen, Iro Laina, and Andrea Vedaldi. Training-free layout control with cross-attention guidance. In WACV, 2024. 1, 2, 3, 6 +[7] Guillaume Couairon, Marlene Careil, Matthieu Cord, Stéphane Lathuiliere, and Jakob Verbeek. Zero-shot spatial layout conditioning for text-to-image diffusion models. In ICCV, 2023. 2, 3 +[8] Marco Cuturi. Sinkhorn distances: Lightspeed computation of optimal transport. NeurIPS, 2013. 3, 5 +[9] Weixi Feng, Xuehai He, Tsu-Jui Fu, Varun Jampani, Arjun Akula, Pradyumna Narayana, Sugato Basu, Xin Eric Wang, and William Yang Wang. Training-free structured diffusion guidance for compositional text-to-image synthesis. arXiv preprint arXiv:2212.05032, 2022. 1, 2, 3, 6 +[10] Tejas Gokhale, Hamid Palangi, Besmira Nushi, Vibhav Vineet, Eric Horvitz, Ece Kamar, Chitta Baral, and Yezhou Yang. Benchmarking spatial relationships in text-to-image generation. arXiv preprint arXiv:2212.10015, 2022. 6 +[11] Biao Gong, Siteng Huang, Yutong Feng, Shiwei Zhang, Yuyuan Li, and Yu Liu. Check locate rectify: A training-free layout calibration system for text-to-image generation. In CVPR, 2024. 3 +[12] Xiefan Guo, Jinlin Liu, Miaomiao Cui, Jiankai Li, Hongyu Yang, and Di Huang. Initno: Boosting text-to-image diffusion models via initial noise optimization. In CVPR, 2024, 2, 3, 6, 7 +[13] Woojung Han, Chanyoung Kim, Dayun Ju, Yumin Shim, and Seong Jae Hwang. Advancing text-driven chest x-ray generation with policy-based reinforcement learning. MICCAI, 2024. 1 +[14] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 3 + +[15] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. NeurIPS, 2020. 2 +[16] Yushi Hu, Benlin Liu, Jungo Kasai, Yizhong Wang, Mari Ostendorf, Ranjay Krishna, and Noah A Smith. Tifa: Accurate and interpretable text-to-image faithfulness evaluation with question answering. In ICCV, 2023. 6, 7 +[17] Kaiyi Huang, Kaiyue Sun, Enze Xie, Zhenguo Li, and Xihui Liu. T2i-compbench: A comprehensive benchmark for open-world compositional text-to-image generation. NeurIPS, 2023. 6 +[18] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. In CVPR, 2017. 2 +[19] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In CVPR, 2019. 2 +[20] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, 2022. 6, 7 +[21] Jiachen Li, Jitesh Jain, and Humphrey Shi. Matting anything. In CVPR, 2024. 2, 3, 5, 6, 7 +[22] Nan Liu, Shuang Li, Yilun Du, Antonio Torralba, and Joshua B Tenenbaum. Compositional visual generation with composable diffusion models. In ECCV, 2022. 1, 2, 3, 6 +[23] Tuna Han Salih Meral, Enis Simsar, Federico Tombari, and Pinar Yanardag. Conform: Contrast is all you need for high-fidelity text-to-image diffusion models. In CVPR, 2024. 2, 3, 5, 6, 7 +[24] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 1, 6 +[25] Weili Nie, Sifei Liu, Morteza Mardani, Chao Liu, Benjamin Eckart, and Arash Vahdat. Compositional text-to-image generation with dense blob representations. In ICML, 2024. 3 +[26] Ji-Hoon Park, Yeong-Joon Ju, and Seong-Whan Lee. Explaining generative diffusion models via visual analysis for interpretable decision-making process. Expert Systems with Applications, 2024. 5 +[27] Taesung Park, Ming-Yu Liu, Ting-Chun Wang, and Jun-Yan Zhu. Semantic image synthesis with spatially-adaptive normalization. In CVPR, 2019. 2 +[28] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 3, 6 +[29] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In ICML, 2021. 2 +[30] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 2 + +[31] Royi Rassin, Eran Hirsch, Daniel Glickman, Shauli Ravfogel, Yoav Goldberg, and Gal Chechik. Linguistic binding in diffusion models: Enhancing attribute correspondence through attention map alignment. NeurIPS, 2024. 3 +[32] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 1, 2, 3, 5, 6, 7 +[33] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In MICCAI, 2015. 3 +[34] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. NeurIPS, 2022. 2 +[35] Wenqiang Sun, Teng Li, Zehong Lin, and Jun Zhang. Spatial-aware latent initialization for controllable image generation. arXiv preprint arXiv:2401.16157, 2024. 3 +[36] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 7 +[37] Cédric Villani et al. Optimal transport: old and new. Springer, 338, 2009. 2, 3 +[38] Qiucheng Wu, Yujuan Liu, Handong Zhao, Trung Bui, Zhe Lin, Yang Zhang, and Shiyu Chang. Harnessing the spatial-temporal attention of diffusion models for high-fidelity text-to-image synthesis. In ICCV, 2023. 2, 3 +[39] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, et al. Scaling autoregressive models for content-rich text-to-image generation. arXiv preprint arXiv:2206.10789, 2022. 2 +[40] Mert Yuksekgonul, Federico Bianchi, Pratyusha Kalluri, Dan Jurafsky, and James Zou. When and why vision-language models behave like bags-of-words, and what to do about it? arXiv preprint arXiv:2210.01936, 2022. 7 +[41] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, 2023. 1 +[42] Tianjun Zhang, Yi Zhang, Vibhav Vineet, Neel Joshi, and Xin Wang. Controllable text-to-image generation with gpt-4. arXiv preprint arXiv:2305.18583, 2023. 6 +[43] Xingyi Zhou, Vladlen Koltun, and Philipp Krahenbuhl. Simple multi-dataset detection. In CVPR, 2022. 6 \ No newline at end of file diff --git a/2025/Spatial Transport Optimization by Repositioning Attention Map for Training-Free Text-to-Image Synthesis/images.zip b/2025/Spatial Transport Optimization by Repositioning Attention Map for Training-Free Text-to-Image Synthesis/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..2713716453921a5eba0acd482772d0546d3016b2 --- /dev/null +++ b/2025/Spatial Transport Optimization by Repositioning Attention Map for Training-Free Text-to-Image Synthesis/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:58684dd18dcf3baafddb9043a89dbfa82875d017a8d3c951f8798b4f199581c0 +size 1070233 diff --git a/2025/Spatial Transport Optimization by Repositioning Attention Map for Training-Free Text-to-Image Synthesis/layout.json b/2025/Spatial Transport Optimization by Repositioning Attention Map for Training-Free Text-to-Image Synthesis/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..e78b2b261e1fe35fa2e9522e47456529aaff62ff --- /dev/null +++ b/2025/Spatial Transport Optimization by Repositioning Attention Map for Training-Free Text-to-Image Synthesis/layout.json @@ -0,0 +1,7563 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 106, + 103, + 504, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 106, + 103, + 504, + 140 + ], + "spans": [ + { + "bbox": [ + 106, + 103, + 504, + 140 + ], + "type": "text", + "content": "Spatial Transport Optimization by Repositioning Attention Map for Training-Free Text-to-Image Synthesis" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 85, + 161, + 529, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 161, + 529, + 190 + ], + "spans": [ + { + "bbox": [ + 85, + 161, + 529, + 190 + ], + "type": "text", + "content": "Woojung Han Yeonkyung Lee Chanyoung Kim Kwanghyun Park Seong Jae Hwang* Yonsei University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 104, + 191, + 504, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 191, + 504, + 204 + ], + "spans": [ + { + "bbox": [ + 104, + 191, + 504, + 204 + ], + "type": "text", + "content": "{dnwjDDL, yeonkyung.lee, chanyoung, kwanghyun.park, seongjae}@yonsei.ac.kr" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 57, + 236, + 553, + 415 + ], + "blocks": [ + { + "bbox": [ + 57, + 236, + 553, + 415 + ], + "lines": [ + { + "bbox": [ + 57, + 236, + 553, + 415 + ], + "spans": [ + { + "bbox": [ + 57, + 236, + 553, + 415 + ], + "type": "image", + "image_path": "4ac46047851eb13618cbfeaeacfa209b6ec31355d9bf81b5dd48867cccff1c19.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 417, + 555, + 464 + ], + "lines": [ + { + "bbox": [ + 55, + 417, + 555, + 464 + ], + "spans": [ + { + "bbox": [ + 55, + 417, + 555, + 464 + ], + "type": "text", + "content": "Figure 1. Three main challenges in training-free text-to-image (T2I) generation: (1) missing objects, (2) mismatched attributes, and (3) mislocated objects. While existing approaches address missing objects and mismatched attributes, effectively controlling object positioning remains an open problem. Our proposed model, STORM, introduces a dynamic approach to aligning relative object positions throughout the denoising process, enabling precise spatial control without additional spatial templates. Red underline highlights errors made by SD." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 152, + 472, + 200, + 484 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 472, + 200, + 484 + ], + "spans": [ + { + "bbox": [ + 152, + 472, + 200, + 484 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 498, + 297, + 691 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 498, + 297, + 691 + ], + "spans": [ + { + "bbox": [ + 54, + 498, + 297, + 691 + ], + "type": "text", + "content": "Diffusion-based text-to-image (T2I) models have recently excelled in high-quality image generation, particularly in a training-free manner, enabling cost-effective adaptability and generalization across diverse tasks. However, while the existing methods have been continuously focusing on several challenges such as \"missing objects\" and \"mismatched attributes,\" another critical issue of \"mislocated objects\" remains where generated spatial positions fail to align with text prompts. Surprisingly, ensuring such seemingly basic functionality remains challenging in popular T2I models due to the inherent difficulty of imposing explicit spatial guidance via text forms. To address this, we propose STORM (Spatial Transport Optimization by Repositioning Attention Map), a novel training-free approach for spatially coherent T2I synthesis. STORM employs Spatial Transport Optimization (STO), rooted in optimal transport theory, to" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 473, + 556, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 473, + 556, + 581 + ], + "spans": [ + { + "bbox": [ + 313, + 473, + 556, + 581 + ], + "type": "text", + "content": "dynamically adjust object attention maps for precise spatial adherence, supported by a Spatial Transport (ST) Cost function that enhances spatial understanding. Our analysis shows that integrating spatial awareness is most effective in the early denoising stages, while later phases refine details. Extensive experiments demonstrate that STORM surpasses existing methods, effectively mitigating mislocated objects while improving missing and mismatched attributes, setting a new benchmark for spatial alignment in T2I synthesis." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 591, + 394, + 602 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 591, + 394, + 602 + ], + "spans": [ + { + "bbox": [ + 314, + 591, + 394, + 602 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "content": "Diffusion-based text-to-image (T2I) models exhibit remarkable versatility across various applications, generating high-quality images from textual prompts [13, 22, 24, 32, 41]. Notably, training-free methods offer efficiency, real-time adaptability, lower costs, and broad task generalization [5, 6, 9]. These training-free methods leverage their advantages while addressing three key challenges in fully capturing textual details. In particular, as shown in Fig. 1, these challenges include (1) missing objects, (2) mismatched at" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 69, + 703, + 280, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 703, + 280, + 713 + ], + "spans": [ + { + "bbox": [ + 69, + 703, + 280, + 713 + ], + "type": "text", + "content": "Project Page: https://micv-yonsei.github.io/storm2025/" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18401" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 108 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 108 + ], + "type": "text", + "content": "tributes, and (3) mislocated objects. These limitations may lead to images that poorly represent the intended scene or convey misleading information, reducing their reliability." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 110, + 295, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 110, + 295, + 314 + ], + "spans": [ + { + "bbox": [ + 54, + 110, + 295, + 314 + ], + "type": "text", + "content": "While there has been substantial progress in addressing issues like missing objects and mismatched attributes in diffusion-based image synthesis [5, 12, 21, 23], the challenge of mislocated objects remains relatively unexplored despite its critical importance. For example, in Fig. 1, the fifth column shows that the backpack should be below a donut, but Stable Diffusion (SD) [32] generates donut to the left of a backpack. Furthermore, SD consistently generates similar images regardless of spatial cues in the text (e.g., same seed with only positional changes for teddy bear and vase in Fig. 2), indicating limited spatial understanding. This limitation highlights the importance of addressing the problem of mislocated objects and demonstrates the need for more advanced methods of interpreting and aligning with spatial cues. Overcoming these challenges is key to improving T2I model reliability and supporting adaptable applications across diverse settings." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 316, + 295, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 316, + 295, + 483 + ], + "spans": [ + { + "bbox": [ + 55, + 316, + 295, + 483 + ], + "type": "text", + "content": "Existing solutions for mislocated objects often rely on fixed spatial templates or predefined layouts [6, 7, 38], which offer some control but impose rigid constraints and require additional inputs. For instance, these templates may fix an object's position in the image (e.g., far-left corner) instead of offering flexible guidance like placing one object \"to the left\" of another. While effective in structured fields like automated design, these approaches present limitations in more dynamic areas, such as digital art or interactive media, by restricting spatial flexibility. Therefore, instead of relying on restrictive templates or requiring additional inputs, we aim to push the boundaries of T2I models by enabling precise alignment with textual guidance, which is essential to unlocking their full potential across applications." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 486, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 486, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 486, + 295, + 714 + ], + "type": "text", + "content": "To this end, we propose Spatial Transport Optimization by Repositioning Attention Map, STORM, a training-free approach that adjusts relative object positions dynamically, without relying on fixed spatial templates. While SD disregards spatial cues (see Fig. 2), STORM achieves various object positioning with the same model weights, ensuring spatial alignment. To implement STORM, we introduce the Spatial Transport Optimization (STO) framework, which interprets each text token's attention map as a distribution, and leverages Optimal Transport (OT) [37] to efficiently reposition distributions. At an early diffusion timestep, STO adjusts the placement of objects based on their relative positioning (see Fig.3 for attention map changes across the denoising process). While OT considers both distribution and distance to minimize transformation costs, we propose the Spatial Transport Cost (ST Cost) to align with our objectives and guide distribution positioning. This function updates and optimizes the latent vector through backward guidance, following methods from prior work [5, 6]." + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 316, + 69, + 550, + 201 + ], + "blocks": [ + { + "bbox": [ + 316, + 69, + 550, + 201 + ], + "lines": [ + { + "bbox": [ + 316, + 69, + 550, + 201 + ], + "spans": [ + { + "bbox": [ + 316, + 69, + 550, + 201 + ], + "type": "image", + "image_path": "516592f1799a6376dd7c3f34af248e0d527c1e134b0644c6d03cbd802ad9e1e9.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 201, + 555, + 269 + ], + "lines": [ + { + "bbox": [ + 313, + 201, + 555, + 269 + ], + "spans": [ + { + "bbox": [ + 313, + 201, + 555, + 269 + ], + "type": "text", + "content": "Figure 2. Comparison of Spatial Awareness. " + }, + { + "bbox": [ + 313, + 201, + 555, + 269 + ], + "type": "inline_equation", + "content": "\\{\\text{position}^*\\}" + }, + { + "bbox": [ + 313, + 201, + 555, + 269 + ], + "type": "text", + "content": " in each prompt denotes the spatial relationship in each column (e.g., \"to the left of\", \"to the right of\", \"above, and below\"). While Stable Diffusion (SD) shows limited spatial awareness by generating similar images regardless of spatial prompts, our model accurately reflects specified positions. (Same seed for all synthesis)." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 280, + 555, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 280, + 555, + 376 + ], + "spans": [ + { + "bbox": [ + 313, + 280, + 555, + 376 + ], + "type": "text", + "content": "Our analysis suggests that integrating spatial awareness yields optimal results in the early denoising stages, whereas the later stages are better suited for focusing on fine details. Furthermore, we found that missing objects can be resolved concurrently as mislocated objects are corrected in the early steps. Extensive experiments show that our approach outperforms existing methods, setting a new benchmark for spatial alignment in diffusion models." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 323, + 379, + 481, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 379, + 481, + 389 + ], + "spans": [ + { + "bbox": [ + 323, + 379, + 481, + 389 + ], + "type": "text", + "content": "Our main contributions are as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 393, + 553, + 525 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 314, + 393, + 553, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 393, + 553, + 429 + ], + "spans": [ + { + "bbox": [ + 314, + 393, + 553, + 429 + ], + "type": "text", + "content": "- We propose STORM, a training-free method that dynamically adjusts relative object positions, achieving accurate spatial alignment without relying on fixed templates." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 430, + 553, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 430, + 553, + 477 + ], + "spans": [ + { + "bbox": [ + 314, + 430, + 553, + 477 + ], + "type": "text", + "content": "- We introduce the Spatial Transport Optimization (STO) framework based on Optimal Transport (OT), which treats attention maps as distributions and repositions objects relative to each other across diffusion timesteps." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 477, + 553, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 477, + 553, + 525 + ], + "spans": [ + { + "bbox": [ + 314, + 477, + 553, + 525 + ], + "type": "text", + "content": "- We develop ST Cost function that penalizes overlaps and progressively enhances spatial accuracy, demonstrating its effectiveness through extensive experiments that set a new benchmark for spatial alignment in T2I models." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 540, + 400, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 540, + 400, + 552 + ], + "spans": [ + { + "bbox": [ + 314, + 540, + 400, + 552 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 557, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 557, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 557, + 555, + 715 + ], + "type": "text", + "content": "Text-to-Image Synthesis. Text-to-image (T2I) synthesis aims to generate visually realistic images that align with given text prompts. Early studies focused on GANs [18, 19, 27] and autoregressive models [29, 39], but diffusion models [15] have taken the lead, demonstrating remarkable performance in T2I synthesis [22, 32]. Despite advances driven by VLMs [32, 34, 39], aligning generated images with text prompts remains a challenge. To address this, some approaches [2, 30, 34] have expanded network architectures, though these methods require retraining and are difficult to adapt to existing frameworks. Alternatively, other research [5, 9, 21, 23] focused on training-free improvement strategies, which aligns closely with our approach." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18402" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 70, + 293, + 240 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 293, + 240 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 293, + 240 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 293, + 240 + ], + "type": "image", + "image_path": "90b79f3547110a014b5a218fc9829f909e771d7e8a89a6371c54e946fc654ddf.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 244, + 296, + 344 + ], + "lines": [ + { + "bbox": [ + 55, + 244, + 296, + 344 + ], + "spans": [ + { + "bbox": [ + 55, + 244, + 296, + 344 + ], + "type": "text", + "content": "Figure 3. Comparison of Attention Map Progression During Denoising. Visualization of the attention maps for \"a red bird to the right of a green plant\" throughout the denoising process for both Stable Diffusion (a) and our model (b). While Stable Diffusion struggles to distinctly capture the spatial relationship between the bird and the plant, our model effectively aligns the objects according to the specified spatial cue (\"to the right of\"). The resulting image from our model demonstrates improved spatial accuracy compared to Stable Diffusion." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 350, + 296, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 350, + 296, + 578 + ], + "spans": [ + { + "bbox": [ + 56, + 350, + 296, + 578 + ], + "type": "text", + "content": "Training-free T2I Models. Previous studies [1, 5, 9, 12, 21, 22, 31] have explored various approaches to enhance how text prompts are aligned within text-to-image diffusion models. Composable Diffusion [22] and Structure Diffusion [9] aim to improve compositionality through multi-model or cross-attention methods, yet they still face issues like subject mixing and attribute binding. To address these, Attend-and-Excite [5] dynamically adjusts attention during inference to capture all subjects, while Divide-and-Bind [21] improves attribute binding and compositional accuracy with targeted loss functions. CONFORM [23] uses a contrastive method to separate objects in attention maps while keeping related attributes together, and INITNO [12] further tackles spatial misalignment by improving control over object placement to reduce overlap. Although these studies focus on the missing object issue by reducing object overlap, they often overlook the crucial task of capturing accurate relative positioning and directional relationships between objects." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "type": "text", + "content": "Training-free T2I Models for Mislocated Objects. The advantages of attention maps for controlling spatial positioning have led to several studies using fixed templates, such as layouts, to improve object placement [6, 7, 11, 25, 35, 38]. Wu et al. [38] use layout predictors to assign pixel regions, while ZestGuide [7] and Chen et al. [6] leverage cross-attention layers for alignment with segmentations or user-defined layouts. However, these methods often require additional layout inputs and lack flexibility for relative positioning. As a result, there is a need for methods that enhance object placement without relying on predefined layouts." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 314, + 71, + 399, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 71, + 399, + 83 + ], + "spans": [ + { + "bbox": [ + 314, + 71, + 399, + 83 + ], + "type": "text", + "content": "3. Preliminaries" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 85, + 555, + 229 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 85, + 555, + 229 + ], + "spans": [ + { + "bbox": [ + 313, + 85, + 555, + 229 + ], + "type": "text", + "content": "Text Conditioned via Cross-Attention. Our approach uses Stable Diffusion (SD) [32], which generates images from text by integrating text conditioning through cross-attention. A pre-trained CLIP encoder [28] converts the text prompt into a conditioning vector " + }, + { + "bbox": [ + 313, + 85, + 555, + 229 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 85, + 555, + 229 + ], + "type": "text", + "content": ", fed into the UNet model [33] via cross-attention layers at different resolutions. At each timestep " + }, + { + "bbox": [ + 313, + 85, + 555, + 229 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 85, + 555, + 229 + ], + "type": "text", + "content": ", the attention map " + }, + { + "bbox": [ + 313, + 85, + 555, + 229 + ], + "type": "inline_equation", + "content": "A_{t} \\in \\mathbb{R}^{P \\times P \\times N}" + }, + { + "bbox": [ + 313, + 85, + 555, + 229 + ], + "type": "text", + "content": " assigns probabilities for transferring token information to spatial positions, where " + }, + { + "bbox": [ + 313, + 85, + 555, + 229 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 313, + 85, + 555, + 229 + ], + "type": "text", + "content": " is the spatial dimension of the feature map and " + }, + { + "bbox": [ + 313, + 85, + 555, + 229 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 85, + 555, + 229 + ], + "type": "text", + "content": " the number of tokens in the text prompt. In our work, we use " + }, + { + "bbox": [ + 313, + 85, + 555, + 229 + ], + "type": "inline_equation", + "content": "16 \\times 16" + }, + { + "bbox": [ + 313, + 85, + 555, + 229 + ], + "type": "text", + "content": " attention maps as they contain the most semantically meaningful information [14]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 231, + 556, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 231, + 556, + 353 + ], + "spans": [ + { + "bbox": [ + 313, + 231, + 556, + 353 + ], + "type": "text", + "content": "Optimal Transport with Sinkhorn Algorithm. Optimal Transport (OT) [37] provides a framework for transforming one probability distribution into another while preserving spatial relationships. Given two discrete distributions " + }, + { + "bbox": [ + 313, + 231, + 556, + 353 + ], + "type": "inline_equation", + "content": "\\mu = \\{\\mu_u\\}_{u=1}^n" + }, + { + "bbox": [ + 313, + 231, + 556, + 353 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 231, + 556, + 353 + ], + "type": "inline_equation", + "content": "\\nu = \\{\\nu_v\\}_{v=1}^n" + }, + { + "bbox": [ + 313, + 231, + 556, + 353 + ], + "type": "text", + "content": ", each with " + }, + { + "bbox": [ + 313, + 231, + 556, + 353 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 231, + 556, + 353 + ], + "type": "text", + "content": " discrete locations, and a cost matrix " + }, + { + "bbox": [ + 313, + 231, + 556, + 353 + ], + "type": "inline_equation", + "content": "\\mathbf{C} \\in \\mathbb{R}^{n \\times n}" + }, + { + "bbox": [ + 313, + 231, + 556, + 353 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 231, + 556, + 353 + ], + "type": "inline_equation", + "content": "\\mathbf{C}_{uv}" + }, + { + "bbox": [ + 313, + 231, + 556, + 353 + ], + "type": "text", + "content": " is the cost of moving distribution from point " + }, + { + "bbox": [ + 313, + 231, + 556, + 353 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 313, + 231, + 556, + 353 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 231, + 556, + 353 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 313, + 231, + 556, + 353 + ], + "type": "text", + "content": ", OT seeks a transport plan " + }, + { + "bbox": [ + 313, + 231, + 556, + 353 + ], + "type": "inline_equation", + "content": "\\mathbf{P} \\in \\mathbb{R}^{n \\times n}" + }, + { + "bbox": [ + 313, + 231, + 556, + 353 + ], + "type": "text", + "content": " that minimizes the total cost, " + }, + { + "bbox": [ + 313, + 231, + 556, + 353 + ], + "type": "inline_equation", + "content": "\\min_{P \\geq 0} \\sum_{u=1}^{n} \\sum_{v=1}^{n} \\mathbf{C}_{uv} \\mathbf{P}_{uv}" + }, + { + "bbox": [ + 313, + 231, + 556, + 353 + ], + "type": "text", + "content": ". For large " + }, + { + "bbox": [ + 313, + 231, + 556, + 353 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 231, + 556, + 353 + ], + "type": "text", + "content": ", the Sinkhorn algorithm [8] applies entropic regularization for a faster, scalable, and differentiable solution:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 323, + 357, + 553, + 384 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 357, + 553, + 384 + ], + "spans": [ + { + "bbox": [ + 323, + 357, + 553, + 384 + ], + "type": "interline_equation", + "content": "\\min _ {P \\geq 0} \\quad \\sum_ {u = 1} ^ {n} \\sum_ {v = 1} ^ {n} \\mathbf {C} _ {u v} \\mathbf {P} _ {u v} - \\varepsilon \\sum_ {u = 1} ^ {n} \\sum_ {v = 1} ^ {n} \\mathbf {P} _ {u v} (\\log \\mathbf {P} _ {u v} - 1), \\tag {1}", + "image_path": "2aa159bd594d1f304d7bb9959cdccb76f5034e6899683806ba0a3783f6318a70.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 383, + 555, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 383, + 555, + 418 + ], + "spans": [ + { + "bbox": [ + 313, + 383, + 555, + 418 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 383, + 555, + 418 + ], + "type": "inline_equation", + "content": "\\varepsilon > 0" + }, + { + "bbox": [ + 313, + 383, + 555, + 418 + ], + "type": "text", + "content": " balances transport cost and entropy. We further integrate spatial alignment objectives into this OT-based approach in our framework." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 422, + 370, + 434 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 422, + 370, + 434 + ], + "spans": [ + { + "bbox": [ + 314, + 422, + 370, + 434 + ], + "type": "text", + "content": "4. Method" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 437, + 555, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 555, + 510 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 555, + 510 + ], + "type": "text", + "content": "We describe the full pipeline of our model, STORM, as shown in Fig. 4. To support the understanding of our Spatial Transport Optimization (STO) framework, we first provide background information in Sec. 4.1. The full details of STO in Sec. 4.2, while Sec. 4.3 covers optimizations for non-spatial text prompts, and Sec. 4.4 details the update process." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 511, + 432, + 524 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 511, + 432, + 524 + ], + "spans": [ + { + "bbox": [ + 313, + 511, + 432, + 524 + ], + "type": "text", + "content": "4.1. Background of STO" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 525, + 556, + 716 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 525, + 556, + 716 + ], + "spans": [ + { + "bbox": [ + 313, + 525, + 556, + 716 + ], + "type": "text", + "content": "Since objects are generated in regions of high attention within the attention map (see Fig. 3), our goal is to reposition the attention map based on the positional information from the text prompt, using relative objects as a reference. To achieve this, our approach involves three types of attention maps: the attention map of the object to be moved, called the source distribution (e.g., blue area in Fig. 4); the attention map of the relative object, which acts as a reference distribution (e.g., red area in Fig. 4); and the target distribution which represents the intended position for the source distribution to move (e.g., gray area in Fig. 4). The target distribution is an arbitrary distribution defined based on the reference point, which is the centroid of the reference distribution, " + }, + { + "bbox": [ + 313, + 525, + 556, + 716 + ], + "type": "inline_equation", + "content": "(j_{\\mathrm{ref}}, i_{\\mathrm{ref}})" + }, + { + "bbox": [ + 313, + 525, + 556, + 716 + ], + "type": "text", + "content": ". It primarily serves to guide the intended position of the source distribution using reference points without exerting significant direct influence." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18403" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 71, + 549, + 259 + ], + "blocks": [ + { + "bbox": [ + 58, + 71, + 549, + 259 + ], + "lines": [ + { + "bbox": [ + 58, + 71, + 549, + 259 + ], + "spans": [ + { + "bbox": [ + 58, + 71, + 549, + 259 + ], + "type": "image", + "image_path": "8096962e3fc7166fc54cbbc04e1742ffb70c4eac9855ddc3a881a154e038120b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 261, + 555, + 340 + ], + "lines": [ + { + "bbox": [ + 55, + 261, + 555, + 340 + ], + "spans": [ + { + "bbox": [ + 55, + 261, + 555, + 340 + ], + "type": "text", + "content": "Figure 4. Overview pipeline of STORM. Our method leverages Optimal Transport in a training-free manner, allowing the model to accurately reflect relative object positions at each step without additional inputs. Given the prompt \"A car to the left of an elephant\", our method dynamically adjusts the attention maps to induce the specified spatial relationship. The process starts with initial attention maps for the \"car\" and \"elephant\" at time step " + }, + { + "bbox": [ + 55, + 261, + 555, + 340 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 55, + 261, + 555, + 340 + ], + "type": "text", + "content": ". Using the centroids of these attention maps, the Spatial Transport Optimization (STO) computes the losses to correct positional relationships (e.g., ensuring the car is to the left of the elephant). The updated attention map is then used to refine the latent representation " + }, + { + "bbox": [ + 55, + 261, + 555, + 340 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 55, + 261, + 555, + 340 + ], + "type": "text", + "content": ", leading to a final image that adheres to the desired spatial arrangement. The comparison of attention maps (before and after STO) shows improved alignment, effectively placing the car to the left of the elephant as instructed in the prompt." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 349, + 284, + 361 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 349, + 284, + 361 + ], + "spans": [ + { + "bbox": [ + 55, + 349, + 284, + 361 + ], + "type": "text", + "content": "4.2. Spatial Transport Optimization Framework" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 363, + 295, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 363, + 295, + 567 + ], + "spans": [ + { + "bbox": [ + 54, + 363, + 295, + 567 + ], + "type": "text", + "content": "In this section, we detail our Spatial Transport Optimization (STO) framework, which dynamically repositions attention maps for accurate object alignment without modifying model weights. This framework consists of two main components: a cost function and a transport plan. The cost function directs the distribution shift from the source in the intended direction, allowing the attention map to adjust positionally according to spatial relationships. Although we build on standard cost functions that measure distance between locations, our Spatial Transport Cost (ST Cost) incorporates additional objectives adapted to our specific spatial alignment goals. The transport plan, which optimally moves the distribution between points, is implemented using the Sinkhorn algorithm. Beginning with an initial attention map, STO iteratively updates the latent vector by applying STO-derived losses, gradually guiding the distribution into the target alignment." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 570, + 244, + 582 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 244, + 582 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 244, + 582 + ], + "type": "text", + "content": "4.2.1 Construction of Target Distributions" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "content": "Our approach focuses on relative positioning adjustments at each timestep, using an arbitrary target distribution (e.g., gray area in Fig. 4) to guide the source distribution based on the reference. Depending on the text-specified direction (e.g., horizontal or vertical), only the relevant reference point of " + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "inline_equation", + "content": "(j_{\\mathrm{ref}},i_{\\mathrm{ref}})" + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "content": " is considered: for 'left' or 'right' positioning, only the horizontal dimension of the reference point " + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "inline_equation", + "content": "j_{\\mathrm{ref}}" + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "content": " is used, while for 'above' or 'below' positioning, only the vertical dimension " + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "inline_equation", + "content": "i_{\\mathrm{ref}}" + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "content": " is considered, allowing the flexibility in the opposite dimension. For example, in Fig. 4a, with the car attention map as a source distribution and the" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 350, + 555, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 350, + 555, + 410 + ], + "spans": [ + { + "bbox": [ + 313, + 350, + 555, + 410 + ], + "type": "text", + "content": "elephant attention map as a reference distribution, the target distribution is placed to the left of a elephant's distribution. This approach similarly applies to another source. Supplementary Materials provide details on computing the target distribution and reference point." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 415, + 485, + 426 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 415, + 485, + 426 + ], + "spans": [ + { + "bbox": [ + 313, + 415, + 485, + 426 + ], + "type": "text", + "content": "4.2.2 Spatial Transport Cost Function" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 426, + 555, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 555, + 533 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 555, + 533 + ], + "type": "text", + "content": "In the STO framework, we aim to find an optimal transport plan that minimizes the cost of transferring the attention map. The cost function measures the expense of moving distribution mass by computing distances between points, where a higher value indicates less desirable locations. While our overall cost function merges the standard OT loss with an additional term (detailed in the supplementary material), here we focus on our newly introduced ST Cost, which is built on two core principles:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 533, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 556, + 713 + ], + "type": "text", + "content": "(I) Positional Cost: We developed a cost function to guide the distribution to its intended position relative to the reference point. In simple terms, when the source distribution is intended to be on the left side of the reference point, it has a low cost when positioned on the left and a high cost when positioned on the right. To quantify this directional positioning, we define a set of positional " + }, + { + "bbox": [ + 313, + 533, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 313, + 533, + 556, + 713 + ], + "type": "text", + "content": " values, representing the distances between the current point and the reference point in each of the four directions. Specifically, the " + }, + { + "bbox": [ + 313, + 533, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 313, + 533, + 556, + 713 + ], + "type": "text", + "content": " values are defined as follows: " + }, + { + "bbox": [ + 313, + 533, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\delta_{ij}^{\\leftarrow} = j_{\\mathrm{ref}} - j,\\delta_{ij}^{\\rightarrow} = j - j_{\\mathrm{ref}},\\delta_{ij}^{\\uparrow} = i_{\\mathrm{ref}} - i,\\delta_{ij}^{\\downarrow} = i - i_{\\mathrm{ref}}" + }, + { + "bbox": [ + 313, + 533, + 556, + 713 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 313, + 533, + 556, + 713 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 533, + 556, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 533, + 556, + 713 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 533, + 556, + 713 + ], + "type": "text", + "content": " representing the current position values, and the arrows indicate the desired and restricted direction for positioning the source attention map relative to the reference point. When constructing the overall cost function that incorporates all directional " + }, + { + "bbox": [ + 313, + 533, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 313, + 533, + 556, + 713 + ], + "type": "text", + "content": " values, each direction is categorized as either de" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18404" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 163 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 163 + ], + "type": "text", + "content": "sired or restricted. The desired direction " + }, + { + "bbox": [ + 55, + 72, + 294, + 163 + ], + "type": "inline_equation", + "content": "(\\delta^{\\mathrm{des}})" + }, + { + "bbox": [ + 55, + 72, + 294, + 163 + ], + "type": "text", + "content": " corresponds to the intended movement of the source distribution, while the restricted direction " + }, + { + "bbox": [ + 55, + 72, + 294, + 163 + ], + "type": "inline_equation", + "content": "(\\delta^{\\mathrm{res}})" + }, + { + "bbox": [ + 55, + 72, + 294, + 163 + ], + "type": "text", + "content": " represents the opposite or undesired direction. For instance, if the source distribution should move to the left, " + }, + { + "bbox": [ + 55, + 72, + 294, + 163 + ], + "type": "inline_equation", + "content": "\\delta_{ij}^{\\leftarrow}" + }, + { + "bbox": [ + 55, + 72, + 294, + 163 + ], + "type": "text", + "content": " is marked as the desired direction " + }, + { + "bbox": [ + 55, + 72, + 294, + 163 + ], + "type": "inline_equation", + "content": "(\\delta^{\\mathrm{des}})" + }, + { + "bbox": [ + 55, + 72, + 294, + 163 + ], + "type": "text", + "content": ", while " + }, + { + "bbox": [ + 55, + 72, + 294, + 163 + ], + "type": "inline_equation", + "content": "\\delta_{ij}^{\\rightarrow}" + }, + { + "bbox": [ + 55, + 72, + 294, + 163 + ], + "type": "text", + "content": " is treated as the restricted direction " + }, + { + "bbox": [ + 55, + 72, + 294, + 163 + ], + "type": "inline_equation", + "content": "(\\delta^{\\mathrm{res}})" + }, + { + "bbox": [ + 55, + 72, + 294, + 163 + ], + "type": "text", + "content": ". The combined cost function, accounting for both desired and restricted positions, can be expressed as:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 166, + 295, + 190 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 166, + 295, + 190 + ], + "spans": [ + { + "bbox": [ + 61, + 166, + 295, + 190 + ], + "type": "interline_equation", + "content": "\\Delta_ {i j} \\left(\\delta_ {i j} ^ {\\mathrm {d e s}}, \\delta_ {i j} ^ {\\mathrm {r e s}}\\right) = \\frac {1}{\\omega \\left(\\delta_ {i j} ^ {\\mathrm {d e s}} + \\epsilon\\right)} \\mathbb {1} _ {+} \\left[ \\delta_ {i j} ^ {\\mathrm {d e s}} \\right] + \\omega \\left(\\delta_ {i j} ^ {\\mathrm {r e s}} + \\epsilon\\right) \\mathbb {1} _ {+} \\left[ \\delta_ {i j} ^ {\\mathrm {r e s}} \\right], \\tag {2}", + "image_path": "dcd5e7fc9eff2b2529dcf7c537774b04cd677c41c6eb1f399276f55922e0f89d.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 196, + 296, + 292 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 196, + 296, + 292 + ], + "spans": [ + { + "bbox": [ + 55, + 196, + 296, + 292 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 196, + 296, + 292 + ], + "type": "inline_equation", + "content": "\\omega (\\cdot) > 1" + }, + { + "bbox": [ + 55, + 196, + 296, + 292 + ], + "type": "text", + "content": " is a progressive adaptive weight that controls the alignment importance across timesteps, and " + }, + { + "bbox": [ + 55, + 196, + 296, + 292 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 55, + 196, + 296, + 292 + ], + "type": "text", + "content": " is a stabilizing factor. " + }, + { + "bbox": [ + 55, + 196, + 296, + 292 + ], + "type": "inline_equation", + "content": "\\mathbb{1}_{+}" + }, + { + "bbox": [ + 55, + 196, + 296, + 292 + ], + "type": "text", + "content": " is an indicator function that equals 1 when the input is positive and 0 otherwise, allowing selective application of computation based on positivity. As the source distribution aligns more closely with the desired direction, the " + }, + { + "bbox": [ + 55, + 196, + 296, + 292 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 55, + 196, + 296, + 292 + ], + "type": "text", + "content": " value approaches zero, indicating minimal cost for correctly aligned positioning (see Example below)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 296, + 296, + 438 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 296, + 296, + 438 + ], + "spans": [ + { + "bbox": [ + 55, + 296, + 296, + 438 + ], + "type": "text", + "content": "(II) Non-overlap Cost: In this process, it is crucial to prevent the source distribution from overlapping with the reference distribution during its movement. This is one of the primary issues in addressing the missing object problem in SD, where avoiding overlap is essential for ensuring that each object is distinctly represented. To achieve this, we simply embed the reference distribution directly into the cost function, guiding objects to occupy separate locations. Specifically, with " + }, + { + "bbox": [ + 55, + 296, + 296, + 438 + ], + "type": "inline_equation", + "content": "A_{ij}" + }, + { + "bbox": [ + 55, + 296, + 296, + 438 + ], + "type": "text", + "content": " representing the attention weight of the reference distribution, we incorporated this distribution into the cost function, assigning high costs to regions occupied by the reference distribution and low costs elsewhere." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 439, + 295, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 439, + 295, + 460 + ], + "spans": [ + { + "bbox": [ + 55, + 439, + 295, + 460 + ], + "type": "text", + "content": "Combining this principle with our core ideas, our cost function can be defined as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 127, + 469, + 294, + 483 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 469, + 294, + 483 + ], + "spans": [ + { + "bbox": [ + 127, + 469, + 294, + 483 + ], + "type": "interline_equation", + "content": "C _ {i j} = A _ {i j} \\Delta_ {i j} \\left(\\delta_ {i j} ^ {\\mathrm {d e s}}, \\delta_ {i j} ^ {\\mathrm {r e s}}\\right). \\tag {3}", + "image_path": "85a2cac5766ec2065270d23d1f20bd5035c3ca68461555e7d04673608bbd4346.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 489, + 296, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 489, + 296, + 657 + ], + "spans": [ + { + "bbox": [ + 55, + 489, + 296, + 657 + ], + "type": "text", + "content": "Example. In Fig. 4, the car should be left of the elephant, and the elephant to the right of the car. The cost functions for each of them are represented by " + }, + { + "bbox": [ + 55, + 489, + 296, + 657 + ], + "type": "inline_equation", + "content": "C_{ij}^{\\leftarrow} = A_{ij}\\Delta_{ij}(\\delta_{ij}^{\\leftarrow},\\delta_{ij}^{\\rightarrow})" + }, + { + "bbox": [ + 55, + 489, + 296, + 657 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 489, + 296, + 657 + ], + "type": "inline_equation", + "content": "C_{ij}^{\\rightarrow} = A_{ij}\\Delta_{ij}(\\delta_{ij}^{\\rightarrow},\\delta_{ij}^{\\leftarrow})" + }, + { + "bbox": [ + 55, + 489, + 296, + 657 + ], + "type": "text", + "content": ", respectively. Taking the car as an example, it needs to be positioned to the left of the elephant relative to the elephant's centroid point. In more detail, if the current point is positioned to the left of the centroid, the " + }, + { + "bbox": [ + 55, + 489, + 296, + 657 + ], + "type": "inline_equation", + "content": "\\delta_{ij}^{\\mathrm{des}}" + }, + { + "bbox": [ + 55, + 489, + 296, + 657 + ], + "type": "text", + "content": " is activated; if positioned to the right of the centroid, the " + }, + { + "bbox": [ + 55, + 489, + 296, + 657 + ], + "type": "inline_equation", + "content": "\\delta_{ij}^{\\mathrm{res}}" + }, + { + "bbox": [ + 55, + 489, + 296, + 657 + ], + "type": "text", + "content": " is activated. Additionally, as the car moves to the left of the elephant, it should not overlap the elephant it passes along the way. The elephant's attention map is used in the cost function, guiding the car to avoid overlapping with the elephant's area. The resulting " + }, + { + "bbox": [ + 55, + 489, + 296, + 657 + ], + "type": "inline_equation", + "content": "C_{ij}" + }, + { + "bbox": [ + 55, + 489, + 296, + 657 + ], + "type": "text", + "content": " is utilized as the cost function for STO." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 666, + 271, + 677 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 666, + 271, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 666, + 271, + 677 + ], + "type": "text", + "content": "4.2.3 Sinkhorn Algorithm-based Transport Plan" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "content": "After combining the standard OT approach with ST components (see Supplementary Material for the complete formulation), we compute the transport plan using the Sinkhorn" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "text", + "content": "algorithm [8], which applies entropic regularization for computational efficiency and stability, especially for high-dimensional attention maps. The regularized OT problem is formulated as in Eq. (1) using target distribution calculated in Sec. 4.2.1. The Sinkhorn algorithm iteratively updates transport plan " + }, + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "text", + "content": " to meet the marginal constraints of source and target distributions, applying the cost function bidirectionally for both objects." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 171, + 525, + 183 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 171, + 525, + 183 + ], + "spans": [ + { + "bbox": [ + 313, + 171, + 525, + 183 + ], + "type": "text", + "content": "4.3. Optimization for Objects and Attributes" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 184, + 553, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 184, + 553, + 363 + ], + "spans": [ + { + "bbox": [ + 313, + 184, + 553, + 363 + ], + "type": "text", + "content": "Our method also extends to cases where the text lacks spatial information. (i) Object-level Optimization: We handle cases where the text prompt specifies objects without explicit spatial guidance by randomly placing the target distribution of each object. This flexible positioning encourages diverse object placement when spatial relationships are not provided. Therefore, the cost function is simplified to focus on solely on preventing overlap, expressed as " + }, + { + "bbox": [ + 313, + 184, + 553, + 363 + ], + "type": "inline_equation", + "content": "C_{ij} = A_{ij} \\otimes \\mathbf{1}_N" + }, + { + "bbox": [ + 313, + 184, + 553, + 363 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 184, + 553, + 363 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 313, + 184, + 553, + 363 + ], + "type": "text", + "content": " is the reference attention map. (ii) Attribute Matching Optimization: The attention map for an object and its corresponding attribute should be identical, reflecting the same attention focus. Therefore, we used the standard OT cost function, calculating Euclidean distance between corresponding positions across the object, and attribute attention maps to quantify differences." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 368, + 495, + 380 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 368, + 495, + 380 + ], + "spans": [ + { + "bbox": [ + 313, + 368, + 495, + 380 + ], + "type": "text", + "content": "4.4. Optimization and Update Process" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 382, + 553, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 382, + 553, + 453 + ], + "spans": [ + { + "bbox": [ + 313, + 382, + 553, + 453 + ], + "type": "text", + "content": "We update the latent vector using STO-based losses to guide the spatial alignment of the attention maps in a differentiable manner. At each timestep, we compute the OT loss " + }, + { + "bbox": [ + 313, + 382, + 553, + 453 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 313, + 382, + 553, + 453 + ], + "type": "text", + "content": " and update the latent vector " + }, + { + "bbox": [ + 313, + 382, + 553, + 453 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 313, + 382, + 553, + 453 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 313, + 382, + 553, + 453 + ], + "type": "inline_equation", + "content": "z_{t}^{\\prime}\\gets z_{t} - \\alpha_{t}\\cdot \\nabla z_{t}\\mathcal{L}" + }, + { + "bbox": [ + 313, + 382, + 553, + 453 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 313, + 382, + 553, + 453 + ], + "type": "inline_equation", + "content": "\\alpha_{t}" + }, + { + "bbox": [ + 313, + 382, + 553, + 453 + ], + "type": "text", + "content": " is the step size. After the update, a forward pass computes " + }, + { + "bbox": [ + 313, + 382, + 553, + 453 + ], + "type": "inline_equation", + "content": "z_{t - 1}" + }, + { + "bbox": [ + 313, + 382, + 553, + 453 + ], + "type": "text", + "content": " for the next denoising step." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 456, + 554, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 456, + 554, + 588 + ], + "spans": [ + { + "bbox": [ + 313, + 456, + 554, + 588 + ], + "type": "text", + "content": "Remark. As shown in Fig.8, addressing the mislocated object problem is most effective in early timesteps, which, according to related work[26], establish the overall image structure while later timesteps refine finer details. Based on this, STORM establishes object positioning and spatial configurations early in the denoising process, minimizing the risk of misplacements and overlaps. Furthermore, as shown in the bottom row of Fig. 8, solving the mislocated object problem before addressing missing objects can naturally resolve the latter, highlighting the importance of prioritizing mislocated objects as a critical step in the synthesis process." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 601, + 394, + 613 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 601, + 394, + 613 + ], + "spans": [ + { + "bbox": [ + 313, + 601, + 394, + 613 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 617, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 554, + 713 + ], + "type": "text", + "content": "Implementation Details. Following methods established in prior research [5, 21, 23], we conducted a comprehensive evaluation of our model. We used the official Stable Diffusion models [32] v1.4 and v2.1, configuring the denoising timestep to 50, and the guidance scale to 7.5 in line with common practices in the field. To smooth the cross-attention map, a Gaussian filter with a kernel size of 3 and a standard deviation of 0.5 was applied. Optimization" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18405" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 137, + 552, + 361 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 137 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 137 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 137 + ], + "type": "text", + "content": "Table 1. Performance comparison between different models on VISOR (\\%) and Object Accuracy (OA) (\\%) metrics, based on Stable Diffusion 1.4 and Stable Diffusion 2.1 [32]. For entries marked with " + }, + { + "bbox": [ + 55, + 70, + 555, + 137 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 55, + 70, + 555, + 137 + ], + "type": "text", + "content": ", official values were unavailable; thus, we directly calculated these values. Bold values indicate the best scores, while underlined values indicate the second-best scores among the fair comparison. Notes on VISOR. VISOR cond: spatial accuracy using the results in which both objects are sufficiently generated; VISOR uncond: spatial accuracy using all results, including results with failed object generations; VISOR " + }, + { + "bbox": [ + 55, + 70, + 555, + 137 + ], + "type": "inline_equation", + "content": "1/2/3/4" + }, + { + "bbox": [ + 55, + 70, + 555, + 137 + ], + "type": "text", + "content": ": ratios of one, two, three, or four spatially accurate results out of four seed results, respectively. (e.g., " + }, + { + "bbox": [ + 55, + 70, + 555, + 137 + ], + "type": "inline_equation", + "content": "\\mathrm{VISOR}_1" + }, + { + "bbox": [ + 55, + 70, + 555, + 137 + ], + "type": "text", + "content": ": % of results with exactly one spatially correct result out of four generated results)" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 137, + 552, + 361 + ], + "lines": [ + { + "bbox": [ + 58, + 137, + 552, + 361 + ], + "spans": [ + { + "bbox": [ + 58, + 137, + 552, + 361 + ], + "type": "table", + "html": "
ModelVenueExtra TrainingExtra InputOA (%)VISOR (%)
uncondcond1234
Stable Diffusion 1.4 Based
SD 1.4 [32]-X29.8618.8162.9846.6020.116.891.63
SD 1.4 + CDM [22]ECCV22✓X23.2714.9964.4139.4414.564.841.12
GLIDE [24]ICML22✓X3.361.9859.066.721.020.170.03
GLIDE + CDM [22]ECCV22✓X10.176.4363.2120.074.690.830.11
Control-GPT [42]arXiv23✓✓48.3344.1765.9769.8051.2035.6720.48
Layout Guidance [6]WACV24X✓40.0138.8095.95----
REVISION [3]ECCV24X✓53.9652.7197.6977.7961.0244.9027.15
Structure Diffusion [9]ICLR23XX28.6517.8762.3644.7018.736.571.46
Attend-and-Excite [5]SIGGRAPH23XX42.0725.7561.2149.2919.334.560.08
Divide-and-Bind†[21]BMVC24XX46.0331.6268.7064.7237.8218.645.30
INITNO†[12]CVPR24XX60.4035.1858.2471.2042.7120.096.72
CONFORM†[23]CVPR24XX60.7338.4862.3373.0145.8225.579.52
Ours (SD 1.4)XX61.0157.5894.3985.9369.7149.0125.70
Stable Diffusion 2.1 Based
SD 2.1 [32]-X47.8330.2563.2464.4235.7416.134.70
SPRIGHT [4]ECCV24✓X60.6842.2371.2471.7851.8833.0916.15
REVISION [3]ECCV24X✓48.2647.1197.6176.0755.7537.1019.53
Ours (SD 2.1)XX62.5559.3594.8888.3471.7552.0325.42
", + "image_path": "855e16195c9c9137b7cf42571bd23f7c7ccc4dcc00637fcc9f96d36e00520bf4.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 372, + 295, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 372, + 295, + 411 + ], + "spans": [ + { + "bbox": [ + 55, + 372, + 295, + 411 + ], + "type": "text", + "content": "was performed iteratively with forward updates at specified timesteps (" + }, + { + "bbox": [ + 55, + 372, + 295, + 411 + ], + "type": "inline_equation", + "content": "t \\in 5, 10, 15, 20" + }, + { + "bbox": [ + 55, + 372, + 295, + 411 + ], + "type": "text", + "content": "), stopping at " + }, + { + "bbox": [ + 55, + 372, + 295, + 411 + ], + "type": "inline_equation", + "content": "t = 25" + }, + { + "bbox": [ + 55, + 372, + 295, + 411 + ], + "type": "text", + "content": " to enhance quality. All experiments use NVIDIA RTX A6000 Ada*." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 420, + 167, + 431 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 420, + 167, + 431 + ], + "spans": [ + { + "bbox": [ + 55, + 420, + 167, + 431 + ], + "type": "text", + "content": "5.1. Evaluation Results" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 434, + 295, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 434, + 295, + 590 + ], + "spans": [ + { + "bbox": [ + 54, + 434, + 295, + 590 + ], + "type": "text", + "content": "We evaluated our method against existing training-free T2I models using both qualitative and quantitative measures, focusing on two key aspects. (i) Spatial Accuracy: To measure alignment with spatial configurations specified in prompts, we used the VISOR [10] and the T2I-CompBench [17], both provide a comprehensive assessment of spatial relationships captured by the model. (ii) Object and Attribute Consistency: We assessed object and attribute representation using T2I-CompBench in combination with CLIP-space distance [28] and TIFA score [16], following established methods [5, 21, 23]. We provide detailed descriptions of each evaluation metric and datasets in the Supplementary Materials for additional clarity." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 594, + 295, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 594, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 54, + 594, + 295, + 689 + ], + "type": "text", + "content": "Quantitative Evaluation: VISOR. To evaluate spatial accuracy, we use the VISOR [10] to measure the ability of a model to position objects based on spatial cues (e.g., above) in text prompts. VISOR benchmark contains over 25K prompts describing spatial relationships between objects. Notably, as seen in Table 1, the increase in " + }, + { + "bbox": [ + 54, + 594, + 295, + 689 + ], + "type": "inline_equation", + "content": "\\mathrm{VISOR}_4" + }, + { + "bbox": [ + 54, + 594, + 295, + 689 + ], + "type": "text", + "content": " demonstrates the model's reliability in generating spatially accurate images across prompts and seeds." + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 316, + 403, + 553, + 512 + ], + "blocks": [ + { + "bbox": [ + 313, + 371, + 555, + 403 + ], + "lines": [ + { + "bbox": [ + 313, + 371, + 555, + 403 + ], + "spans": [ + { + "bbox": [ + 313, + 371, + 555, + 403 + ], + "type": "text", + "content": "Table 2. Comparison of methods on T2I-CompBench, with attribute binding and spatial relationships calculated using models based on Stable Diffusion 1.4 and 2.1." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 316, + 403, + 553, + 512 + ], + "lines": [ + { + "bbox": [ + 316, + 403, + 553, + 512 + ], + "spans": [ + { + "bbox": [ + 316, + 403, + 553, + 512 + ], + "type": "table", + "html": "
MethodAttribute Binding (Color ↑)Attribute Binding (Shape ↑)Attribute Binding (Texture ↑)Object Relationship (Spatial ↑)
Stable Diffusion 1.4 Based
Stable Diffusion-v1.4 [32]0.37650.35760.41560.1246
Ours (SD 1.4)0.64580.59830.75390.1613
Stable Diffusion 2.1 Based
Stable Diffusion-v2.1 [32]0.50650.42210.49220.1342
Composable Diffusion [22]0.40630.32990.36450.0800
Structured Diffusion [9]0.49900.42180.49000.1386
Attend-and-Excite [5]0.64000.45170.59630.1455
Ours (SD 2.1)0.67770.62260.78840.1981
", + "image_path": "72ea1ed6d83bee0e516d508fa5acedd7097248e3268d1b20eeb4fb7535b4a2da.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 518, + 555, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 518, + 555, + 662 + ], + "spans": [ + { + "bbox": [ + 313, + 518, + 555, + 662 + ], + "type": "text", + "content": "Quantitative Evaluation: T2I-CompBench. The T2I-CompBench [17] evaluates object presence, spatial relationships, and attribute alignment from text prompts. We used the attribute and spatial configuration datasets provided by T2I-CompBench, averaging results over more than 10 randomly selected seeds for each dataset. Attribute Binding is evaluated with the BLIP-VQA metric [20], which poses separate questions for each attribute-object pair in the prompt. UniDet-based metric [43] is used for spatial relationships. In Table 2, our score shows that the generated images position the specified objects correctly and accurately reflect the given attributes (e.g., color, shape, texture)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "content": "Quantitative Evaluation: CLIP Score and TIFA Score. Following previous work [5, 12, 23], we evaluate our method using CLIP Image-Text and Text-Text Similarity, generating 64 images per prompt and averaging CLIP co" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 70, + 703, + 266, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 703, + 266, + 712 + ], + "spans": [ + { + "bbox": [ + 70, + 703, + 266, + 712 + ], + "type": "text", + "content": "* Advanced Database System Infrastructure(NFEC-2024-11-300458)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18406" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 70, + 550, + 368 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 550, + 368 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 550, + 368 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 550, + 368 + ], + "type": "image", + "image_path": "6f6563efb61becaf78ec9b87df5e987ca8f097923a40e1b4f8c2da7eda98e414.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 369, + 555, + 392 + ], + "lines": [ + { + "bbox": [ + 55, + 369, + 555, + 392 + ], + "spans": [ + { + "bbox": [ + 55, + 369, + 555, + 392 + ], + "type": "text", + "content": "Figure 5. Qualitative comparison across the custom prompt, which involves attribute and positional information in text, evaluating previous state-of-the-art training-free T2I methods, Attend&Excite [5], Divide&Bind [21], INITNO [12], CONFORM [23], and ours." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 57, + 421, + 294, + 485 + ], + "blocks": [ + { + "bbox": [ + 55, + 398, + 295, + 420 + ], + "lines": [ + { + "bbox": [ + 55, + 398, + 295, + 420 + ], + "spans": [ + { + "bbox": [ + 55, + 398, + 295, + 420 + ], + "type": "text", + "content": "Table 3. User study evaluating model performance on object synthesis, attribute matching, spatial correctness, and overall fidelity." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 421, + 294, + 485 + ], + "lines": [ + { + "bbox": [ + 57, + 421, + 294, + 485 + ], + "spans": [ + { + "bbox": [ + 57, + 421, + 294, + 485 + ], + "type": "table", + "html": "
MethodObject Accuracy (%)Attribute Matching (%)Spatial Correctness (%)Overall Fidelity (%)
Stable Diffusion [32]14.6814.3111.8115.02
Attend-and-Excite [5]16.3416.5013.2016.04
Divide-and-Bind [21]16.3915.6913.0116.52
INITNO [12]15.9816.0313.1515.63
CONFORM [23]15.6416.7912.9114.14
Ours20.9720.6835.9222.65
", + "image_path": "587f27950643b7e7cf1ae84dcc5a10e4ec2b0cf6ab7dd4b47e10ae1605e716b0.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 495, + 295, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 495, + 295, + 685 + ], + "spans": [ + { + "bbox": [ + 54, + 495, + 295, + 685 + ], + "type": "text", + "content": "sine similarity scores across methods. Fig. 7 reports two CLIP Image-Text Similarity metrics: full prompt similarity (cosine similarity between the entire prompt and generated image) and minimum object similarity (minimum similarity between the generated image and each subject prompt) to see the accuracy of each object. We generate captions for the images with BLIP [20] and compute cosine similarity with the original prompt. Our method achieves comparable performance to other leading methods. However, as the reliability of the CLIP score is limited by its \"bag-of-words\" approach [40] (as noted by [5]), we also use the TIFA metric [16] for improved evaluation of complex prompts. TIFA [16] evaluates image faithfulness to the input prompt by generating question-answer pairs using the LLaMA 2 [36] model. Our method performs well compared to other state-of-the-art models." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "type": "text", + "content": "Quantitative Evaluation: User Studies. We create 10 custom text prompts with detailed objects, attributes, and spa" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 320, + 402, + 545, + 556 + ], + "blocks": [ + { + "bbox": [ + 320, + 402, + 545, + 556 + ], + "lines": [ + { + "bbox": [ + 320, + 402, + 545, + 556 + ], + "spans": [ + { + "bbox": [ + 320, + 402, + 545, + 556 + ], + "type": "image", + "image_path": "2bb58b2f12a24e2ab41514c50056e828d9ee092f414d32d83dd9037591a9e270.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 558, + 555, + 581 + ], + "lines": [ + { + "bbox": [ + 313, + 558, + 555, + 581 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 555, + 581 + ], + "type": "text", + "content": "Figure 6. Generated images with complex positioning details in the prompt (Additional results in Supplementary Materials)." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 590, + 555, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 590, + 555, + 651 + ], + "spans": [ + { + "bbox": [ + 313, + 590, + 555, + 651 + ], + "type": "text", + "content": "tial information, then generate images using random seeds. With 30 participants, our user study evaluates (1) object accuracy, (2) attribute matching, (3) spatial correctness, and (4) overall fidelity. In Table 3, our model shows surpassing results across the existing models of all categories." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 653, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 556, + 715 + ], + "type": "text", + "content": "Qualitative Analysis. Fig. 5 provides a qualitative comparison with recent state-of-the-art models [12, 23]. To compare attribute matching and object localization accuracy, we create custom text prompts to evaluate the qualitative performance of our model. Existing methods often misalign" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18407" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 70, + 319, + 152 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 319, + 152 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 319, + 152 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 319, + 152 + ], + "type": "image", + "image_path": "aa791de998ec864ce81dec2e93acea1b5a7b3be5e8b45f5e07858bccb7380be6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 154, + 555, + 176 + ], + "lines": [ + { + "bbox": [ + 55, + 154, + 555, + 176 + ], + "spans": [ + { + "bbox": [ + 55, + 154, + 555, + 176 + ], + "type": "text", + "content": "Figure 7. Comparison of various models on CLIP Image-Text, CLIP Text-Text, and TIFA scores across prompts involving Animal-Animal, Animal-Object, and Object-Object pairs. (The seed is randomly selected.)" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 323, + 71, + 550, + 152 + ], + "blocks": [ + { + "bbox": [ + 323, + 71, + 550, + 152 + ], + "lines": [ + { + "bbox": [ + 323, + 71, + 550, + 152 + ], + "spans": [ + { + "bbox": [ + 323, + 71, + 550, + 152 + ], + "type": "image", + "image_path": "9e207de35df137f1d0c9f77e81141e5f34d45b2f925b244536d31e132b1062d3.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 58, + 190, + 105, + 228 + ], + "blocks": [ + { + "bbox": [ + 58, + 190, + 105, + 228 + ], + "lines": [ + { + "bbox": [ + 58, + 190, + 105, + 228 + ], + "spans": [ + { + "bbox": [ + 58, + 190, + 105, + 228 + ], + "type": "image", + "image_path": "74600838616d46d450a3a651c4ebed241cdabee538f6c15666c23f3d848a38fe.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 59, + 228, + 102, + 236 + ], + "lines": [ + { + "bbox": [ + 59, + 228, + 102, + 236 + ], + "spans": [ + { + "bbox": [ + 59, + 228, + 102, + 236 + ], + "type": "text", + "content": "Exp.#A0 (SD)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 106, + 190, + 153, + 228 + ], + "blocks": [ + { + "bbox": [ + 113, + 182, + 237, + 190 + ], + "lines": [ + { + "bbox": [ + 113, + 182, + 237, + 190 + ], + "spans": [ + { + "bbox": [ + 113, + 182, + 237, + 190 + ], + "type": "text", + "content": "\"A cake to the left of a tie\"" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 106, + 190, + 153, + 228 + ], + "lines": [ + { + "bbox": [ + 106, + 190, + 153, + 228 + ], + "spans": [ + { + "bbox": [ + 106, + 190, + 153, + 228 + ], + "type": "image", + "image_path": "6872d1bd4154ec196101ba94c8f7256b1c13f9735dc134ab3d36c0d12397652d.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 115, + 228, + 141, + 236 + ], + "lines": [ + { + "bbox": [ + 115, + 228, + 141, + 236 + ], + "spans": [ + { + "bbox": [ + 115, + 228, + 141, + 236 + ], + "type": "text", + "content": "Exp.#A1" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 154, + 190, + 200, + 228 + ], + "blocks": [ + { + "bbox": [ + 154, + 190, + 200, + 228 + ], + "lines": [ + { + "bbox": [ + 154, + 190, + 200, + 228 + ], + "spans": [ + { + "bbox": [ + 154, + 190, + 200, + 228 + ], + "type": "image", + "image_path": "b09aef1acabae910e6c694cd23ee3d9f1b474fd6538fbf633845ecc9fffa204d.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 162, + 228, + 189, + 236 + ], + "lines": [ + { + "bbox": [ + 162, + 228, + 189, + 236 + ], + "spans": [ + { + "bbox": [ + 162, + 228, + 189, + 236 + ], + "type": "text", + "content": "Exp.#A2" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 200, + 190, + 246, + 228 + ], + "blocks": [ + { + "bbox": [ + 200, + 190, + 246, + 228 + ], + "lines": [ + { + "bbox": [ + 200, + 190, + 246, + 228 + ], + "spans": [ + { + "bbox": [ + 200, + 190, + 246, + 228 + ], + "type": "image", + "image_path": "c9d5d8e87e46b948fdfed57cbe639efd94cea39461f3f4c541458e0f8bfba06b.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 209, + 228, + 236, + 236 + ], + "lines": [ + { + "bbox": [ + 209, + 228, + 236, + 236 + ], + "spans": [ + { + "bbox": [ + 209, + 228, + 236, + 236 + ], + "type": "text", + "content": "Exp.#A3" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 246, + 190, + 293, + 228 + ], + "blocks": [ + { + "bbox": [ + 246, + 190, + 293, + 228 + ], + "lines": [ + { + "bbox": [ + 246, + 190, + 293, + 228 + ], + "spans": [ + { + "bbox": [ + 246, + 190, + 293, + 228 + ], + "type": "image", + "image_path": "4fec3d3107de492ac659bc0c9d9ffeaaa20e19b00d450510d890f5e0a20b4508.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 247, + 228, + 293, + 236 + ], + "lines": [ + { + "bbox": [ + 247, + 228, + 293, + 236 + ], + "spans": [ + { + "bbox": [ + 247, + 228, + 293, + 236 + ], + "type": "text", + "content": "Exp.#A4(Ours)" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 58, + 247, + 105, + 293 + ], + "blocks": [ + { + "bbox": [ + 58, + 247, + 105, + 293 + ], + "lines": [ + { + "bbox": [ + 58, + 247, + 105, + 293 + ], + "spans": [ + { + "bbox": [ + 58, + 247, + 105, + 293 + ], + "type": "image", + "image_path": "2e4b2b3ed6e933b26ffa99e29d43124a879e0c7ef07a3a09d245dd144fd10e1e.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 59, + 293, + 102, + 300 + ], + "lines": [ + { + "bbox": [ + 59, + 293, + 102, + 300 + ], + "spans": [ + { + "bbox": [ + 59, + 293, + 102, + 300 + ], + "type": "text", + "content": "Exp.#A0 (SD)" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 303, + 296, + 369 + ], + "lines": [ + { + "bbox": [ + 55, + 303, + 296, + 369 + ], + "spans": [ + { + "bbox": [ + 55, + 303, + 296, + 369 + ], + "type": "text", + "content": "Figure 8. Comparison of results when applying STO at different timesteps. Experiments are organized as follows: no STO, STO applied from timesteps 19-24, 13-24, 7-24, and 1-24 (Ours). As seen in the images, earlier STO application improves object positioning and reduces overlap, resulting in more accurately positioned objects. (See Table 4 for quantitative evaluation.)" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 105, + 247, + 153, + 293 + ], + "blocks": [ + { + "bbox": [ + 115, + 238, + 235, + 247 + ], + "lines": [ + { + "bbox": [ + 115, + 238, + 235, + 247 + ], + "spans": [ + { + "bbox": [ + 115, + 238, + 235, + 247 + ], + "type": "text", + "content": "\"An orange above a suitcase\"" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 105, + 247, + 153, + 293 + ], + "lines": [ + { + "bbox": [ + 105, + 247, + 153, + 293 + ], + "spans": [ + { + "bbox": [ + 105, + 247, + 153, + 293 + ], + "type": "image", + "image_path": "21b28c35622515d03f64baec19d1cf117751fb078e0b62d59c28dca30788160d.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 115, + 293, + 141, + 300 + ], + "lines": [ + { + "bbox": [ + 115, + 293, + 141, + 300 + ], + "spans": [ + { + "bbox": [ + 115, + 293, + 141, + 300 + ], + "type": "text", + "content": "Exp.#A1" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 153, + 247, + 199, + 293 + ], + "blocks": [ + { + "bbox": [ + 153, + 247, + 199, + 293 + ], + "lines": [ + { + "bbox": [ + 153, + 247, + 199, + 293 + ], + "spans": [ + { + "bbox": [ + 153, + 247, + 199, + 293 + ], + "type": "image", + "image_path": "7571c6ab5d0366785ebe52fc87797ddda207e51302110d75305fe97f5246e394.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 162, + 293, + 189, + 300 + ], + "lines": [ + { + "bbox": [ + 162, + 293, + 189, + 300 + ], + "spans": [ + { + "bbox": [ + 162, + 293, + 189, + 300 + ], + "type": "text", + "content": "Exp.#A2" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 200, + 247, + 246, + 293 + ], + "blocks": [ + { + "bbox": [ + 200, + 247, + 246, + 293 + ], + "lines": [ + { + "bbox": [ + 200, + 247, + 246, + 293 + ], + "spans": [ + { + "bbox": [ + 200, + 247, + 246, + 293 + ], + "type": "image", + "image_path": "c1bf3e168887c321ef9c5e16dfe12bad45b47318d9b3efe286cb4299f805a121.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 209, + 293, + 236, + 300 + ], + "lines": [ + { + "bbox": [ + 209, + 293, + 236, + 300 + ], + "spans": [ + { + "bbox": [ + 209, + 293, + 236, + 300 + ], + "type": "text", + "content": "Exp.#A3" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 246, + 247, + 293, + 293 + ], + "blocks": [ + { + "bbox": [ + 246, + 247, + 293, + 293 + ], + "lines": [ + { + "bbox": [ + 246, + 247, + 293, + 293 + ], + "spans": [ + { + "bbox": [ + 246, + 247, + 293, + 293 + ], + "type": "image", + "image_path": "3c0a3cf3070b282ad151499bb0a32a2681ebcd07be4c75197d9fc5268a077866.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 247, + 293, + 293, + 300 + ], + "lines": [ + { + "bbox": [ + 247, + 293, + 293, + 300 + ], + "spans": [ + { + "bbox": [ + 247, + 293, + 293, + 300 + ], + "type": "text", + "content": "Exp.#A4 (Ours)" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "bbox": [ + 55, + 376, + 296, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 376, + 296, + 473 + ], + "spans": [ + { + "bbox": [ + 55, + 376, + 296, + 473 + ], + "type": "text", + "content": "spatial details, leading to errors such as a rabbit overlapping a backpack instead of being positioned to its right. In contrast, our method, which is focused on precise positioning, demonstrates a strong understanding of spatial information and generates objects accurately. Our method also can generate images that accurately reflect complex spatial information, even when prompts contain intricate positioning details (refer Fig. 6)." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 55, + 479, + 156, + 490 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 479, + 156, + 490 + ], + "spans": [ + { + "bbox": [ + 55, + 479, + 156, + 490 + ], + "type": "text", + "content": "5.2. Ablation Studies" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 55, + 493, + 296, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 493, + 296, + 649 + ], + "spans": [ + { + "bbox": [ + 55, + 493, + 296, + 649 + ], + "type": "text", + "content": "Effects of Optimizing through Timestep. In Table 4, we compare STO applied at different timestep ranges (Exp.#A1: 19-24, Exp.#A2: 13-24, Exp.#A3: 7-24, Exp.#A4: 1-24). These experiments show that setting spatial configurations in the early stages is essential not only for positional accuracy but also for achieving higher OA. This improvement increases as spatial relationships are established earlier in the process. Notably, all configurations with STO outperform the baseline (Exp.#A0, without STO), demonstrating a substantial improvement in both object positioning and accuracy. In Fig. 8, tie placement is optimal when spatial cues are applied early, while delayed guidance leads to misplacements or suboptimal synthesis." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "type": "text", + "content": "Effects of Cost Function. Table 5 presents an ablation study on methods for computing movement costs between a source and target distribution. Exp.#B0 uses Manhattan distance (only horizontal and vertical movement), while Exp.#B1 considers the mass distribution of the attention" + } + ] + } + ], + "index": 29 + }, + { + "type": "table", + "bbox": [ + 317, + 228, + 553, + 299 + ], + "blocks": [ + { + "bbox": [ + 313, + 182, + 554, + 227 + ], + "lines": [ + { + "bbox": [ + 313, + 182, + 554, + 227 + ], + "spans": [ + { + "bbox": [ + 313, + 182, + 554, + 227 + ], + "type": "text", + "content": "Table 4. Ablation study on the impact of applying STO at different timesteps. Exp.#A0 shows baseline results from SD without STO. From Exp.#A1 to Exp.#A4, STO is progressively applied across increasing timestep ranges (19-24, 13-24, 7-24, and 1-24)." + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 228, + 553, + 299 + ], + "lines": [ + { + "bbox": [ + 317, + 228, + 553, + 299 + ], + "spans": [ + { + "bbox": [ + 317, + 228, + 553, + 299 + ], + "type": "table", + "html": "
# Exp.OA (%)VISOR
uncondcond1234
A029.8618.8162.9846.6020.116.891.63
A146.6237.4880.3874.1145.7322.307.82
A255.6547.9086.0781.6360.3035.8913.81
A359.8553.6289.6083.6167.5344.1219.31
A4 (Ours)61.0157.5894.3985.9369.7149.0125.70
", + "image_path": "f0e6c7f7a885726381c498896adab9009a56d58c1b7bca2f1cca5e0defbbe8c4.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "table_body" + } + ], + "index": 31 + }, + { + "type": "table", + "bbox": [ + 317, + 350, + 552, + 422 + ], + "blocks": [ + { + "bbox": [ + 313, + 305, + 554, + 350 + ], + "lines": [ + { + "bbox": [ + 313, + 305, + 554, + 350 + ], + "spans": [ + { + "bbox": [ + 313, + 305, + 554, + 350 + ], + "type": "text", + "content": "Table 5. Ablation study on computing cost function, highlighting the impact of ST Cost components on OA and VISOR performance. (B0: Manhattan distance; B1: Euclidean distance; B2: B0 + B1; B3: ST cost without overlap prevention.)" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 350, + 552, + 422 + ], + "lines": [ + { + "bbox": [ + 317, + 350, + 552, + 422 + ], + "spans": [ + { + "bbox": [ + 317, + 350, + 552, + 422 + ], + "type": "table", + "html": "
# Exp.OA (%)VISOR
uncondcond1234
B027.3217.9165.5142.8419.257.602.00
B147.4830.3563.9365.2037.7114.034.54
B246.6029.2062.9363.9233.9114.434.60
B331.3827.7888.5361.3032.6113.443.82
B4 (Ours)61.0157.5894.3985.9369.7149.0125.70
", + "image_path": "2057f576cb755ae6570a25ff681ddad80e6982e0a43bd3bccc35704ce03a015d.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "table_body" + } + ], + "index": 33 + }, + { + "bbox": [ + 313, + 432, + 555, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 432, + 555, + 506 + ], + "spans": [ + { + "bbox": [ + 313, + 432, + 555, + 506 + ], + "type": "text", + "content": "map using Euclidean distance. Exp.#B2 combines Exp.#B0 and Exp.#B1. Exp.#B3 excludes the overlap prevention component from ST Cost to examine its impact. Our results (Exp.#B4) show that ST Cost outperforms standard OT functions, with overlap prevention proving critical for maintaining OA, despite minimal impact on VISOR." + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 313, + 512, + 388, + 524 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 512, + 388, + 524 + ], + "spans": [ + { + "bbox": [ + 313, + 512, + 388, + 524 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 313, + 529, + 555, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 529, + 555, + 673 + ], + "spans": [ + { + "bbox": [ + 313, + 529, + 555, + 673 + ], + "type": "text", + "content": "In this work, we introduced STORM, a framework that dynamically addresses spatial misalignment issues in training-free T2I synthesis. By leveraging the STO framework, which combines OT and an ST Cost function, STORM not only resolves misplaced objects but also tackles missing objects and mismatched attributes. Extensive experiments show that STORM significantly improves spatial alignment and object accuracy, surpassing existing state-of-the-art methods. This advancement establishes STORM as a new benchmark for spatial control in T2I generation, pushing the boundaries of adaptability and efficiency in training-free approaches." + } + ] + } + ], + "index": 36 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 313, + 679, + 554, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 679, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 679, + 554, + 713 + ], + "type": "text", + "content": "Acknowledgement. This work was supported in part by IITP RS-2024-00457882 (AI Research Hub Project), IITP 2020-0-01361, NRF RS-2024-00345806, NRF RS-2023-00219019, and RS-2024-00403860 (Korea Basic Science Institute, National Research Facilities and Equipment Center)." + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18408" + } + ] + } + ], + "index": 38 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "type": "text", + "content": "[1] Aishwarya Agarwal, Srikrishna Karanam, KJ Joseph, Apoorv Saxena, Koustava Goswami, and Balaji Vasan Srinivasan. A-star: Test-time attention segregation and retention for text-to-image synthesis. In ICCV, 2023. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 137, + 296, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 137, + 296, + 192 + ], + "spans": [ + { + "bbox": [ + 61, + 137, + 296, + 192 + ], + "type": "text", + "content": "[2] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Ji-aming Song, Qinsheng Zhang, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, et al. ediff-i: Text-to-image diffusion models with an ensemble of expert denoisers. arXiv preprint arXiv:2211.01324, 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 194, + 295, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 194, + 295, + 225 + ], + "spans": [ + { + "bbox": [ + 62, + 194, + 295, + 225 + ], + "type": "text", + "content": "[3] Agneet Chatterjee, Yiran Luo, Tejas Gokhale, Yezhou Yang, and Chitta Baral. Revision: Rendering tools enable spatial fidelity in vision-language models. In ECCV, 2025. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 227, + 295, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 227, + 295, + 281 + ], + "spans": [ + { + "bbox": [ + 62, + 227, + 295, + 281 + ], + "type": "text", + "content": "[4] Agneet Chatterjee, Gabriela Ben Melech Stan, Estelle Aflalo, Sayak Paul, Dhruba Ghosh, Tejas Gokhale, Ludwig Schmidt, Hannaneh Hajishirzi, Vasudev Lal, Chitta Baral, et al. Getting it right: Improving spatial consistency in text-to-image models. In ECCV, 2025. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 284, + 295, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 284, + 295, + 327 + ], + "spans": [ + { + "bbox": [ + 62, + 284, + 295, + 327 + ], + "type": "text", + "content": "[5] Hila Chefer, Yuval Alaluf, Yael Vinker, Lior Wolf, and Daniel Cohen-Or. Attend-and-excite: Attention-based semantic guidance for text-to-image diffusion models. SIGGRAPH, 2023. 1, 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 328, + 295, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 328, + 295, + 360 + ], + "spans": [ + { + "bbox": [ + 62, + 328, + 295, + 360 + ], + "type": "text", + "content": "[6] Minghao Chen, Iro Laina, and Andrea Vedaldi. Training-free layout control with cross-attention guidance. In WACV, 2024. 1, 2, 3, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 363, + 295, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 363, + 295, + 407 + ], + "spans": [ + { + "bbox": [ + 62, + 363, + 295, + 407 + ], + "type": "text", + "content": "[7] Guillaume Couairon, Marlene Careil, Matthieu Cord, Stéphane Lathuiliere, and Jakob Verbeek. Zero-shot spatial layout conditioning for text-to-image diffusion models. In ICCV, 2023. 2, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 409, + 295, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 409, + 295, + 430 + ], + "spans": [ + { + "bbox": [ + 62, + 409, + 295, + 430 + ], + "type": "text", + "content": "[8] Marco Cuturi. Sinkhorn distances: Lightspeed computation of optimal transport. NeurIPS, 2013. 3, 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 432, + 295, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 432, + 295, + 486 + ], + "spans": [ + { + "bbox": [ + 62, + 432, + 295, + 486 + ], + "type": "text", + "content": "[9] Weixi Feng, Xuehai He, Tsu-Jui Fu, Varun Jampani, Arjun Akula, Pradyumna Narayana, Sugato Basu, Xin Eric Wang, and William Yang Wang. Training-free structured diffusion guidance for compositional text-to-image synthesis. arXiv preprint arXiv:2212.05032, 2022. 1, 2, 3, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 488, + 295, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 488, + 295, + 532 + ], + "spans": [ + { + "bbox": [ + 57, + 488, + 295, + 532 + ], + "type": "text", + "content": "[10] Tejas Gokhale, Hamid Palangi, Besmira Nushi, Vibhav Vineet, Eric Horvitz, Ece Kamar, Chitta Baral, and Yezhou Yang. Benchmarking spatial relationships in text-to-image generation. arXiv preprint arXiv:2212.10015, 2022. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 533, + 295, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 533, + 295, + 576 + ], + "spans": [ + { + "bbox": [ + 57, + 533, + 295, + 576 + ], + "type": "text", + "content": "[11] Biao Gong, Siteng Huang, Yutong Feng, Shiwei Zhang, Yuyuan Li, and Yu Liu. Check locate rectify: A training-free layout calibration system for text-to-image generation. In CVPR, 2024. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 578, + 295, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 578, + 295, + 622 + ], + "spans": [ + { + "bbox": [ + 57, + 578, + 295, + 622 + ], + "type": "text", + "content": "[12] Xiefan Guo, Jinlin Liu, Miaomiao Cui, Jiankai Li, Hongyu Yang, and Di Huang. Initno: Boosting text-to-image diffusion models via initial noise optimization. In CVPR, 2024, 2, 3, 6, 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 624, + 295, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 624, + 295, + 666 + ], + "spans": [ + { + "bbox": [ + 57, + 624, + 295, + 666 + ], + "type": "text", + "content": "[13] Woojung Han, Chanyoung Kim, Dayun Ju, Yumin Shim, and Seong Jae Hwang. Advancing text-driven chest x-ray generation with policy-based reinforcement learning. MICCAI, 2024. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 670, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 295, + 712 + ], + "type": "text", + "content": "[14] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 3" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 95 + ], + "type": "text", + "content": "[15] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. NeurIPS, 2020. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 97, + 553, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 97, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 317, + 97, + 553, + 140 + ], + "type": "text", + "content": "[16] Yushi Hu, Benlin Liu, Jungo Kasai, Yizhong Wang, Mari Ostendorf, Ranjay Krishna, and Noah A Smith. Tifa: Accurate and interpretable text-to-image faithfulness evaluation with question answering. In ICCV, 2023. 6, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 143, + 553, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 143, + 553, + 185 + ], + "spans": [ + { + "bbox": [ + 317, + 143, + 553, + 185 + ], + "type": "text", + "content": "[17] Kaiyi Huang, Kaiyue Sun, Enze Xie, Zhenguo Li, and Xihui Liu. T2i-compbench: A comprehensive benchmark for open-world compositional text-to-image generation. NeurIPS, 2023. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 188, + 553, + 220 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 188, + 553, + 220 + ], + "spans": [ + { + "bbox": [ + 317, + 188, + 553, + 220 + ], + "type": "text", + "content": "[18] Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A Efros. Image-to-image translation with conditional adversarial networks. In CVPR, 2017. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 223, + 553, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 223, + 553, + 255 + ], + "spans": [ + { + "bbox": [ + 316, + 223, + 553, + 255 + ], + "type": "text", + "content": "[19] Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In CVPR, 2019. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 258, + 553, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 258, + 553, + 300 + ], + "spans": [ + { + "bbox": [ + 316, + 258, + 553, + 300 + ], + "type": "text", + "content": "[20] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, 2022. 6, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 303, + 553, + 324 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 303, + 553, + 324 + ], + "spans": [ + { + "bbox": [ + 316, + 303, + 553, + 324 + ], + "type": "text", + "content": "[21] Jiachen Li, Jitesh Jain, and Humphrey Shi. Matting anything. In CVPR, 2024. 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 327, + 553, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 327, + 553, + 358 + ], + "spans": [ + { + "bbox": [ + 316, + 327, + 553, + 358 + ], + "type": "text", + "content": "[22] Nan Liu, Shuang Li, Yilun Du, Antonio Torralba, and Joshua B Tenenbaum. Compositional visual generation with composable diffusion models. In ECCV, 2022. 1, 2, 3, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 361, + 553, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 361, + 553, + 404 + ], + "spans": [ + { + "bbox": [ + 316, + 361, + 553, + 404 + ], + "type": "text", + "content": "[23] Tuna Han Salih Meral, Enis Simsar, Federico Tombari, and Pinar Yanardag. Conform: Contrast is all you need for high-fidelity text-to-image diffusion models. In CVPR, 2024. 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 407, + 553, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 407, + 553, + 460 + ], + "spans": [ + { + "bbox": [ + 317, + 407, + 553, + 460 + ], + "type": "text", + "content": "[24] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 1, 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 463, + 553, + 496 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 463, + 553, + 496 + ], + "spans": [ + { + "bbox": [ + 317, + 463, + 553, + 496 + ], + "type": "text", + "content": "[25] Weili Nie, Sifei Liu, Morteza Mardani, Chao Liu, Benjamin Eckart, and Arash Vahdat. Compositional text-to-image generation with dense blob representations. In ICML, 2024. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 498, + 553, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 498, + 553, + 542 + ], + "spans": [ + { + "bbox": [ + 317, + 498, + 553, + 542 + ], + "type": "text", + "content": "[26] Ji-Hoon Park, Yeong-Joon Ju, and Seong-Whan Lee. Explaining generative diffusion models via visual analysis for interpretable decision-making process. Expert Systems with Applications, 2024. 5" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 544, + 553, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 544, + 553, + 576 + ], + "spans": [ + { + "bbox": [ + 316, + 544, + 553, + 576 + ], + "type": "text", + "content": "[27] Taesung Park, Ming-Yu Liu, Ting-Chun Wang, and Jun-Yan Zhu. Semantic image synthesis with spatially-adaptive normalization. In CVPR, 2019. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 317, + 578, + 553, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 578, + 553, + 632 + ], + "spans": [ + { + "bbox": [ + 317, + 578, + 553, + 632 + ], + "type": "text", + "content": "[28] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 3, 6" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 317, + 635, + 553, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 635, + 553, + 667 + ], + "spans": [ + { + "bbox": [ + 317, + 635, + 553, + 667 + ], + "type": "text", + "content": "[29] Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In ICML, 2021. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "text", + "content": "[30] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 2022. 2" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "18409" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 600 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "text", + "content": "[31] Royi Rassin, Eran Hirsch, Daniel Glickman, Shauli Ravfogel, Yoav Goldberg, and Gal Chechik. Linguistic binding in diffusion models: Enhancing attribute correspondence through attention map alignment. NeurIPS, 2024. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 161 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 161 + ], + "type": "text", + "content": "[32] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 1, 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 294, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 294, + 195 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 294, + 195 + ], + "type": "text", + "content": "[33] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In MICCAI, 2015. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 197, + 294, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 197, + 294, + 251 + ], + "spans": [ + { + "bbox": [ + 56, + 197, + 294, + 251 + ], + "type": "text", + "content": "[34] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. NeurIPS, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 253, + 294, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 253, + 294, + 285 + ], + "spans": [ + { + "bbox": [ + 56, + 253, + 294, + 285 + ], + "type": "text", + "content": "[35] Wenqiang Sun, Teng Li, Zehong Lin, and Jun Zhang. Spatial-aware latent initialization for controllable image generation. arXiv preprint arXiv:2401.16157, 2024. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 286, + 294, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 286, + 294, + 341 + ], + "spans": [ + { + "bbox": [ + 56, + 286, + 294, + 341 + ], + "type": "text", + "content": "[36] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 342, + 294, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 342, + 294, + 363 + ], + "spans": [ + { + "bbox": [ + 56, + 342, + 294, + 363 + ], + "type": "text", + "content": "[37] Cédric Villani et al. Optimal transport: old and new. Springer, 338, 2009. 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 365, + 294, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 365, + 294, + 408 + ], + "spans": [ + { + "bbox": [ + 56, + 365, + 294, + 408 + ], + "type": "text", + "content": "[38] Qiucheng Wu, Yujuan Liu, Handong Zhao, Trung Bui, Zhe Lin, Yang Zhang, and Shiyu Chang. Harnessing the spatial-temporal attention of diffusion models for high-fidelity text-to-image synthesis. In ICCV, 2023. 2, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 410, + 294, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 410, + 294, + 464 + ], + "spans": [ + { + "bbox": [ + 56, + 410, + 294, + 464 + ], + "type": "text", + "content": "[39] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, et al. Scaling autoregressive models for content-rich text-to-image generation. arXiv preprint arXiv:2206.10789, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 466, + 294, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 466, + 294, + 509 + ], + "spans": [ + { + "bbox": [ + 56, + 466, + 294, + 509 + ], + "type": "text", + "content": "[40] Mert Yuksekgonul, Federico Bianchi, Pratyusha Kalluri, Dan Jurafsky, and James Zou. When and why vision-language models behave like bags-of-words, and what to do about it? arXiv preprint arXiv:2210.01936, 2022. 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 510, + 294, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 510, + 294, + 542 + ], + "spans": [ + { + "bbox": [ + 56, + 510, + 294, + 542 + ], + "type": "text", + "content": "[41] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, 2023. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 544, + 294, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 544, + 294, + 577 + ], + "spans": [ + { + "bbox": [ + 56, + 544, + 294, + 577 + ], + "type": "text", + "content": "[42] Tianjun Zhang, Yi Zhang, Vibhav Vineet, Neel Joshi, and Xin Wang. Controllable text-to-image generation with gpt-4. arXiv preprint arXiv:2305.18583, 2023. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 578, + 294, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 578, + 294, + 600 + ], + "spans": [ + { + "bbox": [ + 56, + 578, + 294, + 600 + ], + "type": "text", + "content": "[43] Xingyi Zhou, Vladlen Koltun, and Philipp Krahenbuhl. Simple multi-dataset detection. In CVPR, 2022. 6" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "18410" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Spatial-Temporal Graph Diffusion Policy with Kinematic Modeling for Bimanual Robotic Manipulation/8300c221-08df-4f6a-8aec-775be3434ff5_content_list.json b/2025/Spatial-Temporal Graph Diffusion Policy with Kinematic Modeling for Bimanual Robotic Manipulation/8300c221-08df-4f6a-8aec-775be3434ff5_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..608a1e46645561e48a0e60c91f78fb6389863383 --- /dev/null +++ b/2025/Spatial-Temporal Graph Diffusion Policy with Kinematic Modeling for Bimanual Robotic Manipulation/8300c221-08df-4f6a-8aec-775be3434ff5_content_list.json @@ -0,0 +1,1791 @@ +[ + { + "type": "text", + "text": "Spatial-Temporal Graph Diffusion Policy with Kinematic Modeling for Bimanual Robotic Manipulation", + "text_level": 1, + "bbox": [ + 142, + 128, + 854, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Qi Lv $^{1,2,3}$ Hao Li $^{1,3}$ Xiang Deng $^{1\\dagger}$ Rui Shao $^{1}$ Yinchuan Li $^{3}$ Jianye Hao $^{3}$ Longxiang Gao $^{4}$ Michael Yu Wang $^{2}$ Liqiang Nie $^{1\\dagger}$", + "bbox": [ + 166, + 200, + 828, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ School of Computer Science and Technology, Harbin Institute of Technology (Shenzhen) \n $^{2}$ School of Engineering, Greate Bay University \n $^{3}$ Huawei Noah's Ark Lab \n $^{4}$ Qilu University of Technology (Shandong Academy of Sciences)", + "bbox": [ + 143, + 241, + 852, + 295 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "lvqi@stu.hit.edu.cn, dengxiang@hit.edu.cn", + "bbox": [ + 256, + 300, + 733, + 316 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 366, + 326, + 382 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite the significant success of imitation learning in robotic manipulation, its application to bimanual tasks remains highly challenging. Existing approaches mainly learn a policy to predict a distant next-best end-effector pose (NBP) and then compute the corresponding joint rotation angles for motion using inverse kinematics. However, they suffer from two important issues: (1) rarely considering the physical robotic structure, which may cause self-collisions or interferences, and (2) overlooking the kinematics constraint, which may result in the predicted poses not conforming to the actual limitations of the robot joints. In this paper, we propose Kinematics enhanced Spatial-TemporAl gRaph Diffuser (KStar Diffuser). Specifically, (1) to incorporate the physical robot structure information into action prediction, KStar Diffuser maintains a dynamic spatial-temporal graph according to the physical bimanual joint motions at continuous timesteps. This dynamic graph serves as the robot-structure condition for denoising the actions; (2) to make the NBP learning objective consistent with kinematics, we introduce the differentiable kinematics to provide the reference for optimizing KStar Diffuser. This module regularizes the policy to predict more reliable and kinematics-aware next end-effector poses. Experimental results show that our method effectively leverages the physical structural information and generates kinematics-aware actions in both simulation and real-world.", + "bbox": [ + 88, + 398, + 485, + 791 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 806, + 220, + 821 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Bimanual manipulation [34, 36, 42, 45] represents a fundamental capability for robotic systems to perform com", + "bbox": [ + 89, + 833, + 483, + 864 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/00add3a9ccda9ddd24e457da885fb141b35afbbd665f185f3aa0eac03b07d4e5.jpg", + "image_caption": [ + "Figure 1. The self-collision problem in bimanual manipulation tasks due to overlooking the robotic structure." + ], + "image_footnote": [], + "bbox": [ + 545, + 351, + 874, + 512 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "plex tasks requiring two-arm coordination. While imitation learning has demonstrated remarkable success in single-arm manipulation [6, 7, 33], its extension to bimanual scenarios faces unique challenges as robots need to coordinate dual-arm movements while conforming to physical constraints. These challenges significantly impact the reliability and feasibility of predicted actions in real-world applications.", + "bbox": [ + 509, + 566, + 906, + 672 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "A straightforward approach [9, 10, 13, 55, 56, 58] would be to learn the entire motion trajectory by predicting joint positions at consecutive timesteps. However, this poses significant challenges as trajectories are typically long and difficult to learn. Therefore, mainstream approaches [15, 17, 31, 44] typically adopt a two-stage pipeline: first predicting the next-best end-effector pose (NBP), then computing joint rotations through inverse kinematics [21, 26]. However, the dominant NBP-based approaches, though simplify the learning process, often lead to unreliable motion generation. Concretely, the predicted poses may violate physical structure constraints [27, 43], causing potential arm collisions, or exceeding joint limits due to inadequate consideration of kinematic feasibility, as Figure 1 illustrated.", + "bbox": [ + 509, + 674, + 908, + 883 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In real-world bimanual manipulation tasks, existing ap", + "bbox": [ + 531, + 885, + 903, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*This work was done when Qi Lv and Hao Li were intern at Huawei Noah's Ark Lab. † Corresponding author", + "bbox": [ + 89, + 875, + 482, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "17394", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "proaches [9, 15, 17, 44, 57] exhibit significant limitations in motion execution despite their promising performance in trajectory prediction. Empirical observations indicate that while individual end-effector poses are feasible in isolation, their concurrent execution frequently results in interarm collisions [43] and kinematically infeasible configurations [31, 35]. The fundamental challenge lies in the optimization of end-effector poses exclusively in Cartesian space, which introduces a substantial discrepancy between motion prediction and physical constraints. Existing approaches [9, 10, 17] incorporate proprioception information in a simplistic manner, (e.g., gripper rotation angles [17, 44] and end-effector poses [48]), while disregarding the robot's intrinsic structural properties e.g., the kinematic chain and joint configurations during the prediction phase. This oversimplified representation fails to capture the crucial spatial correlations between robotic arms, joint links, and their potential interactions. Consequently, the predicted dual-arm trajectories may satisfy task objectives in Cartesian space while violating crucial spatial constraints during execution.", + "bbox": [ + 89, + 90, + 480, + 393 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Moreover, the conventional paradigm [17, 24, 44, 46] of treating inverse kinematics as a post-processing procedure introduces additional complexities. The disjunction between kinematic constraints and the pose prediction learning objective often yields solutions that exhibit apparent smoothness in Cartesian space while manifesting discontinuous or mechanically infeasible trajectories in the joint space. This limitation becomes particularly pronounced in configurations approaching kinematic singularities or joint limits, where the mapping between Cartesian and joint spaces becomes ill-conditioned. These inherent limitations necessitate the integration of both structural and kinematic constraints into the policy learning framework.", + "bbox": [ + 89, + 397, + 482, + 594 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address the above issues, we propose Kinematics enhanced Spatial-TemporAl gRaph Diffuser (KStar Diffuser), a novel framework that explicitly incorporates both robot structures and kinematics into the bimanual motion generation process. Our key insight is that the physical structure and kinematic properties of the robot should guide the learning process of pose prediction, rather than be treated as independent post-processing constraints. Specifically, (1) to incorporate structural awareness, we construct a dynamic spatial-temporal graph from the robot's URDF specifications, where nodes represent joint properties and edges capture both spatial relations and temporal dependencies. This graph structure is encoded via Graph Convolutional Network (GCN) [25] to provide explicit physical constraints for the diffusion process; (2) for kinematic feasibility, we regularize the NBP learning objective by incorporating joint-space prediction, where the predicted joint positions are mapped to reference end-effector poses through differentiable forward kinematics. These kinematically-feasible poses then serve as conditions", + "bbox": [ + 89, + 598, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "to guide the diffusion process, ensuring that the generated motions satisfy both structural and kinematic constraints. Our main contributions are as follows:", + "bbox": [ + 511, + 90, + 903, + 135 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Different from existing approaches that optimize end-effector poses solely in Cartesian space, we propose a novel spatial-temporal robot graph that explicitly models the robot physical configuration to guide the generative action denoising procedure.", + "- We introduce a kinematics regularizer that augments the NBP learning objective by introducing joint-space supervision. This regularizer leverages forward kinematics to provide kinematically-feasible reference poses, effectively guiding the diffusion process to conform to kinematic constraints.", + "- Extensive experiments show that our proposed KStar Diffuser is superior in both simulation and real-world scenarios, surpassing baselines more than $10\\%$ in success ratio." + ], + "bbox": [ + 511, + 136, + 903, + 349 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 359, + 653, + 373 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Diffusion Models in Bimanual Robotic Controls. Diffusion models such as Denoising Diffusion Probabilistic Models (DDPM) [19] have achieved great success in the fields of image generation [38, 39, 41] and video generation [37, 50, 59]. Thus, recent work [8, 40, 47, 52, 60] has been devoted to applying its powerful generation capabilities to action generation for robotic manipulation tasks in a imitation learning mode. With the 2D images [9, 31, 40, 47, 49] or 3D point cloud observations [8, 20, 23, 55, 56], the policy is trained to output the action sequence of joint positions or end-effector poses by iterative denoising process.", + "bbox": [ + 511, + 380, + 903, + 547 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Bimanual manipulation tasks, which resemble human-like actions more closely, have garnered increasing attention [10, 14, 17, 30, 58]. However, existing research primarily extends single-arm manipulation methods to dual-arm tasks [10, 58], without considering about the distinct challenges specific to bimanual settings. The relative independence of the two arms makes it essential to consider self-collision avoidance [1, 3] when predicting actions. In practice, the robot's structure significantly influences its motion since the joint types determine movement directions and joint angles constrain the range of motion. Although some studies [15, 44, 48] integrate proprioception in robotic control, they often use simply low-dimensional vectors to represent body information, lacking depth in structural exploration and overlooking critical spatial relationships between dual arms. Therefore, we study the physical structure information of the dual-arm robot system, and improve both the accuracy and adaptability of robotic movements.", + "bbox": [ + 511, + 547, + 905, + 820 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Robotic Kinematics Modeling. Robotic kinematics modeling [2, 16, 31, 53, 54] is a classic problem in the robotic control [4, 18, 32, 51] where inverse kinematics (IK) presents a fundamental challenge, i.e., deriving the joint configuration given the end-effector pose. The inher", + "bbox": [ + 511, + 825, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "17395", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ent non-uniqueness of IK solutions as a significant obstacle for learning algorithms [12, 22], often produce inaccurate models by averaging over nonconvex feasible sets. Recent advancements integrate neural solutions [11, 28, 29] to enhance the adaptability and precision of kinematic models. Bócsi et al. [5] leveraged support vector machines to parameterize quadratic programs, aligning solutions with IK in specific workspace regions. Current data-driven imitation learning approaches [10, 13, 15, 17, 58] primarily rely on probabilistic modeling of action distributions to predict subsequent actions, but lack the reliability guarantees from the aspect of kinematics. In this paper, we construct a spatial-temporal graph to learn the robot physical representation, and provide a kinematics-aware latent end-effect embedding for diffusion policy as guidance. In this way, o'ur approach enhances the robot's ability to perform tasks with superious precision and adaptability.", + "bbox": [ + 89, + 90, + 483, + 349 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 366, + 181, + 382 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminary", + "text_level": 1, + "bbox": [ + 89, + 392, + 220, + 409 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Diffusion Policy. Chi et al. [9] propose the diffusion policy which represents a robot's visuomotor policy as a conditional denoising diffusion process. It learns a model distribution $\\pi_{\\phi}(\\pmb{a}_0|\\pmb{o})$ conditioned on observation $\\pmb{o}$ to approximate joint distribution $q(\\pmb{a}_0)$ . The whole procedure consists of a forward process and a reverse process.", + "bbox": [ + 89, + 416, + 483, + 507 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "(1) Forward Process: For the Markov chain with Gaussian transitions parameterized, the policy $\\pi_{\\phi}(\\pmb{a}_0|\\pmb{o})$ learns a fixed inference procedure $q(\\pmb{a}_k|\\pmb{a}_0)$ :", + "bbox": [ + 89, + 508, + 483, + 555 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} q \\left(\\boldsymbol {a} _ {k} \\mid \\boldsymbol {a} _ {0}\\right) = \\int q \\left(\\boldsymbol {a} _ {1: k} \\mid \\boldsymbol {a} _ {0}\\right) \\mathrm {d} \\boldsymbol {a} _ {1: (k - 1)} \\tag {1} \\\\ = \\mathcal {N} \\left(\\boldsymbol {a} _ {k}; \\sqrt {\\bar {\\alpha} _ {k}} \\boldsymbol {a} _ {0}, (1 - \\bar {\\alpha} _ {k}) \\boldsymbol {I}\\right) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 163, + 569, + 480, + 619 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Thus, $a_{k}$ can be expressed as a linear combination of $a_0$ and a noise variable $\\epsilon$ :", + "bbox": [ + 89, + 633, + 483, + 662 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {a} _ {k} = \\sqrt {\\bar {\\alpha} _ {k}} \\boldsymbol {a} _ {0} + \\sqrt {1 - \\bar {\\alpha} _ {k}} \\varepsilon \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 194, + 679, + 482, + 696 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The training loss is obtained:", + "bbox": [ + 89, + 710, + 284, + 726 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathbb {E} _ {\\boldsymbol {a} _ {0} \\sim q (\\boldsymbol {a} _ {0}), \\epsilon_ {k} \\sim \\mathcal {N} (\\boldsymbol {0}, \\boldsymbol {I}), k} \\left[ \\left| \\left| \\epsilon_ {k}, \\pi_ {\\phi} (\\boldsymbol {a} _ {0} + \\epsilon_ {k}, \\boldsymbol {o}, k) \\right| \\right| ^ {2} \\right] \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 104, + 739, + 480, + 772 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\epsilon_{k}$ is the random noise sampling at $k$ iteration.", + "bbox": [ + 89, + 773, + 441, + 787 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "(2) Reverse Process: Starting from a sample $\\pmb{a}_K \\sim \\mathcal{N}(\\pmb{0}, \\pmb{I})$ , the reverse steps are:", + "bbox": [ + 89, + 789, + 483, + 820 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {a} _ {k - 1} = \\sqrt {\\bar {\\alpha} _ {k - 1}} \\pi_ {\\phi} (a _ {k}, k, o) + \\sqrt {1 - \\bar {\\alpha} _ {k - 1}} \\epsilon . \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 133, + 834, + 482, + 852 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "By the iterative denoising process, the policy generates $a_0$ as its next action. The continuous predicted action sequence forms the complete action trajectory.", + "bbox": [ + 89, + 866, + 483, + 912 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Robotic Kinematics. Robotics kinematics describes the relationship between the robot joints and its end-effector. It can be divided into forward kinematics and inverse kinematics for controlling robot movements.", + "bbox": [ + 511, + 90, + 905, + 150 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "(1) Forward Kinematics (FK): Given a specific configuration of joint angles, i.e., $\\theta = [\\theta_{1},\\theta_{2},\\dots,\\theta_{n}]\\in \\Theta$ , it aims to construct a function $f_{\\mathrm{FK}}:\\mathbb{R}^n\\to SE(3)$ , computing the end-effector position and orientation that are represented as a pose matrix $\\mathbf{T}\\in SE(3)$ .", + "bbox": [ + 511, + 151, + 906, + 227 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nF K (\\theta) = \\mathbf {T} \\in \\mathcal {T}, \\text {w h e r e} \\mathbf {T} = \\prod_ {i = 1} ^ {n} \\mathbf {T} _ {i} (\\theta_ {i}), (5)\n$$\n", + "text_format": "latex", + "bbox": [ + 553, + 233, + 906, + 258 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbf{T}_i(\\theta_i)$ is the homogeneous transformation matrix for joint $i$ , and $\\mathcal{T}$ means the task space. $\\Theta \\in \\mathbb{R}^n$ denotes the configuration space, while the configuration refers to the specific arrangement of the robotic joints in a given pose.", + "bbox": [ + 511, + 267, + 905, + 327 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "(2) Inverse Kinematics (IK): In contrast, IK aims to find the joint configurations which achieve a desired end-effector pose. It is defined as a mapping $f_{\\mathrm{IK}}: SE(3) \\to \\mathbb{R}^n$ from the task space $\\mathcal{T}$ to the set of possible configurations $\\Theta$ that satisfy the given pose matrix $\\mathbf{T}$ :", + "bbox": [ + 511, + 327, + 905, + 404 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nI K (\\mathcal {T}) = \\{\\theta \\in \\Theta | F K (\\theta) = \\mathbf {T} \\}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 596, + 414, + 903, + 430 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "It is noted that this mapping is often non-unique, especially for redundant manipulators with more than 6 DoF.", + "bbox": [ + 511, + 440, + 905, + 470 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Task Definition", + "text_level": 1, + "bbox": [ + 511, + 479, + 666, + 494 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given the dataset $D$ with language instructions $\\pmb{l}$ and the RGB-D observations $\\pmb{o}$ , we aim to learn a policy $\\pi_{\\phi}(\\pmb{a}| \\pmb{o}, \\pmb{l})$ which predicts actions $\\pmb{a}$ . Here, $\\pmb{a}$ is composed of a trajectory $\\pmb{a}^{\\mathrm{joint}} = \\{\\pmb{a}_0, \\pmb{a}_1, \\dots, \\pmb{a}_T\\}$ and gripper opening or closing action $\\pmb{a}^{\\mathrm{gripper}}$ , where $T$ denotes the trajectory length and $\\pmb{a}_i \\in \\mathbb{R}^n$ . $n$ represents the number of robot joints. In the bimanual manipulation, $n$ is usually to be 12 or 14 as each robot arm has 6 or 7-DoF. Referring to prior works [17, 44], it is inefficient to train the policy on all trajectory points. Thus, a keyframe discovery method is used to extract a set of $K_{\\xi}$ keyframe indices $\\{k_i\\}_{i=0}^{K_{\\xi}}$ and the prediction actions are the key set of end-effector poses $\\pmb{a}^{\\mathrm{EE}}$ .", + "bbox": [ + 511, + 501, + 906, + 685 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.3. KStar Diffuser", + "text_level": 1, + "bbox": [ + 511, + 693, + 665, + 707 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.3.1. Overview", + "text_level": 1, + "bbox": [ + 511, + 715, + 625, + 729 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Mainstream methods [17, 23, 58] train a policy to predict actions, but take little consideration about the mechanical robot structure which determines its motion. We thus propose a spatial-temporal graph to model both the static physical structure and dynamic history movement information. In addition, to reduce kinematically infeasible predictions for the end-effector pose, we introduce a differentiable kinematics module that provides kinematics-aware references to the policy network. The overview of our proposed Kinematics enhanced Spatial-TemporAl gRaph Diffuser (KStar Diffuser) is shown in Figure 2.", + "bbox": [ + 511, + 734, + 906, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "17396", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/2bb8018e2225916c9c5a23ac6d9af49d89107a0d63faf94afc1e612f07d8fbfe.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 96, + 89, + 906, + 212 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/b1cddb061fecf0b7a76fe375fbb3848943c39b2a62c48ad7f571ecb82841926f.jpg", + "image_caption": [ + "Figure 2. Overview of KStar Diffuser. The top part presents the spatial-temporal graph which is constructed according to the robot architecture. The bottom part shows our backbone and the proposed kinematics regularizer. For the backbone, it extracts the multimodal information which consists of multiview RGB-D observations and language instruction, and then generates bimanual 6D end-effector poses. The kinematics regularizer enhances pose learning by incorporating joint-level predictions, which are mapped to reference end-effector poses through differentiable forward kinematics (FK)." + ], + "image_footnote": [], + "bbox": [ + 94, + 213, + 906, + 377 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3.2. Backbone", + "text_level": 1, + "bbox": [ + 89, + 477, + 205, + 491 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given the language instruction $l$ and multiview RGB-D observation images $o$ , we first adopt the Transformer-based encoders to extract their features $\\mathbf{E}_{\\mathrm{I}}$ and $\\mathbf{E}_{\\mathrm{T}}$ , respectively. Then the features are fused with the $l_{\\mathrm{FiLM}}$ -layers FiLM block to obtain the hidden states $\\mathbf{H}_{\\mathrm{B}}$ . Each layer is combined with an up-sampling 2D convolution layer:", + "bbox": [ + 88, + 496, + 482, + 587 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {H} _ {\\mathrm {B}} ^ {i} = \\operatorname {C o n v 2 D} \\left(\\mathbf {H} _ {\\mathrm {B}} ^ {i}\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 210, + 594, + 480, + 612 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {H} _ {\\mathrm {B}} ^ {i + 1} = \\operatorname {F i L M} ^ {i} \\left(\\mathbf {H} _ {\\mathrm {B}} ^ {i}, \\mathbf {E} _ {\\mathrm {T}}\\right), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 196, + 616, + 480, + 635 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $i\\in [0,l_{\\mathrm{FiLM}})$ and the $\\mathbf{H}_{\\mathrm{B}}^{0}$ is initialized with $\\mathbf{E}_{\\mathrm{I}}$", + "bbox": [ + 89, + 638, + 452, + 652 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our backbone $\\pi_{\\phi}$ uses the last hidden states $\\mathbf{H}_{\\mathrm{B}}$ as the condition to guide a diffusion head to denoise and generate the bimanual end-effector pose. It is noted that we attach the $n$ historical observation images to provide more information for capturing the motion tendency. Following Chi et al. [9], we let the policy predict next $m$ actions during training to alleviate multimodal problems. Here, we set both $n$ and $m$ to 2. The action prediction is as follows:", + "bbox": [ + 89, + 654, + 483, + 773 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {a} _ {k - 1} ^ {\\mathrm {E E}} \\sim \\pi_ {\\phi} \\left(\\boldsymbol {a} _ {k - 1} ^ {\\mathrm {E E}} \\mid \\boldsymbol {a} _ {k} ^ {\\mathrm {E E}}, \\mathbf {H} _ {\\mathrm {B}}\\right), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 187, + 779, + 480, + 799 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\pi_ {\\phi} \\left(\\boldsymbol {a} _ {k - 1} ^ {\\mathrm {E E}} \\mid \\boldsymbol {a} _ {k} ^ {\\mathrm {E E}}, \\mathbf {H} _ {\\mathrm {B}}\\right) \\tag {10} \\\\ = \\mathcal {N} (\\boldsymbol {a} _ {k - 1} ^ {\\mathrm {E E}}; \\sqrt {\\bar {\\alpha} _ {k - 1}} \\pi_ {\\phi} (\\boldsymbol {a} _ {k} ^ {\\mathrm {E E}}, \\mathbf {H} _ {\\mathrm {B}}); (1 - \\bar {\\alpha} _ {k - 1}) I). \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 102, + 804, + 482, + 845 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The learning objective is:", + "bbox": [ + 89, + 848, + 261, + 863 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {E E}} = \\mathbb {E} _ {q} \\left[ \\log \\frac {\\pi_ {\\phi} \\left(\\boldsymbol {a} _ {0 : K} ^ {\\mathrm {E E}} \\mid \\mathbf {H} _ {\\mathrm {B}}\\right)}{q \\left(\\boldsymbol {a} _ {1 : K} ^ {\\mathrm {E E}} \\mid \\boldsymbol {a} _ {0} ^ {\\mathrm {E E}}, \\mathbf {H} _ {\\mathrm {B}}\\right)} \\right]. \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 161, + 869, + 482, + 905 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3.3. Spatial-Temporal Robot Graph", + "text_level": 1, + "bbox": [ + 511, + 477, + 777, + 493 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The physical architecture impacts the motion of the whole robot, determining whether it can complete the task. Meanwhile, the historical spatial information is also important to future movements. Thus, we propose a spatial-temporal graph method to model the robot architecture at each step and the robot motion at continuous timesteps, representing the static spatial information and dynamic motion features.", + "bbox": [ + 511, + 497, + 906, + 603 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Spatial Structure Graph Construction. To represent the robot structure, we first parse the Unified Robot Description Format (URDF) file which is usually used to describe the static physical structure of robot such as joint types, joint limits, and link lengths. Then, we define a dual-arm system as an undirected graph $G_{\\mathrm{s}} = \\langle V,E\\rangle$ based on the joint and link configuration. Here, $V = \\{\\pmb{v}_{\\mathrm{j}_1},\\pmb{v}_{\\mathrm{j}_2},\\dots,\\pmb{v}_{\\mathrm{j}_m}\\}$ and $E = \\{e_{ij}|(\\mathrm{joint}_i,\\mathrm{joint}_j)\\in \\mathrm{links}\\}$ represents the node set of joints and edge set of links respectively, where $m$ denotes the number of joint. We use $f_{i}\\in \\mathbb{R}^{D}$ to denote the value of $i$ -th node feature. It consists of three attributes as follows:", + "bbox": [ + 511, + 612, + 906, + 777 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Joint Coordinate: we use a vector in Cartesian coordinate system, $\\pmb{f}_i^{\\mathrm{JC}} = [\\pmb{x}_i, \\pmb{y}_i, \\pmb{z}_i] \\in \\mathbb{R}^3$ , to denote the absolute coordinates of the $i$ -th joint. The vector is normalized according to the workspace boundary for stable convergence of model training.", + "- Joint Distance: To measure the spatial relationship between the node $\\mathbf{v}_i$ and the other nodes $\\mathbf{v}_j$ , we compute the Euclidean Distance between $\\mathbf{v}_i$ and $\\mathbf{v}_j$ :" + ], + "bbox": [ + 511, + 780, + 905, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "17397", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {f} _ {i} ^ {\\mathrm {J D}} = \\left| \\left| \\boldsymbol {v} _ {i} - \\boldsymbol {v} _ {j} \\right| \\right| _ {2}, \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 223, + 88, + 483, + 108 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\pmb{f}_i^{\\mathrm{JD}}\\in \\mathbb{R}^m$ and $||\\cdot ||_2$ means the Euclidean norm.", + "bbox": [ + 102, + 114, + 475, + 130 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "- Body Label: To discriminate the source of node $\\mathbf{v}_i$ , we use a one-hot vector $\\mathbf{f}_i^{\\mathrm{BL}} = [0,1]$ as one of its features. It can also help the policy capture the motion modes of different robot arms such as the symmetry.", + "bbox": [ + 89, + 131, + 483, + 191 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We concatenate $\\pmb{f}_i^{\\mathrm{JC}}$ , $\\pmb{f}_i^{\\mathrm{JD}}$ , and $\\pmb{f}_i^{\\mathrm{BL}}$ to form the completed feature $\\pmb{f}_i \\in \\mathbb{R}^{D_f}$ , where $D_{f}$ denotes the dimension of node features.", + "bbox": [ + 89, + 191, + 483, + 239 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Spatial-Temporal Graph Learning. Given the same instruction and observation, different historical robot poses result in different predictions. Thus, we build the spatial structure graph with the temporal motion information. Specifically, we build a spatial-temporal graph $G_{\\mathrm{ST}} = \\langle V_{\\mathrm{ST}}, E_{\\mathrm{ST}} \\rangle$ by combining $\\{G_{\\mathrm{S}}^{i} = \\langle V_{\\mathrm{S}}^{i}, E_{\\mathrm{S}}^{i} \\rangle \\}_{i=0}^{T-1}$ from historical timesteps, where $T$ denotes the number of history steps. In $G_{\\mathrm{ST}}$ , the node set $V_{\\mathrm{ST}}$ contains $V_{\\mathrm{S}}^{i}$ from historical static spatial graph. Additionally, we add edges $E'$ linking the same joint node $\\text{joint}_i^t$ at different timesteps, in order to establish the correlation of joint motions at continuous times. It can be formulated as follows:", + "bbox": [ + 89, + 247, + 483, + 429 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nV _ {\\mathrm {S T}} = \\cup \\left\\{V _ {\\mathrm {S}} ^ {i} \\right\\} _ {i = 0} ^ {T - 1}, E _ {\\mathrm {S T}} = \\cup \\left\\{E _ {\\mathrm {S}} ^ {i} \\right\\} _ {i = 0} ^ {T - 1} \\cup E ^ {\\prime}, \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 119, + 438, + 482, + 458 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nE ^ {\\prime} = \\left\\{\\boldsymbol {e} _ {i} ^ {t, t ^ {\\prime}} \\mid t, t ^ {\\prime} \\in \\{0, 1, \\dots , T \\}, t \\neq t ^ {\\prime} \\right\\}. \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 468, + 480, + 487 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this way, we obtain the whole spatial-temporal graph. A Graph Convolutional Network (GCN) is then adopted to propagate and aggregate node features across the graph. The GCN layer updates each node feature $f_{i}$ by aggregating the features of its neighboring nodes, thereby capturing the relational and structural information of the robotic arms. We use the node feature $\\mathbf{H}_{\\mathrm{G}}$ of the last encoder layer as the representation of the robot structure to condition the denoising process.", + "bbox": [ + 89, + 493, + 483, + 628 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3.4. Kinematics Regularizer", + "text_level": 1, + "bbox": [ + 89, + 636, + 302, + 651 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To control the end-effector effectively, the generated pose trajectory must be processed by an Inverse Kinematics (IK) solver, which calculates the joint configurations to achieve the specified poses. However, because the predicted trajectory is generated without considering the robot kinematic constraints, it often falls outside the IK solver's feasible range, resulting in high failure rates during execution. To address this limitation, we propose a kinematics regularizer into the end-effector pose learning objective. This regularizer aligns the predicted poses with the robot kinematic constraints, ensuring that the generated trajectory remains within the solvable space of the IK solver, thus enhancing then reliability of trajectory execution.", + "bbox": [ + 89, + 655, + 483, + 851 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Differentiable Kinematics. Given a joint configuration $\\Theta = [\\theta_1, \\theta_2, \\dots, \\theta_n]$ , the corresponding end-effector pose $\\mathbf{T} \\in SE(3)$ can be computed using forward kinematics,", + "bbox": [ + 89, + 854, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "represented as a mapping $f(\\Theta): \\mathbb{R}^n \\to SE(3)$ . This mapping from joint space to end-effector space is differentiable, namely Differentiable Forward Kinematics (DFK), enabling the use of gradients to optimize our control policy. Leveraging DFK, our policy learns to predict the next joint configuration $\\hat{a}^{\\mathrm{joint}}$ , from which we compute an intermediate end-effector pose $\\mathbf{H}_{\\mathrm{R}}$ . By using $\\mathbf{H}_{\\mathrm{R}}$ as a reference, we guide a denoising process to generate a precise and executable end-effector pose.", + "bbox": [ + 511, + 90, + 903, + 226 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Specifically, we combine structure features $\\mathbf{H}_{\\mathrm{G}}$ with the last hidden state $\\mathbf{H}_{\\mathrm{B}}$ , projecting to the joint space and using the DFK to obtain reference as $\\mathbf{H}_{\\mathrm{R}}$ follows:", + "bbox": [ + 511, + 227, + 903, + 272 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\boldsymbol {a}} ^ {\\text {j o i n t}} = \\operatorname {P r o j} \\left(\\left[ \\mathbf {H} _ {\\mathrm {B}}, \\mathbf {H} _ {\\mathrm {G}} \\right]\\right), \\quad \\mathbf {H} _ {\\mathrm {R}} = \\operatorname {D F K} \\left(\\hat {\\boldsymbol {a}} ^ {\\text {j o i n t}}\\right). \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 531, + 282, + 903, + 301 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To ensure the consistency between predicted and actual joint angles, we minimize the joint loss:", + "bbox": [ + 511, + 313, + 903, + 344 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {j o i n t}} = \\mathbb {E} _ {\\boldsymbol {a} ^ {\\text {j o i n t}} \\sim q (\\boldsymbol {a} _ {0})} \\left[ \\left| \\left| \\boldsymbol {a} ^ {\\text {j o i n t}} - \\hat {\\boldsymbol {a}} ^ {\\text {j o i n t}} \\right| \\right| ^ {2} \\right]. \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 547, + 354, + 903, + 380 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Conditioning Diffusion Process on Kinematics. To enforce kinematic consistency, we condition the diffusion process on the joint angle representation $\\mathbf{H}_{\\mathrm{R}}$ , an auxiliary input encoding kinematic constraints. This allows the predicted pose trajectory to stay within feasible space. Given the diffusion steps from Eq.(9) and Eq.(10), we have:", + "bbox": [ + 511, + 392, + 903, + 483 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {a} _ {k - 1} ^ {\\mathrm {E E}} \\sim \\pi_ {\\phi} \\left(\\boldsymbol {a} _ {k} ^ {\\mathrm {E E}} \\mid \\boldsymbol {a} _ {k - 1} ^ {\\mathrm {E E}}, \\mathbf {H} _ {\\mathrm {B}}, \\mathbf {H} _ {\\mathrm {R}}\\right), \\tag {17}\n$$\n", + "text_format": "latex", + "bbox": [ + 594, + 493, + 903, + 512 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\pi_ {\\phi} \\left(\\boldsymbol {a} _ {k - 1} ^ {\\mathrm {E E}} \\mid \\boldsymbol {a} _ {k} ^ {\\mathrm {E E}}, \\mathbf {H} _ {\\mathrm {B}}, \\mathbf {H} _ {\\mathrm {R}}\\right) =\n$$\n", + "text_format": "latex", + "bbox": [ + 532, + 523, + 718, + 542 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {N} \\left(\\boldsymbol {a} _ {k - 1} ^ {\\mathrm {E E}}; \\sqrt {\\bar {\\alpha} _ {k - 1}} \\pi_ {\\phi} \\left(\\boldsymbol {a} _ {k} ^ {\\mathrm {E E}}, \\mathbf {H} _ {\\mathrm {B}}, \\mathbf {H} _ {\\mathrm {R}}\\right); (1 - \\bar {\\alpha} _ {k - 1}) \\boldsymbol {I}\\right). \\tag {18}\n$$\n", + "text_format": "latex", + "bbox": [ + 532, + 545, + 903, + 579 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Incorporating DFK into the diffusion process allows gradients from the pose loss to propagate back through the kinematic function, ensuring that each denoising step maintains compliance with joint constraints, thereby optimizing the end-effector's control accuracy and robustness.", + "bbox": [ + 511, + 585, + 903, + 661 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Training and Inference", + "text_level": 1, + "bbox": [ + 511, + 671, + 725, + 686 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training. We use the conditional action generation mode to train KStar Diffuser, which is formulated as a conditional denoising diffusion. The loss function is defined as the Mean Square Error (MSE) as follows:", + "bbox": [ + 511, + 694, + 903, + 753 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {a} _ {k} = \\sqrt {\\bar {\\alpha} _ {k}} \\boldsymbol {a} _ {0} + \\sqrt {1 - \\bar {\\alpha} _ {k}} \\epsilon , \\tag {19}\n$$\n", + "text_format": "latex", + "bbox": [ + 612, + 762, + 903, + 780 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {E E}} = \\mathbb {E} _ {\\boldsymbol {a} _ {0} ^ {\\mathrm {E E}} \\sim q (\\boldsymbol {a} _ {0}), k} \\left[ \\left| \\left| \\boldsymbol {a} _ {0} ^ {\\mathrm {E E}} - \\pi_ {\\phi} (\\boldsymbol {a} _ {k} ^ {\\mathrm {E E}}, k, C) \\right| \\right| ^ {2} \\right], \\tag {20}\n$$\n", + "text_format": "latex", + "bbox": [ + 527, + 789, + 903, + 810 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {j o i n t}} = \\mathbb {E} _ {\\boldsymbol {a} _ {0} ^ {\\text {j o i n t}} \\sim q (\\boldsymbol {a} _ {0})} [ \\| \\boldsymbol {a} _ {0} ^ {\\text {j o i n t}} - \\pi_ {\\phi} (\\mathbf {H} _ {\\mathrm {B}}, \\mathbf {H} _ {\\mathrm {G}}) \\| ^ {2} ], \\tag {21}\n$$\n", + "text_format": "latex", + "bbox": [ + 527, + 816, + 903, + 839 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\lambda \\mathcal {L} _ {\\mathrm {E E}} + (1 - \\lambda) \\mathcal {L} _ {\\text {j o i n t}}, \\tag {22}\n$$\n", + "text_format": "latex", + "bbox": [ + 612, + 845, + 903, + 862 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{a}_k$ is obtained by the forward diffusion process and $C$ is the combination of $\\mathbf{H}_{\\mathrm{B}}$ , $\\mathbf{H}_{\\mathrm{G}}$ , $\\mathbf{H}_{\\mathrm{R}}$ . $\\lambda$ is the trade-off coefficient.", + "bbox": [ + 511, + 869, + 906, + 912 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "17398", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Inference. Sampling from a Gaussian noise $\\epsilon_{k}$ , the policy $\\pi_{\\phi}$ performs $K$ iterations to gradually denoise a random noise $\\epsilon_{k}$ into the noise-free action $\\pmb{a}_{0}$ :", + "bbox": [ + 89, + 90, + 482, + 136 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {a} _ {k - 1} ^ {\\mathrm {E E}} = \\sqrt {\\overline {{\\alpha}} _ {k - 1}} \\pi_ {\\phi} (\\boldsymbol {a} _ {k} ^ {\\mathrm {E E}}, k, C) + \\sqrt {1 - \\overline {{\\alpha}} _ {k - 1}} \\epsilon_ {k}. \\tag {23}\n$$\n", + "text_format": "latex", + "bbox": [ + 106, + 142, + 482, + 162 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We use the predicted $a_0^{\\mathrm{EE}}$ as the final action to control robot execution.", + "bbox": [ + 89, + 167, + 482, + 198 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiment", + "text_level": 1, + "bbox": [ + 89, + 210, + 215, + 227 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Dataset and Evaluation Settings", + "text_level": 1, + "bbox": [ + 89, + 234, + 372, + 251 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Dataset. Bimanual manipulation tasks demand high levels of coordination, synchronization, and symmetry awareness between the two robotic arms, making them inherently more challenging than single-arm tasks. To assess the capabilities of KStar Diffuser in these areas, we conducted comprehensive experiments using the RLBench2 benchmark [17], an extended version of RLBench tailored for bimanual manipulation and comprising tasks closely resembling real-world scenarios. Please refer to Appendix for more details about RLBench2 and real-world tasks.", + "bbox": [ + 89, + 257, + 483, + 407 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Settings. To evaluate the policy performance, we employ success rate as the primary metric. Although the policy generates multiple sequential actions during execution, we primarily focus on the final goal achievement rather than intermediate steps. Each task is associated with a specific success criterion defined by its target state. To comprehensively assess the policy's capability, we conduct experiments with varying numbers of demonstrations (20 and 100) during training. Figure 3 illustrates our experimental setup, including both the simulation environment and the Cobot Agilex ALOHA robot. Detailed descriptions of simulation tasks and real-world experimental settings are provided in the Appendix.", + "bbox": [ + 89, + 411, + 483, + 606 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/19ce8d6ab0fd354f2d346adf3d79d5393a255af8688d43ce0902f4732a570abf.jpg", + "image_caption": [ + "Figure 3. The left: the simulation environment of pick_laptop task. The right: the ALOHA device used in the real-world tasks." + ], + "image_footnote": [], + "bbox": [ + 112, + 607, + 483, + 752 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Baselines", + "text_level": 1, + "bbox": [ + 89, + 800, + 197, + 814 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We systematically evaluate KStar Diffuser against state-of-the-art methods in two major categories:", + "bbox": [ + 89, + 821, + 482, + 852 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Transformer-based methods: (1) Action Chunking with Transformers (ACT) [58] employs a Conditional VAE architecture, consisting of an encoder-decoder framework", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "for joint angle sequence prediction; (2) Robotic View Transformer Leader Following (RVT-LF) [17] leverages RVT [15] as its backbone, incorporating a multi-view transformer for cross-view information aggregation and image re-rendering, coupled with a leader following mechanism for action prediction; (3) Perceiver-Actor Leader Following (PerAct-LF) [17] adopts the leader following paradigm based on PerAct [44], utilizing a perceiver Transformer to encode both instructions and voxel observations for optimal voxel action generation; (4) PerAct2 [17] enhances PerAct by implementing a unified feature space for dual-arm actions and employing combined self-attention for synchronized bimanual action prediction.", + "bbox": [ + 511, + 90, + 906, + 287 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Diffusion-based methods: (1) Joint-based Diffusion Policy (DP-J) [9] adopts a diffusion model to robotic manipulation within an imitation learning framework, focusing on joint angle prediction; (2) EndEffector-based Diffusion Policy (DP-EE) is a DP variant which we reimplement Diffusion Policy to predict end-effector poses instead of joint angles, offering an alternative control paradigm; (3) 3D Diffusion Policy (DP3) [56] enhances 3D perception by incorporating point clouds for joint angle prediction.", + "bbox": [ + 511, + 292, + 908, + 430 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Comparison Results with SOTA Methods", + "text_level": 1, + "bbox": [ + 511, + 439, + 864, + 455 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Experimental Results on RLBench2. As shown in Table 1, the KStar Diffuser significantly outperforms other state-of-the-art baselines, achieving more than $20\\%$ higher overall performance with both 20 and 100 training demonstrations. We found that:", + "bbox": [ + 511, + 462, + 905, + 536 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "(1) Similar to learning a single-arm policy, the process of learning a bimanual policy can adapt quickly and achieve a high success rate, given a relatively consistent distribution of task trajectories. In push_box, where the objective is for both arms to push a box toward a specified target along fixed trajectories, our KStar Diffuser and other baseline models perform well. However, as task complexity increases, success rates decrease. For example, in the liftball task, both arms must lift a large ball simultaneously to complete the task. Any asynchrony in movement can cause instability, leading to the ball slipping and ultimately resulting in task failure. Our KStar Diffuser achieves its robust performance on such bimanual tasks by explicitly modeling the spatial and motion relationships between the two arms, surpassing other methods more than $6\\%$ .", + "bbox": [ + 511, + 537, + 906, + 763 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "(2) Distinct from single-arm systems, bimanual robotic systems possess the capability for collaborative manipulation. Methods which are directly adapted from single-arm to bimanual manipulation exhibit high failure rates in tasks, e.g., pick_laptop, as they lack consideration of the spatial and motion relationships between the arms. Specifically, as shown in Figure 4, this task involves picking up a notebook lying flat on a cabinet surface. Given that the notebook rests fully against the tabletop, direct grasping by", + "bbox": [ + 511, + 765, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "17399", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/881164f59b9e47c7a79a073017e63358223d35a5ab0db09282488711775dffb6.jpg", + "table_caption": [ + "Table 1. The experimental result on simulated tasks. We train the policy with the setting of different training demonstrations, i.e. [20, 100], to test its capability comprehensively. The best results are in bold. Each result is reported with three seeds on average." + ], + "table_footnote": [], + "table_body": "
MethodBackbonePush BoxLift BallHandover Item (easy)Pick LaptopSweep DustpanOverall
20 demos
ACT [58]Transformer26.0 ±4.617.3 ±5.80.0 ±0.00.0 ±0.00.0 ±0.08.7 ±1.6
RVT-LF [17]Transformer44.0 ±1.012.3 ±6.70.0 ±0.02.0 ±1.70.0 ±0.011.7 ±1.5
PerAct-LF [17]Transformer63.7 ±3.142.3 ±8.13.3 ±0.67.0 ±2.74.7 ±0.624.2 ±1.6
PerAct2 [17]Transformer59.7 ±7.637.7 ±4.78.0 ±1.017.3 ±0.61.7 ±1.524.9 ±2.1
DP-J [9]Diffusion30.0 ±3.619.7 ±2.10.0 ±0.00.0 ±0.00.0 ±4.010.0 ±1.0
DP-EE [9]Diffusion34.7 ±2.534.3 ±7.40.7 ±0.61.3 ±0.614.0 ±1.717.0 ±1.6
DP3 [56]Diffusion25.3 ±4.034.3 ±7.60.0 ±0.02.7 ±2.90.0 ±0.012.5 ±1.3
KStar Diffuser (Ours)Diffusion79.3 ±3.587.0 ±2.723.7 ±0.617.0 ±2.083.0 ±4.458.0 ±1.4
100 demos
ACT [58]Transformer48.7 ±6.840.7 ±2.10.0 ±0.00.0 ±0.00.0 ±0.017.9 ±1.0
RVT-LF [17]Transformer76.3 ±1.528.7 ±1.00.0 ±0.01.3 ±1.22.3 ±3.221.7 ±0.7
PerAct-LF [17]Transformer66.3 ±3.168.3 ±6.17.3 ±2.114.3 ±3.88.0 ±2.732.9 ±2.5
PerAct2 [17]Transformer77.0 ±3.047.7 ±3.210.3 ±0.634.3 ±4.06.7 ±2.133.9 ±1.0
DP-J [9]Diffusion62.7 ±7.643.3 ±2.10.0 ±0.01.3 ±0.60.0 ±4.021.5 ±1.6
DP-EE [9]Diffusion61.0 ±2.059.7 ±3.52.0 ±1.010.3 ±3.169.7 ±2.540.5 ±1.2
DP3 [56]Diffusion56.0 ±3.664.0 ±2.70.0 ±0.06.3 ±3.11.7 ±2.125.6 ±1.4
KStar Diffuser (Ours)Diffusion83.0 ±1.798.7 ±1.527.0 ±1.743.7 ±4.589.0 ±5.268.2 ±2.1
", + "bbox": [ + 135, + 127, + 864, + 439 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "the robotic arm is not possible. Instead, an effective strategy is to control one arm to push the notebook outward from the cabinet by a short distance, allowing the other arm to pick it up. KStar Diffuser achieves a success rate approximately $9\\%$ higher than other methods, demonstrating its ability to capture the coordinated motion patterns required for collaborative object manipulation between two arms.", + "bbox": [ + 89, + 460, + 483, + 568 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Real-world Experimental Results. To comprehensively evaluate the policy's effectiveness, we build 2 tasks in real-world based on the simulation benchmark. The performances on real-world tasks are illustrated in Table 2.", + "bbox": [ + 89, + 580, + 483, + 642 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Similar to the simulation results, we observe that policies which do not consider about bimanual scenarios, i.e., ACT, DP, and DP3, demonstrate limited capability across all bimanual tasks, achieving around $20\\%$ success ratio on average. Although PerAct2 is designed for bimanual tasks via mapping bimanual actions into a shared learning space, it fails to capture the spatial structure of the bimanual system, leading to ineffective arm coordination during execution. Furthermore, we also found PerAct2 faces significant inverse kinematics issues with its predicted end-effector poses, including joint configuration conflicts and unreachable positions which are shown in Figure 4. It is likely due to PerAct2's limited ability to capture the complex spatial constraints and kinematic relationships within the bimanual robotic system. In contrast, KStar Diffuser achieves superior bimanual coordination, surpassing other methods by over $10\\%$ , as it successfully captures the motion patterns", + "bbox": [ + 89, + 643, + 483, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "between dual arms and predicts feasible end-effector poses.", + "bbox": [ + 511, + 460, + 906, + 478 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/475298fe301cc143007e03a4691ca56ffba3dfefad97d33c54107b3ad89c2988.jpg", + "table_caption": [ + "Table 2. The result of real-world tasks. We train all policies with 100 demonstrations and test 15 times. The best result are in bold." + ], + "table_footnote": [], + "table_body": "
MethodsLift PlateHandoverOverall
ACT [58]37.8 ±8.30.0 ±0.018.9 ±11.2
DP [9]42.1 ±6.50.0 ±0.021.1 ±12.5
DP3 [56]44.0 ±8.30.0 ±0.022.0 ±13.4
PerAct2 [17]51.1 ±6.38.8 ±3.029.9 ±10.2
KStar Diffuser (Ours)66.7 ±5.319.7 ±5.343.1 ±17.8
", + "bbox": [ + 529, + 527, + 893, + 623 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/5448b7cec52f500a71a10722aa82660280095f92ccaccccd7e79a6f9be558bb0.jpg", + "table_caption": [ + "Table 3. Ablation study of model components." + ], + "table_footnote": [], + "table_body": "
ST GraphKRHandover-SHandover-ROverall
✓✓27.0 ±1.719.7 ±5.323.4 ±5.2
✓×18.3 ±1.515.3 ±3.316.8 ±2.1
××16.3 ±3.113.3 ±9.414.8 ±2.1
", + "bbox": [ + 532, + 672, + 890, + 742 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Ablation Studies", + "text_level": 1, + "bbox": [ + 511, + 756, + 679, + 771 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effects of Model Components. To systematically evaluate the contribution of each component in KStar Diffuser, we conduct ablation experiments on the handover-item task in both simulated and real-world environments. We design a progressive ablation process by first removing the Differential Forward Kinematics module while retaining the Spatial-Temporal Graph (ST Graph), and then completely disabling both ST Graph and Kinematics Regularizer (KR).", + "bbox": [ + 511, + 779, + 906, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "17400", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/514c1ca54d213a8ab09bf241b7777aa88f449f206b1e094afc4efcb0626cf4ed.jpg", + "image_caption": [ + "Figure 4. The visualization of bimanual manipulation on simulated RLBench2 and real-world tasks. The blue annotations represent the motion of the robot's left arm, while the green annotations indicate the motion of the right arm." + ], + "image_footnote": [], + "bbox": [ + 101, + 89, + 897, + 407 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The experimental results in Table 3 demonstrate the crucial role of each component. The removal of KR leads to a significant decrease in success rate, particularly pronounced in real-world scenarios. This performance degradation can be attributed to the fundamental differences between simulated and real environments. While simulated environments maintain consistent and noise-free inputs, real-world scenarios introduce various perturbations, e.g., sensor noise and light reflection, making the policy more susceptible to kinematic constraint violations without the regularization effect of KR. Further ablation by removing both ST Graph and KR results in a substantial performance drop across all experimental settings. This observation illustrates two key aspects: First, the ST Graph effectively captures the spatial-temporal dependencies among joints, which is essential for coordinating the relative positioning and interactions between robotic arms. Second, the graph structure's explicit encoding of the robotic physical architecture enhances the policy's robustness against unexpected perturbations by maintaining spatial and temporal coherence.", + "bbox": [ + 88, + 462, + 482, + 765 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Qualitative Analysis", + "text_level": 1, + "bbox": [ + 89, + 772, + 282, + 789 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We further present qualitative analysis in Figure 4. We compare the performance of KStar Diffuser with DP3 and PerAct2 in executing bimanual manipulation tasks within both simulated and real-world environments, respectively.", + "bbox": [ + 89, + 794, + 482, + 854 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In the simulation task, since the laptop rests flat on the cabinet, direct lifting is not feasible. One robotic arm initiates a forward push, creating space, while the other arm", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "concurrently grasps and elevates the laptop. KStar Diffuser effectively models this dual-arm coordination, generating a precise trajectory of synchronized actions. Conversely, DP3, adapted from a single-arm policy to a dual-arm configuration, fails to achieve effective coordination. Concretely, after executing the push motion, the right arm does not halt, obstructing the left arm's lifting process.", + "bbox": [ + 511, + 462, + 906, + 568 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In the real-world task, KStar Diffuser generates an executable item transfer trajectory between the left and right arms, with no collisions throughout the task, reflecting its strong environmental adaptability and collision avoidance capabilities. Conversely, PerAct2 encounters collisions during the handover process (marked in red), indicating less effective handling of dynamic real-world variables and a lack of kinematic awareness on the robot movements.", + "bbox": [ + 511, + 569, + 908, + 690 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 708, + 633, + 723 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we have proposed a novel Kinematics enhanced Spatial-TemporAl gRaph Diffuser (KStar Diffuser) that explicitly incorporates both robot structures and kinematics into the bimanual motion generation process. It consists of a spatial-temporal robot graph that explicitly models the robot physical configuration to guide the generative action denoising procedure, and a kinematics regularizer that augments the NBP learning objective by introducing joint-space supervision. Extensive experiments demonstrate that KStar Diffuser outperforms the baselines by a large margin on both simulation and real-world tasks.", + "bbox": [ + 511, + 734, + 908, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "17401", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Acknowledgements", + "text_level": 1, + "bbox": [ + 89, + 90, + 277, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We would like to thank all co-authors for their efforts and the reviewers for their constructive comments. This work is supported by Shenzhen Science and Technology Program (KQTD20240729102207002), National Natural Science Foundation of China (Grant No. 62406092), National Natural Science Foundation of China (Grant No. U24B20175), Shenzhen Science and Technology Program (Grant No. KJZD20240903100017022), Guangdong Basic and Applied Basic Research Foundation (2025A1515010169), and Research on Efficient Exploration and Self-Evolution of APP Agents & Embodied Intelligent Cerebellum Control Model and Collaborative Feedback Training Project (Grant No. TC20240403047). Additionally, it is also supported by National Natural Science Foundation of China (Grant No. 62306090, No. 62236003), Natural Science Foundation of Guangdong Province of China (Grant No. 2024A1515010147).", + "bbox": [ + 89, + 114, + 485, + 371 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 383, + 187, + 400 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Lucia Angelini, Manuela Uliano, Angela Mazzeo, Mattia Penzotti, and Marco Controzzi. Self-collision avoidance in bimanual teleoperation using collisionik: algorithm revision and usability experiment. In 2022 IEEE-RAS 21st International Conference on Humanoid Robots (Humanoids), pages 112-118, 2022. 2", + "[2] Andreas Aristidou, Joan Lasenby, Yiorgos Chrysanthou, and Ariel Shamir. Inverse kinematics techniques in computer graphics: A survey. In Computer graphics forum, pages 35-58. Wiley Online Library, 2018. 2", + "[3] Weibang Bai, Ningshan Zhang, Baoru Huang, Ziwei Wang, Francesco Cursi, Ya-Yen Tsai, Bo Xiao, and Eric M Yeatman. Dual-arm coordinated manipulation for object twisting with human intelligence. In 2021 IEEE International Conference on Systems, Man, and Cybernetics (SMC), pages 902-908. IEEE, 2021. 2", + "[4] Aaron M Bestick, Samuel A Burden, Giorgia Willits, Nikhil Naikal, S Shankar Sastry, and Ruzena Bajcsy. Personalized kinematics for human-robot collaborative manipulation. In 2015 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 1037–1044. IEEE, 2015. 2", + "[5] Botond Bócsi, Duy Nguyen-Tuong, Lehel Csato, Bernhard Schoelkopf, and Jan Peters. Learning inverse kinematics with structured prediction. In 2011 IEEE/RSJ International Conference on Intelligent Robots and Systems, pages 698-703. IEEE, 2011. 3", + "[6] Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Joseph Dabis, Chelsea Finn, Keerthana Gopalakrishnan, Karol Hausman, Alex Herzog, Jasmine Hsu, et al. Rt-1: Robotics transformer for real-world control at scale. arXiv preprint arXiv:2212.06817, 2022. 1", + "[7] Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Xi Chen, Krzysztof Choromanski, Tianli Ding, Danny Driess, Avinava Dubey, Chelsea Finn, et al. Rt-2:" + ], + "bbox": [ + 99, + 410, + 482, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Vision-language-action models transfer web knowledge to robotic control. arXiv preprint arXiv:2307.15818, 2023. 1", + "[8] Shizhe Chen, Ricardo Garcia, Cordelia Schmid, and Ivan Laptev. Polarnet: 3d point clouds for language-guided robotic manipulation. arXiv preprint arXiv:2309.15596, 2023.2", + "[9] Cheng Chi, Zhenjia Xu, Siyuan Feng, Eric Cousineau, Yilun Du, Benjamin Burchfiel, Russ Tedrake, and Shuran Song. Diffusion policy: Visuomotor policy learning via action diffusion. The International Journal of Robotics Research, page 02783649241273668, 2023. 1, 2, 3, 4, 6, 7", + "[10] Cheng Chi, Zhenjia Xu, Chuer Pan, Eric Cousineau, Benjamin Burchfiel, Siyuan Feng, Russ Tedrake, and Shuran Song. Universal manipulation interface: In-the-wild robot teaching without in-the-wild robots. arXiv preprint arXiv:2402.10329, 2024. 1, 2, 3", + "[11] Akos Csiszar, Jan Eilers, and Alexander Verl. On solving the inverse kinematics problem using neural networks. In 2017 24th International Conference on Mechatronics and Machine Vision in Practice (M2VIP), pages 1-6, 2017. 3", + "[12] Aaron D'Souza, Sethu Vijayakumar, and Stefan Schaal. Learning inverse kinematics. In Proceedings 2001 IEEE/RSJ International Conference on Intelligent Robots and Systems. Expanding the Societal Role of Robotics in the Next Millennium (Cat. No. 01CH37180), pages 298-303. IEEE, 2001. 3", + "[13] Zipeng Fu, Tony Z Zhao, and Chelsea Finn. Mobile aloha: Learning bimanual mobile manipulation with low-cost whole-body teleoperation. arXiv preprint arXiv:2401.02117, 2024. 1, 3", + "[14] Jianfeng Gao, Xiaoshu Jin, Franziska Krebs, Noémie Jaquier, and Tamim Asfour. Bi-kvil: Keypoints-based visual imitation learning of bimanual manipulation tasks. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 16850-16857. IEEE, 2024. 2", + "[15] Ankit Goyal, Jie Xu, Yijie Guo, Valts Blukis, Yu-Wei Chao, and Dieter Fox. Rvt: Robotic view transformer for 3d object manipulation. In Conference on Robot Learning, pages 694–710. PMLR, 2023. 1, 2, 3, 6", + "[16] Reinhard Grassmann, Vincent Modes, and Jessica Burgner-Kahrs. Learning the forward and inverse kinematics of a 6-dof concentric tube continuum robot in se (3). In 2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 5125-5132. IEEE, 2018. 2", + "[17] Markus Grotz, Mohit Shridhar, Tamim Asfour, and Dieter Fox. Peract2: A perceiver actor framework for bimanual manipulation tasks. arXiv preprint arXiv:2407.00278, 2024. 1, 2, 3, 6, 7", + "[18] Abdullah Aamir Hayat, Ratan OM Sadanand, and Subir K Saha. Robot manipulation through inverse kinematics. In Proceedings of the 2015 conference on advances in robotics, pages 1-6, 2015. 2", + "[19] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2", + "[20] Binghao Huang, Yixuan Wang, Xinyi Yang, Yiyue Luo, and Yunzhu Li. 3d-vitac: Learning fine-grained manipulation" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "17402", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "with visuo-tactile sensing. arXiv preprint arXiv:2410.24091, 2024. 2", + "[21] Kenneth H Hunt. Structural kinematics of in-parallel-actuated robot-arms. 1983. 1", + "[22] Michael I Jordan and David E Rumelhart. Forward models: Supervised learning with a distal teacher. In Backpropagation, pages 189-236. Psychology Press, 2013. 3", + "[23] Tsung-Wei Ke, Nikolaos Gkanatsios, and Katerina Fragkiadaki. 3d diffuser actor: Policy diffusion with 3d scene representations. arXiv preprint arXiv:2402.10885, 2024. 2, 3", + "[24] Moo Jin Kim, Karl Pertsch, Siddharth Karamcheti, Ted Xiao, Ashwin Balakrishna, Suraj Nair, Rafael Rafailov, Ethan Foster, Grace Lam, Pannag Sanketi, et al. Openvla: An open-source vision-language-action model. arXiv preprint arXiv:2406.09246, 2024. 2", + "[25] Thomas N Kipf and Max Welling. Semi-supervised classification with graph convolutional networks. arXiv preprint arXiv:1609.02907, 2016. 2", + "[26] CS George Lee. Robot arm kinematics, dynamics, and control. Computer, 15(12):62-80, 1982. 1", + "[27] Maolin Lei, Ting Wang, Chen Yao, Huan Liu, Zhi Wang, and Yongsheng Deng. Real-time kinematics-based self-collision avoidance algorithm for dual-arm robots. Applied Sciences, 10(17):5893, 2020. 1", + "[28] Jiefeng Li, Chao Xu, Zhicun Chen, Siyuan Bian, Lixin Yang, and Cewu Lu. Hybrik: A hybrid analytical-neural inverse kinematics solution for 3d human pose and shape estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3383-3393, 2021. 3", + "[29] Jiefeng Li, Chao Xu, Zhicun Chen, Siyuan Bian, Lixin Yang, and Cewu Lu. Hybrik: A hybrid analytical-neural inverse kinematics solution for 3d human pose and shape estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3383-3393, 2021. 3", + "[30] I Liu, Chun Arthur, Sicheng He, Daniel Seita, and Gaurav Sukhatme. Voxact-b: Voxel-based acting and stabilizing policy for bimanual manipulation. arXiv preprint arXiv:2407.04152, 2024. 2", + "[31] Xiao Ma, Sumit Patidar, Iain Haughton, and Stephen James. Hierarchical diffusion policy for kinematics-aware multitask robotic manipulation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18081-18090, 2024. 1, 2", + "[32] Matthew T Mason. Mechanics of robotic manipulation. MIT press, 2001. 2", + "[33] Yao Mu, Qinglong Zhang, Mengkang Hu, Wenhai Wang, Mingyu Ding, Jun Jin, Bin Wang, Jifeng Dai, Yu Qiao, and Ping Luo. Embodiedgpt: Vision-language pre-training via embodied chain of thought. Advances in Neural Information Processing Systems, 36, 2024. 1", + "[34] Yoshihiko Nakamura, Kiyoshi Nagai, and Tsuneo Yoshikawa. Dynamics and stability in coordination of multiple robotic mechanisms. The International Journal of Robotics Research, 8(2):44-61, 1989. 1", + "[35] Valerio Ortenzi, Naresh Marturi, Michael Mistry, Jeffrey Kuo, and Rustam Stolkin. Vision-based framework" + ], + "bbox": [ + 91, + 90, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "to estimate robot configuration and kinematic constraints. IEEE/ASME Transactions on Mechatronics, 23(5):2402-2412, 2018. 2", + "[36] Eric Paljug, Xiaoping Yun, and Vijay Kumar. Control of rolling contacts in multi-arm manipulation. IEEE Transactions on Robotics and Automation, 10(4):441-452, 1994. 1", + "[37] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4195-4205, 2023. 2", + "[38] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 2", + "[39] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 2", + "[40] Allen Z Ren, Justin Lidard, Lars L Ankile, Anthony Simeonov, Pulkit Agrawal, Anirudha Majumdar, Benjamin Burchfiel, Hongkai Dai, and Max Simchowitz. Diffusion policy policy optimization. arXiv preprint arXiv:2409.00588, 2024. 2", + "[41] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2", + "[42] Nilanjan Sarkar, Xiaoping Yun, and Vijay Kumar. Dynamic control of 3-d rolling contacts in two-arm manipulation. IEEE Transactions on Robotics and Automation, 13(3): 364-376, 1997. 1", + "[43] Youshik Shin and Zeungnam Bien. Collision-free trajectory planning for two robot arms. Robotica, 7(3):205-212, 1989. 1, 2", + "[44] Mohit Shridhar, Lucas Manuelli, and Dieter Fox. Perceiver-actor: A multi-task transformer for robotic manipulation. In Conference on Robot Learning, pages 785–799. PMLR, 2023. 1, 2, 3, 6", + "[45] Christian Smith, Yiannis Karayiannidis, Lazaros Nalpantidis, Xavi Gratal, Peng Qi, Dimos V Dimarogonas, and Danica Kragic. Dual arm manipulation—a survey. Robotics and Autonomous systems, 60(10):1340-1353, 2012. 1", + "[46] Octo Model Team, Dibya Ghosh, Homer Walke, Karl Pertsch, Kevin Black, Oier Mees, Sudeep Dasari, Joel Hejna, Tobias Kreiman, Charles Xu, et al. Octo: An open-source generalist robot policy. arXiv preprint arXiv:2405.12213, 2024. 2", + "[47] Vitalis Vosylius, Younggyo Seo, Jafar Uruç, and Stephen James. Render and diffuse: Aligning image and action spaces for diffusion-based behaviour cloning. arXiv preprint arXiv:2405.18196, 2024. 2", + "[48] Lirui Wang, Xinlei Chen, Jialiang Zhao, and Kaiming He. Scaling proprioceptive-visual learning with heterogeneous pre-trained transformers. arXiv preprint arXiv:2409.20537, 2024. 2" + ], + "bbox": [ + 516, + 92, + 906, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "17403", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[49] Yixiao Wang, Yifei Zhang, Mingxiao Huo, Ran Tian, Xiang Zhang, Yichen Xie, Chenfeng Xu, Pengliang Ji, Wei Zhan, Mingyu Ding, et al. Sparse diffusion policy: A sparse, reusable, and flexible policy for robot learning. arXiv preprint arXiv:2407.01531, 2024. 2", + "[50] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. Advances in Neural Information Processing Systems, 36, 2024. 2", + "[51] Chris Welman. Inverse kinematics and geometric constraints for articulated figure manipulation. 1993. 2", + "[52] Kun Wu, Yichen Zhu, Jinming Li, Junjie Wen, Ning Liu, Zhiyuan Xu, Qinru Qiu, and Jian Tang. Discrete policy: Learning disentangled action space for multi-task robotic manipulation. arXiv preprint arXiv:2409.18707, 2024. 2", + "[53] Wenke Xia, Dong Wang, Xincheng Pang, Zhigang Wang, Bin Zhao, Di Hu, and Xuelong Li. Kinematic-aware prompting for generalizable articulated object manipulation with llms. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 2073-2080. IEEE, 2024. 2", + "[54] Andrea Maria Zanchettin, Paolo Rocco, Luca Bascetta, Ioannis Symeonidis, and Steffen Peldschus. Kinematic analysis and synthesis of the human arm motion during a manipulation task. In 2011 IEEE international conference on robotics and automation, pages 2692-2697. IEEE, 2011. 2", + "[55] Yanjie Ze, Zixuan Chen, Wenhao Wang, Tianyi Chen, Xialin He, Ying Yuan, Xue Bin Peng, and Jiajun Wu. Generalizable humanoid manipulation with improved 3d diffusion policies. arXiv preprint arXiv:2410.10803, 2024. 1, 2", + "[56] Yanjie Ze, Gu Zhang, Kangning Zhang, Chenyuan Hu, Muhan Wang, and Huazhe Xu. 3d diffusion policy. arXiv preprint arXiv:2403.03954, 2024. 1, 2, 6, 7", + "[57] Junjie Zhang, Chenjia Bai, Haoran He, Wenke Xia, Zhigang Wang, Bin Zhao, Xiu Li, and Xuelong Li. Sam-e: Leveraging visual foundation model with sequence imitation for embodied manipulation. arXiv preprint arXiv:2405.19586, 2024. 2", + "[58] Tony Z Zhao, Vikash Kumar, Sergey Levine, and Chelsea Finn. Learning fine-grained bimanual manipulation with low-cost hardware. arXiv preprint arXiv:2304.13705, 2023. 1, 2, 3, 6, 7", + "[59] Zangwei Zheng, Xiangyu Peng, Tianji Yang, Chenhui Shen, Shenggui Li, Hongxin Liu, Yukun Zhou, Tianyi Li, and Yang You. Open-sora: Democratizing efficient video production for all, march 2024. URL https://github.com/hpcaitech/Open-Sora, 1(3):4.2", + "[60] Minjie Zhu, Yichen Zhu, Jinming Li, Junjie Wen, Zhiyuan Xu, Ning Liu, Ran Cheng, Chaomin Shen, Yaxin Peng, Feifei Feng, et al. Scaling diffusion policy in transformer to 1 billion parameters for robotic manipulation. arXiv preprint arXiv:2409.14411, 2024. 2" + ], + "bbox": [ + 91, + 90, + 480, + 825 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "17404", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Spatial-Temporal Graph Diffusion Policy with Kinematic Modeling for Bimanual Robotic Manipulation/8300c221-08df-4f6a-8aec-775be3434ff5_model.json b/2025/Spatial-Temporal Graph Diffusion Policy with Kinematic Modeling for Bimanual Robotic Manipulation/8300c221-08df-4f6a-8aec-775be3434ff5_model.json new file mode 100644 index 0000000000000000000000000000000000000000..5fadfd4b8f5816ae72e0b91ad4e9f5487ec6176e --- /dev/null +++ b/2025/Spatial-Temporal Graph Diffusion Policy with Kinematic Modeling for Bimanual Robotic Manipulation/8300c221-08df-4f6a-8aec-775be3434ff5_model.json @@ -0,0 +1,2477 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.143, + 0.13, + 0.856, + 0.177 + ], + "angle": 0, + "content": "Spatial-Temporal Graph Diffusion Policy with Kinematic Modeling for Bimanual Robotic Manipulation" + }, + { + "type": "text", + "bbox": [ + 0.167, + 0.202, + 0.83, + 0.24 + ], + "angle": 0, + "content": "Qi Lv\\(^{1,2,3}\\) Hao Li\\(^{1,3}\\) Xiang Deng\\(^{1\\dagger}\\) Rui Shao\\(^{1}\\) Yinchuan Li\\(^{3}\\) Jianye Hao\\(^{3}\\) Longxiang Gao\\(^{4}\\) Michael Yu Wang\\(^{2}\\) Liqiang Nie\\(^{1\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.144, + 0.242, + 0.854, + 0.296 + ], + "angle": 0, + "content": "\\(^{1}\\)School of Computer Science and Technology, Harbin Institute of Technology (Shenzhen) \n\\(^{2}\\)School of Engineering, Greate Bay University \n\\(^{3}\\)Huawei Noah's Ark Lab \n\\(^{4}\\)Qilu University of Technology (Shandong Academy of Sciences)" + }, + { + "type": "text", + "bbox": [ + 0.258, + 0.301, + 0.735, + 0.318 + ], + "angle": 0, + "content": "lvqi@stu.hit.edu.cn, dengxiang@hit.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.367, + 0.327, + 0.383 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.399, + 0.486, + 0.792 + ], + "angle": 0, + "content": "Despite the significant success of imitation learning in robotic manipulation, its application to bimanual tasks remains highly challenging. Existing approaches mainly learn a policy to predict a distant next-best end-effector pose (NBP) and then compute the corresponding joint rotation angles for motion using inverse kinematics. However, they suffer from two important issues: (1) rarely considering the physical robotic structure, which may cause self-collisions or interferences, and (2) overlooking the kinematics constraint, which may result in the predicted poses not conforming to the actual limitations of the robot joints. In this paper, we propose Kinematics enhanced Spatial-TemporAl gRaph Diffuser (KStar Diffuser). Specifically, (1) to incorporate the physical robot structure information into action prediction, KStar Diffuser maintains a dynamic spatial-temporal graph according to the physical bimanual joint motions at continuous timesteps. This dynamic graph serves as the robot-structure condition for denoising the actions; (2) to make the NBP learning objective consistent with kinematics, we introduce the differentiable kinematics to provide the reference for optimizing KStar Diffuser. This module regularizes the policy to predict more reliable and kinematics-aware next end-effector poses. Experimental results show that our method effectively leverages the physical structural information and generates kinematics-aware actions in both simulation and real-world." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.808, + 0.222, + 0.823 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.834, + 0.484, + 0.865 + ], + "angle": 0, + "content": "Bimanual manipulation [34, 36, 42, 45] represents a fundamental capability for robotic systems to perform com" + }, + { + "type": "image", + "bbox": [ + 0.547, + 0.352, + 0.875, + 0.513 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.524, + 0.907, + 0.553 + ], + "angle": 0, + "content": "Figure 1. The self-collision problem in bimanual manipulation tasks due to overlooking the robotic structure." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.568, + 0.907, + 0.674 + ], + "angle": 0, + "content": "plex tasks requiring two-arm coordination. While imitation learning has demonstrated remarkable success in single-arm manipulation [6, 7, 33], its extension to bimanual scenarios faces unique challenges as robots need to coordinate dual-arm movements while conforming to physical constraints. These challenges significantly impact the reliability and feasibility of predicted actions in real-world applications." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.675, + 0.909, + 0.885 + ], + "angle": 0, + "content": "A straightforward approach [9, 10, 13, 55, 56, 58] would be to learn the entire motion trajectory by predicting joint positions at consecutive timesteps. However, this poses significant challenges as trajectories are typically long and difficult to learn. Therefore, mainstream approaches [15, 17, 31, 44] typically adopt a two-stage pipeline: first predicting the next-best end-effector pose (NBP), then computing joint rotations through inverse kinematics [21, 26]. However, the dominant NBP-based approaches, though simplify the learning process, often lead to unreliable motion generation. Concretely, the predicted poses may violate physical structure constraints [27, 43], causing potential arm collisions, or exceeding joint limits due to inadequate consideration of kinematic feasibility, as Figure 1 illustrated." + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.886, + 0.905, + 0.902 + ], + "angle": 0, + "content": "In real-world bimanual manipulation tasks, existing ap" + }, + { + "type": "page_footnote", + "bbox": [ + 0.09, + 0.875, + 0.483, + 0.901 + ], + "angle": 0, + "content": "*This work was done when Qi Lv and Hao Li were intern at Huawei Noah's Ark Lab. † Corresponding author" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17394" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.091, + 0.482, + 0.394 + ], + "angle": 0, + "content": "proaches [9, 15, 17, 44, 57] exhibit significant limitations in motion execution despite their promising performance in trajectory prediction. Empirical observations indicate that while individual end-effector poses are feasible in isolation, their concurrent execution frequently results in interarm collisions [43] and kinematically infeasible configurations [31, 35]. The fundamental challenge lies in the optimization of end-effector poses exclusively in Cartesian space, which introduces a substantial discrepancy between motion prediction and physical constraints. Existing approaches [9, 10, 17] incorporate proprioception information in a simplistic manner, (e.g., gripper rotation angles [17, 44] and end-effector poses [48]), while disregarding the robot's intrinsic structural properties e.g., the kinematic chain and joint configurations during the prediction phase. This oversimplified representation fails to capture the crucial spatial correlations between robotic arms, joint links, and their potential interactions. Consequently, the predicted dual-arm trajectories may satisfy task objectives in Cartesian space while violating crucial spatial constraints during execution." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.398, + 0.483, + 0.595 + ], + "angle": 0, + "content": "Moreover, the conventional paradigm [17, 24, 44, 46] of treating inverse kinematics as a post-processing procedure introduces additional complexities. The disjunction between kinematic constraints and the pose prediction learning objective often yields solutions that exhibit apparent smoothness in Cartesian space while manifesting discontinuous or mechanically infeasible trajectories in the joint space. This limitation becomes particularly pronounced in configurations approaching kinematic singularities or joint limits, where the mapping between Cartesian and joint spaces becomes ill-conditioned. These inherent limitations necessitate the integration of both structural and kinematic constraints into the policy learning framework." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.599, + 0.483, + 0.901 + ], + "angle": 0, + "content": "To address the above issues, we propose Kinematics enhanced Spatial-TemporAl gRaph Diffuser (KStar Diffuser), a novel framework that explicitly incorporates both robot structures and kinematics into the bimanual motion generation process. Our key insight is that the physical structure and kinematic properties of the robot should guide the learning process of pose prediction, rather than be treated as independent post-processing constraints. Specifically, (1) to incorporate structural awareness, we construct a dynamic spatial-temporal graph from the robot's URDF specifications, where nodes represent joint properties and edges capture both spatial relations and temporal dependencies. This graph structure is encoded via Graph Convolutional Network (GCN) [25] to provide explicit physical constraints for the diffusion process; (2) for kinematic feasibility, we regularize the NBP learning objective by incorporating joint-space prediction, where the predicted joint positions are mapped to reference end-effector poses through differentiable forward kinematics. These kinematically-feasible poses then serve as conditions" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.136 + ], + "angle": 0, + "content": "to guide the diffusion process, ensuring that the generated motions satisfy both structural and kinematic constraints. Our main contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.137, + 0.905, + 0.213 + ], + "angle": 0, + "content": "- Different from existing approaches that optimize end-effector poses solely in Cartesian space, we propose a novel spatial-temporal robot graph that explicitly models the robot physical configuration to guide the generative action denoising procedure." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.213, + 0.905, + 0.302 + ], + "angle": 0, + "content": "- We introduce a kinematics regularizer that augments the NBP learning objective by introducing joint-space supervision. This regularizer leverages forward kinematics to provide kinematically-feasible reference poses, effectively guiding the diffusion process to conform to kinematic constraints." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.303, + 0.905, + 0.35 + ], + "angle": 0, + "content": "- Extensive experiments show that our proposed KStar Diffuser is superior in both simulation and real-world scenarios, surpassing baselines more than \\(10\\%\\) in success ratio." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.137, + 0.905, + 0.35 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.36, + 0.655, + 0.374 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.381, + 0.905, + 0.548 + ], + "angle": 0, + "content": "Diffusion Models in Bimanual Robotic Controls. Diffusion models such as Denoising Diffusion Probabilistic Models (DDPM) [19] have achieved great success in the fields of image generation [38, 39, 41] and video generation [37, 50, 59]. Thus, recent work [8, 40, 47, 52, 60] has been devoted to applying its powerful generation capabilities to action generation for robotic manipulation tasks in a imitation learning mode. With the 2D images [9, 31, 40, 47, 49] or 3D point cloud observations [8, 20, 23, 55, 56], the policy is trained to output the action sequence of joint positions or end-effector poses by iterative denoising process." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.549, + 0.906, + 0.821 + ], + "angle": 0, + "content": "Bimanual manipulation tasks, which resemble human-like actions more closely, have garnered increasing attention [10, 14, 17, 30, 58]. However, existing research primarily extends single-arm manipulation methods to dual-arm tasks [10, 58], without considering about the distinct challenges specific to bimanual settings. The relative independence of the two arms makes it essential to consider self-collision avoidance [1, 3] when predicting actions. In practice, the robot's structure significantly influences its motion since the joint types determine movement directions and joint angles constrain the range of motion. Although some studies [15, 44, 48] integrate proprioception in robotic control, they often use simply low-dimensional vectors to represent body information, lacking depth in structural exploration and overlooking critical spatial relationships between dual arms. Therefore, we study the physical structure information of the dual-arm robot system, and improve both the accuracy and adaptability of robotic movements." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.826, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Robotic Kinematics Modeling. Robotic kinematics modeling [2, 16, 31, 53, 54] is a classic problem in the robotic control [4, 18, 32, 51] where inverse kinematics (IK) presents a fundamental challenge, i.e., deriving the joint configuration given the end-effector pose. The inher" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "17395" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.351 + ], + "angle": 0, + "content": "ent non-uniqueness of IK solutions as a significant obstacle for learning algorithms [12, 22], often produce inaccurate models by averaging over nonconvex feasible sets. Recent advancements integrate neural solutions [11, 28, 29] to enhance the adaptability and precision of kinematic models. Bócsi et al. [5] leveraged support vector machines to parameterize quadratic programs, aligning solutions with IK in specific workspace regions. Current data-driven imitation learning approaches [10, 13, 15, 17, 58] primarily rely on probabilistic modeling of action distributions to predict subsequent actions, but lack the reliability guarantees from the aspect of kinematics. In this paper, we construct a spatial-temporal graph to learn the robot physical representation, and provide a kinematics-aware latent end-effect embedding for diffusion policy as guidance. In this way, o'ur approach enhances the robot's ability to perform tasks with superious precision and adaptability." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.367, + 0.183, + 0.383 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.393, + 0.221, + 0.41 + ], + "angle": 0, + "content": "3.1. Preliminary" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.417, + 0.484, + 0.508 + ], + "angle": 0, + "content": "Diffusion Policy. Chi et al. [9] propose the diffusion policy which represents a robot's visuomotor policy as a conditional denoising diffusion process. It learns a model distribution \\(\\pi_{\\phi}(\\pmb{a}_0|\\pmb{o})\\) conditioned on observation \\(\\pmb{o}\\) to approximate joint distribution \\(q(\\pmb{a}_0)\\). The whole procedure consists of a forward process and a reverse process." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.51, + 0.484, + 0.556 + ], + "angle": 0, + "content": "(1) Forward Process: For the Markov chain with Gaussian transitions parameterized, the policy \\(\\pi_{\\phi}(\\pmb{a}_0|\\pmb{o})\\) learns a fixed inference procedure \\(q(\\pmb{a}_k|\\pmb{a}_0)\\):" + }, + { + "type": "equation", + "bbox": [ + 0.164, + 0.57, + 0.482, + 0.62 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} q \\left(\\boldsymbol {a} _ {k} \\mid \\boldsymbol {a} _ {0}\\right) = \\int q \\left(\\boldsymbol {a} _ {1: k} \\mid \\boldsymbol {a} _ {0}\\right) \\mathrm {d} \\boldsymbol {a} _ {1: (k - 1)} \\tag {1} \\\\ = \\mathcal {N} \\left(\\boldsymbol {a} _ {k}; \\sqrt {\\bar {\\alpha} _ {k}} \\boldsymbol {a} _ {0}, (1 - \\bar {\\alpha} _ {k}) \\boldsymbol {I}\\right) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.635, + 0.484, + 0.664 + ], + "angle": 0, + "content": "Thus, \\(a_{k}\\) can be expressed as a linear combination of \\(a_0\\) and a noise variable \\(\\epsilon\\):" + }, + { + "type": "equation", + "bbox": [ + 0.196, + 0.68, + 0.483, + 0.697 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {a} _ {k} = \\sqrt {\\bar {\\alpha} _ {k}} \\boldsymbol {a} _ {0} + \\sqrt {1 - \\bar {\\alpha} _ {k}} \\varepsilon \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.712, + 0.285, + 0.727 + ], + "angle": 0, + "content": "The training loss is obtained:" + }, + { + "type": "equation", + "bbox": [ + 0.105, + 0.741, + 0.482, + 0.773 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathbb {E} _ {\\boldsymbol {a} _ {0} \\sim q (\\boldsymbol {a} _ {0}), \\epsilon_ {k} \\sim \\mathcal {N} (\\boldsymbol {0}, \\boldsymbol {I}), k} \\left[ \\left| \\left| \\epsilon_ {k}, \\pi_ {\\phi} (\\boldsymbol {a} _ {0} + \\epsilon_ {k}, \\boldsymbol {o}, k) \\right| \\right| ^ {2} \\right] \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.774, + 0.442, + 0.788 + ], + "angle": 0, + "content": "where \\(\\epsilon_{k}\\) is the random noise sampling at \\(k\\) iteration." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.79, + 0.484, + 0.821 + ], + "angle": 0, + "content": "(2) Reverse Process: Starting from a sample \\( \\pmb{a}_K \\sim \\mathcal{N}(\\pmb{0}, \\pmb{I}) \\), the reverse steps are:" + }, + { + "type": "equation", + "bbox": [ + 0.134, + 0.835, + 0.483, + 0.853 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {a} _ {k - 1} = \\sqrt {\\bar {\\alpha} _ {k - 1}} \\pi_ {\\phi} (a _ {k}, k, o) + \\sqrt {1 - \\bar {\\alpha} _ {k - 1}} \\epsilon . \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.867, + 0.484, + 0.914 + ], + "angle": 0, + "content": "By the iterative denoising process, the policy generates \\( a_0 \\) as its next action. The continuous predicted action sequence forms the complete action trajectory." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.151 + ], + "angle": 0, + "content": "Robotic Kinematics. Robotics kinematics describes the relationship between the robot joints and its end-effector. It can be divided into forward kinematics and inverse kinematics for controlling robot movements." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.152, + 0.907, + 0.228 + ], + "angle": 0, + "content": "(1) Forward Kinematics (FK): Given a specific configuration of joint angles, i.e., \\(\\theta = [\\theta_{1},\\theta_{2},\\dots,\\theta_{n}]\\in \\Theta\\), it aims to construct a function \\(f_{\\mathrm{FK}}:\\mathbb{R}^n\\to SE(3)\\), computing the end-effector position and orientation that are represented as a pose matrix \\(\\mathbf{T}\\in SE(3)\\)." + }, + { + "type": "equation", + "bbox": [ + 0.554, + 0.234, + 0.907, + 0.26 + ], + "angle": 0, + "content": "\\[\nF K (\\theta) = \\mathbf {T} \\in \\mathcal {T}, \\text {w h e r e} \\mathbf {T} = \\prod_ {i = 1} ^ {n} \\mathbf {T} _ {i} (\\theta_ {i}), (5)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.268, + 0.906, + 0.328 + ], + "angle": 0, + "content": "where \\(\\mathbf{T}_i(\\theta_i)\\) is the homogeneous transformation matrix for joint \\(i\\), and \\(\\mathcal{T}\\) means the task space. \\(\\Theta \\in \\mathbb{R}^n\\) denotes the configuration space, while the configuration refers to the specific arrangement of the robotic joints in a given pose." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.328, + 0.906, + 0.405 + ], + "angle": 0, + "content": "(2) Inverse Kinematics (IK): In contrast, IK aims to find the joint configurations which achieve a desired end-effector pose. It is defined as a mapping \\( f_{\\mathrm{IK}}: SE(3) \\to \\mathbb{R}^n \\) from the task space \\( \\mathcal{T} \\) to the set of possible configurations \\( \\Theta \\) that satisfy the given pose matrix \\( \\mathbf{T} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.598, + 0.415, + 0.905, + 0.431 + ], + "angle": 0, + "content": "\\[\nI K (\\mathcal {T}) = \\{\\theta \\in \\Theta | F K (\\theta) = \\mathbf {T} \\}. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.441, + 0.906, + 0.471 + ], + "angle": 0, + "content": "It is noted that this mapping is often non-unique, especially for redundant manipulators with more than 6 DoF." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.48, + 0.667, + 0.495 + ], + "angle": 0, + "content": "3.2. Task Definition" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.502, + 0.907, + 0.686 + ], + "angle": 0, + "content": "Given the dataset \\(D\\) with language instructions \\(\\pmb{l}\\) and the RGB-D observations \\(\\pmb{o}\\), we aim to learn a policy \\(\\pi_{\\phi}(\\pmb{a}| \\pmb{o}, \\pmb{l})\\) which predicts actions \\(\\pmb{a}\\). Here, \\(\\pmb{a}\\) is composed of a trajectory \\(\\pmb{a}^{\\mathrm{joint}} = \\{\\pmb{a}_0, \\pmb{a}_1, \\dots, \\pmb{a}_T\\}\\) and gripper opening or closing action \\(\\pmb{a}^{\\mathrm{gripper}}\\), where \\(T\\) denotes the trajectory length and \\(\\pmb{a}_i \\in \\mathbb{R}^n\\). \\(n\\) represents the number of robot joints. In the bimanual manipulation, \\(n\\) is usually to be 12 or 14 as each robot arm has 6 or 7-DoF. Referring to prior works [17, 44], it is inefficient to train the policy on all trajectory points. Thus, a keyframe discovery method is used to extract a set of \\(K_{\\xi}\\) keyframe indices \\(\\{k_i\\}_{i=0}^{K_{\\xi}}\\) and the prediction actions are the key set of end-effector poses \\(\\pmb{a}^{\\mathrm{EE}}\\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.694, + 0.666, + 0.708 + ], + "angle": 0, + "content": "3.3. KStar Diffuser" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.716, + 0.627, + 0.73 + ], + "angle": 0, + "content": "3.3.1. Overview" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Mainstream methods [17, 23, 58] train a policy to predict actions, but take little consideration about the mechanical robot structure which determines its motion. We thus propose a spatial-temporal graph to model both the static physical structure and dynamic history movement information. In addition, to reduce kinematically infeasible predictions for the end-effector pose, we introduce a differentiable kinematics module that provides kinematics-aware references to the policy network. The overview of our proposed Kinematics enhanced Spatial-TemporAl gRaph Diffuser (KStar Diffuser) is shown in Figure 2." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17396" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.09, + 0.907, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.096, + 0.214, + 0.907, + 0.378 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.39, + 0.907, + 0.462 + ], + "angle": 0, + "content": "Figure 2. Overview of KStar Diffuser. The top part presents the spatial-temporal graph which is constructed according to the robot architecture. The bottom part shows our backbone and the proposed kinematics regularizer. For the backbone, it extracts the multimodal information which consists of multiview RGB-D observations and language instruction, and then generates bimanual 6D end-effector poses. The kinematics regularizer enhances pose learning by incorporating joint-level predictions, which are mapped to reference end-effector poses through differentiable forward kinematics (FK)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.478, + 0.206, + 0.492 + ], + "angle": 0, + "content": "3.3.2. Backbone" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.497, + 0.483, + 0.588 + ], + "angle": 0, + "content": "Given the language instruction \\(l\\) and multiview RGB-D observation images \\(o\\), we first adopt the Transformer-based encoders to extract their features \\(\\mathbf{E}_{\\mathrm{I}}\\) and \\(\\mathbf{E}_{\\mathrm{T}}\\), respectively. Then the features are fused with the \\(l_{\\mathrm{FiLM}}\\)-layers FiLM block to obtain the hidden states \\(\\mathbf{H}_{\\mathrm{B}}\\). Each layer is combined with an up-sampling 2D convolution layer:" + }, + { + "type": "equation", + "bbox": [ + 0.211, + 0.595, + 0.482, + 0.613 + ], + "angle": 0, + "content": "\\[\n\\mathbf {H} _ {\\mathrm {B}} ^ {i} = \\operatorname {C o n v 2 D} \\left(\\mathbf {H} _ {\\mathrm {B}} ^ {i}\\right), \\tag {7}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.197, + 0.617, + 0.482, + 0.636 + ], + "angle": 0, + "content": "\\[\n\\mathbf {H} _ {\\mathrm {B}} ^ {i + 1} = \\operatorname {F i L M} ^ {i} \\left(\\mathbf {H} _ {\\mathrm {B}} ^ {i}, \\mathbf {E} _ {\\mathrm {T}}\\right), \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.639, + 0.453, + 0.654 + ], + "angle": 0, + "content": "where \\(i\\in [0,l_{\\mathrm{FiLM}})\\) and the \\(\\mathbf{H}_{\\mathrm{B}}^{0}\\) is initialized with \\(\\mathbf{E}_{\\mathrm{I}}\\)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.655, + 0.484, + 0.775 + ], + "angle": 0, + "content": "Our backbone \\(\\pi_{\\phi}\\) uses the last hidden states \\(\\mathbf{H}_{\\mathrm{B}}\\) as the condition to guide a diffusion head to denoise and generate the bimanual end-effector pose. It is noted that we attach the \\(n\\) historical observation images to provide more information for capturing the motion tendency. Following Chi et al. [9], we let the policy predict next \\(m\\) actions during training to alleviate multimodal problems. Here, we set both \\(n\\) and \\(m\\) to 2. The action prediction is as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.188, + 0.78, + 0.482, + 0.8 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {a} _ {k - 1} ^ {\\mathrm {E E}} \\sim \\pi_ {\\phi} \\left(\\boldsymbol {a} _ {k - 1} ^ {\\mathrm {E E}} \\mid \\boldsymbol {a} _ {k} ^ {\\mathrm {E E}}, \\mathbf {H} _ {\\mathrm {B}}\\right), \\tag {9}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.104, + 0.805, + 0.483, + 0.846 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\pi_ {\\phi} \\left(\\boldsymbol {a} _ {k - 1} ^ {\\mathrm {E E}} \\mid \\boldsymbol {a} _ {k} ^ {\\mathrm {E E}}, \\mathbf {H} _ {\\mathrm {B}}\\right) \\tag {10} \\\\ = \\mathcal {N} (\\boldsymbol {a} _ {k - 1} ^ {\\mathrm {E E}}; \\sqrt {\\bar {\\alpha} _ {k - 1}} \\pi_ {\\phi} (\\boldsymbol {a} _ {k} ^ {\\mathrm {E E}}, \\mathbf {H} _ {\\mathrm {B}}); (1 - \\bar {\\alpha} _ {k - 1}) I). \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.849, + 0.262, + 0.864 + ], + "angle": 0, + "content": "The learning objective is:" + }, + { + "type": "equation", + "bbox": [ + 0.163, + 0.87, + 0.483, + 0.906 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {E E}} = \\mathbb {E} _ {q} \\left[ \\log \\frac {\\pi_ {\\phi} \\left(\\boldsymbol {a} _ {0 : K} ^ {\\mathrm {E E}} \\mid \\mathbf {H} _ {\\mathrm {B}}\\right)}{q \\left(\\boldsymbol {a} _ {1 : K} ^ {\\mathrm {E E}} \\mid \\boldsymbol {a} _ {0} ^ {\\mathrm {E E}}, \\mathbf {H} _ {\\mathrm {B}}\\right)} \\right]. \\tag {11}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.478, + 0.779, + 0.494 + ], + "angle": 0, + "content": "3.3.3. Spatial-Temporal Robot Graph" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.498, + 0.907, + 0.604 + ], + "angle": 0, + "content": "The physical architecture impacts the motion of the whole robot, determining whether it can complete the task. Meanwhile, the historical spatial information is also important to future movements. Thus, we propose a spatial-temporal graph method to model the robot architecture at each step and the robot motion at continuous timesteps, representing the static spatial information and dynamic motion features." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.613, + 0.907, + 0.779 + ], + "angle": 0, + "content": "Spatial Structure Graph Construction. To represent the robot structure, we first parse the Unified Robot Description Format (URDF) file which is usually used to describe the static physical structure of robot such as joint types, joint limits, and link lengths. Then, we define a dual-arm system as an undirected graph \\( G_{\\mathrm{s}} = \\langle V,E\\rangle \\) based on the joint and link configuration. Here, \\( V = \\{\\pmb{v}_{\\mathrm{j}_1},\\pmb{v}_{\\mathrm{j}_2},\\dots,\\pmb{v}_{\\mathrm{j}_m}\\} \\) and \\( E = \\{e_{ij}|(\\mathrm{joint}_i,\\mathrm{joint}_j)\\in \\mathrm{links}\\} \\) represents the node set of joints and edge set of links respectively, where \\( m \\) denotes the number of joint. We use \\( f_{i}\\in \\mathbb{R}^{D} \\) to denote the value of \\( i \\)-th node feature. It consists of three attributes as follows:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.781, + 0.906, + 0.856 + ], + "angle": 0, + "content": "- Joint Coordinate: we use a vector in Cartesian coordinate system, \\( \\pmb{f}_i^{\\mathrm{JC}} = [\\pmb{x}_i, \\pmb{y}_i, \\pmb{z}_i] \\in \\mathbb{R}^3 \\), to denote the absolute coordinates of the \\( i \\)-th joint. The vector is normalized according to the workspace boundary for stable convergence of model training." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.856, + 0.905, + 0.902 + ], + "angle": 0, + "content": "- Joint Distance: To measure the spatial relationship between the node \\( \\mathbf{v}_i \\) and the other nodes \\( \\mathbf{v}_j \\), we compute the Euclidean Distance between \\( \\mathbf{v}_i \\) and \\( \\mathbf{v}_j \\):" + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.781, + 0.906, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17397" + } + ], + [ + { + "type": "equation", + "bbox": [ + 0.225, + 0.089, + 0.484, + 0.109 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {f} _ {i} ^ {\\mathrm {J D}} = \\left| \\left| \\boldsymbol {v} _ {i} - \\boldsymbol {v} _ {j} \\right| \\right| _ {2}, \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.104, + 0.115, + 0.476, + 0.131 + ], + "angle": 0, + "content": "where \\(\\pmb{f}_i^{\\mathrm{JD}}\\in \\mathbb{R}^m\\) and \\(||\\cdot ||_2\\) means the Euclidean norm." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.132, + 0.484, + 0.192 + ], + "angle": 0, + "content": "- Body Label: To discriminate the source of node \\( \\mathbf{v}_i \\), we use a one-hot vector \\( \\mathbf{f}_i^{\\mathrm{BL}} = [0,1] \\) as one of its features. It can also help the policy capture the motion modes of different robot arms such as the symmetry." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.193, + 0.484, + 0.24 + ], + "angle": 0, + "content": "We concatenate \\( \\pmb{f}_i^{\\mathrm{JC}} \\), \\( \\pmb{f}_i^{\\mathrm{JD}} \\), and \\( \\pmb{f}_i^{\\mathrm{BL}} \\) to form the completed feature \\( \\pmb{f}_i \\in \\mathbb{R}^{D_f} \\), where \\( D_{f} \\) denotes the dimension of node features." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.248, + 0.485, + 0.43 + ], + "angle": 0, + "content": "Spatial-Temporal Graph Learning. Given the same instruction and observation, different historical robot poses result in different predictions. Thus, we build the spatial structure graph with the temporal motion information. Specifically, we build a spatial-temporal graph \\( G_{\\mathrm{ST}} = \\langle V_{\\mathrm{ST}}, E_{\\mathrm{ST}} \\rangle \\) by combining \\( \\{G_{\\mathrm{S}}^{i} = \\langle V_{\\mathrm{S}}^{i}, E_{\\mathrm{S}}^{i} \\rangle \\}_{i=0}^{T-1} \\) from historical timesteps, where \\( T \\) denotes the number of history steps. In \\( G_{\\mathrm{ST}} \\), the node set \\( V_{\\mathrm{ST}} \\) contains \\( V_{\\mathrm{S}}^{i} \\) from historical static spatial graph. Additionally, we add edges \\( E' \\) linking the same joint node \\( \\text{joint}_i^t \\) at different timesteps, in order to establish the correlation of joint motions at continuous times. It can be formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.12, + 0.439, + 0.483, + 0.459 + ], + "angle": 0, + "content": "\\[\nV _ {\\mathrm {S T}} = \\cup \\left\\{V _ {\\mathrm {S}} ^ {i} \\right\\} _ {i = 0} ^ {T - 1}, E _ {\\mathrm {S T}} = \\cup \\left\\{E _ {\\mathrm {S}} ^ {i} \\right\\} _ {i = 0} ^ {T - 1} \\cup E ^ {\\prime}, \\tag {13}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.15, + 0.469, + 0.482, + 0.488 + ], + "angle": 0, + "content": "\\[\nE ^ {\\prime} = \\left\\{\\boldsymbol {e} _ {i} ^ {t, t ^ {\\prime}} \\mid t, t ^ {\\prime} \\in \\{0, 1, \\dots , T \\}, t \\neq t ^ {\\prime} \\right\\}. \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.494, + 0.484, + 0.63 + ], + "angle": 0, + "content": "In this way, we obtain the whole spatial-temporal graph. A Graph Convolutional Network (GCN) is then adopted to propagate and aggregate node features across the graph. The GCN layer updates each node feature \\( f_{i} \\) by aggregating the features of its neighboring nodes, thereby capturing the relational and structural information of the robotic arms. We use the node feature \\( \\mathbf{H}_{\\mathrm{G}} \\) of the last encoder layer as the representation of the robot structure to condition the denoising process." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.637, + 0.303, + 0.652 + ], + "angle": 0, + "content": "3.3.4. Kinematics Regularizer" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.656, + 0.484, + 0.852 + ], + "angle": 0, + "content": "To control the end-effector effectively, the generated pose trajectory must be processed by an Inverse Kinematics (IK) solver, which calculates the joint configurations to achieve the specified poses. However, because the predicted trajectory is generated without considering the robot kinematic constraints, it often falls outside the IK solver's feasible range, resulting in high failure rates during execution. To address this limitation, we propose a kinematics regularizer into the end-effector pose learning objective. This regularizer aligns the predicted poses with the robot kinematic constraints, ensuring that the generated trajectory remains within the solvable space of the IK solver, thus enhancing then reliability of trajectory execution." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.856, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Differentiable Kinematics. Given a joint configuration \\(\\Theta = [\\theta_1, \\theta_2, \\dots, \\theta_n]\\), the corresponding end-effector pose \\(\\mathbf{T} \\in SE(3)\\) can be computed using forward kinematics," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.227 + ], + "angle": 0, + "content": "represented as a mapping \\( f(\\Theta): \\mathbb{R}^n \\to SE(3) \\). This mapping from joint space to end-effector space is differentiable, namely Differentiable Forward Kinematics (DFK), enabling the use of gradients to optimize our control policy. Leveraging DFK, our policy learns to predict the next joint configuration \\( \\hat{a}^{\\mathrm{joint}} \\), from which we compute an intermediate end-effector pose \\( \\mathbf{H}_{\\mathrm{R}} \\). By using \\( \\mathbf{H}_{\\mathrm{R}} \\) as a reference, we guide a denoising process to generate a precise and executable end-effector pose." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.228, + 0.905, + 0.273 + ], + "angle": 0, + "content": "Specifically, we combine structure features \\(\\mathbf{H}_{\\mathrm{G}}\\) with the last hidden state \\(\\mathbf{H}_{\\mathrm{B}}\\), projecting to the joint space and using the DFK to obtain reference as \\(\\mathbf{H}_{\\mathrm{R}}\\) follows:" + }, + { + "type": "equation", + "bbox": [ + 0.532, + 0.284, + 0.905, + 0.302 + ], + "angle": 0, + "content": "\\[\n\\hat {\\boldsymbol {a}} ^ {\\text {j o i n t}} = \\operatorname {P r o j} \\left(\\left[ \\mathbf {H} _ {\\mathrm {B}}, \\mathbf {H} _ {\\mathrm {G}} \\right]\\right), \\quad \\mathbf {H} _ {\\mathrm {R}} = \\operatorname {D F K} \\left(\\hat {\\boldsymbol {a}} ^ {\\text {j o i n t}}\\right). \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.314, + 0.905, + 0.345 + ], + "angle": 0, + "content": "To ensure the consistency between predicted and actual joint angles, we minimize the joint loss:" + }, + { + "type": "equation", + "bbox": [ + 0.548, + 0.355, + 0.905, + 0.381 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {j o i n t}} = \\mathbb {E} _ {\\boldsymbol {a} ^ {\\text {j o i n t}} \\sim q (\\boldsymbol {a} _ {0})} \\left[ \\left| \\left| \\boldsymbol {a} ^ {\\text {j o i n t}} - \\hat {\\boldsymbol {a}} ^ {\\text {j o i n t}} \\right| \\right| ^ {2} \\right]. \\tag {16}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.393, + 0.905, + 0.484 + ], + "angle": 0, + "content": "Conditioning Diffusion Process on Kinematics. To enforce kinematic consistency, we condition the diffusion process on the joint angle representation \\(\\mathbf{H}_{\\mathrm{R}}\\), an auxiliary input encoding kinematic constraints. This allows the predicted pose trajectory to stay within feasible space. Given the diffusion steps from Eq.(9) and Eq.(10), we have:" + }, + { + "type": "equation", + "bbox": [ + 0.595, + 0.494, + 0.905, + 0.513 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {a} _ {k - 1} ^ {\\mathrm {E E}} \\sim \\pi_ {\\phi} \\left(\\boldsymbol {a} _ {k} ^ {\\mathrm {E E}} \\mid \\boldsymbol {a} _ {k - 1} ^ {\\mathrm {E E}}, \\mathbf {H} _ {\\mathrm {B}}, \\mathbf {H} _ {\\mathrm {R}}\\right), \\tag {17}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.534, + 0.524, + 0.72, + 0.543 + ], + "angle": 0, + "content": "\\[\n\\pi_ {\\phi} \\left(\\boldsymbol {a} _ {k - 1} ^ {\\mathrm {E E}} \\mid \\boldsymbol {a} _ {k} ^ {\\mathrm {E E}}, \\mathbf {H} _ {\\mathrm {B}}, \\mathbf {H} _ {\\mathrm {R}}\\right) =\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.534, + 0.546, + 0.905, + 0.58 + ], + "angle": 0, + "content": "\\[\n\\mathcal {N} \\left(\\boldsymbol {a} _ {k - 1} ^ {\\mathrm {E E}}; \\sqrt {\\bar {\\alpha} _ {k - 1}} \\pi_ {\\phi} \\left(\\boldsymbol {a} _ {k} ^ {\\mathrm {E E}}, \\mathbf {H} _ {\\mathrm {B}}, \\mathbf {H} _ {\\mathrm {R}}\\right); (1 - \\bar {\\alpha} _ {k - 1}) \\boldsymbol {I}\\right). \\tag {18}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.587, + 0.905, + 0.662 + ], + "angle": 0, + "content": "Incorporating DFK into the diffusion process allows gradients from the pose loss to propagate back through the kinematic function, ensuring that each denoising step maintains compliance with joint constraints, thereby optimizing the end-effector's control accuracy and robustness." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.672, + 0.727, + 0.687 + ], + "angle": 0, + "content": "3.4. Training and Inference" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.695, + 0.905, + 0.755 + ], + "angle": 0, + "content": "Training. We use the conditional action generation mode to train KStar Diffuser, which is formulated as a conditional denoising diffusion. The loss function is defined as the Mean Square Error (MSE) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.614, + 0.763, + 0.905, + 0.781 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {a} _ {k} = \\sqrt {\\bar {\\alpha} _ {k}} \\boldsymbol {a} _ {0} + \\sqrt {1 - \\bar {\\alpha} _ {k}} \\epsilon , \\tag {19}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.528, + 0.79, + 0.905, + 0.811 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {E E}} = \\mathbb {E} _ {\\boldsymbol {a} _ {0} ^ {\\mathrm {E E}} \\sim q (\\boldsymbol {a} _ {0}), k} \\left[ \\left| \\left| \\boldsymbol {a} _ {0} ^ {\\mathrm {E E}} - \\pi_ {\\phi} (\\boldsymbol {a} _ {k} ^ {\\mathrm {E E}}, k, C) \\right| \\right| ^ {2} \\right], \\tag {20}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.529, + 0.818, + 0.905, + 0.84 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {j o i n t}} = \\mathbb {E} _ {\\boldsymbol {a} _ {0} ^ {\\text {j o i n t}} \\sim q (\\boldsymbol {a} _ {0})} [ \\| \\boldsymbol {a} _ {0} ^ {\\text {j o i n t}} - \\pi_ {\\phi} (\\mathbf {H} _ {\\mathrm {B}}, \\mathbf {H} _ {\\mathrm {G}}) \\| ^ {2} ], \\tag {21}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.614, + 0.847, + 0.905, + 0.863 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\lambda \\mathcal {L} _ {\\mathrm {E E}} + (1 - \\lambda) \\mathcal {L} _ {\\text {j o i n t}}, \\tag {22}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.87, + 0.907, + 0.914 + ], + "angle": 0, + "content": "where \\( \\mathbf{a}_k \\) is obtained by the forward diffusion process and \\( C \\) is the combination of \\( \\mathbf{H}_{\\mathrm{B}} \\), \\( \\mathbf{H}_{\\mathrm{G}} \\), \\( \\mathbf{H}_{\\mathrm{R}} \\). \\( \\lambda \\) is the trade-off coefficient." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17398" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.137 + ], + "angle": 0, + "content": "Inference. Sampling from a Gaussian noise \\(\\epsilon_{k}\\), the policy \\(\\pi_{\\phi}\\) performs \\(K\\) iterations to gradually denoise a random noise \\(\\epsilon_{k}\\) into the noise-free action \\(\\pmb{a}_{0}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.107, + 0.143, + 0.483, + 0.163 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {a} _ {k - 1} ^ {\\mathrm {E E}} = \\sqrt {\\overline {{\\alpha}} _ {k - 1}} \\pi_ {\\phi} (\\boldsymbol {a} _ {k} ^ {\\mathrm {E E}}, k, C) + \\sqrt {1 - \\overline {{\\alpha}} _ {k - 1}} \\epsilon_ {k}. \\tag {23}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.168, + 0.483, + 0.199 + ], + "angle": 0, + "content": "We use the predicted \\( a_0^{\\mathrm{EE}} \\) as the final action to control robot execution." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.211, + 0.216, + 0.228 + ], + "angle": 0, + "content": "4. Experiment" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.236, + 0.373, + 0.252 + ], + "angle": 0, + "content": "4.1. Dataset and Evaluation Settings" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.258, + 0.484, + 0.408 + ], + "angle": 0, + "content": "Dataset. Bimanual manipulation tasks demand high levels of coordination, synchronization, and symmetry awareness between the two robotic arms, making them inherently more challenging than single-arm tasks. To assess the capabilities of KStar Diffuser in these areas, we conducted comprehensive experiments using the RLBench2 benchmark [17], an extended version of RLBench tailored for bimanual manipulation and comprising tasks closely resembling real-world scenarios. Please refer to Appendix for more details about RLBench2 and real-world tasks." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.412, + 0.484, + 0.607 + ], + "angle": 0, + "content": "Evaluation Settings. To evaluate the policy performance, we employ success rate as the primary metric. Although the policy generates multiple sequential actions during execution, we primarily focus on the final goal achievement rather than intermediate steps. Each task is associated with a specific success criterion defined by its target state. To comprehensively assess the policy's capability, we conduct experiments with varying numbers of demonstrations (20 and 100) during training. Figure 3 illustrates our experimental setup, including both the simulation environment and the Cobot Agilex ALOHA robot. Detailed descriptions of simulation tasks and real-world experimental settings are provided in the Appendix." + }, + { + "type": "image", + "bbox": [ + 0.114, + 0.608, + 0.484, + 0.753 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.764, + 0.483, + 0.793 + ], + "angle": 0, + "content": "Figure 3. The left: the simulation environment of pick_laptop task. The right: the ALOHA device used in the real-world tasks." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.801, + 0.199, + 0.815 + ], + "angle": 0, + "content": "4.2. Baselines" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.822, + 0.483, + 0.853 + ], + "angle": 0, + "content": "We systematically evaluate KStar Diffuser against state-of-the-art methods in two major categories:" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Transformer-based methods: (1) Action Chunking with Transformers (ACT) [58] employs a Conditional VAE architecture, consisting of an encoder-decoder framework" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.289 + ], + "angle": 0, + "content": "for joint angle sequence prediction; (2) Robotic View Transformer Leader Following (RVT-LF) [17] leverages RVT [15] as its backbone, incorporating a multi-view transformer for cross-view information aggregation and image re-rendering, coupled with a leader following mechanism for action prediction; (3) Perceiver-Actor Leader Following (PerAct-LF) [17] adopts the leader following paradigm based on PerAct [44], utilizing a perceiver Transformer to encode both instructions and voxel observations for optimal voxel action generation; (4) PerAct2 [17] enhances PerAct by implementing a unified feature space for dual-arm actions and employing combined self-attention for synchronized bimanual action prediction." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.294, + 0.909, + 0.431 + ], + "angle": 0, + "content": "Diffusion-based methods: (1) Joint-based Diffusion Policy (DP-J) [9] adopts a diffusion model to robotic manipulation within an imitation learning framework, focusing on joint angle prediction; (2) EndEffector-based Diffusion Policy (DP-EE) is a DP variant which we reimplement Diffusion Policy to predict end-effector poses instead of joint angles, offering an alternative control paradigm; (3) 3D Diffusion Policy (DP3) [56] enhances 3D perception by incorporating point clouds for joint angle prediction." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.44, + 0.866, + 0.456 + ], + "angle": 0, + "content": "4.3. Comparison Results with SOTA Methods" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.463, + 0.906, + 0.537 + ], + "angle": 0, + "content": "Experimental Results on RLBench2. As shown in Table 1, the KStar Diffuser significantly outperforms other state-of-the-art baselines, achieving more than \\(20\\%\\) higher overall performance with both 20 and 100 training demonstrations. We found that:" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.539, + 0.907, + 0.765 + ], + "angle": 0, + "content": "(1) Similar to learning a single-arm policy, the process of learning a bimanual policy can adapt quickly and achieve a high success rate, given a relatively consistent distribution of task trajectories. In push_box, where the objective is for both arms to push a box toward a specified target along fixed trajectories, our KStar Diffuser and other baseline models perform well. However, as task complexity increases, success rates decrease. For example, in the liftball task, both arms must lift a large ball simultaneously to complete the task. Any asynchrony in movement can cause instability, leading to the ball slipping and ultimately resulting in task failure. Our KStar Diffuser achieves its robust performance on such bimanual tasks by explicitly modeling the spatial and motion relationships between the two arms, surpassing other methods more than \\(6\\%\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.766, + 0.909, + 0.903 + ], + "angle": 0, + "content": "(2) Distinct from single-arm systems, bimanual robotic systems possess the capability for collaborative manipulation. Methods which are directly adapted from single-arm to bimanual manipulation exhibit high failure rates in tasks, e.g., pick_laptop, as they lack consideration of the spatial and motion relationships between the arms. Specifically, as shown in Figure 4, this task involves picking up a notebook lying flat on a cabinet surface. Given that the notebook rests fully against the tabletop, direct grasping by" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17399" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.908, + 0.12 + ], + "angle": 0, + "content": "Table 1. The experimental result on simulated tasks. We train the policy with the setting of different training demonstrations, i.e. [20, 100], to test its capability comprehensively. The best results are in bold. Each result is reported with three seeds on average." + }, + { + "type": "table", + "bbox": [ + 0.137, + 0.128, + 0.865, + 0.44 + ], + "angle": 0, + "content": "
MethodBackbonePush BoxLift BallHandover Item (easy)Pick LaptopSweep DustpanOverall
20 demos
ACT [58]Transformer26.0 ±4.617.3 ±5.80.0 ±0.00.0 ±0.00.0 ±0.08.7 ±1.6
RVT-LF [17]Transformer44.0 ±1.012.3 ±6.70.0 ±0.02.0 ±1.70.0 ±0.011.7 ±1.5
PerAct-LF [17]Transformer63.7 ±3.142.3 ±8.13.3 ±0.67.0 ±2.74.7 ±0.624.2 ±1.6
PerAct2 [17]Transformer59.7 ±7.637.7 ±4.78.0 ±1.017.3 ±0.61.7 ±1.524.9 ±2.1
DP-J [9]Diffusion30.0 ±3.619.7 ±2.10.0 ±0.00.0 ±0.00.0 ±4.010.0 ±1.0
DP-EE [9]Diffusion34.7 ±2.534.3 ±7.40.7 ±0.61.3 ±0.614.0 ±1.717.0 ±1.6
DP3 [56]Diffusion25.3 ±4.034.3 ±7.60.0 ±0.02.7 ±2.90.0 ±0.012.5 ±1.3
KStar Diffuser (Ours)Diffusion79.3 ±3.587.0 ±2.723.7 ±0.617.0 ±2.083.0 ±4.458.0 ±1.4
100 demos
ACT [58]Transformer48.7 ±6.840.7 ±2.10.0 ±0.00.0 ±0.00.0 ±0.017.9 ±1.0
RVT-LF [17]Transformer76.3 ±1.528.7 ±1.00.0 ±0.01.3 ±1.22.3 ±3.221.7 ±0.7
PerAct-LF [17]Transformer66.3 ±3.168.3 ±6.17.3 ±2.114.3 ±3.88.0 ±2.732.9 ±2.5
PerAct2 [17]Transformer77.0 ±3.047.7 ±3.210.3 ±0.634.3 ±4.06.7 ±2.133.9 ±1.0
DP-J [9]Diffusion62.7 ±7.643.3 ±2.10.0 ±0.01.3 ±0.60.0 ±4.021.5 ±1.6
DP-EE [9]Diffusion61.0 ±2.059.7 ±3.52.0 ±1.010.3 ±3.169.7 ±2.540.5 ±1.2
DP3 [56]Diffusion56.0 ±3.664.0 ±2.70.0 ±0.06.3 ±3.11.7 ±2.125.6 ±1.4
KStar Diffuser (Ours)Diffusion83.0 ±1.798.7 ±1.527.0 ±1.743.7 ±4.589.0 ±5.268.2 ±2.1
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.462, + 0.484, + 0.569 + ], + "angle": 0, + "content": "the robotic arm is not possible. Instead, an effective strategy is to control one arm to push the notebook outward from the cabinet by a short distance, allowing the other arm to pick it up. KStar Diffuser achieves a success rate approximately \\(9\\%\\) higher than other methods, demonstrating its ability to capture the coordinated motion patterns required for collaborative object manipulation between two arms." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.582, + 0.484, + 0.643 + ], + "angle": 0, + "content": "Real-world Experimental Results. To comprehensively evaluate the policy's effectiveness, we build 2 tasks in real-world based on the simulation benchmark. The performances on real-world tasks are illustrated in Table 2." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.644, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Similar to the simulation results, we observe that policies which do not consider about bimanual scenarios, i.e., ACT, DP, and DP3, demonstrate limited capability across all bimanual tasks, achieving around \\(20\\%\\) success ratio on average. Although PerAct2 is designed for bimanual tasks via mapping bimanual actions into a shared learning space, it fails to capture the spatial structure of the bimanual system, leading to ineffective arm coordination during execution. Furthermore, we also found PerAct2 faces significant inverse kinematics issues with its predicted end-effector poses, including joint configuration conflicts and unreachable positions which are shown in Figure 4. It is likely due to PerAct2's limited ability to capture the complex spatial constraints and kinematic relationships within the bimanual robotic system. In contrast, KStar Diffuser achieves superior bimanual coordination, surpassing other methods by over \\(10\\%\\), as it successfully captures the motion patterns" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.462, + 0.907, + 0.479 + ], + "angle": 0, + "content": "between dual arms and predicts feasible end-effector poses." + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.493, + 0.908, + 0.522 + ], + "angle": 0, + "content": "Table 2. The result of real-world tasks. We train all policies with 100 demonstrations and test 15 times. The best result are in bold." + }, + { + "type": "table", + "bbox": [ + 0.53, + 0.529, + 0.895, + 0.624 + ], + "angle": 0, + "content": "
MethodsLift PlateHandoverOverall
ACT [58]37.8 ±8.30.0 ±0.018.9 ±11.2
DP [9]42.1 ±6.50.0 ±0.021.1 ±12.5
DP3 [56]44.0 ±8.30.0 ±0.022.0 ±13.4
PerAct2 [17]51.1 ±6.38.8 ±3.029.9 ±10.2
KStar Diffuser (Ours)66.7 ±5.319.7 ±5.343.1 ±17.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.571, + 0.648, + 0.848, + 0.662 + ], + "angle": 0, + "content": "Table 3. Ablation study of model components." + }, + { + "type": "table", + "bbox": [ + 0.534, + 0.673, + 0.891, + 0.743 + ], + "angle": 0, + "content": "
ST GraphKRHandover-SHandover-ROverall
✓✓27.0 ±1.719.7 ±5.323.4 ±5.2
✓×18.3 ±1.515.3 ±3.316.8 ±2.1
××16.3 ±3.113.3 ±9.414.8 ±2.1
" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.757, + 0.681, + 0.772 + ], + "angle": 0, + "content": "4.4. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Effects of Model Components. To systematically evaluate the contribution of each component in KStar Diffuser, we conduct ablation experiments on the handover-item task in both simulated and real-world environments. We design a progressive ablation process by first removing the Differential Forward Kinematics module while retaining the Spatial-Temporal Graph (ST Graph), and then completely disabling both ST Graph and Kinematics Regularizer (KR)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "17400" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.102, + 0.09, + 0.898, + 0.409 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.419, + 0.908, + 0.448 + ], + "angle": 0, + "content": "Figure 4. The visualization of bimanual manipulation on simulated RLBench2 and real-world tasks. The blue annotations represent the motion of the robot's left arm, while the green annotations indicate the motion of the right arm." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.463, + 0.483, + 0.766 + ], + "angle": 0, + "content": "The experimental results in Table 3 demonstrate the crucial role of each component. The removal of KR leads to a significant decrease in success rate, particularly pronounced in real-world scenarios. This performance degradation can be attributed to the fundamental differences between simulated and real environments. While simulated environments maintain consistent and noise-free inputs, real-world scenarios introduce various perturbations, e.g., sensor noise and light reflection, making the policy more susceptible to kinematic constraint violations without the regularization effect of KR. Further ablation by removing both ST Graph and KR results in a substantial performance drop across all experimental settings. This observation illustrates two key aspects: First, the ST Graph effectively captures the spatial-temporal dependencies among joints, which is essential for coordinating the relative positioning and interactions between robotic arms. Second, the graph structure's explicit encoding of the robotic physical architecture enhances the policy's robustness against unexpected perturbations by maintaining spatial and temporal coherence." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.773, + 0.283, + 0.79 + ], + "angle": 0, + "content": "4.5. Qualitative Analysis" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.483, + 0.856 + ], + "angle": 0, + "content": "We further present qualitative analysis in Figure 4. We compare the performance of KStar Diffuser with DP3 and PerAct2 in executing bimanual manipulation tasks within both simulated and real-world environments, respectively." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "In the simulation task, since the laptop rests flat on the cabinet, direct lifting is not feasible. One robotic arm initiates a forward push, creating space, while the other arm" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.463, + 0.907, + 0.569 + ], + "angle": 0, + "content": "concurrently grasps and elevates the laptop. KStar Diffuser effectively models this dual-arm coordination, generating a precise trajectory of synchronized actions. Conversely, DP3, adapted from a single-arm policy to a dual-arm configuration, fails to achieve effective coordination. Concretely, after executing the push motion, the right arm does not halt, obstructing the left arm's lifting process." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.57, + 0.909, + 0.691 + ], + "angle": 0, + "content": "In the real-world task, KStar Diffuser generates an executable item transfer trajectory between the left and right arms, with no collisions throughout the task, reflecting its strong environmental adaptability and collision avoidance capabilities. Conversely, PerAct2 encounters collisions during the handover process (marked in red), indicating less effective handling of dynamic real-world variables and a lack of kinematic awareness on the robot movements." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.709, + 0.634, + 0.724 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.909, + 0.901 + ], + "angle": 0, + "content": "In this paper, we have proposed a novel Kinematics enhanced Spatial-TemporAl gRaph Diffuser (KStar Diffuser) that explicitly incorporates both robot structures and kinematics into the bimanual motion generation process. It consists of a spatial-temporal robot graph that explicitly models the robot physical configuration to guide the generative action denoising procedure, and a kinematics regularizer that augments the NBP learning objective by introducing joint-space supervision. Extensive experiments demonstrate that KStar Diffuser outperforms the baselines by a large margin on both simulation and real-world tasks." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "17401" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.279, + 0.108 + ], + "angle": 0, + "content": "6. Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.486, + 0.372 + ], + "angle": 0, + "content": "We would like to thank all co-authors for their efforts and the reviewers for their constructive comments. This work is supported by Shenzhen Science and Technology Program (KQTD20240729102207002), National Natural Science Foundation of China (Grant No. 62406092), National Natural Science Foundation of China (Grant No. U24B20175), Shenzhen Science and Technology Program (Grant No. KJZD20240903100017022), Guangdong Basic and Applied Basic Research Foundation (2025A1515010169), and Research on Efficient Exploration and Self-Evolution of APP Agents & Embodied Intelligent Cerebellum Control Model and Collaborative Feedback Training Project (Grant No. TC20240403047). Additionally, it is also supported by National Natural Science Foundation of China (Grant No. 62306090, No. 62236003), Natural Science Foundation of Guangdong Province of China (Grant No. 2024A1515010147)." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.385, + 0.188, + 0.401 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.411, + 0.482, + 0.493 + ], + "angle": 0, + "content": "[1] Lucia Angelini, Manuela Uliano, Angela Mazzeo, Mattia Penzotti, and Marco Controzzi. Self-collision avoidance in bimanual teleoperation using collisionik: algorithm revision and usability experiment. In 2022 IEEE-RAS 21st International Conference on Humanoid Robots (Humanoids), pages 112-118, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.495, + 0.483, + 0.55 + ], + "angle": 0, + "content": "[2] Andreas Aristidou, Joan Lasenby, Yiorgos Chrysanthou, and Ariel Shamir. Inverse kinematics techniques in computer graphics: A survey. In Computer graphics forum, pages 35-58. Wiley Online Library, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.552, + 0.483, + 0.633 + ], + "angle": 0, + "content": "[3] Weibang Bai, Ningshan Zhang, Baoru Huang, Ziwei Wang, Francesco Cursi, Ya-Yen Tsai, Bo Xiao, and Eric M Yeatman. Dual-arm coordinated manipulation for object twisting with human intelligence. In 2021 IEEE International Conference on Systems, Man, and Cybernetics (SMC), pages 902-908. IEEE, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.635, + 0.483, + 0.717 + ], + "angle": 0, + "content": "[4] Aaron M Bestick, Samuel A Burden, Giorgia Willits, Nikhil Naikal, S Shankar Sastry, and Ruzena Bajcsy. Personalized kinematics for human-robot collaborative manipulation. In 2015 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 1037–1044. IEEE, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.719, + 0.483, + 0.787 + ], + "angle": 0, + "content": "[5] Botond Bócsi, Duy Nguyen-Tuong, Lehel Csato, Bernhard Schoelkopf, and Jan Peters. Learning inverse kinematics with structured prediction. In 2011 IEEE/RSJ International Conference on Intelligent Robots and Systems, pages 698-703. IEEE, 2011. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.79, + 0.483, + 0.858 + ], + "angle": 0, + "content": "[6] Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Joseph Dabis, Chelsea Finn, Keerthana Gopalakrishnan, Karol Hausman, Alex Herzog, Jasmine Hsu, et al. Rt-1: Robotics transformer for real-world control at scale. arXiv preprint arXiv:2212.06817, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.86, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[7] Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Xi Chen, Krzysztof Choromanski, Tianli Ding, Danny Driess, Avinava Dubey, Chelsea Finn, et al. Rt-2:" + }, + { + "type": "list", + "bbox": [ + 0.1, + 0.411, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "Vision-language-action models transfer web knowledge to robotic control. arXiv preprint arXiv:2307.15818, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.122, + 0.905, + 0.176 + ], + "angle": 0, + "content": "[8] Shizhe Chen, Ricardo Garcia, Cordelia Schmid, and Ivan Laptev. Polarnet: 3d point clouds for language-guided robotic manipulation. arXiv preprint arXiv:2309.15596, 2023.2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.179, + 0.905, + 0.247 + ], + "angle": 0, + "content": "[9] Cheng Chi, Zhenjia Xu, Siyuan Feng, Eric Cousineau, Yilun Du, Benjamin Burchfiel, Russ Tedrake, and Shuran Song. Diffusion policy: Visuomotor policy learning via action diffusion. The International Journal of Robotics Research, page 02783649241273668, 2023. 1, 2, 3, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.25, + 0.905, + 0.317 + ], + "angle": 0, + "content": "[10] Cheng Chi, Zhenjia Xu, Chuer Pan, Eric Cousineau, Benjamin Burchfiel, Siyuan Feng, Russ Tedrake, and Shuran Song. Universal manipulation interface: In-the-wild robot teaching without in-the-wild robots. arXiv preprint arXiv:2402.10329, 2024. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.321, + 0.905, + 0.375 + ], + "angle": 0, + "content": "[11] Akos Csiszar, Jan Eilers, and Alexander Verl. On solving the inverse kinematics problem using neural networks. In 2017 24th International Conference on Mechatronics and Machine Vision in Practice (M2VIP), pages 1-6, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.378, + 0.905, + 0.459 + ], + "angle": 0, + "content": "[12] Aaron D'Souza, Sethu Vijayakumar, and Stefan Schaal. Learning inverse kinematics. In Proceedings 2001 IEEE/RSJ International Conference on Intelligent Robots and Systems. Expanding the Societal Role of Robotics in the Next Millennium (Cat. No. 01CH37180), pages 298-303. IEEE, 2001. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.462, + 0.904, + 0.516 + ], + "angle": 0, + "content": "[13] Zipeng Fu, Tony Z Zhao, and Chelsea Finn. Mobile aloha: Learning bimanual mobile manipulation with low-cost whole-body teleoperation. arXiv preprint arXiv:2401.02117, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.519, + 0.905, + 0.587 + ], + "angle": 0, + "content": "[14] Jianfeng Gao, Xiaoshu Jin, Franziska Krebs, Noémie Jaquier, and Tamim Asfour. Bi-kvil: Keypoints-based visual imitation learning of bimanual manipulation tasks. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 16850-16857. IEEE, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.59, + 0.905, + 0.643 + ], + "angle": 0, + "content": "[15] Ankit Goyal, Jie Xu, Yijie Guo, Valts Blukis, Yu-Wei Chao, and Dieter Fox. Rvt: Robotic view transformer for 3d object manipulation. In Conference on Robot Learning, pages 694–710. PMLR, 2023. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.646, + 0.905, + 0.715 + ], + "angle": 0, + "content": "[16] Reinhard Grassmann, Vincent Modes, and Jessica Burgner-Kahrs. Learning the forward and inverse kinematics of a 6-dof concentric tube continuum robot in se (3). In 2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 5125-5132. IEEE, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.717, + 0.905, + 0.771 + ], + "angle": 0, + "content": "[17] Markus Grotz, Mohit Shridhar, Tamim Asfour, and Dieter Fox. Peract2: A perceiver actor framework for bimanual manipulation tasks. arXiv preprint arXiv:2407.00278, 2024. 1, 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.774, + 0.905, + 0.828 + ], + "angle": 0, + "content": "[18] Abdullah Aamir Hayat, Ratan OM Sadanand, and Subir K Saha. Robot manipulation through inverse kinematics. In Proceedings of the 2015 conference on advances in robotics, pages 1-6, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.831, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[19] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[20] Binghao Huang, Yixuan Wang, Xinyi Yang, Yiyue Luo, and Yunzhu Li. 3d-vitac: Learning fine-grained manipulation" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "17402" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.12 + ], + "angle": 0, + "content": "with visuo-tactile sensing. arXiv preprint arXiv:2410.24091, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.122, + 0.483, + 0.148 + ], + "angle": 0, + "content": "[21] Kenneth H Hunt. Structural kinematics of in-parallel-actuated robot-arms. 1983. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.151, + 0.483, + 0.192 + ], + "angle": 0, + "content": "[22] Michael I Jordan and David E Rumelhart. Forward models: Supervised learning with a distal teacher. In Backpropagation, pages 189-236. Psychology Press, 2013. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.194, + 0.483, + 0.234 + ], + "angle": 0, + "content": "[23] Tsung-Wei Ke, Nikolaos Gkanatsios, and Katerina Fragkiadaki. 3d diffuser actor: Policy diffusion with 3d scene representations. arXiv preprint arXiv:2402.10885, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.236, + 0.483, + 0.304 + ], + "angle": 0, + "content": "[24] Moo Jin Kim, Karl Pertsch, Siddharth Karamcheti, Ted Xiao, Ashwin Balakrishna, Suraj Nair, Rafael Rafailov, Ethan Foster, Grace Lam, Pannag Sanketi, et al. Openvla: An open-source vision-language-action model. arXiv preprint arXiv:2406.09246, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.307, + 0.483, + 0.347 + ], + "angle": 0, + "content": "[25] Thomas N Kipf and Max Welling. Semi-supervised classification with graph convolutional networks. arXiv preprint arXiv:1609.02907, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.35, + 0.483, + 0.377 + ], + "angle": 0, + "content": "[26] CS George Lee. Robot arm kinematics, dynamics, and control. Computer, 15(12):62-80, 1982. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.379, + 0.483, + 0.433 + ], + "angle": 0, + "content": "[27] Maolin Lei, Ting Wang, Chen Yao, Huan Liu, Zhi Wang, and Yongsheng Deng. Real-time kinematics-based self-collision avoidance algorithm for dual-arm robots. Applied Sciences, 10(17):5893, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.435, + 0.483, + 0.517 + ], + "angle": 0, + "content": "[28] Jiefeng Li, Chao Xu, Zhicun Chen, Siyuan Bian, Lixin Yang, and Cewu Lu. Hybrik: A hybrid analytical-neural inverse kinematics solution for 3d human pose and shape estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3383-3393, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.52, + 0.483, + 0.588 + ], + "angle": 0, + "content": "[29] Jiefeng Li, Chao Xu, Zhicun Chen, Siyuan Bian, Lixin Yang, and Cewu Lu. Hybrik: A hybrid analytical-neural inverse kinematics solution for 3d human pose and shape estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3383-3393, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.591, + 0.483, + 0.644 + ], + "angle": 0, + "content": "[30] I Liu, Chun Arthur, Sicheng He, Daniel Seita, and Gaurav Sukhatme. Voxact-b: Voxel-based acting and stabilizing policy for bimanual manipulation. arXiv preprint arXiv:2407.04152, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.647, + 0.483, + 0.716 + ], + "angle": 0, + "content": "[31] Xiao Ma, Sumit Patidar, Iain Haughton, and Stephen James. Hierarchical diffusion policy for kinematics-aware multitask robotic manipulation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18081-18090, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.718, + 0.483, + 0.745 + ], + "angle": 0, + "content": "[32] Matthew T Mason. Mechanics of robotic manipulation. MIT press, 2001. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.747, + 0.483, + 0.814 + ], + "angle": 0, + "content": "[33] Yao Mu, Qinglong Zhang, Mengkang Hu, Wenhai Wang, Mingyu Ding, Jun Jin, Bin Wang, Jifeng Dai, Yu Qiao, and Ping Luo. Embodiedgpt: Vision-language pre-training via embodied chain of thought. Advances in Neural Information Processing Systems, 36, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.817, + 0.483, + 0.871 + ], + "angle": 0, + "content": "[34] Yoshihiko Nakamura, Kiyoshi Nagai, and Tsuneo Yoshikawa. Dynamics and stability in coordination of multiple robotic mechanisms. The International Journal of Robotics Research, 8(2):44-61, 1989. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.874, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[35] Valerio Ortenzi, Naresh Marturi, Michael Mistry, Jeffrey Kuo, and Rustam Stolkin. Vision-based framework" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "to estimate robot configuration and kinematic constraints. IEEE/ASME Transactions on Mechatronics, 23(5):2402-2412, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.177 + ], + "angle": 0, + "content": "[36] Eric Paljug, Xiaoping Yun, and Vijay Kumar. Control of rolling contacts in multi-arm manipulation. IEEE Transactions on Robotics and Automation, 10(4):441-452, 1994. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.905, + 0.233 + ], + "angle": 0, + "content": "[37] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4195-4205, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.235, + 0.905, + 0.304 + ], + "angle": 0, + "content": "[38] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.307, + 0.905, + 0.36 + ], + "angle": 0, + "content": "[39] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.363, + 0.905, + 0.431 + ], + "angle": 0, + "content": "[40] Allen Z Ren, Justin Lidard, Lars L Ankile, Anthony Simeonov, Pulkit Agrawal, Anirudha Majumdar, Benjamin Burchfiel, Hongkai Dai, and Max Simchowitz. Diffusion policy policy optimization. arXiv preprint arXiv:2409.00588, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.434, + 0.907, + 0.503 + ], + "angle": 0, + "content": "[41] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.505, + 0.905, + 0.559 + ], + "angle": 0, + "content": "[42] Nilanjan Sarkar, Xiaoping Yun, and Vijay Kumar. Dynamic control of 3-d rolling contacts in two-arm manipulation. IEEE Transactions on Robotics and Automation, 13(3): 364-376, 1997. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.562, + 0.905, + 0.601 + ], + "angle": 0, + "content": "[43] Youshik Shin and Zeungnam Bien. Collision-free trajectory planning for two robot arms. Robotica, 7(3):205-212, 1989. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.604, + 0.905, + 0.658 + ], + "angle": 0, + "content": "[44] Mohit Shridhar, Lucas Manuelli, and Dieter Fox. Perceiver-actor: A multi-task transformer for robotic manipulation. In Conference on Robot Learning, pages 785–799. PMLR, 2023. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.661, + 0.905, + 0.716 + ], + "angle": 0, + "content": "[45] Christian Smith, Yiannis Karayiannidis, Lazaros Nalpantidis, Xavi Gratal, Peng Qi, Dimos V Dimarogonas, and Danica Kragic. Dual arm manipulation—a survey. Robotics and Autonomous systems, 60(10):1340-1353, 2012. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.718, + 0.905, + 0.786 + ], + "angle": 0, + "content": "[46] Octo Model Team, Dibya Ghosh, Homer Walke, Karl Pertsch, Kevin Black, Oier Mees, Sudeep Dasari, Joel Hejna, Tobias Kreiman, Charles Xu, et al. Octo: An open-source generalist robot policy. arXiv preprint arXiv:2405.12213, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.789, + 0.905, + 0.843 + ], + "angle": 0, + "content": "[47] Vitalis Vosylius, Younggyo Seo, Jafar Uruç, and Stephen James. Render and diffuse: Aligning image and action spaces for diffusion-based behaviour cloning. arXiv preprint arXiv:2405.18196, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.905, + 0.899 + ], + "angle": 0, + "content": "[48] Lirui Wang, Xinlei Chen, Jialiang Zhao, and Kaiming He. Scaling proprioceptive-visual learning with heterogeneous pre-trained transformers. arXiv preprint arXiv:2409.20537, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17403" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.161 + ], + "angle": 0, + "content": "[49] Yixiao Wang, Yifei Zhang, Mingxiao Huo, Ran Tian, Xiang Zhang, Yichen Xie, Chenfeng Xu, Pengliang Ji, Wei Zhan, Mingyu Ding, et al. Sparse diffusion policy: A sparse, reusable, and flexible policy for robot learning. arXiv preprint arXiv:2407.01531, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.164, + 0.482, + 0.232 + ], + "angle": 0, + "content": "[50] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. Advances in Neural Information Processing Systems, 36, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.234, + 0.482, + 0.261 + ], + "angle": 0, + "content": "[51] Chris Welman. Inverse kinematics and geometric constraints for articulated figure manipulation. 1993. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.263, + 0.482, + 0.318 + ], + "angle": 0, + "content": "[52] Kun Wu, Yichen Zhu, Jinming Li, Junjie Wen, Ning Liu, Zhiyuan Xu, Qinru Qiu, and Jian Tang. Discrete policy: Learning disentangled action space for multi-task robotic manipulation. arXiv preprint arXiv:2409.18707, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.32, + 0.482, + 0.389 + ], + "angle": 0, + "content": "[53] Wenke Xia, Dong Wang, Xincheng Pang, Zhigang Wang, Bin Zhao, Di Hu, and Xuelong Li. Kinematic-aware prompting for generalizable articulated object manipulation with llms. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 2073-2080. IEEE, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.391, + 0.482, + 0.459 + ], + "angle": 0, + "content": "[54] Andrea Maria Zanchettin, Paolo Rocco, Luca Bascetta, Ioannis Symeonidis, and Steffen Peldschus. Kinematic analysis and synthesis of the human arm motion during a manipulation task. In 2011 IEEE international conference on robotics and automation, pages 2692-2697. IEEE, 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.461, + 0.482, + 0.515 + ], + "angle": 0, + "content": "[55] Yanjie Ze, Zixuan Chen, Wenhao Wang, Tianyi Chen, Xialin He, Ying Yuan, Xue Bin Peng, and Jiajun Wu. Generalizable humanoid manipulation with improved 3d diffusion policies. arXiv preprint arXiv:2410.10803, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.517, + 0.482, + 0.559 + ], + "angle": 0, + "content": "[56] Yanjie Ze, Gu Zhang, Kangning Zhang, Chenyuan Hu, Muhan Wang, and Huazhe Xu. 3d diffusion policy. arXiv preprint arXiv:2403.03954, 2024. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.561, + 0.482, + 0.627 + ], + "angle": 0, + "content": "[57] Junjie Zhang, Chenjia Bai, Haoran He, Wenke Xia, Zhigang Wang, Bin Zhao, Xiu Li, and Xuelong Li. Sam-e: Leveraging visual foundation model with sequence imitation for embodied manipulation. arXiv preprint arXiv:2405.19586, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.63, + 0.482, + 0.685 + ], + "angle": 0, + "content": "[58] Tony Z Zhao, Vikash Kumar, Sergey Levine, and Chelsea Finn. Learning fine-grained bimanual manipulation with low-cost hardware. arXiv preprint arXiv:2304.13705, 2023. 1, 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.687, + 0.482, + 0.755 + ], + "angle": 0, + "content": "[59] Zangwei Zheng, Xiangyu Peng, Tianji Yang, Chenhui Shen, Shenggui Li, Hongxin Liu, Yukun Zhou, Tianyi Li, and Yang You. Open-sora: Democratizing efficient video production for all, march 2024. URL https://github.com/hpcaitech/Open-Sora, 1(3):4.2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.757, + 0.482, + 0.826 + ], + "angle": 0, + "content": "[60] Minjie Zhu, Yichen Zhu, Jinming Li, Junjie Wen, Zhiyuan Xu, Ning Liu, Ran Cheng, Chaomin Shen, Yaxin Peng, Feifei Feng, et al. Scaling diffusion policy in transformer to 1 billion parameters for robotic manipulation. arXiv preprint arXiv:2409.14411, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.826 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "17404" + } + ] +] \ No newline at end of file diff --git a/2025/Spatial-Temporal Graph Diffusion Policy with Kinematic Modeling for Bimanual Robotic Manipulation/8300c221-08df-4f6a-8aec-775be3434ff5_origin.pdf b/2025/Spatial-Temporal Graph Diffusion Policy with Kinematic Modeling for Bimanual Robotic Manipulation/8300c221-08df-4f6a-8aec-775be3434ff5_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..1f846e583963fbe5e5bd03f5e17b65b9fdb92787 --- /dev/null +++ b/2025/Spatial-Temporal Graph Diffusion Policy with Kinematic Modeling for Bimanual Robotic Manipulation/8300c221-08df-4f6a-8aec-775be3434ff5_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:52fb08d8a5b8dc6d9a099bb76b80bd8af5a9ab7fbecc4d3b68f4a588fe0ab753 +size 1722099 diff --git a/2025/Spatial-Temporal Graph Diffusion Policy with Kinematic Modeling for Bimanual Robotic Manipulation/full.md b/2025/Spatial-Temporal Graph Diffusion Policy with Kinematic Modeling for Bimanual Robotic Manipulation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..5b5929b610dca68a5af503a8c4ca17fd1d771feb --- /dev/null +++ b/2025/Spatial-Temporal Graph Diffusion Policy with Kinematic Modeling for Bimanual Robotic Manipulation/full.md @@ -0,0 +1,389 @@ +# Spatial-Temporal Graph Diffusion Policy with Kinematic Modeling for Bimanual Robotic Manipulation + +Qi Lv $^{1,2,3}$ Hao Li $^{1,3}$ Xiang Deng $^{1\dagger}$ Rui Shao $^{1}$ Yinchuan Li $^{3}$ Jianye Hao $^{3}$ Longxiang Gao $^{4}$ Michael Yu Wang $^{2}$ Liqiang Nie $^{1\dagger}$ + +$^{1}$ School of Computer Science and Technology, Harbin Institute of Technology (Shenzhen) + $^{2}$ School of Engineering, Greate Bay University + $^{3}$ Huawei Noah's Ark Lab + $^{4}$ Qilu University of Technology (Shandong Academy of Sciences) + +lvqi@stu.hit.edu.cn, dengxiang@hit.edu.cn + +# Abstract + +Despite the significant success of imitation learning in robotic manipulation, its application to bimanual tasks remains highly challenging. Existing approaches mainly learn a policy to predict a distant next-best end-effector pose (NBP) and then compute the corresponding joint rotation angles for motion using inverse kinematics. However, they suffer from two important issues: (1) rarely considering the physical robotic structure, which may cause self-collisions or interferences, and (2) overlooking the kinematics constraint, which may result in the predicted poses not conforming to the actual limitations of the robot joints. In this paper, we propose Kinematics enhanced Spatial-TemporAl gRaph Diffuser (KStar Diffuser). Specifically, (1) to incorporate the physical robot structure information into action prediction, KStar Diffuser maintains a dynamic spatial-temporal graph according to the physical bimanual joint motions at continuous timesteps. This dynamic graph serves as the robot-structure condition for denoising the actions; (2) to make the NBP learning objective consistent with kinematics, we introduce the differentiable kinematics to provide the reference for optimizing KStar Diffuser. This module regularizes the policy to predict more reliable and kinematics-aware next end-effector poses. Experimental results show that our method effectively leverages the physical structural information and generates kinematics-aware actions in both simulation and real-world. + +# 1. Introduction + +Bimanual manipulation [34, 36, 42, 45] represents a fundamental capability for robotic systems to perform com + +![](images/00add3a9ccda9ddd24e457da885fb141b35afbbd665f185f3aa0eac03b07d4e5.jpg) +Figure 1. The self-collision problem in bimanual manipulation tasks due to overlooking the robotic structure. + +plex tasks requiring two-arm coordination. While imitation learning has demonstrated remarkable success in single-arm manipulation [6, 7, 33], its extension to bimanual scenarios faces unique challenges as robots need to coordinate dual-arm movements while conforming to physical constraints. These challenges significantly impact the reliability and feasibility of predicted actions in real-world applications. + +A straightforward approach [9, 10, 13, 55, 56, 58] would be to learn the entire motion trajectory by predicting joint positions at consecutive timesteps. However, this poses significant challenges as trajectories are typically long and difficult to learn. Therefore, mainstream approaches [15, 17, 31, 44] typically adopt a two-stage pipeline: first predicting the next-best end-effector pose (NBP), then computing joint rotations through inverse kinematics [21, 26]. However, the dominant NBP-based approaches, though simplify the learning process, often lead to unreliable motion generation. Concretely, the predicted poses may violate physical structure constraints [27, 43], causing potential arm collisions, or exceeding joint limits due to inadequate consideration of kinematic feasibility, as Figure 1 illustrated. + +In real-world bimanual manipulation tasks, existing ap + +proaches [9, 15, 17, 44, 57] exhibit significant limitations in motion execution despite their promising performance in trajectory prediction. Empirical observations indicate that while individual end-effector poses are feasible in isolation, their concurrent execution frequently results in interarm collisions [43] and kinematically infeasible configurations [31, 35]. The fundamental challenge lies in the optimization of end-effector poses exclusively in Cartesian space, which introduces a substantial discrepancy between motion prediction and physical constraints. Existing approaches [9, 10, 17] incorporate proprioception information in a simplistic manner, (e.g., gripper rotation angles [17, 44] and end-effector poses [48]), while disregarding the robot's intrinsic structural properties e.g., the kinematic chain and joint configurations during the prediction phase. This oversimplified representation fails to capture the crucial spatial correlations between robotic arms, joint links, and their potential interactions. Consequently, the predicted dual-arm trajectories may satisfy task objectives in Cartesian space while violating crucial spatial constraints during execution. + +Moreover, the conventional paradigm [17, 24, 44, 46] of treating inverse kinematics as a post-processing procedure introduces additional complexities. The disjunction between kinematic constraints and the pose prediction learning objective often yields solutions that exhibit apparent smoothness in Cartesian space while manifesting discontinuous or mechanically infeasible trajectories in the joint space. This limitation becomes particularly pronounced in configurations approaching kinematic singularities or joint limits, where the mapping between Cartesian and joint spaces becomes ill-conditioned. These inherent limitations necessitate the integration of both structural and kinematic constraints into the policy learning framework. + +To address the above issues, we propose Kinematics enhanced Spatial-TemporAl gRaph Diffuser (KStar Diffuser), a novel framework that explicitly incorporates both robot structures and kinematics into the bimanual motion generation process. Our key insight is that the physical structure and kinematic properties of the robot should guide the learning process of pose prediction, rather than be treated as independent post-processing constraints. Specifically, (1) to incorporate structural awareness, we construct a dynamic spatial-temporal graph from the robot's URDF specifications, where nodes represent joint properties and edges capture both spatial relations and temporal dependencies. This graph structure is encoded via Graph Convolutional Network (GCN) [25] to provide explicit physical constraints for the diffusion process; (2) for kinematic feasibility, we regularize the NBP learning objective by incorporating joint-space prediction, where the predicted joint positions are mapped to reference end-effector poses through differentiable forward kinematics. These kinematically-feasible poses then serve as conditions + +to guide the diffusion process, ensuring that the generated motions satisfy both structural and kinematic constraints. Our main contributions are as follows: + +- Different from existing approaches that optimize end-effector poses solely in Cartesian space, we propose a novel spatial-temporal robot graph that explicitly models the robot physical configuration to guide the generative action denoising procedure. +- We introduce a kinematics regularizer that augments the NBP learning objective by introducing joint-space supervision. This regularizer leverages forward kinematics to provide kinematically-feasible reference poses, effectively guiding the diffusion process to conform to kinematic constraints. +- Extensive experiments show that our proposed KStar Diffuser is superior in both simulation and real-world scenarios, surpassing baselines more than $10\%$ in success ratio. + +# 2. Related Work + +Diffusion Models in Bimanual Robotic Controls. Diffusion models such as Denoising Diffusion Probabilistic Models (DDPM) [19] have achieved great success in the fields of image generation [38, 39, 41] and video generation [37, 50, 59]. Thus, recent work [8, 40, 47, 52, 60] has been devoted to applying its powerful generation capabilities to action generation for robotic manipulation tasks in a imitation learning mode. With the 2D images [9, 31, 40, 47, 49] or 3D point cloud observations [8, 20, 23, 55, 56], the policy is trained to output the action sequence of joint positions or end-effector poses by iterative denoising process. + +Bimanual manipulation tasks, which resemble human-like actions more closely, have garnered increasing attention [10, 14, 17, 30, 58]. However, existing research primarily extends single-arm manipulation methods to dual-arm tasks [10, 58], without considering about the distinct challenges specific to bimanual settings. The relative independence of the two arms makes it essential to consider self-collision avoidance [1, 3] when predicting actions. In practice, the robot's structure significantly influences its motion since the joint types determine movement directions and joint angles constrain the range of motion. Although some studies [15, 44, 48] integrate proprioception in robotic control, they often use simply low-dimensional vectors to represent body information, lacking depth in structural exploration and overlooking critical spatial relationships between dual arms. Therefore, we study the physical structure information of the dual-arm robot system, and improve both the accuracy and adaptability of robotic movements. + +Robotic Kinematics Modeling. Robotic kinematics modeling [2, 16, 31, 53, 54] is a classic problem in the robotic control [4, 18, 32, 51] where inverse kinematics (IK) presents a fundamental challenge, i.e., deriving the joint configuration given the end-effector pose. The inher + +ent non-uniqueness of IK solutions as a significant obstacle for learning algorithms [12, 22], often produce inaccurate models by averaging over nonconvex feasible sets. Recent advancements integrate neural solutions [11, 28, 29] to enhance the adaptability and precision of kinematic models. Bócsi et al. [5] leveraged support vector machines to parameterize quadratic programs, aligning solutions with IK in specific workspace regions. Current data-driven imitation learning approaches [10, 13, 15, 17, 58] primarily rely on probabilistic modeling of action distributions to predict subsequent actions, but lack the reliability guarantees from the aspect of kinematics. In this paper, we construct a spatial-temporal graph to learn the robot physical representation, and provide a kinematics-aware latent end-effect embedding for diffusion policy as guidance. In this way, o'ur approach enhances the robot's ability to perform tasks with superious precision and adaptability. + +# 3. Method + +# 3.1. Preliminary + +Diffusion Policy. Chi et al. [9] propose the diffusion policy which represents a robot's visuomotor policy as a conditional denoising diffusion process. It learns a model distribution $\pi_{\phi}(\pmb{a}_0|\pmb{o})$ conditioned on observation $\pmb{o}$ to approximate joint distribution $q(\pmb{a}_0)$ . The whole procedure consists of a forward process and a reverse process. + +(1) Forward Process: For the Markov chain with Gaussian transitions parameterized, the policy $\pi_{\phi}(\pmb{a}_0|\pmb{o})$ learns a fixed inference procedure $q(\pmb{a}_k|\pmb{a}_0)$ : + +$$ +\begin{array}{l} q \left(\boldsymbol {a} _ {k} \mid \boldsymbol {a} _ {0}\right) = \int q \left(\boldsymbol {a} _ {1: k} \mid \boldsymbol {a} _ {0}\right) \mathrm {d} \boldsymbol {a} _ {1: (k - 1)} \tag {1} \\ = \mathcal {N} \left(\boldsymbol {a} _ {k}; \sqrt {\bar {\alpha} _ {k}} \boldsymbol {a} _ {0}, (1 - \bar {\alpha} _ {k}) \boldsymbol {I}\right) \\ \end{array} +$$ + +Thus, $a_{k}$ can be expressed as a linear combination of $a_0$ and a noise variable $\epsilon$ : + +$$ +\boldsymbol {a} _ {k} = \sqrt {\bar {\alpha} _ {k}} \boldsymbol {a} _ {0} + \sqrt {1 - \bar {\alpha} _ {k}} \varepsilon \tag {2} +$$ + +The training loss is obtained: + +$$ +\mathcal {L} = \mathbb {E} _ {\boldsymbol {a} _ {0} \sim q (\boldsymbol {a} _ {0}), \epsilon_ {k} \sim \mathcal {N} (\boldsymbol {0}, \boldsymbol {I}), k} \left[ \left| \left| \epsilon_ {k}, \pi_ {\phi} (\boldsymbol {a} _ {0} + \epsilon_ {k}, \boldsymbol {o}, k) \right| \right| ^ {2} \right] \tag {3} +$$ + +where $\epsilon_{k}$ is the random noise sampling at $k$ iteration. + +(2) Reverse Process: Starting from a sample $\pmb{a}_K \sim \mathcal{N}(\pmb{0}, \pmb{I})$ , the reverse steps are: + +$$ +\boldsymbol {a} _ {k - 1} = \sqrt {\bar {\alpha} _ {k - 1}} \pi_ {\phi} (a _ {k}, k, o) + \sqrt {1 - \bar {\alpha} _ {k - 1}} \epsilon . \tag {4} +$$ + +By the iterative denoising process, the policy generates $a_0$ as its next action. The continuous predicted action sequence forms the complete action trajectory. + +Robotic Kinematics. Robotics kinematics describes the relationship between the robot joints and its end-effector. It can be divided into forward kinematics and inverse kinematics for controlling robot movements. + +(1) Forward Kinematics (FK): Given a specific configuration of joint angles, i.e., $\theta = [\theta_{1},\theta_{2},\dots,\theta_{n}]\in \Theta$ , it aims to construct a function $f_{\mathrm{FK}}:\mathbb{R}^n\to SE(3)$ , computing the end-effector position and orientation that are represented as a pose matrix $\mathbf{T}\in SE(3)$ . + +$$ +F K (\theta) = \mathbf {T} \in \mathcal {T}, \text {w h e r e} \mathbf {T} = \prod_ {i = 1} ^ {n} \mathbf {T} _ {i} (\theta_ {i}), (5) +$$ + +where $\mathbf{T}_i(\theta_i)$ is the homogeneous transformation matrix for joint $i$ , and $\mathcal{T}$ means the task space. $\Theta \in \mathbb{R}^n$ denotes the configuration space, while the configuration refers to the specific arrangement of the robotic joints in a given pose. + +(2) Inverse Kinematics (IK): In contrast, IK aims to find the joint configurations which achieve a desired end-effector pose. It is defined as a mapping $f_{\mathrm{IK}}: SE(3) \to \mathbb{R}^n$ from the task space $\mathcal{T}$ to the set of possible configurations $\Theta$ that satisfy the given pose matrix $\mathbf{T}$ : + +$$ +I K (\mathcal {T}) = \{\theta \in \Theta | F K (\theta) = \mathbf {T} \}. \tag {6} +$$ + +It is noted that this mapping is often non-unique, especially for redundant manipulators with more than 6 DoF. + +# 3.2. Task Definition + +Given the dataset $D$ with language instructions $\pmb{l}$ and the RGB-D observations $\pmb{o}$ , we aim to learn a policy $\pi_{\phi}(\pmb{a}| \pmb{o}, \pmb{l})$ which predicts actions $\pmb{a}$ . Here, $\pmb{a}$ is composed of a trajectory $\pmb{a}^{\mathrm{joint}} = \{\pmb{a}_0, \pmb{a}_1, \dots, \pmb{a}_T\}$ and gripper opening or closing action $\pmb{a}^{\mathrm{gripper}}$ , where $T$ denotes the trajectory length and $\pmb{a}_i \in \mathbb{R}^n$ . $n$ represents the number of robot joints. In the bimanual manipulation, $n$ is usually to be 12 or 14 as each robot arm has 6 or 7-DoF. Referring to prior works [17, 44], it is inefficient to train the policy on all trajectory points. Thus, a keyframe discovery method is used to extract a set of $K_{\xi}$ keyframe indices $\{k_i\}_{i=0}^{K_{\xi}}$ and the prediction actions are the key set of end-effector poses $\pmb{a}^{\mathrm{EE}}$ . + +# 3.3. KStar Diffuser + +# 3.3.1. Overview + +Mainstream methods [17, 23, 58] train a policy to predict actions, but take little consideration about the mechanical robot structure which determines its motion. We thus propose a spatial-temporal graph to model both the static physical structure and dynamic history movement information. In addition, to reduce kinematically infeasible predictions for the end-effector pose, we introduce a differentiable kinematics module that provides kinematics-aware references to the policy network. The overview of our proposed Kinematics enhanced Spatial-TemporAl gRaph Diffuser (KStar Diffuser) is shown in Figure 2. + +![](images/2bb8018e2225916c9c5a23ac6d9af49d89107a0d63faf94afc1e612f07d8fbfe.jpg) + +![](images/b1cddb061fecf0b7a76fe375fbb3848943c39b2a62c48ad7f571ecb82841926f.jpg) +Figure 2. Overview of KStar Diffuser. The top part presents the spatial-temporal graph which is constructed according to the robot architecture. The bottom part shows our backbone and the proposed kinematics regularizer. For the backbone, it extracts the multimodal information which consists of multiview RGB-D observations and language instruction, and then generates bimanual 6D end-effector poses. The kinematics regularizer enhances pose learning by incorporating joint-level predictions, which are mapped to reference end-effector poses through differentiable forward kinematics (FK). + +# 3.3.2. Backbone + +Given the language instruction $l$ and multiview RGB-D observation images $o$ , we first adopt the Transformer-based encoders to extract their features $\mathbf{E}_{\mathrm{I}}$ and $\mathbf{E}_{\mathrm{T}}$ , respectively. Then the features are fused with the $l_{\mathrm{FiLM}}$ -layers FiLM block to obtain the hidden states $\mathbf{H}_{\mathrm{B}}$ . Each layer is combined with an up-sampling 2D convolution layer: + +$$ +\mathbf {H} _ {\mathrm {B}} ^ {i} = \operatorname {C o n v 2 D} \left(\mathbf {H} _ {\mathrm {B}} ^ {i}\right), \tag {7} +$$ + +$$ +\mathbf {H} _ {\mathrm {B}} ^ {i + 1} = \operatorname {F i L M} ^ {i} \left(\mathbf {H} _ {\mathrm {B}} ^ {i}, \mathbf {E} _ {\mathrm {T}}\right), \tag {8} +$$ + +where $i\in [0,l_{\mathrm{FiLM}})$ and the $\mathbf{H}_{\mathrm{B}}^{0}$ is initialized with $\mathbf{E}_{\mathrm{I}}$ + +Our backbone $\pi_{\phi}$ uses the last hidden states $\mathbf{H}_{\mathrm{B}}$ as the condition to guide a diffusion head to denoise and generate the bimanual end-effector pose. It is noted that we attach the $n$ historical observation images to provide more information for capturing the motion tendency. Following Chi et al. [9], we let the policy predict next $m$ actions during training to alleviate multimodal problems. Here, we set both $n$ and $m$ to 2. The action prediction is as follows: + +$$ +\boldsymbol {a} _ {k - 1} ^ {\mathrm {E E}} \sim \pi_ {\phi} \left(\boldsymbol {a} _ {k - 1} ^ {\mathrm {E E}} \mid \boldsymbol {a} _ {k} ^ {\mathrm {E E}}, \mathbf {H} _ {\mathrm {B}}\right), \tag {9} +$$ + +$$ +\begin{array}{l} \pi_ {\phi} \left(\boldsymbol {a} _ {k - 1} ^ {\mathrm {E E}} \mid \boldsymbol {a} _ {k} ^ {\mathrm {E E}}, \mathbf {H} _ {\mathrm {B}}\right) \tag {10} \\ = \mathcal {N} (\boldsymbol {a} _ {k - 1} ^ {\mathrm {E E}}; \sqrt {\bar {\alpha} _ {k - 1}} \pi_ {\phi} (\boldsymbol {a} _ {k} ^ {\mathrm {E E}}, \mathbf {H} _ {\mathrm {B}}); (1 - \bar {\alpha} _ {k - 1}) I). \\ \end{array} +$$ + +The learning objective is: + +$$ +\mathcal {L} _ {\mathrm {E E}} = \mathbb {E} _ {q} \left[ \log \frac {\pi_ {\phi} \left(\boldsymbol {a} _ {0 : K} ^ {\mathrm {E E}} \mid \mathbf {H} _ {\mathrm {B}}\right)}{q \left(\boldsymbol {a} _ {1 : K} ^ {\mathrm {E E}} \mid \boldsymbol {a} _ {0} ^ {\mathrm {E E}}, \mathbf {H} _ {\mathrm {B}}\right)} \right]. \tag {11} +$$ + +# 3.3.3. Spatial-Temporal Robot Graph + +The physical architecture impacts the motion of the whole robot, determining whether it can complete the task. Meanwhile, the historical spatial information is also important to future movements. Thus, we propose a spatial-temporal graph method to model the robot architecture at each step and the robot motion at continuous timesteps, representing the static spatial information and dynamic motion features. + +Spatial Structure Graph Construction. To represent the robot structure, we first parse the Unified Robot Description Format (URDF) file which is usually used to describe the static physical structure of robot such as joint types, joint limits, and link lengths. Then, we define a dual-arm system as an undirected graph $G_{\mathrm{s}} = \langle V,E\rangle$ based on the joint and link configuration. Here, $V = \{\pmb{v}_{\mathrm{j}_1},\pmb{v}_{\mathrm{j}_2},\dots,\pmb{v}_{\mathrm{j}_m}\}$ and $E = \{e_{ij}|(\mathrm{joint}_i,\mathrm{joint}_j)\in \mathrm{links}\}$ represents the node set of joints and edge set of links respectively, where $m$ denotes the number of joint. We use $f_{i}\in \mathbb{R}^{D}$ to denote the value of $i$ -th node feature. It consists of three attributes as follows: + +- Joint Coordinate: we use a vector in Cartesian coordinate system, $\pmb{f}_i^{\mathrm{JC}} = [\pmb{x}_i, \pmb{y}_i, \pmb{z}_i] \in \mathbb{R}^3$ , to denote the absolute coordinates of the $i$ -th joint. The vector is normalized according to the workspace boundary for stable convergence of model training. +- Joint Distance: To measure the spatial relationship between the node $\mathbf{v}_i$ and the other nodes $\mathbf{v}_j$ , we compute the Euclidean Distance between $\mathbf{v}_i$ and $\mathbf{v}_j$ : + +$$ +\boldsymbol {f} _ {i} ^ {\mathrm {J D}} = \left| \left| \boldsymbol {v} _ {i} - \boldsymbol {v} _ {j} \right| \right| _ {2}, \tag {12} +$$ + +where $\pmb{f}_i^{\mathrm{JD}}\in \mathbb{R}^m$ and $||\cdot ||_2$ means the Euclidean norm. + +- Body Label: To discriminate the source of node $\mathbf{v}_i$ , we use a one-hot vector $\mathbf{f}_i^{\mathrm{BL}} = [0,1]$ as one of its features. It can also help the policy capture the motion modes of different robot arms such as the symmetry. + +We concatenate $\pmb{f}_i^{\mathrm{JC}}$ , $\pmb{f}_i^{\mathrm{JD}}$ , and $\pmb{f}_i^{\mathrm{BL}}$ to form the completed feature $\pmb{f}_i \in \mathbb{R}^{D_f}$ , where $D_{f}$ denotes the dimension of node features. + +Spatial-Temporal Graph Learning. Given the same instruction and observation, different historical robot poses result in different predictions. Thus, we build the spatial structure graph with the temporal motion information. Specifically, we build a spatial-temporal graph $G_{\mathrm{ST}} = \langle V_{\mathrm{ST}}, E_{\mathrm{ST}} \rangle$ by combining $\{G_{\mathrm{S}}^{i} = \langle V_{\mathrm{S}}^{i}, E_{\mathrm{S}}^{i} \rangle \}_{i=0}^{T-1}$ from historical timesteps, where $T$ denotes the number of history steps. In $G_{\mathrm{ST}}$ , the node set $V_{\mathrm{ST}}$ contains $V_{\mathrm{S}}^{i}$ from historical static spatial graph. Additionally, we add edges $E'$ linking the same joint node $\text{joint}_i^t$ at different timesteps, in order to establish the correlation of joint motions at continuous times. It can be formulated as follows: + +$$ +V _ {\mathrm {S T}} = \cup \left\{V _ {\mathrm {S}} ^ {i} \right\} _ {i = 0} ^ {T - 1}, E _ {\mathrm {S T}} = \cup \left\{E _ {\mathrm {S}} ^ {i} \right\} _ {i = 0} ^ {T - 1} \cup E ^ {\prime}, \tag {13} +$$ + +$$ +E ^ {\prime} = \left\{\boldsymbol {e} _ {i} ^ {t, t ^ {\prime}} \mid t, t ^ {\prime} \in \{0, 1, \dots , T \}, t \neq t ^ {\prime} \right\}. \tag {14} +$$ + +In this way, we obtain the whole spatial-temporal graph. A Graph Convolutional Network (GCN) is then adopted to propagate and aggregate node features across the graph. The GCN layer updates each node feature $f_{i}$ by aggregating the features of its neighboring nodes, thereby capturing the relational and structural information of the robotic arms. We use the node feature $\mathbf{H}_{\mathrm{G}}$ of the last encoder layer as the representation of the robot structure to condition the denoising process. + +# 3.3.4. Kinematics Regularizer + +To control the end-effector effectively, the generated pose trajectory must be processed by an Inverse Kinematics (IK) solver, which calculates the joint configurations to achieve the specified poses. However, because the predicted trajectory is generated without considering the robot kinematic constraints, it often falls outside the IK solver's feasible range, resulting in high failure rates during execution. To address this limitation, we propose a kinematics regularizer into the end-effector pose learning objective. This regularizer aligns the predicted poses with the robot kinematic constraints, ensuring that the generated trajectory remains within the solvable space of the IK solver, thus enhancing then reliability of trajectory execution. + +Differentiable Kinematics. Given a joint configuration $\Theta = [\theta_1, \theta_2, \dots, \theta_n]$ , the corresponding end-effector pose $\mathbf{T} \in SE(3)$ can be computed using forward kinematics, + +represented as a mapping $f(\Theta): \mathbb{R}^n \to SE(3)$ . This mapping from joint space to end-effector space is differentiable, namely Differentiable Forward Kinematics (DFK), enabling the use of gradients to optimize our control policy. Leveraging DFK, our policy learns to predict the next joint configuration $\hat{a}^{\mathrm{joint}}$ , from which we compute an intermediate end-effector pose $\mathbf{H}_{\mathrm{R}}$ . By using $\mathbf{H}_{\mathrm{R}}$ as a reference, we guide a denoising process to generate a precise and executable end-effector pose. + +Specifically, we combine structure features $\mathbf{H}_{\mathrm{G}}$ with the last hidden state $\mathbf{H}_{\mathrm{B}}$ , projecting to the joint space and using the DFK to obtain reference as $\mathbf{H}_{\mathrm{R}}$ follows: + +$$ +\hat {\boldsymbol {a}} ^ {\text {j o i n t}} = \operatorname {P r o j} \left(\left[ \mathbf {H} _ {\mathrm {B}}, \mathbf {H} _ {\mathrm {G}} \right]\right), \quad \mathbf {H} _ {\mathrm {R}} = \operatorname {D F K} \left(\hat {\boldsymbol {a}} ^ {\text {j o i n t}}\right). \tag {15} +$$ + +To ensure the consistency between predicted and actual joint angles, we minimize the joint loss: + +$$ +\mathcal {L} _ {\text {j o i n t}} = \mathbb {E} _ {\boldsymbol {a} ^ {\text {j o i n t}} \sim q (\boldsymbol {a} _ {0})} \left[ \left| \left| \boldsymbol {a} ^ {\text {j o i n t}} - \hat {\boldsymbol {a}} ^ {\text {j o i n t}} \right| \right| ^ {2} \right]. \tag {16} +$$ + +Conditioning Diffusion Process on Kinematics. To enforce kinematic consistency, we condition the diffusion process on the joint angle representation $\mathbf{H}_{\mathrm{R}}$ , an auxiliary input encoding kinematic constraints. This allows the predicted pose trajectory to stay within feasible space. Given the diffusion steps from Eq.(9) and Eq.(10), we have: + +$$ +\boldsymbol {a} _ {k - 1} ^ {\mathrm {E E}} \sim \pi_ {\phi} \left(\boldsymbol {a} _ {k} ^ {\mathrm {E E}} \mid \boldsymbol {a} _ {k - 1} ^ {\mathrm {E E}}, \mathbf {H} _ {\mathrm {B}}, \mathbf {H} _ {\mathrm {R}}\right), \tag {17} +$$ + +$$ +\pi_ {\phi} \left(\boldsymbol {a} _ {k - 1} ^ {\mathrm {E E}} \mid \boldsymbol {a} _ {k} ^ {\mathrm {E E}}, \mathbf {H} _ {\mathrm {B}}, \mathbf {H} _ {\mathrm {R}}\right) = +$$ + +$$ +\mathcal {N} \left(\boldsymbol {a} _ {k - 1} ^ {\mathrm {E E}}; \sqrt {\bar {\alpha} _ {k - 1}} \pi_ {\phi} \left(\boldsymbol {a} _ {k} ^ {\mathrm {E E}}, \mathbf {H} _ {\mathrm {B}}, \mathbf {H} _ {\mathrm {R}}\right); (1 - \bar {\alpha} _ {k - 1}) \boldsymbol {I}\right). \tag {18} +$$ + +Incorporating DFK into the diffusion process allows gradients from the pose loss to propagate back through the kinematic function, ensuring that each denoising step maintains compliance with joint constraints, thereby optimizing the end-effector's control accuracy and robustness. + +# 3.4. Training and Inference + +Training. We use the conditional action generation mode to train KStar Diffuser, which is formulated as a conditional denoising diffusion. The loss function is defined as the Mean Square Error (MSE) as follows: + +$$ +\boldsymbol {a} _ {k} = \sqrt {\bar {\alpha} _ {k}} \boldsymbol {a} _ {0} + \sqrt {1 - \bar {\alpha} _ {k}} \epsilon , \tag {19} +$$ + +$$ +\mathcal {L} _ {\mathrm {E E}} = \mathbb {E} _ {\boldsymbol {a} _ {0} ^ {\mathrm {E E}} \sim q (\boldsymbol {a} _ {0}), k} \left[ \left| \left| \boldsymbol {a} _ {0} ^ {\mathrm {E E}} - \pi_ {\phi} (\boldsymbol {a} _ {k} ^ {\mathrm {E E}}, k, C) \right| \right| ^ {2} \right], \tag {20} +$$ + +$$ +\mathcal {L} _ {\text {j o i n t}} = \mathbb {E} _ {\boldsymbol {a} _ {0} ^ {\text {j o i n t}} \sim q (\boldsymbol {a} _ {0})} [ \| \boldsymbol {a} _ {0} ^ {\text {j o i n t}} - \pi_ {\phi} (\mathbf {H} _ {\mathrm {B}}, \mathbf {H} _ {\mathrm {G}}) \| ^ {2} ], \tag {21} +$$ + +$$ +\mathcal {L} = \lambda \mathcal {L} _ {\mathrm {E E}} + (1 - \lambda) \mathcal {L} _ {\text {j o i n t}}, \tag {22} +$$ + +where $\mathbf{a}_k$ is obtained by the forward diffusion process and $C$ is the combination of $\mathbf{H}_{\mathrm{B}}$ , $\mathbf{H}_{\mathrm{G}}$ , $\mathbf{H}_{\mathrm{R}}$ . $\lambda$ is the trade-off coefficient. + +Inference. Sampling from a Gaussian noise $\epsilon_{k}$ , the policy $\pi_{\phi}$ performs $K$ iterations to gradually denoise a random noise $\epsilon_{k}$ into the noise-free action $\pmb{a}_{0}$ : + +$$ +\boldsymbol {a} _ {k - 1} ^ {\mathrm {E E}} = \sqrt {\overline {{\alpha}} _ {k - 1}} \pi_ {\phi} (\boldsymbol {a} _ {k} ^ {\mathrm {E E}}, k, C) + \sqrt {1 - \overline {{\alpha}} _ {k - 1}} \epsilon_ {k}. \tag {23} +$$ + +We use the predicted $a_0^{\mathrm{EE}}$ as the final action to control robot execution. + +# 4. Experiment + +# 4.1. Dataset and Evaluation Settings + +Dataset. Bimanual manipulation tasks demand high levels of coordination, synchronization, and symmetry awareness between the two robotic arms, making them inherently more challenging than single-arm tasks. To assess the capabilities of KStar Diffuser in these areas, we conducted comprehensive experiments using the RLBench2 benchmark [17], an extended version of RLBench tailored for bimanual manipulation and comprising tasks closely resembling real-world scenarios. Please refer to Appendix for more details about RLBench2 and real-world tasks. + +Evaluation Settings. To evaluate the policy performance, we employ success rate as the primary metric. Although the policy generates multiple sequential actions during execution, we primarily focus on the final goal achievement rather than intermediate steps. Each task is associated with a specific success criterion defined by its target state. To comprehensively assess the policy's capability, we conduct experiments with varying numbers of demonstrations (20 and 100) during training. Figure 3 illustrates our experimental setup, including both the simulation environment and the Cobot Agilex ALOHA robot. Detailed descriptions of simulation tasks and real-world experimental settings are provided in the Appendix. + +![](images/19ce8d6ab0fd354f2d346adf3d79d5393a255af8688d43ce0902f4732a570abf.jpg) +Figure 3. The left: the simulation environment of pick_laptop task. The right: the ALOHA device used in the real-world tasks. + +# 4.2. Baselines + +We systematically evaluate KStar Diffuser against state-of-the-art methods in two major categories: + +Transformer-based methods: (1) Action Chunking with Transformers (ACT) [58] employs a Conditional VAE architecture, consisting of an encoder-decoder framework + +for joint angle sequence prediction; (2) Robotic View Transformer Leader Following (RVT-LF) [17] leverages RVT [15] as its backbone, incorporating a multi-view transformer for cross-view information aggregation and image re-rendering, coupled with a leader following mechanism for action prediction; (3) Perceiver-Actor Leader Following (PerAct-LF) [17] adopts the leader following paradigm based on PerAct [44], utilizing a perceiver Transformer to encode both instructions and voxel observations for optimal voxel action generation; (4) PerAct2 [17] enhances PerAct by implementing a unified feature space for dual-arm actions and employing combined self-attention for synchronized bimanual action prediction. + +Diffusion-based methods: (1) Joint-based Diffusion Policy (DP-J) [9] adopts a diffusion model to robotic manipulation within an imitation learning framework, focusing on joint angle prediction; (2) EndEffector-based Diffusion Policy (DP-EE) is a DP variant which we reimplement Diffusion Policy to predict end-effector poses instead of joint angles, offering an alternative control paradigm; (3) 3D Diffusion Policy (DP3) [56] enhances 3D perception by incorporating point clouds for joint angle prediction. + +# 4.3. Comparison Results with SOTA Methods + +Experimental Results on RLBench2. As shown in Table 1, the KStar Diffuser significantly outperforms other state-of-the-art baselines, achieving more than $20\%$ higher overall performance with both 20 and 100 training demonstrations. We found that: + +(1) Similar to learning a single-arm policy, the process of learning a bimanual policy can adapt quickly and achieve a high success rate, given a relatively consistent distribution of task trajectories. In push_box, where the objective is for both arms to push a box toward a specified target along fixed trajectories, our KStar Diffuser and other baseline models perform well. However, as task complexity increases, success rates decrease. For example, in the liftball task, both arms must lift a large ball simultaneously to complete the task. Any asynchrony in movement can cause instability, leading to the ball slipping and ultimately resulting in task failure. Our KStar Diffuser achieves its robust performance on such bimanual tasks by explicitly modeling the spatial and motion relationships between the two arms, surpassing other methods more than $6\%$ . + +(2) Distinct from single-arm systems, bimanual robotic systems possess the capability for collaborative manipulation. Methods which are directly adapted from single-arm to bimanual manipulation exhibit high failure rates in tasks, e.g., pick_laptop, as they lack consideration of the spatial and motion relationships between the arms. Specifically, as shown in Figure 4, this task involves picking up a notebook lying flat on a cabinet surface. Given that the notebook rests fully against the tabletop, direct grasping by + +Table 1. The experimental result on simulated tasks. We train the policy with the setting of different training demonstrations, i.e. [20, 100], to test its capability comprehensively. The best results are in bold. Each result is reported with three seeds on average. + +
MethodBackbonePush BoxLift BallHandover Item (easy)Pick LaptopSweep DustpanOverall
20 demos
ACT [58]Transformer26.0 ±4.617.3 ±5.80.0 ±0.00.0 ±0.00.0 ±0.08.7 ±1.6
RVT-LF [17]Transformer44.0 ±1.012.3 ±6.70.0 ±0.02.0 ±1.70.0 ±0.011.7 ±1.5
PerAct-LF [17]Transformer63.7 ±3.142.3 ±8.13.3 ±0.67.0 ±2.74.7 ±0.624.2 ±1.6
PerAct2 [17]Transformer59.7 ±7.637.7 ±4.78.0 ±1.017.3 ±0.61.7 ±1.524.9 ±2.1
DP-J [9]Diffusion30.0 ±3.619.7 ±2.10.0 ±0.00.0 ±0.00.0 ±4.010.0 ±1.0
DP-EE [9]Diffusion34.7 ±2.534.3 ±7.40.7 ±0.61.3 ±0.614.0 ±1.717.0 ±1.6
DP3 [56]Diffusion25.3 ±4.034.3 ±7.60.0 ±0.02.7 ±2.90.0 ±0.012.5 ±1.3
KStar Diffuser (Ours)Diffusion79.3 ±3.587.0 ±2.723.7 ±0.617.0 ±2.083.0 ±4.458.0 ±1.4
100 demos
ACT [58]Transformer48.7 ±6.840.7 ±2.10.0 ±0.00.0 ±0.00.0 ±0.017.9 ±1.0
RVT-LF [17]Transformer76.3 ±1.528.7 ±1.00.0 ±0.01.3 ±1.22.3 ±3.221.7 ±0.7
PerAct-LF [17]Transformer66.3 ±3.168.3 ±6.17.3 ±2.114.3 ±3.88.0 ±2.732.9 ±2.5
PerAct2 [17]Transformer77.0 ±3.047.7 ±3.210.3 ±0.634.3 ±4.06.7 ±2.133.9 ±1.0
DP-J [9]Diffusion62.7 ±7.643.3 ±2.10.0 ±0.01.3 ±0.60.0 ±4.021.5 ±1.6
DP-EE [9]Diffusion61.0 ±2.059.7 ±3.52.0 ±1.010.3 ±3.169.7 ±2.540.5 ±1.2
DP3 [56]Diffusion56.0 ±3.664.0 ±2.70.0 ±0.06.3 ±3.11.7 ±2.125.6 ±1.4
KStar Diffuser (Ours)Diffusion83.0 ±1.798.7 ±1.527.0 ±1.743.7 ±4.589.0 ±5.268.2 ±2.1
+ +the robotic arm is not possible. Instead, an effective strategy is to control one arm to push the notebook outward from the cabinet by a short distance, allowing the other arm to pick it up. KStar Diffuser achieves a success rate approximately $9\%$ higher than other methods, demonstrating its ability to capture the coordinated motion patterns required for collaborative object manipulation between two arms. + +Real-world Experimental Results. To comprehensively evaluate the policy's effectiveness, we build 2 tasks in real-world based on the simulation benchmark. The performances on real-world tasks are illustrated in Table 2. + +Similar to the simulation results, we observe that policies which do not consider about bimanual scenarios, i.e., ACT, DP, and DP3, demonstrate limited capability across all bimanual tasks, achieving around $20\%$ success ratio on average. Although PerAct2 is designed for bimanual tasks via mapping bimanual actions into a shared learning space, it fails to capture the spatial structure of the bimanual system, leading to ineffective arm coordination during execution. Furthermore, we also found PerAct2 faces significant inverse kinematics issues with its predicted end-effector poses, including joint configuration conflicts and unreachable positions which are shown in Figure 4. It is likely due to PerAct2's limited ability to capture the complex spatial constraints and kinematic relationships within the bimanual robotic system. In contrast, KStar Diffuser achieves superior bimanual coordination, surpassing other methods by over $10\%$ , as it successfully captures the motion patterns + +between dual arms and predicts feasible end-effector poses. + +Table 2. The result of real-world tasks. We train all policies with 100 demonstrations and test 15 times. The best result are in bold. + +
MethodsLift PlateHandoverOverall
ACT [58]37.8 ±8.30.0 ±0.018.9 ±11.2
DP [9]42.1 ±6.50.0 ±0.021.1 ±12.5
DP3 [56]44.0 ±8.30.0 ±0.022.0 ±13.4
PerAct2 [17]51.1 ±6.38.8 ±3.029.9 ±10.2
KStar Diffuser (Ours)66.7 ±5.319.7 ±5.343.1 ±17.8
+ +Table 3. Ablation study of model components. + +
ST GraphKRHandover-SHandover-ROverall
✓✓27.0 ±1.719.7 ±5.323.4 ±5.2
✓×18.3 ±1.515.3 ±3.316.8 ±2.1
××16.3 ±3.113.3 ±9.414.8 ±2.1
+ +# 4.4. Ablation Studies + +Effects of Model Components. To systematically evaluate the contribution of each component in KStar Diffuser, we conduct ablation experiments on the handover-item task in both simulated and real-world environments. We design a progressive ablation process by first removing the Differential Forward Kinematics module while retaining the Spatial-Temporal Graph (ST Graph), and then completely disabling both ST Graph and Kinematics Regularizer (KR). + +![](images/514c1ca54d213a8ab09bf241b7777aa88f449f206b1e094afc4efcb0626cf4ed.jpg) +Figure 4. The visualization of bimanual manipulation on simulated RLBench2 and real-world tasks. The blue annotations represent the motion of the robot's left arm, while the green annotations indicate the motion of the right arm. + +The experimental results in Table 3 demonstrate the crucial role of each component. The removal of KR leads to a significant decrease in success rate, particularly pronounced in real-world scenarios. This performance degradation can be attributed to the fundamental differences between simulated and real environments. While simulated environments maintain consistent and noise-free inputs, real-world scenarios introduce various perturbations, e.g., sensor noise and light reflection, making the policy more susceptible to kinematic constraint violations without the regularization effect of KR. Further ablation by removing both ST Graph and KR results in a substantial performance drop across all experimental settings. This observation illustrates two key aspects: First, the ST Graph effectively captures the spatial-temporal dependencies among joints, which is essential for coordinating the relative positioning and interactions between robotic arms. Second, the graph structure's explicit encoding of the robotic physical architecture enhances the policy's robustness against unexpected perturbations by maintaining spatial and temporal coherence. + +# 4.5. Qualitative Analysis + +We further present qualitative analysis in Figure 4. We compare the performance of KStar Diffuser with DP3 and PerAct2 in executing bimanual manipulation tasks within both simulated and real-world environments, respectively. + +In the simulation task, since the laptop rests flat on the cabinet, direct lifting is not feasible. One robotic arm initiates a forward push, creating space, while the other arm + +concurrently grasps and elevates the laptop. KStar Diffuser effectively models this dual-arm coordination, generating a precise trajectory of synchronized actions. Conversely, DP3, adapted from a single-arm policy to a dual-arm configuration, fails to achieve effective coordination. Concretely, after executing the push motion, the right arm does not halt, obstructing the left arm's lifting process. + +In the real-world task, KStar Diffuser generates an executable item transfer trajectory between the left and right arms, with no collisions throughout the task, reflecting its strong environmental adaptability and collision avoidance capabilities. Conversely, PerAct2 encounters collisions during the handover process (marked in red), indicating less effective handling of dynamic real-world variables and a lack of kinematic awareness on the robot movements. + +# 5. Conclusion + +In this paper, we have proposed a novel Kinematics enhanced Spatial-TemporAl gRaph Diffuser (KStar Diffuser) that explicitly incorporates both robot structures and kinematics into the bimanual motion generation process. It consists of a spatial-temporal robot graph that explicitly models the robot physical configuration to guide the generative action denoising procedure, and a kinematics regularizer that augments the NBP learning objective by introducing joint-space supervision. Extensive experiments demonstrate that KStar Diffuser outperforms the baselines by a large margin on both simulation and real-world tasks. + +# 6. Acknowledgements + +We would like to thank all co-authors for their efforts and the reviewers for their constructive comments. This work is supported by Shenzhen Science and Technology Program (KQTD20240729102207002), National Natural Science Foundation of China (Grant No. 62406092), National Natural Science Foundation of China (Grant No. U24B20175), Shenzhen Science and Technology Program (Grant No. KJZD20240903100017022), Guangdong Basic and Applied Basic Research Foundation (2025A1515010169), and Research on Efficient Exploration and Self-Evolution of APP Agents & Embodied Intelligent Cerebellum Control Model and Collaborative Feedback Training Project (Grant No. TC20240403047). Additionally, it is also supported by National Natural Science Foundation of China (Grant No. 62306090, No. 62236003), Natural Science Foundation of Guangdong Province of China (Grant No. 2024A1515010147). + +# References + +[1] Lucia Angelini, Manuela Uliano, Angela Mazzeo, Mattia Penzotti, and Marco Controzzi. Self-collision avoidance in bimanual teleoperation using collisionik: algorithm revision and usability experiment. In 2022 IEEE-RAS 21st International Conference on Humanoid Robots (Humanoids), pages 112-118, 2022. 2 +[2] Andreas Aristidou, Joan Lasenby, Yiorgos Chrysanthou, and Ariel Shamir. Inverse kinematics techniques in computer graphics: A survey. In Computer graphics forum, pages 35-58. Wiley Online Library, 2018. 2 +[3] Weibang Bai, Ningshan Zhang, Baoru Huang, Ziwei Wang, Francesco Cursi, Ya-Yen Tsai, Bo Xiao, and Eric M Yeatman. Dual-arm coordinated manipulation for object twisting with human intelligence. In 2021 IEEE International Conference on Systems, Man, and Cybernetics (SMC), pages 902-908. IEEE, 2021. 2 +[4] Aaron M Bestick, Samuel A Burden, Giorgia Willits, Nikhil Naikal, S Shankar Sastry, and Ruzena Bajcsy. Personalized kinematics for human-robot collaborative manipulation. In 2015 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 1037–1044. IEEE, 2015. 2 +[5] Botond Bócsi, Duy Nguyen-Tuong, Lehel Csato, Bernhard Schoelkopf, and Jan Peters. Learning inverse kinematics with structured prediction. In 2011 IEEE/RSJ International Conference on Intelligent Robots and Systems, pages 698-703. IEEE, 2011. 3 +[6] Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Joseph Dabis, Chelsea Finn, Keerthana Gopalakrishnan, Karol Hausman, Alex Herzog, Jasmine Hsu, et al. Rt-1: Robotics transformer for real-world control at scale. arXiv preprint arXiv:2212.06817, 2022. 1 +[7] Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Xi Chen, Krzysztof Choromanski, Tianli Ding, Danny Driess, Avinava Dubey, Chelsea Finn, et al. Rt-2: + +Vision-language-action models transfer web knowledge to robotic control. arXiv preprint arXiv:2307.15818, 2023. 1 +[8] Shizhe Chen, Ricardo Garcia, Cordelia Schmid, and Ivan Laptev. Polarnet: 3d point clouds for language-guided robotic manipulation. arXiv preprint arXiv:2309.15596, 2023.2 +[9] Cheng Chi, Zhenjia Xu, Siyuan Feng, Eric Cousineau, Yilun Du, Benjamin Burchfiel, Russ Tedrake, and Shuran Song. Diffusion policy: Visuomotor policy learning via action diffusion. The International Journal of Robotics Research, page 02783649241273668, 2023. 1, 2, 3, 4, 6, 7 +[10] Cheng Chi, Zhenjia Xu, Chuer Pan, Eric Cousineau, Benjamin Burchfiel, Siyuan Feng, Russ Tedrake, and Shuran Song. Universal manipulation interface: In-the-wild robot teaching without in-the-wild robots. arXiv preprint arXiv:2402.10329, 2024. 1, 2, 3 +[11] Akos Csiszar, Jan Eilers, and Alexander Verl. On solving the inverse kinematics problem using neural networks. In 2017 24th International Conference on Mechatronics and Machine Vision in Practice (M2VIP), pages 1-6, 2017. 3 +[12] Aaron D'Souza, Sethu Vijayakumar, and Stefan Schaal. Learning inverse kinematics. In Proceedings 2001 IEEE/RSJ International Conference on Intelligent Robots and Systems. Expanding the Societal Role of Robotics in the Next Millennium (Cat. No. 01CH37180), pages 298-303. IEEE, 2001. 3 +[13] Zipeng Fu, Tony Z Zhao, and Chelsea Finn. Mobile aloha: Learning bimanual mobile manipulation with low-cost whole-body teleoperation. arXiv preprint arXiv:2401.02117, 2024. 1, 3 +[14] Jianfeng Gao, Xiaoshu Jin, Franziska Krebs, Noémie Jaquier, and Tamim Asfour. Bi-kvil: Keypoints-based visual imitation learning of bimanual manipulation tasks. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 16850-16857. IEEE, 2024. 2 +[15] Ankit Goyal, Jie Xu, Yijie Guo, Valts Blukis, Yu-Wei Chao, and Dieter Fox. Rvt: Robotic view transformer for 3d object manipulation. In Conference on Robot Learning, pages 694–710. PMLR, 2023. 1, 2, 3, 6 +[16] Reinhard Grassmann, Vincent Modes, and Jessica Burgner-Kahrs. Learning the forward and inverse kinematics of a 6-dof concentric tube continuum robot in se (3). In 2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 5125-5132. IEEE, 2018. 2 +[17] Markus Grotz, Mohit Shridhar, Tamim Asfour, and Dieter Fox. Peract2: A perceiver actor framework for bimanual manipulation tasks. arXiv preprint arXiv:2407.00278, 2024. 1, 2, 3, 6, 7 +[18] Abdullah Aamir Hayat, Ratan OM Sadanand, and Subir K Saha. Robot manipulation through inverse kinematics. In Proceedings of the 2015 conference on advances in robotics, pages 1-6, 2015. 2 +[19] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2 +[20] Binghao Huang, Yixuan Wang, Xinyi Yang, Yiyue Luo, and Yunzhu Li. 3d-vitac: Learning fine-grained manipulation + +with visuo-tactile sensing. arXiv preprint arXiv:2410.24091, 2024. 2 +[21] Kenneth H Hunt. Structural kinematics of in-parallel-actuated robot-arms. 1983. 1 +[22] Michael I Jordan and David E Rumelhart. Forward models: Supervised learning with a distal teacher. In Backpropagation, pages 189-236. Psychology Press, 2013. 3 +[23] Tsung-Wei Ke, Nikolaos Gkanatsios, and Katerina Fragkiadaki. 3d diffuser actor: Policy diffusion with 3d scene representations. arXiv preprint arXiv:2402.10885, 2024. 2, 3 +[24] Moo Jin Kim, Karl Pertsch, Siddharth Karamcheti, Ted Xiao, Ashwin Balakrishna, Suraj Nair, Rafael Rafailov, Ethan Foster, Grace Lam, Pannag Sanketi, et al. Openvla: An open-source vision-language-action model. arXiv preprint arXiv:2406.09246, 2024. 2 +[25] Thomas N Kipf and Max Welling. Semi-supervised classification with graph convolutional networks. arXiv preprint arXiv:1609.02907, 2016. 2 +[26] CS George Lee. Robot arm kinematics, dynamics, and control. Computer, 15(12):62-80, 1982. 1 +[27] Maolin Lei, Ting Wang, Chen Yao, Huan Liu, Zhi Wang, and Yongsheng Deng. Real-time kinematics-based self-collision avoidance algorithm for dual-arm robots. Applied Sciences, 10(17):5893, 2020. 1 +[28] Jiefeng Li, Chao Xu, Zhicun Chen, Siyuan Bian, Lixin Yang, and Cewu Lu. Hybrik: A hybrid analytical-neural inverse kinematics solution for 3d human pose and shape estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3383-3393, 2021. 3 +[29] Jiefeng Li, Chao Xu, Zhicun Chen, Siyuan Bian, Lixin Yang, and Cewu Lu. Hybrik: A hybrid analytical-neural inverse kinematics solution for 3d human pose and shape estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3383-3393, 2021. 3 +[30] I Liu, Chun Arthur, Sicheng He, Daniel Seita, and Gaurav Sukhatme. Voxact-b: Voxel-based acting and stabilizing policy for bimanual manipulation. arXiv preprint arXiv:2407.04152, 2024. 2 +[31] Xiao Ma, Sumit Patidar, Iain Haughton, and Stephen James. Hierarchical diffusion policy for kinematics-aware multitask robotic manipulation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18081-18090, 2024. 1, 2 +[32] Matthew T Mason. Mechanics of robotic manipulation. MIT press, 2001. 2 +[33] Yao Mu, Qinglong Zhang, Mengkang Hu, Wenhai Wang, Mingyu Ding, Jun Jin, Bin Wang, Jifeng Dai, Yu Qiao, and Ping Luo. Embodiedgpt: Vision-language pre-training via embodied chain of thought. Advances in Neural Information Processing Systems, 36, 2024. 1 +[34] Yoshihiko Nakamura, Kiyoshi Nagai, and Tsuneo Yoshikawa. Dynamics and stability in coordination of multiple robotic mechanisms. The International Journal of Robotics Research, 8(2):44-61, 1989. 1 +[35] Valerio Ortenzi, Naresh Marturi, Michael Mistry, Jeffrey Kuo, and Rustam Stolkin. Vision-based framework + +to estimate robot configuration and kinematic constraints. IEEE/ASME Transactions on Mechatronics, 23(5):2402-2412, 2018. 2 +[36] Eric Paljug, Xiaoping Yun, and Vijay Kumar. Control of rolling contacts in multi-arm manipulation. IEEE Transactions on Robotics and Automation, 10(4):441-452, 1994. 1 +[37] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4195-4205, 2023. 2 +[38] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 2 +[39] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 2 +[40] Allen Z Ren, Justin Lidard, Lars L Ankile, Anthony Simeonov, Pulkit Agrawal, Anirudha Majumdar, Benjamin Burchfiel, Hongkai Dai, and Max Simchowitz. Diffusion policy policy optimization. arXiv preprint arXiv:2409.00588, 2024. 2 +[41] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2 +[42] Nilanjan Sarkar, Xiaoping Yun, and Vijay Kumar. Dynamic control of 3-d rolling contacts in two-arm manipulation. IEEE Transactions on Robotics and Automation, 13(3): 364-376, 1997. 1 +[43] Youshik Shin and Zeungnam Bien. Collision-free trajectory planning for two robot arms. Robotica, 7(3):205-212, 1989. 1, 2 +[44] Mohit Shridhar, Lucas Manuelli, and Dieter Fox. Perceiver-actor: A multi-task transformer for robotic manipulation. In Conference on Robot Learning, pages 785–799. PMLR, 2023. 1, 2, 3, 6 +[45] Christian Smith, Yiannis Karayiannidis, Lazaros Nalpantidis, Xavi Gratal, Peng Qi, Dimos V Dimarogonas, and Danica Kragic. Dual arm manipulation—a survey. Robotics and Autonomous systems, 60(10):1340-1353, 2012. 1 +[46] Octo Model Team, Dibya Ghosh, Homer Walke, Karl Pertsch, Kevin Black, Oier Mees, Sudeep Dasari, Joel Hejna, Tobias Kreiman, Charles Xu, et al. Octo: An open-source generalist robot policy. arXiv preprint arXiv:2405.12213, 2024. 2 +[47] Vitalis Vosylius, Younggyo Seo, Jafar Uruç, and Stephen James. Render and diffuse: Aligning image and action spaces for diffusion-based behaviour cloning. arXiv preprint arXiv:2405.18196, 2024. 2 +[48] Lirui Wang, Xinlei Chen, Jialiang Zhao, and Kaiming He. Scaling proprioceptive-visual learning with heterogeneous pre-trained transformers. arXiv preprint arXiv:2409.20537, 2024. 2 + +[49] Yixiao Wang, Yifei Zhang, Mingxiao Huo, Ran Tian, Xiang Zhang, Yichen Xie, Chenfeng Xu, Pengliang Ji, Wei Zhan, Mingyu Ding, et al. Sparse diffusion policy: A sparse, reusable, and flexible policy for robot learning. arXiv preprint arXiv:2407.01531, 2024. 2 +[50] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. Advances in Neural Information Processing Systems, 36, 2024. 2 +[51] Chris Welman. Inverse kinematics and geometric constraints for articulated figure manipulation. 1993. 2 +[52] Kun Wu, Yichen Zhu, Jinming Li, Junjie Wen, Ning Liu, Zhiyuan Xu, Qinru Qiu, and Jian Tang. Discrete policy: Learning disentangled action space for multi-task robotic manipulation. arXiv preprint arXiv:2409.18707, 2024. 2 +[53] Wenke Xia, Dong Wang, Xincheng Pang, Zhigang Wang, Bin Zhao, Di Hu, and Xuelong Li. Kinematic-aware prompting for generalizable articulated object manipulation with llms. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 2073-2080. IEEE, 2024. 2 +[54] Andrea Maria Zanchettin, Paolo Rocco, Luca Bascetta, Ioannis Symeonidis, and Steffen Peldschus. Kinematic analysis and synthesis of the human arm motion during a manipulation task. In 2011 IEEE international conference on robotics and automation, pages 2692-2697. IEEE, 2011. 2 +[55] Yanjie Ze, Zixuan Chen, Wenhao Wang, Tianyi Chen, Xialin He, Ying Yuan, Xue Bin Peng, and Jiajun Wu. Generalizable humanoid manipulation with improved 3d diffusion policies. arXiv preprint arXiv:2410.10803, 2024. 1, 2 +[56] Yanjie Ze, Gu Zhang, Kangning Zhang, Chenyuan Hu, Muhan Wang, and Huazhe Xu. 3d diffusion policy. arXiv preprint arXiv:2403.03954, 2024. 1, 2, 6, 7 +[57] Junjie Zhang, Chenjia Bai, Haoran He, Wenke Xia, Zhigang Wang, Bin Zhao, Xiu Li, and Xuelong Li. Sam-e: Leveraging visual foundation model with sequence imitation for embodied manipulation. arXiv preprint arXiv:2405.19586, 2024. 2 +[58] Tony Z Zhao, Vikash Kumar, Sergey Levine, and Chelsea Finn. Learning fine-grained bimanual manipulation with low-cost hardware. arXiv preprint arXiv:2304.13705, 2023. 1, 2, 3, 6, 7 +[59] Zangwei Zheng, Xiangyu Peng, Tianji Yang, Chenhui Shen, Shenggui Li, Hongxin Liu, Yukun Zhou, Tianyi Li, and Yang You. Open-sora: Democratizing efficient video production for all, march 2024. URL https://github.com/hpcaitech/Open-Sora, 1(3):4.2 +[60] Minjie Zhu, Yichen Zhu, Jinming Li, Junjie Wen, Zhiyuan Xu, Ning Liu, Ran Cheng, Chaomin Shen, Yaxin Peng, Feifei Feng, et al. Scaling diffusion policy in transformer to 1 billion parameters for robotic manipulation. arXiv preprint arXiv:2409.14411, 2024. 2 \ No newline at end of file diff --git a/2025/Spatial-Temporal Graph Diffusion Policy with Kinematic Modeling for Bimanual Robotic Manipulation/images.zip b/2025/Spatial-Temporal Graph Diffusion Policy with Kinematic Modeling for Bimanual Robotic Manipulation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..e7b7cc4b6adaf7e6cf0c80b98a85b52d9dd3ce1f --- /dev/null +++ b/2025/Spatial-Temporal Graph Diffusion Policy with Kinematic Modeling for Bimanual Robotic Manipulation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3bf4db1985aa28edc53c6f555c5da3f84e88cd196c9d46c6bf657563ef875fb1 +size 606390 diff --git a/2025/Spatial-Temporal Graph Diffusion Policy with Kinematic Modeling for Bimanual Robotic Manipulation/layout.json b/2025/Spatial-Temporal Graph Diffusion Policy with Kinematic Modeling for Bimanual Robotic Manipulation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..0e856a136132bba1fc5f465aabef291c1540cef1 --- /dev/null +++ b/2025/Spatial-Temporal Graph Diffusion Policy with Kinematic Modeling for Bimanual Robotic Manipulation/layout.json @@ -0,0 +1,9946 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 87, + 102, + 523, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 87, + 102, + 523, + 140 + ], + "spans": [ + { + "bbox": [ + 87, + 102, + 523, + 140 + ], + "type": "text", + "content": "Spatial-Temporal Graph Diffusion Policy with Kinematic Modeling for Bimanual Robotic Manipulation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 102, + 159, + 507, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 159, + 507, + 190 + ], + "spans": [ + { + "bbox": [ + 102, + 159, + 507, + 190 + ], + "type": "text", + "content": "Qi Lv" + }, + { + "bbox": [ + 102, + 159, + 507, + 190 + ], + "type": "inline_equation", + "content": "^{1,2,3}" + }, + { + "bbox": [ + 102, + 159, + 507, + 190 + ], + "type": "text", + "content": " Hao Li" + }, + { + "bbox": [ + 102, + 159, + 507, + 190 + ], + "type": "inline_equation", + "content": "^{1,3}" + }, + { + "bbox": [ + 102, + 159, + 507, + 190 + ], + "type": "text", + "content": " Xiang Deng" + }, + { + "bbox": [ + 102, + 159, + 507, + 190 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 102, + 159, + 507, + 190 + ], + "type": "text", + "content": " Rui Shao" + }, + { + "bbox": [ + 102, + 159, + 507, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 102, + 159, + 507, + 190 + ], + "type": "text", + "content": " Yinchuan Li" + }, + { + "bbox": [ + 102, + 159, + 507, + 190 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 102, + 159, + 507, + 190 + ], + "type": "text", + "content": " Jianye Hao" + }, + { + "bbox": [ + 102, + 159, + 507, + 190 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 102, + 159, + 507, + 190 + ], + "type": "text", + "content": " Longxiang Gao" + }, + { + "bbox": [ + 102, + 159, + 507, + 190 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 102, + 159, + 507, + 190 + ], + "type": "text", + "content": " Michael Yu Wang" + }, + { + "bbox": [ + 102, + 159, + 507, + 190 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 102, + 159, + 507, + 190 + ], + "type": "text", + "content": " Liqiang Nie" + }, + { + "bbox": [ + 102, + 159, + 507, + 190 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 88, + 191, + 522, + 234 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 191, + 522, + 234 + ], + "spans": [ + { + "bbox": [ + 88, + 191, + 522, + 234 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 88, + 191, + 522, + 234 + ], + "type": "text", + "content": "School of Computer Science and Technology, Harbin Institute of Technology (Shenzhen) \n" + }, + { + "bbox": [ + 88, + 191, + 522, + 234 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 88, + 191, + 522, + 234 + ], + "type": "text", + "content": "School of Engineering, Greate Bay University \n" + }, + { + "bbox": [ + 88, + 191, + 522, + 234 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 88, + 191, + 522, + 234 + ], + "type": "text", + "content": "Huawei Noah's Ark Lab \n" + }, + { + "bbox": [ + 88, + 191, + 522, + 234 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 88, + 191, + 522, + 234 + ], + "type": "text", + "content": "Qilu University of Technology (Shandong Academy of Sciences)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 157, + 238, + 449, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 238, + 449, + 251 + ], + "spans": [ + { + "bbox": [ + 157, + 238, + 449, + 251 + ], + "type": "text", + "content": "lvqi@stu.hit.edu.cn, dengxiang@hit.edu.cn" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 151, + 290, + 200, + 303 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 290, + 200, + 303 + ], + "spans": [ + { + "bbox": [ + 151, + 290, + 200, + 303 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 316, + 297, + 627 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 316, + 297, + 627 + ], + "spans": [ + { + "bbox": [ + 54, + 316, + 297, + 627 + ], + "type": "text", + "content": "Despite the significant success of imitation learning in robotic manipulation, its application to bimanual tasks remains highly challenging. Existing approaches mainly learn a policy to predict a distant next-best end-effector pose (NBP) and then compute the corresponding joint rotation angles for motion using inverse kinematics. However, they suffer from two important issues: (1) rarely considering the physical robotic structure, which may cause self-collisions or interferences, and (2) overlooking the kinematics constraint, which may result in the predicted poses not conforming to the actual limitations of the robot joints. In this paper, we propose Kinematics enhanced Spatial-TemporAl gRaph Diffuser (KStar Diffuser). Specifically, (1) to incorporate the physical robot structure information into action prediction, KStar Diffuser maintains a dynamic spatial-temporal graph according to the physical bimanual joint motions at continuous timesteps. This dynamic graph serves as the robot-structure condition for denoising the actions; (2) to make the NBP learning objective consistent with kinematics, we introduce the differentiable kinematics to provide the reference for optimizing KStar Diffuser. This module regularizes the policy to predict more reliable and kinematics-aware next end-effector poses. Experimental results show that our method effectively leverages the physical structural information and generates kinematics-aware actions in both simulation and real-world." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 639, + 135, + 651 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 639, + 135, + 651 + ], + "spans": [ + { + "bbox": [ + 55, + 639, + 135, + 651 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 660, + 296, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 660, + 296, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 660, + 296, + 685 + ], + "type": "text", + "content": "Bimanual manipulation [34, 36, 42, 45] represents a fundamental capability for robotic systems to perform com" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 334, + 278, + 535, + 406 + ], + "blocks": [ + { + "bbox": [ + 334, + 278, + 535, + 406 + ], + "lines": [ + { + "bbox": [ + 334, + 278, + 535, + 406 + ], + "spans": [ + { + "bbox": [ + 334, + 278, + 535, + 406 + ], + "type": "image", + "image_path": "00add3a9ccda9ddd24e457da885fb141b35afbbd665f185f3aa0eac03b07d4e5.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 415, + 555, + 437 + ], + "lines": [ + { + "bbox": [ + 313, + 415, + 555, + 437 + ], + "spans": [ + { + "bbox": [ + 313, + 415, + 555, + 437 + ], + "type": "text", + "content": "Figure 1. The self-collision problem in bimanual manipulation tasks due to overlooking the robotic structure." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 312, + 449, + 555, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 449, + 555, + 533 + ], + "spans": [ + { + "bbox": [ + 312, + 449, + 555, + 533 + ], + "type": "text", + "content": "plex tasks requiring two-arm coordination. While imitation learning has demonstrated remarkable success in single-arm manipulation [6, 7, 33], its extension to bimanual scenarios faces unique challenges as robots need to coordinate dual-arm movements while conforming to physical constraints. These challenges significantly impact the reliability and feasibility of predicted actions in real-world applications." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 312, + 534, + 556, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 534, + 556, + 700 + ], + "spans": [ + { + "bbox": [ + 312, + 534, + 556, + 700 + ], + "type": "text", + "content": "A straightforward approach [9, 10, 13, 55, 56, 58] would be to learn the entire motion trajectory by predicting joint positions at consecutive timesteps. However, this poses significant challenges as trajectories are typically long and difficult to learn. Therefore, mainstream approaches [15, 17, 31, 44] typically adopt a two-stage pipeline: first predicting the next-best end-effector pose (NBP), then computing joint rotations through inverse kinematics [21, 26]. However, the dominant NBP-based approaches, though simplify the learning process, often lead to unreliable motion generation. Concretely, the predicted poses may violate physical structure constraints [27, 43], causing potential arm collisions, or exceeding joint limits due to inadequate consideration of kinematic feasibility, as Figure 1 illustrated." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 325, + 701, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 701, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 325, + 701, + 553, + 714 + ], + "type": "text", + "content": "In real-world bimanual manipulation tasks, existing ap" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 693, + 295, + 713 + ], + "type": "text", + "content": "*This work was done when Qi Lv and Hao Li were intern at Huawei Noah's Ark Lab. † Corresponding author" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "17394" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 312 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 312 + ], + "type": "text", + "content": "proaches [9, 15, 17, 44, 57] exhibit significant limitations in motion execution despite their promising performance in trajectory prediction. Empirical observations indicate that while individual end-effector poses are feasible in isolation, their concurrent execution frequently results in interarm collisions [43] and kinematically infeasible configurations [31, 35]. The fundamental challenge lies in the optimization of end-effector poses exclusively in Cartesian space, which introduces a substantial discrepancy between motion prediction and physical constraints. Existing approaches [9, 10, 17] incorporate proprioception information in a simplistic manner, (e.g., gripper rotation angles [17, 44] and end-effector poses [48]), while disregarding the robot's intrinsic structural properties e.g., the kinematic chain and joint configurations during the prediction phase. This oversimplified representation fails to capture the crucial spatial correlations between robotic arms, joint links, and their potential interactions. Consequently, the predicted dual-arm trajectories may satisfy task objectives in Cartesian space while violating crucial spatial constraints during execution." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 315, + 295, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 315, + 295, + 471 + ], + "spans": [ + { + "bbox": [ + 55, + 315, + 295, + 471 + ], + "type": "text", + "content": "Moreover, the conventional paradigm [17, 24, 44, 46] of treating inverse kinematics as a post-processing procedure introduces additional complexities. The disjunction between kinematic constraints and the pose prediction learning objective often yields solutions that exhibit apparent smoothness in Cartesian space while manifesting discontinuous or mechanically infeasible trajectories in the joint space. This limitation becomes particularly pronounced in configurations approaching kinematic singularities or joint limits, where the mapping between Cartesian and joint spaces becomes ill-conditioned. These inherent limitations necessitate the integration of both structural and kinematic constraints into the policy learning framework." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 474, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 474, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 474, + 295, + 713 + ], + "type": "text", + "content": "To address the above issues, we propose Kinematics enhanced Spatial-TemporAl gRaph Diffuser (KStar Diffuser), a novel framework that explicitly incorporates both robot structures and kinematics into the bimanual motion generation process. Our key insight is that the physical structure and kinematic properties of the robot should guide the learning process of pose prediction, rather than be treated as independent post-processing constraints. Specifically, (1) to incorporate structural awareness, we construct a dynamic spatial-temporal graph from the robot's URDF specifications, where nodes represent joint properties and edges capture both spatial relations and temporal dependencies. This graph structure is encoded via Graph Convolutional Network (GCN) [25] to provide explicit physical constraints for the diffusion process; (2) for kinematic feasibility, we regularize the NBP learning objective by incorporating joint-space prediction, where the predicted joint positions are mapped to reference end-effector poses through differentiable forward kinematics. These kinematically-feasible poses then serve as conditions" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 313, + 72, + 553, + 107 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 107 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 107 + ], + "type": "text", + "content": "to guide the diffusion process, ensuring that the generated motions satisfy both structural and kinematic constraints. Our main contributions are as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 108, + 553, + 277 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 314, + 108, + 553, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 108, + 553, + 168 + ], + "spans": [ + { + "bbox": [ + 314, + 108, + 553, + 168 + ], + "type": "text", + "content": "- Different from existing approaches that optimize end-effector poses solely in Cartesian space, we propose a novel spatial-temporal robot graph that explicitly models the robot physical configuration to guide the generative action denoising procedure." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 168, + 553, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 168, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 313, + 168, + 553, + 239 + ], + "type": "text", + "content": "- We introduce a kinematics regularizer that augments the NBP learning objective by introducing joint-space supervision. This regularizer leverages forward kinematics to provide kinematically-feasible reference poses, effectively guiding the diffusion process to conform to kinematic constraints." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 239, + 553, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 239, + 553, + 277 + ], + "spans": [ + { + "bbox": [ + 313, + 239, + 553, + 277 + ], + "type": "text", + "content": "- Extensive experiments show that our proposed KStar Diffuser is superior in both simulation and real-world scenarios, surpassing baselines more than " + }, + { + "bbox": [ + 313, + 239, + 553, + 277 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 313, + 239, + 553, + 277 + ], + "type": "text", + "content": " in success ratio." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 285, + 400, + 296 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 285, + 400, + 296 + ], + "spans": [ + { + "bbox": [ + 314, + 285, + 400, + 296 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 301, + 553, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 301, + 553, + 434 + ], + "spans": [ + { + "bbox": [ + 313, + 301, + 553, + 434 + ], + "type": "text", + "content": "Diffusion Models in Bimanual Robotic Controls. Diffusion models such as Denoising Diffusion Probabilistic Models (DDPM) [19] have achieved great success in the fields of image generation [38, 39, 41] and video generation [37, 50, 59]. Thus, recent work [8, 40, 47, 52, 60] has been devoted to applying its powerful generation capabilities to action generation for robotic manipulation tasks in a imitation learning mode. With the 2D images [9, 31, 40, 47, 49] or 3D point cloud observations [8, 20, 23, 55, 56], the policy is trained to output the action sequence of joint positions or end-effector poses by iterative denoising process." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 434, + 554, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 434, + 554, + 650 + ], + "spans": [ + { + "bbox": [ + 313, + 434, + 554, + 650 + ], + "type": "text", + "content": "Bimanual manipulation tasks, which resemble human-like actions more closely, have garnered increasing attention [10, 14, 17, 30, 58]. However, existing research primarily extends single-arm manipulation methods to dual-arm tasks [10, 58], without considering about the distinct challenges specific to bimanual settings. The relative independence of the two arms makes it essential to consider self-collision avoidance [1, 3] when predicting actions. In practice, the robot's structure significantly influences its motion since the joint types determine movement directions and joint angles constrain the range of motion. Although some studies [15, 44, 48] integrate proprioception in robotic control, they often use simply low-dimensional vectors to represent body information, lacking depth in structural exploration and overlooking critical spatial relationships between dual arms. Therefore, we study the physical structure information of the dual-arm robot system, and improve both the accuracy and adaptability of robotic movements." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "text", + "content": "Robotic Kinematics Modeling. Robotic kinematics modeling [2, 16, 31, 53, 54] is a classic problem in the robotic control [4, 18, 32, 51] where inverse kinematics (IK) presents a fundamental challenge, i.e., deriving the joint configuration given the end-effector pose. The inher" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17395" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 277 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 277 + ], + "type": "text", + "content": "ent non-uniqueness of IK solutions as a significant obstacle for learning algorithms [12, 22], often produce inaccurate models by averaging over nonconvex feasible sets. Recent advancements integrate neural solutions [11, 28, 29] to enhance the adaptability and precision of kinematic models. Bócsi et al. [5] leveraged support vector machines to parameterize quadratic programs, aligning solutions with IK in specific workspace regions. Current data-driven imitation learning approaches [10, 13, 15, 17, 58] primarily rely on probabilistic modeling of action distributions to predict subsequent actions, but lack the reliability guarantees from the aspect of kinematics. In this paper, we construct a spatial-temporal graph to learn the robot physical representation, and provide a kinematics-aware latent end-effect embedding for diffusion policy as guidance. In this way, o'ur approach enhances the robot's ability to perform tasks with superious precision and adaptability." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 290, + 111, + 303 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 290, + 111, + 303 + ], + "spans": [ + { + "bbox": [ + 55, + 290, + 111, + 303 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 311, + 135, + 324 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 311, + 135, + 324 + ], + "spans": [ + { + "bbox": [ + 55, + 311, + 135, + 324 + ], + "type": "text", + "content": "3.1. Preliminary" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 330, + 296, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 330, + 296, + 402 + ], + "spans": [ + { + "bbox": [ + 55, + 330, + 296, + 402 + ], + "type": "text", + "content": "Diffusion Policy. Chi et al. [9] propose the diffusion policy which represents a robot's visuomotor policy as a conditional denoising diffusion process. It learns a model distribution " + }, + { + "bbox": [ + 55, + 330, + 296, + 402 + ], + "type": "inline_equation", + "content": "\\pi_{\\phi}(\\pmb{a}_0|\\pmb{o})" + }, + { + "bbox": [ + 55, + 330, + 296, + 402 + ], + "type": "text", + "content": " conditioned on observation " + }, + { + "bbox": [ + 55, + 330, + 296, + 402 + ], + "type": "inline_equation", + "content": "\\pmb{o}" + }, + { + "bbox": [ + 55, + 330, + 296, + 402 + ], + "type": "text", + "content": " to approximate joint distribution " + }, + { + "bbox": [ + 55, + 330, + 296, + 402 + ], + "type": "inline_equation", + "content": "q(\\pmb{a}_0)" + }, + { + "bbox": [ + 55, + 330, + 296, + 402 + ], + "type": "text", + "content": ". The whole procedure consists of a forward process and a reverse process." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 403, + 296, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 403, + 296, + 440 + ], + "spans": [ + { + "bbox": [ + 55, + 403, + 296, + 440 + ], + "type": "text", + "content": "(1) Forward Process: For the Markov chain with Gaussian transitions parameterized, the policy " + }, + { + "bbox": [ + 55, + 403, + 296, + 440 + ], + "type": "inline_equation", + "content": "\\pi_{\\phi}(\\pmb{a}_0|\\pmb{o})" + }, + { + "bbox": [ + 55, + 403, + 296, + 440 + ], + "type": "text", + "content": " learns a fixed inference procedure " + }, + { + "bbox": [ + 55, + 403, + 296, + 440 + ], + "type": "inline_equation", + "content": "q(\\pmb{a}_k|\\pmb{a}_0)" + }, + { + "bbox": [ + 55, + 403, + 296, + 440 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 100, + 451, + 294, + 491 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 451, + 294, + 491 + ], + "spans": [ + { + "bbox": [ + 100, + 451, + 294, + 491 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} q \\left(\\boldsymbol {a} _ {k} \\mid \\boldsymbol {a} _ {0}\\right) = \\int q \\left(\\boldsymbol {a} _ {1: k} \\mid \\boldsymbol {a} _ {0}\\right) \\mathrm {d} \\boldsymbol {a} _ {1: (k - 1)} \\tag {1} \\\\ = \\mathcal {N} \\left(\\boldsymbol {a} _ {k}; \\sqrt {\\bar {\\alpha} _ {k}} \\boldsymbol {a} _ {0}, (1 - \\bar {\\alpha} _ {k}) \\boldsymbol {I}\\right) \\\\ \\end{array}", + "image_path": "d04638a0834348aaaacc2efc8415eecff43a20fbc895fab15d6e7818d632d2d8.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 502, + 296, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 502, + 296, + 525 + ], + "spans": [ + { + "bbox": [ + 55, + 502, + 296, + 525 + ], + "type": "text", + "content": "Thus, " + }, + { + "bbox": [ + 55, + 502, + 296, + 525 + ], + "type": "inline_equation", + "content": "a_{k}" + }, + { + "bbox": [ + 55, + 502, + 296, + 525 + ], + "type": "text", + "content": " can be expressed as a linear combination of " + }, + { + "bbox": [ + 55, + 502, + 296, + 525 + ], + "type": "inline_equation", + "content": "a_0" + }, + { + "bbox": [ + 55, + 502, + 296, + 525 + ], + "type": "text", + "content": " and a noise variable " + }, + { + "bbox": [ + 55, + 502, + 296, + 525 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 55, + 502, + 296, + 525 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 119, + 538, + 295, + 552 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 119, + 538, + 295, + 552 + ], + "spans": [ + { + "bbox": [ + 119, + 538, + 295, + 552 + ], + "type": "interline_equation", + "content": "\\boldsymbol {a} _ {k} = \\sqrt {\\bar {\\alpha} _ {k}} \\boldsymbol {a} _ {0} + \\sqrt {1 - \\bar {\\alpha} _ {k}} \\varepsilon \\tag {2}", + "image_path": "606698a35caa46deb7f742e472bb2217a9304849281b4de0a3bc2371b3009d0c.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 563, + 174, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 563, + 174, + 575 + ], + "spans": [ + { + "bbox": [ + 55, + 563, + 174, + 575 + ], + "type": "text", + "content": "The training loss is obtained:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 64, + 586, + 294, + 612 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 586, + 294, + 612 + ], + "spans": [ + { + "bbox": [ + 64, + 586, + 294, + 612 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathbb {E} _ {\\boldsymbol {a} _ {0} \\sim q (\\boldsymbol {a} _ {0}), \\epsilon_ {k} \\sim \\mathcal {N} (\\boldsymbol {0}, \\boldsymbol {I}), k} \\left[ \\left| \\left| \\epsilon_ {k}, \\pi_ {\\phi} (\\boldsymbol {a} _ {0} + \\epsilon_ {k}, \\boldsymbol {o}, k) \\right| \\right| ^ {2} \\right] \\tag {3}", + "image_path": "fe0060140df1f56a620220a096bc98a4ddb0c0e1f32ad685aad5e410d51514e6.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 613, + 270, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 613, + 270, + 624 + ], + "spans": [ + { + "bbox": [ + 55, + 613, + 270, + 624 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 613, + 270, + 624 + ], + "type": "inline_equation", + "content": "\\epsilon_{k}" + }, + { + "bbox": [ + 55, + 613, + 270, + 624 + ], + "type": "text", + "content": " is the random noise sampling at " + }, + { + "bbox": [ + 55, + 613, + 270, + 624 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 613, + 270, + 624 + ], + "type": "text", + "content": " iteration." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 625, + 296, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 625, + 296, + 650 + ], + "spans": [ + { + "bbox": [ + 55, + 625, + 296, + 650 + ], + "type": "text", + "content": "(2) Reverse Process: Starting from a sample " + }, + { + "bbox": [ + 55, + 625, + 296, + 650 + ], + "type": "inline_equation", + "content": "\\pmb{a}_K \\sim \\mathcal{N}(\\pmb{0}, \\pmb{I})" + }, + { + "bbox": [ + 55, + 625, + 296, + 650 + ], + "type": "text", + "content": ", the reverse steps are:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 82, + 661, + 295, + 675 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 661, + 295, + 675 + ], + "spans": [ + { + "bbox": [ + 82, + 661, + 295, + 675 + ], + "type": "interline_equation", + "content": "\\boldsymbol {a} _ {k - 1} = \\sqrt {\\bar {\\alpha} _ {k - 1}} \\pi_ {\\phi} (a _ {k}, k, o) + \\sqrt {1 - \\bar {\\alpha} _ {k - 1}} \\epsilon . \\tag {4}", + "image_path": "8632a78a47b93353119991a924ae89e7d7ac903bb06b3f9222be13877b211529.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 686, + 296, + 723 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 686, + 296, + 723 + ], + "spans": [ + { + "bbox": [ + 55, + 686, + 296, + 723 + ], + "type": "text", + "content": "By the iterative denoising process, the policy generates " + }, + { + "bbox": [ + 55, + 686, + 296, + 723 + ], + "type": "inline_equation", + "content": "a_0" + }, + { + "bbox": [ + 55, + 686, + 296, + 723 + ], + "type": "text", + "content": " as its next action. The continuous predicted action sequence forms the complete action trajectory." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 72, + 554, + 119 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 119 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 119 + ], + "type": "text", + "content": "Robotic Kinematics. Robotics kinematics describes the relationship between the robot joints and its end-effector. It can be divided into forward kinematics and inverse kinematics for controlling robot movements." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 120, + 555, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 120, + 555, + 180 + ], + "spans": [ + { + "bbox": [ + 313, + 120, + 555, + 180 + ], + "type": "text", + "content": "(1) Forward Kinematics (FK): Given a specific configuration of joint angles, i.e., " + }, + { + "bbox": [ + 313, + 120, + 555, + 180 + ], + "type": "inline_equation", + "content": "\\theta = [\\theta_{1},\\theta_{2},\\dots,\\theta_{n}]\\in \\Theta" + }, + { + "bbox": [ + 313, + 120, + 555, + 180 + ], + "type": "text", + "content": ", it aims to construct a function " + }, + { + "bbox": [ + 313, + 120, + 555, + 180 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{FK}}:\\mathbb{R}^n\\to SE(3)" + }, + { + "bbox": [ + 313, + 120, + 555, + 180 + ], + "type": "text", + "content": ", computing the end-effector position and orientation that are represented as a pose matrix " + }, + { + "bbox": [ + 313, + 120, + 555, + 180 + ], + "type": "inline_equation", + "content": "\\mathbf{T}\\in SE(3)" + }, + { + "bbox": [ + 313, + 120, + 555, + 180 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 339, + 185, + 555, + 205 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 185, + 555, + 205 + ], + "spans": [ + { + "bbox": [ + 339, + 185, + 555, + 205 + ], + "type": "interline_equation", + "content": "F K (\\theta) = \\mathbf {T} \\in \\mathcal {T}, \\text {w h e r e} \\mathbf {T} = \\prod_ {i = 1} ^ {n} \\mathbf {T} _ {i} (\\theta_ {i}), (5)", + "image_path": "03098c2a424fb985da734b7716f4ef380e34d2db5a8c2842ec7e961cb0bde5cd.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 212, + 554, + 259 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 212, + 554, + 259 + ], + "spans": [ + { + "bbox": [ + 313, + 212, + 554, + 259 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 212, + 554, + 259 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_i(\\theta_i)" + }, + { + "bbox": [ + 313, + 212, + 554, + 259 + ], + "type": "text", + "content": " is the homogeneous transformation matrix for joint " + }, + { + "bbox": [ + 313, + 212, + 554, + 259 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 212, + 554, + 259 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 212, + 554, + 259 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 313, + 212, + 554, + 259 + ], + "type": "text", + "content": " means the task space. " + }, + { + "bbox": [ + 313, + 212, + 554, + 259 + ], + "type": "inline_equation", + "content": "\\Theta \\in \\mathbb{R}^n" + }, + { + "bbox": [ + 313, + 212, + 554, + 259 + ], + "type": "text", + "content": " denotes the configuration space, while the configuration refers to the specific arrangement of the robotic joints in a given pose." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 259, + 554, + 320 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 259, + 554, + 320 + ], + "spans": [ + { + "bbox": [ + 313, + 259, + 554, + 320 + ], + "type": "text", + "content": "(2) Inverse Kinematics (IK): In contrast, IK aims to find the joint configurations which achieve a desired end-effector pose. It is defined as a mapping " + }, + { + "bbox": [ + 313, + 259, + 554, + 320 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{IK}}: SE(3) \\to \\mathbb{R}^n" + }, + { + "bbox": [ + 313, + 259, + 554, + 320 + ], + "type": "text", + "content": " from the task space " + }, + { + "bbox": [ + 313, + 259, + 554, + 320 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 313, + 259, + 554, + 320 + ], + "type": "text", + "content": " to the set of possible configurations " + }, + { + "bbox": [ + 313, + 259, + 554, + 320 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 313, + 259, + 554, + 320 + ], + "type": "text", + "content": " that satisfy the given pose matrix " + }, + { + "bbox": [ + 313, + 259, + 554, + 320 + ], + "type": "inline_equation", + "content": "\\mathbf{T}" + }, + { + "bbox": [ + 313, + 259, + 554, + 320 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 365, + 328, + 553, + 341 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 328, + 553, + 341 + ], + "spans": [ + { + "bbox": [ + 365, + 328, + 553, + 341 + ], + "type": "interline_equation", + "content": "I K (\\mathcal {T}) = \\{\\theta \\in \\Theta | F K (\\theta) = \\mathbf {T} \\}. \\tag {6}", + "image_path": "3d6cb97be7a7069eda5915f784fc560c11ca2527e7f8f3866b09342167412585.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 349, + 554, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 349, + 554, + 373 + ], + "spans": [ + { + "bbox": [ + 313, + 349, + 554, + 373 + ], + "type": "text", + "content": "It is noted that this mapping is often non-unique, especially for redundant manipulators with more than 6 DoF." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 380, + 408, + 392 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 380, + 408, + 392 + ], + "spans": [ + { + "bbox": [ + 313, + 380, + 408, + 392 + ], + "type": "text", + "content": "3.2. Task Definition" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 397, + 555, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 397, + 555, + 543 + ], + "spans": [ + { + "bbox": [ + 313, + 397, + 555, + 543 + ], + "type": "text", + "content": "Given the dataset " + }, + { + "bbox": [ + 313, + 397, + 555, + 543 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 313, + 397, + 555, + 543 + ], + "type": "text", + "content": " with language instructions " + }, + { + "bbox": [ + 313, + 397, + 555, + 543 + ], + "type": "inline_equation", + "content": "\\pmb{l}" + }, + { + "bbox": [ + 313, + 397, + 555, + 543 + ], + "type": "text", + "content": " and the RGB-D observations " + }, + { + "bbox": [ + 313, + 397, + 555, + 543 + ], + "type": "inline_equation", + "content": "\\pmb{o}" + }, + { + "bbox": [ + 313, + 397, + 555, + 543 + ], + "type": "text", + "content": ", we aim to learn a policy " + }, + { + "bbox": [ + 313, + 397, + 555, + 543 + ], + "type": "inline_equation", + "content": "\\pi_{\\phi}(\\pmb{a}| \\pmb{o}, \\pmb{l})" + }, + { + "bbox": [ + 313, + 397, + 555, + 543 + ], + "type": "text", + "content": " which predicts actions " + }, + { + "bbox": [ + 313, + 397, + 555, + 543 + ], + "type": "inline_equation", + "content": "\\pmb{a}" + }, + { + "bbox": [ + 313, + 397, + 555, + 543 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 313, + 397, + 555, + 543 + ], + "type": "inline_equation", + "content": "\\pmb{a}" + }, + { + "bbox": [ + 313, + 397, + 555, + 543 + ], + "type": "text", + "content": " is composed of a trajectory " + }, + { + "bbox": [ + 313, + 397, + 555, + 543 + ], + "type": "inline_equation", + "content": "\\pmb{a}^{\\mathrm{joint}} = \\{\\pmb{a}_0, \\pmb{a}_1, \\dots, \\pmb{a}_T\\}" + }, + { + "bbox": [ + 313, + 397, + 555, + 543 + ], + "type": "text", + "content": " and gripper opening or closing action " + }, + { + "bbox": [ + 313, + 397, + 555, + 543 + ], + "type": "inline_equation", + "content": "\\pmb{a}^{\\mathrm{gripper}}" + }, + { + "bbox": [ + 313, + 397, + 555, + 543 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 397, + 555, + 543 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 397, + 555, + 543 + ], + "type": "text", + "content": " denotes the trajectory length and " + }, + { + "bbox": [ + 313, + 397, + 555, + 543 + ], + "type": "inline_equation", + "content": "\\pmb{a}_i \\in \\mathbb{R}^n" + }, + { + "bbox": [ + 313, + 397, + 555, + 543 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 313, + 397, + 555, + 543 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 397, + 555, + 543 + ], + "type": "text", + "content": " represents the number of robot joints. In the bimanual manipulation, " + }, + { + "bbox": [ + 313, + 397, + 555, + 543 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 397, + 555, + 543 + ], + "type": "text", + "content": " is usually to be 12 or 14 as each robot arm has 6 or 7-DoF. Referring to prior works [17, 44], it is inefficient to train the policy on all trajectory points. Thus, a keyframe discovery method is used to extract a set of " + }, + { + "bbox": [ + 313, + 397, + 555, + 543 + ], + "type": "inline_equation", + "content": "K_{\\xi}" + }, + { + "bbox": [ + 313, + 397, + 555, + 543 + ], + "type": "text", + "content": " keyframe indices " + }, + { + "bbox": [ + 313, + 397, + 555, + 543 + ], + "type": "inline_equation", + "content": "\\{k_i\\}_{i=0}^{K_{\\xi}}" + }, + { + "bbox": [ + 313, + 397, + 555, + 543 + ], + "type": "text", + "content": " and the prediction actions are the key set of end-effector poses " + }, + { + "bbox": [ + 313, + 397, + 555, + 543 + ], + "type": "inline_equation", + "content": "\\pmb{a}^{\\mathrm{EE}}" + }, + { + "bbox": [ + 313, + 397, + 555, + 543 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 549, + 407, + 560 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 549, + 407, + 560 + ], + "spans": [ + { + "bbox": [ + 313, + 549, + 407, + 560 + ], + "type": "text", + "content": "3.3. KStar Diffuser" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 567, + 383, + 578 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 567, + 383, + 578 + ], + "spans": [ + { + "bbox": [ + 313, + 567, + 383, + 578 + ], + "type": "text", + "content": "3.3.1. Overview" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 582, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 555, + 714 + ], + "type": "text", + "content": "Mainstream methods [17, 23, 58] train a policy to predict actions, but take little consideration about the mechanical robot structure which determines its motion. We thus propose a spatial-temporal graph to model both the static physical structure and dynamic history movement information. In addition, to reduce kinematically infeasible predictions for the end-effector pose, we introduce a differentiable kinematics module that provides kinematics-aware references to the policy network. The overview of our proposed Kinematics enhanced Spatial-TemporAl gRaph Diffuser (KStar Diffuser) is shown in Figure 2." + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "17396" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 71, + 555, + 168 + ], + "blocks": [ + { + "bbox": [ + 59, + 71, + 555, + 168 + ], + "lines": [ + { + "bbox": [ + 59, + 71, + 555, + 168 + ], + "spans": [ + { + "bbox": [ + 59, + 71, + 555, + 168 + ], + "type": "image", + "image_path": "2bb8018e2225916c9c5a23ac6d9af49d89107a0d63faf94afc1e612f07d8fbfe.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 58, + 169, + 555, + 299 + ], + "blocks": [ + { + "bbox": [ + 58, + 169, + 555, + 299 + ], + "lines": [ + { + "bbox": [ + 58, + 169, + 555, + 299 + ], + "spans": [ + { + "bbox": [ + 58, + 169, + 555, + 299 + ], + "type": "image", + "image_path": "b1cddb061fecf0b7a76fe375fbb3848943c39b2a62c48ad7f571ecb82841926f.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 308, + 555, + 365 + ], + "lines": [ + { + "bbox": [ + 54, + 308, + 555, + 365 + ], + "spans": [ + { + "bbox": [ + 54, + 308, + 555, + 365 + ], + "type": "text", + "content": "Figure 2. Overview of KStar Diffuser. The top part presents the spatial-temporal graph which is constructed according to the robot architecture. The bottom part shows our backbone and the proposed kinematics regularizer. For the backbone, it extracts the multimodal information which consists of multiview RGB-D observations and language instruction, and then generates bimanual 6D end-effector poses. The kinematics regularizer enhances pose learning by incorporating joint-level predictions, which are mapped to reference end-effector poses through differentiable forward kinematics (FK)." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 378, + 126, + 389 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 378, + 126, + 389 + ], + "spans": [ + { + "bbox": [ + 55, + 378, + 126, + 389 + ], + "type": "text", + "content": "3.3.2. Backbone" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 393, + 295, + 465 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 393, + 295, + 465 + ], + "spans": [ + { + "bbox": [ + 54, + 393, + 295, + 465 + ], + "type": "text", + "content": "Given the language instruction " + }, + { + "bbox": [ + 54, + 393, + 295, + 465 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 54, + 393, + 295, + 465 + ], + "type": "text", + "content": " and multiview RGB-D observation images " + }, + { + "bbox": [ + 54, + 393, + 295, + 465 + ], + "type": "inline_equation", + "content": "o" + }, + { + "bbox": [ + 54, + 393, + 295, + 465 + ], + "type": "text", + "content": ", we first adopt the Transformer-based encoders to extract their features " + }, + { + "bbox": [ + 54, + 393, + 295, + 465 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_{\\mathrm{I}}" + }, + { + "bbox": [ + 54, + 393, + 295, + 465 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 393, + 295, + 465 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_{\\mathrm{T}}" + }, + { + "bbox": [ + 54, + 393, + 295, + 465 + ], + "type": "text", + "content": ", respectively. Then the features are fused with the " + }, + { + "bbox": [ + 54, + 393, + 295, + 465 + ], + "type": "inline_equation", + "content": "l_{\\mathrm{FiLM}}" + }, + { + "bbox": [ + 54, + 393, + 295, + 465 + ], + "type": "text", + "content": "-layers FiLM block to obtain the hidden states " + }, + { + "bbox": [ + 54, + 393, + 295, + 465 + ], + "type": "inline_equation", + "content": "\\mathbf{H}_{\\mathrm{B}}" + }, + { + "bbox": [ + 54, + 393, + 295, + 465 + ], + "type": "text", + "content": ". Each layer is combined with an up-sampling 2D convolution layer:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 129, + 471, + 294, + 485 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 471, + 294, + 485 + ], + "spans": [ + { + "bbox": [ + 129, + 471, + 294, + 485 + ], + "type": "interline_equation", + "content": "\\mathbf {H} _ {\\mathrm {B}} ^ {i} = \\operatorname {C o n v 2 D} \\left(\\mathbf {H} _ {\\mathrm {B}} ^ {i}\\right), \\tag {7}", + "image_path": "a811ac2b96af78394ad124c995ced5a88cca134df5628ee0e7314fd73a8c7fcd.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 120, + 488, + 294, + 503 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 120, + 488, + 294, + 503 + ], + "spans": [ + { + "bbox": [ + 120, + 488, + 294, + 503 + ], + "type": "interline_equation", + "content": "\\mathbf {H} _ {\\mathrm {B}} ^ {i + 1} = \\operatorname {F i L M} ^ {i} \\left(\\mathbf {H} _ {\\mathrm {B}} ^ {i}, \\mathbf {E} _ {\\mathrm {T}}\\right), \\tag {8}", + "image_path": "c59e8fea16503f70763f3ae98e4ef3b4a4a930ce06ccb9d83accb9f6dc4419b1.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 506, + 277, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 506, + 277, + 517 + ], + "spans": [ + { + "bbox": [ + 55, + 506, + 277, + 517 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 506, + 277, + 517 + ], + "type": "inline_equation", + "content": "i\\in [0,l_{\\mathrm{FiLM}})" + }, + { + "bbox": [ + 55, + 506, + 277, + 517 + ], + "type": "text", + "content": " and the " + }, + { + "bbox": [ + 55, + 506, + 277, + 517 + ], + "type": "inline_equation", + "content": "\\mathbf{H}_{\\mathrm{B}}^{0}" + }, + { + "bbox": [ + 55, + 506, + 277, + 517 + ], + "type": "text", + "content": " is initialized with " + }, + { + "bbox": [ + 55, + 506, + 277, + 517 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_{\\mathrm{I}}" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 518, + 296, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 518, + 296, + 613 + ], + "spans": [ + { + "bbox": [ + 55, + 518, + 296, + 613 + ], + "type": "text", + "content": "Our backbone " + }, + { + "bbox": [ + 55, + 518, + 296, + 613 + ], + "type": "inline_equation", + "content": "\\pi_{\\phi}" + }, + { + "bbox": [ + 55, + 518, + 296, + 613 + ], + "type": "text", + "content": " uses the last hidden states " + }, + { + "bbox": [ + 55, + 518, + 296, + 613 + ], + "type": "inline_equation", + "content": "\\mathbf{H}_{\\mathrm{B}}" + }, + { + "bbox": [ + 55, + 518, + 296, + 613 + ], + "type": "text", + "content": " as the condition to guide a diffusion head to denoise and generate the bimanual end-effector pose. It is noted that we attach the " + }, + { + "bbox": [ + 55, + 518, + 296, + 613 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 518, + 296, + 613 + ], + "type": "text", + "content": " historical observation images to provide more information for capturing the motion tendency. Following Chi et al. [9], we let the policy predict next " + }, + { + "bbox": [ + 55, + 518, + 296, + 613 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 55, + 518, + 296, + 613 + ], + "type": "text", + "content": " actions during training to alleviate multimodal problems. Here, we set both " + }, + { + "bbox": [ + 55, + 518, + 296, + 613 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 518, + 296, + 613 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 518, + 296, + 613 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 55, + 518, + 296, + 613 + ], + "type": "text", + "content": " to 2. The action prediction is as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 115, + 617, + 294, + 633 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 617, + 294, + 633 + ], + "spans": [ + { + "bbox": [ + 115, + 617, + 294, + 633 + ], + "type": "interline_equation", + "content": "\\boldsymbol {a} _ {k - 1} ^ {\\mathrm {E E}} \\sim \\pi_ {\\phi} \\left(\\boldsymbol {a} _ {k - 1} ^ {\\mathrm {E E}} \\mid \\boldsymbol {a} _ {k} ^ {\\mathrm {E E}}, \\mathbf {H} _ {\\mathrm {B}}\\right), \\tag {9}", + "image_path": "56b8135bba750413b49fa6386be3979261008794d88d20143b5dddee21055406.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 63, + 637, + 295, + 670 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 637, + 295, + 670 + ], + "spans": [ + { + "bbox": [ + 63, + 637, + 295, + 670 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\pi_ {\\phi} \\left(\\boldsymbol {a} _ {k - 1} ^ {\\mathrm {E E}} \\mid \\boldsymbol {a} _ {k} ^ {\\mathrm {E E}}, \\mathbf {H} _ {\\mathrm {B}}\\right) \\tag {10} \\\\ = \\mathcal {N} (\\boldsymbol {a} _ {k - 1} ^ {\\mathrm {E E}}; \\sqrt {\\bar {\\alpha} _ {k - 1}} \\pi_ {\\phi} (\\boldsymbol {a} _ {k} ^ {\\mathrm {E E}}, \\mathbf {H} _ {\\mathrm {B}}); (1 - \\bar {\\alpha} _ {k - 1}) I). \\\\ \\end{array}", + "image_path": "66dbae768d8b7f697d833947dbf36d40cf35307f0957cffe3caff27741b3c2d4.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 672, + 160, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 672, + 160, + 684 + ], + "spans": [ + { + "bbox": [ + 55, + 672, + 160, + 684 + ], + "type": "text", + "content": "The learning objective is:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 99, + 689, + 295, + 717 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 689, + 295, + 717 + ], + "spans": [ + { + "bbox": [ + 99, + 689, + 295, + 717 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {E E}} = \\mathbb {E} _ {q} \\left[ \\log \\frac {\\pi_ {\\phi} \\left(\\boldsymbol {a} _ {0 : K} ^ {\\mathrm {E E}} \\mid \\mathbf {H} _ {\\mathrm {B}}\\right)}{q \\left(\\boldsymbol {a} _ {1 : K} ^ {\\mathrm {E E}} \\mid \\boldsymbol {a} _ {0} ^ {\\mathrm {E E}}, \\mathbf {H} _ {\\mathrm {B}}\\right)} \\right]. \\tag {11}", + "image_path": "7f9b3f5e75bdf705744e2171385276feead4b230803ff815f290c45f7cc5fce6.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 378, + 476, + 391 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 378, + 476, + 391 + ], + "spans": [ + { + "bbox": [ + 313, + 378, + 476, + 391 + ], + "type": "text", + "content": "3.3.3. Spatial-Temporal Robot Graph" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 394, + 555, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 394, + 555, + 478 + ], + "spans": [ + { + "bbox": [ + 313, + 394, + 555, + 478 + ], + "type": "text", + "content": "The physical architecture impacts the motion of the whole robot, determining whether it can complete the task. Meanwhile, the historical spatial information is also important to future movements. Thus, we propose a spatial-temporal graph method to model the robot architecture at each step and the robot motion at continuous timesteps, representing the static spatial information and dynamic motion features." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 485, + 555, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 485, + 555, + 616 + ], + "spans": [ + { + "bbox": [ + 313, + 485, + 555, + 616 + ], + "type": "text", + "content": "Spatial Structure Graph Construction. To represent the robot structure, we first parse the Unified Robot Description Format (URDF) file which is usually used to describe the static physical structure of robot such as joint types, joint limits, and link lengths. Then, we define a dual-arm system as an undirected graph " + }, + { + "bbox": [ + 313, + 485, + 555, + 616 + ], + "type": "inline_equation", + "content": "G_{\\mathrm{s}} = \\langle V,E\\rangle" + }, + { + "bbox": [ + 313, + 485, + 555, + 616 + ], + "type": "text", + "content": " based on the joint and link configuration. Here, " + }, + { + "bbox": [ + 313, + 485, + 555, + 616 + ], + "type": "inline_equation", + "content": "V = \\{\\pmb{v}_{\\mathrm{j}_1},\\pmb{v}_{\\mathrm{j}_2},\\dots,\\pmb{v}_{\\mathrm{j}_m}\\}" + }, + { + "bbox": [ + 313, + 485, + 555, + 616 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 485, + 555, + 616 + ], + "type": "inline_equation", + "content": "E = \\{e_{ij}|(\\mathrm{joint}_i,\\mathrm{joint}_j)\\in \\mathrm{links}\\}" + }, + { + "bbox": [ + 313, + 485, + 555, + 616 + ], + "type": "text", + "content": " represents the node set of joints and edge set of links respectively, where " + }, + { + "bbox": [ + 313, + 485, + 555, + 616 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 485, + 555, + 616 + ], + "type": "text", + "content": " denotes the number of joint. We use " + }, + { + "bbox": [ + 313, + 485, + 555, + 616 + ], + "type": "inline_equation", + "content": "f_{i}\\in \\mathbb{R}^{D}" + }, + { + "bbox": [ + 313, + 485, + 555, + 616 + ], + "type": "text", + "content": " to denote the value of " + }, + { + "bbox": [ + 313, + 485, + 555, + 616 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 485, + 555, + 616 + ], + "type": "text", + "content": "-th node feature. It consists of three attributes as follows:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 618, + 554, + 714 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 314, + 618, + 554, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 618, + 554, + 677 + ], + "spans": [ + { + "bbox": [ + 314, + 618, + 554, + 677 + ], + "type": "text", + "content": "- Joint Coordinate: we use a vector in Cartesian coordinate system, " + }, + { + "bbox": [ + 314, + 618, + 554, + 677 + ], + "type": "inline_equation", + "content": "\\pmb{f}_i^{\\mathrm{JC}} = [\\pmb{x}_i, \\pmb{y}_i, \\pmb{z}_i] \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 314, + 618, + 554, + 677 + ], + "type": "text", + "content": ", to denote the absolute coordinates of the " + }, + { + "bbox": [ + 314, + 618, + 554, + 677 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 314, + 618, + 554, + 677 + ], + "type": "text", + "content": "-th joint. The vector is normalized according to the workspace boundary for stable convergence of model training." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 677, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 553, + 714 + ], + "type": "text", + "content": "- Joint Distance: To measure the spatial relationship between the node " + }, + { + "bbox": [ + 313, + 677, + 553, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_i" + }, + { + "bbox": [ + 313, + 677, + 553, + 714 + ], + "type": "text", + "content": " and the other nodes " + }, + { + "bbox": [ + 313, + 677, + 553, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_j" + }, + { + "bbox": [ + 313, + 677, + 553, + 714 + ], + "type": "text", + "content": ", we compute the Euclidean Distance between " + }, + { + "bbox": [ + 313, + 677, + 553, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_i" + }, + { + "bbox": [ + 313, + 677, + 553, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 677, + 553, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_j" + }, + { + "bbox": [ + 313, + 677, + 553, + 714 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "17397" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 137, + 70, + 296, + 86 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 70, + 296, + 86 + ], + "spans": [ + { + "bbox": [ + 137, + 70, + 296, + 86 + ], + "type": "interline_equation", + "content": "\\boldsymbol {f} _ {i} ^ {\\mathrm {J D}} = \\left| \\left| \\boldsymbol {v} _ {i} - \\boldsymbol {v} _ {j} \\right| \\right| _ {2}, \\tag {12}", + "image_path": "78f877ae5234e1d2b22ebf72bb2c657139f009429ee769d9e32063db833ddcdb.jpg" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 63, + 91, + 291, + 103 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 91, + 291, + 103 + ], + "spans": [ + { + "bbox": [ + 63, + 91, + 291, + 103 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 63, + 91, + 291, + 103 + ], + "type": "inline_equation", + "content": "\\pmb{f}_i^{\\mathrm{JD}}\\in \\mathbb{R}^m" + }, + { + "bbox": [ + 63, + 91, + 291, + 103 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 63, + 91, + 291, + 103 + ], + "type": "inline_equation", + "content": "||\\cdot ||_2" + }, + { + "bbox": [ + 63, + 91, + 291, + 103 + ], + "type": "text", + "content": " means the Euclidean norm." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 104, + 296, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 104, + 296, + 152 + ], + "spans": [ + { + "bbox": [ + 55, + 104, + 296, + 152 + ], + "type": "text", + "content": "- Body Label: To discriminate the source of node " + }, + { + "bbox": [ + 55, + 104, + 296, + 152 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_i" + }, + { + "bbox": [ + 55, + 104, + 296, + 152 + ], + "type": "text", + "content": ", we use a one-hot vector " + }, + { + "bbox": [ + 55, + 104, + 296, + 152 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_i^{\\mathrm{BL}} = [0,1]" + }, + { + "bbox": [ + 55, + 104, + 296, + 152 + ], + "type": "text", + "content": " as one of its features. It can also help the policy capture the motion modes of different robot arms such as the symmetry." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 152, + 296, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 152, + 296, + 190 + ], + "spans": [ + { + "bbox": [ + 55, + 152, + 296, + 190 + ], + "type": "text", + "content": "We concatenate " + }, + { + "bbox": [ + 55, + 152, + 296, + 190 + ], + "type": "inline_equation", + "content": "\\pmb{f}_i^{\\mathrm{JC}}" + }, + { + "bbox": [ + 55, + 152, + 296, + 190 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 152, + 296, + 190 + ], + "type": "inline_equation", + "content": "\\pmb{f}_i^{\\mathrm{JD}}" + }, + { + "bbox": [ + 55, + 152, + 296, + 190 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 152, + 296, + 190 + ], + "type": "inline_equation", + "content": "\\pmb{f}_i^{\\mathrm{BL}}" + }, + { + "bbox": [ + 55, + 152, + 296, + 190 + ], + "type": "text", + "content": " to form the completed feature " + }, + { + "bbox": [ + 55, + 152, + 296, + 190 + ], + "type": "inline_equation", + "content": "\\pmb{f}_i \\in \\mathbb{R}^{D_f}" + }, + { + "bbox": [ + 55, + 152, + 296, + 190 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 152, + 296, + 190 + ], + "type": "inline_equation", + "content": "D_{f}" + }, + { + "bbox": [ + 55, + 152, + 296, + 190 + ], + "type": "text", + "content": " denotes the dimension of node features." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 196, + 296, + 340 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 196, + 296, + 340 + ], + "spans": [ + { + "bbox": [ + 55, + 196, + 296, + 340 + ], + "type": "text", + "content": "Spatial-Temporal Graph Learning. Given the same instruction and observation, different historical robot poses result in different predictions. Thus, we build the spatial structure graph with the temporal motion information. Specifically, we build a spatial-temporal graph " + }, + { + "bbox": [ + 55, + 196, + 296, + 340 + ], + "type": "inline_equation", + "content": "G_{\\mathrm{ST}} = \\langle V_{\\mathrm{ST}}, E_{\\mathrm{ST}} \\rangle" + }, + { + "bbox": [ + 55, + 196, + 296, + 340 + ], + "type": "text", + "content": " by combining " + }, + { + "bbox": [ + 55, + 196, + 296, + 340 + ], + "type": "inline_equation", + "content": "\\{G_{\\mathrm{S}}^{i} = \\langle V_{\\mathrm{S}}^{i}, E_{\\mathrm{S}}^{i} \\rangle \\}_{i=0}^{T-1}" + }, + { + "bbox": [ + 55, + 196, + 296, + 340 + ], + "type": "text", + "content": " from historical timesteps, where " + }, + { + "bbox": [ + 55, + 196, + 296, + 340 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 196, + 296, + 340 + ], + "type": "text", + "content": " denotes the number of history steps. In " + }, + { + "bbox": [ + 55, + 196, + 296, + 340 + ], + "type": "inline_equation", + "content": "G_{\\mathrm{ST}}" + }, + { + "bbox": [ + 55, + 196, + 296, + 340 + ], + "type": "text", + "content": ", the node set " + }, + { + "bbox": [ + 55, + 196, + 296, + 340 + ], + "type": "inline_equation", + "content": "V_{\\mathrm{ST}}" + }, + { + "bbox": [ + 55, + 196, + 296, + 340 + ], + "type": "text", + "content": " contains " + }, + { + "bbox": [ + 55, + 196, + 296, + 340 + ], + "type": "inline_equation", + "content": "V_{\\mathrm{S}}^{i}" + }, + { + "bbox": [ + 55, + 196, + 296, + 340 + ], + "type": "text", + "content": " from historical static spatial graph. Additionally, we add edges " + }, + { + "bbox": [ + 55, + 196, + 296, + 340 + ], + "type": "inline_equation", + "content": "E'" + }, + { + "bbox": [ + 55, + 196, + 296, + 340 + ], + "type": "text", + "content": " linking the same joint node " + }, + { + "bbox": [ + 55, + 196, + 296, + 340 + ], + "type": "inline_equation", + "content": "\\text{joint}_i^t" + }, + { + "bbox": [ + 55, + 196, + 296, + 340 + ], + "type": "text", + "content": " at different timesteps, in order to establish the correlation of joint motions at continuous times. It can be formulated as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 73, + 347, + 295, + 363 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 347, + 295, + 363 + ], + "spans": [ + { + "bbox": [ + 73, + 347, + 295, + 363 + ], + "type": "interline_equation", + "content": "V _ {\\mathrm {S T}} = \\cup \\left\\{V _ {\\mathrm {S}} ^ {i} \\right\\} _ {i = 0} ^ {T - 1}, E _ {\\mathrm {S T}} = \\cup \\left\\{E _ {\\mathrm {S}} ^ {i} \\right\\} _ {i = 0} ^ {T - 1} \\cup E ^ {\\prime}, \\tag {13}", + "image_path": "9db081944879d7460526095fd79cf32f31c4e743e6210af342e1324fbfef5dfb.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 91, + 371, + 294, + 386 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 371, + 294, + 386 + ], + "spans": [ + { + "bbox": [ + 91, + 371, + 294, + 386 + ], + "type": "interline_equation", + "content": "E ^ {\\prime} = \\left\\{\\boldsymbol {e} _ {i} ^ {t, t ^ {\\prime}} \\mid t, t ^ {\\prime} \\in \\{0, 1, \\dots , T \\}, t \\neq t ^ {\\prime} \\right\\}. \\tag {14}", + "image_path": "93e997ad6f188aa73296f9ae9a15391b08110d3372a6d7b1c11f5054e9d124cd.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 391, + 296, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 391, + 296, + 498 + ], + "spans": [ + { + "bbox": [ + 55, + 391, + 296, + 498 + ], + "type": "text", + "content": "In this way, we obtain the whole spatial-temporal graph. A Graph Convolutional Network (GCN) is then adopted to propagate and aggregate node features across the graph. The GCN layer updates each node feature " + }, + { + "bbox": [ + 55, + 391, + 296, + 498 + ], + "type": "inline_equation", + "content": "f_{i}" + }, + { + "bbox": [ + 55, + 391, + 296, + 498 + ], + "type": "text", + "content": " by aggregating the features of its neighboring nodes, thereby capturing the relational and structural information of the robotic arms. We use the node feature " + }, + { + "bbox": [ + 55, + 391, + 296, + 498 + ], + "type": "inline_equation", + "content": "\\mathbf{H}_{\\mathrm{G}}" + }, + { + "bbox": [ + 55, + 391, + 296, + 498 + ], + "type": "text", + "content": " of the last encoder layer as the representation of the robot structure to condition the denoising process." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 504, + 185, + 516 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 504, + 185, + 516 + ], + "spans": [ + { + "bbox": [ + 55, + 504, + 185, + 516 + ], + "type": "text", + "content": "3.3.4. Kinematics Regularizer" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 519, + 296, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 519, + 296, + 674 + ], + "spans": [ + { + "bbox": [ + 55, + 519, + 296, + 674 + ], + "type": "text", + "content": "To control the end-effector effectively, the generated pose trajectory must be processed by an Inverse Kinematics (IK) solver, which calculates the joint configurations to achieve the specified poses. However, because the predicted trajectory is generated without considering the robot kinematic constraints, it often falls outside the IK solver's feasible range, resulting in high failure rates during execution. To address this limitation, we propose a kinematics regularizer into the end-effector pose learning objective. This regularizer aligns the predicted poses with the robot kinematic constraints, ensuring that the generated trajectory remains within the solvable space of the IK solver, thus enhancing then reliability of trajectory execution." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": "Differentiable Kinematics. Given a joint configuration " + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\Theta = [\\theta_1, \\theta_2, \\dots, \\theta_n]" + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": ", the corresponding end-effector pose " + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{T} \\in SE(3)" + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": " can be computed using forward kinematics," + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "text", + "content": "represented as a mapping " + }, + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "inline_equation", + "content": "f(\\Theta): \\mathbb{R}^n \\to SE(3)" + }, + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "text", + "content": ". This mapping from joint space to end-effector space is differentiable, namely Differentiable Forward Kinematics (DFK), enabling the use of gradients to optimize our control policy. Leveraging DFK, our policy learns to predict the next joint configuration " + }, + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "inline_equation", + "content": "\\hat{a}^{\\mathrm{joint}}" + }, + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "text", + "content": ", from which we compute an intermediate end-effector pose " + }, + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "inline_equation", + "content": "\\mathbf{H}_{\\mathrm{R}}" + }, + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "text", + "content": ". By using " + }, + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "inline_equation", + "content": "\\mathbf{H}_{\\mathrm{R}}" + }, + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "text", + "content": " as a reference, we guide a denoising process to generate a precise and executable end-effector pose." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 180, + 553, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 180, + 553, + 216 + ], + "spans": [ + { + "bbox": [ + 313, + 180, + 553, + 216 + ], + "type": "text", + "content": "Specifically, we combine structure features " + }, + { + "bbox": [ + 313, + 180, + 553, + 216 + ], + "type": "inline_equation", + "content": "\\mathbf{H}_{\\mathrm{G}}" + }, + { + "bbox": [ + 313, + 180, + 553, + 216 + ], + "type": "text", + "content": " with the last hidden state " + }, + { + "bbox": [ + 313, + 180, + 553, + 216 + ], + "type": "inline_equation", + "content": "\\mathbf{H}_{\\mathrm{B}}" + }, + { + "bbox": [ + 313, + 180, + 553, + 216 + ], + "type": "text", + "content": ", projecting to the joint space and using the DFK to obtain reference as " + }, + { + "bbox": [ + 313, + 180, + 553, + 216 + ], + "type": "inline_equation", + "content": "\\mathbf{H}_{\\mathrm{R}}" + }, + { + "bbox": [ + 313, + 180, + 553, + 216 + ], + "type": "text", + "content": " follows:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 325, + 224, + 553, + 239 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 224, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 325, + 224, + 553, + 239 + ], + "type": "interline_equation", + "content": "\\hat {\\boldsymbol {a}} ^ {\\text {j o i n t}} = \\operatorname {P r o j} \\left(\\left[ \\mathbf {H} _ {\\mathrm {B}}, \\mathbf {H} _ {\\mathrm {G}} \\right]\\right), \\quad \\mathbf {H} _ {\\mathrm {R}} = \\operatorname {D F K} \\left(\\hat {\\boldsymbol {a}} ^ {\\text {j o i n t}}\\right). \\tag {15}", + "image_path": "253bfdc7c8eda8946ddfedf3ffc729206c807d4ba22b8e214cc18b5d550c1334.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 248, + 553, + 273 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 248, + 553, + 273 + ], + "spans": [ + { + "bbox": [ + 313, + 248, + 553, + 273 + ], + "type": "text", + "content": "To ensure the consistency between predicted and actual joint angles, we minimize the joint loss:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 335, + 281, + 553, + 301 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 281, + 553, + 301 + ], + "spans": [ + { + "bbox": [ + 335, + 281, + 553, + 301 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {j o i n t}} = \\mathbb {E} _ {\\boldsymbol {a} ^ {\\text {j o i n t}} \\sim q (\\boldsymbol {a} _ {0})} \\left[ \\left| \\left| \\boldsymbol {a} ^ {\\text {j o i n t}} - \\hat {\\boldsymbol {a}} ^ {\\text {j o i n t}} \\right| \\right| ^ {2} \\right]. \\tag {16}", + "image_path": "a41b767306dfa0a467a1c847eec72b035f80cf372fd2e00fcfe53c15799673da.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 311, + 553, + 383 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 311, + 553, + 383 + ], + "spans": [ + { + "bbox": [ + 313, + 311, + 553, + 383 + ], + "type": "text", + "content": "Conditioning Diffusion Process on Kinematics. To enforce kinematic consistency, we condition the diffusion process on the joint angle representation " + }, + { + "bbox": [ + 313, + 311, + 553, + 383 + ], + "type": "inline_equation", + "content": "\\mathbf{H}_{\\mathrm{R}}" + }, + { + "bbox": [ + 313, + 311, + 553, + 383 + ], + "type": "text", + "content": ", an auxiliary input encoding kinematic constraints. This allows the predicted pose trajectory to stay within feasible space. Given the diffusion steps from Eq.(9) and Eq.(10), we have:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 364, + 391, + 553, + 406 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 364, + 391, + 553, + 406 + ], + "spans": [ + { + "bbox": [ + 364, + 391, + 553, + 406 + ], + "type": "interline_equation", + "content": "\\boldsymbol {a} _ {k - 1} ^ {\\mathrm {E E}} \\sim \\pi_ {\\phi} \\left(\\boldsymbol {a} _ {k} ^ {\\mathrm {E E}} \\mid \\boldsymbol {a} _ {k - 1} ^ {\\mathrm {E E}}, \\mathbf {H} _ {\\mathrm {B}}, \\mathbf {H} _ {\\mathrm {R}}\\right), \\tag {17}", + "image_path": "de12a27cb1f1bb261daee2366c2b8d02c5635b2fb010741d8ebda33e88272886.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 326, + 415, + 440, + 430 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 415, + 440, + 430 + ], + "spans": [ + { + "bbox": [ + 326, + 415, + 440, + 430 + ], + "type": "interline_equation", + "content": "\\pi_ {\\phi} \\left(\\boldsymbol {a} _ {k - 1} ^ {\\mathrm {E E}} \\mid \\boldsymbol {a} _ {k} ^ {\\mathrm {E E}}, \\mathbf {H} _ {\\mathrm {B}}, \\mathbf {H} _ {\\mathrm {R}}\\right) =", + "image_path": "baa2e142f0823197d22d0b45d42dc6ea10d6253f7155e641fb9f2b71ebaef56b.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 326, + 432, + 553, + 459 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 432, + 553, + 459 + ], + "spans": [ + { + "bbox": [ + 326, + 432, + 553, + 459 + ], + "type": "interline_equation", + "content": "\\mathcal {N} \\left(\\boldsymbol {a} _ {k - 1} ^ {\\mathrm {E E}}; \\sqrt {\\bar {\\alpha} _ {k - 1}} \\pi_ {\\phi} \\left(\\boldsymbol {a} _ {k} ^ {\\mathrm {E E}}, \\mathbf {H} _ {\\mathrm {B}}, \\mathbf {H} _ {\\mathrm {R}}\\right); (1 - \\bar {\\alpha} _ {k - 1}) \\boldsymbol {I}\\right). \\tag {18}", + "image_path": "0b4258c50e165389e075f2855d3c674475526c7f2da75e8175342b3be02d6455.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 464, + 553, + 524 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 464, + 553, + 524 + ], + "spans": [ + { + "bbox": [ + 313, + 464, + 553, + 524 + ], + "type": "text", + "content": "Incorporating DFK into the diffusion process allows gradients from the pose loss to propagate back through the kinematic function, ensuring that each denoising step maintains compliance with joint constraints, thereby optimizing the end-effector's control accuracy and robustness." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 532, + 444, + 544 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 532, + 444, + 544 + ], + "spans": [ + { + "bbox": [ + 313, + 532, + 444, + 544 + ], + "type": "text", + "content": "3.4. Training and Inference" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 550, + 553, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 550, + 553, + 597 + ], + "spans": [ + { + "bbox": [ + 313, + 550, + 553, + 597 + ], + "type": "text", + "content": "Training. We use the conditional action generation mode to train KStar Diffuser, which is formulated as a conditional denoising diffusion. The loss function is defined as the Mean Square Error (MSE) as follows:" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 375, + 604, + 553, + 618 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 375, + 604, + 553, + 618 + ], + "spans": [ + { + "bbox": [ + 375, + 604, + 553, + 618 + ], + "type": "interline_equation", + "content": "\\boldsymbol {a} _ {k} = \\sqrt {\\bar {\\alpha} _ {k}} \\boldsymbol {a} _ {0} + \\sqrt {1 - \\bar {\\alpha} _ {k}} \\epsilon , \\tag {19}", + "image_path": "71dee5498f62d7f9a5618921ec6ba0a5d579654109614f7194e094be4476db30.jpg" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 323, + 625, + 553, + 642 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 625, + 553, + 642 + ], + "spans": [ + { + "bbox": [ + 323, + 625, + 553, + 642 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {E E}} = \\mathbb {E} _ {\\boldsymbol {a} _ {0} ^ {\\mathrm {E E}} \\sim q (\\boldsymbol {a} _ {0}), k} \\left[ \\left| \\left| \\boldsymbol {a} _ {0} ^ {\\mathrm {E E}} - \\pi_ {\\phi} (\\boldsymbol {a} _ {k} ^ {\\mathrm {E E}}, k, C) \\right| \\right| ^ {2} \\right], \\tag {20}", + "image_path": "34eb57a8118747926de320eabec727f185ff35a07ac497946e3c608aab6e20ff.jpg" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 323, + 647, + 553, + 665 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 647, + 553, + 665 + ], + "spans": [ + { + "bbox": [ + 323, + 647, + 553, + 665 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {j o i n t}} = \\mathbb {E} _ {\\boldsymbol {a} _ {0} ^ {\\text {j o i n t}} \\sim q (\\boldsymbol {a} _ {0})} [ \\| \\boldsymbol {a} _ {0} ^ {\\text {j o i n t}} - \\pi_ {\\phi} (\\mathbf {H} _ {\\mathrm {B}}, \\mathbf {H} _ {\\mathrm {G}}) \\| ^ {2} ], \\tag {21}", + "image_path": "95789b6355c6472dcf0ee62bbd6432606086333ddb9d6c50a4237e78f2042d0b.jpg" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 375, + 670, + 553, + 683 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 375, + 670, + 553, + 683 + ], + "spans": [ + { + "bbox": [ + 375, + 670, + 553, + 683 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\lambda \\mathcal {L} _ {\\mathrm {E E}} + (1 - \\lambda) \\mathcal {L} _ {\\text {j o i n t}}, \\tag {22}", + "image_path": "8b1ca27c6f49a9fa1a73afd0742621a7c8c525247df96516d6ebd09da1465904.jpg" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 689, + 555, + 723 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 723 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 723 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 689, + 555, + 723 + ], + "type": "inline_equation", + "content": "\\mathbf{a}_k" + }, + { + "bbox": [ + 313, + 689, + 555, + 723 + ], + "type": "text", + "content": " is obtained by the forward diffusion process and " + }, + { + "bbox": [ + 313, + 689, + 555, + 723 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 313, + 689, + 555, + 723 + ], + "type": "text", + "content": " is the combination of " + }, + { + "bbox": [ + 313, + 689, + 555, + 723 + ], + "type": "inline_equation", + "content": "\\mathbf{H}_{\\mathrm{B}}" + }, + { + "bbox": [ + 313, + 689, + 555, + 723 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 689, + 555, + 723 + ], + "type": "inline_equation", + "content": "\\mathbf{H}_{\\mathrm{G}}" + }, + { + "bbox": [ + 313, + 689, + 555, + 723 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 689, + 555, + 723 + ], + "type": "inline_equation", + "content": "\\mathbf{H}_{\\mathrm{R}}" + }, + { + "bbox": [ + 313, + 689, + 555, + 723 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 313, + 689, + 555, + 723 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 689, + 555, + 723 + ], + "type": "text", + "content": " is the trade-off coefficient." + } + ] + } + ], + "index": 27 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "17398" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 108 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 108 + ], + "type": "text", + "content": "Inference. Sampling from a Gaussian noise " + }, + { + "bbox": [ + 55, + 72, + 295, + 108 + ], + "type": "inline_equation", + "content": "\\epsilon_{k}" + }, + { + "bbox": [ + 55, + 72, + 295, + 108 + ], + "type": "text", + "content": ", the policy " + }, + { + "bbox": [ + 55, + 72, + 295, + 108 + ], + "type": "inline_equation", + "content": "\\pi_{\\phi}" + }, + { + "bbox": [ + 55, + 72, + 295, + 108 + ], + "type": "text", + "content": " performs " + }, + { + "bbox": [ + 55, + 72, + 295, + 108 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 72, + 295, + 108 + ], + "type": "text", + "content": " iterations to gradually denoise a random noise " + }, + { + "bbox": [ + 55, + 72, + 295, + 108 + ], + "type": "inline_equation", + "content": "\\epsilon_{k}" + }, + { + "bbox": [ + 55, + 72, + 295, + 108 + ], + "type": "text", + "content": " into the noise-free action " + }, + { + "bbox": [ + 55, + 72, + 295, + 108 + ], + "type": "inline_equation", + "content": "\\pmb{a}_{0}" + }, + { + "bbox": [ + 55, + 72, + 295, + 108 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 65, + 113, + 295, + 129 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 113, + 295, + 129 + ], + "spans": [ + { + "bbox": [ + 65, + 113, + 295, + 129 + ], + "type": "interline_equation", + "content": "\\boldsymbol {a} _ {k - 1} ^ {\\mathrm {E E}} = \\sqrt {\\overline {{\\alpha}} _ {k - 1}} \\pi_ {\\phi} (\\boldsymbol {a} _ {k} ^ {\\mathrm {E E}}, k, C) + \\sqrt {1 - \\overline {{\\alpha}} _ {k - 1}} \\epsilon_ {k}. \\tag {23}", + "image_path": "1ee3efe04621ce996a9df8d530db3d4bfeaa696899d43e45e06fac2a49da9f2f.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 133, + 295, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 133, + 295, + 157 + ], + "spans": [ + { + "bbox": [ + 55, + 133, + 295, + 157 + ], + "type": "text", + "content": "We use the predicted " + }, + { + "bbox": [ + 55, + 133, + 295, + 157 + ], + "type": "inline_equation", + "content": "a_0^{\\mathrm{EE}}" + }, + { + "bbox": [ + 55, + 133, + 295, + 157 + ], + "type": "text", + "content": " as the final action to control robot execution." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 167, + 132, + 180 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 167, + 132, + 180 + ], + "spans": [ + { + "bbox": [ + 55, + 167, + 132, + 180 + ], + "type": "text", + "content": "4. Experiment" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 186, + 228, + 199 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 186, + 228, + 199 + ], + "spans": [ + { + "bbox": [ + 55, + 186, + 228, + 199 + ], + "type": "text", + "content": "4.1. Dataset and Evaluation Settings" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 204, + 296, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 204, + 296, + 323 + ], + "spans": [ + { + "bbox": [ + 55, + 204, + 296, + 323 + ], + "type": "text", + "content": "Dataset. Bimanual manipulation tasks demand high levels of coordination, synchronization, and symmetry awareness between the two robotic arms, making them inherently more challenging than single-arm tasks. To assess the capabilities of KStar Diffuser in these areas, we conducted comprehensive experiments using the RLBench2 benchmark [17], an extended version of RLBench tailored for bimanual manipulation and comprising tasks closely resembling real-world scenarios. Please refer to Appendix for more details about RLBench2 and real-world tasks." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 326, + 296, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 326, + 296, + 480 + ], + "spans": [ + { + "bbox": [ + 55, + 326, + 296, + 480 + ], + "type": "text", + "content": "Evaluation Settings. To evaluate the policy performance, we employ success rate as the primary metric. Although the policy generates multiple sequential actions during execution, we primarily focus on the final goal achievement rather than intermediate steps. Each task is associated with a specific success criterion defined by its target state. To comprehensively assess the policy's capability, we conduct experiments with varying numbers of demonstrations (20 and 100) during training. Figure 3 illustrates our experimental setup, including both the simulation environment and the Cobot Agilex ALOHA robot. Detailed descriptions of simulation tasks and real-world experimental settings are provided in the Appendix." + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 69, + 481, + 296, + 596 + ], + "blocks": [ + { + "bbox": [ + 69, + 481, + 296, + 596 + ], + "lines": [ + { + "bbox": [ + 69, + 481, + 296, + 596 + ], + "spans": [ + { + "bbox": [ + 69, + 481, + 296, + 596 + ], + "type": "image", + "image_path": "19ce8d6ab0fd354f2d346adf3d79d5393a255af8688d43ce0902f4732a570abf.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 605, + 295, + 628 + ], + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 628 + ], + "type": "text", + "content": "Figure 3. The left: the simulation environment of pick_laptop task. The right: the ALOHA device used in the real-world tasks." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 634, + 121, + 645 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 634, + 121, + 645 + ], + "spans": [ + { + "bbox": [ + 55, + 634, + 121, + 645 + ], + "type": "text", + "content": "4.2. Baselines" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 651, + 295, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 651, + 295, + 675 + ], + "spans": [ + { + "bbox": [ + 55, + 651, + 295, + 675 + ], + "type": "text", + "content": "We systematically evaluate KStar Diffuser against state-of-the-art methods in two major categories:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "Transformer-based methods: (1) Action Chunking with Transformers (ACT) [58] employs a Conditional VAE architecture, consisting of an encoder-decoder framework" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 72, + 555, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 228 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 228 + ], + "type": "text", + "content": "for joint angle sequence prediction; (2) Robotic View Transformer Leader Following (RVT-LF) [17] leverages RVT [15] as its backbone, incorporating a multi-view transformer for cross-view information aggregation and image re-rendering, coupled with a leader following mechanism for action prediction; (3) Perceiver-Actor Leader Following (PerAct-LF) [17] adopts the leader following paradigm based on PerAct [44], utilizing a perceiver Transformer to encode both instructions and voxel observations for optimal voxel action generation; (4) PerAct2 [17] enhances PerAct by implementing a unified feature space for dual-arm actions and employing combined self-attention for synchronized bimanual action prediction." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 232, + 556, + 341 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 232, + 556, + 341 + ], + "spans": [ + { + "bbox": [ + 313, + 232, + 556, + 341 + ], + "type": "text", + "content": "Diffusion-based methods: (1) Joint-based Diffusion Policy (DP-J) [9] adopts a diffusion model to robotic manipulation within an imitation learning framework, focusing on joint angle prediction; (2) EndEffector-based Diffusion Policy (DP-EE) is a DP variant which we reimplement Diffusion Policy to predict end-effector poses instead of joint angles, offering an alternative control paradigm; (3) 3D Diffusion Policy (DP3) [56] enhances 3D perception by incorporating point clouds for joint angle prediction." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 348, + 529, + 361 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 348, + 529, + 361 + ], + "spans": [ + { + "bbox": [ + 313, + 348, + 529, + 361 + ], + "type": "text", + "content": "4.3. Comparison Results with SOTA Methods" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 366, + 554, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 366, + 554, + 425 + ], + "spans": [ + { + "bbox": [ + 313, + 366, + 554, + 425 + ], + "type": "text", + "content": "Experimental Results on RLBench2. As shown in Table 1, the KStar Diffuser significantly outperforms other state-of-the-art baselines, achieving more than " + }, + { + "bbox": [ + 313, + 366, + 554, + 425 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 313, + 366, + 554, + 425 + ], + "type": "text", + "content": " higher overall performance with both 20 and 100 training demonstrations. We found that:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 426, + 555, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 555, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 555, + 605 + ], + "type": "text", + "content": "(1) Similar to learning a single-arm policy, the process of learning a bimanual policy can adapt quickly and achieve a high success rate, given a relatively consistent distribution of task trajectories. In push_box, where the objective is for both arms to push a box toward a specified target along fixed trajectories, our KStar Diffuser and other baseline models perform well. However, as task complexity increases, success rates decrease. For example, in the liftball task, both arms must lift a large ball simultaneously to complete the task. Any asynchrony in movement can cause instability, leading to the ball slipping and ultimately resulting in task failure. Our KStar Diffuser achieves its robust performance on such bimanual tasks by explicitly modeling the spatial and motion relationships between the two arms, surpassing other methods more than " + }, + { + "bbox": [ + 313, + 426, + 555, + 605 + ], + "type": "inline_equation", + "content": "6\\%" + }, + { + "bbox": [ + 313, + 426, + 555, + 605 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 606, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 606, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 606, + 556, + 715 + ], + "type": "text", + "content": "(2) Distinct from single-arm systems, bimanual robotic systems possess the capability for collaborative manipulation. Methods which are directly adapted from single-arm to bimanual manipulation exhibit high failure rates in tasks, e.g., pick_laptop, as they lack consideration of the spatial and motion relationships between the arms. Specifically, as shown in Figure 4, this task involves picking up a notebook lying flat on a cabinet surface. Given that the notebook rests fully against the tabletop, direct grasping by" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17399" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 83, + 101, + 529, + 348 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 95 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 95 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 95 + ], + "type": "text", + "content": "Table 1. The experimental result on simulated tasks. We train the policy with the setting of different training demonstrations, i.e. [20, 100], to test its capability comprehensively. The best results are in bold. Each result is reported with three seeds on average." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 83, + 101, + 529, + 348 + ], + "lines": [ + { + "bbox": [ + 83, + 101, + 529, + 348 + ], + "spans": [ + { + "bbox": [ + 83, + 101, + 529, + 348 + ], + "type": "table", + "html": "
MethodBackbonePush BoxLift BallHandover Item (easy)Pick LaptopSweep DustpanOverall
20 demos
ACT [58]Transformer26.0 ±4.617.3 ±5.80.0 ±0.00.0 ±0.00.0 ±0.08.7 ±1.6
RVT-LF [17]Transformer44.0 ±1.012.3 ±6.70.0 ±0.02.0 ±1.70.0 ±0.011.7 ±1.5
PerAct-LF [17]Transformer63.7 ±3.142.3 ±8.13.3 ±0.67.0 ±2.74.7 ±0.624.2 ±1.6
PerAct2 [17]Transformer59.7 ±7.637.7 ±4.78.0 ±1.017.3 ±0.61.7 ±1.524.9 ±2.1
DP-J [9]Diffusion30.0 ±3.619.7 ±2.10.0 ±0.00.0 ±0.00.0 ±4.010.0 ±1.0
DP-EE [9]Diffusion34.7 ±2.534.3 ±7.40.7 ±0.61.3 ±0.614.0 ±1.717.0 ±1.6
DP3 [56]Diffusion25.3 ±4.034.3 ±7.60.0 ±0.02.7 ±2.90.0 ±0.012.5 ±1.3
KStar Diffuser (Ours)Diffusion79.3 ±3.587.0 ±2.723.7 ±0.617.0 ±2.083.0 ±4.458.0 ±1.4
100 demos
ACT [58]Transformer48.7 ±6.840.7 ±2.10.0 ±0.00.0 ±0.00.0 ±0.017.9 ±1.0
RVT-LF [17]Transformer76.3 ±1.528.7 ±1.00.0 ±0.01.3 ±1.22.3 ±3.221.7 ±0.7
PerAct-LF [17]Transformer66.3 ±3.168.3 ±6.17.3 ±2.114.3 ±3.88.0 ±2.732.9 ±2.5
PerAct2 [17]Transformer77.0 ±3.047.7 ±3.210.3 ±0.634.3 ±4.06.7 ±2.133.9 ±1.0
DP-J [9]Diffusion62.7 ±7.643.3 ±2.10.0 ±0.01.3 ±0.60.0 ±4.021.5 ±1.6
DP-EE [9]Diffusion61.0 ±2.059.7 ±3.52.0 ±1.010.3 ±3.169.7 ±2.540.5 ±1.2
DP3 [56]Diffusion56.0 ±3.664.0 ±2.70.0 ±0.06.3 ±3.11.7 ±2.125.6 ±1.4
KStar Diffuser (Ours)Diffusion83.0 ±1.798.7 ±1.527.0 ±1.743.7 ±4.589.0 ±5.268.2 ±2.1
", + "image_path": "881164f59b9e47c7a79a073017e63358223d35a5ab0db09282488711775dffb6.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 365, + 296, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 365, + 296, + 450 + ], + "spans": [ + { + "bbox": [ + 55, + 365, + 296, + 450 + ], + "type": "text", + "content": "the robotic arm is not possible. Instead, an effective strategy is to control one arm to push the notebook outward from the cabinet by a short distance, allowing the other arm to pick it up. KStar Diffuser achieves a success rate approximately " + }, + { + "bbox": [ + 55, + 365, + 296, + 450 + ], + "type": "inline_equation", + "content": "9\\%" + }, + { + "bbox": [ + 55, + 365, + 296, + 450 + ], + "type": "text", + "content": " higher than other methods, demonstrating its ability to capture the coordinated motion patterns required for collaborative object manipulation between two arms." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 460, + 296, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 460, + 296, + 509 + ], + "spans": [ + { + "bbox": [ + 55, + 460, + 296, + 509 + ], + "type": "text", + "content": "Real-world Experimental Results. To comprehensively evaluate the policy's effectiveness, we build 2 tasks in real-world based on the simulation benchmark. The performances on real-world tasks are illustrated in Table 2." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 510, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 296, + 715 + ], + "type": "text", + "content": "Similar to the simulation results, we observe that policies which do not consider about bimanual scenarios, i.e., ACT, DP, and DP3, demonstrate limited capability across all bimanual tasks, achieving around " + }, + { + "bbox": [ + 55, + 510, + 296, + 715 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 55, + 510, + 296, + 715 + ], + "type": "text", + "content": " success ratio on average. Although PerAct2 is designed for bimanual tasks via mapping bimanual actions into a shared learning space, it fails to capture the spatial structure of the bimanual system, leading to ineffective arm coordination during execution. Furthermore, we also found PerAct2 faces significant inverse kinematics issues with its predicted end-effector poses, including joint configuration conflicts and unreachable positions which are shown in Figure 4. It is likely due to PerAct2's limited ability to capture the complex spatial constraints and kinematic relationships within the bimanual robotic system. In contrast, KStar Diffuser achieves superior bimanual coordination, surpassing other methods by over " + }, + { + "bbox": [ + 55, + 510, + 296, + 715 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 55, + 510, + 296, + 715 + ], + "type": "text", + "content": ", as it successfully captures the motion patterns" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 365, + 555, + 379 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 365, + 555, + 379 + ], + "spans": [ + { + "bbox": [ + 313, + 365, + 555, + 379 + ], + "type": "text", + "content": "between dual arms and predicts feasible end-effector poses." + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 324, + 418, + 547, + 494 + ], + "blocks": [ + { + "bbox": [ + 313, + 390, + 555, + 413 + ], + "lines": [ + { + "bbox": [ + 313, + 390, + 555, + 413 + ], + "spans": [ + { + "bbox": [ + 313, + 390, + 555, + 413 + ], + "type": "text", + "content": "Table 2. The result of real-world tasks. We train all policies with 100 demonstrations and test 15 times. The best result are in bold." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 324, + 418, + 547, + 494 + ], + "lines": [ + { + "bbox": [ + 324, + 418, + 547, + 494 + ], + "spans": [ + { + "bbox": [ + 324, + 418, + 547, + 494 + ], + "type": "table", + "html": "
MethodsLift PlateHandoverOverall
ACT [58]37.8 ±8.30.0 ±0.018.9 ±11.2
DP [9]42.1 ±6.50.0 ±0.021.1 ±12.5
DP3 [56]44.0 ±8.30.0 ±0.022.0 ±13.4
PerAct2 [17]51.1 ±6.38.8 ±3.029.9 ±10.2
KStar Diffuser (Ours)66.7 ±5.319.7 ±5.343.1 ±17.8
", + "image_path": "475298fe301cc143007e03a4691ca56ffba3dfefad97d33c54107b3ad89c2988.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 326, + 533, + 545, + 588 + ], + "blocks": [ + { + "bbox": [ + 349, + 513, + 518, + 524 + ], + "lines": [ + { + "bbox": [ + 349, + 513, + 518, + 524 + ], + "spans": [ + { + "bbox": [ + 349, + 513, + 518, + 524 + ], + "type": "text", + "content": "Table 3. Ablation study of model components." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 326, + 533, + 545, + 588 + ], + "lines": [ + { + "bbox": [ + 326, + 533, + 545, + 588 + ], + "spans": [ + { + "bbox": [ + 326, + 533, + 545, + 588 + ], + "type": "table", + "html": "
ST GraphKRHandover-SHandover-ROverall
✓✓27.0 ±1.719.7 ±5.323.4 ±5.2
✓×18.3 ±1.515.3 ±3.316.8 ±2.1
××16.3 ±3.113.3 ±9.414.8 ±2.1
", + "image_path": "5448b7cec52f500a71a10722aa82660280095f92ccaccccd7e79a6f9be558bb0.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 599, + 416, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 599, + 416, + 611 + ], + "spans": [ + { + "bbox": [ + 313, + 599, + 416, + 611 + ], + "type": "text", + "content": "4.4. Ablation Studies" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 617, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 555, + 715 + ], + "type": "text", + "content": "Effects of Model Components. To systematically evaluate the contribution of each component in KStar Diffuser, we conduct ablation experiments on the handover-item task in both simulated and real-world environments. We design a progressive ablation process by first removing the Differential Forward Kinematics module while retaining the Spatial-Temporal Graph (ST Graph), and then completely disabling both ST Graph and Kinematics Regularizer (KR)." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "17400" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 62, + 71, + 549, + 323 + ], + "blocks": [ + { + "bbox": [ + 62, + 71, + 549, + 323 + ], + "lines": [ + { + "bbox": [ + 62, + 71, + 549, + 323 + ], + "spans": [ + { + "bbox": [ + 62, + 71, + 549, + 323 + ], + "type": "image", + "image_path": "514c1ca54d213a8ab09bf241b7777aa88f449f206b1e094afc4efcb0626cf4ed.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 331, + 555, + 354 + ], + "lines": [ + { + "bbox": [ + 55, + 331, + 555, + 354 + ], + "spans": [ + { + "bbox": [ + 55, + 331, + 555, + 354 + ], + "type": "text", + "content": "Figure 4. The visualization of bimanual manipulation on simulated RLBench2 and real-world tasks. The blue annotations represent the motion of the robot's left arm, while the green annotations indicate the motion of the right arm." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 366, + 295, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 366, + 295, + 606 + ], + "spans": [ + { + "bbox": [ + 54, + 366, + 295, + 606 + ], + "type": "text", + "content": "The experimental results in Table 3 demonstrate the crucial role of each component. The removal of KR leads to a significant decrease in success rate, particularly pronounced in real-world scenarios. This performance degradation can be attributed to the fundamental differences between simulated and real environments. While simulated environments maintain consistent and noise-free inputs, real-world scenarios introduce various perturbations, e.g., sensor noise and light reflection, making the policy more susceptible to kinematic constraint violations without the regularization effect of KR. Further ablation by removing both ST Graph and KR results in a substantial performance drop across all experimental settings. This observation illustrates two key aspects: First, the ST Graph effectively captures the spatial-temporal dependencies among joints, which is essential for coordinating the relative positioning and interactions between robotic arms. Second, the graph structure's explicit encoding of the robotic physical architecture enhances the policy's robustness against unexpected perturbations by maintaining spatial and temporal coherence." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 612, + 173, + 625 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 612, + 173, + 625 + ], + "spans": [ + { + "bbox": [ + 55, + 612, + 173, + 625 + ], + "type": "text", + "content": "4.5. Qualitative Analysis" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 629, + 295, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 295, + 677 + ], + "type": "text", + "content": "We further present qualitative analysis in Figure 4. We compare the performance of KStar Diffuser with DP3 and PerAct2 in executing bimanual manipulation tasks within both simulated and real-world environments, respectively." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "In the simulation task, since the laptop rests flat on the cabinet, direct lifting is not feasible. One robotic arm initiates a forward push, creating space, while the other arm" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 366, + 555, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 366, + 555, + 450 + ], + "spans": [ + { + "bbox": [ + 313, + 366, + 555, + 450 + ], + "type": "text", + "content": "concurrently grasps and elevates the laptop. KStar Diffuser effectively models this dual-arm coordination, generating a precise trajectory of synchronized actions. Conversely, DP3, adapted from a single-arm policy to a dual-arm configuration, fails to achieve effective coordination. Concretely, after executing the push motion, the right arm does not halt, obstructing the left arm's lifting process." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 451, + 556, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 451, + 556, + 547 + ], + "spans": [ + { + "bbox": [ + 313, + 451, + 556, + 547 + ], + "type": "text", + "content": "In the real-world task, KStar Diffuser generates an executable item transfer trajectory between the left and right arms, with no collisions throughout the task, reflecting its strong environmental adaptability and collision avoidance capabilities. Conversely, PerAct2 encounters collisions during the handover process (marked in red), indicating less effective handling of dynamic real-world variables and a lack of kinematic awareness on the robot movements." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 561, + 388, + 573 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 561, + 388, + 573 + ], + "spans": [ + { + "bbox": [ + 313, + 561, + 388, + 573 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 582, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 556, + 713 + ], + "type": "text", + "content": "In this paper, we have proposed a novel Kinematics enhanced Spatial-TemporAl gRaph Diffuser (KStar Diffuser) that explicitly incorporates both robot structures and kinematics into the bimanual motion generation process. It consists of a spatial-temporal robot graph that explicitly models the robot physical configuration to guide the generative action denoising procedure, and a kinematics regularizer that augments the NBP learning objective by introducing joint-space supervision. Extensive experiments demonstrate that KStar Diffuser outperforms the baselines by a large margin on both simulation and real-world tasks." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17401" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 170, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 170, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 170, + 85 + ], + "type": "text", + "content": "6. Acknowledgements" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 297, + 294 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 297, + 294 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 297, + 294 + ], + "type": "text", + "content": "We would like to thank all co-authors for their efforts and the reviewers for their constructive comments. This work is supported by Shenzhen Science and Technology Program (KQTD20240729102207002), National Natural Science Foundation of China (Grant No. 62406092), National Natural Science Foundation of China (Grant No. U24B20175), Shenzhen Science and Technology Program (Grant No. KJZD20240903100017022), Guangdong Basic and Applied Basic Research Foundation (2025A1515010169), and Research on Efficient Exploration and Self-Evolution of APP Agents & Embodied Intelligent Cerebellum Control Model and Collaborative Feedback Training Project (Grant No. TC20240403047). Additionally, it is also supported by National Natural Science Foundation of China (Grant No. 62306090, No. 62236003), Natural Science Foundation of Guangdong Province of China (Grant No. 2024A1515010147)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 304, + 115, + 317 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 304, + 115, + 317 + ], + "spans": [ + { + "bbox": [ + 56, + 304, + 115, + 317 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 325, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 61, + 325, + 294, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 325, + 294, + 390 + ], + "spans": [ + { + "bbox": [ + 61, + 325, + 294, + 390 + ], + "type": "text", + "content": "[1] Lucia Angelini, Manuela Uliano, Angela Mazzeo, Mattia Penzotti, and Marco Controzzi. Self-collision avoidance in bimanual teleoperation using collisionik: algorithm revision and usability experiment. In 2022 IEEE-RAS 21st International Conference on Humanoid Robots (Humanoids), pages 112-118, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 392, + 295, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 392, + 295, + 435 + ], + "spans": [ + { + "bbox": [ + 61, + 392, + 295, + 435 + ], + "type": "text", + "content": "[2] Andreas Aristidou, Joan Lasenby, Yiorgos Chrysanthou, and Ariel Shamir. Inverse kinematics techniques in computer graphics: A survey. In Computer graphics forum, pages 35-58. Wiley Online Library, 2018. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 437, + 295, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 437, + 295, + 501 + ], + "spans": [ + { + "bbox": [ + 62, + 437, + 295, + 501 + ], + "type": "text", + "content": "[3] Weibang Bai, Ningshan Zhang, Baoru Huang, Ziwei Wang, Francesco Cursi, Ya-Yen Tsai, Bo Xiao, and Eric M Yeatman. Dual-arm coordinated manipulation for object twisting with human intelligence. In 2021 IEEE International Conference on Systems, Man, and Cybernetics (SMC), pages 902-908. IEEE, 2021. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 502, + 295, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 502, + 295, + 567 + ], + "spans": [ + { + "bbox": [ + 62, + 502, + 295, + 567 + ], + "type": "text", + "content": "[4] Aaron M Bestick, Samuel A Burden, Giorgia Willits, Nikhil Naikal, S Shankar Sastry, and Ruzena Bajcsy. Personalized kinematics for human-robot collaborative manipulation. In 2015 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 1037–1044. IEEE, 2015. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 569, + 295, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 569, + 295, + 623 + ], + "spans": [ + { + "bbox": [ + 62, + 569, + 295, + 623 + ], + "type": "text", + "content": "[5] Botond Bócsi, Duy Nguyen-Tuong, Lehel Csato, Bernhard Schoelkopf, and Jan Peters. Learning inverse kinematics with structured prediction. In 2011 IEEE/RSJ International Conference on Intelligent Robots and Systems, pages 698-703. IEEE, 2011. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 625, + 295, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 625, + 295, + 679 + ], + "spans": [ + { + "bbox": [ + 62, + 625, + 295, + 679 + ], + "type": "text", + "content": "[6] Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Joseph Dabis, Chelsea Finn, Keerthana Gopalakrishnan, Karol Hausman, Alex Herzog, Jasmine Hsu, et al. Rt-1: Robotics transformer for real-world control at scale. arXiv preprint arXiv:2212.06817, 2022. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 681, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 681, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 62, + 681, + 295, + 713 + ], + "type": "text", + "content": "[7] Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Xi Chen, Krzysztof Choromanski, Tianli Ding, Danny Driess, Avinava Dubey, Chelsea Finn, et al. Rt-2:" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "Vision-language-action models transfer web knowledge to robotic control. arXiv preprint arXiv:2307.15818, 2023. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 320, + 96, + 553, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 96, + 553, + 139 + ], + "spans": [ + { + "bbox": [ + 320, + 96, + 553, + 139 + ], + "type": "text", + "content": "[8] Shizhe Chen, Ricardo Garcia, Cordelia Schmid, and Ivan Laptev. Polarnet: 3d point clouds for language-guided robotic manipulation. arXiv preprint arXiv:2309.15596, 2023.2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 320, + 141, + 553, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 141, + 553, + 195 + ], + "spans": [ + { + "bbox": [ + 320, + 141, + 553, + 195 + ], + "type": "text", + "content": "[9] Cheng Chi, Zhenjia Xu, Siyuan Feng, Eric Cousineau, Yilun Du, Benjamin Burchfiel, Russ Tedrake, and Shuran Song. Diffusion policy: Visuomotor policy learning via action diffusion. The International Journal of Robotics Research, page 02783649241273668, 2023. 1, 2, 3, 4, 6, 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 198, + 553, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 198, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 316, + 198, + 553, + 251 + ], + "type": "text", + "content": "[10] Cheng Chi, Zhenjia Xu, Chuer Pan, Eric Cousineau, Benjamin Burchfiel, Siyuan Feng, Russ Tedrake, and Shuran Song. Universal manipulation interface: In-the-wild robot teaching without in-the-wild robots. arXiv preprint arXiv:2402.10329, 2024. 1, 2, 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 254, + 553, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 254, + 553, + 297 + ], + "spans": [ + { + "bbox": [ + 316, + 254, + 553, + 297 + ], + "type": "text", + "content": "[11] Akos Csiszar, Jan Eilers, and Alexander Verl. On solving the inverse kinematics problem using neural networks. In 2017 24th International Conference on Mechatronics and Machine Vision in Practice (M2VIP), pages 1-6, 2017. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 299, + 553, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 299, + 553, + 363 + ], + "spans": [ + { + "bbox": [ + 317, + 299, + 553, + 363 + ], + "type": "text", + "content": "[12] Aaron D'Souza, Sethu Vijayakumar, and Stefan Schaal. Learning inverse kinematics. In Proceedings 2001 IEEE/RSJ International Conference on Intelligent Robots and Systems. Expanding the Societal Role of Robotics in the Next Millennium (Cat. No. 01CH37180), pages 298-303. IEEE, 2001. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 365, + 553, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 365, + 553, + 408 + ], + "spans": [ + { + "bbox": [ + 316, + 365, + 553, + 408 + ], + "type": "text", + "content": "[13] Zipeng Fu, Tony Z Zhao, and Chelsea Finn. Mobile aloha: Learning bimanual mobile manipulation with low-cost whole-body teleoperation. arXiv preprint arXiv:2401.02117, 2024. 1, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 411, + 553, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 411, + 553, + 464 + ], + "spans": [ + { + "bbox": [ + 316, + 411, + 553, + 464 + ], + "type": "text", + "content": "[14] Jianfeng Gao, Xiaoshu Jin, Franziska Krebs, Noémie Jaquier, and Tamim Asfour. Bi-kvil: Keypoints-based visual imitation learning of bimanual manipulation tasks. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 16850-16857. IEEE, 2024. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 467, + 553, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 467, + 553, + 509 + ], + "spans": [ + { + "bbox": [ + 317, + 467, + 553, + 509 + ], + "type": "text", + "content": "[15] Ankit Goyal, Jie Xu, Yijie Guo, Valts Blukis, Yu-Wei Chao, and Dieter Fox. Rvt: Robotic view transformer for 3d object manipulation. In Conference on Robot Learning, pages 694–710. PMLR, 2023. 1, 2, 3, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 511, + 553, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 511, + 553, + 566 + ], + "spans": [ + { + "bbox": [ + 317, + 511, + 553, + 566 + ], + "type": "text", + "content": "[16] Reinhard Grassmann, Vincent Modes, and Jessica Burgner-Kahrs. Learning the forward and inverse kinematics of a 6-dof concentric tube continuum robot in se (3). In 2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 5125-5132. IEEE, 2018. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 567, + 553, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 567, + 553, + 610 + ], + "spans": [ + { + "bbox": [ + 317, + 567, + 553, + 610 + ], + "type": "text", + "content": "[17] Markus Grotz, Mohit Shridhar, Tamim Asfour, and Dieter Fox. Peract2: A perceiver actor framework for bimanual manipulation tasks. arXiv preprint arXiv:2407.00278, 2024. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 613, + 553, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 613, + 553, + 655 + ], + "spans": [ + { + "bbox": [ + 317, + 613, + 553, + 655 + ], + "type": "text", + "content": "[18] Abdullah Aamir Hayat, Ratan OM Sadanand, and Subir K Saha. Robot manipulation through inverse kinematics. In Proceedings of the 2015 conference on advances in robotics, pages 1-6, 2015. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 658, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 553, + 689 + ], + "type": "text", + "content": "[19] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "text", + "content": "[20] Binghao Huang, Yixuan Wang, Xinyi Yang, Yiyue Luo, and Yunzhu Li. 3d-vitac: Learning fine-grained manipulation" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "17402" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "text", + "content": "with visuo-tactile sensing. arXiv preprint arXiv:2410.24091, 2024. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 96, + 295, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 96, + 295, + 117 + ], + "spans": [ + { + "bbox": [ + 57, + 96, + 295, + 117 + ], + "type": "text", + "content": "[21] Kenneth H Hunt. Structural kinematics of in-parallel-actuated robot-arms. 1983. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 119, + 295, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 119, + 295, + 152 + ], + "spans": [ + { + "bbox": [ + 56, + 119, + 295, + 152 + ], + "type": "text", + "content": "[22] Michael I Jordan and David E Rumelhart. Forward models: Supervised learning with a distal teacher. In Backpropagation, pages 189-236. Psychology Press, 2013. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 153, + 295, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 153, + 295, + 185 + ], + "spans": [ + { + "bbox": [ + 57, + 153, + 295, + 185 + ], + "type": "text", + "content": "[23] Tsung-Wei Ke, Nikolaos Gkanatsios, and Katerina Fragkiadaki. 3d diffuser actor: Policy diffusion with 3d scene representations. arXiv preprint arXiv:2402.10885, 2024. 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 186, + 295, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 186, + 295, + 240 + ], + "spans": [ + { + "bbox": [ + 56, + 186, + 295, + 240 + ], + "type": "text", + "content": "[24] Moo Jin Kim, Karl Pertsch, Siddharth Karamcheti, Ted Xiao, Ashwin Balakrishna, Suraj Nair, Rafael Rafailov, Ethan Foster, Grace Lam, Pannag Sanketi, et al. Openvla: An open-source vision-language-action model. arXiv preprint arXiv:2406.09246, 2024. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 57, + 243, + 295, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 243, + 295, + 274 + ], + "spans": [ + { + "bbox": [ + 57, + 243, + 295, + 274 + ], + "type": "text", + "content": "[25] Thomas N Kipf and Max Welling. Semi-supervised classification with graph convolutional networks. arXiv preprint arXiv:1609.02907, 2016. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 57, + 277, + 295, + 298 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 277, + 295, + 298 + ], + "spans": [ + { + "bbox": [ + 57, + 277, + 295, + 298 + ], + "type": "text", + "content": "[26] CS George Lee. Robot arm kinematics, dynamics, and control. Computer, 15(12):62-80, 1982. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 300, + 295, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 300, + 295, + 342 + ], + "spans": [ + { + "bbox": [ + 56, + 300, + 295, + 342 + ], + "type": "text", + "content": "[27] Maolin Lei, Ting Wang, Chen Yao, Huan Liu, Zhi Wang, and Yongsheng Deng. Real-time kinematics-based self-collision avoidance algorithm for dual-arm robots. Applied Sciences, 10(17):5893, 2020. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 344, + 295, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 344, + 295, + 409 + ], + "spans": [ + { + "bbox": [ + 56, + 344, + 295, + 409 + ], + "type": "text", + "content": "[28] Jiefeng Li, Chao Xu, Zhicun Chen, Siyuan Bian, Lixin Yang, and Cewu Lu. Hybrik: A hybrid analytical-neural inverse kinematics solution for 3d human pose and shape estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3383-3393, 2021. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 411, + 295, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 411, + 295, + 465 + ], + "spans": [ + { + "bbox": [ + 56, + 411, + 295, + 465 + ], + "type": "text", + "content": "[29] Jiefeng Li, Chao Xu, Zhicun Chen, Siyuan Bian, Lixin Yang, and Cewu Lu. Hybrik: A hybrid analytical-neural inverse kinematics solution for 3d human pose and shape estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3383-3393, 2021. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 468, + 295, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 468, + 295, + 510 + ], + "spans": [ + { + "bbox": [ + 56, + 468, + 295, + 510 + ], + "type": "text", + "content": "[30] I Liu, Chun Arthur, Sicheng He, Daniel Seita, and Gaurav Sukhatme. Voxact-b: Voxel-based acting and stabilizing policy for bimanual manipulation. arXiv preprint arXiv:2407.04152, 2024. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 512, + 295, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 512, + 295, + 567 + ], + "spans": [ + { + "bbox": [ + 56, + 512, + 295, + 567 + ], + "type": "text", + "content": "[31] Xiao Ma, Sumit Patidar, Iain Haughton, and Stephen James. Hierarchical diffusion policy for kinematics-aware multitask robotic manipulation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18081-18090, 2024. 1, 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 568, + 295, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 568, + 295, + 590 + ], + "spans": [ + { + "bbox": [ + 56, + 568, + 295, + 590 + ], + "type": "text", + "content": "[32] Matthew T Mason. Mechanics of robotic manipulation. MIT press, 2001. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 591, + 295, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 591, + 295, + 644 + ], + "spans": [ + { + "bbox": [ + 56, + 591, + 295, + 644 + ], + "type": "text", + "content": "[33] Yao Mu, Qinglong Zhang, Mengkang Hu, Wenhai Wang, Mingyu Ding, Jun Jin, Bin Wang, Jifeng Dai, Yu Qiao, and Ping Luo. Embodiedgpt: Vision-language pre-training via embodied chain of thought. Advances in Neural Information Processing Systems, 36, 2024. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 647, + 295, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 295, + 689 + ], + "type": "text", + "content": "[34] Yoshihiko Nakamura, Kiyoshi Nagai, and Tsuneo Yoshikawa. Dynamics and stability in coordination of multiple robotic mechanisms. The International Journal of Robotics Research, 8(2):44-61, 1989. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 692, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 692, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 692, + 295, + 713 + ], + "type": "text", + "content": "[35] Valerio Ortenzi, Naresh Marturi, Michael Mistry, Jeffrey Kuo, and Rustam Stolkin. Vision-based framework" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 106 + ], + "type": "text", + "content": "to estimate robot configuration and kinematic constraints. IEEE/ASME Transactions on Mechatronics, 23(5):2402-2412, 2018. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 107, + 553, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 140 + ], + "type": "text", + "content": "[36] Eric Paljug, Xiaoping Yun, and Vijay Kumar. Control of rolling contacts in multi-arm manipulation. IEEE Transactions on Robotics and Automation, 10(4):441-452, 1994. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 141, + 553, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 553, + 184 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 553, + 184 + ], + "type": "text", + "content": "[37] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4195-4205, 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 186, + 553, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 186, + 553, + 240 + ], + "spans": [ + { + "bbox": [ + 316, + 186, + 553, + 240 + ], + "type": "text", + "content": "[38] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 243, + 553, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 243, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 316, + 243, + 553, + 285 + ], + "type": "text", + "content": "[39] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 287, + 553, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 287, + 553, + 341 + ], + "spans": [ + { + "bbox": [ + 316, + 287, + 553, + 341 + ], + "type": "text", + "content": "[40] Allen Z Ren, Justin Lidard, Lars L Ankile, Anthony Simeonov, Pulkit Agrawal, Anirudha Majumdar, Benjamin Burchfiel, Hongkai Dai, and Max Simchowitz. Diffusion policy policy optimization. arXiv preprint arXiv:2409.00588, 2024. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 343, + 555, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 343, + 555, + 398 + ], + "spans": [ + { + "bbox": [ + 316, + 343, + 555, + 398 + ], + "type": "text", + "content": "[41] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 399, + 553, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 399, + 553, + 442 + ], + "spans": [ + { + "bbox": [ + 316, + 399, + 553, + 442 + ], + "type": "text", + "content": "[42] Nilanjan Sarkar, Xiaoping Yun, and Vijay Kumar. Dynamic control of 3-d rolling contacts in two-arm manipulation. IEEE Transactions on Robotics and Automation, 13(3): 364-376, 1997. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 445, + 553, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 445, + 553, + 475 + ], + "spans": [ + { + "bbox": [ + 316, + 445, + 553, + 475 + ], + "type": "text", + "content": "[43] Youshik Shin and Zeungnam Bien. Collision-free trajectory planning for two robot arms. Robotica, 7(3):205-212, 1989. 1, 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 478, + 553, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 478, + 553, + 521 + ], + "spans": [ + { + "bbox": [ + 316, + 478, + 553, + 521 + ], + "type": "text", + "content": "[44] Mohit Shridhar, Lucas Manuelli, and Dieter Fox. Perceiver-actor: A multi-task transformer for robotic manipulation. In Conference on Robot Learning, pages 785–799. PMLR, 2023. 1, 2, 3, 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 523, + 553, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 523, + 553, + 567 + ], + "spans": [ + { + "bbox": [ + 316, + 523, + 553, + 567 + ], + "type": "text", + "content": "[45] Christian Smith, Yiannis Karayiannidis, Lazaros Nalpantidis, Xavi Gratal, Peng Qi, Dimos V Dimarogonas, and Danica Kragic. Dual arm manipulation—a survey. Robotics and Autonomous systems, 60(10):1340-1353, 2012. 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 568, + 553, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 568, + 553, + 622 + ], + "spans": [ + { + "bbox": [ + 316, + 568, + 553, + 622 + ], + "type": "text", + "content": "[46] Octo Model Team, Dibya Ghosh, Homer Walke, Karl Pertsch, Kevin Black, Oier Mees, Sudeep Dasari, Joel Hejna, Tobias Kreiman, Charles Xu, et al. Octo: An open-source generalist robot policy. arXiv preprint arXiv:2405.12213, 2024. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 624, + 553, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 553, + 667 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 553, + 667 + ], + "type": "text", + "content": "[47] Vitalis Vosylius, Younggyo Seo, Jafar Uruç, and Stephen James. Render and diffuse: Aligning image and action spaces for diffusion-based behaviour cloning. arXiv preprint arXiv:2405.18196, 2024. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "text", + "content": "[48] Lirui Wang, Xinlei Chen, Jialiang Zhao, and Kaiming He. Scaling proprioceptive-visual learning with heterogeneous pre-trained transformers. arXiv preprint arXiv:2409.20537, 2024. 2" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "17403" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 654 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "text", + "content": "[49] Yixiao Wang, Yifei Zhang, Mingxiao Huo, Ran Tian, Xiang Zhang, Yichen Xie, Chenfeng Xu, Pengliang Ji, Wei Zhan, Mingyu Ding, et al. Sparse diffusion policy: A sparse, reusable, and flexible policy for robot learning. arXiv preprint arXiv:2407.01531, 2024. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 294, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 294, + 183 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 294, + 183 + ], + "type": "text", + "content": "[50] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation. Advances in Neural Information Processing Systems, 36, 2024. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 185, + 294, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 185, + 294, + 206 + ], + "spans": [ + { + "bbox": [ + 56, + 185, + 294, + 206 + ], + "type": "text", + "content": "[51] Chris Welman. Inverse kinematics and geometric constraints for articulated figure manipulation. 1993. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 208, + 294, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 208, + 294, + 251 + ], + "spans": [ + { + "bbox": [ + 56, + 208, + 294, + 251 + ], + "type": "text", + "content": "[52] Kun Wu, Yichen Zhu, Jinming Li, Junjie Wen, Ning Liu, Zhiyuan Xu, Qinru Qiu, and Jian Tang. Discrete policy: Learning disentangled action space for multi-task robotic manipulation. arXiv preprint arXiv:2409.18707, 2024. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 253, + 294, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 253, + 294, + 308 + ], + "spans": [ + { + "bbox": [ + 56, + 253, + 294, + 308 + ], + "type": "text", + "content": "[53] Wenke Xia, Dong Wang, Xincheng Pang, Zhigang Wang, Bin Zhao, Di Hu, and Xuelong Li. Kinematic-aware prompting for generalizable articulated object manipulation with llms. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 2073-2080. IEEE, 2024. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 309, + 294, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 309, + 294, + 363 + ], + "spans": [ + { + "bbox": [ + 56, + 309, + 294, + 363 + ], + "type": "text", + "content": "[54] Andrea Maria Zanchettin, Paolo Rocco, Luca Bascetta, Ioannis Symeonidis, and Steffen Peldschus. Kinematic analysis and synthesis of the human arm motion during a manipulation task. In 2011 IEEE international conference on robotics and automation, pages 2692-2697. IEEE, 2011. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 365, + 294, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 365, + 294, + 407 + ], + "spans": [ + { + "bbox": [ + 56, + 365, + 294, + 407 + ], + "type": "text", + "content": "[55] Yanjie Ze, Zixuan Chen, Wenhao Wang, Tianyi Chen, Xialin He, Ying Yuan, Xue Bin Peng, and Jiajun Wu. Generalizable humanoid manipulation with improved 3d diffusion policies. arXiv preprint arXiv:2410.10803, 2024. 1, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 409, + 294, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 409, + 294, + 442 + ], + "spans": [ + { + "bbox": [ + 56, + 409, + 294, + 442 + ], + "type": "text", + "content": "[56] Yanjie Ze, Gu Zhang, Kangning Zhang, Chenyuan Hu, Muhan Wang, and Huazhe Xu. 3d diffusion policy. arXiv preprint arXiv:2403.03954, 2024. 1, 2, 6, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 444, + 294, + 496 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 444, + 294, + 496 + ], + "spans": [ + { + "bbox": [ + 56, + 444, + 294, + 496 + ], + "type": "text", + "content": "[57] Junjie Zhang, Chenjia Bai, Haoran He, Wenke Xia, Zhigang Wang, Bin Zhao, Xiu Li, and Xuelong Li. Sam-e: Leveraging visual foundation model with sequence imitation for embodied manipulation. arXiv preprint arXiv:2405.19586, 2024. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 498, + 294, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 498, + 294, + 542 + ], + "spans": [ + { + "bbox": [ + 56, + 498, + 294, + 542 + ], + "type": "text", + "content": "[58] Tony Z Zhao, Vikash Kumar, Sergey Levine, and Chelsea Finn. Learning fine-grained bimanual manipulation with low-cost hardware. arXiv preprint arXiv:2304.13705, 2023. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 544, + 294, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 544, + 294, + 597 + ], + "spans": [ + { + "bbox": [ + 56, + 544, + 294, + 597 + ], + "type": "text", + "content": "[59] Zangwei Zheng, Xiangyu Peng, Tianji Yang, Chenhui Shen, Shenggui Li, Hongxin Liu, Yukun Zhou, Tianyi Li, and Yang You. Open-sora: Democratizing efficient video production for all, march 2024. URL https://github.com/hpcaitech/Open-Sora, 1(3):4.2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 599, + 294, + 654 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 599, + 294, + 654 + ], + "spans": [ + { + "bbox": [ + 56, + 599, + 294, + 654 + ], + "type": "text", + "content": "[60] Minjie Zhu, Yichen Zhu, Jinming Li, Junjie Wen, Zhiyuan Xu, Ning Liu, Ran Cheng, Chaomin Shen, Yaxin Peng, Feifei Feng, et al. Scaling diffusion policy in transformer to 1 billion parameters for robotic manipulation. arXiv preprint arXiv:2409.14411, 2024. 2" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17404" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Spatial457_ A Diagnostic Benchmark for 6D Spatial Reasoning of Large Mutimodal Models/96f4562f-945b-497d-9468-cd476fb5f769_content_list.json b/2025/Spatial457_ A Diagnostic Benchmark for 6D Spatial Reasoning of Large Mutimodal Models/96f4562f-945b-497d-9468-cd476fb5f769_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..7ecd41b6210c4309e67f8a0ab68ec02bee5c8081 --- /dev/null +++ b/2025/Spatial457_ A Diagnostic Benchmark for 6D Spatial Reasoning of Large Mutimodal Models/96f4562f-945b-497d-9468-cd476fb5f769_content_list.json @@ -0,0 +1,1790 @@ +[ + { + "type": "text", + "text": "Spatial457: A Diagnostic Benchmark for 6D Spatial Reasoning of Large Multimodal Models", + "text_level": 1, + "bbox": [ + 181, + 130, + 818, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xingrui Wang $^{1}$ , Wufei Ma $^{1}$ , Tiezheng Zhang $^{1}$ , Celso M de Melo $^{2}$ , Jieneng Chen $^{1\\dagger}$ , Alan Yuille $^{1\\dagger}$", + "bbox": [ + 124, + 202, + 879, + 222 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Johns Hopkins University $^{2}$ DEVCOM Army Research Laboratory", + "bbox": [ + 289, + 223, + 704, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 273, + 326, + 290 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Although large multimodal models (LMMs) have demonstrated remarkable capabilities in visual scene interpretation and reasoning, their capacity for complex and precise 3-dimensional spatial reasoning remains uncertain. Existing benchmarks focus predominantly on 2D spatial understanding and lack a framework to comprehensively evaluate 6D spatial reasoning across varying complexities. To address this limitation, we present Spatial457, a scalable and unbiased synthetic dataset designed with 4 key capability for spatial reasoning: multi-object recognition, 2D location, 3D location, and 3D orientation. We develop a cascading evaluation structure, constructing 7 question types across 5 difficulty levels that range from basic single object recognition to our new proposed complex 6D spatial reasoning tasks. We evaluated various large multimodal models (LMMs) on Spatial457, observing a general decline in performance as task complexity increases, particularly in 3D reasoning and 6D spatial tasks. To quantify these challenges, we introduce the Relative Performance Dropping Rate (RPDR), highlighting key weaknesses in 3D reasoning capabilities. Leveraging the unbiased attribute design of our dataset, we also uncover prediction biases across different attributes, with similar patterns observed in real-world image settings. The code is released in https://github.com/XingruiWang/Spatial457.", + "bbox": [ + 91, + 306, + 483, + 684 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 723, + 220, + 739 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In recent years, large multimodal models (LMMs) have demonstrated remarkable reasoning capabilities, enabling them to interpret visual scenes from images and articulate their understanding through natural language [1, 3, 11, 26, 27, 38, 39, 48]. These advancements have been evaluated using a variety of benchmarks that primarily emphasize semantic features—such as object categories and appearances, spatial relationships within the 2D image plane [15,", + "bbox": [ + 89, + 742, + 482, + 864 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "17, 18] (e.g. left-right positioning from the camera view), and limited 2.5D features such as camera depth [40]. Tasks like object detection [13, 22], image captioning [21, 42, 47] and visual question answering [18, 50, 51] have significantly benefited from these evaluations, driving LMMs to new heights in visual understanding.", + "bbox": [ + 511, + 275, + 906, + 366 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "However, a critical gap remains in assessing the ability of LMMs to comprehend and reason about the full six degrees of freedom (6D) of objects in spatial environments, which includes both 3D positions and orientations [43, 46, 53]. Despite its significance in many real-world applications, such as robotics [2], autonomous navigation [14], and augmented reality [5], there are no available benchmarks for LMMs that evaluate 6D spatial reasoning capabilities.", + "bbox": [ + 511, + 367, + 908, + 489 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Building a scalable diagnostic dataset for 6D vision tasks poses considerable challenges due to the extensive annotations required across vast image collections. One strategy is to develop 6D vision question-answering benchmarks on top of existing image datasets [4, 6, 7, 36]. However, real-world image collections often exhibit significant biases in 3D locations and poses, influenced by camera perspective limitations and the prevalence of common object orientations in everyday scenes. For instance, statistics from the SUN-RGBD [36] and nuScenes [7] datasets reveal that over $70\\%$ of objects cluster into a single predominant direction (see Fig. 2), limiting the diversity required for comprehensive 6D reasoning evaluation.", + "bbox": [ + 511, + 489, + 908, + 686 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To overcome these challenges, we introduce an unbiased and scalable synthetic dataset rendered in a realistic style, specifically designed to diagnose the 6D spatial reasoning abilities of LMMs. To comprehensively evaluate diverse aspects of 6D spatial reasoning, we focus on four core capabilities: multi-object recognition, 2D locations, 3D locations, and 3D orientation. Multi-object recognition serves as the foundation of scene understanding, while 2D location represents the simplest way to depict the spatial relationship between objects within the image plane [18, 19]. The 3D location factor extends understanding into three-dimensional space, capturing the depth and distance of objects, which is essential for tasks such as reasoning about occlusions [9, 10, 44]. The 3D orientation introduces object", + "bbox": [ + 511, + 688, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "1Approved for public release: distribution is unlimited.", + "bbox": [ + 107, + 875, + 400, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$\\dagger$ : Equal advising.", + "bbox": [ + 116, + 888, + 212, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "24669", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1864bdae54cca2dc2242be14e7524808da21299f243132d64d8d7e4e7da4802c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 143, + 92, + 326, + 196 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/30d3a8c4f56983c66f5072c4add864c722b739fc85750ef883d5c3ec0535e6fd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 163, + 203, + 186, + 220 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3D Orientations", + "bbox": [ + 192, + 205, + 290, + 218 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/73f4fbe35d85cfb1d81613d5e40465c3a10eb66eb51d47017fe9331b1e3a2e2d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 161, + 234, + 187, + 253 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3D Locations", + "bbox": [ + 192, + 239, + 272, + 250 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/5ca5917c849a9c43b71b82d8ae007c43a6f4ef55085d41bd963f6fcadca79188.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 165, + 270, + 184, + 285 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2D Locations", + "bbox": [ + 192, + 272, + 272, + 282 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/7284700c93f5092b896d396ebb2139ffbb6a005fe1ebc54d1d6b3cadbb359ff1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 163, + 304, + 187, + 316 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Multiple Objects", + "bbox": [ + 194, + 305, + 281, + 316 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/e3c9dcd25aabff842a0391022215b0ee9b9a8075ef389524db34b89ec26de262.jpg", + "image_caption": [ + "4 Capabilities of Comprehensive Spatial Reasoning" + ], + "image_footnote": [], + "bbox": [ + 351, + 92, + 666, + 354 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/4bf3ba052d061251eeae6726d273807e7e69ab2f73ab853dfff8c86a9093edcf.jpg", + "image_caption": [ + "Figure 1. Overview of the Spatial457 benchmark. We define four core capabilities for spatial reasoning. By incorporating these capabilities step-by-step, the benchmark assesses models across five progressive difficulty levels and seven question types, ranging from single-object recognition to advanced 6D spatial relationships and collision prediction. Our evaluation demonstrates a significant performance drop for more advanced questions and highlights the gap between state-of-the-art (SoTA) models and human performance in complex spatial reasoning VQA tasks." + ], + "image_footnote": [], + "bbox": [ + 674, + 92, + 856, + 354 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/991ef26e1fdb5ca8588e848748afa12a64c729a885d18c8c85f9749b0aea54a3.jpg", + "image_caption": [ + "(a) SUN-RGBD" + ], + "image_footnote": [], + "bbox": [ + 125, + 463, + 227, + 545 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/7bf783f866998671f7592a8496fdc91cc480f4b16aea96c8d3040a39fcaf0e6d.jpg", + "image_caption": [ + "(b) nuScenes", + "Figure 2. The distribution of 3D poses in representative 6D datasets (a&b) highlights an imbalance in object orientations. This inspired us to develop a balanced benchmark encompassing all degrees of 3D orientation." + ], + "image_footnote": [], + "bbox": [ + 230, + 463, + 336, + 545 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/2ad47a1c7db930e43d981363ad687305e879358386caf9c6f1703cc066b49da2.jpg", + "image_caption": [ + "(c) Ours" + ], + "image_footnote": [], + "bbox": [ + 344, + 463, + 447, + 545 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "rotation [44, 46], enabling models to reason about precise 3D poses of objects within a scene.", + "bbox": [ + 89, + 643, + 482, + 672 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "By progressively incorporating the four core capabilities, we establish five levels of difficulty, resulting in seven distinct question types. These questions span a spectrum of spatial reasoning complexity, ranging from basic semantic identification and 2D spatial relationships to advanced 3D positional understanding and orientation-based queries, culminating in comprehensive 6D spatial reasoning tasks.", + "bbox": [ + 89, + 672, + 482, + 779 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our evaluation measures performance across varied question types and examines how it is affected by different combinations of the four capabilities. This tiered design provides comprehensive benchmarks for spatial reasoning capabilities, highlighting each model's strengths and areas of vulnerability. To the best of our knowledge, this work presents the first comprehensive benchmark designed to evaluate the 6D spatial reasoning capabilities of vision-", + "bbox": [ + 89, + 780, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "language models.", + "bbox": [ + 511, + 465, + 632, + 479 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Additionally, our dataset's unbiased attribute design allows us to analyze potential biases in model predictions. By systematically exploring model performance under different attributes, we identify tendencies or limitations that may affect real-world applications. Altogether, our benchmark not only provides a foundational tool for evaluating LMMs but also serves as a guide for developing future benchmarks that support models with advanced 6D spatial reasoning.", + "bbox": [ + 511, + 484, + 906, + 604 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our main contributions are threefold:", + "bbox": [ + 531, + 608, + 781, + 622 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. An unbiased and scalable synthetic dataset: We present a realistic, unbiased synthetic dataset designed to assess 6D spatial reasoning in LMMs. We define four core capabilities and propose five difficulty levels for spatial reasoning VQA with seven question types.", + "2. Comprehensive benchmarking of 6D capabilities: We benchmark various LMMs across all difficulty levels. The performance variations across settings reveal potential limitations or strengths, providing valuable insights into which components should be prioritized to improve the spatial reasoning of LMMs.", + "3. Bias analysis in model predictions: Leveraging the unbiased nature of our dataset, we conduct an in-depth bias analysis to explore how LMM predictions may be unbalanced across different attributes, especially 3D-related ones. This reveals potential weaknesses in model predictions, offering insights into limitations that may also appear in real-world scenarios." + ], + "bbox": [ + 513, + 626, + 906, + 897 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "24670", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 91, + 89, + 233, + 104 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1. Synthetic Datasets in VQA", + "text_level": 1, + "bbox": [ + 89, + 114, + 331, + 130 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The use of synthetic data in visual question answering (VQA) has gained traction as a means to circumvent the limitations and biases inherent in real-world datasets. Synthetic environments [16, 30, 34] allow for controlled manipulation of scene elements, enabling researchers to probe specific reasoning abilities in VQA models without the confounding factors often present in natural imagery. Previous works [19, 24, 25] have utilized synthetic environments to evaluate multimodal models across a range of visual reasoning tasks, typically focusing on 2D spatial relationships and object features. For instance, CLEVR [19], a synthetic dataset widely used in the VQA community, evaluates compositional reasoning by introducing artificially generated scenes with diverse object arrangements. Super-CLEVR-3D [45] represents a significant step forward in evaluating 3D locations, it lacks an assessment of 6D spatial relationships that consider both 3D location and 3D orientation. Our work extends this tradition by creating a synthetic dataset tailored for 6D spatial reasoning, filling a gap in current synthetic datasets that do not address the complete spatial understanding of objects in 3D space.", + "bbox": [ + 89, + 131, + 485, + 450 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. LMM Benchmarks", + "text_level": 1, + "bbox": [ + 89, + 454, + 274, + 468 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Existing benchmarks [15, 18, 35, 51] for large multimodal models (LMMs) predominantly assess their performance on tasks involving 2D visual question answering and multimodal reasoning. For instance, datasets like VQAv2 [15] and GQA [18] focus on answering questions based on visual content. However, these benchmarks leave a gap in evaluating models' capabilities in spatial contexts. Recent benchmarks, such as CV-Bench [40], assess spatial understanding, including depth ordering for 2.5D evaluation. More recent work [9, 10, 32] introduces questions regarding either aspect of 3D locations or rotations. Our proposed benchmark goes beyond traditional 2D, 2.5D, and 3D location or rotation assessments by introducing tasks that challenge LMMs to understand 3D positions and orientations together. We also evaluate models in a cascading manner across five difficulty levels and summarize their performance in four reasoning abilities, along with their output biases, providing a deeper analysis of the weaknesses in spatial reasoning within multimodal AI.", + "bbox": [ + 89, + 470, + 483, + 758 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Spatial457: Comprehensive Spatial Reasoning Benchmark", + "text_level": 1, + "bbox": [ + 89, + 768, + 482, + 805 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To comprehensively evaluate the spatial reasoning capabilities of vision-language models, we introduce a synthetic dataset named Spatial457. We define four core capabilities for comprehensive spatial reasoning (Sec. 3.1) and generate scenes with full 6D annotations (Sec. 3.2). The full combination of these capabilities form the new 6D spatial", + "bbox": [ + 89, + 810, + 483, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "relationship VQA tasks (Sec. 3.3).", + "bbox": [ + 511, + 90, + 743, + 106 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To thoroughly assess model capabilities, we define five levels of difficulty in vision question answering (VQA) tasks, converting seven question subsets that range from single-object recognition to the newly proposed 6D spatial reasoning tasks (Sec. 3.4). This structured progression in question difficulty is achieved by sequentially introducing new spatial reasoning capabilities, as illustrated in Fig. 1.", + "bbox": [ + 511, + 107, + 906, + 213 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Four Capabilities for Spatial Reasoning", + "text_level": 1, + "bbox": [ + 511, + 224, + 854, + 242 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We define 4 capabilities for comprehensive spatial understanding in a 3D scene: multiple objects, 2D locations, 3D locations, and 3D orientations.", + "bbox": [ + 511, + 247, + 906, + 292 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "(1) Multiple objects. Objects are fundamental to defining scene structure [52]. Large vision-language models show strong results in aligning semantic features with text [23, 31], handling multiple objects from scene becomes the first toward scene understanding. Models must recognize individual attributes and interpret relationships among them. Our dataset evaluates multi-object reasoning, such as comparison and counting. For example, in Fig. 1, the image shows two cyan objects, requiring the model to verify color simultarity and count cyan objects accurately.", + "(2) 2D locations. 2D spatial relationships are essential in prior image-based VQA tasks [18, 19]. In traditional multimodal models [20, 28, 49], these relationships can be learned through object detection training [8, 33]. However, in large-scale vision-language models, explicit location information is often absent from the latent representations. Following Li et al. [24], we examine four spatial relationships (left, right, front, back) between objects from 2D camera views as the first taks of spatial reasoning. Questions use these relations to filter targets, such as \"What is the shape of the object to the left of the chopper?\" (see Fig. 1).", + "(3) 3D locations. Beyond 2D relationships, 3D locations can incorporate depth, capturing spatial relationships that 2D alone cannot represent. This is crucial for tasks involving occlusions and complex spatial arrangements. Our dataset assesses the 3D location understanding abilities of model through questions about occlusions query as in [45]. For example, as shown in Fig. 1, questions may ask about an object's position relative to others in 3D space or whether one object is occluded, challenging models to judge distances and spatial positions accurately.", + "(4) 3D orientations. The 3D orientation of objects introduces an additional layer of complexity by incorporating rotational attributes. In current computer vision tasks, 3D orientation is often derived by pose estimation methods. However, large-scale vision-language models generally lack explicit 3D orientation information in their latent representations, potentially making it challenging to accurately infer the object's orientations. In our dataset, we eval" + ], + "bbox": [ + 511, + 294, + 906, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "24671", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "uate models' abilities to understand 3D orientations by focusing on object poses, including their facing direction and alignments with others. For example, as depicted in Fig. 1, the model may be asked to identify whether a cyan object is facing left, oriented parallel to a brown object, or positioned in a specific spatial relationship — i.e. left, right, front, or behind— from the perspective of the objects themselves.", + "bbox": [ + 89, + 90, + 483, + 196 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Scene Rendering", + "text_level": 1, + "bbox": [ + 89, + 205, + 256, + 220 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We build upon the generation pipeline from prior VQA datasets [19, 24] and uniformly select 21 object models across 5 categories to construct the dataset as Li et al. [24]. To improve the realism of our rendered dataset, we use real image style environment maps and refined object textures. For benchmarking new 6D spatial understanding ability, we add new 6D spatial relationship annotations among objects.", + "bbox": [ + 89, + 224, + 483, + 330 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Background. To create realistic scene backgrounds, we use HDRI images $^2$ with environment maps and natural lightings. We selected 224 different outdoor scenes with normal brightness levels without high light contrast by their metadata to maintain visual consistency across images.", + "bbox": [ + 89, + 330, + 483, + 406 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Objects. We reconstruct 3D mesh models [24] with realistic textures, introducing more variation and realistic appearances for objects. Similar to Li et al. [24], each object in our dataset is defined by distinct attributes and precise spatial locations. The original attributes include shape, color, and size. For benchmarking 6D spatial reasoning, each object also includes a vector specifying its 2D location $(\\mathbf{x},\\mathbf{y})$ , 3D world coordinates $(\\mathrm{X},\\mathrm{Y},\\mathrm{Z})$ , and pose direction derived from its 3D orientation [45].", + "bbox": [ + 89, + 407, + 483, + 542 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Relationships. As shown in Fig. 3 (a.1 to a.3), we define three types of spatial relationships: 1. 2D Spatial Relationships: These describe the relative positioning of objects within the 2D camera view, using terms such as left, right, front, and behind. 2. 6D Spatial Relationships: Extending spatial descriptions into 3D space, these relationships incorporate both the 3D position and orientation of objects to define relative positions from a target object's perspective, including left, right, front, and behind. 3. Collision Relationships: This relationship describes potential collisions if the target object moves forward or backward. Collision prediction involves assessing the orientation and 3D location of target objects relative to others to determine possible collisions.", + "bbox": [ + 89, + 545, + 483, + 755 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. New Questions for 6D Spatial Reasoning", + "text_level": 1, + "bbox": [ + 89, + 763, + 437, + 780 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Based on the four capabilities defined in Sec. 3.1, we construct new question types for 6D spatial reasoning and collision prediction. These questions require an understanding of both 3D location and 3D orientation across multiple objects to generate accurate answers. This level of complexity represents the L5 difficulty level shown in Fig. 1.", + "bbox": [ + 89, + 782, + 483, + 875 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/f6fcdc111cff4c02fce2f5207fed5572b3d475b0be5ee653357e91a05618a402.jpg", + "image_caption": [ + "Q: How many yellow objects are to the left of the red wagon? A: 1" + ], + "image_footnote": [], + "bbox": [ + 542, + 88, + 648, + 151 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/0a2a36af4790072fcdea3e45232b9148a8f57a15edfce4180dd57bbaa769eeed.jpg", + "image_caption": [ + "Q: How many yellow objects are on the left hand side of the red wagon? A: 3" + ], + "image_footnote": [], + "bbox": [ + 655, + 89, + 761, + 148 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/1871951b8fe3cb2330b8bf2b78e45e41803a43513be1565e8d953b2dead326b3.jpg", + "image_caption": [ + "Q: Is the red wagon will collide with a yellow regular bus if it moving forward?" + ], + "image_footnote": [], + "bbox": [ + 766, + 89, + 872, + 148 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/e5009a32a1bb030559d3c89e9a090fefbcb9f87618f8510654f493c0d8f6fc9d.jpg", + "image_caption": [ + "(a.1) 2D spatial", + "(b.1) Programs of 6D spatial question" + ], + "image_footnote": [], + "bbox": [ + 544, + 217, + 709, + 285 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/53f9d9d3ce6033dd56704d041a039483042b573f2b5f7e60b536bb6ba17b6dd0.jpg", + "image_caption": [ + "(a.2) 6D spatial", + "A: No", + "(a.3) Collision", + "(b.2) Programs of collision prediction question", + "Figure 3. (a.1-3) Example of the 2D spatial, 3D spatial and collision with images and questions. (b.1-2) shows the new operation programs we generated for the new 6D spatial reasoning questions." + ], + "image_footnote": [], + "bbox": [ + 718, + 217, + 869, + 297 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The 6D spatial relationship questions challenge the model to understand the spatial relationship involving both 3D position and orientation from a target object's perspective (shown in Fig. 3 (a.2)). Compared with prior VQA benchmarks focusing on 2D [19, 24], we introduce new 3D spatial reasoning operations, 3D Relate( $\\cdot$ ), to determine relative positions like left, right, front, and behind within a 3D coordinate space (see Fig. 3 (b.1)). For example, a question might ask, \"How many objects are to the right of the red wagon?\" requiring the model to filter and count objects based on their orientation and relative 3D position, which would yield different answers compared to similar 2D left and right questions (Fig. 3 (a.1)). This task is particularly challenging as it requires models to integrate 3D spatial relationships while managing occlusions and depth perception—capabilities not typically covered in simpler 2D spatial tasks.", + "bbox": [ + 511, + 386, + 906, + 643 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The collision prediction questions, as shown in Fig. 3 (a.3), extend 6D spatial reasoning by introducing motion-based inference. These questions assess the model's ability to anticipate future states and potential interactions (collisions) based on accurate estimation of 3D location and direction, denoted as forward or backward. For instance, a question might ask, \"Will the red wagon collide with a yellow bus if it moves forward?\" This requires the model to consider the orientation and direction of movement relative to surrounding objects to predict possible outcomes. Such tasks test the model's advanced spatial reasoning about 3D orientation and position, and its capacity to predict accurate trajectories and directions within the future scene.", + "bbox": [ + 511, + 643, + 908, + 839 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Questions under Other 4 Difficulty Levels", + "text_level": 1, + "bbox": [ + 511, + 848, + 870, + 864 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The L5 questions above represent the most challenging cases for 6D spatial reasoning, testing all four capabilities", + "bbox": [ + 511, + 869, + 906, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "2https://polyhaven.com/hdris", + "bbox": [ + 107, + 886, + 331, + 898 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "24672", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "in Sec. 3.1. However, the purpose of this benchmark is to comprehensively diagnose each factor in spatial reasoning. Thus, we design a structured roadmap to evaluate the model's performance across different levels of spatial reasoning, assessing its capabilities in each specific factor.", + "bbox": [ + 89, + 90, + 480, + 167 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "By composing the four factors of comprehensive spatial reasoning, we define an additional four levels of question difficulty based on the generated scenes, encompassing a total of seven tasks in our benchmark. As the complexity of spatial reasoning increases with the introduction of additional factors, we organize these questions into a structured roadmap, as illustrated in Fig. 1. Although many questions are reimplementations from previous benchmarks [24, 45] on our new images, we aim to integrate these into a unified benchmark to provide deeper insights into the model's spatial reasoning capabilities. The composition of the four capabilities at each level, along with comparisons to previous VQA benchmarks, is shown in Tab. 1.", + "bbox": [ + 89, + 169, + 482, + 364 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Level 1 - single object questions [19, 24]. We begin with the questions with single objects. These questions are generated using a filtered set of templates from Li et al. [24]. We select five templates involving only one object in the question, such as \"What color is the double bus?\" in Fig. 1. This question type provides a base score for the tested vision-language model. The comparison with L1 helps to determine whether the decrease of performance are due to challenge of spatial reasoning or object recognition.", + "bbox": [ + 89, + 366, + 482, + 501 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Level 2 - multiple objects questions [19, 24]. Level 2 questions extend Level 1 by involving multiple objects. These questions involve understanding the relationship between objects and counting. For example, \"Is there another object of the same color as the double bus?\" as shown in Fig. 1, where the question templates follow Li et al. [24]. We evaluate the model's capacity for reasoning in scenes with multiple entities as a critical step toward understanding spatial relationships in next levels.", + "bbox": [ + 89, + 503, + 482, + 640 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Level 3 - 2D spatial questions [19, 24]. Level 3 questions further introduce 2D spatial relationships among multiple objects, requiring the model to understand the position of objects and their relationships from the 2D camera's perspective. For example, \"What shape is the object to the left of the chopper?\" as depicted in Fig. 1. The 2D relationship is the first step of spatial reasoning.", + "bbox": [ + 89, + 642, + 482, + 747 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Level 4 - 3D pose questions [45]. Level 4 questions extend 3D orientation reasoning beyond the 2D spatial relationships in Level 3. Beyond appearance attributes (shape, color, size) or 2D locations, these questions require understanding 3D orientation, including the rotation angle or facing direction of an object from the camera's perspective. For example, \"What shape is the object that is parallel to the brown object?\" as shown in Fig. 1. 3D pose questions test the model's ability to interpret scenes with 3D awareness, involving rotation and alignment with other objects.", + "bbox": [ + 89, + 750, + 482, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Level 4 - occlusion questions [45]. In addition to 3D orientation, incorporating 3D location presents another challenge beyond the Level 3 questions on 2D spatial relationships. A unique aspect of 3D spatial reasoning from images is occlusion. The occlusion questions in Level 4 require the model to comprehend depth perception and the relative positions of objects in 3D space, determining which object is closer to the camera within overlapping areas. For example, an occlusion question might ask, \"Is the red wagon occluded by the yellow bus?\" as shown in Fig. 1. This level evaluates the model's ability to reason about occlusions and depth perception, which are essential for understanding complex spatial relationships in 3D scenes.", + "bbox": [ + 511, + 90, + 903, + 287 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/36adcd6fd157b7710556b58d3f68b1d6483d8b43e2f9a48b9b84a4d0efd8aa22.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Question LevelMulti-object2D location3D location3D OrientationVQA Benchmarks
L1spatial-agnostic bench (e.g., TextVQA [35])
L2✓CLEVR-Math [25], CLEVR-Count [19]
L3✓✓VQA [15], GQA [18], CLEVR [19], Super-CLEVR [24]
L4 (a)✓✓✓CV-Bench (DepthOrder) [40], Super-CLEVR-3D [45]
L4 (b)✓✓✓Super-CLEVR-3D [45]
L5✓✓✓✓Spatial457
", + "bbox": [ + 517, + 287, + 903, + 435 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1. Comparison of VQA Benchmarks by Factors and Question Levels. Compared with existing VQA benchmarks, Spatial457 includes all four spatial reasoning factors, providing a comprehensive evaluation of 6D spatial understanding.", + "bbox": [ + 511, + 446, + 903, + 502 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Evaluation Protocols", + "text_level": 1, + "bbox": [ + 513, + 507, + 709, + 522 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To better evaluate the model's spatial reasoning capabilities, we propose a structured evaluation protocol that measures the relative capability of models on each factor for spatial reasoning. We define this evaluation metric as the Relative Performance Dropping Rate (RPDR), which assesses the decline in performance after introducing a new factor in the questions. Based on the roadmap in Fig. 1, the RPDR for each factor is defined as follows:", + "bbox": [ + 511, + 523, + 903, + 635 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {R P D R} _ {\\text {M u l t i - O b j .}} = A c c _ {\\text {L e v e l 1}} \\Rightarrow A c c _ {\\text {L e v e l 2}}\n$$\n", + "text_format": "latex", + "bbox": [ + 526, + 640, + 764, + 655 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {R P D R} _ {\\mathrm {2 D L o c .}} = A c c _ {\\text {L e v e l 2}} \\Rightarrow A c c _ {\\text {L e v e l 3}}\n$$\n", + "text_format": "latex", + "bbox": [ + 537, + 657, + 764, + 671 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {R P D R} _ {\\mathrm {3 D L o c .}} = \\operatorname {A v g} \\left( \\begin{array}{c} A c c _ {\\text {L e v e l 3}} \\Rightarrow A c c _ {\\text {L e v e l 4 (O c c .)}}, \\\\ A c c _ {\\text {L e v e l 4 (3 D P o s e)}} \\Rightarrow A c c _ {\\text {L e v e l 5 (6 D)}} \\end{array} \\right)\n$$\n", + "text_format": "latex", + "bbox": [ + 535, + 676, + 888, + 705 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {R P D R} _ {\\mathrm {3 D O r i .}} = \\operatorname {A v g} \\left( \\begin{array}{c} A c c _ {\\text {L e v e l 3}} \\Rightarrow A c c _ {\\text {L e v e l 4 (3 D P o s e)}}, \\\\ A c c _ {\\text {L e v e l 4 (O c c .)}} \\Rightarrow A c c _ {\\text {L e v e l 5 (C o l .)}} \\end{array} \\right)\n$$\n", + "text_format": "latex", + "bbox": [ + 537, + 708, + 879, + 739 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $Acc_{\\mathrm{Level}}$ denotes the model's accuracy on the corresponding questions and $(Acc_{\\mathrm{Q1}} \\Rightarrow Acc_{\\mathrm{Q2}}) = \\frac{Acc_{\\mathrm{Q2}} - Acc_{\\mathrm{Q1}}}{Acc_{\\mathrm{Q1}}}$ denotes the relative performance dropping rate from $Q1$ to $Q2$ . The results of RPDR will be discussed in Sec. 4.3.", + "bbox": [ + 511, + 746, + 903, + 808 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 816, + 645, + 832 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 511, + 837, + 702, + 853 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We generate 1,000 images to test the performance of large vision-language models. For the seven types of questions, we generate 1000 single-object questions (L1-Single), 5,000", + "bbox": [ + 511, + 854, + 903, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "24673", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/f29d6be82835acb5026359d636086e59a52279a26b4c6558c545e3449a152355.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Difficulty LevelLevel 1Level 2Level 3Level 4Level 5
Questions typeSingle ObjectMulti-Obj.2D SpatialOcclusion3D PoseCollisions6D Spatial
Random33.0532.7733.4722.0418.9921.0219.41
GPT-4o [29]74.4662.8856.1448.4042.4138.4137.01
GeminiPro 1.5 [37]73.2662.5454.4947.6543.6741.1939.36
Claude 3.5 Sonnet [3]68.2457.4054.1930.8438.4035.3433.48
Qwen2-VL-7B-Inst. [39]71.9661.4455.3427.8734.2936.5833.75
InternVL2-8B [38]58.1158.7657.4033.2532.3234.3034.30
LLaVA-v1.5-7B [26]44.8744.7242.2024.3424.5523.6323.86
LLaVA-NeXT-vicuna-7B [27]50.7249.4746.0129.6829.3531.9531.95
LLaVA-NeXT-Ilama3-8B [41]52.1549.7345.9230.3129.7732.1232.12
PO3D-VQA [45]86.4682.5580.6470.4981.4068.1271.06
Human89.9786.8384.9582.7684.9581.8279.94
", + "bbox": [ + 109, + 88, + 890, + 310 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Performance comparison across all 7 question types under 5 difficulty levels. As task complexity increases from single-object questions to 6D spatial relationships, performance generally decreases in most cases, highlighting the challenges in handling multi-object interactions, 3D orientations, and predictive spatial reasoning. The highest score among all models for each question type is marked in red , and the second-highest scores are underlined. The best performance among open-source models is marked in blue . PO3D-VQA is the state-of-the-art (SOTA) 3D-aware neural symbolic reasoning model as proposed in [45], and is listed as the potential upper bound of evaluated methods. The comparison with human performance also shows an increasing gap for LLMs as task difficulty increases.", + "bbox": [ + 89, + 320, + 906, + 407 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "multiple-object questions (L2-Multi-object.), 4,995 2D spatial relationship questions (L3-2D-Spatial), 3,534 occlusion questions (L4-Occlusion), 4,555 3D pose questions (L4-3D-Pose), 4,357 collision questions (L5-Collision), and 4,995 6D spatial relationship questions (L5-6D-Spatial).", + "bbox": [ + 88, + 431, + 480, + 508 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We evaluate two series of large vision-language models on Spatial457: (1) API-based large language models, including GPT-4o [29], Gemini-1.5-Pro [37], and Claude-3.5-Sonnet [3]; and (2) open-source large vision-language models, including Qwen2-VL-7B-Instruct [39], InternVL2-8B [38], llava-v1.5-7b [26], llava next-vicuna-7b [27], and llava next-llama3-8b [41]. The evaluation follows the format of Vlmealkit [12]. The details of the model implementation including prompts are listed in Appendix.", + "bbox": [ + 88, + 508, + 482, + 645 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Evaluation Results", + "text_level": 1, + "bbox": [ + 89, + 654, + 272, + 667 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In Tab. 2, we present the performance of various large vision-language models across different levels of difficulty on Spatial457. The levels range from basic single-object tasks to complex 6D spatial reasoning, with increasing difficulty as more factors are introduced. We compare both API-based models (e.g., GPT-4o, GeminiPro1.5, Claude3-5V) and open-source models (e.g., Qwen2-VL-7B-Instruct, InternVL2-8B, LLaVA series) to assess their spatial reasoning capabilities.", + "bbox": [ + 89, + 672, + 482, + 808 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Comparison across difficulty levels. We evaluate model performance across different levels of difficulty in Spatial457, ranging from basic single-object tasks to complex 6D spatial reasoning. As shown in Tab. 2, performance generally decreases as the complexity of spatial relationships increases, highlighting the challenges models face", + "bbox": [ + 89, + 810, + 483, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "when handling multi-object interactions, 3D orientations, and predictive spatial reasoning tasks.", + "bbox": [ + 511, + 431, + 903, + 462 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Comparison between API models and open-source models. From Tab. 2, we observe that API models generally outperform open-source models across all difficulty levels. In 2D spatial relationship questions, InternVL2-8B achieves slightly higher results than the API models, while GPT-4o is only $1\\%$ lower. Qwen2-VL-7B-Instruct achieves comparable results, especially in simpler tasks like single-object and multiple-object reasoning. This analysis highlights the strengths of API models in complex spatial reasoning tasks.", + "bbox": [ + 511, + 463, + 906, + 599 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Comparison among API models. API-based models, i.e. GPT-4o, GeminiPro1.5, and Claude3-5V, demonstrate strengths across all difficulty levels. For instance, GPT-4o excels in 2D and 3D spatial tasks, while GeminiPro1.5 shows robust performance in collision prediction tasks. This comparison provides insights into the unique capabilities and limitations of each API model in handling spatial reasoning across different levels of complexity.", + "bbox": [ + 511, + 599, + 905, + 720 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. RPDR Analysis for 4 Capabilities", + "text_level": 1, + "bbox": [ + 511, + 729, + 805, + 744 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As outlined in Sec. 3.5, we use the Relative Performance Dropping Rate (RPDR) to analyze model performance across the four core factors of Spatial457: multi-object reasoning, 2D location understanding, 3D location comprehension, and 3D orientation interpretation. The RPDR analysis provides a structured approach to quantify the decline in accuracy with the introduction of each new factor, as detailed in Tab. 3. Our findings reveal that all models exhibit weaknesses in 3D reasoning. Specifically, GPT-4o and GeminiPro-1.5 perform the worst in 3D orientation tasks,", + "bbox": [ + 511, + 750, + 906, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "24674", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "while the other two struggle most with 3D location.", + "bbox": [ + 91, + 90, + 434, + 106 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/45a20e07cf0ec03b6ed0a9443cd4535e70aaa84db21deaab709ff59b57cc9cb1.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelMulti.2D Loc.3D Ori.3D Loc.
GPT-4o [29]84.4589.2777.4686.74
GeminiPro 1.5 [37]85.3787.1383.2888.79
Claude 3.5 S. [3]84.1294.4185.4372.05
Qwen2-VL-7B [39]85.3990.0680.9974.40
", + "bbox": [ + 102, + 117, + 472, + 194 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Bias Analysis", + "text_level": 1, + "bbox": [ + 89, + 257, + 230, + 272 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we analyze prediction biases in large language models (LLMs) regarding their understanding of key attributes, specifically shape, color, size, and pose. Leveraging the unbiased generation of dataset images and questions in Spatial457, we examine how these models may demonstrate attribute-specific biases in their predictions.", + "bbox": [ + 89, + 273, + 483, + 363 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We first compute the confusion matrices and distribution statistics for both ground truth and model predictions across each attribute (shape, color, size, and pose). For example, the confusion matrix plots for color and pose attributes of GPT-4o and GeminiPro1.5, together with the distribution of prediction and ground truth label are shown in Fig. 4. We find both GPT-4o and GeminiPro1.5 exhibit bias in both color and pose predictions. These results highlight areas for improvement to reduce bias and enhance predictive reliability across models.", + "bbox": [ + 88, + 363, + 483, + 513 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To quantify prediction variability, we calculate the Coefficient of Variation (CV) for each attribute as follows:", + "bbox": [ + 89, + 513, + 483, + 542 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {C V} = \\frac {\\operatorname {S t d} \\left(p _ {1} , \\dots , p _ {k}\\right)}{\\operatorname {M e a n} \\left(p _ {1} , \\dots , p _ {k}\\right)}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 204, + 547, + 482, + 577 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\operatorname{Std}(\\cdot)$ and $\\operatorname{Mean}(\\cdot)$ are the standard deviation and mean of the probability distribution $p_i$ for each attribute label $i$ (e.g., left), \"right\" for pose).", + "bbox": [ + 89, + 579, + 482, + 622 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Tab. 4 shows the CV values for shape, color, size, and pose attributes across different models. Lower CV values indicate more consistent predictions with less bias, which reveals significant biases in model predictions across attributes. Higher CV values, especially in color and pose, indicate consistency challenges.", + "bbox": [ + 89, + 623, + 483, + 714 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Comparison to the 3D Symbolic Model", + "text_level": 1, + "bbox": [ + 89, + 717, + 424, + 733 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As neural symbolic models achieve state-of-the-art (SOTA) performance on many prior synthetic datasets, we implement the 3D-aware neural symbolic model PO3DVQA [45] on our new image domain and all levels of questions. We train PO3D-VQA on a separate set of 20,000 images and questions and report its performance on the same test dataset in Tab. 2. A direct comparison between the 3D neural symbolic model and large language models is not entirely fair, as each module of the neural symbolic model (6D pose estimation, attribute classification, and language reasoning) is trained independently as a sub-task. However, we", + "bbox": [ + 89, + 734, + 483, + 900 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1269a9c5dd214d8344a48f511976c9e6edf5d8b671ec81807079c72f0fcb378c.jpg", + "image_caption": [ + "(b) GeminiPro1-5", + "Figure 4. The distribution of color and pose attributes in the L4-3D-Pose task for GPT-4o and Gemini-Pro1.5. Although the ground truth labels are well balanced, the predicted colors and poses for both models are imbalanced, with \"yellow\" being the most frequent color and \"front\" the most frequent pose." + ], + "image_footnote": [], + "bbox": [ + 535, + 90, + 883, + 366 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/9982cddff640840295fbb2d9c80fad34199c2d6f675254e41a027490fa9986d0.jpg", + "table_caption": [ + "Table 3. RPDR (%) analysis on the four spatial reasoning capabilities, where higher values indicate better performance. For each row, we highlight the lowest RPDR score among all factors." + ], + "table_footnote": [], + "table_body": "
ModelsAvg. ↓ShapeColorSizePose
GPT4o [29]0.4850.6230.5210.0430.752
GeminiPro1.5 [37]0.4980.5670.3660.2970.760
Claude3.5V-S [3]0.7200.7770.4650.5741.063
Qwen2-VL-7B [39]0.5450.8680.5580.0670.688
Human0.0930.1100.0820.0550.126
", + "bbox": [ + 517, + 481, + 903, + 577 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. Coefficient of Variation (CV) for predictions on each attribute in the L4-3D-Pose task. We evaluate three API models and one open-source model, calculating the CV across four attributes (shape, color, size, 3D pose) based on the distribution of predicted labels. Lower CV values indicate less bias. The first column shows the average CV across attributes, and GPT4-o has the lowest bias.", + "bbox": [ + 511, + 585, + 905, + 670 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "include this model to illustrate the upper bound that a data-driven machine learning model can achieve on our dataset.", + "bbox": [ + 511, + 691, + 903, + 722 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.6. Qualitative Results", + "text_level": 1, + "bbox": [ + 511, + 731, + 696, + 746 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Fig. 5 presents an example of GPT-4o's performance across various spatial reasoning tasks on the same image. GPT-4o correctly answers multi-object, 2D spatial, and occlusion tasks but fails on 3D pose and 6D spatial questions, indicating challenges in 3D orientation understanding in this case. As we prompt the models to output the answer as well as the reasoning steps, the models correctly detect the existence of objects (e.g. in Fig. 5 the yellow minivan in L4-Pose or the fighter plane and road bike in L5 questions, but has the wrong understand to the orientations or 3D locations.", + "bbox": [ + 511, + 750, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "24675", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7fa982bf2e5866389df7429e32bea4da7f224784da2851f9cb02347b8566918b.jpg", + "image_caption": [ + "Figure 5. Example of GPT-4o across different spatial reasoning tasks on the same image with the ground truth answer marked in blue. GPT-4o answers multi-object, 2D spatial, and occlusion tasks correctly (in the green block) but fails on 3D pose and 6D spatial questions (the red block)." + ], + "image_footnote": [], + "bbox": [ + 109, + 88, + 460, + 258 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Extension to Real Dataset", + "text_level": 1, + "bbox": [ + 89, + 359, + 331, + 375 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We extend the setting of 3D questions to real-world image data. Due to limitations in fully annotated attributes such as category, color, 3D orientation, position, and occlusion, we focus on questions from the L4-Pose category. Using the scene structure from SUN-RGBD [36], we construct a new subset, L4-Pose-Real, which involves only the category and pose attributes. To reduce ambiguity in object localization, we add a red bounding box around target objects, ensuring clarity in identifying categories within the scene.", + "bbox": [ + 89, + 378, + 482, + 513 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We provide the question along with the image as input to the LLM. In this example in Fig. 6, GPT-4o provides the correct answer, while Gemini-Pro 1.5 is incorrect. However, the reasoning process reveals that both models rely on common sense knowledge or 2D visibility cues to interpret the 3D orientation.", + "bbox": [ + 89, + 513, + 483, + 604 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/760ee8a6c6463ba8e962e210d939301f889828d4dfff2cc5f58051e282124c79.jpg", + "image_caption": [ + "Figure 6. Example of real-world 3D pose reasoning tasks $L4$ -Pose-Real. Here we show the reasoning and answers provided by LLMs where GPT-4o is correct and Gemini-Pro 1.5 is wrong. But the reasoning process reveals these two LLMs are both reasoning 3D orientation from the command sense or 2D visibility." + ], + "image_footnote": [], + "bbox": [ + 109, + 614, + 457, + 772 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As shown in Tab. 5, Gemini Pro 1.5 achieves the highest accuracy (40.50%) in L4-Pose-Real, and the gap between the three API-based models and the open-source", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "model Qwen2-VL-7B-Instruct (26.50%) is larger. However, all these models perform significantly lower than humans, which have accuracy 92.64%. We further analyze the bias of predictions in this real-world question set. Among the models, Claude3-5V Sonnet exhibits the lowest CV (0.374), indicating relatively stable predictions, while Qwen2-VL-7B-Instruct has the highest CV (0.934). The distribution of ground truth and predicted pose and the confusion matrix are shown in Fig. 7.", + "bbox": [ + 511, + 90, + 906, + 227 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/8bd8ef8160a6e4fff4b2a8811718d82fbcf814436744faa44f68f80ba56bda0d.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelsAccuracy (%) ↑CV ↓
GPT-4o [29]38.750.690
GeminiPro1-5 [37]40.500.762
Claude3-5V Sonnet [3]38.750.374
Qwen2-VL-7B-Instruct [39]26.500.934
Human92.640.197
", + "bbox": [ + 526, + 244, + 890, + 351 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. Accuracy and Coefficient of Variation (CV) for L4-Poserel task: Comparison of accuracy and CV for GPT-4o, GeminiPro1.5, Claude3-5V, and Qwen2-VL-7B-Instruct on real-world 3D pose tasks. Human performance is included as a reference.", + "bbox": [ + 511, + 352, + 906, + 409 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a544513755b30e876e51395ca63b4f69cbfe29077150cfa24ebd01acf46cd70d.jpg", + "image_caption": [ + "Figure 7. Distribution of 3D poses for in L4-Pose-Real task." + ], + "image_footnote": [], + "bbox": [ + 535, + 436, + 888, + 588 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 637, + 633, + 652 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we introduce Spatial457, a scalable and unbiased benchmark designed to comprehensively assess 6D spatial reasoning capabilities in LMMs. Spatial457 extends traditional 2D and 3D VQA benchmarks by incorporating complex 6D spatial reasoning tasks, structured across five levels of difficulty. Through evaluations, we observe that current LMMs demonstrate strong performance in basic object recognition and 2D spatial relationships, but struggle significantly with 3D spatial understanding and collision prediction tasks. Additionally, our analysis of prediction biases uncovers attribute-specific inconsistencies that could affect model reliability in real-world applications. Limitations. To real images settings, current evaluation limits to L3 pose questions. We are actively working on building the additional required annotations for these real images to expand the domain of the proposed benchmark.", + "bbox": [ + 511, + 657, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "24676", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 91, + 90, + 250, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work is supported by ONR with award N00014-23-1-2641, from ARL Army Research Laboratory with award W911NF2320008.", + "bbox": [ + 91, + 114, + 482, + 159 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 172, + 187, + 188 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Alibaba DAMO Academy. Gemini: Vision-language model by alibaba. Alibaba DAMO, 2023.", + "[2] Boshi An, Yiran Geng, Kai Chen, Xiaqi Li, Qi Dou, and Hao Dong. Rgbmanip: Monocular image-based robotic manipulation through active object pose estimation. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 7748-7755. IEEE, 2024.", + "[3] Anthropic. Claude: An anthropic language model. Anthropotic, 2023.", + "[4] Gilad Baruch, Zhuoyuan Chen, Afshin Dehghan, Tal Dimry, Yuri Feigin, Peter Fu, Thomas Gebauer, Brandon Joffe, Daniel Kurz, Arik Schwartz, et al. Arkitsscenes: A diverse real-world dataset for 3d indoor scene understanding using mobile rgb-d data. arXiv preprint arXiv:2111.08897, 2021.", + "[5] Mark Billinghurst, Adrian Clark, and Gun Lee. A survey of augmented reality. Foundations and Trends in Human-Computer Interaction, 8(2-3):73-272, 2015.", + "[6] Garrick Brazil, Abhinav Kumar, Julian Straub, Nikhila Ravi, Justin Johnson, and Georgia Gkioxari. Omni3d: A large benchmark and model for 3d object detection in the wild. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13154-13164, 2023.", + "[7] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11621-11631, 2020.", + "[8] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020.", + "[9] Boyuan Chen, Zhuo Xu, Sean Kirmani, Brain Ichter, Dorsa Sadigh, Leonidas Guibas, and Fei Xia. Spatialvm: Endowing vision-language models with spatial reasoning capabilities. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14455-14465, 2024.", + "[10] An-Chieh Cheng, Hongxu Yin, Yang Fu, Qiushan Guo, Ruihan Yang, Jan Kautz, Xiaolong Wang, and Sifei Liu. Spatial-rgpt: Grounded spatial reasoning in vision language model. arXiv preprint arXiv:2406.01584, 2024.", + "[11] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhanghao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E Gonzalez, et al. Vicuna: An open-source chatbot impressing gpt-4 with $90\\%$ chatgpt quality, march 2023. URL https://lmsys.org/blog/2023-03-30-vicuna, 3(5), 2023.", + "[12] Haodong Duan, Junming Yang, Yuxuan Qiao, Xinyu Fang, Lin Chen, Yuan Liu, Xiaoyi Dong, Yuhang Zang, Pan Zhang," + ], + "bbox": [ + 93, + 198, + 483, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Jiaqi Wang, et al. Vlmealkit: An open-source toolkit for evaluating large multi-modality models. In Proceedings of the 32nd ACM international conference on multimedia, pages 11198-11201, 2024.", + "[13] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. International Journal of Computer Vision, 88(2):303-338, 2010.", + "[14] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 3354-3361, 2012.", + "[15] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Batra, and Devi Parikh. Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6904-6913, 2017.", + "[16] Roland Hess. *Blender foundations: The essential guide to learning blender* 2.5. Routledge, 2013.", + "[17] Yushi Hu, Hang Hua, Zhengyuan Yang, Weijia Shi, Noah A Smith, and Jiebo Luo. Promptcap: Prompt-guided task-aware image captioning. arXiv preprint arXiv:2211.09699, 2022.", + "[18] Drew A Hudson and Christopher D Manning. Gqa: A new dataset for real-world visual reasoning and compositional question answering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6700-6709, 2019.", + "[19] Justin Johnson, Bharath Hariharan, Laurens Van Der Maaten, Li Fei-Fei, C Lawrence Zitnick, and Ross Girshick. Clevr: A diagnostic dataset for compositional language and elementary visual reasoning. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2901–2910, 2017.", + "[20] Aishwarya Kamath, Mannat Singh, Yann LeCun, Gabriel Synnaeve, Ishan Misra, and Nicolas Carion. Mdetr-modulated detection for end-to-end multi-modal understanding. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1780-1790, 2021.", + "[21] Dong-Jin Kim, Jinsoo Choi, Tae-Hyun Oh, and In So Kweon. Dense relational captioning: Triple-stream networks for relationship-based captioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6271-6280, 2019.", + "[22] Alina Kuznetsova, Hassan Rom, Neil Alldrin, Jasper Uijlings, Ilya Krasin, Jordi Pont-Tuset, Gleb Muryshkin, Sergey Popov, Marwa Elhachem, Pedro de Carvalho, et al. The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale. International Journal of Computer Vision, 128(7):1956-1981, 2020.", + "[23] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730-19742. PMLR, 2023." + ], + "bbox": [ + 516, + 92, + 906, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "24677", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[24] Zhuowan Li, Xingrui Wang, Elias Stengel-Eskin, Adam Kortylewski, Wufei Ma, Benjamin Van Durme, and Alan L Yuille. Super-clevr: A virtual benchmark to diagnose domain robustness in visual reasoning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14963-14973, 2023.", + "[25] Adam Dahlgren Lindstrom and Savitha Sam Abraham. Clevr-math: A dataset for compositional language, visual and mathematical reasoning. arXiv preprint arXiv:2208.05358, 2022.", + "[26] Haotian Liu et al. Llava: Large language and vision assistant. LLaVA Project, 2023.", + "[27] Haotian Liu et al. Llava-last: Extended large language and vision assistant. LLaVA Project, 2023.", + "[28] Jiayuan Mao, Chuang Gan, Pushmeet Kohli, Joshua B Tenenbaum, and Jiajun Wu. The neuro-symbolic concept learner: Interpreting scenes, words, and sentences from natural supervision. arXiv preprint arXiv:1904.12584, 2019.", + "[29] OpenAI. Gpt-4 technical report. OpenAI, 2023.", + "[30] Weichao Qiu and Alan Yuille. Unrealcv: Connecting computer vision to unreal engine. In Computer Vision-ECCV 2016 Workshops: Amsterdam, The Netherlands, October 8-10 and 15-16, 2016, Proceedings, Part III 14, pages 909-916. Springer, 2016.", + "[31] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021.", + "[32] Arjit Ray, Jiafei Duan, Reuben Tan, Dina Bashkirova, Rose Hendrix, Kiana Ehsani, Aniruddha Kembhavi, Bryan A Plummer, Ranjay Krishna, Kuo-Hao Zeng, et al. Sat: Spatial aptitude training for multimodal language models. arXiv preprint arXiv:2412.07755, 2024.", + "[33] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. IEEE transactions on pattern analysis and machine intelligence, 39(6):1137-1149, 2016.", + "[34] Andrew Sanders. An introduction to Unreal engine 4. AK Peters/CRC Press, 2016.", + "[35] Amanpreet Singh, Vivek Natarajan, Meet Shah, Yu Jiang, Xinlei Chen, Dhruv Batra, Devi Parikh, and Marcus Rohrbach. Towards vqa models that can read. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8317-8326, 2019.", + "[36] Shuran Song, Samuel P Lichtenberg, and Jianxiong Xiao. Sun rgb-d: A rgb-d scene understanding benchmark suite. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 567-576, 2015.", + "[37] Gemini Team, Petko Georgiev, Ving Ian Lei, Ryan Burnell, Libin Bai, Anmol Gulati, Garrett Tanzer, Damien Vincent, Zhufeng Pan, Shibo Wang, et al. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context. arXiv preprint arXiv:2403.05530, 2024.", + "[38] Intern Team. Internvl 2.0: A vision-language model. Intern Foundation, 2023." + ], + "bbox": [ + 91, + 90, + 485, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[39] Qwen Team. Qwen-vl: A large multimodal model. Alibaba Group, 2023.", + "[40] Shengbang Tong, Ellis Brown, Penghao Wu, Sanghyun Woo, Manoj Middepogu, Sai Charitha Akula, Jihan Yang, Shusheng Yang, Adithya Iyer, Xichen Pan, et al. Cambrian1: A fully open, vision-centric exploration of multimodal llms. arXiv preprint arXiv:2406.16860, 2024.", + "[41] Hugo Touvron et al. Llama: Open and efficient foundation language models. Meta AI, 2023.", + "[42] Oriol Vinyals, Alexander Toshev, Samy Bengio, and Dumitru Erhan. Show and tell: A neural image caption generator. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3156-3164, 2015.", + "[43] He Wang, Srinath Sridhar, Jingwei Huang, Julien Valentin, Shuran Song, and Leonidas J Guibas. Normalized object coordinate space for category-level 6d object pose and size estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2642-2651, 2019.", + "[44] Xingrui Wang, Wufei Ma, Zhuowan Li, Adam Kortylewski, and Alan Yuille. 3d-aware visual question answering about parts, poses and occlusions. In Thirty-seventh Conference on Neural Information Processing Systems, 2023.", + "[45] Xingrui Wang, Wufei Ma, Zhuowan Li, Adam Kortylewski, and Alan L Yuille. 3d-aware visual question answering about parts, poses and occlusions. Advances in Neural Information Processing Systems, 36, 2024.", + "[46] Yu Xiang, Roozbeh Mottaghi, and Silvio Savarese. Beyond Pascal: A benchmark for 3d object detection in the wild. In IEEE winter conference on applications of computer vision, pages 75-82. IEEE, 2014.", + "[47] Kelvin Xu, Jimmy Ba, Ryan Kiros, Kyunghyun Cho, Aaron Courville, Ruslan Salakhudinov, Rich Zemel, and Yoshua Bengio. Show, attend and tell: Neural image caption generation with visual attention. In International conference on machine learning, pages 2048-2057. PMLR, 2015.", + "[48] Lin Xu, Yilin Zhao, Daquan Zhou, Zhijie Lin, See Kiong Ng, and Jiashi Feng. Pllava: Parameter-free llava extension from images to videos for video dense captioning. arXiv preprint arXiv:2404.16994, 2024.", + "[49] Kexin Yi, Jiajun Wu, Chuang Gan, Antonio Torralba, Pushmeet Kohli, and Josh Tenenbaum. Neural-symbolic vqa: Disentangling reasoning from vision and language understanding. Advances in neural information processing systems, 31, 2018.", + "[50] Kexin Yi, Chuang Gan, Yunzhu Li, Pushmeet Kohli, Jiajun Wu, Antonio Torralba, and Joshua B Tenenbaum. Clevrer: Collision events for video representation and reasoning. arXiv preprint arXiv:1910.01442, 2019.", + "[51] Xiang Yue, Yuansheng Ni, Kai Zhang, Tianyu Zheng, Ruoqi Liu, Ge Zhang, Samuel Stevens, Dongfu Jiang, Weiming Ren, Yuxuan Sun, et al. Mmmu: A massive multi-discipline multimodal understanding and reasoning benchmark for expert agi. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9556-9567, 2024." + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "24678", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[52] Alan Yuille and Daniel Kersten. Vision as bayesian inference: analysis by synthesis? Trends in cognitive sciences, 10(7):301-308, 2006.", + "[53] Jiyao Zhang, Weiyao Huang, Bo Peng, Mingdong Wu, Fei Hu, Zijian Chen, Bo Zhao, and Hao Dong. Omni6dpose: A benchmark and model for universal 6d object pose estimation and tracking. In European Conference on Computer Vision, pages 199-216. Springer, 2024." + ], + "bbox": [ + 91, + 92, + 480, + 204 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "24679", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Spatial457_ A Diagnostic Benchmark for 6D Spatial Reasoning of Large Mutimodal Models/96f4562f-945b-497d-9468-cd476fb5f769_model.json b/2025/Spatial457_ A Diagnostic Benchmark for 6D Spatial Reasoning of Large Mutimodal Models/96f4562f-945b-497d-9468-cd476fb5f769_model.json new file mode 100644 index 0000000000000000000000000000000000000000..182366a533fa0c90c3709e78fa985ebe03cc3493 --- /dev/null +++ b/2025/Spatial457_ A Diagnostic Benchmark for 6D Spatial Reasoning of Large Mutimodal Models/96f4562f-945b-497d-9468-cd476fb5f769_model.json @@ -0,0 +1,2532 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.182, + 0.131, + 0.819, + 0.177 + ], + "angle": 0, + "content": "Spatial457: A Diagnostic Benchmark for 6D Spatial Reasoning of Large Multimodal Models" + }, + { + "type": "text", + "bbox": [ + 0.125, + 0.203, + 0.88, + 0.223 + ], + "angle": 0, + "content": "Xingrui Wang\\(^{1}\\), Wufei Ma\\(^{1}\\), Tiezheng Zhang\\(^{1}\\), Celso M de Melo\\(^{2}\\), Jieneng Chen\\(^{1\\dagger}\\), Alan Yuille\\(^{1\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.29, + 0.224, + 0.705, + 0.24 + ], + "angle": 0, + "content": "\\(^{1}\\)Johns Hopkins University \\(^{2}\\)DEVCOM Army Research Laboratory" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.274, + 0.327, + 0.291 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.307, + 0.485, + 0.685 + ], + "angle": 0, + "content": "Although large multimodal models (LMMs) have demonstrated remarkable capabilities in visual scene interpretation and reasoning, their capacity for complex and precise 3-dimensional spatial reasoning remains uncertain. Existing benchmarks focus predominantly on 2D spatial understanding and lack a framework to comprehensively evaluate 6D spatial reasoning across varying complexities. To address this limitation, we present Spatial457, a scalable and unbiased synthetic dataset designed with 4 key capability for spatial reasoning: multi-object recognition, 2D location, 3D location, and 3D orientation. We develop a cascading evaluation structure, constructing 7 question types across 5 difficulty levels that range from basic single object recognition to our new proposed complex 6D spatial reasoning tasks. We evaluated various large multimodal models (LMMs) on Spatial457, observing a general decline in performance as task complexity increases, particularly in 3D reasoning and 6D spatial tasks. To quantify these challenges, we introduce the Relative Performance Dropping Rate (RPDR), highlighting key weaknesses in 3D reasoning capabilities. Leveraging the unbiased attribute design of our dataset, we also uncover prediction biases across different attributes, with similar patterns observed in real-world image settings. The code is released in https://github.com/XingruiWang/Spatial457." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.724, + 0.222, + 0.74 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.743, + 0.483, + 0.866 + ], + "angle": 0, + "content": "In recent years, large multimodal models (LMMs) have demonstrated remarkable reasoning capabilities, enabling them to interpret visual scenes from images and articulate their understanding through natural language [1, 3, 11, 26, 27, 38, 39, 48]. These advancements have been evaluated using a variety of benchmarks that primarily emphasize semantic features—such as object categories and appearances, spatial relationships within the 2D image plane [15," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.276, + 0.907, + 0.367 + ], + "angle": 0, + "content": "17, 18] (e.g. left-right positioning from the camera view), and limited 2.5D features such as camera depth [40]. Tasks like object detection [13, 22], image captioning [21, 42, 47] and visual question answering [18, 50, 51] have significantly benefited from these evaluations, driving LMMs to new heights in visual understanding." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.368, + 0.909, + 0.49 + ], + "angle": 0, + "content": "However, a critical gap remains in assessing the ability of LMMs to comprehend and reason about the full six degrees of freedom (6D) of objects in spatial environments, which includes both 3D positions and orientations [43, 46, 53]. Despite its significance in many real-world applications, such as robotics [2], autonomous navigation [14], and augmented reality [5], there are no available benchmarks for LMMs that evaluate 6D spatial reasoning capabilities." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.491, + 0.909, + 0.688 + ], + "angle": 0, + "content": "Building a scalable diagnostic dataset for 6D vision tasks poses considerable challenges due to the extensive annotations required across vast image collections. One strategy is to develop 6D vision question-answering benchmarks on top of existing image datasets [4, 6, 7, 36]. However, real-world image collections often exhibit significant biases in 3D locations and poses, influenced by camera perspective limitations and the prevalence of common object orientations in everyday scenes. For instance, statistics from the SUN-RGBD [36] and nuScenes [7] datasets reveal that over \\(70\\%\\) of objects cluster into a single predominant direction (see Fig. 2), limiting the diversity required for comprehensive 6D reasoning evaluation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.909, + 0.903 + ], + "angle": 0, + "content": "To overcome these challenges, we introduce an unbiased and scalable synthetic dataset rendered in a realistic style, specifically designed to diagnose the 6D spatial reasoning abilities of LMMs. To comprehensively evaluate diverse aspects of 6D spatial reasoning, we focus on four core capabilities: multi-object recognition, 2D locations, 3D locations, and 3D orientation. Multi-object recognition serves as the foundation of scene understanding, while 2D location represents the simplest way to depict the spatial relationship between objects within the image plane [18, 19]. The 3D location factor extends understanding into three-dimensional space, capturing the depth and distance of objects, which is essential for tasks such as reasoning about occlusions [9, 10, 44]. The 3D orientation introduces object" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.875, + 0.401, + 0.888 + ], + "angle": 0, + "content": "1Approved for public release: distribution is unlimited." + }, + { + "type": "page_footnote", + "bbox": [ + 0.117, + 0.889, + 0.214, + 0.901 + ], + "angle": 0, + "content": "\\(\\dagger\\) : Equal advising." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "24669" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.145, + 0.093, + 0.327, + 0.197 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.165, + 0.204, + 0.187, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.193, + 0.207, + 0.292, + 0.219 + ], + "angle": 0, + "content": "3D Orientations" + }, + { + "type": "image", + "bbox": [ + 0.163, + 0.235, + 0.189, + 0.254 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.194, + 0.24, + 0.274, + 0.251 + ], + "angle": 0, + "content": "3D Locations" + }, + { + "type": "image", + "bbox": [ + 0.166, + 0.271, + 0.186, + 0.286 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.194, + 0.273, + 0.274, + 0.284 + ], + "angle": 0, + "content": "2D Locations" + }, + { + "type": "image", + "bbox": [ + 0.164, + 0.305, + 0.188, + 0.317 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.195, + 0.306, + 0.282, + 0.318 + ], + "angle": 0, + "content": "Multiple Objects" + }, + { + "type": "image_caption", + "bbox": [ + 0.146, + 0.332, + 0.34, + 0.358 + ], + "angle": 0, + "content": "4 Capabilities of Comprehensive Spatial Reasoning" + }, + { + "type": "image", + "bbox": [ + 0.352, + 0.093, + 0.668, + 0.355 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.676, + 0.093, + 0.857, + 0.356 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.37, + 0.907, + 0.44 + ], + "angle": 0, + "content": "Figure 1. Overview of the Spatial457 benchmark. We define four core capabilities for spatial reasoning. By incorporating these capabilities step-by-step, the benchmark assesses models across five progressive difficulty levels and seven question types, ranging from single-object recognition to advanced 6D spatial relationships and collision prediction. Our evaluation demonstrates a significant performance drop for more advanced questions and highlights the gap between state-of-the-art (SoTA) models and human performance in complex spatial reasoning VQA tasks." + }, + { + "type": "image", + "bbox": [ + 0.127, + 0.464, + 0.228, + 0.546 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.147, + 0.549, + 0.222, + 0.559 + ], + "angle": 0, + "content": "(a) SUN-RGBD" + }, + { + "type": "image", + "bbox": [ + 0.231, + 0.464, + 0.338, + 0.546 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.258, + 0.549, + 0.319, + 0.559 + ], + "angle": 0, + "content": "(b) nuScenes" + }, + { + "type": "image", + "bbox": [ + 0.346, + 0.464, + 0.449, + 0.546 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.374, + 0.549, + 0.415, + 0.559 + ], + "angle": 0, + "content": "(c) Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.571, + 0.483, + 0.627 + ], + "angle": 0, + "content": "Figure 2. The distribution of 3D poses in representative 6D datasets (a&b) highlights an imbalance in object orientations. This inspired us to develop a balanced benchmark encompassing all degrees of 3D orientation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.644, + 0.483, + 0.673 + ], + "angle": 0, + "content": "rotation [44, 46], enabling models to reason about precise 3D poses of objects within a scene." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.674, + 0.483, + 0.78 + ], + "angle": 0, + "content": "By progressively incorporating the four core capabilities, we establish five levels of difficulty, resulting in seven distinct question types. These questions span a spectrum of spatial reasoning complexity, ranging from basic semantic identification and 2D spatial relationships to advanced 3D positional understanding and orientation-based queries, culminating in comprehensive 6D spatial reasoning tasks." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Our evaluation measures performance across varied question types and examines how it is affected by different combinations of the four capabilities. This tiered design provides comprehensive benchmarks for spatial reasoning capabilities, highlighting each model's strengths and areas of vulnerability. To the best of our knowledge, this work presents the first comprehensive benchmark designed to evaluate the 6D spatial reasoning capabilities of vision-" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.467, + 0.633, + 0.481 + ], + "angle": 0, + "content": "language models." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.485, + 0.907, + 0.606 + ], + "angle": 0, + "content": "Additionally, our dataset's unbiased attribute design allows us to analyze potential biases in model predictions. By systematically exploring model performance under different attributes, we identify tendencies or limitations that may affect real-world applications. Altogether, our benchmark not only provides a foundational tool for evaluating LMMs but also serves as a guide for developing future benchmarks that support models with advanced 6D spatial reasoning." + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.609, + 0.782, + 0.623 + ], + "angle": 0, + "content": "Our main contributions are threefold:" + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.627, + 0.907, + 0.702 + ], + "angle": 0, + "content": "1. An unbiased and scalable synthetic dataset: We present a realistic, unbiased synthetic dataset designed to assess 6D spatial reasoning in LMMs. We define four core capabilities and propose five difficulty levels for spatial reasoning VQA with seven question types." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.703, + 0.906, + 0.792 + ], + "angle": 0, + "content": "2. Comprehensive benchmarking of 6D capabilities: We benchmark various LMMs across all difficulty levels. The performance variations across settings reveal potential limitations or strengths, providing valuable insights into which components should be prioritized to improve the spatial reasoning of LMMs." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.793, + 0.906, + 0.898 + ], + "angle": 0, + "content": "3. Bias analysis in model predictions: Leveraging the unbiased nature of our dataset, we conduct an in-depth bias analysis to explore how LMM predictions may be unbalanced across different attributes, especially 3D-related ones. This reveals potential weaknesses in model predictions, offering insights into limitations that may also appear in real-world scenarios." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.627, + 0.907, + 0.898 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24670" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.09, + 0.234, + 0.106 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.115, + 0.332, + 0.131 + ], + "angle": 0, + "content": "2.1. Synthetic Datasets in VQA" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.132, + 0.486, + 0.451 + ], + "angle": 0, + "content": "The use of synthetic data in visual question answering (VQA) has gained traction as a means to circumvent the limitations and biases inherent in real-world datasets. Synthetic environments [16, 30, 34] allow for controlled manipulation of scene elements, enabling researchers to probe specific reasoning abilities in VQA models without the confounding factors often present in natural imagery. Previous works [19, 24, 25] have utilized synthetic environments to evaluate multimodal models across a range of visual reasoning tasks, typically focusing on 2D spatial relationships and object features. For instance, CLEVR [19], a synthetic dataset widely used in the VQA community, evaluates compositional reasoning by introducing artificially generated scenes with diverse object arrangements. Super-CLEVR-3D [45] represents a significant step forward in evaluating 3D locations, it lacks an assessment of 6D spatial relationships that consider both 3D location and 3D orientation. Our work extends this tradition by creating a synthetic dataset tailored for 6D spatial reasoning, filling a gap in current synthetic datasets that do not address the complete spatial understanding of objects in 3D space." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.455, + 0.276, + 0.469 + ], + "angle": 0, + "content": "2.2. LMM Benchmarks" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.471, + 0.485, + 0.76 + ], + "angle": 0, + "content": "Existing benchmarks [15, 18, 35, 51] for large multimodal models (LMMs) predominantly assess their performance on tasks involving 2D visual question answering and multimodal reasoning. For instance, datasets like VQAv2 [15] and GQA [18] focus on answering questions based on visual content. However, these benchmarks leave a gap in evaluating models' capabilities in spatial contexts. Recent benchmarks, such as CV-Bench [40], assess spatial understanding, including depth ordering for 2.5D evaluation. More recent work [9, 10, 32] introduces questions regarding either aspect of 3D locations or rotations. Our proposed benchmark goes beyond traditional 2D, 2.5D, and 3D location or rotation assessments by introducing tasks that challenge LMMs to understand 3D positions and orientations together. We also evaluate models in a cascading manner across five difficulty levels and summarize their performance in four reasoning abilities, along with their output biases, providing a deeper analysis of the weaknesses in spatial reasoning within multimodal AI." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.77, + 0.483, + 0.806 + ], + "angle": 0, + "content": "3. Spatial457: Comprehensive Spatial Reasoning Benchmark" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.485, + 0.903 + ], + "angle": 0, + "content": "To comprehensively evaluate the spatial reasoning capabilities of vision-language models, we introduce a synthetic dataset named Spatial457. We define four core capabilities for comprehensive spatial reasoning (Sec. 3.1) and generate scenes with full 6D annotations (Sec. 3.2). The full combination of these capabilities form the new 6D spatial" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.744, + 0.107 + ], + "angle": 0, + "content": "relationship VQA tasks (Sec. 3.3)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.108, + 0.907, + 0.214 + ], + "angle": 0, + "content": "To thoroughly assess model capabilities, we define five levels of difficulty in vision question answering (VQA) tasks, converting seven question subsets that range from single-object recognition to the newly proposed 6D spatial reasoning tasks (Sec. 3.4). This structured progression in question difficulty is achieved by sequentially introducing new spatial reasoning capabilities, as illustrated in Fig. 1." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.225, + 0.855, + 0.243 + ], + "angle": 0, + "content": "3.1. Four Capabilities for Spatial Reasoning" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.248, + 0.907, + 0.293 + ], + "angle": 0, + "content": "We define 4 capabilities for comprehensive spatial understanding in a 3D scene: multiple objects, 2D locations, 3D locations, and 3D orientations." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.295, + 0.906, + 0.446 + ], + "angle": 0, + "content": "(1) Multiple objects. Objects are fundamental to defining scene structure [52]. Large vision-language models show strong results in aligning semantic features with text [23, 31], handling multiple objects from scene becomes the first toward scene understanding. Models must recognize individual attributes and interpret relationships among them. Our dataset evaluates multi-object reasoning, such as comparison and counting. For example, in Fig. 1, the image shows two cyan objects, requiring the model to verify color simultarity and count cyan objects accurately." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.447, + 0.906, + 0.627 + ], + "angle": 0, + "content": "(2) 2D locations. 2D spatial relationships are essential in prior image-based VQA tasks [18, 19]. In traditional multimodal models [20, 28, 49], these relationships can be learned through object detection training [8, 33]. However, in large-scale vision-language models, explicit location information is often absent from the latent representations. Following Li et al. [24], we examine four spatial relationships (left, right, front, back) between objects from 2D camera views as the first taks of spatial reasoning. Questions use these relations to filter targets, such as \"What is the shape of the object to the left of the chopper?\" (see Fig. 1)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.628, + 0.906, + 0.78 + ], + "angle": 0, + "content": "(3) 3D locations. Beyond 2D relationships, 3D locations can incorporate depth, capturing spatial relationships that 2D alone cannot represent. This is crucial for tasks involving occlusions and complex spatial arrangements. Our dataset assesses the 3D location understanding abilities of model through questions about occlusions query as in [45]. For example, as shown in Fig. 1, questions may ask about an object's position relative to others in 3D space or whether one object is occluded, challenging models to judge distances and spatial positions accurately." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.781, + 0.907, + 0.902 + ], + "angle": 0, + "content": "(4) 3D orientations. The 3D orientation of objects introduces an additional layer of complexity by incorporating rotational attributes. In current computer vision tasks, 3D orientation is often derived by pose estimation methods. However, large-scale vision-language models generally lack explicit 3D orientation information in their latent representations, potentially making it challenging to accurately infer the object's orientations. In our dataset, we eval" + }, + { + "type": "list", + "bbox": [ + 0.512, + 0.295, + 0.907, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "24671" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.198 + ], + "angle": 0, + "content": "uate models' abilities to understand 3D orientations by focusing on object poses, including their facing direction and alignments with others. For example, as depicted in Fig. 1, the model may be asked to identify whether a cyan object is facing left, oriented parallel to a brown object, or positioned in a specific spatial relationship — i.e. left, right, front, or behind— from the perspective of the objects themselves." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.206, + 0.258, + 0.222 + ], + "angle": 0, + "content": "3.2. Scene Rendering" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.225, + 0.484, + 0.331 + ], + "angle": 0, + "content": "We build upon the generation pipeline from prior VQA datasets [19, 24] and uniformly select 21 object models across 5 categories to construct the dataset as Li et al. [24]. To improve the realism of our rendered dataset, we use real image style environment maps and refined object textures. For benchmarking new 6D spatial understanding ability, we add new 6D spatial relationship annotations among objects." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.332, + 0.484, + 0.407 + ], + "angle": 0, + "content": "Background. To create realistic scene backgrounds, we use HDRI images\\(^2\\) with environment maps and natural lightings. We selected 224 different outdoor scenes with normal brightness levels without high light contrast by their metadata to maintain visual consistency across images." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.408, + 0.484, + 0.544 + ], + "angle": 0, + "content": "Objects. We reconstruct 3D mesh models [24] with realistic textures, introducing more variation and realistic appearances for objects. Similar to Li et al. [24], each object in our dataset is defined by distinct attributes and precise spatial locations. The original attributes include shape, color, and size. For benchmarking 6D spatial reasoning, each object also includes a vector specifying its 2D location \\((\\mathbf{x},\\mathbf{y})\\), 3D world coordinates \\((\\mathrm{X},\\mathrm{Y},\\mathrm{Z})\\), and pose direction derived from its 3D orientation [45]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.546, + 0.484, + 0.756 + ], + "angle": 0, + "content": "Relationships. As shown in Fig. 3 (a.1 to a.3), we define three types of spatial relationships: 1. 2D Spatial Relationships: These describe the relative positioning of objects within the 2D camera view, using terms such as left, right, front, and behind. 2. 6D Spatial Relationships: Extending spatial descriptions into 3D space, these relationships incorporate both the 3D position and orientation of objects to define relative positions from a target object's perspective, including left, right, front, and behind. 3. Collision Relationships: This relationship describes potential collisions if the target object moves forward or backward. Collision prediction involves assessing the orientation and 3D location of target objects relative to others to determine possible collisions." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.765, + 0.439, + 0.781 + ], + "angle": 0, + "content": "3.3. New Questions for 6D Spatial Reasoning" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.784, + 0.484, + 0.875 + ], + "angle": 0, + "content": "Based on the four capabilities defined in Sec. 3.1, we construct new question types for 6D spatial reasoning and collision prediction. These questions require an understanding of both 3D location and 3D orientation across multiple objects to generate accurate answers. This level of complexity represents the L5 difficulty level shown in Fig. 1." + }, + { + "type": "image", + "bbox": [ + 0.544, + 0.089, + 0.649, + 0.152 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.547, + 0.155, + 0.645, + 0.195 + ], + "angle": 0, + "content": "Q: How many yellow objects are to the left of the red wagon? A: 1" + }, + { + "type": "image_caption", + "bbox": [ + 0.559, + 0.2, + 0.642, + 0.213 + ], + "angle": 0, + "content": "(a.1) 2D spatial" + }, + { + "type": "image", + "bbox": [ + 0.656, + 0.09, + 0.762, + 0.15 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.659, + 0.151, + 0.758, + 0.198 + ], + "angle": 0, + "content": "Q: How many yellow objects are on the left hand side of the red wagon? A: 3" + }, + { + "type": "image_caption", + "bbox": [ + 0.671, + 0.2, + 0.754, + 0.212 + ], + "angle": 0, + "content": "(a.2) 6D spatial" + }, + { + "type": "image", + "bbox": [ + 0.767, + 0.09, + 0.874, + 0.15 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.772, + 0.151, + 0.871, + 0.189 + ], + "angle": 0, + "content": "Q: Is the red wagon will collide with a yellow regular bus if it moving forward?" + }, + { + "type": "image_caption", + "bbox": [ + 0.774, + 0.19, + 0.801, + 0.198 + ], + "angle": 0, + "content": "A: No" + }, + { + "type": "image_caption", + "bbox": [ + 0.778, + 0.2, + 0.856, + 0.212 + ], + "angle": 0, + "content": "(a.3) Collision" + }, + { + "type": "image", + "bbox": [ + 0.545, + 0.218, + 0.71, + 0.286 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.58, + 0.3, + 0.672, + 0.323 + ], + "angle": 0, + "content": "(b.1) Programs of 6D spatial question" + }, + { + "type": "image", + "bbox": [ + 0.719, + 0.218, + 0.87, + 0.299 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.724, + 0.301, + 0.852, + 0.322 + ], + "angle": 0, + "content": "(b.2) Programs of collision prediction question" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.332, + 0.906, + 0.375 + ], + "angle": 0, + "content": "Figure 3. (a.1-3) Example of the 2D spatial, 3D spatial and collision with images and questions. (b.1-2) shows the new operation programs we generated for the new 6D spatial reasoning questions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.387, + 0.907, + 0.644 + ], + "angle": 0, + "content": "The 6D spatial relationship questions challenge the model to understand the spatial relationship involving both 3D position and orientation from a target object's perspective (shown in Fig. 3 (a.2)). Compared with prior VQA benchmarks focusing on 2D [19, 24], we introduce new 3D spatial reasoning operations, 3D Relate(\\(\\cdot\\)), to determine relative positions like left, right, front, and behind within a 3D coordinate space (see Fig. 3 (b.1)). For example, a question might ask, \"How many objects are to the right of the red wagon?\" requiring the model to filter and count objects based on their orientation and relative 3D position, which would yield different answers compared to similar 2D left and right questions (Fig. 3 (a.1)). This task is particularly challenging as it requires models to integrate 3D spatial relationships while managing occlusions and depth perception—capabilities not typically covered in simpler 2D spatial tasks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.644, + 0.909, + 0.84 + ], + "angle": 0, + "content": "The collision prediction questions, as shown in Fig. 3 (a.3), extend 6D spatial reasoning by introducing motion-based inference. These questions assess the model's ability to anticipate future states and potential interactions (collisions) based on accurate estimation of 3D location and direction, denoted as forward or backward. For instance, a question might ask, \"Will the red wagon collide with a yellow bus if it moves forward?\" This requires the model to consider the orientation and direction of movement relative to surrounding objects to predict possible outcomes. Such tasks test the model's advanced spatial reasoning about 3D orientation and position, and its capacity to predict accurate trajectories and directions within the future scene." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.849, + 0.871, + 0.865 + ], + "angle": 0, + "content": "3.4. Questions under Other 4 Difficulty Levels" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.908, + 0.902 + ], + "angle": 0, + "content": "The L5 questions above represent the most challenging cases for 6D spatial reasoning, testing all four capabilities" + }, + { + "type": "page_footnote", + "bbox": [ + 0.108, + 0.887, + 0.332, + 0.9 + ], + "angle": 0, + "content": "2https://polyhaven.com/hdris" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "24672" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.168 + ], + "angle": 0, + "content": "in Sec. 3.1. However, the purpose of this benchmark is to comprehensively diagnose each factor in spatial reasoning. Thus, we design a structured roadmap to evaluate the model's performance across different levels of spatial reasoning, assessing its capabilities in each specific factor." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.17, + 0.483, + 0.365 + ], + "angle": 0, + "content": "By composing the four factors of comprehensive spatial reasoning, we define an additional four levels of question difficulty based on the generated scenes, encompassing a total of seven tasks in our benchmark. As the complexity of spatial reasoning increases with the introduction of additional factors, we organize these questions into a structured roadmap, as illustrated in Fig. 1. Although many questions are reimplementations from previous benchmarks [24, 45] on our new images, we aim to integrate these into a unified benchmark to provide deeper insights into the model's spatial reasoning capabilities. The composition of the four capabilities at each level, along with comparisons to previous VQA benchmarks, is shown in Tab. 1." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.367, + 0.483, + 0.502 + ], + "angle": 0, + "content": "Level 1 - single object questions [19, 24]. We begin with the questions with single objects. These questions are generated using a filtered set of templates from Li et al. [24]. We select five templates involving only one object in the question, such as \"What color is the double bus?\" in Fig. 1. This question type provides a base score for the tested vision-language model. The comparison with L1 helps to determine whether the decrease of performance are due to challenge of spatial reasoning or object recognition." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.505, + 0.483, + 0.641 + ], + "angle": 0, + "content": "Level 2 - multiple objects questions [19, 24]. Level 2 questions extend Level 1 by involving multiple objects. These questions involve understanding the relationship between objects and counting. For example, \"Is there another object of the same color as the double bus?\" as shown in Fig. 1, where the question templates follow Li et al. [24]. We evaluate the model's capacity for reasoning in scenes with multiple entities as a critical step toward understanding spatial relationships in next levels." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.643, + 0.483, + 0.748 + ], + "angle": 0, + "content": "Level 3 - 2D spatial questions [19, 24]. Level 3 questions further introduce 2D spatial relationships among multiple objects, requiring the model to understand the position of objects and their relationships from the 2D camera's perspective. For example, \"What shape is the object to the left of the chopper?\" as depicted in Fig. 1. The 2D relationship is the first step of spatial reasoning." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Level 4 - 3D pose questions [45]. Level 4 questions extend 3D orientation reasoning beyond the 2D spatial relationships in Level 3. Beyond appearance attributes (shape, color, size) or 2D locations, these questions require understanding 3D orientation, including the rotation angle or facing direction of an object from the camera's perspective. For example, \"What shape is the object that is parallel to the brown object?\" as shown in Fig. 1. 3D pose questions test the model's ability to interpret scenes with 3D awareness, involving rotation and alignment with other objects." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.288 + ], + "angle": 0, + "content": "Level 4 - occlusion questions [45]. In addition to 3D orientation, incorporating 3D location presents another challenge beyond the Level 3 questions on 2D spatial relationships. A unique aspect of 3D spatial reasoning from images is occlusion. The occlusion questions in Level 4 require the model to comprehend depth perception and the relative positions of objects in 3D space, determining which object is closer to the camera within overlapping areas. For example, an occlusion question might ask, \"Is the red wagon occluded by the yellow bus?\" as shown in Fig. 1. This level evaluates the model's ability to reason about occlusions and depth perception, which are essential for understanding complex spatial relationships in 3D scenes." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.289, + 0.905, + 0.436 + ], + "angle": 0, + "content": "
Question LevelMulti-object2D location3D location3D OrientationVQA Benchmarks
L1spatial-agnostic bench (e.g., TextVQA [35])
L2✓CLEVR-Math [25], CLEVR-Count [19]
L3✓✓VQA [15], GQA [18], CLEVR [19], Super-CLEVR [24]
L4 (a)✓✓✓CV-Bench (DepthOrder) [40], Super-CLEVR-3D [45]
L4 (b)✓✓✓Super-CLEVR-3D [45]
L5✓✓✓✓Spatial457
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.447, + 0.905, + 0.503 + ], + "angle": 0, + "content": "Table 1. Comparison of VQA Benchmarks by Factors and Question Levels. Compared with existing VQA benchmarks, Spatial457 includes all four spatial reasoning factors, providing a comprehensive evaluation of 6D spatial understanding." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.508, + 0.71, + 0.523 + ], + "angle": 0, + "content": "3.5. Evaluation Protocols" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.525, + 0.905, + 0.636 + ], + "angle": 0, + "content": "To better evaluate the model's spatial reasoning capabilities, we propose a structured evaluation protocol that measures the relative capability of models on each factor for spatial reasoning. We define this evaluation metric as the Relative Performance Dropping Rate (RPDR), which assesses the decline in performance after introducing a new factor in the questions. Based on the roadmap in Fig. 1, the RPDR for each factor is defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.527, + 0.641, + 0.765, + 0.656 + ], + "angle": 0, + "content": "\\[\n\\mathrm {R P D R} _ {\\text {M u l t i - O b j .}} = A c c _ {\\text {L e v e l 1}} \\Rightarrow A c c _ {\\text {L e v e l 2}}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.538, + 0.659, + 0.765, + 0.672 + ], + "angle": 0, + "content": "\\[\n\\mathrm {R P D R} _ {\\mathrm {2 D L o c .}} = A c c _ {\\text {L e v e l 2}} \\Rightarrow A c c _ {\\text {L e v e l 3}}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.537, + 0.677, + 0.89, + 0.707 + ], + "angle": 0, + "content": "\\[\n\\mathrm {R P D R} _ {\\mathrm {3 D L o c .}} = \\operatorname {A v g} \\left( \\begin{array}{c} A c c _ {\\text {L e v e l 3}} \\Rightarrow A c c _ {\\text {L e v e l 4 (O c c .)}}, \\\\ A c c _ {\\text {L e v e l 4 (3 D P o s e)}} \\Rightarrow A c c _ {\\text {L e v e l 5 (6 D)}} \\end{array} \\right)\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.539, + 0.709, + 0.88, + 0.74 + ], + "angle": 0, + "content": "\\[\n\\mathrm {R P D R} _ {\\mathrm {3 D O r i .}} = \\operatorname {A v g} \\left( \\begin{array}{c} A c c _ {\\text {L e v e l 3}} \\Rightarrow A c c _ {\\text {L e v e l 4 (3 D P o s e)}}, \\\\ A c c _ {\\text {L e v e l 4 (O c c .)}} \\Rightarrow A c c _ {\\text {L e v e l 5 (C o l .)}} \\end{array} \\right)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.747, + 0.905, + 0.809 + ], + "angle": 0, + "content": "where \\( Acc_{\\mathrm{Level}} \\) denotes the model's accuracy on the corresponding questions and \\( (Acc_{\\mathrm{Q1}} \\Rightarrow Acc_{\\mathrm{Q2}}) = \\frac{Acc_{\\mathrm{Q2}} - Acc_{\\mathrm{Q1}}}{Acc_{\\mathrm{Q1}}} \\) denotes the relative performance dropping rate from \\( Q1 \\) to \\( Q2 \\). The results of RPDR will be discussed in Sec. 4.3." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.817, + 0.646, + 0.833 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.838, + 0.704, + 0.854 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.905, + 0.902 + ], + "angle": 0, + "content": "We generate 1,000 images to test the performance of large vision-language models. For the seven types of questions, we generate 1000 single-object questions (L1-Single), 5,000" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "24673" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.11, + 0.089, + 0.891, + 0.311 + ], + "angle": 0, + "content": "
Difficulty LevelLevel 1Level 2Level 3Level 4Level 5
Questions typeSingle ObjectMulti-Obj.2D SpatialOcclusion3D PoseCollisions6D Spatial
Random33.0532.7733.4722.0418.9921.0219.41
GPT-4o [29]74.4662.8856.1448.4042.4138.4137.01
GeminiPro 1.5 [37]73.2662.5454.4947.6543.6741.1939.36
Claude 3.5 Sonnet [3]68.2457.4054.1930.8438.4035.3433.48
Qwen2-VL-7B-Inst. [39]71.9661.4455.3427.8734.2936.5833.75
InternVL2-8B [38]58.1158.7657.4033.2532.3234.3034.30
LLaVA-v1.5-7B [26]44.8744.7242.2024.3424.5523.6323.86
LLaVA-NeXT-vicuna-7B [27]50.7249.4746.0129.6829.3531.9531.95
LLaVA-NeXT-Ilama3-8B [41]52.1549.7345.9230.3129.7732.1232.12
PO3D-VQA [45]86.4682.5580.6470.4981.4068.1271.06
Human89.9786.8384.9582.7684.9581.8279.94
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.321, + 0.907, + 0.409 + ], + "angle": 0, + "content": "Table 2. Performance comparison across all 7 question types under 5 difficulty levels. As task complexity increases from single-object questions to 6D spatial relationships, performance generally decreases in most cases, highlighting the challenges in handling multi-object interactions, 3D orientations, and predictive spatial reasoning. The highest score among all models for each question type is marked in red , and the second-highest scores are underlined. The best performance among open-source models is marked in blue . PO3D-VQA is the state-of-the-art (SOTA) 3D-aware neural symbolic reasoning model as proposed in [45], and is listed as the potential upper bound of evaluated methods. The comparison with human performance also shows an increasing gap for LLMs as task difficulty increases." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.433, + 0.482, + 0.509 + ], + "angle": 0, + "content": "multiple-object questions (L2-Multi-object.), 4,995 2D spatial relationship questions (L3-2D-Spatial), 3,534 occlusion questions (L4-Occlusion), 4,555 3D pose questions (L4-3D-Pose), 4,357 collision questions (L5-Collision), and 4,995 6D spatial relationship questions (L5-6D-Spatial)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.51, + 0.483, + 0.646 + ], + "angle": 0, + "content": "We evaluate two series of large vision-language models on Spatial457: (1) API-based large language models, including GPT-4o [29], Gemini-1.5-Pro [37], and Claude-3.5-Sonnet [3]; and (2) open-source large vision-language models, including Qwen2-VL-7B-Instruct [39], InternVL2-8B [38], llava-v1.5-7b [26], llava next-vicuna-7b [27], and llava next-llama3-8b [41]. The evaluation follows the format of Vlmealkit [12]. The details of the model implementation including prompts are listed in Appendix." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.655, + 0.273, + 0.669 + ], + "angle": 0, + "content": "4.2. Evaluation Results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.673, + 0.483, + 0.809 + ], + "angle": 0, + "content": "In Tab. 2, we present the performance of various large vision-language models across different levels of difficulty on Spatial457. The levels range from basic single-object tasks to complex 6D spatial reasoning, with increasing difficulty as more factors are introduced. We compare both API-based models (e.g., GPT-4o, GeminiPro1.5, Claude3-5V) and open-source models (e.g., Qwen2-VL-7B-Instruct, InternVL2-8B, LLaVA series) to assess their spatial reasoning capabilities." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Comparison across difficulty levels. We evaluate model performance across different levels of difficulty in Spatial457, ranging from basic single-object tasks to complex 6D spatial reasoning. As shown in Tab. 2, performance generally decreases as the complexity of spatial relationships increases, highlighting the challenges models face" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.433, + 0.905, + 0.463 + ], + "angle": 0, + "content": "when handling multi-object interactions, 3D orientations, and predictive spatial reasoning tasks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.464, + 0.907, + 0.6 + ], + "angle": 0, + "content": "Comparison between API models and open-source models. From Tab. 2, we observe that API models generally outperform open-source models across all difficulty levels. In 2D spatial relationship questions, InternVL2-8B achieves slightly higher results than the API models, while GPT-4o is only \\(1\\%\\) lower. Qwen2-VL-7B-Instruct achieves comparable results, especially in simpler tasks like single-object and multiple-object reasoning. This analysis highlights the strengths of API models in complex spatial reasoning tasks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.601, + 0.906, + 0.722 + ], + "angle": 0, + "content": "Comparison among API models. API-based models, i.e. GPT-4o, GeminiPro1.5, and Claude3-5V, demonstrate strengths across all difficulty levels. For instance, GPT-4o excels in 2D and 3D spatial tasks, while GeminiPro1.5 shows robust performance in collision prediction tasks. This comparison provides insights into the unique capabilities and limitations of each API model in handling spatial reasoning across different levels of complexity." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.73, + 0.807, + 0.746 + ], + "angle": 0, + "content": "4.3. RPDR Analysis for 4 Capabilities" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.907, + 0.901 + ], + "angle": 0, + "content": "As outlined in Sec. 3.5, we use the Relative Performance Dropping Rate (RPDR) to analyze model performance across the four core factors of Spatial457: multi-object reasoning, 2D location understanding, 3D location comprehension, and 3D orientation interpretation. The RPDR analysis provides a structured approach to quantify the decline in accuracy with the introduction of each new factor, as detailed in Tab. 3. Our findings reveal that all models exhibit weaknesses in 3D reasoning. Specifically, GPT-4o and GeminiPro-1.5 perform the worst in 3D orientation tasks," + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "24674" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.092, + 0.092, + 0.435, + 0.107 + ], + "angle": 0, + "content": "while the other two struggle most with 3D location." + }, + { + "type": "table", + "bbox": [ + 0.104, + 0.118, + 0.473, + 0.195 + ], + "angle": 0, + "content": "
ModelMulti.2D Loc.3D Ori.3D Loc.
GPT-4o [29]84.4589.2777.4686.74
GeminiPro 1.5 [37]85.3787.1383.2888.79
Claude 3.5 S. [3]84.1294.4185.4372.05
Qwen2-VL-7B [39]85.3990.0680.9974.40
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.205, + 0.483, + 0.248 + ], + "angle": 0, + "content": "Table 3. RPDR (%) analysis on the four spatial reasoning capabilities, where higher values indicate better performance. For each row, we highlight the lowest RPDR score among all factors." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.258, + 0.231, + 0.273 + ], + "angle": 0, + "content": "4.4. Bias Analysis" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.274, + 0.484, + 0.364 + ], + "angle": 0, + "content": "In this section, we analyze prediction biases in large language models (LLMs) regarding their understanding of key attributes, specifically shape, color, size, and pose. Leveraging the unbiased generation of dataset images and questions in Spatial457, we examine how these models may demonstrate attribute-specific biases in their predictions." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.364, + 0.484, + 0.514 + ], + "angle": 0, + "content": "We first compute the confusion matrices and distribution statistics for both ground truth and model predictions across each attribute (shape, color, size, and pose). For example, the confusion matrix plots for color and pose attributes of GPT-4o and GeminiPro1.5, together with the distribution of prediction and ground truth label are shown in Fig. 4. We find both GPT-4o and GeminiPro1.5 exhibit bias in both color and pose predictions. These results highlight areas for improvement to reduce bias and enhance predictive reliability across models." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.514, + 0.484, + 0.543 + ], + "angle": 0, + "content": "To quantify prediction variability, we calculate the Coefficient of Variation (CV) for each attribute as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.205, + 0.549, + 0.483, + 0.578 + ], + "angle": 0, + "content": "\\[\n\\mathrm {C V} = \\frac {\\operatorname {S t d} \\left(p _ {1} , \\dots , p _ {k}\\right)}{\\operatorname {M e a n} \\left(p _ {1} , \\dots , p _ {k}\\right)}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.58, + 0.483, + 0.623 + ], + "angle": 0, + "content": "where \\(\\operatorname{Std}(\\cdot)\\) and \\(\\operatorname{Mean}(\\cdot)\\) are the standard deviation and mean of the probability distribution \\(p_i\\) for each attribute label \\(i\\) (e.g., left), \"right\" for pose)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.624, + 0.484, + 0.715 + ], + "angle": 0, + "content": "Tab. 4 shows the CV values for shape, color, size, and pose attributes across different models. Lower CV values indicate more consistent predictions with less bias, which reveals significant biases in model predictions across attributes. Higher CV values, especially in color and pose, indicate consistency challenges." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.718, + 0.425, + 0.734 + ], + "angle": 0, + "content": "4.5. Comparison to the 3D Symbolic Model" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.484, + 0.901 + ], + "angle": 0, + "content": "As neural symbolic models achieve state-of-the-art (SOTA) performance on many prior synthetic datasets, we implement the 3D-aware neural symbolic model PO3DVQA [45] on our new image domain and all levels of questions. We train PO3D-VQA on a separate set of 20,000 images and questions and report its performance on the same test dataset in Tab. 2. A direct comparison between the 3D neural symbolic model and large language models is not entirely fair, as each module of the neural symbolic model (6D pose estimation, attribute classification, and language reasoning) is trained independently as a sub-task. However, we" + }, + { + "type": "image", + "bbox": [ + 0.536, + 0.092, + 0.885, + 0.367 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.664, + 0.369, + 0.761, + 0.381 + ], + "angle": 0, + "content": "(b) GeminiPro1-5" + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.393, + 0.907, + 0.464 + ], + "angle": 0, + "content": "Figure 4. The distribution of color and pose attributes in the L4-3D-Pose task for GPT-4o and Gemini-Pro1.5. Although the ground truth labels are well balanced, the predicted colors and poses for both models are imbalanced, with \"yellow\" being the most frequent color and \"front\" the most frequent pose." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.482, + 0.905, + 0.578 + ], + "angle": 0, + "content": "
ModelsAvg. ↓ShapeColorSizePose
GPT4o [29]0.4850.6230.5210.0430.752
GeminiPro1.5 [37]0.4980.5670.3660.2970.760
Claude3.5V-S [3]0.7200.7770.4650.5741.063
Qwen2-VL-7B [39]0.5450.8680.5580.0670.688
Human0.0930.1100.0820.0550.126
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.587, + 0.906, + 0.671 + ], + "angle": 0, + "content": "Table 4. Coefficient of Variation (CV) for predictions on each attribute in the L4-3D-Pose task. We evaluate three API models and one open-source model, calculating the CV across four attributes (shape, color, size, 3D pose) based on the distribution of predicted labels. Lower CV values indicate less bias. The first column shows the average CV across attributes, and GPT4-o has the lowest bias." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.692, + 0.905, + 0.723 + ], + "angle": 0, + "content": "include this model to illustrate the upper bound that a data-driven machine learning model can achieve on our dataset." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.732, + 0.697, + 0.747 + ], + "angle": 0, + "content": "4.6. Qualitative Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Fig. 5 presents an example of GPT-4o's performance across various spatial reasoning tasks on the same image. GPT-4o correctly answers multi-object, 2D spatial, and occlusion tasks but fails on 3D pose and 6D spatial questions, indicating challenges in 3D orientation understanding in this case. As we prompt the models to output the answer as well as the reasoning steps, the models correctly detect the existence of objects (e.g. in Fig. 5 the yellow minivan in L4-Pose or the fighter plane and road bike in L5 questions, but has the wrong understand to the orientations or 3D locations." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24675" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.111, + 0.089, + 0.462, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.267, + 0.485, + 0.337 + ], + "angle": 0, + "content": "Figure 5. Example of GPT-4o across different spatial reasoning tasks on the same image with the ground truth answer marked in blue. GPT-4o answers multi-object, 2D spatial, and occlusion tasks correctly (in the green block) but fails on 3D pose and 6D spatial questions (the red block)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.36, + 0.332, + 0.376 + ], + "angle": 0, + "content": "5. Extension to Real Dataset" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.379, + 0.483, + 0.514 + ], + "angle": 0, + "content": "We extend the setting of 3D questions to real-world image data. Due to limitations in fully annotated attributes such as category, color, 3D orientation, position, and occlusion, we focus on questions from the L4-Pose category. Using the scene structure from SUN-RGBD [36], we construct a new subset, L4-Pose-Real, which involves only the category and pose attributes. To reduce ambiguity in object localization, we add a red bounding box around target objects, ensuring clarity in identifying categories within the scene." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.515, + 0.484, + 0.605 + ], + "angle": 0, + "content": "We provide the question along with the image as input to the LLM. In this example in Fig. 6, GPT-4o provides the correct answer, while Gemini-Pro 1.5 is incorrect. However, the reasoning process reveals that both models rely on common sense knowledge or 2D visibility cues to interpret the 3D orientation." + }, + { + "type": "image", + "bbox": [ + 0.111, + 0.615, + 0.458, + 0.773 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.78, + 0.483, + 0.852 + ], + "angle": 0, + "content": "Figure 6. Example of real-world 3D pose reasoning tasks \\(L4\\)-Pose-Real. Here we show the reasoning and answers provided by LLMs where GPT-4o is correct and Gemini-Pro 1.5 is wrong. But the reasoning process reveals these two LLMs are both reasoning 3D orientation from the command sense or 2D visibility." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "As shown in Tab. 5, Gemini Pro 1.5 achieves the highest accuracy (40.50%) in L4-Pose-Real, and the gap between the three API-based models and the open-source" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.228 + ], + "angle": 0, + "content": "model Qwen2-VL-7B-Instruct (26.50%) is larger. However, all these models perform significantly lower than humans, which have accuracy 92.64%. We further analyze the bias of predictions in this real-world question set. Among the models, Claude3-5V Sonnet exhibits the lowest CV (0.374), indicating relatively stable predictions, while Qwen2-VL-7B-Instruct has the highest CV (0.934). The distribution of ground truth and predicted pose and the confusion matrix are shown in Fig. 7." + }, + { + "type": "table", + "bbox": [ + 0.527, + 0.246, + 0.892, + 0.352 + ], + "angle": 0, + "content": "
ModelsAccuracy (%) ↑CV ↓
GPT-4o [29]38.750.690
GeminiPro1-5 [37]40.500.762
Claude3-5V Sonnet [3]38.750.374
Qwen2-VL-7B-Instruct [39]26.500.934
Human92.640.197
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.353, + 0.907, + 0.41 + ], + "angle": 0, + "content": "Table 5. Accuracy and Coefficient of Variation (CV) for L4-Poserel task: Comparison of accuracy and CV for GPT-4o, GeminiPro1.5, Claude3-5V, and Qwen2-VL-7B-Instruct on real-world 3D pose tasks. Human performance is included as a reference." + }, + { + "type": "image", + "bbox": [ + 0.537, + 0.437, + 0.889, + 0.589 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.531, + 0.601, + 0.888, + 0.615 + ], + "angle": 0, + "content": "Figure 7. Distribution of 3D poses for in L4-Pose-Real task." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.638, + 0.634, + 0.654 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.659, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In this work, we introduce Spatial457, a scalable and unbiased benchmark designed to comprehensively assess 6D spatial reasoning capabilities in LMMs. Spatial457 extends traditional 2D and 3D VQA benchmarks by incorporating complex 6D spatial reasoning tasks, structured across five levels of difficulty. Through evaluations, we observe that current LMMs demonstrate strong performance in basic object recognition and 2D spatial relationships, but struggle significantly with 3D spatial understanding and collision prediction tasks. Additionally, our analysis of prediction biases uncovers attribute-specific inconsistencies that could affect model reliability in real-world applications. Limitations. To real images settings, current evaluation limits to L3 pose questions. We are actively working on building the additional required annotations for these real images to expand the domain of the proposed benchmark." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "24676" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.091, + 0.251, + 0.108 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.115, + 0.483, + 0.16 + ], + "angle": 0, + "content": "This work is supported by ONR with award N00014-23-1-2641, from ARL Army Research Laboratory with award W911NF2320008." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.174, + 0.188, + 0.189 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.199, + 0.483, + 0.226 + ], + "angle": 0, + "content": "[1] Alibaba DAMO Academy. Gemini: Vision-language model by alibaba. Alibaba DAMO, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.228, + 0.484, + 0.296 + ], + "angle": 0, + "content": "[2] Boshi An, Yiran Geng, Kai Chen, Xiaqi Li, Qi Dou, and Hao Dong. Rgbmanip: Monocular image-based robotic manipulation through active object pose estimation. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 7748-7755. IEEE, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.298, + 0.482, + 0.325 + ], + "angle": 0, + "content": "[3] Anthropic. Claude: An anthropic language model. Anthropotic, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.327, + 0.483, + 0.395 + ], + "angle": 0, + "content": "[4] Gilad Baruch, Zhuoyuan Chen, Afshin Dehghan, Tal Dimry, Yuri Feigin, Peter Fu, Thomas Gebauer, Brandon Joffe, Daniel Kurz, Arik Schwartz, et al. Arkitsscenes: A diverse real-world dataset for 3d indoor scene understanding using mobile rgb-d data. arXiv preprint arXiv:2111.08897, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.398, + 0.482, + 0.437 + ], + "angle": 0, + "content": "[5] Mark Billinghurst, Adrian Clark, and Gun Lee. A survey of augmented reality. Foundations and Trends in Human-Computer Interaction, 8(2-3):73-272, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.439, + 0.483, + 0.508 + ], + "angle": 0, + "content": "[6] Garrick Brazil, Abhinav Kumar, Julian Straub, Nikhila Ravi, Justin Johnson, and Georgia Gkioxari. Omni3d: A large benchmark and model for 3d object detection in the wild. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13154-13164, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.51, + 0.484, + 0.592 + ], + "angle": 0, + "content": "[7] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11621-11631, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.594, + 0.482, + 0.648 + ], + "angle": 0, + "content": "[8] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.65, + 0.482, + 0.731 + ], + "angle": 0, + "content": "[9] Boyuan Chen, Zhuo Xu, Sean Kirmani, Brain Ichter, Dorsa Sadigh, Leonidas Guibas, and Fei Xia. Spatialvm: Endowing vision-language models with spatial reasoning capabilities. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14455-14465, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.733, + 0.482, + 0.788 + ], + "angle": 0, + "content": "[10] An-Chieh Cheng, Hongxu Yin, Yang Fu, Qiushan Guo, Ruihan Yang, Jan Kautz, Xiaolong Wang, and Sifei Liu. Spatial-rgpt: Grounded spatial reasoning in vision language model. arXiv preprint arXiv:2406.01584, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.79, + 0.482, + 0.87 + ], + "angle": 0, + "content": "[11] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhanghao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E Gonzalez, et al. Vicuna: An open-source chatbot impressing gpt-4 with \\(90\\%\\) chatgpt quality, march 2023. URL https://lmsys.org/blog/2023-03-30-vicuna, 3(5), 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.482, + 0.902 + ], + "angle": 0, + "content": "[12] Haodong Duan, Junming Yang, Yuxuan Qiao, Xinyu Fang, Lin Chen, Yuan Liu, Xiaoyi Dong, Yuhang Zang, Pan Zhang," + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.199, + 0.484, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.148 + ], + "angle": 0, + "content": "Jiaqi Wang, et al. Vlmealkit: An open-source toolkit for evaluating large multi-modality models. In Proceedings of the 32nd ACM international conference on multimedia, pages 11198-11201, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.151, + 0.905, + 0.205 + ], + "angle": 0, + "content": "[13] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. International Journal of Computer Vision, 88(2):303-338, 2010." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.208, + 0.905, + 0.275 + ], + "angle": 0, + "content": "[14] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 3354-3361, 2012." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.278, + 0.905, + 0.348 + ], + "angle": 0, + "content": "[15] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Batra, and Devi Parikh. Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6904-6913, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.35, + 0.905, + 0.377 + ], + "angle": 0, + "content": "[16] Roland Hess. *Blender foundations: The essential guide to learning blender* 2.5. Routledge, 2013." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.379, + 0.905, + 0.433 + ], + "angle": 0, + "content": "[17] Yushi Hu, Hang Hua, Zhengyuan Yang, Weijia Shi, Noah A Smith, and Jiebo Luo. Promptcap: Prompt-guided task-aware image captioning. arXiv preprint arXiv:2211.09699, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.436, + 0.905, + 0.504 + ], + "angle": 0, + "content": "[18] Drew A Hudson and Christopher D Manning. Gqa: A new dataset for real-world visual reasoning and compositional question answering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6700-6709, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.507, + 0.905, + 0.59 + ], + "angle": 0, + "content": "[19] Justin Johnson, Bharath Hariharan, Laurens Van Der Maaten, Li Fei-Fei, C Lawrence Zitnick, and Ross Girshick. Clevr: A diagnostic dataset for compositional language and elementary visual reasoning. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2901–2910, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.592, + 0.905, + 0.661 + ], + "angle": 0, + "content": "[20] Aishwarya Kamath, Mannat Singh, Yann LeCun, Gabriel Synnaeve, Ishan Misra, and Nicolas Carion. Mdetr-modulated detection for end-to-end multi-modal understanding. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1780-1790, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.662, + 0.907, + 0.732 + ], + "angle": 0, + "content": "[21] Dong-Jin Kim, Jinsoo Choi, Tae-Hyun Oh, and In So Kweon. Dense relational captioning: Triple-stream networks for relationship-based captioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6271-6280, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.734, + 0.905, + 0.829 + ], + "angle": 0, + "content": "[22] Alina Kuznetsova, Hassan Rom, Neil Alldrin, Jasper Uijlings, Ilya Krasin, Jordi Pont-Tuset, Gleb Muryshkin, Sergey Popov, Marwa Elhachem, Pedro de Carvalho, et al. The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale. International Journal of Computer Vision, 128(7):1956-1981, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.832, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[23] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730-19742. PMLR, 2023." + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24677" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.176 + ], + "angle": 0, + "content": "[24] Zhuowan Li, Xingrui Wang, Elias Stengel-Eskin, Adam Kortylewski, Wufei Ma, Benjamin Van Durme, and Alan L Yuille. Super-clevr: A virtual benchmark to diagnose domain robustness in visual reasoning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14963-14973, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.178, + 0.483, + 0.232 + ], + "angle": 0, + "content": "[25] Adam Dahlgren Lindstrom and Savitha Sam Abraham. Clevr-math: A dataset for compositional language, visual and mathematical reasoning. arXiv preprint arXiv:2208.05358, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.234, + 0.482, + 0.262 + ], + "angle": 0, + "content": "[26] Haotian Liu et al. Llava: Large language and vision assistant. LLaVA Project, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.264, + 0.483, + 0.29 + ], + "angle": 0, + "content": "[27] Haotian Liu et al. Llava-last: Extended large language and vision assistant. LLaVA Project, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.293, + 0.483, + 0.347 + ], + "angle": 0, + "content": "[28] Jiayuan Mao, Chuang Gan, Pushmeet Kohli, Joshua B Tenenbaum, and Jiajun Wu. The neuro-symbolic concept learner: Interpreting scenes, words, and sentences from natural supervision. arXiv preprint arXiv:1904.12584, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.349, + 0.41, + 0.363 + ], + "angle": 0, + "content": "[29] OpenAI. Gpt-4 technical report. OpenAI, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.365, + 0.483, + 0.433 + ], + "angle": 0, + "content": "[30] Weichao Qiu and Alan Yuille. Unrealcv: Connecting computer vision to unreal engine. In Computer Vision-ECCV 2016 Workshops: Amsterdam, The Netherlands, October 8-10 and 15-16, 2016, Proceedings, Part III 14, pages 909-916. Springer, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.435, + 0.483, + 0.517 + ], + "angle": 0, + "content": "[31] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.519, + 0.483, + 0.588 + ], + "angle": 0, + "content": "[32] Arjit Ray, Jiafei Duan, Reuben Tan, Dina Bashkirova, Rose Hendrix, Kiana Ehsani, Aniruddha Kembhavi, Bryan A Plummer, Ranjay Krishna, Kuo-Hao Zeng, et al. Sat: Spatial aptitude training for multimodal language models. arXiv preprint arXiv:2412.07755, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.59, + 0.483, + 0.645 + ], + "angle": 0, + "content": "[33] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. IEEE transactions on pattern analysis and machine intelligence, 39(6):1137-1149, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.647, + 0.483, + 0.673 + ], + "angle": 0, + "content": "[34] Andrew Sanders. An introduction to Unreal engine 4. AK Peters/CRC Press, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.676, + 0.483, + 0.744 + ], + "angle": 0, + "content": "[35] Amanpreet Singh, Vivek Natarajan, Meet Shah, Yu Jiang, Xinlei Chen, Dhruv Batra, Devi Parikh, and Marcus Rohrbach. Towards vqa models that can read. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8317-8326, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.746, + 0.483, + 0.802 + ], + "angle": 0, + "content": "[36] Shuran Song, Samuel P Lichtenberg, and Jianxiong Xiao. Sun rgb-d: A rgb-d scene understanding benchmark suite. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 567-576, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.804, + 0.483, + 0.871 + ], + "angle": 0, + "content": "[37] Gemini Team, Petko Georgiev, Ving Ian Lei, Ryan Burnell, Libin Bai, Anmol Gulati, Garrett Tanzer, Damien Vincent, Zhufeng Pan, Shibo Wang, et al. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context. arXiv preprint arXiv:2403.05530, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.483, + 0.9 + ], + "angle": 0, + "content": "[38] Intern Team. Internvl 2.0: A vision-language model. Intern Foundation, 2023." + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.121 + ], + "angle": 0, + "content": "[39] Qwen Team. Qwen-vl: A large multimodal model. Alibaba Group, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.124, + 0.905, + 0.192 + ], + "angle": 0, + "content": "[40] Shengbang Tong, Ellis Brown, Penghao Wu, Sanghyun Woo, Manoj Middepogu, Sai Charitha Akula, Jihan Yang, Shusheng Yang, Adithya Iyer, Xichen Pan, et al. Cambrian1: A fully open, vision-centric exploration of multimodal llms. arXiv preprint arXiv:2406.16860, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.195, + 0.905, + 0.223 + ], + "angle": 0, + "content": "[41] Hugo Touvron et al. Llama: Open and efficient foundation language models. Meta AI, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.226, + 0.905, + 0.28 + ], + "angle": 0, + "content": "[42] Oriol Vinyals, Alexander Toshev, Samy Bengio, and Dumitru Erhan. Show and tell: A neural image caption generator. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3156-3164, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.284, + 0.905, + 0.366 + ], + "angle": 0, + "content": "[43] He Wang, Srinath Sridhar, Jingwei Huang, Julien Valentin, Shuran Song, and Leonidas J Guibas. Normalized object coordinate space for category-level 6d object pose and size estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2642-2651, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.37, + 0.905, + 0.425 + ], + "angle": 0, + "content": "[44] Xingrui Wang, Wufei Ma, Zhuowan Li, Adam Kortylewski, and Alan Yuille. 3d-aware visual question answering about parts, poses and occlusions. In Thirty-seventh Conference on Neural Information Processing Systems, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.428, + 0.905, + 0.482 + ], + "angle": 0, + "content": "[45] Xingrui Wang, Wufei Ma, Zhuowan Li, Adam Kortylewski, and Alan L Yuille. 3d-aware visual question answering about parts, poses and occlusions. Advances in Neural Information Processing Systems, 36, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.486, + 0.905, + 0.541 + ], + "angle": 0, + "content": "[46] Yu Xiang, Roozbeh Mottaghi, and Silvio Savarese. Beyond Pascal: A benchmark for 3d object detection in the wild. In IEEE winter conference on applications of computer vision, pages 75-82. IEEE, 2014." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.544, + 0.905, + 0.613 + ], + "angle": 0, + "content": "[47] Kelvin Xu, Jimmy Ba, Ryan Kiros, Kyunghyun Cho, Aaron Courville, Ruslan Salakhudinov, Rich Zemel, and Yoshua Bengio. Show, attend and tell: Neural image caption generation with visual attention. In International conference on machine learning, pages 2048-2057. PMLR, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.616, + 0.905, + 0.67 + ], + "angle": 0, + "content": "[48] Lin Xu, Yilin Zhao, Daquan Zhou, Zhijie Lin, See Kiong Ng, and Jiashi Feng. Pllava: Parameter-free llava extension from images to videos for video dense captioning. arXiv preprint arXiv:2404.16994, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.674, + 0.905, + 0.742 + ], + "angle": 0, + "content": "[49] Kexin Yi, Jiajun Wu, Chuang Gan, Antonio Torralba, Pushmeet Kohli, and Josh Tenenbaum. Neural-symbolic vqa: Disentangling reasoning from vision and language understanding. Advances in neural information processing systems, 31, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.746, + 0.905, + 0.801 + ], + "angle": 0, + "content": "[50] Kexin Yi, Chuang Gan, Yunzhu Li, Pushmeet Kohli, Jiajun Wu, Antonio Torralba, and Joshua B Tenenbaum. Clevrer: Collision events for video representation and reasoning. arXiv preprint arXiv:1910.01442, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.804, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[51] Xiang Yue, Yuansheng Ni, Kai Zhang, Tianyu Zheng, Ruoqi Liu, Ge Zhang, Samuel Stevens, Dongfu Jiang, Weiming Ren, Yuxuan Sun, et al. Mmmu: A massive multi-discipline multimodal understanding and reasoning benchmark for expert agi. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9556-9567, 2024." + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "24678" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.093, + 0.482, + 0.133 + ], + "angle": 0, + "content": "[52] Alan Yuille and Daniel Kersten. Vision as bayesian inference: analysis by synthesis? Trends in cognitive sciences, 10(7):301-308, 2006." + }, + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.136, + 0.482, + 0.205 + ], + "angle": 0, + "content": "[53] Jiyao Zhang, Weiyao Huang, Bo Peng, Mingdong Wu, Fei Hu, Zijian Chen, Bo Zhao, and Hao Dong. Omni6dpose: A benchmark and model for universal 6d object pose estimation and tracking. In European Conference on Computer Vision, pages 199-216. Springer, 2024." + }, + { + "type": "list", + "bbox": [ + 0.092, + 0.093, + 0.482, + 0.205 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24679" + } + ] +] \ No newline at end of file diff --git a/2025/Spatial457_ A Diagnostic Benchmark for 6D Spatial Reasoning of Large Mutimodal Models/96f4562f-945b-497d-9468-cd476fb5f769_origin.pdf b/2025/Spatial457_ A Diagnostic Benchmark for 6D Spatial Reasoning of Large Mutimodal Models/96f4562f-945b-497d-9468-cd476fb5f769_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..e8c5a51453ab21c6b4005fbf460e8a99adeef714 --- /dev/null +++ b/2025/Spatial457_ A Diagnostic Benchmark for 6D Spatial Reasoning of Large Mutimodal Models/96f4562f-945b-497d-9468-cd476fb5f769_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4597295c2e6f6d7d6e76f799947e4da57a82753cd92205830b016a08973f0a0b +size 1976444 diff --git a/2025/Spatial457_ A Diagnostic Benchmark for 6D Spatial Reasoning of Large Mutimodal Models/full.md b/2025/Spatial457_ A Diagnostic Benchmark for 6D Spatial Reasoning of Large Mutimodal Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..3454e71dd0a116d25cff96930ff79e1e5686e0c5 --- /dev/null +++ b/2025/Spatial457_ A Diagnostic Benchmark for 6D Spatial Reasoning of Large Mutimodal Models/full.md @@ -0,0 +1,344 @@ +# Spatial457: A Diagnostic Benchmark for 6D Spatial Reasoning of Large Multimodal Models + +Xingrui Wang $^{1}$ , Wufei Ma $^{1}$ , Tiezheng Zhang $^{1}$ , Celso M de Melo $^{2}$ , Jieneng Chen $^{1\dagger}$ , Alan Yuille $^{1\dagger}$ + +$^{1}$ Johns Hopkins University $^{2}$ DEVCOM Army Research Laboratory + +# Abstract + +Although large multimodal models (LMMs) have demonstrated remarkable capabilities in visual scene interpretation and reasoning, their capacity for complex and precise 3-dimensional spatial reasoning remains uncertain. Existing benchmarks focus predominantly on 2D spatial understanding and lack a framework to comprehensively evaluate 6D spatial reasoning across varying complexities. To address this limitation, we present Spatial457, a scalable and unbiased synthetic dataset designed with 4 key capability for spatial reasoning: multi-object recognition, 2D location, 3D location, and 3D orientation. We develop a cascading evaluation structure, constructing 7 question types across 5 difficulty levels that range from basic single object recognition to our new proposed complex 6D spatial reasoning tasks. We evaluated various large multimodal models (LMMs) on Spatial457, observing a general decline in performance as task complexity increases, particularly in 3D reasoning and 6D spatial tasks. To quantify these challenges, we introduce the Relative Performance Dropping Rate (RPDR), highlighting key weaknesses in 3D reasoning capabilities. Leveraging the unbiased attribute design of our dataset, we also uncover prediction biases across different attributes, with similar patterns observed in real-world image settings. The code is released in https://github.com/XingruiWang/Spatial457. + +# 1. Introduction + +In recent years, large multimodal models (LMMs) have demonstrated remarkable reasoning capabilities, enabling them to interpret visual scenes from images and articulate their understanding through natural language [1, 3, 11, 26, 27, 38, 39, 48]. These advancements have been evaluated using a variety of benchmarks that primarily emphasize semantic features—such as object categories and appearances, spatial relationships within the 2D image plane [15, + +17, 18] (e.g. left-right positioning from the camera view), and limited 2.5D features such as camera depth [40]. Tasks like object detection [13, 22], image captioning [21, 42, 47] and visual question answering [18, 50, 51] have significantly benefited from these evaluations, driving LMMs to new heights in visual understanding. + +However, a critical gap remains in assessing the ability of LMMs to comprehend and reason about the full six degrees of freedom (6D) of objects in spatial environments, which includes both 3D positions and orientations [43, 46, 53]. Despite its significance in many real-world applications, such as robotics [2], autonomous navigation [14], and augmented reality [5], there are no available benchmarks for LMMs that evaluate 6D spatial reasoning capabilities. + +Building a scalable diagnostic dataset for 6D vision tasks poses considerable challenges due to the extensive annotations required across vast image collections. One strategy is to develop 6D vision question-answering benchmarks on top of existing image datasets [4, 6, 7, 36]. However, real-world image collections often exhibit significant biases in 3D locations and poses, influenced by camera perspective limitations and the prevalence of common object orientations in everyday scenes. For instance, statistics from the SUN-RGBD [36] and nuScenes [7] datasets reveal that over $70\%$ of objects cluster into a single predominant direction (see Fig. 2), limiting the diversity required for comprehensive 6D reasoning evaluation. + +To overcome these challenges, we introduce an unbiased and scalable synthetic dataset rendered in a realistic style, specifically designed to diagnose the 6D spatial reasoning abilities of LMMs. To comprehensively evaluate diverse aspects of 6D spatial reasoning, we focus on four core capabilities: multi-object recognition, 2D locations, 3D locations, and 3D orientation. Multi-object recognition serves as the foundation of scene understanding, while 2D location represents the simplest way to depict the spatial relationship between objects within the image plane [18, 19]. The 3D location factor extends understanding into three-dimensional space, capturing the depth and distance of objects, which is essential for tasks such as reasoning about occlusions [9, 10, 44]. The 3D orientation introduces object + +![](images/1864bdae54cca2dc2242be14e7524808da21299f243132d64d8d7e4e7da4802c.jpg) + +![](images/30d3a8c4f56983c66f5072c4add864c722b739fc85750ef883d5c3ec0535e6fd.jpg) + +3D Orientations + +![](images/73f4fbe35d85cfb1d81613d5e40465c3a10eb66eb51d47017fe9331b1e3a2e2d.jpg) + +3D Locations + +![](images/5ca5917c849a9c43b71b82d8ae007c43a6f4ef55085d41bd963f6fcadca79188.jpg) + +2D Locations + +![](images/7284700c93f5092b896d396ebb2139ffbb6a005fe1ebc54d1d6b3cadbb359ff1.jpg) + +Multiple Objects + +![](images/e3c9dcd25aabff842a0391022215b0ee9b9a8075ef389524db34b89ec26de262.jpg) +4 Capabilities of Comprehensive Spatial Reasoning + +![](images/4bf3ba052d061251eeae6726d273807e7e69ab2f73ab853dfff8c86a9093edcf.jpg) +Figure 1. Overview of the Spatial457 benchmark. We define four core capabilities for spatial reasoning. By incorporating these capabilities step-by-step, the benchmark assesses models across five progressive difficulty levels and seven question types, ranging from single-object recognition to advanced 6D spatial relationships and collision prediction. Our evaluation demonstrates a significant performance drop for more advanced questions and highlights the gap between state-of-the-art (SoTA) models and human performance in complex spatial reasoning VQA tasks. + +![](images/991ef26e1fdb5ca8588e848748afa12a64c729a885d18c8c85f9749b0aea54a3.jpg) +(a) SUN-RGBD + +![](images/7bf783f866998671f7592a8496fdc91cc480f4b16aea96c8d3040a39fcaf0e6d.jpg) +(b) nuScenes +Figure 2. The distribution of 3D poses in representative 6D datasets (a&b) highlights an imbalance in object orientations. This inspired us to develop a balanced benchmark encompassing all degrees of 3D orientation. + +![](images/2ad47a1c7db930e43d981363ad687305e879358386caf9c6f1703cc066b49da2.jpg) +(c) Ours + +rotation [44, 46], enabling models to reason about precise 3D poses of objects within a scene. + +By progressively incorporating the four core capabilities, we establish five levels of difficulty, resulting in seven distinct question types. These questions span a spectrum of spatial reasoning complexity, ranging from basic semantic identification and 2D spatial relationships to advanced 3D positional understanding and orientation-based queries, culminating in comprehensive 6D spatial reasoning tasks. + +Our evaluation measures performance across varied question types and examines how it is affected by different combinations of the four capabilities. This tiered design provides comprehensive benchmarks for spatial reasoning capabilities, highlighting each model's strengths and areas of vulnerability. To the best of our knowledge, this work presents the first comprehensive benchmark designed to evaluate the 6D spatial reasoning capabilities of vision- + +language models. + +Additionally, our dataset's unbiased attribute design allows us to analyze potential biases in model predictions. By systematically exploring model performance under different attributes, we identify tendencies or limitations that may affect real-world applications. Altogether, our benchmark not only provides a foundational tool for evaluating LMMs but also serves as a guide for developing future benchmarks that support models with advanced 6D spatial reasoning. + +Our main contributions are threefold: + +1. An unbiased and scalable synthetic dataset: We present a realistic, unbiased synthetic dataset designed to assess 6D spatial reasoning in LMMs. We define four core capabilities and propose five difficulty levels for spatial reasoning VQA with seven question types. +2. Comprehensive benchmarking of 6D capabilities: We benchmark various LMMs across all difficulty levels. The performance variations across settings reveal potential limitations or strengths, providing valuable insights into which components should be prioritized to improve the spatial reasoning of LMMs. +3. Bias analysis in model predictions: Leveraging the unbiased nature of our dataset, we conduct an in-depth bias analysis to explore how LMM predictions may be unbalanced across different attributes, especially 3D-related ones. This reveals potential weaknesses in model predictions, offering insights into limitations that may also appear in real-world scenarios. + +# 2. Related Work + +# 2.1. Synthetic Datasets in VQA + +The use of synthetic data in visual question answering (VQA) has gained traction as a means to circumvent the limitations and biases inherent in real-world datasets. Synthetic environments [16, 30, 34] allow for controlled manipulation of scene elements, enabling researchers to probe specific reasoning abilities in VQA models without the confounding factors often present in natural imagery. Previous works [19, 24, 25] have utilized synthetic environments to evaluate multimodal models across a range of visual reasoning tasks, typically focusing on 2D spatial relationships and object features. For instance, CLEVR [19], a synthetic dataset widely used in the VQA community, evaluates compositional reasoning by introducing artificially generated scenes with diverse object arrangements. Super-CLEVR-3D [45] represents a significant step forward in evaluating 3D locations, it lacks an assessment of 6D spatial relationships that consider both 3D location and 3D orientation. Our work extends this tradition by creating a synthetic dataset tailored for 6D spatial reasoning, filling a gap in current synthetic datasets that do not address the complete spatial understanding of objects in 3D space. + +# 2.2. LMM Benchmarks + +Existing benchmarks [15, 18, 35, 51] for large multimodal models (LMMs) predominantly assess their performance on tasks involving 2D visual question answering and multimodal reasoning. For instance, datasets like VQAv2 [15] and GQA [18] focus on answering questions based on visual content. However, these benchmarks leave a gap in evaluating models' capabilities in spatial contexts. Recent benchmarks, such as CV-Bench [40], assess spatial understanding, including depth ordering for 2.5D evaluation. More recent work [9, 10, 32] introduces questions regarding either aspect of 3D locations or rotations. Our proposed benchmark goes beyond traditional 2D, 2.5D, and 3D location or rotation assessments by introducing tasks that challenge LMMs to understand 3D positions and orientations together. We also evaluate models in a cascading manner across five difficulty levels and summarize their performance in four reasoning abilities, along with their output biases, providing a deeper analysis of the weaknesses in spatial reasoning within multimodal AI. + +# 3. Spatial457: Comprehensive Spatial Reasoning Benchmark + +To comprehensively evaluate the spatial reasoning capabilities of vision-language models, we introduce a synthetic dataset named Spatial457. We define four core capabilities for comprehensive spatial reasoning (Sec. 3.1) and generate scenes with full 6D annotations (Sec. 3.2). The full combination of these capabilities form the new 6D spatial + +relationship VQA tasks (Sec. 3.3). + +To thoroughly assess model capabilities, we define five levels of difficulty in vision question answering (VQA) tasks, converting seven question subsets that range from single-object recognition to the newly proposed 6D spatial reasoning tasks (Sec. 3.4). This structured progression in question difficulty is achieved by sequentially introducing new spatial reasoning capabilities, as illustrated in Fig. 1. + +# 3.1. Four Capabilities for Spatial Reasoning + +We define 4 capabilities for comprehensive spatial understanding in a 3D scene: multiple objects, 2D locations, 3D locations, and 3D orientations. + +(1) Multiple objects. Objects are fundamental to defining scene structure [52]. Large vision-language models show strong results in aligning semantic features with text [23, 31], handling multiple objects from scene becomes the first toward scene understanding. Models must recognize individual attributes and interpret relationships among them. Our dataset evaluates multi-object reasoning, such as comparison and counting. For example, in Fig. 1, the image shows two cyan objects, requiring the model to verify color simultarity and count cyan objects accurately. +(2) 2D locations. 2D spatial relationships are essential in prior image-based VQA tasks [18, 19]. In traditional multimodal models [20, 28, 49], these relationships can be learned through object detection training [8, 33]. However, in large-scale vision-language models, explicit location information is often absent from the latent representations. Following Li et al. [24], we examine four spatial relationships (left, right, front, back) between objects from 2D camera views as the first taks of spatial reasoning. Questions use these relations to filter targets, such as "What is the shape of the object to the left of the chopper?" (see Fig. 1). +(3) 3D locations. Beyond 2D relationships, 3D locations can incorporate depth, capturing spatial relationships that 2D alone cannot represent. This is crucial for tasks involving occlusions and complex spatial arrangements. Our dataset assesses the 3D location understanding abilities of model through questions about occlusions query as in [45]. For example, as shown in Fig. 1, questions may ask about an object's position relative to others in 3D space or whether one object is occluded, challenging models to judge distances and spatial positions accurately. +(4) 3D orientations. The 3D orientation of objects introduces an additional layer of complexity by incorporating rotational attributes. In current computer vision tasks, 3D orientation is often derived by pose estimation methods. However, large-scale vision-language models generally lack explicit 3D orientation information in their latent representations, potentially making it challenging to accurately infer the object's orientations. In our dataset, we eval + +uate models' abilities to understand 3D orientations by focusing on object poses, including their facing direction and alignments with others. For example, as depicted in Fig. 1, the model may be asked to identify whether a cyan object is facing left, oriented parallel to a brown object, or positioned in a specific spatial relationship — i.e. left, right, front, or behind— from the perspective of the objects themselves. + +# 3.2. Scene Rendering + +We build upon the generation pipeline from prior VQA datasets [19, 24] and uniformly select 21 object models across 5 categories to construct the dataset as Li et al. [24]. To improve the realism of our rendered dataset, we use real image style environment maps and refined object textures. For benchmarking new 6D spatial understanding ability, we add new 6D spatial relationship annotations among objects. + +Background. To create realistic scene backgrounds, we use HDRI images $^2$ with environment maps and natural lightings. We selected 224 different outdoor scenes with normal brightness levels without high light contrast by their metadata to maintain visual consistency across images. + +Objects. We reconstruct 3D mesh models [24] with realistic textures, introducing more variation and realistic appearances for objects. Similar to Li et al. [24], each object in our dataset is defined by distinct attributes and precise spatial locations. The original attributes include shape, color, and size. For benchmarking 6D spatial reasoning, each object also includes a vector specifying its 2D location $(\mathbf{x},\mathbf{y})$ , 3D world coordinates $(\mathrm{X},\mathrm{Y},\mathrm{Z})$ , and pose direction derived from its 3D orientation [45]. + +Relationships. As shown in Fig. 3 (a.1 to a.3), we define three types of spatial relationships: 1. 2D Spatial Relationships: These describe the relative positioning of objects within the 2D camera view, using terms such as left, right, front, and behind. 2. 6D Spatial Relationships: Extending spatial descriptions into 3D space, these relationships incorporate both the 3D position and orientation of objects to define relative positions from a target object's perspective, including left, right, front, and behind. 3. Collision Relationships: This relationship describes potential collisions if the target object moves forward or backward. Collision prediction involves assessing the orientation and 3D location of target objects relative to others to determine possible collisions. + +# 3.3. New Questions for 6D Spatial Reasoning + +Based on the four capabilities defined in Sec. 3.1, we construct new question types for 6D spatial reasoning and collision prediction. These questions require an understanding of both 3D location and 3D orientation across multiple objects to generate accurate answers. This level of complexity represents the L5 difficulty level shown in Fig. 1. + +![](images/f6fcdc111cff4c02fce2f5207fed5572b3d475b0be5ee653357e91a05618a402.jpg) +Q: How many yellow objects are to the left of the red wagon? A: 1 + +![](images/0a2a36af4790072fcdea3e45232b9148a8f57a15edfce4180dd57bbaa769eeed.jpg) +Q: How many yellow objects are on the left hand side of the red wagon? A: 3 + +![](images/1871951b8fe3cb2330b8bf2b78e45e41803a43513be1565e8d953b2dead326b3.jpg) +Q: Is the red wagon will collide with a yellow regular bus if it moving forward? + +![](images/e5009a32a1bb030559d3c89e9a090fefbcb9f87618f8510654f493c0d8f6fc9d.jpg) +(a.1) 2D spatial +(b.1) Programs of 6D spatial question + +![](images/53f9d9d3ce6033dd56704d041a039483042b573f2b5f7e60b536bb6ba17b6dd0.jpg) +(a.2) 6D spatial +A: No +(a.3) Collision +(b.2) Programs of collision prediction question +Figure 3. (a.1-3) Example of the 2D spatial, 3D spatial and collision with images and questions. (b.1-2) shows the new operation programs we generated for the new 6D spatial reasoning questions. + +The 6D spatial relationship questions challenge the model to understand the spatial relationship involving both 3D position and orientation from a target object's perspective (shown in Fig. 3 (a.2)). Compared with prior VQA benchmarks focusing on 2D [19, 24], we introduce new 3D spatial reasoning operations, 3D Relate( $\cdot$ ), to determine relative positions like left, right, front, and behind within a 3D coordinate space (see Fig. 3 (b.1)). For example, a question might ask, "How many objects are to the right of the red wagon?" requiring the model to filter and count objects based on their orientation and relative 3D position, which would yield different answers compared to similar 2D left and right questions (Fig. 3 (a.1)). This task is particularly challenging as it requires models to integrate 3D spatial relationships while managing occlusions and depth perception—capabilities not typically covered in simpler 2D spatial tasks. + +The collision prediction questions, as shown in Fig. 3 (a.3), extend 6D spatial reasoning by introducing motion-based inference. These questions assess the model's ability to anticipate future states and potential interactions (collisions) based on accurate estimation of 3D location and direction, denoted as forward or backward. For instance, a question might ask, "Will the red wagon collide with a yellow bus if it moves forward?" This requires the model to consider the orientation and direction of movement relative to surrounding objects to predict possible outcomes. Such tasks test the model's advanced spatial reasoning about 3D orientation and position, and its capacity to predict accurate trajectories and directions within the future scene. + +# 3.4. Questions under Other 4 Difficulty Levels + +The L5 questions above represent the most challenging cases for 6D spatial reasoning, testing all four capabilities + +in Sec. 3.1. However, the purpose of this benchmark is to comprehensively diagnose each factor in spatial reasoning. Thus, we design a structured roadmap to evaluate the model's performance across different levels of spatial reasoning, assessing its capabilities in each specific factor. + +By composing the four factors of comprehensive spatial reasoning, we define an additional four levels of question difficulty based on the generated scenes, encompassing a total of seven tasks in our benchmark. As the complexity of spatial reasoning increases with the introduction of additional factors, we organize these questions into a structured roadmap, as illustrated in Fig. 1. Although many questions are reimplementations from previous benchmarks [24, 45] on our new images, we aim to integrate these into a unified benchmark to provide deeper insights into the model's spatial reasoning capabilities. The composition of the four capabilities at each level, along with comparisons to previous VQA benchmarks, is shown in Tab. 1. + +Level 1 - single object questions [19, 24]. We begin with the questions with single objects. These questions are generated using a filtered set of templates from Li et al. [24]. We select five templates involving only one object in the question, such as "What color is the double bus?" in Fig. 1. This question type provides a base score for the tested vision-language model. The comparison with L1 helps to determine whether the decrease of performance are due to challenge of spatial reasoning or object recognition. + +Level 2 - multiple objects questions [19, 24]. Level 2 questions extend Level 1 by involving multiple objects. These questions involve understanding the relationship between objects and counting. For example, "Is there another object of the same color as the double bus?" as shown in Fig. 1, where the question templates follow Li et al. [24]. We evaluate the model's capacity for reasoning in scenes with multiple entities as a critical step toward understanding spatial relationships in next levels. + +Level 3 - 2D spatial questions [19, 24]. Level 3 questions further introduce 2D spatial relationships among multiple objects, requiring the model to understand the position of objects and their relationships from the 2D camera's perspective. For example, "What shape is the object to the left of the chopper?" as depicted in Fig. 1. The 2D relationship is the first step of spatial reasoning. + +Level 4 - 3D pose questions [45]. Level 4 questions extend 3D orientation reasoning beyond the 2D spatial relationships in Level 3. Beyond appearance attributes (shape, color, size) or 2D locations, these questions require understanding 3D orientation, including the rotation angle or facing direction of an object from the camera's perspective. For example, "What shape is the object that is parallel to the brown object?" as shown in Fig. 1. 3D pose questions test the model's ability to interpret scenes with 3D awareness, involving rotation and alignment with other objects. + +Level 4 - occlusion questions [45]. In addition to 3D orientation, incorporating 3D location presents another challenge beyond the Level 3 questions on 2D spatial relationships. A unique aspect of 3D spatial reasoning from images is occlusion. The occlusion questions in Level 4 require the model to comprehend depth perception and the relative positions of objects in 3D space, determining which object is closer to the camera within overlapping areas. For example, an occlusion question might ask, "Is the red wagon occluded by the yellow bus?" as shown in Fig. 1. This level evaluates the model's ability to reason about occlusions and depth perception, which are essential for understanding complex spatial relationships in 3D scenes. + +
Question LevelMulti-object2D location3D location3D OrientationVQA Benchmarks
L1spatial-agnostic bench (e.g., TextVQA [35])
L2✓CLEVR-Math [25], CLEVR-Count [19]
L3✓✓VQA [15], GQA [18], CLEVR [19], Super-CLEVR [24]
L4 (a)✓✓✓CV-Bench (DepthOrder) [40], Super-CLEVR-3D [45]
L4 (b)✓✓✓Super-CLEVR-3D [45]
L5✓✓✓✓Spatial457
+ +Table 1. Comparison of VQA Benchmarks by Factors and Question Levels. Compared with existing VQA benchmarks, Spatial457 includes all four spatial reasoning factors, providing a comprehensive evaluation of 6D spatial understanding. + +# 3.5. Evaluation Protocols + +To better evaluate the model's spatial reasoning capabilities, we propose a structured evaluation protocol that measures the relative capability of models on each factor for spatial reasoning. We define this evaluation metric as the Relative Performance Dropping Rate (RPDR), which assesses the decline in performance after introducing a new factor in the questions. Based on the roadmap in Fig. 1, the RPDR for each factor is defined as follows: + +$$ +\mathrm {R P D R} _ {\text {M u l t i - O b j .}} = A c c _ {\text {L e v e l 1}} \Rightarrow A c c _ {\text {L e v e l 2}} +$$ + +$$ +\mathrm {R P D R} _ {\mathrm {2 D L o c .}} = A c c _ {\text {L e v e l 2}} \Rightarrow A c c _ {\text {L e v e l 3}} +$$ + +$$ +\mathrm {R P D R} _ {\mathrm {3 D L o c .}} = \operatorname {A v g} \left( \begin{array}{c} A c c _ {\text {L e v e l 3}} \Rightarrow A c c _ {\text {L e v e l 4 (O c c .)}}, \\ A c c _ {\text {L e v e l 4 (3 D P o s e)}} \Rightarrow A c c _ {\text {L e v e l 5 (6 D)}} \end{array} \right) +$$ + +$$ +\mathrm {R P D R} _ {\mathrm {3 D O r i .}} = \operatorname {A v g} \left( \begin{array}{c} A c c _ {\text {L e v e l 3}} \Rightarrow A c c _ {\text {L e v e l 4 (3 D P o s e)}}, \\ A c c _ {\text {L e v e l 4 (O c c .)}} \Rightarrow A c c _ {\text {L e v e l 5 (C o l .)}} \end{array} \right) +$$ + +where $Acc_{\mathrm{Level}}$ denotes the model's accuracy on the corresponding questions and $(Acc_{\mathrm{Q1}} \Rightarrow Acc_{\mathrm{Q2}}) = \frac{Acc_{\mathrm{Q2}} - Acc_{\mathrm{Q1}}}{Acc_{\mathrm{Q1}}}$ denotes the relative performance dropping rate from $Q1$ to $Q2$ . The results of RPDR will be discussed in Sec. 4.3. + +# 4. Experiments + +# 4.1. Experimental Setup + +We generate 1,000 images to test the performance of large vision-language models. For the seven types of questions, we generate 1000 single-object questions (L1-Single), 5,000 + +
Difficulty LevelLevel 1Level 2Level 3Level 4Level 5
Questions typeSingle ObjectMulti-Obj.2D SpatialOcclusion3D PoseCollisions6D Spatial
Random33.0532.7733.4722.0418.9921.0219.41
GPT-4o [29]74.4662.8856.1448.4042.4138.4137.01
GeminiPro 1.5 [37]73.2662.5454.4947.6543.6741.1939.36
Claude 3.5 Sonnet [3]68.2457.4054.1930.8438.4035.3433.48
Qwen2-VL-7B-Inst. [39]71.9661.4455.3427.8734.2936.5833.75
InternVL2-8B [38]58.1158.7657.4033.2532.3234.3034.30
LLaVA-v1.5-7B [26]44.8744.7242.2024.3424.5523.6323.86
LLaVA-NeXT-vicuna-7B [27]50.7249.4746.0129.6829.3531.9531.95
LLaVA-NeXT-Ilama3-8B [41]52.1549.7345.9230.3129.7732.1232.12
PO3D-VQA [45]86.4682.5580.6470.4981.4068.1271.06
Human89.9786.8384.9582.7684.9581.8279.94
+ +Table 2. Performance comparison across all 7 question types under 5 difficulty levels. As task complexity increases from single-object questions to 6D spatial relationships, performance generally decreases in most cases, highlighting the challenges in handling multi-object interactions, 3D orientations, and predictive spatial reasoning. The highest score among all models for each question type is marked in red , and the second-highest scores are underlined. The best performance among open-source models is marked in blue . PO3D-VQA is the state-of-the-art (SOTA) 3D-aware neural symbolic reasoning model as proposed in [45], and is listed as the potential upper bound of evaluated methods. The comparison with human performance also shows an increasing gap for LLMs as task difficulty increases. + +multiple-object questions (L2-Multi-object.), 4,995 2D spatial relationship questions (L3-2D-Spatial), 3,534 occlusion questions (L4-Occlusion), 4,555 3D pose questions (L4-3D-Pose), 4,357 collision questions (L5-Collision), and 4,995 6D spatial relationship questions (L5-6D-Spatial). + +We evaluate two series of large vision-language models on Spatial457: (1) API-based large language models, including GPT-4o [29], Gemini-1.5-Pro [37], and Claude-3.5-Sonnet [3]; and (2) open-source large vision-language models, including Qwen2-VL-7B-Instruct [39], InternVL2-8B [38], llava-v1.5-7b [26], llava next-vicuna-7b [27], and llava next-llama3-8b [41]. The evaluation follows the format of Vlmealkit [12]. The details of the model implementation including prompts are listed in Appendix. + +# 4.2. Evaluation Results + +In Tab. 2, we present the performance of various large vision-language models across different levels of difficulty on Spatial457. The levels range from basic single-object tasks to complex 6D spatial reasoning, with increasing difficulty as more factors are introduced. We compare both API-based models (e.g., GPT-4o, GeminiPro1.5, Claude3-5V) and open-source models (e.g., Qwen2-VL-7B-Instruct, InternVL2-8B, LLaVA series) to assess their spatial reasoning capabilities. + +Comparison across difficulty levels. We evaluate model performance across different levels of difficulty in Spatial457, ranging from basic single-object tasks to complex 6D spatial reasoning. As shown in Tab. 2, performance generally decreases as the complexity of spatial relationships increases, highlighting the challenges models face + +when handling multi-object interactions, 3D orientations, and predictive spatial reasoning tasks. + +Comparison between API models and open-source models. From Tab. 2, we observe that API models generally outperform open-source models across all difficulty levels. In 2D spatial relationship questions, InternVL2-8B achieves slightly higher results than the API models, while GPT-4o is only $1\%$ lower. Qwen2-VL-7B-Instruct achieves comparable results, especially in simpler tasks like single-object and multiple-object reasoning. This analysis highlights the strengths of API models in complex spatial reasoning tasks. + +Comparison among API models. API-based models, i.e. GPT-4o, GeminiPro1.5, and Claude3-5V, demonstrate strengths across all difficulty levels. For instance, GPT-4o excels in 2D and 3D spatial tasks, while GeminiPro1.5 shows robust performance in collision prediction tasks. This comparison provides insights into the unique capabilities and limitations of each API model in handling spatial reasoning across different levels of complexity. + +# 4.3. RPDR Analysis for 4 Capabilities + +As outlined in Sec. 3.5, we use the Relative Performance Dropping Rate (RPDR) to analyze model performance across the four core factors of Spatial457: multi-object reasoning, 2D location understanding, 3D location comprehension, and 3D orientation interpretation. The RPDR analysis provides a structured approach to quantify the decline in accuracy with the introduction of each new factor, as detailed in Tab. 3. Our findings reveal that all models exhibit weaknesses in 3D reasoning. Specifically, GPT-4o and GeminiPro-1.5 perform the worst in 3D orientation tasks, + +while the other two struggle most with 3D location. + +
ModelMulti.2D Loc.3D Ori.3D Loc.
GPT-4o [29]84.4589.2777.4686.74
GeminiPro 1.5 [37]85.3787.1383.2888.79
Claude 3.5 S. [3]84.1294.4185.4372.05
Qwen2-VL-7B [39]85.3990.0680.9974.40
+ +# 4.4. Bias Analysis + +In this section, we analyze prediction biases in large language models (LLMs) regarding their understanding of key attributes, specifically shape, color, size, and pose. Leveraging the unbiased generation of dataset images and questions in Spatial457, we examine how these models may demonstrate attribute-specific biases in their predictions. + +We first compute the confusion matrices and distribution statistics for both ground truth and model predictions across each attribute (shape, color, size, and pose). For example, the confusion matrix plots for color and pose attributes of GPT-4o and GeminiPro1.5, together with the distribution of prediction and ground truth label are shown in Fig. 4. We find both GPT-4o and GeminiPro1.5 exhibit bias in both color and pose predictions. These results highlight areas for improvement to reduce bias and enhance predictive reliability across models. + +To quantify prediction variability, we calculate the Coefficient of Variation (CV) for each attribute as follows: + +$$ +\mathrm {C V} = \frac {\operatorname {S t d} \left(p _ {1} , \dots , p _ {k}\right)}{\operatorname {M e a n} \left(p _ {1} , \dots , p _ {k}\right)}, \tag {1} +$$ + +where $\operatorname{Std}(\cdot)$ and $\operatorname{Mean}(\cdot)$ are the standard deviation and mean of the probability distribution $p_i$ for each attribute label $i$ (e.g., left), "right" for pose). + +Tab. 4 shows the CV values for shape, color, size, and pose attributes across different models. Lower CV values indicate more consistent predictions with less bias, which reveals significant biases in model predictions across attributes. Higher CV values, especially in color and pose, indicate consistency challenges. + +# 4.5. Comparison to the 3D Symbolic Model + +As neural symbolic models achieve state-of-the-art (SOTA) performance on many prior synthetic datasets, we implement the 3D-aware neural symbolic model PO3DVQA [45] on our new image domain and all levels of questions. We train PO3D-VQA on a separate set of 20,000 images and questions and report its performance on the same test dataset in Tab. 2. A direct comparison between the 3D neural symbolic model and large language models is not entirely fair, as each module of the neural symbolic model (6D pose estimation, attribute classification, and language reasoning) is trained independently as a sub-task. However, we + +![](images/1269a9c5dd214d8344a48f511976c9e6edf5d8b671ec81807079c72f0fcb378c.jpg) +(b) GeminiPro1-5 +Figure 4. The distribution of color and pose attributes in the L4-3D-Pose task for GPT-4o and Gemini-Pro1.5. Although the ground truth labels are well balanced, the predicted colors and poses for both models are imbalanced, with "yellow" being the most frequent color and "front" the most frequent pose. + +Table 3. RPDR (%) analysis on the four spatial reasoning capabilities, where higher values indicate better performance. For each row, we highlight the lowest RPDR score among all factors. + +
ModelsAvg. ↓ShapeColorSizePose
GPT4o [29]0.4850.6230.5210.0430.752
GeminiPro1.5 [37]0.4980.5670.3660.2970.760
Claude3.5V-S [3]0.7200.7770.4650.5741.063
Qwen2-VL-7B [39]0.5450.8680.5580.0670.688
Human0.0930.1100.0820.0550.126
+ +Table 4. Coefficient of Variation (CV) for predictions on each attribute in the L4-3D-Pose task. We evaluate three API models and one open-source model, calculating the CV across four attributes (shape, color, size, 3D pose) based on the distribution of predicted labels. Lower CV values indicate less bias. The first column shows the average CV across attributes, and GPT4-o has the lowest bias. + +include this model to illustrate the upper bound that a data-driven machine learning model can achieve on our dataset. + +# 4.6. Qualitative Results + +Fig. 5 presents an example of GPT-4o's performance across various spatial reasoning tasks on the same image. GPT-4o correctly answers multi-object, 2D spatial, and occlusion tasks but fails on 3D pose and 6D spatial questions, indicating challenges in 3D orientation understanding in this case. As we prompt the models to output the answer as well as the reasoning steps, the models correctly detect the existence of objects (e.g. in Fig. 5 the yellow minivan in L4-Pose or the fighter plane and road bike in L5 questions, but has the wrong understand to the orientations or 3D locations. + +![](images/7fa982bf2e5866389df7429e32bea4da7f224784da2851f9cb02347b8566918b.jpg) +Figure 5. Example of GPT-4o across different spatial reasoning tasks on the same image with the ground truth answer marked in blue. GPT-4o answers multi-object, 2D spatial, and occlusion tasks correctly (in the green block) but fails on 3D pose and 6D spatial questions (the red block). + +# 5. Extension to Real Dataset + +We extend the setting of 3D questions to real-world image data. Due to limitations in fully annotated attributes such as category, color, 3D orientation, position, and occlusion, we focus on questions from the L4-Pose category. Using the scene structure from SUN-RGBD [36], we construct a new subset, L4-Pose-Real, which involves only the category and pose attributes. To reduce ambiguity in object localization, we add a red bounding box around target objects, ensuring clarity in identifying categories within the scene. + +We provide the question along with the image as input to the LLM. In this example in Fig. 6, GPT-4o provides the correct answer, while Gemini-Pro 1.5 is incorrect. However, the reasoning process reveals that both models rely on common sense knowledge or 2D visibility cues to interpret the 3D orientation. + +![](images/760ee8a6c6463ba8e962e210d939301f889828d4dfff2cc5f58051e282124c79.jpg) +Figure 6. Example of real-world 3D pose reasoning tasks $L4$ -Pose-Real. Here we show the reasoning and answers provided by LLMs where GPT-4o is correct and Gemini-Pro 1.5 is wrong. But the reasoning process reveals these two LLMs are both reasoning 3D orientation from the command sense or 2D visibility. + +As shown in Tab. 5, Gemini Pro 1.5 achieves the highest accuracy (40.50%) in L4-Pose-Real, and the gap between the three API-based models and the open-source + +model Qwen2-VL-7B-Instruct (26.50%) is larger. However, all these models perform significantly lower than humans, which have accuracy 92.64%. We further analyze the bias of predictions in this real-world question set. Among the models, Claude3-5V Sonnet exhibits the lowest CV (0.374), indicating relatively stable predictions, while Qwen2-VL-7B-Instruct has the highest CV (0.934). The distribution of ground truth and predicted pose and the confusion matrix are shown in Fig. 7. + +
ModelsAccuracy (%) ↑CV ↓
GPT-4o [29]38.750.690
GeminiPro1-5 [37]40.500.762
Claude3-5V Sonnet [3]38.750.374
Qwen2-VL-7B-Instruct [39]26.500.934
Human92.640.197
+ +Table 5. Accuracy and Coefficient of Variation (CV) for L4-Poserel task: Comparison of accuracy and CV for GPT-4o, GeminiPro1.5, Claude3-5V, and Qwen2-VL-7B-Instruct on real-world 3D pose tasks. Human performance is included as a reference. + +![](images/a544513755b30e876e51395ca63b4f69cbfe29077150cfa24ebd01acf46cd70d.jpg) +Figure 7. Distribution of 3D poses for in L4-Pose-Real task. + +# 6. Conclusion + +In this work, we introduce Spatial457, a scalable and unbiased benchmark designed to comprehensively assess 6D spatial reasoning capabilities in LMMs. Spatial457 extends traditional 2D and 3D VQA benchmarks by incorporating complex 6D spatial reasoning tasks, structured across five levels of difficulty. Through evaluations, we observe that current LMMs demonstrate strong performance in basic object recognition and 2D spatial relationships, but struggle significantly with 3D spatial understanding and collision prediction tasks. Additionally, our analysis of prediction biases uncovers attribute-specific inconsistencies that could affect model reliability in real-world applications. Limitations. To real images settings, current evaluation limits to L3 pose questions. We are actively working on building the additional required annotations for these real images to expand the domain of the proposed benchmark. + +# Acknowledgement + +This work is supported by ONR with award N00014-23-1-2641, from ARL Army Research Laboratory with award W911NF2320008. + +# References + +[1] Alibaba DAMO Academy. Gemini: Vision-language model by alibaba. Alibaba DAMO, 2023. +[2] Boshi An, Yiran Geng, Kai Chen, Xiaqi Li, Qi Dou, and Hao Dong. Rgbmanip: Monocular image-based robotic manipulation through active object pose estimation. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 7748-7755. IEEE, 2024. +[3] Anthropic. Claude: An anthropic language model. Anthropotic, 2023. +[4] Gilad Baruch, Zhuoyuan Chen, Afshin Dehghan, Tal Dimry, Yuri Feigin, Peter Fu, Thomas Gebauer, Brandon Joffe, Daniel Kurz, Arik Schwartz, et al. Arkitsscenes: A diverse real-world dataset for 3d indoor scene understanding using mobile rgb-d data. arXiv preprint arXiv:2111.08897, 2021. +[5] Mark Billinghurst, Adrian Clark, and Gun Lee. A survey of augmented reality. Foundations and Trends in Human-Computer Interaction, 8(2-3):73-272, 2015. +[6] Garrick Brazil, Abhinav Kumar, Julian Straub, Nikhila Ravi, Justin Johnson, and Georgia Gkioxari. Omni3d: A large benchmark and model for 3d object detection in the wild. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13154-13164, 2023. +[7] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11621-11631, 2020. +[8] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. +[9] Boyuan Chen, Zhuo Xu, Sean Kirmani, Brain Ichter, Dorsa Sadigh, Leonidas Guibas, and Fei Xia. Spatialvm: Endowing vision-language models with spatial reasoning capabilities. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14455-14465, 2024. +[10] An-Chieh Cheng, Hongxu Yin, Yang Fu, Qiushan Guo, Ruihan Yang, Jan Kautz, Xiaolong Wang, and Sifei Liu. Spatial-rgpt: Grounded spatial reasoning in vision language model. arXiv preprint arXiv:2406.01584, 2024. +[11] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhanghao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E Gonzalez, et al. Vicuna: An open-source chatbot impressing gpt-4 with $90\%$ chatgpt quality, march 2023. URL https://lmsys.org/blog/2023-03-30-vicuna, 3(5), 2023. +[12] Haodong Duan, Junming Yang, Yuxuan Qiao, Xinyu Fang, Lin Chen, Yuan Liu, Xiaoyi Dong, Yuhang Zang, Pan Zhang, + +Jiaqi Wang, et al. Vlmealkit: An open-source toolkit for evaluating large multi-modality models. In Proceedings of the 32nd ACM international conference on multimedia, pages 11198-11201, 2024. +[13] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. International Journal of Computer Vision, 88(2):303-338, 2010. +[14] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 3354-3361, 2012. +[15] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Batra, and Devi Parikh. Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6904-6913, 2017. +[16] Roland Hess. *Blender foundations: The essential guide to learning blender* 2.5. Routledge, 2013. +[17] Yushi Hu, Hang Hua, Zhengyuan Yang, Weijia Shi, Noah A Smith, and Jiebo Luo. Promptcap: Prompt-guided task-aware image captioning. arXiv preprint arXiv:2211.09699, 2022. +[18] Drew A Hudson and Christopher D Manning. Gqa: A new dataset for real-world visual reasoning and compositional question answering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6700-6709, 2019. +[19] Justin Johnson, Bharath Hariharan, Laurens Van Der Maaten, Li Fei-Fei, C Lawrence Zitnick, and Ross Girshick. Clevr: A diagnostic dataset for compositional language and elementary visual reasoning. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2901–2910, 2017. +[20] Aishwarya Kamath, Mannat Singh, Yann LeCun, Gabriel Synnaeve, Ishan Misra, and Nicolas Carion. Mdetr-modulated detection for end-to-end multi-modal understanding. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1780-1790, 2021. +[21] Dong-Jin Kim, Jinsoo Choi, Tae-Hyun Oh, and In So Kweon. Dense relational captioning: Triple-stream networks for relationship-based captioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6271-6280, 2019. +[22] Alina Kuznetsova, Hassan Rom, Neil Alldrin, Jasper Uijlings, Ilya Krasin, Jordi Pont-Tuset, Gleb Muryshkin, Sergey Popov, Marwa Elhachem, Pedro de Carvalho, et al. The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale. International Journal of Computer Vision, 128(7):1956-1981, 2020. +[23] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730-19742. PMLR, 2023. + +[24] Zhuowan Li, Xingrui Wang, Elias Stengel-Eskin, Adam Kortylewski, Wufei Ma, Benjamin Van Durme, and Alan L Yuille. Super-clevr: A virtual benchmark to diagnose domain robustness in visual reasoning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14963-14973, 2023. +[25] Adam Dahlgren Lindstrom and Savitha Sam Abraham. Clevr-math: A dataset for compositional language, visual and mathematical reasoning. arXiv preprint arXiv:2208.05358, 2022. +[26] Haotian Liu et al. Llava: Large language and vision assistant. LLaVA Project, 2023. +[27] Haotian Liu et al. Llava-last: Extended large language and vision assistant. LLaVA Project, 2023. +[28] Jiayuan Mao, Chuang Gan, Pushmeet Kohli, Joshua B Tenenbaum, and Jiajun Wu. The neuro-symbolic concept learner: Interpreting scenes, words, and sentences from natural supervision. arXiv preprint arXiv:1904.12584, 2019. +[29] OpenAI. Gpt-4 technical report. OpenAI, 2023. +[30] Weichao Qiu and Alan Yuille. Unrealcv: Connecting computer vision to unreal engine. In Computer Vision-ECCV 2016 Workshops: Amsterdam, The Netherlands, October 8-10 and 15-16, 2016, Proceedings, Part III 14, pages 909-916. Springer, 2016. +[31] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. +[32] Arjit Ray, Jiafei Duan, Reuben Tan, Dina Bashkirova, Rose Hendrix, Kiana Ehsani, Aniruddha Kembhavi, Bryan A Plummer, Ranjay Krishna, Kuo-Hao Zeng, et al. Sat: Spatial aptitude training for multimodal language models. arXiv preprint arXiv:2412.07755, 2024. +[33] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. IEEE transactions on pattern analysis and machine intelligence, 39(6):1137-1149, 2016. +[34] Andrew Sanders. An introduction to Unreal engine 4. AK Peters/CRC Press, 2016. +[35] Amanpreet Singh, Vivek Natarajan, Meet Shah, Yu Jiang, Xinlei Chen, Dhruv Batra, Devi Parikh, and Marcus Rohrbach. Towards vqa models that can read. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8317-8326, 2019. +[36] Shuran Song, Samuel P Lichtenberg, and Jianxiong Xiao. Sun rgb-d: A rgb-d scene understanding benchmark suite. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 567-576, 2015. +[37] Gemini Team, Petko Georgiev, Ving Ian Lei, Ryan Burnell, Libin Bai, Anmol Gulati, Garrett Tanzer, Damien Vincent, Zhufeng Pan, Shibo Wang, et al. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context. arXiv preprint arXiv:2403.05530, 2024. +[38] Intern Team. Internvl 2.0: A vision-language model. Intern Foundation, 2023. + +[39] Qwen Team. Qwen-vl: A large multimodal model. Alibaba Group, 2023. +[40] Shengbang Tong, Ellis Brown, Penghao Wu, Sanghyun Woo, Manoj Middepogu, Sai Charitha Akula, Jihan Yang, Shusheng Yang, Adithya Iyer, Xichen Pan, et al. Cambrian1: A fully open, vision-centric exploration of multimodal llms. arXiv preprint arXiv:2406.16860, 2024. +[41] Hugo Touvron et al. Llama: Open and efficient foundation language models. Meta AI, 2023. +[42] Oriol Vinyals, Alexander Toshev, Samy Bengio, and Dumitru Erhan. Show and tell: A neural image caption generator. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3156-3164, 2015. +[43] He Wang, Srinath Sridhar, Jingwei Huang, Julien Valentin, Shuran Song, and Leonidas J Guibas. Normalized object coordinate space for category-level 6d object pose and size estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2642-2651, 2019. +[44] Xingrui Wang, Wufei Ma, Zhuowan Li, Adam Kortylewski, and Alan Yuille. 3d-aware visual question answering about parts, poses and occlusions. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. +[45] Xingrui Wang, Wufei Ma, Zhuowan Li, Adam Kortylewski, and Alan L Yuille. 3d-aware visual question answering about parts, poses and occlusions. Advances in Neural Information Processing Systems, 36, 2024. +[46] Yu Xiang, Roozbeh Mottaghi, and Silvio Savarese. Beyond Pascal: A benchmark for 3d object detection in the wild. In IEEE winter conference on applications of computer vision, pages 75-82. IEEE, 2014. +[47] Kelvin Xu, Jimmy Ba, Ryan Kiros, Kyunghyun Cho, Aaron Courville, Ruslan Salakhudinov, Rich Zemel, and Yoshua Bengio. Show, attend and tell: Neural image caption generation with visual attention. In International conference on machine learning, pages 2048-2057. PMLR, 2015. +[48] Lin Xu, Yilin Zhao, Daquan Zhou, Zhijie Lin, See Kiong Ng, and Jiashi Feng. Pllava: Parameter-free llava extension from images to videos for video dense captioning. arXiv preprint arXiv:2404.16994, 2024. +[49] Kexin Yi, Jiajun Wu, Chuang Gan, Antonio Torralba, Pushmeet Kohli, and Josh Tenenbaum. Neural-symbolic vqa: Disentangling reasoning from vision and language understanding. Advances in neural information processing systems, 31, 2018. +[50] Kexin Yi, Chuang Gan, Yunzhu Li, Pushmeet Kohli, Jiajun Wu, Antonio Torralba, and Joshua B Tenenbaum. Clevrer: Collision events for video representation and reasoning. arXiv preprint arXiv:1910.01442, 2019. +[51] Xiang Yue, Yuansheng Ni, Kai Zhang, Tianyu Zheng, Ruoqi Liu, Ge Zhang, Samuel Stevens, Dongfu Jiang, Weiming Ren, Yuxuan Sun, et al. Mmmu: A massive multi-discipline multimodal understanding and reasoning benchmark for expert agi. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9556-9567, 2024. + +[52] Alan Yuille and Daniel Kersten. Vision as bayesian inference: analysis by synthesis? Trends in cognitive sciences, 10(7):301-308, 2006. +[53] Jiyao Zhang, Weiyao Huang, Bo Peng, Mingdong Wu, Fei Hu, Zijian Chen, Bo Zhao, and Hao Dong. Omni6dpose: A benchmark and model for universal 6d object pose estimation and tracking. In European Conference on Computer Vision, pages 199-216. Springer, 2024. \ No newline at end of file diff --git a/2025/Spatial457_ A Diagnostic Benchmark for 6D Spatial Reasoning of Large Mutimodal Models/images.zip b/2025/Spatial457_ A Diagnostic Benchmark for 6D Spatial Reasoning of Large Mutimodal Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..1dd5988781adc5f84623b500d6d7279774dff72d --- /dev/null +++ b/2025/Spatial457_ A Diagnostic Benchmark for 6D Spatial Reasoning of Large Mutimodal Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9e06818d417d0f88380db3ad1fac21bcc4ee7d1eca6b1959f348dabac979f2b9 +size 525679 diff --git a/2025/Spatial457_ A Diagnostic Benchmark for 6D Spatial Reasoning of Large Mutimodal Models/layout.json b/2025/Spatial457_ A Diagnostic Benchmark for 6D Spatial Reasoning of Large Mutimodal Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..2361ba5188841bd4d7de17d62244cfd3996f9368 --- /dev/null +++ b/2025/Spatial457_ A Diagnostic Benchmark for 6D Spatial Reasoning of Large Mutimodal Models/layout.json @@ -0,0 +1,8286 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 111, + 103, + 501, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 103, + 501, + 140 + ], + "spans": [ + { + "bbox": [ + 111, + 103, + 501, + 140 + ], + "type": "text", + "content": "Spatial457: A Diagnostic Benchmark for 6D Spatial Reasoning of Large Multimodal Models" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 76, + 160, + 538, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 160, + 538, + 176 + ], + "spans": [ + { + "bbox": [ + 76, + 160, + 538, + 176 + ], + "type": "text", + "content": "Xingrui Wang" + }, + { + "bbox": [ + 76, + 160, + 538, + 176 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 76, + 160, + 538, + 176 + ], + "type": "text", + "content": ", Wufei Ma" + }, + { + "bbox": [ + 76, + 160, + 538, + 176 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 76, + 160, + 538, + 176 + ], + "type": "text", + "content": ", Tiezheng Zhang" + }, + { + "bbox": [ + 76, + 160, + 538, + 176 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 76, + 160, + 538, + 176 + ], + "type": "text", + "content": ", Celso M de Melo" + }, + { + "bbox": [ + 76, + 160, + 538, + 176 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 76, + 160, + 538, + 176 + ], + "type": "text", + "content": ", Jieneng Chen" + }, + { + "bbox": [ + 76, + 160, + 538, + 176 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 76, + 160, + 538, + 176 + ], + "type": "text", + "content": ", Alan Yuille" + }, + { + "bbox": [ + 76, + 160, + 538, + 176 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 177, + 177, + 431, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 177, + 177, + 431, + 190 + ], + "spans": [ + { + "bbox": [ + 177, + 177, + 431, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 177, + 177, + 431, + 190 + ], + "type": "text", + "content": "Johns Hopkins University " + }, + { + "bbox": [ + 177, + 177, + 431, + 190 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 177, + 177, + 431, + 190 + ], + "type": "text", + "content": "DEVCOM Army Research Laboratory" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 217, + 200, + 230 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 217, + 200, + 230 + ], + "spans": [ + { + "bbox": [ + 151, + 217, + 200, + 230 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 243, + 296, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 243, + 296, + 542 + ], + "spans": [ + { + "bbox": [ + 56, + 243, + 296, + 542 + ], + "type": "text", + "content": "Although large multimodal models (LMMs) have demonstrated remarkable capabilities in visual scene interpretation and reasoning, their capacity for complex and precise 3-dimensional spatial reasoning remains uncertain. Existing benchmarks focus predominantly on 2D spatial understanding and lack a framework to comprehensively evaluate 6D spatial reasoning across varying complexities. To address this limitation, we present Spatial457, a scalable and unbiased synthetic dataset designed with 4 key capability for spatial reasoning: multi-object recognition, 2D location, 3D location, and 3D orientation. We develop a cascading evaluation structure, constructing 7 question types across 5 difficulty levels that range from basic single object recognition to our new proposed complex 6D spatial reasoning tasks. We evaluated various large multimodal models (LMMs) on Spatial457, observing a general decline in performance as task complexity increases, particularly in 3D reasoning and 6D spatial tasks. To quantify these challenges, we introduce the Relative Performance Dropping Rate (RPDR), highlighting key weaknesses in 3D reasoning capabilities. Leveraging the unbiased attribute design of our dataset, we also uncover prediction biases across different attributes, with similar patterns observed in real-world image settings. The code is released in https://github.com/XingruiWang/Spatial457." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 573, + 135, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 573, + 135, + 586 + ], + "spans": [ + { + "bbox": [ + 56, + 573, + 135, + 586 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 588, + 295, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 588, + 295, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 588, + 295, + 685 + ], + "type": "text", + "content": "In recent years, large multimodal models (LMMs) have demonstrated remarkable reasoning capabilities, enabling them to interpret visual scenes from images and articulate their understanding through natural language [1, 3, 11, 26, 27, 38, 39, 48]. These advancements have been evaluated using a variety of benchmarks that primarily emphasize semantic features—such as object categories and appearances, spatial relationships within the 2D image plane [15," + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 218, + 555, + 290 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 218, + 555, + 290 + ], + "spans": [ + { + "bbox": [ + 313, + 218, + 555, + 290 + ], + "type": "text", + "content": "17, 18] (e.g. left-right positioning from the camera view), and limited 2.5D features such as camera depth [40]. Tasks like object detection [13, 22], image captioning [21, 42, 47] and visual question answering [18, 50, 51] have significantly benefited from these evaluations, driving LMMs to new heights in visual understanding." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 291, + 556, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 291, + 556, + 388 + ], + "spans": [ + { + "bbox": [ + 313, + 291, + 556, + 388 + ], + "type": "text", + "content": "However, a critical gap remains in assessing the ability of LMMs to comprehend and reason about the full six degrees of freedom (6D) of objects in spatial environments, which includes both 3D positions and orientations [43, 46, 53]. Despite its significance in many real-world applications, such as robotics [2], autonomous navigation [14], and augmented reality [5], there are no available benchmarks for LMMs that evaluate 6D spatial reasoning capabilities." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 388, + 556, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 388, + 556, + 544 + ], + "spans": [ + { + "bbox": [ + 313, + 388, + 556, + 544 + ], + "type": "text", + "content": "Building a scalable diagnostic dataset for 6D vision tasks poses considerable challenges due to the extensive annotations required across vast image collections. One strategy is to develop 6D vision question-answering benchmarks on top of existing image datasets [4, 6, 7, 36]. However, real-world image collections often exhibit significant biases in 3D locations and poses, influenced by camera perspective limitations and the prevalence of common object orientations in everyday scenes. For instance, statistics from the SUN-RGBD [36] and nuScenes [7] datasets reveal that over " + }, + { + "bbox": [ + 313, + 388, + 556, + 544 + ], + "type": "inline_equation", + "content": "70\\%" + }, + { + "bbox": [ + 313, + 388, + 556, + 544 + ], + "type": "text", + "content": " of objects cluster into a single predominant direction (see Fig. 2), limiting the diversity required for comprehensive 6D reasoning evaluation." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 545, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 556, + 715 + ], + "type": "text", + "content": "To overcome these challenges, we introduce an unbiased and scalable synthetic dataset rendered in a realistic style, specifically designed to diagnose the 6D spatial reasoning abilities of LMMs. To comprehensively evaluate diverse aspects of 6D spatial reasoning, we focus on four core capabilities: multi-object recognition, 2D locations, 3D locations, and 3D orientation. Multi-object recognition serves as the foundation of scene understanding, while 2D location represents the simplest way to depict the spatial relationship between objects within the image plane [18, 19]. The 3D location factor extends understanding into three-dimensional space, capturing the depth and distance of objects, which is essential for tasks such as reasoning about occlusions [9, 10, 44]. The 3D orientation introduces object" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 693, + 245, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 693, + 245, + 703 + ], + "spans": [ + { + "bbox": [ + 66, + 693, + 245, + 703 + ], + "type": "text", + "content": "1Approved for public release: distribution is unlimited." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 71, + 704, + 130, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 704, + 130, + 713 + ], + "spans": [ + { + "bbox": [ + 71, + 704, + 130, + 713 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 71, + 704, + 130, + 713 + ], + "type": "text", + "content": " : Equal advising." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "24669" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 88, + 73, + 200, + 156 + ], + "blocks": [ + { + "bbox": [ + 88, + 73, + 200, + 156 + ], + "lines": [ + { + "bbox": [ + 88, + 73, + 200, + 156 + ], + "spans": [ + { + "bbox": [ + 88, + 73, + 200, + 156 + ], + "type": "image", + "image_path": "1864bdae54cca2dc2242be14e7524808da21299f243132d64d8d7e4e7da4802c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 100, + 161, + 114, + 175 + ], + "blocks": [ + { + "bbox": [ + 100, + 161, + 114, + 175 + ], + "lines": [ + { + "bbox": [ + 100, + 161, + 114, + 175 + ], + "spans": [ + { + "bbox": [ + 100, + 161, + 114, + 175 + ], + "type": "image", + "image_path": "30d3a8c4f56983c66f5072c4add864c722b739fc85750ef883d5c3ec0535e6fd.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 118, + 163, + 178, + 173 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 163, + 178, + 173 + ], + "spans": [ + { + "bbox": [ + 118, + 163, + 178, + 173 + ], + "type": "text", + "content": "3D Orientations" + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 99, + 186, + 115, + 201 + ], + "blocks": [ + { + "bbox": [ + 99, + 186, + 115, + 201 + ], + "lines": [ + { + "bbox": [ + 99, + 186, + 115, + 201 + ], + "spans": [ + { + "bbox": [ + 99, + 186, + 115, + 201 + ], + "type": "image", + "image_path": "73f4fbe35d85cfb1d81613d5e40465c3a10eb66eb51d47017fe9331b1e3a2e2d.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 118, + 190, + 167, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 190, + 167, + 198 + ], + "spans": [ + { + "bbox": [ + 118, + 190, + 167, + 198 + ], + "type": "text", + "content": "3D Locations" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 101, + 214, + 113, + 226 + ], + "blocks": [ + { + "bbox": [ + 101, + 214, + 113, + 226 + ], + "lines": [ + { + "bbox": [ + 101, + 214, + 113, + 226 + ], + "spans": [ + { + "bbox": [ + 101, + 214, + 113, + 226 + ], + "type": "image", + "image_path": "5ca5917c849a9c43b71b82d8ae007c43a6f4ef55085d41bd963f6fcadca79188.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 118, + 216, + 167, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 216, + 167, + 224 + ], + "spans": [ + { + "bbox": [ + 118, + 216, + 167, + 224 + ], + "type": "text", + "content": "2D Locations" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 100, + 241, + 115, + 251 + ], + "blocks": [ + { + "bbox": [ + 100, + 241, + 115, + 251 + ], + "lines": [ + { + "bbox": [ + 100, + 241, + 115, + 251 + ], + "spans": [ + { + "bbox": [ + 100, + 241, + 115, + 251 + ], + "type": "image", + "image_path": "7284700c93f5092b896d396ebb2139ffbb6a005fe1ebc54d1d6b3cadbb359ff1.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 119, + 242, + 172, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 119, + 242, + 172, + 251 + ], + "spans": [ + { + "bbox": [ + 119, + 242, + 172, + 251 + ], + "type": "text", + "content": "Multiple Objects" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 215, + 73, + 408, + 281 + ], + "blocks": [ + { + "bbox": [ + 89, + 262, + 208, + 283 + ], + "lines": [ + { + "bbox": [ + 89, + 262, + 208, + 283 + ], + "spans": [ + { + "bbox": [ + 89, + 262, + 208, + 283 + ], + "type": "text", + "content": "4 Capabilities of Comprehensive Spatial Reasoning" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 215, + 73, + 408, + 281 + ], + "lines": [ + { + "bbox": [ + 215, + 73, + 408, + 281 + ], + "spans": [ + { + "bbox": [ + 215, + 73, + 408, + 281 + ], + "type": "image", + "image_path": "e3c9dcd25aabff842a0391022215b0ee9b9a8075ef389524db34b89ec26de262.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 413, + 73, + 524, + 281 + ], + "blocks": [ + { + "bbox": [ + 413, + 73, + 524, + 281 + ], + "lines": [ + { + "bbox": [ + 413, + 73, + 524, + 281 + ], + "spans": [ + { + "bbox": [ + 413, + 73, + 524, + 281 + ], + "type": "image", + "image_path": "4bf3ba052d061251eeae6726d273807e7e69ab2f73ab853dfff8c86a9093edcf.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 293, + 555, + 348 + ], + "lines": [ + { + "bbox": [ + 54, + 293, + 555, + 348 + ], + "spans": [ + { + "bbox": [ + 54, + 293, + 555, + 348 + ], + "type": "text", + "content": "Figure 1. Overview of the Spatial457 benchmark. We define four core capabilities for spatial reasoning. By incorporating these capabilities step-by-step, the benchmark assesses models across five progressive difficulty levels and seven question types, ranging from single-object recognition to advanced 6D spatial relationships and collision prediction. Our evaluation demonstrates a significant performance drop for more advanced questions and highlights the gap between state-of-the-art (SoTA) models and human performance in complex spatial reasoning VQA tasks." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 77, + 367, + 139, + 432 + ], + "blocks": [ + { + "bbox": [ + 77, + 367, + 139, + 432 + ], + "lines": [ + { + "bbox": [ + 77, + 367, + 139, + 432 + ], + "spans": [ + { + "bbox": [ + 77, + 367, + 139, + 432 + ], + "type": "image", + "image_path": "991ef26e1fdb5ca8588e848748afa12a64c729a885d18c8c85f9749b0aea54a3.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 89, + 434, + 135, + 442 + ], + "lines": [ + { + "bbox": [ + 89, + 434, + 135, + 442 + ], + "spans": [ + { + "bbox": [ + 89, + 434, + 135, + 442 + ], + "type": "text", + "content": "(a) SUN-RGBD" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 141, + 367, + 206, + 432 + ], + "blocks": [ + { + "bbox": [ + 141, + 367, + 206, + 432 + ], + "lines": [ + { + "bbox": [ + 141, + 367, + 206, + 432 + ], + "spans": [ + { + "bbox": [ + 141, + 367, + 206, + 432 + ], + "type": "image", + "image_path": "7bf783f866998671f7592a8496fdc91cc480f4b16aea96c8d3040a39fcaf0e6d.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 157, + 434, + 195, + 442 + ], + "lines": [ + { + "bbox": [ + 157, + 434, + 195, + 442 + ], + "spans": [ + { + "bbox": [ + 157, + 434, + 195, + 442 + ], + "type": "text", + "content": "(b) nuScenes" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 452, + 295, + 496 + ], + "lines": [ + { + "bbox": [ + 55, + 452, + 295, + 496 + ], + "spans": [ + { + "bbox": [ + 55, + 452, + 295, + 496 + ], + "type": "text", + "content": "Figure 2. The distribution of 3D poses in representative 6D datasets (a&b) highlights an imbalance in object orientations. This inspired us to develop a balanced benchmark encompassing all degrees of 3D orientation." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 211, + 367, + 274, + 432 + ], + "blocks": [ + { + "bbox": [ + 211, + 367, + 274, + 432 + ], + "lines": [ + { + "bbox": [ + 211, + 367, + 274, + 432 + ], + "spans": [ + { + "bbox": [ + 211, + 367, + 274, + 432 + ], + "type": "image", + "image_path": "2ad47a1c7db930e43d981363ad687305e879358386caf9c6f1703cc066b49da2.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 228, + 434, + 253, + 442 + ], + "lines": [ + { + "bbox": [ + 228, + 434, + 253, + 442 + ], + "spans": [ + { + "bbox": [ + 228, + 434, + 253, + 442 + ], + "type": "text", + "content": "(c) Ours" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "bbox": [ + 55, + 510, + 295, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 295, + 533 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 295, + 533 + ], + "type": "text", + "content": "rotation [44, 46], enabling models to reason about precise 3D poses of objects within a scene." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 55, + 533, + 295, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 533, + 295, + 617 + ], + "spans": [ + { + "bbox": [ + 55, + 533, + 295, + 617 + ], + "type": "text", + "content": "By progressively incorporating the four core capabilities, we establish five levels of difficulty, resulting in seven distinct question types. These questions span a spectrum of spatial reasoning complexity, ranging from basic semantic identification and 2D spatial relationships to advanced 3D positional understanding and orientation-based queries, culminating in comprehensive 6D spatial reasoning tasks." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 55, + 618, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 295, + 713 + ], + "type": "text", + "content": "Our evaluation measures performance across varied question types and examines how it is affected by different combinations of the four capabilities. This tiered design provides comprehensive benchmarks for spatial reasoning capabilities, highlighting each model's strengths and areas of vulnerability. To the best of our knowledge, this work presents the first comprehensive benchmark designed to evaluate the 6D spatial reasoning capabilities of vision-" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 369, + 387, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 369, + 387, + 380 + ], + "spans": [ + { + "bbox": [ + 313, + 369, + 387, + 380 + ], + "type": "text", + "content": "language models." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 384, + 555, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 384, + 555, + 479 + ], + "spans": [ + { + "bbox": [ + 313, + 384, + 555, + 479 + ], + "type": "text", + "content": "Additionally, our dataset's unbiased attribute design allows us to analyze potential biases in model predictions. By systematically exploring model performance under different attributes, we identify tendencies or limitations that may affect real-world applications. Altogether, our benchmark not only provides a foundational tool for evaluating LMMs but also serves as a guide for developing future benchmarks that support models with advanced 6D spatial reasoning." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 325, + 482, + 478, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 482, + 478, + 493 + ], + "spans": [ + { + "bbox": [ + 325, + 482, + 478, + 493 + ], + "type": "text", + "content": "Our main contributions are threefold:" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 314, + 496, + 555, + 711 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 315, + 496, + 555, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 496, + 555, + 555 + ], + "spans": [ + { + "bbox": [ + 315, + 496, + 555, + 555 + ], + "type": "text", + "content": "1. An unbiased and scalable synthetic dataset: We present a realistic, unbiased synthetic dataset designed to assess 6D spatial reasoning in LMMs. We define four core capabilities and propose five difficulty levels for spatial reasoning VQA with seven question types." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 314, + 556, + 554, + 627 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 556, + 554, + 627 + ], + "spans": [ + { + "bbox": [ + 314, + 556, + 554, + 627 + ], + "type": "text", + "content": "2. Comprehensive benchmarking of 6D capabilities: We benchmark various LMMs across all difficulty levels. The performance variations across settings reveal potential limitations or strengths, providing valuable insights into which components should be prioritized to improve the spatial reasoning of LMMs." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 314, + 628, + 554, + 711 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 628, + 554, + 711 + ], + "spans": [ + { + "bbox": [ + 314, + 628, + 554, + 711 + ], + "type": "text", + "content": "3. Bias analysis in model predictions: Leveraging the unbiased nature of our dataset, we conduct an in-depth bias analysis to explore how LMM predictions may be unbalanced across different attributes, especially 3D-related ones. This reveals potential weaknesses in model predictions, offering insights into limitations that may also appear in real-world scenarios." + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24670" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 143, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 143, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 143, + 83 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 203, + 103 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 203, + 103 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 203, + 103 + ], + "type": "text", + "content": "2.1. Synthetic Datasets in VQA" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 104, + 297, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 104, + 297, + 357 + ], + "spans": [ + { + "bbox": [ + 55, + 104, + 297, + 357 + ], + "type": "text", + "content": "The use of synthetic data in visual question answering (VQA) has gained traction as a means to circumvent the limitations and biases inherent in real-world datasets. Synthetic environments [16, 30, 34] allow for controlled manipulation of scene elements, enabling researchers to probe specific reasoning abilities in VQA models without the confounding factors often present in natural imagery. Previous works [19, 24, 25] have utilized synthetic environments to evaluate multimodal models across a range of visual reasoning tasks, typically focusing on 2D spatial relationships and object features. For instance, CLEVR [19], a synthetic dataset widely used in the VQA community, evaluates compositional reasoning by introducing artificially generated scenes with diverse object arrangements. Super-CLEVR-3D [45] represents a significant step forward in evaluating 3D locations, it lacks an assessment of 6D spatial relationships that consider both 3D location and 3D orientation. Our work extends this tradition by creating a synthetic dataset tailored for 6D spatial reasoning, filling a gap in current synthetic datasets that do not address the complete spatial understanding of objects in 3D space." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 360, + 168, + 371 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 360, + 168, + 371 + ], + "spans": [ + { + "bbox": [ + 55, + 360, + 168, + 371 + ], + "type": "text", + "content": "2.2. LMM Benchmarks" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 373, + 296, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 373, + 296, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 373, + 296, + 601 + ], + "type": "text", + "content": "Existing benchmarks [15, 18, 35, 51] for large multimodal models (LMMs) predominantly assess their performance on tasks involving 2D visual question answering and multimodal reasoning. For instance, datasets like VQAv2 [15] and GQA [18] focus on answering questions based on visual content. However, these benchmarks leave a gap in evaluating models' capabilities in spatial contexts. Recent benchmarks, such as CV-Bench [40], assess spatial understanding, including depth ordering for 2.5D evaluation. More recent work [9, 10, 32] introduces questions regarding either aspect of 3D locations or rotations. Our proposed benchmark goes beyond traditional 2D, 2.5D, and 3D location or rotation assessments by introducing tasks that challenge LMMs to understand 3D positions and orientations together. We also evaluate models in a cascading manner across five difficulty levels and summarize their performance in four reasoning abilities, along with their output biases, providing a deeper analysis of the weaknesses in spatial reasoning within multimodal AI." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 609, + 295, + 638 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 609, + 295, + 638 + ], + "spans": [ + { + "bbox": [ + 55, + 609, + 295, + 638 + ], + "type": "text", + "content": "3. Spatial457: Comprehensive Spatial Reasoning Benchmark" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 642, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 715 + ], + "type": "text", + "content": "To comprehensively evaluate the spatial reasoning capabilities of vision-language models, we introduce a synthetic dataset named Spatial457. We define four core capabilities for comprehensive spatial reasoning (Sec. 3.1) and generate scenes with full 6D annotations (Sec. 3.2). The full combination of these capabilities form the new 6D spatial" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 455, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 455, + 84 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 455, + 84 + ], + "type": "text", + "content": "relationship VQA tasks (Sec. 3.3)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 85, + 555, + 169 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 85, + 555, + 169 + ], + "spans": [ + { + "bbox": [ + 313, + 85, + 555, + 169 + ], + "type": "text", + "content": "To thoroughly assess model capabilities, we define five levels of difficulty in vision question answering (VQA) tasks, converting seven question subsets that range from single-object recognition to the newly proposed 6D spatial reasoning tasks (Sec. 3.4). This structured progression in question difficulty is achieved by sequentially introducing new spatial reasoning capabilities, as illustrated in Fig. 1." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 178, + 523, + 192 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 178, + 523, + 192 + ], + "spans": [ + { + "bbox": [ + 313, + 178, + 523, + 192 + ], + "type": "text", + "content": "3.1. Four Capabilities for Spatial Reasoning" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 196, + 555, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 196, + 555, + 232 + ], + "spans": [ + { + "bbox": [ + 313, + 196, + 555, + 232 + ], + "type": "text", + "content": "We define 4 capabilities for comprehensive spatial understanding in a 3D scene: multiple objects, 2D locations, 3D locations, and 3D orientations." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 233, + 555, + 714 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 313, + 233, + 554, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 233, + 554, + 353 + ], + "spans": [ + { + "bbox": [ + 313, + 233, + 554, + 353 + ], + "type": "text", + "content": "(1) Multiple objects. Objects are fundamental to defining scene structure [52]. Large vision-language models show strong results in aligning semantic features with text [23, 31], handling multiple objects from scene becomes the first toward scene understanding. Models must recognize individual attributes and interpret relationships among them. Our dataset evaluates multi-object reasoning, such as comparison and counting. For example, in Fig. 1, the image shows two cyan objects, requiring the model to verify color simultarity and count cyan objects accurately." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 354, + 554, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 354, + 554, + 496 + ], + "spans": [ + { + "bbox": [ + 313, + 354, + 554, + 496 + ], + "type": "text", + "content": "(2) 2D locations. 2D spatial relationships are essential in prior image-based VQA tasks [18, 19]. In traditional multimodal models [20, 28, 49], these relationships can be learned through object detection training [8, 33]. However, in large-scale vision-language models, explicit location information is often absent from the latent representations. Following Li et al. [24], we examine four spatial relationships (left, right, front, back) between objects from 2D camera views as the first taks of spatial reasoning. Questions use these relations to filter targets, such as \"What is the shape of the object to the left of the chopper?\" (see Fig. 1)." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 497, + 554, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 497, + 554, + 617 + ], + "spans": [ + { + "bbox": [ + 313, + 497, + 554, + 617 + ], + "type": "text", + "content": "(3) 3D locations. Beyond 2D relationships, 3D locations can incorporate depth, capturing spatial relationships that 2D alone cannot represent. This is crucial for tasks involving occlusions and complex spatial arrangements. Our dataset assesses the 3D location understanding abilities of model through questions about occlusions query as in [45]. For example, as shown in Fig. 1, questions may ask about an object's position relative to others in 3D space or whether one object is occluded, challenging models to judge distances and spatial positions accurately." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 618, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 618, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 618, + 555, + 714 + ], + "type": "text", + "content": "(4) 3D orientations. The 3D orientation of objects introduces an additional layer of complexity by incorporating rotational attributes. In current computer vision tasks, 3D orientation is often derived by pose estimation methods. However, large-scale vision-language models generally lack explicit 3D orientation information in their latent representations, potentially making it challenging to accurately infer the object's orientations. In our dataset, we eval" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24671" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "content": "uate models' abilities to understand 3D orientations by focusing on object poses, including their facing direction and alignments with others. For example, as depicted in Fig. 1, the model may be asked to identify whether a cyan object is facing left, oriented parallel to a brown object, or positioned in a specific spatial relationship — i.e. left, right, front, or behind— from the perspective of the objects themselves." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 163, + 157, + 175 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 163, + 157, + 175 + ], + "spans": [ + { + "bbox": [ + 55, + 163, + 157, + 175 + ], + "type": "text", + "content": "3.2. Scene Rendering" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 178, + 296, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 178, + 296, + 262 + ], + "spans": [ + { + "bbox": [ + 55, + 178, + 296, + 262 + ], + "type": "text", + "content": "We build upon the generation pipeline from prior VQA datasets [19, 24] and uniformly select 21 object models across 5 categories to construct the dataset as Li et al. [24]. To improve the realism of our rendered dataset, we use real image style environment maps and refined object textures. For benchmarking new 6D spatial understanding ability, we add new 6D spatial relationship annotations among objects." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 262, + 296, + 322 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 262, + 296, + 322 + ], + "spans": [ + { + "bbox": [ + 55, + 262, + 296, + 322 + ], + "type": "text", + "content": "Background. To create realistic scene backgrounds, we use HDRI images" + }, + { + "bbox": [ + 55, + 262, + 296, + 322 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 55, + 262, + 296, + 322 + ], + "type": "text", + "content": " with environment maps and natural lightings. We selected 224 different outdoor scenes with normal brightness levels without high light contrast by their metadata to maintain visual consistency across images." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 323, + 296, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 323, + 296, + 430 + ], + "spans": [ + { + "bbox": [ + 55, + 323, + 296, + 430 + ], + "type": "text", + "content": "Objects. We reconstruct 3D mesh models [24] with realistic textures, introducing more variation and realistic appearances for objects. Similar to Li et al. [24], each object in our dataset is defined by distinct attributes and precise spatial locations. The original attributes include shape, color, and size. For benchmarking 6D spatial reasoning, each object also includes a vector specifying its 2D location " + }, + { + "bbox": [ + 55, + 323, + 296, + 430 + ], + "type": "inline_equation", + "content": "(\\mathbf{x},\\mathbf{y})" + }, + { + "bbox": [ + 55, + 323, + 296, + 430 + ], + "type": "text", + "content": ", 3D world coordinates " + }, + { + "bbox": [ + 55, + 323, + 296, + 430 + ], + "type": "inline_equation", + "content": "(\\mathrm{X},\\mathrm{Y},\\mathrm{Z})" + }, + { + "bbox": [ + 55, + 323, + 296, + 430 + ], + "type": "text", + "content": ", and pose direction derived from its 3D orientation [45]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 432, + 296, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 432, + 296, + 598 + ], + "spans": [ + { + "bbox": [ + 55, + 432, + 296, + 598 + ], + "type": "text", + "content": "Relationships. As shown in Fig. 3 (a.1 to a.3), we define three types of spatial relationships: 1. 2D Spatial Relationships: These describe the relative positioning of objects within the 2D camera view, using terms such as left, right, front, and behind. 2. 6D Spatial Relationships: Extending spatial descriptions into 3D space, these relationships incorporate both the 3D position and orientation of objects to define relative positions from a target object's perspective, including left, right, front, and behind. 3. Collision Relationships: This relationship describes potential collisions if the target object moves forward or backward. Collision prediction involves assessing the orientation and 3D location of target objects relative to others to determine possible collisions." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 605, + 268, + 618 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 268, + 618 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 268, + 618 + ], + "type": "text", + "content": "3.3. New Questions for 6D Spatial Reasoning" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 620, + 296, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 620, + 296, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 620, + 296, + 693 + ], + "type": "text", + "content": "Based on the four capabilities defined in Sec. 3.1, we construct new question types for 6D spatial reasoning and collision prediction. These questions require an understanding of both 3D location and 3D orientation across multiple objects to generate accurate answers. This level of complexity represents the L5 difficulty level shown in Fig. 1." + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 332, + 70, + 397, + 120 + ], + "blocks": [ + { + "bbox": [ + 332, + 70, + 397, + 120 + ], + "lines": [ + { + "bbox": [ + 332, + 70, + 397, + 120 + ], + "spans": [ + { + "bbox": [ + 332, + 70, + 397, + 120 + ], + "type": "image", + "image_path": "f6fcdc111cff4c02fce2f5207fed5572b3d475b0be5ee653357e91a05618a402.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 334, + 122, + 394, + 154 + ], + "lines": [ + { + "bbox": [ + 334, + 122, + 394, + 154 + ], + "spans": [ + { + "bbox": [ + 334, + 122, + 394, + 154 + ], + "type": "text", + "content": "Q: How many yellow objects are to the left of the red wagon? A: 1" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 401, + 71, + 466, + 118 + ], + "blocks": [ + { + "bbox": [ + 401, + 71, + 466, + 118 + ], + "lines": [ + { + "bbox": [ + 401, + 71, + 466, + 118 + ], + "spans": [ + { + "bbox": [ + 401, + 71, + 466, + 118 + ], + "type": "image", + "image_path": "0a2a36af4790072fcdea3e45232b9148a8f57a15edfce4180dd57bbaa769eeed.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 403, + 119, + 463, + 156 + ], + "lines": [ + { + "bbox": [ + 403, + 119, + 463, + 156 + ], + "spans": [ + { + "bbox": [ + 403, + 119, + 463, + 156 + ], + "type": "text", + "content": "Q: How many yellow objects are on the left hand side of the red wagon? A: 3" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 469, + 71, + 534, + 118 + ], + "blocks": [ + { + "bbox": [ + 469, + 71, + 534, + 118 + ], + "lines": [ + { + "bbox": [ + 469, + 71, + 534, + 118 + ], + "spans": [ + { + "bbox": [ + 469, + 71, + 534, + 118 + ], + "type": "image", + "image_path": "1871951b8fe3cb2330b8bf2b78e45e41803a43513be1565e8d953b2dead326b3.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 472, + 119, + 533, + 149 + ], + "lines": [ + { + "bbox": [ + 472, + 119, + 533, + 149 + ], + "spans": [ + { + "bbox": [ + 472, + 119, + 533, + 149 + ], + "type": "text", + "content": "Q: Is the red wagon will collide with a yellow regular bus if it moving forward?" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 333, + 172, + 434, + 226 + ], + "blocks": [ + { + "bbox": [ + 342, + 158, + 392, + 168 + ], + "lines": [ + { + "bbox": [ + 342, + 158, + 392, + 168 + ], + "spans": [ + { + "bbox": [ + 342, + 158, + 392, + 168 + ], + "type": "text", + "content": "(a.1) 2D spatial" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 333, + 172, + 434, + 226 + ], + "lines": [ + { + "bbox": [ + 333, + 172, + 434, + 226 + ], + "spans": [ + { + "bbox": [ + 333, + 172, + 434, + 226 + ], + "type": "image", + "image_path": "e5009a32a1bb030559d3c89e9a090fefbcb9f87618f8510654f493c0d8f6fc9d.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 354, + 237, + 411, + 255 + ], + "lines": [ + { + "bbox": [ + 354, + 237, + 411, + 255 + ], + "spans": [ + { + "bbox": [ + 354, + 237, + 411, + 255 + ], + "type": "text", + "content": "(b.1) Programs of 6D spatial question" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 440, + 172, + 532, + 236 + ], + "blocks": [ + { + "bbox": [ + 410, + 158, + 461, + 167 + ], + "lines": [ + { + "bbox": [ + 410, + 158, + 461, + 167 + ], + "spans": [ + { + "bbox": [ + 410, + 158, + 461, + 167 + ], + "type": "text", + "content": "(a.2) 6D spatial" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 473, + 150, + 490, + 156 + ], + "lines": [ + { + "bbox": [ + 473, + 150, + 490, + 156 + ], + "spans": [ + { + "bbox": [ + 473, + 150, + 490, + 156 + ], + "type": "text", + "content": "A: No" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 476, + 158, + 523, + 167 + ], + "lines": [ + { + "bbox": [ + 476, + 158, + 523, + 167 + ], + "spans": [ + { + "bbox": [ + 476, + 158, + 523, + 167 + ], + "type": "text", + "content": "(a.3) Collision" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 440, + 172, + 532, + 236 + ], + "lines": [ + { + "bbox": [ + 440, + 172, + 532, + 236 + ], + "spans": [ + { + "bbox": [ + 440, + 172, + 532, + 236 + ], + "type": "image", + "image_path": "53f9d9d3ce6033dd56704d041a039483042b573f2b5f7e60b536bb6ba17b6dd0.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 443, + 238, + 521, + 255 + ], + "lines": [ + { + "bbox": [ + 443, + 238, + 521, + 255 + ], + "spans": [ + { + "bbox": [ + 443, + 238, + 521, + 255 + ], + "type": "text", + "content": "(b.2) Programs of collision prediction question" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 262, + 554, + 297 + ], + "lines": [ + { + "bbox": [ + 313, + 262, + 554, + 297 + ], + "spans": [ + { + "bbox": [ + 313, + 262, + 554, + 297 + ], + "type": "text", + "content": "Figure 3. (a.1-3) Example of the 2D spatial, 3D spatial and collision with images and questions. (b.1-2) shows the new operation programs we generated for the new 6D spatial reasoning questions." + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 306, + 555, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 306, + 555, + 510 + ], + "spans": [ + { + "bbox": [ + 313, + 306, + 555, + 510 + ], + "type": "text", + "content": "The 6D spatial relationship questions challenge the model to understand the spatial relationship involving both 3D position and orientation from a target object's perspective (shown in Fig. 3 (a.2)). Compared with prior VQA benchmarks focusing on 2D [19, 24], we introduce new 3D spatial reasoning operations, 3D Relate(" + }, + { + "bbox": [ + 313, + 306, + 555, + 510 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 313, + 306, + 555, + 510 + ], + "type": "text", + "content": "), to determine relative positions like left, right, front, and behind within a 3D coordinate space (see Fig. 3 (b.1)). For example, a question might ask, \"How many objects are to the right of the red wagon?\" requiring the model to filter and count objects based on their orientation and relative 3D position, which would yield different answers compared to similar 2D left and right questions (Fig. 3 (a.1)). This task is particularly challenging as it requires models to integrate 3D spatial relationships while managing occlusions and depth perception—capabilities not typically covered in simpler 2D spatial tasks." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 510, + 556, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 556, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 556, + 665 + ], + "type": "text", + "content": "The collision prediction questions, as shown in Fig. 3 (a.3), extend 6D spatial reasoning by introducing motion-based inference. These questions assess the model's ability to anticipate future states and potential interactions (collisions) based on accurate estimation of 3D location and direction, denoted as forward or backward. For instance, a question might ask, \"Will the red wagon collide with a yellow bus if it moves forward?\" This requires the model to consider the orientation and direction of movement relative to surrounding objects to predict possible outcomes. Such tasks test the model's advanced spatial reasoning about 3D orientation and position, and its capacity to predict accurate trajectories and directions within the future scene." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 672, + 533, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 672, + 533, + 685 + ], + "spans": [ + { + "bbox": [ + 313, + 672, + 533, + 685 + ], + "type": "text", + "content": "3.4. Questions under Other 4 Difficulty Levels" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": "The L5 questions above represent the most challenging cases for 6D spatial reasoning, testing all four capabilities" + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 66, + 702, + 203, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 702, + 203, + 712 + ], + "spans": [ + { + "bbox": [ + 66, + 702, + 203, + 712 + ], + "type": "text", + "content": "2https://polyhaven.com/hdris" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "24672" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "type": "text", + "content": "in Sec. 3.1. However, the purpose of this benchmark is to comprehensively diagnose each factor in spatial reasoning. Thus, we design a structured roadmap to evaluate the model's performance across different levels of spatial reasoning, assessing its capabilities in each specific factor." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 134, + 295, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 134, + 295, + 289 + ], + "spans": [ + { + "bbox": [ + 55, + 134, + 295, + 289 + ], + "type": "text", + "content": "By composing the four factors of comprehensive spatial reasoning, we define an additional four levels of question difficulty based on the generated scenes, encompassing a total of seven tasks in our benchmark. As the complexity of spatial reasoning increases with the introduction of additional factors, we organize these questions into a structured roadmap, as illustrated in Fig. 1. Although many questions are reimplementations from previous benchmarks [24, 45] on our new images, we aim to integrate these into a unified benchmark to provide deeper insights into the model's spatial reasoning capabilities. The composition of the four capabilities at each level, along with comparisons to previous VQA benchmarks, is shown in Tab. 1." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 290, + 295, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 290, + 295, + 397 + ], + "spans": [ + { + "bbox": [ + 55, + 290, + 295, + 397 + ], + "type": "text", + "content": "Level 1 - single object questions [19, 24]. We begin with the questions with single objects. These questions are generated using a filtered set of templates from Li et al. [24]. We select five templates involving only one object in the question, such as \"What color is the double bus?\" in Fig. 1. This question type provides a base score for the tested vision-language model. The comparison with L1 helps to determine whether the decrease of performance are due to challenge of spatial reasoning or object recognition." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 399, + 295, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 399, + 295, + 507 + ], + "spans": [ + { + "bbox": [ + 55, + 399, + 295, + 507 + ], + "type": "text", + "content": "Level 2 - multiple objects questions [19, 24]. Level 2 questions extend Level 1 by involving multiple objects. These questions involve understanding the relationship between objects and counting. For example, \"Is there another object of the same color as the double bus?\" as shown in Fig. 1, where the question templates follow Li et al. [24]. We evaluate the model's capacity for reasoning in scenes with multiple entities as a critical step toward understanding spatial relationships in next levels." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 509, + 295, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 509, + 295, + 592 + ], + "spans": [ + { + "bbox": [ + 55, + 509, + 295, + 592 + ], + "type": "text", + "content": "Level 3 - 2D spatial questions [19, 24]. Level 3 questions further introduce 2D spatial relationships among multiple objects, requiring the model to understand the position of objects and their relationships from the 2D camera's perspective. For example, \"What shape is the object to the left of the chopper?\" as depicted in Fig. 1. The 2D relationship is the first step of spatial reasoning." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "content": "Level 4 - 3D pose questions [45]. Level 4 questions extend 3D orientation reasoning beyond the 2D spatial relationships in Level 3. Beyond appearance attributes (shape, color, size) or 2D locations, these questions require understanding 3D orientation, including the rotation angle or facing direction of an object from the camera's perspective. For example, \"What shape is the object that is parallel to the brown object?\" as shown in Fig. 1. 3D pose questions test the model's ability to interpret scenes with 3D awareness, involving rotation and alignment with other objects." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 553, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 228 + ], + "type": "text", + "content": "Level 4 - occlusion questions [45]. In addition to 3D orientation, incorporating 3D location presents another challenge beyond the Level 3 questions on 2D spatial relationships. A unique aspect of 3D spatial reasoning from images is occlusion. The occlusion questions in Level 4 require the model to comprehend depth perception and the relative positions of objects in 3D space, determining which object is closer to the camera within overlapping areas. For example, an occlusion question might ask, \"Is the red wagon occluded by the yellow bus?\" as shown in Fig. 1. This level evaluates the model's ability to reason about occlusions and depth perception, which are essential for understanding complex spatial relationships in 3D scenes." + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 317, + 228, + 553, + 345 + ], + "blocks": [ + { + "bbox": [ + 317, + 228, + 553, + 345 + ], + "lines": [ + { + "bbox": [ + 317, + 228, + 553, + 345 + ], + "spans": [ + { + "bbox": [ + 317, + 228, + 553, + 345 + ], + "type": "table", + "html": "
Question LevelMulti-object2D location3D location3D OrientationVQA Benchmarks
L1spatial-agnostic bench (e.g., TextVQA [35])
L2✓CLEVR-Math [25], CLEVR-Count [19]
L3✓✓VQA [15], GQA [18], CLEVR [19], Super-CLEVR [24]
L4 (a)✓✓✓CV-Bench (DepthOrder) [40], Super-CLEVR-3D [45]
L4 (b)✓✓✓Super-CLEVR-3D [45]
L5✓✓✓✓Spatial457
", + "image_path": "36adcd6fd157b7710556b58d3f68b1d6483d8b43e2f9a48b9b84a4d0efd8aa22.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 354, + 553, + 398 + ], + "lines": [ + { + "bbox": [ + 313, + 354, + 553, + 398 + ], + "spans": [ + { + "bbox": [ + 313, + 354, + 553, + 398 + ], + "type": "text", + "content": "Table 1. Comparison of VQA Benchmarks by Factors and Question Levels. Compared with existing VQA benchmarks, Spatial457 includes all four spatial reasoning factors, providing a comprehensive evaluation of 6D spatial understanding." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 314, + 402, + 434, + 414 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 402, + 434, + 414 + ], + "spans": [ + { + "bbox": [ + 314, + 402, + 434, + 414 + ], + "type": "text", + "content": "3.5. Evaluation Protocols" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 415, + 553, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 415, + 553, + 503 + ], + "spans": [ + { + "bbox": [ + 313, + 415, + 553, + 503 + ], + "type": "text", + "content": "To better evaluate the model's spatial reasoning capabilities, we propose a structured evaluation protocol that measures the relative capability of models on each factor for spatial reasoning. We define this evaluation metric as the Relative Performance Dropping Rate (RPDR), which assesses the decline in performance after introducing a new factor in the questions. Based on the roadmap in Fig. 1, the RPDR for each factor is defined as follows:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 322, + 507, + 468, + 519 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 507, + 468, + 519 + ], + "spans": [ + { + "bbox": [ + 322, + 507, + 468, + 519 + ], + "type": "interline_equation", + "content": "\\mathrm {R P D R} _ {\\text {M u l t i - O b j .}} = A c c _ {\\text {L e v e l 1}} \\Rightarrow A c c _ {\\text {L e v e l 2}}", + "image_path": "2fd87a6c99e50074c9654a9df30ee328f0052f3bdb442f3ead45cde8553c29a2.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 329, + 521, + 468, + 532 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 329, + 521, + 468, + 532 + ], + "spans": [ + { + "bbox": [ + 329, + 521, + 468, + 532 + ], + "type": "interline_equation", + "content": "\\mathrm {R P D R} _ {\\mathrm {2 D L o c .}} = A c c _ {\\text {L e v e l 2}} \\Rightarrow A c c _ {\\text {L e v e l 3}}", + "image_path": "72a34c1d133f8325255778c01942ef99fe89a5ff2ab51e081f1401832417d352.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 328, + 536, + 544, + 559 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 536, + 544, + 559 + ], + "spans": [ + { + "bbox": [ + 328, + 536, + 544, + 559 + ], + "type": "interline_equation", + "content": "\\mathrm {R P D R} _ {\\mathrm {3 D L o c .}} = \\operatorname {A v g} \\left( \\begin{array}{c} A c c _ {\\text {L e v e l 3}} \\Rightarrow A c c _ {\\text {L e v e l 4 (O c c .)}}, \\\\ A c c _ {\\text {L e v e l 4 (3 D P o s e)}} \\Rightarrow A c c _ {\\text {L e v e l 5 (6 D)}} \\end{array} \\right)", + "image_path": "bb5f9176433bbb8971fd7e2a05d1b8afc753456c955ccc0de87d0a1d6ed39480.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 329, + 561, + 538, + 586 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 329, + 561, + 538, + 586 + ], + "spans": [ + { + "bbox": [ + 329, + 561, + 538, + 586 + ], + "type": "interline_equation", + "content": "\\mathrm {R P D R} _ {\\mathrm {3 D O r i .}} = \\operatorname {A v g} \\left( \\begin{array}{c} A c c _ {\\text {L e v e l 3}} \\Rightarrow A c c _ {\\text {L e v e l 4 (3 D P o s e)}}, \\\\ A c c _ {\\text {L e v e l 4 (O c c .)}} \\Rightarrow A c c _ {\\text {L e v e l 5 (C o l .)}} \\end{array} \\right)", + "image_path": "b3490d1149b6725167ff44be7a5f9e2d9b6f825d63c114f01ea5f5a3edfe45ec.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 591, + 553, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 591, + 553, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 591, + 553, + 640 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 591, + 553, + 640 + ], + "type": "inline_equation", + "content": "Acc_{\\mathrm{Level}}" + }, + { + "bbox": [ + 313, + 591, + 553, + 640 + ], + "type": "text", + "content": " denotes the model's accuracy on the corresponding questions and " + }, + { + "bbox": [ + 313, + 591, + 553, + 640 + ], + "type": "inline_equation", + "content": "(Acc_{\\mathrm{Q1}} \\Rightarrow Acc_{\\mathrm{Q2}}) = \\frac{Acc_{\\mathrm{Q2}} - Acc_{\\mathrm{Q1}}}{Acc_{\\mathrm{Q1}}}" + }, + { + "bbox": [ + 313, + 591, + 553, + 640 + ], + "type": "text", + "content": " denotes the relative performance dropping rate from " + }, + { + "bbox": [ + 313, + 591, + 553, + 640 + ], + "type": "inline_equation", + "content": "Q1" + }, + { + "bbox": [ + 313, + 591, + 553, + 640 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 591, + 553, + 640 + ], + "type": "inline_equation", + "content": "Q2" + }, + { + "bbox": [ + 313, + 591, + 553, + 640 + ], + "type": "text", + "content": ". The results of RPDR will be discussed in Sec. 4.3." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 647, + 395, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 647, + 395, + 659 + ], + "spans": [ + { + "bbox": [ + 313, + 647, + 395, + 659 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 663, + 430, + 676 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 663, + 430, + 676 + ], + "spans": [ + { + "bbox": [ + 313, + 663, + 430, + 676 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 677, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 553, + 714 + ], + "type": "text", + "content": "We generate 1,000 images to test the performance of large vision-language models. For the seven types of questions, we generate 1000 single-object questions (L1-Single), 5,000" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24673" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 67, + 70, + 545, + 246 + ], + "blocks": [ + { + "bbox": [ + 67, + 70, + 545, + 246 + ], + "lines": [ + { + "bbox": [ + 67, + 70, + 545, + 246 + ], + "spans": [ + { + "bbox": [ + 67, + 70, + 545, + 246 + ], + "type": "table", + "html": "
Difficulty LevelLevel 1Level 2Level 3Level 4Level 5
Questions typeSingle ObjectMulti-Obj.2D SpatialOcclusion3D PoseCollisions6D Spatial
Random33.0532.7733.4722.0418.9921.0219.41
GPT-4o [29]74.4662.8856.1448.4042.4138.4137.01
GeminiPro 1.5 [37]73.2662.5454.4947.6543.6741.1939.36
Claude 3.5 Sonnet [3]68.2457.4054.1930.8438.4035.3433.48
Qwen2-VL-7B-Inst. [39]71.9661.4455.3427.8734.2936.5833.75
InternVL2-8B [38]58.1158.7657.4033.2532.3234.3034.30
LLaVA-v1.5-7B [26]44.8744.7242.2024.3424.5523.6323.86
LLaVA-NeXT-vicuna-7B [27]50.7249.4746.0129.6829.3531.9531.95
LLaVA-NeXT-Ilama3-8B [41]52.1549.7345.9230.3129.7732.1232.12
PO3D-VQA [45]86.4682.5580.6470.4981.4068.1271.06
Human89.9786.8384.9582.7684.9581.8279.94
", + "image_path": "f29d6be82835acb5026359d636086e59a52279a26b4c6558c545e3449a152355.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 254, + 555, + 323 + ], + "lines": [ + { + "bbox": [ + 55, + 254, + 555, + 323 + ], + "spans": [ + { + "bbox": [ + 55, + 254, + 555, + 323 + ], + "type": "text", + "content": "Table 2. Performance comparison across all 7 question types under 5 difficulty levels. As task complexity increases from single-object questions to 6D spatial relationships, performance generally decreases in most cases, highlighting the challenges in handling multi-object interactions, 3D orientations, and predictive spatial reasoning. The highest score among all models for each question type is marked in red , and the second-highest scores are underlined. The best performance among open-source models is marked in blue . PO3D-VQA is the state-of-the-art (SOTA) 3D-aware neural symbolic reasoning model as proposed in [45], and is listed as the potential upper bound of evaluated methods. The comparison with human performance also shows an increasing gap for LLMs as task difficulty increases." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 342, + 294, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 342, + 294, + 403 + ], + "spans": [ + { + "bbox": [ + 54, + 342, + 294, + 403 + ], + "type": "text", + "content": "multiple-object questions (L2-Multi-object.), 4,995 2D spatial relationship questions (L3-2D-Spatial), 3,534 occlusion questions (L4-Occlusion), 4,555 3D pose questions (L4-3D-Pose), 4,357 collision questions (L5-Collision), and 4,995 6D spatial relationship questions (L5-6D-Spatial)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 403, + 295, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 403, + 295, + 511 + ], + "spans": [ + { + "bbox": [ + 54, + 403, + 295, + 511 + ], + "type": "text", + "content": "We evaluate two series of large vision-language models on Spatial457: (1) API-based large language models, including GPT-4o [29], Gemini-1.5-Pro [37], and Claude-3.5-Sonnet [3]; and (2) open-source large vision-language models, including Qwen2-VL-7B-Instruct [39], InternVL2-8B [38], llava-v1.5-7b [26], llava next-vicuna-7b [27], and llava next-llama3-8b [41]. The evaluation follows the format of Vlmealkit [12]. The details of the model implementation including prompts are listed in Appendix." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 518, + 167, + 529 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 518, + 167, + 529 + ], + "spans": [ + { + "bbox": [ + 55, + 518, + 167, + 529 + ], + "type": "text", + "content": "4.2. Evaluation Results" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 533, + 295, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 533, + 295, + 640 + ], + "spans": [ + { + "bbox": [ + 55, + 533, + 295, + 640 + ], + "type": "text", + "content": "In Tab. 2, we present the performance of various large vision-language models across different levels of difficulty on Spatial457. The levels range from basic single-object tasks to complex 6D spatial reasoning, with increasing difficulty as more factors are introduced. We compare both API-based models (e.g., GPT-4o, GeminiPro1.5, Claude3-5V) and open-source models (e.g., Qwen2-VL-7B-Instruct, InternVL2-8B, LLaVA series) to assess their spatial reasoning capabilities." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": "Comparison across difficulty levels. We evaluate model performance across different levels of difficulty in Spatial457, ranging from basic single-object tasks to complex 6D spatial reasoning. As shown in Tab. 2, performance generally decreases as the complexity of spatial relationships increases, highlighting the challenges models face" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 342, + 553, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 342, + 553, + 366 + ], + "spans": [ + { + "bbox": [ + 313, + 342, + 553, + 366 + ], + "type": "text", + "content": "when handling multi-object interactions, 3D orientations, and predictive spatial reasoning tasks." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 367, + 555, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 367, + 555, + 475 + ], + "spans": [ + { + "bbox": [ + 313, + 367, + 555, + 475 + ], + "type": "text", + "content": "Comparison between API models and open-source models. From Tab. 2, we observe that API models generally outperform open-source models across all difficulty levels. In 2D spatial relationship questions, InternVL2-8B achieves slightly higher results than the API models, while GPT-4o is only " + }, + { + "bbox": [ + 313, + 367, + 555, + 475 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 313, + 367, + 555, + 475 + ], + "type": "text", + "content": " lower. Qwen2-VL-7B-Instruct achieves comparable results, especially in simpler tasks like single-object and multiple-object reasoning. This analysis highlights the strengths of API models in complex spatial reasoning tasks." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 475, + 554, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 475, + 554, + 571 + ], + "spans": [ + { + "bbox": [ + 313, + 475, + 554, + 571 + ], + "type": "text", + "content": "Comparison among API models. API-based models, i.e. GPT-4o, GeminiPro1.5, and Claude3-5V, demonstrate strengths across all difficulty levels. For instance, GPT-4o excels in 2D and 3D spatial tasks, while GeminiPro1.5 shows robust performance in collision prediction tasks. This comparison provides insights into the unique capabilities and limitations of each API model in handling spatial reasoning across different levels of complexity." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 578, + 493, + 590 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 578, + 493, + 590 + ], + "spans": [ + { + "bbox": [ + 313, + 578, + 493, + 590 + ], + "type": "text", + "content": "4.3. RPDR Analysis for 4 Capabilities" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "content": "As outlined in Sec. 3.5, we use the Relative Performance Dropping Rate (RPDR) to analyze model performance across the four core factors of Spatial457: multi-object reasoning, 2D location understanding, 3D location comprehension, and 3D orientation interpretation. The RPDR analysis provides a structured approach to quantify the decline in accuracy with the introduction of each new factor, as detailed in Tab. 3. Our findings reveal that all models exhibit weaknesses in 3D reasoning. Specifically, GPT-4o and GeminiPro-1.5 perform the worst in 3D orientation tasks," + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "24674" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 266, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 266, + 84 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 266, + 84 + ], + "type": "text", + "content": "while the other two struggle most with 3D location." + } + ] + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 63, + 93, + 289, + 154 + ], + "blocks": [ + { + "bbox": [ + 63, + 93, + 289, + 154 + ], + "lines": [ + { + "bbox": [ + 63, + 93, + 289, + 154 + ], + "spans": [ + { + "bbox": [ + 63, + 93, + 289, + 154 + ], + "type": "table", + "html": "
ModelMulti.2D Loc.3D Ori.3D Loc.
GPT-4o [29]84.4589.2777.4686.74
GeminiPro 1.5 [37]85.3787.1383.2888.79
Claude 3.5 S. [3]84.1294.4185.4372.05
Qwen2-VL-7B [39]85.3990.0680.9974.40
", + "image_path": "45a20e07cf0ec03b6ed0a9443cd4535e70aaa84db21deaab709ff59b57cc9cb1.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 204, + 141, + 216 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 204, + 141, + 216 + ], + "spans": [ + { + "bbox": [ + 55, + 204, + 141, + 216 + ], + "type": "text", + "content": "4.4. Bias Analysis" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 217, + 296, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 217, + 296, + 288 + ], + "spans": [ + { + "bbox": [ + 55, + 217, + 296, + 288 + ], + "type": "text", + "content": "In this section, we analyze prediction biases in large language models (LLMs) regarding their understanding of key attributes, specifically shape, color, size, and pose. Leveraging the unbiased generation of dataset images and questions in Spatial457, we examine how these models may demonstrate attribute-specific biases in their predictions." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 288, + 296, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 288, + 296, + 407 + ], + "spans": [ + { + "bbox": [ + 54, + 288, + 296, + 407 + ], + "type": "text", + "content": "We first compute the confusion matrices and distribution statistics for both ground truth and model predictions across each attribute (shape, color, size, and pose). For example, the confusion matrix plots for color and pose attributes of GPT-4o and GeminiPro1.5, together with the distribution of prediction and ground truth label are shown in Fig. 4. We find both GPT-4o and GeminiPro1.5 exhibit bias in both color and pose predictions. These results highlight areas for improvement to reduce bias and enhance predictive reliability across models." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 407, + 296, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 407, + 296, + 430 + ], + "spans": [ + { + "bbox": [ + 55, + 407, + 296, + 430 + ], + "type": "text", + "content": "To quantify prediction variability, we calculate the Coefficient of Variation (CV) for each attribute as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 125, + 434, + 295, + 457 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 125, + 434, + 295, + 457 + ], + "spans": [ + { + "bbox": [ + 125, + 434, + 295, + 457 + ], + "type": "interline_equation", + "content": "\\mathrm {C V} = \\frac {\\operatorname {S t d} \\left(p _ {1} , \\dots , p _ {k}\\right)}{\\operatorname {M e a n} \\left(p _ {1} , \\dots , p _ {k}\\right)}, \\tag {1}", + "image_path": "9194e8d39721d58c99b2a666930a420adbc859588e990400094e6587d5a83772.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 459, + 295, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 459, + 295, + 493 + ], + "spans": [ + { + "bbox": [ + 55, + 459, + 295, + 493 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 459, + 295, + 493 + ], + "type": "inline_equation", + "content": "\\operatorname{Std}(\\cdot)" + }, + { + "bbox": [ + 55, + 459, + 295, + 493 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 459, + 295, + 493 + ], + "type": "inline_equation", + "content": "\\operatorname{Mean}(\\cdot)" + }, + { + "bbox": [ + 55, + 459, + 295, + 493 + ], + "type": "text", + "content": " are the standard deviation and mean of the probability distribution " + }, + { + "bbox": [ + 55, + 459, + 295, + 493 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 55, + 459, + 295, + 493 + ], + "type": "text", + "content": " for each attribute label " + }, + { + "bbox": [ + 55, + 459, + 295, + 493 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 459, + 295, + 493 + ], + "type": "text", + "content": " (e.g., left), \"right\" for pose)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 494, + 296, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 494, + 296, + 566 + ], + "spans": [ + { + "bbox": [ + 55, + 494, + 296, + 566 + ], + "type": "text", + "content": "Tab. 4 shows the CV values for shape, color, size, and pose attributes across different models. Lower CV values indicate more consistent predictions with less bias, which reveals significant biases in model predictions across attributes. Higher CV values, especially in color and pose, indicate consistency challenges." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 568, + 260, + 581 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 568, + 260, + 581 + ], + "spans": [ + { + "bbox": [ + 55, + 568, + 260, + 581 + ], + "type": "text", + "content": "4.5. Comparison to the 3D Symbolic Model" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "content": "As neural symbolic models achieve state-of-the-art (SOTA) performance on many prior synthetic datasets, we implement the 3D-aware neural symbolic model PO3DVQA [45] on our new image domain and all levels of questions. We train PO3D-VQA on a separate set of 20,000 images and questions and report its performance on the same test dataset in Tab. 2. A direct comparison between the 3D neural symbolic model and large language models is not entirely fair, as each module of the neural symbolic model (6D pose estimation, attribute classification, and language reasoning) is trained independently as a sub-task. However, we" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 328, + 72, + 541, + 290 + ], + "blocks": [ + { + "bbox": [ + 328, + 72, + 541, + 290 + ], + "lines": [ + { + "bbox": [ + 328, + 72, + 541, + 290 + ], + "spans": [ + { + "bbox": [ + 328, + 72, + 541, + 290 + ], + "type": "image", + "image_path": "1269a9c5dd214d8344a48f511976c9e6edf5d8b671ec81807079c72f0fcb378c.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 406, + 292, + 465, + 301 + ], + "lines": [ + { + "bbox": [ + 406, + 292, + 465, + 301 + ], + "spans": [ + { + "bbox": [ + 406, + 292, + 465, + 301 + ], + "type": "text", + "content": "(b) GeminiPro1-5" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 311, + 555, + 367 + ], + "lines": [ + { + "bbox": [ + 313, + 311, + 555, + 367 + ], + "spans": [ + { + "bbox": [ + 313, + 311, + 555, + 367 + ], + "type": "text", + "content": "Figure 4. The distribution of color and pose attributes in the L4-3D-Pose task for GPT-4o and Gemini-Pro1.5. Although the ground truth labels are well balanced, the predicted colors and poses for both models are imbalanced, with \"yellow\" being the most frequent color and \"front\" the most frequent pose." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "table", + "bbox": [ + 317, + 381, + 553, + 457 + ], + "blocks": [ + { + "bbox": [ + 55, + 162, + 295, + 196 + ], + "lines": [ + { + "bbox": [ + 55, + 162, + 295, + 196 + ], + "spans": [ + { + "bbox": [ + 55, + 162, + 295, + 196 + ], + "type": "text", + "content": "Table 3. RPDR (%) analysis on the four spatial reasoning capabilities, where higher values indicate better performance. For each row, we highlight the lowest RPDR score among all factors." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 381, + 553, + 457 + ], + "lines": [ + { + "bbox": [ + 317, + 381, + 553, + 457 + ], + "spans": [ + { + "bbox": [ + 317, + 381, + 553, + 457 + ], + "type": "table", + "html": "
ModelsAvg. ↓ShapeColorSizePose
GPT4o [29]0.4850.6230.5210.0430.752
GeminiPro1.5 [37]0.4980.5670.3660.2970.760
Claude3.5V-S [3]0.7200.7770.4650.5741.063
Qwen2-VL-7B [39]0.5450.8680.5580.0670.688
Human0.0930.1100.0820.0550.126
", + "image_path": "9982cddff640840295fbb2d9c80fad34199c2d6f675254e41a027490fa9986d0.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_body" + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 464, + 554, + 531 + ], + "lines": [ + { + "bbox": [ + 313, + 464, + 554, + 531 + ], + "spans": [ + { + "bbox": [ + 313, + 464, + 554, + 531 + ], + "type": "text", + "content": "Table 4. Coefficient of Variation (CV) for predictions on each attribute in the L4-3D-Pose task. We evaluate three API models and one open-source model, calculating the CV across four attributes (shape, color, size, 3D pose) based on the distribution of predicted labels. Lower CV values indicate less bias. The first column shows the average CV across attributes, and GPT4-o has the lowest bias." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 548, + 553, + 572 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 548, + 553, + 572 + ], + "spans": [ + { + "bbox": [ + 313, + 548, + 553, + 572 + ], + "type": "text", + "content": "include this model to illustrate the upper bound that a data-driven machine learning model can achieve on our dataset." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 579, + 426, + 591 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 579, + 426, + 591 + ], + "spans": [ + { + "bbox": [ + 313, + 579, + 426, + 591 + ], + "type": "text", + "content": "4.6. Qualitative Results" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "content": "Fig. 5 presents an example of GPT-4o's performance across various spatial reasoning tasks on the same image. GPT-4o correctly answers multi-object, 2D spatial, and occlusion tasks but fails on 3D pose and 6D spatial questions, indicating challenges in 3D orientation understanding in this case. As we prompt the models to output the answer as well as the reasoning steps, the models correctly detect the existence of objects (e.g. in Fig. 5 the yellow minivan in L4-Pose or the fighter plane and road bike in L5 questions, but has the wrong understand to the orientations or 3D locations." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24675" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 67, + 70, + 282, + 205 + ], + "blocks": [ + { + "bbox": [ + 67, + 70, + 282, + 205 + ], + "lines": [ + { + "bbox": [ + 67, + 70, + 282, + 205 + ], + "spans": [ + { + "bbox": [ + 67, + 70, + 282, + 205 + ], + "type": "image", + "image_path": "7fa982bf2e5866389df7429e32bea4da7f224784da2851f9cb02347b8566918b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 211, + 296, + 266 + ], + "lines": [ + { + "bbox": [ + 55, + 211, + 296, + 266 + ], + "spans": [ + { + "bbox": [ + 55, + 211, + 296, + 266 + ], + "type": "text", + "content": "Figure 5. Example of GPT-4o across different spatial reasoning tasks on the same image with the ground truth answer marked in blue. GPT-4o answers multi-object, 2D spatial, and occlusion tasks correctly (in the green block) but fails on 3D pose and 6D spatial questions (the red block)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 285, + 203, + 297 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 285, + 203, + 297 + ], + "spans": [ + { + "bbox": [ + 55, + 285, + 203, + 297 + ], + "type": "text", + "content": "5. Extension to Real Dataset" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 300, + 295, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 300, + 295, + 407 + ], + "spans": [ + { + "bbox": [ + 55, + 300, + 295, + 407 + ], + "type": "text", + "content": "We extend the setting of 3D questions to real-world image data. Due to limitations in fully annotated attributes such as category, color, 3D orientation, position, and occlusion, we focus on questions from the L4-Pose category. Using the scene structure from SUN-RGBD [36], we construct a new subset, L4-Pose-Real, which involves only the category and pose attributes. To reduce ambiguity in object localization, we add a red bounding box around target objects, ensuring clarity in identifying categories within the scene." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 407, + 296, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 407, + 296, + 479 + ], + "spans": [ + { + "bbox": [ + 55, + 407, + 296, + 479 + ], + "type": "text", + "content": "We provide the question along with the image as input to the LLM. In this example in Fig. 6, GPT-4o provides the correct answer, while Gemini-Pro 1.5 is incorrect. However, the reasoning process reveals that both models rely on common sense knowledge or 2D visibility cues to interpret the 3D orientation." + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 67, + 487, + 280, + 612 + ], + "blocks": [ + { + "bbox": [ + 67, + 487, + 280, + 612 + ], + "lines": [ + { + "bbox": [ + 67, + 487, + 280, + 612 + ], + "spans": [ + { + "bbox": [ + 67, + 487, + 280, + 612 + ], + "type": "image", + "image_path": "760ee8a6c6463ba8e962e210d939301f889828d4dfff2cc5f58051e282124c79.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 617, + 295, + 674 + ], + "lines": [ + { + "bbox": [ + 55, + 617, + 295, + 674 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 295, + 674 + ], + "type": "text", + "content": "Figure 6. Example of real-world 3D pose reasoning tasks " + }, + { + "bbox": [ + 55, + 617, + 295, + 674 + ], + "type": "inline_equation", + "content": "L4" + }, + { + "bbox": [ + 55, + 617, + 295, + 674 + ], + "type": "text", + "content": "-Pose-Real. Here we show the reasoning and answers provided by LLMs where GPT-4o is correct and Gemini-Pro 1.5 is wrong. But the reasoning process reveals these two LLMs are both reasoning 3D orientation from the command sense or 2D visibility." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "As shown in Tab. 5, Gemini Pro 1.5 achieves the highest accuracy (40.50%) in L4-Pose-Real, and the gap between the three API-based models and the open-source" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "text", + "content": "model Qwen2-VL-7B-Instruct (26.50%) is larger. However, all these models perform significantly lower than humans, which have accuracy 92.64%. We further analyze the bias of predictions in this real-world question set. Among the models, Claude3-5V Sonnet exhibits the lowest CV (0.374), indicating relatively stable predictions, while Qwen2-VL-7B-Instruct has the highest CV (0.934). The distribution of ground truth and predicted pose and the confusion matrix are shown in Fig. 7." + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 322, + 194, + 545, + 278 + ], + "blocks": [ + { + "bbox": [ + 322, + 194, + 545, + 278 + ], + "lines": [ + { + "bbox": [ + 322, + 194, + 545, + 278 + ], + "spans": [ + { + "bbox": [ + 322, + 194, + 545, + 278 + ], + "type": "table", + "html": "
ModelsAccuracy (%) ↑CV ↓
GPT-4o [29]38.750.690
GeminiPro1-5 [37]40.500.762
Claude3-5V Sonnet [3]38.750.374
Qwen2-VL-7B-Instruct [39]26.500.934
Human92.640.197
", + "image_path": "8bd8ef8160a6e4fff4b2a8811718d82fbcf814436744faa44f68f80ba56bda0d.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 279, + 555, + 324 + ], + "lines": [ + { + "bbox": [ + 313, + 279, + 555, + 324 + ], + "spans": [ + { + "bbox": [ + 313, + 279, + 555, + 324 + ], + "type": "text", + "content": "Table 5. Accuracy and Coefficient of Variation (CV) for L4-Poserel task: Comparison of accuracy and CV for GPT-4o, GeminiPro1.5, Claude3-5V, and Qwen2-VL-7B-Instruct on real-world 3D pose tasks. Human performance is included as a reference." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 328, + 346, + 544, + 466 + ], + "blocks": [ + { + "bbox": [ + 328, + 346, + 544, + 466 + ], + "lines": [ + { + "bbox": [ + 328, + 346, + 544, + 466 + ], + "spans": [ + { + "bbox": [ + 328, + 346, + 544, + 466 + ], + "type": "image", + "image_path": "a544513755b30e876e51395ca63b4f69cbfe29077150cfa24ebd01acf46cd70d.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 324, + 475, + 543, + 487 + ], + "lines": [ + { + "bbox": [ + 324, + 475, + 543, + 487 + ], + "spans": [ + { + "bbox": [ + 324, + 475, + 543, + 487 + ], + "type": "text", + "content": "Figure 7. Distribution of 3D poses for in L4-Pose-Real task." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 505, + 388, + 517 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 505, + 388, + 517 + ], + "spans": [ + { + "bbox": [ + 313, + 505, + 388, + 517 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "type": "text", + "content": "In this work, we introduce Spatial457, a scalable and unbiased benchmark designed to comprehensively assess 6D spatial reasoning capabilities in LMMs. Spatial457 extends traditional 2D and 3D VQA benchmarks by incorporating complex 6D spatial reasoning tasks, structured across five levels of difficulty. Through evaluations, we observe that current LMMs demonstrate strong performance in basic object recognition and 2D spatial relationships, but struggle significantly with 3D spatial understanding and collision prediction tasks. Additionally, our analysis of prediction biases uncovers attribute-specific inconsistencies that could affect model reliability in real-world applications. Limitations. To real images settings, current evaluation limits to L3 pose questions. We are actively working on building the additional required annotations for these real images to expand the domain of the proposed benchmark." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "24676" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 91, + 295, + 126 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 91, + 295, + 126 + ], + "spans": [ + { + "bbox": [ + 56, + 91, + 295, + 126 + ], + "type": "text", + "content": "This work is supported by ONR with award N00014-23-1-2641, from ARL Army Research Laboratory with award W911NF2320008." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 137, + 115, + 149 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 137, + 115, + 149 + ], + "spans": [ + { + "bbox": [ + 56, + 137, + 115, + 149 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 157, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 157, + 295, + 178 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 157, + 295, + 178 + ], + "spans": [ + { + "bbox": [ + 61, + 157, + 295, + 178 + ], + "type": "text", + "content": "[1] Alibaba DAMO Academy. Gemini: Vision-language model by alibaba. Alibaba DAMO, 2023." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 180, + 296, + 234 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 180, + 296, + 234 + ], + "spans": [ + { + "bbox": [ + 61, + 180, + 296, + 234 + ], + "type": "text", + "content": "[2] Boshi An, Yiran Geng, Kai Chen, Xiaqi Li, Qi Dou, and Hao Dong. Rgbmanip: Monocular image-based robotic manipulation through active object pose estimation. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 7748-7755. IEEE, 2024." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 236, + 294, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 236, + 294, + 257 + ], + "spans": [ + { + "bbox": [ + 62, + 236, + 294, + 257 + ], + "type": "text", + "content": "[3] Anthropic. Claude: An anthropic language model. Anthropotic, 2023." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 258, + 295, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 258, + 295, + 312 + ], + "spans": [ + { + "bbox": [ + 62, + 258, + 295, + 312 + ], + "type": "text", + "content": "[4] Gilad Baruch, Zhuoyuan Chen, Afshin Dehghan, Tal Dimry, Yuri Feigin, Peter Fu, Thomas Gebauer, Brandon Joffe, Daniel Kurz, Arik Schwartz, et al. Arkitsscenes: A diverse real-world dataset for 3d indoor scene understanding using mobile rgb-d data. arXiv preprint arXiv:2111.08897, 2021." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 315, + 294, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 315, + 294, + 346 + ], + "spans": [ + { + "bbox": [ + 62, + 315, + 294, + 346 + ], + "type": "text", + "content": "[5] Mark Billinghurst, Adrian Clark, and Gun Lee. A survey of augmented reality. Foundations and Trends in Human-Computer Interaction, 8(2-3):73-272, 2015." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 347, + 295, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 347, + 295, + 402 + ], + "spans": [ + { + "bbox": [ + 62, + 347, + 295, + 402 + ], + "type": "text", + "content": "[6] Garrick Brazil, Abhinav Kumar, Julian Straub, Nikhila Ravi, Justin Johnson, and Georgia Gkioxari. Omni3d: A large benchmark and model for 3d object detection in the wild. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 13154-13164, 2023." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 403, + 296, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 403, + 296, + 468 + ], + "spans": [ + { + "bbox": [ + 62, + 403, + 296, + 468 + ], + "type": "text", + "content": "[7] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11621-11631, 2020." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 470, + 294, + 513 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 470, + 294, + 513 + ], + "spans": [ + { + "bbox": [ + 62, + 470, + 294, + 513 + ], + "type": "text", + "content": "[8] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 514, + 294, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 514, + 294, + 578 + ], + "spans": [ + { + "bbox": [ + 62, + 514, + 294, + 578 + ], + "type": "text", + "content": "[9] Boyuan Chen, Zhuo Xu, Sean Kirmani, Brain Ichter, Dorsa Sadigh, Leonidas Guibas, and Fei Xia. Spatialvm: Endowing vision-language models with spatial reasoning capabilities. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14455-14465, 2024." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 580, + 294, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 580, + 294, + 624 + ], + "spans": [ + { + "bbox": [ + 57, + 580, + 294, + 624 + ], + "type": "text", + "content": "[10] An-Chieh Cheng, Hongxu Yin, Yang Fu, Qiushan Guo, Ruihan Yang, Jan Kautz, Xiaolong Wang, and Sifei Liu. Spatial-rgpt: Grounded spatial reasoning in vision language model. arXiv preprint arXiv:2406.01584, 2024." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 625, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 625, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 57, + 625, + 294, + 689 + ], + "type": "text", + "content": "[11] Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhanghao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E Gonzalez, et al. Vicuna: An open-source chatbot impressing gpt-4 with " + }, + { + "bbox": [ + 57, + 625, + 294, + 689 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 57, + 625, + 294, + 689 + ], + "type": "text", + "content": " chatgpt quality, march 2023. URL https://lmsys.org/blog/2023-03-30-vicuna, 3(5), 2023." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 691, + 294, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 294, + 714 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 294, + 714 + ], + "type": "text", + "content": "[12] Haodong Duan, Junming Yang, Yuxuan Qiao, Xinyu Fang, Lin Chen, Yuan Liu, Xiaoyi Dong, Yuhang Zang, Pan Zhang," + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 117 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 117 + ], + "type": "text", + "content": "Jiaqi Wang, et al. Vlmealkit: An open-source toolkit for evaluating large multi-modality models. In Proceedings of the 32nd ACM international conference on multimedia, pages 11198-11201, 2024." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 119, + 553, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 119, + 553, + 162 + ], + "spans": [ + { + "bbox": [ + 316, + 119, + 553, + 162 + ], + "type": "text", + "content": "[13] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. International Journal of Computer Vision, 88(2):303-338, 2010." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 164, + 553, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 164, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 317, + 164, + 553, + 217 + ], + "type": "text", + "content": "[14] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 3354-3361, 2012." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 220, + 553, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 220, + 553, + 275 + ], + "spans": [ + { + "bbox": [ + 316, + 220, + 553, + 275 + ], + "type": "text", + "content": "[15] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Batra, and Devi Parikh. Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6904-6913, 2017." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 277, + 553, + 298 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 277, + 553, + 298 + ], + "spans": [ + { + "bbox": [ + 316, + 277, + 553, + 298 + ], + "type": "text", + "content": "[16] Roland Hess. *Blender foundations: The essential guide to learning blender* 2.5. Routledge, 2013." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 300, + 553, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 300, + 553, + 342 + ], + "spans": [ + { + "bbox": [ + 317, + 300, + 553, + 342 + ], + "type": "text", + "content": "[17] Yushi Hu, Hang Hua, Zhengyuan Yang, Weijia Shi, Noah A Smith, and Jiebo Luo. Promptcap: Prompt-guided task-aware image captioning. arXiv preprint arXiv:2211.09699, 2022." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 345, + 553, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 345, + 553, + 399 + ], + "spans": [ + { + "bbox": [ + 316, + 345, + 553, + 399 + ], + "type": "text", + "content": "[18] Drew A Hudson and Christopher D Manning. Gqa: A new dataset for real-world visual reasoning and compositional question answering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6700-6709, 2019." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 401, + 553, + 467 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 401, + 553, + 467 + ], + "spans": [ + { + "bbox": [ + 317, + 401, + 553, + 467 + ], + "type": "text", + "content": "[19] Justin Johnson, Bharath Hariharan, Laurens Van Der Maaten, Li Fei-Fei, C Lawrence Zitnick, and Ross Girshick. Clevr: A diagnostic dataset for compositional language and elementary visual reasoning. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2901–2910, 2017." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 468, + 553, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 468, + 553, + 523 + ], + "spans": [ + { + "bbox": [ + 317, + 468, + 553, + 523 + ], + "type": "text", + "content": "[20] Aishwarya Kamath, Mannat Singh, Yann LeCun, Gabriel Synnaeve, Ishan Misra, and Nicolas Carion. Mdetr-modulated detection for end-to-end multi-modal understanding. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1780-1790, 2021." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 524, + 555, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 524, + 555, + 579 + ], + "spans": [ + { + "bbox": [ + 317, + 524, + 555, + 579 + ], + "type": "text", + "content": "[21] Dong-Jin Kim, Jinsoo Choi, Tae-Hyun Oh, and In So Kweon. Dense relational captioning: Triple-stream networks for relationship-based captioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6271-6280, 2019." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 581, + 553, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 581, + 553, + 656 + ], + "spans": [ + { + "bbox": [ + 317, + 581, + 553, + 656 + ], + "type": "text", + "content": "[22] Alina Kuznetsova, Hassan Rom, Neil Alldrin, Jasper Uijlings, Ilya Krasin, Jordi Pont-Tuset, Gleb Muryshkin, Sergey Popov, Marwa Elhachem, Pedro de Carvalho, et al. The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale. International Journal of Computer Vision, 128(7):1956-1981, 2020." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 658, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 553, + 712 + ], + "type": "text", + "content": "[23] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730-19742. PMLR, 2023." + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24677" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 297, + 139 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 297, + 139 + ], + "type": "text", + "content": "[24] Zhuowan Li, Xingrui Wang, Elias Stengel-Eskin, Adam Kortylewski, Wufei Ma, Benjamin Van Durme, and Alan L Yuille. Super-clevr: A virtual benchmark to diagnose domain robustness in visual reasoning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14963-14973, 2023." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 140, + 295, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 140, + 295, + 183 + ], + "spans": [ + { + "bbox": [ + 56, + 140, + 295, + 183 + ], + "type": "text", + "content": "[25] Adam Dahlgren Lindstrom and Savitha Sam Abraham. Clevr-math: A dataset for compositional language, visual and mathematical reasoning. arXiv preprint arXiv:2208.05358, 2022." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 185, + 294, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 185, + 294, + 207 + ], + "spans": [ + { + "bbox": [ + 57, + 185, + 294, + 207 + ], + "type": "text", + "content": "[26] Haotian Liu et al. Llava: Large language and vision assistant. LLaVA Project, 2023." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 209, + 295, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 209, + 295, + 229 + ], + "spans": [ + { + "bbox": [ + 57, + 209, + 295, + 229 + ], + "type": "text", + "content": "[27] Haotian Liu et al. Llava-last: Extended large language and vision assistant. LLaVA Project, 2023." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 57, + 232, + 295, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 232, + 295, + 274 + ], + "spans": [ + { + "bbox": [ + 57, + 232, + 295, + 274 + ], + "type": "text", + "content": "[28] Jiayuan Mao, Chuang Gan, Pushmeet Kohli, Joshua B Tenenbaum, and Jiajun Wu. The neuro-symbolic concept learner: Interpreting scenes, words, and sentences from natural supervision. arXiv preprint arXiv:1904.12584, 2019." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 57, + 276, + 250, + 287 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 276, + 250, + 287 + ], + "spans": [ + { + "bbox": [ + 57, + 276, + 250, + 287 + ], + "type": "text", + "content": "[29] OpenAI. Gpt-4 technical report. OpenAI, 2023." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 289, + 295, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 289, + 295, + 342 + ], + "spans": [ + { + "bbox": [ + 56, + 289, + 295, + 342 + ], + "type": "text", + "content": "[30] Weichao Qiu and Alan Yuille. Unrealcv: Connecting computer vision to unreal engine. In Computer Vision-ECCV 2016 Workshops: Amsterdam, The Netherlands, October 8-10 and 15-16, 2016, Proceedings, Part III 14, pages 909-916. Springer, 2016." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 344, + 295, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 344, + 295, + 409 + ], + "spans": [ + { + "bbox": [ + 56, + 344, + 295, + 409 + ], + "type": "text", + "content": "[31] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 411, + 295, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 411, + 295, + 465 + ], + "spans": [ + { + "bbox": [ + 56, + 411, + 295, + 465 + ], + "type": "text", + "content": "[32] Arjit Ray, Jiafei Duan, Reuben Tan, Dina Bashkirova, Rose Hendrix, Kiana Ehsani, Aniruddha Kembhavi, Bryan A Plummer, Ranjay Krishna, Kuo-Hao Zeng, et al. Sat: Spatial aptitude training for multimodal language models. arXiv preprint arXiv:2412.07755, 2024." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 467, + 295, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 467, + 295, + 510 + ], + "spans": [ + { + "bbox": [ + 56, + 467, + 295, + 510 + ], + "type": "text", + "content": "[33] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. IEEE transactions on pattern analysis and machine intelligence, 39(6):1137-1149, 2016." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 512, + 295, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 512, + 295, + 533 + ], + "spans": [ + { + "bbox": [ + 56, + 512, + 295, + 533 + ], + "type": "text", + "content": "[34] Andrew Sanders. An introduction to Unreal engine 4. AK Peters/CRC Press, 2016." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 535, + 295, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 535, + 295, + 589 + ], + "spans": [ + { + "bbox": [ + 56, + 535, + 295, + 589 + ], + "type": "text", + "content": "[35] Amanpreet Singh, Vivek Natarajan, Meet Shah, Yu Jiang, Xinlei Chen, Dhruv Batra, Devi Parikh, and Marcus Rohrbach. Towards vqa models that can read. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8317-8326, 2019." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 590, + 295, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 590, + 295, + 635 + ], + "spans": [ + { + "bbox": [ + 56, + 590, + 295, + 635 + ], + "type": "text", + "content": "[36] Shuran Song, Samuel P Lichtenberg, and Jianxiong Xiao. Sun rgb-d: A rgb-d scene understanding benchmark suite. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 567-576, 2015." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 636, + 295, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 636, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 636, + 295, + 689 + ], + "type": "text", + "content": "[37] Gemini Team, Petko Georgiev, Ving Ian Lei, Ryan Burnell, Libin Bai, Anmol Gulati, Garrett Tanzer, Damien Vincent, Zhufeng Pan, Shibo Wang, et al. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context. arXiv preprint arXiv:2403.05530, 2024." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 691, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 295, + 712 + ], + "type": "text", + "content": "[38] Intern Team. Internvl 2.0: A vision-language model. Intern Foundation, 2023." + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 95 + ], + "type": "text", + "content": "[39] Qwen Team. Qwen-vl: A large multimodal model. Alibaba Group, 2023." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 98, + 553, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 98, + 553, + 152 + ], + "spans": [ + { + "bbox": [ + 317, + 98, + 553, + 152 + ], + "type": "text", + "content": "[40] Shengbang Tong, Ellis Brown, Penghao Wu, Sanghyun Woo, Manoj Middepogu, Sai Charitha Akula, Jihan Yang, Shusheng Yang, Adithya Iyer, Xichen Pan, et al. Cambrian1: A fully open, vision-centric exploration of multimodal llms. arXiv preprint arXiv:2406.16860, 2024." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 154, + 553, + 176 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 154, + 553, + 176 + ], + "spans": [ + { + "bbox": [ + 316, + 154, + 553, + 176 + ], + "type": "text", + "content": "[41] Hugo Touvron et al. Llama: Open and efficient foundation language models. Meta AI, 2023." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 178, + 553, + 221 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 178, + 553, + 221 + ], + "spans": [ + { + "bbox": [ + 317, + 178, + 553, + 221 + ], + "type": "text", + "content": "[42] Oriol Vinyals, Alexander Toshev, Samy Bengio, and Dumitru Erhan. Show and tell: A neural image caption generator. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3156-3164, 2015." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 224, + 553, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 224, + 553, + 289 + ], + "spans": [ + { + "bbox": [ + 316, + 224, + 553, + 289 + ], + "type": "text", + "content": "[43] He Wang, Srinath Sridhar, Jingwei Huang, Julien Valentin, Shuran Song, and Leonidas J Guibas. Normalized object coordinate space for category-level 6d object pose and size estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2642-2651, 2019." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 293, + 553, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 293, + 553, + 336 + ], + "spans": [ + { + "bbox": [ + 316, + 293, + 553, + 336 + ], + "type": "text", + "content": "[44] Xingrui Wang, Wufei Ma, Zhuowan Li, Adam Kortylewski, and Alan Yuille. 3d-aware visual question answering about parts, poses and occlusions. In Thirty-seventh Conference on Neural Information Processing Systems, 2023." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 338, + 553, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 338, + 553, + 381 + ], + "spans": [ + { + "bbox": [ + 316, + 338, + 553, + 381 + ], + "type": "text", + "content": "[45] Xingrui Wang, Wufei Ma, Zhuowan Li, Adam Kortylewski, and Alan L Yuille. 3d-aware visual question answering about parts, poses and occlusions. Advances in Neural Information Processing Systems, 36, 2024." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 384, + 553, + 428 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 384, + 553, + 428 + ], + "spans": [ + { + "bbox": [ + 316, + 384, + 553, + 428 + ], + "type": "text", + "content": "[46] Yu Xiang, Roozbeh Mottaghi, and Silvio Savarese. Beyond Pascal: A benchmark for 3d object detection in the wild. In IEEE winter conference on applications of computer vision, pages 75-82. IEEE, 2014." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 430, + 553, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 430, + 553, + 485 + ], + "spans": [ + { + "bbox": [ + 316, + 430, + 553, + 485 + ], + "type": "text", + "content": "[47] Kelvin Xu, Jimmy Ba, Ryan Kiros, Kyunghyun Cho, Aaron Courville, Ruslan Salakhudinov, Rich Zemel, and Yoshua Bengio. Show, attend and tell: Neural image caption generation with visual attention. In International conference on machine learning, pages 2048-2057. PMLR, 2015." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 487, + 553, + 530 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 487, + 553, + 530 + ], + "spans": [ + { + "bbox": [ + 316, + 487, + 553, + 530 + ], + "type": "text", + "content": "[48] Lin Xu, Yilin Zhao, Daquan Zhou, Zhijie Lin, See Kiong Ng, and Jiashi Feng. Pllava: Parameter-free llava extension from images to videos for video dense captioning. arXiv preprint arXiv:2404.16994, 2024." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 533, + 553, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 533, + 553, + 587 + ], + "spans": [ + { + "bbox": [ + 317, + 533, + 553, + 587 + ], + "type": "text", + "content": "[49] Kexin Yi, Jiajun Wu, Chuang Gan, Antonio Torralba, Pushmeet Kohli, and Josh Tenenbaum. Neural-symbolic vqa: Disentangling reasoning from vision and language understanding. Advances in neural information processing systems, 31, 2018." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 590, + 553, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 590, + 553, + 634 + ], + "spans": [ + { + "bbox": [ + 316, + 590, + 553, + 634 + ], + "type": "text", + "content": "[50] Kexin Yi, Chuang Gan, Yunzhu Li, Pushmeet Kohli, Jiajun Wu, Antonio Torralba, and Joshua B Tenenbaum. Clevrer: Collision events for video representation and reasoning. arXiv preprint arXiv:1910.01442, 2019." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 636, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 636, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 636, + 553, + 712 + ], + "type": "text", + "content": "[51] Xiang Yue, Yuansheng Ni, Kai Zhang, Tianyu Zheng, Ruoqi Liu, Ge Zhang, Samuel Stevens, Dongfu Jiang, Weiming Ren, Yuxuan Sun, et al. Mmmu: A massive multi-discipline multimodal understanding and reasoning benchmark for expert agi. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9556-9567, 2024." + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "24678" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 162 + ], + "type": "list", + "angle": 0, + "index": 2, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 105 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 105 + ], + "type": "text", + "content": "[52] Alan Yuille and Daniel Kersten. Vision as bayesian inference: analysis by synthesis? Trends in cognitive sciences, 10(7):301-308, 2006." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 294, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 294, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 294, + 162 + ], + "type": "text", + "content": "[53] Jiyao Zhang, Weiyao Huang, Bo Peng, Mingdong Wu, Fei Hu, Zijian Chen, Bo Zhao, and Hao Dong. Omni6dpose: A benchmark and model for universal 6d object pose estimation and tracking. In European Conference on Computer Vision, pages 199-216. Springer, 2024." + } + ] + } + ], + "index": 1 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "24679" + } + ] + } + ], + "index": 3 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/SpatialCLIP_ Learning 3D-aware Image Representations from Spatially Discriminative Language/dad5650f-d216-4b30-a0c7-dd7ca6658593_content_list.json b/2025/SpatialCLIP_ Learning 3D-aware Image Representations from Spatially Discriminative Language/dad5650f-d216-4b30-a0c7-dd7ca6658593_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..1f2865f4ce8466f7ec273d454e737637f9be642e --- /dev/null +++ b/2025/SpatialCLIP_ Learning 3D-aware Image Representations from Spatially Discriminative Language/dad5650f-d216-4b30-a0c7-dd7ca6658593_content_list.json @@ -0,0 +1,1727 @@ +[ + { + "type": "text", + "text": "SpatialCLIP: Learning 3D-aware Image Representations from Spatially Discriminative Language", + "text_level": 1, + "bbox": [ + 210, + 130, + 785, + 175 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zehan Wang $^{1*}$ , Sahuai Zhou $^{1*}$ , Shaoxuan He $^{1}$ , Haifeng Huang $^{1}$ , Lihe Yang $^{2}$ , Ziang Zhang $^{1}$ , Xize Cheng $^{1}$ , Shengpeng Ji $^{1}$ , Tao Jin $^{1}$ , Hengshuang Zhao $^{2}$ , Zhou Zhao $^{1\\dagger}$", + "bbox": [ + 153, + 191, + 846, + 227 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1Zhejiang University; 2The University of Hong Kong https://github.com/SpatialVision/Spatial-CLIP", + "bbox": [ + 228, + 233, + 759, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 292, + 326, + 309 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Contrastive Language-Image Pre-training (CLIP) learns robust visual models through language supervision, making it a crucial visual encoding technique for various applications. However, CLIP struggles with comprehending spatial concepts in images, potentially restricting the spatial intelligence of CLIP-based AI systems. In this work, we propose SpatialCLIP, an enhanced version of CLIP with better spatial understanding capabilities. To capture the intricate 3D spatial relationships in images, we improve both \"visual model\" and \"language supervision\" of CLIP. Specifically, we design 3D-inspired ViT to replace the standard ViT in CLIP. By lifting 2D image tokens into 3D space and incorporating design insights from point cloud networks, our visual model gains greater potential for spatial perception. Meanwhile, captions with accurate and detailed spatial information are very rare. To explore better language supervision for spatial understanding, we re-caption images and perturb their spatial phrases as negative descriptions, which compels the visual model to seek spatial cues to distinguish these hard negative captions. With the enhanced visual model, we introduce SpatialLLaVA, following the same LLaVA-1.5 training protocol, to investigate the importance of visual representations for MLLM's spatial intelligence. Furthermore, we create SpatialBench, a benchmark specifically designed to evaluate CLIP and MLLM in spatial reasoning. SpatialCLIP and SpatialLLaVA achieve substantial performance improvements, demonstrating stronger capabilities in spatial perception and reasoning, while maintaining comparable results on general-purpose benchmarks.", + "bbox": [ + 89, + 325, + 485, + 781 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 808, + 220, + 823 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, Contrastive Language-Image Pretraining (CLIP) models [13, 22, 47, 53, 73] have demonstrated the ability to", + "bbox": [ + 89, + 833, + 482, + 864 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "learn robust image representations from language supervision. These models serve as foundational visual encoders for various applications, such as open-world visual recognition [26, 36, 40, 46, 49, 52] and Multi-modal Large Language Models (MLLM) [2, 10, 11, 41, 42, 57].", + "bbox": [ + 511, + 295, + 903, + 369 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Although CLIP exhibits strong zero-shot ability across diverse tasks, it struggles with understanding spatial arrangement within images. As shown in Fig. 1, for a correctly matched image-text pair, we introduce perturbations to the caption by modifying the object category, object attribute, and spatial phrases. The results reveal that altering the object category or attribute significantly reduces the image-text similarity in the CLIP space. However, CLIP struggles to distinguish incorrect spatial relationship descriptions and even assigns a higher matching score to captions with wrong spatial descriptions. Spatial intelligence is a key aspect of visual perception, and limitations in spatial understanding may constrain systems built on CLIP from effectively understanding real-world environments.", + "bbox": [ + 511, + 369, + 905, + 582 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we present SpatialCLIP, a CLIP model with enhanced spatial perception capabilities. As titled in [47], CLIP models are \"Learning Transferable Visual Models from Natural Language Supervision\". Our SpatialCLIP advances the spatial understanding from both \"visual models\" and \"language supervision\" perspectives.", + "bbox": [ + 511, + 582, + 905, + 672 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "For \"visual models,\" we propose a 3D-inspired ViT that processes RGB-D input rather than RGB images. By incorporating depth information, our visual model gains the ability to perceive images from the 3D perspective. This design helps the model understand complex 3D spatial relationships like \"behind,\" \"in front of,\" \"under,\" and \"far.\" Specifically, we lift the 2D tokens into 3D space using the extra depth information. In the 3D space, we integrate two structures widely used in point cloud networks: 3D convolution [28, 61] and 3D relative position encoding [34, 62, 67], into the pre-trained ViT model [18]. By integrating design insights from 3D point cloud processing, our 3D-inspired ViT demonstrates enhanced potential for capturing complex 3D spatial structures in images.", + "bbox": [ + 511, + 672, + 905, + 883 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "For \"language supervision\", in current vision-language", + "bbox": [ + 531, + 885, + 903, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal Contribution.", + "bbox": [ + 107, + 875, + 222, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author.", + "bbox": [ + 109, + 887, + 235, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "29656", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2cad005fad5393440a393d83a4b69d570032eb6aca59fec42d5a8fff2dd68352.jpg", + "image_caption": [ + "Figure 1. The preliminary analysis of OpenAI CLIP-L/14@336px. CLIP Score stands for the cosine similarity between image and language description. These examples demonstrate that while CLIP is highly sensitive to errors involving object categories and attributes, it struggles to accurately identify mistakes in spatial descriptions." + ], + "image_footnote": [ + "- Accurate: The darker-haired cat is behind the white cat. \nCLIP Score: 22.28", + "Spatial Error: The darker-haired cat is to the right of the white cat. \nCLIP Score: 23.66 (+1.38)" + ], + "bbox": [ + 101, + 89, + 290, + 181 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/45ca3f9a068d23df4997ccc971c5fa7aa8345d42f117e54f3ee047f37a1f4467.jpg", + "image_caption": [], + "image_footnote": [ + "- Category Error: The darker-haired bird is behind the white cat. \nCLIP Score: 20.43 (-1.85)", + "-Accurate: The white refrigerator is closer to the camera than brown door. \nCLIP Score: 29.74", + "$\\nrightarrow$ Category Error: The white cabinet is closer to the camera brown door. CLIP Score: 25.66 (-4.08)", + "$\\nrightarrow$ Attribute Error: The dark refrigerator is closer to the camera than opened door. CLIP Score: 29.39 (-0.35)", + "Spatial Error: The white refrigerator is further from the camera than brown door. CLIP Score: 29.62 (-0.12)" + ], + "bbox": [ + 302, + 89, + 493, + 181 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/61f6d082312b421fe2fd2ab3ae77f0d8c8effd548c2285aeef62c587c42913d1.jpg", + "image_caption": [], + "image_footnote": [ + "$\\nLeftrightarrow$ Attribute Error: The light-haired cat is behind the white cat. CLIP Score: 21.22 (-1.06)", + "-Accurate: The dark coffee machine is to the left of the mug. CLIP Score: 22.10", + "$\\rightarrow$ Category Error: The ceramic cup is to the left of the mug. CLIP Score: 19.48 (-2.62)", + "$\\nrightarrow$ Attribute Error: The old broken coffee machine is to the left of the mug. CLIP Score: 21.27 (-0.83)", + "Spatial Error: The dark coffee machine is in front of the mug. CLIP Score: 22.53 (+0.43)" + ], + "bbox": [ + 503, + 89, + 692, + 181 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/b800fcb15ff8474828847cc57bb4e5e06c1201a1c420d9e29a4b072a0daf2b6b.jpg", + "image_caption": [], + "image_footnote": [ + "-Accurate: Two people are standing behind a pile of huge-sized pipes. CLIP Score: 26.62", + "$\\nvdash$ Category Error: Two people are standing behind a pile of huge-sized papers. CLIP Score: 16.40 (-10.22)", + "$\\nvdash$ Attribute Error: Two people are standing behind a pile of small pipes. \nCLIP Score: 25.78 (-0.84)", + "Spatial Error: Two people are standing inside a pile of huge-sized pipes. \nCLIP Score: 27.18 (+0.56)" + ], + "bbox": [ + 704, + 90, + 893, + 181 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "datasets [5, 50, 51] for CLIP training, the input captions are primarily simple summary texts that only describe the primary object category and attributes. The absence of precise spatial information in these captions may constrain the visual model's ability to capture spatial relationships. To enhance language supervision, we recapture images using advanced MLLMs [9, 41] and introduce perturbed spatial phrases to captions as hard negative queries. Training the model to differentiate between captions that share the same object categories and attributes but vary in spatial phrases forces it to better capture spatial cues in images.", + "bbox": [ + 88, + 368, + 482, + 535 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Furthermore, we explore whether SpatialCLIP, with its enhanced spatial perception, can contribute stronger spatial intelligence to MLLMs when using the same data. Recent work [7, 12, 14] mainly enhances the spatial perception capabilities of MLLMs by introducing new training data, while lacking the exploration of spatial-focused visual representations. However, as discussed in [58], visual representation is also crucial to the capabilities of MLLMs. We construct SpatialLLaVA using the same training recipe as the standard LLaVA-1.5 [41], enabling a fair and systematical analysis over MLLMs.", + "bbox": [ + 88, + 536, + 482, + 700 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To quantitatively assess the spatial understanding capabilities of CLIP and MLLM, we introduce SpatialBench, a benchmark designed to evaluate spatial comprehension within images. This benchmark challenges models to comprehend spatial arrangements and perform complex spatial reasoning. For thorough evaluation, we curate and annotate images from both outdoor and indoor scenes.", + "bbox": [ + 89, + 703, + 482, + 808 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Experimentally, SpatialCLIP shows substantial improvements across spatial-focused benchmarks, such as Spatial-Bench and BLINK [24], and performs well on general-purpose downstream tasks [39, 69]. Similarly, SpatialLLaVA demonstrates notably enhanced spatial perception while retaining strong general perceptual abilities.", + "bbox": [ + 89, + 810, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our main contributions are as follows:", + "bbox": [ + 529, + 369, + 787, + 382 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose SpatialCLIP, which improves the CLIP's spatial understanding capabilities by refining both \"visual model\" and \"language supervision\".", + "- We develop SpatialLLaVA, by replacing the visual encoding method in LLaVA with our SpatialCLIP while preserving other training protocols. This investigates the value of SpatialCLIP in enhancing the spatial intelligence of CLIP-based MLLM, as shown in Fig. 2.", + "- We construct SpatialBench to evaluate the spatial capabilities of CLIP and MLLMs, collecting and annotating samples from both outdoor and indoor scenarios.", + "- We conduct extensive experiments on both spatial-focused and general-purpose benchmarks to validate the effectiveness of SpatialCLIP, SpatialLLaVA, and each newly proposed component." + ], + "bbox": [ + 513, + 383, + 905, + 611 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 625, + 653, + 640 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Contrastive Language-Image Pre-training", + "text_level": 1, + "bbox": [ + 511, + 650, + 872, + 667 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "CLIP models [13, 22, 47, 53] are trained on web-scale image-text paired data [5, 50, 51] using multimodal contrastive learning as objective. By learning cross-modal semantic alignment from massive datasets, CLIP demonstrates strong zero-shot generalization capabilities and is widely used in various vision-language downstream tasks, such as image object detection [52], segmentation [48, 60], and video action recognition [40]. In recent multimodal large language models (MLLM), CLIP is the most common visual encoding strategy, enabling large language models with the ability to understand images. LLaVA [42] and subsequent works [35, 35, 41, 64] primarily rely on OpenAI-CLIP [47], while other state-of-the-art MLLMs like Emu2 [54], InternVL1.5 [10] and MiniCPM-v [68] also utilize various CLIP variants, such as EVA-CLIP [53], In", + "bbox": [ + 509, + 672, + 906, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "29657", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/addc6756b77131b54ba6cf9c60030b1cdb6f8d02f256910abd1d371015b51030.jpg", + "image_caption": [ + "Select the best caption for the image:" + ], + "image_footnote": [], + "bbox": [ + 99, + 88, + 295, + 212 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/23424863be70cc3457b3dbe5399bca59f30fd6007034bac4614ffab170f308a5.jpg", + "image_caption": [ + "B. Red electric car is closer to the camera.", + "LLaVA-1.5: The best caption is B.", + "SpatialLLaVA: The best caption is A.", + "Select the best caption for the image:", + "A. The fan is facing the camera.", + "B. The fan is facing the white desk.", + "C. The fan is facing the sofa.", + "D. The fan is facing the chair.", + "LLaVA-1.5: The best caption is A.", + "SpatialLLaVA: The best caption is D.", + "Figure 2. Visualization of LLaVA-1.5 and SpatialLLaVA on SpatialBench. The first row displays examples from outdoor scenes, while the second row features cases in indoor environments. LLaVA-1.5 often encounters difficulties with basic spatial concepts, while SpatialLLaVA improves its spatial perception ability by employing enhanced visual representations." + ], + "image_footnote": [], + "bbox": [ + 99, + 279, + 295, + 402 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/4022cc6b053062907a26b5449eac47bdefc6d1a222aa516f872dcd9e7c40557a.jpg", + "image_caption": [ + "A. Yellow electric car is closer to the camera.", + "Select the best caption for the image: A. The bird standing taller is smaller than the other.", + "B. The bird standing taller is bigger than the other.", + "C. The two birds are roughly the same size." + ], + "image_footnote": [], + "bbox": [ + 297, + 88, + 491, + 213 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/38e9ac6842ef8ffd5194849f124aecda9d165f779b8b423f64918cfb6355b7f6.jpg", + "image_caption": [ + "C. Two cars are equally close to the camera.", + "LLaVA-1.5: The best caption is B.", + "SpatialLLaVA: The best caption is C.", + "Select the best caption for the image:", + "A. All the sofas form a circle", + "B. All the sofas form a rectangle.", + "C. All the sofas form an oval.", + "D. All sofas are arranged randomly.", + "LLaVA-1.5: The best caption is A.", + "SpatialLLaVA: The best caption is B." + ], + "image_footnote": [], + "bbox": [ + 297, + 279, + 491, + 402 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/b4589a4fd056cc1756b66a0016a24460867eca3629756fb45916b25e4204ec10.jpg", + "image_caption": [ + "Select the best caption for the image:", + "A. The two trains are perpendicular.", + "B. The two trains are intersecting.", + "C. The two trains are parallel.", + "A. The stop sign is taller than the wall.", + "B. The stop sign is shorter than the wall.", + "C. The stop sign and the wall are of the same height." + ], + "image_footnote": [], + "bbox": [ + 493, + 88, + 687, + 213 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/a8b1b29f9decad6790b671733642171920730598399f53d1d9140a9dc62c48ef.jpg", + "image_caption": [ + "LLaVA-1.5: The best caption is A.", + "SpatialLLaVA: The best caption is C.", + "Select the best caption for the image:", + "A. Ceiling light hangs above the edge of the table.", + "B. Ceiling light hangs above the center of the table." + ], + "image_footnote": [], + "bbox": [ + 491, + 279, + 687, + 402 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/3e1822acb1b71751ba1e2efd3fc7dbeeabae72b2a705aa9ab09d319e1f10830a.jpg", + "image_caption": [ + "Select the best caption for the image:" + ], + "image_footnote": [], + "bbox": [ + 691, + 88, + 883, + 212 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/b458a8c4f2a4f16ef29591dd4b360c22e474b4e7749acad2ebf39a7a3881371e.jpg", + "image_caption": [ + "LLaVA-1.5: The best caption is C.", + "SpatialLLaVA: The best caption is A.", + "Select the best caption for the image:", + "A. The left window is larger than the right window.", + "B. The right window is larger than the left window.", + "C. The Two mirrors are of the same size.", + "LLaVA-1.5: The best caption is A.", + "SpatialLLaVA: The best caption is C." + ], + "image_footnote": [], + "bbox": [ + 691, + 279, + 883, + 402 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ternVL [11], and SigLIP [73]. However, ReCLIP [52] recognizes CLIP lacks the capability of understanding spatial relationships in images. Additionally, as discussed in [58], the shortcoming in visual encoding may limit the capabilities of MLLMs built upon it.", + "bbox": [ + 89, + 534, + 482, + 609 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our work attempts to enhance CLIP's spatial perception and overcome the shortcomings in spatial understanding.", + "bbox": [ + 89, + 609, + 482, + 640 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Spatial Intelligence in Visual Perception", + "text_level": 1, + "bbox": [ + 89, + 650, + 431, + 666 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Spatial understanding [8, 56] and spatial intelligence are fundamental abilities for artificial intelligence systems to perceive and interact with the complex visual world. In our 3D world, spatial understanding involves distinguishing various spatial relationships between objects, such as up and down, left and right, far and near, and inside and outside, as well as perceiving the size, thickness, length, and width of visual elements. The ability to comprehend and reason on these basic spatial concepts is crucial for solving many complex problems, including robot navigation [70, 75], manipulation [25], human-computer interaction [19, 44], and applications in virtual and augmented reality [33].", + "bbox": [ + 89, + 672, + 482, + 854 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "MLLMs have made impressive progress in recent years. Built upon pre-trained visual representation and large language models [20, 59], these methods demonstrate strong", + "bbox": [ + 89, + 856, + 482, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "performance across a variety of downstream tasks, such as image captioning [39], visual question answering [1], visual grounding [30, 45], and action recognition [23]. However, despite their powerful generalization capabilities, most MLLMs still struggle with spatial understanding and reasoning, as discussed in [7].", + "bbox": [ + 511, + 534, + 906, + 625 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "There are several works that attempt to address the shortcomings of MLLMs in spatial understanding. SpatialVLM [7] is the first to focus on this issue, introducing massive automatically-generated 3D spatial reasoning VQA data for training MLLM. Its successor, SpatialRGPT [12], enhances the 3D-aware data curating pipeline and provides extra visual depth and region information for MLLM. CubeLLM [14] combines multiple existing 2D and 3D object detection datasets in a unified format and then trains MLLM with better spatial detection capability.", + "bbox": [ + 511, + 633, + 906, + 786 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Existing works mainly enhance the spatial intelligence of MLLM by improving the training data, while our work explores the issue from a different perspective: improving the visual representations. Utilizing the same general-purpose training data, simply replacing the standard CLIP with our SpatialCLIP enhances the spatial understanding capabilities of LLaVA [41, 42].", + "bbox": [ + 511, + 795, + 906, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "29658", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/357f47b7c6ca1367b8c172971e79c111504f680c575f5bdc4e53808a860abd0a.jpg", + "image_caption": [ + "Figure 3. The overview of SpatialCLIP. The visual encoder is replaced by our newly proposed 3D-Inspired Vision Transformer, which enables more informative RGB-D inputs and excels in modeling the 3D spatial structure. Language supervision is enhanced through recaption and perturbation, which prompt the encoding of visual spatial information." + ], + "image_footnote": [], + "bbox": [ + 130, + 87, + 617, + 344 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/4b3aa839cd88e41a94e5b2a1dd2b497ca1a672099bc85394228d8183089850bb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 627, + 88, + 862, + 339 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3. SpatialCLIP", + "text_level": 1, + "bbox": [ + 89, + 398, + 220, + 416 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our proposed SpatialCLIP enhances standard CLIP from two perspectives: 3D-Inspired ViT for more powerful \"Visual Model\" in Sec. 3.1 and recaption&perturbation strategy for more informative \"Language Supervision\" in Sec. 3.2.", + "bbox": [ + 89, + 425, + 483, + 487 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Visual Model: 3D-Inspired ViT", + "text_level": 1, + "bbox": [ + 89, + 497, + 367, + 513 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In RGB images, the 3D spatial structure of the world is flattened into a 2D plane. Many methods [12, 15, 17, 74] try to introduce depth maps for extra 3D information. However, how to effectively encode such RGB-D inputs while maintaining the pre-trained knowledge for the RGB channel is non-trivial. AlphaCLIP [55] enables an extra input channel with a simple extra input convolution layer, while SpatialRGPT [12] repeats depth map three times as a depth image, and lately fuses the latent of the RGB image and depth image with the pre-trained CLIP image encoder. Our empirical findings show that such simple early and late fusion methods are less effective at integrating depth information. To address this, we propose 3D-inspired ViT for RGB-D inputs to replace the standard vision transformer (ViT) widely used as the visual model in CLIP models.", + "bbox": [ + 89, + 521, + 483, + 748 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Recently, monocular depth estimation models [31, 65, 66] have gained impressive progress, depth maps can be easily predicted for wild-world RGB images with low cost. Moreover, the accurate depth map shows impressive results in lifting the images into 3D point clouds. Meanwhile, the architectures for processing 3D point cloud [61, 62, 67] focus on capturing spatial relationships among features in 3D space. These observations inspire us to lift 2D tokens into 3D space using predicted depth information and thereby incorporate design insights from 3D point cloud architectures", + "bbox": [ + 89, + 750, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "to better model 3D spatial relationships in images.", + "bbox": [ + 511, + 402, + 844, + 416 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To this end, we first dive into the standard ViT, which processes RGB images as input. Specifically, an RGB image $\\mathbf{X} \\in \\mathbb{R}^{H \\times W \\times 3}$ is divided into 2D patches, and each is linearly projected into $C$ -dimension tokens. The resulting 2D tokens map $\\mathbf{Z} \\in \\mathbb{R}^{H_z \\times W_z \\times C}$ is passed into a stack of transformer blocks, where $H_{z}$ and $W_{z}$ represent the tokens map's height and width. The 2D token map for the $L$ -th transformer block can be denoted as $\\mathbf{Z}_{L} \\in \\mathbb{R}^{H_{z} \\times W_{z} \\times C_{L}}$ .", + "bbox": [ + 511, + 417, + 905, + 537 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To integrate the well-established 3D point cloud architecture designs into ViT for images, for each transformer layer, we first lift its 2D input tokens $\\mathbf{Z}_L \\in \\mathbb{R}^{H_z \\times W_z \\times C_L}$ into 3D voxel $\\mathbf{Z}_L^{3\\mathrm{D}} \\in \\mathbb{R}^{H_z \\times W_z \\times D_z \\times C_L}$ based on the average depth value of each patch, where $H_z$ , $W_z$ , and $D_z$ represent the grid sizes along the height, width, and depth axes. Since $\\mathbf{Z}_L^{3\\mathrm{D}}$ is derived from $\\mathbf{Z}_L$ , the resulting 3D voxel grid is highly sparse. The $H_z \\times W_z$ tokens in $\\mathbf{Z}_L$ are distributed along the depth axis based on their corresponding depth values, and extracting tokens along the depth axis can easily compress $\\mathbf{Z}_L^{3\\mathrm{D}}$ back into 2D space.", + "bbox": [ + 511, + 539, + 905, + 705 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For 3D voxelized tokens, 3D depth-wise convolution is widely used to capture spatial patterns across the three dimensions. As shown in the right part of Fig. 3, we integrate this 3D convolution within each transformer block of the pre-trained ViT model, applying it to $\\mathbf{Z}_L^{3\\mathrm{D}}$ and then short-cutting the resulting features back to the original $\\mathbf{Z}_L$ . This process is expressed as:", + "bbox": [ + 511, + 705, + 905, + 811 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {Z} _ {L} = \\mathbf {Z} _ {L} + \\operatorname {C o m p r e s s} \\left(3 \\mathrm {D} - \\operatorname {C o n v} \\left(\\mathbf {Z} _ {\\mathrm {L}} ^ {3 \\mathrm {D}}\\right)\\right) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 571, + 824, + 903, + 842 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where 3D-Conv $(\\cdot)$ is the 3D depth wise convolution and Compress $(\\cdot)$ indicates compressing sparse 3D voxel back into 2D token maps.", + "bbox": [ + 511, + 854, + 905, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "29659", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Moreover, another commonly used technique in point cloud networks is 3D relative positional encoding, which explicitly indicates the relative spatial relationship between every pair of tokens in 3D space. For the self-attention operator of each transformer layer, we employ 3D relative positional biases on the attention map. The attention score $a_{ij}$ between $i$ -th and $j$ -th token can be represented as:", + "bbox": [ + 89, + 90, + 480, + 196 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\na _ {i j} = \\frac {\\left(\\mathbf {z} _ {i} \\mathbf {W} ^ {Q}\\right) \\left(\\mathbf {z} _ {j} \\mathbf {W} ^ {K}\\right) ^ {T} + b _ {i j}}{\\sqrt {C}} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 178, + 205, + 482, + 241 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $b_{ij}$ is a scalar selected from the learnable bias matrix $\\mathbf{B} \\in \\mathbb{R}^{(2H_z - 1) \\times (2W_z - 1) \\times (2D_z - 1)}$ , based on the relative 3D position between the $i$ -th and $j$ -th tokens in 3D space.", + "bbox": [ + 89, + 250, + 482, + 294 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "By adopting the designs inspired by the structure of 3D point cloud networks, the 3D-inspired ViT can perceive 2D images from the 3D perspective, greatly enhancing the potential to model complex spatial relationships in images.", + "bbox": [ + 89, + 294, + 483, + 354 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2. Language Supervision: Recapiction & Perturbation", + "text_level": 1, + "bbox": [ + 89, + 363, + 482, + 393 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the contrastive learning process of CLIP, visual representations of different images are pulled close with the paired caption while pushed away from captions of other images. The guidance information in language captions strongly influences the visual model's capability. For prompting 3D-Inspired ViT to better capture spatial relationships in images, more informative supervision from natural language is necessary. However, current widely-used vision-language datasets for training CLIP, such as Conceptual-12M [5] and LAION [50, 51], are inherently noisy and mainly provide simple summaries of global attributes, lacking detailed caption with accurate spatial description.", + "bbox": [ + 89, + 402, + 482, + 583 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Recent studies [3, 6, 21, 37] have shown that recaptioning images with semantically richer and more aligned descriptions can greatly enhance the performance of vision-language models like CLIP and text-to-image models. However, as discussed in [7, 12], current MLLMs excel at recognizing object categories and attributes but struggle with spatial perception. Consequently, the detailed captions they generated often fail to accurately capture all spatial relationships within the image.", + "bbox": [ + 89, + 584, + 482, + 718 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To address this \"chicken-and-egg\" dilemma, we propose perturbing the spatial descriptions in the dense captions generated by MLLMs. Specifically, we employ a template to randomly substitute phrases indicating spatial relationships with their antonymic counterparts. For instance, a caption like \"two cats are standing in front of a table, while the left cat is bigger than the right one\" may be modified to \"two cats are standing under a table, while the upper cat is smaller than the lower one.\"", + "bbox": [ + 89, + 719, + 482, + 853 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Since the generated and perturbed captions are semantically consistent in terms of object categories and attributes, applying contrastive learning between these two captions", + "bbox": [ + 89, + 854, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "encourages the model to distinguish the entirely incorrect caption from the partially correct one. Such a straightforward perturbation method effectively promotes the visual model's ability to capture detailed spatial relationships within images. The learning objective can be expressed as:", + "bbox": [ + 511, + 90, + 903, + 167 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} L = - \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\left[ \\log \\frac {e ^ {\\left(\\mathbf {t} _ {i} \\cdot \\mathbf {v} _ {i} / \\tau\\right)}}{\\sum_ {j = 1} ^ {N} e ^ {\\left(\\mathbf {t} _ {i} \\cdot \\mathbf {v} _ {j} / \\tau\\right)}} \\right. \\tag {3} \\\\ \\left. + \\log \\frac {e ^ {\\left(\\mathbf {t} _ {i} \\cdot \\mathbf {v} _ {i} / \\tau\\right)}}{\\sum_ {j = 1} ^ {N} \\left(e ^ {\\left(\\mathbf {t} _ {j} \\cdot \\mathbf {v} _ {i} / \\tau\\right)} + e ^ {\\left(\\tilde {\\mathbf {t}} _ {j} \\cdot \\mathbf {v} _ {i} / \\tau\\right)}\\right)} \\right] \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 178, + 903, + 263 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where the $v_{i}$ is the visual representation of $i$ -th image, while $t_{i}$ and $\\tilde{t}_i$ denote the textual representation of the $i$ -th caption and its perturbed version, respectively.", + "bbox": [ + 511, + 273, + 903, + 321 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. SpatialLLaVA", + "text_level": 1, + "bbox": [ + 513, + 334, + 658, + 352 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As discussed in [57, 58], the capabilities of MLLMs may be constrained by the inherent limitations of their visual representations. This raises a crucial question: Can the spatial intelligence of CLIP-based MLLM be enhanced by improving the spatial perception of CLIP's visual representations?", + "bbox": [ + 511, + 359, + 905, + 436 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We utilize LLaVA-1.5, which employs the pre-trained CLIP as its visual encoding model, as the framework to investigate the effectiveness of SpatialCLIP in MLLM. Specifically, we replace the CLIP model in LLaVA-1.5 [41] with our SpatialCLIP. To ensure fair and systematic comparisons between visual encoding methods, we follow the two-stage training strategy used in LLaVA-1.5 and apply the same training datasets and hyperparameters for each stage. The resulting MLLM is named SpatialLLaVA.", + "bbox": [ + 511, + 436, + 905, + 573 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. SpatialBench", + "text_level": 1, + "bbox": [ + 513, + 587, + 648, + 604 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To evaluate the ability of CLIP and MLLM in understanding complex 2D/3D spatial relations in images, we build a benchmark SpatialBench, which focuses entirely on spatial perception. Our benchmark employs the caption matching task, requiring CLIP or MLLM to distinguish the caption that best matches the image from several distractors. By manually crafting the correct description and distractors, the caption-matching task requires the model to distinguish nuanced spatial concepts in the image (e.g. object arrangement, relationships, relative depth, and size).", + "bbox": [ + 511, + 612, + 905, + 763 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To fully assess models in various scenarios, we collect images of both outdoor and indoor scenes. Specifically, outdoor images are collected from the COCO [39] validation set and the Internet, totaling 159 images. Indoor images are sourced from ScanNet [16], consisting of 140 images. Each image is paired with 2 to 4 manually crafted descriptions, of which only one is accurate, with the remaining serving as distractors. Examples of SpatialBench can be found in Fig. 2 and Appendix.", + "bbox": [ + 511, + 763, + 905, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "29660", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/56693c74fc3fc9dbb5a690c1fba83dc004de00a35273f18881b1e99d188460d7.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelParamSpatial-FocusedGeneral-Purposed
SpatialBenchBLINKCOCOFlickr30K
OutDoor.InDoor.SpatialR@1R@5R@1R@5
Random-29.7729.8450.00----
OpenAI CLIP-B@224px149.6M35.2230.7151.0542.1967.2534.4056.65
SpatialCLIP-B@224px149.9M+24.8M41.5134.2857.3448.3373.2738.8662.80
OpenAI CLIP-L@336px427.9M32.0730.0047.5546.7870.5438.2560.95
SpatialCLIP-L@336px428.7M+24.8M35.8433.5757.3450.6275.6642.5467.04
", + "bbox": [ + 119, + 88, + 874, + 250 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/3ed399047d3c63247f1819bcad8ffbc86308bf156d36d501823e0f64b8141f4d.jpg", + "table_caption": [ + "Table 1. Results on CLIP benchmark. The best results are bold." + ], + "table_footnote": [], + "table_body": "
Training DataModelSpatialBenchBLINKCV-Bench3d
OutDoor.InDoor.SpatialDepthCorresJigsawDepthDistance
Spatial-FocusedSpatial-RGPT-7B37.7338.5765.7382.3032.5550.0062.3359.00
Stronger General-PurposedLLaVA-1.6-7B39.6233.5768.5351.6129.6548.0053.3551.12
LLaVA-1.6-13B38.3635.7169.9948.3830.8150.6755.1652.25
LLaVA-interleave-7B33.2737.1469.9356.6934.8851.3357.1462.30
Baseline General-PurposedLLaVA-1.5-7B29.5532.4761.5452.4225.5811.3345.8350.00
Spatial-LLaVA-7B40.2540.1771.3258.0631.9751.3357.3352.16
", + "bbox": [ + 96, + 279, + 906, + 431 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/84b9401335b357714485fd33617f992361f95e269dee2815505873e01fc9d1ce.jpg", + "table_caption": [ + "Table 2. Results on Spatial-Focused MLLM benchmark." + ], + "table_footnote": [], + "table_body": "
POPE [38]MM.Vet [72]VOAv2 [27]GOA [29]RealWorldQA [63]
LLaVA-1.5-7B85.930.578.562.054.8
Spatial-LLaVA-7B86.530.779.960.654.2
", + "bbox": [ + 94, + 470, + 483, + 594 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 3. Results on General-Purposed MLLM benchmark.", + "bbox": [ + 107, + 595, + 464, + 609 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For CLIP, we calculate matching scores between the image and its corresponding captions, selecting the description with the highest matching score as the prediction. For MLLMs, the descriptions are presented as multiple-choice options, and the models are prompted to select the best description in a question-answering manner.", + "bbox": [ + 89, + 623, + 482, + 715 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6. Experiment", + "text_level": 1, + "bbox": [ + 89, + 729, + 215, + 747 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6.1. Implementation Details", + "text_level": 1, + "bbox": [ + 89, + 755, + 307, + 771 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Training setting. We use images in SA1B dataset [32] for training, since most images in it contain multiple objects and showcase complex spatial relationships. Following [6], LLaVA-1.5-13B [41], CoCa [71] and extra ShareGPT4V [9] are used for image reception, the three kinds of captions are randomly used in training for more diverse and informative supervision. A detailed caption perturbation template is provided in the Appendix. The tem", + "bbox": [ + 88, + 779, + 482, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "perature factor $\\tau$ in Eq. 3 is set to 0.01. For optimization, we employ AdamW [43] with a learning rate of 2e-7 for pre-trained CLIP encoders and 4e-6 for the newly introduced components. Models are trained for 8,000 steps with a batch size of 2,048 on 8 A100 GPUs.", + "bbox": [ + 511, + 474, + 906, + 550 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Model setting. We train SpatialCLIP from OpenAI CLIP-B/16@224px and OpenAI CLIP-L/14@336px, leading to SpatialCLIP-B and SpatialCLIP-L, respectively. We utilize the lightweight Depth Anything v2 ViT-Small as our monocular depth estimation model, which brings an extra 24.8M parameters. The kernel size of the 3D CNN in Eq. 1 is 3, and the $D_w$ is 6. The newly added 3D CNN and 3D relative position encoding are both zero-initialized and introduce only nearly 1M extra parameters. We develop SpatialLLaVA based on SpatialCLIP-L, following LLaVA-1.5.", + "bbox": [ + 511, + 566, + 908, + 719 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6.2. Comparison for CLIP", + "text_level": 1, + "bbox": [ + 511, + 726, + 720, + 742 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Results on spatial-focused benchmark. To assess the spatial perception ability for CLIP models, we employ our newly proposed SpatialBench and modify the \"spatial relation\" subset of BLINK to the caption-matching task, as another spatial-focused benchmark. Results on these benchmarks are presented in Tab. 1. There are two key findings: 1) In both the CLIP-B and -L variants, SpatialCLIP demonstrates significant performance advantages over the original CLIP, indicating that SpatialCLIP more effectively captures spatial information in images and distinguishes between", + "bbox": [ + 511, + 750, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "29661", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/7cb0b2214567b453bb2c8385ca2ca038b5a9a7416420443838b8cd938ee37cb8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 96, + 90, + 295, + 231 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "LLaVA: The banana nut bread is on the left side of the picture, with a banana and an egg on the right side. The banana is placed in a blue bowl next to the bread. The eggs are in a glass bowl on the left side of the picture. The banana nut bread is placed in a brown paper bag, with a nut on top.", + "bbox": [ + 305, + 95, + 890, + 157 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "SpatialLLaVA: In the image, a bag of banana nut bread mix is placed next to a banana, an egg, and some nuts. The banana is positioned on the right side of the image, while the egg is located on the left side, and the nuts are scattered around the bag of mix. The bag of mix is placed in the middle of the scene, with the banana, egg, and nuts surrounding it.", + "bbox": [ + 305, + 169, + 893, + 229 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "LLaVA: There are two cars that are oriented in a vertical position. The green bus is on the left side of the picture, while the red bus is on the right side. The front of the green bus is facing toward the viewer, while the red bus is facing away from the viewer.", + "bbox": [ + 99, + 250, + 689, + 295 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "SpatialLLaVA: In the image, the two buses are positioned side by side, with one bus on the left and the other on the right. They are both facing the same direction, which suggests that they are either parked next to each other or driving in the same direction on the road.", + "bbox": [ + 99, + 306, + 689, + 354 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/96df64351e5387fae6b5a8d2baa45d329fdc6ce5053bda13234664619b26d242.jpg", + "image_caption": [ + "Figure 4. Qualitative comparison between LLaVA-1.5 and SpatialLLaVA. We prompt these two MLLMs to provide detailed descriptions of spatial relationships in images. Our model perform better at capturing these spatial concepts." + ], + "image_footnote": [], + "bbox": [ + 702, + 244, + 901, + 357 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "different spatial concepts. 2) For both the original CLIP and SpatialCLIP, performance on spatial-focused benchmarks is not correlated with model size. We conjecture this situation is derived from the scarcity of the spatial concepts in the current CLIP pre-training dataset. Increasing the parameter (i.e., enhancing learning capacity) does not necessarily lead to improved spatial perception. This highlights the value of additional designs for better spatial perception, rather than merely scaling up within the existing paradigm.", + "bbox": [ + 89, + 405, + 483, + 541 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results on general-purposed benchmark. In addition to enhancing spatial concept comprehension, accurate recognition of object categories and attributes is also essential for advanced applications. Since SpatialCLIP is built upon pretrained CLIP and is further trained with high-quality images and informative captions generated by diverse MLLMs, we also observe significant performance improvements on general-purpose benchmarks, like COCO and Flickr30K for zero-shot image-text retrieval.", + "bbox": [ + 89, + 560, + 483, + 696 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6.3. Comparison for MLLM", + "text_level": 1, + "bbox": [ + 89, + 709, + 310, + 724 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results on spatial-focused benchmark. In Tab. 2, we comprehensively validate the spatial perception and reasoning ability of our SpatialLLaVA on various spatial-focused MLLM benchmarks. With the same training data, SpatialLLaVA achieves much better results than LLaVA-1.5. SpatialLLaVA exhibits a stronger capability to differentiate between nuanced spatial concepts across indoor and outdoor scenes on our SpatialBench.", + "bbox": [ + 89, + 732, + 482, + 853 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We also conduct absolute performance comparison with more advanced MLLMs, including Spatial-RGPT (trained on spatial-focused data) and LLaVA-1.6/LLaVA-Interleave", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "(trained on much broader general-purpose data). Despite this \"unfair\" comparison, Spatial-LLaVA consistently shows superior spatial perception. Interestingly, Spatial-RGPT excels in tasks it was explicitly trained for (e.g., depth and distance estimation) but underperforms on tasks beyond its training scope (e.g., spatial question-answering). This observation highlights the need to enhance visual perception capabilities in MLLMs beyond simply expanding training data.", + "bbox": [ + 511, + 406, + 906, + 541 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The advantages on BLINK underscore the importance of capturing spatial structure for high-level applications. For \"visual corresponding\" task, capturing spatial structure is beneficial for identifying the same scene points across different viewpoints, lighting, or times, as the scene structure often remains consistent despite these changes. For \"Jigsaw,\" viewing 2D images in the 3D perspective allows the model to inspect the continuity of the 3D spatial structure rather than only relying on color and texture, aiding in organizing image patches.", + "bbox": [ + 511, + 542, + 908, + 694 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "CV-Bench[57] collects and annotates samples from the 3D object detection dataset Omni3D [4], focusing on 3D understanding. Employing SpatialCLIP also leads to performance improvements for MLLM on this benchmark, further demonstrating the superiority of SpatialCLIP's visual representations in capturing 3D spatial relationships.", + "bbox": [ + 511, + 696, + 908, + 787 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results on general-purposed benchmark. In Tab. 3, we examine how enhancing spatial perception in visual representations affect the performance of MLLMs in general-purposed scenarios. Across five benchmarks, Spatial-LLaVA achieves comparable performance to LLaVA-1.5 for general-purposed questions.", + "bbox": [ + 511, + 810, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "29662", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/33384e405c54da06f1d40eac2945c0c23c96c7d1b1c7938af0975268be10f10c.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
RowTuning3D CNN3D RPEPerturbSpatialBenchBLINKCOCOFlickr30k
OutDoor.InDoor.Spatial.R@1R@1
1×---35.2231.4351.0542.1934.40
2✓×××35.8531.4351.7548.9939.06
3✓✓××38.3631.4353.8549.0437.76
4✓×✓×37.1032.1454.5547.8139.33
5✓✓✓×38.3633.5755.2447.6737.56
6✓✓✓✓41.5134.2857.3448.3438.86
", + "bbox": [ + 143, + 88, + 849, + 237 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/79d1e5349eb5c637e816b30244308893028cf6821106e1683faf46b19f5e29b4.jpg", + "table_caption": [ + "Table 4. Ablation study of our \"Visual Model\" and \"Language Supervision\". All the experiments are conducted on SpatialCLIP-B." + ], + "table_footnote": [], + "table_body": "
SettingSpatialBenchBLINKCOCOFlickr30k
OutDoorInDoorSpatialR@1R@1
OpenAI CLIP-B35.2231.4351.0542.1934.40
Early Fusion39.6232.8551.7543.0133.09
Late Fusion40.2534.2852.4542.3831.80
3D-Inspried ViT41.5134.2857.3448.3438.86
", + "bbox": [ + 94, + 277, + 482, + 377 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. Ablation study of depth channel integration strategies.", + "bbox": [ + 98, + 380, + 472, + 393 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Qualitative comparison. To better demonstrate the improvement of SpatialLLaVA in spatial perception, we provide visualizations of both models in Fig. 2 and 4. The qualitative comparison of their responses reveals that SpatialLLaVA can more accurately and comprehensively capture the spatial relationships between objects in images.", + "bbox": [ + 89, + 421, + 482, + 513 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6.4. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 523, + 243, + 540 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of recaption. In the first two rows of Tab. 4, we present the comparison between the original CLIP and its direct fine-tuning using our recaption dataset. On the general-purposed benchmarks, COCO and Flickr30k, tuning on our dataset leads to significant performance improvements, underscoring the benefits of high information density and semantic alignment in recaption data. However, simply using the recaption data only brings negligible changes to the spatial-focused benchmarks, motivating us to introduce more specific designs to learn spatial concepts.", + "bbox": [ + 89, + 547, + 482, + 699 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of 3D-inspired ViT. We ablate the designs of 3D-Inspired ViT in Row 2-5 of Tab. 4. When fine-tuning on the same recaption data, incorporating 3D convolution (Row 3) or 3D relative position encoding (Row 4) shows improved performance on spatial-focused benchmarks compared to the standard ViT (Row 2). The combination of both techniques results in even greater enhancements. The improvement in spatial perception gained from the 3D-inspired design underscores the importance of integrating 3D point cloud structures into 2D ViT. Encoding images in 3D space with 3D-specific structure enables the visual model to capture complex 3D spatial information more effectively.", + "bbox": [ + 89, + 719, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effect of caption perturbation. The comparison between Rows 5 and 6 in Tab. 4 illustrates the benefits of caption perturbation, which prompts the model to capture spatial concepts and enhances its spatial perception ability. Since MLLMs themselves often struggle in perceiving and reasoning about spatial relationships, the captions generated by them can not fully capture all spatial details. By perturbing the spatial phrases, we create references that are fully incorrect in spatial relationships. This approach drives the model to learn spatial concepts by distinguishing the partially correct captions from the fully incorrect ones.", + "bbox": [ + 511, + 280, + 906, + 448 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Different depth channel integration strategies. In addition to our 3D-Inspired ViT, there are other alternative methods to enable RGB-D inputs. In Tab. 5, we discuss the commonly used early and late fusion schemes, which project the depth map into RGB features before and after the visual encoder, which are used by AlphaCLIP [55] and SpatialRGPT [12], respectively. In most benchmarks, simply projecting the depth channel to the pre-trained RGB features is hard to improve the performance. Considering the additional information provided by the reception data, these schemes may even have a negative impact. Conversely, our 3D-inspired ViT makes effective use of the extra depth information, yielding significant improvements.", + "bbox": [ + 511, + 476, + 908, + 674 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 691, + 633, + 708 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We present SpatialCLIP to overcome the shortcomings of original CLIP models in spatial understanding. By introducing improved \"Visual Model\" (3D-Inspired ViT) and enhanced \"Language Supervision\" (Recap tion & Perturbation), our model achieves significant improvements, not just on spatial-focused benchmarks like our proposed Spatial-Bench and modified BLINK, but also on zero-shot image-text retrieval. Additionally, we develop SpatialLLaVA based on SpatialCLIP, adhering to the LLaVA-1.5 training protocol. This integration of visual representations specifically enhanced for spatial recognition improves the spatial intelligence of MLLM.", + "bbox": [ + 511, + 719, + 908, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "29663", + "bbox": [ + 478, + 944, + 517, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 91, + 90, + 258, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported in part by National Natural Science Foundation of China under Grant No.624B2128 and No.62222211.", + "bbox": [ + 89, + 114, + 483, + 160 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 176, + 187, + 191 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C Lawrence Zitnick, and Devi Parikh. Vqa: Visual question answering. In ICCV, 2015. 3", + "[2] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond. arXiv:2308.12966, 2023. 1", + "[3] James Betker, Gabriel Goh, Li Jing, Tim Brooks, Jianfeng Wang, Linjie Li, Long Ouyang, Juntang Zhuang, Joyce Lee, Yufei Guo, et al. Improving image generation with better captions. https://cdn.openai.com/papers/dall-e-3.pdf, 2023.5", + "[4] Garrick Brazil, Abhinav Kumar, Julian Straub, Nikhila Ravi, Justin Johnson, and Georgia Gkioxari. Omni3d: A large benchmark and model for 3d object detection in the wild. In CVPR, 2023. 7", + "[5] Soravit Changpinyo, Piyush Sharma, Nan Ding, and Radu Soricut. Conceptual 12m: Pushing web-scale image-text pretraining to recognize long-tail visual concepts. In CVPR, 2021. 2, 5", + "[6] Agneet Chatterjee, Gabriela Ben Melech Stan, Estelle Aflalo, Sayak Paul, Dhruba Ghosh, Tejas Gokhale, Ludwig Schmidt, Hannaneh Hajishirzi, Vasudev Lal, Chitta Baral, et al. Getting it right: Improving spatial consistency in text-to-image models. In ECCV, 2025. 5, 6", + "[7] Boyuan Chen, Zhuo Xu, Sean Kirmani, Brain Ichter, Dorsa Sadigh, Leonidas Guibas, and Fei Xia. Spatialvlm: Endowing vision-language models with spatial reasoning capabilities. In CVPR, 2024. 2, 3, 5", + "[8] Dave Zhenyu Chen, Angel X Chang, and Matthias Nießner. Scanrefer: 3d object localization in rgb-d scans using natural language. In ECCV, 2020. 3", + "[9] Lin Chen, Jinsong Li, Xiaoyi Dong, Pan Zhang, Conghui He, Jiaqi Wang, Feng Zhao, and Dahua Lin. Sharegpt4v: Improving large multi-modal models with better captions. arXiv:2311.12793, 2023. 2, 6", + "[10] Zhe Chen, Weiyun Wang, Hao Tian, Shenglong Ye, Zhang-wei Gao, Erfei Cui, Wenwen Tong, Kongzhi Hu, Jiapeng Luo, Zheng Ma, et al. How far are we to gpt-4v? closing the gap to commercial multimodal models with open-source suites. arXiv:2404.16821, 2024. 1, 2", + "[11] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. In CVPR, 2024. 1, 3", + "[12] An-Chieh Cheng, Hongxu Yin, Yang Fu, Qiushan Guo, Ruihan Yang, Jan Kautz, Xiaolong Wang, and Sifei Liu. Spatial-" + ], + "bbox": [ + 93, + 200, + 483, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "rgpt: Grounded spatial reasoning in vision language model. arXiv:2406.01584, 2024. 2, 3, 4, 5, 8", + "[13] Mehdi Cherti, Romain Beaumont, Ross Wightman, Mitchell Wortsman, Gabriel Ilharco, Cade Gordon, Christoph Schuhmann, Ludwig Schmidt, and Jenia Jitsev. Reproducible scaling laws for contrastive language-image learning. In CVPR, 2023. 1, 2", + "[14] Jang Hyun Cho, Boris Ivanovic, Yulong Cao, Edward Schmerling, Yue Wang, Xinshuo Weng, Boyi Li, Yurong You, Philipp Krahenbuhl, Yan Wang, et al. Language-image models with 3d understanding. arXiv:2405.03685, 2024. 2, 3", + "[15] Jaeyoung Chung, Jeongtaek Oh, and Kyoung Mu Lee. Depth-regularized optimization for 3d gaussian splatting in few-shot images. In CVPR, 2024. 4", + "[16] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In CVPR, 2017. 5", + "[17] Kangle Deng, Andrew Liu, Jun-Yan Zhu, and Deva Ramanan. Depth-supervised nerf: Fewer views and faster training for free. In CVPR, 2022. 4", + "[18] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv:2010.11929, 2020.1", + "[19] Danny Driess, Fei Xia, Mehdi SM Sajjadi, Corey Lynch, Aakanksha Chowdhery, Brian Ichter, Ayzaan Wahid, Jonathan Tompson, Quan Vuong, Tianhe Yu, et al. Palm-e: An embodied multimodal language model. arXiv:2303.03378, 2023. 3", + "[20] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv:2407.21783, 2024. 3", + "[21] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In ICML, 2024. 5", + "[22] Alex Fang, Albin Madappally Jose, Amit Jain, Ludwig Schmidt, Alexander Toshev, and Vaishaal Shankar. Data filtering networks. arXiv:2309.17425, 2023. 1, 2", + "[23] Chaoyou Fu, Yuhan Dai, Yondong Luo, Lei Li, Shuhuai Ren, Renrui Zhang, Zihan Wang, Chenyu Zhou, Yunhang Shen, Mengdan Zhang, et al. Video-mme: The first-ever comprehensive evaluation benchmark of multi-modal llms in video analysis. arXiv:2405.21075, 2024. 3", + "[24] Xingyu Fu, Yushi Hu, Bangzheng Li, Yu Feng, Haoyu Wang, Xudong Lin, Dan Roth, Noah A Smith, Wei-Chiu Ma, and Ranjay Krishna. Blink: Multimodal large language models can see but not perceive. arXiv:2404.12390, 2024. 2", + "[25] Jensen Gao, Bidipta Sarkar, Fei Xia, Ted Xiao, Jiajun Wu, Brian Ichter, Anirudha Majumdar, and Dorsa Sadigh. Physically grounded vision-language models for robotic manipulation. In ICRA, 2024. 3", + "[26] Rohit Girdhar, Alaaeldin El-Nouby, Zhuang Liu, Mannat Singh, Kalyan Vasudev Alwala, Armand Joulin, and Ishan" + ], + "bbox": [ + 516, + 92, + 903, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "29664", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Misra. Imagebind: One embedding space to bind them all. In CVPR, 2023. 1", + "[27] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Bartra, and Devi Parikh. Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In CVPR, 2017. 6", + "[28] Benjamin Graham, Martin Engelcke, and Laurens Van Der Maaten. 3d semantic segmentation with submanifold sparse convolutional networks. In CVPR, 2018. 1", + "[29] Drew A Hudson and Christopher D Manning. Gqa: A new dataset for real-world visual reasoning and compositional question answering. In CVPR, 2019. 6", + "[30] Sahar Kazemzadeh, Vicente Ordonez, Mark Matten, and Tamara Berg. Referitag: Referring to objects in photographs of natural scenes. In EMNLP, 2014. 3", + "[31] Bingxin Ke, Anton Obukhov, Shengyu Huang, Nando Metzger, Rodrigo Caye Daudt, and Konrad Schindler. Repurposing diffusion-based image generators for monocular depth estimation. In CVPR, 2024. 4", + "[32] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In ICCV, 2023. 6", + "[33] Mikhail Konenkov, Artem Lykov, Daria Trinitatova, and Dzmitry Tsetserukou. Vr-gpt: Visual language model for intelligent virtual reality applications. arXiv:2405.11537, 2024.3", + "[34] Xin Lai, Jianhui Liu, Li Jiang, Liwei Wang, Hengshuang Zhao, Shu Liu, Xiaojuan Qi, and Jiaya Jia. Stratified transformer for 3d point cloud segmentation. In CVPR, 2022. 1", + "[35] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. Llava-onevision: Easy visual task transfer. arXiv:2408.03326, 2024. 2", + "[36] Liunian Harold Li, Pengchuan Zhang, Haotian Zhang, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, et al. Grounded language-image pre-training. In CVPR, 2022. 1", + "[37] Xianhang Li, Haoqin Tu, Mude Hui, Zeyu Wang, Bingchen Zhao, Junfei Xiao, Sucheng Ren, Jieru Mei, Qing Liu, Huangjie Zheng, et al. What if we recapture billions of web images with llama-3? arXiv:2406.08478, 2024. 5", + "[38] Yifan Li, Yifan Du, Kun Zhou, Jinpeng Wang, Wayne Xin Zhao, and Ji-Rong Wen. Evaluating object hallucination in large vision-language models. arXiv:2305.10355, 2023. 6", + "[39] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 2, 3, 5", + "[40] Ziyi Lin, Shijie Geng, Renrui Zhang, Peng Gao, Gerard De Melo, Xiaogang Wang, Jifeng Dai, Yu Qiao, and Hongsheng Li. Frozen clip models are efficient video learners. In ECCV, 2022. 1, 2", + "[41] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In CVPR, 2024. 1, 2, 3, 5, 6", + "[42] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, 2024. 1, 2, 3" + ], + "bbox": [ + 91, + 90, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[43] I Loshchilov. Decoupled weight decay regularization. arXiv:1711.05101, 2017. 6", + "[44] Soroush Nasiriany, Fei Xia, Wenhao Yu, Ted Xiao, Jacky Liang, Ishita Dasgupta, Annie Xie, Danny Driess, Ayzaan Wahid, Zhuo Xu, et al. Pivot: Iterative visual prompting elicits actionable knowledge for vlms. arXiv:2402.07872, 2024.3", + "[45] Bryan A Plummer, Liwei Wang, Chris M Cervantes, Juan C Caicedo, Julia Hockenmaier, and Svetlana Lazebnik. Flickr30k entities: Collecting region-to-phrase correspondences for richer image-to-sentence models. In ICCV, 2015. 3", + "[46] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv:2307.01952, 2023. 1", + "[47] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 1, 2", + "[48] Yongming Rao, Wenliang Zhao, Guangyi Chen, Yansong Tang, Zheng Zhu, Guan Huang, Jie Zhou, and Jiwen Lu. Denseclip: Language-guided dense prediction with context-aware prompting. In CVPR, 2022. 2", + "[49] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 1", + "[50] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv:2111.02114, 2021. 2, 5", + "[51] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. In NeurIPS, 2022. 2, 5", + "[52] Sanjay Subramanian, William Merrill, Trevor Darrell, Matt Gardner, Sameer Singh, and Anna Rohrbach. Reclip: A strong zero-shot baseline for referring expression comprehension. arXiv:2204.05991, 2022. 1, 2, 3", + "[53] Quan Sun, Yuxin Fang, Ledell Wu, Xinlong Wang, and Yue Cao. Eva-clip: Improved training techniques for clip at scale. arXiv:2303.15389, 2023. 1, 2", + "[54] Quan Sun, Yufeng Cui, Xiaosong Zhang, Fan Zhang, Qiying Yu, Yueze Wang, Yongming Rao, Jingjing Liu, Tiejun Huang, and Xinlong Wang. Generative multimodal models are in-context learners. In CVPR, 2024. 2", + "[55] Zeyi Sun, Ye Fang, Tong Wu, Pan Zhang, Yuhang Zang, Shu Kong, Yuanjun Xiong, Dahua Lin, and Jiaqi Wang. Alphaclip: A clip model focusing on wherever you want. In CVPR, 2024. 4, 8", + "[56] Kaihua Tang, Hanwang Zhang, Baoyuan Wu, Wenhan Luo, and Wei Liu. Learning to compose dynamic tree structures for visual contexts. In CVPR, 2019. 3" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "29665", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[57] Shengbang Tong, Ellis Brown, Penghao Wu, Sanghyun Woo, Manoj Middepogu, Sai Charitha Akula, Jihan Yang, Shusheng Yang, Adithya Iyer, Xichen Pan, et al. Cambrian1: A fully open, vision-centric exploration of multimodal llms. arXiv:2406.16860, 2024. 1, 5, 7", + "[58] Shengbang Tong, Zhuang Liu, Yuexiang Zhai, Yi Ma, Yann LeCun, and Saining Xie. Eyes wide shut? exploring the visual shortcomings of multimodal llms. In CVPR, 2024. 2, 3, 5", + "[59] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv:2307.09288, 2023. 3", + "[60] Haoxiang Wang, Pavan Kumar Anasosalu Vasu, Fartash Faghri, Raviteja Vemulapalli, Mehrdad Farajtabar, Sachin Mehta, Mohammad Rastegari, Oncel Tuzel, and Hadi Pouransari. Sam-clip: Merging vision foundation models towards semantic and spatial understanding. In CVPR, 2024. 2", + "[61] Peng-Shuai Wang. Octformer: Octree-based transformers for 3d point clouds. TOG, 2023. 1, 4", + "[62] Xiaoyang Wu, Li Jiang, Peng-Shuai Wang, Zhijian Liu, Xihui Liu, Yu Qiao, Wanli Ouyang, Tong He, and Hengshuang Zhao. Point transformer v3: Simpler faster stronger. In CVPR, 2024. 1, 4", + "[63] xAI. grok, 2024. 6", + "[64] Lin Xu, Yilin Zhao, Daquan Zhou, Zhijie Lin, See Kiong Ng, and Jiashi Feng. Pllava: Parameter-free llava extension from images to videos for video dense captioning. arXiv:2404.16994, 2024. 2", + "[65] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. In CVPR, 2024. 4", + "[66] Lihe Yang, Bingyi Kang, Zilong Huang, Zhen Zhao, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything v2. arXiv:2406.09414, 2024. 4", + "[67] Yu-Qi Yang, Yu-Xiao Guo, Jian-Yu Xiong, Yang Liu, Hao Pan, Peng-Shuai Wang, Xin Tong, and Baining Guo. Swin3d: A pretrained transformer backbone for 3d indoor scene understanding. arXiv:2304.06906, 2023. 1, 4", + "[68] Yuan Yao, Tianyu Yu, Ao Zhang, Chongyi Wang, Junbo Cui, Hongji Zhu, Tianchi Cai, Haoyu Li, Weilin Zhao, Zhihui He, et al. Minicpm-v: A gpt-4v level mllm on your phone. arXiv:2408.01800, 2024. 2", + "[69] Peter Young, Alice Lai, Micah Hodosh, and Julia Hockenmaier. From image descriptions to visual denotations: New similarity metrics for semantic inference over event descriptions. TACL, 2014. 2", + "[70] Bangguo Yu, Hamidreza Kasaei, and Ming Cao. L3mvn: Leveraging large language models for visual target navigation. In IROS, 2023. 3", + "[71] Jiahui Yu, Zirui Wang, Vijay Vasudevan, Legg Yeung, Mojtaba Seyedhosseini, and Yonghui Wu. Coca: Contrastive captioners are image-text foundation models. arXiv:2205.01917, 2022. 6" + ], + "bbox": [ + 91, + 90, + 480, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[72] Weihao Yu, Zhengyuan Yang, Linjie Li, Jianfeng Wang, Kevin Lin, Zicheng Liu, Xinchao Wang, and Lijuan Wang. Mm-vet: Evaluating large multimodal models for integrated capabilities. arXiv:2308.02490, 2023. 6", + "[73] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In ICCV, 2023. 1, 3", + "[74] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In CVPR, 2023. 4", + "[75] Gengze Zhou, Yicong Hong, and Qi Wu. Navgpt: Explicit reasoning in vision-and-language navigation with large language models. In AAAI, 2024. 3" + ], + "bbox": [ + 514, + 90, + 903, + 275 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "29666", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/SpatialCLIP_ Learning 3D-aware Image Representations from Spatially Discriminative Language/dad5650f-d216-4b30-a0c7-dd7ca6658593_model.json b/2025/SpatialCLIP_ Learning 3D-aware Image Representations from Spatially Discriminative Language/dad5650f-d216-4b30-a0c7-dd7ca6658593_model.json new file mode 100644 index 0000000000000000000000000000000000000000..3b14f294c6f757f318a5afab37c269c0356361f6 --- /dev/null +++ b/2025/SpatialCLIP_ Learning 3D-aware Image Representations from Spatially Discriminative Language/dad5650f-d216-4b30-a0c7-dd7ca6658593_model.json @@ -0,0 +1,3192 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.211, + 0.131, + 0.787, + 0.176 + ], + "angle": 0, + "content": "SpatialCLIP: Learning 3D-aware Image Representations from Spatially Discriminative Language" + }, + { + "type": "text", + "bbox": [ + 0.155, + 0.192, + 0.847, + 0.228 + ], + "angle": 0, + "content": "Zehan Wang\\(^{1*}\\), Sahuai Zhou\\(^{1*}\\), Shaoxuan He\\(^{1}\\), Haifeng Huang\\(^{1}\\), Lihe Yang\\(^{2}\\), Ziang Zhang\\(^{1}\\), Xize Cheng\\(^{1}\\), Shengpeng Ji\\(^{1}\\), Tao Jin\\(^{1}\\), Hengshuang Zhao\\(^{2}\\), Zhou Zhao\\(^{1\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.229, + 0.234, + 0.761, + 0.27 + ], + "angle": 0, + "content": "1Zhejiang University; 2The University of Hong Kong https://github.com/SpatialVision/Spatial-CLIP" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.294, + 0.327, + 0.31 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.326, + 0.486, + 0.782 + ], + "angle": 0, + "content": "Contrastive Language-Image Pre-training (CLIP) learns robust visual models through language supervision, making it a crucial visual encoding technique for various applications. However, CLIP struggles with comprehending spatial concepts in images, potentially restricting the spatial intelligence of CLIP-based AI systems. In this work, we propose SpatialCLIP, an enhanced version of CLIP with better spatial understanding capabilities. To capture the intricate 3D spatial relationships in images, we improve both \"visual model\" and \"language supervision\" of CLIP. Specifically, we design 3D-inspired ViT to replace the standard ViT in CLIP. By lifting 2D image tokens into 3D space and incorporating design insights from point cloud networks, our visual model gains greater potential for spatial perception. Meanwhile, captions with accurate and detailed spatial information are very rare. To explore better language supervision for spatial understanding, we re-caption images and perturb their spatial phrases as negative descriptions, which compels the visual model to seek spatial cues to distinguish these hard negative captions. With the enhanced visual model, we introduce SpatialLLaVA, following the same LLaVA-1.5 training protocol, to investigate the importance of visual representations for MLLM's spatial intelligence. Furthermore, we create SpatialBench, a benchmark specifically designed to evaluate CLIP and MLLM in spatial reasoning. SpatialCLIP and SpatialLLaVA achieve substantial performance improvements, demonstrating stronger capabilities in spatial perception and reasoning, while maintaining comparable results on general-purpose benchmarks." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.809, + 0.222, + 0.824 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.834, + 0.483, + 0.866 + ], + "angle": 0, + "content": "Recently, Contrastive Language-Image Pretraining (CLIP) models [13, 22, 47, 53, 73] have demonstrated the ability to" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.296, + 0.905, + 0.37 + ], + "angle": 0, + "content": "learn robust image representations from language supervision. These models serve as foundational visual encoders for various applications, such as open-world visual recognition [26, 36, 40, 46, 49, 52] and Multi-modal Large Language Models (MLLM) [2, 10, 11, 41, 42, 57]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.371, + 0.906, + 0.583 + ], + "angle": 0, + "content": "Although CLIP exhibits strong zero-shot ability across diverse tasks, it struggles with understanding spatial arrangement within images. As shown in Fig. 1, for a correctly matched image-text pair, we introduce perturbations to the caption by modifying the object category, object attribute, and spatial phrases. The results reveal that altering the object category or attribute significantly reduces the image-text similarity in the CLIP space. However, CLIP struggles to distinguish incorrect spatial relationship descriptions and even assigns a higher matching score to captions with wrong spatial descriptions. Spatial intelligence is a key aspect of visual perception, and limitations in spatial understanding may constrain systems built on CLIP from effectively understanding real-world environments." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.583, + 0.906, + 0.674 + ], + "angle": 0, + "content": "In this paper, we present SpatialCLIP, a CLIP model with enhanced spatial perception capabilities. As titled in [47], CLIP models are \"Learning Transferable Visual Models from Natural Language Supervision\". Our SpatialCLIP advances the spatial understanding from both \"visual models\" and \"language supervision\" perspectives." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.674, + 0.906, + 0.885 + ], + "angle": 0, + "content": "For \"visual models,\" we propose a 3D-inspired ViT that processes RGB-D input rather than RGB images. By incorporating depth information, our visual model gains the ability to perceive images from the 3D perspective. This design helps the model understand complex 3D spatial relationships like \"behind,\" \"in front of,\" \"under,\" and \"far.\" Specifically, we lift the 2D tokens into 3D space using the extra depth information. In the 3D space, we integrate two structures widely used in point cloud networks: 3D convolution [28, 61] and 3D relative position encoding [34, 62, 67], into the pre-trained ViT model [18]. By integrating design insights from 3D point cloud processing, our 3D-inspired ViT demonstrates enhanced potential for capturing complex 3D spatial structures in images." + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.886, + 0.905, + 0.902 + ], + "angle": 0, + "content": "For \"language supervision\", in current vision-language" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.875, + 0.223, + 0.888 + ], + "angle": 0, + "content": "*Equal Contribution." + }, + { + "type": "page_footnote", + "bbox": [ + 0.111, + 0.888, + 0.236, + 0.9 + ], + "angle": 0, + "content": "† Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "29656" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.102, + 0.09, + 0.292, + 0.182 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.104, + 0.184, + 0.258, + 0.213 + ], + "angle": 0, + "content": "- Accurate: The darker-haired cat is behind the white cat. \nCLIP Score: 22.28" + }, + { + "type": "image_footnote", + "bbox": [ + 0.1, + 0.218, + 0.282, + 0.249 + ], + "angle": 0, + "content": "- Category Error: The darker-haired bird is behind the white cat. \nCLIP Score: 20.43 (-1.85)" + }, + { + "type": "image_footnote", + "bbox": [ + 0.098, + 0.253, + 0.28, + 0.283 + ], + "angle": 0, + "content": "\\(\\nLeftrightarrow\\) Attribute Error: The light-haired cat is behind the white cat. CLIP Score: 21.22 (-1.06)" + }, + { + "type": "image_footnote", + "bbox": [ + 0.102, + 0.289, + 0.276, + 0.319 + ], + "angle": 0, + "content": "Spatial Error: The darker-haired cat is to the right of the white cat. \nCLIP Score: 23.66 (+1.38)" + }, + { + "type": "list", + "bbox": [ + 0.098, + 0.09, + 0.292, + 0.319 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.303, + 0.09, + 0.495, + 0.182 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.303, + 0.184, + 0.488, + 0.213 + ], + "angle": 0, + "content": "-Accurate: The white refrigerator is closer to the camera than brown door. \nCLIP Score: 29.74" + }, + { + "type": "image_footnote", + "bbox": [ + 0.297, + 0.219, + 0.473, + 0.249 + ], + "angle": 0, + "content": "\\(\\nrightarrow\\) Category Error: The white cabinet is closer to the camera brown door. CLIP Score: 25.66 (-4.08)" + }, + { + "type": "image_footnote", + "bbox": [ + 0.297, + 0.254, + 0.486, + 0.284 + ], + "angle": 0, + "content": "\\(\\nrightarrow\\) Attribute Error: The dark refrigerator is closer to the camera than opened door. CLIP Score: 29.39 (-0.35)" + }, + { + "type": "image_footnote", + "bbox": [ + 0.297, + 0.289, + 0.485, + 0.318 + ], + "angle": 0, + "content": "Spatial Error: The white refrigerator is further from the camera than brown door. CLIP Score: 29.62 (-0.12)" + }, + { + "type": "list", + "bbox": [ + 0.297, + 0.184, + 0.488, + 0.318 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.09, + 0.693, + 0.182 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.503, + 0.184, + 0.684, + 0.213 + ], + "angle": 0, + "content": "-Accurate: The dark coffee machine is to the left of the mug. CLIP Score: 22.10" + }, + { + "type": "image_footnote", + "bbox": [ + 0.5, + 0.219, + 0.679, + 0.249 + ], + "angle": 0, + "content": "\\(\\rightarrow\\) Category Error: The ceramic cup is to the left of the mug. CLIP Score: 19.48 (-2.62)" + }, + { + "type": "image_footnote", + "bbox": [ + 0.5, + 0.254, + 0.682, + 0.284 + ], + "angle": 0, + "content": "\\(\\nrightarrow\\) Attribute Error: The old broken coffee machine is to the left of the mug. CLIP Score: 21.27 (-0.83)" + }, + { + "type": "image_footnote", + "bbox": [ + 0.5, + 0.289, + 0.684, + 0.319 + ], + "angle": 0, + "content": "Spatial Error: The dark coffee machine is in front of the mug. CLIP Score: 22.53 (+0.43)" + }, + { + "type": "list", + "bbox": [ + 0.5, + 0.184, + 0.684, + 0.319 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.705, + 0.091, + 0.894, + 0.182 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.704, + 0.184, + 0.863, + 0.213 + ], + "angle": 0, + "content": "-Accurate: Two people are standing behind a pile of huge-sized pipes. CLIP Score: 26.62" + }, + { + "type": "image_footnote", + "bbox": [ + 0.703, + 0.219, + 0.894, + 0.249 + ], + "angle": 0, + "content": "\\(\\nvdash\\) Category Error: Two people are standing behind a pile of huge-sized papers. CLIP Score: 16.40 (-10.22)" + }, + { + "type": "image_footnote", + "bbox": [ + 0.703, + 0.254, + 0.893, + 0.284 + ], + "angle": 0, + "content": "\\(\\nvdash\\) Attribute Error: Two people are standing behind a pile of small pipes. \nCLIP Score: 25.78 (-0.84)" + }, + { + "type": "image_footnote", + "bbox": [ + 0.704, + 0.289, + 0.883, + 0.319 + ], + "angle": 0, + "content": "Spatial Error: Two people are standing inside a pile of huge-sized pipes. \nCLIP Score: 27.18 (+0.56)" + }, + { + "type": "list", + "bbox": [ + 0.703, + 0.184, + 0.894, + 0.319 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.321, + 0.907, + 0.363 + ], + "angle": 0, + "content": "Figure 1. The preliminary analysis of OpenAI CLIP-L/14@336px. CLIP Score stands for the cosine similarity between image and language description. These examples demonstrate that while CLIP is highly sensitive to errors involving object categories and attributes, it struggles to accurately identify mistakes in spatial descriptions." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.369, + 0.483, + 0.536 + ], + "angle": 0, + "content": "datasets [5, 50, 51] for CLIP training, the input captions are primarily simple summary texts that only describe the primary object category and attributes. The absence of precise spatial information in these captions may constrain the visual model's ability to capture spatial relationships. To enhance language supervision, we recapture images using advanced MLLMs [9, 41] and introduce perturbed spatial phrases to captions as hard negative queries. Training the model to differentiate between captions that share the same object categories and attributes but vary in spatial phrases forces it to better capture spatial cues in images." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.537, + 0.483, + 0.702 + ], + "angle": 0, + "content": "Furthermore, we explore whether SpatialCLIP, with its enhanced spatial perception, can contribute stronger spatial intelligence to MLLMs when using the same data. Recent work [7, 12, 14] mainly enhances the spatial perception capabilities of MLLMs by introducing new training data, while lacking the exploration of spatial-focused visual representations. However, as discussed in [58], visual representation is also crucial to the capabilities of MLLMs. We construct SpatialLLaVA using the same training recipe as the standard LLaVA-1.5 [41], enabling a fair and systematical analysis over MLLMs." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.704, + 0.483, + 0.809 + ], + "angle": 0, + "content": "To quantitatively assess the spatial understanding capabilities of CLIP and MLLM, we introduce SpatialBench, a benchmark designed to evaluate spatial comprehension within images. This benchmark challenges models to comprehend spatial arrangements and perform complex spatial reasoning. For thorough evaluation, we curate and annotate images from both outdoor and indoor scenes." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Experimentally, SpatialCLIP shows substantial improvements across spatial-focused benchmarks, such as Spatial-Bench and BLINK [24], and performs well on general-purpose downstream tasks [39, 69]. Similarly, SpatialLLaVA demonstrates notably enhanced spatial perception while retaining strong general perceptual abilities." + }, + { + "type": "text", + "bbox": [ + 0.531, + 0.37, + 0.789, + 0.383 + ], + "angle": 0, + "content": "Our main contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.385, + 0.906, + 0.43 + ], + "angle": 0, + "content": "- We propose SpatialCLIP, which improves the CLIP's spatial understanding capabilities by refining both \"visual model\" and \"language supervision\"." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.431, + 0.905, + 0.505 + ], + "angle": 0, + "content": "- We develop SpatialLLaVA, by replacing the visual encoding method in LLaVA with our SpatialCLIP while preserving other training protocols. This investigates the value of SpatialCLIP in enhancing the spatial intelligence of CLIP-based MLLM, as shown in Fig. 2." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.507, + 0.905, + 0.55 + ], + "angle": 0, + "content": "- We construct SpatialBench to evaluate the spatial capabilities of CLIP and MLLMs, collecting and annotating samples from both outdoor and indoor scenarios." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.551, + 0.905, + 0.612 + ], + "angle": 0, + "content": "- We conduct extensive experiments on both spatial-focused and general-purpose benchmarks to validate the effectiveness of SpatialCLIP, SpatialLLaVA, and each newly proposed component." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.385, + 0.906, + 0.612 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.626, + 0.655, + 0.641 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.651, + 0.873, + 0.668 + ], + "angle": 0, + "content": "2.1. Contrastive Language-Image Pre-training" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.674, + 0.907, + 0.901 + ], + "angle": 0, + "content": "CLIP models [13, 22, 47, 53] are trained on web-scale image-text paired data [5, 50, 51] using multimodal contrastive learning as objective. By learning cross-modal semantic alignment from massive datasets, CLIP demonstrates strong zero-shot generalization capabilities and is widely used in various vision-language downstream tasks, such as image object detection [52], segmentation [48, 60], and video action recognition [40]. In recent multimodal large language models (MLLM), CLIP is the most common visual encoding strategy, enabling large language models with the ability to understand images. LLaVA [42] and subsequent works [35, 35, 41, 64] primarily rely on OpenAI-CLIP [47], while other state-of-the-art MLLMs like Emu2 [54], InternVL1.5 [10] and MiniCPM-v [68] also utilize various CLIP variants, such as EVA-CLIP [53], In" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "29657" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.101, + 0.089, + 0.296, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.1, + 0.215, + 0.273, + 0.225 + ], + "angle": 0, + "content": "Select the best caption for the image:" + }, + { + "type": "image_caption", + "bbox": [ + 0.101, + 0.225, + 0.267, + 0.234 + ], + "angle": 0, + "content": "A. Yellow electric car is closer to the camera." + }, + { + "type": "image_caption", + "bbox": [ + 0.101, + 0.234, + 0.254, + 0.242 + ], + "angle": 0, + "content": "B. Red electric car is closer to the camera." + }, + { + "type": "image_caption", + "bbox": [ + 0.101, + 0.242, + 0.263, + 0.251 + ], + "angle": 0, + "content": "C. Two cars are equally close to the camera." + }, + { + "type": "image_caption", + "bbox": [ + 0.108, + 0.252, + 0.246, + 0.262 + ], + "angle": 0, + "content": "LLaVA-1.5: The best caption is B." + }, + { + "type": "image_caption", + "bbox": [ + 0.107, + 0.268, + 0.259, + 0.277 + ], + "angle": 0, + "content": "SpatialLLaVA: The best caption is A." + }, + { + "type": "image", + "bbox": [ + 0.101, + 0.28, + 0.296, + 0.403 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.101, + 0.404, + 0.273, + 0.414 + ], + "angle": 0, + "content": "Select the best caption for the image:" + }, + { + "type": "image_caption", + "bbox": [ + 0.101, + 0.414, + 0.22, + 0.422 + ], + "angle": 0, + "content": "A. The fan is facing the camera." + }, + { + "type": "image_caption", + "bbox": [ + 0.101, + 0.422, + 0.231, + 0.43 + ], + "angle": 0, + "content": "B. The fan is facing the white desk." + }, + { + "type": "image_caption", + "bbox": [ + 0.101, + 0.43, + 0.207, + 0.438 + ], + "angle": 0, + "content": "C. The fan is facing the sofa." + }, + { + "type": "image_caption", + "bbox": [ + 0.101, + 0.438, + 0.21, + 0.447 + ], + "angle": 0, + "content": "D. The fan is facing the chair." + }, + { + "type": "image_caption", + "bbox": [ + 0.108, + 0.449, + 0.246, + 0.459 + ], + "angle": 0, + "content": "LLaVA-1.5: The best caption is A." + }, + { + "type": "image_caption", + "bbox": [ + 0.107, + 0.464, + 0.26, + 0.474 + ], + "angle": 0, + "content": "SpatialLLaVA: The best caption is D." + }, + { + "type": "image", + "bbox": [ + 0.298, + 0.089, + 0.493, + 0.214 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.298, + 0.215, + 0.489, + 0.232 + ], + "angle": 0, + "content": "Select the best caption for the image: A. The bird standing taller is smaller than the other." + }, + { + "type": "image_caption", + "bbox": [ + 0.298, + 0.232, + 0.484, + 0.241 + ], + "angle": 0, + "content": "B. The bird standing taller is bigger than the other." + }, + { + "type": "image_caption", + "bbox": [ + 0.298, + 0.241, + 0.457, + 0.249 + ], + "angle": 0, + "content": "C. The two birds are roughly the same size." + }, + { + "type": "image_caption", + "bbox": [ + 0.304, + 0.251, + 0.442, + 0.262 + ], + "angle": 0, + "content": "LLaVA-1.5: The best caption is B." + }, + { + "type": "image_caption", + "bbox": [ + 0.304, + 0.267, + 0.456, + 0.277 + ], + "angle": 0, + "content": "SpatialLLaVA: The best caption is C." + }, + { + "type": "image", + "bbox": [ + 0.298, + 0.28, + 0.493, + 0.403 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.298, + 0.404, + 0.471, + 0.414 + ], + "angle": 0, + "content": "Select the best caption for the image:" + }, + { + "type": "image_caption", + "bbox": [ + 0.298, + 0.414, + 0.405, + 0.422 + ], + "angle": 0, + "content": "A. All the sofas form a circle" + }, + { + "type": "image_caption", + "bbox": [ + 0.298, + 0.422, + 0.42, + 0.431 + ], + "angle": 0, + "content": "B. All the sofas form a rectangle." + }, + { + "type": "image_caption", + "bbox": [ + 0.298, + 0.431, + 0.406, + 0.439 + ], + "angle": 0, + "content": "C. All the sofas form an oval." + }, + { + "type": "image_caption", + "bbox": [ + 0.298, + 0.44, + 0.43, + 0.447 + ], + "angle": 0, + "content": "D. All sofas are arranged randomly." + }, + { + "type": "image_caption", + "bbox": [ + 0.304, + 0.449, + 0.443, + 0.459 + ], + "angle": 0, + "content": "LLaVA-1.5: The best caption is A." + }, + { + "type": "image_caption", + "bbox": [ + 0.304, + 0.464, + 0.456, + 0.474 + ], + "angle": 0, + "content": "SpatialLLaVA: The best caption is B." + }, + { + "type": "image", + "bbox": [ + 0.495, + 0.089, + 0.689, + 0.214 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.495, + 0.215, + 0.668, + 0.225 + ], + "angle": 0, + "content": "Select the best caption for the image:" + }, + { + "type": "image_caption", + "bbox": [ + 0.495, + 0.225, + 0.627, + 0.234 + ], + "angle": 0, + "content": "A. The two trains are perpendicular." + }, + { + "type": "image_caption", + "bbox": [ + 0.495, + 0.234, + 0.62, + 0.241 + ], + "angle": 0, + "content": "B. The two trains are intersecting." + }, + { + "type": "image_caption", + "bbox": [ + 0.495, + 0.241, + 0.605, + 0.249 + ], + "angle": 0, + "content": "C. The two trains are parallel." + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.252, + 0.64, + 0.262 + ], + "angle": 0, + "content": "LLaVA-1.5: The best caption is A." + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.268, + 0.652, + 0.277 + ], + "angle": 0, + "content": "SpatialLLaVA: The best caption is C." + }, + { + "type": "image", + "bbox": [ + 0.493, + 0.28, + 0.689, + 0.403 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.495, + 0.404, + 0.679, + 0.414 + ], + "angle": 0, + "content": "Select the best caption for the image:" + }, + { + "type": "image_caption", + "bbox": [ + 0.495, + 0.414, + 0.68, + 0.423 + ], + "angle": 0, + "content": "A. Ceiling light hangs above the edge of the table." + }, + { + "type": "image_caption", + "bbox": [ + 0.495, + 0.423, + 0.682, + 0.431 + ], + "angle": 0, + "content": "B. Ceiling light hangs above the center of the table." + }, + { + "type": "image", + "bbox": [ + 0.692, + 0.089, + 0.885, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.692, + 0.215, + 0.865, + 0.225 + ], + "angle": 0, + "content": "Select the best caption for the image:" + }, + { + "type": "image_caption", + "bbox": [ + 0.692, + 0.225, + 0.832, + 0.233 + ], + "angle": 0, + "content": "A. The stop sign is taller than the wall." + }, + { + "type": "image_caption", + "bbox": [ + 0.692, + 0.233, + 0.839, + 0.241 + ], + "angle": 0, + "content": "B. The stop sign is shorter than the wall." + }, + { + "type": "image_caption", + "bbox": [ + 0.692, + 0.241, + 0.883, + 0.249 + ], + "angle": 0, + "content": "C. The stop sign and the wall are of the same height." + }, + { + "type": "image_caption", + "bbox": [ + 0.697, + 0.252, + 0.835, + 0.262 + ], + "angle": 0, + "content": "LLaVA-1.5: The best caption is C." + }, + { + "type": "image_caption", + "bbox": [ + 0.697, + 0.268, + 0.848, + 0.277 + ], + "angle": 0, + "content": "SpatialLLaVA: The best caption is A." + }, + { + "type": "image", + "bbox": [ + 0.692, + 0.28, + 0.885, + 0.403 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.692, + 0.404, + 0.863, + 0.414 + ], + "angle": 0, + "content": "Select the best caption for the image:" + }, + { + "type": "image_caption", + "bbox": [ + 0.692, + 0.414, + 0.878, + 0.423 + ], + "angle": 0, + "content": "A. The left window is larger than the right window." + }, + { + "type": "image_caption", + "bbox": [ + 0.692, + 0.423, + 0.878, + 0.431 + ], + "angle": 0, + "content": "B. The right window is larger than the left window." + }, + { + "type": "image_caption", + "bbox": [ + 0.692, + 0.431, + 0.841, + 0.439 + ], + "angle": 0, + "content": "C. The Two mirrors are of the same size." + }, + { + "type": "image_caption", + "bbox": [ + 0.697, + 0.448, + 0.836, + 0.459 + ], + "angle": 0, + "content": "LLaVA-1.5: The best caption is A." + }, + { + "type": "image_caption", + "bbox": [ + 0.697, + 0.464, + 0.848, + 0.474 + ], + "angle": 0, + "content": "SpatialLLaVA: The best caption is C." + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.481, + 0.907, + 0.523 + ], + "angle": 0, + "content": "Figure 2. Visualization of LLaVA-1.5 and SpatialLLaVA on SpatialBench. The first row displays examples from outdoor scenes, while the second row features cases in indoor environments. LLaVA-1.5 often encounters difficulties with basic spatial concepts, while SpatialLLaVA improves its spatial perception ability by employing enhanced visual representations." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.535, + 0.483, + 0.61 + ], + "angle": 0, + "content": "ternVL [11], and SigLIP [73]. However, ReCLIP [52] recognizes CLIP lacks the capability of understanding spatial relationships in images. Additionally, as discussed in [58], the shortcoming in visual encoding may limit the capabilities of MLLMs built upon it." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.611, + 0.483, + 0.641 + ], + "angle": 0, + "content": "Our work attempts to enhance CLIP's spatial perception and overcome the shortcomings in spatial understanding." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.651, + 0.433, + 0.667 + ], + "angle": 0, + "content": "2.2. Spatial Intelligence in Visual Perception" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.674, + 0.483, + 0.856 + ], + "angle": 0, + "content": "Spatial understanding [8, 56] and spatial intelligence are fundamental abilities for artificial intelligence systems to perceive and interact with the complex visual world. In our 3D world, spatial understanding involves distinguishing various spatial relationships between objects, such as up and down, left and right, far and near, and inside and outside, as well as perceiving the size, thickness, length, and width of visual elements. The ability to comprehend and reason on these basic spatial concepts is crucial for solving many complex problems, including robot navigation [70, 75], manipulation [25], human-computer interaction [19, 44], and applications in virtual and augmented reality [33]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.857, + 0.483, + 0.902 + ], + "angle": 0, + "content": "MLLMs have made impressive progress in recent years. Built upon pre-trained visual representation and large language models [20, 59], these methods demonstrate strong" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.535, + 0.907, + 0.626 + ], + "angle": 0, + "content": "performance across a variety of downstream tasks, such as image captioning [39], visual question answering [1], visual grounding [30, 45], and action recognition [23]. However, despite their powerful generalization capabilities, most MLLMs still struggle with spatial understanding and reasoning, as discussed in [7]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.635, + 0.907, + 0.787 + ], + "angle": 0, + "content": "There are several works that attempt to address the shortcomings of MLLMs in spatial understanding. SpatialVLM [7] is the first to focus on this issue, introducing massive automatically-generated 3D spatial reasoning VQA data for training MLLM. Its successor, SpatialRGPT [12], enhances the 3D-aware data curating pipeline and provides extra visual depth and region information for MLLM. CubeLLM [14] combines multiple existing 2D and 3D object detection datasets in a unified format and then trains MLLM with better spatial detection capability." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Existing works mainly enhance the spatial intelligence of MLLM by improving the training data, while our work explores the issue from a different perspective: improving the visual representations. Utilizing the same general-purpose training data, simply replacing the standard CLIP with our SpatialCLIP enhances the spatial understanding capabilities of LLaVA [41, 42]." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "29658" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.132, + 0.088, + 0.619, + 0.345 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.629, + 0.089, + 0.864, + 0.34 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.348, + 0.908, + 0.391 + ], + "angle": 0, + "content": "Figure 3. The overview of SpatialCLIP. The visual encoder is replaced by our newly proposed 3D-Inspired Vision Transformer, which enables more informative RGB-D inputs and excels in modeling the 3D spatial structure. Language supervision is enhanced through recaption and perturbation, which prompt the encoding of visual spatial information." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.4, + 0.222, + 0.417 + ], + "angle": 0, + "content": "3. SpatialCLIP" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.426, + 0.484, + 0.488 + ], + "angle": 0, + "content": "Our proposed SpatialCLIP enhances standard CLIP from two perspectives: 3D-Inspired ViT for more powerful \"Visual Model\" in Sec. 3.1 and recaption&perturbation strategy for more informative \"Language Supervision\" in Sec. 3.2." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.498, + 0.368, + 0.515 + ], + "angle": 0, + "content": "3.1. Visual Model: 3D-Inspired ViT" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.522, + 0.484, + 0.749 + ], + "angle": 0, + "content": "In RGB images, the 3D spatial structure of the world is flattened into a 2D plane. Many methods [12, 15, 17, 74] try to introduce depth maps for extra 3D information. However, how to effectively encode such RGB-D inputs while maintaining the pre-trained knowledge for the RGB channel is non-trivial. AlphaCLIP [55] enables an extra input channel with a simple extra input convolution layer, while SpatialRGPT [12] repeats depth map three times as a depth image, and lately fuses the latent of the RGB image and depth image with the pre-trained CLIP image encoder. Our empirical findings show that such simple early and late fusion methods are less effective at integrating depth information. To address this, we propose 3D-inspired ViT for RGB-D inputs to replace the standard vision transformer (ViT) widely used as the visual model in CLIP models." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Recently, monocular depth estimation models [31, 65, 66] have gained impressive progress, depth maps can be easily predicted for wild-world RGB images with low cost. Moreover, the accurate depth map shows impressive results in lifting the images into 3D point clouds. Meanwhile, the architectures for processing 3D point cloud [61, 62, 67] focus on capturing spatial relationships among features in 3D space. These observations inspire us to lift 2D tokens into 3D space using predicted depth information and thereby incorporate design insights from 3D point cloud architectures" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.403, + 0.845, + 0.417 + ], + "angle": 0, + "content": "to better model 3D spatial relationships in images." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.418, + 0.906, + 0.538 + ], + "angle": 0, + "content": "To this end, we first dive into the standard ViT, which processes RGB images as input. Specifically, an RGB image \\(\\mathbf{X} \\in \\mathbb{R}^{H \\times W \\times 3}\\) is divided into 2D patches, and each is linearly projected into \\(C\\)-dimension tokens. The resulting 2D tokens map \\(\\mathbf{Z} \\in \\mathbb{R}^{H_z \\times W_z \\times C}\\) is passed into a stack of transformer blocks, where \\(H_{z}\\) and \\(W_{z}\\) represent the tokens map's height and width. The 2D token map for the \\(L\\)-th transformer block can be denoted as \\(\\mathbf{Z}_{L} \\in \\mathbb{R}^{H_{z} \\times W_{z} \\times C_{L}}\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.54, + 0.906, + 0.706 + ], + "angle": 0, + "content": "To integrate the well-established 3D point cloud architecture designs into ViT for images, for each transformer layer, we first lift its 2D input tokens \\(\\mathbf{Z}_L \\in \\mathbb{R}^{H_z \\times W_z \\times C_L}\\) into 3D voxel \\(\\mathbf{Z}_L^{3\\mathrm{D}} \\in \\mathbb{R}^{H_z \\times W_z \\times D_z \\times C_L}\\) based on the average depth value of each patch, where \\(H_z\\), \\(W_z\\), and \\(D_z\\) represent the grid sizes along the height, width, and depth axes. Since \\(\\mathbf{Z}_L^{3\\mathrm{D}}\\) is derived from \\(\\mathbf{Z}_L\\), the resulting 3D voxel grid is highly sparse. The \\(H_z \\times W_z\\) tokens in \\(\\mathbf{Z}_L\\) are distributed along the depth axis based on their corresponding depth values, and extracting tokens along the depth axis can easily compress \\(\\mathbf{Z}_L^{3\\mathrm{D}}\\) back into 2D space." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.706, + 0.906, + 0.812 + ], + "angle": 0, + "content": "For 3D voxelized tokens, 3D depth-wise convolution is widely used to capture spatial patterns across the three dimensions. As shown in the right part of Fig. 3, we integrate this 3D convolution within each transformer block of the pre-trained ViT model, applying it to \\(\\mathbf{Z}_L^{3\\mathrm{D}}\\) and then short-cutting the resulting features back to the original \\(\\mathbf{Z}_L\\). This process is expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.572, + 0.825, + 0.905, + 0.843 + ], + "angle": 0, + "content": "\\[\n\\mathbf {Z} _ {L} = \\mathbf {Z} _ {L} + \\operatorname {C o m p r e s s} \\left(3 \\mathrm {D} - \\operatorname {C o n v} \\left(\\mathbf {Z} _ {\\mathrm {L}} ^ {3 \\mathrm {D}}\\right)\\right) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.906, + 0.902 + ], + "angle": 0, + "content": "where 3D-Conv \\((\\cdot)\\) is the 3D depth wise convolution and Compress \\((\\cdot)\\) indicates compressing sparse 3D voxel back into 2D token maps." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "29659" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.198 + ], + "angle": 0, + "content": "Moreover, another commonly used technique in point cloud networks is 3D relative positional encoding, which explicitly indicates the relative spatial relationship between every pair of tokens in 3D space. For the self-attention operator of each transformer layer, we employ 3D relative positional biases on the attention map. The attention score \\( a_{ij} \\) between \\( i \\)-th and \\( j \\)-th token can be represented as:" + }, + { + "type": "equation", + "bbox": [ + 0.179, + 0.206, + 0.483, + 0.242 + ], + "angle": 0, + "content": "\\[\na _ {i j} = \\frac {\\left(\\mathbf {z} _ {i} \\mathbf {W} ^ {Q}\\right) \\left(\\mathbf {z} _ {j} \\mathbf {W} ^ {K}\\right) ^ {T} + b _ {i j}}{\\sqrt {C}} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.25, + 0.483, + 0.295 + ], + "angle": 0, + "content": "where \\( b_{ij} \\) is a scalar selected from the learnable bias matrix \\( \\mathbf{B} \\in \\mathbb{R}^{(2H_z - 1) \\times (2W_z - 1) \\times (2D_z - 1)} \\), based on the relative 3D position between the \\( i \\)-th and \\( j \\)-th tokens in 3D space." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.295, + 0.484, + 0.356 + ], + "angle": 0, + "content": "By adopting the designs inspired by the structure of 3D point cloud networks, the 3D-inspired ViT can perceive 2D images from the 3D perspective, greatly enhancing the potential to model complex spatial relationships in images." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.364, + 0.483, + 0.395 + ], + "angle": 0, + "content": "3.2. Language Supervision: Recapiction & Perturbation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.403, + 0.483, + 0.584 + ], + "angle": 0, + "content": "In the contrastive learning process of CLIP, visual representations of different images are pulled close with the paired caption while pushed away from captions of other images. The guidance information in language captions strongly influences the visual model's capability. For prompting 3D-Inspired ViT to better capture spatial relationships in images, more informative supervision from natural language is necessary. However, current widely-used vision-language datasets for training CLIP, such as Conceptual-12M [5] and LAION [50, 51], are inherently noisy and mainly provide simple summaries of global attributes, lacking detailed caption with accurate spatial description." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.585, + 0.483, + 0.719 + ], + "angle": 0, + "content": "Recent studies [3, 6, 21, 37] have shown that recaptioning images with semantically richer and more aligned descriptions can greatly enhance the performance of vision-language models like CLIP and text-to-image models. However, as discussed in [7, 12], current MLLMs excel at recognizing object categories and attributes but struggle with spatial perception. Consequently, the detailed captions they generated often fail to accurately capture all spatial relationships within the image." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.483, + 0.854 + ], + "angle": 0, + "content": "To address this \"chicken-and-egg\" dilemma, we propose perturbing the spatial descriptions in the dense captions generated by MLLMs. Specifically, we employ a template to randomly substitute phrases indicating spatial relationships with their antonymic counterparts. For instance, a caption like \"two cats are standing in front of a table, while the left cat is bigger than the right one\" may be modified to \"two cats are standing under a table, while the upper cat is smaller than the lower one.\"" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Since the generated and perturbed captions are semantically consistent in terms of object categories and attributes, applying contrastive learning between these two captions" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.168 + ], + "angle": 0, + "content": "encourages the model to distinguish the entirely incorrect caption from the partially correct one. Such a straightforward perturbation method effectively promotes the visual model's ability to capture detailed spatial relationships within images. The learning objective can be expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.568, + 0.179, + 0.905, + 0.265 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} L = - \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\left[ \\log \\frac {e ^ {\\left(\\mathbf {t} _ {i} \\cdot \\mathbf {v} _ {i} / \\tau\\right)}}{\\sum_ {j = 1} ^ {N} e ^ {\\left(\\mathbf {t} _ {i} \\cdot \\mathbf {v} _ {j} / \\tau\\right)}} \\right. \\tag {3} \\\\ \\left. + \\log \\frac {e ^ {\\left(\\mathbf {t} _ {i} \\cdot \\mathbf {v} _ {i} / \\tau\\right)}}{\\sum_ {j = 1} ^ {N} \\left(e ^ {\\left(\\mathbf {t} _ {j} \\cdot \\mathbf {v} _ {i} / \\tau\\right)} + e ^ {\\left(\\tilde {\\mathbf {t}} _ {j} \\cdot \\mathbf {v} _ {i} / \\tau\\right)}\\right)} \\right] \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.275, + 0.905, + 0.322 + ], + "angle": 0, + "content": "where the \\( v_{i} \\) is the visual representation of \\( i \\)-th image, while \\( t_{i} \\) and \\( \\tilde{t}_i \\) denote the textual representation of the \\( i \\)-th caption and its perturbed version, respectively." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.335, + 0.659, + 0.353 + ], + "angle": 0, + "content": "4. SpatialLLaVA" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.361, + 0.906, + 0.437 + ], + "angle": 0, + "content": "As discussed in [57, 58], the capabilities of MLLMs may be constrained by the inherent limitations of their visual representations. This raises a crucial question: Can the spatial intelligence of CLIP-based MLLM be enhanced by improving the spatial perception of CLIP's visual representations?" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.437, + 0.906, + 0.574 + ], + "angle": 0, + "content": "We utilize LLaVA-1.5, which employs the pre-trained CLIP as its visual encoding model, as the framework to investigate the effectiveness of SpatialCLIP in MLLM. Specifically, we replace the CLIP model in LLaVA-1.5 [41] with our SpatialCLIP. To ensure fair and systematic comparisons between visual encoding methods, we follow the two-stage training strategy used in LLaVA-1.5 and apply the same training datasets and hyperparameters for each stage. The resulting MLLM is named SpatialLLaVA." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.588, + 0.65, + 0.605 + ], + "angle": 0, + "content": "5. SpatialBench" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.613, + 0.906, + 0.764 + ], + "angle": 0, + "content": "To evaluate the ability of CLIP and MLLM in understanding complex 2D/3D spatial relations in images, we build a benchmark SpatialBench, which focuses entirely on spatial perception. Our benchmark employs the caption matching task, requiring CLIP or MLLM to distinguish the caption that best matches the image from several distractors. By manually crafting the correct description and distractors, the caption-matching task requires the model to distinguish nuanced spatial concepts in the image (e.g. object arrangement, relationships, relative depth, and size)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.906, + 0.902 + ], + "angle": 0, + "content": "To fully assess models in various scenarios, we collect images of both outdoor and indoor scenes. Specifically, outdoor images are collected from the COCO [39] validation set and the Internet, totaling 159 images. Indoor images are sourced from ScanNet [16], consisting of 140 images. Each image is paired with 2 to 4 manually crafted descriptions, of which only one is accurate, with the remaining serving as distractors. Examples of SpatialBench can be found in Fig. 2 and Appendix." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "29660" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.12, + 0.089, + 0.875, + 0.251 + ], + "angle": 0, + "content": "
ModelParamSpatial-FocusedGeneral-Purposed
SpatialBenchBLINKCOCOFlickr30K
OutDoor.InDoor.SpatialR@1R@5R@1R@5
Random-29.7729.8450.00----
OpenAI CLIP-B@224px149.6M35.2230.7151.0542.1967.2534.4056.65
SpatialCLIP-B@224px149.9M+24.8M41.5134.2857.3448.3373.2738.8662.80
OpenAI CLIP-L@336px427.9M32.0730.0047.5546.7870.5438.2560.95
SpatialCLIP-L@336px428.7M+24.8M35.8433.5757.3450.6275.6642.5467.04
" + }, + { + "type": "table_caption", + "bbox": [ + 0.308, + 0.253, + 0.688, + 0.266 + ], + "angle": 0, + "content": "Table 1. Results on CLIP benchmark. The best results are bold." + }, + { + "type": "table", + "bbox": [ + 0.097, + 0.28, + 0.907, + 0.432 + ], + "angle": 0, + "content": "
Training DataModelSpatialBenchBLINKCV-Bench3d
OutDoor.InDoor.SpatialDepthCorresJigsawDepthDistance
Spatial-FocusedSpatial-RGPT-7B37.7338.5765.7382.3032.5550.0062.3359.00
Stronger General-PurposedLLaVA-1.6-7B39.6233.5768.5351.6129.6548.0053.3551.12
LLaVA-1.6-13B38.3635.7169.9948.3830.8150.6755.1652.25
LLaVA-interleave-7B33.2737.1469.9356.6934.8851.3357.1462.30
Baseline General-PurposedLLaVA-1.5-7B29.5532.4761.5452.4225.5811.3345.8350.00
Spatial-LLaVA-7B40.2540.1771.3258.0631.9751.3357.3352.16
" + }, + { + "type": "table_caption", + "bbox": [ + 0.328, + 0.435, + 0.668, + 0.448 + ], + "angle": 0, + "content": "Table 2. Results on Spatial-Focused MLLM benchmark." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.472, + 0.485, + 0.595 + ], + "angle": 0, + "content": "
POPE [38]MM.Vet [72]VOAv2 [27]GOA [29]RealWorldQA [63]
LLaVA-1.5-7B85.930.578.562.054.8
Spatial-LLaVA-7B86.530.779.960.654.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.108, + 0.597, + 0.465, + 0.611 + ], + "angle": 0, + "content": "Table 3. Results on General-Purposed MLLM benchmark." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.624, + 0.483, + 0.716 + ], + "angle": 0, + "content": "For CLIP, we calculate matching scores between the image and its corresponding captions, selecting the description with the highest matching score as the prediction. For MLLMs, the descriptions are presented as multiple-choice options, and the models are prompted to select the best description in a question-answering manner." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.731, + 0.216, + 0.748 + ], + "angle": 0, + "content": "6. Experiment" + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.756, + 0.308, + 0.772 + ], + "angle": 0, + "content": "6.1. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.78, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Training setting. We use images in SA1B dataset [32] for training, since most images in it contain multiple objects and showcase complex spatial relationships. Following [6], LLaVA-1.5-13B [41], CoCa [71] and extra ShareGPT4V [9] are used for image reception, the three kinds of captions are randomly used in training for more diverse and informative supervision. A detailed caption perturbation template is provided in the Appendix. The tem" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.476, + 0.907, + 0.551 + ], + "angle": 0, + "content": "perature factor \\(\\tau\\) in Eq. 3 is set to 0.01. For optimization, we employ AdamW [43] with a learning rate of 2e-7 for pre-trained CLIP encoders and 4e-6 for the newly introduced components. Models are trained for 8,000 steps with a batch size of 2,048 on 8 A100 GPUs." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.568, + 0.909, + 0.72 + ], + "angle": 0, + "content": "Model setting. We train SpatialCLIP from OpenAI CLIP-B/16@224px and OpenAI CLIP-L/14@336px, leading to SpatialCLIP-B and SpatialCLIP-L, respectively. We utilize the lightweight Depth Anything v2 ViT-Small as our monocular depth estimation model, which brings an extra 24.8M parameters. The kernel size of the 3D CNN in Eq. 1 is 3, and the \\( D_w \\) is 6. The newly added 3D CNN and 3D relative position encoding are both zero-initialized and introduce only nearly 1M extra parameters. We develop SpatialLLaVA based on SpatialCLIP-L, following LLaVA-1.5." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.727, + 0.722, + 0.743 + ], + "angle": 0, + "content": "6.2. Comparison for CLIP" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Results on spatial-focused benchmark. To assess the spatial perception ability for CLIP models, we employ our newly proposed SpatialBench and modify the \"spatial relation\" subset of BLINK to the caption-matching task, as another spatial-focused benchmark. Results on these benchmarks are presented in Tab. 1. There are two key findings: 1) In both the CLIP-B and -L variants, SpatialCLIP demonstrates significant performance advantages over the original CLIP, indicating that SpatialCLIP more effectively captures spatial information in images and distinguishes between" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "29661" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.092, + 0.296, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.306, + 0.096, + 0.892, + 0.158 + ], + "angle": 0, + "content": "LLaVA: The banana nut bread is on the left side of the picture, with a banana and an egg on the right side. The banana is placed in a blue bowl next to the bread. The eggs are in a glass bowl on the left side of the picture. The banana nut bread is placed in a brown paper bag, with a nut on top." + }, + { + "type": "text", + "bbox": [ + 0.306, + 0.17, + 0.895, + 0.231 + ], + "angle": 0, + "content": "SpatialLLaVA: In the image, a bag of banana nut bread mix is placed next to a banana, an egg, and some nuts. The banana is positioned on the right side of the image, while the egg is located on the left side, and the nuts are scattered around the bag of mix. The bag of mix is placed in the middle of the scene, with the banana, egg, and nuts surrounding it." + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.25, + 0.69, + 0.296 + ], + "angle": 0, + "content": "LLaVA: There are two cars that are oriented in a vertical position. The green bus is on the left side of the picture, while the red bus is on the right side. The front of the green bus is facing toward the viewer, while the red bus is facing away from the viewer." + }, + { + "type": "text", + "bbox": [ + 0.1, + 0.308, + 0.691, + 0.356 + ], + "angle": 0, + "content": "SpatialLLaVA: In the image, the two buses are positioned side by side, with one bus on the left and the other on the right. They are both facing the same direction, which suggests that they are either parked next to each other or driving in the same direction on the road." + }, + { + "type": "image", + "bbox": [ + 0.704, + 0.245, + 0.903, + 0.358 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.363, + 0.908, + 0.392 + ], + "angle": 0, + "content": "Figure 4. Qualitative comparison between LLaVA-1.5 and SpatialLLaVA. We prompt these two MLLMs to provide detailed descriptions of spatial relationships in images. Our model perform better at capturing these spatial concepts." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.406, + 0.484, + 0.542 + ], + "angle": 0, + "content": "different spatial concepts. 2) For both the original CLIP and SpatialCLIP, performance on spatial-focused benchmarks is not correlated with model size. We conjecture this situation is derived from the scarcity of the spatial concepts in the current CLIP pre-training dataset. Increasing the parameter (i.e., enhancing learning capacity) does not necessarily lead to improved spatial perception. This highlights the value of additional designs for better spatial perception, rather than merely scaling up within the existing paradigm." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.561, + 0.484, + 0.697 + ], + "angle": 0, + "content": "Results on general-purposed benchmark. In addition to enhancing spatial concept comprehension, accurate recognition of object categories and attributes is also essential for advanced applications. Since SpatialCLIP is built upon pretrained CLIP and is further trained with high-quality images and informative captions generated by diverse MLLMs, we also observe significant performance improvements on general-purpose benchmarks, like COCO and Flickr30K for zero-shot image-text retrieval." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.71, + 0.312, + 0.725 + ], + "angle": 0, + "content": "6.3. Comparison for MLLM" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.733, + 0.483, + 0.854 + ], + "angle": 0, + "content": "Results on spatial-focused benchmark. In Tab. 2, we comprehensively validate the spatial perception and reasoning ability of our SpatialLLaVA on various spatial-focused MLLM benchmarks. With the same training data, SpatialLLaVA achieves much better results than LLaVA-1.5. SpatialLLaVA exhibits a stronger capability to differentiate between nuanced spatial concepts across indoor and outdoor scenes on our SpatialBench." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "We also conduct absolute performance comparison with more advanced MLLMs, including Spatial-RGPT (trained on spatial-focused data) and LLaVA-1.6/LLaVA-Interleave" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.407, + 0.907, + 0.542 + ], + "angle": 0, + "content": "(trained on much broader general-purpose data). Despite this \"unfair\" comparison, Spatial-LLaVA consistently shows superior spatial perception. Interestingly, Spatial-RGPT excels in tasks it was explicitly trained for (e.g., depth and distance estimation) but underperforms on tasks beyond its training scope (e.g., spatial question-answering). This observation highlights the need to enhance visual perception capabilities in MLLMs beyond simply expanding training data." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.544, + 0.909, + 0.695 + ], + "angle": 0, + "content": "The advantages on BLINK underscore the importance of capturing spatial structure for high-level applications. For \"visual corresponding\" task, capturing spatial structure is beneficial for identifying the same scene points across different viewpoints, lighting, or times, as the scene structure often remains consistent despite these changes. For \"Jigsaw,\" viewing 2D images in the 3D perspective allows the model to inspect the continuity of the 3D spatial structure rather than only relying on color and texture, aiding in organizing image patches." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.697, + 0.909, + 0.789 + ], + "angle": 0, + "content": "CV-Bench[57] collects and annotates samples from the 3D object detection dataset Omni3D [4], focusing on 3D understanding. Employing SpatialCLIP also leads to performance improvements for MLLM on this benchmark, further demonstrating the superiority of SpatialCLIP's visual representations in capturing 3D spatial relationships." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Results on general-purposed benchmark. In Tab. 3, we examine how enhancing spatial perception in visual representations affect the performance of MLLMs in general-purposed scenarios. Across five benchmarks, Spatial-LLaVA achieves comparable performance to LLaVA-1.5 for general-purposed questions." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "29662" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.145, + 0.089, + 0.85, + 0.238 + ], + "angle": 0, + "content": "
RowTuning3D CNN3D RPEPerturbSpatialBenchBLINKCOCOFlickr30k
OutDoor.InDoor.Spatial.R@1R@1
1×---35.2231.4351.0542.1934.40
2✓×××35.8531.4351.7548.9939.06
3✓✓××38.3631.4353.8549.0437.76
4✓×✓×37.1032.1454.5547.8139.33
5✓✓✓×38.3633.5755.2447.6737.56
6✓✓✓✓41.5134.2857.3448.3438.86
" + }, + { + "type": "table_caption", + "bbox": [ + 0.111, + 0.24, + 0.884, + 0.255 + ], + "angle": 0, + "content": "Table 4. Ablation study of our \"Visual Model\" and \"Language Supervision\". All the experiments are conducted on SpatialCLIP-B." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.279, + 0.483, + 0.378 + ], + "angle": 0, + "content": "
SettingSpatialBenchBLINKCOCOFlickr30k
OutDoorInDoorSpatialR@1R@1
OpenAI CLIP-B35.2231.4351.0542.1934.40
Early Fusion39.6232.8551.7543.0133.09
Late Fusion40.2534.2852.4542.3831.80
3D-Inspried ViT41.5134.2857.3448.3438.86
" + }, + { + "type": "table_caption", + "bbox": [ + 0.099, + 0.381, + 0.473, + 0.395 + ], + "angle": 0, + "content": "Table 5. Ablation study of depth channel integration strategies." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.422, + 0.483, + 0.514 + ], + "angle": 0, + "content": "Qualitative comparison. To better demonstrate the improvement of SpatialLLaVA in spatial perception, we provide visualizations of both models in Fig. 2 and 4. The qualitative comparison of their responses reveals that SpatialLLaVA can more accurately and comprehensively capture the spatial relationships between objects in images." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.524, + 0.245, + 0.541 + ], + "angle": 0, + "content": "6.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.548, + 0.483, + 0.7 + ], + "angle": 0, + "content": "Effect of recaption. In the first two rows of Tab. 4, we present the comparison between the original CLIP and its direct fine-tuning using our recaption dataset. On the general-purposed benchmarks, COCO and Flickr30k, tuning on our dataset leads to significant performance improvements, underscoring the benefits of high information density and semantic alignment in recaption data. However, simply using the recaption data only brings negligible changes to the spatial-focused benchmarks, motivating us to introduce more specific designs to learn spatial concepts." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Effect of 3D-inspired ViT. We ablate the designs of 3D-Inspired ViT in Row 2-5 of Tab. 4. When fine-tuning on the same recaption data, incorporating 3D convolution (Row 3) or 3D relative position encoding (Row 4) shows improved performance on spatial-focused benchmarks compared to the standard ViT (Row 2). The combination of both techniques results in even greater enhancements. The improvement in spatial perception gained from the 3D-inspired design underscores the importance of integrating 3D point cloud structures into 2D ViT. Encoding images in 3D space with 3D-specific structure enables the visual model to capture complex 3D spatial information more effectively." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.281, + 0.907, + 0.449 + ], + "angle": 0, + "content": "Effect of caption perturbation. The comparison between Rows 5 and 6 in Tab. 4 illustrates the benefits of caption perturbation, which prompts the model to capture spatial concepts and enhances its spatial perception ability. Since MLLMs themselves often struggle in perceiving and reasoning about spatial relationships, the captions generated by them can not fully capture all spatial details. By perturbing the spatial phrases, we create references that are fully incorrect in spatial relationships. This approach drives the model to learn spatial concepts by distinguishing the partially correct captions from the fully incorrect ones." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.477, + 0.909, + 0.675 + ], + "angle": 0, + "content": "Different depth channel integration strategies. In addition to our 3D-Inspired ViT, there are other alternative methods to enable RGB-D inputs. In Tab. 5, we discuss the commonly used early and late fusion schemes, which project the depth map into RGB features before and after the visual encoder, which are used by AlphaCLIP [55] and SpatialRGPT [12], respectively. In most benchmarks, simply projecting the depth channel to the pre-trained RGB features is hard to improve the performance. Considering the additional information provided by the reception data, these schemes may even have a negative impact. Conversely, our 3D-inspired ViT makes effective use of the extra depth information, yielding significant improvements." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.692, + 0.634, + 0.709 + ], + "angle": 0, + "content": "7. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.909, + 0.901 + ], + "angle": 0, + "content": "We present SpatialCLIP to overcome the shortcomings of original CLIP models in spatial understanding. By introducing improved \"Visual Model\" (3D-Inspired ViT) and enhanced \"Language Supervision\" (Recap tion & Perturbation), our model achieves significant improvements, not just on spatial-focused benchmarks like our proposed Spatial-Bench and modified BLINK, but also on zero-shot image-text retrieval. Additionally, we develop SpatialLLaVA based on SpatialCLIP, adhering to the LLaVA-1.5 training protocol. This integration of visual representations specifically enhanced for spatial recognition improves the spatial intelligence of MLLM." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.958 + ], + "angle": 0, + "content": "29663" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.259, + 0.108 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.116, + 0.484, + 0.161 + ], + "angle": 0, + "content": "This work was supported in part by National Natural Science Foundation of China under Grant No.624B2128 and No.62222211." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.177, + 0.188, + 0.193 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.202, + 0.484, + 0.244 + ], + "angle": 0, + "content": "[1] Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C Lawrence Zitnick, and Devi Parikh. Vqa: Visual question answering. In ICCV, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.246, + 0.484, + 0.315 + ], + "angle": 0, + "content": "[2] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond. arXiv:2308.12966, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.317, + 0.484, + 0.386 + ], + "angle": 0, + "content": "[3] James Betker, Gabriel Goh, Li Jing, Tim Brooks, Jianfeng Wang, Linjie Li, Long Ouyang, Juntang Zhuang, Joyce Lee, Yufei Guo, et al. Improving image generation with better captions. https://cdn.openai.com/papers/dall-e-3.pdf, 2023.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.388, + 0.483, + 0.443 + ], + "angle": 0, + "content": "[4] Garrick Brazil, Abhinav Kumar, Julian Straub, Nikhila Ravi, Justin Johnson, and Georgia Gkioxari. Omni3d: A large benchmark and model for 3d object detection in the wild. In CVPR, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.445, + 0.484, + 0.5 + ], + "angle": 0, + "content": "[5] Soravit Changpinyo, Piyush Sharma, Nan Ding, and Radu Soricut. Conceptual 12m: Pushing web-scale image-text pretraining to recognize long-tail visual concepts. In CVPR, 2021. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.502, + 0.484, + 0.571 + ], + "angle": 0, + "content": "[6] Agneet Chatterjee, Gabriela Ben Melech Stan, Estelle Aflalo, Sayak Paul, Dhruba Ghosh, Tejas Gokhale, Ludwig Schmidt, Hannaneh Hajishirzi, Vasudev Lal, Chitta Baral, et al. Getting it right: Improving spatial consistency in text-to-image models. In ECCV, 2025. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.574, + 0.484, + 0.628 + ], + "angle": 0, + "content": "[7] Boyuan Chen, Zhuo Xu, Sean Kirmani, Brain Ichter, Dorsa Sadigh, Leonidas Guibas, and Fei Xia. Spatialvlm: Endowing vision-language models with spatial reasoning capabilities. In CVPR, 2024. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.631, + 0.484, + 0.672 + ], + "angle": 0, + "content": "[8] Dave Zhenyu Chen, Angel X Chang, and Matthias Nießner. Scanrefer: 3d object localization in rgb-d scans using natural language. In ECCV, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.674, + 0.484, + 0.729 + ], + "angle": 0, + "content": "[9] Lin Chen, Jinsong Li, Xiaoyi Dong, Pan Zhang, Conghui He, Jiaqi Wang, Feng Zhao, and Dahua Lin. Sharegpt4v: Improving large multi-modal models with better captions. arXiv:2311.12793, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.731, + 0.483, + 0.8 + ], + "angle": 0, + "content": "[10] Zhe Chen, Weiyun Wang, Hao Tian, Shenglong Ye, Zhang-wei Gao, Erfei Cui, Wenwen Tong, Kongzhi Hu, Jiapeng Luo, Zheng Ma, et al. How far are we to gpt-4v? closing the gap to commercial multimodal models with open-source suites. arXiv:2404.16821, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.802, + 0.483, + 0.87 + ], + "angle": 0, + "content": "[11] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. In CVPR, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.483, + 0.902 + ], + "angle": 0, + "content": "[12] An-Chieh Cheng, Hongxu Yin, Yang Fu, Qiushan Guo, Ruihan Yang, Jan Kautz, Xiaolong Wang, and Sifei Liu. Spatial-" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.202, + 0.484, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "rgpt: Grounded spatial reasoning in vision language model. arXiv:2406.01584, 2024. 2, 3, 4, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[13] Mehdi Cherti, Romain Beaumont, Ross Wightman, Mitchell Wortsman, Gabriel Ilharco, Cade Gordon, Christoph Schuhmann, Ludwig Schmidt, and Jenia Jitsev. Reproducible scaling laws for contrastive language-image learning. In CVPR, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.192, + 0.905, + 0.261 + ], + "angle": 0, + "content": "[14] Jang Hyun Cho, Boris Ivanovic, Yulong Cao, Edward Schmerling, Yue Wang, Xinshuo Weng, Boyi Li, Yurong You, Philipp Krahenbuhl, Yan Wang, et al. Language-image models with 3d understanding. arXiv:2405.03685, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.263, + 0.905, + 0.305 + ], + "angle": 0, + "content": "[15] Jaeyoung Chung, Jeongtaek Oh, and Kyoung Mu Lee. Depth-regularized optimization for 3d gaussian splatting in few-shot images. In CVPR, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.306, + 0.905, + 0.36 + ], + "angle": 0, + "content": "[16] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In CVPR, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.362, + 0.905, + 0.404 + ], + "angle": 0, + "content": "[17] Kangle Deng, Andrew Liu, Jun-Yan Zhu, and Deva Ramanan. Depth-supervised nerf: Fewer views and faster training for free. In CVPR, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.406, + 0.905, + 0.447 + ], + "angle": 0, + "content": "[18] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv:2010.11929, 2020.1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.449, + 0.905, + 0.517 + ], + "angle": 0, + "content": "[19] Danny Driess, Fei Xia, Mehdi SM Sajjadi, Corey Lynch, Aakanksha Chowdhery, Brian Ichter, Ayzaan Wahid, Jonathan Tompson, Quan Vuong, Tianhe Yu, et al. Palm-e: An embodied multimodal language model. arXiv:2303.03378, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.519, + 0.905, + 0.574 + ], + "angle": 0, + "content": "[20] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv:2407.21783, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.576, + 0.905, + 0.644 + ], + "angle": 0, + "content": "[21] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In ICML, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.646, + 0.905, + 0.688 + ], + "angle": 0, + "content": "[22] Alex Fang, Albin Madappally Jose, Amit Jain, Ludwig Schmidt, Alexander Toshev, and Vaishaal Shankar. Data filtering networks. arXiv:2309.17425, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.689, + 0.905, + 0.758 + ], + "angle": 0, + "content": "[23] Chaoyou Fu, Yuhan Dai, Yondong Luo, Lei Li, Shuhuai Ren, Renrui Zhang, Zihan Wang, Chenyu Zhou, Yunhang Shen, Mengdan Zhang, et al. Video-mme: The first-ever comprehensive evaluation benchmark of multi-modal llms in video analysis. arXiv:2405.21075, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.76, + 0.905, + 0.815 + ], + "angle": 0, + "content": "[24] Xingyu Fu, Yushi Hu, Bangzheng Li, Yu Feng, Haoyu Wang, Xudong Lin, Dan Roth, Noah A Smith, Wei-Chiu Ma, and Ranjay Krishna. Blink: Multimodal large language models can see but not perceive. arXiv:2404.12390, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.817, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[25] Jensen Gao, Bidipta Sarkar, Fei Xia, Ted Xiao, Jiajun Wu, Brian Ichter, Anirudha Majumdar, and Dorsa Sadigh. Physically grounded vision-language models for robotic manipulation. In ICRA, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.902 + ], + "angle": 0, + "content": "[26] Rohit Girdhar, Alaaeldin El-Nouby, Zhuang Liu, Mannat Singh, Kalyan Vasudev Alwala, Armand Joulin, and Ishan" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "29664" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.119 + ], + "angle": 0, + "content": "Misra. Imagebind: One embedding space to bind them all. In CVPR, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.121, + 0.483, + 0.176 + ], + "angle": 0, + "content": "[27] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Bartra, and Devi Parikh. Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In CVPR, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.177, + 0.482, + 0.218 + ], + "angle": 0, + "content": "[28] Benjamin Graham, Martin Engelcke, and Laurens Van Der Maaten. 3d semantic segmentation with submanifold sparse convolutional networks. In CVPR, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.219, + 0.482, + 0.26 + ], + "angle": 0, + "content": "[29] Drew A Hudson and Christopher D Manning. Gqa: A new dataset for real-world visual reasoning and compositional question answering. In CVPR, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.261, + 0.482, + 0.302 + ], + "angle": 0, + "content": "[30] Sahar Kazemzadeh, Vicente Ordonez, Mark Matten, and Tamara Berg. Referitag: Referring to objects in photographs of natural scenes. In EMNLP, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.302, + 0.482, + 0.357 + ], + "angle": 0, + "content": "[31] Bingxin Ke, Anton Obukhov, Shengyu Huang, Nando Metzger, Rodrigo Caye Daudt, and Konrad Schindler. Repurposing diffusion-based image generators for monocular depth estimation. In CVPR, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.359, + 0.482, + 0.413 + ], + "angle": 0, + "content": "[32] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In ICCV, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.414, + 0.482, + 0.467 + ], + "angle": 0, + "content": "[33] Mikhail Konenkov, Artem Lykov, Daria Trinitatova, and Dzmitry Tsetserukou. Vr-gpt: Visual language model for intelligent virtual reality applications. arXiv:2405.11537, 2024.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.469, + 0.482, + 0.511 + ], + "angle": 0, + "content": "[34] Xin Lai, Jianhui Liu, Li Jiang, Liwei Wang, Hengshuang Zhao, Shu Liu, Xiaojuan Qi, and Jiaya Jia. Stratified transformer for 3d point cloud segmentation. In CVPR, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.511, + 0.482, + 0.566 + ], + "angle": 0, + "content": "[35] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. Llava-onevision: Easy visual task transfer. arXiv:2408.03326, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.567, + 0.482, + 0.622 + ], + "angle": 0, + "content": "[36] Liunian Harold Li, Pengchuan Zhang, Haotian Zhang, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, et al. Grounded language-image pre-training. In CVPR, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.622, + 0.482, + 0.678 + ], + "angle": 0, + "content": "[37] Xianhang Li, Haoqin Tu, Mude Hui, Zeyu Wang, Bingchen Zhao, Junfei Xiao, Sucheng Ren, Jieru Mei, Qing Liu, Huangjie Zheng, et al. What if we recapture billions of web images with llama-3? arXiv:2406.08478, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.679, + 0.482, + 0.719 + ], + "angle": 0, + "content": "[38] Yifan Li, Yifan Du, Kun Zhou, Jinpeng Wang, Wayne Xin Zhao, and Ji-Rong Wen. Evaluating object hallucination in large vision-language models. arXiv:2305.10355, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.72, + 0.482, + 0.775 + ], + "angle": 0, + "content": "[39] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.776, + 0.482, + 0.83 + ], + "angle": 0, + "content": "[40] Ziyi Lin, Shijie Geng, Renrui Zhang, Peng Gao, Gerard De Melo, Xiaogang Wang, Jifeng Dai, Yu Qiao, and Hongsheng Li. Frozen clip models are efficient video learners. In ECCV, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.831, + 0.482, + 0.872 + ], + "angle": 0, + "content": "[41] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In CVPR, 2024. 1, 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[42] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, 2024. 1, 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "[43] I Loshchilov. Decoupled weight decay regularization. arXiv:1711.05101, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[44] Soroush Nasiriany, Fei Xia, Wenhao Yu, Ted Xiao, Jacky Liang, Ishita Dasgupta, Annie Xie, Danny Driess, Ayzaan Wahid, Zhuo Xu, et al. Pivot: Iterative visual prompting elicits actionable knowledge for vlms. arXiv:2402.07872, 2024.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.192, + 0.905, + 0.261 + ], + "angle": 0, + "content": "[45] Bryan A Plummer, Liwei Wang, Chris M Cervantes, Juan C Caicedo, Julia Hockenmaier, and Svetlana Lazebnik. Flickr30k entities: Collecting region-to-phrase correspondences for richer image-to-sentence models. In ICCV, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.263, + 0.905, + 0.331 + ], + "angle": 0, + "content": "[46] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv:2307.01952, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.334, + 0.905, + 0.403 + ], + "angle": 0, + "content": "[47] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.405, + 0.905, + 0.461 + ], + "angle": 0, + "content": "[48] Yongming Rao, Wenliang Zhao, Guangyi Chen, Yansong Tang, Zheng Zhu, Guan Huang, Jie Zhou, and Jiwen Lu. Denseclip: Language-guided dense prediction with context-aware prompting. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.462, + 0.905, + 0.503 + ], + "angle": 0, + "content": "[49] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.504, + 0.905, + 0.573 + ], + "angle": 0, + "content": "[50] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv:2111.02114, 2021. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.575, + 0.905, + 0.645 + ], + "angle": 0, + "content": "[51] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. In NeurIPS, 2022. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.646, + 0.905, + 0.701 + ], + "angle": 0, + "content": "[52] Sanjay Subramanian, William Merrill, Trevor Darrell, Matt Gardner, Sameer Singh, and Anna Rohrbach. Reclip: A strong zero-shot baseline for referring expression comprehension. arXiv:2204.05991, 2022. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.702, + 0.905, + 0.744 + ], + "angle": 0, + "content": "[53] Quan Sun, Yuxin Fang, Ledell Wu, Xinlong Wang, and Yue Cao. Eva-clip: Improved training techniques for clip at scale. arXiv:2303.15389, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.745, + 0.905, + 0.8 + ], + "angle": 0, + "content": "[54] Quan Sun, Yufeng Cui, Xiaosong Zhang, Fan Zhang, Qiying Yu, Yueze Wang, Yongming Rao, Jingjing Liu, Tiejun Huang, and Xinlong Wang. Generative multimodal models are in-context learners. In CVPR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.802, + 0.905, + 0.857 + ], + "angle": 0, + "content": "[55] Zeyi Sun, Ye Fang, Tong Wu, Pan Zhang, Yuhang Zang, Shu Kong, Yuanjun Xiong, Dahua Lin, and Jiaqi Wang. Alphaclip: A clip model focusing on wherever you want. In CVPR, 2024. 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.859, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[56] Kaihua Tang, Hanwang Zhang, Baoyuan Wu, Wenhan Luo, and Wei Liu. Learning to compose dynamic tree structures for visual contexts. In CVPR, 2019. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "29665" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.161 + ], + "angle": 0, + "content": "[57] Shengbang Tong, Ellis Brown, Penghao Wu, Sanghyun Woo, Manoj Middepogu, Sai Charitha Akula, Jihan Yang, Shusheng Yang, Adithya Iyer, Xichen Pan, et al. Cambrian1: A fully open, vision-centric exploration of multimodal llms. arXiv:2406.16860, 2024. 1, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.165, + 0.482, + 0.219 + ], + "angle": 0, + "content": "[58] Shengbang Tong, Zhuang Liu, Yuexiang Zhai, Yi Ma, Yann LeCun, and Saining Xie. Eyes wide shut? exploring the visual shortcomings of multimodal llms. In CVPR, 2024. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.222, + 0.482, + 0.29 + ], + "angle": 0, + "content": "[59] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv:2307.09288, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.294, + 0.482, + 0.375 + ], + "angle": 0, + "content": "[60] Haoxiang Wang, Pavan Kumar Anasosalu Vasu, Fartash Faghri, Raviteja Vemulapalli, Mehrdad Farajtabar, Sachin Mehta, Mohammad Rastegari, Oncel Tuzel, and Hadi Pouransari. Sam-clip: Merging vision foundation models towards semantic and spatial understanding. In CVPR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.379, + 0.482, + 0.406 + ], + "angle": 0, + "content": "[61] Peng-Shuai Wang. Octformer: Octree-based transformers for 3d point clouds. TOG, 2023. 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.409, + 0.482, + 0.464 + ], + "angle": 0, + "content": "[62] Xiaoyang Wu, Li Jiang, Peng-Shuai Wang, Zhijian Liu, Xihui Liu, Yu Qiao, Wanli Ouyang, Tong He, and Hengshuang Zhao. Point transformer v3: Simpler faster stronger. In CVPR, 2024. 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.467, + 0.241, + 0.48 + ], + "angle": 0, + "content": "[63] xAI. grok, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.483, + 0.482, + 0.537 + ], + "angle": 0, + "content": "[64] Lin Xu, Yilin Zhao, Daquan Zhou, Zhijie Lin, See Kiong Ng, and Jiashi Feng. Pllava: Parameter-free llava extension from images to videos for video dense captioning. arXiv:2404.16994, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.541, + 0.482, + 0.582 + ], + "angle": 0, + "content": "[65] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. In CVPR, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.585, + 0.482, + 0.626 + ], + "angle": 0, + "content": "[66] Lihe Yang, Bingyi Kang, Zilong Huang, Zhen Zhao, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything v2. arXiv:2406.09414, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.629, + 0.482, + 0.684 + ], + "angle": 0, + "content": "[67] Yu-Qi Yang, Yu-Xiao Guo, Jian-Yu Xiong, Yang Liu, Hao Pan, Peng-Shuai Wang, Xin Tong, and Baining Guo. Swin3d: A pretrained transformer backbone for 3d indoor scene understanding. arXiv:2304.06906, 2023. 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.687, + 0.482, + 0.74 + ], + "angle": 0, + "content": "[68] Yuan Yao, Tianyu Yu, Ao Zhang, Chongyi Wang, Junbo Cui, Hongji Zhu, Tianchi Cai, Haoyu Li, Weilin Zhao, Zhihui He, et al. Minicpm-v: A gpt-4v level mllm on your phone. arXiv:2408.01800, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.744, + 0.482, + 0.798 + ], + "angle": 0, + "content": "[69] Peter Young, Alice Lai, Micah Hodosh, and Julia Hockenmaier. From image descriptions to visual denotations: New similarity metrics for semantic inference over event descriptions. TACL, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.802, + 0.482, + 0.842 + ], + "angle": 0, + "content": "[70] Bangguo Yu, Hamidreza Kasaei, and Ming Cao. L3mvn: Leveraging large language models for visual target navigation. In IROS, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[71] Jiahui Yu, Zirui Wang, Vijay Vasudevan, Legg Yeung, Mojtaba Seyedhosseini, and Yonghui Wu. Coca: Contrastive captioners are image-text foundation models. arXiv:2205.01917, 2022. 6" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.092, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[72] Weihao Yu, Zhengyuan Yang, Linjie Li, Jianfeng Wang, Kevin Lin, Zicheng Liu, Xinchao Wang, and Lijuan Wang. Mm-vet: Evaluating large multimodal models for integrated capabilities. arXiv:2308.02490, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.149, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[73] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In ICCV, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.193, + 0.905, + 0.233 + ], + "angle": 0, + "content": "[74] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In CVPR, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.235, + 0.905, + 0.276 + ], + "angle": 0, + "content": "[75] Gengze Zhou, Yicong Hong, and Qi Wu. Navgpt: Explicit reasoning in vision-and-language navigation with large language models. In AAAI, 2024. 3" + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.092, + 0.905, + 0.276 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "29666" + } + ] +] \ No newline at end of file diff --git a/2025/SpatialCLIP_ Learning 3D-aware Image Representations from Spatially Discriminative Language/dad5650f-d216-4b30-a0c7-dd7ca6658593_origin.pdf b/2025/SpatialCLIP_ Learning 3D-aware Image Representations from Spatially Discriminative Language/dad5650f-d216-4b30-a0c7-dd7ca6658593_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..4b623b0b8a256b8018a8706554dbfc92b92a1603 --- /dev/null +++ b/2025/SpatialCLIP_ Learning 3D-aware Image Representations from Spatially Discriminative Language/dad5650f-d216-4b30-a0c7-dd7ca6658593_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:88b40cc6adb29eca80fa42a082709aa38b42383df6d4c9c65ef3c7773b38052f +size 4728836 diff --git a/2025/SpatialCLIP_ Learning 3D-aware Image Representations from Spatially Discriminative Language/full.md b/2025/SpatialCLIP_ Learning 3D-aware Image Representations from Spatially Discriminative Language/full.md new file mode 100644 index 0000000000000000000000000000000000000000..b54b8c54a161df8ad1ece72c590624d8bbc3680c --- /dev/null +++ b/2025/SpatialCLIP_ Learning 3D-aware Image Representations from Spatially Discriminative Language/full.md @@ -0,0 +1,395 @@ +# SpatialCLIP: Learning 3D-aware Image Representations from Spatially Discriminative Language + +Zehan Wang $^{1*}$ , Sahuai Zhou $^{1*}$ , Shaoxuan He $^{1}$ , Haifeng Huang $^{1}$ , Lihe Yang $^{2}$ , Ziang Zhang $^{1}$ , Xize Cheng $^{1}$ , Shengpeng Ji $^{1}$ , Tao Jin $^{1}$ , Hengshuang Zhao $^{2}$ , Zhou Zhao $^{1\dagger}$ + +1Zhejiang University; 2The University of Hong Kong https://github.com/SpatialVision/Spatial-CLIP + +# Abstract + +Contrastive Language-Image Pre-training (CLIP) learns robust visual models through language supervision, making it a crucial visual encoding technique for various applications. However, CLIP struggles with comprehending spatial concepts in images, potentially restricting the spatial intelligence of CLIP-based AI systems. In this work, we propose SpatialCLIP, an enhanced version of CLIP with better spatial understanding capabilities. To capture the intricate 3D spatial relationships in images, we improve both "visual model" and "language supervision" of CLIP. Specifically, we design 3D-inspired ViT to replace the standard ViT in CLIP. By lifting 2D image tokens into 3D space and incorporating design insights from point cloud networks, our visual model gains greater potential for spatial perception. Meanwhile, captions with accurate and detailed spatial information are very rare. To explore better language supervision for spatial understanding, we re-caption images and perturb their spatial phrases as negative descriptions, which compels the visual model to seek spatial cues to distinguish these hard negative captions. With the enhanced visual model, we introduce SpatialLLaVA, following the same LLaVA-1.5 training protocol, to investigate the importance of visual representations for MLLM's spatial intelligence. Furthermore, we create SpatialBench, a benchmark specifically designed to evaluate CLIP and MLLM in spatial reasoning. SpatialCLIP and SpatialLLaVA achieve substantial performance improvements, demonstrating stronger capabilities in spatial perception and reasoning, while maintaining comparable results on general-purpose benchmarks. + +# 1. Introduction + +Recently, Contrastive Language-Image Pretraining (CLIP) models [13, 22, 47, 53, 73] have demonstrated the ability to + +learn robust image representations from language supervision. These models serve as foundational visual encoders for various applications, such as open-world visual recognition [26, 36, 40, 46, 49, 52] and Multi-modal Large Language Models (MLLM) [2, 10, 11, 41, 42, 57]. + +Although CLIP exhibits strong zero-shot ability across diverse tasks, it struggles with understanding spatial arrangement within images. As shown in Fig. 1, for a correctly matched image-text pair, we introduce perturbations to the caption by modifying the object category, object attribute, and spatial phrases. The results reveal that altering the object category or attribute significantly reduces the image-text similarity in the CLIP space. However, CLIP struggles to distinguish incorrect spatial relationship descriptions and even assigns a higher matching score to captions with wrong spatial descriptions. Spatial intelligence is a key aspect of visual perception, and limitations in spatial understanding may constrain systems built on CLIP from effectively understanding real-world environments. + +In this paper, we present SpatialCLIP, a CLIP model with enhanced spatial perception capabilities. As titled in [47], CLIP models are "Learning Transferable Visual Models from Natural Language Supervision". Our SpatialCLIP advances the spatial understanding from both "visual models" and "language supervision" perspectives. + +For "visual models," we propose a 3D-inspired ViT that processes RGB-D input rather than RGB images. By incorporating depth information, our visual model gains the ability to perceive images from the 3D perspective. This design helps the model understand complex 3D spatial relationships like "behind," "in front of," "under," and "far." Specifically, we lift the 2D tokens into 3D space using the extra depth information. In the 3D space, we integrate two structures widely used in point cloud networks: 3D convolution [28, 61] and 3D relative position encoding [34, 62, 67], into the pre-trained ViT model [18]. By integrating design insights from 3D point cloud processing, our 3D-inspired ViT demonstrates enhanced potential for capturing complex 3D spatial structures in images. + +For "language supervision", in current vision-language + +Figure 1. The preliminary analysis of OpenAI CLIP-L/14@336px. CLIP Score stands for the cosine similarity between image and language description. These examples demonstrate that while CLIP is highly sensitive to errors involving object categories and attributes, it struggles to accurately identify mistakes in spatial descriptions. +![](images/2cad005fad5393440a393d83a4b69d570032eb6aca59fec42d5a8fff2dd68352.jpg) +- Accurate: The darker-haired cat is behind the white cat. +CLIP Score: 22.28 +Spatial Error: The darker-haired cat is to the right of the white cat. +CLIP Score: 23.66 (+1.38) + +![](images/45ca3f9a068d23df4997ccc971c5fa7aa8345d42f117e54f3ee047f37a1f4467.jpg) +- Category Error: The darker-haired bird is behind the white cat. +CLIP Score: 20.43 (-1.85) +-Accurate: The white refrigerator is closer to the camera than brown door. +CLIP Score: 29.74 +$\nrightarrow$ Category Error: The white cabinet is closer to the camera brown door. CLIP Score: 25.66 (-4.08) +$\nrightarrow$ Attribute Error: The dark refrigerator is closer to the camera than opened door. CLIP Score: 29.39 (-0.35) +Spatial Error: The white refrigerator is further from the camera than brown door. CLIP Score: 29.62 (-0.12) + +![](images/61f6d082312b421fe2fd2ab3ae77f0d8c8effd548c2285aeef62c587c42913d1.jpg) +$\nLeftrightarrow$ Attribute Error: The light-haired cat is behind the white cat. CLIP Score: 21.22 (-1.06) +-Accurate: The dark coffee machine is to the left of the mug. CLIP Score: 22.10 +$\rightarrow$ Category Error: The ceramic cup is to the left of the mug. CLIP Score: 19.48 (-2.62) +$\nrightarrow$ Attribute Error: The old broken coffee machine is to the left of the mug. CLIP Score: 21.27 (-0.83) +Spatial Error: The dark coffee machine is in front of the mug. CLIP Score: 22.53 (+0.43) + +![](images/b800fcb15ff8474828847cc57bb4e5e06c1201a1c420d9e29a4b072a0daf2b6b.jpg) +-Accurate: Two people are standing behind a pile of huge-sized pipes. CLIP Score: 26.62 +$\nvdash$ Category Error: Two people are standing behind a pile of huge-sized papers. CLIP Score: 16.40 (-10.22) +$\nvdash$ Attribute Error: Two people are standing behind a pile of small pipes. +CLIP Score: 25.78 (-0.84) +Spatial Error: Two people are standing inside a pile of huge-sized pipes. +CLIP Score: 27.18 (+0.56) + +datasets [5, 50, 51] for CLIP training, the input captions are primarily simple summary texts that only describe the primary object category and attributes. The absence of precise spatial information in these captions may constrain the visual model's ability to capture spatial relationships. To enhance language supervision, we recapture images using advanced MLLMs [9, 41] and introduce perturbed spatial phrases to captions as hard negative queries. Training the model to differentiate between captions that share the same object categories and attributes but vary in spatial phrases forces it to better capture spatial cues in images. + +Furthermore, we explore whether SpatialCLIP, with its enhanced spatial perception, can contribute stronger spatial intelligence to MLLMs when using the same data. Recent work [7, 12, 14] mainly enhances the spatial perception capabilities of MLLMs by introducing new training data, while lacking the exploration of spatial-focused visual representations. However, as discussed in [58], visual representation is also crucial to the capabilities of MLLMs. We construct SpatialLLaVA using the same training recipe as the standard LLaVA-1.5 [41], enabling a fair and systematical analysis over MLLMs. + +To quantitatively assess the spatial understanding capabilities of CLIP and MLLM, we introduce SpatialBench, a benchmark designed to evaluate spatial comprehension within images. This benchmark challenges models to comprehend spatial arrangements and perform complex spatial reasoning. For thorough evaluation, we curate and annotate images from both outdoor and indoor scenes. + +Experimentally, SpatialCLIP shows substantial improvements across spatial-focused benchmarks, such as Spatial-Bench and BLINK [24], and performs well on general-purpose downstream tasks [39, 69]. Similarly, SpatialLLaVA demonstrates notably enhanced spatial perception while retaining strong general perceptual abilities. + +Our main contributions are as follows: + +- We propose SpatialCLIP, which improves the CLIP's spatial understanding capabilities by refining both "visual model" and "language supervision". +- We develop SpatialLLaVA, by replacing the visual encoding method in LLaVA with our SpatialCLIP while preserving other training protocols. This investigates the value of SpatialCLIP in enhancing the spatial intelligence of CLIP-based MLLM, as shown in Fig. 2. +- We construct SpatialBench to evaluate the spatial capabilities of CLIP and MLLMs, collecting and annotating samples from both outdoor and indoor scenarios. +- We conduct extensive experiments on both spatial-focused and general-purpose benchmarks to validate the effectiveness of SpatialCLIP, SpatialLLaVA, and each newly proposed component. + +# 2. Related Work + +# 2.1. Contrastive Language-Image Pre-training + +CLIP models [13, 22, 47, 53] are trained on web-scale image-text paired data [5, 50, 51] using multimodal contrastive learning as objective. By learning cross-modal semantic alignment from massive datasets, CLIP demonstrates strong zero-shot generalization capabilities and is widely used in various vision-language downstream tasks, such as image object detection [52], segmentation [48, 60], and video action recognition [40]. In recent multimodal large language models (MLLM), CLIP is the most common visual encoding strategy, enabling large language models with the ability to understand images. LLaVA [42] and subsequent works [35, 35, 41, 64] primarily rely on OpenAI-CLIP [47], while other state-of-the-art MLLMs like Emu2 [54], InternVL1.5 [10] and MiniCPM-v [68] also utilize various CLIP variants, such as EVA-CLIP [53], In + +![](images/addc6756b77131b54ba6cf9c60030b1cdb6f8d02f256910abd1d371015b51030.jpg) +Select the best caption for the image: + +![](images/23424863be70cc3457b3dbe5399bca59f30fd6007034bac4614ffab170f308a5.jpg) +B. Red electric car is closer to the camera. +LLaVA-1.5: The best caption is B. +SpatialLLaVA: The best caption is A. +Select the best caption for the image: +A. The fan is facing the camera. +B. The fan is facing the white desk. +C. The fan is facing the sofa. +D. The fan is facing the chair. +LLaVA-1.5: The best caption is A. +SpatialLLaVA: The best caption is D. +Figure 2. Visualization of LLaVA-1.5 and SpatialLLaVA on SpatialBench. The first row displays examples from outdoor scenes, while the second row features cases in indoor environments. LLaVA-1.5 often encounters difficulties with basic spatial concepts, while SpatialLLaVA improves its spatial perception ability by employing enhanced visual representations. + +![](images/4022cc6b053062907a26b5449eac47bdefc6d1a222aa516f872dcd9e7c40557a.jpg) +A. Yellow electric car is closer to the camera. +Select the best caption for the image: A. The bird standing taller is smaller than the other. +B. The bird standing taller is bigger than the other. +C. The two birds are roughly the same size. + +![](images/38e9ac6842ef8ffd5194849f124aecda9d165f779b8b423f64918cfb6355b7f6.jpg) +C. Two cars are equally close to the camera. +LLaVA-1.5: The best caption is B. +SpatialLLaVA: The best caption is C. +Select the best caption for the image: +A. All the sofas form a circle +B. All the sofas form a rectangle. +C. All the sofas form an oval. +D. All sofas are arranged randomly. +LLaVA-1.5: The best caption is A. +SpatialLLaVA: The best caption is B. + +![](images/b4589a4fd056cc1756b66a0016a24460867eca3629756fb45916b25e4204ec10.jpg) +Select the best caption for the image: +A. The two trains are perpendicular. +B. The two trains are intersecting. +C. The two trains are parallel. +A. The stop sign is taller than the wall. +B. The stop sign is shorter than the wall. +C. The stop sign and the wall are of the same height. + +![](images/a8b1b29f9decad6790b671733642171920730598399f53d1d9140a9dc62c48ef.jpg) +LLaVA-1.5: The best caption is A. +SpatialLLaVA: The best caption is C. +Select the best caption for the image: +A. Ceiling light hangs above the edge of the table. +B. Ceiling light hangs above the center of the table. + +![](images/3e1822acb1b71751ba1e2efd3fc7dbeeabae72b2a705aa9ab09d319e1f10830a.jpg) +Select the best caption for the image: + +![](images/b458a8c4f2a4f16ef29591dd4b360c22e474b4e7749acad2ebf39a7a3881371e.jpg) +LLaVA-1.5: The best caption is C. +SpatialLLaVA: The best caption is A. +Select the best caption for the image: +A. The left window is larger than the right window. +B. The right window is larger than the left window. +C. The Two mirrors are of the same size. +LLaVA-1.5: The best caption is A. +SpatialLLaVA: The best caption is C. + +ternVL [11], and SigLIP [73]. However, ReCLIP [52] recognizes CLIP lacks the capability of understanding spatial relationships in images. Additionally, as discussed in [58], the shortcoming in visual encoding may limit the capabilities of MLLMs built upon it. + +Our work attempts to enhance CLIP's spatial perception and overcome the shortcomings in spatial understanding. + +# 2.2. Spatial Intelligence in Visual Perception + +Spatial understanding [8, 56] and spatial intelligence are fundamental abilities for artificial intelligence systems to perceive and interact with the complex visual world. In our 3D world, spatial understanding involves distinguishing various spatial relationships between objects, such as up and down, left and right, far and near, and inside and outside, as well as perceiving the size, thickness, length, and width of visual elements. The ability to comprehend and reason on these basic spatial concepts is crucial for solving many complex problems, including robot navigation [70, 75], manipulation [25], human-computer interaction [19, 44], and applications in virtual and augmented reality [33]. + +MLLMs have made impressive progress in recent years. Built upon pre-trained visual representation and large language models [20, 59], these methods demonstrate strong + +performance across a variety of downstream tasks, such as image captioning [39], visual question answering [1], visual grounding [30, 45], and action recognition [23]. However, despite their powerful generalization capabilities, most MLLMs still struggle with spatial understanding and reasoning, as discussed in [7]. + +There are several works that attempt to address the shortcomings of MLLMs in spatial understanding. SpatialVLM [7] is the first to focus on this issue, introducing massive automatically-generated 3D spatial reasoning VQA data for training MLLM. Its successor, SpatialRGPT [12], enhances the 3D-aware data curating pipeline and provides extra visual depth and region information for MLLM. CubeLLM [14] combines multiple existing 2D and 3D object detection datasets in a unified format and then trains MLLM with better spatial detection capability. + +Existing works mainly enhance the spatial intelligence of MLLM by improving the training data, while our work explores the issue from a different perspective: improving the visual representations. Utilizing the same general-purpose training data, simply replacing the standard CLIP with our SpatialCLIP enhances the spatial understanding capabilities of LLaVA [41, 42]. + +![](images/357f47b7c6ca1367b8c172971e79c111504f680c575f5bdc4e53808a860abd0a.jpg) +Figure 3. The overview of SpatialCLIP. The visual encoder is replaced by our newly proposed 3D-Inspired Vision Transformer, which enables more informative RGB-D inputs and excels in modeling the 3D spatial structure. Language supervision is enhanced through recaption and perturbation, which prompt the encoding of visual spatial information. + +![](images/4b3aa839cd88e41a94e5b2a1dd2b497ca1a672099bc85394228d8183089850bb.jpg) + +# 3. SpatialCLIP + +Our proposed SpatialCLIP enhances standard CLIP from two perspectives: 3D-Inspired ViT for more powerful "Visual Model" in Sec. 3.1 and recaption&perturbation strategy for more informative "Language Supervision" in Sec. 3.2. + +# 3.1. Visual Model: 3D-Inspired ViT + +In RGB images, the 3D spatial structure of the world is flattened into a 2D plane. Many methods [12, 15, 17, 74] try to introduce depth maps for extra 3D information. However, how to effectively encode such RGB-D inputs while maintaining the pre-trained knowledge for the RGB channel is non-trivial. AlphaCLIP [55] enables an extra input channel with a simple extra input convolution layer, while SpatialRGPT [12] repeats depth map three times as a depth image, and lately fuses the latent of the RGB image and depth image with the pre-trained CLIP image encoder. Our empirical findings show that such simple early and late fusion methods are less effective at integrating depth information. To address this, we propose 3D-inspired ViT for RGB-D inputs to replace the standard vision transformer (ViT) widely used as the visual model in CLIP models. + +Recently, monocular depth estimation models [31, 65, 66] have gained impressive progress, depth maps can be easily predicted for wild-world RGB images with low cost. Moreover, the accurate depth map shows impressive results in lifting the images into 3D point clouds. Meanwhile, the architectures for processing 3D point cloud [61, 62, 67] focus on capturing spatial relationships among features in 3D space. These observations inspire us to lift 2D tokens into 3D space using predicted depth information and thereby incorporate design insights from 3D point cloud architectures + +to better model 3D spatial relationships in images. + +To this end, we first dive into the standard ViT, which processes RGB images as input. Specifically, an RGB image $\mathbf{X} \in \mathbb{R}^{H \times W \times 3}$ is divided into 2D patches, and each is linearly projected into $C$ -dimension tokens. The resulting 2D tokens map $\mathbf{Z} \in \mathbb{R}^{H_z \times W_z \times C}$ is passed into a stack of transformer blocks, where $H_{z}$ and $W_{z}$ represent the tokens map's height and width. The 2D token map for the $L$ -th transformer block can be denoted as $\mathbf{Z}_{L} \in \mathbb{R}^{H_{z} \times W_{z} \times C_{L}}$ . + +To integrate the well-established 3D point cloud architecture designs into ViT for images, for each transformer layer, we first lift its 2D input tokens $\mathbf{Z}_L \in \mathbb{R}^{H_z \times W_z \times C_L}$ into 3D voxel $\mathbf{Z}_L^{3\mathrm{D}} \in \mathbb{R}^{H_z \times W_z \times D_z \times C_L}$ based on the average depth value of each patch, where $H_z$ , $W_z$ , and $D_z$ represent the grid sizes along the height, width, and depth axes. Since $\mathbf{Z}_L^{3\mathrm{D}}$ is derived from $\mathbf{Z}_L$ , the resulting 3D voxel grid is highly sparse. The $H_z \times W_z$ tokens in $\mathbf{Z}_L$ are distributed along the depth axis based on their corresponding depth values, and extracting tokens along the depth axis can easily compress $\mathbf{Z}_L^{3\mathrm{D}}$ back into 2D space. + +For 3D voxelized tokens, 3D depth-wise convolution is widely used to capture spatial patterns across the three dimensions. As shown in the right part of Fig. 3, we integrate this 3D convolution within each transformer block of the pre-trained ViT model, applying it to $\mathbf{Z}_L^{3\mathrm{D}}$ and then short-cutting the resulting features back to the original $\mathbf{Z}_L$ . This process is expressed as: + +$$ +\mathbf {Z} _ {L} = \mathbf {Z} _ {L} + \operatorname {C o m p r e s s} \left(3 \mathrm {D} - \operatorname {C o n v} \left(\mathbf {Z} _ {\mathrm {L}} ^ {3 \mathrm {D}}\right)\right) \tag {1} +$$ + +where 3D-Conv $(\cdot)$ is the 3D depth wise convolution and Compress $(\cdot)$ indicates compressing sparse 3D voxel back into 2D token maps. + +Moreover, another commonly used technique in point cloud networks is 3D relative positional encoding, which explicitly indicates the relative spatial relationship between every pair of tokens in 3D space. For the self-attention operator of each transformer layer, we employ 3D relative positional biases on the attention map. The attention score $a_{ij}$ between $i$ -th and $j$ -th token can be represented as: + +$$ +a _ {i j} = \frac {\left(\mathbf {z} _ {i} \mathbf {W} ^ {Q}\right) \left(\mathbf {z} _ {j} \mathbf {W} ^ {K}\right) ^ {T} + b _ {i j}}{\sqrt {C}} \tag {2} +$$ + +where $b_{ij}$ is a scalar selected from the learnable bias matrix $\mathbf{B} \in \mathbb{R}^{(2H_z - 1) \times (2W_z - 1) \times (2D_z - 1)}$ , based on the relative 3D position between the $i$ -th and $j$ -th tokens in 3D space. + +By adopting the designs inspired by the structure of 3D point cloud networks, the 3D-inspired ViT can perceive 2D images from the 3D perspective, greatly enhancing the potential to model complex spatial relationships in images. + +# 3.2. Language Supervision: Recapiction & Perturbation + +In the contrastive learning process of CLIP, visual representations of different images are pulled close with the paired caption while pushed away from captions of other images. The guidance information in language captions strongly influences the visual model's capability. For prompting 3D-Inspired ViT to better capture spatial relationships in images, more informative supervision from natural language is necessary. However, current widely-used vision-language datasets for training CLIP, such as Conceptual-12M [5] and LAION [50, 51], are inherently noisy and mainly provide simple summaries of global attributes, lacking detailed caption with accurate spatial description. + +Recent studies [3, 6, 21, 37] have shown that recaptioning images with semantically richer and more aligned descriptions can greatly enhance the performance of vision-language models like CLIP and text-to-image models. However, as discussed in [7, 12], current MLLMs excel at recognizing object categories and attributes but struggle with spatial perception. Consequently, the detailed captions they generated often fail to accurately capture all spatial relationships within the image. + +To address this "chicken-and-egg" dilemma, we propose perturbing the spatial descriptions in the dense captions generated by MLLMs. Specifically, we employ a template to randomly substitute phrases indicating spatial relationships with their antonymic counterparts. For instance, a caption like "two cats are standing in front of a table, while the left cat is bigger than the right one" may be modified to "two cats are standing under a table, while the upper cat is smaller than the lower one." + +Since the generated and perturbed captions are semantically consistent in terms of object categories and attributes, applying contrastive learning between these two captions + +encourages the model to distinguish the entirely incorrect caption from the partially correct one. Such a straightforward perturbation method effectively promotes the visual model's ability to capture detailed spatial relationships within images. The learning objective can be expressed as: + +$$ +\begin{array}{l} L = - \frac {1}{N} \sum_ {i = 1} ^ {N} \left[ \log \frac {e ^ {\left(\mathbf {t} _ {i} \cdot \mathbf {v} _ {i} / \tau\right)}}{\sum_ {j = 1} ^ {N} e ^ {\left(\mathbf {t} _ {i} \cdot \mathbf {v} _ {j} / \tau\right)}} \right. \tag {3} \\ \left. + \log \frac {e ^ {\left(\mathbf {t} _ {i} \cdot \mathbf {v} _ {i} / \tau\right)}}{\sum_ {j = 1} ^ {N} \left(e ^ {\left(\mathbf {t} _ {j} \cdot \mathbf {v} _ {i} / \tau\right)} + e ^ {\left(\tilde {\mathbf {t}} _ {j} \cdot \mathbf {v} _ {i} / \tau\right)}\right)} \right] \\ \end{array} +$$ + +where the $v_{i}$ is the visual representation of $i$ -th image, while $t_{i}$ and $\tilde{t}_i$ denote the textual representation of the $i$ -th caption and its perturbed version, respectively. + +# 4. SpatialLLaVA + +As discussed in [57, 58], the capabilities of MLLMs may be constrained by the inherent limitations of their visual representations. This raises a crucial question: Can the spatial intelligence of CLIP-based MLLM be enhanced by improving the spatial perception of CLIP's visual representations? + +We utilize LLaVA-1.5, which employs the pre-trained CLIP as its visual encoding model, as the framework to investigate the effectiveness of SpatialCLIP in MLLM. Specifically, we replace the CLIP model in LLaVA-1.5 [41] with our SpatialCLIP. To ensure fair and systematic comparisons between visual encoding methods, we follow the two-stage training strategy used in LLaVA-1.5 and apply the same training datasets and hyperparameters for each stage. The resulting MLLM is named SpatialLLaVA. + +# 5. SpatialBench + +To evaluate the ability of CLIP and MLLM in understanding complex 2D/3D spatial relations in images, we build a benchmark SpatialBench, which focuses entirely on spatial perception. Our benchmark employs the caption matching task, requiring CLIP or MLLM to distinguish the caption that best matches the image from several distractors. By manually crafting the correct description and distractors, the caption-matching task requires the model to distinguish nuanced spatial concepts in the image (e.g. object arrangement, relationships, relative depth, and size). + +To fully assess models in various scenarios, we collect images of both outdoor and indoor scenes. Specifically, outdoor images are collected from the COCO [39] validation set and the Internet, totaling 159 images. Indoor images are sourced from ScanNet [16], consisting of 140 images. Each image is paired with 2 to 4 manually crafted descriptions, of which only one is accurate, with the remaining serving as distractors. Examples of SpatialBench can be found in Fig. 2 and Appendix. + +
ModelParamSpatial-FocusedGeneral-Purposed
SpatialBenchBLINKCOCOFlickr30K
OutDoor.InDoor.SpatialR@1R@5R@1R@5
Random-29.7729.8450.00----
OpenAI CLIP-B@224px149.6M35.2230.7151.0542.1967.2534.4056.65
SpatialCLIP-B@224px149.9M+24.8M41.5134.2857.3448.3373.2738.8662.80
OpenAI CLIP-L@336px427.9M32.0730.0047.5546.7870.5438.2560.95
SpatialCLIP-L@336px428.7M+24.8M35.8433.5757.3450.6275.6642.5467.04
+ +Table 1. Results on CLIP benchmark. The best results are bold. + +
Training DataModelSpatialBenchBLINKCV-Bench3d
OutDoor.InDoor.SpatialDepthCorresJigsawDepthDistance
Spatial-FocusedSpatial-RGPT-7B37.7338.5765.7382.3032.5550.0062.3359.00
Stronger General-PurposedLLaVA-1.6-7B39.6233.5768.5351.6129.6548.0053.3551.12
LLaVA-1.6-13B38.3635.7169.9948.3830.8150.6755.1652.25
LLaVA-interleave-7B33.2737.1469.9356.6934.8851.3357.1462.30
Baseline General-PurposedLLaVA-1.5-7B29.5532.4761.5452.4225.5811.3345.8350.00
Spatial-LLaVA-7B40.2540.1771.3258.0631.9751.3357.3352.16
+ +Table 2. Results on Spatial-Focused MLLM benchmark. + +
POPE [38]MM.Vet [72]VOAv2 [27]GOA [29]RealWorldQA [63]
LLaVA-1.5-7B85.930.578.562.054.8
Spatial-LLaVA-7B86.530.779.960.654.2
+ +Table 3. Results on General-Purposed MLLM benchmark. + +For CLIP, we calculate matching scores between the image and its corresponding captions, selecting the description with the highest matching score as the prediction. For MLLMs, the descriptions are presented as multiple-choice options, and the models are prompted to select the best description in a question-answering manner. + +# 6. Experiment + +# 6.1. Implementation Details + +Training setting. We use images in SA1B dataset [32] for training, since most images in it contain multiple objects and showcase complex spatial relationships. Following [6], LLaVA-1.5-13B [41], CoCa [71] and extra ShareGPT4V [9] are used for image reception, the three kinds of captions are randomly used in training for more diverse and informative supervision. A detailed caption perturbation template is provided in the Appendix. The tem + +perature factor $\tau$ in Eq. 3 is set to 0.01. For optimization, we employ AdamW [43] with a learning rate of 2e-7 for pre-trained CLIP encoders and 4e-6 for the newly introduced components. Models are trained for 8,000 steps with a batch size of 2,048 on 8 A100 GPUs. + +Model setting. We train SpatialCLIP from OpenAI CLIP-B/16@224px and OpenAI CLIP-L/14@336px, leading to SpatialCLIP-B and SpatialCLIP-L, respectively. We utilize the lightweight Depth Anything v2 ViT-Small as our monocular depth estimation model, which brings an extra 24.8M parameters. The kernel size of the 3D CNN in Eq. 1 is 3, and the $D_w$ is 6. The newly added 3D CNN and 3D relative position encoding are both zero-initialized and introduce only nearly 1M extra parameters. We develop SpatialLLaVA based on SpatialCLIP-L, following LLaVA-1.5. + +# 6.2. Comparison for CLIP + +Results on spatial-focused benchmark. To assess the spatial perception ability for CLIP models, we employ our newly proposed SpatialBench and modify the "spatial relation" subset of BLINK to the caption-matching task, as another spatial-focused benchmark. Results on these benchmarks are presented in Tab. 1. There are two key findings: 1) In both the CLIP-B and -L variants, SpatialCLIP demonstrates significant performance advantages over the original CLIP, indicating that SpatialCLIP more effectively captures spatial information in images and distinguishes between + +![](images/7cb0b2214567b453bb2c8385ca2ca038b5a9a7416420443838b8cd938ee37cb8.jpg) + +LLaVA: The banana nut bread is on the left side of the picture, with a banana and an egg on the right side. The banana is placed in a blue bowl next to the bread. The eggs are in a glass bowl on the left side of the picture. The banana nut bread is placed in a brown paper bag, with a nut on top. + +SpatialLLaVA: In the image, a bag of banana nut bread mix is placed next to a banana, an egg, and some nuts. The banana is positioned on the right side of the image, while the egg is located on the left side, and the nuts are scattered around the bag of mix. The bag of mix is placed in the middle of the scene, with the banana, egg, and nuts surrounding it. + +LLaVA: There are two cars that are oriented in a vertical position. The green bus is on the left side of the picture, while the red bus is on the right side. The front of the green bus is facing toward the viewer, while the red bus is facing away from the viewer. + +SpatialLLaVA: In the image, the two buses are positioned side by side, with one bus on the left and the other on the right. They are both facing the same direction, which suggests that they are either parked next to each other or driving in the same direction on the road. + +![](images/96df64351e5387fae6b5a8d2baa45d329fdc6ce5053bda13234664619b26d242.jpg) +Figure 4. Qualitative comparison between LLaVA-1.5 and SpatialLLaVA. We prompt these two MLLMs to provide detailed descriptions of spatial relationships in images. Our model perform better at capturing these spatial concepts. + +different spatial concepts. 2) For both the original CLIP and SpatialCLIP, performance on spatial-focused benchmarks is not correlated with model size. We conjecture this situation is derived from the scarcity of the spatial concepts in the current CLIP pre-training dataset. Increasing the parameter (i.e., enhancing learning capacity) does not necessarily lead to improved spatial perception. This highlights the value of additional designs for better spatial perception, rather than merely scaling up within the existing paradigm. + +Results on general-purposed benchmark. In addition to enhancing spatial concept comprehension, accurate recognition of object categories and attributes is also essential for advanced applications. Since SpatialCLIP is built upon pretrained CLIP and is further trained with high-quality images and informative captions generated by diverse MLLMs, we also observe significant performance improvements on general-purpose benchmarks, like COCO and Flickr30K for zero-shot image-text retrieval. + +# 6.3. Comparison for MLLM + +Results on spatial-focused benchmark. In Tab. 2, we comprehensively validate the spatial perception and reasoning ability of our SpatialLLaVA on various spatial-focused MLLM benchmarks. With the same training data, SpatialLLaVA achieves much better results than LLaVA-1.5. SpatialLLaVA exhibits a stronger capability to differentiate between nuanced spatial concepts across indoor and outdoor scenes on our SpatialBench. + +We also conduct absolute performance comparison with more advanced MLLMs, including Spatial-RGPT (trained on spatial-focused data) and LLaVA-1.6/LLaVA-Interleave + +(trained on much broader general-purpose data). Despite this "unfair" comparison, Spatial-LLaVA consistently shows superior spatial perception. Interestingly, Spatial-RGPT excels in tasks it was explicitly trained for (e.g., depth and distance estimation) but underperforms on tasks beyond its training scope (e.g., spatial question-answering). This observation highlights the need to enhance visual perception capabilities in MLLMs beyond simply expanding training data. + +The advantages on BLINK underscore the importance of capturing spatial structure for high-level applications. For "visual corresponding" task, capturing spatial structure is beneficial for identifying the same scene points across different viewpoints, lighting, or times, as the scene structure often remains consistent despite these changes. For "Jigsaw," viewing 2D images in the 3D perspective allows the model to inspect the continuity of the 3D spatial structure rather than only relying on color and texture, aiding in organizing image patches. + +CV-Bench[57] collects and annotates samples from the 3D object detection dataset Omni3D [4], focusing on 3D understanding. Employing SpatialCLIP also leads to performance improvements for MLLM on this benchmark, further demonstrating the superiority of SpatialCLIP's visual representations in capturing 3D spatial relationships. + +Results on general-purposed benchmark. In Tab. 3, we examine how enhancing spatial perception in visual representations affect the performance of MLLMs in general-purposed scenarios. Across five benchmarks, Spatial-LLaVA achieves comparable performance to LLaVA-1.5 for general-purposed questions. + +
RowTuning3D CNN3D RPEPerturbSpatialBenchBLINKCOCOFlickr30k
OutDoor.InDoor.Spatial.R@1R@1
1×---35.2231.4351.0542.1934.40
2✓×××35.8531.4351.7548.9939.06
3✓✓××38.3631.4353.8549.0437.76
4✓×✓×37.1032.1454.5547.8139.33
5✓✓✓×38.3633.5755.2447.6737.56
6✓✓✓✓41.5134.2857.3448.3438.86
+ +Table 4. Ablation study of our "Visual Model" and "Language Supervision". All the experiments are conducted on SpatialCLIP-B. + +
SettingSpatialBenchBLINKCOCOFlickr30k
OutDoorInDoorSpatialR@1R@1
OpenAI CLIP-B35.2231.4351.0542.1934.40
Early Fusion39.6232.8551.7543.0133.09
Late Fusion40.2534.2852.4542.3831.80
3D-Inspried ViT41.5134.2857.3448.3438.86
+ +Table 5. Ablation study of depth channel integration strategies. + +Qualitative comparison. To better demonstrate the improvement of SpatialLLaVA in spatial perception, we provide visualizations of both models in Fig. 2 and 4. The qualitative comparison of their responses reveals that SpatialLLaVA can more accurately and comprehensively capture the spatial relationships between objects in images. + +# 6.4. Ablation Study + +Effect of recaption. In the first two rows of Tab. 4, we present the comparison between the original CLIP and its direct fine-tuning using our recaption dataset. On the general-purposed benchmarks, COCO and Flickr30k, tuning on our dataset leads to significant performance improvements, underscoring the benefits of high information density and semantic alignment in recaption data. However, simply using the recaption data only brings negligible changes to the spatial-focused benchmarks, motivating us to introduce more specific designs to learn spatial concepts. + +Effect of 3D-inspired ViT. We ablate the designs of 3D-Inspired ViT in Row 2-5 of Tab. 4. When fine-tuning on the same recaption data, incorporating 3D convolution (Row 3) or 3D relative position encoding (Row 4) shows improved performance on spatial-focused benchmarks compared to the standard ViT (Row 2). The combination of both techniques results in even greater enhancements. The improvement in spatial perception gained from the 3D-inspired design underscores the importance of integrating 3D point cloud structures into 2D ViT. Encoding images in 3D space with 3D-specific structure enables the visual model to capture complex 3D spatial information more effectively. + +Effect of caption perturbation. The comparison between Rows 5 and 6 in Tab. 4 illustrates the benefits of caption perturbation, which prompts the model to capture spatial concepts and enhances its spatial perception ability. Since MLLMs themselves often struggle in perceiving and reasoning about spatial relationships, the captions generated by them can not fully capture all spatial details. By perturbing the spatial phrases, we create references that are fully incorrect in spatial relationships. This approach drives the model to learn spatial concepts by distinguishing the partially correct captions from the fully incorrect ones. + +Different depth channel integration strategies. In addition to our 3D-Inspired ViT, there are other alternative methods to enable RGB-D inputs. In Tab. 5, we discuss the commonly used early and late fusion schemes, which project the depth map into RGB features before and after the visual encoder, which are used by AlphaCLIP [55] and SpatialRGPT [12], respectively. In most benchmarks, simply projecting the depth channel to the pre-trained RGB features is hard to improve the performance. Considering the additional information provided by the reception data, these schemes may even have a negative impact. Conversely, our 3D-inspired ViT makes effective use of the extra depth information, yielding significant improvements. + +# 7. Conclusion + +We present SpatialCLIP to overcome the shortcomings of original CLIP models in spatial understanding. By introducing improved "Visual Model" (3D-Inspired ViT) and enhanced "Language Supervision" (Recap tion & Perturbation), our model achieves significant improvements, not just on spatial-focused benchmarks like our proposed Spatial-Bench and modified BLINK, but also on zero-shot image-text retrieval. Additionally, we develop SpatialLLaVA based on SpatialCLIP, adhering to the LLaVA-1.5 training protocol. This integration of visual representations specifically enhanced for spatial recognition improves the spatial intelligence of MLLM. + +# Acknowledgements + +This work was supported in part by National Natural Science Foundation of China under Grant No.624B2128 and No.62222211. + +# References + +[1] Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C Lawrence Zitnick, and Devi Parikh. Vqa: Visual question answering. In ICCV, 2015. 3 +[2] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond. arXiv:2308.12966, 2023. 1 +[3] James Betker, Gabriel Goh, Li Jing, Tim Brooks, Jianfeng Wang, Linjie Li, Long Ouyang, Juntang Zhuang, Joyce Lee, Yufei Guo, et al. Improving image generation with better captions. https://cdn.openai.com/papers/dall-e-3.pdf, 2023.5 +[4] Garrick Brazil, Abhinav Kumar, Julian Straub, Nikhila Ravi, Justin Johnson, and Georgia Gkioxari. Omni3d: A large benchmark and model for 3d object detection in the wild. In CVPR, 2023. 7 +[5] Soravit Changpinyo, Piyush Sharma, Nan Ding, and Radu Soricut. Conceptual 12m: Pushing web-scale image-text pretraining to recognize long-tail visual concepts. In CVPR, 2021. 2, 5 +[6] Agneet Chatterjee, Gabriela Ben Melech Stan, Estelle Aflalo, Sayak Paul, Dhruba Ghosh, Tejas Gokhale, Ludwig Schmidt, Hannaneh Hajishirzi, Vasudev Lal, Chitta Baral, et al. Getting it right: Improving spatial consistency in text-to-image models. In ECCV, 2025. 5, 6 +[7] Boyuan Chen, Zhuo Xu, Sean Kirmani, Brain Ichter, Dorsa Sadigh, Leonidas Guibas, and Fei Xia. Spatialvlm: Endowing vision-language models with spatial reasoning capabilities. In CVPR, 2024. 2, 3, 5 +[8] Dave Zhenyu Chen, Angel X Chang, and Matthias Nießner. Scanrefer: 3d object localization in rgb-d scans using natural language. In ECCV, 2020. 3 +[9] Lin Chen, Jinsong Li, Xiaoyi Dong, Pan Zhang, Conghui He, Jiaqi Wang, Feng Zhao, and Dahua Lin. Sharegpt4v: Improving large multi-modal models with better captions. arXiv:2311.12793, 2023. 2, 6 +[10] Zhe Chen, Weiyun Wang, Hao Tian, Shenglong Ye, Zhang-wei Gao, Erfei Cui, Wenwen Tong, Kongzhi Hu, Jiapeng Luo, Zheng Ma, et al. How far are we to gpt-4v? closing the gap to commercial multimodal models with open-source suites. arXiv:2404.16821, 2024. 1, 2 +[11] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. In CVPR, 2024. 1, 3 +[12] An-Chieh Cheng, Hongxu Yin, Yang Fu, Qiushan Guo, Ruihan Yang, Jan Kautz, Xiaolong Wang, and Sifei Liu. Spatial- + +rgpt: Grounded spatial reasoning in vision language model. arXiv:2406.01584, 2024. 2, 3, 4, 5, 8 +[13] Mehdi Cherti, Romain Beaumont, Ross Wightman, Mitchell Wortsman, Gabriel Ilharco, Cade Gordon, Christoph Schuhmann, Ludwig Schmidt, and Jenia Jitsev. Reproducible scaling laws for contrastive language-image learning. In CVPR, 2023. 1, 2 +[14] Jang Hyun Cho, Boris Ivanovic, Yulong Cao, Edward Schmerling, Yue Wang, Xinshuo Weng, Boyi Li, Yurong You, Philipp Krahenbuhl, Yan Wang, et al. Language-image models with 3d understanding. arXiv:2405.03685, 2024. 2, 3 +[15] Jaeyoung Chung, Jeongtaek Oh, and Kyoung Mu Lee. Depth-regularized optimization for 3d gaussian splatting in few-shot images. In CVPR, 2024. 4 +[16] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In CVPR, 2017. 5 +[17] Kangle Deng, Andrew Liu, Jun-Yan Zhu, and Deva Ramanan. Depth-supervised nerf: Fewer views and faster training for free. In CVPR, 2022. 4 +[18] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv:2010.11929, 2020.1 +[19] Danny Driess, Fei Xia, Mehdi SM Sajjadi, Corey Lynch, Aakanksha Chowdhery, Brian Ichter, Ayzaan Wahid, Jonathan Tompson, Quan Vuong, Tianhe Yu, et al. Palm-e: An embodied multimodal language model. arXiv:2303.03378, 2023. 3 +[20] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv:2407.21783, 2024. 3 +[21] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In ICML, 2024. 5 +[22] Alex Fang, Albin Madappally Jose, Amit Jain, Ludwig Schmidt, Alexander Toshev, and Vaishaal Shankar. Data filtering networks. arXiv:2309.17425, 2023. 1, 2 +[23] Chaoyou Fu, Yuhan Dai, Yondong Luo, Lei Li, Shuhuai Ren, Renrui Zhang, Zihan Wang, Chenyu Zhou, Yunhang Shen, Mengdan Zhang, et al. Video-mme: The first-ever comprehensive evaluation benchmark of multi-modal llms in video analysis. arXiv:2405.21075, 2024. 3 +[24] Xingyu Fu, Yushi Hu, Bangzheng Li, Yu Feng, Haoyu Wang, Xudong Lin, Dan Roth, Noah A Smith, Wei-Chiu Ma, and Ranjay Krishna. Blink: Multimodal large language models can see but not perceive. arXiv:2404.12390, 2024. 2 +[25] Jensen Gao, Bidipta Sarkar, Fei Xia, Ted Xiao, Jiajun Wu, Brian Ichter, Anirudha Majumdar, and Dorsa Sadigh. Physically grounded vision-language models for robotic manipulation. In ICRA, 2024. 3 +[26] Rohit Girdhar, Alaaeldin El-Nouby, Zhuang Liu, Mannat Singh, Kalyan Vasudev Alwala, Armand Joulin, and Ishan + +Misra. Imagebind: One embedding space to bind them all. In CVPR, 2023. 1 +[27] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Bartra, and Devi Parikh. Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In CVPR, 2017. 6 +[28] Benjamin Graham, Martin Engelcke, and Laurens Van Der Maaten. 3d semantic segmentation with submanifold sparse convolutional networks. In CVPR, 2018. 1 +[29] Drew A Hudson and Christopher D Manning. Gqa: A new dataset for real-world visual reasoning and compositional question answering. In CVPR, 2019. 6 +[30] Sahar Kazemzadeh, Vicente Ordonez, Mark Matten, and Tamara Berg. Referitag: Referring to objects in photographs of natural scenes. In EMNLP, 2014. 3 +[31] Bingxin Ke, Anton Obukhov, Shengyu Huang, Nando Metzger, Rodrigo Caye Daudt, and Konrad Schindler. Repurposing diffusion-based image generators for monocular depth estimation. In CVPR, 2024. 4 +[32] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In ICCV, 2023. 6 +[33] Mikhail Konenkov, Artem Lykov, Daria Trinitatova, and Dzmitry Tsetserukou. Vr-gpt: Visual language model for intelligent virtual reality applications. arXiv:2405.11537, 2024.3 +[34] Xin Lai, Jianhui Liu, Li Jiang, Liwei Wang, Hengshuang Zhao, Shu Liu, Xiaojuan Qi, and Jiaya Jia. Stratified transformer for 3d point cloud segmentation. In CVPR, 2022. 1 +[35] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. Llava-onevision: Easy visual task transfer. arXiv:2408.03326, 2024. 2 +[36] Liunian Harold Li, Pengchuan Zhang, Haotian Zhang, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, et al. Grounded language-image pre-training. In CVPR, 2022. 1 +[37] Xianhang Li, Haoqin Tu, Mude Hui, Zeyu Wang, Bingchen Zhao, Junfei Xiao, Sucheng Ren, Jieru Mei, Qing Liu, Huangjie Zheng, et al. What if we recapture billions of web images with llama-3? arXiv:2406.08478, 2024. 5 +[38] Yifan Li, Yifan Du, Kun Zhou, Jinpeng Wang, Wayne Xin Zhao, and Ji-Rong Wen. Evaluating object hallucination in large vision-language models. arXiv:2305.10355, 2023. 6 +[39] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 2, 3, 5 +[40] Ziyi Lin, Shijie Geng, Renrui Zhang, Peng Gao, Gerard De Melo, Xiaogang Wang, Jifeng Dai, Yu Qiao, and Hongsheng Li. Frozen clip models are efficient video learners. In ECCV, 2022. 1, 2 +[41] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In CVPR, 2024. 1, 2, 3, 5, 6 +[42] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, 2024. 1, 2, 3 + +[43] I Loshchilov. Decoupled weight decay regularization. arXiv:1711.05101, 2017. 6 +[44] Soroush Nasiriany, Fei Xia, Wenhao Yu, Ted Xiao, Jacky Liang, Ishita Dasgupta, Annie Xie, Danny Driess, Ayzaan Wahid, Zhuo Xu, et al. Pivot: Iterative visual prompting elicits actionable knowledge for vlms. arXiv:2402.07872, 2024.3 +[45] Bryan A Plummer, Liwei Wang, Chris M Cervantes, Juan C Caicedo, Julia Hockenmaier, and Svetlana Lazebnik. Flickr30k entities: Collecting region-to-phrase correspondences for richer image-to-sentence models. In ICCV, 2015. 3 +[46] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv:2307.01952, 2023. 1 +[47] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 1, 2 +[48] Yongming Rao, Wenliang Zhao, Guangyi Chen, Yansong Tang, Zheng Zhu, Guan Huang, Jie Zhou, and Jiwen Lu. Denseclip: Language-guided dense prediction with context-aware prompting. In CVPR, 2022. 2 +[49] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 1 +[50] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv:2111.02114, 2021. 2, 5 +[51] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. In NeurIPS, 2022. 2, 5 +[52] Sanjay Subramanian, William Merrill, Trevor Darrell, Matt Gardner, Sameer Singh, and Anna Rohrbach. Reclip: A strong zero-shot baseline for referring expression comprehension. arXiv:2204.05991, 2022. 1, 2, 3 +[53] Quan Sun, Yuxin Fang, Ledell Wu, Xinlong Wang, and Yue Cao. Eva-clip: Improved training techniques for clip at scale. arXiv:2303.15389, 2023. 1, 2 +[54] Quan Sun, Yufeng Cui, Xiaosong Zhang, Fan Zhang, Qiying Yu, Yueze Wang, Yongming Rao, Jingjing Liu, Tiejun Huang, and Xinlong Wang. Generative multimodal models are in-context learners. In CVPR, 2024. 2 +[55] Zeyi Sun, Ye Fang, Tong Wu, Pan Zhang, Yuhang Zang, Shu Kong, Yuanjun Xiong, Dahua Lin, and Jiaqi Wang. Alphaclip: A clip model focusing on wherever you want. In CVPR, 2024. 4, 8 +[56] Kaihua Tang, Hanwang Zhang, Baoyuan Wu, Wenhan Luo, and Wei Liu. Learning to compose dynamic tree structures for visual contexts. In CVPR, 2019. 3 + +[57] Shengbang Tong, Ellis Brown, Penghao Wu, Sanghyun Woo, Manoj Middepogu, Sai Charitha Akula, Jihan Yang, Shusheng Yang, Adithya Iyer, Xichen Pan, et al. Cambrian1: A fully open, vision-centric exploration of multimodal llms. arXiv:2406.16860, 2024. 1, 5, 7 +[58] Shengbang Tong, Zhuang Liu, Yuexiang Zhai, Yi Ma, Yann LeCun, and Saining Xie. Eyes wide shut? exploring the visual shortcomings of multimodal llms. In CVPR, 2024. 2, 3, 5 +[59] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv:2307.09288, 2023. 3 +[60] Haoxiang Wang, Pavan Kumar Anasosalu Vasu, Fartash Faghri, Raviteja Vemulapalli, Mehrdad Farajtabar, Sachin Mehta, Mohammad Rastegari, Oncel Tuzel, and Hadi Pouransari. Sam-clip: Merging vision foundation models towards semantic and spatial understanding. In CVPR, 2024. 2 +[61] Peng-Shuai Wang. Octformer: Octree-based transformers for 3d point clouds. TOG, 2023. 1, 4 +[62] Xiaoyang Wu, Li Jiang, Peng-Shuai Wang, Zhijian Liu, Xihui Liu, Yu Qiao, Wanli Ouyang, Tong He, and Hengshuang Zhao. Point transformer v3: Simpler faster stronger. In CVPR, 2024. 1, 4 +[63] xAI. grok, 2024. 6 +[64] Lin Xu, Yilin Zhao, Daquan Zhou, Zhijie Lin, See Kiong Ng, and Jiashi Feng. Pllava: Parameter-free llava extension from images to videos for video dense captioning. arXiv:2404.16994, 2024. 2 +[65] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. In CVPR, 2024. 4 +[66] Lihe Yang, Bingyi Kang, Zilong Huang, Zhen Zhao, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything v2. arXiv:2406.09414, 2024. 4 +[67] Yu-Qi Yang, Yu-Xiao Guo, Jian-Yu Xiong, Yang Liu, Hao Pan, Peng-Shuai Wang, Xin Tong, and Baining Guo. Swin3d: A pretrained transformer backbone for 3d indoor scene understanding. arXiv:2304.06906, 2023. 1, 4 +[68] Yuan Yao, Tianyu Yu, Ao Zhang, Chongyi Wang, Junbo Cui, Hongji Zhu, Tianchi Cai, Haoyu Li, Weilin Zhao, Zhihui He, et al. Minicpm-v: A gpt-4v level mllm on your phone. arXiv:2408.01800, 2024. 2 +[69] Peter Young, Alice Lai, Micah Hodosh, and Julia Hockenmaier. From image descriptions to visual denotations: New similarity metrics for semantic inference over event descriptions. TACL, 2014. 2 +[70] Bangguo Yu, Hamidreza Kasaei, and Ming Cao. L3mvn: Leveraging large language models for visual target navigation. In IROS, 2023. 3 +[71] Jiahui Yu, Zirui Wang, Vijay Vasudevan, Legg Yeung, Mojtaba Seyedhosseini, and Yonghui Wu. Coca: Contrastive captioners are image-text foundation models. arXiv:2205.01917, 2022. 6 + +[72] Weihao Yu, Zhengyuan Yang, Linjie Li, Jianfeng Wang, Kevin Lin, Zicheng Liu, Xinchao Wang, and Lijuan Wang. Mm-vet: Evaluating large multimodal models for integrated capabilities. arXiv:2308.02490, 2023. 6 +[73] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In ICCV, 2023. 1, 3 +[74] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In CVPR, 2023. 4 +[75] Gengze Zhou, Yicong Hong, and Qi Wu. Navgpt: Explicit reasoning in vision-and-language navigation with large language models. In AAAI, 2024. 3 \ No newline at end of file diff --git a/2025/SpatialCLIP_ Learning 3D-aware Image Representations from Spatially Discriminative Language/images.zip b/2025/SpatialCLIP_ Learning 3D-aware Image Representations from Spatially Discriminative Language/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..e41210ba42e098c6f547dd76566743ff34a6ea54 --- /dev/null +++ b/2025/SpatialCLIP_ Learning 3D-aware Image Representations from Spatially Discriminative Language/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:02c611b9528908a7d0010b92c5b510a209ce7e577c11a57af0e2e3442c65e552 +size 540204 diff --git a/2025/SpatialCLIP_ Learning 3D-aware Image Representations from Spatially Discriminative Language/layout.json b/2025/SpatialCLIP_ Learning 3D-aware Image Representations from Spatially Discriminative Language/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..86ad0899d80139069b2cd78da155015e5617b628 --- /dev/null +++ b/2025/SpatialCLIP_ Learning 3D-aware Image Representations from Spatially Discriminative Language/layout.json @@ -0,0 +1,10606 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 129, + 103, + 481, + 139 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 103, + 481, + 139 + ], + "spans": [ + { + "bbox": [ + 129, + 103, + 481, + 139 + ], + "type": "text", + "content": "SpatialCLIP: Learning 3D-aware Image Representations from Spatially Discriminative Language" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 94, + 152, + 518, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 152, + 518, + 180 + ], + "spans": [ + { + "bbox": [ + 94, + 152, + 518, + 180 + ], + "type": "text", + "content": "Zehan Wang" + }, + { + "bbox": [ + 94, + 152, + 518, + 180 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 94, + 152, + 518, + 180 + ], + "type": "text", + "content": ", Sahuai Zhou" + }, + { + "bbox": [ + 94, + 152, + 518, + 180 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 94, + 152, + 518, + 180 + ], + "type": "text", + "content": ", Shaoxuan He" + }, + { + "bbox": [ + 94, + 152, + 518, + 180 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 94, + 152, + 518, + 180 + ], + "type": "text", + "content": ", Haifeng Huang" + }, + { + "bbox": [ + 94, + 152, + 518, + 180 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 94, + 152, + 518, + 180 + ], + "type": "text", + "content": ", Lihe Yang" + }, + { + "bbox": [ + 94, + 152, + 518, + 180 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 94, + 152, + 518, + 180 + ], + "type": "text", + "content": ", Ziang Zhang" + }, + { + "bbox": [ + 94, + 152, + 518, + 180 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 94, + 152, + 518, + 180 + ], + "type": "text", + "content": ", Xize Cheng" + }, + { + "bbox": [ + 94, + 152, + 518, + 180 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 94, + 152, + 518, + 180 + ], + "type": "text", + "content": ", Shengpeng Ji" + }, + { + "bbox": [ + 94, + 152, + 518, + 180 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 94, + 152, + 518, + 180 + ], + "type": "text", + "content": ", Tao Jin" + }, + { + "bbox": [ + 94, + 152, + 518, + 180 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 94, + 152, + 518, + 180 + ], + "type": "text", + "content": ", Hengshuang Zhao" + }, + { + "bbox": [ + 94, + 152, + 518, + 180 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 94, + 152, + 518, + 180 + ], + "type": "text", + "content": ", Zhou Zhao" + }, + { + "bbox": [ + 94, + 152, + 518, + 180 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 140, + 185, + 465, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 185, + 465, + 213 + ], + "spans": [ + { + "bbox": [ + 140, + 185, + 465, + 213 + ], + "type": "text", + "content": "1Zhejiang University; 2The University of Hong Kong https://github.com/SpatialVision/Spatial-CLIP" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 232, + 200, + 245 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 232, + 200, + 245 + ], + "spans": [ + { + "bbox": [ + 151, + 232, + 200, + 245 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 258, + 297, + 619 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 258, + 297, + 619 + ], + "spans": [ + { + "bbox": [ + 55, + 258, + 297, + 619 + ], + "type": "text", + "content": "Contrastive Language-Image Pre-training (CLIP) learns robust visual models through language supervision, making it a crucial visual encoding technique for various applications. However, CLIP struggles with comprehending spatial concepts in images, potentially restricting the spatial intelligence of CLIP-based AI systems. In this work, we propose SpatialCLIP, an enhanced version of CLIP with better spatial understanding capabilities. To capture the intricate 3D spatial relationships in images, we improve both \"visual model\" and \"language supervision\" of CLIP. Specifically, we design 3D-inspired ViT to replace the standard ViT in CLIP. By lifting 2D image tokens into 3D space and incorporating design insights from point cloud networks, our visual model gains greater potential for spatial perception. Meanwhile, captions with accurate and detailed spatial information are very rare. To explore better language supervision for spatial understanding, we re-caption images and perturb their spatial phrases as negative descriptions, which compels the visual model to seek spatial cues to distinguish these hard negative captions. With the enhanced visual model, we introduce SpatialLLaVA, following the same LLaVA-1.5 training protocol, to investigate the importance of visual representations for MLLM's spatial intelligence. Furthermore, we create SpatialBench, a benchmark specifically designed to evaluate CLIP and MLLM in spatial reasoning. SpatialCLIP and SpatialLLaVA achieve substantial performance improvements, demonstrating stronger capabilities in spatial perception and reasoning, while maintaining comparable results on general-purpose benchmarks." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 640, + 135, + 652 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 640, + 135, + 652 + ], + "spans": [ + { + "bbox": [ + 56, + 640, + 135, + 652 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 660, + 295, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 660, + 295, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 660, + 295, + 685 + ], + "type": "text", + "content": "Recently, Contrastive Language-Image Pretraining (CLIP) models [13, 22, 47, 53, 73] have demonstrated the ability to" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 234, + 553, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 234, + 553, + 293 + ], + "spans": [ + { + "bbox": [ + 313, + 234, + 553, + 293 + ], + "type": "text", + "content": "learn robust image representations from language supervision. These models serve as foundational visual encoders for various applications, such as open-world visual recognition [26, 36, 40, 46, 49, 52] and Multi-modal Large Language Models (MLLM) [2, 10, 11, 41, 42, 57]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 293, + 554, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 293, + 554, + 461 + ], + "spans": [ + { + "bbox": [ + 313, + 293, + 554, + 461 + ], + "type": "text", + "content": "Although CLIP exhibits strong zero-shot ability across diverse tasks, it struggles with understanding spatial arrangement within images. As shown in Fig. 1, for a correctly matched image-text pair, we introduce perturbations to the caption by modifying the object category, object attribute, and spatial phrases. The results reveal that altering the object category or attribute significantly reduces the image-text similarity in the CLIP space. However, CLIP struggles to distinguish incorrect spatial relationship descriptions and even assigns a higher matching score to captions with wrong spatial descriptions. Spatial intelligence is a key aspect of visual perception, and limitations in spatial understanding may constrain systems built on CLIP from effectively understanding real-world environments." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 461, + 554, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 461, + 554, + 533 + ], + "spans": [ + { + "bbox": [ + 313, + 461, + 554, + 533 + ], + "type": "text", + "content": "In this paper, we present SpatialCLIP, a CLIP model with enhanced spatial perception capabilities. As titled in [47], CLIP models are \"Learning Transferable Visual Models from Natural Language Supervision\". Our SpatialCLIP advances the spatial understanding from both \"visual models\" and \"language supervision\" perspectives." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 533, + 554, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 554, + 700 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 554, + 700 + ], + "type": "text", + "content": "For \"visual models,\" we propose a 3D-inspired ViT that processes RGB-D input rather than RGB images. By incorporating depth information, our visual model gains the ability to perceive images from the 3D perspective. This design helps the model understand complex 3D spatial relationships like \"behind,\" \"in front of,\" \"under,\" and \"far.\" Specifically, we lift the 2D tokens into 3D space using the extra depth information. In the 3D space, we integrate two structures widely used in point cloud networks: 3D convolution [28, 61] and 3D relative position encoding [34, 62, 67], into the pre-trained ViT model [18]. By integrating design insights from 3D point cloud processing, our 3D-inspired ViT demonstrates enhanced potential for capturing complex 3D spatial structures in images." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 325, + 701, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 701, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 325, + 701, + 553, + 714 + ], + "type": "text", + "content": "For \"language supervision\", in current vision-language" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 693, + 136, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 693, + 136, + 703 + ], + "spans": [ + { + "bbox": [ + 66, + 693, + 136, + 703 + ], + "type": "text", + "content": "*Equal Contribution." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "type": "text", + "content": "† Corresponding author." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "29656" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 62, + 71, + 178, + 144 + ], + "blocks": [ + { + "bbox": [ + 62, + 71, + 178, + 144 + ], + "lines": [ + { + "bbox": [ + 62, + 71, + 178, + 144 + ], + "spans": [ + { + "bbox": [ + 62, + 71, + 178, + 144 + ], + "type": "image", + "image_path": "2cad005fad5393440a393d83a4b69d570032eb6aca59fec42d5a8fff2dd68352.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 63, + 145, + 157, + 168 + ], + "lines": [ + { + "bbox": [ + 63, + 145, + 157, + 168 + ], + "spans": [ + { + "bbox": [ + 63, + 145, + 157, + 168 + ], + "type": "text", + "content": "- Accurate: The darker-haired cat is behind the white cat. \nCLIP Score: 22.28" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 62, + 228, + 168, + 252 + ], + "lines": [ + { + "bbox": [ + 62, + 228, + 168, + 252 + ], + "spans": [ + { + "bbox": [ + 62, + 228, + 168, + 252 + ], + "type": "text", + "content": "Spatial Error: The darker-haired cat is to the right of the white cat. \nCLIP Score: 23.66 (+1.38)" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 55, + 254, + 555, + 287 + ], + "lines": [ + { + "bbox": [ + 55, + 254, + 555, + 287 + ], + "spans": [ + { + "bbox": [ + 55, + 254, + 555, + 287 + ], + "type": "text", + "content": "Figure 1. The preliminary analysis of OpenAI CLIP-L/14@336px. CLIP Score stands for the cosine similarity between image and language description. These examples demonstrate that while CLIP is highly sensitive to errors involving object categories and attributes, it struggles to accurately identify mistakes in spatial descriptions." + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 185, + 71, + 302, + 144 + ], + "blocks": [ + { + "bbox": [ + 61, + 172, + 172, + 197 + ], + "lines": [ + { + "bbox": [ + 61, + 172, + 172, + 197 + ], + "spans": [ + { + "bbox": [ + 61, + 172, + 172, + 197 + ], + "type": "text", + "content": "- Category Error: The darker-haired bird is behind the white cat. \nCLIP Score: 20.43 (-1.85)" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 185, + 71, + 302, + 144 + ], + "lines": [ + { + "bbox": [ + 185, + 71, + 302, + 144 + ], + "spans": [ + { + "bbox": [ + 185, + 71, + 302, + 144 + ], + "type": "image", + "image_path": "45ca3f9a068d23df4997ccc971c5fa7aa8345d42f117e54f3ee047f37a1f4467.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 185, + 145, + 298, + 168 + ], + "lines": [ + { + "bbox": [ + 185, + 145, + 298, + 168 + ], + "spans": [ + { + "bbox": [ + 185, + 145, + 298, + 168 + ], + "type": "text", + "content": "-Accurate: The white refrigerator is closer to the camera than brown door. \nCLIP Score: 29.74" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 181, + 173, + 289, + 197 + ], + "lines": [ + { + "bbox": [ + 181, + 173, + 289, + 197 + ], + "spans": [ + { + "bbox": [ + 181, + 173, + 289, + 197 + ], + "type": "inline_equation", + "content": "\\nrightarrow" + }, + { + "bbox": [ + 181, + 173, + 289, + 197 + ], + "type": "text", + "content": " Category Error: The white cabinet is closer to the camera brown door. CLIP Score: 25.66 (-4.08)" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 181, + 201, + 297, + 224 + ], + "lines": [ + { + "bbox": [ + 181, + 201, + 297, + 224 + ], + "spans": [ + { + "bbox": [ + 181, + 201, + 297, + 224 + ], + "type": "inline_equation", + "content": "\\nrightarrow" + }, + { + "bbox": [ + 181, + 201, + 297, + 224 + ], + "type": "text", + "content": " Attribute Error: The dark refrigerator is closer to the camera than opened door. CLIP Score: 29.39 (-0.35)" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 181, + 228, + 296, + 251 + ], + "lines": [ + { + "bbox": [ + 181, + 228, + 296, + 251 + ], + "spans": [ + { + "bbox": [ + 181, + 228, + 296, + 251 + ], + "type": "text", + "content": "Spatial Error: The white refrigerator is further from the camera than brown door. CLIP Score: 29.62 (-0.12)" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 308, + 71, + 424, + 144 + ], + "blocks": [ + { + "bbox": [ + 59, + 200, + 171, + 224 + ], + "lines": [ + { + "bbox": [ + 59, + 200, + 171, + 224 + ], + "spans": [ + { + "bbox": [ + 59, + 200, + 171, + 224 + ], + "type": "inline_equation", + "content": "\\nLeftrightarrow" + }, + { + "bbox": [ + 59, + 200, + 171, + 224 + ], + "type": "text", + "content": " Attribute Error: The light-haired cat is behind the white cat. CLIP Score: 21.22 (-1.06)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 308, + 71, + 424, + 144 + ], + "lines": [ + { + "bbox": [ + 308, + 71, + 424, + 144 + ], + "spans": [ + { + "bbox": [ + 308, + 71, + 424, + 144 + ], + "type": "image", + "image_path": "61f6d082312b421fe2fd2ab3ae77f0d8c8effd548c2285aeef62c587c42913d1.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 307, + 145, + 418, + 168 + ], + "lines": [ + { + "bbox": [ + 307, + 145, + 418, + 168 + ], + "spans": [ + { + "bbox": [ + 307, + 145, + 418, + 168 + ], + "type": "text", + "content": "-Accurate: The dark coffee machine is to the left of the mug. CLIP Score: 22.10" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 306, + 173, + 415, + 197 + ], + "lines": [ + { + "bbox": [ + 306, + 173, + 415, + 197 + ], + "spans": [ + { + "bbox": [ + 306, + 173, + 415, + 197 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 306, + 173, + 415, + 197 + ], + "type": "text", + "content": " Category Error: The ceramic cup is to the left of the mug. CLIP Score: 19.48 (-2.62)" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 306, + 201, + 417, + 224 + ], + "lines": [ + { + "bbox": [ + 306, + 201, + 417, + 224 + ], + "spans": [ + { + "bbox": [ + 306, + 201, + 417, + 224 + ], + "type": "inline_equation", + "content": "\\nrightarrow" + }, + { + "bbox": [ + 306, + 201, + 417, + 224 + ], + "type": "text", + "content": " Attribute Error: The old broken coffee machine is to the left of the mug. CLIP Score: 21.27 (-0.83)" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 306, + 228, + 418, + 252 + ], + "lines": [ + { + "bbox": [ + 306, + 228, + 418, + 252 + ], + "spans": [ + { + "bbox": [ + 306, + 228, + 418, + 252 + ], + "type": "text", + "content": "Spatial Error: The dark coffee machine is in front of the mug. CLIP Score: 22.53 (+0.43)" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 431, + 72, + 547, + 144 + ], + "blocks": [ + { + "bbox": [ + 431, + 72, + 547, + 144 + ], + "lines": [ + { + "bbox": [ + 431, + 72, + 547, + 144 + ], + "spans": [ + { + "bbox": [ + 431, + 72, + 547, + 144 + ], + "type": "image", + "image_path": "b800fcb15ff8474828847cc57bb4e5e06c1201a1c420d9e29a4b072a0daf2b6b.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 430, + 145, + 528, + 168 + ], + "lines": [ + { + "bbox": [ + 430, + 145, + 528, + 168 + ], + "spans": [ + { + "bbox": [ + 430, + 145, + 528, + 168 + ], + "type": "text", + "content": "-Accurate: Two people are standing behind a pile of huge-sized pipes. CLIP Score: 26.62" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 430, + 173, + 547, + 197 + ], + "lines": [ + { + "bbox": [ + 430, + 173, + 547, + 197 + ], + "spans": [ + { + "bbox": [ + 430, + 173, + 547, + 197 + ], + "type": "inline_equation", + "content": "\\nvdash" + }, + { + "bbox": [ + 430, + 173, + 547, + 197 + ], + "type": "text", + "content": " Category Error: Two people are standing behind a pile of huge-sized papers. CLIP Score: 16.40 (-10.22)" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 430, + 201, + 546, + 224 + ], + "lines": [ + { + "bbox": [ + 430, + 201, + 546, + 224 + ], + "spans": [ + { + "bbox": [ + 430, + 201, + 546, + 224 + ], + "type": "inline_equation", + "content": "\\nvdash" + }, + { + "bbox": [ + 430, + 201, + 546, + 224 + ], + "type": "text", + "content": " Attribute Error: Two people are standing behind a pile of small pipes. \nCLIP Score: 25.78 (-0.84)" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 430, + 228, + 540, + 252 + ], + "lines": [ + { + "bbox": [ + 430, + 228, + 540, + 252 + ], + "spans": [ + { + "bbox": [ + 430, + 228, + 540, + 252 + ], + "type": "text", + "content": "Spatial Error: Two people are standing inside a pile of huge-sized pipes. \nCLIP Score: 27.18 (+0.56)" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 18 + }, + { + "bbox": [ + 54, + 292, + 295, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 292, + 295, + 424 + ], + "spans": [ + { + "bbox": [ + 54, + 292, + 295, + 424 + ], + "type": "text", + "content": "datasets [5, 50, 51] for CLIP training, the input captions are primarily simple summary texts that only describe the primary object category and attributes. The absence of precise spatial information in these captions may constrain the visual model's ability to capture spatial relationships. To enhance language supervision, we recapture images using advanced MLLMs [9, 41] and introduce perturbed spatial phrases to captions as hard negative queries. Training the model to differentiate between captions that share the same object categories and attributes but vary in spatial phrases forces it to better capture spatial cues in images." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 54, + 425, + 295, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 425, + 295, + 555 + ], + "spans": [ + { + "bbox": [ + 54, + 425, + 295, + 555 + ], + "type": "text", + "content": "Furthermore, we explore whether SpatialCLIP, with its enhanced spatial perception, can contribute stronger spatial intelligence to MLLMs when using the same data. Recent work [7, 12, 14] mainly enhances the spatial perception capabilities of MLLMs by introducing new training data, while lacking the exploration of spatial-focused visual representations. However, as discussed in [58], visual representation is also crucial to the capabilities of MLLMs. We construct SpatialLLaVA using the same training recipe as the standard LLaVA-1.5 [41], enabling a fair and systematical analysis over MLLMs." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 55, + 557, + 295, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 557, + 295, + 640 + ], + "spans": [ + { + "bbox": [ + 55, + 557, + 295, + 640 + ], + "type": "text", + "content": "To quantitatively assess the spatial understanding capabilities of CLIP and MLLM, we introduce SpatialBench, a benchmark designed to evaluate spatial comprehension within images. This benchmark challenges models to comprehend spatial arrangements and perform complex spatial reasoning. For thorough evaluation, we curate and annotate images from both outdoor and indoor scenes." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "type": "text", + "content": "Experimentally, SpatialCLIP shows substantial improvements across spatial-focused benchmarks, such as Spatial-Bench and BLINK [24], and performs well on general-purpose downstream tasks [39, 69]. Similarly, SpatialLLaVA demonstrates notably enhanced spatial perception while retaining strong general perceptual abilities." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 324, + 293, + 482, + 303 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 293, + 482, + 303 + ], + "spans": [ + { + "bbox": [ + 324, + 293, + 482, + 303 + ], + "type": "text", + "content": "Our main contributions are as follows:" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 314, + 304, + 554, + 484 + ], + "type": "list", + "angle": 0, + "index": 34, + "blocks": [ + { + "bbox": [ + 314, + 304, + 554, + 340 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 304, + 554, + 340 + ], + "spans": [ + { + "bbox": [ + 314, + 304, + 554, + 340 + ], + "type": "text", + "content": "- We propose SpatialCLIP, which improves the CLIP's spatial understanding capabilities by refining both \"visual model\" and \"language supervision\"." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 314, + 341, + 553, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 341, + 553, + 399 + ], + "spans": [ + { + "bbox": [ + 314, + 341, + 553, + 399 + ], + "type": "text", + "content": "- We develop SpatialLLaVA, by replacing the visual encoding method in LLaVA with our SpatialCLIP while preserving other training protocols. This investigates the value of SpatialCLIP in enhancing the spatial intelligence of CLIP-based MLLM, as shown in Fig. 2." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 314, + 401, + 553, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 401, + 553, + 435 + ], + "spans": [ + { + "bbox": [ + 314, + 401, + 553, + 435 + ], + "type": "text", + "content": "- We construct SpatialBench to evaluate the spatial capabilities of CLIP and MLLMs, collecting and annotating samples from both outdoor and indoor scenarios." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 314, + 436, + 553, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 436, + 553, + 484 + ], + "spans": [ + { + "bbox": [ + 314, + 436, + 553, + 484 + ], + "type": "text", + "content": "- We conduct extensive experiments on both spatial-focused and general-purpose benchmarks to validate the effectiveness of SpatialCLIP, SpatialLLaVA, and each newly proposed component." + } + ] + } + ], + "index": 33 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 495, + 400, + 507 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 495, + 400, + 507 + ], + "spans": [ + { + "bbox": [ + 314, + 495, + 400, + 507 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 313, + 515, + 534, + 529 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 515, + 534, + 529 + ], + "spans": [ + { + "bbox": [ + 313, + 515, + 534, + 529 + ], + "type": "text", + "content": "2.1. Contrastive Language-Image Pre-training" + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 312, + 533, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 533, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 312, + 533, + 555, + 713 + ], + "type": "text", + "content": "CLIP models [13, 22, 47, 53] are trained on web-scale image-text paired data [5, 50, 51] using multimodal contrastive learning as objective. By learning cross-modal semantic alignment from massive datasets, CLIP demonstrates strong zero-shot generalization capabilities and is widely used in various vision-language downstream tasks, such as image object detection [52], segmentation [48, 60], and video action recognition [40]. In recent multimodal large language models (MLLM), CLIP is the most common visual encoding strategy, enabling large language models with the ability to understand images. LLaVA [42] and subsequent works [35, 35, 41, 64] primarily rely on OpenAI-CLIP [47], while other state-of-the-art MLLMs like Emu2 [54], InternVL1.5 [10] and MiniCPM-v [68] also utilize various CLIP variants, such as EVA-CLIP [53], In" + } + ] + } + ], + "index": 37 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "29657" + } + ] + } + ], + "index": 38 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 70, + 181, + 168 + ], + "blocks": [ + { + "bbox": [ + 61, + 70, + 181, + 168 + ], + "lines": [ + { + "bbox": [ + 61, + 70, + 181, + 168 + ], + "spans": [ + { + "bbox": [ + 61, + 70, + 181, + 168 + ], + "type": "image", + "image_path": "addc6756b77131b54ba6cf9c60030b1cdb6f8d02f256910abd1d371015b51030.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 61, + 170, + 167, + 178 + ], + "lines": [ + { + "bbox": [ + 61, + 170, + 167, + 178 + ], + "spans": [ + { + "bbox": [ + 61, + 170, + 167, + 178 + ], + "type": "text", + "content": "Select the best caption for the image:" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 61, + 221, + 181, + 319 + ], + "blocks": [ + { + "bbox": [ + 61, + 185, + 155, + 191 + ], + "lines": [ + { + "bbox": [ + 61, + 185, + 155, + 191 + ], + "spans": [ + { + "bbox": [ + 61, + 185, + 155, + 191 + ], + "type": "text", + "content": "B. Red electric car is closer to the camera." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 66, + 199, + 150, + 207 + ], + "lines": [ + { + "bbox": [ + 66, + 199, + 150, + 207 + ], + "spans": [ + { + "bbox": [ + 66, + 199, + 150, + 207 + ], + "type": "text", + "content": "LLaVA-1.5: The best caption is B." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 65, + 212, + 158, + 219 + ], + "lines": [ + { + "bbox": [ + 65, + 212, + 158, + 219 + ], + "spans": [ + { + "bbox": [ + 65, + 212, + 158, + 219 + ], + "type": "text", + "content": "SpatialLLaVA: The best caption is A." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 61, + 221, + 181, + 319 + ], + "lines": [ + { + "bbox": [ + 61, + 221, + 181, + 319 + ], + "spans": [ + { + "bbox": [ + 61, + 221, + 181, + 319 + ], + "type": "image", + "image_path": "23424863be70cc3457b3dbe5399bca59f30fd6007034bac4614ffab170f308a5.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 61, + 319, + 167, + 327 + ], + "lines": [ + { + "bbox": [ + 61, + 319, + 167, + 327 + ], + "spans": [ + { + "bbox": [ + 61, + 319, + 167, + 327 + ], + "type": "text", + "content": "Select the best caption for the image:" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 61, + 327, + 134, + 334 + ], + "lines": [ + { + "bbox": [ + 61, + 327, + 134, + 334 + ], + "spans": [ + { + "bbox": [ + 61, + 327, + 134, + 334 + ], + "type": "text", + "content": "A. The fan is facing the camera." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 61, + 334, + 141, + 340 + ], + "lines": [ + { + "bbox": [ + 61, + 334, + 141, + 340 + ], + "spans": [ + { + "bbox": [ + 61, + 334, + 141, + 340 + ], + "type": "text", + "content": "B. The fan is facing the white desk." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 61, + 340, + 126, + 346 + ], + "lines": [ + { + "bbox": [ + 61, + 340, + 126, + 346 + ], + "spans": [ + { + "bbox": [ + 61, + 340, + 126, + 346 + ], + "type": "text", + "content": "C. The fan is facing the sofa." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 61, + 346, + 128, + 354 + ], + "lines": [ + { + "bbox": [ + 61, + 346, + 128, + 354 + ], + "spans": [ + { + "bbox": [ + 61, + 346, + 128, + 354 + ], + "type": "text", + "content": "D. The fan is facing the chair." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 66, + 355, + 150, + 363 + ], + "lines": [ + { + "bbox": [ + 66, + 355, + 150, + 363 + ], + "spans": [ + { + "bbox": [ + 66, + 355, + 150, + 363 + ], + "type": "text", + "content": "LLaVA-1.5: The best caption is A." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 65, + 367, + 159, + 375 + ], + "lines": [ + { + "bbox": [ + 65, + 367, + 159, + 375 + ], + "spans": [ + { + "bbox": [ + 65, + 367, + 159, + 375 + ], + "type": "text", + "content": "SpatialLLaVA: The best caption is D." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 380, + 555, + 414 + ], + "lines": [ + { + "bbox": [ + 55, + 380, + 555, + 414 + ], + "spans": [ + { + "bbox": [ + 55, + 380, + 555, + 414 + ], + "type": "text", + "content": "Figure 2. Visualization of LLaVA-1.5 and SpatialLLaVA on SpatialBench. The first row displays examples from outdoor scenes, while the second row features cases in indoor environments. LLaVA-1.5 often encounters difficulties with basic spatial concepts, while SpatialLLaVA improves its spatial perception ability by employing enhanced visual representations." + } + ] + } + ], + "index": 54, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 182, + 70, + 301, + 169 + ], + "blocks": [ + { + "bbox": [ + 61, + 178, + 163, + 185 + ], + "lines": [ + { + "bbox": [ + 61, + 178, + 163, + 185 + ], + "spans": [ + { + "bbox": [ + 61, + 178, + 163, + 185 + ], + "type": "text", + "content": "A. Yellow electric car is closer to the camera." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 182, + 70, + 301, + 169 + ], + "lines": [ + { + "bbox": [ + 182, + 70, + 301, + 169 + ], + "spans": [ + { + "bbox": [ + 182, + 70, + 301, + 169 + ], + "type": "image", + "image_path": "4022cc6b053062907a26b5449eac47bdefc6d1a222aa516f872dcd9e7c40557a.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 182, + 170, + 299, + 183 + ], + "lines": [ + { + "bbox": [ + 182, + 170, + 299, + 183 + ], + "spans": [ + { + "bbox": [ + 182, + 170, + 299, + 183 + ], + "type": "text", + "content": "Select the best caption for the image: A. The bird standing taller is smaller than the other." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 182, + 183, + 296, + 190 + ], + "lines": [ + { + "bbox": [ + 182, + 183, + 296, + 190 + ], + "spans": [ + { + "bbox": [ + 182, + 183, + 296, + 190 + ], + "type": "text", + "content": "B. The bird standing taller is bigger than the other." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 182, + 190, + 279, + 197 + ], + "lines": [ + { + "bbox": [ + 182, + 190, + 279, + 197 + ], + "spans": [ + { + "bbox": [ + 182, + 190, + 279, + 197 + ], + "type": "text", + "content": "C. The two birds are roughly the same size." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 182, + 221, + 301, + 319 + ], + "blocks": [ + { + "bbox": [ + 61, + 191, + 160, + 198 + ], + "lines": [ + { + "bbox": [ + 61, + 191, + 160, + 198 + ], + "spans": [ + { + "bbox": [ + 61, + 191, + 160, + 198 + ], + "type": "text", + "content": "C. Two cars are equally close to the camera." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 186, + 198, + 270, + 207 + ], + "lines": [ + { + "bbox": [ + 186, + 198, + 270, + 207 + ], + "spans": [ + { + "bbox": [ + 186, + 198, + 270, + 207 + ], + "type": "text", + "content": "LLaVA-1.5: The best caption is B." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 186, + 211, + 279, + 219 + ], + "lines": [ + { + "bbox": [ + 186, + 211, + 279, + 219 + ], + "spans": [ + { + "bbox": [ + 186, + 211, + 279, + 219 + ], + "type": "text", + "content": "SpatialLLaVA: The best caption is C." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 182, + 221, + 301, + 319 + ], + "lines": [ + { + "bbox": [ + 182, + 221, + 301, + 319 + ], + "spans": [ + { + "bbox": [ + 182, + 221, + 301, + 319 + ], + "type": "image", + "image_path": "38e9ac6842ef8ffd5194849f124aecda9d165f779b8b423f64918cfb6355b7f6.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 182, + 319, + 288, + 327 + ], + "lines": [ + { + "bbox": [ + 182, + 319, + 288, + 327 + ], + "spans": [ + { + "bbox": [ + 182, + 319, + 288, + 327 + ], + "type": "text", + "content": "Select the best caption for the image:" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 182, + 327, + 247, + 334 + ], + "lines": [ + { + "bbox": [ + 182, + 327, + 247, + 334 + ], + "spans": [ + { + "bbox": [ + 182, + 327, + 247, + 334 + ], + "type": "text", + "content": "A. All the sofas form a circle" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 182, + 334, + 257, + 341 + ], + "lines": [ + { + "bbox": [ + 182, + 334, + 257, + 341 + ], + "spans": [ + { + "bbox": [ + 182, + 334, + 257, + 341 + ], + "type": "text", + "content": "B. All the sofas form a rectangle." + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 182, + 341, + 248, + 347 + ], + "lines": [ + { + "bbox": [ + 182, + 341, + 248, + 347 + ], + "spans": [ + { + "bbox": [ + 182, + 341, + 248, + 347 + ], + "type": "text", + "content": "C. All the sofas form an oval." + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 182, + 348, + 263, + 354 + ], + "lines": [ + { + "bbox": [ + 182, + 348, + 263, + 354 + ], + "spans": [ + { + "bbox": [ + 182, + 348, + 263, + 354 + ], + "type": "text", + "content": "D. All sofas are arranged randomly." + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 186, + 355, + 271, + 363 + ], + "lines": [ + { + "bbox": [ + 186, + 355, + 271, + 363 + ], + "spans": [ + { + "bbox": [ + 186, + 355, + 271, + 363 + ], + "type": "text", + "content": "LLaVA-1.5: The best caption is A." + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 186, + 367, + 279, + 375 + ], + "lines": [ + { + "bbox": [ + 186, + 367, + 279, + 375 + ], + "spans": [ + { + "bbox": [ + 186, + 367, + 279, + 375 + ], + "type": "text", + "content": "SpatialLLaVA: The best caption is B." + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 302, + 70, + 421, + 169 + ], + "blocks": [ + { + "bbox": [ + 302, + 70, + 421, + 169 + ], + "lines": [ + { + "bbox": [ + 302, + 70, + 421, + 169 + ], + "spans": [ + { + "bbox": [ + 302, + 70, + 421, + 169 + ], + "type": "image", + "image_path": "b4589a4fd056cc1756b66a0016a24460867eca3629756fb45916b25e4204ec10.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 302, + 170, + 408, + 178 + ], + "lines": [ + { + "bbox": [ + 302, + 170, + 408, + 178 + ], + "spans": [ + { + "bbox": [ + 302, + 170, + 408, + 178 + ], + "type": "text", + "content": "Select the best caption for the image:" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 302, + 178, + 383, + 185 + ], + "lines": [ + { + "bbox": [ + 302, + 178, + 383, + 185 + ], + "spans": [ + { + "bbox": [ + 302, + 178, + 383, + 185 + ], + "type": "text", + "content": "A. The two trains are perpendicular." + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 302, + 185, + 379, + 190 + ], + "lines": [ + { + "bbox": [ + 302, + 185, + 379, + 190 + ], + "spans": [ + { + "bbox": [ + 302, + 185, + 379, + 190 + ], + "type": "text", + "content": "B. The two trains are intersecting." + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 302, + 190, + 370, + 197 + ], + "lines": [ + { + "bbox": [ + 302, + 190, + 370, + 197 + ], + "spans": [ + { + "bbox": [ + 302, + 190, + 370, + 197 + ], + "type": "text", + "content": "C. The two trains are parallel." + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 423, + 178, + 509, + 184 + ], + "lines": [ + { + "bbox": [ + 423, + 178, + 509, + 184 + ], + "spans": [ + { + "bbox": [ + 423, + 178, + 509, + 184 + ], + "type": "text", + "content": "A. The stop sign is taller than the wall." + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 423, + 184, + 513, + 190 + ], + "lines": [ + { + "bbox": [ + 423, + 184, + 513, + 190 + ], + "spans": [ + { + "bbox": [ + 423, + 184, + 513, + 190 + ], + "type": "text", + "content": "B. The stop sign is shorter than the wall." + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 423, + 190, + 540, + 197 + ], + "lines": [ + { + "bbox": [ + 423, + 190, + 540, + 197 + ], + "spans": [ + { + "bbox": [ + 423, + 190, + 540, + 197 + ], + "type": "text", + "content": "C. The stop sign and the wall are of the same height." + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "image_caption" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 301, + 221, + 421, + 319 + ], + "blocks": [ + { + "bbox": [ + 304, + 199, + 391, + 207 + ], + "lines": [ + { + "bbox": [ + 304, + 199, + 391, + 207 + ], + "spans": [ + { + "bbox": [ + 304, + 199, + 391, + 207 + ], + "type": "text", + "content": "LLaVA-1.5: The best caption is A." + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 212, + 399, + 219 + ], + "lines": [ + { + "bbox": [ + 304, + 212, + 399, + 219 + ], + "spans": [ + { + "bbox": [ + 304, + 212, + 399, + 219 + ], + "type": "text", + "content": "SpatialLLaVA: The best caption is C." + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 301, + 221, + 421, + 319 + ], + "lines": [ + { + "bbox": [ + 301, + 221, + 421, + 319 + ], + "spans": [ + { + "bbox": [ + 301, + 221, + 421, + 319 + ], + "type": "image", + "image_path": "a8b1b29f9decad6790b671733642171920730598399f53d1d9140a9dc62c48ef.jpg" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 302, + 319, + 415, + 327 + ], + "lines": [ + { + "bbox": [ + 302, + 319, + 415, + 327 + ], + "spans": [ + { + "bbox": [ + 302, + 319, + 415, + 327 + ], + "type": "text", + "content": "Select the best caption for the image:" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 302, + 327, + 416, + 335 + ], + "lines": [ + { + "bbox": [ + 302, + 327, + 416, + 335 + ], + "spans": [ + { + "bbox": [ + 302, + 327, + 416, + 335 + ], + "type": "text", + "content": "A. Ceiling light hangs above the edge of the table." + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 302, + 335, + 417, + 341 + ], + "lines": [ + { + "bbox": [ + 302, + 335, + 417, + 341 + ], + "spans": [ + { + "bbox": [ + 302, + 335, + 417, + 341 + ], + "type": "text", + "content": "B. Ceiling light hangs above the center of the table." + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_caption" + } + ], + "index": 36 + }, + { + "type": "image", + "bbox": [ + 423, + 70, + 541, + 168 + ], + "blocks": [ + { + "bbox": [ + 423, + 70, + 541, + 168 + ], + "lines": [ + { + "bbox": [ + 423, + 70, + 541, + 168 + ], + "spans": [ + { + "bbox": [ + 423, + 70, + 541, + 168 + ], + "type": "image", + "image_path": "3e1822acb1b71751ba1e2efd3fc7dbeeabae72b2a705aa9ab09d319e1f10830a.jpg" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 423, + 170, + 529, + 178 + ], + "lines": [ + { + "bbox": [ + 423, + 170, + 529, + 178 + ], + "spans": [ + { + "bbox": [ + 423, + 170, + 529, + 178 + ], + "type": "text", + "content": "Select the best caption for the image:" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_caption" + } + ], + "index": 40 + }, + { + "type": "image", + "bbox": [ + 423, + 221, + 541, + 319 + ], + "blocks": [ + { + "bbox": [ + 426, + 199, + 511, + 207 + ], + "lines": [ + { + "bbox": [ + 426, + 199, + 511, + 207 + ], + "spans": [ + { + "bbox": [ + 426, + 199, + 511, + 207 + ], + "type": "text", + "content": "LLaVA-1.5: The best caption is C." + } + ] + } + ], + "index": 45, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 426, + 212, + 518, + 219 + ], + "lines": [ + { + "bbox": [ + 426, + 212, + 518, + 219 + ], + "spans": [ + { + "bbox": [ + 426, + 212, + 518, + 219 + ], + "type": "text", + "content": "SpatialLLaVA: The best caption is A." + } + ] + } + ], + "index": 46, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 423, + 221, + 541, + 319 + ], + "lines": [ + { + "bbox": [ + 423, + 221, + 541, + 319 + ], + "spans": [ + { + "bbox": [ + 423, + 221, + 541, + 319 + ], + "type": "image", + "image_path": "b458a8c4f2a4f16ef29591dd4b360c22e474b4e7749acad2ebf39a7a3881371e.jpg" + } + ] + } + ], + "index": 47, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 423, + 319, + 528, + 327 + ], + "lines": [ + { + "bbox": [ + 423, + 319, + 528, + 327 + ], + "spans": [ + { + "bbox": [ + 423, + 319, + 528, + 327 + ], + "type": "text", + "content": "Select the best caption for the image:" + } + ] + } + ], + "index": 48, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 423, + 327, + 537, + 335 + ], + "lines": [ + { + "bbox": [ + 423, + 327, + 537, + 335 + ], + "spans": [ + { + "bbox": [ + 423, + 327, + 537, + 335 + ], + "type": "text", + "content": "A. The left window is larger than the right window." + } + ] + } + ], + "index": 49, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 423, + 335, + 537, + 341 + ], + "lines": [ + { + "bbox": [ + 423, + 335, + 537, + 341 + ], + "spans": [ + { + "bbox": [ + 423, + 335, + 537, + 341 + ], + "type": "text", + "content": "B. The right window is larger than the left window." + } + ] + } + ], + "index": 50, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 423, + 341, + 514, + 347 + ], + "lines": [ + { + "bbox": [ + 423, + 341, + 514, + 347 + ], + "spans": [ + { + "bbox": [ + 423, + 341, + 514, + 347 + ], + "type": "text", + "content": "C. The Two mirrors are of the same size." + } + ] + } + ], + "index": 51, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 426, + 354, + 511, + 363 + ], + "lines": [ + { + "bbox": [ + 426, + 354, + 511, + 363 + ], + "spans": [ + { + "bbox": [ + 426, + 354, + 511, + 363 + ], + "type": "text", + "content": "LLaVA-1.5: The best caption is A." + } + ] + } + ], + "index": 52, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 426, + 367, + 518, + 375 + ], + "lines": [ + { + "bbox": [ + 426, + 367, + 518, + 375 + ], + "spans": [ + { + "bbox": [ + 426, + 367, + 518, + 375 + ], + "type": "text", + "content": "SpatialLLaVA: The best caption is C." + } + ] + } + ], + "index": 53, + "angle": 0, + "type": "image_caption" + } + ], + "index": 47 + }, + { + "bbox": [ + 55, + 423, + 295, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 423, + 295, + 483 + ], + "spans": [ + { + "bbox": [ + 55, + 423, + 295, + 483 + ], + "type": "text", + "content": "ternVL [11], and SigLIP [73]. However, ReCLIP [52] recognizes CLIP lacks the capability of understanding spatial relationships in images. Additionally, as discussed in [58], the shortcoming in visual encoding may limit the capabilities of MLLMs built upon it." + } + ] + } + ], + "index": 55 + }, + { + "bbox": [ + 55, + 483, + 295, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 483, + 295, + 507 + ], + "spans": [ + { + "bbox": [ + 55, + 483, + 295, + 507 + ], + "type": "text", + "content": "Our work attempts to enhance CLIP's spatial perception and overcome the shortcomings in spatial understanding." + } + ] + } + ], + "index": 56 + }, + { + "bbox": [ + 55, + 515, + 264, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 515, + 264, + 528 + ], + "spans": [ + { + "bbox": [ + 55, + 515, + 264, + 528 + ], + "type": "text", + "content": "2.2. Spatial Intelligence in Visual Perception" + } + ] + } + ], + "index": 57 + }, + { + "bbox": [ + 55, + 533, + 295, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 533, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 533, + 295, + 677 + ], + "type": "text", + "content": "Spatial understanding [8, 56] and spatial intelligence are fundamental abilities for artificial intelligence systems to perceive and interact with the complex visual world. In our 3D world, spatial understanding involves distinguishing various spatial relationships between objects, such as up and down, left and right, far and near, and inside and outside, as well as perceiving the size, thickness, length, and width of visual elements. The ability to comprehend and reason on these basic spatial concepts is crucial for solving many complex problems, including robot navigation [70, 75], manipulation [25], human-computer interaction [19, 44], and applications in virtual and augmented reality [33]." + } + ] + } + ], + "index": 58 + }, + { + "bbox": [ + 55, + 678, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 678, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 678, + 295, + 714 + ], + "type": "text", + "content": "MLLMs have made impressive progress in recent years. Built upon pre-trained visual representation and large language models [20, 59], these methods demonstrate strong" + } + ] + } + ], + "index": 59 + }, + { + "bbox": [ + 313, + 423, + 555, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 423, + 555, + 495 + ], + "spans": [ + { + "bbox": [ + 313, + 423, + 555, + 495 + ], + "type": "text", + "content": "performance across a variety of downstream tasks, such as image captioning [39], visual question answering [1], visual grounding [30, 45], and action recognition [23]. However, despite their powerful generalization capabilities, most MLLMs still struggle with spatial understanding and reasoning, as discussed in [7]." + } + ] + } + ], + "index": 60 + }, + { + "bbox": [ + 313, + 502, + 555, + 623 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 502, + 555, + 623 + ], + "spans": [ + { + "bbox": [ + 313, + 502, + 555, + 623 + ], + "type": "text", + "content": "There are several works that attempt to address the shortcomings of MLLMs in spatial understanding. SpatialVLM [7] is the first to focus on this issue, introducing massive automatically-generated 3D spatial reasoning VQA data for training MLLM. Its successor, SpatialRGPT [12], enhances the 3D-aware data curating pipeline and provides extra visual depth and region information for MLLM. CubeLLM [14] combines multiple existing 2D and 3D object detection datasets in a unified format and then trains MLLM with better spatial detection capability." + } + ] + } + ], + "index": 61 + }, + { + "bbox": [ + 313, + 630, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 555, + 713 + ], + "type": "text", + "content": "Existing works mainly enhance the spatial intelligence of MLLM by improving the training data, while our work explores the issue from a different perspective: improving the visual representations. Utilizing the same general-purpose training data, simply replacing the standard CLIP with our SpatialCLIP enhances the spatial understanding capabilities of LLaVA [41, 42]." + } + ] + } + ], + "index": 62 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "29658" + } + ] + } + ], + "index": 63 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 80, + 69, + 378, + 273 + ], + "blocks": [ + { + "bbox": [ + 80, + 69, + 378, + 273 + ], + "lines": [ + { + "bbox": [ + 80, + 69, + 378, + 273 + ], + "spans": [ + { + "bbox": [ + 80, + 69, + 378, + 273 + ], + "type": "image", + "image_path": "357f47b7c6ca1367b8c172971e79c111504f680c575f5bdc4e53808a860abd0a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 275, + 555, + 309 + ], + "lines": [ + { + "bbox": [ + 54, + 275, + 555, + 309 + ], + "spans": [ + { + "bbox": [ + 54, + 275, + 555, + 309 + ], + "type": "text", + "content": "Figure 3. The overview of SpatialCLIP. The visual encoder is replaced by our newly proposed 3D-Inspired Vision Transformer, which enables more informative RGB-D inputs and excels in modeling the 3D spatial structure. Language supervision is enhanced through recaption and perturbation, which prompt the encoding of visual spatial information." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 384, + 70, + 528, + 269 + ], + "blocks": [ + { + "bbox": [ + 384, + 70, + 528, + 269 + ], + "lines": [ + { + "bbox": [ + 384, + 70, + 528, + 269 + ], + "spans": [ + { + "bbox": [ + 384, + 70, + 528, + 269 + ], + "type": "image", + "image_path": "4b3aa839cd88e41a94e5b2a1dd2b497ca1a672099bc85394228d8183089850bb.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 316, + 135, + 330 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 316, + 135, + 330 + ], + "spans": [ + { + "bbox": [ + 55, + 316, + 135, + 330 + ], + "type": "text", + "content": "3. SpatialCLIP" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 337, + 296, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 337, + 296, + 386 + ], + "spans": [ + { + "bbox": [ + 55, + 337, + 296, + 386 + ], + "type": "text", + "content": "Our proposed SpatialCLIP enhances standard CLIP from two perspectives: 3D-Inspired ViT for more powerful \"Visual Model\" in Sec. 3.1 and recaption&perturbation strategy for more informative \"Language Supervision\" in Sec. 3.2." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 394, + 225, + 407 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 394, + 225, + 407 + ], + "spans": [ + { + "bbox": [ + 55, + 394, + 225, + 407 + ], + "type": "text", + "content": "3.1. Visual Model: 3D-Inspired ViT" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 413, + 296, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 413, + 296, + 593 + ], + "spans": [ + { + "bbox": [ + 55, + 413, + 296, + 593 + ], + "type": "text", + "content": "In RGB images, the 3D spatial structure of the world is flattened into a 2D plane. Many methods [12, 15, 17, 74] try to introduce depth maps for extra 3D information. However, how to effectively encode such RGB-D inputs while maintaining the pre-trained knowledge for the RGB channel is non-trivial. AlphaCLIP [55] enables an extra input channel with a simple extra input convolution layer, while SpatialRGPT [12] repeats depth map three times as a depth image, and lately fuses the latent of the RGB image and depth image with the pre-trained CLIP image encoder. Our empirical findings show that such simple early and late fusion methods are less effective at integrating depth information. To address this, we propose 3D-inspired ViT for RGB-D inputs to replace the standard vision transformer (ViT) widely used as the visual model in CLIP models." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "text", + "content": "Recently, monocular depth estimation models [31, 65, 66] have gained impressive progress, depth maps can be easily predicted for wild-world RGB images with low cost. Moreover, the accurate depth map shows impressive results in lifting the images into 3D point clouds. Meanwhile, the architectures for processing 3D point cloud [61, 62, 67] focus on capturing spatial relationships among features in 3D space. These observations inspire us to lift 2D tokens into 3D space using predicted depth information and thereby incorporate design insights from 3D point cloud architectures" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 319, + 517, + 330 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 319, + 517, + 330 + ], + "spans": [ + { + "bbox": [ + 313, + 319, + 517, + 330 + ], + "type": "text", + "content": "to better model 3D spatial relationships in images." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 331, + 554, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 331, + 554, + 426 + ], + "spans": [ + { + "bbox": [ + 313, + 331, + 554, + 426 + ], + "type": "text", + "content": "To this end, we first dive into the standard ViT, which processes RGB images as input. Specifically, an RGB image " + }, + { + "bbox": [ + 313, + 331, + 554, + 426 + ], + "type": "inline_equation", + "content": "\\mathbf{X} \\in \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 313, + 331, + 554, + 426 + ], + "type": "text", + "content": " is divided into 2D patches, and each is linearly projected into " + }, + { + "bbox": [ + 313, + 331, + 554, + 426 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 313, + 331, + 554, + 426 + ], + "type": "text", + "content": "-dimension tokens. The resulting 2D tokens map " + }, + { + "bbox": [ + 313, + 331, + 554, + 426 + ], + "type": "inline_equation", + "content": "\\mathbf{Z} \\in \\mathbb{R}^{H_z \\times W_z \\times C}" + }, + { + "bbox": [ + 313, + 331, + 554, + 426 + ], + "type": "text", + "content": " is passed into a stack of transformer blocks, where " + }, + { + "bbox": [ + 313, + 331, + 554, + 426 + ], + "type": "inline_equation", + "content": "H_{z}" + }, + { + "bbox": [ + 313, + 331, + 554, + 426 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 331, + 554, + 426 + ], + "type": "inline_equation", + "content": "W_{z}" + }, + { + "bbox": [ + 313, + 331, + 554, + 426 + ], + "type": "text", + "content": " represent the tokens map's height and width. The 2D token map for the " + }, + { + "bbox": [ + 313, + 331, + 554, + 426 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 313, + 331, + 554, + 426 + ], + "type": "text", + "content": "-th transformer block can be denoted as " + }, + { + "bbox": [ + 313, + 331, + 554, + 426 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}_{L} \\in \\mathbb{R}^{H_{z} \\times W_{z} \\times C_{L}}" + }, + { + "bbox": [ + 313, + 331, + 554, + 426 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 427, + 554, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 427, + 554, + 559 + ], + "spans": [ + { + "bbox": [ + 313, + 427, + 554, + 559 + ], + "type": "text", + "content": "To integrate the well-established 3D point cloud architecture designs into ViT for images, for each transformer layer, we first lift its 2D input tokens " + }, + { + "bbox": [ + 313, + 427, + 554, + 559 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}_L \\in \\mathbb{R}^{H_z \\times W_z \\times C_L}" + }, + { + "bbox": [ + 313, + 427, + 554, + 559 + ], + "type": "text", + "content": " into 3D voxel " + }, + { + "bbox": [ + 313, + 427, + 554, + 559 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}_L^{3\\mathrm{D}} \\in \\mathbb{R}^{H_z \\times W_z \\times D_z \\times C_L}" + }, + { + "bbox": [ + 313, + 427, + 554, + 559 + ], + "type": "text", + "content": " based on the average depth value of each patch, where " + }, + { + "bbox": [ + 313, + 427, + 554, + 559 + ], + "type": "inline_equation", + "content": "H_z" + }, + { + "bbox": [ + 313, + 427, + 554, + 559 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 427, + 554, + 559 + ], + "type": "inline_equation", + "content": "W_z" + }, + { + "bbox": [ + 313, + 427, + 554, + 559 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 427, + 554, + 559 + ], + "type": "inline_equation", + "content": "D_z" + }, + { + "bbox": [ + 313, + 427, + 554, + 559 + ], + "type": "text", + "content": " represent the grid sizes along the height, width, and depth axes. Since " + }, + { + "bbox": [ + 313, + 427, + 554, + 559 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}_L^{3\\mathrm{D}}" + }, + { + "bbox": [ + 313, + 427, + 554, + 559 + ], + "type": "text", + "content": " is derived from " + }, + { + "bbox": [ + 313, + 427, + 554, + 559 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}_L" + }, + { + "bbox": [ + 313, + 427, + 554, + 559 + ], + "type": "text", + "content": ", the resulting 3D voxel grid is highly sparse. The " + }, + { + "bbox": [ + 313, + 427, + 554, + 559 + ], + "type": "inline_equation", + "content": "H_z \\times W_z" + }, + { + "bbox": [ + 313, + 427, + 554, + 559 + ], + "type": "text", + "content": " tokens in " + }, + { + "bbox": [ + 313, + 427, + 554, + 559 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}_L" + }, + { + "bbox": [ + 313, + 427, + 554, + 559 + ], + "type": "text", + "content": " are distributed along the depth axis based on their corresponding depth values, and extracting tokens along the depth axis can easily compress " + }, + { + "bbox": [ + 313, + 427, + 554, + 559 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}_L^{3\\mathrm{D}}" + }, + { + "bbox": [ + 313, + 427, + 554, + 559 + ], + "type": "text", + "content": " back into 2D space." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 559, + 554, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 559, + 554, + 643 + ], + "spans": [ + { + "bbox": [ + 313, + 559, + 554, + 643 + ], + "type": "text", + "content": "For 3D voxelized tokens, 3D depth-wise convolution is widely used to capture spatial patterns across the three dimensions. As shown in the right part of Fig. 3, we integrate this 3D convolution within each transformer block of the pre-trained ViT model, applying it to " + }, + { + "bbox": [ + 313, + 559, + 554, + 643 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}_L^{3\\mathrm{D}}" + }, + { + "bbox": [ + 313, + 559, + 554, + 643 + ], + "type": "text", + "content": " and then short-cutting the resulting features back to the original " + }, + { + "bbox": [ + 313, + 559, + 554, + 643 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}_L" + }, + { + "bbox": [ + 313, + 559, + 554, + 643 + ], + "type": "text", + "content": ". This process is expressed as:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 350, + 653, + 553, + 667 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 653, + 553, + 667 + ], + "spans": [ + { + "bbox": [ + 350, + 653, + 553, + 667 + ], + "type": "interline_equation", + "content": "\\mathbf {Z} _ {L} = \\mathbf {Z} _ {L} + \\operatorname {C o m p r e s s} \\left(3 \\mathrm {D} - \\operatorname {C o n v} \\left(\\mathbf {Z} _ {\\mathrm {L}} ^ {3 \\mathrm {D}}\\right)\\right) \\tag {1}", + "image_path": "304f7981970407a9590053ae2a44eb69b86d4ff54bd60fd2032598bae8223ac2.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "text", + "content": "where 3D-Conv " + }, + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "inline_equation", + "content": "(\\cdot)" + }, + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "text", + "content": " is the 3D depth wise convolution and Compress " + }, + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "inline_equation", + "content": "(\\cdot)" + }, + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "text", + "content": " indicates compressing sparse 3D voxel back into 2D token maps." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "29659" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": "Moreover, another commonly used technique in point cloud networks is 3D relative positional encoding, which explicitly indicates the relative spatial relationship between every pair of tokens in 3D space. For the self-attention operator of each transformer layer, we employ 3D relative positional biases on the attention map. The attention score " + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "inline_equation", + "content": "a_{ij}" + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": " between " + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": "-th and " + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": "-th token can be represented as:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 109, + 163, + 295, + 191 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 109, + 163, + 295, + 191 + ], + "spans": [ + { + "bbox": [ + 109, + 163, + 295, + 191 + ], + "type": "interline_equation", + "content": "a _ {i j} = \\frac {\\left(\\mathbf {z} _ {i} \\mathbf {W} ^ {Q}\\right) \\left(\\mathbf {z} _ {j} \\mathbf {W} ^ {K}\\right) ^ {T} + b _ {i j}}{\\sqrt {C}} \\tag {2}", + "image_path": "d99337cf8f5e0e589a3dbeea8491b5bebf438e8a08b7fc1f6029a47d215f1757.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 198, + 295, + 233 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 198, + 295, + 233 + ], + "spans": [ + { + "bbox": [ + 55, + 198, + 295, + 233 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 198, + 295, + 233 + ], + "type": "inline_equation", + "content": "b_{ij}" + }, + { + "bbox": [ + 55, + 198, + 295, + 233 + ], + "type": "text", + "content": " is a scalar selected from the learnable bias matrix " + }, + { + "bbox": [ + 55, + 198, + 295, + 233 + ], + "type": "inline_equation", + "content": "\\mathbf{B} \\in \\mathbb{R}^{(2H_z - 1) \\times (2W_z - 1) \\times (2D_z - 1)}" + }, + { + "bbox": [ + 55, + 198, + 295, + 233 + ], + "type": "text", + "content": ", based on the relative 3D position between the " + }, + { + "bbox": [ + 55, + 198, + 295, + 233 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 198, + 295, + 233 + ], + "type": "text", + "content": "-th and " + }, + { + "bbox": [ + 55, + 198, + 295, + 233 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 55, + 198, + 295, + 233 + ], + "type": "text", + "content": "-th tokens in 3D space." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 233, + 296, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 233, + 296, + 281 + ], + "spans": [ + { + "bbox": [ + 55, + 233, + 296, + 281 + ], + "type": "text", + "content": "By adopting the designs inspired by the structure of 3D point cloud networks, the 3D-inspired ViT can perceive 2D images from the 3D perspective, greatly enhancing the potential to model complex spatial relationships in images." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 288, + 295, + 312 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 288, + 295, + 312 + ], + "spans": [ + { + "bbox": [ + 55, + 288, + 295, + 312 + ], + "type": "text", + "content": "3.2. Language Supervision: Recapiction & Perturbation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 319, + 295, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 319, + 295, + 462 + ], + "spans": [ + { + "bbox": [ + 55, + 319, + 295, + 462 + ], + "type": "text", + "content": "In the contrastive learning process of CLIP, visual representations of different images are pulled close with the paired caption while pushed away from captions of other images. The guidance information in language captions strongly influences the visual model's capability. For prompting 3D-Inspired ViT to better capture spatial relationships in images, more informative supervision from natural language is necessary. However, current widely-used vision-language datasets for training CLIP, such as Conceptual-12M [5] and LAION [50, 51], are inherently noisy and mainly provide simple summaries of global attributes, lacking detailed caption with accurate spatial description." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 463, + 295, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 463, + 295, + 569 + ], + "spans": [ + { + "bbox": [ + 55, + 463, + 295, + 569 + ], + "type": "text", + "content": "Recent studies [3, 6, 21, 37] have shown that recaptioning images with semantically richer and more aligned descriptions can greatly enhance the performance of vision-language models like CLIP and text-to-image models. However, as discussed in [7, 12], current MLLMs excel at recognizing object categories and attributes but struggle with spatial perception. Consequently, the detailed captions they generated often fail to accurately capture all spatial relationships within the image." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 570, + 295, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 295, + 676 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 295, + 676 + ], + "type": "text", + "content": "To address this \"chicken-and-egg\" dilemma, we propose perturbing the spatial descriptions in the dense captions generated by MLLMs. Specifically, we employ a template to randomly substitute phrases indicating spatial relationships with their antonymic counterparts. For instance, a caption like \"two cats are standing in front of a table, while the left cat is bigger than the right one\" may be modified to \"two cats are standing under a table, while the upper cat is smaller than the lower one.\"" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": "Since the generated and perturbed captions are semantically consistent in terms of object categories and attributes, applying contrastive learning between these two captions" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "content": "encourages the model to distinguish the entirely incorrect caption from the partially correct one. Such a straightforward perturbation method effectively promotes the visual model's ability to capture detailed spatial relationships within images. The learning objective can be expressed as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 347, + 141, + 553, + 209 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 141, + 553, + 209 + ], + "spans": [ + { + "bbox": [ + 347, + 141, + 553, + 209 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} L = - \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\left[ \\log \\frac {e ^ {\\left(\\mathbf {t} _ {i} \\cdot \\mathbf {v} _ {i} / \\tau\\right)}}{\\sum_ {j = 1} ^ {N} e ^ {\\left(\\mathbf {t} _ {i} \\cdot \\mathbf {v} _ {j} / \\tau\\right)}} \\right. \\tag {3} \\\\ \\left. + \\log \\frac {e ^ {\\left(\\mathbf {t} _ {i} \\cdot \\mathbf {v} _ {i} / \\tau\\right)}}{\\sum_ {j = 1} ^ {N} \\left(e ^ {\\left(\\mathbf {t} _ {j} \\cdot \\mathbf {v} _ {i} / \\tau\\right)} + e ^ {\\left(\\tilde {\\mathbf {t}} _ {j} \\cdot \\mathbf {v} _ {i} / \\tau\\right)}\\right)} \\right] \\\\ \\end{array}", + "image_path": "c3548c10424fddccfe339a22f27595f81c4e78454e25d3575a073eaffe38540c.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 217, + 553, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 217, + 553, + 255 + ], + "spans": [ + { + "bbox": [ + 313, + 217, + 553, + 255 + ], + "type": "text", + "content": "where the " + }, + { + "bbox": [ + 313, + 217, + 553, + 255 + ], + "type": "inline_equation", + "content": "v_{i}" + }, + { + "bbox": [ + 313, + 217, + 553, + 255 + ], + "type": "text", + "content": " is the visual representation of " + }, + { + "bbox": [ + 313, + 217, + 553, + 255 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 217, + 553, + 255 + ], + "type": "text", + "content": "-th image, while " + }, + { + "bbox": [ + 313, + 217, + 553, + 255 + ], + "type": "inline_equation", + "content": "t_{i}" + }, + { + "bbox": [ + 313, + 217, + 553, + 255 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 217, + 553, + 255 + ], + "type": "inline_equation", + "content": "\\tilde{t}_i" + }, + { + "bbox": [ + 313, + 217, + 553, + 255 + ], + "type": "text", + "content": " denote the textual representation of the " + }, + { + "bbox": [ + 313, + 217, + 553, + 255 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 217, + 553, + 255 + ], + "type": "text", + "content": "-th caption and its perturbed version, respectively." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 265, + 403, + 279 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 265, + 403, + 279 + ], + "spans": [ + { + "bbox": [ + 314, + 265, + 403, + 279 + ], + "type": "text", + "content": "4. SpatialLLaVA" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 285, + 554, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 285, + 554, + 346 + ], + "spans": [ + { + "bbox": [ + 313, + 285, + 554, + 346 + ], + "type": "text", + "content": "As discussed in [57, 58], the capabilities of MLLMs may be constrained by the inherent limitations of their visual representations. This raises a crucial question: Can the spatial intelligence of CLIP-based MLLM be enhanced by improving the spatial perception of CLIP's visual representations?" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 346, + 554, + 454 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 346, + 554, + 454 + ], + "spans": [ + { + "bbox": [ + 313, + 346, + 554, + 454 + ], + "type": "text", + "content": "We utilize LLaVA-1.5, which employs the pre-trained CLIP as its visual encoding model, as the framework to investigate the effectiveness of SpatialCLIP in MLLM. Specifically, we replace the CLIP model in LLaVA-1.5 [41] with our SpatialCLIP. To ensure fair and systematic comparisons between visual encoding methods, we follow the two-stage training strategy used in LLaVA-1.5 and apply the same training datasets and hyperparameters for each stage. The resulting MLLM is named SpatialLLaVA." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 314, + 465, + 397, + 479 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 465, + 397, + 479 + ], + "spans": [ + { + "bbox": [ + 314, + 465, + 397, + 479 + ], + "type": "text", + "content": "5. SpatialBench" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 485, + 554, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 485, + 554, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 485, + 554, + 605 + ], + "type": "text", + "content": "To evaluate the ability of CLIP and MLLM in understanding complex 2D/3D spatial relations in images, we build a benchmark SpatialBench, which focuses entirely on spatial perception. Our benchmark employs the caption matching task, requiring CLIP or MLLM to distinguish the caption that best matches the image from several distractors. By manually crafting the correct description and distractors, the caption-matching task requires the model to distinguish nuanced spatial concepts in the image (e.g. object arrangement, relationships, relative depth, and size)." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 605, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 554, + 714 + ], + "type": "text", + "content": "To fully assess models in various scenarios, we collect images of both outdoor and indoor scenes. Specifically, outdoor images are collected from the COCO [39] validation set and the Internet, totaling 159 images. Indoor images are sourced from ScanNet [16], consisting of 140 images. Each image is paired with 2 to 4 manually crafted descriptions, of which only one is accurate, with the remaining serving as distractors. Examples of SpatialBench can be found in Fig. 2 and Appendix." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "29660" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 73, + 70, + 535, + 198 + ], + "blocks": [ + { + "bbox": [ + 73, + 70, + 535, + 198 + ], + "lines": [ + { + "bbox": [ + 73, + 70, + 535, + 198 + ], + "spans": [ + { + "bbox": [ + 73, + 70, + 535, + 198 + ], + "type": "table", + "html": "
ModelParamSpatial-FocusedGeneral-Purposed
SpatialBenchBLINKCOCOFlickr30K
OutDoor.InDoor.SpatialR@1R@5R@1R@5
Random-29.7729.8450.00----
OpenAI CLIP-B@224px149.6M35.2230.7151.0542.1967.2534.4056.65
SpatialCLIP-B@224px149.9M+24.8M41.5134.2857.3448.3373.2738.8662.80
OpenAI CLIP-L@336px427.9M32.0730.0047.5546.7870.5438.2560.95
SpatialCLIP-L@336px428.7M+24.8M35.8433.5757.3450.6275.6642.5467.04
", + "image_path": "56693c74fc3fc9dbb5a690c1fba83dc004de00a35273f18881b1e99d188460d7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 59, + 221, + 555, + 342 + ], + "blocks": [ + { + "bbox": [ + 188, + 200, + 421, + 210 + ], + "lines": [ + { + "bbox": [ + 188, + 200, + 421, + 210 + ], + "spans": [ + { + "bbox": [ + 188, + 200, + 421, + 210 + ], + "type": "text", + "content": "Table 1. Results on CLIP benchmark. The best results are bold." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 59, + 221, + 555, + 342 + ], + "lines": [ + { + "bbox": [ + 59, + 221, + 555, + 342 + ], + "spans": [ + { + "bbox": [ + 59, + 221, + 555, + 342 + ], + "type": "table", + "html": "
Training DataModelSpatialBenchBLINKCV-Bench3d
OutDoor.InDoor.SpatialDepthCorresJigsawDepthDistance
Spatial-FocusedSpatial-RGPT-7B37.7338.5765.7382.3032.5550.0062.3359.00
Stronger General-PurposedLLaVA-1.6-7B39.6233.5768.5351.6129.6548.0053.3551.12
LLaVA-1.6-13B38.3635.7169.9948.3830.8150.6755.1652.25
LLaVA-interleave-7B33.2737.1469.9356.6934.8851.3357.1462.30
Baseline General-PurposedLLaVA-1.5-7B29.5532.4761.5452.4225.5811.3345.8350.00
Spatial-LLaVA-7B40.2540.1771.3258.0631.9751.3357.3352.16
", + "image_path": "3ed399047d3c63247f1819bcad8ffbc86308bf156d36d501823e0f64b8141f4d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 58, + 373, + 296, + 471 + ], + "blocks": [ + { + "bbox": [ + 200, + 344, + 408, + 354 + ], + "lines": [ + { + "bbox": [ + 200, + 344, + 408, + 354 + ], + "spans": [ + { + "bbox": [ + 200, + 344, + 408, + 354 + ], + "type": "text", + "content": "Table 2. Results on Spatial-Focused MLLM benchmark." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 373, + 296, + 471 + ], + "lines": [ + { + "bbox": [ + 58, + 373, + 296, + 471 + ], + "spans": [ + { + "bbox": [ + 58, + 373, + 296, + 471 + ], + "type": "table", + "html": "
POPE [38]MM.Vet [72]VOAv2 [27]GOA [29]RealWorldQA [63]
LLaVA-1.5-7B85.930.578.562.054.8
Spatial-LLaVA-7B86.530.779.960.654.2
", + "image_path": "84b9401335b357714485fd33617f992361f95e269dee2815505873e01fc9d1ce.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 66, + 472, + 284, + 483 + ], + "lines": [ + { + "bbox": [ + 66, + 472, + 284, + 483 + ], + "spans": [ + { + "bbox": [ + 66, + 472, + 284, + 483 + ], + "type": "text", + "content": "Table 3. Results on General-Purposed MLLM benchmark." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 494, + 295, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 494, + 295, + 567 + ], + "spans": [ + { + "bbox": [ + 55, + 494, + 295, + 567 + ], + "type": "text", + "content": "For CLIP, we calculate matching scores between the image and its corresponding captions, selecting the description with the highest matching score as the prediction. For MLLMs, the descriptions are presented as multiple-choice options, and the models are prompted to select the best description in a question-answering manner." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 578, + 132, + 592 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 578, + 132, + 592 + ], + "spans": [ + { + "bbox": [ + 55, + 578, + 132, + 592 + ], + "type": "text", + "content": "6. Experiment" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 598, + 188, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 598, + 188, + 611 + ], + "spans": [ + { + "bbox": [ + 55, + 598, + 188, + 611 + ], + "type": "text", + "content": "6.1. Implementation Details" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 54, + 617, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 617, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 617, + 295, + 714 + ], + "type": "text", + "content": "Training setting. We use images in SA1B dataset [32] for training, since most images in it contain multiple objects and showcase complex spatial relationships. Following [6], LLaVA-1.5-13B [41], CoCa [71] and extra ShareGPT4V [9] are used for image reception, the three kinds of captions are randomly used in training for more diverse and informative supervision. A detailed caption perturbation template is provided in the Appendix. The tem" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 376, + 555, + 436 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 376, + 555, + 436 + ], + "spans": [ + { + "bbox": [ + 313, + 376, + 555, + 436 + ], + "type": "text", + "content": "perature factor " + }, + { + "bbox": [ + 313, + 376, + 555, + 436 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 313, + 376, + 555, + 436 + ], + "type": "text", + "content": " in Eq. 3 is set to 0.01. For optimization, we employ AdamW [43] with a learning rate of 2e-7 for pre-trained CLIP encoders and 4e-6 for the newly introduced components. Models are trained for 8,000 steps with a batch size of 2,048 on 8 A100 GPUs." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 449, + 556, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 449, + 556, + 570 + ], + "spans": [ + { + "bbox": [ + 313, + 449, + 556, + 570 + ], + "type": "text", + "content": "Model setting. We train SpatialCLIP from OpenAI CLIP-B/16@224px and OpenAI CLIP-L/14@336px, leading to SpatialCLIP-B and SpatialCLIP-L, respectively. We utilize the lightweight Depth Anything v2 ViT-Small as our monocular depth estimation model, which brings an extra 24.8M parameters. The kernel size of the 3D CNN in Eq. 1 is 3, and the " + }, + { + "bbox": [ + 313, + 449, + 556, + 570 + ], + "type": "inline_equation", + "content": "D_w" + }, + { + "bbox": [ + 313, + 449, + 556, + 570 + ], + "type": "text", + "content": " is 6. The newly added 3D CNN and 3D relative position encoding are both zero-initialized and introduce only nearly 1M extra parameters. We develop SpatialLLaVA based on SpatialCLIP-L, following LLaVA-1.5." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 575, + 441, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 575, + 441, + 588 + ], + "spans": [ + { + "bbox": [ + 313, + 575, + 441, + 588 + ], + "type": "text", + "content": "6.2. Comparison for CLIP" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "content": "Results on spatial-focused benchmark. To assess the spatial perception ability for CLIP models, we employ our newly proposed SpatialBench and modify the \"spatial relation\" subset of BLINK to the caption-matching task, as another spatial-focused benchmark. Results on these benchmarks are presented in Tab. 1. There are two key findings: 1) In both the CLIP-B and -L variants, SpatialCLIP demonstrates significant performance advantages over the original CLIP, indicating that SpatialCLIP more effectively captures spatial information in images and distinguishes between" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "29661" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 72, + 181, + 183 + ], + "blocks": [ + { + "bbox": [ + 59, + 72, + 181, + 183 + ], + "lines": [ + { + "bbox": [ + 59, + 72, + 181, + 183 + ], + "spans": [ + { + "bbox": [ + 59, + 72, + 181, + 183 + ], + "type": "image", + "image_path": "7cb0b2214567b453bb2c8385ca2ca038b5a9a7416420443838b8cd938ee37cb8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 187, + 76, + 545, + 125 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 187, + 76, + 545, + 125 + ], + "spans": [ + { + "bbox": [ + 187, + 76, + 545, + 125 + ], + "type": "text", + "content": "LLaVA: The banana nut bread is on the left side of the picture, with a banana and an egg on the right side. The banana is placed in a blue bowl next to the bread. The eggs are in a glass bowl on the left side of the picture. The banana nut bread is placed in a brown paper bag, with a nut on top." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 187, + 134, + 547, + 182 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 187, + 134, + 547, + 182 + ], + "spans": [ + { + "bbox": [ + 187, + 134, + 547, + 182 + ], + "type": "text", + "content": "SpatialLLaVA: In the image, a bag of banana nut bread mix is placed next to a banana, an egg, and some nuts. The banana is positioned on the right side of the image, while the egg is located on the left side, and the nuts are scattered around the bag of mix. The bag of mix is placed in the middle of the scene, with the banana, egg, and nuts surrounding it." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 198, + 422, + 234 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 198, + 422, + 234 + ], + "spans": [ + { + "bbox": [ + 61, + 198, + 422, + 234 + ], + "type": "text", + "content": "LLaVA: There are two cars that are oriented in a vertical position. The green bus is on the left side of the picture, while the red bus is on the right side. The front of the green bus is facing toward the viewer, while the red bus is facing away from the viewer." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 243, + 422, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 243, + 422, + 281 + ], + "spans": [ + { + "bbox": [ + 61, + 243, + 422, + 281 + ], + "type": "text", + "content": "SpatialLLaVA: In the image, the two buses are positioned side by side, with one bus on the left and the other on the right. They are both facing the same direction, which suggests that they are either parked next to each other or driving in the same direction on the road." + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 430, + 194, + 552, + 283 + ], + "blocks": [ + { + "bbox": [ + 430, + 194, + 552, + 283 + ], + "lines": [ + { + "bbox": [ + 430, + 194, + 552, + 283 + ], + "spans": [ + { + "bbox": [ + 430, + 194, + 552, + 283 + ], + "type": "image", + "image_path": "96df64351e5387fae6b5a8d2baa45d329fdc6ce5053bda13234664619b26d242.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 287, + 555, + 310 + ], + "lines": [ + { + "bbox": [ + 55, + 287, + 555, + 310 + ], + "spans": [ + { + "bbox": [ + 55, + 287, + 555, + 310 + ], + "type": "text", + "content": "Figure 4. Qualitative comparison between LLaVA-1.5 and SpatialLLaVA. We prompt these two MLLMs to provide detailed descriptions of spatial relationships in images. Our model perform better at capturing these spatial concepts." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 321, + 296, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 321, + 296, + 429 + ], + "spans": [ + { + "bbox": [ + 55, + 321, + 296, + 429 + ], + "type": "text", + "content": "different spatial concepts. 2) For both the original CLIP and SpatialCLIP, performance on spatial-focused benchmarks is not correlated with model size. We conjecture this situation is derived from the scarcity of the spatial concepts in the current CLIP pre-training dataset. Increasing the parameter (i.e., enhancing learning capacity) does not necessarily lead to improved spatial perception. This highlights the value of additional designs for better spatial perception, rather than merely scaling up within the existing paradigm." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 444, + 296, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 444, + 296, + 552 + ], + "spans": [ + { + "bbox": [ + 55, + 444, + 296, + 552 + ], + "type": "text", + "content": "Results on general-purposed benchmark. In addition to enhancing spatial concept comprehension, accurate recognition of object categories and attributes is also essential for advanced applications. Since SpatialCLIP is built upon pretrained CLIP and is further trained with high-quality images and informative captions generated by diverse MLLMs, we also observe significant performance improvements on general-purpose benchmarks, like COCO and Flickr30K for zero-shot image-text retrieval." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 562, + 190, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 562, + 190, + 574 + ], + "spans": [ + { + "bbox": [ + 55, + 562, + 190, + 574 + ], + "type": "text", + "content": "6.3. Comparison for MLLM" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 580, + 295, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 580, + 295, + 676 + ], + "spans": [ + { + "bbox": [ + 55, + 580, + 295, + 676 + ], + "type": "text", + "content": "Results on spatial-focused benchmark. In Tab. 2, we comprehensively validate the spatial perception and reasoning ability of our SpatialLLaVA on various spatial-focused MLLM benchmarks. With the same training data, SpatialLLaVA achieves much better results than LLaVA-1.5. SpatialLLaVA exhibits a stronger capability to differentiate between nuanced spatial concepts across indoor and outdoor scenes on our SpatialBench." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "We also conduct absolute performance comparison with more advanced MLLMs, including Spatial-RGPT (trained on spatial-focused data) and LLaVA-1.6/LLaVA-Interleave" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 322, + 555, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 322, + 555, + 429 + ], + "spans": [ + { + "bbox": [ + 313, + 322, + 555, + 429 + ], + "type": "text", + "content": "(trained on much broader general-purpose data). Despite this \"unfair\" comparison, Spatial-LLaVA consistently shows superior spatial perception. Interestingly, Spatial-RGPT excels in tasks it was explicitly trained for (e.g., depth and distance estimation) but underperforms on tasks beyond its training scope (e.g., spatial question-answering). This observation highlights the need to enhance visual perception capabilities in MLLMs beyond simply expanding training data." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 430, + 556, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 430, + 556, + 550 + ], + "spans": [ + { + "bbox": [ + 313, + 430, + 556, + 550 + ], + "type": "text", + "content": "The advantages on BLINK underscore the importance of capturing spatial structure for high-level applications. For \"visual corresponding\" task, capturing spatial structure is beneficial for identifying the same scene points across different viewpoints, lighting, or times, as the scene structure often remains consistent despite these changes. For \"Jigsaw,\" viewing 2D images in the 3D perspective allows the model to inspect the continuity of the 3D spatial structure rather than only relying on color and texture, aiding in organizing image patches." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 552, + 556, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 552, + 556, + 624 + ], + "spans": [ + { + "bbox": [ + 313, + 552, + 556, + 624 + ], + "type": "text", + "content": "CV-Bench[57] collects and annotates samples from the 3D object detection dataset Omni3D [4], focusing on 3D understanding. Employing SpatialCLIP also leads to performance improvements for MLLM on this benchmark, further demonstrating the superiority of SpatialCLIP's visual representations in capturing 3D spatial relationships." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "text", + "content": "Results on general-purposed benchmark. In Tab. 3, we examine how enhancing spatial perception in visual representations affect the performance of MLLMs in general-purposed scenarios. Across five benchmarks, Spatial-LLaVA achieves comparable performance to LLaVA-1.5 for general-purposed questions." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "29662" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 88, + 70, + 520, + 188 + ], + "blocks": [ + { + "bbox": [ + 88, + 70, + 520, + 188 + ], + "lines": [ + { + "bbox": [ + 88, + 70, + 520, + 188 + ], + "spans": [ + { + "bbox": [ + 88, + 70, + 520, + 188 + ], + "type": "table", + "html": "
RowTuning3D CNN3D RPEPerturbSpatialBenchBLINKCOCOFlickr30k
OutDoor.InDoor.Spatial.R@1R@1
1×---35.2231.4351.0542.1934.40
2✓×××35.8531.4351.7548.9939.06
3✓✓××38.3631.4353.8549.0437.76
4✓×✓×37.1032.1454.5547.8139.33
5✓✓✓×38.3633.5755.2447.6737.56
6✓✓✓✓41.5134.2857.3448.3438.86
", + "image_path": "33384e405c54da06f1d40eac2945c0c23c96c7d1b1c7938af0975268be10f10c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 58, + 220, + 295, + 299 + ], + "blocks": [ + { + "bbox": [ + 67, + 190, + 541, + 201 + ], + "lines": [ + { + "bbox": [ + 67, + 190, + 541, + 201 + ], + "spans": [ + { + "bbox": [ + 67, + 190, + 541, + 201 + ], + "type": "text", + "content": "Table 4. Ablation study of our \"Visual Model\" and \"Language Supervision\". All the experiments are conducted on SpatialCLIP-B." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 220, + 295, + 299 + ], + "lines": [ + { + "bbox": [ + 58, + 220, + 295, + 299 + ], + "spans": [ + { + "bbox": [ + 58, + 220, + 295, + 299 + ], + "type": "table", + "html": "
SettingSpatialBenchBLINKCOCOFlickr30k
OutDoorInDoorSpatialR@1R@1
OpenAI CLIP-B35.2231.4351.0542.1934.40
Early Fusion39.6232.8551.7543.0133.09
Late Fusion40.2534.2852.4542.3831.80
3D-Inspried ViT41.5134.2857.3448.3438.86
", + "image_path": "79d1e5349eb5c637e816b30244308893028cf6821106e1683faf46b19f5e29b4.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 60, + 301, + 289, + 312 + ], + "lines": [ + { + "bbox": [ + 60, + 301, + 289, + 312 + ], + "spans": [ + { + "bbox": [ + 60, + 301, + 289, + 312 + ], + "type": "text", + "content": "Table 5. Ablation study of depth channel integration strategies." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 334, + 295, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 334, + 295, + 407 + ], + "spans": [ + { + "bbox": [ + 55, + 334, + 295, + 407 + ], + "type": "text", + "content": "Qualitative comparison. To better demonstrate the improvement of SpatialLLaVA in spatial perception, we provide visualizations of both models in Fig. 2 and 4. The qualitative comparison of their responses reveals that SpatialLLaVA can more accurately and comprehensively capture the spatial relationships between objects in images." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 415, + 149, + 428 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 415, + 149, + 428 + ], + "spans": [ + { + "bbox": [ + 55, + 415, + 149, + 428 + ], + "type": "text", + "content": "6.4. Ablation Study" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 434, + 295, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 434, + 295, + 554 + ], + "spans": [ + { + "bbox": [ + 55, + 434, + 295, + 554 + ], + "type": "text", + "content": "Effect of recaption. In the first two rows of Tab. 4, we present the comparison between the original CLIP and its direct fine-tuning using our recaption dataset. On the general-purposed benchmarks, COCO and Flickr30k, tuning on our dataset leads to significant performance improvements, underscoring the benefits of high information density and semantic alignment in recaption data. However, simply using the recaption data only brings negligible changes to the spatial-focused benchmarks, motivating us to introduce more specific designs to learn spatial concepts." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 570, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 295, + 713 + ], + "type": "text", + "content": "Effect of 3D-inspired ViT. We ablate the designs of 3D-Inspired ViT in Row 2-5 of Tab. 4. When fine-tuning on the same recaption data, incorporating 3D convolution (Row 3) or 3D relative position encoding (Row 4) shows improved performance on spatial-focused benchmarks compared to the standard ViT (Row 2). The combination of both techniques results in even greater enhancements. The improvement in spatial perception gained from the 3D-inspired design underscores the importance of integrating 3D point cloud structures into 2D ViT. Encoding images in 3D space with 3D-specific structure enables the visual model to capture complex 3D spatial information more effectively." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 222, + 555, + 355 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 222, + 555, + 355 + ], + "spans": [ + { + "bbox": [ + 313, + 222, + 555, + 355 + ], + "type": "text", + "content": "Effect of caption perturbation. The comparison between Rows 5 and 6 in Tab. 4 illustrates the benefits of caption perturbation, which prompts the model to capture spatial concepts and enhances its spatial perception ability. Since MLLMs themselves often struggle in perceiving and reasoning about spatial relationships, the captions generated by them can not fully capture all spatial details. By perturbing the spatial phrases, we create references that are fully incorrect in spatial relationships. This approach drives the model to learn spatial concepts by distinguishing the partially correct captions from the fully incorrect ones." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 377, + 556, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 377, + 556, + 534 + ], + "spans": [ + { + "bbox": [ + 313, + 377, + 556, + 534 + ], + "type": "text", + "content": "Different depth channel integration strategies. In addition to our 3D-Inspired ViT, there are other alternative methods to enable RGB-D inputs. In Tab. 5, we discuss the commonly used early and late fusion schemes, which project the depth map into RGB features before and after the visual encoder, which are used by AlphaCLIP [55] and SpatialRGPT [12], respectively. In most benchmarks, simply projecting the depth channel to the pre-trained RGB features is hard to improve the performance. Considering the additional information provided by the reception data, these schemes may even have a negative impact. Conversely, our 3D-inspired ViT makes effective use of the extra depth information, yielding significant improvements." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 548, + 388, + 561 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 548, + 388, + 561 + ], + "spans": [ + { + "bbox": [ + 314, + 548, + 388, + 561 + ], + "type": "text", + "content": "7. Conclusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "text", + "content": "We present SpatialCLIP to overcome the shortcomings of original CLIP models in spatial understanding. By introducing improved \"Visual Model\" (3D-Inspired ViT) and enhanced \"Language Supervision\" (Recap tion & Perturbation), our model achieves significant improvements, not just on spatial-focused benchmarks like our proposed Spatial-Bench and modified BLINK, but also on zero-shot image-text retrieval. Additionally, we develop SpatialLLaVA based on SpatialCLIP, adhering to the LLaVA-1.5 training protocol. This integration of visual representations specifically enhanced for spatial recognition improves the spatial intelligence of MLLM." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "text", + "content": "29663" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 127 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 127 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 127 + ], + "type": "text", + "content": "This work was supported in part by National Natural Science Foundation of China under Grant No.624B2128 and No.62222211." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 140, + 115, + 152 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 140, + 115, + 152 + ], + "spans": [ + { + "bbox": [ + 56, + 140, + 115, + 152 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 159, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 159, + 296, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 159, + 296, + 193 + ], + "spans": [ + { + "bbox": [ + 61, + 159, + 296, + 193 + ], + "type": "text", + "content": "[1] Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C Lawrence Zitnick, and Devi Parikh. Vqa: Visual question answering. In ICCV, 2015. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 194, + 296, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 194, + 296, + 249 + ], + "spans": [ + { + "bbox": [ + 61, + 194, + 296, + 249 + ], + "type": "text", + "content": "[2] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond. arXiv:2308.12966, 2023. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 251, + 296, + 305 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 251, + 296, + 305 + ], + "spans": [ + { + "bbox": [ + 61, + 251, + 296, + 305 + ], + "type": "text", + "content": "[3] James Betker, Gabriel Goh, Li Jing, Tim Brooks, Jianfeng Wang, Linjie Li, Long Ouyang, Juntang Zhuang, Joyce Lee, Yufei Guo, et al. Improving image generation with better captions. https://cdn.openai.com/papers/dall-e-3.pdf, 2023.5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 307, + 295, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 307, + 295, + 350 + ], + "spans": [ + { + "bbox": [ + 62, + 307, + 295, + 350 + ], + "type": "text", + "content": "[4] Garrick Brazil, Abhinav Kumar, Julian Straub, Nikhila Ravi, Justin Johnson, and Georgia Gkioxari. Omni3d: A large benchmark and model for 3d object detection in the wild. In CVPR, 2023. 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 352, + 296, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 352, + 296, + 396 + ], + "spans": [ + { + "bbox": [ + 62, + 352, + 296, + 396 + ], + "type": "text", + "content": "[5] Soravit Changpinyo, Piyush Sharma, Nan Ding, and Radu Soricut. Conceptual 12m: Pushing web-scale image-text pretraining to recognize long-tail visual concepts. In CVPR, 2021. 2, 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 397, + 296, + 452 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 397, + 296, + 452 + ], + "spans": [ + { + "bbox": [ + 62, + 397, + 296, + 452 + ], + "type": "text", + "content": "[6] Agneet Chatterjee, Gabriela Ben Melech Stan, Estelle Aflalo, Sayak Paul, Dhruba Ghosh, Tejas Gokhale, Ludwig Schmidt, Hannaneh Hajishirzi, Vasudev Lal, Chitta Baral, et al. Getting it right: Improving spatial consistency in text-to-image models. In ECCV, 2025. 5, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 454, + 296, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 454, + 296, + 497 + ], + "spans": [ + { + "bbox": [ + 62, + 454, + 296, + 497 + ], + "type": "text", + "content": "[7] Boyuan Chen, Zhuo Xu, Sean Kirmani, Brain Ichter, Dorsa Sadigh, Leonidas Guibas, and Fei Xia. Spatialvlm: Endowing vision-language models with spatial reasoning capabilities. In CVPR, 2024. 2, 3, 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 499, + 296, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 499, + 296, + 532 + ], + "spans": [ + { + "bbox": [ + 62, + 499, + 296, + 532 + ], + "type": "text", + "content": "[8] Dave Zhenyu Chen, Angel X Chang, and Matthias Nießner. Scanrefer: 3d object localization in rgb-d scans using natural language. In ECCV, 2020. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 533, + 296, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 533, + 296, + 577 + ], + "spans": [ + { + "bbox": [ + 62, + 533, + 296, + 577 + ], + "type": "text", + "content": "[9] Lin Chen, Jinsong Li, Xiaoyi Dong, Pan Zhang, Conghui He, Jiaqi Wang, Feng Zhao, and Dahua Lin. Sharegpt4v: Improving large multi-modal models with better captions. arXiv:2311.12793, 2023. 2, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 578, + 295, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 578, + 295, + 633 + ], + "spans": [ + { + "bbox": [ + 57, + 578, + 295, + 633 + ], + "type": "text", + "content": "[10] Zhe Chen, Weiyun Wang, Hao Tian, Shenglong Ye, Zhang-wei Gao, Erfei Cui, Wenwen Tong, Kongzhi Hu, Jiapeng Luo, Zheng Ma, et al. How far are we to gpt-4v? closing the gap to commercial multimodal models with open-source suites. arXiv:2404.16821, 2024. 1, 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 635, + 295, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 635, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 57, + 635, + 295, + 689 + ], + "type": "text", + "content": "[11] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. In CVPR, 2024. 1, 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 691, + 295, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 295, + 714 + ], + "type": "text", + "content": "[12] An-Chieh Cheng, Hongxu Yin, Yang Fu, Qiushan Guo, Ruihan Yang, Jan Kautz, Xiaolong Wang, and Sifei Liu. Spatial-" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 714 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "rgpt: Grounded spatial reasoning in vision language model. arXiv:2406.01584, 2024. 2, 3, 4, 5, 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 96, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 150 + ], + "type": "text", + "content": "[13] Mehdi Cherti, Romain Beaumont, Ross Wightman, Mitchell Wortsman, Gabriel Ilharco, Cade Gordon, Christoph Schuhmann, Ludwig Schmidt, and Jenia Jitsev. Reproducible scaling laws for contrastive language-image learning. In CVPR, 2023. 1, 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 152, + 553, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 553, + 206 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 553, + 206 + ], + "type": "text", + "content": "[14] Jang Hyun Cho, Boris Ivanovic, Yulong Cao, Edward Schmerling, Yue Wang, Xinshuo Weng, Boyi Li, Yurong You, Philipp Krahenbuhl, Yan Wang, et al. Language-image models with 3d understanding. arXiv:2405.03685, 2024. 2, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 208, + 553, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 208, + 553, + 241 + ], + "spans": [ + { + "bbox": [ + 316, + 208, + 553, + 241 + ], + "type": "text", + "content": "[15] Jaeyoung Chung, Jeongtaek Oh, and Kyoung Mu Lee. Depth-regularized optimization for 3d gaussian splatting in few-shot images. In CVPR, 2024. 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 242, + 553, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 242, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 316, + 242, + 553, + 285 + ], + "type": "text", + "content": "[16] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In CVPR, 2017. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 286, + 553, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 286, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 316, + 286, + 553, + 319 + ], + "type": "text", + "content": "[17] Kangle Deng, Andrew Liu, Jun-Yan Zhu, and Deva Ramanan. Depth-supervised nerf: Fewer views and faster training for free. In CVPR, 2022. 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 321, + 553, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 321, + 553, + 354 + ], + "spans": [ + { + "bbox": [ + 316, + 321, + 553, + 354 + ], + "type": "text", + "content": "[18] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv:2010.11929, 2020.1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 355, + 553, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 355, + 553, + 409 + ], + "spans": [ + { + "bbox": [ + 316, + 355, + 553, + 409 + ], + "type": "text", + "content": "[19] Danny Driess, Fei Xia, Mehdi SM Sajjadi, Corey Lynch, Aakanksha Chowdhery, Brian Ichter, Ayzaan Wahid, Jonathan Tompson, Quan Vuong, Tianhe Yu, et al. Palm-e: An embodied multimodal language model. arXiv:2303.03378, 2023. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 411, + 553, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 411, + 553, + 454 + ], + "spans": [ + { + "bbox": [ + 316, + 411, + 553, + 454 + ], + "type": "text", + "content": "[20] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv:2407.21783, 2024. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 456, + 553, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 456, + 553, + 510 + ], + "spans": [ + { + "bbox": [ + 316, + 456, + 553, + 510 + ], + "type": "text", + "content": "[21] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In ICML, 2024. 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 511, + 553, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 511, + 553, + 544 + ], + "spans": [ + { + "bbox": [ + 316, + 511, + 553, + 544 + ], + "type": "text", + "content": "[22] Alex Fang, Albin Madappally Jose, Amit Jain, Ludwig Schmidt, Alexander Toshev, and Vaishaal Shankar. Data filtering networks. arXiv:2309.17425, 2023. 1, 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 545, + 553, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 545, + 553, + 600 + ], + "spans": [ + { + "bbox": [ + 316, + 545, + 553, + 600 + ], + "type": "text", + "content": "[23] Chaoyou Fu, Yuhan Dai, Yondong Luo, Lei Li, Shuhuai Ren, Renrui Zhang, Zihan Wang, Chenyu Zhou, Yunhang Shen, Mengdan Zhang, et al. Video-mme: The first-ever comprehensive evaluation benchmark of multi-modal llms in video analysis. arXiv:2405.21075, 2024. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 601, + 553, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 601, + 553, + 645 + ], + "spans": [ + { + "bbox": [ + 316, + 601, + 553, + 645 + ], + "type": "text", + "content": "[24] Xingyu Fu, Yushi Hu, Bangzheng Li, Yu Feng, Haoyu Wang, Xudong Lin, Dan Roth, Noah A Smith, Wei-Chiu Ma, and Ranjay Krishna. Blink: Multimodal large language models can see but not perceive. arXiv:2404.12390, 2024. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "type": "text", + "content": "[25] Jensen Gao, Bidipta Sarkar, Fei Xia, Ted Xiao, Jiajun Wu, Brian Ichter, Anirudha Majumdar, and Dorsa Sadigh. Physically grounded vision-language models for robotic manipulation. In ICRA, 2024. 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 691, + 553, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 714 + ], + "type": "text", + "content": "[26] Rohit Girdhar, Alaaeldin El-Nouby, Zhuang Liu, Mannat Singh, Kalyan Vasudev Alwala, Armand Joulin, and Ishan" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "29664" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "type": "text", + "content": "Misra. Imagebind: One embedding space to bind them all. In CVPR, 2023. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 95, + 295, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 95, + 295, + 139 + ], + "spans": [ + { + "bbox": [ + 56, + 95, + 295, + 139 + ], + "type": "text", + "content": "[27] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Bartra, and Devi Parikh. Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In CVPR, 2017. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 140, + 294, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 140, + 294, + 172 + ], + "spans": [ + { + "bbox": [ + 56, + 140, + 294, + 172 + ], + "type": "text", + "content": "[28] Benjamin Graham, Martin Engelcke, and Laurens Van Der Maaten. 3d semantic segmentation with submanifold sparse convolutional networks. In CVPR, 2018. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 173, + 294, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 173, + 294, + 205 + ], + "spans": [ + { + "bbox": [ + 56, + 173, + 294, + 205 + ], + "type": "text", + "content": "[29] Drew A Hudson and Christopher D Manning. Gqa: A new dataset for real-world visual reasoning and compositional question answering. In CVPR, 2019. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 206, + 294, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 206, + 294, + 239 + ], + "spans": [ + { + "bbox": [ + 56, + 206, + 294, + 239 + ], + "type": "text", + "content": "[30] Sahar Kazemzadeh, Vicente Ordonez, Mark Matten, and Tamara Berg. Referitag: Referring to objects in photographs of natural scenes. In EMNLP, 2014. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 239, + 294, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 239, + 294, + 282 + ], + "spans": [ + { + "bbox": [ + 56, + 239, + 294, + 282 + ], + "type": "text", + "content": "[31] Bingxin Ke, Anton Obukhov, Shengyu Huang, Nando Metzger, Rodrigo Caye Daudt, and Konrad Schindler. Repurposing diffusion-based image generators for monocular depth estimation. In CVPR, 2024. 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 284, + 294, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 284, + 294, + 327 + ], + "spans": [ + { + "bbox": [ + 56, + 284, + 294, + 327 + ], + "type": "text", + "content": "[32] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In ICCV, 2023. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 327, + 294, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 327, + 294, + 369 + ], + "spans": [ + { + "bbox": [ + 56, + 327, + 294, + 369 + ], + "type": "text", + "content": "[33] Mikhail Konenkov, Artem Lykov, Daria Trinitatova, and Dzmitry Tsetserukou. Vr-gpt: Visual language model for intelligent virtual reality applications. arXiv:2405.11537, 2024.3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 371, + 294, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 371, + 294, + 404 + ], + "spans": [ + { + "bbox": [ + 56, + 371, + 294, + 404 + ], + "type": "text", + "content": "[34] Xin Lai, Jianhui Liu, Li Jiang, Liwei Wang, Hengshuang Zhao, Shu Liu, Xiaojuan Qi, and Jiaya Jia. Stratified transformer for 3d point cloud segmentation. In CVPR, 2022. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 404, + 294, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 404, + 294, + 448 + ], + "spans": [ + { + "bbox": [ + 56, + 404, + 294, + 448 + ], + "type": "text", + "content": "[35] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. Llava-onevision: Easy visual task transfer. arXiv:2408.03326, 2024. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 449, + 294, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 449, + 294, + 492 + ], + "spans": [ + { + "bbox": [ + 56, + 449, + 294, + 492 + ], + "type": "text", + "content": "[36] Liunian Harold Li, Pengchuan Zhang, Haotian Zhang, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, et al. Grounded language-image pre-training. In CVPR, 2022. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 492, + 294, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 492, + 294, + 536 + ], + "spans": [ + { + "bbox": [ + 56, + 492, + 294, + 536 + ], + "type": "text", + "content": "[37] Xianhang Li, Haoqin Tu, Mude Hui, Zeyu Wang, Bingchen Zhao, Junfei Xiao, Sucheng Ren, Jieru Mei, Qing Liu, Huangjie Zheng, et al. What if we recapture billions of web images with llama-3? arXiv:2406.08478, 2024. 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 537, + 294, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 537, + 294, + 569 + ], + "spans": [ + { + "bbox": [ + 56, + 537, + 294, + 569 + ], + "type": "text", + "content": "[38] Yifan Li, Yifan Du, Kun Zhou, Jinpeng Wang, Wayne Xin Zhao, and Ji-Rong Wen. Evaluating object hallucination in large vision-language models. arXiv:2305.10355, 2023. 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 570, + 294, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 570, + 294, + 613 + ], + "spans": [ + { + "bbox": [ + 56, + 570, + 294, + 613 + ], + "type": "text", + "content": "[39] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In ECCV, 2014. 2, 3, 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 614, + 294, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 614, + 294, + 657 + ], + "spans": [ + { + "bbox": [ + 56, + 614, + 294, + 657 + ], + "type": "text", + "content": "[40] Ziyi Lin, Shijie Geng, Renrui Zhang, Peng Gao, Gerard De Melo, Xiaogang Wang, Jifeng Dai, Yu Qiao, and Hongsheng Li. Frozen clip models are efficient video learners. In ECCV, 2022. 1, 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 56, + 658, + 294, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 294, + 690 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 294, + 690 + ], + "type": "text", + "content": "[41] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In CVPR, 2024. 1, 2, 3, 5, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "text", + "content": "[42] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, 2024. 1, 2, 3" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 95 + ], + "type": "text", + "content": "[43] I Loshchilov. Decoupled weight decay regularization. arXiv:1711.05101, 2017. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 96, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 150 + ], + "type": "text", + "content": "[44] Soroush Nasiriany, Fei Xia, Wenhao Yu, Ted Xiao, Jacky Liang, Ishita Dasgupta, Annie Xie, Danny Driess, Ayzaan Wahid, Zhuo Xu, et al. Pivot: Iterative visual prompting elicits actionable knowledge for vlms. arXiv:2402.07872, 2024.3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 152, + 553, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 553, + 206 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 553, + 206 + ], + "type": "text", + "content": "[45] Bryan A Plummer, Liwei Wang, Chris M Cervantes, Juan C Caicedo, Julia Hockenmaier, and Svetlana Lazebnik. Flickr30k entities: Collecting region-to-phrase correspondences for richer image-to-sentence models. In ICCV, 2015. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 208, + 553, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 208, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 316, + 208, + 553, + 262 + ], + "type": "text", + "content": "[46] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv:2307.01952, 2023. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 264, + 553, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 264, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 316, + 264, + 553, + 319 + ], + "type": "text", + "content": "[47] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 1, 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 320, + 553, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 320, + 553, + 365 + ], + "spans": [ + { + "bbox": [ + 316, + 320, + 553, + 365 + ], + "type": "text", + "content": "[48] Yongming Rao, Wenliang Zhao, Guangyi Chen, Yansong Tang, Zheng Zhu, Guan Huang, Jie Zhou, and Jiwen Lu. Denseclip: Language-guided dense prediction with context-aware prompting. In CVPR, 2022. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 365, + 553, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 365, + 553, + 398 + ], + "spans": [ + { + "bbox": [ + 316, + 365, + 553, + 398 + ], + "type": "text", + "content": "[49] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 399, + 553, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 399, + 553, + 453 + ], + "spans": [ + { + "bbox": [ + 316, + 399, + 553, + 453 + ], + "type": "text", + "content": "[50] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv:2111.02114, 2021. 2, 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 455, + 553, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 455, + 553, + 510 + ], + "spans": [ + { + "bbox": [ + 316, + 455, + 553, + 510 + ], + "type": "text", + "content": "[51] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. In NeurIPS, 2022. 2, 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 511, + 553, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 511, + 553, + 555 + ], + "spans": [ + { + "bbox": [ + 316, + 511, + 553, + 555 + ], + "type": "text", + "content": "[52] Sanjay Subramanian, William Merrill, Trevor Darrell, Matt Gardner, Sameer Singh, and Anna Rohrbach. Reclip: A strong zero-shot baseline for referring expression comprehension. arXiv:2204.05991, 2022. 1, 2, 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 555, + 553, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 555, + 553, + 589 + ], + "spans": [ + { + "bbox": [ + 316, + 555, + 553, + 589 + ], + "type": "text", + "content": "[53] Quan Sun, Yuxin Fang, Ledell Wu, Xinlong Wang, and Yue Cao. Eva-clip: Improved training techniques for clip at scale. arXiv:2303.15389, 2023. 1, 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 590, + 553, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 590, + 553, + 633 + ], + "spans": [ + { + "bbox": [ + 316, + 590, + 553, + 633 + ], + "type": "text", + "content": "[54] Quan Sun, Yufeng Cui, Xiaosong Zhang, Fan Zhang, Qiying Yu, Yueze Wang, Yongming Rao, Jingjing Liu, Tiejun Huang, and Xinlong Wang. Generative multimodal models are in-context learners. In CVPR, 2024. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 635, + 553, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 553, + 678 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 553, + 678 + ], + "type": "text", + "content": "[55] Zeyi Sun, Ye Fang, Tong Wu, Pan Zhang, Yuhang Zang, Shu Kong, Yuanjun Xiong, Dahua Lin, and Jiaqi Wang. Alphaclip: A clip model focusing on wherever you want. In CVPR, 2024. 4, 8" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 680, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 680, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 680, + 553, + 713 + ], + "type": "text", + "content": "[56] Kaihua Tang, Hanwang Zhang, Baoyuan Wu, Wenhan Luo, and Wei Liu. Learning to compose dynamic tree structures for visual contexts. In CVPR, 2019. 3" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "29665" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "text", + "content": "[57] Shengbang Tong, Ellis Brown, Penghao Wu, Sanghyun Woo, Manoj Middepogu, Sai Charitha Akula, Jihan Yang, Shusheng Yang, Adithya Iyer, Xichen Pan, et al. Cambrian1: A fully open, vision-centric exploration of multimodal llms. arXiv:2406.16860, 2024. 1, 5, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 130, + 294, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 130, + 294, + 173 + ], + "spans": [ + { + "bbox": [ + 56, + 130, + 294, + 173 + ], + "type": "text", + "content": "[58] Shengbang Tong, Zhuang Liu, Yuexiang Zhai, Yi Ma, Yann LeCun, and Saining Xie. Eyes wide shut? exploring the visual shortcomings of multimodal llms. In CVPR, 2024. 2, 3, 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 175, + 294, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 294, + 229 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 294, + 229 + ], + "type": "text", + "content": "[59] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv:2307.09288, 2023. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 232, + 294, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 232, + 294, + 297 + ], + "spans": [ + { + "bbox": [ + 56, + 232, + 294, + 297 + ], + "type": "text", + "content": "[60] Haoxiang Wang, Pavan Kumar Anasosalu Vasu, Fartash Faghri, Raviteja Vemulapalli, Mehrdad Farajtabar, Sachin Mehta, Mohammad Rastegari, Oncel Tuzel, and Hadi Pouransari. Sam-clip: Merging vision foundation models towards semantic and spatial understanding. In CVPR, 2024. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 300, + 294, + 321 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 300, + 294, + 321 + ], + "spans": [ + { + "bbox": [ + 56, + 300, + 294, + 321 + ], + "type": "text", + "content": "[61] Peng-Shuai Wang. Octformer: Octree-based transformers for 3d point clouds. TOG, 2023. 1, 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 323, + 294, + 367 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 323, + 294, + 367 + ], + "spans": [ + { + "bbox": [ + 56, + 323, + 294, + 367 + ], + "type": "text", + "content": "[62] Xiaoyang Wu, Li Jiang, Peng-Shuai Wang, Zhijian Liu, Xihui Liu, Yu Qiao, Wanli Ouyang, Tong He, and Hengshuang Zhao. Point transformer v3: Simpler faster stronger. In CVPR, 2024. 1, 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 369, + 147, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 369, + 147, + 380 + ], + "spans": [ + { + "bbox": [ + 56, + 369, + 147, + 380 + ], + "type": "text", + "content": "[63] xAI. grok, 2024. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 382, + 294, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 382, + 294, + 425 + ], + "spans": [ + { + "bbox": [ + 56, + 382, + 294, + 425 + ], + "type": "text", + "content": "[64] Lin Xu, Yilin Zhao, Daquan Zhou, Zhijie Lin, See Kiong Ng, and Jiashi Feng. Pllava: Parameter-free llava extension from images to videos for video dense captioning. arXiv:2404.16994, 2024. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 428, + 294, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 428, + 294, + 460 + ], + "spans": [ + { + "bbox": [ + 56, + 428, + 294, + 460 + ], + "type": "text", + "content": "[65] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. In CVPR, 2024. 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 463, + 294, + 495 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 463, + 294, + 495 + ], + "spans": [ + { + "bbox": [ + 56, + 463, + 294, + 495 + ], + "type": "text", + "content": "[66] Lihe Yang, Bingyi Kang, Zilong Huang, Zhen Zhao, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything v2. arXiv:2406.09414, 2024. 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 498, + 294, + 541 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 498, + 294, + 541 + ], + "spans": [ + { + "bbox": [ + 56, + 498, + 294, + 541 + ], + "type": "text", + "content": "[67] Yu-Qi Yang, Yu-Xiao Guo, Jian-Yu Xiong, Yang Liu, Hao Pan, Peng-Shuai Wang, Xin Tong, and Baining Guo. Swin3d: A pretrained transformer backbone for 3d indoor scene understanding. arXiv:2304.06906, 2023. 1, 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 544, + 294, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 544, + 294, + 586 + ], + "spans": [ + { + "bbox": [ + 56, + 544, + 294, + 586 + ], + "type": "text", + "content": "[68] Yuan Yao, Tianyu Yu, Ao Zhang, Chongyi Wang, Junbo Cui, Hongji Zhu, Tianchi Cai, Haoyu Li, Weilin Zhao, Zhihui He, et al. Minicpm-v: A gpt-4v level mllm on your phone. arXiv:2408.01800, 2024. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 589, + 294, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 589, + 294, + 632 + ], + "spans": [ + { + "bbox": [ + 56, + 589, + 294, + 632 + ], + "type": "text", + "content": "[69] Peter Young, Alice Lai, Micah Hodosh, and Julia Hockenmaier. From image descriptions to visual denotations: New similarity metrics for semantic inference over event descriptions. TACL, 2014. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 635, + 294, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 294, + 666 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 294, + 666 + ], + "type": "text", + "content": "[70] Bangguo Yu, Hamidreza Kasaei, and Ming Cao. L3mvn: Leveraging large language models for visual target navigation. In IROS, 2023. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "type": "text", + "content": "[71] Jiahui Yu, Zirui Wang, Vijay Vasudevan, Legg Yeung, Mojtaba Seyedhosseini, and Yonghui Wu. Coca: Contrastive captioners are image-text foundation models. arXiv:2205.01917, 2022. 6" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 72, + 553, + 218 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 315, + 72, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 72, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 315, + 72, + 553, + 116 + ], + "type": "text", + "content": "[72] Weihao Yu, Zhengyuan Yang, Linjie Li, Jianfeng Wang, Kevin Lin, Zicheng Liu, Xinchao Wang, and Lijuan Wang. Mm-vet: Evaluating large multimodal models for integrated capabilities. arXiv:2308.02490, 2023. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 315, + 118, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 118, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 315, + 118, + 553, + 150 + ], + "type": "text", + "content": "[73] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In ICCV, 2023. 1, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 315, + 152, + 553, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 152, + 553, + 184 + ], + "spans": [ + { + "bbox": [ + 315, + 152, + 553, + 184 + ], + "type": "text", + "content": "[74] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In CVPR, 2023. 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 315, + 186, + 553, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 186, + 553, + 218 + ], + "spans": [ + { + "bbox": [ + 315, + 186, + 553, + 218 + ], + "type": "text", + "content": "[75] Gengze Zhou, Yicong Hong, and Qi Wu. Navgpt: Explicit reasoning in vision-and-language navigation with large language models. In AAAI, 2024. 3" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "29666" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/SpatialDreamer_ Self-supervised Stereo Video Synthesis from Monocular Input/24175650-223e-41c7-b804-6cc94c2c5191_content_list.json b/2025/SpatialDreamer_ Self-supervised Stereo Video Synthesis from Monocular Input/24175650-223e-41c7-b804-6cc94c2c5191_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..fd6e445ca513cd0838730213ff555da63f5bce53 --- /dev/null +++ b/2025/SpatialDreamer_ Self-supervised Stereo Video Synthesis from Monocular Input/24175650-223e-41c7-b804-6cc94c2c5191_content_list.json @@ -0,0 +1,1506 @@ +[ + { + "type": "text", + "text": "SpatialDreamer: Self-supervised Stereo Video Synthesis from Monocular Input", + "text_level": 1, + "bbox": [ + 98, + 128, + 897, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zhen Lv $^{1}$ , Yangqi Long $^{1}$ , Congzhentao Huang $^{1}$ , Cao Li $^{1}$ , Chengfei Lv $^{1\\dagger}$ , Hao Ren $^{1}$ , Dian Zheng $^{2}$ , $^{1}$ Alibaba Group; $^{2}$ Sun Yat-sen University", + "bbox": [ + 114, + 191, + 879, + 231 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 276, + 326, + 292 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Stereo video synthesis from monocular input is challenging in spatial computing and virtual reality due to the lack of high-quality stereo video pairs for training and the difficulty of maintaining spatio-temporal consistency between frames. Existing methods primarily address these issues by directly applying novel view synthesis (NVS) techniques to video, while facing limitations such as the inability to effectively represent dynamic scenes and the requirement for extensive training data. In this paper, we introduce a novel self-supervised stereo video synthesis paradigm via a video diffusion model, termed SpatialDreamer, which meets the challenges head-on. Firstly, to address the stereo video data insufficiency, we propose a Depth based Video Generation module DVG, which employs a forward-backward rendering mechanism to generate paired videos with geometric and temporal priors. Leveraging data generated by DVG, we propose RefinerNet along with a self-supervised synthetic framework designed to facilitate efficient and dedicated training. More importantly, we devise a consistency control module, which consists of a metric of stereo deviation strength and a Temporal Interaction Learning module TIL for geometric and temporal consistency assurance respectively. We evaluated the proposed method against various benchmark methods, with the results showcasing its superior performance. Our project website is at: https://spatialdreamer.github.io.", + "bbox": [ + 88, + 308, + 485, + 700 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 729, + 220, + 744 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Stereo video synthesis from a monocular input aims to generate the target-view video based on the given view with geometric and spatio-temporal consistency, which has a wide range of applications in 3D model reconstruction [6], 3D movie production [18, 40], and Apple Vision Pro (AVP) [44] like virtual reality content.", + "bbox": [ + 89, + 755, + 482, + 845 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The primary difficulties of this task stem from the lack of adequate high-quality paired stereo videos for training, and", + "bbox": [ + 89, + 845, + 482, + 876 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/615a11c17cd91356c759af8756fb3f396788be871862426545a3a5a91989c04a.jpg", + "image_caption": [ + "Figure 1. Visualization of the temporal consistency in stereo video generated by different methods. We extract the yellow line of each frame and stack them together. A good result should show a natural transition in the $t$ dimension." + ], + "image_footnote": [], + "bbox": [ + 522, + 282, + 903, + 520 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "the challenge of preserving the geometric consistency between two views as well as maintaining the temporal consistency across generated frames. Traditional methods for generating stereo content involve capturing scenes using dual-camera setups [52, 59, 70, 76]. However, the acquisition of these images involves the use of a professional-grade camera, resulting in substantial costs. Recent advancements in deep learning mainly handle these problems by directly applying single to stereo or multi-view image generation techniques to video, while often facing challenges related to geometric and spatio-temporal inconsistencies [21, 32, 41, 50, 66, 67]. As shown in Fig 1, methods based on NVs are challenging to maintain accurate spatio-temporal consistency, and even the recently released AVP 3D photo converter [2] may inevitably result in content flickering and inconsistency when applied to video synthesis. This is due to the difficulty of handling the temporal coherence between frames and the spatial information between paired views in complex dynamic scenes.", + "bbox": [ + 511, + 613, + 906, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author.", + "bbox": [ + 114, + 886, + 243, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "811", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we propose a self-supervised stereo video synthesis paradigm via a video diffusion model, termed SpatialDreamer, which will meet the data insufficiency and spatio-temporal inconsistency at once. Firstly, we design a Depth based Video data Generation module DVG to handle the problem of data insufficiency. Similar to [21, 66], DVG constructs a training pair of two views using forward-backward rendering without the need for data annotation [51, 59, 70], while [21, 66] only render individual images, resulting in flickering effects in generated video [9]. In this paper, we improve the generation of video data for refining stereo occlusion masks by leveraging the inter-frame motion obtained from optical flow [29]. This methodology enables DVG to produce paired video data that maintains both geometric and temporal consistency. Leveraging paired video generated by DVG, we propose Refiner-Net, along with a self-supervised video synthesis framework, to enable efficient and targeted training. More importantly, with access to sufficient paired videos, we devise a consistency control module, which consists of a metric of stereo deviation strength and a Temporal Interaction Learning module TIL. The stereo deviation strength aims to enable the generation of stereo videos in diverse real-world scenarios, with a stereo-aware loss further supervising the model to learn the magnitude of difference between paired view features in the latent space. TIL integrates the latent features from long-temporal frames as a global information to augment the temporal coherence of the generated results.", + "bbox": [ + 93, + 90, + 480, + 512 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Extensive experiments compared with various benchmark methods show that we achieve state-of-the-art performance. Notably, our SpatialDreamer meets the demand of real-world application without jitter, geometric and temporal inconsistency, which even beats AVP 3D converter and any open-source stereo video synthesis methods. We summarize our contributions as follows:", + "bbox": [ + 93, + 515, + 480, + 619 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- A novel self-supervised stereo video synthesis framework, SpatialDreamer, is proposed, which is robust across a wide range of scenes and dynamic content in video. Moreover, equipping the DVG module, we build a stereo video dataset using a self-supervised approach.", + "- A consistency control module is devised, which consists of a metric of stereo deviation strength and a temporal interaction learning module, ensuring geometric and temporal consistency in video sequences.", + "- The sufficient experiments and quantitative results demonstrate that the proposed method outperforms the age-of-the-art methods, even beats AVP." + ], + "bbox": [ + 89, + 622, + 478, + 797 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 94, + 820, + 230, + 834 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Single-image Novel View Synthesis", + "text_level": 1, + "bbox": [ + 94, + 845, + 390, + 859 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Early studies address the single-image novel view synthesis task without relying on explicit 3D representation,", + "bbox": [ + 94, + 871, + 478, + 898 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "such as End-to-end Synthesis [69], Infinite-Nature [37] and Animatable 3D Characters [68]. These methods lack explicit 3D representations for scenes, leading to inconsistent views and causing flickering and blurring, especially in occluded regions. Layer-based methods, such as 3D-photography [56] and SLIDE [27] represent a 3D scene using discrete layers. Yet, the synthesized views may lack accuracy in texture and geometric consistency across different perspectives. SMPI [61], AdaMPI [21] and SinMPI [47] represent the 3D scene using multiple planes, while their rendered results may suffer from depth discretization artifacts and repetitive texture patterns. To reduce the reliance on large labeled datasets, Shi et.al [55], Wang et.al. [66] and NVSVDE-Net [4] explore self-supervised learning for creating novel view images from single image. Although they produce photorealistic results, they struggle with temporal and spatial consistency, especially in dynamic scenes.", + "bbox": [ + 516, + 90, + 903, + 347 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "As for our approach, we design a depth-based video generation module that generates paired videos with geometric and temporal priors. Additionally, Our method ensures multi-view consistency for the entire scene by employing a consistency control module.", + "bbox": [ + 516, + 357, + 903, + 430 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Multi-view Novel View Synthesis", + "text_level": 1, + "bbox": [ + 516, + 468, + 799, + 482 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Early approaches in this field frequently utilized geometric methods, such as multi-view stereo reconstruction [30, 54] and structure-from-motion techniques [46, 53], to generate novel viewpoints from multiple images captured at different angles. NeRF-based methods were proposed to generate novel views based on fewer input images, including SinNeRF [72], CoPoNeRF [24] and NViST [28]. 3D Gaussian Splatting [32] presented an explicit representation of the scene. Gaussian-based methods, such as pixelSplat [12] and MVSplat [14], can produce high-quality synthesis results from multi-view images. Recently, researchers have begun exploring the application of advanced diffusion models for novel view synthesis, driven by their remarkable success in this area. Notable approaches such as Photoconsistent-NVS [75], MultiDiff [42], GCD [63] and NVS-Solver [74] generate image sequences based on predefined camera trajectories using either single or multiple input images. While these methods have shown promising results, they often struggle with dependence on dense scene geometry, handling complex scene dynamics, and generalization challenges [32]. Moreover, the challenge of optimizing NeRF/GS with sparse views hampers the ability of multi-view NVS to produce high-quality results.", + "bbox": [ + 516, + 500, + 903, + 845 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In contrast, we explore a video diffusion model to achieve stereo video synthesis, while preserving geometric structures through a RefinerNet based framework.", + "bbox": [ + 516, + 856, + 903, + 898 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "812", + "bbox": [ + 486, + 946, + 509, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/4dcc85377b5a73a634d9a166cd85995df87c19ba07c84e4620eae5e7e486c4f6.jpg", + "image_caption": [ + "Figure 2. Overview of the proposed method, given left view video, the target view video is rendered, encoded, and concatenated with multi-frame noise, followed by the denoising U-Net architecture (i.e., SVD). The reference view images are fed into RefinerNet, through which the spatial features are extracted. (Notably, the left and right views are the target and reference view for both training and inference). The temporal interaction learning module integrates the latent features from long-temporal frames, and the deviation strength is projected as positional embedding and added to the time step embedding. Finally, the variational autoencoder decoder decodes the result into a video clip." + ], + "image_footnote": [], + "bbox": [ + 117, + 95, + 880, + 348 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Video Synthesis via Diffusion Models", + "text_level": 1, + "bbox": [ + 89, + 454, + 410, + 469 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For video synthesis, previous works mainly expand image models to video, exploiting the image generation capability in the pretrained image diffusion models [38]. The video diffusion model (VDM) [23] and Video LDM [8] are proposed to adopt a 3D U-Net structure for joint image and video training. Subsequent works [15, 19, 19, 25, 57] also follow this insight with temporal attention, and [33, 48, 65, 71] further apply cross-frame attention to improve the consistency of the generated video. Another approach to promote temporal consistency is encoding video into 2D images [3, 45], which has been mainly explored in the video editing area. Although these models produce consistent results, they often require per-video optimization, which takes a long time, and would result in degradation when dealing with large motion. Most recently, [10, 20, 34] employ a transformer [64] architecture to model the temporal and spatial relations, demonstrating excellent video generation capabilities. However, these models need very large scale video datasets for training, which can be difficult to collect.", + "bbox": [ + 91, + 477, + 482, + 763 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Beyond temporal information modeling, some works [13, 17, 67, 77] have introduced aligned spatial guidance to enhance the stability of the generated videos. The proposed framework also uses spatial guidance. Differing from those methods which require spatially aligned controls, a spatially unaligned reference image is used via a cross-attention mechanism. Furthermore, a metric of stereo deviation strength is introduced to strengthen the spatial constraint, resulting in more consistent results.", + "bbox": [ + 89, + 765, + 482, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. SpatialDreamer", + "text_level": 1, + "bbox": [ + 513, + 453, + 671, + 470 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given an image sequence and stereoscopic pose describing the scene depth information, the proposed method generates a stereo pair. In this section, we first provide a brief introduction about SVD [7]. Secondly, we offer detailed information of our video generation module DVG, which can generate sufficient paired video with geometric and temporal consistency. Using DVG, we present the framework for self-supervised learning-based video synthesis to support efficient and dedicated training. More importantly, to ensure consistency between frames, we design a consistency control module for geometric and temporal consistency assurance. Finally, we present the detailed training process.", + "bbox": [ + 511, + 478, + 906, + 660 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminaries", + "text_level": 1, + "bbox": [ + 513, + 667, + 653, + 681 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The challenge in video generation lies in accurately modeling the spatial-temporal dynamics, involving spatial relationships within frames and temporal relationships across frames. To address this challenge, Stable Video Diffusion (SVD) [7] introduces temporal convolution and attention layers. SVD follows the latent diffusion model (LDM) to encode video pixels into the latent space, enabling a more efficient denoising process. Given a video $I_{v}$ , encoder $\\mathcal{E}$ encodes $I_{v}$ into the latent space as $z_{0} = \\mathcal{E}(I_{v})$ . Gaussian noise $\\epsilon$ is then added in the Markov process:", + "bbox": [ + 511, + 690, + 906, + 840 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nz _ {t} = \\sqrt {\\bar {\\alpha} _ {t}} z _ {0} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon_ {t} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 620, + 847, + 903, + 864 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\bar{\\alpha}_0,\\dots ,\\bar{\\alpha}_T$ are the pre-defined noise schedules for T steps in the Markov process. In particular, SVD is trained", + "bbox": [ + 511, + 869, + 906, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "813", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "by the $v$ -prediction formulation with the mean squared error (MSE) between the ground truth and its prediction in the latent space, known as $l_{\\epsilon}$ .", + "bbox": [ + 89, + 90, + 485, + 137 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Depth based Video Generation", + "text_level": 1, + "bbox": [ + 89, + 147, + 362, + 165 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To create training pairs of images along with their corresponding occluded masks, the following self-supervised mechanism is employed utilizing real images. Firstly, the depth of the given video sample is estimated [49]. Secondly, the reference view image $x_{1}$ is rendered into a masked one under the target viewpoint $P_{2}$ , and then the trained inpainting model [58] is used to fill these occluded regions to obtain the novel viewpoint $x_{2}$ [43]. Finally, $x_{2}$ is rendered to the original viewpoint $P_{1}$ via a forward rendering. After this, we can obtain the masked image $\\widetilde{x}_{1}$ with mask $M$ under viewpoint $P_{1}$ , as well as the inpainted image $x_{2}$ under viewpoint $P_{2}$ . The mask $M$ and masked image $\\widetilde{x}_{1}$ are considered to be the conditions, with the image $x_{1}$ serving as the ground truth for the diffusion model. This approach involves encoding the inpainted image $x_{2}$ into latent feature representations and feeding it into RefinerNet.", + "bbox": [ + 88, + 171, + 482, + 412 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "However, the aforementioned approach encounters challenges in addressing inconsistencies among generated data frames, resulting in observable artifacts such as jittering and flickering. To tackle this issue, we leverage optical flow [60] to capture motion between successive video frames. Specifically, by leveraging the optical flow between the reference image $x_{1}^{t}$ at time $t$ and its adjacent frames, we establish the pixel correspondences between the image $x_{1}^{t}$ and the images $x_{1}^{t-1}$ and $x_{1}^{t+1}$ , respectively. Particularly, these optical flows enable computing a confidence map via forward-backward consistency [16], which allows us to more accurately merge the stereo occlusions in the masks between adjacent frames [1]. Subsequently, the refined occlusion mask $m^{t}$ is generated by sampling from the occlusion maps $m^{t-1}$ and $m^{t+1}$ , leveraging the pixel correspondences between consecutive frames under the reference view, as described by the following equation.", + "bbox": [ + 89, + 412, + 482, + 672 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nm ^ {t} (i, j) = \\left\\{ \\begin{array}{l l} 1, & \\sum_ {k \\in \\{t - 1, t + 1 \\}} m ^ {k} (i + u, j + v) \\cdot \\\\ & C (i, j) \\geq 1 \\\\ m ^ {t} (i, j), & e l s e \\end{array} \\right. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 102, + 694, + 480, + 779 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $u$ and $v$ are the $x$ and $y$ components of the optical flow vector, and $C$ represents the optical flow confidence measure. By propagating information from occluded areas across frames, this inter-frame continuity facilitates a more precise refinement of occluded regions in paired view videos and enforces temporal smoothness and coherence within the self-supervised framework, effectively mitigating the undesirable effects of jitter and flicker.", + "bbox": [ + 89, + 780, + 482, + 902 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/804bf090dcca49d6bf77fd03da14a6625ff8ef4378bf10a90929487f83332cff.jpg", + "image_caption": [ + "Figure 3. DVG. The temporal motion can be utilized to refine the occluded regions in the current frame, thereby providing smoother images and more consistent occlusion in the temporal sequence." + ], + "image_footnote": [], + "bbox": [ + 531, + 88, + 890, + 250 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As shown in Figure 3, the difference between the training and inference procedures is that the former uses both backward and forward rendering whereas the latter uses only forward rendering to obtain an image under the target view.", + "bbox": [ + 511, + 339, + 906, + 401 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Self-Supervised Stereo Video Synthesis framework", + "text_level": 1, + "bbox": [ + 511, + 426, + 906, + 458 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Leveraging the generated paired view videos, we propose a self-supervised framework for stereo video synthesis, as illustrated in Figure 2. This framework commences with the rendering of the target view video that incorporates occlusions, followed by the application of a spatio-temporal method to effectively inpaint the missing regions, ultimately yielding a complete non-occluded target view video.", + "bbox": [ + 511, + 470, + 905, + 578 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To learn the distribution of the feature space between the paired view images, we devise RefinerNet, which is modeled after the architecture of the denoising U-Net, excluding the temporal layer. RefinerNet adopts the weight initialization from the original SD2.1 model [50]. As depicted in Figure 2, a spatial attention layer is employed instead of the self-attention layer. Given a feature map $z_{t} \\in \\mathbb{R}^{t \\times h \\times w \\times c}$ from the denoising U-Net and another feature map $z_{r} \\in \\mathbb{R}^{t \\times h \\times w \\times c}$ from RefinerNet, they are concatenated along the h dimension, and then self-attention is applied. Subsequently, the first half of the feature map is retrieved as the output. One great advantage of this approach is that it enables RefinerNet to maintain the feature representation capability inherent in the original SD Unet, extracting reference view image information. In addition, the denoising U-Net can adaptively learn features from RefinerNet that are correlated in the same feature space, owing to the fundamentally similar network architectures between the denoising U-Net and RefinerNet. This not only ensures a robust initialization for the feature representation but also enhances the learning ability of the entire network.", + "bbox": [ + 511, + 583, + 906, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "814", + "bbox": [ + 485, + 945, + 511, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Consistency Control Module", + "text_level": 1, + "bbox": [ + 89, + 90, + 346, + 107 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4.1 Temporal Interaction Learning module", + "text_level": 1, + "bbox": [ + 89, + 113, + 418, + 128 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "While our RefinerNet-based approach providing spatial guidance, paired view features are still merged independently, posing challenges for temporal consistency and potentially causing texture discrepancies or odd results between adjacent frames.", + "bbox": [ + 89, + 136, + 482, + 210 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Inspired by previous studies [11, 22], we propose an attention-based temporal interaction learning module that learns the distribution of the feature space from adjacent frames, named TIL. Specifically, given latent features of $N_r$ adjacent images under reference view $z_i^t$ , where $i = 1, 2, \\dots, N_r$ and the latent feature of the reference view image $z_r^t$ at the time step $t$ , we augment the $z_r^t$ to $aug_r^t$ as:", + "bbox": [ + 89, + 212, + 483, + 319 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\na u g _ {r} ^ {t} = \\lambda \\cdot \\operatorname {A t t n} _ {r, r} + (1 - \\lambda) \\cdot \\frac {1}{N _ {r}} \\sum_ {i = 1} ^ {N _ {r}} \\operatorname {A t t n} _ {r, i} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 127, + 330, + 482, + 371 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda \\in [0,1]$ and $\\mathrm{Attn}_{i,j}$ defines attention between two latent features $z_{i}$ and $z_{j}$ . This module blends self-attention of $z_{r}^{t}$ in RefinerNet's U-Net block with the cross-view attention between $z_{r}^{t}$ and each adjacent view $z_{i}^{t}$ . The cross-view attention aligns the feature of all adjacent frames to the reference view while the self-attention helps each reference image retain distinctiveness. The augmented reference feature $aug_{r}^{t}$ is then fed into spatial attention layer to assist U-Net network learning.", + "bbox": [ + 89, + 381, + 483, + 518 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4.2 Stereo Deviation Strength", + "text_level": 1, + "bbox": [ + 89, + 536, + 323, + 551 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "During the training, we noticed that different viewpoints impact the 3D impression of the scenes in the generated videos. However, relying solely on a given viewpoint as a guide for generating stereoscopic videos is insufficient. This is because videos with the same viewpoint may exhibit diverse depth ranges and variations, depending on the scene [70]. Therefore, pose alone cannot accurately control the 3D effect of a scene. Consequently, we introduce a metric called the stereo deviation strength, which quantitatively assesses the binocular disparity in a scene and facilitates the creation of controllable stereo vision:", + "bbox": [ + 89, + 559, + 483, + 724 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\ns (z) = \\left| z _ {0} - z _ {\\text {r e f}} \\right| \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 222, + 737, + 482, + 753 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where the stereo deviation strength quantifies the latent differences between reference view $z_{ref}$ and target view $z_0$ . Similar to the time step, the deviation strength is projected into a positional embedding and added to each frame in the residual block to ensure uniform application of the deviation strength to every frame. The impact of stereo deviation strength guidance on the results is illustrated in Figure 4.", + "bbox": [ + 89, + 763, + 482, + 869 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Moreover, for better convergence, a stereo-aware loss function is proposed to directly supervise the disparity dif", + "bbox": [ + 89, + 871, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/fc476e0ee20f138f5507c314034e7853356f1d077a6ae97128a8663bfe3f6e0f.jpg", + "image_caption": [ + "Figure 4. Stereo deviation strength guidance examples. Augmenting the deviation strength enhances the 3D photography effect, but excessive deviation strength may lead to image distortions or even lower the quality of stereoscopic image." + ], + "image_footnote": [], + "bbox": [ + 535, + 88, + 885, + 247 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "ference:", + "bbox": [ + 513, + 340, + 571, + 354 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nl _ {d} = \\left| \\left| s \\left(z _ {0}\\right) - s \\left(\\hat {z} _ {0}\\right) \\right| \\right| _ {2} ^ {2} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 630, + 354, + 903, + 372 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\hat{z}_0$ represents the estimated clean video latent $z_0$ , which can be obtained by:", + "bbox": [ + 511, + 378, + 906, + 407 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {z} _ {0} = z _ {0} - \\frac {\\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon_ {\\theta} (z _ {t} , c , t)}{\\sqrt {\\bar {\\alpha} _ {t}}} \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 606, + 419, + 903, + 454 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\overline{\\alpha}_t = \\prod_{i=1}^t (1 - \\beta_t)$ , and the noise prediction loss is combined with the stereo-aware loss by a scaling factor:", + "bbox": [ + 511, + 464, + 905, + 496 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nl = l _ {\\epsilon} + \\lambda \\cdot l _ {d} \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 660, + 507, + 903, + 522 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Training Strategy", + "text_level": 1, + "bbox": [ + 513, + 542, + 686, + 560 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The training process is divided into two stages. In the first training stage, we focus on individual frames from videos. The temporal layer in denoising U-Net is frozen, and cross-view attention in TIL module is removed. The Refiner-Net model and the denoising U-Net are initialized with pretrained weights from SD2.1 [50] and SVD [7] respectively. The weights of the variational autoencoder encoder and decoder, as well as the contrastive language-image pretraining (CLIP) image encoder, are all kept fixed. The aim of this stage is for the model to learn to synthesize new viewpoint images under the condition of a reference image and a new viewpoint pose. In the second stage, we train the temporal layer and cross-view attention in TIL module with video sequences. This enables the model to capture temporal context information efficiently. The input for the model is the 8-frame video clip and $N_r$ for TIL is set to 8.", + "bbox": [ + 511, + 566, + 906, + 809 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 821, + 645, + 839 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Datasets, Baselines and Metrics", + "text_level": 1, + "bbox": [ + 511, + 847, + 790, + 863 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We conduct our experiments on image and video levels with RealEstate10K [78] and self-collected video data.", + "bbox": [ + 511, + 869, + 905, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "815", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "RealEstate10K is a large, widely used dataset with camera poses, corresponding to 10 million frames derived from about 80,000 video clips, gathered from about 10,000 YouTube videos. As there is not open-source stereo video benchmark, we collect 1500 monocular videos with 1400 for training, 100 for testing. For image synthesis, we compared the proposed method with recent open-source SOTA methods: 3D-photography [56], SynSim [69], AdaMPI [21], SinMPI [47], Wang et al. [66], Photoconsistent-NVS(P-NVS) [75], NVSVDE-Net [4], CoPoNeRF [24] and MVsplat [14]. For a fair comparison, the 3D photography method, AdaMPI, SinMPI, and the proposed method all used MiDaS [49] for depth estimation. Following the method of 3D-photography [56], we measured the accuracy of the generated target views with regard to the ground-truth images using the three metrics of the learned perceptual image patch similarity (LPIPS), peak signal-to-noise ratio (PSNR), and structural similarity index measure (SSIM).", + "bbox": [ + 93, + 90, + 480, + 361 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For the video synthesis, we picked several representative approaches, i.e., the method of 3D-photography [56], the open-source code repository webui-depthmap, P-NVS [75], CoPoNeRF [24], NVS-Solver [74] and AVP [44]. Moreover, we conducted the same depth estimation method [49] for a fair comparison of 3D-photography [56], webui-depthmap and NVS-Solver [74]. We employ the FVD [62] score to measure the perceptual similarity between input videos and outputs and report the flow warping error $E_{\\text{warp}}$ [35] to assess the temporal consistency of the resulting video sequences.", + "bbox": [ + 93, + 364, + 480, + 527 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Implementation", + "text_level": 1, + "bbox": [ + 93, + 540, + 248, + 555 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We employed the AdamW [39] optimizer with a learning rate of $1 \\times 10^{-5}$ to train the model on two NVIDIA A800 GPUs. In the first training stage, individual video frames were resized and center-cropped to a resolution of $1024 \\times 1024$ . We trained the model for 100,000 iterations with a batch size of 8. In the second training stage, the videos were first split according to the transitions to ensure that a scene only appeared in one video clip. We then formed each video clip for training with eight frames. Finally, the temporal layer and TIL module was trained for 10,000 steps with a batch size of 2. The stereo-aware loss scaling factor $\\lambda$ was set to 0.001 and $\\lambda$ in Eq. 3 was set as 0.6.", + "bbox": [ + 93, + 563, + 480, + 743 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Qualitative Results", + "text_level": 1, + "bbox": [ + 93, + 755, + 271, + 768 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3.1 Stereo Image Synthesis", + "text_level": 1, + "bbox": [ + 93, + 779, + 303, + 792 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We visually compared different methods for generating new view images from a target viewpoint. We qualitatively compared the newly synthesized images in Figure 5. 3D-photography fails to resolve the image distortion problem, and the image appears to be enhanced only slightly by the", + "bbox": [ + 93, + 803, + 480, + 876 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "rendering. SynSin generates the most blurry details and loses a lot of information, compared with the input view. Neither AdaMPI nor SinMPI can generate fine enough details, and the results contain noticeable artifacts. CoPoN-erF tends to produce discrepancies from the original image in terms of fine-grained details, and it struggles with effectively handling scenes involving motion, as well as MVsplat. In contrast, the proposed method shows a clean and detailed output, especially in the area of depth information discontinuity.", + "bbox": [ + 516, + 90, + 903, + 241 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3.2 Stereo Video Synthesis", + "text_level": 1, + "bbox": [ + 516, + 281, + 723, + 294 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Moreover, we qualitatively compared the results of the stereo video synthesis with other methods. In addition to the quality issues of single-frame generation, the most important aspect in videos is the temporal consistency. We seek to generate videos that are free of flicker. The visualization of generated videos from different methods are illustrated in Figure 6, for each video, we show the scanline (the yellow line in figure) slice through the spatial-temporal volume. Our method achieve the smoothest trajectories, while other methods exhibit significant flickering. The most distinct regions are zoomed in and highlighted under the images to illustrate the details. In the first case, our method attains the smooth trajectories of the background (maintaining the straight line). And the second case, only our method preserves the texture of the trees behind the woman. And the last case, only our method preserves the texture of the black vertical lines. More results of generated frames are also presented in supplementary material.", + "bbox": [ + 516, + 310, + 903, + 580 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Quantitative Results", + "text_level": 1, + "bbox": [ + 516, + 603, + 705, + 617 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1 lists the SSIM, PSNR, and LPIPS scores for the newly synthesized images. The proposed method outperforms 3D-photography and SynSin in the major metrics by a wide margin. The method of Wang et al., P-NVS and NVSVDE-Net perform slightly worse in terms of all metrics compared to the proposed method. Since the proposed method merges the spatial features and fills the occluded regions, it achieves the best LPIPS scores and shows competitive results in PSNR and SSIM when compared with AdaMPI and CoPoNeRF. MVSplat outperforms other methods when $t = 10$ but is inferior to ours when $t = 5$ . Overall, the results exhibit superior visual effects, which is consistent with the performance observed in the qualitative experiments. We further demonstrate the effectiveness of our method on video synthesis, as shown in Table 2. The proposed method achieves the best performance on both FVD and $E_{\\text{warp}}$ , which means that our generated videos attain the best quality and temporal consistency.", + "bbox": [ + 516, + 628, + 903, + 898 + ], + "page_idx": 5 + }, + { + "type": "page_footnote", + "text": "https://github.com/thygate/stable-diffusion-webui-depthmap-script", + "bbox": [ + 116, + 887, + 465, + 898 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "816", + "bbox": [ + 486, + 946, + 511, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/86c7f958e9f5db0238d862fe33cbd5bdb019e9307947a7de00988dc1e58aaee6.jpg", + "image_caption": [ + "Figure 5. Qualitative results on the RealEstate10K dataset. The proposed method generates better-quality information and maintains the edge and textural details of the reference image better than the other methods. More results are available in supplementary material." + ], + "image_footnote": [], + "bbox": [ + 109, + 90, + 883, + 309 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d9b5162f3405508b0ac7cf01973d65d5bca984da229fd3cb8df89e4f02d3c9d7.jpg", + "image_caption": [ + "Figure 6. Visual comparisons in stereo video benchmark. We show the scanline (yellow line in original frame) slice through the spatial-temporal volume. The most distinct regions are zoomed in and highlighted under the images to illustrate the details." + ], + "image_footnote": [], + "bbox": [ + 99, + 354, + 872, + 539 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/24b1891b54fa31e041b0e92c79ce76e931da71388d454afe23cefbdd6c07ea0d.jpg", + "image_caption": [ + "Figure 7. Ablation study of different designs. Only the proposed method ensures consistent preservation of details with the groundtruth target view. More results are available in supplementary material." + ], + "image_footnote": [], + "bbox": [ + 107, + 599, + 473, + 747 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 834, + 243, + 849 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5.1 Effectiveness of each components", + "text_level": 1, + "bbox": [ + 89, + 857, + 375, + 872 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We show the results in Table 3 on the RealEstate10K (stereo image) and collected stereo video datasets. CN means sub", + "bbox": [ + 89, + 881, + 483, + 911 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "stituting our RefinerNet (RN) to ControlNet, SDS means the proposed stereo deviation strength. For image-level comparison, the results show that utilizing U-Net architecture only is not sufficient for stereo image synthesis and our RefinerNet is a better choice than ControlNet. Adding SDS further improves the geometric consistency. As visualized in Figure 7, solely relying on U-Net features results in distortion in the edge area and struggles with generating content properly. It also fails to preserve textural details near a plane boundary. Adding ControlNet often results in content inconsistency, presumably due to the failure of the paired view feature integration. RefinerNet handles this problem, showing it is a better architecture for stereo image generation. SDS further improves the consistency with the input and contain a more detailed image structure. As for video-level comparison, the quantitative results are shown in Table 3, adding TIL and DVG both attain the performance gain, and our final model achieves the best performance.", + "bbox": [ + 511, + 599, + 906, + 871 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "817", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/327cec0a2f4369c3c6adf54baaf4c9f4eed8f7cee3f56154f1cd6858f6753d9d.jpg", + "table_caption": [ + "Table 1. Quantitative comparison on RealEstate10K dataset. $\\dagger$ means re-implementing in our setting for fair comparison. When $t = 5$ , the current frame serves as the left view and the frame at time +5 serves as the right view; the same approach applies for $t = 10$ . The best results are in bold." + ], + "table_footnote": [], + "table_body": "
SSIM ↑PSNR ↑LPIPS ↓
t=5t=10t=5t=10t=5t=10
3D-photography [56]†0.6410.44611.089.5330.1160.386
SynSin [69]†0.7400.64521.2219.550.0870.104
AdaMPI [21]†0.8690.73627.7022.650.0410.087
SinMPI [47]†0.8980.77528.7923.210.0400.080
Wang et al. [66]0.8400.72025.3521.360.0490.095
P-NVS [75]†0.7240.66721.9819.460.2410.279
NVSVDE-Net [4]0.84024.310.233
CoPoNeRF [24]†0.6580.63222.6321.330.1710.209
MVSplat [14]†0.86325.890.132
Proposed0.9160.85732.2624.860.0380.049
", + "bbox": [ + 91, + 169, + 498, + 375 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/830c44823b90b4fa6bc24ce946110991266aff61c09d36fefefd1f934b3b7a63.jpg", + "table_caption": [ + "Table 2. Quantitative comparison with the other methods. The best results are in **bold**. $E_{\\text{warp}}^*$ denotes $E_{\\text{warp}} \\left( \\times 10^{-3} \\right)$ ." + ], + "table_footnote": [], + "table_body": "
FVD ↓E* warp ↓
3D-photography [56]155.03.418
Webui-depthmap83.183.486
P-NVS [75]420.223.02
CoPoNeRF [24]290.07.144
NVS-Solver [74]249.15.842
MVSplat [14]15437.913
AVP [44]99.923.446
Proposed67.093.374
", + "bbox": [ + 122, + 421, + 450, + 575 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/6029d5f9f732f184793bbfb64425fed1e2fcda89b05e4abeef4199e88b806b1c.jpg", + "table_caption": [ + "Table 3. Quantitative comparison for the ablation study. The best results are in bold." + ], + "table_footnote": [], + "table_body": "
MethodCNRNSDSTILDVGStereo ImageStereo Video
SSIM↑PSNR↑LPIPS↓FVD↓
Image-level✓0.88023.730.06-
0.85524.040.183-
✓0.89530.200.043-
✓✓0.91632.260.038-
Video-level✓✓---184.0
✓✓✓----123.5
✓✓✓---85.21
✓✓✓✓---67.09
", + "bbox": [ + 94, + 628, + 480, + 757 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5.2 Data improvement by DVG", + "text_level": 1, + "bbox": [ + 89, + 784, + 336, + 800 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We show the generated stereo occlusion mask with interframe motion refinement in Figure 8. It can be observed that the occlusions appear more complete spatially and exhibit smoother in their temporal distribution. By employing DVG, we can acquire sufficient data to support the generation of spatio-temporally consistent videos. More analyses", + "bbox": [ + 89, + 809, + 483, + 902 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/776aa8844cfaa92050a9122ee32602002cc6d873b8640d0800de504c952bd9b1.jpg", + "image_caption": [ + "Figure 8. Visual comparisons for the improvement of stereo occlusion. We show the scanline of (yellow line in original frame) slice through the spatial-temporal volume. The red region represents the changes of occlusion mask." + ], + "image_footnote": [], + "bbox": [ + 522, + 88, + 647, + 186 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/10e1e42214e722de071131e3b2234a58c99e0ecccd585d0d8b581fe539a449b4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 647, + 88, + 771, + 186 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5666cf02f20695baf24a3ee0a7ab9d6399bce9c90c61d915097c0fe5325b141e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 771, + 88, + 898, + 186 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "are available in the supplementary materials.", + "bbox": [ + 513, + 280, + 808, + 295 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5.3 Exploration of different depth estimation methods", + "text_level": 1, + "bbox": [ + 511, + 316, + 906, + 347 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We further compared different design choices of depth map including DepthAnything [73], Marigold [31], MiDaS [49], ZoeDepth [5] and DepthCrafter [26]. The results are available in supplementary materials. The proposed method can produce new viewpoint images that are realistic, seamless, and rich in detail, regardless of the depth estimation method used. This capability demonstrates the versatility and robustness of the proposed method, ensuring superior performance in various applications and environments.", + "bbox": [ + 511, + 356, + 906, + 492 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.6. Limitations", + "text_level": 1, + "bbox": [ + 513, + 503, + 638, + 518 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "SpatialDreamer effectively creates stereo video from a single input but faces limitations. It relies on depth maps for rendering, affecting stereoscopic quality due to depth accuracy and distribution. Future research could focus on integrating implicit depth to address this [36]. Additionally, the model's large parameter size limits real-time performance.", + "bbox": [ + 511, + 527, + 905, + 619 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 633, + 633, + 648 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we introduce a new self-supervised stereo video synthesis approach using a video diffusion model, termed SpatialDreamer. This method addresses the issues of insufficient data and spatio-temporal inconsistency between frames. To solve the problem of insufficient data, we develop a Depth-based Video Generation (DVG) module that uses a forward-backward rendering mechanism to generate a rendered video with geometric and temporal priors. Additionally, we propose RefinerNet along with a self-supervised synthetic framework to enable efficient and dedicated training using data generated by DVG. Furthermore, we design a consistency control module to ensure geometric and temporal consistency. Our SpatialDreamer outperforms all other open-source stereo image and video synthesis methods and has the potential for future expansion into virtual reality applications.", + "bbox": [ + 511, + 659, + 906, + 901 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "818", + "bbox": [ + 485, + 945, + 511, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 90, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Dmitry Akimov, Alexey Shestov, Alexander Voronov, and Dmitriy Vatolin. Occlusion refinement for stereo video using optical flow. In 2012 International Conference on 3D Imaging (IC3D), pages 1-8, 2012.", + "[2] Andrea Schubert. visionos 2 brings new spatial computing experiences to apple vision pro. https://www.apple.com/newsroom/2024/06/visionos-2-brings-new-spatial-computing-experiences-to-apple-vision-pro, 2024.", + "[3] Omer Bar-Tal, Dolev Ofri-Amar, Rafail Fridman, Yoni Kasten, and Tali Dekel. Text2live: Text-driven layered image and video editing. In European conference on computer vision, pages 707-723. Springer, 2022.", + "[4] Juan Luis Gonzalez Bello and Munchurl Kim. Novel view synthesis with view-dependent effects from a single image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024.", + "[5] Shariq Farooq Bhat, Reiner Birkl, Diana Wofk, Peter Wonka, and Matthias Müller. Zoedepth: Zero-shot transfer by combining relative and metric depth, 2023.", + "[6] Phuong Ngoc Binh Do and Quoc Chi Nguyen. A review of stereo-photogrammetry method for 3-d reconstruction in computer vision. In 2019 19th International Symposium on Communications and Information Technologies (ISCIT), pages 138–143, 2019.", + "[7] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023.", + "[8] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22563-22575, 2023.", + "[9] Aaron Bobick and Stephen Intille. Large occlusion stereo. International Journal of Computer Vision, 33:181-200, 1999.", + "[10] Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Yufei Guo, Li Jing, David Schnurr, Joe Taylor, Troy Luhman, Eric Luhman, Clarence Ng, Ricky Wang, and Aditya Ramesh. Video generation models as world simulators. 2024.", + "[11] Mingdeng Cao, Xintao Wang, Zhongang Qi, Ying Shan, Xiaohu Qie, and Yinqiang Zheng. Masactrl: Tuning-free mutual self-attention control for consistent image synthesis and editing, 2023.", + "[12] David Charatan, Sizhe Li, Andrea Tagliasacchi, and Vincent Sitzmann. pixelsplat: 3d gaussian splats from image pairs for scalable generalizable 3d reconstruction. In arXiv, 2023.", + "[13] Weifeng Chen, Jie Wu, Pan Xie, Hefeng Wu, Jiashi Li, Xin Xia, Xuefeng Xiao, and Liang Lin. Control-a-video: Controllable text-to-video generation with diffusion models. arXiv preprint arXiv:2305.13840, 2023." + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Yuedong Chen, Haofei Xu, Chuanxia Zheng, Bohan Zhuang, Marc Pollefeys, Andreas Geiger, Tat-Jen Cham, and Jianfei Cai. Mvsplat: Efficient 3d gaussian splatting from sparse multi-view images. arXiv preprint arXiv:2403.14627, 2024.", + "[15] Zuozhuo Dai, Zhenghao Zhang, Yao Yao, Bingxue Qiu, Siyu Zhu, Long Qin, and Weizhi Wang. Animateanything: Fine-grained open domain image animation with motion guidance. arXiv e-prints, pages arXiv-2311, 2023.", + "[16] G. Egnal and R.P. Wildes. Detecting binocular half-occlusions: empirical comparisons of five approaches. IEEE Transactions on Pattern Analysis and Machine Intelligence, 24(8):1127-1133, 2002.", + "[17] Patrick Esser, Johnathan Chiu, Parmida Atighehchian, Jonathan Granskog, and Anastasis Germanidis. Structure and content-guided video synthesis with diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7346-7356, 2023.", + "[18] Chen Gao, Ayush Saraf, Johannes Kopf, and Jia-Bin Huang. Dynamic view synthesis from dynamic monocular video, 2021.", + "[19] Yuwei Guo, Ceyuan Yang, Anyi Rao, Zhengyang Liang, Yaohui Wang, Yu Qiao, Maneesh Agrawala, Dahua Lin, and Bo Dai. Animatediff: Imagine your personalized text-to-image diffusion models without specific tuning. International Conference on Learning Representations, 2024.", + "[20] Agrim Gupta, Lijun Yu, Kihyuk Sohn, Xiuye Gu, Meera Hahn, Li Fei-Fei, Irfan Essa, Lu Jiang, and José Lezama. Photorealistic video generation with diffusion models. arXiv preprint arXiv:2312.06662, 2023.", + "[21] Yuxuan Han, Ruicheng Wang, and Jiaolong Yang. Single-view view synthesis in the wild with learned adaptive multiplane images. In ACM SIGGRAPH 2022 Conference Proceedings, pages 1-8, 2022.", + "[22] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control, 2022.", + "[23] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2022.", + "[24] Sunghwan Hong, Jaewoo Jung, Heeseong Shin, Jiaolong Yang, Seungryong Kim, and Chong Luo. Unifying correspondence, pose and nerf for pose-free novel view synthesis from stereo pairs, 2024.", + "[25] Li Hu, Xin Gao, Peng Zhang, Ke Sun, Bang Zhang, and Liefeng Bo. animate anyone: Consistent and controllable image-to-video synthesis for character animation. arXiv preprint arXiv:2311.17117, 2023.", + "[26] Wenbo Hu, Xiangjun Gao, Xiaoyu Li, Sijie Zhao, Xiaodong Cun, Yong Zhang, Long Quan, and Ying Shan. Depthcrafter: Generating consistent long depth sequences for open-world videos. arXiv preprint arXiv:2409.02095, 2024.", + "[27] Varun Jampani, Huiwen Chang, Kyle Sargent, Abhishek Kar, Richard Tucker, Michael Krainin, Dominik Kaeser, William T Freeman, David Salesin, Brian Curless, et al. Slide: Single image 3d photography with soft layering and depth-aware inpainting. In Proceedings of the IEEE/CVF In" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "819", + "bbox": [ + 485, + 945, + 511, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ternational Conference on Computer Vision, pages 12518-12527, 2021.", + "[28] Wonbong Jang and Lourdes Agapito. Nvist: In the wild new view synthesis from a single image with transformers, 2024.", + "[29] Jisoo Jeong, Hong Cai, Risheek Garrepalli, Jamie Menjay Lin, Munawar Hayat, and Fatih Porikli. Ocai: Improving optical flow estimation by occlusion and consistency aware interpolation, 2024.", + "[30] Hailin Jin, Stefano Soatto, and Anthony J Yezzi. Multi-view stereo reconstruction of dense shape and complex appearance. International Journal of Computer Vision, 63:175-189, 2005.", + "[31] Bingxin Ke, Anton Obukhov, Shengyu Huang, Nando Metzger, Rodrigo Caye Daudt, and Konrad Schindler. Repurposing diffusion-based image generators for monocular depth estimation, 2024.", + "[32] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering, 2023.", + "[33] Levon Khachatriyan, Andranik Movsisyan, Vahram Tadevosyan, Roberto Henschel, Zhangyang Wang, Shant Navasardyan, and Humphrey Shi. Text2video-zero: Text-to-image diffusion models are zero-shot video generators. arXiv preprint arXiv:2303.13439, 2023.", + "[34] Dan Kondratyuk, Lijun Yu, Xiuye Gu, José Lezama, Jonathan Huang, Rachel Hornung, Hartwig Adam, Hassan Akbari, Yair Alon, Vighnesh Birodkar, et al. Videopoet: A large language model for zero-shot video generation. arXiv preprint arXiv:2312.14125, 2023.", + "[35] Wei-Sheng Lai, Jia-Bin Huang, Oliver Wang, Eli Shechtman, Ersin Yumer, and Ming-Hsuan Yang. Learning blind video temporal consistency. In Proceedings of the European conference on computer vision (ECCV), pages 170-185, 2018.", + "[36] Hanwen Liang, Junli Cao, Vedit Goel, Guocheng Qian, Sergei Korolev, Demetri Terzopoulos, Konstantinos Plataniotis, Sergey Tulyakov, and Jian Ren. Wonderland: Navigating 3d scenes from a single image. arXiv preprint arXiv:2412.12091, 2024.", + "[37] Andrew Liu, Richard Tucker, Varun Jampani, Ameesh Makadia, Noah Snavely, and Angjoo Kanazawa. Infinite nature: Perpetual view generation of natural scenes from a single image. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021.", + "[38] Yixin Liu, Kai Zhang, Yuan Li, Zhiling Yan, Chujie Gao, Ruoxi Chen, Zhengqing Yuan, Yue Huang, Hanchi Sun, Jianfeng Gao, Lifang He, and Lichao Sun. Sora: A review on background, technology, limitations, and opportunities of large vision models, 2024.", + "[39] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017.", + "[40] Kevin Matzen, Michael F. Cohen, Bryce Evans, Johannes Kopf, and Richard Szeliski. Low-cost 360 stereo photography and video capture. ACM Trans. Graph., 36(4), 2017.", + "[41] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis, 2020." + ], + "bbox": [ + 91, + 90, + 482, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[42] Norman Müller, Katja Schwarz, Barbara Rössle, Lorenzo Porzi, Samuel Rota Bulò, Matthias Nießner, and Peter Kontschieder. Multidiff: Consistent novel view synthesis from a single image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10258-10268, 2024.", + "[43] Simon Niklaus and Feng Liu. Softmax splatting for video frame interpolation, 2020.", + "[44] Onee. What is spatial video on iphone 15 pro and vision pro. https://xreality-zone/en/posts/what-is-spatial-video-on-iphone-15-pro-and/apple-vision-pro, 2024.", + "[45] Hao Ouyang, Qiuyu Wang, Yuxi Xiao, Qingyan Bai, Juntao Zhang, Kecheng Zheng, Xiaowei Zhou, Qifeng Chen, and Yujun Shen. Codef: Content deformation fields for temporally consistent video processing. arXiv preprint arXiv:2308.07926, 2023.", + "[46] Onur Özyeşil, Vladislav Voroninski, Ronen Basri, and Amit Singer. A survey of structure from motion*. Acta Numerica, 26:305-364, 2017.", + "[47] Guo Pu, Peng-Shuai Wang, and Zhouhui Lian. Sinmpi: Novel view synthesis from a single image with expanded multiplane images, 2023.", + "[48] Chenyang Qi, Xiaodong Cun, Yong Zhang, Chenyang Lei, Xintao Wang, Ying Shan, and Qifeng Chen. Fatezero: Fusing attentions for zero-shot text-based video editing. arXiv preprint arXiv:2303.09535, 2023.", + "[49] René Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(3), 2022.", + "[50] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022.", + "[51] Christoph Sager, Christian Janiesch, and Patrick Zschech. A survey of image labelling for computer vision applications. Journal of Business Analytics, 4(2):91-110, 2021.", + "[52] D. Scharstein. View Synthesis Using Stereo Vision. Springer Berlin Heidelberg, 1999.", + "[53] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4104-4113, 2016.", + "[54] S.M. Seitz, B. Curless, J. Diebel, D. Scharstein, and R. Szeliski. A comparison and evaluation of multi-view stereo reconstruction algorithms. In 2006 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR'06), pages 519-528, 2006.", + "[55] Yujiao Shi, Hongdong Li, and Xin Yu. Self-supervised visibility learning for novel view synthesis.*. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2021.", + "[56] Meng-Li Shih, Shih-Yang Su, Johannes Kopf, and Jia-Bin Huang. 3d photography using context-aware layered depth" + ], + "bbox": [ + 516, + 90, + 906, + 901 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "820", + "bbox": [ + 485, + 945, + 511, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "inpainting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8028-8038, 2020.", + "[57] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792, 2022.", + "[58] Lorenzo Stacchio. Train stable diffusion for inpainting, 2023.", + "[59] Wenxiu Sun, Lingfeng Xu, Oscar C Au, Sung Him Chui, and Chun Wing Kwok. An overview of free view-point depth-image-based rendering (dibr). In APSIPA Annual Summit and Conference, pages 1023-1030, 2010.", + "[60] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow, 2020.", + "[61] Richard Tucker and Noah Snavely. Single-view view synthesis with multiplane images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 551-560, 2020.", + "[62] Thomas Unterthiner, Sjoerd van Steenkiste, Karol Kurach, Raphaël Marinier, Marcin Michalski, and Sylvain Gelly. FVD: A new metric for video generation. In Deep Generative Models for Highly Structured Data, ICLR 2019 Workshop, New Orleans, Louisiana, United States, May 6, 2019. OpenReview.net, 2019.", + "[63] Basile Van Hoorick, Rundi Wu, Ege Ozguroglu, Kyle Sargent, Ruoshi Liu, Pavel Tokmakov, Achal Dave, Changxi Zheng, and Carl Vondrick. Generative camera dolly: Extreme monocular dynamic novel view synthesis. European Conference on Computer Vision (ECCV), 2024.", + "[64] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017.", + "[65] Wen Wang, Yan Jiang, Kangyang Xie, Zide Liu, Hao Chen, Yue Cao, Xinlong Wang, and Chunhua Shen. Zero-shot video editing using off-the-shelf image diffusion models. arXiv preprint arXiv:2303.17599, 2023.", + "[66] Xiaodong Wang, Chenfei Wu, Shengming Yin, Minheng Ni, Jianfeng Wang, Linjie Li, Zhengyuan Yang, Fan Yang, Lijuan Wang, Zicheng Liu, et al. Learning 3d photography videos via self-supervised diffusion on single images. arXiv preprint arXiv:2302.10781, 2023.", + "[67] Xiang Wang, Hangjie Yuan, Shiwei Zhang, Dayou Chen, Jiuniu Wang, Yingya Zhang, Yujun Shen, Deli Zhao, and Jingren Zhou. Videocomposer: Compositional video synthesis with motion controllability. Advances in Neural Information Processing Systems, 36, 2024.", + "[68] Chung-Yi Weng, Brian Curless, and Ira Kemelmacher-Shlizerman. Photo wake-up: 3d character animation from a single photo. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5908-5917, 2019.", + "[69] Olivia Wiles, Georgia Gkioxari, Richard Szeliski, and Justin Johnson. Synsin: End-to-end view synthesis from a single image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020." + ], + "bbox": [ + 91, + 90, + 483, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[70] Andrew Woods, Tom Docherty, and Rolf Koch. Image distortions in stereoscopic video systems. Proc SPIE, 1915, 2002.", + "[71] Jay Zhangjie Wu, Yixiao Ge, Xintao Wang, Stan Weixian Lei, Yuchao Gu, Yufei Shi, Wynne Hsu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7623-7633, 2023.", + "[72] Dejia Xu, Yifan Jiang, Peihao Wang, Zhiwen Fan, Humphrey Shi, and Zhangyang Wang. Sinnerf: Training neural radiance fields on complex scenes from a single image, 2022.", + "[73] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data, 2024.", + "[74] Meng You, Zhiyu Zhu, Hui Liu, and Junhui Hou. Nvs-solver: Video diffusion model as zero-shot novel view synthesizer. arXiv preprint arXiv:2405.15364, 2024.", + "[75] Jason J. Yu, Fereshteh Forghani, Konstantinos G. Derpanis, and Marcus A. Brubaker. Long-term photometric consistent novel view synthesis with diffusion models, 2023.", + "[76] Guofeng Zhang, Wei Hua, Xueying Qin, Tien-Tsin Wong, and Hujun Bao. Stereoscopic video synthesis from a monocular video. IEEE Transactions on Visualization and Computer Graphics, 13(4):686-696, 2007.", + "[77] Min Zhao, Rongzhen Wang, Fan Bao, Chongxuan Li, and Jun Zhu. Controlvideo: Adding conditional control for one shot text-to-video editing. arXiv preprint arXiv:2305.17098, 2023.", + "[78] Tinghui Zhou, Richard Tucker, John Flynn, Graham Fyffe, and Noah Snavely. Stereo magnification: Learning view synthesis using multiplane images. CoRR, abs/1805.09817, 2018." + ], + "bbox": [ + 516, + 90, + 903, + 556 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "821", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/SpatialDreamer_ Self-supervised Stereo Video Synthesis from Monocular Input/24175650-223e-41c7-b804-6cc94c2c5191_model.json b/2025/SpatialDreamer_ Self-supervised Stereo Video Synthesis from Monocular Input/24175650-223e-41c7-b804-6cc94c2c5191_model.json new file mode 100644 index 0000000000000000000000000000000000000000..346b4430dfccefc86c1964e7ab972e48059ccfba --- /dev/null +++ b/2025/SpatialDreamer_ Self-supervised Stereo Video Synthesis from Monocular Input/24175650-223e-41c7-b804-6cc94c2c5191_model.json @@ -0,0 +1,2378 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.099, + 0.13, + 0.898, + 0.154 + ], + "angle": 0, + "content": "SpatialDreamer: Self-supervised Stereo Video Synthesis from Monocular Input" + }, + { + "type": "text", + "bbox": [ + 0.115, + 0.192, + 0.88, + 0.232 + ], + "angle": 0, + "content": "Zhen Lv\\(^{1}\\), Yangqi Long\\(^{1}\\), Congzhentao Huang\\(^{1}\\), Cao Li\\(^{1}\\), Chengfei Lv\\(^{1\\dagger}\\), Hao Ren\\(^{1}\\), Dian Zheng\\(^{2}\\), \\(^{1}\\)Alibaba Group; \\(^{2}\\)Sun Yat-sen University" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.277, + 0.327, + 0.293 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.309, + 0.486, + 0.701 + ], + "angle": 0, + "content": "Stereo video synthesis from monocular input is challenging in spatial computing and virtual reality due to the lack of high-quality stereo video pairs for training and the difficulty of maintaining spatio-temporal consistency between frames. Existing methods primarily address these issues by directly applying novel view synthesis (NVS) techniques to video, while facing limitations such as the inability to effectively represent dynamic scenes and the requirement for extensive training data. In this paper, we introduce a novel self-supervised stereo video synthesis paradigm via a video diffusion model, termed SpatialDreamer, which meets the challenges head-on. Firstly, to address the stereo video data insufficiency, we propose a Depth based Video Generation module DVG, which employs a forward-backward rendering mechanism to generate paired videos with geometric and temporal priors. Leveraging data generated by DVG, we propose RefinerNet along with a self-supervised synthetic framework designed to facilitate efficient and dedicated training. More importantly, we devise a consistency control module, which consists of a metric of stereo deviation strength and a Temporal Interaction Learning module TIL for geometric and temporal consistency assurance respectively. We evaluated the proposed method against various benchmark methods, with the results showcasing its superior performance. Our project website is at: https://spatialdreamer.github.io." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.731, + 0.222, + 0.745 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.756, + 0.483, + 0.847 + ], + "angle": 0, + "content": "Stereo video synthesis from a monocular input aims to generate the target-view video based on the given view with geometric and spatio-temporal consistency, which has a wide range of applications in 3D model reconstruction [6], 3D movie production [18, 40], and Apple Vision Pro (AVP) [44] like virtual reality content." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.847, + 0.483, + 0.877 + ], + "angle": 0, + "content": "The primary difficulties of this task stem from the lack of adequate high-quality paired stereo videos for training, and" + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.283, + 0.905, + 0.521 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.535, + 0.907, + 0.591 + ], + "angle": 0, + "content": "Figure 1. Visualization of the temporal consistency in stereo video generated by different methods. We extract the yellow line of each frame and stack them together. A good result should show a natural transition in the \\(t\\) dimension." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.614, + 0.907, + 0.901 + ], + "angle": 0, + "content": "the challenge of preserving the geometric consistency between two views as well as maintaining the temporal consistency across generated frames. Traditional methods for generating stereo content involve capturing scenes using dual-camera setups [52, 59, 70, 76]. However, the acquisition of these images involves the use of a professional-grade camera, resulting in substantial costs. Recent advancements in deep learning mainly handle these problems by directly applying single to stereo or multi-view image generation techniques to video, while often facing challenges related to geometric and spatio-temporal inconsistencies [21, 32, 41, 50, 66, 67]. As shown in Fig 1, methods based on NVs are challenging to maintain accurate spatio-temporal consistency, and even the recently released AVP 3D photo converter [2] may inevitably result in content flickering and inconsistency when applied to video synthesis. This is due to the difficulty of handling the temporal coherence between frames and the spatial information between paired views in complex dynamic scenes." + }, + { + "type": "page_footnote", + "bbox": [ + 0.115, + 0.887, + 0.245, + 0.9 + ], + "angle": 0, + "content": "† Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "811" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.094, + 0.092, + 0.482, + 0.513 + ], + "angle": 0, + "content": "In this paper, we propose a self-supervised stereo video synthesis paradigm via a video diffusion model, termed SpatialDreamer, which will meet the data insufficiency and spatio-temporal inconsistency at once. Firstly, we design a Depth based Video data Generation module DVG to handle the problem of data insufficiency. Similar to [21, 66], DVG constructs a training pair of two views using forward-backward rendering without the need for data annotation [51, 59, 70], while [21, 66] only render individual images, resulting in flickering effects in generated video [9]. In this paper, we improve the generation of video data for refining stereo occlusion masks by leveraging the inter-frame motion obtained from optical flow [29]. This methodology enables DVG to produce paired video data that maintains both geometric and temporal consistency. Leveraging paired video generated by DVG, we propose Refiner-Net, along with a self-supervised video synthesis framework, to enable efficient and targeted training. More importantly, with access to sufficient paired videos, we devise a consistency control module, which consists of a metric of stereo deviation strength and a Temporal Interaction Learning module TIL. The stereo deviation strength aims to enable the generation of stereo videos in diverse real-world scenarios, with a stereo-aware loss further supervising the model to learn the magnitude of difference between paired view features in the latent space. TIL integrates the latent features from long-temporal frames as a global information to augment the temporal coherence of the generated results." + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.516, + 0.482, + 0.62 + ], + "angle": 0, + "content": "Extensive experiments compared with various benchmark methods show that we achieve state-of-the-art performance. Notably, our SpatialDreamer meets the demand of real-world application without jitter, geometric and temporal inconsistency, which even beats AVP 3D converter and any open-source stereo video synthesis methods. We summarize our contributions as follows:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.623, + 0.48, + 0.695 + ], + "angle": 0, + "content": "- A novel self-supervised stereo video synthesis framework, SpatialDreamer, is proposed, which is robust across a wide range of scenes and dynamic content in video. Moreover, equipping the DVG module, we build a stereo video dataset using a self-supervised approach." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.698, + 0.48, + 0.755 + ], + "angle": 0, + "content": "- A consistency control module is devised, which consists of a metric of stereo deviation strength and a temporal interaction learning module, ensuring geometric and temporal consistency in video sequences." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.759, + 0.48, + 0.799 + ], + "angle": 0, + "content": "- The sufficient experiments and quantitative results demonstrate that the proposed method outperforms the age-of-the-art methods, even beats AVP." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.623, + 0.48, + 0.799 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.095, + 0.821, + 0.231, + 0.835 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.095, + 0.847, + 0.391, + 0.861 + ], + "angle": 0, + "content": "2.1. Single-image Novel View Synthesis" + }, + { + "type": "text", + "bbox": [ + 0.095, + 0.872, + 0.48, + 0.9 + ], + "angle": 0, + "content": "Early studies address the single-image novel view synthesis task without relying on explicit 3D representation," + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.092, + 0.904, + 0.348 + ], + "angle": 0, + "content": "such as End-to-end Synthesis [69], Infinite-Nature [37] and Animatable 3D Characters [68]. These methods lack explicit 3D representations for scenes, leading to inconsistent views and causing flickering and blurring, especially in occluded regions. Layer-based methods, such as 3D-photography [56] and SLIDE [27] represent a 3D scene using discrete layers. Yet, the synthesized views may lack accuracy in texture and geometric consistency across different perspectives. SMPI [61], AdaMPI [21] and SinMPI [47] represent the 3D scene using multiple planes, while their rendered results may suffer from depth discretization artifacts and repetitive texture patterns. To reduce the reliance on large labeled datasets, Shi et.al [55], Wang et.al. [66] and NVSVDE-Net [4] explore self-supervised learning for creating novel view images from single image. Although they produce photorealistic results, they struggle with temporal and spatial consistency, especially in dynamic scenes." + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.358, + 0.904, + 0.431 + ], + "angle": 0, + "content": "As for our approach, we design a depth-based video generation module that generates paired videos with geometric and temporal priors. Additionally, Our method ensures multi-view consistency for the entire scene by employing a consistency control module." + }, + { + "type": "title", + "bbox": [ + 0.517, + 0.469, + 0.8, + 0.483 + ], + "angle": 0, + "content": "2.2. Multi-view Novel View Synthesis" + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.501, + 0.904, + 0.846 + ], + "angle": 0, + "content": "Early approaches in this field frequently utilized geometric methods, such as multi-view stereo reconstruction [30, 54] and structure-from-motion techniques [46, 53], to generate novel viewpoints from multiple images captured at different angles. NeRF-based methods were proposed to generate novel views based on fewer input images, including SinNeRF [72], CoPoNeRF [24] and NViST [28]. 3D Gaussian Splatting [32] presented an explicit representation of the scene. Gaussian-based methods, such as pixelSplat [12] and MVSplat [14], can produce high-quality synthesis results from multi-view images. Recently, researchers have begun exploring the application of advanced diffusion models for novel view synthesis, driven by their remarkable success in this area. Notable approaches such as Photoconsistent-NVS [75], MultiDiff [42], GCD [63] and NVS-Solver [74] generate image sequences based on predefined camera trajectories using either single or multiple input images. While these methods have shown promising results, they often struggle with dependence on dense scene geometry, handling complex scene dynamics, and generalization challenges [32]. Moreover, the challenge of optimizing NeRF/GS with sparse views hampers the ability of multi-view NVS to produce high-quality results." + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.857, + 0.904, + 0.9 + ], + "angle": 0, + "content": "In contrast, we explore a video diffusion model to achieve stereo video synthesis, while preserving geometric structures through a RefinerNet based framework." + }, + { + "type": "page_number", + "bbox": [ + 0.488, + 0.947, + 0.511, + 0.957 + ], + "angle": 0, + "content": "812" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.118, + 0.096, + 0.882, + 0.349 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.359, + 0.908, + 0.445 + ], + "angle": 0, + "content": "Figure 2. Overview of the proposed method, given left view video, the target view video is rendered, encoded, and concatenated with multi-frame noise, followed by the denoising U-Net architecture (i.e., SVD). The reference view images are fed into RefinerNet, through which the spatial features are extracted. (Notably, the left and right views are the target and reference view for both training and inference). The temporal interaction learning module integrates the latent features from long-temporal frames, and the deviation strength is projected as positional embedding and added to the time step embedding. Finally, the variational autoencoder decoder decodes the result into a video clip." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.455, + 0.411, + 0.47 + ], + "angle": 0, + "content": "2.3. Video Synthesis via Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.478, + 0.483, + 0.765 + ], + "angle": 0, + "content": "For video synthesis, previous works mainly expand image models to video, exploiting the image generation capability in the pretrained image diffusion models [38]. The video diffusion model (VDM) [23] and Video LDM [8] are proposed to adopt a 3D U-Net structure for joint image and video training. Subsequent works [15, 19, 19, 25, 57] also follow this insight with temporal attention, and [33, 48, 65, 71] further apply cross-frame attention to improve the consistency of the generated video. Another approach to promote temporal consistency is encoding video into 2D images [3, 45], which has been mainly explored in the video editing area. Although these models produce consistent results, they often require per-video optimization, which takes a long time, and would result in degradation when dealing with large motion. Most recently, [10, 20, 34] employ a transformer [64] architecture to model the temporal and spatial relations, demonstrating excellent video generation capabilities. However, these models need very large scale video datasets for training, which can be difficult to collect." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.766, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Beyond temporal information modeling, some works [13, 17, 67, 77] have introduced aligned spatial guidance to enhance the stability of the generated videos. The proposed framework also uses spatial guidance. Differing from those methods which require spatially aligned controls, a spatially unaligned reference image is used via a cross-attention mechanism. Furthermore, a metric of stereo deviation strength is introduced to strengthen the spatial constraint, resulting in more consistent results." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.454, + 0.673, + 0.471 + ], + "angle": 0, + "content": "3. SpatialDreamer" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.479, + 0.907, + 0.661 + ], + "angle": 0, + "content": "Given an image sequence and stereoscopic pose describing the scene depth information, the proposed method generates a stereo pair. In this section, we first provide a brief introduction about SVD [7]. Secondly, we offer detailed information of our video generation module DVG, which can generate sufficient paired video with geometric and temporal consistency. Using DVG, we present the framework for self-supervised learning-based video synthesis to support efficient and dedicated training. More importantly, to ensure consistency between frames, we design a consistency control module for geometric and temporal consistency assurance. Finally, we present the detailed training process." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.668, + 0.655, + 0.683 + ], + "angle": 0, + "content": "3.1. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.691, + 0.907, + 0.842 + ], + "angle": 0, + "content": "The challenge in video generation lies in accurately modeling the spatial-temporal dynamics, involving spatial relationships within frames and temporal relationships across frames. To address this challenge, Stable Video Diffusion (SVD) [7] introduces temporal convolution and attention layers. SVD follows the latent diffusion model (LDM) to encode video pixels into the latent space, enabling a more efficient denoising process. Given a video \\( I_{v} \\), encoder \\( \\mathcal{E} \\) encodes \\( I_{v} \\) into the latent space as \\( z_{0} = \\mathcal{E}(I_{v}) \\). Gaussian noise \\( \\epsilon \\) is then added in the Markov process:" + }, + { + "type": "equation", + "bbox": [ + 0.621, + 0.848, + 0.905, + 0.865 + ], + "angle": 0, + "content": "\\[\nz _ {t} = \\sqrt {\\bar {\\alpha} _ {t}} z _ {0} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon_ {t} \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.907, + 0.902 + ], + "angle": 0, + "content": "where \\(\\bar{\\alpha}_0,\\dots ,\\bar{\\alpha}_T\\) are the pre-defined noise schedules for T steps in the Markov process. In particular, SVD is trained" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "813" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.138 + ], + "angle": 0, + "content": "by the \\( v \\)-prediction formulation with the mean squared error (MSE) between the ground truth and its prediction in the latent space, known as \\( l_{\\epsilon} \\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.148, + 0.364, + 0.166 + ], + "angle": 0, + "content": "3.2. Depth based Video Generation" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.172, + 0.483, + 0.413 + ], + "angle": 0, + "content": "To create training pairs of images along with their corresponding occluded masks, the following self-supervised mechanism is employed utilizing real images. Firstly, the depth of the given video sample is estimated [49]. Secondly, the reference view image \\( x_{1} \\) is rendered into a masked one under the target viewpoint \\( P_{2} \\), and then the trained inpainting model [58] is used to fill these occluded regions to obtain the novel viewpoint \\( x_{2} \\) [43]. Finally, \\( x_{2} \\) is rendered to the original viewpoint \\( P_{1} \\) via a forward rendering. After this, we can obtain the masked image \\( \\widetilde{x}_{1} \\) with mask \\( M \\) under viewpoint \\( P_{1} \\), as well as the inpainted image \\( x_{2} \\) under viewpoint \\( P_{2} \\). The mask \\( M \\) and masked image \\( \\widetilde{x}_{1} \\) are considered to be the conditions, with the image \\( x_{1} \\) serving as the ground truth for the diffusion model. This approach involves encoding the inpainted image \\( x_{2} \\) into latent feature representations and feeding it into RefinerNet." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.414, + 0.483, + 0.673 + ], + "angle": 0, + "content": "However, the aforementioned approach encounters challenges in addressing inconsistencies among generated data frames, resulting in observable artifacts such as jittering and flickering. To tackle this issue, we leverage optical flow [60] to capture motion between successive video frames. Specifically, by leveraging the optical flow between the reference image \\( x_{1}^{t} \\) at time \\( t \\) and its adjacent frames, we establish the pixel correspondences between the image \\( x_{1}^{t} \\) and the images \\( x_{1}^{t-1} \\) and \\( x_{1}^{t+1} \\), respectively. Particularly, these optical flows enable computing a confidence map via forward-backward consistency [16], which allows us to more accurately merge the stereo occlusions in the masks between adjacent frames [1]. Subsequently, the refined occlusion mask \\( m^{t} \\) is generated by sampling from the occlusion maps \\( m^{t-1} \\) and \\( m^{t+1} \\), leveraging the pixel correspondences between consecutive frames under the reference view, as described by the following equation." + }, + { + "type": "equation", + "bbox": [ + 0.104, + 0.695, + 0.482, + 0.78 + ], + "angle": 0, + "content": "\\[\nm ^ {t} (i, j) = \\left\\{ \\begin{array}{l l} 1, & \\sum_ {k \\in \\{t - 1, t + 1 \\}} m ^ {k} (i + u, j + v) \\cdot \\\\ & C (i, j) \\geq 1 \\\\ m ^ {t} (i, j), & e l s e \\end{array} \\right. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.483, + 0.903 + ], + "angle": 0, + "content": "where \\( u \\) and \\( v \\) are the \\( x \\) and \\( y \\) components of the optical flow vector, and \\( C \\) represents the optical flow confidence measure. By propagating information from occluded areas across frames, this inter-frame continuity facilitates a more precise refinement of occluded regions in paired view videos and enforces temporal smoothness and coherence within the self-supervised framework, effectively mitigating the undesirable effects of jitter and flicker." + }, + { + "type": "image", + "bbox": [ + 0.532, + 0.089, + 0.891, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.259, + 0.908, + 0.303 + ], + "angle": 0, + "content": "Figure 3. DVG. The temporal motion can be utilized to refine the occluded regions in the current frame, thereby providing smoother images and more consistent occlusion in the temporal sequence." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.34, + 0.907, + 0.402 + ], + "angle": 0, + "content": "As shown in Figure 3, the difference between the training and inference procedures is that the former uses both backward and forward rendering whereas the latter uses only forward rendering to obtain an image under the target view." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.428, + 0.907, + 0.459 + ], + "angle": 0, + "content": "3.3. Self-Supervised Stereo Video Synthesis framework" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.472, + 0.906, + 0.579 + ], + "angle": 0, + "content": "Leveraging the generated paired view videos, we propose a self-supervised framework for stereo video synthesis, as illustrated in Figure 2. This framework commences with the rendering of the target view video that incorporates occlusions, followed by the application of a spatio-temporal method to effectively inpaint the missing regions, ultimately yielding a complete non-occluded target view video." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.584, + 0.907, + 0.903 + ], + "angle": 0, + "content": "To learn the distribution of the feature space between the paired view images, we devise RefinerNet, which is modeled after the architecture of the denoising U-Net, excluding the temporal layer. RefinerNet adopts the weight initialization from the original SD2.1 model [50]. As depicted in Figure 2, a spatial attention layer is employed instead of the self-attention layer. Given a feature map \\( z_{t} \\in \\mathbb{R}^{t \\times h \\times w \\times c} \\) from the denoising U-Net and another feature map \\( z_{r} \\in \\mathbb{R}^{t \\times h \\times w \\times c} \\) from RefinerNet, they are concatenated along the h dimension, and then self-attention is applied. Subsequently, the first half of the feature map is retrieved as the output. One great advantage of this approach is that it enables RefinerNet to maintain the feature representation capability inherent in the original SD Unet, extracting reference view image information. In addition, the denoising U-Net can adaptively learn features from RefinerNet that are correlated in the same feature space, owing to the fundamentally similar network architectures between the denoising U-Net and RefinerNet. This not only ensures a robust initialization for the feature representation but also enhances the learning ability of the entire network." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.513, + 0.957 + ], + "angle": 0, + "content": "814" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.348, + 0.108 + ], + "angle": 0, + "content": "3.4. Consistency Control Module" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.114, + 0.419, + 0.13 + ], + "angle": 0, + "content": "3.4.1 Temporal Interaction Learning module" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.137, + 0.483, + 0.212 + ], + "angle": 0, + "content": "While our RefinerNet-based approach providing spatial guidance, paired view features are still merged independently, posing challenges for temporal consistency and potentially causing texture discrepancies or odd results between adjacent frames." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.213, + 0.484, + 0.32 + ], + "angle": 0, + "content": "Inspired by previous studies [11, 22], we propose an attention-based temporal interaction learning module that learns the distribution of the feature space from adjacent frames, named TIL. Specifically, given latent features of \\( N_r \\) adjacent images under reference view \\( z_i^t \\), where \\( i = 1, 2, \\dots, N_r \\) and the latent feature of the reference view image \\( z_r^t \\) at the time step \\( t \\), we augment the \\( z_r^t \\) to \\( aug_r^t \\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.129, + 0.331, + 0.483, + 0.372 + ], + "angle": 0, + "content": "\\[\na u g _ {r} ^ {t} = \\lambda \\cdot \\operatorname {A t t n} _ {r, r} + (1 - \\lambda) \\cdot \\frac {1}{N _ {r}} \\sum_ {i = 1} ^ {N _ {r}} \\operatorname {A t t n} _ {r, i} \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.382, + 0.484, + 0.519 + ], + "angle": 0, + "content": "where \\(\\lambda \\in [0,1]\\) and \\(\\mathrm{Attn}_{i,j}\\) defines attention between two latent features \\(z_{i}\\) and \\(z_{j}\\). This module blends self-attention of \\(z_{r}^{t}\\) in RefinerNet's U-Net block with the cross-view attention between \\(z_{r}^{t}\\) and each adjacent view \\(z_{i}^{t}\\). The cross-view attention aligns the feature of all adjacent frames to the reference view while the self-attention helps each reference image retain distinctiveness. The augmented reference feature \\(aug_{r}^{t}\\) is then fed into spatial attention layer to assist U-Net network learning." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.537, + 0.325, + 0.552 + ], + "angle": 0, + "content": "3.4.2 Stereo Deviation Strength" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.56, + 0.484, + 0.726 + ], + "angle": 0, + "content": "During the training, we noticed that different viewpoints impact the 3D impression of the scenes in the generated videos. However, relying solely on a given viewpoint as a guide for generating stereoscopic videos is insufficient. This is because videos with the same viewpoint may exhibit diverse depth ranges and variations, depending on the scene [70]. Therefore, pose alone cannot accurately control the 3D effect of a scene. Consequently, we introduce a metric called the stereo deviation strength, which quantitatively assesses the binocular disparity in a scene and facilitates the creation of controllable stereo vision:" + }, + { + "type": "equation", + "bbox": [ + 0.223, + 0.738, + 0.483, + 0.755 + ], + "angle": 0, + "content": "\\[\ns (z) = \\left| z _ {0} - z _ {\\text {r e f}} \\right| \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.871 + ], + "angle": 0, + "content": "where the stereo deviation strength quantifies the latent differences between reference view \\( z_{ref} \\) and target view \\( z_0 \\). Similar to the time step, the deviation strength is projected into a positional embedding and added to each frame in the residual block to ensure uniform application of the deviation strength to every frame. The impact of stereo deviation strength guidance on the results is illustrated in Figure 4." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.872, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Moreover, for better convergence, a stereo-aware loss function is proposed to directly supervise the disparity dif" + }, + { + "type": "image", + "bbox": [ + 0.536, + 0.089, + 0.887, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.259, + 0.907, + 0.316 + ], + "angle": 0, + "content": "Figure 4. Stereo deviation strength guidance examples. Augmenting the deviation strength enhances the 3D photography effect, but excessive deviation strength may lead to image distortions or even lower the quality of stereoscopic image." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.342, + 0.573, + 0.355 + ], + "angle": 0, + "content": "ference:" + }, + { + "type": "equation", + "bbox": [ + 0.631, + 0.355, + 0.905, + 0.373 + ], + "angle": 0, + "content": "\\[\nl _ {d} = \\left| \\left| s \\left(z _ {0}\\right) - s \\left(\\hat {z} _ {0}\\right) \\right| \\right| _ {2} ^ {2} \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.379, + 0.907, + 0.409 + ], + "angle": 0, + "content": "where \\(\\hat{z}_0\\) represents the estimated clean video latent \\(z_0\\), which can be obtained by:" + }, + { + "type": "equation", + "bbox": [ + 0.607, + 0.42, + 0.905, + 0.455 + ], + "angle": 0, + "content": "\\[\n\\hat {z} _ {0} = z _ {0} - \\frac {\\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon_ {\\theta} (z _ {t} , c , t)}{\\sqrt {\\bar {\\alpha} _ {t}}} \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.465, + 0.906, + 0.497 + ], + "angle": 0, + "content": "where \\(\\overline{\\alpha}_t = \\prod_{i=1}^t (1 - \\beta_t)\\), and the noise prediction loss is combined with the stereo-aware loss by a scaling factor:" + }, + { + "type": "equation", + "bbox": [ + 0.661, + 0.508, + 0.905, + 0.523 + ], + "angle": 0, + "content": "\\[\nl = l _ {\\epsilon} + \\lambda \\cdot l _ {d} \\tag {7}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.544, + 0.687, + 0.561 + ], + "angle": 0, + "content": "3.5. Training Strategy" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.568, + 0.907, + 0.81 + ], + "angle": 0, + "content": "The training process is divided into two stages. In the first training stage, we focus on individual frames from videos. The temporal layer in denoising U-Net is frozen, and cross-view attention in TIL module is removed. The Refiner-Net model and the denoising U-Net are initialized with pretrained weights from SD2.1 [50] and SVD [7] respectively. The weights of the variational autoencoder encoder and decoder, as well as the contrastive language-image pretraining (CLIP) image encoder, are all kept fixed. The aim of this stage is for the model to learn to synthesize new viewpoint images under the condition of a reference image and a new viewpoint pose. In the second stage, we train the temporal layer and cross-view attention in TIL module with video sequences. This enables the model to capture temporal context information efficiently. The input for the model is the 8-frame video clip and \\( N_r \\) for TIL is set to 8." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.823, + 0.646, + 0.84 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.848, + 0.792, + 0.864 + ], + "angle": 0, + "content": "4.1. Datasets, Baselines and Metrics" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.906, + 0.901 + ], + "angle": 0, + "content": "We conduct our experiments on image and video levels with RealEstate10K [78] and self-collected video data." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "815" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.094, + 0.092, + 0.482, + 0.362 + ], + "angle": 0, + "content": "RealEstate10K is a large, widely used dataset with camera poses, corresponding to 10 million frames derived from about 80,000 video clips, gathered from about 10,000 YouTube videos. As there is not open-source stereo video benchmark, we collect 1500 monocular videos with 1400 for training, 100 for testing. For image synthesis, we compared the proposed method with recent open-source SOTA methods: 3D-photography [56], SynSim [69], AdaMPI [21], SinMPI [47], Wang et al. [66], Photoconsistent-NVS(P-NVS) [75], NVSVDE-Net [4], CoPoNeRF [24] and MVsplat [14]. For a fair comparison, the 3D photography method, AdaMPI, SinMPI, and the proposed method all used MiDaS [49] for depth estimation. Following the method of 3D-photography [56], we measured the accuracy of the generated target views with regard to the ground-truth images using the three metrics of the learned perceptual image patch similarity (LPIPS), peak signal-to-noise ratio (PSNR), and structural similarity index measure (SSIM)." + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.365, + 0.482, + 0.529 + ], + "angle": 0, + "content": "For the video synthesis, we picked several representative approaches, i.e., the method of 3D-photography [56], the open-source code repository webui-depthmap, P-NVS [75], CoPoNeRF [24], NVS-Solver [74] and AVP [44]. Moreover, we conducted the same depth estimation method [49] for a fair comparison of 3D-photography [56], webui-depthmap and NVS-Solver [74]. We employ the FVD [62] score to measure the perceptual similarity between input videos and outputs and report the flow warping error \\( E_{\\text{warp}} \\) [35] to assess the temporal consistency of the resulting video sequences." + }, + { + "type": "title", + "bbox": [ + 0.094, + 0.541, + 0.249, + 0.556 + ], + "angle": 0, + "content": "4.2. Implementation" + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.564, + 0.482, + 0.744 + ], + "angle": 0, + "content": "We employed the AdamW [39] optimizer with a learning rate of \\(1 \\times 10^{-5}\\) to train the model on two NVIDIA A800 GPUs. In the first training stage, individual video frames were resized and center-cropped to a resolution of \\(1024 \\times 1024\\). We trained the model for 100,000 iterations with a batch size of 8. In the second training stage, the videos were first split according to the transitions to ensure that a scene only appeared in one video clip. We then formed each video clip for training with eight frames. Finally, the temporal layer and TIL module was trained for 10,000 steps with a batch size of 2. The stereo-aware loss scaling factor \\(\\lambda\\) was set to 0.001 and \\(\\lambda\\) in Eq. 3 was set as 0.6." + }, + { + "type": "title", + "bbox": [ + 0.094, + 0.756, + 0.272, + 0.77 + ], + "angle": 0, + "content": "4.3. Qualitative Results" + }, + { + "type": "title", + "bbox": [ + 0.094, + 0.78, + 0.305, + 0.793 + ], + "angle": 0, + "content": "4.3.1 Stereo Image Synthesis" + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.804, + 0.482, + 0.877 + ], + "angle": 0, + "content": "We visually compared different methods for generating new view images from a target viewpoint. We qualitatively compared the newly synthesized images in Figure 5. 3D-photography fails to resolve the image distortion problem, and the image appears to be enhanced only slightly by the" + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.242 + ], + "angle": 0, + "content": "rendering. SynSin generates the most blurry details and loses a lot of information, compared with the input view. Neither AdaMPI nor SinMPI can generate fine enough details, and the results contain noticeable artifacts. CoPoN-erF tends to produce discrepancies from the original image in terms of fine-grained details, and it struggles with effectively handling scenes involving motion, as well as MVsplat. In contrast, the proposed method shows a clean and detailed output, especially in the area of depth information discontinuity." + }, + { + "type": "title", + "bbox": [ + 0.517, + 0.282, + 0.724, + 0.295 + ], + "angle": 0, + "content": "4.3.2 Stereo Video Synthesis" + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.311, + 0.905, + 0.581 + ], + "angle": 0, + "content": "Moreover, we qualitatively compared the results of the stereo video synthesis with other methods. In addition to the quality issues of single-frame generation, the most important aspect in videos is the temporal consistency. We seek to generate videos that are free of flicker. The visualization of generated videos from different methods are illustrated in Figure 6, for each video, we show the scanline (the yellow line in figure) slice through the spatial-temporal volume. Our method achieve the smoothest trajectories, while other methods exhibit significant flickering. The most distinct regions are zoomed in and highlighted under the images to illustrate the details. In the first case, our method attains the smooth trajectories of the background (maintaining the straight line). And the second case, only our method preserves the texture of the trees behind the woman. And the last case, only our method preserves the texture of the black vertical lines. More results of generated frames are also presented in supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.517, + 0.604, + 0.706, + 0.618 + ], + "angle": 0, + "content": "4.4. Quantitative Results" + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.629, + 0.905, + 0.9 + ], + "angle": 0, + "content": "Table 1 lists the SSIM, PSNR, and LPIPS scores for the newly synthesized images. The proposed method outperforms 3D-photography and SynSin in the major metrics by a wide margin. The method of Wang et al., P-NVS and NVSVDE-Net perform slightly worse in terms of all metrics compared to the proposed method. Since the proposed method merges the spatial features and fills the occluded regions, it achieves the best LPIPS scores and shows competitive results in PSNR and SSIM when compared with AdaMPI and CoPoNeRF. MVSplat outperforms other methods when \\( t = 10 \\) but is inferior to ours when \\( t = 5 \\). Overall, the results exhibit superior visual effects, which is consistent with the performance observed in the qualitative experiments. We further demonstrate the effectiveness of our method on video synthesis, as shown in Table 2. The proposed method achieves the best performance on both FVD and \\( E_{\\text{warp}} \\), which means that our generated videos attain the best quality and temporal consistency." + }, + { + "type": "page_footnote", + "bbox": [ + 0.117, + 0.888, + 0.466, + 0.9 + ], + "angle": 0, + "content": "https://github.com/thygate/stable-diffusion-webui-depthmap-script" + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.947, + 0.512, + 0.957 + ], + "angle": 0, + "content": "816" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.11, + 0.092, + 0.885, + 0.31 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.313, + 0.908, + 0.342 + ], + "angle": 0, + "content": "Figure 5. Qualitative results on the RealEstate10K dataset. The proposed method generates better-quality information and maintains the edge and textural details of the reference image better than the other methods. More results are available in supplementary material." + }, + { + "type": "image", + "bbox": [ + 0.1, + 0.356, + 0.874, + 0.54 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.545, + 0.908, + 0.574 + ], + "angle": 0, + "content": "Figure 6. Visual comparisons in stereo video benchmark. We show the scanline (yellow line in original frame) slice through the spatial-temporal volume. The most distinct regions are zoomed in and highlighted under the images to illustrate the details." + }, + { + "type": "image", + "bbox": [ + 0.108, + 0.6, + 0.474, + 0.748 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.758, + 0.483, + 0.813 + ], + "angle": 0, + "content": "Figure 7. Ablation study of different designs. Only the proposed method ensures consistent preservation of details with the groundtruth target view. More results are available in supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.835, + 0.245, + 0.851 + ], + "angle": 0, + "content": "4.5. Ablation Study" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.858, + 0.377, + 0.873 + ], + "angle": 0, + "content": "4.5.1 Effectiveness of each components" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.882, + 0.484, + 0.912 + ], + "angle": 0, + "content": "We show the results in Table 3 on the RealEstate10K (stereo image) and collected stereo video datasets. CN means sub" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.6, + 0.907, + 0.872 + ], + "angle": 0, + "content": "stituting our RefinerNet (RN) to ControlNet, SDS means the proposed stereo deviation strength. For image-level comparison, the results show that utilizing U-Net architecture only is not sufficient for stereo image synthesis and our RefinerNet is a better choice than ControlNet. Adding SDS further improves the geometric consistency. As visualized in Figure 7, solely relying on U-Net features results in distortion in the edge area and struggles with generating content properly. It also fails to preserve textural details near a plane boundary. Adding ControlNet often results in content inconsistency, presumably due to the failure of the paired view feature integration. RefinerNet handles this problem, showing it is a better architecture for stereo image generation. SDS further improves the consistency with the input and contain a more detailed image structure. As for video-level comparison, the quantitative results are shown in Table 3, adding TIL and DVG both attain the performance gain, and our final model achieves the best performance." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "817" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.486, + 0.16 + ], + "angle": 0, + "content": "Table 1. Quantitative comparison on RealEstate10K dataset. \\(\\dagger\\) means re-implementing in our setting for fair comparison. When \\(t = 5\\), the current frame serves as the left view and the frame at time +5 serves as the right view; the same approach applies for \\(t = 10\\). The best results are in bold." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.17, + 0.499, + 0.375 + ], + "angle": 0, + "content": "
SSIM ↑PSNR ↑LPIPS ↓
t=5t=10t=5t=10t=5t=10
3D-photography [56]†0.6410.44611.089.5330.1160.386
SynSin [69]†0.7400.64521.2219.550.0870.104
AdaMPI [21]†0.8690.73627.7022.650.0410.087
SinMPI [47]†0.8980.77528.7923.210.0400.080
Wang et al. [66]0.8400.72025.3521.360.0490.095
P-NVS [75]†0.7240.66721.9819.460.2410.279
NVSVDE-Net [4]0.84024.310.233
CoPoNeRF [24]†0.6580.63222.6321.330.1710.209
MVSplat [14]†0.86325.890.132
Proposed0.9160.85732.2624.860.0380.049
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.384, + 0.484, + 0.414 + ], + "angle": 0, + "content": "Table 2. Quantitative comparison with the other methods. The best results are in **bold**. \\( E_{\\text{warp}}^* \\) denotes \\( E_{\\text{warp}} \\left( \\times 10^{-3} \\right) \\)." + }, + { + "type": "table", + "bbox": [ + 0.123, + 0.422, + 0.451, + 0.577 + ], + "angle": 0, + "content": "
FVD ↓E* warp ↓
3D-photography [56]155.03.418
Webui-depthmap83.183.486
P-NVS [75]420.223.02
CoPoNeRF [24]290.07.144
NVS-Solver [74]249.15.842
MVSplat [14]15437.913
AVP [44]99.923.446
Proposed67.093.374
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.591, + 0.484, + 0.62 + ], + "angle": 0, + "content": "Table 3. Quantitative comparison for the ablation study. The best results are in bold." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.63, + 0.481, + 0.758 + ], + "angle": 0, + "content": "
MethodCNRNSDSTILDVGStereo ImageStereo Video
SSIM↑PSNR↑LPIPS↓FVD↓
Image-level✓0.88023.730.06-
0.85524.040.183-
✓0.89530.200.043-
✓✓0.91632.260.038-
Video-level✓✓---184.0
✓✓✓----123.5
✓✓✓---85.21
✓✓✓✓---67.09
" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.785, + 0.337, + 0.801 + ], + "angle": 0, + "content": "4.5.2 Data improvement by DVG" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.484, + 0.903 + ], + "angle": 0, + "content": "We show the generated stereo occlusion mask with interframe motion refinement in Figure 8. It can be observed that the occlusions appear more complete spatially and exhibit smoother in their temporal distribution. By employing DVG, we can acquire sufficient data to support the generation of spatio-temporally consistent videos. More analyses" + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.089, + 0.648, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.648, + 0.089, + 0.772, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.772, + 0.089, + 0.899, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.197, + 0.907, + 0.253 + ], + "angle": 0, + "content": "Figure 8. Visual comparisons for the improvement of stereo occlusion. We show the scanline of (yellow line in original frame) slice through the spatial-temporal volume. The red region represents the changes of occlusion mask." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.281, + 0.81, + 0.296 + ], + "angle": 0, + "content": "are available in the supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.317, + 0.907, + 0.348 + ], + "angle": 0, + "content": "4.5.3 Exploration of different depth estimation methods" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.357, + 0.907, + 0.493 + ], + "angle": 0, + "content": "We further compared different design choices of depth map including DepthAnything [73], Marigold [31], MiDaS [49], ZoeDepth [5] and DepthCrafter [26]. The results are available in supplementary materials. The proposed method can produce new viewpoint images that are realistic, seamless, and rich in detail, regardless of the depth estimation method used. This capability demonstrates the versatility and robustness of the proposed method, ensuring superior performance in various applications and environments." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.504, + 0.64, + 0.519 + ], + "angle": 0, + "content": "4.6. Limitations" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.528, + 0.906, + 0.62 + ], + "angle": 0, + "content": "SpatialDreamer effectively creates stereo video from a single input but faces limitations. It relies on depth maps for rendering, affecting stereoscopic quality due to depth accuracy and distribution. Future research could focus on integrating implicit depth to address this [36]. Additionally, the model's large parameter size limits real-time performance." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.634, + 0.634, + 0.649 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.66, + 0.907, + 0.902 + ], + "angle": 0, + "content": "In this paper, we introduce a new self-supervised stereo video synthesis approach using a video diffusion model, termed SpatialDreamer. This method addresses the issues of insufficient data and spatio-temporal inconsistency between frames. To solve the problem of insufficient data, we develop a Depth-based Video Generation (DVG) module that uses a forward-backward rendering mechanism to generate a rendered video with geometric and temporal priors. Additionally, we propose RefinerNet along with a self-supervised synthetic framework to enable efficient and dedicated training using data generated by DVG. Furthermore, we design a consistency control module to ensure geometric and temporal consistency. Our SpatialDreamer outperforms all other open-source stereo image and video synthesis methods and has the potential for future expansion into virtual reality applications." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.513, + 0.957 + ], + "angle": 0, + "content": "818" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.091, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.116, + 0.484, + 0.171 + ], + "angle": 0, + "content": "[1] Dmitry Akimov, Alexey Shestov, Alexander Voronov, and Dmitriy Vatolin. Occlusion refinement for stereo video using optical flow. In 2012 International Conference on 3D Imaging (IC3D), pages 1-8, 2012." + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.173, + 0.484, + 0.243 + ], + "angle": 0, + "content": "[2] Andrea Schubert. visionos 2 brings new spatial computing experiences to apple vision pro. https://www.apple.com/newsroom/2024/06/visionos-2-brings-new-spatial-computing-experiences-to-apple-vision-pro, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.245, + 0.483, + 0.3 + ], + "angle": 0, + "content": "[3] Omer Bar-Tal, Dolev Ofri-Amar, Rafail Fridman, Yoni Kasten, and Tali Dekel. Text2live: Text-driven layered image and video editing. In European conference on computer vision, pages 707-723. Springer, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.302, + 0.483, + 0.357 + ], + "angle": 0, + "content": "[4] Juan Luis Gonzalez Bello and Munchurl Kim. Novel view synthesis with view-dependent effects from a single image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.359, + 0.482, + 0.401 + ], + "angle": 0, + "content": "[5] Shariq Farooq Bhat, Reiner Birkl, Diana Wofk, Peter Wonka, and Matthias Müller. Zoedepth: Zero-shot transfer by combining relative and metric depth, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.403, + 0.483, + 0.471 + ], + "angle": 0, + "content": "[6] Phuong Ngoc Binh Do and Quoc Chi Nguyen. A review of stereo-photogrammetry method for 3-d reconstruction in computer vision. In 2019 19th International Symposium on Communications and Information Technologies (ISCIT), pages 138–143, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.474, + 0.483, + 0.543 + ], + "angle": 0, + "content": "[7] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.545, + 0.483, + 0.627 + ], + "angle": 0, + "content": "[8] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22563-22575, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.63, + 0.483, + 0.67 + ], + "angle": 0, + "content": "[9] Aaron Bobick and Stephen Intille. Large occlusion stereo. International Journal of Computer Vision, 33:181-200, 1999." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.673, + 0.483, + 0.741 + ], + "angle": 0, + "content": "[10] Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Yufei Guo, Li Jing, David Schnurr, Joe Taylor, Troy Luhman, Eric Luhman, Clarence Ng, Ricky Wang, and Aditya Ramesh. Video generation models as world simulators. 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.745, + 0.483, + 0.8 + ], + "angle": 0, + "content": "[11] Mingdeng Cao, Xintao Wang, Zhongang Qi, Ying Shan, Xiaohu Qie, and Yinqiang Zheng. Masactrl: Tuning-free mutual self-attention control for consistent image synthesis and editing, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.802, + 0.483, + 0.843 + ], + "angle": 0, + "content": "[12] David Charatan, Sizhe Li, Andrea Tagliasacchi, and Vincent Sitzmann. pixelsplat: 3d gaussian splats from image pairs for scalable generalizable 3d reconstruction. In arXiv, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[13] Weifeng Chen, Jie Wu, Pan Xie, Hefeng Wu, Jiashi Li, Xin Xia, Xuefeng Xiao, and Liang Lin. Control-a-video: Controllable text-to-video generation with diffusion models. arXiv preprint arXiv:2305.13840, 2023." + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.116, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.148 + ], + "angle": 0, + "content": "[14] Yuedong Chen, Haofei Xu, Chuanxia Zheng, Bohan Zhuang, Marc Pollefeys, Andreas Geiger, Tat-Jen Cham, and Jianfei Cai. Mvsplat: Efficient 3d gaussian splatting from sparse multi-view images. arXiv preprint arXiv:2403.14627, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.151, + 0.905, + 0.205 + ], + "angle": 0, + "content": "[15] Zuozhuo Dai, Zhenghao Zhang, Yao Yao, Bingxue Qiu, Siyu Zhu, Long Qin, and Weizhi Wang. Animateanything: Fine-grained open domain image animation with motion guidance. arXiv e-prints, pages arXiv-2311, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.208, + 0.905, + 0.261 + ], + "angle": 0, + "content": "[16] G. Egnal and R.P. Wildes. Detecting binocular half-occlusions: empirical comparisons of five approaches. IEEE Transactions on Pattern Analysis and Machine Intelligence, 24(8):1127-1133, 2002." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.264, + 0.905, + 0.332 + ], + "angle": 0, + "content": "[17] Patrick Esser, Johnathan Chiu, Parmida Atighehchian, Jonathan Granskog, and Anastasis Germanidis. Structure and content-guided video synthesis with diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7346-7356, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.334, + 0.904, + 0.374 + ], + "angle": 0, + "content": "[18] Chen Gao, Ayush Saraf, Johannes Kopf, and Jia-Bin Huang. Dynamic view synthesis from dynamic monocular video, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.378, + 0.905, + 0.446 + ], + "angle": 0, + "content": "[19] Yuwei Guo, Ceyuan Yang, Anyi Rao, Zhengyang Liang, Yaohui Wang, Yu Qiao, Maneesh Agrawala, Dahua Lin, and Bo Dai. Animatediff: Imagine your personalized text-to-image diffusion models without specific tuning. International Conference on Learning Representations, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.448, + 0.905, + 0.503 + ], + "angle": 0, + "content": "[20] Agrim Gupta, Lijun Yu, Kihyuk Sohn, Xiuye Gu, Meera Hahn, Li Fei-Fei, Irfan Essa, Lu Jiang, and José Lezama. Photorealistic video generation with diffusion models. arXiv preprint arXiv:2312.06662, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.505, + 0.905, + 0.56 + ], + "angle": 0, + "content": "[21] Yuxuan Han, Ruicheng Wang, and Jiaolong Yang. Single-view view synthesis in the wild with learned adaptive multiplane images. In ACM SIGGRAPH 2022 Conference Proceedings, pages 1-8, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.562, + 0.905, + 0.602 + ], + "angle": 0, + "content": "[22] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.604, + 0.905, + 0.659 + ], + "angle": 0, + "content": "[23] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.661, + 0.905, + 0.716 + ], + "angle": 0, + "content": "[24] Sunghwan Hong, Jaewoo Jung, Heeseong Shin, Jiaolong Yang, Seungryong Kim, and Chong Luo. Unifying correspondence, pose and nerf for pose-free novel view synthesis from stereo pairs, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.718, + 0.905, + 0.773 + ], + "angle": 0, + "content": "[25] Li Hu, Xin Gao, Peng Zhang, Ke Sun, Bang Zhang, and Liefeng Bo. animate anyone: Consistent and controllable image-to-video synthesis for character animation. arXiv preprint arXiv:2311.17117, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.776, + 0.905, + 0.83 + ], + "angle": 0, + "content": "[26] Wenbo Hu, Xiangjun Gao, Xiaoyu Li, Sijie Zhao, Xiaodong Cun, Yong Zhang, Long Quan, and Ying Shan. Depthcrafter: Generating consistent long depth sequences for open-world videos. arXiv preprint arXiv:2409.02095, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.832, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[27] Varun Jampani, Huiwen Chang, Kyle Sargent, Abhishek Kar, Richard Tucker, Michael Krainin, Dominik Kaeser, William T Freeman, David Salesin, Brian Curless, et al. Slide: Single image 3d photography with soft layering and depth-aware inpainting. In Proceedings of the IEEE/CVF In" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "819" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.126, + 0.092, + 0.483, + 0.12 + ], + "angle": 0, + "content": "ternational Conference on Computer Vision, pages 12518-12527, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.122, + 0.483, + 0.149 + ], + "angle": 0, + "content": "[28] Wonbong Jang and Lourdes Agapito. Nvist: In the wild new view synthesis from a single image with transformers, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.151, + 0.483, + 0.207 + ], + "angle": 0, + "content": "[29] Jisoo Jeong, Hong Cai, Risheek Garrepalli, Jamie Menjay Lin, Munawar Hayat, and Fatih Porikli. Ocai: Improving optical flow estimation by occlusion and consistency aware interpolation, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.208, + 0.483, + 0.262 + ], + "angle": 0, + "content": "[30] Hailin Jin, Stefano Soatto, and Anthony J Yezzi. Multi-view stereo reconstruction of dense shape and complex appearance. International Journal of Computer Vision, 63:175-189, 2005." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.264, + 0.483, + 0.319 + ], + "angle": 0, + "content": "[31] Bingxin Ke, Anton Obukhov, Shengyu Huang, Nando Metzger, Rodrigo Caye Daudt, and Konrad Schindler. Repurposing diffusion-based image generators for monocular depth estimation, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.321, + 0.483, + 0.363 + ], + "angle": 0, + "content": "[32] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.365, + 0.483, + 0.433 + ], + "angle": 0, + "content": "[33] Levon Khachatriyan, Andranik Movsisyan, Vahram Tadevosyan, Roberto Henschel, Zhangyang Wang, Shant Navasardyan, and Humphrey Shi. Text2video-zero: Text-to-image diffusion models are zero-shot video generators. arXiv preprint arXiv:2303.13439, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.435, + 0.483, + 0.504 + ], + "angle": 0, + "content": "[34] Dan Kondratyuk, Lijun Yu, Xiuye Gu, José Lezama, Jonathan Huang, Rachel Hornung, Hartwig Adam, Hassan Akbari, Yair Alon, Vighnesh Birodkar, et al. Videopoet: A large language model for zero-shot video generation. arXiv preprint arXiv:2312.14125, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.505, + 0.483, + 0.56 + ], + "angle": 0, + "content": "[35] Wei-Sheng Lai, Jia-Bin Huang, Oliver Wang, Eli Shechtman, Ersin Yumer, and Ming-Hsuan Yang. Learning blind video temporal consistency. In Proceedings of the European conference on computer vision (ECCV), pages 170-185, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.562, + 0.483, + 0.63 + ], + "angle": 0, + "content": "[36] Hanwen Liang, Junli Cao, Vedit Goel, Guocheng Qian, Sergei Korolev, Demetri Terzopoulos, Konstantinos Plataniotis, Sergey Tulyakov, and Jian Ren. Wonderland: Navigating 3d scenes from a single image. arXiv preprint arXiv:2412.12091, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.632, + 0.483, + 0.702 + ], + "angle": 0, + "content": "[37] Andrew Liu, Richard Tucker, Varun Jampani, Ameesh Makadia, Noah Snavely, and Angjoo Kanazawa. Infinite nature: Perpetual view generation of natural scenes from a single image. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.703, + 0.483, + 0.771 + ], + "angle": 0, + "content": "[38] Yixin Liu, Kai Zhang, Yuan Li, Zhiling Yan, Chujie Gao, Ruoxi Chen, Zhengqing Yuan, Yue Huang, Hanchi Sun, Jianfeng Gao, Lifang He, and Lichao Sun. Sora: A review on background, technology, limitations, and opportunities of large vision models, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.773, + 0.483, + 0.801 + ], + "angle": 0, + "content": "[39] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.803, + 0.483, + 0.845 + ], + "angle": 0, + "content": "[40] Kevin Matzen, Michael F. Cohen, Bryce Evans, Johannes Kopf, and Richard Szeliski. Low-cost 360 stereo photography and video capture. ACM Trans. Graph., 36(4), 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.483, + 0.9 + ], + "angle": 0, + "content": "[41] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis, 2020." + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.906, + 0.175 + ], + "angle": 0, + "content": "[42] Norman Müller, Katja Schwarz, Barbara Rössle, Lorenzo Porzi, Samuel Rota Bulò, Matthias Nießner, and Peter Kontschieder. Multidiff: Consistent novel view synthesis from a single image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10258-10268, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.178, + 0.905, + 0.204 + ], + "angle": 0, + "content": "[43] Simon Niklaus and Feng Liu. Softmax splatting for video frame interpolation, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.207, + 0.905, + 0.262 + ], + "angle": 0, + "content": "[44] Onee. What is spatial video on iphone 15 pro and vision pro. https://xreality-zone/en/posts/what-is-spatial-video-on-iphone-15-pro-and/apple-vision-pro, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.264, + 0.905, + 0.331 + ], + "angle": 0, + "content": "[45] Hao Ouyang, Qiuyu Wang, Yuxi Xiao, Qingyan Bai, Juntao Zhang, Kecheng Zheng, Xiaowei Zhou, Qifeng Chen, and Yujun Shen. Codef: Content deformation fields for temporally consistent video processing. arXiv preprint arXiv:2308.07926, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.333, + 0.905, + 0.374 + ], + "angle": 0, + "content": "[46] Onur Özyeşil, Vladislav Voroninski, Ronen Basri, and Amit Singer. A survey of structure from motion*. Acta Numerica, 26:305-364, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.377, + 0.905, + 0.418 + ], + "angle": 0, + "content": "[47] Guo Pu, Peng-Shuai Wang, and Zhouhui Lian. Sinmpi: Novel view synthesis from a single image with expanded multiplane images, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.42, + 0.905, + 0.474 + ], + "angle": 0, + "content": "[48] Chenyang Qi, Xiaodong Cun, Yong Zhang, Chenyang Lei, Xintao Wang, Ying Shan, and Qifeng Chen. Fatezero: Fusing attentions for zero-shot text-based video editing. arXiv preprint arXiv:2303.09535, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.476, + 0.905, + 0.545 + ], + "angle": 0, + "content": "[49] René Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(3), 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.547, + 0.907, + 0.616 + ], + "angle": 0, + "content": "[50] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.617, + 0.905, + 0.659 + ], + "angle": 0, + "content": "[51] Christoph Sager, Christian Janiesch, and Patrick Zschech. A survey of image labelling for computer vision applications. Journal of Business Analytics, 4(2):91-110, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.66, + 0.905, + 0.688 + ], + "angle": 0, + "content": "[52] D. Scharstein. View Synthesis Using Stereo Vision. Springer Berlin Heidelberg, 1999." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.689, + 0.905, + 0.744 + ], + "angle": 0, + "content": "[53] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4104-4113, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.746, + 0.905, + 0.815 + ], + "angle": 0, + "content": "[54] S.M. Seitz, B. Curless, J. Diebel, D. Scharstein, and R. Szeliski. A comparison and evaluation of multi-view stereo reconstruction algorithms. In 2006 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR'06), pages 519-528, 2006." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.817, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[55] Yujiao Shi, Hongdong Li, and Xin Yu. Self-supervised visibility learning for novel view synthesis.*. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.902 + ], + "angle": 0, + "content": "[56] Meng-Li Shih, Shih-Yang Su, Johannes Kopf, and Jia-Bin Huang. 3d photography using context-aware layered depth" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.907, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "820" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.484, + 0.133 + ], + "angle": 0, + "content": "inpainting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8028-8038, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.134, + 0.484, + 0.203 + ], + "angle": 0, + "content": "[57] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.204, + 0.483, + 0.232 + ], + "angle": 0, + "content": "[58] Lorenzo Stacchio. Train stable diffusion for inpainting, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.233, + 0.483, + 0.288 + ], + "angle": 0, + "content": "[59] Wenxiu Sun, Lingfeng Xu, Oscar C Au, Sung Him Chui, and Chun Wing Kwok. An overview of free view-point depth-image-based rendering (dibr). In APSIPA Annual Summit and Conference, pages 1023-1030, 2010." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.289, + 0.483, + 0.316 + ], + "angle": 0, + "content": "[60] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.317, + 0.483, + 0.372 + ], + "angle": 0, + "content": "[61] Richard Tucker and Noah Snavely. Single-view view synthesis with multiplane images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 551-560, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.372, + 0.483, + 0.455 + ], + "angle": 0, + "content": "[62] Thomas Unterthiner, Sjoerd van Steenkiste, Karol Kurach, Raphaël Marinier, Marcin Michalski, and Sylvain Gelly. FVD: A new metric for video generation. In Deep Generative Models for Highly Structured Data, ICLR 2019 Workshop, New Orleans, Louisiana, United States, May 6, 2019. OpenReview.net, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.456, + 0.483, + 0.525 + ], + "angle": 0, + "content": "[63] Basile Van Hoorick, Rundi Wu, Ege Ozguroglu, Kyle Sargent, Ruoshi Liu, Pavel Tokmakov, Achal Dave, Changxi Zheng, and Carl Vondrick. Generative camera dolly: Extreme monocular dynamic novel view synthesis. European Conference on Computer Vision (ECCV), 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.526, + 0.483, + 0.58 + ], + "angle": 0, + "content": "[64] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.581, + 0.483, + 0.636 + ], + "angle": 0, + "content": "[65] Wen Wang, Yan Jiang, Kangyang Xie, Zide Liu, Hao Chen, Yue Cao, Xinlong Wang, and Chunhua Shen. Zero-shot video editing using off-the-shelf image diffusion models. arXiv preprint arXiv:2303.17599, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.636, + 0.483, + 0.705 + ], + "angle": 0, + "content": "[66] Xiaodong Wang, Chenfei Wu, Shengming Yin, Minheng Ni, Jianfeng Wang, Linjie Li, Zhengyuan Yang, Fan Yang, Lijuan Wang, Zicheng Liu, et al. Learning 3d photography videos via self-supervised diffusion on single images. arXiv preprint arXiv:2302.10781, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.706, + 0.483, + 0.775 + ], + "angle": 0, + "content": "[67] Xiang Wang, Hangjie Yuan, Shiwei Zhang, Dayou Chen, Jiuniu Wang, Yingya Zhang, Yujun Shen, Deli Zhao, and Jingren Zhou. Videocomposer: Compositional video synthesis with motion controllability. Advances in Neural Information Processing Systems, 36, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.776, + 0.483, + 0.844 + ], + "angle": 0, + "content": "[68] Chung-Yi Weng, Brian Curless, and Ira Kemelmacher-Shlizerman. Photo wake-up: 3d character animation from a single photo. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5908-5917, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.845, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[69] Olivia Wiles, Georgia Gkioxari, Richard Szeliski, and Justin Johnson. Synsin: End-to-end view synthesis from a single image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020." + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.133 + ], + "angle": 0, + "content": "[70] Andrew Woods, Tom Docherty, and Rolf Koch. Image distortions in stereoscopic video systems. Proc SPIE, 1915, 2002." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.134, + 0.905, + 0.219 + ], + "angle": 0, + "content": "[71] Jay Zhangjie Wu, Yixiao Ge, Xintao Wang, Stan Weixian Lei, Yuchao Gu, Yufei Shi, Wynne Hsu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7623-7633, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.22, + 0.905, + 0.261 + ], + "angle": 0, + "content": "[72] Dejia Xu, Yifan Jiang, Peihao Wang, Zhiwen Fan, Humphrey Shi, and Zhangyang Wang. Sinnerf: Training neural radiance fields on complex scenes from a single image, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.262, + 0.905, + 0.304 + ], + "angle": 0, + "content": "[73] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.305, + 0.905, + 0.347 + ], + "angle": 0, + "content": "[74] Meng You, Zhiyu Zhu, Hui Liu, and Junhui Hou. Nvs-solver: Video diffusion model as zero-shot novel view synthesizer. arXiv preprint arXiv:2405.15364, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.348, + 0.905, + 0.389 + ], + "angle": 0, + "content": "[75] Jason J. Yu, Fereshteh Forghani, Konstantinos G. Derpanis, and Marcus A. Brubaker. Long-term photometric consistent novel view synthesis with diffusion models, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.39, + 0.905, + 0.446 + ], + "angle": 0, + "content": "[76] Guofeng Zhang, Wei Hua, Xueying Qin, Tien-Tsin Wong, and Hujun Bao. Stereoscopic video synthesis from a monocular video. IEEE Transactions on Visualization and Computer Graphics, 13(4):686-696, 2007." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.447, + 0.905, + 0.502 + ], + "angle": 0, + "content": "[77] Min Zhao, Rongzhen Wang, Fan Bao, Chongxuan Li, and Jun Zhu. Controlvideo: Adding conditional control for one shot text-to-video editing. arXiv preprint arXiv:2305.17098, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.503, + 0.905, + 0.558 + ], + "angle": 0, + "content": "[78] Tinghui Zhou, Richard Tucker, John Flynn, Graham Fyffe, and Noah Snavely. Stereo magnification: Learning view synthesis using multiplane images. CoRR, abs/1805.09817, 2018." + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.558 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.51, + 0.957 + ], + "angle": 0, + "content": "821" + } + ] +] \ No newline at end of file diff --git a/2025/SpatialDreamer_ Self-supervised Stereo Video Synthesis from Monocular Input/24175650-223e-41c7-b804-6cc94c2c5191_origin.pdf b/2025/SpatialDreamer_ Self-supervised Stereo Video Synthesis from Monocular Input/24175650-223e-41c7-b804-6cc94c2c5191_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..10248e575b36fe05f0c28e5d3fc052b43a89a7c2 --- /dev/null +++ b/2025/SpatialDreamer_ Self-supervised Stereo Video Synthesis from Monocular Input/24175650-223e-41c7-b804-6cc94c2c5191_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c1125da4d0e21afbee7ffc811c3859d4ffd8bfab566616583606cac53a39201c +size 8444547 diff --git a/2025/SpatialDreamer_ Self-supervised Stereo Video Synthesis from Monocular Input/full.md b/2025/SpatialDreamer_ Self-supervised Stereo Video Synthesis from Monocular Input/full.md new file mode 100644 index 0000000000000000000000000000000000000000..fa1ab0d5eade7243f8b23e277f97098ceed83091 --- /dev/null +++ b/2025/SpatialDreamer_ Self-supervised Stereo Video Synthesis from Monocular Input/full.md @@ -0,0 +1,313 @@ +# SpatialDreamer: Self-supervised Stereo Video Synthesis from Monocular Input + +Zhen Lv $^{1}$ , Yangqi Long $^{1}$ , Congzhentao Huang $^{1}$ , Cao Li $^{1}$ , Chengfei Lv $^{1\dagger}$ , Hao Ren $^{1}$ , Dian Zheng $^{2}$ , $^{1}$ Alibaba Group; $^{2}$ Sun Yat-sen University + +# Abstract + +Stereo video synthesis from monocular input is challenging in spatial computing and virtual reality due to the lack of high-quality stereo video pairs for training and the difficulty of maintaining spatio-temporal consistency between frames. Existing methods primarily address these issues by directly applying novel view synthesis (NVS) techniques to video, while facing limitations such as the inability to effectively represent dynamic scenes and the requirement for extensive training data. In this paper, we introduce a novel self-supervised stereo video synthesis paradigm via a video diffusion model, termed SpatialDreamer, which meets the challenges head-on. Firstly, to address the stereo video data insufficiency, we propose a Depth based Video Generation module DVG, which employs a forward-backward rendering mechanism to generate paired videos with geometric and temporal priors. Leveraging data generated by DVG, we propose RefinerNet along with a self-supervised synthetic framework designed to facilitate efficient and dedicated training. More importantly, we devise a consistency control module, which consists of a metric of stereo deviation strength and a Temporal Interaction Learning module TIL for geometric and temporal consistency assurance respectively. We evaluated the proposed method against various benchmark methods, with the results showcasing its superior performance. Our project website is at: https://spatialdreamer.github.io. + +# 1. Introduction + +Stereo video synthesis from a monocular input aims to generate the target-view video based on the given view with geometric and spatio-temporal consistency, which has a wide range of applications in 3D model reconstruction [6], 3D movie production [18, 40], and Apple Vision Pro (AVP) [44] like virtual reality content. + +The primary difficulties of this task stem from the lack of adequate high-quality paired stereo videos for training, and + +![](images/615a11c17cd91356c759af8756fb3f396788be871862426545a3a5a91989c04a.jpg) +Figure 1. Visualization of the temporal consistency in stereo video generated by different methods. We extract the yellow line of each frame and stack them together. A good result should show a natural transition in the $t$ dimension. + +the challenge of preserving the geometric consistency between two views as well as maintaining the temporal consistency across generated frames. Traditional methods for generating stereo content involve capturing scenes using dual-camera setups [52, 59, 70, 76]. However, the acquisition of these images involves the use of a professional-grade camera, resulting in substantial costs. Recent advancements in deep learning mainly handle these problems by directly applying single to stereo or multi-view image generation techniques to video, while often facing challenges related to geometric and spatio-temporal inconsistencies [21, 32, 41, 50, 66, 67]. As shown in Fig 1, methods based on NVs are challenging to maintain accurate spatio-temporal consistency, and even the recently released AVP 3D photo converter [2] may inevitably result in content flickering and inconsistency when applied to video synthesis. This is due to the difficulty of handling the temporal coherence between frames and the spatial information between paired views in complex dynamic scenes. + +In this paper, we propose a self-supervised stereo video synthesis paradigm via a video diffusion model, termed SpatialDreamer, which will meet the data insufficiency and spatio-temporal inconsistency at once. Firstly, we design a Depth based Video data Generation module DVG to handle the problem of data insufficiency. Similar to [21, 66], DVG constructs a training pair of two views using forward-backward rendering without the need for data annotation [51, 59, 70], while [21, 66] only render individual images, resulting in flickering effects in generated video [9]. In this paper, we improve the generation of video data for refining stereo occlusion masks by leveraging the inter-frame motion obtained from optical flow [29]. This methodology enables DVG to produce paired video data that maintains both geometric and temporal consistency. Leveraging paired video generated by DVG, we propose Refiner-Net, along with a self-supervised video synthesis framework, to enable efficient and targeted training. More importantly, with access to sufficient paired videos, we devise a consistency control module, which consists of a metric of stereo deviation strength and a Temporal Interaction Learning module TIL. The stereo deviation strength aims to enable the generation of stereo videos in diverse real-world scenarios, with a stereo-aware loss further supervising the model to learn the magnitude of difference between paired view features in the latent space. TIL integrates the latent features from long-temporal frames as a global information to augment the temporal coherence of the generated results. + +Extensive experiments compared with various benchmark methods show that we achieve state-of-the-art performance. Notably, our SpatialDreamer meets the demand of real-world application without jitter, geometric and temporal inconsistency, which even beats AVP 3D converter and any open-source stereo video synthesis methods. We summarize our contributions as follows: + +- A novel self-supervised stereo video synthesis framework, SpatialDreamer, is proposed, which is robust across a wide range of scenes and dynamic content in video. Moreover, equipping the DVG module, we build a stereo video dataset using a self-supervised approach. +- A consistency control module is devised, which consists of a metric of stereo deviation strength and a temporal interaction learning module, ensuring geometric and temporal consistency in video sequences. +- The sufficient experiments and quantitative results demonstrate that the proposed method outperforms the age-of-the-art methods, even beats AVP. + +# 2. Related Work + +# 2.1. Single-image Novel View Synthesis + +Early studies address the single-image novel view synthesis task without relying on explicit 3D representation, + +such as End-to-end Synthesis [69], Infinite-Nature [37] and Animatable 3D Characters [68]. These methods lack explicit 3D representations for scenes, leading to inconsistent views and causing flickering and blurring, especially in occluded regions. Layer-based methods, such as 3D-photography [56] and SLIDE [27] represent a 3D scene using discrete layers. Yet, the synthesized views may lack accuracy in texture and geometric consistency across different perspectives. SMPI [61], AdaMPI [21] and SinMPI [47] represent the 3D scene using multiple planes, while their rendered results may suffer from depth discretization artifacts and repetitive texture patterns. To reduce the reliance on large labeled datasets, Shi et.al [55], Wang et.al. [66] and NVSVDE-Net [4] explore self-supervised learning for creating novel view images from single image. Although they produce photorealistic results, they struggle with temporal and spatial consistency, especially in dynamic scenes. + +As for our approach, we design a depth-based video generation module that generates paired videos with geometric and temporal priors. Additionally, Our method ensures multi-view consistency for the entire scene by employing a consistency control module. + +# 2.2. Multi-view Novel View Synthesis + +Early approaches in this field frequently utilized geometric methods, such as multi-view stereo reconstruction [30, 54] and structure-from-motion techniques [46, 53], to generate novel viewpoints from multiple images captured at different angles. NeRF-based methods were proposed to generate novel views based on fewer input images, including SinNeRF [72], CoPoNeRF [24] and NViST [28]. 3D Gaussian Splatting [32] presented an explicit representation of the scene. Gaussian-based methods, such as pixelSplat [12] and MVSplat [14], can produce high-quality synthesis results from multi-view images. Recently, researchers have begun exploring the application of advanced diffusion models for novel view synthesis, driven by their remarkable success in this area. Notable approaches such as Photoconsistent-NVS [75], MultiDiff [42], GCD [63] and NVS-Solver [74] generate image sequences based on predefined camera trajectories using either single or multiple input images. While these methods have shown promising results, they often struggle with dependence on dense scene geometry, handling complex scene dynamics, and generalization challenges [32]. Moreover, the challenge of optimizing NeRF/GS with sparse views hampers the ability of multi-view NVS to produce high-quality results. + +In contrast, we explore a video diffusion model to achieve stereo video synthesis, while preserving geometric structures through a RefinerNet based framework. + +![](images/4dcc85377b5a73a634d9a166cd85995df87c19ba07c84e4620eae5e7e486c4f6.jpg) +Figure 2. Overview of the proposed method, given left view video, the target view video is rendered, encoded, and concatenated with multi-frame noise, followed by the denoising U-Net architecture (i.e., SVD). The reference view images are fed into RefinerNet, through which the spatial features are extracted. (Notably, the left and right views are the target and reference view for both training and inference). The temporal interaction learning module integrates the latent features from long-temporal frames, and the deviation strength is projected as positional embedding and added to the time step embedding. Finally, the variational autoencoder decoder decodes the result into a video clip. + +# 2.3. Video Synthesis via Diffusion Models + +For video synthesis, previous works mainly expand image models to video, exploiting the image generation capability in the pretrained image diffusion models [38]. The video diffusion model (VDM) [23] and Video LDM [8] are proposed to adopt a 3D U-Net structure for joint image and video training. Subsequent works [15, 19, 19, 25, 57] also follow this insight with temporal attention, and [33, 48, 65, 71] further apply cross-frame attention to improve the consistency of the generated video. Another approach to promote temporal consistency is encoding video into 2D images [3, 45], which has been mainly explored in the video editing area. Although these models produce consistent results, they often require per-video optimization, which takes a long time, and would result in degradation when dealing with large motion. Most recently, [10, 20, 34] employ a transformer [64] architecture to model the temporal and spatial relations, demonstrating excellent video generation capabilities. However, these models need very large scale video datasets for training, which can be difficult to collect. + +Beyond temporal information modeling, some works [13, 17, 67, 77] have introduced aligned spatial guidance to enhance the stability of the generated videos. The proposed framework also uses spatial guidance. Differing from those methods which require spatially aligned controls, a spatially unaligned reference image is used via a cross-attention mechanism. Furthermore, a metric of stereo deviation strength is introduced to strengthen the spatial constraint, resulting in more consistent results. + +# 3. SpatialDreamer + +Given an image sequence and stereoscopic pose describing the scene depth information, the proposed method generates a stereo pair. In this section, we first provide a brief introduction about SVD [7]. Secondly, we offer detailed information of our video generation module DVG, which can generate sufficient paired video with geometric and temporal consistency. Using DVG, we present the framework for self-supervised learning-based video synthesis to support efficient and dedicated training. More importantly, to ensure consistency between frames, we design a consistency control module for geometric and temporal consistency assurance. Finally, we present the detailed training process. + +# 3.1. Preliminaries + +The challenge in video generation lies in accurately modeling the spatial-temporal dynamics, involving spatial relationships within frames and temporal relationships across frames. To address this challenge, Stable Video Diffusion (SVD) [7] introduces temporal convolution and attention layers. SVD follows the latent diffusion model (LDM) to encode video pixels into the latent space, enabling a more efficient denoising process. Given a video $I_{v}$ , encoder $\mathcal{E}$ encodes $I_{v}$ into the latent space as $z_{0} = \mathcal{E}(I_{v})$ . Gaussian noise $\epsilon$ is then added in the Markov process: + +$$ +z _ {t} = \sqrt {\bar {\alpha} _ {t}} z _ {0} + \sqrt {1 - \bar {\alpha} _ {t}} \epsilon_ {t} \tag {1} +$$ + +where $\bar{\alpha}_0,\dots ,\bar{\alpha}_T$ are the pre-defined noise schedules for T steps in the Markov process. In particular, SVD is trained + +by the $v$ -prediction formulation with the mean squared error (MSE) between the ground truth and its prediction in the latent space, known as $l_{\epsilon}$ . + +# 3.2. Depth based Video Generation + +To create training pairs of images along with their corresponding occluded masks, the following self-supervised mechanism is employed utilizing real images. Firstly, the depth of the given video sample is estimated [49]. Secondly, the reference view image $x_{1}$ is rendered into a masked one under the target viewpoint $P_{2}$ , and then the trained inpainting model [58] is used to fill these occluded regions to obtain the novel viewpoint $x_{2}$ [43]. Finally, $x_{2}$ is rendered to the original viewpoint $P_{1}$ via a forward rendering. After this, we can obtain the masked image $\widetilde{x}_{1}$ with mask $M$ under viewpoint $P_{1}$ , as well as the inpainted image $x_{2}$ under viewpoint $P_{2}$ . The mask $M$ and masked image $\widetilde{x}_{1}$ are considered to be the conditions, with the image $x_{1}$ serving as the ground truth for the diffusion model. This approach involves encoding the inpainted image $x_{2}$ into latent feature representations and feeding it into RefinerNet. + +However, the aforementioned approach encounters challenges in addressing inconsistencies among generated data frames, resulting in observable artifacts such as jittering and flickering. To tackle this issue, we leverage optical flow [60] to capture motion between successive video frames. Specifically, by leveraging the optical flow between the reference image $x_{1}^{t}$ at time $t$ and its adjacent frames, we establish the pixel correspondences between the image $x_{1}^{t}$ and the images $x_{1}^{t-1}$ and $x_{1}^{t+1}$ , respectively. Particularly, these optical flows enable computing a confidence map via forward-backward consistency [16], which allows us to more accurately merge the stereo occlusions in the masks between adjacent frames [1]. Subsequently, the refined occlusion mask $m^{t}$ is generated by sampling from the occlusion maps $m^{t-1}$ and $m^{t+1}$ , leveraging the pixel correspondences between consecutive frames under the reference view, as described by the following equation. + +$$ +m ^ {t} (i, j) = \left\{ \begin{array}{l l} 1, & \sum_ {k \in \{t - 1, t + 1 \}} m ^ {k} (i + u, j + v) \cdot \\ & C (i, j) \geq 1 \\ m ^ {t} (i, j), & e l s e \end{array} \right. \tag {2} +$$ + +where $u$ and $v$ are the $x$ and $y$ components of the optical flow vector, and $C$ represents the optical flow confidence measure. By propagating information from occluded areas across frames, this inter-frame continuity facilitates a more precise refinement of occluded regions in paired view videos and enforces temporal smoothness and coherence within the self-supervised framework, effectively mitigating the undesirable effects of jitter and flicker. + +![](images/804bf090dcca49d6bf77fd03da14a6625ff8ef4378bf10a90929487f83332cff.jpg) +Figure 3. DVG. The temporal motion can be utilized to refine the occluded regions in the current frame, thereby providing smoother images and more consistent occlusion in the temporal sequence. + +As shown in Figure 3, the difference between the training and inference procedures is that the former uses both backward and forward rendering whereas the latter uses only forward rendering to obtain an image under the target view. + +# 3.3. Self-Supervised Stereo Video Synthesis framework + +Leveraging the generated paired view videos, we propose a self-supervised framework for stereo video synthesis, as illustrated in Figure 2. This framework commences with the rendering of the target view video that incorporates occlusions, followed by the application of a spatio-temporal method to effectively inpaint the missing regions, ultimately yielding a complete non-occluded target view video. + +To learn the distribution of the feature space between the paired view images, we devise RefinerNet, which is modeled after the architecture of the denoising U-Net, excluding the temporal layer. RefinerNet adopts the weight initialization from the original SD2.1 model [50]. As depicted in Figure 2, a spatial attention layer is employed instead of the self-attention layer. Given a feature map $z_{t} \in \mathbb{R}^{t \times h \times w \times c}$ from the denoising U-Net and another feature map $z_{r} \in \mathbb{R}^{t \times h \times w \times c}$ from RefinerNet, they are concatenated along the h dimension, and then self-attention is applied. Subsequently, the first half of the feature map is retrieved as the output. One great advantage of this approach is that it enables RefinerNet to maintain the feature representation capability inherent in the original SD Unet, extracting reference view image information. In addition, the denoising U-Net can adaptively learn features from RefinerNet that are correlated in the same feature space, owing to the fundamentally similar network architectures between the denoising U-Net and RefinerNet. This not only ensures a robust initialization for the feature representation but also enhances the learning ability of the entire network. + +# 3.4. Consistency Control Module + +# 3.4.1 Temporal Interaction Learning module + +While our RefinerNet-based approach providing spatial guidance, paired view features are still merged independently, posing challenges for temporal consistency and potentially causing texture discrepancies or odd results between adjacent frames. + +Inspired by previous studies [11, 22], we propose an attention-based temporal interaction learning module that learns the distribution of the feature space from adjacent frames, named TIL. Specifically, given latent features of $N_r$ adjacent images under reference view $z_i^t$ , where $i = 1, 2, \dots, N_r$ and the latent feature of the reference view image $z_r^t$ at the time step $t$ , we augment the $z_r^t$ to $aug_r^t$ as: + +$$ +a u g _ {r} ^ {t} = \lambda \cdot \operatorname {A t t n} _ {r, r} + (1 - \lambda) \cdot \frac {1}{N _ {r}} \sum_ {i = 1} ^ {N _ {r}} \operatorname {A t t n} _ {r, i} \tag {3} +$$ + +where $\lambda \in [0,1]$ and $\mathrm{Attn}_{i,j}$ defines attention between two latent features $z_{i}$ and $z_{j}$ . This module blends self-attention of $z_{r}^{t}$ in RefinerNet's U-Net block with the cross-view attention between $z_{r}^{t}$ and each adjacent view $z_{i}^{t}$ . The cross-view attention aligns the feature of all adjacent frames to the reference view while the self-attention helps each reference image retain distinctiveness. The augmented reference feature $aug_{r}^{t}$ is then fed into spatial attention layer to assist U-Net network learning. + +# 3.4.2 Stereo Deviation Strength + +During the training, we noticed that different viewpoints impact the 3D impression of the scenes in the generated videos. However, relying solely on a given viewpoint as a guide for generating stereoscopic videos is insufficient. This is because videos with the same viewpoint may exhibit diverse depth ranges and variations, depending on the scene [70]. Therefore, pose alone cannot accurately control the 3D effect of a scene. Consequently, we introduce a metric called the stereo deviation strength, which quantitatively assesses the binocular disparity in a scene and facilitates the creation of controllable stereo vision: + +$$ +s (z) = \left| z _ {0} - z _ {\text {r e f}} \right| \tag {4} +$$ + +where the stereo deviation strength quantifies the latent differences between reference view $z_{ref}$ and target view $z_0$ . Similar to the time step, the deviation strength is projected into a positional embedding and added to each frame in the residual block to ensure uniform application of the deviation strength to every frame. The impact of stereo deviation strength guidance on the results is illustrated in Figure 4. + +Moreover, for better convergence, a stereo-aware loss function is proposed to directly supervise the disparity dif + +![](images/fc476e0ee20f138f5507c314034e7853356f1d077a6ae97128a8663bfe3f6e0f.jpg) +Figure 4. Stereo deviation strength guidance examples. Augmenting the deviation strength enhances the 3D photography effect, but excessive deviation strength may lead to image distortions or even lower the quality of stereoscopic image. + +ference: + +$$ +l _ {d} = \left| \left| s \left(z _ {0}\right) - s \left(\hat {z} _ {0}\right) \right| \right| _ {2} ^ {2} \tag {5} +$$ + +where $\hat{z}_0$ represents the estimated clean video latent $z_0$ , which can be obtained by: + +$$ +\hat {z} _ {0} = z _ {0} - \frac {\sqrt {1 - \bar {\alpha} _ {t}} \epsilon_ {\theta} (z _ {t} , c , t)}{\sqrt {\bar {\alpha} _ {t}}} \tag {6} +$$ + +where $\overline{\alpha}_t = \prod_{i=1}^t (1 - \beta_t)$ , and the noise prediction loss is combined with the stereo-aware loss by a scaling factor: + +$$ +l = l _ {\epsilon} + \lambda \cdot l _ {d} \tag {7} +$$ + +# 3.5. Training Strategy + +The training process is divided into two stages. In the first training stage, we focus on individual frames from videos. The temporal layer in denoising U-Net is frozen, and cross-view attention in TIL module is removed. The Refiner-Net model and the denoising U-Net are initialized with pretrained weights from SD2.1 [50] and SVD [7] respectively. The weights of the variational autoencoder encoder and decoder, as well as the contrastive language-image pretraining (CLIP) image encoder, are all kept fixed. The aim of this stage is for the model to learn to synthesize new viewpoint images under the condition of a reference image and a new viewpoint pose. In the second stage, we train the temporal layer and cross-view attention in TIL module with video sequences. This enables the model to capture temporal context information efficiently. The input for the model is the 8-frame video clip and $N_r$ for TIL is set to 8. + +# 4. Experiments + +# 4.1. Datasets, Baselines and Metrics + +We conduct our experiments on image and video levels with RealEstate10K [78] and self-collected video data. + +RealEstate10K is a large, widely used dataset with camera poses, corresponding to 10 million frames derived from about 80,000 video clips, gathered from about 10,000 YouTube videos. As there is not open-source stereo video benchmark, we collect 1500 monocular videos with 1400 for training, 100 for testing. For image synthesis, we compared the proposed method with recent open-source SOTA methods: 3D-photography [56], SynSim [69], AdaMPI [21], SinMPI [47], Wang et al. [66], Photoconsistent-NVS(P-NVS) [75], NVSVDE-Net [4], CoPoNeRF [24] and MVsplat [14]. For a fair comparison, the 3D photography method, AdaMPI, SinMPI, and the proposed method all used MiDaS [49] for depth estimation. Following the method of 3D-photography [56], we measured the accuracy of the generated target views with regard to the ground-truth images using the three metrics of the learned perceptual image patch similarity (LPIPS), peak signal-to-noise ratio (PSNR), and structural similarity index measure (SSIM). + +For the video synthesis, we picked several representative approaches, i.e., the method of 3D-photography [56], the open-source code repository webui-depthmap, P-NVS [75], CoPoNeRF [24], NVS-Solver [74] and AVP [44]. Moreover, we conducted the same depth estimation method [49] for a fair comparison of 3D-photography [56], webui-depthmap and NVS-Solver [74]. We employ the FVD [62] score to measure the perceptual similarity between input videos and outputs and report the flow warping error $E_{\text{warp}}$ [35] to assess the temporal consistency of the resulting video sequences. + +# 4.2. Implementation + +We employed the AdamW [39] optimizer with a learning rate of $1 \times 10^{-5}$ to train the model on two NVIDIA A800 GPUs. In the first training stage, individual video frames were resized and center-cropped to a resolution of $1024 \times 1024$ . We trained the model for 100,000 iterations with a batch size of 8. In the second training stage, the videos were first split according to the transitions to ensure that a scene only appeared in one video clip. We then formed each video clip for training with eight frames. Finally, the temporal layer and TIL module was trained for 10,000 steps with a batch size of 2. The stereo-aware loss scaling factor $\lambda$ was set to 0.001 and $\lambda$ in Eq. 3 was set as 0.6. + +# 4.3. Qualitative Results + +# 4.3.1 Stereo Image Synthesis + +We visually compared different methods for generating new view images from a target viewpoint. We qualitatively compared the newly synthesized images in Figure 5. 3D-photography fails to resolve the image distortion problem, and the image appears to be enhanced only slightly by the + +rendering. SynSin generates the most blurry details and loses a lot of information, compared with the input view. Neither AdaMPI nor SinMPI can generate fine enough details, and the results contain noticeable artifacts. CoPoN-erF tends to produce discrepancies from the original image in terms of fine-grained details, and it struggles with effectively handling scenes involving motion, as well as MVsplat. In contrast, the proposed method shows a clean and detailed output, especially in the area of depth information discontinuity. + +# 4.3.2 Stereo Video Synthesis + +Moreover, we qualitatively compared the results of the stereo video synthesis with other methods. In addition to the quality issues of single-frame generation, the most important aspect in videos is the temporal consistency. We seek to generate videos that are free of flicker. The visualization of generated videos from different methods are illustrated in Figure 6, for each video, we show the scanline (the yellow line in figure) slice through the spatial-temporal volume. Our method achieve the smoothest trajectories, while other methods exhibit significant flickering. The most distinct regions are zoomed in and highlighted under the images to illustrate the details. In the first case, our method attains the smooth trajectories of the background (maintaining the straight line). And the second case, only our method preserves the texture of the trees behind the woman. And the last case, only our method preserves the texture of the black vertical lines. More results of generated frames are also presented in supplementary material. + +# 4.4. Quantitative Results + +Table 1 lists the SSIM, PSNR, and LPIPS scores for the newly synthesized images. The proposed method outperforms 3D-photography and SynSin in the major metrics by a wide margin. The method of Wang et al., P-NVS and NVSVDE-Net perform slightly worse in terms of all metrics compared to the proposed method. Since the proposed method merges the spatial features and fills the occluded regions, it achieves the best LPIPS scores and shows competitive results in PSNR and SSIM when compared with AdaMPI and CoPoNeRF. MVSplat outperforms other methods when $t = 10$ but is inferior to ours when $t = 5$ . Overall, the results exhibit superior visual effects, which is consistent with the performance observed in the qualitative experiments. We further demonstrate the effectiveness of our method on video synthesis, as shown in Table 2. The proposed method achieves the best performance on both FVD and $E_{\text{warp}}$ , which means that our generated videos attain the best quality and temporal consistency. + +![](images/86c7f958e9f5db0238d862fe33cbd5bdb019e9307947a7de00988dc1e58aaee6.jpg) +Figure 5. Qualitative results on the RealEstate10K dataset. The proposed method generates better-quality information and maintains the edge and textural details of the reference image better than the other methods. More results are available in supplementary material. + +![](images/d9b5162f3405508b0ac7cf01973d65d5bca984da229fd3cb8df89e4f02d3c9d7.jpg) +Figure 6. Visual comparisons in stereo video benchmark. We show the scanline (yellow line in original frame) slice through the spatial-temporal volume. The most distinct regions are zoomed in and highlighted under the images to illustrate the details. + +![](images/24b1891b54fa31e041b0e92c79ce76e931da71388d454afe23cefbdd6c07ea0d.jpg) +Figure 7. Ablation study of different designs. Only the proposed method ensures consistent preservation of details with the groundtruth target view. More results are available in supplementary material. + +# 4.5. Ablation Study + +# 4.5.1 Effectiveness of each components + +We show the results in Table 3 on the RealEstate10K (stereo image) and collected stereo video datasets. CN means sub + +stituting our RefinerNet (RN) to ControlNet, SDS means the proposed stereo deviation strength. For image-level comparison, the results show that utilizing U-Net architecture only is not sufficient for stereo image synthesis and our RefinerNet is a better choice than ControlNet. Adding SDS further improves the geometric consistency. As visualized in Figure 7, solely relying on U-Net features results in distortion in the edge area and struggles with generating content properly. It also fails to preserve textural details near a plane boundary. Adding ControlNet often results in content inconsistency, presumably due to the failure of the paired view feature integration. RefinerNet handles this problem, showing it is a better architecture for stereo image generation. SDS further improves the consistency with the input and contain a more detailed image structure. As for video-level comparison, the quantitative results are shown in Table 3, adding TIL and DVG both attain the performance gain, and our final model achieves the best performance. + +Table 1. Quantitative comparison on RealEstate10K dataset. $\dagger$ means re-implementing in our setting for fair comparison. When $t = 5$ , the current frame serves as the left view and the frame at time +5 serves as the right view; the same approach applies for $t = 10$ . The best results are in bold. + +
SSIM ↑PSNR ↑LPIPS ↓
t=5t=10t=5t=10t=5t=10
3D-photography [56]†0.6410.44611.089.5330.1160.386
SynSin [69]†0.7400.64521.2219.550.0870.104
AdaMPI [21]†0.8690.73627.7022.650.0410.087
SinMPI [47]†0.8980.77528.7923.210.0400.080
Wang et al. [66]0.8400.72025.3521.360.0490.095
P-NVS [75]†0.7240.66721.9819.460.2410.279
NVSVDE-Net [4]0.84024.310.233
CoPoNeRF [24]†0.6580.63222.6321.330.1710.209
MVSplat [14]†0.86325.890.132
Proposed0.9160.85732.2624.860.0380.049
+ +Table 2. Quantitative comparison with the other methods. The best results are in **bold**. $E_{\text{warp}}^*$ denotes $E_{\text{warp}} \left( \times 10^{-3} \right)$ . + +
FVD ↓E* warp ↓
3D-photography [56]155.03.418
Webui-depthmap83.183.486
P-NVS [75]420.223.02
CoPoNeRF [24]290.07.144
NVS-Solver [74]249.15.842
MVSplat [14]15437.913
AVP [44]99.923.446
Proposed67.093.374
+ +Table 3. Quantitative comparison for the ablation study. The best results are in bold. + +
MethodCNRNSDSTILDVGStereo ImageStereo Video
SSIM↑PSNR↑LPIPS↓FVD↓
Image-level✓0.88023.730.06-
0.85524.040.183-
✓0.89530.200.043-
✓✓0.91632.260.038-
Video-level✓✓---184.0
✓✓✓----123.5
✓✓✓---85.21
✓✓✓✓---67.09
+ +# 4.5.2 Data improvement by DVG + +We show the generated stereo occlusion mask with interframe motion refinement in Figure 8. It can be observed that the occlusions appear more complete spatially and exhibit smoother in their temporal distribution. By employing DVG, we can acquire sufficient data to support the generation of spatio-temporally consistent videos. More analyses + +![](images/776aa8844cfaa92050a9122ee32602002cc6d873b8640d0800de504c952bd9b1.jpg) +Figure 8. Visual comparisons for the improvement of stereo occlusion. We show the scanline of (yellow line in original frame) slice through the spatial-temporal volume. The red region represents the changes of occlusion mask. + +![](images/10e1e42214e722de071131e3b2234a58c99e0ecccd585d0d8b581fe539a449b4.jpg) + +![](images/5666cf02f20695baf24a3ee0a7ab9d6399bce9c90c61d915097c0fe5325b141e.jpg) + +are available in the supplementary materials. + +# 4.5.3 Exploration of different depth estimation methods + +We further compared different design choices of depth map including DepthAnything [73], Marigold [31], MiDaS [49], ZoeDepth [5] and DepthCrafter [26]. The results are available in supplementary materials. The proposed method can produce new viewpoint images that are realistic, seamless, and rich in detail, regardless of the depth estimation method used. This capability demonstrates the versatility and robustness of the proposed method, ensuring superior performance in various applications and environments. + +# 4.6. Limitations + +SpatialDreamer effectively creates stereo video from a single input but faces limitations. It relies on depth maps for rendering, affecting stereoscopic quality due to depth accuracy and distribution. Future research could focus on integrating implicit depth to address this [36]. Additionally, the model's large parameter size limits real-time performance. + +# 5. Conclusion + +In this paper, we introduce a new self-supervised stereo video synthesis approach using a video diffusion model, termed SpatialDreamer. This method addresses the issues of insufficient data and spatio-temporal inconsistency between frames. To solve the problem of insufficient data, we develop a Depth-based Video Generation (DVG) module that uses a forward-backward rendering mechanism to generate a rendered video with geometric and temporal priors. Additionally, we propose RefinerNet along with a self-supervised synthetic framework to enable efficient and dedicated training using data generated by DVG. Furthermore, we design a consistency control module to ensure geometric and temporal consistency. Our SpatialDreamer outperforms all other open-source stereo image and video synthesis methods and has the potential for future expansion into virtual reality applications. + +# References + +[1] Dmitry Akimov, Alexey Shestov, Alexander Voronov, and Dmitriy Vatolin. Occlusion refinement for stereo video using optical flow. In 2012 International Conference on 3D Imaging (IC3D), pages 1-8, 2012. +[2] Andrea Schubert. visionos 2 brings new spatial computing experiences to apple vision pro. https://www.apple.com/newsroom/2024/06/visionos-2-brings-new-spatial-computing-experiences-to-apple-vision-pro, 2024. +[3] Omer Bar-Tal, Dolev Ofri-Amar, Rafail Fridman, Yoni Kasten, and Tali Dekel. Text2live: Text-driven layered image and video editing. In European conference on computer vision, pages 707-723. Springer, 2022. +[4] Juan Luis Gonzalez Bello and Munchurl Kim. Novel view synthesis with view-dependent effects from a single image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. +[5] Shariq Farooq Bhat, Reiner Birkl, Diana Wofk, Peter Wonka, and Matthias Müller. Zoedepth: Zero-shot transfer by combining relative and metric depth, 2023. +[6] Phuong Ngoc Binh Do and Quoc Chi Nguyen. A review of stereo-photogrammetry method for 3-d reconstruction in computer vision. In 2019 19th International Symposium on Communications and Information Technologies (ISCIT), pages 138–143, 2019. +[7] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. +[8] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22563-22575, 2023. +[9] Aaron Bobick and Stephen Intille. Large occlusion stereo. International Journal of Computer Vision, 33:181-200, 1999. +[10] Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Yufei Guo, Li Jing, David Schnurr, Joe Taylor, Troy Luhman, Eric Luhman, Clarence Ng, Ricky Wang, and Aditya Ramesh. Video generation models as world simulators. 2024. +[11] Mingdeng Cao, Xintao Wang, Zhongang Qi, Ying Shan, Xiaohu Qie, and Yinqiang Zheng. Masactrl: Tuning-free mutual self-attention control for consistent image synthesis and editing, 2023. +[12] David Charatan, Sizhe Li, Andrea Tagliasacchi, and Vincent Sitzmann. pixelsplat: 3d gaussian splats from image pairs for scalable generalizable 3d reconstruction. In arXiv, 2023. +[13] Weifeng Chen, Jie Wu, Pan Xie, Hefeng Wu, Jiashi Li, Xin Xia, Xuefeng Xiao, and Liang Lin. Control-a-video: Controllable text-to-video generation with diffusion models. arXiv preprint arXiv:2305.13840, 2023. + +[14] Yuedong Chen, Haofei Xu, Chuanxia Zheng, Bohan Zhuang, Marc Pollefeys, Andreas Geiger, Tat-Jen Cham, and Jianfei Cai. Mvsplat: Efficient 3d gaussian splatting from sparse multi-view images. arXiv preprint arXiv:2403.14627, 2024. +[15] Zuozhuo Dai, Zhenghao Zhang, Yao Yao, Bingxue Qiu, Siyu Zhu, Long Qin, and Weizhi Wang. Animateanything: Fine-grained open domain image animation with motion guidance. arXiv e-prints, pages arXiv-2311, 2023. +[16] G. Egnal and R.P. Wildes. Detecting binocular half-occlusions: empirical comparisons of five approaches. IEEE Transactions on Pattern Analysis and Machine Intelligence, 24(8):1127-1133, 2002. +[17] Patrick Esser, Johnathan Chiu, Parmida Atighehchian, Jonathan Granskog, and Anastasis Germanidis. Structure and content-guided video synthesis with diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7346-7356, 2023. +[18] Chen Gao, Ayush Saraf, Johannes Kopf, and Jia-Bin Huang. Dynamic view synthesis from dynamic monocular video, 2021. +[19] Yuwei Guo, Ceyuan Yang, Anyi Rao, Zhengyang Liang, Yaohui Wang, Yu Qiao, Maneesh Agrawala, Dahua Lin, and Bo Dai. Animatediff: Imagine your personalized text-to-image diffusion models without specific tuning. International Conference on Learning Representations, 2024. +[20] Agrim Gupta, Lijun Yu, Kihyuk Sohn, Xiuye Gu, Meera Hahn, Li Fei-Fei, Irfan Essa, Lu Jiang, and José Lezama. Photorealistic video generation with diffusion models. arXiv preprint arXiv:2312.06662, 2023. +[21] Yuxuan Han, Ruicheng Wang, and Jiaolong Yang. Single-view view synthesis in the wild with learned adaptive multiplane images. In ACM SIGGRAPH 2022 Conference Proceedings, pages 1-8, 2022. +[22] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control, 2022. +[23] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2022. +[24] Sunghwan Hong, Jaewoo Jung, Heeseong Shin, Jiaolong Yang, Seungryong Kim, and Chong Luo. Unifying correspondence, pose and nerf for pose-free novel view synthesis from stereo pairs, 2024. +[25] Li Hu, Xin Gao, Peng Zhang, Ke Sun, Bang Zhang, and Liefeng Bo. animate anyone: Consistent and controllable image-to-video synthesis for character animation. arXiv preprint arXiv:2311.17117, 2023. +[26] Wenbo Hu, Xiangjun Gao, Xiaoyu Li, Sijie Zhao, Xiaodong Cun, Yong Zhang, Long Quan, and Ying Shan. Depthcrafter: Generating consistent long depth sequences for open-world videos. arXiv preprint arXiv:2409.02095, 2024. +[27] Varun Jampani, Huiwen Chang, Kyle Sargent, Abhishek Kar, Richard Tucker, Michael Krainin, Dominik Kaeser, William T Freeman, David Salesin, Brian Curless, et al. Slide: Single image 3d photography with soft layering and depth-aware inpainting. In Proceedings of the IEEE/CVF In + +ternational Conference on Computer Vision, pages 12518-12527, 2021. +[28] Wonbong Jang and Lourdes Agapito. Nvist: In the wild new view synthesis from a single image with transformers, 2024. +[29] Jisoo Jeong, Hong Cai, Risheek Garrepalli, Jamie Menjay Lin, Munawar Hayat, and Fatih Porikli. Ocai: Improving optical flow estimation by occlusion and consistency aware interpolation, 2024. +[30] Hailin Jin, Stefano Soatto, and Anthony J Yezzi. Multi-view stereo reconstruction of dense shape and complex appearance. International Journal of Computer Vision, 63:175-189, 2005. +[31] Bingxin Ke, Anton Obukhov, Shengyu Huang, Nando Metzger, Rodrigo Caye Daudt, and Konrad Schindler. Repurposing diffusion-based image generators for monocular depth estimation, 2024. +[32] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering, 2023. +[33] Levon Khachatriyan, Andranik Movsisyan, Vahram Tadevosyan, Roberto Henschel, Zhangyang Wang, Shant Navasardyan, and Humphrey Shi. Text2video-zero: Text-to-image diffusion models are zero-shot video generators. arXiv preprint arXiv:2303.13439, 2023. +[34] Dan Kondratyuk, Lijun Yu, Xiuye Gu, José Lezama, Jonathan Huang, Rachel Hornung, Hartwig Adam, Hassan Akbari, Yair Alon, Vighnesh Birodkar, et al. Videopoet: A large language model for zero-shot video generation. arXiv preprint arXiv:2312.14125, 2023. +[35] Wei-Sheng Lai, Jia-Bin Huang, Oliver Wang, Eli Shechtman, Ersin Yumer, and Ming-Hsuan Yang. Learning blind video temporal consistency. In Proceedings of the European conference on computer vision (ECCV), pages 170-185, 2018. +[36] Hanwen Liang, Junli Cao, Vedit Goel, Guocheng Qian, Sergei Korolev, Demetri Terzopoulos, Konstantinos Plataniotis, Sergey Tulyakov, and Jian Ren. Wonderland: Navigating 3d scenes from a single image. arXiv preprint arXiv:2412.12091, 2024. +[37] Andrew Liu, Richard Tucker, Varun Jampani, Ameesh Makadia, Noah Snavely, and Angjoo Kanazawa. Infinite nature: Perpetual view generation of natural scenes from a single image. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021. +[38] Yixin Liu, Kai Zhang, Yuan Li, Zhiling Yan, Chujie Gao, Ruoxi Chen, Zhengqing Yuan, Yue Huang, Hanchi Sun, Jianfeng Gao, Lifang He, and Lichao Sun. Sora: A review on background, technology, limitations, and opportunities of large vision models, 2024. +[39] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. +[40] Kevin Matzen, Michael F. Cohen, Bryce Evans, Johannes Kopf, and Richard Szeliski. Low-cost 360 stereo photography and video capture. ACM Trans. Graph., 36(4), 2017. +[41] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis, 2020. + +[42] Norman Müller, Katja Schwarz, Barbara Rössle, Lorenzo Porzi, Samuel Rota Bulò, Matthias Nießner, and Peter Kontschieder. Multidiff: Consistent novel view synthesis from a single image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10258-10268, 2024. +[43] Simon Niklaus and Feng Liu. Softmax splatting for video frame interpolation, 2020. +[44] Onee. What is spatial video on iphone 15 pro and vision pro. https://xreality-zone/en/posts/what-is-spatial-video-on-iphone-15-pro-and/apple-vision-pro, 2024. +[45] Hao Ouyang, Qiuyu Wang, Yuxi Xiao, Qingyan Bai, Juntao Zhang, Kecheng Zheng, Xiaowei Zhou, Qifeng Chen, and Yujun Shen. Codef: Content deformation fields for temporally consistent video processing. arXiv preprint arXiv:2308.07926, 2023. +[46] Onur Özyeşil, Vladislav Voroninski, Ronen Basri, and Amit Singer. A survey of structure from motion*. Acta Numerica, 26:305-364, 2017. +[47] Guo Pu, Peng-Shuai Wang, and Zhouhui Lian. Sinmpi: Novel view synthesis from a single image with expanded multiplane images, 2023. +[48] Chenyang Qi, Xiaodong Cun, Yong Zhang, Chenyang Lei, Xintao Wang, Ying Shan, and Qifeng Chen. Fatezero: Fusing attentions for zero-shot text-based video editing. arXiv preprint arXiv:2303.09535, 2023. +[49] René Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(3), 2022. +[50] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. +[51] Christoph Sager, Christian Janiesch, and Patrick Zschech. A survey of image labelling for computer vision applications. Journal of Business Analytics, 4(2):91-110, 2021. +[52] D. Scharstein. View Synthesis Using Stereo Vision. Springer Berlin Heidelberg, 1999. +[53] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4104-4113, 2016. +[54] S.M. Seitz, B. Curless, J. Diebel, D. Scharstein, and R. Szeliski. A comparison and evaluation of multi-view stereo reconstruction algorithms. In 2006 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR'06), pages 519-528, 2006. +[55] Yujiao Shi, Hongdong Li, and Xin Yu. Self-supervised visibility learning for novel view synthesis.*. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2021. +[56] Meng-Li Shih, Shih-Yang Su, Johannes Kopf, and Jia-Bin Huang. 3d photography using context-aware layered depth + +inpainting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8028-8038, 2020. +[57] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792, 2022. +[58] Lorenzo Stacchio. Train stable diffusion for inpainting, 2023. +[59] Wenxiu Sun, Lingfeng Xu, Oscar C Au, Sung Him Chui, and Chun Wing Kwok. An overview of free view-point depth-image-based rendering (dibr). In APSIPA Annual Summit and Conference, pages 1023-1030, 2010. +[60] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow, 2020. +[61] Richard Tucker and Noah Snavely. Single-view view synthesis with multiplane images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 551-560, 2020. +[62] Thomas Unterthiner, Sjoerd van Steenkiste, Karol Kurach, Raphaël Marinier, Marcin Michalski, and Sylvain Gelly. FVD: A new metric for video generation. In Deep Generative Models for Highly Structured Data, ICLR 2019 Workshop, New Orleans, Louisiana, United States, May 6, 2019. OpenReview.net, 2019. +[63] Basile Van Hoorick, Rundi Wu, Ege Ozguroglu, Kyle Sargent, Ruoshi Liu, Pavel Tokmakov, Achal Dave, Changxi Zheng, and Carl Vondrick. Generative camera dolly: Extreme monocular dynamic novel view synthesis. European Conference on Computer Vision (ECCV), 2024. +[64] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. +[65] Wen Wang, Yan Jiang, Kangyang Xie, Zide Liu, Hao Chen, Yue Cao, Xinlong Wang, and Chunhua Shen. Zero-shot video editing using off-the-shelf image diffusion models. arXiv preprint arXiv:2303.17599, 2023. +[66] Xiaodong Wang, Chenfei Wu, Shengming Yin, Minheng Ni, Jianfeng Wang, Linjie Li, Zhengyuan Yang, Fan Yang, Lijuan Wang, Zicheng Liu, et al. Learning 3d photography videos via self-supervised diffusion on single images. arXiv preprint arXiv:2302.10781, 2023. +[67] Xiang Wang, Hangjie Yuan, Shiwei Zhang, Dayou Chen, Jiuniu Wang, Yingya Zhang, Yujun Shen, Deli Zhao, and Jingren Zhou. Videocomposer: Compositional video synthesis with motion controllability. Advances in Neural Information Processing Systems, 36, 2024. +[68] Chung-Yi Weng, Brian Curless, and Ira Kemelmacher-Shlizerman. Photo wake-up: 3d character animation from a single photo. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5908-5917, 2019. +[69] Olivia Wiles, Georgia Gkioxari, Richard Szeliski, and Justin Johnson. Synsin: End-to-end view synthesis from a single image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020. + +[70] Andrew Woods, Tom Docherty, and Rolf Koch. Image distortions in stereoscopic video systems. Proc SPIE, 1915, 2002. +[71] Jay Zhangjie Wu, Yixiao Ge, Xintao Wang, Stan Weixian Lei, Yuchao Gu, Yufei Shi, Wynne Hsu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7623-7633, 2023. +[72] Dejia Xu, Yifan Jiang, Peihao Wang, Zhiwen Fan, Humphrey Shi, and Zhangyang Wang. Sinnerf: Training neural radiance fields on complex scenes from a single image, 2022. +[73] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data, 2024. +[74] Meng You, Zhiyu Zhu, Hui Liu, and Junhui Hou. Nvs-solver: Video diffusion model as zero-shot novel view synthesizer. arXiv preprint arXiv:2405.15364, 2024. +[75] Jason J. Yu, Fereshteh Forghani, Konstantinos G. Derpanis, and Marcus A. Brubaker. Long-term photometric consistent novel view synthesis with diffusion models, 2023. +[76] Guofeng Zhang, Wei Hua, Xueying Qin, Tien-Tsin Wong, and Hujun Bao. Stereoscopic video synthesis from a monocular video. IEEE Transactions on Visualization and Computer Graphics, 13(4):686-696, 2007. +[77] Min Zhao, Rongzhen Wang, Fan Bao, Chongxuan Li, and Jun Zhu. Controlvideo: Adding conditional control for one shot text-to-video editing. arXiv preprint arXiv:2305.17098, 2023. +[78] Tinghui Zhou, Richard Tucker, John Flynn, Graham Fyffe, and Noah Snavely. Stereo magnification: Learning view synthesis using multiplane images. CoRR, abs/1805.09817, 2018. \ No newline at end of file diff --git a/2025/SpatialDreamer_ Self-supervised Stereo Video Synthesis from Monocular Input/images.zip b/2025/SpatialDreamer_ Self-supervised Stereo Video Synthesis from Monocular Input/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..7de8013fa90dec0c8c21aca82320269aeff6462b --- /dev/null +++ b/2025/SpatialDreamer_ Self-supervised Stereo Video Synthesis from Monocular Input/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:67d05b33ef7514e4949f77cec6c4753c75cd299df3482188c1191543d91968cc +size 713660 diff --git a/2025/SpatialDreamer_ Self-supervised Stereo Video Synthesis from Monocular Input/layout.json b/2025/SpatialDreamer_ Self-supervised Stereo Video Synthesis from Monocular Input/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..cd549bc7e4e6ac0b6b2693f31652973b5910bbea --- /dev/null +++ b/2025/SpatialDreamer_ Self-supervised Stereo Video Synthesis from Monocular Input/layout.json @@ -0,0 +1,8842 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 60, + 102, + 549, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 102, + 549, + 121 + ], + "spans": [ + { + "bbox": [ + 60, + 102, + 549, + 121 + ], + "type": "text", + "content": "SpatialDreamer: Self-supervised Stereo Video Synthesis from Monocular Input" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 70, + 152, + 538, + 183 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 152, + 538, + 183 + ], + "spans": [ + { + "bbox": [ + 70, + 152, + 538, + 183 + ], + "type": "text", + "content": "Zhen Lv" + }, + { + "bbox": [ + 70, + 152, + 538, + 183 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 70, + 152, + 538, + 183 + ], + "type": "text", + "content": ", Yangqi Long" + }, + { + "bbox": [ + 70, + 152, + 538, + 183 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 70, + 152, + 538, + 183 + ], + "type": "text", + "content": ", Congzhentao Huang" + }, + { + "bbox": [ + 70, + 152, + 538, + 183 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 70, + 152, + 538, + 183 + ], + "type": "text", + "content": ", Cao Li" + }, + { + "bbox": [ + 70, + 152, + 538, + 183 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 70, + 152, + 538, + 183 + ], + "type": "text", + "content": ", Chengfei Lv" + }, + { + "bbox": [ + 70, + 152, + 538, + 183 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 70, + 152, + 538, + 183 + ], + "type": "text", + "content": ", Hao Ren" + }, + { + "bbox": [ + 70, + 152, + 538, + 183 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 70, + 152, + 538, + 183 + ], + "type": "text", + "content": ", Dian Zheng" + }, + { + "bbox": [ + 70, + 152, + 538, + 183 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 70, + 152, + 538, + 183 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 70, + 152, + 538, + 183 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 70, + 152, + 538, + 183 + ], + "type": "text", + "content": "Alibaba Group; " + }, + { + "bbox": [ + 70, + 152, + 538, + 183 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 70, + 152, + 538, + 183 + ], + "type": "text", + "content": "Sun Yat-sen University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 151, + 219, + 200, + 232 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 219, + 200, + 232 + ], + "spans": [ + { + "bbox": [ + 151, + 219, + 200, + 232 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 244, + 297, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 244, + 297, + 555 + ], + "spans": [ + { + "bbox": [ + 54, + 244, + 297, + 555 + ], + "type": "text", + "content": "Stereo video synthesis from monocular input is challenging in spatial computing and virtual reality due to the lack of high-quality stereo video pairs for training and the difficulty of maintaining spatio-temporal consistency between frames. Existing methods primarily address these issues by directly applying novel view synthesis (NVS) techniques to video, while facing limitations such as the inability to effectively represent dynamic scenes and the requirement for extensive training data. In this paper, we introduce a novel self-supervised stereo video synthesis paradigm via a video diffusion model, termed SpatialDreamer, which meets the challenges head-on. Firstly, to address the stereo video data insufficiency, we propose a Depth based Video Generation module DVG, which employs a forward-backward rendering mechanism to generate paired videos with geometric and temporal priors. Leveraging data generated by DVG, we propose RefinerNet along with a self-supervised synthetic framework designed to facilitate efficient and dedicated training. More importantly, we devise a consistency control module, which consists of a metric of stereo deviation strength and a Temporal Interaction Learning module TIL for geometric and temporal consistency assurance respectively. We evaluated the proposed method against various benchmark methods, with the results showcasing its superior performance. Our project website is at: https://spatialdreamer.github.io." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 578, + 135, + 590 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 578, + 135, + 590 + ], + "spans": [ + { + "bbox": [ + 56, + 578, + 135, + 590 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 598, + 295, + 670 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 598, + 295, + 670 + ], + "spans": [ + { + "bbox": [ + 55, + 598, + 295, + 670 + ], + "type": "text", + "content": "Stereo video synthesis from a monocular input aims to generate the target-view video based on the given view with geometric and spatio-temporal consistency, which has a wide range of applications in 3D model reconstruction [6], 3D movie production [18, 40], and Apple Vision Pro (AVP) [44] like virtual reality content." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 670, + 295, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 670, + 295, + 694 + ], + "spans": [ + { + "bbox": [ + 55, + 670, + 295, + 694 + ], + "type": "text", + "content": "The primary difficulties of this task stem from the lack of adequate high-quality paired stereo videos for training, and" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 320, + 224, + 553, + 412 + ], + "blocks": [ + { + "bbox": [ + 320, + 224, + 553, + 412 + ], + "lines": [ + { + "bbox": [ + 320, + 224, + 553, + 412 + ], + "spans": [ + { + "bbox": [ + 320, + 224, + 553, + 412 + ], + "type": "image", + "image_path": "615a11c17cd91356c759af8756fb3f396788be871862426545a3a5a91989c04a.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 423, + 555, + 468 + ], + "lines": [ + { + "bbox": [ + 313, + 423, + 555, + 468 + ], + "spans": [ + { + "bbox": [ + 313, + 423, + 555, + 468 + ], + "type": "text", + "content": "Figure 1. Visualization of the temporal consistency in stereo video generated by different methods. We extract the yellow line of each frame and stack them together. A good result should show a natural transition in the " + }, + { + "bbox": [ + 313, + 423, + 555, + 468 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 423, + 555, + 468 + ], + "type": "text", + "content": " dimension." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 486, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 486, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 486, + 555, + 713 + ], + "type": "text", + "content": "the challenge of preserving the geometric consistency between two views as well as maintaining the temporal consistency across generated frames. Traditional methods for generating stereo content involve capturing scenes using dual-camera setups [52, 59, 70, 76]. However, the acquisition of these images involves the use of a professional-grade camera, resulting in substantial costs. Recent advancements in deep learning mainly handle these problems by directly applying single to stereo or multi-view image generation techniques to video, while often facing challenges related to geometric and spatio-temporal inconsistencies [21, 32, 41, 50, 66, 67]. As shown in Fig 1, methods based on NVs are challenging to maintain accurate spatio-temporal consistency, and even the recently released AVP 3D photo converter [2] may inevitably result in content flickering and inconsistency when applied to video synthesis. This is due to the difficulty of handling the temporal coherence between frames and the spatial information between paired views in complex dynamic scenes." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 70, + 702, + 149, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 702, + 149, + 712 + ], + "spans": [ + { + "bbox": [ + 70, + 702, + 149, + 712 + ], + "type": "text", + "content": "† Corresponding author." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "811" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 57, + 72, + 294, + 406 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 72, + 294, + 406 + ], + "spans": [ + { + "bbox": [ + 57, + 72, + 294, + 406 + ], + "type": "text", + "content": "In this paper, we propose a self-supervised stereo video synthesis paradigm via a video diffusion model, termed SpatialDreamer, which will meet the data insufficiency and spatio-temporal inconsistency at once. Firstly, we design a Depth based Video data Generation module DVG to handle the problem of data insufficiency. Similar to [21, 66], DVG constructs a training pair of two views using forward-backward rendering without the need for data annotation [51, 59, 70], while [21, 66] only render individual images, resulting in flickering effects in generated video [9]. In this paper, we improve the generation of video data for refining stereo occlusion masks by leveraging the inter-frame motion obtained from optical flow [29]. This methodology enables DVG to produce paired video data that maintains both geometric and temporal consistency. Leveraging paired video generated by DVG, we propose Refiner-Net, along with a self-supervised video synthesis framework, to enable efficient and targeted training. More importantly, with access to sufficient paired videos, we devise a consistency control module, which consists of a metric of stereo deviation strength and a Temporal Interaction Learning module TIL. The stereo deviation strength aims to enable the generation of stereo videos in diverse real-world scenarios, with a stereo-aware loss further supervising the model to learn the magnitude of difference between paired view features in the latent space. TIL integrates the latent features from long-temporal frames as a global information to augment the temporal coherence of the generated results." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 408, + 294, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 408, + 294, + 491 + ], + "spans": [ + { + "bbox": [ + 57, + 408, + 294, + 491 + ], + "type": "text", + "content": "Extensive experiments compared with various benchmark methods show that we achieve state-of-the-art performance. Notably, our SpatialDreamer meets the demand of real-world application without jitter, geometric and temporal inconsistency, which even beats AVP 3D converter and any open-source stereo video synthesis methods. We summarize our contributions as follows:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 493, + 293, + 632 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 55, + 493, + 293, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 493, + 293, + 550 + ], + "spans": [ + { + "bbox": [ + 55, + 493, + 293, + 550 + ], + "type": "text", + "content": "- A novel self-supervised stereo video synthesis framework, SpatialDreamer, is proposed, which is robust across a wide range of scenes and dynamic content in video. Moreover, equipping the DVG module, we build a stereo video dataset using a self-supervised approach." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 552, + 293, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 552, + 293, + 597 + ], + "spans": [ + { + "bbox": [ + 55, + 552, + 293, + 597 + ], + "type": "text", + "content": "- A consistency control module is devised, which consists of a metric of stereo deviation strength and a temporal interaction learning module, ensuring geometric and temporal consistency in video sequences." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 601, + 293, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 601, + 293, + 632 + ], + "spans": [ + { + "bbox": [ + 55, + 601, + 293, + 632 + ], + "type": "text", + "content": "- The sufficient experiments and quantitative results demonstrate that the proposed method outperforms the age-of-the-art methods, even beats AVP." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 58, + 650, + 141, + 661 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 650, + 141, + 661 + ], + "spans": [ + { + "bbox": [ + 58, + 650, + 141, + 661 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 58, + 670, + 239, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 670, + 239, + 681 + ], + "spans": [ + { + "bbox": [ + 58, + 670, + 239, + 681 + ], + "type": "text", + "content": "2.1. Single-image Novel View Synthesis" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 58, + 690, + 293, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 690, + 293, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 690, + 293, + 712 + ], + "type": "text", + "content": "Early studies address the single-image novel view synthesis task without relying on explicit 3D representation," + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 316, + 72, + 553, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 275 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 275 + ], + "type": "text", + "content": "such as End-to-end Synthesis [69], Infinite-Nature [37] and Animatable 3D Characters [68]. These methods lack explicit 3D representations for scenes, leading to inconsistent views and causing flickering and blurring, especially in occluded regions. Layer-based methods, such as 3D-photography [56] and SLIDE [27] represent a 3D scene using discrete layers. Yet, the synthesized views may lack accuracy in texture and geometric consistency across different perspectives. SMPI [61], AdaMPI [21] and SinMPI [47] represent the 3D scene using multiple planes, while their rendered results may suffer from depth discretization artifacts and repetitive texture patterns. To reduce the reliance on large labeled datasets, Shi et.al [55], Wang et.al. [66] and NVSVDE-Net [4] explore self-supervised learning for creating novel view images from single image. Although they produce photorealistic results, they struggle with temporal and spatial consistency, especially in dynamic scenes." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 316, + 283, + 553, + 341 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 283, + 553, + 341 + ], + "spans": [ + { + "bbox": [ + 316, + 283, + 553, + 341 + ], + "type": "text", + "content": "As for our approach, we design a depth-based video generation module that generates paired videos with geometric and temporal priors. Additionally, Our method ensures multi-view consistency for the entire scene by employing a consistency control module." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 316, + 371, + 489, + 382 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 371, + 489, + 382 + ], + "spans": [ + { + "bbox": [ + 316, + 371, + 489, + 382 + ], + "type": "text", + "content": "2.2. Multi-view Novel View Synthesis" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 316, + 396, + 553, + 670 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 396, + 553, + 670 + ], + "spans": [ + { + "bbox": [ + 316, + 396, + 553, + 670 + ], + "type": "text", + "content": "Early approaches in this field frequently utilized geometric methods, such as multi-view stereo reconstruction [30, 54] and structure-from-motion techniques [46, 53], to generate novel viewpoints from multiple images captured at different angles. NeRF-based methods were proposed to generate novel views based on fewer input images, including SinNeRF [72], CoPoNeRF [24] and NViST [28]. 3D Gaussian Splatting [32] presented an explicit representation of the scene. Gaussian-based methods, such as pixelSplat [12] and MVSplat [14], can produce high-quality synthesis results from multi-view images. Recently, researchers have begun exploring the application of advanced diffusion models for novel view synthesis, driven by their remarkable success in this area. Notable approaches such as Photoconsistent-NVS [75], MultiDiff [42], GCD [63] and NVS-Solver [74] generate image sequences based on predefined camera trajectories using either single or multiple input images. While these methods have shown promising results, they often struggle with dependence on dense scene geometry, handling complex scene dynamics, and generalization challenges [32]. Moreover, the challenge of optimizing NeRF/GS with sparse views hampers the ability of multi-view NVS to produce high-quality results." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 316, + 678, + 553, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 678, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 678, + 553, + 712 + ], + "type": "text", + "content": "In contrast, we explore a video diffusion model to achieve stereo video synthesis, while preserving geometric structures through a RefinerNet based framework." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 750, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 750, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 750, + 312, + 757 + ], + "type": "text", + "content": "812" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 72, + 76, + 539, + 276 + ], + "blocks": [ + { + "bbox": [ + 72, + 76, + 539, + 276 + ], + "lines": [ + { + "bbox": [ + 72, + 76, + 539, + 276 + ], + "spans": [ + { + "bbox": [ + 72, + 76, + 539, + 276 + ], + "type": "image", + "image_path": "4dcc85377b5a73a634d9a166cd85995df87c19ba07c84e4620eae5e7e486c4f6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 284, + 555, + 352 + ], + "lines": [ + { + "bbox": [ + 54, + 284, + 555, + 352 + ], + "spans": [ + { + "bbox": [ + 54, + 284, + 555, + 352 + ], + "type": "text", + "content": "Figure 2. Overview of the proposed method, given left view video, the target view video is rendered, encoded, and concatenated with multi-frame noise, followed by the denoising U-Net architecture (i.e., SVD). The reference view images are fed into RefinerNet, through which the spatial features are extracted. (Notably, the left and right views are the target and reference view for both training and inference). The temporal interaction learning module integrates the latent features from long-temporal frames, and the deviation strength is projected as positional embedding and added to the time step embedding. Finally, the variational autoencoder decoder decodes the result into a video clip." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 360, + 251, + 372 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 360, + 251, + 372 + ], + "spans": [ + { + "bbox": [ + 55, + 360, + 251, + 372 + ], + "type": "text", + "content": "2.3. Video Synthesis via Diffusion Models" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 378, + 295, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 378, + 295, + 605 + ], + "spans": [ + { + "bbox": [ + 56, + 378, + 295, + 605 + ], + "type": "text", + "content": "For video synthesis, previous works mainly expand image models to video, exploiting the image generation capability in the pretrained image diffusion models [38]. The video diffusion model (VDM) [23] and Video LDM [8] are proposed to adopt a 3D U-Net structure for joint image and video training. Subsequent works [15, 19, 19, 25, 57] also follow this insight with temporal attention, and [33, 48, 65, 71] further apply cross-frame attention to improve the consistency of the generated video. Another approach to promote temporal consistency is encoding video into 2D images [3, 45], which has been mainly explored in the video editing area. Although these models produce consistent results, they often require per-video optimization, which takes a long time, and would result in degradation when dealing with large motion. Most recently, [10, 20, 34] employ a transformer [64] architecture to model the temporal and spatial relations, demonstrating excellent video generation capabilities. However, these models need very large scale video datasets for training, which can be difficult to collect." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 606, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 606, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 606, + 295, + 714 + ], + "type": "text", + "content": "Beyond temporal information modeling, some works [13, 17, 67, 77] have introduced aligned spatial guidance to enhance the stability of the generated videos. The proposed framework also uses spatial guidance. Differing from those methods which require spatially aligned controls, a spatially unaligned reference image is used via a cross-attention mechanism. Furthermore, a metric of stereo deviation strength is introduced to strengthen the spatial constraint, resulting in more consistent results." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 314, + 359, + 411, + 373 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 359, + 411, + 373 + ], + "spans": [ + { + "bbox": [ + 314, + 359, + 411, + 373 + ], + "type": "text", + "content": "3. SpatialDreamer" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 379, + 555, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 379, + 555, + 523 + ], + "spans": [ + { + "bbox": [ + 313, + 379, + 555, + 523 + ], + "type": "text", + "content": "Given an image sequence and stereoscopic pose describing the scene depth information, the proposed method generates a stereo pair. In this section, we first provide a brief introduction about SVD [7]. Secondly, we offer detailed information of our video generation module DVG, which can generate sufficient paired video with geometric and temporal consistency. Using DVG, we present the framework for self-supervised learning-based video synthesis to support efficient and dedicated training. More importantly, to ensure consistency between frames, we design a consistency control module for geometric and temporal consistency assurance. Finally, we present the detailed training process." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 529, + 400, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 529, + 400, + 540 + ], + "spans": [ + { + "bbox": [ + 314, + 529, + 400, + 540 + ], + "type": "text", + "content": "3.1. Preliminaries" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 547, + 555, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 547, + 555, + 666 + ], + "spans": [ + { + "bbox": [ + 313, + 547, + 555, + 666 + ], + "type": "text", + "content": "The challenge in video generation lies in accurately modeling the spatial-temporal dynamics, involving spatial relationships within frames and temporal relationships across frames. To address this challenge, Stable Video Diffusion (SVD) [7] introduces temporal convolution and attention layers. SVD follows the latent diffusion model (LDM) to encode video pixels into the latent space, enabling a more efficient denoising process. Given a video " + }, + { + "bbox": [ + 313, + 547, + 555, + 666 + ], + "type": "inline_equation", + "content": "I_{v}" + }, + { + "bbox": [ + 313, + 547, + 555, + 666 + ], + "type": "text", + "content": ", encoder " + }, + { + "bbox": [ + 313, + 547, + 555, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 313, + 547, + 555, + 666 + ], + "type": "text", + "content": " encodes " + }, + { + "bbox": [ + 313, + 547, + 555, + 666 + ], + "type": "inline_equation", + "content": "I_{v}" + }, + { + "bbox": [ + 313, + 547, + 555, + 666 + ], + "type": "text", + "content": " into the latent space as " + }, + { + "bbox": [ + 313, + 547, + 555, + 666 + ], + "type": "inline_equation", + "content": "z_{0} = \\mathcal{E}(I_{v})" + }, + { + "bbox": [ + 313, + 547, + 555, + 666 + ], + "type": "text", + "content": ". Gaussian noise " + }, + { + "bbox": [ + 313, + 547, + 555, + 666 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 313, + 547, + 555, + 666 + ], + "type": "text", + "content": " is then added in the Markov process:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 380, + 671, + 553, + 685 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 380, + 671, + 553, + 685 + ], + "spans": [ + { + "bbox": [ + 380, + 671, + 553, + 685 + ], + "type": "interline_equation", + "content": "z _ {t} = \\sqrt {\\bar {\\alpha} _ {t}} z _ {0} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon_ {t} \\tag {1}", + "image_path": "f9ab01b44c62a8193550d132830ec54632eee471a4fd86f2c531272fcb07d305.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\bar{\\alpha}_0,\\dots ,\\bar{\\alpha}_T" + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": " are the pre-defined noise schedules for T steps in the Markov process. In particular, SVD is trained" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "813" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 109 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 109 + ], + "type": "text", + "content": "by the " + }, + { + "bbox": [ + 55, + 72, + 297, + 109 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 55, + 72, + 297, + 109 + ], + "type": "text", + "content": "-prediction formulation with the mean squared error (MSE) between the ground truth and its prediction in the latent space, known as " + }, + { + "bbox": [ + 55, + 72, + 297, + 109 + ], + "type": "inline_equation", + "content": "l_{\\epsilon}" + }, + { + "bbox": [ + 55, + 72, + 297, + 109 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 117, + 222, + 131 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 117, + 222, + 131 + ], + "spans": [ + { + "bbox": [ + 55, + 117, + 222, + 131 + ], + "type": "text", + "content": "3.2. Depth based Video Generation" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 136, + 295, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 136, + 295, + 327 + ], + "spans": [ + { + "bbox": [ + 54, + 136, + 295, + 327 + ], + "type": "text", + "content": "To create training pairs of images along with their corresponding occluded masks, the following self-supervised mechanism is employed utilizing real images. Firstly, the depth of the given video sample is estimated [49]. Secondly, the reference view image " + }, + { + "bbox": [ + 54, + 136, + 295, + 327 + ], + "type": "inline_equation", + "content": "x_{1}" + }, + { + "bbox": [ + 54, + 136, + 295, + 327 + ], + "type": "text", + "content": " is rendered into a masked one under the target viewpoint " + }, + { + "bbox": [ + 54, + 136, + 295, + 327 + ], + "type": "inline_equation", + "content": "P_{2}" + }, + { + "bbox": [ + 54, + 136, + 295, + 327 + ], + "type": "text", + "content": ", and then the trained inpainting model [58] is used to fill these occluded regions to obtain the novel viewpoint " + }, + { + "bbox": [ + 54, + 136, + 295, + 327 + ], + "type": "inline_equation", + "content": "x_{2}" + }, + { + "bbox": [ + 54, + 136, + 295, + 327 + ], + "type": "text", + "content": " [43]. Finally, " + }, + { + "bbox": [ + 54, + 136, + 295, + 327 + ], + "type": "inline_equation", + "content": "x_{2}" + }, + { + "bbox": [ + 54, + 136, + 295, + 327 + ], + "type": "text", + "content": " is rendered to the original viewpoint " + }, + { + "bbox": [ + 54, + 136, + 295, + 327 + ], + "type": "inline_equation", + "content": "P_{1}" + }, + { + "bbox": [ + 54, + 136, + 295, + 327 + ], + "type": "text", + "content": " via a forward rendering. After this, we can obtain the masked image " + }, + { + "bbox": [ + 54, + 136, + 295, + 327 + ], + "type": "inline_equation", + "content": "\\widetilde{x}_{1}" + }, + { + "bbox": [ + 54, + 136, + 295, + 327 + ], + "type": "text", + "content": " with mask " + }, + { + "bbox": [ + 54, + 136, + 295, + 327 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 54, + 136, + 295, + 327 + ], + "type": "text", + "content": " under viewpoint " + }, + { + "bbox": [ + 54, + 136, + 295, + 327 + ], + "type": "inline_equation", + "content": "P_{1}" + }, + { + "bbox": [ + 54, + 136, + 295, + 327 + ], + "type": "text", + "content": ", as well as the inpainted image " + }, + { + "bbox": [ + 54, + 136, + 295, + 327 + ], + "type": "inline_equation", + "content": "x_{2}" + }, + { + "bbox": [ + 54, + 136, + 295, + 327 + ], + "type": "text", + "content": " under viewpoint " + }, + { + "bbox": [ + 54, + 136, + 295, + 327 + ], + "type": "inline_equation", + "content": "P_{2}" + }, + { + "bbox": [ + 54, + 136, + 295, + 327 + ], + "type": "text", + "content": ". The mask " + }, + { + "bbox": [ + 54, + 136, + 295, + 327 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 54, + 136, + 295, + 327 + ], + "type": "text", + "content": " and masked image " + }, + { + "bbox": [ + 54, + 136, + 295, + 327 + ], + "type": "inline_equation", + "content": "\\widetilde{x}_{1}" + }, + { + "bbox": [ + 54, + 136, + 295, + 327 + ], + "type": "text", + "content": " are considered to be the conditions, with the image " + }, + { + "bbox": [ + 54, + 136, + 295, + 327 + ], + "type": "inline_equation", + "content": "x_{1}" + }, + { + "bbox": [ + 54, + 136, + 295, + 327 + ], + "type": "text", + "content": " serving as the ground truth for the diffusion model. This approach involves encoding the inpainted image " + }, + { + "bbox": [ + 54, + 136, + 295, + 327 + ], + "type": "inline_equation", + "content": "x_{2}" + }, + { + "bbox": [ + 54, + 136, + 295, + 327 + ], + "type": "text", + "content": " into latent feature representations and feeding it into RefinerNet." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 327, + 295, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 327, + 295, + 533 + ], + "spans": [ + { + "bbox": [ + 55, + 327, + 295, + 533 + ], + "type": "text", + "content": "However, the aforementioned approach encounters challenges in addressing inconsistencies among generated data frames, resulting in observable artifacts such as jittering and flickering. To tackle this issue, we leverage optical flow [60] to capture motion between successive video frames. Specifically, by leveraging the optical flow between the reference image " + }, + { + "bbox": [ + 55, + 327, + 295, + 533 + ], + "type": "inline_equation", + "content": "x_{1}^{t}" + }, + { + "bbox": [ + 55, + 327, + 295, + 533 + ], + "type": "text", + "content": " at time " + }, + { + "bbox": [ + 55, + 327, + 295, + 533 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 327, + 295, + 533 + ], + "type": "text", + "content": " and its adjacent frames, we establish the pixel correspondences between the image " + }, + { + "bbox": [ + 55, + 327, + 295, + 533 + ], + "type": "inline_equation", + "content": "x_{1}^{t}" + }, + { + "bbox": [ + 55, + 327, + 295, + 533 + ], + "type": "text", + "content": " and the images " + }, + { + "bbox": [ + 55, + 327, + 295, + 533 + ], + "type": "inline_equation", + "content": "x_{1}^{t-1}" + }, + { + "bbox": [ + 55, + 327, + 295, + 533 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 327, + 295, + 533 + ], + "type": "inline_equation", + "content": "x_{1}^{t+1}" + }, + { + "bbox": [ + 55, + 327, + 295, + 533 + ], + "type": "text", + "content": ", respectively. Particularly, these optical flows enable computing a confidence map via forward-backward consistency [16], which allows us to more accurately merge the stereo occlusions in the masks between adjacent frames [1]. Subsequently, the refined occlusion mask " + }, + { + "bbox": [ + 55, + 327, + 295, + 533 + ], + "type": "inline_equation", + "content": "m^{t}" + }, + { + "bbox": [ + 55, + 327, + 295, + 533 + ], + "type": "text", + "content": " is generated by sampling from the occlusion maps " + }, + { + "bbox": [ + 55, + 327, + 295, + 533 + ], + "type": "inline_equation", + "content": "m^{t-1}" + }, + { + "bbox": [ + 55, + 327, + 295, + 533 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 327, + 295, + 533 + ], + "type": "inline_equation", + "content": "m^{t+1}" + }, + { + "bbox": [ + 55, + 327, + 295, + 533 + ], + "type": "text", + "content": ", leveraging the pixel correspondences between consecutive frames under the reference view, as described by the following equation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 63, + 550, + 294, + 617 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 550, + 294, + 617 + ], + "spans": [ + { + "bbox": [ + 63, + 550, + 294, + 617 + ], + "type": "interline_equation", + "content": "m ^ {t} (i, j) = \\left\\{ \\begin{array}{l l} 1, & \\sum_ {k \\in \\{t - 1, t + 1 \\}} m ^ {k} (i + u, j + v) \\cdot \\\\ & C (i, j) \\geq 1 \\\\ m ^ {t} (i, j), & e l s e \\end{array} \\right. \\tag {2}", + "image_path": "48b4131b2783a041e0b166607b117131f380417cf8ab2bc308a693cd4bbfc753.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 618, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 295, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 618, + 295, + 715 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 55, + 618, + 295, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 618, + 295, + 715 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 55, + 618, + 295, + 715 + ], + "type": "text", + "content": " are the " + }, + { + "bbox": [ + 55, + 618, + 295, + 715 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 618, + 295, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 618, + 295, + 715 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 55, + 618, + 295, + 715 + ], + "type": "text", + "content": " components of the optical flow vector, and " + }, + { + "bbox": [ + 55, + 618, + 295, + 715 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 55, + 618, + 295, + 715 + ], + "type": "text", + "content": " represents the optical flow confidence measure. By propagating information from occluded areas across frames, this inter-frame continuity facilitates a more precise refinement of occluded regions in paired view videos and enforces temporal smoothness and coherence within the self-supervised framework, effectively mitigating the undesirable effects of jitter and flicker." + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 325, + 70, + 545, + 198 + ], + "blocks": [ + { + "bbox": [ + 325, + 70, + 545, + 198 + ], + "lines": [ + { + "bbox": [ + 325, + 70, + 545, + 198 + ], + "spans": [ + { + "bbox": [ + 325, + 70, + 545, + 198 + ], + "type": "image", + "image_path": "804bf090dcca49d6bf77fd03da14a6625ff8ef4378bf10a90929487f83332cff.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 205, + 555, + 239 + ], + "lines": [ + { + "bbox": [ + 313, + 205, + 555, + 239 + ], + "spans": [ + { + "bbox": [ + 313, + 205, + 555, + 239 + ], + "type": "text", + "content": "Figure 3. DVG. The temporal motion can be utilized to refine the occluded regions in the current frame, thereby providing smoother images and more consistent occlusion in the temporal sequence." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 269, + 555, + 318 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 269, + 555, + 318 + ], + "spans": [ + { + "bbox": [ + 313, + 269, + 555, + 318 + ], + "type": "text", + "content": "As shown in Figure 3, the difference between the training and inference procedures is that the former uses both backward and forward rendering whereas the latter uses only forward rendering to obtain an image under the target view." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 338, + 555, + 363 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 338, + 555, + 363 + ], + "spans": [ + { + "bbox": [ + 313, + 338, + 555, + 363 + ], + "type": "text", + "content": "3.3. Self-Supervised Stereo Video Synthesis framework" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 373, + 554, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 373, + 554, + 458 + ], + "spans": [ + { + "bbox": [ + 313, + 373, + 554, + 458 + ], + "type": "text", + "content": "Leveraging the generated paired view videos, we propose a self-supervised framework for stereo video synthesis, as illustrated in Figure 2. This framework commences with the rendering of the target view video that incorporates occlusions, followed by the application of a spatio-temporal method to effectively inpaint the missing regions, ultimately yielding a complete non-occluded target view video." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 462, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 462, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 462, + 555, + 715 + ], + "type": "text", + "content": "To learn the distribution of the feature space between the paired view images, we devise RefinerNet, which is modeled after the architecture of the denoising U-Net, excluding the temporal layer. RefinerNet adopts the weight initialization from the original SD2.1 model [50]. As depicted in Figure 2, a spatial attention layer is employed instead of the self-attention layer. Given a feature map " + }, + { + "bbox": [ + 313, + 462, + 555, + 715 + ], + "type": "inline_equation", + "content": "z_{t} \\in \\mathbb{R}^{t \\times h \\times w \\times c}" + }, + { + "bbox": [ + 313, + 462, + 555, + 715 + ], + "type": "text", + "content": " from the denoising U-Net and another feature map " + }, + { + "bbox": [ + 313, + 462, + 555, + 715 + ], + "type": "inline_equation", + "content": "z_{r} \\in \\mathbb{R}^{t \\times h \\times w \\times c}" + }, + { + "bbox": [ + 313, + 462, + 555, + 715 + ], + "type": "text", + "content": " from RefinerNet, they are concatenated along the h dimension, and then self-attention is applied. Subsequently, the first half of the feature map is retrieved as the output. One great advantage of this approach is that it enables RefinerNet to maintain the feature representation capability inherent in the original SD Unet, extracting reference view image information. In addition, the denoising U-Net can adaptively learn features from RefinerNet that are correlated in the same feature space, owing to the fundamentally similar network architectures between the denoising U-Net and RefinerNet. This not only ensures a robust initialization for the feature representation but also enhances the learning ability of the entire network." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "text", + "content": "814" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 212, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 212, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 212, + 85 + ], + "type": "text", + "content": "3.4. Consistency Control Module" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 90, + 256, + 102 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 90, + 256, + 102 + ], + "spans": [ + { + "bbox": [ + 55, + 90, + 256, + 102 + ], + "type": "text", + "content": "3.4.1 Temporal Interaction Learning module" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 108, + 295, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 108, + 295, + 167 + ], + "spans": [ + { + "bbox": [ + 55, + 108, + 295, + 167 + ], + "type": "text", + "content": "While our RefinerNet-based approach providing spatial guidance, paired view features are still merged independently, posing challenges for temporal consistency and potentially causing texture discrepancies or odd results between adjacent frames." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 168, + 296, + 253 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 168, + 296, + 253 + ], + "spans": [ + { + "bbox": [ + 55, + 168, + 296, + 253 + ], + "type": "text", + "content": "Inspired by previous studies [11, 22], we propose an attention-based temporal interaction learning module that learns the distribution of the feature space from adjacent frames, named TIL. Specifically, given latent features of " + }, + { + "bbox": [ + 55, + 168, + 296, + 253 + ], + "type": "inline_equation", + "content": "N_r" + }, + { + "bbox": [ + 55, + 168, + 296, + 253 + ], + "type": "text", + "content": " adjacent images under reference view " + }, + { + "bbox": [ + 55, + 168, + 296, + 253 + ], + "type": "inline_equation", + "content": "z_i^t" + }, + { + "bbox": [ + 55, + 168, + 296, + 253 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 168, + 296, + 253 + ], + "type": "inline_equation", + "content": "i = 1, 2, \\dots, N_r" + }, + { + "bbox": [ + 55, + 168, + 296, + 253 + ], + "type": "text", + "content": " and the latent feature of the reference view image " + }, + { + "bbox": [ + 55, + 168, + 296, + 253 + ], + "type": "inline_equation", + "content": "z_r^t" + }, + { + "bbox": [ + 55, + 168, + 296, + 253 + ], + "type": "text", + "content": " at the time step " + }, + { + "bbox": [ + 55, + 168, + 296, + 253 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 168, + 296, + 253 + ], + "type": "text", + "content": ", we augment the " + }, + { + "bbox": [ + 55, + 168, + 296, + 253 + ], + "type": "inline_equation", + "content": "z_r^t" + }, + { + "bbox": [ + 55, + 168, + 296, + 253 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 168, + 296, + 253 + ], + "type": "inline_equation", + "content": "aug_r^t" + }, + { + "bbox": [ + 55, + 168, + 296, + 253 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 78, + 262, + 295, + 294 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 262, + 295, + 294 + ], + "spans": [ + { + "bbox": [ + 78, + 262, + 295, + 294 + ], + "type": "interline_equation", + "content": "a u g _ {r} ^ {t} = \\lambda \\cdot \\operatorname {A t t n} _ {r, r} + (1 - \\lambda) \\cdot \\frac {1}{N _ {r}} \\sum_ {i = 1} ^ {N _ {r}} \\operatorname {A t t n} _ {r, i} \\tag {3}", + "image_path": "09853a4f04311b2acca7e56450c433b7c85af5d550d34bc2a0505c6c44818fd0.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 302, + 296, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 302, + 296, + 411 + ], + "spans": [ + { + "bbox": [ + 55, + 302, + 296, + 411 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 302, + 296, + 411 + ], + "type": "inline_equation", + "content": "\\lambda \\in [0,1]" + }, + { + "bbox": [ + 55, + 302, + 296, + 411 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 302, + 296, + 411 + ], + "type": "inline_equation", + "content": "\\mathrm{Attn}_{i,j}" + }, + { + "bbox": [ + 55, + 302, + 296, + 411 + ], + "type": "text", + "content": " defines attention between two latent features " + }, + { + "bbox": [ + 55, + 302, + 296, + 411 + ], + "type": "inline_equation", + "content": "z_{i}" + }, + { + "bbox": [ + 55, + 302, + 296, + 411 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 302, + 296, + 411 + ], + "type": "inline_equation", + "content": "z_{j}" + }, + { + "bbox": [ + 55, + 302, + 296, + 411 + ], + "type": "text", + "content": ". This module blends self-attention of " + }, + { + "bbox": [ + 55, + 302, + 296, + 411 + ], + "type": "inline_equation", + "content": "z_{r}^{t}" + }, + { + "bbox": [ + 55, + 302, + 296, + 411 + ], + "type": "text", + "content": " in RefinerNet's U-Net block with the cross-view attention between " + }, + { + "bbox": [ + 55, + 302, + 296, + 411 + ], + "type": "inline_equation", + "content": "z_{r}^{t}" + }, + { + "bbox": [ + 55, + 302, + 296, + 411 + ], + "type": "text", + "content": " and each adjacent view " + }, + { + "bbox": [ + 55, + 302, + 296, + 411 + ], + "type": "inline_equation", + "content": "z_{i}^{t}" + }, + { + "bbox": [ + 55, + 302, + 296, + 411 + ], + "type": "text", + "content": ". The cross-view attention aligns the feature of all adjacent frames to the reference view while the self-attention helps each reference image retain distinctiveness. The augmented reference feature " + }, + { + "bbox": [ + 55, + 302, + 296, + 411 + ], + "type": "inline_equation", + "content": "aug_{r}^{t}" + }, + { + "bbox": [ + 55, + 302, + 296, + 411 + ], + "type": "text", + "content": " is then fed into spatial attention layer to assist U-Net network learning." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 425, + 198, + 437 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 425, + 198, + 437 + ], + "spans": [ + { + "bbox": [ + 55, + 425, + 198, + 437 + ], + "type": "text", + "content": "3.4.2 Stereo Deviation Strength" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 443, + 296, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 443, + 296, + 574 + ], + "spans": [ + { + "bbox": [ + 55, + 443, + 296, + 574 + ], + "type": "text", + "content": "During the training, we noticed that different viewpoints impact the 3D impression of the scenes in the generated videos. However, relying solely on a given viewpoint as a guide for generating stereoscopic videos is insufficient. This is because videos with the same viewpoint may exhibit diverse depth ranges and variations, depending on the scene [70]. Therefore, pose alone cannot accurately control the 3D effect of a scene. Consequently, we introduce a metric called the stereo deviation strength, which quantitatively assesses the binocular disparity in a scene and facilitates the creation of controllable stereo vision:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 136, + 584, + 295, + 597 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 584, + 295, + 597 + ], + "spans": [ + { + "bbox": [ + 136, + 584, + 295, + 597 + ], + "type": "interline_equation", + "content": "s (z) = \\left| z _ {0} - z _ {\\text {r e f}} \\right| \\tag {4}", + "image_path": "02a80347daca72a9ad04864adc161a3be493c715e4c88d17e63f92ea9b11452e.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 605, + 295, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 689 + ], + "type": "text", + "content": "where the stereo deviation strength quantifies the latent differences between reference view " + }, + { + "bbox": [ + 55, + 605, + 295, + 689 + ], + "type": "inline_equation", + "content": "z_{ref}" + }, + { + "bbox": [ + 55, + 605, + 295, + 689 + ], + "type": "text", + "content": " and target view " + }, + { + "bbox": [ + 55, + 605, + 295, + 689 + ], + "type": "inline_equation", + "content": "z_0" + }, + { + "bbox": [ + 55, + 605, + 295, + 689 + ], + "type": "text", + "content": ". Similar to the time step, the deviation strength is projected into a positional embedding and added to each frame in the residual block to ensure uniform application of the deviation strength to every frame. The impact of stereo deviation strength guidance on the results is illustrated in Figure 4." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 690, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 690, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 690, + 296, + 714 + ], + "type": "text", + "content": "Moreover, for better convergence, a stereo-aware loss function is proposed to directly supervise the disparity dif" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 328, + 70, + 542, + 196 + ], + "blocks": [ + { + "bbox": [ + 328, + 70, + 542, + 196 + ], + "lines": [ + { + "bbox": [ + 328, + 70, + 542, + 196 + ], + "spans": [ + { + "bbox": [ + 328, + 70, + 542, + 196 + ], + "type": "image", + "image_path": "fc476e0ee20f138f5507c314034e7853356f1d077a6ae97128a8663bfe3f6e0f.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 205, + 555, + 250 + ], + "lines": [ + { + "bbox": [ + 313, + 205, + 555, + 250 + ], + "spans": [ + { + "bbox": [ + 313, + 205, + 555, + 250 + ], + "type": "text", + "content": "Figure 4. Stereo deviation strength guidance examples. Augmenting the deviation strength enhances the 3D photography effect, but excessive deviation strength may lead to image distortions or even lower the quality of stereoscopic image." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 270, + 350, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 270, + 350, + 281 + ], + "spans": [ + { + "bbox": [ + 314, + 270, + 350, + 281 + ], + "type": "text", + "content": "ference:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 386, + 281, + 553, + 295 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 386, + 281, + 553, + 295 + ], + "spans": [ + { + "bbox": [ + 386, + 281, + 553, + 295 + ], + "type": "interline_equation", + "content": "l _ {d} = \\left| \\left| s \\left(z _ {0}\\right) - s \\left(\\hat {z} _ {0}\\right) \\right| \\right| _ {2} ^ {2} \\tag {5}", + "image_path": "f040cb8deace8ee9550732b78691f7d1ab84c3e011b33f80f2feffa11d8af8bf.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 300, + 555, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 300, + 555, + 323 + ], + "spans": [ + { + "bbox": [ + 313, + 300, + 555, + 323 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 300, + 555, + 323 + ], + "type": "inline_equation", + "content": "\\hat{z}_0" + }, + { + "bbox": [ + 313, + 300, + 555, + 323 + ], + "type": "text", + "content": " represents the estimated clean video latent " + }, + { + "bbox": [ + 313, + 300, + 555, + 323 + ], + "type": "inline_equation", + "content": "z_0" + }, + { + "bbox": [ + 313, + 300, + 555, + 323 + ], + "type": "text", + "content": ", which can be obtained by:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 371, + 332, + 553, + 360 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 371, + 332, + 553, + 360 + ], + "spans": [ + { + "bbox": [ + 371, + 332, + 553, + 360 + ], + "type": "interline_equation", + "content": "\\hat {z} _ {0} = z _ {0} - \\frac {\\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon_ {\\theta} (z _ {t} , c , t)}{\\sqrt {\\bar {\\alpha} _ {t}}} \\tag {6}", + "image_path": "3da993401453a57d59382e9a6e37651ff46d5f0d7dd567bf24e26e287ab35050.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 368, + 554, + 393 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 368, + 554, + 393 + ], + "spans": [ + { + "bbox": [ + 313, + 368, + 554, + 393 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 368, + 554, + 393 + ], + "type": "inline_equation", + "content": "\\overline{\\alpha}_t = \\prod_{i=1}^t (1 - \\beta_t)" + }, + { + "bbox": [ + 313, + 368, + 554, + 393 + ], + "type": "text", + "content": ", and the noise prediction loss is combined with the stereo-aware loss by a scaling factor:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 404, + 402, + 553, + 414 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 404, + 402, + 553, + 414 + ], + "spans": [ + { + "bbox": [ + 404, + 402, + 553, + 414 + ], + "type": "interline_equation", + "content": "l = l _ {\\epsilon} + \\lambda \\cdot l _ {d} \\tag {7}", + "image_path": "19ad4cd7e7e8922c0f694f51c0261ecc1fc39dc3b087dc307d07ccd0e7b49136.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 314, + 430, + 420, + 444 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 430, + 420, + 444 + ], + "spans": [ + { + "bbox": [ + 314, + 430, + 420, + 444 + ], + "type": "text", + "content": "3.5. Training Strategy" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 449, + 555, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 449, + 555, + 641 + ], + "spans": [ + { + "bbox": [ + 313, + 449, + 555, + 641 + ], + "type": "text", + "content": "The training process is divided into two stages. In the first training stage, we focus on individual frames from videos. The temporal layer in denoising U-Net is frozen, and cross-view attention in TIL module is removed. The Refiner-Net model and the denoising U-Net are initialized with pretrained weights from SD2.1 [50] and SVD [7] respectively. The weights of the variational autoencoder encoder and decoder, as well as the contrastive language-image pretraining (CLIP) image encoder, are all kept fixed. The aim of this stage is for the model to learn to synthesize new viewpoint images under the condition of a reference image and a new viewpoint pose. In the second stage, we train the temporal layer and cross-view attention in TIL module with video sequences. This enables the model to capture temporal context information efficiently. The input for the model is the 8-frame video clip and " + }, + { + "bbox": [ + 313, + 449, + 555, + 641 + ], + "type": "inline_equation", + "content": "N_r" + }, + { + "bbox": [ + 313, + 449, + 555, + 641 + ], + "type": "text", + "content": " for TIL is set to 8." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 651, + 395, + 665 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 651, + 395, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 651, + 395, + 665 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 671, + 484, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 671, + 484, + 684 + ], + "spans": [ + { + "bbox": [ + 313, + 671, + 484, + 684 + ], + "type": "text", + "content": "4.1. Datasets, Baselines and Metrics" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "content": "We conduct our experiments on image and video levels with RealEstate10K [78] and self-collected video data." + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "815" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 57, + 72, + 294, + 286 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 72, + 294, + 286 + ], + "spans": [ + { + "bbox": [ + 57, + 72, + 294, + 286 + ], + "type": "text", + "content": "RealEstate10K is a large, widely used dataset with camera poses, corresponding to 10 million frames derived from about 80,000 video clips, gathered from about 10,000 YouTube videos. As there is not open-source stereo video benchmark, we collect 1500 monocular videos with 1400 for training, 100 for testing. For image synthesis, we compared the proposed method with recent open-source SOTA methods: 3D-photography [56], SynSim [69], AdaMPI [21], SinMPI [47], Wang et al. [66], Photoconsistent-NVS(P-NVS) [75], NVSVDE-Net [4], CoPoNeRF [24] and MVsplat [14]. For a fair comparison, the 3D photography method, AdaMPI, SinMPI, and the proposed method all used MiDaS [49] for depth estimation. Following the method of 3D-photography [56], we measured the accuracy of the generated target views with regard to the ground-truth images using the three metrics of the learned perceptual image patch similarity (LPIPS), peak signal-to-noise ratio (PSNR), and structural similarity index measure (SSIM)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 289, + 294, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 289, + 294, + 418 + ], + "spans": [ + { + "bbox": [ + 57, + 289, + 294, + 418 + ], + "type": "text", + "content": "For the video synthesis, we picked several representative approaches, i.e., the method of 3D-photography [56], the open-source code repository webui-depthmap, P-NVS [75], CoPoNeRF [24], NVS-Solver [74] and AVP [44]. Moreover, we conducted the same depth estimation method [49] for a fair comparison of 3D-photography [56], webui-depthmap and NVS-Solver [74]. We employ the FVD [62] score to measure the perceptual similarity between input videos and outputs and report the flow warping error " + }, + { + "bbox": [ + 57, + 289, + 294, + 418 + ], + "type": "inline_equation", + "content": "E_{\\text{warp}}" + }, + { + "bbox": [ + 57, + 289, + 294, + 418 + ], + "type": "text", + "content": " [35] to assess the temporal consistency of the resulting video sequences." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 428, + 152, + 440 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 428, + 152, + 440 + ], + "spans": [ + { + "bbox": [ + 57, + 428, + 152, + 440 + ], + "type": "text", + "content": "4.2. Implementation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 446, + 294, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 446, + 294, + 589 + ], + "spans": [ + { + "bbox": [ + 57, + 446, + 294, + 589 + ], + "type": "text", + "content": "We employed the AdamW [39] optimizer with a learning rate of " + }, + { + "bbox": [ + 57, + 446, + 294, + 589 + ], + "type": "inline_equation", + "content": "1 \\times 10^{-5}" + }, + { + "bbox": [ + 57, + 446, + 294, + 589 + ], + "type": "text", + "content": " to train the model on two NVIDIA A800 GPUs. In the first training stage, individual video frames were resized and center-cropped to a resolution of " + }, + { + "bbox": [ + 57, + 446, + 294, + 589 + ], + "type": "inline_equation", + "content": "1024 \\times 1024" + }, + { + "bbox": [ + 57, + 446, + 294, + 589 + ], + "type": "text", + "content": ". We trained the model for 100,000 iterations with a batch size of 8. In the second training stage, the videos were first split according to the transitions to ensure that a scene only appeared in one video clip. We then formed each video clip for training with eight frames. Finally, the temporal layer and TIL module was trained for 10,000 steps with a batch size of 2. The stereo-aware loss scaling factor " + }, + { + "bbox": [ + 57, + 446, + 294, + 589 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 57, + 446, + 294, + 589 + ], + "type": "text", + "content": " was set to 0.001 and " + }, + { + "bbox": [ + 57, + 446, + 294, + 589 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 57, + 446, + 294, + 589 + ], + "type": "text", + "content": " in Eq. 3 was set as 0.6." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 57, + 598, + 166, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 598, + 166, + 609 + ], + "spans": [ + { + "bbox": [ + 57, + 598, + 166, + 609 + ], + "type": "text", + "content": "4.3. Qualitative Results" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 57, + 617, + 186, + 628 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 617, + 186, + 628 + ], + "spans": [ + { + "bbox": [ + 57, + 617, + 186, + 628 + ], + "type": "text", + "content": "4.3.1 Stereo Image Synthesis" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 57, + 636, + 294, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 636, + 294, + 694 + ], + "spans": [ + { + "bbox": [ + 57, + 636, + 294, + 694 + ], + "type": "text", + "content": "We visually compared different methods for generating new view images from a target viewpoint. We qualitatively compared the newly synthesized images in Figure 5. 3D-photography fails to resolve the image distortion problem, and the image appears to be enhanced only slightly by the" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 316, + 72, + 553, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 191 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 191 + ], + "type": "text", + "content": "rendering. SynSin generates the most blurry details and loses a lot of information, compared with the input view. Neither AdaMPI nor SinMPI can generate fine enough details, and the results contain noticeable artifacts. CoPoN-erF tends to produce discrepancies from the original image in terms of fine-grained details, and it struggles with effectively handling scenes involving motion, as well as MVsplat. In contrast, the proposed method shows a clean and detailed output, especially in the area of depth information discontinuity." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 316, + 223, + 443, + 233 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 223, + 443, + 233 + ], + "spans": [ + { + "bbox": [ + 316, + 223, + 443, + 233 + ], + "type": "text", + "content": "4.3.2 Stereo Video Synthesis" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 316, + 246, + 553, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 246, + 553, + 460 + ], + "spans": [ + { + "bbox": [ + 316, + 246, + 553, + 460 + ], + "type": "text", + "content": "Moreover, we qualitatively compared the results of the stereo video synthesis with other methods. In addition to the quality issues of single-frame generation, the most important aspect in videos is the temporal consistency. We seek to generate videos that are free of flicker. The visualization of generated videos from different methods are illustrated in Figure 6, for each video, we show the scanline (the yellow line in figure) slice through the spatial-temporal volume. Our method achieve the smoothest trajectories, while other methods exhibit significant flickering. The most distinct regions are zoomed in and highlighted under the images to illustrate the details. In the first case, our method attains the smooth trajectories of the background (maintaining the straight line). And the second case, only our method preserves the texture of the trees behind the woman. And the last case, only our method preserves the texture of the black vertical lines. More results of generated frames are also presented in supplementary material." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 316, + 478, + 432, + 489 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 478, + 432, + 489 + ], + "spans": [ + { + "bbox": [ + 316, + 478, + 432, + 489 + ], + "type": "text", + "content": "4.4. Quantitative Results" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 316, + 498, + 553, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 498, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 498, + 553, + 712 + ], + "type": "text", + "content": "Table 1 lists the SSIM, PSNR, and LPIPS scores for the newly synthesized images. The proposed method outperforms 3D-photography and SynSin in the major metrics by a wide margin. The method of Wang et al., P-NVS and NVSVDE-Net perform slightly worse in terms of all metrics compared to the proposed method. Since the proposed method merges the spatial features and fills the occluded regions, it achieves the best LPIPS scores and shows competitive results in PSNR and SSIM when compared with AdaMPI and CoPoNeRF. MVSplat outperforms other methods when " + }, + { + "bbox": [ + 316, + 498, + 553, + 712 + ], + "type": "inline_equation", + "content": "t = 10" + }, + { + "bbox": [ + 316, + 498, + 553, + 712 + ], + "type": "text", + "content": " but is inferior to ours when " + }, + { + "bbox": [ + 316, + 498, + 553, + 712 + ], + "type": "inline_equation", + "content": "t = 5" + }, + { + "bbox": [ + 316, + 498, + 553, + 712 + ], + "type": "text", + "content": ". Overall, the results exhibit superior visual effects, which is consistent with the performance observed in the qualitative experiments. We further demonstrate the effectiveness of our method on video synthesis, as shown in Table 2. The proposed method achieves the best performance on both FVD and " + }, + { + "bbox": [ + 316, + 498, + 553, + 712 + ], + "type": "inline_equation", + "content": "E_{\\text{warp}}" + }, + { + "bbox": [ + 316, + 498, + 553, + 712 + ], + "type": "text", + "content": ", which means that our generated videos attain the best quality and temporal consistency." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 71, + 703, + 285, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 703, + 285, + 712 + ], + "spans": [ + { + "bbox": [ + 71, + 703, + 285, + 712 + ], + "type": "text", + "content": "https://github.com/thygate/stable-diffusion-webui-depthmap-script" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 298, + 750, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 750, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 750, + 313, + 757 + ], + "type": "text", + "content": "816" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 67, + 72, + 541, + 245 + ], + "blocks": [ + { + "bbox": [ + 67, + 72, + 541, + 245 + ], + "lines": [ + { + "bbox": [ + 67, + 72, + 541, + 245 + ], + "spans": [ + { + "bbox": [ + 67, + 72, + 541, + 245 + ], + "type": "image", + "image_path": "86c7f958e9f5db0238d862fe33cbd5bdb019e9307947a7de00988dc1e58aaee6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 247, + 555, + 270 + ], + "lines": [ + { + "bbox": [ + 55, + 247, + 555, + 270 + ], + "spans": [ + { + "bbox": [ + 55, + 247, + 555, + 270 + ], + "type": "text", + "content": "Figure 5. Qualitative results on the RealEstate10K dataset. The proposed method generates better-quality information and maintains the edge and textural details of the reference image better than the other methods. More results are available in supplementary material." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 61, + 281, + 534, + 427 + ], + "blocks": [ + { + "bbox": [ + 61, + 281, + 534, + 427 + ], + "lines": [ + { + "bbox": [ + 61, + 281, + 534, + 427 + ], + "spans": [ + { + "bbox": [ + 61, + 281, + 534, + 427 + ], + "type": "image", + "image_path": "d9b5162f3405508b0ac7cf01973d65d5bca984da229fd3cb8df89e4f02d3c9d7.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 431, + 555, + 454 + ], + "lines": [ + { + "bbox": [ + 55, + 431, + 555, + 454 + ], + "spans": [ + { + "bbox": [ + 55, + 431, + 555, + 454 + ], + "type": "text", + "content": "Figure 6. Visual comparisons in stereo video benchmark. We show the scanline (yellow line in original frame) slice through the spatial-temporal volume. The most distinct regions are zoomed in and highlighted under the images to illustrate the details." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 66, + 475, + 290, + 592 + ], + "blocks": [ + { + "bbox": [ + 66, + 475, + 290, + 592 + ], + "lines": [ + { + "bbox": [ + 66, + 475, + 290, + 592 + ], + "spans": [ + { + "bbox": [ + 66, + 475, + 290, + 592 + ], + "type": "image", + "image_path": "24b1891b54fa31e041b0e92c79ce76e931da71388d454afe23cefbdd6c07ea0d.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 600, + 295, + 643 + ], + "lines": [ + { + "bbox": [ + 55, + 600, + 295, + 643 + ], + "spans": [ + { + "bbox": [ + 55, + 600, + 295, + 643 + ], + "type": "text", + "content": "Figure 7. Ablation study of different designs. Only the proposed method ensures consistent preservation of details with the groundtruth target view. More results are available in supplementary material." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 661, + 149, + 673 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 661, + 149, + 673 + ], + "spans": [ + { + "bbox": [ + 55, + 661, + 149, + 673 + ], + "type": "text", + "content": "4.5. Ablation Study" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 679, + 230, + 691 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 679, + 230, + 691 + ], + "spans": [ + { + "bbox": [ + 55, + 679, + 230, + 691 + ], + "type": "text", + "content": "4.5.1 Effectiveness of each components" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 698, + 296, + 722 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 698, + 296, + 722 + ], + "spans": [ + { + "bbox": [ + 55, + 698, + 296, + 722 + ], + "type": "text", + "content": "We show the results in Table 3 on the RealEstate10K (stereo image) and collected stereo video datasets. CN means sub" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 475, + 555, + 690 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 475, + 555, + 690 + ], + "spans": [ + { + "bbox": [ + 313, + 475, + 555, + 690 + ], + "type": "text", + "content": "stituting our RefinerNet (RN) to ControlNet, SDS means the proposed stereo deviation strength. For image-level comparison, the results show that utilizing U-Net architecture only is not sufficient for stereo image synthesis and our RefinerNet is a better choice than ControlNet. Adding SDS further improves the geometric consistency. As visualized in Figure 7, solely relying on U-Net features results in distortion in the edge area and struggles with generating content properly. It also fails to preserve textural details near a plane boundary. Adding ControlNet often results in content inconsistency, presumably due to the failure of the paired view feature integration. RefinerNet handles this problem, showing it is a better architecture for stereo image generation. SDS further improves the consistency with the input and contain a more detailed image structure. As for video-level comparison, the quantitative results are shown in Table 3, adding TIL and DVG both attain the performance gain, and our final model achieves the best performance." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "817" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 134, + 305, + 297 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 297, + 126 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 297, + 126 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 297, + 126 + ], + "type": "text", + "content": "Table 1. Quantitative comparison on RealEstate10K dataset. " + }, + { + "bbox": [ + 55, + 70, + 297, + 126 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 55, + 70, + 297, + 126 + ], + "type": "text", + "content": " means re-implementing in our setting for fair comparison. When " + }, + { + "bbox": [ + 55, + 70, + 297, + 126 + ], + "type": "inline_equation", + "content": "t = 5" + }, + { + "bbox": [ + 55, + 70, + 297, + 126 + ], + "type": "text", + "content": ", the current frame serves as the left view and the frame at time +5 serves as the right view; the same approach applies for " + }, + { + "bbox": [ + 55, + 70, + 297, + 126 + ], + "type": "inline_equation", + "content": "t = 10" + }, + { + "bbox": [ + 55, + 70, + 297, + 126 + ], + "type": "text", + "content": ". The best results are in bold." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 134, + 305, + 297 + ], + "lines": [ + { + "bbox": [ + 56, + 134, + 305, + 297 + ], + "spans": [ + { + "bbox": [ + 56, + 134, + 305, + 297 + ], + "type": "table", + "html": "
SSIM ↑PSNR ↑LPIPS ↓
t=5t=10t=5t=10t=5t=10
3D-photography [56]†0.6410.44611.089.5330.1160.386
SynSin [69]†0.7400.64521.2219.550.0870.104
AdaMPI [21]†0.8690.73627.7022.650.0410.087
SinMPI [47]†0.8980.77528.7923.210.0400.080
Wang et al. [66]0.8400.72025.3521.360.0490.095
P-NVS [75]†0.7240.66721.9819.460.2410.279
NVSVDE-Net [4]0.84024.310.233
CoPoNeRF [24]†0.6580.63222.6321.330.1710.209
MVSplat [14]†0.86325.890.132
Proposed0.9160.85732.2624.860.0380.049
", + "image_path": "327cec0a2f4369c3c6adf54baaf4c9f4eed8f7cee3f56154f1cd6858f6753d9d.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 75, + 334, + 276, + 456 + ], + "blocks": [ + { + "bbox": [ + 55, + 304, + 296, + 327 + ], + "lines": [ + { + "bbox": [ + 55, + 304, + 296, + 327 + ], + "spans": [ + { + "bbox": [ + 55, + 304, + 296, + 327 + ], + "type": "text", + "content": "Table 2. Quantitative comparison with the other methods. The best results are in **bold**. " + }, + { + "bbox": [ + 55, + 304, + 296, + 327 + ], + "type": "inline_equation", + "content": "E_{\\text{warp}}^*" + }, + { + "bbox": [ + 55, + 304, + 296, + 327 + ], + "type": "text", + "content": " denotes " + }, + { + "bbox": [ + 55, + 304, + 296, + 327 + ], + "type": "inline_equation", + "content": "E_{\\text{warp}} \\left( \\times 10^{-3} \\right)" + }, + { + "bbox": [ + 55, + 304, + 296, + 327 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 75, + 334, + 276, + 456 + ], + "lines": [ + { + "bbox": [ + 75, + 334, + 276, + 456 + ], + "spans": [ + { + "bbox": [ + 75, + 334, + 276, + 456 + ], + "type": "table", + "html": "
FVD ↓E* warp ↓
3D-photography [56]155.03.418
Webui-depthmap83.183.486
P-NVS [75]420.223.02
CoPoNeRF [24]290.07.144
NVS-Solver [74]249.15.842
MVSplat [14]15437.913
AVP [44]99.923.446
Proposed67.093.374
", + "image_path": "830c44823b90b4fa6bc24ce946110991266aff61c09d36fefefd1f934b3b7a63.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 58, + 498, + 294, + 600 + ], + "blocks": [ + { + "bbox": [ + 55, + 468, + 296, + 491 + ], + "lines": [ + { + "bbox": [ + 55, + 468, + 296, + 491 + ], + "spans": [ + { + "bbox": [ + 55, + 468, + 296, + 491 + ], + "type": "text", + "content": "Table 3. Quantitative comparison for the ablation study. The best results are in bold." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 498, + 294, + 600 + ], + "lines": [ + { + "bbox": [ + 58, + 498, + 294, + 600 + ], + "spans": [ + { + "bbox": [ + 58, + 498, + 294, + 600 + ], + "type": "table", + "html": "
MethodCNRNSDSTILDVGStereo ImageStereo Video
SSIM↑PSNR↑LPIPS↓FVD↓
Image-level✓0.88023.730.06-
0.85524.040.183-
✓0.89530.200.043-
✓✓0.91632.260.038-
Video-level✓✓---184.0
✓✓✓----123.5
✓✓✓---85.21
✓✓✓✓---67.09
", + "image_path": "6029d5f9f732f184793bbfb64425fed1e2fcda89b05e4abeef4199e88b806b1c.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 621, + 206, + 634 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 621, + 206, + 634 + ], + "spans": [ + { + "bbox": [ + 55, + 621, + 206, + 634 + ], + "type": "text", + "content": "4.5.2 Data improvement by DVG" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 641, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 296, + 715 + ], + "type": "text", + "content": "We show the generated stereo occlusion mask with interframe motion refinement in Figure 8. It can be observed that the occlusions appear more complete spatially and exhibit smoother in their temporal distribution. By employing DVG, we can acquire sufficient data to support the generation of spatio-temporally consistent videos. More analyses" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 320, + 70, + 396, + 148 + ], + "blocks": [ + { + "bbox": [ + 320, + 70, + 396, + 148 + ], + "lines": [ + { + "bbox": [ + 320, + 70, + 396, + 148 + ], + "spans": [ + { + "bbox": [ + 320, + 70, + 396, + 148 + ], + "type": "image", + "image_path": "776aa8844cfaa92050a9122ee32602002cc6d873b8640d0800de504c952bd9b1.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 156, + 555, + 200 + ], + "lines": [ + { + "bbox": [ + 313, + 156, + 555, + 200 + ], + "spans": [ + { + "bbox": [ + 313, + 156, + 555, + 200 + ], + "type": "text", + "content": "Figure 8. Visual comparisons for the improvement of stereo occlusion. We show the scanline of (yellow line in original frame) slice through the spatial-temporal volume. The red region represents the changes of occlusion mask." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 396, + 70, + 472, + 148 + ], + "blocks": [ + { + "bbox": [ + 396, + 70, + 472, + 148 + ], + "lines": [ + { + "bbox": [ + 396, + 70, + 472, + 148 + ], + "spans": [ + { + "bbox": [ + 396, + 70, + 472, + 148 + ], + "type": "image", + "image_path": "10e1e42214e722de071131e3b2234a58c99e0ecccd585d0d8b581fe539a449b4.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 472, + 70, + 550, + 148 + ], + "blocks": [ + { + "bbox": [ + 472, + 70, + 550, + 148 + ], + "lines": [ + { + "bbox": [ + 472, + 70, + 550, + 148 + ], + "spans": [ + { + "bbox": [ + 472, + 70, + 550, + 148 + ], + "type": "image", + "image_path": "5666cf02f20695baf24a3ee0a7ab9d6399bce9c90c61d915097c0fe5325b141e.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 222, + 495, + 234 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 222, + 495, + 234 + ], + "spans": [ + { + "bbox": [ + 314, + 222, + 495, + 234 + ], + "type": "text", + "content": "are available in the supplementary materials." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 251, + 555, + 275 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 251, + 555, + 275 + ], + "spans": [ + { + "bbox": [ + 313, + 251, + 555, + 275 + ], + "type": "text", + "content": "4.5.3 Exploration of different depth estimation methods" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 282, + 555, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 282, + 555, + 390 + ], + "spans": [ + { + "bbox": [ + 313, + 282, + 555, + 390 + ], + "type": "text", + "content": "We further compared different design choices of depth map including DepthAnything [73], Marigold [31], MiDaS [49], ZoeDepth [5] and DepthCrafter [26]. The results are available in supplementary materials. The proposed method can produce new viewpoint images that are realistic, seamless, and rich in detail, regardless of the depth estimation method used. This capability demonstrates the versatility and robustness of the proposed method, ensuring superior performance in various applications and environments." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 314, + 399, + 391, + 411 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 399, + 391, + 411 + ], + "spans": [ + { + "bbox": [ + 314, + 399, + 391, + 411 + ], + "type": "text", + "content": "4.6. Limitations" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 418, + 554, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 418, + 554, + 491 + ], + "spans": [ + { + "bbox": [ + 313, + 418, + 554, + 491 + ], + "type": "text", + "content": "SpatialDreamer effectively creates stereo video from a single input but faces limitations. It relies on depth maps for rendering, affecting stereoscopic quality due to depth accuracy and distribution. Future research could focus on integrating implicit depth to address this [36]. Additionally, the model's large parameter size limits real-time performance." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 314, + 502, + 388, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 502, + 388, + 514 + ], + "spans": [ + { + "bbox": [ + 314, + 502, + 388, + 514 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 522, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 522, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 522, + 555, + 714 + ], + "type": "text", + "content": "In this paper, we introduce a new self-supervised stereo video synthesis approach using a video diffusion model, termed SpatialDreamer. This method addresses the issues of insufficient data and spatio-temporal inconsistency between frames. To solve the problem of insufficient data, we develop a Depth-based Video Generation (DVG) module that uses a forward-backward rendering mechanism to generate a rendered video with geometric and temporal priors. Additionally, we propose RefinerNet along with a self-supervised synthetic framework to enable efficient and dedicated training using data generated by DVG. Furthermore, we design a consistency control module to ensure geometric and temporal consistency. Our SpatialDreamer outperforms all other open-source stereo image and video synthesis methods and has the potential for future expansion into virtual reality applications." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "text", + "content": "818" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "type": "text", + "content": "[1] Dmitry Akimov, Alexey Shestov, Alexander Voronov, and Dmitriy Vatolin. Occlusion refinement for stereo video using optical flow. In 2012 International Conference on 3D Imaging (IC3D), pages 1-8, 2012." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 137, + 296, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 137, + 296, + 192 + ], + "spans": [ + { + "bbox": [ + 61, + 137, + 296, + 192 + ], + "type": "text", + "content": "[2] Andrea Schubert. visionos 2 brings new spatial computing experiences to apple vision pro. https://www.apple.com/newsroom/2024/06/visionos-2-brings-new-spatial-computing-experiences-to-apple-vision-pro, 2024." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 194, + 295, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 194, + 295, + 237 + ], + "spans": [ + { + "bbox": [ + 62, + 194, + 295, + 237 + ], + "type": "text", + "content": "[3] Omer Bar-Tal, Dolev Ofri-Amar, Rafail Fridman, Yoni Kasten, and Tali Dekel. Text2live: Text-driven layered image and video editing. In European conference on computer vision, pages 707-723. Springer, 2022." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 239, + 295, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 239, + 295, + 282 + ], + "spans": [ + { + "bbox": [ + 62, + 239, + 295, + 282 + ], + "type": "text", + "content": "[4] Juan Luis Gonzalez Bello and Munchurl Kim. Novel view synthesis with view-dependent effects from a single image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 284, + 294, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 284, + 294, + 317 + ], + "spans": [ + { + "bbox": [ + 62, + 284, + 294, + 317 + ], + "type": "text", + "content": "[5] Shariq Farooq Bhat, Reiner Birkl, Diana Wofk, Peter Wonka, and Matthias Müller. Zoedepth: Zero-shot transfer by combining relative and metric depth, 2023." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 319, + 295, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 319, + 295, + 373 + ], + "spans": [ + { + "bbox": [ + 62, + 319, + 295, + 373 + ], + "type": "text", + "content": "[6] Phuong Ngoc Binh Do and Quoc Chi Nguyen. A review of stereo-photogrammetry method for 3-d reconstruction in computer vision. In 2019 19th International Symposium on Communications and Information Technologies (ISCIT), pages 138–143, 2019." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 375, + 295, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 375, + 295, + 430 + ], + "spans": [ + { + "bbox": [ + 62, + 375, + 295, + 430 + ], + "type": "text", + "content": "[7] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 431, + 295, + 496 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 431, + 295, + 496 + ], + "spans": [ + { + "bbox": [ + 62, + 431, + 295, + 496 + ], + "type": "text", + "content": "[8] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22563-22575, 2023." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 498, + 295, + 530 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 498, + 295, + 530 + ], + "spans": [ + { + "bbox": [ + 62, + 498, + 295, + 530 + ], + "type": "text", + "content": "[9] Aaron Bobick and Stephen Intille. Large occlusion stereo. International Journal of Computer Vision, 33:181-200, 1999." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 533, + 295, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 533, + 295, + 586 + ], + "spans": [ + { + "bbox": [ + 57, + 533, + 295, + 586 + ], + "type": "text", + "content": "[10] Tim Brooks, Bill Peebles, Connor Holmes, Will DePue, Yufei Guo, Li Jing, David Schnurr, Joe Taylor, Troy Luhman, Eric Luhman, Clarence Ng, Ricky Wang, and Aditya Ramesh. Video generation models as world simulators. 2024." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 590, + 295, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 590, + 295, + 633 + ], + "spans": [ + { + "bbox": [ + 57, + 590, + 295, + 633 + ], + "type": "text", + "content": "[11] Mingdeng Cao, Xintao Wang, Zhongang Qi, Ying Shan, Xiaohu Qie, and Yinqiang Zheng. Masactrl: Tuning-free mutual self-attention control for consistent image synthesis and editing, 2023." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 635, + 295, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 635, + 295, + 667 + ], + "spans": [ + { + "bbox": [ + 57, + 635, + 295, + 667 + ], + "type": "text", + "content": "[12] David Charatan, Sizhe Li, Andrea Tagliasacchi, and Vincent Sitzmann. pixelsplat: 3d gaussian splats from image pairs for scalable generalizable 3d reconstruction. In arXiv, 2023." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "type": "text", + "content": "[13] Weifeng Chen, Jie Wu, Pan Xie, Hefeng Wu, Jiashi Li, Xin Xia, Xuefeng Xiao, and Liang Lin. Control-a-video: Controllable text-to-video generation with diffusion models. arXiv preprint arXiv:2305.13840, 2023." + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 117 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 117 + ], + "type": "text", + "content": "[14] Yuedong Chen, Haofei Xu, Chuanxia Zheng, Bohan Zhuang, Marc Pollefeys, Andreas Geiger, Tat-Jen Cham, and Jianfei Cai. Mvsplat: Efficient 3d gaussian splatting from sparse multi-view images. arXiv preprint arXiv:2403.14627, 2024." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 119, + 553, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 119, + 553, + 162 + ], + "spans": [ + { + "bbox": [ + 317, + 119, + 553, + 162 + ], + "type": "text", + "content": "[15] Zuozhuo Dai, Zhenghao Zhang, Yao Yao, Bingxue Qiu, Siyu Zhu, Long Qin, and Weizhi Wang. Animateanything: Fine-grained open domain image animation with motion guidance. arXiv e-prints, pages arXiv-2311, 2023." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 164, + 553, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 164, + 553, + 206 + ], + "spans": [ + { + "bbox": [ + 317, + 164, + 553, + 206 + ], + "type": "text", + "content": "[16] G. Egnal and R.P. Wildes. Detecting binocular half-occlusions: empirical comparisons of five approaches. IEEE Transactions on Pattern Analysis and Machine Intelligence, 24(8):1127-1133, 2002." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 209, + 553, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 209, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 317, + 209, + 553, + 262 + ], + "type": "text", + "content": "[17] Patrick Esser, Johnathan Chiu, Parmida Atighehchian, Jonathan Granskog, and Anastasis Germanidis. Structure and content-guided video synthesis with diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7346-7356, 2023." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 264, + 553, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 264, + 553, + 296 + ], + "spans": [ + { + "bbox": [ + 316, + 264, + 553, + 296 + ], + "type": "text", + "content": "[18] Chen Gao, Ayush Saraf, Johannes Kopf, and Jia-Bin Huang. Dynamic view synthesis from dynamic monocular video, 2021." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 299, + 553, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 299, + 553, + 353 + ], + "spans": [ + { + "bbox": [ + 317, + 299, + 553, + 353 + ], + "type": "text", + "content": "[19] Yuwei Guo, Ceyuan Yang, Anyi Rao, Zhengyang Liang, Yaohui Wang, Yu Qiao, Maneesh Agrawala, Dahua Lin, and Bo Dai. Animatediff: Imagine your personalized text-to-image diffusion models without specific tuning. International Conference on Learning Representations, 2024." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 354, + 553, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 354, + 553, + 398 + ], + "spans": [ + { + "bbox": [ + 317, + 354, + 553, + 398 + ], + "type": "text", + "content": "[20] Agrim Gupta, Lijun Yu, Kihyuk Sohn, Xiuye Gu, Meera Hahn, Li Fei-Fei, Irfan Essa, Lu Jiang, and José Lezama. Photorealistic video generation with diffusion models. arXiv preprint arXiv:2312.06662, 2023." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 399, + 553, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 399, + 553, + 443 + ], + "spans": [ + { + "bbox": [ + 317, + 399, + 553, + 443 + ], + "type": "text", + "content": "[21] Yuxuan Han, Ruicheng Wang, and Jiaolong Yang. Single-view view synthesis in the wild with learned adaptive multiplane images. In ACM SIGGRAPH 2022 Conference Proceedings, pages 1-8, 2022." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 445, + 553, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 445, + 553, + 476 + ], + "spans": [ + { + "bbox": [ + 317, + 445, + 553, + 476 + ], + "type": "text", + "content": "[22] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control, 2022." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 478, + 553, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 478, + 553, + 521 + ], + "spans": [ + { + "bbox": [ + 317, + 478, + 553, + 521 + ], + "type": "text", + "content": "[23] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2022." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 523, + 553, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 523, + 553, + 567 + ], + "spans": [ + { + "bbox": [ + 317, + 523, + 553, + 567 + ], + "type": "text", + "content": "[24] Sunghwan Hong, Jaewoo Jung, Heeseong Shin, Jiaolong Yang, Seungryong Kim, and Chong Luo. Unifying correspondence, pose and nerf for pose-free novel view synthesis from stereo pairs, 2024." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 568, + 553, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 568, + 553, + 612 + ], + "spans": [ + { + "bbox": [ + 317, + 568, + 553, + 612 + ], + "type": "text", + "content": "[25] Li Hu, Xin Gao, Peng Zhang, Ke Sun, Bang Zhang, and Liefeng Bo. animate anyone: Consistent and controllable image-to-video synthesis for character animation. arXiv preprint arXiv:2311.17117, 2023." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 614, + 553, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 614, + 553, + 657 + ], + "spans": [ + { + "bbox": [ + 317, + 614, + 553, + 657 + ], + "type": "text", + "content": "[26] Wenbo Hu, Xiangjun Gao, Xiaoyu Li, Sijie Zhao, Xiaodong Cun, Yong Zhang, Long Quan, and Ying Shan. Depthcrafter: Generating consistent long depth sequences for open-world videos. arXiv preprint arXiv:2409.02095, 2024." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 658, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 658, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 658, + 553, + 713 + ], + "type": "text", + "content": "[27] Varun Jampani, Huiwen Chang, Kyle Sargent, Abhishek Kar, Richard Tucker, Michael Krainin, Dominik Kaeser, William T Freeman, David Salesin, Brian Curless, et al. Slide: Single image 3d photography with soft layering and depth-aware inpainting. In Proceedings of the IEEE/CVF In" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "text", + "content": "819" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 77, + 72, + 295, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 72, + 295, + 95 + ], + "spans": [ + { + "bbox": [ + 77, + 72, + 295, + 95 + ], + "type": "text", + "content": "ternational Conference on Computer Vision, pages 12518-12527, 2021." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 96, + 295, + 118 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 96, + 295, + 118 + ], + "spans": [ + { + "bbox": [ + 57, + 96, + 295, + 118 + ], + "type": "text", + "content": "[28] Wonbong Jang and Lourdes Agapito. Nvist: In the wild new view synthesis from a single image with transformers, 2024." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 119, + 295, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 119, + 295, + 163 + ], + "spans": [ + { + "bbox": [ + 56, + 119, + 295, + 163 + ], + "type": "text", + "content": "[29] Jisoo Jeong, Hong Cai, Risheek Garrepalli, Jamie Menjay Lin, Munawar Hayat, and Fatih Porikli. Ocai: Improving optical flow estimation by occlusion and consistency aware interpolation, 2024." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 164, + 295, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 164, + 295, + 207 + ], + "spans": [ + { + "bbox": [ + 56, + 164, + 295, + 207 + ], + "type": "text", + "content": "[30] Hailin Jin, Stefano Soatto, and Anthony J Yezzi. Multi-view stereo reconstruction of dense shape and complex appearance. International Journal of Computer Vision, 63:175-189, 2005." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 209, + 295, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 209, + 295, + 252 + ], + "spans": [ + { + "bbox": [ + 56, + 209, + 295, + 252 + ], + "type": "text", + "content": "[31] Bingxin Ke, Anton Obukhov, Shengyu Huang, Nando Metzger, Rodrigo Caye Daudt, and Konrad Schindler. Repurposing diffusion-based image generators for monocular depth estimation, 2024." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 254, + 295, + 287 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 254, + 295, + 287 + ], + "spans": [ + { + "bbox": [ + 56, + 254, + 295, + 287 + ], + "type": "text", + "content": "[32] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering, 2023." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 289, + 295, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 289, + 295, + 342 + ], + "spans": [ + { + "bbox": [ + 56, + 289, + 295, + 342 + ], + "type": "text", + "content": "[33] Levon Khachatriyan, Andranik Movsisyan, Vahram Tadevosyan, Roberto Henschel, Zhangyang Wang, Shant Navasardyan, and Humphrey Shi. Text2video-zero: Text-to-image diffusion models are zero-shot video generators. arXiv preprint arXiv:2303.13439, 2023." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 344, + 295, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 344, + 295, + 399 + ], + "spans": [ + { + "bbox": [ + 56, + 344, + 295, + 399 + ], + "type": "text", + "content": "[34] Dan Kondratyuk, Lijun Yu, Xiuye Gu, José Lezama, Jonathan Huang, Rachel Hornung, Hartwig Adam, Hassan Akbari, Yair Alon, Vighnesh Birodkar, et al. Videopoet: A large language model for zero-shot video generation. arXiv preprint arXiv:2312.14125, 2023." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 399, + 295, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 399, + 295, + 443 + ], + "spans": [ + { + "bbox": [ + 56, + 399, + 295, + 443 + ], + "type": "text", + "content": "[35] Wei-Sheng Lai, Jia-Bin Huang, Oliver Wang, Eli Shechtman, Ersin Yumer, and Ming-Hsuan Yang. Learning blind video temporal consistency. In Proceedings of the European conference on computer vision (ECCV), pages 170-185, 2018." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 445, + 295, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 445, + 295, + 498 + ], + "spans": [ + { + "bbox": [ + 56, + 445, + 295, + 498 + ], + "type": "text", + "content": "[36] Hanwen Liang, Junli Cao, Vedit Goel, Guocheng Qian, Sergei Korolev, Demetri Terzopoulos, Konstantinos Plataniotis, Sergey Tulyakov, and Jian Ren. Wonderland: Navigating 3d scenes from a single image. arXiv preprint arXiv:2412.12091, 2024." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 500, + 295, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 500, + 295, + 555 + ], + "spans": [ + { + "bbox": [ + 56, + 500, + 295, + 555 + ], + "type": "text", + "content": "[37] Andrew Liu, Richard Tucker, Varun Jampani, Ameesh Makadia, Noah Snavely, and Angjoo Kanazawa. Infinite nature: Perpetual view generation of natural scenes from a single image. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 556, + 295, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 556, + 295, + 610 + ], + "spans": [ + { + "bbox": [ + 56, + 556, + 295, + 610 + ], + "type": "text", + "content": "[38] Yixin Liu, Kai Zhang, Yuan Li, Zhiling Yan, Chujie Gao, Ruoxi Chen, Zhengqing Yuan, Yue Huang, Hanchi Sun, Jianfeng Gao, Lifang He, and Lichao Sun. Sora: A review on background, technology, limitations, and opportunities of large vision models, 2024." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 612, + 295, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 612, + 295, + 634 + ], + "spans": [ + { + "bbox": [ + 56, + 612, + 295, + 634 + ], + "type": "text", + "content": "[39] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 635, + 295, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 295, + 669 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 295, + 669 + ], + "type": "text", + "content": "[40] Kevin Matzen, Michael F. Cohen, Bryce Evans, Johannes Kopf, and Richard Szeliski. Low-cost 360 stereo photography and video capture. ACM Trans. Graph., 36(4), 2017." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 670, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 295, + 712 + ], + "type": "text", + "content": "[41] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis, 2020." + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 555, + 714 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 316, + 72, + 554, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 554, + 138 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 554, + 138 + ], + "type": "text", + "content": "[42] Norman Müller, Katja Schwarz, Barbara Rössle, Lorenzo Porzi, Samuel Rota Bulò, Matthias Nießner, and Peter Kontschieder. Multidiff: Consistent novel view synthesis from a single image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10258-10268, 2024." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 140, + 553, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 140, + 553, + 161 + ], + "spans": [ + { + "bbox": [ + 316, + 140, + 553, + 161 + ], + "type": "text", + "content": "[43] Simon Niklaus and Feng Liu. Softmax splatting for video frame interpolation, 2020." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 163, + 553, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 553, + 207 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 553, + 207 + ], + "type": "text", + "content": "[44] Onee. What is spatial video on iphone 15 pro and vision pro. https://xreality-zone/en/posts/what-is-spatial-video-on-iphone-15-pro-and/apple-vision-pro, 2024." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 209, + 553, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 209, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 316, + 209, + 553, + 262 + ], + "type": "text", + "content": "[45] Hao Ouyang, Qiuyu Wang, Yuxi Xiao, Qingyan Bai, Juntao Zhang, Kecheng Zheng, Xiaowei Zhou, Qifeng Chen, and Yujun Shen. Codef: Content deformation fields for temporally consistent video processing. arXiv preprint arXiv:2308.07926, 2023." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 263, + 553, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 263, + 553, + 296 + ], + "spans": [ + { + "bbox": [ + 316, + 263, + 553, + 296 + ], + "type": "text", + "content": "[46] Onur Özyeşil, Vladislav Voroninski, Ronen Basri, and Amit Singer. A survey of structure from motion*. Acta Numerica, 26:305-364, 2017." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 298, + 553, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 298, + 553, + 331 + ], + "spans": [ + { + "bbox": [ + 316, + 298, + 553, + 331 + ], + "type": "text", + "content": "[47] Guo Pu, Peng-Shuai Wang, and Zhouhui Lian. Sinmpi: Novel view synthesis from a single image with expanded multiplane images, 2023." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 332, + 553, + 375 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 332, + 553, + 375 + ], + "spans": [ + { + "bbox": [ + 316, + 332, + 553, + 375 + ], + "type": "text", + "content": "[48] Chenyang Qi, Xiaodong Cun, Yong Zhang, Chenyang Lei, Xintao Wang, Ying Shan, and Qifeng Chen. Fatezero: Fusing attentions for zero-shot text-based video editing. arXiv preprint arXiv:2303.09535, 2023." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 376, + 553, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 376, + 553, + 431 + ], + "spans": [ + { + "bbox": [ + 316, + 376, + 553, + 431 + ], + "type": "text", + "content": "[49] René Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(3), 2022." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 433, + 555, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 433, + 555, + 487 + ], + "spans": [ + { + "bbox": [ + 316, + 433, + 555, + 487 + ], + "type": "text", + "content": "[50] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 488, + 553, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 488, + 553, + 521 + ], + "spans": [ + { + "bbox": [ + 316, + 488, + 553, + 521 + ], + "type": "text", + "content": "[51] Christoph Sager, Christian Janiesch, and Patrick Zschech. A survey of image labelling for computer vision applications. Journal of Business Analytics, 4(2):91-110, 2021." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 522, + 553, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 522, + 553, + 544 + ], + "spans": [ + { + "bbox": [ + 316, + 522, + 553, + 544 + ], + "type": "text", + "content": "[52] D. Scharstein. View Synthesis Using Stereo Vision. Springer Berlin Heidelberg, 1999." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 545, + 553, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 545, + 553, + 589 + ], + "spans": [ + { + "bbox": [ + 316, + 545, + 553, + 589 + ], + "type": "text", + "content": "[53] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4104-4113, 2016." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 590, + 553, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 590, + 553, + 645 + ], + "spans": [ + { + "bbox": [ + 316, + 590, + 553, + 645 + ], + "type": "text", + "content": "[54] S.M. Seitz, B. Curless, J. Diebel, D. Scharstein, and R. Szeliski. A comparison and evaluation of multi-view stereo reconstruction algorithms. In 2006 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR'06), pages 519-528, 2006." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "type": "text", + "content": "[55] Yujiao Shi, Hongdong Li, and Xin Yu. Self-supervised visibility learning for novel view synthesis.*. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2021." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 691, + 553, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 714 + ], + "type": "text", + "content": "[56] Meng-Li Shih, Shih-Yang Su, Johannes Kopf, and Jia-Bin Huang. 3d photography using context-aware layered depth" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "text", + "content": "820" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 76, + 72, + 296, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 296, + 105 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 296, + 105 + ], + "type": "text", + "content": "inpainting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8028-8038, 2020." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 106, + 296, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 106, + 296, + 160 + ], + "spans": [ + { + "bbox": [ + 56, + 106, + 296, + 160 + ], + "type": "text", + "content": "[57] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792, 2022." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 161, + 295, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 161, + 295, + 183 + ], + "spans": [ + { + "bbox": [ + 56, + 161, + 295, + 183 + ], + "type": "text", + "content": "[58] Lorenzo Stacchio. Train stable diffusion for inpainting, 2023." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 184, + 295, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 184, + 295, + 228 + ], + "spans": [ + { + "bbox": [ + 56, + 184, + 295, + 228 + ], + "type": "text", + "content": "[59] Wenxiu Sun, Lingfeng Xu, Oscar C Au, Sung Him Chui, and Chun Wing Kwok. An overview of free view-point depth-image-based rendering (dibr). In APSIPA Annual Summit and Conference, pages 1023-1030, 2010." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 228, + 295, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 228, + 295, + 250 + ], + "spans": [ + { + "bbox": [ + 56, + 228, + 295, + 250 + ], + "type": "text", + "content": "[60] Zachary Teed and Jia Deng. Raft: Recurrent all-pairs field transforms for optical flow, 2020." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 251, + 295, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 251, + 295, + 294 + ], + "spans": [ + { + "bbox": [ + 56, + 251, + 295, + 294 + ], + "type": "text", + "content": "[61] Richard Tucker and Noah Snavely. Single-view view synthesis with multiplane images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 551-560, 2020." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 294, + 295, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 294, + 295, + 360 + ], + "spans": [ + { + "bbox": [ + 56, + 294, + 295, + 360 + ], + "type": "text", + "content": "[62] Thomas Unterthiner, Sjoerd van Steenkiste, Karol Kurach, Raphaël Marinier, Marcin Michalski, and Sylvain Gelly. FVD: A new metric for video generation. In Deep Generative Models for Highly Structured Data, ICLR 2019 Workshop, New Orleans, Louisiana, United States, May 6, 2019. OpenReview.net, 2019." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 361, + 295, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 361, + 295, + 415 + ], + "spans": [ + { + "bbox": [ + 56, + 361, + 295, + 415 + ], + "type": "text", + "content": "[63] Basile Van Hoorick, Rundi Wu, Ege Ozguroglu, Kyle Sargent, Ruoshi Liu, Pavel Tokmakov, Achal Dave, Changxi Zheng, and Carl Vondrick. Generative camera dolly: Extreme monocular dynamic novel view synthesis. European Conference on Computer Vision (ECCV), 2024." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 416, + 295, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 416, + 295, + 459 + ], + "spans": [ + { + "bbox": [ + 56, + 416, + 295, + 459 + ], + "type": "text", + "content": "[64] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 460, + 295, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 460, + 295, + 503 + ], + "spans": [ + { + "bbox": [ + 56, + 460, + 295, + 503 + ], + "type": "text", + "content": "[65] Wen Wang, Yan Jiang, Kangyang Xie, Zide Liu, Hao Chen, Yue Cao, Xinlong Wang, and Chunhua Shen. Zero-shot video editing using off-the-shelf image diffusion models. arXiv preprint arXiv:2303.17599, 2023." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 503, + 295, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 503, + 295, + 558 + ], + "spans": [ + { + "bbox": [ + 56, + 503, + 295, + 558 + ], + "type": "text", + "content": "[66] Xiaodong Wang, Chenfei Wu, Shengming Yin, Minheng Ni, Jianfeng Wang, Linjie Li, Zhengyuan Yang, Fan Yang, Lijuan Wang, Zicheng Liu, et al. Learning 3d photography videos via self-supervised diffusion on single images. arXiv preprint arXiv:2302.10781, 2023." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 559, + 295, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 559, + 295, + 613 + ], + "spans": [ + { + "bbox": [ + 56, + 559, + 295, + 613 + ], + "type": "text", + "content": "[67] Xiang Wang, Hangjie Yuan, Shiwei Zhang, Dayou Chen, Jiuniu Wang, Yingya Zhang, Yujun Shen, Deli Zhao, and Jingren Zhou. Videocomposer: Compositional video synthesis with motion controllability. Advances in Neural Information Processing Systems, 36, 2024." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 614, + 295, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 614, + 295, + 668 + ], + "spans": [ + { + "bbox": [ + 56, + 614, + 295, + 668 + ], + "type": "text", + "content": "[68] Chung-Yi Weng, Brian Curless, and Ira Kemelmacher-Shlizerman. Photo wake-up: 3d character animation from a single photo. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5908-5917, 2019." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 669, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 669, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 669, + 295, + 713 + ], + "type": "text", + "content": "[69] Olivia Wiles, Georgia Gkioxari, Richard Szeliski, and Justin Johnson. Synsin: End-to-end view synthesis from a single image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020." + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 441 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 105 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 105 + ], + "type": "text", + "content": "[70] Andrew Woods, Tom Docherty, and Rolf Koch. Image distortions in stereoscopic video systems. Proc SPIE, 1915, 2002." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 106, + 553, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 106, + 553, + 173 + ], + "spans": [ + { + "bbox": [ + 316, + 106, + 553, + 173 + ], + "type": "text", + "content": "[71] Jay Zhangjie Wu, Yixiao Ge, Xintao Wang, Stan Weixian Lei, Yuchao Gu, Yufei Shi, Wynne Hsu, Ying Shan, Xiaohu Qie, and Mike Zheng Shou. Tune-a-video: One-shot tuning of image diffusion models for text-to-video generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7623-7633, 2023." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 174, + 553, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 174, + 553, + 206 + ], + "spans": [ + { + "bbox": [ + 316, + 174, + 553, + 206 + ], + "type": "text", + "content": "[72] Dejia Xu, Yifan Jiang, Peihao Wang, Zhiwen Fan, Humphrey Shi, and Zhangyang Wang. Sinnerf: Training neural radiance fields on complex scenes from a single image, 2022." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 207, + 553, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 207, + 553, + 240 + ], + "spans": [ + { + "bbox": [ + 316, + 207, + 553, + 240 + ], + "type": "text", + "content": "[73] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data, 2024." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 241, + 553, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 241, + 553, + 274 + ], + "spans": [ + { + "bbox": [ + 316, + 241, + 553, + 274 + ], + "type": "text", + "content": "[74] Meng You, Zhiyu Zhu, Hui Liu, and Junhui Hou. Nvs-solver: Video diffusion model as zero-shot novel view synthesizer. arXiv preprint arXiv:2405.15364, 2024." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 275, + 553, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 275, + 553, + 308 + ], + "spans": [ + { + "bbox": [ + 316, + 275, + 553, + 308 + ], + "type": "text", + "content": "[75] Jason J. Yu, Fereshteh Forghani, Konstantinos G. Derpanis, and Marcus A. Brubaker. Long-term photometric consistent novel view synthesis with diffusion models, 2023." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 308, + 553, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 308, + 553, + 353 + ], + "spans": [ + { + "bbox": [ + 316, + 308, + 553, + 353 + ], + "type": "text", + "content": "[76] Guofeng Zhang, Wei Hua, Xueying Qin, Tien-Tsin Wong, and Hujun Bao. Stereoscopic video synthesis from a monocular video. IEEE Transactions on Visualization and Computer Graphics, 13(4):686-696, 2007." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 354, + 553, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 354, + 553, + 397 + ], + "spans": [ + { + "bbox": [ + 316, + 354, + 553, + 397 + ], + "type": "text", + "content": "[77] Min Zhao, Rongzhen Wang, Fan Bao, Chongxuan Li, and Jun Zhu. Controlvideo: Adding conditional control for one shot text-to-video editing. arXiv preprint arXiv:2305.17098, 2023." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 398, + 553, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 398, + 553, + 441 + ], + "spans": [ + { + "bbox": [ + 316, + 398, + 553, + 441 + ], + "type": "text", + "content": "[78] Tinghui Zhou, Richard Tucker, John Flynn, Graham Fyffe, and Noah Snavely. Stereo magnification: Learning view synthesis using multiplane images. CoRR, abs/1805.09817, 2018." + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "821" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/SpatialLLM_ A Compound 3D-Informed Design towards Spatially-Intelligent Large Multimodal Models/cf47bc9d-8b67-4bc0-a3fe-ba2e3465761d_content_list.json b/2025/SpatialLLM_ A Compound 3D-Informed Design towards Spatially-Intelligent Large Multimodal Models/cf47bc9d-8b67-4bc0-a3fe-ba2e3465761d_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..5557c4b69df58bca2980cac458aa020073bcde16 --- /dev/null +++ b/2025/SpatialLLM_ A Compound 3D-Informed Design towards Spatially-Intelligent Large Multimodal Models/cf47bc9d-8b67-4bc0-a3fe-ba2e3465761d_content_list.json @@ -0,0 +1,1776 @@ +[ + { + "type": "text", + "text": "SpatialLLM: A Compound 3D-Informed Design towards Spatially-Intelligent Large Multimodal Models", + "text_level": 1, + "bbox": [ + 210, + 130, + 785, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Wufei Ma1, Luoxin Ye1, Celso M de Melo2, Alan Yuille1, Jieneng Chen1", + "bbox": [ + 199, + 202, + 795, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Johns Hopkins University $^{2}$ DEVCOM Army Research Laboratory", + "bbox": [ + 289, + 223, + 702, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 273, + 326, + 289 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Humans naturally understand 3D spatial relationships, enabling complex reasoning like predicting collisions of vehicles from different directions. Current large multimodal models (LMMs), however, lack of this capability of 3D spatial reasoning. This limitation stems from the scarcity of 3D training data and the bias in current model designs toward 2D data. In this paper, we systematically study the impact of 3D-informed data, architecture, and training setups, introducing SpatialLLM, a large multi-modal model with advanced 3D spatial reasoning abilities. To address data limitations, we develop two types of 3D-informed training datasets: (1) 3D-informed probing data focused on object's 3D location and orientation, and (2) 3D-informed conversation data for complex spatial relationships. Notably, we are the first to curate VQA data that incorporate 3D orientation relationships on real images. Furthermore, we systematically integrate these two types of training data with the architectural and training designs of LMMs, providing a roadmap for optimal design aimed at achieving superior 3D reasoning capabilities. Our SpatialLLM advances machines toward highly capable 3D-informed reasoning, surpassing GPT-4o performance by $8.7\\%$ . Our systematic empirical design and the resulting findings offer valuable insights for future research in this direction. Our project page is available at this link.", + "bbox": [ + 89, + 305, + 483, + 681 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 709, + 220, + 724 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "When humans observe a scene, they perceive more than isolated objects; they intuitively understand the spatial relationships among these objects [55], enabling complex reasoning and interaction with the environment. For example, recognizing that a car is approaching an obstacle and might collide involves interpreting the 3D object orientations and spatial relationships. This innate ability to comprehend and reason about 3D space is crucial for navigating and interacting with the physical world.", + "bbox": [ + 89, + 734, + 482, + 869 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite significant advancements, current LMMs [1, 4, 39, 41, 57], exhibit limited capability of 3D spatial reason", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/aaad0f38892f2429e3ab001841f512bf0b3f249d95cdb0c266a2b6708fdcdd10.jpg", + "image_caption": [ + "Orientation Relationships", + "Spatial Reasoning Relationships", + "Distance Relationships The pedestrian is closer to the viewer than the cyclist." + ], + "image_footnote": [], + "bbox": [ + 517, + 271, + 772, + 347 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "White van is facing the pedestrian instead of the cyclist", + "bbox": [ + 519, + 378, + 612, + 405 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "From the perspective of the pedestrian, the white van is on the right side of the pedestrian", + "bbox": [ + 625, + 378, + 764, + 405 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6ab07807c283b9bfe2ec939f21710c29d7ab3e485090fe18cd7b6272006c25b8.jpg", + "image_caption": [ + "3D-Informed Architecture" + ], + "image_footnote": [], + "bbox": [ + 776, + 332, + 888, + 377 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/50c06ac2dbca8f308a5a567d29b45d3beb18dc061831de4411333dfecb035bce.jpg", + "image_caption": [ + "3D-Informed Data" + ], + "image_footnote": [], + "bbox": [ + 522, + 411, + 573, + 465 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1M 3D-Informed Probing Data description on depth, distance, & orientation", + "bbox": [ + 581, + 422, + 656, + 467 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1M 3D-Informed Instruction Tuning Data conversations about 3D spatial relationships", + "bbox": [ + 660, + 422, + 767, + 467 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Mixed encoders, 3D-aware visual encoders, etc. 3D-Informed Training", + "bbox": [ + 774, + 422, + 895, + 453 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Pre-pretraining, alignment & finetuning", + "bbox": [ + 776, + 457, + 898, + 476 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Figure 1. 3D spatial reasoning is crucial for LMMs to ground objects in 3D space and infer their 3D spatial relationships, such as distance, orientation, and spatial interactions. We systematically study the impact of 3D-informed data, architecture, and training setups, introducing SpatialLLM, an LMM with advanced 3D spatial reasoning abilities.", + "bbox": [ + 511, + 488, + 906, + 571 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ing. These models excel at identifying objects and generating descriptive captions but struggle with tasks that require intricate 3D reasoning, such as predicting object interactions or understanding physical events.", + "bbox": [ + 511, + 606, + 905, + 666 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The primary challenge lies in the scarcity of high-quality 3D spatial relationship data, which prevents models from learning effective 3D spatial reasoning. Collecting a small set of high-quality 3D-aware data to tackle the first challenge is feasible, albeit labor-intensive, using readily available tools. Recent efforts [2, 14] leverage vision foundation models to recognize [67], detect [42], segment [31] and estimate the depth [10, 63] in images. These approaches form semi-automated pipelines to construct pseudo-2.5D datasets for training LMMs with spatial reasoning capabilities, such as depth ordering [14, 58].", + "bbox": [ + 511, + 670, + 906, + 835 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "However, a significant gap remains: previous works [2, 14, 16] have primarily focused on 3D distance relationships, overlooking the crucial role of 3D object orientation. This is primarily due to the absence of robust 6D pose estimation.", + "bbox": [ + 511, + 839, + 906, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "17249", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tors. Understanding object orientation is essential for complex spatial reasoning tasks, such as determining if two cars are facing the same direction to avoid dangerous collision. Addressing this gap, we aim to incorporate 3D orientation relationships—converted from ImageNet3D [17]—into our data engine, making us the first to enable complex spatial reasoning involving 3D orientation relationships.", + "bbox": [ + 89, + 90, + 480, + 196 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Even with 3D-informed data in hand, fully unlocking its potential remains challenging due to the intricate, multistage development process required for an LMM. For example, SpatialVLM [14] utilizes 3D-informed data during the instruction tuning stage, but it remains uncertain whether such data can enhance other stages like vision encoder pretraining or multimodal alignment. This limitation suggests that a more holistic approach is necessary.", + "bbox": [ + 89, + 198, + 480, + 319 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we present a systematic investigation into building LMMs with enhanced 3D spatial reasoning capabilities through a compound design approach, as illustrated in Fig. 1. Our key contributions are threefold:", + "bbox": [ + 89, + 321, + 480, + 381 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "First, we identify three fundamental types of 3D spatial relationships that LMMs must master: distance relationships, orientation relationships, and complex spatial reasoning that combines both. To evaluate these capabilities, we introduce SpatialVQA, a comprehensive benchmark containing 1,323 questions that target 3D spatial understanding, built upon diverse urban and indoor scenes from Omni3D [11].", + "bbox": [ + 89, + 382, + 480, + 503 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Second, we propose a novel compound 3D-informed design that introduces improvements across multiple dimensions, leading to our proposed SpatialLLM model. On the data front, we develop two types of 3D-informed training datasets: (1) 3D-informed probing data focused on object-level 3D properties, and (2) 3D-informed conversation data for complex spatial relationships. Architecturally, we explore mixing different visual encoders and investigate their 3D awareness capabilities. For training, we introduce 3D-informed data at multiple stages: pre-pretraining, multimodal alignment, and instruction tuning.", + "bbox": [ + 89, + 505, + 480, + 671 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Third, we present the first comprehensive search over the LMM design space for spatial reasoning tasks and propose a roadmap towards developing state-of-the-art models in this domain. The experiment demonstrates that our proposed SpatialLLM achieves state-of-the-art performance on spatial reasoning tasks, surpassing both GPT-4o [1] (by $8.7\\%$ ) and spatial-centric models like SpatialVLM [14] (by $10.5\\%$ ). Our systematic analysis reveals that while architectural improvements contribute a $1.3\\%$ gain, the strategic incorporation of 3D-informed data across different training stages yields a substantial $13.7\\%$ performance improvement.", + "bbox": [ + 89, + 672, + 480, + 852 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our findings highlight the importance of a holistic approach to improving LMMs' 3D spatial reasoning capabilities, demonstrating that both architectural choices and care", + "bbox": [ + 89, + 854, + 480, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "fully curated open-source 3D-informed training data play crucial roles in achieving superior performance. This work establishes a new benchmark for spatial reasoning in LMMs and provides valuable insights for future research in this direction.", + "bbox": [ + 511, + 90, + 903, + 165 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 513, + 180, + 663, + 196 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Large multimodal models (LMMs). Recently, there has been a significant surge in research on multimodal LLMs, exemplified by products like GPT-4 [1] and Gemini [57], as well as open-source models such as LLaVA [34, 39, 41] and Qwen-VL [8]. These open-source models combine pretrained vision encoders [15, 27, 47, 48] with LLMs [7, 28, 59], enabling the language models to handle a wide range of tasks that involve images as input. Typically, these methods have the alignment stage that aligns image features with the input embedding space of the language model using connectors [13, 35, 39], trained with paired images and captions [53]. The instruction fine-tuning [41] stage uses data that contains both images and instructions to enhances LMMs. However, current LMMs are unsuccessful in spatial visual reasoning, due to limited spatial-aware data and stagnant 3D-informed architecture design. We aim to bridge this research gap.", + "bbox": [ + 511, + 205, + 903, + 462 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3D spatial reasoning with LMMs. Understanding spatial relationships between objects in images is crucial for visual reasoning. While this comes naturally to humans, LMMs were found to struggle with recognizing spatial relations [49, 60, 61], due to a lack of 3D-aware spatial-related data during training. While neural symbolic methods [60] achieved improved performance on synthetic benchmarks by involving explicit 3D reasoning modules [26, 44], they struggled to generalize to a broader range of objects found in real images. Fortunately, with foundation models like SAM [31] and depth estimators [10, 63], semi-automatic annotation for 3D-aware VQA data has been adopted in recent studies [14, 16]. SpatialVLM [14] uses the pseudo-annotated VQA data for training only while SpatialRGPT [14, 16] further augments LLaVA with extra input of regional mask and depth map during inference, which is not our focus. Although human-annotated 3D-aware VQA data is rare, we aim to leverage human-annotated 3D object-level datasets (like ImageNet3D [45]) by simply converting image-pose pairs into image-text pair data. However, even with the collected data, a major challenge remains in effectively utilizing 3D-aware data to train optimal models. Current approaches like SpatialVLM [14] use the standard LLaVA and two-stage training. To date, no work has addressed when (which training stage) and where (which network component) are optimal to inject the 3D-awareness to LMMs. We bridge this gap by conducting a compound design that simultaneously considers 3D-aware data, architecture, and the training, leading to a best", + "bbox": [ + 511, + 463, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "17250", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "performing design. Rather than competing, our study has the potential to complement prior works [14, 16] by offering new compound design recipe.", + "bbox": [ + 89, + 90, + 482, + 137 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methods", + "text_level": 1, + "bbox": [ + 89, + 154, + 189, + 169 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section we start by reviewing LMMs in Sec. 3.1, specifically the architecture and the choice of training data and strategies. We present the task of reasoning 3D spatial relationships and explain the challenges LMMs face when answering these questions in Sec. 3.2. Then we study how to address these challenging problems with LMMs and introduce our SpatialLLM in Sec. 3.3.", + "bbox": [ + 89, + 180, + 482, + 286 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminary of LMMs", + "text_level": 1, + "bbox": [ + 89, + 299, + 295, + 316 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Data. The success of deep learning in vision is largely due to well-curated data [20, 56], and LMMs similarly depend on this foundation, albeit with diverse data formats below.", + "bbox": [ + 89, + 323, + 482, + 367 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Image-only data, such as ImageNet [17], is commonly used for MAE [23]/DINOv2 [47] pretraining with a self-supervised objective. This step enables the model to learn rich visual features solely from visual signals.", + "- Noisy image-text pairs: Large-scale image-text pairs [20, 37, 52] are used for pretraining with contrastive loss (e.g., CLIP [48, 65]) to align visual and textual representations.", + "- Multimodal alignment: At this stage, image-caption pairs (e.g., CC3M [53]) are utilized to align visual features with the language model's word embedding space, typically through image-captioning tasks. This pretraining ensures that the model can interpret and relate visual content to textual queries in a shared semantic space.", + "- Multimodal instruction-following data. Visual instruction tuning data is essential but challenging to collect, as it rarely exists in its natural form online. Prior works [39, 41, 46, 58] repurpose VQA benchmarks [22, 32] into instruction-tuning datasets." + ], + "bbox": [ + 91, + 369, + 482, + 641 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Architecture. A standard LMM [39, 41] consists of a visual encoder to process the image, a multimodal connector to transform the visual feature to visual token, and a LLM for reasoning. However, the extent to which these architectural components contribute to the spatial understanding of LMMs remains unknown. We review these architectural components as below.", + "bbox": [ + 89, + 642, + 482, + 747 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Vision encoder: Most LMMs rely on language-supervised models like CLIP [48], leveraging web-scale noisy image-text data. Recent LMMs [43, 58] integrate vision models trained only on visual signals, including self-supervised [23, 47, 69] and generative models [51].", + "- Connector: Features from a visual encoder are mapped into the LLM token space by connectors [3, 13, 35, 41] such as a MLP [39, 41].", + "- LLMs: LLMs [1, 7, 59] serve as the foundational backbone of LMMs, driving their reasoning capabilities." + ], + "bbox": [ + 91, + 750, + 482, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Open-source LLMs like Llama series [18, 59] are frequently adopted to enhance the reasoning capability.", + "bbox": [ + 526, + 90, + 903, + 121 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Training. The aforementioned data supports pretraining for strong visual representation learning, alignment pretraining to synchronize visual and language representations, and instruction tuning to enable multimodal instruction-following and reasoning capabilities", + "bbox": [ + 511, + 122, + 903, + 196 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Pre-pretraining. This stage focuses on developing foundational visual representations, often with reconstruction-based objectives (e.g., MAE [23], DINOv2 [47]). This step prepares the model to handle more complex multimodal tasks by first refining its visual understanding.", + "- Multimodal alignment pretraining. At this stage, the model is trained to describe images in details to align visual and language representations in the same space.", + "- Visual instruction tuning. The final stage involves refining the model's ability to process complex multimodal instructions using visual instruction tuning data. This often incorporates transformed VQA datasets, enhancing the model's capacity in handling tasks that demand coordinated reasoning across language and vision." + ], + "bbox": [ + 513, + 198, + 903, + 409 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Reasoning 3D Spatial Relationships", + "text_level": 1, + "bbox": [ + 511, + 420, + 821, + 436 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3D spatial reasoning analyzes the spatial relationships between objects in the 3D world space, such as objects being far or close to each other, or one object lies to the left or right of another object [60]. It extends beyond standard 3D tasks, such as 3D object detection and 3D pose estimation that studies the 3D nature of individual objects, and focuses on understanding their mutual relationships in 3D. In this work we identify three main types of 3D spatial relationships, i.e., distance, orientation, and spatial reasoning, and study how we can better address these problems with LMMs.", + "bbox": [ + 511, + 443, + 906, + 594 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3D spatial relationships. For distance spatial relationships, we study if models can estimate and compare 3D distances between objects. These spatial relationships require more than depth awareness, which gives the distance between the viewer and the object and can be roughly estimated from the object scale, but also the 3D distances between two objects that are often harder to estimate. On the other hand, orientation spatial relationships require models to understand the 3D orientations of an object, e.g., predicting if two objects are facing the same direction. Lastly, spatial reasoning spatial relationships look into 3D spatial relationships that need a combination of 3D awareness of location and orientation and then performing spatial reasoning on the 3D information.", + "bbox": [ + 511, + 614, + 906, + 825 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2.1. Challenges of 3D spatial reasoning", + "text_level": 1, + "bbox": [ + 511, + 835, + 799, + 851 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Reasoning about 3D spatial relationships remains a challenging task for state-of-the-art open-sourced and proprietary LMMs. We attribute this to two main limitations in", + "bbox": [ + 511, + 854, + 903, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "17251", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/06f567a187c83769ca94b276493b3c978c825a3c7e72462d7e72e61fcf483f85.jpg", + "image_caption": [ + "√3D Location 03D Orientation", + "User" + ], + "image_footnote": [], + "bbox": [ + 114, + 88, + 282, + 183 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/f5d09c711e117ef47c7b87625ca17b09a1a4a1e550c307ebe432c1bc91674eff.jpg", + "image_caption": [ + "The chair in Region [2] is closer to the chair in Region [0].", + "Spatial Reasoning Relationships $\\sqrt{3}$ Location $\\sqrt{3}$ Orientation", + "User", + "If I stand at the location of the car in Region [0] facing where it is facing, where is the bus in Region [1]?", + "The bus in Region [1] is on the left side of the car in Region [0].", + "Figure 2. Examples from our SpatialVQA benchmark featuring a broad range of questions that require 3D spatial reasoning." + ], + "image_footnote": [], + "bbox": [ + 93, + 186, + 303, + 281 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/a41c6ea95a9124d7247030d0b993c6e50fb2f699ac0f49311a988ab340cde7dc.jpg", + "image_caption": [ + "Which object is closer to the chair in Region [0], the bookcase in Region [1] or the chair in Region [2]?", + "3DI-LMM", + "√3D Location √3D Orientation", + "Use", + "Are the car in Region [0] and the car in Region [1] facing the same direction, or different directions?", + "The car in Region [0] and the car in Region [1] are facing different directions." + ], + "image_footnote": [], + "bbox": [ + 501, + 89, + 712, + 183 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/17b5d311c6cbe245dfc4975a68277ca32ad4e9e28fc260aca93e1ccafaac99e4.jpg", + "image_caption": [ + "3DI-LMM", + "Spatial Reasoning Relationships", + "√3D Location √3D Orientation", + "User", + "Is the car in Region [0] oriented towards the car in Region [1] or the car in Region [2]?", + "3DI-LMM", + "The car in Region [0] is facing towards the car in Region [1]." + ], + "image_footnote": [], + "bbox": [ + 501, + 186, + 712, + 281 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "modern LMMs, the lack of 3D awareness and inability to perform 3D spatial reasoning.", + "bbox": [ + 89, + 335, + 482, + 367 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3D awareness refers to the ability of a visual encoder to represent 3D-aware features such as 3D object shapes and 3D orientations. Previous studies resorted to proxy tasks, e.g., part correspondence [19] and pose estimation [45], to quantitatively evaluate the 3D awareness of visual foundation models with linear probing. Results showed that most visual encoder, specifically ones trained with language supervision, demonstrated limited 3D awareness of 3D object poses [45], which inevitably limits the 3D spatial reasoning capabilities of LMMs built upon these encoders.", + "bbox": [ + 89, + 386, + 483, + 537 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Inability to perform 3D spatial reasoning. To infer the 3D spatial relationships, models must also be able to perform 3D spatial reasoning from 3D-aware visual features. Despite the abundant Internet-scale data for multi-modal training, we argue that high-quality 3D spatial reasoning data are scarce. Existing pretraining and visual instruction tuning data for LMMs [41, 58] focused on detailed descriptions and conversations about scenes, appearances, and actions, while being vague about the 3D spatial relationships that build up the high-level physical events. Despite their success on a wide range of image understanding tasks, these LMMs fail to learn the nature of 3D scenes or to perform 3D spatial reasoning.", + "bbox": [ + 89, + 556, + 483, + 753 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.2. SpatialVQA for Evaluation", + "text_level": 1, + "bbox": [ + 89, + 761, + 326, + 775 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Despite the rising interests in spatial reasoning, there lacks a comprehensive open source benchmark for reasoning 3D spatial relationships. Early datasets on spatial reasoning were built on 3D scans rather than images [6, 64]. Recent visual-language benchmarks on spatial reasoning either focused on 2D spatial relationships [14, 16], e.g., left or right in the image plane, or only on distance spatial relationships [58].", + "bbox": [ + 89, + 780, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To enable a thorough evaluation of the 3D spatial reasoning capabilities of LMMs, we introduce SpatialVQA, a visual question answering benchmark that covers a wide variety of 3D spatial relationships. Our SpatialVQA distinguishes itself from all previous spatial reasoning benchmarks in the sense that all questions require different levels of 3D awareness and cannot be answered from 2D spatial reasoning only. Specifically, SpatialVQA consists of 1,323 questions, ranging from questions can be answered from distance-awareness only, e.g., which object is closer to a third object, to questions that require estimating objects' 3D orientations, e.g., if two objects are facing the same direction, and eventually questions that require complex 3D spatial reasoning over a combination of object 3D locations and orientations, e.g., an object is in the front/left/back/right direction of another object.", + "bbox": [ + 511, + 335, + 906, + 578 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We build our SpatialVQA on images from Omni3D [11], with 3D bounding box annotations on diverse objects from both urban [12, 21] and indoor scenes [9, 50, 54]. We follow [16, 58] and develop rule-based methods to generate visual question-answer pairs from the 3D groundtruths. Please refer to Fig. 2 for some examples of SpatialVQA and to appendix Sec.B for details about our benchmark.", + "bbox": [ + 511, + 580, + 905, + 686 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Compound 3D-Informed Design", + "text_level": 1, + "bbox": [ + 511, + 708, + 797, + 724 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Problem formulation. While previous approaches focused on generating 3D-informed instruction tuning data to enable LMMs with the abilities to estimate distances, depths, or spatial relationships [14, 16], it is largely understudied the best recipe for developing LMMs with strong 3D spatial reasoning capabilities. Motivated by the two aforementioned limitations of LMMs in Sec. 3.2.1, we consider two main aspects in our compound 3D-informed design – the architecture design that leads to visual encoders with strong 3D awareness and the training setup that advances the 3D spatial reasoning capabilities of LMMs.", + "bbox": [ + 511, + 734, + 906, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "17252", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/aa085b7ec1966899e356d8745d3fa88c93bdcb56458e2df75f569419bac3f58f.jpg", + "image_caption": [ + "(a) data", + "(b) training stage", + "(c) architecture", + "(d) vision encoder" + ], + "image_footnote": [], + "bbox": [ + 89, + 88, + 480, + 212 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/458bf0902d6b89f5e588a589b82a63450156b0daed0108466939cec495bbf875.jpg", + "image_caption": [ + "Figure 3. Design space for LMMs capable of spatial reasoning. The dashed boxes and lines highlight our new design space compared to LLaVA-v1.5. This compound design simultaneously considers 3D-informed data, architecture, and training methods to search for the best-performing models for spatial reasoning.", + "Figure 4. Comparison of data cards showcasing the curation process and data types: standard LLaVA data (left) vs. our 3D-informed data (right). For our 3D-informed data curation, we label depth and distance of objects in OpenImage [33] using semi-automatic tools, resulting in 3DI-Pb-OI1M. Beyond this, we introduce 3D orientation probing data derived from human-annotated ImageNet3D [45], resulting in 3DI-Pb-IN166K for alignment. We further augment the conversation about spatial relationship, leading to 3DI-Ft-1M for instruction tuning." + ], + "image_footnote": [], + "bbox": [ + 94, + 333, + 282, + 512 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/4dc08d9a73e6c392cac4a36ceb33b3fc99b84c1476af4dd2ec4a1c6a7c9dc97a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 292, + 333, + 478, + 511 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3.1. Design space", + "text_level": 1, + "bbox": [ + 89, + 680, + 227, + 695 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We introduce the design space considered in our work, i.e., choices of training data, model architecture, and training setup that advance the 3D spatial reasoning capabilities of LMMs. Refer to Fig. 3 for an overview of our design space.", + "bbox": [ + 89, + 699, + 483, + 761 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Data. Standard LMMs follow LLaVA [41] and adopte feature alignment data for connector pretraining and visual instruction data for supervised instruction tuning. These datasets focus on detailed descriptions of images or machine-generated multi-turn conversations. Besides standard visual-language datasets, we further consider 3D-informed data, including 3D-aware probing data and 3D-informed instruction tuning data. Inspired by [45], the 3D", + "bbox": [ + 89, + 779, + 483, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "aware probing data involves fundamental object-level 3D properties, such as the azimuth and elevation 3D viewpoint of an object, depth of the object from the viewer, and 3D distances between two objects. The goal of the 3D-aware probing data is to effectively inject fundamental 3D awareness into the models and benefit subsequent reasoning. Meanwhile, the 3D-informed instruction tuning data focuses on high-level 3D spatial relationships between the objects, which are built on a combination of the fundamental 3D properties and 3D spatial reasoning.", + "bbox": [ + 511, + 90, + 903, + 241 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To generate the 3D-aware probing data and the 3D-informed instruction tuning data, we follow the previous data generation pipelines [14, 16] and exploit visual foundation models to extract and tag the objects in 3D space. Furthermore, we adopt a pretrained 6D pose estimator in [45] and estimate 3D orientations of the objects, which provides the crucial 3D information for generating 3D-informed data regarding various orientation-related 3D spatial relationships. We illustrate our 3D-informed data generation pipeline in Fig. 4 and refer the readers to appendex Sec.A for details of our 3D-informed data generation.", + "bbox": [ + 511, + 242, + 906, + 409 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Architecture. Inspired by Probe3D [19] which probes the 3D-awareness of visual foundation models, we study the design of visual encoder to enable better 3D spatial reasoning. We consider two types of visual encoders: (i) Frozen & pretrained visual encoder CLIP [48] following [41] but with the option to mix a wider variety of models, such as MAE [23] and DINOv2 [47], and (ii) 3D-aware visual encoders finetuned with our 3D-informed data, e.g., contrastive pretraining or finetuning with LoRA using language modeling loss.", + "bbox": [ + 511, + 429, + 908, + 566 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training setup. To address the limitations of 3D-aware LMMs discussed in Sec. 3.2.1, we propose new training setups that aim to improve 3D awareness and advance the 3D spatial reasoning capabilities. We consider the following: (i) in Stage 0: use 3D-informed probing data (3DI-Pb-OI1M/1N166K) to tune the Lora layers of CLIP vision encoder (Pre-pretraining); (ii) in Stage 1: 3D-informed multimodal alignment using a combination of standard CC558K [41] and our 3D-informed probing data (3DI-Pb-OI1M/1N166K); and (iii) in Stage 2: 3D-informed instruction tuning with the standard Mix665K [41] and our 3D-informed visual instruction tuning data (3DI-Ft-1M).", + "bbox": [ + 511, + 585, + 908, + 767 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3.2. Designing 3D-Informed LMM: A Roadmap", + "text_level": 1, + "bbox": [ + 511, + 775, + 864, + 791 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Design instantiation and comparison. The mentioned design space can result in several design instantiations, as depicted in Fig. 5. In a multimodal LLM, a pre-trained CLIP visual encoder extracts grid features from the input image. These features are converted into visual tokens via a connector and then processed by the LLM for visual reasoning. Fig. 5 (b) shows that LLaVA is a standard design space", + "bbox": [ + 511, + 795, + 908, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "17253", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/af92f889ccc138fcba6a6605c945b4eb095cf3f72f3750ce067312d2cfe9ce92.jpg", + "image_caption": [ + "(a) our proposed design" + ], + "image_footnote": [], + "bbox": [ + 114, + 103, + 486, + 208 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/be280e1db447da6cdc52ac19ab235d8ee252aacb77c2d6e70818603c0f61f32d.jpg", + "image_caption": [ + "(b) design of LLaVA-v1.5 (CVPR'24)" + ], + "image_footnote": [], + "bbox": [ + 508, + 103, + 882, + 207 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/529895ee21b4cbcfe8f8bc1b75b4b2672e56e3d025a7643e574f089b4329c533.jpg", + "image_caption": [ + "(c) design of SpatialVLM (CVPR'24)" + ], + "image_footnote": [], + "bbox": [ + 114, + 252, + 483, + 354 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d4388b1caeed971136a941795b4946bc2bf24d89a0a4915e4510e3086c738429.jpg", + "image_caption": [ + "(d) data cards" + ], + "image_footnote": [], + "bbox": [ + 521, + 253, + 887, + 349 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "without any new spatial-centric design, meaning that the 3D-awareness is completely missing in LLaVA (Fig. 5 (b)). As in Fig. 5 (c), SpatialVLM improves LLaVA by collecting 3D-informed data that is applied to instruction tuning for LLM, boosting the spatial reasoning capability. Compared to SpatialVLM, our design (in Fig. 5 (a)) inject the 3D-awareness by introduce 3D-informed alignment with ImageNet3D data. Additionally, our 3D-informed intrusion tuning data contain extra 3D orientation data that is missing in spatialVLM.", + "bbox": [ + 88, + 489, + 482, + 638 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Design roadmap. We present the roadmap going from a standard LLaVA to our final model in Fig. 6. We start with LLaVA-v1.5 which only achieves a $47.7\\%$ accuracy on our SpatialVQA. This will be our baseline. We then study a series of design decisions.", + "bbox": [ + 89, + 640, + 483, + 715 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- + mixed vision encoder. To build stronger visual encoder, we mix the original CLIP vision encoder [48] with the DINOv2 [47] vision encoder to extract the hybrid feature maps, lifting a $0.3\\%$ performance gain. This show that the self-supervised feature can enhance spatial reasoning.", + "- Vicuna1.5 $\\rightarrow$ Llama3. We switch our large language model from Vicuna-v1.5-7B [68] to Llama3-8B [18], as the latter achieves $20\\%$ higher accuracy on MMLU. This change results in $1\\%$ gain on spatial reasoning. We analyze that improved language reasoning may enhance spatial reasoning as well.", + "- +3DI-Ft1M data for stage 2 instruction tuning. We col" + ], + "bbox": [ + 89, + 719, + 482, + 901 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/42d503fcb94cf182feed8978f02582e39a84d445a516eb8bc4e6177b68a8594c.jpg", + "image_caption": [ + "Figure 5. Design instantiation and comparison. (a) Architecture and Training of our proposed design. We investigate the 3D-awareness of mixed visual encoders, and incorporate 3D-informed data at each training stage across all architecture components. (b) LLaVA comparison, where only traditional data are used. (c) SpatialVLM comparison, where 3D-informed data are used in the final stage. (d) Traditional data vs. 3D-informed data. Note SpatialVLM lacks of 3D orientation data. Architectures trained with traditional data and 3D-informed data are marked with blue and orange backgrounds, respectively.", + "Figure 6. We modernize a standard LLaVA-v1.5 towards the design of a 3D-informed LMM. The bars are the answer accuracies on the SpatialVQA benchmark, indicating the spatial reasoning capability of each model; results for the GPT-4o is shown with the purple bar." + ], + "image_footnote": [], + "bbox": [ + 516, + 486, + 895, + 672 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "lect the new spatial-aware training set and use it in the instruction tuning stage. Impressively, this 3D-informed instruction tuning results in a $10.7\\%$ performance improvement, highlighting the critical importance of 3D-informed data.", + "bbox": [ + 524, + 795, + 903, + 868 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "- + 3DI-Pb-1N166K data for stage 1 multimodal alignment. We propose leveraging object-level 3D-informed", + "bbox": [ + 513, + 869, + 906, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "17254", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "annotation (e.g., 3D location and 3D orientation of objects) to align visual features with 3D awareness. 3DI-Pb-1N166K are converted from human-annotated ImageNet3D [45]. This experiment results in an additional $3\\%$ performance improvement, underscoring the importance of 3D-informed feature alignment.", + "bbox": [ + 102, + 90, + 480, + 181 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In a nutshell, the performance gain can be summarized as 1) better architecture and 2) 3D-informed data. This roadmap leads to our SpatialLLM.", + "bbox": [ + 89, + 184, + 482, + 229 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 242, + 222, + 258 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section we start by outlining the experimental setup in Sec. 4.1, followed by presenting the results in Sec. 4.2.", + "bbox": [ + 89, + 267, + 482, + 297 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.1. Experimental setup", + "text_level": 1, + "bbox": [ + 89, + 306, + 277, + 321 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We prepare our training data as described in Fig. 4, with 1M paired image and 3D-informed text for pre-pretraining and alignment, and another 1M multi-turn 3D-informed conversation for instruction tuning. Data distributions are supplemented in appendix. Our training setup is built upon LLaVA-v1.5 [39] and all hyperparameters remain unchanged unless explicitly stated otherwise. We briefly overview the training and data setup for the ablated design choices in Sec. 3.3.1. All experiments are conducted in 8 Nvidia A100 GPUs. We evaluate the spatial reasoning capability of LMMs on SpatialVQA as introduced in Sec. 3.2.2.", + "bbox": [ + 89, + 328, + 482, + 494 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Results", + "text_level": 1, + "bbox": [ + 89, + 503, + 184, + 517 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparison with the state-of-the-art. Table 1 presents the performance of our model compared to state-of-the-art proprietary systems such as GPT-4o and Claude 3.5 Sonnet, along with top open-source models. Remark", + "bbox": [ + 89, + 523, + 482, + 585 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/3fea4ccd4fb51bff2437adc68f8de5949693190db575db53503f5c9cee18cfc0.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelAvg.3D Dist.3D Orient.3D Spatial Rel.
Proprietary
GPT-4o [1]54.059.452.450.8
GPT-4o Mini [1]45.446.939.348.5
Claude 3.5 Sonnet [4]49.350.844.551.5
Open-sourced
LLaVA-v1.5-7B [39]47.750.646.446.2
LLaVA-NeXT-8B [40]50.663.549.241.5
Cambrian-8B [58]51.759.846.448.9
Spatial-centric
SpatialVLM-13B [14]52.264.348.545.8
Ours62.786.352.950.8
", + "bbox": [ + 93, + 597, + 478, + 816 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 1. Comparison with the state-of-the-arts including proprietary and open source models.", + "bbox": [ + 89, + 825, + 482, + 854 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ably, our model achieves a performance of $62.7\\%$ , outperforming the top proprietary model by $8.7\\%$ and the best", + "bbox": [ + 89, + 869, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "open-source model by $10.5\\%$ . Besides, our model excels in all aspects of spatial reasoning involving 3D distance, 3D orientation and 3D spatial relationship. Moreover, our model not only enhances spatial reasoning performance, but still preserves general visual reasoning (detailed in appendix). Interestingly, although SpatialVLM [14] (implemented in SpaceLLaVA [2]) outperforms other open-source models in overall performance, it falls short in 3D orientation reasoning compared to LLaVA, due to the lack of tuning with 3D orientation-specific data. We exclude SpatialRGPT from this comparison as it relies on additional inputs such as object bounding box coordinates, segmentation masks, and depth maps. In contrast, our SpatialLLM, along with LLaVA and Cambrian, focuses on learning spatial awareness directly from raw images. We will consider models with additional inputs in future work. In sum, our model achieves state-of-the-art performance on spatial reasoning, surpassing API models [1, 4] and also spatial-centric VLM [14].", + "bbox": [ + 511, + 90, + 903, + 378 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Thorough empirical exploration of the design space. We comprehensively explore the design space of our approach as shown in Table. 2. We analyze the results and elaborate the findings detailed below.", + "bbox": [ + 511, + 380, + 903, + 441 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Upgrading architectures enable better spatial reasoning. In terms of architecture, integrating a mixed vision encoder can improve overall performance especially for the 3D orientation. Then, upgrading the LLM to the more advanced LLama3-8B enhances spatial reasoning performance by $1\\%$ , leveraging its improved language reasoning capabilities.", + "bbox": [ + 511, + 443, + 903, + 547 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3D-informed alignment and tuning matter a lot. In terms of 3D-informed data and training, we find that 3D-informed instruction tuning with our proposed 3DI-Ft1M dataset yields a substantial performance boost of $+10.7\\%$ . To further refine the framework, we explore three data types for stage 1 multimodal alignment: curated 3DI-Ft1M, 3DI-Pb-OI1M, and 3DI-Pb-IN166K. The model trained with 3DI-Pb-IN166K performs best. We attribute this to the use of human-annotated ImageNet3D data, as opposed to the semi-automated methods employed for 3DI-Pb-OI1M. The human-annotated ImageNet3D provides more accurate 3D orientation of objects. This highlights the importance of high-quality data in the alignment stage.", + "bbox": [ + 511, + 550, + 905, + 747 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3D-informed pretraining of vision encoder? The pretrained CLIP vision encoder, originally designed for image-text contrastive learning, lacks 3D-awareness [19], We hypotheszie that the vision encoder has a chance to benefit from 3D-informed data. To verify this, we conduct 3D-informed pretraining for the vision encoder using curated 3DI-Pb-OI1M probing data. We explore both CLIP [25] and next token prediction (NTP) [41] objectives to fine-tune Lora layers [24] of the ViT. We observe that pre-pretraining in stage 0 reduces performance, suggesting that tuning vi", + "bbox": [ + 511, + 750, + 905, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "17255", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/d8bd78a4ed73c9f2459e29ba927f8d7b04fe2282d1710efcd32f2fa972ac5a39.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Model\nVision EncoderLLMStage 0\nEncoder pretrainingStage 1\nAlignmentStage 2\nInstruction TuningSpatialVQA
Avg.Dist.Orient.Spatial Rel.
Baseline (LLaVA-v1.5)
CLIP [48]Vicuna-7B [68]-CC558K [39]Mix665K [39]47.750.646.446.2
Mixed Encoders
CLIP [48]+MAE [23]Vicuna-7B [68]-CC558KMix665K47.952.245.746.1
CLIP [48]+DINOv2 [47]Vicuna-7B [68]-CC558KMix665K48.051.946.945.6
Advanced LLM
CLIP [48]+DINOv2 [47]Llama3-8B [18]-CC558K [39]Mix665K [39]49.056.351.741.5
3D-informed Tuning for Stage 2
CLIP [48]+DINOv2 [47]Llama3-8B [18]-CC558K [39]Mix665K [39] + 3DI-Ft1M59.785.252.144.8
3D-informed Alignment for Stage 1
CLIP [48]+DINOv2 [47]Llama3-8B [18]-CC558K [39] + 3DI-Ft1MMix665K [39] + 3DI-Ft1M57.580.846.946.4
CLIP [48]+DINOv2 [47]Llama3-8B [18]-CC558K [39] + 3DI-Pb-OI1MMix665K [39] + 3DI-Ft1M60.082.146.951.6
CLIP [48]+DINOv2 [47]Llama3-8B [18]-CC558K [39] + 3DI-Pb-IN166KMix665K [39] + 3DI-Ft1M62.786.352.951.0
3D-informed pretraining for vision encoder in stage 0
CLIP [48]+DINOv2 [47]Llama3-8B [18]3DI-Pb-OI1M (CLIP [25])CC558K [39] + 3DI-Pb-IN166KMix665K [39] + 3DI-Ft1M60.385.050.547.5
CLIP [48]+DINOv2 [47]Llama3-8B [18]3DI-Pb-OI1M (NTP [34])CC558K [39] + 3DI-Pb-IN166KMix665K [39] + 3DI-Ft1M62.685.252.951.5
", + "bbox": [ + 94, + 88, + 903, + 361 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 2. Thorough exploration of the design space and roadmap progression. We systematically examine the 3D-informed design space from the aspects of data, architecture and training. The non-gray rows, listed from top to bottom, illustrate the step-by-step progression of our design roadmap. Gray rows correspond to ablations across the full design space. Our final model is highlighted with a light orange background. Our final model's superior performance over other variants confirms the necessity of a 3D-informed compound design.", + "bbox": [ + 89, + 372, + 906, + 429 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "sual foundation models may compromise feature generalizability. This aligns with prior works advocating for frozen or locked visual encoders [36, 38].", + "bbox": [ + 89, + 455, + 482, + 501 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Roadmap progression towards the best-performing model. After a thorough empirical exploration of the design space, we can devise our roadmap progressed towards the best-performing model. The non-gray rows in Table. 2 align with the roadmap illustrated in Fig. 6, highlighting our systematic approach to upgrade the architecture and incorporate 3D-informed data into the training process. Overall, the improved architecture achieves a $1.3\\%$ performance gain, while 3D-informed alignment and instruction tuning further enhance performance by $13.7\\%$ .", + "bbox": [ + 89, + 501, + 483, + 652 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Qualitative comparison. Fig. 7 illustrates a scenario involving spatial reasoning. GPT-4o struggles to reason about the spatial relationships between objects due to its limited ability to comprehend their 3D orientation. In contrast, our model successfully understands these spatial relationships and provides accurate answers, demonstrating superior 3D spatial reasoning capability.", + "bbox": [ + 89, + 652, + 483, + 758 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusions", + "text_level": 1, + "bbox": [ + 89, + 768, + 218, + 784 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we systematically enhance the 3D spatial reasoning capabilities of LMMs through a compound design approach. By introducing SpatialLLM, we demonstrated that integrating 3D-informed data, architectural innovations, and tailored training setups significantly improves an LMM's ability to understand and reason about complex 3D spatial relationships. To tackle data scarcity, we develop", + "bbox": [ + 89, + 794, + 483, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d2fedb8e633046d43ee48872746547787fde33b0cd8a454c0f32921049fa5c04.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 517, + 455, + 763, + 517 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ef9d4b3f8241ae4a604de36f0ab946aea5993a83fadeadb013519bb1842a82c6.jpg", + "image_caption": [ + "User:" + ], + "image_footnote": [], + "bbox": [ + 524, + 530, + 547, + 550 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Consider the real-world 3D locations and orientations of the objects. Which object is closer to the car in Region [0], the car in Region [1] or the car in Region [2]?", + "bbox": [ + 557, + 526, + 759, + 551 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b8790ef8c43ee80660eff58393f58e2c76fa85958ecd1fc9f6a9df105d700aa9.jpg", + "image_caption": [ + "Figure 7. Our model is capable of answering question correctly that needs accurate reasoning on spatial distance and 3D orientation, while GPT-4o either needs more spatial information or gives wrong answer." + ], + "image_footnote": [], + "bbox": [ + 519, + 571, + 549, + 597 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To determine which car is closer to the car in Region [0], we would need to analyze the spatial arrangement of the objects in the image. However, without the ability to visually assess the image, I can't provide a definitive answer. If you can describe the positions of the cars in Regions [1] and [2] relative to Region [0], I can help you deduce which one is closer.", + "bbox": [ + 557, + 558, + 759, + 613 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The car in Region [1].", + "bbox": [ + 620, + 618, + 697, + 628 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3d08707deb9c3b6aaac7cc8c8293544380e97eafa903793ea4d1ef455e79988f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 767, + 455, + 900, + 517 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Consider the real-world 3D orientations of the objects. Are the chair in Region [0] and the chair in Region [1] facing the same (or similar) directions, or very different directions?", + "bbox": [ + 771, + 522, + 898, + 570 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Based on the image, the chair in Region [0] and the chair in Region [1] face similar directions", + "bbox": [ + 771, + 580, + 895, + 606 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Very different directions.", + "bbox": [ + 787, + 618, + 877, + 626 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3D-informed training datasets focused on 3D locations, orientations and complex spatial relationship. Our systematic integration of these datasets with architectural and training designs provide a roadmap for optimizing LMMs for superior 3D reasoning capabilities. Experimental results on the SpatialVQA benchmark highlight SpatialLLM's superiority over state-of-the-art models, setting a new standard for 3D spatial reasoning. We invite the community to build upon our design roadmap and findings to push the boundaries of what LMMs can achieve in understanding the complex 3D spatial relationships of the real world.", + "bbox": [ + 511, + 717, + 906, + 883 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "17256", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Acknowledgements", + "text_level": 1, + "bbox": [ + 91, + 90, + 277, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "WM and AY acknowledges support from ONR with N00014-23-1-2641 and ARL award W911NF2320008.", + "bbox": [ + 89, + 114, + 482, + 146 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 161, + 187, + 176 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 1, 2, 3, 7", + "[2] Remyx AI. SpaceLLaVA. https://huggingface.co/remyxai/SpaceLLaVA.1,7", + "[3] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in neural information processing systems, 35:23716-23736, 2022. 3", + "[4] Anthropic. Claude 3.5 Sonnet. https://www.anthropic.com/news/claude-3-5-sonnet. 1, 7", + "[5] Apollo Team. Apollo synthetic dataset, 2019. 1", + "[6] Daichi Azuma, Taiki Miyanishi, Shuhei Kurita, and Motoaki Kawanabe. Scanqa: 3d question answering for spatial scene understanding. In proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 19129-19139, 2022. 4, 1", + "[7] Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, et al. Qwen technical report. arXiv preprint arXiv:2309.16609, 2023. 2, 3", + "[8] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966, 2023. 2", + "[9] Gilad Baruch, Zhuoyuan Chen, Afshin Dehghan, Tal Dimry, Yuri Feigin, Peter Fu, Thomas Gebauer, Brandon Joffe, Daniel Kurz, Arik Schwartz, et al. Arkitsscenes: A diverse real-world dataset for 3d indoor scene understanding using mobile rgb-d data. arXiv preprint arXiv:2111.08897, 2021. 4, 1", + "[10] Shariq Farooq Bhat, Reiner Birkl, Diana Wofk, Peter Wonka, and Matthias Müller. Zoedepth: Zero-shot transfer by combining relative and metric depth. arXiv preprint arXiv:2302.12288, 2023. 1, 2", + "[11] Garrick Brazil, Abhinav Kumar, Julian Straub, Nikhila Ravi, Justin Johnson, and Georgia Gkioxari. Omni3D: A large benchmark and model for 3D object detection in the wild. In CVPR, Vancouver, Canada, 2023. IEEE. 2, 4, 1", + "[12] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11621-11631, 2020. 4, 1" + ], + "bbox": [ + 93, + 186, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Junbum Cha, Wooyoung Kang, Jonghwan Mun, and Byungseok Roh. Honeybee: Locality-enhanced projector for multimodal llm. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13817-13827, 2024. 2, 3", + "[14] Boyuan Chen, Zhuo Xu, Sean Kirmani, Brain Ichter, Dorsa Sadigh, Leonidas Guibas, and Fei Xia. Spatialvlm: Endowing vision-language models with spatial reasoning capabilities. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14455-14465, 2024. 1, 2, 3, 4, 5, 7", + "[15] Jieneng Chen, Qihang Yu, Xiaohui Shen, Alan Yuille, and Liang-Chieh Chen. Vitamin: Designing scalable vision models in the vision-language era. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 2", + "[16] An-Chieh Cheng, Hongxu Yin, Yang Fu, Qiushan Guo, Ruihan Yang, Jan Kautz, Xiaolong Wang, and Sifei Liu. Spatialrgpt: Grounded spatial reasoning in vision language model. arXiv preprint arXiv:2406.01584, 2024. 1, 2, 3, 4, 5", + "[17] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 2, 3", + "[18] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 3, 6, 8", + "[19] Mohamed El Banani, Amit Raj, Kevis-Kokitsi Maninis, Abhishek Kar, Yuanzhen Li, Michael Rubinstein, Deqing Sun, Leonidas Guibas, Justin Johnson, and Varun Jampani. Probing the 3d awareness of visual foundation models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21795-21806, 2024. 4, 5, 7, 1", + "[20] Samir Yitzhak Gadre, Gabriel Ilharco, Alex Fang, Jonathan Hayase, Georgios Smyrnis, Thao Nguyen, Ryan Marten, Mitchell Wortsman, Dhruba Ghosh, Jieyu Zhang, et al. Dat-acomp: In search of the next generation of multimodal datasets. arXiv preprint arXiv:2304.14108, 2023. 3", + "[21] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In 2012 IEEE conference on computer vision and pattern recognition, pages 3354-3361. IEEE, 2012. 4, 1", + "[22] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Batra, and Devi Parikh. Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6904-6913, 2017. 3", + "[23] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 3, 5, 8", + "[24] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen." + ], + "bbox": [ + 516, + 92, + 905, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "17257", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 7", + "[25] Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. Openclip, 2021. If you use this software, please cite it as below. 7, 8", + "[26] Artur Jesslen, Guofeng Zhang, Angtian Wang, Wufei Ma, Alan Yuille, and Adam Kortylewski. Novum: Neural object volumes for robust object classification. In European Conference on Computer Vision, pages 264-281. Springer, 2024. 2", + "[27] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In ICML, 2021. 2", + "[28] Albert Q Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, et al. Mistral 7b. arXiv preprint arXiv:2310.06825, 2023. 2", + "[29] Linyi Jin, Jianming Zhang, Yannick Hold-Geoffroy, Oliver Wang, Kevin Blackburn-Matzen, Matthew Sticha, and David F Fouhey. Perspective fields for single image camera calibration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17307-17316, 2023. 1", + "[30] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dollar, and Ross Girshick. Segment anything. arXiv:2304.02643, 2023. 1", + "[31] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 1, 2", + "[32] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision, 123:32-73, 2017. 3", + "[33] Alina Kuznetsova, Hassan Rom, Neil Alldrin, Jasper Uijlings, Ivan Krasin, Jordi Pont-Tuset, Shahab Kamali, Stefan Popov, Matteo Malloci, Alexander Kolesnikov, et al. The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale. International journal of computer vision, 128(7):1956-1981, 2020. 5, 1", + "[34] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 2, 8", + "[35] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified" + ], + "bbox": [ + 91, + 90, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "vision-language understanding and generation. In ICML, 2022. 2, 3", + "[36] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730–19742. PMLR, 2023. 8", + "[37] Xianhang Li, Haoqin Tu, Mude Hui, Zeyu Wang, Bingchen Zhao, Junfei Xiao, Sucheng Ren, Jieru Mei, Qing Liu, Huangjie Zheng, et al. What if we recapture billions of web images with llama-3? arXiv preprint arXiv:2406.08478, 2024. 3", + "[38] Haotian Liu, Kilho Son, Jianwei Yang, Ce Liu, Jianfeng Gao, Yong Jae Lee, and Chunyuan Li. Learning customized visual models with retrieval-augmented knowledge. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15148-15158, 2023. 8", + "[39] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26296-26306, 2024. 1, 2, 3, 7, 8", + "[40] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava next: Improved reasoning,ocr,and world knowledge,2024.7", + "[41] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 1, 2, 3, 4, 5, 7", + "[42] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Qing Jiang, Chunyuan Li, Jianwei Yang, Hang Su, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023. 1", + "[43] Haoyu Lu, Wen Liu, Bo Zhang, Bingxuan Wang, Kai Dong, Bo Liu, Jingxiang Sun, Tongzheng Ren, Zhuoshu Li, Hao Yang, et al. Deepseek-vl: towards real-world vision-language understanding. arXiv preprint arXiv:2403.05525, 2024. 3", + "[44] Wufei Ma, Angtian Wang, Alan Yuille, and Adam Kortylewski. Robust category-level 6d pose estimation with coarse-to-fine rendering of neural features. In European Conference on Computer Vision, pages 492-508. Springer, 2022. 2", + "[45] Wufei Ma, Guanning Zeng, Guofeng Zhang, Qihao Liu, Letian Zhang, Adam Kortylewski, Yaoyao Liu, and Alan Yuille. Imagenet3d: Towards general-purpose object-level 3d understanding. In NeurIPS, 2024. 2, 4, 5, 7, 1", + "[46] Brandon McKinzie, Zhe Gan, Jean-Philippe Fauconnier, Sam Dodge, Bowen Zhang, Philipp Dufter, Dhruti Shah, Xianzhi Du, Futang Peng, Floris Weers, et al. Mm1: Methods, analysis & insights from multimodal lIm pre-training. arXiv preprint arXiv:2403.09611, 2024. 3", + "[47] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. TMLR, 2024. 2, 3, 5, 6, 8" + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "17258", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[48] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 2, 3, 5, 6, 8", + "[49] Navid Rajabi and Jana Kosecka. Gsr-bench: A benchmark for grounded spatial reasoning evaluation via multimodal llms. arXiv preprint arXiv:2406.13246, 2024. 2", + "[50] Mike Roberts, Jason Ramapuram, Anurag Ranjan, Atulit Kumar, Miguel Angel Bautista, Nathan Paczan, Russ Webb, and Joshua M. Susskind. Hypersim: A photorealistic synthetic dataset for holistic indoor scene understanding. In International Conference on Computer Vision (ICCV) 2021, 2021. 4, 1", + "[51] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 3", + "[52] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278-25294, 2022. 3", + "[53] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In ACL, 2018. 2, 3", + "[54] Shuran Song, Samuel P Lichtenberg, and Jianxiong Xiao. Sun rgb-d: A rgb-d scene understanding benchmark suite. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 567-576, 2015. 4, 1", + "[55] Elizabeth S Spelke and Katherine D Kinzler. Core knowledge. Developmental science, 10(1):89-96, 2007. 1", + "[56] Chen Sun, Abhinav Shrivastava, Saurabh Singh, and Abhinav Gupta. Revisiting unreasonable effectiveness of data in deep learning era. In Proceedings of the IEEE international conference on computer vision, pages 843-852, 2017. 3", + "[57] Gemini Team, Rohan Anil, Sebastian Borgeaud, Yonghui Wu, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, et al. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805, 2023. 1, 2", + "[58] Shengbang Tong, Ellis Brown, Penghao Wu, Sanghyun Woo, Manoj Middepogu, Sai Charitha Akula, Jihan Yang, Shusheng Yang, Adithya Iyer, Xichen Pan, et al. Cambrian-1: A fully open, vision-centric exploration of multimodal llms. arXiv preprint arXiv:2406.16860, 2024. 1, 3, 4, 7", + "[59] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, Dan Bikel, Lukas Blecher, Cristian Canton Ferrer, Moya Chen, Guillem Cucurull, David Esiobu, Jude Fernandes, Jeremy Fu, Wenyin Fu, Brian Fuller, Cynthia Gao, Vedanuj Goswami, Naman Goyal, Anthony Hartshorn, Saghar Hosseini, Rui Hou, Hakan Inan, Marcin Kardas, Vik" + ], + "bbox": [ + 89, + 90, + 485, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "tor Kerkez, Madian Khabsa, Isabel Kloumann, Artem Korenev, Punit Singh Koura, Marie-Anne Lachaux, Thibaut Lavril, Jenya Lee, Diana Liskovich, Yinghai Lu, Yuning Mao, Xavier Martinet, Todor Mihaylov, Pushkar Mishra, Igor Molybog, Yixin Nie, Andrew Poulton, Jeremy Reizenstein, Rashi Rungta, Kalyan Saladi, Alan Schelten, Ruan Silva, Eric Michael Smith, Ranjan Subramanian, Xiaoqing Ellen Tan, Binh Tang, Ross Taylor, Adina Williams, Jian Xiang Kuan, Puxin Xu, Zheng Yan, Iliyan Zarov, Yuchen Zhang, Angela Fan, Melanie Kambadur, Sharan Narang, Aurelien Rodriguez, Robert Stojnic, Sergey Edunov, and Thomas Scialom. Llama 2: Open foundation and finetuned chat models. arXiv preprint arXiv:2307.09288, 2023. 2,3", + "[60] Xingrui Wang, Wufei Ma, Zhuowan Li, Adam Kortylewski, and Alan L Yuille. 3d-aware visual question answering about parts, poses and occlusions. Advances in Neural Information Processing Systems, 36, 2024. 2, 3", + "[61] Xingrui Wang, Wufei Ma, Angtian Wang, Shuo Chen, Adam Kortylewski, and Alan Yuille. Compositional 4d dynamic scenes understanding with physics priors for video question answering. arXiv preprint arXiv:2406.00622, 2024. 2", + "[62] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. In CVPR, 2024. 1", + "[63] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10371-10381, 2024. 1, 2", + "[64] Shuquan Ye, Dongdong Chen, Songfang Han, and Jing Liao. 3d question answering. arXiv preprint arXiv:2112.08359, 2021. 4, 1", + "[65] Jiahui Yu, Zirui Wang, Vijay Vasudevan, Legg Yeung, Mojtaba Seyedhosseini, and Yonghui Wu. Coca: Contrastive captioners are image-text foundation models. arXiv preprint arXiv:2205.01917, 2022. 3", + "[66] Youcai Zhang, Xinyu Huang, Jinyu Ma, Zhaoyang Li, Zhaochuan Luo, Yanchun Xie, Yuzhuo Qin, Tong Luo, Yaqian Li, Shilong Liu, et al. Recognize anything: A strong image tagging model. arXiv preprint arXiv:2306.03514, 2023.1", + "[67] Youcai Zhang, Xinyu Huang, Jinyu Ma, Zhaoyang Li, Zhaochuan Luo, Yanchun Xie, Yuzhuo Qin, Tong Luo, Yaqian Li, Shilong Liu, et al. Recognize anything: A strong image tagging model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1724-1732, 2024. 1", + "[68] Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric Xing, et al. Judging llm-as-a-judge with mt-bench and chatbot arena. Advances in Neural Information Processing Systems, 36:46595-46623, 2023. 6, 8", + "[69] Jinghao Zhou, Chen Wei, Huiyu Wang, Wei Shen, Cihang Xie, Alan Yuille, and Tao Kong. ibot: Image bert pre-training with online tokenizer. arXiv preprint arXiv:2111.07832, 2021.3" + ], + "bbox": [ + 514, + 92, + 906, + 898 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "17259", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 10 + }, + { + "type": "ref_text", + "text": "[70] Shengjie Zhu, Abhinav Kumar, Masa Hu, and Xiaoming Liu. Tame a wild camera: In-the-wild monocular camera calibration. In NeurIPS, 2023. 1", + "bbox": [ + 91, + 90, + 482, + 133 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "17260", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2025/SpatialLLM_ A Compound 3D-Informed Design towards Spatially-Intelligent Large Multimodal Models/cf47bc9d-8b67-4bc0-a3fe-ba2e3465761d_model.json b/2025/SpatialLLM_ A Compound 3D-Informed Design towards Spatially-Intelligent Large Multimodal Models/cf47bc9d-8b67-4bc0-a3fe-ba2e3465761d_model.json new file mode 100644 index 0000000000000000000000000000000000000000..c490d57e94f4fe9cab759daba04688c8d951fd93 --- /dev/null +++ b/2025/SpatialLLM_ A Compound 3D-Informed Design towards Spatially-Intelligent Large Multimodal Models/cf47bc9d-8b67-4bc0-a3fe-ba2e3465761d_model.json @@ -0,0 +1,2952 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.212, + 0.131, + 0.787, + 0.177 + ], + "angle": 0, + "content": "SpatialLLM: A Compound 3D-Informed Design towards Spatially-Intelligent Large Multimodal Models" + }, + { + "type": "text", + "bbox": [ + 0.2, + 0.203, + 0.796, + 0.222 + ], + "angle": 0, + "content": "Wufei Ma1, Luoxin Ye1, Celso M de Melo2, Alan Yuille1, Jieneng Chen1" + }, + { + "type": "text", + "bbox": [ + 0.29, + 0.224, + 0.704, + 0.24 + ], + "angle": 0, + "content": "\\(^{1}\\)Johns Hopkins University \\(^{2}\\)DEVCOM Army Research Laboratory" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.274, + 0.327, + 0.29 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.306, + 0.485, + 0.683 + ], + "angle": 0, + "content": "Humans naturally understand 3D spatial relationships, enabling complex reasoning like predicting collisions of vehicles from different directions. Current large multimodal models (LMMs), however, lack of this capability of 3D spatial reasoning. This limitation stems from the scarcity of 3D training data and the bias in current model designs toward 2D data. In this paper, we systematically study the impact of 3D-informed data, architecture, and training setups, introducing SpatialLLM, a large multi-modal model with advanced 3D spatial reasoning abilities. To address data limitations, we develop two types of 3D-informed training datasets: (1) 3D-informed probing data focused on object's 3D location and orientation, and (2) 3D-informed conversation data for complex spatial relationships. Notably, we are the first to curate VQA data that incorporate 3D orientation relationships on real images. Furthermore, we systematically integrate these two types of training data with the architectural and training designs of LMMs, providing a roadmap for optimal design aimed at achieving superior 3D reasoning capabilities. Our SpatialLLM advances machines toward highly capable 3D-informed reasoning, surpassing GPT-4o performance by \\(8.7\\%\\). Our systematic empirical design and the resulting findings offer valuable insights for future research in this direction. Our project page is available at this link." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.71, + 0.222, + 0.725 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.483, + 0.87 + ], + "angle": 0, + "content": "When humans observe a scene, they perceive more than isolated objects; they intuitively understand the spatial relationships among these objects [55], enabling complex reasoning and interaction with the environment. For example, recognizing that a car is approaching an obstacle and might collide involves interpreting the 3D object orientations and spatial relationships. This innate ability to comprehend and reason about 3D space is crucial for navigating and interacting with the physical world." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Despite significant advancements, current LMMs [1, 4, 39, 41, 57], exhibit limited capability of 3D spatial reason" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.272, + 0.773, + 0.348 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.538, + 0.356, + 0.6, + 0.373 + ], + "angle": 0, + "content": "Orientation Relationships" + }, + { + "type": "text", + "bbox": [ + 0.521, + 0.379, + 0.614, + 0.406 + ], + "angle": 0, + "content": "White van is facing the pedestrian instead of the cyclist" + }, + { + "type": "image_caption", + "bbox": [ + 0.63, + 0.355, + 0.715, + 0.373 + ], + "angle": 0, + "content": "Spatial Reasoning Relationships" + }, + { + "type": "text", + "bbox": [ + 0.627, + 0.379, + 0.766, + 0.406 + ], + "angle": 0, + "content": "From the perspective of the pedestrian, the white van is on the right side of the pedestrian" + }, + { + "type": "image_caption", + "bbox": [ + 0.776, + 0.275, + 0.887, + 0.327 + ], + "angle": 0, + "content": "Distance Relationships The pedestrian is closer to the viewer than the cyclist." + }, + { + "type": "image", + "bbox": [ + 0.777, + 0.333, + 0.889, + 0.378 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.412, + 0.574, + 0.466 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.582, + 0.411, + 0.667, + 0.42 + ], + "angle": 0, + "content": "3D-Informed Data" + }, + { + "type": "text", + "bbox": [ + 0.582, + 0.424, + 0.658, + 0.468 + ], + "angle": 0, + "content": "1M 3D-Informed Probing Data description on depth, distance, & orientation" + }, + { + "type": "text", + "bbox": [ + 0.661, + 0.424, + 0.768, + 0.468 + ], + "angle": 0, + "content": "1M 3D-Informed Instruction Tuning Data conversations about 3D spatial relationships" + }, + { + "type": "image_caption", + "bbox": [ + 0.775, + 0.411, + 0.897, + 0.42 + ], + "angle": 0, + "content": "3D-Informed Architecture" + }, + { + "type": "text", + "bbox": [ + 0.776, + 0.424, + 0.897, + 0.454 + ], + "angle": 0, + "content": "Mixed encoders, 3D-aware visual encoders, etc. 3D-Informed Training" + }, + { + "type": "text", + "bbox": [ + 0.777, + 0.458, + 0.9, + 0.477 + ], + "angle": 0, + "content": "Pre-pretraining, alignment & finetuning" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.489, + 0.907, + 0.572 + ], + "angle": 0, + "content": "Figure 1. 3D spatial reasoning is crucial for LMMs to ground objects in 3D space and infer their 3D spatial relationships, such as distance, orientation, and spatial interactions. We systematically study the impact of 3D-informed data, architecture, and training setups, introducing SpatialLLM, an LMM with advanced 3D spatial reasoning abilities." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.607, + 0.906, + 0.667 + ], + "angle": 0, + "content": "ing. These models excel at identifying objects and generating descriptive captions but struggle with tasks that require intricate 3D reasoning, such as predicting object interactions or understanding physical events." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.671, + 0.907, + 0.837 + ], + "angle": 0, + "content": "The primary challenge lies in the scarcity of high-quality 3D spatial relationship data, which prevents models from learning effective 3D spatial reasoning. Collecting a small set of high-quality 3D-aware data to tackle the first challenge is feasible, albeit labor-intensive, using readily available tools. Recent efforts [2, 14] leverage vision foundation models to recognize [67], detect [42], segment [31] and estimate the depth [10, 63] in images. These approaches form semi-automated pipelines to construct pseudo-2.5D datasets for training LMMs with spatial reasoning capabilities, such as depth ordering [14, 58]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.907, + 0.901 + ], + "angle": 0, + "content": "However, a significant gap remains: previous works [2, 14, 16] have primarily focused on 3D distance relationships, overlooking the crucial role of 3D object orientation. This is primarily due to the absence of robust 6D pose estimation." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17249" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.198 + ], + "angle": 0, + "content": "tors. Understanding object orientation is essential for complex spatial reasoning tasks, such as determining if two cars are facing the same direction to avoid dangerous collision. Addressing this gap, we aim to incorporate 3D orientation relationships—converted from ImageNet3D [17]—into our data engine, making us the first to enable complex spatial reasoning involving 3D orientation relationships." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.199, + 0.482, + 0.32 + ], + "angle": 0, + "content": "Even with 3D-informed data in hand, fully unlocking its potential remains challenging due to the intricate, multistage development process required for an LMM. For example, SpatialVLM [14] utilizes 3D-informed data during the instruction tuning stage, but it remains uncertain whether such data can enhance other stages like vision encoder pretraining or multimodal alignment. This limitation suggests that a more holistic approach is necessary." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.322, + 0.482, + 0.382 + ], + "angle": 0, + "content": "In this paper, we present a systematic investigation into building LMMs with enhanced 3D spatial reasoning capabilities through a compound design approach, as illustrated in Fig. 1. Our key contributions are threefold:" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.383, + 0.482, + 0.504 + ], + "angle": 0, + "content": "First, we identify three fundamental types of 3D spatial relationships that LMMs must master: distance relationships, orientation relationships, and complex spatial reasoning that combines both. To evaluate these capabilities, we introduce SpatialVQA, a comprehensive benchmark containing 1,323 questions that target 3D spatial understanding, built upon diverse urban and indoor scenes from Omni3D [11]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.506, + 0.482, + 0.672 + ], + "angle": 0, + "content": "Second, we propose a novel compound 3D-informed design that introduces improvements across multiple dimensions, leading to our proposed SpatialLLM model. On the data front, we develop two types of 3D-informed training datasets: (1) 3D-informed probing data focused on object-level 3D properties, and (2) 3D-informed conversation data for complex spatial relationships. Architecturally, we explore mixing different visual encoders and investigate their 3D awareness capabilities. For training, we introduce 3D-informed data at multiple stages: pre-pretraining, multimodal alignment, and instruction tuning." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.673, + 0.482, + 0.853 + ], + "angle": 0, + "content": "Third, we present the first comprehensive search over the LMM design space for spatial reasoning tasks and propose a roadmap towards developing state-of-the-art models in this domain. The experiment demonstrates that our proposed SpatialLLM achieves state-of-the-art performance on spatial reasoning tasks, surpassing both GPT-4o [1] (by \\(8.7\\%\\) ) and spatial-centric models like SpatialVLM [14] (by \\(10.5\\%\\)). Our systematic analysis reveals that while architectural improvements contribute a \\(1.3\\%\\) gain, the strategic incorporation of 3D-informed data across different training stages yields a substantial \\(13.7\\%\\) performance improvement." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.482, + 0.901 + ], + "angle": 0, + "content": "Our findings highlight the importance of a holistic approach to improving LMMs' 3D spatial reasoning capabilities, demonstrating that both architectural choices and care" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.166 + ], + "angle": 0, + "content": "fully curated open-source 3D-informed training data play crucial roles in achieving superior performance. This work establishes a new benchmark for spatial reasoning in LMMs and provides valuable insights for future research in this direction." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.181, + 0.664, + 0.198 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.206, + 0.905, + 0.463 + ], + "angle": 0, + "content": "Large multimodal models (LMMs). Recently, there has been a significant surge in research on multimodal LLMs, exemplified by products like GPT-4 [1] and Gemini [57], as well as open-source models such as LLaVA [34, 39, 41] and Qwen-VL [8]. These open-source models combine pretrained vision encoders [15, 27, 47, 48] with LLMs [7, 28, 59], enabling the language models to handle a wide range of tasks that involve images as input. Typically, these methods have the alignment stage that aligns image features with the input embedding space of the language model using connectors [13, 35, 39], trained with paired images and captions [53]. The instruction fine-tuning [41] stage uses data that contains both images and instructions to enhances LMMs. However, current LMMs are unsuccessful in spatial visual reasoning, due to limited spatial-aware data and stagnant 3D-informed architecture design. We aim to bridge this research gap." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.464, + 0.905, + 0.901 + ], + "angle": 0, + "content": "3D spatial reasoning with LMMs. Understanding spatial relationships between objects in images is crucial for visual reasoning. While this comes naturally to humans, LMMs were found to struggle with recognizing spatial relations [49, 60, 61], due to a lack of 3D-aware spatial-related data during training. While neural symbolic methods [60] achieved improved performance on synthetic benchmarks by involving explicit 3D reasoning modules [26, 44], they struggled to generalize to a broader range of objects found in real images. Fortunately, with foundation models like SAM [31] and depth estimators [10, 63], semi-automatic annotation for 3D-aware VQA data has been adopted in recent studies [14, 16]. SpatialVLM [14] uses the pseudo-annotated VQA data for training only while SpatialRGPT [14, 16] further augments LLaVA with extra input of regional mask and depth map during inference, which is not our focus. Although human-annotated 3D-aware VQA data is rare, we aim to leverage human-annotated 3D object-level datasets (like ImageNet3D [45]) by simply converting image-pose pairs into image-text pair data. However, even with the collected data, a major challenge remains in effectively utilizing 3D-aware data to train optimal models. Current approaches like SpatialVLM [14] use the standard LLaVA and two-stage training. To date, no work has addressed when (which training stage) and where (which network component) are optimal to inject the 3D-awareness to LMMs. We bridge this gap by conducting a compound design that simultaneously considers 3D-aware data, architecture, and the training, leading to a best" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "17250" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.138 + ], + "angle": 0, + "content": "performing design. Rather than competing, our study has the potential to complement prior works [14, 16] by offering new compound design recipe." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.155, + 0.191, + 0.17 + ], + "angle": 0, + "content": "3. Methods" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.181, + 0.483, + 0.287 + ], + "angle": 0, + "content": "In this section we start by reviewing LMMs in Sec. 3.1, specifically the architecture and the choice of training data and strategies. We present the task of reasoning 3D spatial relationships and explain the challenges LMMs face when answering these questions in Sec. 3.2. Then we study how to address these challenging problems with LMMs and introduce our SpatialLLM in Sec. 3.3." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.3, + 0.296, + 0.317 + ], + "angle": 0, + "content": "3.1. Preliminary of LMMs" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.324, + 0.483, + 0.368 + ], + "angle": 0, + "content": "Data. The success of deep learning in vision is largely due to well-curated data [20, 56], and LMMs similarly depend on this foundation, albeit with diverse data formats below." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.37, + 0.483, + 0.431 + ], + "angle": 0, + "content": "- Image-only data, such as ImageNet [17], is commonly used for MAE [23]/DINOv2 [47] pretraining with a self-supervised objective. This step enables the model to learn rich visual features solely from visual signals." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.431, + 0.483, + 0.476 + ], + "angle": 0, + "content": "- Noisy image-text pairs: Large-scale image-text pairs [20, 37, 52] are used for pretraining with contrastive loss (e.g., CLIP [48, 65]) to align visual and textual representations." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.477, + 0.483, + 0.567 + ], + "angle": 0, + "content": "- Multimodal alignment: At this stage, image-caption pairs (e.g., CC3M [53]) are utilized to align visual features with the language model's word embedding space, typically through image-captioning tasks. This pretraining ensures that the model can interpret and relate visual content to textual queries in a shared semantic space." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.567, + 0.483, + 0.642 + ], + "angle": 0, + "content": "- Multimodal instruction-following data. Visual instruction tuning data is essential but challenging to collect, as it rarely exists in its natural form online. Prior works [39, 41, 46, 58] repurpose VQA benchmarks [22, 32] into instruction-tuning datasets." + }, + { + "type": "list", + "bbox": [ + 0.092, + 0.37, + 0.483, + 0.642 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.643, + 0.483, + 0.748 + ], + "angle": 0, + "content": "Architecture. A standard LMM [39, 41] consists of a visual encoder to process the image, a multimodal connector to transform the visual feature to visual token, and a LLM for reasoning. However, the extent to which these architectural components contribute to the spatial understanding of LMMs remains unknown. We review these architectural components as below." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.75, + 0.483, + 0.826 + ], + "angle": 0, + "content": "- Vision encoder: Most LMMs rely on language-supervised models like CLIP [48], leveraging web-scale noisy image-text data. Recent LMMs [43, 58] integrate vision models trained only on visual signals, including self-supervised [23, 47, 69] and generative models [51]." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.826, + 0.483, + 0.871 + ], + "angle": 0, + "content": "- Connector: Features from a visual encoder are mapped into the LLM token space by connectors [3, 13, 35, 41] such as a MLP [39, 41]." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.871, + 0.483, + 0.902 + ], + "angle": 0, + "content": "- LLMs: LLMs [1, 7, 59] serve as the foundational backbone of LMMs, driving their reasoning capabilities." + }, + { + "type": "list", + "bbox": [ + 0.092, + 0.75, + 0.483, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.527, + 0.092, + 0.905, + 0.122 + ], + "angle": 0, + "content": "Open-source LLMs like Llama series [18, 59] are frequently adopted to enhance the reasoning capability." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.123, + 0.905, + 0.198 + ], + "angle": 0, + "content": "Training. The aforementioned data supports pretraining for strong visual representation learning, alignment pretraining to synchronize visual and language representations, and instruction tuning to enable multimodal instruction-following and reasoning capabilities" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.199, + 0.905, + 0.274 + ], + "angle": 0, + "content": "- Pre-pretraining. This stage focuses on developing foundational visual representations, often with reconstruction-based objectives (e.g., MAE [23], DINOv2 [47]). This step prepares the model to handle more complex multimodal tasks by first refining its visual understanding." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.275, + 0.905, + 0.319 + ], + "angle": 0, + "content": "- Multimodal alignment pretraining. At this stage, the model is trained to describe images in details to align visual and language representations in the same space." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.32, + 0.905, + 0.41 + ], + "angle": 0, + "content": "- Visual instruction tuning. The final stage involves refining the model's ability to process complex multimodal instructions using visual instruction tuning data. This often incorporates transformed VQA datasets, enhancing the model's capacity in handling tasks that demand coordinated reasoning across language and vision." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.199, + 0.905, + 0.41 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.421, + 0.822, + 0.437 + ], + "angle": 0, + "content": "3.2. Reasoning 3D Spatial Relationships" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.444, + 0.907, + 0.595 + ], + "angle": 0, + "content": "3D spatial reasoning analyzes the spatial relationships between objects in the 3D world space, such as objects being far or close to each other, or one object lies to the left or right of another object [60]. It extends beyond standard 3D tasks, such as 3D object detection and 3D pose estimation that studies the 3D nature of individual objects, and focuses on understanding their mutual relationships in 3D. In this work we identify three main types of 3D spatial relationships, i.e., distance, orientation, and spatial reasoning, and study how we can better address these problems with LMMs." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.616, + 0.907, + 0.827 + ], + "angle": 0, + "content": "3D spatial relationships. For distance spatial relationships, we study if models can estimate and compare 3D distances between objects. These spatial relationships require more than depth awareness, which gives the distance between the viewer and the object and can be roughly estimated from the object scale, but also the 3D distances between two objects that are often harder to estimate. On the other hand, orientation spatial relationships require models to understand the 3D orientations of an object, e.g., predicting if two objects are facing the same direction. Lastly, spatial reasoning spatial relationships look into 3D spatial relationships that need a combination of 3D awareness of location and orientation and then performing spatial reasoning on the 3D information." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.836, + 0.8, + 0.852 + ], + "angle": 0, + "content": "3.2.1. Challenges of 3D spatial reasoning" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Reasoning about 3D spatial relationships remains a challenging task for state-of-the-art open-sourced and proprietary LMMs. We attribute this to two main limitations in" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "17251" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.116, + 0.089, + 0.283, + 0.184 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.312, + 0.098, + 0.46, + 0.108 + ], + "angle": 0, + "content": "√3D Location 03D Orientation" + }, + { + "type": "image_caption", + "bbox": [ + 0.328, + 0.112, + 0.348, + 0.121 + ], + "angle": 0, + "content": "User" + }, + { + "type": "image_caption", + "bbox": [ + 0.318, + 0.124, + 0.488, + 0.148 + ], + "angle": 0, + "content": "Which object is closer to the chair in Region [0], the bookcase in Region [1] or the chair in Region [2]?" + }, + { + "type": "image_caption", + "bbox": [ + 0.444, + 0.153, + 0.488, + 0.162 + ], + "angle": 0, + "content": "3DI-LMM" + }, + { + "type": "image_caption", + "bbox": [ + 0.319, + 0.165, + 0.479, + 0.182 + ], + "angle": 0, + "content": "The chair in Region [2] is closer to the chair in Region [0]." + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.188, + 0.305, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.312, + 0.19, + 0.488, + 0.208 + ], + "angle": 0, + "content": "Spatial Reasoning Relationships \\(\\sqrt{3}\\) Location \\(\\sqrt{3}\\) Orientation" + }, + { + "type": "image_caption", + "bbox": [ + 0.328, + 0.211, + 0.37, + 0.22 + ], + "angle": 0, + "content": "User" + }, + { + "type": "image_caption", + "bbox": [ + 0.319, + 0.223, + 0.483, + 0.248 + ], + "angle": 0, + "content": "If I stand at the location of the car in Region [0] facing where it is facing, where is the bus in Region [1]?" + }, + { + "type": "image_caption", + "bbox": [ + 0.444, + 0.253, + 0.488, + 0.261 + ], + "angle": 0, + "content": "3DI-LMM" + }, + { + "type": "image_caption", + "bbox": [ + 0.319, + 0.264, + 0.478, + 0.281 + ], + "angle": 0, + "content": "The bus in Region [1] is on the left side of the car in Region [0]." + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.09, + 0.714, + 0.184 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.721, + 0.098, + 0.869, + 0.107 + ], + "angle": 0, + "content": "√3D Location √3D Orientation" + }, + { + "type": "image_caption", + "bbox": [ + 0.736, + 0.12, + 0.754, + 0.128 + ], + "angle": 0, + "content": "Use" + }, + { + "type": "image_caption", + "bbox": [ + 0.727, + 0.131, + 0.899, + 0.149 + ], + "angle": 0, + "content": "Are the car in Region [0] and the car in Region [1] facing the same direction, or different directions?" + }, + { + "type": "image_caption", + "bbox": [ + 0.728, + 0.165, + 0.887, + 0.182 + ], + "angle": 0, + "content": "The car in Region [0] and the car in Region [1] are facing different directions." + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.188, + 0.714, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.721, + 0.191, + 0.854, + 0.199 + ], + "angle": 0, + "content": "Spatial Reasoning Relationships" + }, + { + "type": "image_caption", + "bbox": [ + 0.721, + 0.199, + 0.869, + 0.207 + ], + "angle": 0, + "content": "√3D Location √3D Orientation" + }, + { + "type": "image_caption", + "bbox": [ + 0.736, + 0.219, + 0.757, + 0.227 + ], + "angle": 0, + "content": "User" + }, + { + "type": "image_caption", + "bbox": [ + 0.727, + 0.231, + 0.895, + 0.248 + ], + "angle": 0, + "content": "Is the car in Region [0] oriented towards the car in Region [1] or the car in Region [2]?" + }, + { + "type": "image_caption", + "bbox": [ + 0.853, + 0.253, + 0.897, + 0.26 + ], + "angle": 0, + "content": "3DI-LMM" + }, + { + "type": "image_caption", + "bbox": [ + 0.727, + 0.264, + 0.895, + 0.281 + ], + "angle": 0, + "content": "The car in Region [0] is facing towards the car in Region [1]." + }, + { + "type": "image_caption", + "bbox": [ + 0.13, + 0.296, + 0.865, + 0.311 + ], + "angle": 0, + "content": "Figure 2. Examples from our SpatialVQA benchmark featuring a broad range of questions that require 3D spatial reasoning." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.337, + 0.483, + 0.368 + ], + "angle": 0, + "content": "modern LMMs, the lack of 3D awareness and inability to perform 3D spatial reasoning." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.387, + 0.484, + 0.538 + ], + "angle": 0, + "content": "3D awareness refers to the ability of a visual encoder to represent 3D-aware features such as 3D object shapes and 3D orientations. Previous studies resorted to proxy tasks, e.g., part correspondence [19] and pose estimation [45], to quantitatively evaluate the 3D awareness of visual foundation models with linear probing. Results showed that most visual encoder, specifically ones trained with language supervision, demonstrated limited 3D awareness of 3D object poses [45], which inevitably limits the 3D spatial reasoning capabilities of LMMs built upon these encoders." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.557, + 0.484, + 0.754 + ], + "angle": 0, + "content": "Inability to perform 3D spatial reasoning. To infer the 3D spatial relationships, models must also be able to perform 3D spatial reasoning from 3D-aware visual features. Despite the abundant Internet-scale data for multi-modal training, we argue that high-quality 3D spatial reasoning data are scarce. Existing pretraining and visual instruction tuning data for LMMs [41, 58] focused on detailed descriptions and conversations about scenes, appearances, and actions, while being vague about the 3D spatial relationships that build up the high-level physical events. Despite their success on a wide range of image understanding tasks, these LMMs fail to learn the nature of 3D scenes or to perform 3D spatial reasoning." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.762, + 0.327, + 0.776 + ], + "angle": 0, + "content": "3.2.2. SpatialVQA for Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Despite the rising interests in spatial reasoning, there lacks a comprehensive open source benchmark for reasoning 3D spatial relationships. Early datasets on spatial reasoning were built on 3D scans rather than images [6, 64]. Recent visual-language benchmarks on spatial reasoning either focused on 2D spatial relationships [14, 16], e.g., left or right in the image plane, or only on distance spatial relationships [58]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.337, + 0.907, + 0.579 + ], + "angle": 0, + "content": "To enable a thorough evaluation of the 3D spatial reasoning capabilities of LMMs, we introduce SpatialVQA, a visual question answering benchmark that covers a wide variety of 3D spatial relationships. Our SpatialVQA distinguishes itself from all previous spatial reasoning benchmarks in the sense that all questions require different levels of 3D awareness and cannot be answered from 2D spatial reasoning only. Specifically, SpatialVQA consists of 1,323 questions, ranging from questions can be answered from distance-awareness only, e.g., which object is closer to a third object, to questions that require estimating objects' 3D orientations, e.g., if two objects are facing the same direction, and eventually questions that require complex 3D spatial reasoning over a combination of object 3D locations and orientations, e.g., an object is in the front/left/back/right direction of another object." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.582, + 0.906, + 0.688 + ], + "angle": 0, + "content": "We build our SpatialVQA on images from Omni3D [11], with 3D bounding box annotations on diverse objects from both urban [12, 21] and indoor scenes [9, 50, 54]. We follow [16, 58] and develop rule-based methods to generate visual question-answer pairs from the 3D groundtruths. Please refer to Fig. 2 for some examples of SpatialVQA and to appendix Sec.B for details about our benchmark." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.709, + 0.798, + 0.725 + ], + "angle": 0, + "content": "3.3. Compound 3D-Informed Design" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Problem formulation. While previous approaches focused on generating 3D-informed instruction tuning data to enable LMMs with the abilities to estimate distances, depths, or spatial relationships [14, 16], it is largely understudied the best recipe for developing LMMs with strong 3D spatial reasoning capabilities. Motivated by the two aforementioned limitations of LMMs in Sec. 3.2.1, we consider two main aspects in our compound 3D-informed design – the architecture design that leads to visual encoders with strong 3D awareness and the training setup that advances the 3D spatial reasoning capabilities of LMMs." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17252" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.091, + 0.089, + 0.482, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.104, + 0.22, + 0.148, + 0.232 + ], + "angle": 0, + "content": "(a) data" + }, + { + "type": "image_caption", + "bbox": [ + 0.182, + 0.22, + 0.276, + 0.233 + ], + "angle": 0, + "content": "(b) training stage" + }, + { + "type": "image_caption", + "bbox": [ + 0.291, + 0.22, + 0.374, + 0.233 + ], + "angle": 0, + "content": "(c) architecture" + }, + { + "type": "image_caption", + "bbox": [ + 0.384, + 0.22, + 0.483, + 0.233 + ], + "angle": 0, + "content": "(d) vision encoder" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.248, + 0.485, + 0.318 + ], + "angle": 0, + "content": "Figure 3. Design space for LMMs capable of spatial reasoning. The dashed boxes and lines highlight our new design space compared to LLaVA-v1.5. This compound design simultaneously considers 3D-informed data, architecture, and training methods to search for the best-performing models for spatial reasoning." + }, + { + "type": "image", + "bbox": [ + 0.096, + 0.334, + 0.283, + 0.513 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.293, + 0.334, + 0.48, + 0.512 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.528, + 0.484, + 0.653 + ], + "angle": 0, + "content": "Figure 4. Comparison of data cards showcasing the curation process and data types: standard LLaVA data (left) vs. our 3D-informed data (right). For our 3D-informed data curation, we label depth and distance of objects in OpenImage [33] using semi-automatic tools, resulting in 3DI-Pb-OI1M. Beyond this, we introduce 3D orientation probing data derived from human-annotated ImageNet3D [45], resulting in 3DI-Pb-IN166K for alignment. We further augment the conversation about spatial relationship, leading to 3DI-Ft-1M for instruction tuning." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.681, + 0.228, + 0.696 + ], + "angle": 0, + "content": "3.3.1. Design space" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.7, + 0.484, + 0.762 + ], + "angle": 0, + "content": "We introduce the design space considered in our work, i.e., choices of training data, model architecture, and training setup that advance the 3D spatial reasoning capabilities of LMMs. Refer to Fig. 3 for an overview of our design space." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Data. Standard LMMs follow LLaVA [41] and adopte feature alignment data for connector pretraining and visual instruction data for supervised instruction tuning. These datasets focus on detailed descriptions of images or machine-generated multi-turn conversations. Besides standard visual-language datasets, we further consider 3D-informed data, including 3D-aware probing data and 3D-informed instruction tuning data. Inspired by [45], the 3D" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.242 + ], + "angle": 0, + "content": "aware probing data involves fundamental object-level 3D properties, such as the azimuth and elevation 3D viewpoint of an object, depth of the object from the viewer, and 3D distances between two objects. The goal of the 3D-aware probing data is to effectively inject fundamental 3D awareness into the models and benefit subsequent reasoning. Meanwhile, the 3D-informed instruction tuning data focuses on high-level 3D spatial relationships between the objects, which are built on a combination of the fundamental 3D properties and 3D spatial reasoning." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.243, + 0.907, + 0.41 + ], + "angle": 0, + "content": "To generate the 3D-aware probing data and the 3D-informed instruction tuning data, we follow the previous data generation pipelines [14, 16] and exploit visual foundation models to extract and tag the objects in 3D space. Furthermore, we adopt a pretrained 6D pose estimator in [45] and estimate 3D orientations of the objects, which provides the crucial 3D information for generating 3D-informed data regarding various orientation-related 3D spatial relationships. We illustrate our 3D-informed data generation pipeline in Fig. 4 and refer the readers to appendex Sec.A for details of our 3D-informed data generation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.43, + 0.909, + 0.568 + ], + "angle": 0, + "content": "Architecture. Inspired by Probe3D [19] which probes the 3D-awareness of visual foundation models, we study the design of visual encoder to enable better 3D spatial reasoning. We consider two types of visual encoders: (i) Frozen & pretrained visual encoder CLIP [48] following [41] but with the option to mix a wider variety of models, such as MAE [23] and DINOv2 [47], and (ii) 3D-aware visual encoders finetuned with our 3D-informed data, e.g., contrastive pretraining or finetuning with LoRA using language modeling loss." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.587, + 0.909, + 0.768 + ], + "angle": 0, + "content": "Training setup. To address the limitations of 3D-aware LMMs discussed in Sec. 3.2.1, we propose new training setups that aim to improve 3D awareness and advance the 3D spatial reasoning capabilities. We consider the following: (i) in Stage 0: use 3D-informed probing data (3DI-Pb-OI1M/1N166K) to tune the Lora layers of CLIP vision encoder (Pre-pretraining); (ii) in Stage 1: 3D-informed multimodal alignment using a combination of standard CC558K [41] and our 3D-informed probing data (3DI-Pb-OI1M/1N166K); and (iii) in Stage 2: 3D-informed instruction tuning with the standard Mix665K [41] and our 3D-informed visual instruction tuning data (3DI-Ft-1M)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.776, + 0.866, + 0.792 + ], + "angle": 0, + "content": "3.3.2. Designing 3D-Informed LMM: A Roadmap" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Design instantiation and comparison. The mentioned design space can result in several design instantiations, as depicted in Fig. 5. In a multimodal LLM, a pre-trained CLIP visual encoder extracts grid features from the input image. These features are converted into visual tokens via a connector and then processed by the LLM for visual reasoning. Fig. 5 (b) shows that LLaVA is a standard design space" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "17253" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.115, + 0.104, + 0.488, + 0.209 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.226, + 0.216, + 0.368, + 0.23 + ], + "angle": 0, + "content": "(a) our proposed design" + }, + { + "type": "image", + "bbox": [ + 0.509, + 0.104, + 0.883, + 0.208 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.605, + 0.215, + 0.802, + 0.228 + ], + "angle": 0, + "content": "(b) design of LLaVA-v1.5 (CVPR'24)" + }, + { + "type": "image", + "bbox": [ + 0.115, + 0.253, + 0.485, + 0.355 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.207, + 0.363, + 0.404, + 0.377 + ], + "angle": 0, + "content": "(c) design of SpatialVLM (CVPR'24)" + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.254, + 0.888, + 0.35 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.65, + 0.363, + 0.732, + 0.376 + ], + "angle": 0, + "content": "(d) data cards" + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.393, + 0.907, + 0.463 + ], + "angle": 0, + "content": "Figure 5. Design instantiation and comparison. (a) Architecture and Training of our proposed design. We investigate the 3D-awareness of mixed visual encoders, and incorporate 3D-informed data at each training stage across all architecture components. (b) LLaVA comparison, where only traditional data are used. (c) SpatialVLM comparison, where 3D-informed data are used in the final stage. (d) Traditional data vs. 3D-informed data. Note SpatialVLM lacks of 3D orientation data. Architectures trained with traditional data and 3D-informed data are marked with blue and orange backgrounds, respectively." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.49, + 0.483, + 0.639 + ], + "angle": 0, + "content": "without any new spatial-centric design, meaning that the 3D-awareness is completely missing in LLaVA (Fig. 5 (b)). As in Fig. 5 (c), SpatialVLM improves LLaVA by collecting 3D-informed data that is applied to instruction tuning for LLM, boosting the spatial reasoning capability. Compared to SpatialVLM, our design (in Fig. 5 (a)) inject the 3D-awareness by introduce 3D-informed alignment with ImageNet3D data. Additionally, our 3D-informed intrusion tuning data contain extra 3D orientation data that is missing in spatialVLM." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.641, + 0.484, + 0.716 + ], + "angle": 0, + "content": "Design roadmap. We present the roadmap going from a standard LLaVA to our final model in Fig. 6. We start with LLaVA-v1.5 which only achieves a \\(47.7\\%\\) accuracy on our SpatialVQA. This will be our baseline. We then study a series of design decisions." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.72, + 0.483, + 0.796 + ], + "angle": 0, + "content": "- + mixed vision encoder. To build stronger visual encoder, we mix the original CLIP vision encoder [48] with the DINOv2 [47] vision encoder to extract the hybrid feature maps, lifting a \\(0.3\\%\\) performance gain. This show that the self-supervised feature can enhance spatial reasoning." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.797, + 0.483, + 0.886 + ], + "angle": 0, + "content": "- Vicuna1.5 \\(\\rightarrow\\) Llama3. We switch our large language model from Vicuna-v1.5-7B [68] to Llama3-8B [18], as the latter achieves \\(20\\%\\) higher accuracy on MMLU. This change results in \\(1\\%\\) gain on spatial reasoning. We analyze that improved language reasoning may enhance spatial reasoning as well." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.886, + 0.483, + 0.902 + ], + "angle": 0, + "content": "- +3DI-Ft1M data for stage 2 instruction tuning. We col" + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.72, + 0.483, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.487, + 0.897, + 0.673 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.688, + 0.907, + 0.757 + ], + "angle": 0, + "content": "Figure 6. We modernize a standard LLaVA-v1.5 towards the design of a 3D-informed LMM. The bars are the answer accuracies on the SpatialVQA benchmark, indicating the spatial reasoning capability of each model; results for the GPT-4o is shown with the purple bar." + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.796, + 0.905, + 0.869 + ], + "angle": 0, + "content": "lect the new spatial-aware training set and use it in the instruction tuning stage. Impressively, this 3D-informed instruction tuning results in a \\(10.7\\%\\) performance improvement, highlighting the critical importance of 3D-informed data." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.871, + 0.907, + 0.902 + ], + "angle": 0, + "content": "- + 3DI-Pb-1N166K data for stage 1 multimodal alignment. We propose leveraging object-level 3D-informed" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17254" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.104, + 0.092, + 0.482, + 0.182 + ], + "angle": 0, + "content": "annotation (e.g., 3D location and 3D orientation of objects) to align visual features with 3D awareness. 3DI-Pb-1N166K are converted from human-annotated ImageNet3D [45]. This experiment results in an additional \\(3\\%\\) performance improvement, underscoring the importance of 3D-informed feature alignment." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.185, + 0.483, + 0.23 + ], + "angle": 0, + "content": "In a nutshell, the performance gain can be summarized as 1) better architecture and 2) 3D-informed data. This roadmap leads to our SpatialLLM." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.243, + 0.223, + 0.259 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.268, + 0.483, + 0.299 + ], + "angle": 0, + "content": "In this section we start by outlining the experimental setup in Sec. 4.1, followed by presenting the results in Sec. 4.2." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.307, + 0.279, + 0.323 + ], + "angle": 0, + "content": "4.1. Experimental setup" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.329, + 0.483, + 0.495 + ], + "angle": 0, + "content": "We prepare our training data as described in Fig. 4, with 1M paired image and 3D-informed text for pre-pretraining and alignment, and another 1M multi-turn 3D-informed conversation for instruction tuning. Data distributions are supplemented in appendix. Our training setup is built upon LLaVA-v1.5 [39] and all hyperparameters remain unchanged unless explicitly stated otherwise. We briefly overview the training and data setup for the ablated design choices in Sec. 3.3.1. All experiments are conducted in 8 Nvidia A100 GPUs. We evaluate the spatial reasoning capability of LMMs on SpatialVQA as introduced in Sec. 3.2.2." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.504, + 0.185, + 0.518 + ], + "angle": 0, + "content": "4.2. Results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.525, + 0.483, + 0.587 + ], + "angle": 0, + "content": "Comparison with the state-of-the-art. Table 1 presents the performance of our model compared to state-of-the-art proprietary systems such as GPT-4o and Claude 3.5 Sonnet, along with top open-source models. Remark" + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.598, + 0.48, + 0.817 + ], + "angle": 0, + "content": "
ModelAvg.3D Dist.3D Orient.3D Spatial Rel.
Proprietary
GPT-4o [1]54.059.452.450.8
GPT-4o Mini [1]45.446.939.348.5
Claude 3.5 Sonnet [4]49.350.844.551.5
Open-sourced
LLaVA-v1.5-7B [39]47.750.646.446.2
LLaVA-NeXT-8B [40]50.663.549.241.5
Cambrian-8B [58]51.759.846.448.9
Spatial-centric
SpatialVLM-13B [14]52.264.348.545.8
Ours62.786.352.950.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.827, + 0.483, + 0.856 + ], + "angle": 0, + "content": "Table 1. Comparison with the state-of-the-arts including proprietary and open source models." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.901 + ], + "angle": 0, + "content": "ably, our model achieves a performance of \\(62.7\\%\\), outperforming the top proprietary model by \\(8.7\\%\\) and the best" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.379 + ], + "angle": 0, + "content": "open-source model by \\(10.5\\%\\). Besides, our model excels in all aspects of spatial reasoning involving 3D distance, 3D orientation and 3D spatial relationship. Moreover, our model not only enhances spatial reasoning performance, but still preserves general visual reasoning (detailed in appendix). Interestingly, although SpatialVLM [14] (implemented in SpaceLLaVA [2]) outperforms other open-source models in overall performance, it falls short in 3D orientation reasoning compared to LLaVA, due to the lack of tuning with 3D orientation-specific data. We exclude SpatialRGPT from this comparison as it relies on additional inputs such as object bounding box coordinates, segmentation masks, and depth maps. In contrast, our SpatialLLM, along with LLaVA and Cambrian, focuses on learning spatial awareness directly from raw images. We will consider models with additional inputs in future work. In sum, our model achieves state-of-the-art performance on spatial reasoning, surpassing API models [1, 4] and also spatial-centric VLM [14]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.381, + 0.905, + 0.442 + ], + "angle": 0, + "content": "Thorough empirical exploration of the design space. We comprehensively explore the design space of our approach as shown in Table. 2. We analyze the results and elaborate the findings detailed below." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.444, + 0.905, + 0.548 + ], + "angle": 0, + "content": "Upgrading architectures enable better spatial reasoning. In terms of architecture, integrating a mixed vision encoder can improve overall performance especially for the 3D orientation. Then, upgrading the LLM to the more advanced LLama3-8B enhances spatial reasoning performance by \\(1\\%\\), leveraging its improved language reasoning capabilities." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.551, + 0.906, + 0.748 + ], + "angle": 0, + "content": "3D-informed alignment and tuning matter a lot. In terms of 3D-informed data and training, we find that 3D-informed instruction tuning with our proposed 3DI-Ft1M dataset yields a substantial performance boost of \\(+10.7\\%\\). To further refine the framework, we explore three data types for stage 1 multimodal alignment: curated 3DI-Ft1M, 3DI-Pb-OI1M, and 3DI-Pb-IN166K. The model trained with 3DI-Pb-IN166K performs best. We attribute this to the use of human-annotated ImageNet3D data, as opposed to the semi-automated methods employed for 3DI-Pb-OI1M. The human-annotated ImageNet3D provides more accurate 3D orientation of objects. This highlights the importance of high-quality data in the alignment stage." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.906, + 0.902 + ], + "angle": 0, + "content": "3D-informed pretraining of vision encoder? The pretrained CLIP vision encoder, originally designed for image-text contrastive learning, lacks 3D-awareness [19], We hypotheszie that the vision encoder has a chance to benefit from 3D-informed data. To verify this, we conduct 3D-informed pretraining for the vision encoder using curated 3DI-Pb-OI1M probing data. We explore both CLIP [25] and next token prediction (NTP) [41] objectives to fine-tune Lora layers [24] of the ViT. We observe that pre-pretraining in stage 0 reduces performance, suggesting that tuning vi" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17255" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.089, + 0.904, + 0.362 + ], + "angle": 0, + "content": "
Model\nVision EncoderLLMStage 0\nEncoder pretrainingStage 1\nAlignmentStage 2\nInstruction TuningSpatialVQA
Avg.Dist.Orient.Spatial Rel.
Baseline (LLaVA-v1.5)
CLIP [48]Vicuna-7B [68]-CC558K [39]Mix665K [39]47.750.646.446.2
Mixed Encoders
CLIP [48]+MAE [23]Vicuna-7B [68]-CC558KMix665K47.952.245.746.1
CLIP [48]+DINOv2 [47]Vicuna-7B [68]-CC558KMix665K48.051.946.945.6
Advanced LLM
CLIP [48]+DINOv2 [47]Llama3-8B [18]-CC558K [39]Mix665K [39]49.056.351.741.5
3D-informed Tuning for Stage 2
CLIP [48]+DINOv2 [47]Llama3-8B [18]-CC558K [39]Mix665K [39] + 3DI-Ft1M59.785.252.144.8
3D-informed Alignment for Stage 1
CLIP [48]+DINOv2 [47]Llama3-8B [18]-CC558K [39] + 3DI-Ft1MMix665K [39] + 3DI-Ft1M57.580.846.946.4
CLIP [48]+DINOv2 [47]Llama3-8B [18]-CC558K [39] + 3DI-Pb-OI1MMix665K [39] + 3DI-Ft1M60.082.146.951.6
CLIP [48]+DINOv2 [47]Llama3-8B [18]-CC558K [39] + 3DI-Pb-IN166KMix665K [39] + 3DI-Ft1M62.786.352.951.0
3D-informed pretraining for vision encoder in stage 0
CLIP [48]+DINOv2 [47]Llama3-8B [18]3DI-Pb-OI1M (CLIP [25])CC558K [39] + 3DI-Pb-IN166KMix665K [39] + 3DI-Ft1M60.385.050.547.5
CLIP [48]+DINOv2 [47]Llama3-8B [18]3DI-Pb-OI1M (NTP [34])CC558K [39] + 3DI-Pb-IN166KMix665K [39] + 3DI-Ft1M62.685.252.951.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.373, + 0.907, + 0.43 + ], + "angle": 0, + "content": "Table 2. Thorough exploration of the design space and roadmap progression. We systematically examine the 3D-informed design space from the aspects of data, architecture and training. The non-gray rows, listed from top to bottom, illustrate the step-by-step progression of our design roadmap. Gray rows correspond to ablations across the full design space. Our final model is highlighted with a light orange background. Our final model's superior performance over other variants confirms the necessity of a 3D-informed compound design." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.456, + 0.483, + 0.502 + ], + "angle": 0, + "content": "sual foundation models may compromise feature generalizability. This aligns with prior works advocating for frozen or locked visual encoders [36, 38]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.502, + 0.484, + 0.653 + ], + "angle": 0, + "content": "Roadmap progression towards the best-performing model. After a thorough empirical exploration of the design space, we can devise our roadmap progressed towards the best-performing model. The non-gray rows in Table. 2 align with the roadmap illustrated in Fig. 6, highlighting our systematic approach to upgrade the architecture and incorporate 3D-informed data into the training process. Overall, the improved architecture achieves a \\(1.3\\%\\) performance gain, while 3D-informed alignment and instruction tuning further enhance performance by \\(13.7\\%\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.653, + 0.484, + 0.759 + ], + "angle": 0, + "content": "Qualitative comparison. Fig. 7 illustrates a scenario involving spatial reasoning. GPT-4o struggles to reason about the spatial relationships between objects due to its limited ability to comprehend their 3D orientation. In contrast, our model successfully understands these spatial relationships and provides accurate answers, demonstrating superior 3D spatial reasoning capability." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.77, + 0.219, + 0.785 + ], + "angle": 0, + "content": "5. Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.484, + 0.902 + ], + "angle": 0, + "content": "In this work, we systematically enhance the 3D spatial reasoning capabilities of LMMs through a compound design approach. By introducing SpatialLLM, we demonstrated that integrating 3D-informed data, architectural innovations, and tailored training setups significantly improves an LMM's ability to understand and reason about complex 3D spatial relationships. To tackle data scarcity, we develop" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.456, + 0.764, + 0.518 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.529, + 0.523, + 0.549, + 0.53 + ], + "angle": 0, + "content": "User:" + }, + { + "type": "image", + "bbox": [ + 0.525, + 0.531, + 0.549, + 0.551 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.558, + 0.527, + 0.76, + 0.552 + ], + "angle": 0, + "content": "Consider the real-world 3D locations and orientations of the objects. Which object is closer to the car in Region [0], the car in Region [1] or the car in Region [2]?" + }, + { + "type": "image", + "bbox": [ + 0.521, + 0.572, + 0.55, + 0.598 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.558, + 0.559, + 0.76, + 0.614 + ], + "angle": 0, + "content": "To determine which car is closer to the car in Region [0], we would need to analyze the spatial arrangement of the objects in the image. However, without the ability to visually assess the image, I can't provide a definitive answer. If you can describe the positions of the cars in Regions [1] and [2] relative to Region [0], I can help you deduce which one is closer." + }, + { + "type": "text", + "bbox": [ + 0.621, + 0.619, + 0.698, + 0.629 + ], + "angle": 0, + "content": "The car in Region [1]." + }, + { + "type": "image", + "bbox": [ + 0.768, + 0.457, + 0.901, + 0.518 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.772, + 0.523, + 0.899, + 0.571 + ], + "angle": 0, + "content": "Consider the real-world 3D orientations of the objects. Are the chair in Region [0] and the chair in Region [1] facing the same (or similar) directions, or very different directions?" + }, + { + "type": "text", + "bbox": [ + 0.772, + 0.582, + 0.896, + 0.607 + ], + "angle": 0, + "content": "Based on the image, the chair in Region [0] and the chair in Region [1] face similar directions" + }, + { + "type": "text", + "bbox": [ + 0.788, + 0.619, + 0.879, + 0.627 + ], + "angle": 0, + "content": "Very different directions." + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.646, + 0.907, + 0.703 + ], + "angle": 0, + "content": "Figure 7. Our model is capable of answering question correctly that needs accurate reasoning on spatial distance and 3D orientation, while GPT-4o either needs more spatial information or gives wrong answer." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.718, + 0.907, + 0.885 + ], + "angle": 0, + "content": "3D-informed training datasets focused on 3D locations, orientations and complex spatial relationship. Our systematic integration of these datasets with architectural and training designs provide a roadmap for optimizing LMMs for superior 3D reasoning capabilities. Experimental results on the SpatialVQA benchmark highlight SpatialLLM's superiority over state-of-the-art models, setting a new standard for 3D spatial reasoning. We invite the community to build upon our design roadmap and findings to push the boundaries of what LMMs can achieve in understanding the complex 3D spatial relationships of the real world." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17256" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.279, + 0.108 + ], + "angle": 0, + "content": "6. Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.116, + 0.483, + 0.147 + ], + "angle": 0, + "content": "WM and AY acknowledges support from ONR with N00014-23-1-2641 and ARL award W911NF2320008." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.162, + 0.188, + 0.178 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.187, + 0.483, + 0.256 + ], + "angle": 0, + "content": "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 1, 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.259, + 0.48, + 0.286 + ], + "angle": 0, + "content": "[2] Remyx AI. SpaceLLaVA. https://huggingface.co/remyxai/SpaceLLaVA.1,7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.289, + 0.483, + 0.37 + ], + "angle": 0, + "content": "[3] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in neural information processing systems, 35:23716-23736, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.374, + 0.482, + 0.413 + ], + "angle": 0, + "content": "[4] Anthropic. Claude 3.5 Sonnet. https://www.anthropic.com/news/claude-3-5-sonnet. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.417, + 0.405, + 0.431 + ], + "angle": 0, + "content": "[5] Apollo Team. Apollo synthetic dataset, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.433, + 0.483, + 0.5 + ], + "angle": 0, + "content": "[6] Daichi Azuma, Taiki Miyanishi, Shuhei Kurita, and Motoaki Kawanabe. Scanqa: 3d question answering for spatial scene understanding. In proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 19129-19139, 2022. 4, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.504, + 0.483, + 0.559 + ], + "angle": 0, + "content": "[7] Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, et al. Qwen technical report. arXiv preprint arXiv:2309.16609, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.562, + 0.483, + 0.616 + ], + "angle": 0, + "content": "[8] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.619, + 0.483, + 0.7 + ], + "angle": 0, + "content": "[9] Gilad Baruch, Zhuoyuan Chen, Afshin Dehghan, Tal Dimry, Yuri Feigin, Peter Fu, Thomas Gebauer, Brandon Joffe, Daniel Kurz, Arik Schwartz, et al. Arkitsscenes: A diverse real-world dataset for 3d indoor scene understanding using mobile rgb-d data. arXiv preprint arXiv:2111.08897, 2021. 4, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.704, + 0.483, + 0.758 + ], + "angle": 0, + "content": "[10] Shariq Farooq Bhat, Reiner Birkl, Diana Wofk, Peter Wonka, and Matthias Müller. Zoedepth: Zero-shot transfer by combining relative and metric depth. arXiv preprint arXiv:2302.12288, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.761, + 0.483, + 0.815 + ], + "angle": 0, + "content": "[11] Garrick Brazil, Abhinav Kumar, Julian Straub, Nikhila Ravi, Justin Johnson, and Georgia Gkioxari. Omni3D: A large benchmark and model for 3D object detection in the wild. In CVPR, Vancouver, Canada, 2023. IEEE. 2, 4, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.818, + 0.484, + 0.9 + ], + "angle": 0, + "content": "[12] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11621-11631, 2020. 4, 1" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.187, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.161 + ], + "angle": 0, + "content": "[13] Junbum Cha, Wooyoung Kang, Jonghwan Mun, and Byungseok Roh. Honeybee: Locality-enhanced projector for multimodal llm. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13817-13827, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.165, + 0.906, + 0.245 + ], + "angle": 0, + "content": "[14] Boyuan Chen, Zhuo Xu, Sean Kirmani, Brain Ichter, Dorsa Sadigh, Leonidas Guibas, and Fei Xia. Spatialvlm: Endowing vision-language models with spatial reasoning capabilities. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14455-14465, 2024. 1, 2, 3, 4, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.249, + 0.905, + 0.317 + ], + "angle": 0, + "content": "[15] Jieneng Chen, Qihang Yu, Xiaohui Shen, Alan Yuille, and Liang-Chieh Chen. Vitamin: Designing scalable vision models in the vision-language era. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.32, + 0.905, + 0.374 + ], + "angle": 0, + "content": "[16] An-Chieh Cheng, Hongxu Yin, Yang Fu, Qiushan Guo, Ruihan Yang, Jan Kautz, Xiaolong Wang, and Sifei Liu. Spatialrgpt: Grounded spatial reasoning in vision language model. arXiv preprint arXiv:2406.01584, 2024. 1, 2, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.377, + 0.906, + 0.432 + ], + "angle": 0, + "content": "[17] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.435, + 0.906, + 0.501 + ], + "angle": 0, + "content": "[18] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 3, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.505, + 0.906, + 0.599 + ], + "angle": 0, + "content": "[19] Mohamed El Banani, Amit Raj, Kevis-Kokitsi Maninis, Abhishek Kar, Yuanzhen Li, Michael Rubinstein, Deqing Sun, Leonidas Guibas, Justin Johnson, and Varun Jampani. Probing the 3d awareness of visual foundation models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21795-21806, 2024. 4, 5, 7, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.604, + 0.905, + 0.671 + ], + "angle": 0, + "content": "[20] Samir Yitzhak Gadre, Gabriel Ilharco, Alex Fang, Jonathan Hayase, Georgios Smyrnis, Thao Nguyen, Ryan Marten, Mitchell Wortsman, Dhruba Ghosh, Jieyu Zhang, et al. Dat-acomp: In search of the next generation of multimodal datasets. arXiv preprint arXiv:2304.14108, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.675, + 0.905, + 0.73 + ], + "angle": 0, + "content": "[21] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In 2012 IEEE conference on computer vision and pattern recognition, pages 3354-3361. IEEE, 2012. 4, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.732, + 0.905, + 0.8 + ], + "angle": 0, + "content": "[22] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Batra, and Devi Parikh. Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6904-6913, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.802, + 0.905, + 0.87 + ], + "angle": 0, + "content": "[23] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 3, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[24] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen." + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "17257" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.483, + 0.12 + ], + "angle": 0, + "content": "Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.205 + ], + "angle": 0, + "content": "[25] Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. Openclip, 2021. If you use this software, please cite it as below. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.207, + 0.482, + 0.274 + ], + "angle": 0, + "content": "[26] Artur Jesslen, Guofeng Zhang, Angtian Wang, Wufei Ma, Alan Yuille, and Adam Kortylewski. Novum: Neural object volumes for robust object classification. In European Conference on Computer Vision, pages 264-281. Springer, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.278, + 0.482, + 0.332 + ], + "angle": 0, + "content": "[27] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In ICML, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.335, + 0.482, + 0.403 + ], + "angle": 0, + "content": "[28] Albert Q Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, et al. Mistral 7b. arXiv preprint arXiv:2310.06825, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.406, + 0.482, + 0.488 + ], + "angle": 0, + "content": "[29] Linyi Jin, Jianming Zhang, Yannick Hold-Geoffroy, Oliver Wang, Kevin Blackburn-Matzen, Matthew Sticha, and David F Fouhey. Perspective fields for single image camera calibration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17307-17316, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.491, + 0.482, + 0.558 + ], + "angle": 0, + "content": "[30] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dollar, and Ross Girshick. Segment anything. arXiv:2304.02643, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.562, + 0.482, + 0.631 + ], + "angle": 0, + "content": "[31] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.633, + 0.482, + 0.714 + ], + "angle": 0, + "content": "[32] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision, 123:32-73, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.717, + 0.482, + 0.813 + ], + "angle": 0, + "content": "[33] Alina Kuznetsova, Hassan Rom, Neil Alldrin, Jasper Uijlings, Ivan Krasin, Jordi Pont-Tuset, Shahab Kamali, Stefan Popov, Matteo Malloci, Alexander Kolesnikov, et al. The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale. International journal of computer vision, 128(7):1956-1981, 2020. 5, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.816, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[34] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[35] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.093, + 0.905, + 0.119 + ], + "angle": 0, + "content": "vision-language understanding and generation. In ICML, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[36] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730–19742. PMLR, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.193, + 0.905, + 0.261 + ], + "angle": 0, + "content": "[37] Xianhang Li, Haoqin Tu, Mude Hui, Zeyu Wang, Bingchen Zhao, Junfei Xiao, Sucheng Ren, Jieru Mei, Qing Liu, Huangjie Zheng, et al. What if we recapture billions of web images with llama-3? arXiv preprint arXiv:2406.08478, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.264, + 0.905, + 0.332 + ], + "angle": 0, + "content": "[38] Haotian Liu, Kilho Son, Jianwei Yang, Ce Liu, Jianfeng Gao, Yong Jae Lee, and Chunyuan Li. Learning customized visual models with retrieval-augmented knowledge. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15148-15158, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.334, + 0.905, + 0.403 + ], + "angle": 0, + "content": "[39] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26296-26306, 2024. 1, 2, 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.406, + 0.905, + 0.447 + ], + "angle": 0, + "content": "[40] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava next: Improved reasoning,ocr,and world knowledge,2024.7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.449, + 0.905, + 0.49 + ], + "angle": 0, + "content": "[41] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 1, 2, 3, 4, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.492, + 0.905, + 0.56 + ], + "angle": 0, + "content": "[42] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Qing Jiang, Chunyuan Li, Jianwei Yang, Hang Su, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.563, + 0.905, + 0.63 + ], + "angle": 0, + "content": "[43] Haoyu Lu, Wen Liu, Bo Zhang, Bingxuan Wang, Kai Dong, Bo Liu, Jingxiang Sun, Tongzheng Ren, Zhuoshu Li, Hao Yang, et al. Deepseek-vl: towards real-world vision-language understanding. arXiv preprint arXiv:2403.05525, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.633, + 0.905, + 0.701 + ], + "angle": 0, + "content": "[44] Wufei Ma, Angtian Wang, Alan Yuille, and Adam Kortylewski. Robust category-level 6d pose estimation with coarse-to-fine rendering of neural features. In European Conference on Computer Vision, pages 492-508. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.704, + 0.905, + 0.759 + ], + "angle": 0, + "content": "[45] Wufei Ma, Guanning Zeng, Guofeng Zhang, Qihao Liu, Letian Zhang, Adam Kortylewski, Yaoyao Liu, and Alan Yuille. Imagenet3d: Towards general-purpose object-level 3d understanding. In NeurIPS, 2024. 2, 4, 5, 7, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.761, + 0.905, + 0.83 + ], + "angle": 0, + "content": "[46] Brandon McKinzie, Zhe Gan, Jean-Philippe Fauconnier, Sam Dodge, Bowen Zhang, Philipp Dufter, Dhruti Shah, Xianzhi Du, Futang Peng, Floris Weers, et al. Mm1: Methods, analysis & insights from multimodal lIm pre-training. arXiv preprint arXiv:2403.09611, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.832, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[47] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. TMLR, 2024. 2, 3, 5, 6, 8" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "17258" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.092, + 0.486, + 0.161 + ], + "angle": 0, + "content": "[48] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 2, 3, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.091, + 0.163, + 0.485, + 0.207 + ], + "angle": 0, + "content": "[49] Navid Rajabi and Jana Kosecka. Gsr-bench: A benchmark for grounded spatial reasoning evaluation via multimodal llms. arXiv preprint arXiv:2406.13246, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.207, + 0.484, + 0.29 + ], + "angle": 0, + "content": "[50] Mike Roberts, Jason Ramapuram, Anurag Ranjan, Atulit Kumar, Miguel Angel Bautista, Nathan Paczan, Russ Webb, and Joshua M. Susskind. Hypersim: A photorealistic synthetic dataset for holistic indoor scene understanding. In International Conference on Computer Vision (ICCV) 2021, 2021. 4, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.291, + 0.484, + 0.362 + ], + "angle": 0, + "content": "[51] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.363, + 0.484, + 0.446 + ], + "angle": 0, + "content": "[52] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278-25294, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.447, + 0.484, + 0.502 + ], + "angle": 0, + "content": "[53] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In ACL, 2018. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.504, + 0.485, + 0.56 + ], + "angle": 0, + "content": "[54] Shuran Song, Samuel P Lichtenberg, and Jianxiong Xiao. Sun rgb-d: A rgb-d scene understanding benchmark suite. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 567-576, 2015. 4, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.561, + 0.484, + 0.59 + ], + "angle": 0, + "content": "[55] Elizabeth S Spelke and Katherine D Kinzler. Core knowledge. Developmental science, 10(1):89-96, 2007. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.591, + 0.484, + 0.647 + ], + "angle": 0, + "content": "[56] Chen Sun, Abhinav Shrivastava, Saurabh Singh, and Abhinav Gupta. Revisiting unreasonable effectiveness of data in deep learning era. In Proceedings of the IEEE international conference on computer vision, pages 843-852, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.649, + 0.484, + 0.717 + ], + "angle": 0, + "content": "[57] Gemini Team, Rohan Anil, Sebastian Borgeaud, Yonghui Wu, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, et al. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.719, + 0.484, + 0.789 + ], + "angle": 0, + "content": "[58] Shengbang Tong, Ellis Brown, Penghao Wu, Sanghyun Woo, Manoj Middepogu, Sai Charitha Akula, Jihan Yang, Shusheng Yang, Adithya Iyer, Xichen Pan, et al. Cambrian-1: A fully open, vision-centric exploration of multimodal llms. arXiv preprint arXiv:2406.16860, 2024. 1, 3, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.79, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[59] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, Dan Bikel, Lukas Blecher, Cristian Canton Ferrer, Moya Chen, Guillem Cucurull, David Esiobu, Jude Fernandes, Jeremy Fu, Wenyin Fu, Brian Fuller, Cynthia Gao, Vedanuj Goswami, Naman Goyal, Anthony Hartshorn, Saghar Hosseini, Rui Hou, Hakan Inan, Marcin Kardas, Vik" + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.092, + 0.486, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.547, + 0.093, + 0.905, + 0.285 + ], + "angle": 0, + "content": "tor Kerkez, Madian Khabsa, Isabel Kloumann, Artem Korenev, Punit Singh Koura, Marie-Anne Lachaux, Thibaut Lavril, Jenya Lee, Diana Liskovich, Yinghai Lu, Yuning Mao, Xavier Martinet, Todor Mihaylov, Pushkar Mishra, Igor Molybog, Yixin Nie, Andrew Poulton, Jeremy Reizenstein, Rashi Rungta, Kalyan Saladi, Alan Schelten, Ruan Silva, Eric Michael Smith, Ranjan Subramanian, Xiaoqing Ellen Tan, Binh Tang, Ross Taylor, Adina Williams, Jian Xiang Kuan, Puxin Xu, Zheng Yan, Iliyan Zarov, Yuchen Zhang, Angela Fan, Melanie Kambadur, Sharan Narang, Aurelien Rodriguez, Robert Stojnic, Sergey Edunov, and Thomas Scialom. Llama 2: Open foundation and finetuned chat models. arXiv preprint arXiv:2307.09288, 2023. 2,3" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.288, + 0.905, + 0.344 + ], + "angle": 0, + "content": "[60] Xingrui Wang, Wufei Ma, Zhuowan Li, Adam Kortylewski, and Alan L Yuille. 3d-aware visual question answering about parts, poses and occlusions. Advances in Neural Information Processing Systems, 36, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.346, + 0.905, + 0.401 + ], + "angle": 0, + "content": "[61] Xingrui Wang, Wufei Ma, Angtian Wang, Shuo Chen, Adam Kortylewski, and Alan Yuille. Compositional 4d dynamic scenes understanding with physics priors for video question answering. arXiv preprint arXiv:2406.00622, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.402, + 0.905, + 0.445 + ], + "angle": 0, + "content": "[62] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. In CVPR, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.447, + 0.908, + 0.516 + ], + "angle": 0, + "content": "[63] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10371-10381, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.517, + 0.905, + 0.559 + ], + "angle": 0, + "content": "[64] Shuquan Ye, Dongdong Chen, Songfang Han, and Jing Liao. 3d question answering. arXiv preprint arXiv:2112.08359, 2021. 4, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.561, + 0.905, + 0.616 + ], + "angle": 0, + "content": "[65] Jiahui Yu, Zirui Wang, Vijay Vasudevan, Legg Yeung, Mojtaba Seyedhosseini, and Yonghui Wu. Coca: Contrastive captioners are image-text foundation models. arXiv preprint arXiv:2205.01917, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.618, + 0.905, + 0.687 + ], + "angle": 0, + "content": "[66] Youcai Zhang, Xinyu Huang, Jinyu Ma, Zhaoyang Li, Zhaochuan Luo, Yanchun Xie, Yuzhuo Qin, Tong Luo, Yaqian Li, Shilong Liu, et al. Recognize anything: A strong image tagging model. arXiv preprint arXiv:2306.03514, 2023.1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.689, + 0.905, + 0.771 + ], + "angle": 0, + "content": "[67] Youcai Zhang, Xinyu Huang, Jinyu Ma, Zhaoyang Li, Zhaochuan Luo, Yanchun Xie, Yuzhuo Qin, Tong Luo, Yaqian Li, Shilong Liu, et al. Recognize anything: A strong image tagging model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1724-1732, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.774, + 0.905, + 0.844 + ], + "angle": 0, + "content": "[68] Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric Xing, et al. Judging llm-as-a-judge with mt-bench and chatbot arena. Advances in Neural Information Processing Systems, 36:46595-46623, 2023. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.845, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[69] Jinghao Zhou, Chen Wei, Huiyu Wang, Wei Shen, Cihang Xie, Alan Yuille, and Tao Kong. ibot: Image bert pre-training with online tokenizer. arXiv preprint arXiv:2111.07832, 2021.3" + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.093, + 0.908, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "17259" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.134 + ], + "angle": 0, + "content": "[70] Shengjie Zhu, Abhinav Kumar, Masa Hu, and Xiaoming Liu. Tame a wild camera: In-the-wild monocular camera calibration. In NeurIPS, 2023. 1" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.956 + ], + "angle": 0, + "content": "17260" + } + ] +] \ No newline at end of file diff --git a/2025/SpatialLLM_ A Compound 3D-Informed Design towards Spatially-Intelligent Large Multimodal Models/cf47bc9d-8b67-4bc0-a3fe-ba2e3465761d_origin.pdf b/2025/SpatialLLM_ A Compound 3D-Informed Design towards Spatially-Intelligent Large Multimodal Models/cf47bc9d-8b67-4bc0-a3fe-ba2e3465761d_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..394424aa4c5575854b94e777d4fa939cef08d64a --- /dev/null +++ b/2025/SpatialLLM_ A Compound 3D-Informed Design towards Spatially-Intelligent Large Multimodal Models/cf47bc9d-8b67-4bc0-a3fe-ba2e3465761d_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:879dfe8e630e42903dcfa5b369003042c2f32401c1a2bbbf85dea7476d4f48cd +size 2915665 diff --git a/2025/SpatialLLM_ A Compound 3D-Informed Design towards Spatially-Intelligent Large Multimodal Models/full.md b/2025/SpatialLLM_ A Compound 3D-Informed Design towards Spatially-Intelligent Large Multimodal Models/full.md new file mode 100644 index 0000000000000000000000000000000000000000..1674aab53e1400fe6de36b7f220af05de4c53c52 --- /dev/null +++ b/2025/SpatialLLM_ A Compound 3D-Informed Design towards Spatially-Intelligent Large Multimodal Models/full.md @@ -0,0 +1,372 @@ +# SpatialLLM: A Compound 3D-Informed Design towards Spatially-Intelligent Large Multimodal Models + +Wufei Ma1, Luoxin Ye1, Celso M de Melo2, Alan Yuille1, Jieneng Chen1 + +$^{1}$ Johns Hopkins University $^{2}$ DEVCOM Army Research Laboratory + +# Abstract + +Humans naturally understand 3D spatial relationships, enabling complex reasoning like predicting collisions of vehicles from different directions. Current large multimodal models (LMMs), however, lack of this capability of 3D spatial reasoning. This limitation stems from the scarcity of 3D training data and the bias in current model designs toward 2D data. In this paper, we systematically study the impact of 3D-informed data, architecture, and training setups, introducing SpatialLLM, a large multi-modal model with advanced 3D spatial reasoning abilities. To address data limitations, we develop two types of 3D-informed training datasets: (1) 3D-informed probing data focused on object's 3D location and orientation, and (2) 3D-informed conversation data for complex spatial relationships. Notably, we are the first to curate VQA data that incorporate 3D orientation relationships on real images. Furthermore, we systematically integrate these two types of training data with the architectural and training designs of LMMs, providing a roadmap for optimal design aimed at achieving superior 3D reasoning capabilities. Our SpatialLLM advances machines toward highly capable 3D-informed reasoning, surpassing GPT-4o performance by $8.7\%$ . Our systematic empirical design and the resulting findings offer valuable insights for future research in this direction. Our project page is available at this link. + +# 1. Introduction + +When humans observe a scene, they perceive more than isolated objects; they intuitively understand the spatial relationships among these objects [55], enabling complex reasoning and interaction with the environment. For example, recognizing that a car is approaching an obstacle and might collide involves interpreting the 3D object orientations and spatial relationships. This innate ability to comprehend and reason about 3D space is crucial for navigating and interacting with the physical world. + +Despite significant advancements, current LMMs [1, 4, 39, 41, 57], exhibit limited capability of 3D spatial reason + +![](images/aaad0f38892f2429e3ab001841f512bf0b3f249d95cdb0c266a2b6708fdcdd10.jpg) +Orientation Relationships +Spatial Reasoning Relationships +Distance Relationships The pedestrian is closer to the viewer than the cyclist. + +White van is facing the pedestrian instead of the cyclist + +From the perspective of the pedestrian, the white van is on the right side of the pedestrian + +![](images/6ab07807c283b9bfe2ec939f21710c29d7ab3e485090fe18cd7b6272006c25b8.jpg) +3D-Informed Architecture + +![](images/50c06ac2dbca8f308a5a567d29b45d3beb18dc061831de4411333dfecb035bce.jpg) +3D-Informed Data + +1M 3D-Informed Probing Data description on depth, distance, & orientation + +1M 3D-Informed Instruction Tuning Data conversations about 3D spatial relationships + +Mixed encoders, 3D-aware visual encoders, etc. 3D-Informed Training + +Pre-pretraining, alignment & finetuning + +Figure 1. 3D spatial reasoning is crucial for LMMs to ground objects in 3D space and infer their 3D spatial relationships, such as distance, orientation, and spatial interactions. We systematically study the impact of 3D-informed data, architecture, and training setups, introducing SpatialLLM, an LMM with advanced 3D spatial reasoning abilities. + +ing. These models excel at identifying objects and generating descriptive captions but struggle with tasks that require intricate 3D reasoning, such as predicting object interactions or understanding physical events. + +The primary challenge lies in the scarcity of high-quality 3D spatial relationship data, which prevents models from learning effective 3D spatial reasoning. Collecting a small set of high-quality 3D-aware data to tackle the first challenge is feasible, albeit labor-intensive, using readily available tools. Recent efforts [2, 14] leverage vision foundation models to recognize [67], detect [42], segment [31] and estimate the depth [10, 63] in images. These approaches form semi-automated pipelines to construct pseudo-2.5D datasets for training LMMs with spatial reasoning capabilities, such as depth ordering [14, 58]. + +However, a significant gap remains: previous works [2, 14, 16] have primarily focused on 3D distance relationships, overlooking the crucial role of 3D object orientation. This is primarily due to the absence of robust 6D pose estimation. + +tors. Understanding object orientation is essential for complex spatial reasoning tasks, such as determining if two cars are facing the same direction to avoid dangerous collision. Addressing this gap, we aim to incorporate 3D orientation relationships—converted from ImageNet3D [17]—into our data engine, making us the first to enable complex spatial reasoning involving 3D orientation relationships. + +Even with 3D-informed data in hand, fully unlocking its potential remains challenging due to the intricate, multistage development process required for an LMM. For example, SpatialVLM [14] utilizes 3D-informed data during the instruction tuning stage, but it remains uncertain whether such data can enhance other stages like vision encoder pretraining or multimodal alignment. This limitation suggests that a more holistic approach is necessary. + +In this paper, we present a systematic investigation into building LMMs with enhanced 3D spatial reasoning capabilities through a compound design approach, as illustrated in Fig. 1. Our key contributions are threefold: + +First, we identify three fundamental types of 3D spatial relationships that LMMs must master: distance relationships, orientation relationships, and complex spatial reasoning that combines both. To evaluate these capabilities, we introduce SpatialVQA, a comprehensive benchmark containing 1,323 questions that target 3D spatial understanding, built upon diverse urban and indoor scenes from Omni3D [11]. + +Second, we propose a novel compound 3D-informed design that introduces improvements across multiple dimensions, leading to our proposed SpatialLLM model. On the data front, we develop two types of 3D-informed training datasets: (1) 3D-informed probing data focused on object-level 3D properties, and (2) 3D-informed conversation data for complex spatial relationships. Architecturally, we explore mixing different visual encoders and investigate their 3D awareness capabilities. For training, we introduce 3D-informed data at multiple stages: pre-pretraining, multimodal alignment, and instruction tuning. + +Third, we present the first comprehensive search over the LMM design space for spatial reasoning tasks and propose a roadmap towards developing state-of-the-art models in this domain. The experiment demonstrates that our proposed SpatialLLM achieves state-of-the-art performance on spatial reasoning tasks, surpassing both GPT-4o [1] (by $8.7\%$ ) and spatial-centric models like SpatialVLM [14] (by $10.5\%$ ). Our systematic analysis reveals that while architectural improvements contribute a $1.3\%$ gain, the strategic incorporation of 3D-informed data across different training stages yields a substantial $13.7\%$ performance improvement. + +Our findings highlight the importance of a holistic approach to improving LMMs' 3D spatial reasoning capabilities, demonstrating that both architectural choices and care + +fully curated open-source 3D-informed training data play crucial roles in achieving superior performance. This work establishes a new benchmark for spatial reasoning in LMMs and provides valuable insights for future research in this direction. + +# 2. Related Works + +Large multimodal models (LMMs). Recently, there has been a significant surge in research on multimodal LLMs, exemplified by products like GPT-4 [1] and Gemini [57], as well as open-source models such as LLaVA [34, 39, 41] and Qwen-VL [8]. These open-source models combine pretrained vision encoders [15, 27, 47, 48] with LLMs [7, 28, 59], enabling the language models to handle a wide range of tasks that involve images as input. Typically, these methods have the alignment stage that aligns image features with the input embedding space of the language model using connectors [13, 35, 39], trained with paired images and captions [53]. The instruction fine-tuning [41] stage uses data that contains both images and instructions to enhances LMMs. However, current LMMs are unsuccessful in spatial visual reasoning, due to limited spatial-aware data and stagnant 3D-informed architecture design. We aim to bridge this research gap. + +3D spatial reasoning with LMMs. Understanding spatial relationships between objects in images is crucial for visual reasoning. While this comes naturally to humans, LMMs were found to struggle with recognizing spatial relations [49, 60, 61], due to a lack of 3D-aware spatial-related data during training. While neural symbolic methods [60] achieved improved performance on synthetic benchmarks by involving explicit 3D reasoning modules [26, 44], they struggled to generalize to a broader range of objects found in real images. Fortunately, with foundation models like SAM [31] and depth estimators [10, 63], semi-automatic annotation for 3D-aware VQA data has been adopted in recent studies [14, 16]. SpatialVLM [14] uses the pseudo-annotated VQA data for training only while SpatialRGPT [14, 16] further augments LLaVA with extra input of regional mask and depth map during inference, which is not our focus. Although human-annotated 3D-aware VQA data is rare, we aim to leverage human-annotated 3D object-level datasets (like ImageNet3D [45]) by simply converting image-pose pairs into image-text pair data. However, even with the collected data, a major challenge remains in effectively utilizing 3D-aware data to train optimal models. Current approaches like SpatialVLM [14] use the standard LLaVA and two-stage training. To date, no work has addressed when (which training stage) and where (which network component) are optimal to inject the 3D-awareness to LMMs. We bridge this gap by conducting a compound design that simultaneously considers 3D-aware data, architecture, and the training, leading to a best + +performing design. Rather than competing, our study has the potential to complement prior works [14, 16] by offering new compound design recipe. + +# 3. Methods + +In this section we start by reviewing LMMs in Sec. 3.1, specifically the architecture and the choice of training data and strategies. We present the task of reasoning 3D spatial relationships and explain the challenges LMMs face when answering these questions in Sec. 3.2. Then we study how to address these challenging problems with LMMs and introduce our SpatialLLM in Sec. 3.3. + +# 3.1. Preliminary of LMMs + +Data. The success of deep learning in vision is largely due to well-curated data [20, 56], and LMMs similarly depend on this foundation, albeit with diverse data formats below. + +- Image-only data, such as ImageNet [17], is commonly used for MAE [23]/DINOv2 [47] pretraining with a self-supervised objective. This step enables the model to learn rich visual features solely from visual signals. +- Noisy image-text pairs: Large-scale image-text pairs [20, 37, 52] are used for pretraining with contrastive loss (e.g., CLIP [48, 65]) to align visual and textual representations. +- Multimodal alignment: At this stage, image-caption pairs (e.g., CC3M [53]) are utilized to align visual features with the language model's word embedding space, typically through image-captioning tasks. This pretraining ensures that the model can interpret and relate visual content to textual queries in a shared semantic space. +- Multimodal instruction-following data. Visual instruction tuning data is essential but challenging to collect, as it rarely exists in its natural form online. Prior works [39, 41, 46, 58] repurpose VQA benchmarks [22, 32] into instruction-tuning datasets. + +Architecture. A standard LMM [39, 41] consists of a visual encoder to process the image, a multimodal connector to transform the visual feature to visual token, and a LLM for reasoning. However, the extent to which these architectural components contribute to the spatial understanding of LMMs remains unknown. We review these architectural components as below. + +- Vision encoder: Most LMMs rely on language-supervised models like CLIP [48], leveraging web-scale noisy image-text data. Recent LMMs [43, 58] integrate vision models trained only on visual signals, including self-supervised [23, 47, 69] and generative models [51]. +- Connector: Features from a visual encoder are mapped into the LLM token space by connectors [3, 13, 35, 41] such as a MLP [39, 41]. +- LLMs: LLMs [1, 7, 59] serve as the foundational backbone of LMMs, driving their reasoning capabilities. + +Open-source LLMs like Llama series [18, 59] are frequently adopted to enhance the reasoning capability. + +Training. The aforementioned data supports pretraining for strong visual representation learning, alignment pretraining to synchronize visual and language representations, and instruction tuning to enable multimodal instruction-following and reasoning capabilities + +- Pre-pretraining. This stage focuses on developing foundational visual representations, often with reconstruction-based objectives (e.g., MAE [23], DINOv2 [47]). This step prepares the model to handle more complex multimodal tasks by first refining its visual understanding. +- Multimodal alignment pretraining. At this stage, the model is trained to describe images in details to align visual and language representations in the same space. +- Visual instruction tuning. The final stage involves refining the model's ability to process complex multimodal instructions using visual instruction tuning data. This often incorporates transformed VQA datasets, enhancing the model's capacity in handling tasks that demand coordinated reasoning across language and vision. + +# 3.2. Reasoning 3D Spatial Relationships + +3D spatial reasoning analyzes the spatial relationships between objects in the 3D world space, such as objects being far or close to each other, or one object lies to the left or right of another object [60]. It extends beyond standard 3D tasks, such as 3D object detection and 3D pose estimation that studies the 3D nature of individual objects, and focuses on understanding their mutual relationships in 3D. In this work we identify three main types of 3D spatial relationships, i.e., distance, orientation, and spatial reasoning, and study how we can better address these problems with LMMs. + +3D spatial relationships. For distance spatial relationships, we study if models can estimate and compare 3D distances between objects. These spatial relationships require more than depth awareness, which gives the distance between the viewer and the object and can be roughly estimated from the object scale, but also the 3D distances between two objects that are often harder to estimate. On the other hand, orientation spatial relationships require models to understand the 3D orientations of an object, e.g., predicting if two objects are facing the same direction. Lastly, spatial reasoning spatial relationships look into 3D spatial relationships that need a combination of 3D awareness of location and orientation and then performing spatial reasoning on the 3D information. + +# 3.2.1. Challenges of 3D spatial reasoning + +Reasoning about 3D spatial relationships remains a challenging task for state-of-the-art open-sourced and proprietary LMMs. We attribute this to two main limitations in + +![](images/06f567a187c83769ca94b276493b3c978c825a3c7e72462d7e72e61fcf483f85.jpg) +√3D Location 03D Orientation +User + +![](images/f5d09c711e117ef47c7b87625ca17b09a1a4a1e550c307ebe432c1bc91674eff.jpg) +The chair in Region [2] is closer to the chair in Region [0]. +Spatial Reasoning Relationships $\sqrt{3}$ Location $\sqrt{3}$ Orientation +User +If I stand at the location of the car in Region [0] facing where it is facing, where is the bus in Region [1]? +The bus in Region [1] is on the left side of the car in Region [0]. +Figure 2. Examples from our SpatialVQA benchmark featuring a broad range of questions that require 3D spatial reasoning. + +![](images/a41c6ea95a9124d7247030d0b993c6e50fb2f699ac0f49311a988ab340cde7dc.jpg) +Which object is closer to the chair in Region [0], the bookcase in Region [1] or the chair in Region [2]? +3DI-LMM +√3D Location √3D Orientation +Use +Are the car in Region [0] and the car in Region [1] facing the same direction, or different directions? +The car in Region [0] and the car in Region [1] are facing different directions. + +![](images/17b5d311c6cbe245dfc4975a68277ca32ad4e9e28fc260aca93e1ccafaac99e4.jpg) +3DI-LMM +Spatial Reasoning Relationships +√3D Location √3D Orientation +User +Is the car in Region [0] oriented towards the car in Region [1] or the car in Region [2]? +3DI-LMM +The car in Region [0] is facing towards the car in Region [1]. + +modern LMMs, the lack of 3D awareness and inability to perform 3D spatial reasoning. + +3D awareness refers to the ability of a visual encoder to represent 3D-aware features such as 3D object shapes and 3D orientations. Previous studies resorted to proxy tasks, e.g., part correspondence [19] and pose estimation [45], to quantitatively evaluate the 3D awareness of visual foundation models with linear probing. Results showed that most visual encoder, specifically ones trained with language supervision, demonstrated limited 3D awareness of 3D object poses [45], which inevitably limits the 3D spatial reasoning capabilities of LMMs built upon these encoders. + +Inability to perform 3D spatial reasoning. To infer the 3D spatial relationships, models must also be able to perform 3D spatial reasoning from 3D-aware visual features. Despite the abundant Internet-scale data for multi-modal training, we argue that high-quality 3D spatial reasoning data are scarce. Existing pretraining and visual instruction tuning data for LMMs [41, 58] focused on detailed descriptions and conversations about scenes, appearances, and actions, while being vague about the 3D spatial relationships that build up the high-level physical events. Despite their success on a wide range of image understanding tasks, these LMMs fail to learn the nature of 3D scenes or to perform 3D spatial reasoning. + +# 3.2.2. SpatialVQA for Evaluation + +Despite the rising interests in spatial reasoning, there lacks a comprehensive open source benchmark for reasoning 3D spatial relationships. Early datasets on spatial reasoning were built on 3D scans rather than images [6, 64]. Recent visual-language benchmarks on spatial reasoning either focused on 2D spatial relationships [14, 16], e.g., left or right in the image plane, or only on distance spatial relationships [58]. + +To enable a thorough evaluation of the 3D spatial reasoning capabilities of LMMs, we introduce SpatialVQA, a visual question answering benchmark that covers a wide variety of 3D spatial relationships. Our SpatialVQA distinguishes itself from all previous spatial reasoning benchmarks in the sense that all questions require different levels of 3D awareness and cannot be answered from 2D spatial reasoning only. Specifically, SpatialVQA consists of 1,323 questions, ranging from questions can be answered from distance-awareness only, e.g., which object is closer to a third object, to questions that require estimating objects' 3D orientations, e.g., if two objects are facing the same direction, and eventually questions that require complex 3D spatial reasoning over a combination of object 3D locations and orientations, e.g., an object is in the front/left/back/right direction of another object. + +We build our SpatialVQA on images from Omni3D [11], with 3D bounding box annotations on diverse objects from both urban [12, 21] and indoor scenes [9, 50, 54]. We follow [16, 58] and develop rule-based methods to generate visual question-answer pairs from the 3D groundtruths. Please refer to Fig. 2 for some examples of SpatialVQA and to appendix Sec.B for details about our benchmark. + +# 3.3. Compound 3D-Informed Design + +Problem formulation. While previous approaches focused on generating 3D-informed instruction tuning data to enable LMMs with the abilities to estimate distances, depths, or spatial relationships [14, 16], it is largely understudied the best recipe for developing LMMs with strong 3D spatial reasoning capabilities. Motivated by the two aforementioned limitations of LMMs in Sec. 3.2.1, we consider two main aspects in our compound 3D-informed design – the architecture design that leads to visual encoders with strong 3D awareness and the training setup that advances the 3D spatial reasoning capabilities of LMMs. + +![](images/aa085b7ec1966899e356d8745d3fa88c93bdcb56458e2df75f569419bac3f58f.jpg) +(a) data +(b) training stage +(c) architecture +(d) vision encoder + +![](images/458bf0902d6b89f5e588a589b82a63450156b0daed0108466939cec495bbf875.jpg) +Figure 3. Design space for LMMs capable of spatial reasoning. The dashed boxes and lines highlight our new design space compared to LLaVA-v1.5. This compound design simultaneously considers 3D-informed data, architecture, and training methods to search for the best-performing models for spatial reasoning. +Figure 4. Comparison of data cards showcasing the curation process and data types: standard LLaVA data (left) vs. our 3D-informed data (right). For our 3D-informed data curation, we label depth and distance of objects in OpenImage [33] using semi-automatic tools, resulting in 3DI-Pb-OI1M. Beyond this, we introduce 3D orientation probing data derived from human-annotated ImageNet3D [45], resulting in 3DI-Pb-IN166K for alignment. We further augment the conversation about spatial relationship, leading to 3DI-Ft-1M for instruction tuning. + +![](images/4dc08d9a73e6c392cac4a36ceb33b3fc99b84c1476af4dd2ec4a1c6a7c9dc97a.jpg) + +# 3.3.1. Design space + +We introduce the design space considered in our work, i.e., choices of training data, model architecture, and training setup that advance the 3D spatial reasoning capabilities of LMMs. Refer to Fig. 3 for an overview of our design space. + +Data. Standard LMMs follow LLaVA [41] and adopte feature alignment data for connector pretraining and visual instruction data for supervised instruction tuning. These datasets focus on detailed descriptions of images or machine-generated multi-turn conversations. Besides standard visual-language datasets, we further consider 3D-informed data, including 3D-aware probing data and 3D-informed instruction tuning data. Inspired by [45], the 3D + +aware probing data involves fundamental object-level 3D properties, such as the azimuth and elevation 3D viewpoint of an object, depth of the object from the viewer, and 3D distances between two objects. The goal of the 3D-aware probing data is to effectively inject fundamental 3D awareness into the models and benefit subsequent reasoning. Meanwhile, the 3D-informed instruction tuning data focuses on high-level 3D spatial relationships between the objects, which are built on a combination of the fundamental 3D properties and 3D spatial reasoning. + +To generate the 3D-aware probing data and the 3D-informed instruction tuning data, we follow the previous data generation pipelines [14, 16] and exploit visual foundation models to extract and tag the objects in 3D space. Furthermore, we adopt a pretrained 6D pose estimator in [45] and estimate 3D orientations of the objects, which provides the crucial 3D information for generating 3D-informed data regarding various orientation-related 3D spatial relationships. We illustrate our 3D-informed data generation pipeline in Fig. 4 and refer the readers to appendex Sec.A for details of our 3D-informed data generation. + +Architecture. Inspired by Probe3D [19] which probes the 3D-awareness of visual foundation models, we study the design of visual encoder to enable better 3D spatial reasoning. We consider two types of visual encoders: (i) Frozen & pretrained visual encoder CLIP [48] following [41] but with the option to mix a wider variety of models, such as MAE [23] and DINOv2 [47], and (ii) 3D-aware visual encoders finetuned with our 3D-informed data, e.g., contrastive pretraining or finetuning with LoRA using language modeling loss. + +Training setup. To address the limitations of 3D-aware LMMs discussed in Sec. 3.2.1, we propose new training setups that aim to improve 3D awareness and advance the 3D spatial reasoning capabilities. We consider the following: (i) in Stage 0: use 3D-informed probing data (3DI-Pb-OI1M/1N166K) to tune the Lora layers of CLIP vision encoder (Pre-pretraining); (ii) in Stage 1: 3D-informed multimodal alignment using a combination of standard CC558K [41] and our 3D-informed probing data (3DI-Pb-OI1M/1N166K); and (iii) in Stage 2: 3D-informed instruction tuning with the standard Mix665K [41] and our 3D-informed visual instruction tuning data (3DI-Ft-1M). + +# 3.3.2. Designing 3D-Informed LMM: A Roadmap + +Design instantiation and comparison. The mentioned design space can result in several design instantiations, as depicted in Fig. 5. In a multimodal LLM, a pre-trained CLIP visual encoder extracts grid features from the input image. These features are converted into visual tokens via a connector and then processed by the LLM for visual reasoning. Fig. 5 (b) shows that LLaVA is a standard design space + +![](images/af92f889ccc138fcba6a6605c945b4eb095cf3f72f3750ce067312d2cfe9ce92.jpg) +(a) our proposed design + +![](images/be280e1db447da6cdc52ac19ab235d8ee252aacb77c2d6e70818603c0f61f32d.jpg) +(b) design of LLaVA-v1.5 (CVPR'24) + +![](images/529895ee21b4cbcfe8f8bc1b75b4b2672e56e3d025a7643e574f089b4329c533.jpg) +(c) design of SpatialVLM (CVPR'24) + +![](images/d4388b1caeed971136a941795b4946bc2bf24d89a0a4915e4510e3086c738429.jpg) +(d) data cards + +without any new spatial-centric design, meaning that the 3D-awareness is completely missing in LLaVA (Fig. 5 (b)). As in Fig. 5 (c), SpatialVLM improves LLaVA by collecting 3D-informed data that is applied to instruction tuning for LLM, boosting the spatial reasoning capability. Compared to SpatialVLM, our design (in Fig. 5 (a)) inject the 3D-awareness by introduce 3D-informed alignment with ImageNet3D data. Additionally, our 3D-informed intrusion tuning data contain extra 3D orientation data that is missing in spatialVLM. + +Design roadmap. We present the roadmap going from a standard LLaVA to our final model in Fig. 6. We start with LLaVA-v1.5 which only achieves a $47.7\%$ accuracy on our SpatialVQA. This will be our baseline. We then study a series of design decisions. + +- + mixed vision encoder. To build stronger visual encoder, we mix the original CLIP vision encoder [48] with the DINOv2 [47] vision encoder to extract the hybrid feature maps, lifting a $0.3\%$ performance gain. This show that the self-supervised feature can enhance spatial reasoning. +- Vicuna1.5 $\rightarrow$ Llama3. We switch our large language model from Vicuna-v1.5-7B [68] to Llama3-8B [18], as the latter achieves $20\%$ higher accuracy on MMLU. This change results in $1\%$ gain on spatial reasoning. We analyze that improved language reasoning may enhance spatial reasoning as well. +- +3DI-Ft1M data for stage 2 instruction tuning. We col + +![](images/42d503fcb94cf182feed8978f02582e39a84d445a516eb8bc4e6177b68a8594c.jpg) +Figure 5. Design instantiation and comparison. (a) Architecture and Training of our proposed design. We investigate the 3D-awareness of mixed visual encoders, and incorporate 3D-informed data at each training stage across all architecture components. (b) LLaVA comparison, where only traditional data are used. (c) SpatialVLM comparison, where 3D-informed data are used in the final stage. (d) Traditional data vs. 3D-informed data. Note SpatialVLM lacks of 3D orientation data. Architectures trained with traditional data and 3D-informed data are marked with blue and orange backgrounds, respectively. +Figure 6. We modernize a standard LLaVA-v1.5 towards the design of a 3D-informed LMM. The bars are the answer accuracies on the SpatialVQA benchmark, indicating the spatial reasoning capability of each model; results for the GPT-4o is shown with the purple bar. + +lect the new spatial-aware training set and use it in the instruction tuning stage. Impressively, this 3D-informed instruction tuning results in a $10.7\%$ performance improvement, highlighting the critical importance of 3D-informed data. + +- + 3DI-Pb-1N166K data for stage 1 multimodal alignment. We propose leveraging object-level 3D-informed + +annotation (e.g., 3D location and 3D orientation of objects) to align visual features with 3D awareness. 3DI-Pb-1N166K are converted from human-annotated ImageNet3D [45]. This experiment results in an additional $3\%$ performance improvement, underscoring the importance of 3D-informed feature alignment. + +In a nutshell, the performance gain can be summarized as 1) better architecture and 2) 3D-informed data. This roadmap leads to our SpatialLLM. + +# 4. Experiments + +In this section we start by outlining the experimental setup in Sec. 4.1, followed by presenting the results in Sec. 4.2. + +# 4.1. Experimental setup + +We prepare our training data as described in Fig. 4, with 1M paired image and 3D-informed text for pre-pretraining and alignment, and another 1M multi-turn 3D-informed conversation for instruction tuning. Data distributions are supplemented in appendix. Our training setup is built upon LLaVA-v1.5 [39] and all hyperparameters remain unchanged unless explicitly stated otherwise. We briefly overview the training and data setup for the ablated design choices in Sec. 3.3.1. All experiments are conducted in 8 Nvidia A100 GPUs. We evaluate the spatial reasoning capability of LMMs on SpatialVQA as introduced in Sec. 3.2.2. + +# 4.2. Results + +Comparison with the state-of-the-art. Table 1 presents the performance of our model compared to state-of-the-art proprietary systems such as GPT-4o and Claude 3.5 Sonnet, along with top open-source models. Remark + +
ModelAvg.3D Dist.3D Orient.3D Spatial Rel.
Proprietary
GPT-4o [1]54.059.452.450.8
GPT-4o Mini [1]45.446.939.348.5
Claude 3.5 Sonnet [4]49.350.844.551.5
Open-sourced
LLaVA-v1.5-7B [39]47.750.646.446.2
LLaVA-NeXT-8B [40]50.663.549.241.5
Cambrian-8B [58]51.759.846.448.9
Spatial-centric
SpatialVLM-13B [14]52.264.348.545.8
Ours62.786.352.950.8
+ +Table 1. Comparison with the state-of-the-arts including proprietary and open source models. + +ably, our model achieves a performance of $62.7\%$ , outperforming the top proprietary model by $8.7\%$ and the best + +open-source model by $10.5\%$ . Besides, our model excels in all aspects of spatial reasoning involving 3D distance, 3D orientation and 3D spatial relationship. Moreover, our model not only enhances spatial reasoning performance, but still preserves general visual reasoning (detailed in appendix). Interestingly, although SpatialVLM [14] (implemented in SpaceLLaVA [2]) outperforms other open-source models in overall performance, it falls short in 3D orientation reasoning compared to LLaVA, due to the lack of tuning with 3D orientation-specific data. We exclude SpatialRGPT from this comparison as it relies on additional inputs such as object bounding box coordinates, segmentation masks, and depth maps. In contrast, our SpatialLLM, along with LLaVA and Cambrian, focuses on learning spatial awareness directly from raw images. We will consider models with additional inputs in future work. In sum, our model achieves state-of-the-art performance on spatial reasoning, surpassing API models [1, 4] and also spatial-centric VLM [14]. + +Thorough empirical exploration of the design space. We comprehensively explore the design space of our approach as shown in Table. 2. We analyze the results and elaborate the findings detailed below. + +Upgrading architectures enable better spatial reasoning. In terms of architecture, integrating a mixed vision encoder can improve overall performance especially for the 3D orientation. Then, upgrading the LLM to the more advanced LLama3-8B enhances spatial reasoning performance by $1\%$ , leveraging its improved language reasoning capabilities. + +3D-informed alignment and tuning matter a lot. In terms of 3D-informed data and training, we find that 3D-informed instruction tuning with our proposed 3DI-Ft1M dataset yields a substantial performance boost of $+10.7\%$ . To further refine the framework, we explore three data types for stage 1 multimodal alignment: curated 3DI-Ft1M, 3DI-Pb-OI1M, and 3DI-Pb-IN166K. The model trained with 3DI-Pb-IN166K performs best. We attribute this to the use of human-annotated ImageNet3D data, as opposed to the semi-automated methods employed for 3DI-Pb-OI1M. The human-annotated ImageNet3D provides more accurate 3D orientation of objects. This highlights the importance of high-quality data in the alignment stage. + +3D-informed pretraining of vision encoder? The pretrained CLIP vision encoder, originally designed for image-text contrastive learning, lacks 3D-awareness [19], We hypotheszie that the vision encoder has a chance to benefit from 3D-informed data. To verify this, we conduct 3D-informed pretraining for the vision encoder using curated 3DI-Pb-OI1M probing data. We explore both CLIP [25] and next token prediction (NTP) [41] objectives to fine-tune Lora layers [24] of the ViT. We observe that pre-pretraining in stage 0 reduces performance, suggesting that tuning vi + +
Model +Vision EncoderLLMStage 0 +Encoder pretrainingStage 1 +AlignmentStage 2 +Instruction TuningSpatialVQA
Avg.Dist.Orient.Spatial Rel.
Baseline (LLaVA-v1.5)
CLIP [48]Vicuna-7B [68]-CC558K [39]Mix665K [39]47.750.646.446.2
Mixed Encoders
CLIP [48]+MAE [23]Vicuna-7B [68]-CC558KMix665K47.952.245.746.1
CLIP [48]+DINOv2 [47]Vicuna-7B [68]-CC558KMix665K48.051.946.945.6
Advanced LLM
CLIP [48]+DINOv2 [47]Llama3-8B [18]-CC558K [39]Mix665K [39]49.056.351.741.5
3D-informed Tuning for Stage 2
CLIP [48]+DINOv2 [47]Llama3-8B [18]-CC558K [39]Mix665K [39] + 3DI-Ft1M59.785.252.144.8
3D-informed Alignment for Stage 1
CLIP [48]+DINOv2 [47]Llama3-8B [18]-CC558K [39] + 3DI-Ft1MMix665K [39] + 3DI-Ft1M57.580.846.946.4
CLIP [48]+DINOv2 [47]Llama3-8B [18]-CC558K [39] + 3DI-Pb-OI1MMix665K [39] + 3DI-Ft1M60.082.146.951.6
CLIP [48]+DINOv2 [47]Llama3-8B [18]-CC558K [39] + 3DI-Pb-IN166KMix665K [39] + 3DI-Ft1M62.786.352.951.0
3D-informed pretraining for vision encoder in stage 0
CLIP [48]+DINOv2 [47]Llama3-8B [18]3DI-Pb-OI1M (CLIP [25])CC558K [39] + 3DI-Pb-IN166KMix665K [39] + 3DI-Ft1M60.385.050.547.5
CLIP [48]+DINOv2 [47]Llama3-8B [18]3DI-Pb-OI1M (NTP [34])CC558K [39] + 3DI-Pb-IN166KMix665K [39] + 3DI-Ft1M62.685.252.951.5
+ +Table 2. Thorough exploration of the design space and roadmap progression. We systematically examine the 3D-informed design space from the aspects of data, architecture and training. The non-gray rows, listed from top to bottom, illustrate the step-by-step progression of our design roadmap. Gray rows correspond to ablations across the full design space. Our final model is highlighted with a light orange background. Our final model's superior performance over other variants confirms the necessity of a 3D-informed compound design. + +sual foundation models may compromise feature generalizability. This aligns with prior works advocating for frozen or locked visual encoders [36, 38]. + +Roadmap progression towards the best-performing model. After a thorough empirical exploration of the design space, we can devise our roadmap progressed towards the best-performing model. The non-gray rows in Table. 2 align with the roadmap illustrated in Fig. 6, highlighting our systematic approach to upgrade the architecture and incorporate 3D-informed data into the training process. Overall, the improved architecture achieves a $1.3\%$ performance gain, while 3D-informed alignment and instruction tuning further enhance performance by $13.7\%$ . + +Qualitative comparison. Fig. 7 illustrates a scenario involving spatial reasoning. GPT-4o struggles to reason about the spatial relationships between objects due to its limited ability to comprehend their 3D orientation. In contrast, our model successfully understands these spatial relationships and provides accurate answers, demonstrating superior 3D spatial reasoning capability. + +# 5. Conclusions + +In this work, we systematically enhance the 3D spatial reasoning capabilities of LMMs through a compound design approach. By introducing SpatialLLM, we demonstrated that integrating 3D-informed data, architectural innovations, and tailored training setups significantly improves an LMM's ability to understand and reason about complex 3D spatial relationships. To tackle data scarcity, we develop + +![](images/d2fedb8e633046d43ee48872746547787fde33b0cd8a454c0f32921049fa5c04.jpg) + +![](images/ef9d4b3f8241ae4a604de36f0ab946aea5993a83fadeadb013519bb1842a82c6.jpg) +User: + +Consider the real-world 3D locations and orientations of the objects. Which object is closer to the car in Region [0], the car in Region [1] or the car in Region [2]? + +![](images/b8790ef8c43ee80660eff58393f58e2c76fa85958ecd1fc9f6a9df105d700aa9.jpg) +Figure 7. Our model is capable of answering question correctly that needs accurate reasoning on spatial distance and 3D orientation, while GPT-4o either needs more spatial information or gives wrong answer. + +To determine which car is closer to the car in Region [0], we would need to analyze the spatial arrangement of the objects in the image. However, without the ability to visually assess the image, I can't provide a definitive answer. If you can describe the positions of the cars in Regions [1] and [2] relative to Region [0], I can help you deduce which one is closer. + +The car in Region [1]. + +![](images/3d08707deb9c3b6aaac7cc8c8293544380e97eafa903793ea4d1ef455e79988f.jpg) + +Consider the real-world 3D orientations of the objects. Are the chair in Region [0] and the chair in Region [1] facing the same (or similar) directions, or very different directions? + +Based on the image, the chair in Region [0] and the chair in Region [1] face similar directions + +Very different directions. + +3D-informed training datasets focused on 3D locations, orientations and complex spatial relationship. Our systematic integration of these datasets with architectural and training designs provide a roadmap for optimizing LMMs for superior 3D reasoning capabilities. Experimental results on the SpatialVQA benchmark highlight SpatialLLM's superiority over state-of-the-art models, setting a new standard for 3D spatial reasoning. We invite the community to build upon our design roadmap and findings to push the boundaries of what LMMs can achieve in understanding the complex 3D spatial relationships of the real world. + +# 6. Acknowledgements + +WM and AY acknowledges support from ONR with N00014-23-1-2641 and ARL award W911NF2320008. + +# References + +[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 1, 2, 3, 7 +[2] Remyx AI. SpaceLLaVA. https://huggingface.co/remyxai/SpaceLLaVA.1,7 +[3] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in neural information processing systems, 35:23716-23736, 2022. 3 +[4] Anthropic. Claude 3.5 Sonnet. https://www.anthropic.com/news/claude-3-5-sonnet. 1, 7 +[5] Apollo Team. Apollo synthetic dataset, 2019. 1 +[6] Daichi Azuma, Taiki Miyanishi, Shuhei Kurita, and Motoaki Kawanabe. Scanqa: 3d question answering for spatial scene understanding. In proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 19129-19139, 2022. 4, 1 +[7] Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, et al. Qwen technical report. arXiv preprint arXiv:2309.16609, 2023. 2, 3 +[8] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966, 2023. 2 +[9] Gilad Baruch, Zhuoyuan Chen, Afshin Dehghan, Tal Dimry, Yuri Feigin, Peter Fu, Thomas Gebauer, Brandon Joffe, Daniel Kurz, Arik Schwartz, et al. Arkitsscenes: A diverse real-world dataset for 3d indoor scene understanding using mobile rgb-d data. arXiv preprint arXiv:2111.08897, 2021. 4, 1 +[10] Shariq Farooq Bhat, Reiner Birkl, Diana Wofk, Peter Wonka, and Matthias Müller. Zoedepth: Zero-shot transfer by combining relative and metric depth. arXiv preprint arXiv:2302.12288, 2023. 1, 2 +[11] Garrick Brazil, Abhinav Kumar, Julian Straub, Nikhila Ravi, Justin Johnson, and Georgia Gkioxari. Omni3D: A large benchmark and model for 3D object detection in the wild. In CVPR, Vancouver, Canada, 2023. IEEE. 2, 4, 1 +[12] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11621-11631, 2020. 4, 1 + +[13] Junbum Cha, Wooyoung Kang, Jonghwan Mun, and Byungseok Roh. Honeybee: Locality-enhanced projector for multimodal llm. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13817-13827, 2024. 2, 3 +[14] Boyuan Chen, Zhuo Xu, Sean Kirmani, Brain Ichter, Dorsa Sadigh, Leonidas Guibas, and Fei Xia. Spatialvlm: Endowing vision-language models with spatial reasoning capabilities. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14455-14465, 2024. 1, 2, 3, 4, 5, 7 +[15] Jieneng Chen, Qihang Yu, Xiaohui Shen, Alan Yuille, and Liang-Chieh Chen. Vitamin: Designing scalable vision models in the vision-language era. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 2 +[16] An-Chieh Cheng, Hongxu Yin, Yang Fu, Qiushan Guo, Ruihan Yang, Jan Kautz, Xiaolong Wang, and Sifei Liu. Spatialrgpt: Grounded spatial reasoning in vision language model. arXiv preprint arXiv:2406.01584, 2024. 1, 2, 3, 4, 5 +[17] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 2, 3 +[18] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 3, 6, 8 +[19] Mohamed El Banani, Amit Raj, Kevis-Kokitsi Maninis, Abhishek Kar, Yuanzhen Li, Michael Rubinstein, Deqing Sun, Leonidas Guibas, Justin Johnson, and Varun Jampani. Probing the 3d awareness of visual foundation models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21795-21806, 2024. 4, 5, 7, 1 +[20] Samir Yitzhak Gadre, Gabriel Ilharco, Alex Fang, Jonathan Hayase, Georgios Smyrnis, Thao Nguyen, Ryan Marten, Mitchell Wortsman, Dhruba Ghosh, Jieyu Zhang, et al. Dat-acomp: In search of the next generation of multimodal datasets. arXiv preprint arXiv:2304.14108, 2023. 3 +[21] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In 2012 IEEE conference on computer vision and pattern recognition, pages 3354-3361. IEEE, 2012. 4, 1 +[22] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Batra, and Devi Parikh. Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6904-6913, 2017. 3 +[23] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 3, 5, 8 +[24] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. + +Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 7 +[25] Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. Openclip, 2021. If you use this software, please cite it as below. 7, 8 +[26] Artur Jesslen, Guofeng Zhang, Angtian Wang, Wufei Ma, Alan Yuille, and Adam Kortylewski. Novum: Neural object volumes for robust object classification. In European Conference on Computer Vision, pages 264-281. Springer, 2024. 2 +[27] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In ICML, 2021. 2 +[28] Albert Q Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, et al. Mistral 7b. arXiv preprint arXiv:2310.06825, 2023. 2 +[29] Linyi Jin, Jianming Zhang, Yannick Hold-Geoffroy, Oliver Wang, Kevin Blackburn-Matzen, Matthew Sticha, and David F Fouhey. Perspective fields for single image camera calibration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17307-17316, 2023. 1 +[30] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dollar, and Ross Girshick. Segment anything. arXiv:2304.02643, 2023. 1 +[31] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 1, 2 +[32] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision, 123:32-73, 2017. 3 +[33] Alina Kuznetsova, Hassan Rom, Neil Alldrin, Jasper Uijlings, Ivan Krasin, Jordi Pont-Tuset, Shahab Kamali, Stefan Popov, Matteo Malloci, Alexander Kolesnikov, et al. The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale. International journal of computer vision, 128(7):1956-1981, 2020. 5, 1 +[34] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 2, 8 +[35] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified + +vision-language understanding and generation. In ICML, 2022. 2, 3 +[36] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730–19742. PMLR, 2023. 8 +[37] Xianhang Li, Haoqin Tu, Mude Hui, Zeyu Wang, Bingchen Zhao, Junfei Xiao, Sucheng Ren, Jieru Mei, Qing Liu, Huangjie Zheng, et al. What if we recapture billions of web images with llama-3? arXiv preprint arXiv:2406.08478, 2024. 3 +[38] Haotian Liu, Kilho Son, Jianwei Yang, Ce Liu, Jianfeng Gao, Yong Jae Lee, and Chunyuan Li. Learning customized visual models with retrieval-augmented knowledge. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15148-15158, 2023. 8 +[39] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26296-26306, 2024. 1, 2, 3, 7, 8 +[40] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava next: Improved reasoning,ocr,and world knowledge,2024.7 +[41] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 1, 2, 3, 4, 5, 7 +[42] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Qing Jiang, Chunyuan Li, Jianwei Yang, Hang Su, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023. 1 +[43] Haoyu Lu, Wen Liu, Bo Zhang, Bingxuan Wang, Kai Dong, Bo Liu, Jingxiang Sun, Tongzheng Ren, Zhuoshu Li, Hao Yang, et al. Deepseek-vl: towards real-world vision-language understanding. arXiv preprint arXiv:2403.05525, 2024. 3 +[44] Wufei Ma, Angtian Wang, Alan Yuille, and Adam Kortylewski. Robust category-level 6d pose estimation with coarse-to-fine rendering of neural features. In European Conference on Computer Vision, pages 492-508. Springer, 2022. 2 +[45] Wufei Ma, Guanning Zeng, Guofeng Zhang, Qihao Liu, Letian Zhang, Adam Kortylewski, Yaoyao Liu, and Alan Yuille. Imagenet3d: Towards general-purpose object-level 3d understanding. In NeurIPS, 2024. 2, 4, 5, 7, 1 +[46] Brandon McKinzie, Zhe Gan, Jean-Philippe Fauconnier, Sam Dodge, Bowen Zhang, Philipp Dufter, Dhruti Shah, Xianzhi Du, Futang Peng, Floris Weers, et al. Mm1: Methods, analysis & insights from multimodal lIm pre-training. arXiv preprint arXiv:2403.09611, 2024. 3 +[47] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. TMLR, 2024. 2, 3, 5, 6, 8 + +[48] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 2, 3, 5, 6, 8 +[49] Navid Rajabi and Jana Kosecka. Gsr-bench: A benchmark for grounded spatial reasoning evaluation via multimodal llms. arXiv preprint arXiv:2406.13246, 2024. 2 +[50] Mike Roberts, Jason Ramapuram, Anurag Ranjan, Atulit Kumar, Miguel Angel Bautista, Nathan Paczan, Russ Webb, and Joshua M. Susskind. Hypersim: A photorealistic synthetic dataset for holistic indoor scene understanding. In International Conference on Computer Vision (ICCV) 2021, 2021. 4, 1 +[51] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 3 +[52] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278-25294, 2022. 3 +[53] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In ACL, 2018. 2, 3 +[54] Shuran Song, Samuel P Lichtenberg, and Jianxiong Xiao. Sun rgb-d: A rgb-d scene understanding benchmark suite. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 567-576, 2015. 4, 1 +[55] Elizabeth S Spelke and Katherine D Kinzler. Core knowledge. Developmental science, 10(1):89-96, 2007. 1 +[56] Chen Sun, Abhinav Shrivastava, Saurabh Singh, and Abhinav Gupta. Revisiting unreasonable effectiveness of data in deep learning era. In Proceedings of the IEEE international conference on computer vision, pages 843-852, 2017. 3 +[57] Gemini Team, Rohan Anil, Sebastian Borgeaud, Yonghui Wu, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, et al. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805, 2023. 1, 2 +[58] Shengbang Tong, Ellis Brown, Penghao Wu, Sanghyun Woo, Manoj Middepogu, Sai Charitha Akula, Jihan Yang, Shusheng Yang, Adithya Iyer, Xichen Pan, et al. Cambrian-1: A fully open, vision-centric exploration of multimodal llms. arXiv preprint arXiv:2406.16860, 2024. 1, 3, 4, 7 +[59] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, Dan Bikel, Lukas Blecher, Cristian Canton Ferrer, Moya Chen, Guillem Cucurull, David Esiobu, Jude Fernandes, Jeremy Fu, Wenyin Fu, Brian Fuller, Cynthia Gao, Vedanuj Goswami, Naman Goyal, Anthony Hartshorn, Saghar Hosseini, Rui Hou, Hakan Inan, Marcin Kardas, Vik + +tor Kerkez, Madian Khabsa, Isabel Kloumann, Artem Korenev, Punit Singh Koura, Marie-Anne Lachaux, Thibaut Lavril, Jenya Lee, Diana Liskovich, Yinghai Lu, Yuning Mao, Xavier Martinet, Todor Mihaylov, Pushkar Mishra, Igor Molybog, Yixin Nie, Andrew Poulton, Jeremy Reizenstein, Rashi Rungta, Kalyan Saladi, Alan Schelten, Ruan Silva, Eric Michael Smith, Ranjan Subramanian, Xiaoqing Ellen Tan, Binh Tang, Ross Taylor, Adina Williams, Jian Xiang Kuan, Puxin Xu, Zheng Yan, Iliyan Zarov, Yuchen Zhang, Angela Fan, Melanie Kambadur, Sharan Narang, Aurelien Rodriguez, Robert Stojnic, Sergey Edunov, and Thomas Scialom. Llama 2: Open foundation and finetuned chat models. arXiv preprint arXiv:2307.09288, 2023. 2,3 +[60] Xingrui Wang, Wufei Ma, Zhuowan Li, Adam Kortylewski, and Alan L Yuille. 3d-aware visual question answering about parts, poses and occlusions. Advances in Neural Information Processing Systems, 36, 2024. 2, 3 +[61] Xingrui Wang, Wufei Ma, Angtian Wang, Shuo Chen, Adam Kortylewski, and Alan Yuille. Compositional 4d dynamic scenes understanding with physics priors for video question answering. arXiv preprint arXiv:2406.00622, 2024. 2 +[62] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. In CVPR, 2024. 1 +[63] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10371-10381, 2024. 1, 2 +[64] Shuquan Ye, Dongdong Chen, Songfang Han, and Jing Liao. 3d question answering. arXiv preprint arXiv:2112.08359, 2021. 4, 1 +[65] Jiahui Yu, Zirui Wang, Vijay Vasudevan, Legg Yeung, Mojtaba Seyedhosseini, and Yonghui Wu. Coca: Contrastive captioners are image-text foundation models. arXiv preprint arXiv:2205.01917, 2022. 3 +[66] Youcai Zhang, Xinyu Huang, Jinyu Ma, Zhaoyang Li, Zhaochuan Luo, Yanchun Xie, Yuzhuo Qin, Tong Luo, Yaqian Li, Shilong Liu, et al. Recognize anything: A strong image tagging model. arXiv preprint arXiv:2306.03514, 2023.1 +[67] Youcai Zhang, Xinyu Huang, Jinyu Ma, Zhaoyang Li, Zhaochuan Luo, Yanchun Xie, Yuzhuo Qin, Tong Luo, Yaqian Li, Shilong Liu, et al. Recognize anything: A strong image tagging model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1724-1732, 2024. 1 +[68] Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric Xing, et al. Judging llm-as-a-judge with mt-bench and chatbot arena. Advances in Neural Information Processing Systems, 36:46595-46623, 2023. 6, 8 +[69] Jinghao Zhou, Chen Wei, Huiyu Wang, Wei Shen, Cihang Xie, Alan Yuille, and Tao Kong. ibot: Image bert pre-training with online tokenizer. arXiv preprint arXiv:2111.07832, 2021.3 + +[70] Shengjie Zhu, Abhinav Kumar, Masa Hu, and Xiaoming Liu. Tame a wild camera: In-the-wild monocular camera calibration. In NeurIPS, 2023. 1 \ No newline at end of file diff --git a/2025/SpatialLLM_ A Compound 3D-Informed Design towards Spatially-Intelligent Large Multimodal Models/images.zip b/2025/SpatialLLM_ A Compound 3D-Informed Design towards Spatially-Intelligent Large Multimodal Models/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..3e1670cc6d9cedd7635ed3053b9942811fc4cebb --- /dev/null +++ b/2025/SpatialLLM_ A Compound 3D-Informed Design towards Spatially-Intelligent Large Multimodal Models/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:28e7e23d4ab52deb074f18e15af4879bad6e662eb8b29504e26927ed7a7ee9fe +size 395876 diff --git a/2025/SpatialLLM_ A Compound 3D-Informed Design towards Spatially-Intelligent Large Multimodal Models/layout.json b/2025/SpatialLLM_ A Compound 3D-Informed Design towards Spatially-Intelligent Large Multimodal Models/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..ab024f534fbbb9508ad3b896285196c07113328f --- /dev/null +++ b/2025/SpatialLLM_ A Compound 3D-Informed Design towards Spatially-Intelligent Large Multimodal Models/layout.json @@ -0,0 +1,9430 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 129, + 103, + 481, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 103, + 481, + 140 + ], + "spans": [ + { + "bbox": [ + 129, + 103, + 481, + 140 + ], + "type": "text", + "content": "SpatialLLM: A Compound 3D-Informed Design towards Spatially-Intelligent Large Multimodal Models" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 122, + 160, + 487, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 122, + 160, + 487, + 175 + ], + "spans": [ + { + "bbox": [ + 122, + 160, + 487, + 175 + ], + "type": "text", + "content": "Wufei Ma1, Luoxin Ye1, Celso M de Melo2, Alan Yuille1, Jieneng Chen1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 177, + 177, + 430, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 177, + 177, + 430, + 190 + ], + "spans": [ + { + "bbox": [ + 177, + 177, + 430, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 177, + 177, + 430, + 190 + ], + "type": "text", + "content": "Johns Hopkins University " + }, + { + "bbox": [ + 177, + 177, + 430, + 190 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 177, + 177, + 430, + 190 + ], + "type": "text", + "content": "DEVCOM Army Research Laboratory" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 217, + 200, + 229 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 217, + 200, + 229 + ], + "spans": [ + { + "bbox": [ + 151, + 217, + 200, + 229 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 242, + 296, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 242, + 296, + 540 + ], + "spans": [ + { + "bbox": [ + 55, + 242, + 296, + 540 + ], + "type": "text", + "content": "Humans naturally understand 3D spatial relationships, enabling complex reasoning like predicting collisions of vehicles from different directions. Current large multimodal models (LMMs), however, lack of this capability of 3D spatial reasoning. This limitation stems from the scarcity of 3D training data and the bias in current model designs toward 2D data. In this paper, we systematically study the impact of 3D-informed data, architecture, and training setups, introducing SpatialLLM, a large multi-modal model with advanced 3D spatial reasoning abilities. To address data limitations, we develop two types of 3D-informed training datasets: (1) 3D-informed probing data focused on object's 3D location and orientation, and (2) 3D-informed conversation data for complex spatial relationships. Notably, we are the first to curate VQA data that incorporate 3D orientation relationships on real images. Furthermore, we systematically integrate these two types of training data with the architectural and training designs of LMMs, providing a roadmap for optimal design aimed at achieving superior 3D reasoning capabilities. Our SpatialLLM advances machines toward highly capable 3D-informed reasoning, surpassing GPT-4o performance by " + }, + { + "bbox": [ + 55, + 242, + 296, + 540 + ], + "type": "inline_equation", + "content": "8.7\\%" + }, + { + "bbox": [ + 55, + 242, + 296, + 540 + ], + "type": "text", + "content": ". Our systematic empirical design and the resulting findings offer valuable insights for future research in this direction. Our project page is available at this link." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 562, + 135, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 562, + 135, + 574 + ], + "spans": [ + { + "bbox": [ + 56, + 562, + 135, + 574 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 582, + 295, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 295, + 689 + ], + "type": "text", + "content": "When humans observe a scene, they perceive more than isolated objects; they intuitively understand the spatial relationships among these objects [55], enabling complex reasoning and interaction with the environment. For example, recognizing that a car is approaching an obstacle and might collide involves interpreting the 3D object orientations and spatial relationships. This innate ability to comprehend and reason about 3D space is crucial for navigating and interacting with the physical world." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "Despite significant advancements, current LMMs [1, 4, 39, 41, 57], exhibit limited capability of 3D spatial reason" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 317, + 215, + 473, + 275 + ], + "blocks": [ + { + "bbox": [ + 317, + 215, + 473, + 275 + ], + "lines": [ + { + "bbox": [ + 317, + 215, + 473, + 275 + ], + "spans": [ + { + "bbox": [ + 317, + 215, + 473, + 275 + ], + "type": "image", + "image_path": "aaad0f38892f2429e3ab001841f512bf0b3f249d95cdb0c266a2b6708fdcdd10.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 329, + 281, + 367, + 295 + ], + "lines": [ + { + "bbox": [ + 329, + 281, + 367, + 295 + ], + "spans": [ + { + "bbox": [ + 329, + 281, + 367, + 295 + ], + "type": "text", + "content": "Orientation Relationships" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 385, + 281, + 437, + 295 + ], + "lines": [ + { + "bbox": [ + 385, + 281, + 437, + 295 + ], + "spans": [ + { + "bbox": [ + 385, + 281, + 437, + 295 + ], + "type": "text", + "content": "Spatial Reasoning Relationships" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 474, + 217, + 542, + 258 + ], + "lines": [ + { + "bbox": [ + 474, + 217, + 542, + 258 + ], + "spans": [ + { + "bbox": [ + 474, + 217, + 542, + 258 + ], + "type": "text", + "content": "Distance Relationships The pedestrian is closer to the viewer than the cyclist." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 318, + 300, + 375, + 321 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 300, + 375, + 321 + ], + "spans": [ + { + "bbox": [ + 318, + 300, + 375, + 321 + ], + "type": "text", + "content": "White van is facing the pedestrian instead of the cyclist" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 383, + 300, + 468, + 321 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 383, + 300, + 468, + 321 + ], + "spans": [ + { + "bbox": [ + 383, + 300, + 468, + 321 + ], + "type": "text", + "content": "From the perspective of the pedestrian, the white van is on the right side of the pedestrian" + } + ] + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 475, + 263, + 544, + 299 + ], + "blocks": [ + { + "bbox": [ + 475, + 263, + 544, + 299 + ], + "lines": [ + { + "bbox": [ + 475, + 263, + 544, + 299 + ], + "spans": [ + { + "bbox": [ + 475, + 263, + 544, + 299 + ], + "type": "image", + "image_path": "6ab07807c283b9bfe2ec939f21710c29d7ab3e485090fe18cd7b6272006c25b8.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 474, + 325, + 548, + 332 + ], + "lines": [ + { + "bbox": [ + 474, + 325, + 548, + 332 + ], + "spans": [ + { + "bbox": [ + 474, + 325, + 548, + 332 + ], + "type": "text", + "content": "3D-Informed Architecture" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 320, + 326, + 351, + 369 + ], + "blocks": [ + { + "bbox": [ + 320, + 326, + 351, + 369 + ], + "lines": [ + { + "bbox": [ + 320, + 326, + 351, + 369 + ], + "spans": [ + { + "bbox": [ + 320, + 326, + 351, + 369 + ], + "type": "image", + "image_path": "50c06ac2dbca8f308a5a567d29b45d3beb18dc061831de4411333dfecb035bce.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 356, + 325, + 408, + 332 + ], + "lines": [ + { + "bbox": [ + 356, + 325, + 408, + 332 + ], + "spans": [ + { + "bbox": [ + 356, + 325, + 408, + 332 + ], + "type": "text", + "content": "3D-Informed Data" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "bbox": [ + 356, + 335, + 402, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 356, + 335, + 402, + 370 + ], + "spans": [ + { + "bbox": [ + 356, + 335, + 402, + 370 + ], + "type": "text", + "content": "1M 3D-Informed Probing Data description on depth, distance, & orientation" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 404, + 335, + 470, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 404, + 335, + 470, + 370 + ], + "spans": [ + { + "bbox": [ + 404, + 335, + 470, + 370 + ], + "type": "text", + "content": "1M 3D-Informed Instruction Tuning Data conversations about 3D spatial relationships" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 474, + 335, + 548, + 359 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 474, + 335, + 548, + 359 + ], + "spans": [ + { + "bbox": [ + 474, + 335, + 548, + 359 + ], + "type": "text", + "content": "Mixed encoders, 3D-aware visual encoders, etc. 3D-Informed Training" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 475, + 362, + 550, + 377 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 475, + 362, + 550, + 377 + ], + "spans": [ + { + "bbox": [ + 475, + 362, + 550, + 377 + ], + "type": "text", + "content": "Pre-pretraining, alignment & finetuning" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 387, + 555, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 387, + 555, + 453 + ], + "spans": [ + { + "bbox": [ + 313, + 387, + 555, + 453 + ], + "type": "text", + "content": "Figure 1. 3D spatial reasoning is crucial for LMMs to ground objects in 3D space and infer their 3D spatial relationships, such as distance, orientation, and spatial interactions. We systematically study the impact of 3D-informed data, architecture, and training setups, introducing SpatialLLM, an LMM with advanced 3D spatial reasoning abilities." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 480, + 554, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 480, + 554, + 528 + ], + "spans": [ + { + "bbox": [ + 313, + 480, + 554, + 528 + ], + "type": "text", + "content": "ing. These models excel at identifying objects and generating descriptive captions but struggle with tasks that require intricate 3D reasoning, such as predicting object interactions or understanding physical events." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 531, + 555, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 531, + 555, + 662 + ], + "spans": [ + { + "bbox": [ + 313, + 531, + 555, + 662 + ], + "type": "text", + "content": "The primary challenge lies in the scarcity of high-quality 3D spatial relationship data, which prevents models from learning effective 3D spatial reasoning. Collecting a small set of high-quality 3D-aware data to tackle the first challenge is feasible, albeit labor-intensive, using readily available tools. Recent efforts [2, 14] leverage vision foundation models to recognize [67], detect [42], segment [31] and estimate the depth [10, 63] in images. These approaches form semi-automated pipelines to construct pseudo-2.5D datasets for training LMMs with spatial reasoning capabilities, such as depth ordering [14, 58]." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "type": "text", + "content": "However, a significant gap remains: previous works [2, 14, 16] have primarily focused on 3D distance relationships, overlooking the crucial role of 3D object orientation. This is primarily due to the absence of robust 6D pose estimation." + } + ] + } + ], + "index": 27 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17249" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": "tors. Understanding object orientation is essential for complex spatial reasoning tasks, such as determining if two cars are facing the same direction to avoid dangerous collision. Addressing this gap, we aim to incorporate 3D orientation relationships—converted from ImageNet3D [17]—into our data engine, making us the first to enable complex spatial reasoning involving 3D orientation relationships." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 157, + 294, + 253 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 157, + 294, + 253 + ], + "spans": [ + { + "bbox": [ + 55, + 157, + 294, + 253 + ], + "type": "text", + "content": "Even with 3D-informed data in hand, fully unlocking its potential remains challenging due to the intricate, multistage development process required for an LMM. For example, SpatialVLM [14] utilizes 3D-informed data during the instruction tuning stage, but it remains uncertain whether such data can enhance other stages like vision encoder pretraining or multimodal alignment. This limitation suggests that a more holistic approach is necessary." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 255, + 294, + 302 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 255, + 294, + 302 + ], + "spans": [ + { + "bbox": [ + 55, + 255, + 294, + 302 + ], + "type": "text", + "content": "In this paper, we present a systematic investigation into building LMMs with enhanced 3D spatial reasoning capabilities through a compound design approach, as illustrated in Fig. 1. Our key contributions are threefold:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 303, + 294, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 303, + 294, + 399 + ], + "spans": [ + { + "bbox": [ + 55, + 303, + 294, + 399 + ], + "type": "text", + "content": "First, we identify three fundamental types of 3D spatial relationships that LMMs must master: distance relationships, orientation relationships, and complex spatial reasoning that combines both. To evaluate these capabilities, we introduce SpatialVQA, a comprehensive benchmark containing 1,323 questions that target 3D spatial understanding, built upon diverse urban and indoor scenes from Omni3D [11]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 400, + 294, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 400, + 294, + 532 + ], + "spans": [ + { + "bbox": [ + 55, + 400, + 294, + 532 + ], + "type": "text", + "content": "Second, we propose a novel compound 3D-informed design that introduces improvements across multiple dimensions, leading to our proposed SpatialLLM model. On the data front, we develop two types of 3D-informed training datasets: (1) 3D-informed probing data focused on object-level 3D properties, and (2) 3D-informed conversation data for complex spatial relationships. Architecturally, we explore mixing different visual encoders and investigate their 3D awareness capabilities. For training, we introduce 3D-informed data at multiple stages: pre-pretraining, multimodal alignment, and instruction tuning." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 533, + 294, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 533, + 294, + 675 + ], + "spans": [ + { + "bbox": [ + 55, + 533, + 294, + 675 + ], + "type": "text", + "content": "Third, we present the first comprehensive search over the LMM design space for spatial reasoning tasks and propose a roadmap towards developing state-of-the-art models in this domain. The experiment demonstrates that our proposed SpatialLLM achieves state-of-the-art performance on spatial reasoning tasks, surpassing both GPT-4o [1] (by " + }, + { + "bbox": [ + 55, + 533, + 294, + 675 + ], + "type": "inline_equation", + "content": "8.7\\%" + }, + { + "bbox": [ + 55, + 533, + 294, + 675 + ], + "type": "text", + "content": " ) and spatial-centric models like SpatialVLM [14] (by " + }, + { + "bbox": [ + 55, + 533, + 294, + 675 + ], + "type": "inline_equation", + "content": "10.5\\%" + }, + { + "bbox": [ + 55, + 533, + 294, + 675 + ], + "type": "text", + "content": "). Our systematic analysis reveals that while architectural improvements contribute a " + }, + { + "bbox": [ + 55, + 533, + 294, + 675 + ], + "type": "inline_equation", + "content": "1.3\\%" + }, + { + "bbox": [ + 55, + 533, + 294, + 675 + ], + "type": "text", + "content": " gain, the strategic incorporation of 3D-informed data across different training stages yields a substantial " + }, + { + "bbox": [ + 55, + 533, + 294, + 675 + ], + "type": "inline_equation", + "content": "13.7\\%" + }, + { + "bbox": [ + 55, + 533, + 294, + 675 + ], + "type": "text", + "content": " performance improvement." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 677, + 294, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 294, + 713 + ], + "type": "text", + "content": "Our findings highlight the importance of a holistic approach to improving LMMs' 3D spatial reasoning capabilities, demonstrating that both architectural choices and care" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 553, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 131 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 131 + ], + "type": "text", + "content": "fully curated open-source 3D-informed training data play crucial roles in achieving superior performance. This work establishes a new benchmark for spatial reasoning in LMMs and provides valuable insights for future research in this direction." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 143, + 406, + 156 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 143, + 406, + 156 + ], + "spans": [ + { + "bbox": [ + 314, + 143, + 406, + 156 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 163, + 553, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 163, + 553, + 366 + ], + "spans": [ + { + "bbox": [ + 313, + 163, + 553, + 366 + ], + "type": "text", + "content": "Large multimodal models (LMMs). Recently, there has been a significant surge in research on multimodal LLMs, exemplified by products like GPT-4 [1] and Gemini [57], as well as open-source models such as LLaVA [34, 39, 41] and Qwen-VL [8]. These open-source models combine pretrained vision encoders [15, 27, 47, 48] with LLMs [7, 28, 59], enabling the language models to handle a wide range of tasks that involve images as input. Typically, these methods have the alignment stage that aligns image features with the input embedding space of the language model using connectors [13, 35, 39], trained with paired images and captions [53]. The instruction fine-tuning [41] stage uses data that contains both images and instructions to enhances LMMs. However, current LMMs are unsuccessful in spatial visual reasoning, due to limited spatial-aware data and stagnant 3D-informed architecture design. We aim to bridge this research gap." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 367, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 367, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 367, + 553, + 713 + ], + "type": "text", + "content": "3D spatial reasoning with LMMs. Understanding spatial relationships between objects in images is crucial for visual reasoning. While this comes naturally to humans, LMMs were found to struggle with recognizing spatial relations [49, 60, 61], due to a lack of 3D-aware spatial-related data during training. While neural symbolic methods [60] achieved improved performance on synthetic benchmarks by involving explicit 3D reasoning modules [26, 44], they struggled to generalize to a broader range of objects found in real images. Fortunately, with foundation models like SAM [31] and depth estimators [10, 63], semi-automatic annotation for 3D-aware VQA data has been adopted in recent studies [14, 16]. SpatialVLM [14] uses the pseudo-annotated VQA data for training only while SpatialRGPT [14, 16] further augments LLaVA with extra input of regional mask and depth map during inference, which is not our focus. Although human-annotated 3D-aware VQA data is rare, we aim to leverage human-annotated 3D object-level datasets (like ImageNet3D [45]) by simply converting image-pose pairs into image-text pair data. However, even with the collected data, a major challenge remains in effectively utilizing 3D-aware data to train optimal models. Current approaches like SpatialVLM [14] use the standard LLaVA and two-stage training. To date, no work has addressed when (which training stage) and where (which network component) are optimal to inject the 3D-awareness to LMMs. We bridge this gap by conducting a compound design that simultaneously considers 3D-aware data, architecture, and the training, leading to a best" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17250" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 109 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 109 + ], + "type": "text", + "content": "performing design. Rather than competing, our study has the potential to complement prior works [14, 16] by offering new compound design recipe." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 122, + 116, + 134 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 122, + 116, + 134 + ], + "spans": [ + { + "bbox": [ + 55, + 122, + 116, + 134 + ], + "type": "text", + "content": "3. Methods" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 143, + 295, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 143, + 295, + 227 + ], + "spans": [ + { + "bbox": [ + 55, + 143, + 295, + 227 + ], + "type": "text", + "content": "In this section we start by reviewing LMMs in Sec. 3.1, specifically the architecture and the choice of training data and strategies. We present the task of reasoning 3D spatial relationships and explain the challenges LMMs face when answering these questions in Sec. 3.2. Then we study how to address these challenging problems with LMMs and introduce our SpatialLLM in Sec. 3.3." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 237, + 181, + 251 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 237, + 181, + 251 + ], + "spans": [ + { + "bbox": [ + 55, + 237, + 181, + 251 + ], + "type": "text", + "content": "3.1. Preliminary of LMMs" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 256, + 295, + 291 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 256, + 295, + 291 + ], + "spans": [ + { + "bbox": [ + 55, + 256, + 295, + 291 + ], + "type": "text", + "content": "Data. The success of deep learning in vision is largely due to well-curated data [20, 56], and LMMs similarly depend on this foundation, albeit with diverse data formats below." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 293, + 295, + 508 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 56, + 293, + 295, + 341 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 293, + 295, + 341 + ], + "spans": [ + { + "bbox": [ + 56, + 293, + 295, + 341 + ], + "type": "text", + "content": "- Image-only data, such as ImageNet [17], is commonly used for MAE [23]/DINOv2 [47] pretraining with a self-supervised objective. This step enables the model to learn rich visual features solely from visual signals." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 341, + 295, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 341, + 295, + 376 + ], + "spans": [ + { + "bbox": [ + 56, + 341, + 295, + 376 + ], + "type": "text", + "content": "- Noisy image-text pairs: Large-scale image-text pairs [20, 37, 52] are used for pretraining with contrastive loss (e.g., CLIP [48, 65]) to align visual and textual representations." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 377, + 295, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 377, + 295, + 449 + ], + "spans": [ + { + "bbox": [ + 56, + 377, + 295, + 449 + ], + "type": "text", + "content": "- Multimodal alignment: At this stage, image-caption pairs (e.g., CC3M [53]) are utilized to align visual features with the language model's word embedding space, typically through image-captioning tasks. This pretraining ensures that the model can interpret and relate visual content to textual queries in a shared semantic space." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 449, + 295, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 449, + 295, + 508 + ], + "spans": [ + { + "bbox": [ + 56, + 449, + 295, + 508 + ], + "type": "text", + "content": "- Multimodal instruction-following data. Visual instruction tuning data is essential but challenging to collect, as it rarely exists in its natural form online. Prior works [39, 41, 46, 58] repurpose VQA benchmarks [22, 32] into instruction-tuning datasets." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 509, + 295, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 509, + 295, + 592 + ], + "spans": [ + { + "bbox": [ + 55, + 509, + 295, + 592 + ], + "type": "text", + "content": "Architecture. A standard LMM [39, 41] consists of a visual encoder to process the image, a multimodal connector to transform the visual feature to visual token, and a LLM for reasoning. However, the extent to which these architectural components contribute to the spatial understanding of LMMs remains unknown. We review these architectural components as below." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 594, + 295, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 56, + 594, + 295, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 594, + 295, + 654 + ], + "spans": [ + { + "bbox": [ + 56, + 594, + 295, + 654 + ], + "type": "text", + "content": "- Vision encoder: Most LMMs rely on language-supervised models like CLIP [48], leveraging web-scale noisy image-text data. Recent LMMs [43, 58] integrate vision models trained only on visual signals, including self-supervised [23, 47, 69] and generative models [51]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 654, + 295, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 654, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 654, + 295, + 689 + ], + "type": "text", + "content": "- Connector: Features from a visual encoder are mapped into the LLM token space by connectors [3, 13, 35, 41] such as a MLP [39, 41]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 689, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 689, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 56, + 689, + 295, + 714 + ], + "type": "text", + "content": "- LLMs: LLMs [1, 7, 59] serve as the foundational backbone of LMMs, driving their reasoning capabilities." + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 322, + 72, + 553, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 72, + 553, + 96 + ], + "spans": [ + { + "bbox": [ + 322, + 72, + 553, + 96 + ], + "type": "text", + "content": "Open-source LLMs like Llama series [18, 59] are frequently adopted to enhance the reasoning capability." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 97, + 553, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 97, + 553, + 156 + ], + "spans": [ + { + "bbox": [ + 313, + 97, + 553, + 156 + ], + "type": "text", + "content": "Training. The aforementioned data supports pretraining for strong visual representation learning, alignment pretraining to synchronize visual and language representations, and instruction tuning to enable multimodal instruction-following and reasoning capabilities" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 314, + 157, + 553, + 324 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 314, + 157, + 553, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 157, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 314, + 157, + 553, + 217 + ], + "type": "text", + "content": "- Pre-pretraining. This stage focuses on developing foundational visual representations, often with reconstruction-based objectives (e.g., MAE [23], DINOv2 [47]). This step prepares the model to handle more complex multimodal tasks by first refining its visual understanding." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 314, + 217, + 553, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 217, + 553, + 252 + ], + "spans": [ + { + "bbox": [ + 314, + 217, + 553, + 252 + ], + "type": "text", + "content": "- Multimodal alignment pretraining. At this stage, the model is trained to describe images in details to align visual and language representations in the same space." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 314, + 253, + 553, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 253, + 553, + 324 + ], + "spans": [ + { + "bbox": [ + 314, + 253, + 553, + 324 + ], + "type": "text", + "content": "- Visual instruction tuning. The final stage involves refining the model's ability to process complex multimodal instructions using visual instruction tuning data. This often incorporates transformed VQA datasets, enhancing the model's capacity in handling tasks that demand coordinated reasoning across language and vision." + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 333, + 503, + 346 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 333, + 503, + 346 + ], + "spans": [ + { + "bbox": [ + 313, + 333, + 503, + 346 + ], + "type": "text", + "content": "3.2. Reasoning 3D Spatial Relationships" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 351, + 555, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 351, + 555, + 471 + ], + "spans": [ + { + "bbox": [ + 313, + 351, + 555, + 471 + ], + "type": "text", + "content": "3D spatial reasoning analyzes the spatial relationships between objects in the 3D world space, such as objects being far or close to each other, or one object lies to the left or right of another object [60]. It extends beyond standard 3D tasks, such as 3D object detection and 3D pose estimation that studies the 3D nature of individual objects, and focuses on understanding their mutual relationships in 3D. In this work we identify three main types of 3D spatial relationships, i.e., distance, orientation, and spatial reasoning, and study how we can better address these problems with LMMs." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 487, + 555, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 487, + 555, + 654 + ], + "spans": [ + { + "bbox": [ + 313, + 487, + 555, + 654 + ], + "type": "text", + "content": "3D spatial relationships. For distance spatial relationships, we study if models can estimate and compare 3D distances between objects. These spatial relationships require more than depth awareness, which gives the distance between the viewer and the object and can be roughly estimated from the object scale, but also the 3D distances between two objects that are often harder to estimate. On the other hand, orientation spatial relationships require models to understand the 3D orientations of an object, e.g., predicting if two objects are facing the same direction. Lastly, spatial reasoning spatial relationships look into 3D spatial relationships that need a combination of 3D awareness of location and orientation and then performing spatial reasoning on the 3D information." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 662, + 489, + 674 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 662, + 489, + 674 + ], + "spans": [ + { + "bbox": [ + 313, + 662, + 489, + 674 + ], + "type": "text", + "content": "3.2.1. Challenges of 3D spatial reasoning" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "content": "Reasoning about 3D spatial relationships remains a challenging task for state-of-the-art open-sourced and proprietary LMMs. We attribute this to two main limitations in" + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17251" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 70, + 70, + 173, + 145 + ], + "blocks": [ + { + "bbox": [ + 70, + 70, + 173, + 145 + ], + "lines": [ + { + "bbox": [ + 70, + 70, + 173, + 145 + ], + "spans": [ + { + "bbox": [ + 70, + 70, + 173, + 145 + ], + "type": "image", + "image_path": "06f567a187c83769ca94b276493b3c978c825a3c7e72462d7e72e61fcf483f85.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 190, + 77, + 281, + 85 + ], + "lines": [ + { + "bbox": [ + 190, + 77, + 281, + 85 + ], + "spans": [ + { + "bbox": [ + 190, + 77, + 281, + 85 + ], + "type": "text", + "content": "√3D Location 03D Orientation" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 200, + 88, + 212, + 95 + ], + "lines": [ + { + "bbox": [ + 200, + 88, + 212, + 95 + ], + "spans": [ + { + "bbox": [ + 200, + 88, + 212, + 95 + ], + "type": "text", + "content": "User" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 57, + 148, + 186, + 223 + ], + "blocks": [ + { + "bbox": [ + 195, + 130, + 293, + 144 + ], + "lines": [ + { + "bbox": [ + 195, + 130, + 293, + 144 + ], + "spans": [ + { + "bbox": [ + 195, + 130, + 293, + 144 + ], + "type": "text", + "content": "The chair in Region [2] is closer to the chair in Region [0]." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 57, + 148, + 186, + 223 + ], + "lines": [ + { + "bbox": [ + 57, + 148, + 186, + 223 + ], + "spans": [ + { + "bbox": [ + 57, + 148, + 186, + 223 + ], + "type": "image", + "image_path": "f5d09c711e117ef47c7b87625ca17b09a1a4a1e550c307ebe432c1bc91674eff.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 190, + 150, + 298, + 164 + ], + "lines": [ + { + "bbox": [ + 190, + 150, + 298, + 164 + ], + "spans": [ + { + "bbox": [ + 190, + 150, + 298, + 164 + ], + "type": "text", + "content": "Spatial Reasoning Relationships " + }, + { + "bbox": [ + 190, + 150, + 298, + 164 + ], + "type": "inline_equation", + "content": "\\sqrt{3}" + }, + { + "bbox": [ + 190, + 150, + 298, + 164 + ], + "type": "text", + "content": " Location " + }, + { + "bbox": [ + 190, + 150, + 298, + 164 + ], + "type": "inline_equation", + "content": "\\sqrt{3}" + }, + { + "bbox": [ + 190, + 150, + 298, + 164 + ], + "type": "text", + "content": " Orientation" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 200, + 167, + 226, + 174 + ], + "lines": [ + { + "bbox": [ + 200, + 167, + 226, + 174 + ], + "spans": [ + { + "bbox": [ + 200, + 167, + 226, + 174 + ], + "type": "text", + "content": "User" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 195, + 176, + 295, + 196 + ], + "lines": [ + { + "bbox": [ + 195, + 176, + 295, + 196 + ], + "spans": [ + { + "bbox": [ + 195, + 176, + 295, + 196 + ], + "type": "text", + "content": "If I stand at the location of the car in Region [0] facing where it is facing, where is the bus in Region [1]?" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 195, + 209, + 292, + 222 + ], + "lines": [ + { + "bbox": [ + 195, + 209, + 292, + 222 + ], + "spans": [ + { + "bbox": [ + 195, + 209, + 292, + 222 + ], + "type": "text", + "content": "The bus in Region [1] is on the left side of the car in Region [0]." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 79, + 234, + 529, + 246 + ], + "lines": [ + { + "bbox": [ + 79, + 234, + 529, + 246 + ], + "spans": [ + { + "bbox": [ + 79, + 234, + 529, + 246 + ], + "type": "text", + "content": "Figure 2. Examples from our SpatialVQA benchmark featuring a broad range of questions that require 3D spatial reasoning." + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 307, + 71, + 436, + 145 + ], + "blocks": [ + { + "bbox": [ + 194, + 98, + 298, + 117 + ], + "lines": [ + { + "bbox": [ + 194, + 98, + 298, + 117 + ], + "spans": [ + { + "bbox": [ + 194, + 98, + 298, + 117 + ], + "type": "text", + "content": "Which object is closer to the chair in Region [0], the bookcase in Region [1] or the chair in Region [2]?" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 271, + 121, + 298, + 128 + ], + "lines": [ + { + "bbox": [ + 271, + 121, + 298, + 128 + ], + "spans": [ + { + "bbox": [ + 271, + 121, + 298, + 128 + ], + "type": "text", + "content": "3DI-LMM" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 307, + 71, + 436, + 145 + ], + "lines": [ + { + "bbox": [ + 307, + 71, + 436, + 145 + ], + "spans": [ + { + "bbox": [ + 307, + 71, + 436, + 145 + ], + "type": "image", + "image_path": "a41c6ea95a9124d7247030d0b993c6e50fb2f699ac0f49311a988ab340cde7dc.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 441, + 77, + 531, + 84 + ], + "lines": [ + { + "bbox": [ + 441, + 77, + 531, + 84 + ], + "spans": [ + { + "bbox": [ + 441, + 77, + 531, + 84 + ], + "type": "text", + "content": "√3D Location √3D Orientation" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 450, + 95, + 461, + 101 + ], + "lines": [ + { + "bbox": [ + 450, + 95, + 461, + 101 + ], + "spans": [ + { + "bbox": [ + 450, + 95, + 461, + 101 + ], + "type": "text", + "content": "Use" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 444, + 103, + 550, + 118 + ], + "lines": [ + { + "bbox": [ + 444, + 103, + 550, + 118 + ], + "spans": [ + { + "bbox": [ + 444, + 103, + 550, + 118 + ], + "type": "text", + "content": "Are the car in Region [0] and the car in Region [1] facing the same direction, or different directions?" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 445, + 130, + 542, + 144 + ], + "lines": [ + { + "bbox": [ + 445, + 130, + 542, + 144 + ], + "spans": [ + { + "bbox": [ + 445, + 130, + 542, + 144 + ], + "type": "text", + "content": "The car in Region [0] and the car in Region [1] are facing different directions." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 307, + 148, + 436, + 223 + ], + "blocks": [ + { + "bbox": [ + 271, + 200, + 298, + 206 + ], + "lines": [ + { + "bbox": [ + 271, + 200, + 298, + 206 + ], + "spans": [ + { + "bbox": [ + 271, + 200, + 298, + 206 + ], + "type": "text", + "content": "3DI-LMM" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 307, + 148, + 436, + 223 + ], + "lines": [ + { + "bbox": [ + 307, + 148, + 436, + 223 + ], + "spans": [ + { + "bbox": [ + 307, + 148, + 436, + 223 + ], + "type": "image", + "image_path": "17b5d311c6cbe245dfc4975a68277ca32ad4e9e28fc260aca93e1ccafaac99e4.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 441, + 151, + 522, + 157 + ], + "lines": [ + { + "bbox": [ + 441, + 151, + 522, + 157 + ], + "spans": [ + { + "bbox": [ + 441, + 151, + 522, + 157 + ], + "type": "text", + "content": "Spatial Reasoning Relationships" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 441, + 157, + 531, + 163 + ], + "lines": [ + { + "bbox": [ + 441, + 157, + 531, + 163 + ], + "spans": [ + { + "bbox": [ + 441, + 157, + 531, + 163 + ], + "type": "text", + "content": "√3D Location √3D Orientation" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 450, + 173, + 463, + 179 + ], + "lines": [ + { + "bbox": [ + 450, + 173, + 463, + 179 + ], + "spans": [ + { + "bbox": [ + 450, + 173, + 463, + 179 + ], + "type": "text", + "content": "User" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 444, + 182, + 547, + 196 + ], + "lines": [ + { + "bbox": [ + 444, + 182, + 547, + 196 + ], + "spans": [ + { + "bbox": [ + 444, + 182, + 547, + 196 + ], + "type": "text", + "content": "Is the car in Region [0] oriented towards the car in Region [1] or the car in Region [2]?" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 522, + 200, + 548, + 205 + ], + "lines": [ + { + "bbox": [ + 522, + 200, + 548, + 205 + ], + "spans": [ + { + "bbox": [ + 522, + 200, + 548, + 205 + ], + "type": "text", + "content": "3DI-LMM" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 444, + 209, + 547, + 222 + ], + "lines": [ + { + "bbox": [ + 444, + 209, + 547, + 222 + ], + "spans": [ + { + "bbox": [ + 444, + 209, + 547, + 222 + ], + "type": "text", + "content": "The car in Region [0] is facing towards the car in Region [1]." + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "bbox": [ + 55, + 266, + 295, + 291 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 266, + 295, + 291 + ], + "spans": [ + { + "bbox": [ + 55, + 266, + 295, + 291 + ], + "type": "text", + "content": "modern LMMs, the lack of 3D awareness and inability to perform 3D spatial reasoning." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 55, + 306, + 296, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 306, + 296, + 426 + ], + "spans": [ + { + "bbox": [ + 55, + 306, + 296, + 426 + ], + "type": "text", + "content": "3D awareness refers to the ability of a visual encoder to represent 3D-aware features such as 3D object shapes and 3D orientations. Previous studies resorted to proxy tasks, e.g., part correspondence [19] and pose estimation [45], to quantitatively evaluate the 3D awareness of visual foundation models with linear probing. Results showed that most visual encoder, specifically ones trained with language supervision, demonstrated limited 3D awareness of 3D object poses [45], which inevitably limits the 3D spatial reasoning capabilities of LMMs built upon these encoders." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 55, + 441, + 296, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 441, + 296, + 597 + ], + "spans": [ + { + "bbox": [ + 55, + 441, + 296, + 597 + ], + "type": "text", + "content": "Inability to perform 3D spatial reasoning. To infer the 3D spatial relationships, models must also be able to perform 3D spatial reasoning from 3D-aware visual features. Despite the abundant Internet-scale data for multi-modal training, we argue that high-quality 3D spatial reasoning data are scarce. Existing pretraining and visual instruction tuning data for LMMs [41, 58] focused on detailed descriptions and conversations about scenes, appearances, and actions, while being vague about the 3D spatial relationships that build up the high-level physical events. Despite their success on a wide range of image understanding tasks, these LMMs fail to learn the nature of 3D scenes or to perform 3D spatial reasoning." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 55, + 603, + 200, + 614 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 603, + 200, + 614 + ], + "spans": [ + { + "bbox": [ + 55, + 603, + 200, + 614 + ], + "type": "text", + "content": "3.2.2. SpatialVQA for Evaluation" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 55, + 618, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 296, + 713 + ], + "type": "text", + "content": "Despite the rising interests in spatial reasoning, there lacks a comprehensive open source benchmark for reasoning 3D spatial relationships. Early datasets on spatial reasoning were built on 3D scans rather than images [6, 64]. Recent visual-language benchmarks on spatial reasoning either focused on 2D spatial relationships [14, 16], e.g., left or right in the image plane, or only on distance spatial relationships [58]." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 313, + 266, + 555, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 266, + 555, + 458 + ], + "spans": [ + { + "bbox": [ + 313, + 266, + 555, + 458 + ], + "type": "text", + "content": "To enable a thorough evaluation of the 3D spatial reasoning capabilities of LMMs, we introduce SpatialVQA, a visual question answering benchmark that covers a wide variety of 3D spatial relationships. Our SpatialVQA distinguishes itself from all previous spatial reasoning benchmarks in the sense that all questions require different levels of 3D awareness and cannot be answered from 2D spatial reasoning only. Specifically, SpatialVQA consists of 1,323 questions, ranging from questions can be answered from distance-awareness only, e.g., which object is closer to a third object, to questions that require estimating objects' 3D orientations, e.g., if two objects are facing the same direction, and eventually questions that require complex 3D spatial reasoning over a combination of object 3D locations and orientations, e.g., an object is in the front/left/back/right direction of another object." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 313, + 460, + 554, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 460, + 554, + 544 + ], + "spans": [ + { + "bbox": [ + 313, + 460, + 554, + 544 + ], + "type": "text", + "content": "We build our SpatialVQA on images from Omni3D [11], with 3D bounding box annotations on diverse objects from both urban [12, 21] and indoor scenes [9, 50, 54]. We follow [16, 58] and develop rule-based methods to generate visual question-answer pairs from the 3D groundtruths. Please refer to Fig. 2 for some examples of SpatialVQA and to appendix Sec.B for details about our benchmark." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 313, + 561, + 488, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 561, + 488, + 574 + ], + "spans": [ + { + "bbox": [ + 313, + 561, + 488, + 574 + ], + "type": "text", + "content": "3.3. Compound 3D-Informed Design" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "content": "Problem formulation. While previous approaches focused on generating 3D-informed instruction tuning data to enable LMMs with the abilities to estimate distances, depths, or spatial relationships [14, 16], it is largely understudied the best recipe for developing LMMs with strong 3D spatial reasoning capabilities. Motivated by the two aforementioned limitations of LMMs in Sec. 3.2.1, we consider two main aspects in our compound 3D-informed design – the architecture design that leads to visual encoders with strong 3D awareness and the training setup that advances the 3D spatial reasoning capabilities of LMMs." + } + ] + } + ], + "index": 33 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17252" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 55, + 70, + 294, + 168 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 294, + 168 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 294, + 168 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 294, + 168 + ], + "type": "image", + "image_path": "aa085b7ec1966899e356d8745d3fa88c93bdcb56458e2df75f569419bac3f58f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 63, + 174, + 90, + 183 + ], + "lines": [ + { + "bbox": [ + 63, + 174, + 90, + 183 + ], + "spans": [ + { + "bbox": [ + 63, + 174, + 90, + 183 + ], + "type": "text", + "content": "(a) data" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 111, + 174, + 168, + 184 + ], + "lines": [ + { + "bbox": [ + 111, + 174, + 168, + 184 + ], + "spans": [ + { + "bbox": [ + 111, + 174, + 168, + 184 + ], + "type": "text", + "content": "(b) training stage" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 178, + 174, + 228, + 184 + ], + "lines": [ + { + "bbox": [ + 178, + 174, + 228, + 184 + ], + "spans": [ + { + "bbox": [ + 178, + 174, + 228, + 184 + ], + "type": "text", + "content": "(c) architecture" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 235, + 174, + 295, + 184 + ], + "lines": [ + { + "bbox": [ + 235, + 174, + 295, + 184 + ], + "spans": [ + { + "bbox": [ + 235, + 174, + 295, + 184 + ], + "type": "text", + "content": "(d) vision encoder" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 58, + 264, + 173, + 406 + ], + "blocks": [ + { + "bbox": [ + 55, + 196, + 296, + 251 + ], + "lines": [ + { + "bbox": [ + 55, + 196, + 296, + 251 + ], + "spans": [ + { + "bbox": [ + 55, + 196, + 296, + 251 + ], + "type": "text", + "content": "Figure 3. Design space for LMMs capable of spatial reasoning. The dashed boxes and lines highlight our new design space compared to LLaVA-v1.5. This compound design simultaneously considers 3D-informed data, architecture, and training methods to search for the best-performing models for spatial reasoning." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 58, + 264, + 173, + 406 + ], + "lines": [ + { + "bbox": [ + 58, + 264, + 173, + 406 + ], + "spans": [ + { + "bbox": [ + 58, + 264, + 173, + 406 + ], + "type": "image", + "image_path": "458bf0902d6b89f5e588a589b82a63450156b0daed0108466939cec495bbf875.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 418, + 296, + 517 + ], + "lines": [ + { + "bbox": [ + 55, + 418, + 296, + 517 + ], + "spans": [ + { + "bbox": [ + 55, + 418, + 296, + 517 + ], + "type": "text", + "content": "Figure 4. Comparison of data cards showcasing the curation process and data types: standard LLaVA data (left) vs. our 3D-informed data (right). For our 3D-informed data curation, we label depth and distance of objects in OpenImage [33] using semi-automatic tools, resulting in 3DI-Pb-OI1M. Beyond this, we introduce 3D orientation probing data derived from human-annotated ImageNet3D [45], resulting in 3DI-Pb-IN166K for alignment. We further augment the conversation about spatial relationship, leading to 3DI-Ft-1M for instruction tuning." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 179, + 264, + 293, + 405 + ], + "blocks": [ + { + "bbox": [ + 179, + 264, + 293, + 405 + ], + "lines": [ + { + "bbox": [ + 179, + 264, + 293, + 405 + ], + "spans": [ + { + "bbox": [ + 179, + 264, + 293, + 405 + ], + "type": "image", + "image_path": "4dc08d9a73e6c392cac4a36ceb33b3fc99b84c1476af4dd2ec4a1c6a7c9dc97a.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 539, + 139, + 551 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 539, + 139, + 551 + ], + "spans": [ + { + "bbox": [ + 55, + 539, + 139, + 551 + ], + "type": "text", + "content": "3.3.1. Design space" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 554, + 296, + 603 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 554, + 296, + 603 + ], + "spans": [ + { + "bbox": [ + 55, + 554, + 296, + 603 + ], + "type": "text", + "content": "We introduce the design space considered in our work, i.e., choices of training data, model architecture, and training setup that advance the 3D spatial reasoning capabilities of LMMs. Refer to Fig. 3 for an overview of our design space." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 617, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 296, + 715 + ], + "type": "text", + "content": "Data. Standard LMMs follow LLaVA [41] and adopte feature alignment data for connector pretraining and visual instruction data for supervised instruction tuning. These datasets focus on detailed descriptions of images or machine-generated multi-turn conversations. Besides standard visual-language datasets, we further consider 3D-informed data, including 3D-aware probing data and 3D-informed instruction tuning data. Inspired by [45], the 3D" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 72, + 553, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 191 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 191 + ], + "type": "text", + "content": "aware probing data involves fundamental object-level 3D properties, such as the azimuth and elevation 3D viewpoint of an object, depth of the object from the viewer, and 3D distances between two objects. The goal of the 3D-aware probing data is to effectively inject fundamental 3D awareness into the models and benefit subsequent reasoning. Meanwhile, the 3D-informed instruction tuning data focuses on high-level 3D spatial relationships between the objects, which are built on a combination of the fundamental 3D properties and 3D spatial reasoning." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 192, + 555, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 192, + 555, + 324 + ], + "spans": [ + { + "bbox": [ + 313, + 192, + 555, + 324 + ], + "type": "text", + "content": "To generate the 3D-aware probing data and the 3D-informed instruction tuning data, we follow the previous data generation pipelines [14, 16] and exploit visual foundation models to extract and tag the objects in 3D space. Furthermore, we adopt a pretrained 6D pose estimator in [45] and estimate 3D orientations of the objects, which provides the crucial 3D information for generating 3D-informed data regarding various orientation-related 3D spatial relationships. We illustrate our 3D-informed data generation pipeline in Fig. 4 and refer the readers to appendex Sec.A for details of our 3D-informed data generation." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 340, + 556, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 340, + 556, + 449 + ], + "spans": [ + { + "bbox": [ + 313, + 340, + 556, + 449 + ], + "type": "text", + "content": "Architecture. Inspired by Probe3D [19] which probes the 3D-awareness of visual foundation models, we study the design of visual encoder to enable better 3D spatial reasoning. We consider two types of visual encoders: (i) Frozen & pretrained visual encoder CLIP [48] following [41] but with the option to mix a wider variety of models, such as MAE [23] and DINOv2 [47], and (ii) 3D-aware visual encoders finetuned with our 3D-informed data, e.g., contrastive pretraining or finetuning with LoRA using language modeling loss." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 464, + 556, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 464, + 556, + 608 + ], + "spans": [ + { + "bbox": [ + 313, + 464, + 556, + 608 + ], + "type": "text", + "content": "Training setup. To address the limitations of 3D-aware LMMs discussed in Sec. 3.2.1, we propose new training setups that aim to improve 3D awareness and advance the 3D spatial reasoning capabilities. We consider the following: (i) in Stage 0: use 3D-informed probing data (3DI-Pb-OI1M/1N166K) to tune the Lora layers of CLIP vision encoder (Pre-pretraining); (ii) in Stage 1: 3D-informed multimodal alignment using a combination of standard CC558K [41] and our 3D-informed probing data (3DI-Pb-OI1M/1N166K); and (iii) in Stage 2: 3D-informed instruction tuning with the standard Mix665K [41] and our 3D-informed visual instruction tuning data (3DI-Ft-1M)." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 614, + 529, + 627 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 614, + 529, + 627 + ], + "spans": [ + { + "bbox": [ + 313, + 614, + 529, + 627 + ], + "type": "text", + "content": "3.3.2. Designing 3D-Informed LMM: A Roadmap" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "content": "Design instantiation and comparison. The mentioned design space can result in several design instantiations, as depicted in Fig. 5. In a multimodal LLM, a pre-trained CLIP visual encoder extracts grid features from the input image. These features are converted into visual tokens via a connector and then processed by the LLM for visual reasoning. Fig. 5 (b) shows that LLaVA is a standard design space" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "17253" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 70, + 82, + 298, + 165 + ], + "blocks": [ + { + "bbox": [ + 70, + 82, + 298, + 165 + ], + "lines": [ + { + "bbox": [ + 70, + 82, + 298, + 165 + ], + "spans": [ + { + "bbox": [ + 70, + 82, + 298, + 165 + ], + "type": "image", + "image_path": "af92f889ccc138fcba6a6605c945b4eb095cf3f72f3750ce067312d2cfe9ce92.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 138, + 171, + 225, + 182 + ], + "lines": [ + { + "bbox": [ + 138, + 171, + 225, + 182 + ], + "spans": [ + { + "bbox": [ + 138, + 171, + 225, + 182 + ], + "type": "text", + "content": "(a) our proposed design" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 311, + 82, + 540, + 164 + ], + "blocks": [ + { + "bbox": [ + 311, + 82, + 540, + 164 + ], + "lines": [ + { + "bbox": [ + 311, + 82, + 540, + 164 + ], + "spans": [ + { + "bbox": [ + 311, + 82, + 540, + 164 + ], + "type": "image", + "image_path": "be280e1db447da6cdc52ac19ab235d8ee252aacb77c2d6e70818603c0f61f32d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 370, + 170, + 490, + 180 + ], + "lines": [ + { + "bbox": [ + 370, + 170, + 490, + 180 + ], + "spans": [ + { + "bbox": [ + 370, + 170, + 490, + 180 + ], + "type": "text", + "content": "(b) design of LLaVA-v1.5 (CVPR'24)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 70, + 200, + 296, + 281 + ], + "blocks": [ + { + "bbox": [ + 70, + 200, + 296, + 281 + ], + "lines": [ + { + "bbox": [ + 70, + 200, + 296, + 281 + ], + "spans": [ + { + "bbox": [ + 70, + 200, + 296, + 281 + ], + "type": "image", + "image_path": "529895ee21b4cbcfe8f8bc1b75b4b2672e56e3d025a7643e574f089b4329c533.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 126, + 287, + 247, + 298 + ], + "lines": [ + { + "bbox": [ + 126, + 287, + 247, + 298 + ], + "spans": [ + { + "bbox": [ + 126, + 287, + 247, + 298 + ], + "type": "text", + "content": "(c) design of SpatialVLM (CVPR'24)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 319, + 201, + 543, + 277 + ], + "blocks": [ + { + "bbox": [ + 319, + 201, + 543, + 277 + ], + "lines": [ + { + "bbox": [ + 319, + 201, + 543, + 277 + ], + "spans": [ + { + "bbox": [ + 319, + 201, + 543, + 277 + ], + "type": "image", + "image_path": "d4388b1caeed971136a941795b4946bc2bf24d89a0a4915e4510e3086c738429.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 397, + 287, + 447, + 297 + ], + "lines": [ + { + "bbox": [ + 397, + 287, + 447, + 297 + ], + "spans": [ + { + "bbox": [ + 397, + 287, + 447, + 297 + ], + "type": "text", + "content": "(d) data cards" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 388, + 295, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 388, + 295, + 506 + ], + "spans": [ + { + "bbox": [ + 54, + 388, + 295, + 506 + ], + "type": "text", + "content": "without any new spatial-centric design, meaning that the 3D-awareness is completely missing in LLaVA (Fig. 5 (b)). As in Fig. 5 (c), SpatialVLM improves LLaVA by collecting 3D-informed data that is applied to instruction tuning for LLM, boosting the spatial reasoning capability. Compared to SpatialVLM, our design (in Fig. 5 (a)) inject the 3D-awareness by introduce 3D-informed alignment with ImageNet3D data. Additionally, our 3D-informed intrusion tuning data contain extra 3D orientation data that is missing in spatialVLM." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 507, + 296, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 507, + 296, + 567 + ], + "spans": [ + { + "bbox": [ + 55, + 507, + 296, + 567 + ], + "type": "text", + "content": "Design roadmap. We present the roadmap going from a standard LLaVA to our final model in Fig. 6. We start with LLaVA-v1.5 which only achieves a " + }, + { + "bbox": [ + 55, + 507, + 296, + 567 + ], + "type": "inline_equation", + "content": "47.7\\%" + }, + { + "bbox": [ + 55, + 507, + 296, + 567 + ], + "type": "text", + "content": " accuracy on our SpatialVQA. This will be our baseline. We then study a series of design decisions." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 55, + 570, + 295, + 630 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 295, + 630 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 295, + 630 + ], + "type": "text", + "content": "- + mixed vision encoder. To build stronger visual encoder, we mix the original CLIP vision encoder [48] with the DINOv2 [47] vision encoder to extract the hybrid feature maps, lifting a " + }, + { + "bbox": [ + 55, + 570, + 295, + 630 + ], + "type": "inline_equation", + "content": "0.3\\%" + }, + { + "bbox": [ + 55, + 570, + 295, + 630 + ], + "type": "text", + "content": " performance gain. This show that the self-supervised feature can enhance spatial reasoning." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 631, + 295, + 701 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 631, + 295, + 701 + ], + "spans": [ + { + "bbox": [ + 55, + 631, + 295, + 701 + ], + "type": "text", + "content": "- Vicuna1.5 " + }, + { + "bbox": [ + 55, + 631, + 295, + 701 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 55, + 631, + 295, + 701 + ], + "type": "text", + "content": " Llama3. We switch our large language model from Vicuna-v1.5-7B [68] to Llama3-8B [18], as the latter achieves " + }, + { + "bbox": [ + 55, + 631, + 295, + 701 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 55, + 631, + 295, + 701 + ], + "type": "text", + "content": " higher accuracy on MMLU. This change results in " + }, + { + "bbox": [ + 55, + 631, + 295, + 701 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 55, + 631, + 295, + 701 + ], + "type": "text", + "content": " gain on spatial reasoning. We analyze that improved language reasoning may enhance spatial reasoning as well." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 701, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 701, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 701, + 295, + 714 + ], + "type": "text", + "content": "- +3DI-Ft1M data for stage 2 instruction tuning. We col" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "text" + }, + { + "type": "image", + "bbox": [ + 316, + 385, + 548, + 533 + ], + "blocks": [ + { + "bbox": [ + 54, + 311, + 555, + 366 + ], + "lines": [ + { + "bbox": [ + 54, + 311, + 555, + 366 + ], + "spans": [ + { + "bbox": [ + 54, + 311, + 555, + 366 + ], + "type": "text", + "content": "Figure 5. Design instantiation and comparison. (a) Architecture and Training of our proposed design. We investigate the 3D-awareness of mixed visual encoders, and incorporate 3D-informed data at each training stage across all architecture components. (b) LLaVA comparison, where only traditional data are used. (c) SpatialVLM comparison, where 3D-informed data are used in the final stage. (d) Traditional data vs. 3D-informed data. Note SpatialVLM lacks of 3D orientation data. Architectures trained with traditional data and 3D-informed data are marked with blue and orange backgrounds, respectively." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 316, + 385, + 548, + 533 + ], + "lines": [ + { + "bbox": [ + 316, + 385, + 548, + 533 + ], + "spans": [ + { + "bbox": [ + 316, + 385, + 548, + 533 + ], + "type": "image", + "image_path": "42d503fcb94cf182feed8978f02582e39a84d445a516eb8bc4e6177b68a8594c.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 544, + 555, + 599 + ], + "lines": [ + { + "bbox": [ + 313, + 544, + 555, + 599 + ], + "spans": [ + { + "bbox": [ + 313, + 544, + 555, + 599 + ], + "type": "text", + "content": "Figure 6. We modernize a standard LLaVA-v1.5 towards the design of a 3D-informed LMM. The bars are the answer accuracies on the SpatialVQA benchmark, indicating the spatial reasoning capability of each model; results for the GPT-4o is shown with the purple bar." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 321, + 630, + 553, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 630, + 553, + 688 + ], + "spans": [ + { + "bbox": [ + 321, + 630, + 553, + 688 + ], + "type": "text", + "content": "lect the new spatial-aware training set and use it in the instruction tuning stage. Impressively, this 3D-informed instruction tuning results in a " + }, + { + "bbox": [ + 321, + 630, + 553, + 688 + ], + "type": "inline_equation", + "content": "10.7\\%" + }, + { + "bbox": [ + 321, + 630, + 553, + 688 + ], + "type": "text", + "content": " performance improvement, highlighting the critical importance of 3D-informed data." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 314, + 689, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 689, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 314, + 689, + 555, + 714 + ], + "type": "text", + "content": "- + 3DI-Pb-1N166K data for stage 1 multimodal alignment. We propose leveraging object-level 3D-informed" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17254" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 63, + 72, + 294, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 72, + 294, + 144 + ], + "spans": [ + { + "bbox": [ + 63, + 72, + 294, + 144 + ], + "type": "text", + "content": "annotation (e.g., 3D location and 3D orientation of objects) to align visual features with 3D awareness. 3DI-Pb-1N166K are converted from human-annotated ImageNet3D [45]. This experiment results in an additional " + }, + { + "bbox": [ + 63, + 72, + 294, + 144 + ], + "type": "inline_equation", + "content": "3\\%" + }, + { + "bbox": [ + 63, + 72, + 294, + 144 + ], + "type": "text", + "content": " performance improvement, underscoring the importance of 3D-informed feature alignment." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 146, + 295, + 182 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 146, + 295, + 182 + ], + "spans": [ + { + "bbox": [ + 55, + 146, + 295, + 182 + ], + "type": "text", + "content": "In a nutshell, the performance gain can be summarized as 1) better architecture and 2) 3D-informed data. This roadmap leads to our SpatialLLM." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 192, + 136, + 205 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 192, + 136, + 205 + ], + "spans": [ + { + "bbox": [ + 55, + 192, + 136, + 205 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 212, + 295, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 212, + 295, + 236 + ], + "spans": [ + { + "bbox": [ + 55, + 212, + 295, + 236 + ], + "type": "text", + "content": "In this section we start by outlining the experimental setup in Sec. 4.1, followed by presenting the results in Sec. 4.2." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 243, + 170, + 255 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 243, + 170, + 255 + ], + "spans": [ + { + "bbox": [ + 55, + 243, + 170, + 255 + ], + "type": "text", + "content": "4.1. Experimental setup" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 260, + 295, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 260, + 295, + 392 + ], + "spans": [ + { + "bbox": [ + 55, + 260, + 295, + 392 + ], + "type": "text", + "content": "We prepare our training data as described in Fig. 4, with 1M paired image and 3D-informed text for pre-pretraining and alignment, and another 1M multi-turn 3D-informed conversation for instruction tuning. Data distributions are supplemented in appendix. Our training setup is built upon LLaVA-v1.5 [39] and all hyperparameters remain unchanged unless explicitly stated otherwise. We briefly overview the training and data setup for the ablated design choices in Sec. 3.3.1. All experiments are conducted in 8 Nvidia A100 GPUs. We evaluate the spatial reasoning capability of LMMs on SpatialVQA as introduced in Sec. 3.2.2." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 399, + 113, + 410 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 399, + 113, + 410 + ], + "spans": [ + { + "bbox": [ + 55, + 399, + 113, + 410 + ], + "type": "text", + "content": "4.2. Results" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 415, + 295, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 415, + 295, + 464 + ], + "spans": [ + { + "bbox": [ + 55, + 415, + 295, + 464 + ], + "type": "text", + "content": "Comparison with the state-of-the-art. Table 1 presents the performance of our model compared to state-of-the-art proprietary systems such as GPT-4o and Claude 3.5 Sonnet, along with top open-source models. Remark" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 57, + 473, + 293, + 647 + ], + "blocks": [ + { + "bbox": [ + 57, + 473, + 293, + 647 + ], + "lines": [ + { + "bbox": [ + 57, + 473, + 293, + 647 + ], + "spans": [ + { + "bbox": [ + 57, + 473, + 293, + 647 + ], + "type": "table", + "html": "
ModelAvg.3D Dist.3D Orient.3D Spatial Rel.
Proprietary
GPT-4o [1]54.059.452.450.8
GPT-4o Mini [1]45.446.939.348.5
Claude 3.5 Sonnet [4]49.350.844.551.5
Open-sourced
LLaVA-v1.5-7B [39]47.750.646.446.2
LLaVA-NeXT-8B [40]50.663.549.241.5
Cambrian-8B [58]51.759.846.448.9
Spatial-centric
SpatialVLM-13B [14]52.264.348.545.8
Ours62.786.352.950.8
", + "image_path": "3fea4ccd4fb51bff2437adc68f8de5949693190db575db53503f5c9cee18cfc0.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 654, + 295, + 677 + ], + "lines": [ + { + "bbox": [ + 55, + 654, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 295, + 677 + ], + "type": "text", + "content": "Table 1. Comparison with the state-of-the-arts including proprietary and open source models." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "content": "ably, our model achieves a performance of " + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "inline_equation", + "content": "62.7\\%" + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "content": ", outperforming the top proprietary model by " + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "inline_equation", + "content": "8.7\\%" + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "content": " and the best" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 72, + 553, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 300 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 300 + ], + "type": "text", + "content": "open-source model by " + }, + { + "bbox": [ + 313, + 72, + 553, + 300 + ], + "type": "inline_equation", + "content": "10.5\\%" + }, + { + "bbox": [ + 313, + 72, + 553, + 300 + ], + "type": "text", + "content": ". Besides, our model excels in all aspects of spatial reasoning involving 3D distance, 3D orientation and 3D spatial relationship. Moreover, our model not only enhances spatial reasoning performance, but still preserves general visual reasoning (detailed in appendix). Interestingly, although SpatialVLM [14] (implemented in SpaceLLaVA [2]) outperforms other open-source models in overall performance, it falls short in 3D orientation reasoning compared to LLaVA, due to the lack of tuning with 3D orientation-specific data. We exclude SpatialRGPT from this comparison as it relies on additional inputs such as object bounding box coordinates, segmentation masks, and depth maps. In contrast, our SpatialLLM, along with LLaVA and Cambrian, focuses on learning spatial awareness directly from raw images. We will consider models with additional inputs in future work. In sum, our model achieves state-of-the-art performance on spatial reasoning, surpassing API models [1, 4] and also spatial-centric VLM [14]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 301, + 553, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 301, + 553, + 350 + ], + "spans": [ + { + "bbox": [ + 313, + 301, + 553, + 350 + ], + "type": "text", + "content": "Thorough empirical exploration of the design space. We comprehensively explore the design space of our approach as shown in Table. 2. We analyze the results and elaborate the findings detailed below." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 351, + 553, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 351, + 553, + 434 + ], + "spans": [ + { + "bbox": [ + 313, + 351, + 553, + 434 + ], + "type": "text", + "content": "Upgrading architectures enable better spatial reasoning. In terms of architecture, integrating a mixed vision encoder can improve overall performance especially for the 3D orientation. Then, upgrading the LLM to the more advanced LLama3-8B enhances spatial reasoning performance by " + }, + { + "bbox": [ + 313, + 351, + 553, + 434 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 313, + 351, + 553, + 434 + ], + "type": "text", + "content": ", leveraging its improved language reasoning capabilities." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 436, + 554, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 436, + 554, + 592 + ], + "spans": [ + { + "bbox": [ + 313, + 436, + 554, + 592 + ], + "type": "text", + "content": "3D-informed alignment and tuning matter a lot. In terms of 3D-informed data and training, we find that 3D-informed instruction tuning with our proposed 3DI-Ft1M dataset yields a substantial performance boost of " + }, + { + "bbox": [ + 313, + 436, + 554, + 592 + ], + "type": "inline_equation", + "content": "+10.7\\%" + }, + { + "bbox": [ + 313, + 436, + 554, + 592 + ], + "type": "text", + "content": ". To further refine the framework, we explore three data types for stage 1 multimodal alignment: curated 3DI-Ft1M, 3DI-Pb-OI1M, and 3DI-Pb-IN166K. The model trained with 3DI-Pb-IN166K performs best. We attribute this to the use of human-annotated ImageNet3D data, as opposed to the semi-automated methods employed for 3DI-Pb-OI1M. The human-annotated ImageNet3D provides more accurate 3D orientation of objects. This highlights the importance of high-quality data in the alignment stage." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 594, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 554, + 714 + ], + "type": "text", + "content": "3D-informed pretraining of vision encoder? The pretrained CLIP vision encoder, originally designed for image-text contrastive learning, lacks 3D-awareness [19], We hypotheszie that the vision encoder has a chance to benefit from 3D-informed data. To verify this, we conduct 3D-informed pretraining for the vision encoder using curated 3DI-Pb-OI1M probing data. We explore both CLIP [25] and next token prediction (NTP) [41] objectives to fine-tune Lora layers [24] of the ViT. We observe that pre-pretraining in stage 0 reduces performance, suggesting that tuning vi" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17255" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 70, + 553, + 286 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 553, + 286 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 553, + 286 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 553, + 286 + ], + "type": "table", + "html": "
Model\nVision EncoderLLMStage 0\nEncoder pretrainingStage 1\nAlignmentStage 2\nInstruction TuningSpatialVQA
Avg.Dist.Orient.Spatial Rel.
Baseline (LLaVA-v1.5)
CLIP [48]Vicuna-7B [68]-CC558K [39]Mix665K [39]47.750.646.446.2
Mixed Encoders
CLIP [48]+MAE [23]Vicuna-7B [68]-CC558KMix665K47.952.245.746.1
CLIP [48]+DINOv2 [47]Vicuna-7B [68]-CC558KMix665K48.051.946.945.6
Advanced LLM
CLIP [48]+DINOv2 [47]Llama3-8B [18]-CC558K [39]Mix665K [39]49.056.351.741.5
3D-informed Tuning for Stage 2
CLIP [48]+DINOv2 [47]Llama3-8B [18]-CC558K [39]Mix665K [39] + 3DI-Ft1M59.785.252.144.8
3D-informed Alignment for Stage 1
CLIP [48]+DINOv2 [47]Llama3-8B [18]-CC558K [39] + 3DI-Ft1MMix665K [39] + 3DI-Ft1M57.580.846.946.4
CLIP [48]+DINOv2 [47]Llama3-8B [18]-CC558K [39] + 3DI-Pb-OI1MMix665K [39] + 3DI-Ft1M60.082.146.951.6
CLIP [48]+DINOv2 [47]Llama3-8B [18]-CC558K [39] + 3DI-Pb-IN166KMix665K [39] + 3DI-Ft1M62.786.352.951.0
3D-informed pretraining for vision encoder in stage 0
CLIP [48]+DINOv2 [47]Llama3-8B [18]3DI-Pb-OI1M (CLIP [25])CC558K [39] + 3DI-Pb-IN166KMix665K [39] + 3DI-Ft1M60.385.050.547.5
CLIP [48]+DINOv2 [47]Llama3-8B [18]3DI-Pb-OI1M (NTP [34])CC558K [39] + 3DI-Pb-IN166KMix665K [39] + 3DI-Ft1M62.685.252.951.5
", + "image_path": "d8bd78a4ed73c9f2459e29ba927f8d7b04fe2282d1710efcd32f2fa972ac5a39.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 295, + 555, + 340 + ], + "lines": [ + { + "bbox": [ + 55, + 295, + 555, + 340 + ], + "spans": [ + { + "bbox": [ + 55, + 295, + 555, + 340 + ], + "type": "text", + "content": "Table 2. Thorough exploration of the design space and roadmap progression. We systematically examine the 3D-informed design space from the aspects of data, architecture and training. The non-gray rows, listed from top to bottom, illustrate the step-by-step progression of our design roadmap. Gray rows correspond to ablations across the full design space. Our final model is highlighted with a light orange background. Our final model's superior performance over other variants confirms the necessity of a 3D-informed compound design." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 361, + 295, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 361, + 295, + 397 + ], + "spans": [ + { + "bbox": [ + 55, + 361, + 295, + 397 + ], + "type": "text", + "content": "sual foundation models may compromise feature generalizability. This aligns with prior works advocating for frozen or locked visual encoders [36, 38]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 397, + 296, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 397, + 296, + 517 + ], + "spans": [ + { + "bbox": [ + 55, + 397, + 296, + 517 + ], + "type": "text", + "content": "Roadmap progression towards the best-performing model. After a thorough empirical exploration of the design space, we can devise our roadmap progressed towards the best-performing model. The non-gray rows in Table. 2 align with the roadmap illustrated in Fig. 6, highlighting our systematic approach to upgrade the architecture and incorporate 3D-informed data into the training process. Overall, the improved architecture achieves a " + }, + { + "bbox": [ + 55, + 397, + 296, + 517 + ], + "type": "inline_equation", + "content": "1.3\\%" + }, + { + "bbox": [ + 55, + 397, + 296, + 517 + ], + "type": "text", + "content": " performance gain, while 3D-informed alignment and instruction tuning further enhance performance by " + }, + { + "bbox": [ + 55, + 397, + 296, + 517 + ], + "type": "inline_equation", + "content": "13.7\\%" + }, + { + "bbox": [ + 55, + 397, + 296, + 517 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 517, + 296, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 517, + 296, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 517, + 296, + 601 + ], + "type": "text", + "content": "Qualitative comparison. Fig. 7 illustrates a scenario involving spatial reasoning. GPT-4o struggles to reason about the spatial relationships between objects due to its limited ability to comprehend their 3D orientation. In contrast, our model successfully understands these spatial relationships and provides accurate answers, demonstrating superior 3D spatial reasoning capability." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 609, + 134, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 609, + 134, + 621 + ], + "spans": [ + { + "bbox": [ + 55, + 609, + 134, + 621 + ], + "type": "text", + "content": "5. Conclusions" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 629, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 296, + 714 + ], + "type": "text", + "content": "In this work, we systematically enhance the 3D spatial reasoning capabilities of LMMs through a compound design approach. By introducing SpatialLLM, we demonstrated that integrating 3D-informed data, architectural innovations, and tailored training setups significantly improves an LMM's ability to understand and reason about complex 3D spatial relationships. To tackle data scarcity, we develop" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 317, + 361, + 467, + 410 + ], + "blocks": [ + { + "bbox": [ + 317, + 361, + 467, + 410 + ], + "lines": [ + { + "bbox": [ + 317, + 361, + 467, + 410 + ], + "spans": [ + { + "bbox": [ + 317, + 361, + 467, + 410 + ], + "type": "image", + "image_path": "d2fedb8e633046d43ee48872746547787fde33b0cd8a454c0f32921049fa5c04.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 321, + 420, + 335, + 436 + ], + "blocks": [ + { + "bbox": [ + 323, + 414, + 335, + 419 + ], + "lines": [ + { + "bbox": [ + 323, + 414, + 335, + 419 + ], + "spans": [ + { + "bbox": [ + 323, + 414, + 335, + 419 + ], + "type": "text", + "content": "User:" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 321, + 420, + 335, + 436 + ], + "lines": [ + { + "bbox": [ + 321, + 420, + 335, + 436 + ], + "spans": [ + { + "bbox": [ + 321, + 420, + 335, + 436 + ], + "type": "image", + "image_path": "ef9d4b3f8241ae4a604de36f0ab946aea5993a83fadeadb013519bb1842a82c6.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 341, + 417, + 465, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 341, + 417, + 465, + 437 + ], + "spans": [ + { + "bbox": [ + 341, + 417, + 465, + 437 + ], + "type": "text", + "content": "Consider the real-world 3D locations and orientations of the objects. Which object is closer to the car in Region [0], the car in Region [1] or the car in Region [2]?" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 318, + 453, + 336, + 473 + ], + "blocks": [ + { + "bbox": [ + 318, + 453, + 336, + 473 + ], + "lines": [ + { + "bbox": [ + 318, + 453, + 336, + 473 + ], + "spans": [ + { + "bbox": [ + 318, + 453, + 336, + 473 + ], + "type": "image", + "image_path": "b8790ef8c43ee80660eff58393f58e2c76fa85958ecd1fc9f6a9df105d700aa9.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 511, + 555, + 556 + ], + "lines": [ + { + "bbox": [ + 313, + 511, + 555, + 556 + ], + "spans": [ + { + "bbox": [ + 313, + 511, + 555, + 556 + ], + "type": "text", + "content": "Figure 7. Our model is capable of answering question correctly that needs accurate reasoning on spatial distance and 3D orientation, while GPT-4o either needs more spatial information or gives wrong answer." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 341, + 442, + 465, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 341, + 442, + 465, + 486 + ], + "spans": [ + { + "bbox": [ + 341, + 442, + 465, + 486 + ], + "type": "text", + "content": "To determine which car is closer to the car in Region [0], we would need to analyze the spatial arrangement of the objects in the image. However, without the ability to visually assess the image, I can't provide a definitive answer. If you can describe the positions of the cars in Regions [1] and [2] relative to Region [0], I can help you deduce which one is closer." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 380, + 490, + 427, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 380, + 490, + 427, + 498 + ], + "spans": [ + { + "bbox": [ + 380, + 490, + 427, + 498 + ], + "type": "text", + "content": "The car in Region [1]." + } + ] + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 470, + 361, + 551, + 410 + ], + "blocks": [ + { + "bbox": [ + 470, + 361, + 551, + 410 + ], + "lines": [ + { + "bbox": [ + 470, + 361, + 551, + 410 + ], + "spans": [ + { + "bbox": [ + 470, + 361, + 551, + 410 + ], + "type": "image", + "image_path": "3d08707deb9c3b6aaac7cc8c8293544380e97eafa903793ea4d1ef455e79988f.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 472, + 414, + 550, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 472, + 414, + 550, + 452 + ], + "spans": [ + { + "bbox": [ + 472, + 414, + 550, + 452 + ], + "type": "text", + "content": "Consider the real-world 3D orientations of the objects. Are the chair in Region [0] and the chair in Region [1] facing the same (or similar) directions, or very different directions?" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 472, + 460, + 548, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 472, + 460, + 548, + 480 + ], + "spans": [ + { + "bbox": [ + 472, + 460, + 548, + 480 + ], + "type": "text", + "content": "Based on the image, the chair in Region [0] and the chair in Region [1] face similar directions" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 482, + 490, + 537, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 482, + 490, + 537, + 496 + ], + "spans": [ + { + "bbox": [ + 482, + 490, + 537, + 496 + ], + "type": "text", + "content": "Very different directions." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 568, + 555, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 568, + 555, + 700 + ], + "spans": [ + { + "bbox": [ + 313, + 568, + 555, + 700 + ], + "type": "text", + "content": "3D-informed training datasets focused on 3D locations, orientations and complex spatial relationship. Our systematic integration of these datasets with architectural and training designs provide a roadmap for optimizing LMMs for superior 3D reasoning capabilities. Experimental results on the SpatialVQA benchmark highlight SpatialLLM's superiority over state-of-the-art models, setting a new standard for 3D spatial reasoning. We invite the community to build upon our design roadmap and findings to push the boundaries of what LMMs can achieve in understanding the complex 3D spatial relationships of the real world." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "17256" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 170, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 170, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 170, + 85 + ], + "type": "text", + "content": "6. Acknowledgements" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 295, + 116 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 295, + 116 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 295, + 116 + ], + "type": "text", + "content": "WM and AY acknowledges support from ONR with N00014-23-1-2641 and ARL award W911NF2320008." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 128, + 115, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 128, + 115, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 128, + 115, + 140 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 148, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 62, + 148, + 295, + 202 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 148, + 295, + 202 + ], + "spans": [ + { + "bbox": [ + 62, + 148, + 295, + 202 + ], + "type": "text", + "content": "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 1, 2, 3, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 205, + 293, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 205, + 293, + 226 + ], + "spans": [ + { + "bbox": [ + 61, + 205, + 293, + 226 + ], + "type": "text", + "content": "[2] Remyx AI. SpaceLLaVA. https://huggingface.co/remyxai/SpaceLLaVA.1,7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 228, + 295, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 228, + 295, + 293 + ], + "spans": [ + { + "bbox": [ + 62, + 228, + 295, + 293 + ], + "type": "text", + "content": "[3] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in neural information processing systems, 35:23716-23736, 2022. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 296, + 294, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 296, + 294, + 327 + ], + "spans": [ + { + "bbox": [ + 62, + 296, + 294, + 327 + ], + "type": "text", + "content": "[4] Anthropic. Claude 3.5 Sonnet. https://www.anthropic.com/news/claude-3-5-sonnet. 1, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 330, + 247, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 330, + 247, + 341 + ], + "spans": [ + { + "bbox": [ + 62, + 330, + 247, + 341 + ], + "type": "text", + "content": "[5] Apollo Team. Apollo synthetic dataset, 2019. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 342, + 295, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 342, + 295, + 396 + ], + "spans": [ + { + "bbox": [ + 62, + 342, + 295, + 396 + ], + "type": "text", + "content": "[6] Daichi Azuma, Taiki Miyanishi, Shuhei Kurita, and Motoaki Kawanabe. Scanqa: 3d question answering for spatial scene understanding. In proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 19129-19139, 2022. 4, 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 399, + 295, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 399, + 295, + 442 + ], + "spans": [ + { + "bbox": [ + 62, + 399, + 295, + 442 + ], + "type": "text", + "content": "[7] Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, et al. Qwen technical report. arXiv preprint arXiv:2309.16609, 2023. 2, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 445, + 295, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 445, + 295, + 487 + ], + "spans": [ + { + "bbox": [ + 62, + 445, + 295, + 487 + ], + "type": "text", + "content": "[8] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 490, + 295, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 490, + 295, + 554 + ], + "spans": [ + { + "bbox": [ + 62, + 490, + 295, + 554 + ], + "type": "text", + "content": "[9] Gilad Baruch, Zhuoyuan Chen, Afshin Dehghan, Tal Dimry, Yuri Feigin, Peter Fu, Thomas Gebauer, Brandon Joffe, Daniel Kurz, Arik Schwartz, et al. Arkitsscenes: A diverse real-world dataset for 3d indoor scene understanding using mobile rgb-d data. arXiv preprint arXiv:2111.08897, 2021. 4, 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 557, + 295, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 557, + 295, + 600 + ], + "spans": [ + { + "bbox": [ + 57, + 557, + 295, + 600 + ], + "type": "text", + "content": "[10] Shariq Farooq Bhat, Reiner Birkl, Diana Wofk, Peter Wonka, and Matthias Müller. Zoedepth: Zero-shot transfer by combining relative and metric depth. arXiv preprint arXiv:2302.12288, 2023. 1, 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 602, + 295, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 602, + 295, + 645 + ], + "spans": [ + { + "bbox": [ + 57, + 602, + 295, + 645 + ], + "type": "text", + "content": "[11] Garrick Brazil, Abhinav Kumar, Julian Straub, Nikhila Ravi, Justin Johnson, and Georgia Gkioxari. Omni3D: A large benchmark and model for 3D object detection in the wild. In CVPR, Vancouver, Canada, 2023. IEEE. 2, 4, 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 647, + 296, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 647, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 647, + 296, + 712 + ], + "type": "text", + "content": "[12] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11621-11631, 2020. 4, 1" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 554, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 316, + 73, + 554, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 127 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 127 + ], + "type": "text", + "content": "[13] Junbum Cha, Wooyoung Kang, Jonghwan Mun, and Byungseok Roh. Honeybee: Locality-enhanced projector for multimodal llm. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13817-13827, 2024. 2, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 130, + 554, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 130, + 554, + 194 + ], + "spans": [ + { + "bbox": [ + 317, + 130, + 554, + 194 + ], + "type": "text", + "content": "[14] Boyuan Chen, Zhuo Xu, Sean Kirmani, Brain Ichter, Dorsa Sadigh, Leonidas Guibas, and Fei Xia. Spatialvlm: Endowing vision-language models with spatial reasoning capabilities. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14455-14465, 2024. 1, 2, 3, 4, 5, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 197, + 553, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 197, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 316, + 197, + 553, + 251 + ], + "type": "text", + "content": "[15] Jieneng Chen, Qihang Yu, Xiaohui Shen, Alan Yuille, and Liang-Chieh Chen. Vitamin: Designing scalable vision models in the vision-language era. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 253, + 553, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 253, + 553, + 296 + ], + "spans": [ + { + "bbox": [ + 317, + 253, + 553, + 296 + ], + "type": "text", + "content": "[16] An-Chieh Cheng, Hongxu Yin, Yang Fu, Qiushan Guo, Ruihan Yang, Jan Kautz, Xiaolong Wang, and Sifei Liu. Spatialrgpt: Grounded spatial reasoning in vision language model. arXiv preprint arXiv:2406.01584, 2024. 1, 2, 3, 4, 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 298, + 554, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 298, + 554, + 342 + ], + "spans": [ + { + "bbox": [ + 317, + 298, + 554, + 342 + ], + "type": "text", + "content": "[17] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 2, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 344, + 554, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 344, + 554, + 396 + ], + "spans": [ + { + "bbox": [ + 317, + 344, + 554, + 396 + ], + "type": "text", + "content": "[18] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 3, 6, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 399, + 554, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 399, + 554, + 474 + ], + "spans": [ + { + "bbox": [ + 317, + 399, + 554, + 474 + ], + "type": "text", + "content": "[19] Mohamed El Banani, Amit Raj, Kevis-Kokitsi Maninis, Abhishek Kar, Yuanzhen Li, Michael Rubinstein, Deqing Sun, Leonidas Guibas, Justin Johnson, and Varun Jampani. Probing the 3d awareness of visual foundation models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21795-21806, 2024. 4, 5, 7, 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 478, + 553, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 478, + 553, + 531 + ], + "spans": [ + { + "bbox": [ + 317, + 478, + 553, + 531 + ], + "type": "text", + "content": "[20] Samir Yitzhak Gadre, Gabriel Ilharco, Alex Fang, Jonathan Hayase, Georgios Smyrnis, Thao Nguyen, Ryan Marten, Mitchell Wortsman, Dhruba Ghosh, Jieyu Zhang, et al. Dat-acomp: In search of the next generation of multimodal datasets. arXiv preprint arXiv:2304.14108, 2023. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 534, + 553, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 534, + 553, + 578 + ], + "spans": [ + { + "bbox": [ + 317, + 534, + 553, + 578 + ], + "type": "text", + "content": "[21] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In 2012 IEEE conference on computer vision and pattern recognition, pages 3354-3361. IEEE, 2012. 4, 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 579, + 553, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 579, + 553, + 633 + ], + "spans": [ + { + "bbox": [ + 317, + 579, + 553, + 633 + ], + "type": "text", + "content": "[22] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Batra, and Devi Parikh. Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6904-6913, 2017. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 635, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 635, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 317, + 635, + 553, + 689 + ], + "type": "text", + "content": "[23] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 3, 5, 8" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "text", + "content": "[24] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen." + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17257" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 76, + 72, + 295, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 295, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 295, + 95 + ], + "type": "text", + "content": "Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 162 + ], + "type": "text", + "content": "[25] Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. Openclip, 2021. If you use this software, please cite it as below. 7, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 294, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 294, + 217 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 294, + 217 + ], + "type": "text", + "content": "[26] Artur Jesslen, Guofeng Zhang, Angtian Wang, Wufei Ma, Alan Yuille, and Adam Kortylewski. Novum: Neural object volumes for robust object classification. In European Conference on Computer Vision, pages 264-281. Springer, 2024. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 220, + 294, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 220, + 294, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 220, + 294, + 262 + ], + "type": "text", + "content": "[27] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In ICML, 2021. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 265, + 294, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 265, + 294, + 319 + ], + "spans": [ + { + "bbox": [ + 56, + 265, + 294, + 319 + ], + "type": "text", + "content": "[28] Albert Q Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, et al. Mistral 7b. arXiv preprint arXiv:2310.06825, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 321, + 294, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 321, + 294, + 386 + ], + "spans": [ + { + "bbox": [ + 56, + 321, + 294, + 386 + ], + "type": "text", + "content": "[29] Linyi Jin, Jianming Zhang, Yannick Hold-Geoffroy, Oliver Wang, Kevin Blackburn-Matzen, Matthew Sticha, and David F Fouhey. Perspective fields for single image camera calibration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17307-17316, 2023. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 388, + 294, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 388, + 294, + 441 + ], + "spans": [ + { + "bbox": [ + 56, + 388, + 294, + 441 + ], + "type": "text", + "content": "[30] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dollar, and Ross Girshick. Segment anything. arXiv:2304.02643, 2023. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 445, + 294, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 445, + 294, + 499 + ], + "spans": [ + { + "bbox": [ + 56, + 445, + 294, + 499 + ], + "type": "text", + "content": "[31] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 501, + 294, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 501, + 294, + 565 + ], + "spans": [ + { + "bbox": [ + 56, + 501, + 294, + 565 + ], + "type": "text", + "content": "[32] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision, 123:32-73, 2017. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 567, + 294, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 567, + 294, + 643 + ], + "spans": [ + { + "bbox": [ + 56, + 567, + 294, + 643 + ], + "type": "text", + "content": "[33] Alina Kuznetsova, Hassan Rom, Neil Alldrin, Jasper Uijlings, Ivan Krasin, Jordi Pont-Tuset, Shahab Kamali, Stefan Popov, Matteo Malloci, Alexander Kolesnikov, et al. The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale. International journal of computer vision, 128(7):1956-1981, 2020. 5, 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 646, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 646, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 646, + 294, + 689 + ], + "type": "text", + "content": "[34] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 2, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "text", + "content": "[35] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "type": "text", + "content": "vision-language understanding and generation. In ICML, 2022. 2, 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 96, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 150 + ], + "type": "text", + "content": "[36] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730–19742. PMLR, 2023. 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 152, + 553, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 553, + 206 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 553, + 206 + ], + "type": "text", + "content": "[37] Xianhang Li, Haoqin Tu, Mude Hui, Zeyu Wang, Bingchen Zhao, Junfei Xiao, Sucheng Ren, Jieru Mei, Qing Liu, Huangjie Zheng, et al. What if we recapture billions of web images with llama-3? arXiv preprint arXiv:2406.08478, 2024. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 209, + 553, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 209, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 316, + 209, + 553, + 262 + ], + "type": "text", + "content": "[38] Haotian Liu, Kilho Son, Jianwei Yang, Ce Liu, Jianfeng Gao, Yong Jae Lee, and Chunyuan Li. Learning customized visual models with retrieval-augmented knowledge. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15148-15158, 2023. 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 264, + 553, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 264, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 316, + 264, + 553, + 319 + ], + "type": "text", + "content": "[39] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26296-26306, 2024. 1, 2, 3, 7, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 321, + 553, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 321, + 553, + 354 + ], + "spans": [ + { + "bbox": [ + 316, + 321, + 553, + 354 + ], + "type": "text", + "content": "[40] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava next: Improved reasoning,ocr,and world knowledge,2024.7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 355, + 553, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 355, + 553, + 388 + ], + "spans": [ + { + "bbox": [ + 316, + 355, + 553, + 388 + ], + "type": "text", + "content": "[41] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 1, 2, 3, 4, 5, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 389, + 553, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 389, + 553, + 443 + ], + "spans": [ + { + "bbox": [ + 316, + 389, + 553, + 443 + ], + "type": "text", + "content": "[42] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Qing Jiang, Chunyuan Li, Jianwei Yang, Hang Su, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 445, + 553, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 445, + 553, + 498 + ], + "spans": [ + { + "bbox": [ + 316, + 445, + 553, + 498 + ], + "type": "text", + "content": "[43] Haoyu Lu, Wen Liu, Bo Zhang, Bingxuan Wang, Kai Dong, Bo Liu, Jingxiang Sun, Tongzheng Ren, Zhuoshu Li, Hao Yang, et al. Deepseek-vl: towards real-world vision-language understanding. arXiv preprint arXiv:2403.05525, 2024. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 501, + 553, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 501, + 553, + 555 + ], + "spans": [ + { + "bbox": [ + 316, + 501, + 553, + 555 + ], + "type": "text", + "content": "[44] Wufei Ma, Angtian Wang, Alan Yuille, and Adam Kortylewski. Robust category-level 6d pose estimation with coarse-to-fine rendering of neural features. In European Conference on Computer Vision, pages 492-508. Springer, 2022. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 557, + 553, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 557, + 553, + 601 + ], + "spans": [ + { + "bbox": [ + 316, + 557, + 553, + 601 + ], + "type": "text", + "content": "[45] Wufei Ma, Guanning Zeng, Guofeng Zhang, Qihao Liu, Letian Zhang, Adam Kortylewski, Yaoyao Liu, and Alan Yuille. Imagenet3d: Towards general-purpose object-level 3d understanding. In NeurIPS, 2024. 2, 4, 5, 7, 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 602, + 553, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 602, + 553, + 657 + ], + "spans": [ + { + "bbox": [ + 316, + 602, + 553, + 657 + ], + "type": "text", + "content": "[46] Brandon McKinzie, Zhe Gan, Jean-Philippe Fauconnier, Sam Dodge, Bowen Zhang, Philipp Dufter, Dhruti Shah, Xianzhi Du, Futang Peng, Floris Weers, et al. Mm1: Methods, analysis & insights from multimodal lIm pre-training. arXiv preprint arXiv:2403.09611, 2024. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 658, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 553, + 712 + ], + "type": "text", + "content": "[47] Maxime Oquab, Timothee Darcet, Theo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. TMLR, 2024. 2, 3, 5, 6, 8" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "17258" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 297, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 297, + 127 + ], + "type": "text", + "content": "[48] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 2, 3, 5, 6, 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 129, + 296, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 129, + 296, + 163 + ], + "spans": [ + { + "bbox": [ + 55, + 129, + 296, + 163 + ], + "type": "text", + "content": "[49] Navid Rajabi and Jana Kosecka. Gsr-bench: A benchmark for grounded spatial reasoning evaluation via multimodal llms. arXiv preprint arXiv:2406.13246, 2024. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 296, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 296, + 229 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 296, + 229 + ], + "type": "text", + "content": "[50] Mike Roberts, Jason Ramapuram, Anurag Ranjan, Atulit Kumar, Miguel Angel Bautista, Nathan Paczan, Russ Webb, and Joshua M. Susskind. Hypersim: A photorealistic synthetic dataset for holistic indoor scene understanding. In International Conference on Computer Vision (ICCV) 2021, 2021. 4, 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 230, + 296, + 286 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 230, + 296, + 286 + ], + "spans": [ + { + "bbox": [ + 56, + 230, + 296, + 286 + ], + "type": "text", + "content": "[51] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 287, + 296, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 287, + 296, + 353 + ], + "spans": [ + { + "bbox": [ + 56, + 287, + 296, + 353 + ], + "type": "text", + "content": "[52] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278-25294, 2022. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 354, + 296, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 354, + 296, + 397 + ], + "spans": [ + { + "bbox": [ + 56, + 354, + 296, + 397 + ], + "type": "text", + "content": "[53] Piyush Sharma, Nan Ding, Sebastian Goodman, and Radu Soricut. Conceptual captions: A cleaned, hypernymed, image alt-text dataset for automatic image captioning. In ACL, 2018. 2, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 399, + 296, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 399, + 296, + 443 + ], + "spans": [ + { + "bbox": [ + 56, + 399, + 296, + 443 + ], + "type": "text", + "content": "[54] Shuran Song, Samuel P Lichtenberg, and Jianxiong Xiao. Sun rgb-d: A rgb-d scene understanding benchmark suite. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 567-576, 2015. 4, 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 444, + 296, + 467 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 444, + 296, + 467 + ], + "spans": [ + { + "bbox": [ + 56, + 444, + 296, + 467 + ], + "type": "text", + "content": "[55] Elizabeth S Spelke and Katherine D Kinzler. Core knowledge. Developmental science, 10(1):89-96, 2007. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 468, + 296, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 468, + 296, + 512 + ], + "spans": [ + { + "bbox": [ + 56, + 468, + 296, + 512 + ], + "type": "text", + "content": "[56] Chen Sun, Abhinav Shrivastava, Saurabh Singh, and Abhinav Gupta. Revisiting unreasonable effectiveness of data in deep learning era. In Proceedings of the IEEE international conference on computer vision, pages 843-852, 2017. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 514, + 296, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 514, + 296, + 567 + ], + "spans": [ + { + "bbox": [ + 56, + 514, + 296, + 567 + ], + "type": "text", + "content": "[57] Gemini Team, Rohan Anil, Sebastian Borgeaud, Yonghui Wu, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, et al. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805, 2023. 1, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 569, + 296, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 569, + 296, + 624 + ], + "spans": [ + { + "bbox": [ + 56, + 569, + 296, + 624 + ], + "type": "text", + "content": "[58] Shengbang Tong, Ellis Brown, Penghao Wu, Sanghyun Woo, Manoj Middepogu, Sai Charitha Akula, Jihan Yang, Shusheng Yang, Adithya Iyer, Xichen Pan, et al. Cambrian-1: A fully open, vision-centric exploration of multimodal llms. arXiv preprint arXiv:2406.16860, 2024. 1, 3, 4, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 625, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 625, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 625, + 296, + 713 + ], + "type": "text", + "content": "[59] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, Dan Bikel, Lukas Blecher, Cristian Canton Ferrer, Moya Chen, Guillem Cucurull, David Esiobu, Jude Fernandes, Jeremy Fu, Wenyin Fu, Brian Fuller, Cynthia Gao, Vedanuj Goswami, Naman Goyal, Anthony Hartshorn, Saghar Hosseini, Rui Hou, Hakan Inan, Marcin Kardas, Vik" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 73, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 334, + 73, + 553, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 334, + 73, + 553, + 225 + ], + "spans": [ + { + "bbox": [ + 334, + 73, + 553, + 225 + ], + "type": "text", + "content": "tor Kerkez, Madian Khabsa, Isabel Kloumann, Artem Korenev, Punit Singh Koura, Marie-Anne Lachaux, Thibaut Lavril, Jenya Lee, Diana Liskovich, Yinghai Lu, Yuning Mao, Xavier Martinet, Todor Mihaylov, Pushkar Mishra, Igor Molybog, Yixin Nie, Andrew Poulton, Jeremy Reizenstein, Rashi Rungta, Kalyan Saladi, Alan Schelten, Ruan Silva, Eric Michael Smith, Ranjan Subramanian, Xiaoqing Ellen Tan, Binh Tang, Ross Taylor, Adina Williams, Jian Xiang Kuan, Puxin Xu, Zheng Yan, Iliyan Zarov, Yuchen Zhang, Angela Fan, Melanie Kambadur, Sharan Narang, Aurelien Rodriguez, Robert Stojnic, Sergey Edunov, and Thomas Scialom. Llama 2: Open foundation and finetuned chat models. arXiv preprint arXiv:2307.09288, 2023. 2,3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 315, + 228, + 553, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 228, + 553, + 272 + ], + "spans": [ + { + "bbox": [ + 315, + 228, + 553, + 272 + ], + "type": "text", + "content": "[60] Xingrui Wang, Wufei Ma, Zhuowan Li, Adam Kortylewski, and Alan L Yuille. 3d-aware visual question answering about parts, poses and occlusions. Advances in Neural Information Processing Systems, 36, 2024. 2, 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 315, + 274, + 553, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 274, + 553, + 317 + ], + "spans": [ + { + "bbox": [ + 315, + 274, + 553, + 317 + ], + "type": "text", + "content": "[61] Xingrui Wang, Wufei Ma, Angtian Wang, Shuo Chen, Adam Kortylewski, and Alan Yuille. Compositional 4d dynamic scenes understanding with physics priors for video question answering. arXiv preprint arXiv:2406.00622, 2024. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 315, + 318, + 553, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 318, + 553, + 352 + ], + "spans": [ + { + "bbox": [ + 315, + 318, + 553, + 352 + ], + "type": "text", + "content": "[62] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. In CVPR, 2024. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 354, + 555, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 354, + 555, + 408 + ], + "spans": [ + { + "bbox": [ + 316, + 354, + 555, + 408 + ], + "type": "text", + "content": "[63] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10371-10381, 2024. 1, 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 409, + 553, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 409, + 553, + 442 + ], + "spans": [ + { + "bbox": [ + 316, + 409, + 553, + 442 + ], + "type": "text", + "content": "[64] Shuquan Ye, Dongdong Chen, Songfang Han, and Jing Liao. 3d question answering. arXiv preprint arXiv:2112.08359, 2021. 4, 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 444, + 553, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 444, + 553, + 487 + ], + "spans": [ + { + "bbox": [ + 317, + 444, + 553, + 487 + ], + "type": "text", + "content": "[65] Jiahui Yu, Zirui Wang, Vijay Vasudevan, Legg Yeung, Mojtaba Seyedhosseini, and Yonghui Wu. Coca: Contrastive captioners are image-text foundation models. arXiv preprint arXiv:2205.01917, 2022. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 489, + 553, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 489, + 553, + 544 + ], + "spans": [ + { + "bbox": [ + 317, + 489, + 553, + 544 + ], + "type": "text", + "content": "[66] Youcai Zhang, Xinyu Huang, Jinyu Ma, Zhaoyang Li, Zhaochuan Luo, Yanchun Xie, Yuzhuo Qin, Tong Luo, Yaqian Li, Shilong Liu, et al. Recognize anything: A strong image tagging model. arXiv preprint arXiv:2306.03514, 2023.1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 545, + 553, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 545, + 553, + 610 + ], + "spans": [ + { + "bbox": [ + 317, + 545, + 553, + 610 + ], + "type": "text", + "content": "[67] Youcai Zhang, Xinyu Huang, Jinyu Ma, Zhaoyang Li, Zhaochuan Luo, Yanchun Xie, Yuzhuo Qin, Tong Luo, Yaqian Li, Shilong Liu, et al. Recognize anything: A strong image tagging model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1724-1732, 2024. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 613, + 553, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 613, + 553, + 668 + ], + "spans": [ + { + "bbox": [ + 317, + 613, + 553, + 668 + ], + "type": "text", + "content": "[68] Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric Xing, et al. Judging llm-as-a-judge with mt-bench and chatbot arena. Advances in Neural Information Processing Systems, 36:46595-46623, 2023. 6, 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 669, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 669, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 669, + 553, + 712 + ], + "type": "text", + "content": "[69] Jinghao Zhou, Chen Wei, Huiyu Wang, Wei Shen, Cihang Xie, Alan Yuille, and Tao Kong. ibot: Image bert pre-training with online tokenizer. arXiv preprint arXiv:2111.07832, 2021.3" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "17259" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 295, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 295, + 106 + ], + "type": "text", + "content": "[70] Shengjie Zhu, Abhinav Kumar, Masa Hu, and Xiaoming Liu. Tame a wild camera: In-the-wild monocular camera calibration. In NeurIPS, 2023. 1" + } + ] + } + ], + "index": 0 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "17260" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Spatiotemporal Decoupling for Efficient Vision-Based Occupancy Forecasting/9ae564d7-647a-4ca8-a4dc-af9f575b1e21_content_list.json b/2025/Spatiotemporal Decoupling for Efficient Vision-Based Occupancy Forecasting/9ae564d7-647a-4ca8-a4dc-af9f575b1e21_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..399acdf284546774c0db968298755e4c60761005 --- /dev/null +++ b/2025/Spatiotemporal Decoupling for Efficient Vision-Based Occupancy Forecasting/9ae564d7-647a-4ca8-a4dc-af9f575b1e21_content_list.json @@ -0,0 +1,1350 @@ +[ + { + "type": "text", + "text": "Spatiotemporal Decoupling for Efficient Vision-Based Occupancy Forecasting", + "text_level": 1, + "bbox": [ + 93, + 130, + 875, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jingyi Xu $^{1*}$ Xieyuanli Chen $^{2*}$ Junyi Ma $^{3}$ Jiawei Huang $^{4}$ Jintao Xu $^{4}$ Yue Wang $^{5}$ Ling Pei $^{1\\dagger}$ $^{1}$ Shanghai Key Laboratory of Navigation and Location Based Services, Shanghai Jiao Tong University \n $^{2}$ College of Intelligence Science and Technology, National University of Defense Technology \n $^{3}$ Department of Automation, Shanghai Jiao Tong University ${}^{4}$ HAOMO.AI \n $^{5}$ State Key Laboratory of Industrial Control and Technology, Zhejiang University", + "bbox": [ + 86, + 179, + 893, + 271 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 303, + 312, + 319 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The task of occupancy forecasting (OCF) involves utilizing past and present perception data to predict future occupancy states of autonomous vehicle surrounding environments, which is critical for downstream tasks such as obstacle avoidance and path planning. Existing 3D OCF approaches struggle to predict plausible spatial details for movable objects and suffer from slow inference speeds due to neglecting the bias and uneven distribution of changing occupancy states in both space and time. In this paper, we propose a novel spatiotemporal decoupling vision-based paradigm to explicitly tackle the bias and achieve both effective and efficient 3D OCF. To tackle spatial bias in empty areas, we introduce a novel spatial representation that decouples the conventional dense 3D format into 2D bird's-eye view (BEV) occupancy with corresponding height values, enabling 3D OCF derived only from 2D predictions thus enhancing efficiency. To reduce temporal bias on static voxels, we design temporal decoupling to improve end-to-end OCF by temporally associating instances via predicted flows. We develop an efficient multi-head network EfficientOCF to achieve 3D OCF with our devised spatiotemporally decoupled representation. A new metric, conditional IoU (C-IoU), is also introduced to provide a robust 3D OCF performance assessment, especially in datasets with missing or incomplete annotations. The experimental results demonstrate that EfficientOCF surpasses existing baseline methods on accuracy and efficiency, achieving state-of-the-art performance with a fast inference time of 82.33 ms with a single GPU. Our code is released at: https://github.com/BIT-XJY/EfficientOCF.", + "bbox": [ + 76, + 343, + 472, + 796 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3a01581574a1a087b3e7055fc58d3f8f03b290257fd43593599f65a257d68a67.jpg", + "image_caption": [ + "Figure 1. Unlike the traditional 3D vision-based OCF pipeline, our proposed paradigm implements spatiotemporal decoupling, thus achieving the efficiency of 2D OCF methods with the instance-aware capability of 3D OCF approaches." + ], + "image_footnote": [], + "bbox": [ + 503, + 303, + 890, + 532 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 501, + 598, + 630, + 613 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Accurately estimating occupancy states in the surrounding environment based on visual inputs is crucial for autonomous mobile systems, as it enables the identification of general obstacles, including objects absent from semantic databases. Most existing vision-based occupancy estimation methods [3, 7, 9, 10, 17, 34, 39] rely on complex model architectures but often overlook temporal changes in occupancy states. To address this, occupancy forecasting (OCF) approaches [1, 8, 12, 15, 25, 38] incorporate the prediction of surrounding object movements over time. These methods use past observations to forecast occupancy states in the present and near future, providing valuable information in advance for downstream tasks such as planning and control in autonomous driving.", + "bbox": [ + 496, + 626, + 890, + 837 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Existing vision-based OCF methods predict future occupancy states for the entire 3D scene, as illustrated in Fig. 1(a). These methods process all voxel grids in each scene through end-to-end 3D feature encoding and decod", + "bbox": [ + 496, + 839, + 890, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 44 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution", + "bbox": [ + 94, + 851, + 202, + 863 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author", + "bbox": [ + 96, + 863, + 220, + 875 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "This work was funded by Science and Technology Commission of Shanghai Municipality (24DZ3101300, 24TS1402600, 24TS1402800).", + "bbox": [ + 78, + 876, + 468, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "22338", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ing [21, 25]. However, this indiscriminate dense representation often leads to inefficient and less accurate OCF due to two main biases: (1) a spatial bias toward empty voxels, which dominate the 3D space; (2) a temporal changing bias toward static objects, as only a small subset of movable objects change positions in near future. These biases not only result in high computational costs and wasteful resource use but also increase uncertainty in predicting movable object occupancy, leading to rapid error accumulation in forecasts.", + "bbox": [ + 75, + 90, + 468, + 226 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To tackle these issues, we propose a novel spatiotemporal decoupling representation and an efficient network EfficientOCF for vision-based OCF. As shown in Fig. 1(b), it decouples 3D OCF both spatially and temporally. We design a new spatial representation that decouples bird's-eye view (BEV) occupancy with height values. By storing only heights for 2D occupied grids, it achieves 3D OCF while avoiding the traditional memory-intensive 3D occupancy representation [20, 29, 34, 35]. For temporal changing forecasting, we decouple current instance segmentation from future occupancy estimation, leveraging instance movement understanding to enhance OCF rather than relying on end-to-end forecasting. To achieve these goals, we design EfficientOCF, a novel network with flow, segmentation, and height prediction heads. This network compresses 3D voxel features into 2D BEV features and employs lightweight 2D encoder-decoder structures to predict BEV occupancy. Simultaneously, it predicts 2D backward centripetal flow and height values for grids containing movable objects at each timestep. Flow vectors establish instance associations between current observations and future forecasts, refining OCF results at the instance level for enhanced accuracy and efficiency. The final 3D OCF is obtained by lifting refined 2D OCF with height estimation. We further introduce new metrics and conduct extensive experiments on the nuScenes [2], nuScenes-Occupancy [34], and Lyft-Level5 [11] datasets to demonstrate that our EfficientOCF generates high-quality OCF results and achieves superior real-time performance compared to the SOTA.", + "bbox": [ + 75, + 227, + 472, + 666 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, the contributions of our work are threefold:", + "bbox": [ + 94, + 667, + 467, + 683 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a spatiotemporal decoupling framework for effective and efficient vision-based 3D OCF. Spatial decoupling is designed with a representation that combines BEV occupancy with height values to avoid 3D voxel redundancy. Temporal decoupling is achieved via instance-aware associations, obtaining enhanced 3D OCF results.", + "- We introduce EfficientOCF, a novel efficient vision-based 3D OCF network, that predicts the future occupancy states of movable objects using our proposed spatiotemporally decoupled representation. EfficientOCF employs lightweight 2D encoder-decoder structures rather than 3D counterparts and incorporates an adaptive dual pooling strategy for 3D-2D feature transformation, achieving SOTA 3D OCF performance and efficiency." + ], + "bbox": [ + 76, + 689, + 467, + 898 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- We propose new evaluation metrics tailored for OCF on fine-grained general movable objects. These metrics address the shortcomings of traditional intersection over union (IoU) by minimizing penalties for false positives within the bounding box of movable objects, which enhances the rationality of assessments, especially in cases of missing or incomplete occupancy annotations.", + "bbox": [ + 500, + 90, + 890, + 196 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 210, + 640, + 226 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3D occupancy prediction. 3D occupancy prediction aims to estimate whether each voxel in the surrounding environment is currently occupied, representing general obstacles and enhancing the expressiveness of complex scenes. MonoScene, proposed by Cao et al. [3], first introduces a 3D semantic scene completion framework using only camera inputs, but it primarily focuses on front-view scenarios. To address this, TPVFormer by Huang et al. [9] proposes a tri-perspective view (TPV) representation to lift image features, achieving comparable performance with other LiDAR methods. Building on the TPV representation, they further propose SelfOcc [10], employing a self-supervised scheme to learn 3D occupancy states. Similar to TPVFormer, transformer modules [32] are also utilized in Occ3D [31], OccFormer [39], and SurroundOcc [35]. Occ3D [31] proposes CTF-Occ, a novel transformer-based course-to-fine occupancy network, achieving 3D occupancy predictions with high spatial resolution. OccFormer [39] proposes a dual-path transformer network to fully integrate 2D global features and 3D local features, while SurroundOcc proposed by Wei et al. [35] adopts spatial 2D-3D attention to lift images to the 3D volume space. Wang et al. [34] propose OpenOccupancy, the first surrounding semantic occupancy perception benchmark, and they introduce nuScenes-Occupancy dataset extending the large-scale nuScenes dataset [2] with fine-grained occupancy annotations. These approaches consider the complex challenge of inferring the spatial distribution of obstacles but ignore their temporal evolution. However, understanding the detailed and dynamic environment is crucial for planning and decision-making in autonomous driving, ensuring robust and reliable performance under various driving conditions.", + "bbox": [ + 496, + 234, + 890, + 718 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "BEV occupancy forecasting. BEV representation has been widely used in autonomous mobile systems due to its lightweight 2D format [16, 18, 19, 22, 26]. To overcome the challenge of capturing the temporal evolution in 2D format, some occupancy perception methods are proposed to achieve occupancy forecasting and instance prediction in BEV. FIERY, proposed by Hu et al. [8], is the first to use surround-view camera inputs to directly predict future instance segmentation and the motion of dynamic agents using BEV representation. Based on FIERY, StretchBEV by Akan et al. [1] incorporates a stochastic temporal model to refine residual updates. Zhang et al. [38] propose BE", + "bbox": [ + 496, + 719, + 890, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "22339", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Verse, which also uses BEV occupancy to jointly reason about multiple tasks, including object detection, semantic map construction, and motion prediction. Similarly, MotionNet by Wu et al. [36] encodes object category and motion information in each BEV grid to predict scene flow and detect objects. Compared to MotionNet, MP3 proposed by Casas et al. [4] enhances occupancy flow parameterization with probabilistic modeling for multi-modal behavior. In contrast, PowerBEV by Li et al. [15] relies sorely on a semantic segmentation head alongside a backward centripetal flow head to achieve future instance prediction. Despite their effectiveness, these BEV-based methods lack a comprehensive understanding of spatial structures along the z-axis, which is crucial for accurately capturing the full 3D states of obstacles.", + "bbox": [ + 75, + 90, + 472, + 316 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3D occupancy forecasting. Recently, researchers have increasingly explored end-to-end occupancy forecasting in 3D space. As depicted in the previous work [25], occupancy states can be directly generated from point cloud prediction [5, 23, 24, 27, 28] by trivial voxelization, but cannot achieve acceptable occupancy forecasting accuracy. OccWorld [40] and OccSora [33] use generation models to simultaneously predict the movement of the ego car and the evolution of the surrounding scenes. Ma et al. [25] introduce Cam4DOcc, the first benchmark to standardize evaluation protocol of vision-only occupancy forecasting in 3D space, and propose OCFNet, an end-to-end network to forecast dense 3D occupancy states. Compared to these 3D occupancy forecasting methods, in this work, we decouple spatiotemporal representation, where BEV occupancy with height values is utilized for spatial decoupling and instance-aware refinement in consecutive frames is designed for temporal decoupling. Based on the spatiotemporally decoupled representation, we propose a novel lightweight network EfficientOCF. Our experiments show that our proposed method achieves the SOTA performance with fast inference speed.", + "bbox": [ + 75, + 318, + 472, + 636 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Proposed Method", + "text_level": 1, + "bbox": [ + 76, + 652, + 250, + 667 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Task Definition", + "text_level": 1, + "bbox": [ + 76, + 676, + 230, + 691 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given $N_{\\mathrm{p}}$ past and present surrounding visual images $\\{\\mathcal{I}|I_t\\in \\mathbb{R}^{N_{\\mathrm{cam}}\\times 3\\times H_{\\mathrm{v}}\\times W_{\\mathrm{v}}}\\}$ $t\\in [-N_{\\mathrm{p}},0]$ as input, the vision-based OCF task aims to predict the occupancy states (occupied or free) of general movable objects in the present frame and $N_{\\mathrm{f}}$ future frames $\\{\\mathcal{O}^{\\mathrm{3D}}|O_t^{\\mathrm{3D}}\\in \\mathbb{R}^{1\\times H\\times W\\times L}\\}$ $t\\in [0,N_{\\mathrm{f}}]$ . $N_{\\mathrm{cam}}$ represents surrounding camera numbers. $H_{\\mathrm{v}}$ , $W_{\\mathrm{v}}$ represent the height and width of input images, and $H,W,L$ represent the height, width, and length of the specific range defined in the present $(t = 0)$ coordinate system.", + "bbox": [ + 76, + 700, + 468, + 838 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Compared to traditional prediction methods, our work first spatially decouples OCF in 2D BEV occupancy $\\{\\mathcal{O}^{2\\mathrm{D}}|O_t^{2\\mathrm{D}}\\in \\mathbb{R}^{1\\times H\\times W}\\}$ and corresponding height prediction $\\{\\mathcal{O}^{\\mathrm{height}}|O_t^{\\mathrm{height}}\\in \\mathbb{R}^{1\\times H\\times W}\\}$ , $t\\in [0,N_{\\mathrm{f}}]$ . It pre", + "bbox": [ + 76, + 838, + 468, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "dicts the future height values of each voxel column and then lifts $\\mathcal{O}^{2\\mathrm{D}}$ to $\\mathcal{O}^{3\\mathrm{D}}$ . Unlike static objects that remain unchanged over short periods, general movable objects exhibit more dynamic motion characteristics. Therefore, following [25], we focus on forecasting the occupancy states of general movable objects, e.g., bicycles, buses, cars, etc., in the nuScenes dataset [2]. This emphasis on movable objects enhances the relevance and applicability of the forecasting task to real-world autonomous driving scenarios, where accurately predicting the motion of these objects is vital for both safety and efficiency. While existing 3D OCF methods [25] forecast movable objects end-to-end, we propose temporal decoupling to propagate current instance segmentation through flow association and achieve refined 3D OCF.", + "bbox": [ + 496, + 90, + 893, + 303 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Network Architecture", + "text_level": 1, + "bbox": [ + 500, + 315, + 705, + 330 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The pipeline of our EfficientOCF is shown in Fig. 2, which consists of four main components: a perception module, an aggregation module, a prediction module, and a refinement module. The perception module extracts features from the input surrounding images and lifts them from 2D to 3D space following Lift-Splat-Shoot [30] to obtain 3D spatial voxel features. The aggregation module then aggregates voxel features from sequential frames and transforms them into 2D BEV features, which are fed into the following prediction module. Next, the prediction module, composed of multiple 2D branches, decodes the aggregated 2D BEV features to initial 2D OCF, backward centripetal flow, and height values respectively. Then, the refinement module propagates instance segmentation into each future timestep, which is further used to refine initial OCF results in 2D BEV. Ultimately, we use height values predicted by the prediction module to lift the refined 2D occupancy to 3D space.", + "bbox": [ + 496, + 339, + 893, + 597 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Perception module. To extract 3D voxel features from the input surrounding images, we adopt the approach used in previous studies [15, 25], applying a shared ResNet [6] backbone to derive 2D features from both past and present images. These features are represented as $\\{\\mathcal{F}^{2\\mathrm{D}}|F_t^{2\\mathrm{D}}\\in$ $\\mathbb{R}^{c_{\\mathrm{v}}\\times h_{\\mathrm{v}}\\times w_{\\mathrm{v}}}\\}$ , where $t\\in [-N_{\\mathrm{p}},0]$ , $c_{\\mathrm{v}}$ represents the 2D feature channel, and $h_\\mathrm{v}\\times w_\\mathrm{v}$ denotes the 2D spatial resolution. These 2D features are then lifted and integrated into 3D voxel features $\\{\\mathcal{F}^{\\mathrm{3D}}|F_t^{\\mathrm{3D}}\\in \\mathbb{R}^{c\\times h\\times w\\times l}\\}$ using a 2D-3D lifting module [30], where $c$ is the 3D feature channel, and $h\\times w\\times l$ represents the 3D spatial resolution.", + "bbox": [ + 496, + 598, + 893, + 763 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Aggregation module. All the obtained 3D voxel features are projected into the present $(t = 0)$ coordinate system exploiting 6-DOF ego-car poses. Here we propose an adaptive dual pooling strategy to further transform 3D voxel features into 2D BEV features. We first focus on the holistic information of each voxel column and employ average pooling to generate 2D features $\\{\\mathcal{F}^{\\mathrm{avg}}|F_t^{\\mathrm{avg}}\\in \\mathbb{R}^{c\\times h\\times w}\\}$ , where $t\\in [-N_{\\mathrm{p}},0]$ . Additionally, we use max pooling to generate the concurrent 2D features $\\{\\mathcal{F}^{\\mathrm{max}}|F_t^{\\mathrm{max}}\\in \\mathbb{R}^{c\\times h\\times w}\\}$ , cap", + "bbox": [ + 496, + 765, + 893, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "22340", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/74afc6068fb6ac97f377a9eefb4309b557db3b3cab47310e49c4120b68d32800.jpg", + "image_caption": [ + "Figure 2. Architecture of our proposed EfficientOCF. The perception module (gray area) extracts 2D visual features from the input surrounding images and lifts them to 3D voxel features. Subsequently, the voxel features are further compressed into BEV features and aggregated by the aggregation module (yellow area). Next, the prediction module (blue area) with three branches sharing the same 2D structure, achieves initial OCF in 2D BEV and predicts backward centripetal flow and height values for BEV occupancy. The refinement module (green area) forecasts instances using estimated flow to refine initial 2D OCF results, which are ultimately lifted to 3D space." + ], + "image_footnote": [], + "bbox": [ + 101, + 88, + 869, + 255 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/f3aa9cc965fa38737d2b5284bf0a9b325cf80a30b52f512688089e71d7b678a4.jpg", + "image_caption": [ + "Figure 3. The process of refining initial OCF results following temporal decoupling. We exploit step-wise association along the time axis to improve forecasting quality." + ], + "image_footnote": [], + "bbox": [ + 98, + 332, + 450, + 470 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "turing the voxel grid with prominent occupancy features implicitly corresponding to height values. Learnable weights, $\\alpha_{\\mathrm{avg}}$ and $\\alpha_{\\mathrm{max}}$ , are designed to adaptively adjust the contribution of these two types of 2D BEV features $\\mathcal{F}^{\\mathrm{avg}}$ and $\\mathcal{F}^{\\mathrm{max}}$ . Finally, we aggregate features from both $N_{\\mathrm{p}}$ past and present inputs to obtain the aggregated 2D BEV features $F^{\\mathrm{agg}} \\in \\mathbb{R}^{(N_{\\mathrm{p}} + 1) \\times c \\times h \\times w}$ .", + "bbox": [ + 75, + 521, + 468, + 626 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Prediction module. The proposed prediction module consists of three 2D convolution-based branches for forecasting initial 2D occupancy, height values, and backward centripetal flow vectors respectively. The aggregated 2D BEV features are fed into 2D encoder-decoder structures, and then processed by three predictive heads. The segmentation head estimates initial present and future occupancy states $\\{\\mathcal{O}^{2\\mathrm{D}}|O_t^{2\\mathrm{D}}\\in \\mathbb{R}^{1\\times H\\times W}\\}$ , $t\\in [0,N_{\\mathrm{f}}]$ in 2D space. Concurrently, the height head determines the height values of 2D voxel grids, denoted as $\\{\\mathcal{O}^{\\mathrm{height}}|O_t^{\\mathrm{height}}\\in \\mathbb{R}^{1\\times H\\times W}\\}$ for $t\\in [0,N_{\\mathrm{f}}]$ , based on our proposed representation with spatial decoupling. In addition, the flow head generates 2D backward centripetal flow [15] for each voxel grid between consecutive frames, denoted as $\\{\\mathcal{O}^{\\mathrm{flow}}|O_t^{\\mathrm{flow}}\\in \\mathbb{R}^{2\\times H\\times W}\\}$ , $t\\in [0,N_{\\mathrm{f}}]$ , which points from a voxel at time $t$ to its corresponding potential 2D instance center at $t - 1$ .", + "bbox": [ + 75, + 628, + 470, + 869 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Refinement module. We find that existing OCF methods [15, 25] suffer from shape divergence over time due to", + "bbox": [ + 75, + 869, + 470, + 900 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/65a1ef89433b1c5640d5160a23e11edf785d1ec4a4da3574d09e04d9ab06ad84.jpg", + "image_caption": [ + "Figure 4. EfficientOCF achieves 3D OCF by assigning 2D grids with refined occupancy states and heights changing with time." + ], + "image_footnote": [], + "bbox": [ + 521, + 333, + 870, + 469 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "insufficient instance consistency across temporal frames. To address this, we propose decoupling current instance segmentation from future occupancy forecasts and refining 3D OCF through instance association, rather than learning 3D OCF end-to-end. The motivation is that instance segmentation based on current observations is often more accurate than end-to-end forecasts for future unobserved frames. By separating segmentation from occupancy forecasting, we simplify the learning process, making it easier compared to end-to-end 3D OCF. This allows us to leverage more accurate instance shapes to improve 3D OCF results. To this end, we design a refinement module that first performs instance segmentation in 2D BEV and then uses associated instances to mask inaccurate occupancy predictions for movable objects in the initial 2D OCF results, which are then lifted by height information to obtain refined 3D OCF. Specifically, we extract the centers of instances by nonmaximum suppression (NMS) at $t = -1$ and associate pixelwise instance ID between adjacent frames iteratively over time $t \\in [0, N_{\\mathrm{f}}]$ using the predicted 2D backward centripetal flow. Here we denote the instance forecasting results as $\\{\\mathcal{M}^{2\\mathrm{D}}|M_t^{2\\mathrm{D}} \\in \\mathbb{R}^{1 \\times H \\times W}\\}$ , $t \\in [0, N_{\\mathrm{f}}]$ . For each timestep $t$ , we convert $M_t^{2\\mathrm{D}}$ to the instance mask $\\bar{M}_t^{2\\mathrm{D}}$ by $\\mathrm{CLIP}(M_t^{2\\mathrm{D}})$ , which converts voxels of instances to occupied. Ultimately, we generate the final refined 2D OCF results $\\bar{\\mathcal{O}}^{2\\mathrm{D}}(i,j) = 1$ , if $O_t^{2\\mathrm{D}}(i,j) = \\bar{M}_t^{2\\mathrm{D}}(i,j) = 1$ , else 0, where $(i,j)$ is the", + "bbox": [ + 496, + 508, + 892, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "22341", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "voxel index in BEV, as Fig. 3 shows. We therefore achieve temporal decoupling by splitting the end-to-end OCF into step-wise instance associations along the time axis. By assigning each 2D voxel grid with corresponding refined occupancy states and height values over time, we achieve lifting 2D OCF $\\bar{\\mathcal{O}}^{2\\mathrm{D}}$ to 3D OCF $\\{\\bar{\\mathcal{O}}^{3\\mathrm{D}}|\\bar{\\mathcal{O}}_t^{3\\mathrm{D}}\\in \\mathbb{R}^{1\\times H\\times W\\times L}\\}$ , $t\\in [0,N_{\\mathrm{f}}]$ , as shown in Fig. 4.", + "bbox": [ + 75, + 90, + 472, + 198 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Training and Inference", + "text_level": 1, + "bbox": [ + 76, + 205, + 290, + 222 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our approach is trained and evaluated on three publicly available datasets, including the nuScenes [2], nuScenes-Occupancy [34], and Lyft-Level5 [11] datasets.", + "bbox": [ + 75, + 228, + 468, + 273 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Data generation. We develop a novel data generation method based on three publicly available datasets. Initially, we split these datasets into sequences with a time length of $N = N_{\\mathrm{p}} + N_{\\mathrm{f}} + 1$ . Sequential semantic and instance annotations of general movable objects are then extracted for each sequence. Based on the extracted semantic information, we set the voxel grids corresponding to the general movable objects to 1 and the other positions to 0. We thereby obtain the bounding-box-aware 3D occupancy states $\\{\\hat{\\mathcal{O}}^{\\mathrm{bb}}|\\hat{\\mathcal{O}}_t^{\\mathrm{bb}}\\in \\mathbb{R}^{1\\times H\\times W\\times L}\\}$ (where 1 indicates occupied and 0 indicates free), which corresponds to the inflated annotations mentioned in [25]. In this case, we assign the heights of ground-truth bounding boxes to the voxel grids within them. Subsequently, we compress 3D voxel annotations into 2D BEV counterparts $\\{\\hat{\\mathcal{O}}^{\\mathrm{bev}}|\\hat{\\mathcal{O}}_t^{\\mathrm{bev}}\\in \\mathbb{R}^{1\\times H\\times W}\\}$ . This is achieved under the principle that if any voxel in a vertical voxel column has a value of 1, the corresponding position in the BEV is set to 1. We extract ground-truth height values of each voxel column $\\{\\hat{\\mathcal{O}}^{\\mathrm{height}}|\\hat{\\mathcal{O}}_t^{\\mathrm{height}}\\in \\mathbb{R}^{1\\times H\\times W}\\}$ by selecting the highest grid with the semantics of general movable objects. We then conduct fine-grained 3D occupancy states $\\{\\hat{\\mathcal{O}}^{\\mathrm{fg}}|\\hat{\\mathcal{O}}_t^{\\mathrm{fg}}\\in \\mathbb{R}^{1\\times H\\times W\\times L}\\}$ by assigning height values to 2D BEV occupancy annotations.", + "bbox": [ + 75, + 275, + 472, + 621 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Loss functions. We employ the cross-entropy loss as the OCF loss $\\mathcal{L}_{\\mathrm{occ}}$ , and use the smooth L1 loss for both the height prediction loss $\\mathcal{L}_{\\mathrm{height}}$ and the flow prediction loss $\\mathcal{L}_{\\mathrm{flow}}$ . The total loss used to train our EfficientOCF is the weighted sum of these three losses, computed as follows:", + "bbox": [ + 76, + 622, + 468, + 696 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {\\text {a l l}} = \\frac {1}{N _ {\\mathrm {f}} + 1} \\sum_ {t = 0} ^ {N _ {\\mathrm {f}}} \\left(\\lambda_ {1} \\mathcal {L} _ {\\text {o c c}} \\left(\\hat {O} _ {t} ^ {\\text {b e v}}, \\bar {O} _ {t} ^ {\\text {2 D}}\\right) \\right. \\tag {1} \\\\ + \\lambda_ {2} \\mathcal {L} _ {\\text {h e i g h t}} \\left(\\hat {O} _ {t} ^ {\\text {h e i g h t}}, O _ {t} ^ {\\text {h e i g h t}}\\right) + \\lambda_ {3} \\mathcal {L} _ {\\text {f l o w}} \\left(\\hat {O} _ {t} ^ {\\text {f l o w}}, O _ {t} ^ {\\text {f l o w}}\\right)), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 94, + 698, + 468, + 753 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda_{1}$ , $\\lambda_{2}$ , and $\\lambda_{3}$ are the weighting parameters to balance the supervision for different heads.", + "bbox": [ + 75, + 763, + 468, + 794 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation metrics. We employ intersection-over-union (IoU), and conditional IoU (C-IoU) to evaluate OCF performance. We also use video panoptic quality (VPQ) [13] to assess the performance of 3D instance forecasting. We utilize $\\mathrm{IoU_c}$ , $\\mathrm{IoU_f}$ , and $\\tilde{\\mathrm{IoU}}_{\\mathrm{f}}$ to separately evaluate the occupancy estimation at the present moment ( $t = 0$ ), the forecasting results for future timesteps ( $t \\in [1, N_{\\mathrm{f}}]$ ), and", + "bbox": [ + 75, + 795, + 468, + 901 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/d528d0b78a4cb304845c032d58f0e2eb2cd1ae04dd1ed9a89a8918ecd1f99b3f.jpg", + "image_caption": [ + "Figure 5. Examples of incomplete 3D occupancy labels due to LiDAR sparsity in nuScenes-Occupancy [34], which cannot comprehensively assess occupancy forecasting methods." + ], + "image_footnote": [], + "bbox": [ + 544, + 87, + 851, + 234 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "the occupancy estimation over the entire time horizon $(t\\in [0,N_{\\mathrm{f}}])$ . These evaluation metrics are computed by", + "bbox": [ + 500, + 282, + 890, + 311 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {I o U} = \\frac {1}{N _ {\\text {a l l}}} \\sum_ {t = t _ {\\mathrm {s}}} ^ {t _ {\\mathrm {e}}} \\frac {\\sum \\hat {O} _ {t} ^ {\\mathrm {o c c}} \\cdot \\bar {O} _ {t} ^ {\\mathrm {o c c}}}{\\sum \\hat {O} _ {t} ^ {\\mathrm {o c c}} + \\bar {O} _ {t} ^ {\\mathrm {o c c}} - \\hat {O} _ {t} ^ {\\mathrm {o c c}} \\cdot \\bar {O} _ {t} ^ {\\mathrm {o c c}}}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 540, + 311, + 890, + 349 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We set $\\hat{O}_t^{\\mathrm{occ}} / \\bar{O}_t^{\\mathrm{occ}}$ as $\\hat{O}_t^{\\mathrm{bev}} / \\bar{O}_t^{2\\mathrm{D}}$ to access 2D OCF performance. To evaluate the performance of forecasting 3D inflated and fine-grained occupancy respectively, we calculate IoU by setting $\\hat{O}_t^{\\mathrm{occ}} / \\bar{O}_t^{\\mathrm{occ}}$ as $\\hat{O}_t^{\\mathrm{bb}} / \\bar{O}_t^{3\\mathrm{D}}$ and $\\hat{O}_t^{\\mathrm{fg}} / \\bar{O}_t^{3\\mathrm{D}}$ . We set $t_{\\mathrm{s}} = t_{\\mathrm{e}} = 0$ and $N_{\\mathrm{all}} = 1$ to calculate IoUc, $t_{\\mathrm{s}} = 1$ , $t_{\\mathrm{e}} = N_{\\mathrm{f}}$ , and $N_{\\mathrm{all}} = N_{\\mathrm{f}} + 1$ for IoU.", + "bbox": [ + 496, + 358, + 890, + 464 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "While conventional IoU metrics can evaluate forecasting results within bounding boxes or fine-grained voxels [25], they struggle with scenarios involving incomplete 3D annotations due to LiDAR data sparsity, as illustrated in Fig. 5. The nuScenes-Occupancy dataset [34] extends the original nuScenes dataset [2] with dense semantic occupancy annotation using the proposed augmenting and purifying (AAP) pipeline. However, during the AAP's automatic labeling process, some annotation omissions are inevitable, resulting in missed 3D occupancy labels due to factors such as occlusion or sparse LiDAR observations, and even introducing pseudo occupancy labels with self-training [37]. Additionally, during the manual purifying phase, while noise removal is effective, high-precision leak detection remains challenging. Therefore, we propose a new evaluation metric for 3D occupancy forecasting, conditional IoU (C-IoU):", + "bbox": [ + 496, + 465, + 892, + 708 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {C} - \\mathrm {I o U} = \\frac {1}{N _ {\\text {a l l}}} \\sum_ {t = t _ {\\mathrm {s}}} ^ {t _ {\\mathrm {e}}} \\frac {\\left| T P _ {t} \\right| + \\left| \\widetilde {F P} _ {t} \\right|}{\\left| T P _ {t} \\right| + \\left| F N _ {t} \\right| + \\left(\\left| F P _ {t} \\right| - \\left| \\widetilde {F P} _ {t} \\right|\\right)}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 511, + 715, + 890, + 771 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $TP_{t}$ , $FN_{t}$ , and $FP_{t}$ represent true positives, false negatives, and false positives at timestep $t$ , computed in the 3D fine-grained format. $\\widetilde{FP}_{t}$ represents partial false positives of $FP_{t}$ , which lies in the corresponding annotated bounding boxes. Note that we only use this metric for nuScenes and nuScenes-Occupancy datasets because Lyft-Level5 lacks fine-grained voxel-wise annotations. C-IoU comprehensively considers predictions within", + "bbox": [ + 496, + 777, + 892, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "22342", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/56d2c0418ceac4e9519b2fbdac62564f3e64fa86c341076d39b2201ba7cc393c.jpg", + "table_caption": [ + "Table 1. Comparison of performance on 3D occupancy forecasting" + ], + "table_footnote": [], + "table_body": "
ApproachnuScenesLyft-Level5nuScenes & nuScenes-Occupancy
2D3D2D3D3D
IoUcIoUfIoUIoUcIoUfIoUIoUcIoUfIoUIoUcIoUfIoUIoUcIoUfIoUC-IoUcC-IoUfC-IoUf
PowerBEV33.6230.0730.7728.3425.4726.0545.7843.4443.9136.1534.1834.58------
OpenOccupancy28.7725.9326.5024.2622.6923.1538.1635.1135.7932.9529.7330.4014.9913.0613.4727.1124.6825.17
OccFormer38.9831.0532.7634.7528.0829.5647.0032.1035.0241.6828.5531.0015.8011.3012.1844.5635.9837.67
OCFNet-30.3426.2727.4727.3224.4024.9836.0233.1233.6931.0428.9229.3612.6610.6511.0529.6427.4727.91
OCFNet35.5032.0232.7133.2429.3629.9445.7443.6744.0839.7138.1438.4614.7312.5713.0034.8333.1433.49
EfficientOCF-(ours)31.6929.0029.5630.2428.1728.6036.3933.2433.8333.9531.1231.7015.9814.0514.4635.8934.2034.56
EfficientOCF (ours)39.9336.1536.9335.6032.7333.3247.4244.9645.4644.2741.9042.3921.2819.0219.5047.5345.5745.99
", + "bbox": [ + 78, + 104, + 885, + 234 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "bonding boxes and fine-grained occupancy predictions, accounting for the discrepancies between the nuScenes and nuScenes-Occupancy datasets to avoid penalizing false positives in bounding boxes unfairly. This mitigates the effect of missing and incomplete occupancy annotations within bounding boxes of movable objects. Additionally, we use VPQ [13] to evaluate 3D instance forecasting lifted from 2D counterparts by height values, which is calculated by", + "bbox": [ + 75, + 242, + 468, + 363 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {V P Q} \\left(\\hat {\\mathcal {O}} ^ {\\mathrm {3 D}}, \\bar {\\mathcal {O}} ^ {\\mathrm {3 D}}\\right) = \\sum_ {t = 0} ^ {N _ {\\mathrm {f}}} \\frac {\\sum_ {T P _ {t}} \\mathrm {I o U} _ {t}}{\\left| T P _ {t} \\right| + \\frac {1}{2} \\left| F P _ {t} \\right| + \\frac {1}{2} \\left| F N _ {t} \\right|}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 367, + 468, + 407 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We set $\\hat{\\mathcal{O}}^{3\\mathrm{D}}$ as $\\hat{\\mathcal{O}}^{\\mathrm{bb}}$ and $\\hat{\\mathcal{O}}^{\\mathrm{fg}}$ for inflated and fine-grained formats, denoted as $\\mathrm{VPQ}^{\\mathrm{bb}}$ and $\\mathrm{VPQ}^{\\mathrm{fg}}$ respectively.", + "bbox": [ + 76, + 421, + 468, + 455 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 76, + 470, + 209, + 488 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experimental Setups", + "text_level": 1, + "bbox": [ + 76, + 497, + 274, + 513 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Dataset details. We evaluate our proposed EfficientOCF alongside the SOTA baselines using the nuScenes [2], nuScenes-Occupancy [34] and Lyft-Level5 [11] datasets. Following existing works [15, 25, 34], we split the nuScenes and nuScenes-Occupancy datasets into training and test sets with 700 and 150 scenes respectively, and the Lyft-Level5 dataset into 130 training scenes and 50 test scenes. Each sequence has a time length $N$ of 7 frames ( $N_{\\mathrm{p}} = 2$ , $N_{\\mathrm{f}} = 4$ ), meaning that we use 2 past frames and 1 present observation to estimate the present occupancy states and forecast the counterparts of the next 4 frames. The nuScenes and nuScenes-Occupancy datasets are labeled at a frequency of $2\\mathrm{Hz}$ , and the Lyft-Level5 dataset at $5\\mathrm{Hz}$ , resulting in different forecasting time durations across these datasets.", + "bbox": [ + 75, + 521, + 468, + 733 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation details. Following [25], we set the evaluation range to $[-51.2\\mathrm{m}, 51.2\\mathrm{m}]$ for both the x-axis and y-axis, and $[-5\\mathrm{m}, 3\\mathrm{m}]$ for the z-axis. The voxel resolution is set to $0.2\\mathrm{m}$ , resulting in 3D occupancy voxels of size $512 \\times 512 \\times 40$ . We compare our EfficientOCF against four SOTA methods: PowerBEV [15], OpenOccupancy [34], OccFormer [39], and OCFNet [25]. All models are trained for 15 epochs using AdamW optimizer [14] with an initial learning rate $3 \\times 10^{-4}$ and a weight decay of 0.01. Eight A100 GPUs are used for training with a batch size of 8.", + "bbox": [ + 75, + 734, + 470, + 898 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Occupancy Forecasting Assessment", + "text_level": 1, + "bbox": [ + 498, + 241, + 808, + 258 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We first assess the occupancy forecasting performance of EfficientOCF and baseline methods, as shown in Tab. 1. The nuScenes-Occupancy dataset provides fine-grained labels for the nuScenes dataset, allowing us to conduct evaluations based on fine-grained voxels. In contrast, the Lyft-Level5 dataset only contains bounding box annotations for movable objects, and thus the evaluation on this dataset is based on inflated objects where the C-IoU metric is not available. Notably, the vanilla PowerBEV only forecasts occupancy states in 2D space. Therefore, we lift its 2D results to 3D space by assigning a fixed height value to compute the 3D metrics. This fixed height for each test scene is determined by averaging the heights of all general movable objects. For OpenOccupancy and OccFormer, we estimate the occupancy states of the present frame and then use it as the prediction of all future timesteps based on the static-world model [25]. Besides, for OpenOccupancy, OccFormer, and OCFNet, which are end-to-end 3D occupancy perception methods, we first calculate 3D metrics and then compress the 3D results to BEV space to calculate the 2D metrics. Following [25], we additionally report the performance of OCFNet and our EfficientOCF when trained on limited $1/6$ of the training sequences, denoted as OCFNet $^{-}$ and EfficientOCF $^{-}$ . Tab. 1 shows that our proposed EfficientOCF estimates the present occupancy states and forecasts future counterparts better than all the other baselines on the nuScenes, nuScenes-Occupancy, and Lyft-Level5 datasets in both 2D and 3D evaluation metrics. In particular, EfficientOCF improves the proposed C-IoU of the baselines more significantly than the other existing metrics. It is also notable that EfficientOCF $^{-}$ trained on limited data even outperforms the baseline OCFNet supervised by the holistic training set on all 3D fine-grained metrics. This validates that our method more effectively captures diverse motion patterns, even with only partial scenario perception, making it practically useful when training data is limited.", + "bbox": [ + 496, + 265, + 890, + 809 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We further report the forecasting performance with different future time horizons. As shown in Tab. 2, our EfficientOCF remains the best performance within the future time horizons $\\{0.5\\mathrm{s},1.0\\mathrm{s},1.5\\mathrm{s},2.0\\mathrm{s}\\}$ on the nuScenes and nuScenes-Occupancy datasets, and $\\{0.2\\mathrm{s},0.4\\mathrm{s},0.6\\mathrm{s},0.8\\mathrm{s}\\}$ on the Lyft-Level5 dataset. Our", + "bbox": [ + 498, + 810, + 892, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "22343", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/11baf4d8cf137d53b64911960d9f5b3f18d44abf2f82b97c5da699cd119ba0c0.jpg", + "table_caption": [ + "Table 2. Comparison of performance on 3D OCF in different future time horizons" + ], + "table_footnote": [], + "table_body": "
ApproachnuScenesLyft-Level5nuScenes & nuScenes-Occupancy
IoUf (2D)IoUf (3D)IoUf (2D)IoUf (3D)IoUf (3D)C-IoUf (3D)
0.5s1.0s1.5s2.0s0.5s1.0s1.5s2.0s0.2s0.4s0.6s0.8s0.2s0.4s0.6s0.8s0.5s1.0s1.5s2.0s0.5s1.0s1.5s2.0s
PowerBEV32.1331.3130.6130.0727.1726.4925.9225.4745.0444.6444.0543.4435.5235.1834.6834.18--------
OpenOccupancy27.8927.1126.4625.9324.2023.7923.2622.6937.2436.6235.8635.1132.0331.1930.4529.7314.0813.7113.3613.0626.6225.8325.1624.68
OccFormer34.3532.5231.6031.0531.0929.4928.6228.0839.7735.6733.4832.1034.9231.5629.7428.5512.8512.0711.6111.3039.3537.5236.5635.98
OCFNet-29.1128.1127.3826.7726.3125.5124.9024.4035.0534.4633.8033.1230.3929.9729.4528.9211.6511.2510.9410.6529.0928.3927.8827.47
OCFNet34.4533.6532.8432.0231.4030.7330.0429.3645.1844.7744.2743.6739.4039.0238.5938.1413.6413.3412.9712.5734.6434.1733.6633.14
EfficientOCF-30.9030.1629.5229.0029.7429.1328.6028.1735.3534.6433.9333.2432.8832.4331.7831.1215.0714.7014.3514.0535.8635.1634.6134.20
EfficientOCF38.8037.8536.9836.1534.8134.0733.3732.7346.7046.2045.5944.9643.5443.0742.4941.9020.2119.8619.4619.0247.6346.9146.2445.57
", + "bbox": [ + 78, + 104, + 885, + 241 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4018c459a76367314a027632f7ca7b826828f1979fc9d059029362e2466974ef.jpg", + "image_caption": [ + "Figure 6. Visualization of OCF results of EfficientOCF and OCFNet, and ground-truth from nuScenes (nusc), nuScenes-Occupancy (nusc-occ), and Lyft-Level5 datasets. The OCF results and ground-truth labels from timesteps 0 to $N_{\\mathrm{f}}$ are assigned colors from dark to light." + ], + "image_footnote": [], + "bbox": [ + 99, + 247, + 222, + 468 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a2f034abc47480312af93ff6cd6ff0bf571b94d38173d71467d1de8a8f57781f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 225, + 248, + 346, + 468 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4a43ad1722483a3c8d2fe70520e3001b10cae52195f15a20afc7b349634e026e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 361, + 248, + 480, + 468 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6743b106ca56dc9ba35c8a2af2620852b9584a659a64a2fd9628a5247ffa9921.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 486, + 248, + 607, + 468 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ac815a6c0efd2993c557daf83bd643085b383472a6d262f174122fd60c032a4f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 620, + 248, + 743, + 468 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/82919a370796b9037aa3de4286366279a885a4e428233bc0b6ac9ac60f0c437a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 746, + 248, + 867, + 468 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "temporal decoupling leads to more accurate OCF through instance forecasting associated with predicted flow at each timestep. As can be seen in Fig. 6, EfficientOCF captures better contours of movable objects (e.g., the first and the second samples) compared to the SOTA baseline OCFNet which basically overlooks shape details. This is achieved by spatial decoupling, which provides more detailed 3D structures inherent in the combination of BEV occupancy and heights. Moreover, we demonstrate that EfficientOCF can generate more TP and fewer FP predictions in 2D space by Fig. 7, which also contributes to the higher forecasting accuracy of our EfficientOCF. Besides, it can be seen from the nusc-occ ground-truth (GT) of Fig. 6 that the publicly available fine-grained GT annotations are not always accurate due to LiDAR sparsity. This reflects the need for our new metrics C-IoU.", + "bbox": [ + 75, + 503, + 470, + 744 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Studies", + "text_level": 1, + "bbox": [ + 76, + 755, + 240, + 768 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We conduct several ablation studies to validate the effectiveness of different components in EfficientOCF. All the 3D evaluation metrics are derived by comparing fine-grained predictions with corresponding ground-truth labels.", + "bbox": [ + 75, + 779, + 468, + 840 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Spatiotemporal decoupling. We refer the ablation study on spatial decoupling to Tab. 1, where EfficientOCF outperforms OCFNet that directly implements occupancy forecasting in dense 3D representation. We further ablate tem", + "bbox": [ + 75, + 840, + 468, + 900 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3db5ca1f30a3bf59e2167de13e477707ccf2b2431ce4429309a536548351143e.jpg", + "image_caption": [ + "Figure 7. Visualization of TP (green), FP (red), and FN (yellow) results against GT at continuous timesteps." + ], + "image_footnote": [], + "bbox": [ + 521, + 502, + 696, + 674 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d9ab4dd03786b1bbd83ea98ed37c06fc78d4f015cccc3fb3bdf406cde48cf232.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 696, + 502, + 870, + 674 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "poral decoupling by removing the instance-aware refinement in EfficientOCF shown in Tab. 3. We conduct a baseline EfficientOCF $\\ddagger$ that directly lifts the initial 2D OCF results with height values. As shown, our instance-aware temporal decoupling improves the initial OCF performance of the prediction module by absolute 1.64 and 2.55 percentage points in 3D $\\tilde{\\mathrm{IoU}}_{\\mathrm{f}}$ and C- $\\tilde{\\mathrm{IoU}}_{\\mathrm{f}}$ on the nuScenes-Occupancy dataset. This demonstrates that instance forecasting encompassing flow-based association provides the more accurate occupancy estimation for each decoupled timestep.", + "bbox": [ + 496, + 714, + 890, + 866 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Adaptive dual pooling. We compare our adaptive dual pooling (ADP) to different pooling strategies that compress", + "bbox": [ + 498, + 869, + 890, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "22344", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/77bae1b4d06e72888f4c44423c59a248a8d92c58d460562af38aac04f560dc10.jpg", + "table_caption": [ + "Table 3. Ablation on temporal decoupling" + ], + "table_footnote": [], + "table_body": "
Approach2D3D
IoUcIoUfIoUfIoUcIoUfIoUfC-IoUcC-IoUfC-IoUf
EfficientOCF‡39.4435.8936.6119.9817.3117.8645.5242.9043.44
EfficientOCF39.9336.1536.9321.2819.0219.5047.5345.5745.99
", + "bbox": [ + 78, + 104, + 467, + 165 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/b7767f612bc3924732130b66939cdd611c29774f21c80948fb2ffa1d894277ad.jpg", + "table_caption": [ + "Table 4. Ablation on adaptive dual pooling" + ], + "table_footnote": [], + "table_body": "
Approach2D3D
IoUcIoUfIoUfIoUcIoUfIoUfC-IoUcC-IoUfC-IoUf
Average Pooling38.0434.7935.4519.2916.7717.2944.0241.7142.18
Max Pooling38.8435.5336.2019.4216.9017.4144.6942.3342.81
ADP39.9336.1536.9321.2819.0219.5047.5345.5745.99
", + "bbox": [ + 78, + 184, + 467, + 253 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/ca72ae52ea4a3e9e075347d3404c849b8be63b1fb57ba3f5e342c772944d85ad.jpg", + "table_caption": [ + "Table 5. Ablation on segmentation/flow/height heads" + ], + "table_footnote": [], + "table_body": "
Seg. \nHeadHeight \nHeadFlow \nHead2D3D
IoUcIoUfIoUfIoUcIoUfIoUfC-IoUcC-IoUfC-IoUf
✓38.4835.2635.93------
✓✓38.5535.3335.9819.4116.8917.4144.0941.8342.29
✓✓✓39.9336.1536.9321.2819.0219.5047.5345.5745.99
", + "bbox": [ + 78, + 271, + 467, + 340 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "3D features to 2D space. Tab. 4 shows that EfficientOCF with adaptive weighting for the combination of average and max pooling outperforms the baseline with only one single pooling strategy for 3D-2D transformation. The max pooling contributes more to forecasting accuracy than the average pooling, likely because height values and max-pooled features both capture prominent structural changes along the z-axis. We also deploy our ADP strategy as the 3D-2D transformation module in PowerBEV, and it enhances the performance of the vanilla PowerBEV in 2D $\\mathrm{IoU_c}$ and $\\mathrm{IoU_f}$ (2.0s) by absolute 5.16 and 5.42 percentage points. This validates that other OCF methods can also benefit from our proposed ADP strategy.", + "bbox": [ + 75, + 349, + 467, + 546 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Multiple heads. We ablate the devised heads in the prediction module of EfficientOCF in Tab. 5. The inclusion of the height head allows the transformation of forecasting results from 2D to 3D space, and provides more structure-aware supervision signals. It thus significantly improves OCF performance on all the metrics. Furthermore, the introduction of the flow head also provides additional enhancements because flow prediction enables the model to be aware of the concrete motion pattern of each voxel grid. Besides, the flow head helps to refine the initial occupancy following our proposed temporal decoupling.", + "bbox": [ + 75, + 547, + 467, + 714 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Instance Prediction Assessment", + "text_level": 1, + "bbox": [ + 76, + 726, + 354, + 739 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This experiment presents the 3D instance prediction performance of SOTA OCFNet and our EfficientOCF with the above-mentioned VPQ metric. As shown in Tab. 6, EfficientOCF improves the VPQ of OCFNet forecasting by absolute 1.42 and 4.08 percentage points in $\\mathrm{VPQ^{bb}}$ and $\\mathrm{VPQ^{fg}}$ respectively on nuScenes, and by 8.40 percentage points on Lyft-Level5. EfficientOCF $^{-}$ still outperforms OCFNet $^{-}$ for both datasets. EfficientOCF forecasts instances in 2D space which are further lifted to 3D counterparts following our proposed spatiotemporal decoupling. OCFNet directly im", + "bbox": [ + 75, + 750, + 467, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/1d084ce9ee23c558a7a166d0f76de9d3c827fa49697f0d0ef5226646ef5e6ab4.jpg", + "table_caption": [ + "Table 6. Evaluation on 3D instance prediction" + ], + "table_footnote": [], + "table_body": "
ApproachnuScenesLyft-Level5
VPQbbVPQfgVPQBb
OCFNet-13.684.2524.11
OCFNet18.536.6128.35
EfficientOCF-(ours)14.116.5027.06
EfficientOCF(ours)19.9510.6936.75
", + "bbox": [ + 506, + 104, + 879, + 191 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/2bf19c81d2c601ace7f3a6e292084a690819ab1c8ab6fc09a3f8c175a8f27c21.jpg", + "table_caption": [ + "Table 7. Comparison of inference time" + ], + "table_footnote": [], + "table_body": "
ApproachInference time [ms]
PerceptionProcessingPredictionTotal
OCFNet107.054.1813.15124.38
EfficientOCF (ours)64.785.0512.5082.33
", + "bbox": [ + 506, + 210, + 877, + 268 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "plements instance association in dense 3D formats, leading to more prediction artifacts and uncertainties.", + "bbox": [ + 498, + 277, + 890, + 306 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Inference Time", + "text_level": 1, + "bbox": [ + 500, + 316, + 651, + 330 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We compare the runtime of OCFNet and our EfficientOCF on 3D OCF shown in Tab. 7. As seen, the processing time of EfficientOCF is slightly higher than OCFNet due to the additional 3D-2D transformation, while EfficientOCF achieves a faster overall inference speed $(82.33\\mathrm{ms} / 12.15\\mathrm{Hz})$ compared to OCFNet. Moreover, we find that EfficientOCF reduces memory consumption by $478\\mathrm{MB}$ compared to OCFNet. This is because EfficientOCF utilizes lightweight 2D encoders and decoders in the perception module to achieve 3D occupancy forecasting by lifting 2D results to 3D space using height values. In contrast, OCFNet uses heavy 3D encoders and decoders with dense 3D representation, leading to more inference time.", + "bbox": [ + 496, + 339, + 890, + 535 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 549, + 617, + 564 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper proposes spatiotemporal decoupling for efficient vision-based occupancy forecasting (OCF). We achieve spatial decoupling by transforming 3D dense representations into compact 2D BEV occupancy with height values. We also propose using temporal decoupling to refine the end-to-end OCF by step-wise instance forecasting along the time axis. Based on our proposed spatiotemporally decoupled representation, we further develop EfficientOCF, a lightweight network that concurrently forecasts future 2D occupancy states and height values of general movable objects, effectively bridging 2D predictions into the 3D space. It also refines initial OCF results with instance forecasting at each individual timestep. Additionally, we introduce new evaluation metrics providing a comprehensive assessment of OCF performance for datasets with incomplete annotations. Our experiments on extensive evaluation metrics demonstrate that EfficientOCF outperforms existing methods in both accuracy and efficiency, achieving SOTA results in both 2D and 3D spaces. We hope that the spatiotemporal decoupling pipeline proposed in this work can be used as a foundation paradigm in the field of vision-based OCF.", + "bbox": [ + 496, + 574, + 890, + 891 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "22345", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 89, + 173, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Adil Kaan Akan and Fatma Güney. Stretchbev: Stretching future instance prediction spatially and temporally. In European Conference on Computer Vision, pages 444-460. Springer, 2022. 1, 2", + "[2] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11621-11631, 2020. 2, 3, 5, 6", + "[3] Anh-Quan Cao and Raoul De Charette. Monoscene: Monocular 3d semantic scene completion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3991–4001, 2022. 1, 2", + "[4] Sergio Casas, Abbas Sadat, and Raquel Urtasun. Mp3: A unified model to map, perceive, predict and plan. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14403-14412, 2021. 3", + "[5] Hehe Fan and Yi Yang. Point recurrent neural network for moving point cloud processing. arXiv preprint arXiv:1910.08287, 2019. 3", + "[6] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 3", + "[7] Yu Hong, Qian Liu, Huayuan Cheng, Danjiao Ma, Hang Dai, Yu Wang, Guangzhi Cao, and Yong Ding. Univision: A unified framework for vision-centric 3d perception. arXiv preprint arXiv:2401.06994, 2024. 1", + "[8] Anthony Hu, Zak Murez, Nikhil Mohan, Sofia Dudas, Jeffrey Hawke, Vijay Badrinarayanan, Roberto Cipolla, and Alex Kendall. Fiery: Future instance prediction in bird's-eye view from surround monocular cameras. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15273-15282, 2021. 1, 2", + "[9] Yuanhui Huang, Wenzhao Zheng, Yunpeng Zhang, Jie Zhou, and Jiwen Lu. Tri-perspective view for vision-based 3d semantic occupancy prediction. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9223-9232, 2023. 1, 2", + "[10] Yuanhui Huang, Wenzhao Zheng, Borui Zhang, Jie Zhou, and Jiwen Lu. Selfocc: Self-supervised vision-based 3d occupancy prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19946-19956, 2024. 1, 2", + "[11] R. Kesten, M. Usman, J. Houston, T. Pandya, K. Nadhamuni, A. Ferreira, M. Yuan, B. Low, A. Jain, P. Ondruska, S. Omari, S. Shah, A. Kulkarni, A. Kazakova, C. Tao, L. Platinsky, W. Jiang, and V. Shet. Lyft level 5 perception dataset 2020, 2019, 2, 5, 6", + "[12] Tarasha Khurana, Peiyun Hu, David Held, and Deva Ramanan. Point cloud forecasting as a proxy for 4d occupancy forecasting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1116-1124, 2023. 1" + ], + "bbox": [ + 78, + 114, + 470, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Dahun Kim, Sanghyun Woo, Joon-Young Lee, and In So Kweon. Video panoptic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9859–9868, 2020. 5, 6", + "[14] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6", + "[15] Peizheng Li, Shuxiao Ding, Xieyuanli Chen, Niklas Hanselmann, Marius Cordts, and Juergen Gall. Powerbev: a powerful yet lightweight framework for instance prediction in bird's-eye view. In Proceedings of the Thirty-Second International Joint Conference on Artificial Intelligence, pages 1080–1088, 2023. 1, 3, 4, 6", + "[16] Yinhao Li, Zheng Ge, Guanyi Yu, Jinrong Yang, Zengran Wang, Yukang Shi, Jianjian Sun, and Zeming Li. Bevdepth: Acquisition of reliable depth for multi-view 3d object detection. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1477-1485, 2023. 2", + "[17] Yiming Li, Zhiding Yu, Christopher Choy, Chaowei Xiao, Jose M Alvarez, Sanja Fidler, Chen Feng, and Anima Anandkumar. Voxformer: Sparse voxel transformer for camera-based 3d semantic scene completion. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9087-9098, 2023. 1", + "[18] Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chonghao Sima, Tong Lu, Yu Qiao, and Jifeng Dai. Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In European conference on computer vision, pages 1-18. Springer, 2022. 2", + "[19] Haochen Liu, Zhiyu Huang, and Chen Lv. Multi-modal hierarchical transformer for occupancy flow field prediction in autonomous driving. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 1449-1455. IEEE, 2023. 2", + "[20] Jian Liu, Sipeng Zhang, Chuixin Kong, Wenyuan Zhang, Yuhang Wu, Yikang Ding, Borun Xu, Ruibo Ming, Donglai Wei, and Xianming Liu. Occtransformer: Improving bevformer for 3d camera-only occupancy prediction. arXiv preprint arXiv:2402.18140, 2024. 2", + "[21] Xinhao Liu, Moonjun Gong, Qi Fang, Haoyu Xie, Yiming Li, Hang Zhao, and Chen Feng. Lidar-based 4d occupancy completion and forecasting. arXiv preprint arXiv:2310.11239, 2023. 2", + "[22] Zhijian Liu, Haotian Tang, Alexander Amini, Xinyu Yang, Huizi Mao, Daniela L Rus, and Song Han. Bevfusion: Multi-task multi-sensor fusion with unified bird's-eye view representation. In 2023 IEEE international conference on robotics and automation (ICRA), pages 2774-2781. IEEE, 2023. 2", + "[23] Fan Lu, Guang Chen, Zhijun Li, Lijun Zhang, Yinlong Liu, Sanqing Qu, and Alois Knoll. Monet: Motion-based point cloud prediction network. IEEE Transactions on Intelligent Transportation Systems, 23(8):13794-13804, 2021. 3", + "[24] Zhen Luo, Junyi Ma, Zijie Zhou, and Guangming Xiong. Pcpcnet: An efficient and semantic-enhanced transformer network for point cloud prediction. IEEE Robotics and Automation Letters, 8(7):4267-4274, 2023. 3", + "[25] Junyi Ma, Xieyuanli Chen, Jiawei Huang, Jingyi Xu, Zhen Luo, Jintao Xu, Weihao Gu, Rui Ai, and Hesheng Wang." + ], + "bbox": [ + 501, + 92, + 893, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "22346", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Cam4docc: Benchmark for camera-only 4d occupancy forecasting in autonomous driving applications. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21486-21495, 2024. 1, 2, 3, 4, 5, 6", + "[26] Reza Mahjourian, Jinkyu Kim, Yuning Chai, Mingxing Tan, Ben Sapp, and Dragomir Anguelov. Occupancy flow fields for motion forecasting in autonomous driving. IEEE Robotics and Automation Letters, 7(2):5639-5646, 2022. 2", + "[27] Benedikt Mersch, Xieyuanli Chen, Jens Behley, and Cyril Stachniss. Self-supervised point cloud prediction using 3d spatio-temporal convolutional networks. In Conference on Robot Learning, pages 1444-1454. PMLR, 2022. 3", + "[28] Kaustab Pal, Aditya Sharma, Avinash Sharma, and K Madhava Krishna. Atppnet: Attention based temporal point cloud prediction network. arXiv preprint arXiv:2401.17399, 2024. 3", + "[29] Mingjie Pan, Li Liu, Jiaming Liu, Peixiang Huang, Longlong Wang, Shanghai Zhang, Shaoqing Xu, Zhiyi Lai, and Kuiyuan Yang. Uniocc: Unifying vision-centric 3d occupancy prediction with geometric and semantic rendering. arXiv preprint arXiv:2306.09117, 2023. 2", + "[30] Jonah Philion and Sanja Fidler. Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIV 16, pages 194-210. Springer, 2020. 3", + "[31] Xiaoyu Tian, Tao Jiang, Longfei Yun, Yucheng Mao, Huitong Yang, Yue Wang, Yilun Wang, and Hang Zhao. Occ3d: A large-scale 3d occupancy prediction benchmark for autonomous driving. Advances in Neural Information Processing Systems, 36, 2024. 2", + "[32] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 2", + "[33] Lening Wang, Wenzhao Zheng, Yilong Ren, Han Jiang, Zhiyong Cui, Haiyang Yu, and Jiwen Lu. Occsora: 4d occupancy generation models as world simulators for autonomous driving. arXiv preprint arXiv:2405.20337, 2024. 3", + "[34] Xiaofeng Wang, Zheng Zhu, Wenbo Xu, Yunpeng Zhang, Yi Wei, Xu Chi, Yun Ye, Dalong Du, Jiwen Lu, and Xingang Wang. Openoccupancy: A large scale benchmark for surrounding semantic occupancy perception. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17850-17859, 2023. 1, 2, 5, 6", + "[35] Yi Wei, Linqing Zhao, Wenzhao Zheng, Zheng Zhu, Jie Zhou, and Jiwen Lu. Surroundocc: Multi-camera 3d occupancy prediction for autonomous driving. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 21729-21740, 2023. 2", + "[36] Pengxiang Wu, Siheng Chen, and Dimitris N Metaxas. Motionnet: Joint perception and motion prediction for autonomous driving based on bird's eye view maps. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11385-11395, 2020. 3", + "[37] Qizhe Xie, Minh-Thang Luong, Eduard Hovy, and Quoc V Le. Self-training with noisy student improvesImagenet" + ], + "bbox": [ + 78, + 90, + 468, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "classification. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10687-10698, 2020. 5", + "[38] Yunpeng Zhang, Zheng Zhu, Wenzhao Zheng, Junjie Huang, Guan Huang, Jie Zhou, and Jiwen Lu. Beverse: Unified perception and prediction in birds-eye-view for vision-centric autonomous driving. arXiv preprint arXiv:2205.09743, 2022.1,2", + "[39] Yunpeng Zhang, Zheng Zhu, and Dalong Du. Occformer: Dual-path transformer for vision-based 3d semantic occupancy prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9433-9443, 2023. 1, 2, 6", + "[40] Wenzhao Zheng, Weiliang Chen, Yuanhui Huang, Borui Zhang, Yueqi Duan, and Jiwen Lu. Occworld: Learning a 3d occupancy world model for autonomous driving. arXiv preprint arXiv:2311.16038, 2023. 3" + ], + "bbox": [ + 501, + 92, + 890, + 330 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "22347", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Spatiotemporal Decoupling for Efficient Vision-Based Occupancy Forecasting/9ae564d7-647a-4ca8-a4dc-af9f575b1e21_model.json b/2025/Spatiotemporal Decoupling for Efficient Vision-Based Occupancy Forecasting/9ae564d7-647a-4ca8-a4dc-af9f575b1e21_model.json new file mode 100644 index 0000000000000000000000000000000000000000..12079990b25bd9efd012eeb827d1e8463b3e638c --- /dev/null +++ b/2025/Spatiotemporal Decoupling for Efficient Vision-Based Occupancy Forecasting/9ae564d7-647a-4ca8-a4dc-af9f575b1e21_model.json @@ -0,0 +1,1870 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.045 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.094, + 0.131, + 0.877, + 0.153 + ], + "angle": 0, + "content": "Spatiotemporal Decoupling for Efficient Vision-Based Occupancy Forecasting" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.18, + 0.895, + 0.272 + ], + "angle": 0, + "content": "Jingyi Xu\\(^{1*}\\) Xieyuanli Chen\\(^{2*}\\) Junyi Ma\\(^{3}\\) Jiawei Huang\\(^{4}\\) Jintao Xu\\(^{4}\\) Yue Wang\\(^{5}\\) Ling Pei\\(^{1\\dagger}\\) \n\\(^{1}\\)Shanghai Key Laboratory of Navigation and Location Based Services, Shanghai Jiao Tong University \n\\(^{2}\\)College of Intelligence Science and Technology, National University of Defense Technology \n\\(^{3}\\)Department of Automation, Shanghai Jiao Tong University \\({}^{4}\\)HAOMO.AI \n\\(^{5}\\)State Key Laboratory of Industrial Control and Technology, Zhejiang University" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.304, + 0.313, + 0.32 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.344, + 0.473, + 0.797 + ], + "angle": 0, + "content": "The task of occupancy forecasting (OCF) involves utilizing past and present perception data to predict future occupancy states of autonomous vehicle surrounding environments, which is critical for downstream tasks such as obstacle avoidance and path planning. Existing 3D OCF approaches struggle to predict plausible spatial details for movable objects and suffer from slow inference speeds due to neglecting the bias and uneven distribution of changing occupancy states in both space and time. In this paper, we propose a novel spatiotemporal decoupling vision-based paradigm to explicitly tackle the bias and achieve both effective and efficient 3D OCF. To tackle spatial bias in empty areas, we introduce a novel spatial representation that decouples the conventional dense 3D format into 2D bird's-eye view (BEV) occupancy with corresponding height values, enabling 3D OCF derived only from 2D predictions thus enhancing efficiency. To reduce temporal bias on static voxels, we design temporal decoupling to improve end-to-end OCF by temporally associating instances via predicted flows. We develop an efficient multi-head network EfficientOCF to achieve 3D OCF with our devised spatiotemporally decoupled representation. A new metric, conditional IoU (C-IoU), is also introduced to provide a robust 3D OCF performance assessment, especially in datasets with missing or incomplete annotations. The experimental results demonstrate that EfficientOCF surpasses existing baseline methods on accuracy and efficiency, achieving state-of-the-art performance with a fast inference time of 82.33 ms with a single GPU. Our code is released at: https://github.com/BIT-XJY/EfficientOCF." + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.304, + 0.892, + 0.534 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.534, + 0.892, + 0.59 + ], + "angle": 0, + "content": "Figure 1. Unlike the traditional 3D vision-based OCF pipeline, our proposed paradigm implements spatiotemporal decoupling, thus achieving the efficiency of 2D OCF methods with the instance-aware capability of 3D OCF approaches." + }, + { + "type": "title", + "bbox": [ + 0.502, + 0.599, + 0.631, + 0.614 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.627, + 0.892, + 0.838 + ], + "angle": 0, + "content": "Accurately estimating occupancy states in the surrounding environment based on visual inputs is crucial for autonomous mobile systems, as it enables the identification of general obstacles, including objects absent from semantic databases. Most existing vision-based occupancy estimation methods [3, 7, 9, 10, 17, 34, 39] rely on complex model architectures but often overlook temporal changes in occupancy states. To address this, occupancy forecasting (OCF) approaches [1, 8, 12, 15, 25, 38] incorporate the prediction of surrounding object movements over time. These methods use past observations to forecast occupancy states in the present and near future, providing valuable information in advance for downstream tasks such as planning and control in autonomous driving." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.892, + 0.901 + ], + "angle": 0, + "content": "Existing vision-based OCF methods predict future occupancy states for the entire 3D scene, as illustrated in Fig. 1(a). These methods process all voxel grids in each scene through end-to-end 3D feature encoding and decod" + }, + { + "type": "page_footnote", + "bbox": [ + 0.096, + 0.852, + 0.204, + 0.864 + ], + "angle": 0, + "content": "*Equal contribution" + }, + { + "type": "page_footnote", + "bbox": [ + 0.098, + 0.864, + 0.221, + 0.876 + ], + "angle": 0, + "content": "† Corresponding author" + }, + { + "type": "page_footnote", + "bbox": [ + 0.079, + 0.877, + 0.469, + 0.899 + ], + "angle": 0, + "content": "This work was funded by Science and Technology Commission of Shanghai Municipality (24DZ3101300, 24TS1402600, 24TS1402800)." + }, + { + "type": "list", + "bbox": [ + 0.079, + 0.852, + 0.469, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22338" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.47, + 0.227 + ], + "angle": 0, + "content": "ing [21, 25]. However, this indiscriminate dense representation often leads to inefficient and less accurate OCF due to two main biases: (1) a spatial bias toward empty voxels, which dominate the 3D space; (2) a temporal changing bias toward static objects, as only a small subset of movable objects change positions in near future. These biases not only result in high computational costs and wasteful resource use but also increase uncertainty in predicting movable object occupancy, leading to rapid error accumulation in forecasts." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.228, + 0.473, + 0.667 + ], + "angle": 0, + "content": "To tackle these issues, we propose a novel spatiotemporal decoupling representation and an efficient network EfficientOCF for vision-based OCF. As shown in Fig. 1(b), it decouples 3D OCF both spatially and temporally. We design a new spatial representation that decouples bird's-eye view (BEV) occupancy with height values. By storing only heights for 2D occupied grids, it achieves 3D OCF while avoiding the traditional memory-intensive 3D occupancy representation [20, 29, 34, 35]. For temporal changing forecasting, we decouple current instance segmentation from future occupancy estimation, leveraging instance movement understanding to enhance OCF rather than relying on end-to-end forecasting. To achieve these goals, we design EfficientOCF, a novel network with flow, segmentation, and height prediction heads. This network compresses 3D voxel features into 2D BEV features and employs lightweight 2D encoder-decoder structures to predict BEV occupancy. Simultaneously, it predicts 2D backward centripetal flow and height values for grids containing movable objects at each timestep. Flow vectors establish instance associations between current observations and future forecasts, refining OCF results at the instance level for enhanced accuracy and efficiency. The final 3D OCF is obtained by lifting refined 2D OCF with height estimation. We further introduce new metrics and conduct extensive experiments on the nuScenes [2], nuScenes-Occupancy [34], and Lyft-Level5 [11] datasets to demonstrate that our EfficientOCF generates high-quality OCF results and achieves superior real-time performance compared to the SOTA." + }, + { + "type": "text", + "bbox": [ + 0.096, + 0.669, + 0.468, + 0.684 + ], + "angle": 0, + "content": "In summary, the contributions of our work are threefold:" + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.69, + 0.468, + 0.779 + ], + "angle": 0, + "content": "- We propose a spatiotemporal decoupling framework for effective and efficient vision-based 3D OCF. Spatial decoupling is designed with a representation that combines BEV occupancy with height values to avoid 3D voxel redundancy. Temporal decoupling is achieved via instance-aware associations, obtaining enhanced 3D OCF results." + }, + { + "type": "text", + "bbox": [ + 0.078, + 0.781, + 0.468, + 0.9 + ], + "angle": 0, + "content": "- We introduce EfficientOCF, a novel efficient vision-based 3D OCF network, that predicts the future occupancy states of movable objects using our proposed spatiotemporally decoupled representation. EfficientOCF employs lightweight 2D encoder-decoder structures rather than 3D counterparts and incorporates an adaptive dual pooling strategy for 3D-2D feature transformation, achieving SOTA 3D OCF performance and efficiency." + }, + { + "type": "list", + "bbox": [ + 0.078, + 0.69, + 0.468, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.092, + 0.892, + 0.198 + ], + "angle": 0, + "content": "- We propose new evaluation metrics tailored for OCF on fine-grained general movable objects. These metrics address the shortcomings of traditional intersection over union (IoU) by minimizing penalties for false positives within the bounding box of movable objects, which enhances the rationality of assessments, especially in cases of missing or incomplete occupancy annotations." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.211, + 0.642, + 0.227 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.236, + 0.892, + 0.719 + ], + "angle": 0, + "content": "3D occupancy prediction. 3D occupancy prediction aims to estimate whether each voxel in the surrounding environment is currently occupied, representing general obstacles and enhancing the expressiveness of complex scenes. MonoScene, proposed by Cao et al. [3], first introduces a 3D semantic scene completion framework using only camera inputs, but it primarily focuses on front-view scenarios. To address this, TPVFormer by Huang et al. [9] proposes a tri-perspective view (TPV) representation to lift image features, achieving comparable performance with other LiDAR methods. Building on the TPV representation, they further propose SelfOcc [10], employing a self-supervised scheme to learn 3D occupancy states. Similar to TPVFormer, transformer modules [32] are also utilized in Occ3D [31], OccFormer [39], and SurroundOcc [35]. Occ3D [31] proposes CTF-Occ, a novel transformer-based course-to-fine occupancy network, achieving 3D occupancy predictions with high spatial resolution. OccFormer [39] proposes a dual-path transformer network to fully integrate 2D global features and 3D local features, while SurroundOcc proposed by Wei et al. [35] adopts spatial 2D-3D attention to lift images to the 3D volume space. Wang et al. [34] propose OpenOccupancy, the first surrounding semantic occupancy perception benchmark, and they introduce nuScenes-Occupancy dataset extending the large-scale nuScenes dataset [2] with fine-grained occupancy annotations. These approaches consider the complex challenge of inferring the spatial distribution of obstacles but ignore their temporal evolution. However, understanding the detailed and dynamic environment is crucial for planning and decision-making in autonomous driving, ensuring robust and reliable performance under various driving conditions." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.72, + 0.892, + 0.901 + ], + "angle": 0, + "content": "BEV occupancy forecasting. BEV representation has been widely used in autonomous mobile systems due to its lightweight 2D format [16, 18, 19, 22, 26]. To overcome the challenge of capturing the temporal evolution in 2D format, some occupancy perception methods are proposed to achieve occupancy forecasting and instance prediction in BEV. FIERY, proposed by Hu et al. [8], is the first to use surround-view camera inputs to directly predict future instance segmentation and the motion of dynamic agents using BEV representation. Based on FIERY, StretchBEV by Akan et al. [1] incorporates a stochastic temporal model to refine residual updates. Zhang et al. [38] propose BE" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22339" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.317 + ], + "angle": 0, + "content": "Verse, which also uses BEV occupancy to jointly reason about multiple tasks, including object detection, semantic map construction, and motion prediction. Similarly, MotionNet by Wu et al. [36] encodes object category and motion information in each BEV grid to predict scene flow and detect objects. Compared to MotionNet, MP3 proposed by Casas et al. [4] enhances occupancy flow parameterization with probabilistic modeling for multi-modal behavior. In contrast, PowerBEV by Li et al. [15] relies sorely on a semantic segmentation head alongside a backward centripetal flow head to achieve future instance prediction. Despite their effectiveness, these BEV-based methods lack a comprehensive understanding of spatial structures along the z-axis, which is crucial for accurately capturing the full 3D states of obstacles." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.319, + 0.473, + 0.637 + ], + "angle": 0, + "content": "3D occupancy forecasting. Recently, researchers have increasingly explored end-to-end occupancy forecasting in 3D space. As depicted in the previous work [25], occupancy states can be directly generated from point cloud prediction [5, 23, 24, 27, 28] by trivial voxelization, but cannot achieve acceptable occupancy forecasting accuracy. OccWorld [40] and OccSora [33] use generation models to simultaneously predict the movement of the ego car and the evolution of the surrounding scenes. Ma et al. [25] introduce Cam4DOcc, the first benchmark to standardize evaluation protocol of vision-only occupancy forecasting in 3D space, and propose OCFNet, an end-to-end network to forecast dense 3D occupancy states. Compared to these 3D occupancy forecasting methods, in this work, we decouple spatiotemporal representation, where BEV occupancy with height values is utilized for spatial decoupling and instance-aware refinement in consecutive frames is designed for temporal decoupling. Based on the spatiotemporally decoupled representation, we propose a novel lightweight network EfficientOCF. Our experiments show that our proposed method achieves the SOTA performance with fast inference speed." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.653, + 0.251, + 0.669 + ], + "angle": 0, + "content": "3. Proposed Method" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.678, + 0.231, + 0.693 + ], + "angle": 0, + "content": "3.1. Task Definition" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.702, + 0.47, + 0.839 + ], + "angle": 0, + "content": "Given \\(N_{\\mathrm{p}}\\) past and present surrounding visual images \\(\\{\\mathcal{I}|I_t\\in \\mathbb{R}^{N_{\\mathrm{cam}}\\times 3\\times H_{\\mathrm{v}}\\times W_{\\mathrm{v}}}\\}\\) \\(t\\in [-N_{\\mathrm{p}},0]\\) as input, the vision-based OCF task aims to predict the occupancy states (occupied or free) of general movable objects in the present frame and \\(N_{\\mathrm{f}}\\) future frames \\(\\{\\mathcal{O}^{\\mathrm{3D}}|O_t^{\\mathrm{3D}}\\in \\mathbb{R}^{1\\times H\\times W\\times L}\\}\\) \\(t\\in [0,N_{\\mathrm{f}}]\\). \\(N_{\\mathrm{cam}}\\) represents surrounding camera numbers. \\(H_{\\mathrm{v}}\\), \\(W_{\\mathrm{v}}\\) represent the height and width of input images, and \\(H,W,L\\) represent the height, width, and length of the specific range defined in the present \\((t = 0)\\) coordinate system." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.839, + 0.47, + 0.903 + ], + "angle": 0, + "content": "Compared to traditional prediction methods, our work first spatially decouples OCF in 2D BEV occupancy \\(\\{\\mathcal{O}^{2\\mathrm{D}}|O_t^{2\\mathrm{D}}\\in \\mathbb{R}^{1\\times H\\times W}\\}\\) and corresponding height prediction \\(\\{\\mathcal{O}^{\\mathrm{height}}|O_t^{\\mathrm{height}}\\in \\mathbb{R}^{1\\times H\\times W}\\}\\), \\(t\\in [0,N_{\\mathrm{f}}]\\). It pre" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.304 + ], + "angle": 0, + "content": "dicts the future height values of each voxel column and then lifts \\(\\mathcal{O}^{2\\mathrm{D}}\\) to \\(\\mathcal{O}^{3\\mathrm{D}}\\). Unlike static objects that remain unchanged over short periods, general movable objects exhibit more dynamic motion characteristics. Therefore, following [25], we focus on forecasting the occupancy states of general movable objects, e.g., bicycles, buses, cars, etc., in the nuScenes dataset [2]. This emphasis on movable objects enhances the relevance and applicability of the forecasting task to real-world autonomous driving scenarios, where accurately predicting the motion of these objects is vital for both safety and efficiency. While existing 3D OCF methods [25] forecast movable objects end-to-end, we propose temporal decoupling to propagate current instance segmentation through flow association and achieve refined 3D OCF." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.316, + 0.706, + 0.331 + ], + "angle": 0, + "content": "3.2. Network Architecture" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.34, + 0.895, + 0.598 + ], + "angle": 0, + "content": "The pipeline of our EfficientOCF is shown in Fig. 2, which consists of four main components: a perception module, an aggregation module, a prediction module, and a refinement module. The perception module extracts features from the input surrounding images and lifts them from 2D to 3D space following Lift-Splat-Shoot [30] to obtain 3D spatial voxel features. The aggregation module then aggregates voxel features from sequential frames and transforms them into 2D BEV features, which are fed into the following prediction module. Next, the prediction module, composed of multiple 2D branches, decodes the aggregated 2D BEV features to initial 2D OCF, backward centripetal flow, and height values respectively. Then, the refinement module propagates instance segmentation into each future timestep, which is further used to refine initial OCF results in 2D BEV. Ultimately, we use height values predicted by the prediction module to lift the refined 2D occupancy to 3D space." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.599, + 0.895, + 0.765 + ], + "angle": 0, + "content": "Perception module. To extract 3D voxel features from the input surrounding images, we adopt the approach used in previous studies [15, 25], applying a shared ResNet [6] backbone to derive 2D features from both past and present images. These features are represented as \\(\\{\\mathcal{F}^{2\\mathrm{D}}|F_t^{2\\mathrm{D}}\\in\\) \\(\\mathbb{R}^{c_{\\mathrm{v}}\\times h_{\\mathrm{v}}\\times w_{\\mathrm{v}}}\\}\\), where \\(t\\in [-N_{\\mathrm{p}},0]\\), \\(c_{\\mathrm{v}}\\) represents the 2D feature channel, and \\(h_\\mathrm{v}\\times w_\\mathrm{v}\\) denotes the 2D spatial resolution. These 2D features are then lifted and integrated into 3D voxel features \\(\\{\\mathcal{F}^{\\mathrm{3D}}|F_t^{\\mathrm{3D}}\\in \\mathbb{R}^{c\\times h\\times w\\times l}\\}\\) using a 2D-3D lifting module [30], where \\(c\\) is the 3D feature channel, and \\(h\\times w\\times l\\) represents the 3D spatial resolution." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.766, + 0.895, + 0.903 + ], + "angle": 0, + "content": "Aggregation module. All the obtained 3D voxel features are projected into the present \\((t = 0)\\) coordinate system exploiting 6-DOF ego-car poses. Here we propose an adaptive dual pooling strategy to further transform 3D voxel features into 2D BEV features. We first focus on the holistic information of each voxel column and employ average pooling to generate 2D features \\(\\{\\mathcal{F}^{\\mathrm{avg}}|F_t^{\\mathrm{avg}}\\in \\mathbb{R}^{c\\times h\\times w}\\}\\), where \\(t\\in [-N_{\\mathrm{p}},0]\\). Additionally, we use max pooling to generate the concurrent 2D features \\(\\{\\mathcal{F}^{\\mathrm{max}}|F_t^{\\mathrm{max}}\\in \\mathbb{R}^{c\\times h\\times w}\\}\\), cap" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22340" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.102, + 0.089, + 0.87, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.258, + 0.895, + 0.329 + ], + "angle": 0, + "content": "Figure 2. Architecture of our proposed EfficientOCF. The perception module (gray area) extracts 2D visual features from the input surrounding images and lifts them to 3D voxel features. Subsequently, the voxel features are further compressed into BEV features and aggregated by the aggregation module (yellow area). Next, the prediction module (blue area) with three branches sharing the same 2D structure, achieves initial OCF in 2D BEV and predicts backward centripetal flow and height values for BEV occupancy. The refinement module (green area) forecasts instances using estimated flow to refine initial 2D OCF results, which are ultimately lifted to 3D space." + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.333, + 0.451, + 0.472 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.474, + 0.47, + 0.516 + ], + "angle": 0, + "content": "Figure 3. The process of refining initial OCF results following temporal decoupling. We exploit step-wise association along the time axis to improve forecasting quality." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.522, + 0.47, + 0.627 + ], + "angle": 0, + "content": "turing the voxel grid with prominent occupancy features implicitly corresponding to height values. Learnable weights, \\(\\alpha_{\\mathrm{avg}}\\) and \\(\\alpha_{\\mathrm{max}}\\), are designed to adaptively adjust the contribution of these two types of 2D BEV features \\(\\mathcal{F}^{\\mathrm{avg}}\\) and \\(\\mathcal{F}^{\\mathrm{max}}\\). Finally, we aggregate features from both \\(N_{\\mathrm{p}}\\) past and present inputs to obtain the aggregated 2D BEV features \\(F^{\\mathrm{agg}} \\in \\mathbb{R}^{(N_{\\mathrm{p}} + 1) \\times c \\times h \\times w}\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.629, + 0.471, + 0.87 + ], + "angle": 0, + "content": "Prediction module. The proposed prediction module consists of three 2D convolution-based branches for forecasting initial 2D occupancy, height values, and backward centripetal flow vectors respectively. The aggregated 2D BEV features are fed into 2D encoder-decoder structures, and then processed by three predictive heads. The segmentation head estimates initial present and future occupancy states \\(\\{\\mathcal{O}^{2\\mathrm{D}}|O_t^{2\\mathrm{D}}\\in \\mathbb{R}^{1\\times H\\times W}\\}\\), \\(t\\in [0,N_{\\mathrm{f}}]\\) in 2D space. Concurrently, the height head determines the height values of 2D voxel grids, denoted as \\(\\{\\mathcal{O}^{\\mathrm{height}}|O_t^{\\mathrm{height}}\\in \\mathbb{R}^{1\\times H\\times W}\\}\\) for \\(t\\in [0,N_{\\mathrm{f}}]\\), based on our proposed representation with spatial decoupling. In addition, the flow head generates 2D backward centripetal flow [15] for each voxel grid between consecutive frames, denoted as \\(\\{\\mathcal{O}^{\\mathrm{flow}}|O_t^{\\mathrm{flow}}\\in \\mathbb{R}^{2\\times H\\times W}\\}\\), \\(t\\in [0,N_{\\mathrm{f}}]\\), which points from a voxel at time \\(t\\) to its corresponding potential 2D instance center at \\(t - 1\\)." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Refinement module. We find that existing OCF methods [15, 25] suffer from shape divergence over time due to" + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.334, + 0.872, + 0.47 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.472, + 0.892, + 0.501 + ], + "angle": 0, + "content": "Figure 4. EfficientOCF achieves 3D OCF by assigning 2D grids with refined occupancy states and heights changing with time." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.509, + 0.893, + 0.903 + ], + "angle": 0, + "content": "insufficient instance consistency across temporal frames. To address this, we propose decoupling current instance segmentation from future occupancy forecasts and refining 3D OCF through instance association, rather than learning 3D OCF end-to-end. The motivation is that instance segmentation based on current observations is often more accurate than end-to-end forecasts for future unobserved frames. By separating segmentation from occupancy forecasting, we simplify the learning process, making it easier compared to end-to-end 3D OCF. This allows us to leverage more accurate instance shapes to improve 3D OCF results. To this end, we design a refinement module that first performs instance segmentation in 2D BEV and then uses associated instances to mask inaccurate occupancy predictions for movable objects in the initial 2D OCF results, which are then lifted by height information to obtain refined 3D OCF. Specifically, we extract the centers of instances by nonmaximum suppression (NMS) at \\( t = -1 \\) and associate pixelwise instance ID between adjacent frames iteratively over time \\( t \\in [0, N_{\\mathrm{f}}] \\) using the predicted 2D backward centripetal flow. Here we denote the instance forecasting results as \\( \\{\\mathcal{M}^{2\\mathrm{D}}|M_t^{2\\mathrm{D}} \\in \\mathbb{R}^{1 \\times H \\times W}\\} \\), \\( t \\in [0, N_{\\mathrm{f}}] \\). For each timestep \\( t \\), we convert \\( M_t^{2\\mathrm{D}} \\) to the instance mask \\( \\bar{M}_t^{2\\mathrm{D}} \\) by \\( \\mathrm{CLIP}(M_t^{2\\mathrm{D}}) \\), which converts voxels of instances to occupied. Ultimately, we generate the final refined 2D OCF results \\( \\bar{\\mathcal{O}}^{2\\mathrm{D}}(i,j) = 1 \\), if \\( O_t^{2\\mathrm{D}}(i,j) = \\bar{M}_t^{2\\mathrm{D}}(i,j) = 1 \\), else 0, where \\( (i,j) \\) is the" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "22341" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.091, + 0.473, + 0.199 + ], + "angle": 0, + "content": "voxel index in BEV, as Fig. 3 shows. We therefore achieve temporal decoupling by splitting the end-to-end OCF into step-wise instance associations along the time axis. By assigning each 2D voxel grid with corresponding refined occupancy states and height values over time, we achieve lifting 2D OCF \\(\\bar{\\mathcal{O}}^{2\\mathrm{D}}\\) to 3D OCF \\(\\{\\bar{\\mathcal{O}}^{3\\mathrm{D}}|\\bar{\\mathcal{O}}_t^{3\\mathrm{D}}\\in \\mathbb{R}^{1\\times H\\times W\\times L}\\}\\), \\(t\\in [0,N_{\\mathrm{f}}]\\), as shown in Fig. 4." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.207, + 0.292, + 0.223 + ], + "angle": 0, + "content": "3.3. Training and Inference" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.229, + 0.47, + 0.275 + ], + "angle": 0, + "content": "Our approach is trained and evaluated on three publicly available datasets, including the nuScenes [2], nuScenes-Occupancy [34], and Lyft-Level5 [11] datasets." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.276, + 0.473, + 0.622 + ], + "angle": 0, + "content": "Data generation. We develop a novel data generation method based on three publicly available datasets. Initially, we split these datasets into sequences with a time length of \\(N = N_{\\mathrm{p}} + N_{\\mathrm{f}} + 1\\). Sequential semantic and instance annotations of general movable objects are then extracted for each sequence. Based on the extracted semantic information, we set the voxel grids corresponding to the general movable objects to 1 and the other positions to 0. We thereby obtain the bounding-box-aware 3D occupancy states \\(\\{\\hat{\\mathcal{O}}^{\\mathrm{bb}}|\\hat{\\mathcal{O}}_t^{\\mathrm{bb}}\\in \\mathbb{R}^{1\\times H\\times W\\times L}\\}\\) (where 1 indicates occupied and 0 indicates free), which corresponds to the inflated annotations mentioned in [25]. In this case, we assign the heights of ground-truth bounding boxes to the voxel grids within them. Subsequently, we compress 3D voxel annotations into 2D BEV counterparts \\(\\{\\hat{\\mathcal{O}}^{\\mathrm{bev}}|\\hat{\\mathcal{O}}_t^{\\mathrm{bev}}\\in \\mathbb{R}^{1\\times H\\times W}\\}\\). This is achieved under the principle that if any voxel in a vertical voxel column has a value of 1, the corresponding position in the BEV is set to 1. We extract ground-truth height values of each voxel column \\(\\{\\hat{\\mathcal{O}}^{\\mathrm{height}}|\\hat{\\mathcal{O}}_t^{\\mathrm{height}}\\in \\mathbb{R}^{1\\times H\\times W}\\}\\) by selecting the highest grid with the semantics of general movable objects. We then conduct fine-grained 3D occupancy states \\(\\{\\hat{\\mathcal{O}}^{\\mathrm{fg}}|\\hat{\\mathcal{O}}_t^{\\mathrm{fg}}\\in \\mathbb{R}^{1\\times H\\times W\\times L}\\}\\) by assigning height values to 2D BEV occupancy annotations." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.623, + 0.47, + 0.698 + ], + "angle": 0, + "content": "Loss functions. We employ the cross-entropy loss as the OCF loss \\(\\mathcal{L}_{\\mathrm{occ}}\\), and use the smooth L1 loss for both the height prediction loss \\(\\mathcal{L}_{\\mathrm{height}}\\) and the flow prediction loss \\(\\mathcal{L}_{\\mathrm{flow}}\\). The total loss used to train our EfficientOCF is the weighted sum of these three losses, computed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.095, + 0.699, + 0.47, + 0.755 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {\\text {a l l}} = \\frac {1}{N _ {\\mathrm {f}} + 1} \\sum_ {t = 0} ^ {N _ {\\mathrm {f}}} \\left(\\lambda_ {1} \\mathcal {L} _ {\\text {o c c}} \\left(\\hat {O} _ {t} ^ {\\text {b e v}}, \\bar {O} _ {t} ^ {\\text {2 D}}\\right) \\right. \\tag {1} \\\\ + \\lambda_ {2} \\mathcal {L} _ {\\text {h e i g h t}} \\left(\\hat {O} _ {t} ^ {\\text {h e i g h t}}, O _ {t} ^ {\\text {h e i g h t}}\\right) + \\lambda_ {3} \\mathcal {L} _ {\\text {f l o w}} \\left(\\hat {O} _ {t} ^ {\\text {f l o w}}, O _ {t} ^ {\\text {f l o w}}\\right)), \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.765, + 0.469, + 0.795 + ], + "angle": 0, + "content": "where \\(\\lambda_{1}\\), \\(\\lambda_{2}\\), and \\(\\lambda_{3}\\) are the weighting parameters to balance the supervision for different heads." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.47, + 0.902 + ], + "angle": 0, + "content": "Evaluation metrics. We employ intersection-over-union (IoU), and conditional IoU (C-IoU) to evaluate OCF performance. We also use video panoptic quality (VPQ) [13] to assess the performance of 3D instance forecasting. We utilize \\(\\mathrm{IoU_c}\\), \\(\\mathrm{IoU_f}\\), and \\(\\tilde{\\mathrm{IoU}}_{\\mathrm{f}}\\) to separately evaluate the occupancy estimation at the present moment (\\(t = 0\\)), the forecasting results for future timesteps (\\(t \\in [1, N_{\\mathrm{f}}]\\)), and" + }, + { + "type": "image", + "bbox": [ + 0.545, + 0.088, + 0.852, + 0.235 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.236, + 0.892, + 0.278 + ], + "angle": 0, + "content": "Figure 5. Examples of incomplete 3D occupancy labels due to LiDAR sparsity in nuScenes-Occupancy [34], which cannot comprehensively assess occupancy forecasting methods." + }, + { + "type": "text", + "bbox": [ + 0.5, + 0.283, + 0.892, + 0.313 + ], + "angle": 0, + "content": "the occupancy estimation over the entire time horizon \\((t\\in [0,N_{\\mathrm{f}}])\\). These evaluation metrics are computed by" + }, + { + "type": "equation", + "bbox": [ + 0.542, + 0.312, + 0.892, + 0.351 + ], + "angle": 0, + "content": "\\[\n\\mathrm {I o U} = \\frac {1}{N _ {\\text {a l l}}} \\sum_ {t = t _ {\\mathrm {s}}} ^ {t _ {\\mathrm {e}}} \\frac {\\sum \\hat {O} _ {t} ^ {\\mathrm {o c c}} \\cdot \\bar {O} _ {t} ^ {\\mathrm {o c c}}}{\\sum \\hat {O} _ {t} ^ {\\mathrm {o c c}} + \\bar {O} _ {t} ^ {\\mathrm {o c c}} - \\hat {O} _ {t} ^ {\\mathrm {o c c}} \\cdot \\bar {O} _ {t} ^ {\\mathrm {o c c}}}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.359, + 0.892, + 0.465 + ], + "angle": 0, + "content": "We set \\(\\hat{O}_t^{\\mathrm{occ}} / \\bar{O}_t^{\\mathrm{occ}}\\) as \\(\\hat{O}_t^{\\mathrm{bev}} / \\bar{O}_t^{2\\mathrm{D}}\\) to access 2D OCF performance. To evaluate the performance of forecasting 3D inflated and fine-grained occupancy respectively, we calculate IoU by setting \\(\\hat{O}_t^{\\mathrm{occ}} / \\bar{O}_t^{\\mathrm{occ}}\\) as \\(\\hat{O}_t^{\\mathrm{bb}} / \\bar{O}_t^{3\\mathrm{D}}\\) and \\(\\hat{O}_t^{\\mathrm{fg}} / \\bar{O}_t^{3\\mathrm{D}}\\). We set \\(t_{\\mathrm{s}} = t_{\\mathrm{e}} = 0\\) and \\(N_{\\mathrm{all}} = 1\\) to calculate IoUc, \\(t_{\\mathrm{s}} = 1\\), \\(t_{\\mathrm{e}} = N_{\\mathrm{f}}\\), and \\(N_{\\mathrm{all}} = N_{\\mathrm{f}} + 1\\) for IoU." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.467, + 0.893, + 0.709 + ], + "angle": 0, + "content": "While conventional IoU metrics can evaluate forecasting results within bounding boxes or fine-grained voxels [25], they struggle with scenarios involving incomplete 3D annotations due to LiDAR data sparsity, as illustrated in Fig. 5. The nuScenes-Occupancy dataset [34] extends the original nuScenes dataset [2] with dense semantic occupancy annotation using the proposed augmenting and purifying (AAP) pipeline. However, during the AAP's automatic labeling process, some annotation omissions are inevitable, resulting in missed 3D occupancy labels due to factors such as occlusion or sparse LiDAR observations, and even introducing pseudo occupancy labels with self-training [37]. Additionally, during the manual purifying phase, while noise removal is effective, high-precision leak detection remains challenging. Therefore, we propose a new evaluation metric for 3D occupancy forecasting, conditional IoU (C-IoU):" + }, + { + "type": "equation", + "bbox": [ + 0.512, + 0.716, + 0.892, + 0.772 + ], + "angle": 0, + "content": "\\[\n\\mathrm {C} - \\mathrm {I o U} = \\frac {1}{N _ {\\text {a l l}}} \\sum_ {t = t _ {\\mathrm {s}}} ^ {t _ {\\mathrm {e}}} \\frac {\\left| T P _ {t} \\right| + \\left| \\widetilde {F P} _ {t} \\right|}{\\left| T P _ {t} \\right| + \\left| F N _ {t} \\right| + \\left(\\left| F P _ {t} \\right| - \\left| \\widetilde {F P} _ {t} \\right|\\right)}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.778, + 0.893, + 0.901 + ], + "angle": 0, + "content": "where \\( TP_{t} \\), \\( FN_{t} \\), and \\( FP_{t} \\) represent true positives, false negatives, and false positives at timestep \\( t \\), computed in the 3D fine-grained format. \\( \\widetilde{FP}_{t} \\) represents partial false positives of \\( FP_{t} \\), which lies in the corresponding annotated bounding boxes. Note that we only use this metric for nuScenes and nuScenes-Occupancy datasets because Lyft-Level5 lacks fine-grained voxel-wise annotations. C-IoU comprehensively considers predictions within" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22342" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.287, + 0.09, + 0.684, + 0.104 + ], + "angle": 0, + "content": "Table 1. Comparison of performance on 3D occupancy forecasting" + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.105, + 0.887, + 0.235 + ], + "angle": 0, + "content": "
ApproachnuScenesLyft-Level5nuScenes & nuScenes-Occupancy
2D3D2D3D3D
IoUcIoUfIoUIoUcIoUfIoUIoUcIoUfIoUIoUcIoUfIoUIoUcIoUfIoUC-IoUcC-IoUfC-IoUf
PowerBEV33.6230.0730.7728.3425.4726.0545.7843.4443.9136.1534.1834.58------
OpenOccupancy28.7725.9326.5024.2622.6923.1538.1635.1135.7932.9529.7330.4014.9913.0613.4727.1124.6825.17
OccFormer38.9831.0532.7634.7528.0829.5647.0032.1035.0241.6828.5531.0015.8011.3012.1844.5635.9837.67
OCFNet-30.3426.2727.4727.3224.4024.9836.0233.1233.6931.0428.9229.3612.6610.6511.0529.6427.4727.91
OCFNet35.5032.0232.7133.2429.3629.9445.7443.6744.0839.7138.1438.4614.7312.5713.0034.8333.1433.49
EfficientOCF-(ours)31.6929.0029.5630.2428.1728.6036.3933.2433.8333.9531.1231.7015.9814.0514.4635.8934.2034.56
EfficientOCF (ours)39.9336.1536.9335.6032.7333.3247.4244.9645.4644.2741.9042.3921.2819.0219.5047.5345.5745.99
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.243, + 0.47, + 0.364 + ], + "angle": 0, + "content": "bonding boxes and fine-grained occupancy predictions, accounting for the discrepancies between the nuScenes and nuScenes-Occupancy datasets to avoid penalizing false positives in bounding boxes unfairly. This mitigates the effect of missing and incomplete occupancy annotations within bounding boxes of movable objects. Additionally, we use VPQ [13] to evaluate 3D instance forecasting lifted from 2D counterparts by height values, which is calculated by" + }, + { + "type": "equation", + "bbox": [ + 0.091, + 0.368, + 0.47, + 0.408 + ], + "angle": 0, + "content": "\\[\n\\mathrm {V P Q} \\left(\\hat {\\mathcal {O}} ^ {\\mathrm {3 D}}, \\bar {\\mathcal {O}} ^ {\\mathrm {3 D}}\\right) = \\sum_ {t = 0} ^ {N _ {\\mathrm {f}}} \\frac {\\sum_ {T P _ {t}} \\mathrm {I o U} _ {t}}{\\left| T P _ {t} \\right| + \\frac {1}{2} \\left| F P _ {t} \\right| + \\frac {1}{2} \\left| F N _ {t} \\right|}. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.422, + 0.47, + 0.456 + ], + "angle": 0, + "content": "We set \\(\\hat{\\mathcal{O}}^{3\\mathrm{D}}\\) as \\(\\hat{\\mathcal{O}}^{\\mathrm{bb}}\\) and \\(\\hat{\\mathcal{O}}^{\\mathrm{fg}}\\) for inflated and fine-grained formats, denoted as \\(\\mathrm{VPQ}^{\\mathrm{bb}}\\) and \\(\\mathrm{VPQ}^{\\mathrm{fg}}\\) respectively." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.471, + 0.21, + 0.489 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.498, + 0.275, + 0.515 + ], + "angle": 0, + "content": "4.1. Experimental Setups" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.522, + 0.47, + 0.734 + ], + "angle": 0, + "content": "Dataset details. We evaluate our proposed EfficientOCF alongside the SOTA baselines using the nuScenes [2], nuScenes-Occupancy [34] and Lyft-Level5 [11] datasets. Following existing works [15, 25, 34], we split the nuScenes and nuScenes-Occupancy datasets into training and test sets with 700 and 150 scenes respectively, and the Lyft-Level5 dataset into 130 training scenes and 50 test scenes. Each sequence has a time length \\(N\\) of 7 frames (\\(N_{\\mathrm{p}} = 2\\), \\(N_{\\mathrm{f}} = 4\\)), meaning that we use 2 past frames and 1 present observation to estimate the present occupancy states and forecast the counterparts of the next 4 frames. The nuScenes and nuScenes-Occupancy datasets are labeled at a frequency of \\(2\\mathrm{Hz}\\), and the Lyft-Level5 dataset at \\(5\\mathrm{Hz}\\), resulting in different forecasting time durations across these datasets." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.735, + 0.471, + 0.9 + ], + "angle": 0, + "content": "Implementation details. Following [25], we set the evaluation range to \\([-51.2\\mathrm{m}, 51.2\\mathrm{m}]\\) for both the x-axis and y-axis, and \\([-5\\mathrm{m}, 3\\mathrm{m}]\\) for the z-axis. The voxel resolution is set to \\(0.2\\mathrm{m}\\), resulting in 3D occupancy voxels of size \\(512 \\times 512 \\times 40\\). We compare our EfficientOCF against four SOTA methods: PowerBEV [15], OpenOccupancy [34], OccFormer [39], and OCFNet [25]. All models are trained for 15 epochs using AdamW optimizer [14] with an initial learning rate \\(3 \\times 10^{-4}\\) and a weight decay of 0.01. Eight A100 GPUs are used for training with a batch size of 8." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.242, + 0.809, + 0.259 + ], + "angle": 0, + "content": "4.2. Occupancy Forecasting Assessment" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.266, + 0.892, + 0.81 + ], + "angle": 0, + "content": "We first assess the occupancy forecasting performance of EfficientOCF and baseline methods, as shown in Tab. 1. The nuScenes-Occupancy dataset provides fine-grained labels for the nuScenes dataset, allowing us to conduct evaluations based on fine-grained voxels. In contrast, the Lyft-Level5 dataset only contains bounding box annotations for movable objects, and thus the evaluation on this dataset is based on inflated objects where the C-IoU metric is not available. Notably, the vanilla PowerBEV only forecasts occupancy states in 2D space. Therefore, we lift its 2D results to 3D space by assigning a fixed height value to compute the 3D metrics. This fixed height for each test scene is determined by averaging the heights of all general movable objects. For OpenOccupancy and OccFormer, we estimate the occupancy states of the present frame and then use it as the prediction of all future timesteps based on the static-world model [25]. Besides, for OpenOccupancy, OccFormer, and OCFNet, which are end-to-end 3D occupancy perception methods, we first calculate 3D metrics and then compress the 3D results to BEV space to calculate the 2D metrics. Following [25], we additionally report the performance of OCFNet and our EfficientOCF when trained on limited \\(1/6\\) of the training sequences, denoted as OCFNet\\(^{-}\\) and EfficientOCF\\(^{-}\\). Tab. 1 shows that our proposed EfficientOCF estimates the present occupancy states and forecasts future counterparts better than all the other baselines on the nuScenes, nuScenes-Occupancy, and Lyft-Level5 datasets in both 2D and 3D evaluation metrics. In particular, EfficientOCF improves the proposed C-IoU of the baselines more significantly than the other existing metrics. It is also notable that EfficientOCF\\(^{-}\\) trained on limited data even outperforms the baseline OCFNet supervised by the holistic training set on all 3D fine-grained metrics. This validates that our method more effectively captures diverse motion patterns, even with only partial scenario perception, making it practically useful when training data is limited." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.811, + 0.893, + 0.902 + ], + "angle": 0, + "content": "We further report the forecasting performance with different future time horizons. As shown in Tab. 2, our EfficientOCF remains the best performance within the future time horizons \\(\\{0.5\\mathrm{s},1.0\\mathrm{s},1.5\\mathrm{s},2.0\\mathrm{s}\\}\\) on the nuScenes and nuScenes-Occupancy datasets, and \\(\\{0.2\\mathrm{s},0.4\\mathrm{s},0.6\\mathrm{s},0.8\\mathrm{s}\\}\\) on the Lyft-Level5 dataset. Our" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22343" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.241, + 0.09, + 0.726, + 0.103 + ], + "angle": 0, + "content": "Table 2. Comparison of performance on 3D OCF in different future time horizons" + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.105, + 0.887, + 0.242 + ], + "angle": 0, + "content": "
ApproachnuScenesLyft-Level5nuScenes & nuScenes-Occupancy
IoUf (2D)IoUf (3D)IoUf (2D)IoUf (3D)IoUf (3D)C-IoUf (3D)
0.5s1.0s1.5s2.0s0.5s1.0s1.5s2.0s0.2s0.4s0.6s0.8s0.2s0.4s0.6s0.8s0.5s1.0s1.5s2.0s0.5s1.0s1.5s2.0s
PowerBEV32.1331.3130.6130.0727.1726.4925.9225.4745.0444.6444.0543.4435.5235.1834.6834.18--------
OpenOccupancy27.8927.1126.4625.9324.2023.7923.2622.6937.2436.6235.8635.1132.0331.1930.4529.7314.0813.7113.3613.0626.6225.8325.1624.68
OccFormer34.3532.5231.6031.0531.0929.4928.6228.0839.7735.6733.4832.1034.9231.5629.7428.5512.8512.0711.6111.3039.3537.5236.5635.98
OCFNet-29.1128.1127.3826.7726.3125.5124.9024.4035.0534.4633.8033.1230.3929.9729.4528.9211.6511.2510.9410.6529.0928.3927.8827.47
OCFNet34.4533.6532.8432.0231.4030.7330.0429.3645.1844.7744.2743.6739.4039.0238.5938.1413.6413.3412.9712.5734.6434.1733.6633.14
EfficientOCF-30.9030.1629.5229.0029.7429.1328.6028.1735.3534.6433.9333.2432.8832.4331.7831.1215.0714.7014.3514.0535.8635.1634.6134.20
EfficientOCF38.8037.8536.9836.1534.8134.0733.3732.7346.7046.2045.5944.9643.5443.0742.4941.9020.2119.8619.4619.0247.6346.9146.2445.57
" + }, + { + "type": "image", + "bbox": [ + 0.101, + 0.248, + 0.223, + 0.469 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.227, + 0.249, + 0.348, + 0.469 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.362, + 0.249, + 0.482, + 0.469 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.487, + 0.249, + 0.609, + 0.469 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.622, + 0.249, + 0.744, + 0.469 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.748, + 0.249, + 0.869, + 0.469 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.472, + 0.891, + 0.499 + ], + "angle": 0, + "content": "Figure 6. Visualization of OCF results of EfficientOCF and OCFNet, and ground-truth from nuScenes (nusc), nuScenes-Occupancy (nusc-occ), and Lyft-Level5 datasets. The OCF results and ground-truth labels from timesteps 0 to \\(N_{\\mathrm{f}}\\) are assigned colors from dark to light." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.505, + 0.471, + 0.745 + ], + "angle": 0, + "content": "temporal decoupling leads to more accurate OCF through instance forecasting associated with predicted flow at each timestep. As can be seen in Fig. 6, EfficientOCF captures better contours of movable objects (e.g., the first and the second samples) compared to the SOTA baseline OCFNet which basically overlooks shape details. This is achieved by spatial decoupling, which provides more detailed 3D structures inherent in the combination of BEV occupancy and heights. Moreover, we demonstrate that EfficientOCF can generate more TP and fewer FP predictions in 2D space by Fig. 7, which also contributes to the higher forecasting accuracy of our EfficientOCF. Besides, it can be seen from the nusc-occ ground-truth (GT) of Fig. 6 that the publicly available fine-grained GT annotations are not always accurate due to LiDAR sparsity. This reflects the need for our new metrics C-IoU." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.756, + 0.241, + 0.77 + ], + "angle": 0, + "content": "4.3. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.78, + 0.469, + 0.841 + ], + "angle": 0, + "content": "We conduct several ablation studies to validate the effectiveness of different components in EfficientOCF. All the 3D evaluation metrics are derived by comparing fine-grained predictions with corresponding ground-truth labels." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.841, + 0.469, + 0.901 + ], + "angle": 0, + "content": "Spatiotemporal decoupling. We refer the ablation study on spatial decoupling to Tab. 1, where EfficientOCF outperforms OCFNet that directly implements occupancy forecasting in dense 3D representation. We further ablate tem" + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.503, + 0.697, + 0.675 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.697, + 0.503, + 0.871, + 0.675 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.676, + 0.891, + 0.704 + ], + "angle": 0, + "content": "Figure 7. Visualization of TP (green), FP (red), and FN (yellow) results against GT at continuous timesteps." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.715, + 0.892, + 0.867 + ], + "angle": 0, + "content": "poral decoupling by removing the instance-aware refinement in EfficientOCF shown in Tab. 3. We conduct a baseline EfficientOCF\\(\\ddagger\\) that directly lifts the initial 2D OCF results with height values. As shown, our instance-aware temporal decoupling improves the initial OCF performance of the prediction module by absolute 1.64 and 2.55 percentage points in 3D \\(\\tilde{\\mathrm{IoU}}_{\\mathrm{f}}\\) and C-\\(\\tilde{\\mathrm{IoU}}_{\\mathrm{f}}\\) on the nuScenes-Occupancy dataset. This demonstrates that instance forecasting encompassing flow-based association provides the more accurate occupancy estimation for each decoupled timestep." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.871, + 0.891, + 0.901 + ], + "angle": 0, + "content": "Adaptive dual pooling. We compare our adaptive dual pooling (ADP) to different pooling strategies that compress" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22344" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.148, + 0.09, + 0.398, + 0.103 + ], + "angle": 0, + "content": "Table 3. Ablation on temporal decoupling" + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.106, + 0.468, + 0.166 + ], + "angle": 0, + "content": "
Approach2D3D
IoUcIoUfIoUfIoUcIoUfIoUfC-IoUcC-IoUfC-IoUf
EfficientOCF‡39.4435.8936.6119.9817.3117.8645.5242.9043.44
EfficientOCF39.9336.1536.9321.2819.0219.5047.5345.5745.99
" + }, + { + "type": "table_caption", + "bbox": [ + 0.147, + 0.17, + 0.402, + 0.183 + ], + "angle": 0, + "content": "Table 4. Ablation on adaptive dual pooling" + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.185, + 0.468, + 0.254 + ], + "angle": 0, + "content": "
Approach2D3D
IoUcIoUfIoUfIoUcIoUfIoUfC-IoUcC-IoUfC-IoUf
Average Pooling38.0434.7935.4519.2916.7717.2944.0241.7142.18
Max Pooling38.8435.5336.2019.4216.9017.4144.6942.3342.81
ADP39.9336.1536.9321.2819.0219.5047.5345.5745.99
" + }, + { + "type": "table_caption", + "bbox": [ + 0.117, + 0.256, + 0.43, + 0.269 + ], + "angle": 0, + "content": "Table 5. Ablation on segmentation/flow/height heads" + }, + { + "type": "table", + "bbox": [ + 0.08, + 0.272, + 0.468, + 0.341 + ], + "angle": 0, + "content": "
Seg. \nHeadHeight \nHeadFlow \nHead2D3D
IoUcIoUfIoUfIoUcIoUfIoUfC-IoUcC-IoUfC-IoUf
✓38.4835.2635.93------
✓✓38.5535.3335.9819.4116.8917.4144.0941.8342.29
✓✓✓39.9336.1536.9321.2819.0219.5047.5345.5745.99
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.351, + 0.468, + 0.547 + ], + "angle": 0, + "content": "3D features to 2D space. Tab. 4 shows that EfficientOCF with adaptive weighting for the combination of average and max pooling outperforms the baseline with only one single pooling strategy for 3D-2D transformation. The max pooling contributes more to forecasting accuracy than the average pooling, likely because height values and max-pooled features both capture prominent structural changes along the z-axis. We also deploy our ADP strategy as the 3D-2D transformation module in PowerBEV, and it enhances the performance of the vanilla PowerBEV in 2D \\(\\mathrm{IoU_c}\\) and \\(\\mathrm{IoU_f}\\) (2.0s) by absolute 5.16 and 5.42 percentage points. This validates that other OCF methods can also benefit from our proposed ADP strategy." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.548, + 0.468, + 0.715 + ], + "angle": 0, + "content": "Multiple heads. We ablate the devised heads in the prediction module of EfficientOCF in Tab. 5. The inclusion of the height head allows the transformation of forecasting results from 2D to 3D space, and provides more structure-aware supervision signals. It thus significantly improves OCF performance on all the metrics. Furthermore, the introduction of the flow head also provides additional enhancements because flow prediction enables the model to be aware of the concrete motion pattern of each voxel grid. Besides, the flow head helps to refine the initial occupancy following our proposed temporal decoupling." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.727, + 0.355, + 0.74 + ], + "angle": 0, + "content": "4.4. Instance Prediction Assessment" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.468, + 0.901 + ], + "angle": 0, + "content": "This experiment presents the 3D instance prediction performance of SOTA OCFNet and our EfficientOCF with the above-mentioned VPQ metric. As shown in Tab. 6, EfficientOCF improves the VPQ of OCFNet forecasting by absolute 1.42 and 4.08 percentage points in \\(\\mathrm{VPQ^{bb}}\\) and \\(\\mathrm{VPQ^{fg}}\\) respectively on nuScenes, and by 8.40 percentage points on Lyft-Level5. EfficientOCF\\(^{-}\\) still outperforms OCFNet\\(^{-}\\) for both datasets. EfficientOCF forecasts instances in 2D space which are further lifted to 3D counterparts following our proposed spatiotemporal decoupling. OCFNet directly im" + }, + { + "type": "table_caption", + "bbox": [ + 0.559, + 0.09, + 0.833, + 0.103 + ], + "angle": 0, + "content": "Table 6. Evaluation on 3D instance prediction" + }, + { + "type": "table", + "bbox": [ + 0.508, + 0.106, + 0.88, + 0.193 + ], + "angle": 0, + "content": "
ApproachnuScenesLyft-Level5
VPQbbVPQfgVPQBb
OCFNet-13.684.2524.11
OCFNet18.536.6128.35
EfficientOCF-(ours)14.116.5027.06
EfficientOCF(ours)19.9510.6936.75
" + }, + { + "type": "table_caption", + "bbox": [ + 0.581, + 0.196, + 0.812, + 0.209 + ], + "angle": 0, + "content": "Table 7. Comparison of inference time" + }, + { + "type": "table", + "bbox": [ + 0.508, + 0.212, + 0.879, + 0.269 + ], + "angle": 0, + "content": "
ApproachInference time [ms]
PerceptionProcessingPredictionTotal
OCFNet107.054.1813.15124.38
EfficientOCF (ours)64.785.0512.5082.33
" + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.278, + 0.892, + 0.308 + ], + "angle": 0, + "content": "plements instance association in dense 3D formats, leading to more prediction artifacts and uncertainties." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.317, + 0.653, + 0.332 + ], + "angle": 0, + "content": "4.5. Inference Time" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.34, + 0.892, + 0.536 + ], + "angle": 0, + "content": "We compare the runtime of OCFNet and our EfficientOCF on 3D OCF shown in Tab. 7. As seen, the processing time of EfficientOCF is slightly higher than OCFNet due to the additional 3D-2D transformation, while EfficientOCF achieves a faster overall inference speed \\((82.33\\mathrm{ms} / 12.15\\mathrm{Hz})\\) compared to OCFNet. Moreover, we find that EfficientOCF reduces memory consumption by \\(478\\mathrm{MB}\\) compared to OCFNet. This is because EfficientOCF utilizes lightweight 2D encoders and decoders in the perception module to achieve 3D occupancy forecasting by lifting 2D results to 3D space using height values. In contrast, OCFNet uses heavy 3D encoders and decoders with dense 3D representation, leading to more inference time." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.55, + 0.618, + 0.565 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.575, + 0.892, + 0.892 + ], + "angle": 0, + "content": "This paper proposes spatiotemporal decoupling for efficient vision-based occupancy forecasting (OCF). We achieve spatial decoupling by transforming 3D dense representations into compact 2D BEV occupancy with height values. We also propose using temporal decoupling to refine the end-to-end OCF by step-wise instance forecasting along the time axis. Based on our proposed spatiotemporally decoupled representation, we further develop EfficientOCF, a lightweight network that concurrently forecasts future 2D occupancy states and height values of general movable objects, effectively bridging 2D predictions into the 3D space. It also refines initial OCF results with instance forecasting at each individual timestep. Additionally, we introduce new evaluation metrics providing a comprehensive assessment of OCF performance for datasets with incomplete annotations. Our experiments on extensive evaluation metrics demonstrate that EfficientOCF outperforms existing methods in both accuracy and efficiency, achieving SOTA results in both 2D and 3D spaces. We hope that the spatiotemporal decoupling pipeline proposed in this work can be used as a foundation paradigm in the field of vision-based OCF." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "22345" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.08, + 0.09, + 0.174, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.115, + 0.47, + 0.171 + ], + "angle": 0, + "content": "[1] Adil Kaan Akan and Fatma Güney. Stretchbev: Stretching future instance prediction spatially and temporally. In European Conference on Computer Vision, pages 444-460. Springer, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.173, + 0.472, + 0.256 + ], + "angle": 0, + "content": "[2] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11621-11631, 2020. 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.259, + 0.472, + 0.314 + ], + "angle": 0, + "content": "[3] Anh-Quan Cao and Raoul De Charette. Monoscene: Monocular 3d semantic scene completion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3991–4001, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.316, + 0.47, + 0.372 + ], + "angle": 0, + "content": "[4] Sergio Casas, Abbas Sadat, and Raquel Urtasun. Mp3: A unified model to map, perceive, predict and plan. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14403-14412, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.373, + 0.47, + 0.415 + ], + "angle": 0, + "content": "[5] Hehe Fan and Yi Yang. Point recurrent neural network for moving point cloud processing. arXiv preprint arXiv:1910.08287, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.417, + 0.47, + 0.473 + ], + "angle": 0, + "content": "[6] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.475, + 0.47, + 0.53 + ], + "angle": 0, + "content": "[7] Yu Hong, Qian Liu, Huayuan Cheng, Danjiao Ma, Hang Dai, Yu Wang, Guangzhi Cao, and Yong Ding. Univision: A unified framework for vision-centric 3d perception. arXiv preprint arXiv:2401.06994, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.532, + 0.47, + 0.616 + ], + "angle": 0, + "content": "[8] Anthony Hu, Zak Murez, Nikhil Mohan, Sofia Dudas, Jeffrey Hawke, Vijay Badrinarayanan, Roberto Cipolla, and Alex Kendall. Fiery: Future instance prediction in bird's-eye view from surround monocular cameras. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15273-15282, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.617, + 0.472, + 0.687 + ], + "angle": 0, + "content": "[9] Yuanhui Huang, Wenzhao Zheng, Yunpeng Zhang, Jie Zhou, and Jiwen Lu. Tri-perspective view for vision-based 3d semantic occupancy prediction. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9223-9232, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.689, + 0.47, + 0.758 + ], + "angle": 0, + "content": "[10] Yuanhui Huang, Wenzhao Zheng, Borui Zhang, Jie Zhou, and Jiwen Lu. Selfocc: Self-supervised vision-based 3d occupancy prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19946-19956, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.76, + 0.47, + 0.829 + ], + "angle": 0, + "content": "[11] R. Kesten, M. Usman, J. Houston, T. Pandya, K. Nadhamuni, A. Ferreira, M. Yuan, B. Low, A. Jain, P. Ondruska, S. Omari, S. Shah, A. Kulkarni, A. Kazakova, C. Tao, L. Platinsky, W. Jiang, and V. Shet. Lyft level 5 perception dataset 2020, 2019, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.831, + 0.47, + 0.9 + ], + "angle": 0, + "content": "[12] Tarasha Khurana, Peiyun Hu, David Held, and Deva Ramanan. Point cloud forecasting as a proxy for 4d occupancy forecasting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1116-1124, 2023. 1" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.115, + 0.472, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.148 + ], + "angle": 0, + "content": "[13] Dahun Kim, Sanghyun Woo, Joon-Young Lee, and In So Kweon. Video panoptic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9859–9868, 2020. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.149, + 0.894, + 0.189 + ], + "angle": 0, + "content": "[14] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.191, + 0.894, + 0.272 + ], + "angle": 0, + "content": "[15] Peizheng Li, Shuxiao Ding, Xieyuanli Chen, Niklas Hanselmann, Marius Cordts, and Juergen Gall. Powerbev: a powerful yet lightweight framework for instance prediction in bird's-eye view. In Proceedings of the Thirty-Second International Joint Conference on Artificial Intelligence, pages 1080–1088, 2023. 1, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.274, + 0.894, + 0.343 + ], + "angle": 0, + "content": "[16] Yinhao Li, Zheng Ge, Guanyi Yu, Jinrong Yang, Zengran Wang, Yukang Shi, Jianjian Sun, and Zeming Li. Bevdepth: Acquisition of reliable depth for multi-view 3d object detection. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1477-1485, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.344, + 0.894, + 0.427 + ], + "angle": 0, + "content": "[17] Yiming Li, Zhiding Yu, Christopher Choy, Chaowei Xiao, Jose M Alvarez, Sanja Fidler, Chen Feng, and Anima Anandkumar. Voxformer: Sparse voxel transformer for camera-based 3d semantic scene completion. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9087-9098, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.428, + 0.893, + 0.497 + ], + "angle": 0, + "content": "[18] Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chonghao Sima, Tong Lu, Yu Qiao, and Jifeng Dai. Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In European conference on computer vision, pages 1-18. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.498, + 0.893, + 0.564 + ], + "angle": 0, + "content": "[19] Haochen Liu, Zhiyu Huang, and Chen Lv. Multi-modal hierarchical transformer for occupancy flow field prediction in autonomous driving. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 1449-1455. IEEE, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.567, + 0.893, + 0.636 + ], + "angle": 0, + "content": "[20] Jian Liu, Sipeng Zhang, Chuixin Kong, Wenyuan Zhang, Yuhang Wu, Yikang Ding, Borun Xu, Ruibo Ming, Donglai Wei, and Xianming Liu. Occtransformer: Improving bevformer for 3d camera-only occupancy prediction. arXiv preprint arXiv:2402.18140, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.637, + 0.893, + 0.691 + ], + "angle": 0, + "content": "[21] Xinhao Liu, Moonjun Gong, Qi Fang, Haoyu Xie, Yiming Li, Hang Zhao, and Chen Feng. Lidar-based 4d occupancy completion and forecasting. arXiv preprint arXiv:2310.11239, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.692, + 0.893, + 0.762 + ], + "angle": 0, + "content": "[22] Zhijian Liu, Haotian Tang, Alexander Amini, Xinyu Yang, Huizi Mao, Daniela L Rus, and Song Han. Bevfusion: Multi-task multi-sensor fusion with unified bird's-eye view representation. In 2023 IEEE international conference on robotics and automation (ICRA), pages 2774-2781. IEEE, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.762, + 0.893, + 0.817 + ], + "angle": 0, + "content": "[23] Fan Lu, Guang Chen, Zhijun Li, Lijun Zhang, Yinlong Liu, Sanqing Qu, and Alois Knoll. Monet: Motion-based point cloud prediction network. IEEE Transactions on Intelligent Transportation Systems, 23(8):13794-13804, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.818, + 0.893, + 0.872 + ], + "angle": 0, + "content": "[24] Zhen Luo, Junyi Ma, Zijie Zhou, and Guangming Xiong. Pcpcnet: An efficient and semantic-enhanced transformer network for point cloud prediction. IEEE Robotics and Automation Letters, 8(7):4267-4274, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.873, + 0.893, + 0.902 + ], + "angle": 0, + "content": "[25] Junyi Ma, Xieyuanli Chen, Jiawei Huang, Jingyi Xu, Zhen Luo, Jintao Xu, Weihao Gu, Rui Ai, and Hesheng Wang." + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "22346" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.092, + 0.47, + 0.161 + ], + "angle": 0, + "content": "Cam4docc: Benchmark for camera-only 4d occupancy forecasting in autonomous driving applications. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21486-21495, 2024. 1, 2, 3, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.164, + 0.47, + 0.219 + ], + "angle": 0, + "content": "[26] Reza Mahjourian, Jinkyu Kim, Yuning Chai, Mingxing Tan, Ben Sapp, and Dragomir Anguelov. Occupancy flow fields for motion forecasting in autonomous driving. IEEE Robotics and Automation Letters, 7(2):5639-5646, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.221, + 0.47, + 0.276 + ], + "angle": 0, + "content": "[27] Benedikt Mersch, Xieyuanli Chen, Jens Behley, and Cyril Stachniss. Self-supervised point cloud prediction using 3d spatio-temporal convolutional networks. In Conference on Robot Learning, pages 1444-1454. PMLR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.278, + 0.47, + 0.332 + ], + "angle": 0, + "content": "[28] Kaustab Pal, Aditya Sharma, Avinash Sharma, and K Madhava Krishna. Atppnet: Attention based temporal point cloud prediction network. arXiv preprint arXiv:2401.17399, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.334, + 0.47, + 0.404 + ], + "angle": 0, + "content": "[29] Mingjie Pan, Li Liu, Jiaming Liu, Peixiang Huang, Longlong Wang, Shanghai Zhang, Shaoqing Xu, Zhiyi Lai, and Kuiyuan Yang. Uniocc: Unifying vision-centric 3d occupancy prediction with geometric and semantic rendering. arXiv preprint arXiv:2306.09117, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.405, + 0.47, + 0.475 + ], + "angle": 0, + "content": "[30] Jonah Philion and Sanja Fidler. Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIV 16, pages 194-210. Springer, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.477, + 0.47, + 0.546 + ], + "angle": 0, + "content": "[31] Xiaoyu Tian, Tao Jiang, Longfei Yun, Yucheng Mao, Huitong Yang, Yue Wang, Yilun Wang, and Hang Zhao. Occ3d: A large-scale 3d occupancy prediction benchmark for autonomous driving. Advances in Neural Information Processing Systems, 36, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.548, + 0.47, + 0.574 + ], + "angle": 0, + "content": "[32] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.576, + 0.47, + 0.645 + ], + "angle": 0, + "content": "[33] Lening Wang, Wenzhao Zheng, Yilong Ren, Han Jiang, Zhiyong Cui, Haiyang Yu, and Jiwen Lu. Occsora: 4d occupancy generation models as world simulators for autonomous driving. arXiv preprint arXiv:2405.20337, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.647, + 0.47, + 0.73 + ], + "angle": 0, + "content": "[34] Xiaofeng Wang, Zheng Zhu, Wenbo Xu, Yunpeng Zhang, Yi Wei, Xu Chi, Yun Ye, Dalong Du, Jiwen Lu, and Xingang Wang. Openoccupancy: A large scale benchmark for surrounding semantic occupancy perception. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17850-17859, 2023. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.732, + 0.47, + 0.801 + ], + "angle": 0, + "content": "[35] Yi Wei, Linqing Zhao, Wenzhao Zheng, Zheng Zhu, Jie Zhou, and Jiwen Lu. Surroundocc: Multi-camera 3d occupancy prediction for autonomous driving. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 21729-21740, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.803, + 0.47, + 0.872 + ], + "angle": 0, + "content": "[36] Pengxiang Wu, Siheng Chen, and Dimitris N Metaxas. Motionnet: Joint perception and motion prediction for autonomous driving based on bird's eye view maps. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11385-11395, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.47, + 0.901 + ], + "angle": 0, + "content": "[37] Qizhe Xie, Minh-Thang Luong, Eduard Hovy, and Quoc V Le. Self-training with noisy student improvesImagenet" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.892, + 0.134 + ], + "angle": 0, + "content": "classification. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10687-10698, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.892, + 0.204 + ], + "angle": 0, + "content": "[38] Yunpeng Zhang, Zheng Zhu, Wenzhao Zheng, Junjie Huang, Guan Huang, Jie Zhou, and Jiwen Lu. Beverse: Unified perception and prediction in birds-eye-view for vision-centric autonomous driving. arXiv preprint arXiv:2205.09743, 2022.1,2" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.206, + 0.892, + 0.275 + ], + "angle": 0, + "content": "[39] Yunpeng Zhang, Zheng Zhu, and Dalong Du. Occformer: Dual-path transformer for vision-based 3d semantic occupancy prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9433-9443, 2023. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.277, + 0.892, + 0.332 + ], + "angle": 0, + "content": "[40] Wenzhao Zheng, Weiliang Chen, Yuanhui Huang, Borui Zhang, Yueqi Duan, and Jiwen Lu. Occworld: Learning a 3d occupancy world model for autonomous driving. arXiv preprint arXiv:2311.16038, 2023. 3" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.892, + 0.332 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22347" + } + ] +] \ No newline at end of file diff --git a/2025/Spatiotemporal Decoupling for Efficient Vision-Based Occupancy Forecasting/9ae564d7-647a-4ca8-a4dc-af9f575b1e21_origin.pdf b/2025/Spatiotemporal Decoupling for Efficient Vision-Based Occupancy Forecasting/9ae564d7-647a-4ca8-a4dc-af9f575b1e21_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..5c57bf58c8dbd066e931e3116d8f1580e4611906 --- /dev/null +++ b/2025/Spatiotemporal Decoupling for Efficient Vision-Based Occupancy Forecasting/9ae564d7-647a-4ca8-a4dc-af9f575b1e21_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2c9dbca67ba163d8b099001e711f446c1b83330ace6e54ed2fb5953898b11f6f +size 4850404 diff --git a/2025/Spatiotemporal Decoupling for Efficient Vision-Based Occupancy Forecasting/full.md b/2025/Spatiotemporal Decoupling for Efficient Vision-Based Occupancy Forecasting/full.md new file mode 100644 index 0000000000000000000000000000000000000000..3ba170ba09d2e8edc341f244a08fa10c5a751a81 --- /dev/null +++ b/2025/Spatiotemporal Decoupling for Efficient Vision-Based Occupancy Forecasting/full.md @@ -0,0 +1,261 @@ +# Spatiotemporal Decoupling for Efficient Vision-Based Occupancy Forecasting + +Jingyi Xu $^{1*}$ Xieyuanli Chen $^{2*}$ Junyi Ma $^{3}$ Jiawei Huang $^{4}$ Jintao Xu $^{4}$ Yue Wang $^{5}$ Ling Pei $^{1\dagger}$ $^{1}$ Shanghai Key Laboratory of Navigation and Location Based Services, Shanghai Jiao Tong University + $^{2}$ College of Intelligence Science and Technology, National University of Defense Technology + $^{3}$ Department of Automation, Shanghai Jiao Tong University ${}^{4}$ HAOMO.AI + $^{5}$ State Key Laboratory of Industrial Control and Technology, Zhejiang University + +# Abstract + +The task of occupancy forecasting (OCF) involves utilizing past and present perception data to predict future occupancy states of autonomous vehicle surrounding environments, which is critical for downstream tasks such as obstacle avoidance and path planning. Existing 3D OCF approaches struggle to predict plausible spatial details for movable objects and suffer from slow inference speeds due to neglecting the bias and uneven distribution of changing occupancy states in both space and time. In this paper, we propose a novel spatiotemporal decoupling vision-based paradigm to explicitly tackle the bias and achieve both effective and efficient 3D OCF. To tackle spatial bias in empty areas, we introduce a novel spatial representation that decouples the conventional dense 3D format into 2D bird's-eye view (BEV) occupancy with corresponding height values, enabling 3D OCF derived only from 2D predictions thus enhancing efficiency. To reduce temporal bias on static voxels, we design temporal decoupling to improve end-to-end OCF by temporally associating instances via predicted flows. We develop an efficient multi-head network EfficientOCF to achieve 3D OCF with our devised spatiotemporally decoupled representation. A new metric, conditional IoU (C-IoU), is also introduced to provide a robust 3D OCF performance assessment, especially in datasets with missing or incomplete annotations. The experimental results demonstrate that EfficientOCF surpasses existing baseline methods on accuracy and efficiency, achieving state-of-the-art performance with a fast inference time of 82.33 ms with a single GPU. Our code is released at: https://github.com/BIT-XJY/EfficientOCF. + +![](images/3a01581574a1a087b3e7055fc58d3f8f03b290257fd43593599f65a257d68a67.jpg) +Figure 1. Unlike the traditional 3D vision-based OCF pipeline, our proposed paradigm implements spatiotemporal decoupling, thus achieving the efficiency of 2D OCF methods with the instance-aware capability of 3D OCF approaches. + +# 1. Introduction + +Accurately estimating occupancy states in the surrounding environment based on visual inputs is crucial for autonomous mobile systems, as it enables the identification of general obstacles, including objects absent from semantic databases. Most existing vision-based occupancy estimation methods [3, 7, 9, 10, 17, 34, 39] rely on complex model architectures but often overlook temporal changes in occupancy states. To address this, occupancy forecasting (OCF) approaches [1, 8, 12, 15, 25, 38] incorporate the prediction of surrounding object movements over time. These methods use past observations to forecast occupancy states in the present and near future, providing valuable information in advance for downstream tasks such as planning and control in autonomous driving. + +Existing vision-based OCF methods predict future occupancy states for the entire 3D scene, as illustrated in Fig. 1(a). These methods process all voxel grids in each scene through end-to-end 3D feature encoding and decod + +ing [21, 25]. However, this indiscriminate dense representation often leads to inefficient and less accurate OCF due to two main biases: (1) a spatial bias toward empty voxels, which dominate the 3D space; (2) a temporal changing bias toward static objects, as only a small subset of movable objects change positions in near future. These biases not only result in high computational costs and wasteful resource use but also increase uncertainty in predicting movable object occupancy, leading to rapid error accumulation in forecasts. + +To tackle these issues, we propose a novel spatiotemporal decoupling representation and an efficient network EfficientOCF for vision-based OCF. As shown in Fig. 1(b), it decouples 3D OCF both spatially and temporally. We design a new spatial representation that decouples bird's-eye view (BEV) occupancy with height values. By storing only heights for 2D occupied grids, it achieves 3D OCF while avoiding the traditional memory-intensive 3D occupancy representation [20, 29, 34, 35]. For temporal changing forecasting, we decouple current instance segmentation from future occupancy estimation, leveraging instance movement understanding to enhance OCF rather than relying on end-to-end forecasting. To achieve these goals, we design EfficientOCF, a novel network with flow, segmentation, and height prediction heads. This network compresses 3D voxel features into 2D BEV features and employs lightweight 2D encoder-decoder structures to predict BEV occupancy. Simultaneously, it predicts 2D backward centripetal flow and height values for grids containing movable objects at each timestep. Flow vectors establish instance associations between current observations and future forecasts, refining OCF results at the instance level for enhanced accuracy and efficiency. The final 3D OCF is obtained by lifting refined 2D OCF with height estimation. We further introduce new metrics and conduct extensive experiments on the nuScenes [2], nuScenes-Occupancy [34], and Lyft-Level5 [11] datasets to demonstrate that our EfficientOCF generates high-quality OCF results and achieves superior real-time performance compared to the SOTA. + +In summary, the contributions of our work are threefold: + +- We propose a spatiotemporal decoupling framework for effective and efficient vision-based 3D OCF. Spatial decoupling is designed with a representation that combines BEV occupancy with height values to avoid 3D voxel redundancy. Temporal decoupling is achieved via instance-aware associations, obtaining enhanced 3D OCF results. +- We introduce EfficientOCF, a novel efficient vision-based 3D OCF network, that predicts the future occupancy states of movable objects using our proposed spatiotemporally decoupled representation. EfficientOCF employs lightweight 2D encoder-decoder structures rather than 3D counterparts and incorporates an adaptive dual pooling strategy for 3D-2D feature transformation, achieving SOTA 3D OCF performance and efficiency. + +- We propose new evaluation metrics tailored for OCF on fine-grained general movable objects. These metrics address the shortcomings of traditional intersection over union (IoU) by minimizing penalties for false positives within the bounding box of movable objects, which enhances the rationality of assessments, especially in cases of missing or incomplete occupancy annotations. + +# 2. Related Work + +3D occupancy prediction. 3D occupancy prediction aims to estimate whether each voxel in the surrounding environment is currently occupied, representing general obstacles and enhancing the expressiveness of complex scenes. MonoScene, proposed by Cao et al. [3], first introduces a 3D semantic scene completion framework using only camera inputs, but it primarily focuses on front-view scenarios. To address this, TPVFormer by Huang et al. [9] proposes a tri-perspective view (TPV) representation to lift image features, achieving comparable performance with other LiDAR methods. Building on the TPV representation, they further propose SelfOcc [10], employing a self-supervised scheme to learn 3D occupancy states. Similar to TPVFormer, transformer modules [32] are also utilized in Occ3D [31], OccFormer [39], and SurroundOcc [35]. Occ3D [31] proposes CTF-Occ, a novel transformer-based course-to-fine occupancy network, achieving 3D occupancy predictions with high spatial resolution. OccFormer [39] proposes a dual-path transformer network to fully integrate 2D global features and 3D local features, while SurroundOcc proposed by Wei et al. [35] adopts spatial 2D-3D attention to lift images to the 3D volume space. Wang et al. [34] propose OpenOccupancy, the first surrounding semantic occupancy perception benchmark, and they introduce nuScenes-Occupancy dataset extending the large-scale nuScenes dataset [2] with fine-grained occupancy annotations. These approaches consider the complex challenge of inferring the spatial distribution of obstacles but ignore their temporal evolution. However, understanding the detailed and dynamic environment is crucial for planning and decision-making in autonomous driving, ensuring robust and reliable performance under various driving conditions. + +BEV occupancy forecasting. BEV representation has been widely used in autonomous mobile systems due to its lightweight 2D format [16, 18, 19, 22, 26]. To overcome the challenge of capturing the temporal evolution in 2D format, some occupancy perception methods are proposed to achieve occupancy forecasting and instance prediction in BEV. FIERY, proposed by Hu et al. [8], is the first to use surround-view camera inputs to directly predict future instance segmentation and the motion of dynamic agents using BEV representation. Based on FIERY, StretchBEV by Akan et al. [1] incorporates a stochastic temporal model to refine residual updates. Zhang et al. [38] propose BE + +Verse, which also uses BEV occupancy to jointly reason about multiple tasks, including object detection, semantic map construction, and motion prediction. Similarly, MotionNet by Wu et al. [36] encodes object category and motion information in each BEV grid to predict scene flow and detect objects. Compared to MotionNet, MP3 proposed by Casas et al. [4] enhances occupancy flow parameterization with probabilistic modeling for multi-modal behavior. In contrast, PowerBEV by Li et al. [15] relies sorely on a semantic segmentation head alongside a backward centripetal flow head to achieve future instance prediction. Despite their effectiveness, these BEV-based methods lack a comprehensive understanding of spatial structures along the z-axis, which is crucial for accurately capturing the full 3D states of obstacles. + +3D occupancy forecasting. Recently, researchers have increasingly explored end-to-end occupancy forecasting in 3D space. As depicted in the previous work [25], occupancy states can be directly generated from point cloud prediction [5, 23, 24, 27, 28] by trivial voxelization, but cannot achieve acceptable occupancy forecasting accuracy. OccWorld [40] and OccSora [33] use generation models to simultaneously predict the movement of the ego car and the evolution of the surrounding scenes. Ma et al. [25] introduce Cam4DOcc, the first benchmark to standardize evaluation protocol of vision-only occupancy forecasting in 3D space, and propose OCFNet, an end-to-end network to forecast dense 3D occupancy states. Compared to these 3D occupancy forecasting methods, in this work, we decouple spatiotemporal representation, where BEV occupancy with height values is utilized for spatial decoupling and instance-aware refinement in consecutive frames is designed for temporal decoupling. Based on the spatiotemporally decoupled representation, we propose a novel lightweight network EfficientOCF. Our experiments show that our proposed method achieves the SOTA performance with fast inference speed. + +# 3. Proposed Method + +# 3.1. Task Definition + +Given $N_{\mathrm{p}}$ past and present surrounding visual images $\{\mathcal{I}|I_t\in \mathbb{R}^{N_{\mathrm{cam}}\times 3\times H_{\mathrm{v}}\times W_{\mathrm{v}}}\}$ $t\in [-N_{\mathrm{p}},0]$ as input, the vision-based OCF task aims to predict the occupancy states (occupied or free) of general movable objects in the present frame and $N_{\mathrm{f}}$ future frames $\{\mathcal{O}^{\mathrm{3D}}|O_t^{\mathrm{3D}}\in \mathbb{R}^{1\times H\times W\times L}\}$ $t\in [0,N_{\mathrm{f}}]$ . $N_{\mathrm{cam}}$ represents surrounding camera numbers. $H_{\mathrm{v}}$ , $W_{\mathrm{v}}$ represent the height and width of input images, and $H,W,L$ represent the height, width, and length of the specific range defined in the present $(t = 0)$ coordinate system. + +Compared to traditional prediction methods, our work first spatially decouples OCF in 2D BEV occupancy $\{\mathcal{O}^{2\mathrm{D}}|O_t^{2\mathrm{D}}\in \mathbb{R}^{1\times H\times W}\}$ and corresponding height prediction $\{\mathcal{O}^{\mathrm{height}}|O_t^{\mathrm{height}}\in \mathbb{R}^{1\times H\times W}\}$ , $t\in [0,N_{\mathrm{f}}]$ . It pre + +dicts the future height values of each voxel column and then lifts $\mathcal{O}^{2\mathrm{D}}$ to $\mathcal{O}^{3\mathrm{D}}$ . Unlike static objects that remain unchanged over short periods, general movable objects exhibit more dynamic motion characteristics. Therefore, following [25], we focus on forecasting the occupancy states of general movable objects, e.g., bicycles, buses, cars, etc., in the nuScenes dataset [2]. This emphasis on movable objects enhances the relevance and applicability of the forecasting task to real-world autonomous driving scenarios, where accurately predicting the motion of these objects is vital for both safety and efficiency. While existing 3D OCF methods [25] forecast movable objects end-to-end, we propose temporal decoupling to propagate current instance segmentation through flow association and achieve refined 3D OCF. + +# 3.2. Network Architecture + +The pipeline of our EfficientOCF is shown in Fig. 2, which consists of four main components: a perception module, an aggregation module, a prediction module, and a refinement module. The perception module extracts features from the input surrounding images and lifts them from 2D to 3D space following Lift-Splat-Shoot [30] to obtain 3D spatial voxel features. The aggregation module then aggregates voxel features from sequential frames and transforms them into 2D BEV features, which are fed into the following prediction module. Next, the prediction module, composed of multiple 2D branches, decodes the aggregated 2D BEV features to initial 2D OCF, backward centripetal flow, and height values respectively. Then, the refinement module propagates instance segmentation into each future timestep, which is further used to refine initial OCF results in 2D BEV. Ultimately, we use height values predicted by the prediction module to lift the refined 2D occupancy to 3D space. + +Perception module. To extract 3D voxel features from the input surrounding images, we adopt the approach used in previous studies [15, 25], applying a shared ResNet [6] backbone to derive 2D features from both past and present images. These features are represented as $\{\mathcal{F}^{2\mathrm{D}}|F_t^{2\mathrm{D}}\in$ $\mathbb{R}^{c_{\mathrm{v}}\times h_{\mathrm{v}}\times w_{\mathrm{v}}}\}$ , where $t\in [-N_{\mathrm{p}},0]$ , $c_{\mathrm{v}}$ represents the 2D feature channel, and $h_\mathrm{v}\times w_\mathrm{v}$ denotes the 2D spatial resolution. These 2D features are then lifted and integrated into 3D voxel features $\{\mathcal{F}^{\mathrm{3D}}|F_t^{\mathrm{3D}}\in \mathbb{R}^{c\times h\times w\times l}\}$ using a 2D-3D lifting module [30], where $c$ is the 3D feature channel, and $h\times w\times l$ represents the 3D spatial resolution. + +Aggregation module. All the obtained 3D voxel features are projected into the present $(t = 0)$ coordinate system exploiting 6-DOF ego-car poses. Here we propose an adaptive dual pooling strategy to further transform 3D voxel features into 2D BEV features. We first focus on the holistic information of each voxel column and employ average pooling to generate 2D features $\{\mathcal{F}^{\mathrm{avg}}|F_t^{\mathrm{avg}}\in \mathbb{R}^{c\times h\times w}\}$ , where $t\in [-N_{\mathrm{p}},0]$ . Additionally, we use max pooling to generate the concurrent 2D features $\{\mathcal{F}^{\mathrm{max}}|F_t^{\mathrm{max}}\in \mathbb{R}^{c\times h\times w}\}$ , cap + +![](images/74afc6068fb6ac97f377a9eefb4309b557db3b3cab47310e49c4120b68d32800.jpg) +Figure 2. Architecture of our proposed EfficientOCF. The perception module (gray area) extracts 2D visual features from the input surrounding images and lifts them to 3D voxel features. Subsequently, the voxel features are further compressed into BEV features and aggregated by the aggregation module (yellow area). Next, the prediction module (blue area) with three branches sharing the same 2D structure, achieves initial OCF in 2D BEV and predicts backward centripetal flow and height values for BEV occupancy. The refinement module (green area) forecasts instances using estimated flow to refine initial 2D OCF results, which are ultimately lifted to 3D space. + +![](images/f3aa9cc965fa38737d2b5284bf0a9b325cf80a30b52f512688089e71d7b678a4.jpg) +Figure 3. The process of refining initial OCF results following temporal decoupling. We exploit step-wise association along the time axis to improve forecasting quality. + +turing the voxel grid with prominent occupancy features implicitly corresponding to height values. Learnable weights, $\alpha_{\mathrm{avg}}$ and $\alpha_{\mathrm{max}}$ , are designed to adaptively adjust the contribution of these two types of 2D BEV features $\mathcal{F}^{\mathrm{avg}}$ and $\mathcal{F}^{\mathrm{max}}$ . Finally, we aggregate features from both $N_{\mathrm{p}}$ past and present inputs to obtain the aggregated 2D BEV features $F^{\mathrm{agg}} \in \mathbb{R}^{(N_{\mathrm{p}} + 1) \times c \times h \times w}$ . + +Prediction module. The proposed prediction module consists of three 2D convolution-based branches for forecasting initial 2D occupancy, height values, and backward centripetal flow vectors respectively. The aggregated 2D BEV features are fed into 2D encoder-decoder structures, and then processed by three predictive heads. The segmentation head estimates initial present and future occupancy states $\{\mathcal{O}^{2\mathrm{D}}|O_t^{2\mathrm{D}}\in \mathbb{R}^{1\times H\times W}\}$ , $t\in [0,N_{\mathrm{f}}]$ in 2D space. Concurrently, the height head determines the height values of 2D voxel grids, denoted as $\{\mathcal{O}^{\mathrm{height}}|O_t^{\mathrm{height}}\in \mathbb{R}^{1\times H\times W}\}$ for $t\in [0,N_{\mathrm{f}}]$ , based on our proposed representation with spatial decoupling. In addition, the flow head generates 2D backward centripetal flow [15] for each voxel grid between consecutive frames, denoted as $\{\mathcal{O}^{\mathrm{flow}}|O_t^{\mathrm{flow}}\in \mathbb{R}^{2\times H\times W}\}$ , $t\in [0,N_{\mathrm{f}}]$ , which points from a voxel at time $t$ to its corresponding potential 2D instance center at $t - 1$ . + +Refinement module. We find that existing OCF methods [15, 25] suffer from shape divergence over time due to + +![](images/65a1ef89433b1c5640d5160a23e11edf785d1ec4a4da3574d09e04d9ab06ad84.jpg) +Figure 4. EfficientOCF achieves 3D OCF by assigning 2D grids with refined occupancy states and heights changing with time. + +insufficient instance consistency across temporal frames. To address this, we propose decoupling current instance segmentation from future occupancy forecasts and refining 3D OCF through instance association, rather than learning 3D OCF end-to-end. The motivation is that instance segmentation based on current observations is often more accurate than end-to-end forecasts for future unobserved frames. By separating segmentation from occupancy forecasting, we simplify the learning process, making it easier compared to end-to-end 3D OCF. This allows us to leverage more accurate instance shapes to improve 3D OCF results. To this end, we design a refinement module that first performs instance segmentation in 2D BEV and then uses associated instances to mask inaccurate occupancy predictions for movable objects in the initial 2D OCF results, which are then lifted by height information to obtain refined 3D OCF. Specifically, we extract the centers of instances by nonmaximum suppression (NMS) at $t = -1$ and associate pixelwise instance ID between adjacent frames iteratively over time $t \in [0, N_{\mathrm{f}}]$ using the predicted 2D backward centripetal flow. Here we denote the instance forecasting results as $\{\mathcal{M}^{2\mathrm{D}}|M_t^{2\mathrm{D}} \in \mathbb{R}^{1 \times H \times W}\}$ , $t \in [0, N_{\mathrm{f}}]$ . For each timestep $t$ , we convert $M_t^{2\mathrm{D}}$ to the instance mask $\bar{M}_t^{2\mathrm{D}}$ by $\mathrm{CLIP}(M_t^{2\mathrm{D}})$ , which converts voxels of instances to occupied. Ultimately, we generate the final refined 2D OCF results $\bar{\mathcal{O}}^{2\mathrm{D}}(i,j) = 1$ , if $O_t^{2\mathrm{D}}(i,j) = \bar{M}_t^{2\mathrm{D}}(i,j) = 1$ , else 0, where $(i,j)$ is the + +voxel index in BEV, as Fig. 3 shows. We therefore achieve temporal decoupling by splitting the end-to-end OCF into step-wise instance associations along the time axis. By assigning each 2D voxel grid with corresponding refined occupancy states and height values over time, we achieve lifting 2D OCF $\bar{\mathcal{O}}^{2\mathrm{D}}$ to 3D OCF $\{\bar{\mathcal{O}}^{3\mathrm{D}}|\bar{\mathcal{O}}_t^{3\mathrm{D}}\in \mathbb{R}^{1\times H\times W\times L}\}$ , $t\in [0,N_{\mathrm{f}}]$ , as shown in Fig. 4. + +# 3.3. Training and Inference + +Our approach is trained and evaluated on three publicly available datasets, including the nuScenes [2], nuScenes-Occupancy [34], and Lyft-Level5 [11] datasets. + +Data generation. We develop a novel data generation method based on three publicly available datasets. Initially, we split these datasets into sequences with a time length of $N = N_{\mathrm{p}} + N_{\mathrm{f}} + 1$ . Sequential semantic and instance annotations of general movable objects are then extracted for each sequence. Based on the extracted semantic information, we set the voxel grids corresponding to the general movable objects to 1 and the other positions to 0. We thereby obtain the bounding-box-aware 3D occupancy states $\{\hat{\mathcal{O}}^{\mathrm{bb}}|\hat{\mathcal{O}}_t^{\mathrm{bb}}\in \mathbb{R}^{1\times H\times W\times L}\}$ (where 1 indicates occupied and 0 indicates free), which corresponds to the inflated annotations mentioned in [25]. In this case, we assign the heights of ground-truth bounding boxes to the voxel grids within them. Subsequently, we compress 3D voxel annotations into 2D BEV counterparts $\{\hat{\mathcal{O}}^{\mathrm{bev}}|\hat{\mathcal{O}}_t^{\mathrm{bev}}\in \mathbb{R}^{1\times H\times W}\}$ . This is achieved under the principle that if any voxel in a vertical voxel column has a value of 1, the corresponding position in the BEV is set to 1. We extract ground-truth height values of each voxel column $\{\hat{\mathcal{O}}^{\mathrm{height}}|\hat{\mathcal{O}}_t^{\mathrm{height}}\in \mathbb{R}^{1\times H\times W}\}$ by selecting the highest grid with the semantics of general movable objects. We then conduct fine-grained 3D occupancy states $\{\hat{\mathcal{O}}^{\mathrm{fg}}|\hat{\mathcal{O}}_t^{\mathrm{fg}}\in \mathbb{R}^{1\times H\times W\times L}\}$ by assigning height values to 2D BEV occupancy annotations. + +Loss functions. We employ the cross-entropy loss as the OCF loss $\mathcal{L}_{\mathrm{occ}}$ , and use the smooth L1 loss for both the height prediction loss $\mathcal{L}_{\mathrm{height}}$ and the flow prediction loss $\mathcal{L}_{\mathrm{flow}}$ . The total loss used to train our EfficientOCF is the weighted sum of these three losses, computed as follows: + +$$ +\begin{array}{l} \mathcal {L} _ {\text {a l l}} = \frac {1}{N _ {\mathrm {f}} + 1} \sum_ {t = 0} ^ {N _ {\mathrm {f}}} \left(\lambda_ {1} \mathcal {L} _ {\text {o c c}} \left(\hat {O} _ {t} ^ {\text {b e v}}, \bar {O} _ {t} ^ {\text {2 D}}\right) \right. \tag {1} \\ + \lambda_ {2} \mathcal {L} _ {\text {h e i g h t}} \left(\hat {O} _ {t} ^ {\text {h e i g h t}}, O _ {t} ^ {\text {h e i g h t}}\right) + \lambda_ {3} \mathcal {L} _ {\text {f l o w}} \left(\hat {O} _ {t} ^ {\text {f l o w}}, O _ {t} ^ {\text {f l o w}}\right)), \\ \end{array} +$$ + +where $\lambda_{1}$ , $\lambda_{2}$ , and $\lambda_{3}$ are the weighting parameters to balance the supervision for different heads. + +Evaluation metrics. We employ intersection-over-union (IoU), and conditional IoU (C-IoU) to evaluate OCF performance. We also use video panoptic quality (VPQ) [13] to assess the performance of 3D instance forecasting. We utilize $\mathrm{IoU_c}$ , $\mathrm{IoU_f}$ , and $\tilde{\mathrm{IoU}}_{\mathrm{f}}$ to separately evaluate the occupancy estimation at the present moment ( $t = 0$ ), the forecasting results for future timesteps ( $t \in [1, N_{\mathrm{f}}]$ ), and + +![](images/d528d0b78a4cb304845c032d58f0e2eb2cd1ae04dd1ed9a89a8918ecd1f99b3f.jpg) +Figure 5. Examples of incomplete 3D occupancy labels due to LiDAR sparsity in nuScenes-Occupancy [34], which cannot comprehensively assess occupancy forecasting methods. + +the occupancy estimation over the entire time horizon $(t\in [0,N_{\mathrm{f}}])$ . These evaluation metrics are computed by + +$$ +\mathrm {I o U} = \frac {1}{N _ {\text {a l l}}} \sum_ {t = t _ {\mathrm {s}}} ^ {t _ {\mathrm {e}}} \frac {\sum \hat {O} _ {t} ^ {\mathrm {o c c}} \cdot \bar {O} _ {t} ^ {\mathrm {o c c}}}{\sum \hat {O} _ {t} ^ {\mathrm {o c c}} + \bar {O} _ {t} ^ {\mathrm {o c c}} - \hat {O} _ {t} ^ {\mathrm {o c c}} \cdot \bar {O} _ {t} ^ {\mathrm {o c c}}}. \tag {2} +$$ + +We set $\hat{O}_t^{\mathrm{occ}} / \bar{O}_t^{\mathrm{occ}}$ as $\hat{O}_t^{\mathrm{bev}} / \bar{O}_t^{2\mathrm{D}}$ to access 2D OCF performance. To evaluate the performance of forecasting 3D inflated and fine-grained occupancy respectively, we calculate IoU by setting $\hat{O}_t^{\mathrm{occ}} / \bar{O}_t^{\mathrm{occ}}$ as $\hat{O}_t^{\mathrm{bb}} / \bar{O}_t^{3\mathrm{D}}$ and $\hat{O}_t^{\mathrm{fg}} / \bar{O}_t^{3\mathrm{D}}$ . We set $t_{\mathrm{s}} = t_{\mathrm{e}} = 0$ and $N_{\mathrm{all}} = 1$ to calculate IoUc, $t_{\mathrm{s}} = 1$ , $t_{\mathrm{e}} = N_{\mathrm{f}}$ , and $N_{\mathrm{all}} = N_{\mathrm{f}} + 1$ for IoU. + +While conventional IoU metrics can evaluate forecasting results within bounding boxes or fine-grained voxels [25], they struggle with scenarios involving incomplete 3D annotations due to LiDAR data sparsity, as illustrated in Fig. 5. The nuScenes-Occupancy dataset [34] extends the original nuScenes dataset [2] with dense semantic occupancy annotation using the proposed augmenting and purifying (AAP) pipeline. However, during the AAP's automatic labeling process, some annotation omissions are inevitable, resulting in missed 3D occupancy labels due to factors such as occlusion or sparse LiDAR observations, and even introducing pseudo occupancy labels with self-training [37]. Additionally, during the manual purifying phase, while noise removal is effective, high-precision leak detection remains challenging. Therefore, we propose a new evaluation metric for 3D occupancy forecasting, conditional IoU (C-IoU): + +$$ +\mathrm {C} - \mathrm {I o U} = \frac {1}{N _ {\text {a l l}}} \sum_ {t = t _ {\mathrm {s}}} ^ {t _ {\mathrm {e}}} \frac {\left| T P _ {t} \right| + \left| \widetilde {F P} _ {t} \right|}{\left| T P _ {t} \right| + \left| F N _ {t} \right| + \left(\left| F P _ {t} \right| - \left| \widetilde {F P} _ {t} \right|\right)}, \tag {3} +$$ + +where $TP_{t}$ , $FN_{t}$ , and $FP_{t}$ represent true positives, false negatives, and false positives at timestep $t$ , computed in the 3D fine-grained format. $\widetilde{FP}_{t}$ represents partial false positives of $FP_{t}$ , which lies in the corresponding annotated bounding boxes. Note that we only use this metric for nuScenes and nuScenes-Occupancy datasets because Lyft-Level5 lacks fine-grained voxel-wise annotations. C-IoU comprehensively considers predictions within + +Table 1. Comparison of performance on 3D occupancy forecasting + +
ApproachnuScenesLyft-Level5nuScenes & nuScenes-Occupancy
2D3D2D3D3D
IoUcIoUfIoUIoUcIoUfIoUIoUcIoUfIoUIoUcIoUfIoUIoUcIoUfIoUC-IoUcC-IoUfC-IoUf
PowerBEV33.6230.0730.7728.3425.4726.0545.7843.4443.9136.1534.1834.58------
OpenOccupancy28.7725.9326.5024.2622.6923.1538.1635.1135.7932.9529.7330.4014.9913.0613.4727.1124.6825.17
OccFormer38.9831.0532.7634.7528.0829.5647.0032.1035.0241.6828.5531.0015.8011.3012.1844.5635.9837.67
OCFNet-30.3426.2727.4727.3224.4024.9836.0233.1233.6931.0428.9229.3612.6610.6511.0529.6427.4727.91
OCFNet35.5032.0232.7133.2429.3629.9445.7443.6744.0839.7138.1438.4614.7312.5713.0034.8333.1433.49
EfficientOCF-(ours)31.6929.0029.5630.2428.1728.6036.3933.2433.8333.9531.1231.7015.9814.0514.4635.8934.2034.56
EfficientOCF (ours)39.9336.1536.9335.6032.7333.3247.4244.9645.4644.2741.9042.3921.2819.0219.5047.5345.5745.99
+ +bonding boxes and fine-grained occupancy predictions, accounting for the discrepancies between the nuScenes and nuScenes-Occupancy datasets to avoid penalizing false positives in bounding boxes unfairly. This mitigates the effect of missing and incomplete occupancy annotations within bounding boxes of movable objects. Additionally, we use VPQ [13] to evaluate 3D instance forecasting lifted from 2D counterparts by height values, which is calculated by + +$$ +\mathrm {V P Q} \left(\hat {\mathcal {O}} ^ {\mathrm {3 D}}, \bar {\mathcal {O}} ^ {\mathrm {3 D}}\right) = \sum_ {t = 0} ^ {N _ {\mathrm {f}}} \frac {\sum_ {T P _ {t}} \mathrm {I o U} _ {t}}{\left| T P _ {t} \right| + \frac {1}{2} \left| F P _ {t} \right| + \frac {1}{2} \left| F N _ {t} \right|}. \tag {4} +$$ + +We set $\hat{\mathcal{O}}^{3\mathrm{D}}$ as $\hat{\mathcal{O}}^{\mathrm{bb}}$ and $\hat{\mathcal{O}}^{\mathrm{fg}}$ for inflated and fine-grained formats, denoted as $\mathrm{VPQ}^{\mathrm{bb}}$ and $\mathrm{VPQ}^{\mathrm{fg}}$ respectively. + +# 4. Experiments + +# 4.1. Experimental Setups + +Dataset details. We evaluate our proposed EfficientOCF alongside the SOTA baselines using the nuScenes [2], nuScenes-Occupancy [34] and Lyft-Level5 [11] datasets. Following existing works [15, 25, 34], we split the nuScenes and nuScenes-Occupancy datasets into training and test sets with 700 and 150 scenes respectively, and the Lyft-Level5 dataset into 130 training scenes and 50 test scenes. Each sequence has a time length $N$ of 7 frames ( $N_{\mathrm{p}} = 2$ , $N_{\mathrm{f}} = 4$ ), meaning that we use 2 past frames and 1 present observation to estimate the present occupancy states and forecast the counterparts of the next 4 frames. The nuScenes and nuScenes-Occupancy datasets are labeled at a frequency of $2\mathrm{Hz}$ , and the Lyft-Level5 dataset at $5\mathrm{Hz}$ , resulting in different forecasting time durations across these datasets. + +Implementation details. Following [25], we set the evaluation range to $[-51.2\mathrm{m}, 51.2\mathrm{m}]$ for both the x-axis and y-axis, and $[-5\mathrm{m}, 3\mathrm{m}]$ for the z-axis. The voxel resolution is set to $0.2\mathrm{m}$ , resulting in 3D occupancy voxels of size $512 \times 512 \times 40$ . We compare our EfficientOCF against four SOTA methods: PowerBEV [15], OpenOccupancy [34], OccFormer [39], and OCFNet [25]. All models are trained for 15 epochs using AdamW optimizer [14] with an initial learning rate $3 \times 10^{-4}$ and a weight decay of 0.01. Eight A100 GPUs are used for training with a batch size of 8. + +# 4.2. Occupancy Forecasting Assessment + +We first assess the occupancy forecasting performance of EfficientOCF and baseline methods, as shown in Tab. 1. The nuScenes-Occupancy dataset provides fine-grained labels for the nuScenes dataset, allowing us to conduct evaluations based on fine-grained voxels. In contrast, the Lyft-Level5 dataset only contains bounding box annotations for movable objects, and thus the evaluation on this dataset is based on inflated objects where the C-IoU metric is not available. Notably, the vanilla PowerBEV only forecasts occupancy states in 2D space. Therefore, we lift its 2D results to 3D space by assigning a fixed height value to compute the 3D metrics. This fixed height for each test scene is determined by averaging the heights of all general movable objects. For OpenOccupancy and OccFormer, we estimate the occupancy states of the present frame and then use it as the prediction of all future timesteps based on the static-world model [25]. Besides, for OpenOccupancy, OccFormer, and OCFNet, which are end-to-end 3D occupancy perception methods, we first calculate 3D metrics and then compress the 3D results to BEV space to calculate the 2D metrics. Following [25], we additionally report the performance of OCFNet and our EfficientOCF when trained on limited $1/6$ of the training sequences, denoted as OCFNet $^{-}$ and EfficientOCF $^{-}$ . Tab. 1 shows that our proposed EfficientOCF estimates the present occupancy states and forecasts future counterparts better than all the other baselines on the nuScenes, nuScenes-Occupancy, and Lyft-Level5 datasets in both 2D and 3D evaluation metrics. In particular, EfficientOCF improves the proposed C-IoU of the baselines more significantly than the other existing metrics. It is also notable that EfficientOCF $^{-}$ trained on limited data even outperforms the baseline OCFNet supervised by the holistic training set on all 3D fine-grained metrics. This validates that our method more effectively captures diverse motion patterns, even with only partial scenario perception, making it practically useful when training data is limited. + +We further report the forecasting performance with different future time horizons. As shown in Tab. 2, our EfficientOCF remains the best performance within the future time horizons $\{0.5\mathrm{s},1.0\mathrm{s},1.5\mathrm{s},2.0\mathrm{s}\}$ on the nuScenes and nuScenes-Occupancy datasets, and $\{0.2\mathrm{s},0.4\mathrm{s},0.6\mathrm{s},0.8\mathrm{s}\}$ on the Lyft-Level5 dataset. Our + +Table 2. Comparison of performance on 3D OCF in different future time horizons + +
ApproachnuScenesLyft-Level5nuScenes & nuScenes-Occupancy
IoUf (2D)IoUf (3D)IoUf (2D)IoUf (3D)IoUf (3D)C-IoUf (3D)
0.5s1.0s1.5s2.0s0.5s1.0s1.5s2.0s0.2s0.4s0.6s0.8s0.2s0.4s0.6s0.8s0.5s1.0s1.5s2.0s0.5s1.0s1.5s2.0s
PowerBEV32.1331.3130.6130.0727.1726.4925.9225.4745.0444.6444.0543.4435.5235.1834.6834.18--------
OpenOccupancy27.8927.1126.4625.9324.2023.7923.2622.6937.2436.6235.8635.1132.0331.1930.4529.7314.0813.7113.3613.0626.6225.8325.1624.68
OccFormer34.3532.5231.6031.0531.0929.4928.6228.0839.7735.6733.4832.1034.9231.5629.7428.5512.8512.0711.6111.3039.3537.5236.5635.98
OCFNet-29.1128.1127.3826.7726.3125.5124.9024.4035.0534.4633.8033.1230.3929.9729.4528.9211.6511.2510.9410.6529.0928.3927.8827.47
OCFNet34.4533.6532.8432.0231.4030.7330.0429.3645.1844.7744.2743.6739.4039.0238.5938.1413.6413.3412.9712.5734.6434.1733.6633.14
EfficientOCF-30.9030.1629.5229.0029.7429.1328.6028.1735.3534.6433.9333.2432.8832.4331.7831.1215.0714.7014.3514.0535.8635.1634.6134.20
EfficientOCF38.8037.8536.9836.1534.8134.0733.3732.7346.7046.2045.5944.9643.5443.0742.4941.9020.2119.8619.4619.0247.6346.9146.2445.57
+ +![](images/4018c459a76367314a027632f7ca7b826828f1979fc9d059029362e2466974ef.jpg) +Figure 6. Visualization of OCF results of EfficientOCF and OCFNet, and ground-truth from nuScenes (nusc), nuScenes-Occupancy (nusc-occ), and Lyft-Level5 datasets. The OCF results and ground-truth labels from timesteps 0 to $N_{\mathrm{f}}$ are assigned colors from dark to light. + +![](images/a2f034abc47480312af93ff6cd6ff0bf571b94d38173d71467d1de8a8f57781f.jpg) + +![](images/4a43ad1722483a3c8d2fe70520e3001b10cae52195f15a20afc7b349634e026e.jpg) + +![](images/6743b106ca56dc9ba35c8a2af2620852b9584a659a64a2fd9628a5247ffa9921.jpg) + +![](images/ac815a6c0efd2993c557daf83bd643085b383472a6d262f174122fd60c032a4f.jpg) + +![](images/82919a370796b9037aa3de4286366279a885a4e428233bc0b6ac9ac60f0c437a.jpg) + +temporal decoupling leads to more accurate OCF through instance forecasting associated with predicted flow at each timestep. As can be seen in Fig. 6, EfficientOCF captures better contours of movable objects (e.g., the first and the second samples) compared to the SOTA baseline OCFNet which basically overlooks shape details. This is achieved by spatial decoupling, which provides more detailed 3D structures inherent in the combination of BEV occupancy and heights. Moreover, we demonstrate that EfficientOCF can generate more TP and fewer FP predictions in 2D space by Fig. 7, which also contributes to the higher forecasting accuracy of our EfficientOCF. Besides, it can be seen from the nusc-occ ground-truth (GT) of Fig. 6 that the publicly available fine-grained GT annotations are not always accurate due to LiDAR sparsity. This reflects the need for our new metrics C-IoU. + +# 4.3. Ablation Studies + +We conduct several ablation studies to validate the effectiveness of different components in EfficientOCF. All the 3D evaluation metrics are derived by comparing fine-grained predictions with corresponding ground-truth labels. + +Spatiotemporal decoupling. We refer the ablation study on spatial decoupling to Tab. 1, where EfficientOCF outperforms OCFNet that directly implements occupancy forecasting in dense 3D representation. We further ablate tem + +![](images/3db5ca1f30a3bf59e2167de13e477707ccf2b2431ce4429309a536548351143e.jpg) +Figure 7. Visualization of TP (green), FP (red), and FN (yellow) results against GT at continuous timesteps. + +![](images/d9ab4dd03786b1bbd83ea98ed37c06fc78d4f015cccc3fb3bdf406cde48cf232.jpg) + +poral decoupling by removing the instance-aware refinement in EfficientOCF shown in Tab. 3. We conduct a baseline EfficientOCF $\ddagger$ that directly lifts the initial 2D OCF results with height values. As shown, our instance-aware temporal decoupling improves the initial OCF performance of the prediction module by absolute 1.64 and 2.55 percentage points in 3D $\tilde{\mathrm{IoU}}_{\mathrm{f}}$ and C- $\tilde{\mathrm{IoU}}_{\mathrm{f}}$ on the nuScenes-Occupancy dataset. This demonstrates that instance forecasting encompassing flow-based association provides the more accurate occupancy estimation for each decoupled timestep. + +Adaptive dual pooling. We compare our adaptive dual pooling (ADP) to different pooling strategies that compress + +Table 3. Ablation on temporal decoupling + +
Approach2D3D
IoUcIoUfIoUfIoUcIoUfIoUfC-IoUcC-IoUfC-IoUf
EfficientOCF‡39.4435.8936.6119.9817.3117.8645.5242.9043.44
EfficientOCF39.9336.1536.9321.2819.0219.5047.5345.5745.99
+ +Table 4. Ablation on adaptive dual pooling + +
Approach2D3D
IoUcIoUfIoUfIoUcIoUfIoUfC-IoUcC-IoUfC-IoUf
Average Pooling38.0434.7935.4519.2916.7717.2944.0241.7142.18
Max Pooling38.8435.5336.2019.4216.9017.4144.6942.3342.81
ADP39.9336.1536.9321.2819.0219.5047.5345.5745.99
+ +Table 5. Ablation on segmentation/flow/height heads + +
Seg. +HeadHeight +HeadFlow +Head2D3D
IoUcIoUfIoUfIoUcIoUfIoUfC-IoUcC-IoUfC-IoUf
✓38.4835.2635.93------
✓✓38.5535.3335.9819.4116.8917.4144.0941.8342.29
✓✓✓39.9336.1536.9321.2819.0219.5047.5345.5745.99
+ +3D features to 2D space. Tab. 4 shows that EfficientOCF with adaptive weighting for the combination of average and max pooling outperforms the baseline with only one single pooling strategy for 3D-2D transformation. The max pooling contributes more to forecasting accuracy than the average pooling, likely because height values and max-pooled features both capture prominent structural changes along the z-axis. We also deploy our ADP strategy as the 3D-2D transformation module in PowerBEV, and it enhances the performance of the vanilla PowerBEV in 2D $\mathrm{IoU_c}$ and $\mathrm{IoU_f}$ (2.0s) by absolute 5.16 and 5.42 percentage points. This validates that other OCF methods can also benefit from our proposed ADP strategy. + +Multiple heads. We ablate the devised heads in the prediction module of EfficientOCF in Tab. 5. The inclusion of the height head allows the transformation of forecasting results from 2D to 3D space, and provides more structure-aware supervision signals. It thus significantly improves OCF performance on all the metrics. Furthermore, the introduction of the flow head also provides additional enhancements because flow prediction enables the model to be aware of the concrete motion pattern of each voxel grid. Besides, the flow head helps to refine the initial occupancy following our proposed temporal decoupling. + +# 4.4. Instance Prediction Assessment + +This experiment presents the 3D instance prediction performance of SOTA OCFNet and our EfficientOCF with the above-mentioned VPQ metric. As shown in Tab. 6, EfficientOCF improves the VPQ of OCFNet forecasting by absolute 1.42 and 4.08 percentage points in $\mathrm{VPQ^{bb}}$ and $\mathrm{VPQ^{fg}}$ respectively on nuScenes, and by 8.40 percentage points on Lyft-Level5. EfficientOCF $^{-}$ still outperforms OCFNet $^{-}$ for both datasets. EfficientOCF forecasts instances in 2D space which are further lifted to 3D counterparts following our proposed spatiotemporal decoupling. OCFNet directly im + +Table 6. Evaluation on 3D instance prediction + +
ApproachnuScenesLyft-Level5
VPQbbVPQfgVPQBb
OCFNet-13.684.2524.11
OCFNet18.536.6128.35
EfficientOCF-(ours)14.116.5027.06
EfficientOCF(ours)19.9510.6936.75
+ +Table 7. Comparison of inference time + +
ApproachInference time [ms]
PerceptionProcessingPredictionTotal
OCFNet107.054.1813.15124.38
EfficientOCF (ours)64.785.0512.5082.33
+ +plements instance association in dense 3D formats, leading to more prediction artifacts and uncertainties. + +# 4.5. Inference Time + +We compare the runtime of OCFNet and our EfficientOCF on 3D OCF shown in Tab. 7. As seen, the processing time of EfficientOCF is slightly higher than OCFNet due to the additional 3D-2D transformation, while EfficientOCF achieves a faster overall inference speed $(82.33\mathrm{ms} / 12.15\mathrm{Hz})$ compared to OCFNet. Moreover, we find that EfficientOCF reduces memory consumption by $478\mathrm{MB}$ compared to OCFNet. This is because EfficientOCF utilizes lightweight 2D encoders and decoders in the perception module to achieve 3D occupancy forecasting by lifting 2D results to 3D space using height values. In contrast, OCFNet uses heavy 3D encoders and decoders with dense 3D representation, leading to more inference time. + +# 5. Conclusion + +This paper proposes spatiotemporal decoupling for efficient vision-based occupancy forecasting (OCF). We achieve spatial decoupling by transforming 3D dense representations into compact 2D BEV occupancy with height values. We also propose using temporal decoupling to refine the end-to-end OCF by step-wise instance forecasting along the time axis. Based on our proposed spatiotemporally decoupled representation, we further develop EfficientOCF, a lightweight network that concurrently forecasts future 2D occupancy states and height values of general movable objects, effectively bridging 2D predictions into the 3D space. It also refines initial OCF results with instance forecasting at each individual timestep. Additionally, we introduce new evaluation metrics providing a comprehensive assessment of OCF performance for datasets with incomplete annotations. Our experiments on extensive evaluation metrics demonstrate that EfficientOCF outperforms existing methods in both accuracy and efficiency, achieving SOTA results in both 2D and 3D spaces. We hope that the spatiotemporal decoupling pipeline proposed in this work can be used as a foundation paradigm in the field of vision-based OCF. + +# References + +[1] Adil Kaan Akan and Fatma Güney. Stretchbev: Stretching future instance prediction spatially and temporally. In European Conference on Computer Vision, pages 444-460. Springer, 2022. 1, 2 +[2] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11621-11631, 2020. 2, 3, 5, 6 +[3] Anh-Quan Cao and Raoul De Charette. Monoscene: Monocular 3d semantic scene completion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3991–4001, 2022. 1, 2 +[4] Sergio Casas, Abbas Sadat, and Raquel Urtasun. Mp3: A unified model to map, perceive, predict and plan. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14403-14412, 2021. 3 +[5] Hehe Fan and Yi Yang. Point recurrent neural network for moving point cloud processing. arXiv preprint arXiv:1910.08287, 2019. 3 +[6] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 3 +[7] Yu Hong, Qian Liu, Huayuan Cheng, Danjiao Ma, Hang Dai, Yu Wang, Guangzhi Cao, and Yong Ding. Univision: A unified framework for vision-centric 3d perception. arXiv preprint arXiv:2401.06994, 2024. 1 +[8] Anthony Hu, Zak Murez, Nikhil Mohan, Sofia Dudas, Jeffrey Hawke, Vijay Badrinarayanan, Roberto Cipolla, and Alex Kendall. Fiery: Future instance prediction in bird's-eye view from surround monocular cameras. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15273-15282, 2021. 1, 2 +[9] Yuanhui Huang, Wenzhao Zheng, Yunpeng Zhang, Jie Zhou, and Jiwen Lu. Tri-perspective view for vision-based 3d semantic occupancy prediction. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9223-9232, 2023. 1, 2 +[10] Yuanhui Huang, Wenzhao Zheng, Borui Zhang, Jie Zhou, and Jiwen Lu. Selfocc: Self-supervised vision-based 3d occupancy prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19946-19956, 2024. 1, 2 +[11] R. Kesten, M. Usman, J. Houston, T. Pandya, K. Nadhamuni, A. Ferreira, M. Yuan, B. Low, A. Jain, P. Ondruska, S. Omari, S. Shah, A. Kulkarni, A. Kazakova, C. Tao, L. Platinsky, W. Jiang, and V. Shet. Lyft level 5 perception dataset 2020, 2019, 2, 5, 6 +[12] Tarasha Khurana, Peiyun Hu, David Held, and Deva Ramanan. Point cloud forecasting as a proxy for 4d occupancy forecasting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1116-1124, 2023. 1 + +[13] Dahun Kim, Sanghyun Woo, Joon-Young Lee, and In So Kweon. Video panoptic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9859–9868, 2020. 5, 6 +[14] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6 +[15] Peizheng Li, Shuxiao Ding, Xieyuanli Chen, Niklas Hanselmann, Marius Cordts, and Juergen Gall. Powerbev: a powerful yet lightweight framework for instance prediction in bird's-eye view. In Proceedings of the Thirty-Second International Joint Conference on Artificial Intelligence, pages 1080–1088, 2023. 1, 3, 4, 6 +[16] Yinhao Li, Zheng Ge, Guanyi Yu, Jinrong Yang, Zengran Wang, Yukang Shi, Jianjian Sun, and Zeming Li. Bevdepth: Acquisition of reliable depth for multi-view 3d object detection. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1477-1485, 2023. 2 +[17] Yiming Li, Zhiding Yu, Christopher Choy, Chaowei Xiao, Jose M Alvarez, Sanja Fidler, Chen Feng, and Anima Anandkumar. Voxformer: Sparse voxel transformer for camera-based 3d semantic scene completion. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9087-9098, 2023. 1 +[18] Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chonghao Sima, Tong Lu, Yu Qiao, and Jifeng Dai. Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In European conference on computer vision, pages 1-18. Springer, 2022. 2 +[19] Haochen Liu, Zhiyu Huang, and Chen Lv. Multi-modal hierarchical transformer for occupancy flow field prediction in autonomous driving. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 1449-1455. IEEE, 2023. 2 +[20] Jian Liu, Sipeng Zhang, Chuixin Kong, Wenyuan Zhang, Yuhang Wu, Yikang Ding, Borun Xu, Ruibo Ming, Donglai Wei, and Xianming Liu. Occtransformer: Improving bevformer for 3d camera-only occupancy prediction. arXiv preprint arXiv:2402.18140, 2024. 2 +[21] Xinhao Liu, Moonjun Gong, Qi Fang, Haoyu Xie, Yiming Li, Hang Zhao, and Chen Feng. Lidar-based 4d occupancy completion and forecasting. arXiv preprint arXiv:2310.11239, 2023. 2 +[22] Zhijian Liu, Haotian Tang, Alexander Amini, Xinyu Yang, Huizi Mao, Daniela L Rus, and Song Han. Bevfusion: Multi-task multi-sensor fusion with unified bird's-eye view representation. In 2023 IEEE international conference on robotics and automation (ICRA), pages 2774-2781. IEEE, 2023. 2 +[23] Fan Lu, Guang Chen, Zhijun Li, Lijun Zhang, Yinlong Liu, Sanqing Qu, and Alois Knoll. Monet: Motion-based point cloud prediction network. IEEE Transactions on Intelligent Transportation Systems, 23(8):13794-13804, 2021. 3 +[24] Zhen Luo, Junyi Ma, Zijie Zhou, and Guangming Xiong. Pcpcnet: An efficient and semantic-enhanced transformer network for point cloud prediction. IEEE Robotics and Automation Letters, 8(7):4267-4274, 2023. 3 +[25] Junyi Ma, Xieyuanli Chen, Jiawei Huang, Jingyi Xu, Zhen Luo, Jintao Xu, Weihao Gu, Rui Ai, and Hesheng Wang. + +Cam4docc: Benchmark for camera-only 4d occupancy forecasting in autonomous driving applications. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21486-21495, 2024. 1, 2, 3, 4, 5, 6 +[26] Reza Mahjourian, Jinkyu Kim, Yuning Chai, Mingxing Tan, Ben Sapp, and Dragomir Anguelov. Occupancy flow fields for motion forecasting in autonomous driving. IEEE Robotics and Automation Letters, 7(2):5639-5646, 2022. 2 +[27] Benedikt Mersch, Xieyuanli Chen, Jens Behley, and Cyril Stachniss. Self-supervised point cloud prediction using 3d spatio-temporal convolutional networks. In Conference on Robot Learning, pages 1444-1454. PMLR, 2022. 3 +[28] Kaustab Pal, Aditya Sharma, Avinash Sharma, and K Madhava Krishna. Atppnet: Attention based temporal point cloud prediction network. arXiv preprint arXiv:2401.17399, 2024. 3 +[29] Mingjie Pan, Li Liu, Jiaming Liu, Peixiang Huang, Longlong Wang, Shanghai Zhang, Shaoqing Xu, Zhiyi Lai, and Kuiyuan Yang. Uniocc: Unifying vision-centric 3d occupancy prediction with geometric and semantic rendering. arXiv preprint arXiv:2306.09117, 2023. 2 +[30] Jonah Philion and Sanja Fidler. Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIV 16, pages 194-210. Springer, 2020. 3 +[31] Xiaoyu Tian, Tao Jiang, Longfei Yun, Yucheng Mao, Huitong Yang, Yue Wang, Yilun Wang, and Hang Zhao. Occ3d: A large-scale 3d occupancy prediction benchmark for autonomous driving. Advances in Neural Information Processing Systems, 36, 2024. 2 +[32] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 2 +[33] Lening Wang, Wenzhao Zheng, Yilong Ren, Han Jiang, Zhiyong Cui, Haiyang Yu, and Jiwen Lu. Occsora: 4d occupancy generation models as world simulators for autonomous driving. arXiv preprint arXiv:2405.20337, 2024. 3 +[34] Xiaofeng Wang, Zheng Zhu, Wenbo Xu, Yunpeng Zhang, Yi Wei, Xu Chi, Yun Ye, Dalong Du, Jiwen Lu, and Xingang Wang. Openoccupancy: A large scale benchmark for surrounding semantic occupancy perception. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17850-17859, 2023. 1, 2, 5, 6 +[35] Yi Wei, Linqing Zhao, Wenzhao Zheng, Zheng Zhu, Jie Zhou, and Jiwen Lu. Surroundocc: Multi-camera 3d occupancy prediction for autonomous driving. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 21729-21740, 2023. 2 +[36] Pengxiang Wu, Siheng Chen, and Dimitris N Metaxas. Motionnet: Joint perception and motion prediction for autonomous driving based on bird's eye view maps. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11385-11395, 2020. 3 +[37] Qizhe Xie, Minh-Thang Luong, Eduard Hovy, and Quoc V Le. Self-training with noisy student improvesImagenet + +classification. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10687-10698, 2020. 5 +[38] Yunpeng Zhang, Zheng Zhu, Wenzhao Zheng, Junjie Huang, Guan Huang, Jie Zhou, and Jiwen Lu. Beverse: Unified perception and prediction in birds-eye-view for vision-centric autonomous driving. arXiv preprint arXiv:2205.09743, 2022.1,2 +[39] Yunpeng Zhang, Zheng Zhu, and Dalong Du. Occformer: Dual-path transformer for vision-based 3d semantic occupancy prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9433-9443, 2023. 1, 2, 6 +[40] Wenzhao Zheng, Weiliang Chen, Yuanhui Huang, Borui Zhang, Yueqi Duan, and Jiwen Lu. Occworld: Learning a 3d occupancy world model for autonomous driving. arXiv preprint arXiv:2311.16038, 2023. 3 \ No newline at end of file diff --git a/2025/Spatiotemporal Decoupling for Efficient Vision-Based Occupancy Forecasting/images.zip b/2025/Spatiotemporal Decoupling for Efficient Vision-Based Occupancy Forecasting/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..69c1d019e68aa95a122ba0b951e44c63a01640d8 --- /dev/null +++ b/2025/Spatiotemporal Decoupling for Efficient Vision-Based Occupancy Forecasting/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d8ff688c18a2ba344df65c4fb60661c1c4bcfdb8488953bf83d44f150c468081 +size 644728 diff --git a/2025/Spatiotemporal Decoupling for Efficient Vision-Based Occupancy Forecasting/layout.json b/2025/Spatiotemporal Decoupling for Efficient Vision-Based Occupancy Forecasting/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..cd6e20d5a3654ee1c9229e8d6af1aab1b9b16c20 --- /dev/null +++ b/2025/Spatiotemporal Decoupling for Efficient Vision-Based Occupancy Forecasting/layout.json @@ -0,0 +1,8423 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 57, + 103, + 536, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 103, + 536, + 121 + ], + "spans": [ + { + "bbox": [ + 57, + 103, + 536, + 121 + ], + "type": "text", + "content": "Spatiotemporal Decoupling for Efficient Vision-Based Occupancy Forecasting" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 142, + 547, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 142, + 547, + 215 + ], + "spans": [ + { + "bbox": [ + 53, + 142, + 547, + 215 + ], + "type": "text", + "content": "Jingyi Xu" + }, + { + "bbox": [ + 53, + 142, + 547, + 215 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 53, + 142, + 547, + 215 + ], + "type": "text", + "content": " Xieyuanli Chen" + }, + { + "bbox": [ + 53, + 142, + 547, + 215 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 53, + 142, + 547, + 215 + ], + "type": "text", + "content": " Junyi Ma" + }, + { + "bbox": [ + 53, + 142, + 547, + 215 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 53, + 142, + 547, + 215 + ], + "type": "text", + "content": " Jiawei Huang" + }, + { + "bbox": [ + 53, + 142, + 547, + 215 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 53, + 142, + 547, + 215 + ], + "type": "text", + "content": " Jintao Xu" + }, + { + "bbox": [ + 53, + 142, + 547, + 215 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 53, + 142, + 547, + 215 + ], + "type": "text", + "content": " Yue Wang" + }, + { + "bbox": [ + 53, + 142, + 547, + 215 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 53, + 142, + 547, + 215 + ], + "type": "text", + "content": " Ling Pei" + }, + { + "bbox": [ + 53, + 142, + 547, + 215 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 53, + 142, + 547, + 215 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 53, + 142, + 547, + 215 + ], + "type": "text", + "content": "Shanghai Key Laboratory of Navigation and Location Based Services, Shanghai Jiao Tong University \n" + }, + { + "bbox": [ + 53, + 142, + 547, + 215 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 53, + 142, + 547, + 215 + ], + "type": "text", + "content": "College of Intelligence Science and Technology, National University of Defense Technology \n" + }, + { + "bbox": [ + 53, + 142, + 547, + 215 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 53, + 142, + 547, + 215 + ], + "type": "text", + "content": "Department of Automation, Shanghai Jiao Tong University " + }, + { + "bbox": [ + 53, + 142, + 547, + 215 + ], + "type": "inline_equation", + "content": "{}^{4}" + }, + { + "bbox": [ + 53, + 142, + 547, + 215 + ], + "type": "text", + "content": "HAOMO.AI \n" + }, + { + "bbox": [ + 53, + 142, + 547, + 215 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 53, + 142, + 547, + 215 + ], + "type": "text", + "content": "State Key Laboratory of Industrial Control and Technology, Zhejiang University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 240, + 191, + 253 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 240, + 191, + 253 + ], + "spans": [ + { + "bbox": [ + 143, + 240, + 191, + 253 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 272, + 289, + 631 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 272, + 289, + 631 + ], + "spans": [ + { + "bbox": [ + 47, + 272, + 289, + 631 + ], + "type": "text", + "content": "The task of occupancy forecasting (OCF) involves utilizing past and present perception data to predict future occupancy states of autonomous vehicle surrounding environments, which is critical for downstream tasks such as obstacle avoidance and path planning. Existing 3D OCF approaches struggle to predict plausible spatial details for movable objects and suffer from slow inference speeds due to neglecting the bias and uneven distribution of changing occupancy states in both space and time. In this paper, we propose a novel spatiotemporal decoupling vision-based paradigm to explicitly tackle the bias and achieve both effective and efficient 3D OCF. To tackle spatial bias in empty areas, we introduce a novel spatial representation that decouples the conventional dense 3D format into 2D bird's-eye view (BEV) occupancy with corresponding height values, enabling 3D OCF derived only from 2D predictions thus enhancing efficiency. To reduce temporal bias on static voxels, we design temporal decoupling to improve end-to-end OCF by temporally associating instances via predicted flows. We develop an efficient multi-head network EfficientOCF to achieve 3D OCF with our devised spatiotemporally decoupled representation. A new metric, conditional IoU (C-IoU), is also introduced to provide a robust 3D OCF performance assessment, especially in datasets with missing or incomplete annotations. The experimental results demonstrate that EfficientOCF surpasses existing baseline methods on accuracy and efficiency, achieving state-of-the-art performance with a fast inference time of 82.33 ms with a single GPU. Our code is released at: https://github.com/BIT-XJY/EfficientOCF." + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 308, + 240, + 545, + 422 + ], + "blocks": [ + { + "bbox": [ + 308, + 240, + 545, + 422 + ], + "lines": [ + { + "bbox": [ + 308, + 240, + 545, + 422 + ], + "spans": [ + { + "bbox": [ + 308, + 240, + 545, + 422 + ], + "type": "image", + "image_path": "3a01581574a1a087b3e7055fc58d3f8f03b290257fd43593599f65a257d68a67.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 422, + 545, + 467 + ], + "lines": [ + { + "bbox": [ + 306, + 422, + 545, + 467 + ], + "spans": [ + { + "bbox": [ + 306, + 422, + 545, + 467 + ], + "type": "text", + "content": "Figure 1. Unlike the traditional 3D vision-based OCF pipeline, our proposed paradigm implements spatiotemporal decoupling, thus achieving the efficiency of 2D OCF methods with the instance-aware capability of 3D OCF approaches." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 307, + 474, + 386, + 486 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 474, + 386, + 486 + ], + "spans": [ + { + "bbox": [ + 307, + 474, + 386, + 486 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 496, + 545, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 496, + 545, + 663 + ], + "spans": [ + { + "bbox": [ + 304, + 496, + 545, + 663 + ], + "type": "text", + "content": "Accurately estimating occupancy states in the surrounding environment based on visual inputs is crucial for autonomous mobile systems, as it enables the identification of general obstacles, including objects absent from semantic databases. Most existing vision-based occupancy estimation methods [3, 7, 9, 10, 17, 34, 39] rely on complex model architectures but often overlook temporal changes in occupancy states. To address this, occupancy forecasting (OCF) approaches [1, 8, 12, 15, 25, 38] incorporate the prediction of surrounding object movements over time. These methods use past observations to forecast occupancy states in the present and near future, providing valuable information in advance for downstream tasks such as planning and control in autonomous driving." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 545, + 713 + ], + "type": "text", + "content": "Existing vision-based OCF methods predict future occupancy states for the entire 3D scene, as illustrated in Fig. 1(a). These methods process all voxel grids in each scene through end-to-end 3D feature encoding and decod" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 35 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 35 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 35 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 674, + 124, + 684 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 674, + 124, + 684 + ], + "spans": [ + { + "bbox": [ + 58, + 674, + 124, + 684 + ], + "type": "text", + "content": "*Equal contribution" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 59, + 684, + 135, + 693 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 684, + 135, + 693 + ], + "spans": [ + { + "bbox": [ + 59, + 684, + 135, + 693 + ], + "type": "text", + "content": "† Corresponding author" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 694, + 287, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 694, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 694, + 287, + 712 + ], + "type": "text", + "content": "This work was funded by Science and Technology Commission of Shanghai Municipality (24DZ3101300, 24TS1402600, 24TS1402800)." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22338" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 287, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 287, + 179 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 287, + 179 + ], + "type": "text", + "content": "ing [21, 25]. However, this indiscriminate dense representation often leads to inefficient and less accurate OCF due to two main biases: (1) a spatial bias toward empty voxels, which dominate the 3D space; (2) a temporal changing bias toward static objects, as only a small subset of movable objects change positions in near future. These biases not only result in high computational costs and wasteful resource use but also increase uncertainty in predicting movable object occupancy, leading to rapid error accumulation in forecasts." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 180, + 289, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 180, + 289, + 528 + ], + "spans": [ + { + "bbox": [ + 46, + 180, + 289, + 528 + ], + "type": "text", + "content": "To tackle these issues, we propose a novel spatiotemporal decoupling representation and an efficient network EfficientOCF for vision-based OCF. As shown in Fig. 1(b), it decouples 3D OCF both spatially and temporally. We design a new spatial representation that decouples bird's-eye view (BEV) occupancy with height values. By storing only heights for 2D occupied grids, it achieves 3D OCF while avoiding the traditional memory-intensive 3D occupancy representation [20, 29, 34, 35]. For temporal changing forecasting, we decouple current instance segmentation from future occupancy estimation, leveraging instance movement understanding to enhance OCF rather than relying on end-to-end forecasting. To achieve these goals, we design EfficientOCF, a novel network with flow, segmentation, and height prediction heads. This network compresses 3D voxel features into 2D BEV features and employs lightweight 2D encoder-decoder structures to predict BEV occupancy. Simultaneously, it predicts 2D backward centripetal flow and height values for grids containing movable objects at each timestep. Flow vectors establish instance associations between current observations and future forecasts, refining OCF results at the instance level for enhanced accuracy and efficiency. The final 3D OCF is obtained by lifting refined 2D OCF with height estimation. We further introduce new metrics and conduct extensive experiments on the nuScenes [2], nuScenes-Occupancy [34], and Lyft-Level5 [11] datasets to demonstrate that our EfficientOCF generates high-quality OCF results and achieves superior real-time performance compared to the SOTA." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 529, + 286, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 529, + 286, + 541 + ], + "spans": [ + { + "bbox": [ + 58, + 529, + 286, + 541 + ], + "type": "text", + "content": "In summary, the contributions of our work are threefold:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 546, + 286, + 712 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 47, + 546, + 286, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 546, + 286, + 616 + ], + "spans": [ + { + "bbox": [ + 47, + 546, + 286, + 616 + ], + "type": "text", + "content": "- We propose a spatiotemporal decoupling framework for effective and efficient vision-based 3D OCF. Spatial decoupling is designed with a representation that combines BEV occupancy with height values to avoid 3D voxel redundancy. Temporal decoupling is achieved via instance-aware associations, obtaining enhanced 3D OCF results." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 618, + 286, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 618, + 286, + 712 + ], + "spans": [ + { + "bbox": [ + 47, + 618, + 286, + 712 + ], + "type": "text", + "content": "- We introduce EfficientOCF, a novel efficient vision-based 3D OCF network, that predicts the future occupancy states of movable objects using our proposed spatiotemporally decoupled representation. EfficientOCF employs lightweight 2D encoder-decoder structures rather than 3D counterparts and incorporates an adaptive dual pooling strategy for 3D-2D feature transformation, achieving SOTA 3D OCF performance and efficiency." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 72, + 545, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 72, + 545, + 156 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 545, + 156 + ], + "type": "text", + "content": "- We propose new evaluation metrics tailored for OCF on fine-grained general movable objects. These metrics address the shortcomings of traditional intersection over union (IoU) by minimizing penalties for false positives within the bounding box of movable objects, which enhances the rationality of assessments, especially in cases of missing or incomplete occupancy annotations." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 167, + 392, + 179 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 167, + 392, + 179 + ], + "spans": [ + { + "bbox": [ + 306, + 167, + 392, + 179 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 186, + 545, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 186, + 545, + 569 + ], + "spans": [ + { + "bbox": [ + 304, + 186, + 545, + 569 + ], + "type": "text", + "content": "3D occupancy prediction. 3D occupancy prediction aims to estimate whether each voxel in the surrounding environment is currently occupied, representing general obstacles and enhancing the expressiveness of complex scenes. MonoScene, proposed by Cao et al. [3], first introduces a 3D semantic scene completion framework using only camera inputs, but it primarily focuses on front-view scenarios. To address this, TPVFormer by Huang et al. [9] proposes a tri-perspective view (TPV) representation to lift image features, achieving comparable performance with other LiDAR methods. Building on the TPV representation, they further propose SelfOcc [10], employing a self-supervised scheme to learn 3D occupancy states. Similar to TPVFormer, transformer modules [32] are also utilized in Occ3D [31], OccFormer [39], and SurroundOcc [35]. Occ3D [31] proposes CTF-Occ, a novel transformer-based course-to-fine occupancy network, achieving 3D occupancy predictions with high spatial resolution. OccFormer [39] proposes a dual-path transformer network to fully integrate 2D global features and 3D local features, while SurroundOcc proposed by Wei et al. [35] adopts spatial 2D-3D attention to lift images to the 3D volume space. Wang et al. [34] propose OpenOccupancy, the first surrounding semantic occupancy perception benchmark, and they introduce nuScenes-Occupancy dataset extending the large-scale nuScenes dataset [2] with fine-grained occupancy annotations. These approaches consider the complex challenge of inferring the spatial distribution of obstacles but ignore their temporal evolution. However, understanding the detailed and dynamic environment is crucial for planning and decision-making in autonomous driving, ensuring robust and reliable performance under various driving conditions." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 570, + 545, + 713 + ], + "type": "text", + "content": "BEV occupancy forecasting. BEV representation has been widely used in autonomous mobile systems due to its lightweight 2D format [16, 18, 19, 22, 26]. To overcome the challenge of capturing the temporal evolution in 2D format, some occupancy perception methods are proposed to achieve occupancy forecasting and instance prediction in BEV. FIERY, proposed by Hu et al. [8], is the first to use surround-view camera inputs to directly predict future instance segmentation and the motion of dynamic agents using BEV representation. Based on FIERY, StretchBEV by Akan et al. [1] incorporates a stochastic temporal model to refine residual updates. Zhang et al. [38] propose BE" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22339" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 251 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 251 + ], + "type": "text", + "content": "Verse, which also uses BEV occupancy to jointly reason about multiple tasks, including object detection, semantic map construction, and motion prediction. Similarly, MotionNet by Wu et al. [36] encodes object category and motion information in each BEV grid to predict scene flow and detect objects. Compared to MotionNet, MP3 proposed by Casas et al. [4] enhances occupancy flow parameterization with probabilistic modeling for multi-modal behavior. In contrast, PowerBEV by Li et al. [15] relies sorely on a semantic segmentation head alongside a backward centripetal flow head to achieve future instance prediction. Despite their effectiveness, these BEV-based methods lack a comprehensive understanding of spatial structures along the z-axis, which is crucial for accurately capturing the full 3D states of obstacles." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 252, + 289, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 252, + 289, + 504 + ], + "spans": [ + { + "bbox": [ + 46, + 252, + 289, + 504 + ], + "type": "text", + "content": "3D occupancy forecasting. Recently, researchers have increasingly explored end-to-end occupancy forecasting in 3D space. As depicted in the previous work [25], occupancy states can be directly generated from point cloud prediction [5, 23, 24, 27, 28] by trivial voxelization, but cannot achieve acceptable occupancy forecasting accuracy. OccWorld [40] and OccSora [33] use generation models to simultaneously predict the movement of the ego car and the evolution of the surrounding scenes. Ma et al. [25] introduce Cam4DOcc, the first benchmark to standardize evaluation protocol of vision-only occupancy forecasting in 3D space, and propose OCFNet, an end-to-end network to forecast dense 3D occupancy states. Compared to these 3D occupancy forecasting methods, in this work, we decouple spatiotemporal representation, where BEV occupancy with height values is utilized for spatial decoupling and instance-aware refinement in consecutive frames is designed for temporal decoupling. Based on the spatiotemporally decoupled representation, we propose a novel lightweight network EfficientOCF. Our experiments show that our proposed method achieves the SOTA performance with fast inference speed." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 47, + 517, + 153, + 529 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 517, + 153, + 529 + ], + "spans": [ + { + "bbox": [ + 47, + 517, + 153, + 529 + ], + "type": "text", + "content": "3. Proposed Method" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 536, + 141, + 548 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 536, + 141, + 548 + ], + "spans": [ + { + "bbox": [ + 47, + 536, + 141, + 548 + ], + "type": "text", + "content": "3.1. Task Definition" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 555, + 287, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 555, + 287, + 664 + ], + "spans": [ + { + "bbox": [ + 47, + 555, + 287, + 664 + ], + "type": "text", + "content": "Given " + }, + { + "bbox": [ + 47, + 555, + 287, + 664 + ], + "type": "inline_equation", + "content": "N_{\\mathrm{p}}" + }, + { + "bbox": [ + 47, + 555, + 287, + 664 + ], + "type": "text", + "content": " past and present surrounding visual images " + }, + { + "bbox": [ + 47, + 555, + 287, + 664 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{I}|I_t\\in \\mathbb{R}^{N_{\\mathrm{cam}}\\times 3\\times H_{\\mathrm{v}}\\times W_{\\mathrm{v}}}\\}" + }, + { + "bbox": [ + 47, + 555, + 287, + 664 + ], + "type": "inline_equation", + "content": "t\\in [-N_{\\mathrm{p}},0]" + }, + { + "bbox": [ + 47, + 555, + 287, + 664 + ], + "type": "text", + "content": " as input, the vision-based OCF task aims to predict the occupancy states (occupied or free) of general movable objects in the present frame and " + }, + { + "bbox": [ + 47, + 555, + 287, + 664 + ], + "type": "inline_equation", + "content": "N_{\\mathrm{f}}" + }, + { + "bbox": [ + 47, + 555, + 287, + 664 + ], + "type": "text", + "content": " future frames " + }, + { + "bbox": [ + 47, + 555, + 287, + 664 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{O}^{\\mathrm{3D}}|O_t^{\\mathrm{3D}}\\in \\mathbb{R}^{1\\times H\\times W\\times L}\\}" + }, + { + "bbox": [ + 47, + 555, + 287, + 664 + ], + "type": "inline_equation", + "content": "t\\in [0,N_{\\mathrm{f}}]" + }, + { + "bbox": [ + 47, + 555, + 287, + 664 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 47, + 555, + 287, + 664 + ], + "type": "inline_equation", + "content": "N_{\\mathrm{cam}}" + }, + { + "bbox": [ + 47, + 555, + 287, + 664 + ], + "type": "text", + "content": " represents surrounding camera numbers. " + }, + { + "bbox": [ + 47, + 555, + 287, + 664 + ], + "type": "inline_equation", + "content": "H_{\\mathrm{v}}" + }, + { + "bbox": [ + 47, + 555, + 287, + 664 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 555, + 287, + 664 + ], + "type": "inline_equation", + "content": "W_{\\mathrm{v}}" + }, + { + "bbox": [ + 47, + 555, + 287, + 664 + ], + "type": "text", + "content": " represent the height and width of input images, and " + }, + { + "bbox": [ + 47, + 555, + 287, + 664 + ], + "type": "inline_equation", + "content": "H,W,L" + }, + { + "bbox": [ + 47, + 555, + 287, + 664 + ], + "type": "text", + "content": " represent the height, width, and length of the specific range defined in the present " + }, + { + "bbox": [ + 47, + 555, + 287, + 664 + ], + "type": "inline_equation", + "content": "(t = 0)" + }, + { + "bbox": [ + 47, + 555, + 287, + 664 + ], + "type": "text", + "content": " coordinate system." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 664, + 287, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 664, + 287, + 715 + ], + "spans": [ + { + "bbox": [ + 47, + 664, + 287, + 715 + ], + "type": "text", + "content": "Compared to traditional prediction methods, our work first spatially decouples OCF in 2D BEV occupancy " + }, + { + "bbox": [ + 47, + 664, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{O}^{2\\mathrm{D}}|O_t^{2\\mathrm{D}}\\in \\mathbb{R}^{1\\times H\\times W}\\}" + }, + { + "bbox": [ + 47, + 664, + 287, + 715 + ], + "type": "text", + "content": " and corresponding height prediction " + }, + { + "bbox": [ + 47, + 664, + 287, + 715 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{O}^{\\mathrm{height}}|O_t^{\\mathrm{height}}\\in \\mathbb{R}^{1\\times H\\times W}\\}" + }, + { + "bbox": [ + 47, + 664, + 287, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 47, + 664, + 287, + 715 + ], + "type": "inline_equation", + "content": "t\\in [0,N_{\\mathrm{f}}]" + }, + { + "bbox": [ + 47, + 664, + 287, + 715 + ], + "type": "text", + "content": ". It pre" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 72, + 547, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 240 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 240 + ], + "type": "text", + "content": "dicts the future height values of each voxel column and then lifts " + }, + { + "bbox": [ + 304, + 72, + 547, + 240 + ], + "type": "inline_equation", + "content": "\\mathcal{O}^{2\\mathrm{D}}" + }, + { + "bbox": [ + 304, + 72, + 547, + 240 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 304, + 72, + 547, + 240 + ], + "type": "inline_equation", + "content": "\\mathcal{O}^{3\\mathrm{D}}" + }, + { + "bbox": [ + 304, + 72, + 547, + 240 + ], + "type": "text", + "content": ". Unlike static objects that remain unchanged over short periods, general movable objects exhibit more dynamic motion characteristics. Therefore, following [25], we focus on forecasting the occupancy states of general movable objects, e.g., bicycles, buses, cars, etc., in the nuScenes dataset [2]. This emphasis on movable objects enhances the relevance and applicability of the forecasting task to real-world autonomous driving scenarios, where accurately predicting the motion of these objects is vital for both safety and efficiency. While existing 3D OCF methods [25] forecast movable objects end-to-end, we propose temporal decoupling to propagate current instance segmentation through flow association and achieve refined 3D OCF." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 306, + 250, + 432, + 262 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 250, + 432, + 262 + ], + "spans": [ + { + "bbox": [ + 306, + 250, + 432, + 262 + ], + "type": "text", + "content": "3.2. Network Architecture" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 269, + 547, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 269, + 547, + 473 + ], + "spans": [ + { + "bbox": [ + 304, + 269, + 547, + 473 + ], + "type": "text", + "content": "The pipeline of our EfficientOCF is shown in Fig. 2, which consists of four main components: a perception module, an aggregation module, a prediction module, and a refinement module. The perception module extracts features from the input surrounding images and lifts them from 2D to 3D space following Lift-Splat-Shoot [30] to obtain 3D spatial voxel features. The aggregation module then aggregates voxel features from sequential frames and transforms them into 2D BEV features, which are fed into the following prediction module. Next, the prediction module, composed of multiple 2D branches, decodes the aggregated 2D BEV features to initial 2D OCF, backward centripetal flow, and height values respectively. Then, the refinement module propagates instance segmentation into each future timestep, which is further used to refine initial OCF results in 2D BEV. Ultimately, we use height values predicted by the prediction module to lift the refined 2D occupancy to 3D space." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 474, + 547, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 474, + 547, + 605 + ], + "spans": [ + { + "bbox": [ + 304, + 474, + 547, + 605 + ], + "type": "text", + "content": "Perception module. To extract 3D voxel features from the input surrounding images, we adopt the approach used in previous studies [15, 25], applying a shared ResNet [6] backbone to derive 2D features from both past and present images. These features are represented as " + }, + { + "bbox": [ + 304, + 474, + 547, + 605 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{F}^{2\\mathrm{D}}|F_t^{2\\mathrm{D}}\\in" + }, + { + "bbox": [ + 304, + 474, + 547, + 605 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{c_{\\mathrm{v}}\\times h_{\\mathrm{v}}\\times w_{\\mathrm{v}}}\\}" + }, + { + "bbox": [ + 304, + 474, + 547, + 605 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 474, + 547, + 605 + ], + "type": "inline_equation", + "content": "t\\in [-N_{\\mathrm{p}},0]" + }, + { + "bbox": [ + 304, + 474, + 547, + 605 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 474, + 547, + 605 + ], + "type": "inline_equation", + "content": "c_{\\mathrm{v}}" + }, + { + "bbox": [ + 304, + 474, + 547, + 605 + ], + "type": "text", + "content": " represents the 2D feature channel, and " + }, + { + "bbox": [ + 304, + 474, + 547, + 605 + ], + "type": "inline_equation", + "content": "h_\\mathrm{v}\\times w_\\mathrm{v}" + }, + { + "bbox": [ + 304, + 474, + 547, + 605 + ], + "type": "text", + "content": " denotes the 2D spatial resolution. These 2D features are then lifted and integrated into 3D voxel features " + }, + { + "bbox": [ + 304, + 474, + 547, + 605 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{F}^{\\mathrm{3D}}|F_t^{\\mathrm{3D}}\\in \\mathbb{R}^{c\\times h\\times w\\times l}\\}" + }, + { + "bbox": [ + 304, + 474, + 547, + 605 + ], + "type": "text", + "content": " using a 2D-3D lifting module [30], where " + }, + { + "bbox": [ + 304, + 474, + 547, + 605 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 304, + 474, + 547, + 605 + ], + "type": "text", + "content": " is the 3D feature channel, and " + }, + { + "bbox": [ + 304, + 474, + 547, + 605 + ], + "type": "inline_equation", + "content": "h\\times w\\times l" + }, + { + "bbox": [ + 304, + 474, + 547, + 605 + ], + "type": "text", + "content": " represents the 3D spatial resolution." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 606, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 606, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 606, + 547, + 715 + ], + "type": "text", + "content": "Aggregation module. All the obtained 3D voxel features are projected into the present " + }, + { + "bbox": [ + 304, + 606, + 547, + 715 + ], + "type": "inline_equation", + "content": "(t = 0)" + }, + { + "bbox": [ + 304, + 606, + 547, + 715 + ], + "type": "text", + "content": " coordinate system exploiting 6-DOF ego-car poses. Here we propose an adaptive dual pooling strategy to further transform 3D voxel features into 2D BEV features. We first focus on the holistic information of each voxel column and employ average pooling to generate 2D features " + }, + { + "bbox": [ + 304, + 606, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{F}^{\\mathrm{avg}}|F_t^{\\mathrm{avg}}\\in \\mathbb{R}^{c\\times h\\times w}\\}" + }, + { + "bbox": [ + 304, + 606, + 547, + 715 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 606, + 547, + 715 + ], + "type": "inline_equation", + "content": "t\\in [-N_{\\mathrm{p}},0]" + }, + { + "bbox": [ + 304, + 606, + 547, + 715 + ], + "type": "text", + "content": ". Additionally, we use max pooling to generate the concurrent 2D features " + }, + { + "bbox": [ + 304, + 606, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{F}^{\\mathrm{max}}|F_t^{\\mathrm{max}}\\in \\mathbb{R}^{c\\times h\\times w}\\}" + }, + { + "bbox": [ + 304, + 606, + 547, + 715 + ], + "type": "text", + "content": ", cap" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22340" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 62, + 70, + 532, + 202 + ], + "blocks": [ + { + "bbox": [ + 62, + 70, + 532, + 202 + ], + "lines": [ + { + "bbox": [ + 62, + 70, + 532, + 202 + ], + "spans": [ + { + "bbox": [ + 62, + 70, + 532, + 202 + ], + "type": "image", + "image_path": "74afc6068fb6ac97f377a9eefb4309b557db3b3cab47310e49c4120b68d32800.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 204, + 547, + 260 + ], + "lines": [ + { + "bbox": [ + 46, + 204, + 547, + 260 + ], + "spans": [ + { + "bbox": [ + 46, + 204, + 547, + 260 + ], + "type": "text", + "content": "Figure 2. Architecture of our proposed EfficientOCF. The perception module (gray area) extracts 2D visual features from the input surrounding images and lifts them to 3D voxel features. Subsequently, the voxel features are further compressed into BEV features and aggregated by the aggregation module (yellow area). Next, the prediction module (blue area) with three branches sharing the same 2D structure, achieves initial OCF in 2D BEV and predicts backward centripetal flow and height values for BEV occupancy. The refinement module (green area) forecasts instances using estimated flow to refine initial 2D OCF results, which are ultimately lifted to 3D space." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 60, + 263, + 276, + 373 + ], + "blocks": [ + { + "bbox": [ + 60, + 263, + 276, + 373 + ], + "lines": [ + { + "bbox": [ + 60, + 263, + 276, + 373 + ], + "spans": [ + { + "bbox": [ + 60, + 263, + 276, + 373 + ], + "type": "image", + "image_path": "f3aa9cc965fa38737d2b5284bf0a9b325cf80a30b52f512688089e71d7b678a4.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 375, + 287, + 408 + ], + "lines": [ + { + "bbox": [ + 46, + 375, + 287, + 408 + ], + "spans": [ + { + "bbox": [ + 46, + 375, + 287, + 408 + ], + "type": "text", + "content": "Figure 3. The process of refining initial OCF results following temporal decoupling. We exploit step-wise association along the time axis to improve forecasting quality." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 413, + 287, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 413, + 287, + 496 + ], + "spans": [ + { + "bbox": [ + 46, + 413, + 287, + 496 + ], + "type": "text", + "content": "turing the voxel grid with prominent occupancy features implicitly corresponding to height values. Learnable weights, " + }, + { + "bbox": [ + 46, + 413, + 287, + 496 + ], + "type": "inline_equation", + "content": "\\alpha_{\\mathrm{avg}}" + }, + { + "bbox": [ + 46, + 413, + 287, + 496 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 413, + 287, + 496 + ], + "type": "inline_equation", + "content": "\\alpha_{\\mathrm{max}}" + }, + { + "bbox": [ + 46, + 413, + 287, + 496 + ], + "type": "text", + "content": ", are designed to adaptively adjust the contribution of these two types of 2D BEV features " + }, + { + "bbox": [ + 46, + 413, + 287, + 496 + ], + "type": "inline_equation", + "content": "\\mathcal{F}^{\\mathrm{avg}}" + }, + { + "bbox": [ + 46, + 413, + 287, + 496 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 413, + 287, + 496 + ], + "type": "inline_equation", + "content": "\\mathcal{F}^{\\mathrm{max}}" + }, + { + "bbox": [ + 46, + 413, + 287, + 496 + ], + "type": "text", + "content": ". Finally, we aggregate features from both " + }, + { + "bbox": [ + 46, + 413, + 287, + 496 + ], + "type": "inline_equation", + "content": "N_{\\mathrm{p}}" + }, + { + "bbox": [ + 46, + 413, + 287, + 496 + ], + "type": "text", + "content": " past and present inputs to obtain the aggregated 2D BEV features " + }, + { + "bbox": [ + 46, + 413, + 287, + 496 + ], + "type": "inline_equation", + "content": "F^{\\mathrm{agg}} \\in \\mathbb{R}^{(N_{\\mathrm{p}} + 1) \\times c \\times h \\times w}" + }, + { + "bbox": [ + 46, + 413, + 287, + 496 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 498, + 288, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 288, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 288, + 689 + ], + "type": "text", + "content": "Prediction module. The proposed prediction module consists of three 2D convolution-based branches for forecasting initial 2D occupancy, height values, and backward centripetal flow vectors respectively. The aggregated 2D BEV features are fed into 2D encoder-decoder structures, and then processed by three predictive heads. The segmentation head estimates initial present and future occupancy states " + }, + { + "bbox": [ + 46, + 498, + 288, + 689 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{O}^{2\\mathrm{D}}|O_t^{2\\mathrm{D}}\\in \\mathbb{R}^{1\\times H\\times W}\\}" + }, + { + "bbox": [ + 46, + 498, + 288, + 689 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 498, + 288, + 689 + ], + "type": "inline_equation", + "content": "t\\in [0,N_{\\mathrm{f}}]" + }, + { + "bbox": [ + 46, + 498, + 288, + 689 + ], + "type": "text", + "content": " in 2D space. Concurrently, the height head determines the height values of 2D voxel grids, denoted as " + }, + { + "bbox": [ + 46, + 498, + 288, + 689 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{O}^{\\mathrm{height}}|O_t^{\\mathrm{height}}\\in \\mathbb{R}^{1\\times H\\times W}\\}" + }, + { + "bbox": [ + 46, + 498, + 288, + 689 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 46, + 498, + 288, + 689 + ], + "type": "inline_equation", + "content": "t\\in [0,N_{\\mathrm{f}}]" + }, + { + "bbox": [ + 46, + 498, + 288, + 689 + ], + "type": "text", + "content": ", based on our proposed representation with spatial decoupling. In addition, the flow head generates 2D backward centripetal flow [15] for each voxel grid between consecutive frames, denoted as " + }, + { + "bbox": [ + 46, + 498, + 288, + 689 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{O}^{\\mathrm{flow}}|O_t^{\\mathrm{flow}}\\in \\mathbb{R}^{2\\times H\\times W}\\}" + }, + { + "bbox": [ + 46, + 498, + 288, + 689 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 498, + 288, + 689 + ], + "type": "inline_equation", + "content": "t\\in [0,N_{\\mathrm{f}}]" + }, + { + "bbox": [ + 46, + 498, + 288, + 689 + ], + "type": "text", + "content": ", which points from a voxel at time " + }, + { + "bbox": [ + 46, + 498, + 288, + 689 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 498, + 288, + 689 + ], + "type": "text", + "content": " to its corresponding potential 2D instance center at " + }, + { + "bbox": [ + 46, + 498, + 288, + 689 + ], + "type": "inline_equation", + "content": "t - 1" + }, + { + "bbox": [ + 46, + 498, + 288, + 689 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 689, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 288, + 713 + ], + "type": "text", + "content": "Refinement module. We find that existing OCF methods [15, 25] suffer from shape divergence over time due to" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 319, + 264, + 533, + 372 + ], + "blocks": [ + { + "bbox": [ + 319, + 264, + 533, + 372 + ], + "lines": [ + { + "bbox": [ + 319, + 264, + 533, + 372 + ], + "spans": [ + { + "bbox": [ + 319, + 264, + 533, + 372 + ], + "type": "image", + "image_path": "65a1ef89433b1c5640d5160a23e11edf785d1ec4a4da3574d09e04d9ab06ad84.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 373, + 545, + 396 + ], + "lines": [ + { + "bbox": [ + 306, + 373, + 545, + 396 + ], + "spans": [ + { + "bbox": [ + 306, + 373, + 545, + 396 + ], + "type": "text", + "content": "Figure 4. EfficientOCF achieves 3D OCF by assigning 2D grids with refined occupancy states and heights changing with time." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 403, + 546, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 403, + 546, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 403, + 546, + 715 + ], + "type": "text", + "content": "insufficient instance consistency across temporal frames. To address this, we propose decoupling current instance segmentation from future occupancy forecasts and refining 3D OCF through instance association, rather than learning 3D OCF end-to-end. The motivation is that instance segmentation based on current observations is often more accurate than end-to-end forecasts for future unobserved frames. By separating segmentation from occupancy forecasting, we simplify the learning process, making it easier compared to end-to-end 3D OCF. This allows us to leverage more accurate instance shapes to improve 3D OCF results. To this end, we design a refinement module that first performs instance segmentation in 2D BEV and then uses associated instances to mask inaccurate occupancy predictions for movable objects in the initial 2D OCF results, which are then lifted by height information to obtain refined 3D OCF. Specifically, we extract the centers of instances by nonmaximum suppression (NMS) at " + }, + { + "bbox": [ + 304, + 403, + 546, + 715 + ], + "type": "inline_equation", + "content": "t = -1" + }, + { + "bbox": [ + 304, + 403, + 546, + 715 + ], + "type": "text", + "content": " and associate pixelwise instance ID between adjacent frames iteratively over time " + }, + { + "bbox": [ + 304, + 403, + 546, + 715 + ], + "type": "inline_equation", + "content": "t \\in [0, N_{\\mathrm{f}}]" + }, + { + "bbox": [ + 304, + 403, + 546, + 715 + ], + "type": "text", + "content": " using the predicted 2D backward centripetal flow. Here we denote the instance forecasting results as " + }, + { + "bbox": [ + 304, + 403, + 546, + 715 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{M}^{2\\mathrm{D}}|M_t^{2\\mathrm{D}} \\in \\mathbb{R}^{1 \\times H \\times W}\\}" + }, + { + "bbox": [ + 304, + 403, + 546, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 403, + 546, + 715 + ], + "type": "inline_equation", + "content": "t \\in [0, N_{\\mathrm{f}}]" + }, + { + "bbox": [ + 304, + 403, + 546, + 715 + ], + "type": "text", + "content": ". For each timestep " + }, + { + "bbox": [ + 304, + 403, + 546, + 715 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 403, + 546, + 715 + ], + "type": "text", + "content": ", we convert " + }, + { + "bbox": [ + 304, + 403, + 546, + 715 + ], + "type": "inline_equation", + "content": "M_t^{2\\mathrm{D}}" + }, + { + "bbox": [ + 304, + 403, + 546, + 715 + ], + "type": "text", + "content": " to the instance mask " + }, + { + "bbox": [ + 304, + 403, + 546, + 715 + ], + "type": "inline_equation", + "content": "\\bar{M}_t^{2\\mathrm{D}}" + }, + { + "bbox": [ + 304, + 403, + 546, + 715 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 304, + 403, + 546, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP}(M_t^{2\\mathrm{D}})" + }, + { + "bbox": [ + 304, + 403, + 546, + 715 + ], + "type": "text", + "content": ", which converts voxels of instances to occupied. Ultimately, we generate the final refined 2D OCF results " + }, + { + "bbox": [ + 304, + 403, + 546, + 715 + ], + "type": "inline_equation", + "content": "\\bar{\\mathcal{O}}^{2\\mathrm{D}}(i,j) = 1" + }, + { + "bbox": [ + 304, + 403, + 546, + 715 + ], + "type": "text", + "content": ", if " + }, + { + "bbox": [ + 304, + 403, + 546, + 715 + ], + "type": "inline_equation", + "content": "O_t^{2\\mathrm{D}}(i,j) = \\bar{M}_t^{2\\mathrm{D}}(i,j) = 1" + }, + { + "bbox": [ + 304, + 403, + 546, + 715 + ], + "type": "text", + "content": ", else 0, where " + }, + { + "bbox": [ + 304, + 403, + 546, + 715 + ], + "type": "inline_equation", + "content": "(i,j)" + }, + { + "bbox": [ + 304, + 403, + 546, + 715 + ], + "type": "text", + "content": " is the" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22341" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 157 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 157 + ], + "type": "text", + "content": "voxel index in BEV, as Fig. 3 shows. We therefore achieve temporal decoupling by splitting the end-to-end OCF into step-wise instance associations along the time axis. By assigning each 2D voxel grid with corresponding refined occupancy states and height values over time, we achieve lifting 2D OCF " + }, + { + "bbox": [ + 46, + 72, + 289, + 157 + ], + "type": "inline_equation", + "content": "\\bar{\\mathcal{O}}^{2\\mathrm{D}}" + }, + { + "bbox": [ + 46, + 72, + 289, + 157 + ], + "type": "text", + "content": " to 3D OCF " + }, + { + "bbox": [ + 46, + 72, + 289, + 157 + ], + "type": "inline_equation", + "content": "\\{\\bar{\\mathcal{O}}^{3\\mathrm{D}}|\\bar{\\mathcal{O}}_t^{3\\mathrm{D}}\\in \\mathbb{R}^{1\\times H\\times W\\times L}\\}" + }, + { + "bbox": [ + 46, + 72, + 289, + 157 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 72, + 289, + 157 + ], + "type": "inline_equation", + "content": "t\\in [0,N_{\\mathrm{f}}]" + }, + { + "bbox": [ + 46, + 72, + 289, + 157 + ], + "type": "text", + "content": ", as shown in Fig. 4." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 163, + 178, + 176 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 163, + 178, + 176 + ], + "spans": [ + { + "bbox": [ + 47, + 163, + 178, + 176 + ], + "type": "text", + "content": "3.3. Training and Inference" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 181, + 287, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 181, + 287, + 217 + ], + "spans": [ + { + "bbox": [ + 46, + 181, + 287, + 217 + ], + "type": "text", + "content": "Our approach is trained and evaluated on three publicly available datasets, including the nuScenes [2], nuScenes-Occupancy [34], and Lyft-Level5 [11] datasets." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 218, + 289, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 218, + 289, + 492 + ], + "spans": [ + { + "bbox": [ + 46, + 218, + 289, + 492 + ], + "type": "text", + "content": "Data generation. We develop a novel data generation method based on three publicly available datasets. Initially, we split these datasets into sequences with a time length of " + }, + { + "bbox": [ + 46, + 218, + 289, + 492 + ], + "type": "inline_equation", + "content": "N = N_{\\mathrm{p}} + N_{\\mathrm{f}} + 1" + }, + { + "bbox": [ + 46, + 218, + 289, + 492 + ], + "type": "text", + "content": ". Sequential semantic and instance annotations of general movable objects are then extracted for each sequence. Based on the extracted semantic information, we set the voxel grids corresponding to the general movable objects to 1 and the other positions to 0. We thereby obtain the bounding-box-aware 3D occupancy states " + }, + { + "bbox": [ + 46, + 218, + 289, + 492 + ], + "type": "inline_equation", + "content": "\\{\\hat{\\mathcal{O}}^{\\mathrm{bb}}|\\hat{\\mathcal{O}}_t^{\\mathrm{bb}}\\in \\mathbb{R}^{1\\times H\\times W\\times L}\\}" + }, + { + "bbox": [ + 46, + 218, + 289, + 492 + ], + "type": "text", + "content": " (where 1 indicates occupied and 0 indicates free), which corresponds to the inflated annotations mentioned in [25]. In this case, we assign the heights of ground-truth bounding boxes to the voxel grids within them. Subsequently, we compress 3D voxel annotations into 2D BEV counterparts " + }, + { + "bbox": [ + 46, + 218, + 289, + 492 + ], + "type": "inline_equation", + "content": "\\{\\hat{\\mathcal{O}}^{\\mathrm{bev}}|\\hat{\\mathcal{O}}_t^{\\mathrm{bev}}\\in \\mathbb{R}^{1\\times H\\times W}\\}" + }, + { + "bbox": [ + 46, + 218, + 289, + 492 + ], + "type": "text", + "content": ". This is achieved under the principle that if any voxel in a vertical voxel column has a value of 1, the corresponding position in the BEV is set to 1. We extract ground-truth height values of each voxel column " + }, + { + "bbox": [ + 46, + 218, + 289, + 492 + ], + "type": "inline_equation", + "content": "\\{\\hat{\\mathcal{O}}^{\\mathrm{height}}|\\hat{\\mathcal{O}}_t^{\\mathrm{height}}\\in \\mathbb{R}^{1\\times H\\times W}\\}" + }, + { + "bbox": [ + 46, + 218, + 289, + 492 + ], + "type": "text", + "content": " by selecting the highest grid with the semantics of general movable objects. We then conduct fine-grained 3D occupancy states " + }, + { + "bbox": [ + 46, + 218, + 289, + 492 + ], + "type": "inline_equation", + "content": "\\{\\hat{\\mathcal{O}}^{\\mathrm{fg}}|\\hat{\\mathcal{O}}_t^{\\mathrm{fg}}\\in \\mathbb{R}^{1\\times H\\times W\\times L}\\}" + }, + { + "bbox": [ + 46, + 218, + 289, + 492 + ], + "type": "text", + "content": " by assigning height values to 2D BEV occupancy annotations." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 493, + 287, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 493, + 287, + 552 + ], + "spans": [ + { + "bbox": [ + 47, + 493, + 287, + 552 + ], + "type": "text", + "content": "Loss functions. We employ the cross-entropy loss as the OCF loss " + }, + { + "bbox": [ + 47, + 493, + 287, + 552 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{occ}}" + }, + { + "bbox": [ + 47, + 493, + 287, + 552 + ], + "type": "text", + "content": ", and use the smooth L1 loss for both the height prediction loss " + }, + { + "bbox": [ + 47, + 493, + 287, + 552 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{height}}" + }, + { + "bbox": [ + 47, + 493, + 287, + 552 + ], + "type": "text", + "content": " and the flow prediction loss " + }, + { + "bbox": [ + 47, + 493, + 287, + 552 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{flow}}" + }, + { + "bbox": [ + 47, + 493, + 287, + 552 + ], + "type": "text", + "content": ". The total loss used to train our EfficientOCF is the weighted sum of these three losses, computed as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 58, + 553, + 287, + 597 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 553, + 287, + 597 + ], + "spans": [ + { + "bbox": [ + 58, + 553, + 287, + 597 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {\\text {a l l}} = \\frac {1}{N _ {\\mathrm {f}} + 1} \\sum_ {t = 0} ^ {N _ {\\mathrm {f}}} \\left(\\lambda_ {1} \\mathcal {L} _ {\\text {o c c}} \\left(\\hat {O} _ {t} ^ {\\text {b e v}}, \\bar {O} _ {t} ^ {\\text {2 D}}\\right) \\right. \\tag {1} \\\\ + \\lambda_ {2} \\mathcal {L} _ {\\text {h e i g h t}} \\left(\\hat {O} _ {t} ^ {\\text {h e i g h t}}, O _ {t} ^ {\\text {h e i g h t}}\\right) + \\lambda_ {3} \\mathcal {L} _ {\\text {f l o w}} \\left(\\hat {O} _ {t} ^ {\\text {f l o w}}, O _ {t} ^ {\\text {f l o w}}\\right)), \\\\ \\end{array}", + "image_path": "d5bd8a403e256c4e7759ff67bea1fc750dce0cc3533a154ede7cbb9c965a044c.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 605, + 287, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 605, + 287, + 629 + ], + "spans": [ + { + "bbox": [ + 46, + 605, + 287, + 629 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 605, + 287, + 629 + ], + "type": "inline_equation", + "content": "\\lambda_{1}" + }, + { + "bbox": [ + 46, + 605, + 287, + 629 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 605, + 287, + 629 + ], + "type": "inline_equation", + "content": "\\lambda_{2}" + }, + { + "bbox": [ + 46, + 605, + 287, + 629 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 605, + 287, + 629 + ], + "type": "inline_equation", + "content": "\\lambda_{3}" + }, + { + "bbox": [ + 46, + 605, + 287, + 629 + ], + "type": "text", + "content": " are the weighting parameters to balance the supervision for different heads." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 630, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 287, + 714 + ], + "type": "text", + "content": "Evaluation metrics. We employ intersection-over-union (IoU), and conditional IoU (C-IoU) to evaluate OCF performance. We also use video panoptic quality (VPQ) [13] to assess the performance of 3D instance forecasting. We utilize " + }, + { + "bbox": [ + 46, + 630, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{IoU_c}" + }, + { + "bbox": [ + 46, + 630, + 287, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 630, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{IoU_f}" + }, + { + "bbox": [ + 46, + 630, + 287, + 714 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 46, + 630, + 287, + 714 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathrm{IoU}}_{\\mathrm{f}}" + }, + { + "bbox": [ + 46, + 630, + 287, + 714 + ], + "type": "text", + "content": " to separately evaluate the occupancy estimation at the present moment (" + }, + { + "bbox": [ + 46, + 630, + 287, + 714 + ], + "type": "inline_equation", + "content": "t = 0" + }, + { + "bbox": [ + 46, + 630, + 287, + 714 + ], + "type": "text", + "content": "), the forecasting results for future timesteps (" + }, + { + "bbox": [ + 46, + 630, + 287, + 714 + ], + "type": "inline_equation", + "content": "t \\in [1, N_{\\mathrm{f}}]" + }, + { + "bbox": [ + 46, + 630, + 287, + 714 + ], + "type": "text", + "content": "), and" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 333, + 69, + 521, + 186 + ], + "blocks": [ + { + "bbox": [ + 333, + 69, + 521, + 186 + ], + "lines": [ + { + "bbox": [ + 333, + 69, + 521, + 186 + ], + "spans": [ + { + "bbox": [ + 333, + 69, + 521, + 186 + ], + "type": "image", + "image_path": "d528d0b78a4cb304845c032d58f0e2eb2cd1ae04dd1ed9a89a8918ecd1f99b3f.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 186, + 545, + 220 + ], + "lines": [ + { + "bbox": [ + 304, + 186, + 545, + 220 + ], + "spans": [ + { + "bbox": [ + 304, + 186, + 545, + 220 + ], + "type": "text", + "content": "Figure 5. Examples of incomplete 3D occupancy labels due to LiDAR sparsity in nuScenes-Occupancy [34], which cannot comprehensively assess occupancy forecasting methods." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 224, + 545, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 224, + 545, + 247 + ], + "spans": [ + { + "bbox": [ + 306, + 224, + 545, + 247 + ], + "type": "text", + "content": "the occupancy estimation over the entire time horizon " + }, + { + "bbox": [ + 306, + 224, + 545, + 247 + ], + "type": "inline_equation", + "content": "(t\\in [0,N_{\\mathrm{f}}])" + }, + { + "bbox": [ + 306, + 224, + 545, + 247 + ], + "type": "text", + "content": ". These evaluation metrics are computed by" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 331, + 247, + 545, + 277 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 247, + 545, + 277 + ], + "spans": [ + { + "bbox": [ + 331, + 247, + 545, + 277 + ], + "type": "interline_equation", + "content": "\\mathrm {I o U} = \\frac {1}{N _ {\\text {a l l}}} \\sum_ {t = t _ {\\mathrm {s}}} ^ {t _ {\\mathrm {e}}} \\frac {\\sum \\hat {O} _ {t} ^ {\\mathrm {o c c}} \\cdot \\bar {O} _ {t} ^ {\\mathrm {o c c}}}{\\sum \\hat {O} _ {t} ^ {\\mathrm {o c c}} + \\bar {O} _ {t} ^ {\\mathrm {o c c}} - \\hat {O} _ {t} ^ {\\mathrm {o c c}} \\cdot \\bar {O} _ {t} ^ {\\mathrm {o c c}}}. \\tag {2}", + "image_path": "f1301155ac6935872b2d45f6b616693c492cafc4363edc54d3dc6523d4b7f3ba.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 284, + 545, + 368 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 284, + 545, + 368 + ], + "spans": [ + { + "bbox": [ + 304, + 284, + 545, + 368 + ], + "type": "text", + "content": "We set " + }, + { + "bbox": [ + 304, + 284, + 545, + 368 + ], + "type": "inline_equation", + "content": "\\hat{O}_t^{\\mathrm{occ}} / \\bar{O}_t^{\\mathrm{occ}}" + }, + { + "bbox": [ + 304, + 284, + 545, + 368 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 304, + 284, + 545, + 368 + ], + "type": "inline_equation", + "content": "\\hat{O}_t^{\\mathrm{bev}} / \\bar{O}_t^{2\\mathrm{D}}" + }, + { + "bbox": [ + 304, + 284, + 545, + 368 + ], + "type": "text", + "content": " to access 2D OCF performance. To evaluate the performance of forecasting 3D inflated and fine-grained occupancy respectively, we calculate IoU by setting " + }, + { + "bbox": [ + 304, + 284, + 545, + 368 + ], + "type": "inline_equation", + "content": "\\hat{O}_t^{\\mathrm{occ}} / \\bar{O}_t^{\\mathrm{occ}}" + }, + { + "bbox": [ + 304, + 284, + 545, + 368 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 304, + 284, + 545, + 368 + ], + "type": "inline_equation", + "content": "\\hat{O}_t^{\\mathrm{bb}} / \\bar{O}_t^{3\\mathrm{D}}" + }, + { + "bbox": [ + 304, + 284, + 545, + 368 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 284, + 545, + 368 + ], + "type": "inline_equation", + "content": "\\hat{O}_t^{\\mathrm{fg}} / \\bar{O}_t^{3\\mathrm{D}}" + }, + { + "bbox": [ + 304, + 284, + 545, + 368 + ], + "type": "text", + "content": ". We set " + }, + { + "bbox": [ + 304, + 284, + 545, + 368 + ], + "type": "inline_equation", + "content": "t_{\\mathrm{s}} = t_{\\mathrm{e}} = 0" + }, + { + "bbox": [ + 304, + 284, + 545, + 368 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 284, + 545, + 368 + ], + "type": "inline_equation", + "content": "N_{\\mathrm{all}} = 1" + }, + { + "bbox": [ + 304, + 284, + 545, + 368 + ], + "type": "text", + "content": " to calculate IoUc, " + }, + { + "bbox": [ + 304, + 284, + 545, + 368 + ], + "type": "inline_equation", + "content": "t_{\\mathrm{s}} = 1" + }, + { + "bbox": [ + 304, + 284, + 545, + 368 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 284, + 545, + 368 + ], + "type": "inline_equation", + "content": "t_{\\mathrm{e}} = N_{\\mathrm{f}}" + }, + { + "bbox": [ + 304, + 284, + 545, + 368 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 284, + 545, + 368 + ], + "type": "inline_equation", + "content": "N_{\\mathrm{all}} = N_{\\mathrm{f}} + 1" + }, + { + "bbox": [ + 304, + 284, + 545, + 368 + ], + "type": "text", + "content": " for IoU." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 369, + 546, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 369, + 546, + 561 + ], + "spans": [ + { + "bbox": [ + 304, + 369, + 546, + 561 + ], + "type": "text", + "content": "While conventional IoU metrics can evaluate forecasting results within bounding boxes or fine-grained voxels [25], they struggle with scenarios involving incomplete 3D annotations due to LiDAR data sparsity, as illustrated in Fig. 5. The nuScenes-Occupancy dataset [34] extends the original nuScenes dataset [2] with dense semantic occupancy annotation using the proposed augmenting and purifying (AAP) pipeline. However, during the AAP's automatic labeling process, some annotation omissions are inevitable, resulting in missed 3D occupancy labels due to factors such as occlusion or sparse LiDAR observations, and even introducing pseudo occupancy labels with self-training [37]. Additionally, during the manual purifying phase, while noise removal is effective, high-precision leak detection remains challenging. Therefore, we propose a new evaluation metric for 3D occupancy forecasting, conditional IoU (C-IoU):" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 567, + 545, + 611 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 567, + 545, + 611 + ], + "spans": [ + { + "bbox": [ + 313, + 567, + 545, + 611 + ], + "type": "interline_equation", + "content": "\\mathrm {C} - \\mathrm {I o U} = \\frac {1}{N _ {\\text {a l l}}} \\sum_ {t = t _ {\\mathrm {s}}} ^ {t _ {\\mathrm {e}}} \\frac {\\left| T P _ {t} \\right| + \\left| \\widetilde {F P} _ {t} \\right|}{\\left| T P _ {t} \\right| + \\left| F N _ {t} \\right| + \\left(\\left| F P _ {t} \\right| - \\left| \\widetilde {F P} _ {t} \\right|\\right)}, \\tag {3}", + "image_path": "e5e07d46d93417d58967b188b877b7a98ad62a9d6c5e6a9994d3d4769efbb5c1.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 616, + 546, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 616, + 546, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 616, + 546, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 616, + 546, + 713 + ], + "type": "inline_equation", + "content": "TP_{t}" + }, + { + "bbox": [ + 304, + 616, + 546, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 304, + 616, + 546, + 713 + ], + "type": "inline_equation", + "content": "FN_{t}" + }, + { + "bbox": [ + 304, + 616, + 546, + 713 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 304, + 616, + 546, + 713 + ], + "type": "inline_equation", + "content": "FP_{t}" + }, + { + "bbox": [ + 304, + 616, + 546, + 713 + ], + "type": "text", + "content": " represent true positives, false negatives, and false positives at timestep " + }, + { + "bbox": [ + 304, + 616, + 546, + 713 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 304, + 616, + 546, + 713 + ], + "type": "text", + "content": ", computed in the 3D fine-grained format. " + }, + { + "bbox": [ + 304, + 616, + 546, + 713 + ], + "type": "inline_equation", + "content": "\\widetilde{FP}_{t}" + }, + { + "bbox": [ + 304, + 616, + 546, + 713 + ], + "type": "text", + "content": " represents partial false positives of " + }, + { + "bbox": [ + 304, + 616, + 546, + 713 + ], + "type": "inline_equation", + "content": "FP_{t}" + }, + { + "bbox": [ + 304, + 616, + 546, + 713 + ], + "type": "text", + "content": ", which lies in the corresponding annotated bounding boxes. Note that we only use this metric for nuScenes and nuScenes-Occupancy datasets because Lyft-Level5 lacks fine-grained voxel-wise annotations. C-IoU comprehensively considers predictions within" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22342" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 83, + 542, + 186 + ], + "blocks": [ + { + "bbox": [ + 175, + 71, + 418, + 82 + ], + "lines": [ + { + "bbox": [ + 175, + 71, + 418, + 82 + ], + "spans": [ + { + "bbox": [ + 175, + 71, + 418, + 82 + ], + "type": "text", + "content": "Table 1. Comparison of performance on 3D occupancy forecasting" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 83, + 542, + 186 + ], + "lines": [ + { + "bbox": [ + 48, + 83, + 542, + 186 + ], + "spans": [ + { + "bbox": [ + 48, + 83, + 542, + 186 + ], + "type": "table", + "html": "
ApproachnuScenesLyft-Level5nuScenes & nuScenes-Occupancy
2D3D2D3D3D
IoUcIoUfIoUIoUcIoUfIoUIoUcIoUfIoUIoUcIoUfIoUIoUcIoUfIoUC-IoUcC-IoUfC-IoUf
PowerBEV33.6230.0730.7728.3425.4726.0545.7843.4443.9136.1534.1834.58------
OpenOccupancy28.7725.9326.5024.2622.6923.1538.1635.1135.7932.9529.7330.4014.9913.0613.4727.1124.6825.17
OccFormer38.9831.0532.7634.7528.0829.5647.0032.1035.0241.6828.5531.0015.8011.3012.1844.5635.9837.67
OCFNet-30.3426.2727.4727.3224.4024.9836.0233.1233.6931.0428.9229.3612.6610.6511.0529.6427.4727.91
OCFNet35.5032.0232.7133.2429.3629.9445.7443.6744.0839.7138.1438.4614.7312.5713.0034.8333.1433.49
EfficientOCF-(ours)31.6929.0029.5630.2428.1728.6036.3933.2433.8333.9531.1231.7015.9814.0514.4635.8934.2034.56
EfficientOCF (ours)39.9336.1536.9335.6032.7333.3247.4244.9645.4644.2741.9042.3921.2819.0219.5047.5345.5745.99
", + "image_path": "56d2c0418ceac4e9519b2fbdac62564f3e64fa86c341076d39b2201ba7cc393c.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 192, + 287, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 192, + 287, + 288 + ], + "spans": [ + { + "bbox": [ + 46, + 192, + 287, + 288 + ], + "type": "text", + "content": "bonding boxes and fine-grained occupancy predictions, accounting for the discrepancies between the nuScenes and nuScenes-Occupancy datasets to avoid penalizing false positives in bounding boxes unfairly. This mitigates the effect of missing and incomplete occupancy annotations within bounding boxes of movable objects. Additionally, we use VPQ [13] to evaluate 3D instance forecasting lifted from 2D counterparts by height values, which is calculated by" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 291, + 287, + 323 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 291, + 287, + 323 + ], + "spans": [ + { + "bbox": [ + 55, + 291, + 287, + 323 + ], + "type": "interline_equation", + "content": "\\mathrm {V P Q} \\left(\\hat {\\mathcal {O}} ^ {\\mathrm {3 D}}, \\bar {\\mathcal {O}} ^ {\\mathrm {3 D}}\\right) = \\sum_ {t = 0} ^ {N _ {\\mathrm {f}}} \\frac {\\sum_ {T P _ {t}} \\mathrm {I o U} _ {t}}{\\left| T P _ {t} \\right| + \\frac {1}{2} \\left| F P _ {t} \\right| + \\frac {1}{2} \\left| F N _ {t} \\right|}. \\tag {4}", + "image_path": "95b6ebb5eab281d2b9dd3a649aca256623a9afa2e8e30f73519ad3ff959c443c.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 334, + 287, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 334, + 287, + 361 + ], + "spans": [ + { + "bbox": [ + 47, + 334, + 287, + 361 + ], + "type": "text", + "content": "We set " + }, + { + "bbox": [ + 47, + 334, + 287, + 361 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{O}}^{3\\mathrm{D}}" + }, + { + "bbox": [ + 47, + 334, + 287, + 361 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 47, + 334, + 287, + 361 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{O}}^{\\mathrm{bb}}" + }, + { + "bbox": [ + 47, + 334, + 287, + 361 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 334, + 287, + 361 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{O}}^{\\mathrm{fg}}" + }, + { + "bbox": [ + 47, + 334, + 287, + 361 + ], + "type": "text", + "content": " for inflated and fine-grained formats, denoted as " + }, + { + "bbox": [ + 47, + 334, + 287, + 361 + ], + "type": "inline_equation", + "content": "\\mathrm{VPQ}^{\\mathrm{bb}}" + }, + { + "bbox": [ + 47, + 334, + 287, + 361 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 47, + 334, + 287, + 361 + ], + "type": "inline_equation", + "content": "\\mathrm{VPQ}^{\\mathrm{fg}}" + }, + { + "bbox": [ + 47, + 334, + 287, + 361 + ], + "type": "text", + "content": " respectively." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 373, + 128, + 387 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 373, + 128, + 387 + ], + "spans": [ + { + "bbox": [ + 47, + 373, + 128, + 387 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 394, + 168, + 407 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 394, + 168, + 407 + ], + "spans": [ + { + "bbox": [ + 47, + 394, + 168, + 407 + ], + "type": "text", + "content": "4.1. Experimental Setups" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 413, + 287, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 413, + 287, + 581 + ], + "spans": [ + { + "bbox": [ + 46, + 413, + 287, + 581 + ], + "type": "text", + "content": "Dataset details. We evaluate our proposed EfficientOCF alongside the SOTA baselines using the nuScenes [2], nuScenes-Occupancy [34] and Lyft-Level5 [11] datasets. Following existing works [15, 25, 34], we split the nuScenes and nuScenes-Occupancy datasets into training and test sets with 700 and 150 scenes respectively, and the Lyft-Level5 dataset into 130 training scenes and 50 test scenes. Each sequence has a time length " + }, + { + "bbox": [ + 46, + 413, + 287, + 581 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 46, + 413, + 287, + 581 + ], + "type": "text", + "content": " of 7 frames (" + }, + { + "bbox": [ + 46, + 413, + 287, + 581 + ], + "type": "inline_equation", + "content": "N_{\\mathrm{p}} = 2" + }, + { + "bbox": [ + 46, + 413, + 287, + 581 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 413, + 287, + 581 + ], + "type": "inline_equation", + "content": "N_{\\mathrm{f}} = 4" + }, + { + "bbox": [ + 46, + 413, + 287, + 581 + ], + "type": "text", + "content": "), meaning that we use 2 past frames and 1 present observation to estimate the present occupancy states and forecast the counterparts of the next 4 frames. The nuScenes and nuScenes-Occupancy datasets are labeled at a frequency of " + }, + { + "bbox": [ + 46, + 413, + 287, + 581 + ], + "type": "inline_equation", + "content": "2\\mathrm{Hz}" + }, + { + "bbox": [ + 46, + 413, + 287, + 581 + ], + "type": "text", + "content": ", and the Lyft-Level5 dataset at " + }, + { + "bbox": [ + 46, + 413, + 287, + 581 + ], + "type": "inline_equation", + "content": "5\\mathrm{Hz}" + }, + { + "bbox": [ + 46, + 413, + 287, + 581 + ], + "type": "text", + "content": ", resulting in different forecasting time durations across these datasets." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 582, + 288, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 582, + 288, + 712 + ], + "spans": [ + { + "bbox": [ + 46, + 582, + 288, + 712 + ], + "type": "text", + "content": "Implementation details. Following [25], we set the evaluation range to " + }, + { + "bbox": [ + 46, + 582, + 288, + 712 + ], + "type": "inline_equation", + "content": "[-51.2\\mathrm{m}, 51.2\\mathrm{m}]" + }, + { + "bbox": [ + 46, + 582, + 288, + 712 + ], + "type": "text", + "content": " for both the x-axis and y-axis, and " + }, + { + "bbox": [ + 46, + 582, + 288, + 712 + ], + "type": "inline_equation", + "content": "[-5\\mathrm{m}, 3\\mathrm{m}]" + }, + { + "bbox": [ + 46, + 582, + 288, + 712 + ], + "type": "text", + "content": " for the z-axis. The voxel resolution is set to " + }, + { + "bbox": [ + 46, + 582, + 288, + 712 + ], + "type": "inline_equation", + "content": "0.2\\mathrm{m}" + }, + { + "bbox": [ + 46, + 582, + 288, + 712 + ], + "type": "text", + "content": ", resulting in 3D occupancy voxels of size " + }, + { + "bbox": [ + 46, + 582, + 288, + 712 + ], + "type": "inline_equation", + "content": "512 \\times 512 \\times 40" + }, + { + "bbox": [ + 46, + 582, + 288, + 712 + ], + "type": "text", + "content": ". We compare our EfficientOCF against four SOTA methods: PowerBEV [15], OpenOccupancy [34], OccFormer [39], and OCFNet [25]. All models are trained for 15 epochs using AdamW optimizer [14] with an initial learning rate " + }, + { + "bbox": [ + 46, + 582, + 288, + 712 + ], + "type": "inline_equation", + "content": "3 \\times 10^{-4}" + }, + { + "bbox": [ + 46, + 582, + 288, + 712 + ], + "type": "text", + "content": " and a weight decay of 0.01. Eight A100 GPUs are used for training with a batch size of 8." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 305, + 191, + 495, + 205 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 191, + 495, + 205 + ], + "spans": [ + { + "bbox": [ + 305, + 191, + 495, + 205 + ], + "type": "text", + "content": "4.2. Occupancy Forecasting Assessment" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 210, + 545, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 210, + 545, + 641 + ], + "spans": [ + { + "bbox": [ + 304, + 210, + 545, + 641 + ], + "type": "text", + "content": "We first assess the occupancy forecasting performance of EfficientOCF and baseline methods, as shown in Tab. 1. The nuScenes-Occupancy dataset provides fine-grained labels for the nuScenes dataset, allowing us to conduct evaluations based on fine-grained voxels. In contrast, the Lyft-Level5 dataset only contains bounding box annotations for movable objects, and thus the evaluation on this dataset is based on inflated objects where the C-IoU metric is not available. Notably, the vanilla PowerBEV only forecasts occupancy states in 2D space. Therefore, we lift its 2D results to 3D space by assigning a fixed height value to compute the 3D metrics. This fixed height for each test scene is determined by averaging the heights of all general movable objects. For OpenOccupancy and OccFormer, we estimate the occupancy states of the present frame and then use it as the prediction of all future timesteps based on the static-world model [25]. Besides, for OpenOccupancy, OccFormer, and OCFNet, which are end-to-end 3D occupancy perception methods, we first calculate 3D metrics and then compress the 3D results to BEV space to calculate the 2D metrics. Following [25], we additionally report the performance of OCFNet and our EfficientOCF when trained on limited " + }, + { + "bbox": [ + 304, + 210, + 545, + 641 + ], + "type": "inline_equation", + "content": "1/6" + }, + { + "bbox": [ + 304, + 210, + 545, + 641 + ], + "type": "text", + "content": " of the training sequences, denoted as OCFNet" + }, + { + "bbox": [ + 304, + 210, + 545, + 641 + ], + "type": "inline_equation", + "content": "^{-}" + }, + { + "bbox": [ + 304, + 210, + 545, + 641 + ], + "type": "text", + "content": " and EfficientOCF" + }, + { + "bbox": [ + 304, + 210, + 545, + 641 + ], + "type": "inline_equation", + "content": "^{-}" + }, + { + "bbox": [ + 304, + 210, + 545, + 641 + ], + "type": "text", + "content": ". Tab. 1 shows that our proposed EfficientOCF estimates the present occupancy states and forecasts future counterparts better than all the other baselines on the nuScenes, nuScenes-Occupancy, and Lyft-Level5 datasets in both 2D and 3D evaluation metrics. In particular, EfficientOCF improves the proposed C-IoU of the baselines more significantly than the other existing metrics. It is also notable that EfficientOCF" + }, + { + "bbox": [ + 304, + 210, + 545, + 641 + ], + "type": "inline_equation", + "content": "^{-}" + }, + { + "bbox": [ + 304, + 210, + 545, + 641 + ], + "type": "text", + "content": " trained on limited data even outperforms the baseline OCFNet supervised by the holistic training set on all 3D fine-grained metrics. This validates that our method more effectively captures diverse motion patterns, even with only partial scenario perception, making it practically useful when training data is limited." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 305, + 642, + 546, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 642, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 305, + 642, + 546, + 714 + ], + "type": "text", + "content": "We further report the forecasting performance with different future time horizons. As shown in Tab. 2, our EfficientOCF remains the best performance within the future time horizons " + }, + { + "bbox": [ + 305, + 642, + 546, + 714 + ], + "type": "inline_equation", + "content": "\\{0.5\\mathrm{s},1.0\\mathrm{s},1.5\\mathrm{s},2.0\\mathrm{s}\\}" + }, + { + "bbox": [ + 305, + 642, + 546, + 714 + ], + "type": "text", + "content": " on the nuScenes and nuScenes-Occupancy datasets, and " + }, + { + "bbox": [ + 305, + 642, + 546, + 714 + ], + "type": "inline_equation", + "content": "\\{0.2\\mathrm{s},0.4\\mathrm{s},0.6\\mathrm{s},0.8\\mathrm{s}\\}" + }, + { + "bbox": [ + 305, + 642, + 546, + 714 + ], + "type": "text", + "content": " on the Lyft-Level5 dataset. Our" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22343" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 83, + 542, + 191 + ], + "blocks": [ + { + "bbox": [ + 147, + 71, + 444, + 81 + ], + "lines": [ + { + "bbox": [ + 147, + 71, + 444, + 81 + ], + "spans": [ + { + "bbox": [ + 147, + 71, + 444, + 81 + ], + "type": "text", + "content": "Table 2. Comparison of performance on 3D OCF in different future time horizons" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 83, + 542, + 191 + ], + "lines": [ + { + "bbox": [ + 48, + 83, + 542, + 191 + ], + "spans": [ + { + "bbox": [ + 48, + 83, + 542, + 191 + ], + "type": "table", + "html": "
ApproachnuScenesLyft-Level5nuScenes & nuScenes-Occupancy
IoUf (2D)IoUf (3D)IoUf (2D)IoUf (3D)IoUf (3D)C-IoUf (3D)
0.5s1.0s1.5s2.0s0.5s1.0s1.5s2.0s0.2s0.4s0.6s0.8s0.2s0.4s0.6s0.8s0.5s1.0s1.5s2.0s0.5s1.0s1.5s2.0s
PowerBEV32.1331.3130.6130.0727.1726.4925.9225.4745.0444.6444.0543.4435.5235.1834.6834.18--------
OpenOccupancy27.8927.1126.4625.9324.2023.7923.2622.6937.2436.6235.8635.1132.0331.1930.4529.7314.0813.7113.3613.0626.6225.8325.1624.68
OccFormer34.3532.5231.6031.0531.0929.4928.6228.0839.7735.6733.4832.1034.9231.5629.7428.5512.8512.0711.6111.3039.3537.5236.5635.98
OCFNet-29.1128.1127.3826.7726.3125.5124.9024.4035.0534.4633.8033.1230.3929.9729.4528.9211.6511.2510.9410.6529.0928.3927.8827.47
OCFNet34.4533.6532.8432.0231.4030.7330.0429.3645.1844.7744.2743.6739.4039.0238.5938.1413.6413.3412.9712.5734.6434.1733.6633.14
EfficientOCF-30.9030.1629.5229.0029.7429.1328.6028.1735.3534.6433.9333.2432.8832.4331.7831.1215.0714.7014.3514.0535.8635.1634.6134.20
EfficientOCF38.8037.8536.9836.1534.8134.0733.3732.7346.7046.2045.5944.9643.5443.0742.4941.9020.2119.8619.4619.0247.6346.9146.2445.57
", + "image_path": "11baf4d8cf137d53b64911960d9f5b3f18d44abf2f82b97c5da699cd119ba0c0.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 61, + 196, + 136, + 371 + ], + "blocks": [ + { + "bbox": [ + 61, + 196, + 136, + 371 + ], + "lines": [ + { + "bbox": [ + 61, + 196, + 136, + 371 + ], + "spans": [ + { + "bbox": [ + 61, + 196, + 136, + 371 + ], + "type": "image", + "image_path": "4018c459a76367314a027632f7ca7b826828f1979fc9d059029362e2466974ef.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 373, + 545, + 395 + ], + "lines": [ + { + "bbox": [ + 46, + 373, + 545, + 395 + ], + "spans": [ + { + "bbox": [ + 46, + 373, + 545, + 395 + ], + "type": "text", + "content": "Figure 6. Visualization of OCF results of EfficientOCF and OCFNet, and ground-truth from nuScenes (nusc), nuScenes-Occupancy (nusc-occ), and Lyft-Level5 datasets. The OCF results and ground-truth labels from timesteps 0 to " + }, + { + "bbox": [ + 46, + 373, + 545, + 395 + ], + "type": "inline_equation", + "content": "N_{\\mathrm{f}}" + }, + { + "bbox": [ + 46, + 373, + 545, + 395 + ], + "type": "text", + "content": " are assigned colors from dark to light." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 138, + 197, + 212, + 371 + ], + "blocks": [ + { + "bbox": [ + 138, + 197, + 212, + 371 + ], + "lines": [ + { + "bbox": [ + 138, + 197, + 212, + 371 + ], + "spans": [ + { + "bbox": [ + 138, + 197, + 212, + 371 + ], + "type": "image", + "image_path": "a2f034abc47480312af93ff6cd6ff0bf571b94d38173d71467d1de8a8f57781f.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 221, + 197, + 294, + 371 + ], + "blocks": [ + { + "bbox": [ + 221, + 197, + 294, + 371 + ], + "lines": [ + { + "bbox": [ + 221, + 197, + 294, + 371 + ], + "spans": [ + { + "bbox": [ + 221, + 197, + 294, + 371 + ], + "type": "image", + "image_path": "4a43ad1722483a3c8d2fe70520e3001b10cae52195f15a20afc7b349634e026e.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 298, + 197, + 372, + 371 + ], + "blocks": [ + { + "bbox": [ + 298, + 197, + 372, + 371 + ], + "lines": [ + { + "bbox": [ + 298, + 197, + 372, + 371 + ], + "spans": [ + { + "bbox": [ + 298, + 197, + 372, + 371 + ], + "type": "image", + "image_path": "6743b106ca56dc9ba35c8a2af2620852b9584a659a64a2fd9628a5247ffa9921.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 380, + 197, + 455, + 371 + ], + "blocks": [ + { + "bbox": [ + 380, + 197, + 455, + 371 + ], + "lines": [ + { + "bbox": [ + 380, + 197, + 455, + 371 + ], + "spans": [ + { + "bbox": [ + 380, + 197, + 455, + 371 + ], + "type": "image", + "image_path": "ac815a6c0efd2993c557daf83bd643085b383472a6d262f174122fd60c032a4f.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 457, + 197, + 531, + 371 + ], + "blocks": [ + { + "bbox": [ + 457, + 197, + 531, + 371 + ], + "lines": [ + { + "bbox": [ + 457, + 197, + 531, + 371 + ], + "spans": [ + { + "bbox": [ + 457, + 197, + 531, + 371 + ], + "type": "image", + "image_path": "82919a370796b9037aa3de4286366279a885a4e428233bc0b6ac9ac60f0c437a.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 46, + 399, + 288, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 399, + 288, + 590 + ], + "spans": [ + { + "bbox": [ + 46, + 399, + 288, + 590 + ], + "type": "text", + "content": "temporal decoupling leads to more accurate OCF through instance forecasting associated with predicted flow at each timestep. As can be seen in Fig. 6, EfficientOCF captures better contours of movable objects (e.g., the first and the second samples) compared to the SOTA baseline OCFNet which basically overlooks shape details. This is achieved by spatial decoupling, which provides more detailed 3D structures inherent in the combination of BEV occupancy and heights. Moreover, we demonstrate that EfficientOCF can generate more TP and fewer FP predictions in 2D space by Fig. 7, which also contributes to the higher forecasting accuracy of our EfficientOCF. Besides, it can be seen from the nusc-occ ground-truth (GT) of Fig. 6 that the publicly available fine-grained GT annotations are not always accurate due to LiDAR sparsity. This reflects the need for our new metrics C-IoU." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 598, + 147, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 598, + 147, + 609 + ], + "spans": [ + { + "bbox": [ + 47, + 598, + 147, + 609 + ], + "type": "text", + "content": "4.3. Ablation Studies" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 617, + 287, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 617, + 287, + 666 + ], + "spans": [ + { + "bbox": [ + 46, + 617, + 287, + 666 + ], + "type": "text", + "content": "We conduct several ablation studies to validate the effectiveness of different components in EfficientOCF. All the 3D evaluation metrics are derived by comparing fine-grained predictions with corresponding ground-truth labels." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 666, + 287, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 666, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 666, + 287, + 713 + ], + "type": "text", + "content": "Spatiotemporal decoupling. We refer the ablation study on spatial decoupling to Tab. 1, where EfficientOCF outperforms OCFNet that directly implements occupancy forecasting in dense 3D representation. We further ablate tem" + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 319, + 398, + 426, + 534 + ], + "blocks": [ + { + "bbox": [ + 319, + 398, + 426, + 534 + ], + "lines": [ + { + "bbox": [ + 319, + 398, + 426, + 534 + ], + "spans": [ + { + "bbox": [ + 319, + 398, + 426, + 534 + ], + "type": "image", + "image_path": "3db5ca1f30a3bf59e2167de13e477707ccf2b2431ce4429309a536548351143e.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 535, + 545, + 557 + ], + "lines": [ + { + "bbox": [ + 305, + 535, + 545, + 557 + ], + "spans": [ + { + "bbox": [ + 305, + 535, + 545, + 557 + ], + "type": "text", + "content": "Figure 7. Visualization of TP (green), FP (red), and FN (yellow) results against GT at continuous timesteps." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 426, + 398, + 533, + 534 + ], + "blocks": [ + { + "bbox": [ + 426, + 398, + 533, + 534 + ], + "lines": [ + { + "bbox": [ + 426, + 398, + 533, + 534 + ], + "spans": [ + { + "bbox": [ + 426, + 398, + 533, + 534 + ], + "type": "image", + "image_path": "d9ab4dd03786b1bbd83ea98ed37c06fc78d4f015cccc3fb3bdf406cde48cf232.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 566, + 545, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 566, + 545, + 686 + ], + "spans": [ + { + "bbox": [ + 304, + 566, + 545, + 686 + ], + "type": "text", + "content": "poral decoupling by removing the instance-aware refinement in EfficientOCF shown in Tab. 3. We conduct a baseline EfficientOCF" + }, + { + "bbox": [ + 304, + 566, + 545, + 686 + ], + "type": "inline_equation", + "content": "\\ddagger" + }, + { + "bbox": [ + 304, + 566, + 545, + 686 + ], + "type": "text", + "content": " that directly lifts the initial 2D OCF results with height values. As shown, our instance-aware temporal decoupling improves the initial OCF performance of the prediction module by absolute 1.64 and 2.55 percentage points in 3D " + }, + { + "bbox": [ + 304, + 566, + 545, + 686 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathrm{IoU}}_{\\mathrm{f}}" + }, + { + "bbox": [ + 304, + 566, + 545, + 686 + ], + "type": "text", + "content": " and C-" + }, + { + "bbox": [ + 304, + 566, + 545, + 686 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathrm{IoU}}_{\\mathrm{f}}" + }, + { + "bbox": [ + 304, + 566, + 545, + 686 + ], + "type": "text", + "content": " on the nuScenes-Occupancy dataset. This demonstrates that instance forecasting encompassing flow-based association provides the more accurate occupancy estimation for each decoupled timestep." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 305, + 689, + 545, + 713 + ], + "type": "text", + "content": "Adaptive dual pooling. We compare our adaptive dual pooling (ADP) to different pooling strategies that compress" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22344" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 48, + 83, + 286, + 131 + ], + "blocks": [ + { + "bbox": [ + 90, + 71, + 243, + 81 + ], + "lines": [ + { + "bbox": [ + 90, + 71, + 243, + 81 + ], + "spans": [ + { + "bbox": [ + 90, + 71, + 243, + 81 + ], + "type": "text", + "content": "Table 3. Ablation on temporal decoupling" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 83, + 286, + 131 + ], + "lines": [ + { + "bbox": [ + 48, + 83, + 286, + 131 + ], + "spans": [ + { + "bbox": [ + 48, + 83, + 286, + 131 + ], + "type": "table", + "html": "
Approach2D3D
IoUcIoUfIoUfIoUcIoUfIoUfC-IoUcC-IoUfC-IoUf
EfficientOCF‡39.4435.8936.6119.9817.3117.8645.5242.9043.44
EfficientOCF39.9336.1536.9321.2819.0219.5047.5345.5745.99
", + "image_path": "77bae1b4d06e72888f4c44423c59a248a8d92c58d460562af38aac04f560dc10.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 48, + 146, + 286, + 201 + ], + "blocks": [ + { + "bbox": [ + 89, + 134, + 246, + 144 + ], + "lines": [ + { + "bbox": [ + 89, + 134, + 246, + 144 + ], + "spans": [ + { + "bbox": [ + 89, + 134, + 246, + 144 + ], + "type": "text", + "content": "Table 4. Ablation on adaptive dual pooling" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 146, + 286, + 201 + ], + "lines": [ + { + "bbox": [ + 48, + 146, + 286, + 201 + ], + "spans": [ + { + "bbox": [ + 48, + 146, + 286, + 201 + ], + "type": "table", + "html": "
Approach2D3D
IoUcIoUfIoUfIoUcIoUfIoUfC-IoUcC-IoUfC-IoUf
Average Pooling38.0434.7935.4519.2916.7717.2944.0241.7142.18
Max Pooling38.8435.5336.2019.4216.9017.4144.6942.3342.81
ADP39.9336.1536.9321.2819.0219.5047.5345.5745.99
", + "image_path": "b7767f612bc3924732130b66939cdd611c29774f21c80948fb2ffa1d894277ad.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 48, + 215, + 286, + 270 + ], + "blocks": [ + { + "bbox": [ + 71, + 202, + 263, + 213 + ], + "lines": [ + { + "bbox": [ + 71, + 202, + 263, + 213 + ], + "spans": [ + { + "bbox": [ + 71, + 202, + 263, + 213 + ], + "type": "text", + "content": "Table 5. Ablation on segmentation/flow/height heads" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 48, + 215, + 286, + 270 + ], + "lines": [ + { + "bbox": [ + 48, + 215, + 286, + 270 + ], + "spans": [ + { + "bbox": [ + 48, + 215, + 286, + 270 + ], + "type": "table", + "html": "
Seg. \nHeadHeight \nHeadFlow \nHead2D3D
IoUcIoUfIoUfIoUcIoUfIoUfC-IoUcC-IoUfC-IoUf
✓38.4835.2635.93------
✓✓38.5535.3335.9819.4116.8917.4144.0941.8342.29
✓✓✓39.9336.1536.9321.2819.0219.5047.5345.5745.99
", + "image_path": "ca72ae52ea4a3e9e075347d3404c849b8be63b1fb57ba3f5e342c772944d85ad.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 46, + 277, + 286, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 277, + 286, + 433 + ], + "spans": [ + { + "bbox": [ + 46, + 277, + 286, + 433 + ], + "type": "text", + "content": "3D features to 2D space. Tab. 4 shows that EfficientOCF with adaptive weighting for the combination of average and max pooling outperforms the baseline with only one single pooling strategy for 3D-2D transformation. The max pooling contributes more to forecasting accuracy than the average pooling, likely because height values and max-pooled features both capture prominent structural changes along the z-axis. We also deploy our ADP strategy as the 3D-2D transformation module in PowerBEV, and it enhances the performance of the vanilla PowerBEV in 2D " + }, + { + "bbox": [ + 46, + 277, + 286, + 433 + ], + "type": "inline_equation", + "content": "\\mathrm{IoU_c}" + }, + { + "bbox": [ + 46, + 277, + 286, + 433 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 277, + 286, + 433 + ], + "type": "inline_equation", + "content": "\\mathrm{IoU_f}" + }, + { + "bbox": [ + 46, + 277, + 286, + 433 + ], + "type": "text", + "content": " (2.0s) by absolute 5.16 and 5.42 percentage points. This validates that other OCF methods can also benefit from our proposed ADP strategy." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 434, + 286, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 434, + 286, + 566 + ], + "spans": [ + { + "bbox": [ + 46, + 434, + 286, + 566 + ], + "type": "text", + "content": "Multiple heads. We ablate the devised heads in the prediction module of EfficientOCF in Tab. 5. The inclusion of the height head allows the transformation of forecasting results from 2D to 3D space, and provides more structure-aware supervision signals. It thus significantly improves OCF performance on all the metrics. Furthermore, the introduction of the flow head also provides additional enhancements because flow prediction enables the model to be aware of the concrete motion pattern of each voxel grid. Besides, the flow head helps to refine the initial occupancy following our proposed temporal decoupling." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 575, + 217, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 575, + 217, + 586 + ], + "spans": [ + { + "bbox": [ + 47, + 575, + 217, + 586 + ], + "type": "text", + "content": "4.4. Instance Prediction Assessment" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "type": "text", + "content": "This experiment presents the 3D instance prediction performance of SOTA OCFNet and our EfficientOCF with the above-mentioned VPQ metric. As shown in Tab. 6, EfficientOCF improves the VPQ of OCFNet forecasting by absolute 1.42 and 4.08 percentage points in " + }, + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{VPQ^{bb}}" + }, + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{VPQ^{fg}}" + }, + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "type": "text", + "content": " respectively on nuScenes, and by 8.40 percentage points on Lyft-Level5. EfficientOCF" + }, + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "type": "inline_equation", + "content": "^{-}" + }, + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "type": "text", + "content": " still outperforms OCFNet" + }, + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "type": "inline_equation", + "content": "^{-}" + }, + { + "bbox": [ + 46, + 594, + 286, + 713 + ], + "type": "text", + "content": " for both datasets. EfficientOCF forecasts instances in 2D space which are further lifted to 3D counterparts following our proposed spatiotemporal decoupling. OCFNet directly im" + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 310, + 83, + 538, + 152 + ], + "blocks": [ + { + "bbox": [ + 342, + 71, + 509, + 81 + ], + "lines": [ + { + "bbox": [ + 342, + 71, + 509, + 81 + ], + "spans": [ + { + "bbox": [ + 342, + 71, + 509, + 81 + ], + "type": "text", + "content": "Table 6. Evaluation on 3D instance prediction" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 310, + 83, + 538, + 152 + ], + "lines": [ + { + "bbox": [ + 310, + 83, + 538, + 152 + ], + "spans": [ + { + "bbox": [ + 310, + 83, + 538, + 152 + ], + "type": "table", + "html": "
ApproachnuScenesLyft-Level5
VPQbbVPQfgVPQBb
OCFNet-13.684.2524.11
OCFNet18.536.6128.35
EfficientOCF-(ours)14.116.5027.06
EfficientOCF(ours)19.9510.6936.75
", + "image_path": "1d084ce9ee23c558a7a166d0f76de9d3c827fa49697f0d0ef5226646ef5e6ab4.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 310, + 167, + 537, + 213 + ], + "blocks": [ + { + "bbox": [ + 355, + 155, + 496, + 165 + ], + "lines": [ + { + "bbox": [ + 355, + 155, + 496, + 165 + ], + "spans": [ + { + "bbox": [ + 355, + 155, + 496, + 165 + ], + "type": "text", + "content": "Table 7. Comparison of inference time" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 310, + 167, + 537, + 213 + ], + "lines": [ + { + "bbox": [ + 310, + 167, + 537, + 213 + ], + "spans": [ + { + "bbox": [ + 310, + 167, + 537, + 213 + ], + "type": "table", + "html": "
ApproachInference time [ms]
PerceptionProcessingPredictionTotal
OCFNet107.054.1813.15124.38
EfficientOCF (ours)64.785.0512.5082.33
", + "image_path": "2bf19c81d2c601ace7f3a6e292084a690819ab1c8ab6fc09a3f8c175a8f27c21.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 305, + 220, + 545, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 220, + 545, + 243 + ], + "spans": [ + { + "bbox": [ + 305, + 220, + 545, + 243 + ], + "type": "text", + "content": "plements instance association in dense 3D formats, leading to more prediction artifacts and uncertainties." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 306, + 251, + 399, + 262 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 251, + 399, + 262 + ], + "spans": [ + { + "bbox": [ + 306, + 251, + 399, + 262 + ], + "type": "text", + "content": "4.5. Inference Time" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 269, + 545, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 269, + 545, + 424 + ], + "spans": [ + { + "bbox": [ + 304, + 269, + 545, + 424 + ], + "type": "text", + "content": "We compare the runtime of OCFNet and our EfficientOCF on 3D OCF shown in Tab. 7. As seen, the processing time of EfficientOCF is slightly higher than OCFNet due to the additional 3D-2D transformation, while EfficientOCF achieves a faster overall inference speed " + }, + { + "bbox": [ + 304, + 269, + 545, + 424 + ], + "type": "inline_equation", + "content": "(82.33\\mathrm{ms} / 12.15\\mathrm{Hz})" + }, + { + "bbox": [ + 304, + 269, + 545, + 424 + ], + "type": "text", + "content": " compared to OCFNet. Moreover, we find that EfficientOCF reduces memory consumption by " + }, + { + "bbox": [ + 304, + 269, + 545, + 424 + ], + "type": "inline_equation", + "content": "478\\mathrm{MB}" + }, + { + "bbox": [ + 304, + 269, + 545, + 424 + ], + "type": "text", + "content": " compared to OCFNet. This is because EfficientOCF utilizes lightweight 2D encoders and decoders in the perception module to achieve 3D occupancy forecasting by lifting 2D results to 3D space using height values. In contrast, OCFNet uses heavy 3D encoders and decoders with dense 3D representation, leading to more inference time." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 306, + 435, + 378, + 447 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 435, + 378, + 447 + ], + "spans": [ + { + "bbox": [ + 306, + 435, + 378, + 447 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 455, + 545, + 706 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 455, + 545, + 706 + ], + "spans": [ + { + "bbox": [ + 304, + 455, + 545, + 706 + ], + "type": "text", + "content": "This paper proposes spatiotemporal decoupling for efficient vision-based occupancy forecasting (OCF). We achieve spatial decoupling by transforming 3D dense representations into compact 2D BEV occupancy with height values. We also propose using temporal decoupling to refine the end-to-end OCF by step-wise instance forecasting along the time axis. Based on our proposed spatiotemporally decoupled representation, we further develop EfficientOCF, a lightweight network that concurrently forecasts future 2D occupancy states and height values of general movable objects, effectively bridging 2D predictions into the 3D space. It also refines initial OCF results with instance forecasting at each individual timestep. Additionally, we introduce new evaluation metrics providing a comprehensive assessment of OCF performance for datasets with incomplete annotations. Our experiments on extensive evaluation metrics demonstrate that EfficientOCF outperforms existing methods in both accuracy and efficiency, achieving SOTA results in both 2D and 3D spaces. We hope that the spatiotemporal decoupling pipeline proposed in this work can be used as a foundation paradigm in the field of vision-based OCF." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "22345" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "spans": [ + { + "bbox": [ + 48, + 71, + 106, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 91, + 288, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "spans": [ + { + "bbox": [ + 53, + 91, + 287, + 135 + ], + "type": "text", + "content": "[1] Adil Kaan Akan and Fatma Güney. Stretchbev: Stretching future instance prediction spatially and temporally. In European Conference on Computer Vision, pages 444-460. Springer, 2022. 1, 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 53, + 137, + 288, + 202 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 137, + 288, + 202 + ], + "spans": [ + { + "bbox": [ + 53, + 137, + 288, + 202 + ], + "type": "text", + "content": "[2] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11621-11631, 2020. 2, 3, 5, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 53, + 205, + 288, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 205, + 288, + 248 + ], + "spans": [ + { + "bbox": [ + 53, + 205, + 288, + 248 + ], + "type": "text", + "content": "[3] Anh-Quan Cao and Raoul De Charette. Monoscene: Monocular 3d semantic scene completion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3991–4001, 2022. 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 250, + 287, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 250, + 287, + 294 + ], + "spans": [ + { + "bbox": [ + 53, + 250, + 287, + 294 + ], + "type": "text", + "content": "[4] Sergio Casas, Abbas Sadat, and Raquel Urtasun. Mp3: A unified model to map, perceive, predict and plan. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14403-14412, 2021. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 295, + 287, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 295, + 287, + 328 + ], + "spans": [ + { + "bbox": [ + 53, + 295, + 287, + 328 + ], + "type": "text", + "content": "[5] Hehe Fan and Yi Yang. Point recurrent neural network for moving point cloud processing. arXiv preprint arXiv:1910.08287, 2019. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 330, + 287, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 330, + 287, + 374 + ], + "spans": [ + { + "bbox": [ + 53, + 330, + 287, + 374 + ], + "type": "text", + "content": "[6] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 376, + 287, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 376, + 287, + 419 + ], + "spans": [ + { + "bbox": [ + 53, + 376, + 287, + 419 + ], + "type": "text", + "content": "[7] Yu Hong, Qian Liu, Huayuan Cheng, Danjiao Ma, Hang Dai, Yu Wang, Guangzhi Cao, and Yong Ding. Univision: A unified framework for vision-centric 3d perception. arXiv preprint arXiv:2401.06994, 2024. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 421, + 287, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 421, + 287, + 487 + ], + "spans": [ + { + "bbox": [ + 53, + 421, + 287, + 487 + ], + "type": "text", + "content": "[8] Anthony Hu, Zak Murez, Nikhil Mohan, Sofia Dudas, Jeffrey Hawke, Vijay Badrinarayanan, Roberto Cipolla, and Alex Kendall. Fiery: Future instance prediction in bird's-eye view from surround monocular cameras. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15273-15282, 2021. 1, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 488, + 288, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 488, + 288, + 544 + ], + "spans": [ + { + "bbox": [ + 53, + 488, + 288, + 544 + ], + "type": "text", + "content": "[9] Yuanhui Huang, Wenzhao Zheng, Yunpeng Zhang, Jie Zhou, and Jiwen Lu. Tri-perspective view for vision-based 3d semantic occupancy prediction. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9223-9232, 2023. 1, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 545, + 287, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 545, + 287, + 600 + ], + "spans": [ + { + "bbox": [ + 48, + 545, + 287, + 600 + ], + "type": "text", + "content": "[10] Yuanhui Huang, Wenzhao Zheng, Borui Zhang, Jie Zhou, and Jiwen Lu. Selfocc: Self-supervised vision-based 3d occupancy prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19946-19956, 2024. 1, 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 601, + 287, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 601, + 287, + 656 + ], + "spans": [ + { + "bbox": [ + 48, + 601, + 287, + 656 + ], + "type": "text", + "content": "[11] R. Kesten, M. Usman, J. Houston, T. Pandya, K. Nadhamuni, A. Ferreira, M. Yuan, B. Low, A. Jain, P. Ondruska, S. Omari, S. Shah, A. Kulkarni, A. Kazakova, C. Tao, L. Platinsky, W. Jiang, and V. Shet. Lyft level 5 perception dataset 2020, 2019, 2, 5, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 287, + 712 + ], + "type": "text", + "content": "[12] Tarasha Khurana, Peiyun Hu, David Held, and Deva Ramanan. Point cloud forecasting as a proxy for 4d occupancy forecasting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1116-1124, 2023. 1" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 714 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 307, + 73, + 547, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 73, + 547, + 117 + ], + "spans": [ + { + "bbox": [ + 307, + 73, + 547, + 117 + ], + "type": "text", + "content": "[13] Dahun Kim, Sanghyun Woo, Joon-Young Lee, and In So Kweon. Video panoptic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9859–9868, 2020. 5, 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 118, + 547, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 118, + 547, + 149 + ], + "spans": [ + { + "bbox": [ + 307, + 118, + 547, + 149 + ], + "type": "text", + "content": "[14] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 308, + 151, + 547, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 151, + 547, + 215 + ], + "spans": [ + { + "bbox": [ + 308, + 151, + 547, + 215 + ], + "type": "text", + "content": "[15] Peizheng Li, Shuxiao Ding, Xieyuanli Chen, Niklas Hanselmann, Marius Cordts, and Juergen Gall. Powerbev: a powerful yet lightweight framework for instance prediction in bird's-eye view. In Proceedings of the Thirty-Second International Joint Conference on Artificial Intelligence, pages 1080–1088, 2023. 1, 3, 4, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 308, + 217, + 547, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 217, + 547, + 271 + ], + "spans": [ + { + "bbox": [ + 308, + 217, + 547, + 271 + ], + "type": "text", + "content": "[16] Yinhao Li, Zheng Ge, Guanyi Yu, Jinrong Yang, Zengran Wang, Yukang Shi, Jianjian Sun, and Zeming Li. Bevdepth: Acquisition of reliable depth for multi-view 3d object detection. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1477-1485, 2023. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 308, + 272, + 547, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 272, + 547, + 338 + ], + "spans": [ + { + "bbox": [ + 308, + 272, + 547, + 338 + ], + "type": "text", + "content": "[17] Yiming Li, Zhiding Yu, Christopher Choy, Chaowei Xiao, Jose M Alvarez, Sanja Fidler, Chen Feng, and Anima Anandkumar. Voxformer: Sparse voxel transformer for camera-based 3d semantic scene completion. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9087-9098, 2023. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 338, + 546, + 393 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 338, + 546, + 393 + ], + "spans": [ + { + "bbox": [ + 308, + 338, + 546, + 393 + ], + "type": "text", + "content": "[18] Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chonghao Sima, Tong Lu, Yu Qiao, and Jifeng Dai. Bevformer: Learning bird's-eye-view representation from multi-camera images via spatiotemporal transformers. In European conference on computer vision, pages 1-18. Springer, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 308, + 394, + 546, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 394, + 546, + 446 + ], + "spans": [ + { + "bbox": [ + 308, + 394, + 546, + 446 + ], + "type": "text", + "content": "[19] Haochen Liu, Zhiyu Huang, and Chen Lv. Multi-modal hierarchical transformer for occupancy flow field prediction in autonomous driving. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 1449-1455. IEEE, 2023. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 449, + 546, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 449, + 546, + 503 + ], + "spans": [ + { + "bbox": [ + 308, + 449, + 546, + 503 + ], + "type": "text", + "content": "[20] Jian Liu, Sipeng Zhang, Chuixin Kong, Wenyuan Zhang, Yuhang Wu, Yikang Ding, Borun Xu, Ruibo Ming, Donglai Wei, and Xianming Liu. Occtransformer: Improving bevformer for 3d camera-only occupancy prediction. arXiv preprint arXiv:2402.18140, 2024. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 504, + 546, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 504, + 546, + 547 + ], + "spans": [ + { + "bbox": [ + 308, + 504, + 546, + 547 + ], + "type": "text", + "content": "[21] Xinhao Liu, Moonjun Gong, Qi Fang, Haoyu Xie, Yiming Li, Hang Zhao, and Chen Feng. Lidar-based 4d occupancy completion and forecasting. arXiv preprint arXiv:2310.11239, 2023. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 548, + 546, + 603 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 548, + 546, + 603 + ], + "spans": [ + { + "bbox": [ + 308, + 548, + 546, + 603 + ], + "type": "text", + "content": "[22] Zhijian Liu, Haotian Tang, Alexander Amini, Xinyu Yang, Huizi Mao, Daniela L Rus, and Song Han. Bevfusion: Multi-task multi-sensor fusion with unified bird's-eye view representation. In 2023 IEEE international conference on robotics and automation (ICRA), pages 2774-2781. IEEE, 2023. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 603, + 546, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 603, + 546, + 647 + ], + "spans": [ + { + "bbox": [ + 308, + 603, + 546, + 647 + ], + "type": "text", + "content": "[23] Fan Lu, Guang Chen, Zhijun Li, Lijun Zhang, Yinlong Liu, Sanqing Qu, and Alois Knoll. Monet: Motion-based point cloud prediction network. IEEE Transactions on Intelligent Transportation Systems, 23(8):13794-13804, 2021. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 647, + 546, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 647, + 546, + 690 + ], + "spans": [ + { + "bbox": [ + 308, + 647, + 546, + 690 + ], + "type": "text", + "content": "[24] Zhen Luo, Junyi Ma, Zijie Zhou, and Guangming Xiong. Pcpcnet: An efficient and semantic-enhanced transformer network for point cloud prediction. IEEE Robotics and Automation Letters, 8(7):4267-4274, 2023. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 691, + 546, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 691, + 546, + 714 + ], + "spans": [ + { + "bbox": [ + 308, + 691, + 546, + 714 + ], + "type": "text", + "content": "[25] Junyi Ma, Xieyuanli Chen, Jiawei Huang, Jingyi Xu, Zhen Luo, Jintao Xu, Weihao Gu, Rui Ai, and Hesheng Wang." + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "22346" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 66, + 72, + 287, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 72, + 287, + 127 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 287, + 127 + ], + "type": "text", + "content": "Cam4docc: Benchmark for camera-only 4d occupancy forecasting in autonomous driving applications. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21486-21495, 2024. 1, 2, 3, 4, 5, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 48, + 129, + 287, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 129, + 287, + 173 + ], + "spans": [ + { + "bbox": [ + 48, + 129, + 287, + 173 + ], + "type": "text", + "content": "[26] Reza Mahjourian, Jinkyu Kim, Yuning Chai, Mingxing Tan, Ben Sapp, and Dragomir Anguelov. Occupancy flow fields for motion forecasting in autonomous driving. IEEE Robotics and Automation Letters, 7(2):5639-5646, 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 175, + 287, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 175, + 287, + 218 + ], + "spans": [ + { + "bbox": [ + 48, + 175, + 287, + 218 + ], + "type": "text", + "content": "[27] Benedikt Mersch, Xieyuanli Chen, Jens Behley, and Cyril Stachniss. Self-supervised point cloud prediction using 3d spatio-temporal convolutional networks. In Conference on Robot Learning, pages 1444-1454. PMLR, 2022. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 220, + 287, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 220, + 287, + 262 + ], + "spans": [ + { + "bbox": [ + 48, + 220, + 287, + 262 + ], + "type": "text", + "content": "[28] Kaustab Pal, Aditya Sharma, Avinash Sharma, and K Madhava Krishna. Atppnet: Attention based temporal point cloud prediction network. arXiv preprint arXiv:2401.17399, 2024. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 264, + 287, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 264, + 287, + 319 + ], + "spans": [ + { + "bbox": [ + 48, + 264, + 287, + 319 + ], + "type": "text", + "content": "[29] Mingjie Pan, Li Liu, Jiaming Liu, Peixiang Huang, Longlong Wang, Shanghai Zhang, Shaoqing Xu, Zhiyi Lai, and Kuiyuan Yang. Uniocc: Unifying vision-centric 3d occupancy prediction with geometric and semantic rendering. arXiv preprint arXiv:2306.09117, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 320, + 287, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 320, + 287, + 376 + ], + "spans": [ + { + "bbox": [ + 48, + 320, + 287, + 376 + ], + "type": "text", + "content": "[30] Jonah Philion and Sanja Fidler. Lift, splat, shoot: Encoding images from arbitrary camera rigs by implicitly unprojecting to 3d. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XIV 16, pages 194-210. Springer, 2020. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 377, + 287, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 377, + 287, + 432 + ], + "spans": [ + { + "bbox": [ + 48, + 377, + 287, + 432 + ], + "type": "text", + "content": "[31] Xiaoyu Tian, Tao Jiang, Longfei Yun, Yucheng Mao, Huitong Yang, Yue Wang, Yilun Wang, and Hang Zhao. Occ3d: A large-scale 3d occupancy prediction benchmark for autonomous driving. Advances in Neural Information Processing Systems, 36, 2024. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 434, + 287, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 434, + 287, + 454 + ], + "spans": [ + { + "bbox": [ + 48, + 434, + 287, + 454 + ], + "type": "text", + "content": "[32] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 456, + 287, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 456, + 287, + 510 + ], + "spans": [ + { + "bbox": [ + 48, + 456, + 287, + 510 + ], + "type": "text", + "content": "[33] Lening Wang, Wenzhao Zheng, Yilong Ren, Han Jiang, Zhiyong Cui, Haiyang Yu, and Jiwen Lu. Occsora: 4d occupancy generation models as world simulators for autonomous driving. arXiv preprint arXiv:2405.20337, 2024. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 512, + 287, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 512, + 287, + 578 + ], + "spans": [ + { + "bbox": [ + 48, + 512, + 287, + 578 + ], + "type": "text", + "content": "[34] Xiaofeng Wang, Zheng Zhu, Wenbo Xu, Yunpeng Zhang, Yi Wei, Xu Chi, Yun Ye, Dalong Du, Jiwen Lu, and Xingang Wang. Openoccupancy: A large scale benchmark for surrounding semantic occupancy perception. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17850-17859, 2023. 1, 2, 5, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 579, + 287, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 579, + 287, + 634 + ], + "spans": [ + { + "bbox": [ + 48, + 579, + 287, + 634 + ], + "type": "text", + "content": "[35] Yi Wei, Linqing Zhao, Wenzhao Zheng, Zheng Zhu, Jie Zhou, and Jiwen Lu. Surroundocc: Multi-camera 3d occupancy prediction for autonomous driving. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 21729-21740, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 635, + 287, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 635, + 287, + 690 + ], + "spans": [ + { + "bbox": [ + 48, + 635, + 287, + 690 + ], + "type": "text", + "content": "[36] Pengxiang Wu, Siheng Chen, and Dimitris N Metaxas. Motionnet: Joint perception and motion prediction for autonomous driving based on bird's eye view maps. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11385-11395, 2020. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 287, + 713 + ], + "type": "text", + "content": "[37] Qizhe Xie, Minh-Thang Luong, Eduard Hovy, and Quoc V Le. Self-training with noisy student improvesImagenet" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 545, + 262 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 327, + 73, + 545, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 545, + 106 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 545, + 106 + ], + "type": "text", + "content": "classification. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10687-10698, 2020. 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 307, + 107, + 545, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 545, + 161 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 545, + 161 + ], + "type": "text", + "content": "[38] Yunpeng Zhang, Zheng Zhu, Wenzhao Zheng, Junjie Huang, Guan Huang, Jie Zhou, and Jiwen Lu. Beverse: Unified perception and prediction in birds-eye-view for vision-centric autonomous driving. arXiv preprint arXiv:2205.09743, 2022.1,2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 307, + 163, + 545, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 163, + 545, + 217 + ], + "spans": [ + { + "bbox": [ + 307, + 163, + 545, + 217 + ], + "type": "text", + "content": "[39] Yunpeng Zhang, Zheng Zhu, and Dalong Du. Occformer: Dual-path transformer for vision-based 3d semantic occupancy prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9433-9443, 2023. 1, 2, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 307, + 219, + 545, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 219, + 545, + 262 + ], + "spans": [ + { + "bbox": [ + 307, + 219, + 545, + 262 + ], + "type": "text", + "content": "[40] Wenzhao Zheng, Weiliang Chen, Yuanhui Huang, Borui Zhang, Yueqi Duan, and Jiwen Lu. Occworld: Learning a 3d occupancy world model for autonomous driving. arXiv preprint arXiv:2311.16038, 2023. 3" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22347" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Spatiotemporal Skip Guidance for Enhanced Video Diffusion Sampling/4ec7da94-3031-45c8-a8d6-19bf226e3cba_content_list.json b/2025/Spatiotemporal Skip Guidance for Enhanced Video Diffusion Sampling/4ec7da94-3031-45c8-a8d6-19bf226e3cba_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..d551cdf917af3196557f2c7b12cc18ba2248a45d --- /dev/null +++ b/2025/Spatiotemporal Skip Guidance for Enhanced Video Diffusion Sampling/4ec7da94-3031-45c8-a8d6-19bf226e3cba_content_list.json @@ -0,0 +1,1804 @@ +[ + { + "type": "text", + "text": "Spatiotemporal Skip Guidance for Enhanced Video Diffusion Sampling", + "text_level": 1, + "bbox": [ + 138, + 130, + 859, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Junha Hyung $^{*1}$ Kinam Kim $^{*1}$ Susung Hong $^{2}$ Min-Jung Kim $^{1}$ Jaegul Choo $^{1}$", + "bbox": [ + 169, + 179, + 823, + 200 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ KAIST AI $^{2}$ University of Washington", + "bbox": [ + 334, + 214, + 661, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{sharpeeee, kinamplify, emjay73, jchoo}@kaist.ac.kr, susung@cs.washington.edu", + "bbox": [ + 156, + 236, + 836, + 252 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "\"A close-up shot of a butterfly landing on the nose of a woman, highlighting her smile and the details of the butterfly's wings.\"", + "bbox": [ + 112, + 664, + 480, + 694 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "\"A close-up of a woman's face with colored powder exploding around her, creating an abstract splash of vibrant hues.\"", + "bbox": [ + 524, + 664, + 880, + 693 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Figure 1. Visual comparison of video quality between CFG (top row) and our STG method (bottom row). Best viewed in Acrobat Reader; click on the images to watch the videos.", + "bbox": [ + 88, + 705, + 906, + 733 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 279, + 753, + 361, + 771 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Diffusion models have emerged as a powerful tool for generating high-quality images, videos, and 3D content. While sampling guidance techniques like CFG improve quality, they reduce diversity and motion. Autoguidance mitigates these issues but demands extra weak model training, limiting its practicality for large-scale models. In this", + "bbox": [ + 89, + 787, + 485, + 878 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "work, we introduce Spatiotemporal Skip Guidance (STG), a simple training-free sampling guidance method for enhancing transformer-based video diffusion models. STG employs an implicit weak model via self-perturbation, avoiding the need for external models or additional training. By selectively skipping spatiotemporal layers, STG produces an aligned, degraded version of the original model to boost sample quality without compromising diversity or dynamic degree. Our contributions include: (1) introducing STG as", + "bbox": [ + 511, + 756, + 908, + 893 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* indicates equal contribution.", + "bbox": [ + 114, + 888, + 277, + 901 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "11006", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "an efficient, high-performing guidance technique for video diffusion models, (2) eliminating the need for auxiliary models by simulating a weak model through layer skipping, and (3) ensuring quality-enhanced guidance without compromising sample diversity or dynamics unlike CFG. For additional results, visit https://junhahyung.github.io/STGuidance/.", + "bbox": [ + 89, + 90, + 483, + 196 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 224, + 222, + 242 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Diffusion models [11, 30-32] are a successful class of generative models known for their flexibility in modeling complex data distributions, achieving impressive results in image, video, and 3D generation. By progressively denoising random noise, they enable robust generalization, making them a leading choice for realistic content generation and often surpassing GAN-based methods [5, 9, 18, 29]. Building on this success, video diffusion models [4, 12, 27, 33, 38] generate high-quality videos by using temporal or 3D attention layers to handle sequential frames.", + "bbox": [ + 89, + 250, + 482, + 402 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Meanwhile, to enhance sample quality, sampling guidance techniques such as Classifier-Free Guidance (CFG) [10] and Autoguidance [19] have been introduced to guide the denoising process. These techniques employ weak models to predict poor trajectories, steering the main model away from them and pushing samples toward high-quality regions on the data manifold. However, CFG often reduces diversity, leading to saturated or overly simplified results [6, 19].", + "bbox": [ + 89, + 402, + 482, + 537 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Autoguidance [19] addresses this issue by using a weak model trained on the same task, conditioning, and data distribution as the main model. The main drawback of this approach, however, is the need to train an additional weak model, which is impractical for large-scale models. Alternative methods such as PAG [1] and SEG [13] use self-perturbation to implicitly mimic a weak model, avoiding the need for extra training. Yet, these methods are designed specifically for image generation diffusion models, applying self-perturbation to 2D spatial attention maps.", + "bbox": [ + 89, + 537, + 482, + 688 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we propose Spatiotemporal Skip Guidance (STG), a simple and effective sampling guidance method for video diffusion models that significantly enhances the performance of any transformer-based video diffusion model without additional training. Specifically, we use an implicit weak model for guidance through self-perturbation, eliminating the need for explicit weak models and their associated training costs. This is especially crucial for video diffusion models, where training costs are high.", + "bbox": [ + 89, + 688, + 482, + 824 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our implicit weak model, a key component of our framework, is deliberately designed as a degraded but \"aligned\" version of the original video generation model. As demonstrated in Autoguidance [19], having the weak model share the same task, conditioning, and data distribution as the", + "bbox": [ + 89, + 825, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/a9ff040f709f6d6ceae9ae479c4eebc42adde4b366b565ff287709d970e93b04.jpg", + "image_caption": [ + "CFG" + ], + "image_footnote": [], + "bbox": [ + 524, + 89, + 725, + 222 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/4437acdf9142fc0f4a69a3b9477ef08b528ab6cae40df116cb29b121c290f3cb.jpg", + "image_caption": [ + "Ours", + "Figure 2. Comparison between CFG and STG, with the band conceptually representing the noisy data manifold. In STG, the weak model and the main model are aligned along the direction of increasing quality. In contrast, the two models in CFG differ not only in quality but also in aspects such as diversity and prompt alignment capabilities." + ], + "image_footnote": [], + "bbox": [ + 756, + 89, + 900, + 222 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "main model is essential for quality improvement, as we expect both models to exhibit similar, aligned errors in the same parts of samples. We conceptually illustrate this alignment in Fig. 2, where our weak model and main model are arranged in a direction of increasing quality. To achieve this, we apply spatiotemporal perturbation to both spatial and temporal attention layers—or, in the case of 3D attention, to the entire layer—by selectively skipping certain layers. This straightforward approach effectively nullifies specific residual or attention layers, generating a lower-quality version of the model that simulates an aligned weak model.", + "bbox": [ + 511, + 363, + 905, + 529 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To ensure samples remain on the data manifold, even when using large guidance scales, we employ optional techniques such as the rescaling [21] and restart [36] methods. Rescaling the latent code constrains its variance, addressing the issue of larger variance causing saturation in the results [21]. Meanwhile, the restart sampling technique leverages the error contraction property of forward SDEs [36] to keep the sampling trajectory on the manifold. These techniques help prevent overshooting in the sampling trajectory, which could otherwise push samples off the manifold and result in saturated or distorted outputs.", + "bbox": [ + 511, + 541, + 908, + 708 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our key contributions are as follows:", + "bbox": [ + 531, + 719, + 777, + 734 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose STG—a surprisingly simple sampling guidance framework that significantly boosts the performance of video diffusion models.", + "- Our method introduces an implicit weak model by skipping spatiotemporal layers in video diffusion models, eliminating the need for additional training or external models.", + "- Our method enhances sample quality during guidance without reducing diversity or limiting the dynamics of generated videos." + ], + "bbox": [ + 511, + 746, + 903, + 897 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "11007", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 91, + 89, + 233, + 104 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Guidance with trained weak model Classifier-Free Guidance (CFG) [10] improves conditional generation in Diffusion Models by using an implicit unconditional model as a weak model. However, differences in tasks between the unconditional and conditional models can reduce sample diversity [6, 19] and increase sampling trajectory curvature [6], leading to overshooting the data manifold and producing skewed or oversaturated images.", + "bbox": [ + 89, + 114, + 483, + 236 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Autoguidance [19] mitigates these issues by employing a bad version of the main model as a weak model, trained with reduced capacity and compute to ensure alignment. This alignment allows the guidance algorithm to correct errors by analyzing prediction differences. While effective, it requires additional training, which is challenging for large-scale video diffusion models.", + "bbox": [ + 89, + 237, + 483, + 342 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Guidance with training-free weak model Another line of work avoids additional training by using self-perturbation of the main model to mimic a weak model. Self-Attention Guidance (SAG) [14] blurs high-attention regions, Perturbed Attention Guidance (PAG) [1] replaces attention maps with identity matrices, and Smoothed Energy Guidance (SEG) [13] applies Gaussian blur to the attention weights to smooth the energy landscape. These methods guide sampling toward high-quality outputs by leveraging differences in predictions from the weakened model. While effective, they are primarily designed for image diffusion models with 2D self-attention. We aim to extend this approach to video diffusion models, which require handling temporal dynamics with additional temporal or 3D spatiotemporal attention layers.", + "bbox": [ + 89, + 364, + 483, + 592 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Preliminaries", + "text_level": 1, + "bbox": [ + 89, + 607, + 230, + 622 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Diffusion models", + "text_level": 1, + "bbox": [ + 89, + 632, + 256, + 647 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Diffusion models generate samples by progressively removing noise from noisy data, restoring the original data distribution. Song et al. [32] defined the process of adding noise to the data using a stochastic differential equation (SDE):", + "bbox": [ + 89, + 656, + 483, + 717 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nd x = - \\frac {\\beta (t)}{2} x d t + \\sqrt {\\beta (t)} d w, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 181, + 728, + 480, + 758 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\beta (t)$ is a time-dependent noise schedule, and $w$ represents the standard Wiener process. Corresponding reverse-time SDE is:", + "bbox": [ + 89, + 770, + 483, + 815 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nd x = \\left[ - \\frac {\\beta (t)}{2} x - \\beta (t) \\nabla_ {x} \\log p _ {t} (x) \\right] d t + \\sqrt {\\beta (t)} d \\bar {w} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 825, + 483, + 859 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $d\\bar{w}$ is the Wiener process running backward in time. Here, the score function $\\nabla_x\\log p_t(x)$ is approximated by", + "bbox": [ + 89, + 869, + 483, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "a neural network $s_{\\theta}(x(t))$ trained using denoising score matching [35]:", + "bbox": [ + 511, + 90, + 906, + 121 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\theta^ {*} = \\arg \\min _ {\\theta} \\mathbb {E} _ {t} \\left\\{\\lambda (t) \\mathbb {E} _ {x _ {0}} \\mathbb {E} _ {x _ {t} | x _ {0}} \\right[ \\\\ \\left. \\left\\| s _ {\\theta} \\left(x _ {t}\\right) - \\nabla_ {x _ {t}} \\log p _ {t} \\left(x _ {t} \\mid x _ {0}\\right) \\| _ {2} ^ {2} \\right] \\right\\}. \\tag {3} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 527, + 133, + 906, + 179 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Classifier Guidance", + "text_level": 1, + "bbox": [ + 513, + 186, + 700, + 200 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Classifier Guidance (CG) [25] reformulates the reverse process of a diffusion model by incorporating an external classifier $p_{\\phi}$ as", + "bbox": [ + 511, + 208, + 905, + 255 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {p _ {\\theta}} \\left(x _ {t} | y\\right) \\propto p _ {\\theta} \\left(x _ {t}\\right) p _ {\\phi} \\left(y \\mid x _ {t}\\right) ^ {\\lambda}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 609, + 263, + 906, + 282 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $y$ is the desired class label and $\\lambda$ controls the guidance strength. The score function is then derived as:", + "bbox": [ + 511, + 291, + 905, + 321 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\nabla_ {x _ {t}} \\log \\tilde {p} _ {\\theta} (x _ {t} | y) = \\nabla_ {x _ {t}} \\log p _ {\\theta} (x _ {t}) \\\\ + \\lambda \\nabla_ {x _ {t}} \\log p _ {\\phi} (y | x _ {t}). \\tag {5} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 527, + 335, + 905, + 371 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "By substituting the score function in Eq. 2 with Eq. 5, sampling from the desired class condition $y$ becomes possible.", + "bbox": [ + 511, + 380, + 905, + 411 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.3. Classifier-Free Guidance", + "text_level": 1, + "bbox": [ + 513, + 419, + 741, + 434 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Classifier-Free Guidance (CFG) [10] uses Bayes' rule to replace a classifier-guided score with a linear combination of conditional and unconditional score estimates:", + "bbox": [ + 511, + 441, + 906, + 486 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\epsilon} _ {\\theta} ^ {\\lambda} \\left(x _ {t}\\right) = \\epsilon_ {\\theta} \\left(x _ {t}\\right) + \\lambda \\left(\\epsilon_ {\\theta} \\left(x _ {t}\\right) - \\epsilon_ {\\theta} \\left(x _ {t} \\mid \\phi\\right)\\right). \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 496, + 905, + 513 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "CFG jointly trains the unconditional model $\\epsilon_{\\theta}(x_t|\\phi)$ and the conditional model $\\epsilon_{\\theta}(x_t|c) (= \\epsilon_{\\theta}(x_t))$ within a single model by setting the condition $c$ to a null token $\\phi$ . Using this guided denoising process, the model better captures the conditions and often generates high-fidelity outputs.", + "bbox": [ + 511, + 523, + 905, + 599 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Method", + "text_level": 1, + "bbox": [ + 513, + 612, + 604, + 627 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1. Optimal Weak Model Design", + "text_level": 1, + "bbox": [ + 513, + 636, + 772, + 652 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our goal is to construct an aligned weak model that captures a distribution similar to the original model while generating slightly lower-quality samples. This ensures that the guidance gradient points toward improved quality, as shown in Fig. 2. Misaligned models, as seen in CFG [10], may lead to unintended outcomes like reduced diversity, while Autoguidance [19] requires additional weak model training, making it impractical for large-scale models.", + "bbox": [ + 511, + 657, + 905, + 779 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To address these limitations, we design an implicit weak model by leveraging the original model itself, eliminating the need for external training. This training-free approach ensures computational efficiency and better alignment, as the weak model is derived directly from the original, preserving most network weights.", + "bbox": [ + 511, + 779, + 905, + 869 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To achieve this, we apply perturbation methods directly to the main model's forward pass, creating an implicit weak", + "bbox": [ + 511, + 869, + 905, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "11008", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/c53eada5524d1d2d8f2a9402b231be54c8289affd8ac4b1d675727d8ce3f5235.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelsImaging QualityAesthetic QualityMotion SmoothnessDynamic DegreeTemporal Flickering
Mochi (CFG)0.5240.5070.9850.870.976
Mochi (STG)0.6280.5540.9880.860.978
Open-Sora (CFG)0.5610.4930.9820.9020.975
Open-Sora (STG)0.6060.5090.9870.8950.976
", + "bbox": [ + 112, + 88, + 883, + 179 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/e86e0514a44704a104cc86a9e026a51b07779d48f640e801fca00a85ba7ebee4.jpg", + "table_caption": [ + "Table 1. Quantitative results for Mochi [33] and Open-Sora [38] on VBench [15] T2V benchmarks." + ], + "table_footnote": [], + "table_body": "
ModelsFVD (↓)ISImaging QualityAesthetic QualityMotion SmoothnessDynamic Degree
SVD (CFG)151.338.00.6870.6370.9660.562
SVD (STG)128.738.50.6940.6390.9680.694
", + "bbox": [ + 173, + 218, + 823, + 272 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Table 2. Quantitative results for SVD [4] on FVD, IS, and VBenchmark [15] I2V benchmarks.", + "bbox": [ + 232, + 281, + 761, + 296 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "model. Specifically, we use spatiotemporal perturbation, a key component for aligning the weak model with video diffusion models, which will be elaborated upon in detail.", + "bbox": [ + 89, + 323, + 483, + 369 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Sampling from High Quality Samples", + "text_level": 1, + "bbox": [ + 89, + 380, + 415, + 397 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Similar to CG [25], we define our goal as conditioning the model on an imaginary label $y_{g}$ that represents high-quality samples, leading to the sampling distribution", + "bbox": [ + 89, + 402, + 483, + 450 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {p} _ {\\theta} \\left(x _ {t} \\mid y _ {g}\\right) \\propto p _ {\\theta} \\left(x _ {t}\\right) p _ {\\phi} \\left(y _ {g} \\mid x _ {t}\\right) ^ {w}. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 179, + 460, + 482, + 479 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Using $w > 0$ sharpens the distribution, promoting the generation of high-quality samples. From Eq. 7, the score is derived as", + "bbox": [ + 89, + 492, + 483, + 536 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\nabla_ {x _ {t}} \\log \\tilde {p} _ {\\theta} (x _ {t} | y _ {g}) \\\\ = \\nabla_ {x _ {t}} \\log p _ {\\theta} (x _ {t}) + w \\nabla_ {x _ {t}} \\log p _ {\\phi} (y _ {g} | x _ {t}). \\tag {8} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 106, + 554, + 480, + 590 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Rather than using an external classifier $p_{\\phi}$ , we propose using our model $p_{\\theta}$ as an implicit classifier. We design this classifier to be inversely proportional to the probability of an imaginary \"bad\" label $y_{b}$ , expressed as", + "bbox": [ + 89, + 603, + 483, + 664 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\np _ {\\phi} \\left(y _ {g} \\mid x _ {t}\\right) \\propto \\frac {1}{p _ {\\theta} \\left(y _ {b} \\mid x _ {t}\\right)}. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 204, + 676, + 482, + 709 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Using Bayes' rule, the function can be expressed in terms of the marginal posterior as follows:", + "bbox": [ + 89, + 720, + 483, + 751 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\nabla_ {x _ {t}} \\log \\frac {1}{p _ {\\theta} (y _ {b} | x _ {t})} = \\nabla_ {x _ {t}} \\log \\frac {p _ {\\theta} (x _ {t})}{p _ {\\theta} (y _ {b}) p _ {\\theta} (x _ {t} | y _ {b})} \\\\ = \\nabla_ {x _ {t}} \\left(\\log p _ {\\theta} \\left(x _ {t}\\right) - \\log p _ {\\theta} \\left(x _ {t} \\mid y _ {b}\\right)\\right), \\tag {10} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 130, + 763, + 482, + 816 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "leading to the score of the target distribution as:", + "bbox": [ + 89, + 829, + 406, + 844 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\nabla_ {x _ {t}} \\log \\tilde {p} _ {\\theta} (x _ {t} | y _ {g}) = \\nabla_ {x _ {t}} \\log p _ {\\theta} (x _ {t}) \\\\ + w \\nabla_ {x _ {t}} \\left(\\log p _ {\\theta} \\left(x _ {t}\\right) - \\log p _ {\\theta} \\left(x _ {t} \\mid y _ {b}\\right)\\right). \\tag {11} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 106, + 861, + 482, + 897 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We can sample from $\\tilde{p}_{\\theta}(x_t|y_g)$ by substituting the score function in Eq. 2 with Eq. 11, resulting in:", + "bbox": [ + 511, + 323, + 903, + 354 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} d x = \\left[ - \\frac {\\beta (t)}{2} x - \\beta (t) \\left(\\nabla_ {x _ {t}} \\log p _ {t} \\left(x _ {t}\\right) \\right. \\right. \\tag {12} \\\\ \\left. + w \\nabla_ {x _ {t}} (\\log p _ {\\theta} (x _ {t}) - \\log p _ {\\theta} (x _ {t} | y _ {b})) \\right] d t + \\sqrt {\\beta (t)} d \\bar {w}, \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 517, + 364, + 903, + 435 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "and solving the reverse SDE. Since the score function is approximated using the neural network, we can generate samples using", + "bbox": [ + 511, + 444, + 905, + 491 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\epsilon} _ {\\theta} ^ {w} (x _ {t}) = \\epsilon_ {\\theta} (x _ {t}) + w \\left(\\epsilon_ {\\theta} (x _ {t}) - \\epsilon_ {\\theta} ^ {b} (x _ {t})\\right). \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 575, + 502, + 903, + 521 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "An interesting approach here is to model $\\nabla_{x_t}\\log p_\\theta (x_t|y_b)$ by perturbing the forward pass of $\\epsilon_{\\theta}(x_{t})$ , denoted as $\\epsilon_{\\theta}^{b}(x_{t})$ . Now the main focus is designing a perturbation that effectively yields the weak model capable of estimating $\\epsilon_{\\theta}^{b}(x_{t})$ , aligning closely with $\\epsilon_{\\theta}(x_{t})$ as discussed in Sec. 4.1. Ideally, we want a distribution that deviates minimally from $\\epsilon_{\\theta}(x_{t})$ while producing slightly lower-quality samples.", + "bbox": [ + 511, + 532, + 905, + 655 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.3. Spatiotemporal Skip Guidance (STG)", + "text_level": 1, + "bbox": [ + 511, + 665, + 839, + 681 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We introduce Spatiotemporal Skip Guidance (STG), a simple guidance method designed for video diffusion models that generates diverse, high-fidelity samples while addressing the limitations of existing methods. STG implicitly simulates an aligned weak model through spatiotemporal perturbation, capturing the spatiotemporal dynamics of video data.", + "bbox": [ + 511, + 688, + 905, + 792 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "An interesting discovery of this paper is that skipping layers within the network is an effective approach for constructing an aligned weak model. Modern neural network architectures for video diffusion models, such as ADM [25], DiT [26], and SiT [23], are partially or fully transformer-based and contain attention layers and residual blocks. These architectures are well-suited for layer skipping, as", + "bbox": [ + 511, + 795, + 905, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "11009", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/5070aabb09be4e8193c02ce195eebaa51b7ce9f83b565d3636ae12aadeb0475b.jpg", + "image_caption": [ + "Figure 3. Selected frames from videos generated by Mochi [33] with increasing STG scales." + ], + "image_footnote": [], + "bbox": [ + 93, + 101, + 906, + 210 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/3ec1b25622f72d9107003e686df759da57fe5afc1b25b233f755826228815b4e.jpg", + "image_caption": [ + "Figure 4. Comparison of CFG and STG across varying scales in terms of Imaging Quality and FVD." + ], + "image_footnote": [], + "bbox": [ + 114, + 255, + 457, + 412 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "they can still produce plausible outputs without significant degradation, even when a few layers are removed.", + "bbox": [ + 89, + 460, + 482, + 491 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Layer skipping is advantageous for generating an aligned weak model because it retains most of the neural network's weights and forward pass, resulting in similar predictions and distributions. This approach offers a clear advantage over methods that rely on external models, such as Autoguidance [19], or alternative objectives like CFG [10], where forward passes differ significantly.", + "bbox": [ + 89, + 491, + 482, + 595 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Moreover, our method applies perturbations to both spatial and temporal layers (or spatiotemporal layers), unlike existing image-based perturbation methods [13, 14] limited to 2D spatial attention layers. This dual-layer perturbation is essential for aligning the weak model with the main video diffusion model. We now discuss various STG configurations that can be applied across different video diffusion model architectures.", + "bbox": [ + 89, + 597, + 482, + 715 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Residual skip To skip an entire residual block, we modify $\\operatorname{Res}(z_l)$ to $\\operatorname{Res}'(z_l)$ ,", + "bbox": [ + 89, + 724, + 482, + 758 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {R e s} \\left(z _ {l}\\right) = z _ {l + 1} = f _ {l} \\left(z _ {l}\\right) + z _ {l}, \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 184, + 770, + 482, + 786 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {R e s} ^ {\\prime} \\left(z _ {l}\\right) = z _ {l + 1} = z _ {l}, \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 210, + 797, + 482, + 816 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $z_{l}$ and $z_{l + 1}$ represent the features at the $l^{\\mathrm{th}}$ and $(l + 1)^{\\mathrm{th}}$ layers, respectively, and $f_{l}$ denotes the $l^{\\mathrm{th}}$ neural net layer. The residual layers, which add small residuals $f_{l}(z_{l})$ to the original feature $z_{l}$ , ensure that the perturbed $z_{l + 1}$ does not deviate significantly from the original $z_{l + 1}$ .", + "bbox": [ + 89, + 823, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This reduces out-of-distribution issues in consecutive layers, generating perturbed but aligned samples.", + "bbox": [ + 511, + 255, + 903, + 286 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Attention skip Self-attention computes a linear combination of value tensors as", + "bbox": [ + 511, + 295, + 903, + 324 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {S A} (Q, K, V) = \\operatorname {S o f t m a x} \\left(\\frac {Q K ^ {T}}{\\sqrt {d}}\\right) V = \\mathbf {A} V, \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 539, + 333, + 903, + 369 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $Q \\in \\mathbb{R}^{(h \\times w \\times f) \\times d}$ , $K \\in \\mathbb{R}^{(h \\times w \\times f) \\times d}$ , $V \\in \\mathbb{R}^{(h \\times w \\times f) \\times d}$ are the query, key, and value matrices, respectively. Here, $h$ , $w$ , $f$ , and $d$ represent the height, width, frame number, and channel dimensions.", + "bbox": [ + 511, + 378, + 903, + 441 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We can skip this layer partially by passing the value matrix directly to the next layer without computing its linear combination. This is equivalent to replacing the attention matrix $\\mathbf{A}$ with an identity matrix $\\mathbf{I} \\in \\mathbb{R}^{hwf \\times hwf}$ , resulting in", + "bbox": [ + 511, + 441, + 905, + 516 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {S A} ^ {\\prime} (Q, K, V) = \\mathbf {I} V. \\tag {17}\n$$\n", + "text_format": "latex", + "bbox": [ + 637, + 517, + 903, + 534 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This represents a 3D extension of PAG [1].", + "bbox": [ + 511, + 542, + 799, + 558 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Factorized attention While recent models like Movie Gen [27] and Mochi [33] utilize full 3D spatiotemporal attention layers, many models, such as SVD [4] and OpenSora [38], still use factorized attention layers for efficiency. These models employ sequential 2D spatial attention and 1D temporal attention to approximate 3D spatiotemporal attention. For factorized models, we apply skip perturbation of Eq. 17 to spatial and temporal layers separately and use their linear combination for the final guidance.", + "bbox": [ + 511, + 566, + 905, + 703 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We denote the spatial and temporal perturbation labels as $y_{sb}$ and $y_{tb}$ , respectively. In practice, $y_{sb}$ and $y_{tb}$ are rarely independent, as they can influence each other. For instance, random spatial perturbations that create varied color adjustments across frames may disrupt temporal continuity. Similarly, random temporal perturbations can affect spatial consistency, leading to distortions in specific frames. However, for our skip perturbations, we can loosely assume independence, as skipping layers primarily reduces details in residual networks, which may not inherently affect temporal consistency.", + "bbox": [ + 511, + 704, + 905, + 869 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Therefore, to simplify the derivation, we initially assume independence between spatial and temporal perturbations.", + "bbox": [ + 511, + 869, + 903, + 900 + ], + "page_idx": 4 + }, + { + "type": "header", + "text": "A macro cinematography animation showing a butterfly emerging from its chrysalis, filmed with side-lit lighting ...", + "bbox": [ + 158, + 87, + 834, + 99 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "11010", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "(1) A romantic scene of a couple dancing under string lights in a backyard, with warm, golden tones highlighting their laughter.", + "(2) An animation showing a floating castle drifting above the clouds, with birds flying around it and sunlight casting golden rays ...", + "(3) A realistic documentary-style video of artisans crafting pottery, with the scene unfolding and transforming as hands shape clay ...", + "(4) A ghost in a white bedsheet faces a mirror. The ghost's reflection can be seen in the mirror. The ghost is in a dusty attic ..." + ], + "bbox": [ + 89, + 89, + 872, + 143 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/70c7a672903779f88253a7cf57412338b5e58688bb386b7bc4f3512985c07ec8.jpg", + "image_caption": [ + "Figure 5. Qualitative comparison between CFG and STG on videos generated by Mochi [33]." + ], + "image_footnote": [], + "bbox": [ + 104, + 146, + 900, + 883 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "11011", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We will revisit this assumption and propose an adjustment afterward. Under this independence assumption, the joint distribution can be expressed as", + "bbox": [ + 89, + 90, + 483, + 137 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\np _ {\\theta} \\left(x _ {t} \\mid y _ {b}\\right) = p _ {\\theta} \\left(x _ {t} \\mid y _ {\\mathrm {s b}}\\right) p _ {\\theta} \\left(x _ {t} \\mid y _ {\\mathrm {f b}}\\right). \\tag {18}\n$$\n", + "text_format": "latex", + "bbox": [ + 171, + 151, + 482, + 167 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Following the same approach as in Eq. 9 and using Eq. 11, we modify the score in Eq. 8 as follows:", + "bbox": [ + 89, + 180, + 482, + 210 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\nabla_ {x _ {t}} \\log \\tilde {p} _ {\\theta} (x _ {t} | y _ {g}) = \\nabla_ {x _ {t}} \\log p _ {\\theta} (x _ {t}) \\\\ + w \\nabla_ {x _ {t}} \\left(\\log p _ {\\theta} (x _ {t}) - \\log p _ {\\theta} (x _ {t} | y _ {s})\\right) \\\\ + w \\nabla_ {x _ {t}} \\left(\\log p _ {\\theta} \\left(x _ {t}\\right) - \\log p _ {\\theta} \\left(x _ {t} \\mid y _ {t}\\right)\\right). \\tag {19} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 119, + 223, + 482, + 280 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "By replacing the score with the estimated denoiser and utilizing separate scales for spatial and temporal terms, we obtain the following equation:", + "bbox": [ + 89, + 292, + 483, + 339 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\tilde {\\epsilon} _ {\\theta} ^ {w} (x _ {t}) = \\epsilon_ {\\theta} (x _ {t}) + w _ {1} \\left(\\epsilon_ {\\theta} (x _ {t}) - \\epsilon_ {\\theta} ^ {s} (x _ {t})\\right) \\\\ + w _ {2} \\left(\\epsilon_ {\\theta} \\left(x _ {t}\\right) - \\epsilon_ {\\theta} ^ {t} \\left(x _ {t}\\right)\\right) \\tag {20} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 150, + 351, + 480, + 387 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\epsilon_{\\theta}^{\\mathrm{s}}(x_t)$ and $\\epsilon_{\\theta}^{\\mathrm{t}}(x_t)$ represent spatially and temporally perturbed models, respectively.", + "bbox": [ + 89, + 400, + 482, + 431 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Next, we revisit our assumption of independence between spatial and temporal perturbations. While Eq. 20 works well in practice, we can derive an alternative STG formulation that uses orthogonalization to isolate the independent components of the spatial and temporal guidance. Inspired by a negative prompting technique [3], we can modify the Eq. 20 as follows:", + "bbox": [ + 89, + 431, + 483, + 537 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\tilde {\\epsilon} _ {\\theta} ^ {w} (x _ {t}) = \\epsilon_ {\\theta} (x _ {t}) + w _ {1} \\Delta_ {\\mathrm {s}} \\\\ + w _ {2} \\left(\\Delta_ {\\mathrm {t}} - \\frac {\\left\\langle \\Delta_ {\\mathrm {s}} , \\Delta_ {\\mathrm {t}} \\right\\rangle}{\\left\\| \\Delta_ {\\mathrm {s}} \\right\\| ^ {2}} \\Delta_ {\\mathrm {s}}\\right), \\tag {21} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 104, + 554, + 482, + 609 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\Delta_{\\mathrm{s}} = \\epsilon_{\\theta}(x_{t}) - \\epsilon_{\\theta}^{\\mathrm{s}}(x_{t})$ and $\\Delta_{\\mathrm{t}} = \\epsilon_{\\theta}(x_{t}) - \\epsilon_{\\theta}^{\\mathrm{t}}(x_{t})$", + "bbox": [ + 89, + 619, + 460, + 638 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Manifold Constrained Guidance Even with a well-aligned weak model, larger guidance scales $w$ can drive samples off the data manifold, resulting in poor quality and oversaturation. To address error accumulation at high guidance scales, we explore optional techniques to keep samples constrained to the manifold. Rescaling the latent code [21] helps mitigate this issue by constraining its variance, as larger variance is known to cause saturation in the results. Additionally, Restart sampling [36] demonstrates that introducing stochasticity can correct off-manifold deviations. Building on this idea, we incorporate stochastic forward processes into our sampling guidance framework as an optional method. While this approach modestly improves final sample quality and reduces saturation, it introduces additional computational overhead. Further details are provided in Appendix A5.1.", + "bbox": [ + 89, + 657, + 483, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 513, + 89, + 645, + 107 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.1. Overview", + "text_level": 1, + "bbox": [ + 513, + 114, + 624, + 128 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We employ three models for our experiments:", + "bbox": [ + 513, + 136, + 816, + 151 + ], + "page_idx": 6 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Mochi [33] is a text-to-video model built on AsymmDiT blocks, containing a total of 10 billion parameters and utilizing 3D self-attention in its spatiotemporal layers.", + "- Open-Sora [38] is a text-to-video model built on STDiT blocks with 1.1 billion parameters, employing factorized spatial and temporal attention layers.", + "- SVD [4] is an image-to-video model with 1.5 billion parameters that leverages factorized spatial and temporal attention within a UNet architecture." + ], + "bbox": [ + 513, + 152, + 903, + 287 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We evaluate the proposed method using widely adopted datasets and metrics to ensure a comprehensive performance analysis.", + "bbox": [ + 511, + 291, + 903, + 335 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "- UCF-101 The UCF-101 dataset comprises 13,320 videos organized into 101 action classes. Using this dataset, we assess the Fréchet Video Distance (FVD) and Inception Score (IS) of the image-to-video model SVD [4]. Following DIGAN [37], we calculate FVD and IS on 2,048 and 10,000 samples, respectively. For conditioning, initial frames from UCF-101 videos are used as inputs to the SVD model, with each input frame serving as the starting frame of the generated videos.", + "bbox": [ + 513, + 338, + 905, + 474 + ], + "page_idx": 6 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- VBench We use VBench [15] for automatic evaluation across various video metrics. SVD is evaluated using the image-to-video (I2V) framework on 355 samples from the VBench I2V dataset with 5 random seeds. Mochi and Open-Sora are evaluated with the text-to-video (T2V) framework: Open-Sora uses the standard VBench prompt list, while Mochi uses 100 randomly selected prompts due to computational limits.", + "- EvalCrafter We perform human evaluations using 700 prompts from the EvalCrafter dataset [22].", + "- LLM-Generated Prompts We use a set of 100 selected prompts generated by Claude 3.5 Sonnet [2] for our demo and qualitative comparisons." + ], + "bbox": [ + 513, + 474, + 905, + 671 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For evaluation, CFG is used together with STG for our evaluation. We did not use Restart sampling or guidance orthogonalization; additional results with these techniques are in the Appendix sections A5.1 and A5.2.", + "bbox": [ + 511, + 672, + 905, + 734 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Results", + "text_level": 1, + "bbox": [ + 513, + 742, + 607, + 757 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Qualitative Comparison Fig. 5 compares CFG and our method. Videos from the naive CFG model often show blurry objects with indistinct shapes, while STG produces clearer, more vivid frames with sharper image quality. STG also reduces temporal inconsistency and flickering, especially in dynamic videos with large motion where CFG frequently fails. Spatial guidance enhances object structure, and temporal guidance improves consistency. Please refer to the Appendix sections A5 and A6 for additional results.", + "bbox": [ + 511, + 763, + 905, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "11012", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/f7b0fe222a09f8bfbfbe8ae259fdb0fe99d5dabbd1449264c4793f988e6f6d5d.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelsFVD (↓)ISImaging QualityAesthetic QualityMotion SmoothnessDynamic DegreeTemporal Flickering
Mochi (STG-R)--0.6280.5540.9880.860.978
Mochi (STG-A)--0.5550.5410.9870.860.976
Open-Sora (STG-R)--0.5500.4740.9810.8940.977
Open-Sora (STG-A)--0.6060.5090.9870.8950.976
SVD (STG-R)155.939.30.6870.6370.9650.641-
SVD (STG-A)128.738.50.6940.6390.9680.694-
", + "bbox": [ + 93, + 88, + 903, + 200 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/9f7e52e1289188afa717d95934d6a7cff5163d5d61408c290e7bcddd1b86a3b5.jpg", + "table_caption": [ + "Table 3. Comparison of STG-R (residual skip) and STG-A (attention skip) across Mochi [33], Open-Sora [38], and SVD [4]. STG-R shows stronger performance on Mochi, while STG-A yields better results on Open-Sora and SVD." + ], + "table_footnote": [], + "table_body": "
ModelsFVD (↓)ISImaging QualityAesthetic QualityMotion SmoothnessDynamic Degree
CFG151.338.00.6870.6370.9660.562
+ Spatial133.838.30.6910.6390.9670.659
+ Temporal128.738.50.6940.6380.9680.694
", + "bbox": [ + 135, + 252, + 859, + 327 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4. Ablation study results on SVD [4] factorized attention, showing the impact of adding spatial and temporal guidance.", + "bbox": [ + 127, + 337, + 864, + 351 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Quantitative Comparison We compare CFG and STG using the FVD-Imaging Quality (VBench) plot across different scales in Fig. 4. FVD measures video distribution, while Imaging Quality assesses frame clarity. Higher CFG scales improve Imaging Quality but reduce diversity, as reflected in higher FVD. STG avoids this trade-off, maintaining diversity at increased scales.", + "bbox": [ + 88, + 369, + 480, + 474 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "T2V and I2V VBenchmark metrics in Tab.1 and Tab.2 show notable improvements in frame-level quality (Imaging and Aesthetic Quality). Temporal quality improves qualitatively, though metrics like Motion Smoothness and Temporal Flickering show marginal gains, as these scores are near saturation $(\\sim 0.9x)$", + "bbox": [ + 88, + 476, + 482, + 566 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "For Dynamic Degree, we aim to keep it unchanged, as it may not correlate with video quality. This is achieved in the T2V metric, but in the I2V model, CFG increases the influence of the conditioned image, reducing motion. STG, while not directly affecting Dynamic Degree, mitigates CFG's impact, leading to increased Dynamic Degree in I2V when used together.", + "bbox": [ + 88, + 566, + 482, + 672 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Human Evaluation We provide human evaluation results on 700 prompts from the EvalCrafter dataset [22] in the Appendix A4.", + "bbox": [ + 88, + 672, + 482, + 718 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.3. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 727, + 243, + 743 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Fig. 3 displays selected frames from videos generated by Mochi using various STG scales, where $w = 0$ corresponds to the CFG-only model without STG. Increasing the STG scale results in more vivid colors and finer details compared to the monotonous colors at $w = 0$ . Notably, unlike with CFG, sampling diversity is preserved as the STG scale increases, as shown in Fig. 4.", + "bbox": [ + 88, + 750, + 482, + 854 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We performed an ablation study on SVD to evaluate spatial and temporal guidance (Tab. 4). Spatial guidance notably reduced FVD and improved metrics, while tempo", + "bbox": [ + 88, + 854, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ral guidance further boosted performance, confirming their combined effectiveness for spatiotemporal generation.", + "bbox": [ + 511, + 369, + 903, + 401 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We also test two STG variants—residual skip (STGR) and attention skip (STG-A)—on SVD, Open-Sora, and Mochi. As shown in Tab. 3, STG-R performs well with Mochi, while STG-A is more effective for the other models. This is likely due to Mochi's higher parameter count and layer depth, enabling more extensive skipping without triggering out-of-distribution (OOD) issues in consecutive layers. Furthermore, Mochi's spatiotemporal layers consist of a single 3D attention layer, so STG-R skips a residual layer with just one attention layer. In contrast, SVD and Open-Sora use factorized attention, meaning STG-R skips a residual layer with two attention layers, which may cause excessive perturbation and lead to OOD issues in subsequent layers. More results are available in the Appendix A5.", + "bbox": [ + 511, + 402, + 905, + 614 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 632, + 633, + 647 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We proposed Spatiotemporal Skip Guidance (STG), a simple, training-free method for video diffusion models. By simulating an aligned weak model via spatiotemporal skipping, STG offers strong guidance for high-fidelity video generation, achieving notable qualitative and quantitative improvements. We hope this work advances video diffusion models and inspires further research in the field.", + "bbox": [ + 511, + 659, + 905, + 765 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitation and Ethical Considerations STG's performance depends on scale and layer selection, with the optimal configuration varying across models, requiring users to set these through heuristic tuning. While video quality improvements are notable, they also raise ethical concerns about misuse, underscoring the importance of using this technology responsibly and constructively.", + "bbox": [ + 511, + 794, + 905, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "11013", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements This work was supported by Institute for Information & communications Technology Planning & Evaluation(IITP) grant funded by the Korea government(MSIT) (RS-2019-II190075, Artificial Intelligence Graduate School Program(KAIST)). This work was supported by the National Research Foundation of Korea(NRF) grant funded by the Korea government(MSIT) (No. RS2025-00555621).", + "bbox": [ + 89, + 90, + 485, + 210 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 224, + 187, + 239 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Donghoon Ahn, Hyoungwon Cho, Jaewon Min, Wooseok Jang, Jungwoo Kim, SeonHwa Kim, Hyun Hee Park, Kyong Hwan Jin, and Seungryong Kim. Self-rectifying diffusion sampling with perturbed-attention guidance. arXiv preprint arXiv:2403.17377, 2024. 2, 3, 5, 7", + "[2] Anthropic. The claude 3 model family: Opus, sonnet, haiku. In-, 2024. 7", + "[3] Mohammadreza Armandpour, Ali Sadeghian, Huangjie Zheng, Amir Sadeghian, and Mingyuan Zhou. Re-imagine the negative prompt algorithm: Transform 2d diffusion into 3d, alleviate janus problem and beyond. arXiv preprint arXiv:2304.04968, 2023. 7", + "[4] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 2, 4, 5, 7, 8, 1, 9, 11, 13, 14, 17, 18, 23", + "[5] Andrew Brock. Large scale gan training for high fidelity natural image synthesis. arXiv preprint arXiv:1809.11096, 2018. 2", + "[6] Hyungjin Chung, Jeongsol Kim, Geon Yeong Park, Hyelin Nam, and Jong Chul Ye. Cfg++: Manifold-constrained classifier free guidance for diffusion models. arXiv preprint arXiv:2406.08070, 2024. 2, 3", + "[7] Zihan Ding, Xiao-Yang Liu, Miao Yin, and Linghe Kong. Tgan: Deep tensor generative adversarial nets for large image generation. arXiv preprint arXiv:1901.09953, 2019. 1", + "[8] Songwei Ge, Aniruddha Mahapatra, Gaurav Parmar, JunYan Zhu, and Jia-Bin Huang. On the content bias in fréchet video distance. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7277-7288, 2024. 1", + "[9] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139-144, 2020. 2", + "[10] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 2, 3, 5", + "[11] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2", + "[12] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. Advances in Neural Information Processing Systems, 35:8633-8646, 2022. 2" + ], + "bbox": [ + 93, + 250, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Susung Hong. Smoothed energy guidance: Guiding diffusion models with reduced energy curvature of attention. arXiv preprint arXiv:2408.00760, 2024. 2, 3, 5, 8, 18", + "[14] Susung Hong, Gyuseong Lee, Wooseok Jang, and Seunglyong Kim. Improving sample quality of diffusion models using self-attention guidance. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7462-7471, 2023. 3, 5", + "[15] Ziqi Huang, Yinan He, Jiashuo Yu, Fan Zhang, Chenyang Si, Yuming Jiang, Yuanhan Zhang, Tianxing Wu, Qingyang Jin, Nattapol Chanpaisit, et al. Vbench: Comprehensive benchmark suite for video generative models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21807-21818, 2024. 4, 7", + "[16] Yangqing Jia, Evan Shelhamer, Jeff Donahue, Sergey Karayev, Jonathan Long, Ross Girshick, Sergio Guadarrama, and Trevor Darrell. Caffe: Convolutional architecture for fast feature embedding. In Proceedings of the 22nd ACM international conference on Multimedia, pages 675-678, 2014.", + "[17] Andrej Karpathy, George Toderici, Sanketh Shetty, Thomas Leung, Rahul Sukthankar, and Li Fei-Fei. Large-scale video classification with convolutional neural networks. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 1725–1732, 2014. 1", + "[18] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8110-8119, 2020. 2", + "[19] Tero Karras, Miika Aittala, Tuomas Kynkänniemi, Jaakko Lehtinen, Timo Aila, and Samuli Laine. Guiding a diffusion model with a bad version of itself. arXiv preprint arXiv:2406.02507, 2024. 2, 3, 5", + "[20] Black Forest Labs. Flux: Github repository. https://github.com/black-forest-labs/flux, 2024. Accessed: 2024-11-22. 5, 6", + "[21] Shanchuan Lin, Bingchen Liu, Jiashi Li, and Xiao Yang. Common diffusion noise schedules and sample steps are flawed. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 5404-5411, 2024. 2, 7, 5, 10", + "[22] Yaofang Liu, Xiaodong Cun, Xuebo Liu, Xintao Wang, Yong Zhang, Haoxin Chen, Yang Liu, Tieyong Zeng, Raymond Chan, and Ying Shan. Evalcrafter: Benchmarking and evaluating large video generation models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22139-22149, 2024. 7, 8, 5, 6", + "[23] Nanye Ma, Mark Goldstein, Michael S Albergo, Nicholas M Boffi, Eric Vanden-Eijnden, and Saining Xie. Sit: Exploring flow and diffusion-based generative models with scalable interpolant transformers. arXiv preprint arXiv:2401.08740, 2024. 4", + "[24] Xinyin Ma, Gongfan Fang, Michael Bi Mi, and Xinchao Wang. Learning-to-cache: Accelerating diffusion transformer via layer caching. arXiv preprint arXiv:2406.01733, 2024. 4" + ], + "bbox": [ + 516, + 92, + 906, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "11014", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[25] Soumik Mukhopadhyay, Matthew Gwilliam, Vatsal Agarwal, Namitha Padmanabhan, Archana Swaminathan, Srinidhi Hegde, Tianyi Zhou, and Abhinav Shrivastava. Diffusion models beat gans on image classification. arXiv preprint arXiv:2307.08702, 2023. 3, 4", + "[26] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4195-4205, 2023. 4", + "[27] Adam Polyak, Amit Zohar, Andrew Brown, Andros Tjandra, Animesh Sinha, Ann Lee, Apoorv Vyas, Bowen Shi, Chih-Yao Ma, Ching-Yao Chuang, et al. Movie gen: A cast of media foundation models. arXiv preprint arXiv:2410.13720, 2024. 2, 5", + "[28] Hubert Ramsauer, Bernhard Schäfl, Johannes Lehner, Philipp Seidl, Michael Widrich, Thomas Adler, Lukas Gruber, Markus Holzleitner, Milena Pavlovic, Geir Kjetil Sandve, et al. Hopfield networks is all you need. arXiv preprint arXiv:2008.02217, 2020. 4", + "[29] Ali Razavi, Aaron Van den Oord, and Oriol Vinyals. Generating diverse high-fidelity images with vq-vae-2. Advances in neural information processing systems, 32, 2019. 2", + "[30] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2", + "[31] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International conference on machine learning, pages 2256-2265. PMLR, 2015.", + "[32] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 2, 3", + "[33] Genmo Team. Mochi 1. https://github.com/ genmoai/models, 2024. 2, 4, 5, 6, 7, 8, 1, 9, 10, 15, 19, 20", + "[34] Du Tran, Lubomir Bourdev, Rob Fergus, Lorenzo Torresani, and Manohar Paluri. Learning spatiotemporal features with 3d convolutional networks. In Proceedings of the IEEE international conference on computer vision, pages 4489-4497, 2015. 1", + "[35] Pascal Vincent. A connection between score matching and denoising autoencoders. Neural computation, 23(7):1661-1674, 2011. 3", + "[36] Yilun Xu, Mingyang Deng, Xiang Cheng, Yonglong Tian, Ziming Liu, and Tommi Jaakkola. Restart sampling for improving generative processes. Advances in Neural Information Processing Systems, 36:76806-76838, 2023. 2, 7, 5", + "[37] Sihyun Yu, Jihoon Tack, Sangwoo Mo, Hyunsu Kim, Junho Kim, Jung-Woo Ha, and Jinwoo Shin. Generating videos with dynamics-aware implicit generative adversarial networks. arXiv preprint arXiv:2202.10571, 2022. 7", + "[38] Zangwei Zheng, Xiangyu Peng, Tianji Yang, Chenhui Shen, Shenggui Li, Hongxin Liu, Yukun Zhou, Tianyi Li, and Yang You. Open-sora: Democratizing efficient video production" + ], + "bbox": [ + 89, + 90, + 485, + 901 + ], + "page_idx": 9 + }, + { + "type": "text", + "text": "for all. https://github.com/hpcaitech/Open-Sora, 2024. 2, 4, 5, 7, 8, 1, 12, 16, 21, 22", + "bbox": [ + 545, + 90, + 906, + 119 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "11015", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Spatiotemporal Skip Guidance for Enhanced Video Diffusion Sampling/4ec7da94-3031-45c8-a8d6-19bf226e3cba_model.json b/2025/Spatiotemporal Skip Guidance for Enhanced Video Diffusion Sampling/4ec7da94-3031-45c8-a8d6-19bf226e3cba_model.json new file mode 100644 index 0000000000000000000000000000000000000000..9987dedf6b1afd4d14194e16e5fe8e9ab054b4be --- /dev/null +++ b/2025/Spatiotemporal Skip Guidance for Enhanced Video Diffusion Sampling/4ec7da94-3031-45c8-a8d6-19bf226e3cba_model.json @@ -0,0 +1,2332 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.139, + 0.131, + 0.86, + 0.154 + ], + "angle": 0, + "content": "Spatiotemporal Skip Guidance for Enhanced Video Diffusion Sampling" + }, + { + "type": "text", + "bbox": [ + 0.171, + 0.18, + 0.825, + 0.201 + ], + "angle": 0, + "content": "Junha Hyung\\(^{*1}\\) Kinam Kim\\(^{*1}\\) Susung Hong\\(^{2}\\) Min-Jung Kim\\(^{1}\\) Jaegul Choo\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.336, + 0.215, + 0.662, + 0.235 + ], + "angle": 0, + "content": "\\(^{1}\\)KAIST AI \\(^{2}\\)University of Washington" + }, + { + "type": "text", + "bbox": [ + 0.157, + 0.237, + 0.838, + 0.253 + ], + "angle": 0, + "content": "{sharpeeee, kinamplify, emjay73, jchoo}@kaist.ac.kr, susung@cs.washington.edu" + }, + { + "type": "text", + "bbox": [ + 0.114, + 0.665, + 0.481, + 0.695 + ], + "angle": 0, + "content": "\"A close-up shot of a butterfly landing on the nose of a woman, highlighting her smile and the details of the butterfly's wings.\"" + }, + { + "type": "text", + "bbox": [ + 0.525, + 0.665, + 0.882, + 0.694 + ], + "angle": 0, + "content": "\"A close-up of a woman's face with colored powder exploding around her, creating an abstract splash of vibrant hues.\"" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.707, + 0.908, + 0.734 + ], + "angle": 0, + "content": "Figure 1. Visual comparison of video quality between CFG (top row) and our STG method (bottom row). Best viewed in Acrobat Reader; click on the images to watch the videos." + }, + { + "type": "title", + "bbox": [ + 0.281, + 0.755, + 0.362, + 0.772 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.788, + 0.486, + 0.879 + ], + "angle": 0, + "content": "Diffusion models have emerged as a powerful tool for generating high-quality images, videos, and 3D content. While sampling guidance techniques like CFG improve quality, they reduce diversity and motion. Autoguidance mitigates these issues but demands extra weak model training, limiting its practicality for large-scale models. In this" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.757, + 0.909, + 0.894 + ], + "angle": 0, + "content": "work, we introduce Spatiotemporal Skip Guidance (STG), a simple training-free sampling guidance method for enhancing transformer-based video diffusion models. STG employs an implicit weak model via self-perturbation, avoiding the need for external models or additional training. By selectively skipping spatiotemporal layers, STG produces an aligned, degraded version of the original model to boost sample quality without compromising diversity or dynamic degree. Our contributions include: (1) introducing STG as" + }, + { + "type": "page_footnote", + "bbox": [ + 0.115, + 0.889, + 0.279, + 0.902 + ], + "angle": 0, + "content": "* indicates equal contribution." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.521, + 0.957 + ], + "angle": 0, + "content": "11006" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.198 + ], + "angle": 0, + "content": "an efficient, high-performing guidance technique for video diffusion models, (2) eliminating the need for auxiliary models by simulating a weak model through layer skipping, and (3) ensuring quality-enhanced guidance without compromising sample diversity or dynamics unlike CFG. For additional results, visit https://junhahyung.github.io/STGuidance/." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.226, + 0.223, + 0.243 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.251, + 0.483, + 0.403 + ], + "angle": 0, + "content": "Diffusion models [11, 30-32] are a successful class of generative models known for their flexibility in modeling complex data distributions, achieving impressive results in image, video, and 3D generation. By progressively denoising random noise, they enable robust generalization, making them a leading choice for realistic content generation and often surpassing GAN-based methods [5, 9, 18, 29]. Building on this success, video diffusion models [4, 12, 27, 33, 38] generate high-quality videos by using temporal or 3D attention layers to handle sequential frames." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.403, + 0.483, + 0.538 + ], + "angle": 0, + "content": "Meanwhile, to enhance sample quality, sampling guidance techniques such as Classifier-Free Guidance (CFG) [10] and Autoguidance [19] have been introduced to guide the denoising process. These techniques employ weak models to predict poor trajectories, steering the main model away from them and pushing samples toward high-quality regions on the data manifold. However, CFG often reduces diversity, leading to saturated or overly simplified results [6, 19]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.539, + 0.483, + 0.689 + ], + "angle": 0, + "content": "Autoguidance [19] addresses this issue by using a weak model trained on the same task, conditioning, and data distribution as the main model. The main drawback of this approach, however, is the need to train an additional weak model, which is impractical for large-scale models. Alternative methods such as PAG [1] and SEG [13] use self-perturbation to implicitly mimic a weak model, avoiding the need for extra training. Yet, these methods are designed specifically for image generation diffusion models, applying self-perturbation to 2D spatial attention maps." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.483, + 0.825 + ], + "angle": 0, + "content": "In this work, we propose Spatiotemporal Skip Guidance (STG), a simple and effective sampling guidance method for video diffusion models that significantly enhances the performance of any transformer-based video diffusion model without additional training. Specifically, we use an implicit weak model for guidance through self-perturbation, eliminating the need for explicit weak models and their associated training costs. This is especially crucial for video diffusion models, where training costs are high." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Our implicit weak model, a key component of our framework, is deliberately designed as a degraded but \"aligned\" version of the original video generation model. As demonstrated in Autoguidance [19], having the weak model share the same task, conditioning, and data distribution as the" + }, + { + "type": "image", + "bbox": [ + 0.526, + 0.09, + 0.727, + 0.223 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.619, + 0.227, + 0.651, + 0.238 + ], + "angle": 0, + "content": "CFG" + }, + { + "type": "image", + "bbox": [ + 0.757, + 0.09, + 0.901, + 0.223 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.818, + 0.227, + 0.851, + 0.238 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.25, + 0.906, + 0.335 + ], + "angle": 0, + "content": "Figure 2. Comparison between CFG and STG, with the band conceptually representing the noisy data manifold. In STG, the weak model and the main model are aligned along the direction of increasing quality. In contrast, the two models in CFG differ not only in quality but also in aspects such as diversity and prompt alignment capabilities." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.364, + 0.906, + 0.53 + ], + "angle": 0, + "content": "main model is essential for quality improvement, as we expect both models to exhibit similar, aligned errors in the same parts of samples. We conceptually illustrate this alignment in Fig. 2, where our weak model and main model are arranged in a direction of increasing quality. To achieve this, we apply spatiotemporal perturbation to both spatial and temporal attention layers—or, in the case of 3D attention, to the entire layer—by selectively skipping certain layers. This straightforward approach effectively nullifies specific residual or attention layers, generating a lower-quality version of the model that simulates an aligned weak model." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.542, + 0.909, + 0.709 + ], + "angle": 0, + "content": "To ensure samples remain on the data manifold, even when using large guidance scales, we employ optional techniques such as the rescaling [21] and restart [36] methods. Rescaling the latent code constrains its variance, addressing the issue of larger variance causing saturation in the results [21]. Meanwhile, the restart sampling technique leverages the error contraction property of forward SDEs [36] to keep the sampling trajectory on the manifold. These techniques help prevent overshooting in the sampling trajectory, which could otherwise push samples off the manifold and result in saturated or distorted outputs." + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.72, + 0.779, + 0.735 + ], + "angle": 0, + "content": "Our key contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.747, + 0.905, + 0.791 + ], + "angle": 0, + "content": "- We propose STG—a surprisingly simple sampling guidance framework that significantly boosts the performance of video diffusion models." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.793, + 0.905, + 0.851 + ], + "angle": 0, + "content": "- Our method introduces an implicit weak model by skipping spatiotemporal layers in video diffusion models, eliminating the need for additional training or external models." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.853, + 0.905, + 0.898 + ], + "angle": 0, + "content": "- Our method enhances sample quality during guidance without reducing diversity or limiting the dynamics of generated videos." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.747, + 0.905, + 0.898 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11007" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.09, + 0.234, + 0.106 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.116, + 0.484, + 0.237 + ], + "angle": 0, + "content": "Guidance with trained weak model Classifier-Free Guidance (CFG) [10] improves conditional generation in Diffusion Models by using an implicit unconditional model as a weak model. However, differences in tasks between the unconditional and conditional models can reduce sample diversity [6, 19] and increase sampling trajectory curvature [6], leading to overshooting the data manifold and producing skewed or oversaturated images." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.238, + 0.484, + 0.343 + ], + "angle": 0, + "content": "Autoguidance [19] mitigates these issues by employing a bad version of the main model as a weak model, trained with reduced capacity and compute to ensure alignment. This alignment allows the guidance algorithm to correct errors by analyzing prediction differences. While effective, it requires additional training, which is challenging for large-scale video diffusion models." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.366, + 0.484, + 0.593 + ], + "angle": 0, + "content": "Guidance with training-free weak model Another line of work avoids additional training by using self-perturbation of the main model to mimic a weak model. Self-Attention Guidance (SAG) [14] blurs high-attention regions, Perturbed Attention Guidance (PAG) [1] replaces attention maps with identity matrices, and Smoothed Energy Guidance (SEG) [13] applies Gaussian blur to the attention weights to smooth the energy landscape. These methods guide sampling toward high-quality outputs by leveraging differences in predictions from the weakened model. While effective, they are primarily designed for image diffusion models with 2D self-attention. We aim to extend this approach to video diffusion models, which require handling temporal dynamics with additional temporal or 3D spatiotemporal attention layers." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.608, + 0.231, + 0.623 + ], + "angle": 0, + "content": "3. Preliminaries" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.633, + 0.257, + 0.648 + ], + "angle": 0, + "content": "3.1. Diffusion models" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.657, + 0.484, + 0.718 + ], + "angle": 0, + "content": "Diffusion models generate samples by progressively removing noise from noisy data, restoring the original data distribution. Song et al. [32] defined the process of adding noise to the data using a stochastic differential equation (SDE):" + }, + { + "type": "equation", + "bbox": [ + 0.182, + 0.729, + 0.482, + 0.76 + ], + "angle": 0, + "content": "\\[\nd x = - \\frac {\\beta (t)}{2} x d t + \\sqrt {\\beta (t)} d w, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.771, + 0.484, + 0.816 + ], + "angle": 0, + "content": "where \\(\\beta (t)\\) is a time-dependent noise schedule, and \\(w\\) represents the standard Wiener process. Corresponding reverse-time SDE is:" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.826, + 0.484, + 0.86 + ], + "angle": 0, + "content": "\\[\nd x = \\left[ - \\frac {\\beta (t)}{2} x - \\beta (t) \\nabla_ {x} \\log p _ {t} (x) \\right] d t + \\sqrt {\\beta (t)} d \\bar {w} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.484, + 0.903 + ], + "angle": 0, + "content": "where \\(d\\bar{w}\\) is the Wiener process running backward in time. Here, the score function \\(\\nabla_x\\log p_t(x)\\) is approximated by" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.907, + 0.122 + ], + "angle": 0, + "content": "a neural network \\( s_{\\theta}(x(t)) \\) trained using denoising score matching [35]:" + }, + { + "type": "equation", + "bbox": [ + 0.529, + 0.135, + 0.907, + 0.18 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\theta^ {*} = \\arg \\min _ {\\theta} \\mathbb {E} _ {t} \\left\\{\\lambda (t) \\mathbb {E} _ {x _ {0}} \\mathbb {E} _ {x _ {t} | x _ {0}} \\right[ \\\\ \\left. \\left\\| s _ {\\theta} \\left(x _ {t}\\right) - \\nabla_ {x _ {t}} \\log p _ {t} \\left(x _ {t} \\mid x _ {0}\\right) \\| _ {2} ^ {2} \\right] \\right\\}. \\tag {3} \\\\ \\end{array}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.187, + 0.702, + 0.202 + ], + "angle": 0, + "content": "3.2. Classifier Guidance" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.209, + 0.906, + 0.256 + ], + "angle": 0, + "content": "Classifier Guidance (CG) [25] reformulates the reverse process of a diffusion model by incorporating an external classifier \\( p_{\\phi} \\) as" + }, + { + "type": "equation", + "bbox": [ + 0.61, + 0.265, + 0.907, + 0.283 + ], + "angle": 0, + "content": "\\[\n\\tilde {p _ {\\theta}} \\left(x _ {t} | y\\right) \\propto p _ {\\theta} \\left(x _ {t}\\right) p _ {\\phi} \\left(y \\mid x _ {t}\\right) ^ {\\lambda}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.292, + 0.906, + 0.322 + ], + "angle": 0, + "content": "where \\(y\\) is the desired class label and \\(\\lambda\\) controls the guidance strength. The score function is then derived as:" + }, + { + "type": "equation", + "bbox": [ + 0.529, + 0.336, + 0.906, + 0.372 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\nabla_ {x _ {t}} \\log \\tilde {p} _ {\\theta} (x _ {t} | y) = \\nabla_ {x _ {t}} \\log p _ {\\theta} (x _ {t}) \\\\ + \\lambda \\nabla_ {x _ {t}} \\log p _ {\\phi} (y | x _ {t}). \\tag {5} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.381, + 0.906, + 0.412 + ], + "angle": 0, + "content": "By substituting the score function in Eq. 2 with Eq. 5, sampling from the desired class condition \\(y\\) becomes possible." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.42, + 0.742, + 0.435 + ], + "angle": 0, + "content": "3.3. Classifier-Free Guidance" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.442, + 0.907, + 0.487 + ], + "angle": 0, + "content": "Classifier-Free Guidance (CFG) [10] uses Bayes' rule to replace a classifier-guided score with a linear combination of conditional and unconditional score estimates:" + }, + { + "type": "equation", + "bbox": [ + 0.568, + 0.497, + 0.906, + 0.515 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\epsilon} _ {\\theta} ^ {\\lambda} \\left(x _ {t}\\right) = \\epsilon_ {\\theta} \\left(x _ {t}\\right) + \\lambda \\left(\\epsilon_ {\\theta} \\left(x _ {t}\\right) - \\epsilon_ {\\theta} \\left(x _ {t} \\mid \\phi\\right)\\right). \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.524, + 0.906, + 0.6 + ], + "angle": 0, + "content": "CFG jointly trains the unconditional model \\(\\epsilon_{\\theta}(x_t|\\phi)\\) and the conditional model \\(\\epsilon_{\\theta}(x_t|c) (= \\epsilon_{\\theta}(x_t))\\) within a single model by setting the condition \\(c\\) to a null token \\(\\phi\\). Using this guided denoising process, the model better captures the conditions and often generates high-fidelity outputs." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.613, + 0.605, + 0.628 + ], + "angle": 0, + "content": "4. Method" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.637, + 0.773, + 0.654 + ], + "angle": 0, + "content": "4.1. Optimal Weak Model Design" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.659, + 0.906, + 0.78 + ], + "angle": 0, + "content": "Our goal is to construct an aligned weak model that captures a distribution similar to the original model while generating slightly lower-quality samples. This ensures that the guidance gradient points toward improved quality, as shown in Fig. 2. Misaligned models, as seen in CFG [10], may lead to unintended outcomes like reduced diversity, while Autoguidance [19] requires additional weak model training, making it impractical for large-scale models." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.906, + 0.871 + ], + "angle": 0, + "content": "To address these limitations, we design an implicit weak model by leveraging the original model itself, eliminating the need for external training. This training-free approach ensures computational efficiency and better alignment, as the weak model is derived directly from the original, preserving most network weights." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.906, + 0.901 + ], + "angle": 0, + "content": "To achieve this, we apply perturbation methods directly to the main model's forward pass, creating an implicit weak" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11008" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.113, + 0.089, + 0.885, + 0.18 + ], + "angle": 0, + "content": "
ModelsImaging QualityAesthetic QualityMotion SmoothnessDynamic DegreeTemporal Flickering
Mochi (CFG)0.5240.5070.9850.870.976
Mochi (STG)0.6280.5540.9880.860.978
Open-Sora (CFG)0.5610.4930.9820.9020.975
Open-Sora (STG)0.6060.5090.9870.8950.976
" + }, + { + "type": "table_caption", + "bbox": [ + 0.203, + 0.191, + 0.794, + 0.207 + ], + "angle": 0, + "content": "Table 1. Quantitative results for Mochi [33] and Open-Sora [38] on VBench [15] T2V benchmarks." + }, + { + "type": "table", + "bbox": [ + 0.174, + 0.219, + 0.825, + 0.273 + ], + "angle": 0, + "content": "
ModelsFVD (↓)ISImaging QualityAesthetic QualityMotion SmoothnessDynamic Degree
SVD (CFG)151.338.00.6870.6370.9660.562
SVD (STG)128.738.50.6940.6390.9680.694
" + }, + { + "type": "table_caption", + "bbox": [ + 0.233, + 0.282, + 0.763, + 0.297 + ], + "angle": 0, + "content": "Table 2. Quantitative results for SVD [4] on FVD, IS, and VBenchmark [15] I2V benchmarks." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.324, + 0.484, + 0.371 + ], + "angle": 0, + "content": "model. Specifically, we use spatiotemporal perturbation, a key component for aligning the weak model with video diffusion models, which will be elaborated upon in detail." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.381, + 0.416, + 0.398 + ], + "angle": 0, + "content": "4.2. Sampling from High Quality Samples" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.404, + 0.484, + 0.451 + ], + "angle": 0, + "content": "Similar to CG [25], we define our goal as conditioning the model on an imaginary label \\( y_{g} \\) that represents high-quality samples, leading to the sampling distribution" + }, + { + "type": "equation", + "bbox": [ + 0.18, + 0.462, + 0.483, + 0.481 + ], + "angle": 0, + "content": "\\[\n\\tilde {p} _ {\\theta} \\left(x _ {t} \\mid y _ {g}\\right) \\propto p _ {\\theta} \\left(x _ {t}\\right) p _ {\\phi} \\left(y _ {g} \\mid x _ {t}\\right) ^ {w}. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.493, + 0.484, + 0.537 + ], + "angle": 0, + "content": "Using \\( w > 0 \\) sharpens the distribution, promoting the generation of high-quality samples. From Eq. 7, the score is derived as" + }, + { + "type": "equation", + "bbox": [ + 0.107, + 0.555, + 0.482, + 0.592 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\nabla_ {x _ {t}} \\log \\tilde {p} _ {\\theta} (x _ {t} | y _ {g}) \\\\ = \\nabla_ {x _ {t}} \\log p _ {\\theta} (x _ {t}) + w \\nabla_ {x _ {t}} \\log p _ {\\phi} (y _ {g} | x _ {t}). \\tag {8} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.604, + 0.484, + 0.665 + ], + "angle": 0, + "content": "Rather than using an external classifier \\( p_{\\phi} \\), we propose using our model \\( p_{\\theta} \\) as an implicit classifier. We design this classifier to be inversely proportional to the probability of an imaginary \"bad\" label \\( y_{b} \\), expressed as" + }, + { + "type": "equation", + "bbox": [ + 0.205, + 0.677, + 0.483, + 0.71 + ], + "angle": 0, + "content": "\\[\np _ {\\phi} \\left(y _ {g} \\mid x _ {t}\\right) \\propto \\frac {1}{p _ {\\theta} \\left(y _ {b} \\mid x _ {t}\\right)}. \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.722, + 0.484, + 0.752 + ], + "angle": 0, + "content": "Using Bayes' rule, the function can be expressed in terms of the marginal posterior as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.131, + 0.765, + 0.483, + 0.817 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\nabla_ {x _ {t}} \\log \\frac {1}{p _ {\\theta} (y _ {b} | x _ {t})} = \\nabla_ {x _ {t}} \\log \\frac {p _ {\\theta} (x _ {t})}{p _ {\\theta} (y _ {b}) p _ {\\theta} (x _ {t} | y _ {b})} \\\\ = \\nabla_ {x _ {t}} \\left(\\log p _ {\\theta} \\left(x _ {t}\\right) - \\log p _ {\\theta} \\left(x _ {t} \\mid y _ {b}\\right)\\right), \\tag {10} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.83, + 0.408, + 0.845 + ], + "angle": 0, + "content": "leading to the score of the target distribution as:" + }, + { + "type": "equation", + "bbox": [ + 0.107, + 0.862, + 0.483, + 0.898 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\nabla_ {x _ {t}} \\log \\tilde {p} _ {\\theta} (x _ {t} | y _ {g}) = \\nabla_ {x _ {t}} \\log p _ {\\theta} (x _ {t}) \\\\ + w \\nabla_ {x _ {t}} \\left(\\log p _ {\\theta} \\left(x _ {t}\\right) - \\log p _ {\\theta} \\left(x _ {t} \\mid y _ {b}\\right)\\right). \\tag {11} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.324, + 0.905, + 0.355 + ], + "angle": 0, + "content": "We can sample from \\(\\tilde{p}_{\\theta}(x_t|y_g)\\) by substituting the score function in Eq. 2 with Eq. 11, resulting in:" + }, + { + "type": "equation", + "bbox": [ + 0.518, + 0.365, + 0.905, + 0.436 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} d x = \\left[ - \\frac {\\beta (t)}{2} x - \\beta (t) \\left(\\nabla_ {x _ {t}} \\log p _ {t} \\left(x _ {t}\\right) \\right. \\right. \\tag {12} \\\\ \\left. + w \\nabla_ {x _ {t}} (\\log p _ {\\theta} (x _ {t}) - \\log p _ {\\theta} (x _ {t} | y _ {b})) \\right] d t + \\sqrt {\\beta (t)} d \\bar {w}, \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.445, + 0.906, + 0.492 + ], + "angle": 0, + "content": "and solving the reverse SDE. Since the score function is approximated using the neural network, we can generate samples using" + }, + { + "type": "equation", + "bbox": [ + 0.576, + 0.503, + 0.905, + 0.522 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\epsilon} _ {\\theta} ^ {w} (x _ {t}) = \\epsilon_ {\\theta} (x _ {t}) + w \\left(\\epsilon_ {\\theta} (x _ {t}) - \\epsilon_ {\\theta} ^ {b} (x _ {t})\\right). \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.534, + 0.906, + 0.656 + ], + "angle": 0, + "content": "An interesting approach here is to model \\(\\nabla_{x_t}\\log p_\\theta (x_t|y_b)\\) by perturbing the forward pass of \\(\\epsilon_{\\theta}(x_{t})\\), denoted as \\(\\epsilon_{\\theta}^{b}(x_{t})\\). Now the main focus is designing a perturbation that effectively yields the weak model capable of estimating \\(\\epsilon_{\\theta}^{b}(x_{t})\\), aligning closely with \\(\\epsilon_{\\theta}(x_{t})\\) as discussed in Sec. 4.1. Ideally, we want a distribution that deviates minimally from \\(\\epsilon_{\\theta}(x_{t})\\) while producing slightly lower-quality samples." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.666, + 0.84, + 0.683 + ], + "angle": 0, + "content": "4.3. Spatiotemporal Skip Guidance (STG)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.906, + 0.793 + ], + "angle": 0, + "content": "We introduce Spatiotemporal Skip Guidance (STG), a simple guidance method designed for video diffusion models that generates diverse, high-fidelity samples while addressing the limitations of existing methods. STG implicitly simulates an aligned weak model through spatiotemporal perturbation, capturing the spatiotemporal dynamics of video data." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.906, + 0.902 + ], + "angle": 0, + "content": "An interesting discovery of this paper is that skipping layers within the network is an effective approach for constructing an aligned weak model. Modern neural network architectures for video diffusion models, such as ADM [25], DiT [26], and SiT [23], are partially or fully transformer-based and contain attention layers and residual blocks. These architectures are well-suited for layer skipping, as" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "11009" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.16, + 0.088, + 0.835, + 0.101 + ], + "angle": 0, + "content": "A macro cinematography animation showing a butterfly emerging from its chrysalis, filmed with side-lit lighting ..." + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.102, + 0.907, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.224, + 0.229, + 0.773, + 0.243 + ], + "angle": 0, + "content": "Figure 3. Selected frames from videos generated by Mochi [33] with increasing STG scales." + }, + { + "type": "image", + "bbox": [ + 0.116, + 0.256, + 0.459, + 0.414 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.428, + 0.483, + 0.456 + ], + "angle": 0, + "content": "Figure 4. Comparison of CFG and STG across varying scales in terms of Imaging Quality and FVD." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.461, + 0.483, + 0.492 + ], + "angle": 0, + "content": "they can still produce plausible outputs without significant degradation, even when a few layers are removed." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.492, + 0.483, + 0.597 + ], + "angle": 0, + "content": "Layer skipping is advantageous for generating an aligned weak model because it retains most of the neural network's weights and forward pass, resulting in similar predictions and distributions. This approach offers a clear advantage over methods that rely on external models, such as Autoguidance [19], or alternative objectives like CFG [10], where forward passes differ significantly." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.598, + 0.483, + 0.716 + ], + "angle": 0, + "content": "Moreover, our method applies perturbations to both spatial and temporal layers (or spatiotemporal layers), unlike existing image-based perturbation methods [13, 14] limited to 2D spatial attention layers. This dual-layer perturbation is essential for aligning the weak model with the main video diffusion model. We now discuss various STG configurations that can be applied across different video diffusion model architectures." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.726, + 0.483, + 0.759 + ], + "angle": 0, + "content": "Residual skip To skip an entire residual block, we modify \\(\\operatorname{Res}(z_l)\\) to \\(\\operatorname{Res}'(z_l)\\)," + }, + { + "type": "equation", + "bbox": [ + 0.185, + 0.771, + 0.483, + 0.787 + ], + "angle": 0, + "content": "\\[\n\\operatorname {R e s} \\left(z _ {l}\\right) = z _ {l + 1} = f _ {l} \\left(z _ {l}\\right) + z _ {l}, \\tag {14}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.212, + 0.799, + 0.483, + 0.817 + ], + "angle": 0, + "content": "\\[\n\\operatorname {R e s} ^ {\\prime} \\left(z _ {l}\\right) = z _ {l + 1} = z _ {l}, \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.824, + 0.484, + 0.902 + ], + "angle": 0, + "content": "where \\( z_{l} \\) and \\( z_{l + 1} \\) represent the features at the \\( l^{\\mathrm{th}} \\) and \\( (l + 1)^{\\mathrm{th}} \\) layers, respectively, and \\( f_{l} \\) denotes the \\( l^{\\mathrm{th}} \\) neural net layer. The residual layers, which add small residuals \\( f_{l}(z_{l}) \\) to the original feature \\( z_{l} \\), ensure that the perturbed \\( z_{l + 1} \\) does not deviate significantly from the original \\( z_{l + 1} \\)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.256, + 0.905, + 0.287 + ], + "angle": 0, + "content": "This reduces out-of-distribution issues in consecutive layers, generating perturbed but aligned samples." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.296, + 0.905, + 0.325 + ], + "angle": 0, + "content": "Attention skip Self-attention computes a linear combination of value tensors as" + }, + { + "type": "equation", + "bbox": [ + 0.54, + 0.334, + 0.905, + 0.37 + ], + "angle": 0, + "content": "\\[\n\\operatorname {S A} (Q, K, V) = \\operatorname {S o f t m a x} \\left(\\frac {Q K ^ {T}}{\\sqrt {d}}\\right) V = \\mathbf {A} V, \\tag {16}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.38, + 0.905, + 0.442 + ], + "angle": 0, + "content": "where \\(Q \\in \\mathbb{R}^{(h \\times w \\times f) \\times d}\\), \\(K \\in \\mathbb{R}^{(h \\times w \\times f) \\times d}\\), \\(V \\in \\mathbb{R}^{(h \\times w \\times f) \\times d}\\) are the query, key, and value matrices, respectively. Here, \\(h\\), \\(w\\), \\(f\\), and \\(d\\) represent the height, width, frame number, and channel dimensions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.443, + 0.906, + 0.517 + ], + "angle": 0, + "content": "We can skip this layer partially by passing the value matrix directly to the next layer without computing its linear combination. This is equivalent to replacing the attention matrix \\(\\mathbf{A}\\) with an identity matrix \\(\\mathbf{I} \\in \\mathbb{R}^{hwf \\times hwf}\\), resulting in" + }, + { + "type": "equation", + "bbox": [ + 0.638, + 0.518, + 0.905, + 0.535 + ], + "angle": 0, + "content": "\\[\n\\mathrm {S A} ^ {\\prime} (Q, K, V) = \\mathbf {I} V. \\tag {17}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.544, + 0.8, + 0.559 + ], + "angle": 0, + "content": "This represents a 3D extension of PAG [1]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.568, + 0.906, + 0.704 + ], + "angle": 0, + "content": "Factorized attention While recent models like Movie Gen [27] and Mochi [33] utilize full 3D spatiotemporal attention layers, many models, such as SVD [4] and OpenSora [38], still use factorized attention layers for efficiency. These models employ sequential 2D spatial attention and 1D temporal attention to approximate 3D spatiotemporal attention. For factorized models, we apply skip perturbation of Eq. 17 to spatial and temporal layers separately and use their linear combination for the final guidance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.906, + 0.87 + ], + "angle": 0, + "content": "We denote the spatial and temporal perturbation labels as \\( y_{sb} \\) and \\( y_{tb} \\), respectively. In practice, \\( y_{sb} \\) and \\( y_{tb} \\) are rarely independent, as they can influence each other. For instance, random spatial perturbations that create varied color adjustments across frames may disrupt temporal continuity. Similarly, random temporal perturbations can affect spatial consistency, leading to distortions in specific frames. However, for our skip perturbations, we can loosely assume independence, as skipping layers primarily reduces details in residual networks, which may not inherently affect temporal consistency." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Therefore, to simplify the derivation, we initially assume independence between spatial and temporal perturbations." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11010" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.09, + 0.845, + 0.104 + ], + "angle": 0, + "content": "(1) A romantic scene of a couple dancing under string lights in a backyard, with warm, golden tones highlighting their laughter." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.104, + 0.864, + 0.117 + ], + "angle": 0, + "content": "(2) An animation showing a floating castle drifting above the clouds, with birds flying around it and sunlight casting golden rays ..." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.117, + 0.874, + 0.131 + ], + "angle": 0, + "content": "(3) A realistic documentary-style video of artisans crafting pottery, with the scene unfolding and transforming as hands shape clay ..." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.131, + 0.829, + 0.145 + ], + "angle": 0, + "content": "(4) A ghost in a white bedsheet faces a mirror. The ghost's reflection can be seen in the mirror. The ghost is in a dusty attic ..." + }, + { + "type": "list", + "bbox": [ + 0.09, + 0.09, + 0.874, + 0.145 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.105, + 0.147, + 0.901, + 0.884 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.221, + 0.895, + 0.776, + 0.91 + ], + "angle": 0, + "content": "Figure 5. Qualitative comparison between CFG and STG on videos generated by Mochi [33]." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "11011" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.138 + ], + "angle": 0, + "content": "We will revisit this assumption and propose an adjustment afterward. Under this independence assumption, the joint distribution can be expressed as" + }, + { + "type": "equation", + "bbox": [ + 0.173, + 0.152, + 0.483, + 0.168 + ], + "angle": 0, + "content": "\\[\np _ {\\theta} \\left(x _ {t} \\mid y _ {b}\\right) = p _ {\\theta} \\left(x _ {t} \\mid y _ {\\mathrm {s b}}\\right) p _ {\\theta} \\left(x _ {t} \\mid y _ {\\mathrm {f b}}\\right). \\tag {18}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.181, + 0.483, + 0.211 + ], + "angle": 0, + "content": "Following the same approach as in Eq. 9 and using Eq. 11, we modify the score in Eq. 8 as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.12, + 0.224, + 0.483, + 0.281 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\nabla_ {x _ {t}} \\log \\tilde {p} _ {\\theta} (x _ {t} | y _ {g}) = \\nabla_ {x _ {t}} \\log p _ {\\theta} (x _ {t}) \\\\ + w \\nabla_ {x _ {t}} \\left(\\log p _ {\\theta} (x _ {t}) - \\log p _ {\\theta} (x _ {t} | y _ {s})\\right) \\\\ + w \\nabla_ {x _ {t}} \\left(\\log p _ {\\theta} \\left(x _ {t}\\right) - \\log p _ {\\theta} \\left(x _ {t} \\mid y _ {t}\\right)\\right). \\tag {19} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.293, + 0.484, + 0.34 + ], + "angle": 0, + "content": "By replacing the score with the estimated denoiser and utilizing separate scales for spatial and temporal terms, we obtain the following equation:" + }, + { + "type": "equation", + "bbox": [ + 0.151, + 0.352, + 0.482, + 0.388 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\tilde {\\epsilon} _ {\\theta} ^ {w} (x _ {t}) = \\epsilon_ {\\theta} (x _ {t}) + w _ {1} \\left(\\epsilon_ {\\theta} (x _ {t}) - \\epsilon_ {\\theta} ^ {s} (x _ {t})\\right) \\\\ + w _ {2} \\left(\\epsilon_ {\\theta} \\left(x _ {t}\\right) - \\epsilon_ {\\theta} ^ {t} \\left(x _ {t}\\right)\\right) \\tag {20} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.401, + 0.483, + 0.432 + ], + "angle": 0, + "content": "where \\(\\epsilon_{\\theta}^{\\mathrm{s}}(x_t)\\) and \\(\\epsilon_{\\theta}^{\\mathrm{t}}(x_t)\\) represent spatially and temporally perturbed models, respectively." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.433, + 0.484, + 0.538 + ], + "angle": 0, + "content": "Next, we revisit our assumption of independence between spatial and temporal perturbations. While Eq. 20 works well in practice, we can derive an alternative STG formulation that uses orthogonalization to isolate the independent components of the spatial and temporal guidance. Inspired by a negative prompting technique [3], we can modify the Eq. 20 as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.106, + 0.555, + 0.483, + 0.61 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\tilde {\\epsilon} _ {\\theta} ^ {w} (x _ {t}) = \\epsilon_ {\\theta} (x _ {t}) + w _ {1} \\Delta_ {\\mathrm {s}} \\\\ + w _ {2} \\left(\\Delta_ {\\mathrm {t}} - \\frac {\\left\\langle \\Delta_ {\\mathrm {s}} , \\Delta_ {\\mathrm {t}} \\right\\rangle}{\\left\\| \\Delta_ {\\mathrm {s}} \\right\\| ^ {2}} \\Delta_ {\\mathrm {s}}\\right), \\tag {21} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.621, + 0.462, + 0.639 + ], + "angle": 0, + "content": "where \\(\\Delta_{\\mathrm{s}} = \\epsilon_{\\theta}(x_{t}) - \\epsilon_{\\theta}^{\\mathrm{s}}(x_{t})\\) and \\(\\Delta_{\\mathrm{t}} = \\epsilon_{\\theta}(x_{t}) - \\epsilon_{\\theta}^{\\mathrm{t}}(x_{t})\\)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.659, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Manifold Constrained Guidance Even with a well-aligned weak model, larger guidance scales \\( w \\) can drive samples off the data manifold, resulting in poor quality and oversaturation. To address error accumulation at high guidance scales, we explore optional techniques to keep samples constrained to the manifold. Rescaling the latent code [21] helps mitigate this issue by constraining its variance, as larger variance is known to cause saturation in the results. Additionally, Restart sampling [36] demonstrates that introducing stochasticity can correct off-manifold deviations. Building on this idea, we incorporate stochastic forward processes into our sampling guidance framework as an optional method. While this approach modestly improves final sample quality and reduces saturation, it introduces additional computational overhead. Further details are provided in Appendix A5.1." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.09, + 0.646, + 0.108 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.115, + 0.625, + 0.13 + ], + "angle": 0, + "content": "5.1. Overview" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.137, + 0.817, + 0.152 + ], + "angle": 0, + "content": "We employ three models for our experiments:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.154, + 0.905, + 0.199 + ], + "angle": 0, + "content": "- Mochi [33] is a text-to-video model built on AsymmDiT blocks, containing a total of 10 billion parameters and utilizing 3D self-attention in its spatiotemporal layers." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.2, + 0.905, + 0.244 + ], + "angle": 0, + "content": "- Open-Sora [38] is a text-to-video model built on STDiT blocks with 1.1 billion parameters, employing factorized spatial and temporal attention layers." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.245, + 0.905, + 0.289 + ], + "angle": 0, + "content": "- SVD [4] is an image-to-video model with 1.5 billion parameters that leverages factorized spatial and temporal attention within a UNet architecture." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.154, + 0.905, + 0.289 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.292, + 0.905, + 0.337 + ], + "angle": 0, + "content": "We evaluate the proposed method using widely adopted datasets and metrics to ensure a comprehensive performance analysis." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.339, + 0.906, + 0.475 + ], + "angle": 0, + "content": "- UCF-101 The UCF-101 dataset comprises 13,320 videos organized into 101 action classes. Using this dataset, we assess the Fréchet Video Distance (FVD) and Inception Score (IS) of the image-to-video model SVD [4]. Following DIGAN [37], we calculate FVD and IS on 2,048 and 10,000 samples, respectively. For conditioning, initial frames from UCF-101 videos are used as inputs to the SVD model, with each input frame serving as the starting frame of the generated videos." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.476, + 0.906, + 0.596 + ], + "angle": 0, + "content": "- VBench We use VBench [15] for automatic evaluation across various video metrics. SVD is evaluated using the image-to-video (I2V) framework on 355 samples from the VBench I2V dataset with 5 random seeds. Mochi and Open-Sora are evaluated with the text-to-video (T2V) framework: Open-Sora uses the standard VBench prompt list, while Mochi uses 100 randomly selected prompts due to computational limits." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.597, + 0.905, + 0.626 + ], + "angle": 0, + "content": "- EvalCrafter We perform human evaluations using 700 prompts from the EvalCrafter dataset [22]." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.627, + 0.905, + 0.672 + ], + "angle": 0, + "content": "- LLM-Generated Prompts We use a set of 100 selected prompts generated by Claude 3.5 Sonnet [2] for our demo and qualitative comparisons." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.476, + 0.906, + 0.672 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.674, + 0.906, + 0.735 + ], + "angle": 0, + "content": "For evaluation, CFG is used together with STG for our evaluation. We did not use Restart sampling or guidance orthogonalization; additional results with these techniques are in the Appendix sections A5.1 and A5.2." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.743, + 0.608, + 0.758 + ], + "angle": 0, + "content": "5.2. Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Qualitative Comparison Fig. 5 compares CFG and our method. Videos from the naive CFG model often show blurry objects with indistinct shapes, while STG produces clearer, more vivid frames with sharper image quality. STG also reduces temporal inconsistency and flickering, especially in dynamic videos with large motion where CFG frequently fails. Spatial guidance enhances object structure, and temporal guidance improves consistency. Please refer to the Appendix sections A5 and A6 for additional results." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11012" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.094, + 0.089, + 0.905, + 0.201 + ], + "angle": 0, + "content": "
ModelsFVD (↓)ISImaging QualityAesthetic QualityMotion SmoothnessDynamic DegreeTemporal Flickering
Mochi (STG-R)--0.6280.5540.9880.860.978
Mochi (STG-A)--0.5550.5410.9870.860.976
Open-Sora (STG-R)--0.5500.4740.9810.8940.977
Open-Sora (STG-A)--0.6060.5090.9870.8950.976
SVD (STG-R)155.939.30.6870.6370.9650.641-
SVD (STG-A)128.738.50.6940.6390.9680.694-
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.211, + 0.907, + 0.24 + ], + "angle": 0, + "content": "Table 3. Comparison of STG-R (residual skip) and STG-A (attention skip) across Mochi [33], Open-Sora [38], and SVD [4]. STG-R shows stronger performance on Mochi, while STG-A yields better results on Open-Sora and SVD." + }, + { + "type": "table", + "bbox": [ + 0.136, + 0.253, + 0.861, + 0.328 + ], + "angle": 0, + "content": "
ModelsFVD (↓)ISImaging QualityAesthetic QualityMotion SmoothnessDynamic Degree
CFG151.338.00.6870.6370.9660.562
+ Spatial133.838.30.6910.6390.9670.659
+ Temporal128.738.50.6940.6380.9680.694
" + }, + { + "type": "table_caption", + "bbox": [ + 0.128, + 0.338, + 0.866, + 0.352 + ], + "angle": 0, + "content": "Table 4. Ablation study results on SVD [4] factorized attention, showing the impact of adding spatial and temporal guidance." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.371, + 0.482, + 0.476 + ], + "angle": 0, + "content": "Quantitative Comparison We compare CFG and STG using the FVD-Imaging Quality (VBench) plot across different scales in Fig. 4. FVD measures video distribution, while Imaging Quality assesses frame clarity. Higher CFG scales improve Imaging Quality but reduce diversity, as reflected in higher FVD. STG avoids this trade-off, maintaining diversity at increased scales." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.477, + 0.483, + 0.568 + ], + "angle": 0, + "content": "T2V and I2V VBenchmark metrics in Tab.1 and Tab.2 show notable improvements in frame-level quality (Imaging and Aesthetic Quality). Temporal quality improves qualitatively, though metrics like Motion Smoothness and Temporal Flickering show marginal gains, as these scores are near saturation \\((\\sim 0.9x)\\)" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.568, + 0.483, + 0.673 + ], + "angle": 0, + "content": "For Dynamic Degree, we aim to keep it unchanged, as it may not correlate with video quality. This is achieved in the T2V metric, but in the I2V model, CFG increases the influence of the conditioned image, reducing motion. STG, while not directly affecting Dynamic Degree, mitigates CFG's impact, leading to increased Dynamic Degree in I2V when used together." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.673, + 0.483, + 0.719 + ], + "angle": 0, + "content": "Human Evaluation We provide human evaluation results on 700 prompts from the EvalCrafter dataset [22] in the Appendix A4." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.728, + 0.245, + 0.744 + ], + "angle": 0, + "content": "5.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.75, + 0.483, + 0.856 + ], + "angle": 0, + "content": "Fig. 3 displays selected frames from videos generated by Mochi using various STG scales, where \\( w = 0 \\) corresponds to the CFG-only model without STG. Increasing the STG scale results in more vivid colors and finer details compared to the monotonous colors at \\( w = 0 \\). Notably, unlike with CFG, sampling diversity is preserved as the STG scale increases, as shown in Fig. 4." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.856, + 0.483, + 0.901 + ], + "angle": 0, + "content": "We performed an ablation study on SVD to evaluate spatial and temporal guidance (Tab. 4). Spatial guidance notably reduced FVD and improved metrics, while tempo" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.371, + 0.905, + 0.402 + ], + "angle": 0, + "content": "ral guidance further boosted performance, confirming their combined effectiveness for spatiotemporal generation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.403, + 0.906, + 0.616 + ], + "angle": 0, + "content": "We also test two STG variants—residual skip (STGR) and attention skip (STG-A)—on SVD, Open-Sora, and Mochi. As shown in Tab. 3, STG-R performs well with Mochi, while STG-A is more effective for the other models. This is likely due to Mochi's higher parameter count and layer depth, enabling more extensive skipping without triggering out-of-distribution (OOD) issues in consecutive layers. Furthermore, Mochi's spatiotemporal layers consist of a single 3D attention layer, so STG-R skips a residual layer with just one attention layer. In contrast, SVD and Open-Sora use factorized attention, meaning STG-R skips a residual layer with two attention layers, which may cause excessive perturbation and lead to OOD issues in subsequent layers. More results are available in the Appendix A5." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.633, + 0.634, + 0.648 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.66, + 0.906, + 0.766 + ], + "angle": 0, + "content": "We proposed Spatiotemporal Skip Guidance (STG), a simple, training-free method for video diffusion models. By simulating an aligned weak model via spatiotemporal skipping, STG offers strong guidance for high-fidelity video generation, achieving notable qualitative and quantitative improvements. We hope this work advances video diffusion models and inspires further research in the field." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Limitation and Ethical Considerations STG's performance depends on scale and layer selection, with the optimal configuration varying across models, requiring users to set these through heuristic tuning. While video quality improvements are notable, they also raise ethical concerns about misuse, underscoring the importance of using this technology responsibly and constructively." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11013" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.212 + ], + "angle": 0, + "content": "Acknowledgements This work was supported by Institute for Information & communications Technology Planning & Evaluation(IITP) grant funded by the Korea government(MSIT) (RS-2019-II190075, Artificial Intelligence Graduate School Program(KAIST)). This work was supported by the National Research Foundation of Korea(NRF) grant funded by the Korea government(MSIT) (No. RS2025-00555621)." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.226, + 0.188, + 0.241 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.251, + 0.484, + 0.32 + ], + "angle": 0, + "content": "[1] Donghoon Ahn, Hyoungwon Cho, Jaewon Min, Wooseok Jang, Jungwoo Kim, SeonHwa Kim, Hyun Hee Park, Kyong Hwan Jin, and Seungryong Kim. Self-rectifying diffusion sampling with perturbed-attention guidance. arXiv preprint arXiv:2403.17377, 2024. 2, 3, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.322, + 0.482, + 0.348 + ], + "angle": 0, + "content": "[2] Anthropic. The claude 3 model family: Opus, sonnet, haiku. In-, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.351, + 0.483, + 0.419 + ], + "angle": 0, + "content": "[3] Mohammadreza Armandpour, Ali Sadeghian, Huangjie Zheng, Amir Sadeghian, and Mingyuan Zhou. Re-imagine the negative prompt algorithm: Transform 2d diffusion into 3d, alleviate janus problem and beyond. arXiv preprint arXiv:2304.04968, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.421, + 0.483, + 0.504 + ], + "angle": 0, + "content": "[4] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 2, 4, 5, 7, 8, 1, 9, 11, 13, 14, 17, 18, 23" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.506, + 0.482, + 0.545 + ], + "angle": 0, + "content": "[5] Andrew Brock. Large scale gan training for high fidelity natural image synthesis. arXiv preprint arXiv:1809.11096, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.548, + 0.483, + 0.603 + ], + "angle": 0, + "content": "[6] Hyungjin Chung, Jeongsol Kim, Geon Yeong Park, Hyelin Nam, and Jong Chul Ye. Cfg++: Manifold-constrained classifier free guidance for diffusion models. arXiv preprint arXiv:2406.08070, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.605, + 0.482, + 0.646 + ], + "angle": 0, + "content": "[7] Zihan Ding, Xiao-Yang Liu, Miao Yin, and Linghe Kong. Tgan: Deep tensor generative adversarial nets for large image generation. arXiv preprint arXiv:1901.09953, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.648, + 0.483, + 0.715 + ], + "angle": 0, + "content": "[8] Songwei Ge, Aniruddha Mahapatra, Gaurav Parmar, JunYan Zhu, and Jia-Bin Huang. On the content bias in fréchet video distance. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7277-7288, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.718, + 0.483, + 0.773 + ], + "angle": 0, + "content": "[9] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139-144, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.775, + 0.482, + 0.802 + ], + "angle": 0, + "content": "[10] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.804, + 0.483, + 0.845 + ], + "angle": 0, + "content": "[11] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.847, + 0.483, + 0.9 + ], + "angle": 0, + "content": "[12] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. Advances in Neural Information Processing Systems, 35:8633-8646, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.251, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.135 + ], + "angle": 0, + "content": "[13] Susung Hong. Smoothed energy guidance: Guiding diffusion models with reduced energy curvature of attention. arXiv preprint arXiv:2408.00760, 2024. 2, 3, 5, 8, 18" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.137, + 0.905, + 0.205 + ], + "angle": 0, + "content": "[14] Susung Hong, Gyuseong Lee, Wooseok Jang, and Seunglyong Kim. Improving sample quality of diffusion models using self-attention guidance. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7462-7471, 2023. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.208, + 0.907, + 0.29 + ], + "angle": 0, + "content": "[15] Ziqi Huang, Yinan He, Jiashuo Yu, Fan Zhang, Chenyang Si, Yuming Jiang, Yuanhan Zhang, Tianxing Wu, Qingyang Jin, Nattapol Chanpaisit, et al. Vbench: Comprehensive benchmark suite for video generative models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21807-21818, 2024. 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.292, + 0.905, + 0.373 + ], + "angle": 0, + "content": "[16] Yangqing Jia, Evan Shelhamer, Jeff Donahue, Sergey Karayev, Jonathan Long, Ross Girshick, Sergio Guadarrama, and Trevor Darrell. Caffe: Convolutional architecture for fast feature embedding. In Proceedings of the 22nd ACM international conference on Multimedia, pages 675-678, 2014." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.376, + 0.905, + 0.446 + ], + "angle": 0, + "content": "[17] Andrej Karpathy, George Toderici, Sanketh Shetty, Thomas Leung, Rahul Sukthankar, and Li Fei-Fei. Large-scale video classification with convolutional neural networks. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 1725–1732, 2014. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.448, + 0.907, + 0.517 + ], + "angle": 0, + "content": "[18] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8110-8119, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.519, + 0.905, + 0.573 + ], + "angle": 0, + "content": "[19] Tero Karras, Miika Aittala, Tuomas Kynkänniemi, Jaakko Lehtinen, Timo Aila, and Samuli Laine. Guiding a diffusion model with a bad version of itself. arXiv preprint arXiv:2406.02507, 2024. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.576, + 0.904, + 0.616 + ], + "angle": 0, + "content": "[20] Black Forest Labs. Flux: Github repository. https://github.com/black-forest-labs/flux, 2024. Accessed: 2024-11-22. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.619, + 0.905, + 0.687 + ], + "angle": 0, + "content": "[21] Shanchuan Lin, Bingchen Liu, Jiashi Li, and Xiao Yang. Common diffusion noise schedules and sample steps are flawed. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 5404-5411, 2024. 2, 7, 5, 10" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.69, + 0.907, + 0.773 + ], + "angle": 0, + "content": "[22] Yaofang Liu, Xiaodong Cun, Xuebo Liu, Xintao Wang, Yong Zhang, Haoxin Chen, Yang Liu, Tieyong Zeng, Raymond Chan, and Ying Shan. Evalcrafter: Benchmarking and evaluating large video generation models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22139-22149, 2024. 7, 8, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.775, + 0.905, + 0.842 + ], + "angle": 0, + "content": "[23] Nanye Ma, Mark Goldstein, Michael S Albergo, Nicholas M Boffi, Eric Vanden-Eijnden, and Saining Xie. Sit: Exploring flow and diffusion-based generative models with scalable interpolant transformers. arXiv preprint arXiv:2401.08740, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.905, + 0.899 + ], + "angle": 0, + "content": "[24] Xinyin Ma, Gongfan Fang, Michael Bi Mi, and Xinchao Wang. Learning-to-cache: Accelerating diffusion transformer via layer caching. arXiv preprint arXiv:2406.01733, 2024. 4" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11014" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.092, + 0.486, + 0.161 + ], + "angle": 0, + "content": "[25] Soumik Mukhopadhyay, Matthew Gwilliam, Vatsal Agarwal, Namitha Padmanabhan, Archana Swaminathan, Srinidhi Hegde, Tianyi Zhou, and Abhinav Shrivastava. Diffusion models beat gans on image classification. arXiv preprint arXiv:2307.08702, 2023. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.091, + 0.163, + 0.485, + 0.218 + ], + "angle": 0, + "content": "[26] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4195-4205, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.22, + 0.486, + 0.288 + ], + "angle": 0, + "content": "[27] Adam Polyak, Amit Zohar, Andrew Brown, Andros Tjandra, Animesh Sinha, Ann Lee, Apoorv Vyas, Bowen Shi, Chih-Yao Ma, Ching-Yao Chuang, et al. Movie gen: A cast of media foundation models. arXiv preprint arXiv:2410.13720, 2024. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.291, + 0.485, + 0.36 + ], + "angle": 0, + "content": "[28] Hubert Ramsauer, Bernhard Schäfl, Johannes Lehner, Philipp Seidl, Michael Widrich, Thomas Adler, Lukas Gruber, Markus Holzleitner, Milena Pavlovic, Geir Kjetil Sandve, et al. Hopfield networks is all you need. arXiv preprint arXiv:2008.02217, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.361, + 0.484, + 0.403 + ], + "angle": 0, + "content": "[29] Ali Razavi, Aaron Van den Oord, and Oriol Vinyals. Generating diverse high-fidelity images with vq-vae-2. Advances in neural information processing systems, 32, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.404, + 0.485, + 0.473 + ], + "angle": 0, + "content": "[30] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.475, + 0.485, + 0.532 + ], + "angle": 0, + "content": "[31] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International conference on machine learning, pages 2256-2265. PMLR, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.532, + 0.485, + 0.588 + ], + "angle": 0, + "content": "[32] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.589, + 0.485, + 0.63 + ], + "angle": 0, + "content": "[33] Genmo Team. Mochi 1. https://github.com/ genmoai/models, 2024. 2, 4, 5, 6, 7, 8, 1, 9, 10, 15, 19, 20" + }, + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.632, + 0.485, + 0.7 + ], + "angle": 0, + "content": "[34] Du Tran, Lubomir Bourdev, Rob Fergus, Lorenzo Torresani, and Manohar Paluri. Learning spatiotemporal features with 3d convolutional networks. In Proceedings of the IEEE international conference on computer vision, pages 4489-4497, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.702, + 0.485, + 0.743 + ], + "angle": 0, + "content": "[35] Pascal Vincent. A connection between score matching and denoising autoencoders. Neural computation, 23(7):1661-1674, 2011. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.745, + 0.485, + 0.801 + ], + "angle": 0, + "content": "[36] Yilun Xu, Mingyang Deng, Xiang Cheng, Yonglong Tian, Ziming Liu, and Tommi Jaakkola. Restart sampling for improving generative processes. Advances in Neural Information Processing Systems, 36:76806-76838, 2023. 2, 7, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.802, + 0.485, + 0.858 + ], + "angle": 0, + "content": "[37] Sihyun Yu, Jihoon Tack, Sangwoo Mo, Hyunsu Kim, Junho Kim, Jung-Woo Ha, and Jinwoo Shin. Generating videos with dynamics-aware implicit generative adversarial networks. arXiv preprint arXiv:2202.10571, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.859, + 0.485, + 0.902 + ], + "angle": 0, + "content": "[38] Zangwei Zheng, Xiangyu Peng, Tianji Yang, Chenhui Shen, Shenggui Li, Hongxin Liu, Yukun Zhou, Tianyi Li, and Yang You. Open-sora: Democratizing efficient video production" + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.092, + 0.486, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.547, + 0.092, + 0.907, + 0.12 + ], + "angle": 0, + "content": "for all. https://github.com/hpcaitech/Open-Sora, 2024. 2, 4, 5, 7, 8, 1, 12, 16, 21, 22" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11015" + } + ] +] \ No newline at end of file diff --git a/2025/Spatiotemporal Skip Guidance for Enhanced Video Diffusion Sampling/4ec7da94-3031-45c8-a8d6-19bf226e3cba_origin.pdf b/2025/Spatiotemporal Skip Guidance for Enhanced Video Diffusion Sampling/4ec7da94-3031-45c8-a8d6-19bf226e3cba_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..62dab9dc0e080632af8bf4ab58c1a97e50974979 --- /dev/null +++ b/2025/Spatiotemporal Skip Guidance for Enhanced Video Diffusion Sampling/4ec7da94-3031-45c8-a8d6-19bf226e3cba_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:14a6024e3d254f63f0f6fb025b1e5f2ea42e491a32bddd25e4d7fdcf67fb5b3a +size 9995422 diff --git a/2025/Spatiotemporal Skip Guidance for Enhanced Video Diffusion Sampling/full.md b/2025/Spatiotemporal Skip Guidance for Enhanced Video Diffusion Sampling/full.md new file mode 100644 index 0000000000000000000000000000000000000000..21721475e5b33d0119a30a5a26581b667fcb2aee --- /dev/null +++ b/2025/Spatiotemporal Skip Guidance for Enhanced Video Diffusion Sampling/full.md @@ -0,0 +1,369 @@ +# Spatiotemporal Skip Guidance for Enhanced Video Diffusion Sampling + +Junha Hyung $^{*1}$ Kinam Kim $^{*1}$ Susung Hong $^{2}$ Min-Jung Kim $^{1}$ Jaegul Choo $^{1}$ + +$^{1}$ KAIST AI $^{2}$ University of Washington + +{sharpeeee, kinamplify, emjay73, jchoo}@kaist.ac.kr, susung@cs.washington.edu + +"A close-up shot of a butterfly landing on the nose of a woman, highlighting her smile and the details of the butterfly's wings." + +"A close-up of a woman's face with colored powder exploding around her, creating an abstract splash of vibrant hues." + +Figure 1. Visual comparison of video quality between CFG (top row) and our STG method (bottom row). Best viewed in Acrobat Reader; click on the images to watch the videos. + +# Abstract + +Diffusion models have emerged as a powerful tool for generating high-quality images, videos, and 3D content. While sampling guidance techniques like CFG improve quality, they reduce diversity and motion. Autoguidance mitigates these issues but demands extra weak model training, limiting its practicality for large-scale models. In this + +work, we introduce Spatiotemporal Skip Guidance (STG), a simple training-free sampling guidance method for enhancing transformer-based video diffusion models. STG employs an implicit weak model via self-perturbation, avoiding the need for external models or additional training. By selectively skipping spatiotemporal layers, STG produces an aligned, degraded version of the original model to boost sample quality without compromising diversity or dynamic degree. Our contributions include: (1) introducing STG as + +an efficient, high-performing guidance technique for video diffusion models, (2) eliminating the need for auxiliary models by simulating a weak model through layer skipping, and (3) ensuring quality-enhanced guidance without compromising sample diversity or dynamics unlike CFG. For additional results, visit https://junhahyung.github.io/STGuidance/. + +# 1. Introduction + +Diffusion models [11, 30-32] are a successful class of generative models known for their flexibility in modeling complex data distributions, achieving impressive results in image, video, and 3D generation. By progressively denoising random noise, they enable robust generalization, making them a leading choice for realistic content generation and often surpassing GAN-based methods [5, 9, 18, 29]. Building on this success, video diffusion models [4, 12, 27, 33, 38] generate high-quality videos by using temporal or 3D attention layers to handle sequential frames. + +Meanwhile, to enhance sample quality, sampling guidance techniques such as Classifier-Free Guidance (CFG) [10] and Autoguidance [19] have been introduced to guide the denoising process. These techniques employ weak models to predict poor trajectories, steering the main model away from them and pushing samples toward high-quality regions on the data manifold. However, CFG often reduces diversity, leading to saturated or overly simplified results [6, 19]. + +Autoguidance [19] addresses this issue by using a weak model trained on the same task, conditioning, and data distribution as the main model. The main drawback of this approach, however, is the need to train an additional weak model, which is impractical for large-scale models. Alternative methods such as PAG [1] and SEG [13] use self-perturbation to implicitly mimic a weak model, avoiding the need for extra training. Yet, these methods are designed specifically for image generation diffusion models, applying self-perturbation to 2D spatial attention maps. + +In this work, we propose Spatiotemporal Skip Guidance (STG), a simple and effective sampling guidance method for video diffusion models that significantly enhances the performance of any transformer-based video diffusion model without additional training. Specifically, we use an implicit weak model for guidance through self-perturbation, eliminating the need for explicit weak models and their associated training costs. This is especially crucial for video diffusion models, where training costs are high. + +Our implicit weak model, a key component of our framework, is deliberately designed as a degraded but "aligned" version of the original video generation model. As demonstrated in Autoguidance [19], having the weak model share the same task, conditioning, and data distribution as the + +![](images/a9ff040f709f6d6ceae9ae479c4eebc42adde4b366b565ff287709d970e93b04.jpg) +CFG + +![](images/4437acdf9142fc0f4a69a3b9477ef08b528ab6cae40df116cb29b121c290f3cb.jpg) +Ours +Figure 2. Comparison between CFG and STG, with the band conceptually representing the noisy data manifold. In STG, the weak model and the main model are aligned along the direction of increasing quality. In contrast, the two models in CFG differ not only in quality but also in aspects such as diversity and prompt alignment capabilities. + +main model is essential for quality improvement, as we expect both models to exhibit similar, aligned errors in the same parts of samples. We conceptually illustrate this alignment in Fig. 2, where our weak model and main model are arranged in a direction of increasing quality. To achieve this, we apply spatiotemporal perturbation to both spatial and temporal attention layers—or, in the case of 3D attention, to the entire layer—by selectively skipping certain layers. This straightforward approach effectively nullifies specific residual or attention layers, generating a lower-quality version of the model that simulates an aligned weak model. + +To ensure samples remain on the data manifold, even when using large guidance scales, we employ optional techniques such as the rescaling [21] and restart [36] methods. Rescaling the latent code constrains its variance, addressing the issue of larger variance causing saturation in the results [21]. Meanwhile, the restart sampling technique leverages the error contraction property of forward SDEs [36] to keep the sampling trajectory on the manifold. These techniques help prevent overshooting in the sampling trajectory, which could otherwise push samples off the manifold and result in saturated or distorted outputs. + +Our key contributions are as follows: + +- We propose STG—a surprisingly simple sampling guidance framework that significantly boosts the performance of video diffusion models. +- Our method introduces an implicit weak model by skipping spatiotemporal layers in video diffusion models, eliminating the need for additional training or external models. +- Our method enhances sample quality during guidance without reducing diversity or limiting the dynamics of generated videos. + +# 2. Related Work + +Guidance with trained weak model Classifier-Free Guidance (CFG) [10] improves conditional generation in Diffusion Models by using an implicit unconditional model as a weak model. However, differences in tasks between the unconditional and conditional models can reduce sample diversity [6, 19] and increase sampling trajectory curvature [6], leading to overshooting the data manifold and producing skewed or oversaturated images. + +Autoguidance [19] mitigates these issues by employing a bad version of the main model as a weak model, trained with reduced capacity and compute to ensure alignment. This alignment allows the guidance algorithm to correct errors by analyzing prediction differences. While effective, it requires additional training, which is challenging for large-scale video diffusion models. + +Guidance with training-free weak model Another line of work avoids additional training by using self-perturbation of the main model to mimic a weak model. Self-Attention Guidance (SAG) [14] blurs high-attention regions, Perturbed Attention Guidance (PAG) [1] replaces attention maps with identity matrices, and Smoothed Energy Guidance (SEG) [13] applies Gaussian blur to the attention weights to smooth the energy landscape. These methods guide sampling toward high-quality outputs by leveraging differences in predictions from the weakened model. While effective, they are primarily designed for image diffusion models with 2D self-attention. We aim to extend this approach to video diffusion models, which require handling temporal dynamics with additional temporal or 3D spatiotemporal attention layers. + +# 3. Preliminaries + +# 3.1. Diffusion models + +Diffusion models generate samples by progressively removing noise from noisy data, restoring the original data distribution. Song et al. [32] defined the process of adding noise to the data using a stochastic differential equation (SDE): + +$$ +d x = - \frac {\beta (t)}{2} x d t + \sqrt {\beta (t)} d w, \tag {1} +$$ + +where $\beta (t)$ is a time-dependent noise schedule, and $w$ represents the standard Wiener process. Corresponding reverse-time SDE is: + +$$ +d x = \left[ - \frac {\beta (t)}{2} x - \beta (t) \nabla_ {x} \log p _ {t} (x) \right] d t + \sqrt {\beta (t)} d \bar {w} \tag {2} +$$ + +where $d\bar{w}$ is the Wiener process running backward in time. Here, the score function $\nabla_x\log p_t(x)$ is approximated by + +a neural network $s_{\theta}(x(t))$ trained using denoising score matching [35]: + +$$ +\begin{array}{l} \theta^ {*} = \arg \min _ {\theta} \mathbb {E} _ {t} \left\{\lambda (t) \mathbb {E} _ {x _ {0}} \mathbb {E} _ {x _ {t} | x _ {0}} \right[ \\ \left. \left\| s _ {\theta} \left(x _ {t}\right) - \nabla_ {x _ {t}} \log p _ {t} \left(x _ {t} \mid x _ {0}\right) \| _ {2} ^ {2} \right] \right\}. \tag {3} \\ \end{array} +$$ + +# 3.2. Classifier Guidance + +Classifier Guidance (CG) [25] reformulates the reverse process of a diffusion model by incorporating an external classifier $p_{\phi}$ as + +$$ +\tilde {p _ {\theta}} \left(x _ {t} | y\right) \propto p _ {\theta} \left(x _ {t}\right) p _ {\phi} \left(y \mid x _ {t}\right) ^ {\lambda}, \tag {4} +$$ + +where $y$ is the desired class label and $\lambda$ controls the guidance strength. The score function is then derived as: + +$$ +\begin{array}{l} \nabla_ {x _ {t}} \log \tilde {p} _ {\theta} (x _ {t} | y) = \nabla_ {x _ {t}} \log p _ {\theta} (x _ {t}) \\ + \lambda \nabla_ {x _ {t}} \log p _ {\phi} (y | x _ {t}). \tag {5} \\ \end{array} +$$ + +By substituting the score function in Eq. 2 with Eq. 5, sampling from the desired class condition $y$ becomes possible. + +# 3.3. Classifier-Free Guidance + +Classifier-Free Guidance (CFG) [10] uses Bayes' rule to replace a classifier-guided score with a linear combination of conditional and unconditional score estimates: + +$$ +\tilde {\epsilon} _ {\theta} ^ {\lambda} \left(x _ {t}\right) = \epsilon_ {\theta} \left(x _ {t}\right) + \lambda \left(\epsilon_ {\theta} \left(x _ {t}\right) - \epsilon_ {\theta} \left(x _ {t} \mid \phi\right)\right). \tag {6} +$$ + +CFG jointly trains the unconditional model $\epsilon_{\theta}(x_t|\phi)$ and the conditional model $\epsilon_{\theta}(x_t|c) (= \epsilon_{\theta}(x_t))$ within a single model by setting the condition $c$ to a null token $\phi$ . Using this guided denoising process, the model better captures the conditions and often generates high-fidelity outputs. + +# 4. Method + +# 4.1. Optimal Weak Model Design + +Our goal is to construct an aligned weak model that captures a distribution similar to the original model while generating slightly lower-quality samples. This ensures that the guidance gradient points toward improved quality, as shown in Fig. 2. Misaligned models, as seen in CFG [10], may lead to unintended outcomes like reduced diversity, while Autoguidance [19] requires additional weak model training, making it impractical for large-scale models. + +To address these limitations, we design an implicit weak model by leveraging the original model itself, eliminating the need for external training. This training-free approach ensures computational efficiency and better alignment, as the weak model is derived directly from the original, preserving most network weights. + +To achieve this, we apply perturbation methods directly to the main model's forward pass, creating an implicit weak + +
ModelsImaging QualityAesthetic QualityMotion SmoothnessDynamic DegreeTemporal Flickering
Mochi (CFG)0.5240.5070.9850.870.976
Mochi (STG)0.6280.5540.9880.860.978
Open-Sora (CFG)0.5610.4930.9820.9020.975
Open-Sora (STG)0.6060.5090.9870.8950.976
+ +Table 1. Quantitative results for Mochi [33] and Open-Sora [38] on VBench [15] T2V benchmarks. + +
ModelsFVD (↓)ISImaging QualityAesthetic QualityMotion SmoothnessDynamic Degree
SVD (CFG)151.338.00.6870.6370.9660.562
SVD (STG)128.738.50.6940.6390.9680.694
+ +Table 2. Quantitative results for SVD [4] on FVD, IS, and VBenchmark [15] I2V benchmarks. + +model. Specifically, we use spatiotemporal perturbation, a key component for aligning the weak model with video diffusion models, which will be elaborated upon in detail. + +# 4.2. Sampling from High Quality Samples + +Similar to CG [25], we define our goal as conditioning the model on an imaginary label $y_{g}$ that represents high-quality samples, leading to the sampling distribution + +$$ +\tilde {p} _ {\theta} \left(x _ {t} \mid y _ {g}\right) \propto p _ {\theta} \left(x _ {t}\right) p _ {\phi} \left(y _ {g} \mid x _ {t}\right) ^ {w}. \tag {7} +$$ + +Using $w > 0$ sharpens the distribution, promoting the generation of high-quality samples. From Eq. 7, the score is derived as + +$$ +\begin{array}{l} \nabla_ {x _ {t}} \log \tilde {p} _ {\theta} (x _ {t} | y _ {g}) \\ = \nabla_ {x _ {t}} \log p _ {\theta} (x _ {t}) + w \nabla_ {x _ {t}} \log p _ {\phi} (y _ {g} | x _ {t}). \tag {8} \\ \end{array} +$$ + +Rather than using an external classifier $p_{\phi}$ , we propose using our model $p_{\theta}$ as an implicit classifier. We design this classifier to be inversely proportional to the probability of an imaginary "bad" label $y_{b}$ , expressed as + +$$ +p _ {\phi} \left(y _ {g} \mid x _ {t}\right) \propto \frac {1}{p _ {\theta} \left(y _ {b} \mid x _ {t}\right)}. \tag {9} +$$ + +Using Bayes' rule, the function can be expressed in terms of the marginal posterior as follows: + +$$ +\begin{array}{l} \nabla_ {x _ {t}} \log \frac {1}{p _ {\theta} (y _ {b} | x _ {t})} = \nabla_ {x _ {t}} \log \frac {p _ {\theta} (x _ {t})}{p _ {\theta} (y _ {b}) p _ {\theta} (x _ {t} | y _ {b})} \\ = \nabla_ {x _ {t}} \left(\log p _ {\theta} \left(x _ {t}\right) - \log p _ {\theta} \left(x _ {t} \mid y _ {b}\right)\right), \tag {10} \\ \end{array} +$$ + +leading to the score of the target distribution as: + +$$ +\begin{array}{l} \nabla_ {x _ {t}} \log \tilde {p} _ {\theta} (x _ {t} | y _ {g}) = \nabla_ {x _ {t}} \log p _ {\theta} (x _ {t}) \\ + w \nabla_ {x _ {t}} \left(\log p _ {\theta} \left(x _ {t}\right) - \log p _ {\theta} \left(x _ {t} \mid y _ {b}\right)\right). \tag {11} \\ \end{array} +$$ + +We can sample from $\tilde{p}_{\theta}(x_t|y_g)$ by substituting the score function in Eq. 2 with Eq. 11, resulting in: + +$$ +\begin{array}{l} d x = \left[ - \frac {\beta (t)}{2} x - \beta (t) \left(\nabla_ {x _ {t}} \log p _ {t} \left(x _ {t}\right) \right. \right. \tag {12} \\ \left. + w \nabla_ {x _ {t}} (\log p _ {\theta} (x _ {t}) - \log p _ {\theta} (x _ {t} | y _ {b})) \right] d t + \sqrt {\beta (t)} d \bar {w}, \\ \end{array} +$$ + +and solving the reverse SDE. Since the score function is approximated using the neural network, we can generate samples using + +$$ +\tilde {\epsilon} _ {\theta} ^ {w} (x _ {t}) = \epsilon_ {\theta} (x _ {t}) + w \left(\epsilon_ {\theta} (x _ {t}) - \epsilon_ {\theta} ^ {b} (x _ {t})\right). \tag {13} +$$ + +An interesting approach here is to model $\nabla_{x_t}\log p_\theta (x_t|y_b)$ by perturbing the forward pass of $\epsilon_{\theta}(x_{t})$ , denoted as $\epsilon_{\theta}^{b}(x_{t})$ . Now the main focus is designing a perturbation that effectively yields the weak model capable of estimating $\epsilon_{\theta}^{b}(x_{t})$ , aligning closely with $\epsilon_{\theta}(x_{t})$ as discussed in Sec. 4.1. Ideally, we want a distribution that deviates minimally from $\epsilon_{\theta}(x_{t})$ while producing slightly lower-quality samples. + +# 4.3. Spatiotemporal Skip Guidance (STG) + +We introduce Spatiotemporal Skip Guidance (STG), a simple guidance method designed for video diffusion models that generates diverse, high-fidelity samples while addressing the limitations of existing methods. STG implicitly simulates an aligned weak model through spatiotemporal perturbation, capturing the spatiotemporal dynamics of video data. + +An interesting discovery of this paper is that skipping layers within the network is an effective approach for constructing an aligned weak model. Modern neural network architectures for video diffusion models, such as ADM [25], DiT [26], and SiT [23], are partially or fully transformer-based and contain attention layers and residual blocks. These architectures are well-suited for layer skipping, as + +![](images/5070aabb09be4e8193c02ce195eebaa51b7ce9f83b565d3636ae12aadeb0475b.jpg) +Figure 3. Selected frames from videos generated by Mochi [33] with increasing STG scales. + +![](images/3ec1b25622f72d9107003e686df759da57fe5afc1b25b233f755826228815b4e.jpg) +Figure 4. Comparison of CFG and STG across varying scales in terms of Imaging Quality and FVD. + +they can still produce plausible outputs without significant degradation, even when a few layers are removed. + +Layer skipping is advantageous for generating an aligned weak model because it retains most of the neural network's weights and forward pass, resulting in similar predictions and distributions. This approach offers a clear advantage over methods that rely on external models, such as Autoguidance [19], or alternative objectives like CFG [10], where forward passes differ significantly. + +Moreover, our method applies perturbations to both spatial and temporal layers (or spatiotemporal layers), unlike existing image-based perturbation methods [13, 14] limited to 2D spatial attention layers. This dual-layer perturbation is essential for aligning the weak model with the main video diffusion model. We now discuss various STG configurations that can be applied across different video diffusion model architectures. + +Residual skip To skip an entire residual block, we modify $\operatorname{Res}(z_l)$ to $\operatorname{Res}'(z_l)$ , + +$$ +\operatorname {R e s} \left(z _ {l}\right) = z _ {l + 1} = f _ {l} \left(z _ {l}\right) + z _ {l}, \tag {14} +$$ + +$$ +\operatorname {R e s} ^ {\prime} \left(z _ {l}\right) = z _ {l + 1} = z _ {l}, \tag {15} +$$ + +where $z_{l}$ and $z_{l + 1}$ represent the features at the $l^{\mathrm{th}}$ and $(l + 1)^{\mathrm{th}}$ layers, respectively, and $f_{l}$ denotes the $l^{\mathrm{th}}$ neural net layer. The residual layers, which add small residuals $f_{l}(z_{l})$ to the original feature $z_{l}$ , ensure that the perturbed $z_{l + 1}$ does not deviate significantly from the original $z_{l + 1}$ . + +This reduces out-of-distribution issues in consecutive layers, generating perturbed but aligned samples. + +Attention skip Self-attention computes a linear combination of value tensors as + +$$ +\operatorname {S A} (Q, K, V) = \operatorname {S o f t m a x} \left(\frac {Q K ^ {T}}{\sqrt {d}}\right) V = \mathbf {A} V, \tag {16} +$$ + +where $Q \in \mathbb{R}^{(h \times w \times f) \times d}$ , $K \in \mathbb{R}^{(h \times w \times f) \times d}$ , $V \in \mathbb{R}^{(h \times w \times f) \times d}$ are the query, key, and value matrices, respectively. Here, $h$ , $w$ , $f$ , and $d$ represent the height, width, frame number, and channel dimensions. + +We can skip this layer partially by passing the value matrix directly to the next layer without computing its linear combination. This is equivalent to replacing the attention matrix $\mathbf{A}$ with an identity matrix $\mathbf{I} \in \mathbb{R}^{hwf \times hwf}$ , resulting in + +$$ +\mathrm {S A} ^ {\prime} (Q, K, V) = \mathbf {I} V. \tag {17} +$$ + +This represents a 3D extension of PAG [1]. + +Factorized attention While recent models like Movie Gen [27] and Mochi [33] utilize full 3D spatiotemporal attention layers, many models, such as SVD [4] and OpenSora [38], still use factorized attention layers for efficiency. These models employ sequential 2D spatial attention and 1D temporal attention to approximate 3D spatiotemporal attention. For factorized models, we apply skip perturbation of Eq. 17 to spatial and temporal layers separately and use their linear combination for the final guidance. + +We denote the spatial and temporal perturbation labels as $y_{sb}$ and $y_{tb}$ , respectively. In practice, $y_{sb}$ and $y_{tb}$ are rarely independent, as they can influence each other. For instance, random spatial perturbations that create varied color adjustments across frames may disrupt temporal continuity. Similarly, random temporal perturbations can affect spatial consistency, leading to distortions in specific frames. However, for our skip perturbations, we can loosely assume independence, as skipping layers primarily reduces details in residual networks, which may not inherently affect temporal consistency. + +Therefore, to simplify the derivation, we initially assume independence between spatial and temporal perturbations. + +(1) A romantic scene of a couple dancing under string lights in a backyard, with warm, golden tones highlighting their laughter. +(2) An animation showing a floating castle drifting above the clouds, with birds flying around it and sunlight casting golden rays ... +(3) A realistic documentary-style video of artisans crafting pottery, with the scene unfolding and transforming as hands shape clay ... +(4) A ghost in a white bedsheet faces a mirror. The ghost's reflection can be seen in the mirror. The ghost is in a dusty attic ... + +![](images/70c7a672903779f88253a7cf57412338b5e58688bb386b7bc4f3512985c07ec8.jpg) +Figure 5. Qualitative comparison between CFG and STG on videos generated by Mochi [33]. + +We will revisit this assumption and propose an adjustment afterward. Under this independence assumption, the joint distribution can be expressed as + +$$ +p _ {\theta} \left(x _ {t} \mid y _ {b}\right) = p _ {\theta} \left(x _ {t} \mid y _ {\mathrm {s b}}\right) p _ {\theta} \left(x _ {t} \mid y _ {\mathrm {f b}}\right). \tag {18} +$$ + +Following the same approach as in Eq. 9 and using Eq. 11, we modify the score in Eq. 8 as follows: + +$$ +\begin{array}{l} \nabla_ {x _ {t}} \log \tilde {p} _ {\theta} (x _ {t} | y _ {g}) = \nabla_ {x _ {t}} \log p _ {\theta} (x _ {t}) \\ + w \nabla_ {x _ {t}} \left(\log p _ {\theta} (x _ {t}) - \log p _ {\theta} (x _ {t} | y _ {s})\right) \\ + w \nabla_ {x _ {t}} \left(\log p _ {\theta} \left(x _ {t}\right) - \log p _ {\theta} \left(x _ {t} \mid y _ {t}\right)\right). \tag {19} \\ \end{array} +$$ + +By replacing the score with the estimated denoiser and utilizing separate scales for spatial and temporal terms, we obtain the following equation: + +$$ +\begin{array}{l} \tilde {\epsilon} _ {\theta} ^ {w} (x _ {t}) = \epsilon_ {\theta} (x _ {t}) + w _ {1} \left(\epsilon_ {\theta} (x _ {t}) - \epsilon_ {\theta} ^ {s} (x _ {t})\right) \\ + w _ {2} \left(\epsilon_ {\theta} \left(x _ {t}\right) - \epsilon_ {\theta} ^ {t} \left(x _ {t}\right)\right) \tag {20} \\ \end{array} +$$ + +where $\epsilon_{\theta}^{\mathrm{s}}(x_t)$ and $\epsilon_{\theta}^{\mathrm{t}}(x_t)$ represent spatially and temporally perturbed models, respectively. + +Next, we revisit our assumption of independence between spatial and temporal perturbations. While Eq. 20 works well in practice, we can derive an alternative STG formulation that uses orthogonalization to isolate the independent components of the spatial and temporal guidance. Inspired by a negative prompting technique [3], we can modify the Eq. 20 as follows: + +$$ +\begin{array}{l} \tilde {\epsilon} _ {\theta} ^ {w} (x _ {t}) = \epsilon_ {\theta} (x _ {t}) + w _ {1} \Delta_ {\mathrm {s}} \\ + w _ {2} \left(\Delta_ {\mathrm {t}} - \frac {\left\langle \Delta_ {\mathrm {s}} , \Delta_ {\mathrm {t}} \right\rangle}{\left\| \Delta_ {\mathrm {s}} \right\| ^ {2}} \Delta_ {\mathrm {s}}\right), \tag {21} \\ \end{array} +$$ + +where $\Delta_{\mathrm{s}} = \epsilon_{\theta}(x_{t}) - \epsilon_{\theta}^{\mathrm{s}}(x_{t})$ and $\Delta_{\mathrm{t}} = \epsilon_{\theta}(x_{t}) - \epsilon_{\theta}^{\mathrm{t}}(x_{t})$ + +Manifold Constrained Guidance Even with a well-aligned weak model, larger guidance scales $w$ can drive samples off the data manifold, resulting in poor quality and oversaturation. To address error accumulation at high guidance scales, we explore optional techniques to keep samples constrained to the manifold. Rescaling the latent code [21] helps mitigate this issue by constraining its variance, as larger variance is known to cause saturation in the results. Additionally, Restart sampling [36] demonstrates that introducing stochasticity can correct off-manifold deviations. Building on this idea, we incorporate stochastic forward processes into our sampling guidance framework as an optional method. While this approach modestly improves final sample quality and reduces saturation, it introduces additional computational overhead. Further details are provided in Appendix A5.1. + +# 5. Experiments + +# 5.1. Overview + +We employ three models for our experiments: + +- Mochi [33] is a text-to-video model built on AsymmDiT blocks, containing a total of 10 billion parameters and utilizing 3D self-attention in its spatiotemporal layers. +- Open-Sora [38] is a text-to-video model built on STDiT blocks with 1.1 billion parameters, employing factorized spatial and temporal attention layers. +- SVD [4] is an image-to-video model with 1.5 billion parameters that leverages factorized spatial and temporal attention within a UNet architecture. + +We evaluate the proposed method using widely adopted datasets and metrics to ensure a comprehensive performance analysis. + +- UCF-101 The UCF-101 dataset comprises 13,320 videos organized into 101 action classes. Using this dataset, we assess the Fréchet Video Distance (FVD) and Inception Score (IS) of the image-to-video model SVD [4]. Following DIGAN [37], we calculate FVD and IS on 2,048 and 10,000 samples, respectively. For conditioning, initial frames from UCF-101 videos are used as inputs to the SVD model, with each input frame serving as the starting frame of the generated videos. + +- VBench We use VBench [15] for automatic evaluation across various video metrics. SVD is evaluated using the image-to-video (I2V) framework on 355 samples from the VBench I2V dataset with 5 random seeds. Mochi and Open-Sora are evaluated with the text-to-video (T2V) framework: Open-Sora uses the standard VBench prompt list, while Mochi uses 100 randomly selected prompts due to computational limits. +- EvalCrafter We perform human evaluations using 700 prompts from the EvalCrafter dataset [22]. +- LLM-Generated Prompts We use a set of 100 selected prompts generated by Claude 3.5 Sonnet [2] for our demo and qualitative comparisons. + +For evaluation, CFG is used together with STG for our evaluation. We did not use Restart sampling or guidance orthogonalization; additional results with these techniques are in the Appendix sections A5.1 and A5.2. + +# 5.2. Results + +Qualitative Comparison Fig. 5 compares CFG and our method. Videos from the naive CFG model often show blurry objects with indistinct shapes, while STG produces clearer, more vivid frames with sharper image quality. STG also reduces temporal inconsistency and flickering, especially in dynamic videos with large motion where CFG frequently fails. Spatial guidance enhances object structure, and temporal guidance improves consistency. Please refer to the Appendix sections A5 and A6 for additional results. + +
ModelsFVD (↓)ISImaging QualityAesthetic QualityMotion SmoothnessDynamic DegreeTemporal Flickering
Mochi (STG-R)--0.6280.5540.9880.860.978
Mochi (STG-A)--0.5550.5410.9870.860.976
Open-Sora (STG-R)--0.5500.4740.9810.8940.977
Open-Sora (STG-A)--0.6060.5090.9870.8950.976
SVD (STG-R)155.939.30.6870.6370.9650.641-
SVD (STG-A)128.738.50.6940.6390.9680.694-
+ +Table 3. Comparison of STG-R (residual skip) and STG-A (attention skip) across Mochi [33], Open-Sora [38], and SVD [4]. STG-R shows stronger performance on Mochi, while STG-A yields better results on Open-Sora and SVD. + +
ModelsFVD (↓)ISImaging QualityAesthetic QualityMotion SmoothnessDynamic Degree
CFG151.338.00.6870.6370.9660.562
+ Spatial133.838.30.6910.6390.9670.659
+ Temporal128.738.50.6940.6380.9680.694
+ +Table 4. Ablation study results on SVD [4] factorized attention, showing the impact of adding spatial and temporal guidance. + +Quantitative Comparison We compare CFG and STG using the FVD-Imaging Quality (VBench) plot across different scales in Fig. 4. FVD measures video distribution, while Imaging Quality assesses frame clarity. Higher CFG scales improve Imaging Quality but reduce diversity, as reflected in higher FVD. STG avoids this trade-off, maintaining diversity at increased scales. + +T2V and I2V VBenchmark metrics in Tab.1 and Tab.2 show notable improvements in frame-level quality (Imaging and Aesthetic Quality). Temporal quality improves qualitatively, though metrics like Motion Smoothness and Temporal Flickering show marginal gains, as these scores are near saturation $(\sim 0.9x)$ + +For Dynamic Degree, we aim to keep it unchanged, as it may not correlate with video quality. This is achieved in the T2V metric, but in the I2V model, CFG increases the influence of the conditioned image, reducing motion. STG, while not directly affecting Dynamic Degree, mitigates CFG's impact, leading to increased Dynamic Degree in I2V when used together. + +Human Evaluation We provide human evaluation results on 700 prompts from the EvalCrafter dataset [22] in the Appendix A4. + +# 5.3. Ablation Study + +Fig. 3 displays selected frames from videos generated by Mochi using various STG scales, where $w = 0$ corresponds to the CFG-only model without STG. Increasing the STG scale results in more vivid colors and finer details compared to the monotonous colors at $w = 0$ . Notably, unlike with CFG, sampling diversity is preserved as the STG scale increases, as shown in Fig. 4. + +We performed an ablation study on SVD to evaluate spatial and temporal guidance (Tab. 4). Spatial guidance notably reduced FVD and improved metrics, while tempo + +ral guidance further boosted performance, confirming their combined effectiveness for spatiotemporal generation. + +We also test two STG variants—residual skip (STGR) and attention skip (STG-A)—on SVD, Open-Sora, and Mochi. As shown in Tab. 3, STG-R performs well with Mochi, while STG-A is more effective for the other models. This is likely due to Mochi's higher parameter count and layer depth, enabling more extensive skipping without triggering out-of-distribution (OOD) issues in consecutive layers. Furthermore, Mochi's spatiotemporal layers consist of a single 3D attention layer, so STG-R skips a residual layer with just one attention layer. In contrast, SVD and Open-Sora use factorized attention, meaning STG-R skips a residual layer with two attention layers, which may cause excessive perturbation and lead to OOD issues in subsequent layers. More results are available in the Appendix A5. + +# 6. Conclusion + +We proposed Spatiotemporal Skip Guidance (STG), a simple, training-free method for video diffusion models. By simulating an aligned weak model via spatiotemporal skipping, STG offers strong guidance for high-fidelity video generation, achieving notable qualitative and quantitative improvements. We hope this work advances video diffusion models and inspires further research in the field. + +Limitation and Ethical Considerations STG's performance depends on scale and layer selection, with the optimal configuration varying across models, requiring users to set these through heuristic tuning. While video quality improvements are notable, they also raise ethical concerns about misuse, underscoring the importance of using this technology responsibly and constructively. + +Acknowledgements This work was supported by Institute for Information & communications Technology Planning & Evaluation(IITP) grant funded by the Korea government(MSIT) (RS-2019-II190075, Artificial Intelligence Graduate School Program(KAIST)). This work was supported by the National Research Foundation of Korea(NRF) grant funded by the Korea government(MSIT) (No. RS2025-00555621). + +# References + +[1] Donghoon Ahn, Hyoungwon Cho, Jaewon Min, Wooseok Jang, Jungwoo Kim, SeonHwa Kim, Hyun Hee Park, Kyong Hwan Jin, and Seungryong Kim. Self-rectifying diffusion sampling with perturbed-attention guidance. arXiv preprint arXiv:2403.17377, 2024. 2, 3, 5, 7 +[2] Anthropic. The claude 3 model family: Opus, sonnet, haiku. In-, 2024. 7 +[3] Mohammadreza Armandpour, Ali Sadeghian, Huangjie Zheng, Amir Sadeghian, and Mingyuan Zhou. Re-imagine the negative prompt algorithm: Transform 2d diffusion into 3d, alleviate janus problem and beyond. arXiv preprint arXiv:2304.04968, 2023. 7 +[4] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 2, 4, 5, 7, 8, 1, 9, 11, 13, 14, 17, 18, 23 +[5] Andrew Brock. Large scale gan training for high fidelity natural image synthesis. arXiv preprint arXiv:1809.11096, 2018. 2 +[6] Hyungjin Chung, Jeongsol Kim, Geon Yeong Park, Hyelin Nam, and Jong Chul Ye. Cfg++: Manifold-constrained classifier free guidance for diffusion models. arXiv preprint arXiv:2406.08070, 2024. 2, 3 +[7] Zihan Ding, Xiao-Yang Liu, Miao Yin, and Linghe Kong. Tgan: Deep tensor generative adversarial nets for large image generation. arXiv preprint arXiv:1901.09953, 2019. 1 +[8] Songwei Ge, Aniruddha Mahapatra, Gaurav Parmar, JunYan Zhu, and Jia-Bin Huang. On the content bias in fréchet video distance. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7277-7288, 2024. 1 +[9] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139-144, 2020. 2 +[10] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 2, 3, 5 +[11] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2 +[12] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. Advances in Neural Information Processing Systems, 35:8633-8646, 2022. 2 + +[13] Susung Hong. Smoothed energy guidance: Guiding diffusion models with reduced energy curvature of attention. arXiv preprint arXiv:2408.00760, 2024. 2, 3, 5, 8, 18 +[14] Susung Hong, Gyuseong Lee, Wooseok Jang, and Seunglyong Kim. Improving sample quality of diffusion models using self-attention guidance. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7462-7471, 2023. 3, 5 +[15] Ziqi Huang, Yinan He, Jiashuo Yu, Fan Zhang, Chenyang Si, Yuming Jiang, Yuanhan Zhang, Tianxing Wu, Qingyang Jin, Nattapol Chanpaisit, et al. Vbench: Comprehensive benchmark suite for video generative models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21807-21818, 2024. 4, 7 +[16] Yangqing Jia, Evan Shelhamer, Jeff Donahue, Sergey Karayev, Jonathan Long, Ross Girshick, Sergio Guadarrama, and Trevor Darrell. Caffe: Convolutional architecture for fast feature embedding. In Proceedings of the 22nd ACM international conference on Multimedia, pages 675-678, 2014. +[17] Andrej Karpathy, George Toderici, Sanketh Shetty, Thomas Leung, Rahul Sukthankar, and Li Fei-Fei. Large-scale video classification with convolutional neural networks. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 1725–1732, 2014. 1 +[18] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8110-8119, 2020. 2 +[19] Tero Karras, Miika Aittala, Tuomas Kynkänniemi, Jaakko Lehtinen, Timo Aila, and Samuli Laine. Guiding a diffusion model with a bad version of itself. arXiv preprint arXiv:2406.02507, 2024. 2, 3, 5 +[20] Black Forest Labs. Flux: Github repository. https://github.com/black-forest-labs/flux, 2024. Accessed: 2024-11-22. 5, 6 +[21] Shanchuan Lin, Bingchen Liu, Jiashi Li, and Xiao Yang. Common diffusion noise schedules and sample steps are flawed. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 5404-5411, 2024. 2, 7, 5, 10 +[22] Yaofang Liu, Xiaodong Cun, Xuebo Liu, Xintao Wang, Yong Zhang, Haoxin Chen, Yang Liu, Tieyong Zeng, Raymond Chan, and Ying Shan. Evalcrafter: Benchmarking and evaluating large video generation models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22139-22149, 2024. 7, 8, 5, 6 +[23] Nanye Ma, Mark Goldstein, Michael S Albergo, Nicholas M Boffi, Eric Vanden-Eijnden, and Saining Xie. Sit: Exploring flow and diffusion-based generative models with scalable interpolant transformers. arXiv preprint arXiv:2401.08740, 2024. 4 +[24] Xinyin Ma, Gongfan Fang, Michael Bi Mi, and Xinchao Wang. Learning-to-cache: Accelerating diffusion transformer via layer caching. arXiv preprint arXiv:2406.01733, 2024. 4 + +[25] Soumik Mukhopadhyay, Matthew Gwilliam, Vatsal Agarwal, Namitha Padmanabhan, Archana Swaminathan, Srinidhi Hegde, Tianyi Zhou, and Abhinav Shrivastava. Diffusion models beat gans on image classification. arXiv preprint arXiv:2307.08702, 2023. 3, 4 +[26] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4195-4205, 2023. 4 +[27] Adam Polyak, Amit Zohar, Andrew Brown, Andros Tjandra, Animesh Sinha, Ann Lee, Apoorv Vyas, Bowen Shi, Chih-Yao Ma, Ching-Yao Chuang, et al. Movie gen: A cast of media foundation models. arXiv preprint arXiv:2410.13720, 2024. 2, 5 +[28] Hubert Ramsauer, Bernhard Schäfl, Johannes Lehner, Philipp Seidl, Michael Widrich, Thomas Adler, Lukas Gruber, Markus Holzleitner, Milena Pavlovic, Geir Kjetil Sandve, et al. Hopfield networks is all you need. arXiv preprint arXiv:2008.02217, 2020. 4 +[29] Ali Razavi, Aaron Van den Oord, and Oriol Vinyals. Generating diverse high-fidelity images with vq-vae-2. Advances in neural information processing systems, 32, 2019. 2 +[30] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2 +[31] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International conference on machine learning, pages 2256-2265. PMLR, 2015. +[32] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 2, 3 +[33] Genmo Team. Mochi 1. https://github.com/ genmoai/models, 2024. 2, 4, 5, 6, 7, 8, 1, 9, 10, 15, 19, 20 +[34] Du Tran, Lubomir Bourdev, Rob Fergus, Lorenzo Torresani, and Manohar Paluri. Learning spatiotemporal features with 3d convolutional networks. In Proceedings of the IEEE international conference on computer vision, pages 4489-4497, 2015. 1 +[35] Pascal Vincent. A connection between score matching and denoising autoencoders. Neural computation, 23(7):1661-1674, 2011. 3 +[36] Yilun Xu, Mingyang Deng, Xiang Cheng, Yonglong Tian, Ziming Liu, and Tommi Jaakkola. Restart sampling for improving generative processes. Advances in Neural Information Processing Systems, 36:76806-76838, 2023. 2, 7, 5 +[37] Sihyun Yu, Jihoon Tack, Sangwoo Mo, Hyunsu Kim, Junho Kim, Jung-Woo Ha, and Jinwoo Shin. Generating videos with dynamics-aware implicit generative adversarial networks. arXiv preprint arXiv:2202.10571, 2022. 7 +[38] Zangwei Zheng, Xiangyu Peng, Tianji Yang, Chenhui Shen, Shenggui Li, Hongxin Liu, Yukun Zhou, Tianyi Li, and Yang You. Open-sora: Democratizing efficient video production + +for all. https://github.com/hpcaitech/Open-Sora, 2024. 2, 4, 5, 7, 8, 1, 12, 16, 21, 22 \ No newline at end of file diff --git a/2025/Spatiotemporal Skip Guidance for Enhanced Video Diffusion Sampling/images.zip b/2025/Spatiotemporal Skip Guidance for Enhanced Video Diffusion Sampling/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..bedc0ea9b8eceb485bba37e05c0684cb044d2850 --- /dev/null +++ b/2025/Spatiotemporal Skip Guidance for Enhanced Video Diffusion Sampling/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4129cfa42a5a3c54ee082d47a0f833f50eb77466a17d7bcc8e8d1ce009778e4d +size 617357 diff --git a/2025/Spatiotemporal Skip Guidance for Enhanced Video Diffusion Sampling/layout.json b/2025/Spatiotemporal Skip Guidance for Enhanced Video Diffusion Sampling/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..7990b8ea6f6eca75ae69f446746a716dc1257aeb --- /dev/null +++ b/2025/Spatiotemporal Skip Guidance for Enhanced Video Diffusion Sampling/layout.json @@ -0,0 +1,8306 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 85, + 103, + 526, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 103, + 526, + 121 + ], + "spans": [ + { + "bbox": [ + 85, + 103, + 526, + 121 + ], + "type": "text", + "content": "Spatiotemporal Skip Guidance for Enhanced Video Diffusion Sampling" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 104, + 142, + 504, + 159 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 142, + 504, + 159 + ], + "spans": [ + { + "bbox": [ + 104, + 142, + 504, + 159 + ], + "type": "text", + "content": "Junha Hyung" + }, + { + "bbox": [ + 104, + 142, + 504, + 159 + ], + "type": "inline_equation", + "content": "^{*1}" + }, + { + "bbox": [ + 104, + 142, + 504, + 159 + ], + "type": "text", + "content": " Kinam Kim" + }, + { + "bbox": [ + 104, + 142, + 504, + 159 + ], + "type": "inline_equation", + "content": "^{*1}" + }, + { + "bbox": [ + 104, + 142, + 504, + 159 + ], + "type": "text", + "content": " Susung Hong" + }, + { + "bbox": [ + 104, + 142, + 504, + 159 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 104, + 142, + 504, + 159 + ], + "type": "text", + "content": " Min-Jung Kim" + }, + { + "bbox": [ + 104, + 142, + 504, + 159 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 104, + 142, + 504, + 159 + ], + "type": "text", + "content": " Jaegul Choo" + }, + { + "bbox": [ + 104, + 142, + 504, + 159 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 205, + 170, + 405, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 205, + 170, + 405, + 186 + ], + "spans": [ + { + "bbox": [ + 205, + 170, + 405, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 205, + 170, + 405, + 186 + ], + "type": "text", + "content": "KAIST AI " + }, + { + "bbox": [ + 205, + 170, + 405, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 205, + 170, + 405, + 186 + ], + "type": "text", + "content": "University of Washington" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 96, + 187, + 512, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 187, + 512, + 200 + ], + "spans": [ + { + "bbox": [ + 96, + 187, + 512, + 200 + ], + "type": "text", + "content": "{sharpeeee, kinamplify, emjay73, jchoo}@kaist.ac.kr, susung@cs.washington.edu" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 69, + 526, + 294, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 526, + 294, + 550 + ], + "spans": [ + { + "bbox": [ + 69, + 526, + 294, + 550 + ], + "type": "text", + "content": "\"A close-up shot of a butterfly landing on the nose of a woman, highlighting her smile and the details of the butterfly's wings.\"" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 321, + 526, + 539, + 549 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 526, + 539, + 549 + ], + "spans": [ + { + "bbox": [ + 321, + 526, + 539, + 549 + ], + "type": "text", + "content": "\"A close-up of a woman's face with colored powder exploding around her, creating an abstract splash of vibrant hues.\"" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 559, + 555, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 559, + 555, + 581 + ], + "spans": [ + { + "bbox": [ + 54, + 559, + 555, + 581 + ], + "type": "text", + "content": "Figure 1. Visual comparison of video quality between CFG (top row) and our STG method (bottom row). Best viewed in Acrobat Reader; click on the images to watch the videos." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 171, + 597, + 221, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 171, + 597, + 221, + 611 + ], + "spans": [ + { + "bbox": [ + 171, + 597, + 221, + 611 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 624, + 297, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 624, + 297, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 624, + 297, + 696 + ], + "type": "text", + "content": "Diffusion models have emerged as a powerful tool for generating high-quality images, videos, and 3D content. While sampling guidance techniques like CFG improve quality, they reduce diversity and motion. Autoguidance mitigates these issues but demands extra weak model training, limiting its practicality for large-scale models. In this" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 599, + 556, + 708 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 599, + 556, + 708 + ], + "spans": [ + { + "bbox": [ + 313, + 599, + 556, + 708 + ], + "type": "text", + "content": "work, we introduce Spatiotemporal Skip Guidance (STG), a simple training-free sampling guidance method for enhancing transformer-based video diffusion models. STG employs an implicit weak model via self-perturbation, avoiding the need for external models or additional training. By selectively skipping spatiotemporal layers, STG produces an aligned, degraded version of the original model to boost sample quality without compromising diversity or dynamic degree. Our contributions include: (1) introducing STG as" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 70, + 704, + 170, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 704, + 170, + 714 + ], + "spans": [ + { + "bbox": [ + 70, + 704, + 170, + 714 + ], + "type": "text", + "content": "* indicates equal contribution." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "11006" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "content": "an efficient, high-performing guidance technique for video diffusion models, (2) eliminating the need for auxiliary models by simulating a weak model through layer skipping, and (3) ensuring quality-enhanced guidance without compromising sample diversity or dynamics unlike CFG. For additional results, visit https://junhahyung.github.io/STGuidance/." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 178, + 136, + 192 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 178, + 136, + 192 + ], + "spans": [ + { + "bbox": [ + 56, + 178, + 136, + 192 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 198, + 295, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 198, + 295, + 319 + ], + "spans": [ + { + "bbox": [ + 55, + 198, + 295, + 319 + ], + "type": "text", + "content": "Diffusion models [11, 30-32] are a successful class of generative models known for their flexibility in modeling complex data distributions, achieving impressive results in image, video, and 3D generation. By progressively denoising random noise, they enable robust generalization, making them a leading choice for realistic content generation and often surpassing GAN-based methods [5, 9, 18, 29]. Building on this success, video diffusion models [4, 12, 27, 33, 38] generate high-quality videos by using temporal or 3D attention layers to handle sequential frames." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 319, + 295, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 319, + 295, + 426 + ], + "spans": [ + { + "bbox": [ + 55, + 319, + 295, + 426 + ], + "type": "text", + "content": "Meanwhile, to enhance sample quality, sampling guidance techniques such as Classifier-Free Guidance (CFG) [10] and Autoguidance [19] have been introduced to guide the denoising process. These techniques employ weak models to predict poor trajectories, steering the main model away from them and pushing samples toward high-quality regions on the data manifold. However, CFG often reduces diversity, leading to saturated or overly simplified results [6, 19]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 426, + 295, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 426, + 295, + 545 + ], + "spans": [ + { + "bbox": [ + 55, + 426, + 295, + 545 + ], + "type": "text", + "content": "Autoguidance [19] addresses this issue by using a weak model trained on the same task, conditioning, and data distribution as the main model. The main drawback of this approach, however, is the need to train an additional weak model, which is impractical for large-scale models. Alternative methods such as PAG [1] and SEG [13] use self-perturbation to implicitly mimic a weak model, avoiding the need for extra training. Yet, these methods are designed specifically for image generation diffusion models, applying self-perturbation to 2D spatial attention maps." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 545, + 295, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 295, + 653 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 295, + 653 + ], + "type": "text", + "content": "In this work, we propose Spatiotemporal Skip Guidance (STG), a simple and effective sampling guidance method for video diffusion models that significantly enhances the performance of any transformer-based video diffusion model without additional training. Specifically, we use an implicit weak model for guidance through self-perturbation, eliminating the need for explicit weak models and their associated training costs. This is especially crucial for video diffusion models, where training costs are high." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 654, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 295, + 713 + ], + "type": "text", + "content": "Our implicit weak model, a key component of our framework, is deliberately designed as a degraded but \"aligned\" version of the original video generation model. As demonstrated in Autoguidance [19], having the weak model share the same task, conditioning, and data distribution as the" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 321, + 71, + 444, + 176 + ], + "blocks": [ + { + "bbox": [ + 321, + 71, + 444, + 176 + ], + "lines": [ + { + "bbox": [ + 321, + 71, + 444, + 176 + ], + "spans": [ + { + "bbox": [ + 321, + 71, + 444, + 176 + ], + "type": "image", + "image_path": "a9ff040f709f6d6ceae9ae479c4eebc42adde4b366b565ff287709d970e93b04.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 378, + 179, + 398, + 188 + ], + "lines": [ + { + "bbox": [ + 378, + 179, + 398, + 188 + ], + "spans": [ + { + "bbox": [ + 378, + 179, + 398, + 188 + ], + "type": "text", + "content": "CFG" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 463, + 71, + 551, + 176 + ], + "blocks": [ + { + "bbox": [ + 463, + 71, + 551, + 176 + ], + "lines": [ + { + "bbox": [ + 463, + 71, + 551, + 176 + ], + "spans": [ + { + "bbox": [ + 463, + 71, + 551, + 176 + ], + "type": "image", + "image_path": "4437acdf9142fc0f4a69a3b9477ef08b528ab6cae40df116cb29b121c290f3cb.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 500, + 179, + 520, + 188 + ], + "lines": [ + { + "bbox": [ + 500, + 179, + 520, + 188 + ], + "spans": [ + { + "bbox": [ + 500, + 179, + 520, + 188 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 198, + 554, + 265 + ], + "lines": [ + { + "bbox": [ + 313, + 198, + 554, + 265 + ], + "spans": [ + { + "bbox": [ + 313, + 198, + 554, + 265 + ], + "type": "text", + "content": "Figure 2. Comparison between CFG and STG, with the band conceptually representing the noisy data manifold. In STG, the weak model and the main model are aligned along the direction of increasing quality. In contrast, the two models in CFG differ not only in quality but also in aspects such as diversity and prompt alignment capabilities." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 288, + 554, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 288, + 554, + 419 + ], + "spans": [ + { + "bbox": [ + 313, + 288, + 554, + 419 + ], + "type": "text", + "content": "main model is essential for quality improvement, as we expect both models to exhibit similar, aligned errors in the same parts of samples. We conceptually illustrate this alignment in Fig. 2, where our weak model and main model are arranged in a direction of increasing quality. To achieve this, we apply spatiotemporal perturbation to both spatial and temporal attention layers—or, in the case of 3D attention, to the entire layer—by selectively skipping certain layers. This straightforward approach effectively nullifies specific residual or attention layers, generating a lower-quality version of the model that simulates an aligned weak model." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 429, + 556, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 429, + 556, + 561 + ], + "spans": [ + { + "bbox": [ + 313, + 429, + 556, + 561 + ], + "type": "text", + "content": "To ensure samples remain on the data manifold, even when using large guidance scales, we employ optional techniques such as the rescaling [21] and restart [36] methods. Rescaling the latent code constrains its variance, addressing the issue of larger variance causing saturation in the results [21]. Meanwhile, the restart sampling technique leverages the error contraction property of forward SDEs [36] to keep the sampling trajectory on the manifold. These techniques help prevent overshooting in the sampling trajectory, which could otherwise push samples off the manifold and result in saturated or distorted outputs." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 325, + 570, + 476, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 570, + 476, + 582 + ], + "spans": [ + { + "bbox": [ + 325, + 570, + 476, + 582 + ], + "type": "text", + "content": "Our key contributions are as follows:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 591, + 553, + 711 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 313, + 591, + 553, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 591, + 553, + 626 + ], + "spans": [ + { + "bbox": [ + 313, + 591, + 553, + 626 + ], + "type": "text", + "content": "- We propose STG—a surprisingly simple sampling guidance framework that significantly boosts the performance of video diffusion models." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 628, + 553, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 628, + 553, + 673 + ], + "spans": [ + { + "bbox": [ + 313, + 628, + 553, + 673 + ], + "type": "text", + "content": "- Our method introduces an implicit weak model by skipping spatiotemporal layers in video diffusion models, eliminating the need for additional training or external models." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 675, + 553, + 711 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 675, + 553, + 711 + ], + "spans": [ + { + "bbox": [ + 313, + 675, + 553, + 711 + ], + "type": "text", + "content": "- Our method enhances sample quality during guidance without reducing diversity or limiting the dynamics of generated videos." + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11007" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 143, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 143, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 143, + 83 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 187 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 187 + ], + "type": "text", + "content": "Guidance with trained weak model Classifier-Free Guidance (CFG) [10] improves conditional generation in Diffusion Models by using an implicit unconditional model as a weak model. However, differences in tasks between the unconditional and conditional models can reduce sample diversity [6, 19] and increase sampling trajectory curvature [6], leading to overshooting the data manifold and producing skewed or oversaturated images." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 188, + 296, + 271 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 188, + 296, + 271 + ], + "spans": [ + { + "bbox": [ + 55, + 188, + 296, + 271 + ], + "type": "text", + "content": "Autoguidance [19] mitigates these issues by employing a bad version of the main model as a weak model, trained with reduced capacity and compute to ensure alignment. This alignment allows the guidance algorithm to correct errors by analyzing prediction differences. While effective, it requires additional training, which is challenging for large-scale video diffusion models." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 289, + 296, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 289, + 296, + 469 + ], + "spans": [ + { + "bbox": [ + 55, + 289, + 296, + 469 + ], + "type": "text", + "content": "Guidance with training-free weak model Another line of work avoids additional training by using self-perturbation of the main model to mimic a weak model. Self-Attention Guidance (SAG) [14] blurs high-attention regions, Perturbed Attention Guidance (PAG) [1] replaces attention maps with identity matrices, and Smoothed Energy Guidance (SEG) [13] applies Gaussian blur to the attention weights to smooth the energy landscape. These methods guide sampling toward high-quality outputs by leveraging differences in predictions from the weakened model. While effective, they are primarily designed for image diffusion models with 2D self-attention. We aim to extend this approach to video diffusion models, which require handling temporal dynamics with additional temporal or 3D spatiotemporal attention layers." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 481, + 141, + 493 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 481, + 141, + 493 + ], + "spans": [ + { + "bbox": [ + 55, + 481, + 141, + 493 + ], + "type": "text", + "content": "3. Preliminaries" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 501, + 157, + 513 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 501, + 157, + 513 + ], + "spans": [ + { + "bbox": [ + 55, + 501, + 157, + 513 + ], + "type": "text", + "content": "3.1. Diffusion models" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 520, + 296, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 520, + 296, + 568 + ], + "spans": [ + { + "bbox": [ + 55, + 520, + 296, + 568 + ], + "type": "text", + "content": "Diffusion models generate samples by progressively removing noise from noisy data, restoring the original data distribution. Song et al. [32] defined the process of adding noise to the data using a stochastic differential equation (SDE):" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 111, + 577, + 294, + 601 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 577, + 294, + 601 + ], + "spans": [ + { + "bbox": [ + 111, + 577, + 294, + 601 + ], + "type": "interline_equation", + "content": "d x = - \\frac {\\beta (t)}{2} x d t + \\sqrt {\\beta (t)} d w, \\tag {1}", + "image_path": "f63feb7b1ac0937660c0d1c7f9e91f607478c813f262a0e1957aa7fb942b8070.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 610, + 296, + 646 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 610, + 296, + 646 + ], + "spans": [ + { + "bbox": [ + 55, + 610, + 296, + 646 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 610, + 296, + 646 + ], + "type": "inline_equation", + "content": "\\beta (t)" + }, + { + "bbox": [ + 55, + 610, + 296, + 646 + ], + "type": "text", + "content": " is a time-dependent noise schedule, and " + }, + { + "bbox": [ + 55, + 610, + 296, + 646 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 55, + 610, + 296, + 646 + ], + "type": "text", + "content": " represents the standard Wiener process. Corresponding reverse-time SDE is:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 61, + 654, + 296, + 681 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 654, + 296, + 681 + ], + "spans": [ + { + "bbox": [ + 61, + 654, + 296, + 681 + ], + "type": "interline_equation", + "content": "d x = \\left[ - \\frac {\\beta (t)}{2} x - \\beta (t) \\nabla_ {x} \\log p _ {t} (x) \\right] d t + \\sqrt {\\beta (t)} d \\bar {w} \\tag {2}", + "image_path": "7b09cafbf36bed1f7017a4140052a69a441de6d617d8803b65b0203937b136c1.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "inline_equation", + "content": "d\\bar{w}" + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": " is the Wiener process running backward in time. Here, the score function " + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\nabla_x\\log p_t(x)" + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": " is approximated by" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "type": "text", + "content": "a neural network " + }, + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "type": "inline_equation", + "content": "s_{\\theta}(x(t))" + }, + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "type": "text", + "content": " trained using denoising score matching [35]:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 323, + 106, + 555, + 142 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 106, + 555, + 142 + ], + "spans": [ + { + "bbox": [ + 323, + 106, + 555, + 142 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\theta^ {*} = \\arg \\min _ {\\theta} \\mathbb {E} _ {t} \\left\\{\\lambda (t) \\mathbb {E} _ {x _ {0}} \\mathbb {E} _ {x _ {t} | x _ {0}} \\right[ \\\\ \\left. \\left\\| s _ {\\theta} \\left(x _ {t}\\right) - \\nabla_ {x _ {t}} \\log p _ {t} \\left(x _ {t} \\mid x _ {0}\\right) \\| _ {2} ^ {2} \\right] \\right\\}. \\tag {3} \\\\ \\end{array}", + "image_path": "1f037c2076f37733e903fa338dbd29ad1c52dc498abccde8ed028e777421ca2a.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 148, + 429, + 159 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 148, + 429, + 159 + ], + "spans": [ + { + "bbox": [ + 314, + 148, + 429, + 159 + ], + "type": "text", + "content": "3.2. Classifier Guidance" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 165, + 554, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 165, + 554, + 202 + ], + "spans": [ + { + "bbox": [ + 313, + 165, + 554, + 202 + ], + "type": "text", + "content": "Classifier Guidance (CG) [25] reformulates the reverse process of a diffusion model by incorporating an external classifier " + }, + { + "bbox": [ + 313, + 165, + 554, + 202 + ], + "type": "inline_equation", + "content": "p_{\\phi}" + }, + { + "bbox": [ + 313, + 165, + 554, + 202 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 373, + 209, + 555, + 224 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 373, + 209, + 555, + 224 + ], + "spans": [ + { + "bbox": [ + 373, + 209, + 555, + 224 + ], + "type": "interline_equation", + "content": "\\tilde {p _ {\\theta}} \\left(x _ {t} | y\\right) \\propto p _ {\\theta} \\left(x _ {t}\\right) p _ {\\phi} \\left(y \\mid x _ {t}\\right) ^ {\\lambda}, \\tag {4}", + "image_path": "39dc4de084081a770ba437e863616226c0eb8167a4c8a437a401493bd20972e4.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 231, + 554, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 231, + 554, + 255 + ], + "spans": [ + { + "bbox": [ + 313, + 231, + 554, + 255 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 231, + 554, + 255 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 313, + 231, + 554, + 255 + ], + "type": "text", + "content": " is the desired class label and " + }, + { + "bbox": [ + 313, + 231, + 554, + 255 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 231, + 554, + 255 + ], + "type": "text", + "content": " controls the guidance strength. The score function is then derived as:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 323, + 266, + 554, + 294 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 266, + 554, + 294 + ], + "spans": [ + { + "bbox": [ + 323, + 266, + 554, + 294 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\nabla_ {x _ {t}} \\log \\tilde {p} _ {\\theta} (x _ {t} | y) = \\nabla_ {x _ {t}} \\log p _ {\\theta} (x _ {t}) \\\\ + \\lambda \\nabla_ {x _ {t}} \\log p _ {\\phi} (y | x _ {t}). \\tag {5} \\\\ \\end{array}", + "image_path": "dd47670e1d15be5ab6c73ae1a115ac6274e0ec36ba7139e7007b3954ae953980.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 301, + 554, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 301, + 554, + 326 + ], + "spans": [ + { + "bbox": [ + 313, + 301, + 554, + 326 + ], + "type": "text", + "content": "By substituting the score function in Eq. 2 with Eq. 5, sampling from the desired class condition " + }, + { + "bbox": [ + 313, + 301, + 554, + 326 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 313, + 301, + 554, + 326 + ], + "type": "text", + "content": " becomes possible." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 314, + 332, + 454, + 344 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 332, + 454, + 344 + ], + "spans": [ + { + "bbox": [ + 314, + 332, + 454, + 344 + ], + "type": "text", + "content": "3.3. Classifier-Free Guidance" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 350, + 555, + 385 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 350, + 555, + 385 + ], + "spans": [ + { + "bbox": [ + 313, + 350, + 555, + 385 + ], + "type": "text", + "content": "Classifier-Free Guidance (CFG) [10] uses Bayes' rule to replace a classifier-guided score with a linear combination of conditional and unconditional score estimates:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 347, + 393, + 554, + 407 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 393, + 554, + 407 + ], + "spans": [ + { + "bbox": [ + 347, + 393, + 554, + 407 + ], + "type": "interline_equation", + "content": "\\tilde {\\epsilon} _ {\\theta} ^ {\\lambda} \\left(x _ {t}\\right) = \\epsilon_ {\\theta} \\left(x _ {t}\\right) + \\lambda \\left(\\epsilon_ {\\theta} \\left(x _ {t}\\right) - \\epsilon_ {\\theta} \\left(x _ {t} \\mid \\phi\\right)\\right). \\tag {6}", + "image_path": "340ab4fecf67724d8fbf502d96da0e4829beb8244fc9e2e9af2034e86e4d28e0.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 415, + 554, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 415, + 554, + 475 + ], + "spans": [ + { + "bbox": [ + 313, + 415, + 554, + 475 + ], + "type": "text", + "content": "CFG jointly trains the unconditional model " + }, + { + "bbox": [ + 313, + 415, + 554, + 475 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(x_t|\\phi)" + }, + { + "bbox": [ + 313, + 415, + 554, + 475 + ], + "type": "text", + "content": " and the conditional model " + }, + { + "bbox": [ + 313, + 415, + 554, + 475 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(x_t|c) (= \\epsilon_{\\theta}(x_t))" + }, + { + "bbox": [ + 313, + 415, + 554, + 475 + ], + "type": "text", + "content": " within a single model by setting the condition " + }, + { + "bbox": [ + 313, + 415, + 554, + 475 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 415, + 554, + 475 + ], + "type": "text", + "content": " to a null token " + }, + { + "bbox": [ + 313, + 415, + 554, + 475 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 313, + 415, + 554, + 475 + ], + "type": "text", + "content": ". Using this guided denoising process, the model better captures the conditions and often generates high-fidelity outputs." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 314, + 485, + 370, + 497 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 485, + 370, + 497 + ], + "spans": [ + { + "bbox": [ + 314, + 485, + 370, + 497 + ], + "type": "text", + "content": "4. Method" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 314, + 504, + 473, + 517 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 504, + 473, + 517 + ], + "spans": [ + { + "bbox": [ + 314, + 504, + 473, + 517 + ], + "type": "text", + "content": "4.1. Optimal Weak Model Design" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 521, + 554, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 554, + 617 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 554, + 617 + ], + "type": "text", + "content": "Our goal is to construct an aligned weak model that captures a distribution similar to the original model while generating slightly lower-quality samples. This ensures that the guidance gradient points toward improved quality, as shown in Fig. 2. Misaligned models, as seen in CFG [10], may lead to unintended outcomes like reduced diversity, while Autoguidance [19] requires additional weak model training, making it impractical for large-scale models." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 617, + 554, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 554, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 554, + 689 + ], + "type": "text", + "content": "To address these limitations, we design an implicit weak model by leveraging the original model itself, eliminating the need for external training. This training-free approach ensures computational efficiency and better alignment, as the weak model is derived directly from the original, preserving most network weights." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "content": "To achieve this, we apply perturbation methods directly to the main model's forward pass, creating an implicit weak" + } + ] + } + ], + "index": 27 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11008" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 69, + 70, + 541, + 142 + ], + "blocks": [ + { + "bbox": [ + 69, + 70, + 541, + 142 + ], + "lines": [ + { + "bbox": [ + 69, + 70, + 541, + 142 + ], + "spans": [ + { + "bbox": [ + 69, + 70, + 541, + 142 + ], + "type": "table", + "html": "
ModelsImaging QualityAesthetic QualityMotion SmoothnessDynamic DegreeTemporal Flickering
Mochi (CFG)0.5240.5070.9850.870.976
Mochi (STG)0.6280.5540.9880.860.978
Open-Sora (CFG)0.5610.4930.9820.9020.975
Open-Sora (STG)0.6060.5090.9870.8950.976
", + "image_path": "c53eada5524d1d2d8f2a9402b231be54c8289affd8ac4b1d675727d8ce3f5235.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 106, + 173, + 504, + 216 + ], + "blocks": [ + { + "bbox": [ + 124, + 151, + 485, + 163 + ], + "lines": [ + { + "bbox": [ + 124, + 151, + 485, + 163 + ], + "spans": [ + { + "bbox": [ + 124, + 151, + 485, + 163 + ], + "type": "text", + "content": "Table 1. Quantitative results for Mochi [33] and Open-Sora [38] on VBench [15] T2V benchmarks." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 106, + 173, + 504, + 216 + ], + "lines": [ + { + "bbox": [ + 106, + 173, + 504, + 216 + ], + "spans": [ + { + "bbox": [ + 106, + 173, + 504, + 216 + ], + "type": "table", + "html": "
ModelsFVD (↓)ISImaging QualityAesthetic QualityMotion SmoothnessDynamic Degree
SVD (CFG)151.338.00.6870.6370.9660.562
SVD (STG)128.738.50.6940.6390.9680.694
", + "image_path": "e86e0514a44704a104cc86a9e026a51b07779d48f640e801fca00a85ba7ebee4.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 142, + 223, + 466, + 235 + ], + "lines": [ + { + "bbox": [ + 142, + 223, + 466, + 235 + ], + "spans": [ + { + "bbox": [ + 142, + 223, + 466, + 235 + ], + "type": "text", + "content": "Table 2. Quantitative results for SVD [4] on FVD, IS, and VBenchmark [15] I2V benchmarks." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 256, + 296, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 256, + 296, + 293 + ], + "spans": [ + { + "bbox": [ + 55, + 256, + 296, + 293 + ], + "type": "text", + "content": "model. Specifically, we use spatiotemporal perturbation, a key component for aligning the weak model with video diffusion models, which will be elaborated upon in detail." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 301, + 254, + 315 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 301, + 254, + 315 + ], + "spans": [ + { + "bbox": [ + 55, + 301, + 254, + 315 + ], + "type": "text", + "content": "4.2. Sampling from High Quality Samples" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 319, + 296, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 319, + 296, + 357 + ], + "spans": [ + { + "bbox": [ + 55, + 319, + 296, + 357 + ], + "type": "text", + "content": "Similar to CG [25], we define our goal as conditioning the model on an imaginary label " + }, + { + "bbox": [ + 55, + 319, + 296, + 357 + ], + "type": "inline_equation", + "content": "y_{g}" + }, + { + "bbox": [ + 55, + 319, + 296, + 357 + ], + "type": "text", + "content": " that represents high-quality samples, leading to the sampling distribution" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 110, + 365, + 295, + 380 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 365, + 295, + 380 + ], + "spans": [ + { + "bbox": [ + 110, + 365, + 295, + 380 + ], + "type": "interline_equation", + "content": "\\tilde {p} _ {\\theta} \\left(x _ {t} \\mid y _ {g}\\right) \\propto p _ {\\theta} \\left(x _ {t}\\right) p _ {\\phi} \\left(y _ {g} \\mid x _ {t}\\right) ^ {w}. \\tag {7}", + "image_path": "3ae48446460b8b2cb65a1db3bf78f23b0506a981000d607e9651b8d37cb1abd1.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 390, + 296, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 390, + 296, + 425 + ], + "spans": [ + { + "bbox": [ + 55, + 390, + 296, + 425 + ], + "type": "text", + "content": "Using " + }, + { + "bbox": [ + 55, + 390, + 296, + 425 + ], + "type": "inline_equation", + "content": "w > 0" + }, + { + "bbox": [ + 55, + 390, + 296, + 425 + ], + "type": "text", + "content": " sharpens the distribution, promoting the generation of high-quality samples. From Eq. 7, the score is derived as" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 65, + 439, + 294, + 468 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 439, + 294, + 468 + ], + "spans": [ + { + "bbox": [ + 65, + 439, + 294, + 468 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\nabla_ {x _ {t}} \\log \\tilde {p} _ {\\theta} (x _ {t} | y _ {g}) \\\\ = \\nabla_ {x _ {t}} \\log p _ {\\theta} (x _ {t}) + w \\nabla_ {x _ {t}} \\log p _ {\\phi} (y _ {g} | x _ {t}). \\tag {8} \\\\ \\end{array}", + "image_path": "e615508d1471fcd3f1680001816ff444dc0fc78cd8683bd34d0d8def2c282766.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 478, + 296, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 478, + 296, + 526 + ], + "spans": [ + { + "bbox": [ + 55, + 478, + 296, + 526 + ], + "type": "text", + "content": "Rather than using an external classifier " + }, + { + "bbox": [ + 55, + 478, + 296, + 526 + ], + "type": "inline_equation", + "content": "p_{\\phi}" + }, + { + "bbox": [ + 55, + 478, + 296, + 526 + ], + "type": "text", + "content": ", we propose using our model " + }, + { + "bbox": [ + 55, + 478, + 296, + 526 + ], + "type": "inline_equation", + "content": "p_{\\theta}" + }, + { + "bbox": [ + 55, + 478, + 296, + 526 + ], + "type": "text", + "content": " as an implicit classifier. We design this classifier to be inversely proportional to the probability of an imaginary \"bad\" label " + }, + { + "bbox": [ + 55, + 478, + 296, + 526 + ], + "type": "inline_equation", + "content": "y_{b}" + }, + { + "bbox": [ + 55, + 478, + 296, + 526 + ], + "type": "text", + "content": ", expressed as" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 125, + 536, + 295, + 562 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 125, + 536, + 295, + 562 + ], + "spans": [ + { + "bbox": [ + 125, + 536, + 295, + 562 + ], + "type": "interline_equation", + "content": "p _ {\\phi} \\left(y _ {g} \\mid x _ {t}\\right) \\propto \\frac {1}{p _ {\\theta} \\left(y _ {b} \\mid x _ {t}\\right)}. \\tag {9}", + "image_path": "e6097b1f6b1fcb5a367bb6684800adafb940fa947a97d272978429a4df66b05d.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 571, + 296, + 595 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 571, + 296, + 595 + ], + "spans": [ + { + "bbox": [ + 55, + 571, + 296, + 595 + ], + "type": "text", + "content": "Using Bayes' rule, the function can be expressed in terms of the marginal posterior as follows:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 80, + 605, + 295, + 647 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 605, + 295, + 647 + ], + "spans": [ + { + "bbox": [ + 80, + 605, + 295, + 647 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\nabla_ {x _ {t}} \\log \\frac {1}{p _ {\\theta} (y _ {b} | x _ {t})} = \\nabla_ {x _ {t}} \\log \\frac {p _ {\\theta} (x _ {t})}{p _ {\\theta} (y _ {b}) p _ {\\theta} (x _ {t} | y _ {b})} \\\\ = \\nabla_ {x _ {t}} \\left(\\log p _ {\\theta} \\left(x _ {t}\\right) - \\log p _ {\\theta} \\left(x _ {t} \\mid y _ {b}\\right)\\right), \\tag {10} \\\\ \\end{array}", + "image_path": "0bccc17130b7d275495ba2118cc707e2fafc237d6bbdc9e97a79f534819c682b.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 657, + 249, + 669 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 657, + 249, + 669 + ], + "spans": [ + { + "bbox": [ + 55, + 657, + 249, + 669 + ], + "type": "text", + "content": "leading to the score of the target distribution as:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 65, + 682, + 295, + 711 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 682, + 295, + 711 + ], + "spans": [ + { + "bbox": [ + 65, + 682, + 295, + 711 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\nabla_ {x _ {t}} \\log \\tilde {p} _ {\\theta} (x _ {t} | y _ {g}) = \\nabla_ {x _ {t}} \\log p _ {\\theta} (x _ {t}) \\\\ + w \\nabla_ {x _ {t}} \\left(\\log p _ {\\theta} \\left(x _ {t}\\right) - \\log p _ {\\theta} \\left(x _ {t} \\mid y _ {b}\\right)\\right). \\tag {11} \\\\ \\end{array}", + "image_path": "19f6c30623f77b1dc13fd4c1d60bce44c724f47fb2943f8f1720e3774ef833ef.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 256, + 553, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 256, + 553, + 281 + ], + "spans": [ + { + "bbox": [ + 313, + 256, + 553, + 281 + ], + "type": "text", + "content": "We can sample from " + }, + { + "bbox": [ + 313, + 256, + 553, + 281 + ], + "type": "inline_equation", + "content": "\\tilde{p}_{\\theta}(x_t|y_g)" + }, + { + "bbox": [ + 313, + 256, + 553, + 281 + ], + "type": "text", + "content": " by substituting the score function in Eq. 2 with Eq. 11, resulting in:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 289, + 553, + 345 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 289, + 553, + 345 + ], + "spans": [ + { + "bbox": [ + 317, + 289, + 553, + 345 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} d x = \\left[ - \\frac {\\beta (t)}{2} x - \\beta (t) \\left(\\nabla_ {x _ {t}} \\log p _ {t} \\left(x _ {t}\\right) \\right. \\right. \\tag {12} \\\\ \\left. + w \\nabla_ {x _ {t}} (\\log p _ {\\theta} (x _ {t}) - \\log p _ {\\theta} (x _ {t} | y _ {b})) \\right] d t + \\sqrt {\\beta (t)} d \\bar {w}, \\\\ \\end{array}", + "image_path": "297cf6988088f6218097df9060908159a34af1aa9e6cce2e5ca69514ff844e4f.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 352, + 554, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 352, + 554, + 389 + ], + "spans": [ + { + "bbox": [ + 313, + 352, + 554, + 389 + ], + "type": "text", + "content": "and solving the reverse SDE. Since the score function is approximated using the neural network, we can generate samples using" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 352, + 398, + 553, + 413 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 352, + 398, + 553, + 413 + ], + "spans": [ + { + "bbox": [ + 352, + 398, + 553, + 413 + ], + "type": "interline_equation", + "content": "\\tilde {\\epsilon} _ {\\theta} ^ {w} (x _ {t}) = \\epsilon_ {\\theta} (x _ {t}) + w \\left(\\epsilon_ {\\theta} (x _ {t}) - \\epsilon_ {\\theta} ^ {b} (x _ {t})\\right). \\tag {13}", + "image_path": "df9f1a90bc05450ca32fe9cb2d57c7e1579249defec01476a4f5aafb3a123bdb.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 422, + 554, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 422, + 554, + 519 + ], + "spans": [ + { + "bbox": [ + 313, + 422, + 554, + 519 + ], + "type": "text", + "content": "An interesting approach here is to model " + }, + { + "bbox": [ + 313, + 422, + 554, + 519 + ], + "type": "inline_equation", + "content": "\\nabla_{x_t}\\log p_\\theta (x_t|y_b)" + }, + { + "bbox": [ + 313, + 422, + 554, + 519 + ], + "type": "text", + "content": " by perturbing the forward pass of " + }, + { + "bbox": [ + 313, + 422, + 554, + 519 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(x_{t})" + }, + { + "bbox": [ + 313, + 422, + 554, + 519 + ], + "type": "text", + "content": ", denoted as " + }, + { + "bbox": [ + 313, + 422, + 554, + 519 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}^{b}(x_{t})" + }, + { + "bbox": [ + 313, + 422, + 554, + 519 + ], + "type": "text", + "content": ". Now the main focus is designing a perturbation that effectively yields the weak model capable of estimating " + }, + { + "bbox": [ + 313, + 422, + 554, + 519 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}^{b}(x_{t})" + }, + { + "bbox": [ + 313, + 422, + 554, + 519 + ], + "type": "text", + "content": ", aligning closely with " + }, + { + "bbox": [ + 313, + 422, + 554, + 519 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(x_{t})" + }, + { + "bbox": [ + 313, + 422, + 554, + 519 + ], + "type": "text", + "content": " as discussed in Sec. 4.1. Ideally, we want a distribution that deviates minimally from " + }, + { + "bbox": [ + 313, + 422, + 554, + 519 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(x_{t})" + }, + { + "bbox": [ + 313, + 422, + 554, + 519 + ], + "type": "text", + "content": " while producing slightly lower-quality samples." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 527, + 514, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 527, + 514, + 540 + ], + "spans": [ + { + "bbox": [ + 313, + 527, + 514, + 540 + ], + "type": "text", + "content": "4.3. Spatiotemporal Skip Guidance (STG)" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 545, + 554, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 554, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 554, + 628 + ], + "type": "text", + "content": "We introduce Spatiotemporal Skip Guidance (STG), a simple guidance method designed for video diffusion models that generates diverse, high-fidelity samples while addressing the limitations of existing methods. STG implicitly simulates an aligned weak model through spatiotemporal perturbation, capturing the spatiotemporal dynamics of video data." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 630, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 554, + 714 + ], + "type": "text", + "content": "An interesting discovery of this paper is that skipping layers within the network is an effective approach for constructing an aligned weak model. Modern neural network architectures for video diffusion models, such as ADM [25], DiT [26], and SiT [23], are partially or fully transformer-based and contain attention layers and residual blocks. These architectures are well-suited for layer skipping, as" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "11009" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 80, + 555, + 167 + ], + "blocks": [ + { + "bbox": [ + 57, + 80, + 555, + 167 + ], + "lines": [ + { + "bbox": [ + 57, + 80, + 555, + 167 + ], + "spans": [ + { + "bbox": [ + 57, + 80, + 555, + 167 + ], + "type": "image", + "image_path": "5070aabb09be4e8193c02ce195eebaa51b7ce9f83b565d3636ae12aadeb0475b.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 137, + 181, + 473, + 192 + ], + "lines": [ + { + "bbox": [ + 137, + 181, + 473, + 192 + ], + "spans": [ + { + "bbox": [ + 137, + 181, + 473, + 192 + ], + "type": "text", + "content": "Figure 3. Selected frames from videos generated by Mochi [33] with increasing STG scales." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 70, + 202, + 280, + 327 + ], + "blocks": [ + { + "bbox": [ + 70, + 202, + 280, + 327 + ], + "lines": [ + { + "bbox": [ + 70, + 202, + 280, + 327 + ], + "spans": [ + { + "bbox": [ + 70, + 202, + 280, + 327 + ], + "type": "image", + "image_path": "3ec1b25622f72d9107003e686df759da57fe5afc1b25b233f755826228815b4e.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 338, + 295, + 361 + ], + "lines": [ + { + "bbox": [ + 55, + 338, + 295, + 361 + ], + "spans": [ + { + "bbox": [ + 55, + 338, + 295, + 361 + ], + "type": "text", + "content": "Figure 4. Comparison of CFG and STG across varying scales in terms of Imaging Quality and FVD." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 365, + 295, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 365, + 295, + 389 + ], + "spans": [ + { + "bbox": [ + 55, + 365, + 295, + 389 + ], + "type": "text", + "content": "they can still produce plausible outputs without significant degradation, even when a few layers are removed." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 389, + 295, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 389, + 295, + 472 + ], + "spans": [ + { + "bbox": [ + 55, + 389, + 295, + 472 + ], + "type": "text", + "content": "Layer skipping is advantageous for generating an aligned weak model because it retains most of the neural network's weights and forward pass, resulting in similar predictions and distributions. This approach offers a clear advantage over methods that rely on external models, such as Autoguidance [19], or alternative objectives like CFG [10], where forward passes differ significantly." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 473, + 295, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 473, + 295, + 567 + ], + "spans": [ + { + "bbox": [ + 55, + 473, + 295, + 567 + ], + "type": "text", + "content": "Moreover, our method applies perturbations to both spatial and temporal layers (or spatiotemporal layers), unlike existing image-based perturbation methods [13, 14] limited to 2D spatial attention layers. This dual-layer perturbation is essential for aligning the weak model with the main video diffusion model. We now discuss various STG configurations that can be applied across different video diffusion model architectures." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 574, + 295, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 574, + 295, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 574, + 295, + 601 + ], + "type": "text", + "content": "Residual skip To skip an entire residual block, we modify " + }, + { + "bbox": [ + 55, + 574, + 295, + 601 + ], + "type": "inline_equation", + "content": "\\operatorname{Res}(z_l)" + }, + { + "bbox": [ + 55, + 574, + 295, + 601 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 574, + 295, + 601 + ], + "type": "inline_equation", + "content": "\\operatorname{Res}'(z_l)" + }, + { + "bbox": [ + 55, + 574, + 295, + 601 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 113, + 610, + 295, + 623 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 610, + 295, + 623 + ], + "spans": [ + { + "bbox": [ + 113, + 610, + 295, + 623 + ], + "type": "interline_equation", + "content": "\\operatorname {R e s} \\left(z _ {l}\\right) = z _ {l + 1} = f _ {l} \\left(z _ {l}\\right) + z _ {l}, \\tag {14}", + "image_path": "7678f9fe64f0a1f97213363ed4bbfe8796ffa49940694380fdce3a70112a2ba9.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 129, + 632, + 295, + 647 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 632, + 295, + 647 + ], + "spans": [ + { + "bbox": [ + 129, + 632, + 295, + 647 + ], + "type": "interline_equation", + "content": "\\operatorname {R e s} ^ {\\prime} \\left(z _ {l}\\right) = z _ {l + 1} = z _ {l}, \\tag {15}", + "image_path": "cd04c77bd63662909cf575b8e722f463878c3da47b38e3d352e29ab65d809736.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 652, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 652, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 652, + 296, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 652, + 296, + 714 + ], + "type": "inline_equation", + "content": "z_{l}" + }, + { + "bbox": [ + 55, + 652, + 296, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 652, + 296, + 714 + ], + "type": "inline_equation", + "content": "z_{l + 1}" + }, + { + "bbox": [ + 55, + 652, + 296, + 714 + ], + "type": "text", + "content": " represent the features at the " + }, + { + "bbox": [ + 55, + 652, + 296, + 714 + ], + "type": "inline_equation", + "content": "l^{\\mathrm{th}}" + }, + { + "bbox": [ + 55, + 652, + 296, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 652, + 296, + 714 + ], + "type": "inline_equation", + "content": "(l + 1)^{\\mathrm{th}}" + }, + { + "bbox": [ + 55, + 652, + 296, + 714 + ], + "type": "text", + "content": " layers, respectively, and " + }, + { + "bbox": [ + 55, + 652, + 296, + 714 + ], + "type": "inline_equation", + "content": "f_{l}" + }, + { + "bbox": [ + 55, + 652, + 296, + 714 + ], + "type": "text", + "content": " denotes the " + }, + { + "bbox": [ + 55, + 652, + 296, + 714 + ], + "type": "inline_equation", + "content": "l^{\\mathrm{th}}" + }, + { + "bbox": [ + 55, + 652, + 296, + 714 + ], + "type": "text", + "content": " neural net layer. The residual layers, which add small residuals " + }, + { + "bbox": [ + 55, + 652, + 296, + 714 + ], + "type": "inline_equation", + "content": "f_{l}(z_{l})" + }, + { + "bbox": [ + 55, + 652, + 296, + 714 + ], + "type": "text", + "content": " to the original feature " + }, + { + "bbox": [ + 55, + 652, + 296, + 714 + ], + "type": "inline_equation", + "content": "z_{l}" + }, + { + "bbox": [ + 55, + 652, + 296, + 714 + ], + "type": "text", + "content": ", ensure that the perturbed " + }, + { + "bbox": [ + 55, + 652, + 296, + 714 + ], + "type": "inline_equation", + "content": "z_{l + 1}" + }, + { + "bbox": [ + 55, + 652, + 296, + 714 + ], + "type": "text", + "content": " does not deviate significantly from the original " + }, + { + "bbox": [ + 55, + 652, + 296, + 714 + ], + "type": "inline_equation", + "content": "z_{l + 1}" + }, + { + "bbox": [ + 55, + 652, + 296, + 714 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 202, + 553, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 202, + 553, + 227 + ], + "spans": [ + { + "bbox": [ + 313, + 202, + 553, + 227 + ], + "type": "text", + "content": "This reduces out-of-distribution issues in consecutive layers, generating perturbed but aligned samples." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 234, + 553, + 257 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 234, + 553, + 257 + ], + "spans": [ + { + "bbox": [ + 313, + 234, + 553, + 257 + ], + "type": "text", + "content": "Attention skip Self-attention computes a linear combination of value tensors as" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 330, + 264, + 553, + 293 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 330, + 264, + 553, + 293 + ], + "spans": [ + { + "bbox": [ + 330, + 264, + 553, + 293 + ], + "type": "interline_equation", + "content": "\\operatorname {S A} (Q, K, V) = \\operatorname {S o f t m a x} \\left(\\frac {Q K ^ {T}}{\\sqrt {d}}\\right) V = \\mathbf {A} V, \\tag {16}", + "image_path": "d917503572ae2aa94f95a29c23b9d328142f464ebc859fd2c6b69311dd720e53.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 300, + 553, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 300, + 553, + 350 + ], + "spans": [ + { + "bbox": [ + 313, + 300, + 553, + 350 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 300, + 553, + 350 + ], + "type": "inline_equation", + "content": "Q \\in \\mathbb{R}^{(h \\times w \\times f) \\times d}" + }, + { + "bbox": [ + 313, + 300, + 553, + 350 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 300, + 553, + 350 + ], + "type": "inline_equation", + "content": "K \\in \\mathbb{R}^{(h \\times w \\times f) \\times d}" + }, + { + "bbox": [ + 313, + 300, + 553, + 350 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 300, + 553, + 350 + ], + "type": "inline_equation", + "content": "V \\in \\mathbb{R}^{(h \\times w \\times f) \\times d}" + }, + { + "bbox": [ + 313, + 300, + 553, + 350 + ], + "type": "text", + "content": " are the query, key, and value matrices, respectively. Here, " + }, + { + "bbox": [ + 313, + 300, + 553, + 350 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 313, + 300, + 553, + 350 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 300, + 553, + 350 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 313, + 300, + 553, + 350 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 300, + 553, + 350 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 313, + 300, + 553, + 350 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 300, + 553, + 350 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 313, + 300, + 553, + 350 + ], + "type": "text", + "content": " represent the height, width, frame number, and channel dimensions." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 350, + 554, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 350, + 554, + 409 + ], + "spans": [ + { + "bbox": [ + 313, + 350, + 554, + 409 + ], + "type": "text", + "content": "We can skip this layer partially by passing the value matrix directly to the next layer without computing its linear combination. This is equivalent to replacing the attention matrix " + }, + { + "bbox": [ + 313, + 350, + 554, + 409 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 313, + 350, + 554, + 409 + ], + "type": "text", + "content": " with an identity matrix " + }, + { + "bbox": [ + 313, + 350, + 554, + 409 + ], + "type": "inline_equation", + "content": "\\mathbf{I} \\in \\mathbb{R}^{hwf \\times hwf}" + }, + { + "bbox": [ + 313, + 350, + 554, + 409 + ], + "type": "text", + "content": ", resulting in" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 390, + 410, + 553, + 423 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 390, + 410, + 553, + 423 + ], + "spans": [ + { + "bbox": [ + 390, + 410, + 553, + 423 + ], + "type": "interline_equation", + "content": "\\mathrm {S A} ^ {\\prime} (Q, K, V) = \\mathbf {I} V. \\tag {17}", + "image_path": "e0bb4fdf15a7d0df7198eb4713589467729a7e1c174f1347cb47047382bc5bae.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 430, + 489, + 442 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 430, + 489, + 442 + ], + "spans": [ + { + "bbox": [ + 313, + 430, + 489, + 442 + ], + "type": "text", + "content": "This represents a 3D extension of PAG [1]." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 449, + 554, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 449, + 554, + 557 + ], + "spans": [ + { + "bbox": [ + 313, + 449, + 554, + 557 + ], + "type": "text", + "content": "Factorized attention While recent models like Movie Gen [27] and Mochi [33] utilize full 3D spatiotemporal attention layers, many models, such as SVD [4] and OpenSora [38], still use factorized attention layers for efficiency. These models employ sequential 2D spatial attention and 1D temporal attention to approximate 3D spatiotemporal attention. For factorized models, we apply skip perturbation of Eq. 17 to spatial and temporal layers separately and use their linear combination for the final guidance." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 558, + 554, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 554, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 554, + 689 + ], + "type": "text", + "content": "We denote the spatial and temporal perturbation labels as " + }, + { + "bbox": [ + 313, + 558, + 554, + 689 + ], + "type": "inline_equation", + "content": "y_{sb}" + }, + { + "bbox": [ + 313, + 558, + 554, + 689 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 558, + 554, + 689 + ], + "type": "inline_equation", + "content": "y_{tb}" + }, + { + "bbox": [ + 313, + 558, + 554, + 689 + ], + "type": "text", + "content": ", respectively. In practice, " + }, + { + "bbox": [ + 313, + 558, + 554, + 689 + ], + "type": "inline_equation", + "content": "y_{sb}" + }, + { + "bbox": [ + 313, + 558, + 554, + 689 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 558, + 554, + 689 + ], + "type": "inline_equation", + "content": "y_{tb}" + }, + { + "bbox": [ + 313, + 558, + 554, + 689 + ], + "type": "text", + "content": " are rarely independent, as they can influence each other. For instance, random spatial perturbations that create varied color adjustments across frames may disrupt temporal continuity. Similarly, random temporal perturbations can affect spatial consistency, leading to distortions in specific frames. However, for our skip perturbations, we can loosely assume independence, as skipping layers primarily reduces details in residual networks, which may not inherently affect temporal consistency." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "content": "Therefore, to simplify the derivation, we initially assume independence between spatial and temporal perturbations." + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 97, + 69, + 511, + 79 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 69, + 511, + 79 + ], + "spans": [ + { + "bbox": [ + 97, + 69, + 511, + 79 + ], + "type": "text", + "content": "A macro cinematography animation showing a butterfly emerging from its chrysalis, filmed with side-lit lighting ..." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "11010" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 71, + 534, + 114 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 55, + 71, + 517, + 82 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 71, + 517, + 82 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 517, + 82 + ], + "type": "text", + "content": "(1) A romantic scene of a couple dancing under string lights in a backyard, with warm, golden tones highlighting their laughter." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 82, + 528, + 92 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 82, + 528, + 92 + ], + "spans": [ + { + "bbox": [ + 55, + 82, + 528, + 92 + ], + "type": "text", + "content": "(2) An animation showing a floating castle drifting above the clouds, with birds flying around it and sunlight casting golden rays ..." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 92, + 534, + 103 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 92, + 534, + 103 + ], + "spans": [ + { + "bbox": [ + 55, + 92, + 534, + 103 + ], + "type": "text", + "content": "(3) A realistic documentary-style video of artisans crafting pottery, with the scene unfolding and transforming as hands shape clay ..." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 103, + 507, + 114 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 103, + 507, + 114 + ], + "spans": [ + { + "bbox": [ + 55, + 103, + 507, + 114 + ], + "type": "text", + "content": "(4) A ghost in a white bedsheet faces a mirror. The ghost's reflection can be seen in the mirror. The ghost is in a dusty attic ..." + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "text" + }, + { + "type": "image", + "bbox": [ + 64, + 116, + 551, + 700 + ], + "blocks": [ + { + "bbox": [ + 64, + 116, + 551, + 700 + ], + "lines": [ + { + "bbox": [ + 64, + 116, + 551, + 700 + ], + "spans": [ + { + "bbox": [ + 64, + 116, + 551, + 700 + ], + "type": "image", + "image_path": "70c7a672903779f88253a7cf57412338b5e58688bb386b7bc4f3512985c07ec8.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 135, + 708, + 474, + 720 + ], + "lines": [ + { + "bbox": [ + 135, + 708, + 474, + 720 + ], + "spans": [ + { + "bbox": [ + 135, + 708, + 474, + 720 + ], + "type": "text", + "content": "Figure 5. Qualitative comparison between CFG and STG on videos generated by Mochi [33]." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "11011" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 109 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 109 + ], + "type": "text", + "content": "We will revisit this assumption and propose an adjustment afterward. Under this independence assumption, the joint distribution can be expressed as" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 120, + 295, + 133 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 120, + 295, + 133 + ], + "spans": [ + { + "bbox": [ + 105, + 120, + 295, + 133 + ], + "type": "interline_equation", + "content": "p _ {\\theta} \\left(x _ {t} \\mid y _ {b}\\right) = p _ {\\theta} \\left(x _ {t} \\mid y _ {\\mathrm {s b}}\\right) p _ {\\theta} \\left(x _ {t} \\mid y _ {\\mathrm {f b}}\\right). \\tag {18}", + "image_path": "826318f003c8bfc1dcb52fb7bd37dc39b4a35334146e982aad47e841e061a08b.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 143, + 295, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 143, + 295, + 167 + ], + "spans": [ + { + "bbox": [ + 55, + 143, + 295, + 167 + ], + "type": "text", + "content": "Following the same approach as in Eq. 9 and using Eq. 11, we modify the score in Eq. 8 as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 73, + 177, + 295, + 222 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 177, + 295, + 222 + ], + "spans": [ + { + "bbox": [ + 73, + 177, + 295, + 222 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\nabla_ {x _ {t}} \\log \\tilde {p} _ {\\theta} (x _ {t} | y _ {g}) = \\nabla_ {x _ {t}} \\log p _ {\\theta} (x _ {t}) \\\\ + w \\nabla_ {x _ {t}} \\left(\\log p _ {\\theta} (x _ {t}) - \\log p _ {\\theta} (x _ {t} | y _ {s})\\right) \\\\ + w \\nabla_ {x _ {t}} \\left(\\log p _ {\\theta} \\left(x _ {t}\\right) - \\log p _ {\\theta} \\left(x _ {t} \\mid y _ {t}\\right)\\right). \\tag {19} \\\\ \\end{array}", + "image_path": "78f138fdff2fda97a31194cea4ff130ab332feaf4296c7c5dfb38fee6f878775.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 232, + 296, + 269 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 232, + 296, + 269 + ], + "spans": [ + { + "bbox": [ + 55, + 232, + 296, + 269 + ], + "type": "text", + "content": "By replacing the score with the estimated denoiser and utilizing separate scales for spatial and temporal terms, we obtain the following equation:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 92, + 278, + 294, + 307 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 92, + 278, + 294, + 307 + ], + "spans": [ + { + "bbox": [ + 92, + 278, + 294, + 307 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\tilde {\\epsilon} _ {\\theta} ^ {w} (x _ {t}) = \\epsilon_ {\\theta} (x _ {t}) + w _ {1} \\left(\\epsilon_ {\\theta} (x _ {t}) - \\epsilon_ {\\theta} ^ {s} (x _ {t})\\right) \\\\ + w _ {2} \\left(\\epsilon_ {\\theta} \\left(x _ {t}\\right) - \\epsilon_ {\\theta} ^ {t} \\left(x _ {t}\\right)\\right) \\tag {20} \\\\ \\end{array}", + "image_path": "77d6f628b15b6a19e75413e6d5d5c604d00961d3dbd09bd6ea51858675b38ee0.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 317, + 295, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 317, + 295, + 342 + ], + "spans": [ + { + "bbox": [ + 55, + 317, + 295, + 342 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 317, + 295, + 342 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}^{\\mathrm{s}}(x_t)" + }, + { + "bbox": [ + 55, + 317, + 295, + 342 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 317, + 295, + 342 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}^{\\mathrm{t}}(x_t)" + }, + { + "bbox": [ + 55, + 317, + 295, + 342 + ], + "type": "text", + "content": " represent spatially and temporally perturbed models, respectively." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 342, + 296, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 342, + 296, + 426 + ], + "spans": [ + { + "bbox": [ + 55, + 342, + 296, + 426 + ], + "type": "text", + "content": "Next, we revisit our assumption of independence between spatial and temporal perturbations. While Eq. 20 works well in practice, we can derive an alternative STG formulation that uses orthogonalization to isolate the independent components of the spatial and temporal guidance. Inspired by a negative prompting technique [3], we can modify the Eq. 20 as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 64, + 439, + 295, + 483 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 439, + 295, + 483 + ], + "spans": [ + { + "bbox": [ + 64, + 439, + 295, + 483 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\tilde {\\epsilon} _ {\\theta} ^ {w} (x _ {t}) = \\epsilon_ {\\theta} (x _ {t}) + w _ {1} \\Delta_ {\\mathrm {s}} \\\\ + w _ {2} \\left(\\Delta_ {\\mathrm {t}} - \\frac {\\left\\langle \\Delta_ {\\mathrm {s}} , \\Delta_ {\\mathrm {t}} \\right\\rangle}{\\left\\| \\Delta_ {\\mathrm {s}} \\right\\| ^ {2}} \\Delta_ {\\mathrm {s}}\\right), \\tag {21} \\\\ \\end{array}", + "image_path": "b82c18ef47b80f3d08b8c3a5bb4c0f25cc8fd63bd7de4c0408a281f6b8e15ad1.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 491, + 282, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 491, + 282, + 506 + ], + "spans": [ + { + "bbox": [ + 55, + 491, + 282, + 506 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 491, + 282, + 506 + ], + "type": "inline_equation", + "content": "\\Delta_{\\mathrm{s}} = \\epsilon_{\\theta}(x_{t}) - \\epsilon_{\\theta}^{\\mathrm{s}}(x_{t})" + }, + { + "bbox": [ + 55, + 491, + 282, + 506 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 491, + 282, + 506 + ], + "type": "inline_equation", + "content": "\\Delta_{\\mathrm{t}} = \\epsilon_{\\theta}(x_{t}) - \\epsilon_{\\theta}^{\\mathrm{t}}(x_{t})" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 521, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 296, + 714 + ], + "type": "text", + "content": "Manifold Constrained Guidance Even with a well-aligned weak model, larger guidance scales " + }, + { + "bbox": [ + 55, + 521, + 296, + 714 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 55, + 521, + 296, + 714 + ], + "type": "text", + "content": " can drive samples off the data manifold, resulting in poor quality and oversaturation. To address error accumulation at high guidance scales, we explore optional techniques to keep samples constrained to the manifold. Rescaling the latent code [21] helps mitigate this issue by constraining its variance, as larger variance is known to cause saturation in the results. Additionally, Restart sampling [36] demonstrates that introducing stochasticity can correct off-manifold deviations. Building on this idea, we incorporate stochastic forward processes into our sampling guidance framework as an optional method. While this approach modestly improves final sample quality and reduces saturation, it introduces additional computational overhead. Further details are provided in Appendix A5.1." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 71, + 395, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 71, + 395, + 85 + ], + "spans": [ + { + "bbox": [ + 314, + 71, + 395, + 85 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 91, + 382, + 102 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 91, + 382, + 102 + ], + "spans": [ + { + "bbox": [ + 314, + 91, + 382, + 102 + ], + "type": "text", + "content": "5.1. Overview" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 108, + 500, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 108, + 500, + 120 + ], + "spans": [ + { + "bbox": [ + 314, + 108, + 500, + 120 + ], + "type": "text", + "content": "We employ three models for our experiments:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 314, + 121, + 553, + 228 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 314, + 121, + 553, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 121, + 553, + 157 + ], + "spans": [ + { + "bbox": [ + 314, + 121, + 553, + 157 + ], + "type": "text", + "content": "- Mochi [33] is a text-to-video model built on AsymmDiT blocks, containing a total of 10 billion parameters and utilizing 3D self-attention in its spatiotemporal layers." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 314, + 158, + 553, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 158, + 553, + 193 + ], + "spans": [ + { + "bbox": [ + 314, + 158, + 553, + 193 + ], + "type": "text", + "content": "- Open-Sora [38] is a text-to-video model built on STDiT blocks with 1.1 billion parameters, employing factorized spatial and temporal attention layers." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 314, + 194, + 553, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 194, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 314, + 194, + 553, + 228 + ], + "type": "text", + "content": "- SVD [4] is an image-to-video model with 1.5 billion parameters that leverages factorized spatial and temporal attention within a UNet architecture." + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 231, + 553, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 231, + 553, + 266 + ], + "spans": [ + { + "bbox": [ + 313, + 231, + 553, + 266 + ], + "type": "text", + "content": "We evaluate the proposed method using widely adopted datasets and metrics to ensure a comprehensive performance analysis." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 314, + 268, + 554, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 268, + 554, + 376 + ], + "spans": [ + { + "bbox": [ + 314, + 268, + 554, + 376 + ], + "type": "text", + "content": "- UCF-101 The UCF-101 dataset comprises 13,320 videos organized into 101 action classes. Using this dataset, we assess the Fréchet Video Distance (FVD) and Inception Score (IS) of the image-to-video model SVD [4]. Following DIGAN [37], we calculate FVD and IS on 2,048 and 10,000 samples, respectively. For conditioning, initial frames from UCF-101 videos are used as inputs to the SVD model, with each input frame serving as the starting frame of the generated videos." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 314, + 376, + 554, + 532 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 314, + 376, + 554, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 376, + 554, + 472 + ], + "spans": [ + { + "bbox": [ + 314, + 376, + 554, + 472 + ], + "type": "text", + "content": "- VBench We use VBench [15] for automatic evaluation across various video metrics. SVD is evaluated using the image-to-video (I2V) framework on 355 samples from the VBench I2V dataset with 5 random seeds. Mochi and Open-Sora are evaluated with the text-to-video (T2V) framework: Open-Sora uses the standard VBench prompt list, while Mochi uses 100 randomly selected prompts due to computational limits." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 314, + 472, + 553, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 472, + 553, + 495 + ], + "spans": [ + { + "bbox": [ + 314, + 472, + 553, + 495 + ], + "type": "text", + "content": "- EvalCrafter We perform human evaluations using 700 prompts from the EvalCrafter dataset [22]." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 314, + 496, + 553, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 496, + 553, + 532 + ], + "spans": [ + { + "bbox": [ + 314, + 496, + 553, + 532 + ], + "type": "text", + "content": "- LLM-Generated Prompts We use a set of 100 selected prompts generated by Claude 3.5 Sonnet [2] for our demo and qualitative comparisons." + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 533, + 554, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 554, + 582 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 554, + 582 + ], + "type": "text", + "content": "For evaluation, CFG is used together with STG for our evaluation. We did not use Restart sampling or guidance orthogonalization; additional results with these techniques are in the Appendix sections A5.1 and A5.2." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 314, + 588, + 372, + 600 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 588, + 372, + 600 + ], + "spans": [ + { + "bbox": [ + 314, + 588, + 372, + 600 + ], + "type": "text", + "content": "5.2. Results" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 605, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 554, + 713 + ], + "type": "text", + "content": "Qualitative Comparison Fig. 5 compares CFG and our method. Videos from the naive CFG model often show blurry objects with indistinct shapes, while STG produces clearer, more vivid frames with sharper image quality. STG also reduces temporal inconsistency and flickering, especially in dynamic videos with large motion where CFG frequently fails. Spatial guidance enhances object structure, and temporal guidance improves consistency. Please refer to the Appendix sections A5 and A6 for additional results." + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "11012" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 57, + 70, + 553, + 159 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 553, + 159 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 553, + 159 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 553, + 159 + ], + "type": "table", + "html": "
ModelsFVD (↓)ISImaging QualityAesthetic QualityMotion SmoothnessDynamic DegreeTemporal Flickering
Mochi (STG-R)--0.6280.5540.9880.860.978
Mochi (STG-A)--0.5550.5410.9870.860.976
Open-Sora (STG-R)--0.5500.4740.9810.8940.977
Open-Sora (STG-A)--0.6060.5090.9870.8950.976
SVD (STG-R)155.939.30.6870.6370.9650.641-
SVD (STG-A)128.738.50.6940.6390.9680.694-
", + "image_path": "f7b0fe222a09f8bfbfbe8ae259fdb0fe99d5dabbd1449264c4793f988e6f6d5d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 83, + 200, + 526, + 259 + ], + "blocks": [ + { + "bbox": [ + 55, + 167, + 555, + 190 + ], + "lines": [ + { + "bbox": [ + 55, + 167, + 555, + 190 + ], + "spans": [ + { + "bbox": [ + 55, + 167, + 555, + 190 + ], + "type": "text", + "content": "Table 3. Comparison of STG-R (residual skip) and STG-A (attention skip) across Mochi [33], Open-Sora [38], and SVD [4]. STG-R shows stronger performance on Mochi, while STG-A yields better results on Open-Sora and SVD." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 83, + 200, + 526, + 259 + ], + "lines": [ + { + "bbox": [ + 83, + 200, + 526, + 259 + ], + "spans": [ + { + "bbox": [ + 83, + 200, + 526, + 259 + ], + "type": "table", + "html": "
ModelsFVD (↓)ISImaging QualityAesthetic QualityMotion SmoothnessDynamic Degree
CFG151.338.00.6870.6370.9660.562
+ Spatial133.838.30.6910.6390.9670.659
+ Temporal128.738.50.6940.6380.9680.694
", + "image_path": "9f7e52e1289188afa717d95934d6a7cff5163d5d61408c290e7bcddd1b86a3b5.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 78, + 267, + 529, + 278 + ], + "lines": [ + { + "bbox": [ + 78, + 267, + 529, + 278 + ], + "spans": [ + { + "bbox": [ + 78, + 267, + 529, + 278 + ], + "type": "text", + "content": "Table 4. Ablation study results on SVD [4] factorized attention, showing the impact of adding spatial and temporal guidance." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 293, + 294, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 293, + 294, + 376 + ], + "spans": [ + { + "bbox": [ + 54, + 293, + 294, + 376 + ], + "type": "text", + "content": "Quantitative Comparison We compare CFG and STG using the FVD-Imaging Quality (VBench) plot across different scales in Fig. 4. FVD measures video distribution, while Imaging Quality assesses frame clarity. Higher CFG scales improve Imaging Quality but reduce diversity, as reflected in higher FVD. STG avoids this trade-off, maintaining diversity at increased scales." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 377, + 295, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 377, + 295, + 449 + ], + "spans": [ + { + "bbox": [ + 54, + 377, + 295, + 449 + ], + "type": "text", + "content": "T2V and I2V VBenchmark metrics in Tab.1 and Tab.2 show notable improvements in frame-level quality (Imaging and Aesthetic Quality). Temporal quality improves qualitatively, though metrics like Motion Smoothness and Temporal Flickering show marginal gains, as these scores are near saturation " + }, + { + "bbox": [ + 54, + 377, + 295, + 449 + ], + "type": "inline_equation", + "content": "(\\sim 0.9x)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 449, + 295, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 449, + 295, + 533 + ], + "spans": [ + { + "bbox": [ + 54, + 449, + 295, + 533 + ], + "type": "text", + "content": "For Dynamic Degree, we aim to keep it unchanged, as it may not correlate with video quality. This is achieved in the T2V metric, but in the I2V model, CFG increases the influence of the conditioned image, reducing motion. STG, while not directly affecting Dynamic Degree, mitigates CFG's impact, leading to increased Dynamic Degree in I2V when used together." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 533, + 295, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 533, + 295, + 569 + ], + "spans": [ + { + "bbox": [ + 54, + 533, + 295, + 569 + ], + "type": "text", + "content": "Human Evaluation We provide human evaluation results on 700 prompts from the EvalCrafter dataset [22] in the Appendix A4." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 576, + 149, + 589 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 576, + 149, + 589 + ], + "spans": [ + { + "bbox": [ + 55, + 576, + 149, + 589 + ], + "type": "text", + "content": "5.3. Ablation Study" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 54, + 594, + 295, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 594, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 54, + 594, + 295, + 677 + ], + "type": "text", + "content": "Fig. 3 displays selected frames from videos generated by Mochi using various STG scales, where " + }, + { + "bbox": [ + 54, + 594, + 295, + 677 + ], + "type": "inline_equation", + "content": "w = 0" + }, + { + "bbox": [ + 54, + 594, + 295, + 677 + ], + "type": "text", + "content": " corresponds to the CFG-only model without STG. Increasing the STG scale results in more vivid colors and finer details compared to the monotonous colors at " + }, + { + "bbox": [ + 54, + 594, + 295, + 677 + ], + "type": "inline_equation", + "content": "w = 0" + }, + { + "bbox": [ + 54, + 594, + 295, + 677 + ], + "type": "text", + "content": ". Notably, unlike with CFG, sampling diversity is preserved as the STG scale increases, as shown in Fig. 4." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 54, + 677, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 677, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 54, + 677, + 295, + 713 + ], + "type": "text", + "content": "We performed an ablation study on SVD to evaluate spatial and temporal guidance (Tab. 4). Spatial guidance notably reduced FVD and improved metrics, while tempo" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 293, + 553, + 318 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 293, + 553, + 318 + ], + "spans": [ + { + "bbox": [ + 313, + 293, + 553, + 318 + ], + "type": "text", + "content": "ral guidance further boosted performance, confirming their combined effectiveness for spatiotemporal generation." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 319, + 554, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 319, + 554, + 487 + ], + "spans": [ + { + "bbox": [ + 313, + 319, + 554, + 487 + ], + "type": "text", + "content": "We also test two STG variants—residual skip (STGR) and attention skip (STG-A)—on SVD, Open-Sora, and Mochi. As shown in Tab. 3, STG-R performs well with Mochi, while STG-A is more effective for the other models. This is likely due to Mochi's higher parameter count and layer depth, enabling more extensive skipping without triggering out-of-distribution (OOD) issues in consecutive layers. Furthermore, Mochi's spatiotemporal layers consist of a single 3D attention layer, so STG-R skips a residual layer with just one attention layer. In contrast, SVD and Open-Sora use factorized attention, meaning STG-R skips a residual layer with two attention layers, which may cause excessive perturbation and lead to OOD issues in subsequent layers. More results are available in the Appendix A5." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 501, + 388, + 513 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 501, + 388, + 513 + ], + "spans": [ + { + "bbox": [ + 313, + 501, + 388, + 513 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 522, + 554, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 522, + 554, + 606 + ], + "spans": [ + { + "bbox": [ + 313, + 522, + 554, + 606 + ], + "type": "text", + "content": "We proposed Spatiotemporal Skip Guidance (STG), a simple, training-free method for video diffusion models. By simulating an aligned weak model via spatiotemporal skipping, STG offers strong guidance for high-fidelity video generation, achieving notable qualitative and quantitative improvements. We hope this work advances video diffusion models and inspires further research in the field." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 629, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 554, + 713 + ], + "type": "text", + "content": "Limitation and Ethical Considerations STG's performance depends on scale and layer selection, with the optimal configuration varying across models, requiring users to set these through heuristic tuning. While video quality improvements are notable, they also raise ethical concerns about misuse, underscoring the importance of using this technology responsibly and constructively." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "11013" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 167 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 167 + ], + "type": "text", + "content": "Acknowledgements This work was supported by Institute for Information & communications Technology Planning & Evaluation(IITP) grant funded by the Korea government(MSIT) (RS-2019-II190075, Artificial Intelligence Graduate School Program(KAIST)). This work was supported by the National Research Foundation of Korea(NRF) grant funded by the Korea government(MSIT) (No. RS2025-00555621)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 178, + 115, + 190 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 178, + 115, + 190 + ], + "spans": [ + { + "bbox": [ + 56, + 178, + 115, + 190 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 198, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 198, + 296, + 253 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 198, + 296, + 253 + ], + "spans": [ + { + "bbox": [ + 61, + 198, + 296, + 253 + ], + "type": "text", + "content": "[1] Donghoon Ahn, Hyoungwon Cho, Jaewon Min, Wooseok Jang, Jungwoo Kim, SeonHwa Kim, Hyun Hee Park, Kyong Hwan Jin, and Seungryong Kim. Self-rectifying diffusion sampling with perturbed-attention guidance. arXiv preprint arXiv:2403.17377, 2024. 2, 3, 5, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 255, + 294, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 255, + 294, + 275 + ], + "spans": [ + { + "bbox": [ + 62, + 255, + 294, + 275 + ], + "type": "text", + "content": "[2] Anthropic. The claude 3 model family: Opus, sonnet, haiku. In-, 2024. 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 277, + 295, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 277, + 295, + 331 + ], + "spans": [ + { + "bbox": [ + 62, + 277, + 295, + 331 + ], + "type": "text", + "content": "[3] Mohammadreza Armandpour, Ali Sadeghian, Huangjie Zheng, Amir Sadeghian, and Mingyuan Zhou. Re-imagine the negative prompt algorithm: Transform 2d diffusion into 3d, alleviate janus problem and beyond. arXiv preprint arXiv:2304.04968, 2023. 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 333, + 295, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 333, + 295, + 399 + ], + "spans": [ + { + "bbox": [ + 62, + 333, + 295, + 399 + ], + "type": "text", + "content": "[4] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 2, 4, 5, 7, 8, 1, 9, 11, 13, 14, 17, 18, 23" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 400, + 294, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 400, + 294, + 431 + ], + "spans": [ + { + "bbox": [ + 62, + 400, + 294, + 431 + ], + "type": "text", + "content": "[5] Andrew Brock. Large scale gan training for high fidelity natural image synthesis. arXiv preprint arXiv:1809.11096, 2018. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 434, + 295, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 434, + 295, + 477 + ], + "spans": [ + { + "bbox": [ + 62, + 434, + 295, + 477 + ], + "type": "text", + "content": "[6] Hyungjin Chung, Jeongsol Kim, Geon Yeong Park, Hyelin Nam, and Jong Chul Ye. Cfg++: Manifold-constrained classifier free guidance for diffusion models. arXiv preprint arXiv:2406.08070, 2024. 2, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 479, + 294, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 479, + 294, + 511 + ], + "spans": [ + { + "bbox": [ + 62, + 479, + 294, + 511 + ], + "type": "text", + "content": "[7] Zihan Ding, Xiao-Yang Liu, Miao Yin, and Linghe Kong. Tgan: Deep tensor generative adversarial nets for large image generation. arXiv preprint arXiv:1901.09953, 2019. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 513, + 295, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 513, + 295, + 566 + ], + "spans": [ + { + "bbox": [ + 62, + 513, + 295, + 566 + ], + "type": "text", + "content": "[8] Songwei Ge, Aniruddha Mahapatra, Gaurav Parmar, JunYan Zhu, and Jia-Bin Huang. On the content bias in fréchet video distance. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7277-7288, 2024. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 568, + 295, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 568, + 295, + 612 + ], + "spans": [ + { + "bbox": [ + 62, + 568, + 295, + 612 + ], + "type": "text", + "content": "[9] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. Communications of the ACM, 63(11):139-144, 2020. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 613, + 294, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 613, + 294, + 635 + ], + "spans": [ + { + "bbox": [ + 57, + 613, + 294, + 635 + ], + "type": "text", + "content": "[10] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 2, 3, 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 636, + 295, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 636, + 295, + 669 + ], + "spans": [ + { + "bbox": [ + 57, + 636, + 295, + 669 + ], + "type": "text", + "content": "[11] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 670, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 295, + 712 + ], + "type": "text", + "content": "[12] Jonathan Ho, Tim Salimans, Alexey Gritsenko, William Chan, Mohammad Norouzi, and David J Fleet. Video diffusion models. Advances in Neural Information Processing Systems, 35:8633-8646, 2022. 2" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "text", + "content": "[13] Susung Hong. Smoothed energy guidance: Guiding diffusion models with reduced energy curvature of attention. arXiv preprint arXiv:2408.00760, 2024. 2, 3, 5, 8, 18" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 108, + 553, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 108, + 553, + 162 + ], + "spans": [ + { + "bbox": [ + 317, + 108, + 553, + 162 + ], + "type": "text", + "content": "[14] Susung Hong, Gyuseong Lee, Wooseok Jang, and Seunglyong Kim. Improving sample quality of diffusion models using self-attention guidance. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7462-7471, 2023. 3, 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 164, + 555, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 164, + 555, + 229 + ], + "spans": [ + { + "bbox": [ + 317, + 164, + 555, + 229 + ], + "type": "text", + "content": "[15] Ziqi Huang, Yinan He, Jiashuo Yu, Fan Zhang, Chenyang Si, Yuming Jiang, Yuanhan Zhang, Tianxing Wu, Qingyang Jin, Nattapol Chanpaisit, et al. Vbench: Comprehensive benchmark suite for video generative models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21807-21818, 2024. 4, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 231, + 553, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 231, + 553, + 295 + ], + "spans": [ + { + "bbox": [ + 317, + 231, + 553, + 295 + ], + "type": "text", + "content": "[16] Yangqing Jia, Evan Shelhamer, Jeff Donahue, Sergey Karayev, Jonathan Long, Ross Girshick, Sergio Guadarrama, and Trevor Darrell. Caffe: Convolutional architecture for fast feature embedding. In Proceedings of the 22nd ACM international conference on Multimedia, pages 675-678, 2014." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 297, + 553, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 297, + 553, + 353 + ], + "spans": [ + { + "bbox": [ + 317, + 297, + 553, + 353 + ], + "type": "text", + "content": "[17] Andrej Karpathy, George Toderici, Sanketh Shetty, Thomas Leung, Rahul Sukthankar, and Li Fei-Fei. Large-scale video classification with convolutional neural networks. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 1725–1732, 2014. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 354, + 555, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 354, + 555, + 409 + ], + "spans": [ + { + "bbox": [ + 317, + 354, + 555, + 409 + ], + "type": "text", + "content": "[18] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of stylegan. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8110-8119, 2020. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 411, + 553, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 411, + 553, + 453 + ], + "spans": [ + { + "bbox": [ + 317, + 411, + 553, + 453 + ], + "type": "text", + "content": "[19] Tero Karras, Miika Aittala, Tuomas Kynkänniemi, Jaakko Lehtinen, Timo Aila, and Samuli Laine. Guiding a diffusion model with a bad version of itself. arXiv preprint arXiv:2406.02507, 2024. 2, 3, 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 456, + 553, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 456, + 553, + 487 + ], + "spans": [ + { + "bbox": [ + 316, + 456, + 553, + 487 + ], + "type": "text", + "content": "[20] Black Forest Labs. Flux: Github repository. https://github.com/black-forest-labs/flux, 2024. Accessed: 2024-11-22. 5, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 490, + 553, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 490, + 553, + 544 + ], + "spans": [ + { + "bbox": [ + 317, + 490, + 553, + 544 + ], + "type": "text", + "content": "[21] Shanchuan Lin, Bingchen Liu, Jiashi Li, and Xiao Yang. Common diffusion noise schedules and sample steps are flawed. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 5404-5411, 2024. 2, 7, 5, 10" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 546, + 555, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 546, + 555, + 612 + ], + "spans": [ + { + "bbox": [ + 317, + 546, + 555, + 612 + ], + "type": "text", + "content": "[22] Yaofang Liu, Xiaodong Cun, Xuebo Liu, Xintao Wang, Yong Zhang, Haoxin Chen, Yang Liu, Tieyong Zeng, Raymond Chan, and Ying Shan. Evalcrafter: Benchmarking and evaluating large video generation models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22139-22149, 2024. 7, 8, 5, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 613, + 553, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 613, + 553, + 666 + ], + "spans": [ + { + "bbox": [ + 317, + 613, + 553, + 666 + ], + "type": "text", + "content": "[23] Nanye Ma, Mark Goldstein, Michael S Albergo, Nicholas M Boffi, Eric Vanden-Eijnden, and Saining Xie. Sit: Exploring flow and diffusion-based generative models with scalable interpolant transformers. arXiv preprint arXiv:2401.08740, 2024. 4" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "text", + "content": "[24] Xinyin Ma, Gongfan Fang, Michael Bi Mi, and Xinchao Wang. Learning-to-cache: Accelerating diffusion transformer via layer caching. arXiv preprint arXiv:2406.01733, 2024. 4" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "11014" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 297, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 297, + 127 + ], + "type": "text", + "content": "[25] Soumik Mukhopadhyay, Matthew Gwilliam, Vatsal Agarwal, Namitha Padmanabhan, Archana Swaminathan, Srinidhi Hegde, Tianyi Zhou, and Abhinav Shrivastava. Diffusion models beat gans on image classification. arXiv preprint arXiv:2307.08702, 2023. 3, 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 129, + 296, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 129, + 296, + 172 + ], + "spans": [ + { + "bbox": [ + 55, + 129, + 296, + 172 + ], + "type": "text", + "content": "[26] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4195-4205, 2023. 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 174, + 297, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 174, + 297, + 228 + ], + "spans": [ + { + "bbox": [ + 56, + 174, + 297, + 228 + ], + "type": "text", + "content": "[27] Adam Polyak, Amit Zohar, Andrew Brown, Andros Tjandra, Animesh Sinha, Ann Lee, Apoorv Vyas, Bowen Shi, Chih-Yao Ma, Ching-Yao Chuang, et al. Movie gen: A cast of media foundation models. arXiv preprint arXiv:2410.13720, 2024. 2, 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 230, + 296, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 230, + 296, + 285 + ], + "spans": [ + { + "bbox": [ + 56, + 230, + 296, + 285 + ], + "type": "text", + "content": "[28] Hubert Ramsauer, Bernhard Schäfl, Johannes Lehner, Philipp Seidl, Michael Widrich, Thomas Adler, Lukas Gruber, Markus Holzleitner, Milena Pavlovic, Geir Kjetil Sandve, et al. Hopfield networks is all you need. arXiv preprint arXiv:2008.02217, 2020. 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 285, + 296, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 285, + 296, + 319 + ], + "spans": [ + { + "bbox": [ + 56, + 285, + 296, + 319 + ], + "type": "text", + "content": "[29] Ali Razavi, Aaron Van den Oord, and Oriol Vinyals. Generating diverse high-fidelity images with vq-vae-2. Advances in neural information processing systems, 32, 2019. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 319, + 296, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 319, + 296, + 374 + ], + "spans": [ + { + "bbox": [ + 56, + 319, + 296, + 374 + ], + "type": "text", + "content": "[30] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 376, + 296, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 376, + 296, + 421 + ], + "spans": [ + { + "bbox": [ + 56, + 376, + 296, + 421 + ], + "type": "text", + "content": "[31] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International conference on machine learning, pages 2256-2265. PMLR, 2015." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 421, + 296, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 421, + 296, + 465 + ], + "spans": [ + { + "bbox": [ + 56, + 421, + 296, + 465 + ], + "type": "text", + "content": "[32] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 2, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 466, + 296, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 466, + 296, + 498 + ], + "spans": [ + { + "bbox": [ + 56, + 466, + 296, + 498 + ], + "type": "text", + "content": "[33] Genmo Team. Mochi 1. https://github.com/ genmoai/models, 2024. 2, 4, 5, 6, 7, 8, 1, 9, 10, 15, 19, 20" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 500, + 296, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 500, + 296, + 554 + ], + "spans": [ + { + "bbox": [ + 56, + 500, + 296, + 554 + ], + "type": "text", + "content": "[34] Du Tran, Lubomir Bourdev, Rob Fergus, Lorenzo Torresani, and Manohar Paluri. Learning spatiotemporal features with 3d convolutional networks. In Proceedings of the IEEE international conference on computer vision, pages 4489-4497, 2015. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 555, + 296, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 555, + 296, + 588 + ], + "spans": [ + { + "bbox": [ + 56, + 555, + 296, + 588 + ], + "type": "text", + "content": "[35] Pascal Vincent. A connection between score matching and denoising autoencoders. Neural computation, 23(7):1661-1674, 2011. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 590, + 296, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 590, + 296, + 634 + ], + "spans": [ + { + "bbox": [ + 56, + 590, + 296, + 634 + ], + "type": "text", + "content": "[36] Yilun Xu, Mingyang Deng, Xiang Cheng, Yonglong Tian, Ziming Liu, and Tommi Jaakkola. Restart sampling for improving generative processes. Advances in Neural Information Processing Systems, 36:76806-76838, 2023. 2, 7, 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 635, + 296, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 296, + 679 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 296, + 679 + ], + "type": "text", + "content": "[37] Sihyun Yu, Jihoon Tack, Sangwoo Mo, Hyunsu Kim, Junho Kim, Jung-Woo Ha, and Jinwoo Shin. Generating videos with dynamics-aware implicit generative adversarial networks. arXiv preprint arXiv:2202.10571, 2022. 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 680, + 296, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 680, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 56, + 680, + 296, + 714 + ], + "type": "text", + "content": "[38] Zangwei Zheng, Xiangyu Peng, Tianji Yang, Chenhui Shen, Shenggui Li, Hongxin Liu, Yukun Zhou, Tianyi Li, and Yang You. Open-sora: Democratizing efficient video production" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 334, + 72, + 555, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 334, + 72, + 555, + 95 + ], + "spans": [ + { + "bbox": [ + 334, + 72, + 555, + 95 + ], + "type": "text", + "content": "for all. https://github.com/hpcaitech/Open-Sora, 2024. 2, 4, 5, 7, 8, 1, 12, 16, 21, 22" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "11015" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/SpecTRe-GS_ Modeling Highly Specular Surfaces with Reflected Nearby Objects by Tracing Rays in 3D Gaussian Splatting/05871852-d282-4279-bc18-a4b518c95c0a_content_list.json b/2025/SpecTRe-GS_ Modeling Highly Specular Surfaces with Reflected Nearby Objects by Tracing Rays in 3D Gaussian Splatting/05871852-d282-4279-bc18-a4b518c95c0a_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..2223b94192a827da31f9c8fbe881e98f598953cc --- /dev/null +++ b/2025/SpecTRe-GS_ Modeling Highly Specular Surfaces with Reflected Nearby Objects by Tracing Rays in 3D Gaussian Splatting/05871852-d282-4279-bc18-a4b518c95c0a_content_list.json @@ -0,0 +1,1598 @@ +[ + { + "type": "text", + "text": "SpecTRe-GS: Modeling Highly Specular Surfaces with Reflected Nearby Objects by Tracing Rays in 3D Gaussian Splatting", + "text_level": 1, + "bbox": [ + 91, + 128, + 906, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jiajun Tang $^{1,2}$", + "bbox": [ + 215, + 202, + 328, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Youyu Chen", + "bbox": [ + 173, + 220, + 281, + 238 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Fan Fei $^{1,2}$", + "bbox": [ + 348, + 203, + 428, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Binxiao Huang", + "bbox": [ + 303, + 220, + 431, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zhihao Li3", + "bbox": [ + 447, + 203, + 537, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zhenyu Chen", + "bbox": [ + 452, + 220, + 568, + 238 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xiao Tang3", + "bbox": [ + 558, + 203, + 650, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xiaofei Wu3", + "bbox": [ + 594, + 220, + 696, + 237 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "lyong Liu3", + "bbox": [ + 697, + 203, + 779, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Boxin Shi $^{1,2\\#}$", + "bbox": [ + 718, + 220, + 821, + 237 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ State Key Laboratory for Multimedia Information Processing, School of Computer Science, Peking University", + "bbox": [ + 133, + 239, + 862, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ National Engineering Research Center of Visual Technology, School of Computer Science, Peking University", + "bbox": [ + 133, + 256, + 859, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{3}$ Huawei Noah's Ark Lab $^{4}$ Harbin Institute of Technology $^{5}$ University of Hong Kong", + "bbox": [ + 202, + 268, + 790, + 284 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/212d433fd9ee9f5cce3729380e35988aa00901f5e7a9731a1b9ac80863b7f2f7.jpg", + "image_caption": [ + "Figure 1. Our SpecTRe-GS effectively captures the appearance of reflectors with highly specular surfaces, even in the presence of nearby objects. In the real-world scene shown above, it performs favorably against previous methods [13, 15, 30, 39, 42], both visually (as the highlighted reflections within the insets) and quantitatively (with PSNR values within the reflective region annotated in the corners)." + ], + "image_footnote": [], + "bbox": [ + 93, + 299, + 903, + 460 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 517, + 326, + 532 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3D Gaussian Splatting (3DGS), a recently emerged multiview 3D reconstruction technique, has shown significant advantages in real-time rendering and explicit editing. However, 3DGS encounters challenges in the accurate modeling of both high-frequency view-dependent appearances and global illumination effects, including inter-reflection. This paper introduces SpecTRe-GS, which addresses these challenges and models highly Specular surfaces that reflect nearby objects through Tracing Rays in 3D Gaussian Splatting. SpecTRe-GS separately models reflections from highly specular and rough surfaces to leverage the distinctions between their reflective properties and integrates an efficient ray tracer within the 3DGS framework for querying secondary rays, thus achieving fast and accurate rendering. Also, it incorporates normal prior guidance and joint geometry optimization at various stages of the training process to enhance geometry reconstruction for undistorted reflections. Experiments on both synthetic and real-world scenes demonstrate the superiority of SpecTRe-GS compared to existing 3DGS-based methods in capturing highly specular inter-reflections and also showcase its editing applications.", + "bbox": [ + 88, + 547, + 485, + 867 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 514, + 517, + 643, + 532 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, 3D Gaussian Splatting (3DGS) [15] has emerged as an innovative paradigm for multi-view 3D reconstruction. Unlike its predecessor, neural radiance fields (NeRF) [23], 3DGS adopts an explicit point-based scene representation with tile-based rasterization rather than relying on ray casting through an implicit neural field for rendering. This transition endows 3DGS with significant advantages in terms of real-time rendering speed and allows for convenient post-reconstruction editing. Nevertheless, challenges arise when the scene to be reconstructed includes highly specular (i.e., mirror-like) surfaces which exhibit strong and high-frequency view-dependent appearances. The original 3DGS struggles to accurately model such high-frequency reflections, as it utilizes per-Gaussian third-order spherical harmonics (SH) to represent view-dependent appearances. This approach is restricted to low-frequency radiance variations in angular space and is susceptible to overfitting when the training views that observe the Gaussian are not sufficiently dense [17, 37].", + "bbox": [ + 511, + 547, + 906, + 835 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To address this issue, a natural strategy is to adopt a physically-based shading model and optimize reflective properties for each Gaussian [2, 7, 8, 10, 11, 13, 18, 29, 35, 36, 39, 40, 43, 47], rather than merely recording the out", + "bbox": [ + 511, + 835, + 908, + 897 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "The work was done during an internship in Huawei.", + "bbox": [ + 112, + 875, + 388, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "Corresponding author. E-mail: shiboxin@pku.edu.cn.", + "bbox": [ + 112, + 887, + 447, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "16133", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "going radiance. This strategy enables the computation of reflected radiance by evaluating the physically constrained rendering equation [14], taking into account local scene components including geometry, material properties, and incident lighting, thereby yielding more accurate, realistic, and reliable view-dependent effects. Among the three local scene components, addressing local incident lighting is perhaps the most intricate. Adopting a suboptimal approach to addressing local incident lighting may undermine global illumination effects such as inter-reflections, resulting in an unrealistic scene appearance. Notably, this issue is particularly pronounced in scenes characterized by highly specular surfaces; for instance, the absence of reflections of nearby objects would be especially noticeable. Therefore, it is crucial for methods aimed at modeling such high-frequency view-dependent appearances through the rendering equation to carefully address local incident lighting.", + "bbox": [ + 89, + 90, + 480, + 347 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Unfortunately, effectively acquiring local incident lighting poses a substantial challenge for 3DGS-based methods. The advantages of 3DGS come with a trade-off: The exclusive reliance on the tile-based rasterizer limits its ability to efficiently render secondary rays emanating from vastly different origins towards arbitrary directions, which is essential for querying local incident lighting. As a compromise, a majority of methods neglect secondary effects [36, 39, 47] or periodically bake secondary effects including local occlusions [8, 40] and incident radiance [18, 29] as low-order SH stored in Gaussians or probes. However, low-order SH can only capture low-frequency variations in incident lighting, and the requirement of baking not only incurs additional storage costs but also hinders post-reconstruction editing, such as inserting or deleting nearby objects or relighting. Employing screen space ray tracing (SSRT) [2] produces high-frequency but incomplete inter-reflections as SSRT struggles to trace out-of-screen and occluded rays.", + "bbox": [ + 89, + 347, + 482, + 619 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We propose SpecTRe-GS to model highly Specular surfaces with reflected nearby objects through Tracing Rays in the 3D space within the Gaussian Splatting framework, overcoming the aforementioned challenges. We first made the observation that, although highly specular surfaces exhibit reflections that can be extremely high-frequency and thus hard to capture using SH color or neural network, they are also favorable for modeling in the sense that the specular lobe of their bidirectional reflectance distribution functions (BRDF) are highly concentrated. This fact facilitates the evaluation of the rendering equation [14] on these surfaces because tracing only a few secondary rays already suffices, in contrast to rougher surfaces with more scattered specular lobes. To make use of this fact, we separately model highly specular and rough surfaces, ray-tracing the appearance of the former while retaining the low-order SH color to well-approximate the latter, significantly reducing the amount of secondary rays to trace. To further speed up the time", + "bbox": [ + 89, + 619, + 482, + 891 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "consuming ray tracing process, we implement a ray tracer within the 3DGS framework, inspired by the recent progresses [24, 42]. We also propose a training strategy to enhance the scene geometry reconstruction, the quality of which directly decides the quality of the reconstructed reflection. To summarize, our contributions are as follows:", + "bbox": [ + 511, + 92, + 903, + 180 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- proposing a shading model tailored for indirect mirror reflections combined with an efficient ray tracer for fast and accurate rendering;", + "- designing a training strategy to enhance geometry reconstruction for precise incident direction computing; and", + "- achieving state-of-the-art novel view synthesis results on scenes with highly specular reflections (shown in Fig. 1)." + ], + "bbox": [ + 511, + 181, + 903, + 287 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 301, + 653, + 316 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Improving view-dependent appearances in 3DGS. The original 3DGS [15] uses low-order SH color that captures only low-frequency angular variations, a limitation that many follow-up methods were trying to address (see Tab. 1). While a few methods [30, 35] (row 2) adopt neural networks to decode local features, the majority [2, 8, 10, 13, 18, 29, 36, 39, 40, 47] (rows 3-8) employ the physically-based rendering equation [14]. This equation offers the most accurate approach for modeling view-dependent appearances but requires explicit modeling of surface normals, BRDFs, and the most demanding component - incident lighting. The rendering equation can accommodate all-frequency BRDFs; thus, the frequency of the reflected radiance is now bounded by that of the modeled incident lighting. While an environment map suffices to model direct lighting, it is significantly more challenging to model indirect lighting responsible for global illumination effects including occlusions and inter-reflections. Some methods circumvent this issue by ignoring secondary effects [36] (row 3) or by using a residual color [13] (row 4). For methods that explicitly model indirect lighting, a common strategy is to periodically bake occlusions [8] or inter-reflections [18, 29] into low-order SH, thereby limiting their incident radiance to low-frequency variations (rows 5-7). GI-GS [2] (row 8) uses screen space ray tracing to simulate high-frequency inter-reflections; however, it only models those whose sources are directly visible in the viewpoint. There are also specialized methods for plane mirrors that reflect the point cloud or camera about the plane [19, 22]. Our SpecTRe-GS traces the incident lighting in 3D space, thus is able to render high-frequency and complete inter-reflections for highly specular surfaces with convex shapes.", + "bbox": [ + 511, + 325, + 903, + 808 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Ray-based 3D Gaussian rendering. It is noteworthy that tracing secondary rays for incoming local lighting has almost become a common practice in recent NeRF-based methods that explicitly model specular reflections or interreflections [9, 20, 32, 33, 45]. However, it is challenging to achieve this in 3DGS-based frameworks. 3DGRT [24]", + "bbox": [ + 511, + 810, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "16134", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "table", + "img_path": "images/aa3dc18f044511a489a76805f8b3704d98e953d3d8e96918ba36c8d914e69328.jpg", + "table_caption": [ + "Table 1. We summarize works improving the view-dependent appearances upon 3DGS [15] based on: How do they model the view-dependent appearances (i.e., the outgoing radiance) and the supported frequency in angular space; how do they accommodate the global illumination effects (i.e., the local incident lighting) including occlusions and inter-reflections and the supported frequency in angular space; and whether is it feasible to edit the scene appearance under (in)direct lighting after reconstruction without any re-baking." + ], + "table_footnote": [], + "table_body": "
MethodView-dependent AppearanceGlobal Illumination EffectEditable Appearance
ModelingFrequencyOcclusionInter-reflectionFrequency
(1) 3DGS [15]Low-order SHLowJointly baked in outgoing SH colorLowNone
(2) 3iGS [30]; [35]Neural networkNet.-boundedJointly baked in the neural featureNet.-boundedNone
(3) DeferredGS [36]IgnoredNone
(4) GShader [13]; [39]Arbitrary, thus bounded by the frequency of incident lightingJointly baked in outgoing SH residualDirect
(5) GS-ID [8]; [40]Physically-based rendering equation [14]Baked as SHIgnoredLow
(6) GS-IR [18]; [10]Baked as SH
(7) GIR [29]Ray-traced in 3D spaceBaked as SH
(8) GI-GS [2]Ray-traced in screen space, incompleteHigh(In)direct
(9) OursRay-traced in 3D spaceHigh(In)direct
", + "bbox": [ + 94, + 148, + 901, + 345 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "first demonstrates efficient ray tracing in 3D Gaussians, enabling ray-based lighting effects such as non-projective views. RayGauss [1] and EVER [21] employ ray sampling in 3D Gaussians, mitigating common artifacts in 3DGS, such as aliasing [41], popping artifacts [27] and projection errors [12]. Yang et al. [46] adopt path tracing to simulate global illumination with Gaussian primitives. Meanwhile, Jorge et al. [4] utilizes ray sampling to simulate volumetric scattering and emissive media. Our SpecTRe-GS integrates ray tracing techniques to evaluate accurate incident radiance in the 3DGS framework at an affordable cost.", + "bbox": [ + 88, + 353, + 480, + 518 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Proposed Method", + "text_level": 1, + "bbox": [ + 89, + 536, + 264, + 553 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our method is specifically designed to accurately model the high-frequency reflections of highly specular surfaces. We build upon the recent advancements in 3DGS (Sec. 3.1) and adapt the reflection model to efficiently handle scenes comprising both highly specular objects and rougher objects (Sec. 3.2). To precisely and reliably query the incident radiance necessary for rendering both direct and indirect highly specular reflections, we incorporate ray tracing within the 3DGS framework (Sec. 3.3). We employ a dedicated training strategy to enhance geometry reconstruction, which is crucial for rendering undistorted and realistic reflections (Sec. 3.4). Our pipeline is outlined in Fig. 2.", + "bbox": [ + 88, + 561, + 482, + 744 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminary", + "text_level": 1, + "bbox": [ + 89, + 756, + 220, + 772 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The seminal work of 3DGS [15] represents the scene as a translucent Gaussian point cloud, where each 3D Gaussian is described with central point $\\pmb{\\mu} \\in \\mathbb{R}^3$ , scales $\\pmb{s} \\in \\mathbb{R}^3$ , rotation quaternion $\\pmb{q} \\in \\mathbb{R}^4$ , and opacity $\\sigma \\in [0,1]$ . The distribution of each Gaussian in 3D space is given by $G(\\pmb{x}) = e^{-\\frac{1}{2} (\\pmb{x} - \\pmb{\\mu})^\\top \\pmb{\\Sigma}^{-1}(\\pmb{x} - \\pmb{\\mu})}$ , where the covariance $\\pmb{\\Sigma} = \\mathbf{RSS}^\\top \\mathbf{R}^\\top$ is compactly parameterized as $\\pmb{s}$ and $\\pmb{q}$ . By performing local affine transformation [48], 3D Gauss", + "bbox": [ + 89, + 777, + 482, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "sians are projected as 2D Gaussians on the image plane for rasterization. Then for each ray corresponding to an image pixel, the contributing 2D Gaussians are accumulated from front to back for volumetric rendering:", + "bbox": [ + 511, + 353, + 906, + 412 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nI \\left(\\boldsymbol {\\omega} _ {\\mathrm {o}}\\right) = \\sum_ {i} \\boldsymbol {c} _ {i} \\left(\\boldsymbol {\\omega} _ {\\mathrm {o}}\\right) \\alpha_ {i} \\prod_ {j = 1} ^ {i - 1} \\left(1 - \\alpha_ {j}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 578, + 417, + 906, + 438 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\alpha_{i} = \\sigma_{i}G_{i}$ and the view-dependent color $c_{i}(\\omega_{\\mathrm{o}})\\in$ $\\mathbb{R}^3$ is modeled by third-order SH coefficients $\\varphi_{i}\\in \\mathbb{R}^{48}$", + "bbox": [ + 511, + 441, + 905, + 470 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Besides the view-dependent color $c_{i}$ , many other opt. mizable attributes including surface normal, diffuse albedo, or occlusion information, can be attached to each Gaussian. These attributes can be used locally in each Gaussian, or be accumulated to a map in the same way as Eq. (1) for later applications such as deferred shading [39].", + "bbox": [ + 511, + 472, + 905, + 564 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Highly Specular Reflection Modeling", + "text_level": 1, + "bbox": [ + 511, + 575, + 834, + 593 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To model high-frequency reflections, we follow the convention to incorporate the physically-based rendering equation [14]. As commonly observed phenomena in real-world life, the inter-reflections from the nearby objects on highly specular surfaces should be modeled for a photorealistic scene appearance. The common practice with distant environment maps [13] ignores such inter-reflections and does not suffice for this task, thus we employ raytracing (detailed in Sec. 3.3) in 3D space for a complete and high-frequency inter-reflection. However, we found that performing ray tracing on the entire scene, including highly specular surfaces and rougher surfaces, is highly time-consuming. Evaluating the rendering equation [14] on rougher surfaces typically needs larger samples of the incident lighting even with multiple importance sampling [31], thus heavily increases the computational cost. Fortunately, previous works [15] have proven that rough surfaces can be well-captured by directly optimizing the outgoing color. Therefore, we propose the following reflection model that treats highly specular surfaces and rougher surfaces sepa", + "bbox": [ + 509, + 598, + 906, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "16135", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/90c711e9d2b22af64bed4a863207a89fae0fefe7c5cc400050b466b8a2575f18.jpg", + "image_caption": [ + "Figure 2. Overview of the SpecTRe-GS pipeline. The scene is represented as a Gaussian point cloud, where each Gaussian is associated with geometry and shading attributes used to render the depth map $D$ , normal map $N$ , diffuse reflection $I_{\\mathrm{diff}}$ , specular reflectance $A_{\\mathrm{spec}}$ and rough surface reflection $I_{\\mathrm{rs}}$ through splatting (Sec. 3.1). Highly specular surface reflection $I_{\\mathrm{ss}}$ is formulated as the sum of the diffuse reflection $I_{\\mathrm{diff}}$ and the specular reflection computed from the incident radiance $I_{\\mathrm{i}}$ and specular reflectance $A_{\\mathrm{spec}}$ (Sec. 3.2). The incident radiance $I_{\\mathrm{i}}$ is determined along reflected secondary rays given by the scene geometry $D$ and $N$ , with the indirect component $I_{\\mathrm{ind}}$ and visibility $V_{\\mathrm{i}}$ evaluated via efficient ray tracing within the Gaussian point cloud, and the direct component $I_{\\mathrm{dir}}$ modeled by an estimizable environment map (Sec. 3.3). The scene geometry is enhanced using normal prior guidance, numerical gradients from ray-traced incident radiance, and other training signals to achieve undistorted reflections (Sec. 3.4)." + ], + "image_footnote": [], + "bbox": [ + 89, + 85, + 906, + 344 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "rately to enhance the specular reflections while maintaining the rendering speed. Also, we follow deferred shading [5] to shade for each pixel instead of each Gaussian, as it reduces the computation and gives sharp surface normals [47].", + "bbox": [ + 88, + 474, + 482, + 535 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Modeling highly specular surfaces. To model surfaces that may exhibit both highly specular and diffuse reflections, such as polished marble, we define the reflection $I_{\\mathrm{ss}}$ for these highly specular surfaces at surface point $x$ , observed from view direction $\\omega_{\\mathrm{o}}$ as follows:", + "bbox": [ + 88, + 536, + 482, + 609 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nI _ {\\mathrm {s s}} (\\boldsymbol {x}, \\omega_ {\\mathrm {o}}) = I _ {\\text {d i f f}} (\\boldsymbol {x}) + I _ {\\text {s p e c}} (\\boldsymbol {x}, \\omega_ {\\mathrm {o}}). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 158, + 617, + 482, + 633 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "While the diffuse reflection $I_{\\mathrm{diff}}$ is the view-independent outgoing radiance directly optimized as $d$ in each Gaussian, the specular reflection $I_{\\mathrm{spec}}$ is approximated with an ideal mirror reflection:", + "bbox": [ + 89, + 638, + 483, + 698 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nI _ {\\operatorname {s p e c}} (\\boldsymbol {x}, \\omega_ {\\mathrm {o}}) = A _ {\\operatorname {s p e c}} (\\boldsymbol {x}, \\omega_ {\\mathrm {r}}) I _ {\\mathrm {i}} (\\boldsymbol {x}, \\omega_ {\\mathrm {r}}), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 156, + 705, + 482, + 722 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where the reflected incident direction $\\omega_{\\mathrm{r}}$ is given by $\\omega_{\\mathrm{o}} - 2\\omega_{\\mathrm{o}}^{\\top}\\pmb{n}\\cdot \\pmb{n}$ in which $\\pmb{n}$ denotes the local surface normal, and $A_{\\mathrm{spec}}$ is the specular reflectance considering Fresnel effects through Schlick's approximation [28]:", + "bbox": [ + 89, + 727, + 483, + 787 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nA _ {\\mathrm {s p e c}} (\\pmb {x}, \\pmb {\\omega} _ {\\mathrm {r}}) = F _ {0} (\\pmb {x}) + (1 - F _ {0} (\\pmb {x})) (1 - \\pmb {n} ^ {\\top} \\pmb {\\omega} _ {\\mathrm {r}}) ^ {5}. \\quad (4)\n$$\n", + "text_format": "latex", + "bbox": [ + 102, + 791, + 482, + 811 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here, $F_{0}$ is the specular reflectance at normal incidence optimized as $\\pmb{f_0}$ in each Gaussian and is 3-channel to support metallic materials. The incident radiance $I_{\\mathrm{i}}$ in $\\omega_{\\mathrm{r}}$ originates", + "bbox": [ + 89, + 814, + 483, + 861 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "from both direct and indirect sources:", + "bbox": [ + 511, + 474, + 763, + 489 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "(5)", + "bbox": [ + 880, + 482, + 903, + 496 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nI _ {\\mathrm {i}} (\\boldsymbol {x}, \\omega_ {\\mathrm {r}}) = V _ {\\mathrm {i}} (\\boldsymbol {x}, \\omega_ {\\mathrm {r}}) I _ {\\mathrm {d i r}} (\\omega_ {\\mathrm {r}}) + (1 - V _ {\\mathrm {i}} (\\boldsymbol {x}, \\omega_ {\\mathrm {r}})) I _ {\\mathrm {i n d}} (\\boldsymbol {x}, \\omega_ {\\mathrm {r}}),\n$$\n", + "text_format": "latex", + "bbox": [ + 511, + 498, + 916, + 515 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $V_{\\mathrm{i}} \\in [0,1]$ quantifies the proportion of the incident lighting occluded by the scene itself, $I_{\\mathrm{dir}}$ is the direct incident lighting from a distant environment map, and $I_{\\mathrm{ind}}$ is the indirect incident radiance from the scene.", + "bbox": [ + 511, + 521, + 906, + 580 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Defining depth and surface normal of Gaussians. It is essential to accurately acquire the depth and surface normal of each Gaussian in our method, as they directly determine the origin and direction of the reflection. However, they are ambiguous because the Gaussians are translucent ellipsoids in 3D space. Instead of treating the depth of the projected center and the shortest axis direction facing the camera as the depth and normal of each Gaussian [13, 17, 39], we adopt the formulation in GOF [42] to calculate for each Gaussian the \"intersection point\" along the ray achieving the maximum response, using its depth and normal of the intersection plane. This formulation gives a better and more consistent geometry reconstruction with 3D Gaussians.", + "bbox": [ + 511, + 582, + 908, + 777 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Modeling the whole image. We model the part of the scene containing highly specular surfaces using Eq. (2). For the other part which exhibits only low-frequency view dependencies, we follow GOF [42] to model the appearance $I_{\\mathrm{rs}}$ of the rougher surfaces using per-Gaussian SH coefficients $\\varphi$ . We combine $I_{\\mathrm{ss}}$ and $I_{\\mathrm{rs}}$ using a binary mask $M$ (which can be handily acquired through pre-trained segmentation models such as SAM [16]) specifying the highly specular", + "bbox": [ + 511, + 777, + 908, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "1We omit the distribution and mask-shadowing term in the microfacet model as we assume ideal mirror reflection.", + "bbox": [ + 89, + 875, + 482, + 898 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "16136", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "surfaces within the scene ( $\\odot$ is pixel-wise multiplication):", + "bbox": [ + 89, + 90, + 473, + 106 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nI \\left(\\omega_ {\\mathrm {o}}\\right) = I _ {\\mathrm {s s}} \\left(\\omega_ {\\mathrm {o}}\\right) \\odot M + I _ {\\mathrm {r s}} \\left(\\omega_ {\\mathrm {o}}\\right) \\odot (1 - M). \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 133, + 112, + 482, + 128 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Efficient Ray Tracing for Incident Radiance", + "text_level": 1, + "bbox": [ + 89, + 138, + 464, + 154 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To accurately compute the reflection from highly specular surfaces, it is crucial to obtain the incident radiance at the reflected view direction as modeled by Eq. (5), which incorporates both direct and indirect lighting. However, due to the incoherent origins and directions of these secondary rays, the incident radiance cannot be efficiently rendered using the rasterizer from the original 3DGS [15], as it relies on the locality of the rays, nor can it be computed merely using an environment map, which considers only direct lighting [13]. Inspired by the recent work 3DGRT [24], we implement an efficient ray tracer within the 3DGS framework to compute the radiance of secondary rays, thus achieving high-frequency and accurate reflections.", + "bbox": [ + 89, + 161, + 483, + 354 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Geometry acceleration structure (GAS). For fast scene traversal, we construct a GAS within the OptiX programming interface [25], which uses ray tracing hardware to accelerate ray-primitive intersection tests and is several times faster than BVH [3] in CUDA. Instead of using axis-aligned bounding boxes which inefficiently encapsulate irregular Gaussians and lead to false positives, we adopt stretched icosahedrons as proxy meshes for more efficient ray intersection [24]. The vertex $j$ of the icosahedron approximating Gaussian $i$ is given by:", + "bbox": [ + 89, + 357, + 483, + 508 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {v} _ {i, j} = \\sqrt {2 \\log \\left(\\sigma_ {i} / \\alpha_ {\\min }\\right)} \\mathbf {R} _ {i} ^ {\\top} \\mathbf {S} _ {i} ^ {\\top} \\boldsymbol {v} _ {j} ^ {\\text {u n i t}} + \\boldsymbol {\\mu} _ {i}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 142, + 513, + 480, + 532 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\pmb{v}_j^{\\mathrm{unit}}$ is the vertex of an inscribed icosahedron in a unit sphere. Here, $\\alpha_{\\mathrm{min}}$ serves as a threshold that adaptively scales each icosahedron by dictating the minimal contribution of the Gaussian at the transformed vertices, thereby reducing the computational overhead during intersection and radiance evaluation. We regularly update the GAS after each densification operation [15] and every $n_{\\mathrm{up}}$ step as the Gaussian point cloud changes along the training.", + "bbox": [ + 89, + 537, + 483, + 659 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Ray queries with GAS. To evaluate the incident radiance $I_{\\mathrm{ind}}(\\pmb{x}_0, \\pmb{\\omega}_{\\mathrm{r}})$ , secondary rays $\\pmb{x}(t) = \\pmb{x}_0 + t\\pmb{\\omega}_{\\mathrm{r}}$ are traced against proxy meshes within the GAS. Upon tracing each ray, we first collect the $k$ closest valid hits, each corresponding to a Gaussian with $\\alpha_i \\geq \\alpha_{\\min}$ . The depth of the \"true\" intersection point for each Gaussian is defined as:", + "bbox": [ + 89, + 659, + 483, + 750 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nt _ {\\max , i} = \\frac {\\left(\\boldsymbol {\\mu} _ {i} - \\boldsymbol {x} _ {0}\\right) ^ {\\top} \\boldsymbol {\\Sigma} _ {i} ^ {- 1} \\boldsymbol {\\omega} _ {\\mathrm {r}}}{\\boldsymbol {\\omega} _ {\\mathrm {r}} ^ {\\top} \\boldsymbol {\\Sigma} _ {i} ^ {- 1} \\boldsymbol {\\omega} _ {\\mathrm {r}}}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 184, + 753, + 482, + 789 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "which maximizes the Gaussian's contribution $\\sigma_{i}G_{i}(\\pmb{x}_{0} + t_{\\max,i}\\pmb{\\omega}_{\\mathrm{r}})$ . We then sort the Gaussians by their depth and update the ray depth $D_{\\mathrm{ind}}$ , visibility (i.e., transmittance) $V_{\\mathrm{i}}$ and color $I_{\\mathrm{ind}}$ (using the Gaussians' outgoing SH colors assuming that they model rough surfaces), following Eq. (1). The above process is repeated after advancing the ray origin until no more valid hits remain or the ray transmittance", + "bbox": [ + 89, + 795, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "drops below a threshold. For the transmittance that remains unoccluded by the reconstructed scene itself, we adopt an estimable environment map for calculating the direct incident radiance $I_{\\mathrm{dir}}(\\omega)$ . According to Eq. (5), the final incident radiance $I_{\\mathrm{i}}(\\pmb{x}, \\pmb{\\omega}_{\\mathrm{r}})$ combines direct and indirect components, weighted by visibility $V_{\\mathrm{i}}(\\pmb{x}, \\pmb{\\omega})$ .", + "bbox": [ + 511, + 90, + 903, + 181 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Thanks to the efficient and hardware-accelerated ray tracing techniques elaborated above, as well as our separate modeling of highly specular and rough surfaces, the average rendering time of our method is only $50\\%$ to $100\\%$ slower than that of a pure rasterizer on each of our test scenes. This varies according to the density of objects in the scene and the proportion of highly specular surfaces.", + "bbox": [ + 511, + 181, + 906, + 287 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Training", + "text_level": 1, + "bbox": [ + 513, + 306, + 617, + 321 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We propose a training strategy that especially focuses on enhancing the geometry reconstruction, as unfaithful scene geometry induces inaccurate secondary rays and results in distorted reflections. We optimize our model from an initial sparse point cloud using the following loss terms:", + "bbox": [ + 511, + 327, + 905, + 402 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {\\mathrm {c}} + \\lambda_ {\\mathrm {d}} \\mathcal {L} _ {\\mathrm {d}} + \\lambda_ {\\mathrm {d n}} \\mathcal {L} _ {\\mathrm {d n}} + \\lambda_ {\\mathrm {s m}} \\mathcal {L} _ {\\mathrm {s m}} + \\lambda_ {\\mathrm {n}} \\mathcal {L} _ {\\mathrm {n}} + \\lambda_ {\\mathrm {r}} \\mathcal {L} _ {\\mathrm {r}}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 521, + 409, + 903, + 425 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_{\\mathrm{c}}$ is a color reconstruction loss [15] differentiable to the geometry, $\\mathcal{L}_{\\mathrm{d}}$ and $\\mathcal{L}_{\\mathrm{dn}}$ are the distortion and depth-normal consistency loss used in GOF [42], $\\mathcal{L}_{\\mathrm{sm}}$ is an edge-aware smoothness term for the normal and specular reflectance for the highly specular surfaces, $\\mathcal{L}_{\\mathrm{n}}$ is a normal prior term to employ data-driven clues, and $\\mathcal{L}_{\\mathrm{r}}$ is a residual-suppressing term for progressive learning. This section describes these terms and more details are in the supplement.", + "bbox": [ + 511, + 430, + 905, + 551 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Normal prior guidance. To tackle the challenging problem of reconstructing geometry for highly specular objects, we employ StableNormal [38] to provide monocular normal priors $\\hat{N}$ , offering initial geometric constraints:", + "bbox": [ + 511, + 551, + 903, + 611 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {n}} = 1 / | N | \\sum \\left| \\left| (\\hat {N} - N) \\odot M \\right| \\right| _ {1}. \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 617, + 903, + 638 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Joint geometry optimization. Recognizing potential inaccuracies and inconsistencies across views in the normal prior, we concurrently optimize the scene geometry during training. This is achieved by differentiating the incident radiance $I_{\\mathrm{i}}$ , which relates to the final image through Eq. (2) and Eq. (3), with respect to the scene geometry. Computing the gradient of the direct incident radiance $I_{\\mathrm{dir}}$ is straightforward. To compute the gradient of the indirect incident radiance $I_{\\mathrm{ind}}$ concerning the local surface normal $n$ , we adopt a simple yet effective finite difference method to numerically evaluate the gradient of $I_{\\mathrm{ind}}$ with respect to the reflected direction $\\omega$ that depends on $n$ :", + "bbox": [ + 511, + 643, + 906, + 824 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left. \\frac {\\partial I _ {\\text {i n d}}}{\\partial \\omega^ {i}} \\right| _ {\\omega_ {0}} \\approx \\frac {I _ {\\text {i n d}} \\left(\\boldsymbol {x} , \\omega_ {0} + \\Delta_ {\\omega} ^ {i}\\right) - I _ {\\text {i n d}} \\left(\\boldsymbol {x} , \\omega_ {0} - \\Delta_ {\\omega} ^ {i}\\right)}{2 \\Delta_ {\\omega} ^ {i}}, \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 828, + 903, + 864 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\Delta_{\\omega}^{i}$ denotes the perturbation of the $i$ -th dimension in $\\omega$ . Considering that using the same perturbation for sec-", + "bbox": [ + 511, + 869, + 903, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "16137", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/9bb6cbf2041863355c58631c42f8ac6bbc2e4788f3a3268cf8128cb005cf2305.jpg", + "table_caption": [ + "Table 2. Quantitative comparisons of novel view synthesis results with state-of-the-art methods on both synthetic and real-world scenes. We report the scores of PSNR, SSIM [34], and LPIPS [44] for entire images and within reflective regions. $\\uparrow (\\downarrow)$ means higher (lower) is better. We mark the best and the second best results in each column." + ], + "table_footnote": [], + "table_body": "
DataSynthetic (entire image)Synthetic (reflective)Real-world (entire image)Real-world (reflective)
MethodPSNR↓SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↓SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
3DGS [15]27.710.9100.13121.510.9400.07824.820.8230.22721.310.9650.047
GOF [42]27.970.9170.11721.380.9400.07624.860.8260.21921.250.9630.047
GOF* [42]27.570.9130.12520.670.9360.08524.690.8250.22320.440.9620.051
3iGS [30]28.580.9120.12521.980.9390.07624.370.8090.23321.220.9630.048
GShader [13]26.650.8880.16020.790.9350.08623.930.8100.25520.230.9630.053
GShader* [13]27.140.8940.15321.220.9380.08424.030.8120.25320.360.9640.053
3DGS-DR [39]27.620.8900.17321.570.9380.08524.790.8240.24321.180.9650.050
3DGS-DR* [39]26.970.8830.18221.180.9360.08824.600.8230.25120.380.9640.055
Ours29.410.9220.10923.280.9490.06225.110.8280.21322.770.9670.035
", + "bbox": [ + 94, + 133, + 898, + 287 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ondary rays with different depths $D_{\\mathrm{ind}}$ leads to different amounts of intersection movements, making the numerical gradient less accurate, we use a depth-aware ray perturbation that adjusts to $D_{\\mathrm{ind}}$ for more informative gradients:", + "bbox": [ + 89, + 292, + 482, + 354 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\Delta_ {\\omega} ^ {i} = \\theta_ {\\mathrm {d}} \\max \\left(\\min \\left(D _ {\\text {i n d}} / D _ {0}, \\sigma_ {\\mathrm {d}} ^ {\\max }\\right), \\sigma_ {\\mathrm {d}} ^ {\\min }\\right), \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 122, + 358, + 482, + 376 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\theta_{\\mathrm{d}}$ and $D_0$ define a reference angular perturbation and $\\sigma_{\\mathrm{d}}^{\\mathrm{min}}$ and $\\sigma_{\\mathrm{d}}^{\\mathrm{max}}$ ensure numerical stability. For computational efficiency, we disregard the gradient with respect to the point position $\\pmb{p}$ , as we find reflections are distorted mainly due to normal errors. We detach visibility $V_{\\mathrm{ind}}$ and depth $D_{\\mathrm{ind}}$ because the sparsity of the visibility gradient and abrupt changes in depth damage numerical stability.", + "bbox": [ + 89, + 380, + 482, + 487 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Progressive learning. As training progresses, monocular normal priors may hinder the accurate geometry reconstruction, thus we gradually decay the weight of normal prior guidance $\\lambda_{\\mathrm{n}}$ , letting gradients from physics-based rendering take over the fine-tuning of scene geometry. Besides, for smoother training and consequently better reconstruction within the reflective regions, we replace the hard mask $M$ used in Eq. (6) with a relaxed version $1 - I_{\\mathrm{r}}$ rendered with another per-Gaussian attribute $r \\in [0,1]$ using Eq. (1). We encourage the reflective region to be modeled with highly specular reflections by suppressing the residual within $M$ :", + "bbox": [ + 89, + 487, + 482, + 652 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {r}} = 1 / | M | \\sum \\left| \\left| I _ {\\mathrm {r}} \\odot M \\right| \\right| _ {1}. \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 187, + 657, + 480, + 680 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 691, + 223, + 709 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we compare our method on both synthetic and real-world scenes (Sec. 4.2) against state-of-the-art methods (Sec. 4.3). We also perform ablation studies to validate the key components of our method (Sec. 4.4) and explore applications in scene editing (Sec. 4.5). Please check the supplementary material for more implementation details, data creation details, and experimental results.", + "bbox": [ + 89, + 715, + 482, + 821 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Implementation Details", + "text_level": 1, + "bbox": [ + 89, + 833, + 308, + 849 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We implement our method in PyTorch [26] with CUDA kernels for rasterization based on GOF [42] and custom OptiX programs for ray tracing. We conduct all experiments on a", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "GPU with 24 GB VRAM, 83 shader core TFLOPS, and 191 ray tracing core TFLOPS. We set $\\lambda_{\\mathrm{sm}} = 0.5$ , $\\lambda_{\\mathrm{r}} = 1.0$ , and $\\lambda_{\\mathrm{n}} = 0.5$ , exponentially decaying to 0.0025 from 4k to 10k iterations. We use $r = 0.9$ as the initialization. For other hyperparameters, we use the settings in GOF [42].", + "bbox": [ + 511, + 292, + 906, + 369 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Datasets", + "text_level": 1, + "bbox": [ + 511, + 378, + 614, + 392 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Synthetic scenes. We build 6 synthetic scenes, each containing multiple surrounding objects and a highly specular surface with distinctive properties: HELMET, MARBLETABLE, VASE, POT, TOASTER, and MIRROR. For each scene, we use the Blender Cycles engine [6] to render about 300 images of resolution $1024 \\times 1024$ , of which 2/3 are used for training and the rest are left for testing.", + "bbox": [ + 511, + 400, + 905, + 506 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Real-world scenes. For real-world validation, we capture two scenes: REALBOWL and REALPOT, consisting of 138 and 139 images of resolution $1440 \\times 1080$ respectively, with every eighth image reserved in the test set.", + "bbox": [ + 511, + 507, + 905, + 566 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Comparison", + "text_level": 1, + "bbox": [ + 511, + 575, + 645, + 593 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Comparing methods. We compare our method with 5 open-source state-of-the-art methods: the original 3DGS [15]; GOF [42] which focus on geometry reconstruction with 3D Gaussians; 3iGS [30] which improves view-dependent effects through neural features; and 3DGS-DR [39] and GShader [13] which are adapted for specular reflection employing physically-based rendering (PBR). For a fair comparison, we provide GOF [42], 3DGS-DR [39], and GShader [13] with the normal prior guidance, and indicate these versions by “*” throughout this paper.", + "bbox": [ + 511, + 598, + 906, + 750 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation protocol. We report peak signal-to-noise ratio (PSNR), structural similarity index measure (SSIM) [34], and learned perceptual image patch similarity (LPIPS) [44] measured on both entire images and reflective regions.", + "bbox": [ + 511, + 750, + 905, + 810 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Quantitative evaluation. As shown in Tab. 2, our method outperforms comparing methods on both synthetic and real-world scenes, especially in reflective regions. It is worth noting that using normal prior guidance usually leads to degradation of GOF [42] which relies solely on SH color, as it suppresses the ability to simulate high-frequency appear", + "bbox": [ + 511, + 810, + 905, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "16138", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a6a031519f5cb53855c07d3865df6f09606179703e6a7e213ace45fdb1793c05.jpg", + "image_caption": [ + "Figure 3. Novel view synthesis comparison on synthetic scenes. We evaluate SpecTRe-GS against relevant 3DGS-based methods modeling view-dependent appearances: 3DGS-DR [39], GShader [13], 3iGS [30], GOF [42], and 3DGS [15]. The insets highlight the improvement of SpecTRe-GS in rendering high-frequency specular reflections, showcasing the precise positioning of reflected content and sharp details." + ], + "image_footnote": [], + "bbox": [ + 94, + 88, + 906, + 366 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7baf3e4fd74eccd4a0c84a39c794fa03898ad53ef0e6b00498dcf536a0bc35ee.jpg", + "image_caption": [ + "Figure 4. Novel view synthesis and image decomposition results on synthetic scenes, comparing our SpecTRe-GS with 3DGS-DR [39] and GShader [13]. SpecTRe-GS better distinguishes diffuse and specular components in reflective regions, with the least baked-in view-dependent appearance in diffuse components. The masks of reflective regions are shown as insets in the reference image." + ], + "image_footnote": [], + "bbox": [ + 94, + 414, + 901, + 671 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ances through false geometry. In contrast, the influence of this guidance on 3DGS-DR [39] and GShader [13] is twofold: It facilitates the reconstruction of appearance under direct lighting, but also prevents these methods from imitating specular inter-reflections by faking geometry.", + "bbox": [ + 89, + 722, + 482, + 797 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Qualitative evaluation. In visual comparisons, we show results of GOF [42], GShader* [13], and 3DGS-DR [39] as they are variants with higher qualities. As shown in Fig. 3, our method performs better than the baseline methods for both planar reflectors (rows 1-2) and curved reflectors (row 3), capturing clean, accurate, and high-frequency specular", + "bbox": [ + 89, + 797, + 482, + 888 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "reflections. Our method also surpasses baseline methods in real-world scenes, as shown in Fig. 1. By visualizing the separate components of PBR-based methods in Fig. 4, we find that incident radiance modeling for both direct and indirect sources is vital for physically meaningful scene appearance decomposition. Otherwise, to fit the overall color in training images, indirect incident radiance is heavily baked into the diffuse component (row 1 for 3DGS-DR [39] and row 3 for GShader [13]). The baked-in incident radiance loses high-frequency details and prevents realistic scene editing in a physically-accurate manner. It should be noted", + "bbox": [ + 511, + 722, + 906, + 888 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "16139", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/dbb97ada72e67ea5c45c0be24a591082a042ae59defeb6ae0acd8626bcb09d38.jpg", + "image_caption": [ + "Figure 5. Our reconstruction results on two real-world scenes. SpecTRe-GS accurately reconstructs the geometry of highly specular surfaces alongside the reflection of nearby objects." + ], + "image_footnote": [], + "bbox": [ + 91, + 89, + 480, + 323 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "that 3DGS-DR [39] performs better than GShader [13] in reconstructing the direct part of incident radiance from the environment, showing the superiority of surface-based deferred shading over volumetric rendering in PBR modeling of high-frequency reflections. Our method, with well-optimized geometry and consequent accurate incident radiance modeling, achieves satisfactory novel view synthesis of specular reflection on both synthetic and real-world scenes, as also indicated in Fig. 5.", + "bbox": [ + 89, + 372, + 483, + 508 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 516, + 243, + 532 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "On a subset of our synthetic scenes, we validate the following key component of our method: (1) Normal prior guidance; (2) Joint geometry optimization from incident lighting; (3) ray-traced Indirect incident radiance; (4) Progressive learning; and (5) Depth-aware ray perturbation. The results are shown in Tab. 3. Our full method (Ours) achieves the highest scores, confirming the effectiveness of our method designs. Without the initial normal prior guidance (Ours w/o N.), the optimization process is prone to be trapped into local minima far away from the true underneath geometry of the reflective region. On the other hand, only relying on normal priors without further joint optimization for fine-tuning geometry (Ours w/o J.) also fails in reconstructing accurate geometry for physically-based high-frequency reflections. As discussed above, using only direct incident lighting modeling (Ours w/o I.) leads to baked-in indirect incident lighting and damages the quality of high-frequency components. The progressive learning scheme (Ours w/o P.) and depth-aware ray perturbation (Ours w/o D.) also further helps to faithfully reconstruct the scene.", + "bbox": [ + 89, + 537, + 483, + 840 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Applications", + "text_level": 1, + "bbox": [ + 89, + 848, + 225, + 864 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our Gaussian point cloud representation with explicit geometry, material, and incident lighting facilitates various", + "bbox": [ + 89, + 869, + 482, + 901 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/806bc9376edaa48fa32966e6dc84f446293fd67985c22ccec4bbd4416455bb24.jpg", + "table_caption": [ + "Table 3. Ablation study with variants of our proposed method excluding: Normal prior guidance, Joint geometry optimization, Indirect incident lighting modeling, Progressive learning, and Depth-aware ray perturbation. $\\uparrow (\\downarrow)$ means higher (lower) is better." + ], + "table_footnote": [], + "table_body": "
MethodEntire ImageReflective Region
PSNR↓SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
Ours w/o N.26.700.8850.14919.800.9130.094
Ours w/o J.26.370.8780.15419.520.9070.100
Ours w/o I.28.490.9050.13022.190.9350.074
Ours w/o P.29.260.9100.11823.160.9400.062
Ours w/o D.29.750.9140.11423.880.9440.056
Ours29.900.9140.11224.050.9440.056
", + "bbox": [ + 516, + 147, + 903, + 266 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/194d38599ece68cb60acf280c4538d4f287a2ddc648f9cf9cfff1183081baa16.jpg", + "image_caption": [ + "Figure 6. Our explicit representation of geometry, material, and incident lighting in the Gaussian point cloud enables various editing applications. Insets have been brightened for clarity." + ], + "image_footnote": [], + "bbox": [ + 517, + 268, + 903, + 431 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "editing applications. As illustrated in Fig. 6, specular objects of interest can be faithfully captured as assets in the presence of unwanted reflections from nearby objects, enabling further usages such as relighting (column 1) and material editing (column 4). We can also manipulate the scene geometry such as deletion (column 2) and insertion (column 3), with specular inter-reflections adjusted accordingly.", + "bbox": [ + 511, + 484, + 906, + 592 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 604, + 633, + 619 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We introduce SpecTRe-GS to model highly specular surfaces with inter-reflections, distinguished by its separate modeling of specular and rough surfaces, the efficient ray tracer, and a geometry-enhancing training strategy for fast and accurate specular inter-reflection rendering.", + "bbox": [ + 511, + 630, + 905, + 705 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations. We only model specular inter-reflections from rough surfaces to avoid multi-bounce path tracing, and we only trace one secondary ray for each pixel in the reflective region, approximating the specular reflection with an ideal mirror. Also, the monocular normal prior may fail in the existence of strong semantic ambiguity (such as mirrors with hidden borders), making our geometry fail to initialize.", + "bbox": [ + 511, + 705, + 905, + 811 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 511, + 818, + 671, + 834 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work is supported by the National Natural Science Foundation of China under Grant No. 62136001. PKU-affiliated authors thank openbayes.com for providing computing resource.", + "bbox": [ + 511, + 838, + 905, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "16140", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Hugo Blanc, Jean-Emmanuel Deschaud, and Alexis Paljic. RayGauss: Volumetric Gaussian-based ray casting for photorealistic novel view synthesis. In Proc. of IEEE Winter Conference on Applications of Computer Vision (WACV), 2024. 3", + "[2] Hongze Chen, Zehong Lin, and Jun Zhang. GI-GS: Global illumination decomposition on Gaussian splatting for inverse rendering. In Proc. of International Conference on Learning Representations (ICLR), 2025. 1, 2, 3", + "[3] James H. Clark. Hierarchical geometric models for visible surface algorithms. Communications of the ACM (CACM), 19(10):547-554, 1976. 5", + "[4] Jorge Condor, Sebastien Speierer, Lukas Bode, Aljaz Bozic, Simon Green, Piotr Didyk, and Adrian Jarabo. Don't splat your Gaussians: Volumetric ray-traced primitives for modeling and rendering scattering and emissive media. ACM Transactions on Graphics (TOG), 44(1):10:1-10:17, 2025. 3", + "[5] Michael Deering, Stephanie Winner, Bic Schediwy, Chris Duffy, and Neil Hunt. The triangle processor and normal vector Shader: a VLSI system for high performance graphics. In Proc. of the ACM SIGGRAPH Conference and Exhibition On Computer Graphics and Interactive Techniques (SIGGRAPH), 1988. 4", + "[6] Cycles Developers. Cycles: Open source production rendering, 2024. https://www.cycles-renderer.org/. Accessed: 2024-11-06. 6", + "[7] Jan-Niklas Dihlmann, Arjun Majumdar, Andreas Engelhardt, Raphael Braun, and Hendrik P. A. Lensch. Subsurface scattering for 3D Gaussian splattering. In Proc. of Neural Information Processing Systems (NeurIPS), 2024. 1", + "[8] Kang Du, Zhihao Liang, and Zeyu Wang. GS-ID: Illumination decomposition on Gaussian splattering via diffusion prior and parametric light source optimization. In arXiv preprint arXiv:2408.08524, 2024. 1, 2, 3", + "[9] Fan Fei, Jiajun Tang, Ping Tan, and Boxin Shi. VMINer: Versatile multi-view inverse rendering with near- and farfield light sources. In Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2", + "[10] Jian Gao, Chun Gu, Youtian Lin, Hao Zhu, Xun Cao, Li Zhang, and Yao Yao. Relightable 3D Gaussian: Real-time point cloud relighting with BRDF decomposition and ray tracing. In Proc. of European Conference on Computer Vision (ECCV), 2024. 1, 2, 3", + "[11] Yijia Guo, Yuanxi Bai, Liwen Hu, Ziyi Guo, Mianzhi Liu, Yu Cai, Tiejun Huang, and Lei Ma. PRTGS: Precomputed radiance transfer of Gaussian splats for real-time high-quality relighting. In Proc. of ACM International Conference on Multimedia (ACM MM), 2024. 1", + "[12] Letian Huang, Jiayang Bai, Jie Guo, Yuanqi Li, and Yanwen Guo. On the error analysis of 3D Gaussian splattering and an optimal projection strategy. In Proc. of European Conference on Computer Vision (ECCV), 2024. 3", + "[13] Yingwenqi Jiang, Jiadong Tu, Yuan Liu, Xifeng Gao, Xiaoxiao Long, Wenping Wang, and Yuexin Ma. GaussianShader:" + ], + "bbox": [ + 91, + 114, + 483, + 902 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "3D Gaussian splattering with shading functions for reflective surfaces. In Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 1, 2, 3, 4, 5, 6, 7, 8", + "[14] James T. Kajiya. The rendering equation. In Proc. of the ACM SIGGRAPH Conference and Exhibition On Computer Graphics and Interactive Techniques (SIGGRAPH), 1986. 2, 3", + "[15] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3D Gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics (TOG), 42(4):139:1-139:14, 2023. 1, 2, 3, 5, 6, 7", + "[16] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dollar, and Ross B. Girshick. Segment anything. In Proc. of IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 4", + "[17] Jiahe Li, Jiawei Zhang, Xiao Bai, Jin Zheng, Xin Ning, Jun Zhou, and Lin Gu. DNGaussian: Optimizing sparse-view 3D Gaussian radiance fields with global-local depth normalization. In Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 1, 4", + "[18] Zhihao Liang, Qi Zhang, Ying Feng, Ying Shan, and Kui Jia. GS-IR: 3D Gaussian splatting for inverse rendering. In Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 1, 2, 3", + "[19] Jiayue Liu, Xiao Tang, Freeman Cheng, Roy Yang, Zhihao Li, Jianzhuang Liu, Yi Huang, Jiaqi Lin, Shiyong Liu, Xiaofei Wu, Songcen Xu, and Chun Yuan. MirrorGaussian: Reflecting 3D Gaussians for reconstructing mirror reflections. In Proc. of European Conference on Computer Vision (ECCV), 2024. 2", + "[20] Yuan Liu, Peng Wang, Cheng Lin, Xiaoxiao Long, Jiepeng Wang, Lingjie Liu, Taku Komura, and Wenping Wang. NeRO: Neural geometry and BRDF reconstruction of reflective objects from multiview images. ACM Transactions on Graphics (TOG), 42(4):114:1-114:22, 2023. 2", + "[21] Alexander Mai, Peter Hedman, George Kopanas, Dor Verbin, David Futschik, Qiangeng Xu, Falko Kuester, Jonathan T. Barron, and Yinda Zhang. EVER: Exact volumetric ellipsoid rendering for real-time view synthesis. In arXiv preprint arXiv:2410.01804, 2024. 3", + "[22] Jiarui Meng, Haijie Li, Yanmin Wu, Qiankun Gao, Shuzhou Yang, Jian Zhang, and Siwei Ma. Mirror-3DGS: Incorporating mirror reflections into 3D Gaussian splattering. In Proc. of IEEE International Conference on Visual Communications and Image Processing, 2024. 2", + "[23] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing scenes as neural radiance fields for view synthesis. In Proc. of European Conference on Computer Vision (ECCV), 2020. 1", + "[24] Nicolas Moënne-Loccoz, Ashkan Mirzaei, Or Perel, Ricardo de Lutio, Janick Martinez Esturo, Gavriel State, Sanja Fidler, Nicholas Sharp, and Zan Gojcic. 3D Gaussian ray tracing: Fast tracing of particle scenes. ACM Transactions on Graphics (TOG), 43(6):232:1-232:19, 2024. 2, 5" + ], + "bbox": [ + 516, + 92, + 906, + 902 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "16141", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[25] Steven G. Parker, James Bigler, Andreas Dietrich, Heiko Friedrich, Jared Hoberock, David P. Luebke, David K. McAllister, Morgan McGuire, R. Keith Morley, Austin Robison, and Martin Stich. OptiX: a general purpose ray tracing engine. In ACM Transactions on Graphics (Proc. of ACM SIGGRAPH), 2010. 5", + "[26] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Köpf, Edward Z. Yang, Zachary DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. PyTorch: An imperative style, high-performance deep learning library. In Proc. of Neural Information Processing Systems (NeurIPS), 2019. 6", + "[27] Lukas Radl, Michael Steiner, Mathias Parger, Alexander Weinrauch, Bernhard Kerbl, and Markus Steinberger. StopThePop: Sorted Gaussian splatting for view-consistent real-time rendering. ACM Transactions on Graphics (TOG), 43(4):64:1-64:17, 2024. 3", + "[28] Christophe Schlick. An inexpensive BRDF model for physically-based rendering. Computer Graphics Forum (CGF), 13(3):233-246, 1994. 4", + "[29] Yahao Shi, Yanmin Wu, Chenming Wu, Xing Liu, Chen Zhao, Haocheng Feng, Jingtuo Liu, Liangjun Zhang, Jian Zhang, Bin Zhou, Errui Ding, and Jingdong Wang. GIR: 3D Gaussian inverse rendering for reightable scene factorization. In arXiv preprint arXiv:2312.05133, 2023. 1, 2, 3", + "[30] Zhe Jun Tang and Tat-Jen Cham. 3iGS: Factorised tensorial illumination for 3D Gaussian splatting. In Proc. of European Conference on Computer Vision (ECCV), 2024. 1, 2, 3, 6, 7", + "[31] Eric Veach and Leonidas J. Guibas. Optimally combining sampling techniques for Monte Carlo rendering. In Proc. of the ACM SIGGRAPH Conference and Exhibition On Computer Graphics and Interactive Techniques (SIGGRAPH), 1995. 3", + "[32] Dor Verbin, Peter Hedman, Ben Mildenhall, Todd Zickler, Jonathan T. Barron, and Pratul P. Srinivasan. Ref-NeRF: Structured view-dependent appearance for neural radiance fields. In Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2", + "[33] Dor Verbin, Pratul P. Srinivasan, Peter Hedman, Ben Mildenhall, Benjamin Attal, Richard Szeliski, and Jonathan T. Barron. NeRF-Casting: Improved view-dependent appearance with consistent reflections. In Proc. of the ACM SIGGRAPH Conference and Exhibition on Computer Graphics and Interactive Techniques in Asia (SIGGRAPH Asia), 2024. 2", + "[34] Zhou Wang, A.C. Bovik, H.R. Sheikh, and E.P. Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing (TIP), 13(4):600-612, 2004. 6", + "[35] Zhiru Wang, Shiyun Xie, Chengwei Pan, and Guoping Wang. SpecGaussian with latent features: A high-quality modeling of the view-dependent appearance for 3D Gaussian splatting. In Proc. of ACM International Conference on Multimedia (ACM MM), 2024. 1, 2, 3", + "[36] Tong Wu, Jia-Mu Sun, Yu-Kun Lai, Yuewen Ma, Leif Kobbelt, and Lin Gao. DeferredGS: Decoupled and editable" + ], + "bbox": [ + 91, + 92, + 480, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Gaussian splatting with deferred shading. In arXiv preprint arXiv:2404.09412, 2024. 1, 2, 3", + "[37] Wangze Xu, Huachen Gao, Shihe Shen, Rui Peng, Jianbo Jiao, and Ronggang Wang. MVPGS: Excavating multi-view priors for Gaussian splatting from sparse input views. In Proc. of European Conference on Computer Vision (ECCV), 2024. 1", + "[38] Chongjie Ye, Lingteng Qiu, Xiaodong Gu, Qi Zuo, Yushuang Wu, Zilong Dong, Liefeng Bo, Yuliang Xiu, and Xiaoguang Han. StableNormal: Reducing diffusion variance for stable and sharp normal. In Proc. of the ACM SIGGRAPH Conference and Exhibition on Computer Graphics and Interactive Techniques in Asia (SIGGRAPH Asia), 2024. 5", + "[39] Keyang Ye, Qiming Hou, and Kun Zhou. 3D Gaussian splating with deferred reflection. In Proc. of the ACM SIGGRAPH Conference and Exhibition On Computer Graphics and Interactive Techniques (SIGGRAPH), 2024. 1, 2, 3, 4, 6, 7, 8", + "[40] Keyang Ye, Qiming Hou, and Kun Zhou. Progressive radiance distillation for inverse rendering with Gaussian splatting. In arXiv preprint arXiv:2408.07595, 2024. 1, 2, 3", + "[41] Zehao Yu, Anpei Chen, Binbin Huang, Torsten Sattler, and Andreas Geiger. Mip-Splatting: Alias-free 3D Gaussian splatting. In Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 3", + "[42] Zehao Yu, Torsten Sattler, and Andreas Geiger. Gaussian opacity fields: Efficient adaptive surface reconstruction in unbounded scenes. ACM Transactions on Graphics (TOG), 43(6):271:1-271:13, 2024. 1, 2, 4, 5, 6, 7", + "[43] Libo Zhang, Yuxuan Han, Wenbin Lin, Jingwang Ling, and Feng Xu. PRTGaussian: Efficient relighting using 3D Gaussians with precomputed radiance transfer. In Asia Pacific Signal and Information Processing Association Annual Summit and Conference (APSIPA ASC), 2024. 1", + "[44] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 6", + "[45] Yuanqing Zhang, Jiaming Sun, Xingyi He, Huan Fu, Rongfei Jia, and Xiaowei Zhou. Modeling indirect illumination for inverse rendering. In Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2", + "[46] Yang Zhou, Songyin Wu, and Ling-Qi Yan. Unified Gaussian primitives for scene representation and rendering. In arXiv preprint arXiv:2406.09733, 2024. 3", + "[47] Zuo-Liang Zhu, Beibei Wang, and Jian Yang. GS-ROR: 3D Gaussian splatting for reflective object relighting via SDF priors. In arXiv preprint arXiv:2406.18544, 2024. 1, 2, 4", + "[48] Matthias Zwicker, Hanspeter Pfister, Jeroen van Baar, and Markus H. Gross. EWA volume splatting. In Proc. of IEEE Visualization Conference (VIS), 2001. 3" + ], + "bbox": [ + 516, + 92, + 903, + 825 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "16142", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/SpecTRe-GS_ Modeling Highly Specular Surfaces with Reflected Nearby Objects by Tracing Rays in 3D Gaussian Splatting/05871852-d282-4279-bc18-a4b518c95c0a_model.json b/2025/SpecTRe-GS_ Modeling Highly Specular Surfaces with Reflected Nearby Objects by Tracing Rays in 3D Gaussian Splatting/05871852-d282-4279-bc18-a4b518c95c0a_model.json new file mode 100644 index 0000000000000000000000000000000000000000..f0df16b5a4a1a4c12bc485b777da3d3081aa1268 --- /dev/null +++ b/2025/SpecTRe-GS_ Modeling Highly Specular Surfaces with Reflected Nearby Objects by Tracing Rays in 3D Gaussian Splatting/05871852-d282-4279-bc18-a4b518c95c0a_model.json @@ -0,0 +1,2178 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.13, + 0.908, + 0.178 + ], + "angle": 0, + "content": "SpecTRe-GS: Modeling Highly Specular Surfaces with Reflected Nearby Objects by Tracing Rays in 3D Gaussian Splatting" + }, + { + "type": "text", + "bbox": [ + 0.216, + 0.203, + 0.329, + 0.222 + ], + "angle": 0, + "content": "Jiajun Tang\\(^{1,2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.174, + 0.222, + 0.282, + 0.239 + ], + "angle": 0, + "content": "Youyu Chen" + }, + { + "type": "text", + "bbox": [ + 0.349, + 0.204, + 0.429, + 0.222 + ], + "angle": 0, + "content": "Fan Fei\\(^{1,2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.304, + 0.222, + 0.432, + 0.24 + ], + "angle": 0, + "content": "Binxiao Huang" + }, + { + "type": "text", + "bbox": [ + 0.449, + 0.204, + 0.538, + 0.222 + ], + "angle": 0, + "content": "Zhihao Li3" + }, + { + "type": "text", + "bbox": [ + 0.453, + 0.222, + 0.569, + 0.239 + ], + "angle": 0, + "content": "Zhenyu Chen" + }, + { + "type": "text", + "bbox": [ + 0.56, + 0.204, + 0.651, + 0.222 + ], + "angle": 0, + "content": "Xiao Tang3" + }, + { + "type": "text", + "bbox": [ + 0.596, + 0.222, + 0.697, + 0.238 + ], + "angle": 0, + "content": "Xiaofei Wu3" + }, + { + "type": "text", + "bbox": [ + 0.699, + 0.204, + 0.781, + 0.222 + ], + "angle": 0, + "content": "lyong Liu3" + }, + { + "type": "text", + "bbox": [ + 0.72, + 0.222, + 0.823, + 0.238 + ], + "angle": 0, + "content": "Boxin Shi\\(^{1,2\\#}\\)" + }, + { + "type": "text", + "bbox": [ + 0.135, + 0.241, + 0.864, + 0.257 + ], + "angle": 0, + "content": "\\(^{1}\\)State Key Laboratory for Multimedia Information Processing, School of Computer Science, Peking University" + }, + { + "type": "text", + "bbox": [ + 0.135, + 0.257, + 0.86, + 0.27 + ], + "angle": 0, + "content": "\\(^{2}\\)National Engineering Research Center of Visual Technology, School of Computer Science, Peking University" + }, + { + "type": "text", + "bbox": [ + 0.204, + 0.269, + 0.792, + 0.285 + ], + "angle": 0, + "content": "\\(^{3}\\)Huawei Noah's Ark Lab \\(^{4}\\)Harbin Institute of Technology \\(^{5}\\)University of Hong Kong" + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.3, + 0.905, + 0.462 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.464, + 0.908, + 0.507 + ], + "angle": 0, + "content": "Figure 1. Our SpecTRe-GS effectively captures the appearance of reflectors with highly specular surfaces, even in the presence of nearby objects. In the real-world scene shown above, it performs favorably against previous methods [13, 15, 30, 39, 42], both visually (as the highlighted reflections within the insets) and quantitatively (with PSNR values within the reflective region annotated in the corners)." + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.518, + 0.327, + 0.533 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.549, + 0.486, + 0.868 + ], + "angle": 0, + "content": "3D Gaussian Splatting (3DGS), a recently emerged multiview 3D reconstruction technique, has shown significant advantages in real-time rendering and explicit editing. However, 3DGS encounters challenges in the accurate modeling of both high-frequency view-dependent appearances and global illumination effects, including inter-reflection. This paper introduces SpecTRe-GS, which addresses these challenges and models highly Specular surfaces that reflect nearby objects through Tracing Rays in 3D Gaussian Splatting. SpecTRe-GS separately models reflections from highly specular and rough surfaces to leverage the distinctions between their reflective properties and integrates an efficient ray tracer within the 3DGS framework for querying secondary rays, thus achieving fast and accurate rendering. Also, it incorporates normal prior guidance and joint geometry optimization at various stages of the training process to enhance geometry reconstruction for undistorted reflections. Experiments on both synthetic and real-world scenes demonstrate the superiority of SpecTRe-GS compared to existing 3DGS-based methods in capturing highly specular inter-reflections and also showcase its editing applications." + }, + { + "type": "title", + "bbox": [ + 0.515, + 0.518, + 0.645, + 0.533 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.549, + 0.907, + 0.836 + ], + "angle": 0, + "content": "Recently, 3D Gaussian Splatting (3DGS) [15] has emerged as an innovative paradigm for multi-view 3D reconstruction. Unlike its predecessor, neural radiance fields (NeRF) [23], 3DGS adopts an explicit point-based scene representation with tile-based rasterization rather than relying on ray casting through an implicit neural field for rendering. This transition endows 3DGS with significant advantages in terms of real-time rendering speed and allows for convenient post-reconstruction editing. Nevertheless, challenges arise when the scene to be reconstructed includes highly specular (i.e., mirror-like) surfaces which exhibit strong and high-frequency view-dependent appearances. The original 3DGS struggles to accurately model such high-frequency reflections, as it utilizes per-Gaussian third-order spherical harmonics (SH) to represent view-dependent appearances. This approach is restricted to low-frequency radiance variations in angular space and is susceptible to overfitting when the training views that observe the Gaussian are not sufficiently dense [17, 37]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.837, + 0.909, + 0.898 + ], + "angle": 0, + "content": "To address this issue, a natural strategy is to adopt a physically-based shading model and optimize reflective properties for each Gaussian [2, 7, 8, 10, 11, 13, 18, 29, 35, 36, 39, 40, 43, 47], rather than merely recording the out" + }, + { + "type": "page_footnote", + "bbox": [ + 0.114, + 0.875, + 0.39, + 0.888 + ], + "angle": 0, + "content": "The work was done during an internship in Huawei." + }, + { + "type": "page_footnote", + "bbox": [ + 0.114, + 0.888, + 0.448, + 0.9 + ], + "angle": 0, + "content": "Corresponding author. E-mail: shiboxin@pku.edu.cn." + }, + { + "type": "list", + "bbox": [ + 0.114, + 0.875, + 0.448, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16133" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.348 + ], + "angle": 0, + "content": "going radiance. This strategy enables the computation of reflected radiance by evaluating the physically constrained rendering equation [14], taking into account local scene components including geometry, material properties, and incident lighting, thereby yielding more accurate, realistic, and reliable view-dependent effects. Among the three local scene components, addressing local incident lighting is perhaps the most intricate. Adopting a suboptimal approach to addressing local incident lighting may undermine global illumination effects such as inter-reflections, resulting in an unrealistic scene appearance. Notably, this issue is particularly pronounced in scenes characterized by highly specular surfaces; for instance, the absence of reflections of nearby objects would be especially noticeable. Therefore, it is crucial for methods aimed at modeling such high-frequency view-dependent appearances through the rendering equation to carefully address local incident lighting." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.348, + 0.483, + 0.62 + ], + "angle": 0, + "content": "Unfortunately, effectively acquiring local incident lighting poses a substantial challenge for 3DGS-based methods. The advantages of 3DGS come with a trade-off: The exclusive reliance on the tile-based rasterizer limits its ability to efficiently render secondary rays emanating from vastly different origins towards arbitrary directions, which is essential for querying local incident lighting. As a compromise, a majority of methods neglect secondary effects [36, 39, 47] or periodically bake secondary effects including local occlusions [8, 40] and incident radiance [18, 29] as low-order SH stored in Gaussians or probes. However, low-order SH can only capture low-frequency variations in incident lighting, and the requirement of baking not only incurs additional storage costs but also hinders post-reconstruction editing, such as inserting or deleting nearby objects or relighting. Employing screen space ray tracing (SSRT) [2] produces high-frequency but incomplete inter-reflections as SSRT struggles to trace out-of-screen and occluded rays." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.62, + 0.483, + 0.892 + ], + "angle": 0, + "content": "We propose SpecTRe-GS to model highly Specular surfaces with reflected nearby objects through Tracing Rays in the 3D space within the Gaussian Splatting framework, overcoming the aforementioned challenges. We first made the observation that, although highly specular surfaces exhibit reflections that can be extremely high-frequency and thus hard to capture using SH color or neural network, they are also favorable for modeling in the sense that the specular lobe of their bidirectional reflectance distribution functions (BRDF) are highly concentrated. This fact facilitates the evaluation of the rendering equation [14] on these surfaces because tracing only a few secondary rays already suffices, in contrast to rougher surfaces with more scattered specular lobes. To make use of this fact, we separately model highly specular and rough surfaces, ray-tracing the appearance of the former while retaining the low-order SH color to well-approximate the latter, significantly reducing the amount of secondary rays to trace. To further speed up the time" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.093, + 0.905, + 0.181 + ], + "angle": 0, + "content": "consuming ray tracing process, we implement a ray tracer within the 3DGS framework, inspired by the recent progresses [24, 42]. We also propose a training strategy to enhance the scene geometry reconstruction, the quality of which directly decides the quality of the reconstructed reflection. To summarize, our contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.183, + 0.905, + 0.227 + ], + "angle": 0, + "content": "- proposing a shading model tailored for indirect mirror reflections combined with an efficient ray tracer for fast and accurate rendering;" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.228, + 0.904, + 0.258 + ], + "angle": 0, + "content": "- designing a training strategy to enhance geometry reconstruction for precise incident direction computing; and" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.258, + 0.904, + 0.288 + ], + "angle": 0, + "content": "- achieving state-of-the-art novel view synthesis results on scenes with highly specular reflections (shown in Fig. 1)." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.183, + 0.905, + 0.288 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.302, + 0.655, + 0.318 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.327, + 0.905, + 0.809 + ], + "angle": 0, + "content": "Improving view-dependent appearances in 3DGS. The original 3DGS [15] uses low-order SH color that captures only low-frequency angular variations, a limitation that many follow-up methods were trying to address (see Tab. 1). While a few methods [30, 35] (row 2) adopt neural networks to decode local features, the majority [2, 8, 10, 13, 18, 29, 36, 39, 40, 47] (rows 3-8) employ the physically-based rendering equation [14]. This equation offers the most accurate approach for modeling view-dependent appearances but requires explicit modeling of surface normals, BRDFs, and the most demanding component - incident lighting. The rendering equation can accommodate all-frequency BRDFs; thus, the frequency of the reflected radiance is now bounded by that of the modeled incident lighting. While an environment map suffices to model direct lighting, it is significantly more challenging to model indirect lighting responsible for global illumination effects including occlusions and inter-reflections. Some methods circumvent this issue by ignoring secondary effects [36] (row 3) or by using a residual color [13] (row 4). For methods that explicitly model indirect lighting, a common strategy is to periodically bake occlusions [8] or inter-reflections [18, 29] into low-order SH, thereby limiting their incident radiance to low-frequency variations (rows 5-7). GI-GS [2] (row 8) uses screen space ray tracing to simulate high-frequency inter-reflections; however, it only models those whose sources are directly visible in the viewpoint. There are also specialized methods for plane mirrors that reflect the point cloud or camera about the plane [19, 22]. Our SpecTRe-GS traces the incident lighting in 3D space, thus is able to render high-frequency and complete inter-reflections for highly specular surfaces with convex shapes." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.811, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Ray-based 3D Gaussian rendering. It is noteworthy that tracing secondary rays for incoming local lighting has almost become a common practice in recent NeRF-based methods that explicitly model specular reflections or interreflections [9, 20, 32, 33, 45]. However, it is challenging to achieve this in 3DGS-based frameworks. 3DGRT [24]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16134" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.907, + 0.146 + ], + "angle": 0, + "content": "Table 1. We summarize works improving the view-dependent appearances upon 3DGS [15] based on: How do they model the view-dependent appearances (i.e., the outgoing radiance) and the supported frequency in angular space; how do they accommodate the global illumination effects (i.e., the local incident lighting) including occlusions and inter-reflections and the supported frequency in angular space; and whether is it feasible to edit the scene appearance under (in)direct lighting after reconstruction without any re-baking." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.149, + 0.902, + 0.346 + ], + "angle": 0, + "content": "
MethodView-dependent AppearanceGlobal Illumination EffectEditable Appearance
ModelingFrequencyOcclusionInter-reflectionFrequency
(1) 3DGS [15]Low-order SHLowJointly baked in outgoing SH colorLowNone
(2) 3iGS [30]; [35]Neural networkNet.-boundedJointly baked in the neural featureNet.-boundedNone
(3) DeferredGS [36]IgnoredNone
(4) GShader [13]; [39]Arbitrary, thus bounded by the frequency of incident lightingJointly baked in outgoing SH residualDirect
(5) GS-ID [8]; [40]Physically-based rendering equation [14]Baked as SHIgnoredLow
(6) GS-IR [18]; [10]Baked as SH
(7) GIR [29]Ray-traced in 3D spaceBaked as SH
(8) GI-GS [2]Ray-traced in screen space, incompleteHigh(In)direct
(9) OursRay-traced in 3D spaceHigh(In)direct
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.354, + 0.482, + 0.52 + ], + "angle": 0, + "content": "first demonstrates efficient ray tracing in 3D Gaussians, enabling ray-based lighting effects such as non-projective views. RayGauss [1] and EVER [21] employ ray sampling in 3D Gaussians, mitigating common artifacts in 3DGS, such as aliasing [41], popping artifacts [27] and projection errors [12]. Yang et al. [46] adopt path tracing to simulate global illumination with Gaussian primitives. Meanwhile, Jorge et al. [4] utilizes ray sampling to simulate volumetric scattering and emissive media. Our SpecTRe-GS integrates ray tracing techniques to evaluate accurate incident radiance in the 3DGS framework at an affordable cost." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.537, + 0.265, + 0.554 + ], + "angle": 0, + "content": "3. Proposed Method" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.562, + 0.483, + 0.745 + ], + "angle": 0, + "content": "Our method is specifically designed to accurately model the high-frequency reflections of highly specular surfaces. We build upon the recent advancements in 3DGS (Sec. 3.1) and adapt the reflection model to efficiently handle scenes comprising both highly specular objects and rougher objects (Sec. 3.2). To precisely and reliably query the incident radiance necessary for rendering both direct and indirect highly specular reflections, we incorporate ray tracing within the 3DGS framework (Sec. 3.3). We employ a dedicated training strategy to enhance geometry reconstruction, which is crucial for rendering undistorted and realistic reflections (Sec. 3.4). Our pipeline is outlined in Fig. 2." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.757, + 0.221, + 0.773 + ], + "angle": 0, + "content": "3.1. Preliminary" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.778, + 0.483, + 0.903 + ], + "angle": 0, + "content": "The seminal work of 3DGS [15] represents the scene as a translucent Gaussian point cloud, where each 3D Gaussian is described with central point \\(\\pmb{\\mu} \\in \\mathbb{R}^3\\), scales \\(\\pmb{s} \\in \\mathbb{R}^3\\), rotation quaternion \\(\\pmb{q} \\in \\mathbb{R}^4\\), and opacity \\(\\sigma \\in [0,1]\\). The distribution of each Gaussian in 3D space is given by \\(G(\\pmb{x}) = e^{-\\frac{1}{2} (\\pmb{x} - \\pmb{\\mu})^\\top \\pmb{\\Sigma}^{-1}(\\pmb{x} - \\pmb{\\mu})}\\), where the covariance \\(\\pmb{\\Sigma} = \\mathbf{RSS}^\\top \\mathbf{R}^\\top\\) is compactly parameterized as \\(\\pmb{s}\\) and \\(\\pmb{q}\\). By performing local affine transformation [48], 3D Gauss" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.354, + 0.907, + 0.414 + ], + "angle": 0, + "content": "sians are projected as 2D Gaussians on the image plane for rasterization. Then for each ray corresponding to an image pixel, the contributing 2D Gaussians are accumulated from front to back for volumetric rendering:" + }, + { + "type": "equation", + "bbox": [ + 0.58, + 0.419, + 0.907, + 0.439 + ], + "angle": 0, + "content": "\\[\nI \\left(\\boldsymbol {\\omega} _ {\\mathrm {o}}\\right) = \\sum_ {i} \\boldsymbol {c} _ {i} \\left(\\boldsymbol {\\omega} _ {\\mathrm {o}}\\right) \\alpha_ {i} \\prod_ {j = 1} ^ {i - 1} \\left(1 - \\alpha_ {j}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.442, + 0.906, + 0.472 + ], + "angle": 0, + "content": "where \\(\\alpha_{i} = \\sigma_{i}G_{i}\\) and the view-dependent color \\(c_{i}(\\omega_{\\mathrm{o}})\\in\\) \\(\\mathbb{R}^3\\) is modeled by third-order SH coefficients \\(\\varphi_{i}\\in \\mathbb{R}^{48}\\)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.473, + 0.906, + 0.565 + ], + "angle": 0, + "content": "Besides the view-dependent color \\( c_{i} \\), many other opt. mizable attributes including surface normal, diffuse albedo, or occlusion information, can be attached to each Gaussian. These attributes can be used locally in each Gaussian, or be accumulated to a map in the same way as Eq. (1) for later applications such as deferred shading [39]." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.577, + 0.835, + 0.594 + ], + "angle": 0, + "content": "3.2. Highly Specular Reflection Modeling" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.599, + 0.907, + 0.903 + ], + "angle": 0, + "content": "To model high-frequency reflections, we follow the convention to incorporate the physically-based rendering equation [14]. As commonly observed phenomena in real-world life, the inter-reflections from the nearby objects on highly specular surfaces should be modeled for a photorealistic scene appearance. The common practice with distant environment maps [13] ignores such inter-reflections and does not suffice for this task, thus we employ raytracing (detailed in Sec. 3.3) in 3D space for a complete and high-frequency inter-reflection. However, we found that performing ray tracing on the entire scene, including highly specular surfaces and rougher surfaces, is highly time-consuming. Evaluating the rendering equation [14] on rougher surfaces typically needs larger samples of the incident lighting even with multiple importance sampling [31], thus heavily increases the computational cost. Fortunately, previous works [15] have proven that rough surfaces can be well-captured by directly optimizing the outgoing color. Therefore, we propose the following reflection model that treats highly specular surfaces and rougher surfaces sepa" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16135" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.091, + 0.087, + 0.907, + 0.345 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.348, + 0.907, + 0.461 + ], + "angle": 0, + "content": "Figure 2. Overview of the SpecTRe-GS pipeline. The scene is represented as a Gaussian point cloud, where each Gaussian is associated with geometry and shading attributes used to render the depth map \\(D\\), normal map \\(N\\), diffuse reflection \\(I_{\\mathrm{diff}}\\), specular reflectance \\(A_{\\mathrm{spec}}\\) and rough surface reflection \\(I_{\\mathrm{rs}}\\) through splatting (Sec. 3.1). Highly specular surface reflection \\(I_{\\mathrm{ss}}\\) is formulated as the sum of the diffuse reflection \\(I_{\\mathrm{diff}}\\) and the specular reflection computed from the incident radiance \\(I_{\\mathrm{i}}\\) and specular reflectance \\(A_{\\mathrm{spec}}\\) (Sec. 3.2). The incident radiance \\(I_{\\mathrm{i}}\\) is determined along reflected secondary rays given by the scene geometry \\(D\\) and \\(N\\), with the indirect component \\(I_{\\mathrm{ind}}\\) and visibility \\(V_{\\mathrm{i}}\\) evaluated via efficient ray tracing within the Gaussian point cloud, and the direct component \\(I_{\\mathrm{dir}}\\) modeled by an estimizable environment map (Sec. 3.3). The scene geometry is enhanced using normal prior guidance, numerical gradients from ray-traced incident radiance, and other training signals to achieve undistorted reflections (Sec. 3.4)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.476, + 0.483, + 0.536 + ], + "angle": 0, + "content": "rately to enhance the specular reflections while maintaining the rendering speed. Also, we follow deferred shading [5] to shade for each pixel instead of each Gaussian, as it reduces the computation and gives sharp surface normals [47]." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.537, + 0.483, + 0.611 + ], + "angle": 0, + "content": "Modeling highly specular surfaces. To model surfaces that may exhibit both highly specular and diffuse reflections, such as polished marble, we define the reflection \\( I_{\\mathrm{ss}} \\) for these highly specular surfaces at surface point \\( x \\), observed from view direction \\( \\omega_{\\mathrm{o}} \\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.16, + 0.618, + 0.483, + 0.635 + ], + "angle": 0, + "content": "\\[\nI _ {\\mathrm {s s}} (\\boldsymbol {x}, \\omega_ {\\mathrm {o}}) = I _ {\\text {d i f f}} (\\boldsymbol {x}) + I _ {\\text {s p e c}} (\\boldsymbol {x}, \\omega_ {\\mathrm {o}}). \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.64, + 0.484, + 0.699 + ], + "angle": 0, + "content": "While the diffuse reflection \\( I_{\\mathrm{diff}} \\) is the view-independent outgoing radiance directly optimized as \\( d \\) in each Gaussian, the specular reflection \\( I_{\\mathrm{spec}} \\) is approximated with an ideal mirror reflection:" + }, + { + "type": "equation", + "bbox": [ + 0.157, + 0.706, + 0.483, + 0.723 + ], + "angle": 0, + "content": "\\[\nI _ {\\operatorname {s p e c}} (\\boldsymbol {x}, \\omega_ {\\mathrm {o}}) = A _ {\\operatorname {s p e c}} (\\boldsymbol {x}, \\omega_ {\\mathrm {r}}) I _ {\\mathrm {i}} (\\boldsymbol {x}, \\omega_ {\\mathrm {r}}), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.728, + 0.484, + 0.789 + ], + "angle": 0, + "content": "where the reflected incident direction \\(\\omega_{\\mathrm{r}}\\) is given by \\(\\omega_{\\mathrm{o}} - 2\\omega_{\\mathrm{o}}^{\\top}\\pmb{n}\\cdot \\pmb{n}\\) in which \\(\\pmb{n}\\) denotes the local surface normal, and \\(A_{\\mathrm{spec}}\\) is the specular reflectance considering Fresnel effects through Schlick's approximation [28]:" + }, + { + "type": "equation", + "bbox": [ + 0.104, + 0.792, + 0.483, + 0.812 + ], + "angle": 0, + "content": "\\[\nA _ {\\mathrm {s p e c}} (\\pmb {x}, \\pmb {\\omega} _ {\\mathrm {r}}) = F _ {0} (\\pmb {x}) + (1 - F _ {0} (\\pmb {x})) (1 - \\pmb {n} ^ {\\top} \\pmb {\\omega} _ {\\mathrm {r}}) ^ {5}. \\quad (4)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.815, + 0.484, + 0.862 + ], + "angle": 0, + "content": "Here, \\(F_{0}\\) is the specular reflectance at normal incidence optimized as \\(\\pmb{f_0}\\) in each Gaussian and is 3-channel to support metallic materials. The incident radiance \\(I_{\\mathrm{i}}\\) in \\(\\omega_{\\mathrm{r}}\\) originates" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.476, + 0.764, + 0.491 + ], + "angle": 0, + "content": "from both direct and indirect sources:" + }, + { + "type": "text", + "bbox": [ + 0.882, + 0.483, + 0.905, + 0.497 + ], + "angle": 0, + "content": "(5)" + }, + { + "type": "equation", + "bbox": [ + 0.513, + 0.499, + 0.918, + 0.516 + ], + "angle": 0, + "content": "\\[\nI _ {\\mathrm {i}} (\\boldsymbol {x}, \\omega_ {\\mathrm {r}}) = V _ {\\mathrm {i}} (\\boldsymbol {x}, \\omega_ {\\mathrm {r}}) I _ {\\mathrm {d i r}} (\\omega_ {\\mathrm {r}}) + (1 - V _ {\\mathrm {i}} (\\boldsymbol {x}, \\omega_ {\\mathrm {r}})) I _ {\\mathrm {i n d}} (\\boldsymbol {x}, \\omega_ {\\mathrm {r}}),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.522, + 0.907, + 0.581 + ], + "angle": 0, + "content": "where \\( V_{\\mathrm{i}} \\in [0,1] \\) quantifies the proportion of the incident lighting occluded by the scene itself, \\( I_{\\mathrm{dir}} \\) is the direct incident lighting from a distant environment map, and \\( I_{\\mathrm{ind}} \\) is the indirect incident radiance from the scene." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.583, + 0.909, + 0.779 + ], + "angle": 0, + "content": "Defining depth and surface normal of Gaussians. It is essential to accurately acquire the depth and surface normal of each Gaussian in our method, as they directly determine the origin and direction of the reflection. However, they are ambiguous because the Gaussians are translucent ellipsoids in 3D space. Instead of treating the depth of the projected center and the shortest axis direction facing the camera as the depth and normal of each Gaussian [13, 17, 39], we adopt the formulation in GOF [42] to calculate for each Gaussian the \"intersection point\" along the ray achieving the maximum response, using its depth and normal of the intersection plane. This formulation gives a better and more consistent geometry reconstruction with 3D Gaussians." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.779, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Modeling the whole image. We model the part of the scene containing highly specular surfaces using Eq. (2). For the other part which exhibits only low-frequency view dependencies, we follow GOF [42] to model the appearance \\( I_{\\mathrm{rs}} \\) of the rougher surfaces using per-Gaussian SH coefficients \\( \\varphi \\). We combine \\( I_{\\mathrm{ss}} \\) and \\( I_{\\mathrm{rs}} \\) using a binary mask \\( M \\) (which can be handily acquired through pre-trained segmentation models such as SAM [16]) specifying the highly specular" + }, + { + "type": "page_footnote", + "bbox": [ + 0.091, + 0.875, + 0.483, + 0.9 + ], + "angle": 0, + "content": "1We omit the distribution and mask-shadowing term in the microfacet model as we assume ideal mirror reflection." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16136" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.474, + 0.107 + ], + "angle": 0, + "content": "surfaces within the scene ( \\(\\odot\\) is pixel-wise multiplication):" + }, + { + "type": "equation", + "bbox": [ + 0.134, + 0.113, + 0.483, + 0.129 + ], + "angle": 0, + "content": "\\[\nI \\left(\\omega_ {\\mathrm {o}}\\right) = I _ {\\mathrm {s s}} \\left(\\omega_ {\\mathrm {o}}\\right) \\odot M + I _ {\\mathrm {r s}} \\left(\\omega_ {\\mathrm {o}}\\right) \\odot (1 - M). \\tag {6}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.14, + 0.465, + 0.155 + ], + "angle": 0, + "content": "3.3. Efficient Ray Tracing for Incident Radiance" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.162, + 0.484, + 0.356 + ], + "angle": 0, + "content": "To accurately compute the reflection from highly specular surfaces, it is crucial to obtain the incident radiance at the reflected view direction as modeled by Eq. (5), which incorporates both direct and indirect lighting. However, due to the incoherent origins and directions of these secondary rays, the incident radiance cannot be efficiently rendered using the rasterizer from the original 3DGS [15], as it relies on the locality of the rays, nor can it be computed merely using an environment map, which considers only direct lighting [13]. Inspired by the recent work 3DGRT [24], we implement an efficient ray tracer within the 3DGS framework to compute the radiance of secondary rays, thus achieving high-frequency and accurate reflections." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.358, + 0.484, + 0.509 + ], + "angle": 0, + "content": "Geometry acceleration structure (GAS). For fast scene traversal, we construct a GAS within the OptiX programming interface [25], which uses ray tracing hardware to accelerate ray-primitive intersection tests and is several times faster than BVH [3] in CUDA. Instead of using axis-aligned bounding boxes which inefficiently encapsulate irregular Gaussians and lead to false positives, we adopt stretched icosahedrons as proxy meshes for more efficient ray intersection [24]. The vertex \\(j\\) of the icosahedron approximating Gaussian \\(i\\) is given by:" + }, + { + "type": "equation", + "bbox": [ + 0.143, + 0.514, + 0.482, + 0.534 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {v} _ {i, j} = \\sqrt {2 \\log \\left(\\sigma_ {i} / \\alpha_ {\\min }\\right)} \\mathbf {R} _ {i} ^ {\\top} \\mathbf {S} _ {i} ^ {\\top} \\boldsymbol {v} _ {j} ^ {\\text {u n i t}} + \\boldsymbol {\\mu} _ {i}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.539, + 0.484, + 0.66 + ], + "angle": 0, + "content": "where \\(\\pmb{v}_j^{\\mathrm{unit}}\\) is the vertex of an inscribed icosahedron in a unit sphere. Here, \\(\\alpha_{\\mathrm{min}}\\) serves as a threshold that adaptively scales each icosahedron by dictating the minimal contribution of the Gaussian at the transformed vertices, thereby reducing the computational overhead during intersection and radiance evaluation. We regularly update the GAS after each densification operation [15] and every \\(n_{\\mathrm{up}}\\) step as the Gaussian point cloud changes along the training." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.66, + 0.484, + 0.751 + ], + "angle": 0, + "content": "Ray queries with GAS. To evaluate the incident radiance \\( I_{\\mathrm{ind}}(\\pmb{x}_0, \\pmb{\\omega}_{\\mathrm{r}}) \\), secondary rays \\( \\pmb{x}(t) = \\pmb{x}_0 + t\\pmb{\\omega}_{\\mathrm{r}} \\) are traced against proxy meshes within the GAS. Upon tracing each ray, we first collect the \\( k \\) closest valid hits, each corresponding to a Gaussian with \\( \\alpha_i \\geq \\alpha_{\\min} \\). The depth of the \"true\" intersection point for each Gaussian is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.185, + 0.755, + 0.483, + 0.79 + ], + "angle": 0, + "content": "\\[\nt _ {\\max , i} = \\frac {\\left(\\boldsymbol {\\mu} _ {i} - \\boldsymbol {x} _ {0}\\right) ^ {\\top} \\boldsymbol {\\Sigma} _ {i} ^ {- 1} \\boldsymbol {\\omega} _ {\\mathrm {r}}}{\\boldsymbol {\\omega} _ {\\mathrm {r}} ^ {\\top} \\boldsymbol {\\Sigma} _ {i} ^ {- 1} \\boldsymbol {\\omega} _ {\\mathrm {r}}}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.484, + 0.901 + ], + "angle": 0, + "content": "which maximizes the Gaussian's contribution \\(\\sigma_{i}G_{i}(\\pmb{x}_{0} + t_{\\max,i}\\pmb{\\omega}_{\\mathrm{r}})\\). We then sort the Gaussians by their depth and update the ray depth \\(D_{\\mathrm{ind}}\\), visibility (i.e., transmittance) \\(V_{\\mathrm{i}}\\) and color \\(I_{\\mathrm{ind}}\\) (using the Gaussians' outgoing SH colors assuming that they model rough surfaces), following Eq. (1). The above process is repeated after advancing the ray origin until no more valid hits remain or the ray transmittance" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.182 + ], + "angle": 0, + "content": "drops below a threshold. For the transmittance that remains unoccluded by the reconstructed scene itself, we adopt an estimable environment map for calculating the direct incident radiance \\( I_{\\mathrm{dir}}(\\omega) \\). According to Eq. (5), the final incident radiance \\( I_{\\mathrm{i}}(\\pmb{x}, \\pmb{\\omega}_{\\mathrm{r}}) \\) combines direct and indirect components, weighted by visibility \\( V_{\\mathrm{i}}(\\pmb{x}, \\pmb{\\omega}) \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.183, + 0.907, + 0.289 + ], + "angle": 0, + "content": "Thanks to the efficient and hardware-accelerated ray tracing techniques elaborated above, as well as our separate modeling of highly specular and rough surfaces, the average rendering time of our method is only \\(50\\%\\) to \\(100\\%\\) slower than that of a pure rasterizer on each of our test scenes. This varies according to the density of objects in the scene and the proportion of highly specular surfaces." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.307, + 0.618, + 0.322 + ], + "angle": 0, + "content": "3.4. Training" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.328, + 0.906, + 0.404 + ], + "angle": 0, + "content": "We propose a training strategy that especially focuses on enhancing the geometry reconstruction, as unfaithful scene geometry induces inaccurate secondary rays and results in distorted reflections. We optimize our model from an initial sparse point cloud using the following loss terms:" + }, + { + "type": "equation", + "bbox": [ + 0.522, + 0.41, + 0.905, + 0.426 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {\\mathrm {c}} + \\lambda_ {\\mathrm {d}} \\mathcal {L} _ {\\mathrm {d}} + \\lambda_ {\\mathrm {d n}} \\mathcal {L} _ {\\mathrm {d n}} + \\lambda_ {\\mathrm {s m}} \\mathcal {L} _ {\\mathrm {s m}} + \\lambda_ {\\mathrm {n}} \\mathcal {L} _ {\\mathrm {n}} + \\lambda_ {\\mathrm {r}} \\mathcal {L} _ {\\mathrm {r}}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.431, + 0.906, + 0.552 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_{\\mathrm{c}}\\) is a color reconstruction loss [15] differentiable to the geometry, \\(\\mathcal{L}_{\\mathrm{d}}\\) and \\(\\mathcal{L}_{\\mathrm{dn}}\\) are the distortion and depth-normal consistency loss used in GOF [42], \\(\\mathcal{L}_{\\mathrm{sm}}\\) is an edge-aware smoothness term for the normal and specular reflectance for the highly specular surfaces, \\(\\mathcal{L}_{\\mathrm{n}}\\) is a normal prior term to employ data-driven clues, and \\(\\mathcal{L}_{\\mathrm{r}}\\) is a residual-suppressing term for progressive learning. This section describes these terms and more details are in the supplement." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.552, + 0.905, + 0.612 + ], + "angle": 0, + "content": "Normal prior guidance. To tackle the challenging problem of reconstructing geometry for highly specular objects, we employ StableNormal [38] to provide monocular normal priors \\(\\hat{N}\\), offering initial geometric constraints:" + }, + { + "type": "equation", + "bbox": [ + 0.587, + 0.618, + 0.905, + 0.639 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {n}} = 1 / | N | \\sum \\left| \\left| (\\hat {N} - N) \\odot M \\right| \\right| _ {1}. \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.644, + 0.907, + 0.825 + ], + "angle": 0, + "content": "Joint geometry optimization. Recognizing potential inaccuracies and inconsistencies across views in the normal prior, we concurrently optimize the scene geometry during training. This is achieved by differentiating the incident radiance \\( I_{\\mathrm{i}} \\), which relates to the final image through Eq. (2) and Eq. (3), with respect to the scene geometry. Computing the gradient of the direct incident radiance \\( I_{\\mathrm{dir}} \\) is straightforward. To compute the gradient of the indirect incident radiance \\( I_{\\mathrm{ind}} \\) concerning the local surface normal \\( n \\), we adopt a simple yet effective finite difference method to numerically evaluate the gradient of \\( I_{\\mathrm{ind}} \\) with respect to the reflected direction \\( \\omega \\) that depends on \\( n \\):" + }, + { + "type": "equation", + "bbox": [ + 0.52, + 0.829, + 0.905, + 0.866 + ], + "angle": 0, + "content": "\\[\n\\left. \\frac {\\partial I _ {\\text {i n d}}}{\\partial \\omega^ {i}} \\right| _ {\\omega_ {0}} \\approx \\frac {I _ {\\text {i n d}} \\left(\\boldsymbol {x} , \\omega_ {0} + \\Delta_ {\\omega} ^ {i}\\right) - I _ {\\text {i n d}} \\left(\\boldsymbol {x} , \\omega_ {0} - \\Delta_ {\\omega} ^ {i}\\right)}{2 \\Delta_ {\\omega} ^ {i}}, \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.905, + 0.901 + ], + "angle": 0, + "content": "where \\(\\Delta_{\\omega}^{i}\\) denotes the perturbation of the \\(i\\)-th dimension in \\(\\omega\\). Considering that using the same perturbation for sec-" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16137" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.907, + 0.131 + ], + "angle": 0, + "content": "Table 2. Quantitative comparisons of novel view synthesis results with state-of-the-art methods on both synthetic and real-world scenes. We report the scores of PSNR, SSIM [34], and LPIPS [44] for entire images and within reflective regions. \\(\\uparrow (\\downarrow)\\) means higher (lower) is better. We mark the best and the second best results in each column." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.134, + 0.9, + 0.289 + ], + "angle": 0, + "content": "
DataSynthetic (entire image)Synthetic (reflective)Real-world (entire image)Real-world (reflective)
MethodPSNR↓SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↓SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
3DGS [15]27.710.9100.13121.510.9400.07824.820.8230.22721.310.9650.047
GOF [42]27.970.9170.11721.380.9400.07624.860.8260.21921.250.9630.047
GOF* [42]27.570.9130.12520.670.9360.08524.690.8250.22320.440.9620.051
3iGS [30]28.580.9120.12521.980.9390.07624.370.8090.23321.220.9630.048
GShader [13]26.650.8880.16020.790.9350.08623.930.8100.25520.230.9630.053
GShader* [13]27.140.8940.15321.220.9380.08424.030.8120.25320.360.9640.053
3DGS-DR [39]27.620.8900.17321.570.9380.08524.790.8240.24321.180.9650.050
3DGS-DR* [39]26.970.8830.18221.180.9360.08824.600.8230.25120.380.9640.055
Ours29.410.9220.10923.280.9490.06225.110.8280.21322.770.9670.035
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.293, + 0.483, + 0.355 + ], + "angle": 0, + "content": "ondary rays with different depths \\( D_{\\mathrm{ind}} \\) leads to different amounts of intersection movements, making the numerical gradient less accurate, we use a depth-aware ray perturbation that adjusts to \\( D_{\\mathrm{ind}} \\) for more informative gradients:" + }, + { + "type": "equation", + "bbox": [ + 0.124, + 0.359, + 0.483, + 0.377 + ], + "angle": 0, + "content": "\\[\n\\Delta_ {\\omega} ^ {i} = \\theta_ {\\mathrm {d}} \\max \\left(\\min \\left(D _ {\\text {i n d}} / D _ {0}, \\sigma_ {\\mathrm {d}} ^ {\\max }\\right), \\sigma_ {\\mathrm {d}} ^ {\\min }\\right), \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.381, + 0.483, + 0.488 + ], + "angle": 0, + "content": "where \\(\\theta_{\\mathrm{d}}\\) and \\(D_0\\) define a reference angular perturbation and \\(\\sigma_{\\mathrm{d}}^{\\mathrm{min}}\\) and \\(\\sigma_{\\mathrm{d}}^{\\mathrm{max}}\\) ensure numerical stability. For computational efficiency, we disregard the gradient with respect to the point position \\(\\pmb{p}\\), as we find reflections are distorted mainly due to normal errors. We detach visibility \\(V_{\\mathrm{ind}}\\) and depth \\(D_{\\mathrm{ind}}\\) because the sparsity of the visibility gradient and abrupt changes in depth damage numerical stability." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.488, + 0.483, + 0.653 + ], + "angle": 0, + "content": "Progressive learning. As training progresses, monocular normal priors may hinder the accurate geometry reconstruction, thus we gradually decay the weight of normal prior guidance \\(\\lambda_{\\mathrm{n}}\\), letting gradients from physics-based rendering take over the fine-tuning of scene geometry. Besides, for smoother training and consequently better reconstruction within the reflective regions, we replace the hard mask \\(M\\) used in Eq. (6) with a relaxed version \\(1 - I_{\\mathrm{r}}\\) rendered with another per-Gaussian attribute \\(r \\in [0,1]\\) using Eq. (1). We encourage the reflective region to be modeled with highly specular reflections by suppressing the residual within \\(M\\):" + }, + { + "type": "equation", + "bbox": [ + 0.188, + 0.658, + 0.482, + 0.681 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {r}} = 1 / | M | \\sum \\left| \\left| I _ {\\mathrm {r}} \\odot M \\right| \\right| _ {1}. \\tag {13}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.692, + 0.224, + 0.71 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.717, + 0.483, + 0.822 + ], + "angle": 0, + "content": "In this section, we compare our method on both synthetic and real-world scenes (Sec. 4.2) against state-of-the-art methods (Sec. 4.3). We also perform ablation studies to validate the key components of our method (Sec. 4.4) and explore applications in scene editing (Sec. 4.5). Please check the supplementary material for more implementation details, data creation details, and experimental results." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.834, + 0.31, + 0.85 + ], + "angle": 0, + "content": "4.1. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "We implement our method in PyTorch [26] with CUDA kernels for rasterization based on GOF [42] and custom OptiX programs for ray tracing. We conduct all experiments on a" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.293, + 0.907, + 0.37 + ], + "angle": 0, + "content": "GPU with 24 GB VRAM, 83 shader core TFLOPS, and 191 ray tracing core TFLOPS. We set \\(\\lambda_{\\mathrm{sm}} = 0.5\\), \\(\\lambda_{\\mathrm{r}} = 1.0\\), and \\(\\lambda_{\\mathrm{n}} = 0.5\\), exponentially decaying to 0.0025 from 4k to 10k iterations. We use \\(r = 0.9\\) as the initialization. For other hyperparameters, we use the settings in GOF [42]." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.379, + 0.616, + 0.393 + ], + "angle": 0, + "content": "4.2. Datasets" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.401, + 0.906, + 0.507 + ], + "angle": 0, + "content": "Synthetic scenes. We build 6 synthetic scenes, each containing multiple surrounding objects and a highly specular surface with distinctive properties: HELMET, MARBLETABLE, VASE, POT, TOASTER, and MIRROR. For each scene, we use the Blender Cycles engine [6] to render about 300 images of resolution \\(1024 \\times 1024\\), of which 2/3 are used for training and the rest are left for testing." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.508, + 0.906, + 0.568 + ], + "angle": 0, + "content": "Real-world scenes. For real-world validation, we capture two scenes: REALBOWL and REALPOT, consisting of 138 and 139 images of resolution \\(1440 \\times 1080\\) respectively, with every eighth image reserved in the test set." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.577, + 0.646, + 0.594 + ], + "angle": 0, + "content": "4.3. Comparison" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.599, + 0.907, + 0.75 + ], + "angle": 0, + "content": "Comparing methods. We compare our method with 5 open-source state-of-the-art methods: the original 3DGS [15]; GOF [42] which focus on geometry reconstruction with 3D Gaussians; 3iGS [30] which improves view-dependent effects through neural features; and 3DGS-DR [39] and GShader [13] which are adapted for specular reflection employing physically-based rendering (PBR). For a fair comparison, we provide GOF [42], 3DGS-DR [39], and GShader [13] with the normal prior guidance, and indicate these versions by “*” throughout this paper." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.751, + 0.906, + 0.811 + ], + "angle": 0, + "content": "Evaluation protocol. We report peak signal-to-noise ratio (PSNR), structural similarity index measure (SSIM) [34], and learned perceptual image patch similarity (LPIPS) [44] measured on both entire images and reflective regions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.906, + 0.902 + ], + "angle": 0, + "content": "Quantitative evaluation. As shown in Tab. 2, our method outperforms comparing methods on both synthetic and real-world scenes, especially in reflective regions. It is worth noting that using normal prior guidance usually leads to degradation of GOF [42] which relies solely on SH color, as it suppresses the ability to simulate high-frequency appear" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16138" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.089, + 0.907, + 0.367 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.369, + 0.907, + 0.412 + ], + "angle": 0, + "content": "Figure 3. Novel view synthesis comparison on synthetic scenes. We evaluate SpecTRe-GS against relevant 3DGS-based methods modeling view-dependent appearances: 3DGS-DR [39], GShader [13], 3iGS [30], GOF [42], and 3DGS [15]. The insets highlight the improvement of SpecTRe-GS in rendering high-frequency specular reflections, showcasing the precise positioning of reflected content and sharp details." + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.415, + 0.903, + 0.672 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.674, + 0.907, + 0.716 + ], + "angle": 0, + "content": "Figure 4. Novel view synthesis and image decomposition results on synthetic scenes, comparing our SpecTRe-GS with 3DGS-DR [39] and GShader [13]. SpecTRe-GS better distinguishes diffuse and specular components in reflective regions, with the least baked-in view-dependent appearance in diffuse components. The masks of reflective regions are shown as insets in the reference image." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.723, + 0.483, + 0.798 + ], + "angle": 0, + "content": "ances through false geometry. In contrast, the influence of this guidance on 3DGS-DR [39] and GShader [13] is twofold: It facilitates the reconstruction of appearance under direct lighting, but also prevents these methods from imitating specular inter-reflections by faking geometry." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.798, + 0.483, + 0.889 + ], + "angle": 0, + "content": "Qualitative evaluation. In visual comparisons, we show results of GOF [42], GShader* [13], and 3DGS-DR [39] as they are variants with higher qualities. As shown in Fig. 3, our method performs better than the baseline methods for both planar reflectors (rows 1-2) and curved reflectors (row 3), capturing clean, accurate, and high-frequency specular" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.723, + 0.907, + 0.89 + ], + "angle": 0, + "content": "reflections. Our method also surpasses baseline methods in real-world scenes, as shown in Fig. 1. By visualizing the separate components of PBR-based methods in Fig. 4, we find that incident radiance modeling for both direct and indirect sources is vital for physically meaningful scene appearance decomposition. Otherwise, to fit the overall color in training images, indirect incident radiance is heavily baked into the diffuse component (row 1 for 3DGS-DR [39] and row 3 for GShader [13]). The baked-in incident radiance loses high-frequency details and prevents realistic scene editing in a physically-accurate manner. It should be noted" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16139" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.093, + 0.09, + 0.482, + 0.324 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.327, + 0.483, + 0.368 + ], + "angle": 0, + "content": "Figure 5. Our reconstruction results on two real-world scenes. SpecTRe-GS accurately reconstructs the geometry of highly specular surfaces alongside the reflection of nearby objects." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.373, + 0.484, + 0.51 + ], + "angle": 0, + "content": "that 3DGS-DR [39] performs better than GShader [13] in reconstructing the direct part of incident radiance from the environment, showing the superiority of surface-based deferred shading over volumetric rendering in PBR modeling of high-frequency reflections. Our method, with well-optimized geometry and consequent accurate incident radiance modeling, achieves satisfactory novel view synthesis of specular reflection on both synthetic and real-world scenes, as also indicated in Fig. 5." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.517, + 0.245, + 0.533 + ], + "angle": 0, + "content": "4.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.539, + 0.484, + 0.842 + ], + "angle": 0, + "content": "On a subset of our synthetic scenes, we validate the following key component of our method: (1) Normal prior guidance; (2) Joint geometry optimization from incident lighting; (3) ray-traced Indirect incident radiance; (4) Progressive learning; and (5) Depth-aware ray perturbation. The results are shown in Tab. 3. Our full method (Ours) achieves the highest scores, confirming the effectiveness of our method designs. Without the initial normal prior guidance (Ours w/o N.), the optimization process is prone to be trapped into local minima far away from the true underneath geometry of the reflective region. On the other hand, only relying on normal priors without further joint optimization for fine-tuning geometry (Ours w/o J.) also fails in reconstructing accurate geometry for physically-based high-frequency reflections. As discussed above, using only direct incident lighting modeling (Ours w/o I.) leads to baked-in indirect incident lighting and damages the quality of high-frequency components. The progressive learning scheme (Ours w/o P.) and depth-aware ray perturbation (Ours w/o D.) also further helps to faithfully reconstruct the scene." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.849, + 0.226, + 0.866 + ], + "angle": 0, + "content": "4.5. Applications" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Our Gaussian point cloud representation with explicit geometry, material, and incident lighting facilitates various" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.09, + 0.907, + 0.146 + ], + "angle": 0, + "content": "Table 3. Ablation study with variants of our proposed method excluding: Normal prior guidance, Joint geometry optimization, Indirect incident lighting modeling, Progressive learning, and Depth-aware ray perturbation. \\(\\uparrow (\\downarrow)\\) means higher (lower) is better." + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.148, + 0.905, + 0.267 + ], + "angle": 0, + "content": "
MethodEntire ImageReflective Region
PSNR↓SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
Ours w/o N.26.700.8850.14919.800.9130.094
Ours w/o J.26.370.8780.15419.520.9070.100
Ours w/o I.28.490.9050.13022.190.9350.074
Ours w/o P.29.260.9100.11823.160.9400.062
Ours w/o D.29.750.9140.11423.880.9440.056
Ours29.900.9140.11224.050.9440.056
" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.269, + 0.905, + 0.432 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.435, + 0.907, + 0.477 + ], + "angle": 0, + "content": "Figure 6. Our explicit representation of geometry, material, and incident lighting in the Gaussian point cloud enables various editing applications. Insets have been brightened for clarity." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.486, + 0.907, + 0.593 + ], + "angle": 0, + "content": "editing applications. As illustrated in Fig. 6, specular objects of interest can be faithfully captured as assets in the presence of unwanted reflections from nearby objects, enabling further usages such as relighting (column 1) and material editing (column 4). We can also manipulate the scene geometry such as deletion (column 2) and insertion (column 3), with specular inter-reflections adjusted accordingly." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.606, + 0.634, + 0.621 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.631, + 0.906, + 0.707 + ], + "angle": 0, + "content": "We introduce SpecTRe-GS to model highly specular surfaces with inter-reflections, distinguished by its separate modeling of specular and rough surfaces, the efficient ray tracer, and a geometry-enhancing training strategy for fast and accurate specular inter-reflection rendering." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.707, + 0.906, + 0.813 + ], + "angle": 0, + "content": "Limitations. We only model specular inter-reflections from rough surfaces to avoid multi-bounce path tracing, and we only trace one secondary ray for each pixel in the reflective region, approximating the specular reflection with an ideal mirror. Also, the monocular normal prior may fail in the existence of strong semantic ambiguity (such as mirrors with hidden borders), making our geometry fail to initialize." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.819, + 0.673, + 0.835 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.839, + 0.906, + 0.9 + ], + "angle": 0, + "content": "This work is supported by the National Natural Science Foundation of China under Grant No. 62136001. PKU-affiliated authors thank openbayes.com for providing computing resource." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16140" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.184 + ], + "angle": 0, + "content": "[1] Hugo Blanc, Jean-Emmanuel Deschaud, and Alexis Paljic. RayGauss: Volumetric Gaussian-based ray casting for photorealistic novel view synthesis. In Proc. of IEEE Winter Conference on Applications of Computer Vision (WACV), 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.185, + 0.484, + 0.242 + ], + "angle": 0, + "content": "[2] Hongze Chen, Zehong Lin, and Jun Zhang. GI-GS: Global illumination decomposition on Gaussian splatting for inverse rendering. In Proc. of International Conference on Learning Representations (ICLR), 2025. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.242, + 0.483, + 0.283 + ], + "angle": 0, + "content": "[3] James H. Clark. Hierarchical geometric models for visible surface algorithms. Communications of the ACM (CACM), 19(10):547-554, 1976. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.283, + 0.484, + 0.367 + ], + "angle": 0, + "content": "[4] Jorge Condor, Sebastien Speierer, Lukas Bode, Aljaz Bozic, Simon Green, Piotr Didyk, and Adrian Jarabo. Don't splat your Gaussians: Volumetric ray-traced primitives for modeling and rendering scattering and emissive media. ACM Transactions on Graphics (TOG), 44(1):10:1-10:17, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.368, + 0.484, + 0.45 + ], + "angle": 0, + "content": "[5] Michael Deering, Stephanie Winner, Bic Schediwy, Chris Duffy, and Neil Hunt. The triangle processor and normal vector Shader: a VLSI system for high performance graphics. In Proc. of the ACM SIGGRAPH Conference and Exhibition On Computer Graphics and Interactive Techniques (SIGGRAPH), 1988. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.452, + 0.483, + 0.493 + ], + "angle": 0, + "content": "[6] Cycles Developers. Cycles: Open source production rendering, 2024. https://www.cycles-renderer.org/. Accessed: 2024-11-06. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.494, + 0.484, + 0.55 + ], + "angle": 0, + "content": "[7] Jan-Niklas Dihlmann, Arjun Majumdar, Andreas Engelhardt, Raphael Braun, and Hendrik P. A. Lensch. Subsurface scattering for 3D Gaussian splattering. In Proc. of Neural Information Processing Systems (NeurIPS), 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.55, + 0.484, + 0.606 + ], + "angle": 0, + "content": "[8] Kang Du, Zhihao Liang, and Zeyu Wang. GS-ID: Illumination decomposition on Gaussian splattering via diffusion prior and parametric light source optimization. In arXiv preprint arXiv:2408.08524, 2024. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.606, + 0.484, + 0.675 + ], + "angle": 0, + "content": "[9] Fan Fei, Jiajun Tang, Ping Tan, and Boxin Shi. VMINer: Versatile multi-view inverse rendering with near- and farfield light sources. In Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.677, + 0.483, + 0.746 + ], + "angle": 0, + "content": "[10] Jian Gao, Chun Gu, Youtian Lin, Hao Zhu, Xun Cao, Li Zhang, and Yao Yao. Relightable 3D Gaussian: Real-time point cloud relighting with BRDF decomposition and ray tracing. In Proc. of European Conference on Computer Vision (ECCV), 2024. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.746, + 0.483, + 0.816 + ], + "angle": 0, + "content": "[11] Yijia Guo, Yuanxi Bai, Liwen Hu, Ziyi Guo, Mianzhi Liu, Yu Cai, Tiejun Huang, and Lei Ma. PRTGS: Precomputed radiance transfer of Gaussian splats for real-time high-quality relighting. In Proc. of ACM International Conference on Multimedia (ACM MM), 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.816, + 0.483, + 0.873 + ], + "angle": 0, + "content": "[12] Letian Huang, Jiayang Bai, Jie Guo, Yuanqi Li, and Yanwen Guo. On the error analysis of 3D Gaussian splattering and an optimal projection strategy. In Proc. of European Conference on Computer Vision (ECCV), 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.483, + 0.903 + ], + "angle": 0, + "content": "[13] Yingwenqi Jiang, Jiadong Tu, Yuan Liu, Xifeng Gao, Xiaoxiao Long, Wenping Wang, and Yuexin Ma. GaussianShader:" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.115, + 0.484, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.907, + 0.147 + ], + "angle": 0, + "content": "3D Gaussian splattering with shading functions for reflective surfaces. In Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 1, 2, 3, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.149, + 0.907, + 0.203 + ], + "angle": 0, + "content": "[14] James T. Kajiya. The rendering equation. In Proc. of the ACM SIGGRAPH Conference and Exhibition On Computer Graphics and Interactive Techniques (SIGGRAPH), 1986. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.204, + 0.906, + 0.26 + ], + "angle": 0, + "content": "[15] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3D Gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics (TOG), 42(4):139:1-139:14, 2023. 1, 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.26, + 0.906, + 0.342 + ], + "angle": 0, + "content": "[16] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dollar, and Ross B. Girshick. Segment anything. In Proc. of IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.342, + 0.906, + 0.413 + ], + "angle": 0, + "content": "[17] Jiahe Li, Jiawei Zhang, Xiao Bai, Jin Zheng, Xin Ning, Jun Zhou, and Lin Gu. DNGaussian: Optimizing sparse-view 3D Gaussian radiance fields with global-local depth normalization. In Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.413, + 0.906, + 0.469 + ], + "angle": 0, + "content": "[18] Zhihao Liang, Qi Zhang, Ying Feng, Ying Shan, and Kui Jia. GS-IR: 3D Gaussian splatting for inverse rendering. In Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.469, + 0.906, + 0.552 + ], + "angle": 0, + "content": "[19] Jiayue Liu, Xiao Tang, Freeman Cheng, Roy Yang, Zhihao Li, Jianzhuang Liu, Yi Huang, Jiaqi Lin, Shiyong Liu, Xiaofei Wu, Songcen Xu, and Chun Yuan. MirrorGaussian: Reflecting 3D Gaussians for reconstructing mirror reflections. In Proc. of European Conference on Computer Vision (ECCV), 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.553, + 0.906, + 0.622 + ], + "angle": 0, + "content": "[20] Yuan Liu, Peng Wang, Cheng Lin, Xiaoxiao Long, Jiepeng Wang, Lingjie Liu, Taku Komura, and Wenping Wang. NeRO: Neural geometry and BRDF reconstruction of reflective objects from multiview images. ACM Transactions on Graphics (TOG), 42(4):114:1-114:22, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.622, + 0.906, + 0.692 + ], + "angle": 0, + "content": "[21] Alexander Mai, Peter Hedman, George Kopanas, Dor Verbin, David Futschik, Qiangeng Xu, Falko Kuester, Jonathan T. Barron, and Yinda Zhang. EVER: Exact volumetric ellipsoid rendering for real-time view synthesis. In arXiv preprint arXiv:2410.01804, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.692, + 0.907, + 0.762 + ], + "angle": 0, + "content": "[22] Jiarui Meng, Haijie Li, Yanmin Wu, Qiankun Gao, Shuzhou Yang, Jian Zhang, and Siwei Ma. Mirror-3DGS: Incorporating mirror reflections into 3D Gaussian splattering. In Proc. of IEEE International Conference on Visual Communications and Image Processing, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.762, + 0.906, + 0.831 + ], + "angle": 0, + "content": "[23] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing scenes as neural radiance fields for view synthesis. In Proc. of European Conference on Computer Vision (ECCV), 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.831, + 0.906, + 0.903 + ], + "angle": 0, + "content": "[24] Nicolas Moënne-Loccoz, Ashkan Mirzaei, Or Perel, Ricardo de Lutio, Janick Martinez Esturo, Gavriel State, Sanja Fidler, Nicholas Sharp, and Zan Gojcic. 3D Gaussian ray tracing: Fast tracing of particle scenes. ACM Transactions on Graphics (TOG), 43(6):232:1-232:19, 2024. 2, 5" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "16141" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.175 + ], + "angle": 0, + "content": "[25] Steven G. Parker, James Bigler, Andreas Dietrich, Heiko Friedrich, Jared Hoberock, David P. Luebke, David K. McAllister, Morgan McGuire, R. Keith Morley, Austin Robison, and Martin Stich. OptiX: a general purpose ray tracing engine. In ACM Transactions on Graphics (Proc. of ACM SIGGRAPH), 2010. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.177, + 0.482, + 0.3 + ], + "angle": 0, + "content": "[26] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Köpf, Edward Z. Yang, Zachary DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. PyTorch: An imperative style, high-performance deep learning library. In Proc. of Neural Information Processing Systems (NeurIPS), 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.302, + 0.482, + 0.37 + ], + "angle": 0, + "content": "[27] Lukas Radl, Michael Steiner, Mathias Parger, Alexander Weinrauch, Bernhard Kerbl, and Markus Steinberger. StopThePop: Sorted Gaussian splatting for view-consistent real-time rendering. ACM Transactions on Graphics (TOG), 43(4):64:1-64:17, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.372, + 0.482, + 0.412 + ], + "angle": 0, + "content": "[28] Christophe Schlick. An inexpensive BRDF model for physically-based rendering. Computer Graphics Forum (CGF), 13(3):233-246, 1994. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.414, + 0.482, + 0.482 + ], + "angle": 0, + "content": "[29] Yahao Shi, Yanmin Wu, Chenming Wu, Xing Liu, Chen Zhao, Haocheng Feng, Jingtuo Liu, Liangjun Zhang, Jian Zhang, Bin Zhou, Errui Ding, and Jingdong Wang. GIR: 3D Gaussian inverse rendering for reightable scene factorization. In arXiv preprint arXiv:2312.05133, 2023. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.484, + 0.482, + 0.524 + ], + "angle": 0, + "content": "[30] Zhe Jun Tang and Tat-Jen Cham. 3iGS: Factorised tensorial illumination for 3D Gaussian splatting. In Proc. of European Conference on Computer Vision (ECCV), 2024. 1, 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.526, + 0.482, + 0.592 + ], + "angle": 0, + "content": "[31] Eric Veach and Leonidas J. Guibas. Optimally combining sampling techniques for Monte Carlo rendering. In Proc. of the ACM SIGGRAPH Conference and Exhibition On Computer Graphics and Interactive Techniques (SIGGRAPH), 1995. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.595, + 0.482, + 0.663 + ], + "angle": 0, + "content": "[32] Dor Verbin, Peter Hedman, Ben Mildenhall, Todd Zickler, Jonathan T. Barron, and Pratul P. Srinivasan. Ref-NeRF: Structured view-dependent appearance for neural radiance fields. In Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.665, + 0.482, + 0.747 + ], + "angle": 0, + "content": "[33] Dor Verbin, Pratul P. Srinivasan, Peter Hedman, Ben Mildenhall, Benjamin Attal, Richard Szeliski, and Jonathan T. Barron. NeRF-Casting: Improved view-dependent appearance with consistent reflections. In Proc. of the ACM SIGGRAPH Conference and Exhibition on Computer Graphics and Interactive Techniques in Asia (SIGGRAPH Asia), 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.749, + 0.482, + 0.802 + ], + "angle": 0, + "content": "[34] Zhou Wang, A.C. Bovik, H.R. Sheikh, and E.P. Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing (TIP), 13(4):600-612, 2004. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.804, + 0.482, + 0.872 + ], + "angle": 0, + "content": "[35] Zhiru Wang, Shiyun Xie, Chengwei Pan, and Guoping Wang. SpecGaussian with latent features: A high-quality modeling of the view-dependent appearance for 3D Gaussian splatting. In Proc. of ACM International Conference on Multimedia (ACM MM), 2024. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.874, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[36] Tong Wu, Jia-Mu Sun, Yu-Kun Lai, Yuewen Ma, Leif Kobbelt, and Lin Gao. DeferredGS: Decoupled and editable" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "Gaussian splatting with deferred shading. In arXiv preprint arXiv:2404.09412, 2024. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.189 + ], + "angle": 0, + "content": "[37] Wangze Xu, Huachen Gao, Shihe Shen, Rui Peng, Jianbo Jiao, and Ronggang Wang. MVPGS: Excavating multi-view priors for Gaussian splatting from sparse input views. In Proc. of European Conference on Computer Vision (ECCV), 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.192, + 0.905, + 0.275 + ], + "angle": 0, + "content": "[38] Chongjie Ye, Lingteng Qiu, Xiaodong Gu, Qi Zuo, Yushuang Wu, Zilong Dong, Liefeng Bo, Yuliang Xiu, and Xiaoguang Han. StableNormal: Reducing diffusion variance for stable and sharp normal. In Proc. of the ACM SIGGRAPH Conference and Exhibition on Computer Graphics and Interactive Techniques in Asia (SIGGRAPH Asia), 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.277, + 0.905, + 0.344 + ], + "angle": 0, + "content": "[39] Keyang Ye, Qiming Hou, and Kun Zhou. 3D Gaussian splating with deferred reflection. In Proc. of the ACM SIGGRAPH Conference and Exhibition On Computer Graphics and Interactive Techniques (SIGGRAPH), 2024. 1, 2, 3, 4, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.347, + 0.905, + 0.388 + ], + "angle": 0, + "content": "[40] Keyang Ye, Qiming Hou, and Kun Zhou. Progressive radiance distillation for inverse rendering with Gaussian splatting. In arXiv preprint arXiv:2408.07595, 2024. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.39, + 0.905, + 0.445 + ], + "angle": 0, + "content": "[41] Zehao Yu, Anpei Chen, Binbin Huang, Torsten Sattler, and Andreas Geiger. Mip-Splatting: Alias-free 3D Gaussian splatting. In Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.447, + 0.905, + 0.501 + ], + "angle": 0, + "content": "[42] Zehao Yu, Torsten Sattler, and Andreas Geiger. Gaussian opacity fields: Efficient adaptive surface reconstruction in unbounded scenes. ACM Transactions on Graphics (TOG), 43(6):271:1-271:13, 2024. 1, 2, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.503, + 0.905, + 0.572 + ], + "angle": 0, + "content": "[43] Libo Zhang, Yuxuan Han, Wenbin Lin, Jingwang Ling, and Feng Xu. PRTGaussian: Efficient relighting using 3D Gaussians with precomputed radiance transfer. In Asia Pacific Signal and Information Processing Association Annual Summit and Conference (APSIPA ASC), 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.574, + 0.905, + 0.642 + ], + "angle": 0, + "content": "[44] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.644, + 0.905, + 0.699 + ], + "angle": 0, + "content": "[45] Yuanqing Zhang, Jiaming Sun, Xingyi He, Huan Fu, Rongfei Jia, and Xiaowei Zhou. Modeling indirect illumination for inverse rendering. In Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.701, + 0.905, + 0.741 + ], + "angle": 0, + "content": "[46] Yang Zhou, Songyin Wu, and Ling-Qi Yan. Unified Gaussian primitives for scene representation and rendering. In arXiv preprint arXiv:2406.09733, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.743, + 0.905, + 0.785 + ], + "angle": 0, + "content": "[47] Zuo-Liang Zhu, Beibei Wang, and Jian Yang. GS-ROR: 3D Gaussian splatting for reflective object relighting via SDF priors. In arXiv preprint arXiv:2406.18544, 2024. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.787, + 0.905, + 0.827 + ], + "angle": 0, + "content": "[48] Matthias Zwicker, Hanspeter Pfister, Jeroen van Baar, and Markus H. Gross. EWA volume splatting. In Proc. of IEEE Visualization Conference (VIS), 2001. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.827 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16142" + } + ] +] \ No newline at end of file diff --git a/2025/SpecTRe-GS_ Modeling Highly Specular Surfaces with Reflected Nearby Objects by Tracing Rays in 3D Gaussian Splatting/05871852-d282-4279-bc18-a4b518c95c0a_origin.pdf b/2025/SpecTRe-GS_ Modeling Highly Specular Surfaces with Reflected Nearby Objects by Tracing Rays in 3D Gaussian Splatting/05871852-d282-4279-bc18-a4b518c95c0a_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..924a7f7d90914d68a142583dca18ee848479af80 --- /dev/null +++ b/2025/SpecTRe-GS_ Modeling Highly Specular Surfaces with Reflected Nearby Objects by Tracing Rays in 3D Gaussian Splatting/05871852-d282-4279-bc18-a4b518c95c0a_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fab94e32dfcf8713e025f956fe99305dc41acd06515fa8546fabea6e0a9280e8 +size 6753569 diff --git a/2025/SpecTRe-GS_ Modeling Highly Specular Surfaces with Reflected Nearby Objects by Tracing Rays in 3D Gaussian Splatting/full.md b/2025/SpecTRe-GS_ Modeling Highly Specular Surfaces with Reflected Nearby Objects by Tracing Rays in 3D Gaussian Splatting/full.md new file mode 100644 index 0000000000000000000000000000000000000000..9efe7a662da618d7e5fee31ff4656fa4c483945f --- /dev/null +++ b/2025/SpecTRe-GS_ Modeling Highly Specular Surfaces with Reflected Nearby Objects by Tracing Rays in 3D Gaussian Splatting/full.md @@ -0,0 +1,321 @@ +# SpecTRe-GS: Modeling Highly Specular Surfaces with Reflected Nearby Objects by Tracing Rays in 3D Gaussian Splatting + +Jiajun Tang $^{1,2}$ + +Youyu Chen + +Fan Fei $^{1,2}$ + +Binxiao Huang + +Zhihao Li3 + +Zhenyu Chen + +Xiao Tang3 + +Xiaofei Wu3 + +lyong Liu3 + +Boxin Shi $^{1,2\#}$ + +$^{1}$ State Key Laboratory for Multimedia Information Processing, School of Computer Science, Peking University + +$^{2}$ National Engineering Research Center of Visual Technology, School of Computer Science, Peking University + +$^{3}$ Huawei Noah's Ark Lab $^{4}$ Harbin Institute of Technology $^{5}$ University of Hong Kong + +![](images/212d433fd9ee9f5cce3729380e35988aa00901f5e7a9731a1b9ac80863b7f2f7.jpg) +Figure 1. Our SpecTRe-GS effectively captures the appearance of reflectors with highly specular surfaces, even in the presence of nearby objects. In the real-world scene shown above, it performs favorably against previous methods [13, 15, 30, 39, 42], both visually (as the highlighted reflections within the insets) and quantitatively (with PSNR values within the reflective region annotated in the corners). + +# Abstract + +3D Gaussian Splatting (3DGS), a recently emerged multiview 3D reconstruction technique, has shown significant advantages in real-time rendering and explicit editing. However, 3DGS encounters challenges in the accurate modeling of both high-frequency view-dependent appearances and global illumination effects, including inter-reflection. This paper introduces SpecTRe-GS, which addresses these challenges and models highly Specular surfaces that reflect nearby objects through Tracing Rays in 3D Gaussian Splatting. SpecTRe-GS separately models reflections from highly specular and rough surfaces to leverage the distinctions between their reflective properties and integrates an efficient ray tracer within the 3DGS framework for querying secondary rays, thus achieving fast and accurate rendering. Also, it incorporates normal prior guidance and joint geometry optimization at various stages of the training process to enhance geometry reconstruction for undistorted reflections. Experiments on both synthetic and real-world scenes demonstrate the superiority of SpecTRe-GS compared to existing 3DGS-based methods in capturing highly specular inter-reflections and also showcase its editing applications. + +# 1. Introduction + +Recently, 3D Gaussian Splatting (3DGS) [15] has emerged as an innovative paradigm for multi-view 3D reconstruction. Unlike its predecessor, neural radiance fields (NeRF) [23], 3DGS adopts an explicit point-based scene representation with tile-based rasterization rather than relying on ray casting through an implicit neural field for rendering. This transition endows 3DGS with significant advantages in terms of real-time rendering speed and allows for convenient post-reconstruction editing. Nevertheless, challenges arise when the scene to be reconstructed includes highly specular (i.e., mirror-like) surfaces which exhibit strong and high-frequency view-dependent appearances. The original 3DGS struggles to accurately model such high-frequency reflections, as it utilizes per-Gaussian third-order spherical harmonics (SH) to represent view-dependent appearances. This approach is restricted to low-frequency radiance variations in angular space and is susceptible to overfitting when the training views that observe the Gaussian are not sufficiently dense [17, 37]. + +To address this issue, a natural strategy is to adopt a physically-based shading model and optimize reflective properties for each Gaussian [2, 7, 8, 10, 11, 13, 18, 29, 35, 36, 39, 40, 43, 47], rather than merely recording the out + +going radiance. This strategy enables the computation of reflected radiance by evaluating the physically constrained rendering equation [14], taking into account local scene components including geometry, material properties, and incident lighting, thereby yielding more accurate, realistic, and reliable view-dependent effects. Among the three local scene components, addressing local incident lighting is perhaps the most intricate. Adopting a suboptimal approach to addressing local incident lighting may undermine global illumination effects such as inter-reflections, resulting in an unrealistic scene appearance. Notably, this issue is particularly pronounced in scenes characterized by highly specular surfaces; for instance, the absence of reflections of nearby objects would be especially noticeable. Therefore, it is crucial for methods aimed at modeling such high-frequency view-dependent appearances through the rendering equation to carefully address local incident lighting. + +Unfortunately, effectively acquiring local incident lighting poses a substantial challenge for 3DGS-based methods. The advantages of 3DGS come with a trade-off: The exclusive reliance on the tile-based rasterizer limits its ability to efficiently render secondary rays emanating from vastly different origins towards arbitrary directions, which is essential for querying local incident lighting. As a compromise, a majority of methods neglect secondary effects [36, 39, 47] or periodically bake secondary effects including local occlusions [8, 40] and incident radiance [18, 29] as low-order SH stored in Gaussians or probes. However, low-order SH can only capture low-frequency variations in incident lighting, and the requirement of baking not only incurs additional storage costs but also hinders post-reconstruction editing, such as inserting or deleting nearby objects or relighting. Employing screen space ray tracing (SSRT) [2] produces high-frequency but incomplete inter-reflections as SSRT struggles to trace out-of-screen and occluded rays. + +We propose SpecTRe-GS to model highly Specular surfaces with reflected nearby objects through Tracing Rays in the 3D space within the Gaussian Splatting framework, overcoming the aforementioned challenges. We first made the observation that, although highly specular surfaces exhibit reflections that can be extremely high-frequency and thus hard to capture using SH color or neural network, they are also favorable for modeling in the sense that the specular lobe of their bidirectional reflectance distribution functions (BRDF) are highly concentrated. This fact facilitates the evaluation of the rendering equation [14] on these surfaces because tracing only a few secondary rays already suffices, in contrast to rougher surfaces with more scattered specular lobes. To make use of this fact, we separately model highly specular and rough surfaces, ray-tracing the appearance of the former while retaining the low-order SH color to well-approximate the latter, significantly reducing the amount of secondary rays to trace. To further speed up the time + +consuming ray tracing process, we implement a ray tracer within the 3DGS framework, inspired by the recent progresses [24, 42]. We also propose a training strategy to enhance the scene geometry reconstruction, the quality of which directly decides the quality of the reconstructed reflection. To summarize, our contributions are as follows: + +- proposing a shading model tailored for indirect mirror reflections combined with an efficient ray tracer for fast and accurate rendering; +- designing a training strategy to enhance geometry reconstruction for precise incident direction computing; and +- achieving state-of-the-art novel view synthesis results on scenes with highly specular reflections (shown in Fig. 1). + +# 2. Related Work + +Improving view-dependent appearances in 3DGS. The original 3DGS [15] uses low-order SH color that captures only low-frequency angular variations, a limitation that many follow-up methods were trying to address (see Tab. 1). While a few methods [30, 35] (row 2) adopt neural networks to decode local features, the majority [2, 8, 10, 13, 18, 29, 36, 39, 40, 47] (rows 3-8) employ the physically-based rendering equation [14]. This equation offers the most accurate approach for modeling view-dependent appearances but requires explicit modeling of surface normals, BRDFs, and the most demanding component - incident lighting. The rendering equation can accommodate all-frequency BRDFs; thus, the frequency of the reflected radiance is now bounded by that of the modeled incident lighting. While an environment map suffices to model direct lighting, it is significantly more challenging to model indirect lighting responsible for global illumination effects including occlusions and inter-reflections. Some methods circumvent this issue by ignoring secondary effects [36] (row 3) or by using a residual color [13] (row 4). For methods that explicitly model indirect lighting, a common strategy is to periodically bake occlusions [8] or inter-reflections [18, 29] into low-order SH, thereby limiting their incident radiance to low-frequency variations (rows 5-7). GI-GS [2] (row 8) uses screen space ray tracing to simulate high-frequency inter-reflections; however, it only models those whose sources are directly visible in the viewpoint. There are also specialized methods for plane mirrors that reflect the point cloud or camera about the plane [19, 22]. Our SpecTRe-GS traces the incident lighting in 3D space, thus is able to render high-frequency and complete inter-reflections for highly specular surfaces with convex shapes. + +Ray-based 3D Gaussian rendering. It is noteworthy that tracing secondary rays for incoming local lighting has almost become a common practice in recent NeRF-based methods that explicitly model specular reflections or interreflections [9, 20, 32, 33, 45]. However, it is challenging to achieve this in 3DGS-based frameworks. 3DGRT [24] + +Table 1. We summarize works improving the view-dependent appearances upon 3DGS [15] based on: How do they model the view-dependent appearances (i.e., the outgoing radiance) and the supported frequency in angular space; how do they accommodate the global illumination effects (i.e., the local incident lighting) including occlusions and inter-reflections and the supported frequency in angular space; and whether is it feasible to edit the scene appearance under (in)direct lighting after reconstruction without any re-baking. + +
MethodView-dependent AppearanceGlobal Illumination EffectEditable Appearance
ModelingFrequencyOcclusionInter-reflectionFrequency
(1) 3DGS [15]Low-order SHLowJointly baked in outgoing SH colorLowNone
(2) 3iGS [30]; [35]Neural networkNet.-boundedJointly baked in the neural featureNet.-boundedNone
(3) DeferredGS [36]IgnoredNone
(4) GShader [13]; [39]Arbitrary, thus bounded by the frequency of incident lightingJointly baked in outgoing SH residualDirect
(5) GS-ID [8]; [40]Physically-based rendering equation [14]Baked as SHIgnoredLow
(6) GS-IR [18]; [10]Baked as SH
(7) GIR [29]Ray-traced in 3D spaceBaked as SH
(8) GI-GS [2]Ray-traced in screen space, incompleteHigh(In)direct
(9) OursRay-traced in 3D spaceHigh(In)direct
+ +first demonstrates efficient ray tracing in 3D Gaussians, enabling ray-based lighting effects such as non-projective views. RayGauss [1] and EVER [21] employ ray sampling in 3D Gaussians, mitigating common artifacts in 3DGS, such as aliasing [41], popping artifacts [27] and projection errors [12]. Yang et al. [46] adopt path tracing to simulate global illumination with Gaussian primitives. Meanwhile, Jorge et al. [4] utilizes ray sampling to simulate volumetric scattering and emissive media. Our SpecTRe-GS integrates ray tracing techniques to evaluate accurate incident radiance in the 3DGS framework at an affordable cost. + +# 3. Proposed Method + +Our method is specifically designed to accurately model the high-frequency reflections of highly specular surfaces. We build upon the recent advancements in 3DGS (Sec. 3.1) and adapt the reflection model to efficiently handle scenes comprising both highly specular objects and rougher objects (Sec. 3.2). To precisely and reliably query the incident radiance necessary for rendering both direct and indirect highly specular reflections, we incorporate ray tracing within the 3DGS framework (Sec. 3.3). We employ a dedicated training strategy to enhance geometry reconstruction, which is crucial for rendering undistorted and realistic reflections (Sec. 3.4). Our pipeline is outlined in Fig. 2. + +# 3.1. Preliminary + +The seminal work of 3DGS [15] represents the scene as a translucent Gaussian point cloud, where each 3D Gaussian is described with central point $\pmb{\mu} \in \mathbb{R}^3$ , scales $\pmb{s} \in \mathbb{R}^3$ , rotation quaternion $\pmb{q} \in \mathbb{R}^4$ , and opacity $\sigma \in [0,1]$ . The distribution of each Gaussian in 3D space is given by $G(\pmb{x}) = e^{-\frac{1}{2} (\pmb{x} - \pmb{\mu})^\top \pmb{\Sigma}^{-1}(\pmb{x} - \pmb{\mu})}$ , where the covariance $\pmb{\Sigma} = \mathbf{RSS}^\top \mathbf{R}^\top$ is compactly parameterized as $\pmb{s}$ and $\pmb{q}$ . By performing local affine transformation [48], 3D Gauss + +sians are projected as 2D Gaussians on the image plane for rasterization. Then for each ray corresponding to an image pixel, the contributing 2D Gaussians are accumulated from front to back for volumetric rendering: + +$$ +I \left(\boldsymbol {\omega} _ {\mathrm {o}}\right) = \sum_ {i} \boldsymbol {c} _ {i} \left(\boldsymbol {\omega} _ {\mathrm {o}}\right) \alpha_ {i} \prod_ {j = 1} ^ {i - 1} \left(1 - \alpha_ {j}\right), \tag {1} +$$ + +where $\alpha_{i} = \sigma_{i}G_{i}$ and the view-dependent color $c_{i}(\omega_{\mathrm{o}})\in$ $\mathbb{R}^3$ is modeled by third-order SH coefficients $\varphi_{i}\in \mathbb{R}^{48}$ + +Besides the view-dependent color $c_{i}$ , many other opt. mizable attributes including surface normal, diffuse albedo, or occlusion information, can be attached to each Gaussian. These attributes can be used locally in each Gaussian, or be accumulated to a map in the same way as Eq. (1) for later applications such as deferred shading [39]. + +# 3.2. Highly Specular Reflection Modeling + +To model high-frequency reflections, we follow the convention to incorporate the physically-based rendering equation [14]. As commonly observed phenomena in real-world life, the inter-reflections from the nearby objects on highly specular surfaces should be modeled for a photorealistic scene appearance. The common practice with distant environment maps [13] ignores such inter-reflections and does not suffice for this task, thus we employ raytracing (detailed in Sec. 3.3) in 3D space for a complete and high-frequency inter-reflection. However, we found that performing ray tracing on the entire scene, including highly specular surfaces and rougher surfaces, is highly time-consuming. Evaluating the rendering equation [14] on rougher surfaces typically needs larger samples of the incident lighting even with multiple importance sampling [31], thus heavily increases the computational cost. Fortunately, previous works [15] have proven that rough surfaces can be well-captured by directly optimizing the outgoing color. Therefore, we propose the following reflection model that treats highly specular surfaces and rougher surfaces sepa + +![](images/90c711e9d2b22af64bed4a863207a89fae0fefe7c5cc400050b466b8a2575f18.jpg) +Figure 2. Overview of the SpecTRe-GS pipeline. The scene is represented as a Gaussian point cloud, where each Gaussian is associated with geometry and shading attributes used to render the depth map $D$ , normal map $N$ , diffuse reflection $I_{\mathrm{diff}}$ , specular reflectance $A_{\mathrm{spec}}$ and rough surface reflection $I_{\mathrm{rs}}$ through splatting (Sec. 3.1). Highly specular surface reflection $I_{\mathrm{ss}}$ is formulated as the sum of the diffuse reflection $I_{\mathrm{diff}}$ and the specular reflection computed from the incident radiance $I_{\mathrm{i}}$ and specular reflectance $A_{\mathrm{spec}}$ (Sec. 3.2). The incident radiance $I_{\mathrm{i}}$ is determined along reflected secondary rays given by the scene geometry $D$ and $N$ , with the indirect component $I_{\mathrm{ind}}$ and visibility $V_{\mathrm{i}}$ evaluated via efficient ray tracing within the Gaussian point cloud, and the direct component $I_{\mathrm{dir}}$ modeled by an estimizable environment map (Sec. 3.3). The scene geometry is enhanced using normal prior guidance, numerical gradients from ray-traced incident radiance, and other training signals to achieve undistorted reflections (Sec. 3.4). + +rately to enhance the specular reflections while maintaining the rendering speed. Also, we follow deferred shading [5] to shade for each pixel instead of each Gaussian, as it reduces the computation and gives sharp surface normals [47]. + +Modeling highly specular surfaces. To model surfaces that may exhibit both highly specular and diffuse reflections, such as polished marble, we define the reflection $I_{\mathrm{ss}}$ for these highly specular surfaces at surface point $x$ , observed from view direction $\omega_{\mathrm{o}}$ as follows: + +$$ +I _ {\mathrm {s s}} (\boldsymbol {x}, \omega_ {\mathrm {o}}) = I _ {\text {d i f f}} (\boldsymbol {x}) + I _ {\text {s p e c}} (\boldsymbol {x}, \omega_ {\mathrm {o}}). \tag {2} +$$ + +While the diffuse reflection $I_{\mathrm{diff}}$ is the view-independent outgoing radiance directly optimized as $d$ in each Gaussian, the specular reflection $I_{\mathrm{spec}}$ is approximated with an ideal mirror reflection: + +$$ +I _ {\operatorname {s p e c}} (\boldsymbol {x}, \omega_ {\mathrm {o}}) = A _ {\operatorname {s p e c}} (\boldsymbol {x}, \omega_ {\mathrm {r}}) I _ {\mathrm {i}} (\boldsymbol {x}, \omega_ {\mathrm {r}}), \tag {3} +$$ + +where the reflected incident direction $\omega_{\mathrm{r}}$ is given by $\omega_{\mathrm{o}} - 2\omega_{\mathrm{o}}^{\top}\pmb{n}\cdot \pmb{n}$ in which $\pmb{n}$ denotes the local surface normal, and $A_{\mathrm{spec}}$ is the specular reflectance considering Fresnel effects through Schlick's approximation [28]: + +$$ +A _ {\mathrm {s p e c}} (\pmb {x}, \pmb {\omega} _ {\mathrm {r}}) = F _ {0} (\pmb {x}) + (1 - F _ {0} (\pmb {x})) (1 - \pmb {n} ^ {\top} \pmb {\omega} _ {\mathrm {r}}) ^ {5}. \quad (4) +$$ + +Here, $F_{0}$ is the specular reflectance at normal incidence optimized as $\pmb{f_0}$ in each Gaussian and is 3-channel to support metallic materials. The incident radiance $I_{\mathrm{i}}$ in $\omega_{\mathrm{r}}$ originates + +from both direct and indirect sources: + +(5) + +$$ +I _ {\mathrm {i}} (\boldsymbol {x}, \omega_ {\mathrm {r}}) = V _ {\mathrm {i}} (\boldsymbol {x}, \omega_ {\mathrm {r}}) I _ {\mathrm {d i r}} (\omega_ {\mathrm {r}}) + (1 - V _ {\mathrm {i}} (\boldsymbol {x}, \omega_ {\mathrm {r}})) I _ {\mathrm {i n d}} (\boldsymbol {x}, \omega_ {\mathrm {r}}), +$$ + +where $V_{\mathrm{i}} \in [0,1]$ quantifies the proportion of the incident lighting occluded by the scene itself, $I_{\mathrm{dir}}$ is the direct incident lighting from a distant environment map, and $I_{\mathrm{ind}}$ is the indirect incident radiance from the scene. + +Defining depth and surface normal of Gaussians. It is essential to accurately acquire the depth and surface normal of each Gaussian in our method, as they directly determine the origin and direction of the reflection. However, they are ambiguous because the Gaussians are translucent ellipsoids in 3D space. Instead of treating the depth of the projected center and the shortest axis direction facing the camera as the depth and normal of each Gaussian [13, 17, 39], we adopt the formulation in GOF [42] to calculate for each Gaussian the "intersection point" along the ray achieving the maximum response, using its depth and normal of the intersection plane. This formulation gives a better and more consistent geometry reconstruction with 3D Gaussians. + +Modeling the whole image. We model the part of the scene containing highly specular surfaces using Eq. (2). For the other part which exhibits only low-frequency view dependencies, we follow GOF [42] to model the appearance $I_{\mathrm{rs}}$ of the rougher surfaces using per-Gaussian SH coefficients $\varphi$ . We combine $I_{\mathrm{ss}}$ and $I_{\mathrm{rs}}$ using a binary mask $M$ (which can be handily acquired through pre-trained segmentation models such as SAM [16]) specifying the highly specular + +surfaces within the scene ( $\odot$ is pixel-wise multiplication): + +$$ +I \left(\omega_ {\mathrm {o}}\right) = I _ {\mathrm {s s}} \left(\omega_ {\mathrm {o}}\right) \odot M + I _ {\mathrm {r s}} \left(\omega_ {\mathrm {o}}\right) \odot (1 - M). \tag {6} +$$ + +# 3.3. Efficient Ray Tracing for Incident Radiance + +To accurately compute the reflection from highly specular surfaces, it is crucial to obtain the incident radiance at the reflected view direction as modeled by Eq. (5), which incorporates both direct and indirect lighting. However, due to the incoherent origins and directions of these secondary rays, the incident radiance cannot be efficiently rendered using the rasterizer from the original 3DGS [15], as it relies on the locality of the rays, nor can it be computed merely using an environment map, which considers only direct lighting [13]. Inspired by the recent work 3DGRT [24], we implement an efficient ray tracer within the 3DGS framework to compute the radiance of secondary rays, thus achieving high-frequency and accurate reflections. + +Geometry acceleration structure (GAS). For fast scene traversal, we construct a GAS within the OptiX programming interface [25], which uses ray tracing hardware to accelerate ray-primitive intersection tests and is several times faster than BVH [3] in CUDA. Instead of using axis-aligned bounding boxes which inefficiently encapsulate irregular Gaussians and lead to false positives, we adopt stretched icosahedrons as proxy meshes for more efficient ray intersection [24]. The vertex $j$ of the icosahedron approximating Gaussian $i$ is given by: + +$$ +\boldsymbol {v} _ {i, j} = \sqrt {2 \log \left(\sigma_ {i} / \alpha_ {\min }\right)} \mathbf {R} _ {i} ^ {\top} \mathbf {S} _ {i} ^ {\top} \boldsymbol {v} _ {j} ^ {\text {u n i t}} + \boldsymbol {\mu} _ {i}, \tag {7} +$$ + +where $\pmb{v}_j^{\mathrm{unit}}$ is the vertex of an inscribed icosahedron in a unit sphere. Here, $\alpha_{\mathrm{min}}$ serves as a threshold that adaptively scales each icosahedron by dictating the minimal contribution of the Gaussian at the transformed vertices, thereby reducing the computational overhead during intersection and radiance evaluation. We regularly update the GAS after each densification operation [15] and every $n_{\mathrm{up}}$ step as the Gaussian point cloud changes along the training. + +Ray queries with GAS. To evaluate the incident radiance $I_{\mathrm{ind}}(\pmb{x}_0, \pmb{\omega}_{\mathrm{r}})$ , secondary rays $\pmb{x}(t) = \pmb{x}_0 + t\pmb{\omega}_{\mathrm{r}}$ are traced against proxy meshes within the GAS. Upon tracing each ray, we first collect the $k$ closest valid hits, each corresponding to a Gaussian with $\alpha_i \geq \alpha_{\min}$ . The depth of the "true" intersection point for each Gaussian is defined as: + +$$ +t _ {\max , i} = \frac {\left(\boldsymbol {\mu} _ {i} - \boldsymbol {x} _ {0}\right) ^ {\top} \boldsymbol {\Sigma} _ {i} ^ {- 1} \boldsymbol {\omega} _ {\mathrm {r}}}{\boldsymbol {\omega} _ {\mathrm {r}} ^ {\top} \boldsymbol {\Sigma} _ {i} ^ {- 1} \boldsymbol {\omega} _ {\mathrm {r}}}, \tag {8} +$$ + +which maximizes the Gaussian's contribution $\sigma_{i}G_{i}(\pmb{x}_{0} + t_{\max,i}\pmb{\omega}_{\mathrm{r}})$ . We then sort the Gaussians by their depth and update the ray depth $D_{\mathrm{ind}}$ , visibility (i.e., transmittance) $V_{\mathrm{i}}$ and color $I_{\mathrm{ind}}$ (using the Gaussians' outgoing SH colors assuming that they model rough surfaces), following Eq. (1). The above process is repeated after advancing the ray origin until no more valid hits remain or the ray transmittance + +drops below a threshold. For the transmittance that remains unoccluded by the reconstructed scene itself, we adopt an estimable environment map for calculating the direct incident radiance $I_{\mathrm{dir}}(\omega)$ . According to Eq. (5), the final incident radiance $I_{\mathrm{i}}(\pmb{x}, \pmb{\omega}_{\mathrm{r}})$ combines direct and indirect components, weighted by visibility $V_{\mathrm{i}}(\pmb{x}, \pmb{\omega})$ . + +Thanks to the efficient and hardware-accelerated ray tracing techniques elaborated above, as well as our separate modeling of highly specular and rough surfaces, the average rendering time of our method is only $50\%$ to $100\%$ slower than that of a pure rasterizer on each of our test scenes. This varies according to the density of objects in the scene and the proportion of highly specular surfaces. + +# 3.4. Training + +We propose a training strategy that especially focuses on enhancing the geometry reconstruction, as unfaithful scene geometry induces inaccurate secondary rays and results in distorted reflections. We optimize our model from an initial sparse point cloud using the following loss terms: + +$$ +\mathcal {L} = \mathcal {L} _ {\mathrm {c}} + \lambda_ {\mathrm {d}} \mathcal {L} _ {\mathrm {d}} + \lambda_ {\mathrm {d n}} \mathcal {L} _ {\mathrm {d n}} + \lambda_ {\mathrm {s m}} \mathcal {L} _ {\mathrm {s m}} + \lambda_ {\mathrm {n}} \mathcal {L} _ {\mathrm {n}} + \lambda_ {\mathrm {r}} \mathcal {L} _ {\mathrm {r}}, \tag {9} +$$ + +where $\mathcal{L}_{\mathrm{c}}$ is a color reconstruction loss [15] differentiable to the geometry, $\mathcal{L}_{\mathrm{d}}$ and $\mathcal{L}_{\mathrm{dn}}$ are the distortion and depth-normal consistency loss used in GOF [42], $\mathcal{L}_{\mathrm{sm}}$ is an edge-aware smoothness term for the normal and specular reflectance for the highly specular surfaces, $\mathcal{L}_{\mathrm{n}}$ is a normal prior term to employ data-driven clues, and $\mathcal{L}_{\mathrm{r}}$ is a residual-suppressing term for progressive learning. This section describes these terms and more details are in the supplement. + +Normal prior guidance. To tackle the challenging problem of reconstructing geometry for highly specular objects, we employ StableNormal [38] to provide monocular normal priors $\hat{N}$ , offering initial geometric constraints: + +$$ +\mathcal {L} _ {\mathrm {n}} = 1 / | N | \sum \left| \left| (\hat {N} - N) \odot M \right| \right| _ {1}. \tag {10} +$$ + +Joint geometry optimization. Recognizing potential inaccuracies and inconsistencies across views in the normal prior, we concurrently optimize the scene geometry during training. This is achieved by differentiating the incident radiance $I_{\mathrm{i}}$ , which relates to the final image through Eq. (2) and Eq. (3), with respect to the scene geometry. Computing the gradient of the direct incident radiance $I_{\mathrm{dir}}$ is straightforward. To compute the gradient of the indirect incident radiance $I_{\mathrm{ind}}$ concerning the local surface normal $n$ , we adopt a simple yet effective finite difference method to numerically evaluate the gradient of $I_{\mathrm{ind}}$ with respect to the reflected direction $\omega$ that depends on $n$ : + +$$ +\left. \frac {\partial I _ {\text {i n d}}}{\partial \omega^ {i}} \right| _ {\omega_ {0}} \approx \frac {I _ {\text {i n d}} \left(\boldsymbol {x} , \omega_ {0} + \Delta_ {\omega} ^ {i}\right) - I _ {\text {i n d}} \left(\boldsymbol {x} , \omega_ {0} - \Delta_ {\omega} ^ {i}\right)}{2 \Delta_ {\omega} ^ {i}}, \tag {11} +$$ + +where $\Delta_{\omega}^{i}$ denotes the perturbation of the $i$ -th dimension in $\omega$ . Considering that using the same perturbation for sec- + +Table 2. Quantitative comparisons of novel view synthesis results with state-of-the-art methods on both synthetic and real-world scenes. We report the scores of PSNR, SSIM [34], and LPIPS [44] for entire images and within reflective regions. $\uparrow (\downarrow)$ means higher (lower) is better. We mark the best and the second best results in each column. + +
DataSynthetic (entire image)Synthetic (reflective)Real-world (entire image)Real-world (reflective)
MethodPSNR↓SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↓SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
3DGS [15]27.710.9100.13121.510.9400.07824.820.8230.22721.310.9650.047
GOF [42]27.970.9170.11721.380.9400.07624.860.8260.21921.250.9630.047
GOF* [42]27.570.9130.12520.670.9360.08524.690.8250.22320.440.9620.051
3iGS [30]28.580.9120.12521.980.9390.07624.370.8090.23321.220.9630.048
GShader [13]26.650.8880.16020.790.9350.08623.930.8100.25520.230.9630.053
GShader* [13]27.140.8940.15321.220.9380.08424.030.8120.25320.360.9640.053
3DGS-DR [39]27.620.8900.17321.570.9380.08524.790.8240.24321.180.9650.050
3DGS-DR* [39]26.970.8830.18221.180.9360.08824.600.8230.25120.380.9640.055
Ours29.410.9220.10923.280.9490.06225.110.8280.21322.770.9670.035
+ +ondary rays with different depths $D_{\mathrm{ind}}$ leads to different amounts of intersection movements, making the numerical gradient less accurate, we use a depth-aware ray perturbation that adjusts to $D_{\mathrm{ind}}$ for more informative gradients: + +$$ +\Delta_ {\omega} ^ {i} = \theta_ {\mathrm {d}} \max \left(\min \left(D _ {\text {i n d}} / D _ {0}, \sigma_ {\mathrm {d}} ^ {\max }\right), \sigma_ {\mathrm {d}} ^ {\min }\right), \tag {12} +$$ + +where $\theta_{\mathrm{d}}$ and $D_0$ define a reference angular perturbation and $\sigma_{\mathrm{d}}^{\mathrm{min}}$ and $\sigma_{\mathrm{d}}^{\mathrm{max}}$ ensure numerical stability. For computational efficiency, we disregard the gradient with respect to the point position $\pmb{p}$ , as we find reflections are distorted mainly due to normal errors. We detach visibility $V_{\mathrm{ind}}$ and depth $D_{\mathrm{ind}}$ because the sparsity of the visibility gradient and abrupt changes in depth damage numerical stability. + +Progressive learning. As training progresses, monocular normal priors may hinder the accurate geometry reconstruction, thus we gradually decay the weight of normal prior guidance $\lambda_{\mathrm{n}}$ , letting gradients from physics-based rendering take over the fine-tuning of scene geometry. Besides, for smoother training and consequently better reconstruction within the reflective regions, we replace the hard mask $M$ used in Eq. (6) with a relaxed version $1 - I_{\mathrm{r}}$ rendered with another per-Gaussian attribute $r \in [0,1]$ using Eq. (1). We encourage the reflective region to be modeled with highly specular reflections by suppressing the residual within $M$ : + +$$ +\mathcal {L} _ {\mathrm {r}} = 1 / | M | \sum \left| \left| I _ {\mathrm {r}} \odot M \right| \right| _ {1}. \tag {13} +$$ + +# 4. Experiments + +In this section, we compare our method on both synthetic and real-world scenes (Sec. 4.2) against state-of-the-art methods (Sec. 4.3). We also perform ablation studies to validate the key components of our method (Sec. 4.4) and explore applications in scene editing (Sec. 4.5). Please check the supplementary material for more implementation details, data creation details, and experimental results. + +# 4.1. Implementation Details + +We implement our method in PyTorch [26] with CUDA kernels for rasterization based on GOF [42] and custom OptiX programs for ray tracing. We conduct all experiments on a + +GPU with 24 GB VRAM, 83 shader core TFLOPS, and 191 ray tracing core TFLOPS. We set $\lambda_{\mathrm{sm}} = 0.5$ , $\lambda_{\mathrm{r}} = 1.0$ , and $\lambda_{\mathrm{n}} = 0.5$ , exponentially decaying to 0.0025 from 4k to 10k iterations. We use $r = 0.9$ as the initialization. For other hyperparameters, we use the settings in GOF [42]. + +# 4.2. Datasets + +Synthetic scenes. We build 6 synthetic scenes, each containing multiple surrounding objects and a highly specular surface with distinctive properties: HELMET, MARBLETABLE, VASE, POT, TOASTER, and MIRROR. For each scene, we use the Blender Cycles engine [6] to render about 300 images of resolution $1024 \times 1024$ , of which 2/3 are used for training and the rest are left for testing. + +Real-world scenes. For real-world validation, we capture two scenes: REALBOWL and REALPOT, consisting of 138 and 139 images of resolution $1440 \times 1080$ respectively, with every eighth image reserved in the test set. + +# 4.3. Comparison + +Comparing methods. We compare our method with 5 open-source state-of-the-art methods: the original 3DGS [15]; GOF [42] which focus on geometry reconstruction with 3D Gaussians; 3iGS [30] which improves view-dependent effects through neural features; and 3DGS-DR [39] and GShader [13] which are adapted for specular reflection employing physically-based rendering (PBR). For a fair comparison, we provide GOF [42], 3DGS-DR [39], and GShader [13] with the normal prior guidance, and indicate these versions by “*” throughout this paper. + +Evaluation protocol. We report peak signal-to-noise ratio (PSNR), structural similarity index measure (SSIM) [34], and learned perceptual image patch similarity (LPIPS) [44] measured on both entire images and reflective regions. + +Quantitative evaluation. As shown in Tab. 2, our method outperforms comparing methods on both synthetic and real-world scenes, especially in reflective regions. It is worth noting that using normal prior guidance usually leads to degradation of GOF [42] which relies solely on SH color, as it suppresses the ability to simulate high-frequency appear + +![](images/a6a031519f5cb53855c07d3865df6f09606179703e6a7e213ace45fdb1793c05.jpg) +Figure 3. Novel view synthesis comparison on synthetic scenes. We evaluate SpecTRe-GS against relevant 3DGS-based methods modeling view-dependent appearances: 3DGS-DR [39], GShader [13], 3iGS [30], GOF [42], and 3DGS [15]. The insets highlight the improvement of SpecTRe-GS in rendering high-frequency specular reflections, showcasing the precise positioning of reflected content and sharp details. + +![](images/7baf3e4fd74eccd4a0c84a39c794fa03898ad53ef0e6b00498dcf536a0bc35ee.jpg) +Figure 4. Novel view synthesis and image decomposition results on synthetic scenes, comparing our SpecTRe-GS with 3DGS-DR [39] and GShader [13]. SpecTRe-GS better distinguishes diffuse and specular components in reflective regions, with the least baked-in view-dependent appearance in diffuse components. The masks of reflective regions are shown as insets in the reference image. + +ances through false geometry. In contrast, the influence of this guidance on 3DGS-DR [39] and GShader [13] is twofold: It facilitates the reconstruction of appearance under direct lighting, but also prevents these methods from imitating specular inter-reflections by faking geometry. + +Qualitative evaluation. In visual comparisons, we show results of GOF [42], GShader* [13], and 3DGS-DR [39] as they are variants with higher qualities. As shown in Fig. 3, our method performs better than the baseline methods for both planar reflectors (rows 1-2) and curved reflectors (row 3), capturing clean, accurate, and high-frequency specular + +reflections. Our method also surpasses baseline methods in real-world scenes, as shown in Fig. 1. By visualizing the separate components of PBR-based methods in Fig. 4, we find that incident radiance modeling for both direct and indirect sources is vital for physically meaningful scene appearance decomposition. Otherwise, to fit the overall color in training images, indirect incident radiance is heavily baked into the diffuse component (row 1 for 3DGS-DR [39] and row 3 for GShader [13]). The baked-in incident radiance loses high-frequency details and prevents realistic scene editing in a physically-accurate manner. It should be noted + +![](images/dbb97ada72e67ea5c45c0be24a591082a042ae59defeb6ae0acd8626bcb09d38.jpg) +Figure 5. Our reconstruction results on two real-world scenes. SpecTRe-GS accurately reconstructs the geometry of highly specular surfaces alongside the reflection of nearby objects. + +that 3DGS-DR [39] performs better than GShader [13] in reconstructing the direct part of incident radiance from the environment, showing the superiority of surface-based deferred shading over volumetric rendering in PBR modeling of high-frequency reflections. Our method, with well-optimized geometry and consequent accurate incident radiance modeling, achieves satisfactory novel view synthesis of specular reflection on both synthetic and real-world scenes, as also indicated in Fig. 5. + +# 4.4. Ablation Study + +On a subset of our synthetic scenes, we validate the following key component of our method: (1) Normal prior guidance; (2) Joint geometry optimization from incident lighting; (3) ray-traced Indirect incident radiance; (4) Progressive learning; and (5) Depth-aware ray perturbation. The results are shown in Tab. 3. Our full method (Ours) achieves the highest scores, confirming the effectiveness of our method designs. Without the initial normal prior guidance (Ours w/o N.), the optimization process is prone to be trapped into local minima far away from the true underneath geometry of the reflective region. On the other hand, only relying on normal priors without further joint optimization for fine-tuning geometry (Ours w/o J.) also fails in reconstructing accurate geometry for physically-based high-frequency reflections. As discussed above, using only direct incident lighting modeling (Ours w/o I.) leads to baked-in indirect incident lighting and damages the quality of high-frequency components. The progressive learning scheme (Ours w/o P.) and depth-aware ray perturbation (Ours w/o D.) also further helps to faithfully reconstruct the scene. + +# 4.5. Applications + +Our Gaussian point cloud representation with explicit geometry, material, and incident lighting facilitates various + +Table 3. Ablation study with variants of our proposed method excluding: Normal prior guidance, Joint geometry optimization, Indirect incident lighting modeling, Progressive learning, and Depth-aware ray perturbation. $\uparrow (\downarrow)$ means higher (lower) is better. + +
MethodEntire ImageReflective Region
PSNR↓SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
Ours w/o N.26.700.8850.14919.800.9130.094
Ours w/o J.26.370.8780.15419.520.9070.100
Ours w/o I.28.490.9050.13022.190.9350.074
Ours w/o P.29.260.9100.11823.160.9400.062
Ours w/o D.29.750.9140.11423.880.9440.056
Ours29.900.9140.11224.050.9440.056
+ +![](images/194d38599ece68cb60acf280c4538d4f287a2ddc648f9cf9cfff1183081baa16.jpg) +Figure 6. Our explicit representation of geometry, material, and incident lighting in the Gaussian point cloud enables various editing applications. Insets have been brightened for clarity. + +editing applications. As illustrated in Fig. 6, specular objects of interest can be faithfully captured as assets in the presence of unwanted reflections from nearby objects, enabling further usages such as relighting (column 1) and material editing (column 4). We can also manipulate the scene geometry such as deletion (column 2) and insertion (column 3), with specular inter-reflections adjusted accordingly. + +# 5. Conclusion + +We introduce SpecTRe-GS to model highly specular surfaces with inter-reflections, distinguished by its separate modeling of specular and rough surfaces, the efficient ray tracer, and a geometry-enhancing training strategy for fast and accurate specular inter-reflection rendering. + +Limitations. We only model specular inter-reflections from rough surfaces to avoid multi-bounce path tracing, and we only trace one secondary ray for each pixel in the reflective region, approximating the specular reflection with an ideal mirror. Also, the monocular normal prior may fail in the existence of strong semantic ambiguity (such as mirrors with hidden borders), making our geometry fail to initialize. + +# Acknowledgments + +This work is supported by the National Natural Science Foundation of China under Grant No. 62136001. PKU-affiliated authors thank openbayes.com for providing computing resource. + +# References + +[1] Hugo Blanc, Jean-Emmanuel Deschaud, and Alexis Paljic. RayGauss: Volumetric Gaussian-based ray casting for photorealistic novel view synthesis. In Proc. of IEEE Winter Conference on Applications of Computer Vision (WACV), 2024. 3 +[2] Hongze Chen, Zehong Lin, and Jun Zhang. GI-GS: Global illumination decomposition on Gaussian splatting for inverse rendering. In Proc. of International Conference on Learning Representations (ICLR), 2025. 1, 2, 3 +[3] James H. Clark. Hierarchical geometric models for visible surface algorithms. Communications of the ACM (CACM), 19(10):547-554, 1976. 5 +[4] Jorge Condor, Sebastien Speierer, Lukas Bode, Aljaz Bozic, Simon Green, Piotr Didyk, and Adrian Jarabo. Don't splat your Gaussians: Volumetric ray-traced primitives for modeling and rendering scattering and emissive media. ACM Transactions on Graphics (TOG), 44(1):10:1-10:17, 2025. 3 +[5] Michael Deering, Stephanie Winner, Bic Schediwy, Chris Duffy, and Neil Hunt. The triangle processor and normal vector Shader: a VLSI system for high performance graphics. In Proc. of the ACM SIGGRAPH Conference and Exhibition On Computer Graphics and Interactive Techniques (SIGGRAPH), 1988. 4 +[6] Cycles Developers. Cycles: Open source production rendering, 2024. https://www.cycles-renderer.org/. Accessed: 2024-11-06. 6 +[7] Jan-Niklas Dihlmann, Arjun Majumdar, Andreas Engelhardt, Raphael Braun, and Hendrik P. A. Lensch. Subsurface scattering for 3D Gaussian splattering. In Proc. of Neural Information Processing Systems (NeurIPS), 2024. 1 +[8] Kang Du, Zhihao Liang, and Zeyu Wang. GS-ID: Illumination decomposition on Gaussian splattering via diffusion prior and parametric light source optimization. In arXiv preprint arXiv:2408.08524, 2024. 1, 2, 3 +[9] Fan Fei, Jiajun Tang, Ping Tan, and Boxin Shi. VMINer: Versatile multi-view inverse rendering with near- and farfield light sources. In Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2 +[10] Jian Gao, Chun Gu, Youtian Lin, Hao Zhu, Xun Cao, Li Zhang, and Yao Yao. Relightable 3D Gaussian: Real-time point cloud relighting with BRDF decomposition and ray tracing. In Proc. of European Conference on Computer Vision (ECCV), 2024. 1, 2, 3 +[11] Yijia Guo, Yuanxi Bai, Liwen Hu, Ziyi Guo, Mianzhi Liu, Yu Cai, Tiejun Huang, and Lei Ma. PRTGS: Precomputed radiance transfer of Gaussian splats for real-time high-quality relighting. In Proc. of ACM International Conference on Multimedia (ACM MM), 2024. 1 +[12] Letian Huang, Jiayang Bai, Jie Guo, Yuanqi Li, and Yanwen Guo. On the error analysis of 3D Gaussian splattering and an optimal projection strategy. In Proc. of European Conference on Computer Vision (ECCV), 2024. 3 +[13] Yingwenqi Jiang, Jiadong Tu, Yuan Liu, Xifeng Gao, Xiaoxiao Long, Wenping Wang, and Yuexin Ma. GaussianShader: + +3D Gaussian splattering with shading functions for reflective surfaces. In Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 1, 2, 3, 4, 5, 6, 7, 8 +[14] James T. Kajiya. The rendering equation. In Proc. of the ACM SIGGRAPH Conference and Exhibition On Computer Graphics and Interactive Techniques (SIGGRAPH), 1986. 2, 3 +[15] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3D Gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics (TOG), 42(4):139:1-139:14, 2023. 1, 2, 3, 5, 6, 7 +[16] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dollar, and Ross B. Girshick. Segment anything. In Proc. of IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 4 +[17] Jiahe Li, Jiawei Zhang, Xiao Bai, Jin Zheng, Xin Ning, Jun Zhou, and Lin Gu. DNGaussian: Optimizing sparse-view 3D Gaussian radiance fields with global-local depth normalization. In Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 1, 4 +[18] Zhihao Liang, Qi Zhang, Ying Feng, Ying Shan, and Kui Jia. GS-IR: 3D Gaussian splatting for inverse rendering. In Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 1, 2, 3 +[19] Jiayue Liu, Xiao Tang, Freeman Cheng, Roy Yang, Zhihao Li, Jianzhuang Liu, Yi Huang, Jiaqi Lin, Shiyong Liu, Xiaofei Wu, Songcen Xu, and Chun Yuan. MirrorGaussian: Reflecting 3D Gaussians for reconstructing mirror reflections. In Proc. of European Conference on Computer Vision (ECCV), 2024. 2 +[20] Yuan Liu, Peng Wang, Cheng Lin, Xiaoxiao Long, Jiepeng Wang, Lingjie Liu, Taku Komura, and Wenping Wang. NeRO: Neural geometry and BRDF reconstruction of reflective objects from multiview images. ACM Transactions on Graphics (TOG), 42(4):114:1-114:22, 2023. 2 +[21] Alexander Mai, Peter Hedman, George Kopanas, Dor Verbin, David Futschik, Qiangeng Xu, Falko Kuester, Jonathan T. Barron, and Yinda Zhang. EVER: Exact volumetric ellipsoid rendering for real-time view synthesis. In arXiv preprint arXiv:2410.01804, 2024. 3 +[22] Jiarui Meng, Haijie Li, Yanmin Wu, Qiankun Gao, Shuzhou Yang, Jian Zhang, and Siwei Ma. Mirror-3DGS: Incorporating mirror reflections into 3D Gaussian splattering. In Proc. of IEEE International Conference on Visual Communications and Image Processing, 2024. 2 +[23] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing scenes as neural radiance fields for view synthesis. In Proc. of European Conference on Computer Vision (ECCV), 2020. 1 +[24] Nicolas Moënne-Loccoz, Ashkan Mirzaei, Or Perel, Ricardo de Lutio, Janick Martinez Esturo, Gavriel State, Sanja Fidler, Nicholas Sharp, and Zan Gojcic. 3D Gaussian ray tracing: Fast tracing of particle scenes. ACM Transactions on Graphics (TOG), 43(6):232:1-232:19, 2024. 2, 5 + +[25] Steven G. Parker, James Bigler, Andreas Dietrich, Heiko Friedrich, Jared Hoberock, David P. Luebke, David K. McAllister, Morgan McGuire, R. Keith Morley, Austin Robison, and Martin Stich. OptiX: a general purpose ray tracing engine. In ACM Transactions on Graphics (Proc. of ACM SIGGRAPH), 2010. 5 +[26] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Köpf, Edward Z. Yang, Zachary DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. PyTorch: An imperative style, high-performance deep learning library. In Proc. of Neural Information Processing Systems (NeurIPS), 2019. 6 +[27] Lukas Radl, Michael Steiner, Mathias Parger, Alexander Weinrauch, Bernhard Kerbl, and Markus Steinberger. StopThePop: Sorted Gaussian splatting for view-consistent real-time rendering. ACM Transactions on Graphics (TOG), 43(4):64:1-64:17, 2024. 3 +[28] Christophe Schlick. An inexpensive BRDF model for physically-based rendering. Computer Graphics Forum (CGF), 13(3):233-246, 1994. 4 +[29] Yahao Shi, Yanmin Wu, Chenming Wu, Xing Liu, Chen Zhao, Haocheng Feng, Jingtuo Liu, Liangjun Zhang, Jian Zhang, Bin Zhou, Errui Ding, and Jingdong Wang. GIR: 3D Gaussian inverse rendering for reightable scene factorization. In arXiv preprint arXiv:2312.05133, 2023. 1, 2, 3 +[30] Zhe Jun Tang and Tat-Jen Cham. 3iGS: Factorised tensorial illumination for 3D Gaussian splatting. In Proc. of European Conference on Computer Vision (ECCV), 2024. 1, 2, 3, 6, 7 +[31] Eric Veach and Leonidas J. Guibas. Optimally combining sampling techniques for Monte Carlo rendering. In Proc. of the ACM SIGGRAPH Conference and Exhibition On Computer Graphics and Interactive Techniques (SIGGRAPH), 1995. 3 +[32] Dor Verbin, Peter Hedman, Ben Mildenhall, Todd Zickler, Jonathan T. Barron, and Pratul P. Srinivasan. Ref-NeRF: Structured view-dependent appearance for neural radiance fields. In Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2 +[33] Dor Verbin, Pratul P. Srinivasan, Peter Hedman, Ben Mildenhall, Benjamin Attal, Richard Szeliski, and Jonathan T. Barron. NeRF-Casting: Improved view-dependent appearance with consistent reflections. In Proc. of the ACM SIGGRAPH Conference and Exhibition on Computer Graphics and Interactive Techniques in Asia (SIGGRAPH Asia), 2024. 2 +[34] Zhou Wang, A.C. Bovik, H.R. Sheikh, and E.P. Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing (TIP), 13(4):600-612, 2004. 6 +[35] Zhiru Wang, Shiyun Xie, Chengwei Pan, and Guoping Wang. SpecGaussian with latent features: A high-quality modeling of the view-dependent appearance for 3D Gaussian splatting. In Proc. of ACM International Conference on Multimedia (ACM MM), 2024. 1, 2, 3 +[36] Tong Wu, Jia-Mu Sun, Yu-Kun Lai, Yuewen Ma, Leif Kobbelt, and Lin Gao. DeferredGS: Decoupled and editable + +Gaussian splatting with deferred shading. In arXiv preprint arXiv:2404.09412, 2024. 1, 2, 3 +[37] Wangze Xu, Huachen Gao, Shihe Shen, Rui Peng, Jianbo Jiao, and Ronggang Wang. MVPGS: Excavating multi-view priors for Gaussian splatting from sparse input views. In Proc. of European Conference on Computer Vision (ECCV), 2024. 1 +[38] Chongjie Ye, Lingteng Qiu, Xiaodong Gu, Qi Zuo, Yushuang Wu, Zilong Dong, Liefeng Bo, Yuliang Xiu, and Xiaoguang Han. StableNormal: Reducing diffusion variance for stable and sharp normal. In Proc. of the ACM SIGGRAPH Conference and Exhibition on Computer Graphics and Interactive Techniques in Asia (SIGGRAPH Asia), 2024. 5 +[39] Keyang Ye, Qiming Hou, and Kun Zhou. 3D Gaussian splating with deferred reflection. In Proc. of the ACM SIGGRAPH Conference and Exhibition On Computer Graphics and Interactive Techniques (SIGGRAPH), 2024. 1, 2, 3, 4, 6, 7, 8 +[40] Keyang Ye, Qiming Hou, and Kun Zhou. Progressive radiance distillation for inverse rendering with Gaussian splatting. In arXiv preprint arXiv:2408.07595, 2024. 1, 2, 3 +[41] Zehao Yu, Anpei Chen, Binbin Huang, Torsten Sattler, and Andreas Geiger. Mip-Splatting: Alias-free 3D Gaussian splatting. In Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 3 +[42] Zehao Yu, Torsten Sattler, and Andreas Geiger. Gaussian opacity fields: Efficient adaptive surface reconstruction in unbounded scenes. ACM Transactions on Graphics (TOG), 43(6):271:1-271:13, 2024. 1, 2, 4, 5, 6, 7 +[43] Libo Zhang, Yuxuan Han, Wenbin Lin, Jingwang Ling, and Feng Xu. PRTGaussian: Efficient relighting using 3D Gaussians with precomputed radiance transfer. In Asia Pacific Signal and Information Processing Association Annual Summit and Conference (APSIPA ASC), 2024. 1 +[44] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 6 +[45] Yuanqing Zhang, Jiaming Sun, Xingyi He, Huan Fu, Rongfei Jia, and Xiaowei Zhou. Modeling indirect illumination for inverse rendering. In Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2 +[46] Yang Zhou, Songyin Wu, and Ling-Qi Yan. Unified Gaussian primitives for scene representation and rendering. In arXiv preprint arXiv:2406.09733, 2024. 3 +[47] Zuo-Liang Zhu, Beibei Wang, and Jian Yang. GS-ROR: 3D Gaussian splatting for reflective object relighting via SDF priors. In arXiv preprint arXiv:2406.18544, 2024. 1, 2, 4 +[48] Matthias Zwicker, Hanspeter Pfister, Jeroen van Baar, and Markus H. Gross. EWA volume splatting. In Proc. of IEEE Visualization Conference (VIS), 2001. 3 \ No newline at end of file diff --git a/2025/SpecTRe-GS_ Modeling Highly Specular Surfaces with Reflected Nearby Objects by Tracing Rays in 3D Gaussian Splatting/images.zip b/2025/SpecTRe-GS_ Modeling Highly Specular Surfaces with Reflected Nearby Objects by Tracing Rays in 3D Gaussian Splatting/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..1a0028a66a061c9f0033cab6fb31ae8ed4b1d1f3 --- /dev/null +++ b/2025/SpecTRe-GS_ Modeling Highly Specular Surfaces with Reflected Nearby Objects by Tracing Rays in 3D Gaussian Splatting/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5b456e47e9580e2e495c3998c16006aabafd0cadb53554071c333cb25b0b2067 +size 916953 diff --git a/2025/SpecTRe-GS_ Modeling Highly Specular Surfaces with Reflected Nearby Objects by Tracing Rays in 3D Gaussian Splatting/layout.json b/2025/SpecTRe-GS_ Modeling Highly Specular Surfaces with Reflected Nearby Objects by Tracing Rays in 3D Gaussian Splatting/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..6988ddc7414f65f757389ce6679458de257083c2 --- /dev/null +++ b/2025/SpecTRe-GS_ Modeling Highly Specular Surfaces with Reflected Nearby Objects by Tracing Rays in 3D Gaussian Splatting/layout.json @@ -0,0 +1,8820 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 56, + 102, + 555, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 102, + 555, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 102, + 555, + 140 + ], + "type": "text", + "content": "SpecTRe-GS: Modeling Highly Specular Surfaces with Reflected Nearby Objects by Tracing Rays in 3D Gaussian Splatting" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 160, + 201, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 160, + 201, + 175 + ], + "spans": [ + { + "bbox": [ + 132, + 160, + 201, + 175 + ], + "type": "text", + "content": "Jiajun Tang" + }, + { + "bbox": [ + 132, + 160, + 201, + 175 + ], + "type": "inline_equation", + "content": "^{1,2}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 175, + 172, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 106, + 175, + 172, + 189 + ], + "spans": [ + { + "bbox": [ + 106, + 175, + 172, + 189 + ], + "type": "text", + "content": "Youyu Chen" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 213, + 161, + 262, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 213, + 161, + 262, + 175 + ], + "spans": [ + { + "bbox": [ + 213, + 161, + 262, + 175 + ], + "type": "text", + "content": "Fan Fei" + }, + { + "bbox": [ + 213, + 161, + 262, + 175 + ], + "type": "inline_equation", + "content": "^{1,2}" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 186, + 175, + 264, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 186, + 175, + 264, + 190 + ], + "spans": [ + { + "bbox": [ + 186, + 175, + 264, + 190 + ], + "type": "text", + "content": "Binxiao Huang" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 274, + 161, + 329, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 274, + 161, + 329, + 175 + ], + "spans": [ + { + "bbox": [ + 274, + 161, + 329, + 175 + ], + "type": "text", + "content": "Zhihao Li3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 277, + 175, + 348, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 277, + 175, + 348, + 189 + ], + "spans": [ + { + "bbox": [ + 277, + 175, + 348, + 189 + ], + "type": "text", + "content": "Zhenyu Chen" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 342, + 161, + 398, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 342, + 161, + 398, + 175 + ], + "spans": [ + { + "bbox": [ + 342, + 161, + 398, + 175 + ], + "type": "text", + "content": "Xiao Tang3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 364, + 175, + 426, + 188 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 364, + 175, + 426, + 188 + ], + "spans": [ + { + "bbox": [ + 364, + 175, + 426, + 188 + ], + "type": "text", + "content": "Xiaofei Wu3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 427, + 161, + 477, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 427, + 161, + 477, + 175 + ], + "spans": [ + { + "bbox": [ + 427, + 161, + 477, + 175 + ], + "type": "text", + "content": "lyong Liu3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 440, + 175, + 503, + 188 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 440, + 175, + 503, + 188 + ], + "spans": [ + { + "bbox": [ + 440, + 175, + 503, + 188 + ], + "type": "text", + "content": "Boxin Shi" + }, + { + "bbox": [ + 440, + 175, + 503, + 188 + ], + "type": "inline_equation", + "content": "^{1,2\\#}" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 82, + 190, + 528, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 190, + 528, + 203 + ], + "spans": [ + { + "bbox": [ + 82, + 190, + 528, + 203 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 82, + 190, + 528, + 203 + ], + "type": "text", + "content": "State Key Laboratory for Multimedia Information Processing, School of Computer Science, Peking University" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 82, + 203, + 526, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 203, + 526, + 213 + ], + "spans": [ + { + "bbox": [ + 82, + 203, + 526, + 213 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 82, + 203, + 526, + 213 + ], + "type": "text", + "content": "National Engineering Research Center of Visual Technology, School of Computer Science, Peking University" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 124, + 213, + 484, + 225 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 124, + 213, + 484, + 225 + ], + "spans": [ + { + "bbox": [ + 124, + 213, + 484, + 225 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 124, + 213, + 484, + 225 + ], + "type": "text", + "content": "Huawei Noah's Ark Lab " + }, + { + "bbox": [ + 124, + 213, + 484, + 225 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 124, + 213, + 484, + 225 + ], + "type": "text", + "content": "Harbin Institute of Technology " + }, + { + "bbox": [ + 124, + 213, + 484, + 225 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 124, + 213, + 484, + 225 + ], + "type": "text", + "content": "University of Hong Kong" + } + ] + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 57, + 237, + 553, + 365 + ], + "blocks": [ + { + "bbox": [ + 57, + 237, + 553, + 365 + ], + "lines": [ + { + "bbox": [ + 57, + 237, + 553, + 365 + ], + "spans": [ + { + "bbox": [ + 57, + 237, + 553, + 365 + ], + "type": "image", + "image_path": "212d433fd9ee9f5cce3729380e35988aa00901f5e7a9731a1b9ac80863b7f2f7.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 367, + 555, + 401 + ], + "lines": [ + { + "bbox": [ + 55, + 367, + 555, + 401 + ], + "spans": [ + { + "bbox": [ + 55, + 367, + 555, + 401 + ], + "type": "text", + "content": "Figure 1. Our SpecTRe-GS effectively captures the appearance of reflectors with highly specular surfaces, even in the presence of nearby objects. In the real-world scene shown above, it performs favorably against previous methods [13, 15, 30, 39, 42], both visually (as the highlighted reflections within the insets) and quantitatively (with PSNR values within the reflective region annotated in the corners)." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "bbox": [ + 152, + 410, + 200, + 422 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 410, + 200, + 422 + ], + "spans": [ + { + "bbox": [ + 152, + 410, + 200, + 422 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 54, + 434, + 297, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 434, + 297, + 687 + ], + "spans": [ + { + "bbox": [ + 54, + 434, + 297, + 687 + ], + "type": "text", + "content": "3D Gaussian Splatting (3DGS), a recently emerged multiview 3D reconstruction technique, has shown significant advantages in real-time rendering and explicit editing. However, 3DGS encounters challenges in the accurate modeling of both high-frequency view-dependent appearances and global illumination effects, including inter-reflection. This paper introduces SpecTRe-GS, which addresses these challenges and models highly Specular surfaces that reflect nearby objects through Tracing Rays in 3D Gaussian Splatting. SpecTRe-GS separately models reflections from highly specular and rough surfaces to leverage the distinctions between their reflective properties and integrates an efficient ray tracer within the 3DGS framework for querying secondary rays, thus achieving fast and accurate rendering. Also, it incorporates normal prior guidance and joint geometry optimization at various stages of the training process to enhance geometry reconstruction for undistorted reflections. Experiments on both synthetic and real-world scenes demonstrate the superiority of SpecTRe-GS compared to existing 3DGS-based methods in capturing highly specular inter-reflections and also showcase its editing applications." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 315, + 410, + 394, + 422 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 410, + 394, + 422 + ], + "spans": [ + { + "bbox": [ + 315, + 410, + 394, + 422 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 434, + 555, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 434, + 555, + 662 + ], + "spans": [ + { + "bbox": [ + 313, + 434, + 555, + 662 + ], + "type": "text", + "content": "Recently, 3D Gaussian Splatting (3DGS) [15] has emerged as an innovative paradigm for multi-view 3D reconstruction. Unlike its predecessor, neural radiance fields (NeRF) [23], 3DGS adopts an explicit point-based scene representation with tile-based rasterization rather than relying on ray casting through an implicit neural field for rendering. This transition endows 3DGS with significant advantages in terms of real-time rendering speed and allows for convenient post-reconstruction editing. Nevertheless, challenges arise when the scene to be reconstructed includes highly specular (i.e., mirror-like) surfaces which exhibit strong and high-frequency view-dependent appearances. The original 3DGS struggles to accurately model such high-frequency reflections, as it utilizes per-Gaussian third-order spherical harmonics (SH) to represent view-dependent appearances. This approach is restricted to low-frequency radiance variations in angular space and is susceptible to overfitting when the training views that observe the Gaussian are not sufficiently dense [17, 37]." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 662, + 556, + 711 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 662, + 556, + 711 + ], + "spans": [ + { + "bbox": [ + 313, + 662, + 556, + 711 + ], + "type": "text", + "content": "To address this issue, a natural strategy is to adopt a physically-based shading model and optimize reflective properties for each Gaussian [2, 7, 8, 10, 11, 13, 18, 29, 35, 36, 39, 40, 43, 47], rather than merely recording the out" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 69, + 693, + 238, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 693, + 238, + 703 + ], + "spans": [ + { + "bbox": [ + 69, + 693, + 238, + 703 + ], + "type": "text", + "content": "The work was done during an internship in Huawei." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 69, + 703, + 274, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 703, + 274, + 712 + ], + "spans": [ + { + "bbox": [ + 69, + 703, + 274, + 712 + ], + "type": "text", + "content": "Corresponding author. E-mail: shiboxin@pku.edu.cn." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16133" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 275 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 275 + ], + "type": "text", + "content": "going radiance. This strategy enables the computation of reflected radiance by evaluating the physically constrained rendering equation [14], taking into account local scene components including geometry, material properties, and incident lighting, thereby yielding more accurate, realistic, and reliable view-dependent effects. Among the three local scene components, addressing local incident lighting is perhaps the most intricate. Adopting a suboptimal approach to addressing local incident lighting may undermine global illumination effects such as inter-reflections, resulting in an unrealistic scene appearance. Notably, this issue is particularly pronounced in scenes characterized by highly specular surfaces; for instance, the absence of reflections of nearby objects would be especially noticeable. Therefore, it is crucial for methods aimed at modeling such high-frequency view-dependent appearances through the rendering equation to carefully address local incident lighting." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 275, + 295, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 275, + 295, + 491 + ], + "spans": [ + { + "bbox": [ + 55, + 275, + 295, + 491 + ], + "type": "text", + "content": "Unfortunately, effectively acquiring local incident lighting poses a substantial challenge for 3DGS-based methods. The advantages of 3DGS come with a trade-off: The exclusive reliance on the tile-based rasterizer limits its ability to efficiently render secondary rays emanating from vastly different origins towards arbitrary directions, which is essential for querying local incident lighting. As a compromise, a majority of methods neglect secondary effects [36, 39, 47] or periodically bake secondary effects including local occlusions [8, 40] and incident radiance [18, 29] as low-order SH stored in Gaussians or probes. However, low-order SH can only capture low-frequency variations in incident lighting, and the requirement of baking not only incurs additional storage costs but also hinders post-reconstruction editing, such as inserting or deleting nearby objects or relighting. Employing screen space ray tracing (SSRT) [2] produces high-frequency but incomplete inter-reflections as SSRT struggles to trace out-of-screen and occluded rays." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 491, + 295, + 706 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 491, + 295, + 706 + ], + "spans": [ + { + "bbox": [ + 55, + 491, + 295, + 706 + ], + "type": "text", + "content": "We propose SpecTRe-GS to model highly Specular surfaces with reflected nearby objects through Tracing Rays in the 3D space within the Gaussian Splatting framework, overcoming the aforementioned challenges. We first made the observation that, although highly specular surfaces exhibit reflections that can be extremely high-frequency and thus hard to capture using SH color or neural network, they are also favorable for modeling in the sense that the specular lobe of their bidirectional reflectance distribution functions (BRDF) are highly concentrated. This fact facilitates the evaluation of the rendering equation [14] on these surfaces because tracing only a few secondary rays already suffices, in contrast to rougher surfaces with more scattered specular lobes. To make use of this fact, we separately model highly specular and rough surfaces, ray-tracing the appearance of the former while retaining the low-order SH color to well-approximate the latter, significantly reducing the amount of secondary rays to trace. To further speed up the time" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 313, + 73, + 553, + 143 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 73, + 553, + 143 + ], + "spans": [ + { + "bbox": [ + 313, + 73, + 553, + 143 + ], + "type": "text", + "content": "consuming ray tracing process, we implement a ray tracer within the 3DGS framework, inspired by the recent progresses [24, 42]. We also propose a training strategy to enhance the scene geometry reconstruction, the quality of which directly decides the quality of the reconstructed reflection. To summarize, our contributions are as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 144, + 553, + 228 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 314, + 144, + 553, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 144, + 553, + 179 + ], + "spans": [ + { + "bbox": [ + 314, + 144, + 553, + 179 + ], + "type": "text", + "content": "- proposing a shading model tailored for indirect mirror reflections combined with an efficient ray tracer for fast and accurate rendering;" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 180, + 553, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 180, + 553, + 204 + ], + "spans": [ + { + "bbox": [ + 313, + 180, + 553, + 204 + ], + "type": "text", + "content": "- designing a training strategy to enhance geometry reconstruction for precise incident direction computing; and" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 204, + 553, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 204, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 313, + 204, + 553, + 228 + ], + "type": "text", + "content": "- achieving state-of-the-art novel view synthesis results on scenes with highly specular reflections (shown in Fig. 1)." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 239, + 400, + 251 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 239, + 400, + 251 + ], + "spans": [ + { + "bbox": [ + 314, + 239, + 400, + 251 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 258, + 553, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 258, + 553, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 258, + 553, + 640 + ], + "type": "text", + "content": "Improving view-dependent appearances in 3DGS. The original 3DGS [15] uses low-order SH color that captures only low-frequency angular variations, a limitation that many follow-up methods were trying to address (see Tab. 1). While a few methods [30, 35] (row 2) adopt neural networks to decode local features, the majority [2, 8, 10, 13, 18, 29, 36, 39, 40, 47] (rows 3-8) employ the physically-based rendering equation [14]. This equation offers the most accurate approach for modeling view-dependent appearances but requires explicit modeling of surface normals, BRDFs, and the most demanding component - incident lighting. The rendering equation can accommodate all-frequency BRDFs; thus, the frequency of the reflected radiance is now bounded by that of the modeled incident lighting. While an environment map suffices to model direct lighting, it is significantly more challenging to model indirect lighting responsible for global illumination effects including occlusions and inter-reflections. Some methods circumvent this issue by ignoring secondary effects [36] (row 3) or by using a residual color [13] (row 4). For methods that explicitly model indirect lighting, a common strategy is to periodically bake occlusions [8] or inter-reflections [18, 29] into low-order SH, thereby limiting their incident radiance to low-frequency variations (rows 5-7). GI-GS [2] (row 8) uses screen space ray tracing to simulate high-frequency inter-reflections; however, it only models those whose sources are directly visible in the viewpoint. There are also specialized methods for plane mirrors that reflect the point cloud or camera about the plane [19, 22]. Our SpecTRe-GS traces the incident lighting in 3D space, thus is able to render high-frequency and complete inter-reflections for highly specular surfaces with convex shapes." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 642, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 553, + 713 + ], + "type": "text", + "content": "Ray-based 3D Gaussian rendering. It is noteworthy that tracing secondary rays for incoming local lighting has almost become a common practice in recent NeRF-based methods that explicitly model specular reflections or interreflections [9, 20, 32, 33, 45]. However, it is challenging to achieve this in 3DGS-based frameworks. 3DGRT [24]" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16134" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 118, + 552, + 274 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 115 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 115 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 115 + ], + "type": "text", + "content": "Table 1. We summarize works improving the view-dependent appearances upon 3DGS [15] based on: How do they model the view-dependent appearances (i.e., the outgoing radiance) and the supported frequency in angular space; how do they accommodate the global illumination effects (i.e., the local incident lighting) including occlusions and inter-reflections and the supported frequency in angular space; and whether is it feasible to edit the scene appearance under (in)direct lighting after reconstruction without any re-baking." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 118, + 552, + 274 + ], + "lines": [ + { + "bbox": [ + 58, + 118, + 552, + 274 + ], + "spans": [ + { + "bbox": [ + 58, + 118, + 552, + 274 + ], + "type": "table", + "html": "
MethodView-dependent AppearanceGlobal Illumination EffectEditable Appearance
ModelingFrequencyOcclusionInter-reflectionFrequency
(1) 3DGS [15]Low-order SHLowJointly baked in outgoing SH colorLowNone
(2) 3iGS [30]; [35]Neural networkNet.-boundedJointly baked in the neural featureNet.-boundedNone
(3) DeferredGS [36]IgnoredNone
(4) GShader [13]; [39]Arbitrary, thus bounded by the frequency of incident lightingJointly baked in outgoing SH residualDirect
(5) GS-ID [8]; [40]Physically-based rendering equation [14]Baked as SHIgnoredLow
(6) GS-IR [18]; [10]Baked as SH
(7) GIR [29]Ray-traced in 3D spaceBaked as SH
(8) GI-GS [2]Ray-traced in screen space, incompleteHigh(In)direct
(9) OursRay-traced in 3D spaceHigh(In)direct
", + "image_path": "aa3dc18f044511a489a76805f8b3704d98e953d3d8e96918ba36c8d914e69328.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 280, + 294, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 280, + 294, + 411 + ], + "spans": [ + { + "bbox": [ + 54, + 280, + 294, + 411 + ], + "type": "text", + "content": "first demonstrates efficient ray tracing in 3D Gaussians, enabling ray-based lighting effects such as non-projective views. RayGauss [1] and EVER [21] employ ray sampling in 3D Gaussians, mitigating common artifacts in 3DGS, such as aliasing [41], popping artifacts [27] and projection errors [12]. Yang et al. [46] adopt path tracing to simulate global illumination with Gaussian primitives. Meanwhile, Jorge et al. [4] utilizes ray sampling to simulate volumetric scattering and emissive media. Our SpecTRe-GS integrates ray tracing techniques to evaluate accurate incident radiance in the 3DGS framework at an affordable cost." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 425, + 162, + 438 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 425, + 162, + 438 + ], + "spans": [ + { + "bbox": [ + 55, + 425, + 162, + 438 + ], + "type": "text", + "content": "3. Proposed Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 445, + 295, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 445, + 295, + 590 + ], + "spans": [ + { + "bbox": [ + 54, + 445, + 295, + 590 + ], + "type": "text", + "content": "Our method is specifically designed to accurately model the high-frequency reflections of highly specular surfaces. We build upon the recent advancements in 3DGS (Sec. 3.1) and adapt the reflection model to efficiently handle scenes comprising both highly specular objects and rougher objects (Sec. 3.2). To precisely and reliably query the incident radiance necessary for rendering both direct and indirect highly specular reflections, we incorporate ray tracing within the 3DGS framework (Sec. 3.3). We employ a dedicated training strategy to enhance geometry reconstruction, which is crucial for rendering undistorted and realistic reflections (Sec. 3.4). Our pipeline is outlined in Fig. 2." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 599, + 135, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 599, + 135, + 612 + ], + "spans": [ + { + "bbox": [ + 55, + 599, + 135, + 612 + ], + "type": "text", + "content": "3.1. Preliminary" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 616, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 616, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 616, + 295, + 715 + ], + "type": "text", + "content": "The seminal work of 3DGS [15] represents the scene as a translucent Gaussian point cloud, where each 3D Gaussian is described with central point " + }, + { + "bbox": [ + 55, + 616, + 295, + 715 + ], + "type": "inline_equation", + "content": "\\pmb{\\mu} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 55, + 616, + 295, + 715 + ], + "type": "text", + "content": ", scales " + }, + { + "bbox": [ + 55, + 616, + 295, + 715 + ], + "type": "inline_equation", + "content": "\\pmb{s} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 55, + 616, + 295, + 715 + ], + "type": "text", + "content": ", rotation quaternion " + }, + { + "bbox": [ + 55, + 616, + 295, + 715 + ], + "type": "inline_equation", + "content": "\\pmb{q} \\in \\mathbb{R}^4" + }, + { + "bbox": [ + 55, + 616, + 295, + 715 + ], + "type": "text", + "content": ", and opacity " + }, + { + "bbox": [ + 55, + 616, + 295, + 715 + ], + "type": "inline_equation", + "content": "\\sigma \\in [0,1]" + }, + { + "bbox": [ + 55, + 616, + 295, + 715 + ], + "type": "text", + "content": ". The distribution of each Gaussian in 3D space is given by " + }, + { + "bbox": [ + 55, + 616, + 295, + 715 + ], + "type": "inline_equation", + "content": "G(\\pmb{x}) = e^{-\\frac{1}{2} (\\pmb{x} - \\pmb{\\mu})^\\top \\pmb{\\Sigma}^{-1}(\\pmb{x} - \\pmb{\\mu})}" + }, + { + "bbox": [ + 55, + 616, + 295, + 715 + ], + "type": "text", + "content": ", where the covariance " + }, + { + "bbox": [ + 55, + 616, + 295, + 715 + ], + "type": "inline_equation", + "content": "\\pmb{\\Sigma} = \\mathbf{RSS}^\\top \\mathbf{R}^\\top" + }, + { + "bbox": [ + 55, + 616, + 295, + 715 + ], + "type": "text", + "content": " is compactly parameterized as " + }, + { + "bbox": [ + 55, + 616, + 295, + 715 + ], + "type": "inline_equation", + "content": "\\pmb{s}" + }, + { + "bbox": [ + 55, + 616, + 295, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 616, + 295, + 715 + ], + "type": "inline_equation", + "content": "\\pmb{q}" + }, + { + "bbox": [ + 55, + 616, + 295, + 715 + ], + "type": "text", + "content": ". By performing local affine transformation [48], 3D Gauss" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 280, + 555, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 280, + 555, + 327 + ], + "spans": [ + { + "bbox": [ + 313, + 280, + 555, + 327 + ], + "type": "text", + "content": "sians are projected as 2D Gaussians on the image plane for rasterization. Then for each ray corresponding to an image pixel, the contributing 2D Gaussians are accumulated from front to back for volumetric rendering:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 354, + 331, + 555, + 347 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 331, + 555, + 347 + ], + "spans": [ + { + "bbox": [ + 354, + 331, + 555, + 347 + ], + "type": "interline_equation", + "content": "I \\left(\\boldsymbol {\\omega} _ {\\mathrm {o}}\\right) = \\sum_ {i} \\boldsymbol {c} _ {i} \\left(\\boldsymbol {\\omega} _ {\\mathrm {o}}\\right) \\alpha_ {i} \\prod_ {j = 1} ^ {i - 1} \\left(1 - \\alpha_ {j}\\right), \\tag {1}", + "image_path": "fdf22d396d371f178d6d54101ad4fb3311c466d6d6e0a9a1486775eee50efb40.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 350, + 554, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 350, + 554, + 373 + ], + "spans": [ + { + "bbox": [ + 313, + 350, + 554, + 373 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 350, + 554, + 373 + ], + "type": "inline_equation", + "content": "\\alpha_{i} = \\sigma_{i}G_{i}" + }, + { + "bbox": [ + 313, + 350, + 554, + 373 + ], + "type": "text", + "content": " and the view-dependent color " + }, + { + "bbox": [ + 313, + 350, + 554, + 373 + ], + "type": "inline_equation", + "content": "c_{i}(\\omega_{\\mathrm{o}})\\in" + }, + { + "bbox": [ + 313, + 350, + 554, + 373 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^3" + }, + { + "bbox": [ + 313, + 350, + 554, + 373 + ], + "type": "text", + "content": " is modeled by third-order SH coefficients " + }, + { + "bbox": [ + 313, + 350, + 554, + 373 + ], + "type": "inline_equation", + "content": "\\varphi_{i}\\in \\mathbb{R}^{48}" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 374, + 554, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 374, + 554, + 447 + ], + "spans": [ + { + "bbox": [ + 313, + 374, + 554, + 447 + ], + "type": "text", + "content": "Besides the view-dependent color " + }, + { + "bbox": [ + 313, + 374, + 554, + 447 + ], + "type": "inline_equation", + "content": "c_{i}" + }, + { + "bbox": [ + 313, + 374, + 554, + 447 + ], + "type": "text", + "content": ", many other opt. mizable attributes including surface normal, diffuse albedo, or occlusion information, can be attached to each Gaussian. These attributes can be used locally in each Gaussian, or be accumulated to a map in the same way as Eq. (1) for later applications such as deferred shading [39]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 456, + 511, + 470 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 456, + 511, + 470 + ], + "spans": [ + { + "bbox": [ + 313, + 456, + 511, + 470 + ], + "type": "text", + "content": "3.2. Highly Specular Reflection Modeling" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 312, + 474, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 474, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 312, + 474, + 555, + 715 + ], + "type": "text", + "content": "To model high-frequency reflections, we follow the convention to incorporate the physically-based rendering equation [14]. As commonly observed phenomena in real-world life, the inter-reflections from the nearby objects on highly specular surfaces should be modeled for a photorealistic scene appearance. The common practice with distant environment maps [13] ignores such inter-reflections and does not suffice for this task, thus we employ raytracing (detailed in Sec. 3.3) in 3D space for a complete and high-frequency inter-reflection. However, we found that performing ray tracing on the entire scene, including highly specular surfaces and rougher surfaces, is highly time-consuming. Evaluating the rendering equation [14] on rougher surfaces typically needs larger samples of the incident lighting even with multiple importance sampling [31], thus heavily increases the computational cost. Fortunately, previous works [15] have proven that rough surfaces can be well-captured by directly optimizing the outgoing color. Therefore, we propose the following reflection model that treats highly specular surfaces and rougher surfaces sepa" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16135" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 55, + 68, + 555, + 273 + ], + "blocks": [ + { + "bbox": [ + 55, + 68, + 555, + 273 + ], + "lines": [ + { + "bbox": [ + 55, + 68, + 555, + 273 + ], + "spans": [ + { + "bbox": [ + 55, + 68, + 555, + 273 + ], + "type": "image", + "image_path": "90c711e9d2b22af64bed4a863207a89fae0fefe7c5cc400050b466b8a2575f18.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 275, + 555, + 365 + ], + "lines": [ + { + "bbox": [ + 54, + 275, + 555, + 365 + ], + "spans": [ + { + "bbox": [ + 54, + 275, + 555, + 365 + ], + "type": "text", + "content": "Figure 2. Overview of the SpecTRe-GS pipeline. The scene is represented as a Gaussian point cloud, where each Gaussian is associated with geometry and shading attributes used to render the depth map " + }, + { + "bbox": [ + 54, + 275, + 555, + 365 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 54, + 275, + 555, + 365 + ], + "type": "text", + "content": ", normal map " + }, + { + "bbox": [ + 54, + 275, + 555, + 365 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 54, + 275, + 555, + 365 + ], + "type": "text", + "content": ", diffuse reflection " + }, + { + "bbox": [ + 54, + 275, + 555, + 365 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{diff}}" + }, + { + "bbox": [ + 54, + 275, + 555, + 365 + ], + "type": "text", + "content": ", specular reflectance " + }, + { + "bbox": [ + 54, + 275, + 555, + 365 + ], + "type": "inline_equation", + "content": "A_{\\mathrm{spec}}" + }, + { + "bbox": [ + 54, + 275, + 555, + 365 + ], + "type": "text", + "content": " and rough surface reflection " + }, + { + "bbox": [ + 54, + 275, + 555, + 365 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{rs}}" + }, + { + "bbox": [ + 54, + 275, + 555, + 365 + ], + "type": "text", + "content": " through splatting (Sec. 3.1). Highly specular surface reflection " + }, + { + "bbox": [ + 54, + 275, + 555, + 365 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{ss}}" + }, + { + "bbox": [ + 54, + 275, + 555, + 365 + ], + "type": "text", + "content": " is formulated as the sum of the diffuse reflection " + }, + { + "bbox": [ + 54, + 275, + 555, + 365 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{diff}}" + }, + { + "bbox": [ + 54, + 275, + 555, + 365 + ], + "type": "text", + "content": " and the specular reflection computed from the incident radiance " + }, + { + "bbox": [ + 54, + 275, + 555, + 365 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{i}}" + }, + { + "bbox": [ + 54, + 275, + 555, + 365 + ], + "type": "text", + "content": " and specular reflectance " + }, + { + "bbox": [ + 54, + 275, + 555, + 365 + ], + "type": "inline_equation", + "content": "A_{\\mathrm{spec}}" + }, + { + "bbox": [ + 54, + 275, + 555, + 365 + ], + "type": "text", + "content": " (Sec. 3.2). The incident radiance " + }, + { + "bbox": [ + 54, + 275, + 555, + 365 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{i}}" + }, + { + "bbox": [ + 54, + 275, + 555, + 365 + ], + "type": "text", + "content": " is determined along reflected secondary rays given by the scene geometry " + }, + { + "bbox": [ + 54, + 275, + 555, + 365 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 54, + 275, + 555, + 365 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 275, + 555, + 365 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 54, + 275, + 555, + 365 + ], + "type": "text", + "content": ", with the indirect component " + }, + { + "bbox": [ + 54, + 275, + 555, + 365 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{ind}}" + }, + { + "bbox": [ + 54, + 275, + 555, + 365 + ], + "type": "text", + "content": " and visibility " + }, + { + "bbox": [ + 54, + 275, + 555, + 365 + ], + "type": "inline_equation", + "content": "V_{\\mathrm{i}}" + }, + { + "bbox": [ + 54, + 275, + 555, + 365 + ], + "type": "text", + "content": " evaluated via efficient ray tracing within the Gaussian point cloud, and the direct component " + }, + { + "bbox": [ + 54, + 275, + 555, + 365 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{dir}}" + }, + { + "bbox": [ + 54, + 275, + 555, + 365 + ], + "type": "text", + "content": " modeled by an estimizable environment map (Sec. 3.3). The scene geometry is enhanced using normal prior guidance, numerical gradients from ray-traced incident radiance, and other training signals to achieve undistorted reflections (Sec. 3.4)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 376, + 295, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 376, + 295, + 424 + ], + "spans": [ + { + "bbox": [ + 54, + 376, + 295, + 424 + ], + "type": "text", + "content": "rately to enhance the specular reflections while maintaining the rendering speed. Also, we follow deferred shading [5] to shade for each pixel instead of each Gaussian, as it reduces the computation and gives sharp surface normals [47]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 425, + 295, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 425, + 295, + 483 + ], + "spans": [ + { + "bbox": [ + 54, + 425, + 295, + 483 + ], + "type": "text", + "content": "Modeling highly specular surfaces. To model surfaces that may exhibit both highly specular and diffuse reflections, such as polished marble, we define the reflection " + }, + { + "bbox": [ + 54, + 425, + 295, + 483 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{ss}}" + }, + { + "bbox": [ + 54, + 425, + 295, + 483 + ], + "type": "text", + "content": " for these highly specular surfaces at surface point " + }, + { + "bbox": [ + 54, + 425, + 295, + 483 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 54, + 425, + 295, + 483 + ], + "type": "text", + "content": ", observed from view direction " + }, + { + "bbox": [ + 54, + 425, + 295, + 483 + ], + "type": "inline_equation", + "content": "\\omega_{\\mathrm{o}}" + }, + { + "bbox": [ + 54, + 425, + 295, + 483 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 97, + 489, + 295, + 502 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 489, + 295, + 502 + ], + "spans": [ + { + "bbox": [ + 97, + 489, + 295, + 502 + ], + "type": "interline_equation", + "content": "I _ {\\mathrm {s s}} (\\boldsymbol {x}, \\omega_ {\\mathrm {o}}) = I _ {\\text {d i f f}} (\\boldsymbol {x}) + I _ {\\text {s p e c}} (\\boldsymbol {x}, \\omega_ {\\mathrm {o}}). \\tag {2}", + "image_path": "d59fbc88d3ed8f2697de8ec389c8d7b69b49c4a6b0c5892c191e562e19ca1366.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 506, + 296, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 506, + 296, + 553 + ], + "spans": [ + { + "bbox": [ + 55, + 506, + 296, + 553 + ], + "type": "text", + "content": "While the diffuse reflection " + }, + { + "bbox": [ + 55, + 506, + 296, + 553 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{diff}}" + }, + { + "bbox": [ + 55, + 506, + 296, + 553 + ], + "type": "text", + "content": " is the view-independent outgoing radiance directly optimized as " + }, + { + "bbox": [ + 55, + 506, + 296, + 553 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 55, + 506, + 296, + 553 + ], + "type": "text", + "content": " in each Gaussian, the specular reflection " + }, + { + "bbox": [ + 55, + 506, + 296, + 553 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{spec}}" + }, + { + "bbox": [ + 55, + 506, + 296, + 553 + ], + "type": "text", + "content": " is approximated with an ideal mirror reflection:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 96, + 559, + 295, + 572 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 559, + 295, + 572 + ], + "spans": [ + { + "bbox": [ + 96, + 559, + 295, + 572 + ], + "type": "interline_equation", + "content": "I _ {\\operatorname {s p e c}} (\\boldsymbol {x}, \\omega_ {\\mathrm {o}}) = A _ {\\operatorname {s p e c}} (\\boldsymbol {x}, \\omega_ {\\mathrm {r}}) I _ {\\mathrm {i}} (\\boldsymbol {x}, \\omega_ {\\mathrm {r}}), \\tag {3}", + "image_path": "f1dbe26e1625adb46e0c0ccf55d30b13a12310e22baea6a84f00dba12d47b028.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 576, + 296, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 576, + 296, + 624 + ], + "spans": [ + { + "bbox": [ + 55, + 576, + 296, + 624 + ], + "type": "text", + "content": "where the reflected incident direction " + }, + { + "bbox": [ + 55, + 576, + 296, + 624 + ], + "type": "inline_equation", + "content": "\\omega_{\\mathrm{r}}" + }, + { + "bbox": [ + 55, + 576, + 296, + 624 + ], + "type": "text", + "content": " is given by " + }, + { + "bbox": [ + 55, + 576, + 296, + 624 + ], + "type": "inline_equation", + "content": "\\omega_{\\mathrm{o}} - 2\\omega_{\\mathrm{o}}^{\\top}\\pmb{n}\\cdot \\pmb{n}" + }, + { + "bbox": [ + 55, + 576, + 296, + 624 + ], + "type": "text", + "content": " in which " + }, + { + "bbox": [ + 55, + 576, + 296, + 624 + ], + "type": "inline_equation", + "content": "\\pmb{n}" + }, + { + "bbox": [ + 55, + 576, + 296, + 624 + ], + "type": "text", + "content": " denotes the local surface normal, and " + }, + { + "bbox": [ + 55, + 576, + 296, + 624 + ], + "type": "inline_equation", + "content": "A_{\\mathrm{spec}}" + }, + { + "bbox": [ + 55, + 576, + 296, + 624 + ], + "type": "text", + "content": " is the specular reflectance considering Fresnel effects through Schlick's approximation [28]:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 63, + 627, + 295, + 643 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 627, + 295, + 643 + ], + "spans": [ + { + "bbox": [ + 63, + 627, + 295, + 643 + ], + "type": "interline_equation", + "content": "A _ {\\mathrm {s p e c}} (\\pmb {x}, \\pmb {\\omega} _ {\\mathrm {r}}) = F _ {0} (\\pmb {x}) + (1 - F _ {0} (\\pmb {x})) (1 - \\pmb {n} ^ {\\top} \\pmb {\\omega} _ {\\mathrm {r}}) ^ {5}. \\quad (4)", + "image_path": "928355ccad3c0e209f759088680f08f976e1eb6ac3cbf017eb8ef8c95890c402.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 645, + 296, + 682 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 645, + 296, + 682 + ], + "spans": [ + { + "bbox": [ + 55, + 645, + 296, + 682 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 55, + 645, + 296, + 682 + ], + "type": "inline_equation", + "content": "F_{0}" + }, + { + "bbox": [ + 55, + 645, + 296, + 682 + ], + "type": "text", + "content": " is the specular reflectance at normal incidence optimized as " + }, + { + "bbox": [ + 55, + 645, + 296, + 682 + ], + "type": "inline_equation", + "content": "\\pmb{f_0}" + }, + { + "bbox": [ + 55, + 645, + 296, + 682 + ], + "type": "text", + "content": " in each Gaussian and is 3-channel to support metallic materials. The incident radiance " + }, + { + "bbox": [ + 55, + 645, + 296, + 682 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{i}}" + }, + { + "bbox": [ + 55, + 645, + 296, + 682 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 55, + 645, + 296, + 682 + ], + "type": "inline_equation", + "content": "\\omega_{\\mathrm{r}}" + }, + { + "bbox": [ + 55, + 645, + 296, + 682 + ], + "type": "text", + "content": " originates" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 376, + 467, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 376, + 467, + 388 + ], + "spans": [ + { + "bbox": [ + 313, + 376, + 467, + 388 + ], + "type": "text", + "content": "from both direct and indirect sources:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 539, + 382, + 553, + 393 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 539, + 382, + 553, + 393 + ], + "spans": [ + { + "bbox": [ + 539, + 382, + 553, + 393 + ], + "type": "text", + "content": "(5)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 395, + 561, + 408 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 395, + 561, + 408 + ], + "spans": [ + { + "bbox": [ + 313, + 395, + 561, + 408 + ], + "type": "interline_equation", + "content": "I _ {\\mathrm {i}} (\\boldsymbol {x}, \\omega_ {\\mathrm {r}}) = V _ {\\mathrm {i}} (\\boldsymbol {x}, \\omega_ {\\mathrm {r}}) I _ {\\mathrm {d i r}} (\\omega_ {\\mathrm {r}}) + (1 - V _ {\\mathrm {i}} (\\boldsymbol {x}, \\omega_ {\\mathrm {r}})) I _ {\\mathrm {i n d}} (\\boldsymbol {x}, \\omega_ {\\mathrm {r}}),", + "image_path": "d9692e163aa50a3cb616c29edbcf5ca4af9a068d6de4f08d608941eab725f136.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 413, + 555, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 413, + 555, + 460 + ], + "spans": [ + { + "bbox": [ + 313, + 413, + 555, + 460 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 413, + 555, + 460 + ], + "type": "inline_equation", + "content": "V_{\\mathrm{i}} \\in [0,1]" + }, + { + "bbox": [ + 313, + 413, + 555, + 460 + ], + "type": "text", + "content": " quantifies the proportion of the incident lighting occluded by the scene itself, " + }, + { + "bbox": [ + 313, + 413, + 555, + 460 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{dir}}" + }, + { + "bbox": [ + 313, + 413, + 555, + 460 + ], + "type": "text", + "content": " is the direct incident lighting from a distant environment map, and " + }, + { + "bbox": [ + 313, + 413, + 555, + 460 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{ind}}" + }, + { + "bbox": [ + 313, + 413, + 555, + 460 + ], + "type": "text", + "content": " is the indirect incident radiance from the scene." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 461, + 556, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 461, + 556, + 616 + ], + "spans": [ + { + "bbox": [ + 313, + 461, + 556, + 616 + ], + "type": "text", + "content": "Defining depth and surface normal of Gaussians. It is essential to accurately acquire the depth and surface normal of each Gaussian in our method, as they directly determine the origin and direction of the reflection. However, they are ambiguous because the Gaussians are translucent ellipsoids in 3D space. Instead of treating the depth of the projected center and the shortest axis direction facing the camera as the depth and normal of each Gaussian [13, 17, 39], we adopt the formulation in GOF [42] to calculate for each Gaussian the \"intersection point\" along the ray achieving the maximum response, using its depth and normal of the intersection plane. This formulation gives a better and more consistent geometry reconstruction with 3D Gaussians." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 616, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 616, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 616, + 556, + 713 + ], + "type": "text", + "content": "Modeling the whole image. We model the part of the scene containing highly specular surfaces using Eq. (2). For the other part which exhibits only low-frequency view dependencies, we follow GOF [42] to model the appearance " + }, + { + "bbox": [ + 313, + 616, + 556, + 713 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{rs}}" + }, + { + "bbox": [ + 313, + 616, + 556, + 713 + ], + "type": "text", + "content": " of the rougher surfaces using per-Gaussian SH coefficients " + }, + { + "bbox": [ + 313, + 616, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\varphi" + }, + { + "bbox": [ + 313, + 616, + 556, + 713 + ], + "type": "text", + "content": ". We combine " + }, + { + "bbox": [ + 313, + 616, + 556, + 713 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{ss}}" + }, + { + "bbox": [ + 313, + 616, + 556, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 616, + 556, + 713 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{rs}}" + }, + { + "bbox": [ + 313, + 616, + 556, + 713 + ], + "type": "text", + "content": " using a binary mask " + }, + { + "bbox": [ + 313, + 616, + 556, + 713 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 616, + 556, + 713 + ], + "type": "text", + "content": " (which can be handily acquired through pre-trained segmentation models such as SAM [16]) specifying the highly specular" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 55, + 693, + 295, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 693, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 693, + 295, + 712 + ], + "type": "text", + "content": "1We omit the distribution and mask-shadowing term in the microfacet model as we assume ideal mirror reflection." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16136" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 290, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 290, + 84 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 290, + 84 + ], + "type": "text", + "content": "surfaces within the scene ( " + }, + { + "bbox": [ + 55, + 72, + 290, + 84 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 55, + 72, + 290, + 84 + ], + "type": "text", + "content": " is pixel-wise multiplication):" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 82, + 89, + 295, + 102 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 89, + 295, + 102 + ], + "spans": [ + { + "bbox": [ + 82, + 89, + 295, + 102 + ], + "type": "interline_equation", + "content": "I \\left(\\omega_ {\\mathrm {o}}\\right) = I _ {\\mathrm {s s}} \\left(\\omega_ {\\mathrm {o}}\\right) \\odot M + I _ {\\mathrm {r s}} \\left(\\omega_ {\\mathrm {o}}\\right) \\odot (1 - M). \\tag {6}", + "image_path": "2a01b5899e0ca1cb9e685c6c970a309c47d9d9d611dc778c67087e0a5c7b7125.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 110, + 284, + 122 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 110, + 284, + 122 + ], + "spans": [ + { + "bbox": [ + 55, + 110, + 284, + 122 + ], + "type": "text", + "content": "3.3. Efficient Ray Tracing for Incident Radiance" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 128, + 296, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 128, + 296, + 281 + ], + "spans": [ + { + "bbox": [ + 55, + 128, + 296, + 281 + ], + "type": "text", + "content": "To accurately compute the reflection from highly specular surfaces, it is crucial to obtain the incident radiance at the reflected view direction as modeled by Eq. (5), which incorporates both direct and indirect lighting. However, due to the incoherent origins and directions of these secondary rays, the incident radiance cannot be efficiently rendered using the rasterizer from the original 3DGS [15], as it relies on the locality of the rays, nor can it be computed merely using an environment map, which considers only direct lighting [13]. Inspired by the recent work 3DGRT [24], we implement an efficient ray tracer within the 3DGS framework to compute the radiance of secondary rays, thus achieving high-frequency and accurate reflections." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 283, + 296, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 283, + 296, + 403 + ], + "spans": [ + { + "bbox": [ + 55, + 283, + 296, + 403 + ], + "type": "text", + "content": "Geometry acceleration structure (GAS). For fast scene traversal, we construct a GAS within the OptiX programming interface [25], which uses ray tracing hardware to accelerate ray-primitive intersection tests and is several times faster than BVH [3] in CUDA. Instead of using axis-aligned bounding boxes which inefficiently encapsulate irregular Gaussians and lead to false positives, we adopt stretched icosahedrons as proxy meshes for more efficient ray intersection [24]. The vertex " + }, + { + "bbox": [ + 55, + 283, + 296, + 403 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 55, + 283, + 296, + 403 + ], + "type": "text", + "content": " of the icosahedron approximating Gaussian " + }, + { + "bbox": [ + 55, + 283, + 296, + 403 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 283, + 296, + 403 + ], + "type": "text", + "content": " is given by:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 87, + 407, + 294, + 422 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 87, + 407, + 294, + 422 + ], + "spans": [ + { + "bbox": [ + 87, + 407, + 294, + 422 + ], + "type": "interline_equation", + "content": "\\boldsymbol {v} _ {i, j} = \\sqrt {2 \\log \\left(\\sigma_ {i} / \\alpha_ {\\min }\\right)} \\mathbf {R} _ {i} ^ {\\top} \\mathbf {S} _ {i} ^ {\\top} \\boldsymbol {v} _ {j} ^ {\\text {u n i t}} + \\boldsymbol {\\mu} _ {i}, \\tag {7}", + "image_path": "e6245095e7e5be9dc49a8c1d43cbe45edab1be653bc3cd284f3129a1bae68437.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 426, + 296, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 426, + 296, + 522 + ], + "spans": [ + { + "bbox": [ + 55, + 426, + 296, + 522 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 426, + 296, + 522 + ], + "type": "inline_equation", + "content": "\\pmb{v}_j^{\\mathrm{unit}}" + }, + { + "bbox": [ + 55, + 426, + 296, + 522 + ], + "type": "text", + "content": " is the vertex of an inscribed icosahedron in a unit sphere. Here, " + }, + { + "bbox": [ + 55, + 426, + 296, + 522 + ], + "type": "inline_equation", + "content": "\\alpha_{\\mathrm{min}}" + }, + { + "bbox": [ + 55, + 426, + 296, + 522 + ], + "type": "text", + "content": " serves as a threshold that adaptively scales each icosahedron by dictating the minimal contribution of the Gaussian at the transformed vertices, thereby reducing the computational overhead during intersection and radiance evaluation. We regularly update the GAS after each densification operation [15] and every " + }, + { + "bbox": [ + 55, + 426, + 296, + 522 + ], + "type": "inline_equation", + "content": "n_{\\mathrm{up}}" + }, + { + "bbox": [ + 55, + 426, + 296, + 522 + ], + "type": "text", + "content": " step as the Gaussian point cloud changes along the training." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 522, + 296, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 522, + 296, + 594 + ], + "spans": [ + { + "bbox": [ + 55, + 522, + 296, + 594 + ], + "type": "text", + "content": "Ray queries with GAS. To evaluate the incident radiance " + }, + { + "bbox": [ + 55, + 522, + 296, + 594 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{ind}}(\\pmb{x}_0, \\pmb{\\omega}_{\\mathrm{r}})" + }, + { + "bbox": [ + 55, + 522, + 296, + 594 + ], + "type": "text", + "content": ", secondary rays " + }, + { + "bbox": [ + 55, + 522, + 296, + 594 + ], + "type": "inline_equation", + "content": "\\pmb{x}(t) = \\pmb{x}_0 + t\\pmb{\\omega}_{\\mathrm{r}}" + }, + { + "bbox": [ + 55, + 522, + 296, + 594 + ], + "type": "text", + "content": " are traced against proxy meshes within the GAS. Upon tracing each ray, we first collect the " + }, + { + "bbox": [ + 55, + 522, + 296, + 594 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 522, + 296, + 594 + ], + "type": "text", + "content": " closest valid hits, each corresponding to a Gaussian with " + }, + { + "bbox": [ + 55, + 522, + 296, + 594 + ], + "type": "inline_equation", + "content": "\\alpha_i \\geq \\alpha_{\\min}" + }, + { + "bbox": [ + 55, + 522, + 296, + 594 + ], + "type": "text", + "content": ". The depth of the \"true\" intersection point for each Gaussian is defined as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 113, + 597, + 295, + 625 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 597, + 295, + 625 + ], + "spans": [ + { + "bbox": [ + 113, + 597, + 295, + 625 + ], + "type": "interline_equation", + "content": "t _ {\\max , i} = \\frac {\\left(\\boldsymbol {\\mu} _ {i} - \\boldsymbol {x} _ {0}\\right) ^ {\\top} \\boldsymbol {\\Sigma} _ {i} ^ {- 1} \\boldsymbol {\\omega} _ {\\mathrm {r}}}{\\boldsymbol {\\omega} _ {\\mathrm {r}} ^ {\\top} \\boldsymbol {\\Sigma} _ {i} ^ {- 1} \\boldsymbol {\\omega} _ {\\mathrm {r}}}, \\tag {8}", + "image_path": "bf3169e3873fec33a0e77724ba29baf61b693e2c034fe7478907f640e48e4cec.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "text", + "content": "which maximizes the Gaussian's contribution " + }, + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\sigma_{i}G_{i}(\\pmb{x}_{0} + t_{\\max,i}\\pmb{\\omega}_{\\mathrm{r}})" + }, + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "text", + "content": ". We then sort the Gaussians by their depth and update the ray depth " + }, + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "inline_equation", + "content": "D_{\\mathrm{ind}}" + }, + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "text", + "content": ", visibility (i.e., transmittance) " + }, + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "inline_equation", + "content": "V_{\\mathrm{i}}" + }, + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "text", + "content": " and color " + }, + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{ind}}" + }, + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "text", + "content": " (using the Gaussians' outgoing SH colors assuming that they model rough surfaces), following Eq. (1). The above process is repeated after advancing the ray origin until no more valid hits remain or the ray transmittance" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": "drops below a threshold. For the transmittance that remains unoccluded by the reconstructed scene itself, we adopt an estimable environment map for calculating the direct incident radiance " + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{dir}}(\\omega)" + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": ". According to Eq. (5), the final incident radiance " + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{i}}(\\pmb{x}, \\pmb{\\omega}_{\\mathrm{r}})" + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": " combines direct and indirect components, weighted by visibility " + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "inline_equation", + "content": "V_{\\mathrm{i}}(\\pmb{x}, \\pmb{\\omega})" + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 144, + 555, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 144, + 555, + 228 + ], + "spans": [ + { + "bbox": [ + 313, + 144, + 555, + 228 + ], + "type": "text", + "content": "Thanks to the efficient and hardware-accelerated ray tracing techniques elaborated above, as well as our separate modeling of highly specular and rough surfaces, the average rendering time of our method is only " + }, + { + "bbox": [ + 313, + 144, + 555, + 228 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 313, + 144, + 555, + 228 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 144, + 555, + 228 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 313, + 144, + 555, + 228 + ], + "type": "text", + "content": " slower than that of a pure rasterizer on each of our test scenes. This varies according to the density of objects in the scene and the proportion of highly specular surfaces." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 243, + 378, + 255 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 243, + 378, + 255 + ], + "spans": [ + { + "bbox": [ + 314, + 243, + 378, + 255 + ], + "type": "text", + "content": "3.4. Training" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 259, + 554, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 259, + 554, + 319 + ], + "spans": [ + { + "bbox": [ + 313, + 259, + 554, + 319 + ], + "type": "text", + "content": "We propose a training strategy that especially focuses on enhancing the geometry reconstruction, as unfaithful scene geometry induces inaccurate secondary rays and results in distorted reflections. We optimize our model from an initial sparse point cloud using the following loss terms:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 319, + 324, + 553, + 337 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 324, + 553, + 337 + ], + "spans": [ + { + "bbox": [ + 319, + 324, + 553, + 337 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {\\mathrm {c}} + \\lambda_ {\\mathrm {d}} \\mathcal {L} _ {\\mathrm {d}} + \\lambda_ {\\mathrm {d n}} \\mathcal {L} _ {\\mathrm {d n}} + \\lambda_ {\\mathrm {s m}} \\mathcal {L} _ {\\mathrm {s m}} + \\lambda_ {\\mathrm {n}} \\mathcal {L} _ {\\mathrm {n}} + \\lambda_ {\\mathrm {r}} \\mathcal {L} _ {\\mathrm {r}}, \\tag {9}", + "image_path": "e8e6079db2f8f77c67d22389d7b07e57c4aa2fc5a42f709b6c1583c2d1faddf3.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 341, + 554, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 341, + 554, + 437 + ], + "spans": [ + { + "bbox": [ + 313, + 341, + 554, + 437 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 341, + 554, + 437 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{c}}" + }, + { + "bbox": [ + 313, + 341, + 554, + 437 + ], + "type": "text", + "content": " is a color reconstruction loss [15] differentiable to the geometry, " + }, + { + "bbox": [ + 313, + 341, + 554, + 437 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{d}}" + }, + { + "bbox": [ + 313, + 341, + 554, + 437 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 341, + 554, + 437 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{dn}}" + }, + { + "bbox": [ + 313, + 341, + 554, + 437 + ], + "type": "text", + "content": " are the distortion and depth-normal consistency loss used in GOF [42], " + }, + { + "bbox": [ + 313, + 341, + 554, + 437 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{sm}}" + }, + { + "bbox": [ + 313, + 341, + 554, + 437 + ], + "type": "text", + "content": " is an edge-aware smoothness term for the normal and specular reflectance for the highly specular surfaces, " + }, + { + "bbox": [ + 313, + 341, + 554, + 437 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{n}}" + }, + { + "bbox": [ + 313, + 341, + 554, + 437 + ], + "type": "text", + "content": " is a normal prior term to employ data-driven clues, and " + }, + { + "bbox": [ + 313, + 341, + 554, + 437 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{r}}" + }, + { + "bbox": [ + 313, + 341, + 554, + 437 + ], + "type": "text", + "content": " is a residual-suppressing term for progressive learning. This section describes these terms and more details are in the supplement." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 437, + 553, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 553, + 484 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 553, + 484 + ], + "type": "text", + "content": "Normal prior guidance. To tackle the challenging problem of reconstructing geometry for highly specular objects, we employ StableNormal [38] to provide monocular normal priors " + }, + { + "bbox": [ + 313, + 437, + 553, + 484 + ], + "type": "inline_equation", + "content": "\\hat{N}" + }, + { + "bbox": [ + 313, + 437, + 553, + 484 + ], + "type": "text", + "content": ", offering initial geometric constraints:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 359, + 489, + 553, + 506 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 489, + 553, + 506 + ], + "spans": [ + { + "bbox": [ + 359, + 489, + 553, + 506 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {n}} = 1 / | N | \\sum \\left| \\left| (\\hat {N} - N) \\odot M \\right| \\right| _ {1}. \\tag {10}", + "image_path": "83f0217defcf45c0ef3adcbe2d30a1774daca2f2a1abfa20d7cdca561f20c89f.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 510, + 555, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 555, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 555, + 653 + ], + "type": "text", + "content": "Joint geometry optimization. Recognizing potential inaccuracies and inconsistencies across views in the normal prior, we concurrently optimize the scene geometry during training. This is achieved by differentiating the incident radiance " + }, + { + "bbox": [ + 313, + 510, + 555, + 653 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{i}}" + }, + { + "bbox": [ + 313, + 510, + 555, + 653 + ], + "type": "text", + "content": ", which relates to the final image through Eq. (2) and Eq. (3), with respect to the scene geometry. Computing the gradient of the direct incident radiance " + }, + { + "bbox": [ + 313, + 510, + 555, + 653 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{dir}}" + }, + { + "bbox": [ + 313, + 510, + 555, + 653 + ], + "type": "text", + "content": " is straightforward. To compute the gradient of the indirect incident radiance " + }, + { + "bbox": [ + 313, + 510, + 555, + 653 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{ind}}" + }, + { + "bbox": [ + 313, + 510, + 555, + 653 + ], + "type": "text", + "content": " concerning the local surface normal " + }, + { + "bbox": [ + 313, + 510, + 555, + 653 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 510, + 555, + 653 + ], + "type": "text", + "content": ", we adopt a simple yet effective finite difference method to numerically evaluate the gradient of " + }, + { + "bbox": [ + 313, + 510, + 555, + 653 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{ind}}" + }, + { + "bbox": [ + 313, + 510, + 555, + 653 + ], + "type": "text", + "content": " with respect to the reflected direction " + }, + { + "bbox": [ + 313, + 510, + 555, + 653 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 313, + 510, + 555, + 653 + ], + "type": "text", + "content": " that depends on " + }, + { + "bbox": [ + 313, + 510, + 555, + 653 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 510, + 555, + 653 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 318, + 656, + 553, + 685 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 656, + 553, + 685 + ], + "spans": [ + { + "bbox": [ + 318, + 656, + 553, + 685 + ], + "type": "interline_equation", + "content": "\\left. \\frac {\\partial I _ {\\text {i n d}}}{\\partial \\omega^ {i}} \\right| _ {\\omega_ {0}} \\approx \\frac {I _ {\\text {i n d}} \\left(\\boldsymbol {x} , \\omega_ {0} + \\Delta_ {\\omega} ^ {i}\\right) - I _ {\\text {i n d}} \\left(\\boldsymbol {x} , \\omega_ {0} - \\Delta_ {\\omega} ^ {i}\\right)}{2 \\Delta_ {\\omega} ^ {i}}, \\tag {11}", + "image_path": "f0d6191a6d8a6386824210684f706c2030f5e72ba166cd3dffff86591a689b9c.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\Delta_{\\omega}^{i}" + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "content": " denotes the perturbation of the " + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "content": "-th dimension in " + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "content": ". Considering that using the same perturbation for sec-" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16137" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 106, + 550, + 228 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 103 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 103 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 103 + ], + "type": "text", + "content": "Table 2. Quantitative comparisons of novel view synthesis results with state-of-the-art methods on both synthetic and real-world scenes. We report the scores of PSNR, SSIM [34], and LPIPS [44] for entire images and within reflective regions. " + }, + { + "bbox": [ + 55, + 70, + 555, + 103 + ], + "type": "inline_equation", + "content": "\\uparrow (\\downarrow)" + }, + { + "bbox": [ + 55, + 70, + 555, + 103 + ], + "type": "text", + "content": " means higher (lower) is better. We mark the best and the second best results in each column." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 106, + 550, + 228 + ], + "lines": [ + { + "bbox": [ + 58, + 106, + 550, + 228 + ], + "spans": [ + { + "bbox": [ + 58, + 106, + 550, + 228 + ], + "type": "table", + "html": "
DataSynthetic (entire image)Synthetic (reflective)Real-world (entire image)Real-world (reflective)
MethodPSNR↓SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↓SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
3DGS [15]27.710.9100.13121.510.9400.07824.820.8230.22721.310.9650.047
GOF [42]27.970.9170.11721.380.9400.07624.860.8260.21921.250.9630.047
GOF* [42]27.570.9130.12520.670.9360.08524.690.8250.22320.440.9620.051
3iGS [30]28.580.9120.12521.980.9390.07624.370.8090.23321.220.9630.048
GShader [13]26.650.8880.16020.790.9350.08623.930.8100.25520.230.9630.053
GShader* [13]27.140.8940.15321.220.9380.08424.030.8120.25320.360.9640.053
3DGS-DR [39]27.620.8900.17321.570.9380.08524.790.8240.24321.180.9650.050
3DGS-DR* [39]26.970.8830.18221.180.9360.08824.600.8230.25120.380.9640.055
Ours29.410.9220.10923.280.9490.06225.110.8280.21322.770.9670.035
", + "image_path": "9bb6cbf2041863355c58631c42f8ac6bbc2e4788f3a3268cf8128cb005cf2305.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 232, + 295, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 232, + 295, + 281 + ], + "spans": [ + { + "bbox": [ + 55, + 232, + 295, + 281 + ], + "type": "text", + "content": "ondary rays with different depths " + }, + { + "bbox": [ + 55, + 232, + 295, + 281 + ], + "type": "inline_equation", + "content": "D_{\\mathrm{ind}}" + }, + { + "bbox": [ + 55, + 232, + 295, + 281 + ], + "type": "text", + "content": " leads to different amounts of intersection movements, making the numerical gradient less accurate, we use a depth-aware ray perturbation that adjusts to " + }, + { + "bbox": [ + 55, + 232, + 295, + 281 + ], + "type": "inline_equation", + "content": "D_{\\mathrm{ind}}" + }, + { + "bbox": [ + 55, + 232, + 295, + 281 + ], + "type": "text", + "content": " for more informative gradients:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 75, + 284, + 295, + 298 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 284, + 295, + 298 + ], + "spans": [ + { + "bbox": [ + 75, + 284, + 295, + 298 + ], + "type": "interline_equation", + "content": "\\Delta_ {\\omega} ^ {i} = \\theta_ {\\mathrm {d}} \\max \\left(\\min \\left(D _ {\\text {i n d}} / D _ {0}, \\sigma_ {\\mathrm {d}} ^ {\\max }\\right), \\sigma_ {\\mathrm {d}} ^ {\\min }\\right), \\tag {12}", + "image_path": "02846a2e016106caa5ad94359de2c7f53be9906d0ad8499dd84b3ffdd87626b7.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 301, + 295, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 301, + 295, + 386 + ], + "spans": [ + { + "bbox": [ + 55, + 301, + 295, + 386 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 301, + 295, + 386 + ], + "type": "inline_equation", + "content": "\\theta_{\\mathrm{d}}" + }, + { + "bbox": [ + 55, + 301, + 295, + 386 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 301, + 295, + 386 + ], + "type": "inline_equation", + "content": "D_0" + }, + { + "bbox": [ + 55, + 301, + 295, + 386 + ], + "type": "text", + "content": " define a reference angular perturbation and " + }, + { + "bbox": [ + 55, + 301, + 295, + 386 + ], + "type": "inline_equation", + "content": "\\sigma_{\\mathrm{d}}^{\\mathrm{min}}" + }, + { + "bbox": [ + 55, + 301, + 295, + 386 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 301, + 295, + 386 + ], + "type": "inline_equation", + "content": "\\sigma_{\\mathrm{d}}^{\\mathrm{max}}" + }, + { + "bbox": [ + 55, + 301, + 295, + 386 + ], + "type": "text", + "content": " ensure numerical stability. For computational efficiency, we disregard the gradient with respect to the point position " + }, + { + "bbox": [ + 55, + 301, + 295, + 386 + ], + "type": "inline_equation", + "content": "\\pmb{p}" + }, + { + "bbox": [ + 55, + 301, + 295, + 386 + ], + "type": "text", + "content": ", as we find reflections are distorted mainly due to normal errors. We detach visibility " + }, + { + "bbox": [ + 55, + 301, + 295, + 386 + ], + "type": "inline_equation", + "content": "V_{\\mathrm{ind}}" + }, + { + "bbox": [ + 55, + 301, + 295, + 386 + ], + "type": "text", + "content": " and depth " + }, + { + "bbox": [ + 55, + 301, + 295, + 386 + ], + "type": "inline_equation", + "content": "D_{\\mathrm{ind}}" + }, + { + "bbox": [ + 55, + 301, + 295, + 386 + ], + "type": "text", + "content": " because the sparsity of the visibility gradient and abrupt changes in depth damage numerical stability." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 386, + 295, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 386, + 295, + 517 + ], + "spans": [ + { + "bbox": [ + 55, + 386, + 295, + 517 + ], + "type": "text", + "content": "Progressive learning. As training progresses, monocular normal priors may hinder the accurate geometry reconstruction, thus we gradually decay the weight of normal prior guidance " + }, + { + "bbox": [ + 55, + 386, + 295, + 517 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{n}}" + }, + { + "bbox": [ + 55, + 386, + 295, + 517 + ], + "type": "text", + "content": ", letting gradients from physics-based rendering take over the fine-tuning of scene geometry. Besides, for smoother training and consequently better reconstruction within the reflective regions, we replace the hard mask " + }, + { + "bbox": [ + 55, + 386, + 295, + 517 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 386, + 295, + 517 + ], + "type": "text", + "content": " used in Eq. (6) with a relaxed version " + }, + { + "bbox": [ + 55, + 386, + 295, + 517 + ], + "type": "inline_equation", + "content": "1 - I_{\\mathrm{r}}" + }, + { + "bbox": [ + 55, + 386, + 295, + 517 + ], + "type": "text", + "content": " rendered with another per-Gaussian attribute " + }, + { + "bbox": [ + 55, + 386, + 295, + 517 + ], + "type": "inline_equation", + "content": "r \\in [0,1]" + }, + { + "bbox": [ + 55, + 386, + 295, + 517 + ], + "type": "text", + "content": " using Eq. (1). We encourage the reflective region to be modeled with highly specular reflections by suppressing the residual within " + }, + { + "bbox": [ + 55, + 386, + 295, + 517 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 386, + 295, + 517 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 115, + 521, + 294, + 539 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 521, + 294, + 539 + ], + "spans": [ + { + "bbox": [ + 115, + 521, + 294, + 539 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {r}} = 1 / | M | \\sum \\left| \\left| I _ {\\mathrm {r}} \\odot M \\right| \\right| _ {1}. \\tag {13}", + "image_path": "c5966ddf165ed5c26c82e0cf444dade6f1b7bb7c3e0e009edc8e89e3feb7c162.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 548, + 137, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 548, + 137, + 562 + ], + "spans": [ + { + "bbox": [ + 55, + 548, + 137, + 562 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 567, + 295, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 567, + 295, + 651 + ], + "spans": [ + { + "bbox": [ + 55, + 567, + 295, + 651 + ], + "type": "text", + "content": "In this section, we compare our method on both synthetic and real-world scenes (Sec. 4.2) against state-of-the-art methods (Sec. 4.3). We also perform ablation studies to validate the key components of our method (Sec. 4.4) and explore applications in scene editing (Sec. 4.5). Please check the supplementary material for more implementation details, data creation details, and experimental results." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 660, + 189, + 673 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 660, + 189, + 673 + ], + "spans": [ + { + "bbox": [ + 55, + 660, + 189, + 673 + ], + "type": "text", + "content": "4.1. Implementation Details" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "We implement our method in PyTorch [26] with CUDA kernels for rasterization based on GOF [42] and custom OptiX programs for ray tracing. We conduct all experiments on a" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 232, + 555, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 232, + 555, + 293 + ], + "spans": [ + { + "bbox": [ + 313, + 232, + 555, + 293 + ], + "type": "text", + "content": "GPU with 24 GB VRAM, 83 shader core TFLOPS, and 191 ray tracing core TFLOPS. We set " + }, + { + "bbox": [ + 313, + 232, + 555, + 293 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{sm}} = 0.5" + }, + { + "bbox": [ + 313, + 232, + 555, + 293 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 232, + 555, + 293 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{r}} = 1.0" + }, + { + "bbox": [ + 313, + 232, + 555, + 293 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 232, + 555, + 293 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{n}} = 0.5" + }, + { + "bbox": [ + 313, + 232, + 555, + 293 + ], + "type": "text", + "content": ", exponentially decaying to 0.0025 from 4k to 10k iterations. We use " + }, + { + "bbox": [ + 313, + 232, + 555, + 293 + ], + "type": "inline_equation", + "content": "r = 0.9" + }, + { + "bbox": [ + 313, + 232, + 555, + 293 + ], + "type": "text", + "content": " as the initialization. For other hyperparameters, we use the settings in GOF [42]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 300, + 376, + 311 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 300, + 376, + 311 + ], + "spans": [ + { + "bbox": [ + 313, + 300, + 376, + 311 + ], + "type": "text", + "content": "4.2. Datasets" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 317, + 554, + 401 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 317, + 554, + 401 + ], + "spans": [ + { + "bbox": [ + 313, + 317, + 554, + 401 + ], + "type": "text", + "content": "Synthetic scenes. We build 6 synthetic scenes, each containing multiple surrounding objects and a highly specular surface with distinctive properties: HELMET, MARBLETABLE, VASE, POT, TOASTER, and MIRROR. For each scene, we use the Blender Cycles engine [6] to render about 300 images of resolution " + }, + { + "bbox": [ + 313, + 317, + 554, + 401 + ], + "type": "inline_equation", + "content": "1024 \\times 1024" + }, + { + "bbox": [ + 313, + 317, + 554, + 401 + ], + "type": "text", + "content": ", of which 2/3 are used for training and the rest are left for testing." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 402, + 554, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 402, + 554, + 449 + ], + "spans": [ + { + "bbox": [ + 313, + 402, + 554, + 449 + ], + "type": "text", + "content": "Real-world scenes. For real-world validation, we capture two scenes: REALBOWL and REALPOT, consisting of 138 and 139 images of resolution " + }, + { + "bbox": [ + 313, + 402, + 554, + 449 + ], + "type": "inline_equation", + "content": "1440 \\times 1080" + }, + { + "bbox": [ + 313, + 402, + 554, + 449 + ], + "type": "text", + "content": " respectively, with every eighth image reserved in the test set." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 456, + 395, + 470 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 456, + 395, + 470 + ], + "spans": [ + { + "bbox": [ + 313, + 456, + 395, + 470 + ], + "type": "text", + "content": "4.3. Comparison" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 474, + 555, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 474, + 555, + 594 + ], + "spans": [ + { + "bbox": [ + 313, + 474, + 555, + 594 + ], + "type": "text", + "content": "Comparing methods. We compare our method with 5 open-source state-of-the-art methods: the original 3DGS [15]; GOF [42] which focus on geometry reconstruction with 3D Gaussians; 3iGS [30] which improves view-dependent effects through neural features; and 3DGS-DR [39] and GShader [13] which are adapted for specular reflection employing physically-based rendering (PBR). For a fair comparison, we provide GOF [42], 3DGS-DR [39], and GShader [13] with the normal prior guidance, and indicate these versions by “*” throughout this paper." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 594, + 554, + 642 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 554, + 642 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 554, + 642 + ], + "type": "text", + "content": "Evaluation protocol. We report peak signal-to-noise ratio (PSNR), structural similarity index measure (SSIM) [34], and learned perceptual image patch similarity (LPIPS) [44] measured on both entire images and reflective regions." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 642, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 554, + 714 + ], + "type": "text", + "content": "Quantitative evaluation. As shown in Tab. 2, our method outperforms comparing methods on both synthetic and real-world scenes, especially in reflective regions. It is worth noting that using normal prior guidance usually leads to degradation of GOF [42] which relies solely on SH color, as it suppresses the ability to simulate high-frequency appear" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16138" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 70, + 555, + 290 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 555, + 290 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 555, + 290 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 555, + 290 + ], + "type": "image", + "image_path": "a6a031519f5cb53855c07d3865df6f09606179703e6a7e213ace45fdb1793c05.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 292, + 555, + 326 + ], + "lines": [ + { + "bbox": [ + 55, + 292, + 555, + 326 + ], + "spans": [ + { + "bbox": [ + 55, + 292, + 555, + 326 + ], + "type": "text", + "content": "Figure 3. Novel view synthesis comparison on synthetic scenes. We evaluate SpecTRe-GS against relevant 3DGS-based methods modeling view-dependent appearances: 3DGS-DR [39], GShader [13], 3iGS [30], GOF [42], and 3DGS [15]. The insets highlight the improvement of SpecTRe-GS in rendering high-frequency specular reflections, showcasing the precise positioning of reflected content and sharp details." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 58, + 328, + 552, + 532 + ], + "blocks": [ + { + "bbox": [ + 58, + 328, + 552, + 532 + ], + "lines": [ + { + "bbox": [ + 58, + 328, + 552, + 532 + ], + "spans": [ + { + "bbox": [ + 58, + 328, + 552, + 532 + ], + "type": "image", + "image_path": "7baf3e4fd74eccd4a0c84a39c794fa03898ad53ef0e6b00498dcf536a0bc35ee.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 533, + 555, + 567 + ], + "lines": [ + { + "bbox": [ + 55, + 533, + 555, + 567 + ], + "spans": [ + { + "bbox": [ + 55, + 533, + 555, + 567 + ], + "type": "text", + "content": "Figure 4. Novel view synthesis and image decomposition results on synthetic scenes, comparing our SpecTRe-GS with 3DGS-DR [39] and GShader [13]. SpecTRe-GS better distinguishes diffuse and specular components in reflective regions, with the least baked-in view-dependent appearance in diffuse components. The masks of reflective regions are shown as insets in the reference image." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 572, + 295, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 572, + 295, + 632 + ], + "spans": [ + { + "bbox": [ + 55, + 572, + 295, + 632 + ], + "type": "text", + "content": "ances through false geometry. In contrast, the influence of this guidance on 3DGS-DR [39] and GShader [13] is twofold: It facilitates the reconstruction of appearance under direct lighting, but also prevents these methods from imitating specular inter-reflections by faking geometry." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 632, + 295, + 704 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 632, + 295, + 704 + ], + "spans": [ + { + "bbox": [ + 55, + 632, + 295, + 704 + ], + "type": "text", + "content": "Qualitative evaluation. In visual comparisons, we show results of GOF [42], GShader* [13], and 3DGS-DR [39] as they are variants with higher qualities. As shown in Fig. 3, our method performs better than the baseline methods for both planar reflectors (rows 1-2) and curved reflectors (row 3), capturing clean, accurate, and high-frequency specular" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 572, + 555, + 704 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 572, + 555, + 704 + ], + "spans": [ + { + "bbox": [ + 313, + 572, + 555, + 704 + ], + "type": "text", + "content": "reflections. Our method also surpasses baseline methods in real-world scenes, as shown in Fig. 1. By visualizing the separate components of PBR-based methods in Fig. 4, we find that incident radiance modeling for both direct and indirect sources is vital for physically meaningful scene appearance decomposition. Otherwise, to fit the overall color in training images, indirect incident radiance is heavily baked into the diffuse component (row 1 for 3DGS-DR [39] and row 3 for GShader [13]). The baked-in incident radiance loses high-frequency details and prevents realistic scene editing in a physically-accurate manner. It should be noted" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16139" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 71, + 294, + 256 + ], + "blocks": [ + { + "bbox": [ + 56, + 71, + 294, + 256 + ], + "lines": [ + { + "bbox": [ + 56, + 71, + 294, + 256 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 294, + 256 + ], + "type": "image", + "image_path": "dbb97ada72e67ea5c45c0be24a591082a042ae59defeb6ae0acd8626bcb09d38.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 258, + 295, + 291 + ], + "lines": [ + { + "bbox": [ + 55, + 258, + 295, + 291 + ], + "spans": [ + { + "bbox": [ + 55, + 258, + 295, + 291 + ], + "type": "text", + "content": "Figure 5. Our reconstruction results on two real-world scenes. SpecTRe-GS accurately reconstructs the geometry of highly specular surfaces alongside the reflection of nearby objects." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 295, + 296, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 295, + 296, + 403 + ], + "spans": [ + { + "bbox": [ + 55, + 295, + 296, + 403 + ], + "type": "text", + "content": "that 3DGS-DR [39] performs better than GShader [13] in reconstructing the direct part of incident radiance from the environment, showing the superiority of surface-based deferred shading over volumetric rendering in PBR modeling of high-frequency reflections. Our method, with well-optimized geometry and consequent accurate incident radiance modeling, achieves satisfactory novel view synthesis of specular reflection on both synthetic and real-world scenes, as also indicated in Fig. 5." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 409, + 149, + 422 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 409, + 149, + 422 + ], + "spans": [ + { + "bbox": [ + 55, + 409, + 149, + 422 + ], + "type": "text", + "content": "4.4. Ablation Study" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 426, + 296, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 426, + 296, + 666 + ], + "spans": [ + { + "bbox": [ + 55, + 426, + 296, + 666 + ], + "type": "text", + "content": "On a subset of our synthetic scenes, we validate the following key component of our method: (1) Normal prior guidance; (2) Joint geometry optimization from incident lighting; (3) ray-traced Indirect incident radiance; (4) Progressive learning; and (5) Depth-aware ray perturbation. The results are shown in Tab. 3. Our full method (Ours) achieves the highest scores, confirming the effectiveness of our method designs. Without the initial normal prior guidance (Ours w/o N.), the optimization process is prone to be trapped into local minima far away from the true underneath geometry of the reflective region. On the other hand, only relying on normal priors without further joint optimization for fine-tuning geometry (Ours w/o J.) also fails in reconstructing accurate geometry for physically-based high-frequency reflections. As discussed above, using only direct incident lighting modeling (Ours w/o I.) leads to baked-in indirect incident lighting and damages the quality of high-frequency components. The progressive learning scheme (Ours w/o P.) and depth-aware ray perturbation (Ours w/o D.) also further helps to faithfully reconstruct the scene." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 672, + 138, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 672, + 138, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 672, + 138, + 685 + ], + "type": "text", + "content": "4.5. Applications" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": "Our Gaussian point cloud representation with explicit geometry, material, and incident lighting facilitates various" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 316, + 117, + 553, + 211 + ], + "blocks": [ + { + "bbox": [ + 313, + 71, + 555, + 115 + ], + "lines": [ + { + "bbox": [ + 313, + 71, + 555, + 115 + ], + "spans": [ + { + "bbox": [ + 313, + 71, + 555, + 115 + ], + "type": "text", + "content": "Table 3. Ablation study with variants of our proposed method excluding: Normal prior guidance, Joint geometry optimization, Indirect incident lighting modeling, Progressive learning, and Depth-aware ray perturbation. " + }, + { + "bbox": [ + 313, + 71, + 555, + 115 + ], + "type": "inline_equation", + "content": "\\uparrow (\\downarrow)" + }, + { + "bbox": [ + 313, + 71, + 555, + 115 + ], + "type": "text", + "content": " means higher (lower) is better." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 316, + 117, + 553, + 211 + ], + "lines": [ + { + "bbox": [ + 316, + 117, + 553, + 211 + ], + "spans": [ + { + "bbox": [ + 316, + 117, + 553, + 211 + ], + "type": "table", + "html": "
MethodEntire ImageReflective Region
PSNR↓SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
Ours w/o N.26.700.8850.14919.800.9130.094
Ours w/o J.26.370.8780.15419.520.9070.100
Ours w/o I.28.490.9050.13022.190.9350.074
Ours w/o P.29.260.9100.11823.160.9400.062
Ours w/o D.29.750.9140.11423.880.9440.056
Ours29.900.9140.11224.050.9440.056
", + "image_path": "806bc9376edaa48fa32966e6dc84f446293fd67985c22ccec4bbd4416455bb24.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 317, + 213, + 553, + 342 + ], + "blocks": [ + { + "bbox": [ + 317, + 213, + 553, + 342 + ], + "lines": [ + { + "bbox": [ + 317, + 213, + 553, + 342 + ], + "spans": [ + { + "bbox": [ + 317, + 213, + 553, + 342 + ], + "type": "image", + "image_path": "194d38599ece68cb60acf280c4538d4f287a2ddc648f9cf9cfff1183081baa16.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 344, + 555, + 377 + ], + "lines": [ + { + "bbox": [ + 313, + 344, + 555, + 377 + ], + "spans": [ + { + "bbox": [ + 313, + 344, + 555, + 377 + ], + "type": "text", + "content": "Figure 6. Our explicit representation of geometry, material, and incident lighting in the Gaussian point cloud enables various editing applications. Insets have been brightened for clarity." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 384, + 555, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 384, + 555, + 469 + ], + "spans": [ + { + "bbox": [ + 313, + 384, + 555, + 469 + ], + "type": "text", + "content": "editing applications. As illustrated in Fig. 6, specular objects of interest can be faithfully captured as assets in the presence of unwanted reflections from nearby objects, enabling further usages such as relighting (column 1) and material editing (column 4). We can also manipulate the scene geometry such as deletion (column 2) and insertion (column 3), with specular inter-reflections adjusted accordingly." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 479, + 388, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 479, + 388, + 491 + ], + "spans": [ + { + "bbox": [ + 314, + 479, + 388, + 491 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 499, + 554, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 499, + 554, + 559 + ], + "spans": [ + { + "bbox": [ + 313, + 499, + 554, + 559 + ], + "type": "text", + "content": "We introduce SpecTRe-GS to model highly specular surfaces with inter-reflections, distinguished by its separate modeling of specular and rough surfaces, the efficient ray tracer, and a geometry-enhancing training strategy for fast and accurate specular inter-reflection rendering." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 559, + 554, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 559, + 554, + 643 + ], + "spans": [ + { + "bbox": [ + 313, + 559, + 554, + 643 + ], + "type": "text", + "content": "Limitations. We only model specular inter-reflections from rough surfaces to avoid multi-bounce path tracing, and we only trace one secondary ray for each pixel in the reflective region, approximating the specular reflection with an ideal mirror. Also, the monocular normal prior may fail in the existence of strong semantic ambiguity (such as mirrors with hidden borders), making our geometry fail to initialize." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 648, + 411, + 661 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 648, + 411, + 661 + ], + "spans": [ + { + "bbox": [ + 313, + 648, + 411, + 661 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 664, + 554, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 664, + 554, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 664, + 554, + 712 + ], + "type": "text", + "content": "This work is supported by the National Natural Science Foundation of China under Grant No. 62136001. PKU-affiliated authors thank openbayes.com for providing computing resource." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16140" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 91, + 296, + 715 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "text", + "content": "[1] Hugo Blanc, Jean-Emmanuel Deschaud, and Alexis Paljic. RayGauss: Volumetric Gaussian-based ray casting for photorealistic novel view synthesis. In Proc. of IEEE Winter Conference on Applications of Computer Vision (WACV), 2024. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 146, + 296, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 146, + 296, + 191 + ], + "spans": [ + { + "bbox": [ + 61, + 146, + 296, + 191 + ], + "type": "text", + "content": "[2] Hongze Chen, Zehong Lin, and Jun Zhang. GI-GS: Global illumination decomposition on Gaussian splatting for inverse rendering. In Proc. of International Conference on Learning Representations (ICLR), 2025. 1, 2, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 191, + 295, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 191, + 295, + 224 + ], + "spans": [ + { + "bbox": [ + 62, + 191, + 295, + 224 + ], + "type": "text", + "content": "[3] James H. Clark. Hierarchical geometric models for visible surface algorithms. Communications of the ACM (CACM), 19(10):547-554, 1976. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 224, + 296, + 290 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 224, + 296, + 290 + ], + "spans": [ + { + "bbox": [ + 62, + 224, + 296, + 290 + ], + "type": "text", + "content": "[4] Jorge Condor, Sebastien Speierer, Lukas Bode, Aljaz Bozic, Simon Green, Piotr Didyk, and Adrian Jarabo. Don't splat your Gaussians: Volumetric ray-traced primitives for modeling and rendering scattering and emissive media. ACM Transactions on Graphics (TOG), 44(1):10:1-10:17, 2025. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 291, + 296, + 356 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 291, + 296, + 356 + ], + "spans": [ + { + "bbox": [ + 62, + 291, + 296, + 356 + ], + "type": "text", + "content": "[5] Michael Deering, Stephanie Winner, Bic Schediwy, Chris Duffy, and Neil Hunt. The triangle processor and normal vector Shader: a VLSI system for high performance graphics. In Proc. of the ACM SIGGRAPH Conference and Exhibition On Computer Graphics and Interactive Techniques (SIGGRAPH), 1988. 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 357, + 295, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 357, + 295, + 390 + ], + "spans": [ + { + "bbox": [ + 62, + 357, + 295, + 390 + ], + "type": "text", + "content": "[6] Cycles Developers. Cycles: Open source production rendering, 2024. https://www.cycles-renderer.org/. Accessed: 2024-11-06. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 391, + 296, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 391, + 296, + 435 + ], + "spans": [ + { + "bbox": [ + 62, + 391, + 296, + 435 + ], + "type": "text", + "content": "[7] Jan-Niklas Dihlmann, Arjun Majumdar, Andreas Engelhardt, Raphael Braun, and Hendrik P. A. Lensch. Subsurface scattering for 3D Gaussian splattering. In Proc. of Neural Information Processing Systems (NeurIPS), 2024. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 435, + 296, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 435, + 296, + 479 + ], + "spans": [ + { + "bbox": [ + 62, + 435, + 296, + 479 + ], + "type": "text", + "content": "[8] Kang Du, Zhihao Liang, and Zeyu Wang. GS-ID: Illumination decomposition on Gaussian splattering via diffusion prior and parametric light source optimization. In arXiv preprint arXiv:2408.08524, 2024. 1, 2, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 479, + 296, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 479, + 296, + 534 + ], + "spans": [ + { + "bbox": [ + 62, + 479, + 296, + 534 + ], + "type": "text", + "content": "[9] Fan Fei, Jiajun Tang, Ping Tan, and Boxin Shi. VMINer: Versatile multi-view inverse rendering with near- and farfield light sources. In Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 536, + 295, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 536, + 295, + 590 + ], + "spans": [ + { + "bbox": [ + 57, + 536, + 295, + 590 + ], + "type": "text", + "content": "[10] Jian Gao, Chun Gu, Youtian Lin, Hao Zhu, Xun Cao, Li Zhang, and Yao Yao. Relightable 3D Gaussian: Real-time point cloud relighting with BRDF decomposition and ray tracing. In Proc. of European Conference on Computer Vision (ECCV), 2024. 1, 2, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 590, + 295, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 590, + 295, + 646 + ], + "spans": [ + { + "bbox": [ + 57, + 590, + 295, + 646 + ], + "type": "text", + "content": "[11] Yijia Guo, Yuanxi Bai, Liwen Hu, Ziyi Guo, Mianzhi Liu, Yu Cai, Tiejun Huang, and Lei Ma. PRTGS: Precomputed radiance transfer of Gaussian splats for real-time high-quality relighting. In Proc. of ACM International Conference on Multimedia (ACM MM), 2024. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 646, + 295, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 646, + 295, + 691 + ], + "spans": [ + { + "bbox": [ + 56, + 646, + 295, + 691 + ], + "type": "text", + "content": "[12] Letian Huang, Jiayang Bai, Jie Guo, Yuanqi Li, and Yanwen Guo. On the error analysis of 3D Gaussian splattering and an optimal projection strategy. In Proc. of European Conference on Computer Vision (ECCV), 2024. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 691, + 295, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 295, + 715 + ], + "type": "text", + "content": "[13] Yingwenqi Jiang, Jiadong Tu, Yuan Liu, Xifeng Gao, Xiaoxiao Long, Wenping Wang, and Yuexin Ma. GaussianShader:" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 715 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 333, + 73, + 555, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 555, + 116 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 555, + 116 + ], + "type": "text", + "content": "3D Gaussian splattering with shading functions for reflective surfaces. In Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 1, 2, 3, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 118, + 555, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 555, + 160 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 555, + 160 + ], + "type": "text", + "content": "[14] James T. Kajiya. The rendering equation. In Proc. of the ACM SIGGRAPH Conference and Exhibition On Computer Graphics and Interactive Techniques (SIGGRAPH), 1986. 2, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 161, + 554, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 161, + 554, + 205 + ], + "spans": [ + { + "bbox": [ + 316, + 161, + 554, + 205 + ], + "type": "text", + "content": "[15] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3D Gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics (TOG), 42(4):139:1-139:14, 2023. 1, 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 205, + 554, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 205, + 554, + 270 + ], + "spans": [ + { + "bbox": [ + 316, + 205, + 554, + 270 + ], + "type": "text", + "content": "[16] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dollar, and Ross B. Girshick. Segment anything. In Proc. of IEEE/CVF International Conference on Computer Vision (ICCV), 2023. 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 270, + 554, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 270, + 554, + 327 + ], + "spans": [ + { + "bbox": [ + 316, + 270, + 554, + 327 + ], + "type": "text", + "content": "[17] Jiahe Li, Jiawei Zhang, Xiao Bai, Jin Zheng, Xin Ning, Jun Zhou, and Lin Gu. DNGaussian: Optimizing sparse-view 3D Gaussian radiance fields with global-local depth normalization. In Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 1, 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 327, + 554, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 327, + 554, + 371 + ], + "spans": [ + { + "bbox": [ + 316, + 327, + 554, + 371 + ], + "type": "text", + "content": "[18] Zhihao Liang, Qi Zhang, Ying Feng, Ying Shan, and Kui Jia. GS-IR: 3D Gaussian splatting for inverse rendering. In Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 1, 2, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 371, + 554, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 371, + 554, + 437 + ], + "spans": [ + { + "bbox": [ + 316, + 371, + 554, + 437 + ], + "type": "text", + "content": "[19] Jiayue Liu, Xiao Tang, Freeman Cheng, Roy Yang, Zhihao Li, Jianzhuang Liu, Yi Huang, Jiaqi Lin, Shiyong Liu, Xiaofei Wu, Songcen Xu, and Chun Yuan. MirrorGaussian: Reflecting 3D Gaussians for reconstructing mirror reflections. In Proc. of European Conference on Computer Vision (ECCV), 2024. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 437, + 554, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 437, + 554, + 492 + ], + "spans": [ + { + "bbox": [ + 316, + 437, + 554, + 492 + ], + "type": "text", + "content": "[20] Yuan Liu, Peng Wang, Cheng Lin, Xiaoxiao Long, Jiepeng Wang, Lingjie Liu, Taku Komura, and Wenping Wang. NeRO: Neural geometry and BRDF reconstruction of reflective objects from multiview images. ACM Transactions on Graphics (TOG), 42(4):114:1-114:22, 2023. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 492, + 554, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 492, + 554, + 548 + ], + "spans": [ + { + "bbox": [ + 316, + 492, + 554, + 548 + ], + "type": "text", + "content": "[21] Alexander Mai, Peter Hedman, George Kopanas, Dor Verbin, David Futschik, Qiangeng Xu, Falko Kuester, Jonathan T. Barron, and Yinda Zhang. EVER: Exact volumetric ellipsoid rendering for real-time view synthesis. In arXiv preprint arXiv:2410.01804, 2024. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 548, + 555, + 603 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 548, + 555, + 603 + ], + "spans": [ + { + "bbox": [ + 316, + 548, + 555, + 603 + ], + "type": "text", + "content": "[22] Jiarui Meng, Haijie Li, Yanmin Wu, Qiankun Gao, Shuzhou Yang, Jian Zhang, and Siwei Ma. Mirror-3DGS: Incorporating mirror reflections into 3D Gaussian splattering. In Proc. of IEEE International Conference on Visual Communications and Image Processing, 2024. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 603, + 554, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 603, + 554, + 658 + ], + "spans": [ + { + "bbox": [ + 316, + 603, + 554, + 658 + ], + "type": "text", + "content": "[23] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing scenes as neural radiance fields for view synthesis. In Proc. of European Conference on Computer Vision (ECCV), 2020. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 658, + 554, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 554, + 715 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 554, + 715 + ], + "type": "text", + "content": "[24] Nicolas Moënne-Loccoz, Ashkan Mirzaei, Or Perel, Ricardo de Lutio, Janick Martinez Esturo, Gavriel State, Sanja Fidler, Nicholas Sharp, and Zan Gojcic. 3D Gaussian ray tracing: Fast tracing of particle scenes. ACM Transactions on Graphics (TOG), 43(6):232:1-232:19, 2024. 2, 5" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16141" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 138 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 138 + ], + "type": "text", + "content": "[25] Steven G. Parker, James Bigler, Andreas Dietrich, Heiko Friedrich, Jared Hoberock, David P. Luebke, David K. McAllister, Morgan McGuire, R. Keith Morley, Austin Robison, and Martin Stich. OptiX: a general purpose ray tracing engine. In ACM Transactions on Graphics (Proc. of ACM SIGGRAPH), 2010. 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 140, + 294, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 140, + 294, + 237 + ], + "spans": [ + { + "bbox": [ + 56, + 140, + 294, + 237 + ], + "type": "text", + "content": "[26] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Köpf, Edward Z. Yang, Zachary DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. PyTorch: An imperative style, high-performance deep learning library. In Proc. of Neural Information Processing Systems (NeurIPS), 2019. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 239, + 294, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 239, + 294, + 293 + ], + "spans": [ + { + "bbox": [ + 56, + 239, + 294, + 293 + ], + "type": "text", + "content": "[27] Lukas Radl, Michael Steiner, Mathias Parger, Alexander Weinrauch, Bernhard Kerbl, and Markus Steinberger. StopThePop: Sorted Gaussian splatting for view-consistent real-time rendering. ACM Transactions on Graphics (TOG), 43(4):64:1-64:17, 2024. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 294, + 294, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 294, + 294, + 326 + ], + "spans": [ + { + "bbox": [ + 56, + 294, + 294, + 326 + ], + "type": "text", + "content": "[28] Christophe Schlick. An inexpensive BRDF model for physically-based rendering. Computer Graphics Forum (CGF), 13(3):233-246, 1994. 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 327, + 294, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 327, + 294, + 381 + ], + "spans": [ + { + "bbox": [ + 56, + 327, + 294, + 381 + ], + "type": "text", + "content": "[29] Yahao Shi, Yanmin Wu, Chenming Wu, Xing Liu, Chen Zhao, Haocheng Feng, Jingtuo Liu, Liangjun Zhang, Jian Zhang, Bin Zhou, Errui Ding, and Jingdong Wang. GIR: 3D Gaussian inverse rendering for reightable scene factorization. In arXiv preprint arXiv:2312.05133, 2023. 1, 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 383, + 294, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 383, + 294, + 415 + ], + "spans": [ + { + "bbox": [ + 56, + 383, + 294, + 415 + ], + "type": "text", + "content": "[30] Zhe Jun Tang and Tat-Jen Cham. 3iGS: Factorised tensorial illumination for 3D Gaussian splatting. In Proc. of European Conference on Computer Vision (ECCV), 2024. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 416, + 294, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 416, + 294, + 468 + ], + "spans": [ + { + "bbox": [ + 56, + 416, + 294, + 468 + ], + "type": "text", + "content": "[31] Eric Veach and Leonidas J. Guibas. Optimally combining sampling techniques for Monte Carlo rendering. In Proc. of the ACM SIGGRAPH Conference and Exhibition On Computer Graphics and Interactive Techniques (SIGGRAPH), 1995. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 471, + 294, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 471, + 294, + 525 + ], + "spans": [ + { + "bbox": [ + 56, + 471, + 294, + 525 + ], + "type": "text", + "content": "[32] Dor Verbin, Peter Hedman, Ben Mildenhall, Todd Zickler, Jonathan T. Barron, and Pratul P. Srinivasan. Ref-NeRF: Structured view-dependent appearance for neural radiance fields. In Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 526, + 294, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 526, + 294, + 591 + ], + "spans": [ + { + "bbox": [ + 56, + 526, + 294, + 591 + ], + "type": "text", + "content": "[33] Dor Verbin, Pratul P. Srinivasan, Peter Hedman, Ben Mildenhall, Benjamin Attal, Richard Szeliski, and Jonathan T. Barron. NeRF-Casting: Improved view-dependent appearance with consistent reflections. In Proc. of the ACM SIGGRAPH Conference and Exhibition on Computer Graphics and Interactive Techniques in Asia (SIGGRAPH Asia), 2024. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 593, + 294, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 593, + 294, + 635 + ], + "spans": [ + { + "bbox": [ + 56, + 593, + 294, + 635 + ], + "type": "text", + "content": "[34] Zhou Wang, A.C. Bovik, H.R. Sheikh, and E.P. Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing (TIP), 13(4):600-612, 2004. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 636, + 294, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 636, + 294, + 690 + ], + "spans": [ + { + "bbox": [ + 56, + 636, + 294, + 690 + ], + "type": "text", + "content": "[35] Zhiru Wang, Shiyun Xie, Chengwei Pan, and Guoping Wang. SpecGaussian with latent features: A high-quality modeling of the view-dependent appearance for 3D Gaussian splatting. In Proc. of ACM International Conference on Multimedia (ACM MM), 2024. 1, 2, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 692, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 692, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 692, + 294, + 713 + ], + "type": "text", + "content": "[36] Tong Wu, Jia-Mu Sun, Yu-Kun Lai, Yuewen Ma, Leif Kobbelt, and Lin Gao. DeferredGS: Decoupled and editable" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 654 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "Gaussian splatting with deferred shading. In arXiv preprint arXiv:2404.09412, 2024. 1, 2, 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 96, + 553, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 149 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 149 + ], + "type": "text", + "content": "[37] Wangze Xu, Huachen Gao, Shihe Shen, Rui Peng, Jianbo Jiao, and Ronggang Wang. MVPGS: Excavating multi-view priors for Gaussian splatting from sparse input views. In Proc. of European Conference on Computer Vision (ECCV), 2024. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 152, + 553, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 553, + 217 + ], + "type": "text", + "content": "[38] Chongjie Ye, Lingteng Qiu, Xiaodong Gu, Qi Zuo, Yushuang Wu, Zilong Dong, Liefeng Bo, Yuliang Xiu, and Xiaoguang Han. StableNormal: Reducing diffusion variance for stable and sharp normal. In Proc. of the ACM SIGGRAPH Conference and Exhibition on Computer Graphics and Interactive Techniques in Asia (SIGGRAPH Asia), 2024. 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 219, + 553, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 219, + 553, + 272 + ], + "spans": [ + { + "bbox": [ + 316, + 219, + 553, + 272 + ], + "type": "text", + "content": "[39] Keyang Ye, Qiming Hou, and Kun Zhou. 3D Gaussian splating with deferred reflection. In Proc. of the ACM SIGGRAPH Conference and Exhibition On Computer Graphics and Interactive Techniques (SIGGRAPH), 2024. 1, 2, 3, 4, 6, 7, 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 274, + 553, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 274, + 553, + 307 + ], + "spans": [ + { + "bbox": [ + 316, + 274, + 553, + 307 + ], + "type": "text", + "content": "[40] Keyang Ye, Qiming Hou, and Kun Zhou. Progressive radiance distillation for inverse rendering with Gaussian splatting. In arXiv preprint arXiv:2408.07595, 2024. 1, 2, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 308, + 553, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 308, + 553, + 352 + ], + "spans": [ + { + "bbox": [ + 316, + 308, + 553, + 352 + ], + "type": "text", + "content": "[41] Zehao Yu, Anpei Chen, Binbin Huang, Torsten Sattler, and Andreas Geiger. Mip-Splatting: Alias-free 3D Gaussian splatting. In Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 354, + 553, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 354, + 553, + 396 + ], + "spans": [ + { + "bbox": [ + 316, + 354, + 553, + 396 + ], + "type": "text", + "content": "[42] Zehao Yu, Torsten Sattler, and Andreas Geiger. Gaussian opacity fields: Efficient adaptive surface reconstruction in unbounded scenes. ACM Transactions on Graphics (TOG), 43(6):271:1-271:13, 2024. 1, 2, 4, 5, 6, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 398, + 553, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 398, + 553, + 453 + ], + "spans": [ + { + "bbox": [ + 316, + 398, + 553, + 453 + ], + "type": "text", + "content": "[43] Libo Zhang, Yuxuan Han, Wenbin Lin, Jingwang Ling, and Feng Xu. PRTGaussian: Efficient relighting using 3D Gaussians with precomputed radiance transfer. In Asia Pacific Signal and Information Processing Association Annual Summit and Conference (APSIPA ASC), 2024. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 454, + 553, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 454, + 553, + 508 + ], + "spans": [ + { + "bbox": [ + 316, + 454, + 553, + 508 + ], + "type": "text", + "content": "[44] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 510, + 553, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 510, + 553, + 553 + ], + "spans": [ + { + "bbox": [ + 316, + 510, + 553, + 553 + ], + "type": "text", + "content": "[45] Yuanqing Zhang, Jiaming Sun, Xingyi He, Huan Fu, Rongfei Jia, and Xiaowei Zhou. Modeling indirect illumination for inverse rendering. In Proc. of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 555, + 553, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 555, + 553, + 586 + ], + "spans": [ + { + "bbox": [ + 316, + 555, + 553, + 586 + ], + "type": "text", + "content": "[46] Yang Zhou, Songyin Wu, and Ling-Qi Yan. Unified Gaussian primitives for scene representation and rendering. In arXiv preprint arXiv:2406.09733, 2024. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 588, + 553, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 588, + 553, + 621 + ], + "spans": [ + { + "bbox": [ + 316, + 588, + 553, + 621 + ], + "type": "text", + "content": "[47] Zuo-Liang Zhu, Beibei Wang, and Jian Yang. GS-ROR: 3D Gaussian splatting for reflective object relighting via SDF priors. In arXiv preprint arXiv:2406.18544, 2024. 1, 2, 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 623, + 553, + 654 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 623, + 553, + 654 + ], + "spans": [ + { + "bbox": [ + 316, + 623, + 553, + 654 + ], + "type": "text", + "content": "[48] Matthias Zwicker, Hanspeter Pfister, Jeroen van Baar, and Markus H. Gross. EWA volume splatting. In Proc. of IEEE Visualization Conference (VIS), 2001. 3" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16142" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Spectral Informed Mamba for Robust Point Cloud Processing/9acc5c2f-4c09-4911-b4a0-485ec3edcc53_content_list.json b/2025/Spectral Informed Mamba for Robust Point Cloud Processing/9acc5c2f-4c09-4911-b4a0-485ec3edcc53_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..cbaf4563e9ca67f7db828fdf006661be6f970567 --- /dev/null +++ b/2025/Spectral Informed Mamba for Robust Point Cloud Processing/9acc5c2f-4c09-4911-b4a0-485ec3edcc53_content_list.json @@ -0,0 +1,1402 @@ +[ + { + "type": "text", + "text": "Spectral Informed Mamba for Robust Point Cloud Processing", + "text_level": 1, + "bbox": [ + 186, + 130, + 810, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ali Bahri * Moslem Yazdanpanah Mehrdad Noori Sahar Dastani \nMilad Cheraghalikhani Gustavo Adolfo Vargas Hakim David Osowiechi Farzad Beizaee Ismail Ben Ayed Christian Desrosiers", + "bbox": [ + 166, + 180, + 830, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "LIVIA, ÉTS Montréal, Canada International Laboratory on Learning Systems (ILLS)", + "bbox": [ + 285, + 241, + 710, + 277 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 311, + 326, + 328 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "State Space Models (SSMs) have shown significant promise in Natural Language Processing (NLP) and, more recently, computer vision. This paper introduces a new methodology that leverages Mamba and Masked Autoencoder (MAE) networks for point-cloud data in both supervised and self-supervised learning. We propose three key contributions to enhance Mamba's capability in processing complex pointcloud structures. First, we exploit the spectrum of a graph Laplacian to capture patch connectivity, defining an isometry-invariant traversal order that is robust to viewpoints and captures shape manifolds better than traditional 3D grid-based traversals. Second, we adapt segmentation via a recursive patch partitioning strategy informed by Laplacian spectral components, allowing finer integration and segment analysis. Third, we address token placement in MAE for Mamba by restoring tokens to their original positions, which preserves essential order and improves learning. Extensive experiments demonstrate the improvements that our approach brings over state-of-the-art baselines in classification, segmentation, and few-shot tasks. The implementation is available at: https://github.com/AliBahri94/SI-Mamba.git.", + "bbox": [ + 89, + 344, + 483, + 676 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 704, + 220, + 719 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The analysis of 3D point cloud data is fundamental for various applications, including autonomous driving [36, 41], VR/AR [18], and robotics [39]. Compared to the organized structure of 2D images, point clouds consist of 3D coordinates without direct adjacency information forming an unordered bag. In recent years, considerable efforts have been dedicated to adapt deep learning models such as convolutional neural networks (CNNs) and Transformers to this type of data [2, 3, 32, 35, 54, 55]. Due to their permutation-invariant self-attention mechanism, Transformer networks", + "bbox": [ + 89, + 729, + 483, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c1ffb0f79fd99dd1c2b963811007e840cd3bd8e4e1fb88d7fea676113098dd50.jpg", + "image_caption": [ + "Figure 1. (a) Surface-Aware Spectral Traversing (SAST) applied to the patched point clouds of a mesh surface. (b) Traversal from left to right, based on the first to fourth non-constant smallest eigenvectors. (c) Traversal based on the largest eigenvector, forming a non-continuous sequence of tokens." + ], + "image_footnote": [], + "bbox": [ + 516, + 311, + 906, + 368 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "are particularly well suited for the unordered nature of point clouds. However, the quadratic complexity of this mechanism, requiring the computation of a weight between each pair of tokens, impedes the application of these networks to large-sized inputs (e.g., 2D images or 3D point clouds represented by many patches). This has prompted researchers to explore more efficient solutions, including the Set Transformer [24], Sparse Transformer [7], Longformer [4] and Sinkhorn Transformer [43].", + "bbox": [ + 511, + 460, + 906, + 595 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, methods based on Structured State Space Sequence (S4) [16] such as Mamba [14] have gained significant traction as a more efficient alternative to Transformers [29, 60]. So far, very few studies have investigated the potential of S4 approaches such as Mamba for 3D point clouds. Existing methods like Point-Mamba [26] and PCM [57] extend the 2D grid-based traversal used for images to a 3D grid. However, this straightforward adaptation to point clouds suffers from three crucial problems. First: whereas patches from 2D images have adjacency information, which could be exploited by the grid-based traversal, the 3D point patches in point clouds offer a sparse representation of the object's surface, and nearby patches on a 3D grid are not necessarily adjacent on this surface. Second: in the absence of self-attention, task-specific performance is highly influenced by the nature of the token traversal strategy. For example, a traversal suitable for point cloud classification may not be effective for a local task such as point-level classification (i.e., segmentation). Third: due to the \"direction-sensitive\" nature of Mamba, the self-supervised MAE pre", + "bbox": [ + 511, + 599, + 908, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Correspondence to ali.bahri.1@ens.etsmtl.ca", + "bbox": [ + 109, + 887, + 352, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "11799", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "training step of leading point cloud models like Point-MAE [32] and Point-M2AE [55] cannot be used directly as there is no attention mechanism to learn the masked tokens' positions.", + "bbox": [ + 89, + 90, + 480, + 150 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The contribution of our work focuses on addressing these problems as follows:", + "bbox": [ + 89, + 152, + 482, + 181 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. We introduce a Surface-Aware Spectral Traversing (SAST) strategy based on the Laplacian spectrum of a patch-connectivity graph. Compared to the 3D grid traversal of current approaches like Point-Mamba, our strategy is invariant to isometric transformations (e.g., choice of viewpoint) and better captures the object's surface manifold.", + "2. We also present a Hierarchical Local Traversing (HLT) for point-level classification (segmentation) that partitions patches recursively based on their spectral coordinates. Unlike our SAST strategy for classification, which considers Laplacian eigenvectors separately in different traversals, this HLT combines them in a single ordering for a more precise modeling of geometry.", + "3. During the MAE-based Self-Supervised Learning (SSL), we propose a Traverse-Aware Repositioning (TAR) strategy to align the masked tokens according to their spectral adjacency. This strategy addresses the critical issue of spatial adjacency preservation unique to Mamba networks." + ], + "bbox": [ + 89, + 184, + 482, + 488 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 89, + 506, + 238, + 522 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Deep Point Cloud Learning. Deep neural networks have increasingly been applied to point clouds, with early architectures like PointNet [34] and PointNet++ [35] inspiring further work [1, 10, 23, 25, 58] focused on capturing local context. Transformer-based frameworks [46], including the Point Transformer (v1-v3) [50, 51, 59] and Stratified Transformer [22], have set new benchmarks by integrating both local and global features. Recently, self-supervised pre-training has gained traction by leveraging masked point modeling [28, 42] to train Transformers on unlabeled data.", + "bbox": [ + 89, + 536, + 482, + 686 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Building on the effectiveness of MAE in Text and Image domains, Point-BERT [54] presented a revolutionary method inspired by BERT [11], tailoring Transformers to 3D point cloud processing. Point-MAE [32] applied MAE-style pre-training to 3D point clouds using a custom Transformer-based Autoencoder (AE) designed to reconstruct masked irregular patches. The use of multiscale masking and local spatial self-attention mechanisms in Point-M2AE [55] has led to State-Of-The-Art (SOTA) results in 3D representation learning. Furthermore, I2P-MAE [56] improved self-supervised point cloud processing with a masking strategy leveraging pre-trained 2D models through an Image-to-Point transformation. Point-GPT [6] introduced an auto-regressive generative pretraining (GPT)", + "bbox": [ + 89, + 688, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "approach to address the unordered nature and low information density of point clouds. ACT [12] proposed a cross-modal knowledge transfer method using pretrained 2D or natural language Transformers as teachers for 3D representation learning. Finally, GeoMask3D [2] is a self-supervised point cloud method that uses a teacher-student model to select harder, geometrically intricate regions for masking, thereby improving feature representations.", + "bbox": [ + 511, + 90, + 903, + 212 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. State Space Models", + "text_level": 1, + "bbox": [ + 511, + 220, + 697, + 237 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "State Space Models. State-space models (SSMs) have been foundational in control theory and signal processing for modeling dynamic systems. Recently, they have been adapted to deep learning, with advancements like the Linear State-Space Layer (LSSL), which uses a continuous-time memorization framework based on the HiPPO operator [15] to model long-range dependencies. However, LSSL's high computational demands limit its practicality. To address this, the S4 model [16] introduced parameter normalization techniques, paving the way for structured SSMs with enhancements like complex-diagonal structures [17, 19], support for multiple-input/output [17], and low-rank decomposition [20]. Recent developments include Mamba [14], which achieves linear-time inference and efficient training through selection mechanisms and hardware-aware algorithms. MoE-Mamba further boosts efficiency by integrating a Mixture of Experts (MoE), outperforming both Mamba and Transformer-MoE models [33].", + "bbox": [ + 511, + 244, + 903, + 517 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "SSMs for Vision Tasks The above-mentioned works primarily focused on the application of SSMs to long-range or causal data types such as language and speech. In the field of vision, a notable study [29] proposed the VMamba model which features a Cross-Scan Module (CSM) for enhanced 1D selective scanning in 2D spaces and architectural optimizations that significantly improve its performance and speed across various visual tasks. Another significant paper is Vision Mamba [60] which introduces a novel vision backbone called Vim utilizing bidirectional Mamba blocks.", + "bbox": [ + 511, + 520, + 903, + 671 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "For point cloud analysis based on Mamba, two key works are Point-Mamba [26] and PCM [57]. PointMamba introduces a simple approach to token reordering for point cloud analysis by strategically organizing point tokens based on a 3D grid. Similarly, PCM enhances Mamba with a Consistent Traverse Serialization (CTS) technique that converts 3D point clouds into 1D point sequences while maintaining spatial adjacency. Building upon these approaches, our method introduces the Spectral Spatial Traversing (SST) strategy, which improves token ordering and maintains spatial adjacency within Mamba networks.", + "bbox": [ + 511, + 672, + 903, + 838 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 511, + 851, + 602, + 866 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We begin by outlining the fundamental concepts of SSMs and spectral graph analysis which are at the core of our", + "bbox": [ + 511, + 869, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "11800", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/96e7d1d4fec27302183146eb7b4f1db41038e8bd4dc37c256de2af823c5b5507.jpg", + "image_caption": [ + "Figure 2. Overview of the proposed Spectral Spatial Traversing (SST) method. (a) Point cloud, (b) Patchification, (c) Forming the adjacency graph, (d) Traversal based on SAST using $s$ non-constant smallest eigenvectors, (e) HLT for segmentation tasks, (f) TAR strategy for Masked Autoencoders. The process includes reverse and concatenation operations, with learnable tokens, representations, and masked tokens highlighted. (g) The classification task involves sorting tokens by different eigenvectors, concatenating them, and then feeding them into the network. (h) The segmentation task where HLT is applied on the tokens $(\\vec{q})$ and $\\vec{q}$ is fed into the network. (i) A flowchart visualizing the techniques used in self-supervised learning and various downstream tasks." + ], + "image_footnote": [], + "bbox": [ + 169, + 88, + 836, + 518 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "work. We then give an in-depth presentation of our Surface-Aware Spectral Traversing (SAST) strategy for point cloud processing that improves the model's robustness to isometric transformations and better captures the underlying manifold of the point cloud. Thereafter, we provide detailed specifications of our Hierarchical Local Traversing (HLT) strategy for point-level classification, which defines a more structured patch traversal order based on the recursive partitioning of spectral information. Finally, we introduce our Traverse-Aware Repositioning (TAR), which improves the handling of learnable tokens in masked autoencoders within Mamba networks. Fig. 2 illustrates the overview of the proposed Spectral Spatial Traversing (SST) method.", + "bbox": [ + 89, + 636, + 485, + 833 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminaries", + "text_level": 1, + "bbox": [ + 89, + 844, + 232, + 859 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "State Space Models (SSMs) use a series of first-order differential equations to describe how the state of the linear,", + "bbox": [ + 89, + 869, + 483, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "time-invariant system evolves over time:", + "bbox": [ + 513, + 636, + 782, + 651 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\dot {h} (t) = A h (t) + B x (t), \\quad y (t) = C \\dot {h} (t) + D x (t), \\quad (1)\n$$\n", + "text_format": "latex", + "bbox": [ + 527, + 661, + 906, + 681 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Here, $\\dot{h}(t)$ denotes the time derivative of the state vector $h(t)$ . The matrices $A$ , $B$ , $C$ , and $D$ are the weighting parameters.", + "bbox": [ + 511, + 691, + 906, + 736 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Due to their reliance on continuous data streams $x(t)$ , SSMs are not natively equipped to handle discrete inputs represented as $\\{x_0, x_1, \\ldots\\}$ . This necessitates the use of a discretized SSM version for practical applications:", + "bbox": [ + 511, + 738, + 906, + 799 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nh _ {k} = \\bar {A} h _ {k - 1} + \\bar {B} x _ {k}, \\quad y _ {k} = \\bar {C} h _ {k} + \\bar {D} x _ {k}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 563, + 810, + 906, + 828 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The state space model in its discrete version utilizes a recursive function to link each state $h_k$ to its preceding state, encapsulated by the matrices $\\bar{A} \\in \\mathbb{R}^{N \\times N}$ , $\\bar{B} \\in \\mathbb{R}^{N \\times 1}$ , and $\\bar{C} \\in \\mathbb{R}^{N \\times 1}$ , which are tuned parameter matrices. While", + "bbox": [ + 511, + 840, + 908, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "11801", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/c3d67f5050addff6fa87af5f531b71eb6f0618300350fd11ef6b1ee75ca2beb9.jpg", + "image_caption": [ + "Figure 3. (a) Visualization of the four non-constant smallest Laplacian eigenvectors $(v^{(k)}, k = 1, \\ldots, 4)$ and (b) the discrete partitioning $(q)$ of our HLT strategy combining the information of all four eigenvectors. Note: we assumed that patches contain a single point for better visualization." + ], + "image_footnote": [], + "bbox": [ + 98, + 88, + 478, + 236 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "matrix $\\bar{D} \\in \\mathbb{R}^{N \\times 1}$ may be employed as a residual connection, we follow previous work and exclude it from our model. The transition from a continuous signal representation $x(t)$ to a discrete sequence involves sampling $x(t)$ at intervals defined by $\\Delta$ , setting each discrete input as $x_{k} = x(k\\Delta)$ . This adjustment to a discrete framework results in revised matrix definitions:", + "bbox": [ + 89, + 323, + 483, + 431 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\bar {A} = \\left(I - \\frac {\\Delta}{2} A\\right) ^ {- 1} \\left(I + \\frac {\\Delta}{2} A\\right), \\bar {B} = \\left(I - \\frac {\\Delta}{2} A\\right) ^ {- 1} \\Delta B, \\bar {C} = C. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 436, + 483, + 477 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "However, the fixed dynamics of Linear Time-Invariant (LTI) models, exemplified by the constant parameters $A$ , $B$ , and $C$ in Eq. (3), restrict their capacity to selectively retain or discard relevant information, thereby limiting their contextual awareness. To improve content-aware reasoning, we use Mamba's selection mechanism that manages the propagation and interaction of information across the sequence dimension [14].", + "bbox": [ + 89, + 479, + 483, + 599 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Spectral Graph Analysis. Popularized by Chung in the 90s [8], spectral graph analysis characterizes the properties of a graph $G = (V, E)$ by the spectrum (eigenvalues and corresponding eigenvectors) of its Laplacian matrix $L$ . This analysis can be understood as a discretized version of the Laplace-Beltrami Operator $\\Delta$ of a function $f$ defined on a Riemannian manifold:", + "bbox": [ + 89, + 606, + 483, + 712 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\Delta f = \\operatorname {d i v} (\\operatorname {g r a d} f) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 222, + 722, + 482, + 739 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\operatorname{grad}f$ is the gradient of $f$ and $\\operatorname{div}$ the divergence on the manifold. The solution to the Laplacian eigenvalue problem $\\Delta f = -\\lambda f$ , known as Helmholtz wave equation, is an eigenfunction corresponding to the natural vibration form of a homogeneous membrane with eigenvalue $\\lambda$ [38].", + "bbox": [ + 89, + 750, + 482, + 824 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Following methods for spectral clustering [31] and normalized cuts [40], we consider a weighted adjacency matrix $W: V \\times V \\to \\mathbb{R}_+$ where $W_{ij} = 0$ if $(i,j) \\notin E$ to model the Euclidean distance of nearby patches (see Section 3.3). The Laplacian matrix of $G$ is defined as $L = D - W$ where", + "bbox": [ + 89, + 825, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$D$ is the diagonal degree matrix such that $D_{ii} = \\sum_{j}W_{ij}$ . To account for variability in the scale of weights $W_{ij}$ or the distribution of node degrees $D_{ii}$ , it is preferable to employ a normalized version of the Laplacian. In this work, we use the Random Walk Laplacian $L_{rw} = I - D^{-1}W$ which has the following useful properties:", + "bbox": [ + 511, + 90, + 903, + 181 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. $L_{rw}$ is positive semi-definite and has $|V|$ non-negative real-valued eigenvalues $0 = \\lambda_1 \\leq \\ldots \\leq \\lambda_{|V|}$ ;", + "2. 0 is an eigenvalue of $L_{rw}$ with the constant vector as eigenvector and its multiplicity equals the number of connected components in the graph;", + "3. Following Courant's Nodal Line Theorem [9], the $n$ -th eigenmode of $L_{rw}$ has at most $n$ poles of vibration;", + "4. The representation of a shape by the spectrum of $L_{rw}$ is invariant to isometry (i.e., distance-preserving transformation)." + ], + "bbox": [ + 511, + 181, + 903, + 332 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our method uses the $s$ first non-constant eigenvectors of $L_{rw}$ (i.e., the eigenvectors corresponding to the $s$ smallest non-zero eigenvalues) to define traversal orders for classification (Section 3.3) and segmentation (Section 3.4) that are robust to the viewpoint (due to isometry invariance) and provide a smooth parametrization of the surface manifold. We consider the first eigenvectors as they encode low frequency information (by Courant's Nodal Line Theorem), making the resulting traversal more robust to shape variability and noise. Fig. 1 illustrates this concept: (a) shows the original mesh, (b) shows traversals based on the first to fourth non-constant smallest eigenvectors, and (c) shows traversal based on the largest eigenvector forming a non-continuous sequence of tokens.", + "bbox": [ + 511, + 334, + 906, + 546 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Point Cloud Patchification", + "text_level": 1, + "bbox": [ + 513, + 556, + 751, + 571 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given a point cloud $\\mathcal{P} = \\{p_i\\}_{i=1}^{N_p}$ , each point represented by 3D coordinates, we convert $\\mathcal{P}$ to a reduced set of patches that can be processed more efficiently. Toward this goal, we employ the Farthest Point Sampling (FPS) algorithm to select a subset $\\mathcal{C} \\subset \\mathcal{P}$ of $N_c$ points offering a good coverage of the entire point cloud. These selected points will act as the centers of local patches within the point cloud. For each center point $p_{s_i} \\in \\mathcal{C}$ , we then identify $N_n$ nearest points $\\mathcal{N}(p_{s_i}) \\subset \\mathcal{P}$ using the K-Nearest Neighbours (KNN) algorithm. Following this, each patch is defined as a center $p_{s_i}$ and its corresponding nearest-neighbors $\\mathcal{N}(p_{s_i})$ .", + "bbox": [ + 511, + 577, + 906, + 746 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Surface-Aware Spectral Traversing (SAST)", + "text_level": 1, + "bbox": [ + 511, + 756, + 880, + 773 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Current point cloud processing approaches using Mamba, such as Point-Mamba [26] and PCM [57], simply extend the 2D grid-based traversal for images to a 3D grid. As mentioned before, this naive strategy suffers from two issues: 1) the 3D grid is view dependent, thus rotating the point cloud or moving the camera yields a different traversal order; 2) unrelated patches may be adjacent in 3D space, hence can be traversed subsequently. To address these problems, we", + "bbox": [ + 511, + 779, + 906, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "11802", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "define a traversal order based on the Laplacian spectrum of the patch-connectivity graph.", + "bbox": [ + 89, + 90, + 485, + 121 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this graph, each node corresponds to a patch and the weighted adjacency matrix $W$ is defined using the Euclidean distance between patch centers. For patches $i$ and $j$ defined by center points $p_{s_i}$ and $p_{s_j}$ , we add an edge $(i,j)$ if $p_{s_j}$ is among the $K$ nearest neighbors of $p_{s_i}$ or vice-versa. The weight of this edge is computed using a Gaussian kernel: $W_{ij} = \\exp \\left(-\\| p_{s_i} - p_{s_j}\\| _2^2 /\\sigma\\right)$ where $\\sigma$ is a hyperparameter controlling the kernel width.", + "bbox": [ + 89, + 122, + 483, + 243 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Following Section 3.1, we compute the $s$ first nonconstant eigenvectors of the Random Walk Laplacian $L_{rw}$ . This can be achieved efficiently using an iterative method like the Arnoldi algorithm [13] by exploiting the following facts: 1) matrix $W$ is very sparse, and 2) only the first few eigenvectors need to be computed. Eigenvector $v^{(k)} \\in \\mathbb{R}^{N_c}$ , $k \\in \\{1, \\dots, s\\}$ , assigns an eigenfunction value $v_i^{(k)}$ to each patch $i$ . In each Mamba block of our model, we perform two separate traversals of tokens (each token corresponds to an input patch) for every eigenvector: a forward traversal by increasing value of $v_i^{(k)}$ and a reserve traversal by decreasing value of $v_i^{(k)}$ . At the end of the block, we concatenate the tokens from the $s \\times 2$ traversals. The visual effect of each traversal is illustrated in Fig. 1(b) and Fig. 3(a).", + "bbox": [ + 89, + 244, + 483, + 464 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Canonicalization of spectrum. Although the spectrum of $L_{rw}$ forms an isometry-invariant representation of the surface manifold, this representation may be impacted by two sources of ambiguity: 1) the sign of eigenvectors is undetermined (i.e., if $v^{(k)}$ is an eigenvector, then so is $-v^{(k)}$ ), and 2) the order of eigenvectors with similar eigenvalues may vary. We address these two sources of ambiguity with the following canonicalization procedure. For the first one, we flip the sign of an eigenvector $v^{(k)}$ (i.e., $v^{(k)} \\coloneqq -v^{(k)}$ ) if its first element is negative (i.e., $v_{1}^{(k)} < 0$ ). To handle the second ambiguity, we first sort the eigenvectors by nondecreasing eigenvalue. We deal with eigenvalues having a multiplicity greater than one by finding pairs of consecutive eigenvectors $v^{(k)}$ , $v^{(k+1)}$ with near-identical eigenvalues (i.e., $|\\lambda_k - \\lambda_{k+1}| \\leq \\epsilon$ ). For such pairs, we flip the order if $v_{1}^{(k)} > v_{1}^{(k+1)}$ . This reordering process is repeated until no further change occurs.", + "bbox": [ + 89, + 465, + 483, + 729 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Hierarchical Local Traversing (HLT) for Segmentation", + "text_level": 1, + "bbox": [ + 89, + 739, + 482, + 771 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "While effective for classification tasks, the SAST strategy considering each eigenvector in a separate traversal may not capture the precise relationship between patches needed for segmentation. To address this issue, we introduce a Hierarchical Local Traversal (HLT) strategy that considers the full spectrum (all $s$ non-constant eigenvectors) simultaneously.", + "bbox": [ + 89, + 777, + 482, + 869 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our strategy is inspired by the recursive binary partitioning technique of normalized cuts [40]. Starting from", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "the canonicalized spectrum (see Section 3.3), tokens are first split based on the first eigenvector $v^{(1)}$ , by comparing their corresponding value in $v^{(1)}$ with the mean value $\\overline{v}^{(1)}$ . This yields a binary partition of tokens $b_{i}^{(1)} = \\mathbb{1}\\big(v_{i}^{(1)}\\geq \\overline{v}^{(1)}\\big)\\in \\{0,1\\}$ where $\\mathbb{1}$ is the indicator function. Each subgroup is then divided based on the mean of the second eigenvector $v^{(2)}$ , and so on for other eigenvectors. This partitioning process can be seen as building a binary tree, where each level corresponds to a different eigenvector and leaf nodes $i$ are uniquely identified by the sequence of bits $b_{i} = [b_{i}^{(1)},\\dots ,b_{i}^{(s)}]$ on the path from the root to the leaf. Our HLT method traverses groups of leaf nodes (groups of tokens) sequentially based on the lexicographic order of their binary code (e.g., [0000], [0001], [0010], [0011], ... in the case of four eigenvectors). For convenience, we convert binary codes $b_{i}$ to a non-negative integer $q_{i}$ (e.g., bin2Int([0011]) = 3) and define two traversal orders, by increasing or decreasing values of $q_{i}$ .", + "bbox": [ + 511, + 90, + 903, + 369 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For $s$ eigenvectors, the HLT strategy described above divides tokens into $2^{s}$ segments which are traversed sequentially. In the best case scenario, $\\lceil \\log_2(N_c)\\rceil$ eigenvectors are thus needed to split tokens into individual segments. However, it may happen that multiple tokens fall in the same segment, especially when using fewer eigenvectors. In such case, one can further sort tokens within each segment, for example, using the values of the first eigenvector (i.e., $v^{(1)}$ ). In our implementation, we simply sort these tokens randomly to add stochasticity in the training. This section is illustrated in Fig. 2 (e), and Fig. 3 (b).", + "bbox": [ + 511, + 369, + 903, + 535 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As shown in Fig. 3, the first Laplacian eigenvectors encode high-level spatial relations (e.g., bottom vs. top, left vs. right, torso vs. limbs, etc.). In the SAST, because these eigenvectors are used in separate traversals, the network may not be able to differentiate specific regions/parts of the point cloud. In contrast, our HLT strategy can capture such specific parts (e.g., head, left or right arm/thigh/alf, pelvis, etc.).", + "bbox": [ + 511, + 536, + 905, + 655 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Traverse-Aware Repositioning (TAR) for Masked Autoencoders", + "text_level": 1, + "bbox": [ + 511, + 667, + 903, + 696 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Following state-of-art Transformers for point cloud processing, such as Point-MAE [32] and Point-M2AE [55], our method leverages self-supervised pretraining based on MAE to boost performance. In Transformer-based approaches, the learnable tokens of masked patches can be inserted in any position of the sequence (typically at the end) as the self-attention mechanism can still attend to all tokens irrespective of their positions. However, this approach presents a significant problem in Mamba networks which are sensitive to the traversal order of tokens. We handle this problem via a TAR strategy that improves the placement of learnable tokens in MAE within Mamba networks. Specifically, we restore the learnable tokens to their", + "bbox": [ + 511, + 704, + 905, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "11803", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "original positions rather than appending them at the end of the sequence. This ensures that the essential order of tokens is maintained, preserving spatial adjacency and enhancing learning effectiveness within Mamba networks.", + "bbox": [ + 88, + 90, + 480, + 150 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The proposed TAR strategy selects an arbitrary traversal order and randomly masks a subset of $N_{m}$ tokens with the same masking ratio as the transformer-based MAEs. These tokens are then removed from the sequence, and their positions are recorded. Afterwards, the remaining (visible) tokens are fed to the encoder that outputs their representation. Before reconstructing the point cloud using the decoder, we reinsert the learnable tokens in the sequence at their recorded position. The set of masked patches is used for other traversal orders. This procedure can be seen in Fig. 2 (f). Following previous work, we measure the reconstruction error for masked patches using the Chamfer distance:", + "bbox": [ + 89, + 151, + 482, + 345 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {r e c} = \\frac {1}{N _ {m}} \\sum_ {i = 1} ^ {N _ {m}} \\operatorname {C h a m f e r} \\left(\\mathcal {S} _ {i}, \\hat {\\mathcal {S}} _ {i}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 166, + 351, + 480, + 392 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $S_{i}\\in \\mathbb{R}^{N_{n}\\times 3}$ is the set of points forming the $i$ -th masked patch and $\\hat{S}_i$ the reconstructed output for these points. The Chamfer distance between two sets of points $S$ and $\\hat{S}$ is defined as", + "bbox": [ + 89, + 398, + 482, + 459 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\operatorname {C h a m f e r} \\left(\\mathcal {S}, \\mathcal {S} ^ {\\prime}\\right) = \\sum_ {p \\in \\mathcal {S}} \\min _ {p ^ {\\prime} \\in \\mathcal {S} ^ {\\prime}} \\| p - p ^ {\\prime} \\| _ {2} ^ {2} \\\\ + \\sum_ {p ^ {\\prime} \\in \\mathcal {S} ^ {\\prime}} \\min _ {p \\in \\mathcal {S}} \\| p - p ^ {\\prime} \\| _ {2} ^ {2}. \\tag {6} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 137, + 464, + 480, + 534 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 550, + 222, + 566 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Several experiments are conducted to evaluate the proposed method. First, we pretrain the Point-Mamba network us", + "bbox": [ + 89, + 575, + 482, + 606 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/696cd3e92ac7f6ae4b03d9860d1207fff992acbeb86df0a7ced8d3443d8f20ac.jpg", + "table_caption": [ + "Table 1. Few-shot classification on ModelNet40. We report the average accuracy (\\%) and standard deviation (\\%) of 10 independent experiments. \\* denotes reproduced results. A $\\equiv$ represents the average (A) and standard deviation (std), respectively." + ], + "table_footnote": [], + "table_body": "
Method5-way10-way
10-shot20-shot10-shot20-shot
DGCNN [48]91.83793.43286.3390.937
DGCNN + OcCo [47]91.93393.93186.4391.346
Transformer [54]87.83293.3384.6389.43
Transf. + OcCo [54]94.03695.9389.4392.446
Point-BERT [54]94.63196.3391.0392.73
Point-M2AE [55]96.83898.3492.3395.039
Point-MAE [32]96.33597.8392.6395.039
PointMamba* [26]95.93197.3391.6394.535
Ours96.43598.5392.0395.13
", + "bbox": [ + 116, + 691, + 452, + 898 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ing our techniques on the ShapeNet [5] training dataset. We then assess the performance of these pretrained models across a variety of standard benchmarks, including object classification, few-shot learning, and segmentation. Additionally, we train the model from scratch on downstream datasets to demonstrate the robustness and versatility of our method. To have a fair comparison, we adopt the masking ratio (60%) that was used in the Point-Mamba model. Moreover, a comprehensive analysis of the computational efficiency, runtime, and memory usage of our SAST approach is provided in the Supplementary Material.", + "bbox": [ + 511, + 90, + 903, + 257 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Pretraining Setup", + "text_level": 1, + "bbox": [ + 511, + 267, + 687, + 284 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Following Point-Mamba, we adopt the ShapeNet [5] dataset for the pretraining and assess the quality of the 3D representations produced by our approach through a linear evaluation on the ModelNet40 [52] dataset. The linear evaluation is performed by a Support Vector Machine (SVM) fitted on these features. This classification performance is quantified by the accuracy metric. Augmentation in pretraining is scale and transform, while in finetuning it is rotation.", + "bbox": [ + 511, + 290, + 903, + 411 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Downstream Tasks", + "text_level": 1, + "bbox": [ + 511, + 422, + 694, + 436 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Object Classification on Real-World Dataset. To evaluate our method for point clouds, we test it on the ScanObjectNN dataset [44], as described in previous studies. The augmentation used during training is random rotation. The results, presented in Tab. 3, show that our strategy signifi", + "bbox": [ + 511, + 446, + 903, + 523 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/c14631b690e81c460571edd80050a8d2e1315bf1167adfefa4c51e92e2428fa7.jpg", + "table_caption": [ + "Table 2. Part segmentation on the ShapeNetPart [53]. The mIoU for all instances (Inst.) is reported. HLT is Hierarchical Local Traversing for segmentation." + ], + "table_footnote": [], + "table_body": "
MethodsBackboneFLOPs (G)mIoU
Training from scratch
PointNet [34]--83.7
PointNet++ [35]--85.1
DGCNN [48]--85.2
APES [49]--85.8
Point-MAE [32]Transformer15.585.7
PointMamba [26]Mamba14.385.8
Ours (HLT)Mamba14.385.9
Training from pretrained
Transformer [54]Transformer15.585.1
Point-BERT [54]Transformer15.585.6
Point-MAE [32]Transformer15.586.1
Point-M2AE [55]Transformer15.586.5
Point-GPT-S [6]Transformer-86.2
ACT [12]Transformer-86.2
I2P-MAE [56]Transformer-86.8
PointMamba [26]Mamba14.386.0
Ours (SAST)Mamba14.385.7
Ours (HLT)Mamba14.386.1
", + "bbox": [ + 539, + 601, + 875, + 906 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "11804", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/e88eb1893beaeaa53821f6e421b1c48a635d35d4e242ac6fa19df4779be50d53.jpg", + "table_caption": [ + "Table 3. Object classification on ScanObjectNN [45]. Accuracy (%) is reported. ${}^{ \\dagger }$ indicates that this method was fine-tuned without rotation augmentation." + ], + "table_footnote": [], + "table_body": "
MethodsBackboneParam. (M)FLOPs (G)OBJ-BGOBJ-ONLYPB-T50-RSModelNet40
Training from scratch
PointNet [34]-3.50.573.379.268.089.2
PointNet++ [35]-1.51.782.384.377.990.7
DGCNN [48]-1.82.482.886.278.192.9
PointNeXt [37]-1.41.6--87.792.9
PointMLP [30]-13.231.4--85.4-
ADS [21]-----87.5-
Transformer [54]Transformer22.14.879.8680.5577.24-
Point-MAE [32]Transformer22.14.886.7586.9280.7892.3
PointMamba [26]Mamba12.33.690.8790.1885.6092.4
OursMamba12.33.692.2591.3987.3092.7
Training from pretrained
Transformer [54]Transformer22.14.879.8680.5577.2492.1
Transformer-OcCo [54]Transformer--84.8585.5478.79-
Point-BERT [54]Transformer22.14.887.4388.1283.0792.7
Point-M2AE† [55]Transformer--91.2288.8186.4394.0
Point-MAE [32]Transformer22.14.892.7791.2289.0493.2
PointMamba [26]Mamba12.33.693.2991.5688.1792.8
PCM [57]Mamba12.33.6--86.9-
OursMamba12.33.694.3291.9189.1093.4
", + "bbox": [ + 143, + 128, + 854, + 462 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "cantly improves object classification accuracy in both training from scratch and fine-tuning scenarios. These findings highlight the effectiveness of our approach in enhancing the model's ability to identify and classify objects across various backgrounds, demonstrating its robustness in complex real-world scenarios.", + "bbox": [ + 88, + 486, + 480, + 575 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Object Classification on Clean Objects Dataset. We also evaluated our method on the ModelNet40 [52] dataset, following the protocols established in previous works. The augmentation used during training is scale and transform. As shown in Tab. 3, our approach achieves notable enhancements on this challenging dataset compared to both the original Point-Mamba and the Transformer-based Point-MAE. This demonstrates the robustness and effectiveness of our method when applied to the Point-Mamba network.", + "bbox": [ + 88, + 585, + 482, + 722 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Few-shot Learning. We conducted few-shot learning experiments on ModelNet40 [52] dataset, adhering to the protocols of previous studies [27, 32, 55]. The results of our few-shot learning experiments are presented in Tab. 1. Despite the competitive nature of this benchmark, our method demonstrated outstanding performance across all tested scenarios. As shown in Tab. 1, our Mamba-based method achieves results comparable to or exceeding those of transformer-based methods (Point-MAE and PointM2AE).", + "bbox": [ + 88, + 731, + 482, + 881 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Part Segmentation. We evaluated our method's repre", + "bbox": [ + 89, + 885, + 482, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "sensation learning ability on the ShapeNetPart dataset [53] using standard experimental settings [34, 35, 54]. Tab. 2 shows that performance gains in prior methods are minimal, highlighting the dataset's challenge. For instance, PointGPT [6] and ACT [12], despite being state-of-the-art and complex methods, show only marginal improvements over each other and other state-of-the-art approaches. Similarly, I2P-MAE [56], which combines 3D data with 2D information, shows limited improvement over Point-M2AE. For the \"Training from scratch\" setting, our method outperforms several state-of-the-art approaches. In the \"Training from pretrained\" setting, we further demonstrate the effectiveness of HLT strategy compared to SAST in the segmentation task.", + "bbox": [ + 511, + 486, + 906, + 696 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Studies", + "text_level": 1, + "bbox": [ + 511, + 705, + 679, + 720 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we aim to investigate the effects of different parameters on our method. We will focus on two key aspects: the effect of the number of non-constant smallest eigenvectors and the adjacency matrix used in the SAST strategy, and the analysis of the TAR strategy.", + "bbox": [ + 511, + 727, + 905, + 804 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Analysis of Eigenvectors and Graph. One of our ablation studies investigates the impact of the number of nonconstant smallest eigenvectors used in the SAST and TAR strategies. As depicted in Fig. 4 (Left), the best performance is achieved when using the four non-constant smallest eigenvectors (blue line) for traversing. When the num", + "bbox": [ + 511, + 809, + 906, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "11805", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a183ad70f343dd6cb6973cdc3ae5ce3f391de0338854a0167e48131c1bdcd942.jpg", + "image_caption": [ + "Figure 4. Analysis of the number of non-constant smallest eigenvectors and comparison with previous methods (Left) and analysis of the number of nearest neighbors $K$ (Right)." + ], + "image_footnote": [], + "bbox": [ + 133, + 90, + 465, + 250 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/45fb3d852d239418c06ea2d6bdb397a47f2fc819eb286f4e5a144bce4c88264f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 537, + 92, + 861, + 244 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b9858906576a19268085fad3e2faaf70055d2e773b96a86802ba0a297698ed8f.jpg", + "image_caption": [ + "Figure 5. The effect of the TAR strategy in the pretraining phase (Left) and in finetuning (Right)." + ], + "image_footnote": [], + "bbox": [ + 133, + 303, + 460, + 460 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/439b0cdc4c61d8661fdf6bca699605ea9dc2092531996e82d66ce847a0c8b077.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 537, + 308, + 861, + 455 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ber of non-constant smallest eigenvectors increases beyond four, the performance drops. This is because the additional eigenvectors are closer to the largest eigenvectors, which are less smooth and do not capture the most significant structural variations effectively. The significantly lower performance of Point-MAE [32], Point-Mamba [26], and PointMamba with random traversing (Random-Mamba) compared to ours underscores the importance of appropriate traversing for Mamba networks.", + "bbox": [ + 88, + 500, + 482, + 633 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Another study examines the impact of the number of nearest neighbors used in creating the adjacency matrix. As shown in Fig. 4 (Right), the best performance (blue line) is achieved with 20 nearest neighbors. The model's accuracy increases as the number of nearest neighbors increases from 10 to 20, reaching its peak at 20 nearest neighbors. Beyond this point, the performance starts to decline. All these experiments are trained and evaluated from scratch on the ScanObjectNN dataset [44] (OBJ-BG) with scale and transform augmentation.", + "bbox": [ + 88, + 635, + 482, + 787 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Analysis of TAR Strategy. One of the studies examines the impact of the TAR strategy on the model's performance. As shown in Fig. 5 (Left), the accuracy of the model with and without the TAR strategy is plotted against the number of epochs. The model incorporating the TAR strategy (light blue line) demonstrates superior performance compared to the model without TAR (dark blue line). This figure relates", + "bbox": [ + 88, + 794, + 482, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "to the pretraining phase on the ShapeNet [5] dataset, which is subsequently tested on the ModelNet [52] dataset using a SVM.", + "bbox": [ + 511, + 500, + 906, + 541 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Additionally, we evaluate the effect of the TAR strategy by fine-tuning pretrained models (with or without TAR) on four downstream tasks, each corresponding to a distinct dataset. As shown in Fig. 5 (Right), the model pretrained with TAR (light blue line) consistently achieves higher accuracy than the model trained without TAR (dark blue line). This demonstrates that the model pretrained with the TAR strategy has learned more meaningful features, leading to better performance in the downstream tasks.", + "bbox": [ + 511, + 544, + 908, + 681 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion and Limitation", + "text_level": 1, + "bbox": [ + 511, + 694, + 764, + 709 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We introduced three strategies to enhance Mamba networks for point cloud data: isometry-invariant token traversal, recursive patch partitioning for segmentation, and improved learnable token placement, each contributing to a robust feature extraction framework. Our methods demonstrate superior performance over state-of-the-art baselines in classification, segmentation, and few-shot tasks. A limitation of our approach lies in the selection of the neighborhood size $K$ , which is treated as a hyperparameter. While tuning $K$ per dataset helps optimize performance, it may limit generalizability across diverse data distributions and require careful adjustment to maintain robustness.", + "bbox": [ + 511, + 719, + 908, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "11806", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Matan Atzmon, Haggai Maron, and Yaron Lipman. Point convolutional neural networks by extension operators. arXiv preprint arXiv:1803.10091, 2018. 2", + "[2] Ali Bahri, Moslem Yazdanpanah, Mehrdad Noori, Milad Cheraghalikhani, Gustavo Adolfo Vargas Hakim, David Osowiechi, Farzad Beizaee, Ismail Ben Ayed, and Christian Desrosiers. Geomask3d: Geometrically informed mask selection for self-supervised point cloud learning in 3d. arXiv preprint arXiv:2405.12419, 2024. 1, 2", + "[3] Ali Bahri, Moslem Yazdanpanah, Mehrdad Noori, Sahar Dastani, Milad Cheraghalikhani, David Osowiech, Farzad Beizaee, Ismail Ben Ayed, Christian Desrosiers, et al. Test-time adaptation in point clouds: Leveraging sampling variation with weight averaging. arXiv preprint arXiv:2411.01116, 2024. 1", + "[4] Iz Beltagy, Matthew E Peters, and Arman Cohan. Longformer: The long-document transformer. arXiv preprint arXiv:2004.05150, 2020. 1", + "[5] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 6, 8", + "[6] Guangyan Chen, Meiling Wang, Yi Yang, Kai Yu, Li Yuan, and Yufeng Yue. PointGPT: Auto-regressively generative pre-training from point clouds. Advances in Neural Information Processing Systems, 36, 2024. 2, 6, 7", + "[7] Rewon Child, Scott Gray, Alec Radford, and Ilya Sutskever. Generating long sequences with sparse transformers. arXiv preprint arXiv:1904.10509, 2019. 1", + "[8] Fan RK Chung. Spectral graph theory. American Mathematical Soc., 1997. 4", + "[9] Richard Courant and David Hilbert. Methods of mathematical physics: partial differential equations. John Wiley & Sons, 2008. 4", + "[10] Xin Deng, WenYu Zhang, Qing Ding, and XinMing Zhang. Pointvector: a vector representation in point cloud analysis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9455-9465, 2023. 2", + "[11] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 2", + "[12] Runpei Dong, Zekun Qi, Linfeng Zhang, Junbo Zhang, Jianjian Sun, Zheng Ge, Li Yi, and Kaisheng Ma. Autoencoders as cross-modal teachers: Can pretrained 2D image transformers help 3D representation learning? arXiv preprint arXiv:2212.08320, 2022. 2, 6, 7", + "[13] Gene H Golub and Charles F Van Loan. Matrix computations. JHU press, 2013. 5", + "[14] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752, 2023. 1, 2, 4", + "[15] Albert Gu, Tri Dao, Stefano Ermon, Atri Rudra, and Christopher Ré. Hippo: Recurrent memory with optimal polynomial" + ], + "bbox": [ + 93, + 114, + 483, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "projections. Advances in neural information processing systems, 33:1474-1487, 2020. 2", + "[16] Albert Gu, Karan Goel, and Christopher Ré. Efficiently modeling long sequences with structured state spaces. arXiv preprint arXiv:2111.00396, 2021. 1, 2", + "[17] Albert Gu, Karan Goel, Ankit Gupta, and Christopher Ré. On the parameterization and initialization of diagonal state space models. Advances in Neural Information Processing Systems, 35:35971-35983, 2022. 2", + "[18] Yulan Guo, Hanyun Wang, Qingyong Hu, Hao Liu, Li Liu, and Mohammed Bennamoun. Deep learning for 3d point clouds: A survey. IEEE transactions on pattern analysis and machine intelligence, 43(12):4338-4364, 2020. 1", + "[19] Ankit Gupta, Albert Gu, and Jonathan Berant. Diagonal state spaces are as effective as structured state spaces. Advances in Neural Information Processing Systems, 35:22982-22994, 2022. 2", + "[20] Ramin Hasani, Mathias Lechner, Tsun-Hsuan Wang, Makram Chahine, Alexander Amini, and Daniela Rus. Liquid structural state-space models. arXiv preprint arXiv:2209.12951, 2022. 2", + "[21] Cheng-Yao Hong, Yu-Ying Chou, and Tyng-Luh Liu. Attention discriminant sampling for point clouds. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14429-14440, 2023. 7", + "[22] Xin Lai, Jianhui Liu, Li Jiang, Liwei Wang, Hengshuang Zhao, Shu Liu, Xiaojuan Qi, and Jiaya Jia. Stratified transformer for 3d point cloud segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8500-8509, 2022. 2", + "[23] Loic Landrieu and Martin Simonovsky. Large-scale point cloud semantic segmentation with superpoint graphs. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4558-4567, 2018. 2", + "[24] Juho Lee, Yoonho Lee, Jungtaek Kim, Adam Kosierek, Seungjin Choi, and Yee Whye Teh. Set transformer: A framework for attention-based permutation-invariant neural networks. In International conference on machine learning, pages 3744-3753. PMLR, 2019. 1", + "[25] Yangyan Li, Rui Bu, Mingchao Sun, Wei Wu, Xinhan Di, and Baoquan Chen. Pointcnn: Convolution on x-transformed points. Advances in neural information processing systems, 31, 2018. 2", + "[26] Dingkang Liang, Xin Zhou, Xinyu Wang, Xingkui Zhu, Wei Xu, Zhikang Zou, Xiaqing Ye, and Xiang Bai. Pointmamba: A simple state space model for point cloud analysis. arXiv preprint arXiv:2402.10739, 2024. 1, 2, 4, 6, 7, 8", + "[27] Haotian Liu, Mu Cai, and Yong Jae Lee. Masked discrimination for self-supervised learning on point clouds. In European Conference on Computer Vision, pages 657-675. Springer, 2022. 7", + "[28] Yang Liu, Chen Chen, Can Wang, Xulin King, and Mengyuan Liu. Regress before construct: Regress autoencoder for point cloud self-supervised learning. In Proceedings of the 31st ACM International Conference on Multimedia, pages 1738-1749, 2023. 2", + "[29] Yue Liu, Yunjie Tian, Yuzhong Zhao, Hongtian Yu, Lingxi Xie, Yaowei Wang, Qixiang Ye, and Yunfan Liu. Vmamba:" + ], + "bbox": [ + 516, + 93, + 906, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "11807", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Visual state space model. arXiv preprint arXiv:2401.10166, 2024. 1, 2", + "[30] Xu Ma, Can Qin, Haoxuan You, Haoxi Ran, and Yun Fu. Rethinking network design and local geometry in point cloud: A simple residual mlp framework. arXiv preprint arXiv:2202.07123, 2022. 7", + "[31] Andrew Ng, Michael Jordan, and Yair Weiss. On spectral clustering: Analysis and an algorithm. Advances in neural information processing systems, 14, 2001. 4", + "[32] Yatian Pang, Wenxiao Wang, Francis EH Tay, Wei Liu, Yonghong Tian, and Li Yuan. Masked autoencoders for point cloud self-supervised learning. In European conference on computer vision, pages 604-621. Springer, 2022. 1, 2, 5, 6, 7, 8", + "[33] Maciej Pioro, Kamil Ciebiera, Krystian Król, Jan Ludziejewski, and Sebastian Jaszczur. Moe-mamba: Efficient selective state space models with mixture of experts. arXiv preprint arXiv:2401.04081, 2024. 2", + "[34] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 652-660, 2017. 2, 6, 7", + "[35] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems, 30, 2017. 1, 2, 6, 7", + "[36] Charles R Qi, Yin Zhou, Mahyar Najibi, Pei Sun, Khoa Vo, Boyang Deng, and Dragomir Anguelov. Offboard 3d object detection from point cloud sequences. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6134-6144, 2021. 1", + "[37] Guocheng Qian, Yuchen Li, Houwen Peng, Jinjie Mai, Hasan Hammoud, Mohamed Elhoseiny, and Bernard Ghanem. Pointnext: Revisiting pointnet++ with improved training and scaling strategies. Advances in Neural Information Processing Systems, 35:23192-23204, 2022. 7", + "[38] Martin Reuter, Franz-Erich Wolter, and Niklas Peinecke. Laplace-spectra as fingerprints for shape matching. In Proceedings of the 2005 ACM symposium on Solid and physical modeling, pages 101-106, 2005. 4", + "[39] Radu Bogdan Rusu and Steve Cousins. 3d is here: Point cloud library (pcl). In 2011 IEEE international conference on robotics and automation, pages 1-4. IEEE, 2011. 1", + "[40] Jianbo Shi and Jitendra Malik. Normalized cuts and image segmentation. IEEE Transactions on pattern analysis and machine intelligence, 22(8):888-905, 2000. 4, 5", + "[41] Shaoshuai Shi, Xiaogang Wang, and Hongsheng Li. Poincaré: 3d object proposal generation and detection from point cloud. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 770-779, 2019. 1", + "[42] Yuan Tang, Xianzhi Li, Jinfeng Xu, Qiao Yu, Long Hu, Yixue Hao, and Min Chen. Point-lgmask: Local and global contexts embedding for point cloud pre-training with multi-ratio masking. IEEE Transactions on Multimedia, 2023. 2", + "[43] Yi Tay, Dara Bahri, Liu Yang, Donald Metzler, and Da-Cheng Juan. Sparse sinkhorn attention. In International" + ], + "bbox": [ + 91, + 90, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Conference on Machine Learning, pages 9438-9447. PMLR, 2020.1", + "[44] Mikaela Angelina Uy, Quang-Hieu Pham, Binh-Son Hua, Duc Thanh Nguyen, and Sai-Kit Yeung. Revisiting point cloud classification: A new benchmark dataset and classification model on real-world data. In International Conference on Computer Vision (ICCV), 2019. 6, 8", + "[45] Mikaela Angelina Uy, Quang-Hieu Pham, Binh-Son Hua, Thanh Nguyen, and Sai-Kit Yeung. Revisiting point cloud classification: A new benchmark dataset and classification model on real-world data. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1588–1597, 2019. 7", + "[46] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 2", + "[47] Hanchen Wang, Qi Liu, Xiangyu Yue, Joan Lasenby, and Matt J Kusner. Unsupervised point cloud pre-training via occlusion completion. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9782-9792, 2021. 6", + "[48] Yue Wang, Yongbin Sun, Ziwei Liu, Sanjay E Sarma, Michael M Bronstein, and Justin M Solomon. Dynamic graph cnn for learning on point clouds. ACM Transactions on Graphics (tog), 38(5):1-12, 2019. 6, 7", + "[49] Chengzhi Wu, Junwei Zheng, Julius Pfrommer, and Jürgen Beyerer. Attention-based point cloud edge sampling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5333-5343, 2023. 6", + "[50] Xiaoyang Wu, Yixing Lao, Li Jiang, Xihui Liu, and Hengshuang Zhao. Point transformer v2: Grouped vector attention and partition-based pooling. Advances in Neural Information Processing Systems, 35:33330-33342, 2022. 2", + "[51] Xiaoyang Wu, Li Jiang, Peng-Shuai Wang, Zhijian Liu, Xihui Liu, Yu Qiao, Wanli Ouyang, Tong He, and Hengshuang Zhao. Point transformer v3: Simpler, faster, stronger. arXiv preprint arXiv:2312.10035, 2023. 2", + "[52] Zhirong Wu, Shuran Song, Aditya Khosla, Fisher Yu, Linguang Zhang, Xiaou Tang, and Jianxiong Xiao. 3d shapenets: A deep representation for volumetric shapes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1912-1920, 2015. 6, 7, 8", + "[53] Li Yi, Vladimir G Kim, Duygu Ceylan, I-Chao Shen, Mengyan Yan, Hao Su, Cewu Lu, Qixing Huang, Alla Sheffer, and Leonidas Guibas. A scalable active framework for region annotation in 3d shape collections. ACM Transactions on Graphics (ToG), 35(6):1-12, 2016. 6, 7", + "[54] Xumin Yu, Lulu Tang, Yongming Rao, Tiejun Huang, Jie Zhou, and Jiwen Lu. Point-bert: Pre-training 3d point cloud transformers with masked point modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19313-19322, 2022. 1, 2, 6, 7", + "[55] Renrui Zhang, Ziyu Guo, Peng Gao, Rongyao Fang, Bin Zhao, Dong Wang, Yu Qiao, and Hongsheng Li. Point-m2ae: multi-scale masked autoencoders for hierarchical point cloud pre-training. Advances in neural information processing systems, 35:27061-27074, 2022. 1, 2, 5, 6, 7" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "11808", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[56] Renrui Zhang, Liuhui Wang, Yu Qiao, Peng Gao, and Hongsheng Li. Learning 3d representations from 2d pre-trained models via image-to-point masked autoencoders. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21769-21780, 2023. 2, 6, 7", + "[57] Tao Zhang, Xiangtai Li, Haobo Yuan, Shunping Ji, and Shuicheng Yan. Point could mamba: Point cloud learning via state space model. arXiv preprint arXiv:2403.00762, 2024.1, 2, 4, 7", + "[58] Hengshuang Zhao, Li Jiang, Chi-Wing Fu, and Jiaya Jia. Pointweb: Enhancing local neighborhood features for point cloud processing. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5565-5573, 2019. 2", + "[59] Hengshuang Zhao, Li Jiang, Jiaya Jia, Philip HS Torr, and Vladlen Koltun. Point transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 16259-16268, 2021. 2", + "[60] Lianghui Zhu, Bencheng Liao, Qian Zhang, Xinlong Wang, Wenyu Liu, and Xinggang Wang. Vision mamba: Efficient visual representation learning with bidirectional state space model. arXiv preprint arXiv:2401.09417, 2024. 1, 2" + ], + "bbox": [ + 91, + 90, + 480, + 401 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11809", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Spectral Informed Mamba for Robust Point Cloud Processing/9acc5c2f-4c09-4911-b4a0-485ec3edcc53_model.json b/2025/Spectral Informed Mamba for Robust Point Cloud Processing/9acc5c2f-4c09-4911-b4a0-485ec3edcc53_model.json new file mode 100644 index 0000000000000000000000000000000000000000..6fdee7dd8804b1ecf496aabd5d9dd9c3536b5b20 --- /dev/null +++ b/2025/Spectral Informed Mamba for Robust Point Cloud Processing/9acc5c2f-4c09-4911-b4a0-485ec3edcc53_model.json @@ -0,0 +1,2136 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.187, + 0.131, + 0.812, + 0.154 + ], + "angle": 0, + "content": "Spectral Informed Mamba for Robust Point Cloud Processing" + }, + { + "type": "text", + "bbox": [ + 0.168, + 0.181, + 0.831, + 0.234 + ], + "angle": 0, + "content": "Ali Bahri * Moslem Yazdanpanah Mehrdad Noori Sahar Dastani \nMilad Cheraghalikhani Gustavo Adolfo Vargas Hakim David Osowiechi Farzad Beizaee Ismail Ben Ayed Christian Desrosiers" + }, + { + "type": "text", + "bbox": [ + 0.286, + 0.242, + 0.712, + 0.279 + ], + "angle": 0, + "content": "LIVIA, ÉTS Montréal, Canada International Laboratory on Learning Systems (ILLS)" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.313, + 0.327, + 0.329 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.345, + 0.485, + 0.678 + ], + "angle": 0, + "content": "State Space Models (SSMs) have shown significant promise in Natural Language Processing (NLP) and, more recently, computer vision. This paper introduces a new methodology that leverages Mamba and Masked Autoencoder (MAE) networks for point-cloud data in both supervised and self-supervised learning. We propose three key contributions to enhance Mamba's capability in processing complex pointcloud structures. First, we exploit the spectrum of a graph Laplacian to capture patch connectivity, defining an isometry-invariant traversal order that is robust to viewpoints and captures shape manifolds better than traditional 3D grid-based traversals. Second, we adapt segmentation via a recursive patch partitioning strategy informed by Laplacian spectral components, allowing finer integration and segment analysis. Third, we address token placement in MAE for Mamba by restoring tokens to their original positions, which preserves essential order and improves learning. Extensive experiments demonstrate the improvements that our approach brings over state-of-the-art baselines in classification, segmentation, and few-shot tasks. The implementation is available at: https://github.com/AliBahri94/SI-Mamba.git." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.705, + 0.222, + 0.72 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.73, + 0.484, + 0.88 + ], + "angle": 0, + "content": "The analysis of 3D point cloud data is fundamental for various applications, including autonomous driving [36, 41], VR/AR [18], and robotics [39]. Compared to the organized structure of 2D images, point clouds consist of 3D coordinates without direct adjacency information forming an unordered bag. In recent years, considerable efforts have been dedicated to adapt deep learning models such as convolutional neural networks (CNNs) and Transformers to this type of data [2, 3, 32, 35, 54, 55]. Due to their permutation-invariant self-attention mechanism, Transformer networks" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.313, + 0.907, + 0.369 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.379, + 0.907, + 0.449 + ], + "angle": 0, + "content": "Figure 1. (a) Surface-Aware Spectral Traversing (SAST) applied to the patched point clouds of a mesh surface. (b) Traversal from left to right, based on the first to fourth non-constant smallest eigenvectors. (c) Traversal based on the largest eigenvector, forming a non-continuous sequence of tokens." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.461, + 0.907, + 0.597 + ], + "angle": 0, + "content": "are particularly well suited for the unordered nature of point clouds. However, the quadratic complexity of this mechanism, requiring the computation of a weight between each pair of tokens, impedes the application of these networks to large-sized inputs (e.g., 2D images or 3D point clouds represented by many patches). This has prompted researchers to explore more efficient solutions, including the Set Transformer [24], Sparse Transformer [7], Longformer [4] and Sinkhorn Transformer [43]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.6, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Recently, methods based on Structured State Space Sequence (S4) [16] such as Mamba [14] have gained significant traction as a more efficient alternative to Transformers [29, 60]. So far, very few studies have investigated the potential of S4 approaches such as Mamba for 3D point clouds. Existing methods like Point-Mamba [26] and PCM [57] extend the 2D grid-based traversal used for images to a 3D grid. However, this straightforward adaptation to point clouds suffers from three crucial problems. First: whereas patches from 2D images have adjacency information, which could be exploited by the grid-based traversal, the 3D point patches in point clouds offer a sparse representation of the object's surface, and nearby patches on a 3D grid are not necessarily adjacent on this surface. Second: in the absence of self-attention, task-specific performance is highly influenced by the nature of the token traversal strategy. For example, a traversal suitable for point cloud classification may not be effective for a local task such as point-level classification (i.e., segmentation). Third: due to the \"direction-sensitive\" nature of Mamba, the self-supervised MAE pre" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.353, + 0.9 + ], + "angle": 0, + "content": "*Correspondence to ali.bahri.1@ens.etsmtl.ca" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11799" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.151 + ], + "angle": 0, + "content": "training step of leading point cloud models like Point-MAE [32] and Point-M2AE [55] cannot be used directly as there is no attention mechanism to learn the masked tokens' positions." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.154, + 0.483, + 0.183 + ], + "angle": 0, + "content": "The contribution of our work focuses on addressing these problems as follows:" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.185, + 0.483, + 0.29 + ], + "angle": 0, + "content": "1. We introduce a Surface-Aware Spectral Traversing (SAST) strategy based on the Laplacian spectrum of a patch-connectivity graph. Compared to the 3D grid traversal of current approaches like Point-Mamba, our strategy is invariant to isometric transformations (e.g., choice of viewpoint) and better captures the object's surface manifold." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.292, + 0.482, + 0.398 + ], + "angle": 0, + "content": "2. We also present a Hierarchical Local Traversing (HLT) for point-level classification (segmentation) that partitions patches recursively based on their spectral coordinates. Unlike our SAST strategy for classification, which considers Laplacian eigenvectors separately in different traversals, this HLT combines them in a single ordering for a more precise modeling of geometry." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.4, + 0.482, + 0.489 + ], + "angle": 0, + "content": "3. During the MAE-based Self-Supervised Learning (SSL), we propose a Traverse-Aware Repositioning (TAR) strategy to align the masked tokens according to their spectral adjacency. This strategy addresses the critical issue of spatial adjacency preservation unique to Mamba networks." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.185, + 0.483, + 0.489 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.507, + 0.24, + 0.523 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.537, + 0.483, + 0.688 + ], + "angle": 0, + "content": "Deep Point Cloud Learning. Deep neural networks have increasingly been applied to point clouds, with early architectures like PointNet [34] and PointNet++ [35] inspiring further work [1, 10, 23, 25, 58] focused on capturing local context. Transformer-based frameworks [46], including the Point Transformer (v1-v3) [50, 51, 59] and Stratified Transformer [22], have set new benchmarks by integrating both local and global features. Recently, self-supervised pre-training has gained traction by leveraging masked point modeling [28, 42] to train Transformers on unlabeled data." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Building on the effectiveness of MAE in Text and Image domains, Point-BERT [54] presented a revolutionary method inspired by BERT [11], tailoring Transformers to 3D point cloud processing. Point-MAE [32] applied MAE-style pre-training to 3D point clouds using a custom Transformer-based Autoencoder (AE) designed to reconstruct masked irregular patches. The use of multiscale masking and local spatial self-attention mechanisms in Point-M2AE [55] has led to State-Of-The-Art (SOTA) results in 3D representation learning. Furthermore, I2P-MAE [56] improved self-supervised point cloud processing with a masking strategy leveraging pre-trained 2D models through an Image-to-Point transformation. Point-GPT [6] introduced an auto-regressive generative pretraining (GPT)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.213 + ], + "angle": 0, + "content": "approach to address the unordered nature and low information density of point clouds. ACT [12] proposed a cross-modal knowledge transfer method using pretrained 2D or natural language Transformers as teachers for 3D representation learning. Finally, GeoMask3D [2] is a self-supervised point cloud method that uses a teacher-student model to select harder, geometrically intricate regions for masking, thereby improving feature representations." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.222, + 0.699, + 0.238 + ], + "angle": 0, + "content": "2.1. State Space Models" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.246, + 0.905, + 0.518 + ], + "angle": 0, + "content": "State Space Models. State-space models (SSMs) have been foundational in control theory and signal processing for modeling dynamic systems. Recently, they have been adapted to deep learning, with advancements like the Linear State-Space Layer (LSSL), which uses a continuous-time memorization framework based on the HiPPO operator [15] to model long-range dependencies. However, LSSL's high computational demands limit its practicality. To address this, the S4 model [16] introduced parameter normalization techniques, paving the way for structured SSMs with enhancements like complex-diagonal structures [17, 19], support for multiple-input/output [17], and low-rank decomposition [20]. Recent developments include Mamba [14], which achieves linear-time inference and efficient training through selection mechanisms and hardware-aware algorithms. MoE-Mamba further boosts efficiency by integrating a Mixture of Experts (MoE), outperforming both Mamba and Transformer-MoE models [33]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.521, + 0.905, + 0.672 + ], + "angle": 0, + "content": "SSMs for Vision Tasks The above-mentioned works primarily focused on the application of SSMs to long-range or causal data types such as language and speech. In the field of vision, a notable study [29] proposed the VMamba model which features a Cross-Scan Module (CSM) for enhanced 1D selective scanning in 2D spaces and architectural optimizations that significantly improve its performance and speed across various visual tasks. Another significant paper is Vision Mamba [60] which introduces a novel vision backbone called Vim utilizing bidirectional Mamba blocks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.673, + 0.905, + 0.839 + ], + "angle": 0, + "content": "For point cloud analysis based on Mamba, two key works are Point-Mamba [26] and PCM [57]. PointMamba introduces a simple approach to token reordering for point cloud analysis by strategically organizing point tokens based on a 3D grid. Similarly, PCM enhances Mamba with a Consistent Traverse Serialization (CTS) technique that converts 3D point clouds into 1D point sequences while maintaining spatial adjacency. Building upon these approaches, our method introduces the Spectral Spatial Traversing (SST) strategy, which improves token ordering and maintains spatial adjacency within Mamba networks." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.852, + 0.604, + 0.867 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.905, + 0.901 + ], + "angle": 0, + "content": "We begin by outlining the fundamental concepts of SSMs and spectral graph analysis which are at the core of our" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11800" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.171, + 0.089, + 0.837, + 0.52 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.532, + 0.908, + 0.618 + ], + "angle": 0, + "content": "Figure 2. Overview of the proposed Spectral Spatial Traversing (SST) method. (a) Point cloud, (b) Patchification, (c) Forming the adjacency graph, (d) Traversal based on SAST using \\( s \\) non-constant smallest eigenvectors, (e) HLT for segmentation tasks, (f) TAR strategy for Masked Autoencoders. The process includes reverse and concatenation operations, with learnable tokens, representations, and masked tokens highlighted. (g) The classification task involves sorting tokens by different eigenvectors, concatenating them, and then feeding them into the network. (h) The segmentation task where HLT is applied on the tokens \\( (\\vec{q}) \\) and \\( \\vec{q} \\) is fed into the network. (i) A flowchart visualizing the techniques used in self-supervised learning and various downstream tasks." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.637, + 0.486, + 0.834 + ], + "angle": 0, + "content": "work. We then give an in-depth presentation of our Surface-Aware Spectral Traversing (SAST) strategy for point cloud processing that improves the model's robustness to isometric transformations and better captures the underlying manifold of the point cloud. Thereafter, we provide detailed specifications of our Hierarchical Local Traversing (HLT) strategy for point-level classification, which defines a more structured patch traversal order based on the recursive partitioning of spectral information. Finally, we introduce our Traverse-Aware Repositioning (TAR), which improves the handling of learnable tokens in masked autoencoders within Mamba networks. Fig. 2 illustrates the overview of the proposed Spectral Spatial Traversing (SST) method." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.845, + 0.233, + 0.86 + ], + "angle": 0, + "content": "3.1. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.485, + 0.903 + ], + "angle": 0, + "content": "State Space Models (SSMs) use a series of first-order differential equations to describe how the state of the linear," + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.637, + 0.783, + 0.652 + ], + "angle": 0, + "content": "time-invariant system evolves over time:" + }, + { + "type": "equation", + "bbox": [ + 0.529, + 0.662, + 0.907, + 0.683 + ], + "angle": 0, + "content": "\\[\n\\dot {h} (t) = A h (t) + B x (t), \\quad y (t) = C \\dot {h} (t) + D x (t), \\quad (1)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.692, + 0.907, + 0.737 + ], + "angle": 0, + "content": "Here, \\(\\dot{h}(t)\\) denotes the time derivative of the state vector \\(h(t)\\). The matrices \\(A\\), \\(B\\), \\(C\\), and \\(D\\) are the weighting parameters." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.739, + 0.907, + 0.8 + ], + "angle": 0, + "content": "Due to their reliance on continuous data streams \\( x(t) \\), SSMs are not natively equipped to handle discrete inputs represented as \\( \\{x_0, x_1, \\ldots\\} \\). This necessitates the use of a discretized SSM version for practical applications:" + }, + { + "type": "equation", + "bbox": [ + 0.565, + 0.811, + 0.907, + 0.829 + ], + "angle": 0, + "content": "\\[\nh _ {k} = \\bar {A} h _ {k - 1} + \\bar {B} x _ {k}, \\quad y _ {k} = \\bar {C} h _ {k} + \\bar {D} x _ {k}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.841, + 0.909, + 0.903 + ], + "angle": 0, + "content": "The state space model in its discrete version utilizes a recursive function to link each state \\( h_k \\) to its preceding state, encapsulated by the matrices \\( \\bar{A} \\in \\mathbb{R}^{N \\times N} \\), \\( \\bar{B} \\in \\mathbb{R}^{N \\times 1} \\), and \\( \\bar{C} \\in \\mathbb{R}^{N \\times 1} \\), which are tuned parameter matrices. While" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "11801" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.089, + 0.48, + 0.237 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.248, + 0.485, + 0.318 + ], + "angle": 0, + "content": "Figure 3. (a) Visualization of the four non-constant smallest Laplacian eigenvectors \\((v^{(k)}, k = 1, \\ldots, 4)\\) and (b) the discrete partitioning \\((q)\\) of our HLT strategy combining the information of all four eigenvectors. Note: we assumed that patches contain a single point for better visualization." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.324, + 0.484, + 0.432 + ], + "angle": 0, + "content": "matrix \\(\\bar{D} \\in \\mathbb{R}^{N \\times 1}\\) may be employed as a residual connection, we follow previous work and exclude it from our model. The transition from a continuous signal representation \\(x(t)\\) to a discrete sequence involves sampling \\(x(t)\\) at intervals defined by \\(\\Delta\\), setting each discrete input as \\(x_{k} = x(k\\Delta)\\). This adjustment to a discrete framework results in revised matrix definitions:" + }, + { + "type": "equation", + "bbox": [ + 0.091, + 0.437, + 0.484, + 0.478 + ], + "angle": 0, + "content": "\\[\n\\bar {A} = \\left(I - \\frac {\\Delta}{2} A\\right) ^ {- 1} \\left(I + \\frac {\\Delta}{2} A\\right), \\bar {B} = \\left(I - \\frac {\\Delta}{2} A\\right) ^ {- 1} \\Delta B, \\bar {C} = C. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.48, + 0.484, + 0.6 + ], + "angle": 0, + "content": "However, the fixed dynamics of Linear Time-Invariant (LTI) models, exemplified by the constant parameters \\(A\\), \\(B\\), and \\(C\\) in Eq. (3), restrict their capacity to selectively retain or discard relevant information, thereby limiting their contextual awareness. To improve content-aware reasoning, we use Mamba's selection mechanism that manages the propagation and interaction of information across the sequence dimension [14]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.607, + 0.484, + 0.713 + ], + "angle": 0, + "content": "Spectral Graph Analysis. Popularized by Chung in the 90s [8], spectral graph analysis characterizes the properties of a graph \\( G = (V, E) \\) by the spectrum (eigenvalues and corresponding eigenvectors) of its Laplacian matrix \\( L \\). This analysis can be understood as a discretized version of the Laplace-Beltrami Operator \\( \\Delta \\) of a function \\( f \\) defined on a Riemannian manifold:" + }, + { + "type": "equation", + "bbox": [ + 0.223, + 0.723, + 0.483, + 0.74 + ], + "angle": 0, + "content": "\\[\n\\Delta f = \\operatorname {d i v} (\\operatorname {g r a d} f) \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.825 + ], + "angle": 0, + "content": "where \\(\\operatorname{grad}f\\) is the gradient of \\(f\\) and \\(\\operatorname{div}\\) the divergence on the manifold. The solution to the Laplacian eigenvalue problem \\(\\Delta f = -\\lambda f\\), known as Helmholtz wave equation, is an eigenfunction corresponding to the natural vibration form of a homogeneous membrane with eigenvalue \\(\\lambda\\) [38]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Following methods for spectral clustering [31] and normalized cuts [40], we consider a weighted adjacency matrix \\( W: V \\times V \\to \\mathbb{R}_+ \\) where \\( W_{ij} = 0 \\) if \\( (i,j) \\notin E \\) to model the Euclidean distance of nearby patches (see Section 3.3). The Laplacian matrix of \\( G \\) is defined as \\( L = D - W \\) where" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.182 + ], + "angle": 0, + "content": "\\(D\\) is the diagonal degree matrix such that \\(D_{ii} = \\sum_{j}W_{ij}\\). To account for variability in the scale of weights \\(W_{ij}\\) or the distribution of node degrees \\(D_{ii}\\), it is preferable to employ a normalized version of the Laplacian. In this work, we use the Random Walk Laplacian \\(L_{rw} = I - D^{-1}W\\) which has the following useful properties:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.183, + 0.905, + 0.214 + ], + "angle": 0, + "content": "1. \\( L_{rw} \\) is positive semi-definite and has \\( |V| \\) non-negative real-valued eigenvalues \\( 0 = \\lambda_1 \\leq \\ldots \\leq \\lambda_{|V|} \\);" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.214, + 0.905, + 0.258 + ], + "angle": 0, + "content": "2. 0 is an eigenvalue of \\( L_{rw} \\) with the constant vector as eigenvector and its multiplicity equals the number of connected components in the graph;" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.259, + 0.904, + 0.288 + ], + "angle": 0, + "content": "3. Following Courant's Nodal Line Theorem [9], the \\(n\\)-th eigenmode of \\(L_{rw}\\) has at most \\(n\\) poles of vibration;" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.289, + 0.905, + 0.333 + ], + "angle": 0, + "content": "4. The representation of a shape by the spectrum of \\( L_{rw} \\) is invariant to isometry (i.e., distance-preserving transformation)." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.183, + 0.905, + 0.333 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.335, + 0.907, + 0.547 + ], + "angle": 0, + "content": "Our method uses the \\(s\\) first non-constant eigenvectors of \\(L_{rw}\\) (i.e., the eigenvectors corresponding to the \\(s\\) smallest non-zero eigenvalues) to define traversal orders for classification (Section 3.3) and segmentation (Section 3.4) that are robust to the viewpoint (due to isometry invariance) and provide a smooth parametrization of the surface manifold. We consider the first eigenvectors as they encode low frequency information (by Courant's Nodal Line Theorem), making the resulting traversal more robust to shape variability and noise. Fig. 1 illustrates this concept: (a) shows the original mesh, (b) shows traversals based on the first to fourth non-constant smallest eigenvectors, and (c) shows traversal based on the largest eigenvector forming a non-continuous sequence of tokens." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.558, + 0.753, + 0.572 + ], + "angle": 0, + "content": "3.2. Point Cloud Patchification" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.578, + 0.907, + 0.747 + ], + "angle": 0, + "content": "Given a point cloud \\(\\mathcal{P} = \\{p_i\\}_{i=1}^{N_p}\\), each point represented by 3D coordinates, we convert \\(\\mathcal{P}\\) to a reduced set of patches that can be processed more efficiently. Toward this goal, we employ the Farthest Point Sampling (FPS) algorithm to select a subset \\(\\mathcal{C} \\subset \\mathcal{P}\\) of \\(N_c\\) points offering a good coverage of the entire point cloud. These selected points will act as the centers of local patches within the point cloud. For each center point \\(p_{s_i} \\in \\mathcal{C}\\), we then identify \\(N_n\\) nearest points \\(\\mathcal{N}(p_{s_i}) \\subset \\mathcal{P}\\) using the K-Nearest Neighbours (KNN) algorithm. Following this, each patch is defined as a center \\(p_{s_i}\\) and its corresponding nearest-neighbors \\(\\mathcal{N}(p_{s_i})\\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.757, + 0.881, + 0.774 + ], + "angle": 0, + "content": "3.3. Surface-Aware Spectral Traversing (SAST)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Current point cloud processing approaches using Mamba, such as Point-Mamba [26] and PCM [57], simply extend the 2D grid-based traversal for images to a 3D grid. As mentioned before, this naive strategy suffers from two issues: 1) the 3D grid is view dependent, thus rotating the point cloud or moving the camera yields a different traversal order; 2) unrelated patches may be adjacent in 3D space, hence can be traversed subsequently. To address these problems, we" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11802" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.122 + ], + "angle": 0, + "content": "define a traversal order based on the Laplacian spectrum of the patch-connectivity graph." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.123, + 0.484, + 0.244 + ], + "angle": 0, + "content": "In this graph, each node corresponds to a patch and the weighted adjacency matrix \\( W \\) is defined using the Euclidean distance between patch centers. For patches \\( i \\) and \\( j \\) defined by center points \\( p_{s_i} \\) and \\( p_{s_j} \\), we add an edge \\( (i,j) \\) if \\( p_{s_j} \\) is among the \\( K \\) nearest neighbors of \\( p_{s_i} \\) or vice-versa. The weight of this edge is computed using a Gaussian kernel: \\( W_{ij} = \\exp \\left(-\\| p_{s_i} - p_{s_j}\\| _2^2 /\\sigma\\right) \\) where \\( \\sigma \\) is a hyperparameter controlling the kernel width." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.245, + 0.484, + 0.465 + ], + "angle": 0, + "content": "Following Section 3.1, we compute the \\(s\\) first nonconstant eigenvectors of the Random Walk Laplacian \\(L_{rw}\\). This can be achieved efficiently using an iterative method like the Arnoldi algorithm [13] by exploiting the following facts: 1) matrix \\(W\\) is very sparse, and 2) only the first few eigenvectors need to be computed. Eigenvector \\(v^{(k)} \\in \\mathbb{R}^{N_c}\\), \\(k \\in \\{1, \\dots, s\\}\\), assigns an eigenfunction value \\(v_i^{(k)}\\) to each patch \\(i\\). In each Mamba block of our model, we perform two separate traversals of tokens (each token corresponds to an input patch) for every eigenvector: a forward traversal by increasing value of \\(v_i^{(k)}\\) and a reserve traversal by decreasing value of \\(v_i^{(k)}\\). At the end of the block, we concatenate the tokens from the \\(s \\times 2\\) traversals. The visual effect of each traversal is illustrated in Fig. 1(b) and Fig. 3(a)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.467, + 0.484, + 0.73 + ], + "angle": 0, + "content": "Canonicalization of spectrum. Although the spectrum of \\( L_{rw} \\) forms an isometry-invariant representation of the surface manifold, this representation may be impacted by two sources of ambiguity: 1) the sign of eigenvectors is undetermined (i.e., if \\( v^{(k)} \\) is an eigenvector, then so is \\( -v^{(k)} \\)), and 2) the order of eigenvectors with similar eigenvalues may vary. We address these two sources of ambiguity with the following canonicalization procedure. For the first one, we flip the sign of an eigenvector \\( v^{(k)} \\) (i.e., \\( v^{(k)} \\coloneqq -v^{(k)} \\)) if its first element is negative (i.e., \\( v_{1}^{(k)} < 0 \\)). To handle the second ambiguity, we first sort the eigenvectors by nondecreasing eigenvalue. We deal with eigenvalues having a multiplicity greater than one by finding pairs of consecutive eigenvectors \\( v^{(k)} \\), \\( v^{(k+1)} \\) with near-identical eigenvalues (i.e., \\( |\\lambda_k - \\lambda_{k+1}| \\leq \\epsilon \\)). For such pairs, we flip the order if \\( v_{1}^{(k)} > v_{1}^{(k+1)} \\). This reordering process is repeated until no further change occurs." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.741, + 0.483, + 0.772 + ], + "angle": 0, + "content": "3.4. Hierarchical Local Traversing (HLT) for Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.779, + 0.483, + 0.87 + ], + "angle": 0, + "content": "While effective for classification tasks, the SAST strategy considering each eigenvector in a separate traversal may not capture the precise relationship between patches needed for segmentation. To address this issue, we introduce a Hierarchical Local Traversal (HLT) strategy that considers the full spectrum (all \\(s\\) non-constant eigenvectors) simultaneously." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Our strategy is inspired by the recursive binary partitioning technique of normalized cuts [40]. Starting from" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.37 + ], + "angle": 0, + "content": "the canonicalized spectrum (see Section 3.3), tokens are first split based on the first eigenvector \\( v^{(1)} \\), by comparing their corresponding value in \\( v^{(1)} \\) with the mean value \\( \\overline{v}^{(1)} \\). This yields a binary partition of tokens \\( b_{i}^{(1)} = \\mathbb{1}\\big(v_{i}^{(1)}\\geq \\overline{v}^{(1)}\\big)\\in \\{0,1\\} \\) where \\( \\mathbb{1} \\) is the indicator function. Each subgroup is then divided based on the mean of the second eigenvector \\( v^{(2)} \\), and so on for other eigenvectors. This partitioning process can be seen as building a binary tree, where each level corresponds to a different eigenvector and leaf nodes \\( i \\) are uniquely identified by the sequence of bits \\( b_{i} = [b_{i}^{(1)},\\dots ,b_{i}^{(s)}] \\) on the path from the root to the leaf. Our HLT method traverses groups of leaf nodes (groups of tokens) sequentially based on the lexicographic order of their binary code (e.g., [0000], [0001], [0010], [0011], ... in the case of four eigenvectors). For convenience, we convert binary codes \\( b_{i} \\) to a non-negative integer \\( q_{i} \\) (e.g., bin2Int([0011]) = 3) and define two traversal orders, by increasing or decreasing values of \\( q_{i} \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.37, + 0.905, + 0.536 + ], + "angle": 0, + "content": "For \\(s\\) eigenvectors, the HLT strategy described above divides tokens into \\(2^{s}\\) segments which are traversed sequentially. In the best case scenario, \\(\\lceil \\log_2(N_c)\\rceil\\) eigenvectors are thus needed to split tokens into individual segments. However, it may happen that multiple tokens fall in the same segment, especially when using fewer eigenvectors. In such case, one can further sort tokens within each segment, for example, using the values of the first eigenvector (i.e., \\(v^{(1)}\\)). In our implementation, we simply sort these tokens randomly to add stochasticity in the training. This section is illustrated in Fig. 2 (e), and Fig. 3 (b)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.537, + 0.906, + 0.656 + ], + "angle": 0, + "content": "As shown in Fig. 3, the first Laplacian eigenvectors encode high-level spatial relations (e.g., bottom vs. top, left vs. right, torso vs. limbs, etc.). In the SAST, because these eigenvectors are used in separate traversals, the network may not be able to differentiate specific regions/parts of the point cloud. In contrast, our HLT strategy can capture such specific parts (e.g., head, left or right arm/thigh/alf, pelvis, etc.)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.668, + 0.905, + 0.698 + ], + "angle": 0, + "content": "3.5. Traverse-Aware Repositioning (TAR) for Masked Autoencoders" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Following state-of-art Transformers for point cloud processing, such as Point-MAE [32] and Point-M2AE [55], our method leverages self-supervised pretraining based on MAE to boost performance. In Transformer-based approaches, the learnable tokens of masked patches can be inserted in any position of the sequence (typically at the end) as the self-attention mechanism can still attend to all tokens irrespective of their positions. However, this approach presents a significant problem in Mamba networks which are sensitive to the traversal order of tokens. We handle this problem via a TAR strategy that improves the placement of learnable tokens in MAE within Mamba networks. Specifically, we restore the learnable tokens to their" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.956 + ], + "angle": 0, + "content": "11803" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.089, + 0.092, + 0.482, + 0.151 + ], + "angle": 0, + "content": "original positions rather than appending them at the end of the sequence. This ensures that the essential order of tokens is maintained, preserving spatial adjacency and enhancing learning effectiveness within Mamba networks." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.152, + 0.483, + 0.346 + ], + "angle": 0, + "content": "The proposed TAR strategy selects an arbitrary traversal order and randomly masks a subset of \\( N_{m} \\) tokens with the same masking ratio as the transformer-based MAEs. These tokens are then removed from the sequence, and their positions are recorded. Afterwards, the remaining (visible) tokens are fed to the encoder that outputs their representation. Before reconstructing the point cloud using the decoder, we reinsert the learnable tokens in the sequence at their recorded position. The set of masked patches is used for other traversal orders. This procedure can be seen in Fig. 2 (f). Following previous work, we measure the reconstruction error for masked patches using the Chamfer distance:" + }, + { + "type": "equation", + "bbox": [ + 0.167, + 0.352, + 0.482, + 0.393 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {r e c} = \\frac {1}{N _ {m}} \\sum_ {i = 1} ^ {N _ {m}} \\operatorname {C h a m f e r} \\left(\\mathcal {S} _ {i}, \\hat {\\mathcal {S}} _ {i}\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.4, + 0.483, + 0.46 + ], + "angle": 0, + "content": "where \\(S_{i}\\in \\mathbb{R}^{N_{n}\\times 3}\\) is the set of points forming the \\(i\\)-th masked patch and \\(\\hat{S}_i\\) the reconstructed output for these points. The Chamfer distance between two sets of points \\(S\\) and \\(\\hat{S}\\) is defined as" + }, + { + "type": "equation", + "bbox": [ + 0.138, + 0.465, + 0.482, + 0.535 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\operatorname {C h a m f e r} \\left(\\mathcal {S}, \\mathcal {S} ^ {\\prime}\\right) = \\sum_ {p \\in \\mathcal {S}} \\min _ {p ^ {\\prime} \\in \\mathcal {S} ^ {\\prime}} \\| p - p ^ {\\prime} \\| _ {2} ^ {2} \\\\ + \\sum_ {p ^ {\\prime} \\in \\mathcal {S} ^ {\\prime}} \\min _ {p \\in \\mathcal {S}} \\| p - p ^ {\\prime} \\| _ {2} ^ {2}. \\tag {6} \\\\ \\end{array}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.551, + 0.223, + 0.568 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.576, + 0.483, + 0.607 + ], + "angle": 0, + "content": "Several experiments are conducted to evaluate the proposed method. First, we pretrain the Point-Mamba network us" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.625, + 0.483, + 0.681 + ], + "angle": 0, + "content": "Table 1. Few-shot classification on ModelNet40. We report the average accuracy (\\%) and standard deviation (\\%) of 10 independent experiments. \\* denotes reproduced results. A \\(\\equiv\\) represents the average (A) and standard deviation (std), respectively." + }, + { + "type": "table", + "bbox": [ + 0.117, + 0.692, + 0.454, + 0.899 + ], + "angle": 0, + "content": "
Method5-way10-way
10-shot20-shot10-shot20-shot
DGCNN [48]91.83793.43286.3390.937
DGCNN + OcCo [47]91.93393.93186.4391.346
Transformer [54]87.83293.3384.6389.43
Transf. + OcCo [54]94.03695.9389.4392.446
Point-BERT [54]94.63196.3391.0392.73
Point-M2AE [55]96.83898.3492.3395.039
Point-MAE [32]96.33597.8392.6395.039
PointMamba* [26]95.93197.3391.6394.535
Ours96.43598.5392.0395.13
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.258 + ], + "angle": 0, + "content": "ing our techniques on the ShapeNet [5] training dataset. We then assess the performance of these pretrained models across a variety of standard benchmarks, including object classification, few-shot learning, and segmentation. Additionally, we train the model from scratch on downstream datasets to demonstrate the robustness and versatility of our method. To have a fair comparison, we adopt the masking ratio (60%) that was used in the Point-Mamba model. Moreover, a comprehensive analysis of the computational efficiency, runtime, and memory usage of our SAST approach is provided in the Supplementary Material." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.268, + 0.688, + 0.285 + ], + "angle": 0, + "content": "4.1. Pretraining Setup" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.291, + 0.905, + 0.412 + ], + "angle": 0, + "content": "Following Point-Mamba, we adopt the ShapeNet [5] dataset for the pretraining and assess the quality of the 3D representations produced by our approach through a linear evaluation on the ModelNet40 [52] dataset. The linear evaluation is performed by a Support Vector Machine (SVM) fitted on these features. This classification performance is quantified by the accuracy metric. Augmentation in pretraining is scale and transform, while in finetuning it is rotation." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.423, + 0.695, + 0.437 + ], + "angle": 0, + "content": "4.2. Downstream Tasks" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.448, + 0.905, + 0.525 + ], + "angle": 0, + "content": "Object Classification on Real-World Dataset. To evaluate our method for point clouds, we test it on the ScanObjectNN dataset [44], as described in previous studies. The augmentation used during training is random rotation. The results, presented in Tab. 3, show that our strategy signifi" + }, + { + "type": "table_caption", + "bbox": [ + 0.512, + 0.548, + 0.905, + 0.59 + ], + "angle": 0, + "content": "Table 2. Part segmentation on the ShapeNetPart [53]. The mIoU for all instances (Inst.) is reported. HLT is Hierarchical Local Traversing for segmentation." + }, + { + "type": "table", + "bbox": [ + 0.54, + 0.602, + 0.876, + 0.907 + ], + "angle": 0, + "content": "
MethodsBackboneFLOPs (G)mIoU
Training from scratch
PointNet [34]--83.7
PointNet++ [35]--85.1
DGCNN [48]--85.2
APES [49]--85.8
Point-MAE [32]Transformer15.585.7
PointMamba [26]Mamba14.385.8
Ours (HLT)Mamba14.385.9
Training from pretrained
Transformer [54]Transformer15.585.1
Point-BERT [54]Transformer15.585.6
Point-MAE [32]Transformer15.586.1
Point-M2AE [55]Transformer15.586.5
Point-GPT-S [6]Transformer-86.2
ACT [12]Transformer-86.2
I2P-MAE [56]Transformer-86.8
PointMamba [26]Mamba14.386.0
Ours (SAST)Mamba14.385.7
Ours (HLT)Mamba14.386.1
" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "11804" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.09, + 0.908, + 0.119 + ], + "angle": 0, + "content": "Table 3. Object classification on ScanObjectNN [45]. Accuracy (%) is reported. \\( {}^{ \\dagger } \\) indicates that this method was fine-tuned without rotation augmentation." + }, + { + "type": "table", + "bbox": [ + 0.144, + 0.13, + 0.855, + 0.463 + ], + "angle": 0, + "content": "
MethodsBackboneParam. (M)FLOPs (G)OBJ-BGOBJ-ONLYPB-T50-RSModelNet40
Training from scratch
PointNet [34]-3.50.573.379.268.089.2
PointNet++ [35]-1.51.782.384.377.990.7
DGCNN [48]-1.82.482.886.278.192.9
PointNeXt [37]-1.41.6--87.792.9
PointMLP [30]-13.231.4--85.4-
ADS [21]-----87.5-
Transformer [54]Transformer22.14.879.8680.5577.24-
Point-MAE [32]Transformer22.14.886.7586.9280.7892.3
PointMamba [26]Mamba12.33.690.8790.1885.6092.4
OursMamba12.33.692.2591.3987.3092.7
Training from pretrained
Transformer [54]Transformer22.14.879.8680.5577.2492.1
Transformer-OcCo [54]Transformer--84.8585.5478.79-
Point-BERT [54]Transformer22.14.887.4388.1283.0792.7
Point-M2AE† [55]Transformer--91.2288.8186.4394.0
Point-MAE [32]Transformer22.14.892.7791.2289.0493.2
PointMamba [26]Mamba12.33.693.2991.5688.1792.8
PCM [57]Mamba12.33.6--86.9-
OursMamba12.33.694.3291.9189.1093.4
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.487, + 0.482, + 0.577 + ], + "angle": 0, + "content": "cantly improves object classification accuracy in both training from scratch and fine-tuning scenarios. These findings highlight the effectiveness of our approach in enhancing the model's ability to identify and classify objects across various backgrounds, demonstrating its robustness in complex real-world scenarios." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.587, + 0.483, + 0.723 + ], + "angle": 0, + "content": "Object Classification on Clean Objects Dataset. We also evaluated our method on the ModelNet40 [52] dataset, following the protocols established in previous works. The augmentation used during training is scale and transform. As shown in Tab. 3, our approach achieves notable enhancements on this challenging dataset compared to both the original Point-Mamba and the Transformer-based Point-MAE. This demonstrates the robustness and effectiveness of our method when applied to the Point-Mamba network." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.732, + 0.483, + 0.882 + ], + "angle": 0, + "content": "Few-shot Learning. We conducted few-shot learning experiments on ModelNet40 [52] dataset, adhering to the protocols of previous studies [27, 32, 55]. The results of our few-shot learning experiments are presented in Tab. 1. Despite the competitive nature of this benchmark, our method demonstrated outstanding performance across all tested scenarios. As shown in Tab. 1, our Mamba-based method achieves results comparable to or exceeding those of transformer-based methods (Point-MAE and PointM2AE)." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.886, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Part Segmentation. We evaluated our method's repre" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.487, + 0.907, + 0.698 + ], + "angle": 0, + "content": "sensation learning ability on the ShapeNetPart dataset [53] using standard experimental settings [34, 35, 54]. Tab. 2 shows that performance gains in prior methods are minimal, highlighting the dataset's challenge. For instance, PointGPT [6] and ACT [12], despite being state-of-the-art and complex methods, show only marginal improvements over each other and other state-of-the-art approaches. Similarly, I2P-MAE [56], which combines 3D data with 2D information, shows limited improvement over Point-M2AE. For the \"Training from scratch\" setting, our method outperforms several state-of-the-art approaches. In the \"Training from pretrained\" setting, we further demonstrate the effectiveness of HLT strategy compared to SAST in the segmentation task." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.706, + 0.68, + 0.721 + ], + "angle": 0, + "content": "4.3. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.728, + 0.906, + 0.805 + ], + "angle": 0, + "content": "In this section, we aim to investigate the effects of different parameters on our method. We will focus on two key aspects: the effect of the number of non-constant smallest eigenvectors and the adjacency matrix used in the SAST strategy, and the analysis of the TAR strategy." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Analysis of Eigenvectors and Graph. One of our ablation studies investigates the impact of the number of nonconstant smallest eigenvectors used in the SAST and TAR strategies. As depicted in Fig. 4 (Left), the best performance is achieved when using the four non-constant smallest eigenvectors (blue line) for traversing. When the num" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "11805" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.135, + 0.091, + 0.466, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.539, + 0.093, + 0.862, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.261, + 0.908, + 0.292 + ], + "angle": 0, + "content": "Figure 4. Analysis of the number of non-constant smallest eigenvectors and comparison with previous methods (Left) and analysis of the number of nearest neighbors \\( K \\) (Right)." + }, + { + "type": "image", + "bbox": [ + 0.134, + 0.304, + 0.462, + 0.462 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.539, + 0.309, + 0.862, + 0.457 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.212, + 0.473, + 0.786, + 0.49 + ], + "angle": 0, + "content": "Figure 5. The effect of the TAR strategy in the pretraining phase (Left) and in finetuning (Right)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.5, + 0.483, + 0.635 + ], + "angle": 0, + "content": "ber of non-constant smallest eigenvectors increases beyond four, the performance drops. This is because the additional eigenvectors are closer to the largest eigenvectors, which are less smooth and do not capture the most significant structural variations effectively. The significantly lower performance of Point-MAE [32], Point-Mamba [26], and PointMamba with random traversing (Random-Mamba) compared to ours underscores the importance of appropriate traversing for Mamba networks." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.636, + 0.483, + 0.789 + ], + "angle": 0, + "content": "Another study examines the impact of the number of nearest neighbors used in creating the adjacency matrix. As shown in Fig. 4 (Right), the best performance (blue line) is achieved with 20 nearest neighbors. The model's accuracy increases as the number of nearest neighbors increases from 10 to 20, reaching its peak at 20 nearest neighbors. Beyond this point, the performance starts to decline. All these experiments are trained and evaluated from scratch on the ScanObjectNN dataset [44] (OBJ-BG) with scale and transform augmentation." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.795, + 0.483, + 0.903 + ], + "angle": 0, + "content": "Analysis of TAR Strategy. One of the studies examines the impact of the TAR strategy on the model's performance. As shown in Fig. 5 (Left), the accuracy of the model with and without the TAR strategy is plotted against the number of epochs. The model incorporating the TAR strategy (light blue line) demonstrates superior performance compared to the model without TAR (dark blue line). This figure relates" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.5, + 0.907, + 0.542 + ], + "angle": 0, + "content": "to the pretraining phase on the ShapeNet [5] dataset, which is subsequently tested on the ModelNet [52] dataset using a SVM." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.545, + 0.909, + 0.682 + ], + "angle": 0, + "content": "Additionally, we evaluate the effect of the TAR strategy by fine-tuning pretrained models (with or without TAR) on four downstream tasks, each corresponding to a distinct dataset. As shown in Fig. 5 (Right), the model pretrained with TAR (light blue line) consistently achieves higher accuracy than the model trained without TAR (dark blue line). This demonstrates that the model pretrained with the TAR strategy has learned more meaningful features, leading to better performance in the downstream tasks." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.695, + 0.766, + 0.71 + ], + "angle": 0, + "content": "5. Conclusion and Limitation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.909, + 0.901 + ], + "angle": 0, + "content": "We introduced three strategies to enhance Mamba networks for point cloud data: isometry-invariant token traversal, recursive patch partitioning for segmentation, and improved learnable token placement, each contributing to a robust feature extraction framework. Our methods demonstrate superior performance over state-of-the-art baselines in classification, segmentation, and few-shot tasks. A limitation of our approach lies in the selection of the neighborhood size \\( K \\), which is treated as a hyperparameter. While tuning \\( K \\) per dataset helps optimize performance, it may limit generalizability across diverse data distributions and require careful adjustment to maintain robustness." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "11806" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.157 + ], + "angle": 0, + "content": "[1] Matan Atzmon, Haggai Maron, and Yaron Lipman. Point convolutional neural networks by extension operators. arXiv preprint arXiv:1803.10091, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.159, + 0.484, + 0.242 + ], + "angle": 0, + "content": "[2] Ali Bahri, Moslem Yazdanpanah, Mehrdad Noori, Milad Cheraghalikhani, Gustavo Adolfo Vargas Hakim, David Osowiechi, Farzad Beizaee, Ismail Ben Ayed, and Christian Desrosiers. Geomask3d: Geometrically informed mask selection for self-supervised point cloud learning in 3d. arXiv preprint arXiv:2405.12419, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.244, + 0.484, + 0.325 + ], + "angle": 0, + "content": "[3] Ali Bahri, Moslem Yazdanpanah, Mehrdad Noori, Sahar Dastani, Milad Cheraghalikhani, David Osowiech, Farzad Beizaee, Ismail Ben Ayed, Christian Desrosiers, et al. Test-time adaptation in point clouds: Leveraging sampling variation with weight averaging. arXiv preprint arXiv:2411.01116, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.328, + 0.483, + 0.369 + ], + "angle": 0, + "content": "[4] Iz Beltagy, Matthew E Peters, and Arman Cohan. Longformer: The long-document transformer. arXiv preprint arXiv:2004.05150, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.372, + 0.483, + 0.44 + ], + "angle": 0, + "content": "[5] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.443, + 0.483, + 0.497 + ], + "angle": 0, + "content": "[6] Guangyan Chen, Meiling Wang, Yi Yang, Kai Yu, Li Yuan, and Yufeng Yue. PointGPT: Auto-regressively generative pre-training from point clouds. Advances in Neural Information Processing Systems, 36, 2024. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.5, + 0.483, + 0.541 + ], + "angle": 0, + "content": "[7] Rewon Child, Scott Gray, Alec Radford, and Ilya Sutskever. Generating long sequences with sparse transformers. arXiv preprint arXiv:1904.10509, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.543, + 0.483, + 0.571 + ], + "angle": 0, + "content": "[8] Fan RK Chung. Spectral graph theory. American Mathematical Soc., 1997. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.572, + 0.483, + 0.613 + ], + "angle": 0, + "content": "[9] Richard Courant and David Hilbert. Methods of mathematical physics: partial differential equations. John Wiley & Sons, 2008. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.615, + 0.483, + 0.671 + ], + "angle": 0, + "content": "[10] Xin Deng, WenYu Zhang, Qing Ding, and XinMing Zhang. Pointvector: a vector representation in point cloud analysis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9455-9465, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.673, + 0.483, + 0.727 + ], + "angle": 0, + "content": "[11] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.73, + 0.483, + 0.798 + ], + "angle": 0, + "content": "[12] Runpei Dong, Zekun Qi, Linfeng Zhang, Junbo Zhang, Jianjian Sun, Zheng Ge, Li Yi, and Kaisheng Ma. Autoencoders as cross-modal teachers: Can pretrained 2D image transformers help 3D representation learning? arXiv preprint arXiv:2212.08320, 2022. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.801, + 0.483, + 0.828 + ], + "angle": 0, + "content": "[13] Gene H Golub and Charles F Van Loan. Matrix computations. JHU press, 2013. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.83, + 0.483, + 0.871 + ], + "angle": 0, + "content": "[14] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752, 2023. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.483, + 0.902 + ], + "angle": 0, + "content": "[15] Albert Gu, Tri Dao, Stefano Ermon, Atri Rudra, and Christopher Ré. Hippo: Recurrent memory with optimal polynomial" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.094, + 0.905, + 0.12 + ], + "angle": 0, + "content": "projections. Advances in neural information processing systems, 33:1474-1487, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.122, + 0.905, + 0.162 + ], + "angle": 0, + "content": "[16] Albert Gu, Karan Goel, and Christopher Ré. Efficiently modeling long sequences with structured state spaces. arXiv preprint arXiv:2111.00396, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.164, + 0.905, + 0.218 + ], + "angle": 0, + "content": "[17] Albert Gu, Karan Goel, Ankit Gupta, and Christopher Ré. On the parameterization and initialization of diagonal state space models. Advances in Neural Information Processing Systems, 35:35971-35983, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.22, + 0.905, + 0.274 + ], + "angle": 0, + "content": "[18] Yulan Guo, Hanyun Wang, Qingyong Hu, Hao Liu, Li Liu, and Mohammed Bennamoun. Deep learning for 3d point clouds: A survey. IEEE transactions on pattern analysis and machine intelligence, 43(12):4338-4364, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.275, + 0.905, + 0.327 + ], + "angle": 0, + "content": "[19] Ankit Gupta, Albert Gu, and Jonathan Berant. Diagonal state spaces are as effective as structured state spaces. Advances in Neural Information Processing Systems, 35:22982-22994, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.33, + 0.905, + 0.384 + ], + "angle": 0, + "content": "[20] Ramin Hasani, Mathias Lechner, Tsun-Hsuan Wang, Makram Chahine, Alexander Amini, and Daniela Rus. Liquid structural state-space models. arXiv preprint arXiv:2209.12951, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.386, + 0.905, + 0.441 + ], + "angle": 0, + "content": "[21] Cheng-Yao Hong, Yu-Ying Chou, and Tyng-Luh Liu. Attention discriminant sampling for point clouds. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14429-14440, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.443, + 0.907, + 0.511 + ], + "angle": 0, + "content": "[22] Xin Lai, Jianhui Liu, Li Jiang, Liwei Wang, Hengshuang Zhao, Shu Liu, Xiaojuan Qi, and Jiaya Jia. Stratified transformer for 3d point cloud segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8500-8509, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.511, + 0.905, + 0.567 + ], + "angle": 0, + "content": "[23] Loic Landrieu and Martin Simonovsky. Large-scale point cloud semantic segmentation with superpoint graphs. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4558-4567, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.568, + 0.905, + 0.636 + ], + "angle": 0, + "content": "[24] Juho Lee, Yoonho Lee, Jungtaek Kim, Adam Kosierek, Seungjin Choi, and Yee Whye Teh. Set transformer: A framework for attention-based permutation-invariant neural networks. In International conference on machine learning, pages 3744-3753. PMLR, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.637, + 0.905, + 0.691 + ], + "angle": 0, + "content": "[25] Yangyan Li, Rui Bu, Mingchao Sun, Wei Wu, Xinhan Di, and Baoquan Chen. Pointcnn: Convolution on x-transformed points. Advances in neural information processing systems, 31, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.692, + 0.905, + 0.748 + ], + "angle": 0, + "content": "[26] Dingkang Liang, Xin Zhou, Xinyu Wang, Xingkui Zhu, Wei Xu, Zhikang Zou, Xiaqing Ye, and Xiang Bai. Pointmamba: A simple state space model for point cloud analysis. arXiv preprint arXiv:2402.10739, 2024. 1, 2, 4, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.749, + 0.905, + 0.803 + ], + "angle": 0, + "content": "[27] Haotian Liu, Mu Cai, and Yong Jae Lee. Masked discrimination for self-supervised learning on point clouds. In European Conference on Computer Vision, pages 657-675. Springer, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.804, + 0.905, + 0.872 + ], + "angle": 0, + "content": "[28] Yang Liu, Chen Chen, Can Wang, Xulin King, and Mengyuan Liu. Regress before construct: Regress autoencoder for point cloud self-supervised learning. In Proceedings of the 31st ACM International Conference on Multimedia, pages 1738-1749, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.902 + ], + "angle": 0, + "content": "[29] Yue Liu, Yunjie Tian, Yuzhong Zhao, Hongtian Yu, Lingxi Xie, Yaowei Wang, Qixiang Ye, and Yunfan Liu. Vmamba:" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.094, + 0.907, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11807" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.119 + ], + "angle": 0, + "content": "Visual state space model. arXiv preprint arXiv:2401.10166, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.121, + 0.483, + 0.175 + ], + "angle": 0, + "content": "[30] Xu Ma, Can Qin, Haoxuan You, Haoxi Ran, and Yun Fu. Rethinking network design and local geometry in point cloud: A simple residual mlp framework. arXiv preprint arXiv:2202.07123, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.177, + 0.482, + 0.218 + ], + "angle": 0, + "content": "[31] Andrew Ng, Michael Jordan, and Yair Weiss. On spectral clustering: Analysis and an algorithm. Advances in neural information processing systems, 14, 2001. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.219, + 0.482, + 0.286 + ], + "angle": 0, + "content": "[32] Yatian Pang, Wenxiao Wang, Francis EH Tay, Wei Liu, Yonghong Tian, and Li Yuan. Masked autoencoders for point cloud self-supervised learning. In European conference on computer vision, pages 604-621. Springer, 2022. 1, 2, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.288, + 0.482, + 0.343 + ], + "angle": 0, + "content": "[33] Maciej Pioro, Kamil Ciebiera, Krystian Król, Jan Ludziejewski, and Sebastian Jaszczur. Moe-mamba: Efficient selective state space models with mixture of experts. arXiv preprint arXiv:2401.04081, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.344, + 0.482, + 0.412 + ], + "angle": 0, + "content": "[34] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 652-660, 2017. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.413, + 0.482, + 0.468 + ], + "angle": 0, + "content": "[35] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems, 30, 2017. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.469, + 0.482, + 0.538 + ], + "angle": 0, + "content": "[36] Charles R Qi, Yin Zhou, Mahyar Najibi, Pei Sun, Khoa Vo, Boyang Deng, and Dragomir Anguelov. Offboard 3d object detection from point cloud sequences. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6134-6144, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.539, + 0.482, + 0.607 + ], + "angle": 0, + "content": "[37] Guocheng Qian, Yuchen Li, Houwen Peng, Jinjie Mai, Hasan Hammoud, Mohamed Elhoseiny, and Bernard Ghanem. Pointnext: Revisiting pointnet++ with improved training and scaling strategies. Advances in Neural Information Processing Systems, 35:23192-23204, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.608, + 0.482, + 0.663 + ], + "angle": 0, + "content": "[38] Martin Reuter, Franz-Erich Wolter, and Niklas Peinecke. Laplace-spectra as fingerprints for shape matching. In Proceedings of the 2005 ACM symposium on Solid and physical modeling, pages 101-106, 2005. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.664, + 0.482, + 0.705 + ], + "angle": 0, + "content": "[39] Radu Bogdan Rusu and Steve Cousins. 3d is here: Point cloud library (pcl). In 2011 IEEE international conference on robotics and automation, pages 1-4. IEEE, 2011. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.706, + 0.482, + 0.747 + ], + "angle": 0, + "content": "[40] Jianbo Shi and Jitendra Malik. Normalized cuts and image segmentation. IEEE Transactions on pattern analysis and machine intelligence, 22(8):888-905, 2000. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.748, + 0.482, + 0.815 + ], + "angle": 0, + "content": "[41] Shaoshuai Shi, Xiaogang Wang, and Hongsheng Li. Poincaré: 3d object proposal generation and detection from point cloud. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 770-779, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.817, + 0.482, + 0.872 + ], + "angle": 0, + "content": "[42] Yuan Tang, Xianzhi Li, Jinfeng Xu, Qiao Yu, Long Hu, Yixue Hao, and Min Chen. Point-lgmask: Local and global contexts embedding for point cloud pre-training with multi-ratio masking. IEEE Transactions on Multimedia, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[43] Yi Tay, Dara Bahri, Liu Yang, Donald Metzler, and Da-Cheng Juan. Sparse sinkhorn attention. In International" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.093, + 0.905, + 0.119 + ], + "angle": 0, + "content": "Conference on Machine Learning, pages 9438-9447. PMLR, 2020.1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.121, + 0.905, + 0.189 + ], + "angle": 0, + "content": "[44] Mikaela Angelina Uy, Quang-Hieu Pham, Binh-Son Hua, Duc Thanh Nguyen, and Sai-Kit Yeung. Revisiting point cloud classification: A new benchmark dataset and classification model on real-world data. In International Conference on Computer Vision (ICCV), 2019. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.19, + 0.905, + 0.272 + ], + "angle": 0, + "content": "[45] Mikaela Angelina Uy, Quang-Hieu Pham, Binh-Son Hua, Thanh Nguyen, and Sai-Kit Yeung. Revisiting point cloud classification: A new benchmark dataset and classification model on real-world data. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1588–1597, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.274, + 0.905, + 0.329 + ], + "angle": 0, + "content": "[46] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.33, + 0.905, + 0.397 + ], + "angle": 0, + "content": "[47] Hanchen Wang, Qi Liu, Xiangyu Yue, Joan Lasenby, and Matt J Kusner. Unsupervised point cloud pre-training via occlusion completion. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9782-9792, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.399, + 0.905, + 0.455 + ], + "angle": 0, + "content": "[48] Yue Wang, Yongbin Sun, Ziwei Liu, Sanjay E Sarma, Michael M Bronstein, and Justin M Solomon. Dynamic graph cnn for learning on point clouds. ACM Transactions on Graphics (tog), 38(5):1-12, 2019. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.456, + 0.905, + 0.511 + ], + "angle": 0, + "content": "[49] Chengzhi Wu, Junwei Zheng, Julius Pfrommer, and Jürgen Beyerer. Attention-based point cloud edge sampling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5333-5343, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.511, + 0.905, + 0.566 + ], + "angle": 0, + "content": "[50] Xiaoyang Wu, Yixing Lao, Li Jiang, Xihui Liu, and Hengshuang Zhao. Point transformer v2: Grouped vector attention and partition-based pooling. Advances in Neural Information Processing Systems, 35:33330-33342, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.567, + 0.905, + 0.622 + ], + "angle": 0, + "content": "[51] Xiaoyang Wu, Li Jiang, Peng-Shuai Wang, Zhijian Liu, Xihui Liu, Yu Qiao, Wanli Ouyang, Tong He, and Hengshuang Zhao. Point transformer v3: Simpler, faster, stronger. arXiv preprint arXiv:2312.10035, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.623, + 0.905, + 0.692 + ], + "angle": 0, + "content": "[52] Zhirong Wu, Shuran Song, Aditya Khosla, Fisher Yu, Linguang Zhang, Xiaou Tang, and Jianxiong Xiao. 3d shapenets: A deep representation for volumetric shapes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1912-1920, 2015. 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.692, + 0.905, + 0.761 + ], + "angle": 0, + "content": "[53] Li Yi, Vladimir G Kim, Duygu Ceylan, I-Chao Shen, Mengyan Yan, Hao Su, Cewu Lu, Qixing Huang, Alla Sheffer, and Leonidas Guibas. A scalable active framework for region annotation in 3d shape collections. ACM Transactions on Graphics (ToG), 35(6):1-12, 2016. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.762, + 0.905, + 0.831 + ], + "angle": 0, + "content": "[54] Xumin Yu, Lulu Tang, Yongming Rao, Tiejun Huang, Jie Zhou, and Jiwen Lu. Point-bert: Pre-training 3d point cloud transformers with masked point modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19313-19322, 2022. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.832, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[55] Renrui Zhang, Ziyu Guo, Peng Gao, Rongyao Fang, Bin Zhao, Dong Wang, Yu Qiao, and Hongsheng Li. Point-m2ae: multi-scale masked autoencoders for hierarchical point cloud pre-training. Advances in neural information processing systems, 35:27061-27074, 2022. 1, 2, 5, 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "11808" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.161 + ], + "angle": 0, + "content": "[56] Renrui Zhang, Liuhui Wang, Yu Qiao, Peng Gao, and Hongsheng Li. Learning 3d representations from 2d pre-trained models via image-to-point masked autoencoders. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21769-21780, 2023. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.163, + 0.482, + 0.218 + ], + "angle": 0, + "content": "[57] Tao Zhang, Xiangtai Li, Haobo Yuan, Shunping Ji, and Shuicheng Yan. Point could mamba: Point cloud learning via state space model. arXiv preprint arXiv:2403.00762, 2024.1, 2, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.22, + 0.482, + 0.288 + ], + "angle": 0, + "content": "[58] Hengshuang Zhao, Li Jiang, Chi-Wing Fu, and Jiaya Jia. Pointweb: Enhancing local neighborhood features for point cloud processing. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5565-5573, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.29, + 0.482, + 0.346 + ], + "angle": 0, + "content": "[59] Hengshuang Zhao, Li Jiang, Jiaya Jia, Philip HS Torr, and Vladlen Koltun. Point transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 16259-16268, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.348, + 0.482, + 0.402 + ], + "angle": 0, + "content": "[60] Lianghui Zhu, Bencheng Liao, Qian Zhang, Xinlong Wang, Wenyu Liu, and Xinggang Wang. Vision mamba: Efficient visual representation learning with bidirectional state space model. arXiv preprint arXiv:2401.09417, 2024. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.402 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.956 + ], + "angle": 0, + "content": "11809" + } + ] +] \ No newline at end of file diff --git a/2025/Spectral Informed Mamba for Robust Point Cloud Processing/9acc5c2f-4c09-4911-b4a0-485ec3edcc53_origin.pdf b/2025/Spectral Informed Mamba for Robust Point Cloud Processing/9acc5c2f-4c09-4911-b4a0-485ec3edcc53_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..5e00114bd55ec4511523018de0ea34dd9fc6e985 --- /dev/null +++ b/2025/Spectral Informed Mamba for Robust Point Cloud Processing/9acc5c2f-4c09-4911-b4a0-485ec3edcc53_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:23b806383530865080e5cb36ce1a8e559950bbedc64071a121e3a87757b9b618 +size 1941539 diff --git a/2025/Spectral Informed Mamba for Robust Point Cloud Processing/full.md b/2025/Spectral Informed Mamba for Robust Point Cloud Processing/full.md new file mode 100644 index 0000000000000000000000000000000000000000..3354786af0fddc5e6cee5152675683fc6c67e011 --- /dev/null +++ b/2025/Spectral Informed Mamba for Robust Point Cloud Processing/full.md @@ -0,0 +1,286 @@ +# Spectral Informed Mamba for Robust Point Cloud Processing + +Ali Bahri * Moslem Yazdanpanah Mehrdad Noori Sahar Dastani +Milad Cheraghalikhani Gustavo Adolfo Vargas Hakim David Osowiechi Farzad Beizaee Ismail Ben Ayed Christian Desrosiers + +LIVIA, ÉTS Montréal, Canada International Laboratory on Learning Systems (ILLS) + +# Abstract + +State Space Models (SSMs) have shown significant promise in Natural Language Processing (NLP) and, more recently, computer vision. This paper introduces a new methodology that leverages Mamba and Masked Autoencoder (MAE) networks for point-cloud data in both supervised and self-supervised learning. We propose three key contributions to enhance Mamba's capability in processing complex pointcloud structures. First, we exploit the spectrum of a graph Laplacian to capture patch connectivity, defining an isometry-invariant traversal order that is robust to viewpoints and captures shape manifolds better than traditional 3D grid-based traversals. Second, we adapt segmentation via a recursive patch partitioning strategy informed by Laplacian spectral components, allowing finer integration and segment analysis. Third, we address token placement in MAE for Mamba by restoring tokens to their original positions, which preserves essential order and improves learning. Extensive experiments demonstrate the improvements that our approach brings over state-of-the-art baselines in classification, segmentation, and few-shot tasks. The implementation is available at: https://github.com/AliBahri94/SI-Mamba.git. + +# 1. Introduction + +The analysis of 3D point cloud data is fundamental for various applications, including autonomous driving [36, 41], VR/AR [18], and robotics [39]. Compared to the organized structure of 2D images, point clouds consist of 3D coordinates without direct adjacency information forming an unordered bag. In recent years, considerable efforts have been dedicated to adapt deep learning models such as convolutional neural networks (CNNs) and Transformers to this type of data [2, 3, 32, 35, 54, 55]. Due to their permutation-invariant self-attention mechanism, Transformer networks + +![](images/c1ffb0f79fd99dd1c2b963811007e840cd3bd8e4e1fb88d7fea676113098dd50.jpg) +Figure 1. (a) Surface-Aware Spectral Traversing (SAST) applied to the patched point clouds of a mesh surface. (b) Traversal from left to right, based on the first to fourth non-constant smallest eigenvectors. (c) Traversal based on the largest eigenvector, forming a non-continuous sequence of tokens. + +are particularly well suited for the unordered nature of point clouds. However, the quadratic complexity of this mechanism, requiring the computation of a weight between each pair of tokens, impedes the application of these networks to large-sized inputs (e.g., 2D images or 3D point clouds represented by many patches). This has prompted researchers to explore more efficient solutions, including the Set Transformer [24], Sparse Transformer [7], Longformer [4] and Sinkhorn Transformer [43]. + +Recently, methods based on Structured State Space Sequence (S4) [16] such as Mamba [14] have gained significant traction as a more efficient alternative to Transformers [29, 60]. So far, very few studies have investigated the potential of S4 approaches such as Mamba for 3D point clouds. Existing methods like Point-Mamba [26] and PCM [57] extend the 2D grid-based traversal used for images to a 3D grid. However, this straightforward adaptation to point clouds suffers from three crucial problems. First: whereas patches from 2D images have adjacency information, which could be exploited by the grid-based traversal, the 3D point patches in point clouds offer a sparse representation of the object's surface, and nearby patches on a 3D grid are not necessarily adjacent on this surface. Second: in the absence of self-attention, task-specific performance is highly influenced by the nature of the token traversal strategy. For example, a traversal suitable for point cloud classification may not be effective for a local task such as point-level classification (i.e., segmentation). Third: due to the "direction-sensitive" nature of Mamba, the self-supervised MAE pre + +training step of leading point cloud models like Point-MAE [32] and Point-M2AE [55] cannot be used directly as there is no attention mechanism to learn the masked tokens' positions. + +The contribution of our work focuses on addressing these problems as follows: + +1. We introduce a Surface-Aware Spectral Traversing (SAST) strategy based on the Laplacian spectrum of a patch-connectivity graph. Compared to the 3D grid traversal of current approaches like Point-Mamba, our strategy is invariant to isometric transformations (e.g., choice of viewpoint) and better captures the object's surface manifold. +2. We also present a Hierarchical Local Traversing (HLT) for point-level classification (segmentation) that partitions patches recursively based on their spectral coordinates. Unlike our SAST strategy for classification, which considers Laplacian eigenvectors separately in different traversals, this HLT combines them in a single ordering for a more precise modeling of geometry. +3. During the MAE-based Self-Supervised Learning (SSL), we propose a Traverse-Aware Repositioning (TAR) strategy to align the masked tokens according to their spectral adjacency. This strategy addresses the critical issue of spatial adjacency preservation unique to Mamba networks. + +# 2. Related Works + +Deep Point Cloud Learning. Deep neural networks have increasingly been applied to point clouds, with early architectures like PointNet [34] and PointNet++ [35] inspiring further work [1, 10, 23, 25, 58] focused on capturing local context. Transformer-based frameworks [46], including the Point Transformer (v1-v3) [50, 51, 59] and Stratified Transformer [22], have set new benchmarks by integrating both local and global features. Recently, self-supervised pre-training has gained traction by leveraging masked point modeling [28, 42] to train Transformers on unlabeled data. + +Building on the effectiveness of MAE in Text and Image domains, Point-BERT [54] presented a revolutionary method inspired by BERT [11], tailoring Transformers to 3D point cloud processing. Point-MAE [32] applied MAE-style pre-training to 3D point clouds using a custom Transformer-based Autoencoder (AE) designed to reconstruct masked irregular patches. The use of multiscale masking and local spatial self-attention mechanisms in Point-M2AE [55] has led to State-Of-The-Art (SOTA) results in 3D representation learning. Furthermore, I2P-MAE [56] improved self-supervised point cloud processing with a masking strategy leveraging pre-trained 2D models through an Image-to-Point transformation. Point-GPT [6] introduced an auto-regressive generative pretraining (GPT) + +approach to address the unordered nature and low information density of point clouds. ACT [12] proposed a cross-modal knowledge transfer method using pretrained 2D or natural language Transformers as teachers for 3D representation learning. Finally, GeoMask3D [2] is a self-supervised point cloud method that uses a teacher-student model to select harder, geometrically intricate regions for masking, thereby improving feature representations. + +# 2.1. State Space Models + +State Space Models. State-space models (SSMs) have been foundational in control theory and signal processing for modeling dynamic systems. Recently, they have been adapted to deep learning, with advancements like the Linear State-Space Layer (LSSL), which uses a continuous-time memorization framework based on the HiPPO operator [15] to model long-range dependencies. However, LSSL's high computational demands limit its practicality. To address this, the S4 model [16] introduced parameter normalization techniques, paving the way for structured SSMs with enhancements like complex-diagonal structures [17, 19], support for multiple-input/output [17], and low-rank decomposition [20]. Recent developments include Mamba [14], which achieves linear-time inference and efficient training through selection mechanisms and hardware-aware algorithms. MoE-Mamba further boosts efficiency by integrating a Mixture of Experts (MoE), outperforming both Mamba and Transformer-MoE models [33]. + +SSMs for Vision Tasks The above-mentioned works primarily focused on the application of SSMs to long-range or causal data types such as language and speech. In the field of vision, a notable study [29] proposed the VMamba model which features a Cross-Scan Module (CSM) for enhanced 1D selective scanning in 2D spaces and architectural optimizations that significantly improve its performance and speed across various visual tasks. Another significant paper is Vision Mamba [60] which introduces a novel vision backbone called Vim utilizing bidirectional Mamba blocks. + +For point cloud analysis based on Mamba, two key works are Point-Mamba [26] and PCM [57]. PointMamba introduces a simple approach to token reordering for point cloud analysis by strategically organizing point tokens based on a 3D grid. Similarly, PCM enhances Mamba with a Consistent Traverse Serialization (CTS) technique that converts 3D point clouds into 1D point sequences while maintaining spatial adjacency. Building upon these approaches, our method introduces the Spectral Spatial Traversing (SST) strategy, which improves token ordering and maintains spatial adjacency within Mamba networks. + +# 3. Method + +We begin by outlining the fundamental concepts of SSMs and spectral graph analysis which are at the core of our + +![](images/96e7d1d4fec27302183146eb7b4f1db41038e8bd4dc37c256de2af823c5b5507.jpg) +Figure 2. Overview of the proposed Spectral Spatial Traversing (SST) method. (a) Point cloud, (b) Patchification, (c) Forming the adjacency graph, (d) Traversal based on SAST using $s$ non-constant smallest eigenvectors, (e) HLT for segmentation tasks, (f) TAR strategy for Masked Autoencoders. The process includes reverse and concatenation operations, with learnable tokens, representations, and masked tokens highlighted. (g) The classification task involves sorting tokens by different eigenvectors, concatenating them, and then feeding them into the network. (h) The segmentation task where HLT is applied on the tokens $(\vec{q})$ and $\vec{q}$ is fed into the network. (i) A flowchart visualizing the techniques used in self-supervised learning and various downstream tasks. + +work. We then give an in-depth presentation of our Surface-Aware Spectral Traversing (SAST) strategy for point cloud processing that improves the model's robustness to isometric transformations and better captures the underlying manifold of the point cloud. Thereafter, we provide detailed specifications of our Hierarchical Local Traversing (HLT) strategy for point-level classification, which defines a more structured patch traversal order based on the recursive partitioning of spectral information. Finally, we introduce our Traverse-Aware Repositioning (TAR), which improves the handling of learnable tokens in masked autoencoders within Mamba networks. Fig. 2 illustrates the overview of the proposed Spectral Spatial Traversing (SST) method. + +# 3.1. Preliminaries + +State Space Models (SSMs) use a series of first-order differential equations to describe how the state of the linear, + +time-invariant system evolves over time: + +$$ +\dot {h} (t) = A h (t) + B x (t), \quad y (t) = C \dot {h} (t) + D x (t), \quad (1) +$$ + +Here, $\dot{h}(t)$ denotes the time derivative of the state vector $h(t)$ . The matrices $A$ , $B$ , $C$ , and $D$ are the weighting parameters. + +Due to their reliance on continuous data streams $x(t)$ , SSMs are not natively equipped to handle discrete inputs represented as $\{x_0, x_1, \ldots\}$ . This necessitates the use of a discretized SSM version for practical applications: + +$$ +h _ {k} = \bar {A} h _ {k - 1} + \bar {B} x _ {k}, \quad y _ {k} = \bar {C} h _ {k} + \bar {D} x _ {k}. \tag {2} +$$ + +The state space model in its discrete version utilizes a recursive function to link each state $h_k$ to its preceding state, encapsulated by the matrices $\bar{A} \in \mathbb{R}^{N \times N}$ , $\bar{B} \in \mathbb{R}^{N \times 1}$ , and $\bar{C} \in \mathbb{R}^{N \times 1}$ , which are tuned parameter matrices. While + +![](images/c3d67f5050addff6fa87af5f531b71eb6f0618300350fd11ef6b1ee75ca2beb9.jpg) +Figure 3. (a) Visualization of the four non-constant smallest Laplacian eigenvectors $(v^{(k)}, k = 1, \ldots, 4)$ and (b) the discrete partitioning $(q)$ of our HLT strategy combining the information of all four eigenvectors. Note: we assumed that patches contain a single point for better visualization. + +matrix $\bar{D} \in \mathbb{R}^{N \times 1}$ may be employed as a residual connection, we follow previous work and exclude it from our model. The transition from a continuous signal representation $x(t)$ to a discrete sequence involves sampling $x(t)$ at intervals defined by $\Delta$ , setting each discrete input as $x_{k} = x(k\Delta)$ . This adjustment to a discrete framework results in revised matrix definitions: + +$$ +\bar {A} = \left(I - \frac {\Delta}{2} A\right) ^ {- 1} \left(I + \frac {\Delta}{2} A\right), \bar {B} = \left(I - \frac {\Delta}{2} A\right) ^ {- 1} \Delta B, \bar {C} = C. \tag {3} +$$ + +However, the fixed dynamics of Linear Time-Invariant (LTI) models, exemplified by the constant parameters $A$ , $B$ , and $C$ in Eq. (3), restrict their capacity to selectively retain or discard relevant information, thereby limiting their contextual awareness. To improve content-aware reasoning, we use Mamba's selection mechanism that manages the propagation and interaction of information across the sequence dimension [14]. + +Spectral Graph Analysis. Popularized by Chung in the 90s [8], spectral graph analysis characterizes the properties of a graph $G = (V, E)$ by the spectrum (eigenvalues and corresponding eigenvectors) of its Laplacian matrix $L$ . This analysis can be understood as a discretized version of the Laplace-Beltrami Operator $\Delta$ of a function $f$ defined on a Riemannian manifold: + +$$ +\Delta f = \operatorname {d i v} (\operatorname {g r a d} f) \tag {4} +$$ + +where $\operatorname{grad}f$ is the gradient of $f$ and $\operatorname{div}$ the divergence on the manifold. The solution to the Laplacian eigenvalue problem $\Delta f = -\lambda f$ , known as Helmholtz wave equation, is an eigenfunction corresponding to the natural vibration form of a homogeneous membrane with eigenvalue $\lambda$ [38]. + +Following methods for spectral clustering [31] and normalized cuts [40], we consider a weighted adjacency matrix $W: V \times V \to \mathbb{R}_+$ where $W_{ij} = 0$ if $(i,j) \notin E$ to model the Euclidean distance of nearby patches (see Section 3.3). The Laplacian matrix of $G$ is defined as $L = D - W$ where + +$D$ is the diagonal degree matrix such that $D_{ii} = \sum_{j}W_{ij}$ . To account for variability in the scale of weights $W_{ij}$ or the distribution of node degrees $D_{ii}$ , it is preferable to employ a normalized version of the Laplacian. In this work, we use the Random Walk Laplacian $L_{rw} = I - D^{-1}W$ which has the following useful properties: + +1. $L_{rw}$ is positive semi-definite and has $|V|$ non-negative real-valued eigenvalues $0 = \lambda_1 \leq \ldots \leq \lambda_{|V|}$ ; +2. 0 is an eigenvalue of $L_{rw}$ with the constant vector as eigenvector and its multiplicity equals the number of connected components in the graph; +3. Following Courant's Nodal Line Theorem [9], the $n$ -th eigenmode of $L_{rw}$ has at most $n$ poles of vibration; +4. The representation of a shape by the spectrum of $L_{rw}$ is invariant to isometry (i.e., distance-preserving transformation). + +Our method uses the $s$ first non-constant eigenvectors of $L_{rw}$ (i.e., the eigenvectors corresponding to the $s$ smallest non-zero eigenvalues) to define traversal orders for classification (Section 3.3) and segmentation (Section 3.4) that are robust to the viewpoint (due to isometry invariance) and provide a smooth parametrization of the surface manifold. We consider the first eigenvectors as they encode low frequency information (by Courant's Nodal Line Theorem), making the resulting traversal more robust to shape variability and noise. Fig. 1 illustrates this concept: (a) shows the original mesh, (b) shows traversals based on the first to fourth non-constant smallest eigenvectors, and (c) shows traversal based on the largest eigenvector forming a non-continuous sequence of tokens. + +# 3.2. Point Cloud Patchification + +Given a point cloud $\mathcal{P} = \{p_i\}_{i=1}^{N_p}$ , each point represented by 3D coordinates, we convert $\mathcal{P}$ to a reduced set of patches that can be processed more efficiently. Toward this goal, we employ the Farthest Point Sampling (FPS) algorithm to select a subset $\mathcal{C} \subset \mathcal{P}$ of $N_c$ points offering a good coverage of the entire point cloud. These selected points will act as the centers of local patches within the point cloud. For each center point $p_{s_i} \in \mathcal{C}$ , we then identify $N_n$ nearest points $\mathcal{N}(p_{s_i}) \subset \mathcal{P}$ using the K-Nearest Neighbours (KNN) algorithm. Following this, each patch is defined as a center $p_{s_i}$ and its corresponding nearest-neighbors $\mathcal{N}(p_{s_i})$ . + +# 3.3. Surface-Aware Spectral Traversing (SAST) + +Current point cloud processing approaches using Mamba, such as Point-Mamba [26] and PCM [57], simply extend the 2D grid-based traversal for images to a 3D grid. As mentioned before, this naive strategy suffers from two issues: 1) the 3D grid is view dependent, thus rotating the point cloud or moving the camera yields a different traversal order; 2) unrelated patches may be adjacent in 3D space, hence can be traversed subsequently. To address these problems, we + +define a traversal order based on the Laplacian spectrum of the patch-connectivity graph. + +In this graph, each node corresponds to a patch and the weighted adjacency matrix $W$ is defined using the Euclidean distance between patch centers. For patches $i$ and $j$ defined by center points $p_{s_i}$ and $p_{s_j}$ , we add an edge $(i,j)$ if $p_{s_j}$ is among the $K$ nearest neighbors of $p_{s_i}$ or vice-versa. The weight of this edge is computed using a Gaussian kernel: $W_{ij} = \exp \left(-\| p_{s_i} - p_{s_j}\| _2^2 /\sigma\right)$ where $\sigma$ is a hyperparameter controlling the kernel width. + +Following Section 3.1, we compute the $s$ first nonconstant eigenvectors of the Random Walk Laplacian $L_{rw}$ . This can be achieved efficiently using an iterative method like the Arnoldi algorithm [13] by exploiting the following facts: 1) matrix $W$ is very sparse, and 2) only the first few eigenvectors need to be computed. Eigenvector $v^{(k)} \in \mathbb{R}^{N_c}$ , $k \in \{1, \dots, s\}$ , assigns an eigenfunction value $v_i^{(k)}$ to each patch $i$ . In each Mamba block of our model, we perform two separate traversals of tokens (each token corresponds to an input patch) for every eigenvector: a forward traversal by increasing value of $v_i^{(k)}$ and a reserve traversal by decreasing value of $v_i^{(k)}$ . At the end of the block, we concatenate the tokens from the $s \times 2$ traversals. The visual effect of each traversal is illustrated in Fig. 1(b) and Fig. 3(a). + +Canonicalization of spectrum. Although the spectrum of $L_{rw}$ forms an isometry-invariant representation of the surface manifold, this representation may be impacted by two sources of ambiguity: 1) the sign of eigenvectors is undetermined (i.e., if $v^{(k)}$ is an eigenvector, then so is $-v^{(k)}$ ), and 2) the order of eigenvectors with similar eigenvalues may vary. We address these two sources of ambiguity with the following canonicalization procedure. For the first one, we flip the sign of an eigenvector $v^{(k)}$ (i.e., $v^{(k)} \coloneqq -v^{(k)}$ ) if its first element is negative (i.e., $v_{1}^{(k)} < 0$ ). To handle the second ambiguity, we first sort the eigenvectors by nondecreasing eigenvalue. We deal with eigenvalues having a multiplicity greater than one by finding pairs of consecutive eigenvectors $v^{(k)}$ , $v^{(k+1)}$ with near-identical eigenvalues (i.e., $|\lambda_k - \lambda_{k+1}| \leq \epsilon$ ). For such pairs, we flip the order if $v_{1}^{(k)} > v_{1}^{(k+1)}$ . This reordering process is repeated until no further change occurs. + +# 3.4. Hierarchical Local Traversing (HLT) for Segmentation + +While effective for classification tasks, the SAST strategy considering each eigenvector in a separate traversal may not capture the precise relationship between patches needed for segmentation. To address this issue, we introduce a Hierarchical Local Traversal (HLT) strategy that considers the full spectrum (all $s$ non-constant eigenvectors) simultaneously. + +Our strategy is inspired by the recursive binary partitioning technique of normalized cuts [40]. Starting from + +the canonicalized spectrum (see Section 3.3), tokens are first split based on the first eigenvector $v^{(1)}$ , by comparing their corresponding value in $v^{(1)}$ with the mean value $\overline{v}^{(1)}$ . This yields a binary partition of tokens $b_{i}^{(1)} = \mathbb{1}\big(v_{i}^{(1)}\geq \overline{v}^{(1)}\big)\in \{0,1\}$ where $\mathbb{1}$ is the indicator function. Each subgroup is then divided based on the mean of the second eigenvector $v^{(2)}$ , and so on for other eigenvectors. This partitioning process can be seen as building a binary tree, where each level corresponds to a different eigenvector and leaf nodes $i$ are uniquely identified by the sequence of bits $b_{i} = [b_{i}^{(1)},\dots ,b_{i}^{(s)}]$ on the path from the root to the leaf. Our HLT method traverses groups of leaf nodes (groups of tokens) sequentially based on the lexicographic order of their binary code (e.g., [0000], [0001], [0010], [0011], ... in the case of four eigenvectors). For convenience, we convert binary codes $b_{i}$ to a non-negative integer $q_{i}$ (e.g., bin2Int([0011]) = 3) and define two traversal orders, by increasing or decreasing values of $q_{i}$ . + +For $s$ eigenvectors, the HLT strategy described above divides tokens into $2^{s}$ segments which are traversed sequentially. In the best case scenario, $\lceil \log_2(N_c)\rceil$ eigenvectors are thus needed to split tokens into individual segments. However, it may happen that multiple tokens fall in the same segment, especially when using fewer eigenvectors. In such case, one can further sort tokens within each segment, for example, using the values of the first eigenvector (i.e., $v^{(1)}$ ). In our implementation, we simply sort these tokens randomly to add stochasticity in the training. This section is illustrated in Fig. 2 (e), and Fig. 3 (b). + +As shown in Fig. 3, the first Laplacian eigenvectors encode high-level spatial relations (e.g., bottom vs. top, left vs. right, torso vs. limbs, etc.). In the SAST, because these eigenvectors are used in separate traversals, the network may not be able to differentiate specific regions/parts of the point cloud. In contrast, our HLT strategy can capture such specific parts (e.g., head, left or right arm/thigh/alf, pelvis, etc.). + +# 3.5. Traverse-Aware Repositioning (TAR) for Masked Autoencoders + +Following state-of-art Transformers for point cloud processing, such as Point-MAE [32] and Point-M2AE [55], our method leverages self-supervised pretraining based on MAE to boost performance. In Transformer-based approaches, the learnable tokens of masked patches can be inserted in any position of the sequence (typically at the end) as the self-attention mechanism can still attend to all tokens irrespective of their positions. However, this approach presents a significant problem in Mamba networks which are sensitive to the traversal order of tokens. We handle this problem via a TAR strategy that improves the placement of learnable tokens in MAE within Mamba networks. Specifically, we restore the learnable tokens to their + +original positions rather than appending them at the end of the sequence. This ensures that the essential order of tokens is maintained, preserving spatial adjacency and enhancing learning effectiveness within Mamba networks. + +The proposed TAR strategy selects an arbitrary traversal order and randomly masks a subset of $N_{m}$ tokens with the same masking ratio as the transformer-based MAEs. These tokens are then removed from the sequence, and their positions are recorded. Afterwards, the remaining (visible) tokens are fed to the encoder that outputs their representation. Before reconstructing the point cloud using the decoder, we reinsert the learnable tokens in the sequence at their recorded position. The set of masked patches is used for other traversal orders. This procedure can be seen in Fig. 2 (f). Following previous work, we measure the reconstruction error for masked patches using the Chamfer distance: + +$$ +\mathcal {L} _ {r e c} = \frac {1}{N _ {m}} \sum_ {i = 1} ^ {N _ {m}} \operatorname {C h a m f e r} \left(\mathcal {S} _ {i}, \hat {\mathcal {S}} _ {i}\right), \tag {5} +$$ + +where $S_{i}\in \mathbb{R}^{N_{n}\times 3}$ is the set of points forming the $i$ -th masked patch and $\hat{S}_i$ the reconstructed output for these points. The Chamfer distance between two sets of points $S$ and $\hat{S}$ is defined as + +$$ +\begin{array}{l} \operatorname {C h a m f e r} \left(\mathcal {S}, \mathcal {S} ^ {\prime}\right) = \sum_ {p \in \mathcal {S}} \min _ {p ^ {\prime} \in \mathcal {S} ^ {\prime}} \| p - p ^ {\prime} \| _ {2} ^ {2} \\ + \sum_ {p ^ {\prime} \in \mathcal {S} ^ {\prime}} \min _ {p \in \mathcal {S}} \| p - p ^ {\prime} \| _ {2} ^ {2}. \tag {6} \\ \end{array} +$$ + +# 4. Experiments + +Several experiments are conducted to evaluate the proposed method. First, we pretrain the Point-Mamba network us + +Table 1. Few-shot classification on ModelNet40. We report the average accuracy (\%) and standard deviation (\%) of 10 independent experiments. \* denotes reproduced results. A $\equiv$ represents the average (A) and standard deviation (std), respectively. + +
Method5-way10-way
10-shot20-shot10-shot20-shot
DGCNN [48]91.83793.43286.3390.937
DGCNN + OcCo [47]91.93393.93186.4391.346
Transformer [54]87.83293.3384.6389.43
Transf. + OcCo [54]94.03695.9389.4392.446
Point-BERT [54]94.63196.3391.0392.73
Point-M2AE [55]96.83898.3492.3395.039
Point-MAE [32]96.33597.8392.6395.039
PointMamba* [26]95.93197.3391.6394.535
Ours96.43598.5392.0395.13
+ +ing our techniques on the ShapeNet [5] training dataset. We then assess the performance of these pretrained models across a variety of standard benchmarks, including object classification, few-shot learning, and segmentation. Additionally, we train the model from scratch on downstream datasets to demonstrate the robustness and versatility of our method. To have a fair comparison, we adopt the masking ratio (60%) that was used in the Point-Mamba model. Moreover, a comprehensive analysis of the computational efficiency, runtime, and memory usage of our SAST approach is provided in the Supplementary Material. + +# 4.1. Pretraining Setup + +Following Point-Mamba, we adopt the ShapeNet [5] dataset for the pretraining and assess the quality of the 3D representations produced by our approach through a linear evaluation on the ModelNet40 [52] dataset. The linear evaluation is performed by a Support Vector Machine (SVM) fitted on these features. This classification performance is quantified by the accuracy metric. Augmentation in pretraining is scale and transform, while in finetuning it is rotation. + +# 4.2. Downstream Tasks + +Object Classification on Real-World Dataset. To evaluate our method for point clouds, we test it on the ScanObjectNN dataset [44], as described in previous studies. The augmentation used during training is random rotation. The results, presented in Tab. 3, show that our strategy signifi + +Table 2. Part segmentation on the ShapeNetPart [53]. The mIoU for all instances (Inst.) is reported. HLT is Hierarchical Local Traversing for segmentation. + +
MethodsBackboneFLOPs (G)mIoU
Training from scratch
PointNet [34]--83.7
PointNet++ [35]--85.1
DGCNN [48]--85.2
APES [49]--85.8
Point-MAE [32]Transformer15.585.7
PointMamba [26]Mamba14.385.8
Ours (HLT)Mamba14.385.9
Training from pretrained
Transformer [54]Transformer15.585.1
Point-BERT [54]Transformer15.585.6
Point-MAE [32]Transformer15.586.1
Point-M2AE [55]Transformer15.586.5
Point-GPT-S [6]Transformer-86.2
ACT [12]Transformer-86.2
I2P-MAE [56]Transformer-86.8
PointMamba [26]Mamba14.386.0
Ours (SAST)Mamba14.385.7
Ours (HLT)Mamba14.386.1
+ +Table 3. Object classification on ScanObjectNN [45]. Accuracy (%) is reported. ${}^{ \dagger }$ indicates that this method was fine-tuned without rotation augmentation. + +
MethodsBackboneParam. (M)FLOPs (G)OBJ-BGOBJ-ONLYPB-T50-RSModelNet40
Training from scratch
PointNet [34]-3.50.573.379.268.089.2
PointNet++ [35]-1.51.782.384.377.990.7
DGCNN [48]-1.82.482.886.278.192.9
PointNeXt [37]-1.41.6--87.792.9
PointMLP [30]-13.231.4--85.4-
ADS [21]-----87.5-
Transformer [54]Transformer22.14.879.8680.5577.24-
Point-MAE [32]Transformer22.14.886.7586.9280.7892.3
PointMamba [26]Mamba12.33.690.8790.1885.6092.4
OursMamba12.33.692.2591.3987.3092.7
Training from pretrained
Transformer [54]Transformer22.14.879.8680.5577.2492.1
Transformer-OcCo [54]Transformer--84.8585.5478.79-
Point-BERT [54]Transformer22.14.887.4388.1283.0792.7
Point-M2AE† [55]Transformer--91.2288.8186.4394.0
Point-MAE [32]Transformer22.14.892.7791.2289.0493.2
PointMamba [26]Mamba12.33.693.2991.5688.1792.8
PCM [57]Mamba12.33.6--86.9-
OursMamba12.33.694.3291.9189.1093.4
+ +cantly improves object classification accuracy in both training from scratch and fine-tuning scenarios. These findings highlight the effectiveness of our approach in enhancing the model's ability to identify and classify objects across various backgrounds, demonstrating its robustness in complex real-world scenarios. + +Object Classification on Clean Objects Dataset. We also evaluated our method on the ModelNet40 [52] dataset, following the protocols established in previous works. The augmentation used during training is scale and transform. As shown in Tab. 3, our approach achieves notable enhancements on this challenging dataset compared to both the original Point-Mamba and the Transformer-based Point-MAE. This demonstrates the robustness and effectiveness of our method when applied to the Point-Mamba network. + +Few-shot Learning. We conducted few-shot learning experiments on ModelNet40 [52] dataset, adhering to the protocols of previous studies [27, 32, 55]. The results of our few-shot learning experiments are presented in Tab. 1. Despite the competitive nature of this benchmark, our method demonstrated outstanding performance across all tested scenarios. As shown in Tab. 1, our Mamba-based method achieves results comparable to or exceeding those of transformer-based methods (Point-MAE and PointM2AE). + +Part Segmentation. We evaluated our method's repre + +sensation learning ability on the ShapeNetPart dataset [53] using standard experimental settings [34, 35, 54]. Tab. 2 shows that performance gains in prior methods are minimal, highlighting the dataset's challenge. For instance, PointGPT [6] and ACT [12], despite being state-of-the-art and complex methods, show only marginal improvements over each other and other state-of-the-art approaches. Similarly, I2P-MAE [56], which combines 3D data with 2D information, shows limited improvement over Point-M2AE. For the "Training from scratch" setting, our method outperforms several state-of-the-art approaches. In the "Training from pretrained" setting, we further demonstrate the effectiveness of HLT strategy compared to SAST in the segmentation task. + +# 4.3. Ablation Studies + +In this section, we aim to investigate the effects of different parameters on our method. We will focus on two key aspects: the effect of the number of non-constant smallest eigenvectors and the adjacency matrix used in the SAST strategy, and the analysis of the TAR strategy. + +Analysis of Eigenvectors and Graph. One of our ablation studies investigates the impact of the number of nonconstant smallest eigenvectors used in the SAST and TAR strategies. As depicted in Fig. 4 (Left), the best performance is achieved when using the four non-constant smallest eigenvectors (blue line) for traversing. When the num + +![](images/a183ad70f343dd6cb6973cdc3ae5ce3f391de0338854a0167e48131c1bdcd942.jpg) +Figure 4. Analysis of the number of non-constant smallest eigenvectors and comparison with previous methods (Left) and analysis of the number of nearest neighbors $K$ (Right). + +![](images/45fb3d852d239418c06ea2d6bdb397a47f2fc819eb286f4e5a144bce4c88264f.jpg) + +![](images/b9858906576a19268085fad3e2faaf70055d2e773b96a86802ba0a297698ed8f.jpg) +Figure 5. The effect of the TAR strategy in the pretraining phase (Left) and in finetuning (Right). + +![](images/439b0cdc4c61d8661fdf6bca699605ea9dc2092531996e82d66ce847a0c8b077.jpg) + +ber of non-constant smallest eigenvectors increases beyond four, the performance drops. This is because the additional eigenvectors are closer to the largest eigenvectors, which are less smooth and do not capture the most significant structural variations effectively. The significantly lower performance of Point-MAE [32], Point-Mamba [26], and PointMamba with random traversing (Random-Mamba) compared to ours underscores the importance of appropriate traversing for Mamba networks. + +Another study examines the impact of the number of nearest neighbors used in creating the adjacency matrix. As shown in Fig. 4 (Right), the best performance (blue line) is achieved with 20 nearest neighbors. The model's accuracy increases as the number of nearest neighbors increases from 10 to 20, reaching its peak at 20 nearest neighbors. Beyond this point, the performance starts to decline. All these experiments are trained and evaluated from scratch on the ScanObjectNN dataset [44] (OBJ-BG) with scale and transform augmentation. + +Analysis of TAR Strategy. One of the studies examines the impact of the TAR strategy on the model's performance. As shown in Fig. 5 (Left), the accuracy of the model with and without the TAR strategy is plotted against the number of epochs. The model incorporating the TAR strategy (light blue line) demonstrates superior performance compared to the model without TAR (dark blue line). This figure relates + +to the pretraining phase on the ShapeNet [5] dataset, which is subsequently tested on the ModelNet [52] dataset using a SVM. + +Additionally, we evaluate the effect of the TAR strategy by fine-tuning pretrained models (with or without TAR) on four downstream tasks, each corresponding to a distinct dataset. As shown in Fig. 5 (Right), the model pretrained with TAR (light blue line) consistently achieves higher accuracy than the model trained without TAR (dark blue line). This demonstrates that the model pretrained with the TAR strategy has learned more meaningful features, leading to better performance in the downstream tasks. + +# 5. Conclusion and Limitation + +We introduced three strategies to enhance Mamba networks for point cloud data: isometry-invariant token traversal, recursive patch partitioning for segmentation, and improved learnable token placement, each contributing to a robust feature extraction framework. Our methods demonstrate superior performance over state-of-the-art baselines in classification, segmentation, and few-shot tasks. A limitation of our approach lies in the selection of the neighborhood size $K$ , which is treated as a hyperparameter. While tuning $K$ per dataset helps optimize performance, it may limit generalizability across diverse data distributions and require careful adjustment to maintain robustness. + +# References + +[1] Matan Atzmon, Haggai Maron, and Yaron Lipman. Point convolutional neural networks by extension operators. arXiv preprint arXiv:1803.10091, 2018. 2 +[2] Ali Bahri, Moslem Yazdanpanah, Mehrdad Noori, Milad Cheraghalikhani, Gustavo Adolfo Vargas Hakim, David Osowiechi, Farzad Beizaee, Ismail Ben Ayed, and Christian Desrosiers. Geomask3d: Geometrically informed mask selection for self-supervised point cloud learning in 3d. arXiv preprint arXiv:2405.12419, 2024. 1, 2 +[3] Ali Bahri, Moslem Yazdanpanah, Mehrdad Noori, Sahar Dastani, Milad Cheraghalikhani, David Osowiech, Farzad Beizaee, Ismail Ben Ayed, Christian Desrosiers, et al. Test-time adaptation in point clouds: Leveraging sampling variation with weight averaging. arXiv preprint arXiv:2411.01116, 2024. 1 +[4] Iz Beltagy, Matthew E Peters, and Arman Cohan. Longformer: The long-document transformer. arXiv preprint arXiv:2004.05150, 2020. 1 +[5] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 6, 8 +[6] Guangyan Chen, Meiling Wang, Yi Yang, Kai Yu, Li Yuan, and Yufeng Yue. PointGPT: Auto-regressively generative pre-training from point clouds. Advances in Neural Information Processing Systems, 36, 2024. 2, 6, 7 +[7] Rewon Child, Scott Gray, Alec Radford, and Ilya Sutskever. Generating long sequences with sparse transformers. arXiv preprint arXiv:1904.10509, 2019. 1 +[8] Fan RK Chung. Spectral graph theory. American Mathematical Soc., 1997. 4 +[9] Richard Courant and David Hilbert. Methods of mathematical physics: partial differential equations. John Wiley & Sons, 2008. 4 +[10] Xin Deng, WenYu Zhang, Qing Ding, and XinMing Zhang. Pointvector: a vector representation in point cloud analysis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9455-9465, 2023. 2 +[11] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 2 +[12] Runpei Dong, Zekun Qi, Linfeng Zhang, Junbo Zhang, Jianjian Sun, Zheng Ge, Li Yi, and Kaisheng Ma. Autoencoders as cross-modal teachers: Can pretrained 2D image transformers help 3D representation learning? arXiv preprint arXiv:2212.08320, 2022. 2, 6, 7 +[13] Gene H Golub and Charles F Van Loan. Matrix computations. JHU press, 2013. 5 +[14] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752, 2023. 1, 2, 4 +[15] Albert Gu, Tri Dao, Stefano Ermon, Atri Rudra, and Christopher Ré. Hippo: Recurrent memory with optimal polynomial + +projections. Advances in neural information processing systems, 33:1474-1487, 2020. 2 +[16] Albert Gu, Karan Goel, and Christopher Ré. Efficiently modeling long sequences with structured state spaces. arXiv preprint arXiv:2111.00396, 2021. 1, 2 +[17] Albert Gu, Karan Goel, Ankit Gupta, and Christopher Ré. On the parameterization and initialization of diagonal state space models. Advances in Neural Information Processing Systems, 35:35971-35983, 2022. 2 +[18] Yulan Guo, Hanyun Wang, Qingyong Hu, Hao Liu, Li Liu, and Mohammed Bennamoun. Deep learning for 3d point clouds: A survey. IEEE transactions on pattern analysis and machine intelligence, 43(12):4338-4364, 2020. 1 +[19] Ankit Gupta, Albert Gu, and Jonathan Berant. Diagonal state spaces are as effective as structured state spaces. Advances in Neural Information Processing Systems, 35:22982-22994, 2022. 2 +[20] Ramin Hasani, Mathias Lechner, Tsun-Hsuan Wang, Makram Chahine, Alexander Amini, and Daniela Rus. Liquid structural state-space models. arXiv preprint arXiv:2209.12951, 2022. 2 +[21] Cheng-Yao Hong, Yu-Ying Chou, and Tyng-Luh Liu. Attention discriminant sampling for point clouds. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14429-14440, 2023. 7 +[22] Xin Lai, Jianhui Liu, Li Jiang, Liwei Wang, Hengshuang Zhao, Shu Liu, Xiaojuan Qi, and Jiaya Jia. Stratified transformer for 3d point cloud segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8500-8509, 2022. 2 +[23] Loic Landrieu and Martin Simonovsky. Large-scale point cloud semantic segmentation with superpoint graphs. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4558-4567, 2018. 2 +[24] Juho Lee, Yoonho Lee, Jungtaek Kim, Adam Kosierek, Seungjin Choi, and Yee Whye Teh. Set transformer: A framework for attention-based permutation-invariant neural networks. In International conference on machine learning, pages 3744-3753. PMLR, 2019. 1 +[25] Yangyan Li, Rui Bu, Mingchao Sun, Wei Wu, Xinhan Di, and Baoquan Chen. Pointcnn: Convolution on x-transformed points. Advances in neural information processing systems, 31, 2018. 2 +[26] Dingkang Liang, Xin Zhou, Xinyu Wang, Xingkui Zhu, Wei Xu, Zhikang Zou, Xiaqing Ye, and Xiang Bai. Pointmamba: A simple state space model for point cloud analysis. arXiv preprint arXiv:2402.10739, 2024. 1, 2, 4, 6, 7, 8 +[27] Haotian Liu, Mu Cai, and Yong Jae Lee. Masked discrimination for self-supervised learning on point clouds. In European Conference on Computer Vision, pages 657-675. Springer, 2022. 7 +[28] Yang Liu, Chen Chen, Can Wang, Xulin King, and Mengyuan Liu. Regress before construct: Regress autoencoder for point cloud self-supervised learning. In Proceedings of the 31st ACM International Conference on Multimedia, pages 1738-1749, 2023. 2 +[29] Yue Liu, Yunjie Tian, Yuzhong Zhao, Hongtian Yu, Lingxi Xie, Yaowei Wang, Qixiang Ye, and Yunfan Liu. Vmamba: + +Visual state space model. arXiv preprint arXiv:2401.10166, 2024. 1, 2 +[30] Xu Ma, Can Qin, Haoxuan You, Haoxi Ran, and Yun Fu. Rethinking network design and local geometry in point cloud: A simple residual mlp framework. arXiv preprint arXiv:2202.07123, 2022. 7 +[31] Andrew Ng, Michael Jordan, and Yair Weiss. On spectral clustering: Analysis and an algorithm. Advances in neural information processing systems, 14, 2001. 4 +[32] Yatian Pang, Wenxiao Wang, Francis EH Tay, Wei Liu, Yonghong Tian, and Li Yuan. Masked autoencoders for point cloud self-supervised learning. In European conference on computer vision, pages 604-621. Springer, 2022. 1, 2, 5, 6, 7, 8 +[33] Maciej Pioro, Kamil Ciebiera, Krystian Król, Jan Ludziejewski, and Sebastian Jaszczur. Moe-mamba: Efficient selective state space models with mixture of experts. arXiv preprint arXiv:2401.04081, 2024. 2 +[34] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 652-660, 2017. 2, 6, 7 +[35] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems, 30, 2017. 1, 2, 6, 7 +[36] Charles R Qi, Yin Zhou, Mahyar Najibi, Pei Sun, Khoa Vo, Boyang Deng, and Dragomir Anguelov. Offboard 3d object detection from point cloud sequences. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6134-6144, 2021. 1 +[37] Guocheng Qian, Yuchen Li, Houwen Peng, Jinjie Mai, Hasan Hammoud, Mohamed Elhoseiny, and Bernard Ghanem. Pointnext: Revisiting pointnet++ with improved training and scaling strategies. Advances in Neural Information Processing Systems, 35:23192-23204, 2022. 7 +[38] Martin Reuter, Franz-Erich Wolter, and Niklas Peinecke. Laplace-spectra as fingerprints for shape matching. In Proceedings of the 2005 ACM symposium on Solid and physical modeling, pages 101-106, 2005. 4 +[39] Radu Bogdan Rusu and Steve Cousins. 3d is here: Point cloud library (pcl). In 2011 IEEE international conference on robotics and automation, pages 1-4. IEEE, 2011. 1 +[40] Jianbo Shi and Jitendra Malik. Normalized cuts and image segmentation. IEEE Transactions on pattern analysis and machine intelligence, 22(8):888-905, 2000. 4, 5 +[41] Shaoshuai Shi, Xiaogang Wang, and Hongsheng Li. Poincaré: 3d object proposal generation and detection from point cloud. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 770-779, 2019. 1 +[42] Yuan Tang, Xianzhi Li, Jinfeng Xu, Qiao Yu, Long Hu, Yixue Hao, and Min Chen. Point-lgmask: Local and global contexts embedding for point cloud pre-training with multi-ratio masking. IEEE Transactions on Multimedia, 2023. 2 +[43] Yi Tay, Dara Bahri, Liu Yang, Donald Metzler, and Da-Cheng Juan. Sparse sinkhorn attention. In International + +Conference on Machine Learning, pages 9438-9447. PMLR, 2020.1 +[44] Mikaela Angelina Uy, Quang-Hieu Pham, Binh-Son Hua, Duc Thanh Nguyen, and Sai-Kit Yeung. Revisiting point cloud classification: A new benchmark dataset and classification model on real-world data. In International Conference on Computer Vision (ICCV), 2019. 6, 8 +[45] Mikaela Angelina Uy, Quang-Hieu Pham, Binh-Son Hua, Thanh Nguyen, and Sai-Kit Yeung. Revisiting point cloud classification: A new benchmark dataset and classification model on real-world data. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1588–1597, 2019. 7 +[46] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 2 +[47] Hanchen Wang, Qi Liu, Xiangyu Yue, Joan Lasenby, and Matt J Kusner. Unsupervised point cloud pre-training via occlusion completion. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9782-9792, 2021. 6 +[48] Yue Wang, Yongbin Sun, Ziwei Liu, Sanjay E Sarma, Michael M Bronstein, and Justin M Solomon. Dynamic graph cnn for learning on point clouds. ACM Transactions on Graphics (tog), 38(5):1-12, 2019. 6, 7 +[49] Chengzhi Wu, Junwei Zheng, Julius Pfrommer, and Jürgen Beyerer. Attention-based point cloud edge sampling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5333-5343, 2023. 6 +[50] Xiaoyang Wu, Yixing Lao, Li Jiang, Xihui Liu, and Hengshuang Zhao. Point transformer v2: Grouped vector attention and partition-based pooling. Advances in Neural Information Processing Systems, 35:33330-33342, 2022. 2 +[51] Xiaoyang Wu, Li Jiang, Peng-Shuai Wang, Zhijian Liu, Xihui Liu, Yu Qiao, Wanli Ouyang, Tong He, and Hengshuang Zhao. Point transformer v3: Simpler, faster, stronger. arXiv preprint arXiv:2312.10035, 2023. 2 +[52] Zhirong Wu, Shuran Song, Aditya Khosla, Fisher Yu, Linguang Zhang, Xiaou Tang, and Jianxiong Xiao. 3d shapenets: A deep representation for volumetric shapes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1912-1920, 2015. 6, 7, 8 +[53] Li Yi, Vladimir G Kim, Duygu Ceylan, I-Chao Shen, Mengyan Yan, Hao Su, Cewu Lu, Qixing Huang, Alla Sheffer, and Leonidas Guibas. A scalable active framework for region annotation in 3d shape collections. ACM Transactions on Graphics (ToG), 35(6):1-12, 2016. 6, 7 +[54] Xumin Yu, Lulu Tang, Yongming Rao, Tiejun Huang, Jie Zhou, and Jiwen Lu. Point-bert: Pre-training 3d point cloud transformers with masked point modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19313-19322, 2022. 1, 2, 6, 7 +[55] Renrui Zhang, Ziyu Guo, Peng Gao, Rongyao Fang, Bin Zhao, Dong Wang, Yu Qiao, and Hongsheng Li. Point-m2ae: multi-scale masked autoencoders for hierarchical point cloud pre-training. Advances in neural information processing systems, 35:27061-27074, 2022. 1, 2, 5, 6, 7 + +[56] Renrui Zhang, Liuhui Wang, Yu Qiao, Peng Gao, and Hongsheng Li. Learning 3d representations from 2d pre-trained models via image-to-point masked autoencoders. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21769-21780, 2023. 2, 6, 7 +[57] Tao Zhang, Xiangtai Li, Haobo Yuan, Shunping Ji, and Shuicheng Yan. Point could mamba: Point cloud learning via state space model. arXiv preprint arXiv:2403.00762, 2024.1, 2, 4, 7 +[58] Hengshuang Zhao, Li Jiang, Chi-Wing Fu, and Jiaya Jia. Pointweb: Enhancing local neighborhood features for point cloud processing. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5565-5573, 2019. 2 +[59] Hengshuang Zhao, Li Jiang, Jiaya Jia, Philip HS Torr, and Vladlen Koltun. Point transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 16259-16268, 2021. 2 +[60] Lianghui Zhu, Bencheng Liao, Qian Zhang, Xinlong Wang, Wenyu Liu, and Xinggang Wang. Vision mamba: Efficient visual representation learning with bidirectional state space model. arXiv preprint arXiv:2401.09417, 2024. 1, 2 \ No newline at end of file diff --git a/2025/Spectral Informed Mamba for Robust Point Cloud Processing/images.zip b/2025/Spectral Informed Mamba for Robust Point Cloud Processing/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..c745fa20b2a5758ece865da080dce450096f6ed8 --- /dev/null +++ b/2025/Spectral Informed Mamba for Robust Point Cloud Processing/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0cf790f790c4b7b95730065568ce0b2af0ff00bb109b45043c8e3277938a66d7 +size 512792 diff --git a/2025/Spectral Informed Mamba for Robust Point Cloud Processing/layout.json b/2025/Spectral Informed Mamba for Robust Point Cloud Processing/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..4ff8fd0873a126ed8dabdfd4ebd131b54d4868e1 --- /dev/null +++ b/2025/Spectral Informed Mamba for Robust Point Cloud Processing/layout.json @@ -0,0 +1,8889 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 114, + 103, + 496, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 114, + 103, + 496, + 121 + ], + "spans": [ + { + "bbox": [ + 114, + 103, + 496, + 121 + ], + "type": "text", + "content": "Spectral Informed Mamba for Robust Point Cloud Processing" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 102, + 143, + 508, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 143, + 508, + 185 + ], + "spans": [ + { + "bbox": [ + 102, + 143, + 508, + 185 + ], + "type": "text", + "content": "Ali Bahri * Moslem Yazdanpanah Mehrdad Noori Sahar Dastani \nMilad Cheraghalikhani Gustavo Adolfo Vargas Hakim David Osowiechi Farzad Beizaee Ismail Ben Ayed Christian Desrosiers" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 175, + 191, + 435, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 175, + 191, + 435, + 220 + ], + "spans": [ + { + "bbox": [ + 175, + 191, + 435, + 220 + ], + "type": "text", + "content": "LIVIA, ÉTS Montréal, Canada International Laboratory on Learning Systems (ILLS)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 247, + 200, + 260 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 247, + 200, + 260 + ], + "spans": [ + { + "bbox": [ + 151, + 247, + 200, + 260 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 273, + 296, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 273, + 296, + 536 + ], + "spans": [ + { + "bbox": [ + 55, + 273, + 296, + 536 + ], + "type": "text", + "content": "State Space Models (SSMs) have shown significant promise in Natural Language Processing (NLP) and, more recently, computer vision. This paper introduces a new methodology that leverages Mamba and Masked Autoencoder (MAE) networks for point-cloud data in both supervised and self-supervised learning. We propose three key contributions to enhance Mamba's capability in processing complex pointcloud structures. First, we exploit the spectrum of a graph Laplacian to capture patch connectivity, defining an isometry-invariant traversal order that is robust to viewpoints and captures shape manifolds better than traditional 3D grid-based traversals. Second, we adapt segmentation via a recursive patch partitioning strategy informed by Laplacian spectral components, allowing finer integration and segment analysis. Third, we address token placement in MAE for Mamba by restoring tokens to their original positions, which preserves essential order and improves learning. Extensive experiments demonstrate the improvements that our approach brings over state-of-the-art baselines in classification, segmentation, and few-shot tasks. The implementation is available at: https://github.com/AliBahri94/SI-Mamba.git." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 558, + 135, + 570 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 558, + 135, + 570 + ], + "spans": [ + { + "bbox": [ + 56, + 558, + 135, + 570 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 578, + 296, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 578, + 296, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 578, + 296, + 696 + ], + "type": "text", + "content": "The analysis of 3D point cloud data is fundamental for various applications, including autonomous driving [36, 41], VR/AR [18], and robotics [39]. Compared to the organized structure of 2D images, point clouds consist of 3D coordinates without direct adjacency information forming an unordered bag. In recent years, considerable efforts have been dedicated to adapt deep learning models such as convolutional neural networks (CNNs) and Transformers to this type of data [2, 3, 32, 35, 54, 55]. Due to their permutation-invariant self-attention mechanism, Transformer networks" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 316, + 247, + 555, + 292 + ], + "blocks": [ + { + "bbox": [ + 316, + 247, + 555, + 292 + ], + "lines": [ + { + "bbox": [ + 316, + 247, + 555, + 292 + ], + "spans": [ + { + "bbox": [ + 316, + 247, + 555, + 292 + ], + "type": "image", + "image_path": "c1ffb0f79fd99dd1c2b963811007e840cd3bd8e4e1fb88d7fea676113098dd50.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 300, + 555, + 355 + ], + "lines": [ + { + "bbox": [ + 313, + 300, + 555, + 355 + ], + "spans": [ + { + "bbox": [ + 313, + 300, + 555, + 355 + ], + "type": "text", + "content": "Figure 1. (a) Surface-Aware Spectral Traversing (SAST) applied to the patched point clouds of a mesh surface. (b) Traversal from left to right, based on the first to fourth non-constant smallest eigenvectors. (c) Traversal based on the largest eigenvector, forming a non-continuous sequence of tokens." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 365, + 555, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 365, + 555, + 472 + ], + "spans": [ + { + "bbox": [ + 313, + 365, + 555, + 472 + ], + "type": "text", + "content": "are particularly well suited for the unordered nature of point clouds. However, the quadratic complexity of this mechanism, requiring the computation of a weight between each pair of tokens, impedes the application of these networks to large-sized inputs (e.g., 2D images or 3D point clouds represented by many patches). This has prompted researchers to explore more efficient solutions, including the Set Transformer [24], Sparse Transformer [7], Longformer [4] and Sinkhorn Transformer [43]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 475, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 475, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 475, + 556, + 714 + ], + "type": "text", + "content": "Recently, methods based on Structured State Space Sequence (S4) [16] such as Mamba [14] have gained significant traction as a more efficient alternative to Transformers [29, 60]. So far, very few studies have investigated the potential of S4 approaches such as Mamba for 3D point clouds. Existing methods like Point-Mamba [26] and PCM [57] extend the 2D grid-based traversal used for images to a 3D grid. However, this straightforward adaptation to point clouds suffers from three crucial problems. First: whereas patches from 2D images have adjacency information, which could be exploited by the grid-based traversal, the 3D point patches in point clouds offer a sparse representation of the object's surface, and nearby patches on a 3D grid are not necessarily adjacent on this surface. Second: in the absence of self-attention, task-specific performance is highly influenced by the nature of the token traversal strategy. For example, a traversal suitable for point cloud classification may not be effective for a local task such as point-level classification (i.e., segmentation). Third: due to the \"direction-sensitive\" nature of Mamba, the self-supervised MAE pre" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 494, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 703, + 216, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 216, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 216, + 712 + ], + "type": "text", + "content": "*Correspondence to ali.bahri.1@ens.etsmtl.ca" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11799" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 119 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 119 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 119 + ], + "type": "text", + "content": "training step of leading point cloud models like Point-MAE [32] and Point-M2AE [55] cannot be used directly as there is no attention mechanism to learn the masked tokens' positions." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 121, + 295, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 121, + 295, + 144 + ], + "spans": [ + { + "bbox": [ + 55, + 121, + 295, + 144 + ], + "type": "text", + "content": "The contribution of our work focuses on addressing these problems as follows:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 146, + 295, + 387 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 56, + 146, + 295, + 229 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 146, + 295, + 229 + ], + "spans": [ + { + "bbox": [ + 56, + 146, + 295, + 229 + ], + "type": "text", + "content": "1. We introduce a Surface-Aware Spectral Traversing (SAST) strategy based on the Laplacian spectrum of a patch-connectivity graph. Compared to the 3D grid traversal of current approaches like Point-Mamba, our strategy is invariant to isometric transformations (e.g., choice of viewpoint) and better captures the object's surface manifold." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 231, + 294, + 315 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 231, + 294, + 315 + ], + "spans": [ + { + "bbox": [ + 55, + 231, + 294, + 315 + ], + "type": "text", + "content": "2. We also present a Hierarchical Local Traversing (HLT) for point-level classification (segmentation) that partitions patches recursively based on their spectral coordinates. Unlike our SAST strategy for classification, which considers Laplacian eigenvectors separately in different traversals, this HLT combines them in a single ordering for a more precise modeling of geometry." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 316, + 294, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 316, + 294, + 387 + ], + "spans": [ + { + "bbox": [ + 55, + 316, + 294, + 387 + ], + "type": "text", + "content": "3. During the MAE-based Self-Supervised Learning (SSL), we propose a Traverse-Aware Repositioning (TAR) strategy to align the masked tokens according to their spectral adjacency. This strategy addresses the critical issue of spatial adjacency preservation unique to Mamba networks." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 401, + 146, + 414 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 401, + 146, + 414 + ], + "spans": [ + { + "bbox": [ + 55, + 401, + 146, + 414 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 425, + 295, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 425, + 295, + 544 + ], + "spans": [ + { + "bbox": [ + 55, + 425, + 295, + 544 + ], + "type": "text", + "content": "Deep Point Cloud Learning. Deep neural networks have increasingly been applied to point clouds, with early architectures like PointNet [34] and PointNet++ [35] inspiring further work [1, 10, 23, 25, 58] focused on capturing local context. Transformer-based frameworks [46], including the Point Transformer (v1-v3) [50, 51, 59] and Stratified Transformer [22], have set new benchmarks by integrating both local and global features. Recently, self-supervised pre-training has gained traction by leveraging masked point modeling [28, 42] to train Transformers on unlabeled data." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "text", + "content": "Building on the effectiveness of MAE in Text and Image domains, Point-BERT [54] presented a revolutionary method inspired by BERT [11], tailoring Transformers to 3D point cloud processing. Point-MAE [32] applied MAE-style pre-training to 3D point clouds using a custom Transformer-based Autoencoder (AE) designed to reconstruct masked irregular patches. The use of multiscale masking and local spatial self-attention mechanisms in Point-M2AE [55] has led to State-Of-The-Art (SOTA) results in 3D representation learning. Furthermore, I2P-MAE [56] improved self-supervised point cloud processing with a masking strategy leveraging pre-trained 2D models through an Image-to-Point transformation. Point-GPT [6] introduced an auto-regressive generative pretraining (GPT)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 553, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 168 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 168 + ], + "type": "text", + "content": "approach to address the unordered nature and low information density of point clouds. ACT [12] proposed a cross-modal knowledge transfer method using pretrained 2D or natural language Transformers as teachers for 3D representation learning. Finally, GeoMask3D [2] is a self-supervised point cloud method that uses a teacher-student model to select harder, geometrically intricate regions for masking, thereby improving feature representations." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 175, + 427, + 188 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 175, + 427, + 188 + ], + "spans": [ + { + "bbox": [ + 313, + 175, + 427, + 188 + ], + "type": "text", + "content": "2.1. State Space Models" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 194, + 553, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 194, + 553, + 410 + ], + "spans": [ + { + "bbox": [ + 313, + 194, + 553, + 410 + ], + "type": "text", + "content": "State Space Models. State-space models (SSMs) have been foundational in control theory and signal processing for modeling dynamic systems. Recently, they have been adapted to deep learning, with advancements like the Linear State-Space Layer (LSSL), which uses a continuous-time memorization framework based on the HiPPO operator [15] to model long-range dependencies. However, LSSL's high computational demands limit its practicality. To address this, the S4 model [16] introduced parameter normalization techniques, paving the way for structured SSMs with enhancements like complex-diagonal structures [17, 19], support for multiple-input/output [17], and low-rank decomposition [20]. Recent developments include Mamba [14], which achieves linear-time inference and efficient training through selection mechanisms and hardware-aware algorithms. MoE-Mamba further boosts efficiency by integrating a Mixture of Experts (MoE), outperforming both Mamba and Transformer-MoE models [33]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 412, + 553, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 412, + 553, + 532 + ], + "spans": [ + { + "bbox": [ + 313, + 412, + 553, + 532 + ], + "type": "text", + "content": "SSMs for Vision Tasks The above-mentioned works primarily focused on the application of SSMs to long-range or causal data types such as language and speech. In the field of vision, a notable study [29] proposed the VMamba model which features a Cross-Scan Module (CSM) for enhanced 1D selective scanning in 2D spaces and architectural optimizations that significantly improve its performance and speed across various visual tasks. Another significant paper is Vision Mamba [60] which introduces a novel vision backbone called Vim utilizing bidirectional Mamba blocks." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 533, + 553, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 553, + 664 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 553, + 664 + ], + "type": "text", + "content": "For point cloud analysis based on Mamba, two key works are Point-Mamba [26] and PCM [57]. PointMamba introduces a simple approach to token reordering for point cloud analysis by strategically organizing point tokens based on a 3D grid. Similarly, PCM enhances Mamba with a Consistent Traverse Serialization (CTS) technique that converts 3D point clouds into 1D point sequences while maintaining spatial adjacency. Building upon these approaches, our method introduces the Spectral Spatial Traversing (SST) strategy, which improves token ordering and maintains spatial adjacency within Mamba networks." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 674, + 369, + 686 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 674, + 369, + 686 + ], + "spans": [ + { + "bbox": [ + 313, + 674, + 369, + 686 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "content": "We begin by outlining the fundamental concepts of SSMs and spectral graph analysis which are at the core of our" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11800" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 104, + 70, + 512, + 411 + ], + "blocks": [ + { + "bbox": [ + 104, + 70, + 512, + 411 + ], + "lines": [ + { + "bbox": [ + 104, + 70, + 512, + 411 + ], + "spans": [ + { + "bbox": [ + 104, + 70, + 512, + 411 + ], + "type": "image", + "image_path": "96e7d1d4fec27302183146eb7b4f1db41038e8bd4dc37c256de2af823c5b5507.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 421, + 555, + 489 + ], + "lines": [ + { + "bbox": [ + 54, + 421, + 555, + 489 + ], + "spans": [ + { + "bbox": [ + 54, + 421, + 555, + 489 + ], + "type": "text", + "content": "Figure 2. Overview of the proposed Spectral Spatial Traversing (SST) method. (a) Point cloud, (b) Patchification, (c) Forming the adjacency graph, (d) Traversal based on SAST using " + }, + { + "bbox": [ + 54, + 421, + 555, + 489 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 54, + 421, + 555, + 489 + ], + "type": "text", + "content": " non-constant smallest eigenvectors, (e) HLT for segmentation tasks, (f) TAR strategy for Masked Autoencoders. The process includes reverse and concatenation operations, with learnable tokens, representations, and masked tokens highlighted. (g) The classification task involves sorting tokens by different eigenvectors, concatenating them, and then feeding them into the network. (h) The segmentation task where HLT is applied on the tokens " + }, + { + "bbox": [ + 54, + 421, + 555, + 489 + ], + "type": "inline_equation", + "content": "(\\vec{q})" + }, + { + "bbox": [ + 54, + 421, + 555, + 489 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 421, + 555, + 489 + ], + "type": "inline_equation", + "content": "\\vec{q}" + }, + { + "bbox": [ + 54, + 421, + 555, + 489 + ], + "type": "text", + "content": " is fed into the network. (i) A flowchart visualizing the techniques used in self-supervised learning and various downstream tasks." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 504, + 297, + 660 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 504, + 297, + 660 + ], + "spans": [ + { + "bbox": [ + 55, + 504, + 297, + 660 + ], + "type": "text", + "content": "work. We then give an in-depth presentation of our Surface-Aware Spectral Traversing (SAST) strategy for point cloud processing that improves the model's robustness to isometric transformations and better captures the underlying manifold of the point cloud. Thereafter, we provide detailed specifications of our Hierarchical Local Traversing (HLT) strategy for point-level classification, which defines a more structured patch traversal order based on the recursive partitioning of spectral information. Finally, we introduce our Traverse-Aware Repositioning (TAR), which improves the handling of learnable tokens in masked autoencoders within Mamba networks. Fig. 2 illustrates the overview of the proposed Spectral Spatial Traversing (SST) method." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 669, + 142, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 669, + 142, + 681 + ], + "spans": [ + { + "bbox": [ + 55, + 669, + 142, + 681 + ], + "type": "text", + "content": "3.1. Preliminaries" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": "State Space Models (SSMs) use a series of first-order differential equations to describe how the state of the linear," + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 314, + 504, + 479, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 504, + 479, + 516 + ], + "spans": [ + { + "bbox": [ + 314, + 504, + 479, + 516 + ], + "type": "text", + "content": "time-invariant system evolves over time:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 323, + 524, + 555, + 540 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 524, + 555, + 540 + ], + "spans": [ + { + "bbox": [ + 323, + 524, + 555, + 540 + ], + "type": "interline_equation", + "content": "\\dot {h} (t) = A h (t) + B x (t), \\quad y (t) = C \\dot {h} (t) + D x (t), \\quad (1)", + "image_path": "d87da4b0bac57c4a599cd9b70d547526fe2a3ea0b1cc455b13a060d62eebad51.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 548, + 555, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 548, + 555, + 583 + ], + "spans": [ + { + "bbox": [ + 313, + 548, + 555, + 583 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 313, + 548, + 555, + 583 + ], + "type": "inline_equation", + "content": "\\dot{h}(t)" + }, + { + "bbox": [ + 313, + 548, + 555, + 583 + ], + "type": "text", + "content": " denotes the time derivative of the state vector " + }, + { + "bbox": [ + 313, + 548, + 555, + 583 + ], + "type": "inline_equation", + "content": "h(t)" + }, + { + "bbox": [ + 313, + 548, + 555, + 583 + ], + "type": "text", + "content": ". The matrices " + }, + { + "bbox": [ + 313, + 548, + 555, + 583 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 313, + 548, + 555, + 583 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 548, + 555, + 583 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 313, + 548, + 555, + 583 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 548, + 555, + 583 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 313, + 548, + 555, + 583 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 548, + 555, + 583 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 313, + 548, + 555, + 583 + ], + "type": "text", + "content": " are the weighting parameters." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 585, + 555, + 633 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 585, + 555, + 633 + ], + "spans": [ + { + "bbox": [ + 313, + 585, + 555, + 633 + ], + "type": "text", + "content": "Due to their reliance on continuous data streams " + }, + { + "bbox": [ + 313, + 585, + 555, + 633 + ], + "type": "inline_equation", + "content": "x(t)" + }, + { + "bbox": [ + 313, + 585, + 555, + 633 + ], + "type": "text", + "content": ", SSMs are not natively equipped to handle discrete inputs represented as " + }, + { + "bbox": [ + 313, + 585, + 555, + 633 + ], + "type": "inline_equation", + "content": "\\{x_0, x_1, \\ldots\\}" + }, + { + "bbox": [ + 313, + 585, + 555, + 633 + ], + "type": "text", + "content": ". This necessitates the use of a discretized SSM version for practical applications:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 345, + 642, + 555, + 656 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 642, + 555, + 656 + ], + "spans": [ + { + "bbox": [ + 345, + 642, + 555, + 656 + ], + "type": "interline_equation", + "content": "h _ {k} = \\bar {A} h _ {k - 1} + \\bar {B} x _ {k}, \\quad y _ {k} = \\bar {C} h _ {k} + \\bar {D} x _ {k}. \\tag {2}", + "image_path": "7ac8b50f38df25a988e1e9febd4f4370ff9096b524d451aa53f7a2c9603b4426.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 666, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 666, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 666, + 556, + 715 + ], + "type": "text", + "content": "The state space model in its discrete version utilizes a recursive function to link each state " + }, + { + "bbox": [ + 313, + 666, + 556, + 715 + ], + "type": "inline_equation", + "content": "h_k" + }, + { + "bbox": [ + 313, + 666, + 556, + 715 + ], + "type": "text", + "content": " to its preceding state, encapsulated by the matrices " + }, + { + "bbox": [ + 313, + 666, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\bar{A} \\in \\mathbb{R}^{N \\times N}" + }, + { + "bbox": [ + 313, + 666, + 556, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 666, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\bar{B} \\in \\mathbb{R}^{N \\times 1}" + }, + { + "bbox": [ + 313, + 666, + 556, + 715 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 666, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\bar{C} \\in \\mathbb{R}^{N \\times 1}" + }, + { + "bbox": [ + 313, + 666, + 556, + 715 + ], + "type": "text", + "content": ", which are tuned parameter matrices. While" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11801" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 70, + 293, + 187 + ], + "blocks": [ + { + "bbox": [ + 60, + 70, + 293, + 187 + ], + "lines": [ + { + "bbox": [ + 60, + 70, + 293, + 187 + ], + "spans": [ + { + "bbox": [ + 60, + 70, + 293, + 187 + ], + "type": "image", + "image_path": "c3d67f5050addff6fa87af5f531b71eb6f0618300350fd11ef6b1ee75ca2beb9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 196, + 296, + 251 + ], + "lines": [ + { + "bbox": [ + 55, + 196, + 296, + 251 + ], + "spans": [ + { + "bbox": [ + 55, + 196, + 296, + 251 + ], + "type": "text", + "content": "Figure 3. (a) Visualization of the four non-constant smallest Laplacian eigenvectors " + }, + { + "bbox": [ + 55, + 196, + 296, + 251 + ], + "type": "inline_equation", + "content": "(v^{(k)}, k = 1, \\ldots, 4)" + }, + { + "bbox": [ + 55, + 196, + 296, + 251 + ], + "type": "text", + "content": " and (b) the discrete partitioning " + }, + { + "bbox": [ + 55, + 196, + 296, + 251 + ], + "type": "inline_equation", + "content": "(q)" + }, + { + "bbox": [ + 55, + 196, + 296, + 251 + ], + "type": "text", + "content": " of our HLT strategy combining the information of all four eigenvectors. Note: we assumed that patches contain a single point for better visualization." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 256, + 296, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 256, + 296, + 342 + ], + "spans": [ + { + "bbox": [ + 55, + 256, + 296, + 342 + ], + "type": "text", + "content": "matrix " + }, + { + "bbox": [ + 55, + 256, + 296, + 342 + ], + "type": "inline_equation", + "content": "\\bar{D} \\in \\mathbb{R}^{N \\times 1}" + }, + { + "bbox": [ + 55, + 256, + 296, + 342 + ], + "type": "text", + "content": " may be employed as a residual connection, we follow previous work and exclude it from our model. The transition from a continuous signal representation " + }, + { + "bbox": [ + 55, + 256, + 296, + 342 + ], + "type": "inline_equation", + "content": "x(t)" + }, + { + "bbox": [ + 55, + 256, + 296, + 342 + ], + "type": "text", + "content": " to a discrete sequence involves sampling " + }, + { + "bbox": [ + 55, + 256, + 296, + 342 + ], + "type": "inline_equation", + "content": "x(t)" + }, + { + "bbox": [ + 55, + 256, + 296, + 342 + ], + "type": "text", + "content": " at intervals defined by " + }, + { + "bbox": [ + 55, + 256, + 296, + 342 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 55, + 256, + 296, + 342 + ], + "type": "text", + "content": ", setting each discrete input as " + }, + { + "bbox": [ + 55, + 256, + 296, + 342 + ], + "type": "inline_equation", + "content": "x_{k} = x(k\\Delta)" + }, + { + "bbox": [ + 55, + 256, + 296, + 342 + ], + "type": "text", + "content": ". This adjustment to a discrete framework results in revised matrix definitions:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 346, + 296, + 378 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 346, + 296, + 378 + ], + "spans": [ + { + "bbox": [ + 55, + 346, + 296, + 378 + ], + "type": "interline_equation", + "content": "\\bar {A} = \\left(I - \\frac {\\Delta}{2} A\\right) ^ {- 1} \\left(I + \\frac {\\Delta}{2} A\\right), \\bar {B} = \\left(I - \\frac {\\Delta}{2} A\\right) ^ {- 1} \\Delta B, \\bar {C} = C. \\tag {3}", + "image_path": "d36f95d5816fe1b9e74d6524eaed1af9ff3fd952c919d74b3b246cdb0d105519.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 380, + 296, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 380, + 296, + 475 + ], + "spans": [ + { + "bbox": [ + 55, + 380, + 296, + 475 + ], + "type": "text", + "content": "However, the fixed dynamics of Linear Time-Invariant (LTI) models, exemplified by the constant parameters " + }, + { + "bbox": [ + 55, + 380, + 296, + 475 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 55, + 380, + 296, + 475 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 380, + 296, + 475 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 55, + 380, + 296, + 475 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 380, + 296, + 475 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 55, + 380, + 296, + 475 + ], + "type": "text", + "content": " in Eq. (3), restrict their capacity to selectively retain or discard relevant information, thereby limiting their contextual awareness. To improve content-aware reasoning, we use Mamba's selection mechanism that manages the propagation and interaction of information across the sequence dimension [14]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 480, + 296, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 480, + 296, + 564 + ], + "spans": [ + { + "bbox": [ + 55, + 480, + 296, + 564 + ], + "type": "text", + "content": "Spectral Graph Analysis. Popularized by Chung in the 90s [8], spectral graph analysis characterizes the properties of a graph " + }, + { + "bbox": [ + 55, + 480, + 296, + 564 + ], + "type": "inline_equation", + "content": "G = (V, E)" + }, + { + "bbox": [ + 55, + 480, + 296, + 564 + ], + "type": "text", + "content": " by the spectrum (eigenvalues and corresponding eigenvectors) of its Laplacian matrix " + }, + { + "bbox": [ + 55, + 480, + 296, + 564 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 55, + 480, + 296, + 564 + ], + "type": "text", + "content": ". This analysis can be understood as a discretized version of the Laplace-Beltrami Operator " + }, + { + "bbox": [ + 55, + 480, + 296, + 564 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 55, + 480, + 296, + 564 + ], + "type": "text", + "content": " of a function " + }, + { + "bbox": [ + 55, + 480, + 296, + 564 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 55, + 480, + 296, + 564 + ], + "type": "text", + "content": " defined on a Riemannian manifold:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 136, + 572, + 295, + 586 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 136, + 572, + 295, + 586 + ], + "spans": [ + { + "bbox": [ + 136, + 572, + 295, + 586 + ], + "type": "interline_equation", + "content": "\\Delta f = \\operatorname {d i v} (\\operatorname {g r a d} f) \\tag {4}", + "image_path": "96647eebdc5ebcce174f5ed748886fb2e2288004d239480eb7ba4278414f287d.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 594, + 295, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 653 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 653 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 594, + 295, + 653 + ], + "type": "inline_equation", + "content": "\\operatorname{grad}f" + }, + { + "bbox": [ + 55, + 594, + 295, + 653 + ], + "type": "text", + "content": " is the gradient of " + }, + { + "bbox": [ + 55, + 594, + 295, + 653 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 55, + 594, + 295, + 653 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 594, + 295, + 653 + ], + "type": "inline_equation", + "content": "\\operatorname{div}" + }, + { + "bbox": [ + 55, + 594, + 295, + 653 + ], + "type": "text", + "content": " the divergence on the manifold. The solution to the Laplacian eigenvalue problem " + }, + { + "bbox": [ + 55, + 594, + 295, + 653 + ], + "type": "inline_equation", + "content": "\\Delta f = -\\lambda f" + }, + { + "bbox": [ + 55, + 594, + 295, + 653 + ], + "type": "text", + "content": ", known as Helmholtz wave equation, is an eigenfunction corresponding to the natural vibration form of a homogeneous membrane with eigenvalue " + }, + { + "bbox": [ + 55, + 594, + 295, + 653 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 594, + 295, + 653 + ], + "type": "text", + "content": " [38]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "content": "Following methods for spectral clustering [31] and normalized cuts [40], we consider a weighted adjacency matrix " + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "inline_equation", + "content": "W: V \\times V \\to \\mathbb{R}_+" + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "inline_equation", + "content": "W_{ij} = 0" + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "content": " if " + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "inline_equation", + "content": "(i,j) \\notin E" + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "content": " to model the Euclidean distance of nearby patches (see Section 3.3). The Laplacian matrix of " + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "content": " is defined as " + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "inline_equation", + "content": "L = D - W" + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "content": " where" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": " is the diagonal degree matrix such that " + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "inline_equation", + "content": "D_{ii} = \\sum_{j}W_{ij}" + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": ". To account for variability in the scale of weights " + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "inline_equation", + "content": "W_{ij}" + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": " or the distribution of node degrees " + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "inline_equation", + "content": "D_{ii}" + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": ", it is preferable to employ a normalized version of the Laplacian. In this work, we use the Random Walk Laplacian " + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "inline_equation", + "content": "L_{rw} = I - D^{-1}W" + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": " which has the following useful properties:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 144, + 553, + 263 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 314, + 144, + 553, + 169 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 144, + 553, + 169 + ], + "spans": [ + { + "bbox": [ + 314, + 144, + 553, + 169 + ], + "type": "text", + "content": "1. " + }, + { + "bbox": [ + 314, + 144, + 553, + 169 + ], + "type": "inline_equation", + "content": "L_{rw}" + }, + { + "bbox": [ + 314, + 144, + 553, + 169 + ], + "type": "text", + "content": " is positive semi-definite and has " + }, + { + "bbox": [ + 314, + 144, + 553, + 169 + ], + "type": "inline_equation", + "content": "|V|" + }, + { + "bbox": [ + 314, + 144, + 553, + 169 + ], + "type": "text", + "content": " non-negative real-valued eigenvalues " + }, + { + "bbox": [ + 314, + 144, + 553, + 169 + ], + "type": "inline_equation", + "content": "0 = \\lambda_1 \\leq \\ldots \\leq \\lambda_{|V|}" + }, + { + "bbox": [ + 314, + 144, + 553, + 169 + ], + "type": "text", + "content": ";" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 169, + 553, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 169, + 553, + 204 + ], + "spans": [ + { + "bbox": [ + 313, + 169, + 553, + 204 + ], + "type": "text", + "content": "2. 0 is an eigenvalue of " + }, + { + "bbox": [ + 313, + 169, + 553, + 204 + ], + "type": "inline_equation", + "content": "L_{rw}" + }, + { + "bbox": [ + 313, + 169, + 553, + 204 + ], + "type": "text", + "content": " with the constant vector as eigenvector and its multiplicity equals the number of connected components in the graph;" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 205, + 553, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 205, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 313, + 205, + 553, + 228 + ], + "type": "text", + "content": "3. Following Courant's Nodal Line Theorem [9], the " + }, + { + "bbox": [ + 313, + 205, + 553, + 228 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 205, + 553, + 228 + ], + "type": "text", + "content": "-th eigenmode of " + }, + { + "bbox": [ + 313, + 205, + 553, + 228 + ], + "type": "inline_equation", + "content": "L_{rw}" + }, + { + "bbox": [ + 313, + 205, + 553, + 228 + ], + "type": "text", + "content": " has at most " + }, + { + "bbox": [ + 313, + 205, + 553, + 228 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 205, + 553, + 228 + ], + "type": "text", + "content": " poles of vibration;" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 228, + 553, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 228, + 553, + 263 + ], + "spans": [ + { + "bbox": [ + 313, + 228, + 553, + 263 + ], + "type": "text", + "content": "4. The representation of a shape by the spectrum of " + }, + { + "bbox": [ + 313, + 228, + 553, + 263 + ], + "type": "inline_equation", + "content": "L_{rw}" + }, + { + "bbox": [ + 313, + 228, + 553, + 263 + ], + "type": "text", + "content": " is invariant to isometry (i.e., distance-preserving transformation)." + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 265, + 555, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 265, + 555, + 433 + ], + "spans": [ + { + "bbox": [ + 313, + 265, + 555, + 433 + ], + "type": "text", + "content": "Our method uses the " + }, + { + "bbox": [ + 313, + 265, + 555, + 433 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 313, + 265, + 555, + 433 + ], + "type": "text", + "content": " first non-constant eigenvectors of " + }, + { + "bbox": [ + 313, + 265, + 555, + 433 + ], + "type": "inline_equation", + "content": "L_{rw}" + }, + { + "bbox": [ + 313, + 265, + 555, + 433 + ], + "type": "text", + "content": " (i.e., the eigenvectors corresponding to the " + }, + { + "bbox": [ + 313, + 265, + 555, + 433 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 313, + 265, + 555, + 433 + ], + "type": "text", + "content": " smallest non-zero eigenvalues) to define traversal orders for classification (Section 3.3) and segmentation (Section 3.4) that are robust to the viewpoint (due to isometry invariance) and provide a smooth parametrization of the surface manifold. We consider the first eigenvectors as they encode low frequency information (by Courant's Nodal Line Theorem), making the resulting traversal more robust to shape variability and noise. Fig. 1 illustrates this concept: (a) shows the original mesh, (b) shows traversals based on the first to fourth non-constant smallest eigenvectors, and (c) shows traversal based on the largest eigenvector forming a non-continuous sequence of tokens." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 314, + 441, + 460, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 441, + 460, + 453 + ], + "spans": [ + { + "bbox": [ + 314, + 441, + 460, + 453 + ], + "type": "text", + "content": "3.2. Point Cloud Patchification" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 457, + 555, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 457, + 555, + 591 + ], + "spans": [ + { + "bbox": [ + 313, + 457, + 555, + 591 + ], + "type": "text", + "content": "Given a point cloud " + }, + { + "bbox": [ + 313, + 457, + 555, + 591 + ], + "type": "inline_equation", + "content": "\\mathcal{P} = \\{p_i\\}_{i=1}^{N_p}" + }, + { + "bbox": [ + 313, + 457, + 555, + 591 + ], + "type": "text", + "content": ", each point represented by 3D coordinates, we convert " + }, + { + "bbox": [ + 313, + 457, + 555, + 591 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 313, + 457, + 555, + 591 + ], + "type": "text", + "content": " to a reduced set of patches that can be processed more efficiently. Toward this goal, we employ the Farthest Point Sampling (FPS) algorithm to select a subset " + }, + { + "bbox": [ + 313, + 457, + 555, + 591 + ], + "type": "inline_equation", + "content": "\\mathcal{C} \\subset \\mathcal{P}" + }, + { + "bbox": [ + 313, + 457, + 555, + 591 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 313, + 457, + 555, + 591 + ], + "type": "inline_equation", + "content": "N_c" + }, + { + "bbox": [ + 313, + 457, + 555, + 591 + ], + "type": "text", + "content": " points offering a good coverage of the entire point cloud. These selected points will act as the centers of local patches within the point cloud. For each center point " + }, + { + "bbox": [ + 313, + 457, + 555, + 591 + ], + "type": "inline_equation", + "content": "p_{s_i} \\in \\mathcal{C}" + }, + { + "bbox": [ + 313, + 457, + 555, + 591 + ], + "type": "text", + "content": ", we then identify " + }, + { + "bbox": [ + 313, + 457, + 555, + 591 + ], + "type": "inline_equation", + "content": "N_n" + }, + { + "bbox": [ + 313, + 457, + 555, + 591 + ], + "type": "text", + "content": " nearest points " + }, + { + "bbox": [ + 313, + 457, + 555, + 591 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(p_{s_i}) \\subset \\mathcal{P}" + }, + { + "bbox": [ + 313, + 457, + 555, + 591 + ], + "type": "text", + "content": " using the K-Nearest Neighbours (KNN) algorithm. Following this, each patch is defined as a center " + }, + { + "bbox": [ + 313, + 457, + 555, + 591 + ], + "type": "inline_equation", + "content": "p_{s_i}" + }, + { + "bbox": [ + 313, + 457, + 555, + 591 + ], + "type": "text", + "content": " and its corresponding nearest-neighbors " + }, + { + "bbox": [ + 313, + 457, + 555, + 591 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(p_{s_i})" + }, + { + "bbox": [ + 313, + 457, + 555, + 591 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 599, + 539, + 613 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 599, + 539, + 613 + ], + "spans": [ + { + "bbox": [ + 313, + 599, + 539, + 613 + ], + "type": "text", + "content": "3.3. Surface-Aware Spectral Traversing (SAST)" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "type": "text", + "content": "Current point cloud processing approaches using Mamba, such as Point-Mamba [26] and PCM [57], simply extend the 2D grid-based traversal for images to a 3D grid. As mentioned before, this naive strategy suffers from two issues: 1) the 3D grid is view dependent, thus rotating the point cloud or moving the camera yields a different traversal order; 2) unrelated patches may be adjacent in 3D space, hence can be traversed subsequently. To address these problems, we" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11802" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 96 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 96 + ], + "type": "text", + "content": "define a traversal order based on the Laplacian spectrum of the patch-connectivity graph." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 97, + 296, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 97, + 296, + 193 + ], + "spans": [ + { + "bbox": [ + 55, + 97, + 296, + 193 + ], + "type": "text", + "content": "In this graph, each node corresponds to a patch and the weighted adjacency matrix " + }, + { + "bbox": [ + 55, + 97, + 296, + 193 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 55, + 97, + 296, + 193 + ], + "type": "text", + "content": " is defined using the Euclidean distance between patch centers. For patches " + }, + { + "bbox": [ + 55, + 97, + 296, + 193 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 97, + 296, + 193 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 97, + 296, + 193 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 55, + 97, + 296, + 193 + ], + "type": "text", + "content": " defined by center points " + }, + { + "bbox": [ + 55, + 97, + 296, + 193 + ], + "type": "inline_equation", + "content": "p_{s_i}" + }, + { + "bbox": [ + 55, + 97, + 296, + 193 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 97, + 296, + 193 + ], + "type": "inline_equation", + "content": "p_{s_j}" + }, + { + "bbox": [ + 55, + 97, + 296, + 193 + ], + "type": "text", + "content": ", we add an edge " + }, + { + "bbox": [ + 55, + 97, + 296, + 193 + ], + "type": "inline_equation", + "content": "(i,j)" + }, + { + "bbox": [ + 55, + 97, + 296, + 193 + ], + "type": "text", + "content": " if " + }, + { + "bbox": [ + 55, + 97, + 296, + 193 + ], + "type": "inline_equation", + "content": "p_{s_j}" + }, + { + "bbox": [ + 55, + 97, + 296, + 193 + ], + "type": "text", + "content": " is among the " + }, + { + "bbox": [ + 55, + 97, + 296, + 193 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 97, + 296, + 193 + ], + "type": "text", + "content": " nearest neighbors of " + }, + { + "bbox": [ + 55, + 97, + 296, + 193 + ], + "type": "inline_equation", + "content": "p_{s_i}" + }, + { + "bbox": [ + 55, + 97, + 296, + 193 + ], + "type": "text", + "content": " or vice-versa. The weight of this edge is computed using a Gaussian kernel: " + }, + { + "bbox": [ + 55, + 97, + 296, + 193 + ], + "type": "inline_equation", + "content": "W_{ij} = \\exp \\left(-\\| p_{s_i} - p_{s_j}\\| _2^2 /\\sigma\\right)" + }, + { + "bbox": [ + 55, + 97, + 296, + 193 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 55, + 97, + 296, + 193 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 55, + 97, + 296, + 193 + ], + "type": "text", + "content": " is a hyperparameter controlling the kernel width." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 194, + 296, + 368 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 194, + 296, + 368 + ], + "spans": [ + { + "bbox": [ + 55, + 194, + 296, + 368 + ], + "type": "text", + "content": "Following Section 3.1, we compute the " + }, + { + "bbox": [ + 55, + 194, + 296, + 368 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 55, + 194, + 296, + 368 + ], + "type": "text", + "content": " first nonconstant eigenvectors of the Random Walk Laplacian " + }, + { + "bbox": [ + 55, + 194, + 296, + 368 + ], + "type": "inline_equation", + "content": "L_{rw}" + }, + { + "bbox": [ + 55, + 194, + 296, + 368 + ], + "type": "text", + "content": ". This can be achieved efficiently using an iterative method like the Arnoldi algorithm [13] by exploiting the following facts: 1) matrix " + }, + { + "bbox": [ + 55, + 194, + 296, + 368 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 55, + 194, + 296, + 368 + ], + "type": "text", + "content": " is very sparse, and 2) only the first few eigenvectors need to be computed. Eigenvector " + }, + { + "bbox": [ + 55, + 194, + 296, + 368 + ], + "type": "inline_equation", + "content": "v^{(k)} \\in \\mathbb{R}^{N_c}" + }, + { + "bbox": [ + 55, + 194, + 296, + 368 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 194, + 296, + 368 + ], + "type": "inline_equation", + "content": "k \\in \\{1, \\dots, s\\}" + }, + { + "bbox": [ + 55, + 194, + 296, + 368 + ], + "type": "text", + "content": ", assigns an eigenfunction value " + }, + { + "bbox": [ + 55, + 194, + 296, + 368 + ], + "type": "inline_equation", + "content": "v_i^{(k)}" + }, + { + "bbox": [ + 55, + 194, + 296, + 368 + ], + "type": "text", + "content": " to each patch " + }, + { + "bbox": [ + 55, + 194, + 296, + 368 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 194, + 296, + 368 + ], + "type": "text", + "content": ". In each Mamba block of our model, we perform two separate traversals of tokens (each token corresponds to an input patch) for every eigenvector: a forward traversal by increasing value of " + }, + { + "bbox": [ + 55, + 194, + 296, + 368 + ], + "type": "inline_equation", + "content": "v_i^{(k)}" + }, + { + "bbox": [ + 55, + 194, + 296, + 368 + ], + "type": "text", + "content": " and a reserve traversal by decreasing value of " + }, + { + "bbox": [ + 55, + 194, + 296, + 368 + ], + "type": "inline_equation", + "content": "v_i^{(k)}" + }, + { + "bbox": [ + 55, + 194, + 296, + 368 + ], + "type": "text", + "content": ". At the end of the block, we concatenate the tokens from the " + }, + { + "bbox": [ + 55, + 194, + 296, + 368 + ], + "type": "inline_equation", + "content": "s \\times 2" + }, + { + "bbox": [ + 55, + 194, + 296, + 368 + ], + "type": "text", + "content": " traversals. The visual effect of each traversal is illustrated in Fig. 1(b) and Fig. 3(a)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 369, + 296, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 369, + 296, + 578 + ], + "spans": [ + { + "bbox": [ + 55, + 369, + 296, + 578 + ], + "type": "text", + "content": "Canonicalization of spectrum. Although the spectrum of " + }, + { + "bbox": [ + 55, + 369, + 296, + 578 + ], + "type": "inline_equation", + "content": "L_{rw}" + }, + { + "bbox": [ + 55, + 369, + 296, + 578 + ], + "type": "text", + "content": " forms an isometry-invariant representation of the surface manifold, this representation may be impacted by two sources of ambiguity: 1) the sign of eigenvectors is undetermined (i.e., if " + }, + { + "bbox": [ + 55, + 369, + 296, + 578 + ], + "type": "inline_equation", + "content": "v^{(k)}" + }, + { + "bbox": [ + 55, + 369, + 296, + 578 + ], + "type": "text", + "content": " is an eigenvector, then so is " + }, + { + "bbox": [ + 55, + 369, + 296, + 578 + ], + "type": "inline_equation", + "content": "-v^{(k)}" + }, + { + "bbox": [ + 55, + 369, + 296, + 578 + ], + "type": "text", + "content": "), and 2) the order of eigenvectors with similar eigenvalues may vary. We address these two sources of ambiguity with the following canonicalization procedure. For the first one, we flip the sign of an eigenvector " + }, + { + "bbox": [ + 55, + 369, + 296, + 578 + ], + "type": "inline_equation", + "content": "v^{(k)}" + }, + { + "bbox": [ + 55, + 369, + 296, + 578 + ], + "type": "text", + "content": " (i.e., " + }, + { + "bbox": [ + 55, + 369, + 296, + 578 + ], + "type": "inline_equation", + "content": "v^{(k)} \\coloneqq -v^{(k)}" + }, + { + "bbox": [ + 55, + 369, + 296, + 578 + ], + "type": "text", + "content": ") if its first element is negative (i.e., " + }, + { + "bbox": [ + 55, + 369, + 296, + 578 + ], + "type": "inline_equation", + "content": "v_{1}^{(k)} < 0" + }, + { + "bbox": [ + 55, + 369, + 296, + 578 + ], + "type": "text", + "content": "). To handle the second ambiguity, we first sort the eigenvectors by nondecreasing eigenvalue. We deal with eigenvalues having a multiplicity greater than one by finding pairs of consecutive eigenvectors " + }, + { + "bbox": [ + 55, + 369, + 296, + 578 + ], + "type": "inline_equation", + "content": "v^{(k)}" + }, + { + "bbox": [ + 55, + 369, + 296, + 578 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 369, + 296, + 578 + ], + "type": "inline_equation", + "content": "v^{(k+1)}" + }, + { + "bbox": [ + 55, + 369, + 296, + 578 + ], + "type": "text", + "content": " with near-identical eigenvalues (i.e., " + }, + { + "bbox": [ + 55, + 369, + 296, + 578 + ], + "type": "inline_equation", + "content": "|\\lambda_k - \\lambda_{k+1}| \\leq \\epsilon" + }, + { + "bbox": [ + 55, + 369, + 296, + 578 + ], + "type": "text", + "content": "). For such pairs, we flip the order if " + }, + { + "bbox": [ + 55, + 369, + 296, + 578 + ], + "type": "inline_equation", + "content": "v_{1}^{(k)} > v_{1}^{(k+1)}" + }, + { + "bbox": [ + 55, + 369, + 296, + 578 + ], + "type": "text", + "content": ". This reordering process is repeated until no further change occurs." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 586, + 295, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 586, + 295, + 611 + ], + "spans": [ + { + "bbox": [ + 55, + 586, + 295, + 611 + ], + "type": "text", + "content": "3.4. Hierarchical Local Traversing (HLT) for Segmentation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 616, + 295, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 616, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 616, + 295, + 689 + ], + "type": "text", + "content": "While effective for classification tasks, the SAST strategy considering each eigenvector in a separate traversal may not capture the precise relationship between patches needed for segmentation. To address this issue, we introduce a Hierarchical Local Traversal (HLT) strategy that considers the full spectrum (all " + }, + { + "bbox": [ + 55, + 616, + 295, + 689 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 55, + 616, + 295, + 689 + ], + "type": "text", + "content": " non-constant eigenvectors) simultaneously." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "Our strategy is inspired by the recursive binary partitioning technique of normalized cuts [40]. Starting from" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 553, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 293 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 293 + ], + "type": "text", + "content": "the canonicalized spectrum (see Section 3.3), tokens are first split based on the first eigenvector " + }, + { + "bbox": [ + 313, + 72, + 553, + 293 + ], + "type": "inline_equation", + "content": "v^{(1)}" + }, + { + "bbox": [ + 313, + 72, + 553, + 293 + ], + "type": "text", + "content": ", by comparing their corresponding value in " + }, + { + "bbox": [ + 313, + 72, + 553, + 293 + ], + "type": "inline_equation", + "content": "v^{(1)}" + }, + { + "bbox": [ + 313, + 72, + 553, + 293 + ], + "type": "text", + "content": " with the mean value " + }, + { + "bbox": [ + 313, + 72, + 553, + 293 + ], + "type": "inline_equation", + "content": "\\overline{v}^{(1)}" + }, + { + "bbox": [ + 313, + 72, + 553, + 293 + ], + "type": "text", + "content": ". This yields a binary partition of tokens " + }, + { + "bbox": [ + 313, + 72, + 553, + 293 + ], + "type": "inline_equation", + "content": "b_{i}^{(1)} = \\mathbb{1}\\big(v_{i}^{(1)}\\geq \\overline{v}^{(1)}\\big)\\in \\{0,1\\}" + }, + { + "bbox": [ + 313, + 72, + 553, + 293 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 313, + 72, + 553, + 293 + ], + "type": "inline_equation", + "content": "\\mathbb{1}" + }, + { + "bbox": [ + 313, + 72, + 553, + 293 + ], + "type": "text", + "content": " is the indicator function. Each subgroup is then divided based on the mean of the second eigenvector " + }, + { + "bbox": [ + 313, + 72, + 553, + 293 + ], + "type": "inline_equation", + "content": "v^{(2)}" + }, + { + "bbox": [ + 313, + 72, + 553, + 293 + ], + "type": "text", + "content": ", and so on for other eigenvectors. This partitioning process can be seen as building a binary tree, where each level corresponds to a different eigenvector and leaf nodes " + }, + { + "bbox": [ + 313, + 72, + 553, + 293 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 72, + 553, + 293 + ], + "type": "text", + "content": " are uniquely identified by the sequence of bits " + }, + { + "bbox": [ + 313, + 72, + 553, + 293 + ], + "type": "inline_equation", + "content": "b_{i} = [b_{i}^{(1)},\\dots ,b_{i}^{(s)}]" + }, + { + "bbox": [ + 313, + 72, + 553, + 293 + ], + "type": "text", + "content": " on the path from the root to the leaf. Our HLT method traverses groups of leaf nodes (groups of tokens) sequentially based on the lexicographic order of their binary code (e.g., [0000], [0001], [0010], [0011], ... in the case of four eigenvectors). For convenience, we convert binary codes " + }, + { + "bbox": [ + 313, + 72, + 553, + 293 + ], + "type": "inline_equation", + "content": "b_{i}" + }, + { + "bbox": [ + 313, + 72, + 553, + 293 + ], + "type": "text", + "content": " to a non-negative integer " + }, + { + "bbox": [ + 313, + 72, + 553, + 293 + ], + "type": "inline_equation", + "content": "q_{i}" + }, + { + "bbox": [ + 313, + 72, + 553, + 293 + ], + "type": "text", + "content": " (e.g., bin2Int([0011]) = 3) and define two traversal orders, by increasing or decreasing values of " + }, + { + "bbox": [ + 313, + 72, + 553, + 293 + ], + "type": "inline_equation", + "content": "q_{i}" + }, + { + "bbox": [ + 313, + 72, + 553, + 293 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 293, + 553, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 293, + 553, + 424 + ], + "spans": [ + { + "bbox": [ + 313, + 293, + 553, + 424 + ], + "type": "text", + "content": "For " + }, + { + "bbox": [ + 313, + 293, + 553, + 424 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 313, + 293, + 553, + 424 + ], + "type": "text", + "content": " eigenvectors, the HLT strategy described above divides tokens into " + }, + { + "bbox": [ + 313, + 293, + 553, + 424 + ], + "type": "inline_equation", + "content": "2^{s}" + }, + { + "bbox": [ + 313, + 293, + 553, + 424 + ], + "type": "text", + "content": " segments which are traversed sequentially. In the best case scenario, " + }, + { + "bbox": [ + 313, + 293, + 553, + 424 + ], + "type": "inline_equation", + "content": "\\lceil \\log_2(N_c)\\rceil" + }, + { + "bbox": [ + 313, + 293, + 553, + 424 + ], + "type": "text", + "content": " eigenvectors are thus needed to split tokens into individual segments. However, it may happen that multiple tokens fall in the same segment, especially when using fewer eigenvectors. In such case, one can further sort tokens within each segment, for example, using the values of the first eigenvector (i.e., " + }, + { + "bbox": [ + 313, + 293, + 553, + 424 + ], + "type": "inline_equation", + "content": "v^{(1)}" + }, + { + "bbox": [ + 313, + 293, + 553, + 424 + ], + "type": "text", + "content": "). In our implementation, we simply sort these tokens randomly to add stochasticity in the training. This section is illustrated in Fig. 2 (e), and Fig. 3 (b)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 425, + 554, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 425, + 554, + 519 + ], + "spans": [ + { + "bbox": [ + 313, + 425, + 554, + 519 + ], + "type": "text", + "content": "As shown in Fig. 3, the first Laplacian eigenvectors encode high-level spatial relations (e.g., bottom vs. top, left vs. right, torso vs. limbs, etc.). In the SAST, because these eigenvectors are used in separate traversals, the network may not be able to differentiate specific regions/parts of the point cloud. In contrast, our HLT strategy can capture such specific parts (e.g., head, left or right arm/thigh/alf, pelvis, etc.)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 529, + 553, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 529, + 553, + 552 + ], + "spans": [ + { + "bbox": [ + 313, + 529, + 553, + 552 + ], + "type": "text", + "content": "3.5. Traverse-Aware Repositioning (TAR) for Masked Autoencoders" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 558, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 554, + 713 + ], + "type": "text", + "content": "Following state-of-art Transformers for point cloud processing, such as Point-MAE [32] and Point-M2AE [55], our method leverages self-supervised pretraining based on MAE to boost performance. In Transformer-based approaches, the learnable tokens of masked patches can be inserted in any position of the sequence (typically at the end) as the self-attention mechanism can still attend to all tokens irrespective of their positions. However, this approach presents a significant problem in Mamba networks which are sensitive to the traversal order of tokens. We handle this problem via a TAR strategy that improves the placement of learnable tokens in MAE within Mamba networks. Specifically, we restore the learnable tokens to their" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11803" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 54, + 72, + 294, + 119 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 72, + 294, + 119 + ], + "spans": [ + { + "bbox": [ + 54, + 72, + 294, + 119 + ], + "type": "text", + "content": "original positions rather than appending them at the end of the sequence. This ensures that the essential order of tokens is maintained, preserving spatial adjacency and enhancing learning effectiveness within Mamba networks." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 120, + 295, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 120, + 295, + 274 + ], + "spans": [ + { + "bbox": [ + 55, + 120, + 295, + 274 + ], + "type": "text", + "content": "The proposed TAR strategy selects an arbitrary traversal order and randomly masks a subset of " + }, + { + "bbox": [ + 55, + 120, + 295, + 274 + ], + "type": "inline_equation", + "content": "N_{m}" + }, + { + "bbox": [ + 55, + 120, + 295, + 274 + ], + "type": "text", + "content": " tokens with the same masking ratio as the transformer-based MAEs. These tokens are then removed from the sequence, and their positions are recorded. Afterwards, the remaining (visible) tokens are fed to the encoder that outputs their representation. Before reconstructing the point cloud using the decoder, we reinsert the learnable tokens in the sequence at their recorded position. The set of masked patches is used for other traversal orders. This procedure can be seen in Fig. 2 (f). Following previous work, we measure the reconstruction error for masked patches using the Chamfer distance:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 102, + 278, + 294, + 311 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 278, + 294, + 311 + ], + "spans": [ + { + "bbox": [ + 102, + 278, + 294, + 311 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {r e c} = \\frac {1}{N _ {m}} \\sum_ {i = 1} ^ {N _ {m}} \\operatorname {C h a m f e r} \\left(\\mathcal {S} _ {i}, \\hat {\\mathcal {S}} _ {i}\\right), \\tag {5}", + "image_path": "a35f5ef5edbd3b3fa672f165cd078059df367959a94ef859f5a07d6d3176f967.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 316, + 295, + 364 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 316, + 295, + 364 + ], + "spans": [ + { + "bbox": [ + 55, + 316, + 295, + 364 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 316, + 295, + 364 + ], + "type": "inline_equation", + "content": "S_{i}\\in \\mathbb{R}^{N_{n}\\times 3}" + }, + { + "bbox": [ + 55, + 316, + 295, + 364 + ], + "type": "text", + "content": " is the set of points forming the " + }, + { + "bbox": [ + 55, + 316, + 295, + 364 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 316, + 295, + 364 + ], + "type": "text", + "content": "-th masked patch and " + }, + { + "bbox": [ + 55, + 316, + 295, + 364 + ], + "type": "inline_equation", + "content": "\\hat{S}_i" + }, + { + "bbox": [ + 55, + 316, + 295, + 364 + ], + "type": "text", + "content": " the reconstructed output for these points. The Chamfer distance between two sets of points " + }, + { + "bbox": [ + 55, + 316, + 295, + 364 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 55, + 316, + 295, + 364 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 316, + 295, + 364 + ], + "type": "inline_equation", + "content": "\\hat{S}" + }, + { + "bbox": [ + 55, + 316, + 295, + 364 + ], + "type": "text", + "content": " is defined as" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 84, + 368, + 294, + 423 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 84, + 368, + 294, + 423 + ], + "spans": [ + { + "bbox": [ + 84, + 368, + 294, + 423 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\operatorname {C h a m f e r} \\left(\\mathcal {S}, \\mathcal {S} ^ {\\prime}\\right) = \\sum_ {p \\in \\mathcal {S}} \\min _ {p ^ {\\prime} \\in \\mathcal {S} ^ {\\prime}} \\| p - p ^ {\\prime} \\| _ {2} ^ {2} \\\\ + \\sum_ {p ^ {\\prime} \\in \\mathcal {S} ^ {\\prime}} \\min _ {p \\in \\mathcal {S}} \\| p - p ^ {\\prime} \\| _ {2} ^ {2}. \\tag {6} \\\\ \\end{array}", + "image_path": "6598a2da17d25d18ef49dc41742bc01af6fea3926ab9967d0f507478436fa774.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 436, + 136, + 449 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 436, + 136, + 449 + ], + "spans": [ + { + "bbox": [ + 55, + 436, + 136, + 449 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 456, + 295, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 456, + 295, + 480 + ], + "spans": [ + { + "bbox": [ + 55, + 456, + 295, + 480 + ], + "type": "text", + "content": "Several experiments are conducted to evaluate the proposed method. First, we pretrain the Point-Mamba network us" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 71, + 548, + 277, + 712 + ], + "blocks": [ + { + "bbox": [ + 55, + 495, + 295, + 539 + ], + "lines": [ + { + "bbox": [ + 55, + 495, + 295, + 539 + ], + "spans": [ + { + "bbox": [ + 55, + 495, + 295, + 539 + ], + "type": "text", + "content": "Table 1. Few-shot classification on ModelNet40. We report the average accuracy (\\%) and standard deviation (\\%) of 10 independent experiments. \\* denotes reproduced results. A " + }, + { + "bbox": [ + 55, + 495, + 295, + 539 + ], + "type": "inline_equation", + "content": "\\equiv" + }, + { + "bbox": [ + 55, + 495, + 295, + 539 + ], + "type": "text", + "content": " represents the average (A) and standard deviation (std), respectively." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 71, + 548, + 277, + 712 + ], + "lines": [ + { + "bbox": [ + 71, + 548, + 277, + 712 + ], + "spans": [ + { + "bbox": [ + 71, + 548, + 277, + 712 + ], + "type": "table", + "html": "
Method5-way10-way
10-shot20-shot10-shot20-shot
DGCNN [48]91.83793.43286.3390.937
DGCNN + OcCo [47]91.93393.93186.4391.346
Transformer [54]87.83293.3384.6389.43
Transf. + OcCo [54]94.03695.9389.4392.446
Point-BERT [54]94.63196.3391.0392.73
Point-M2AE [55]96.83898.3492.3395.039
Point-MAE [32]96.33597.8392.6395.039
PointMamba* [26]95.93197.3391.6394.535
Ours96.43598.5392.0395.13
", + "image_path": "696cd3e92ac7f6ae4b03d9860d1207fff992acbeb86df0a7ced8d3443d8f20ac.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 553, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 204 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 204 + ], + "type": "text", + "content": "ing our techniques on the ShapeNet [5] training dataset. We then assess the performance of these pretrained models across a variety of standard benchmarks, including object classification, few-shot learning, and segmentation. Additionally, we train the model from scratch on downstream datasets to demonstrate the robustness and versatility of our method. To have a fair comparison, we adopt the masking ratio (60%) that was used in the Point-Mamba model. Moreover, a comprehensive analysis of the computational efficiency, runtime, and memory usage of our SAST approach is provided in the Supplementary Material." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 212, + 421, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 212, + 421, + 225 + ], + "spans": [ + { + "bbox": [ + 313, + 212, + 421, + 225 + ], + "type": "text", + "content": "4.1. Pretraining Setup" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 230, + 553, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 230, + 553, + 326 + ], + "spans": [ + { + "bbox": [ + 313, + 230, + 553, + 326 + ], + "type": "text", + "content": "Following Point-Mamba, we adopt the ShapeNet [5] dataset for the pretraining and assess the quality of the 3D representations produced by our approach through a linear evaluation on the ModelNet40 [52] dataset. The linear evaluation is performed by a Support Vector Machine (SVM) fitted on these features. This classification performance is quantified by the accuracy metric. Augmentation in pretraining is scale and transform, while in finetuning it is rotation." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 335, + 425, + 346 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 335, + 425, + 346 + ], + "spans": [ + { + "bbox": [ + 313, + 335, + 425, + 346 + ], + "type": "text", + "content": "4.2. Downstream Tasks" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 354, + 553, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 354, + 553, + 415 + ], + "spans": [ + { + "bbox": [ + 313, + 354, + 553, + 415 + ], + "type": "text", + "content": "Object Classification on Real-World Dataset. To evaluate our method for point clouds, we test it on the ScanObjectNN dataset [44], as described in previous studies. The augmentation used during training is random rotation. The results, presented in Tab. 3, show that our strategy signifi" + } + ] + } + ], + "index": 13 + }, + { + "type": "table", + "bbox": [ + 330, + 476, + 536, + 718 + ], + "blocks": [ + { + "bbox": [ + 313, + 434, + 553, + 467 + ], + "lines": [ + { + "bbox": [ + 313, + 434, + 553, + 467 + ], + "spans": [ + { + "bbox": [ + 313, + 434, + 553, + 467 + ], + "type": "text", + "content": "Table 2. Part segmentation on the ShapeNetPart [53]. The mIoU for all instances (Inst.) is reported. HLT is Hierarchical Local Traversing for segmentation." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 330, + 476, + 536, + 718 + ], + "lines": [ + { + "bbox": [ + 330, + 476, + 536, + 718 + ], + "spans": [ + { + "bbox": [ + 330, + 476, + 536, + 718 + ], + "type": "table", + "html": "
MethodsBackboneFLOPs (G)mIoU
Training from scratch
PointNet [34]--83.7
PointNet++ [35]--85.1
DGCNN [48]--85.2
APES [49]--85.8
Point-MAE [32]Transformer15.585.7
PointMamba [26]Mamba14.385.8
Ours (HLT)Mamba14.385.9
Training from pretrained
Transformer [54]Transformer15.585.1
Point-BERT [54]Transformer15.585.6
Point-MAE [32]Transformer15.586.1
Point-M2AE [55]Transformer15.586.5
Point-GPT-S [6]Transformer-86.2
ACT [12]Transformer-86.2
I2P-MAE [56]Transformer-86.8
PointMamba [26]Mamba14.386.0
Ours (SAST)Mamba14.385.7
Ours (HLT)Mamba14.386.1
", + "image_path": "c14631b690e81c460571edd80050a8d2e1315bf1167adfefa4c51e92e2428fa7.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_body" + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11804" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 88, + 102, + 523, + 366 + ], + "blocks": [ + { + "bbox": [ + 55, + 71, + 555, + 94 + ], + "lines": [ + { + "bbox": [ + 55, + 71, + 555, + 94 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 555, + 94 + ], + "type": "text", + "content": "Table 3. Object classification on ScanObjectNN [45]. Accuracy (%) is reported. " + }, + { + "bbox": [ + 55, + 71, + 555, + 94 + ], + "type": "inline_equation", + "content": "{}^{ \\dagger }" + }, + { + "bbox": [ + 55, + 71, + 555, + 94 + ], + "type": "text", + "content": " indicates that this method was fine-tuned without rotation augmentation." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 88, + 102, + 523, + 366 + ], + "lines": [ + { + "bbox": [ + 88, + 102, + 523, + 366 + ], + "spans": [ + { + "bbox": [ + 88, + 102, + 523, + 366 + ], + "type": "table", + "html": "
MethodsBackboneParam. (M)FLOPs (G)OBJ-BGOBJ-ONLYPB-T50-RSModelNet40
Training from scratch
PointNet [34]-3.50.573.379.268.089.2
PointNet++ [35]-1.51.782.384.377.990.7
DGCNN [48]-1.82.482.886.278.192.9
PointNeXt [37]-1.41.6--87.792.9
PointMLP [30]-13.231.4--85.4-
ADS [21]-----87.5-
Transformer [54]Transformer22.14.879.8680.5577.24-
Point-MAE [32]Transformer22.14.886.7586.9280.7892.3
PointMamba [26]Mamba12.33.690.8790.1885.6092.4
OursMamba12.33.692.2591.3987.3092.7
Training from pretrained
Transformer [54]Transformer22.14.879.8680.5577.2492.1
Transformer-OcCo [54]Transformer--84.8585.5478.79-
Point-BERT [54]Transformer22.14.887.4388.1283.0792.7
Point-M2AE† [55]Transformer--91.2288.8186.4394.0
Point-MAE [32]Transformer22.14.892.7791.2289.0493.2
PointMamba [26]Mamba12.33.693.2991.5688.1792.8
PCM [57]Mamba12.33.6--86.9-
OursMamba12.33.694.3291.9189.1093.4
", + "image_path": "e88eb1893beaeaa53821f6e421b1c48a635d35d4e242ac6fa19df4779be50d53.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 385, + 294, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 385, + 294, + 456 + ], + "spans": [ + { + "bbox": [ + 54, + 385, + 294, + 456 + ], + "type": "text", + "content": "cantly improves object classification accuracy in both training from scratch and fine-tuning scenarios. These findings highlight the effectiveness of our approach in enhancing the model's ability to identify and classify objects across various backgrounds, demonstrating its robustness in complex real-world scenarios." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 464, + 295, + 572 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 464, + 295, + 572 + ], + "spans": [ + { + "bbox": [ + 54, + 464, + 295, + 572 + ], + "type": "text", + "content": "Object Classification on Clean Objects Dataset. We also evaluated our method on the ModelNet40 [52] dataset, following the protocols established in previous works. The augmentation used during training is scale and transform. As shown in Tab. 3, our approach achieves notable enhancements on this challenging dataset compared to both the original Point-Mamba and the Transformer-based Point-MAE. This demonstrates the robustness and effectiveness of our method when applied to the Point-Mamba network." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 579, + 295, + 698 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 579, + 295, + 698 + ], + "spans": [ + { + "bbox": [ + 54, + 579, + 295, + 698 + ], + "type": "text", + "content": "Few-shot Learning. We conducted few-shot learning experiments on ModelNet40 [52] dataset, adhering to the protocols of previous studies [27, 32, 55]. The results of our few-shot learning experiments are presented in Tab. 1. Despite the competitive nature of this benchmark, our method demonstrated outstanding performance across all tested scenarios. As shown in Tab. 1, our Mamba-based method achieves results comparable to or exceeding those of transformer-based methods (Point-MAE and PointM2AE)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 701, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 701, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 701, + 295, + 714 + ], + "type": "text", + "content": "Part Segmentation. We evaluated our method's repre" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 385, + 555, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 385, + 555, + 552 + ], + "spans": [ + { + "bbox": [ + 313, + 385, + 555, + 552 + ], + "type": "text", + "content": "sensation learning ability on the ShapeNetPart dataset [53] using standard experimental settings [34, 35, 54]. Tab. 2 shows that performance gains in prior methods are minimal, highlighting the dataset's challenge. For instance, PointGPT [6] and ACT [12], despite being state-of-the-art and complex methods, show only marginal improvements over each other and other state-of-the-art approaches. Similarly, I2P-MAE [56], which combines 3D data with 2D information, shows limited improvement over Point-M2AE. For the \"Training from scratch\" setting, our method outperforms several state-of-the-art approaches. In the \"Training from pretrained\" setting, we further demonstrate the effectiveness of HLT strategy compared to SAST in the segmentation task." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 559, + 416, + 571 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 559, + 416, + 571 + ], + "spans": [ + { + "bbox": [ + 313, + 559, + 416, + 571 + ], + "type": "text", + "content": "4.3. Ablation Studies" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 576, + 554, + 637 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 576, + 554, + 637 + ], + "spans": [ + { + "bbox": [ + 313, + 576, + 554, + 637 + ], + "type": "text", + "content": "In this section, we aim to investigate the effects of different parameters on our method. We will focus on two key aspects: the effect of the number of non-constant smallest eigenvectors and the adjacency matrix used in the SAST strategy, and the analysis of the TAR strategy." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 641, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 555, + 715 + ], + "type": "text", + "content": "Analysis of Eigenvectors and Graph. One of our ablation studies investigates the impact of the number of nonconstant smallest eigenvectors used in the SAST and TAR strategies. As depicted in Fig. 4 (Left), the best performance is achieved when using the four non-constant smallest eigenvectors (blue line) for traversing. When the num" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11805" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 82, + 72, + 285, + 198 + ], + "blocks": [ + { + "bbox": [ + 82, + 72, + 285, + 198 + ], + "lines": [ + { + "bbox": [ + 82, + 72, + 285, + 198 + ], + "spans": [ + { + "bbox": [ + 82, + 72, + 285, + 198 + ], + "type": "image", + "image_path": "a183ad70f343dd6cb6973cdc3ae5ce3f391de0338854a0167e48131c1bdcd942.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 206, + 555, + 231 + ], + "lines": [ + { + "bbox": [ + 55, + 206, + 555, + 231 + ], + "spans": [ + { + "bbox": [ + 55, + 206, + 555, + 231 + ], + "type": "text", + "content": "Figure 4. Analysis of the number of non-constant smallest eigenvectors and comparison with previous methods (Left) and analysis of the number of nearest neighbors " + }, + { + "bbox": [ + 55, + 206, + 555, + 231 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 206, + 555, + 231 + ], + "type": "text", + "content": " (Right)." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 329, + 73, + 527, + 194 + ], + "blocks": [ + { + "bbox": [ + 329, + 73, + 527, + 194 + ], + "lines": [ + { + "bbox": [ + 329, + 73, + 527, + 194 + ], + "spans": [ + { + "bbox": [ + 329, + 73, + 527, + 194 + ], + "type": "image", + "image_path": "45fb3d852d239418c06ea2d6bdb397a47f2fc819eb286f4e5a144bce4c88264f.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 82, + 240, + 282, + 365 + ], + "blocks": [ + { + "bbox": [ + 82, + 240, + 282, + 365 + ], + "lines": [ + { + "bbox": [ + 82, + 240, + 282, + 365 + ], + "spans": [ + { + "bbox": [ + 82, + 240, + 282, + 365 + ], + "type": "image", + "image_path": "b9858906576a19268085fad3e2faaf70055d2e773b96a86802ba0a297698ed8f.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 129, + 374, + 481, + 388 + ], + "lines": [ + { + "bbox": [ + 129, + 374, + 481, + 388 + ], + "spans": [ + { + "bbox": [ + 129, + 374, + 481, + 388 + ], + "type": "text", + "content": "Figure 5. The effect of the TAR strategy in the pretraining phase (Left) and in finetuning (Right)." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 329, + 244, + 527, + 361 + ], + "blocks": [ + { + "bbox": [ + 329, + 244, + 527, + 361 + ], + "lines": [ + { + "bbox": [ + 329, + 244, + 527, + 361 + ], + "spans": [ + { + "bbox": [ + 329, + 244, + 527, + 361 + ], + "type": "image", + "image_path": "439b0cdc4c61d8661fdf6bca699605ea9dc2092531996e82d66ce847a0c8b077.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 396, + 295, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 396, + 295, + 502 + ], + "spans": [ + { + "bbox": [ + 54, + 396, + 295, + 502 + ], + "type": "text", + "content": "ber of non-constant smallest eigenvectors increases beyond four, the performance drops. This is because the additional eigenvectors are closer to the largest eigenvectors, which are less smooth and do not capture the most significant structural variations effectively. The significantly lower performance of Point-MAE [32], Point-Mamba [26], and PointMamba with random traversing (Random-Mamba) compared to ours underscores the importance of appropriate traversing for Mamba networks." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 503, + 295, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 503, + 295, + 624 + ], + "spans": [ + { + "bbox": [ + 54, + 503, + 295, + 624 + ], + "type": "text", + "content": "Another study examines the impact of the number of nearest neighbors used in creating the adjacency matrix. As shown in Fig. 4 (Right), the best performance (blue line) is achieved with 20 nearest neighbors. The model's accuracy increases as the number of nearest neighbors increases from 10 to 20, reaching its peak at 20 nearest neighbors. Beyond this point, the performance starts to decline. All these experiments are trained and evaluated from scratch on the ScanObjectNN dataset [44] (OBJ-BG) with scale and transform augmentation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 629, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 629, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 54, + 629, + 295, + 715 + ], + "type": "text", + "content": "Analysis of TAR Strategy. One of the studies examines the impact of the TAR strategy on the model's performance. As shown in Fig. 5 (Left), the accuracy of the model with and without the TAR strategy is plotted against the number of epochs. The model incorporating the TAR strategy (light blue line) demonstrates superior performance compared to the model without TAR (dark blue line). This figure relates" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 396, + 555, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 396, + 555, + 429 + ], + "spans": [ + { + "bbox": [ + 313, + 396, + 555, + 429 + ], + "type": "text", + "content": "to the pretraining phase on the ShapeNet [5] dataset, which is subsequently tested on the ModelNet [52] dataset using a SVM." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 431, + 556, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 431, + 556, + 540 + ], + "spans": [ + { + "bbox": [ + 313, + 431, + 556, + 540 + ], + "type": "text", + "content": "Additionally, we evaluate the effect of the TAR strategy by fine-tuning pretrained models (with or without TAR) on four downstream tasks, each corresponding to a distinct dataset. As shown in Fig. 5 (Right), the model pretrained with TAR (light blue line) consistently achieves higher accuracy than the model trained without TAR (dark blue line). This demonstrates that the model pretrained with the TAR strategy has learned more meaningful features, leading to better performance in the downstream tasks." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 550, + 468, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 550, + 468, + 562 + ], + "spans": [ + { + "bbox": [ + 313, + 550, + 468, + 562 + ], + "type": "text", + "content": "5. Conclusion and Limitation" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "text", + "content": "We introduced three strategies to enhance Mamba networks for point cloud data: isometry-invariant token traversal, recursive patch partitioning for segmentation, and improved learnable token placement, each contributing to a robust feature extraction framework. Our methods demonstrate superior performance over state-of-the-art baselines in classification, segmentation, and few-shot tasks. A limitation of our approach lies in the selection of the neighborhood size " + }, + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "text", + "content": ", which is treated as a hyperparameter. While tuning " + }, + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "text", + "content": " per dataset helps optimize performance, it may limit generalizability across diverse data distributions and require careful adjustment to maintain robustness." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "11806" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 124 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 124 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 124 + ], + "type": "text", + "content": "[1] Matan Atzmon, Haggai Maron, and Yaron Lipman. Point convolutional neural networks by extension operators. arXiv preprint arXiv:1803.10091, 2018. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 125, + 296, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 125, + 296, + 191 + ], + "spans": [ + { + "bbox": [ + 61, + 125, + 296, + 191 + ], + "type": "text", + "content": "[2] Ali Bahri, Moslem Yazdanpanah, Mehrdad Noori, Milad Cheraghalikhani, Gustavo Adolfo Vargas Hakim, David Osowiechi, Farzad Beizaee, Ismail Ben Ayed, and Christian Desrosiers. Geomask3d: Geometrically informed mask selection for self-supervised point cloud learning in 3d. arXiv preprint arXiv:2405.12419, 2024. 1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 193, + 296, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 193, + 296, + 257 + ], + "spans": [ + { + "bbox": [ + 62, + 193, + 296, + 257 + ], + "type": "text", + "content": "[3] Ali Bahri, Moslem Yazdanpanah, Mehrdad Noori, Sahar Dastani, Milad Cheraghalikhani, David Osowiech, Farzad Beizaee, Ismail Ben Ayed, Christian Desrosiers, et al. Test-time adaptation in point clouds: Leveraging sampling variation with weight averaging. arXiv preprint arXiv:2411.01116, 2024. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 259, + 295, + 292 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 259, + 295, + 292 + ], + "spans": [ + { + "bbox": [ + 62, + 259, + 295, + 292 + ], + "type": "text", + "content": "[4] Iz Beltagy, Matthew E Peters, and Arman Cohan. Longformer: The long-document transformer. arXiv preprint arXiv:2004.05150, 2020. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 294, + 295, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 294, + 295, + 348 + ], + "spans": [ + { + "bbox": [ + 62, + 294, + 295, + 348 + ], + "type": "text", + "content": "[5] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 6, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 350, + 295, + 393 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 350, + 295, + 393 + ], + "spans": [ + { + "bbox": [ + 62, + 350, + 295, + 393 + ], + "type": "text", + "content": "[6] Guangyan Chen, Meiling Wang, Yi Yang, Kai Yu, Li Yuan, and Yufeng Yue. PointGPT: Auto-regressively generative pre-training from point clouds. Advances in Neural Information Processing Systems, 36, 2024. 2, 6, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 396, + 295, + 428 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 396, + 295, + 428 + ], + "spans": [ + { + "bbox": [ + 62, + 396, + 295, + 428 + ], + "type": "text", + "content": "[7] Rewon Child, Scott Gray, Alec Radford, and Ilya Sutskever. Generating long sequences with sparse transformers. arXiv preprint arXiv:1904.10509, 2019. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 430, + 295, + 452 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 430, + 295, + 452 + ], + "spans": [ + { + "bbox": [ + 62, + 430, + 295, + 452 + ], + "type": "text", + "content": "[8] Fan RK Chung. Spectral graph theory. American Mathematical Soc., 1997. 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 453, + 295, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 453, + 295, + 485 + ], + "spans": [ + { + "bbox": [ + 62, + 453, + 295, + 485 + ], + "type": "text", + "content": "[9] Richard Courant and David Hilbert. Methods of mathematical physics: partial differential equations. John Wiley & Sons, 2008. 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 487, + 295, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 487, + 295, + 531 + ], + "spans": [ + { + "bbox": [ + 57, + 487, + 295, + 531 + ], + "type": "text", + "content": "[10] Xin Deng, WenYu Zhang, Qing Ding, and XinMing Zhang. Pointvector: a vector representation in point cloud analysis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9455-9465, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 533, + 295, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 533, + 295, + 575 + ], + "spans": [ + { + "bbox": [ + 57, + 533, + 295, + 575 + ], + "type": "text", + "content": "[11] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 578, + 295, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 578, + 295, + 632 + ], + "spans": [ + { + "bbox": [ + 57, + 578, + 295, + 632 + ], + "type": "text", + "content": "[12] Runpei Dong, Zekun Qi, Linfeng Zhang, Junbo Zhang, Jianjian Sun, Zheng Ge, Li Yi, and Kaisheng Ma. Autoencoders as cross-modal teachers: Can pretrained 2D image transformers help 3D representation learning? arXiv preprint arXiv:2212.08320, 2022. 2, 6, 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 634, + 295, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 634, + 295, + 655 + ], + "spans": [ + { + "bbox": [ + 57, + 634, + 295, + 655 + ], + "type": "text", + "content": "[13] Gene H Golub and Charles F Van Loan. Matrix computations. JHU press, 2013. 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 657, + 295, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 657, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 57, + 657, + 295, + 689 + ], + "type": "text", + "content": "[14] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752, 2023. 1, 2, 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 691, + 295, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 295, + 714 + ], + "type": "text", + "content": "[15] Albert Gu, Tri Dao, Stefano Ermon, Atri Rudra, and Christopher Ré. Hippo: Recurrent memory with optimal polynomial" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 74, + 555, + 714 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 335, + 74, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 74, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 74, + 553, + 95 + ], + "type": "text", + "content": "projections. Advances in neural information processing systems, 33:1474-1487, 2020. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 96, + 553, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 96, + 553, + 128 + ], + "spans": [ + { + "bbox": [ + 317, + 96, + 553, + 128 + ], + "type": "text", + "content": "[16] Albert Gu, Karan Goel, and Christopher Ré. Efficiently modeling long sequences with structured state spaces. arXiv preprint arXiv:2111.00396, 2021. 1, 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 129, + 553, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 129, + 553, + 172 + ], + "spans": [ + { + "bbox": [ + 317, + 129, + 553, + 172 + ], + "type": "text", + "content": "[17] Albert Gu, Karan Goel, Ankit Gupta, and Christopher Ré. On the parameterization and initialization of diagonal state space models. Advances in Neural Information Processing Systems, 35:35971-35983, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 174, + 553, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 174, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 317, + 174, + 553, + 217 + ], + "type": "text", + "content": "[18] Yulan Guo, Hanyun Wang, Qingyong Hu, Hao Liu, Li Liu, and Mohammed Bennamoun. Deep learning for 3d point clouds: A survey. IEEE transactions on pattern analysis and machine intelligence, 43(12):4338-4364, 2020. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 217, + 553, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 217, + 553, + 258 + ], + "spans": [ + { + "bbox": [ + 317, + 217, + 553, + 258 + ], + "type": "text", + "content": "[19] Ankit Gupta, Albert Gu, and Jonathan Berant. Diagonal state spaces are as effective as structured state spaces. Advances in Neural Information Processing Systems, 35:22982-22994, 2022. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 261, + 553, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 261, + 553, + 304 + ], + "spans": [ + { + "bbox": [ + 317, + 261, + 553, + 304 + ], + "type": "text", + "content": "[20] Ramin Hasani, Mathias Lechner, Tsun-Hsuan Wang, Makram Chahine, Alexander Amini, and Daniela Rus. Liquid structural state-space models. arXiv preprint arXiv:2209.12951, 2022. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 305, + 553, + 349 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 305, + 553, + 349 + ], + "spans": [ + { + "bbox": [ + 317, + 305, + 553, + 349 + ], + "type": "text", + "content": "[21] Cheng-Yao Hong, Yu-Ying Chou, and Tyng-Luh Liu. Attention discriminant sampling for point clouds. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14429-14440, 2023. 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 350, + 555, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 350, + 555, + 404 + ], + "spans": [ + { + "bbox": [ + 317, + 350, + 555, + 404 + ], + "type": "text", + "content": "[22] Xin Lai, Jianhui Liu, Li Jiang, Liwei Wang, Hengshuang Zhao, Shu Liu, Xiaojuan Qi, and Jiaya Jia. Stratified transformer for 3d point cloud segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8500-8509, 2022. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 404, + 553, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 404, + 553, + 449 + ], + "spans": [ + { + "bbox": [ + 317, + 404, + 553, + 449 + ], + "type": "text", + "content": "[23] Loic Landrieu and Martin Simonovsky. Large-scale point cloud semantic segmentation with superpoint graphs. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4558-4567, 2018. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 449, + 553, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 449, + 553, + 503 + ], + "spans": [ + { + "bbox": [ + 317, + 449, + 553, + 503 + ], + "type": "text", + "content": "[24] Juho Lee, Yoonho Lee, Jungtaek Kim, Adam Kosierek, Seungjin Choi, and Yee Whye Teh. Set transformer: A framework for attention-based permutation-invariant neural networks. In International conference on machine learning, pages 3744-3753. PMLR, 2019. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 504, + 553, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 504, + 553, + 547 + ], + "spans": [ + { + "bbox": [ + 317, + 504, + 553, + 547 + ], + "type": "text", + "content": "[25] Yangyan Li, Rui Bu, Mingchao Sun, Wei Wu, Xinhan Di, and Baoquan Chen. Pointcnn: Convolution on x-transformed points. Advances in neural information processing systems, 31, 2018. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 548, + 553, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 548, + 553, + 592 + ], + "spans": [ + { + "bbox": [ + 317, + 548, + 553, + 592 + ], + "type": "text", + "content": "[26] Dingkang Liang, Xin Zhou, Xinyu Wang, Xingkui Zhu, Wei Xu, Zhikang Zou, Xiaqing Ye, and Xiang Bai. Pointmamba: A simple state space model for point cloud analysis. arXiv preprint arXiv:2402.10739, 2024. 1, 2, 4, 6, 7, 8" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 317, + 593, + 553, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 593, + 553, + 635 + ], + "spans": [ + { + "bbox": [ + 317, + 593, + 553, + 635 + ], + "type": "text", + "content": "[27] Haotian Liu, Mu Cai, and Yong Jae Lee. Masked discrimination for self-supervised learning on point clouds. In European Conference on Computer Vision, pages 657-675. Springer, 2022. 7" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 317, + 636, + 553, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 636, + 553, + 690 + ], + "spans": [ + { + "bbox": [ + 317, + 636, + 553, + 690 + ], + "type": "text", + "content": "[28] Yang Liu, Chen Chen, Can Wang, Xulin King, and Mengyuan Liu. Regress before construct: Regress autoencoder for point cloud self-supervised learning. In Proceedings of the 31st ACM International Conference on Multimedia, pages 1738-1749, 2023. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 691, + 553, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 714 + ], + "type": "text", + "content": "[29] Yue Liu, Yunjie Tian, Yuzhong Zhao, Hongtian Yu, Lingxi Xie, Yaowei Wang, Qixiang Ye, and Yunfan Liu. Vmamba:" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11807" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "type": "text", + "content": "Visual state space model. arXiv preprint arXiv:2401.10166, 2024. 1, 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 95, + 295, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 95, + 295, + 138 + ], + "spans": [ + { + "bbox": [ + 56, + 95, + 295, + 138 + ], + "type": "text", + "content": "[30] Xu Ma, Can Qin, Haoxuan You, Haoxi Ran, and Yun Fu. Rethinking network design and local geometry in point cloud: A simple residual mlp framework. arXiv preprint arXiv:2202.07123, 2022. 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 140, + 294, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 140, + 294, + 172 + ], + "spans": [ + { + "bbox": [ + 56, + 140, + 294, + 172 + ], + "type": "text", + "content": "[31] Andrew Ng, Michael Jordan, and Yair Weiss. On spectral clustering: Analysis and an algorithm. Advances in neural information processing systems, 14, 2001. 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 173, + 294, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 173, + 294, + 226 + ], + "spans": [ + { + "bbox": [ + 56, + 173, + 294, + 226 + ], + "type": "text", + "content": "[32] Yatian Pang, Wenxiao Wang, Francis EH Tay, Wei Liu, Yonghong Tian, and Li Yuan. Masked autoencoders for point cloud self-supervised learning. In European conference on computer vision, pages 604-621. Springer, 2022. 1, 2, 5, 6, 7, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 228, + 294, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 228, + 294, + 271 + ], + "spans": [ + { + "bbox": [ + 56, + 228, + 294, + 271 + ], + "type": "text", + "content": "[33] Maciej Pioro, Kamil Ciebiera, Krystian Król, Jan Ludziejewski, and Sebastian Jaszczur. Moe-mamba: Efficient selective state space models with mixture of experts. arXiv preprint arXiv:2401.04081, 2024. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 272, + 294, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 272, + 294, + 326 + ], + "spans": [ + { + "bbox": [ + 56, + 272, + 294, + 326 + ], + "type": "text", + "content": "[34] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 652-660, 2017. 2, 6, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 327, + 294, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 327, + 294, + 370 + ], + "spans": [ + { + "bbox": [ + 56, + 327, + 294, + 370 + ], + "type": "text", + "content": "[35] Charles Ruizhongtai Qi, Li Yi, Hao Su, and Leonidas J Guibas. Pointnet++: Deep hierarchical feature learning on point sets in a metric space. Advances in neural information processing systems, 30, 2017. 1, 2, 6, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 371, + 294, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 371, + 294, + 426 + ], + "spans": [ + { + "bbox": [ + 56, + 371, + 294, + 426 + ], + "type": "text", + "content": "[36] Charles R Qi, Yin Zhou, Mahyar Najibi, Pei Sun, Khoa Vo, Boyang Deng, and Dragomir Anguelov. Offboard 3d object detection from point cloud sequences. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6134-6144, 2021. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 426, + 294, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 426, + 294, + 480 + ], + "spans": [ + { + "bbox": [ + 56, + 426, + 294, + 480 + ], + "type": "text", + "content": "[37] Guocheng Qian, Yuchen Li, Houwen Peng, Jinjie Mai, Hasan Hammoud, Mohamed Elhoseiny, and Bernard Ghanem. Pointnext: Revisiting pointnet++ with improved training and scaling strategies. Advances in Neural Information Processing Systems, 35:23192-23204, 2022. 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 481, + 294, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 481, + 294, + 525 + ], + "spans": [ + { + "bbox": [ + 56, + 481, + 294, + 525 + ], + "type": "text", + "content": "[38] Martin Reuter, Franz-Erich Wolter, and Niklas Peinecke. Laplace-spectra as fingerprints for shape matching. In Proceedings of the 2005 ACM symposium on Solid and physical modeling, pages 101-106, 2005. 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 525, + 294, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 525, + 294, + 558 + ], + "spans": [ + { + "bbox": [ + 56, + 525, + 294, + 558 + ], + "type": "text", + "content": "[39] Radu Bogdan Rusu and Steve Cousins. 3d is here: Point cloud library (pcl). In 2011 IEEE international conference on robotics and automation, pages 1-4. IEEE, 2011. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 559, + 294, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 559, + 294, + 591 + ], + "spans": [ + { + "bbox": [ + 56, + 559, + 294, + 591 + ], + "type": "text", + "content": "[40] Jianbo Shi and Jitendra Malik. Normalized cuts and image segmentation. IEEE Transactions on pattern analysis and machine intelligence, 22(8):888-905, 2000. 4, 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 592, + 294, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 592, + 294, + 645 + ], + "spans": [ + { + "bbox": [ + 56, + 592, + 294, + 645 + ], + "type": "text", + "content": "[41] Shaoshuai Shi, Xiaogang Wang, and Hongsheng Li. Poincaré: 3d object proposal generation and detection from point cloud. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 770-779, 2019. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 647, + 294, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 294, + 690 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 294, + 690 + ], + "type": "text", + "content": "[42] Yuan Tang, Xianzhi Li, Jinfeng Xu, Qiao Yu, Long Hu, Yixue Hao, and Min Chen. Point-lgmask: Local and global contexts embedding for point cloud pre-training with multi-ratio masking. IEEE Transactions on Multimedia, 2023. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "text", + "content": "[43] Yi Tay, Dara Bahri, Liu Yang, Donald Metzler, and Da-Cheng Juan. Sparse sinkhorn attention. In International" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "type": "text", + "content": "Conference on Machine Learning, pages 9438-9447. PMLR, 2020.1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 95, + 553, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 95, + 553, + 149 + ], + "spans": [ + { + "bbox": [ + 316, + 95, + 553, + 149 + ], + "type": "text", + "content": "[44] Mikaela Angelina Uy, Quang-Hieu Pham, Binh-Son Hua, Duc Thanh Nguyen, and Sai-Kit Yeung. Revisiting point cloud classification: A new benchmark dataset and classification model on real-world data. In International Conference on Computer Vision (ICCV), 2019. 6, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 150, + 553, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 150, + 553, + 215 + ], + "spans": [ + { + "bbox": [ + 316, + 150, + 553, + 215 + ], + "type": "text", + "content": "[45] Mikaela Angelina Uy, Quang-Hieu Pham, Binh-Son Hua, Thanh Nguyen, and Sai-Kit Yeung. Revisiting point cloud classification: A new benchmark dataset and classification model on real-world data. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1588–1597, 2019. 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 217, + 553, + 260 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 217, + 553, + 260 + ], + "spans": [ + { + "bbox": [ + 316, + 217, + 553, + 260 + ], + "type": "text", + "content": "[46] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 261, + 553, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 261, + 553, + 314 + ], + "spans": [ + { + "bbox": [ + 316, + 261, + 553, + 314 + ], + "type": "text", + "content": "[47] Hanchen Wang, Qi Liu, Xiangyu Yue, Joan Lasenby, and Matt J Kusner. Unsupervised point cloud pre-training via occlusion completion. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9782-9792, 2021. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 316, + 553, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 316, + 553, + 360 + ], + "spans": [ + { + "bbox": [ + 316, + 316, + 553, + 360 + ], + "type": "text", + "content": "[48] Yue Wang, Yongbin Sun, Ziwei Liu, Sanjay E Sarma, Michael M Bronstein, and Justin M Solomon. Dynamic graph cnn for learning on point clouds. ACM Transactions on Graphics (tog), 38(5):1-12, 2019. 6, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 361, + 553, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 361, + 553, + 404 + ], + "spans": [ + { + "bbox": [ + 316, + 361, + 553, + 404 + ], + "type": "text", + "content": "[49] Chengzhi Wu, Junwei Zheng, Julius Pfrommer, and Jürgen Beyerer. Attention-based point cloud edge sampling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5333-5343, 2023. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 404, + 553, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 404, + 553, + 448 + ], + "spans": [ + { + "bbox": [ + 316, + 404, + 553, + 448 + ], + "type": "text", + "content": "[50] Xiaoyang Wu, Yixing Lao, Li Jiang, Xihui Liu, and Hengshuang Zhao. Point transformer v2: Grouped vector attention and partition-based pooling. Advances in Neural Information Processing Systems, 35:33330-33342, 2022. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 449, + 553, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 449, + 553, + 492 + ], + "spans": [ + { + "bbox": [ + 316, + 449, + 553, + 492 + ], + "type": "text", + "content": "[51] Xiaoyang Wu, Li Jiang, Peng-Shuai Wang, Zhijian Liu, Xihui Liu, Yu Qiao, Wanli Ouyang, Tong He, and Hengshuang Zhao. Point transformer v3: Simpler, faster, stronger. arXiv preprint arXiv:2312.10035, 2023. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 493, + 553, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 493, + 553, + 548 + ], + "spans": [ + { + "bbox": [ + 316, + 493, + 553, + 548 + ], + "type": "text", + "content": "[52] Zhirong Wu, Shuran Song, Aditya Khosla, Fisher Yu, Linguang Zhang, Xiaou Tang, and Jianxiong Xiao. 3d shapenets: A deep representation for volumetric shapes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1912-1920, 2015. 6, 7, 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 548, + 553, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 548, + 553, + 602 + ], + "spans": [ + { + "bbox": [ + 316, + 548, + 553, + 602 + ], + "type": "text", + "content": "[53] Li Yi, Vladimir G Kim, Duygu Ceylan, I-Chao Shen, Mengyan Yan, Hao Su, Cewu Lu, Qixing Huang, Alla Sheffer, and Leonidas Guibas. A scalable active framework for region annotation in 3d shape collections. ACM Transactions on Graphics (ToG), 35(6):1-12, 2016. 6, 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 603, + 553, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 603, + 553, + 658 + ], + "spans": [ + { + "bbox": [ + 316, + 603, + 553, + 658 + ], + "type": "text", + "content": "[54] Xumin Yu, Lulu Tang, Yongming Rao, Tiejun Huang, Jie Zhou, and Jiwen Lu. Point-bert: Pre-training 3d point cloud transformers with masked point modeling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19313-19322, 2022. 1, 2, 6, 7" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 658, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 553, + 713 + ], + "type": "text", + "content": "[55] Renrui Zhang, Ziyu Guo, Peng Gao, Rongyao Fang, Bin Zhao, Dong Wang, Yu Qiao, and Hongsheng Li. Point-m2ae: multi-scale masked autoencoders for hierarchical point cloud pre-training. Advances in neural information processing systems, 35:27061-27074, 2022. 1, 2, 5, 6, 7" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11808" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 318 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "text", + "content": "[56] Renrui Zhang, Liuhui Wang, Yu Qiao, Peng Gao, and Hongsheng Li. Learning 3d representations from 2d pre-trained models via image-to-point masked autoencoders. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21769-21780, 2023. 2, 6, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 294, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 294, + 172 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 294, + 172 + ], + "type": "text", + "content": "[57] Tao Zhang, Xiangtai Li, Haobo Yuan, Shunping Ji, and Shuicheng Yan. Point could mamba: Point cloud learning via state space model. arXiv preprint arXiv:2403.00762, 2024.1, 2, 4, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 174, + 294, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 174, + 294, + 228 + ], + "spans": [ + { + "bbox": [ + 56, + 174, + 294, + 228 + ], + "type": "text", + "content": "[58] Hengshuang Zhao, Li Jiang, Chi-Wing Fu, and Jiaya Jia. Pointweb: Enhancing local neighborhood features for point cloud processing. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5565-5573, 2019. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 229, + 294, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 229, + 294, + 274 + ], + "spans": [ + { + "bbox": [ + 56, + 229, + 294, + 274 + ], + "type": "text", + "content": "[59] Hengshuang Zhao, Li Jiang, Jiaya Jia, Philip HS Torr, and Vladlen Koltun. Point transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 16259-16268, 2021. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 275, + 294, + 318 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 275, + 294, + 318 + ], + "spans": [ + { + "bbox": [ + 56, + 275, + 294, + 318 + ], + "type": "text", + "content": "[60] Lianghui Zhu, Bencheng Liao, Qian Zhang, Xinlong Wang, Wenyu Liu, and Xinggang Wang. Vision mamba: Efficient visual representation learning with bidirectional state space model. arXiv preprint arXiv:2401.09417, 2024. 1, 2" + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "11809" + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Spectral State Space Model for Rotation-Invariant Visual Representation Learning/da806762-c7b4-4ed7-80ed-cc0683ab510c_content_list.json b/2025/Spectral State Space Model for Rotation-Invariant Visual Representation Learning/da806762-c7b4-4ed7-80ed-cc0683ab510c_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..83303ebacc77008ea7c2be0a018ec4b8a9e216fc --- /dev/null +++ b/2025/Spectral State Space Model for Rotation-Invariant Visual Representation Learning/da806762-c7b4-4ed7-80ed-cc0683ab510c_content_list.json @@ -0,0 +1,1531 @@ +[ + { + "type": "text", + "text": "Spectral State Space Model for Rotation-Invariant Visual Representation Learning", + "text_level": 1, + "bbox": [ + 225, + 130, + 746, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Sahar Dastani $^{1,2*}$ Ali Bahri $^{1}$ Moslem Yazdanpanah $^{1}$ Mehrdad Noori $^{1}$ David Osowiechi $^{1}$ Gustavo Adolfo Vargas Hakim $^{1}$ Farzad Beizaee $^{1}$ Milad Cheraghalikhani $^{1}$ Arnab Kumar Mondal $^{3\\dagger}$ Herve Lombaert $^{2,4}$ Christian Desrosiers $^{1}$ $^{1}$ LIVIA, ILLS, ÉTS Montréal, Canada, $^{2}$ Mila - Quebec AI Institute, $^{3}$ Apple, $^{4}$ Polytechnique Montreal", + "bbox": [ + 148, + 202, + 821, + 313 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 233, + 345, + 313, + 359 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "State Space Models (SSMs) have recently emerged as an alternative to Vision Transformers (ViTs) due to their unique ability of modeling global relationships with linear complexity. SSMs are specifically designed to capture spatially proximate relationships of image patches. However, they fail to identify relationships between conceptually related yet not adjacent patches. This limitation arises from the non-causal nature of image data, which lacks inherent directional relationships. Additionally, current vision-based SSMs are highly sensitive to transformations such as rotation. Their predefined scanning directions depend on the original image orientation, which can cause the model to produce inconsistent patch-processing sequences after rotation. To address these limitations, we introduce Spectral VMamba, a novel approach that effectively captures the global structure within an image by leveraging spectral information derived from the graph Laplacian of image patches. Through spectral decomposition, our approach encodes patch relationships independently of image orientation, achieving patch traversal rotation invariance with our Rotational Feature Normalizer (RFN) module. Our experiments on classification tasks show that Spectral VMamba outperforms the leading SSM models in vision, such as VMamba, while maintaining invariance to rotations and a providing a similar runtime efficiency. The implementation is available at: https://github.com/Sahardastani/spectral_vmamba.git.", + "bbox": [ + 76, + 377, + 473, + 770 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 76, + 775, + 209, + 790 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Visual representation learning is fundamental to computer vision, where precise feature extraction is essential for tasks such as image classification, detection, and segmentation. The ability to extract meaningful patterns directly impacts", + "bbox": [ + 75, + 800, + 470, + 862 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6a7b98e5cf954a278021cc10f1816ed6025156841f4bb4fcff9c30a87c0c7441.jpg", + "image_caption": [ + "Figure 1. Effect of image rotation on patch processing in VMamba and Spectral VMamba networks. In VMamba networks (first row), patches are traversed using predefined horizontal and vertical scanning routes. As a result, rotating the image by $90^{\\circ}$ significantly changes the sequence in which patches are processed. In contrast, Spectral VMamba (second row) organizes patches using spectral information derived from a graph laplacian of the image patches. As shown in the second row, Spectral VMamba consistently processes background patches before snake patches in both the original and rotated images. The snake image in this analysis is AI-generated." + ], + "image_footnote": [], + "bbox": [ + 506, + 344, + 885, + 636 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "performance and unlocks deeper insights into visual tasks. Significant advances in representation learning have been driven by the development of Convolutional Neural Networks (CNNs) [14, 16, 24, 34, 36] and Vision Transformers (ViTs) [7, 23, 38, 44]. Although CNNs are effective at capturing local patterns with linear scaling, they struggle to", + "bbox": [ + 496, + 809, + 895, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "\\*correspondence: sahar.dastani-oghani.1@ens.etsmtl.ca", + "bbox": [ + 93, + 875, + 473, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$\\dagger$ work done during author's PhD at Mila/McGill", + "bbox": [ + 98, + 887, + 352, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "23881", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "efficiently model global relationships. ViTs address this limitation by introducing self-attention mechanisms, which capture global dependencies by analyzing relationships between all input elements [7, 39]. However, the quadratic complexity of ViTs poses challenges for large-scale data processing, as computational costs increase with input size.", + "bbox": [ + 75, + 90, + 470, + 181 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "State Space Models (SSMs) [8, 11, 35] have recently gained attention as a compelling alternative to traditional architectures, offering global receptive fields with linear scalability. Inspired by the achievements of Mamba [10] in language modeling, recent studies [13, 21, 33, 46] have explored the application of linearly complex global receptive fields to the visual domain. However, unlike textual data, image pixels do not have a sequential dependency. Building on the parallelized selective scan operation in Mamba, known as S6 [10], which processes sequential one-dimensional data, VMamba [21] introduced 2D Selective Scan (SS2D), a novel four-way scanning mechanism tailored for spatial domain traversal. Vision Mamba (ViM) [46] further enhanced sequence modeling by employing a bidirectional processing strategy that sums both forward and reverse passes. This process enables the model to capture comprehensive dependencies across an image. Expanding upon these approaches, Multi-Scale VMamba (MSVMamba) [33] incorporates a multiscale 2D scanning technique, which processes both original and down-sampled feature maps. This strategy improves capturing long-range dependency while minimizing computational costs. Finally, MambaVision [13] introduces a hybrid approach. Early stages employ CNNs for fast feature extraction and later stages use Mamba and Transformer blocks to capture both local and global visual features.", + "bbox": [ + 75, + 181, + 473, + 559 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Vision-based SSMs are primarily designed to capture relationships between spatially adjacent patches, focusing on spatial proximity or neighborhood-based interactions. However, SSMs are not inherently equipped to identify relationships between patches that are conceptually related but not spatially adjacent. This limitation arises due to the non-causal nature of image data, where pixels lack inherent temporal order or directional relationships.", + "bbox": [ + 75, + 559, + 470, + 679 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In addition, current vision-based SSMs tend to exhibit high sensitivity to spatial information. For instance, VMamba uses predefined traversal routes that cause the network to indiscriminately switch between foreground and background regions within an image. This approach makes the network vulnerable to perturbations and leads to confusion in assessing feature importance. By treating all regions equally, VMamba fails to prioritize the features most relevant to the task, particularly when the foreground contains critical information. This can lead to poor performance when analyzing images with complex compositions.", + "bbox": [ + 75, + 680, + 470, + 845 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Another notable challenge emerges when SSMs attempt to process images that undergo transformations, such as rotation. Vision-based SSMs typically process images by (1) un", + "bbox": [ + 75, + 845, + 470, + 892 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "folding image patches into sequences, and then (2) scanning these sequences in specific directions, such as horizontally and vertically. The sequence of image patches depends on the image orientation, meaning that rotating the image alters the order of the sequence. Similarly, the predefined scanning directions, which are fixed to the image original orientation, result in different scan paths when the image is rotated. Consequently, SS2D struggle to accommodate transformations such as rotation. Consider Figure 1 (first row): when the image is rotated by 90 degrees, the predefined and fixed traversal mechanism of VMamba significantly alters the sequence in which patches are processed.", + "bbox": [ + 496, + 90, + 893, + 271 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address these limitations, we propose Spectral VMamba, which organizes input patches based on the spectral information derived from a graph Laplacian of image patches. By leveraging the eigenvectors from spectral decomposition of an affinity matrix, our approach captures the global structure and relationships within the image. This eigenvector-driven traversal clusters similar patches together, even if they are not spatially adjacent, thereby preventing indiscriminate transitions between foreground and background regions. Moreover, spectral decomposition encodes patch relationships independently of image orientation, enabling the extraction of rotation-invariant features. Hence, our approach defines relationships between patches by their content and mutual associations rather than by their spatial arrangement. Figure 1 (second row) demonstrates that rotation does not affect the order in which patches are processed. As we can see, the patches representing the background are processed first, followed by those representing the foreground (the snake in Fig. 1).", + "bbox": [ + 496, + 272, + 895, + 558 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our main contributions are summarized as follow:", + "bbox": [ + 517, + 559, + 852, + 573 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a novel technique named Spectral VMamba based on spectral graph analysis for traversing image patches. By leveraging the eigenstructure of graphs that model the relationships between patches, our approach ensures that spatially-coherent structures in image data are maintained. Our method also improves the performance of SSMs in capturing and interpreting complex data patterns.", + "- We introduce the Rotational Feature Normalizer (RFN) module, designed to enhance our model robustness to isometric transformations, such as rotations.", + "- Our method demonstrates a notable improvement on the image classification task, outperforming the current state-of-the-art State Space Model (SSM)-based model in computer vision." + ], + "bbox": [ + 500, + 574, + 893, + 785 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 500, + 794, + 640, + 810 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "State Space Models (SSMs). Transformers have significantly improved the ability to manage long-range dependencies. However, their self-attention mechanism is computationally expensive, especially for high-resolution images, due to the quadratic scaling with input size. To address these", + "bbox": [ + 496, + 824, + 893, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "23882", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/9a1854732eefe77a4be2ffe96affb2c6fcddf37acd86be4fce43d7996b072450.jpg", + "image_caption": [ + "Figure 2. The Spectral Traversal Scan (STS) architecture begins by representing image patches as a graph. The Laplacian spectrum of this graph is then generated, and its eigenvectors are computed. The traversal path for the image patches is determined by the order of eigenvectors. The 1st eigenvector traverses the foreground/background regions, the 2nd eigenvector focuses on the next most salient structures, and each subsequent eigenvector progressively captures finer details in the image." + ], + "image_footnote": [], + "bbox": [ + 86, + 95, + 885, + 275 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "challenges, SSMs have emerged as an effective tool for long sequence modeling [11, 12, 26, 35]. In particular, the structured state-space sequence (S4) [11] model stands out for its efficient handling of long-range dependencies through a diagonal parameterization, which alleviates computational issues seen in earlier models.", + "bbox": [ + 75, + 357, + 470, + 446 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Recent advancements in sequence modeling have been driven by innovations in parallel scanning techniques, inspired by models such as S4, S5 [35] and the H3 model [8]. Mamba [10] improved the S4 model by integrating a data-dependent selection mechanism. This mechanism allows Mamba to achieve linear scalability, ensuring robust performance while accommodating longer sequence lengths without the exponential increase in resource demands. Such improvements are essential for applications requiring real-time or large-scale sequence processing, including Natural Language Processing (NLP) and time-series forecasting.", + "bbox": [ + 75, + 448, + 473, + 614 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "SSMs in vision. The success of Mamba models in the NLP domain has stimulated significant progress in computer vision, leading to various adaptations tailored for visual data processing. Several studies have extended Mamba architectures to address specific challenges in vision tasks. For instance, VMamba [21] introduced an SS2D module to tackle direction sensitivity between non-causal 2D images and ordered 1D sequences. LocalMamba [17] focuses on localized feature extraction, utilizing state-space layers to enhance neighborhood interactions and reduce dependence on self-attention, leading to efficient high-resolution image analysis. QuadMamba [43] uses a quadrilateral traversal mechanism to manage directional dependencies in 2D images, integrating features from different image segments.", + "bbox": [ + 75, + 618, + 470, + 830 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Hybrid architectures. Recent work has also focused on combining SSMs with traditional CNNs or ViTs, leveraging the respective strengths of these models. For example, U-Mamba [25] merges CNNs and state space models to address", + "bbox": [ + 76, + 835, + 472, + 897 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "long-range dependencies in biomedical image segmentation. MambaVision [13] integrates Mamba architectures with convolutional neural networks to improve performance across various vision tasks by leveraging both sequential modeling and spatial feature extraction. Finally, the TranS4mer model [35] integrates S4 with self-attention mechanisms to achieve state-of-the-art performance in movie scene detection. These advancements collectively demonstrate the versatility and scalability of structured state space models across various applications in the vision domain.", + "bbox": [ + 496, + 357, + 893, + 508 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Rotation-invariance in vision. Numerous approaches have been developed to embed equivariance and invariance directly within neural network architectures. Group-equivariant convolutional networks explicitly encode rotation symmetry into their design, providing intrinsic equivariance to rotations through steerable or group-convolutional filters [3, 4, 41, 42]. Learned canonicalization methods, such as Spatial Transformer Networks, address invariance by predicting transformations to consistently align input images into a canonical orientation, effectively standardizing pose variability [18, 19, 27]. Alternatively, symmetrization strategies explicitly aggregate feature responses across multiple rotated inputs or filters, achieving invariance by pooling orientation-specific activations [6, 20, 29, 45]. Additionally, capsule networks explicitly represent the pose of the object, leveraging pose-aware activations to enhance robustness against rotations [15, 31]. Within this landscape, our current work aligns with learned canonicalization approaches by leveraging spectral decomposition to canonicalize patch ordering and achieve patch traversal rotation invariance.", + "bbox": [ + 496, + 513, + 895, + 816 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 500, + 829, + 591, + 845 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This section begins with an overview of key concepts in SSMs and spectral graph analysis in Section 3.1, forming the foundation for our approach. Next, in Section 3.2, we in-", + "bbox": [ + 496, + 854, + 895, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "23883", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "troduce our proposed architecture, Spectral VMamba, along with the RFN module designed to achieve rotation invariance (Section 3.3). Finally, we discuss the robustness of our method to rotational variations in Section 3.6.", + "bbox": [ + 75, + 90, + 472, + 151 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Preliminaries", + "text_level": 1, + "bbox": [ + 76, + 161, + 217, + 176 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "SSM formulation. SSM-based models [10, 11] map a one-dimensional function or an input sequence $x(t) \\in \\mathbb{R}$ to an output sequence $y(t) \\in \\mathbb{R}$ through a learnable hidden state $\\mathbf{h}(t) \\in \\mathbb{R}^N$ . This mapping follows a continuous dynamical system formulated as", + "bbox": [ + 75, + 188, + 470, + 265 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\frac {d}{d t} \\mathbf {h} (t) = \\mathbf {A h} (t) + \\mathbf {B} x (t), \\quad y (t) = \\mathbf {C h} (t) \\qquad (1)\n$$\n", + "text_format": "latex", + "bbox": [ + 127, + 275, + 470, + 306 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{A} \\in \\mathbb{R}^{N \\times N}$ , $\\mathbf{B} \\in \\mathbb{R}^{N \\times 1}$ and $\\mathbf{C} \\in \\mathbb{R}^{1 \\times N}$ are parameters governing the evolution of system.", + "bbox": [ + 75, + 315, + 472, + 345 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To adapt continuous SSMs in deep learning frameworks, discretization is required. This can be achieved using the zero-order hold (ZOH) rule with a time scale parameter $\\Delta \\in \\mathbb{R}$ . Based on this rule, the discretized versions of $\\mathbf{A}$ and $\\mathbf{B}$ are defined as", + "bbox": [ + 75, + 347, + 472, + 421 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\overline {{\\mathbf {A}}} = \\exp (\\Delta \\mathbf {A}), \\\\ - \\end{array} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 129, + 431, + 468, + 455 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\overline {{\\mathbf {B}}} = (\\Delta \\mathbf {A}) ^ {- 1} \\left(\\exp (\\Delta \\mathbf {A}) - \\mathbf {I}\\right) \\Delta \\mathbf {B} \\approx \\Delta \\mathbf {B}\n$$\n", + "text_format": "latex", + "bbox": [ + 132, + 450, + 413, + 469 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "and the system becomes", + "bbox": [ + 76, + 479, + 240, + 494 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {h} (t) = \\overline {{\\mathbf {A}}} \\mathbf {h} (t - 1) + \\overline {{\\mathbf {B}}} x (t), \\quad y (t) = \\mathbf {C h} (t). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 120, + 506, + 468, + 523 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As $\\overline{\\mathbf{A}}$ and $\\overline{\\mathbf{B}}$ are linear operators, the auto-regressive formulation of Eq. (3), computing the output at each time step $t$ in a recurrent manner, can be unrolled and expressed as single convolution with kernel $\\overline{\\mathbf{K}}\\in \\mathbb{R}^{L}$ :", + "bbox": [ + 75, + 534, + 472, + 594 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {y} = \\mathbf {x} \\odot \\overline {{\\mathbf {K}}}, \\quad \\overline {{\\mathbf {K}}} = \\left(\\mathbf {C} \\overline {{\\mathbf {B}}}, \\mathbf {C} \\overline {{\\mathbf {A B}}}, \\dots , \\mathbf {C} \\overline {{\\mathbf {A}}} ^ {L - 1} \\overline {{\\mathbf {B}}}\\right). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 603, + 468, + 630 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "During training, this convolutional can be employed to compute all output values in parallel.", + "bbox": [ + 75, + 638, + 472, + 669 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Spectral Graph Analysis studies the properties of a graph by analyzing the eigen-spectrum of its Laplacian matrix, which can be seen as a discrete version of the continuous Laplace-Beltrami operator obtained by a finite difference method [30]. Given a real-valued function $f \\in C^2$ on a Riemannian manifold $\\mathcal{M}$ , the Laplace-Beltrami operator $\\Delta$ of $f$ is defined as $\\Delta f = \\text{div}(\\text{grad} f)$ where with $\\text{grad} f$ is the gradient of $f$ and $\\text{div}$ denotes the divergence.", + "bbox": [ + 75, + 672, + 470, + 794 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The Laplacian Eigenvalue Problem corresponds to solving the Helmholtz equation $\\Delta f = -\\lambda f$ , whose solutions (eigenfunctions) are directly related to the vibration of a physical membrane [2]. Following Courant's Nodal Line Theorem [5], the Laplacian eigen-spectrum is composed of non-negative eigenvalues $0 \\leq \\lambda_1 \\leq \\lambda_2 \\leq \\ldots$ , each one repeating according to its multiplicity. Moreover, the", + "bbox": [ + 75, + 795, + 470, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "eigenfunction corresponding to $\\lambda_{n}$ has at most $n$ poles of vibration, hence smaller eigenvalues and their associated eigenvector encode lower frequency structural information which captures relationships at a more global scale.", + "bbox": [ + 496, + 90, + 893, + 151 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The Laplacian matrix of a graph $G = (V, E)$ with nodes $V$ and edges $E$ is defined as $\\mathbf{L} = \\mathbf{D} - \\mathbf{W}$ , where $\\mathbf{D}$ is the diagonal degree matrix with entries $D_{ii} = \\sum_{j} W_{ij}$ . A normalized version of the Laplacian is typically employed to address differences in the weight scales $W_{ij}$ and variations in the distribution of node degrees $D_{ii}$ . In this work, we utilize the symmetric normalized Laplacian, defined as", + "bbox": [ + 496, + 151, + 893, + 257 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {L} _ {\\text {s y m}} = \\mathbf {I} - \\mathbf {D} ^ {- \\frac {1}{2}} \\mathbf {W D} ^ {- \\frac {1}{2}}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 602, + 266, + 892, + 286 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "which is positive semi-definite and possesses $|V|$ non-negative real-valued eigenvalues $0 = \\lambda_{1} \\leq \\ldots \\leq \\lambda_{|V|}$ .", + "bbox": [ + 496, + 294, + 893, + 325 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "A key property of the Laplace eigen-spectrum is that it is invariant to transformations preserving distance (i.e., isometry), which include translations and rotations. In our work, we use this property to design an SSM that enhances rotational robustness, where the order in which patches are traversed is determined based on the eigenvectors corresponding to the smallest eigenvalues of the Laplacian matrix.", + "bbox": [ + 496, + 325, + 893, + 431 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Spectral VMamba", + "text_level": 1, + "bbox": [ + 500, + 439, + 678, + 455 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Existing vision-based Mamba methods, such as VMamba [21], are based on SS2D, a four-way scanning mechanism designed for spatial domain traversal. SS2D organizes input patches along four predefined scanning routes. However, two key challenges emerge: (1) connections between semantically related but non- neighboring patches are overlooked; and (2) scanning is sensitive to isometric transformations, such as rotation. To address these limitations, we introduce the Spectral VMamba model whose architecture is summarized in Figure 3. This architecture consists of two main components: a RFN module ensuring consistent feature representation across orientations and a sequence of Spectral VMamba blocks forming the network body.", + "bbox": [ + 496, + 462, + 893, + 657 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Data forwarding in our Spectral VMamba block comprises three key steps: Spectral Traversal Scan (STS), S6 block (selective scan), and Spectral Traversal Merge (STM). In the STS step, input patches are reordered based on the Laplacian spectrum of their underlying graph structure, ensuring traversal aligns with spectral coherence (see Section 3.4). The selective scan step then processes each reordered patch sequence in parallel using dedicated S6 blocks, effectively capturing localized patterns. Finally, in the STM step, the reordered traversal paths are restored to their original spatial format, merging the resulting sequences to generate the final output map. The detailed architecture of STS is depicted in Figure 2. The sequence of patch processing is determined by the order of eigenvectors derived from the spectral decomposition of the graph Laplacian. It is important to note that STS is only performed once for the initial layer.", + "bbox": [ + 496, + 659, + 895, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "23884", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/60c5b827b37c1191418304ec6b87b25d4419ce133099069e87a45111f0b3974d.jpg", + "image_caption": [ + "Figure 3. An overview of the proposed method. Our network consists of two primary components: the RFN module, which normalizes feature orientation to ensure consistency across different perspectives, and a series of Spectral VMamba blocks that process data in three stages: (1) Spectral Traversal Scan (STS), (2) S6 block (selective scan), and (3) Spectral Traversal Merge (STM). In STS, patches are reordered based on spectral coherence derived from the Laplacian spectrum. Each sequence is then processed in parallel by dedicated S6 blocks, capturing localized patterns, and finally merged in STM to reconstruct the spatial layout, producing the final feature map." + ], + "image_footnote": [], + "bbox": [ + 86, + 95, + 890, + 324 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In subsequent layers, we apply a downsampling strategy for generating the traversal path. For a detailed explanation of the downsampling process, please refer to the supplementary material.", + "bbox": [ + 75, + 417, + 470, + 477 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The graph Laplacian eigen-spectrum is isometric invariant, guaranteeing that the order in which patches are traversed is not affected by changes to their position in the image. However, the values in the Laplacian matrix are computed using patch features which are themselves sensitive to rotation of the patch. A simple solution to this problem is to use patches containing a single pixel, however this would be impractical for two reasons. First, individual pixels encode little information (RGB color) and thus using their similarity to construct the Laplacian matrix would not be useful. More importantly, this would severely increase the computational cost of downstream operations such as the spectral decomposition of the Laplacian. To address this issue, we compute patch features using our RFN module (Section 3.3) designed to normalize the feature extraction process by aggregating features from multiple orientations of the input image. This approach allows for effectively handling of feature changes during rotations, preserving the desired order and mitigating orientation-dependent discrepancies.", + "bbox": [ + 75, + 479, + 473, + 767 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Rotational Feature Normalizer (RFN)", + "text_level": 1, + "bbox": [ + 76, + 784, + 406, + 800 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "An overview of the module is illustrated in Figure 3. This module rotates the input image by a pre-determined set of angles $\\{\\theta_i,\\dots ,\\theta_R\\}$ and processes them with a stem module which serves as a feature extraction backbone for patchification. This backbone partitions the input image $\\mathbf{I}\\in \\mathbb{R}^{H\\times W\\times 3}$ into patches $\\mathbf{x}\\in \\mathbb{R}_{\\frac{H}{p}\\times \\frac{W}{p}\\times C}$ , where $(H,W)$ denote the di", + "bbox": [ + 75, + 806, + 472, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "mensions of the input image, $C$ is the number of feature channels, and $p$ represents the patch size. To consolidate features from all rotated versions, each feature map is rotated back to its original orientation. This unrotation ensures that all extracted features align with the original image coordinate system. Finally, an element-wise max operation is applied across the unrotated feature maps to generate a single aggregated feature. The final aggregated feature map can be expressed as:", + "bbox": [ + 496, + 417, + 893, + 554 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {F} _ {i, j} = \\max _ {r \\in \\{1, \\dots , R \\}} \\left[ \\mathcal {R} _ {- \\theta_ {r}} \\left(\\operatorname {P a t c h i f y} \\left(\\mathcal {R} _ {\\theta_ {r}} (\\mathbf {I})\\right)\\right) \\right] _ {i, j} \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 539, + 561, + 893, + 589 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\theta_r$ is the angle of the $r$ -th rotation and $\\mathcal{R}_{\\theta_r}$ is the transformation applying this rotation.", + "bbox": [ + 496, + 597, + 890, + 627 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The choice of rotation angles for the RFN module represents a trade-off between rotation invariance and computational overhead. Using a larger number of angles ensures a greater coverage in terms of rotation, hence improving the invariance of the model at the cost of increased computations. In our implementation, we used four canonical angles, $\\{0^{\\circ}, 90^{\\circ}, 180^{\\circ}, 270^{\\circ}\\}$ , corresponding to quarter turns. By rotating the image at these specific angles, the pixel grid alignment is preserved, eliminating the need for interpolation. This alignment not only simplifies the back-rotation process but also guarantees that the extracted features from all rotated images are comparable.", + "bbox": [ + 496, + 627, + 893, + 808 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Spectral Traversal Scan", + "text_level": 1, + "bbox": [ + 500, + 816, + 720, + 832 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we describe how the graph Laplacian eigenspectrum is leveraged to define the traversal order of patches.", + "bbox": [ + 496, + 839, + 893, + 869 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We begin by representing the image as a graph $G = (V, E)$ , where each node $v_i \\in V$ corresponds to a patch $\\mathbf{x}_i$ .", + "bbox": [ + 496, + 869, + 893, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "23885", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "of the image, and edges $e_{ij} \\in E$ connect pairs of nodes based on their spatial proximity. Following methods for spectral clustering [28] and normalized cuts [32], we consider a weighted adjacency matrix $\\mathbf{W}$ of the graph, which is constructed using Euclidean distances between patches. Specifically, the weight $W_{ij}$ between the nodes $v_i$ and $v_j$ is given by", + "bbox": [ + 75, + 90, + 472, + 196 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nW _ {i j} = \\exp \\left(- \\frac {1}{2 \\sigma^ {2}} \\| \\mathbf {f} _ {i} - \\mathbf {f} _ {j} \\| ^ {2}\\right) \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 165, + 202, + 470, + 236 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathbf{f}_i$ is the feature vector of patch $\\mathbf{x}_i$ and $\\sigma$ is a scaling parameter determined as the average of the Euclidean distances between patches. We set $W_{ij} = 0$ for pairs $(v_i, v_j)$ that are not among the top- $k$ nearest neighbors of each other, resulting in a sparse adjacency matrix.", + "bbox": [ + 75, + 241, + 470, + 316 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We compute the first $m$ smallest eigenvectors of the symmetric Laplacian $\\mathbf{L}_{\\mathrm{sym}}$ using an iterative method known as the Arnoldi algorithm [9], which leverages the sparsity of the adjacency matrix to provide a low complexity. Each eigenvector $\\mathbf{u}^{(j)} \\in \\mathbb{R}^{N_p}$ , where $j$ ranges from 1 to $m$ , assigns an eigenfunction value $\\mathbf{u}_i^{(j)}$ to each image patch $i$ . In every Mamba block of our model, we perform two traversals per eigenvector: one in the order of increasing $\\mathbf{u}_i^{(j)}$ values and another in decreasing order. At the end of the block, the resulting features from these $2m$ traversals are concatenated together.", + "bbox": [ + 75, + 316, + 468, + 487 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Patches $\\mathbf{x}_i$ are ordered based on these eigenvectors, generating multiple traversal sequences:", + "bbox": [ + 76, + 487, + 470, + 517 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {X} ^ {(j, \\text {r e v})} = \\left\\{\\mathbf {x} _ {i _ {1}} ^ {(j, \\text {r e v})}, \\mathbf {x} _ {i _ {2}} ^ {(j, \\text {r e v})}, \\dots , \\mathbf {x} _ {i _ {m}} ^ {(j, \\text {r e v})} \\right\\}, \\text {w h e r e} \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 94, + 522, + 468, + 549 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\ni _ {k} ^ {(j, \\operatorname {r e v})} = \\arg \\min _ {i \\in \\{1, \\dots , m \\}} \\left\\{\\left(- 1\\right) ^ {\\operatorname {r e v}} \\cdot u _ {i} ^ {(j)} \\mid i \\notin \\left\\{i _ {1}, \\dots , i _ {k - 1} \\right\\} \\right\\}\n$$\n", + "text_format": "latex", + "bbox": [ + 84, + 554, + 459, + 579 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\operatorname{rev} \\in \\{0,1\\}$ indicates if the traversal is reversed $(\\operatorname{rev} = 1)$ or not $(\\operatorname{rev} = 0)$ . This concatenated sequence $\\mathbf{X}$ is then fed into the Mamba network for further processing, allowing the model to incorporate rich spectral relationships while preserving global coherence.", + "bbox": [ + 75, + 582, + 468, + 657 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.5. Spectral Canonicalization", + "text_level": 1, + "bbox": [ + 76, + 665, + 310, + 681 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Eigenvectors inherently have a sign ambiguity as, if $\\mathbf{u}$ is an eigenvector of a matrix $\\mathbf{A}$ with eigenvalue $\\lambda$ , then $-\\mathbf{u}$ is also an eigenvector with the same eigenvalue. To resolve this ambiguity, we standardize the sign of eigenvectors by ensuring that the first non-zero element of each eigenvector is positive. If the first element is negative, we flip the sign of the entire eigenvector. This adjustment does not alter the underlying structure captured by the eigenvectors, as flipping the sign consistently maintains the same eigenvalue $\\lambda$ and preserves the relationships between eigenvectors.", + "bbox": [ + 75, + 688, + 468, + 839 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.6. Rotation Invariance", + "text_level": 1, + "bbox": [ + 76, + 847, + 267, + 861 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Spectral decomposition leverages the eigen-structure of the Laplacian matrix, which remains consistent regardless of", + "bbox": [ + 75, + 869, + 470, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "rotation. As our traversal strategy is directly guided by the spectral decomposition of input features, the rotation-invariant nature of spectral decomposition extends to our method. This property is demonstrated in the following theorem.", + "bbox": [ + 496, + 90, + 893, + 165 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Theorem 3.1. Under the assumption that patch features $\\mathbf{f}$ are invariant to rotation, i.e. $\\mathbf{f}(\\mathbf{I}) = \\mathbf{f}(\\mathcal{R}_{\\theta}(\\mathbf{I}))$ for any $\\theta$ , our spectral traversal scan is also rotation-invariant.", + "bbox": [ + 496, + 176, + 893, + 222 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Proof. As rotations are a type of isometry, we can prove the more general property of isometric invariance. This invariance derives from two properties: 1) the nearest neighbors and values of the weighted adjacency matrix are defined based on distance alone, i.e., $W_{ij} = \\exp \\left(-\\frac{1}{2\\sigma}\\right)^2\\| \\mathbf{f}_i - \\mathbf{f}_j\\|^2$ , thus they do not change if distances remain the same; 2) the eigen-spectrum is invariant under cyclic permutation. This second property can be demonstrated as follows. Let $\\mathbf{P}$ be a permutation matrix and $\\mathbf{u}$ be an eigenvector of Laplacian matrix $\\mathbf{L}$ corresponding to eigenvalue $\\lambda$ . Using $\\mathbf{u}' = \\mathbf{P}\\mathbf{u}$ (thus $\\mathbf{u} = \\mathbf{P}^T\\mathbf{u}'$ ) we have that", + "bbox": [ + 496, + 232, + 893, + 400 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathbf {L} \\mathbf {u} = \\lambda \\mathbf {u} \\Rightarrow \\mathbf {L P} ^ {T} \\mathbf {u} ^ {\\prime} = \\lambda \\mathbf {P} ^ {T} \\mathbf {u} ^ {\\prime} \\\\ \\Rightarrow \\mathbf {P L P} ^ {T} \\mathbf {u} ^ {\\prime} = \\mathbf {P} (\\lambda \\mathbf {P} ^ {T} \\mathbf {u} ^ {\\prime}) \\tag {9} \\\\ \\Rightarrow \\mathbf {P L P} ^ {T} \\mathbf {u} ^ {\\prime} = \\lambda \\mathbf {u} ^ {\\prime} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 411, + 890, + 468 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Hence, $\\mathbf{Pu}$ is an eigenvector of the permuted Laplacian matrix $\\mathbf{PLP}^T$ . Since we traverse patches by their eigenvector values (from smallest to largest or vice-versa), we will therefore get the same ordering regardless of the permutation.", + "bbox": [ + 496, + 482, + 893, + 542 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 498, + 555, + 632, + 571 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We start by comparing our method against recently-proposed ViTs and SSMs on an image classification task. We then present an ablation study to thoroughly investigate the impact of various components on the performance of our method.", + "bbox": [ + 496, + 580, + 890, + 641 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Image Classification", + "text_level": 1, + "bbox": [ + 498, + 650, + 691, + 666 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We conduct classification experiments on the miniImageNet [40] dataset. This dataset includes 50,000 training images and 10,000 validation images across 100 categories. Following previous works [21, 22, 46], we train our models for 300 epochs with a batch size of 128 per GPU. The models are optimized using the AdamW optimizer with a momentum of 0.9. A cosine decay scheduler manages the learning rate, which starts at $5 \\times 10^{-4}$ , alongside a weight decay of 0.05. Additionally, we apply an exponential moving average (EMA) to stabilize training. For input images of size $224 \\times 224$ , our data augmentation techniques include color jittering, AutoAugment, random erasing, mixup, and cutmix.", + "bbox": [ + 496, + 674, + 893, + 854 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1 compares our Spectral VMamba method against ViTs and SSM-based models on the miniImageNet dataset. As can be seen, our method outperforms all other approaches", + "bbox": [ + 496, + 854, + 893, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "23886", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/c26fcbb45d6fada828cb704572d5ba7b55529b4b870f83fc9d148949a47cc450.jpg", + "table_caption": [ + "Table 1. Classification performance comparison on miniImageNet. All images are of size $224 \\times 224$ . T, S, and B denote the tiny, small, and base scales, respectively." + ], + "table_footnote": [], + "table_body": "
ModelFLOPs (G)Top-1 (%)Top-5 (%)
Transformer-Based
DeiT-S [37]4.6G70.8389.74
DeiT-B [37]17.5G72.4390.14
Swin-T [22]4.5G83.2595.54
Swin-S [22]8.7G84.1095.53
Swin-B [22]15.4G82.7795.33
XCiT-S24 [1]9.2G85.7996.31
XCiT-M24 [1]16.2G86.8096.38
SSM-Based
Vim-T [46]1.5G67.3087.97
Vim-S [46]5.1G79.7093.20
LocalVim-T [17]1.5G82.1294.60
LocalVim-S [17]4.8G81.6893.63
MSVMamba-N [33]0.9G82.1695.10
MSVMamba-M [33]1.5G83.7295.81
MSVMamba-T [33]4.6G86.4896.43
VMamba-T [21]4.9G86.2596.64
VMamba-S [21]8.7G86.4896.79
VMamba-B [21]8.7G87.1796.96
Ours-T3.9G87.86 (+1.61)97.25 (+0.61)
Ours-S6.3G88.09 (+1.61)97.08 (+0.29)
Ours-B6.3G88.17 (+1.00)97.27 (+0.31)
", + "bbox": [ + 101, + 125, + 441, + 433 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "in both top-1 and top-5 EMA accuracy. Compared to VMamba, its closest competitor, our method improves top-1 accuracy by $1.61\\%$ for the tiny and small models (T and S) and by $1\\%$ for the base model (B), while requiring less FLOPs (see 4.2 for explanation). Moreover, our Spectral VMamba (B) model yields a $1.37\\%$ higher top-1 accuracy than the strongest ViT-based model, XCiT-M24, and an improvement of $5.4\\%$ over Swin-B, one of the most popular architectures based on ViT.", + "bbox": [ + 75, + 455, + 470, + 590 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Ablation Study", + "text_level": 1, + "bbox": [ + 76, + 599, + 230, + 617 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Rotational invariance. To evaluate the robustness to rotation of our method trained on unrotated images, we evaluated its performance on test images rotated at angles from $0^{\\circ}$ to $360^{\\circ}$ with $15^{\\circ}$ increments. As a reminder, our RFN module extracts and aggregates features for four canonical rotations: $\\{0^{\\circ}, 90^{\\circ}, 180^{\\circ}, 270^{\\circ}\\}$ . We therefore expect a greater robustness at those rotation angles, for which our method is fully invariant.", + "bbox": [ + 75, + 628, + 470, + 748 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As shown in Figure 4, our model with the RFN (blue line) maintains a high accuracy near $87\\%$ at those canonical angles, demonstrating its rotational invariance property. In contrast, VMamba (orange line) suffers from a significant drop in performance for images rotated at $90^{\\circ}$ ( $\\sim 30\\%$ drop), $180^{\\circ}$ ( $\\sim 20\\%$ drop) and $270^{\\circ}$ ( $\\sim 30\\%$ drop). Although our method does not guarantee full rotational invariance for non-canonical angles, we observe a remarkably consistent performance across all rotation angles, stabilizing at around $78\\%$ in top-1 accuracy. On the other hand, the performance", + "bbox": [ + 75, + 750, + 470, + 901 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/02e8a067c6b7c6ec97104fc4f505baf212546bc32b3ab321af1337e5599511c8.jpg", + "image_caption": [ + "Figure 4. Model performance comparison across various rotation angles. Our method demonstrates consistent accuracy across all rotation angles, while VMamba exhibits significant fluctuations." + ], + "image_footnote": [], + "bbox": [ + 504, + 90, + 885, + 258 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "of VMamba fluctuates drastically, reaching accuracy values near $50\\%$ at certain angles (e.g., $105^{\\circ}$ or $255^{\\circ}$ ).", + "bbox": [ + 498, + 340, + 890, + 372 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Impact of RFN. To assess the usefulness of our RFN module, we also tested an ablation variant of our method without this module (Ours w/o RFN in Figure 4). As mentioned before, while the Laplacian eigen-spectrum is isometric-invariant, and thus invariant to rotation, the Laplacian matrix is constructed using patch features which are themselves sensitive to rotation. As we expect, removing the RFN causes the performance to drop for all angles, this drop growing as we increase the rotation (away from $0^{\\circ}$ or $360^{\\circ}$ ). Interestingly, our method without the RFN module still achieves a higher accuracy than VMamba for angles between $60^{\\circ}$ and $300^{\\circ}$ , which represent severe rotations. This could be explained by the small size of patches $(16 \\times 16)$ compared to the image, minimizing the impact of rotations (full invariance would be achieved for $1 \\times 1$ patches).", + "bbox": [ + 496, + 376, + 893, + 603 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Impact of STS. Since the STS is a critical part of our Spectral VMamba method, we cannot remove it. To evaluate its impact, we instead add the RFN to VMamba, the resulting approach now only differing from our method by not having a STS. As shown in Figure 4, the VMamba + RFN model (green line) offers a low robustness to rotation, its accuracy similar to that of VMamba. These results demonstrate that improving rotational invariance at the patch level is not sufficient to achieve a high performance.", + "bbox": [ + 496, + 608, + 893, + 744 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Number of eigenvectors. In Figure 5, we assess the effect on performance of varying the number of eigenvectors used in the STS ( $m \\in \\{1,2,3,4\\}$ ). As can be seen, accuracy improves consistently with more eigenvectors, peaking at $87.48\\%$ for $m = 4$ . This suggests that eigenvectors offer complementary information and that their combination enhances the recognition of complex global structures. In the same plot, we also report the performance of the standard VMamba model and a variant using a random traversal strategy. As can be observed, the improvement provided by", + "bbox": [ + 496, + 750, + 895, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "23887", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "our STS over VMamba, for any number of eigenvectors, is greater than the one offered by the raster scan of VMamba over a random traversal.", + "bbox": [ + 76, + 90, + 468, + 133 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/133a9b668575984eefb6abf1857bd45d68854cecddde40485323a43e56c3d175.jpg", + "image_caption": [ + "Figure 5. Model performance for different number of eigenvectors. Four eigenvectors exhibit the best performance." + ], + "image_footnote": [], + "bbox": [ + 91, + 148, + 454, + 314 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Number of nearest neighbors. The K-Nearest Neighbors (KNN) algorithm constructs a sparse adjacency matrix by selecting the top $k$ nearest neighbors for each image patch. In Tab. 2, we report our model's performance across $k$ values from 5 to 25. As can be seen, our model is not very sensitive to this parameter and achieves its highest accuracy of $87.48\\%$ for $k = 5$ neighbors. This suggests that only a few neighbors are necessary to capture the global structure of an image. This has a positive impact on the efficiency of our method as the computational complexity of the Laplacian eigen-decomposition step, using the Arnoldi algorithm, is proportional to the number of non-zero entries in the Laplacian matrix (which is directly proportional to $k$ ).", + "bbox": [ + 75, + 366, + 470, + 564 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/a8fe63bde651ddcabcba54b8997481c6c9bb5f07a3715ed42ae715ca0745f86f.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
k510152025
Top-1 Accuracy (%)87.4886.9887.1486.9187.15
", + "bbox": [ + 81, + 574, + 465, + 612 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 2. Model performance for different values of $k$ (number of nearest neighbors). The best performance is achieved at $k = 5$ .", + "bbox": [ + 76, + 622, + 468, + 651 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Memory Usage and Runtime. Next, we analyze the computational cost and time efficiency of the STS. As mentioned before, the proposed strategy minimizes computations by leveraging the sparsity of the Laplacian matrix and limiting the number of computed eigenvectors. In the following experiments, we use $m = 4$ eigenvectors and batch size = 1. Memory Usage: As depicted in Fig. 6, our STS approach (blue line) is highly efficient in terms of memory usage compared to the VMamba (orange line). Even when the number of patches increases, the STS based on a sparse eigen-solver, requires only a modest increase in memory compared to the VMamba. The stars along each line indicate the token lengths used in our method, highlighting the specific points where processing occurs across different sequence lengths.", + "bbox": [ + 75, + 657, + 468, + 869 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Runtime: Our STS strategy, which can run on a CPU, also shows a favorable runtime scaling with increased patch", + "bbox": [ + 76, + 869, + 468, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6e7566914c0176e2784e903068e1af7c19337fd073c00091bb4d92aae7a3d432.jpg", + "image_caption": [ + "Figure 6. Runtime and memory usage for different patches lengths." + ], + "image_footnote": [], + "bbox": [ + 504, + 89, + 890, + 247 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "length (Fig. 6). Specifically, for a patch length of 196 (marked by a star), which is the setting used in our main experiments, the increase in runtime is minimal when compared to the VMamba backbone, highlighting the efficiency of our approach.", + "bbox": [ + 496, + 292, + 893, + 369 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "FLOPs: The amount of FLOPs for our traversing strategy include the computations required for constructing the adjacency matrix, calculating the Laplacian, and obtaining its first eigenvectors. This calculation results in approximately 2 MB of FLOPs, which is significantly lower than the 3.9 GB FLOPs needed by VMamba's backbone. Interestingly, as reported in Tab. 1, the total number of FLOPs performed by our Spectral VMamba method is less than that of VMamba when using the same model size. This is because, after processing patches, the input to our Spectral VMamba has a size of $14 \\times 14$ tokens, compared to $56 \\times 56$ tokens for VMamba.", + "bbox": [ + 496, + 369, + 893, + 535 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 500, + 558, + 617, + 573 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we introduce Spectral VMamba, a novel approach to visual state space models that achieves patch traversal rotation invariance. Our method leverages spectral decomposition from the graph Laplacian of image patches to define traversal paths for patch processing. In addition, we develop the RFN module to ensure consistent feature representation in different orientations. This module serves as a foundation for our STS module, enabling patch features that are robust to rotational variations. By incorporating the RFN and STS modules, Spectral VMamba achieves a robust transformation-invariant feature extraction pipeline that maintains high accuracy across a variety of orientations. Our experiments demonstrate that Spectral VMamba consistently outperform popular models including Swin transformer, VMamba, MSVmamba, and LocalVim in image classification tasks. Our method does not yet address generalization to the medical domain. In future work, we aim to handle disconnected graph components, common in medical images with distinct anatomical structures or pathologies. Adapting spectral traversal for these regions will enhance pattern capture and robustness across imaging modalities.", + "bbox": [ + 496, + 583, + 893, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "23888", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 78, + 90, + 236, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We appreciate the computational resources and support provided by Compute Canada and the Digital Research Alliance of Canada.", + "bbox": [ + 76, + 114, + 470, + 159 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 78, + 174, + 174, + 190 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Alaaeldin Ali, Hugo Touvron, Mathilde Caron, Piotr Bojanowski, Matthijs Douze, Armand Joulin, Ivan Laptev, Natalia Neverova, Gabriel Synnaeve, Jakob Verbeek, et al. Xcit: Cross-covariance image transformers. NeurIPS, 34:2001-20027, 2021. 7", + "[2] Alvin Bayliss, Charles I Goldstein, and Eli Turkel. An iterative method for the helmholtz equation. Journal of Computational Physics, 49(3):443-457, 1983. 4", + "[3] Taco Cohen and Max Welling. Group equivariant convolutional networks. In International Conference on Machine Learning (ICML), pages 2990-2999, 2016. 3", + "[4] Taco S Cohen and Max Welling. Steerable cnns. In International Conference on Learning Representations (ICLR), 2017. 3", + "[5] Richard Courant and David Hilbert. Methods of mathematical physics: partial differential equations. Interscience Publishers, 1937. 4", + "[6] Sander Dieleman, Jeffrey De Fauw, and Koray Kavukcuoglu. Exploiting cyclic symmetry in convolutional neural networks. International Conference on Machine Learning (ICML), 2016. 3", + "[7] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 1, 2", + "[8] Daniel Y Fu, Tri Dao, Khaled K Saab, Armin W Thomas, Atri Rudra, and Christopher Ré. Hungry hungry hippos: Towards language modeling with state space models. arXiv preprint arXiv:2212.14052, 2022. 2, 3", + "[9] Gene H Golub and Charles F Van Loan. Matrix computations. JHU press, 2013. 6", + "[10] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752, 2023. 2, 3, 4", + "[11] Albert Gu, Karan Goel, and Christopher Ré. Efficiently modeling long sequences with structured state spaces. arXiv preprint arXiv:2111.00396, 2021. 2, 3, 4", + "[12] Albert Gu, Isys Johnson, Karan Goel, Khaled Saab, Tri Dao, Atri Rudra, and Christopher Ré. Combining recurrent, convolutional, and continuous-time models with linear state space layers. Advances in neural information processing systems, 34:572-585, 2021. 3", + "[13] Ali Hatamizadeh and Jan Kautz. Mambavision: A hybrid mamba-transformer vision backbone. arXiv preprint arXiv:2407.08083, 2024. 2, 3", + "[14] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceed-" + ], + "bbox": [ + 78, + 199, + 470, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 1", + "[15] Geoffrey E Hinton, Alex Krizhevsky, and Sida D Wang. Matrix capsules with em routing. In International Conference on Learning Representations (ICLR), 2018. 3", + "[16] Gao Huang, Zhuang Liu, Laurens Van Der Maaten, and Kilian Q Weinberger. Densely connected convolutional networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4700-4708, 2017. 1", + "[17] Tao Huang, Xiaohuan Pei, Shan You, Fei Wang, Chen Qian, and Chang Xu. Localmamba: Visual state space model with windowed selective scan. arXiv preprint arXiv:2403.09338, 2024. 3, 7", + "[18] Max Jaderberg, Karen Simonyan, Andrew Zisserman, and Koray Kavukcuoglu. Spatial transformer networks. In Advances in Neural Information Processing Systems (NeurIPS), pages 2017-2025, 2015. 3", + "[19] Sekou-Oumar Kaba, Arnab Kumar Mondal, Yan Zhang, Yoshua Bengio, and Siamak Ravanbakhsh. Equivariance with learned canonicalization functions. In International Conference on Learning Representations (ICLR), 2023. 3", + "[20] Dmitry Laptev, Nikolay Savinov, Joachim M Buhmann, and Marc Pollefeys. Ti-pooling: transformation-invariant pooling for feature learning in convolutional neural networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 289-297, 2016. 3", + "[21] Yue Liu, Yunjie Tian, Yuzhong Zhao, Hongtian Yu, Lingxi Xie, Yaowei Wang, Qixiang Ye, and Yunfan Liu. Vmamba: Visual state space model. arXiv preprint arXiv:2401.10166, 2024. 2, 3, 4, 6, 7", + "[22] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, pages 10012-10022, 2021. 6, 7", + "[23] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 1", + "[24] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11976-11986, 2022. 1", + "[25] Jun Ma, Feifei Li, and Bo Wang. U-mamba: Enhancing long-range dependency for biomedical image segmentation. arXiv preprint arXiv:2401.04722, 2024. 3", + "[26] Harsh Mehta, Ankit Gupta, Ashok Cutkosky, and Behnam Neyshabur. Long range language modeling via gated state spaces. arXiv preprint arXiv:2206.13947, 2022. 3", + "[27] Arnab Kumar Mondal, Siba Smarak Panigrahi, Oumar Kaba, Sai Rajeswar Mudumba, and Siamak Ravanbakhsh. Equivariant adaptation of large pretrained models. Advances in Neural Information Processing Systems, 36:50293-50309, 2023. 3", + "[28] Andrew Ng, Michael Jordan, and Yair Weiss. On spectral clustering: Analysis and an algorithm. Advances in neural information processing systems, 14, 2001. 6" + ], + "bbox": [ + 501, + 92, + 893, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "23889", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[29] Omri Puny, Matan Atzmon, Heli Ben-Hamu, Ishan Misra, Aditya Grover, Edward J Smith, and Yaron Lipman. Frame averaging for invariant and equivariant network design. arXiv preprint arXiv:2110.03336, 2021. 3", + "[30] Martin Reuter, Franz-Erich Wolter, and Niklas Peinecke. Laplace-beltrami spectra as 'shape-dna' of surfaces and solids. Computer-Aided Design, 38(4):342-366, 2006. 4", + "[31] Sara Sabour, Nicholas Frosst, and Geoffrey E Hinton. Dynamic routing between capsules. In Advances in Neural Information Processing Systems (NeurIPS), 2017. 3", + "[32] Jianbo Shi and Jitendra Malik. Normalized cuts and image segmentation. IEEE Transactions on pattern analysis and machine intelligence, 22(8):888-905, 2000. 6", + "[33] Yuheng Shi, Minjing Dong, and Chang Xu. Multi-scale vmamba: Hierarchy in hierarchy visual state space model. arXiv preprint arXiv:2405.14174, 2024. 2, 7", + "[34] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. In International conference on learning representations, 2014. 1", + "[35] Jimmy TH Smith, Andrew Warrington, and Scott W Linderman. Simplified state space layers for sequence modeling. arXiv preprint arXiv:2208.04933, 2022. 2, 3", + "[36] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In International conference on machine learning, pages 6105-6114. PMLR, 2019. 1", + "[37] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In Int. Conf. Machine Learning, pages 10347-10357, 2021. 7", + "[38] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In International conference on machine learning, pages 10347-10357. PMLR, 2021. 1", + "[39] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 2", + "[40] Oriol Vinyals, Charles Blundell, Timothy Lillicrap, Daan Wierstra, et al. Matching networks for one shot learning. Advances in neural information processing systems, 29, 2016. 6", + "[41] Maurice Weiler and Gabriele Cesa. General E(2)-Equivariant Steerable CNNs. In Conference on Neural Information Processing Systems (NeurIPS), 2019. 3", + "[42] Maurice Weiler, Fred A. Hamprecht, and Martin Storath. Learning steerable filters for rotation equivariant cnns. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 849-858, 2018. 3", + "[43] Fei Xie, Weijia Zhang, Zhongdao Wang, and Chao Ma. Quadmamba: Learning quadtree-based selective scan for visual state space model. arXiv preprint arXiv:2410.06806, 2024. 3", + "[44] Xiaosong Zhang, Yunjie Tian, Lingxi Xie, Wei Huang, Qi Dai, Qixiang Ye, and Qi Tian. Hivit: A simpler and more efficient" + ], + "bbox": [ + 78, + 90, + 470, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "design of hierarchical vision transformer. In The Eleventh International Conference on Learning Representations, 2023. 1", + "[45] Yanzhao Zhou, Qixiang Ye, Qiang Qiu, and Jianbin Jiao. Oriented response networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 519-528, 2017. 3", + "[46] Lianghui Zhu, Bencheng Liao, Qian Zhang, Xinlong Wang, Wenyu Liu, and Xinggang Wang. Vision mamba: Efficient visual representation learning with bidirectional state space model. arXiv preprint arXiv:2401.09417, 2024. 2, 6, 7" + ], + "bbox": [ + 501, + 92, + 893, + 246 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "23890", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Spectral State Space Model for Rotation-Invariant Visual Representation Learning/da806762-c7b4-4ed7-80ed-cc0683ab510c_model.json b/2025/Spectral State Space Model for Rotation-Invariant Visual Representation Learning/da806762-c7b4-4ed7-80ed-cc0683ab510c_model.json new file mode 100644 index 0000000000000000000000000000000000000000..5a6f34668a8934e23dfc6eeb7e2db6e9590f6244 --- /dev/null +++ b/2025/Spectral State Space Model for Rotation-Invariant Visual Representation Learning/da806762-c7b4-4ed7-80ed-cc0683ab510c_model.json @@ -0,0 +1,2079 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.227, + 0.131, + 0.748, + 0.177 + ], + "angle": 0, + "content": "Spectral State Space Model for Rotation-Invariant Visual Representation Learning" + }, + { + "type": "text", + "bbox": [ + 0.149, + 0.203, + 0.822, + 0.314 + ], + "angle": 0, + "content": "Sahar Dastani\\(^{1,2*}\\) Ali Bahri\\(^{1}\\) Moslem Yazdanpanah\\(^{1}\\) Mehrdad Noori\\(^{1}\\) David Osowiechi\\(^{1}\\) Gustavo Adolfo Vargas Hakim\\(^{1}\\) Farzad Beizaee\\(^{1}\\) Milad Cheraghalikhani\\(^{1}\\) Arnab Kumar Mondal\\(^{3\\dagger}\\) Herve Lombaert\\(^{2,4}\\) Christian Desrosiers\\(^{1}\\) \n\\(^{1}\\)LIVIA, ILLS, ÉTS Montréal, Canada, \\(^{2}\\)Mila - Quebec AI Institute, \\(^{3}\\)Apple, \\(^{4}\\)Polytechnique Montreal" + }, + { + "type": "title", + "bbox": [ + 0.235, + 0.346, + 0.314, + 0.361 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.378, + 0.474, + 0.771 + ], + "angle": 0, + "content": "State Space Models (SSMs) have recently emerged as an alternative to Vision Transformers (ViTs) due to their unique ability of modeling global relationships with linear complexity. SSMs are specifically designed to capture spatially proximate relationships of image patches. However, they fail to identify relationships between conceptually related yet not adjacent patches. This limitation arises from the non-causal nature of image data, which lacks inherent directional relationships. Additionally, current vision-based SSMs are highly sensitive to transformations such as rotation. Their predefined scanning directions depend on the original image orientation, which can cause the model to produce inconsistent patch-processing sequences after rotation. To address these limitations, we introduce Spectral VMamba, a novel approach that effectively captures the global structure within an image by leveraging spectral information derived from the graph Laplacian of image patches. Through spectral decomposition, our approach encodes patch relationships independently of image orientation, achieving patch traversal rotation invariance with our Rotational Feature Normalizer (RFN) module. Our experiments on classification tasks show that Spectral VMamba outperforms the leading SSM models in vision, such as VMamba, while maintaining invariance to rotations and a providing a similar runtime efficiency. The implementation is available at: https://github.com/Sahardastani/spectral_vmamba.git." + }, + { + "type": "title", + "bbox": [ + 0.078, + 0.776, + 0.21, + 0.791 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.801, + 0.471, + 0.863 + ], + "angle": 0, + "content": "Visual representation learning is fundamental to computer vision, where precise feature extraction is essential for tasks such as image classification, detection, and segmentation. The ability to extract meaningful patterns directly impacts" + }, + { + "type": "image", + "bbox": [ + 0.507, + 0.345, + 0.887, + 0.637 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.498, + 0.648, + 0.896, + 0.788 + ], + "angle": 0, + "content": "Figure 1. Effect of image rotation on patch processing in VMamba and Spectral VMamba networks. In VMamba networks (first row), patches are traversed using predefined horizontal and vertical scanning routes. As a result, rotating the image by \\(90^{\\circ}\\) significantly changes the sequence in which patches are processed. In contrast, Spectral VMamba (second row) organizes patches using spectral information derived from a graph laplacian of the image patches. As shown in the second row, Spectral VMamba consistently processes background patches before snake patches in both the original and rotated images. The snake image in this analysis is AI-generated." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.81, + 0.897, + 0.903 + ], + "angle": 0, + "content": "performance and unlocks deeper insights into visual tasks. Significant advances in representation learning have been driven by the development of Convolutional Neural Networks (CNNs) [14, 16, 24, 34, 36] and Vision Transformers (ViTs) [7, 23, 38, 44]. Although CNNs are effective at capturing local patterns with linear scaling, they struggle to" + }, + { + "type": "page_footnote", + "bbox": [ + 0.094, + 0.875, + 0.475, + 0.888 + ], + "angle": 0, + "content": "\\*correspondence: sahar.dastani-oghani.1@ens.etsmtl.ca" + }, + { + "type": "page_footnote", + "bbox": [ + 0.099, + 0.888, + 0.353, + 0.9 + ], + "angle": 0, + "content": "\\(\\dagger\\) work done during author's PhD at Mila/McGill" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.875, + 0.475, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "23881" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.471, + 0.182 + ], + "angle": 0, + "content": "efficiently model global relationships. ViTs address this limitation by introducing self-attention mechanisms, which capture global dependencies by analyzing relationships between all input elements [7, 39]. However, the quadratic complexity of ViTs poses challenges for large-scale data processing, as computational costs increase with input size." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.183, + 0.474, + 0.56 + ], + "angle": 0, + "content": "State Space Models (SSMs) [8, 11, 35] have recently gained attention as a compelling alternative to traditional architectures, offering global receptive fields with linear scalability. Inspired by the achievements of Mamba [10] in language modeling, recent studies [13, 21, 33, 46] have explored the application of linearly complex global receptive fields to the visual domain. However, unlike textual data, image pixels do not have a sequential dependency. Building on the parallelized selective scan operation in Mamba, known as S6 [10], which processes sequential one-dimensional data, VMamba [21] introduced 2D Selective Scan (SS2D), a novel four-way scanning mechanism tailored for spatial domain traversal. Vision Mamba (ViM) [46] further enhanced sequence modeling by employing a bidirectional processing strategy that sums both forward and reverse passes. This process enables the model to capture comprehensive dependencies across an image. Expanding upon these approaches, Multi-Scale VMamba (MSVMamba) [33] incorporates a multiscale 2D scanning technique, which processes both original and down-sampled feature maps. This strategy improves capturing long-range dependency while minimizing computational costs. Finally, MambaVision [13] introduces a hybrid approach. Early stages employ CNNs for fast feature extraction and later stages use Mamba and Transformer blocks to capture both local and global visual features." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.56, + 0.472, + 0.68 + ], + "angle": 0, + "content": "Vision-based SSMs are primarily designed to capture relationships between spatially adjacent patches, focusing on spatial proximity or neighborhood-based interactions. However, SSMs are not inherently equipped to identify relationships between patches that are conceptually related but not spatially adjacent. This limitation arises due to the non-causal nature of image data, where pixels lack inherent temporal order or directional relationships." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.681, + 0.472, + 0.846 + ], + "angle": 0, + "content": "In addition, current vision-based SSMs tend to exhibit high sensitivity to spatial information. For instance, VMamba uses predefined traversal routes that cause the network to indiscriminately switch between foreground and background regions within an image. This approach makes the network vulnerable to perturbations and leads to confusion in assessing feature importance. By treating all regions equally, VMamba fails to prioritize the features most relevant to the task, particularly when the foreground contains critical information. This can lead to poor performance when analyzing images with complex compositions." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.847, + 0.472, + 0.893 + ], + "angle": 0, + "content": "Another notable challenge emerges when SSMs attempt to process images that undergo transformations, such as rotation. Vision-based SSMs typically process images by (1) un" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.895, + 0.272 + ], + "angle": 0, + "content": "folding image patches into sequences, and then (2) scanning these sequences in specific directions, such as horizontally and vertically. The sequence of image patches depends on the image orientation, meaning that rotating the image alters the order of the sequence. Similarly, the predefined scanning directions, which are fixed to the image original orientation, result in different scan paths when the image is rotated. Consequently, SS2D struggle to accommodate transformations such as rotation. Consider Figure 1 (first row): when the image is rotated by 90 degrees, the predefined and fixed traversal mechanism of VMamba significantly alters the sequence in which patches are processed." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.273, + 0.897, + 0.559 + ], + "angle": 0, + "content": "To address these limitations, we propose Spectral VMamba, which organizes input patches based on the spectral information derived from a graph Laplacian of image patches. By leveraging the eigenvectors from spectral decomposition of an affinity matrix, our approach captures the global structure and relationships within the image. This eigenvector-driven traversal clusters similar patches together, even if they are not spatially adjacent, thereby preventing indiscriminate transitions between foreground and background regions. Moreover, spectral decomposition encodes patch relationships independently of image orientation, enabling the extraction of rotation-invariant features. Hence, our approach defines relationships between patches by their content and mutual associations rather than by their spatial arrangement. Figure 1 (second row) demonstrates that rotation does not affect the order in which patches are processed. As we can see, the patches representing the background are processed first, followed by those representing the foreground (the snake in Fig. 1)." + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.56, + 0.854, + 0.574 + ], + "angle": 0, + "content": "Our main contributions are summarized as follow:" + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.575, + 0.892, + 0.68 + ], + "angle": 0, + "content": "- We propose a novel technique named Spectral VMamba based on spectral graph analysis for traversing image patches. By leveraging the eigenstructure of graphs that model the relationships between patches, our approach ensures that spatially-coherent structures in image data are maintained. Our method also improves the performance of SSMs in capturing and interpreting complex data patterns." + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.681, + 0.892, + 0.725 + ], + "angle": 0, + "content": "- We introduce the Rotational Feature Normalizer (RFN) module, designed to enhance our model robustness to isometric transformations, such as rotations." + }, + { + "type": "text", + "bbox": [ + 0.501, + 0.726, + 0.894, + 0.786 + ], + "angle": 0, + "content": "- Our method demonstrates a notable improvement on the image classification task, outperforming the current state-of-the-art State Space Model (SSM)-based model in computer vision." + }, + { + "type": "list", + "bbox": [ + 0.501, + 0.575, + 0.894, + 0.786 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.795, + 0.642, + 0.811 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.825, + 0.895, + 0.901 + ], + "angle": 0, + "content": "State Space Models (SSMs). Transformers have significantly improved the ability to manage long-range dependencies. However, their self-attention mechanism is computationally expensive, especially for high-resolution images, due to the quadratic scaling with input size. To address these" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "23882" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.088, + 0.096, + 0.887, + 0.276 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.288, + 0.896, + 0.348 + ], + "angle": 0, + "content": "Figure 2. The Spectral Traversal Scan (STS) architecture begins by representing image patches as a graph. The Laplacian spectrum of this graph is then generated, and its eigenvectors are computed. The traversal path for the image patches is determined by the order of eigenvectors. The 1st eigenvector traverses the foreground/background regions, the 2nd eigenvector focuses on the next most salient structures, and each subsequent eigenvector progressively captures finer details in the image." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.358, + 0.471, + 0.447 + ], + "angle": 0, + "content": "challenges, SSMs have emerged as an effective tool for long sequence modeling [11, 12, 26, 35]. In particular, the structured state-space sequence (S4) [11] model stands out for its efficient handling of long-range dependencies through a diagonal parameterization, which alleviates computational issues seen in earlier models." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.449, + 0.474, + 0.616 + ], + "angle": 0, + "content": "Recent advancements in sequence modeling have been driven by innovations in parallel scanning techniques, inspired by models such as S4, S5 [35] and the H3 model [8]. Mamba [10] improved the S4 model by integrating a data-dependent selection mechanism. This mechanism allows Mamba to achieve linear scalability, ensuring robust performance while accommodating longer sequence lengths without the exponential increase in resource demands. Such improvements are essential for applications requiring real-time or large-scale sequence processing, including Natural Language Processing (NLP) and time-series forecasting." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.619, + 0.472, + 0.832 + ], + "angle": 0, + "content": "SSMs in vision. The success of Mamba models in the NLP domain has stimulated significant progress in computer vision, leading to various adaptations tailored for visual data processing. Several studies have extended Mamba architectures to address specific challenges in vision tasks. For instance, VMamba [21] introduced an SS2D module to tackle direction sensitivity between non-causal 2D images and ordered 1D sequences. LocalMamba [17] focuses on localized feature extraction, utilizing state-space layers to enhance neighborhood interactions and reduce dependence on self-attention, leading to efficient high-resolution image analysis. QuadMamba [43] uses a quadrilateral traversal mechanism to manage directional dependencies in 2D images, integrating features from different image segments." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.836, + 0.473, + 0.898 + ], + "angle": 0, + "content": "Hybrid architectures. Recent work has also focused on combining SSMs with traditional CNNs or ViTs, leveraging the respective strengths of these models. For example, U-Mamba [25] merges CNNs and state space models to address" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.358, + 0.895, + 0.509 + ], + "angle": 0, + "content": "long-range dependencies in biomedical image segmentation. MambaVision [13] integrates Mamba architectures with convolutional neural networks to improve performance across various vision tasks by leveraging both sequential modeling and spatial feature extraction. Finally, the TranS4mer model [35] integrates S4 with self-attention mechanisms to achieve state-of-the-art performance in movie scene detection. These advancements collectively demonstrate the versatility and scalability of structured state space models across various applications in the vision domain." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.514, + 0.897, + 0.817 + ], + "angle": 0, + "content": "Rotation-invariance in vision. Numerous approaches have been developed to embed equivariance and invariance directly within neural network architectures. Group-equivariant convolutional networks explicitly encode rotation symmetry into their design, providing intrinsic equivariance to rotations through steerable or group-convolutional filters [3, 4, 41, 42]. Learned canonicalization methods, such as Spatial Transformer Networks, address invariance by predicting transformations to consistently align input images into a canonical orientation, effectively standardizing pose variability [18, 19, 27]. Alternatively, symmetrization strategies explicitly aggregate feature responses across multiple rotated inputs or filters, achieving invariance by pooling orientation-specific activations [6, 20, 29, 45]. Additionally, capsule networks explicitly represent the pose of the object, leveraging pose-aware activations to enhance robustness against rotations [15, 31]. Within this landscape, our current work aligns with learned canonicalization approaches by leveraging spectral decomposition to canonicalize patch ordering and achieve patch traversal rotation invariance." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.83, + 0.593, + 0.846 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.896, + 0.903 + ], + "angle": 0, + "content": "This section begins with an overview of key concepts in SSMs and spectral graph analysis in Section 3.1, forming the foundation for our approach. Next, in Section 3.2, we in-" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23883" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.152 + ], + "angle": 0, + "content": "troduce our proposed architecture, Spectral VMamba, along with the RFN module designed to achieve rotation invariance (Section 3.3). Finally, we discuss the robustness of our method to rotational variations in Section 3.6." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.162, + 0.218, + 0.177 + ], + "angle": 0, + "content": "3.1. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.189, + 0.471, + 0.266 + ], + "angle": 0, + "content": "SSM formulation. SSM-based models [10, 11] map a one-dimensional function or an input sequence \\( x(t) \\in \\mathbb{R} \\) to an output sequence \\( y(t) \\in \\mathbb{R} \\) through a learnable hidden state \\( \\mathbf{h}(t) \\in \\mathbb{R}^N \\). This mapping follows a continuous dynamical system formulated as" + }, + { + "type": "equation", + "bbox": [ + 0.128, + 0.276, + 0.471, + 0.307 + ], + "angle": 0, + "content": "\\[\n\\frac {d}{d t} \\mathbf {h} (t) = \\mathbf {A h} (t) + \\mathbf {B} x (t), \\quad y (t) = \\mathbf {C h} (t) \\qquad (1)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.316, + 0.473, + 0.347 + ], + "angle": 0, + "content": "where \\(\\mathbf{A} \\in \\mathbb{R}^{N \\times N}\\), \\(\\mathbf{B} \\in \\mathbb{R}^{N \\times 1}\\) and \\(\\mathbf{C} \\in \\mathbb{R}^{1 \\times N}\\) are parameters governing the evolution of system." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.348, + 0.473, + 0.422 + ], + "angle": 0, + "content": "To adapt continuous SSMs in deep learning frameworks, discretization is required. This can be achieved using the zero-order hold (ZOH) rule with a time scale parameter \\(\\Delta \\in \\mathbb{R}\\). Based on this rule, the discretized versions of \\(\\mathbf{A}\\) and \\(\\mathbf{B}\\) are defined as" + }, + { + "type": "equation", + "bbox": [ + 0.13, + 0.432, + 0.47, + 0.457 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\overline {{\\mathbf {A}}} = \\exp (\\Delta \\mathbf {A}), \\\\ - \\end{array} \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.133, + 0.452, + 0.415, + 0.47 + ], + "angle": 0, + "content": "\\[\n\\overline {{\\mathbf {B}}} = (\\Delta \\mathbf {A}) ^ {- 1} \\left(\\exp (\\Delta \\mathbf {A}) - \\mathbf {I}\\right) \\Delta \\mathbf {B} \\approx \\Delta \\mathbf {B}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.481, + 0.241, + 0.495 + ], + "angle": 0, + "content": "and the system becomes" + }, + { + "type": "equation", + "bbox": [ + 0.121, + 0.507, + 0.47, + 0.525 + ], + "angle": 0, + "content": "\\[\n\\mathbf {h} (t) = \\overline {{\\mathbf {A}}} \\mathbf {h} (t - 1) + \\overline {{\\mathbf {B}}} x (t), \\quad y (t) = \\mathbf {C h} (t). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.535, + 0.473, + 0.595 + ], + "angle": 0, + "content": "As \\(\\overline{\\mathbf{A}}\\) and \\(\\overline{\\mathbf{B}}\\) are linear operators, the auto-regressive formulation of Eq. (3), computing the output at each time step \\(t\\) in a recurrent manner, can be unrolled and expressed as single convolution with kernel \\(\\overline{\\mathbf{K}}\\in \\mathbb{R}^{L}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.09, + 0.604, + 0.47, + 0.631 + ], + "angle": 0, + "content": "\\[\n\\mathbf {y} = \\mathbf {x} \\odot \\overline {{\\mathbf {K}}}, \\quad \\overline {{\\mathbf {K}}} = \\left(\\mathbf {C} \\overline {{\\mathbf {B}}}, \\mathbf {C} \\overline {{\\mathbf {A B}}}, \\dots , \\mathbf {C} \\overline {{\\mathbf {A}}} ^ {L - 1} \\overline {{\\mathbf {B}}}\\right). \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.64, + 0.473, + 0.67 + ], + "angle": 0, + "content": "During training, this convolutional can be employed to compute all output values in parallel." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.674, + 0.471, + 0.795 + ], + "angle": 0, + "content": "Spectral Graph Analysis studies the properties of a graph by analyzing the eigen-spectrum of its Laplacian matrix, which can be seen as a discrete version of the continuous Laplace-Beltrami operator obtained by a finite difference method [30]. Given a real-valued function \\( f \\in C^2 \\) on a Riemannian manifold \\( \\mathcal{M} \\), the Laplace-Beltrami operator \\( \\Delta \\) of \\( f \\) is defined as \\( \\Delta f = \\text{div}(\\text{grad} f) \\) where with \\( \\text{grad} f \\) is the gradient of \\( f \\) and \\( \\text{div} \\) denotes the divergence." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.796, + 0.472, + 0.902 + ], + "angle": 0, + "content": "The Laplacian Eigenvalue Problem corresponds to solving the Helmholtz equation \\(\\Delta f = -\\lambda f\\), whose solutions (eigenfunctions) are directly related to the vibration of a physical membrane [2]. Following Courant's Nodal Line Theorem [5], the Laplacian eigen-spectrum is composed of non-negative eigenvalues \\(0 \\leq \\lambda_1 \\leq \\lambda_2 \\leq \\ldots\\), each one repeating according to its multiplicity. Moreover, the" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.894, + 0.152 + ], + "angle": 0, + "content": "eigenfunction corresponding to \\(\\lambda_{n}\\) has at most \\(n\\) poles of vibration, hence smaller eigenvalues and their associated eigenvector encode lower frequency structural information which captures relationships at a more global scale." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.152, + 0.894, + 0.258 + ], + "angle": 0, + "content": "The Laplacian matrix of a graph \\( G = (V, E) \\) with nodes \\( V \\) and edges \\( E \\) is defined as \\( \\mathbf{L} = \\mathbf{D} - \\mathbf{W} \\), where \\( \\mathbf{D} \\) is the diagonal degree matrix with entries \\( D_{ii} = \\sum_{j} W_{ij} \\). A normalized version of the Laplacian is typically employed to address differences in the weight scales \\( W_{ij} \\) and variations in the distribution of node degrees \\( D_{ii} \\). In this work, we utilize the symmetric normalized Laplacian, defined as" + }, + { + "type": "equation", + "bbox": [ + 0.603, + 0.267, + 0.893, + 0.287 + ], + "angle": 0, + "content": "\\[\n\\mathbf {L} _ {\\text {s y m}} = \\mathbf {I} - \\mathbf {D} ^ {- \\frac {1}{2}} \\mathbf {W D} ^ {- \\frac {1}{2}}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.295, + 0.894, + 0.326 + ], + "angle": 0, + "content": "which is positive semi-definite and possesses \\(|V|\\) non-negative real-valued eigenvalues \\(0 = \\lambda_{1} \\leq \\ldots \\leq \\lambda_{|V|}\\)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.326, + 0.894, + 0.432 + ], + "angle": 0, + "content": "A key property of the Laplace eigen-spectrum is that it is invariant to transformations preserving distance (i.e., isometry), which include translations and rotations. In our work, we use this property to design an SSM that enhances rotational robustness, where the order in which patches are traversed is determined based on the eigenvectors corresponding to the smallest eigenvalues of the Laplacian matrix." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.44, + 0.679, + 0.456 + ], + "angle": 0, + "content": "3.2. Spectral VMamba" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.463, + 0.895, + 0.659 + ], + "angle": 0, + "content": "Existing vision-based Mamba methods, such as VMamba [21], are based on SS2D, a four-way scanning mechanism designed for spatial domain traversal. SS2D organizes input patches along four predefined scanning routes. However, two key challenges emerge: (1) connections between semantically related but non- neighboring patches are overlooked; and (2) scanning is sensitive to isometric transformations, such as rotation. To address these limitations, we introduce the Spectral VMamba model whose architecture is summarized in Figure 3. This architecture consists of two main components: a RFN module ensuring consistent feature representation across orientations and a sequence of Spectral VMamba blocks forming the network body." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.66, + 0.897, + 0.902 + ], + "angle": 0, + "content": "Data forwarding in our Spectral VMamba block comprises three key steps: Spectral Traversal Scan (STS), S6 block (selective scan), and Spectral Traversal Merge (STM). In the STS step, input patches are reordered based on the Laplacian spectrum of their underlying graph structure, ensuring traversal aligns with spectral coherence (see Section 3.4). The selective scan step then processes each reordered patch sequence in parallel using dedicated S6 blocks, effectively capturing localized patterns. Finally, in the STM step, the reordered traversal paths are restored to their original spatial format, merging the resulting sequences to generate the final output map. The detailed architecture of STS is depicted in Figure 2. The sequence of patch processing is determined by the order of eigenvectors derived from the spectral decomposition of the graph Laplacian. It is important to note that STS is only performed once for the initial layer." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "23884" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.087, + 0.096, + 0.891, + 0.325 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.076, + 0.335, + 0.896, + 0.409 + ], + "angle": 0, + "content": "Figure 3. An overview of the proposed method. Our network consists of two primary components: the RFN module, which normalizes feature orientation to ensure consistency across different perspectives, and a series of Spectral VMamba blocks that process data in three stages: (1) Spectral Traversal Scan (STS), (2) S6 block (selective scan), and (3) Spectral Traversal Merge (STM). In STS, patches are reordered based on spectral coherence derived from the Laplacian spectrum. Each sequence is then processed in parallel by dedicated S6 blocks, capturing localized patterns, and finally merged in STM to reconstruct the spatial layout, producing the final feature map." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.419, + 0.471, + 0.478 + ], + "angle": 0, + "content": "In subsequent layers, we apply a downsampling strategy for generating the traversal path. For a detailed explanation of the downsampling process, please refer to the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.48, + 0.474, + 0.768 + ], + "angle": 0, + "content": "The graph Laplacian eigen-spectrum is isometric invariant, guaranteeing that the order in which patches are traversed is not affected by changes to their position in the image. However, the values in the Laplacian matrix are computed using patch features which are themselves sensitive to rotation of the patch. A simple solution to this problem is to use patches containing a single pixel, however this would be impractical for two reasons. First, individual pixels encode little information (RGB color) and thus using their similarity to construct the Laplacian matrix would not be useful. More importantly, this would severely increase the computational cost of downstream operations such as the spectral decomposition of the Laplacian. To address this issue, we compute patch features using our RFN module (Section 3.3) designed to normalize the feature extraction process by aggregating features from multiple orientations of the input image. This approach allows for effectively handling of feature changes during rotations, preserving the desired order and mitigating orientation-dependent discrepancies." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.785, + 0.407, + 0.801 + ], + "angle": 0, + "content": "3.3. Rotational Feature Normalizer (RFN)" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.808, + 0.473, + 0.903 + ], + "angle": 0, + "content": "An overview of the module is illustrated in Figure 3. This module rotates the input image by a pre-determined set of angles \\(\\{\\theta_i,\\dots ,\\theta_R\\}\\) and processes them with a stem module which serves as a feature extraction backbone for patchification. This backbone partitions the input image \\(\\mathbf{I}\\in \\mathbb{R}^{H\\times W\\times 3}\\) into patches \\(\\mathbf{x}\\in \\mathbb{R}_{\\frac{H}{p}\\times \\frac{W}{p}\\times C}\\), where \\((H,W)\\) denote the di" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.419, + 0.895, + 0.555 + ], + "angle": 0, + "content": "mensions of the input image, \\( C \\) is the number of feature channels, and \\( p \\) represents the patch size. To consolidate features from all rotated versions, each feature map is rotated back to its original orientation. This unrotation ensures that all extracted features align with the original image coordinate system. Finally, an element-wise max operation is applied across the unrotated feature maps to generate a single aggregated feature. The final aggregated feature map can be expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.54, + 0.563, + 0.895, + 0.59 + ], + "angle": 0, + "content": "\\[\n\\mathbf {F} _ {i, j} = \\max _ {r \\in \\{1, \\dots , R \\}} \\left[ \\mathcal {R} _ {- \\theta_ {r}} \\left(\\operatorname {P a t c h i f y} \\left(\\mathcal {R} _ {\\theta_ {r}} (\\mathbf {I})\\right)\\right) \\right] _ {i, j} \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.598, + 0.892, + 0.628 + ], + "angle": 0, + "content": "where \\(\\theta_r\\) is the angle of the \\(r\\)-th rotation and \\(\\mathcal{R}_{\\theta_r}\\) is the transformation applying this rotation." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.628, + 0.895, + 0.809 + ], + "angle": 0, + "content": "The choice of rotation angles for the RFN module represents a trade-off between rotation invariance and computational overhead. Using a larger number of angles ensures a greater coverage in terms of rotation, hence improving the invariance of the model at the cost of increased computations. In our implementation, we used four canonical angles, \\(\\{0^{\\circ}, 90^{\\circ}, 180^{\\circ}, 270^{\\circ}\\}\\), corresponding to quarter turns. By rotating the image at these specific angles, the pixel grid alignment is preserved, eliminating the need for interpolation. This alignment not only simplifies the back-rotation process but also guarantees that the extracted features from all rotated images are comparable." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.818, + 0.722, + 0.833 + ], + "angle": 0, + "content": "3.4. Spectral Traversal Scan" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.84, + 0.895, + 0.871 + ], + "angle": 0, + "content": "In this section, we describe how the graph Laplacian eigenspectrum is leveraged to define the traversal order of patches." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.871, + 0.895, + 0.903 + ], + "angle": 0, + "content": "We begin by representing the image as a graph \\( G = (V, E) \\), where each node \\( v_i \\in V \\) corresponds to a patch \\( \\mathbf{x}_i \\)." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23885" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.076, + 0.092, + 0.473, + 0.198 + ], + "angle": 0, + "content": "of the image, and edges \\( e_{ij} \\in E \\) connect pairs of nodes based on their spatial proximity. Following methods for spectral clustering [28] and normalized cuts [32], we consider a weighted adjacency matrix \\( \\mathbf{W} \\) of the graph, which is constructed using Euclidean distances between patches. Specifically, the weight \\( W_{ij} \\) between the nodes \\( v_i \\) and \\( v_j \\) is given by" + }, + { + "type": "equation", + "bbox": [ + 0.166, + 0.203, + 0.471, + 0.237 + ], + "angle": 0, + "content": "\\[\nW _ {i j} = \\exp \\left(- \\frac {1}{2 \\sigma^ {2}} \\| \\mathbf {f} _ {i} - \\mathbf {f} _ {j} \\| ^ {2}\\right) \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.242, + 0.471, + 0.317 + ], + "angle": 0, + "content": "where \\(\\mathbf{f}_i\\) is the feature vector of patch \\(\\mathbf{x}_i\\) and \\(\\sigma\\) is a scaling parameter determined as the average of the Euclidean distances between patches. We set \\(W_{ij} = 0\\) for pairs \\((v_i, v_j)\\) that are not among the top-\\(k\\) nearest neighbors of each other, resulting in a sparse adjacency matrix." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.318, + 0.47, + 0.488 + ], + "angle": 0, + "content": "We compute the first \\(m\\) smallest eigenvectors of the symmetric Laplacian \\(\\mathbf{L}_{\\mathrm{sym}}\\) using an iterative method known as the Arnoldi algorithm [9], which leverages the sparsity of the adjacency matrix to provide a low complexity. Each eigenvector \\(\\mathbf{u}^{(j)} \\in \\mathbb{R}^{N_p}\\), where \\(j\\) ranges from 1 to \\(m\\), assigns an eigenfunction value \\(\\mathbf{u}_i^{(j)}\\) to each image patch \\(i\\). In every Mamba block of our model, we perform two traversals per eigenvector: one in the order of increasing \\(\\mathbf{u}_i^{(j)}\\) values and another in decreasing order. At the end of the block, the resulting features from these \\(2m\\) traversals are concatenated together." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.488, + 0.471, + 0.518 + ], + "angle": 0, + "content": "Patches \\(\\mathbf{x}_i\\) are ordered based on these eigenvectors, generating multiple traversal sequences:" + }, + { + "type": "equation", + "bbox": [ + 0.095, + 0.523, + 0.47, + 0.55 + ], + "angle": 0, + "content": "\\[\n\\mathbf {X} ^ {(j, \\text {r e v})} = \\left\\{\\mathbf {x} _ {i _ {1}} ^ {(j, \\text {r e v})}, \\mathbf {x} _ {i _ {2}} ^ {(j, \\text {r e v})}, \\dots , \\mathbf {x} _ {i _ {m}} ^ {(j, \\text {r e v})} \\right\\}, \\text {w h e r e} \\tag {8}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.086, + 0.555, + 0.46, + 0.58 + ], + "angle": 0, + "content": "\\[\ni _ {k} ^ {(j, \\operatorname {r e v})} = \\arg \\min _ {i \\in \\{1, \\dots , m \\}} \\left\\{\\left(- 1\\right) ^ {\\operatorname {r e v}} \\cdot u _ {i} ^ {(j)} \\mid i \\notin \\left\\{i _ {1}, \\dots , i _ {k - 1} \\right\\} \\right\\}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.583, + 0.47, + 0.659 + ], + "angle": 0, + "content": "where \\( \\operatorname{rev} \\in \\{0,1\\} \\) indicates if the traversal is reversed \\( (\\operatorname{rev} = 1) \\) or not \\( (\\operatorname{rev} = 0) \\). This concatenated sequence \\( \\mathbf{X} \\) is then fed into the Mamba network for further processing, allowing the model to incorporate rich spectral relationships while preserving global coherence." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.666, + 0.312, + 0.682 + ], + "angle": 0, + "content": "3.5. Spectral Canonicalization" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.689, + 0.47, + 0.84 + ], + "angle": 0, + "content": "Eigenvectors inherently have a sign ambiguity as, if \\(\\mathbf{u}\\) is an eigenvector of a matrix \\(\\mathbf{A}\\) with eigenvalue \\(\\lambda\\), then \\(-\\mathbf{u}\\) is also an eigenvector with the same eigenvalue. To resolve this ambiguity, we standardize the sign of eigenvectors by ensuring that the first non-zero element of each eigenvector is positive. If the first element is negative, we flip the sign of the entire eigenvector. This adjustment does not alter the underlying structure captured by the eigenvectors, as flipping the sign consistently maintains the same eigenvalue \\(\\lambda\\) and preserves the relationships between eigenvectors." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.848, + 0.268, + 0.862 + ], + "angle": 0, + "content": "3.6. Rotation Invariance" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.871, + 0.471, + 0.901 + ], + "angle": 0, + "content": "Spectral decomposition leverages the eigen-structure of the Laplacian matrix, which remains consistent regardless of" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.092, + 0.894, + 0.166 + ], + "angle": 0, + "content": "rotation. As our traversal strategy is directly guided by the spectral decomposition of input features, the rotation-invariant nature of spectral decomposition extends to our method. This property is demonstrated in the following theorem." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.178, + 0.894, + 0.223 + ], + "angle": 0, + "content": "Theorem 3.1. Under the assumption that patch features \\(\\mathbf{f}\\) are invariant to rotation, i.e. \\(\\mathbf{f}(\\mathbf{I}) = \\mathbf{f}(\\mathcal{R}_{\\theta}(\\mathbf{I}))\\) for any \\(\\theta\\), our spectral traversal scan is also rotation-invariant." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.233, + 0.895, + 0.401 + ], + "angle": 0, + "content": "Proof. As rotations are a type of isometry, we can prove the more general property of isometric invariance. This invariance derives from two properties: 1) the nearest neighbors and values of the weighted adjacency matrix are defined based on distance alone, i.e., \\( W_{ij} = \\exp \\left(-\\frac{1}{2\\sigma}\\right)^2\\| \\mathbf{f}_i - \\mathbf{f}_j\\|^2 \\), thus they do not change if distances remain the same; 2) the eigen-spectrum is invariant under cyclic permutation. This second property can be demonstrated as follows. Let \\(\\mathbf{P}\\) be a permutation matrix and \\(\\mathbf{u}\\) be an eigenvector of Laplacian matrix \\(\\mathbf{L}\\) corresponding to eigenvalue \\(\\lambda\\). Using \\(\\mathbf{u}' = \\mathbf{P}\\mathbf{u}\\) (thus \\(\\mathbf{u} = \\mathbf{P}^T\\mathbf{u}'\\)) we have that" + }, + { + "type": "equation", + "bbox": [ + 0.567, + 0.412, + 0.892, + 0.469 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathbf {L} \\mathbf {u} = \\lambda \\mathbf {u} \\Rightarrow \\mathbf {L P} ^ {T} \\mathbf {u} ^ {\\prime} = \\lambda \\mathbf {P} ^ {T} \\mathbf {u} ^ {\\prime} \\\\ \\Rightarrow \\mathbf {P L P} ^ {T} \\mathbf {u} ^ {\\prime} = \\mathbf {P} (\\lambda \\mathbf {P} ^ {T} \\mathbf {u} ^ {\\prime}) \\tag {9} \\\\ \\Rightarrow \\mathbf {P L P} ^ {T} \\mathbf {u} ^ {\\prime} = \\lambda \\mathbf {u} ^ {\\prime} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.483, + 0.894, + 0.544 + ], + "angle": 0, + "content": "Hence, \\(\\mathbf{Pu}\\) is an eigenvector of the permuted Laplacian matrix \\(\\mathbf{PLP}^T\\). Since we traverse patches by their eigenvector values (from smallest to largest or vice-versa), we will therefore get the same ordering regardless of the permutation." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.556, + 0.633, + 0.573 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.581, + 0.892, + 0.642 + ], + "angle": 0, + "content": "We start by comparing our method against recently-proposed ViTs and SSMs on an image classification task. We then present an ablation study to thoroughly investigate the impact of various components on the performance of our method." + }, + { + "type": "title", + "bbox": [ + 0.499, + 0.651, + 0.692, + 0.667 + ], + "angle": 0, + "content": "4.1. Image Classification" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.675, + 0.894, + 0.856 + ], + "angle": 0, + "content": "We conduct classification experiments on the miniImageNet [40] dataset. This dataset includes 50,000 training images and 10,000 validation images across 100 categories. Following previous works [21, 22, 46], we train our models for 300 epochs with a batch size of 128 per GPU. The models are optimized using the AdamW optimizer with a momentum of 0.9. A cosine decay scheduler manages the learning rate, which starts at \\(5 \\times 10^{-4}\\), alongside a weight decay of 0.05. Additionally, we apply an exponential moving average (EMA) to stabilize training. For input images of size \\(224 \\times 224\\), our data augmentation techniques include color jittering, AutoAugment, random erasing, mixup, and cutmix." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.856, + 0.894, + 0.901 + ], + "angle": 0, + "content": "Table 1 compares our Spectral VMamba method against ViTs and SSM-based models on the miniImageNet dataset. As can be seen, our method outperforms all other approaches" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "23886" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.077, + 0.472, + 0.12 + ], + "angle": 0, + "content": "Table 1. Classification performance comparison on miniImageNet. All images are of size \\(224 \\times 224\\). T, S, and B denote the tiny, small, and base scales, respectively." + }, + { + "type": "table", + "bbox": [ + 0.102, + 0.125, + 0.442, + 0.434 + ], + "angle": 0, + "content": "
ModelFLOPs (G)Top-1 (%)Top-5 (%)
Transformer-Based
DeiT-S [37]4.6G70.8389.74
DeiT-B [37]17.5G72.4390.14
Swin-T [22]4.5G83.2595.54
Swin-S [22]8.7G84.1095.53
Swin-B [22]15.4G82.7795.33
XCiT-S24 [1]9.2G85.7996.31
XCiT-M24 [1]16.2G86.8096.38
SSM-Based
Vim-T [46]1.5G67.3087.97
Vim-S [46]5.1G79.7093.20
LocalVim-T [17]1.5G82.1294.60
LocalVim-S [17]4.8G81.6893.63
MSVMamba-N [33]0.9G82.1695.10
MSVMamba-M [33]1.5G83.7295.81
MSVMamba-T [33]4.6G86.4896.43
VMamba-T [21]4.9G86.2596.64
VMamba-S [21]8.7G86.4896.79
VMamba-B [21]8.7G87.1796.96
Ours-T3.9G87.86 (+1.61)97.25 (+0.61)
Ours-S6.3G88.09 (+1.61)97.08 (+0.29)
Ours-B6.3G88.17 (+1.00)97.27 (+0.31)
" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.456, + 0.472, + 0.592 + ], + "angle": 0, + "content": "in both top-1 and top-5 EMA accuracy. Compared to VMamba, its closest competitor, our method improves top-1 accuracy by \\(1.61\\%\\) for the tiny and small models (T and S) and by \\(1\\%\\) for the base model (B), while requiring less FLOPs (see 4.2 for explanation). Moreover, our Spectral VMamba (B) model yields a \\(1.37\\%\\) higher top-1 accuracy than the strongest ViT-based model, XCiT-M24, and an improvement of \\(5.4\\%\\) over Swin-B, one of the most popular architectures based on ViT." + }, + { + "type": "title", + "bbox": [ + 0.077, + 0.601, + 0.231, + 0.618 + ], + "angle": 0, + "content": "4.2. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.629, + 0.471, + 0.749 + ], + "angle": 0, + "content": "Rotational invariance. To evaluate the robustness to rotation of our method trained on unrotated images, we evaluated its performance on test images rotated at angles from \\(0^{\\circ}\\) to \\(360^{\\circ}\\) with \\(15^{\\circ}\\) increments. As a reminder, our RFN module extracts and aggregates features for four canonical rotations: \\(\\{0^{\\circ}, 90^{\\circ}, 180^{\\circ}, 270^{\\circ}\\}\\). We therefore expect a greater robustness at those rotation angles, for which our method is fully invariant." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.75, + 0.472, + 0.902 + ], + "angle": 0, + "content": "As shown in Figure 4, our model with the RFN (blue line) maintains a high accuracy near \\(87\\%\\) at those canonical angles, demonstrating its rotational invariance property. In contrast, VMamba (orange line) suffers from a significant drop in performance for images rotated at \\(90^{\\circ}\\) (\\(\\sim 30\\%\\) drop), \\(180^{\\circ}\\) (\\(\\sim 20\\%\\) drop) and \\(270^{\\circ}\\) (\\(\\sim 30\\%\\) drop). Although our method does not guarantee full rotational invariance for non-canonical angles, we observe a remarkably consistent performance across all rotation angles, stabilizing at around \\(78\\%\\) in top-1 accuracy. On the other hand, the performance" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.092, + 0.887, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.273, + 0.892, + 0.316 + ], + "angle": 0, + "content": "Figure 4. Model performance comparison across various rotation angles. Our method demonstrates consistent accuracy across all rotation angles, while VMamba exhibits significant fluctuations." + }, + { + "type": "text", + "bbox": [ + 0.499, + 0.342, + 0.892, + 0.373 + ], + "angle": 0, + "content": "of VMamba fluctuates drastically, reaching accuracy values near \\(50\\%\\) at certain angles (e.g., \\(105^{\\circ}\\) or \\(255^{\\circ}\\))." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.377, + 0.895, + 0.604 + ], + "angle": 0, + "content": "Impact of RFN. To assess the usefulness of our RFN module, we also tested an ablation variant of our method without this module (Ours w/o RFN in Figure 4). As mentioned before, while the Laplacian eigen-spectrum is isometric-invariant, and thus invariant to rotation, the Laplacian matrix is constructed using patch features which are themselves sensitive to rotation. As we expect, removing the RFN causes the performance to drop for all angles, this drop growing as we increase the rotation (away from \\(0^{\\circ}\\) or \\(360^{\\circ}\\)). Interestingly, our method without the RFN module still achieves a higher accuracy than VMamba for angles between \\(60^{\\circ}\\) and \\(300^{\\circ}\\), which represent severe rotations. This could be explained by the small size of patches \\((16 \\times 16)\\) compared to the image, minimizing the impact of rotations (full invariance would be achieved for \\(1 \\times 1\\) patches)." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.609, + 0.895, + 0.746 + ], + "angle": 0, + "content": "Impact of STS. Since the STS is a critical part of our Spectral VMamba method, we cannot remove it. To evaluate its impact, we instead add the RFN to VMamba, the resulting approach now only differing from our method by not having a STS. As shown in Figure 4, the VMamba + RFN model (green line) offers a low robustness to rotation, its accuracy similar to that of VMamba. These results demonstrate that improving rotational invariance at the patch level is not sufficient to achieve a high performance." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.75, + 0.897, + 0.903 + ], + "angle": 0, + "content": "Number of eigenvectors. In Figure 5, we assess the effect on performance of varying the number of eigenvectors used in the STS (\\(m \\in \\{1,2,3,4\\}\\)). As can be seen, accuracy improves consistently with more eigenvectors, peaking at \\(87.48\\%\\) for \\(m = 4\\). This suggests that eigenvectors offer complementary information and that their combination enhances the recognition of complex global structures. In the same plot, we also report the performance of the standard VMamba model and a variant using a random traversal strategy. As can be observed, the improvement provided by" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23887" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.077, + 0.092, + 0.47, + 0.135 + ], + "angle": 0, + "content": "our STS over VMamba, for any number of eigenvectors, is greater than the one offered by the raster scan of VMamba over a random traversal." + }, + { + "type": "image", + "bbox": [ + 0.092, + 0.149, + 0.455, + 0.315 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.077, + 0.329, + 0.47, + 0.358 + ], + "angle": 0, + "content": "Figure 5. Model performance for different number of eigenvectors. Four eigenvectors exhibit the best performance." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.367, + 0.471, + 0.565 + ], + "angle": 0, + "content": "Number of nearest neighbors. The K-Nearest Neighbors (KNN) algorithm constructs a sparse adjacency matrix by selecting the top \\( k \\) nearest neighbors for each image patch. In Tab. 2, we report our model's performance across \\( k \\) values from 5 to 25. As can be seen, our model is not very sensitive to this parameter and achieves its highest accuracy of \\( 87.48\\% \\) for \\( k = 5 \\) neighbors. This suggests that only a few neighbors are necessary to capture the global structure of an image. This has a positive impact on the efficiency of our method as the computational complexity of the Laplacian eigen-decomposition step, using the Arnoldi algorithm, is proportional to the number of non-zero entries in the Laplacian matrix (which is directly proportional to \\( k \\))." + }, + { + "type": "table", + "bbox": [ + 0.082, + 0.575, + 0.466, + 0.613 + ], + "angle": 0, + "content": "
k510152025
Top-1 Accuracy (%)87.4886.9887.1486.9187.15
" + }, + { + "type": "table_caption", + "bbox": [ + 0.077, + 0.623, + 0.47, + 0.652 + ], + "angle": 0, + "content": "Table 2. Model performance for different values of \\( k \\) (number of nearest neighbors). The best performance is achieved at \\( k = 5 \\) ." + }, + { + "type": "text", + "bbox": [ + 0.076, + 0.659, + 0.47, + 0.871 + ], + "angle": 0, + "content": "Memory Usage and Runtime. Next, we analyze the computational cost and time efficiency of the STS. As mentioned before, the proposed strategy minimizes computations by leveraging the sparsity of the Laplacian matrix and limiting the number of computed eigenvectors. In the following experiments, we use \\( m = 4 \\) eigenvectors and batch size = 1. Memory Usage: As depicted in Fig. 6, our STS approach (blue line) is highly efficient in terms of memory usage compared to the VMamba (orange line). Even when the number of patches increases, the STS based on a sparse eigen-solver, requires only a modest increase in memory compared to the VMamba. The stars along each line indicate the token lengths used in our method, highlighting the specific points where processing occurs across different sequence lengths." + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.871, + 0.469, + 0.902 + ], + "angle": 0, + "content": "Runtime: Our STS strategy, which can run on a CPU, also shows a favorable runtime scaling with increased patch" + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.09, + 0.891, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.257, + 0.89, + 0.273 + ], + "angle": 0, + "content": "Figure 6. Runtime and memory usage for different patches lengths." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.294, + 0.894, + 0.37 + ], + "angle": 0, + "content": "length (Fig. 6). Specifically, for a patch length of 196 (marked by a star), which is the setting used in our main experiments, the increase in runtime is minimal when compared to the VMamba backbone, highlighting the efficiency of our approach." + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.37, + 0.895, + 0.536 + ], + "angle": 0, + "content": "FLOPs: The amount of FLOPs for our traversing strategy include the computations required for constructing the adjacency matrix, calculating the Laplacian, and obtaining its first eigenvectors. This calculation results in approximately 2 MB of FLOPs, which is significantly lower than the 3.9 GB FLOPs needed by VMamba's backbone. Interestingly, as reported in Tab. 1, the total number of FLOPs performed by our Spectral VMamba method is less than that of VMamba when using the same model size. This is because, after processing patches, the input to our Spectral VMamba has a size of \\(14 \\times 14\\) tokens, compared to \\(56 \\times 56\\) tokens for VMamba." + }, + { + "type": "title", + "bbox": [ + 0.5, + 0.559, + 0.619, + 0.574 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.498, + 0.584, + 0.895, + 0.903 + ], + "angle": 0, + "content": "In this paper, we introduce Spectral VMamba, a novel approach to visual state space models that achieves patch traversal rotation invariance. Our method leverages spectral decomposition from the graph Laplacian of image patches to define traversal paths for patch processing. In addition, we develop the RFN module to ensure consistent feature representation in different orientations. This module serves as a foundation for our STS module, enabling patch features that are robust to rotational variations. By incorporating the RFN and STS modules, Spectral VMamba achieves a robust transformation-invariant feature extraction pipeline that maintains high accuracy across a variety of orientations. Our experiments demonstrate that Spectral VMamba consistently outperform popular models including Swin transformer, VMamba, MSVmamba, and LocalVim in image classification tasks. Our method does not yet address generalization to the medical domain. In future work, we aim to handle disconnected graph components, common in medical images with distinct anatomical structures or pathologies. Adapting spectral traversal for these regions will enhance pattern capture and robustness across imaging modalities." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "23888" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.079, + 0.091, + 0.238, + 0.108 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.077, + 0.115, + 0.472, + 0.16 + ], + "angle": 0, + "content": "We appreciate the computational resources and support provided by Compute Canada and the Digital Research Alliance of Canada." + }, + { + "type": "title", + "bbox": [ + 0.08, + 0.175, + 0.175, + 0.191 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.2, + 0.472, + 0.269 + ], + "angle": 0, + "content": "[1] Alaaeldin Ali, Hugo Touvron, Mathilde Caron, Piotr Bojanowski, Matthijs Douze, Armand Joulin, Ivan Laptev, Natalia Neverova, Gabriel Synnaeve, Jakob Verbeek, et al. Xcit: Cross-covariance image transformers. NeurIPS, 34:2001-20027, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.271, + 0.472, + 0.313 + ], + "angle": 0, + "content": "[2] Alvin Bayliss, Charles I Goldstein, and Eli Turkel. An iterative method for the helmholtz equation. Journal of Computational Physics, 49(3):443-457, 1983. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.315, + 0.472, + 0.357 + ], + "angle": 0, + "content": "[3] Taco Cohen and Max Welling. Group equivariant convolutional networks. In International Conference on Machine Learning (ICML), pages 2990-2999, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.359, + 0.472, + 0.398 + ], + "angle": 0, + "content": "[4] Taco S Cohen and Max Welling. Steerable cnns. In International Conference on Learning Representations (ICLR), 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.401, + 0.472, + 0.441 + ], + "angle": 0, + "content": "[5] Richard Courant and David Hilbert. Methods of mathematical physics: partial differential equations. Interscience Publishers, 1937. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.444, + 0.472, + 0.497 + ], + "angle": 0, + "content": "[6] Sander Dieleman, Jeffrey De Fauw, and Koray Kavukcuoglu. Exploiting cyclic symmetry in convolutional neural networks. International Conference on Machine Learning (ICML), 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.501, + 0.472, + 0.584 + ], + "angle": 0, + "content": "[7] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.586, + 0.472, + 0.64 + ], + "angle": 0, + "content": "[8] Daniel Y Fu, Tri Dao, Khaled K Saab, Armin W Thomas, Atri Rudra, and Christopher Ré. Hungry hungry hippos: Towards language modeling with state space models. arXiv preprint arXiv:2212.14052, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.087, + 0.643, + 0.472, + 0.671 + ], + "angle": 0, + "content": "[9] Gene H Golub and Charles F Van Loan. Matrix computations. JHU press, 2013. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.673, + 0.472, + 0.713 + ], + "angle": 0, + "content": "[10] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752, 2023. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.716, + 0.472, + 0.757 + ], + "angle": 0, + "content": "[11] Albert Gu, Karan Goel, and Christopher Ré. Efficiently modeling long sequences with structured state spaces. arXiv preprint arXiv:2111.00396, 2021. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.759, + 0.472, + 0.827 + ], + "angle": 0, + "content": "[12] Albert Gu, Isys Johnson, Karan Goel, Khaled Saab, Tri Dao, Atri Rudra, and Christopher Ré. Combining recurrent, convolutional, and continuous-time models with linear state space layers. Advances in neural information processing systems, 34:572-585, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.83, + 0.472, + 0.871 + ], + "angle": 0, + "content": "[13] Ali Hatamizadeh and Jan Kautz. Mambavision: A hybrid mamba-transformer vision backbone. arXiv preprint arXiv:2407.08083, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.873, + 0.472, + 0.902 + ], + "angle": 0, + "content": "[14] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceed-" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.2, + 0.472, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.533, + 0.093, + 0.892, + 0.12 + ], + "angle": 0, + "content": "ings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.122, + 0.894, + 0.163 + ], + "angle": 0, + "content": "[15] Geoffrey E Hinton, Alex Krizhevsky, and Sida D Wang. Matrix capsules with em routing. In International Conference on Learning Representations (ICLR), 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.165, + 0.894, + 0.22 + ], + "angle": 0, + "content": "[16] Gao Huang, Zhuang Liu, Laurens Van Der Maaten, and Kilian Q Weinberger. Densely connected convolutional networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4700-4708, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.222, + 0.894, + 0.275 + ], + "angle": 0, + "content": "[17] Tao Huang, Xiaohuan Pei, Shan You, Fei Wang, Chen Qian, and Chang Xu. Localmamba: Visual state space model with windowed selective scan. arXiv preprint arXiv:2403.09338, 2024. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.278, + 0.894, + 0.333 + ], + "angle": 0, + "content": "[18] Max Jaderberg, Karen Simonyan, Andrew Zisserman, and Koray Kavukcuoglu. Spatial transformer networks. In Advances in Neural Information Processing Systems (NeurIPS), pages 2017-2025, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.335, + 0.894, + 0.39 + ], + "angle": 0, + "content": "[19] Sekou-Oumar Kaba, Arnab Kumar Mondal, Yan Zhang, Yoshua Bengio, and Siamak Ravanbakhsh. Equivariance with learned canonicalization functions. In International Conference on Learning Representations (ICLR), 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.392, + 0.892, + 0.461 + ], + "angle": 0, + "content": "[20] Dmitry Laptev, Nikolay Savinov, Joachim M Buhmann, and Marc Pollefeys. Ti-pooling: transformation-invariant pooling for feature learning in convolutional neural networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 289-297, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.463, + 0.892, + 0.517 + ], + "angle": 0, + "content": "[21] Yue Liu, Yunjie Tian, Yuzhong Zhao, Hongtian Yu, Lingxi Xie, Yaowei Wang, Qixiang Ye, and Yunfan Liu. Vmamba: Visual state space model. arXiv preprint arXiv:2401.10166, 2024. 2, 3, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.519, + 0.892, + 0.574 + ], + "angle": 0, + "content": "[22] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, pages 10012-10022, 2021. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.576, + 0.892, + 0.645 + ], + "angle": 0, + "content": "[23] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.647, + 0.894, + 0.714 + ], + "angle": 0, + "content": "[24] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11976-11986, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.717, + 0.894, + 0.758 + ], + "angle": 0, + "content": "[25] Jun Ma, Feifei Li, and Bo Wang. U-mamba: Enhancing long-range dependency for biomedical image segmentation. arXiv preprint arXiv:2401.04722, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.76, + 0.892, + 0.801 + ], + "angle": 0, + "content": "[26] Harsh Mehta, Ankit Gupta, Ashok Cutkosky, and Behnam Neyshabur. Long range language modeling via gated state spaces. arXiv preprint arXiv:2206.13947, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.803, + 0.894, + 0.858 + ], + "angle": 0, + "content": "[27] Arnab Kumar Mondal, Siba Smarak Panigrahi, Oumar Kaba, Sai Rajeswar Mudumba, and Siamak Ravanbakhsh. Equivariant adaptation of large pretrained models. Advances in Neural Information Processing Systems, 36:50293-50309, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.504, + 0.86, + 0.892, + 0.901 + ], + "angle": 0, + "content": "[28] Andrew Ng, Michael Jordan, and Yair Weiss. On spectral clustering: Analysis and an algorithm. Advances in neural information processing systems, 14, 2001. 6" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "23889" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.092, + 0.47, + 0.147 + ], + "angle": 0, + "content": "[29] Omri Puny, Matan Atzmon, Heli Ben-Hamu, Ishan Misra, Aditya Grover, Edward J Smith, and Yaron Lipman. Frame averaging for invariant and equivariant network design. arXiv preprint arXiv:2110.03336, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.081, + 0.149, + 0.471, + 0.19 + ], + "angle": 0, + "content": "[30] Martin Reuter, Franz-Erich Wolter, and Niklas Peinecke. Laplace-beltrami spectra as 'shape-dna' of surfaces and solids. Computer-Aided Design, 38(4):342-366, 2006. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.192, + 0.471, + 0.234 + ], + "angle": 0, + "content": "[31] Sara Sabour, Nicholas Frosst, and Geoffrey E Hinton. Dynamic routing between capsules. In Advances in Neural Information Processing Systems (NeurIPS), 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.236, + 0.471, + 0.275 + ], + "angle": 0, + "content": "[32] Jianbo Shi and Jitendra Malik. Normalized cuts and image segmentation. IEEE Transactions on pattern analysis and machine intelligence, 22(8):888-905, 2000. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.278, + 0.471, + 0.319 + ], + "angle": 0, + "content": "[33] Yuheng Shi, Minjing Dong, and Chang Xu. Multi-scale vmamba: Hierarchy in hierarchy visual state space model. arXiv preprint arXiv:2405.14174, 2024. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.321, + 0.471, + 0.374 + ], + "angle": 0, + "content": "[34] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. In International conference on learning representations, 2014. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.377, + 0.471, + 0.418 + ], + "angle": 0, + "content": "[35] Jimmy TH Smith, Andrew Warrington, and Scott W Linderman. Simplified state space layers for sequence modeling. arXiv preprint arXiv:2208.04933, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.421, + 0.471, + 0.473 + ], + "angle": 0, + "content": "[36] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In International conference on machine learning, pages 6105-6114. PMLR, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.477, + 0.471, + 0.545 + ], + "angle": 0, + "content": "[37] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In Int. Conf. Machine Learning, pages 10347-10357, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.548, + 0.471, + 0.615 + ], + "angle": 0, + "content": "[38] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In International conference on machine learning, pages 10347-10357. PMLR, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.618, + 0.471, + 0.673 + ], + "angle": 0, + "content": "[39] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.675, + 0.471, + 0.728 + ], + "angle": 0, + "content": "[40] Oriol Vinyals, Charles Blundell, Timothy Lillicrap, Daan Wierstra, et al. Matching networks for one shot learning. Advances in neural information processing systems, 29, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.732, + 0.471, + 0.772 + ], + "angle": 0, + "content": "[41] Maurice Weiler and Gabriele Cesa. General E(2)-Equivariant Steerable CNNs. In Conference on Neural Information Processing Systems (NeurIPS), 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.774, + 0.471, + 0.829 + ], + "angle": 0, + "content": "[42] Maurice Weiler, Fred A. Hamprecht, and Martin Storath. Learning steerable filters for rotation equivariant cnns. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 849-858, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.831, + 0.471, + 0.871 + ], + "angle": 0, + "content": "[43] Fei Xie, Weijia Zhang, Zhongdao Wang, and Chao Ma. Quadmamba: Learning quadtree-based selective scan for visual state space model. arXiv preprint arXiv:2410.06806, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.08, + 0.874, + 0.471, + 0.901 + ], + "angle": 0, + "content": "[44] Xiaosong Zhang, Yunjie Tian, Lingxi Xie, Wei Huang, Qi Dai, Qixiang Ye, and Qi Tian. Hivit: A simpler and more efficient" + }, + { + "type": "list", + "bbox": [ + 0.08, + 0.092, + 0.471, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.535, + 0.093, + 0.894, + 0.133 + ], + "angle": 0, + "content": "design of hierarchical vision transformer. In The Eleventh International Conference on Learning Representations, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.136, + 0.894, + 0.19 + ], + "angle": 0, + "content": "[45] Yanzhao Zhou, Qixiang Ye, Qiang Qiu, and Jianbin Jiao. Oriented response networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 519-528, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.503, + 0.192, + 0.894, + 0.247 + ], + "angle": 0, + "content": "[46] Lianghui Zhu, Bencheng Liao, Qian Zhang, Xinlong Wang, Wenyu Liu, and Xinggang Wang. Vision mamba: Efficient visual representation learning with bidirectional state space model. arXiv preprint arXiv:2401.09417, 2024. 2, 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.503, + 0.093, + 0.894, + 0.247 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "23890" + } + ] +] \ No newline at end of file diff --git a/2025/Spectral State Space Model for Rotation-Invariant Visual Representation Learning/da806762-c7b4-4ed7-80ed-cc0683ab510c_origin.pdf b/2025/Spectral State Space Model for Rotation-Invariant Visual Representation Learning/da806762-c7b4-4ed7-80ed-cc0683ab510c_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..fc14ccb581408a3ea430321556646df587af4177 --- /dev/null +++ b/2025/Spectral State Space Model for Rotation-Invariant Visual Representation Learning/da806762-c7b4-4ed7-80ed-cc0683ab510c_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c83f9e3dc2b58cadd6fa93964e1719d7d2e6397672607e08e5b9ec7c7c287813 +size 3153445 diff --git a/2025/Spectral State Space Model for Rotation-Invariant Visual Representation Learning/full.md b/2025/Spectral State Space Model for Rotation-Invariant Visual Representation Learning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..5a07bd7fdec18ea26bce192e4e5df9fc49767a8b --- /dev/null +++ b/2025/Spectral State Space Model for Rotation-Invariant Visual Representation Learning/full.md @@ -0,0 +1,303 @@ +# Spectral State Space Model for Rotation-Invariant Visual Representation Learning + +Sahar Dastani $^{1,2*}$ Ali Bahri $^{1}$ Moslem Yazdanpanah $^{1}$ Mehrdad Noori $^{1}$ David Osowiechi $^{1}$ Gustavo Adolfo Vargas Hakim $^{1}$ Farzad Beizaee $^{1}$ Milad Cheraghalikhani $^{1}$ Arnab Kumar Mondal $^{3\dagger}$ Herve Lombaert $^{2,4}$ Christian Desrosiers $^{1}$ $^{1}$ LIVIA, ILLS, ÉTS Montréal, Canada, $^{2}$ Mila - Quebec AI Institute, $^{3}$ Apple, $^{4}$ Polytechnique Montreal + +# Abstract + +State Space Models (SSMs) have recently emerged as an alternative to Vision Transformers (ViTs) due to their unique ability of modeling global relationships with linear complexity. SSMs are specifically designed to capture spatially proximate relationships of image patches. However, they fail to identify relationships between conceptually related yet not adjacent patches. This limitation arises from the non-causal nature of image data, which lacks inherent directional relationships. Additionally, current vision-based SSMs are highly sensitive to transformations such as rotation. Their predefined scanning directions depend on the original image orientation, which can cause the model to produce inconsistent patch-processing sequences after rotation. To address these limitations, we introduce Spectral VMamba, a novel approach that effectively captures the global structure within an image by leveraging spectral information derived from the graph Laplacian of image patches. Through spectral decomposition, our approach encodes patch relationships independently of image orientation, achieving patch traversal rotation invariance with our Rotational Feature Normalizer (RFN) module. Our experiments on classification tasks show that Spectral VMamba outperforms the leading SSM models in vision, such as VMamba, while maintaining invariance to rotations and a providing a similar runtime efficiency. The implementation is available at: https://github.com/Sahardastani/spectral_vmamba.git. + +# 1. Introduction + +Visual representation learning is fundamental to computer vision, where precise feature extraction is essential for tasks such as image classification, detection, and segmentation. The ability to extract meaningful patterns directly impacts + +![](images/6a7b98e5cf954a278021cc10f1816ed6025156841f4bb4fcff9c30a87c0c7441.jpg) +Figure 1. Effect of image rotation on patch processing in VMamba and Spectral VMamba networks. In VMamba networks (first row), patches are traversed using predefined horizontal and vertical scanning routes. As a result, rotating the image by $90^{\circ}$ significantly changes the sequence in which patches are processed. In contrast, Spectral VMamba (second row) organizes patches using spectral information derived from a graph laplacian of the image patches. As shown in the second row, Spectral VMamba consistently processes background patches before snake patches in both the original and rotated images. The snake image in this analysis is AI-generated. + +performance and unlocks deeper insights into visual tasks. Significant advances in representation learning have been driven by the development of Convolutional Neural Networks (CNNs) [14, 16, 24, 34, 36] and Vision Transformers (ViTs) [7, 23, 38, 44]. Although CNNs are effective at capturing local patterns with linear scaling, they struggle to + +efficiently model global relationships. ViTs address this limitation by introducing self-attention mechanisms, which capture global dependencies by analyzing relationships between all input elements [7, 39]. However, the quadratic complexity of ViTs poses challenges for large-scale data processing, as computational costs increase with input size. + +State Space Models (SSMs) [8, 11, 35] have recently gained attention as a compelling alternative to traditional architectures, offering global receptive fields with linear scalability. Inspired by the achievements of Mamba [10] in language modeling, recent studies [13, 21, 33, 46] have explored the application of linearly complex global receptive fields to the visual domain. However, unlike textual data, image pixels do not have a sequential dependency. Building on the parallelized selective scan operation in Mamba, known as S6 [10], which processes sequential one-dimensional data, VMamba [21] introduced 2D Selective Scan (SS2D), a novel four-way scanning mechanism tailored for spatial domain traversal. Vision Mamba (ViM) [46] further enhanced sequence modeling by employing a bidirectional processing strategy that sums both forward and reverse passes. This process enables the model to capture comprehensive dependencies across an image. Expanding upon these approaches, Multi-Scale VMamba (MSVMamba) [33] incorporates a multiscale 2D scanning technique, which processes both original and down-sampled feature maps. This strategy improves capturing long-range dependency while minimizing computational costs. Finally, MambaVision [13] introduces a hybrid approach. Early stages employ CNNs for fast feature extraction and later stages use Mamba and Transformer blocks to capture both local and global visual features. + +Vision-based SSMs are primarily designed to capture relationships between spatially adjacent patches, focusing on spatial proximity or neighborhood-based interactions. However, SSMs are not inherently equipped to identify relationships between patches that are conceptually related but not spatially adjacent. This limitation arises due to the non-causal nature of image data, where pixels lack inherent temporal order or directional relationships. + +In addition, current vision-based SSMs tend to exhibit high sensitivity to spatial information. For instance, VMamba uses predefined traversal routes that cause the network to indiscriminately switch between foreground and background regions within an image. This approach makes the network vulnerable to perturbations and leads to confusion in assessing feature importance. By treating all regions equally, VMamba fails to prioritize the features most relevant to the task, particularly when the foreground contains critical information. This can lead to poor performance when analyzing images with complex compositions. + +Another notable challenge emerges when SSMs attempt to process images that undergo transformations, such as rotation. Vision-based SSMs typically process images by (1) un + +folding image patches into sequences, and then (2) scanning these sequences in specific directions, such as horizontally and vertically. The sequence of image patches depends on the image orientation, meaning that rotating the image alters the order of the sequence. Similarly, the predefined scanning directions, which are fixed to the image original orientation, result in different scan paths when the image is rotated. Consequently, SS2D struggle to accommodate transformations such as rotation. Consider Figure 1 (first row): when the image is rotated by 90 degrees, the predefined and fixed traversal mechanism of VMamba significantly alters the sequence in which patches are processed. + +To address these limitations, we propose Spectral VMamba, which organizes input patches based on the spectral information derived from a graph Laplacian of image patches. By leveraging the eigenvectors from spectral decomposition of an affinity matrix, our approach captures the global structure and relationships within the image. This eigenvector-driven traversal clusters similar patches together, even if they are not spatially adjacent, thereby preventing indiscriminate transitions between foreground and background regions. Moreover, spectral decomposition encodes patch relationships independently of image orientation, enabling the extraction of rotation-invariant features. Hence, our approach defines relationships between patches by their content and mutual associations rather than by their spatial arrangement. Figure 1 (second row) demonstrates that rotation does not affect the order in which patches are processed. As we can see, the patches representing the background are processed first, followed by those representing the foreground (the snake in Fig. 1). + +Our main contributions are summarized as follow: + +- We propose a novel technique named Spectral VMamba based on spectral graph analysis for traversing image patches. By leveraging the eigenstructure of graphs that model the relationships between patches, our approach ensures that spatially-coherent structures in image data are maintained. Our method also improves the performance of SSMs in capturing and interpreting complex data patterns. +- We introduce the Rotational Feature Normalizer (RFN) module, designed to enhance our model robustness to isometric transformations, such as rotations. +- Our method demonstrates a notable improvement on the image classification task, outperforming the current state-of-the-art State Space Model (SSM)-based model in computer vision. + +# 2. Related Work + +State Space Models (SSMs). Transformers have significantly improved the ability to manage long-range dependencies. However, their self-attention mechanism is computationally expensive, especially for high-resolution images, due to the quadratic scaling with input size. To address these + +![](images/9a1854732eefe77a4be2ffe96affb2c6fcddf37acd86be4fce43d7996b072450.jpg) +Figure 2. The Spectral Traversal Scan (STS) architecture begins by representing image patches as a graph. The Laplacian spectrum of this graph is then generated, and its eigenvectors are computed. The traversal path for the image patches is determined by the order of eigenvectors. The 1st eigenvector traverses the foreground/background regions, the 2nd eigenvector focuses on the next most salient structures, and each subsequent eigenvector progressively captures finer details in the image. + +challenges, SSMs have emerged as an effective tool for long sequence modeling [11, 12, 26, 35]. In particular, the structured state-space sequence (S4) [11] model stands out for its efficient handling of long-range dependencies through a diagonal parameterization, which alleviates computational issues seen in earlier models. + +Recent advancements in sequence modeling have been driven by innovations in parallel scanning techniques, inspired by models such as S4, S5 [35] and the H3 model [8]. Mamba [10] improved the S4 model by integrating a data-dependent selection mechanism. This mechanism allows Mamba to achieve linear scalability, ensuring robust performance while accommodating longer sequence lengths without the exponential increase in resource demands. Such improvements are essential for applications requiring real-time or large-scale sequence processing, including Natural Language Processing (NLP) and time-series forecasting. + +SSMs in vision. The success of Mamba models in the NLP domain has stimulated significant progress in computer vision, leading to various adaptations tailored for visual data processing. Several studies have extended Mamba architectures to address specific challenges in vision tasks. For instance, VMamba [21] introduced an SS2D module to tackle direction sensitivity between non-causal 2D images and ordered 1D sequences. LocalMamba [17] focuses on localized feature extraction, utilizing state-space layers to enhance neighborhood interactions and reduce dependence on self-attention, leading to efficient high-resolution image analysis. QuadMamba [43] uses a quadrilateral traversal mechanism to manage directional dependencies in 2D images, integrating features from different image segments. + +Hybrid architectures. Recent work has also focused on combining SSMs with traditional CNNs or ViTs, leveraging the respective strengths of these models. For example, U-Mamba [25] merges CNNs and state space models to address + +long-range dependencies in biomedical image segmentation. MambaVision [13] integrates Mamba architectures with convolutional neural networks to improve performance across various vision tasks by leveraging both sequential modeling and spatial feature extraction. Finally, the TranS4mer model [35] integrates S4 with self-attention mechanisms to achieve state-of-the-art performance in movie scene detection. These advancements collectively demonstrate the versatility and scalability of structured state space models across various applications in the vision domain. + +Rotation-invariance in vision. Numerous approaches have been developed to embed equivariance and invariance directly within neural network architectures. Group-equivariant convolutional networks explicitly encode rotation symmetry into their design, providing intrinsic equivariance to rotations through steerable or group-convolutional filters [3, 4, 41, 42]. Learned canonicalization methods, such as Spatial Transformer Networks, address invariance by predicting transformations to consistently align input images into a canonical orientation, effectively standardizing pose variability [18, 19, 27]. Alternatively, symmetrization strategies explicitly aggregate feature responses across multiple rotated inputs or filters, achieving invariance by pooling orientation-specific activations [6, 20, 29, 45]. Additionally, capsule networks explicitly represent the pose of the object, leveraging pose-aware activations to enhance robustness against rotations [15, 31]. Within this landscape, our current work aligns with learned canonicalization approaches by leveraging spectral decomposition to canonicalize patch ordering and achieve patch traversal rotation invariance. + +# 3. Method + +This section begins with an overview of key concepts in SSMs and spectral graph analysis in Section 3.1, forming the foundation for our approach. Next, in Section 3.2, we in- + +troduce our proposed architecture, Spectral VMamba, along with the RFN module designed to achieve rotation invariance (Section 3.3). Finally, we discuss the robustness of our method to rotational variations in Section 3.6. + +# 3.1. Preliminaries + +SSM formulation. SSM-based models [10, 11] map a one-dimensional function or an input sequence $x(t) \in \mathbb{R}$ to an output sequence $y(t) \in \mathbb{R}$ through a learnable hidden state $\mathbf{h}(t) \in \mathbb{R}^N$ . This mapping follows a continuous dynamical system formulated as + +$$ +\frac {d}{d t} \mathbf {h} (t) = \mathbf {A h} (t) + \mathbf {B} x (t), \quad y (t) = \mathbf {C h} (t) \qquad (1) +$$ + +where $\mathbf{A} \in \mathbb{R}^{N \times N}$ , $\mathbf{B} \in \mathbb{R}^{N \times 1}$ and $\mathbf{C} \in \mathbb{R}^{1 \times N}$ are parameters governing the evolution of system. + +To adapt continuous SSMs in deep learning frameworks, discretization is required. This can be achieved using the zero-order hold (ZOH) rule with a time scale parameter $\Delta \in \mathbb{R}$ . Based on this rule, the discretized versions of $\mathbf{A}$ and $\mathbf{B}$ are defined as + +$$ +\begin{array}{l} \overline {{\mathbf {A}}} = \exp (\Delta \mathbf {A}), \\ - \end{array} \tag {2} +$$ + +$$ +\overline {{\mathbf {B}}} = (\Delta \mathbf {A}) ^ {- 1} \left(\exp (\Delta \mathbf {A}) - \mathbf {I}\right) \Delta \mathbf {B} \approx \Delta \mathbf {B} +$$ + +and the system becomes + +$$ +\mathbf {h} (t) = \overline {{\mathbf {A}}} \mathbf {h} (t - 1) + \overline {{\mathbf {B}}} x (t), \quad y (t) = \mathbf {C h} (t). \tag {3} +$$ + +As $\overline{\mathbf{A}}$ and $\overline{\mathbf{B}}$ are linear operators, the auto-regressive formulation of Eq. (3), computing the output at each time step $t$ in a recurrent manner, can be unrolled and expressed as single convolution with kernel $\overline{\mathbf{K}}\in \mathbb{R}^{L}$ : + +$$ +\mathbf {y} = \mathbf {x} \odot \overline {{\mathbf {K}}}, \quad \overline {{\mathbf {K}}} = \left(\mathbf {C} \overline {{\mathbf {B}}}, \mathbf {C} \overline {{\mathbf {A B}}}, \dots , \mathbf {C} \overline {{\mathbf {A}}} ^ {L - 1} \overline {{\mathbf {B}}}\right). \tag {4} +$$ + +During training, this convolutional can be employed to compute all output values in parallel. + +Spectral Graph Analysis studies the properties of a graph by analyzing the eigen-spectrum of its Laplacian matrix, which can be seen as a discrete version of the continuous Laplace-Beltrami operator obtained by a finite difference method [30]. Given a real-valued function $f \in C^2$ on a Riemannian manifold $\mathcal{M}$ , the Laplace-Beltrami operator $\Delta$ of $f$ is defined as $\Delta f = \text{div}(\text{grad} f)$ where with $\text{grad} f$ is the gradient of $f$ and $\text{div}$ denotes the divergence. + +The Laplacian Eigenvalue Problem corresponds to solving the Helmholtz equation $\Delta f = -\lambda f$ , whose solutions (eigenfunctions) are directly related to the vibration of a physical membrane [2]. Following Courant's Nodal Line Theorem [5], the Laplacian eigen-spectrum is composed of non-negative eigenvalues $0 \leq \lambda_1 \leq \lambda_2 \leq \ldots$ , each one repeating according to its multiplicity. Moreover, the + +eigenfunction corresponding to $\lambda_{n}$ has at most $n$ poles of vibration, hence smaller eigenvalues and their associated eigenvector encode lower frequency structural information which captures relationships at a more global scale. + +The Laplacian matrix of a graph $G = (V, E)$ with nodes $V$ and edges $E$ is defined as $\mathbf{L} = \mathbf{D} - \mathbf{W}$ , where $\mathbf{D}$ is the diagonal degree matrix with entries $D_{ii} = \sum_{j} W_{ij}$ . A normalized version of the Laplacian is typically employed to address differences in the weight scales $W_{ij}$ and variations in the distribution of node degrees $D_{ii}$ . In this work, we utilize the symmetric normalized Laplacian, defined as + +$$ +\mathbf {L} _ {\text {s y m}} = \mathbf {I} - \mathbf {D} ^ {- \frac {1}{2}} \mathbf {W D} ^ {- \frac {1}{2}}, \tag {5} +$$ + +which is positive semi-definite and possesses $|V|$ non-negative real-valued eigenvalues $0 = \lambda_{1} \leq \ldots \leq \lambda_{|V|}$ . + +A key property of the Laplace eigen-spectrum is that it is invariant to transformations preserving distance (i.e., isometry), which include translations and rotations. In our work, we use this property to design an SSM that enhances rotational robustness, where the order in which patches are traversed is determined based on the eigenvectors corresponding to the smallest eigenvalues of the Laplacian matrix. + +# 3.2. Spectral VMamba + +Existing vision-based Mamba methods, such as VMamba [21], are based on SS2D, a four-way scanning mechanism designed for spatial domain traversal. SS2D organizes input patches along four predefined scanning routes. However, two key challenges emerge: (1) connections between semantically related but non- neighboring patches are overlooked; and (2) scanning is sensitive to isometric transformations, such as rotation. To address these limitations, we introduce the Spectral VMamba model whose architecture is summarized in Figure 3. This architecture consists of two main components: a RFN module ensuring consistent feature representation across orientations and a sequence of Spectral VMamba blocks forming the network body. + +Data forwarding in our Spectral VMamba block comprises three key steps: Spectral Traversal Scan (STS), S6 block (selective scan), and Spectral Traversal Merge (STM). In the STS step, input patches are reordered based on the Laplacian spectrum of their underlying graph structure, ensuring traversal aligns with spectral coherence (see Section 3.4). The selective scan step then processes each reordered patch sequence in parallel using dedicated S6 blocks, effectively capturing localized patterns. Finally, in the STM step, the reordered traversal paths are restored to their original spatial format, merging the resulting sequences to generate the final output map. The detailed architecture of STS is depicted in Figure 2. The sequence of patch processing is determined by the order of eigenvectors derived from the spectral decomposition of the graph Laplacian. It is important to note that STS is only performed once for the initial layer. + +![](images/60c5b827b37c1191418304ec6b87b25d4419ce133099069e87a45111f0b3974d.jpg) +Figure 3. An overview of the proposed method. Our network consists of two primary components: the RFN module, which normalizes feature orientation to ensure consistency across different perspectives, and a series of Spectral VMamba blocks that process data in three stages: (1) Spectral Traversal Scan (STS), (2) S6 block (selective scan), and (3) Spectral Traversal Merge (STM). In STS, patches are reordered based on spectral coherence derived from the Laplacian spectrum. Each sequence is then processed in parallel by dedicated S6 blocks, capturing localized patterns, and finally merged in STM to reconstruct the spatial layout, producing the final feature map. + +In subsequent layers, we apply a downsampling strategy for generating the traversal path. For a detailed explanation of the downsampling process, please refer to the supplementary material. + +The graph Laplacian eigen-spectrum is isometric invariant, guaranteeing that the order in which patches are traversed is not affected by changes to their position in the image. However, the values in the Laplacian matrix are computed using patch features which are themselves sensitive to rotation of the patch. A simple solution to this problem is to use patches containing a single pixel, however this would be impractical for two reasons. First, individual pixels encode little information (RGB color) and thus using their similarity to construct the Laplacian matrix would not be useful. More importantly, this would severely increase the computational cost of downstream operations such as the spectral decomposition of the Laplacian. To address this issue, we compute patch features using our RFN module (Section 3.3) designed to normalize the feature extraction process by aggregating features from multiple orientations of the input image. This approach allows for effectively handling of feature changes during rotations, preserving the desired order and mitigating orientation-dependent discrepancies. + +# 3.3. Rotational Feature Normalizer (RFN) + +An overview of the module is illustrated in Figure 3. This module rotates the input image by a pre-determined set of angles $\{\theta_i,\dots ,\theta_R\}$ and processes them with a stem module which serves as a feature extraction backbone for patchification. This backbone partitions the input image $\mathbf{I}\in \mathbb{R}^{H\times W\times 3}$ into patches $\mathbf{x}\in \mathbb{R}_{\frac{H}{p}\times \frac{W}{p}\times C}$ , where $(H,W)$ denote the di + +mensions of the input image, $C$ is the number of feature channels, and $p$ represents the patch size. To consolidate features from all rotated versions, each feature map is rotated back to its original orientation. This unrotation ensures that all extracted features align with the original image coordinate system. Finally, an element-wise max operation is applied across the unrotated feature maps to generate a single aggregated feature. The final aggregated feature map can be expressed as: + +$$ +\mathbf {F} _ {i, j} = \max _ {r \in \{1, \dots , R \}} \left[ \mathcal {R} _ {- \theta_ {r}} \left(\operatorname {P a t c h i f y} \left(\mathcal {R} _ {\theta_ {r}} (\mathbf {I})\right)\right) \right] _ {i, j} \tag {6} +$$ + +where $\theta_r$ is the angle of the $r$ -th rotation and $\mathcal{R}_{\theta_r}$ is the transformation applying this rotation. + +The choice of rotation angles for the RFN module represents a trade-off between rotation invariance and computational overhead. Using a larger number of angles ensures a greater coverage in terms of rotation, hence improving the invariance of the model at the cost of increased computations. In our implementation, we used four canonical angles, $\{0^{\circ}, 90^{\circ}, 180^{\circ}, 270^{\circ}\}$ , corresponding to quarter turns. By rotating the image at these specific angles, the pixel grid alignment is preserved, eliminating the need for interpolation. This alignment not only simplifies the back-rotation process but also guarantees that the extracted features from all rotated images are comparable. + +# 3.4. Spectral Traversal Scan + +In this section, we describe how the graph Laplacian eigenspectrum is leveraged to define the traversal order of patches. + +We begin by representing the image as a graph $G = (V, E)$ , where each node $v_i \in V$ corresponds to a patch $\mathbf{x}_i$ . + +of the image, and edges $e_{ij} \in E$ connect pairs of nodes based on their spatial proximity. Following methods for spectral clustering [28] and normalized cuts [32], we consider a weighted adjacency matrix $\mathbf{W}$ of the graph, which is constructed using Euclidean distances between patches. Specifically, the weight $W_{ij}$ between the nodes $v_i$ and $v_j$ is given by + +$$ +W _ {i j} = \exp \left(- \frac {1}{2 \sigma^ {2}} \| \mathbf {f} _ {i} - \mathbf {f} _ {j} \| ^ {2}\right) \tag {7} +$$ + +where $\mathbf{f}_i$ is the feature vector of patch $\mathbf{x}_i$ and $\sigma$ is a scaling parameter determined as the average of the Euclidean distances between patches. We set $W_{ij} = 0$ for pairs $(v_i, v_j)$ that are not among the top- $k$ nearest neighbors of each other, resulting in a sparse adjacency matrix. + +We compute the first $m$ smallest eigenvectors of the symmetric Laplacian $\mathbf{L}_{\mathrm{sym}}$ using an iterative method known as the Arnoldi algorithm [9], which leverages the sparsity of the adjacency matrix to provide a low complexity. Each eigenvector $\mathbf{u}^{(j)} \in \mathbb{R}^{N_p}$ , where $j$ ranges from 1 to $m$ , assigns an eigenfunction value $\mathbf{u}_i^{(j)}$ to each image patch $i$ . In every Mamba block of our model, we perform two traversals per eigenvector: one in the order of increasing $\mathbf{u}_i^{(j)}$ values and another in decreasing order. At the end of the block, the resulting features from these $2m$ traversals are concatenated together. + +Patches $\mathbf{x}_i$ are ordered based on these eigenvectors, generating multiple traversal sequences: + +$$ +\mathbf {X} ^ {(j, \text {r e v})} = \left\{\mathbf {x} _ {i _ {1}} ^ {(j, \text {r e v})}, \mathbf {x} _ {i _ {2}} ^ {(j, \text {r e v})}, \dots , \mathbf {x} _ {i _ {m}} ^ {(j, \text {r e v})} \right\}, \text {w h e r e} \tag {8} +$$ + +$$ +i _ {k} ^ {(j, \operatorname {r e v})} = \arg \min _ {i \in \{1, \dots , m \}} \left\{\left(- 1\right) ^ {\operatorname {r e v}} \cdot u _ {i} ^ {(j)} \mid i \notin \left\{i _ {1}, \dots , i _ {k - 1} \right\} \right\} +$$ + +where $\operatorname{rev} \in \{0,1\}$ indicates if the traversal is reversed $(\operatorname{rev} = 1)$ or not $(\operatorname{rev} = 0)$ . This concatenated sequence $\mathbf{X}$ is then fed into the Mamba network for further processing, allowing the model to incorporate rich spectral relationships while preserving global coherence. + +# 3.5. Spectral Canonicalization + +Eigenvectors inherently have a sign ambiguity as, if $\mathbf{u}$ is an eigenvector of a matrix $\mathbf{A}$ with eigenvalue $\lambda$ , then $-\mathbf{u}$ is also an eigenvector with the same eigenvalue. To resolve this ambiguity, we standardize the sign of eigenvectors by ensuring that the first non-zero element of each eigenvector is positive. If the first element is negative, we flip the sign of the entire eigenvector. This adjustment does not alter the underlying structure captured by the eigenvectors, as flipping the sign consistently maintains the same eigenvalue $\lambda$ and preserves the relationships between eigenvectors. + +# 3.6. Rotation Invariance + +Spectral decomposition leverages the eigen-structure of the Laplacian matrix, which remains consistent regardless of + +rotation. As our traversal strategy is directly guided by the spectral decomposition of input features, the rotation-invariant nature of spectral decomposition extends to our method. This property is demonstrated in the following theorem. + +Theorem 3.1. Under the assumption that patch features $\mathbf{f}$ are invariant to rotation, i.e. $\mathbf{f}(\mathbf{I}) = \mathbf{f}(\mathcal{R}_{\theta}(\mathbf{I}))$ for any $\theta$ , our spectral traversal scan is also rotation-invariant. + +Proof. As rotations are a type of isometry, we can prove the more general property of isometric invariance. This invariance derives from two properties: 1) the nearest neighbors and values of the weighted adjacency matrix are defined based on distance alone, i.e., $W_{ij} = \exp \left(-\frac{1}{2\sigma}\right)^2\| \mathbf{f}_i - \mathbf{f}_j\|^2$ , thus they do not change if distances remain the same; 2) the eigen-spectrum is invariant under cyclic permutation. This second property can be demonstrated as follows. Let $\mathbf{P}$ be a permutation matrix and $\mathbf{u}$ be an eigenvector of Laplacian matrix $\mathbf{L}$ corresponding to eigenvalue $\lambda$ . Using $\mathbf{u}' = \mathbf{P}\mathbf{u}$ (thus $\mathbf{u} = \mathbf{P}^T\mathbf{u}'$ ) we have that + +$$ +\begin{array}{l} \mathbf {L} \mathbf {u} = \lambda \mathbf {u} \Rightarrow \mathbf {L P} ^ {T} \mathbf {u} ^ {\prime} = \lambda \mathbf {P} ^ {T} \mathbf {u} ^ {\prime} \\ \Rightarrow \mathbf {P L P} ^ {T} \mathbf {u} ^ {\prime} = \mathbf {P} (\lambda \mathbf {P} ^ {T} \mathbf {u} ^ {\prime}) \tag {9} \\ \Rightarrow \mathbf {P L P} ^ {T} \mathbf {u} ^ {\prime} = \lambda \mathbf {u} ^ {\prime} \\ \end{array} +$$ + +Hence, $\mathbf{Pu}$ is an eigenvector of the permuted Laplacian matrix $\mathbf{PLP}^T$ . Since we traverse patches by their eigenvector values (from smallest to largest or vice-versa), we will therefore get the same ordering regardless of the permutation. + +# 4. Experiments + +We start by comparing our method against recently-proposed ViTs and SSMs on an image classification task. We then present an ablation study to thoroughly investigate the impact of various components on the performance of our method. + +# 4.1. Image Classification + +We conduct classification experiments on the miniImageNet [40] dataset. This dataset includes 50,000 training images and 10,000 validation images across 100 categories. Following previous works [21, 22, 46], we train our models for 300 epochs with a batch size of 128 per GPU. The models are optimized using the AdamW optimizer with a momentum of 0.9. A cosine decay scheduler manages the learning rate, which starts at $5 \times 10^{-4}$ , alongside a weight decay of 0.05. Additionally, we apply an exponential moving average (EMA) to stabilize training. For input images of size $224 \times 224$ , our data augmentation techniques include color jittering, AutoAugment, random erasing, mixup, and cutmix. + +Table 1 compares our Spectral VMamba method against ViTs and SSM-based models on the miniImageNet dataset. As can be seen, our method outperforms all other approaches + +Table 1. Classification performance comparison on miniImageNet. All images are of size $224 \times 224$ . T, S, and B denote the tiny, small, and base scales, respectively. + +
ModelFLOPs (G)Top-1 (%)Top-5 (%)
Transformer-Based
DeiT-S [37]4.6G70.8389.74
DeiT-B [37]17.5G72.4390.14
Swin-T [22]4.5G83.2595.54
Swin-S [22]8.7G84.1095.53
Swin-B [22]15.4G82.7795.33
XCiT-S24 [1]9.2G85.7996.31
XCiT-M24 [1]16.2G86.8096.38
SSM-Based
Vim-T [46]1.5G67.3087.97
Vim-S [46]5.1G79.7093.20
LocalVim-T [17]1.5G82.1294.60
LocalVim-S [17]4.8G81.6893.63
MSVMamba-N [33]0.9G82.1695.10
MSVMamba-M [33]1.5G83.7295.81
MSVMamba-T [33]4.6G86.4896.43
VMamba-T [21]4.9G86.2596.64
VMamba-S [21]8.7G86.4896.79
VMamba-B [21]8.7G87.1796.96
Ours-T3.9G87.86 (+1.61)97.25 (+0.61)
Ours-S6.3G88.09 (+1.61)97.08 (+0.29)
Ours-B6.3G88.17 (+1.00)97.27 (+0.31)
+ +in both top-1 and top-5 EMA accuracy. Compared to VMamba, its closest competitor, our method improves top-1 accuracy by $1.61\%$ for the tiny and small models (T and S) and by $1\%$ for the base model (B), while requiring less FLOPs (see 4.2 for explanation). Moreover, our Spectral VMamba (B) model yields a $1.37\%$ higher top-1 accuracy than the strongest ViT-based model, XCiT-M24, and an improvement of $5.4\%$ over Swin-B, one of the most popular architectures based on ViT. + +# 4.2. Ablation Study + +Rotational invariance. To evaluate the robustness to rotation of our method trained on unrotated images, we evaluated its performance on test images rotated at angles from $0^{\circ}$ to $360^{\circ}$ with $15^{\circ}$ increments. As a reminder, our RFN module extracts and aggregates features for four canonical rotations: $\{0^{\circ}, 90^{\circ}, 180^{\circ}, 270^{\circ}\}$ . We therefore expect a greater robustness at those rotation angles, for which our method is fully invariant. + +As shown in Figure 4, our model with the RFN (blue line) maintains a high accuracy near $87\%$ at those canonical angles, demonstrating its rotational invariance property. In contrast, VMamba (orange line) suffers from a significant drop in performance for images rotated at $90^{\circ}$ ( $\sim 30\%$ drop), $180^{\circ}$ ( $\sim 20\%$ drop) and $270^{\circ}$ ( $\sim 30\%$ drop). Although our method does not guarantee full rotational invariance for non-canonical angles, we observe a remarkably consistent performance across all rotation angles, stabilizing at around $78\%$ in top-1 accuracy. On the other hand, the performance + +![](images/02e8a067c6b7c6ec97104fc4f505baf212546bc32b3ab321af1337e5599511c8.jpg) +Figure 4. Model performance comparison across various rotation angles. Our method demonstrates consistent accuracy across all rotation angles, while VMamba exhibits significant fluctuations. + +of VMamba fluctuates drastically, reaching accuracy values near $50\%$ at certain angles (e.g., $105^{\circ}$ or $255^{\circ}$ ). + +Impact of RFN. To assess the usefulness of our RFN module, we also tested an ablation variant of our method without this module (Ours w/o RFN in Figure 4). As mentioned before, while the Laplacian eigen-spectrum is isometric-invariant, and thus invariant to rotation, the Laplacian matrix is constructed using patch features which are themselves sensitive to rotation. As we expect, removing the RFN causes the performance to drop for all angles, this drop growing as we increase the rotation (away from $0^{\circ}$ or $360^{\circ}$ ). Interestingly, our method without the RFN module still achieves a higher accuracy than VMamba for angles between $60^{\circ}$ and $300^{\circ}$ , which represent severe rotations. This could be explained by the small size of patches $(16 \times 16)$ compared to the image, minimizing the impact of rotations (full invariance would be achieved for $1 \times 1$ patches). + +Impact of STS. Since the STS is a critical part of our Spectral VMamba method, we cannot remove it. To evaluate its impact, we instead add the RFN to VMamba, the resulting approach now only differing from our method by not having a STS. As shown in Figure 4, the VMamba + RFN model (green line) offers a low robustness to rotation, its accuracy similar to that of VMamba. These results demonstrate that improving rotational invariance at the patch level is not sufficient to achieve a high performance. + +Number of eigenvectors. In Figure 5, we assess the effect on performance of varying the number of eigenvectors used in the STS ( $m \in \{1,2,3,4\}$ ). As can be seen, accuracy improves consistently with more eigenvectors, peaking at $87.48\%$ for $m = 4$ . This suggests that eigenvectors offer complementary information and that their combination enhances the recognition of complex global structures. In the same plot, we also report the performance of the standard VMamba model and a variant using a random traversal strategy. As can be observed, the improvement provided by + +our STS over VMamba, for any number of eigenvectors, is greater than the one offered by the raster scan of VMamba over a random traversal. + +![](images/133a9b668575984eefb6abf1857bd45d68854cecddde40485323a43e56c3d175.jpg) +Figure 5. Model performance for different number of eigenvectors. Four eigenvectors exhibit the best performance. + +Number of nearest neighbors. The K-Nearest Neighbors (KNN) algorithm constructs a sparse adjacency matrix by selecting the top $k$ nearest neighbors for each image patch. In Tab. 2, we report our model's performance across $k$ values from 5 to 25. As can be seen, our model is not very sensitive to this parameter and achieves its highest accuracy of $87.48\%$ for $k = 5$ neighbors. This suggests that only a few neighbors are necessary to capture the global structure of an image. This has a positive impact on the efficiency of our method as the computational complexity of the Laplacian eigen-decomposition step, using the Arnoldi algorithm, is proportional to the number of non-zero entries in the Laplacian matrix (which is directly proportional to $k$ ). + +
k510152025
Top-1 Accuracy (%)87.4886.9887.1486.9187.15
+ +Table 2. Model performance for different values of $k$ (number of nearest neighbors). The best performance is achieved at $k = 5$ . + +Memory Usage and Runtime. Next, we analyze the computational cost and time efficiency of the STS. As mentioned before, the proposed strategy minimizes computations by leveraging the sparsity of the Laplacian matrix and limiting the number of computed eigenvectors. In the following experiments, we use $m = 4$ eigenvectors and batch size = 1. Memory Usage: As depicted in Fig. 6, our STS approach (blue line) is highly efficient in terms of memory usage compared to the VMamba (orange line). Even when the number of patches increases, the STS based on a sparse eigen-solver, requires only a modest increase in memory compared to the VMamba. The stars along each line indicate the token lengths used in our method, highlighting the specific points where processing occurs across different sequence lengths. + +Runtime: Our STS strategy, which can run on a CPU, also shows a favorable runtime scaling with increased patch + +![](images/6e7566914c0176e2784e903068e1af7c19337fd073c00091bb4d92aae7a3d432.jpg) +Figure 6. Runtime and memory usage for different patches lengths. + +length (Fig. 6). Specifically, for a patch length of 196 (marked by a star), which is the setting used in our main experiments, the increase in runtime is minimal when compared to the VMamba backbone, highlighting the efficiency of our approach. + +FLOPs: The amount of FLOPs for our traversing strategy include the computations required for constructing the adjacency matrix, calculating the Laplacian, and obtaining its first eigenvectors. This calculation results in approximately 2 MB of FLOPs, which is significantly lower than the 3.9 GB FLOPs needed by VMamba's backbone. Interestingly, as reported in Tab. 1, the total number of FLOPs performed by our Spectral VMamba method is less than that of VMamba when using the same model size. This is because, after processing patches, the input to our Spectral VMamba has a size of $14 \times 14$ tokens, compared to $56 \times 56$ tokens for VMamba. + +# 5. Conclusion + +In this paper, we introduce Spectral VMamba, a novel approach to visual state space models that achieves patch traversal rotation invariance. Our method leverages spectral decomposition from the graph Laplacian of image patches to define traversal paths for patch processing. In addition, we develop the RFN module to ensure consistent feature representation in different orientations. This module serves as a foundation for our STS module, enabling patch features that are robust to rotational variations. By incorporating the RFN and STS modules, Spectral VMamba achieves a robust transformation-invariant feature extraction pipeline that maintains high accuracy across a variety of orientations. Our experiments demonstrate that Spectral VMamba consistently outperform popular models including Swin transformer, VMamba, MSVmamba, and LocalVim in image classification tasks. Our method does not yet address generalization to the medical domain. In future work, we aim to handle disconnected graph components, common in medical images with distinct anatomical structures or pathologies. Adapting spectral traversal for these regions will enhance pattern capture and robustness across imaging modalities. + +# Acknowledgement + +We appreciate the computational resources and support provided by Compute Canada and the Digital Research Alliance of Canada. + +# References + +[1] Alaaeldin Ali, Hugo Touvron, Mathilde Caron, Piotr Bojanowski, Matthijs Douze, Armand Joulin, Ivan Laptev, Natalia Neverova, Gabriel Synnaeve, Jakob Verbeek, et al. Xcit: Cross-covariance image transformers. NeurIPS, 34:2001-20027, 2021. 7 +[2] Alvin Bayliss, Charles I Goldstein, and Eli Turkel. An iterative method for the helmholtz equation. Journal of Computational Physics, 49(3):443-457, 1983. 4 +[3] Taco Cohen and Max Welling. Group equivariant convolutional networks. In International Conference on Machine Learning (ICML), pages 2990-2999, 2016. 3 +[4] Taco S Cohen and Max Welling. Steerable cnns. In International Conference on Learning Representations (ICLR), 2017. 3 +[5] Richard Courant and David Hilbert. Methods of mathematical physics: partial differential equations. Interscience Publishers, 1937. 4 +[6] Sander Dieleman, Jeffrey De Fauw, and Koray Kavukcuoglu. Exploiting cyclic symmetry in convolutional neural networks. International Conference on Machine Learning (ICML), 2016. 3 +[7] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 1, 2 +[8] Daniel Y Fu, Tri Dao, Khaled K Saab, Armin W Thomas, Atri Rudra, and Christopher Ré. Hungry hungry hippos: Towards language modeling with state space models. arXiv preprint arXiv:2212.14052, 2022. 2, 3 +[9] Gene H Golub and Charles F Van Loan. Matrix computations. JHU press, 2013. 6 +[10] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752, 2023. 2, 3, 4 +[11] Albert Gu, Karan Goel, and Christopher Ré. Efficiently modeling long sequences with structured state spaces. arXiv preprint arXiv:2111.00396, 2021. 2, 3, 4 +[12] Albert Gu, Isys Johnson, Karan Goel, Khaled Saab, Tri Dao, Atri Rudra, and Christopher Ré. Combining recurrent, convolutional, and continuous-time models with linear state space layers. Advances in neural information processing systems, 34:572-585, 2021. 3 +[13] Ali Hatamizadeh and Jan Kautz. Mambavision: A hybrid mamba-transformer vision backbone. arXiv preprint arXiv:2407.08083, 2024. 2, 3 +[14] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceed- + +ings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 1 +[15] Geoffrey E Hinton, Alex Krizhevsky, and Sida D Wang. Matrix capsules with em routing. In International Conference on Learning Representations (ICLR), 2018. 3 +[16] Gao Huang, Zhuang Liu, Laurens Van Der Maaten, and Kilian Q Weinberger. Densely connected convolutional networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4700-4708, 2017. 1 +[17] Tao Huang, Xiaohuan Pei, Shan You, Fei Wang, Chen Qian, and Chang Xu. Localmamba: Visual state space model with windowed selective scan. arXiv preprint arXiv:2403.09338, 2024. 3, 7 +[18] Max Jaderberg, Karen Simonyan, Andrew Zisserman, and Koray Kavukcuoglu. Spatial transformer networks. In Advances in Neural Information Processing Systems (NeurIPS), pages 2017-2025, 2015. 3 +[19] Sekou-Oumar Kaba, Arnab Kumar Mondal, Yan Zhang, Yoshua Bengio, and Siamak Ravanbakhsh. Equivariance with learned canonicalization functions. In International Conference on Learning Representations (ICLR), 2023. 3 +[20] Dmitry Laptev, Nikolay Savinov, Joachim M Buhmann, and Marc Pollefeys. Ti-pooling: transformation-invariant pooling for feature learning in convolutional neural networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 289-297, 2016. 3 +[21] Yue Liu, Yunjie Tian, Yuzhong Zhao, Hongtian Yu, Lingxi Xie, Yaowei Wang, Qixiang Ye, and Yunfan Liu. Vmamba: Visual state space model. arXiv preprint arXiv:2401.10166, 2024. 2, 3, 4, 6, 7 +[22] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, pages 10012-10022, 2021. 6, 7 +[23] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 1 +[24] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11976-11986, 2022. 1 +[25] Jun Ma, Feifei Li, and Bo Wang. U-mamba: Enhancing long-range dependency for biomedical image segmentation. arXiv preprint arXiv:2401.04722, 2024. 3 +[26] Harsh Mehta, Ankit Gupta, Ashok Cutkosky, and Behnam Neyshabur. Long range language modeling via gated state spaces. arXiv preprint arXiv:2206.13947, 2022. 3 +[27] Arnab Kumar Mondal, Siba Smarak Panigrahi, Oumar Kaba, Sai Rajeswar Mudumba, and Siamak Ravanbakhsh. Equivariant adaptation of large pretrained models. Advances in Neural Information Processing Systems, 36:50293-50309, 2023. 3 +[28] Andrew Ng, Michael Jordan, and Yair Weiss. On spectral clustering: Analysis and an algorithm. Advances in neural information processing systems, 14, 2001. 6 + +[29] Omri Puny, Matan Atzmon, Heli Ben-Hamu, Ishan Misra, Aditya Grover, Edward J Smith, and Yaron Lipman. Frame averaging for invariant and equivariant network design. arXiv preprint arXiv:2110.03336, 2021. 3 +[30] Martin Reuter, Franz-Erich Wolter, and Niklas Peinecke. Laplace-beltrami spectra as 'shape-dna' of surfaces and solids. Computer-Aided Design, 38(4):342-366, 2006. 4 +[31] Sara Sabour, Nicholas Frosst, and Geoffrey E Hinton. Dynamic routing between capsules. In Advances in Neural Information Processing Systems (NeurIPS), 2017. 3 +[32] Jianbo Shi and Jitendra Malik. Normalized cuts and image segmentation. IEEE Transactions on pattern analysis and machine intelligence, 22(8):888-905, 2000. 6 +[33] Yuheng Shi, Minjing Dong, and Chang Xu. Multi-scale vmamba: Hierarchy in hierarchy visual state space model. arXiv preprint arXiv:2405.14174, 2024. 2, 7 +[34] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. In International conference on learning representations, 2014. 1 +[35] Jimmy TH Smith, Andrew Warrington, and Scott W Linderman. Simplified state space layers for sequence modeling. arXiv preprint arXiv:2208.04933, 2022. 2, 3 +[36] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In International conference on machine learning, pages 6105-6114. PMLR, 2019. 1 +[37] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In Int. Conf. Machine Learning, pages 10347-10357, 2021. 7 +[38] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In International conference on machine learning, pages 10347-10357. PMLR, 2021. 1 +[39] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 2 +[40] Oriol Vinyals, Charles Blundell, Timothy Lillicrap, Daan Wierstra, et al. Matching networks for one shot learning. Advances in neural information processing systems, 29, 2016. 6 +[41] Maurice Weiler and Gabriele Cesa. General E(2)-Equivariant Steerable CNNs. In Conference on Neural Information Processing Systems (NeurIPS), 2019. 3 +[42] Maurice Weiler, Fred A. Hamprecht, and Martin Storath. Learning steerable filters for rotation equivariant cnns. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 849-858, 2018. 3 +[43] Fei Xie, Weijia Zhang, Zhongdao Wang, and Chao Ma. Quadmamba: Learning quadtree-based selective scan for visual state space model. arXiv preprint arXiv:2410.06806, 2024. 3 +[44] Xiaosong Zhang, Yunjie Tian, Lingxi Xie, Wei Huang, Qi Dai, Qixiang Ye, and Qi Tian. Hivit: A simpler and more efficient + +design of hierarchical vision transformer. In The Eleventh International Conference on Learning Representations, 2023. 1 +[45] Yanzhao Zhou, Qixiang Ye, Qiang Qiu, and Jianbin Jiao. Oriented response networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 519-528, 2017. 3 +[46] Lianghui Zhu, Bencheng Liao, Qian Zhang, Xinlong Wang, Wenyu Liu, and Xinggang Wang. Vision mamba: Efficient visual representation learning with bidirectional state space model. arXiv preprint arXiv:2401.09417, 2024. 2, 6, 7 \ No newline at end of file diff --git a/2025/Spectral State Space Model for Rotation-Invariant Visual Representation Learning/images.zip b/2025/Spectral State Space Model for Rotation-Invariant Visual Representation Learning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..fe8cb2e2999786b26657a4083262079adac05527 --- /dev/null +++ b/2025/Spectral State Space Model for Rotation-Invariant Visual Representation Learning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c43b77689cad507656ac22eb9173832cf82aa6e68d7c53736dad77f5d3187fe7 +size 463511 diff --git a/2025/Spectral State Space Model for Rotation-Invariant Visual Representation Learning/layout.json b/2025/Spectral State Space Model for Rotation-Invariant Visual Representation Learning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..bb412d922dbd8f9ef06c6d8d592eea92781630ee --- /dev/null +++ b/2025/Spectral State Space Model for Rotation-Invariant Visual Representation Learning/layout.json @@ -0,0 +1,9208 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 138, + 103, + 457, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 103, + 457, + 140 + ], + "spans": [ + { + "bbox": [ + 138, + 103, + 457, + 140 + ], + "type": "text", + "content": "Spectral State Space Model for Rotation-Invariant Visual Representation Learning" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 91, + 160, + 503, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 160, + 503, + 248 + ], + "spans": [ + { + "bbox": [ + 91, + 160, + 503, + 248 + ], + "type": "text", + "content": "Sahar Dastani" + }, + { + "bbox": [ + 91, + 160, + 503, + 248 + ], + "type": "inline_equation", + "content": "^{1,2*}" + }, + { + "bbox": [ + 91, + 160, + 503, + 248 + ], + "type": "text", + "content": " Ali Bahri" + }, + { + "bbox": [ + 91, + 160, + 503, + 248 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 91, + 160, + 503, + 248 + ], + "type": "text", + "content": " Moslem Yazdanpanah" + }, + { + "bbox": [ + 91, + 160, + 503, + 248 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 91, + 160, + 503, + 248 + ], + "type": "text", + "content": " Mehrdad Noori" + }, + { + "bbox": [ + 91, + 160, + 503, + 248 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 91, + 160, + 503, + 248 + ], + "type": "text", + "content": " David Osowiechi" + }, + { + "bbox": [ + 91, + 160, + 503, + 248 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 91, + 160, + 503, + 248 + ], + "type": "text", + "content": " Gustavo Adolfo Vargas Hakim" + }, + { + "bbox": [ + 91, + 160, + 503, + 248 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 91, + 160, + 503, + 248 + ], + "type": "text", + "content": " Farzad Beizaee" + }, + { + "bbox": [ + 91, + 160, + 503, + 248 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 91, + 160, + 503, + 248 + ], + "type": "text", + "content": " Milad Cheraghalikhani" + }, + { + "bbox": [ + 91, + 160, + 503, + 248 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 91, + 160, + 503, + 248 + ], + "type": "text", + "content": " Arnab Kumar Mondal" + }, + { + "bbox": [ + 91, + 160, + 503, + 248 + ], + "type": "inline_equation", + "content": "^{3\\dagger}" + }, + { + "bbox": [ + 91, + 160, + 503, + 248 + ], + "type": "text", + "content": " Herve Lombaert" + }, + { + "bbox": [ + 91, + 160, + 503, + 248 + ], + "type": "inline_equation", + "content": "^{2,4}" + }, + { + "bbox": [ + 91, + 160, + 503, + 248 + ], + "type": "text", + "content": " Christian Desrosiers" + }, + { + "bbox": [ + 91, + 160, + 503, + 248 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 91, + 160, + 503, + 248 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 91, + 160, + 503, + 248 + ], + "type": "text", + "content": "LIVIA, ILLS, ÉTS Montréal, Canada, " + }, + { + "bbox": [ + 91, + 160, + 503, + 248 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 91, + 160, + 503, + 248 + ], + "type": "text", + "content": "Mila - Quebec AI Institute, " + }, + { + "bbox": [ + 91, + 160, + 503, + 248 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 91, + 160, + 503, + 248 + ], + "type": "text", + "content": "Apple, " + }, + { + "bbox": [ + 91, + 160, + 503, + 248 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 91, + 160, + 503, + 248 + ], + "type": "text", + "content": "Polytechnique Montreal" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 274, + 192, + 285 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 274, + 192, + 285 + ], + "spans": [ + { + "bbox": [ + 143, + 274, + 192, + 285 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 299, + 290, + 610 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 299, + 290, + 610 + ], + "spans": [ + { + "bbox": [ + 47, + 299, + 290, + 610 + ], + "type": "text", + "content": "State Space Models (SSMs) have recently emerged as an alternative to Vision Transformers (ViTs) due to their unique ability of modeling global relationships with linear complexity. SSMs are specifically designed to capture spatially proximate relationships of image patches. However, they fail to identify relationships between conceptually related yet not adjacent patches. This limitation arises from the non-causal nature of image data, which lacks inherent directional relationships. Additionally, current vision-based SSMs are highly sensitive to transformations such as rotation. Their predefined scanning directions depend on the original image orientation, which can cause the model to produce inconsistent patch-processing sequences after rotation. To address these limitations, we introduce Spectral VMamba, a novel approach that effectively captures the global structure within an image by leveraging spectral information derived from the graph Laplacian of image patches. Through spectral decomposition, our approach encodes patch relationships independently of image orientation, achieving patch traversal rotation invariance with our Rotational Feature Normalizer (RFN) module. Our experiments on classification tasks show that Spectral VMamba outperforms the leading SSM models in vision, such as VMamba, while maintaining invariance to rotations and a providing a similar runtime efficiency. The implementation is available at: https://github.com/Sahardastani/spectral_vmamba.git." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 47, + 614, + 128, + 626 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 614, + 128, + 626 + ], + "spans": [ + { + "bbox": [ + 47, + 614, + 128, + 626 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 634, + 288, + 683 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 634, + 288, + 683 + ], + "spans": [ + { + "bbox": [ + 46, + 634, + 288, + 683 + ], + "type": "text", + "content": "Visual representation learning is fundamental to computer vision, where precise feature extraction is essential for tasks such as image classification, detection, and segmentation. The ability to extract meaningful patterns directly impacts" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 310, + 273, + 542, + 504 + ], + "blocks": [ + { + "bbox": [ + 310, + 273, + 542, + 504 + ], + "lines": [ + { + "bbox": [ + 310, + 273, + 542, + 504 + ], + "spans": [ + { + "bbox": [ + 310, + 273, + 542, + 504 + ], + "type": "image", + "image_path": "6a7b98e5cf954a278021cc10f1816ed6025156841f4bb4fcff9c30a87c0c7441.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 513, + 548, + 624 + ], + "lines": [ + { + "bbox": [ + 304, + 513, + 548, + 624 + ], + "spans": [ + { + "bbox": [ + 304, + 513, + 548, + 624 + ], + "type": "text", + "content": "Figure 1. Effect of image rotation on patch processing in VMamba and Spectral VMamba networks. In VMamba networks (first row), patches are traversed using predefined horizontal and vertical scanning routes. As a result, rotating the image by " + }, + { + "bbox": [ + 304, + 513, + 548, + 624 + ], + "type": "inline_equation", + "content": "90^{\\circ}" + }, + { + "bbox": [ + 304, + 513, + 548, + 624 + ], + "type": "text", + "content": " significantly changes the sequence in which patches are processed. In contrast, Spectral VMamba (second row) organizes patches using spectral information derived from a graph laplacian of the image patches. As shown in the second row, Spectral VMamba consistently processes background patches before snake patches in both the original and rotated images. The snake image in this analysis is AI-generated." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 641, + 548, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 641, + 548, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 641, + 548, + 715 + ], + "type": "text", + "content": "performance and unlocks deeper insights into visual tasks. Significant advances in representation learning have been driven by the development of Convolutional Neural Networks (CNNs) [14, 16, 24, 34, 36] and Vision Transformers (ViTs) [7, 23, 38, 44]. Although CNNs are effective at capturing local patterns with linear scaling, they struggle to" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 693, + 290, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 693, + 290, + 703 + ], + "spans": [ + { + "bbox": [ + 57, + 693, + 290, + 703 + ], + "type": "text", + "content": "\\*correspondence: sahar.dastani-oghani.1@ens.etsmtl.ca" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 60, + 703, + 216, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 703, + 216, + 712 + ], + "spans": [ + { + "bbox": [ + 60, + 703, + 216, + 712 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 60, + 703, + 216, + 712 + ], + "type": "text", + "content": " work done during author's PhD at Mila/McGill" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23881" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 288, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 288, + 144 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 288, + 144 + ], + "type": "text", + "content": "efficiently model global relationships. ViTs address this limitation by introducing self-attention mechanisms, which capture global dependencies by analyzing relationships between all input elements [7, 39]. However, the quadratic complexity of ViTs poses challenges for large-scale data processing, as computational costs increase with input size." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 144, + 290, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 144, + 290, + 443 + ], + "spans": [ + { + "bbox": [ + 46, + 144, + 290, + 443 + ], + "type": "text", + "content": "State Space Models (SSMs) [8, 11, 35] have recently gained attention as a compelling alternative to traditional architectures, offering global receptive fields with linear scalability. Inspired by the achievements of Mamba [10] in language modeling, recent studies [13, 21, 33, 46] have explored the application of linearly complex global receptive fields to the visual domain. However, unlike textual data, image pixels do not have a sequential dependency. Building on the parallelized selective scan operation in Mamba, known as S6 [10], which processes sequential one-dimensional data, VMamba [21] introduced 2D Selective Scan (SS2D), a novel four-way scanning mechanism tailored for spatial domain traversal. Vision Mamba (ViM) [46] further enhanced sequence modeling by employing a bidirectional processing strategy that sums both forward and reverse passes. This process enables the model to capture comprehensive dependencies across an image. Expanding upon these approaches, Multi-Scale VMamba (MSVMamba) [33] incorporates a multiscale 2D scanning technique, which processes both original and down-sampled feature maps. This strategy improves capturing long-range dependency while minimizing computational costs. Finally, MambaVision [13] introduces a hybrid approach. Early stages employ CNNs for fast feature extraction and later stages use Mamba and Transformer blocks to capture both local and global visual features." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 443, + 288, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 443, + 288, + 538 + ], + "spans": [ + { + "bbox": [ + 46, + 443, + 288, + 538 + ], + "type": "text", + "content": "Vision-based SSMs are primarily designed to capture relationships between spatially adjacent patches, focusing on spatial proximity or neighborhood-based interactions. However, SSMs are not inherently equipped to identify relationships between patches that are conceptually related but not spatially adjacent. This limitation arises due to the non-causal nature of image data, where pixels lack inherent temporal order or directional relationships." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 539, + 288, + 670 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 539, + 288, + 670 + ], + "spans": [ + { + "bbox": [ + 46, + 539, + 288, + 670 + ], + "type": "text", + "content": "In addition, current vision-based SSMs tend to exhibit high sensitivity to spatial information. For instance, VMamba uses predefined traversal routes that cause the network to indiscriminately switch between foreground and background regions within an image. This approach makes the network vulnerable to perturbations and leads to confusion in assessing feature importance. By treating all regions equally, VMamba fails to prioritize the features most relevant to the task, particularly when the foreground contains critical information. This can lead to poor performance when analyzing images with complex compositions." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 670, + 288, + 707 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 670, + 288, + 707 + ], + "spans": [ + { + "bbox": [ + 46, + 670, + 288, + 707 + ], + "type": "text", + "content": "Another notable challenge emerges when SSMs attempt to process images that undergo transformations, such as rotation. Vision-based SSMs typically process images by (1) un" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 304, + 72, + 547, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 215 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 215 + ], + "type": "text", + "content": "folding image patches into sequences, and then (2) scanning these sequences in specific directions, such as horizontally and vertically. The sequence of image patches depends on the image orientation, meaning that rotating the image alters the order of the sequence. Similarly, the predefined scanning directions, which are fixed to the image original orientation, result in different scan paths when the image is rotated. Consequently, SS2D struggle to accommodate transformations such as rotation. Consider Figure 1 (first row): when the image is rotated by 90 degrees, the predefined and fixed traversal mechanism of VMamba significantly alters the sequence in which patches are processed." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 216, + 548, + 442 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 216, + 548, + 442 + ], + "spans": [ + { + "bbox": [ + 304, + 216, + 548, + 442 + ], + "type": "text", + "content": "To address these limitations, we propose Spectral VMamba, which organizes input patches based on the spectral information derived from a graph Laplacian of image patches. By leveraging the eigenvectors from spectral decomposition of an affinity matrix, our approach captures the global structure and relationships within the image. This eigenvector-driven traversal clusters similar patches together, even if they are not spatially adjacent, thereby preventing indiscriminate transitions between foreground and background regions. Moreover, spectral decomposition encodes patch relationships independently of image orientation, enabling the extraction of rotation-invariant features. Hence, our approach defines relationships between patches by their content and mutual associations rather than by their spatial arrangement. Figure 1 (second row) demonstrates that rotation does not affect the order in which patches are processed. As we can see, the patches representing the background are processed first, followed by those representing the foreground (the snake in Fig. 1)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 317, + 443, + 522, + 454 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 443, + 522, + 454 + ], + "spans": [ + { + "bbox": [ + 317, + 443, + 522, + 454 + ], + "type": "text", + "content": "Our main contributions are summarized as follow:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 455, + 547, + 622 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 306, + 455, + 545, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 455, + 545, + 538 + ], + "spans": [ + { + "bbox": [ + 306, + 455, + 545, + 538 + ], + "type": "text", + "content": "- We propose a novel technique named Spectral VMamba based on spectral graph analysis for traversing image patches. By leveraging the eigenstructure of graphs that model the relationships between patches, our approach ensures that spatially-coherent structures in image data are maintained. Our method also improves the performance of SSMs in capturing and interpreting complex data patterns." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 306, + 539, + 545, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 539, + 545, + 574 + ], + "spans": [ + { + "bbox": [ + 306, + 539, + 545, + 574 + ], + "type": "text", + "content": "- We introduce the Rotational Feature Normalizer (RFN) module, designed to enhance our model robustness to isometric transformations, such as rotations." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 574, + 547, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 574, + 547, + 622 + ], + "spans": [ + { + "bbox": [ + 306, + 574, + 547, + 622 + ], + "type": "text", + "content": "- Our method demonstrates a notable improvement on the image classification task, outperforming the current state-of-the-art State Space Model (SSM)-based model in computer vision." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 306, + 629, + 392, + 642 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 629, + 392, + 642 + ], + "spans": [ + { + "bbox": [ + 306, + 629, + 392, + 642 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 653, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 653, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 653, + 547, + 713 + ], + "type": "text", + "content": "State Space Models (SSMs). Transformers have significantly improved the ability to manage long-range dependencies. However, their self-attention mechanism is computationally expensive, especially for high-resolution images, due to the quadratic scaling with input size. To address these" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "23882" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 76, + 542, + 218 + ], + "blocks": [ + { + "bbox": [ + 53, + 76, + 542, + 218 + ], + "lines": [ + { + "bbox": [ + 53, + 76, + 542, + 218 + ], + "spans": [ + { + "bbox": [ + 53, + 76, + 542, + 218 + ], + "type": "image", + "image_path": "9a1854732eefe77a4be2ffe96affb2c6fcddf37acd86be4fce43d7996b072450.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 228, + 548, + 275 + ], + "lines": [ + { + "bbox": [ + 46, + 228, + 548, + 275 + ], + "spans": [ + { + "bbox": [ + 46, + 228, + 548, + 275 + ], + "type": "text", + "content": "Figure 2. The Spectral Traversal Scan (STS) architecture begins by representing image patches as a graph. The Laplacian spectrum of this graph is then generated, and its eigenvectors are computed. The traversal path for the image patches is determined by the order of eigenvectors. The 1st eigenvector traverses the foreground/background regions, the 2nd eigenvector focuses on the next most salient structures, and each subsequent eigenvector progressively captures finer details in the image." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 283, + 288, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 283, + 288, + 354 + ], + "spans": [ + { + "bbox": [ + 46, + 283, + 288, + 354 + ], + "type": "text", + "content": "challenges, SSMs have emerged as an effective tool for long sequence modeling [11, 12, 26, 35]. In particular, the structured state-space sequence (S4) [11] model stands out for its efficient handling of long-range dependencies through a diagonal parameterization, which alleviates computational issues seen in earlier models." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 355, + 290, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 355, + 290, + 487 + ], + "spans": [ + { + "bbox": [ + 46, + 355, + 290, + 487 + ], + "type": "text", + "content": "Recent advancements in sequence modeling have been driven by innovations in parallel scanning techniques, inspired by models such as S4, S5 [35] and the H3 model [8]. Mamba [10] improved the S4 model by integrating a data-dependent selection mechanism. This mechanism allows Mamba to achieve linear scalability, ensuring robust performance while accommodating longer sequence lengths without the exponential increase in resource demands. Such improvements are essential for applications requiring real-time or large-scale sequence processing, including Natural Language Processing (NLP) and time-series forecasting." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 490, + 288, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 490, + 288, + 658 + ], + "spans": [ + { + "bbox": [ + 46, + 490, + 288, + 658 + ], + "type": "text", + "content": "SSMs in vision. The success of Mamba models in the NLP domain has stimulated significant progress in computer vision, leading to various adaptations tailored for visual data processing. Several studies have extended Mamba architectures to address specific challenges in vision tasks. For instance, VMamba [21] introduced an SS2D module to tackle direction sensitivity between non-causal 2D images and ordered 1D sequences. LocalMamba [17] focuses on localized feature extraction, utilizing state-space layers to enhance neighborhood interactions and reduce dependence on self-attention, leading to efficient high-resolution image analysis. QuadMamba [43] uses a quadrilateral traversal mechanism to manage directional dependencies in 2D images, integrating features from different image segments." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 662, + 289, + 711 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 662, + 289, + 711 + ], + "spans": [ + { + "bbox": [ + 47, + 662, + 289, + 711 + ], + "type": "text", + "content": "Hybrid architectures. Recent work has also focused on combining SSMs with traditional CNNs or ViTs, leveraging the respective strengths of these models. For example, U-Mamba [25] merges CNNs and state space models to address" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 283, + 547, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 283, + 547, + 403 + ], + "spans": [ + { + "bbox": [ + 304, + 283, + 547, + 403 + ], + "type": "text", + "content": "long-range dependencies in biomedical image segmentation. MambaVision [13] integrates Mamba architectures with convolutional neural networks to improve performance across various vision tasks by leveraging both sequential modeling and spatial feature extraction. Finally, the TranS4mer model [35] integrates S4 with self-attention mechanisms to achieve state-of-the-art performance in movie scene detection. These advancements collectively demonstrate the versatility and scalability of structured state space models across various applications in the vision domain." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 304, + 407, + 548, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 407, + 548, + 647 + ], + "spans": [ + { + "bbox": [ + 304, + 407, + 548, + 647 + ], + "type": "text", + "content": "Rotation-invariance in vision. Numerous approaches have been developed to embed equivariance and invariance directly within neural network architectures. Group-equivariant convolutional networks explicitly encode rotation symmetry into their design, providing intrinsic equivariance to rotations through steerable or group-convolutional filters [3, 4, 41, 42]. Learned canonicalization methods, such as Spatial Transformer Networks, address invariance by predicting transformations to consistently align input images into a canonical orientation, effectively standardizing pose variability [18, 19, 27]. Alternatively, symmetrization strategies explicitly aggregate feature responses across multiple rotated inputs or filters, achieving invariance by pooling orientation-specific activations [6, 20, 29, 45]. Additionally, capsule networks explicitly represent the pose of the object, leveraging pose-aware activations to enhance robustness against rotations [15, 31]. Within this landscape, our current work aligns with learned canonicalization approaches by leveraging spectral decomposition to canonicalize patch ordering and achieve patch traversal rotation invariance." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 306, + 657, + 362, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 657, + 362, + 670 + ], + "spans": [ + { + "bbox": [ + 306, + 657, + 362, + 670 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 677, + 548, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 548, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 548, + 715 + ], + "type": "text", + "content": "This section begins with an overview of key concepts in SSMs and spectral graph analysis in Section 3.1, forming the foundation for our approach. Next, in Section 3.2, we in-" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23883" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 120 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 120 + ], + "type": "text", + "content": "troduce our proposed architecture, Spectral VMamba, along with the RFN module designed to achieve rotation invariance (Section 3.3). Finally, we discuss the robustness of our method to rotational variations in Section 3.6." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 128, + 133, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 128, + 133, + 140 + ], + "spans": [ + { + "bbox": [ + 47, + 128, + 133, + 140 + ], + "type": "text", + "content": "3.1. Preliminaries" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 149, + 288, + 210 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 149, + 288, + 210 + ], + "spans": [ + { + "bbox": [ + 46, + 149, + 288, + 210 + ], + "type": "text", + "content": "SSM formulation. SSM-based models [10, 11] map a one-dimensional function or an input sequence " + }, + { + "bbox": [ + 46, + 149, + 288, + 210 + ], + "type": "inline_equation", + "content": "x(t) \\in \\mathbb{R}" + }, + { + "bbox": [ + 46, + 149, + 288, + 210 + ], + "type": "text", + "content": " to an output sequence " + }, + { + "bbox": [ + 46, + 149, + 288, + 210 + ], + "type": "inline_equation", + "content": "y(t) \\in \\mathbb{R}" + }, + { + "bbox": [ + 46, + 149, + 288, + 210 + ], + "type": "text", + "content": " through a learnable hidden state " + }, + { + "bbox": [ + 46, + 149, + 288, + 210 + ], + "type": "inline_equation", + "content": "\\mathbf{h}(t) \\in \\mathbb{R}^N" + }, + { + "bbox": [ + 46, + 149, + 288, + 210 + ], + "type": "text", + "content": ". This mapping follows a continuous dynamical system formulated as" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 78, + 218, + 288, + 243 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 218, + 288, + 243 + ], + "spans": [ + { + "bbox": [ + 78, + 218, + 288, + 243 + ], + "type": "interline_equation", + "content": "\\frac {d}{d t} \\mathbf {h} (t) = \\mathbf {A h} (t) + \\mathbf {B} x (t), \\quad y (t) = \\mathbf {C h} (t) \\qquad (1)", + "image_path": "057f1d0baec898e2795c6ed4d58cb0cd2c38256f611e40e96a6d52c1fe5db987.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 250, + 289, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 250, + 289, + 274 + ], + "spans": [ + { + "bbox": [ + 46, + 250, + 289, + 274 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 250, + 289, + 274 + ], + "type": "inline_equation", + "content": "\\mathbf{A} \\in \\mathbb{R}^{N \\times N}" + }, + { + "bbox": [ + 46, + 250, + 289, + 274 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 46, + 250, + 289, + 274 + ], + "type": "inline_equation", + "content": "\\mathbf{B} \\in \\mathbb{R}^{N \\times 1}" + }, + { + "bbox": [ + 46, + 250, + 289, + 274 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 250, + 289, + 274 + ], + "type": "inline_equation", + "content": "\\mathbf{C} \\in \\mathbb{R}^{1 \\times N}" + }, + { + "bbox": [ + 46, + 250, + 289, + 274 + ], + "type": "text", + "content": " are parameters governing the evolution of system." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 275, + 289, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 275, + 289, + 334 + ], + "spans": [ + { + "bbox": [ + 46, + 275, + 289, + 334 + ], + "type": "text", + "content": "To adapt continuous SSMs in deep learning frameworks, discretization is required. This can be achieved using the zero-order hold (ZOH) rule with a time scale parameter " + }, + { + "bbox": [ + 46, + 275, + 289, + 334 + ], + "type": "inline_equation", + "content": "\\Delta \\in \\mathbb{R}" + }, + { + "bbox": [ + 46, + 275, + 289, + 334 + ], + "type": "text", + "content": ". Based on this rule, the discretized versions of " + }, + { + "bbox": [ + 46, + 275, + 289, + 334 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 46, + 275, + 289, + 334 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 275, + 289, + 334 + ], + "type": "inline_equation", + "content": "\\mathbf{B}" + }, + { + "bbox": [ + 46, + 275, + 289, + 334 + ], + "type": "text", + "content": " are defined as" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 79, + 342, + 287, + 361 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 342, + 287, + 361 + ], + "spans": [ + { + "bbox": [ + 79, + 342, + 287, + 361 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\overline {{\\mathbf {A}}} = \\exp (\\Delta \\mathbf {A}), \\\\ - \\end{array} \\tag {2}", + "image_path": "b482289f99d6c7d8d2437a8b1f9e52a64ecb9cf945ce3b8630c9d8e76584d814.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 81, + 357, + 253, + 372 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 357, + 253, + 372 + ], + "spans": [ + { + "bbox": [ + 81, + 357, + 253, + 372 + ], + "type": "interline_equation", + "content": "\\overline {{\\mathbf {B}}} = (\\Delta \\mathbf {A}) ^ {- 1} \\left(\\exp (\\Delta \\mathbf {A}) - \\mathbf {I}\\right) \\Delta \\mathbf {B} \\approx \\Delta \\mathbf {B}", + "image_path": "2c93b69e3116359eda2867ba53d8112fc7b4fd09f87a9defab6e08cd1fc5e87d.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 380, + 147, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 380, + 147, + 392 + ], + "spans": [ + { + "bbox": [ + 47, + 380, + 147, + 392 + ], + "type": "text", + "content": "and the system becomes" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 74, + 401, + 287, + 415 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 401, + 287, + 415 + ], + "spans": [ + { + "bbox": [ + 74, + 401, + 287, + 415 + ], + "type": "interline_equation", + "content": "\\mathbf {h} (t) = \\overline {{\\mathbf {A}}} \\mathbf {h} (t - 1) + \\overline {{\\mathbf {B}}} x (t), \\quad y (t) = \\mathbf {C h} (t). \\tag {3}", + "image_path": "c695d7135082d0a1d640e7672eba83367b04a888c458a8575de1a040c39d206d.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 46, + 423, + 289, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 423, + 289, + 471 + ], + "spans": [ + { + "bbox": [ + 46, + 423, + 289, + 471 + ], + "type": "text", + "content": "As " + }, + { + "bbox": [ + 46, + 423, + 289, + 471 + ], + "type": "inline_equation", + "content": "\\overline{\\mathbf{A}}" + }, + { + "bbox": [ + 46, + 423, + 289, + 471 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 423, + 289, + 471 + ], + "type": "inline_equation", + "content": "\\overline{\\mathbf{B}}" + }, + { + "bbox": [ + 46, + 423, + 289, + 471 + ], + "type": "text", + "content": " are linear operators, the auto-regressive formulation of Eq. (3), computing the output at each time step " + }, + { + "bbox": [ + 46, + 423, + 289, + 471 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 46, + 423, + 289, + 471 + ], + "type": "text", + "content": " in a recurrent manner, can be unrolled and expressed as single convolution with kernel " + }, + { + "bbox": [ + 46, + 423, + 289, + 471 + ], + "type": "inline_equation", + "content": "\\overline{\\mathbf{K}}\\in \\mathbb{R}^{L}" + }, + { + "bbox": [ + 46, + 423, + 289, + 471 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 478, + 287, + 499 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 478, + 287, + 499 + ], + "spans": [ + { + "bbox": [ + 55, + 478, + 287, + 499 + ], + "type": "interline_equation", + "content": "\\mathbf {y} = \\mathbf {x} \\odot \\overline {{\\mathbf {K}}}, \\quad \\overline {{\\mathbf {K}}} = \\left(\\mathbf {C} \\overline {{\\mathbf {B}}}, \\mathbf {C} \\overline {{\\mathbf {A B}}}, \\dots , \\mathbf {C} \\overline {{\\mathbf {A}}} ^ {L - 1} \\overline {{\\mathbf {B}}}\\right). \\tag {4}", + "image_path": "e4cc8a7b1c15be961dd363470d2b4eb34eba219596272b50c10d2ea3d43daadc.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 46, + 506, + 289, + 530 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 506, + 289, + 530 + ], + "spans": [ + { + "bbox": [ + 46, + 506, + 289, + 530 + ], + "type": "text", + "content": "During training, this convolutional can be employed to compute all output values in parallel." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 46, + 533, + 288, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 533, + 288, + 629 + ], + "spans": [ + { + "bbox": [ + 46, + 533, + 288, + 629 + ], + "type": "text", + "content": "Spectral Graph Analysis studies the properties of a graph by analyzing the eigen-spectrum of its Laplacian matrix, which can be seen as a discrete version of the continuous Laplace-Beltrami operator obtained by a finite difference method [30]. Given a real-valued function " + }, + { + "bbox": [ + 46, + 533, + 288, + 629 + ], + "type": "inline_equation", + "content": "f \\in C^2" + }, + { + "bbox": [ + 46, + 533, + 288, + 629 + ], + "type": "text", + "content": " on a Riemannian manifold " + }, + { + "bbox": [ + 46, + 533, + 288, + 629 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 46, + 533, + 288, + 629 + ], + "type": "text", + "content": ", the Laplace-Beltrami operator " + }, + { + "bbox": [ + 46, + 533, + 288, + 629 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 46, + 533, + 288, + 629 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 46, + 533, + 288, + 629 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 46, + 533, + 288, + 629 + ], + "type": "text", + "content": " is defined as " + }, + { + "bbox": [ + 46, + 533, + 288, + 629 + ], + "type": "inline_equation", + "content": "\\Delta f = \\text{div}(\\text{grad} f)" + }, + { + "bbox": [ + 46, + 533, + 288, + 629 + ], + "type": "text", + "content": " where with " + }, + { + "bbox": [ + 46, + 533, + 288, + 629 + ], + "type": "inline_equation", + "content": "\\text{grad} f" + }, + { + "bbox": [ + 46, + 533, + 288, + 629 + ], + "type": "text", + "content": " is the gradient of " + }, + { + "bbox": [ + 46, + 533, + 288, + 629 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 46, + 533, + 288, + 629 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 533, + 288, + 629 + ], + "type": "inline_equation", + "content": "\\text{div}" + }, + { + "bbox": [ + 46, + 533, + 288, + 629 + ], + "type": "text", + "content": " denotes the divergence." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 46, + 630, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 630, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 630, + 288, + 714 + ], + "type": "text", + "content": "The Laplacian Eigenvalue Problem corresponds to solving the Helmholtz equation " + }, + { + "bbox": [ + 46, + 630, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\Delta f = -\\lambda f" + }, + { + "bbox": [ + 46, + 630, + 288, + 714 + ], + "type": "text", + "content": ", whose solutions (eigenfunctions) are directly related to the vibration of a physical membrane [2]. Following Courant's Nodal Line Theorem [5], the Laplacian eigen-spectrum is composed of non-negative eigenvalues " + }, + { + "bbox": [ + 46, + 630, + 288, + 714 + ], + "type": "inline_equation", + "content": "0 \\leq \\lambda_1 \\leq \\lambda_2 \\leq \\ldots" + }, + { + "bbox": [ + 46, + 630, + 288, + 714 + ], + "type": "text", + "content": ", each one repeating according to its multiplicity. Moreover, the" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 304, + 72, + 547, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 120 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 120 + ], + "type": "text", + "content": "eigenfunction corresponding to " + }, + { + "bbox": [ + 304, + 72, + 547, + 120 + ], + "type": "inline_equation", + "content": "\\lambda_{n}" + }, + { + "bbox": [ + 304, + 72, + 547, + 120 + ], + "type": "text", + "content": " has at most " + }, + { + "bbox": [ + 304, + 72, + 547, + 120 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 304, + 72, + 547, + 120 + ], + "type": "text", + "content": " poles of vibration, hence smaller eigenvalues and their associated eigenvector encode lower frequency structural information which captures relationships at a more global scale." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 120, + 547, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 120, + 547, + 204 + ], + "spans": [ + { + "bbox": [ + 304, + 120, + 547, + 204 + ], + "type": "text", + "content": "The Laplacian matrix of a graph " + }, + { + "bbox": [ + 304, + 120, + 547, + 204 + ], + "type": "inline_equation", + "content": "G = (V, E)" + }, + { + "bbox": [ + 304, + 120, + 547, + 204 + ], + "type": "text", + "content": " with nodes " + }, + { + "bbox": [ + 304, + 120, + 547, + 204 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 304, + 120, + 547, + 204 + ], + "type": "text", + "content": " and edges " + }, + { + "bbox": [ + 304, + 120, + 547, + 204 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 304, + 120, + 547, + 204 + ], + "type": "text", + "content": " is defined as " + }, + { + "bbox": [ + 304, + 120, + 547, + 204 + ], + "type": "inline_equation", + "content": "\\mathbf{L} = \\mathbf{D} - \\mathbf{W}" + }, + { + "bbox": [ + 304, + 120, + 547, + 204 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 304, + 120, + 547, + 204 + ], + "type": "inline_equation", + "content": "\\mathbf{D}" + }, + { + "bbox": [ + 304, + 120, + 547, + 204 + ], + "type": "text", + "content": " is the diagonal degree matrix with entries " + }, + { + "bbox": [ + 304, + 120, + 547, + 204 + ], + "type": "inline_equation", + "content": "D_{ii} = \\sum_{j} W_{ij}" + }, + { + "bbox": [ + 304, + 120, + 547, + 204 + ], + "type": "text", + "content": ". A normalized version of the Laplacian is typically employed to address differences in the weight scales " + }, + { + "bbox": [ + 304, + 120, + 547, + 204 + ], + "type": "inline_equation", + "content": "W_{ij}" + }, + { + "bbox": [ + 304, + 120, + 547, + 204 + ], + "type": "text", + "content": " and variations in the distribution of node degrees " + }, + { + "bbox": [ + 304, + 120, + 547, + 204 + ], + "type": "inline_equation", + "content": "D_{ii}" + }, + { + "bbox": [ + 304, + 120, + 547, + 204 + ], + "type": "text", + "content": ". In this work, we utilize the symmetric normalized Laplacian, defined as" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 369, + 211, + 546, + 227 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 369, + 211, + 546, + 227 + ], + "spans": [ + { + "bbox": [ + 369, + 211, + 546, + 227 + ], + "type": "interline_equation", + "content": "\\mathbf {L} _ {\\text {s y m}} = \\mathbf {I} - \\mathbf {D} ^ {- \\frac {1}{2}} \\mathbf {W D} ^ {- \\frac {1}{2}}, \\tag {5}", + "image_path": "cd2c2c24ca98c269e0fe1fbd0471449047e6daac1349f2bebb19bb227059bf90.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 233, + 547, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 233, + 547, + 258 + ], + "spans": [ + { + "bbox": [ + 304, + 233, + 547, + 258 + ], + "type": "text", + "content": "which is positive semi-definite and possesses " + }, + { + "bbox": [ + 304, + 233, + 547, + 258 + ], + "type": "inline_equation", + "content": "|V|" + }, + { + "bbox": [ + 304, + 233, + 547, + 258 + ], + "type": "text", + "content": " non-negative real-valued eigenvalues " + }, + { + "bbox": [ + 304, + 233, + 547, + 258 + ], + "type": "inline_equation", + "content": "0 = \\lambda_{1} \\leq \\ldots \\leq \\lambda_{|V|}" + }, + { + "bbox": [ + 304, + 233, + 547, + 258 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 304, + 258, + 547, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 258, + 547, + 342 + ], + "spans": [ + { + "bbox": [ + 304, + 258, + 547, + 342 + ], + "type": "text", + "content": "A key property of the Laplace eigen-spectrum is that it is invariant to transformations preserving distance (i.e., isometry), which include translations and rotations. In our work, we use this property to design an SSM that enhances rotational robustness, where the order in which patches are traversed is determined based on the eigenvectors corresponding to the smallest eigenvalues of the Laplacian matrix." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 306, + 348, + 415, + 361 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 348, + 415, + 361 + ], + "spans": [ + { + "bbox": [ + 306, + 348, + 415, + 361 + ], + "type": "text", + "content": "3.2. Spectral VMamba" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 366, + 547, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 366, + 547, + 521 + ], + "spans": [ + { + "bbox": [ + 304, + 366, + 547, + 521 + ], + "type": "text", + "content": "Existing vision-based Mamba methods, such as VMamba [21], are based on SS2D, a four-way scanning mechanism designed for spatial domain traversal. SS2D organizes input patches along four predefined scanning routes. However, two key challenges emerge: (1) connections between semantically related but non- neighboring patches are overlooked; and (2) scanning is sensitive to isometric transformations, such as rotation. To address these limitations, we introduce the Spectral VMamba model whose architecture is summarized in Figure 3. This architecture consists of two main components: a RFN module ensuring consistent feature representation across orientations and a sequence of Spectral VMamba blocks forming the network body." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 304, + 522, + 548, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 522, + 548, + 714 + ], + "spans": [ + { + "bbox": [ + 304, + 522, + 548, + 714 + ], + "type": "text", + "content": "Data forwarding in our Spectral VMamba block comprises three key steps: Spectral Traversal Scan (STS), S6 block (selective scan), and Spectral Traversal Merge (STM). In the STS step, input patches are reordered based on the Laplacian spectrum of their underlying graph structure, ensuring traversal aligns with spectral coherence (see Section 3.4). The selective scan step then processes each reordered patch sequence in parallel using dedicated S6 blocks, effectively capturing localized patterns. Finally, in the STM step, the reordered traversal paths are restored to their original spatial format, merging the resulting sequences to generate the final output map. The detailed architecture of STS is depicted in Figure 2. The sequence of patch processing is determined by the order of eigenvectors derived from the spectral decomposition of the graph Laplacian. It is important to note that STS is only performed once for the initial layer." + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "23884" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 53, + 76, + 545, + 257 + ], + "blocks": [ + { + "bbox": [ + 53, + 76, + 545, + 257 + ], + "lines": [ + { + "bbox": [ + 53, + 76, + 545, + 257 + ], + "spans": [ + { + "bbox": [ + 53, + 76, + 545, + 257 + ], + "type": "image", + "image_path": "60c5b827b37c1191418304ec6b87b25d4419ce133099069e87a45111f0b3974d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 46, + 265, + 548, + 323 + ], + "lines": [ + { + "bbox": [ + 46, + 265, + 548, + 323 + ], + "spans": [ + { + "bbox": [ + 46, + 265, + 548, + 323 + ], + "type": "text", + "content": "Figure 3. An overview of the proposed method. Our network consists of two primary components: the RFN module, which normalizes feature orientation to ensure consistency across different perspectives, and a series of Spectral VMamba blocks that process data in three stages: (1) Spectral Traversal Scan (STS), (2) S6 block (selective scan), and (3) Spectral Traversal Merge (STM). In STS, patches are reordered based on spectral coherence derived from the Laplacian spectrum. Each sequence is then processed in parallel by dedicated S6 blocks, capturing localized patterns, and finally merged in STM to reconstruct the spatial layout, producing the final feature map." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 46, + 331, + 288, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 331, + 288, + 378 + ], + "spans": [ + { + "bbox": [ + 46, + 331, + 288, + 378 + ], + "type": "text", + "content": "In subsequent layers, we apply a downsampling strategy for generating the traversal path. For a detailed explanation of the downsampling process, please refer to the supplementary material." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 380, + 290, + 608 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 380, + 290, + 608 + ], + "spans": [ + { + "bbox": [ + 46, + 380, + 290, + 608 + ], + "type": "text", + "content": "The graph Laplacian eigen-spectrum is isometric invariant, guaranteeing that the order in which patches are traversed is not affected by changes to their position in the image. However, the values in the Laplacian matrix are computed using patch features which are themselves sensitive to rotation of the patch. A simple solution to this problem is to use patches containing a single pixel, however this would be impractical for two reasons. First, individual pixels encode little information (RGB color) and thus using their similarity to construct the Laplacian matrix would not be useful. More importantly, this would severely increase the computational cost of downstream operations such as the spectral decomposition of the Laplacian. To address this issue, we compute patch features using our RFN module (Section 3.3) designed to normalize the feature extraction process by aggregating features from multiple orientations of the input image. This approach allows for effectively handling of feature changes during rotations, preserving the desired order and mitigating orientation-dependent discrepancies." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 621, + 249, + 634 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 621, + 249, + 634 + ], + "spans": [ + { + "bbox": [ + 47, + 621, + 249, + 634 + ], + "type": "text", + "content": "3.3. Rotational Feature Normalizer (RFN)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 639, + 289, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 639, + 289, + 715 + ], + "spans": [ + { + "bbox": [ + 46, + 639, + 289, + 715 + ], + "type": "text", + "content": "An overview of the module is illustrated in Figure 3. This module rotates the input image by a pre-determined set of angles " + }, + { + "bbox": [ + 46, + 639, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\{\\theta_i,\\dots ,\\theta_R\\}" + }, + { + "bbox": [ + 46, + 639, + 289, + 715 + ], + "type": "text", + "content": " and processes them with a stem module which serves as a feature extraction backbone for patchification. This backbone partitions the input image " + }, + { + "bbox": [ + 46, + 639, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{I}\\in \\mathbb{R}^{H\\times W\\times 3}" + }, + { + "bbox": [ + 46, + 639, + 289, + 715 + ], + "type": "text", + "content": " into patches " + }, + { + "bbox": [ + 46, + 639, + 289, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{x}\\in \\mathbb{R}_{\\frac{H}{p}\\times \\frac{W}{p}\\times C}" + }, + { + "bbox": [ + 46, + 639, + 289, + 715 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 639, + 289, + 715 + ], + "type": "inline_equation", + "content": "(H,W)" + }, + { + "bbox": [ + 46, + 639, + 289, + 715 + ], + "type": "text", + "content": " denote the di" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 304, + 331, + 547, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 331, + 547, + 439 + ], + "spans": [ + { + "bbox": [ + 304, + 331, + 547, + 439 + ], + "type": "text", + "content": "mensions of the input image, " + }, + { + "bbox": [ + 304, + 331, + 547, + 439 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 304, + 331, + 547, + 439 + ], + "type": "text", + "content": " is the number of feature channels, and " + }, + { + "bbox": [ + 304, + 331, + 547, + 439 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 304, + 331, + 547, + 439 + ], + "type": "text", + "content": " represents the patch size. To consolidate features from all rotated versions, each feature map is rotated back to its original orientation. This unrotation ensures that all extracted features align with the original image coordinate system. Finally, an element-wise max operation is applied across the unrotated feature maps to generate a single aggregated feature. The final aggregated feature map can be expressed as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 330, + 445, + 547, + 467 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 330, + 445, + 547, + 467 + ], + "spans": [ + { + "bbox": [ + 330, + 445, + 547, + 467 + ], + "type": "interline_equation", + "content": "\\mathbf {F} _ {i, j} = \\max _ {r \\in \\{1, \\dots , R \\}} \\left[ \\mathcal {R} _ {- \\theta_ {r}} \\left(\\operatorname {P a t c h i f y} \\left(\\mathcal {R} _ {\\theta_ {r}} (\\mathbf {I})\\right)\\right) \\right] _ {i, j} \\tag {6}", + "image_path": "787104fb1bf1d1f47df293502014ead2c8373c025c8b59bd0e865623daa6489a.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 304, + 473, + 545, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 473, + 545, + 497 + ], + "spans": [ + { + "bbox": [ + 304, + 473, + 545, + 497 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 304, + 473, + 545, + 497 + ], + "type": "inline_equation", + "content": "\\theta_r" + }, + { + "bbox": [ + 304, + 473, + 545, + 497 + ], + "type": "text", + "content": " is the angle of the " + }, + { + "bbox": [ + 304, + 473, + 545, + 497 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 304, + 473, + 545, + 497 + ], + "type": "text", + "content": "-th rotation and " + }, + { + "bbox": [ + 304, + 473, + 545, + 497 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_{\\theta_r}" + }, + { + "bbox": [ + 304, + 473, + 545, + 497 + ], + "type": "text", + "content": " is the transformation applying this rotation." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 497, + 547, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 497, + 547, + 640 + ], + "spans": [ + { + "bbox": [ + 304, + 497, + 547, + 640 + ], + "type": "text", + "content": "The choice of rotation angles for the RFN module represents a trade-off between rotation invariance and computational overhead. Using a larger number of angles ensures a greater coverage in terms of rotation, hence improving the invariance of the model at the cost of increased computations. In our implementation, we used four canonical angles, " + }, + { + "bbox": [ + 304, + 497, + 547, + 640 + ], + "type": "inline_equation", + "content": "\\{0^{\\circ}, 90^{\\circ}, 180^{\\circ}, 270^{\\circ}\\}" + }, + { + "bbox": [ + 304, + 497, + 547, + 640 + ], + "type": "text", + "content": ", corresponding to quarter turns. By rotating the image at these specific angles, the pixel grid alignment is preserved, eliminating the need for interpolation. This alignment not only simplifies the back-rotation process but also guarantees that the extracted features from all rotated images are comparable." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 306, + 647, + 441, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 647, + 441, + 659 + ], + "spans": [ + { + "bbox": [ + 306, + 647, + 441, + 659 + ], + "type": "text", + "content": "3.4. Spectral Traversal Scan" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 665, + 547, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 665, + 547, + 689 + ], + "spans": [ + { + "bbox": [ + 304, + 665, + 547, + 689 + ], + "type": "text", + "content": "In this section, we describe how the graph Laplacian eigenspectrum is leveraged to define the traversal order of patches." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 689, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 689, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 689, + 547, + 715 + ], + "type": "text", + "content": "We begin by representing the image as a graph " + }, + { + "bbox": [ + 304, + 689, + 547, + 715 + ], + "type": "inline_equation", + "content": "G = (V, E)" + }, + { + "bbox": [ + 304, + 689, + 547, + 715 + ], + "type": "text", + "content": ", where each node " + }, + { + "bbox": [ + 304, + 689, + 547, + 715 + ], + "type": "inline_equation", + "content": "v_i \\in V" + }, + { + "bbox": [ + 304, + 689, + 547, + 715 + ], + "type": "text", + "content": " corresponds to a patch " + }, + { + "bbox": [ + 304, + 689, + 547, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i" + }, + { + "bbox": [ + 304, + 689, + 547, + 715 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23885" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "spans": [ + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "content": "of the image, and edges " + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "inline_equation", + "content": "e_{ij} \\in E" + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "content": " connect pairs of nodes based on their spatial proximity. Following methods for spectral clustering [28] and normalized cuts [32], we consider a weighted adjacency matrix " + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "inline_equation", + "content": "\\mathbf{W}" + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "content": " of the graph, which is constructed using Euclidean distances between patches. Specifically, the weight " + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "inline_equation", + "content": "W_{ij}" + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "content": " between the nodes " + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "inline_equation", + "content": "v_i" + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "inline_equation", + "content": "v_j" + }, + { + "bbox": [ + 46, + 72, + 289, + 156 + ], + "type": "text", + "content": " is given by" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 101, + 160, + 288, + 187 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 101, + 160, + 288, + 187 + ], + "spans": [ + { + "bbox": [ + 101, + 160, + 288, + 187 + ], + "type": "interline_equation", + "content": "W _ {i j} = \\exp \\left(- \\frac {1}{2 \\sigma^ {2}} \\| \\mathbf {f} _ {i} - \\mathbf {f} _ {j} \\| ^ {2}\\right) \\tag {7}", + "image_path": "7d5f567a1be668bc6ecd813e2db1edb03e4d6a4fcacf8477a9f5544a1610e99d.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 191, + 288, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 191, + 288, + 251 + ], + "spans": [ + { + "bbox": [ + 46, + 191, + 288, + 251 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 191, + 288, + 251 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_i" + }, + { + "bbox": [ + 46, + 191, + 288, + 251 + ], + "type": "text", + "content": " is the feature vector of patch " + }, + { + "bbox": [ + 46, + 191, + 288, + 251 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i" + }, + { + "bbox": [ + 46, + 191, + 288, + 251 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 46, + 191, + 288, + 251 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 46, + 191, + 288, + 251 + ], + "type": "text", + "content": " is a scaling parameter determined as the average of the Euclidean distances between patches. We set " + }, + { + "bbox": [ + 46, + 191, + 288, + 251 + ], + "type": "inline_equation", + "content": "W_{ij} = 0" + }, + { + "bbox": [ + 46, + 191, + 288, + 251 + ], + "type": "text", + "content": " for pairs " + }, + { + "bbox": [ + 46, + 191, + 288, + 251 + ], + "type": "inline_equation", + "content": "(v_i, v_j)" + }, + { + "bbox": [ + 46, + 191, + 288, + 251 + ], + "type": "text", + "content": " that are not among the top-" + }, + { + "bbox": [ + 46, + 191, + 288, + 251 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 191, + 288, + 251 + ], + "type": "text", + "content": " nearest neighbors of each other, resulting in a sparse adjacency matrix." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 46, + 251, + 287, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 251, + 287, + 386 + ], + "spans": [ + { + "bbox": [ + 46, + 251, + 287, + 386 + ], + "type": "text", + "content": "We compute the first " + }, + { + "bbox": [ + 46, + 251, + 287, + 386 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 46, + 251, + 287, + 386 + ], + "type": "text", + "content": " smallest eigenvectors of the symmetric Laplacian " + }, + { + "bbox": [ + 46, + 251, + 287, + 386 + ], + "type": "inline_equation", + "content": "\\mathbf{L}_{\\mathrm{sym}}" + }, + { + "bbox": [ + 46, + 251, + 287, + 386 + ], + "type": "text", + "content": " using an iterative method known as the Arnoldi algorithm [9], which leverages the sparsity of the adjacency matrix to provide a low complexity. Each eigenvector " + }, + { + "bbox": [ + 46, + 251, + 287, + 386 + ], + "type": "inline_equation", + "content": "\\mathbf{u}^{(j)} \\in \\mathbb{R}^{N_p}" + }, + { + "bbox": [ + 46, + 251, + 287, + 386 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 46, + 251, + 287, + 386 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 46, + 251, + 287, + 386 + ], + "type": "text", + "content": " ranges from 1 to " + }, + { + "bbox": [ + 46, + 251, + 287, + 386 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 46, + 251, + 287, + 386 + ], + "type": "text", + "content": ", assigns an eigenfunction value " + }, + { + "bbox": [ + 46, + 251, + 287, + 386 + ], + "type": "inline_equation", + "content": "\\mathbf{u}_i^{(j)}" + }, + { + "bbox": [ + 46, + 251, + 287, + 386 + ], + "type": "text", + "content": " to each image patch " + }, + { + "bbox": [ + 46, + 251, + 287, + 386 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 46, + 251, + 287, + 386 + ], + "type": "text", + "content": ". In every Mamba block of our model, we perform two traversals per eigenvector: one in the order of increasing " + }, + { + "bbox": [ + 46, + 251, + 287, + 386 + ], + "type": "inline_equation", + "content": "\\mathbf{u}_i^{(j)}" + }, + { + "bbox": [ + 46, + 251, + 287, + 386 + ], + "type": "text", + "content": " values and another in decreasing order. At the end of the block, the resulting features from these " + }, + { + "bbox": [ + 46, + 251, + 287, + 386 + ], + "type": "inline_equation", + "content": "2m" + }, + { + "bbox": [ + 46, + 251, + 287, + 386 + ], + "type": "text", + "content": " traversals are concatenated together." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 47, + 386, + 288, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 386, + 288, + 410 + ], + "spans": [ + { + "bbox": [ + 47, + 386, + 288, + 410 + ], + "type": "text", + "content": "Patches " + }, + { + "bbox": [ + 47, + 386, + 288, + 410 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i" + }, + { + "bbox": [ + 47, + 386, + 288, + 410 + ], + "type": "text", + "content": " are ordered based on these eigenvectors, generating multiple traversal sequences:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 58, + 414, + 287, + 435 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 414, + 287, + 435 + ], + "spans": [ + { + "bbox": [ + 58, + 414, + 287, + 435 + ], + "type": "interline_equation", + "content": "\\mathbf {X} ^ {(j, \\text {r e v})} = \\left\\{\\mathbf {x} _ {i _ {1}} ^ {(j, \\text {r e v})}, \\mathbf {x} _ {i _ {2}} ^ {(j, \\text {r e v})}, \\dots , \\mathbf {x} _ {i _ {m}} ^ {(j, \\text {r e v})} \\right\\}, \\text {w h e r e} \\tag {8}", + "image_path": "e81409dbb574836e1887d50d0a7e350241e641da176f231a87937edb1b2cb950.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 52, + 439, + 281, + 459 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 52, + 439, + 281, + 459 + ], + "spans": [ + { + "bbox": [ + 52, + 439, + 281, + 459 + ], + "type": "interline_equation", + "content": "i _ {k} ^ {(j, \\operatorname {r e v})} = \\arg \\min _ {i \\in \\{1, \\dots , m \\}} \\left\\{\\left(- 1\\right) ^ {\\operatorname {r e v}} \\cdot u _ {i} ^ {(j)} \\mid i \\notin \\left\\{i _ {1}, \\dots , i _ {k - 1} \\right\\} \\right\\}", + "image_path": "25500b99c10f3daab99af1b0849ee23ee4bb64163a26f9581c4712160564946b.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 46, + 461, + 287, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 461, + 287, + 521 + ], + "spans": [ + { + "bbox": [ + 46, + 461, + 287, + 521 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 46, + 461, + 287, + 521 + ], + "type": "inline_equation", + "content": "\\operatorname{rev} \\in \\{0,1\\}" + }, + { + "bbox": [ + 46, + 461, + 287, + 521 + ], + "type": "text", + "content": " indicates if the traversal is reversed " + }, + { + "bbox": [ + 46, + 461, + 287, + 521 + ], + "type": "inline_equation", + "content": "(\\operatorname{rev} = 1)" + }, + { + "bbox": [ + 46, + 461, + 287, + 521 + ], + "type": "text", + "content": " or not " + }, + { + "bbox": [ + 46, + 461, + 287, + 521 + ], + "type": "inline_equation", + "content": "(\\operatorname{rev} = 0)" + }, + { + "bbox": [ + 46, + 461, + 287, + 521 + ], + "type": "text", + "content": ". This concatenated sequence " + }, + { + "bbox": [ + 46, + 461, + 287, + 521 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 46, + 461, + 287, + 521 + ], + "type": "text", + "content": " is then fed into the Mamba network for further processing, allowing the model to incorporate rich spectral relationships while preserving global coherence." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 47, + 527, + 190, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 527, + 190, + 540 + ], + "spans": [ + { + "bbox": [ + 47, + 527, + 190, + 540 + ], + "type": "text", + "content": "3.5. Spectral Canonicalization" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 46, + 545, + 287, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 545, + 287, + 665 + ], + "spans": [ + { + "bbox": [ + 46, + 545, + 287, + 665 + ], + "type": "text", + "content": "Eigenvectors inherently have a sign ambiguity as, if " + }, + { + "bbox": [ + 46, + 545, + 287, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{u}" + }, + { + "bbox": [ + 46, + 545, + 287, + 665 + ], + "type": "text", + "content": " is an eigenvector of a matrix " + }, + { + "bbox": [ + 46, + 545, + 287, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 46, + 545, + 287, + 665 + ], + "type": "text", + "content": " with eigenvalue " + }, + { + "bbox": [ + 46, + 545, + 287, + 665 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 545, + 287, + 665 + ], + "type": "text", + "content": ", then " + }, + { + "bbox": [ + 46, + 545, + 287, + 665 + ], + "type": "inline_equation", + "content": "-\\mathbf{u}" + }, + { + "bbox": [ + 46, + 545, + 287, + 665 + ], + "type": "text", + "content": " is also an eigenvector with the same eigenvalue. To resolve this ambiguity, we standardize the sign of eigenvectors by ensuring that the first non-zero element of each eigenvector is positive. If the first element is negative, we flip the sign of the entire eigenvector. This adjustment does not alter the underlying structure captured by the eigenvectors, as flipping the sign consistently maintains the same eigenvalue " + }, + { + "bbox": [ + 46, + 545, + 287, + 665 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 46, + 545, + 287, + 665 + ], + "type": "text", + "content": " and preserves the relationships between eigenvectors." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 47, + 671, + 164, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 671, + 164, + 682 + ], + "spans": [ + { + "bbox": [ + 47, + 671, + 164, + 682 + ], + "type": "text", + "content": "3.6. Rotation Invariance" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 46, + 689, + 288, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 689, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 46, + 689, + 288, + 713 + ], + "type": "text", + "content": "Spectral decomposition leverages the eigen-structure of the Laplacian matrix, which remains consistent regardless of" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 304, + 72, + 547, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 72, + 547, + 131 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 547, + 131 + ], + "type": "text", + "content": "rotation. As our traversal strategy is directly guided by the spectral decomposition of input features, the rotation-invariant nature of spectral decomposition extends to our method. This property is demonstrated in the following theorem." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 140, + 547, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 140, + 547, + 176 + ], + "spans": [ + { + "bbox": [ + 304, + 140, + 547, + 176 + ], + "type": "text", + "content": "Theorem 3.1. Under the assumption that patch features " + }, + { + "bbox": [ + 304, + 140, + 547, + 176 + ], + "type": "inline_equation", + "content": "\\mathbf{f}" + }, + { + "bbox": [ + 304, + 140, + 547, + 176 + ], + "type": "text", + "content": " are invariant to rotation, i.e. " + }, + { + "bbox": [ + 304, + 140, + 547, + 176 + ], + "type": "inline_equation", + "content": "\\mathbf{f}(\\mathbf{I}) = \\mathbf{f}(\\mathcal{R}_{\\theta}(\\mathbf{I}))" + }, + { + "bbox": [ + 304, + 140, + 547, + 176 + ], + "type": "text", + "content": " for any " + }, + { + "bbox": [ + 304, + 140, + 547, + 176 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 304, + 140, + 547, + 176 + ], + "type": "text", + "content": ", our spectral traversal scan is also rotation-invariant." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 304, + 184, + 547, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 184, + 547, + 317 + ], + "spans": [ + { + "bbox": [ + 304, + 184, + 547, + 317 + ], + "type": "text", + "content": "Proof. As rotations are a type of isometry, we can prove the more general property of isometric invariance. This invariance derives from two properties: 1) the nearest neighbors and values of the weighted adjacency matrix are defined based on distance alone, i.e., " + }, + { + "bbox": [ + 304, + 184, + 547, + 317 + ], + "type": "inline_equation", + "content": "W_{ij} = \\exp \\left(-\\frac{1}{2\\sigma}\\right)^2\\| \\mathbf{f}_i - \\mathbf{f}_j\\|^2" + }, + { + "bbox": [ + 304, + 184, + 547, + 317 + ], + "type": "text", + "content": ", thus they do not change if distances remain the same; 2) the eigen-spectrum is invariant under cyclic permutation. This second property can be demonstrated as follows. Let " + }, + { + "bbox": [ + 304, + 184, + 547, + 317 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 304, + 184, + 547, + 317 + ], + "type": "text", + "content": " be a permutation matrix and " + }, + { + "bbox": [ + 304, + 184, + 547, + 317 + ], + "type": "inline_equation", + "content": "\\mathbf{u}" + }, + { + "bbox": [ + 304, + 184, + 547, + 317 + ], + "type": "text", + "content": " be an eigenvector of Laplacian matrix " + }, + { + "bbox": [ + 304, + 184, + 547, + 317 + ], + "type": "inline_equation", + "content": "\\mathbf{L}" + }, + { + "bbox": [ + 304, + 184, + 547, + 317 + ], + "type": "text", + "content": " corresponding to eigenvalue " + }, + { + "bbox": [ + 304, + 184, + 547, + 317 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 304, + 184, + 547, + 317 + ], + "type": "text", + "content": ". Using " + }, + { + "bbox": [ + 304, + 184, + 547, + 317 + ], + "type": "inline_equation", + "content": "\\mathbf{u}' = \\mathbf{P}\\mathbf{u}" + }, + { + "bbox": [ + 304, + 184, + 547, + 317 + ], + "type": "text", + "content": " (thus " + }, + { + "bbox": [ + 304, + 184, + 547, + 317 + ], + "type": "inline_equation", + "content": "\\mathbf{u} = \\mathbf{P}^T\\mathbf{u}'" + }, + { + "bbox": [ + 304, + 184, + 547, + 317 + ], + "type": "text", + "content": ") we have that" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 347, + 326, + 545, + 371 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 326, + 545, + 371 + ], + "spans": [ + { + "bbox": [ + 347, + 326, + 545, + 371 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathbf {L} \\mathbf {u} = \\lambda \\mathbf {u} \\Rightarrow \\mathbf {L P} ^ {T} \\mathbf {u} ^ {\\prime} = \\lambda \\mathbf {P} ^ {T} \\mathbf {u} ^ {\\prime} \\\\ \\Rightarrow \\mathbf {P L P} ^ {T} \\mathbf {u} ^ {\\prime} = \\mathbf {P} (\\lambda \\mathbf {P} ^ {T} \\mathbf {u} ^ {\\prime}) \\tag {9} \\\\ \\Rightarrow \\mathbf {P L P} ^ {T} \\mathbf {u} ^ {\\prime} = \\lambda \\mathbf {u} ^ {\\prime} \\\\ \\end{array}", + "image_path": "bfb703c5965e4e6c5c510e92ee2c03335f911121a06919f4df995fd62b05cdac.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 304, + 382, + 547, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 382, + 547, + 430 + ], + "spans": [ + { + "bbox": [ + 304, + 382, + 547, + 430 + ], + "type": "text", + "content": "Hence, " + }, + { + "bbox": [ + 304, + 382, + 547, + 430 + ], + "type": "inline_equation", + "content": "\\mathbf{Pu}" + }, + { + "bbox": [ + 304, + 382, + 547, + 430 + ], + "type": "text", + "content": " is an eigenvector of the permuted Laplacian matrix " + }, + { + "bbox": [ + 304, + 382, + 547, + 430 + ], + "type": "inline_equation", + "content": "\\mathbf{PLP}^T" + }, + { + "bbox": [ + 304, + 382, + 547, + 430 + ], + "type": "text", + "content": ". Since we traverse patches by their eigenvector values (from smallest to largest or vice-versa), we will therefore get the same ordering regardless of the permutation." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 305, + 440, + 387, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 440, + 387, + 453 + ], + "spans": [ + { + "bbox": [ + 305, + 440, + 387, + 453 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 304, + 460, + 545, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 460, + 545, + 508 + ], + "spans": [ + { + "bbox": [ + 304, + 460, + 545, + 508 + ], + "type": "text", + "content": "We start by comparing our method against recently-proposed ViTs and SSMs on an image classification task. We then present an ablation study to thoroughly investigate the impact of various components on the performance of our method." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 305, + 515, + 423, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 515, + 423, + 528 + ], + "spans": [ + { + "bbox": [ + 305, + 515, + 423, + 528 + ], + "type": "text", + "content": "4.1. Image Classification" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 304, + 534, + 547, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 534, + 547, + 677 + ], + "spans": [ + { + "bbox": [ + 304, + 534, + 547, + 677 + ], + "type": "text", + "content": "We conduct classification experiments on the miniImageNet [40] dataset. This dataset includes 50,000 training images and 10,000 validation images across 100 categories. Following previous works [21, 22, 46], we train our models for 300 epochs with a batch size of 128 per GPU. The models are optimized using the AdamW optimizer with a momentum of 0.9. A cosine decay scheduler manages the learning rate, which starts at " + }, + { + "bbox": [ + 304, + 534, + 547, + 677 + ], + "type": "inline_equation", + "content": "5 \\times 10^{-4}" + }, + { + "bbox": [ + 304, + 534, + 547, + 677 + ], + "type": "text", + "content": ", alongside a weight decay of 0.05. Additionally, we apply an exponential moving average (EMA) to stabilize training. For input images of size " + }, + { + "bbox": [ + 304, + 534, + 547, + 677 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 304, + 534, + 547, + 677 + ], + "type": "text", + "content": ", our data augmentation techniques include color jittering, AutoAugment, random erasing, mixup, and cutmix." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "spans": [ + { + "bbox": [ + 304, + 677, + 547, + 713 + ], + "type": "text", + "content": "Table 1 compares our Spectral VMamba method against ViTs and SSM-based models on the miniImageNet dataset. As can be seen, our method outperforms all other approaches" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "23886" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 62, + 99, + 270, + 343 + ], + "blocks": [ + { + "bbox": [ + 47, + 60, + 288, + 95 + ], + "lines": [ + { + "bbox": [ + 47, + 60, + 288, + 95 + ], + "spans": [ + { + "bbox": [ + 47, + 60, + 288, + 95 + ], + "type": "text", + "content": "Table 1. Classification performance comparison on miniImageNet. All images are of size " + }, + { + "bbox": [ + 47, + 60, + 288, + 95 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 47, + 60, + 288, + 95 + ], + "type": "text", + "content": ". T, S, and B denote the tiny, small, and base scales, respectively." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 62, + 99, + 270, + 343 + ], + "lines": [ + { + "bbox": [ + 62, + 99, + 270, + 343 + ], + "spans": [ + { + "bbox": [ + 62, + 99, + 270, + 343 + ], + "type": "table", + "html": "
ModelFLOPs (G)Top-1 (%)Top-5 (%)
Transformer-Based
DeiT-S [37]4.6G70.8389.74
DeiT-B [37]17.5G72.4390.14
Swin-T [22]4.5G83.2595.54
Swin-S [22]8.7G84.1095.53
Swin-B [22]15.4G82.7795.33
XCiT-S24 [1]9.2G85.7996.31
XCiT-M24 [1]16.2G86.8096.38
SSM-Based
Vim-T [46]1.5G67.3087.97
Vim-S [46]5.1G79.7093.20
LocalVim-T [17]1.5G82.1294.60
LocalVim-S [17]4.8G81.6893.63
MSVMamba-N [33]0.9G82.1695.10
MSVMamba-M [33]1.5G83.7295.81
MSVMamba-T [33]4.6G86.4896.43
VMamba-T [21]4.9G86.2596.64
VMamba-S [21]8.7G86.4896.79
VMamba-B [21]8.7G87.1796.96
Ours-T3.9G87.86 (+1.61)97.25 (+0.61)
Ours-S6.3G88.09 (+1.61)97.08 (+0.29)
Ours-B6.3G88.17 (+1.00)97.27 (+0.31)
", + "image_path": "c26fcbb45d6fada828cb704572d5ba7b55529b4b870f83fc9d148949a47cc450.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 361, + 288, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 361, + 288, + 468 + ], + "spans": [ + { + "bbox": [ + 46, + 361, + 288, + 468 + ], + "type": "text", + "content": "in both top-1 and top-5 EMA accuracy. Compared to VMamba, its closest competitor, our method improves top-1 accuracy by " + }, + { + "bbox": [ + 46, + 361, + 288, + 468 + ], + "type": "inline_equation", + "content": "1.61\\%" + }, + { + "bbox": [ + 46, + 361, + 288, + 468 + ], + "type": "text", + "content": " for the tiny and small models (T and S) and by " + }, + { + "bbox": [ + 46, + 361, + 288, + 468 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 46, + 361, + 288, + 468 + ], + "type": "text", + "content": " for the base model (B), while requiring less FLOPs (see 4.2 for explanation). Moreover, our Spectral VMamba (B) model yields a " + }, + { + "bbox": [ + 46, + 361, + 288, + 468 + ], + "type": "inline_equation", + "content": "1.37\\%" + }, + { + "bbox": [ + 46, + 361, + 288, + 468 + ], + "type": "text", + "content": " higher top-1 accuracy than the strongest ViT-based model, XCiT-M24, and an improvement of " + }, + { + "bbox": [ + 46, + 361, + 288, + 468 + ], + "type": "inline_equation", + "content": "5.4\\%" + }, + { + "bbox": [ + 46, + 361, + 288, + 468 + ], + "type": "text", + "content": " over Swin-B, one of the most popular architectures based on ViT." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 47, + 475, + 141, + 489 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 475, + 141, + 489 + ], + "spans": [ + { + "bbox": [ + 47, + 475, + 141, + 489 + ], + "type": "text", + "content": "4.2. Ablation Study" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 46, + 498, + 288, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 498, + 288, + 593 + ], + "spans": [ + { + "bbox": [ + 46, + 498, + 288, + 593 + ], + "type": "text", + "content": "Rotational invariance. To evaluate the robustness to rotation of our method trained on unrotated images, we evaluated its performance on test images rotated at angles from " + }, + { + "bbox": [ + 46, + 498, + 288, + 593 + ], + "type": "inline_equation", + "content": "0^{\\circ}" + }, + { + "bbox": [ + 46, + 498, + 288, + 593 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 46, + 498, + 288, + 593 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 46, + 498, + 288, + 593 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 46, + 498, + 288, + 593 + ], + "type": "inline_equation", + "content": "15^{\\circ}" + }, + { + "bbox": [ + 46, + 498, + 288, + 593 + ], + "type": "text", + "content": " increments. As a reminder, our RFN module extracts and aggregates features for four canonical rotations: " + }, + { + "bbox": [ + 46, + 498, + 288, + 593 + ], + "type": "inline_equation", + "content": "\\{0^{\\circ}, 90^{\\circ}, 180^{\\circ}, 270^{\\circ}\\}" + }, + { + "bbox": [ + 46, + 498, + 288, + 593 + ], + "type": "text", + "content": ". We therefore expect a greater robustness at those rotation angles, for which our method is fully invariant." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "content": "As shown in Figure 4, our model with the RFN (blue line) maintains a high accuracy near " + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "inline_equation", + "content": "87\\%" + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "content": " at those canonical angles, demonstrating its rotational invariance property. In contrast, VMamba (orange line) suffers from a significant drop in performance for images rotated at " + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "inline_equation", + "content": "90^{\\circ}" + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\sim 30\\%" + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "content": " drop), " + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "inline_equation", + "content": "180^{\\circ}" + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\sim 20\\%" + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "content": " drop) and " + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "inline_equation", + "content": "270^{\\circ}" + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "inline_equation", + "content": "\\sim 30\\%" + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "content": " drop). Although our method does not guarantee full rotational invariance for non-canonical angles, we observe a remarkably consistent performance across all rotation angles, stabilizing at around " + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "inline_equation", + "content": "78\\%" + }, + { + "bbox": [ + 46, + 594, + 288, + 714 + ], + "type": "text", + "content": " in top-1 accuracy. On the other hand, the performance" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 309, + 72, + 542, + 205 + ], + "blocks": [ + { + "bbox": [ + 309, + 72, + 542, + 205 + ], + "lines": [ + { + "bbox": [ + 309, + 72, + 542, + 205 + ], + "spans": [ + { + "bbox": [ + 309, + 72, + 542, + 205 + ], + "type": "image", + "image_path": "02e8a067c6b7c6ec97104fc4f505baf212546bc32b3ab321af1337e5599511c8.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 305, + 216, + 545, + 250 + ], + "lines": [ + { + "bbox": [ + 305, + 216, + 545, + 250 + ], + "spans": [ + { + "bbox": [ + 305, + 216, + 545, + 250 + ], + "type": "text", + "content": "Figure 4. Model performance comparison across various rotation angles. Our method demonstrates consistent accuracy across all rotation angles, while VMamba exhibits significant fluctuations." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 305, + 270, + 545, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 305, + 270, + 545, + 295 + ], + "spans": [ + { + "bbox": [ + 305, + 270, + 545, + 295 + ], + "type": "text", + "content": "of VMamba fluctuates drastically, reaching accuracy values near " + }, + { + "bbox": [ + 305, + 270, + 545, + 295 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 305, + 270, + 545, + 295 + ], + "type": "text", + "content": " at certain angles (e.g., " + }, + { + "bbox": [ + 305, + 270, + 545, + 295 + ], + "type": "inline_equation", + "content": "105^{\\circ}" + }, + { + "bbox": [ + 305, + 270, + 545, + 295 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 305, + 270, + 545, + 295 + ], + "type": "inline_equation", + "content": "255^{\\circ}" + }, + { + "bbox": [ + 305, + 270, + 545, + 295 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 298, + 547, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 298, + 547, + 478 + ], + "spans": [ + { + "bbox": [ + 304, + 298, + 547, + 478 + ], + "type": "text", + "content": "Impact of RFN. To assess the usefulness of our RFN module, we also tested an ablation variant of our method without this module (Ours w/o RFN in Figure 4). As mentioned before, while the Laplacian eigen-spectrum is isometric-invariant, and thus invariant to rotation, the Laplacian matrix is constructed using patch features which are themselves sensitive to rotation. As we expect, removing the RFN causes the performance to drop for all angles, this drop growing as we increase the rotation (away from " + }, + { + "bbox": [ + 304, + 298, + 547, + 478 + ], + "type": "inline_equation", + "content": "0^{\\circ}" + }, + { + "bbox": [ + 304, + 298, + 547, + 478 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 304, + 298, + 547, + 478 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 304, + 298, + 547, + 478 + ], + "type": "text", + "content": "). Interestingly, our method without the RFN module still achieves a higher accuracy than VMamba for angles between " + }, + { + "bbox": [ + 304, + 298, + 547, + 478 + ], + "type": "inline_equation", + "content": "60^{\\circ}" + }, + { + "bbox": [ + 304, + 298, + 547, + 478 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 304, + 298, + 547, + 478 + ], + "type": "inline_equation", + "content": "300^{\\circ}" + }, + { + "bbox": [ + 304, + 298, + 547, + 478 + ], + "type": "text", + "content": ", which represent severe rotations. This could be explained by the small size of patches " + }, + { + "bbox": [ + 304, + 298, + 547, + 478 + ], + "type": "inline_equation", + "content": "(16 \\times 16)" + }, + { + "bbox": [ + 304, + 298, + 547, + 478 + ], + "type": "text", + "content": " compared to the image, minimizing the impact of rotations (full invariance would be achieved for " + }, + { + "bbox": [ + 304, + 298, + 547, + 478 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 304, + 298, + 547, + 478 + ], + "type": "text", + "content": " patches)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 304, + 482, + 547, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 482, + 547, + 590 + ], + "spans": [ + { + "bbox": [ + 304, + 482, + 547, + 590 + ], + "type": "text", + "content": "Impact of STS. Since the STS is a critical part of our Spectral VMamba method, we cannot remove it. To evaluate its impact, we instead add the RFN to VMamba, the resulting approach now only differing from our method by not having a STS. As shown in Figure 4, the VMamba + RFN model (green line) offers a low robustness to rotation, its accuracy similar to that of VMamba. These results demonstrate that improving rotational invariance at the patch level is not sufficient to achieve a high performance." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 594, + 548, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 594, + 548, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 594, + 548, + 715 + ], + "type": "text", + "content": "Number of eigenvectors. In Figure 5, we assess the effect on performance of varying the number of eigenvectors used in the STS (" + }, + { + "bbox": [ + 304, + 594, + 548, + 715 + ], + "type": "inline_equation", + "content": "m \\in \\{1,2,3,4\\}" + }, + { + "bbox": [ + 304, + 594, + 548, + 715 + ], + "type": "text", + "content": "). As can be seen, accuracy improves consistently with more eigenvectors, peaking at " + }, + { + "bbox": [ + 304, + 594, + 548, + 715 + ], + "type": "inline_equation", + "content": "87.48\\%" + }, + { + "bbox": [ + 304, + 594, + 548, + 715 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 304, + 594, + 548, + 715 + ], + "type": "inline_equation", + "content": "m = 4" + }, + { + "bbox": [ + 304, + 594, + 548, + 715 + ], + "type": "text", + "content": ". This suggests that eigenvectors offer complementary information and that their combination enhances the recognition of complex global structures. In the same plot, we also report the performance of the standard VMamba model and a variant using a random traversal strategy. As can be observed, the improvement provided by" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23887" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 47, + 72, + 287, + 106 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 72, + 287, + 106 + ], + "spans": [ + { + "bbox": [ + 47, + 72, + 287, + 106 + ], + "type": "text", + "content": "our STS over VMamba, for any number of eigenvectors, is greater than the one offered by the raster scan of VMamba over a random traversal." + } + ] + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 56, + 118, + 278, + 249 + ], + "blocks": [ + { + "bbox": [ + 56, + 118, + 278, + 249 + ], + "lines": [ + { + "bbox": [ + 56, + 118, + 278, + 249 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 278, + 249 + ], + "type": "image", + "image_path": "133a9b668575984eefb6abf1857bd45d68854cecddde40485323a43e56c3d175.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 47, + 260, + 287, + 283 + ], + "lines": [ + { + "bbox": [ + 47, + 260, + 287, + 283 + ], + "spans": [ + { + "bbox": [ + 47, + 260, + 287, + 283 + ], + "type": "text", + "content": "Figure 5. Model performance for different number of eigenvectors. Four eigenvectors exhibit the best performance." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "bbox": [ + 46, + 290, + 288, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 290, + 288, + 447 + ], + "spans": [ + { + "bbox": [ + 46, + 290, + 288, + 447 + ], + "type": "text", + "content": "Number of nearest neighbors. The K-Nearest Neighbors (KNN) algorithm constructs a sparse adjacency matrix by selecting the top " + }, + { + "bbox": [ + 46, + 290, + 288, + 447 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 290, + 288, + 447 + ], + "type": "text", + "content": " nearest neighbors for each image patch. In Tab. 2, we report our model's performance across " + }, + { + "bbox": [ + 46, + 290, + 288, + 447 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 290, + 288, + 447 + ], + "type": "text", + "content": " values from 5 to 25. As can be seen, our model is not very sensitive to this parameter and achieves its highest accuracy of " + }, + { + "bbox": [ + 46, + 290, + 288, + 447 + ], + "type": "inline_equation", + "content": "87.48\\%" + }, + { + "bbox": [ + 46, + 290, + 288, + 447 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 46, + 290, + 288, + 447 + ], + "type": "inline_equation", + "content": "k = 5" + }, + { + "bbox": [ + 46, + 290, + 288, + 447 + ], + "type": "text", + "content": " neighbors. This suggests that only a few neighbors are necessary to capture the global structure of an image. This has a positive impact on the efficiency of our method as the computational complexity of the Laplacian eigen-decomposition step, using the Arnoldi algorithm, is proportional to the number of non-zero entries in the Laplacian matrix (which is directly proportional to " + }, + { + "bbox": [ + 46, + 290, + 288, + 447 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 46, + 290, + 288, + 447 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 50, + 455, + 285, + 485 + ], + "blocks": [ + { + "bbox": [ + 50, + 455, + 285, + 485 + ], + "lines": [ + { + "bbox": [ + 50, + 455, + 285, + 485 + ], + "spans": [ + { + "bbox": [ + 50, + 455, + 285, + 485 + ], + "type": "table", + "html": "
k510152025
Top-1 Accuracy (%)87.4886.9887.1486.9187.15
", + "image_path": "a8fe63bde651ddcabcba54b8997481c6c9bb5f07a3715ed42ae715ca0745f86f.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 47, + 493, + 287, + 516 + ], + "lines": [ + { + "bbox": [ + 47, + 493, + 287, + 516 + ], + "spans": [ + { + "bbox": [ + 47, + 493, + 287, + 516 + ], + "type": "text", + "content": "Table 2. Model performance for different values of " + }, + { + "bbox": [ + 47, + 493, + 287, + 516 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 47, + 493, + 287, + 516 + ], + "type": "text", + "content": " (number of nearest neighbors). The best performance is achieved at " + }, + { + "bbox": [ + 47, + 493, + 287, + 516 + ], + "type": "inline_equation", + "content": "k = 5" + }, + { + "bbox": [ + 47, + 493, + 287, + 516 + ], + "type": "text", + "content": " ." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 46, + 521, + 287, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 46, + 521, + 287, + 689 + ], + "spans": [ + { + "bbox": [ + 46, + 521, + 287, + 689 + ], + "type": "text", + "content": "Memory Usage and Runtime. Next, we analyze the computational cost and time efficiency of the STS. As mentioned before, the proposed strategy minimizes computations by leveraging the sparsity of the Laplacian matrix and limiting the number of computed eigenvectors. In the following experiments, we use " + }, + { + "bbox": [ + 46, + 521, + 287, + 689 + ], + "type": "inline_equation", + "content": "m = 4" + }, + { + "bbox": [ + 46, + 521, + 287, + 689 + ], + "type": "text", + "content": " eigenvectors and batch size = 1. Memory Usage: As depicted in Fig. 6, our STS approach (blue line) is highly efficient in terms of memory usage compared to the VMamba (orange line). Even when the number of patches increases, the STS based on a sparse eigen-solver, requires only a modest increase in memory compared to the VMamba. The stars along each line indicate the token lengths used in our method, highlighting the specific points where processing occurs across different sequence lengths." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "spans": [ + { + "bbox": [ + 47, + 689, + 287, + 714 + ], + "type": "text", + "content": "Runtime: Our STS strategy, which can run on a CPU, also shows a favorable runtime scaling with increased patch" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 309, + 71, + 545, + 196 + ], + "blocks": [ + { + "bbox": [ + 309, + 71, + 545, + 196 + ], + "lines": [ + { + "bbox": [ + 309, + 71, + 545, + 196 + ], + "spans": [ + { + "bbox": [ + 309, + 71, + 545, + 196 + ], + "type": "image", + "image_path": "6e7566914c0176e2784e903068e1af7c19337fd073c00091bb4d92aae7a3d432.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 203, + 544, + 216 + ], + "lines": [ + { + "bbox": [ + 306, + 203, + 544, + 216 + ], + "spans": [ + { + "bbox": [ + 306, + 203, + 544, + 216 + ], + "type": "text", + "content": "Figure 6. Runtime and memory usage for different patches lengths." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 304, + 232, + 547, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 232, + 547, + 293 + ], + "spans": [ + { + "bbox": [ + 304, + 232, + 547, + 293 + ], + "type": "text", + "content": "length (Fig. 6). Specifically, for a patch length of 196 (marked by a star), which is the setting used in our main experiments, the increase in runtime is minimal when compared to the VMamba backbone, highlighting the efficiency of our approach." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 304, + 293, + 547, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 293, + 547, + 424 + ], + "spans": [ + { + "bbox": [ + 304, + 293, + 547, + 424 + ], + "type": "text", + "content": "FLOPs: The amount of FLOPs for our traversing strategy include the computations required for constructing the adjacency matrix, calculating the Laplacian, and obtaining its first eigenvectors. This calculation results in approximately 2 MB of FLOPs, which is significantly lower than the 3.9 GB FLOPs needed by VMamba's backbone. Interestingly, as reported in Tab. 1, the total number of FLOPs performed by our Spectral VMamba method is less than that of VMamba when using the same model size. This is because, after processing patches, the input to our Spectral VMamba has a size of " + }, + { + "bbox": [ + 304, + 293, + 547, + 424 + ], + "type": "inline_equation", + "content": "14 \\times 14" + }, + { + "bbox": [ + 304, + 293, + 547, + 424 + ], + "type": "text", + "content": " tokens, compared to " + }, + { + "bbox": [ + 304, + 293, + 547, + 424 + ], + "type": "inline_equation", + "content": "56 \\times 56" + }, + { + "bbox": [ + 304, + 293, + 547, + 424 + ], + "type": "text", + "content": " tokens for VMamba." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 306, + 442, + 378, + 454 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 306, + 442, + 378, + 454 + ], + "spans": [ + { + "bbox": [ + 306, + 442, + 378, + 454 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 304, + 462, + 547, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 304, + 462, + 547, + 715 + ], + "spans": [ + { + "bbox": [ + 304, + 462, + 547, + 715 + ], + "type": "text", + "content": "In this paper, we introduce Spectral VMamba, a novel approach to visual state space models that achieves patch traversal rotation invariance. Our method leverages spectral decomposition from the graph Laplacian of image patches to define traversal paths for patch processing. In addition, we develop the RFN module to ensure consistent feature representation in different orientations. This module serves as a foundation for our STS module, enabling patch features that are robust to rotational variations. By incorporating the RFN and STS modules, Spectral VMamba achieves a robust transformation-invariant feature extraction pipeline that maintains high accuracy across a variety of orientations. Our experiments demonstrate that Spectral VMamba consistently outperform popular models including Swin transformer, VMamba, MSVmamba, and LocalVim in image classification tasks. Our method does not yet address generalization to the medical domain. In future work, we aim to handle disconnected graph components, common in medical images with distinct anatomical structures or pathologies. Adapting spectral traversal for these regions will enhance pattern capture and robustness across imaging modalities." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23888" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 145, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 145, + 85 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 145, + 85 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 47, + 91, + 288, + 126 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 47, + 91, + 288, + 126 + ], + "spans": [ + { + "bbox": [ + 47, + 91, + 288, + 126 + ], + "type": "text", + "content": "We appreciate the computational resources and support provided by Compute Canada and the Digital Research Alliance of Canada." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 138, + 107, + 151 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 138, + 107, + 151 + ], + "spans": [ + { + "bbox": [ + 48, + 138, + 107, + 151 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 158, + 288, + 714 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 53, + 158, + 288, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 158, + 288, + 213 + ], + "spans": [ + { + "bbox": [ + 53, + 158, + 288, + 213 + ], + "type": "text", + "content": "[1] Alaaeldin Ali, Hugo Touvron, Mathilde Caron, Piotr Bojanowski, Matthijs Douze, Armand Joulin, Ivan Laptev, Natalia Neverova, Gabriel Synnaeve, Jakob Verbeek, et al. Xcit: Cross-covariance image transformers. NeurIPS, 34:2001-20027, 2021. 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 53, + 214, + 288, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 214, + 288, + 247 + ], + "spans": [ + { + "bbox": [ + 53, + 214, + 288, + 247 + ], + "type": "text", + "content": "[2] Alvin Bayliss, Charles I Goldstein, and Eli Turkel. An iterative method for the helmholtz equation. Journal of Computational Physics, 49(3):443-457, 1983. 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 53, + 249, + 288, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 249, + 288, + 282 + ], + "spans": [ + { + "bbox": [ + 53, + 249, + 288, + 282 + ], + "type": "text", + "content": "[3] Taco Cohen and Max Welling. Group equivariant convolutional networks. In International Conference on Machine Learning (ICML), pages 2990-2999, 2016. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 284, + 288, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 284, + 288, + 315 + ], + "spans": [ + { + "bbox": [ + 53, + 284, + 288, + 315 + ], + "type": "text", + "content": "[4] Taco S Cohen and Max Welling. Steerable cnns. In International Conference on Learning Representations (ICLR), 2017. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 53, + 317, + 288, + 349 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 317, + 288, + 349 + ], + "spans": [ + { + "bbox": [ + 53, + 317, + 288, + 349 + ], + "type": "text", + "content": "[5] Richard Courant and David Hilbert. Methods of mathematical physics: partial differential equations. Interscience Publishers, 1937. 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 53, + 351, + 288, + 393 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 351, + 288, + 393 + ], + "spans": [ + { + "bbox": [ + 53, + 351, + 288, + 393 + ], + "type": "text", + "content": "[6] Sander Dieleman, Jeffrey De Fauw, and Koray Kavukcuoglu. Exploiting cyclic symmetry in convolutional neural networks. International Conference on Machine Learning (ICML), 2016. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 396, + 288, + 462 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 396, + 288, + 462 + ], + "spans": [ + { + "bbox": [ + 53, + 396, + 288, + 462 + ], + "type": "text", + "content": "[7] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 1, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 464, + 288, + 506 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 464, + 288, + 506 + ], + "spans": [ + { + "bbox": [ + 53, + 464, + 288, + 506 + ], + "type": "text", + "content": "[8] Daniel Y Fu, Tri Dao, Khaled K Saab, Armin W Thomas, Atri Rudra, and Christopher Ré. Hungry hungry hippos: Towards language modeling with state space models. arXiv preprint arXiv:2212.14052, 2022. 2, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 53, + 509, + 288, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 509, + 288, + 531 + ], + "spans": [ + { + "bbox": [ + 53, + 509, + 288, + 531 + ], + "type": "text", + "content": "[9] Gene H Golub and Charles F Van Loan. Matrix computations. JHU press, 2013. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 533, + 288, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 533, + 288, + 564 + ], + "spans": [ + { + "bbox": [ + 48, + 533, + 288, + 564 + ], + "type": "text", + "content": "[10] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752, 2023. 2, 3, 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 567, + 288, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 567, + 288, + 599 + ], + "spans": [ + { + "bbox": [ + 48, + 567, + 288, + 599 + ], + "type": "text", + "content": "[11] Albert Gu, Karan Goel, and Christopher Ré. Efficiently modeling long sequences with structured state spaces. arXiv preprint arXiv:2111.00396, 2021. 2, 3, 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 601, + 288, + 654 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 601, + 288, + 654 + ], + "spans": [ + { + "bbox": [ + 48, + 601, + 288, + 654 + ], + "type": "text", + "content": "[12] Albert Gu, Isys Johnson, Karan Goel, Khaled Saab, Tri Dao, Atri Rudra, and Christopher Ré. Combining recurrent, convolutional, and continuous-time models with linear state space layers. Advances in neural information processing systems, 34:572-585, 2021. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 657, + 288, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 657, + 288, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 657, + 288, + 689 + ], + "type": "text", + "content": "[13] Ali Hatamizadeh and Jan Kautz. Mambavision: A hybrid mamba-transformer vision backbone. arXiv preprint arXiv:2407.08083, 2024. 2, 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 48, + 691, + 288, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 691, + 288, + 714 + ], + "spans": [ + { + "bbox": [ + 48, + 691, + 288, + 714 + ], + "type": "text", + "content": "[14] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceed-" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 713 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "spans": [ + { + "bbox": [ + 326, + 73, + 545, + 95 + ], + "type": "text", + "content": "ings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 308, + 96, + 547, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 96, + 547, + 129 + ], + "spans": [ + { + "bbox": [ + 308, + 96, + 547, + 129 + ], + "type": "text", + "content": "[15] Geoffrey E Hinton, Alex Krizhevsky, and Sida D Wang. Matrix capsules with em routing. In International Conference on Learning Representations (ICLR), 2018. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 307, + 130, + 547, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 130, + 547, + 174 + ], + "spans": [ + { + "bbox": [ + 307, + 130, + 547, + 174 + ], + "type": "text", + "content": "[16] Gao Huang, Zhuang Liu, Laurens Van Der Maaten, and Kilian Q Weinberger. Densely connected convolutional networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4700-4708, 2017. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 308, + 175, + 547, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 175, + 547, + 217 + ], + "spans": [ + { + "bbox": [ + 308, + 175, + 547, + 217 + ], + "type": "text", + "content": "[17] Tao Huang, Xiaohuan Pei, Shan You, Fei Wang, Chen Qian, and Chang Xu. Localmamba: Visual state space model with windowed selective scan. arXiv preprint arXiv:2403.09338, 2024. 3, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 308, + 220, + 547, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 220, + 547, + 263 + ], + "spans": [ + { + "bbox": [ + 308, + 220, + 547, + 263 + ], + "type": "text", + "content": "[18] Max Jaderberg, Karen Simonyan, Andrew Zisserman, and Koray Kavukcuoglu. Spatial transformer networks. In Advances in Neural Information Processing Systems (NeurIPS), pages 2017-2025, 2015. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 308, + 265, + 547, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 265, + 547, + 308 + ], + "spans": [ + { + "bbox": [ + 308, + 265, + 547, + 308 + ], + "type": "text", + "content": "[19] Sekou-Oumar Kaba, Arnab Kumar Mondal, Yan Zhang, Yoshua Bengio, and Siamak Ravanbakhsh. Equivariance with learned canonicalization functions. In International Conference on Learning Representations (ICLR), 2023. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 308, + 310, + 545, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 310, + 545, + 365 + ], + "spans": [ + { + "bbox": [ + 308, + 310, + 545, + 365 + ], + "type": "text", + "content": "[20] Dmitry Laptev, Nikolay Savinov, Joachim M Buhmann, and Marc Pollefeys. Ti-pooling: transformation-invariant pooling for feature learning in convolutional neural networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 289-297, 2016. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 308, + 366, + 545, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 366, + 545, + 409 + ], + "spans": [ + { + "bbox": [ + 308, + 366, + 545, + 409 + ], + "type": "text", + "content": "[21] Yue Liu, Yunjie Tian, Yuzhong Zhao, Hongtian Yu, Lingxi Xie, Yaowei Wang, Qixiang Ye, and Yunfan Liu. Vmamba: Visual state space model. arXiv preprint arXiv:2401.10166, 2024. 2, 3, 4, 6, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 308, + 411, + 545, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 411, + 545, + 454 + ], + "spans": [ + { + "bbox": [ + 308, + 411, + 545, + 454 + ], + "type": "text", + "content": "[22] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, pages 10012-10022, 2021. 6, 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 308, + 456, + 545, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 456, + 545, + 510 + ], + "spans": [ + { + "bbox": [ + 308, + 456, + 545, + 510 + ], + "type": "text", + "content": "[23] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 308, + 512, + 547, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 512, + 547, + 565 + ], + "spans": [ + { + "bbox": [ + 308, + 512, + 547, + 565 + ], + "type": "text", + "content": "[24] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11976-11986, 2022. 1" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 308, + 567, + 547, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 567, + 547, + 600 + ], + "spans": [ + { + "bbox": [ + 308, + 567, + 547, + 600 + ], + "type": "text", + "content": "[25] Jun Ma, Feifei Li, and Bo Wang. U-mamba: Enhancing long-range dependency for biomedical image segmentation. arXiv preprint arXiv:2401.04722, 2024. 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 308, + 601, + 545, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 601, + 545, + 634 + ], + "spans": [ + { + "bbox": [ + 308, + 601, + 545, + 634 + ], + "type": "text", + "content": "[26] Harsh Mehta, Ankit Gupta, Ashok Cutkosky, and Behnam Neyshabur. Long range language modeling via gated state spaces. arXiv preprint arXiv:2206.13947, 2022. 3" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 308, + 635, + 547, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 635, + 547, + 679 + ], + "spans": [ + { + "bbox": [ + 308, + 635, + 547, + 679 + ], + "type": "text", + "content": "[27] Arnab Kumar Mondal, Siba Smarak Panigrahi, Oumar Kaba, Sai Rajeswar Mudumba, and Siamak Ravanbakhsh. Equivariant adaptation of large pretrained models. Advances in Neural Information Processing Systems, 36:50293-50309, 2023. 3" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 308, + 681, + 545, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 308, + 681, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 308, + 681, + 545, + 713 + ], + "type": "text", + "content": "[28] Andrew Ng, Michael Jordan, and Yair Weiss. On spectral clustering: Analysis and an algorithm. Advances in neural information processing systems, 14, 2001. 6" + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "23889" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 48, + 72, + 288, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "spans": [ + { + "bbox": [ + 48, + 72, + 287, + 116 + ], + "type": "text", + "content": "[29] Omri Puny, Matan Atzmon, Heli Ben-Hamu, Ishan Misra, Aditya Grover, Edward J Smith, and Yaron Lipman. Frame averaging for invariant and equivariant network design. arXiv preprint arXiv:2110.03336, 2021. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 49, + 118, + 288, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 49, + 118, + 288, + 150 + ], + "spans": [ + { + "bbox": [ + 49, + 118, + 288, + 150 + ], + "type": "text", + "content": "[30] Martin Reuter, Franz-Erich Wolter, and Niklas Peinecke. Laplace-beltrami spectra as 'shape-dna' of surfaces and solids. Computer-Aided Design, 38(4):342-366, 2006. 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 48, + 152, + 288, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 152, + 288, + 185 + ], + "spans": [ + { + "bbox": [ + 48, + 152, + 288, + 185 + ], + "type": "text", + "content": "[31] Sara Sabour, Nicholas Frosst, and Geoffrey E Hinton. Dynamic routing between capsules. In Advances in Neural Information Processing Systems (NeurIPS), 2017. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 48, + 186, + 288, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 186, + 288, + 217 + ], + "spans": [ + { + "bbox": [ + 48, + 186, + 288, + 217 + ], + "type": "text", + "content": "[32] Jianbo Shi and Jitendra Malik. Normalized cuts and image segmentation. IEEE Transactions on pattern analysis and machine intelligence, 22(8):888-905, 2000. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 48, + 220, + 288, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 220, + 288, + 252 + ], + "spans": [ + { + "bbox": [ + 48, + 220, + 288, + 252 + ], + "type": "text", + "content": "[33] Yuheng Shi, Minjing Dong, and Chang Xu. Multi-scale vmamba: Hierarchy in hierarchy visual state space model. arXiv preprint arXiv:2405.14174, 2024. 2, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 48, + 254, + 288, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 254, + 288, + 296 + ], + "spans": [ + { + "bbox": [ + 48, + 254, + 288, + 296 + ], + "type": "text", + "content": "[34] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. In International conference on learning representations, 2014. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 48, + 298, + 288, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 298, + 288, + 331 + ], + "spans": [ + { + "bbox": [ + 48, + 298, + 288, + 331 + ], + "type": "text", + "content": "[35] Jimmy TH Smith, Andrew Warrington, and Scott W Linderman. Simplified state space layers for sequence modeling. arXiv preprint arXiv:2208.04933, 2022. 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 48, + 333, + 288, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 333, + 288, + 374 + ], + "spans": [ + { + "bbox": [ + 48, + 333, + 288, + 374 + ], + "type": "text", + "content": "[36] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In International conference on machine learning, pages 6105-6114. PMLR, 2019. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 48, + 377, + 288, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 377, + 288, + 431 + ], + "spans": [ + { + "bbox": [ + 48, + 377, + 288, + 431 + ], + "type": "text", + "content": "[37] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In Int. Conf. Machine Learning, pages 10347-10357, 2021. 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 48, + 434, + 288, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 434, + 288, + 487 + ], + "spans": [ + { + "bbox": [ + 48, + 434, + 288, + 487 + ], + "type": "text", + "content": "[38] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In International conference on machine learning, pages 10347-10357. PMLR, 2021. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 48, + 489, + 288, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 489, + 288, + 533 + ], + "spans": [ + { + "bbox": [ + 48, + 489, + 288, + 533 + ], + "type": "text", + "content": "[39] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 48, + 534, + 288, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 534, + 288, + 576 + ], + "spans": [ + { + "bbox": [ + 48, + 534, + 288, + 576 + ], + "type": "text", + "content": "[40] Oriol Vinyals, Charles Blundell, Timothy Lillicrap, Daan Wierstra, et al. Matching networks for one shot learning. Advances in neural information processing systems, 29, 2016. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 48, + 579, + 288, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 579, + 288, + 611 + ], + "spans": [ + { + "bbox": [ + 48, + 579, + 288, + 611 + ], + "type": "text", + "content": "[41] Maurice Weiler and Gabriele Cesa. General E(2)-Equivariant Steerable CNNs. In Conference on Neural Information Processing Systems (NeurIPS), 2019. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 48, + 613, + 288, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 613, + 288, + 656 + ], + "spans": [ + { + "bbox": [ + 48, + 613, + 288, + 656 + ], + "type": "text", + "content": "[42] Maurice Weiler, Fred A. Hamprecht, and Martin Storath. Learning steerable filters for rotation equivariant cnns. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 849-858, 2018. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 48, + 658, + 288, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 658, + 288, + 689 + ], + "spans": [ + { + "bbox": [ + 48, + 658, + 288, + 689 + ], + "type": "text", + "content": "[43] Fei Xie, Weijia Zhang, Zhongdao Wang, and Chao Ma. Quadmamba: Learning quadtree-based selective scan for visual state space model. arXiv preprint arXiv:2410.06806, 2024. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 48, + 692, + 288, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 48, + 692, + 288, + 713 + ], + "spans": [ + { + "bbox": [ + 48, + 692, + 288, + 713 + ], + "type": "text", + "content": "[44] Xiaosong Zhang, Yunjie Tian, Lingxi Xie, Wei Huang, Qi Dai, Qixiang Ye, and Qi Tian. Hivit: A simpler and more efficient" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 307, + 73, + 547, + 195 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 327, + 73, + 547, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 73, + 547, + 105 + ], + "spans": [ + { + "bbox": [ + 327, + 73, + 547, + 105 + ], + "type": "text", + "content": "design of hierarchical vision transformer. In The Eleventh International Conference on Learning Representations, 2023. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 307, + 107, + 547, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 107, + 547, + 150 + ], + "spans": [ + { + "bbox": [ + 307, + 107, + 547, + 150 + ], + "type": "text", + "content": "[45] Yanzhao Zhou, Qixiang Ye, Qiang Qiu, and Jianbin Jiao. Oriented response networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 519-528, 2017. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 307, + 152, + 547, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 307, + 152, + 547, + 195 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 547, + 195 + ], + "type": "text", + "content": "[46] Lianghui Zhu, Bencheng Liao, Qian Zhang, Xinlong Wang, Wenyu Liu, and Xinggang Wang. Vision mamba: Efficient visual representation learning with bidirectional state space model. arXiv preprint arXiv:2401.09417, 2024. 2, 6, 7" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "23890" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/SpectroMotion_ Dynamic 3D Reconstruction of Specular Scenes/5055bce6-6fec-402b-adcc-413e1a1becdd_content_list.json b/2025/SpectroMotion_ Dynamic 3D Reconstruction of Specular Scenes/5055bce6-6fec-402b-adcc-413e1a1becdd_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..438004d0c43102bda0f061dcf79d0e05a4e7685a --- /dev/null +++ b/2025/SpectroMotion_ Dynamic 3D Reconstruction of Specular Scenes/5055bce6-6fec-402b-adcc-413e1a1becdd_content_list.json @@ -0,0 +1,1843 @@ +[ + { + "type": "text", + "text": "SpectroMotion: Dynamic 3D Reconstruction of Specular Scenes", + "text_level": 1, + "bbox": [ + 176, + 130, + 818, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Cheng-De Fan1 Chen-Wei Chang1 Yi-Ruei Liu1,2 Jie-Ying Lee1 \nJiun-Long Huang1 Yu-Chee Tseng1 Yu-Lun Liu1", + "bbox": [ + 228, + 179, + 790, + 218 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ National Yang Ming Chiao Tung University $^{2}$ University of Illinois Urbana-Champaign", + "bbox": [ + 153, + 226, + 861, + 244 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 279, + 326, + 295 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We present SpectroMotion, a novel approach that combines 3D Gaussian Splatting (3DGS) with physically-based rendering (PBR) and deformation fields to reconstruct dynamic specular scenes. Previous methods extending 3DGS to model dynamic scenes have struggled to represent specular surfaces accurately. Our method addresses this limitation by introducing a residual correction technique for accurate surface normal computation during deformation, complemented by a deformable environment map that adapts to time-varying lighting conditions. We implement a coarse-to-fine training strategy significantly enhancing scene geometry and specular color prediction. It is the only existing 3DGS method capable of synthesizing photorealistic real-world dynamic specular scenes, outperforming state-of-the-art methods in rendering complex, dynamic, and specular scenes. Please see our project page at cdfan0627.github.io/spectromotion.", + "bbox": [ + 89, + 313, + 485, + 570 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 601, + 220, + 616 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3D Gaussian Splitting (3DGS) [16] has emerged as a groundbreaking technique in 3D scene reconstruction, offering fast training and real-time rendering capabilities. By representing 3D scene using a collection of 3D Gaussians and employing a point-based rendering approach, 3DGS has significantly improved efficiency in novel view synthesis. However, extending 3DGS to model dynamic scenes, especially those containing specular surfaces accurately, has remained a significant challenge.", + "bbox": [ + 89, + 626, + 483, + 762 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Existing extensions of 3DGS have made progress in either dynamic scene reconstruction or specular object rendering, but none have successfully combined both aspects. Methods tackling dynamic scenes often struggle with accurately representing specular surfaces, while those focusing on specular rendering are limited to static scenes. This capability gap has hindered the application of 3DGS to real-world scenarios where both motion and specular reflections are present.", + "bbox": [ + 89, + 763, + 483, + 883 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We present SpectroMotion, a novel approach that ad", + "bbox": [ + 109, + 885, + 483, + 900 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c78dbd9e6a09ad3b87ada87fe470b475a575a7b82e7cb685a9ec4005fe669403.jpg", + "image_caption": [ + "Figure 1. Our method, SpectroMotion, recovers and renders dynamic scenes with higher-quality reflections compared to prior work. It introduces physical normal estimation, deformable environment maps, and a coarse-to-fine training strategy to achieve superior results in rendering dynamic scenes with reflections. Here, we present a rendered test image, corresponding normal maps, and a ground-truth image, where the ground-truth normal map (used as a reference) is generated using a pre-trained normal estimator [6]. For Deformable 3DGS, we use the shortest axes of the deformed 3D Gaussians as the normals. We have highlighted the specular regions to demonstrate the effectiveness of our approach." + ], + "image_footnote": [], + "bbox": [ + 517, + 280, + 903, + 503 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "dresses these limitations by combining 3D Gaussian Splating with physically based rendering (PBR) and deformation fields. Our method introduces three key innovations: a residual correction technique for accurate surface normal computation during deformation, a deformable environment map that adapts to time-varying lighting conditions, and a coarse-to-fine training strategy that significantly enhances scene geometry and specular color prediction.", + "bbox": [ + 511, + 699, + 906, + 820 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Our evaluations demonstrate that SpectroMotion outperforms prior methods in view synthesis of scenes containing dynamic specular objects, as illustrated in Fig. 1. It is the only 3DGS method capable of synthesizing photorealistic real-world dynamic specular scenes, surpassing state-of-the", + "bbox": [ + 511, + 825, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 808, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "21328", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "art techniques in rendering complex, dynamic, and specular content. This advancement represents a significant leap in 3D scene reconstruction, particularly for challenging scenarios involving moving specular objects.", + "bbox": [ + 89, + 90, + 483, + 151 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, we make the following contributions:", + "bbox": [ + 109, + 152, + 444, + 167 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose SpectroMotion, a physically-based rendering (PBR) approach combining deformation fields and 3D Gaussian Splatting for real-world dynamic specular scenes.", + "- We introduce a residual correction method for accurate surface normals during deformation, coupled with a deformable environment map to handle time-varying lighting conditions in dynamic scenes.", + "- We develop a coarse-to-fine training strategy enhancing scene geometry and specular color prediction, outperforming state-of-the-art methods." + ], + "bbox": [ + 89, + 167, + 483, + 334 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 349, + 230, + 364 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Dynamic Scene Reconstruction", + "text_level": 1, + "bbox": [ + 89, + 375, + 364, + 390 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recent works have leveraged NeRF representations to jointly solve for canonical space and deformation fields in dynamic scenes using RGB supervision [4, 9, 19, 26, 27, 31, 32, 34, 39, 45, 47]. Further advancements in dynamic neural rendering include object segmentation [36], incorporation of depth information [1], utilization of 2D CNNs for scene priors [24, 33], and multi-view video compression [18]. However, these NeRF-based methods are computationally intensive, limiting their practical applications. To address this, 3D Gaussian Splatting [16] has emerged as a promising alternative, offering real-time rendering capabilities while maintaining high visual quality. Building upon this efficient representation, recent research has adapted 3D Gaussians for dynamic scenes [11, 22, 29, 38, 44, 46, 49]. Nevertheless, these approaches do not explicitly account for changes in surface normal during the dynamic process. Our work extends this line of research by combining specular object rendering based on normal estimation with a deformation field, enabling each 3D Gaussian to model dynamic specular scenes effectively.", + "bbox": [ + 89, + 397, + 483, + 700 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Reflective Object Rendering", + "text_level": 1, + "bbox": [ + 89, + 710, + 344, + 728 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "While significant progress has been made in rendering reflective objects, challenges from complex light interactions persist. Recent years have seen numerous studies addressing these issues, primarily by decomposing appearance into lighting and material properties [2, 3, 17, 30, 37, 41, 54-56]. Building on this foundation, some research has focused on improving the capture and reproduction of specular reflections [28, 40, 42, 51]. In contrast, others have leveraged signed distance functions (SDFs) to enhance normal estimation [8, 20, 21, 25, 53]. The emergence of 3D Gaussian Splatting (3DGS) has sparked a new wave of techniques", + "bbox": [ + 89, + 734, + 485, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "[7, 12, 23, 35, 50, 57] that integrate Gaussian splating with physically-based rendering. Nevertheless, accurately modeling dynamic environments and time-varying specular reflections remains a significant challenge. To address this limitation, our work introduces a novel approach incorporating a deformable environment map and additional explicit Gaussian attributes specifically designed to capture specular color changes over time.", + "bbox": [ + 511, + 90, + 908, + 212 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 513, + 226, + 604, + 242 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Overview of the approach. The overview of our method is illustrated in Fig. 2. Given an input monocular video sequence of frames and corresponding camera poses, we design a three-stage approach to reconstruct the dynamic specular scene, as detailed in Sec. 3.2. Accurate specular reflection requires precise normal estimation, so Sec. 3.3 elaborates on how we estimate normals in dynamic scenes. Finally, we introduce the losses used throughout the training process in Sec. 3.4.", + "bbox": [ + 511, + 251, + 908, + 388 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3.1. Preliminary", + "text_level": 1, + "bbox": [ + 513, + 398, + 642, + 414 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3D Gaussian Splatting. Each 3D Gaussian is defined by a center position $\\boldsymbol{x} \\in \\mathbb{R}^3$ and a covariance matrix $\\Sigma$ . 3D Gaussian Splatting [16] optimizes the covariance matrix using scaling factors $s \\in \\mathbb{R}^3$ and rotation unit quaternion $r \\in \\mathbb{R}^4$ . For novel-view rendering, 3D Gaussians are projected onto 2D camera planes using differentiable splatting [52]:", + "bbox": [ + 511, + 422, + 906, + 515 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\Sigma} ^ {\\prime} = \\mathbf {J} \\mathbf {W} \\boldsymbol {\\Sigma} \\mathbf {W} ^ {T} \\mathbf {J} ^ {T}. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 642, + 525, + 906, + 542 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Pixel colors are computed using point-based volumetric rendering:", + "bbox": [ + 511, + 555, + 908, + 585 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\nC = \\sum_ {i \\in N} T _ {i} \\alpha_ {i} c _ {i}, \\quad \\alpha_ {i} = \\sigma_ {i} e ^ {- \\frac {1}{2} (\\boldsymbol {x}) ^ {T} \\boldsymbol {\\Sigma} ^ {\\prime} (\\boldsymbol {x})}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 595, + 906, + 630 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "where $T_{i} = \\prod_{j=1}^{i-1}(1 - \\alpha_{j})$ is the transmittance, $\\sigma_{i}$ is the opacity, and $\\mathbf{c}_{i}$ is the color of each 3D Gaussian.", + "bbox": [ + 511, + 643, + 906, + 675 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3.2. Specular Rendering", + "text_level": 1, + "bbox": [ + 511, + 684, + 705, + 700 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Since accurate reflections depend heavily on precise geometry, we implement a three-stage coarse-to-fine training strategy: static, dynamic, and specular stages. This approach ensures both stable scene geometry and accurate specular rendering.", + "bbox": [ + 511, + 707, + 908, + 784 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3.2.1. Coarse-to-Fine Training Strategy", + "text_level": 1, + "bbox": [ + 511, + 791, + 790, + 806 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Static stage. In the static stage, we employ vanilla 3DGS [16] for static scene reconstruction to stabilize the geometry of the static scene. Specifically, we optimize the position $\\mathbf{x}$ , scaling $s$ , rotation $r$ , opacity $\\alpha$ , and coefficients of spherical harmonics (SH) of the 3D Gaussians by minimizing the photometric loss $\\mathcal{L}_{\\mathrm{color}}$ identical to 3DGS [16].", + "bbox": [ + 511, + 810, + 908, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "21329", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/2fe4de96c4e1a699f603f34df7fee1b9f4e35d3c3ff1b8bcea1ebf92689f0f0f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 94, + 88, + 336, + 265 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/7c2262654f0f8b253faefba07eab885d64d180700e9c207ad6f7c94f2122b8a4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 343, + 88, + 905, + 266 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/4cf7c210cddea336779c2a496eb78646127e1383a8ba0777c5f44f6ea13d1a4a.jpg", + "image_caption": [ + "Figure 2. Method Overview. Our method stabilizes the scene geometry through three stages. In the static stage, we stabilize the geometry of the static scene by minimizing photometric loss $\\mathcal{L}_{\\mathrm{color}}$ between vanilla 3DGS renders and ground truth images. The dynamic stage combines canonical 3D Gaussians $\\mathbf{G}$ with a deformable Gaussian MLP to model dynamic scenes while simultaneously minimizing normal loss $\\mathcal{L}_{\\mathrm{normal}}$ between rendered normal map $\\mathbf{N}^t$ and gradient normal map from depth map $\\mathbf{D}^t$ , thus further enhancing the overall scene geometry. Finally, the specular stage introduces a deformable reflection MLP to handle changing environment lighting, deforming reflection directions $\\omega_r^t$ to query a canonical environment map for specular color $\\mathbf{c}_s^t$ . It is then combined with diffuse color $\\mathbf{c_d}$ (using zero-order spherical harmonics) and learnable specular tint $\\mathbf{s}_{\\mathrm{int}}$ per 3D Gaussian to obtain the final color $\\mathbf{c}_{\\mathrm{final}}^t$ . This approach enables the modeling of dynamic specular scenes and high-quality novel view rendering." + ], + "image_footnote": [], + "bbox": [ + 94, + 270, + 905, + 454 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Dynamic stage. Following the static stage, we address dynamic objects using Deformable 3DGS [49]. For each 3D Gaussian in canonical 3D Gaussians $\\mathbf{G}$ , we input its position $\\pmb{x}$ and time $t$ into a deformable Gaussian MLP with parameters $\\theta_{G}$ to predict position, rotation, and scaling residuals: $(\\Delta \\pmb{x}^{t}, \\Delta \\pmb{r}^{t}, \\Delta \\pmb{s}^{t}) = F_{\\theta_{G}}(\\gamma(x), \\gamma(t))$ , where $\\gamma$ denotes positional encoding. Attributes of the corresponding 3D Gaussian in deformed 3D Gaussians $\\mathbf{G}^{t}$ at time $t$ is obtained by $(\\pmb{x}^{t}, \\pmb{r}^{t}, \\pmb{s}^{t}) = (\\Delta \\pmb{x}^{t}, \\Delta \\pmb{r}^{t}, \\Delta \\pmb{s}^{t}) + (\\pmb{x}, \\pmb{r}, \\pmb{s})$ .", + "bbox": [ + 88, + 603, + 482, + 738 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This approach separates motion and geometric structural learning, allowing accurate simulation of dynamic behaviors while maintaining a stable geometric reference. To further enhance scene geometry, we estimate normals of deformed 3D Gaussians and optimize them using:", + "bbox": [ + 89, + 739, + 482, + 814 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {n o r m a l}} = 1 - \\mathbf {N} ^ {t} \\cdot \\hat {\\mathbf {N}} ^ {t}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 207, + 825, + 482, + 843 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbf{N}^t$ is the rendered normal map and $\\hat{\\mathbf{N}}^t$ is the normal map derived from the rendered depth map $\\mathbf{D}^t$ . This process improves local associations among 3D Gaussians and opti", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "mizes both depth and normal information across the entire scene.", + "bbox": [ + 511, + 604, + 905, + 632 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Specular stage. We adopt an image-based lighting (IBL) model, where the environment light is given by a learnable cubemap. Following the rendering equation [14], splitsum approximation [15, 30], and Cook-Torrance reflectance model [5], the outgoing radiance of the specular component $L_{s}$ is expressed as:", + "bbox": [ + 511, + 637, + 906, + 728 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} L _ {s} = \\int_ {\\Omega} \\frac {D G F}{4 (\\omega_ {o} ^ {t} \\cdot \\mathbf {n} ^ {t}) (\\omega_ {i} \\cdot \\mathbf {n} ^ {t})} (\\boldsymbol {\\omega} _ {i} \\cdot \\mathbf {n} ^ {t}) d \\boldsymbol {\\omega} _ {i} \\\\ \\times \\int_ {\\Omega} L _ {i} \\left(\\omega_ {i}\\right) D \\left(\\omega_ {i}, \\omega_ {o} ^ {t}\\right) \\left(\\omega_ {i} \\cdot \\mathbf {n} ^ {t}\\right) d \\omega_ {i}, \\tag {4} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 557, + 734, + 906, + 801 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\Omega$ is the hemisphere around the surface normal $\\mathbf{n}^t$ (describe in Sec. 3.3.) $D$ , $G$ , and $F$ represent the GGX normal distribution function [43], geometric attenuation, and Fresnel term, respectively. $\\omega_o^t$ is the view direction, and $L_i(\\omega_i)$ is the incident radiance. In the first term, we follow the GaussianShader [13] directly computed by $\\mathbf{s}_{\\mathrm{int}} * F_1 + F_2$ ,", + "bbox": [ + 511, + 809, + 906, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "21330", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $F_{1}$ and $F_{2}$ are two pre-computed scalars depending on roughness $\\rho$ , view direction $\\omega_{o}^{t}$ and normal $\\mathbf{n}^t$ . Roughness $\\rho \\in [0,1]$ and specular tint $\\mathbf{s}_{\\mathrm{int}} \\in [0,1]^3$ are learnable parameters for each 3D Gaussian. The second term is pre-integrated in a filtered learnable cubemap, where each mip-level corresponds to a specific roughness value. The cubemap can be queried using the reflection direction to obtain the value of the second term. After the static and dynamic stages, the geometry is well-defined. This allows us to calculate reflection directions $\\omega_{r}^{t} = 2(\\omega_{o}^{t} \\cdot \\mathbf{n}^{t})\\mathbf{n}^{t} - \\omega_{o}^{t}$ accurately.", + "bbox": [ + 89, + 90, + 483, + 256 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$L_{s}$ represents only the specular color of the static environment light. To handle time-varying lighting in dynamic scenes, we introduce a deformable environment map, detailed in the following section.", + "bbox": [ + 89, + 257, + 483, + 316 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.2. Deformable Environment Map for Dynamic Lighting.", + "text_level": 1, + "bbox": [ + 89, + 325, + 483, + 354 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The concept of a deformable environment map involves treating the vanilla environment cubemap as a canonical environment map and combining it with a deformation field. This approach enables us to model time-varying lighting conditions effectively. We first apply positional encoding to the reflection direction $\\omega_r^t$ and time $t$ to implement this. These encoded values are then input into a deformable reflection MLP with parameters $\\theta_R$ . This process allows us to obtain the deformed reflection residual $\\Delta \\bar{\\omega}_r^t = F_{\\theta_R}(\\gamma (\\omega_r^t),\\gamma (t))$ for each specified time $t$ .", + "bbox": [ + 89, + 359, + 483, + 508 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Subsequently, we add the deformed reflection residual $\\Delta \\bar{\\omega}_r^t$ to the reflection direction $\\omega_r^t$ , yielding the deformed reflection direction $\\bar{\\omega}_r^t = \\Delta \\bar{\\omega}_r^t + \\omega_r^t$ .", + "bbox": [ + 89, + 511, + 482, + 556 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We can then use this deformed reflection direction $\\bar{\\omega}_r^t$ to query the canonical environment map. The queried value is then multiplied by the first term of Equation 4, allowing us to obtain time-varying specular color $\\mathbf{c}_s^t$ . This approach effectively captures the dynamic nature of lighting in the scene while maintaining a stable canonical reference.", + "bbox": [ + 89, + 556, + 483, + 647 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.3. Color Decomposition and Staged Training Strategy.", + "text_level": 1, + "bbox": [ + 89, + 655, + 482, + 685 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We decompose the final color $\\mathbf{c}_{\\mathrm{final}}^t$ into diffuse and specular components to better distinguish between high and low-frequency information: $\\mathbf{c}_{\\mathrm{final}}^t = \\mathbf{c}_d + \\mathbf{c}_s^t$ , where $\\mathbf{c_d}$ is the diffuse color (using zero-order spherical harmonics as view-independent color), and $\\mathbf{c_s}^t$ is the view-dependent color component. To manage the transition from spherical harmonics to $\\mathbf{c}_{\\mathrm{final}}^t$ and mitigate potential geometry disruptions, in the early specular stage, we fix the deformable Gaussian MLP and most 3D Gaussian attributes, optimizing only zero-order SH, specular tint, and roughness. We temporarily suspend densification during this phase. As $\\mathbf{c}_{\\mathrm{final}}^t$ becomes more complete, we gradually resume optimization of all parameters and reinstate the densification process.", + "bbox": [ + 89, + 688, + 483, + 883 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We further split the specular stage into two parts, applying", + "bbox": [ + 109, + 886, + 482, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "a coarse-to-fine strategy to the environment map. In the first part, we focus on optimizing the canonical environment map for time-invariant lighting. This establishes a stable foundation for the overall lighting structure. In the second part, we proceed to optimize the deformable reflection MLP for dynamic elements. This approach ensures a more robust learning process, allowing us to capture the static lighting conditions before introducing the complexities of dynamic components.", + "bbox": [ + 511, + 90, + 906, + 227 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Physical Normal Estimation", + "text_level": 1, + "bbox": [ + 511, + 234, + 764, + 250 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Challenges in normal estimation for 3D Gaussians. Normal estimation is essential for modeling specular objects in 3D Gaussians, where GaussianShader [12] initially used the shortest axis combined with a residual normal for approximation. While this works for static scenes, it becomes problematic with deformed Gaussians because the residual should vary at each time step. A straightforward approach of rotating the residual normal based on quaternion differences between canonical and deformed states proves insufficient, as it does not account for shape changes during deformation. When deformation alters the relative axis lengths, the shortest axis assumption breaks down. This highlights the need for a more comprehensive approach that considers both rotational and shape deformation effects to achieve accurate normal estimation for dynamic specular objects.", + "bbox": [ + 511, + 256, + 906, + 482 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Improved rotation calculation for deformed 3D Gaussians. To overcome the limitations of naive methods and accurately model the normal of deformed 3D Gaussians, we propose using both the shortest and longest axes of canonical and deformed Gaussians to compute the rotation matrix. This approach accounts for both rotation and shape changes during deformation. We first align the deformed Gaussian's axes with those of the canonical Gaussian using the following method:", + "bbox": [ + 511, + 486, + 908, + 622 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {v} _ {s} ^ {t} = \\left\\{ \\begin{array}{l l} \\mathbf {v} _ {s} ^ {t} & \\text {i f} \\mathbf {v} _ {s} \\cdot \\mathbf {v} _ {s} ^ {t} > 0, \\\\ - \\mathbf {v} _ {s} ^ {t} & \\text {o t h e r w i s e .} \\end{array} , \\quad \\mathbf {v} _ {l} ^ {t} = \\left\\{ \\begin{array}{l l} \\mathbf {v} _ {l} ^ {t} & \\text {i f} \\mathbf {v} _ {l} \\cdot \\mathbf {v} _ {l} ^ {t} > 0, \\\\ - \\mathbf {v} _ {l} ^ {t} & \\text {o t h e r w i s e .} \\end{array} , \\right. \\right. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 511, + 628, + 934, + 683 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{v}_s$ and $\\mathbf{v}_l$ represent the shortest and longest axes of canonical 3D Gaussians, while $\\mathbf{v}_s^t$ and $\\mathbf{v}_l^t$ denote the same for deformed 3D Gaussians. We then construct orthogonal matrices using these aligned axes and their cross-products:", + "bbox": [ + 511, + 690, + 906, + 751 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {U} = \\left[ \\begin{array}{l l l} \\mathbf {v} _ {s} & \\mathbf {v} _ {l} & \\mathbf {v} _ {s} \\times \\mathbf {v} _ {l} \\end{array} \\right], \\quad \\mathbf {V} ^ {t} = \\left[ \\begin{array}{l l l} \\mathbf {v} _ {s} ^ {t} & \\mathbf {v} _ {l} ^ {t} & \\mathbf {v} _ {s} ^ {t} \\times \\mathbf {v} _ {l} ^ {t} \\end{array} \\right]. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 755, + 906, + 787 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Finally, we derive the rotation matrix $\\mathbf{R}^t = \\mathbf{V}^t\\mathbf{U}^\\top$", + "bbox": [ + 511, + 787, + 856, + 801 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Adjusting normal residuals and ensuring accuracy. To account for shape changes during deformation, we scale the normal residual based on the ratio of oblateness $\\frac{\\beta}{\\beta^t}$ between canonical and deformed 3D Gaussians.", + "bbox": [ + 511, + 806, + 906, + 866 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\beta = \\frac {\\left| \\mathbf {v} _ {l} \\right| - \\left| \\mathbf {v} _ {s} \\right|}{\\left| \\mathbf {v} _ {l} \\right|}, \\quad \\beta^ {t} = \\frac {\\left| \\mathbf {v} _ {l} ^ {t} \\right| - \\left| \\mathbf {v} _ {s} ^ {t} \\right|}{\\left| \\mathbf {v} _ {l} ^ {t} \\right|}, \\qquad (7)\n$$\n", + "text_format": "latex", + "bbox": [ + 583, + 869, + 906, + 905 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "21331", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/7d81696dc4d71bf5a92c447c20b447b27990ec5ce755edc11e1da7a34f623d56.jpg", + "image_caption": [ + "Figure 3. Normal estimation. (a) shows that flatter 3D Gaussians align better with scene surfaces, their shortest axis closely matching the surface normal. In contrast, less flat 3D Gaussians fit less accurately, with their shortest axis diverging from the surface normal. (b) shows that when the deformed 3D Gaussian becomes flatter ( $t = t_1$ ), normal residual $\\Delta \\mathbf{n}$ is rotated by $\\mathbf{R}_1^t$ and scaled down by $\\frac{\\beta}{\\beta_1^t}$ , as flatter Gaussians require smaller normal residuals. Conversely, when the deformation results in a less flat shape ( $t = t_2$ ), $\\Delta \\mathbf{n}$ is rotated by $\\mathbf{R}_2^t$ and amplified by $\\frac{\\beta}{\\beta_2^t}$ , requiring a larger correction to align the shortest axis with the surface normal. (c) shows how $\\gamma^k$ changes with $w$ (where $w = \\frac{|v_s^t|}{|v_l^t|}$ ) for $k = 1, k = 5$ , and $k = 50$ . Larger $w$ indicates less flat Gaussians, while smaller $w$ represents flatter Gaussians. As $k$ increases, $\\gamma^k$ decreases more steeply as $w$ rises. For $k = 5$ , we observe a balanced behavior: $\\gamma^k$ approaches 1 for low $w$ and 0 for high $w$ , providing a nuanced penalty adjustment across different Gaussian shapes." + ], + "image_footnote": [], + "bbox": [ + 91, + 87, + 377, + 204 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/005bf15805ae5f13c0468f2a75422b9d04cc850c1889841e0d7532cde013b4f5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 372, + 85, + 725, + 203 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/da8685942c3cea737a4584c8c01ed14ef155082cd41d1dfe232bab939e1bbb72.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 727, + 85, + 908, + 203 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\beta$ and $\\beta^t$ represent the oblateness of canonical and deformed 3D Gaussians, respectively. This is because flatter 3D Gaussians tend to align more closely with the surface, meaning their shortest axis becomes more aligned with the surface normal, as shown in Fig. 3 (a). In such cases, less compensation from the normal residual is needed. Conversely, less flat Gaussians require more compensation, as illustrated in Fig. 3 (b). We then obtain deformed normal residuals:", + "bbox": [ + 89, + 367, + 483, + 488 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\Delta \\mathbf {n} ^ {t} = \\frac {\\beta}{\\beta^ {t}} \\mathbf {R} ^ {t} \\Delta \\mathbf {n}. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 223, + 497, + 483, + 527 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The final normal $\\mathbf{n}^t$ is computed by adding this residual to the shortest axis and ensuring outward orientation:", + "bbox": [ + 89, + 539, + 482, + 570 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {n} ^ {t} = \\Delta \\mathbf {n} ^ {t} + \\mathbf {v} _ {s} ^ {t}, \\quad \\mathbf {n} ^ {t} = \\left\\{ \\begin{array}{l l} \\mathbf {n} ^ {t} & \\text {i f} \\mathbf {n} ^ {t} \\cdot \\omega_ {o} ^ {t} > 0, \\\\ - \\mathbf {n} ^ {t} & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 114, + 579, + 483, + 619 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This approach adjusts for Gaussian flatness and ensures accurate normal estimation.", + "bbox": [ + 89, + 628, + 483, + 657 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Loss Functions", + "text_level": 1, + "bbox": [ + 89, + 667, + 243, + 681 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Normal regularization. To allow the normal residual to correct the normal while not excessively influencing the optimization of the shortest axis towards the surface normal, we introduce a penalty term for the normal residual:", + "bbox": [ + 89, + 689, + 483, + 750 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {r e g}} = \\gamma^ {k} \\| \\Delta \\mathbf {n} \\| _ {2} ^ {2} \\quad \\text {w h e r e} \\quad \\gamma = \\sqrt {1 - \\frac {\\left| \\mathbf {v} _ {s} ^ {t} \\right| ^ {2}}{\\left| \\mathbf {v} _ {l} ^ {t} \\right| ^ {2}}}. \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 119, + 761, + 483, + 801 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In our experiments, we set $k = 5$ . When $k = 5$ , less flatter 3D Gaussians have $\\gamma^k$ approaching 0. Their shortest axis aligns poorly with the surface normal, requiring more normal residual correction and smaller penalties. Conversely, flatter Gaussians have $\\gamma^k$ near 1. Their shortest axis aligns better with the surface normal, needing less normal residual", + "bbox": [ + 89, + 809, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "correction and allowing larger penalties, as shown in Fig. 3 (c).", + "bbox": [ + 511, + 368, + 903, + 397 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Total training loss. To refine all parameters in the dynamic and specular stages, we employ the total training loss:", + "bbox": [ + 511, + 401, + 905, + 433 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {\\text {c o l o r}} + \\lambda_ {\\text {n o r m a l}} \\mathcal {L} _ {\\text {n o r m a l}} + \\mathcal {L} _ {\\text {r e g}}, \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 593, + 444, + 906, + 460 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_{\\mathrm{color}}$ and $\\mathcal{L}_{\\mathrm{normal}}$ are obtained as described in Section 3.2.1. In our experiments, we set $\\lambda_{\\mathrm{normal}} = 0.01$ . Due to space constraints, complete implementation details are provided in the supplementary materials.", + "bbox": [ + 511, + 470, + 906, + 531 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 544, + 645, + 560 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Evaluation Results", + "text_level": 1, + "bbox": [ + 511, + 568, + 694, + 583 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We evaluate our method on two real-world datasets: NeRF-DS dataset [48] and HyperNeRF dataset [32]. While GaussianShader [12] and GS-IR [23] are originally designed for static scenes and are included here only as reference baselines, we train our method and all baseline approaches for 40,000 iterations to ensure fair comparison.", + "bbox": [ + 511, + 590, + 906, + 681 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "NeRF-DS dataset. The NeRF-DS dataset [48] is a monocular video dataset comprising seven real-world scenes from daily life featuring various types of moving or deforming specular objects. We compare our method with the most relevant state-of-the-art approaches. As shown in Tab. 1 and Fig. 4, the quantitative results demonstrate that our method decisively outperforms baselines in reconstructing and rendering real-world highly reflective dynamic specular scenes.", + "bbox": [ + 511, + 685, + 906, + 805 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The rendering speed is correlated with the quantity of 3D Gaussians. When the number of 3D Gaussians is below 178k, our method can achieve real-time rendering over 30 FPS on an NVIDIA RTX 4090.", + "bbox": [ + 511, + 806, + 906, + 866 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "HyperNeRF dataset. The HyperNeRF dataset contains real-world dynamic scenes and does not include specular objects.", + "bbox": [ + 511, + 869, + 908, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "21332", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/9e04e376026e56e96851e11ab40d15e92496b13e5082ebad777731dbbc9257ef.jpg", + "table_caption": [ + "Table 1. Quantitative comparison on the NeRF-DS [48] dataset. We report the average PSNR, SSIM, and LPIPS (VGG) of several previous models on test images. The best, the second best, and third best results are denoted by red, orange, yellow." + ], + "table_footnote": [], + "table_body": "
MethodAsBasinBellCup
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
Deformable 3DGS [49]26.040.88050.185019.530.78550.192423.960.79450.276724.490.88220.1658
4DGS [46]24.850.86320.203819.260.76700.219622.860.80150.206123.820.86950.1792
GaussianShader [12]21.890.77390.362017.790.66700.418720.690.81690.302420.400.74370.3385
GS-IR [23]21.580.80330.303318.060.72480.313520.660.78290.260320.340.81930.2719
NeRF-DS [48]25.340.88030.215020.230.80530.250822.570.78110.292124.510.88020.1707
HyperNeRF [32]17.590.85180.239022.580.81560.249719.800.76500.299915.450.82950.2302
Ours26.800.88430.176119.750.79150.189625.460.84900.160024.650.88710.1588
PlatePressSieveMean
MethodPSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
Deformable 3DGS [49]19.070.73520.359925.520.85940.196425.370.86160.164323.430.82840.2201
4DGS [46]18.770.77090.272124.820.83550.225525.160.85660.174522.790.82350.2115
GaussianShader [12]14.550.64230.495519.970.72440.450722.580.78620.305719.700.73630.3819
GS-IR [23]15.980.69690.420022.280.80880.306722.840.82120.223620.250.77960.2999
NeRF-DS [48]19.700.78130.297425.350.87030.255224.990.87050.200123.240.83840.2402
HyperNeRF [32]21.220.78290.316616.540.82000.281019.920.85210.214219.010.81670.2615
Ours20.840.81720.219826.490.86570.188925.220.87050.151324.170.85220.1778
", + "bbox": [ + 91, + 122, + 906, + 412 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5c94518554cd7ab1cae7b8cc2db0a1dec111098f4882477ab2630a9bae46c691.jpg", + "image_caption": [ + "Figure 4. Qualitative comparison on the NeRF-DS [48] dataset." + ], + "image_footnote": [], + "bbox": [ + 96, + 422, + 906, + 571 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/0ec95ac145092451ffb141850a17a7431abca813b89dc5ca7cec43a44164c66b.jpg", + "table_caption": [ + "Table 2. Quantitative comparison on the HyperNeRF [32] dataset. We report the average PSNR, SSIM, and LPIPS (VGG) of several previous models. The best, the second best, and third best results are denoted by red, orange, yellow." + ], + "table_footnote": [], + "table_body": "
MethodPSNR ↑SSIM ↑LPIPS ↓
Deformable 3DGS [49]22.780.62010.3380
4DGS [46]24.890.67810.3422
GaussianShader [12]18.550.54520.4795
GS-IR [23]19.870.57290.4498
NeRF-DS [48]23.650.64050.3972
HyperNeRF [32]23.110.63870.3968
Ours22.220.60880.3295
", + "bbox": [ + 102, + 686, + 470, + 823 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As shown in Tab. 2 and Fig. 5, the results demonstrate that we are on par with state-of-the-art techniques for rendering novel views, and our method's performance is not limited to", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "shiny scenes.", + "bbox": [ + 513, + 625, + 604, + 638 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In Fig. 6, we compare our method's normal maps with those from Deformable 3DGS [49] and NeRF-DS [48]. For Deformable 3DGS [49], we obtain the normals by using the shortest axes of the deformed 3D Gaussians. As demonstrated, our method produces significantly better quality normal maps compared to Deformable 3DGS [49] and NeRF-DS [48].", + "bbox": [ + 511, + 638, + 906, + 744 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Ablation Study", + "text_level": 1, + "bbox": [ + 511, + 753, + 666, + 768 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For a fair comparison, we train our method and all ablation experiments for 40,000 iterations.", + "bbox": [ + 511, + 775, + 906, + 805 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Different coarse to fine training strategy stages. As shown in Tab. 3 and Fig. 7, each stage contributes effectively to the model's performance. The Dynamic stage enhances dynamic object stability compared to the Static stage alone, while the Specular stage improves reflection clarity beyond the combined Static and Dynamic stages.", + "bbox": [ + 511, + 810, + 908, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "21333", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a9e9bf5cd13deac75326c85d20ebd03f34c97807ec784fbb025581469de96a27.jpg", + "image_caption": [ + "Figure 5. Qualitative comparison on the HyperNeRF [32] dataset." + ], + "image_footnote": [], + "bbox": [ + 93, + 85, + 906, + 234 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e6e0a71cdf767f7ab03aa7f2bce2dc89b6f74a93fbdf58beb32aae0c2aedc0e3.jpg", + "image_caption": [ + "Figure 6. Qualitative comparison of normal maps between our method, Deformable 3DGS, and NeRF-DS." + ], + "image_footnote": [], + "bbox": [ + 91, + 282, + 480, + 378 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/e9c99d99eefa05bef6c03075610fb50d719758f1ad8aa2e4cf8c514404c4a8ca.jpg", + "table_caption": [ + "Table 3. Ablation studies on different coarse to fine training strategy stages." + ], + "table_footnote": [], + "table_body": "
StagePSNR ↑SSIM ↑LPIPS ↓
Static20.260.77850.2953
St. + Dynamic24.020.85080.1831
St. + Dy. + Specular24.170.85220.1778
", + "bbox": [ + 114, + 467, + 459, + 540 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ef87ff0f4d145e3cb09e040c7177b4a7489c7e58d8f38394243b6eac9f4e6bf5.jpg", + "image_caption": [ + "Figure 7. Qualitative comparison of each training stage in our coarse-to-fine approach." + ], + "image_footnote": [], + "bbox": [ + 94, + 555, + 482, + 756 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Ablation study on coarse-to-fine, and loss function. The model's performance was evaluated without key components: the coarse-to-fine training strategy, normal loss $\\mathcal{L}_{\\mathrm{normal}}$ , normal regularization $\\mathcal{L}_{\\mathrm{reg}}$ , and $\\gamma^k$ . Fig. 8 and Tab. 4 illustrate the effects of these omissions. Without the coarse-to-fine", + "bbox": [ + 89, + 825, + 483, + 898 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/3bcf46684a25df798a6b89f76d28381ea0646a702dee521d88af429c0a0711af.jpg", + "table_caption": [ + "Table 4. Ablation studies on different coarse to fine training strategy stages." + ], + "table_footnote": [], + "table_body": "
C2F\\( \\mathcal{L}_{\\text{normal}} \\)\\( \\mathcal{L}_{\\text{reg}} \\)\\( \\gamma^k \\)PSNR↑SSIM↑LPIPS↓
✓✓✓23.160.82940.2156
✓23.400.82770.2278
✓✓24.150.85100.1845
✓✓✓24.090.85150.1818
✓✓✓✓24.170.85220.1778
", + "bbox": [ + 521, + 316, + 898, + 416 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/2455c6f27b02f2e46f403ec99b16c89988cfe0d040adff6baf9a8aaf072e87f6.jpg", + "table_caption": [ + "Table 5. Ablation studies on SH, Static and Deformable environment map." + ], + "table_footnote": [], + "table_body": "
PSNR ↑SSIM ↑LPIPS ↓
SH23.630.84530.1844
Static Env. map24.020.85080.1831
Deformable Env. map24.170.85220.1778
", + "bbox": [ + 532, + 463, + 887, + 536 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "approach, the model, trained directly at the specular stage, produces incomplete scene geometry, affecting environment map queries for specular color. Omitting normal loss $\\mathcal{L}_{\\mathrm{normal}}$ removes direct supervision on normals and leads to inaccurate reflection directions and less precise specular colors. Removing normal regularization $\\mathcal{L}_{\\mathrm{reg}}$ allows the normal residual to dominate normal optimization, resulting in insufficient optimization of the 3D Gaussians' shortest axis towards the correct normal, which in turn reduces the rendering quality. The normal residual decreases for non-flattened and flat Gaussians without $\\gamma^k$ in normal regularization. This particularly affects less flat 3D Gaussians whose shortest axis significantly deviates from the surface normal. The insufficient normal residual correction causes these 3D Gaussians' shortest axes to deviate greatly from their original direction in an attempt to align with the surface normal, ultimately reducing rendering quality.", + "bbox": [ + 511, + 563, + 908, + 820 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Ablation study on SH, Static environment map, and Deformable environment map. Fig. 9 and Tab. 5 demonstrate the superiority of the deformable environment map over the static environment map, which in turn outperforms Spherical Harmonics (SH). SH struggles to accurately model high-", + "bbox": [ + 511, + 825, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "21334", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/cffa3e1f530ca2a1e31da73d46e0acb77bd2e7b60976bd39a99347eeabe638b9.jpg", + "image_caption": [ + "Full model" + ], + "image_footnote": [], + "bbox": [ + 91, + 87, + 292, + 175 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/95c082d7f288a73c6d4f1f2c91f5181be473758f935c5d81a79fa8cda99a60c1.jpg", + "image_caption": [ + "Full model" + ], + "image_footnote": [], + "bbox": [ + 297, + 88, + 496, + 175 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5da24300c8e3a74f08660cd5651e4bd6c9ee6b60318adcc5dc42f60eb366ba51.jpg", + "image_caption": [ + "Full model" + ], + "image_footnote": [], + "bbox": [ + 501, + 88, + 700, + 175 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2eca99811ece7f867e60b06ed6fe26e77cefe9dd0b6134a391bdce9bedaaaeb6.jpg", + "image_caption": [ + "Full model" + ], + "image_footnote": [], + "bbox": [ + 705, + 88, + 903, + 175 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8ff991c9f481437702c238f47b10cb7a4b6a3056e370040673254d41ce57184f.jpg", + "image_caption": [ + "w/o Coarse-to-fine" + ], + "image_footnote": [], + "bbox": [ + 91, + 191, + 292, + 277 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2b453a3c19c78fc582a06ef7f0a0929f6db17565e8f91274536a8a6851fa0864.jpg", + "image_caption": [ + "w/o $\\mathcal{L}_{\\mathrm{normal}}$" + ], + "image_footnote": [], + "bbox": [ + 297, + 191, + 496, + 277 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/eec227e92ea0cabeefb69d6c189faa5ec1094d4d7bf84f8574636a5c2e535ec8.jpg", + "image_caption": [ + "w/o $\\mathcal{L}_{\\mathrm{reg}}$" + ], + "image_footnote": [], + "bbox": [ + 501, + 191, + 700, + 277 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1c8e2e5a51f8e8c9bb1864b1608aa6f82369ce52f10be35973f72ef3cf64c6ea.jpg", + "image_caption": [ + "w/o $\\gamma^k$" + ], + "image_footnote": [], + "bbox": [ + 705, + 191, + 903, + 277 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d4f16fca7dd438c0cf659a75042f2c5d9301b0323856ac744642bf07a4faddfb.jpg", + "image_caption": [ + "Figure 8. Qualitative comparison of ablation study without different components." + ], + "image_footnote": [], + "bbox": [ + 94, + 345, + 287, + 431 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/be59e70aff98002103d86392026d30d7d08df8392e828199988f14bac915566c.jpg", + "image_caption": [ + "Static Env. map" + ], + "image_footnote": [], + "bbox": [ + 290, + 347, + 480, + 431 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e5af51a886f5b77d2ec9e5c3e3fd2e2e64e46d29bb37ec71043c89146b15949e.jpg", + "image_caption": [ + "Spherical Harmonics (SH)", + "Deformable Env. map (Full model)", + "Figure 9. Qualitative comparison of ablation study on SH, Static environment map, and Deformable environment map." + ], + "image_footnote": [], + "bbox": [ + 94, + 446, + 287, + 530 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/109096eadb039a1b5aa84b90b20f94196bd493e4927be2ad8303ac10b9b55b1d.jpg", + "image_caption": [ + "Ground-truth" + ], + "image_footnote": [], + "bbox": [ + 290, + 446, + 480, + 530 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/5a2eaca7063cb7f8003f83e613e68d9445c8042d5b42edde6b1995f197bb45af.jpg", + "table_caption": [ + "Table 6. Ablation studies on 2DGS and without Physical Normal Estimation." + ], + "table_footnote": [], + "table_body": "
PSNR ↑SSIM ↑LPIPS ↓
2DGS [10]23.220.82190.2283
w/o N.E.23.890.84900.1837
Full model24.170.85220.1778
", + "bbox": [ + 142, + 638, + 431, + 712 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "frequency specular colors. While the static environment map can model high-frequency colors, it is best suited for static lighting conditions. In contrast, the deformable environment map models time-varying lighting, offering superior performance for dynamic scenes.", + "bbox": [ + 89, + 743, + 483, + 818 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ablation study on 2DGS [10] and without Physical Normal Estimation. In Fig. 10 and Tab. 6, \"2DGS\" represents replacing our 3D Gaussian with 2D Gaussian representation. Since 2DGS inherently includes normals, we omit physical normal estimation. \"w/o N.E.\" means skipping physical nor", + "bbox": [ + 89, + 824, + 485, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f33d20c3f784bfbe7f0ff3bb9781853c0752766e462530489b8bd01d0fee52ac.jpg", + "image_caption": [ + "Full model" + ], + "image_footnote": [], + "bbox": [ + 516, + 345, + 707, + 430 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ccb980a7b0c4ceb91f053141c76fe75402e5d4140022965b63d68daeb8f5ab6d.jpg", + "image_caption": [ + "Full model" + ], + "image_footnote": [], + "bbox": [ + 710, + 345, + 903, + 430 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/cedd9e62c8c0dcc98d5badf8505cda37ed6e25f28bb5f30bef17e0e11f2e4ae2.jpg", + "image_caption": [ + "2DGS", + "Figure 10. Qualitative comparison of ablation study on 2DGS and without Physical Normal Estimation." + ], + "image_footnote": [], + "bbox": [ + 516, + 446, + 707, + 530 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2068735c5657228fe4714a273f581b1856a11fbe8005e1a250b0ef63e614d843.jpg", + "image_caption": [ + "w/o N.E." + ], + "image_footnote": [], + "bbox": [ + 710, + 446, + 903, + 530 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "mal estimation and using the shortest axis of 3D Gaussians as the normal. This causes the normal loss $\\mathcal{L}_{\\mathrm{normal}}$ to directly supervise the shortest axes, making some axes deviate significantly to align with surface normals, resulting in degraded rendering quality.", + "bbox": [ + 511, + 614, + 906, + 689 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 704, + 632, + 719 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "SpectroMotion enhances 3D Gaussian Splatting for dynamic specular scenes by combining specular rendering with deformation fields. Using normal residual correction, coarse-to-fine training, and a deformable environment map, it achieves superior accuracy and visual quality in novel view synthesis, outperforming existing methods while maintaining geometric consistency.", + "bbox": [ + 511, + 729, + 906, + 835 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations. Though we stabilize the entire scene's geometry using a coarse-to-fine training strategy, some failure cases still occur. Please refer to the supplementary materials for visual results of failure cases.", + "bbox": [ + 511, + 839, + 908, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "21335", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements. This research was funded by the National Science and Technology Council, Taiwan, under Grants NSTC 112-2222-E-A49-004-MY2 and 113-2628-E-A49-023-. The authors are grateful to Google, NVIDIA, and MediaTek Inc. for their generous donations. Yu-Lun Liu acknowledges the Yushan Young Fellow Program by the MOE in Taiwan.", + "bbox": [ + 89, + 90, + 485, + 196 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 210, + 187, + 226 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Benjamin Attal, Eliot Laidlaw, Aaron Gokaslan, Changil Kim, Christian Richardt, James Tompkin, and Matthew O'Toole. Törf: Time-of-flight radiance fields for dynamic scene view synthesis, 2021. 2", + "[2] Sai Bi, Zexiang Xu, Pratul Srinivasan, Ben Mildenhall, Kalyan Sunkavalli, Miloš Hasan, Yannick Hold-Geoffroy, David Kriegman, and Ravi Ramamoorthi. Neural reflectance fields for appearance acquisition, 2020. 2", + "[3] Mark Boss, Varun Jampani, Raphael Braun, Ce Liu, Jonathan T. Barron, and Hendrik P. A. Lensch. Neural-pil: Neural pre-integrated lighting for reflectance decomposition, 2021. 2", + "[4] Ting-Hsuan Chen, Jie Wen Chan, Hau-Shiang Shiu, Shih-Han Yen, Changhan Yeh, and Yu-Lun Liu. Narcan: Natural refined canonical image with integration of diffusion prior for video editing. In NeurIPS, 2024. 2", + "[5] Robert L Cook and Kenneth E. Torrance. A reflectance model for computer graphics. ACM Transactions on Graphics (ToG), 1(1):7-24, 1982. 3", + "[6] Ainaz Eftekhar, Alexander Sax, Jitendra Malik, and Amir Zamir. Omnidata: A scalable pipeline for making multi-task mid-level vision datasets from 3d scans. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10786-10796, 2021. 1", + "[7] Jian Gao, Chun Gu, Youtian Lin, Hao Zhu, Xun Cao, Li Zhang, and Yao Yao. Relightable 3d gaussian: Real-time point cloud relighting with brdf decomposition and ray tracing. arXiv preprint arXiv:2311.16043, 2023. 2", + "[8] Wenhang Ge, Tao Hu, Haoyu Zhao, Shu Liu, and Ying-Cong Chen. Ref-neus: Ambiguity-reduced neural implicit surface learning for multi-view reconstruction with reflection, 2023. 2", + "[9] Xiang Guo, Jiadai Sun, Yuchao Dai, Guanying Chen, Xiaoqing Ye, Xiao Tan, Errui Ding, Yumeng Zhang, and Jingdong Wang. Forward flow for novel view synthesis of dynamic scenes, 2023. 2", + "[10] Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2d gaussian splatting for geometrically accurate radiance fields. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 8", + "[11] Yi-Hua Huang, Yang-Tian Sun, Ziyi Yang, Xiaoyang Lyu, Yan-Pei Cao, and Xiaojuan Qi. Sc-gs: Sparse-controlled gaussian splatting for editable dynamic scenes, 2024. 2", + "[12] Yingwenqi Jiang, Jiadong Tu, Yuan Liu, Xifeng Gao, Xiaoxiao Long, Wenping Wang, and Yuexin Ma. Gaussianshader: 3d gaussian splatting with shading functions for reflective surfaces. arXiv preprint arXiv:2311.17977, 2023. 2, 4, 5, 6" + ], + "bbox": [ + 93, + 234, + 485, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Yingwenqi Jiang, Jiadong Tu, Yuan Liu, Xifeng Gao, Xiaoxiao Long, Wenping Wang, and Yuexin Ma. Gaussianshader: 3d gaussian splatting with shading functions for reflective surfaces. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5322-5332, 2024. 3", + "[14] James T Kajiya. The rendering equation. In Proceedings of the 13th annual conference on Computer graphics and interactive techniques, pages 143-150, 1986. 3", + "[15] Brian Karis and Epic Games. Real shading in unreal engine 4. Proc. Physically Based Shading Theory Practice, 4(3):1, 2013. 3", + "[16] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering, 2023. 1, 2", + "[17] Junxuan Li and Hongdong Li. Neural reflectance for shape recovery with shadow handling, 2022. 2", + "[18] Tianye Li, Mira Slavcheva, Michael Zollhoefer, Simon Green, Christoph Lassner, Changil Kim, Tanner Schmidt, Steven Lovegrove, Michael Goesele, Richard Newcombe, and Zhaoyang Lv. Neural 3d video synthesis from multi-view video, 2022. 2", + "[19] Zhengqi Li, Simon Niklaus, Noah Snively, and Oliver Wang. Neural scene flow fields for space-time view synthesis of dynamic scenes, 2021. 2", + "[20] Ruofan Liang, Huiting Chen, Chunlin Li, Fan Chen, Selvakumar Panneer, and Nandita Vijaykumar. Envidr: Implicit differentiable renderer with neural environment lighting, 2023. 2", + "[21] Ruofan Liang, Jiahao Zhang, Haoda Li, Chen Yang, Yushi Guan, and Nandita Vijaykumar. Spidr: Sdf-based neural point fields for illumination and deformation, 2023. 2", + "[22] Yiqing Liang, Numair Khan, Zhengqin Li, Thu Nguyen-Phuoc, Douglas Lanman, James Tompkin, and Lei Xiao. Gaufre: Gaussian deformation fields for real-time dynamic novel view synthesis, 2023. 2", + "[23] Zhihao Liang, Qi Zhang, Ying Feng, Ying Shan, and Kui Jia. Gs-ir: 3d gaussian splatting for inverse rendering. arXiv preprint arXiv:2311.16473, 2023. 2, 5, 6", + "[24] Haotong Lin, Sida Peng, Zhen Xu, Yunzhi Yan, Qing Shuai, Hujun Bao, and Xiaowei Zhou. Efficient neural radiance fields for interactive free-viewpoint video, 2022. 2", + "[25] Yuan Liu, Peng Wang, Cheng Lin, Xiaoxiao Long, Jiepeng Wang, Lingjie Liu, Taku Komura, and Wenping Wang. Nano: Neural geometry and brdf reconstruction of reflective objects from multiview images, 2023. 2", + "[26] Yu-Lun Liu, Chen Gao, Andreas Meuleman, Hung-Yu Tseng, Ayush Saraf, Changil Kim, Yung-Yu Chuang, Johannes Kopf, and Jia-Bin Huang. Robust dynamic radiance fields. In CVPR, 2023. 2", + "[27] Caoyuan Ma, Yu-Lun Liu, Zhixiang Wang, Wu Liu, Xinchen Liu, and Zheng Wang. Humannerf-se: A simple yet effective approach to animate humannerf with diverse poses. In CVPR, 2024. 2", + "[28] Li Ma, Vasu Agrawal, Haithem Turki, Changil Kim, Chen Gao, Pedro Sander, Michael Zollhöfer, and Christian Richardt. Specnerf: Gaussian directional encoding for specular reflections. arXiv preprint arXiv:2312.13102, 2023. 2" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "21336", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[29] Marko Mihajlovic, Sergey Prokudin, Siyu Tang, Robert Maier, Federica Bogo, Tony Tung, and Edmond Boyer. Splatfields: Neural gaussian splats for sparse 3d and 4d reconstruction. arXiv preprint arXiv:2409.11211, 2024. 2", + "[30] Jacob Munkberg, Jon Hasselgren, Tianchang Shen, Jun Gao, Wenzheng Chen, Alex Evans, Thomas Müller, and Sanja Fidler. Extracting Triangular 3D Models, Materials, and Lighting From Images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8280-8290, 2022. 2, 3", + "[31] Keunhong Park, Utkarsh Sinha, Jonathan T. Barron, Sofien Bouaziz, Dan B Goldman, Steven M. Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields, 2021. 2", + "[32] Keunhong Park, Utkarsh Sinha, Peter Hedman, Jonathan T. Barron, Sofien Bouaziz, Dan B Goldman, Ricardo MartinBrualla, and Steven M. Seitz. Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields, 2021. 2, 5, 6, 7", + "[33] Sida Peng, Yunzhi Yan, Qing Shuai, Hujun Bao, and Xiaowei Zhou. Representing volumetric videos as dynamic mlp maps, 2023. 2", + "[34] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-nerf: Neural radiance fields for dynamic scenes, 2020. 2", + "[35] Yahao Shi, Yanmin Wu, Chenming Wu, Xing Liu, Chen Zhao, Haocheng Feng, Jingtuo Liu, Liangjun Zhang, Jian Zhang, Bin Zhou, et al. Gir: 3d gaussian inverse rendering for relightable scene factorization. arXiv preprint arXiv:2312.05133, 2023. 2", + "[36] Liangchen Song, Anpei Chen, Zhong Li, Zhang Chen, Lele Chen, Junsong Yuan, Yi Xu, and Andreas Geiger. Nerfplayer: A streamable dynamic scene representation with decomposed neural radiance fields, 2023. 2", + "[37] Pratul P. Srinivasan, Boyang Deng, Xiuming Zhang, Matthew Tancik, Ben Mildenhall, and Jonathan T. Barron. Nerv: Neural reflectance and visibility fields for relighting and view synthesis, 2020. 2", + "[38] Colton Stearns, Adam Harley, Mikaela Uy, Florian Dubost, Federico Tombari, Gordon Wetzstein, and Leonidas Guibas. Dynamic gaussian marbles for novel view synthesis of casual monocular videos. arXiv preprint arXiv:2406.18717, 2024. 2", + "[39] Edgar Tretschk, Ayush Tewari, Vladislav Golyanik, Michael Zollhöfer, Christoph Lassner, and Christian Theobalt. Non-rigid neural radiance fields: Reconstruction and novel view synthesis of a dynamic scene from monocular video, 2021. 2", + "[40] Dor Verbin, Peter Hedman, Ben Mildenhall, Todd Zickler, Jonathan T Barron, and Pratul P Srinivasan. Ref-nerf: Structured view-dependent appearance for neural radiance fields. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5481–5490. IEEE, 2022. 2", + "[41] Dor Verbin, Ben Mildenhall, Peter Hedman, Jonathan T Barron, Todd Zickler, and Pratul P Srinivasan. Eclipse: Disambiguating illumination and materials using unintended shadows. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 77-86, 2024. 2" + ], + "bbox": [ + 91, + 90, + 485, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[42] Dor Verbin, Pratul P Srinivasan, Peter Hedman, Ben Mildenhall, Benjamin Attal, Richard Szeliski, and Jonathan T Barron. Nerf-casting: Improved view-dependent appearance with consistent reflections. arXiv preprint arXiv:2405.14871, 2024. 2", + "[43] Bruce Walter, Stephen R Marschner, Hongsong Li, and Kenneth E Torrance. Microfacet models for refraction through rough surfaces. Rendering techniques, 2007:18th, 2007. 3", + "[44] Qianqian Wang, Vickie Ye, Hang Gao, Jake Austin, Zhengqi Li, and Angjoo Kanazawa. Shape of motion: 4d reconstruction from a single video. arXiv preprint arXiv:2407.13764, 2024. 2", + "[45] Chun-Hung Wu, Shih-Hong Chen, Chih-Yao Hu, Hsin-Yu Wu, Kai-Hsin Chen, Yu-You Chen, Chih-Hai Su, Chih-Kuo Lee, and Yu-Lun Liu. Denver: Deformable neural vessel representations for unsupervised video vessel segmentation. In CVPR, 2025. 2", + "[46] Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Xinggang Wang. 4d gaussian splatting for real-time dynamic scene rendering. arXiv preprint arXiv:2310.08528, 2023. 2, 6", + "[47] Wenqi Xian, Jia-Bin Huang, Johannes Kopf, and Changil Kim. Space-time neural irradiance fields for free-viewpoint video, 2021. 2", + "[48] Zhiwen Yan, Chen Li, and Gim Hee Lee. Nerf-ds: Neural radiance fields for dynamic specular objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8285-8295, 2023. 5, 6", + "[49] Ziyi Yang, Xinyu Gao, Wen Zhou, Shaohui Jiao, Yuqing Zhang, and Xiaogang Jin. Deformable 3d gaussians for high-fidelity monocular dynamic scene reconstruction. arXiv preprint arXiv:2309.13101, 2023. 2, 3, 6", + "[50] Ziyi Yang, Xinyu Gao, Yangtian Sun, Yihua Huang, Xiaoyang Lyu, Wen Zhou, Shaohui Jiao, Xiaojuan Qi, and Xiaogang Jin. Spec-gaussian: Anisotropic view-dependent appearance for 3d gaussian splatting, 2024. 2", + "[51] Keyang Ye, Qiming Hou, and Kun Zhou. 3d gaussian splatting with deferred reflection. In ACM SIGGRAPH 2024 Conference Papers, pages 1-10, 2024. 2", + "[52] Wang Yifan, Felice Serena, Shihao Wu, Cengiz Oztireli, and Olga Sorkine-Hornung. Differentiable surface splatting for point-based geometry processing. ACM Transactions on Graphics (TOG), 38(6):1-14, 2019. 2", + "[53] Jingyang Zhang, Yao Yao, Shiwei Li, Jingbo Liu, Tian Fang, David McKinnon, Yanghai Tsin, and Long Quan. Neilf++: Inter-reflectable light fields for geometry and material estimation, 2023. 2", + "[54] Kai Zhang, Fujun Luan, Qianqian Wang, Kavita Bala, and Noah Snavely. Physg: Inverse rendering with spherical gaussians for physics-based material editing and relighting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5453-5462, 2021. 2", + "[55] Xiuming Zhang, Pratul P. Srinivasan, Boyang Deng, Paul Debevec, William T. Freeman, and Jonathan T. Barron. Ner-factor: neural factorization of shape and reflectance under an unknown illumination. ACM Transactions on Graphics, 40 (6):1-18, 2021." + ], + "bbox": [ + 514, + 90, + 906, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "21337", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[56] Xiaoming Zhao, Pratul P Srinivasan, Dor Verbin, Keunhong Park, Ricardo Martin Brualla, and Philipp Henzler. Illuminerf: 3d relighting without inverse rendering. arXiv preprint arXiv:2406.06527, 2024. 2", + "[57] Zuo-Liang Zhu, Beibei Wang, and Jian Yang. Gs-ror: 3d gaussian splatting for reflective object relighting via sdf priors. arXiv preprint arXiv:2406.18544, 2024. 2" + ], + "bbox": [ + 91, + 90, + 486, + 191 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "21338", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/SpectroMotion_ Dynamic 3D Reconstruction of Specular Scenes/5055bce6-6fec-402b-adcc-413e1a1becdd_model.json b/2025/SpectroMotion_ Dynamic 3D Reconstruction of Specular Scenes/5055bce6-6fec-402b-adcc-413e1a1becdd_model.json new file mode 100644 index 0000000000000000000000000000000000000000..28a16d73d948d7b9f180e0a740a8d000268ab7ef --- /dev/null +++ b/2025/SpectroMotion_ Dynamic 3D Reconstruction of Specular Scenes/5055bce6-6fec-402b-adcc-413e1a1becdd_model.json @@ -0,0 +1,2642 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.809, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.177, + 0.131, + 0.82, + 0.153 + ], + "angle": 0, + "content": "SpectroMotion: Dynamic 3D Reconstruction of Specular Scenes" + }, + { + "type": "text", + "bbox": [ + 0.229, + 0.18, + 0.792, + 0.219 + ], + "angle": 0, + "content": "Cheng-De Fan1 Chen-Wei Chang1 Yi-Ruei Liu1,2 Jie-Ying Lee1 \nJiun-Long Huang1 Yu-Chee Tseng1 Yu-Lun Liu1" + }, + { + "type": "text", + "bbox": [ + 0.154, + 0.227, + 0.862, + 0.246 + ], + "angle": 0, + "content": "\\(^{1}\\)National Yang Ming Chiao Tung University \\(^{2}\\)University of Illinois Urbana-Champaign" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.28, + 0.327, + 0.296 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.314, + 0.486, + 0.571 + ], + "angle": 0, + "content": "We present SpectroMotion, a novel approach that combines 3D Gaussian Splatting (3DGS) with physically-based rendering (PBR) and deformation fields to reconstruct dynamic specular scenes. Previous methods extending 3DGS to model dynamic scenes have struggled to represent specular surfaces accurately. Our method addresses this limitation by introducing a residual correction technique for accurate surface normal computation during deformation, complemented by a deformable environment map that adapts to time-varying lighting conditions. We implement a coarse-to-fine training strategy significantly enhancing scene geometry and specular color prediction. It is the only existing 3DGS method capable of synthesizing photorealistic real-world dynamic specular scenes, outperforming state-of-the-art methods in rendering complex, dynamic, and specular scenes. Please see our project page at cdfan0627.github.io/spectromotion." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.602, + 0.221, + 0.617 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.627, + 0.484, + 0.763 + ], + "angle": 0, + "content": "3D Gaussian Splitting (3DGS) [16] has emerged as a groundbreaking technique in 3D scene reconstruction, offering fast training and real-time rendering capabilities. By representing 3D scene using a collection of 3D Gaussians and employing a point-based rendering approach, 3DGS has significantly improved efficiency in novel view synthesis. However, extending 3DGS to model dynamic scenes, especially those containing specular surfaces accurately, has remained a significant challenge." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.484, + 0.885 + ], + "angle": 0, + "content": "Existing extensions of 3DGS have made progress in either dynamic scene reconstruction or specular object rendering, but none have successfully combined both aspects. Methods tackling dynamic scenes often struggle with accurately representing specular surfaces, while those focusing on specular rendering are limited to static scenes. This capability gap has hindered the application of 3DGS to real-world scenarios where both motion and specular reflections are present." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.886, + 0.484, + 0.901 + ], + "angle": 0, + "content": "We present SpectroMotion, a novel approach that ad" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.281, + 0.905, + 0.505 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.518, + 0.908, + 0.671 + ], + "angle": 0, + "content": "Figure 1. Our method, SpectroMotion, recovers and renders dynamic scenes with higher-quality reflections compared to prior work. It introduces physical normal estimation, deformable environment maps, and a coarse-to-fine training strategy to achieve superior results in rendering dynamic scenes with reflections. Here, we present a rendered test image, corresponding normal maps, and a ground-truth image, where the ground-truth normal map (used as a reference) is generated using a pre-trained normal estimator [6]. For Deformable 3DGS, we use the shortest axes of the deformed 3D Gaussians as the normals. We have highlighted the specular regions to demonstrate the effectiveness of our approach." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.7, + 0.907, + 0.821 + ], + "angle": 0, + "content": "dresses these limitations by combining 3D Gaussian Splating with physically based rendering (PBR) and deformation fields. Our method introduces three key innovations: a residual correction technique for accurate surface normal computation during deformation, a deformable environment map that adapts to time-varying lighting conditions, and a coarse-to-fine training strategy that significantly enhances scene geometry and specular color prediction." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Our evaluations demonstrate that SpectroMotion outperforms prior methods in view synthesis of scenes containing dynamic specular objects, as illustrated in Fig. 1. It is the only 3DGS method capable of synthesizing photorealistic real-world dynamic specular scenes, surpassing state-of-the" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21328" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.152 + ], + "angle": 0, + "content": "art techniques in rendering complex, dynamic, and specular content. This advancement represents a significant leap in 3D scene reconstruction, particularly for challenging scenarios involving moving specular objects." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.153, + 0.445, + 0.168 + ], + "angle": 0, + "content": "In summary, we make the following contributions:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.169, + 0.483, + 0.228 + ], + "angle": 0, + "content": "- We propose SpectroMotion, a physically-based rendering (PBR) approach combining deformation fields and 3D Gaussian Splatting for real-world dynamic specular scenes." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.229, + 0.484, + 0.289 + ], + "angle": 0, + "content": "- We introduce a residual correction method for accurate surface normals during deformation, coupled with a deformable environment map to handle time-varying lighting conditions in dynamic scenes." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.29, + 0.484, + 0.335 + ], + "angle": 0, + "content": "- We develop a coarse-to-fine training strategy enhancing scene geometry and specular color prediction, outperforming state-of-the-art methods." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.169, + 0.484, + 0.335 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.35, + 0.232, + 0.366 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.376, + 0.366, + 0.391 + ], + "angle": 0, + "content": "2.1. Dynamic Scene Reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.398, + 0.485, + 0.701 + ], + "angle": 0, + "content": "Recent works have leveraged NeRF representations to jointly solve for canonical space and deformation fields in dynamic scenes using RGB supervision [4, 9, 19, 26, 27, 31, 32, 34, 39, 45, 47]. Further advancements in dynamic neural rendering include object segmentation [36], incorporation of depth information [1], utilization of 2D CNNs for scene priors [24, 33], and multi-view video compression [18]. However, these NeRF-based methods are computationally intensive, limiting their practical applications. To address this, 3D Gaussian Splatting [16] has emerged as a promising alternative, offering real-time rendering capabilities while maintaining high visual quality. Building upon this efficient representation, recent research has adapted 3D Gaussians for dynamic scenes [11, 22, 29, 38, 44, 46, 49]. Nevertheless, these approaches do not explicitly account for changes in surface normal during the dynamic process. Our work extends this line of research by combining specular object rendering based on normal estimation with a deformation field, enabling each 3D Gaussian to model dynamic specular scenes effectively." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.712, + 0.345, + 0.729 + ], + "angle": 0, + "content": "2.2. Reflective Object Rendering" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.486, + 0.903 + ], + "angle": 0, + "content": "While significant progress has been made in rendering reflective objects, challenges from complex light interactions persist. Recent years have seen numerous studies addressing these issues, primarily by decomposing appearance into lighting and material properties [2, 3, 17, 30, 37, 41, 54-56]. Building on this foundation, some research has focused on improving the capture and reproduction of specular reflections [28, 40, 42, 51]. In contrast, others have leveraged signed distance functions (SDFs) to enhance normal estimation [8, 20, 21, 25, 53]. The emergence of 3D Gaussian Splatting (3DGS) has sparked a new wave of techniques" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.909, + 0.213 + ], + "angle": 0, + "content": "[7, 12, 23, 35, 50, 57] that integrate Gaussian splating with physically-based rendering. Nevertheless, accurately modeling dynamic environments and time-varying specular reflections remains a significant challenge. To address this limitation, our work introduces a novel approach incorporating a deformable environment map and additional explicit Gaussian attributes specifically designed to capture specular color changes over time." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.227, + 0.605, + 0.243 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.252, + 0.909, + 0.389 + ], + "angle": 0, + "content": "Overview of the approach. The overview of our method is illustrated in Fig. 2. Given an input monocular video sequence of frames and corresponding camera poses, we design a three-stage approach to reconstruct the dynamic specular scene, as detailed in Sec. 3.2. Accurate specular reflection requires precise normal estimation, so Sec. 3.3 elaborates on how we estimate normals in dynamic scenes. Finally, we introduce the losses used throughout the training process in Sec. 3.4." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.399, + 0.643, + 0.415 + ], + "angle": 0, + "content": "3.1. Preliminary" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.424, + 0.907, + 0.516 + ], + "angle": 0, + "content": "3D Gaussian Splatting. Each 3D Gaussian is defined by a center position \\( \\boldsymbol{x} \\in \\mathbb{R}^3 \\) and a covariance matrix \\( \\Sigma \\). 3D Gaussian Splatting [16] optimizes the covariance matrix using scaling factors \\( s \\in \\mathbb{R}^3 \\) and rotation unit quaternion \\( r \\in \\mathbb{R}^4 \\). For novel-view rendering, 3D Gaussians are projected onto 2D camera planes using differentiable splatting [52]:" + }, + { + "type": "equation", + "bbox": [ + 0.643, + 0.526, + 0.907, + 0.544 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\Sigma} ^ {\\prime} = \\mathbf {J} \\mathbf {W} \\boldsymbol {\\Sigma} \\mathbf {W} ^ {T} \\mathbf {J} ^ {T}. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.556, + 0.909, + 0.587 + ], + "angle": 0, + "content": "Pixel colors are computed using point-based volumetric rendering:" + }, + { + "type": "equation", + "bbox": [ + 0.568, + 0.597, + 0.907, + 0.631 + ], + "angle": 0, + "content": "\\[\nC = \\sum_ {i \\in N} T _ {i} \\alpha_ {i} c _ {i}, \\quad \\alpha_ {i} = \\sigma_ {i} e ^ {- \\frac {1}{2} (\\boldsymbol {x}) ^ {T} \\boldsymbol {\\Sigma} ^ {\\prime} (\\boldsymbol {x})}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.644, + 0.908, + 0.676 + ], + "angle": 0, + "content": "where \\( T_{i} = \\prod_{j=1}^{i-1}(1 - \\alpha_{j}) \\) is the transmittance, \\( \\sigma_{i} \\) is the opacity, and \\( \\mathbf{c}_{i} \\) is the color of each 3D Gaussian." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.685, + 0.706, + 0.702 + ], + "angle": 0, + "content": "3.2. Specular Rendering" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.708, + 0.909, + 0.785 + ], + "angle": 0, + "content": "Since accurate reflections depend heavily on precise geometry, we implement a three-stage coarse-to-fine training strategy: static, dynamic, and specular stages. This approach ensures both stable scene geometry and accurate specular rendering." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.792, + 0.792, + 0.807 + ], + "angle": 0, + "content": "3.2.1. Coarse-to-Fine Training Strategy" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Static stage. In the static stage, we employ vanilla 3DGS [16] for static scene reconstruction to stabilize the geometry of the static scene. Specifically, we optimize the position \\( \\mathbf{x} \\), scaling \\( s \\), rotation \\( r \\), opacity \\( \\alpha \\), and coefficients of spherical harmonics (SH) of the 3D Gaussians by minimizing the photometric loss \\( \\mathcal{L}_{\\mathrm{color}} \\) identical to 3DGS [16]." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "21329" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.089, + 0.337, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.344, + 0.089, + 0.906, + 0.267 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.271, + 0.906, + 0.455 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.466, + 0.907, + 0.578 + ], + "angle": 0, + "content": "Figure 2. Method Overview. Our method stabilizes the scene geometry through three stages. In the static stage, we stabilize the geometry of the static scene by minimizing photometric loss \\(\\mathcal{L}_{\\mathrm{color}}\\) between vanilla 3DGS renders and ground truth images. The dynamic stage combines canonical 3D Gaussians \\(\\mathbf{G}\\) with a deformable Gaussian MLP to model dynamic scenes while simultaneously minimizing normal loss \\(\\mathcal{L}_{\\mathrm{normal}}\\) between rendered normal map \\(\\mathbf{N}^t\\) and gradient normal map from depth map \\(\\mathbf{D}^t\\), thus further enhancing the overall scene geometry. Finally, the specular stage introduces a deformable reflection MLP to handle changing environment lighting, deforming reflection directions \\(\\omega_r^t\\) to query a canonical environment map for specular color \\(\\mathbf{c}_s^t\\). It is then combined with diffuse color \\(\\mathbf{c_d}\\) (using zero-order spherical harmonics) and learnable specular tint \\(\\mathbf{s}_{\\mathrm{int}}\\) per 3D Gaussian to obtain the final color \\(\\mathbf{c}_{\\mathrm{final}}^t\\). This approach enables the modeling of dynamic specular scenes and high-quality novel view rendering." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.604, + 0.483, + 0.739 + ], + "angle": 0, + "content": "Dynamic stage. Following the static stage, we address dynamic objects using Deformable 3DGS [49]. For each 3D Gaussian in canonical 3D Gaussians \\(\\mathbf{G}\\), we input its position \\(\\pmb{x}\\) and time \\(t\\) into a deformable Gaussian MLP with parameters \\(\\theta_{G}\\) to predict position, rotation, and scaling residuals: \\((\\Delta \\pmb{x}^{t}, \\Delta \\pmb{r}^{t}, \\Delta \\pmb{s}^{t}) = F_{\\theta_{G}}(\\gamma(x), \\gamma(t))\\), where \\(\\gamma\\) denotes positional encoding. Attributes of the corresponding 3D Gaussian in deformed 3D Gaussians \\(\\mathbf{G}^{t}\\) at time \\(t\\) is obtained by \\((\\pmb{x}^{t}, \\pmb{r}^{t}, \\pmb{s}^{t}) = (\\Delta \\pmb{x}^{t}, \\Delta \\pmb{r}^{t}, \\Delta \\pmb{s}^{t}) + (\\pmb{x}, \\pmb{r}, \\pmb{s})\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.741, + 0.483, + 0.815 + ], + "angle": 0, + "content": "This approach separates motion and geometric structural learning, allowing accurate simulation of dynamic behaviors while maintaining a stable geometric reference. To further enhance scene geometry, we estimate normals of deformed 3D Gaussians and optimize them using:" + }, + { + "type": "equation", + "bbox": [ + 0.208, + 0.827, + 0.483, + 0.844 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {n o r m a l}} = 1 - \\mathbf {N} ^ {t} \\cdot \\hat {\\mathbf {N}} ^ {t}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.855, + 0.484, + 0.901 + ], + "angle": 0, + "content": "where \\(\\mathbf{N}^t\\) is the rendered normal map and \\(\\hat{\\mathbf{N}}^t\\) is the normal map derived from the rendered depth map \\(\\mathbf{D}^t\\). This process improves local associations among 3D Gaussians and opti" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.605, + 0.906, + 0.633 + ], + "angle": 0, + "content": "mizes both depth and normal information across the entire scene." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.638, + 0.907, + 0.729 + ], + "angle": 0, + "content": "Specular stage. We adopt an image-based lighting (IBL) model, where the environment light is given by a learnable cubemap. Following the rendering equation [14], splitsum approximation [15, 30], and Cook-Torrance reflectance model [5], the outgoing radiance of the specular component \\( L_{s} \\) is expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.558, + 0.736, + 0.907, + 0.802 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} L _ {s} = \\int_ {\\Omega} \\frac {D G F}{4 (\\omega_ {o} ^ {t} \\cdot \\mathbf {n} ^ {t}) (\\omega_ {i} \\cdot \\mathbf {n} ^ {t})} (\\boldsymbol {\\omega} _ {i} \\cdot \\mathbf {n} ^ {t}) d \\boldsymbol {\\omega} _ {i} \\\\ \\times \\int_ {\\Omega} L _ {i} \\left(\\omega_ {i}\\right) D \\left(\\omega_ {i}, \\omega_ {o} ^ {t}\\right) \\left(\\omega_ {i} \\cdot \\mathbf {n} ^ {t}\\right) d \\omega_ {i}, \\tag {4} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.907, + 0.901 + ], + "angle": 0, + "content": "where \\(\\Omega\\) is the hemisphere around the surface normal \\(\\mathbf{n}^t\\) (describe in Sec. 3.3.) \\(D\\), \\(G\\), and \\(F\\) represent the GGX normal distribution function [43], geometric attenuation, and Fresnel term, respectively. \\(\\omega_o^t\\) is the view direction, and \\(L_i(\\omega_i)\\) is the incident radiance. In the first term, we follow the GaussianShader [13] directly computed by \\(\\mathbf{s}_{\\mathrm{int}} * F_1 + F_2\\)," + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21330" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.484, + 0.257 + ], + "angle": 0, + "content": "where \\(F_{1}\\) and \\(F_{2}\\) are two pre-computed scalars depending on roughness \\(\\rho\\), view direction \\(\\omega_{o}^{t}\\) and normal \\(\\mathbf{n}^t\\). Roughness \\(\\rho \\in [0,1]\\) and specular tint \\(\\mathbf{s}_{\\mathrm{int}} \\in [0,1]^3\\) are learnable parameters for each 3D Gaussian. The second term is pre-integrated in a filtered learnable cubemap, where each mip-level corresponds to a specific roughness value. The cubemap can be queried using the reflection direction to obtain the value of the second term. After the static and dynamic stages, the geometry is well-defined. This allows us to calculate reflection directions \\(\\omega_{r}^{t} = 2(\\omega_{o}^{t} \\cdot \\mathbf{n}^{t})\\mathbf{n}^{t} - \\omega_{o}^{t}\\) accurately." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.258, + 0.484, + 0.318 + ], + "angle": 0, + "content": "\\(L_{s}\\) represents only the specular color of the static environment light. To handle time-varying lighting in dynamic scenes, we introduce a deformable environment map, detailed in the following section." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.327, + 0.484, + 0.356 + ], + "angle": 0, + "content": "3.2.2. Deformable Environment Map for Dynamic Lighting." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.36, + 0.484, + 0.51 + ], + "angle": 0, + "content": "The concept of a deformable environment map involves treating the vanilla environment cubemap as a canonical environment map and combining it with a deformation field. This approach enables us to model time-varying lighting conditions effectively. We first apply positional encoding to the reflection direction \\(\\omega_r^t\\) and time \\(t\\) to implement this. These encoded values are then input into a deformable reflection MLP with parameters \\(\\theta_R\\). This process allows us to obtain the deformed reflection residual \\(\\Delta \\bar{\\omega}_r^t = F_{\\theta_R}(\\gamma (\\omega_r^t),\\gamma (t))\\) for each specified time \\(t\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.512, + 0.483, + 0.557 + ], + "angle": 0, + "content": "Subsequently, we add the deformed reflection residual \\(\\Delta \\bar{\\omega}_r^t\\) to the reflection direction \\(\\omega_r^t\\), yielding the deformed reflection direction \\(\\bar{\\omega}_r^t = \\Delta \\bar{\\omega}_r^t + \\omega_r^t\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.557, + 0.484, + 0.648 + ], + "angle": 0, + "content": "We can then use this deformed reflection direction \\(\\bar{\\omega}_r^t\\) to query the canonical environment map. The queried value is then multiplied by the first term of Equation 4, allowing us to obtain time-varying specular color \\(\\mathbf{c}_s^t\\). This approach effectively captures the dynamic nature of lighting in the scene while maintaining a stable canonical reference." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.656, + 0.483, + 0.686 + ], + "angle": 0, + "content": "3.2.3. Color Decomposition and Staged Training Strategy." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.484, + 0.885 + ], + "angle": 0, + "content": "We decompose the final color \\(\\mathbf{c}_{\\mathrm{final}}^t\\) into diffuse and specular components to better distinguish between high and low-frequency information: \\(\\mathbf{c}_{\\mathrm{final}}^t = \\mathbf{c}_d + \\mathbf{c}_s^t\\), where \\(\\mathbf{c_d}\\) is the diffuse color (using zero-order spherical harmonics as view-independent color), and \\(\\mathbf{c_s}^t\\) is the view-dependent color component. To manage the transition from spherical harmonics to \\(\\mathbf{c}_{\\mathrm{final}}^t\\) and mitigate potential geometry disruptions, in the early specular stage, we fix the deformable Gaussian MLP and most 3D Gaussian attributes, optimizing only zero-order SH, specular tint, and roughness. We temporarily suspend densification during this phase. As \\(\\mathbf{c}_{\\mathrm{final}}^t\\) becomes more complete, we gradually resume optimization of all parameters and reinstate the densification process." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.887, + 0.483, + 0.902 + ], + "angle": 0, + "content": "We further split the specular stage into two parts, applying" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.228 + ], + "angle": 0, + "content": "a coarse-to-fine strategy to the environment map. In the first part, we focus on optimizing the canonical environment map for time-invariant lighting. This establishes a stable foundation for the overall lighting structure. In the second part, we proceed to optimize the deformable reflection MLP for dynamic elements. This approach ensures a more robust learning process, allowing us to capture the static lighting conditions before introducing the complexities of dynamic components." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.235, + 0.766, + 0.25 + ], + "angle": 0, + "content": "3.3. Physical Normal Estimation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.257, + 0.908, + 0.483 + ], + "angle": 0, + "content": "Challenges in normal estimation for 3D Gaussians. Normal estimation is essential for modeling specular objects in 3D Gaussians, where GaussianShader [12] initially used the shortest axis combined with a residual normal for approximation. While this works for static scenes, it becomes problematic with deformed Gaussians because the residual should vary at each time step. A straightforward approach of rotating the residual normal based on quaternion differences between canonical and deformed states proves insufficient, as it does not account for shape changes during deformation. When deformation alters the relative axis lengths, the shortest axis assumption breaks down. This highlights the need for a more comprehensive approach that considers both rotational and shape deformation effects to achieve accurate normal estimation for dynamic specular objects." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.487, + 0.909, + 0.623 + ], + "angle": 0, + "content": "Improved rotation calculation for deformed 3D Gaussians. To overcome the limitations of naive methods and accurately model the normal of deformed 3D Gaussians, we propose using both the shortest and longest axes of canonical and deformed Gaussians to compute the rotation matrix. This approach accounts for both rotation and shape changes during deformation. We first align the deformed Gaussian's axes with those of the canonical Gaussian using the following method:" + }, + { + "type": "equation", + "bbox": [ + 0.513, + 0.629, + 0.935, + 0.684 + ], + "angle": 0, + "content": "\\[\n\\mathbf {v} _ {s} ^ {t} = \\left\\{ \\begin{array}{l l} \\mathbf {v} _ {s} ^ {t} & \\text {i f} \\mathbf {v} _ {s} \\cdot \\mathbf {v} _ {s} ^ {t} > 0, \\\\ - \\mathbf {v} _ {s} ^ {t} & \\text {o t h e r w i s e .} \\end{array} , \\quad \\mathbf {v} _ {l} ^ {t} = \\left\\{ \\begin{array}{l l} \\mathbf {v} _ {l} ^ {t} & \\text {i f} \\mathbf {v} _ {l} \\cdot \\mathbf {v} _ {l} ^ {t} > 0, \\\\ - \\mathbf {v} _ {l} ^ {t} & \\text {o t h e r w i s e .} \\end{array} , \\right. \\right. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.691, + 0.907, + 0.752 + ], + "angle": 0, + "content": "where \\(\\mathbf{v}_s\\) and \\(\\mathbf{v}_l\\) represent the shortest and longest axes of canonical 3D Gaussians, while \\(\\mathbf{v}_s^t\\) and \\(\\mathbf{v}_l^t\\) denote the same for deformed 3D Gaussians. We then construct orthogonal matrices using these aligned axes and their cross-products:" + }, + { + "type": "equation", + "bbox": [ + 0.52, + 0.756, + 0.907, + 0.788 + ], + "angle": 0, + "content": "\\[\n\\mathbf {U} = \\left[ \\begin{array}{l l l} \\mathbf {v} _ {s} & \\mathbf {v} _ {l} & \\mathbf {v} _ {s} \\times \\mathbf {v} _ {l} \\end{array} \\right], \\quad \\mathbf {V} ^ {t} = \\left[ \\begin{array}{l l l} \\mathbf {v} _ {s} ^ {t} & \\mathbf {v} _ {l} ^ {t} & \\mathbf {v} _ {s} ^ {t} \\times \\mathbf {v} _ {l} ^ {t} \\end{array} \\right]. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.788, + 0.857, + 0.803 + ], + "angle": 0, + "content": "Finally, we derive the rotation matrix \\(\\mathbf{R}^t = \\mathbf{V}^t\\mathbf{U}^\\top\\)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.807, + 0.907, + 0.867 + ], + "angle": 0, + "content": "Adjusting normal residuals and ensuring accuracy. To account for shape changes during deformation, we scale the normal residual based on the ratio of oblateness \\(\\frac{\\beta}{\\beta^t}\\) between canonical and deformed 3D Gaussians." + }, + { + "type": "equation", + "bbox": [ + 0.584, + 0.871, + 0.907, + 0.906 + ], + "angle": 0, + "content": "\\[\n\\beta = \\frac {\\left| \\mathbf {v} _ {l} \\right| - \\left| \\mathbf {v} _ {s} \\right|}{\\left| \\mathbf {v} _ {l} \\right|}, \\quad \\beta^ {t} = \\frac {\\left| \\mathbf {v} _ {l} ^ {t} \\right| - \\left| \\mathbf {v} _ {s} ^ {t} \\right|}{\\left| \\mathbf {v} _ {l} ^ {t} \\right|}, \\qquad (7)\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "21331" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.088, + 0.379, + 0.205 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.373, + 0.087, + 0.726, + 0.204 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.728, + 0.087, + 0.909, + 0.204 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.213, + 0.907, + 0.343 + ], + "angle": 0, + "content": "Figure 3. Normal estimation. (a) shows that flatter 3D Gaussians align better with scene surfaces, their shortest axis closely matching the surface normal. In contrast, less flat 3D Gaussians fit less accurately, with their shortest axis diverging from the surface normal. (b) shows that when the deformed 3D Gaussian becomes flatter (\\( t = t_1 \\)), normal residual \\( \\Delta \\mathbf{n} \\) is rotated by \\( \\mathbf{R}_1^t \\) and scaled down by \\( \\frac{\\beta}{\\beta_1^t} \\), as flatter Gaussians require smaller normal residuals. Conversely, when the deformation results in a less flat shape (\\( t = t_2 \\)), \\( \\Delta \\mathbf{n} \\) is rotated by \\( \\mathbf{R}_2^t \\) and amplified by \\( \\frac{\\beta}{\\beta_2^t} \\), requiring a larger correction to align the shortest axis with the surface normal. (c) shows how \\( \\gamma^k \\) changes with \\( w \\) (where \\( w = \\frac{|v_s^t|}{|v_l^t|} \\)) for \\( k = 1, k = 5 \\), and \\( k = 50 \\). Larger \\( w \\) indicates less flat Gaussians, while smaller \\( w \\) represents flatter Gaussians. As \\( k \\) increases, \\( \\gamma^k \\) decreases more steeply as \\( w \\) rises. For \\( k = 5 \\), we observe a balanced behavior: \\( \\gamma^k \\) approaches 1 for low \\( w \\) and 0 for high \\( w \\), providing a nuanced penalty adjustment across different Gaussian shapes." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.368, + 0.484, + 0.489 + ], + "angle": 0, + "content": "where \\(\\beta\\) and \\(\\beta^t\\) represent the oblateness of canonical and deformed 3D Gaussians, respectively. This is because flatter 3D Gaussians tend to align more closely with the surface, meaning their shortest axis becomes more aligned with the surface normal, as shown in Fig. 3 (a). In such cases, less compensation from the normal residual is needed. Conversely, less flat Gaussians require more compensation, as illustrated in Fig. 3 (b). We then obtain deformed normal residuals:" + }, + { + "type": "equation", + "bbox": [ + 0.225, + 0.498, + 0.484, + 0.529 + ], + "angle": 0, + "content": "\\[\n\\Delta \\mathbf {n} ^ {t} = \\frac {\\beta}{\\beta^ {t}} \\mathbf {R} ^ {t} \\Delta \\mathbf {n}. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.54, + 0.483, + 0.571 + ], + "angle": 0, + "content": "The final normal \\(\\mathbf{n}^t\\) is computed by adding this residual to the shortest axis and ensuring outward orientation:" + }, + { + "type": "equation", + "bbox": [ + 0.115, + 0.58, + 0.484, + 0.621 + ], + "angle": 0, + "content": "\\[\n\\mathbf {n} ^ {t} = \\Delta \\mathbf {n} ^ {t} + \\mathbf {v} _ {s} ^ {t}, \\quad \\mathbf {n} ^ {t} = \\left\\{ \\begin{array}{l l} \\mathbf {n} ^ {t} & \\text {i f} \\mathbf {n} ^ {t} \\cdot \\omega_ {o} ^ {t} > 0, \\\\ - \\mathbf {n} ^ {t} & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.629, + 0.485, + 0.658 + ], + "angle": 0, + "content": "This approach adjusts for Gaussian flatness and ensures accurate normal estimation." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.668, + 0.245, + 0.683 + ], + "angle": 0, + "content": "3.4. Loss Functions" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.69, + 0.484, + 0.751 + ], + "angle": 0, + "content": "Normal regularization. To allow the normal residual to correct the normal while not excessively influencing the optimization of the shortest axis towards the surface normal, we introduce a penalty term for the normal residual:" + }, + { + "type": "equation", + "bbox": [ + 0.12, + 0.762, + 0.484, + 0.802 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {r e g}} = \\gamma^ {k} \\| \\Delta \\mathbf {n} \\| _ {2} ^ {2} \\quad \\text {w h e r e} \\quad \\gamma = \\sqrt {1 - \\frac {\\left| \\mathbf {v} _ {s} ^ {t} \\right| ^ {2}}{\\left| \\mathbf {v} _ {l} ^ {t} \\right| ^ {2}}}. \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.485, + 0.901 + ], + "angle": 0, + "content": "In our experiments, we set \\( k = 5 \\). When \\( k = 5 \\), less flatter 3D Gaussians have \\( \\gamma^k \\) approaching 0. Their shortest axis aligns poorly with the surface normal, requiring more normal residual correction and smaller penalties. Conversely, flatter Gaussians have \\( \\gamma^k \\) near 1. Their shortest axis aligns better with the surface normal, needing less normal residual" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.369, + 0.905, + 0.398 + ], + "angle": 0, + "content": "correction and allowing larger penalties, as shown in Fig. 3 (c)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.402, + 0.906, + 0.434 + ], + "angle": 0, + "content": "Total training loss. To refine all parameters in the dynamic and specular stages, we employ the total training loss:" + }, + { + "type": "equation", + "bbox": [ + 0.594, + 0.445, + 0.907, + 0.462 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {\\text {c o l o r}} + \\lambda_ {\\text {n o r m a l}} \\mathcal {L} _ {\\text {n o r m a l}} + \\mathcal {L} _ {\\text {r e g}}, \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.471, + 0.907, + 0.532 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_{\\mathrm{color}}\\) and \\(\\mathcal{L}_{\\mathrm{normal}}\\) are obtained as described in Section 3.2.1. In our experiments, we set \\(\\lambda_{\\mathrm{normal}} = 0.01\\). Due to space constraints, complete implementation details are provided in the supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.545, + 0.646, + 0.561 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.569, + 0.695, + 0.584 + ], + "angle": 0, + "content": "4.1. Evaluation Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.591, + 0.907, + 0.682 + ], + "angle": 0, + "content": "We evaluate our method on two real-world datasets: NeRF-DS dataset [48] and HyperNeRF dataset [32]. While GaussianShader [12] and GS-IR [23] are originally designed for static scenes and are included here only as reference baselines, we train our method and all baseline approaches for 40,000 iterations to ensure fair comparison." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.686, + 0.907, + 0.806 + ], + "angle": 0, + "content": "NeRF-DS dataset. The NeRF-DS dataset [48] is a monocular video dataset comprising seven real-world scenes from daily life featuring various types of moving or deforming specular objects. We compare our method with the most relevant state-of-the-art approaches. As shown in Tab. 1 and Fig. 4, the quantitative results demonstrate that our method decisively outperforms baselines in reconstructing and rendering real-world highly reflective dynamic specular scenes." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.807, + 0.907, + 0.867 + ], + "angle": 0, + "content": "The rendering speed is correlated with the quantity of 3D Gaussians. When the number of 3D Gaussians is below 178k, our method can achieve real-time rendering over 30 FPS on an NVIDIA RTX 4090." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.909, + 0.902 + ], + "angle": 0, + "content": "HyperNeRF dataset. The HyperNeRF dataset contains real-world dynamic scenes and does not include specular objects." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "21332" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.908, + 0.12 + ], + "angle": 0, + "content": "Table 1. Quantitative comparison on the NeRF-DS [48] dataset. We report the average PSNR, SSIM, and LPIPS (VGG) of several previous models on test images. The best, the second best, and third best results are denoted by red, orange, yellow." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.123, + 0.907, + 0.413 + ], + "angle": 0, + "content": "
MethodAsBasinBellCup
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
Deformable 3DGS [49]26.040.88050.185019.530.78550.192423.960.79450.276724.490.88220.1658
4DGS [46]24.850.86320.203819.260.76700.219622.860.80150.206123.820.86950.1792
GaussianShader [12]21.890.77390.362017.790.66700.418720.690.81690.302420.400.74370.3385
GS-IR [23]21.580.80330.303318.060.72480.313520.660.78290.260320.340.81930.2719
NeRF-DS [48]25.340.88030.215020.230.80530.250822.570.78110.292124.510.88020.1707
HyperNeRF [32]17.590.85180.239022.580.81560.249719.800.76500.299915.450.82950.2302
Ours26.800.88430.176119.750.79150.189625.460.84900.160024.650.88710.1588
PlatePressSieveMean
MethodPSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
Deformable 3DGS [49]19.070.73520.359925.520.85940.196425.370.86160.164323.430.82840.2201
4DGS [46]18.770.77090.272124.820.83550.225525.160.85660.174522.790.82350.2115
GaussianShader [12]14.550.64230.495519.970.72440.450722.580.78620.305719.700.73630.3819
GS-IR [23]15.980.69690.420022.280.80880.306722.840.82120.223620.250.77960.2999
NeRF-DS [48]19.700.78130.297425.350.87030.255224.990.87050.200123.240.83840.2402
HyperNeRF [32]21.220.78290.316616.540.82000.281019.920.85210.214219.010.81670.2615
Ours20.840.81720.219826.490.86570.188925.220.87050.151324.170.85220.1778
" + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.423, + 0.907, + 0.573 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.3, + 0.584, + 0.696, + 0.598 + ], + "angle": 0, + "content": "Figure 4. Qualitative comparison on the NeRF-DS [48] dataset." + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.623, + 0.483, + 0.683 + ], + "angle": 0, + "content": "Table 2. Quantitative comparison on the HyperNeRF [32] dataset. We report the average PSNR, SSIM, and LPIPS (VGG) of several previous models. The best, the second best, and third best results are denoted by red, orange, yellow." + }, + { + "type": "table", + "bbox": [ + 0.104, + 0.687, + 0.472, + 0.824 + ], + "angle": 0, + "content": "
MethodPSNR ↑SSIM ↑LPIPS ↓
Deformable 3DGS [49]22.780.62010.3380
4DGS [46]24.890.67810.3422
GaussianShader [12]18.550.54520.4795
GS-IR [23]19.870.57290.4498
NeRF-DS [48]23.650.64050.3972
HyperNeRF [32]23.110.63870.3968
Ours22.220.60880.3295
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "As shown in Tab. 2 and Fig. 5, the results demonstrate that we are on par with state-of-the-art techniques for rendering novel views, and our method's performance is not limited to" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.626, + 0.605, + 0.639 + ], + "angle": 0, + "content": "shiny scenes." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.64, + 0.908, + 0.746 + ], + "angle": 0, + "content": "In Fig. 6, we compare our method's normal maps with those from Deformable 3DGS [49] and NeRF-DS [48]. For Deformable 3DGS [49], we obtain the normals by using the shortest axes of the deformed 3D Gaussians. As demonstrated, our method produces significantly better quality normal maps compared to Deformable 3DGS [49] and NeRF-DS [48]." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.754, + 0.667, + 0.77 + ], + "angle": 0, + "content": "4.2. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.776, + 0.907, + 0.806 + ], + "angle": 0, + "content": "For a fair comparison, we train our method and all ablation experiments for 40,000 iterations." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Different coarse to fine training strategy stages. As shown in Tab. 3 and Fig. 7, each stage contributes effectively to the model's performance. The Dynamic stage enhances dynamic object stability compared to the Static stage alone, while the Specular stage improves reflection clarity beyond the combined Static and Dynamic stages." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "21333" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.087, + 0.907, + 0.235 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.293, + 0.245, + 0.704, + 0.26 + ], + "angle": 0, + "content": "Figure 5. Qualitative comparison on the HyperNeRF [32] dataset." + }, + { + "type": "image", + "bbox": [ + 0.092, + 0.283, + 0.482, + 0.379 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.391, + 0.483, + 0.418 + ], + "angle": 0, + "content": "Figure 6. Qualitative comparison of normal maps between our method, Deformable 3DGS, and NeRF-DS." + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.436, + 0.483, + 0.464 + ], + "angle": 0, + "content": "Table 3. Ablation studies on different coarse to fine training strategy stages." + }, + { + "type": "table", + "bbox": [ + 0.115, + 0.468, + 0.46, + 0.541 + ], + "angle": 0, + "content": "
StagePSNR ↑SSIM ↑LPIPS ↓
Static20.260.77850.2953
St. + Dynamic24.020.85080.1831
St. + Dy. + Specular24.170.85220.1778
" + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.556, + 0.483, + 0.757 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.768, + 0.483, + 0.797 + ], + "angle": 0, + "content": "Figure 7. Qualitative comparison of each training stage in our coarse-to-fine approach." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.484, + 0.9 + ], + "angle": 0, + "content": "Ablation study on coarse-to-fine, and loss function. The model's performance was evaluated without key components: the coarse-to-fine training strategy, normal loss \\(\\mathcal{L}_{\\mathrm{normal}}\\), normal regularization \\(\\mathcal{L}_{\\mathrm{reg}}\\), and \\(\\gamma^k\\). Fig. 8 and Tab. 4 illustrate the effects of these omissions. Without the coarse-to-fine" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.284, + 0.907, + 0.313 + ], + "angle": 0, + "content": "Table 4. Ablation studies on different coarse to fine training strategy stages." + }, + { + "type": "table", + "bbox": [ + 0.522, + 0.317, + 0.899, + 0.417 + ], + "angle": 0, + "content": "
C2F\\( \\mathcal{L}_{\\text{normal}} \\)\\( \\mathcal{L}_{\\text{reg}} \\)\\( \\gamma^k \\)PSNR↑SSIM↑LPIPS↓
✓✓✓23.160.82940.2156
✓23.400.82770.2278
✓✓24.150.85100.1845
✓✓✓24.090.85150.1818
✓✓✓✓24.170.85220.1778
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.432, + 0.907, + 0.46 + ], + "angle": 0, + "content": "Table 5. Ablation studies on SH, Static and Deformable environment map." + }, + { + "type": "table", + "bbox": [ + 0.533, + 0.464, + 0.888, + 0.537 + ], + "angle": 0, + "content": "
PSNR ↑SSIM ↑LPIPS ↓
SH23.630.84530.1844
Static Env. map24.020.85080.1831
Deformable Env. map24.170.85220.1778
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.564, + 0.909, + 0.821 + ], + "angle": 0, + "content": "approach, the model, trained directly at the specular stage, produces incomplete scene geometry, affecting environment map queries for specular color. Omitting normal loss \\(\\mathcal{L}_{\\mathrm{normal}}\\) removes direct supervision on normals and leads to inaccurate reflection directions and less precise specular colors. Removing normal regularization \\(\\mathcal{L}_{\\mathrm{reg}}\\) allows the normal residual to dominate normal optimization, resulting in insufficient optimization of the 3D Gaussians' shortest axis towards the correct normal, which in turn reduces the rendering quality. The normal residual decreases for non-flattened and flat Gaussians without \\(\\gamma^k\\) in normal regularization. This particularly affects less flat 3D Gaussians whose shortest axis significantly deviates from the surface normal. The insufficient normal residual correction causes these 3D Gaussians' shortest axes to deviate greatly from their original direction in an attempt to align with the surface normal, ultimately reducing rendering quality." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Ablation study on SH, Static environment map, and Deformable environment map. Fig. 9 and Tab. 5 demonstrate the superiority of the deformable environment map over the static environment map, which in turn outperforms Spherical Harmonics (SH). SH struggles to accurately model high-" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21334" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.093, + 0.088, + 0.293, + 0.176 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.151, + 0.177, + 0.225, + 0.189 + ], + "angle": 0, + "content": "Full model" + }, + { + "type": "image", + "bbox": [ + 0.298, + 0.089, + 0.497, + 0.176 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.361, + 0.177, + 0.434, + 0.189 + ], + "angle": 0, + "content": "Full model" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.089, + 0.701, + 0.176 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.562, + 0.177, + 0.635, + 0.189 + ], + "angle": 0, + "content": "Full model" + }, + { + "type": "image", + "bbox": [ + 0.706, + 0.089, + 0.905, + 0.176 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.774, + 0.177, + 0.847, + 0.189 + ], + "angle": 0, + "content": "Full model" + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.192, + 0.293, + 0.279 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.121, + 0.28, + 0.248, + 0.292 + ], + "angle": 0, + "content": "w/o Coarse-to-fine" + }, + { + "type": "image", + "bbox": [ + 0.298, + 0.192, + 0.497, + 0.279 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.353, + 0.28, + 0.437, + 0.294 + ], + "angle": 0, + "content": "w/o \\(\\mathcal{L}_{\\mathrm{normal}}\\)" + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.192, + 0.701, + 0.279 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.571, + 0.28, + 0.633, + 0.295 + ], + "angle": 0, + "content": "w/o \\(\\mathcal{L}_{\\mathrm{reg}}\\)" + }, + { + "type": "image", + "bbox": [ + 0.706, + 0.192, + 0.905, + 0.279 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.783, + 0.28, + 0.832, + 0.294 + ], + "angle": 0, + "content": "w/o \\(\\gamma^k\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.243, + 0.309, + 0.753, + 0.324 + ], + "angle": 0, + "content": "Figure 8. Qualitative comparison of ablation study without different components." + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.347, + 0.289, + 0.432 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.119, + 0.434, + 0.266, + 0.446 + ], + "angle": 0, + "content": "Spherical Harmonics (SH)" + }, + { + "type": "image", + "bbox": [ + 0.291, + 0.348, + 0.482, + 0.433 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.341, + 0.434, + 0.429, + 0.446 + ], + "angle": 0, + "content": "Static Env. map" + }, + { + "type": "image", + "bbox": [ + 0.096, + 0.447, + 0.288, + 0.531 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.096, + 0.533, + 0.289, + 0.546 + ], + "angle": 0, + "content": "Deformable Env. map (Full model)" + }, + { + "type": "image", + "bbox": [ + 0.291, + 0.447, + 0.482, + 0.531 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.344, + 0.533, + 0.419, + 0.544 + ], + "angle": 0, + "content": "Ground-truth" + }, + { + "type": "image_caption", + "bbox": [ + 0.091, + 0.56, + 0.483, + 0.589 + ], + "angle": 0, + "content": "Figure 9. Qualitative comparison of ablation study on SH, Static environment map, and Deformable environment map." + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.608, + 0.483, + 0.636 + ], + "angle": 0, + "content": "Table 6. Ablation studies on 2DGS and without Physical Normal Estimation." + }, + { + "type": "table", + "bbox": [ + 0.143, + 0.64, + 0.432, + 0.713 + ], + "angle": 0, + "content": "
PSNR ↑SSIM ↑LPIPS ↓
2DGS [10]23.220.82190.2283
w/o N.E.23.890.84900.1837
Full model24.170.85220.1778
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.744, + 0.484, + 0.819 + ], + "angle": 0, + "content": "frequency specular colors. While the static environment map can model high-frequency colors, it is best suited for static lighting conditions. In contrast, the deformable environment map models time-varying lighting, offering superior performance for dynamic scenes." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.486, + 0.902 + ], + "angle": 0, + "content": "Ablation study on 2DGS [10] and without Physical Normal Estimation. In Fig. 10 and Tab. 6, \"2DGS\" represents replacing our 3D Gaussian with 2D Gaussian representation. Since 2DGS inherently includes normals, we omit physical normal estimation. \"w/o N.E.\" means skipping physical nor" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.347, + 0.708, + 0.431 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.571, + 0.432, + 0.643, + 0.444 + ], + "angle": 0, + "content": "Full model" + }, + { + "type": "image", + "bbox": [ + 0.712, + 0.347, + 0.905, + 0.431 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.773, + 0.432, + 0.845, + 0.443 + ], + "angle": 0, + "content": "Full model" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.447, + 0.708, + 0.531 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.591, + 0.532, + 0.632, + 0.543 + ], + "angle": 0, + "content": "2DGS" + }, + { + "type": "image", + "bbox": [ + 0.712, + 0.447, + 0.905, + 0.531 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.777, + 0.531, + 0.84, + 0.543 + ], + "angle": 0, + "content": "w/o N.E." + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.559, + 0.907, + 0.587 + ], + "angle": 0, + "content": "Figure 10. Qualitative comparison of ablation study on 2DGS and without Physical Normal Estimation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.615, + 0.907, + 0.69 + ], + "angle": 0, + "content": "mal estimation and using the shortest axis of 3D Gaussians as the normal. This causes the normal loss \\(\\mathcal{L}_{\\mathrm{normal}}\\) to directly supervise the shortest axes, making some axes deviate significantly to align with surface normals, resulting in degraded rendering quality." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.705, + 0.633, + 0.72 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.731, + 0.908, + 0.837 + ], + "angle": 0, + "content": "SpectroMotion enhances 3D Gaussian Splatting for dynamic specular scenes by combining specular rendering with deformation fields. Using normal residual correction, coarse-to-fine training, and a deformable environment map, it achieves superior accuracy and visual quality in novel view synthesis, outperforming existing methods while maintaining geometric consistency." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.909, + 0.9 + ], + "angle": 0, + "content": "Limitations. Though we stabilize the entire scene's geometry using a coarse-to-fine training strategy, some failure cases still occur. Please refer to the supplementary materials for visual results of failure cases." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21335" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.198 + ], + "angle": 0, + "content": "Acknowledgements. This research was funded by the National Science and Technology Council, Taiwan, under Grants NSTC 112-2222-E-A49-004-MY2 and 113-2628-E-A49-023-. The authors are grateful to Google, NVIDIA, and MediaTek Inc. for their generous donations. Yu-Lun Liu acknowledges the Yushan Young Fellow Program by the MOE in Taiwan." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.211, + 0.188, + 0.227 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.236, + 0.486, + 0.292 + ], + "angle": 0, + "content": "[1] Benjamin Attal, Eliot Laidlaw, Aaron Gokaslan, Changil Kim, Christian Richardt, James Tompkin, and Matthew O'Toole. Törf: Time-of-flight radiance fields for dynamic scene view synthesis, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.293, + 0.484, + 0.348 + ], + "angle": 0, + "content": "[2] Sai Bi, Zexiang Xu, Pratul Srinivasan, Ben Mildenhall, Kalyan Sunkavalli, Miloš Hasan, Yannick Hold-Geoffroy, David Kriegman, and Ravi Ramamoorthi. Neural reflectance fields for appearance acquisition, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.35, + 0.484, + 0.404 + ], + "angle": 0, + "content": "[3] Mark Boss, Varun Jampani, Raphael Braun, Ce Liu, Jonathan T. Barron, and Hendrik P. A. Lensch. Neural-pil: Neural pre-integrated lighting for reflectance decomposition, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.406, + 0.484, + 0.462 + ], + "angle": 0, + "content": "[4] Ting-Hsuan Chen, Jie Wen Chan, Hau-Shiang Shiu, Shih-Han Yen, Changhan Yeh, and Yu-Lun Liu. Narcan: Natural refined canonical image with integration of diffusion prior for video editing. In NeurIPS, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.463, + 0.484, + 0.504 + ], + "angle": 0, + "content": "[5] Robert L Cook and Kenneth E. Torrance. A reflectance model for computer graphics. ACM Transactions on Graphics (ToG), 1(1):7-24, 1982. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.506, + 0.485, + 0.575 + ], + "angle": 0, + "content": "[6] Ainaz Eftekhar, Alexander Sax, Jitendra Malik, and Amir Zamir. Omnidata: A scalable pipeline for making multi-task mid-level vision datasets from 3d scans. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10786-10796, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.577, + 0.484, + 0.631 + ], + "angle": 0, + "content": "[7] Jian Gao, Chun Gu, Youtian Lin, Hao Zhu, Xun Cao, Li Zhang, and Yao Yao. Relightable 3d gaussian: Real-time point cloud relighting with brdf decomposition and ray tracing. arXiv preprint arXiv:2311.16043, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.633, + 0.484, + 0.687 + ], + "angle": 0, + "content": "[8] Wenhang Ge, Tao Hu, Haoyu Zhao, Shu Liu, and Ying-Cong Chen. Ref-neus: Ambiguity-reduced neural implicit surface learning for multi-view reconstruction with reflection, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.689, + 0.484, + 0.744 + ], + "angle": 0, + "content": "[9] Xiang Guo, Jiadai Sun, Yuchao Dai, Guanying Chen, Xiaoqing Ye, Xiao Tan, Errui Ding, Yumeng Zhang, and Jingdong Wang. Forward flow for novel view synthesis of dynamic scenes, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.746, + 0.484, + 0.802 + ], + "angle": 0, + "content": "[10] Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2d gaussian splatting for geometrically accurate radiance fields. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.804, + 0.484, + 0.845 + ], + "angle": 0, + "content": "[11] Yi-Hua Huang, Yang-Tian Sun, Ziyi Yang, Xiaoyang Lyu, Yan-Pei Cao, and Xiaojuan Qi. Sc-gs: Sparse-controlled gaussian splatting for editable dynamic scenes, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[12] Yingwenqi Jiang, Jiadong Tu, Yuan Liu, Xifeng Gao, Xiaoxiao Long, Wenping Wang, and Yuexin Ma. Gaussianshader: 3d gaussian splatting with shading functions for reflective surfaces. arXiv preprint arXiv:2311.17977, 2023. 2, 4, 5, 6" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.236, + 0.486, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.174 + ], + "angle": 0, + "content": "[13] Yingwenqi Jiang, Jiadong Tu, Yuan Liu, Xifeng Gao, Xiaoxiao Long, Wenping Wang, and Yuexin Ma. Gaussianshader: 3d gaussian splatting with shading functions for reflective surfaces. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5322-5332, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.176, + 0.907, + 0.218 + ], + "angle": 0, + "content": "[14] James T Kajiya. The rendering equation. In Proceedings of the 13th annual conference on Computer graphics and interactive techniques, pages 143-150, 1986. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.219, + 0.907, + 0.258 + ], + "angle": 0, + "content": "[15] Brian Karis and Epic Games. Real shading in unreal engine 4. Proc. Physically Based Shading Theory Practice, 4(3):1, 2013. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.26, + 0.906, + 0.301 + ], + "angle": 0, + "content": "[16] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.302, + 0.906, + 0.329 + ], + "angle": 0, + "content": "[17] Junxuan Li and Hongdong Li. Neural reflectance for shape recovery with shadow handling, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.33, + 0.907, + 0.398 + ], + "angle": 0, + "content": "[18] Tianye Li, Mira Slavcheva, Michael Zollhoefer, Simon Green, Christoph Lassner, Changil Kim, Tanner Schmidt, Steven Lovegrove, Michael Goesele, Richard Newcombe, and Zhaoyang Lv. Neural 3d video synthesis from multi-view video, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.4, + 0.907, + 0.44 + ], + "angle": 0, + "content": "[19] Zhengqi Li, Simon Niklaus, Noah Snively, and Oliver Wang. Neural scene flow fields for space-time view synthesis of dynamic scenes, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.442, + 0.908, + 0.495 + ], + "angle": 0, + "content": "[20] Ruofan Liang, Huiting Chen, Chunlin Li, Fan Chen, Selvakumar Panneer, and Nandita Vijaykumar. Envidr: Implicit differentiable renderer with neural environment lighting, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.497, + 0.906, + 0.538 + ], + "angle": 0, + "content": "[21] Ruofan Liang, Jiahao Zhang, Haoda Li, Chen Yang, Yushi Guan, and Nandita Vijaykumar. Spidr: Sdf-based neural point fields for illumination and deformation, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.539, + 0.908, + 0.594 + ], + "angle": 0, + "content": "[22] Yiqing Liang, Numair Khan, Zhengqin Li, Thu Nguyen-Phuoc, Douglas Lanman, James Tompkin, and Lei Xiao. Gaufre: Gaussian deformation fields for real-time dynamic novel view synthesis, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.596, + 0.906, + 0.637 + ], + "angle": 0, + "content": "[23] Zhihao Liang, Qi Zhang, Ying Feng, Ying Shan, and Kui Jia. Gs-ir: 3d gaussian splatting for inverse rendering. arXiv preprint arXiv:2311.16473, 2023. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.638, + 0.906, + 0.678 + ], + "angle": 0, + "content": "[24] Haotong Lin, Sida Peng, Zhen Xu, Yunzhi Yan, Qing Shuai, Hujun Bao, and Xiaowei Zhou. Efficient neural radiance fields for interactive free-viewpoint video, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.679, + 0.906, + 0.734 + ], + "angle": 0, + "content": "[25] Yuan Liu, Peng Wang, Cheng Lin, Xiaoxiao Long, Jiepeng Wang, Lingjie Liu, Taku Komura, and Wenping Wang. Nano: Neural geometry and brdf reconstruction of reflective objects from multiview images, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.735, + 0.907, + 0.788 + ], + "angle": 0, + "content": "[26] Yu-Lun Liu, Chen Gao, Andreas Meuleman, Hung-Yu Tseng, Ayush Saraf, Changil Kim, Yung-Yu Chuang, Johannes Kopf, and Jia-Bin Huang. Robust dynamic radiance fields. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.79, + 0.907, + 0.843 + ], + "angle": 0, + "content": "[27] Caoyuan Ma, Yu-Lun Liu, Zhixiang Wang, Wu Liu, Xinchen Liu, and Zheng Wang. Humannerf-se: A simple yet effective approach to animate humannerf with diverse poses. In CVPR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.845, + 0.907, + 0.901 + ], + "angle": 0, + "content": "[28] Li Ma, Vasu Agrawal, Haithem Turki, Changil Kim, Chen Gao, Pedro Sander, Michael Zollhöfer, and Christian Richardt. Specnerf: Gaussian directional encoding for specular reflections. arXiv preprint arXiv:2312.13102, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "21336" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.148 + ], + "angle": 0, + "content": "[29] Marko Mihajlovic, Sergey Prokudin, Siyu Tang, Robert Maier, Federica Bogo, Tony Tung, and Edmond Boyer. Splatfields: Neural gaussian splats for sparse 3d and 4d reconstruction. arXiv preprint arXiv:2409.11211, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.485, + 0.234 + ], + "angle": 0, + "content": "[30] Jacob Munkberg, Jon Hasselgren, Tianchang Shen, Jun Gao, Wenzheng Chen, Alex Evans, Thomas Müller, and Sanja Fidler. Extracting Triangular 3D Models, Materials, and Lighting From Images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8280-8290, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.234, + 0.484, + 0.289 + ], + "angle": 0, + "content": "[31] Keunhong Park, Utkarsh Sinha, Jonathan T. Barron, Sofien Bouaziz, Dan B Goldman, Steven M. Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.291, + 0.484, + 0.36 + ], + "angle": 0, + "content": "[32] Keunhong Park, Utkarsh Sinha, Peter Hedman, Jonathan T. Barron, Sofien Bouaziz, Dan B Goldman, Ricardo MartinBrualla, and Steven M. Seitz. Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields, 2021. 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.361, + 0.484, + 0.403 + ], + "angle": 0, + "content": "[33] Sida Peng, Yunzhi Yan, Qing Shuai, Hujun Bao, and Xiaowei Zhou. Representing volumetric videos as dynamic mlp maps, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.404, + 0.484, + 0.447 + ], + "angle": 0, + "content": "[34] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-nerf: Neural radiance fields for dynamic scenes, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.448, + 0.484, + 0.517 + ], + "angle": 0, + "content": "[35] Yahao Shi, Yanmin Wu, Chenming Wu, Xing Liu, Chen Zhao, Haocheng Feng, Jingtuo Liu, Liangjun Zhang, Jian Zhang, Bin Zhou, et al. Gir: 3d gaussian inverse rendering for relightable scene factorization. arXiv preprint arXiv:2312.05133, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.518, + 0.484, + 0.573 + ], + "angle": 0, + "content": "[36] Liangchen Song, Anpei Chen, Zhong Li, Zhang Chen, Lele Chen, Junsong Yuan, Yi Xu, and Andreas Geiger. Nerfplayer: A streamable dynamic scene representation with decomposed neural radiance fields, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.575, + 0.484, + 0.631 + ], + "angle": 0, + "content": "[37] Pratul P. Srinivasan, Boyang Deng, Xiuming Zhang, Matthew Tancik, Ben Mildenhall, and Jonathan T. Barron. Nerv: Neural reflectance and visibility fields for relighting and view synthesis, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.632, + 0.484, + 0.689 + ], + "angle": 0, + "content": "[38] Colton Stearns, Adam Harley, Mikaela Uy, Florian Dubost, Federico Tombari, Gordon Wetzstein, and Leonidas Guibas. Dynamic gaussian marbles for novel view synthesis of casual monocular videos. arXiv preprint arXiv:2406.18717, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.689, + 0.484, + 0.746 + ], + "angle": 0, + "content": "[39] Edgar Tretschk, Ayush Tewari, Vladislav Golyanik, Michael Zollhöfer, Christoph Lassner, and Christian Theobalt. Non-rigid neural radiance fields: Reconstruction and novel view synthesis of a dynamic scene from monocular video, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.746, + 0.484, + 0.817 + ], + "angle": 0, + "content": "[40] Dor Verbin, Peter Hedman, Ben Mildenhall, Todd Zickler, Jonathan T Barron, and Pratul P Srinivasan. Ref-nerf: Structured view-dependent appearance for neural radiance fields. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5481–5490. IEEE, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.818, + 0.484, + 0.899 + ], + "angle": 0, + "content": "[41] Dor Verbin, Ben Mildenhall, Peter Hedman, Jonathan T Barron, Todd Zickler, and Pratul P Srinivasan. Eclipse: Disambiguating illumination and materials using unintended shadows. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 77-86, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.092, + 0.908, + 0.161 + ], + "angle": 0, + "content": "[42] Dor Verbin, Pratul P Srinivasan, Peter Hedman, Ben Mildenhall, Benjamin Attal, Richard Szeliski, and Jonathan T Barron. Nerf-casting: Improved view-dependent appearance with consistent reflections. arXiv preprint arXiv:2405.14871, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.162, + 0.908, + 0.207 + ], + "angle": 0, + "content": "[43] Bruce Walter, Stephen R Marschner, Hongsong Li, and Kenneth E Torrance. Microfacet models for refraction through rough surfaces. Rendering techniques, 2007:18th, 2007. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.207, + 0.908, + 0.261 + ], + "angle": 0, + "content": "[44] Qianqian Wang, Vickie Ye, Hang Gao, Jake Austin, Zhengqi Li, and Angjoo Kanazawa. Shape of motion: 4d reconstruction from a single video. arXiv preprint arXiv:2407.13764, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.262, + 0.907, + 0.332 + ], + "angle": 0, + "content": "[45] Chun-Hung Wu, Shih-Hong Chen, Chih-Yao Hu, Hsin-Yu Wu, Kai-Hsin Chen, Yu-You Chen, Chih-Hai Su, Chih-Kuo Lee, and Yu-Lun Liu. Denver: Deformable neural vessel representations for unsupervised video vessel segmentation. In CVPR, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.332, + 0.908, + 0.39 + ], + "angle": 0, + "content": "[46] Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Xinggang Wang. 4d gaussian splatting for real-time dynamic scene rendering. arXiv preprint arXiv:2310.08528, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.39, + 0.907, + 0.432 + ], + "angle": 0, + "content": "[47] Wenqi Xian, Jia-Bin Huang, Johannes Kopf, and Changil Kim. Space-time neural irradiance fields for free-viewpoint video, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.432, + 0.906, + 0.49 + ], + "angle": 0, + "content": "[48] Zhiwen Yan, Chen Li, and Gim Hee Lee. Nerf-ds: Neural radiance fields for dynamic specular objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8285-8295, 2023. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.49, + 0.906, + 0.546 + ], + "angle": 0, + "content": "[49] Ziyi Yang, Xinyu Gao, Wen Zhou, Shaohui Jiao, Yuqing Zhang, and Xiaogang Jin. Deformable 3d gaussians for high-fidelity monocular dynamic scene reconstruction. arXiv preprint arXiv:2309.13101, 2023. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.546, + 0.906, + 0.604 + ], + "angle": 0, + "content": "[50] Ziyi Yang, Xinyu Gao, Yangtian Sun, Yihua Huang, Xiaoyang Lyu, Wen Zhou, Shaohui Jiao, Xiaojuan Qi, and Xiaogang Jin. Spec-gaussian: Anisotropic view-dependent appearance for 3d gaussian splatting, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.604, + 0.907, + 0.647 + ], + "angle": 0, + "content": "[51] Keyang Ye, Qiming Hou, and Kun Zhou. 3d gaussian splatting with deferred reflection. In ACM SIGGRAPH 2024 Conference Papers, pages 1-10, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.647, + 0.906, + 0.703 + ], + "angle": 0, + "content": "[52] Wang Yifan, Felice Serena, Shihao Wu, Cengiz Oztireli, and Olga Sorkine-Hornung. Differentiable surface splatting for point-based geometry processing. ACM Transactions on Graphics (TOG), 38(6):1-14, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.703, + 0.907, + 0.759 + ], + "angle": 0, + "content": "[53] Jingyang Zhang, Yao Yao, Shiwei Li, Jingbo Liu, Tian Fang, David McKinnon, Yanghai Tsin, and Long Quan. Neilf++: Inter-reflectable light fields for geometry and material estimation, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.759, + 0.907, + 0.831 + ], + "angle": 0, + "content": "[54] Kai Zhang, Fujun Luan, Qianqian Wang, Kavita Bala, and Noah Snavely. Physg: Inverse rendering with spherical gaussians for physics-based material editing and relighting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5453-5462, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.831, + 0.907, + 0.901 + ], + "angle": 0, + "content": "[55] Xiuming Zhang, Pratul P. Srinivasan, Boyang Deng, Paul Debevec, William T. Freeman, and Jonathan T. Barron. Ner-factor: neural factorization of shape and reflectance under an unknown illumination. ACM Transactions on Graphics, 40 (6):1-18, 2021." + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.092, + 0.908, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "21337" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.487, + 0.147 + ], + "angle": 0, + "content": "[56] Xiaoming Zhao, Pratul P Srinivasan, Dor Verbin, Keunhong Park, Ricardo Martin Brualla, and Philipp Henzler. Illuminerf: 3d relighting without inverse rendering. arXiv preprint arXiv:2406.06527, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.486, + 0.192 + ], + "angle": 0, + "content": "[57] Zuo-Liang Zhu, Beibei Wang, and Jian Yang. Gs-ror: 3d gaussian splatting for reflective object relighting via sdf priors. arXiv preprint arXiv:2406.18544, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.487, + 0.192 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.956 + ], + "angle": 0, + "content": "21338" + } + ] +] \ No newline at end of file diff --git a/2025/SpectroMotion_ Dynamic 3D Reconstruction of Specular Scenes/5055bce6-6fec-402b-adcc-413e1a1becdd_origin.pdf b/2025/SpectroMotion_ Dynamic 3D Reconstruction of Specular Scenes/5055bce6-6fec-402b-adcc-413e1a1becdd_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..6650cec79b1e3ef67c2442ec220e04811d555605 --- /dev/null +++ b/2025/SpectroMotion_ Dynamic 3D Reconstruction of Specular Scenes/5055bce6-6fec-402b-adcc-413e1a1becdd_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8aa8458ed0b1296d0e850b92a7bde0c1852dc892cfe158d6e12b42bbfc4f64b1 +size 3601280 diff --git a/2025/SpectroMotion_ Dynamic 3D Reconstruction of Specular Scenes/full.md b/2025/SpectroMotion_ Dynamic 3D Reconstruction of Specular Scenes/full.md new file mode 100644 index 0000000000000000000000000000000000000000..93ac48ca014ed6fa2f7a9b96f37f97329add25c2 --- /dev/null +++ b/2025/SpectroMotion_ Dynamic 3D Reconstruction of Specular Scenes/full.md @@ -0,0 +1,380 @@ +# SpectroMotion: Dynamic 3D Reconstruction of Specular Scenes + +Cheng-De Fan1 Chen-Wei Chang1 Yi-Ruei Liu1,2 Jie-Ying Lee1 +Jiun-Long Huang1 Yu-Chee Tseng1 Yu-Lun Liu1 + +$^{1}$ National Yang Ming Chiao Tung University $^{2}$ University of Illinois Urbana-Champaign + +# Abstract + +We present SpectroMotion, a novel approach that combines 3D Gaussian Splatting (3DGS) with physically-based rendering (PBR) and deformation fields to reconstruct dynamic specular scenes. Previous methods extending 3DGS to model dynamic scenes have struggled to represent specular surfaces accurately. Our method addresses this limitation by introducing a residual correction technique for accurate surface normal computation during deformation, complemented by a deformable environment map that adapts to time-varying lighting conditions. We implement a coarse-to-fine training strategy significantly enhancing scene geometry and specular color prediction. It is the only existing 3DGS method capable of synthesizing photorealistic real-world dynamic specular scenes, outperforming state-of-the-art methods in rendering complex, dynamic, and specular scenes. Please see our project page at cdfan0627.github.io/spectromotion. + +# 1. Introduction + +3D Gaussian Splitting (3DGS) [16] has emerged as a groundbreaking technique in 3D scene reconstruction, offering fast training and real-time rendering capabilities. By representing 3D scene using a collection of 3D Gaussians and employing a point-based rendering approach, 3DGS has significantly improved efficiency in novel view synthesis. However, extending 3DGS to model dynamic scenes, especially those containing specular surfaces accurately, has remained a significant challenge. + +Existing extensions of 3DGS have made progress in either dynamic scene reconstruction or specular object rendering, but none have successfully combined both aspects. Methods tackling dynamic scenes often struggle with accurately representing specular surfaces, while those focusing on specular rendering are limited to static scenes. This capability gap has hindered the application of 3DGS to real-world scenarios where both motion and specular reflections are present. + +We present SpectroMotion, a novel approach that ad + +![](images/c78dbd9e6a09ad3b87ada87fe470b475a575a7b82e7cb685a9ec4005fe669403.jpg) +Figure 1. Our method, SpectroMotion, recovers and renders dynamic scenes with higher-quality reflections compared to prior work. It introduces physical normal estimation, deformable environment maps, and a coarse-to-fine training strategy to achieve superior results in rendering dynamic scenes with reflections. Here, we present a rendered test image, corresponding normal maps, and a ground-truth image, where the ground-truth normal map (used as a reference) is generated using a pre-trained normal estimator [6]. For Deformable 3DGS, we use the shortest axes of the deformed 3D Gaussians as the normals. We have highlighted the specular regions to demonstrate the effectiveness of our approach. + +dresses these limitations by combining 3D Gaussian Splating with physically based rendering (PBR) and deformation fields. Our method introduces three key innovations: a residual correction technique for accurate surface normal computation during deformation, a deformable environment map that adapts to time-varying lighting conditions, and a coarse-to-fine training strategy that significantly enhances scene geometry and specular color prediction. + +Our evaluations demonstrate that SpectroMotion outperforms prior methods in view synthesis of scenes containing dynamic specular objects, as illustrated in Fig. 1. It is the only 3DGS method capable of synthesizing photorealistic real-world dynamic specular scenes, surpassing state-of-the + +art techniques in rendering complex, dynamic, and specular content. This advancement represents a significant leap in 3D scene reconstruction, particularly for challenging scenarios involving moving specular objects. + +In summary, we make the following contributions: + +- We propose SpectroMotion, a physically-based rendering (PBR) approach combining deformation fields and 3D Gaussian Splatting for real-world dynamic specular scenes. +- We introduce a residual correction method for accurate surface normals during deformation, coupled with a deformable environment map to handle time-varying lighting conditions in dynamic scenes. +- We develop a coarse-to-fine training strategy enhancing scene geometry and specular color prediction, outperforming state-of-the-art methods. + +# 2. Related Work + +# 2.1. Dynamic Scene Reconstruction + +Recent works have leveraged NeRF representations to jointly solve for canonical space and deformation fields in dynamic scenes using RGB supervision [4, 9, 19, 26, 27, 31, 32, 34, 39, 45, 47]. Further advancements in dynamic neural rendering include object segmentation [36], incorporation of depth information [1], utilization of 2D CNNs for scene priors [24, 33], and multi-view video compression [18]. However, these NeRF-based methods are computationally intensive, limiting their practical applications. To address this, 3D Gaussian Splatting [16] has emerged as a promising alternative, offering real-time rendering capabilities while maintaining high visual quality. Building upon this efficient representation, recent research has adapted 3D Gaussians for dynamic scenes [11, 22, 29, 38, 44, 46, 49]. Nevertheless, these approaches do not explicitly account for changes in surface normal during the dynamic process. Our work extends this line of research by combining specular object rendering based on normal estimation with a deformation field, enabling each 3D Gaussian to model dynamic specular scenes effectively. + +# 2.2. Reflective Object Rendering + +While significant progress has been made in rendering reflective objects, challenges from complex light interactions persist. Recent years have seen numerous studies addressing these issues, primarily by decomposing appearance into lighting and material properties [2, 3, 17, 30, 37, 41, 54-56]. Building on this foundation, some research has focused on improving the capture and reproduction of specular reflections [28, 40, 42, 51]. In contrast, others have leveraged signed distance functions (SDFs) to enhance normal estimation [8, 20, 21, 25, 53]. The emergence of 3D Gaussian Splatting (3DGS) has sparked a new wave of techniques + +[7, 12, 23, 35, 50, 57] that integrate Gaussian splating with physically-based rendering. Nevertheless, accurately modeling dynamic environments and time-varying specular reflections remains a significant challenge. To address this limitation, our work introduces a novel approach incorporating a deformable environment map and additional explicit Gaussian attributes specifically designed to capture specular color changes over time. + +# 3. Method + +Overview of the approach. The overview of our method is illustrated in Fig. 2. Given an input monocular video sequence of frames and corresponding camera poses, we design a three-stage approach to reconstruct the dynamic specular scene, as detailed in Sec. 3.2. Accurate specular reflection requires precise normal estimation, so Sec. 3.3 elaborates on how we estimate normals in dynamic scenes. Finally, we introduce the losses used throughout the training process in Sec. 3.4. + +# 3.1. Preliminary + +3D Gaussian Splatting. Each 3D Gaussian is defined by a center position $\boldsymbol{x} \in \mathbb{R}^3$ and a covariance matrix $\Sigma$ . 3D Gaussian Splatting [16] optimizes the covariance matrix using scaling factors $s \in \mathbb{R}^3$ and rotation unit quaternion $r \in \mathbb{R}^4$ . For novel-view rendering, 3D Gaussians are projected onto 2D camera planes using differentiable splatting [52]: + +$$ +\boldsymbol {\Sigma} ^ {\prime} = \mathbf {J} \mathbf {W} \boldsymbol {\Sigma} \mathbf {W} ^ {T} \mathbf {J} ^ {T}. \tag {1} +$$ + +Pixel colors are computed using point-based volumetric rendering: + +$$ +C = \sum_ {i \in N} T _ {i} \alpha_ {i} c _ {i}, \quad \alpha_ {i} = \sigma_ {i} e ^ {- \frac {1}{2} (\boldsymbol {x}) ^ {T} \boldsymbol {\Sigma} ^ {\prime} (\boldsymbol {x})}, \tag {2} +$$ + +where $T_{i} = \prod_{j=1}^{i-1}(1 - \alpha_{j})$ is the transmittance, $\sigma_{i}$ is the opacity, and $\mathbf{c}_{i}$ is the color of each 3D Gaussian. + +# 3.2. Specular Rendering + +Since accurate reflections depend heavily on precise geometry, we implement a three-stage coarse-to-fine training strategy: static, dynamic, and specular stages. This approach ensures both stable scene geometry and accurate specular rendering. + +# 3.2.1. Coarse-to-Fine Training Strategy + +Static stage. In the static stage, we employ vanilla 3DGS [16] for static scene reconstruction to stabilize the geometry of the static scene. Specifically, we optimize the position $\mathbf{x}$ , scaling $s$ , rotation $r$ , opacity $\alpha$ , and coefficients of spherical harmonics (SH) of the 3D Gaussians by minimizing the photometric loss $\mathcal{L}_{\mathrm{color}}$ identical to 3DGS [16]. + +![](images/2fe4de96c4e1a699f603f34df7fee1b9f4e35d3c3ff1b8bcea1ebf92689f0f0f.jpg) + +![](images/7c2262654f0f8b253faefba07eab885d64d180700e9c207ad6f7c94f2122b8a4.jpg) + +![](images/4cf7c210cddea336779c2a496eb78646127e1383a8ba0777c5f44f6ea13d1a4a.jpg) +Figure 2. Method Overview. Our method stabilizes the scene geometry through three stages. In the static stage, we stabilize the geometry of the static scene by minimizing photometric loss $\mathcal{L}_{\mathrm{color}}$ between vanilla 3DGS renders and ground truth images. The dynamic stage combines canonical 3D Gaussians $\mathbf{G}$ with a deformable Gaussian MLP to model dynamic scenes while simultaneously minimizing normal loss $\mathcal{L}_{\mathrm{normal}}$ between rendered normal map $\mathbf{N}^t$ and gradient normal map from depth map $\mathbf{D}^t$ , thus further enhancing the overall scene geometry. Finally, the specular stage introduces a deformable reflection MLP to handle changing environment lighting, deforming reflection directions $\omega_r^t$ to query a canonical environment map for specular color $\mathbf{c}_s^t$ . It is then combined with diffuse color $\mathbf{c_d}$ (using zero-order spherical harmonics) and learnable specular tint $\mathbf{s}_{\mathrm{int}}$ per 3D Gaussian to obtain the final color $\mathbf{c}_{\mathrm{final}}^t$ . This approach enables the modeling of dynamic specular scenes and high-quality novel view rendering. + +Dynamic stage. Following the static stage, we address dynamic objects using Deformable 3DGS [49]. For each 3D Gaussian in canonical 3D Gaussians $\mathbf{G}$ , we input its position $\pmb{x}$ and time $t$ into a deformable Gaussian MLP with parameters $\theta_{G}$ to predict position, rotation, and scaling residuals: $(\Delta \pmb{x}^{t}, \Delta \pmb{r}^{t}, \Delta \pmb{s}^{t}) = F_{\theta_{G}}(\gamma(x), \gamma(t))$ , where $\gamma$ denotes positional encoding. Attributes of the corresponding 3D Gaussian in deformed 3D Gaussians $\mathbf{G}^{t}$ at time $t$ is obtained by $(\pmb{x}^{t}, \pmb{r}^{t}, \pmb{s}^{t}) = (\Delta \pmb{x}^{t}, \Delta \pmb{r}^{t}, \Delta \pmb{s}^{t}) + (\pmb{x}, \pmb{r}, \pmb{s})$ . + +This approach separates motion and geometric structural learning, allowing accurate simulation of dynamic behaviors while maintaining a stable geometric reference. To further enhance scene geometry, we estimate normals of deformed 3D Gaussians and optimize them using: + +$$ +\mathcal {L} _ {\text {n o r m a l}} = 1 - \mathbf {N} ^ {t} \cdot \hat {\mathbf {N}} ^ {t}, \tag {3} +$$ + +where $\mathbf{N}^t$ is the rendered normal map and $\hat{\mathbf{N}}^t$ is the normal map derived from the rendered depth map $\mathbf{D}^t$ . This process improves local associations among 3D Gaussians and opti + +mizes both depth and normal information across the entire scene. + +Specular stage. We adopt an image-based lighting (IBL) model, where the environment light is given by a learnable cubemap. Following the rendering equation [14], splitsum approximation [15, 30], and Cook-Torrance reflectance model [5], the outgoing radiance of the specular component $L_{s}$ is expressed as: + +$$ +\begin{array}{l} L _ {s} = \int_ {\Omega} \frac {D G F}{4 (\omega_ {o} ^ {t} \cdot \mathbf {n} ^ {t}) (\omega_ {i} \cdot \mathbf {n} ^ {t})} (\boldsymbol {\omega} _ {i} \cdot \mathbf {n} ^ {t}) d \boldsymbol {\omega} _ {i} \\ \times \int_ {\Omega} L _ {i} \left(\omega_ {i}\right) D \left(\omega_ {i}, \omega_ {o} ^ {t}\right) \left(\omega_ {i} \cdot \mathbf {n} ^ {t}\right) d \omega_ {i}, \tag {4} \\ \end{array} +$$ + +where $\Omega$ is the hemisphere around the surface normal $\mathbf{n}^t$ (describe in Sec. 3.3.) $D$ , $G$ , and $F$ represent the GGX normal distribution function [43], geometric attenuation, and Fresnel term, respectively. $\omega_o^t$ is the view direction, and $L_i(\omega_i)$ is the incident radiance. In the first term, we follow the GaussianShader [13] directly computed by $\mathbf{s}_{\mathrm{int}} * F_1 + F_2$ , + +where $F_{1}$ and $F_{2}$ are two pre-computed scalars depending on roughness $\rho$ , view direction $\omega_{o}^{t}$ and normal $\mathbf{n}^t$ . Roughness $\rho \in [0,1]$ and specular tint $\mathbf{s}_{\mathrm{int}} \in [0,1]^3$ are learnable parameters for each 3D Gaussian. The second term is pre-integrated in a filtered learnable cubemap, where each mip-level corresponds to a specific roughness value. The cubemap can be queried using the reflection direction to obtain the value of the second term. After the static and dynamic stages, the geometry is well-defined. This allows us to calculate reflection directions $\omega_{r}^{t} = 2(\omega_{o}^{t} \cdot \mathbf{n}^{t})\mathbf{n}^{t} - \omega_{o}^{t}$ accurately. + +$L_{s}$ represents only the specular color of the static environment light. To handle time-varying lighting in dynamic scenes, we introduce a deformable environment map, detailed in the following section. + +# 3.2.2. Deformable Environment Map for Dynamic Lighting. + +The concept of a deformable environment map involves treating the vanilla environment cubemap as a canonical environment map and combining it with a deformation field. This approach enables us to model time-varying lighting conditions effectively. We first apply positional encoding to the reflection direction $\omega_r^t$ and time $t$ to implement this. These encoded values are then input into a deformable reflection MLP with parameters $\theta_R$ . This process allows us to obtain the deformed reflection residual $\Delta \bar{\omega}_r^t = F_{\theta_R}(\gamma (\omega_r^t),\gamma (t))$ for each specified time $t$ . + +Subsequently, we add the deformed reflection residual $\Delta \bar{\omega}_r^t$ to the reflection direction $\omega_r^t$ , yielding the deformed reflection direction $\bar{\omega}_r^t = \Delta \bar{\omega}_r^t + \omega_r^t$ . + +We can then use this deformed reflection direction $\bar{\omega}_r^t$ to query the canonical environment map. The queried value is then multiplied by the first term of Equation 4, allowing us to obtain time-varying specular color $\mathbf{c}_s^t$ . This approach effectively captures the dynamic nature of lighting in the scene while maintaining a stable canonical reference. + +# 3.2.3. Color Decomposition and Staged Training Strategy. + +We decompose the final color $\mathbf{c}_{\mathrm{final}}^t$ into diffuse and specular components to better distinguish between high and low-frequency information: $\mathbf{c}_{\mathrm{final}}^t = \mathbf{c}_d + \mathbf{c}_s^t$ , where $\mathbf{c_d}$ is the diffuse color (using zero-order spherical harmonics as view-independent color), and $\mathbf{c_s}^t$ is the view-dependent color component. To manage the transition from spherical harmonics to $\mathbf{c}_{\mathrm{final}}^t$ and mitigate potential geometry disruptions, in the early specular stage, we fix the deformable Gaussian MLP and most 3D Gaussian attributes, optimizing only zero-order SH, specular tint, and roughness. We temporarily suspend densification during this phase. As $\mathbf{c}_{\mathrm{final}}^t$ becomes more complete, we gradually resume optimization of all parameters and reinstate the densification process. + +We further split the specular stage into two parts, applying + +a coarse-to-fine strategy to the environment map. In the first part, we focus on optimizing the canonical environment map for time-invariant lighting. This establishes a stable foundation for the overall lighting structure. In the second part, we proceed to optimize the deformable reflection MLP for dynamic elements. This approach ensures a more robust learning process, allowing us to capture the static lighting conditions before introducing the complexities of dynamic components. + +# 3.3. Physical Normal Estimation + +Challenges in normal estimation for 3D Gaussians. Normal estimation is essential for modeling specular objects in 3D Gaussians, where GaussianShader [12] initially used the shortest axis combined with a residual normal for approximation. While this works for static scenes, it becomes problematic with deformed Gaussians because the residual should vary at each time step. A straightforward approach of rotating the residual normal based on quaternion differences between canonical and deformed states proves insufficient, as it does not account for shape changes during deformation. When deformation alters the relative axis lengths, the shortest axis assumption breaks down. This highlights the need for a more comprehensive approach that considers both rotational and shape deformation effects to achieve accurate normal estimation for dynamic specular objects. + +Improved rotation calculation for deformed 3D Gaussians. To overcome the limitations of naive methods and accurately model the normal of deformed 3D Gaussians, we propose using both the shortest and longest axes of canonical and deformed Gaussians to compute the rotation matrix. This approach accounts for both rotation and shape changes during deformation. We first align the deformed Gaussian's axes with those of the canonical Gaussian using the following method: + +$$ +\mathbf {v} _ {s} ^ {t} = \left\{ \begin{array}{l l} \mathbf {v} _ {s} ^ {t} & \text {i f} \mathbf {v} _ {s} \cdot \mathbf {v} _ {s} ^ {t} > 0, \\ - \mathbf {v} _ {s} ^ {t} & \text {o t h e r w i s e .} \end{array} , \quad \mathbf {v} _ {l} ^ {t} = \left\{ \begin{array}{l l} \mathbf {v} _ {l} ^ {t} & \text {i f} \mathbf {v} _ {l} \cdot \mathbf {v} _ {l} ^ {t} > 0, \\ - \mathbf {v} _ {l} ^ {t} & \text {o t h e r w i s e .} \end{array} , \right. \right. \tag {5} +$$ + +where $\mathbf{v}_s$ and $\mathbf{v}_l$ represent the shortest and longest axes of canonical 3D Gaussians, while $\mathbf{v}_s^t$ and $\mathbf{v}_l^t$ denote the same for deformed 3D Gaussians. We then construct orthogonal matrices using these aligned axes and their cross-products: + +$$ +\mathbf {U} = \left[ \begin{array}{l l l} \mathbf {v} _ {s} & \mathbf {v} _ {l} & \mathbf {v} _ {s} \times \mathbf {v} _ {l} \end{array} \right], \quad \mathbf {V} ^ {t} = \left[ \begin{array}{l l l} \mathbf {v} _ {s} ^ {t} & \mathbf {v} _ {l} ^ {t} & \mathbf {v} _ {s} ^ {t} \times \mathbf {v} _ {l} ^ {t} \end{array} \right]. \tag {6} +$$ + +Finally, we derive the rotation matrix $\mathbf{R}^t = \mathbf{V}^t\mathbf{U}^\top$ + +Adjusting normal residuals and ensuring accuracy. To account for shape changes during deformation, we scale the normal residual based on the ratio of oblateness $\frac{\beta}{\beta^t}$ between canonical and deformed 3D Gaussians. + +$$ +\beta = \frac {\left| \mathbf {v} _ {l} \right| - \left| \mathbf {v} _ {s} \right|}{\left| \mathbf {v} _ {l} \right|}, \quad \beta^ {t} = \frac {\left| \mathbf {v} _ {l} ^ {t} \right| - \left| \mathbf {v} _ {s} ^ {t} \right|}{\left| \mathbf {v} _ {l} ^ {t} \right|}, \qquad (7) +$$ + +![](images/7d81696dc4d71bf5a92c447c20b447b27990ec5ce755edc11e1da7a34f623d56.jpg) +Figure 3. Normal estimation. (a) shows that flatter 3D Gaussians align better with scene surfaces, their shortest axis closely matching the surface normal. In contrast, less flat 3D Gaussians fit less accurately, with their shortest axis diverging from the surface normal. (b) shows that when the deformed 3D Gaussian becomes flatter ( $t = t_1$ ), normal residual $\Delta \mathbf{n}$ is rotated by $\mathbf{R}_1^t$ and scaled down by $\frac{\beta}{\beta_1^t}$ , as flatter Gaussians require smaller normal residuals. Conversely, when the deformation results in a less flat shape ( $t = t_2$ ), $\Delta \mathbf{n}$ is rotated by $\mathbf{R}_2^t$ and amplified by $\frac{\beta}{\beta_2^t}$ , requiring a larger correction to align the shortest axis with the surface normal. (c) shows how $\gamma^k$ changes with $w$ (where $w = \frac{|v_s^t|}{|v_l^t|}$ ) for $k = 1, k = 5$ , and $k = 50$ . Larger $w$ indicates less flat Gaussians, while smaller $w$ represents flatter Gaussians. As $k$ increases, $\gamma^k$ decreases more steeply as $w$ rises. For $k = 5$ , we observe a balanced behavior: $\gamma^k$ approaches 1 for low $w$ and 0 for high $w$ , providing a nuanced penalty adjustment across different Gaussian shapes. + +![](images/005bf15805ae5f13c0468f2a75422b9d04cc850c1889841e0d7532cde013b4f5.jpg) + +![](images/da8685942c3cea737a4584c8c01ed14ef155082cd41d1dfe232bab939e1bbb72.jpg) + +where $\beta$ and $\beta^t$ represent the oblateness of canonical and deformed 3D Gaussians, respectively. This is because flatter 3D Gaussians tend to align more closely with the surface, meaning their shortest axis becomes more aligned with the surface normal, as shown in Fig. 3 (a). In such cases, less compensation from the normal residual is needed. Conversely, less flat Gaussians require more compensation, as illustrated in Fig. 3 (b). We then obtain deformed normal residuals: + +$$ +\Delta \mathbf {n} ^ {t} = \frac {\beta}{\beta^ {t}} \mathbf {R} ^ {t} \Delta \mathbf {n}. \tag {8} +$$ + +The final normal $\mathbf{n}^t$ is computed by adding this residual to the shortest axis and ensuring outward orientation: + +$$ +\mathbf {n} ^ {t} = \Delta \mathbf {n} ^ {t} + \mathbf {v} _ {s} ^ {t}, \quad \mathbf {n} ^ {t} = \left\{ \begin{array}{l l} \mathbf {n} ^ {t} & \text {i f} \mathbf {n} ^ {t} \cdot \omega_ {o} ^ {t} > 0, \\ - \mathbf {n} ^ {t} & \text {o t h e r w i s e .} \end{array} \right. \tag {9} +$$ + +This approach adjusts for Gaussian flatness and ensures accurate normal estimation. + +# 3.4. Loss Functions + +Normal regularization. To allow the normal residual to correct the normal while not excessively influencing the optimization of the shortest axis towards the surface normal, we introduce a penalty term for the normal residual: + +$$ +\mathcal {L} _ {\mathrm {r e g}} = \gamma^ {k} \| \Delta \mathbf {n} \| _ {2} ^ {2} \quad \text {w h e r e} \quad \gamma = \sqrt {1 - \frac {\left| \mathbf {v} _ {s} ^ {t} \right| ^ {2}}{\left| \mathbf {v} _ {l} ^ {t} \right| ^ {2}}}. \tag {10} +$$ + +In our experiments, we set $k = 5$ . When $k = 5$ , less flatter 3D Gaussians have $\gamma^k$ approaching 0. Their shortest axis aligns poorly with the surface normal, requiring more normal residual correction and smaller penalties. Conversely, flatter Gaussians have $\gamma^k$ near 1. Their shortest axis aligns better with the surface normal, needing less normal residual + +correction and allowing larger penalties, as shown in Fig. 3 (c). + +Total training loss. To refine all parameters in the dynamic and specular stages, we employ the total training loss: + +$$ +\mathcal {L} = \mathcal {L} _ {\text {c o l o r}} + \lambda_ {\text {n o r m a l}} \mathcal {L} _ {\text {n o r m a l}} + \mathcal {L} _ {\text {r e g}}, \tag {11} +$$ + +where $\mathcal{L}_{\mathrm{color}}$ and $\mathcal{L}_{\mathrm{normal}}$ are obtained as described in Section 3.2.1. In our experiments, we set $\lambda_{\mathrm{normal}} = 0.01$ . Due to space constraints, complete implementation details are provided in the supplementary materials. + +# 4. Experiments + +# 4.1. Evaluation Results + +We evaluate our method on two real-world datasets: NeRF-DS dataset [48] and HyperNeRF dataset [32]. While GaussianShader [12] and GS-IR [23] are originally designed for static scenes and are included here only as reference baselines, we train our method and all baseline approaches for 40,000 iterations to ensure fair comparison. + +NeRF-DS dataset. The NeRF-DS dataset [48] is a monocular video dataset comprising seven real-world scenes from daily life featuring various types of moving or deforming specular objects. We compare our method with the most relevant state-of-the-art approaches. As shown in Tab. 1 and Fig. 4, the quantitative results demonstrate that our method decisively outperforms baselines in reconstructing and rendering real-world highly reflective dynamic specular scenes. + +The rendering speed is correlated with the quantity of 3D Gaussians. When the number of 3D Gaussians is below 178k, our method can achieve real-time rendering over 30 FPS on an NVIDIA RTX 4090. + +HyperNeRF dataset. The HyperNeRF dataset contains real-world dynamic scenes and does not include specular objects. + +Table 1. Quantitative comparison on the NeRF-DS [48] dataset. We report the average PSNR, SSIM, and LPIPS (VGG) of several previous models on test images. The best, the second best, and third best results are denoted by red, orange, yellow. + +
MethodAsBasinBellCup
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
Deformable 3DGS [49]26.040.88050.185019.530.78550.192423.960.79450.276724.490.88220.1658
4DGS [46]24.850.86320.203819.260.76700.219622.860.80150.206123.820.86950.1792
GaussianShader [12]21.890.77390.362017.790.66700.418720.690.81690.302420.400.74370.3385
GS-IR [23]21.580.80330.303318.060.72480.313520.660.78290.260320.340.81930.2719
NeRF-DS [48]25.340.88030.215020.230.80530.250822.570.78110.292124.510.88020.1707
HyperNeRF [32]17.590.85180.239022.580.81560.249719.800.76500.299915.450.82950.2302
Ours26.800.88430.176119.750.79150.189625.460.84900.160024.650.88710.1588
PlatePressSieveMean
MethodPSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
Deformable 3DGS [49]19.070.73520.359925.520.85940.196425.370.86160.164323.430.82840.2201
4DGS [46]18.770.77090.272124.820.83550.225525.160.85660.174522.790.82350.2115
GaussianShader [12]14.550.64230.495519.970.72440.450722.580.78620.305719.700.73630.3819
GS-IR [23]15.980.69690.420022.280.80880.306722.840.82120.223620.250.77960.2999
NeRF-DS [48]19.700.78130.297425.350.87030.255224.990.87050.200123.240.83840.2402
HyperNeRF [32]21.220.78290.316616.540.82000.281019.920.85210.214219.010.81670.2615
Ours20.840.81720.219826.490.86570.188925.220.87050.151324.170.85220.1778
+ +![](images/5c94518554cd7ab1cae7b8cc2db0a1dec111098f4882477ab2630a9bae46c691.jpg) +Figure 4. Qualitative comparison on the NeRF-DS [48] dataset. + +Table 2. Quantitative comparison on the HyperNeRF [32] dataset. We report the average PSNR, SSIM, and LPIPS (VGG) of several previous models. The best, the second best, and third best results are denoted by red, orange, yellow. + +
MethodPSNR ↑SSIM ↑LPIPS ↓
Deformable 3DGS [49]22.780.62010.3380
4DGS [46]24.890.67810.3422
GaussianShader [12]18.550.54520.4795
GS-IR [23]19.870.57290.4498
NeRF-DS [48]23.650.64050.3972
HyperNeRF [32]23.110.63870.3968
Ours22.220.60880.3295
+ +As shown in Tab. 2 and Fig. 5, the results demonstrate that we are on par with state-of-the-art techniques for rendering novel views, and our method's performance is not limited to + +shiny scenes. + +In Fig. 6, we compare our method's normal maps with those from Deformable 3DGS [49] and NeRF-DS [48]. For Deformable 3DGS [49], we obtain the normals by using the shortest axes of the deformed 3D Gaussians. As demonstrated, our method produces significantly better quality normal maps compared to Deformable 3DGS [49] and NeRF-DS [48]. + +# 4.2. Ablation Study + +For a fair comparison, we train our method and all ablation experiments for 40,000 iterations. + +Different coarse to fine training strategy stages. As shown in Tab. 3 and Fig. 7, each stage contributes effectively to the model's performance. The Dynamic stage enhances dynamic object stability compared to the Static stage alone, while the Specular stage improves reflection clarity beyond the combined Static and Dynamic stages. + +![](images/a9e9bf5cd13deac75326c85d20ebd03f34c97807ec784fbb025581469de96a27.jpg) +Figure 5. Qualitative comparison on the HyperNeRF [32] dataset. + +![](images/e6e0a71cdf767f7ab03aa7f2bce2dc89b6f74a93fbdf58beb32aae0c2aedc0e3.jpg) +Figure 6. Qualitative comparison of normal maps between our method, Deformable 3DGS, and NeRF-DS. + +Table 3. Ablation studies on different coarse to fine training strategy stages. + +
StagePSNR ↑SSIM ↑LPIPS ↓
Static20.260.77850.2953
St. + Dynamic24.020.85080.1831
St. + Dy. + Specular24.170.85220.1778
+ +![](images/ef87ff0f4d145e3cb09e040c7177b4a7489c7e58d8f38394243b6eac9f4e6bf5.jpg) +Figure 7. Qualitative comparison of each training stage in our coarse-to-fine approach. + +Ablation study on coarse-to-fine, and loss function. The model's performance was evaluated without key components: the coarse-to-fine training strategy, normal loss $\mathcal{L}_{\mathrm{normal}}$ , normal regularization $\mathcal{L}_{\mathrm{reg}}$ , and $\gamma^k$ . Fig. 8 and Tab. 4 illustrate the effects of these omissions. Without the coarse-to-fine + +Table 4. Ablation studies on different coarse to fine training strategy stages. + +
C2F\( \mathcal{L}_{\text{normal}} \)\( \mathcal{L}_{\text{reg}} \)\( \gamma^k \)PSNR↑SSIM↑LPIPS↓
✓✓✓23.160.82940.2156
✓23.400.82770.2278
✓✓24.150.85100.1845
✓✓✓24.090.85150.1818
✓✓✓✓24.170.85220.1778
+ +Table 5. Ablation studies on SH, Static and Deformable environment map. + +
PSNR ↑SSIM ↑LPIPS ↓
SH23.630.84530.1844
Static Env. map24.020.85080.1831
Deformable Env. map24.170.85220.1778
+ +approach, the model, trained directly at the specular stage, produces incomplete scene geometry, affecting environment map queries for specular color. Omitting normal loss $\mathcal{L}_{\mathrm{normal}}$ removes direct supervision on normals and leads to inaccurate reflection directions and less precise specular colors. Removing normal regularization $\mathcal{L}_{\mathrm{reg}}$ allows the normal residual to dominate normal optimization, resulting in insufficient optimization of the 3D Gaussians' shortest axis towards the correct normal, which in turn reduces the rendering quality. The normal residual decreases for non-flattened and flat Gaussians without $\gamma^k$ in normal regularization. This particularly affects less flat 3D Gaussians whose shortest axis significantly deviates from the surface normal. The insufficient normal residual correction causes these 3D Gaussians' shortest axes to deviate greatly from their original direction in an attempt to align with the surface normal, ultimately reducing rendering quality. + +Ablation study on SH, Static environment map, and Deformable environment map. Fig. 9 and Tab. 5 demonstrate the superiority of the deformable environment map over the static environment map, which in turn outperforms Spherical Harmonics (SH). SH struggles to accurately model high- + +![](images/cffa3e1f530ca2a1e31da73d46e0acb77bd2e7b60976bd39a99347eeabe638b9.jpg) +Full model + +![](images/95c082d7f288a73c6d4f1f2c91f5181be473758f935c5d81a79fa8cda99a60c1.jpg) +Full model + +![](images/5da24300c8e3a74f08660cd5651e4bd6c9ee6b60318adcc5dc42f60eb366ba51.jpg) +Full model + +![](images/2eca99811ece7f867e60b06ed6fe26e77cefe9dd0b6134a391bdce9bedaaaeb6.jpg) +Full model + +![](images/8ff991c9f481437702c238f47b10cb7a4b6a3056e370040673254d41ce57184f.jpg) +w/o Coarse-to-fine + +![](images/2b453a3c19c78fc582a06ef7f0a0929f6db17565e8f91274536a8a6851fa0864.jpg) +w/o $\mathcal{L}_{\mathrm{normal}}$ + +![](images/eec227e92ea0cabeefb69d6c189faa5ec1094d4d7bf84f8574636a5c2e535ec8.jpg) +w/o $\mathcal{L}_{\mathrm{reg}}$ + +![](images/1c8e2e5a51f8e8c9bb1864b1608aa6f82369ce52f10be35973f72ef3cf64c6ea.jpg) +w/o $\gamma^k$ + +![](images/d4f16fca7dd438c0cf659a75042f2c5d9301b0323856ac744642bf07a4faddfb.jpg) +Figure 8. Qualitative comparison of ablation study without different components. + +![](images/be59e70aff98002103d86392026d30d7d08df8392e828199988f14bac915566c.jpg) +Static Env. map + +![](images/e5af51a886f5b77d2ec9e5c3e3fd2e2e64e46d29bb37ec71043c89146b15949e.jpg) +Spherical Harmonics (SH) +Deformable Env. map (Full model) +Figure 9. Qualitative comparison of ablation study on SH, Static environment map, and Deformable environment map. + +![](images/109096eadb039a1b5aa84b90b20f94196bd493e4927be2ad8303ac10b9b55b1d.jpg) +Ground-truth + +Table 6. Ablation studies on 2DGS and without Physical Normal Estimation. + +
PSNR ↑SSIM ↑LPIPS ↓
2DGS [10]23.220.82190.2283
w/o N.E.23.890.84900.1837
Full model24.170.85220.1778
+ +frequency specular colors. While the static environment map can model high-frequency colors, it is best suited for static lighting conditions. In contrast, the deformable environment map models time-varying lighting, offering superior performance for dynamic scenes. + +Ablation study on 2DGS [10] and without Physical Normal Estimation. In Fig. 10 and Tab. 6, "2DGS" represents replacing our 3D Gaussian with 2D Gaussian representation. Since 2DGS inherently includes normals, we omit physical normal estimation. "w/o N.E." means skipping physical nor + +![](images/f33d20c3f784bfbe7f0ff3bb9781853c0752766e462530489b8bd01d0fee52ac.jpg) +Full model + +![](images/ccb980a7b0c4ceb91f053141c76fe75402e5d4140022965b63d68daeb8f5ab6d.jpg) +Full model + +![](images/cedd9e62c8c0dcc98d5badf8505cda37ed6e25f28bb5f30bef17e0e11f2e4ae2.jpg) +2DGS +Figure 10. Qualitative comparison of ablation study on 2DGS and without Physical Normal Estimation. + +![](images/2068735c5657228fe4714a273f581b1856a11fbe8005e1a250b0ef63e614d843.jpg) +w/o N.E. + +mal estimation and using the shortest axis of 3D Gaussians as the normal. This causes the normal loss $\mathcal{L}_{\mathrm{normal}}$ to directly supervise the shortest axes, making some axes deviate significantly to align with surface normals, resulting in degraded rendering quality. + +# 5. Conclusion + +SpectroMotion enhances 3D Gaussian Splatting for dynamic specular scenes by combining specular rendering with deformation fields. Using normal residual correction, coarse-to-fine training, and a deformable environment map, it achieves superior accuracy and visual quality in novel view synthesis, outperforming existing methods while maintaining geometric consistency. + +Limitations. Though we stabilize the entire scene's geometry using a coarse-to-fine training strategy, some failure cases still occur. Please refer to the supplementary materials for visual results of failure cases. + +Acknowledgements. This research was funded by the National Science and Technology Council, Taiwan, under Grants NSTC 112-2222-E-A49-004-MY2 and 113-2628-E-A49-023-. The authors are grateful to Google, NVIDIA, and MediaTek Inc. for their generous donations. Yu-Lun Liu acknowledges the Yushan Young Fellow Program by the MOE in Taiwan. + +# References + +[1] Benjamin Attal, Eliot Laidlaw, Aaron Gokaslan, Changil Kim, Christian Richardt, James Tompkin, and Matthew O'Toole. Törf: Time-of-flight radiance fields for dynamic scene view synthesis, 2021. 2 +[2] Sai Bi, Zexiang Xu, Pratul Srinivasan, Ben Mildenhall, Kalyan Sunkavalli, Miloš Hasan, Yannick Hold-Geoffroy, David Kriegman, and Ravi Ramamoorthi. Neural reflectance fields for appearance acquisition, 2020. 2 +[3] Mark Boss, Varun Jampani, Raphael Braun, Ce Liu, Jonathan T. Barron, and Hendrik P. A. Lensch. Neural-pil: Neural pre-integrated lighting for reflectance decomposition, 2021. 2 +[4] Ting-Hsuan Chen, Jie Wen Chan, Hau-Shiang Shiu, Shih-Han Yen, Changhan Yeh, and Yu-Lun Liu. Narcan: Natural refined canonical image with integration of diffusion prior for video editing. In NeurIPS, 2024. 2 +[5] Robert L Cook and Kenneth E. Torrance. A reflectance model for computer graphics. ACM Transactions on Graphics (ToG), 1(1):7-24, 1982. 3 +[6] Ainaz Eftekhar, Alexander Sax, Jitendra Malik, and Amir Zamir. Omnidata: A scalable pipeline for making multi-task mid-level vision datasets from 3d scans. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10786-10796, 2021. 1 +[7] Jian Gao, Chun Gu, Youtian Lin, Hao Zhu, Xun Cao, Li Zhang, and Yao Yao. Relightable 3d gaussian: Real-time point cloud relighting with brdf decomposition and ray tracing. arXiv preprint arXiv:2311.16043, 2023. 2 +[8] Wenhang Ge, Tao Hu, Haoyu Zhao, Shu Liu, and Ying-Cong Chen. Ref-neus: Ambiguity-reduced neural implicit surface learning for multi-view reconstruction with reflection, 2023. 2 +[9] Xiang Guo, Jiadai Sun, Yuchao Dai, Guanying Chen, Xiaoqing Ye, Xiao Tan, Errui Ding, Yumeng Zhang, and Jingdong Wang. Forward flow for novel view synthesis of dynamic scenes, 2023. 2 +[10] Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2d gaussian splatting for geometrically accurate radiance fields. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 8 +[11] Yi-Hua Huang, Yang-Tian Sun, Ziyi Yang, Xiaoyang Lyu, Yan-Pei Cao, and Xiaojuan Qi. Sc-gs: Sparse-controlled gaussian splatting for editable dynamic scenes, 2024. 2 +[12] Yingwenqi Jiang, Jiadong Tu, Yuan Liu, Xifeng Gao, Xiaoxiao Long, Wenping Wang, and Yuexin Ma. Gaussianshader: 3d gaussian splatting with shading functions for reflective surfaces. arXiv preprint arXiv:2311.17977, 2023. 2, 4, 5, 6 + +[13] Yingwenqi Jiang, Jiadong Tu, Yuan Liu, Xifeng Gao, Xiaoxiao Long, Wenping Wang, and Yuexin Ma. Gaussianshader: 3d gaussian splatting with shading functions for reflective surfaces. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5322-5332, 2024. 3 +[14] James T Kajiya. The rendering equation. In Proceedings of the 13th annual conference on Computer graphics and interactive techniques, pages 143-150, 1986. 3 +[15] Brian Karis and Epic Games. Real shading in unreal engine 4. Proc. Physically Based Shading Theory Practice, 4(3):1, 2013. 3 +[16] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering, 2023. 1, 2 +[17] Junxuan Li and Hongdong Li. Neural reflectance for shape recovery with shadow handling, 2022. 2 +[18] Tianye Li, Mira Slavcheva, Michael Zollhoefer, Simon Green, Christoph Lassner, Changil Kim, Tanner Schmidt, Steven Lovegrove, Michael Goesele, Richard Newcombe, and Zhaoyang Lv. Neural 3d video synthesis from multi-view video, 2022. 2 +[19] Zhengqi Li, Simon Niklaus, Noah Snively, and Oliver Wang. Neural scene flow fields for space-time view synthesis of dynamic scenes, 2021. 2 +[20] Ruofan Liang, Huiting Chen, Chunlin Li, Fan Chen, Selvakumar Panneer, and Nandita Vijaykumar. Envidr: Implicit differentiable renderer with neural environment lighting, 2023. 2 +[21] Ruofan Liang, Jiahao Zhang, Haoda Li, Chen Yang, Yushi Guan, and Nandita Vijaykumar. Spidr: Sdf-based neural point fields for illumination and deformation, 2023. 2 +[22] Yiqing Liang, Numair Khan, Zhengqin Li, Thu Nguyen-Phuoc, Douglas Lanman, James Tompkin, and Lei Xiao. Gaufre: Gaussian deformation fields for real-time dynamic novel view synthesis, 2023. 2 +[23] Zhihao Liang, Qi Zhang, Ying Feng, Ying Shan, and Kui Jia. Gs-ir: 3d gaussian splatting for inverse rendering. arXiv preprint arXiv:2311.16473, 2023. 2, 5, 6 +[24] Haotong Lin, Sida Peng, Zhen Xu, Yunzhi Yan, Qing Shuai, Hujun Bao, and Xiaowei Zhou. Efficient neural radiance fields for interactive free-viewpoint video, 2022. 2 +[25] Yuan Liu, Peng Wang, Cheng Lin, Xiaoxiao Long, Jiepeng Wang, Lingjie Liu, Taku Komura, and Wenping Wang. Nano: Neural geometry and brdf reconstruction of reflective objects from multiview images, 2023. 2 +[26] Yu-Lun Liu, Chen Gao, Andreas Meuleman, Hung-Yu Tseng, Ayush Saraf, Changil Kim, Yung-Yu Chuang, Johannes Kopf, and Jia-Bin Huang. Robust dynamic radiance fields. In CVPR, 2023. 2 +[27] Caoyuan Ma, Yu-Lun Liu, Zhixiang Wang, Wu Liu, Xinchen Liu, and Zheng Wang. Humannerf-se: A simple yet effective approach to animate humannerf with diverse poses. In CVPR, 2024. 2 +[28] Li Ma, Vasu Agrawal, Haithem Turki, Changil Kim, Chen Gao, Pedro Sander, Michael Zollhöfer, and Christian Richardt. Specnerf: Gaussian directional encoding for specular reflections. arXiv preprint arXiv:2312.13102, 2023. 2 + +[29] Marko Mihajlovic, Sergey Prokudin, Siyu Tang, Robert Maier, Federica Bogo, Tony Tung, and Edmond Boyer. Splatfields: Neural gaussian splats for sparse 3d and 4d reconstruction. arXiv preprint arXiv:2409.11211, 2024. 2 +[30] Jacob Munkberg, Jon Hasselgren, Tianchang Shen, Jun Gao, Wenzheng Chen, Alex Evans, Thomas Müller, and Sanja Fidler. Extracting Triangular 3D Models, Materials, and Lighting From Images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8280-8290, 2022. 2, 3 +[31] Keunhong Park, Utkarsh Sinha, Jonathan T. Barron, Sofien Bouaziz, Dan B Goldman, Steven M. Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields, 2021. 2 +[32] Keunhong Park, Utkarsh Sinha, Peter Hedman, Jonathan T. Barron, Sofien Bouaziz, Dan B Goldman, Ricardo MartinBrualla, and Steven M. Seitz. Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields, 2021. 2, 5, 6, 7 +[33] Sida Peng, Yunzhi Yan, Qing Shuai, Hujun Bao, and Xiaowei Zhou. Representing volumetric videos as dynamic mlp maps, 2023. 2 +[34] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-nerf: Neural radiance fields for dynamic scenes, 2020. 2 +[35] Yahao Shi, Yanmin Wu, Chenming Wu, Xing Liu, Chen Zhao, Haocheng Feng, Jingtuo Liu, Liangjun Zhang, Jian Zhang, Bin Zhou, et al. Gir: 3d gaussian inverse rendering for relightable scene factorization. arXiv preprint arXiv:2312.05133, 2023. 2 +[36] Liangchen Song, Anpei Chen, Zhong Li, Zhang Chen, Lele Chen, Junsong Yuan, Yi Xu, and Andreas Geiger. Nerfplayer: A streamable dynamic scene representation with decomposed neural radiance fields, 2023. 2 +[37] Pratul P. Srinivasan, Boyang Deng, Xiuming Zhang, Matthew Tancik, Ben Mildenhall, and Jonathan T. Barron. Nerv: Neural reflectance and visibility fields for relighting and view synthesis, 2020. 2 +[38] Colton Stearns, Adam Harley, Mikaela Uy, Florian Dubost, Federico Tombari, Gordon Wetzstein, and Leonidas Guibas. Dynamic gaussian marbles for novel view synthesis of casual monocular videos. arXiv preprint arXiv:2406.18717, 2024. 2 +[39] Edgar Tretschk, Ayush Tewari, Vladislav Golyanik, Michael Zollhöfer, Christoph Lassner, and Christian Theobalt. Non-rigid neural radiance fields: Reconstruction and novel view synthesis of a dynamic scene from monocular video, 2021. 2 +[40] Dor Verbin, Peter Hedman, Ben Mildenhall, Todd Zickler, Jonathan T Barron, and Pratul P Srinivasan. Ref-nerf: Structured view-dependent appearance for neural radiance fields. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5481–5490. IEEE, 2022. 2 +[41] Dor Verbin, Ben Mildenhall, Peter Hedman, Jonathan T Barron, Todd Zickler, and Pratul P Srinivasan. Eclipse: Disambiguating illumination and materials using unintended shadows. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 77-86, 2024. 2 + +[42] Dor Verbin, Pratul P Srinivasan, Peter Hedman, Ben Mildenhall, Benjamin Attal, Richard Szeliski, and Jonathan T Barron. Nerf-casting: Improved view-dependent appearance with consistent reflections. arXiv preprint arXiv:2405.14871, 2024. 2 +[43] Bruce Walter, Stephen R Marschner, Hongsong Li, and Kenneth E Torrance. Microfacet models for refraction through rough surfaces. Rendering techniques, 2007:18th, 2007. 3 +[44] Qianqian Wang, Vickie Ye, Hang Gao, Jake Austin, Zhengqi Li, and Angjoo Kanazawa. Shape of motion: 4d reconstruction from a single video. arXiv preprint arXiv:2407.13764, 2024. 2 +[45] Chun-Hung Wu, Shih-Hong Chen, Chih-Yao Hu, Hsin-Yu Wu, Kai-Hsin Chen, Yu-You Chen, Chih-Hai Su, Chih-Kuo Lee, and Yu-Lun Liu. Denver: Deformable neural vessel representations for unsupervised video vessel segmentation. In CVPR, 2025. 2 +[46] Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Xinggang Wang. 4d gaussian splatting for real-time dynamic scene rendering. arXiv preprint arXiv:2310.08528, 2023. 2, 6 +[47] Wenqi Xian, Jia-Bin Huang, Johannes Kopf, and Changil Kim. Space-time neural irradiance fields for free-viewpoint video, 2021. 2 +[48] Zhiwen Yan, Chen Li, and Gim Hee Lee. Nerf-ds: Neural radiance fields for dynamic specular objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8285-8295, 2023. 5, 6 +[49] Ziyi Yang, Xinyu Gao, Wen Zhou, Shaohui Jiao, Yuqing Zhang, and Xiaogang Jin. Deformable 3d gaussians for high-fidelity monocular dynamic scene reconstruction. arXiv preprint arXiv:2309.13101, 2023. 2, 3, 6 +[50] Ziyi Yang, Xinyu Gao, Yangtian Sun, Yihua Huang, Xiaoyang Lyu, Wen Zhou, Shaohui Jiao, Xiaojuan Qi, and Xiaogang Jin. Spec-gaussian: Anisotropic view-dependent appearance for 3d gaussian splatting, 2024. 2 +[51] Keyang Ye, Qiming Hou, and Kun Zhou. 3d gaussian splatting with deferred reflection. In ACM SIGGRAPH 2024 Conference Papers, pages 1-10, 2024. 2 +[52] Wang Yifan, Felice Serena, Shihao Wu, Cengiz Oztireli, and Olga Sorkine-Hornung. Differentiable surface splatting for point-based geometry processing. ACM Transactions on Graphics (TOG), 38(6):1-14, 2019. 2 +[53] Jingyang Zhang, Yao Yao, Shiwei Li, Jingbo Liu, Tian Fang, David McKinnon, Yanghai Tsin, and Long Quan. Neilf++: Inter-reflectable light fields for geometry and material estimation, 2023. 2 +[54] Kai Zhang, Fujun Luan, Qianqian Wang, Kavita Bala, and Noah Snavely. Physg: Inverse rendering with spherical gaussians for physics-based material editing and relighting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5453-5462, 2021. 2 +[55] Xiuming Zhang, Pratul P. Srinivasan, Boyang Deng, Paul Debevec, William T. Freeman, and Jonathan T. Barron. Ner-factor: neural factorization of shape and reflectance under an unknown illumination. ACM Transactions on Graphics, 40 (6):1-18, 2021. + +[56] Xiaoming Zhao, Pratul P Srinivasan, Dor Verbin, Keunhong Park, Ricardo Martin Brualla, and Philipp Henzler. Illuminerf: 3d relighting without inverse rendering. arXiv preprint arXiv:2406.06527, 2024. 2 +[57] Zuo-Liang Zhu, Beibei Wang, and Jian Yang. Gs-ror: 3d gaussian splatting for reflective object relighting via sdf priors. arXiv preprint arXiv:2406.18544, 2024. 2 \ No newline at end of file diff --git a/2025/SpectroMotion_ Dynamic 3D Reconstruction of Specular Scenes/images.zip b/2025/SpectroMotion_ Dynamic 3D Reconstruction of Specular Scenes/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..ba972b1409908cb7204721e4e57eaccd5ab9b037 --- /dev/null +++ b/2025/SpectroMotion_ Dynamic 3D Reconstruction of Specular Scenes/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:16afda24576349146eeca1e845d8b272ce5d4e3df5f31768c5bee198f6274c8b +size 964911 diff --git a/2025/SpectroMotion_ Dynamic 3D Reconstruction of Specular Scenes/layout.json b/2025/SpectroMotion_ Dynamic 3D Reconstruction of Specular Scenes/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..3abf2f4d08ad0fa5b5ab46c5aefa2a6980d64a34 --- /dev/null +++ b/2025/SpectroMotion_ Dynamic 3D Reconstruction of Specular Scenes/layout.json @@ -0,0 +1,10580 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 108, + 103, + 501, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 103, + 501, + 121 + ], + "spans": [ + { + "bbox": [ + 108, + 103, + 501, + 121 + ], + "type": "text", + "content": "SpectroMotion: Dynamic 3D Reconstruction of Specular Scenes" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 140, + 142, + 484, + 173 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 142, + 484, + 173 + ], + "spans": [ + { + "bbox": [ + 140, + 142, + 484, + 173 + ], + "type": "text", + "content": "Cheng-De Fan1 Chen-Wei Chang1 Yi-Ruei Liu1,2 Jie-Ying Lee1 \nJiun-Long Huang1 Yu-Chee Tseng1 Yu-Lun Liu1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 94, + 179, + 527, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 179, + 527, + 194 + ], + "spans": [ + { + "bbox": [ + 94, + 179, + 527, + 194 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 94, + 179, + 527, + 194 + ], + "type": "text", + "content": "National Yang Ming Chiao Tung University " + }, + { + "bbox": [ + 94, + 179, + 527, + 194 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 94, + 179, + 527, + 194 + ], + "type": "text", + "content": "University of Illinois Urbana-Champaign" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 221, + 200, + 234 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 221, + 200, + 234 + ], + "spans": [ + { + "bbox": [ + 151, + 221, + 200, + 234 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 248, + 297, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 248, + 297, + 452 + ], + "spans": [ + { + "bbox": [ + 55, + 248, + 297, + 452 + ], + "type": "text", + "content": "We present SpectroMotion, a novel approach that combines 3D Gaussian Splatting (3DGS) with physically-based rendering (PBR) and deformation fields to reconstruct dynamic specular scenes. Previous methods extending 3DGS to model dynamic scenes have struggled to represent specular surfaces accurately. Our method addresses this limitation by introducing a residual correction technique for accurate surface normal computation during deformation, complemented by a deformable environment map that adapts to time-varying lighting conditions. We implement a coarse-to-fine training strategy significantly enhancing scene geometry and specular color prediction. It is the only existing 3DGS method capable of synthesizing photorealistic real-world dynamic specular scenes, outperforming state-of-the-art methods in rendering complex, dynamic, and specular scenes. Please see our project page at cdfan0627.github.io/spectromotion." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 476, + 135, + 488 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 476, + 135, + 488 + ], + "spans": [ + { + "bbox": [ + 55, + 476, + 135, + 488 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 496, + 296, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 496, + 296, + 604 + ], + "spans": [ + { + "bbox": [ + 55, + 496, + 296, + 604 + ], + "type": "text", + "content": "3D Gaussian Splitting (3DGS) [16] has emerged as a groundbreaking technique in 3D scene reconstruction, offering fast training and real-time rendering capabilities. By representing 3D scene using a collection of 3D Gaussians and employing a point-based rendering approach, 3DGS has significantly improved efficiency in novel view synthesis. However, extending 3DGS to model dynamic scenes, especially those containing specular surfaces accurately, has remained a significant challenge." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 605, + 296, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 296, + 700 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 296, + 700 + ], + "type": "text", + "content": "Existing extensions of 3DGS have made progress in either dynamic scene reconstruction or specular object rendering, but none have successfully combined both aspects. Methods tackling dynamic scenes often struggle with accurately representing specular surfaces, while those focusing on specular rendering are limited to static scenes. This capability gap has hindered the application of 3DGS to real-world scenarios where both motion and specular reflections are present." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 67, + 701, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 701, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 701, + 296, + 713 + ], + "type": "text", + "content": "We present SpectroMotion, a novel approach that ad" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 317, + 222, + 553, + 399 + ], + "blocks": [ + { + "bbox": [ + 317, + 222, + 553, + 399 + ], + "lines": [ + { + "bbox": [ + 317, + 222, + 553, + 399 + ], + "spans": [ + { + "bbox": [ + 317, + 222, + 553, + 399 + ], + "type": "image", + "image_path": "c78dbd9e6a09ad3b87ada87fe470b475a575a7b82e7cb685a9ec4005fe669403.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 410, + 555, + 531 + ], + "lines": [ + { + "bbox": [ + 313, + 410, + 555, + 531 + ], + "spans": [ + { + "bbox": [ + 313, + 410, + 555, + 531 + ], + "type": "text", + "content": "Figure 1. Our method, SpectroMotion, recovers and renders dynamic scenes with higher-quality reflections compared to prior work. It introduces physical normal estimation, deformable environment maps, and a coarse-to-fine training strategy to achieve superior results in rendering dynamic scenes with reflections. Here, we present a rendered test image, corresponding normal maps, and a ground-truth image, where the ground-truth normal map (used as a reference) is generated using a pre-trained normal estimator [6]. For Deformable 3DGS, we use the shortest axes of the deformed 3D Gaussians as the normals. We have highlighted the specular regions to demonstrate the effectiveness of our approach." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 554, + 555, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 554, + 555, + 650 + ], + "spans": [ + { + "bbox": [ + 313, + 554, + 555, + 650 + ], + "type": "text", + "content": "dresses these limitations by combining 3D Gaussian Splating with physically based rendering (PBR) and deformation fields. Our method introduces three key innovations: a residual correction technique for accurate surface normal computation during deformation, a deformable environment map that adapts to time-varying lighting conditions, and a coarse-to-fine training strategy that significantly enhances scene geometry and specular color prediction." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "text", + "content": "Our evaluations demonstrate that SpectroMotion outperforms prior methods in view synthesis of scenes containing dynamic specular objects, as illustrated in Fig. 1. It is the only 3DGS method capable of synthesizing photorealistic real-world dynamic specular scenes, surpassing state-of-the" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 495, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 495, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 495, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21328" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "text", + "content": "art techniques in rendering complex, dynamic, and specular content. This advancement represents a significant leap in 3D scene reconstruction, particularly for challenging scenarios involving moving specular objects." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 121, + 272, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 121, + 272, + 133 + ], + "spans": [ + { + "bbox": [ + 67, + 121, + 272, + 133 + ], + "type": "text", + "content": "In summary, we make the following contributions:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 133, + 296, + 265 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 55, + 133, + 295, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 133, + 295, + 180 + ], + "spans": [ + { + "bbox": [ + 55, + 133, + 295, + 180 + ], + "type": "text", + "content": "- We propose SpectroMotion, a physically-based rendering (PBR) approach combining deformation fields and 3D Gaussian Splatting for real-world dynamic specular scenes." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 181, + 296, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 181, + 296, + 228 + ], + "spans": [ + { + "bbox": [ + 55, + 181, + 296, + 228 + ], + "type": "text", + "content": "- We introduce a residual correction method for accurate surface normals during deformation, coupled with a deformable environment map to handle time-varying lighting conditions in dynamic scenes." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 229, + 296, + 265 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 229, + 296, + 265 + ], + "spans": [ + { + "bbox": [ + 55, + 229, + 296, + 265 + ], + "type": "text", + "content": "- We develop a coarse-to-fine training strategy enhancing scene geometry and specular color prediction, outperforming state-of-the-art methods." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 277, + 141, + 289 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 277, + 141, + 289 + ], + "spans": [ + { + "bbox": [ + 55, + 277, + 141, + 289 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 297, + 223, + 309 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 297, + 223, + 309 + ], + "spans": [ + { + "bbox": [ + 55, + 297, + 223, + 309 + ], + "type": "text", + "content": "2.1. Dynamic Scene Reconstruction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 315, + 296, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 315, + 296, + 555 + ], + "spans": [ + { + "bbox": [ + 55, + 315, + 296, + 555 + ], + "type": "text", + "content": "Recent works have leveraged NeRF representations to jointly solve for canonical space and deformation fields in dynamic scenes using RGB supervision [4, 9, 19, 26, 27, 31, 32, 34, 39, 45, 47]. Further advancements in dynamic neural rendering include object segmentation [36], incorporation of depth information [1], utilization of 2D CNNs for scene priors [24, 33], and multi-view video compression [18]. However, these NeRF-based methods are computationally intensive, limiting their practical applications. To address this, 3D Gaussian Splatting [16] has emerged as a promising alternative, offering real-time rendering capabilities while maintaining high visual quality. Building upon this efficient representation, recent research has adapted 3D Gaussians for dynamic scenes [11, 22, 29, 38, 44, 46, 49]. Nevertheless, these approaches do not explicitly account for changes in surface normal during the dynamic process. Our work extends this line of research by combining specular object rendering based on normal estimation with a deformation field, enabling each 3D Gaussian to model dynamic specular scenes effectively." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 563, + 211, + 577 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 563, + 211, + 577 + ], + "spans": [ + { + "bbox": [ + 55, + 563, + 211, + 577 + ], + "type": "text", + "content": "2.2. Reflective Object Rendering" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 582, + 297, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 297, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 297, + 715 + ], + "type": "text", + "content": "While significant progress has been made in rendering reflective objects, challenges from complex light interactions persist. Recent years have seen numerous studies addressing these issues, primarily by decomposing appearance into lighting and material properties [2, 3, 17, 30, 37, 41, 54-56]. Building on this foundation, some research has focused on improving the capture and reproduction of specular reflections [28, 40, 42, 51]. In contrast, others have leveraged signed distance functions (SDFs) to enhance normal estimation [8, 20, 21, 25, 53]. The emergence of 3D Gaussian Splatting (3DGS) has sparked a new wave of techniques" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 72, + 556, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 556, + 168 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 556, + 168 + ], + "type": "text", + "content": "[7, 12, 23, 35, 50, 57] that integrate Gaussian splating with physically-based rendering. Nevertheless, accurately modeling dynamic environments and time-varying specular reflections remains a significant challenge. To address this limitation, our work introduces a novel approach incorporating a deformable environment map and additional explicit Gaussian attributes specifically designed to capture specular color changes over time." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 179, + 370, + 192 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 179, + 370, + 192 + ], + "spans": [ + { + "bbox": [ + 314, + 179, + 370, + 192 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 199, + 556, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 199, + 556, + 308 + ], + "spans": [ + { + "bbox": [ + 313, + 199, + 556, + 308 + ], + "type": "text", + "content": "Overview of the approach. The overview of our method is illustrated in Fig. 2. Given an input monocular video sequence of frames and corresponding camera poses, we design a three-stage approach to reconstruct the dynamic specular scene, as detailed in Sec. 3.2. Accurate specular reflection requires precise normal estimation, so Sec. 3.3 elaborates on how we estimate normals in dynamic scenes. Finally, we introduce the losses used throughout the training process in Sec. 3.4." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 314, + 316, + 393, + 328 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 316, + 393, + 328 + ], + "spans": [ + { + "bbox": [ + 314, + 316, + 393, + 328 + ], + "type": "text", + "content": "3.1. Preliminary" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 335, + 555, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 335, + 555, + 408 + ], + "spans": [ + { + "bbox": [ + 313, + 335, + 555, + 408 + ], + "type": "text", + "content": "3D Gaussian Splatting. Each 3D Gaussian is defined by a center position " + }, + { + "bbox": [ + 313, + 335, + 555, + 408 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 313, + 335, + 555, + 408 + ], + "type": "text", + "content": " and a covariance matrix " + }, + { + "bbox": [ + 313, + 335, + 555, + 408 + ], + "type": "inline_equation", + "content": "\\Sigma" + }, + { + "bbox": [ + 313, + 335, + 555, + 408 + ], + "type": "text", + "content": ". 3D Gaussian Splatting [16] optimizes the covariance matrix using scaling factors " + }, + { + "bbox": [ + 313, + 335, + 555, + 408 + ], + "type": "inline_equation", + "content": "s \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 313, + 335, + 555, + 408 + ], + "type": "text", + "content": " and rotation unit quaternion " + }, + { + "bbox": [ + 313, + 335, + 555, + 408 + ], + "type": "inline_equation", + "content": "r \\in \\mathbb{R}^4" + }, + { + "bbox": [ + 313, + 335, + 555, + 408 + ], + "type": "text", + "content": ". For novel-view rendering, 3D Gaussians are projected onto 2D camera planes using differentiable splatting [52]:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 393, + 416, + 555, + 430 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 393, + 416, + 555, + 430 + ], + "spans": [ + { + "bbox": [ + 393, + 416, + 555, + 430 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\Sigma} ^ {\\prime} = \\mathbf {J} \\mathbf {W} \\boldsymbol {\\Sigma} \\mathbf {W} ^ {T} \\mathbf {J} ^ {T}. \\tag {1}", + "image_path": "617f55b752856c994fad9e18fce13f546b0123210e9d1ac250965939f1401024.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 440, + 556, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 440, + 556, + 464 + ], + "spans": [ + { + "bbox": [ + 313, + 440, + 556, + 464 + ], + "type": "text", + "content": "Pixel colors are computed using point-based volumetric rendering:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 347, + 472, + 555, + 499 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 472, + 555, + 499 + ], + "spans": [ + { + "bbox": [ + 347, + 472, + 555, + 499 + ], + "type": "interline_equation", + "content": "C = \\sum_ {i \\in N} T _ {i} \\alpha_ {i} c _ {i}, \\quad \\alpha_ {i} = \\sigma_ {i} e ^ {- \\frac {1}{2} (\\boldsymbol {x}) ^ {T} \\boldsymbol {\\Sigma} ^ {\\prime} (\\boldsymbol {x})}, \\tag {2}", + "image_path": "02426b23813e4c566bab6442a5a3bcc59f405fc631b5d437285c04fb25cb305b.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 510, + 555, + 535 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 555, + 535 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 555, + 535 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 510, + 555, + 535 + ], + "type": "inline_equation", + "content": "T_{i} = \\prod_{j=1}^{i-1}(1 - \\alpha_{j})" + }, + { + "bbox": [ + 313, + 510, + 555, + 535 + ], + "type": "text", + "content": " is the transmittance, " + }, + { + "bbox": [ + 313, + 510, + 555, + 535 + ], + "type": "inline_equation", + "content": "\\sigma_{i}" + }, + { + "bbox": [ + 313, + 510, + 555, + 535 + ], + "type": "text", + "content": " is the opacity, and " + }, + { + "bbox": [ + 313, + 510, + 555, + 535 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_{i}" + }, + { + "bbox": [ + 313, + 510, + 555, + 535 + ], + "type": "text", + "content": " is the color of each 3D Gaussian." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 542, + 432, + 555 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 542, + 432, + 555 + ], + "spans": [ + { + "bbox": [ + 313, + 542, + 432, + 555 + ], + "type": "text", + "content": "3.2. Specular Rendering" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 560, + 556, + 621 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 560, + 556, + 621 + ], + "spans": [ + { + "bbox": [ + 313, + 560, + 556, + 621 + ], + "type": "text", + "content": "Since accurate reflections depend heavily on precise geometry, we implement a three-stage coarse-to-fine training strategy: static, dynamic, and specular stages. This approach ensures both stable scene geometry and accurate specular rendering." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 627, + 484, + 639 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 627, + 484, + 639 + ], + "spans": [ + { + "bbox": [ + 313, + 627, + 484, + 639 + ], + "type": "text", + "content": "3.2.1. Coarse-to-Fine Training Strategy" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "text", + "content": "Static stage. In the static stage, we employ vanilla 3DGS [16] for static scene reconstruction to stabilize the geometry of the static scene. Specifically, we optimize the position " + }, + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "text", + "content": ", scaling " + }, + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "text", + "content": ", rotation " + }, + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "text", + "content": ", opacity " + }, + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "text", + "content": ", and coefficients of spherical harmonics (SH) of the 3D Gaussians by minimizing the photometric loss " + }, + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{color}}" + }, + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "text", + "content": " identical to 3DGS [16]." + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "21329" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 70, + 206, + 210 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 206, + 210 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 206, + 210 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 206, + 210 + ], + "type": "image", + "image_path": "2fe4de96c4e1a699f603f34df7fee1b9f4e35d3c3ff1b8bcea1ebf92689f0f0f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 210, + 70, + 554, + 211 + ], + "blocks": [ + { + "bbox": [ + 210, + 70, + 554, + 211 + ], + "lines": [ + { + "bbox": [ + 210, + 70, + 554, + 211 + ], + "spans": [ + { + "bbox": [ + 210, + 70, + 554, + 211 + ], + "type": "image", + "image_path": "7c2262654f0f8b253faefba07eab885d64d180700e9c207ad6f7c94f2122b8a4.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 58, + 214, + 554, + 360 + ], + "blocks": [ + { + "bbox": [ + 58, + 214, + 554, + 360 + ], + "lines": [ + { + "bbox": [ + 58, + 214, + 554, + 360 + ], + "spans": [ + { + "bbox": [ + 58, + 214, + 554, + 360 + ], + "type": "image", + "image_path": "4cf7c210cddea336779c2a496eb78646127e1383a8ba0777c5f44f6ea13d1a4a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 369, + 555, + 457 + ], + "lines": [ + { + "bbox": [ + 55, + 369, + 555, + 457 + ], + "spans": [ + { + "bbox": [ + 55, + 369, + 555, + 457 + ], + "type": "text", + "content": "Figure 2. Method Overview. Our method stabilizes the scene geometry through three stages. In the static stage, we stabilize the geometry of the static scene by minimizing photometric loss " + }, + { + "bbox": [ + 55, + 369, + 555, + 457 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{color}}" + }, + { + "bbox": [ + 55, + 369, + 555, + 457 + ], + "type": "text", + "content": " between vanilla 3DGS renders and ground truth images. The dynamic stage combines canonical 3D Gaussians " + }, + { + "bbox": [ + 55, + 369, + 555, + 457 + ], + "type": "inline_equation", + "content": "\\mathbf{G}" + }, + { + "bbox": [ + 55, + 369, + 555, + 457 + ], + "type": "text", + "content": " with a deformable Gaussian MLP to model dynamic scenes while simultaneously minimizing normal loss " + }, + { + "bbox": [ + 55, + 369, + 555, + 457 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{normal}}" + }, + { + "bbox": [ + 55, + 369, + 555, + 457 + ], + "type": "text", + "content": " between rendered normal map " + }, + { + "bbox": [ + 55, + 369, + 555, + 457 + ], + "type": "inline_equation", + "content": "\\mathbf{N}^t" + }, + { + "bbox": [ + 55, + 369, + 555, + 457 + ], + "type": "text", + "content": " and gradient normal map from depth map " + }, + { + "bbox": [ + 55, + 369, + 555, + 457 + ], + "type": "inline_equation", + "content": "\\mathbf{D}^t" + }, + { + "bbox": [ + 55, + 369, + 555, + 457 + ], + "type": "text", + "content": ", thus further enhancing the overall scene geometry. Finally, the specular stage introduces a deformable reflection MLP to handle changing environment lighting, deforming reflection directions " + }, + { + "bbox": [ + 55, + 369, + 555, + 457 + ], + "type": "inline_equation", + "content": "\\omega_r^t" + }, + { + "bbox": [ + 55, + 369, + 555, + 457 + ], + "type": "text", + "content": " to query a canonical environment map for specular color " + }, + { + "bbox": [ + 55, + 369, + 555, + 457 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_s^t" + }, + { + "bbox": [ + 55, + 369, + 555, + 457 + ], + "type": "text", + "content": ". It is then combined with diffuse color " + }, + { + "bbox": [ + 55, + 369, + 555, + 457 + ], + "type": "inline_equation", + "content": "\\mathbf{c_d}" + }, + { + "bbox": [ + 55, + 369, + 555, + 457 + ], + "type": "text", + "content": " (using zero-order spherical harmonics) and learnable specular tint " + }, + { + "bbox": [ + 55, + 369, + 555, + 457 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_{\\mathrm{int}}" + }, + { + "bbox": [ + 55, + 369, + 555, + 457 + ], + "type": "text", + "content": " per 3D Gaussian to obtain the final color " + }, + { + "bbox": [ + 55, + 369, + 555, + 457 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_{\\mathrm{final}}^t" + }, + { + "bbox": [ + 55, + 369, + 555, + 457 + ], + "type": "text", + "content": ". This approach enables the modeling of dynamic specular scenes and high-quality novel view rendering." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 478, + 295, + 585 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 478, + 295, + 585 + ], + "spans": [ + { + "bbox": [ + 54, + 478, + 295, + 585 + ], + "type": "text", + "content": "Dynamic stage. Following the static stage, we address dynamic objects using Deformable 3DGS [49]. For each 3D Gaussian in canonical 3D Gaussians " + }, + { + "bbox": [ + 54, + 478, + 295, + 585 + ], + "type": "inline_equation", + "content": "\\mathbf{G}" + }, + { + "bbox": [ + 54, + 478, + 295, + 585 + ], + "type": "text", + "content": ", we input its position " + }, + { + "bbox": [ + 54, + 478, + 295, + 585 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 54, + 478, + 295, + 585 + ], + "type": "text", + "content": " and time " + }, + { + "bbox": [ + 54, + 478, + 295, + 585 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 54, + 478, + 295, + 585 + ], + "type": "text", + "content": " into a deformable Gaussian MLP with parameters " + }, + { + "bbox": [ + 54, + 478, + 295, + 585 + ], + "type": "inline_equation", + "content": "\\theta_{G}" + }, + { + "bbox": [ + 54, + 478, + 295, + 585 + ], + "type": "text", + "content": " to predict position, rotation, and scaling residuals: " + }, + { + "bbox": [ + 54, + 478, + 295, + 585 + ], + "type": "inline_equation", + "content": "(\\Delta \\pmb{x}^{t}, \\Delta \\pmb{r}^{t}, \\Delta \\pmb{s}^{t}) = F_{\\theta_{G}}(\\gamma(x), \\gamma(t))" + }, + { + "bbox": [ + 54, + 478, + 295, + 585 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 54, + 478, + 295, + 585 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 54, + 478, + 295, + 585 + ], + "type": "text", + "content": " denotes positional encoding. Attributes of the corresponding 3D Gaussian in deformed 3D Gaussians " + }, + { + "bbox": [ + 54, + 478, + 295, + 585 + ], + "type": "inline_equation", + "content": "\\mathbf{G}^{t}" + }, + { + "bbox": [ + 54, + 478, + 295, + 585 + ], + "type": "text", + "content": " at time " + }, + { + "bbox": [ + 54, + 478, + 295, + 585 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 54, + 478, + 295, + 585 + ], + "type": "text", + "content": " is obtained by " + }, + { + "bbox": [ + 54, + 478, + 295, + 585 + ], + "type": "inline_equation", + "content": "(\\pmb{x}^{t}, \\pmb{r}^{t}, \\pmb{s}^{t}) = (\\Delta \\pmb{x}^{t}, \\Delta \\pmb{r}^{t}, \\Delta \\pmb{s}^{t}) + (\\pmb{x}, \\pmb{r}, \\pmb{s})" + }, + { + "bbox": [ + 54, + 478, + 295, + 585 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 586, + 295, + 645 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 586, + 295, + 645 + ], + "spans": [ + { + "bbox": [ + 55, + 586, + 295, + 645 + ], + "type": "text", + "content": "This approach separates motion and geometric structural learning, allowing accurate simulation of dynamic behaviors while maintaining a stable geometric reference. To further enhance scene geometry, we estimate normals of deformed 3D Gaussians and optimize them using:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 127, + 654, + 295, + 668 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 654, + 295, + 668 + ], + "spans": [ + { + "bbox": [ + 127, + 654, + 295, + 668 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {n o r m a l}} = 1 - \\mathbf {N} ^ {t} \\cdot \\hat {\\mathbf {N}} ^ {t}, \\tag {3}", + "image_path": "dac48337c17489d7db3a478b8c9b8c23616ea9296524a2b540e3b7b5ac0d39e7.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{N}^t" + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": " is the rendered normal map and " + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{N}}^t" + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": " is the normal map derived from the rendered depth map " + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{D}^t" + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": ". This process improves local associations among 3D Gaussians and opti" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 479, + 554, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 479, + 554, + 501 + ], + "spans": [ + { + "bbox": [ + 313, + 479, + 554, + 501 + ], + "type": "text", + "content": "mizes both depth and normal information across the entire scene." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 505, + 555, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 505, + 555, + 577 + ], + "spans": [ + { + "bbox": [ + 313, + 505, + 555, + 577 + ], + "type": "text", + "content": "Specular stage. We adopt an image-based lighting (IBL) model, where the environment light is given by a learnable cubemap. Following the rendering equation [14], splitsum approximation [15, 30], and Cook-Torrance reflectance model [5], the outgoing radiance of the specular component " + }, + { + "bbox": [ + 313, + 505, + 555, + 577 + ], + "type": "inline_equation", + "content": "L_{s}" + }, + { + "bbox": [ + 313, + 505, + 555, + 577 + ], + "type": "text", + "content": " is expressed as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 341, + 582, + 555, + 635 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 341, + 582, + 555, + 635 + ], + "spans": [ + { + "bbox": [ + 341, + 582, + 555, + 635 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} L _ {s} = \\int_ {\\Omega} \\frac {D G F}{4 (\\omega_ {o} ^ {t} \\cdot \\mathbf {n} ^ {t}) (\\omega_ {i} \\cdot \\mathbf {n} ^ {t})} (\\boldsymbol {\\omega} _ {i} \\cdot \\mathbf {n} ^ {t}) d \\boldsymbol {\\omega} _ {i} \\\\ \\times \\int_ {\\Omega} L _ {i} \\left(\\omega_ {i}\\right) D \\left(\\omega_ {i}, \\omega_ {o} ^ {t}\\right) \\left(\\omega_ {i} \\cdot \\mathbf {n} ^ {t}\\right) d \\omega_ {i}, \\tag {4} \\\\ \\end{array}", + "image_path": "12860998c334b55dc046a2e4cfa98900207022788bc92fd15322a87ee91c2c69.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 641, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 555, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 641, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\Omega" + }, + { + "bbox": [ + 313, + 641, + 555, + 713 + ], + "type": "text", + "content": " is the hemisphere around the surface normal " + }, + { + "bbox": [ + 313, + 641, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{n}^t" + }, + { + "bbox": [ + 313, + 641, + 555, + 713 + ], + "type": "text", + "content": " (describe in Sec. 3.3.) " + }, + { + "bbox": [ + 313, + 641, + 555, + 713 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 313, + 641, + 555, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 641, + 555, + 713 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 313, + 641, + 555, + 713 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 641, + 555, + 713 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 313, + 641, + 555, + 713 + ], + "type": "text", + "content": " represent the GGX normal distribution function [43], geometric attenuation, and Fresnel term, respectively. " + }, + { + "bbox": [ + 313, + 641, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\omega_o^t" + }, + { + "bbox": [ + 313, + 641, + 555, + 713 + ], + "type": "text", + "content": " is the view direction, and " + }, + { + "bbox": [ + 313, + 641, + 555, + 713 + ], + "type": "inline_equation", + "content": "L_i(\\omega_i)" + }, + { + "bbox": [ + 313, + 641, + 555, + 713 + ], + "type": "text", + "content": " is the incident radiance. In the first term, we follow the GaussianShader [13] directly computed by " + }, + { + "bbox": [ + 313, + 641, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_{\\mathrm{int}} * F_1 + F_2" + }, + { + "bbox": [ + 313, + 641, + 555, + 713 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21330" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 203 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 203 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 72, + 296, + 203 + ], + "type": "inline_equation", + "content": "F_{1}" + }, + { + "bbox": [ + 55, + 72, + 296, + 203 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 72, + 296, + 203 + ], + "type": "inline_equation", + "content": "F_{2}" + }, + { + "bbox": [ + 55, + 72, + 296, + 203 + ], + "type": "text", + "content": " are two pre-computed scalars depending on roughness " + }, + { + "bbox": [ + 55, + 72, + 296, + 203 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 55, + 72, + 296, + 203 + ], + "type": "text", + "content": ", view direction " + }, + { + "bbox": [ + 55, + 72, + 296, + 203 + ], + "type": "inline_equation", + "content": "\\omega_{o}^{t}" + }, + { + "bbox": [ + 55, + 72, + 296, + 203 + ], + "type": "text", + "content": " and normal " + }, + { + "bbox": [ + 55, + 72, + 296, + 203 + ], + "type": "inline_equation", + "content": "\\mathbf{n}^t" + }, + { + "bbox": [ + 55, + 72, + 296, + 203 + ], + "type": "text", + "content": ". Roughness " + }, + { + "bbox": [ + 55, + 72, + 296, + 203 + ], + "type": "inline_equation", + "content": "\\rho \\in [0,1]" + }, + { + "bbox": [ + 55, + 72, + 296, + 203 + ], + "type": "text", + "content": " and specular tint " + }, + { + "bbox": [ + 55, + 72, + 296, + 203 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_{\\mathrm{int}} \\in [0,1]^3" + }, + { + "bbox": [ + 55, + 72, + 296, + 203 + ], + "type": "text", + "content": " are learnable parameters for each 3D Gaussian. The second term is pre-integrated in a filtered learnable cubemap, where each mip-level corresponds to a specific roughness value. The cubemap can be queried using the reflection direction to obtain the value of the second term. After the static and dynamic stages, the geometry is well-defined. This allows us to calculate reflection directions " + }, + { + "bbox": [ + 55, + 72, + 296, + 203 + ], + "type": "inline_equation", + "content": "\\omega_{r}^{t} = 2(\\omega_{o}^{t} \\cdot \\mathbf{n}^{t})\\mathbf{n}^{t} - \\omega_{o}^{t}" + }, + { + "bbox": [ + 55, + 72, + 296, + 203 + ], + "type": "text", + "content": " accurately." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 204, + 296, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 204, + 296, + 251 + ], + "spans": [ + { + "bbox": [ + 55, + 204, + 296, + 251 + ], + "type": "inline_equation", + "content": "L_{s}" + }, + { + "bbox": [ + 55, + 204, + 296, + 251 + ], + "type": "text", + "content": " represents only the specular color of the static environment light. To handle time-varying lighting in dynamic scenes, we introduce a deformable environment map, detailed in the following section." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 258, + 296, + 281 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 258, + 296, + 281 + ], + "spans": [ + { + "bbox": [ + 55, + 258, + 296, + 281 + ], + "type": "text", + "content": "3.2.2. Deformable Environment Map for Dynamic Lighting." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 285, + 296, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 285, + 296, + 403 + ], + "spans": [ + { + "bbox": [ + 55, + 285, + 296, + 403 + ], + "type": "text", + "content": "The concept of a deformable environment map involves treating the vanilla environment cubemap as a canonical environment map and combining it with a deformation field. This approach enables us to model time-varying lighting conditions effectively. We first apply positional encoding to the reflection direction " + }, + { + "bbox": [ + 55, + 285, + 296, + 403 + ], + "type": "inline_equation", + "content": "\\omega_r^t" + }, + { + "bbox": [ + 55, + 285, + 296, + 403 + ], + "type": "text", + "content": " and time " + }, + { + "bbox": [ + 55, + 285, + 296, + 403 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 285, + 296, + 403 + ], + "type": "text", + "content": " to implement this. These encoded values are then input into a deformable reflection MLP with parameters " + }, + { + "bbox": [ + 55, + 285, + 296, + 403 + ], + "type": "inline_equation", + "content": "\\theta_R" + }, + { + "bbox": [ + 55, + 285, + 296, + 403 + ], + "type": "text", + "content": ". This process allows us to obtain the deformed reflection residual " + }, + { + "bbox": [ + 55, + 285, + 296, + 403 + ], + "type": "inline_equation", + "content": "\\Delta \\bar{\\omega}_r^t = F_{\\theta_R}(\\gamma (\\omega_r^t),\\gamma (t))" + }, + { + "bbox": [ + 55, + 285, + 296, + 403 + ], + "type": "text", + "content": " for each specified time " + }, + { + "bbox": [ + 55, + 285, + 296, + 403 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 285, + 296, + 403 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 405, + 295, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 405, + 295, + 441 + ], + "spans": [ + { + "bbox": [ + 55, + 405, + 295, + 441 + ], + "type": "text", + "content": "Subsequently, we add the deformed reflection residual " + }, + { + "bbox": [ + 55, + 405, + 295, + 441 + ], + "type": "inline_equation", + "content": "\\Delta \\bar{\\omega}_r^t" + }, + { + "bbox": [ + 55, + 405, + 295, + 441 + ], + "type": "text", + "content": " to the reflection direction " + }, + { + "bbox": [ + 55, + 405, + 295, + 441 + ], + "type": "inline_equation", + "content": "\\omega_r^t" + }, + { + "bbox": [ + 55, + 405, + 295, + 441 + ], + "type": "text", + "content": ", yielding the deformed reflection direction " + }, + { + "bbox": [ + 55, + 405, + 295, + 441 + ], + "type": "inline_equation", + "content": "\\bar{\\omega}_r^t = \\Delta \\bar{\\omega}_r^t + \\omega_r^t" + }, + { + "bbox": [ + 55, + 405, + 295, + 441 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 441, + 296, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 441, + 296, + 513 + ], + "spans": [ + { + "bbox": [ + 55, + 441, + 296, + 513 + ], + "type": "text", + "content": "We can then use this deformed reflection direction " + }, + { + "bbox": [ + 55, + 441, + 296, + 513 + ], + "type": "inline_equation", + "content": "\\bar{\\omega}_r^t" + }, + { + "bbox": [ + 55, + 441, + 296, + 513 + ], + "type": "text", + "content": " to query the canonical environment map. The queried value is then multiplied by the first term of Equation 4, allowing us to obtain time-varying specular color " + }, + { + "bbox": [ + 55, + 441, + 296, + 513 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_s^t" + }, + { + "bbox": [ + 55, + 441, + 296, + 513 + ], + "type": "text", + "content": ". This approach effectively captures the dynamic nature of lighting in the scene while maintaining a stable canonical reference." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 519, + 295, + 543 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 519, + 295, + 543 + ], + "spans": [ + { + "bbox": [ + 55, + 519, + 295, + 543 + ], + "type": "text", + "content": "3.2.3. Color Decomposition and Staged Training Strategy." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 545, + 296, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 296, + 700 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 296, + 700 + ], + "type": "text", + "content": "We decompose the final color " + }, + { + "bbox": [ + 55, + 545, + 296, + 700 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_{\\mathrm{final}}^t" + }, + { + "bbox": [ + 55, + 545, + 296, + 700 + ], + "type": "text", + "content": " into diffuse and specular components to better distinguish between high and low-frequency information: " + }, + { + "bbox": [ + 55, + 545, + 296, + 700 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_{\\mathrm{final}}^t = \\mathbf{c}_d + \\mathbf{c}_s^t" + }, + { + "bbox": [ + 55, + 545, + 296, + 700 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 545, + 296, + 700 + ], + "type": "inline_equation", + "content": "\\mathbf{c_d}" + }, + { + "bbox": [ + 55, + 545, + 296, + 700 + ], + "type": "text", + "content": " is the diffuse color (using zero-order spherical harmonics as view-independent color), and " + }, + { + "bbox": [ + 55, + 545, + 296, + 700 + ], + "type": "inline_equation", + "content": "\\mathbf{c_s}^t" + }, + { + "bbox": [ + 55, + 545, + 296, + 700 + ], + "type": "text", + "content": " is the view-dependent color component. To manage the transition from spherical harmonics to " + }, + { + "bbox": [ + 55, + 545, + 296, + 700 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_{\\mathrm{final}}^t" + }, + { + "bbox": [ + 55, + 545, + 296, + 700 + ], + "type": "text", + "content": " and mitigate potential geometry disruptions, in the early specular stage, we fix the deformable Gaussian MLP and most 3D Gaussian attributes, optimizing only zero-order SH, specular tint, and roughness. We temporarily suspend densification during this phase. As " + }, + { + "bbox": [ + 55, + 545, + 296, + 700 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_{\\mathrm{final}}^t" + }, + { + "bbox": [ + 55, + 545, + 296, + 700 + ], + "type": "text", + "content": " becomes more complete, we gradually resume optimization of all parameters and reinstate the densification process." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 67, + 702, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 702, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 67, + 702, + 295, + 714 + ], + "type": "text", + "content": "We further split the specular stage into two parts, applying" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "text", + "content": "a coarse-to-fine strategy to the environment map. In the first part, we focus on optimizing the canonical environment map for time-invariant lighting. This establishes a stable foundation for the overall lighting structure. In the second part, we proceed to optimize the deformable reflection MLP for dynamic elements. This approach ensures a more robust learning process, allowing us to capture the static lighting conditions before introducing the complexities of dynamic components." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 186, + 468, + 198 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 186, + 468, + 198 + ], + "spans": [ + { + "bbox": [ + 313, + 186, + 468, + 198 + ], + "type": "text", + "content": "3.3. Physical Normal Estimation" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 203, + 555, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 203, + 555, + 382 + ], + "spans": [ + { + "bbox": [ + 313, + 203, + 555, + 382 + ], + "type": "text", + "content": "Challenges in normal estimation for 3D Gaussians. Normal estimation is essential for modeling specular objects in 3D Gaussians, where GaussianShader [12] initially used the shortest axis combined with a residual normal for approximation. While this works for static scenes, it becomes problematic with deformed Gaussians because the residual should vary at each time step. A straightforward approach of rotating the residual normal based on quaternion differences between canonical and deformed states proves insufficient, as it does not account for shape changes during deformation. When deformation alters the relative axis lengths, the shortest axis assumption breaks down. This highlights the need for a more comprehensive approach that considers both rotational and shape deformation effects to achieve accurate normal estimation for dynamic specular objects." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 385, + 556, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 385, + 556, + 493 + ], + "spans": [ + { + "bbox": [ + 313, + 385, + 556, + 493 + ], + "type": "text", + "content": "Improved rotation calculation for deformed 3D Gaussians. To overcome the limitations of naive methods and accurately model the normal of deformed 3D Gaussians, we propose using both the shortest and longest axes of canonical and deformed Gaussians to compute the rotation matrix. This approach accounts for both rotation and shape changes during deformation. We first align the deformed Gaussian's axes with those of the canonical Gaussian using the following method:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 498, + 572, + 541 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 572, + 541 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 572, + 541 + ], + "type": "interline_equation", + "content": "\\mathbf {v} _ {s} ^ {t} = \\left\\{ \\begin{array}{l l} \\mathbf {v} _ {s} ^ {t} & \\text {i f} \\mathbf {v} _ {s} \\cdot \\mathbf {v} _ {s} ^ {t} > 0, \\\\ - \\mathbf {v} _ {s} ^ {t} & \\text {o t h e r w i s e .} \\end{array} , \\quad \\mathbf {v} _ {l} ^ {t} = \\left\\{ \\begin{array}{l l} \\mathbf {v} _ {l} ^ {t} & \\text {i f} \\mathbf {v} _ {l} \\cdot \\mathbf {v} _ {l} ^ {t} > 0, \\\\ - \\mathbf {v} _ {l} ^ {t} & \\text {o t h e r w i s e .} \\end{array} , \\right. \\right. \\tag {5}", + "image_path": "60ddd0ca3c0634bee92ed2a5cb9388b04a9eebb5cd81559df95f3d6dbcf6e8eb.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 547, + 555, + 595 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 547, + 555, + 595 + ], + "spans": [ + { + "bbox": [ + 313, + 547, + 555, + 595 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 547, + 555, + 595 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_s" + }, + { + "bbox": [ + 313, + 547, + 555, + 595 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 547, + 555, + 595 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_l" + }, + { + "bbox": [ + 313, + 547, + 555, + 595 + ], + "type": "text", + "content": " represent the shortest and longest axes of canonical 3D Gaussians, while " + }, + { + "bbox": [ + 313, + 547, + 555, + 595 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_s^t" + }, + { + "bbox": [ + 313, + 547, + 555, + 595 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 547, + 555, + 595 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_l^t" + }, + { + "bbox": [ + 313, + 547, + 555, + 595 + ], + "type": "text", + "content": " denote the same for deformed 3D Gaussians. We then construct orthogonal matrices using these aligned axes and their cross-products:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 318, + 598, + 555, + 624 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 598, + 555, + 624 + ], + "spans": [ + { + "bbox": [ + 318, + 598, + 555, + 624 + ], + "type": "interline_equation", + "content": "\\mathbf {U} = \\left[ \\begin{array}{l l l} \\mathbf {v} _ {s} & \\mathbf {v} _ {l} & \\mathbf {v} _ {s} \\times \\mathbf {v} _ {l} \\end{array} \\right], \\quad \\mathbf {V} ^ {t} = \\left[ \\begin{array}{l l l} \\mathbf {v} _ {s} ^ {t} & \\mathbf {v} _ {l} ^ {t} & \\mathbf {v} _ {s} ^ {t} \\times \\mathbf {v} _ {l} ^ {t} \\end{array} \\right]. \\tag {6}", + "image_path": "f7709a5f01966ef1f7945860de894601ba59b545ff875ba5d98cd8c4b029bf21.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 624, + 524, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 624, + 524, + 635 + ], + "spans": [ + { + "bbox": [ + 313, + 624, + 524, + 635 + ], + "type": "text", + "content": "Finally, we derive the rotation matrix " + }, + { + "bbox": [ + 313, + 624, + 524, + 635 + ], + "type": "inline_equation", + "content": "\\mathbf{R}^t = \\mathbf{V}^t\\mathbf{U}^\\top" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 639, + 555, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 639, + 555, + 686 + ], + "spans": [ + { + "bbox": [ + 313, + 639, + 555, + 686 + ], + "type": "text", + "content": "Adjusting normal residuals and ensuring accuracy. To account for shape changes during deformation, we scale the normal residual based on the ratio of oblateness " + }, + { + "bbox": [ + 313, + 639, + 555, + 686 + ], + "type": "inline_equation", + "content": "\\frac{\\beta}{\\beta^t}" + }, + { + "bbox": [ + 313, + 639, + 555, + 686 + ], + "type": "text", + "content": " between canonical and deformed 3D Gaussians." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 357, + 689, + 555, + 717 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 357, + 689, + 555, + 717 + ], + "spans": [ + { + "bbox": [ + 357, + 689, + 555, + 717 + ], + "type": "interline_equation", + "content": "\\beta = \\frac {\\left| \\mathbf {v} _ {l} \\right| - \\left| \\mathbf {v} _ {s} \\right|}{\\left| \\mathbf {v} _ {l} \\right|}, \\quad \\beta^ {t} = \\frac {\\left| \\mathbf {v} _ {l} ^ {t} \\right| - \\left| \\mathbf {v} _ {s} ^ {t} \\right|}{\\left| \\mathbf {v} _ {l} ^ {t} \\right|}, \\qquad (7)", + "image_path": "147cd76854811b76238b4f15da8928ccb159d5264e9924a4ef337c1fb39ead12.jpg" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "21331" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 69, + 231, + 162 + ], + "blocks": [ + { + "bbox": [ + 56, + 69, + 231, + 162 + ], + "lines": [ + { + "bbox": [ + 56, + 69, + 231, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 69, + 231, + 162 + ], + "type": "image", + "image_path": "7d81696dc4d71bf5a92c447c20b447b27990ec5ce755edc11e1da7a34f623d56.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 168, + 555, + 271 + ], + "lines": [ + { + "bbox": [ + 55, + 168, + 555, + 271 + ], + "spans": [ + { + "bbox": [ + 55, + 168, + 555, + 271 + ], + "type": "text", + "content": "Figure 3. Normal estimation. (a) shows that flatter 3D Gaussians align better with scene surfaces, their shortest axis closely matching the surface normal. In contrast, less flat 3D Gaussians fit less accurately, with their shortest axis diverging from the surface normal. (b) shows that when the deformed 3D Gaussian becomes flatter (" + }, + { + "bbox": [ + 55, + 168, + 555, + 271 + ], + "type": "inline_equation", + "content": "t = t_1" + }, + { + "bbox": [ + 55, + 168, + 555, + 271 + ], + "type": "text", + "content": "), normal residual " + }, + { + "bbox": [ + 55, + 168, + 555, + 271 + ], + "type": "inline_equation", + "content": "\\Delta \\mathbf{n}" + }, + { + "bbox": [ + 55, + 168, + 555, + 271 + ], + "type": "text", + "content": " is rotated by " + }, + { + "bbox": [ + 55, + 168, + 555, + 271 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_1^t" + }, + { + "bbox": [ + 55, + 168, + 555, + 271 + ], + "type": "text", + "content": " and scaled down by " + }, + { + "bbox": [ + 55, + 168, + 555, + 271 + ], + "type": "inline_equation", + "content": "\\frac{\\beta}{\\beta_1^t}" + }, + { + "bbox": [ + 55, + 168, + 555, + 271 + ], + "type": "text", + "content": ", as flatter Gaussians require smaller normal residuals. Conversely, when the deformation results in a less flat shape (" + }, + { + "bbox": [ + 55, + 168, + 555, + 271 + ], + "type": "inline_equation", + "content": "t = t_2" + }, + { + "bbox": [ + 55, + 168, + 555, + 271 + ], + "type": "text", + "content": "), " + }, + { + "bbox": [ + 55, + 168, + 555, + 271 + ], + "type": "inline_equation", + "content": "\\Delta \\mathbf{n}" + }, + { + "bbox": [ + 55, + 168, + 555, + 271 + ], + "type": "text", + "content": " is rotated by " + }, + { + "bbox": [ + 55, + 168, + 555, + 271 + ], + "type": "inline_equation", + "content": "\\mathbf{R}_2^t" + }, + { + "bbox": [ + 55, + 168, + 555, + 271 + ], + "type": "text", + "content": " and amplified by " + }, + { + "bbox": [ + 55, + 168, + 555, + 271 + ], + "type": "inline_equation", + "content": "\\frac{\\beta}{\\beta_2^t}" + }, + { + "bbox": [ + 55, + 168, + 555, + 271 + ], + "type": "text", + "content": ", requiring a larger correction to align the shortest axis with the surface normal. (c) shows how " + }, + { + "bbox": [ + 55, + 168, + 555, + 271 + ], + "type": "inline_equation", + "content": "\\gamma^k" + }, + { + "bbox": [ + 55, + 168, + 555, + 271 + ], + "type": "text", + "content": " changes with " + }, + { + "bbox": [ + 55, + 168, + 555, + 271 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 55, + 168, + 555, + 271 + ], + "type": "text", + "content": " (where " + }, + { + "bbox": [ + 55, + 168, + 555, + 271 + ], + "type": "inline_equation", + "content": "w = \\frac{|v_s^t|}{|v_l^t|}" + }, + { + "bbox": [ + 55, + 168, + 555, + 271 + ], + "type": "text", + "content": ") for " + }, + { + "bbox": [ + 55, + 168, + 555, + 271 + ], + "type": "inline_equation", + "content": "k = 1, k = 5" + }, + { + "bbox": [ + 55, + 168, + 555, + 271 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 168, + 555, + 271 + ], + "type": "inline_equation", + "content": "k = 50" + }, + { + "bbox": [ + 55, + 168, + 555, + 271 + ], + "type": "text", + "content": ". Larger " + }, + { + "bbox": [ + 55, + 168, + 555, + 271 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 55, + 168, + 555, + 271 + ], + "type": "text", + "content": " indicates less flat Gaussians, while smaller " + }, + { + "bbox": [ + 55, + 168, + 555, + 271 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 55, + 168, + 555, + 271 + ], + "type": "text", + "content": " represents flatter Gaussians. As " + }, + { + "bbox": [ + 55, + 168, + 555, + 271 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 168, + 555, + 271 + ], + "type": "text", + "content": " increases, " + }, + { + "bbox": [ + 55, + 168, + 555, + 271 + ], + "type": "inline_equation", + "content": "\\gamma^k" + }, + { + "bbox": [ + 55, + 168, + 555, + 271 + ], + "type": "text", + "content": " decreases more steeply as " + }, + { + "bbox": [ + 55, + 168, + 555, + 271 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 55, + 168, + 555, + 271 + ], + "type": "text", + "content": " rises. For " + }, + { + "bbox": [ + 55, + 168, + 555, + 271 + ], + "type": "inline_equation", + "content": "k = 5" + }, + { + "bbox": [ + 55, + 168, + 555, + 271 + ], + "type": "text", + "content": ", we observe a balanced behavior: " + }, + { + "bbox": [ + 55, + 168, + 555, + 271 + ], + "type": "inline_equation", + "content": "\\gamma^k" + }, + { + "bbox": [ + 55, + 168, + 555, + 271 + ], + "type": "text", + "content": " approaches 1 for low " + }, + { + "bbox": [ + 55, + 168, + 555, + 271 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 55, + 168, + 555, + 271 + ], + "type": "text", + "content": " and 0 for high " + }, + { + "bbox": [ + 55, + 168, + 555, + 271 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 55, + 168, + 555, + 271 + ], + "type": "text", + "content": ", providing a nuanced penalty adjustment across different Gaussian shapes." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 228, + 68, + 444, + 161 + ], + "blocks": [ + { + "bbox": [ + 228, + 68, + 444, + 161 + ], + "lines": [ + { + "bbox": [ + 228, + 68, + 444, + 161 + ], + "spans": [ + { + "bbox": [ + 228, + 68, + 444, + 161 + ], + "type": "image", + "image_path": "005bf15805ae5f13c0468f2a75422b9d04cc850c1889841e0d7532cde013b4f5.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 445, + 68, + 556, + 161 + ], + "blocks": [ + { + "bbox": [ + 445, + 68, + 556, + 161 + ], + "lines": [ + { + "bbox": [ + 445, + 68, + 556, + 161 + ], + "spans": [ + { + "bbox": [ + 445, + 68, + 556, + 161 + ], + "type": "image", + "image_path": "da8685942c3cea737a4584c8c01ed14ef155082cd41d1dfe232bab939e1bbb72.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 291, + 296, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 291, + 296, + 387 + ], + "spans": [ + { + "bbox": [ + 55, + 291, + 296, + 387 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 291, + 296, + 387 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 55, + 291, + 296, + 387 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 291, + 296, + 387 + ], + "type": "inline_equation", + "content": "\\beta^t" + }, + { + "bbox": [ + 55, + 291, + 296, + 387 + ], + "type": "text", + "content": " represent the oblateness of canonical and deformed 3D Gaussians, respectively. This is because flatter 3D Gaussians tend to align more closely with the surface, meaning their shortest axis becomes more aligned with the surface normal, as shown in Fig. 3 (a). In such cases, less compensation from the normal residual is needed. Conversely, less flat Gaussians require more compensation, as illustrated in Fig. 3 (b). We then obtain deformed normal residuals:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 137, + 394, + 296, + 418 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 394, + 296, + 418 + ], + "spans": [ + { + "bbox": [ + 137, + 394, + 296, + 418 + ], + "type": "interline_equation", + "content": "\\Delta \\mathbf {n} ^ {t} = \\frac {\\beta}{\\beta^ {t}} \\mathbf {R} ^ {t} \\Delta \\mathbf {n}. \\tag {8}", + "image_path": "2953c4a5d209c1df4353f14f25a195974c076607d94b5981b6142a2ccc93fb0d.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 427, + 295, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 427, + 295, + 452 + ], + "spans": [ + { + "bbox": [ + 55, + 427, + 295, + 452 + ], + "type": "text", + "content": "The final normal " + }, + { + "bbox": [ + 55, + 427, + 295, + 452 + ], + "type": "inline_equation", + "content": "\\mathbf{n}^t" + }, + { + "bbox": [ + 55, + 427, + 295, + 452 + ], + "type": "text", + "content": " is computed by adding this residual to the shortest axis and ensuring outward orientation:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 70, + 459, + 296, + 491 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 459, + 296, + 491 + ], + "spans": [ + { + "bbox": [ + 70, + 459, + 296, + 491 + ], + "type": "interline_equation", + "content": "\\mathbf {n} ^ {t} = \\Delta \\mathbf {n} ^ {t} + \\mathbf {v} _ {s} ^ {t}, \\quad \\mathbf {n} ^ {t} = \\left\\{ \\begin{array}{l l} \\mathbf {n} ^ {t} & \\text {i f} \\mathbf {n} ^ {t} \\cdot \\omega_ {o} ^ {t} > 0, \\\\ - \\mathbf {n} ^ {t} & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {9}", + "image_path": "84fefdbe13094a4695ebbd5c72b80cc788b52579d3731e7ae5d7fadca69da3fc.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 498, + 296, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 296, + 521 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 296, + 521 + ], + "type": "text", + "content": "This approach adjusts for Gaussian flatness and ensures accurate normal estimation." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 529, + 149, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 529, + 149, + 540 + ], + "spans": [ + { + "bbox": [ + 55, + 529, + 149, + 540 + ], + "type": "text", + "content": "3.4. Loss Functions" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 546, + 296, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 546, + 296, + 594 + ], + "spans": [ + { + "bbox": [ + 55, + 546, + 296, + 594 + ], + "type": "text", + "content": "Normal regularization. To allow the normal residual to correct the normal while not excessively influencing the optimization of the shortest axis towards the surface normal, we introduce a penalty term for the normal residual:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 73, + 603, + 296, + 635 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 603, + 296, + 635 + ], + "spans": [ + { + "bbox": [ + 73, + 603, + 296, + 635 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {r e g}} = \\gamma^ {k} \\| \\Delta \\mathbf {n} \\| _ {2} ^ {2} \\quad \\text {w h e r e} \\quad \\gamma = \\sqrt {1 - \\frac {\\left| \\mathbf {v} _ {s} ^ {t} \\right| ^ {2}}{\\left| \\mathbf {v} _ {l} ^ {t} \\right| ^ {2}}}. \\tag {10}", + "image_path": "9b5a65c21c9ea6f2fc4a57b8b3dd5a03eb886acf949385b06b174ff80cf6558a.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 641, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 296, + 713 + ], + "type": "text", + "content": "In our experiments, we set " + }, + { + "bbox": [ + 55, + 641, + 296, + 713 + ], + "type": "inline_equation", + "content": "k = 5" + }, + { + "bbox": [ + 55, + 641, + 296, + 713 + ], + "type": "text", + "content": ". When " + }, + { + "bbox": [ + 55, + 641, + 296, + 713 + ], + "type": "inline_equation", + "content": "k = 5" + }, + { + "bbox": [ + 55, + 641, + 296, + 713 + ], + "type": "text", + "content": ", less flatter 3D Gaussians have " + }, + { + "bbox": [ + 55, + 641, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\gamma^k" + }, + { + "bbox": [ + 55, + 641, + 296, + 713 + ], + "type": "text", + "content": " approaching 0. Their shortest axis aligns poorly with the surface normal, requiring more normal residual correction and smaller penalties. Conversely, flatter Gaussians have " + }, + { + "bbox": [ + 55, + 641, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\gamma^k" + }, + { + "bbox": [ + 55, + 641, + 296, + 713 + ], + "type": "text", + "content": " near 1. Their shortest axis aligns better with the surface normal, needing less normal residual" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 292, + 553, + 315 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 292, + 553, + 315 + ], + "spans": [ + { + "bbox": [ + 313, + 292, + 553, + 315 + ], + "type": "text", + "content": "correction and allowing larger penalties, as shown in Fig. 3 (c)." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 318, + 554, + 343 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 318, + 554, + 343 + ], + "spans": [ + { + "bbox": [ + 313, + 318, + 554, + 343 + ], + "type": "text", + "content": "Total training loss. To refine all parameters in the dynamic and specular stages, we employ the total training loss:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 363, + 352, + 555, + 365 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 363, + 352, + 555, + 365 + ], + "spans": [ + { + "bbox": [ + 363, + 352, + 555, + 365 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {\\text {c o l o r}} + \\lambda_ {\\text {n o r m a l}} \\mathcal {L} _ {\\text {n o r m a l}} + \\mathcal {L} _ {\\text {r e g}}, \\tag {11}", + "image_path": "5c66122967519d03a124e90c299e954144094bbd51180c02d6b19fdc95926fdb.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 373, + 555, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 373, + 555, + 421 + ], + "spans": [ + { + "bbox": [ + 313, + 373, + 555, + 421 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 373, + 555, + 421 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{color}}" + }, + { + "bbox": [ + 313, + 373, + 555, + 421 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 373, + 555, + 421 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{normal}}" + }, + { + "bbox": [ + 313, + 373, + 555, + 421 + ], + "type": "text", + "content": " are obtained as described in Section 3.2.1. In our experiments, we set " + }, + { + "bbox": [ + 313, + 373, + 555, + 421 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{normal}} = 0.01" + }, + { + "bbox": [ + 313, + 373, + 555, + 421 + ], + "type": "text", + "content": ". Due to space constraints, complete implementation details are provided in the supplementary materials." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 431, + 395, + 444 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 431, + 395, + 444 + ], + "spans": [ + { + "bbox": [ + 313, + 431, + 395, + 444 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 450, + 425, + 462 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 450, + 425, + 462 + ], + "spans": [ + { + "bbox": [ + 313, + 450, + 425, + 462 + ], + "type": "text", + "content": "4.1. Evaluation Results" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 468, + 555, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 468, + 555, + 540 + ], + "spans": [ + { + "bbox": [ + 313, + 468, + 555, + 540 + ], + "type": "text", + "content": "We evaluate our method on two real-world datasets: NeRF-DS dataset [48] and HyperNeRF dataset [32]. While GaussianShader [12] and GS-IR [23] are originally designed for static scenes and are included here only as reference baselines, we train our method and all baseline approaches for 40,000 iterations to ensure fair comparison." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 543, + 555, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 543, + 555, + 638 + ], + "spans": [ + { + "bbox": [ + 313, + 543, + 555, + 638 + ], + "type": "text", + "content": "NeRF-DS dataset. The NeRF-DS dataset [48] is a monocular video dataset comprising seven real-world scenes from daily life featuring various types of moving or deforming specular objects. We compare our method with the most relevant state-of-the-art approaches. As shown in Tab. 1 and Fig. 4, the quantitative results demonstrate that our method decisively outperforms baselines in reconstructing and rendering real-world highly reflective dynamic specular scenes." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 639, + 555, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 639, + 555, + 686 + ], + "spans": [ + { + "bbox": [ + 313, + 639, + 555, + 686 + ], + "type": "text", + "content": "The rendering speed is correlated with the quantity of 3D Gaussians. When the number of 3D Gaussians is below 178k, our method can achieve real-time rendering over 30 FPS on an NVIDIA RTX 4090." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 689, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 556, + 714 + ], + "type": "text", + "content": "HyperNeRF dataset. The HyperNeRF dataset contains real-world dynamic scenes and does not include specular objects." + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21332" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 97, + 555, + 327 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 95 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 95 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 95 + ], + "type": "text", + "content": "Table 1. Quantitative comparison on the NeRF-DS [48] dataset. We report the average PSNR, SSIM, and LPIPS (VGG) of several previous models on test images. The best, the second best, and third best results are denoted by red, orange, yellow." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 97, + 555, + 327 + ], + "lines": [ + { + "bbox": [ + 56, + 97, + 555, + 327 + ], + "spans": [ + { + "bbox": [ + 56, + 97, + 555, + 327 + ], + "type": "table", + "html": "
MethodAsBasinBellCup
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
Deformable 3DGS [49]26.040.88050.185019.530.78550.192423.960.79450.276724.490.88220.1658
4DGS [46]24.850.86320.203819.260.76700.219622.860.80150.206123.820.86950.1792
GaussianShader [12]21.890.77390.362017.790.66700.418720.690.81690.302420.400.74370.3385
GS-IR [23]21.580.80330.303318.060.72480.313520.660.78290.260320.340.81930.2719
NeRF-DS [48]25.340.88030.215020.230.80530.250822.570.78110.292124.510.88020.1707
HyperNeRF [32]17.590.85180.239022.580.81560.249719.800.76500.299915.450.82950.2302
Ours26.800.88430.176119.750.79150.189625.460.84900.160024.650.88710.1588
PlatePressSieveMean
MethodPSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
Deformable 3DGS [49]19.070.73520.359925.520.85940.196425.370.86160.164323.430.82840.2201
4DGS [46]18.770.77090.272124.820.83550.225525.160.85660.174522.790.82350.2115
GaussianShader [12]14.550.64230.495519.970.72440.450722.580.78620.305719.700.73630.3819
GS-IR [23]15.980.69690.420022.280.80880.306722.840.82120.223620.250.77960.2999
NeRF-DS [48]19.700.78130.297425.350.87030.255224.990.87050.200123.240.83840.2402
HyperNeRF [32]21.220.78290.316616.540.82000.281019.920.85210.214219.010.81670.2615
Ours20.840.81720.219826.490.86570.188925.220.87050.151324.170.85220.1778
", + "image_path": "9e04e376026e56e96851e11ab40d15e92496b13e5082ebad777731dbbc9257ef.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 59, + 335, + 555, + 453 + ], + "blocks": [ + { + "bbox": [ + 59, + 335, + 555, + 453 + ], + "lines": [ + { + "bbox": [ + 59, + 335, + 555, + 453 + ], + "spans": [ + { + "bbox": [ + 59, + 335, + 555, + 453 + ], + "type": "image", + "image_path": "5c94518554cd7ab1cae7b8cc2db0a1dec111098f4882477ab2630a9bae46c691.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 183, + 462, + 425, + 473 + ], + "lines": [ + { + "bbox": [ + 183, + 462, + 425, + 473 + ], + "spans": [ + { + "bbox": [ + 183, + 462, + 425, + 473 + ], + "type": "text", + "content": "Figure 4. Qualitative comparison on the NeRF-DS [48] dataset." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 63, + 544, + 288, + 652 + ], + "blocks": [ + { + "bbox": [ + 55, + 493, + 295, + 540 + ], + "lines": [ + { + "bbox": [ + 55, + 493, + 295, + 540 + ], + "spans": [ + { + "bbox": [ + 55, + 493, + 295, + 540 + ], + "type": "text", + "content": "Table 2. Quantitative comparison on the HyperNeRF [32] dataset. We report the average PSNR, SSIM, and LPIPS (VGG) of several previous models. The best, the second best, and third best results are denoted by red, orange, yellow." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 63, + 544, + 288, + 652 + ], + "lines": [ + { + "bbox": [ + 63, + 544, + 288, + 652 + ], + "spans": [ + { + "bbox": [ + 63, + 544, + 288, + 652 + ], + "type": "table", + "html": "
MethodPSNR ↑SSIM ↑LPIPS ↓
Deformable 3DGS [49]22.780.62010.3380
4DGS [46]24.890.67810.3422
GaussianShader [12]18.550.54520.4795
GS-IR [23]19.870.57290.4498
NeRF-DS [48]23.650.64050.3972
HyperNeRF [32]23.110.63870.3968
Ours22.220.60880.3295
", + "image_path": "0ec95ac145092451ffb141850a17a7431abca813b89dc5ca7cec43a44164c66b.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "As shown in Tab. 2 and Fig. 5, the results demonstrate that we are on par with state-of-the-art techniques for rendering novel views, and our method's performance is not limited to" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 495, + 370, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 495, + 370, + 506 + ], + "spans": [ + { + "bbox": [ + 314, + 495, + 370, + 506 + ], + "type": "text", + "content": "shiny scenes." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 506, + 555, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 506, + 555, + 590 + ], + "spans": [ + { + "bbox": [ + 313, + 506, + 555, + 590 + ], + "type": "text", + "content": "In Fig. 6, we compare our method's normal maps with those from Deformable 3DGS [49] and NeRF-DS [48]. For Deformable 3DGS [49], we obtain the normals by using the shortest axes of the deformed 3D Gaussians. As demonstrated, our method produces significantly better quality normal maps compared to Deformable 3DGS [49] and NeRF-DS [48]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 597, + 408, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 597, + 408, + 609 + ], + "spans": [ + { + "bbox": [ + 313, + 597, + 408, + 609 + ], + "type": "text", + "content": "4.2. Ablation Study" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 614, + 555, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 614, + 555, + 638 + ], + "spans": [ + { + "bbox": [ + 313, + 614, + 555, + 638 + ], + "type": "text", + "content": "For a fair comparison, we train our method and all ablation experiments for 40,000 iterations." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "content": "Different coarse to fine training strategy stages. As shown in Tab. 3 and Fig. 7, each stage contributes effectively to the model's performance. The Dynamic stage enhances dynamic object stability compared to the Static stage alone, while the Specular stage improves reflection clarity beyond the combined Static and Dynamic stages." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21333" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 68, + 555, + 186 + ], + "blocks": [ + { + "bbox": [ + 57, + 68, + 555, + 186 + ], + "lines": [ + { + "bbox": [ + 57, + 68, + 555, + 186 + ], + "spans": [ + { + "bbox": [ + 57, + 68, + 555, + 186 + ], + "type": "image", + "image_path": "a9e9bf5cd13deac75326c85d20ebd03f34c97807ec784fbb025581469de96a27.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 179, + 194, + 430, + 205 + ], + "lines": [ + { + "bbox": [ + 179, + 194, + 430, + 205 + ], + "spans": [ + { + "bbox": [ + 179, + 194, + 430, + 205 + ], + "type": "text", + "content": "Figure 5. Qualitative comparison on the HyperNeRF [32] dataset." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 56, + 224, + 294, + 300 + ], + "blocks": [ + { + "bbox": [ + 56, + 224, + 294, + 300 + ], + "lines": [ + { + "bbox": [ + 56, + 224, + 294, + 300 + ], + "spans": [ + { + "bbox": [ + 56, + 224, + 294, + 300 + ], + "type": "image", + "image_path": "e6e0a71cdf767f7ab03aa7f2bce2dc89b6f74a93fbdf58beb32aae0c2aedc0e3.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 309, + 295, + 331 + ], + "lines": [ + { + "bbox": [ + 55, + 309, + 295, + 331 + ], + "spans": [ + { + "bbox": [ + 55, + 309, + 295, + 331 + ], + "type": "text", + "content": "Figure 6. Qualitative comparison of normal maps between our method, Deformable 3DGS, and NeRF-DS." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 70, + 370, + 281, + 428 + ], + "blocks": [ + { + "bbox": [ + 55, + 345, + 295, + 367 + ], + "lines": [ + { + "bbox": [ + 55, + 345, + 295, + 367 + ], + "spans": [ + { + "bbox": [ + 55, + 345, + 295, + 367 + ], + "type": "text", + "content": "Table 3. Ablation studies on different coarse to fine training strategy stages." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 70, + 370, + 281, + 428 + ], + "lines": [ + { + "bbox": [ + 70, + 370, + 281, + 428 + ], + "spans": [ + { + "bbox": [ + 70, + 370, + 281, + 428 + ], + "type": "table", + "html": "
StagePSNR ↑SSIM ↑LPIPS ↓
Static20.260.77850.2953
St. + Dynamic24.020.85080.1831
St. + Dy. + Specular24.170.85220.1778
", + "image_path": "e9c99d99eefa05bef6c03075610fb50d719758f1ad8aa2e4cf8c514404c4a8ca.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 58, + 440, + 295, + 599 + ], + "blocks": [ + { + "bbox": [ + 58, + 440, + 295, + 599 + ], + "lines": [ + { + "bbox": [ + 58, + 440, + 295, + 599 + ], + "spans": [ + { + "bbox": [ + 58, + 440, + 295, + 599 + ], + "type": "image", + "image_path": "ef87ff0f4d145e3cb09e040c7177b4a7489c7e58d8f38394243b6eac9f4e6bf5.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 608, + 295, + 631 + ], + "lines": [ + { + "bbox": [ + 55, + 608, + 295, + 631 + ], + "spans": [ + { + "bbox": [ + 55, + 608, + 295, + 631 + ], + "type": "text", + "content": "Figure 7. Qualitative comparison of each training stage in our coarse-to-fine approach." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 654, + 296, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 712 + ], + "type": "text", + "content": "Ablation study on coarse-to-fine, and loss function. The model's performance was evaluated without key components: the coarse-to-fine training strategy, normal loss " + }, + { + "bbox": [ + 55, + 654, + 296, + 712 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{normal}}" + }, + { + "bbox": [ + 55, + 654, + 296, + 712 + ], + "type": "text", + "content": ", normal regularization " + }, + { + "bbox": [ + 55, + 654, + 296, + 712 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{reg}}" + }, + { + "bbox": [ + 55, + 654, + 296, + 712 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 654, + 296, + 712 + ], + "type": "inline_equation", + "content": "\\gamma^k" + }, + { + "bbox": [ + 55, + 654, + 296, + 712 + ], + "type": "text", + "content": ". Fig. 8 and Tab. 4 illustrate the effects of these omissions. Without the coarse-to-fine" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 319, + 251, + 550, + 330 + ], + "blocks": [ + { + "bbox": [ + 313, + 224, + 555, + 247 + ], + "lines": [ + { + "bbox": [ + 313, + 224, + 555, + 247 + ], + "spans": [ + { + "bbox": [ + 313, + 224, + 555, + 247 + ], + "type": "text", + "content": "Table 4. Ablation studies on different coarse to fine training strategy stages." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 319, + 251, + 550, + 330 + ], + "lines": [ + { + "bbox": [ + 319, + 251, + 550, + 330 + ], + "spans": [ + { + "bbox": [ + 319, + 251, + 550, + 330 + ], + "type": "table", + "html": "
C2F\\( \\mathcal{L}_{\\text{normal}} \\)\\( \\mathcal{L}_{\\text{reg}} \\)\\( \\gamma^k \\)PSNR↑SSIM↑LPIPS↓
✓✓✓23.160.82940.2156
✓23.400.82770.2278
✓✓24.150.85100.1845
✓✓✓24.090.85150.1818
✓✓✓✓24.170.85220.1778
", + "image_path": "3bcf46684a25df798a6b89f76d28381ea0646a702dee521d88af429c0a0711af.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 326, + 367, + 543, + 425 + ], + "blocks": [ + { + "bbox": [ + 313, + 342, + 555, + 364 + ], + "lines": [ + { + "bbox": [ + 313, + 342, + 555, + 364 + ], + "spans": [ + { + "bbox": [ + 313, + 342, + 555, + 364 + ], + "type": "text", + "content": "Table 5. Ablation studies on SH, Static and Deformable environment map." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 326, + 367, + 543, + 425 + ], + "lines": [ + { + "bbox": [ + 326, + 367, + 543, + 425 + ], + "spans": [ + { + "bbox": [ + 326, + 367, + 543, + 425 + ], + "type": "table", + "html": "
PSNR ↑SSIM ↑LPIPS ↓
SH23.630.84530.1844
Static Env. map24.020.85080.1831
Deformable Env. map24.170.85220.1778
", + "image_path": "2455c6f27b02f2e46f403ec99b16c89988cfe0d040adff6baf9a8aaf072e87f6.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 446, + 556, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 446, + 556, + 650 + ], + "spans": [ + { + "bbox": [ + 313, + 446, + 556, + 650 + ], + "type": "text", + "content": "approach, the model, trained directly at the specular stage, produces incomplete scene geometry, affecting environment map queries for specular color. Omitting normal loss " + }, + { + "bbox": [ + 313, + 446, + 556, + 650 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{normal}}" + }, + { + "bbox": [ + 313, + 446, + 556, + 650 + ], + "type": "text", + "content": " removes direct supervision on normals and leads to inaccurate reflection directions and less precise specular colors. Removing normal regularization " + }, + { + "bbox": [ + 313, + 446, + 556, + 650 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{reg}}" + }, + { + "bbox": [ + 313, + 446, + 556, + 650 + ], + "type": "text", + "content": " allows the normal residual to dominate normal optimization, resulting in insufficient optimization of the 3D Gaussians' shortest axis towards the correct normal, which in turn reduces the rendering quality. The normal residual decreases for non-flattened and flat Gaussians without " + }, + { + "bbox": [ + 313, + 446, + 556, + 650 + ], + "type": "inline_equation", + "content": "\\gamma^k" + }, + { + "bbox": [ + 313, + 446, + 556, + 650 + ], + "type": "text", + "content": " in normal regularization. This particularly affects less flat 3D Gaussians whose shortest axis significantly deviates from the surface normal. The insufficient normal residual correction causes these 3D Gaussians' shortest axes to deviate greatly from their original direction in an attempt to align with the surface normal, ultimately reducing rendering quality." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "text", + "content": "Ablation study on SH, Static environment map, and Deformable environment map. Fig. 9 and Tab. 5 demonstrate the superiority of the deformable environment map over the static environment map, which in turn outperforms Spherical Harmonics (SH). SH struggles to accurately model high-" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21334" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 69, + 179, + 139 + ], + "blocks": [ + { + "bbox": [ + 56, + 69, + 179, + 139 + ], + "lines": [ + { + "bbox": [ + 56, + 69, + 179, + 139 + ], + "spans": [ + { + "bbox": [ + 56, + 69, + 179, + 139 + ], + "type": "image", + "image_path": "cffa3e1f530ca2a1e31da73d46e0acb77bd2e7b60976bd39a99347eeabe638b9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 92, + 140, + 137, + 149 + ], + "lines": [ + { + "bbox": [ + 92, + 140, + 137, + 149 + ], + "spans": [ + { + "bbox": [ + 92, + 140, + 137, + 149 + ], + "type": "text", + "content": "Full model" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 182, + 70, + 304, + 139 + ], + "blocks": [ + { + "bbox": [ + 182, + 70, + 304, + 139 + ], + "lines": [ + { + "bbox": [ + 182, + 70, + 304, + 139 + ], + "spans": [ + { + "bbox": [ + 182, + 70, + 304, + 139 + ], + "type": "image", + "image_path": "95c082d7f288a73c6d4f1f2c91f5181be473758f935c5d81a79fa8cda99a60c1.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 220, + 140, + 265, + 149 + ], + "lines": [ + { + "bbox": [ + 220, + 140, + 265, + 149 + ], + "spans": [ + { + "bbox": [ + 220, + 140, + 265, + 149 + ], + "type": "text", + "content": "Full model" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 307, + 70, + 429, + 139 + ], + "blocks": [ + { + "bbox": [ + 307, + 70, + 429, + 139 + ], + "lines": [ + { + "bbox": [ + 307, + 70, + 429, + 139 + ], + "spans": [ + { + "bbox": [ + 307, + 70, + 429, + 139 + ], + "type": "image", + "image_path": "5da24300c8e3a74f08660cd5651e4bd6c9ee6b60318adcc5dc42f60eb366ba51.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 343, + 140, + 388, + 149 + ], + "lines": [ + { + "bbox": [ + 343, + 140, + 388, + 149 + ], + "spans": [ + { + "bbox": [ + 343, + 140, + 388, + 149 + ], + "type": "text", + "content": "Full model" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 432, + 70, + 553, + 139 + ], + "blocks": [ + { + "bbox": [ + 432, + 70, + 553, + 139 + ], + "lines": [ + { + "bbox": [ + 432, + 70, + 553, + 139 + ], + "spans": [ + { + "bbox": [ + 432, + 70, + 553, + 139 + ], + "type": "image", + "image_path": "2eca99811ece7f867e60b06ed6fe26e77cefe9dd0b6134a391bdce9bedaaaeb6.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 473, + 140, + 518, + 149 + ], + "lines": [ + { + "bbox": [ + 473, + 140, + 518, + 149 + ], + "spans": [ + { + "bbox": [ + 473, + 140, + 518, + 149 + ], + "type": "text", + "content": "Full model" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 56, + 152, + 179, + 220 + ], + "blocks": [ + { + "bbox": [ + 56, + 152, + 179, + 220 + ], + "lines": [ + { + "bbox": [ + 56, + 152, + 179, + 220 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 179, + 220 + ], + "type": "image", + "image_path": "8ff991c9f481437702c238f47b10cb7a4b6a3056e370040673254d41ce57184f.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 74, + 221, + 151, + 231 + ], + "lines": [ + { + "bbox": [ + 74, + 221, + 151, + 231 + ], + "spans": [ + { + "bbox": [ + 74, + 221, + 151, + 231 + ], + "type": "text", + "content": "w/o Coarse-to-fine" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 182, + 152, + 304, + 220 + ], + "blocks": [ + { + "bbox": [ + 182, + 152, + 304, + 220 + ], + "lines": [ + { + "bbox": [ + 182, + 152, + 304, + 220 + ], + "spans": [ + { + "bbox": [ + 182, + 152, + 304, + 220 + ], + "type": "image", + "image_path": "2b453a3c19c78fc582a06ef7f0a0929f6db17565e8f91274536a8a6851fa0864.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 216, + 221, + 267, + 232 + ], + "lines": [ + { + "bbox": [ + 216, + 221, + 267, + 232 + ], + "spans": [ + { + "bbox": [ + 216, + 221, + 267, + 232 + ], + "type": "text", + "content": "w/o " + }, + { + "bbox": [ + 216, + 221, + 267, + 232 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{normal}}" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 307, + 152, + 429, + 220 + ], + "blocks": [ + { + "bbox": [ + 307, + 152, + 429, + 220 + ], + "lines": [ + { + "bbox": [ + 307, + 152, + 429, + 220 + ], + "spans": [ + { + "bbox": [ + 307, + 152, + 429, + 220 + ], + "type": "image", + "image_path": "eec227e92ea0cabeefb69d6c189faa5ec1094d4d7bf84f8574636a5c2e535ec8.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 349, + 221, + 387, + 233 + ], + "lines": [ + { + "bbox": [ + 349, + 221, + 387, + 233 + ], + "spans": [ + { + "bbox": [ + 349, + 221, + 387, + 233 + ], + "type": "text", + "content": "w/o " + }, + { + "bbox": [ + 349, + 221, + 387, + 233 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{reg}}" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 432, + 152, + 553, + 220 + ], + "blocks": [ + { + "bbox": [ + 432, + 152, + 553, + 220 + ], + "lines": [ + { + "bbox": [ + 432, + 152, + 553, + 220 + ], + "spans": [ + { + "bbox": [ + 432, + 152, + 553, + 220 + ], + "type": "image", + "image_path": "1c8e2e5a51f8e8c9bb1864b1608aa6f82369ce52f10be35973f72ef3cf64c6ea.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 479, + 221, + 509, + 232 + ], + "lines": [ + { + "bbox": [ + 479, + 221, + 509, + 232 + ], + "spans": [ + { + "bbox": [ + 479, + 221, + 509, + 232 + ], + "type": "text", + "content": "w/o " + }, + { + "bbox": [ + 479, + 221, + 509, + 232 + ], + "type": "inline_equation", + "content": "\\gamma^k" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 58, + 274, + 176, + 342 + ], + "blocks": [ + { + "bbox": [ + 148, + 244, + 460, + 256 + ], + "lines": [ + { + "bbox": [ + 148, + 244, + 460, + 256 + ], + "spans": [ + { + "bbox": [ + 148, + 244, + 460, + 256 + ], + "type": "text", + "content": "Figure 8. Qualitative comparison of ablation study without different components." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 58, + 274, + 176, + 342 + ], + "lines": [ + { + "bbox": [ + 58, + 274, + 176, + 342 + ], + "spans": [ + { + "bbox": [ + 58, + 274, + 176, + 342 + ], + "type": "image", + "image_path": "d4f16fca7dd438c0cf659a75042f2c5d9301b0323856ac744642bf07a4faddfb.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 178, + 275, + 294, + 342 + ], + "blocks": [ + { + "bbox": [ + 178, + 275, + 294, + 342 + ], + "lines": [ + { + "bbox": [ + 178, + 275, + 294, + 342 + ], + "spans": [ + { + "bbox": [ + 178, + 275, + 294, + 342 + ], + "type": "image", + "image_path": "be59e70aff98002103d86392026d30d7d08df8392e828199988f14bac915566c.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 208, + 343, + 262, + 353 + ], + "lines": [ + { + "bbox": [ + 208, + 343, + 262, + 353 + ], + "spans": [ + { + "bbox": [ + 208, + 343, + 262, + 353 + ], + "type": "text", + "content": "Static Env. map" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 58, + 354, + 176, + 420 + ], + "blocks": [ + { + "bbox": [ + 72, + 343, + 162, + 353 + ], + "lines": [ + { + "bbox": [ + 72, + 343, + 162, + 353 + ], + "spans": [ + { + "bbox": [ + 72, + 343, + 162, + 353 + ], + "type": "text", + "content": "Spherical Harmonics (SH)" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 58, + 354, + 176, + 420 + ], + "lines": [ + { + "bbox": [ + 58, + 354, + 176, + 420 + ], + "spans": [ + { + "bbox": [ + 58, + 354, + 176, + 420 + ], + "type": "image", + "image_path": "e5af51a886f5b77d2ec9e5c3e3fd2e2e64e46d29bb37ec71043c89146b15949e.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 58, + 422, + 176, + 432 + ], + "lines": [ + { + "bbox": [ + 58, + 422, + 176, + 432 + ], + "spans": [ + { + "bbox": [ + 58, + 422, + 176, + 432 + ], + "type": "text", + "content": "Deformable Env. map (Full model)" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 443, + 295, + 466 + ], + "lines": [ + { + "bbox": [ + 55, + 443, + 295, + 466 + ], + "spans": [ + { + "bbox": [ + 55, + 443, + 295, + 466 + ], + "type": "text", + "content": "Figure 9. Qualitative comparison of ablation study on SH, Static environment map, and Deformable environment map." + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 178, + 354, + 294, + 420 + ], + "blocks": [ + { + "bbox": [ + 178, + 354, + 294, + 420 + ], + "lines": [ + { + "bbox": [ + 178, + 354, + 294, + 420 + ], + "spans": [ + { + "bbox": [ + 178, + 354, + 294, + 420 + ], + "type": "image", + "image_path": "109096eadb039a1b5aa84b90b20f94196bd493e4927be2ad8303ac10b9b55b1d.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 210, + 422, + 256, + 430 + ], + "lines": [ + { + "bbox": [ + 210, + 422, + 256, + 430 + ], + "spans": [ + { + "bbox": [ + 210, + 422, + 256, + 430 + ], + "type": "text", + "content": "Ground-truth" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "type": "table", + "bbox": [ + 87, + 506, + 264, + 564 + ], + "blocks": [ + { + "bbox": [ + 55, + 481, + 295, + 503 + ], + "lines": [ + { + "bbox": [ + 55, + 481, + 295, + 503 + ], + "spans": [ + { + "bbox": [ + 55, + 481, + 295, + 503 + ], + "type": "text", + "content": "Table 6. Ablation studies on 2DGS and without Physical Normal Estimation." + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 87, + 506, + 264, + 564 + ], + "lines": [ + { + "bbox": [ + 87, + 506, + 264, + 564 + ], + "spans": [ + { + "bbox": [ + 87, + 506, + 264, + 564 + ], + "type": "table", + "html": "
PSNR ↑SSIM ↑LPIPS ↓
2DGS [10]23.220.82190.2283
w/o N.E.23.890.84900.1837
Full model24.170.85220.1778
", + "image_path": "5a2eaca7063cb7f8003f83e613e68d9445c8042d5b42edde6b1995f197bb45af.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "table_body" + } + ], + "index": 27 + }, + { + "bbox": [ + 55, + 589, + 296, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 589, + 296, + 648 + ], + "spans": [ + { + "bbox": [ + 55, + 589, + 296, + 648 + ], + "type": "text", + "content": "frequency specular colors. While the static environment map can model high-frequency colors, it is best suited for static lighting conditions. In contrast, the deformable environment map models time-varying lighting, offering superior performance for dynamic scenes." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 55, + 653, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 297, + 714 + ], + "type": "text", + "content": "Ablation study on 2DGS [10] and without Physical Normal Estimation. In Fig. 10 and Tab. 6, \"2DGS\" represents replacing our 3D Gaussian with 2D Gaussian representation. Since 2DGS inherently includes normals, we omit physical normal estimation. \"w/o N.E.\" means skipping physical nor" + } + ] + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 316, + 274, + 433, + 341 + ], + "blocks": [ + { + "bbox": [ + 316, + 274, + 433, + 341 + ], + "lines": [ + { + "bbox": [ + 316, + 274, + 433, + 341 + ], + "spans": [ + { + "bbox": [ + 316, + 274, + 433, + 341 + ], + "type": "image", + "image_path": "f33d20c3f784bfbe7f0ff3bb9781853c0752766e462530489b8bd01d0fee52ac.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 349, + 342, + 393, + 351 + ], + "lines": [ + { + "bbox": [ + 349, + 342, + 393, + 351 + ], + "spans": [ + { + "bbox": [ + 349, + 342, + 393, + 351 + ], + "type": "text", + "content": "Full model" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 435, + 274, + 553, + 341 + ], + "blocks": [ + { + "bbox": [ + 435, + 274, + 553, + 341 + ], + "lines": [ + { + "bbox": [ + 435, + 274, + 553, + 341 + ], + "spans": [ + { + "bbox": [ + 435, + 274, + 553, + 341 + ], + "type": "image", + "image_path": "ccb980a7b0c4ceb91f053141c76fe75402e5d4140022965b63d68daeb8f5ab6d.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 473, + 342, + 517, + 350 + ], + "lines": [ + { + "bbox": [ + 473, + 342, + 517, + 350 + ], + "spans": [ + { + "bbox": [ + 473, + 342, + 517, + 350 + ], + "type": "text", + "content": "Full model" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 316, + 354, + 433, + 420 + ], + "blocks": [ + { + "bbox": [ + 316, + 354, + 433, + 420 + ], + "lines": [ + { + "bbox": [ + 316, + 354, + 433, + 420 + ], + "spans": [ + { + "bbox": [ + 316, + 354, + 433, + 420 + ], + "type": "image", + "image_path": "cedd9e62c8c0dcc98d5badf8505cda37ed6e25f28bb5f30bef17e0e11f2e4ae2.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 361, + 421, + 386, + 430 + ], + "lines": [ + { + "bbox": [ + 361, + 421, + 386, + 430 + ], + "spans": [ + { + "bbox": [ + 361, + 421, + 386, + 430 + ], + "type": "text", + "content": "2DGS" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 442, + 555, + 464 + ], + "lines": [ + { + "bbox": [ + 313, + 442, + 555, + 464 + ], + "spans": [ + { + "bbox": [ + 313, + 442, + 555, + 464 + ], + "type": "text", + "content": "Figure 10. Qualitative comparison of ablation study on 2DGS and without Physical Normal Estimation." + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_caption" + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 435, + 354, + 553, + 420 + ], + "blocks": [ + { + "bbox": [ + 435, + 354, + 553, + 420 + ], + "lines": [ + { + "bbox": [ + 435, + 354, + 553, + 420 + ], + "spans": [ + { + "bbox": [ + 435, + 354, + 553, + 420 + ], + "type": "image", + "image_path": "2068735c5657228fe4714a273f581b1856a11fbe8005e1a250b0ef63e614d843.jpg" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 475, + 420, + 514, + 430 + ], + "lines": [ + { + "bbox": [ + 475, + 420, + 514, + 430 + ], + "spans": [ + { + "bbox": [ + 475, + 420, + 514, + 430 + ], + "type": "text", + "content": "w/o N.E." + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_caption" + } + ], + "index": 36 + }, + { + "bbox": [ + 313, + 487, + 555, + 546 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 487, + 555, + 546 + ], + "spans": [ + { + "bbox": [ + 313, + 487, + 555, + 546 + ], + "type": "text", + "content": "mal estimation and using the shortest axis of 3D Gaussians as the normal. This causes the normal loss " + }, + { + "bbox": [ + 313, + 487, + 555, + 546 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{normal}}" + }, + { + "bbox": [ + 313, + 487, + 555, + 546 + ], + "type": "text", + "content": " to directly supervise the shortest axes, making some axes deviate significantly to align with surface normals, resulting in degraded rendering quality." + } + ] + } + ], + "index": 39 + }, + { + "bbox": [ + 313, + 558, + 387, + 570 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 387, + 570 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 387, + 570 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 40 + }, + { + "bbox": [ + 313, + 578, + 555, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 578, + 555, + 662 + ], + "spans": [ + { + "bbox": [ + 313, + 578, + 555, + 662 + ], + "type": "text", + "content": "SpectroMotion enhances 3D Gaussian Splatting for dynamic specular scenes by combining specular rendering with deformation fields. Using normal residual correction, coarse-to-fine training, and a deformable environment map, it achieves superior accuracy and visual quality in novel view synthesis, outperforming existing methods while maintaining geometric consistency." + } + ] + } + ], + "index": 41 + }, + { + "bbox": [ + 313, + 665, + 556, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 556, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 556, + 712 + ], + "type": "text", + "content": "Limitations. Though we stabilize the entire scene's geometry using a coarse-to-fine training strategy, some failure cases still occur. Please refer to the supplementary materials for visual results of failure cases." + } + ] + } + ], + "index": 42 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21335" + } + ] + } + ], + "index": 43 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 156 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 156 + ], + "type": "text", + "content": "Acknowledgements. This research was funded by the National Science and Technology Council, Taiwan, under Grants NSTC 112-2222-E-A49-004-MY2 and 113-2628-E-A49-023-. The authors are grateful to Google, NVIDIA, and MediaTek Inc. for their generous donations. Yu-Lun Liu acknowledges the Yushan Young Fellow Program by the MOE in Taiwan." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 167, + 115, + 179 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 167, + 115, + 179 + ], + "spans": [ + { + "bbox": [ + 56, + 167, + 115, + 179 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 186, + 297, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 186, + 297, + 231 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 186, + 297, + 231 + ], + "spans": [ + { + "bbox": [ + 61, + 186, + 297, + 231 + ], + "type": "text", + "content": "[1] Benjamin Attal, Eliot Laidlaw, Aaron Gokaslan, Changil Kim, Christian Richardt, James Tompkin, and Matthew O'Toole. Törf: Time-of-flight radiance fields for dynamic scene view synthesis, 2021. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 232, + 296, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 232, + 296, + 275 + ], + "spans": [ + { + "bbox": [ + 61, + 232, + 296, + 275 + ], + "type": "text", + "content": "[2] Sai Bi, Zexiang Xu, Pratul Srinivasan, Ben Mildenhall, Kalyan Sunkavalli, Miloš Hasan, Yannick Hold-Geoffroy, David Kriegman, and Ravi Ramamoorthi. Neural reflectance fields for appearance acquisition, 2020. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 277, + 296, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 277, + 296, + 319 + ], + "spans": [ + { + "bbox": [ + 62, + 277, + 296, + 319 + ], + "type": "text", + "content": "[3] Mark Boss, Varun Jampani, Raphael Braun, Ce Liu, Jonathan T. Barron, and Hendrik P. A. Lensch. Neural-pil: Neural pre-integrated lighting for reflectance decomposition, 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 321, + 296, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 321, + 296, + 365 + ], + "spans": [ + { + "bbox": [ + 62, + 321, + 296, + 365 + ], + "type": "text", + "content": "[4] Ting-Hsuan Chen, Jie Wen Chan, Hau-Shiang Shiu, Shih-Han Yen, Changhan Yeh, and Yu-Lun Liu. Narcan: Natural refined canonical image with integration of diffusion prior for video editing. In NeurIPS, 2024. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 366, + 296, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 366, + 296, + 399 + ], + "spans": [ + { + "bbox": [ + 62, + 366, + 296, + 399 + ], + "type": "text", + "content": "[5] Robert L Cook and Kenneth E. Torrance. A reflectance model for computer graphics. ACM Transactions on Graphics (ToG), 1(1):7-24, 1982. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 400, + 296, + 455 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 400, + 296, + 455 + ], + "spans": [ + { + "bbox": [ + 62, + 400, + 296, + 455 + ], + "type": "text", + "content": "[6] Ainaz Eftekhar, Alexander Sax, Jitendra Malik, and Amir Zamir. Omnidata: A scalable pipeline for making multi-task mid-level vision datasets from 3d scans. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10786-10796, 2021. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 456, + 296, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 456, + 296, + 499 + ], + "spans": [ + { + "bbox": [ + 62, + 456, + 296, + 499 + ], + "type": "text", + "content": "[7] Jian Gao, Chun Gu, Youtian Lin, Hao Zhu, Xun Cao, Li Zhang, and Yao Yao. Relightable 3d gaussian: Real-time point cloud relighting with brdf decomposition and ray tracing. arXiv preprint arXiv:2311.16043, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 501, + 296, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 501, + 296, + 544 + ], + "spans": [ + { + "bbox": [ + 62, + 501, + 296, + 544 + ], + "type": "text", + "content": "[8] Wenhang Ge, Tao Hu, Haoyu Zhao, Shu Liu, and Ying-Cong Chen. Ref-neus: Ambiguity-reduced neural implicit surface learning for multi-view reconstruction with reflection, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 545, + 296, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 545, + 296, + 589 + ], + "spans": [ + { + "bbox": [ + 62, + 545, + 296, + 589 + ], + "type": "text", + "content": "[9] Xiang Guo, Jiadai Sun, Yuchao Dai, Guanying Chen, Xiaoqing Ye, Xiao Tan, Errui Ding, Yumeng Zhang, and Jingdong Wang. Forward flow for novel view synthesis of dynamic scenes, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 590, + 296, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 590, + 296, + 635 + ], + "spans": [ + { + "bbox": [ + 57, + 590, + 296, + 635 + ], + "type": "text", + "content": "[10] Binbin Huang, Zehao Yu, Anpei Chen, Andreas Geiger, and Shenghua Gao. 2d gaussian splatting for geometrically accurate radiance fields. In ACM SIGGRAPH 2024 Conference Papers, pages 1-11, 2024. 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 636, + 296, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 636, + 296, + 669 + ], + "spans": [ + { + "bbox": [ + 57, + 636, + 296, + 669 + ], + "type": "text", + "content": "[11] Yi-Hua Huang, Yang-Tian Sun, Ziyi Yang, Xiaoyang Lyu, Yan-Pei Cao, and Xiaojuan Qi. Sc-gs: Sparse-controlled gaussian splatting for editable dynamic scenes, 2024. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 670, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 296, + 713 + ], + "type": "text", + "content": "[12] Yingwenqi Jiang, Jiadong Tu, Yuan Liu, Xifeng Gao, Xiaoxiao Long, Wenping Wang, and Yuexin Ma. Gaussianshader: 3d gaussian splatting with shading functions for reflective surfaces. arXiv preprint arXiv:2311.17977, 2023. 2, 4, 5, 6" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 316, + 73, + 555, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 555, + 137 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 555, + 137 + ], + "type": "text", + "content": "[13] Yingwenqi Jiang, Jiadong Tu, Yuan Liu, Xifeng Gao, Xiaoxiao Long, Wenping Wang, and Yuexin Ma. Gaussianshader: 3d gaussian splatting with shading functions for reflective surfaces. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5322-5332, 2024. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 139, + 555, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 139, + 555, + 172 + ], + "spans": [ + { + "bbox": [ + 316, + 139, + 555, + 172 + ], + "type": "text", + "content": "[14] James T Kajiya. The rendering equation. In Proceedings of the 13th annual conference on Computer graphics and interactive techniques, pages 143-150, 1986. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 173, + 555, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 173, + 555, + 204 + ], + "spans": [ + { + "bbox": [ + 317, + 173, + 555, + 204 + ], + "type": "text", + "content": "[15] Brian Karis and Epic Games. Real shading in unreal engine 4. Proc. Physically Based Shading Theory Practice, 4(3):1, 2013. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 205, + 554, + 238 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 205, + 554, + 238 + ], + "spans": [ + { + "bbox": [ + 316, + 205, + 554, + 238 + ], + "type": "text", + "content": "[16] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering, 2023. 1, 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 239, + 554, + 260 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 239, + 554, + 260 + ], + "spans": [ + { + "bbox": [ + 316, + 239, + 554, + 260 + ], + "type": "text", + "content": "[17] Junxuan Li and Hongdong Li. Neural reflectance for shape recovery with shadow handling, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 261, + 555, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 261, + 555, + 315 + ], + "spans": [ + { + "bbox": [ + 316, + 261, + 555, + 315 + ], + "type": "text", + "content": "[18] Tianye Li, Mira Slavcheva, Michael Zollhoefer, Simon Green, Christoph Lassner, Changil Kim, Tanner Schmidt, Steven Lovegrove, Michael Goesele, Richard Newcombe, and Zhaoyang Lv. Neural 3d video synthesis from multi-view video, 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 316, + 555, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 316, + 555, + 348 + ], + "spans": [ + { + "bbox": [ + 316, + 316, + 555, + 348 + ], + "type": "text", + "content": "[19] Zhengqi Li, Simon Niklaus, Noah Snively, and Oliver Wang. Neural scene flow fields for space-time view synthesis of dynamic scenes, 2021. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 350, + 555, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 350, + 555, + 392 + ], + "spans": [ + { + "bbox": [ + 316, + 350, + 555, + 392 + ], + "type": "text", + "content": "[20] Ruofan Liang, Huiting Chen, Chunlin Li, Fan Chen, Selvakumar Panneer, and Nandita Vijaykumar. Envidr: Implicit differentiable renderer with neural environment lighting, 2023. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 393, + 554, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 393, + 554, + 426 + ], + "spans": [ + { + "bbox": [ + 316, + 393, + 554, + 426 + ], + "type": "text", + "content": "[21] Ruofan Liang, Jiahao Zhang, Haoda Li, Chen Yang, Yushi Guan, and Nandita Vijaykumar. Spidr: Sdf-based neural point fields for illumination and deformation, 2023. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 426, + 555, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 426, + 555, + 470 + ], + "spans": [ + { + "bbox": [ + 316, + 426, + 555, + 470 + ], + "type": "text", + "content": "[22] Yiqing Liang, Numair Khan, Zhengqin Li, Thu Nguyen-Phuoc, Douglas Lanman, James Tompkin, and Lei Xiao. Gaufre: Gaussian deformation fields for real-time dynamic novel view synthesis, 2023. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 472, + 554, + 504 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 472, + 554, + 504 + ], + "spans": [ + { + "bbox": [ + 316, + 472, + 554, + 504 + ], + "type": "text", + "content": "[23] Zhihao Liang, Qi Zhang, Ying Feng, Ying Shan, and Kui Jia. Gs-ir: 3d gaussian splatting for inverse rendering. arXiv preprint arXiv:2311.16473, 2023. 2, 5, 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 505, + 554, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 505, + 554, + 536 + ], + "spans": [ + { + "bbox": [ + 316, + 505, + 554, + 536 + ], + "type": "text", + "content": "[24] Haotong Lin, Sida Peng, Zhen Xu, Yunzhi Yan, Qing Shuai, Hujun Bao, and Xiaowei Zhou. Efficient neural radiance fields for interactive free-viewpoint video, 2022. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 537, + 554, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 537, + 554, + 581 + ], + "spans": [ + { + "bbox": [ + 317, + 537, + 554, + 581 + ], + "type": "text", + "content": "[25] Yuan Liu, Peng Wang, Cheng Lin, Xiaoxiao Long, Jiepeng Wang, Lingjie Liu, Taku Komura, and Wenping Wang. Nano: Neural geometry and brdf reconstruction of reflective objects from multiview images, 2023. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 582, + 555, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 582, + 555, + 624 + ], + "spans": [ + { + "bbox": [ + 316, + 582, + 555, + 624 + ], + "type": "text", + "content": "[26] Yu-Lun Liu, Chen Gao, Andreas Meuleman, Hung-Yu Tseng, Ayush Saraf, Changil Kim, Yung-Yu Chuang, Johannes Kopf, and Jia-Bin Huang. Robust dynamic radiance fields. In CVPR, 2023. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 625, + 555, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 625, + 555, + 667 + ], + "spans": [ + { + "bbox": [ + 316, + 625, + 555, + 667 + ], + "type": "text", + "content": "[27] Caoyuan Ma, Yu-Lun Liu, Zhixiang Wang, Wu Liu, Xinchen Liu, and Zheng Wang. Humannerf-se: A simple yet effective approach to animate humannerf with diverse poses. In CVPR, 2024. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 669, + 555, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 669, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 669, + 555, + 713 + ], + "type": "text", + "content": "[28] Li Ma, Vasu Agrawal, Haithem Turki, Changil Kim, Chen Gao, Pedro Sander, Michael Zollhöfer, and Christian Richardt. Specnerf: Gaussian directional encoding for specular reflections. arXiv preprint arXiv:2312.13102, 2023. 2" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21336" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 297, + 117 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 297, + 117 + ], + "type": "text", + "content": "[29] Marko Mihajlovic, Sergey Prokudin, Siyu Tang, Robert Maier, Federica Bogo, Tony Tung, and Edmond Boyer. Splatfields: Neural gaussian splats for sparse 3d and 4d reconstruction. arXiv preprint arXiv:2409.11211, 2024. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 296, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 296, + 185 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 296, + 185 + ], + "type": "text", + "content": "[30] Jacob Munkberg, Jon Hasselgren, Tianchang Shen, Jun Gao, Wenzheng Chen, Alex Evans, Thomas Müller, and Sanja Fidler. Extracting Triangular 3D Models, Materials, and Lighting From Images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8280-8290, 2022. 2, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 185, + 296, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 185, + 296, + 228 + ], + "spans": [ + { + "bbox": [ + 56, + 185, + 296, + 228 + ], + "type": "text", + "content": "[31] Keunhong Park, Utkarsh Sinha, Jonathan T. Barron, Sofien Bouaziz, Dan B Goldman, Steven M. Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields, 2021. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 230, + 296, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 230, + 296, + 285 + ], + "spans": [ + { + "bbox": [ + 56, + 230, + 296, + 285 + ], + "type": "text", + "content": "[32] Keunhong Park, Utkarsh Sinha, Peter Hedman, Jonathan T. Barron, Sofien Bouaziz, Dan B Goldman, Ricardo MartinBrualla, and Steven M. Seitz. Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields, 2021. 2, 5, 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 285, + 296, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 285, + 296, + 319 + ], + "spans": [ + { + "bbox": [ + 56, + 285, + 296, + 319 + ], + "type": "text", + "content": "[33] Sida Peng, Yunzhi Yan, Qing Shuai, Hujun Bao, and Xiaowei Zhou. Representing volumetric videos as dynamic mlp maps, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 319, + 296, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 319, + 296, + 354 + ], + "spans": [ + { + "bbox": [ + 56, + 319, + 296, + 354 + ], + "type": "text", + "content": "[34] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-nerf: Neural radiance fields for dynamic scenes, 2020. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 354, + 296, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 354, + 296, + 409 + ], + "spans": [ + { + "bbox": [ + 56, + 354, + 296, + 409 + ], + "type": "text", + "content": "[35] Yahao Shi, Yanmin Wu, Chenming Wu, Xing Liu, Chen Zhao, Haocheng Feng, Jingtuo Liu, Liangjun Zhang, Jian Zhang, Bin Zhou, et al. Gir: 3d gaussian inverse rendering for relightable scene factorization. arXiv preprint arXiv:2312.05133, 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 410, + 296, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 410, + 296, + 453 + ], + "spans": [ + { + "bbox": [ + 56, + 410, + 296, + 453 + ], + "type": "text", + "content": "[36] Liangchen Song, Anpei Chen, Zhong Li, Zhang Chen, Lele Chen, Junsong Yuan, Yi Xu, and Andreas Geiger. Nerfplayer: A streamable dynamic scene representation with decomposed neural radiance fields, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 455, + 296, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 455, + 296, + 499 + ], + "spans": [ + { + "bbox": [ + 56, + 455, + 296, + 499 + ], + "type": "text", + "content": "[37] Pratul P. Srinivasan, Boyang Deng, Xiuming Zhang, Matthew Tancik, Ben Mildenhall, and Jonathan T. Barron. Nerv: Neural reflectance and visibility fields for relighting and view synthesis, 2020. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 500, + 296, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 500, + 296, + 545 + ], + "spans": [ + { + "bbox": [ + 56, + 500, + 296, + 545 + ], + "type": "text", + "content": "[38] Colton Stearns, Adam Harley, Mikaela Uy, Florian Dubost, Federico Tombari, Gordon Wetzstein, and Leonidas Guibas. Dynamic gaussian marbles for novel view synthesis of casual monocular videos. arXiv preprint arXiv:2406.18717, 2024. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 545, + 296, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 545, + 296, + 590 + ], + "spans": [ + { + "bbox": [ + 56, + 545, + 296, + 590 + ], + "type": "text", + "content": "[39] Edgar Tretschk, Ayush Tewari, Vladislav Golyanik, Michael Zollhöfer, Christoph Lassner, and Christian Theobalt. Non-rigid neural radiance fields: Reconstruction and novel view synthesis of a dynamic scene from monocular video, 2021. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 590, + 296, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 590, + 296, + 647 + ], + "spans": [ + { + "bbox": [ + 56, + 590, + 296, + 647 + ], + "type": "text", + "content": "[40] Dor Verbin, Peter Hedman, Ben Mildenhall, Todd Zickler, Jonathan T Barron, and Pratul P Srinivasan. Ref-nerf: Structured view-dependent appearance for neural radiance fields. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5481–5490. IEEE, 2022. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 647, + 296, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 296, + 712 + ], + "type": "text", + "content": "[41] Dor Verbin, Ben Mildenhall, Peter Hedman, Jonathan T Barron, Todd Zickler, and Pratul P Srinivasan. Eclipse: Disambiguating illumination and materials using unintended shadows. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 77-86, 2024. 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 72, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 315, + 72, + 555, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 72, + 555, + 127 + ], + "spans": [ + { + "bbox": [ + 315, + 72, + 555, + 127 + ], + "type": "text", + "content": "[42] Dor Verbin, Pratul P Srinivasan, Peter Hedman, Ben Mildenhall, Benjamin Attal, Richard Szeliski, and Jonathan T Barron. Nerf-casting: Improved view-dependent appearance with consistent reflections. arXiv preprint arXiv:2405.14871, 2024. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 315, + 128, + 555, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 128, + 555, + 163 + ], + "spans": [ + { + "bbox": [ + 315, + 128, + 555, + 163 + ], + "type": "text", + "content": "[43] Bruce Walter, Stephen R Marschner, Hongsong Li, and Kenneth E Torrance. Microfacet models for refraction through rough surfaces. Rendering techniques, 2007:18th, 2007. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 315, + 163, + 555, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 163, + 555, + 206 + ], + "spans": [ + { + "bbox": [ + 315, + 163, + 555, + 206 + ], + "type": "text", + "content": "[44] Qianqian Wang, Vickie Ye, Hang Gao, Jake Austin, Zhengqi Li, and Angjoo Kanazawa. Shape of motion: 4d reconstruction from a single video. arXiv preprint arXiv:2407.13764, 2024. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 207, + 555, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 207, + 555, + 262 + ], + "spans": [ + { + "bbox": [ + 316, + 207, + 555, + 262 + ], + "type": "text", + "content": "[45] Chun-Hung Wu, Shih-Hong Chen, Chih-Yao Hu, Hsin-Yu Wu, Kai-Hsin Chen, Yu-You Chen, Chih-Hai Su, Chih-Kuo Lee, and Yu-Lun Liu. Denver: Deformable neural vessel representations for unsupervised video vessel segmentation. In CVPR, 2025. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 262, + 555, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 262, + 555, + 308 + ], + "spans": [ + { + "bbox": [ + 316, + 262, + 555, + 308 + ], + "type": "text", + "content": "[46] Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Xinggang Wang. 4d gaussian splatting for real-time dynamic scene rendering. arXiv preprint arXiv:2310.08528, 2023. 2, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 308, + 555, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 308, + 555, + 342 + ], + "spans": [ + { + "bbox": [ + 316, + 308, + 555, + 342 + ], + "type": "text", + "content": "[47] Wenqi Xian, Jia-Bin Huang, Johannes Kopf, and Changil Kim. Space-time neural irradiance fields for free-viewpoint video, 2021. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 342, + 554, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 342, + 554, + 388 + ], + "spans": [ + { + "bbox": [ + 316, + 342, + 554, + 388 + ], + "type": "text", + "content": "[48] Zhiwen Yan, Chen Li, and Gim Hee Lee. Nerf-ds: Neural radiance fields for dynamic specular objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8285-8295, 2023. 5, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 388, + 554, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 388, + 554, + 432 + ], + "spans": [ + { + "bbox": [ + 316, + 388, + 554, + 432 + ], + "type": "text", + "content": "[49] Ziyi Yang, Xinyu Gao, Wen Zhou, Shaohui Jiao, Yuqing Zhang, and Xiaogang Jin. Deformable 3d gaussians for high-fidelity monocular dynamic scene reconstruction. arXiv preprint arXiv:2309.13101, 2023. 2, 3, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 432, + 554, + 478 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 432, + 554, + 478 + ], + "spans": [ + { + "bbox": [ + 316, + 432, + 554, + 478 + ], + "type": "text", + "content": "[50] Ziyi Yang, Xinyu Gao, Yangtian Sun, Yihua Huang, Xiaoyang Lyu, Wen Zhou, Shaohui Jiao, Xiaojuan Qi, and Xiaogang Jin. Spec-gaussian: Anisotropic view-dependent appearance for 3d gaussian splatting, 2024. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 478, + 555, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 478, + 555, + 512 + ], + "spans": [ + { + "bbox": [ + 316, + 478, + 555, + 512 + ], + "type": "text", + "content": "[51] Keyang Ye, Qiming Hou, and Kun Zhou. 3d gaussian splatting with deferred reflection. In ACM SIGGRAPH 2024 Conference Papers, pages 1-10, 2024. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 512, + 554, + 556 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 512, + 554, + 556 + ], + "spans": [ + { + "bbox": [ + 316, + 512, + 554, + 556 + ], + "type": "text", + "content": "[52] Wang Yifan, Felice Serena, Shihao Wu, Cengiz Oztireli, and Olga Sorkine-Hornung. Differentiable surface splatting for point-based geometry processing. ACM Transactions on Graphics (TOG), 38(6):1-14, 2019. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 556, + 555, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 556, + 555, + 601 + ], + "spans": [ + { + "bbox": [ + 316, + 556, + 555, + 601 + ], + "type": "text", + "content": "[53] Jingyang Zhang, Yao Yao, Shiwei Li, Jingbo Liu, Tian Fang, David McKinnon, Yanghai Tsin, and Long Quan. Neilf++: Inter-reflectable light fields for geometry and material estimation, 2023. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 601, + 555, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 601, + 555, + 658 + ], + "spans": [ + { + "bbox": [ + 316, + 601, + 555, + 658 + ], + "type": "text", + "content": "[54] Kai Zhang, Fujun Luan, Qianqian Wang, Kavita Bala, and Noah Snavely. Physg: Inverse rendering with spherical gaussians for physics-based material editing and relighting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5453-5462, 2021. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 658, + 555, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 555, + 713 + ], + "type": "text", + "content": "[55] Xiuming Zhang, Pratul P. Srinivasan, Boyang Deng, Paul Debevec, William T. Freeman, and Jonathan T. Barron. Ner-factor: neural factorization of shape and reflectance under an unknown illumination. ACM Transactions on Graphics, 40 (6):1-18, 2021." + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21337" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 298, + 152 + ], + "type": "list", + "angle": 0, + "index": 2, + "blocks": [ + { + "bbox": [ + 56, + 72, + 298, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 298, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 298, + 116 + ], + "type": "text", + "content": "[56] Xiaoming Zhao, Pratul P Srinivasan, Dor Verbin, Keunhong Park, Ricardo Martin Brualla, and Philipp Henzler. Illuminerf: 3d relighting without inverse rendering. arXiv preprint arXiv:2406.06527, 2024. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 297, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 297, + 152 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 297, + 152 + ], + "type": "text", + "content": "[57] Zuo-Liang Zhu, Beibei Wang, and Jian Yang. Gs-ror: 3d gaussian splatting for reflective object relighting via sdf priors. arXiv preprint arXiv:2406.18544, 2024. 2" + } + ] + } + ], + "index": 1 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "21338" + } + ] + } + ], + "index": 3 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Speedy-Splat_ Fast 3D Gaussian Splatting with Sparse Pixels and Sparse Primitives/08166209-021b-4f8f-8611-9f7790e8bc2b_content_list.json b/2025/Speedy-Splat_ Fast 3D Gaussian Splatting with Sparse Pixels and Sparse Primitives/08166209-021b-4f8f-8611-9f7790e8bc2b_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..7bd62c7bf7395836edf914ff5492d1f0e1551c11 --- /dev/null +++ b/2025/Speedy-Splat_ Fast 3D Gaussian Splatting with Sparse Pixels and Sparse Primitives/08166209-021b-4f8f-8611-9f7790e8bc2b_content_list.json @@ -0,0 +1,2151 @@ +[ + { + "type": "text", + "text": "Speedy-Splat: Fast 3D Gaussian Splating with Sparse Pixels and Sparse Primitives", + "text_level": 1, + "bbox": [ + 261, + 130, + 736, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Alex Hanson", + "bbox": [ + 253, + 203, + 359, + 219 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Matthias Zwicker", + "bbox": [ + 325, + 222, + 468, + 237 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Allen Tu", + "bbox": [ + 405, + 205, + 478, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Geng Lin", + "bbox": [ + 522, + 204, + 601, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Vasu Singla", + "bbox": [ + 647, + 204, + 743, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Tom Goldstein", + "bbox": [ + 553, + 222, + 671, + 237 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "University of Maryland, College Park", + "bbox": [ + 349, + 250, + 647, + 267 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/80c8529a039791fdbf6f182b0191918f7f9585ea1fe19c1b57530b82780aa8d3.jpg", + "image_caption": [ + "GT" + ], + "image_footnote": [], + "bbox": [ + 94, + 300, + 356, + 412 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/22afe16fab5e67b7e86920db6ee04ce9f8c1b3df5ce2042922144023e376646d.jpg", + "image_caption": [ + "3DGS", + "2.6M Gaussians", + "25.39 PSNR", + "184.87 FPS", + "Figure 1. We reduce the number of Gaussians by over $90\\%$ , only marginally decrease PSNR, and accelerate rendering speed by $6.2\\times$ in the Tanks & Temples truck scene when compared to 3D Gaussian Splatting (3D-GS). Additionally, we speed up training time by $1.38\\times$ ." + ], + "image_footnote": [], + "bbox": [ + 364, + 300, + 624, + 412 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/78fb40f2c812639d67118378d4bc2cff166da1aa572f4c0b2d66ca5de40ec0c8.jpg", + "image_caption": [ + "Ours", + "0.26M Gaussians", + "25.34 PSNR", + "1148.77 FPS" + ], + "image_footnote": [], + "bbox": [ + 630, + 299, + 893, + 412 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 477, + 326, + 492 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3D Gaussian Splatting (3D-GS) is a recent 3D scene reconstruction technique that enables real-time rendering of novel views by modeling scenes as parametric point clouds of differentiable 3D Gaussians. However, its rendering speed and model size still present bottlenecks, especially in resource-constrained settings. In this paper, we identify and address two key inefficiencies in 3D-GS to substantially improve rendering speed. These improvements also yield the ancillary benefits of reduced model size and training time. First, we optimize the rendering pipeline to precisely localize Gaussians in the scene, boosting rendering speed without altering visual fidelity. Second, we introduce a novel pruning technique and integrate it into the training pipeline, significantly reducing model size and training time while further raising rendering speed. Our Speedy-Splat approach combines these techniques to accelerate average rendering speed by a drastic $6.71 \\times$ across scenes from the Mip-NeRF 360, Tanks & Temples, and Deep Blending datasets. Code and additional visualizations are available at https://speedysplat.github.io/.", + "bbox": [ + 88, + 508, + 485, + 811 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 830, + 220, + 844 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Fast rendering of photorealistic novel views has been a long-standing goal in computer vision and graphics. Neural Radiance Fields (NeRF) [23] and its variants have made sig-", + "bbox": [ + 89, + 854, + 483, + 902 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "nificant strides in photorealistic 3D scene reconstruction by representing scenes as continuous neural volumetric models that encode scene density and color at spatial coordinates. However, despite recent efforts [4, 24, 29], fast rendering in NeRF remains challenging because the volumetric sampling used in ray-marching is computationally expensive. Recently, 3D Gaussian Splatting (3D-GS) [14] has emerged as a promising alternative that enables real-time rendering by modeling scenes as parametric point clouds of differentiable 3D Gaussians. Nevertheless, its rendering speed is still limited by high parameter counts and certain algorithmic inefficiencies. Efficient rendering is essential in applications such as virtual reality, networked systems, and multi-view streaming. Real-time rendering on resource-constrained edge devices, such as mobile phones, has yet to be achieved [19].", + "bbox": [ + 511, + 478, + 906, + 719 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Although recent works on compressing 3D-GS models [6, 7, 11, 25] achieve some speed-ups by reducing the number of parameters, few approaches directly target rendering speed [8, 19]. In this paper, we specifically address this gap by demonstrating that the rendering speed of 3D-GS models can be drastically increased while maintaining competitive image quality. Additionally, we show that our methods reduce training time and substantially decrease model size.", + "bbox": [ + 511, + 727, + 908, + 849 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We begin by observing that the cost of 3D-GS rendering is proportional to both the number of Gaussians in the scene and the number of pixels processed per Gaussian. Our ap-", + "bbox": [ + 511, + 854, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "21537", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "proach optimizes both factors. First, we reduce the number of pixels that are processed for each Gaussian by efficiently and accurately localizing it in the rendered image. Second, we reduce the total number of Gaussians in the model through a novel approach that maintains rendering quality.", + "bbox": [ + 89, + 90, + 480, + 167 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3D-GS implements tile rendering to localize Gaussians in the image plane by assigning each Gaussian to the tiles that it intersects. We find the existing algorithm to be overly conservative; in Section 4.1.1, we address this by introducing our SnugBox algorithm to precisely localize Gaussians by computing a tight bounding box around their extent. Then, in Section 4.1.2, we extend SnugBox with our AccuTile algorithm to identify exact tile intersections. Both approaches are plug-and-play, lead to respective inference speed-ups of $1.82 \\times$ and $1.99 \\times$ on average, and do not change the 3D-GS renderings.", + "bbox": [ + 88, + 167, + 480, + 332 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To reduce the total number of Gaussians while preserving visual fidelity, we extend an existing pruning method [11] by reducing its memory requirement by $36\\times$ and incorporating it into the 3D-GS training pipeline. During the densification stage of 3D-GS training, Gaussians are regularly replicated and pruned. In Section 4.2.2, we augment the densification stage with our Soft Pruning method to prune $80\\%$ of Gaussians at regular intervals. After the densification stage, our Hard Pruning method, described in Section 4.2.3, prunes an additional $30\\%$ of Gaussians at set intervals.", + "bbox": [ + 89, + 333, + 482, + 498 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our Speedy-Splat approach integrates both techniques into the 3D-GS training pipeline. On average, rendering speed is accelerated by $6.71 \\times$ , model size is reduced by $10.6 \\times$ , and training speed is improved by $1.4 \\times$ across all evaluated scenes while maintaining high image quality.", + "bbox": [ + 89, + 500, + 482, + 575 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, we propose the following contributions:", + "bbox": [ + 109, + 577, + 460, + 590 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. SnugBox: A precise algorithm for computing Gaussian-tile bounding box intersections.", + "2. AccuTile: An extension of SnugBox for computing exact Gaussian-tile intersections.", + "3. Soft Pruning: An augmentation for pruning Gaussians during densification.", + "4. Hard Pruning: An augmentation for pruning Gaussians post-densification." + ], + "bbox": [ + 89, + 595, + 482, + 715 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 89, + 734, + 227, + 750 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The real-time rendering speed of 3D-GS [14] on desktop GPUs has inspired research focused on further accelerating both its training and inference in resource-constrained environments. In this section, we review related works that specifically target these performance improvements.", + "bbox": [ + 89, + 761, + 482, + 838 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Pruning", + "text_level": 1, + "bbox": [ + 89, + 848, + 191, + 864 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "A large portion of Gaussians in vanilla 3D-GS models are redundant [6, 7, 25], motivating a recent line of", + "bbox": [ + 89, + 869, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "work focused on pruning Gaussians from 3D-GS models to boost rendering speed with minimal loss of visual fidelity [1, 2, 6, 7, 10, 11, 18, 19, 25]. Nearly all approaches assign a significance score to each Gaussian that is used to rank and prune them. Several works compute the aggregated ray contribution for each Gaussian across all input images [6, 7, 18, 25], while others combine opacity with additional information, such as gradients per Gaussian, to calculate their pruning criterion [1, 2, 20]. Papantonakis et al. [26] use resolution and scale-aware redundancy metrics, EAGLES [10] calculates the total transmittance per Gaussian, and Lin et al. [19] accelerate inference speed through an efficiency-aware strategy and foveated rendering.", + "bbox": [ + 511, + 90, + 903, + 287 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "PUP 3D-GS [11] introduces a more principled approach by deriving a Hessian for each Gaussian that represents its sensitivity to the reconstruction error. While PUP 3D-GS achieves state-of-the-art post-hoc pruning results, computing its sensitivity score incurs considerable storage requirements that limit its viability for use during training. Our pruning approach directly extends PUP 3D-GS by improving its memory efficiency by $36\\times$ .", + "bbox": [ + 511, + 287, + 903, + 409 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Other Methods", + "text_level": 1, + "bbox": [ + 511, + 420, + 666, + 434 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In addition to pruning, several other strategies have been explored to enhance the rendering and training speed of 3D-GS. Mini-Splatting [7] modifies its densification strategies and adds a simplification stage to constrain the number of Gaussians. 3DGS-MCMC [15] models training dynamics as an MCMC process. Revisiting Densification [30] introduces an error-based densification strategy. 3DGS-LM [13] replaces Adam [16] with a tailored Levenberg-Marquardt optimizer [9]. Taming 3DGS [22] proposes a constructive optimization process that limits the number of Gaussians to a pre-defined threshold set by the user. DISTWAR [5] dives into the low-level implementation of GPU thread scheduling and optimizes atomic processing with a novel primitive. StopThePop [27] introduces a precise tile intersection method, but we find that our approach is notably faster in Appendix A.4. FlashGS [8] implements a tile intersection method similar to StopThePop. Several works reduce training time and memory requirements by enforcing geometric constraints [21, 28, 31, 32]. Most of these approaches are orthogonal to ours and can be applied alongside it.", + "bbox": [ + 511, + 443, + 906, + 746 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Background", + "text_level": 1, + "bbox": [ + 511, + 761, + 640, + 777 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3.1. 3D Gaussian Splitting Overview", + "text_level": 1, + "bbox": [ + 511, + 787, + 799, + 803 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3D Gaussian Splatting (3D-GS) [14] models scenes as parametric, point-based representations that use differentiable 3D Gaussians as primitives. Given a set of ground truth training images $\\mathcal{I}_{gt} = \\{\\pmb{I}_i\\in \\mathbb{R}^{H\\times W}\\}_{i = 1}^K$ , the scene is initialized by using Structure from Motion (SfM) to produce a sparse point cloud that serves as the initial means for the", + "bbox": [ + 511, + 809, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "21538", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "table", + "img_path": "images/72327e30568d0ffb3f3c89e03eb332008a3ae91c21bd6b6aed453f5a17def6ea.jpg", + "table_caption": [ + "Table 1. Average execution time (milliseconds) of each function across all scenes in Section 5.1. The operation in each row is applied cumulatively to all of the following rows. For each model, accurate measurements are collected by averaging execution times across three runs that each render the test set 20 times to reduce variance. The fastest and second fastest times are color coded." + ], + "table_footnote": [], + "table_body": "
MethodPreprocessInclusive SumDuplicate with KeysRadix SortIdentify Tile RangesRenderOverall
Baseline0.6650.0450.5701.5510.0824.4837.478
+SnugBox0.6560.0460.208 (2.738×)0.729 (2.126×)0.041 (1.980×)2.344 (1.913×)4.102 (1.823×)
+AccuTile0.6680.0460.221 (2.575×)0.612 (2.533×)0.035 (2.326×)2.062 (2.175×)3.748 (1.995×)
+Soft Pruning0.370 (1.798×)0.030 (1.494×)0.146 (3.906×)0.404 (3.843×)0.024 (3.422×)1.337 (3.354×)2.381 (3.141×)
+Hard Pruning0.091 (7.293×)0.016 (2.769×)0.090 (6.325×)0.215 (7.217×)0.013 (6.537×)0.619 (7.247×)1.114 (6.712×)
", + "bbox": [ + 93, + 142, + 900, + 229 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3D Gaussians. The estimated camera poses $\\mathcal{P}_{gt} = \\{\\phi_i\\in$ $\\mathbb{R}^{3\\times 4}\\}_{i = 1}^{K}$ are paired with their corresponding images and are used to optimize the scene.", + "bbox": [ + 89, + 247, + 482, + 291 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Each 3D Gaussian primitive $\\mathcal{G}_i$ is parameterized by three geometry parameters - mean $\\mu_i\\in \\mathbb{R}^3$ , scale $s_i\\in \\mathbb{R}^3$ , and rotation $r_i\\in \\mathbb{R}^4$ - and two color parameters - view-dependent spherical harmonics $h_i\\in \\mathbb{R}^{16\\times 3}$ and opacity $\\sigma_{i}\\in \\mathbb{R}$ . The set of all parameters can be described as:", + "bbox": [ + 89, + 292, + 483, + 369 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {G} = \\left\\{\\mathcal {G} _ {i} = \\left\\{\\mu_ {i}, s _ {i}, r _ {i}, h _ {i}, \\sigma_ {i} \\right\\} \\right\\} _ {i = 1} ^ {N}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 168, + 380, + 482, + 398 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $N$ is the number of Gaussians in the model.", + "bbox": [ + 89, + 410, + 424, + 425 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given camera pose $\\phi$ , the scene is rendered by projecting all Gaussians to image space and applying alpha blending to each pixel. Models are optimized via stochastic gradient descent on image reconstruction losses:", + "bbox": [ + 89, + 426, + 483, + 486 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nL (\\mathcal {G} | \\phi , I _ {g t}) = \\| I _ {\\mathcal {G}} (\\phi) - I _ {g t} \\| _ {1} + L _ {\\mathrm {D - S S I M}} (I _ {\\mathcal {G}} (\\phi), I _ {g t}), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 500, + 482, + 517 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $I_{\\mathcal{G}(\\phi)}$ is the rendered image for pose $\\phi$", + "bbox": [ + 89, + 529, + 393, + 545 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "During optimization, the scene is periodically densified by cloning and splitting uncertain Gaussians and pruned by removing large and transparent Gaussians. The opacities of the Gaussians are also periodically reset.", + "bbox": [ + 89, + 545, + 483, + 604 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. 3D Gaussian Splitting Rendering Specifics", + "text_level": 1, + "bbox": [ + 89, + 618, + 454, + 635 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3D-GS uses a tile-based rendering strategy that divides the rendered image into $16 \\times 16$ pixel tiles. Each Gaussian is projected into image space, where its intersection with these tiles is computed. Then, these Gaussian-to-tile mappings are sorted to collect and order the Gaussians by depth for pixel-wise rendering.", + "bbox": [ + 89, + 641, + 482, + 731 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Rendering runtime is dominated by six key functions. Table 1 empirically analyzes the execution time of each function and highlights the improvements achieved by our methods. Descriptions of each function are provided in the following sections.", + "bbox": [ + 89, + 732, + 483, + 808 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2.1. Preprocessing", + "text_level": 1, + "bbox": [ + 89, + 820, + 235, + 835 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The preprocess kernel is parallelized such that each thread processes a single Gaussian $\\mathcal{G}_i$ . It computes a 2D projection of $\\mathcal{G}_i$ to image space and obtains a count of tiles that it intersects.", + "bbox": [ + 89, + 839, + 483, + 898 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The mean $\\mu_{i}$ is projected to image space using a viewing transform $W$ and a perspective projection, yielding 2D mean $\\mu_{i_{2D}}$ and depth that are stored for later processing. The scale $s_i$ and rotation $r_i$ parameters are converted to the diagonal scale $S_{i}$ and rotation $R_{i}$ matrices. The 3D covariance is then defined as:", + "bbox": [ + 511, + 247, + 906, + 335 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\Sigma} _ {i _ {3 D}} = \\boldsymbol {R} _ {i} \\boldsymbol {S} _ {i} \\boldsymbol {S} _ {i} ^ {T} \\boldsymbol {R} _ {i} ^ {T}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 633, + 353, + 903, + 372 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "which is projected via:", + "bbox": [ + 532, + 383, + 686, + 398 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\boldsymbol {\\Sigma}} _ {i _ {3 D}} = \\boldsymbol {J} \\boldsymbol {W} \\boldsymbol {\\Sigma} _ {i _ {3 D}} \\boldsymbol {W} ^ {T} \\boldsymbol {J} ^ {T}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 612, + 414, + 903, + 431 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $J$ is the Jacobian of the first order approximation of the perspective projection. Dropping the last row and column of $\\hat{\\Sigma}_{i_{3D}}$ gives $\\Sigma_{i_{2D}}$ [14, 33]. The largest eigenvalue of $\\Sigma_{i_{2D}}$ is used to compute the count of tiles intersected by this Gaussian $\\mathcal{G}_i$ as shown in Figure 2a. $\\Sigma_{i_{2D}}^{-1}$ is computed, along with a view-dependent color $c_{i}$ , derived from $W$ and $h_i$ . All three values are stored for later processing.", + "bbox": [ + 511, + 443, + 905, + 549 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2.2. Sorting", + "text_level": 1, + "bbox": [ + 513, + 559, + 612, + 574 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "After preprocessing, four functions process the Gaussians for pixel-wise rendering:", + "bbox": [ + 511, + 579, + 905, + 609 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- InclusiveSum: A CUDA primitive that computes the prefix sum over all Gaussian-tile counts to allocate key and value arrays for Gaussian-to-tile mapping.", + "- duplicateWithKeys: A Gaussian-parallel kernel that recomputes Gaussian-tile intersections to generate a key for each intersecting tile, consisting of the tile index and Gaussian depth.", + "- RadixSort: A CUDA primitive that sorts the key array, ordering Gaussian indices by tile and then depth.", + "- identifyTileRanges: A kernel parallelized across the key array to post-process keys before pixel rendering." + ], + "bbox": [ + 511, + 609, + 905, + 776 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2.3. Rendering", + "text_level": 1, + "bbox": [ + 513, + 786, + 633, + 801 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The render kernel is parallelized across pixels. For each pixel $p$ , all Gaussians within its corresponding tile are loaded and processed in depth order as determined by RadixSort. An alpha value:", + "bbox": [ + 511, + 806, + 905, + 867 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\alpha_ {i} (p) = \\sigma_ {i} g _ {i} (p) \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 651, + 885, + 903, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "21539", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "is computed for each Gaussian $\\mathcal{G}_i$ , where $g_i$ is the value of the projected 2D Gaussian at pixel $p$ :", + "bbox": [ + 89, + 90, + 482, + 122 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ng _ {i} = e ^ {q}, q = - \\frac {1}{2} (p - \\mu_ {i _ {2 D}}) \\boldsymbol {\\Sigma} _ {\\boldsymbol {i} _ {2 D}} ^ {- 1} (p - \\mu_ {i _ {2 D}}) ^ {T}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 111, + 128, + 482, + 159 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "If $\\alpha_{i} > \\frac{1}{255}$ , then the Gaussian is included in the alpha compositing of the pixel color $C$ , given by:", + "bbox": [ + 89, + 167, + 482, + 199 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nC (p) = \\sum_ {i \\in \\mathcal {N}} c _ {i} \\alpha_ {i} (p) \\prod_ {j = 1} ^ {i - 1} (1 - \\alpha_ {j} (p)). \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 161, + 209, + 482, + 252 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Methods", + "text_level": 1, + "bbox": [ + 89, + 261, + 191, + 277 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our Speedy-Splat methods are motivated by two key insights into inefficiencies within the 3D-GS rendering pipeline. First, Gaussian Splatting grossly overestimates the extent of Gaussians in the image. Second, as demonstrated by [11] and other recent pruning works, 3D-GS models are heavily overparameterized.", + "bbox": [ + 89, + 287, + 483, + 378 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Precise Tile Intersect", + "text_level": 1, + "bbox": [ + 89, + 386, + 289, + 401 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Gaussian Splitting identifies tiles intersected by Gaussian $\\mathcal{G}_i$ by calculating the maximum eigenvalue $\\lambda_{\\mathrm{max}}$ of its projected 2D covariance $\\Sigma_{i_2D}$ , then selecting all tiles that intersect the square inscribing the circle defined by center $\\mu_{i_{2D}}$ and radius:", + "bbox": [ + 89, + 407, + 483, + 484 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nr = \\left\\lceil 3 \\sqrt {\\lambda_ {\\max }} \\right\\rceil . \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 225, + 494, + 482, + 518 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This approach neglects opacity $\\sigma_{i}$ in its calculation and generally overestimates the Gaussian extent, as illustrated by Figure 2a. The actual extent of Gaussian $\\mathcal{G}_i$ , shown in Figure 2b, is given by the threshold placed on its alpha value $\\alpha_{i}$ . Specifically, $\\mathcal{G}_i$ does not contribute to the rendering of pixel $p$ if $\\alpha_{i} < \\frac{1}{255}$ for $p$ . By applying the actual extent in tile intersection calculations, we arrive at a far more concise set of intersected tiles.", + "bbox": [ + 89, + 523, + 483, + 643 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We now show the derivation of this extent. The furthest pixel extent of Gaussian $\\mathcal{G}_i$ can be determined by directly substituting this threshold into Equation 5. Rearranging terms gives:", + "bbox": [ + 89, + 643, + 483, + 704 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n2 \\log (2 5 5 \\sigma_ {i}) = (p - \\mu_ {i _ {2 D}}) \\boldsymbol {\\Sigma} _ {\\boldsymbol {i _ {2 D}}} ^ {- 1} (p - \\mu_ {i _ {2 D}}) ^ {T}. \\qquad (9)\n$$\n", + "text_format": "latex", + "bbox": [ + 114, + 712, + 482, + 732 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We can rewrite:", + "bbox": [ + 89, + 739, + 197, + 753 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\np = \\binom {p _ {x}} {p _ {y}}, \\mu_ {i _ {2 D}} = \\binom {\\mu_ {x}} {\\mu_ {y}}, \\boldsymbol {\\Sigma} _ {\\boldsymbol {i} _ {2 D}} ^ {- 1} = \\left( \\begin{array}{c c} a & b \\\\ b & c \\end{array} \\right). \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 106, + 761, + 482, + 796 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Specifying threshold $t$ and centered coordinates $x_{d}$ and $y_{d}$ :", + "bbox": [ + 89, + 803, + 482, + 818 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nt = 2 \\log (2 5 5 \\sigma_ {i}), \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 225, + 843, + 482, + 859 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nx _ {d} = p _ {x} - \\mu_ {x}, \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 235, + 863, + 482, + 878 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ny _ {d} = p _ {y} - \\mu_ {y}, \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 235, + 882, + 482, + 897 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "gives the pixel extent with coordinates $x_{d}$ and $y_{d}$ that satisfy the ellipse equation:", + "bbox": [ + 511, + 90, + 903, + 121 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nt = a x _ {d} ^ {2} + 2 b x _ {d} y _ {d} + c y _ {d} ^ {2}. \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 620, + 133, + 903, + 152 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our approach uses this pixel extent to reduce the number of tiles contained in the Gaussian-to-tile mappings for each Gaussian. We propose two methods for computing precise tile intersections. First, our SnugBox algorithm produces a tight bounding box around each Gaussian. Second, our AccuTile algorithm extends it to identify the exact set of tiles intersected by the Gaussian.", + "bbox": [ + 511, + 165, + 906, + 270 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1.1. SnugBox", + "text_level": 1, + "bbox": [ + 513, + 280, + 622, + 295 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our SnugBox method uses this elliptical extent to compute an axis-aligned bounding box that more precisely identifies tiles intersected by Gaussian $\\mathcal{G}_i$ . To derive this bounding box, we rearrange Equation 14 to solve for $y_d$ :", + "bbox": [ + 511, + 300, + 905, + 359 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ny _ {d} = \\frac {- b x _ {d} \\pm \\sqrt {\\left(b ^ {2} - a c\\right) x _ {d} ^ {2} + t c}}{c}. \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 372, + 903, + 406 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To find the $y$ -coordinate bounding box edges $y_{\\mathrm{min}}$ and $y_{\\mathrm{max}}$ , we identify the values of $x_d$ where $\\partial y_d / \\partial x_d = 0$ . We refer to these $x_d$ values as $x_{d_{\\mathrm{arg}s}}$ to specify that they are the $\\arg \\min y_d$ and $\\arg \\max y_d$ values. Differentiating Equation 15 and solving for $x_{d_{\\mathrm{arg}s}}$ gives:", + "bbox": [ + 511, + 415, + 905, + 494 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nx _ {d _ {\\text {a r g s}}} = \\pm \\sqrt {\\frac {- b ^ {2} t}{\\left(b ^ {2} - a c\\right) a}}. \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 620, + 505, + 903, + 546 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Substituting $x_{d_{\\text{args}}}$ into Equation 15 and adding $\\mu_y$ gives $y_{\\min}$ and $y_{\\max}$ . Due to the symmetry of Equation 14, we can find the $x$ -coordinate bounding box edges $x_{\\min}$ and $x_{\\max}$ by swapping $y_d$ and $x_d$ and constants $a$ and $c$ to rewrite Equations 15 and 16 in terms of $x_d$ and $y_{d_{\\text{args}}}$ .", + "bbox": [ + 511, + 553, + 905, + 630 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "After the bounding box edges are identified, our method follows 3D-GS and converts these edges to tile indices by dividing by tile size, rounding, and clipping to the image boundary. As depicted in Figure 2, Snugbox can produce a significantly tighter bounding box than 3D-GS. Meanwhile, its computational overhead is small as it performs a constant number of operations and is only called twice in the rendering pipeline – once to count the Gaussians intersecting tiles in preprocess and once to populate the Gaussian-to-tile arrays in duplicateWithkeys. Table 1 reports that SnugBox improves the efficiency of all downstream functions and produces an average overall speed-up of $1.82 \\times$ .", + "bbox": [ + 511, + 630, + 906, + 811 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1.2. AccuTile", + "text_level": 1, + "bbox": [ + 513, + 820, + 622, + 834 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our AccuTile method, outlined in Algorithm 1, extends SnugBox to identify the exact tiles intersected by the Gaussian. It takes as input the tight bounding box produced by Snugbox and its rectangular tile extent - depicted as the", + "bbox": [ + 511, + 839, + 905, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "21540", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/498c1943b015373d3a9f7aa4619224253d212be3fdf54a4eeef91620a0d6558b.jpg", + "image_caption": [ + "(a) 3D Gaussian Splitting" + ], + "image_footnote": [], + "bbox": [ + 112, + 94, + 361, + 273 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/54a1fbdd743500bbb5da29510d8a0eec048bfc1ea78239c3a8b7a1127ab99138.jpg", + "image_caption": [ + "(b) SnugBox", + "Figure 2. Gaussian tile allocation by method. (a) 3D Gaussian Splating allocates a Gaussian to a tile when that tile intersects the square inscribing the circle with radius $\\lceil 3\\sqrt{\\lambda_{\\max}}\\rceil$ defined in Equation 8. (b) Our SnugBox method allocates a Gaussian to a tile when that tile intersects the tight bounding box defined by the axis-aligned minima and maxima of the ellipse given by Equation 14. (c) Our AccuTile method allocates a Gaussian to a tile only if that tile intersects the ellipse via Algorithm 1, which computes the minimum and maximum tiles by iterating over the shorter side of the rectangular tile extent given by SnugBox. In this example, our AccuTile algorithm iterates over the tile rows; the only points that are processed are $x_{min}$ , $x_{max}$ , A, B, C, and D." + ], + "image_footnote": [], + "bbox": [ + 372, + 94, + 624, + 273 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/13cf56fb2296d2a06c9f0264d39d6b0ac0948a8d93ab199d86136919acaae2b4.jpg", + "image_caption": [ + "(c) AccuTile" + ], + "image_footnote": [], + "bbox": [ + 638, + 94, + 885, + 273 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "blue box and yellow tiles, respectively, in Figure 2b. Depending on which dimension is smaller, AccuTile then processes either the rows or columns of this tile extent to determine the exact tiles that intersect the Gaussian. Specifically, we identify the minimum and maximum extent of the ellipse within a given row or column, then convert those points to the corresponding touched tiles. All tiles between the minimum and maximum tiles intersect the ellipse.", + "bbox": [ + 88, + 407, + 483, + 527 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "AccuTile's key insight is that calculating the minimum and maximum extent of the ellipse within each row or column requires computing only two points per iteration. The only possible inflection points of the elliptical curve are the bounding box minimum and maximum points identified by Snugbox. If one or both of these points lie within the tile row or column, then they represent the minimum or maximum extent of the ellipse there. If neither point is within the row or column, then the points along that boundary side are monotonically decreasing or increasing – the minimum or maximum point must lie on one of the boundary lines of that row or column. Since the boundary of the last row or column is shared with the next one, we only need to compute the intersection of the ellipse with the next boundary line in each iteration. Thus, our AccuTile algorithm counts tiles in time proportional to the shorter side of the tile extent and processes tiles in time proportional to the tile count.", + "bbox": [ + 89, + 532, + 483, + 790 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Figure 2 illustrates how Accutile restricts the tight bounding box produced by SnugBox to the exact tiles that the Gaussian touches. The tile extent rows from Snugbox, shown in Figure 2b, are processed starting from the bottom. The lower tile row boundary falls below the bounding box, so no initial intersection is computed. The upper boundary lies below the top of the bounding box, so intersection", + "bbox": [ + 89, + 795, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "points $A$ and $B$ are calculated using Equation 15. $x_{\\mathrm{min}}$ is assigned as this row's minimum ellipse extent $e_{\\mathrm{min}}$ because it lies within it, and $B$ is assigned as its maximum ellipse extent $e_{\\mathrm{max}}$ because it is the maximal point. Consequently, this row's tile extent is from the tile containing $x_{\\mathrm{min}}$ to the tile containing $B$ . For the next row, we keep points $A$ and $B$ and compute upper boundary points $C$ and $D$ . $A$ and $x_{\\mathrm{max}}$ are assigned as the row's $e_{\\mathrm{min}}$ and $e_{\\mathrm{max}}$ , and the tiles containing them are its tile extent. Finally, for the last row, $C$ and $D$ are kept; no additional points are computed because the row's upper boundary is above the bounding box. Thus, the tile extent of this row is between $C$ and $D$ .", + "bbox": [ + 511, + 407, + 906, + 588 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The number of tiles is further reduced from Figure 2b to Figure 2c, much less than the original 3D-GS method in Figure 2a. In Table 1, we report that AccuTile further accelerates all downstream functions, culminating in an average overall speed-up of $1.99 \\times$ .", + "bbox": [ + 511, + 589, + 906, + 666 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Efficient Pruning", + "text_level": 1, + "bbox": [ + 511, + 678, + 684, + 694 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "PUP 3D-GS [11] is a pruning method that quantifies the sensitivity of each Gaussian to training views, removing a set percentage with the lowest sensitivities. Sensitivity is computed by approximating the Hessian of the $L_{2}$ loss:", + "bbox": [ + 511, + 700, + 905, + 762 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nH = \\nabla_ {\\mathcal {G}} ^ {2} L _ {2} = \\sum_ {\\phi \\in \\mathcal {P} _ {g t}} \\nabla_ {\\mathcal {G}} I _ {\\mathcal {G}} (\\phi) \\nabla_ {\\mathcal {G}} I _ {\\mathcal {G}} (\\phi) ^ {T}, \\tag {17}\n$$\n", + "text_format": "latex", + "bbox": [ + 553, + 775, + 906, + 809 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\nabla_{\\mathcal{G}}I_{\\mathcal{G}}(\\phi)$ is the gradient over all Gaussian parameters on the rendered image $I_{\\mathcal{G}}$ for pose $\\phi$ . $H$ is shown to be exact when the $L_{1}$ residual error vanishes [11].", + "bbox": [ + 511, + 823, + 905, + 869 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "A per-Gaussian sensitivity can be derived by splitting $H$ into the block diagonal elements that only capture inter", + "bbox": [ + 511, + 869, + 905, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "21541", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Algorithm 1 The AccuTile Algorithm. For simplicity, the algorithm outlined here is applied to the rows of the Snug-Box tile extent bounding box, matching the example in Figure 2c. In practice, it is applied along the smaller side of the tile extent. The subscripts $t$ , $b$ , $l$ , and $r$ represent the top, bottom, left, and right sides, respectively. A proof of correctness sketch is presented in Appendix A.1.", + "bbox": [ + 89, + 90, + 483, + 196 + ], + "page_idx": 5 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [], + "code_body": "Require: Ellipse E $\\triangleright$ Eq. 14 \nRequire: SnugBox Bounding Box B \nRequire: SnugBox Tile Extent Rectangle R \nInitialize: Tile count $C\\gets 0$ \nline $\\mathrm{min}\\gets \\mathbf{R}_b$ \nif line $\\mathrm{min}\\geq \\mathbf{B}_b$ then \n $\\mathbf{i}_{\\mathrm{min}}\\gets$ Intersections( line $\\mathrm{min}$ ,E) $\\triangleright$ Eq. 15 \nend if \nfor row r in R do \nline $\\mathrm{max}\\gets \\mathrm{r}_t$ \nif line $\\mathrm{max}\\leq \\mathbf{B}_t$ then \n $\\mathbf{i}_{\\mathrm{max}}\\gets$ Intersections( line $\\mathrm{max}$ ,E) $\\triangleright$ Eq. 15 \nend if \ne $\\mathrm{min}\\gets \\mathbf{B}_l$ if Bl in r else min(i $\\mathrm{min}$ ,i max) \ne $\\mathrm{max}\\gets$ Br if Br in r else max(i $\\mathrm{min}$ ,i max) \ntile $\\mathrm{min}$ ,tile $\\mathrm{max}\\gets$ Convert(e $\\mathrm{min}$ ,e max) \nC $\\leftarrow$ C + (tile $\\mathrm{max}$ -tile $\\mathrm{min}$ ) \nProcess( tile $\\mathrm{min}$ ,tile $\\mathrm{max}$ ) \n $\\mathbf{i}_{\\mathrm{min}}\\gets \\mathbf{i}_{\\mathrm{max}}$ \nend for \nreturn C", + "bbox": [ + 89, + 207, + 483, + 580 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Gaussian parameter relationships:", + "bbox": [ + 89, + 614, + 316, + 631 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nH _ {i} = \\sum_ {\\phi \\in \\mathcal {P} _ {g t}} \\nabla_ {\\mathcal {G} _ {i}} I _ {\\mathcal {G}} (\\phi) \\nabla_ {\\mathcal {G} _ {i}} I _ {\\mathcal {G}} (\\phi) ^ {T}, \\tag {18}\n$$\n", + "text_format": "latex", + "bbox": [ + 169, + 638, + 482, + 674 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\nabla_{\\mathcal{G}_i}$ is the gradient with respect to only $\\mathcal{G}_i$ . This measures the sensitivity of the $L_2$ loss with respect to Gaussian $\\mathcal{G}_i$ , assuming all other Gaussians are held constant.", + "bbox": [ + 89, + 681, + 482, + 728 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "$H_{i}$ is again approximated by only using the six mean $\\mu_{i}$ and scale $s_i$ parameters to specifically capture geometric sensitivity. The log determinant is taken to provide a representative scalar score $U_{i}$ :", + "bbox": [ + 89, + 728, + 483, + 789 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nU _ {i} = \\log | \\nabla_ {\\mu_ {i}, s _ {i}} I _ {\\mathcal {G}} \\nabla_ {\\mu_ {i}, s _ {i}} I _ {\\mathcal {G}} ^ {T} |. \\tag {19}\n$$\n", + "text_format": "latex", + "bbox": [ + 184, + 797, + 482, + 816 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Using this score, up to $90\\%$ of Gaussians can be robustly pruned from the model while retaining high visual quality.", + "bbox": [ + 89, + 824, + 482, + 854 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Although PUP 3D-GS touts high compression ratios and rendering speeds, we identify two key drawbacks in its formulation. First, computing the Hessian requires storage", + "bbox": [ + 89, + 854, + 483, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "proportional to $N \\times 36$ , where $N$ is the number of Gaussians. In comparison, the 3D-GS model has a memory footprint proportional to $N \\times 59$ because it stores 59 parameters per Gaussian. While this score is effective for post-hoc pruning, using it during training is impractical.", + "bbox": [ + 511, + 90, + 903, + 167 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Second, computing the Hessian requires the pixel-wise gradients of $\\mu$ and $s$ . Since these are 3D parameters of the Gaussian primitives, obtaining their gradients requires back-propagating through the render kernel parallelized per pixel, then back-propagating each Gaussian contributing to that pixel in its thread. This breaks the efficient flow of gradients in 3D-GS, where the per-pixel gradients from the render kernel are parallelized and aggregated to the 2D $\\mu_{2D}$ and $\\Sigma_{2D}$ parameters, which are then parallelized across Gaussians to compute gradients for $\\mu$ and $s$ .", + "bbox": [ + 511, + 167, + 906, + 316 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our approach builds on PUP 3D-GS by introducing an efficient pruning score that we incorporate into the 3D-GS training pipeline. We also define two distinct pruning modalities: Soft Pruning, which takes place during densification in the first 15000 iterations, and Hard Pruning, which is applied after densification is completed after iteration 15000.", + "bbox": [ + 511, + 316, + 905, + 422 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2.1. Efficient Pruning Score", + "text_level": 1, + "bbox": [ + 511, + 429, + 723, + 444 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our insight is that both drawbacks can be alleviated by reparameterizing the Hessian. Concretely, we express the influence of all spatial parameters of Gaussian $\\mathcal{G}_i$ by computing the Hessian approximation with respect to the 2D projected value of $\\mathcal{G}_i$ at pixel $p$ , given by $g_i(p)$ in Equation 6. In doing so, the pruning score $U_i$ from Equation 19 can be expressed as:", + "bbox": [ + 511, + 446, + 905, + 553 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {U} _ {i} = \\log | \\nabla_ {g _ {i}} I _ {\\mathcal {G}} \\nabla_ {g _ {i}} I _ {\\mathcal {G}} ^ {T} |. \\tag {20}\n$$\n", + "text_format": "latex", + "bbox": [ + 624, + 566, + 903, + 585 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Since $g_{i}$ is a scalar and log is monotonically increasing, we can rewrite this score as:", + "bbox": [ + 511, + 588, + 905, + 617 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {U} _ {i} = \\left(\\nabla_ {g _ {i}} I _ {\\mathcal {G}}\\right) ^ {2}. \\tag {21}\n$$\n", + "text_format": "latex", + "bbox": [ + 653, + 625, + 903, + 643 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Gradient $\\nabla_{g_i}I_{\\mathcal{G}}$ is already computed in the backward pass of render and can be efficiently squared and aggregated across all pixels. Moreover, the maximum space requirement for this score is proportional to the number of Gaussians $N$ , reducing the storage requirement by $36\\times$ and allowing this score to be used during training.", + "bbox": [ + 511, + 648, + 906, + 741 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2.2. Soft Pruning", + "text_level": 1, + "bbox": [ + 511, + 746, + 650, + 761 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To maintain a robust Hessian approximation, we observe that the $L_{1}$ loss becomes quite small by iteration 6000 and remains low except after an opacity reset is performed. As such, we augment the densification pipeline to include our Soft Pruning method, where the model is pruned immediately before the three opacity resets at 6000, 9000, and 12000 iterations. Surprisingly, we find that we can set extremely high Soft Pruning ratios – in our experiments, visual fidelity is preserved at $80\\%$ pruning.", + "bbox": [ + 511, + 763, + 906, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "21542", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ffa852edb8b496f4992351b54752d23519b28289c3ede99ec305b29d3979ba23.jpg", + "image_caption": [ + "GT" + ], + "image_footnote": [], + "bbox": [ + 96, + 108, + 282, + 204 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ab1bba5ffc9241fd393b849b78777938923af29863d8da268f51a99e0ac94bff.jpg", + "image_caption": [ + "3D-GS", + "2.5M Gaussians", + "172.15 FPS" + ], + "image_footnote": [], + "bbox": [ + 305, + 108, + 491, + 203 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7b86d26216953725c49b5f1abe752c107466b4e116a939948888647730f80a15.jpg", + "image_caption": [ + "PUP", + "0.25M Gaussians" + ], + "image_footnote": [], + "bbox": [ + 500, + 108, + 684, + 203 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6d9b80c08e645f24d151969b745cc25e543f78c066e072c94ad7237210beb122.jpg", + "image_caption": [ + "0.22M Gaussians", + "1277.20 FPS" + ], + "image_footnote": [], + "bbox": [ + 697, + 90, + 890, + 203 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/cbe29eccb3828a4900ea58195c9e98fefb2d7b2212762b7e7fe9e21bc15f88d7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 96, + 223, + 282, + 320 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a69a2b1aad2b45f92c9a61c3e42068077ae5f9fbe80c3678fa2a3821d4f83abb.jpg", + "image_caption": [ + "4.5M Gaussians", + "141.47 FPS" + ], + "image_footnote": [], + "bbox": [ + 305, + 223, + 491, + 319 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ad195892fca34ea72da077a4a4f19021307eedbdbd2ef40ff1c190e191dec090.jpg", + "image_caption": [ + "0.45M Gaussians", + "261.67 FPS" + ], + "image_footnote": [], + "bbox": [ + 500, + 223, + 684, + 319 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c71c96ee3ad9661c7f598d20039af9416d6e007f20505ed785a468ca4f3cb411.jpg", + "image_caption": [ + "0.46M Gaussians", + "723.79 FPS" + ], + "image_footnote": [], + "bbox": [ + 699, + 223, + 887, + 319 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/74135626c2d30c63985b5840546ec5c0b9e0e6365fe9b37eba0069f9ec92a466.jpg", + "image_caption": [ + "Figure 3. Visual comparison of 3D-GS, PUP 3D-GS, and our method. Notice that, while reaching similar compression ratios to PUP 3D-GS, our Speedy-Splat method delivers vastly faster rendering speeds. Top: playroom from the Deep Blending dataset. Middle: bicycle from the Mip-NeRF 360 dataset. Bottom: drjohnson from the Deep Blending dataset." + ], + "image_footnote": [], + "bbox": [ + 96, + 339, + 284, + 435 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7112c46269c14c0a67c0488d1b942f3961a5977bd336a434a2abe37d0756173f.jpg", + "image_caption": [ + "3.4M Gaussians", + "125.97 FPS" + ], + "image_footnote": [], + "bbox": [ + 305, + 340, + 491, + 434 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4a705b7976af512de25292283ff91eacf5f38e44ef4ca3a5da8b1f75a549a2c3.jpg", + "image_caption": [ + "0.34M Gaussians", + "413.30 FPS" + ], + "image_footnote": [], + "bbox": [ + 500, + 340, + 684, + 435 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6b4a83cf5c4ed81986b2c0dd4ed0c0bd3617f27a78f1d30868f7659030603af2.jpg", + "image_caption": [ + "0.32M Gaussians", + "1121.97 FPS" + ], + "image_footnote": [], + "bbox": [ + 699, + 340, + 888, + 435 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2.3. Hard Pruning", + "text_level": 1, + "bbox": [ + 89, + 520, + 235, + 536 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We also observe that the model's performance after densification closely matches that of the fully-trained model. The iterations after the densification stage essentially finetune the model and can be used to further \"refine\" it after pruning, similar to PUP 3D-GS [11] and LightGaussian [6]. In practice, our Hard Pruning method prunes the model by a constant ratio every 3000 iterations starting at iteration 15000. We Hard Prune $30\\%$ of Gaussians in each interval, which, when paired with Soft Pruning, empirically reduces the total number of Gaussians across scenes by $10.6\\times$ .", + "bbox": [ + 89, + 539, + 483, + 690 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 89, + 702, + 220, + 719 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.1. Datasets", + "text_level": 1, + "bbox": [ + 89, + 727, + 191, + 742 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our evaluation uses the same set of challenging real-world scenes as 3D-GS [14]. This includes nine Mip-Nerf 360 scenes [3] – four indoor and five outdoor – that each feature a complex central object or area with a detailed background. We also include the outdoor train and truck scenes from the Tanks & Temples dataset [17] and the indoor drjohnson and playroom scenes from the Deep Blending dataset [12]. For consistency across experiments, we use the COLMAP pose estimates and sparse point clouds provided by the dataset authors.", + "bbox": [ + 89, + 750, + 483, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/c1deb9d05335c70f11526d2b0757f1aa3b019cfb682981339b047659350e9848.jpg", + "table_caption": [ + "Table 2. Average Gaussian count, FPS, and training time across all scenes in Section 5.1. Ratios for model size compression, rendering speed-up, and training speed-up are reported in (parentheses). The operation in each row is applied cumulatively to all of the following rows. The best and second best value for each metric are color coded." + ], + "table_footnote": [], + "table_body": "
Method# Gaussians ↓FPS ↑Training Time ↓
Baseline2.93M13423.2
+SnugBox2.97M244 (1.82×)21.2 (1.09×)
+AccuTile2.97M267 (1.99×)21.0 (1.10×)
+Soft Pruning1.64M (1.79×)420 (3.14×)17.5 (1.32×)
+Hard Pruning0.28M (10.6×)898 (6.71×)15.7 (1.47×)
", + "bbox": [ + 517, + 609, + 898, + 700 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Implementation Details", + "text_level": 1, + "bbox": [ + 513, + 733, + 728, + 750 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our code builds on the differentiable renderer provided by 3D-GS [14] and modifies the Python training pipeline for pruning schedules and execution. To ensure consistent and precise timing, all times in Table 1 and FPS values in Tables 2 and 3 are measured using CUDA events at the start and end of the forward rendering procedure. All experiments are conducted on an Nvidia RTXA5000 GPU, and the reported Speedy-Splat results represent the average metrics across three independent runs for each scene.", + "bbox": [ + 511, + 763, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "21543", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/407015e335dd8bb6670faceae6133ef9bbba3da783d3357d291e6c30e867c35b.jpg", + "table_caption": [ + "Table 3. Average reported metrics for each pruning method across all scenes in the Mip-NeRF 360 dataset. The Comp column reports model size compression in terms of Gaussian count, FPS reports rendering speed-up, and Train reports training time speed-up, all with respect to the baseline 3D-GS model. PSNR, SSIM, and LPIPS are also recorded. For a fair comparison, we report the published results of each method and use '-' to denote missing metrics. The best and second best value for each metric are color coded; lossless methods are underlined. Results for the Tanks & Temples and Deep Blending datasets are reported in Appendix A.3." + ], + "table_footnote": [], + "table_body": "
MethodComp ↑FPS ↑Train ↑PSNR ↑SSIM ↑LPIPS ↓
3D-GS [14]1.00×1.00×1.00×27.550.8140.222
Trimming [2]4.00×--27.130.7980.248
Compact [18]2.28×1.07×0.73×27.080.7980.247
EAGLES [10]3.68×1.51×1.37×26.940.8000.250
Reducing [26]2.33×1.60×1.23×27.100.8090.226
Light [6]2.94×1.76×-27.280.8050.243
ELMGS [1]5.00×2.69×-27.000.7790.286
PUP [11]8.65×2.55×-26.830.7920.268
Mini-Splat [7]6.84×3.20×1.26×27.340.8220.217
+SnugBox0.99×1.81×1.08×27.550.8140.221
+AccuTile0.99×1.99×1.10×27.570.8140.221
+Soft Pruning1.79×3.14×1.30×27.320.8070.246
+Hard Pruning10.6×6.51×1.45×26.940.7820.296
", + "bbox": [ + 93, + 238, + 478, + 416 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.3. Results", + "text_level": 1, + "bbox": [ + 89, + 435, + 184, + 450 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.3.1. Additive method performance", + "text_level": 1, + "bbox": [ + 89, + 458, + 346, + 473 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The efficacy of Speedy-Splat is demonstrated by Table 1, where we record the average execution times of each function across all scenes when additively applying our methods. SnugBox and AccuTile each introduce minimal additional computation time to preprocess and InclusiveSum. However, limiting the number of tiles touched accelerates all downstream functions, culminating in an overall speed-up of $1.82 \\times$ by SnugBox that is raised to $1.99 \\times$ by AccuTile. Applying soft pruning reduces the runtime of all functions by reducing the number of Gaussians, leading to a $3.14 \\times$ overall speed-up. Finally, performing hard pruning improves overall speed by a whopping $6.71 \\times$ over the baseline 3D-GS model.", + "bbox": [ + 88, + 478, + 482, + 674 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.3.2. Overall Performance", + "text_level": 1, + "bbox": [ + 89, + 684, + 284, + 698 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In Figures 1 and 3, we report qualitative results on two outdoor and two indoor scenes from all three datasets. The magnified regions highlight that our method preserves the fine details in the baseline 3D-GS scene and closely models the ground truth view. Despite touting similar compression ratios and rendering nearly identical images, Speedy-Splat achieves over double the FPS of PUP 3D-GS.", + "bbox": [ + 89, + 703, + 482, + 808 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3 compares our methods against other methods that reduce the number of Gaussians and increase inference speed using the mean of each metric across all scenes in the Mip-NeRF 360 dataset. The underlined methods are \"lossless\", meaning that they avoid degrading visual fidelity at all. SnugBox and AccuTile, our lossless methods, improve", + "bbox": [ + 89, + 810, + 483, + 901 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/a1e75758a8b812f8058f85f2d71856dd6b90270b2d7f2c28e6d2383cae6c0953.jpg", + "table_caption": [ + "Table 4. Average metrics across all scenes in Section 5.1 when using the Speedy-Splat and PUP 3D-GS pruning scores to prune $88.44\\%$ of Gaussians using the PUP 3D-GS pipeline. Two rounds of prune-refine are performed on each baseline 3D-GS model, pruning $66\\%$ of Gaussians and then fine-tuning for 5,000 iterations in each one. The best value for each metric is color coded." + ], + "table_footnote": [], + "table_body": "
Method# Gaussians ↓PSNR ↑SSIM ↑LPIPS ↓FPS ↑
Baseline2.92M27.15030.82960.2238107.53
PUP [11]0.34M26.21360.80440.2731378.57
Ours0.34M26.86580.80220.2840345.52
", + "bbox": [ + 517, + 183, + 903, + 247 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "rendering and training speed while leaving image quality metrics essentially unchanged or slightly better. Our full pipeline, labeled as “+Hard Pruning” boasts the highest compression ratios, rendering speeds, and training speed-ups across all datasets. Furthermore, its image quality metrics are competitive with the other methods.", + "bbox": [ + 511, + 268, + 906, + 359 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.4. Pruning Score Comparison", + "text_level": 1, + "bbox": [ + 511, + 377, + 759, + 393 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Although the primary focus of our work is rendering speed, we find that our efficient pruning score, described in Section 4.2, also performs well when applied in other compression pipelines. In Table 4, we ablate our efficient pruning score with the PUP 3D-GS sensitivity score across all scenes in their post-hoc pruning pipeline. Notably, Speedy-Splat's efficient pruning score outperforms PUP 3D-GS on PSNR and is competitive across the other metrics.", + "bbox": [ + 511, + 401, + 906, + 521 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Limitations", + "text_level": 1, + "bbox": [ + 511, + 544, + 635, + 559 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "A limitation of Speedy-Splat is that it produces slightly lower image quality than 3D-GS. However, this degradation is expected at high compression ratios and is also observed in comparable techniques. Additionally, a direct comparison of our efficient pruning score to the PUP 3D-GS pruning score illuminates a slight, yet noticeable, gap in performance. Future work could explore the possibility of another efficient pruning score that delivers higher performance.", + "bbox": [ + 511, + 566, + 906, + 689 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 710, + 633, + 724 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we present Speedy-Splat: a new 3D-GS technique that accurately localizes Gaussians during rendering and significantly improves inference speed, model size, and training time. Enhanced localization is achieved by our SnugBox and AccuTile methods, while model size reduction is accomplished by our Soft and Hard Pruning approaches. Together, our Speedy-Splat methods accelerate rendering speed by an average of $6.71 \\times$ , reduce model size by $10.6 \\times$ , and improve training time by $1.47 \\times$ across all scenes from the Mip-NeRF 360, Tanks & Temples, and Deep Blending datasets.", + "bbox": [ + 511, + 734, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "21544", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "8. Acknowledgements", + "text_level": 1, + "bbox": [ + 89, + 90, + 277, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was made possible by the IARPA WRIVA Program, the ONR MURI program, and DAPRA TIAMAT. Commercial support was provided by Capital One Bank, the Amazon Research Award program, and Open Philanthropy. Further support was provided by the National Science Foundation (IIS-2212182), and by the NSF TRAILS Institute (2229885). Zwicker was additionally supported by the National Science Foundation (IIS-2126407).", + "bbox": [ + 89, + 108, + 483, + 229 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 242, + 187, + 258 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Muhammad Salman Ali, Sung-Ho Bae, and Enzo Tartaglione. Elmgs: Enhancing memory and computation scalability through compression for 3d gaussian splatting. arXiv preprint arXiv:2410.23213, 2024. 2, 8, 12", + "[2] Muhammad Salman Ali, Maryam Qamar, Sung-Ho Bae, and Enzo Tartaglione. Trimming the fat: Efficient compression of 3d gaussian splats through pruning. arXiv preprint arXiv:2406.18214, 2024. 2, 8, 12", + "[3] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5470–5479, 2022. 7", + "[4] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In European Conference on Computer Vision (ECCV), 2022. 1", + "[5] Sankeerth Durvasula, Adrian Zhao, Fan Chen, Ruofan Liang, Pawan Kumar Sanjaya, and Nandita Vijaykumar. Distwar: Fast differentiable rendering on raster-based rendering pipelines. arXiv preprint arXiv:2401.05345, 2023. 2", + "[6] Zhiwen Fan, Kevin Wang, Kairun Wen, Zehao Zhu, Dejia Xu, and Zhangyang Wang. Lightgaussian: Unbounded 3d gaussian compression with 15x reduction and $200+$ fps. arXiv preprint arXiv:2311.17245, 2023. 1, 2, 7, 8, 12", + "[7] Guangchi Fang and Bing Wang. Mini-splatting: Representing scenes with a constrained number of gaussians. arXiv preprint arXiv:2403.14166, 2024. 1, 2, 8, 12", + "[8] Guofeng Feng, Siyan Chen, Rong Fu, Zimu Liao, Yi Wang, Tao Liu, Zhilin Pei, Hengjie Li, Xingcheng Zhang, and Bo Dai. Flashgs: Efficient 3d gaussian splatting for large-scale and high-resolution rendering. arXiv preprint arXiv:2408.07967, 2024.1, 2", + "[9] Henri P. Gavin. The levenberg-marquardt method for nonlinear least squares curve-fitting problems c ©. 2013. 2", + "[10] Sharath Girish, Kamal Gupta, and Abhinav Shrivastava. Eagles: Efficient accelerated 3d gaussians with lightweight encodings. arXiv preprint arXiv:2312.04564, 2023. 2, 8, 12", + "[11] Alex Hanson, Allen Tu, Vasu Singla, Mayuka Jayawardhana, Matthias Zwicker, and Tom Goldstein. Pup 3d-gs: Principled uncertainty pruning for 3d gaussian splattering. arXiv preprint arXiv:2406.10219, 2024. 1, 2, 4, 5, 7, 8, 12", + "[12] Peter Hedman, Julien Philip, True Price, Jan-Michael Frahm, George Drettakis, and Gabriel Brostow. Deep blending for free-viewpoint image-based rendering. ACM Transactions on Graphics, 37(6):1-15, 2018. 7" + ], + "bbox": [ + 93, + 267, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Lukas Hollein, Aljaž Božić, Michael Zollhöfer, and Matthias Nießner. 3dgs-lm: Faster gaussian-splatting optimization with levenberg-marquardt. arXiv preprint arXiv:2409.12892, 2024. 2", + "[14] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42 (4):1-14, 2023. 1, 2, 3, 7, 8, 12, 13", + "[15] Shakiba Kheradmand, Daniel Rebain, Gopal Sharma, Weiwei Sun, Yang-Che Tseng, Hossam Isack, Abhishek Kar, Andrea Tagliasacchi, and Kwang Moo Yi. 3d gaussian splatting as markov chain monte carlo. In Advances in Neural Information Processing Systems (NeurIPS), 2024. Spotlight Presentation. 2", + "[16] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. CoRR, abs/1412.6980, 2014. 2", + "[17] Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and temples: Benchmarking large-scale scene reconstruction. ACM Transactions on Graphics, 36(4):1-13, 2017. 7", + "[18] Joo Chan Lee, Daniel Rho, Xiangyu Sun, Jong Hwan Ko, and Eunbyung Park. Compact 3d gaussian representation for radiance field. arXiv preprint arXiv:2311.13681, 2023. 2, 8, 12", + "[19] Weikai Lin, Yu Feng, and Yuhao Zhu. Rtgs: Enabling real-time gaussian splatting on mobile devices using efficiency-guided pruning and foveated rendering. arXiv preprint arXiv:2407.00435, 2024. 1, 2", + "[20] Xiangrui Liu, Xinju Wu, Pingping Zhang, Shiqi Wang, Zhu Li, and Sam Kwong. Compgs: Efficient 3d scene representation via compressed gaussian splatting. arXiv preprint arXiv:2404.09458, 2024. 2", + "[21] Tao Lu, Mulin Yu, Linning Xu, Yuanbo Xiangli, Limin Wang, Dahua Lin, and Bo Dai. Scaffold-gs: Structured 3d gaussians for view-adaptive rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20654-20664, 2024. 2", + "[22] Saswat Subhajyoti Mallick, Rahul Goel, Bernhard Kerbl, Markus Steinberger, Francisco Vicente Carrasco, and Fernando De La Torre. Taming 3dgs: High-quality radiance fields with limited resources. In SIGGRAPH Asia 2024 Conference Papers, New York, NY, USA, 2024. Association for Computing Machinery. 2", + "[23] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 1", + "[24] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph., 41(4):102:1-102:15, 2022. 1", + "[25] Michael Niemeyer, Fabian Manhardt, Marie-Julie Rakoto-saona, Michael Oechsle, Daniel Duckworth, Rama Gosula, Keisuke Tateno, John Bates, Dominik Kaeser, and Federico Tombari. Radsplat: Radiance field-informed gaussian splatting for robust real-time rendering with $900+$ fps. arXiv preprint arXiv:2403.13806, 2024. 1, 2" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "21545", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[26] Panagiotis Papantonakis, Georgios Kopanas, Bernhard Kerbl, Alexandre Lanvin, and George Drettakis. Reducing the memory footprint of 3d gaussian splatting. Proceedings of the ACM on Computer Graphics and Interactive Techniques, 7(1):1-17, 2024. 2, 8, 12", + "[27] Lukas Radl, Michael Steiner, Mathias Parger, Alexander Weinrauch, Bernhard Kerbl, and Markus Steinberger. StopThePop: Sorted Gaussian Splatting for View-Consistent Real-time Rendering. ACM Transactions on Graphics, 4 (43), 2024. 2, 13", + "[28] Kerui Ren, Lihan Jiang, Tao Lu, Mulin Yu, Linning Xu, Zhangkai Ni, and Bo Dai. Octree-gs: Towards consistent real-time rendering with lod-structured 3d gaussians. arXiv preprint arXiv:2403.17898, 2024. 2", + "[29] Fernando Rivas-Manzaneque, Jorge Sierra-Acosta, Adrian Penate-Sanchez, Francesc Moreno-Noguer, and Angela Ribeiro. Nerflight: Fast and light neural radiance fields using a shared feature grid. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12417-12427, 2023. 1", + "[30] Samuel Rota Bulò, Lorenzo Porzi, and Peter Kontschieder. Revising densification in gaussian splatting. In European Conference on Computer Vision, pages 347-362. Springer, 2024. 2", + "[31] Evangelos Ververas, Rolandos Alexandros Potamias, Jifei Song, Jiankang Deng, and Stefanos Zafeiriou. Sags: structure-aware 3d gaussian splatting. In European Conference on Computer Vision, pages 221-238. Springer, 2024. 2", + "[32] Meng Wei, Qianyi Wu, Jianmin Zheng, Hamid Rezatofighi, and Jianfei Cai. Normal-gs: 3d gaussian splatting with normal-involved rendering. In Advances in Neural Information Processing Systems (NeurIPS), 2024. 2", + "[33] Matthias Zwicker, Hanspeter Pfister, Jeroen Van Baar, and Markus Gross. Ewa splatting. IEEE Transactions on Visualization and Computer Graphics, 8(3):223-238, 2002. 3" + ], + "bbox": [ + 91, + 92, + 480, + 598 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "21546", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Speedy-Splat_ Fast 3D Gaussian Splatting with Sparse Pixels and Sparse Primitives/08166209-021b-4f8f-8611-9f7790e8bc2b_model.json b/2025/Speedy-Splat_ Fast 3D Gaussian Splatting with Sparse Pixels and Sparse Primitives/08166209-021b-4f8f-8611-9f7790e8bc2b_model.json new file mode 100644 index 0000000000000000000000000000000000000000..aab883a6b0e5677bfd175171ba22becf624b1fc0 --- /dev/null +++ b/2025/Speedy-Splat_ Fast 3D Gaussian Splatting with Sparse Pixels and Sparse Primitives/08166209-021b-4f8f-8611-9f7790e8bc2b_model.json @@ -0,0 +1,2838 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.262, + 0.131, + 0.737, + 0.177 + ], + "angle": 0, + "content": "Speedy-Splat: Fast 3D Gaussian Splating with Sparse Pixels and Sparse Primitives" + }, + { + "type": "text", + "bbox": [ + 0.254, + 0.204, + 0.361, + 0.22 + ], + "angle": 0, + "content": "Alex Hanson" + }, + { + "type": "text", + "bbox": [ + 0.326, + 0.223, + 0.469, + 0.238 + ], + "angle": 0, + "content": "Matthias Zwicker" + }, + { + "type": "text", + "bbox": [ + 0.406, + 0.206, + 0.479, + 0.221 + ], + "angle": 0, + "content": "Allen Tu" + }, + { + "type": "text", + "bbox": [ + 0.524, + 0.205, + 0.602, + 0.221 + ], + "angle": 0, + "content": "Geng Lin" + }, + { + "type": "text", + "bbox": [ + 0.648, + 0.205, + 0.744, + 0.222 + ], + "angle": 0, + "content": "Vasu Singla" + }, + { + "type": "text", + "bbox": [ + 0.554, + 0.223, + 0.672, + 0.238 + ], + "angle": 0, + "content": "Tom Goldstein" + }, + { + "type": "text", + "bbox": [ + 0.35, + 0.25, + 0.648, + 0.268 + ], + "angle": 0, + "content": "University of Maryland, College Park" + }, + { + "type": "image_caption", + "bbox": [ + 0.212, + 0.284, + 0.24, + 0.297 + ], + "angle": 0, + "content": "GT" + }, + { + "type": "image", + "bbox": [ + 0.096, + 0.301, + 0.357, + 0.414 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.468, + 0.283, + 0.519, + 0.298 + ], + "angle": 0, + "content": "3DGS" + }, + { + "type": "image", + "bbox": [ + 0.365, + 0.301, + 0.625, + 0.413 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.369, + 0.415, + 0.466, + 0.427 + ], + "angle": 0, + "content": "2.6M Gaussians" + }, + { + "type": "image_caption", + "bbox": [ + 0.471, + 0.415, + 0.543, + 0.427 + ], + "angle": 0, + "content": "25.39 PSNR" + }, + { + "type": "image_caption", + "bbox": [ + 0.551, + 0.415, + 0.617, + 0.427 + ], + "angle": 0, + "content": "184.87 FPS" + }, + { + "type": "image_caption", + "bbox": [ + 0.746, + 0.284, + 0.789, + 0.298 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.631, + 0.3, + 0.895, + 0.413 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.635, + 0.415, + 0.736, + 0.427 + ], + "angle": 0, + "content": "0.26M Gaussians" + }, + { + "type": "image_caption", + "bbox": [ + 0.742, + 0.415, + 0.814, + 0.427 + ], + "angle": 0, + "content": "25.34 PSNR" + }, + { + "type": "image_caption", + "bbox": [ + 0.821, + 0.415, + 0.894, + 0.427 + ], + "angle": 0, + "content": "1148.77 FPS" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.437, + 0.907, + 0.465 + ], + "angle": 0, + "content": "Figure 1. We reduce the number of Gaussians by over \\(90\\%\\), only marginally decrease PSNR, and accelerate rendering speed by \\(6.2\\times\\) in the Tanks & Temples truck scene when compared to 3D Gaussian Splatting (3D-GS). Additionally, we speed up training time by \\(1.38\\times\\)." + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.478, + 0.327, + 0.493 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.509, + 0.486, + 0.812 + ], + "angle": 0, + "content": "3D Gaussian Splatting (3D-GS) is a recent 3D scene reconstruction technique that enables real-time rendering of novel views by modeling scenes as parametric point clouds of differentiable 3D Gaussians. However, its rendering speed and model size still present bottlenecks, especially in resource-constrained settings. In this paper, we identify and address two key inefficiencies in 3D-GS to substantially improve rendering speed. These improvements also yield the ancillary benefits of reduced model size and training time. First, we optimize the rendering pipeline to precisely localize Gaussians in the scene, boosting rendering speed without altering visual fidelity. Second, we introduce a novel pruning technique and integrate it into the training pipeline, significantly reducing model size and training time while further raising rendering speed. Our Speedy-Splat approach combines these techniques to accelerate average rendering speed by a drastic \\(6.71 \\times\\) across scenes from the Mip-NeRF 360, Tanks & Temples, and Deep Blending datasets. Code and additional visualizations are available at https://speedysplat.github.io/." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.831, + 0.222, + 0.845 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.485, + 0.903 + ], + "angle": 0, + "content": "Fast rendering of photorealistic novel views has been a long-standing goal in computer vision and graphics. Neural Radiance Fields (NeRF) [23] and its variants have made sig-" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.479, + 0.907, + 0.72 + ], + "angle": 0, + "content": "nificant strides in photorealistic 3D scene reconstruction by representing scenes as continuous neural volumetric models that encode scene density and color at spatial coordinates. However, despite recent efforts [4, 24, 29], fast rendering in NeRF remains challenging because the volumetric sampling used in ray-marching is computationally expensive. Recently, 3D Gaussian Splatting (3D-GS) [14] has emerged as a promising alternative that enables real-time rendering by modeling scenes as parametric point clouds of differentiable 3D Gaussians. Nevertheless, its rendering speed is still limited by high parameter counts and certain algorithmic inefficiencies. Efficient rendering is essential in applications such as virtual reality, networked systems, and multi-view streaming. Real-time rendering on resource-constrained edge devices, such as mobile phones, has yet to be achieved [19]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.728, + 0.909, + 0.85 + ], + "angle": 0, + "content": "Although recent works on compressing 3D-GS models [6, 7, 11, 25] achieve some speed-ups by reducing the number of parameters, few approaches directly target rendering speed [8, 19]. In this paper, we specifically address this gap by demonstrating that the rendering speed of 3D-GS models can be drastically increased while maintaining competitive image quality. Additionally, we show that our methods reduce training time and substantially decrease model size." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.909, + 0.903 + ], + "angle": 0, + "content": "We begin by observing that the cost of 3D-GS rendering is proportional to both the number of Gaussians in the scene and the number of pixels processed per Gaussian. Our ap-" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21537" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.168 + ], + "angle": 0, + "content": "proach optimizes both factors. First, we reduce the number of pixels that are processed for each Gaussian by efficiently and accurately localizing it in the rendered image. Second, we reduce the total number of Gaussians in the model through a novel approach that maintains rendering quality." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.168, + 0.482, + 0.333 + ], + "angle": 0, + "content": "3D-GS implements tile rendering to localize Gaussians in the image plane by assigning each Gaussian to the tiles that it intersects. We find the existing algorithm to be overly conservative; in Section 4.1.1, we address this by introducing our SnugBox algorithm to precisely localize Gaussians by computing a tight bounding box around their extent. Then, in Section 4.1.2, we extend SnugBox with our AccuTile algorithm to identify exact tile intersections. Both approaches are plug-and-play, lead to respective inference speed-ups of \\(1.82 \\times\\) and \\(1.99 \\times\\) on average, and do not change the 3D-GS renderings." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.334, + 0.483, + 0.499 + ], + "angle": 0, + "content": "To reduce the total number of Gaussians while preserving visual fidelity, we extend an existing pruning method [11] by reducing its memory requirement by \\(36\\times\\) and incorporating it into the 3D-GS training pipeline. During the densification stage of 3D-GS training, Gaussians are regularly replicated and pruned. In Section 4.2.2, we augment the densification stage with our Soft Pruning method to prune \\(80\\%\\) of Gaussians at regular intervals. After the densification stage, our Hard Pruning method, described in Section 4.2.3, prunes an additional \\(30\\%\\) of Gaussians at set intervals." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.501, + 0.483, + 0.577 + ], + "angle": 0, + "content": "Our Speedy-Splat approach integrates both techniques into the 3D-GS training pipeline. On average, rendering speed is accelerated by \\(6.71 \\times\\), model size is reduced by \\(10.6 \\times\\), and training speed is improved by \\(1.4 \\times\\) across all evaluated scenes while maintaining high image quality." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.578, + 0.462, + 0.592 + ], + "angle": 0, + "content": "In summary, we propose the following contributions:" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.596, + 0.483, + 0.625 + ], + "angle": 0, + "content": "1. SnugBox: A precise algorithm for computing Gaussian-tile bounding box intersections." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.626, + 0.483, + 0.654 + ], + "angle": 0, + "content": "2. AccuTile: An extension of SnugBox for computing exact Gaussian-tile intersections." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.656, + 0.483, + 0.685 + ], + "angle": 0, + "content": "3. Soft Pruning: An augmentation for pruning Gaussians during densification." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.686, + 0.483, + 0.716 + ], + "angle": 0, + "content": "4. Hard Pruning: An augmentation for pruning Gaussians post-densification." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.596, + 0.483, + 0.716 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.736, + 0.228, + 0.751 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.762, + 0.483, + 0.839 + ], + "angle": 0, + "content": "The real-time rendering speed of 3D-GS [14] on desktop GPUs has inspired research focused on further accelerating both its training and inference in resource-constrained environments. In this section, we review related works that specifically target these performance improvements." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.849, + 0.192, + 0.865 + ], + "angle": 0, + "content": "2.1. Pruning" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.901 + ], + "angle": 0, + "content": "A large portion of Gaussians in vanilla 3D-GS models are redundant [6, 7, 25], motivating a recent line of" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.288 + ], + "angle": 0, + "content": "work focused on pruning Gaussians from 3D-GS models to boost rendering speed with minimal loss of visual fidelity [1, 2, 6, 7, 10, 11, 18, 19, 25]. Nearly all approaches assign a significance score to each Gaussian that is used to rank and prune them. Several works compute the aggregated ray contribution for each Gaussian across all input images [6, 7, 18, 25], while others combine opacity with additional information, such as gradients per Gaussian, to calculate their pruning criterion [1, 2, 20]. Papantonakis et al. [26] use resolution and scale-aware redundancy metrics, EAGLES [10] calculates the total transmittance per Gaussian, and Lin et al. [19] accelerate inference speed through an efficiency-aware strategy and foveated rendering." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.289, + 0.905, + 0.41 + ], + "angle": 0, + "content": "PUP 3D-GS [11] introduces a more principled approach by deriving a Hessian for each Gaussian that represents its sensitivity to the reconstruction error. While PUP 3D-GS achieves state-of-the-art post-hoc pruning results, computing its sensitivity score incurs considerable storage requirements that limit its viability for use during training. Our pruning approach directly extends PUP 3D-GS by improving its memory efficiency by \\(36\\times\\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.421, + 0.668, + 0.435 + ], + "angle": 0, + "content": "2.2. Other Methods" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.444, + 0.907, + 0.747 + ], + "angle": 0, + "content": "In addition to pruning, several other strategies have been explored to enhance the rendering and training speed of 3D-GS. Mini-Splatting [7] modifies its densification strategies and adds a simplification stage to constrain the number of Gaussians. 3DGS-MCMC [15] models training dynamics as an MCMC process. Revisiting Densification [30] introduces an error-based densification strategy. 3DGS-LM [13] replaces Adam [16] with a tailored Levenberg-Marquardt optimizer [9]. Taming 3DGS [22] proposes a constructive optimization process that limits the number of Gaussians to a pre-defined threshold set by the user. DISTWAR [5] dives into the low-level implementation of GPU thread scheduling and optimizes atomic processing with a novel primitive. StopThePop [27] introduces a precise tile intersection method, but we find that our approach is notably faster in Appendix A.4. FlashGS [8] implements a tile intersection method similar to StopThePop. Several works reduce training time and memory requirements by enforcing geometric constraints [21, 28, 31, 32]. Most of these approaches are orthogonal to ours and can be applied alongside it." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.762, + 0.642, + 0.779 + ], + "angle": 0, + "content": "3. Background" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.788, + 0.8, + 0.804 + ], + "angle": 0, + "content": "3.1. 3D Gaussian Splitting Overview" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.905, + 0.901 + ], + "angle": 0, + "content": "3D Gaussian Splatting (3D-GS) [14] models scenes as parametric, point-based representations that use differentiable 3D Gaussians as primitives. Given a set of ground truth training images \\(\\mathcal{I}_{gt} = \\{\\pmb{I}_i\\in \\mathbb{R}^{H\\times W}\\}_{i = 1}^K\\), the scene is initialized by using Structure from Motion (SfM) to produce a sparse point cloud that serves as the initial means for the" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21538" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.908, + 0.132 + ], + "angle": 0, + "content": "Table 1. Average execution time (milliseconds) of each function across all scenes in Section 5.1. The operation in each row is applied cumulatively to all of the following rows. For each model, accurate measurements are collected by averaging execution times across three runs that each render the test set 20 times to reduce variance. The fastest and second fastest times are color coded." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.143, + 0.901, + 0.23 + ], + "angle": 0, + "content": "
MethodPreprocessInclusive SumDuplicate with KeysRadix SortIdentify Tile RangesRenderOverall
Baseline0.6650.0450.5701.5510.0824.4837.478
+SnugBox0.6560.0460.208 (2.738×)0.729 (2.126×)0.041 (1.980×)2.344 (1.913×)4.102 (1.823×)
+AccuTile0.6680.0460.221 (2.575×)0.612 (2.533×)0.035 (2.326×)2.062 (2.175×)3.748 (1.995×)
+Soft Pruning0.370 (1.798×)0.030 (1.494×)0.146 (3.906×)0.404 (3.843×)0.024 (3.422×)1.337 (3.354×)2.381 (3.141×)
+Hard Pruning0.091 (7.293×)0.016 (2.769×)0.090 (6.325×)0.215 (7.217×)0.013 (6.537×)0.619 (7.247×)1.114 (6.712×)
" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.248, + 0.483, + 0.292 + ], + "angle": 0, + "content": "3D Gaussians. The estimated camera poses \\(\\mathcal{P}_{gt} = \\{\\phi_i\\in\\) \\(\\mathbb{R}^{3\\times 4}\\}_{i = 1}^{K}\\) are paired with their corresponding images and are used to optimize the scene." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.294, + 0.484, + 0.37 + ], + "angle": 0, + "content": "Each 3D Gaussian primitive \\(\\mathcal{G}_i\\) is parameterized by three geometry parameters - mean \\(\\mu_i\\in \\mathbb{R}^3\\), scale \\(s_i\\in \\mathbb{R}^3\\), and rotation \\(r_i\\in \\mathbb{R}^4\\) - and two color parameters - view-dependent spherical harmonics \\(h_i\\in \\mathbb{R}^{16\\times 3}\\) and opacity \\(\\sigma_{i}\\in \\mathbb{R}\\). The set of all parameters can be described as:" + }, + { + "type": "equation", + "bbox": [ + 0.169, + 0.381, + 0.483, + 0.4 + ], + "angle": 0, + "content": "\\[\n\\mathcal {G} = \\left\\{\\mathcal {G} _ {i} = \\left\\{\\mu_ {i}, s _ {i}, r _ {i}, h _ {i}, \\sigma_ {i} \\right\\} \\right\\} _ {i = 1} ^ {N}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.411, + 0.425, + 0.426 + ], + "angle": 0, + "content": "where \\(N\\) is the number of Gaussians in the model." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.427, + 0.484, + 0.487 + ], + "angle": 0, + "content": "Given camera pose \\(\\phi\\), the scene is rendered by projecting all Gaussians to image space and applying alpha blending to each pixel. Models are optimized via stochastic gradient descent on image reconstruction losses:" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.501, + 0.483, + 0.518 + ], + "angle": 0, + "content": "\\[\nL (\\mathcal {G} | \\phi , I _ {g t}) = \\| I _ {\\mathcal {G}} (\\phi) - I _ {g t} \\| _ {1} + L _ {\\mathrm {D - S S I M}} (I _ {\\mathcal {G}} (\\phi), I _ {g t}), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.53, + 0.395, + 0.546 + ], + "angle": 0, + "content": "where \\(I_{\\mathcal{G}(\\phi)}\\) is the rendered image for pose \\(\\phi\\)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.546, + 0.484, + 0.606 + ], + "angle": 0, + "content": "During optimization, the scene is periodically densified by cloning and splitting uncertain Gaussians and pruned by removing large and transparent Gaussians. The opacities of the Gaussians are also periodically reset." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.619, + 0.455, + 0.636 + ], + "angle": 0, + "content": "3.2. 3D Gaussian Splitting Rendering Specifics" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.642, + 0.483, + 0.732 + ], + "angle": 0, + "content": "3D-GS uses a tile-based rendering strategy that divides the rendered image into \\(16 \\times 16\\) pixel tiles. Each Gaussian is projected into image space, where its intersection with these tiles is computed. Then, these Gaussian-to-tile mappings are sorted to collect and order the Gaussians by depth for pixel-wise rendering." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.733, + 0.484, + 0.809 + ], + "angle": 0, + "content": "Rendering runtime is dominated by six key functions. Table 1 empirically analyzes the execution time of each function and highlights the improvements achieved by our methods. Descriptions of each function are provided in the following sections." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.821, + 0.236, + 0.837 + ], + "angle": 0, + "content": "3.2.1. Preprocessing" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.9 + ], + "angle": 0, + "content": "The preprocess kernel is parallelized such that each thread processes a single Gaussian \\(\\mathcal{G}_i\\). It computes a 2D projection of \\(\\mathcal{G}_i\\) to image space and obtains a count of tiles that it intersects." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.248, + 0.907, + 0.337 + ], + "angle": 0, + "content": "The mean \\(\\mu_{i}\\) is projected to image space using a viewing transform \\(W\\) and a perspective projection, yielding 2D mean \\(\\mu_{i_{2D}}\\) and depth that are stored for later processing. The scale \\(s_i\\) and rotation \\(r_i\\) parameters are converted to the diagonal scale \\(S_{i}\\) and rotation \\(R_{i}\\) matrices. The 3D covariance is then defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.634, + 0.354, + 0.905, + 0.373 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\Sigma} _ {i _ {3 D}} = \\boldsymbol {R} _ {i} \\boldsymbol {S} _ {i} \\boldsymbol {S} _ {i} ^ {T} \\boldsymbol {R} _ {i} ^ {T}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.534, + 0.384, + 0.687, + 0.399 + ], + "angle": 0, + "content": "which is projected via:" + }, + { + "type": "equation", + "bbox": [ + 0.614, + 0.415, + 0.905, + 0.433 + ], + "angle": 0, + "content": "\\[\n\\hat {\\boldsymbol {\\Sigma}} _ {i _ {3 D}} = \\boldsymbol {J} \\boldsymbol {W} \\boldsymbol {\\Sigma} _ {i _ {3 D}} \\boldsymbol {W} ^ {T} \\boldsymbol {J} ^ {T}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.444, + 0.906, + 0.55 + ], + "angle": 0, + "content": "where \\(J\\) is the Jacobian of the first order approximation of the perspective projection. Dropping the last row and column of \\(\\hat{\\Sigma}_{i_{3D}}\\) gives \\(\\Sigma_{i_{2D}}\\) [14, 33]. The largest eigenvalue of \\(\\Sigma_{i_{2D}}\\) is used to compute the count of tiles intersected by this Gaussian \\(\\mathcal{G}_i\\) as shown in Figure 2a. \\(\\Sigma_{i_{2D}}^{-1}\\) is computed, along with a view-dependent color \\(c_{i}\\), derived from \\(W\\) and \\(h_i\\). All three values are stored for later processing." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.56, + 0.613, + 0.575 + ], + "angle": 0, + "content": "3.2.2. Sorting" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.58, + 0.906, + 0.61 + ], + "angle": 0, + "content": "After preprocessing, four functions process the Gaussians for pixel-wise rendering:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.611, + 0.906, + 0.656 + ], + "angle": 0, + "content": "- InclusiveSum: A CUDA primitive that computes the prefix sum over all Gaussian-tile counts to allocate key and value arrays for Gaussian-to-tile mapping." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.656, + 0.906, + 0.717 + ], + "angle": 0, + "content": "- duplicateWithKeys: A Gaussian-parallel kernel that recomputes Gaussian-tile intersections to generate a key for each intersecting tile, consisting of the tile index and Gaussian depth." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.717, + 0.905, + 0.747 + ], + "angle": 0, + "content": "- RadixSort: A CUDA primitive that sorts the key array, ordering Gaussian indices by tile and then depth." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.747, + 0.905, + 0.777 + ], + "angle": 0, + "content": "- identifyTileRanges: A kernel parallelized across the key array to post-process keys before pixel rendering." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.611, + 0.906, + 0.777 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.787, + 0.635, + 0.802 + ], + "angle": 0, + "content": "3.2.3. Rendering" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.807, + 0.906, + 0.868 + ], + "angle": 0, + "content": "The render kernel is parallelized across pixels. For each pixel \\( p \\), all Gaussians within its corresponding tile are loaded and processed in depth order as determined by RadixSort. An alpha value:" + }, + { + "type": "equation", + "bbox": [ + 0.653, + 0.886, + 0.905, + 0.902 + ], + "angle": 0, + "content": "\\[\n\\alpha_ {i} (p) = \\sigma_ {i} g _ {i} (p) \\tag {5}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "21539" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.123 + ], + "angle": 0, + "content": "is computed for each Gaussian \\(\\mathcal{G}_i\\), where \\(g_i\\) is the value of the projected 2D Gaussian at pixel \\(p\\):" + }, + { + "type": "equation", + "bbox": [ + 0.112, + 0.13, + 0.483, + 0.16 + ], + "angle": 0, + "content": "\\[\ng _ {i} = e ^ {q}, q = - \\frac {1}{2} (p - \\mu_ {i _ {2 D}}) \\boldsymbol {\\Sigma} _ {\\boldsymbol {i} _ {2 D}} ^ {- 1} (p - \\mu_ {i _ {2 D}}) ^ {T}. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.168, + 0.483, + 0.2 + ], + "angle": 0, + "content": "If \\(\\alpha_{i} > \\frac{1}{255}\\), then the Gaussian is included in the alpha compositing of the pixel color \\(C\\), given by:" + }, + { + "type": "equation", + "bbox": [ + 0.163, + 0.21, + 0.483, + 0.253 + ], + "angle": 0, + "content": "\\[\nC (p) = \\sum_ {i \\in \\mathcal {N}} c _ {i} \\alpha_ {i} (p) \\prod_ {j = 1} ^ {i - 1} (1 - \\alpha_ {j} (p)). \\tag {7}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.262, + 0.192, + 0.278 + ], + "angle": 0, + "content": "4. Methods" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.288, + 0.484, + 0.38 + ], + "angle": 0, + "content": "Our Speedy-Splat methods are motivated by two key insights into inefficiencies within the 3D-GS rendering pipeline. First, Gaussian Splatting grossly overestimates the extent of Gaussians in the image. Second, as demonstrated by [11] and other recent pruning works, 3D-GS models are heavily overparameterized." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.387, + 0.29, + 0.402 + ], + "angle": 0, + "content": "4.1. Precise Tile Intersect" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.409, + 0.484, + 0.485 + ], + "angle": 0, + "content": "Gaussian Splitting identifies tiles intersected by Gaussian \\(\\mathcal{G}_i\\) by calculating the maximum eigenvalue \\(\\lambda_{\\mathrm{max}}\\) of its projected 2D covariance \\(\\Sigma_{i_2D}\\), then selecting all tiles that intersect the square inscribing the circle defined by center \\(\\mu_{i_{2D}}\\) and radius:" + }, + { + "type": "equation", + "bbox": [ + 0.227, + 0.495, + 0.483, + 0.52 + ], + "angle": 0, + "content": "\\[\nr = \\left\\lceil 3 \\sqrt {\\lambda_ {\\max }} \\right\\rceil . \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.524, + 0.484, + 0.644 + ], + "angle": 0, + "content": "This approach neglects opacity \\(\\sigma_{i}\\) in its calculation and generally overestimates the Gaussian extent, as illustrated by Figure 2a. The actual extent of Gaussian \\(\\mathcal{G}_i\\), shown in Figure 2b, is given by the threshold placed on its alpha value \\(\\alpha_{i}\\). Specifically, \\(\\mathcal{G}_i\\) does not contribute to the rendering of pixel \\(p\\) if \\(\\alpha_{i} < \\frac{1}{255}\\) for \\(p\\). By applying the actual extent in tile intersection calculations, we arrive at a far more concise set of intersected tiles." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.645, + 0.484, + 0.705 + ], + "angle": 0, + "content": "We now show the derivation of this extent. The furthest pixel extent of Gaussian \\(\\mathcal{G}_i\\) can be determined by directly substituting this threshold into Equation 5. Rearranging terms gives:" + }, + { + "type": "equation", + "bbox": [ + 0.115, + 0.713, + 0.483, + 0.733 + ], + "angle": 0, + "content": "\\[\n2 \\log (2 5 5 \\sigma_ {i}) = (p - \\mu_ {i _ {2 D}}) \\boldsymbol {\\Sigma} _ {\\boldsymbol {i _ {2 D}}} ^ {- 1} (p - \\mu_ {i _ {2 D}}) ^ {T}. \\qquad (9)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.741, + 0.199, + 0.755 + ], + "angle": 0, + "content": "We can rewrite:" + }, + { + "type": "equation", + "bbox": [ + 0.107, + 0.762, + 0.483, + 0.797 + ], + "angle": 0, + "content": "\\[\np = \\binom {p _ {x}} {p _ {y}}, \\mu_ {i _ {2 D}} = \\binom {\\mu_ {x}} {\\mu_ {y}}, \\boldsymbol {\\Sigma} _ {\\boldsymbol {i} _ {2 D}} ^ {- 1} = \\left( \\begin{array}{c c} a & b \\\\ b & c \\end{array} \\right). \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.804, + 0.483, + 0.819 + ], + "angle": 0, + "content": "Specifying threshold \\(t\\) and centered coordinates \\(x_{d}\\) and \\(y_{d}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.226, + 0.844, + 0.483, + 0.86 + ], + "angle": 0, + "content": "\\[\nt = 2 \\log (2 5 5 \\sigma_ {i}), \\tag {11}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.236, + 0.864, + 0.483, + 0.879 + ], + "angle": 0, + "content": "\\[\nx _ {d} = p _ {x} - \\mu_ {x}, \\tag {12}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.236, + 0.883, + 0.483, + 0.898 + ], + "angle": 0, + "content": "\\[\ny _ {d} = p _ {y} - \\mu_ {y}, \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.122 + ], + "angle": 0, + "content": "gives the pixel extent with coordinates \\( x_{d} \\) and \\( y_{d} \\) that satisfy the ellipse equation:" + }, + { + "type": "equation", + "bbox": [ + 0.622, + 0.134, + 0.905, + 0.153 + ], + "angle": 0, + "content": "\\[\nt = a x _ {d} ^ {2} + 2 b x _ {d} y _ {d} + c y _ {d} ^ {2}. \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.166, + 0.907, + 0.271 + ], + "angle": 0, + "content": "Our approach uses this pixel extent to reduce the number of tiles contained in the Gaussian-to-tile mappings for each Gaussian. We propose two methods for computing precise tile intersections. First, our SnugBox algorithm produces a tight bounding box around each Gaussian. Second, our AccuTile algorithm extends it to identify the exact set of tiles intersected by the Gaussian." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.281, + 0.623, + 0.296 + ], + "angle": 0, + "content": "4.1.1. SnugBox" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.301, + 0.906, + 0.361 + ], + "angle": 0, + "content": "Our SnugBox method uses this elliptical extent to compute an axis-aligned bounding box that more precisely identifies tiles intersected by Gaussian \\(\\mathcal{G}_i\\). To derive this bounding box, we rearrange Equation 14 to solve for \\(y_d\\):" + }, + { + "type": "equation", + "bbox": [ + 0.591, + 0.373, + 0.905, + 0.407 + ], + "angle": 0, + "content": "\\[\ny _ {d} = \\frac {- b x _ {d} \\pm \\sqrt {\\left(b ^ {2} - a c\\right) x _ {d} ^ {2} + t c}}{c}. \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.416, + 0.906, + 0.495 + ], + "angle": 0, + "content": "To find the \\(y\\)-coordinate bounding box edges \\(y_{\\mathrm{min}}\\) and \\(y_{\\mathrm{max}}\\), we identify the values of \\(x_d\\) where \\(\\partial y_d / \\partial x_d = 0\\). We refer to these \\(x_d\\) values as \\(x_{d_{\\mathrm{arg}s}}\\) to specify that they are the \\(\\arg \\min y_d\\) and \\(\\arg \\max y_d\\) values. Differentiating Equation 15 and solving for \\(x_{d_{\\mathrm{arg}s}}\\) gives:" + }, + { + "type": "equation", + "bbox": [ + 0.621, + 0.506, + 0.905, + 0.547 + ], + "angle": 0, + "content": "\\[\nx _ {d _ {\\text {a r g s}}} = \\pm \\sqrt {\\frac {- b ^ {2} t}{\\left(b ^ {2} - a c\\right) a}}. \\tag {16}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.554, + 0.906, + 0.631 + ], + "angle": 0, + "content": "Substituting \\( x_{d_{\\text{args}}} \\) into Equation 15 and adding \\( \\mu_y \\) gives \\( y_{\\min} \\) and \\( y_{\\max} \\). Due to the symmetry of Equation 14, we can find the \\( x \\)-coordinate bounding box edges \\( x_{\\min} \\) and \\( x_{\\max} \\) by swapping \\( y_d \\) and \\( x_d \\) and constants \\( a \\) and \\( c \\) to rewrite Equations 15 and 16 in terms of \\( x_d \\) and \\( y_{d_{\\text{args}}} \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.631, + 0.907, + 0.812 + ], + "angle": 0, + "content": "After the bounding box edges are identified, our method follows 3D-GS and converts these edges to tile indices by dividing by tile size, rounding, and clipping to the image boundary. As depicted in Figure 2, Snugbox can produce a significantly tighter bounding box than 3D-GS. Meanwhile, its computational overhead is small as it performs a constant number of operations and is only called twice in the rendering pipeline – once to count the Gaussians intersecting tiles in preprocess and once to populate the Gaussian-to-tile arrays in duplicateWithkeys. Table 1 reports that SnugBox improves the efficiency of all downstream functions and produces an average overall speed-up of \\(1.82 \\times\\)." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.821, + 0.623, + 0.835 + ], + "angle": 0, + "content": "4.1.2. AccuTile" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.906, + 0.902 + ], + "angle": 0, + "content": "Our AccuTile method, outlined in Algorithm 1, extends SnugBox to identify the exact tiles intersected by the Gaussian. It takes as input the tight bounding box produced by Snugbox and its rectangular tile extent - depicted as the" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "21540" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.113, + 0.095, + 0.362, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.167, + 0.28, + 0.306, + 0.293 + ], + "angle": 0, + "content": "(a) 3D Gaussian Splitting" + }, + { + "type": "image", + "bbox": [ + 0.373, + 0.095, + 0.625, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.465, + 0.28, + 0.534, + 0.293 + ], + "angle": 0, + "content": "(b) SnugBox" + }, + { + "type": "image", + "bbox": [ + 0.639, + 0.095, + 0.887, + 0.274 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.728, + 0.28, + 0.796, + 0.292 + ], + "angle": 0, + "content": "(c) AccuTile" + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.304, + 0.907, + 0.388 + ], + "angle": 0, + "content": "Figure 2. Gaussian tile allocation by method. (a) 3D Gaussian Splating allocates a Gaussian to a tile when that tile intersects the square inscribing the circle with radius \\(\\lceil 3\\sqrt{\\lambda_{\\max}}\\rceil\\) defined in Equation 8. (b) Our SnugBox method allocates a Gaussian to a tile when that tile intersects the tight bounding box defined by the axis-aligned minima and maxima of the ellipse given by Equation 14. (c) Our AccuTile method allocates a Gaussian to a tile only if that tile intersects the ellipse via Algorithm 1, which computes the minimum and maximum tiles by iterating over the shorter side of the rectangular tile extent given by SnugBox. In this example, our AccuTile algorithm iterates over the tile rows; the only points that are processed are \\(x_{min}\\), \\(x_{max}\\), A, B, C, and D." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.408, + 0.484, + 0.528 + ], + "angle": 0, + "content": "blue box and yellow tiles, respectively, in Figure 2b. Depending on which dimension is smaller, AccuTile then processes either the rows or columns of this tile extent to determine the exact tiles that intersect the Gaussian. Specifically, we identify the minimum and maximum extent of the ellipse within a given row or column, then convert those points to the corresponding touched tiles. All tiles between the minimum and maximum tiles intersect the ellipse." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.534, + 0.484, + 0.791 + ], + "angle": 0, + "content": "AccuTile's key insight is that calculating the minimum and maximum extent of the ellipse within each row or column requires computing only two points per iteration. The only possible inflection points of the elliptical curve are the bounding box minimum and maximum points identified by Snugbox. If one or both of these points lie within the tile row or column, then they represent the minimum or maximum extent of the ellipse there. If neither point is within the row or column, then the points along that boundary side are monotonically decreasing or increasing – the minimum or maximum point must lie on one of the boundary lines of that row or column. Since the boundary of the last row or column is shared with the next one, we only need to compute the intersection of the ellipse with the next boundary line in each iteration. Thus, our AccuTile algorithm counts tiles in time proportional to the shorter side of the tile extent and processes tiles in time proportional to the tile count." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Figure 2 illustrates how Accutile restricts the tight bounding box produced by SnugBox to the exact tiles that the Gaussian touches. The tile extent rows from Snugbox, shown in Figure 2b, are processed starting from the bottom. The lower tile row boundary falls below the bounding box, so no initial intersection is computed. The upper boundary lies below the top of the bounding box, so intersection" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.408, + 0.907, + 0.589 + ], + "angle": 0, + "content": "points \\( A \\) and \\( B \\) are calculated using Equation 15. \\( x_{\\mathrm{min}} \\) is assigned as this row's minimum ellipse extent \\( e_{\\mathrm{min}} \\) because it lies within it, and \\( B \\) is assigned as its maximum ellipse extent \\( e_{\\mathrm{max}} \\) because it is the maximal point. Consequently, this row's tile extent is from the tile containing \\( x_{\\mathrm{min}} \\) to the tile containing \\( B \\). For the next row, we keep points \\( A \\) and \\( B \\) and compute upper boundary points \\( C \\) and \\( D \\). \\( A \\) and \\( x_{\\mathrm{max}} \\) are assigned as the row's \\( e_{\\mathrm{min}} \\) and \\( e_{\\mathrm{max}} \\), and the tiles containing them are its tile extent. Finally, for the last row, \\( C \\) and \\( D \\) are kept; no additional points are computed because the row's upper boundary is above the bounding box. Thus, the tile extent of this row is between \\( C \\) and \\( D \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.59, + 0.907, + 0.667 + ], + "angle": 0, + "content": "The number of tiles is further reduced from Figure 2b to Figure 2c, much less than the original 3D-GS method in Figure 2a. In Table 1, we report that AccuTile further accelerates all downstream functions, culminating in an average overall speed-up of \\(1.99 \\times\\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.679, + 0.685, + 0.695 + ], + "angle": 0, + "content": "4.2. Efficient Pruning" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.702, + 0.906, + 0.763 + ], + "angle": 0, + "content": "PUP 3D-GS [11] is a pruning method that quantifies the sensitivity of each Gaussian to training views, removing a set percentage with the lowest sensitivities. Sensitivity is computed by approximating the Hessian of the \\( L_{2} \\) loss:" + }, + { + "type": "equation", + "bbox": [ + 0.555, + 0.776, + 0.907, + 0.81 + ], + "angle": 0, + "content": "\\[\nH = \\nabla_ {\\mathcal {G}} ^ {2} L _ {2} = \\sum_ {\\phi \\in \\mathcal {P} _ {g t}} \\nabla_ {\\mathcal {G}} I _ {\\mathcal {G}} (\\phi) \\nabla_ {\\mathcal {G}} I _ {\\mathcal {G}} (\\phi) ^ {T}, \\tag {17}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.824, + 0.906, + 0.87 + ], + "angle": 0, + "content": "where \\(\\nabla_{\\mathcal{G}}I_{\\mathcal{G}}(\\phi)\\) is the gradient over all Gaussian parameters on the rendered image \\(I_{\\mathcal{G}}\\) for pose \\(\\phi\\). \\(H\\) is shown to be exact when the \\(L_{1}\\) residual error vanishes [11]." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.906, + 0.902 + ], + "angle": 0, + "content": "A per-Gaussian sensitivity can be derived by splitting \\( H \\) into the block diagonal elements that only capture inter" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "21541" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.091, + 0.485, + 0.198 + ], + "angle": 0, + "content": "Algorithm 1 The AccuTile Algorithm. For simplicity, the algorithm outlined here is applied to the rows of the Snug-Box tile extent bounding box, matching the example in Figure 2c. In practice, it is applied along the smaller side of the tile extent. The subscripts \\( t \\), \\( b \\), \\( l \\), and \\( r \\) represent the top, bottom, left, and right sides, respectively. A proof of correctness sketch is presented in Appendix A.1." + }, + { + "type": "algorithm", + "bbox": [ + 0.09, + 0.208, + 0.484, + 0.581 + ], + "angle": 0, + "content": "Require: Ellipse E \\(\\triangleright\\) Eq. 14 \nRequire: SnugBox Bounding Box B \nRequire: SnugBox Tile Extent Rectangle R \nInitialize: Tile count \\(C\\gets 0\\) \nline \\(\\mathrm{min}\\gets \\mathbf{R}_b\\) \nif line \\(\\mathrm{min}\\geq \\mathbf{B}_b\\) then \n\\(\\mathbf{i}_{\\mathrm{min}}\\gets\\) Intersections( line \\(\\mathrm{min}\\) ,E) \\(\\triangleright\\) Eq. 15 \nend if \nfor row r in R do \nline \\(\\mathrm{max}\\gets \\mathrm{r}_t\\) \nif line \\(\\mathrm{max}\\leq \\mathbf{B}_t\\) then \n\\(\\mathbf{i}_{\\mathrm{max}}\\gets\\) Intersections( line \\(\\mathrm{max}\\) ,E) \\(\\triangleright\\) Eq. 15 \nend if \ne \\(\\mathrm{min}\\gets \\mathbf{B}_l\\) if Bl in r else min(i \\(\\mathrm{min}\\) ,i max) \ne \\(\\mathrm{max}\\gets\\) Br if Br in r else max(i \\(\\mathrm{min}\\) ,i max) \ntile \\(\\mathrm{min}\\) ,tile \\(\\mathrm{max}\\gets\\) Convert(e \\(\\mathrm{min}\\) ,e max) \nC \\(\\leftarrow\\) C + (tile \\(\\mathrm{max}\\) -tile \\(\\mathrm{min}\\) ) \nProcess( tile \\(\\mathrm{min}\\) ,tile \\(\\mathrm{max}\\) ) \n\\(\\mathbf{i}_{\\mathrm{min}}\\gets \\mathbf{i}_{\\mathrm{max}}\\) \nend for \nreturn C" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.616, + 0.318, + 0.632 + ], + "angle": 0, + "content": "Gaussian parameter relationships:" + }, + { + "type": "equation", + "bbox": [ + 0.17, + 0.64, + 0.483, + 0.675 + ], + "angle": 0, + "content": "\\[\nH _ {i} = \\sum_ {\\phi \\in \\mathcal {P} _ {g t}} \\nabla_ {\\mathcal {G} _ {i}} I _ {\\mathcal {G}} (\\phi) \\nabla_ {\\mathcal {G} _ {i}} I _ {\\mathcal {G}} (\\phi) ^ {T}, \\tag {18}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.683, + 0.483, + 0.729 + ], + "angle": 0, + "content": "where \\(\\nabla_{\\mathcal{G}_i}\\) is the gradient with respect to only \\(\\mathcal{G}_i\\). This measures the sensitivity of the \\(L_2\\) loss with respect to Gaussian \\(\\mathcal{G}_i\\), assuming all other Gaussians are held constant." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.729, + 0.484, + 0.79 + ], + "angle": 0, + "content": "\\(H_{i}\\) is again approximated by only using the six mean \\(\\mu_{i}\\) and scale \\(s_i\\) parameters to specifically capture geometric sensitivity. The log determinant is taken to provide a representative scalar score \\(U_{i}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.186, + 0.798, + 0.483, + 0.817 + ], + "angle": 0, + "content": "\\[\nU _ {i} = \\log | \\nabla_ {\\mu_ {i}, s _ {i}} I _ {\\mathcal {G}} \\nabla_ {\\mu_ {i}, s _ {i}} I _ {\\mathcal {G}} ^ {T} |. \\tag {19}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.483, + 0.856 + ], + "angle": 0, + "content": "Using this score, up to \\(90\\%\\) of Gaussians can be robustly pruned from the model while retaining high visual quality." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Although PUP 3D-GS touts high compression ratios and rendering speeds, we identify two key drawbacks in its formulation. First, computing the Hessian requires storage" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.168 + ], + "angle": 0, + "content": "proportional to \\( N \\times 36 \\), where \\( N \\) is the number of Gaussians. In comparison, the 3D-GS model has a memory footprint proportional to \\( N \\times 59 \\) because it stores 59 parameters per Gaussian. While this score is effective for post-hoc pruning, using it during training is impractical." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.168, + 0.907, + 0.318 + ], + "angle": 0, + "content": "Second, computing the Hessian requires the pixel-wise gradients of \\(\\mu\\) and \\(s\\). Since these are 3D parameters of the Gaussian primitives, obtaining their gradients requires back-propagating through the render kernel parallelized per pixel, then back-propagating each Gaussian contributing to that pixel in its thread. This breaks the efficient flow of gradients in 3D-GS, where the per-pixel gradients from the render kernel are parallelized and aggregated to the 2D \\(\\mu_{2D}\\) and \\(\\Sigma_{2D}\\) parameters, which are then parallelized across Gaussians to compute gradients for \\(\\mu\\) and \\(s\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.318, + 0.906, + 0.424 + ], + "angle": 0, + "content": "Our approach builds on PUP 3D-GS by introducing an efficient pruning score that we incorporate into the 3D-GS training pipeline. We also define two distinct pruning modalities: Soft Pruning, which takes place during densification in the first 15000 iterations, and Hard Pruning, which is applied after densification is completed after iteration 15000." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.43, + 0.724, + 0.445 + ], + "angle": 0, + "content": "4.2.1. Efficient Pruning Score" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.448, + 0.906, + 0.554 + ], + "angle": 0, + "content": "Our insight is that both drawbacks can be alleviated by reparameterizing the Hessian. Concretely, we express the influence of all spatial parameters of Gaussian \\(\\mathcal{G}_i\\) by computing the Hessian approximation with respect to the 2D projected value of \\(\\mathcal{G}_i\\) at pixel \\(p\\), given by \\(g_i(p)\\) in Equation 6. In doing so, the pruning score \\(U_i\\) from Equation 19 can be expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.625, + 0.567, + 0.905, + 0.586 + ], + "angle": 0, + "content": "\\[\n\\tilde {U} _ {i} = \\log | \\nabla_ {g _ {i}} I _ {\\mathcal {G}} \\nabla_ {g _ {i}} I _ {\\mathcal {G}} ^ {T} |. \\tag {20}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.589, + 0.906, + 0.618 + ], + "angle": 0, + "content": "Since \\( g_{i} \\) is a scalar and log is monotonically increasing, we can rewrite this score as:" + }, + { + "type": "equation", + "bbox": [ + 0.654, + 0.625, + 0.905, + 0.644 + ], + "angle": 0, + "content": "\\[\n\\tilde {U} _ {i} = \\left(\\nabla_ {g _ {i}} I _ {\\mathcal {G}}\\right) ^ {2}. \\tag {21}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.65, + 0.907, + 0.742 + ], + "angle": 0, + "content": "Gradient \\(\\nabla_{g_i}I_{\\mathcal{G}}\\) is already computed in the backward pass of render and can be efficiently squared and aggregated across all pixels. Moreover, the maximum space requirement for this score is proportional to the number of Gaussians \\(N\\), reducing the storage requirement by \\(36\\times\\) and allowing this score to be used during training." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.747, + 0.651, + 0.762 + ], + "angle": 0, + "content": "4.2.2. Soft Pruning" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.907, + 0.902 + ], + "angle": 0, + "content": "To maintain a robust Hessian approximation, we observe that the \\( L_{1} \\) loss becomes quite small by iteration 6000 and remains low except after an opacity reset is performed. As such, we augment the densification pipeline to include our Soft Pruning method, where the model is pruned immediately before the three opacity resets at 6000, 9000, and 12000 iterations. Surprisingly, we find that we can set extremely high Soft Pruning ratios – in our experiments, visual fidelity is preserved at \\( 80\\% \\) pruning." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "21542" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.176, + 0.091, + 0.203, + 0.105 + ], + "angle": 0, + "content": "GT" + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.109, + 0.284, + 0.205 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.37, + 0.092, + 0.425, + 0.106 + ], + "angle": 0, + "content": "3D-GS" + }, + { + "type": "image", + "bbox": [ + 0.306, + 0.109, + 0.493, + 0.204 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.312, + 0.204, + 0.408, + 0.216 + ], + "angle": 0, + "content": "2.5M Gaussians" + }, + { + "type": "image_caption", + "bbox": [ + 0.411, + 0.204, + 0.48, + 0.215 + ], + "angle": 0, + "content": "172.15 FPS" + }, + { + "type": "image_caption", + "bbox": [ + 0.577, + 0.092, + 0.614, + 0.107 + ], + "angle": 0, + "content": "PUP" + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.109, + 0.686, + 0.204 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.508, + 0.204, + 0.606, + 0.216 + ], + "angle": 0, + "content": "0.25M Gaussians" + }, + { + "type": "image", + "bbox": [ + 0.698, + 0.092, + 0.892, + 0.204 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.706, + 0.204, + 0.805, + 0.216 + ], + "angle": 0, + "content": "0.22M Gaussians" + }, + { + "type": "image_caption", + "bbox": [ + 0.813, + 0.204, + 0.884, + 0.215 + ], + "angle": 0, + "content": "1277.20 FPS" + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.224, + 0.284, + 0.321 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.306, + 0.224, + 0.493, + 0.32 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.312, + 0.321, + 0.408, + 0.332 + ], + "angle": 0, + "content": "4.5M Gaussians" + }, + { + "type": "image_caption", + "bbox": [ + 0.411, + 0.321, + 0.48, + 0.332 + ], + "angle": 0, + "content": "141.47 FPS" + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.224, + 0.686, + 0.32 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.508, + 0.321, + 0.606, + 0.332 + ], + "angle": 0, + "content": "0.45M Gaussians" + }, + { + "type": "image_caption", + "bbox": [ + 0.612, + 0.321, + 0.68, + 0.331 + ], + "angle": 0, + "content": "261.67 FPS" + }, + { + "type": "image", + "bbox": [ + 0.7, + 0.224, + 0.888, + 0.32 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.707, + 0.321, + 0.805, + 0.331 + ], + "angle": 0, + "content": "0.46M Gaussians" + }, + { + "type": "image_caption", + "bbox": [ + 0.813, + 0.321, + 0.882, + 0.331 + ], + "angle": 0, + "content": "723.79 FPS" + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.34, + 0.285, + 0.436 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.306, + 0.341, + 0.493, + 0.435 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.312, + 0.436, + 0.408, + 0.447 + ], + "angle": 0, + "content": "3.4M Gaussians" + }, + { + "type": "image_caption", + "bbox": [ + 0.411, + 0.436, + 0.48, + 0.447 + ], + "angle": 0, + "content": "125.97 FPS" + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.341, + 0.686, + 0.436 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.508, + 0.436, + 0.605, + 0.447 + ], + "angle": 0, + "content": "0.34M Gaussians" + }, + { + "type": "image_caption", + "bbox": [ + 0.612, + 0.436, + 0.679, + 0.447 + ], + "angle": 0, + "content": "413.30 FPS" + }, + { + "type": "image", + "bbox": [ + 0.7, + 0.341, + 0.889, + 0.436 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.707, + 0.436, + 0.805, + 0.447 + ], + "angle": 0, + "content": "0.32M Gaussians" + }, + { + "type": "image_caption", + "bbox": [ + 0.813, + 0.436, + 0.882, + 0.447 + ], + "angle": 0, + "content": "1121.97 FPS" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.459, + 0.907, + 0.502 + ], + "angle": 0, + "content": "Figure 3. Visual comparison of 3D-GS, PUP 3D-GS, and our method. Notice that, while reaching similar compression ratios to PUP 3D-GS, our Speedy-Splat method delivers vastly faster rendering speeds. Top: playroom from the Deep Blending dataset. Middle: bicycle from the Mip-NeRF 360 dataset. Bottom: drjohnson from the Deep Blending dataset." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.521, + 0.236, + 0.537 + ], + "angle": 0, + "content": "4.2.3. Hard Pruning" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.54, + 0.484, + 0.691 + ], + "angle": 0, + "content": "We also observe that the model's performance after densification closely matches that of the fully-trained model. The iterations after the densification stage essentially finetune the model and can be used to further \"refine\" it after pruning, similar to PUP 3D-GS [11] and LightGaussian [6]. In practice, our Hard Pruning method prunes the model by a constant ratio every 3000 iterations starting at iteration 15000. We Hard Prune \\(30\\%\\) of Gaussians in each interval, which, when paired with Soft Pruning, empirically reduces the total number of Gaussians across scenes by \\(10.6\\times\\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.703, + 0.222, + 0.72 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.728, + 0.192, + 0.743 + ], + "angle": 0, + "content": "5.1. Datasets" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Our evaluation uses the same set of challenging real-world scenes as 3D-GS [14]. This includes nine Mip-Nerf 360 scenes [3] – four indoor and five outdoor – that each feature a complex central object or area with a detailed background. We also include the outdoor train and truck scenes from the Tanks & Temples dataset [17] and the indoor drjohnson and playroom scenes from the Deep Blending dataset [12]. For consistency across experiments, we use the COLMAP pose estimates and sparse point clouds provided by the dataset authors." + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.519, + 0.907, + 0.602 + ], + "angle": 0, + "content": "Table 2. Average Gaussian count, FPS, and training time across all scenes in Section 5.1. Ratios for model size compression, rendering speed-up, and training speed-up are reported in (parentheses). The operation in each row is applied cumulatively to all of the following rows. The best and second best value for each metric are color coded." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.61, + 0.9, + 0.701 + ], + "angle": 0, + "content": "
Method# Gaussians ↓FPS ↑Training Time ↓
Baseline2.93M13423.2
+SnugBox2.97M244 (1.82×)21.2 (1.09×)
+AccuTile2.97M267 (1.99×)21.0 (1.10×)
+Soft Pruning1.64M (1.79×)420 (3.14×)17.5 (1.32×)
+Hard Pruning0.28M (10.6×)898 (6.71×)15.7 (1.47×)
" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.734, + 0.73, + 0.75 + ], + "angle": 0, + "content": "5.2. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Our code builds on the differentiable renderer provided by 3D-GS [14] and modifies the Python training pipeline for pruning schedules and execution. To ensure consistent and precise timing, all times in Table 1 and FPS values in Tables 2 and 3 are measured using CUDA events at the start and end of the forward rendering procedure. All experiments are conducted on an Nvidia RTXA5000 GPU, and the reported Speedy-Splat results represent the average metrics across three independent runs for each scene." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "21543" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.482, + 0.228 + ], + "angle": 0, + "content": "Table 3. Average reported metrics for each pruning method across all scenes in the Mip-NeRF 360 dataset. The Comp column reports model size compression in terms of Gaussian count, FPS reports rendering speed-up, and Train reports training time speed-up, all with respect to the baseline 3D-GS model. PSNR, SSIM, and LPIPS are also recorded. For a fair comparison, we report the published results of each method and use '-' to denote missing metrics. The best and second best value for each metric are color coded; lossless methods are underlined. Results for the Tanks & Temples and Deep Blending datasets are reported in Appendix A.3." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.239, + 0.479, + 0.417 + ], + "angle": 0, + "content": "
MethodComp ↑FPS ↑Train ↑PSNR ↑SSIM ↑LPIPS ↓
3D-GS [14]1.00×1.00×1.00×27.550.8140.222
Trimming [2]4.00×--27.130.7980.248
Compact [18]2.28×1.07×0.73×27.080.7980.247
EAGLES [10]3.68×1.51×1.37×26.940.8000.250
Reducing [26]2.33×1.60×1.23×27.100.8090.226
Light [6]2.94×1.76×-27.280.8050.243
ELMGS [1]5.00×2.69×-27.000.7790.286
PUP [11]8.65×2.55×-26.830.7920.268
Mini-Splat [7]6.84×3.20×1.26×27.340.8220.217
+SnugBox0.99×1.81×1.08×27.550.8140.221
+AccuTile0.99×1.99×1.10×27.570.8140.221
+Soft Pruning1.79×3.14×1.30×27.320.8070.246
+Hard Pruning10.6×6.51×1.45×26.940.7820.296
" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.436, + 0.185, + 0.451 + ], + "angle": 0, + "content": "5.3. Results" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.459, + 0.347, + 0.474 + ], + "angle": 0, + "content": "5.3.1. Additive method performance" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.479, + 0.483, + 0.675 + ], + "angle": 0, + "content": "The efficacy of Speedy-Splat is demonstrated by Table 1, where we record the average execution times of each function across all scenes when additively applying our methods. SnugBox and AccuTile each introduce minimal additional computation time to preprocess and InclusiveSum. However, limiting the number of tiles touched accelerates all downstream functions, culminating in an overall speed-up of \\(1.82 \\times\\) by SnugBox that is raised to \\(1.99 \\times\\) by AccuTile. Applying soft pruning reduces the runtime of all functions by reducing the number of Gaussians, leading to a \\(3.14 \\times\\) overall speed-up. Finally, performing hard pruning improves overall speed by a whopping \\(6.71 \\times\\) over the baseline 3D-GS model." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.685, + 0.285, + 0.699 + ], + "angle": 0, + "content": "5.3.2. Overall Performance" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.704, + 0.483, + 0.809 + ], + "angle": 0, + "content": "In Figures 1 and 3, we report qualitative results on two outdoor and two indoor scenes from all three datasets. The magnified regions highlight that our method preserves the fine details in the baseline 3D-GS scene and closely models the ground truth view. Despite touting similar compression ratios and rendering nearly identical images, Speedy-Splat achieves over double the FPS of PUP 3D-GS." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Table 3 compares our methods against other methods that reduce the number of Gaussians and increase inference speed using the mean of each metric across all scenes in the Mip-NeRF 360 dataset. The underlined methods are \"lossless\", meaning that they avoid degrading visual fidelity at all. SnugBox and AccuTile, our lossless methods, improve" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.089, + 0.905, + 0.173 + ], + "angle": 0, + "content": "Table 4. Average metrics across all scenes in Section 5.1 when using the Speedy-Splat and PUP 3D-GS pruning scores to prune \\(88.44\\%\\) of Gaussians using the PUP 3D-GS pipeline. Two rounds of prune-refine are performed on each baseline 3D-GS model, pruning \\(66\\%\\) of Gaussians and then fine-tuning for 5,000 iterations in each one. The best value for each metric is color coded." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.184, + 0.905, + 0.248 + ], + "angle": 0, + "content": "
Method# Gaussians ↓PSNR ↑SSIM ↑LPIPS ↓FPS ↑
Baseline2.92M27.15030.82960.2238107.53
PUP [11]0.34M26.21360.80440.2731378.57
Ours0.34M26.86580.80220.2840345.52
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.269, + 0.907, + 0.36 + ], + "angle": 0, + "content": "rendering and training speed while leaving image quality metrics essentially unchanged or slightly better. Our full pipeline, labeled as “+Hard Pruning” boasts the highest compression ratios, rendering speeds, and training speed-ups across all datasets. Furthermore, its image quality metrics are competitive with the other methods." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.378, + 0.761, + 0.395 + ], + "angle": 0, + "content": "5.4. Pruning Score Comparison" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.402, + 0.907, + 0.522 + ], + "angle": 0, + "content": "Although the primary focus of our work is rendering speed, we find that our efficient pruning score, described in Section 4.2, also performs well when applied in other compression pipelines. In Table 4, we ablate our efficient pruning score with the PUP 3D-GS sensitivity score across all scenes in their post-hoc pruning pipeline. Notably, Speedy-Splat's efficient pruning score outperforms PUP 3D-GS on PSNR and is competitive across the other metrics." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.545, + 0.637, + 0.56 + ], + "angle": 0, + "content": "6. Limitations" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.568, + 0.907, + 0.69 + ], + "angle": 0, + "content": "A limitation of Speedy-Splat is that it produces slightly lower image quality than 3D-GS. However, this degradation is expected at high compression ratios and is also observed in comparable techniques. Additionally, a direct comparison of our efficient pruning score to the PUP 3D-GS pruning score illuminates a slight, yet noticeable, gap in performance. Future work could explore the possibility of another efficient pruning score that delivers higher performance." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.711, + 0.634, + 0.726 + ], + "angle": 0, + "content": "7. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In this work, we present Speedy-Splat: a new 3D-GS technique that accurately localizes Gaussians during rendering and significantly improves inference speed, model size, and training time. Enhanced localization is achieved by our SnugBox and AccuTile methods, while model size reduction is accomplished by our Soft and Hard Pruning approaches. Together, our Speedy-Splat methods accelerate rendering speed by an average of \\(6.71 \\times\\), reduce model size by \\(10.6 \\times\\), and improve training time by \\(1.47 \\times\\) across all scenes from the Mip-NeRF 360, Tanks & Temples, and Deep Blending datasets." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "21544" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.279, + 0.108 + ], + "angle": 0, + "content": "8. Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.109, + 0.484, + 0.231 + ], + "angle": 0, + "content": "This work was made possible by the IARPA WRIVA Program, the ONR MURI program, and DAPRA TIAMAT. Commercial support was provided by Capital One Bank, the Amazon Research Award program, and Open Philanthropy. Further support was provided by the National Science Foundation (IIS-2212182), and by the NSF TRAILS Institute (2229885). Zwicker was additionally supported by the National Science Foundation (IIS-2126407)." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.243, + 0.188, + 0.259 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.268, + 0.484, + 0.324 + ], + "angle": 0, + "content": "[1] Muhammad Salman Ali, Sung-Ho Bae, and Enzo Tartaglione. Elmgs: Enhancing memory and computation scalability through compression for 3d gaussian splatting. arXiv preprint arXiv:2410.23213, 2024. 2, 8, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.325, + 0.484, + 0.379 + ], + "angle": 0, + "content": "[2] Muhammad Salman Ali, Maryam Qamar, Sung-Ho Bae, and Enzo Tartaglione. Trimming the fat: Efficient compression of 3d gaussian splats through pruning. arXiv preprint arXiv:2406.18214, 2024. 2, 8, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.381, + 0.484, + 0.45 + ], + "angle": 0, + "content": "[3] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5470–5479, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.451, + 0.483, + 0.492 + ], + "angle": 0, + "content": "[4] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In European Conference on Computer Vision (ECCV), 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.494, + 0.484, + 0.548 + ], + "angle": 0, + "content": "[5] Sankeerth Durvasula, Adrian Zhao, Fan Chen, Ruofan Liang, Pawan Kumar Sanjaya, and Nandita Vijaykumar. Distwar: Fast differentiable rendering on raster-based rendering pipelines. arXiv preprint arXiv:2401.05345, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.55, + 0.484, + 0.605 + ], + "angle": 0, + "content": "[6] Zhiwen Fan, Kevin Wang, Kairun Wen, Zehao Zhu, Dejia Xu, and Zhangyang Wang. Lightgaussian: Unbounded 3d gaussian compression with 15x reduction and \\(200+\\) fps. arXiv preprint arXiv:2311.17245, 2023. 1, 2, 7, 8, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.607, + 0.484, + 0.647 + ], + "angle": 0, + "content": "[7] Guangchi Fang and Bing Wang. Mini-splatting: Representing scenes with a constrained number of gaussians. arXiv preprint arXiv:2403.14166, 2024. 1, 2, 8, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.649, + 0.484, + 0.716 + ], + "angle": 0, + "content": "[8] Guofeng Feng, Siyan Chen, Rong Fu, Zimu Liao, Yi Wang, Tao Liu, Zhilin Pei, Hengjie Li, Xingcheng Zhang, and Bo Dai. Flashgs: Efficient 3d gaussian splatting for large-scale and high-resolution rendering. arXiv preprint arXiv:2408.07967, 2024.1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.719, + 0.484, + 0.746 + ], + "angle": 0, + "content": "[9] Henri P. Gavin. The levenberg-marquardt method for nonlinear least squares curve-fitting problems c ©. 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.748, + 0.483, + 0.788 + ], + "angle": 0, + "content": "[10] Sharath Girish, Kamal Gupta, and Abhinav Shrivastava. Eagles: Efficient accelerated 3d gaussians with lightweight encodings. arXiv preprint arXiv:2312.04564, 2023. 2, 8, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.79, + 0.484, + 0.844 + ], + "angle": 0, + "content": "[11] Alex Hanson, Allen Tu, Vasu Singla, Mayuka Jayawardhana, Matthias Zwicker, and Tom Goldstein. Pup 3d-gs: Principled uncertainty pruning for 3d gaussian splattering. arXiv preprint arXiv:2406.10219, 2024. 1, 2, 4, 5, 7, 8, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[12] Peter Hedman, Julien Philip, True Price, Jan-Michael Frahm, George Drettakis, and Gabriel Brostow. Deep blending for free-viewpoint image-based rendering. ACM Transactions on Graphics, 37(6):1-15, 2018. 7" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.268, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.147 + ], + "angle": 0, + "content": "[13] Lukas Hollein, Aljaž Božić, Michael Zollhöfer, and Matthias Nießner. 3dgs-lm: Faster gaussian-splatting optimization with levenberg-marquardt. arXiv preprint arXiv:2409.12892, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.149, + 0.906, + 0.204 + ], + "angle": 0, + "content": "[14] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42 (4):1-14, 2023. 1, 2, 3, 7, 8, 12, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.205, + 0.907, + 0.286 + ], + "angle": 0, + "content": "[15] Shakiba Kheradmand, Daniel Rebain, Gopal Sharma, Weiwei Sun, Yang-Che Tseng, Hossam Isack, Abhishek Kar, Andrea Tagliasacchi, and Kwang Moo Yi. 3d gaussian splatting as markov chain monte carlo. In Advances in Neural Information Processing Systems (NeurIPS), 2024. Spotlight Presentation. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.288, + 0.906, + 0.315 + ], + "angle": 0, + "content": "[16] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. CoRR, abs/1412.6980, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.317, + 0.906, + 0.37 + ], + "angle": 0, + "content": "[17] Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and temples: Benchmarking large-scale scene reconstruction. ACM Transactions on Graphics, 36(4):1-13, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.372, + 0.906, + 0.425 + ], + "angle": 0, + "content": "[18] Joo Chan Lee, Daniel Rho, Xiangyu Sun, Jong Hwan Ko, and Eunbyung Park. Compact 3d gaussian representation for radiance field. arXiv preprint arXiv:2311.13681, 2023. 2, 8, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.428, + 0.906, + 0.482 + ], + "angle": 0, + "content": "[19] Weikai Lin, Yu Feng, and Yuhao Zhu. Rtgs: Enabling real-time gaussian splatting on mobile devices using efficiency-guided pruning and foveated rendering. arXiv preprint arXiv:2407.00435, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.484, + 0.906, + 0.537 + ], + "angle": 0, + "content": "[20] Xiangrui Liu, Xinju Wu, Pingping Zhang, Shiqi Wang, Zhu Li, and Sam Kwong. Compgs: Efficient 3d scene representation via compressed gaussian splatting. arXiv preprint arXiv:2404.09458, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.54, + 0.908, + 0.609 + ], + "angle": 0, + "content": "[21] Tao Lu, Mulin Yu, Linning Xu, Yuanbo Xiangli, Limin Wang, Dahua Lin, and Bo Dai. Scaffold-gs: Structured 3d gaussians for view-adaptive rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20654-20664, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.61, + 0.907, + 0.692 + ], + "angle": 0, + "content": "[22] Saswat Subhajyoti Mallick, Rahul Goel, Bernhard Kerbl, Markus Steinberger, Francisco Vicente Carrasco, and Fernando De La Torre. Taming 3dgs: High-quality radiance fields with limited resources. In SIGGRAPH Asia 2024 Conference Papers, New York, NY, USA, 2024. Association for Computing Machinery. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.693, + 0.906, + 0.76 + ], + "angle": 0, + "content": "[23] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.762, + 0.906, + 0.816 + ], + "angle": 0, + "content": "[24] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph., 41(4):102:1-102:15, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.818, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[25] Michael Niemeyer, Fabian Manhardt, Marie-Julie Rakoto-saona, Michael Oechsle, Daniel Duckworth, Rama Gosula, Keisuke Tateno, John Bates, Dominik Kaeser, and Federico Tombari. Radsplat: Radiance field-informed gaussian splatting for robust real-time rendering with \\(900+\\) fps. arXiv preprint arXiv:2403.13806, 2024. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "21545" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.161 + ], + "angle": 0, + "content": "[26] Panagiotis Papantonakis, Georgios Kopanas, Bernhard Kerbl, Alexandre Lanvin, and George Drettakis. Reducing the memory footprint of 3d gaussian splatting. Proceedings of the ACM on Computer Graphics and Interactive Techniques, 7(1):1-17, 2024. 2, 8, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.164, + 0.482, + 0.232 + ], + "angle": 0, + "content": "[27] Lukas Radl, Michael Steiner, Mathias Parger, Alexander Weinrauch, Bernhard Kerbl, and Markus Steinberger. StopThePop: Sorted Gaussian Splatting for View-Consistent Real-time Rendering. ACM Transactions on Graphics, 4 (43), 2024. 2, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.234, + 0.482, + 0.289 + ], + "angle": 0, + "content": "[28] Kerui Ren, Lihan Jiang, Tao Lu, Mulin Yu, Linning Xu, Zhangkai Ni, and Bo Dai. Octree-gs: Towards consistent real-time rendering with lod-structured 3d gaussians. arXiv preprint arXiv:2403.17898, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.291, + 0.482, + 0.373 + ], + "angle": 0, + "content": "[29] Fernando Rivas-Manzaneque, Jorge Sierra-Acosta, Adrian Penate-Sanchez, Francesc Moreno-Noguer, and Angela Ribeiro. Nerflight: Fast and light neural radiance fields using a shared feature grid. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12417-12427, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.375, + 0.482, + 0.429 + ], + "angle": 0, + "content": "[30] Samuel Rota Bulò, Lorenzo Porzi, and Peter Kontschieder. Revising densification in gaussian splatting. In European Conference on Computer Vision, pages 347-362. Springer, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.431, + 0.482, + 0.498 + ], + "angle": 0, + "content": "[31] Evangelos Ververas, Rolandos Alexandros Potamias, Jifei Song, Jiankang Deng, and Stefanos Zafeiriou. Sags: structure-aware 3d gaussian splatting. In European Conference on Computer Vision, pages 221-238. Springer, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.502, + 0.482, + 0.557 + ], + "angle": 0, + "content": "[32] Meng Wei, Qianyi Wu, Jianmin Zheng, Hamid Rezatofighi, and Jianfei Cai. Normal-gs: 3d gaussian splatting with normal-involved rendering. In Advances in Neural Information Processing Systems (NeurIPS), 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.559, + 0.482, + 0.599 + ], + "angle": 0, + "content": "[33] Matthias Zwicker, Hanspeter Pfister, Jeroen Van Baar, and Markus Gross. Ewa splatting. IEEE Transactions on Visualization and Computer Graphics, 8(3):223-238, 2002. 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.599 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.956 + ], + "angle": 0, + "content": "21546" + } + ] +] \ No newline at end of file diff --git a/2025/Speedy-Splat_ Fast 3D Gaussian Splatting with Sparse Pixels and Sparse Primitives/08166209-021b-4f8f-8611-9f7790e8bc2b_origin.pdf b/2025/Speedy-Splat_ Fast 3D Gaussian Splatting with Sparse Pixels and Sparse Primitives/08166209-021b-4f8f-8611-9f7790e8bc2b_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..f2d8aacfc0f322f708e255f5024ef916408711a7 --- /dev/null +++ b/2025/Speedy-Splat_ Fast 3D Gaussian Splatting with Sparse Pixels and Sparse Primitives/08166209-021b-4f8f-8611-9f7790e8bc2b_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ac501f2d4b93bef70252b36eecc9704e17d01b3fc3c6f833eda0770b189fa345 +size 3469907 diff --git a/2025/Speedy-Splat_ Fast 3D Gaussian Splatting with Sparse Pixels and Sparse Primitives/full.md b/2025/Speedy-Splat_ Fast 3D Gaussian Splatting with Sparse Pixels and Sparse Primitives/full.md new file mode 100644 index 0000000000000000000000000000000000000000..75cec216a005b02c9c419de1bbf42ab7335136e5 --- /dev/null +++ b/2025/Speedy-Splat_ Fast 3D Gaussian Splatting with Sparse Pixels and Sparse Primitives/full.md @@ -0,0 +1,468 @@ +# Speedy-Splat: Fast 3D Gaussian Splating with Sparse Pixels and Sparse Primitives + +Alex Hanson + +Matthias Zwicker + +Allen Tu + +Geng Lin + +Vasu Singla + +Tom Goldstein + +University of Maryland, College Park + +![](images/80c8529a039791fdbf6f182b0191918f7f9585ea1fe19c1b57530b82780aa8d3.jpg) +GT + +![](images/22afe16fab5e67b7e86920db6ee04ce9f8c1b3df5ce2042922144023e376646d.jpg) +3DGS +2.6M Gaussians +25.39 PSNR +184.87 FPS +Figure 1. We reduce the number of Gaussians by over $90\%$ , only marginally decrease PSNR, and accelerate rendering speed by $6.2\times$ in the Tanks & Temples truck scene when compared to 3D Gaussian Splatting (3D-GS). Additionally, we speed up training time by $1.38\times$ . + +![](images/78fb40f2c812639d67118378d4bc2cff166da1aa572f4c0b2d66ca5de40ec0c8.jpg) +Ours +0.26M Gaussians +25.34 PSNR +1148.77 FPS + +# Abstract + +3D Gaussian Splatting (3D-GS) is a recent 3D scene reconstruction technique that enables real-time rendering of novel views by modeling scenes as parametric point clouds of differentiable 3D Gaussians. However, its rendering speed and model size still present bottlenecks, especially in resource-constrained settings. In this paper, we identify and address two key inefficiencies in 3D-GS to substantially improve rendering speed. These improvements also yield the ancillary benefits of reduced model size and training time. First, we optimize the rendering pipeline to precisely localize Gaussians in the scene, boosting rendering speed without altering visual fidelity. Second, we introduce a novel pruning technique and integrate it into the training pipeline, significantly reducing model size and training time while further raising rendering speed. Our Speedy-Splat approach combines these techniques to accelerate average rendering speed by a drastic $6.71 \times$ across scenes from the Mip-NeRF 360, Tanks & Temples, and Deep Blending datasets. Code and additional visualizations are available at https://speedysplat.github.io/. + +# 1. Introduction + +Fast rendering of photorealistic novel views has been a long-standing goal in computer vision and graphics. Neural Radiance Fields (NeRF) [23] and its variants have made sig- + +nificant strides in photorealistic 3D scene reconstruction by representing scenes as continuous neural volumetric models that encode scene density and color at spatial coordinates. However, despite recent efforts [4, 24, 29], fast rendering in NeRF remains challenging because the volumetric sampling used in ray-marching is computationally expensive. Recently, 3D Gaussian Splatting (3D-GS) [14] has emerged as a promising alternative that enables real-time rendering by modeling scenes as parametric point clouds of differentiable 3D Gaussians. Nevertheless, its rendering speed is still limited by high parameter counts and certain algorithmic inefficiencies. Efficient rendering is essential in applications such as virtual reality, networked systems, and multi-view streaming. Real-time rendering on resource-constrained edge devices, such as mobile phones, has yet to be achieved [19]. + +Although recent works on compressing 3D-GS models [6, 7, 11, 25] achieve some speed-ups by reducing the number of parameters, few approaches directly target rendering speed [8, 19]. In this paper, we specifically address this gap by demonstrating that the rendering speed of 3D-GS models can be drastically increased while maintaining competitive image quality. Additionally, we show that our methods reduce training time and substantially decrease model size. + +We begin by observing that the cost of 3D-GS rendering is proportional to both the number of Gaussians in the scene and the number of pixels processed per Gaussian. Our ap- + +proach optimizes both factors. First, we reduce the number of pixels that are processed for each Gaussian by efficiently and accurately localizing it in the rendered image. Second, we reduce the total number of Gaussians in the model through a novel approach that maintains rendering quality. + +3D-GS implements tile rendering to localize Gaussians in the image plane by assigning each Gaussian to the tiles that it intersects. We find the existing algorithm to be overly conservative; in Section 4.1.1, we address this by introducing our SnugBox algorithm to precisely localize Gaussians by computing a tight bounding box around their extent. Then, in Section 4.1.2, we extend SnugBox with our AccuTile algorithm to identify exact tile intersections. Both approaches are plug-and-play, lead to respective inference speed-ups of $1.82 \times$ and $1.99 \times$ on average, and do not change the 3D-GS renderings. + +To reduce the total number of Gaussians while preserving visual fidelity, we extend an existing pruning method [11] by reducing its memory requirement by $36\times$ and incorporating it into the 3D-GS training pipeline. During the densification stage of 3D-GS training, Gaussians are regularly replicated and pruned. In Section 4.2.2, we augment the densification stage with our Soft Pruning method to prune $80\%$ of Gaussians at regular intervals. After the densification stage, our Hard Pruning method, described in Section 4.2.3, prunes an additional $30\%$ of Gaussians at set intervals. + +Our Speedy-Splat approach integrates both techniques into the 3D-GS training pipeline. On average, rendering speed is accelerated by $6.71 \times$ , model size is reduced by $10.6 \times$ , and training speed is improved by $1.4 \times$ across all evaluated scenes while maintaining high image quality. + +In summary, we propose the following contributions: + +1. SnugBox: A precise algorithm for computing Gaussian-tile bounding box intersections. +2. AccuTile: An extension of SnugBox for computing exact Gaussian-tile intersections. +3. Soft Pruning: An augmentation for pruning Gaussians during densification. +4. Hard Pruning: An augmentation for pruning Gaussians post-densification. + +# 2. Related work + +The real-time rendering speed of 3D-GS [14] on desktop GPUs has inspired research focused on further accelerating both its training and inference in resource-constrained environments. In this section, we review related works that specifically target these performance improvements. + +# 2.1. Pruning + +A large portion of Gaussians in vanilla 3D-GS models are redundant [6, 7, 25], motivating a recent line of + +work focused on pruning Gaussians from 3D-GS models to boost rendering speed with minimal loss of visual fidelity [1, 2, 6, 7, 10, 11, 18, 19, 25]. Nearly all approaches assign a significance score to each Gaussian that is used to rank and prune them. Several works compute the aggregated ray contribution for each Gaussian across all input images [6, 7, 18, 25], while others combine opacity with additional information, such as gradients per Gaussian, to calculate their pruning criterion [1, 2, 20]. Papantonakis et al. [26] use resolution and scale-aware redundancy metrics, EAGLES [10] calculates the total transmittance per Gaussian, and Lin et al. [19] accelerate inference speed through an efficiency-aware strategy and foveated rendering. + +PUP 3D-GS [11] introduces a more principled approach by deriving a Hessian for each Gaussian that represents its sensitivity to the reconstruction error. While PUP 3D-GS achieves state-of-the-art post-hoc pruning results, computing its sensitivity score incurs considerable storage requirements that limit its viability for use during training. Our pruning approach directly extends PUP 3D-GS by improving its memory efficiency by $36\times$ . + +# 2.2. Other Methods + +In addition to pruning, several other strategies have been explored to enhance the rendering and training speed of 3D-GS. Mini-Splatting [7] modifies its densification strategies and adds a simplification stage to constrain the number of Gaussians. 3DGS-MCMC [15] models training dynamics as an MCMC process. Revisiting Densification [30] introduces an error-based densification strategy. 3DGS-LM [13] replaces Adam [16] with a tailored Levenberg-Marquardt optimizer [9]. Taming 3DGS [22] proposes a constructive optimization process that limits the number of Gaussians to a pre-defined threshold set by the user. DISTWAR [5] dives into the low-level implementation of GPU thread scheduling and optimizes atomic processing with a novel primitive. StopThePop [27] introduces a precise tile intersection method, but we find that our approach is notably faster in Appendix A.4. FlashGS [8] implements a tile intersection method similar to StopThePop. Several works reduce training time and memory requirements by enforcing geometric constraints [21, 28, 31, 32]. Most of these approaches are orthogonal to ours and can be applied alongside it. + +# 3. Background + +# 3.1. 3D Gaussian Splitting Overview + +3D Gaussian Splatting (3D-GS) [14] models scenes as parametric, point-based representations that use differentiable 3D Gaussians as primitives. Given a set of ground truth training images $\mathcal{I}_{gt} = \{\pmb{I}_i\in \mathbb{R}^{H\times W}\}_{i = 1}^K$ , the scene is initialized by using Structure from Motion (SfM) to produce a sparse point cloud that serves as the initial means for the + +Table 1. Average execution time (milliseconds) of each function across all scenes in Section 5.1. The operation in each row is applied cumulatively to all of the following rows. For each model, accurate measurements are collected by averaging execution times across three runs that each render the test set 20 times to reduce variance. The fastest and second fastest times are color coded. + +
MethodPreprocessInclusive SumDuplicate with KeysRadix SortIdentify Tile RangesRenderOverall
Baseline0.6650.0450.5701.5510.0824.4837.478
+SnugBox0.6560.0460.208 (2.738×)0.729 (2.126×)0.041 (1.980×)2.344 (1.913×)4.102 (1.823×)
+AccuTile0.6680.0460.221 (2.575×)0.612 (2.533×)0.035 (2.326×)2.062 (2.175×)3.748 (1.995×)
+Soft Pruning0.370 (1.798×)0.030 (1.494×)0.146 (3.906×)0.404 (3.843×)0.024 (3.422×)1.337 (3.354×)2.381 (3.141×)
+Hard Pruning0.091 (7.293×)0.016 (2.769×)0.090 (6.325×)0.215 (7.217×)0.013 (6.537×)0.619 (7.247×)1.114 (6.712×)
+ +3D Gaussians. The estimated camera poses $\mathcal{P}_{gt} = \{\phi_i\in$ $\mathbb{R}^{3\times 4}\}_{i = 1}^{K}$ are paired with their corresponding images and are used to optimize the scene. + +Each 3D Gaussian primitive $\mathcal{G}_i$ is parameterized by three geometry parameters - mean $\mu_i\in \mathbb{R}^3$ , scale $s_i\in \mathbb{R}^3$ , and rotation $r_i\in \mathbb{R}^4$ - and two color parameters - view-dependent spherical harmonics $h_i\in \mathbb{R}^{16\times 3}$ and opacity $\sigma_{i}\in \mathbb{R}$ . The set of all parameters can be described as: + +$$ +\mathcal {G} = \left\{\mathcal {G} _ {i} = \left\{\mu_ {i}, s _ {i}, r _ {i}, h _ {i}, \sigma_ {i} \right\} \right\} _ {i = 1} ^ {N}, \tag {1} +$$ + +where $N$ is the number of Gaussians in the model. + +Given camera pose $\phi$ , the scene is rendered by projecting all Gaussians to image space and applying alpha blending to each pixel. Models are optimized via stochastic gradient descent on image reconstruction losses: + +$$ +L (\mathcal {G} | \phi , I _ {g t}) = \| I _ {\mathcal {G}} (\phi) - I _ {g t} \| _ {1} + L _ {\mathrm {D - S S I M}} (I _ {\mathcal {G}} (\phi), I _ {g t}), \tag {2} +$$ + +where $I_{\mathcal{G}(\phi)}$ is the rendered image for pose $\phi$ + +During optimization, the scene is periodically densified by cloning and splitting uncertain Gaussians and pruned by removing large and transparent Gaussians. The opacities of the Gaussians are also periodically reset. + +# 3.2. 3D Gaussian Splitting Rendering Specifics + +3D-GS uses a tile-based rendering strategy that divides the rendered image into $16 \times 16$ pixel tiles. Each Gaussian is projected into image space, where its intersection with these tiles is computed. Then, these Gaussian-to-tile mappings are sorted to collect and order the Gaussians by depth for pixel-wise rendering. + +Rendering runtime is dominated by six key functions. Table 1 empirically analyzes the execution time of each function and highlights the improvements achieved by our methods. Descriptions of each function are provided in the following sections. + +# 3.2.1. Preprocessing + +The preprocess kernel is parallelized such that each thread processes a single Gaussian $\mathcal{G}_i$ . It computes a 2D projection of $\mathcal{G}_i$ to image space and obtains a count of tiles that it intersects. + +The mean $\mu_{i}$ is projected to image space using a viewing transform $W$ and a perspective projection, yielding 2D mean $\mu_{i_{2D}}$ and depth that are stored for later processing. The scale $s_i$ and rotation $r_i$ parameters are converted to the diagonal scale $S_{i}$ and rotation $R_{i}$ matrices. The 3D covariance is then defined as: + +$$ +\boldsymbol {\Sigma} _ {i _ {3 D}} = \boldsymbol {R} _ {i} \boldsymbol {S} _ {i} \boldsymbol {S} _ {i} ^ {T} \boldsymbol {R} _ {i} ^ {T}, \tag {3} +$$ + +which is projected via: + +$$ +\hat {\boldsymbol {\Sigma}} _ {i _ {3 D}} = \boldsymbol {J} \boldsymbol {W} \boldsymbol {\Sigma} _ {i _ {3 D}} \boldsymbol {W} ^ {T} \boldsymbol {J} ^ {T}, \tag {4} +$$ + +where $J$ is the Jacobian of the first order approximation of the perspective projection. Dropping the last row and column of $\hat{\Sigma}_{i_{3D}}$ gives $\Sigma_{i_{2D}}$ [14, 33]. The largest eigenvalue of $\Sigma_{i_{2D}}$ is used to compute the count of tiles intersected by this Gaussian $\mathcal{G}_i$ as shown in Figure 2a. $\Sigma_{i_{2D}}^{-1}$ is computed, along with a view-dependent color $c_{i}$ , derived from $W$ and $h_i$ . All three values are stored for later processing. + +# 3.2.2. Sorting + +After preprocessing, four functions process the Gaussians for pixel-wise rendering: + +- InclusiveSum: A CUDA primitive that computes the prefix sum over all Gaussian-tile counts to allocate key and value arrays for Gaussian-to-tile mapping. +- duplicateWithKeys: A Gaussian-parallel kernel that recomputes Gaussian-tile intersections to generate a key for each intersecting tile, consisting of the tile index and Gaussian depth. +- RadixSort: A CUDA primitive that sorts the key array, ordering Gaussian indices by tile and then depth. +- identifyTileRanges: A kernel parallelized across the key array to post-process keys before pixel rendering. + +# 3.2.3. Rendering + +The render kernel is parallelized across pixels. For each pixel $p$ , all Gaussians within its corresponding tile are loaded and processed in depth order as determined by RadixSort. An alpha value: + +$$ +\alpha_ {i} (p) = \sigma_ {i} g _ {i} (p) \tag {5} +$$ + +is computed for each Gaussian $\mathcal{G}_i$ , where $g_i$ is the value of the projected 2D Gaussian at pixel $p$ : + +$$ +g _ {i} = e ^ {q}, q = - \frac {1}{2} (p - \mu_ {i _ {2 D}}) \boldsymbol {\Sigma} _ {\boldsymbol {i} _ {2 D}} ^ {- 1} (p - \mu_ {i _ {2 D}}) ^ {T}. \tag {6} +$$ + +If $\alpha_{i} > \frac{1}{255}$ , then the Gaussian is included in the alpha compositing of the pixel color $C$ , given by: + +$$ +C (p) = \sum_ {i \in \mathcal {N}} c _ {i} \alpha_ {i} (p) \prod_ {j = 1} ^ {i - 1} (1 - \alpha_ {j} (p)). \tag {7} +$$ + +# 4. Methods + +Our Speedy-Splat methods are motivated by two key insights into inefficiencies within the 3D-GS rendering pipeline. First, Gaussian Splatting grossly overestimates the extent of Gaussians in the image. Second, as demonstrated by [11] and other recent pruning works, 3D-GS models are heavily overparameterized. + +# 4.1. Precise Tile Intersect + +Gaussian Splitting identifies tiles intersected by Gaussian $\mathcal{G}_i$ by calculating the maximum eigenvalue $\lambda_{\mathrm{max}}$ of its projected 2D covariance $\Sigma_{i_2D}$ , then selecting all tiles that intersect the square inscribing the circle defined by center $\mu_{i_{2D}}$ and radius: + +$$ +r = \left\lceil 3 \sqrt {\lambda_ {\max }} \right\rceil . \tag {8} +$$ + +This approach neglects opacity $\sigma_{i}$ in its calculation and generally overestimates the Gaussian extent, as illustrated by Figure 2a. The actual extent of Gaussian $\mathcal{G}_i$ , shown in Figure 2b, is given by the threshold placed on its alpha value $\alpha_{i}$ . Specifically, $\mathcal{G}_i$ does not contribute to the rendering of pixel $p$ if $\alpha_{i} < \frac{1}{255}$ for $p$ . By applying the actual extent in tile intersection calculations, we arrive at a far more concise set of intersected tiles. + +We now show the derivation of this extent. The furthest pixel extent of Gaussian $\mathcal{G}_i$ can be determined by directly substituting this threshold into Equation 5. Rearranging terms gives: + +$$ +2 \log (2 5 5 \sigma_ {i}) = (p - \mu_ {i _ {2 D}}) \boldsymbol {\Sigma} _ {\boldsymbol {i _ {2 D}}} ^ {- 1} (p - \mu_ {i _ {2 D}}) ^ {T}. \qquad (9) +$$ + +We can rewrite: + +$$ +p = \binom {p _ {x}} {p _ {y}}, \mu_ {i _ {2 D}} = \binom {\mu_ {x}} {\mu_ {y}}, \boldsymbol {\Sigma} _ {\boldsymbol {i} _ {2 D}} ^ {- 1} = \left( \begin{array}{c c} a & b \\ b & c \end{array} \right). \tag {10} +$$ + +Specifying threshold $t$ and centered coordinates $x_{d}$ and $y_{d}$ : + +$$ +t = 2 \log (2 5 5 \sigma_ {i}), \tag {11} +$$ + +$$ +x _ {d} = p _ {x} - \mu_ {x}, \tag {12} +$$ + +$$ +y _ {d} = p _ {y} - \mu_ {y}, \tag {13} +$$ + +gives the pixel extent with coordinates $x_{d}$ and $y_{d}$ that satisfy the ellipse equation: + +$$ +t = a x _ {d} ^ {2} + 2 b x _ {d} y _ {d} + c y _ {d} ^ {2}. \tag {14} +$$ + +Our approach uses this pixel extent to reduce the number of tiles contained in the Gaussian-to-tile mappings for each Gaussian. We propose two methods for computing precise tile intersections. First, our SnugBox algorithm produces a tight bounding box around each Gaussian. Second, our AccuTile algorithm extends it to identify the exact set of tiles intersected by the Gaussian. + +# 4.1.1. SnugBox + +Our SnugBox method uses this elliptical extent to compute an axis-aligned bounding box that more precisely identifies tiles intersected by Gaussian $\mathcal{G}_i$ . To derive this bounding box, we rearrange Equation 14 to solve for $y_d$ : + +$$ +y _ {d} = \frac {- b x _ {d} \pm \sqrt {\left(b ^ {2} - a c\right) x _ {d} ^ {2} + t c}}{c}. \tag {15} +$$ + +To find the $y$ -coordinate bounding box edges $y_{\mathrm{min}}$ and $y_{\mathrm{max}}$ , we identify the values of $x_d$ where $\partial y_d / \partial x_d = 0$ . We refer to these $x_d$ values as $x_{d_{\mathrm{arg}s}}$ to specify that they are the $\arg \min y_d$ and $\arg \max y_d$ values. Differentiating Equation 15 and solving for $x_{d_{\mathrm{arg}s}}$ gives: + +$$ +x _ {d _ {\text {a r g s}}} = \pm \sqrt {\frac {- b ^ {2} t}{\left(b ^ {2} - a c\right) a}}. \tag {16} +$$ + +Substituting $x_{d_{\text{args}}}$ into Equation 15 and adding $\mu_y$ gives $y_{\min}$ and $y_{\max}$ . Due to the symmetry of Equation 14, we can find the $x$ -coordinate bounding box edges $x_{\min}$ and $x_{\max}$ by swapping $y_d$ and $x_d$ and constants $a$ and $c$ to rewrite Equations 15 and 16 in terms of $x_d$ and $y_{d_{\text{args}}}$ . + +After the bounding box edges are identified, our method follows 3D-GS and converts these edges to tile indices by dividing by tile size, rounding, and clipping to the image boundary. As depicted in Figure 2, Snugbox can produce a significantly tighter bounding box than 3D-GS. Meanwhile, its computational overhead is small as it performs a constant number of operations and is only called twice in the rendering pipeline – once to count the Gaussians intersecting tiles in preprocess and once to populate the Gaussian-to-tile arrays in duplicateWithkeys. Table 1 reports that SnugBox improves the efficiency of all downstream functions and produces an average overall speed-up of $1.82 \times$ . + +# 4.1.2. AccuTile + +Our AccuTile method, outlined in Algorithm 1, extends SnugBox to identify the exact tiles intersected by the Gaussian. It takes as input the tight bounding box produced by Snugbox and its rectangular tile extent - depicted as the + +![](images/498c1943b015373d3a9f7aa4619224253d212be3fdf54a4eeef91620a0d6558b.jpg) +(a) 3D Gaussian Splitting + +![](images/54a1fbdd743500bbb5da29510d8a0eec048bfc1ea78239c3a8b7a1127ab99138.jpg) +(b) SnugBox +Figure 2. Gaussian tile allocation by method. (a) 3D Gaussian Splating allocates a Gaussian to a tile when that tile intersects the square inscribing the circle with radius $\lceil 3\sqrt{\lambda_{\max}}\rceil$ defined in Equation 8. (b) Our SnugBox method allocates a Gaussian to a tile when that tile intersects the tight bounding box defined by the axis-aligned minima and maxima of the ellipse given by Equation 14. (c) Our AccuTile method allocates a Gaussian to a tile only if that tile intersects the ellipse via Algorithm 1, which computes the minimum and maximum tiles by iterating over the shorter side of the rectangular tile extent given by SnugBox. In this example, our AccuTile algorithm iterates over the tile rows; the only points that are processed are $x_{min}$ , $x_{max}$ , A, B, C, and D. + +![](images/13cf56fb2296d2a06c9f0264d39d6b0ac0948a8d93ab199d86136919acaae2b4.jpg) +(c) AccuTile + +blue box and yellow tiles, respectively, in Figure 2b. Depending on which dimension is smaller, AccuTile then processes either the rows or columns of this tile extent to determine the exact tiles that intersect the Gaussian. Specifically, we identify the minimum and maximum extent of the ellipse within a given row or column, then convert those points to the corresponding touched tiles. All tiles between the minimum and maximum tiles intersect the ellipse. + +AccuTile's key insight is that calculating the minimum and maximum extent of the ellipse within each row or column requires computing only two points per iteration. The only possible inflection points of the elliptical curve are the bounding box minimum and maximum points identified by Snugbox. If one or both of these points lie within the tile row or column, then they represent the minimum or maximum extent of the ellipse there. If neither point is within the row or column, then the points along that boundary side are monotonically decreasing or increasing – the minimum or maximum point must lie on one of the boundary lines of that row or column. Since the boundary of the last row or column is shared with the next one, we only need to compute the intersection of the ellipse with the next boundary line in each iteration. Thus, our AccuTile algorithm counts tiles in time proportional to the shorter side of the tile extent and processes tiles in time proportional to the tile count. + +Figure 2 illustrates how Accutile restricts the tight bounding box produced by SnugBox to the exact tiles that the Gaussian touches. The tile extent rows from Snugbox, shown in Figure 2b, are processed starting from the bottom. The lower tile row boundary falls below the bounding box, so no initial intersection is computed. The upper boundary lies below the top of the bounding box, so intersection + +points $A$ and $B$ are calculated using Equation 15. $x_{\mathrm{min}}$ is assigned as this row's minimum ellipse extent $e_{\mathrm{min}}$ because it lies within it, and $B$ is assigned as its maximum ellipse extent $e_{\mathrm{max}}$ because it is the maximal point. Consequently, this row's tile extent is from the tile containing $x_{\mathrm{min}}$ to the tile containing $B$ . For the next row, we keep points $A$ and $B$ and compute upper boundary points $C$ and $D$ . $A$ and $x_{\mathrm{max}}$ are assigned as the row's $e_{\mathrm{min}}$ and $e_{\mathrm{max}}$ , and the tiles containing them are its tile extent. Finally, for the last row, $C$ and $D$ are kept; no additional points are computed because the row's upper boundary is above the bounding box. Thus, the tile extent of this row is between $C$ and $D$ . + +The number of tiles is further reduced from Figure 2b to Figure 2c, much less than the original 3D-GS method in Figure 2a. In Table 1, we report that AccuTile further accelerates all downstream functions, culminating in an average overall speed-up of $1.99 \times$ . + +# 4.2. Efficient Pruning + +PUP 3D-GS [11] is a pruning method that quantifies the sensitivity of each Gaussian to training views, removing a set percentage with the lowest sensitivities. Sensitivity is computed by approximating the Hessian of the $L_{2}$ loss: + +$$ +H = \nabla_ {\mathcal {G}} ^ {2} L _ {2} = \sum_ {\phi \in \mathcal {P} _ {g t}} \nabla_ {\mathcal {G}} I _ {\mathcal {G}} (\phi) \nabla_ {\mathcal {G}} I _ {\mathcal {G}} (\phi) ^ {T}, \tag {17} +$$ + +where $\nabla_{\mathcal{G}}I_{\mathcal{G}}(\phi)$ is the gradient over all Gaussian parameters on the rendered image $I_{\mathcal{G}}$ for pose $\phi$ . $H$ is shown to be exact when the $L_{1}$ residual error vanishes [11]. + +A per-Gaussian sensitivity can be derived by splitting $H$ into the block diagonal elements that only capture inter + +Algorithm 1 The AccuTile Algorithm. For simplicity, the algorithm outlined here is applied to the rows of the Snug-Box tile extent bounding box, matching the example in Figure 2c. In practice, it is applied along the smaller side of the tile extent. The subscripts $t$ , $b$ , $l$ , and $r$ represent the top, bottom, left, and right sides, respectively. A proof of correctness sketch is presented in Appendix A.1. + +Require: Ellipse E $\triangleright$ Eq. 14 +Require: SnugBox Bounding Box B +Require: SnugBox Tile Extent Rectangle R +Initialize: Tile count $C\gets 0$ +line $\mathrm{min}\gets \mathbf{R}_b$ +if line $\mathrm{min}\geq \mathbf{B}_b$ then + $\mathbf{i}_{\mathrm{min}}\gets$ Intersections( line $\mathrm{min}$ ,E) $\triangleright$ Eq. 15 +end if +for row r in R do +line $\mathrm{max}\gets \mathrm{r}_t$ +if line $\mathrm{max}\leq \mathbf{B}_t$ then + $\mathbf{i}_{\mathrm{max}}\gets$ Intersections( line $\mathrm{max}$ ,E) $\triangleright$ Eq. 15 +end if +e $\mathrm{min}\gets \mathbf{B}_l$ if Bl in r else min(i $\mathrm{min}$ ,i max) +e $\mathrm{max}\gets$ Br if Br in r else max(i $\mathrm{min}$ ,i max) +tile $\mathrm{min}$ ,tile $\mathrm{max}\gets$ Convert(e $\mathrm{min}$ ,e max) +C $\leftarrow$ C + (tile $\mathrm{max}$ -tile $\mathrm{min}$ ) +Process( tile $\mathrm{min}$ ,tile $\mathrm{max}$ ) + $\mathbf{i}_{\mathrm{min}}\gets \mathbf{i}_{\mathrm{max}}$ +end for +return C + +Gaussian parameter relationships: + +$$ +H _ {i} = \sum_ {\phi \in \mathcal {P} _ {g t}} \nabla_ {\mathcal {G} _ {i}} I _ {\mathcal {G}} (\phi) \nabla_ {\mathcal {G} _ {i}} I _ {\mathcal {G}} (\phi) ^ {T}, \tag {18} +$$ + +where $\nabla_{\mathcal{G}_i}$ is the gradient with respect to only $\mathcal{G}_i$ . This measures the sensitivity of the $L_2$ loss with respect to Gaussian $\mathcal{G}_i$ , assuming all other Gaussians are held constant. + +$H_{i}$ is again approximated by only using the six mean $\mu_{i}$ and scale $s_i$ parameters to specifically capture geometric sensitivity. The log determinant is taken to provide a representative scalar score $U_{i}$ : + +$$ +U _ {i} = \log | \nabla_ {\mu_ {i}, s _ {i}} I _ {\mathcal {G}} \nabla_ {\mu_ {i}, s _ {i}} I _ {\mathcal {G}} ^ {T} |. \tag {19} +$$ + +Using this score, up to $90\%$ of Gaussians can be robustly pruned from the model while retaining high visual quality. + +Although PUP 3D-GS touts high compression ratios and rendering speeds, we identify two key drawbacks in its formulation. First, computing the Hessian requires storage + +proportional to $N \times 36$ , where $N$ is the number of Gaussians. In comparison, the 3D-GS model has a memory footprint proportional to $N \times 59$ because it stores 59 parameters per Gaussian. While this score is effective for post-hoc pruning, using it during training is impractical. + +Second, computing the Hessian requires the pixel-wise gradients of $\mu$ and $s$ . Since these are 3D parameters of the Gaussian primitives, obtaining their gradients requires back-propagating through the render kernel parallelized per pixel, then back-propagating each Gaussian contributing to that pixel in its thread. This breaks the efficient flow of gradients in 3D-GS, where the per-pixel gradients from the render kernel are parallelized and aggregated to the 2D $\mu_{2D}$ and $\Sigma_{2D}$ parameters, which are then parallelized across Gaussians to compute gradients for $\mu$ and $s$ . + +Our approach builds on PUP 3D-GS by introducing an efficient pruning score that we incorporate into the 3D-GS training pipeline. We also define two distinct pruning modalities: Soft Pruning, which takes place during densification in the first 15000 iterations, and Hard Pruning, which is applied after densification is completed after iteration 15000. + +# 4.2.1. Efficient Pruning Score + +Our insight is that both drawbacks can be alleviated by reparameterizing the Hessian. Concretely, we express the influence of all spatial parameters of Gaussian $\mathcal{G}_i$ by computing the Hessian approximation with respect to the 2D projected value of $\mathcal{G}_i$ at pixel $p$ , given by $g_i(p)$ in Equation 6. In doing so, the pruning score $U_i$ from Equation 19 can be expressed as: + +$$ +\tilde {U} _ {i} = \log | \nabla_ {g _ {i}} I _ {\mathcal {G}} \nabla_ {g _ {i}} I _ {\mathcal {G}} ^ {T} |. \tag {20} +$$ + +Since $g_{i}$ is a scalar and log is monotonically increasing, we can rewrite this score as: + +$$ +\tilde {U} _ {i} = \left(\nabla_ {g _ {i}} I _ {\mathcal {G}}\right) ^ {2}. \tag {21} +$$ + +Gradient $\nabla_{g_i}I_{\mathcal{G}}$ is already computed in the backward pass of render and can be efficiently squared and aggregated across all pixels. Moreover, the maximum space requirement for this score is proportional to the number of Gaussians $N$ , reducing the storage requirement by $36\times$ and allowing this score to be used during training. + +# 4.2.2. Soft Pruning + +To maintain a robust Hessian approximation, we observe that the $L_{1}$ loss becomes quite small by iteration 6000 and remains low except after an opacity reset is performed. As such, we augment the densification pipeline to include our Soft Pruning method, where the model is pruned immediately before the three opacity resets at 6000, 9000, and 12000 iterations. Surprisingly, we find that we can set extremely high Soft Pruning ratios – in our experiments, visual fidelity is preserved at $80\%$ pruning. + +![](images/ffa852edb8b496f4992351b54752d23519b28289c3ede99ec305b29d3979ba23.jpg) +GT + +![](images/ab1bba5ffc9241fd393b849b78777938923af29863d8da268f51a99e0ac94bff.jpg) +3D-GS +2.5M Gaussians +172.15 FPS + +![](images/7b86d26216953725c49b5f1abe752c107466b4e116a939948888647730f80a15.jpg) +PUP +0.25M Gaussians + +![](images/6d9b80c08e645f24d151969b745cc25e543f78c066e072c94ad7237210beb122.jpg) +0.22M Gaussians +1277.20 FPS + +![](images/cbe29eccb3828a4900ea58195c9e98fefb2d7b2212762b7e7fe9e21bc15f88d7.jpg) + +![](images/a69a2b1aad2b45f92c9a61c3e42068077ae5f9fbe80c3678fa2a3821d4f83abb.jpg) +4.5M Gaussians +141.47 FPS + +![](images/ad195892fca34ea72da077a4a4f19021307eedbdbd2ef40ff1c190e191dec090.jpg) +0.45M Gaussians +261.67 FPS + +![](images/c71c96ee3ad9661c7f598d20039af9416d6e007f20505ed785a468ca4f3cb411.jpg) +0.46M Gaussians +723.79 FPS + +![](images/74135626c2d30c63985b5840546ec5c0b9e0e6365fe9b37eba0069f9ec92a466.jpg) +Figure 3. Visual comparison of 3D-GS, PUP 3D-GS, and our method. Notice that, while reaching similar compression ratios to PUP 3D-GS, our Speedy-Splat method delivers vastly faster rendering speeds. Top: playroom from the Deep Blending dataset. Middle: bicycle from the Mip-NeRF 360 dataset. Bottom: drjohnson from the Deep Blending dataset. + +![](images/7112c46269c14c0a67c0488d1b942f3961a5977bd336a434a2abe37d0756173f.jpg) +3.4M Gaussians +125.97 FPS + +![](images/4a705b7976af512de25292283ff91eacf5f38e44ef4ca3a5da8b1f75a549a2c3.jpg) +0.34M Gaussians +413.30 FPS + +![](images/6b4a83cf5c4ed81986b2c0dd4ed0c0bd3617f27a78f1d30868f7659030603af2.jpg) +0.32M Gaussians +1121.97 FPS + +# 4.2.3. Hard Pruning + +We also observe that the model's performance after densification closely matches that of the fully-trained model. The iterations after the densification stage essentially finetune the model and can be used to further "refine" it after pruning, similar to PUP 3D-GS [11] and LightGaussian [6]. In practice, our Hard Pruning method prunes the model by a constant ratio every 3000 iterations starting at iteration 15000. We Hard Prune $30\%$ of Gaussians in each interval, which, when paired with Soft Pruning, empirically reduces the total number of Gaussians across scenes by $10.6\times$ . + +# 5. Experiments + +# 5.1. Datasets + +Our evaluation uses the same set of challenging real-world scenes as 3D-GS [14]. This includes nine Mip-Nerf 360 scenes [3] – four indoor and five outdoor – that each feature a complex central object or area with a detailed background. We also include the outdoor train and truck scenes from the Tanks & Temples dataset [17] and the indoor drjohnson and playroom scenes from the Deep Blending dataset [12]. For consistency across experiments, we use the COLMAP pose estimates and sparse point clouds provided by the dataset authors. + +Table 2. Average Gaussian count, FPS, and training time across all scenes in Section 5.1. Ratios for model size compression, rendering speed-up, and training speed-up are reported in (parentheses). The operation in each row is applied cumulatively to all of the following rows. The best and second best value for each metric are color coded. + +
Method# Gaussians ↓FPS ↑Training Time ↓
Baseline2.93M13423.2
+SnugBox2.97M244 (1.82×)21.2 (1.09×)
+AccuTile2.97M267 (1.99×)21.0 (1.10×)
+Soft Pruning1.64M (1.79×)420 (3.14×)17.5 (1.32×)
+Hard Pruning0.28M (10.6×)898 (6.71×)15.7 (1.47×)
+ +# 5.2. Implementation Details + +Our code builds on the differentiable renderer provided by 3D-GS [14] and modifies the Python training pipeline for pruning schedules and execution. To ensure consistent and precise timing, all times in Table 1 and FPS values in Tables 2 and 3 are measured using CUDA events at the start and end of the forward rendering procedure. All experiments are conducted on an Nvidia RTXA5000 GPU, and the reported Speedy-Splat results represent the average metrics across three independent runs for each scene. + +Table 3. Average reported metrics for each pruning method across all scenes in the Mip-NeRF 360 dataset. The Comp column reports model size compression in terms of Gaussian count, FPS reports rendering speed-up, and Train reports training time speed-up, all with respect to the baseline 3D-GS model. PSNR, SSIM, and LPIPS are also recorded. For a fair comparison, we report the published results of each method and use '-' to denote missing metrics. The best and second best value for each metric are color coded; lossless methods are underlined. Results for the Tanks & Temples and Deep Blending datasets are reported in Appendix A.3. + +
MethodComp ↑FPS ↑Train ↑PSNR ↑SSIM ↑LPIPS ↓
3D-GS [14]1.00×1.00×1.00×27.550.8140.222
Trimming [2]4.00×--27.130.7980.248
Compact [18]2.28×1.07×0.73×27.080.7980.247
EAGLES [10]3.68×1.51×1.37×26.940.8000.250
Reducing [26]2.33×1.60×1.23×27.100.8090.226
Light [6]2.94×1.76×-27.280.8050.243
ELMGS [1]5.00×2.69×-27.000.7790.286
PUP [11]8.65×2.55×-26.830.7920.268
Mini-Splat [7]6.84×3.20×1.26×27.340.8220.217
+SnugBox0.99×1.81×1.08×27.550.8140.221
+AccuTile0.99×1.99×1.10×27.570.8140.221
+Soft Pruning1.79×3.14×1.30×27.320.8070.246
+Hard Pruning10.6×6.51×1.45×26.940.7820.296
+ +# 5.3. Results + +# 5.3.1. Additive method performance + +The efficacy of Speedy-Splat is demonstrated by Table 1, where we record the average execution times of each function across all scenes when additively applying our methods. SnugBox and AccuTile each introduce minimal additional computation time to preprocess and InclusiveSum. However, limiting the number of tiles touched accelerates all downstream functions, culminating in an overall speed-up of $1.82 \times$ by SnugBox that is raised to $1.99 \times$ by AccuTile. Applying soft pruning reduces the runtime of all functions by reducing the number of Gaussians, leading to a $3.14 \times$ overall speed-up. Finally, performing hard pruning improves overall speed by a whopping $6.71 \times$ over the baseline 3D-GS model. + +# 5.3.2. Overall Performance + +In Figures 1 and 3, we report qualitative results on two outdoor and two indoor scenes from all three datasets. The magnified regions highlight that our method preserves the fine details in the baseline 3D-GS scene and closely models the ground truth view. Despite touting similar compression ratios and rendering nearly identical images, Speedy-Splat achieves over double the FPS of PUP 3D-GS. + +Table 3 compares our methods against other methods that reduce the number of Gaussians and increase inference speed using the mean of each metric across all scenes in the Mip-NeRF 360 dataset. The underlined methods are "lossless", meaning that they avoid degrading visual fidelity at all. SnugBox and AccuTile, our lossless methods, improve + +Table 4. Average metrics across all scenes in Section 5.1 when using the Speedy-Splat and PUP 3D-GS pruning scores to prune $88.44\%$ of Gaussians using the PUP 3D-GS pipeline. Two rounds of prune-refine are performed on each baseline 3D-GS model, pruning $66\%$ of Gaussians and then fine-tuning for 5,000 iterations in each one. The best value for each metric is color coded. + +
Method# Gaussians ↓PSNR ↑SSIM ↑LPIPS ↓FPS ↑
Baseline2.92M27.15030.82960.2238107.53
PUP [11]0.34M26.21360.80440.2731378.57
Ours0.34M26.86580.80220.2840345.52
+ +rendering and training speed while leaving image quality metrics essentially unchanged or slightly better. Our full pipeline, labeled as “+Hard Pruning” boasts the highest compression ratios, rendering speeds, and training speed-ups across all datasets. Furthermore, its image quality metrics are competitive with the other methods. + +# 5.4. Pruning Score Comparison + +Although the primary focus of our work is rendering speed, we find that our efficient pruning score, described in Section 4.2, also performs well when applied in other compression pipelines. In Table 4, we ablate our efficient pruning score with the PUP 3D-GS sensitivity score across all scenes in their post-hoc pruning pipeline. Notably, Speedy-Splat's efficient pruning score outperforms PUP 3D-GS on PSNR and is competitive across the other metrics. + +# 6. Limitations + +A limitation of Speedy-Splat is that it produces slightly lower image quality than 3D-GS. However, this degradation is expected at high compression ratios and is also observed in comparable techniques. Additionally, a direct comparison of our efficient pruning score to the PUP 3D-GS pruning score illuminates a slight, yet noticeable, gap in performance. Future work could explore the possibility of another efficient pruning score that delivers higher performance. + +# 7. Conclusion + +In this work, we present Speedy-Splat: a new 3D-GS technique that accurately localizes Gaussians during rendering and significantly improves inference speed, model size, and training time. Enhanced localization is achieved by our SnugBox and AccuTile methods, while model size reduction is accomplished by our Soft and Hard Pruning approaches. Together, our Speedy-Splat methods accelerate rendering speed by an average of $6.71 \times$ , reduce model size by $10.6 \times$ , and improve training time by $1.47 \times$ across all scenes from the Mip-NeRF 360, Tanks & Temples, and Deep Blending datasets. + +# 8. Acknowledgements + +This work was made possible by the IARPA WRIVA Program, the ONR MURI program, and DAPRA TIAMAT. Commercial support was provided by Capital One Bank, the Amazon Research Award program, and Open Philanthropy. Further support was provided by the National Science Foundation (IIS-2212182), and by the NSF TRAILS Institute (2229885). Zwicker was additionally supported by the National Science Foundation (IIS-2126407). + +# References + +[1] Muhammad Salman Ali, Sung-Ho Bae, and Enzo Tartaglione. Elmgs: Enhancing memory and computation scalability through compression for 3d gaussian splatting. arXiv preprint arXiv:2410.23213, 2024. 2, 8, 12 +[2] Muhammad Salman Ali, Maryam Qamar, Sung-Ho Bae, and Enzo Tartaglione. Trimming the fat: Efficient compression of 3d gaussian splats through pruning. arXiv preprint arXiv:2406.18214, 2024. 2, 8, 12 +[3] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5470–5479, 2022. 7 +[4] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In European Conference on Computer Vision (ECCV), 2022. 1 +[5] Sankeerth Durvasula, Adrian Zhao, Fan Chen, Ruofan Liang, Pawan Kumar Sanjaya, and Nandita Vijaykumar. Distwar: Fast differentiable rendering on raster-based rendering pipelines. arXiv preprint arXiv:2401.05345, 2023. 2 +[6] Zhiwen Fan, Kevin Wang, Kairun Wen, Zehao Zhu, Dejia Xu, and Zhangyang Wang. Lightgaussian: Unbounded 3d gaussian compression with 15x reduction and $200+$ fps. arXiv preprint arXiv:2311.17245, 2023. 1, 2, 7, 8, 12 +[7] Guangchi Fang and Bing Wang. Mini-splatting: Representing scenes with a constrained number of gaussians. arXiv preprint arXiv:2403.14166, 2024. 1, 2, 8, 12 +[8] Guofeng Feng, Siyan Chen, Rong Fu, Zimu Liao, Yi Wang, Tao Liu, Zhilin Pei, Hengjie Li, Xingcheng Zhang, and Bo Dai. Flashgs: Efficient 3d gaussian splatting for large-scale and high-resolution rendering. arXiv preprint arXiv:2408.07967, 2024.1, 2 +[9] Henri P. Gavin. The levenberg-marquardt method for nonlinear least squares curve-fitting problems c ©. 2013. 2 +[10] Sharath Girish, Kamal Gupta, and Abhinav Shrivastava. Eagles: Efficient accelerated 3d gaussians with lightweight encodings. arXiv preprint arXiv:2312.04564, 2023. 2, 8, 12 +[11] Alex Hanson, Allen Tu, Vasu Singla, Mayuka Jayawardhana, Matthias Zwicker, and Tom Goldstein. Pup 3d-gs: Principled uncertainty pruning for 3d gaussian splattering. arXiv preprint arXiv:2406.10219, 2024. 1, 2, 4, 5, 7, 8, 12 +[12] Peter Hedman, Julien Philip, True Price, Jan-Michael Frahm, George Drettakis, and Gabriel Brostow. Deep blending for free-viewpoint image-based rendering. ACM Transactions on Graphics, 37(6):1-15, 2018. 7 + +[13] Lukas Hollein, Aljaž Božić, Michael Zollhöfer, and Matthias Nießner. 3dgs-lm: Faster gaussian-splatting optimization with levenberg-marquardt. arXiv preprint arXiv:2409.12892, 2024. 2 +[14] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42 (4):1-14, 2023. 1, 2, 3, 7, 8, 12, 13 +[15] Shakiba Kheradmand, Daniel Rebain, Gopal Sharma, Weiwei Sun, Yang-Che Tseng, Hossam Isack, Abhishek Kar, Andrea Tagliasacchi, and Kwang Moo Yi. 3d gaussian splatting as markov chain monte carlo. In Advances in Neural Information Processing Systems (NeurIPS), 2024. Spotlight Presentation. 2 +[16] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. CoRR, abs/1412.6980, 2014. 2 +[17] Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and temples: Benchmarking large-scale scene reconstruction. ACM Transactions on Graphics, 36(4):1-13, 2017. 7 +[18] Joo Chan Lee, Daniel Rho, Xiangyu Sun, Jong Hwan Ko, and Eunbyung Park. Compact 3d gaussian representation for radiance field. arXiv preprint arXiv:2311.13681, 2023. 2, 8, 12 +[19] Weikai Lin, Yu Feng, and Yuhao Zhu. Rtgs: Enabling real-time gaussian splatting on mobile devices using efficiency-guided pruning and foveated rendering. arXiv preprint arXiv:2407.00435, 2024. 1, 2 +[20] Xiangrui Liu, Xinju Wu, Pingping Zhang, Shiqi Wang, Zhu Li, and Sam Kwong. Compgs: Efficient 3d scene representation via compressed gaussian splatting. arXiv preprint arXiv:2404.09458, 2024. 2 +[21] Tao Lu, Mulin Yu, Linning Xu, Yuanbo Xiangli, Limin Wang, Dahua Lin, and Bo Dai. Scaffold-gs: Structured 3d gaussians for view-adaptive rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20654-20664, 2024. 2 +[22] Saswat Subhajyoti Mallick, Rahul Goel, Bernhard Kerbl, Markus Steinberger, Francisco Vicente Carrasco, and Fernando De La Torre. Taming 3dgs: High-quality radiance fields with limited resources. In SIGGRAPH Asia 2024 Conference Papers, New York, NY, USA, 2024. Association for Computing Machinery. 2 +[23] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 1 +[24] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph., 41(4):102:1-102:15, 2022. 1 +[25] Michael Niemeyer, Fabian Manhardt, Marie-Julie Rakoto-saona, Michael Oechsle, Daniel Duckworth, Rama Gosula, Keisuke Tateno, John Bates, Dominik Kaeser, and Federico Tombari. Radsplat: Radiance field-informed gaussian splatting for robust real-time rendering with $900+$ fps. arXiv preprint arXiv:2403.13806, 2024. 1, 2 + +[26] Panagiotis Papantonakis, Georgios Kopanas, Bernhard Kerbl, Alexandre Lanvin, and George Drettakis. Reducing the memory footprint of 3d gaussian splatting. Proceedings of the ACM on Computer Graphics and Interactive Techniques, 7(1):1-17, 2024. 2, 8, 12 +[27] Lukas Radl, Michael Steiner, Mathias Parger, Alexander Weinrauch, Bernhard Kerbl, and Markus Steinberger. StopThePop: Sorted Gaussian Splatting for View-Consistent Real-time Rendering. ACM Transactions on Graphics, 4 (43), 2024. 2, 13 +[28] Kerui Ren, Lihan Jiang, Tao Lu, Mulin Yu, Linning Xu, Zhangkai Ni, and Bo Dai. Octree-gs: Towards consistent real-time rendering with lod-structured 3d gaussians. arXiv preprint arXiv:2403.17898, 2024. 2 +[29] Fernando Rivas-Manzaneque, Jorge Sierra-Acosta, Adrian Penate-Sanchez, Francesc Moreno-Noguer, and Angela Ribeiro. Nerflight: Fast and light neural radiance fields using a shared feature grid. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12417-12427, 2023. 1 +[30] Samuel Rota Bulò, Lorenzo Porzi, and Peter Kontschieder. Revising densification in gaussian splatting. In European Conference on Computer Vision, pages 347-362. Springer, 2024. 2 +[31] Evangelos Ververas, Rolandos Alexandros Potamias, Jifei Song, Jiankang Deng, and Stefanos Zafeiriou. Sags: structure-aware 3d gaussian splatting. In European Conference on Computer Vision, pages 221-238. Springer, 2024. 2 +[32] Meng Wei, Qianyi Wu, Jianmin Zheng, Hamid Rezatofighi, and Jianfei Cai. Normal-gs: 3d gaussian splatting with normal-involved rendering. In Advances in Neural Information Processing Systems (NeurIPS), 2024. 2 +[33] Matthias Zwicker, Hanspeter Pfister, Jeroen Van Baar, and Markus Gross. Ewa splatting. IEEE Transactions on Visualization and Computer Graphics, 8(3):223-238, 2002. 3 \ No newline at end of file diff --git a/2025/Speedy-Splat_ Fast 3D Gaussian Splatting with Sparse Pixels and Sparse Primitives/images.zip b/2025/Speedy-Splat_ Fast 3D Gaussian Splatting with Sparse Pixels and Sparse Primitives/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..c1dae7cc516b6e6e15bffba39985f4687ba9e7c8 --- /dev/null +++ b/2025/Speedy-Splat_ Fast 3D Gaussian Splatting with Sparse Pixels and Sparse Primitives/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ae6d117899a9f56dabeed1e506489dc745f3cc48ef77ec2dab96c6c69dad32f6 +size 554422 diff --git a/2025/Speedy-Splat_ Fast 3D Gaussian Splatting with Sparse Pixels and Sparse Primitives/layout.json b/2025/Speedy-Splat_ Fast 3D Gaussian Splatting with Sparse Pixels and Sparse Primitives/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..38648e81b551a4ecd511d025bf6aff550a86e807 --- /dev/null +++ b/2025/Speedy-Splat_ Fast 3D Gaussian Splatting with Sparse Pixels and Sparse Primitives/layout.json @@ -0,0 +1,12661 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 160, + 103, + 451, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 103, + 451, + 140 + ], + "spans": [ + { + "bbox": [ + 160, + 103, + 451, + 140 + ], + "type": "text", + "content": "Speedy-Splat: Fast 3D Gaussian Splating with Sparse Pixels and Sparse Primitives" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 155, + 161, + 220, + 174 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 155, + 161, + 220, + 174 + ], + "spans": [ + { + "bbox": [ + 155, + 161, + 220, + 174 + ], + "type": "text", + "content": "Alex Hanson" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 199, + 176, + 287, + 188 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 199, + 176, + 287, + 188 + ], + "spans": [ + { + "bbox": [ + 199, + 176, + 287, + 188 + ], + "type": "text", + "content": "Matthias Zwicker" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 248, + 163, + 293, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 248, + 163, + 293, + 175 + ], + "spans": [ + { + "bbox": [ + 248, + 163, + 293, + 175 + ], + "type": "text", + "content": "Allen Tu" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 320, + 162, + 368, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 162, + 368, + 175 + ], + "spans": [ + { + "bbox": [ + 320, + 162, + 368, + 175 + ], + "type": "text", + "content": "Geng Lin" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 396, + 162, + 455, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 396, + 162, + 455, + 175 + ], + "spans": [ + { + "bbox": [ + 396, + 162, + 455, + 175 + ], + "type": "text", + "content": "Vasu Singla" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 339, + 176, + 411, + 188 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 176, + 411, + 188 + ], + "spans": [ + { + "bbox": [ + 339, + 176, + 411, + 188 + ], + "type": "text", + "content": "Tom Goldstein" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 214, + 198, + 396, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 214, + 198, + 396, + 212 + ], + "spans": [ + { + "bbox": [ + 214, + 198, + 396, + 212 + ], + "type": "text", + "content": "University of Maryland, College Park" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 58, + 238, + 218, + 327 + ], + "blocks": [ + { + "bbox": [ + 129, + 224, + 146, + 235 + ], + "lines": [ + { + "bbox": [ + 129, + 224, + 146, + 235 + ], + "spans": [ + { + "bbox": [ + 129, + 224, + 146, + 235 + ], + "type": "text", + "content": "GT" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 58, + 238, + 218, + 327 + ], + "lines": [ + { + "bbox": [ + 58, + 238, + 218, + 327 + ], + "spans": [ + { + "bbox": [ + 58, + 238, + 218, + 327 + ], + "type": "image", + "image_path": "80c8529a039791fdbf6f182b0191918f7f9585ea1fe19c1b57530b82780aa8d3.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 223, + 238, + 382, + 327 + ], + "blocks": [ + { + "bbox": [ + 286, + 224, + 317, + 236 + ], + "lines": [ + { + "bbox": [ + 286, + 224, + 317, + 236 + ], + "spans": [ + { + "bbox": [ + 286, + 224, + 317, + 236 + ], + "type": "text", + "content": "3DGS" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 223, + 238, + 382, + 327 + ], + "lines": [ + { + "bbox": [ + 223, + 238, + 382, + 327 + ], + "spans": [ + { + "bbox": [ + 223, + 238, + 382, + 327 + ], + "type": "image", + "image_path": "22afe16fab5e67b7e86920db6ee04ce9f8c1b3df5ce2042922144023e376646d.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 225, + 328, + 285, + 338 + ], + "lines": [ + { + "bbox": [ + 225, + 328, + 285, + 338 + ], + "spans": [ + { + "bbox": [ + 225, + 328, + 285, + 338 + ], + "type": "text", + "content": "2.6M Gaussians" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 288, + 328, + 332, + 338 + ], + "lines": [ + { + "bbox": [ + 288, + 328, + 332, + 338 + ], + "spans": [ + { + "bbox": [ + 288, + 328, + 332, + 338 + ], + "type": "text", + "content": "25.39 PSNR" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 337, + 328, + 377, + 338 + ], + "lines": [ + { + "bbox": [ + 337, + 328, + 377, + 338 + ], + "spans": [ + { + "bbox": [ + 337, + 328, + 377, + 338 + ], + "type": "text", + "content": "184.87 FPS" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 346, + 555, + 368 + ], + "lines": [ + { + "bbox": [ + 55, + 346, + 555, + 368 + ], + "spans": [ + { + "bbox": [ + 55, + 346, + 555, + 368 + ], + "type": "text", + "content": "Figure 1. We reduce the number of Gaussians by over " + }, + { + "bbox": [ + 55, + 346, + 555, + 368 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 55, + 346, + 555, + 368 + ], + "type": "text", + "content": ", only marginally decrease PSNR, and accelerate rendering speed by " + }, + { + "bbox": [ + 55, + 346, + 555, + 368 + ], + "type": "inline_equation", + "content": "6.2\\times" + }, + { + "bbox": [ + 55, + 346, + 555, + 368 + ], + "type": "text", + "content": " in the Tanks & Temples truck scene when compared to 3D Gaussian Splatting (3D-GS). Additionally, we speed up training time by " + }, + { + "bbox": [ + 55, + 346, + 555, + 368 + ], + "type": "inline_equation", + "content": "1.38\\times" + }, + { + "bbox": [ + 55, + 346, + 555, + 368 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 386, + 237, + 547, + 327 + ], + "blocks": [ + { + "bbox": [ + 456, + 224, + 482, + 236 + ], + "lines": [ + { + "bbox": [ + 456, + 224, + 482, + 236 + ], + "spans": [ + { + "bbox": [ + 456, + 224, + 482, + 236 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 386, + 237, + 547, + 327 + ], + "lines": [ + { + "bbox": [ + 386, + 237, + 547, + 327 + ], + "spans": [ + { + "bbox": [ + 386, + 237, + 547, + 327 + ], + "type": "image", + "image_path": "78fb40f2c812639d67118378d4bc2cff166da1aa572f4c0b2d66ca5de40ec0c8.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 388, + 328, + 450, + 338 + ], + "lines": [ + { + "bbox": [ + 388, + 328, + 450, + 338 + ], + "spans": [ + { + "bbox": [ + 388, + 328, + 450, + 338 + ], + "type": "text", + "content": "0.26M Gaussians" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 454, + 328, + 498, + 338 + ], + "lines": [ + { + "bbox": [ + 454, + 328, + 498, + 338 + ], + "spans": [ + { + "bbox": [ + 454, + 328, + 498, + 338 + ], + "type": "text", + "content": "25.34 PSNR" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 502, + 328, + 547, + 338 + ], + "lines": [ + { + "bbox": [ + 502, + 328, + 547, + 338 + ], + "spans": [ + { + "bbox": [ + 502, + 328, + 547, + 338 + ], + "type": "text", + "content": "1148.77 FPS" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "bbox": [ + 152, + 378, + 200, + 390 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 378, + 200, + 390 + ], + "spans": [ + { + "bbox": [ + 152, + 378, + 200, + 390 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 54, + 403, + 297, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 403, + 297, + 643 + ], + "spans": [ + { + "bbox": [ + 54, + 403, + 297, + 643 + ], + "type": "text", + "content": "3D Gaussian Splatting (3D-GS) is a recent 3D scene reconstruction technique that enables real-time rendering of novel views by modeling scenes as parametric point clouds of differentiable 3D Gaussians. However, its rendering speed and model size still present bottlenecks, especially in resource-constrained settings. In this paper, we identify and address two key inefficiencies in 3D-GS to substantially improve rendering speed. These improvements also yield the ancillary benefits of reduced model size and training time. First, we optimize the rendering pipeline to precisely localize Gaussians in the scene, boosting rendering speed without altering visual fidelity. Second, we introduce a novel pruning technique and integrate it into the training pipeline, significantly reducing model size and training time while further raising rendering speed. Our Speedy-Splat approach combines these techniques to accelerate average rendering speed by a drastic " + }, + { + "bbox": [ + 54, + 403, + 297, + 643 + ], + "type": "inline_equation", + "content": "6.71 \\times" + }, + { + "bbox": [ + 54, + 403, + 297, + 643 + ], + "type": "text", + "content": " across scenes from the Mip-NeRF 360, Tanks & Temples, and Deep Blending datasets. Code and additional visualizations are available at https://speedysplat.github.io/." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 55, + 658, + 135, + 669 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 658, + 135, + 669 + ], + "spans": [ + { + "bbox": [ + 55, + 658, + 135, + 669 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": "Fast rendering of photorealistic novel views has been a long-standing goal in computer vision and graphics. Neural Radiance Fields (NeRF) [23] and its variants have made sig-" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 379, + 555, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 379, + 555, + 570 + ], + "spans": [ + { + "bbox": [ + 313, + 379, + 555, + 570 + ], + "type": "text", + "content": "nificant strides in photorealistic 3D scene reconstruction by representing scenes as continuous neural volumetric models that encode scene density and color at spatial coordinates. However, despite recent efforts [4, 24, 29], fast rendering in NeRF remains challenging because the volumetric sampling used in ray-marching is computationally expensive. Recently, 3D Gaussian Splatting (3D-GS) [14] has emerged as a promising alternative that enables real-time rendering by modeling scenes as parametric point clouds of differentiable 3D Gaussians. Nevertheless, its rendering speed is still limited by high parameter counts and certain algorithmic inefficiencies. Efficient rendering is essential in applications such as virtual reality, networked systems, and multi-view streaming. Real-time rendering on resource-constrained edge devices, such as mobile phones, has yet to be achieved [19]." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 313, + 576, + 556, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 576, + 556, + 673 + ], + "spans": [ + { + "bbox": [ + 313, + 576, + 556, + 673 + ], + "type": "text", + "content": "Although recent works on compressing 3D-GS models [6, 7, 11, 25] achieve some speed-ups by reducing the number of parameters, few approaches directly target rendering speed [8, 19]. In this paper, we specifically address this gap by demonstrating that the rendering speed of 3D-GS models can be drastically increased while maintaining competitive image quality. Additionally, we show that our methods reduce training time and substantially decrease model size." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "text", + "content": "We begin by observing that the cost of 3D-GS rendering is proportional to both the number of Gaussians in the scene and the number of pixels processed per Gaussian. Our ap-" + } + ] + } + ], + "index": 29 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21537" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "type": "text", + "content": "proach optimizes both factors. First, we reduce the number of pixels that are processed for each Gaussian by efficiently and accurately localizing it in the rendered image. Second, we reduce the total number of Gaussians in the model through a novel approach that maintains rendering quality." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 133, + 294, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 133, + 294, + 263 + ], + "spans": [ + { + "bbox": [ + 54, + 133, + 294, + 263 + ], + "type": "text", + "content": "3D-GS implements tile rendering to localize Gaussians in the image plane by assigning each Gaussian to the tiles that it intersects. We find the existing algorithm to be overly conservative; in Section 4.1.1, we address this by introducing our SnugBox algorithm to precisely localize Gaussians by computing a tight bounding box around their extent. Then, in Section 4.1.2, we extend SnugBox with our AccuTile algorithm to identify exact tile intersections. Both approaches are plug-and-play, lead to respective inference speed-ups of " + }, + { + "bbox": [ + 54, + 133, + 294, + 263 + ], + "type": "inline_equation", + "content": "1.82 \\times" + }, + { + "bbox": [ + 54, + 133, + 294, + 263 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 133, + 294, + 263 + ], + "type": "inline_equation", + "content": "1.99 \\times" + }, + { + "bbox": [ + 54, + 133, + 294, + 263 + ], + "type": "text", + "content": " on average, and do not change the 3D-GS renderings." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 264, + 295, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 264, + 295, + 395 + ], + "spans": [ + { + "bbox": [ + 55, + 264, + 295, + 395 + ], + "type": "text", + "content": "To reduce the total number of Gaussians while preserving visual fidelity, we extend an existing pruning method [11] by reducing its memory requirement by " + }, + { + "bbox": [ + 55, + 264, + 295, + 395 + ], + "type": "inline_equation", + "content": "36\\times" + }, + { + "bbox": [ + 55, + 264, + 295, + 395 + ], + "type": "text", + "content": " and incorporating it into the 3D-GS training pipeline. During the densification stage of 3D-GS training, Gaussians are regularly replicated and pruned. In Section 4.2.2, we augment the densification stage with our Soft Pruning method to prune " + }, + { + "bbox": [ + 55, + 264, + 295, + 395 + ], + "type": "inline_equation", + "content": "80\\%" + }, + { + "bbox": [ + 55, + 264, + 295, + 395 + ], + "type": "text", + "content": " of Gaussians at regular intervals. After the densification stage, our Hard Pruning method, described in Section 4.2.3, prunes an additional " + }, + { + "bbox": [ + 55, + 264, + 295, + 395 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 55, + 264, + 295, + 395 + ], + "type": "text", + "content": " of Gaussians at set intervals." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 396, + 295, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 396, + 295, + 456 + ], + "spans": [ + { + "bbox": [ + 55, + 396, + 295, + 456 + ], + "type": "text", + "content": "Our Speedy-Splat approach integrates both techniques into the 3D-GS training pipeline. On average, rendering speed is accelerated by " + }, + { + "bbox": [ + 55, + 396, + 295, + 456 + ], + "type": "inline_equation", + "content": "6.71 \\times" + }, + { + "bbox": [ + 55, + 396, + 295, + 456 + ], + "type": "text", + "content": ", model size is reduced by " + }, + { + "bbox": [ + 55, + 396, + 295, + 456 + ], + "type": "inline_equation", + "content": "10.6 \\times" + }, + { + "bbox": [ + 55, + 396, + 295, + 456 + ], + "type": "text", + "content": ", and training speed is improved by " + }, + { + "bbox": [ + 55, + 396, + 295, + 456 + ], + "type": "inline_equation", + "content": "1.4 \\times" + }, + { + "bbox": [ + 55, + 396, + 295, + 456 + ], + "type": "text", + "content": " across all evaluated scenes while maintaining high image quality." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 457, + 282, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 457, + 282, + 468 + ], + "spans": [ + { + "bbox": [ + 67, + 457, + 282, + 468 + ], + "type": "text", + "content": "In summary, we propose the following contributions:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 472, + 295, + 567 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 56, + 472, + 295, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 472, + 295, + 495 + ], + "spans": [ + { + "bbox": [ + 56, + 472, + 295, + 495 + ], + "type": "text", + "content": "1. SnugBox: A precise algorithm for computing Gaussian-tile bounding box intersections." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 495, + 295, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 495, + 295, + 517 + ], + "spans": [ + { + "bbox": [ + 55, + 495, + 295, + 517 + ], + "type": "text", + "content": "2. AccuTile: An extension of SnugBox for computing exact Gaussian-tile intersections." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 519, + 295, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 519, + 295, + 542 + ], + "spans": [ + { + "bbox": [ + 55, + 519, + 295, + 542 + ], + "type": "text", + "content": "3. Soft Pruning: An augmentation for pruning Gaussians during densification." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 543, + 295, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 543, + 295, + 567 + ], + "spans": [ + { + "bbox": [ + 55, + 543, + 295, + 567 + ], + "type": "text", + "content": "4. Hard Pruning: An augmentation for pruning Gaussians post-densification." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 582, + 139, + 594 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 139, + 594 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 139, + 594 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 603, + 295, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 603, + 295, + 664 + ], + "spans": [ + { + "bbox": [ + 55, + 603, + 295, + 664 + ], + "type": "text", + "content": "The real-time rendering speed of 3D-GS [14] on desktop GPUs has inspired research focused on further accelerating both its training and inference in resource-constrained environments. In this section, we review related works that specifically target these performance improvements." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 672, + 117, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 672, + 117, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 672, + 117, + 685 + ], + "type": "text", + "content": "2.1. Pruning" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "content": "A large portion of Gaussians in vanilla 3D-GS models are redundant [6, 7, 25], motivating a recent line of" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 72, + 553, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 228 + ], + "type": "text", + "content": "work focused on pruning Gaussians from 3D-GS models to boost rendering speed with minimal loss of visual fidelity [1, 2, 6, 7, 10, 11, 18, 19, 25]. Nearly all approaches assign a significance score to each Gaussian that is used to rank and prune them. Several works compute the aggregated ray contribution for each Gaussian across all input images [6, 7, 18, 25], while others combine opacity with additional information, such as gradients per Gaussian, to calculate their pruning criterion [1, 2, 20]. Papantonakis et al. [26] use resolution and scale-aware redundancy metrics, EAGLES [10] calculates the total transmittance per Gaussian, and Lin et al. [19] accelerate inference speed through an efficiency-aware strategy and foveated rendering." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 228, + 553, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 228, + 553, + 324 + ], + "spans": [ + { + "bbox": [ + 313, + 228, + 553, + 324 + ], + "type": "text", + "content": "PUP 3D-GS [11] introduces a more principled approach by deriving a Hessian for each Gaussian that represents its sensitivity to the reconstruction error. While PUP 3D-GS achieves state-of-the-art post-hoc pruning results, computing its sensitivity score incurs considerable storage requirements that limit its viability for use during training. Our pruning approach directly extends PUP 3D-GS by improving its memory efficiency by " + }, + { + "bbox": [ + 313, + 228, + 553, + 324 + ], + "type": "inline_equation", + "content": "36\\times" + }, + { + "bbox": [ + 313, + 228, + 553, + 324 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 333, + 408, + 344 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 333, + 408, + 344 + ], + "spans": [ + { + "bbox": [ + 313, + 333, + 408, + 344 + ], + "type": "text", + "content": "2.2. Other Methods" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 351, + 555, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 351, + 555, + 591 + ], + "spans": [ + { + "bbox": [ + 313, + 351, + 555, + 591 + ], + "type": "text", + "content": "In addition to pruning, several other strategies have been explored to enhance the rendering and training speed of 3D-GS. Mini-Splatting [7] modifies its densification strategies and adds a simplification stage to constrain the number of Gaussians. 3DGS-MCMC [15] models training dynamics as an MCMC process. Revisiting Densification [30] introduces an error-based densification strategy. 3DGS-LM [13] replaces Adam [16] with a tailored Levenberg-Marquardt optimizer [9]. Taming 3DGS [22] proposes a constructive optimization process that limits the number of Gaussians to a pre-defined threshold set by the user. DISTWAR [5] dives into the low-level implementation of GPU thread scheduling and optimizes atomic processing with a novel primitive. StopThePop [27] introduces a precise tile intersection method, but we find that our approach is notably faster in Appendix A.4. FlashGS [8] implements a tile intersection method similar to StopThePop. Several works reduce training time and memory requirements by enforcing geometric constraints [21, 28, 31, 32]. Most of these approaches are orthogonal to ours and can be applied alongside it." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 603, + 392, + 616 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 603, + 392, + 616 + ], + "spans": [ + { + "bbox": [ + 313, + 603, + 392, + 616 + ], + "type": "text", + "content": "3. Background" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 624, + 489, + 636 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 624, + 489, + 636 + ], + "spans": [ + { + "bbox": [ + 313, + 624, + 489, + 636 + ], + "type": "text", + "content": "3.1. 3D Gaussian Splitting Overview" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 641, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 553, + 713 + ], + "type": "text", + "content": "3D Gaussian Splatting (3D-GS) [14] models scenes as parametric, point-based representations that use differentiable 3D Gaussians as primitives. Given a set of ground truth training images " + }, + { + "bbox": [ + 313, + 641, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_{gt} = \\{\\pmb{I}_i\\in \\mathbb{R}^{H\\times W}\\}_{i = 1}^K" + }, + { + "bbox": [ + 313, + 641, + 553, + 713 + ], + "type": "text", + "content": ", the scene is initialized by using Structure from Motion (SfM) to produce a sparse point cloud that serves as the initial means for the" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21538" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 57, + 113, + 551, + 182 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 104 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 104 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 104 + ], + "type": "text", + "content": "Table 1. Average execution time (milliseconds) of each function across all scenes in Section 5.1. The operation in each row is applied cumulatively to all of the following rows. For each model, accurate measurements are collected by averaging execution times across three runs that each render the test set 20 times to reduce variance. The fastest and second fastest times are color coded." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 113, + 551, + 182 + ], + "lines": [ + { + "bbox": [ + 57, + 113, + 551, + 182 + ], + "spans": [ + { + "bbox": [ + 57, + 113, + 551, + 182 + ], + "type": "table", + "html": "
MethodPreprocessInclusive SumDuplicate with KeysRadix SortIdentify Tile RangesRenderOverall
Baseline0.6650.0450.5701.5510.0824.4837.478
+SnugBox0.6560.0460.208 (2.738×)0.729 (2.126×)0.041 (1.980×)2.344 (1.913×)4.102 (1.823×)
+AccuTile0.6680.0460.221 (2.575×)0.612 (2.533×)0.035 (2.326×)2.062 (2.175×)3.748 (1.995×)
+Soft Pruning0.370 (1.798×)0.030 (1.494×)0.146 (3.906×)0.404 (3.843×)0.024 (3.422×)1.337 (3.354×)2.381 (3.141×)
+Hard Pruning0.091 (7.293×)0.016 (2.769×)0.090 (6.325×)0.215 (7.217×)0.013 (6.537×)0.619 (7.247×)1.114 (6.712×)
", + "image_path": "72327e30568d0ffb3f3c89e03eb332008a3ae91c21bd6b6aed453f5a17def6ea.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 196, + 295, + 231 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 196, + 295, + 231 + ], + "spans": [ + { + "bbox": [ + 55, + 196, + 295, + 231 + ], + "type": "text", + "content": "3D Gaussians. The estimated camera poses " + }, + { + "bbox": [ + 55, + 196, + 295, + 231 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{gt} = \\{\\phi_i\\in" + }, + { + "bbox": [ + 55, + 196, + 295, + 231 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{3\\times 4}\\}_{i = 1}^{K}" + }, + { + "bbox": [ + 55, + 196, + 295, + 231 + ], + "type": "text", + "content": " are paired with their corresponding images and are used to optimize the scene." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 232, + 296, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 232, + 296, + 293 + ], + "spans": [ + { + "bbox": [ + 55, + 232, + 296, + 293 + ], + "type": "text", + "content": "Each 3D Gaussian primitive " + }, + { + "bbox": [ + 55, + 232, + 296, + 293 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_i" + }, + { + "bbox": [ + 55, + 232, + 296, + 293 + ], + "type": "text", + "content": " is parameterized by three geometry parameters - mean " + }, + { + "bbox": [ + 55, + 232, + 296, + 293 + ], + "type": "inline_equation", + "content": "\\mu_i\\in \\mathbb{R}^3" + }, + { + "bbox": [ + 55, + 232, + 296, + 293 + ], + "type": "text", + "content": ", scale " + }, + { + "bbox": [ + 55, + 232, + 296, + 293 + ], + "type": "inline_equation", + "content": "s_i\\in \\mathbb{R}^3" + }, + { + "bbox": [ + 55, + 232, + 296, + 293 + ], + "type": "text", + "content": ", and rotation " + }, + { + "bbox": [ + 55, + 232, + 296, + 293 + ], + "type": "inline_equation", + "content": "r_i\\in \\mathbb{R}^4" + }, + { + "bbox": [ + 55, + 232, + 296, + 293 + ], + "type": "text", + "content": " - and two color parameters - view-dependent spherical harmonics " + }, + { + "bbox": [ + 55, + 232, + 296, + 293 + ], + "type": "inline_equation", + "content": "h_i\\in \\mathbb{R}^{16\\times 3}" + }, + { + "bbox": [ + 55, + 232, + 296, + 293 + ], + "type": "text", + "content": " and opacity " + }, + { + "bbox": [ + 55, + 232, + 296, + 293 + ], + "type": "inline_equation", + "content": "\\sigma_{i}\\in \\mathbb{R}" + }, + { + "bbox": [ + 55, + 232, + 296, + 293 + ], + "type": "text", + "content": ". The set of all parameters can be described as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 103, + 301, + 295, + 316 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 103, + 301, + 295, + 316 + ], + "spans": [ + { + "bbox": [ + 103, + 301, + 295, + 316 + ], + "type": "interline_equation", + "content": "\\mathcal {G} = \\left\\{\\mathcal {G} _ {i} = \\left\\{\\mu_ {i}, s _ {i}, r _ {i}, h _ {i}, \\sigma_ {i} \\right\\} \\right\\} _ {i = 1} ^ {N}, \\tag {1}", + "image_path": "9e250b99343ac697f724c37b61fed6ad79fad8fa4d615ab1778f59f136ced930.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 325, + 260, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 325, + 260, + 337 + ], + "spans": [ + { + "bbox": [ + 55, + 325, + 260, + 337 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 325, + 260, + 337 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 325, + 260, + 337 + ], + "type": "text", + "content": " is the number of Gaussians in the model." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 338, + 296, + 385 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 338, + 296, + 385 + ], + "spans": [ + { + "bbox": [ + 55, + 338, + 296, + 385 + ], + "type": "text", + "content": "Given camera pose " + }, + { + "bbox": [ + 55, + 338, + 296, + 385 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 55, + 338, + 296, + 385 + ], + "type": "text", + "content": ", the scene is rendered by projecting all Gaussians to image space and applying alpha blending to each pixel. Models are optimized via stochastic gradient descent on image reconstruction losses:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 61, + 396, + 295, + 410 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 396, + 295, + 410 + ], + "spans": [ + { + "bbox": [ + 61, + 396, + 295, + 410 + ], + "type": "interline_equation", + "content": "L (\\mathcal {G} | \\phi , I _ {g t}) = \\| I _ {\\mathcal {G}} (\\phi) - I _ {g t} \\| _ {1} + L _ {\\mathrm {D - S S I M}} (I _ {\\mathcal {G}} (\\phi), I _ {g t}), \\tag {2}", + "image_path": "fd1de816a06cea0a5d26c08914c053e30dc3a055c8aa07984ea7ee1404c76d3b.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 419, + 241, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 419, + 241, + 432 + ], + "spans": [ + { + "bbox": [ + 55, + 419, + 241, + 432 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 419, + 241, + 432 + ], + "type": "inline_equation", + "content": "I_{\\mathcal{G}(\\phi)}" + }, + { + "bbox": [ + 55, + 419, + 241, + 432 + ], + "type": "text", + "content": " is the rendered image for pose " + }, + { + "bbox": [ + 55, + 419, + 241, + 432 + ], + "type": "inline_equation", + "content": "\\phi" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 432, + 296, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 432, + 296, + 479 + ], + "spans": [ + { + "bbox": [ + 55, + 432, + 296, + 479 + ], + "type": "text", + "content": "During optimization, the scene is periodically densified by cloning and splitting uncertain Gaussians and pruned by removing large and transparent Gaussians. The opacities of the Gaussians are also periodically reset." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 490, + 278, + 503 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 490, + 278, + 503 + ], + "spans": [ + { + "bbox": [ + 55, + 490, + 278, + 503 + ], + "type": "text", + "content": "3.2. 3D Gaussian Splitting Rendering Specifics" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 508, + 295, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 508, + 295, + 579 + ], + "spans": [ + { + "bbox": [ + 55, + 508, + 295, + 579 + ], + "type": "text", + "content": "3D-GS uses a tile-based rendering strategy that divides the rendered image into " + }, + { + "bbox": [ + 55, + 508, + 295, + 579 + ], + "type": "inline_equation", + "content": "16 \\times 16" + }, + { + "bbox": [ + 55, + 508, + 295, + 579 + ], + "type": "text", + "content": " pixel tiles. Each Gaussian is projected into image space, where its intersection with these tiles is computed. Then, these Gaussian-to-tile mappings are sorted to collect and order the Gaussians by depth for pixel-wise rendering." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 580, + 296, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 580, + 296, + 640 + ], + "spans": [ + { + "bbox": [ + 55, + 580, + 296, + 640 + ], + "type": "text", + "content": "Rendering runtime is dominated by six key functions. Table 1 empirically analyzes the execution time of each function and highlights the improvements achieved by our methods. Descriptions of each function are provided in the following sections." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 650, + 144, + 662 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 650, + 144, + 662 + ], + "spans": [ + { + "bbox": [ + 55, + 650, + 144, + 662 + ], + "type": "text", + "content": "3.2.1. Preprocessing" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 665, + 296, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 712 + ], + "type": "text", + "content": "The preprocess kernel is parallelized such that each thread processes a single Gaussian " + }, + { + "bbox": [ + 55, + 665, + 296, + 712 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_i" + }, + { + "bbox": [ + 55, + 665, + 296, + 712 + ], + "type": "text", + "content": ". It computes a 2D projection of " + }, + { + "bbox": [ + 55, + 665, + 296, + 712 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_i" + }, + { + "bbox": [ + 55, + 665, + 296, + 712 + ], + "type": "text", + "content": " to image space and obtains a count of tiles that it intersects." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 196, + 555, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 196, + 555, + 266 + ], + "spans": [ + { + "bbox": [ + 313, + 196, + 555, + 266 + ], + "type": "text", + "content": "The mean " + }, + { + "bbox": [ + 313, + 196, + 555, + 266 + ], + "type": "inline_equation", + "content": "\\mu_{i}" + }, + { + "bbox": [ + 313, + 196, + 555, + 266 + ], + "type": "text", + "content": " is projected to image space using a viewing transform " + }, + { + "bbox": [ + 313, + 196, + 555, + 266 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 313, + 196, + 555, + 266 + ], + "type": "text", + "content": " and a perspective projection, yielding 2D mean " + }, + { + "bbox": [ + 313, + 196, + 555, + 266 + ], + "type": "inline_equation", + "content": "\\mu_{i_{2D}}" + }, + { + "bbox": [ + 313, + 196, + 555, + 266 + ], + "type": "text", + "content": " and depth that are stored for later processing. The scale " + }, + { + "bbox": [ + 313, + 196, + 555, + 266 + ], + "type": "inline_equation", + "content": "s_i" + }, + { + "bbox": [ + 313, + 196, + 555, + 266 + ], + "type": "text", + "content": " and rotation " + }, + { + "bbox": [ + 313, + 196, + 555, + 266 + ], + "type": "inline_equation", + "content": "r_i" + }, + { + "bbox": [ + 313, + 196, + 555, + 266 + ], + "type": "text", + "content": " parameters are converted to the diagonal scale " + }, + { + "bbox": [ + 313, + 196, + 555, + 266 + ], + "type": "inline_equation", + "content": "S_{i}" + }, + { + "bbox": [ + 313, + 196, + 555, + 266 + ], + "type": "text", + "content": " and rotation " + }, + { + "bbox": [ + 313, + 196, + 555, + 266 + ], + "type": "inline_equation", + "content": "R_{i}" + }, + { + "bbox": [ + 313, + 196, + 555, + 266 + ], + "type": "text", + "content": " matrices. The 3D covariance is then defined as:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 388, + 280, + 553, + 295 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 388, + 280, + 553, + 295 + ], + "spans": [ + { + "bbox": [ + 388, + 280, + 553, + 295 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\Sigma} _ {i _ {3 D}} = \\boldsymbol {R} _ {i} \\boldsymbol {S} _ {i} \\boldsymbol {S} _ {i} ^ {T} \\boldsymbol {R} _ {i} ^ {T}, \\tag {3}", + "image_path": "2f6f13fc54a0c5509e88b34e60d25dcb0f63070055640217875a51152161d89e.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 326, + 304, + 420, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 304, + 420, + 316 + ], + "spans": [ + { + "bbox": [ + 326, + 304, + 420, + 316 + ], + "type": "text", + "content": "which is projected via:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 375, + 328, + 553, + 342 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 375, + 328, + 553, + 342 + ], + "spans": [ + { + "bbox": [ + 375, + 328, + 553, + 342 + ], + "type": "interline_equation", + "content": "\\hat {\\boldsymbol {\\Sigma}} _ {i _ {3 D}} = \\boldsymbol {J} \\boldsymbol {W} \\boldsymbol {\\Sigma} _ {i _ {3 D}} \\boldsymbol {W} ^ {T} \\boldsymbol {J} ^ {T}, \\tag {4}", + "image_path": "82b40a64762c743fe3671f1326ff23efb8deb704fc89928c5fc4d440cb423938.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 351, + 554, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 351, + 554, + 435 + ], + "spans": [ + { + "bbox": [ + 313, + 351, + 554, + 435 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 351, + 554, + 435 + ], + "type": "inline_equation", + "content": "J" + }, + { + "bbox": [ + 313, + 351, + 554, + 435 + ], + "type": "text", + "content": " is the Jacobian of the first order approximation of the perspective projection. Dropping the last row and column of " + }, + { + "bbox": [ + 313, + 351, + 554, + 435 + ], + "type": "inline_equation", + "content": "\\hat{\\Sigma}_{i_{3D}}" + }, + { + "bbox": [ + 313, + 351, + 554, + 435 + ], + "type": "text", + "content": " gives " + }, + { + "bbox": [ + 313, + 351, + 554, + 435 + ], + "type": "inline_equation", + "content": "\\Sigma_{i_{2D}}" + }, + { + "bbox": [ + 313, + 351, + 554, + 435 + ], + "type": "text", + "content": " [14, 33]. The largest eigenvalue of " + }, + { + "bbox": [ + 313, + 351, + 554, + 435 + ], + "type": "inline_equation", + "content": "\\Sigma_{i_{2D}}" + }, + { + "bbox": [ + 313, + 351, + 554, + 435 + ], + "type": "text", + "content": " is used to compute the count of tiles intersected by this Gaussian " + }, + { + "bbox": [ + 313, + 351, + 554, + 435 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_i" + }, + { + "bbox": [ + 313, + 351, + 554, + 435 + ], + "type": "text", + "content": " as shown in Figure 2a. " + }, + { + "bbox": [ + 313, + 351, + 554, + 435 + ], + "type": "inline_equation", + "content": "\\Sigma_{i_{2D}}^{-1}" + }, + { + "bbox": [ + 313, + 351, + 554, + 435 + ], + "type": "text", + "content": " is computed, along with a view-dependent color " + }, + { + "bbox": [ + 313, + 351, + 554, + 435 + ], + "type": "inline_equation", + "content": "c_{i}" + }, + { + "bbox": [ + 313, + 351, + 554, + 435 + ], + "type": "text", + "content": ", derived from " + }, + { + "bbox": [ + 313, + 351, + 554, + 435 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 313, + 351, + 554, + 435 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 351, + 554, + 435 + ], + "type": "inline_equation", + "content": "h_i" + }, + { + "bbox": [ + 313, + 351, + 554, + 435 + ], + "type": "text", + "content": ". All three values are stored for later processing." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 314, + 443, + 375, + 455 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 443, + 375, + 455 + ], + "spans": [ + { + "bbox": [ + 314, + 443, + 375, + 455 + ], + "type": "text", + "content": "3.2.2. Sorting" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 459, + 554, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 459, + 554, + 483 + ], + "spans": [ + { + "bbox": [ + 313, + 459, + 554, + 483 + ], + "type": "text", + "content": "After preprocessing, four functions process the Gaussians for pixel-wise rendering:" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 483, + 554, + 615 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 314, + 483, + 554, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 483, + 554, + 519 + ], + "spans": [ + { + "bbox": [ + 314, + 483, + 554, + 519 + ], + "type": "text", + "content": "- InclusiveSum: A CUDA primitive that computes the prefix sum over all Gaussian-tile counts to allocate key and value arrays for Gaussian-to-tile mapping." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 314, + 519, + 554, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 519, + 554, + 567 + ], + "spans": [ + { + "bbox": [ + 314, + 519, + 554, + 567 + ], + "type": "text", + "content": "- duplicateWithKeys: A Gaussian-parallel kernel that recomputes Gaussian-tile intersections to generate a key for each intersecting tile, consisting of the tile index and Gaussian depth." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 567, + 553, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 567, + 553, + 591 + ], + "spans": [ + { + "bbox": [ + 313, + 567, + 553, + 591 + ], + "type": "text", + "content": "- RadixSort: A CUDA primitive that sorts the key array, ordering Gaussian indices by tile and then depth." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 591, + 553, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 591, + 553, + 615 + ], + "spans": [ + { + "bbox": [ + 313, + 591, + 553, + 615 + ], + "type": "text", + "content": "- identifyTileRanges: A kernel parallelized across the key array to post-process keys before pixel rendering." + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 623, + 388, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 623, + 388, + 635 + ], + "spans": [ + { + "bbox": [ + 314, + 623, + 388, + 635 + ], + "type": "text", + "content": "3.2.3. Rendering" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 313, + 639, + 554, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 639, + 554, + 687 + ], + "spans": [ + { + "bbox": [ + 313, + 639, + 554, + 687 + ], + "type": "text", + "content": "The render kernel is parallelized across pixels. For each pixel " + }, + { + "bbox": [ + 313, + 639, + 554, + 687 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 313, + 639, + 554, + 687 + ], + "type": "text", + "content": ", all Gaussians within its corresponding tile are loaded and processed in depth order as determined by RadixSort. An alpha value:" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 399, + 701, + 553, + 714 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 399, + 701, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 399, + 701, + 553, + 714 + ], + "type": "interline_equation", + "content": "\\alpha_ {i} (p) = \\sigma_ {i} g _ {i} (p) \\tag {5}", + "image_path": "cac5c0c9424eaa0563b0aa18c3278dcb5c3e19a825df0290cfc0d9558a184633.jpg" + } + ] + } + ], + "index": 29 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21539" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "text", + "content": "is computed for each Gaussian " + }, + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_i" + }, + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "inline_equation", + "content": "g_i" + }, + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "text", + "content": " is the value of the projected 2D Gaussian at pixel " + }, + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 68, + 102, + 295, + 126 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 102, + 295, + 126 + ], + "spans": [ + { + "bbox": [ + 68, + 102, + 295, + 126 + ], + "type": "interline_equation", + "content": "g _ {i} = e ^ {q}, q = - \\frac {1}{2} (p - \\mu_ {i _ {2 D}}) \\boldsymbol {\\Sigma} _ {\\boldsymbol {i} _ {2 D}} ^ {- 1} (p - \\mu_ {i _ {2 D}}) ^ {T}. \\tag {6}", + "image_path": "e843af55769dce8628dff65180d5042a72ab0514582370d0cb8a4c7b4a24934c.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 133, + 295, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 133, + 295, + 158 + ], + "spans": [ + { + "bbox": [ + 55, + 133, + 295, + 158 + ], + "type": "text", + "content": "If " + }, + { + "bbox": [ + 55, + 133, + 295, + 158 + ], + "type": "inline_equation", + "content": "\\alpha_{i} > \\frac{1}{255}" + }, + { + "bbox": [ + 55, + 133, + 295, + 158 + ], + "type": "text", + "content": ", then the Gaussian is included in the alpha compositing of the pixel color " + }, + { + "bbox": [ + 55, + 133, + 295, + 158 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 55, + 133, + 295, + 158 + ], + "type": "text", + "content": ", given by:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 99, + 166, + 295, + 200 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 166, + 295, + 200 + ], + "spans": [ + { + "bbox": [ + 99, + 166, + 295, + 200 + ], + "type": "interline_equation", + "content": "C (p) = \\sum_ {i \\in \\mathcal {N}} c _ {i} \\alpha_ {i} (p) \\prod_ {j = 1} ^ {i - 1} (1 - \\alpha_ {j} (p)). \\tag {7}", + "image_path": "620f1a24395dd746c8efcd501e8cb645d8522d9c45abf7f1bd722af7f9377014.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 207, + 117, + 220 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 207, + 117, + 220 + ], + "spans": [ + { + "bbox": [ + 55, + 207, + 117, + 220 + ], + "type": "text", + "content": "4. Methods" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 228, + 296, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 228, + 296, + 300 + ], + "spans": [ + { + "bbox": [ + 55, + 228, + 296, + 300 + ], + "type": "text", + "content": "Our Speedy-Splat methods are motivated by two key insights into inefficiencies within the 3D-GS rendering pipeline. First, Gaussian Splatting grossly overestimates the extent of Gaussians in the image. Second, as demonstrated by [11] and other recent pruning works, 3D-GS models are heavily overparameterized." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 306, + 177, + 318 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 306, + 177, + 318 + ], + "spans": [ + { + "bbox": [ + 55, + 306, + 177, + 318 + ], + "type": "text", + "content": "4.1. Precise Tile Intersect" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 323, + 296, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 323, + 296, + 384 + ], + "spans": [ + { + "bbox": [ + 55, + 323, + 296, + 384 + ], + "type": "text", + "content": "Gaussian Splitting identifies tiles intersected by Gaussian " + }, + { + "bbox": [ + 55, + 323, + 296, + 384 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_i" + }, + { + "bbox": [ + 55, + 323, + 296, + 384 + ], + "type": "text", + "content": " by calculating the maximum eigenvalue " + }, + { + "bbox": [ + 55, + 323, + 296, + 384 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{max}}" + }, + { + "bbox": [ + 55, + 323, + 296, + 384 + ], + "type": "text", + "content": " of its projected 2D covariance " + }, + { + "bbox": [ + 55, + 323, + 296, + 384 + ], + "type": "inline_equation", + "content": "\\Sigma_{i_2D}" + }, + { + "bbox": [ + 55, + 323, + 296, + 384 + ], + "type": "text", + "content": ", then selecting all tiles that intersect the square inscribing the circle defined by center " + }, + { + "bbox": [ + 55, + 323, + 296, + 384 + ], + "type": "inline_equation", + "content": "\\mu_{i_{2D}}" + }, + { + "bbox": [ + 55, + 323, + 296, + 384 + ], + "type": "text", + "content": " and radius:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 138, + 392, + 295, + 411 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 392, + 295, + 411 + ], + "spans": [ + { + "bbox": [ + 138, + 392, + 295, + 411 + ], + "type": "interline_equation", + "content": "r = \\left\\lceil 3 \\sqrt {\\lambda_ {\\max }} \\right\\rceil . \\tag {8}", + "image_path": "941bb0a5b15ed4ffe9d92d091b66eb42e7adc3e017a49757ab8991d3ff488f24.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 415, + 296, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 415, + 296, + 510 + ], + "spans": [ + { + "bbox": [ + 55, + 415, + 296, + 510 + ], + "type": "text", + "content": "This approach neglects opacity " + }, + { + "bbox": [ + 55, + 415, + 296, + 510 + ], + "type": "inline_equation", + "content": "\\sigma_{i}" + }, + { + "bbox": [ + 55, + 415, + 296, + 510 + ], + "type": "text", + "content": " in its calculation and generally overestimates the Gaussian extent, as illustrated by Figure 2a. The actual extent of Gaussian " + }, + { + "bbox": [ + 55, + 415, + 296, + 510 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_i" + }, + { + "bbox": [ + 55, + 415, + 296, + 510 + ], + "type": "text", + "content": ", shown in Figure 2b, is given by the threshold placed on its alpha value " + }, + { + "bbox": [ + 55, + 415, + 296, + 510 + ], + "type": "inline_equation", + "content": "\\alpha_{i}" + }, + { + "bbox": [ + 55, + 415, + 296, + 510 + ], + "type": "text", + "content": ". Specifically, " + }, + { + "bbox": [ + 55, + 415, + 296, + 510 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_i" + }, + { + "bbox": [ + 55, + 415, + 296, + 510 + ], + "type": "text", + "content": " does not contribute to the rendering of pixel " + }, + { + "bbox": [ + 55, + 415, + 296, + 510 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 55, + 415, + 296, + 510 + ], + "type": "text", + "content": " if " + }, + { + "bbox": [ + 55, + 415, + 296, + 510 + ], + "type": "inline_equation", + "content": "\\alpha_{i} < \\frac{1}{255}" + }, + { + "bbox": [ + 55, + 415, + 296, + 510 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 55, + 415, + 296, + 510 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 55, + 415, + 296, + 510 + ], + "type": "text", + "content": ". By applying the actual extent in tile intersection calculations, we arrive at a far more concise set of intersected tiles." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 510, + 296, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 296, + 558 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 296, + 558 + ], + "type": "text", + "content": "We now show the derivation of this extent. The furthest pixel extent of Gaussian " + }, + { + "bbox": [ + 55, + 510, + 296, + 558 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_i" + }, + { + "bbox": [ + 55, + 510, + 296, + 558 + ], + "type": "text", + "content": " can be determined by directly substituting this threshold into Equation 5. Rearranging terms gives:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 70, + 564, + 295, + 580 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 564, + 295, + 580 + ], + "spans": [ + { + "bbox": [ + 70, + 564, + 295, + 580 + ], + "type": "interline_equation", + "content": "2 \\log (2 5 5 \\sigma_ {i}) = (p - \\mu_ {i _ {2 D}}) \\boldsymbol {\\Sigma} _ {\\boldsymbol {i _ {2 D}}} ^ {- 1} (p - \\mu_ {i _ {2 D}}) ^ {T}. \\qquad (9)", + "image_path": "500aecb9f1fd53bbf82340668c1457e43fbc1638d0f8a5ae1ad08013c9d4376c.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 586, + 121, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 586, + 121, + 597 + ], + "spans": [ + { + "bbox": [ + 55, + 586, + 121, + 597 + ], + "type": "text", + "content": "We can rewrite:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 65, + 603, + 295, + 631 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 603, + 295, + 631 + ], + "spans": [ + { + "bbox": [ + 65, + 603, + 295, + 631 + ], + "type": "interline_equation", + "content": "p = \\binom {p _ {x}} {p _ {y}}, \\mu_ {i _ {2 D}} = \\binom {\\mu_ {x}} {\\mu_ {y}}, \\boldsymbol {\\Sigma} _ {\\boldsymbol {i} _ {2 D}} ^ {- 1} = \\left( \\begin{array}{c c} a & b \\\\ b & c \\end{array} \\right). \\tag {10}", + "image_path": "27cade18b1638e24a4c642af44018ba97f8f5c541c84149f6d27d8db74a0def7.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 636, + 295, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 636, + 295, + 648 + ], + "spans": [ + { + "bbox": [ + 55, + 636, + 295, + 648 + ], + "type": "text", + "content": "Specifying threshold " + }, + { + "bbox": [ + 55, + 636, + 295, + 648 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 636, + 295, + 648 + ], + "type": "text", + "content": " and centered coordinates " + }, + { + "bbox": [ + 55, + 636, + 295, + 648 + ], + "type": "inline_equation", + "content": "x_{d}" + }, + { + "bbox": [ + 55, + 636, + 295, + 648 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 636, + 295, + 648 + ], + "type": "inline_equation", + "content": "y_{d}" + }, + { + "bbox": [ + 55, + 636, + 295, + 648 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 138, + 668, + 295, + 681 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 668, + 295, + 681 + ], + "spans": [ + { + "bbox": [ + 138, + 668, + 295, + 681 + ], + "type": "interline_equation", + "content": "t = 2 \\log (2 5 5 \\sigma_ {i}), \\tag {11}", + "image_path": "8c70ed9419fbd7b287418379bc36b573579988191a61dad2d1af51b5bcb4ed7f.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 144, + 684, + 295, + 696 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 144, + 684, + 295, + 696 + ], + "spans": [ + { + "bbox": [ + 144, + 684, + 295, + 696 + ], + "type": "interline_equation", + "content": "x _ {d} = p _ {x} - \\mu_ {x}, \\tag {12}", + "image_path": "883a022058cbcac9920d0bbb3ed4ae31d192e22eb406eea17158d6ef0c74b359.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 144, + 699, + 295, + 711 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 144, + 699, + 295, + 711 + ], + "spans": [ + { + "bbox": [ + 144, + 699, + 295, + 711 + ], + "type": "interline_equation", + "content": "y _ {d} = p _ {y} - \\mu_ {y}, \\tag {13}", + "image_path": "5e86d350907a64cabbd71cf0dbd17713e237826313914242e7986b1fc7bed57d.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "content": "gives the pixel extent with coordinates " + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "inline_equation", + "content": "x_{d}" + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "inline_equation", + "content": "y_{d}" + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "content": " that satisfy the ellipse equation:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 380, + 106, + 553, + 121 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 380, + 106, + 553, + 121 + ], + "spans": [ + { + "bbox": [ + 380, + 106, + 553, + 121 + ], + "type": "interline_equation", + "content": "t = a x _ {d} ^ {2} + 2 b x _ {d} y _ {d} + c y _ {d} ^ {2}. \\tag {14}", + "image_path": "b071a555c1f0f38a497ac853f496c258c5c6eed1304cf98baa3e1be71f61ea0c.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 131, + 555, + 214 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 131, + 555, + 214 + ], + "spans": [ + { + "bbox": [ + 313, + 131, + 555, + 214 + ], + "type": "text", + "content": "Our approach uses this pixel extent to reduce the number of tiles contained in the Gaussian-to-tile mappings for each Gaussian. We propose two methods for computing precise tile intersections. First, our SnugBox algorithm produces a tight bounding box around each Gaussian. Second, our AccuTile algorithm extends it to identify the exact set of tiles intersected by the Gaussian." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 314, + 222, + 381, + 234 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 222, + 381, + 234 + ], + "spans": [ + { + "bbox": [ + 314, + 222, + 381, + 234 + ], + "type": "text", + "content": "4.1.1. SnugBox" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 238, + 554, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 238, + 554, + 285 + ], + "spans": [ + { + "bbox": [ + 313, + 238, + 554, + 285 + ], + "type": "text", + "content": "Our SnugBox method uses this elliptical extent to compute an axis-aligned bounding box that more precisely identifies tiles intersected by Gaussian " + }, + { + "bbox": [ + 313, + 238, + 554, + 285 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_i" + }, + { + "bbox": [ + 313, + 238, + 554, + 285 + ], + "type": "text", + "content": ". To derive this bounding box, we rearrange Equation 14 to solve for " + }, + { + "bbox": [ + 313, + 238, + 554, + 285 + ], + "type": "inline_equation", + "content": "y_d" + }, + { + "bbox": [ + 313, + 238, + 554, + 285 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 361, + 295, + 553, + 322 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 295, + 553, + 322 + ], + "spans": [ + { + "bbox": [ + 361, + 295, + 553, + 322 + ], + "type": "interline_equation", + "content": "y _ {d} = \\frac {- b x _ {d} \\pm \\sqrt {\\left(b ^ {2} - a c\\right) x _ {d} ^ {2} + t c}}{c}. \\tag {15}", + "image_path": "ec7b7e32c96d4f36f024a67e17accc906bef3995f8030fe8e018b05944fcd16b.jpg" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 329, + 554, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 329, + 554, + 392 + ], + "spans": [ + { + "bbox": [ + 313, + 329, + 554, + 392 + ], + "type": "text", + "content": "To find the " + }, + { + "bbox": [ + 313, + 329, + 554, + 392 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 313, + 329, + 554, + 392 + ], + "type": "text", + "content": "-coordinate bounding box edges " + }, + { + "bbox": [ + 313, + 329, + 554, + 392 + ], + "type": "inline_equation", + "content": "y_{\\mathrm{min}}" + }, + { + "bbox": [ + 313, + 329, + 554, + 392 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 329, + 554, + 392 + ], + "type": "inline_equation", + "content": "y_{\\mathrm{max}}" + }, + { + "bbox": [ + 313, + 329, + 554, + 392 + ], + "type": "text", + "content": ", we identify the values of " + }, + { + "bbox": [ + 313, + 329, + 554, + 392 + ], + "type": "inline_equation", + "content": "x_d" + }, + { + "bbox": [ + 313, + 329, + 554, + 392 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 313, + 329, + 554, + 392 + ], + "type": "inline_equation", + "content": "\\partial y_d / \\partial x_d = 0" + }, + { + "bbox": [ + 313, + 329, + 554, + 392 + ], + "type": "text", + "content": ". We refer to these " + }, + { + "bbox": [ + 313, + 329, + 554, + 392 + ], + "type": "inline_equation", + "content": "x_d" + }, + { + "bbox": [ + 313, + 329, + 554, + 392 + ], + "type": "text", + "content": " values as " + }, + { + "bbox": [ + 313, + 329, + 554, + 392 + ], + "type": "inline_equation", + "content": "x_{d_{\\mathrm{arg}s}}" + }, + { + "bbox": [ + 313, + 329, + 554, + 392 + ], + "type": "text", + "content": " to specify that they are the " + }, + { + "bbox": [ + 313, + 329, + 554, + 392 + ], + "type": "inline_equation", + "content": "\\arg \\min y_d" + }, + { + "bbox": [ + 313, + 329, + 554, + 392 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 329, + 554, + 392 + ], + "type": "inline_equation", + "content": "\\arg \\max y_d" + }, + { + "bbox": [ + 313, + 329, + 554, + 392 + ], + "type": "text", + "content": " values. Differentiating Equation 15 and solving for " + }, + { + "bbox": [ + 313, + 329, + 554, + 392 + ], + "type": "inline_equation", + "content": "x_{d_{\\mathrm{arg}s}}" + }, + { + "bbox": [ + 313, + 329, + 554, + 392 + ], + "type": "text", + "content": " gives:" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 380, + 400, + 553, + 433 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 380, + 400, + 553, + 433 + ], + "spans": [ + { + "bbox": [ + 380, + 400, + 553, + 433 + ], + "type": "interline_equation", + "content": "x _ {d _ {\\text {a r g s}}} = \\pm \\sqrt {\\frac {- b ^ {2} t}{\\left(b ^ {2} - a c\\right) a}}. \\tag {16}", + "image_path": "ad4ae141ea8c48d1033bc373112e75a780757867df3ed6b124da13397932b73c.jpg" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 438, + 554, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 438, + 554, + 499 + ], + "spans": [ + { + "bbox": [ + 313, + 438, + 554, + 499 + ], + "type": "text", + "content": "Substituting " + }, + { + "bbox": [ + 313, + 438, + 554, + 499 + ], + "type": "inline_equation", + "content": "x_{d_{\\text{args}}}" + }, + { + "bbox": [ + 313, + 438, + 554, + 499 + ], + "type": "text", + "content": " into Equation 15 and adding " + }, + { + "bbox": [ + 313, + 438, + 554, + 499 + ], + "type": "inline_equation", + "content": "\\mu_y" + }, + { + "bbox": [ + 313, + 438, + 554, + 499 + ], + "type": "text", + "content": " gives " + }, + { + "bbox": [ + 313, + 438, + 554, + 499 + ], + "type": "inline_equation", + "content": "y_{\\min}" + }, + { + "bbox": [ + 313, + 438, + 554, + 499 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 438, + 554, + 499 + ], + "type": "inline_equation", + "content": "y_{\\max}" + }, + { + "bbox": [ + 313, + 438, + 554, + 499 + ], + "type": "text", + "content": ". Due to the symmetry of Equation 14, we can find the " + }, + { + "bbox": [ + 313, + 438, + 554, + 499 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 438, + 554, + 499 + ], + "type": "text", + "content": "-coordinate bounding box edges " + }, + { + "bbox": [ + 313, + 438, + 554, + 499 + ], + "type": "inline_equation", + "content": "x_{\\min}" + }, + { + "bbox": [ + 313, + 438, + 554, + 499 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 438, + 554, + 499 + ], + "type": "inline_equation", + "content": "x_{\\max}" + }, + { + "bbox": [ + 313, + 438, + 554, + 499 + ], + "type": "text", + "content": " by swapping " + }, + { + "bbox": [ + 313, + 438, + 554, + 499 + ], + "type": "inline_equation", + "content": "y_d" + }, + { + "bbox": [ + 313, + 438, + 554, + 499 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 438, + 554, + 499 + ], + "type": "inline_equation", + "content": "x_d" + }, + { + "bbox": [ + 313, + 438, + 554, + 499 + ], + "type": "text", + "content": " and constants " + }, + { + "bbox": [ + 313, + 438, + 554, + 499 + ], + "type": "inline_equation", + "content": "a" + }, + { + "bbox": [ + 313, + 438, + 554, + 499 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 438, + 554, + 499 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 438, + 554, + 499 + ], + "type": "text", + "content": " to rewrite Equations 15 and 16 in terms of " + }, + { + "bbox": [ + 313, + 438, + 554, + 499 + ], + "type": "inline_equation", + "content": "x_d" + }, + { + "bbox": [ + 313, + 438, + 554, + 499 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 438, + 554, + 499 + ], + "type": "inline_equation", + "content": "y_{d_{\\text{args}}}" + }, + { + "bbox": [ + 313, + 438, + 554, + 499 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 499, + 555, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 499, + 555, + 643 + ], + "spans": [ + { + "bbox": [ + 313, + 499, + 555, + 643 + ], + "type": "text", + "content": "After the bounding box edges are identified, our method follows 3D-GS and converts these edges to tile indices by dividing by tile size, rounding, and clipping to the image boundary. As depicted in Figure 2, Snugbox can produce a significantly tighter bounding box than 3D-GS. Meanwhile, its computational overhead is small as it performs a constant number of operations and is only called twice in the rendering pipeline – once to count the Gaussians intersecting tiles in preprocess and once to populate the Gaussian-to-tile arrays in duplicateWithkeys. Table 1 reports that SnugBox improves the efficiency of all downstream functions and produces an average overall speed-up of " + }, + { + "bbox": [ + 313, + 499, + 555, + 643 + ], + "type": "inline_equation", + "content": "1.82 \\times" + }, + { + "bbox": [ + 313, + 499, + 555, + 643 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 314, + 650, + 381, + 661 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 650, + 381, + 661 + ], + "spans": [ + { + "bbox": [ + 314, + 650, + 381, + 661 + ], + "type": "text", + "content": "4.1.2. AccuTile" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "type": "text", + "content": "Our AccuTile method, outlined in Algorithm 1, extends SnugBox to identify the exact tiles intersected by the Gaussian. It takes as input the tight bounding box produced by Snugbox and its rectangular tile extent - depicted as the" + } + ] + } + ], + "index": 29 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21540" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 69, + 75, + 221, + 217 + ], + "blocks": [ + { + "bbox": [ + 69, + 75, + 221, + 217 + ], + "lines": [ + { + "bbox": [ + 69, + 75, + 221, + 217 + ], + "spans": [ + { + "bbox": [ + 69, + 75, + 221, + 217 + ], + "type": "image", + "image_path": "498c1943b015373d3a9f7aa4619224253d212be3fdf54a4eeef91620a0d6558b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 102, + 221, + 187, + 232 + ], + "lines": [ + { + "bbox": [ + 102, + 221, + 187, + 232 + ], + "spans": [ + { + "bbox": [ + 102, + 221, + 187, + 232 + ], + "type": "text", + "content": "(a) 3D Gaussian Splitting" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 228, + 75, + 382, + 217 + ], + "blocks": [ + { + "bbox": [ + 228, + 75, + 382, + 217 + ], + "lines": [ + { + "bbox": [ + 228, + 75, + 382, + 217 + ], + "spans": [ + { + "bbox": [ + 228, + 75, + 382, + 217 + ], + "type": "image", + "image_path": "54a1fbdd743500bbb5da29510d8a0eec048bfc1ea78239c3a8b7a1127ab99138.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 284, + 221, + 326, + 232 + ], + "lines": [ + { + "bbox": [ + 284, + 221, + 326, + 232 + ], + "spans": [ + { + "bbox": [ + 284, + 221, + 326, + 232 + ], + "type": "text", + "content": "(b) SnugBox" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 54, + 240, + 555, + 307 + ], + "lines": [ + { + "bbox": [ + 54, + 240, + 555, + 307 + ], + "spans": [ + { + "bbox": [ + 54, + 240, + 555, + 307 + ], + "type": "text", + "content": "Figure 2. Gaussian tile allocation by method. (a) 3D Gaussian Splating allocates a Gaussian to a tile when that tile intersects the square inscribing the circle with radius " + }, + { + "bbox": [ + 54, + 240, + 555, + 307 + ], + "type": "inline_equation", + "content": "\\lceil 3\\sqrt{\\lambda_{\\max}}\\rceil" + }, + { + "bbox": [ + 54, + 240, + 555, + 307 + ], + "type": "text", + "content": " defined in Equation 8. (b) Our SnugBox method allocates a Gaussian to a tile when that tile intersects the tight bounding box defined by the axis-aligned minima and maxima of the ellipse given by Equation 14. (c) Our AccuTile method allocates a Gaussian to a tile only if that tile intersects the ellipse via Algorithm 1, which computes the minimum and maximum tiles by iterating over the shorter side of the rectangular tile extent given by SnugBox. In this example, our AccuTile algorithm iterates over the tile rows; the only points that are processed are " + }, + { + "bbox": [ + 54, + 240, + 555, + 307 + ], + "type": "inline_equation", + "content": "x_{min}" + }, + { + "bbox": [ + 54, + 240, + 555, + 307 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 54, + 240, + 555, + 307 + ], + "type": "inline_equation", + "content": "x_{max}" + }, + { + "bbox": [ + 54, + 240, + 555, + 307 + ], + "type": "text", + "content": ", A, B, C, and D." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 391, + 75, + 542, + 217 + ], + "blocks": [ + { + "bbox": [ + 391, + 75, + 542, + 217 + ], + "lines": [ + { + "bbox": [ + 391, + 75, + 542, + 217 + ], + "spans": [ + { + "bbox": [ + 391, + 75, + 542, + 217 + ], + "type": "image", + "image_path": "13cf56fb2296d2a06c9f0264d39d6b0ac0948a8d93ab199d86136919acaae2b4.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 445, + 221, + 487, + 231 + ], + "lines": [ + { + "bbox": [ + 445, + 221, + 487, + 231 + ], + "spans": [ + { + "bbox": [ + 445, + 221, + 487, + 231 + ], + "type": "text", + "content": "(c) AccuTile" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 323, + 296, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 323, + 296, + 418 + ], + "spans": [ + { + "bbox": [ + 54, + 323, + 296, + 418 + ], + "type": "text", + "content": "blue box and yellow tiles, respectively, in Figure 2b. Depending on which dimension is smaller, AccuTile then processes either the rows or columns of this tile extent to determine the exact tiles that intersect the Gaussian. Specifically, we identify the minimum and maximum extent of the ellipse within a given row or column, then convert those points to the corresponding touched tiles. All tiles between the minimum and maximum tiles intersect the ellipse." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 422, + 296, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 422, + 296, + 626 + ], + "spans": [ + { + "bbox": [ + 55, + 422, + 296, + 626 + ], + "type": "text", + "content": "AccuTile's key insight is that calculating the minimum and maximum extent of the ellipse within each row or column requires computing only two points per iteration. The only possible inflection points of the elliptical curve are the bounding box minimum and maximum points identified by Snugbox. If one or both of these points lie within the tile row or column, then they represent the minimum or maximum extent of the ellipse there. If neither point is within the row or column, then the points along that boundary side are monotonically decreasing or increasing – the minimum or maximum point must lie on one of the boundary lines of that row or column. Since the boundary of the last row or column is shared with the next one, we only need to compute the intersection of the ellipse with the next boundary line in each iteration. Thus, our AccuTile algorithm counts tiles in time proportional to the shorter side of the tile extent and processes tiles in time proportional to the tile count." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "text", + "content": "Figure 2 illustrates how Accutile restricts the tight bounding box produced by SnugBox to the exact tiles that the Gaussian touches. The tile extent rows from Snugbox, shown in Figure 2b, are processed starting from the bottom. The lower tile row boundary falls below the bounding box, so no initial intersection is computed. The upper boundary lies below the top of the bounding box, so intersection" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 323, + 555, + 466 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 323, + 555, + 466 + ], + "spans": [ + { + "bbox": [ + 313, + 323, + 555, + 466 + ], + "type": "text", + "content": "points " + }, + { + "bbox": [ + 313, + 323, + 555, + 466 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 313, + 323, + 555, + 466 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 323, + 555, + 466 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 313, + 323, + 555, + 466 + ], + "type": "text", + "content": " are calculated using Equation 15. " + }, + { + "bbox": [ + 313, + 323, + 555, + 466 + ], + "type": "inline_equation", + "content": "x_{\\mathrm{min}}" + }, + { + "bbox": [ + 313, + 323, + 555, + 466 + ], + "type": "text", + "content": " is assigned as this row's minimum ellipse extent " + }, + { + "bbox": [ + 313, + 323, + 555, + 466 + ], + "type": "inline_equation", + "content": "e_{\\mathrm{min}}" + }, + { + "bbox": [ + 313, + 323, + 555, + 466 + ], + "type": "text", + "content": " because it lies within it, and " + }, + { + "bbox": [ + 313, + 323, + 555, + 466 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 313, + 323, + 555, + 466 + ], + "type": "text", + "content": " is assigned as its maximum ellipse extent " + }, + { + "bbox": [ + 313, + 323, + 555, + 466 + ], + "type": "inline_equation", + "content": "e_{\\mathrm{max}}" + }, + { + "bbox": [ + 313, + 323, + 555, + 466 + ], + "type": "text", + "content": " because it is the maximal point. Consequently, this row's tile extent is from the tile containing " + }, + { + "bbox": [ + 313, + 323, + 555, + 466 + ], + "type": "inline_equation", + "content": "x_{\\mathrm{min}}" + }, + { + "bbox": [ + 313, + 323, + 555, + 466 + ], + "type": "text", + "content": " to the tile containing " + }, + { + "bbox": [ + 313, + 323, + 555, + 466 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 313, + 323, + 555, + 466 + ], + "type": "text", + "content": ". For the next row, we keep points " + }, + { + "bbox": [ + 313, + 323, + 555, + 466 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 313, + 323, + 555, + 466 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 323, + 555, + 466 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 313, + 323, + 555, + 466 + ], + "type": "text", + "content": " and compute upper boundary points " + }, + { + "bbox": [ + 313, + 323, + 555, + 466 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 313, + 323, + 555, + 466 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 323, + 555, + 466 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 313, + 323, + 555, + 466 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 313, + 323, + 555, + 466 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 313, + 323, + 555, + 466 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 323, + 555, + 466 + ], + "type": "inline_equation", + "content": "x_{\\mathrm{max}}" + }, + { + "bbox": [ + 313, + 323, + 555, + 466 + ], + "type": "text", + "content": " are assigned as the row's " + }, + { + "bbox": [ + 313, + 323, + 555, + 466 + ], + "type": "inline_equation", + "content": "e_{\\mathrm{min}}" + }, + { + "bbox": [ + 313, + 323, + 555, + 466 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 323, + 555, + 466 + ], + "type": "inline_equation", + "content": "e_{\\mathrm{max}}" + }, + { + "bbox": [ + 313, + 323, + 555, + 466 + ], + "type": "text", + "content": ", and the tiles containing them are its tile extent. Finally, for the last row, " + }, + { + "bbox": [ + 313, + 323, + 555, + 466 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 313, + 323, + 555, + 466 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 323, + 555, + 466 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 313, + 323, + 555, + 466 + ], + "type": "text", + "content": " are kept; no additional points are computed because the row's upper boundary is above the bounding box. Thus, the tile extent of this row is between " + }, + { + "bbox": [ + 313, + 323, + 555, + 466 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 313, + 323, + 555, + 466 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 323, + 555, + 466 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 313, + 323, + 555, + 466 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 467, + 555, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 467, + 555, + 528 + ], + "spans": [ + { + "bbox": [ + 313, + 467, + 555, + 528 + ], + "type": "text", + "content": "The number of tiles is further reduced from Figure 2b to Figure 2c, much less than the original 3D-GS method in Figure 2a. In Table 1, we report that AccuTile further accelerates all downstream functions, culminating in an average overall speed-up of " + }, + { + "bbox": [ + 313, + 467, + 555, + 528 + ], + "type": "inline_equation", + "content": "1.99 \\times" + }, + { + "bbox": [ + 313, + 467, + 555, + 528 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 537, + 419, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 537, + 419, + 550 + ], + "spans": [ + { + "bbox": [ + 313, + 537, + 419, + 550 + ], + "type": "text", + "content": "4.2. Efficient Pruning" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 555, + 554, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 555, + 554, + 604 + ], + "spans": [ + { + "bbox": [ + 313, + 555, + 554, + 604 + ], + "type": "text", + "content": "PUP 3D-GS [11] is a pruning method that quantifies the sensitivity of each Gaussian to training views, removing a set percentage with the lowest sensitivities. Sensitivity is computed by approximating the Hessian of the " + }, + { + "bbox": [ + 313, + 555, + 554, + 604 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 313, + 555, + 554, + 604 + ], + "type": "text", + "content": " loss:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 339, + 614, + 555, + 641 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 614, + 555, + 641 + ], + "spans": [ + { + "bbox": [ + 339, + 614, + 555, + 641 + ], + "type": "interline_equation", + "content": "H = \\nabla_ {\\mathcal {G}} ^ {2} L _ {2} = \\sum_ {\\phi \\in \\mathcal {P} _ {g t}} \\nabla_ {\\mathcal {G}} I _ {\\mathcal {G}} (\\phi) \\nabla_ {\\mathcal {G}} I _ {\\mathcal {G}} (\\phi) ^ {T}, \\tag {17}", + "image_path": "b4590467f5548170aff65812bfc2f557997fe032eef2f04763a3ac90edf17040.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 652, + 554, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 652, + 554, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 652, + 554, + 689 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 652, + 554, + 689 + ], + "type": "inline_equation", + "content": "\\nabla_{\\mathcal{G}}I_{\\mathcal{G}}(\\phi)" + }, + { + "bbox": [ + 313, + 652, + 554, + 689 + ], + "type": "text", + "content": " is the gradient over all Gaussian parameters on the rendered image " + }, + { + "bbox": [ + 313, + 652, + 554, + 689 + ], + "type": "inline_equation", + "content": "I_{\\mathcal{G}}" + }, + { + "bbox": [ + 313, + 652, + 554, + 689 + ], + "type": "text", + "content": " for pose " + }, + { + "bbox": [ + 313, + 652, + 554, + 689 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 313, + 652, + 554, + 689 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 313, + 652, + 554, + 689 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 313, + 652, + 554, + 689 + ], + "type": "text", + "content": " is shown to be exact when the " + }, + { + "bbox": [ + 313, + 652, + 554, + 689 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 313, + 652, + 554, + 689 + ], + "type": "text", + "content": " residual error vanishes [11]." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "content": "A per-Gaussian sensitivity can be derived by splitting " + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "content": " into the block diagonal elements that only capture inter" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21541" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "content": "Algorithm 1 The AccuTile Algorithm. For simplicity, the algorithm outlined here is applied to the rows of the Snug-Box tile extent bounding box, matching the example in Figure 2c. In practice, it is applied along the smaller side of the tile extent. The subscripts " + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "content": " represent the top, bottom, left, and right sides, respectively. A proof of correctness sketch is presented in Appendix A.1." + } + ] + } + ], + "index": 0 + }, + { + "type": "code", + "bbox": [ + 55, + 164, + 296, + 460 + ], + "blocks": [ + { + "bbox": [ + 55, + 164, + 296, + 460 + ], + "lines": [ + { + "bbox": [ + 55, + 164, + 296, + 460 + ], + "spans": [ + { + "bbox": [ + 55, + 164, + 296, + 460 + ], + "type": "text", + "content": "Require: Ellipse E " + }, + { + "bbox": [ + 55, + 164, + 296, + 460 + ], + "type": "inline_equation", + "content": "\\triangleright" + }, + { + "bbox": [ + 55, + 164, + 296, + 460 + ], + "type": "text", + "content": " Eq. 14 \nRequire: SnugBox Bounding Box B \nRequire: SnugBox Tile Extent Rectangle R \nInitialize: Tile count " + }, + { + "bbox": [ + 55, + 164, + 296, + 460 + ], + "type": "inline_equation", + "content": "C\\gets 0" + }, + { + "bbox": [ + 55, + 164, + 296, + 460 + ], + "type": "text", + "content": " \nline " + }, + { + "bbox": [ + 55, + 164, + 296, + 460 + ], + "type": "inline_equation", + "content": "\\mathrm{min}\\gets \\mathbf{R}_b" + }, + { + "bbox": [ + 55, + 164, + 296, + 460 + ], + "type": "text", + "content": " \nif line " + }, + { + "bbox": [ + 55, + 164, + 296, + 460 + ], + "type": "inline_equation", + "content": "\\mathrm{min}\\geq \\mathbf{B}_b" + }, + { + "bbox": [ + 55, + 164, + 296, + 460 + ], + "type": "text", + "content": " then \n" + }, + { + "bbox": [ + 55, + 164, + 296, + 460 + ], + "type": "inline_equation", + "content": "\\mathbf{i}_{\\mathrm{min}}\\gets" + }, + { + "bbox": [ + 55, + 164, + 296, + 460 + ], + "type": "text", + "content": " Intersections( line " + }, + { + "bbox": [ + 55, + 164, + 296, + 460 + ], + "type": "inline_equation", + "content": "\\mathrm{min}" + }, + { + "bbox": [ + 55, + 164, + 296, + 460 + ], + "type": "text", + "content": " ,E) " + }, + { + "bbox": [ + 55, + 164, + 296, + 460 + ], + "type": "inline_equation", + "content": "\\triangleright" + }, + { + "bbox": [ + 55, + 164, + 296, + 460 + ], + "type": "text", + "content": " Eq. 15 \nend if \nfor row r in R do \nline " + }, + { + "bbox": [ + 55, + 164, + 296, + 460 + ], + "type": "inline_equation", + "content": "\\mathrm{max}\\gets \\mathrm{r}_t" + }, + { + "bbox": [ + 55, + 164, + 296, + 460 + ], + "type": "text", + "content": " \nif line " + }, + { + "bbox": [ + 55, + 164, + 296, + 460 + ], + "type": "inline_equation", + "content": "\\mathrm{max}\\leq \\mathbf{B}_t" + }, + { + "bbox": [ + 55, + 164, + 296, + 460 + ], + "type": "text", + "content": " then \n" + }, + { + "bbox": [ + 55, + 164, + 296, + 460 + ], + "type": "inline_equation", + "content": "\\mathbf{i}_{\\mathrm{max}}\\gets" + }, + { + "bbox": [ + 55, + 164, + 296, + 460 + ], + "type": "text", + "content": " Intersections( line " + }, + { + "bbox": [ + 55, + 164, + 296, + 460 + ], + "type": "inline_equation", + "content": "\\mathrm{max}" + }, + { + "bbox": [ + 55, + 164, + 296, + 460 + ], + "type": "text", + "content": " ,E) " + }, + { + "bbox": [ + 55, + 164, + 296, + 460 + ], + "type": "inline_equation", + "content": "\\triangleright" + }, + { + "bbox": [ + 55, + 164, + 296, + 460 + ], + "type": "text", + "content": " Eq. 15 \nend if \ne " + }, + { + "bbox": [ + 55, + 164, + 296, + 460 + ], + "type": "inline_equation", + "content": "\\mathrm{min}\\gets \\mathbf{B}_l" + }, + { + "bbox": [ + 55, + 164, + 296, + 460 + ], + "type": "text", + "content": " if Bl in r else min(i " + }, + { + "bbox": [ + 55, + 164, + 296, + 460 + ], + "type": "inline_equation", + "content": "\\mathrm{min}" + }, + { + "bbox": [ + 55, + 164, + 296, + 460 + ], + "type": "text", + "content": " ,i max) \ne " + }, + { + "bbox": [ + 55, + 164, + 296, + 460 + ], + "type": "inline_equation", + "content": "\\mathrm{max}\\gets" + }, + { + "bbox": [ + 55, + 164, + 296, + 460 + ], + "type": "text", + "content": " Br if Br in r else max(i " + }, + { + "bbox": [ + 55, + 164, + 296, + 460 + ], + "type": "inline_equation", + "content": "\\mathrm{min}" + }, + { + "bbox": [ + 55, + 164, + 296, + 460 + ], + "type": "text", + "content": " ,i max) \ntile " + }, + { + "bbox": [ + 55, + 164, + 296, + 460 + ], + "type": "inline_equation", + "content": "\\mathrm{min}" + }, + { + "bbox": [ + 55, + 164, + 296, + 460 + ], + "type": "text", + "content": " ,tile " + }, + { + "bbox": [ + 55, + 164, + 296, + 460 + ], + "type": "inline_equation", + "content": "\\mathrm{max}\\gets" + }, + { + "bbox": [ + 55, + 164, + 296, + 460 + ], + "type": "text", + "content": " Convert(e " + }, + { + "bbox": [ + 55, + 164, + 296, + 460 + ], + "type": "inline_equation", + "content": "\\mathrm{min}" + }, + { + "bbox": [ + 55, + 164, + 296, + 460 + ], + "type": "text", + "content": " ,e max) \nC " + }, + { + "bbox": [ + 55, + 164, + 296, + 460 + ], + "type": "inline_equation", + "content": "\\leftarrow" + }, + { + "bbox": [ + 55, + 164, + 296, + 460 + ], + "type": "text", + "content": " C + (tile " + }, + { + "bbox": [ + 55, + 164, + 296, + 460 + ], + "type": "inline_equation", + "content": "\\mathrm{max}" + }, + { + "bbox": [ + 55, + 164, + 296, + 460 + ], + "type": "text", + "content": " -tile " + }, + { + "bbox": [ + 55, + 164, + 296, + 460 + ], + "type": "inline_equation", + "content": "\\mathrm{min}" + }, + { + "bbox": [ + 55, + 164, + 296, + 460 + ], + "type": "text", + "content": " ) \nProcess( tile " + }, + { + "bbox": [ + 55, + 164, + 296, + 460 + ], + "type": "inline_equation", + "content": "\\mathrm{min}" + }, + { + "bbox": [ + 55, + 164, + 296, + 460 + ], + "type": "text", + "content": " ,tile " + }, + { + "bbox": [ + 55, + 164, + 296, + 460 + ], + "type": "inline_equation", + "content": "\\mathrm{max}" + }, + { + "bbox": [ + 55, + 164, + 296, + 460 + ], + "type": "text", + "content": " ) \n" + }, + { + "bbox": [ + 55, + 164, + 296, + 460 + ], + "type": "inline_equation", + "content": "\\mathbf{i}_{\\mathrm{min}}\\gets \\mathbf{i}_{\\mathrm{max}}" + }, + { + "bbox": [ + 55, + 164, + 296, + 460 + ], + "type": "text", + "content": " \nend for \nreturn C" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "code_body" + } + ], + "index": 1, + "sub_type": "algorithm" + }, + { + "bbox": [ + 55, + 487, + 194, + 500 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 487, + 194, + 500 + ], + "spans": [ + { + "bbox": [ + 55, + 487, + 194, + 500 + ], + "type": "text", + "content": "Gaussian parameter relationships:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 104, + 506, + 295, + 534 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 506, + 295, + 534 + ], + "spans": [ + { + "bbox": [ + 104, + 506, + 295, + 534 + ], + "type": "interline_equation", + "content": "H _ {i} = \\sum_ {\\phi \\in \\mathcal {P} _ {g t}} \\nabla_ {\\mathcal {G} _ {i}} I _ {\\mathcal {G}} (\\phi) \\nabla_ {\\mathcal {G} _ {i}} I _ {\\mathcal {G}} (\\phi) ^ {T}, \\tag {18}", + "image_path": "1768d4aed0edae1c394c97ee10c1762c58b4fce928204ebd1d554ee05bf78bd7.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 540, + 295, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 540, + 295, + 577 + ], + "spans": [ + { + "bbox": [ + 55, + 540, + 295, + 577 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 540, + 295, + 577 + ], + "type": "inline_equation", + "content": "\\nabla_{\\mathcal{G}_i}" + }, + { + "bbox": [ + 55, + 540, + 295, + 577 + ], + "type": "text", + "content": " is the gradient with respect to only " + }, + { + "bbox": [ + 55, + 540, + 295, + 577 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_i" + }, + { + "bbox": [ + 55, + 540, + 295, + 577 + ], + "type": "text", + "content": ". This measures the sensitivity of the " + }, + { + "bbox": [ + 55, + 540, + 295, + 577 + ], + "type": "inline_equation", + "content": "L_2" + }, + { + "bbox": [ + 55, + 540, + 295, + 577 + ], + "type": "text", + "content": " loss with respect to Gaussian " + }, + { + "bbox": [ + 55, + 540, + 295, + 577 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_i" + }, + { + "bbox": [ + 55, + 540, + 295, + 577 + ], + "type": "text", + "content": ", assuming all other Gaussians are held constant." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 577, + 296, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 577, + 296, + 625 + ], + "spans": [ + { + "bbox": [ + 55, + 577, + 296, + 625 + ], + "type": "inline_equation", + "content": "H_{i}" + }, + { + "bbox": [ + 55, + 577, + 296, + 625 + ], + "type": "text", + "content": " is again approximated by only using the six mean " + }, + { + "bbox": [ + 55, + 577, + 296, + 625 + ], + "type": "inline_equation", + "content": "\\mu_{i}" + }, + { + "bbox": [ + 55, + 577, + 296, + 625 + ], + "type": "text", + "content": " and scale " + }, + { + "bbox": [ + 55, + 577, + 296, + 625 + ], + "type": "inline_equation", + "content": "s_i" + }, + { + "bbox": [ + 55, + 577, + 296, + 625 + ], + "type": "text", + "content": " parameters to specifically capture geometric sensitivity. The log determinant is taken to provide a representative scalar score " + }, + { + "bbox": [ + 55, + 577, + 296, + 625 + ], + "type": "inline_equation", + "content": "U_{i}" + }, + { + "bbox": [ + 55, + 577, + 296, + 625 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 113, + 632, + 295, + 647 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 632, + 295, + 647 + ], + "spans": [ + { + "bbox": [ + 113, + 632, + 295, + 647 + ], + "type": "interline_equation", + "content": "U _ {i} = \\log | \\nabla_ {\\mu_ {i}, s _ {i}} I _ {\\mathcal {G}} \\nabla_ {\\mu_ {i}, s _ {i}} I _ {\\mathcal {G}} ^ {T} |. \\tag {19}", + "image_path": "23250ace783439785b6a15129be175dac115c26625584f1bf17c9d16205d177a.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 653, + 295, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 295, + 677 + ], + "type": "text", + "content": "Using this score, up to " + }, + { + "bbox": [ + 55, + 653, + 295, + 677 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 55, + 653, + 295, + 677 + ], + "type": "text", + "content": " of Gaussians can be robustly pruned from the model while retaining high visual quality." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": "Although PUP 3D-GS touts high compression ratios and rendering speeds, we identify two key drawbacks in its formulation. First, computing the Hessian requires storage" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "content": "proportional to " + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "inline_equation", + "content": "N \\times 36" + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "content": " is the number of Gaussians. In comparison, the 3D-GS model has a memory footprint proportional to " + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "inline_equation", + "content": "N \\times 59" + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "content": " because it stores 59 parameters per Gaussian. While this score is effective for post-hoc pruning, using it during training is impractical." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 133, + 555, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 133, + 555, + 251 + ], + "spans": [ + { + "bbox": [ + 313, + 133, + 555, + 251 + ], + "type": "text", + "content": "Second, computing the Hessian requires the pixel-wise gradients of " + }, + { + "bbox": [ + 313, + 133, + 555, + 251 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 313, + 133, + 555, + 251 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 133, + 555, + 251 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 313, + 133, + 555, + 251 + ], + "type": "text", + "content": ". Since these are 3D parameters of the Gaussian primitives, obtaining their gradients requires back-propagating through the render kernel parallelized per pixel, then back-propagating each Gaussian contributing to that pixel in its thread. This breaks the efficient flow of gradients in 3D-GS, where the per-pixel gradients from the render kernel are parallelized and aggregated to the 2D " + }, + { + "bbox": [ + 313, + 133, + 555, + 251 + ], + "type": "inline_equation", + "content": "\\mu_{2D}" + }, + { + "bbox": [ + 313, + 133, + 555, + 251 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 133, + 555, + 251 + ], + "type": "inline_equation", + "content": "\\Sigma_{2D}" + }, + { + "bbox": [ + 313, + 133, + 555, + 251 + ], + "type": "text", + "content": " parameters, which are then parallelized across Gaussians to compute gradients for " + }, + { + "bbox": [ + 313, + 133, + 555, + 251 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 313, + 133, + 555, + 251 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 133, + 555, + 251 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 313, + 133, + 555, + 251 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 251, + 554, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 251, + 554, + 335 + ], + "spans": [ + { + "bbox": [ + 313, + 251, + 554, + 335 + ], + "type": "text", + "content": "Our approach builds on PUP 3D-GS by introducing an efficient pruning score that we incorporate into the 3D-GS training pipeline. We also define two distinct pruning modalities: Soft Pruning, which takes place during densification in the first 15000 iterations, and Hard Pruning, which is applied after densification is completed after iteration 15000." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 340, + 443, + 352 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 340, + 443, + 352 + ], + "spans": [ + { + "bbox": [ + 313, + 340, + 443, + 352 + ], + "type": "text", + "content": "4.2.1. Efficient Pruning Score" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 354, + 554, + 438 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 354, + 554, + 438 + ], + "spans": [ + { + "bbox": [ + 313, + 354, + 554, + 438 + ], + "type": "text", + "content": "Our insight is that both drawbacks can be alleviated by reparameterizing the Hessian. Concretely, we express the influence of all spatial parameters of Gaussian " + }, + { + "bbox": [ + 313, + 354, + 554, + 438 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_i" + }, + { + "bbox": [ + 313, + 354, + 554, + 438 + ], + "type": "text", + "content": " by computing the Hessian approximation with respect to the 2D projected value of " + }, + { + "bbox": [ + 313, + 354, + 554, + 438 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_i" + }, + { + "bbox": [ + 313, + 354, + 554, + 438 + ], + "type": "text", + "content": " at pixel " + }, + { + "bbox": [ + 313, + 354, + 554, + 438 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 313, + 354, + 554, + 438 + ], + "type": "text", + "content": ", given by " + }, + { + "bbox": [ + 313, + 354, + 554, + 438 + ], + "type": "inline_equation", + "content": "g_i(p)" + }, + { + "bbox": [ + 313, + 354, + 554, + 438 + ], + "type": "text", + "content": " in Equation 6. In doing so, the pruning score " + }, + { + "bbox": [ + 313, + 354, + 554, + 438 + ], + "type": "inline_equation", + "content": "U_i" + }, + { + "bbox": [ + 313, + 354, + 554, + 438 + ], + "type": "text", + "content": " from Equation 19 can be expressed as:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 382, + 449, + 553, + 464 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 382, + 449, + 553, + 464 + ], + "spans": [ + { + "bbox": [ + 382, + 449, + 553, + 464 + ], + "type": "interline_equation", + "content": "\\tilde {U} _ {i} = \\log | \\nabla_ {g _ {i}} I _ {\\mathcal {G}} \\nabla_ {g _ {i}} I _ {\\mathcal {G}} ^ {T} |. \\tag {20}", + "image_path": "ab59006ceac50333b85eb4588980b6aa81cfd0a7a3060116deeb8fd3d17ca579.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 466, + 554, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 466, + 554, + 489 + ], + "spans": [ + { + "bbox": [ + 313, + 466, + 554, + 489 + ], + "type": "text", + "content": "Since " + }, + { + "bbox": [ + 313, + 466, + 554, + 489 + ], + "type": "inline_equation", + "content": "g_{i}" + }, + { + "bbox": [ + 313, + 466, + 554, + 489 + ], + "type": "text", + "content": " is a scalar and log is monotonically increasing, we can rewrite this score as:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 400, + 495, + 553, + 510 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 400, + 495, + 553, + 510 + ], + "spans": [ + { + "bbox": [ + 400, + 495, + 553, + 510 + ], + "type": "interline_equation", + "content": "\\tilde {U} _ {i} = \\left(\\nabla_ {g _ {i}} I _ {\\mathcal {G}}\\right) ^ {2}. \\tag {21}", + "image_path": "d979e5e3052db40e5c0083fed4ababde524cf8948e237e5b8bd9221e7961e8b8.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 514, + 555, + 587 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 514, + 555, + 587 + ], + "spans": [ + { + "bbox": [ + 313, + 514, + 555, + 587 + ], + "type": "text", + "content": "Gradient " + }, + { + "bbox": [ + 313, + 514, + 555, + 587 + ], + "type": "inline_equation", + "content": "\\nabla_{g_i}I_{\\mathcal{G}}" + }, + { + "bbox": [ + 313, + 514, + 555, + 587 + ], + "type": "text", + "content": " is already computed in the backward pass of render and can be efficiently squared and aggregated across all pixels. Moreover, the maximum space requirement for this score is proportional to the number of Gaussians " + }, + { + "bbox": [ + 313, + 514, + 555, + 587 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 514, + 555, + 587 + ], + "type": "text", + "content": ", reducing the storage requirement by " + }, + { + "bbox": [ + 313, + 514, + 555, + 587 + ], + "type": "inline_equation", + "content": "36\\times" + }, + { + "bbox": [ + 313, + 514, + 555, + 587 + ], + "type": "text", + "content": " and allowing this score to be used during training." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 591, + 398, + 603 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 591, + 398, + 603 + ], + "spans": [ + { + "bbox": [ + 313, + 591, + 398, + 603 + ], + "type": "text", + "content": "4.2.2. Soft Pruning" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "text", + "content": "To maintain a robust Hessian approximation, we observe that the " + }, + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "text", + "content": " loss becomes quite small by iteration 6000 and remains low except after an opacity reset is performed. As such, we augment the densification pipeline to include our Soft Pruning method, where the model is pruned immediately before the three opacity resets at 6000, 9000, and 12000 iterations. Surprisingly, we find that we can set extremely high Soft Pruning ratios – in our experiments, visual fidelity is preserved at " + }, + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "inline_equation", + "content": "80\\%" + }, + { + "bbox": [ + 313, + 605, + 555, + 714 + ], + "type": "text", + "content": " pruning." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "21542" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 86, + 173, + 162 + ], + "blocks": [ + { + "bbox": [ + 107, + 72, + 124, + 83 + ], + "lines": [ + { + "bbox": [ + 107, + 72, + 124, + 83 + ], + "spans": [ + { + "bbox": [ + 107, + 72, + 124, + 83 + ], + "type": "text", + "content": "GT" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 59, + 86, + 173, + 162 + ], + "lines": [ + { + "bbox": [ + 59, + 86, + 173, + 162 + ], + "spans": [ + { + "bbox": [ + 59, + 86, + 173, + 162 + ], + "type": "image", + "image_path": "ffa852edb8b496f4992351b54752d23519b28289c3ede99ec305b29d3979ba23.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 187, + 86, + 301, + 161 + ], + "blocks": [ + { + "bbox": [ + 226, + 72, + 260, + 83 + ], + "lines": [ + { + "bbox": [ + 226, + 72, + 260, + 83 + ], + "spans": [ + { + "bbox": [ + 226, + 72, + 260, + 83 + ], + "type": "text", + "content": "3D-GS" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 187, + 86, + 301, + 161 + ], + "lines": [ + { + "bbox": [ + 187, + 86, + 301, + 161 + ], + "spans": [ + { + "bbox": [ + 187, + 86, + 301, + 161 + ], + "type": "image", + "image_path": "ab1bba5ffc9241fd393b849b78777938923af29863d8da268f51a99e0ac94bff.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 190, + 161, + 249, + 171 + ], + "lines": [ + { + "bbox": [ + 190, + 161, + 249, + 171 + ], + "spans": [ + { + "bbox": [ + 190, + 161, + 249, + 171 + ], + "type": "text", + "content": "2.5M Gaussians" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 251, + 161, + 293, + 170 + ], + "lines": [ + { + "bbox": [ + 251, + 161, + 293, + 170 + ], + "spans": [ + { + "bbox": [ + 251, + 161, + 293, + 170 + ], + "type": "text", + "content": "172.15 FPS" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 306, + 86, + 419, + 161 + ], + "blocks": [ + { + "bbox": [ + 353, + 72, + 375, + 84 + ], + "lines": [ + { + "bbox": [ + 353, + 72, + 375, + 84 + ], + "spans": [ + { + "bbox": [ + 353, + 72, + 375, + 84 + ], + "type": "text", + "content": "PUP" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 306, + 86, + 419, + 161 + ], + "lines": [ + { + "bbox": [ + 306, + 86, + 419, + 161 + ], + "spans": [ + { + "bbox": [ + 306, + 86, + 419, + 161 + ], + "type": "image", + "image_path": "7b86d26216953725c49b5f1abe752c107466b4e116a939948888647730f80a15.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 310, + 161, + 370, + 171 + ], + "lines": [ + { + "bbox": [ + 310, + 161, + 370, + 171 + ], + "spans": [ + { + "bbox": [ + 310, + 161, + 370, + 171 + ], + "type": "text", + "content": "0.25M Gaussians" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 427, + 72, + 545, + 161 + ], + "blocks": [ + { + "bbox": [ + 427, + 72, + 545, + 161 + ], + "lines": [ + { + "bbox": [ + 427, + 72, + 545, + 161 + ], + "spans": [ + { + "bbox": [ + 427, + 72, + 545, + 161 + ], + "type": "image", + "image_path": "6d9b80c08e645f24d151969b745cc25e543f78c066e072c94ad7237210beb122.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 432, + 161, + 492, + 171 + ], + "lines": [ + { + "bbox": [ + 432, + 161, + 492, + 171 + ], + "spans": [ + { + "bbox": [ + 432, + 161, + 492, + 171 + ], + "type": "text", + "content": "0.22M Gaussians" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 497, + 161, + 541, + 170 + ], + "lines": [ + { + "bbox": [ + 497, + 161, + 541, + 170 + ], + "spans": [ + { + "bbox": [ + 497, + 161, + 541, + 170 + ], + "type": "text", + "content": "1277.20 FPS" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 59, + 177, + 173, + 254 + ], + "blocks": [ + { + "bbox": [ + 59, + 177, + 173, + 254 + ], + "lines": [ + { + "bbox": [ + 59, + 177, + 173, + 254 + ], + "spans": [ + { + "bbox": [ + 59, + 177, + 173, + 254 + ], + "type": "image", + "image_path": "cbe29eccb3828a4900ea58195c9e98fefb2d7b2212762b7e7fe9e21bc15f88d7.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 187, + 177, + 301, + 253 + ], + "blocks": [ + { + "bbox": [ + 187, + 177, + 301, + 253 + ], + "lines": [ + { + "bbox": [ + 187, + 177, + 301, + 253 + ], + "spans": [ + { + "bbox": [ + 187, + 177, + 301, + 253 + ], + "type": "image", + "image_path": "a69a2b1aad2b45f92c9a61c3e42068077ae5f9fbe80c3678fa2a3821d4f83abb.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 190, + 254, + 249, + 262 + ], + "lines": [ + { + "bbox": [ + 190, + 254, + 249, + 262 + ], + "spans": [ + { + "bbox": [ + 190, + 254, + 249, + 262 + ], + "type": "text", + "content": "4.5M Gaussians" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 251, + 254, + 293, + 262 + ], + "lines": [ + { + "bbox": [ + 251, + 254, + 293, + 262 + ], + "spans": [ + { + "bbox": [ + 251, + 254, + 293, + 262 + ], + "type": "text", + "content": "141.47 FPS" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 306, + 177, + 419, + 253 + ], + "blocks": [ + { + "bbox": [ + 306, + 177, + 419, + 253 + ], + "lines": [ + { + "bbox": [ + 306, + 177, + 419, + 253 + ], + "spans": [ + { + "bbox": [ + 306, + 177, + 419, + 253 + ], + "type": "image", + "image_path": "ad195892fca34ea72da077a4a4f19021307eedbdbd2ef40ff1c190e191dec090.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 310, + 254, + 370, + 262 + ], + "lines": [ + { + "bbox": [ + 310, + 254, + 370, + 262 + ], + "spans": [ + { + "bbox": [ + 310, + 254, + 370, + 262 + ], + "type": "text", + "content": "0.45M Gaussians" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 374, + 254, + 416, + 262 + ], + "lines": [ + { + "bbox": [ + 374, + 254, + 416, + 262 + ], + "spans": [ + { + "bbox": [ + 374, + 254, + 416, + 262 + ], + "type": "text", + "content": "261.67 FPS" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 428, + 177, + 543, + 253 + ], + "blocks": [ + { + "bbox": [ + 428, + 177, + 543, + 253 + ], + "lines": [ + { + "bbox": [ + 428, + 177, + 543, + 253 + ], + "spans": [ + { + "bbox": [ + 428, + 177, + 543, + 253 + ], + "type": "image", + "image_path": "c71c96ee3ad9661c7f598d20039af9416d6e007f20505ed785a468ca4f3cb411.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 432, + 254, + 492, + 262 + ], + "lines": [ + { + "bbox": [ + 432, + 254, + 492, + 262 + ], + "spans": [ + { + "bbox": [ + 432, + 254, + 492, + 262 + ], + "type": "text", + "content": "0.46M Gaussians" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 497, + 254, + 539, + 262 + ], + "lines": [ + { + "bbox": [ + 497, + 254, + 539, + 262 + ], + "spans": [ + { + "bbox": [ + 497, + 254, + 539, + 262 + ], + "type": "text", + "content": "723.79 FPS" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 59, + 269, + 174, + 345 + ], + "blocks": [ + { + "bbox": [ + 59, + 269, + 174, + 345 + ], + "lines": [ + { + "bbox": [ + 59, + 269, + 174, + 345 + ], + "spans": [ + { + "bbox": [ + 59, + 269, + 174, + 345 + ], + "type": "image", + "image_path": "74135626c2d30c63985b5840546ec5c0b9e0e6365fe9b37eba0069f9ec92a466.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 363, + 555, + 397 + ], + "lines": [ + { + "bbox": [ + 55, + 363, + 555, + 397 + ], + "spans": [ + { + "bbox": [ + 55, + 363, + 555, + 397 + ], + "type": "text", + "content": "Figure 3. Visual comparison of 3D-GS, PUP 3D-GS, and our method. Notice that, while reaching similar compression ratios to PUP 3D-GS, our Speedy-Splat method delivers vastly faster rendering speeds. Top: playroom from the Deep Blending dataset. Middle: bicycle from the Mip-NeRF 360 dataset. Bottom: drjohnson from the Deep Blending dataset." + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 187, + 270, + 301, + 344 + ], + "blocks": [ + { + "bbox": [ + 187, + 270, + 301, + 344 + ], + "lines": [ + { + "bbox": [ + 187, + 270, + 301, + 344 + ], + "spans": [ + { + "bbox": [ + 187, + 270, + 301, + 344 + ], + "type": "image", + "image_path": "7112c46269c14c0a67c0488d1b942f3961a5977bd336a434a2abe37d0756173f.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 190, + 345, + 249, + 354 + ], + "lines": [ + { + "bbox": [ + 190, + 345, + 249, + 354 + ], + "spans": [ + { + "bbox": [ + 190, + 345, + 249, + 354 + ], + "type": "text", + "content": "3.4M Gaussians" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 251, + 345, + 293, + 354 + ], + "lines": [ + { + "bbox": [ + 251, + 345, + 293, + 354 + ], + "spans": [ + { + "bbox": [ + 251, + 345, + 293, + 354 + ], + "type": "text", + "content": "125.97 FPS" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 306, + 270, + 419, + 345 + ], + "blocks": [ + { + "bbox": [ + 306, + 270, + 419, + 345 + ], + "lines": [ + { + "bbox": [ + 306, + 270, + 419, + 345 + ], + "spans": [ + { + "bbox": [ + 306, + 270, + 419, + 345 + ], + "type": "image", + "image_path": "4a705b7976af512de25292283ff91eacf5f38e44ef4ca3a5da8b1f75a549a2c3.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 310, + 345, + 370, + 354 + ], + "lines": [ + { + "bbox": [ + 310, + 345, + 370, + 354 + ], + "spans": [ + { + "bbox": [ + 310, + 345, + 370, + 354 + ], + "type": "text", + "content": "0.34M Gaussians" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 374, + 345, + 415, + 354 + ], + "lines": [ + { + "bbox": [ + 374, + 345, + 415, + 354 + ], + "spans": [ + { + "bbox": [ + 374, + 345, + 415, + 354 + ], + "type": "text", + "content": "413.30 FPS" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 428, + 270, + 544, + 345 + ], + "blocks": [ + { + "bbox": [ + 428, + 270, + 544, + 345 + ], + "lines": [ + { + "bbox": [ + 428, + 270, + 544, + 345 + ], + "spans": [ + { + "bbox": [ + 428, + 270, + 544, + 345 + ], + "type": "image", + "image_path": "6b4a83cf5c4ed81986b2c0dd4ed0c0bd3617f27a78f1d30868f7659030603af2.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 432, + 345, + 492, + 354 + ], + "lines": [ + { + "bbox": [ + 432, + 345, + 492, + 354 + ], + "spans": [ + { + "bbox": [ + 432, + 345, + 492, + 354 + ], + "type": "text", + "content": "0.32M Gaussians" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 497, + 345, + 539, + 354 + ], + "lines": [ + { + "bbox": [ + 497, + 345, + 539, + 354 + ], + "spans": [ + { + "bbox": [ + 497, + 345, + 539, + 354 + ], + "type": "text", + "content": "1121.97 FPS" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + } + ], + "index": 29 + }, + { + "bbox": [ + 55, + 412, + 144, + 425 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 412, + 144, + 425 + ], + "spans": [ + { + "bbox": [ + 55, + 412, + 144, + 425 + ], + "type": "text", + "content": "4.2.3. Hard Pruning" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 55, + 427, + 296, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 427, + 296, + 547 + ], + "spans": [ + { + "bbox": [ + 55, + 427, + 296, + 547 + ], + "type": "text", + "content": "We also observe that the model's performance after densification closely matches that of the fully-trained model. The iterations after the densification stage essentially finetune the model and can be used to further \"refine\" it after pruning, similar to PUP 3D-GS [11] and LightGaussian [6]. In practice, our Hard Pruning method prunes the model by a constant ratio every 3000 iterations starting at iteration 15000. We Hard Prune " + }, + { + "bbox": [ + 55, + 427, + 296, + 547 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 55, + 427, + 296, + 547 + ], + "type": "text", + "content": " of Gaussians in each interval, which, when paired with Soft Pruning, empirically reduces the total number of Gaussians across scenes by " + }, + { + "bbox": [ + 55, + 427, + 296, + 547 + ], + "type": "inline_equation", + "content": "10.6\\times" + }, + { + "bbox": [ + 55, + 427, + 296, + 547 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 55, + 556, + 135, + 570 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 556, + 135, + 570 + ], + "spans": [ + { + "bbox": [ + 55, + 556, + 135, + 570 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 55, + 576, + 117, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 576, + 117, + 588 + ], + "spans": [ + { + "bbox": [ + 55, + 576, + 117, + 588 + ], + "type": "text", + "content": "5.1. Datasets" + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "content": "Our evaluation uses the same set of challenging real-world scenes as 3D-GS [14]. This includes nine Mip-Nerf 360 scenes [3] – four indoor and five outdoor – that each feature a complex central object or area with a detailed background. We also include the outdoor train and truck scenes from the Tanks & Temples dataset [17] and the indoor drjohnson and playroom scenes from the Deep Blending dataset [12]. For consistency across experiments, we use the COLMAP pose estimates and sparse point clouds provided by the dataset authors." + } + ] + } + ], + "index": 37 + }, + { + "type": "table", + "bbox": [ + 317, + 483, + 550, + 555 + ], + "blocks": [ + { + "bbox": [ + 313, + 411, + 555, + 476 + ], + "lines": [ + { + "bbox": [ + 313, + 411, + 555, + 476 + ], + "spans": [ + { + "bbox": [ + 313, + 411, + 555, + 476 + ], + "type": "text", + "content": "Table 2. Average Gaussian count, FPS, and training time across all scenes in Section 5.1. Ratios for model size compression, rendering speed-up, and training speed-up are reported in (parentheses). The operation in each row is applied cumulatively to all of the following rows. The best and second best value for each metric are color coded." + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 483, + 550, + 555 + ], + "lines": [ + { + "bbox": [ + 317, + 483, + 550, + 555 + ], + "spans": [ + { + "bbox": [ + 317, + 483, + 550, + 555 + ], + "type": "table", + "html": "
Method# Gaussians ↓FPS ↑Training Time ↓
Baseline2.93M13423.2
+SnugBox2.97M244 (1.82×)21.2 (1.09×)
+AccuTile2.97M267 (1.99×)21.0 (1.10×)
+Soft Pruning1.64M (1.79×)420 (3.14×)17.5 (1.32×)
+Hard Pruning0.28M (10.6×)898 (6.71×)15.7 (1.47×)
", + "image_path": "c1deb9d05335c70f11526d2b0757f1aa3b019cfb682981339b047659350e9848.jpg" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "table_body" + } + ], + "index": 39 + }, + { + "bbox": [ + 314, + 581, + 446, + 594 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 581, + 446, + 594 + ], + "spans": [ + { + "bbox": [ + 314, + 581, + 446, + 594 + ], + "type": "text", + "content": "5.2. Implementation Details" + } + ] + } + ], + "index": 40 + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "content": "Our code builds on the differentiable renderer provided by 3D-GS [14] and modifies the Python training pipeline for pruning schedules and execution. To ensure consistent and precise timing, all times in Table 1 and FPS values in Tables 2 and 3 are measured using CUDA events at the start and end of the forward rendering procedure. All experiments are conducted on an Nvidia RTXA5000 GPU, and the reported Speedy-Splat results represent the average metrics across three independent runs for each scene." + } + ] + } + ], + "index": 41 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21543" + } + ] + } + ], + "index": 42 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 57, + 189, + 293, + 330 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 294, + 180 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 294, + 180 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 294, + 180 + ], + "type": "text", + "content": "Table 3. Average reported metrics for each pruning method across all scenes in the Mip-NeRF 360 dataset. The Comp column reports model size compression in terms of Gaussian count, FPS reports rendering speed-up, and Train reports training time speed-up, all with respect to the baseline 3D-GS model. PSNR, SSIM, and LPIPS are also recorded. For a fair comparison, we report the published results of each method and use '-' to denote missing metrics. The best and second best value for each metric are color coded; lossless methods are underlined. Results for the Tanks & Temples and Deep Blending datasets are reported in Appendix A.3." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 189, + 293, + 330 + ], + "lines": [ + { + "bbox": [ + 57, + 189, + 293, + 330 + ], + "spans": [ + { + "bbox": [ + 57, + 189, + 293, + 330 + ], + "type": "table", + "html": "
MethodComp ↑FPS ↑Train ↑PSNR ↑SSIM ↑LPIPS ↓
3D-GS [14]1.00×1.00×1.00×27.550.8140.222
Trimming [2]4.00×--27.130.7980.248
Compact [18]2.28×1.07×0.73×27.080.7980.247
EAGLES [10]3.68×1.51×1.37×26.940.8000.250
Reducing [26]2.33×1.60×1.23×27.100.8090.226
Light [6]2.94×1.76×-27.280.8050.243
ELMGS [1]5.00×2.69×-27.000.7790.286
PUP [11]8.65×2.55×-26.830.7920.268
Mini-Splat [7]6.84×3.20×1.26×27.340.8220.217
+SnugBox0.99×1.81×1.08×27.550.8140.221
+AccuTile0.99×1.99×1.10×27.570.8140.221
+Soft Pruning1.79×3.14×1.30×27.320.8070.246
+Hard Pruning10.6×6.51×1.45×26.940.7820.296
", + "image_path": "407015e335dd8bb6670faceae6133ef9bbba3da783d3357d291e6c30e867c35b.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 345, + 113, + 357 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 345, + 113, + 357 + ], + "spans": [ + { + "bbox": [ + 55, + 345, + 113, + 357 + ], + "type": "text", + "content": "5.3. Results" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 363, + 212, + 375 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 363, + 212, + 375 + ], + "spans": [ + { + "bbox": [ + 55, + 363, + 212, + 375 + ], + "type": "text", + "content": "5.3.1. Additive method performance" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 379, + 295, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 379, + 295, + 534 + ], + "spans": [ + { + "bbox": [ + 54, + 379, + 295, + 534 + ], + "type": "text", + "content": "The efficacy of Speedy-Splat is demonstrated by Table 1, where we record the average execution times of each function across all scenes when additively applying our methods. SnugBox and AccuTile each introduce minimal additional computation time to preprocess and InclusiveSum. However, limiting the number of tiles touched accelerates all downstream functions, culminating in an overall speed-up of " + }, + { + "bbox": [ + 54, + 379, + 295, + 534 + ], + "type": "inline_equation", + "content": "1.82 \\times" + }, + { + "bbox": [ + 54, + 379, + 295, + 534 + ], + "type": "text", + "content": " by SnugBox that is raised to " + }, + { + "bbox": [ + 54, + 379, + 295, + 534 + ], + "type": "inline_equation", + "content": "1.99 \\times" + }, + { + "bbox": [ + 54, + 379, + 295, + 534 + ], + "type": "text", + "content": " by AccuTile. Applying soft pruning reduces the runtime of all functions by reducing the number of Gaussians, leading to a " + }, + { + "bbox": [ + 54, + 379, + 295, + 534 + ], + "type": "inline_equation", + "content": "3.14 \\times" + }, + { + "bbox": [ + 54, + 379, + 295, + 534 + ], + "type": "text", + "content": " overall speed-up. Finally, performing hard pruning improves overall speed by a whopping " + }, + { + "bbox": [ + 54, + 379, + 295, + 534 + ], + "type": "inline_equation", + "content": "6.71 \\times" + }, + { + "bbox": [ + 54, + 379, + 295, + 534 + ], + "type": "text", + "content": " over the baseline 3D-GS model." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 542, + 174, + 553 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 542, + 174, + 553 + ], + "spans": [ + { + "bbox": [ + 55, + 542, + 174, + 553 + ], + "type": "text", + "content": "5.3.2. Overall Performance" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 557, + 295, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 557, + 295, + 640 + ], + "spans": [ + { + "bbox": [ + 55, + 557, + 295, + 640 + ], + "type": "text", + "content": "In Figures 1 and 3, we report qualitative results on two outdoor and two indoor scenes from all three datasets. The magnified regions highlight that our method preserves the fine details in the baseline 3D-GS scene and closely models the ground truth view. Despite touting similar compression ratios and rendering nearly identical images, Speedy-Splat achieves over double the FPS of PUP 3D-GS." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": "Table 3 compares our methods against other methods that reduce the number of Gaussians and increase inference speed using the mean of each metric across all scenes in the Mip-NeRF 360 dataset. The underlined methods are \"lossless\", meaning that they avoid degrading visual fidelity at all. SnugBox and AccuTile, our lossless methods, improve" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 317, + 145, + 553, + 196 + ], + "blocks": [ + { + "bbox": [ + 313, + 70, + 553, + 137 + ], + "lines": [ + { + "bbox": [ + 313, + 70, + 553, + 137 + ], + "spans": [ + { + "bbox": [ + 313, + 70, + 553, + 137 + ], + "type": "text", + "content": "Table 4. Average metrics across all scenes in Section 5.1 when using the Speedy-Splat and PUP 3D-GS pruning scores to prune " + }, + { + "bbox": [ + 313, + 70, + 553, + 137 + ], + "type": "inline_equation", + "content": "88.44\\%" + }, + { + "bbox": [ + 313, + 70, + 553, + 137 + ], + "type": "text", + "content": " of Gaussians using the PUP 3D-GS pipeline. Two rounds of prune-refine are performed on each baseline 3D-GS model, pruning " + }, + { + "bbox": [ + 313, + 70, + 553, + 137 + ], + "type": "inline_equation", + "content": "66\\%" + }, + { + "bbox": [ + 313, + 70, + 553, + 137 + ], + "type": "text", + "content": " of Gaussians and then fine-tuning for 5,000 iterations in each one. The best value for each metric is color coded." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 145, + 553, + 196 + ], + "lines": [ + { + "bbox": [ + 317, + 145, + 553, + 196 + ], + "spans": [ + { + "bbox": [ + 317, + 145, + 553, + 196 + ], + "type": "table", + "html": "
Method# Gaussians ↓PSNR ↑SSIM ↑LPIPS ↓FPS ↑
Baseline2.92M27.15030.82960.2238107.53
PUP [11]0.34M26.21360.80440.2731378.57
Ours0.34M26.86580.80220.2840345.52
", + "image_path": "a1e75758a8b812f8058f85f2d71856dd6b90270b2d7f2c28e6d2383cae6c0953.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 213, + 555, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 213, + 555, + 285 + ], + "spans": [ + { + "bbox": [ + 313, + 213, + 555, + 285 + ], + "type": "text", + "content": "rendering and training speed while leaving image quality metrics essentially unchanged or slightly better. Our full pipeline, labeled as “+Hard Pruning” boasts the highest compression ratios, rendering speeds, and training speed-ups across all datasets. Furthermore, its image quality metrics are competitive with the other methods." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 299, + 465, + 312 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 299, + 465, + 312 + ], + "spans": [ + { + "bbox": [ + 313, + 299, + 465, + 312 + ], + "type": "text", + "content": "5.4. Pruning Score Comparison" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 318, + 555, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 318, + 555, + 413 + ], + "spans": [ + { + "bbox": [ + 313, + 318, + 555, + 413 + ], + "type": "text", + "content": "Although the primary focus of our work is rendering speed, we find that our efficient pruning score, described in Section 4.2, also performs well when applied in other compression pipelines. In Table 4, we ablate our efficient pruning score with the PUP 3D-GS sensitivity score across all scenes in their post-hoc pruning pipeline. Notably, Speedy-Splat's efficient pruning score outperforms PUP 3D-GS on PSNR and is competitive across the other metrics." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 431, + 389, + 443 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 431, + 389, + 443 + ], + "spans": [ + { + "bbox": [ + 313, + 431, + 389, + 443 + ], + "type": "text", + "content": "6. Limitations" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 449, + 555, + 546 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 449, + 555, + 546 + ], + "spans": [ + { + "bbox": [ + 313, + 449, + 555, + 546 + ], + "type": "text", + "content": "A limitation of Speedy-Splat is that it produces slightly lower image quality than 3D-GS. However, this degradation is expected at high compression ratios and is also observed in comparable techniques. Additionally, a direct comparison of our efficient pruning score to the PUP 3D-GS pruning score illuminates a slight, yet noticeable, gap in performance. Future work could explore the possibility of another efficient pruning score that delivers higher performance." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 563, + 388, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 563, + 388, + 574 + ], + "spans": [ + { + "bbox": [ + 313, + 563, + 388, + 574 + ], + "type": "text", + "content": "7. Conclusion" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "content": "In this work, we present Speedy-Splat: a new 3D-GS technique that accurately localizes Gaussians during rendering and significantly improves inference speed, model size, and training time. Enhanced localization is achieved by our SnugBox and AccuTile methods, while model size reduction is accomplished by our Soft and Hard Pruning approaches. Together, our Speedy-Splat methods accelerate rendering speed by an average of " + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "inline_equation", + "content": "6.71 \\times" + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "content": ", reduce model size by " + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "inline_equation", + "content": "10.6 \\times" + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "content": ", and improve training time by " + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "inline_equation", + "content": "1.47 \\times" + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "content": " across all scenes from the Mip-NeRF 360, Tanks & Temples, and Deep Blending datasets." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21544" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 170, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 170, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 170, + 85 + ], + "type": "text", + "content": "8. Acknowledgements" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 86, + 296, + 182 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 86, + 296, + 182 + ], + "spans": [ + { + "bbox": [ + 55, + 86, + 296, + 182 + ], + "type": "text", + "content": "This work was made possible by the IARPA WRIVA Program, the ONR MURI program, and DAPRA TIAMAT. Commercial support was provided by Capital One Bank, the Amazon Research Award program, and Open Philanthropy. Further support was provided by the National Science Foundation (IIS-2212182), and by the NSF TRAILS Institute (2229885). Zwicker was additionally supported by the National Science Foundation (IIS-2126407)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 192, + 115, + 205 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 192, + 115, + 205 + ], + "spans": [ + { + "bbox": [ + 56, + 192, + 115, + 205 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 212, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 212, + 296, + 256 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 212, + 296, + 256 + ], + "spans": [ + { + "bbox": [ + 61, + 212, + 296, + 256 + ], + "type": "text", + "content": "[1] Muhammad Salman Ali, Sung-Ho Bae, and Enzo Tartaglione. Elmgs: Enhancing memory and computation scalability through compression for 3d gaussian splatting. arXiv preprint arXiv:2410.23213, 2024. 2, 8, 12" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 257, + 296, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 257, + 296, + 300 + ], + "spans": [ + { + "bbox": [ + 61, + 257, + 296, + 300 + ], + "type": "text", + "content": "[2] Muhammad Salman Ali, Maryam Qamar, Sung-Ho Bae, and Enzo Tartaglione. Trimming the fat: Efficient compression of 3d gaussian splats through pruning. arXiv preprint arXiv:2406.18214, 2024. 2, 8, 12" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 301, + 296, + 356 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 301, + 296, + 356 + ], + "spans": [ + { + "bbox": [ + 61, + 301, + 296, + 356 + ], + "type": "text", + "content": "[3] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5470–5479, 2022. 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 357, + 295, + 389 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 357, + 295, + 389 + ], + "spans": [ + { + "bbox": [ + 62, + 357, + 295, + 389 + ], + "type": "text", + "content": "[4] Anpei Chen, Zexiang Xu, Andreas Geiger, Jingyi Yu, and Hao Su. Tensorf: Tensorial radiance fields. In European Conference on Computer Vision (ECCV), 2022. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 391, + 296, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 391, + 296, + 434 + ], + "spans": [ + { + "bbox": [ + 62, + 391, + 296, + 434 + ], + "type": "text", + "content": "[5] Sankeerth Durvasula, Adrian Zhao, Fan Chen, Ruofan Liang, Pawan Kumar Sanjaya, and Nandita Vijaykumar. Distwar: Fast differentiable rendering on raster-based rendering pipelines. arXiv preprint arXiv:2401.05345, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 435, + 296, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 435, + 296, + 479 + ], + "spans": [ + { + "bbox": [ + 62, + 435, + 296, + 479 + ], + "type": "text", + "content": "[6] Zhiwen Fan, Kevin Wang, Kairun Wen, Zehao Zhu, Dejia Xu, and Zhangyang Wang. Lightgaussian: Unbounded 3d gaussian compression with 15x reduction and " + }, + { + "bbox": [ + 62, + 435, + 296, + 479 + ], + "type": "inline_equation", + "content": "200+" + }, + { + "bbox": [ + 62, + 435, + 296, + 479 + ], + "type": "text", + "content": " fps. arXiv preprint arXiv:2311.17245, 2023. 1, 2, 7, 8, 12" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 480, + 296, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 480, + 296, + 512 + ], + "spans": [ + { + "bbox": [ + 62, + 480, + 296, + 512 + ], + "type": "text", + "content": "[7] Guangchi Fang and Bing Wang. Mini-splatting: Representing scenes with a constrained number of gaussians. arXiv preprint arXiv:2403.14166, 2024. 1, 2, 8, 12" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 514, + 296, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 514, + 296, + 567 + ], + "spans": [ + { + "bbox": [ + 62, + 514, + 296, + 567 + ], + "type": "text", + "content": "[8] Guofeng Feng, Siyan Chen, Rong Fu, Zimu Liao, Yi Wang, Tao Liu, Zhilin Pei, Hengjie Li, Xingcheng Zhang, and Bo Dai. Flashgs: Efficient 3d gaussian splatting for large-scale and high-resolution rendering. arXiv preprint arXiv:2408.07967, 2024.1, 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 569, + 296, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 569, + 296, + 590 + ], + "spans": [ + { + "bbox": [ + 62, + 569, + 296, + 590 + ], + "type": "text", + "content": "[9] Henri P. Gavin. The levenberg-marquardt method for nonlinear least squares curve-fitting problems c ©. 2013. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 592, + 295, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 592, + 295, + 624 + ], + "spans": [ + { + "bbox": [ + 57, + 592, + 295, + 624 + ], + "type": "text", + "content": "[10] Sharath Girish, Kamal Gupta, and Abhinav Shrivastava. Eagles: Efficient accelerated 3d gaussians with lightweight encodings. arXiv preprint arXiv:2312.04564, 2023. 2, 8, 12" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 625, + 296, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 625, + 296, + 668 + ], + "spans": [ + { + "bbox": [ + 57, + 625, + 296, + 668 + ], + "type": "text", + "content": "[11] Alex Hanson, Allen Tu, Vasu Singla, Mayuka Jayawardhana, Matthias Zwicker, and Tom Goldstein. Pup 3d-gs: Principled uncertainty pruning for 3d gaussian splattering. arXiv preprint arXiv:2406.10219, 2024. 1, 2, 4, 5, 7, 8, 12" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 670, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 296, + 713 + ], + "type": "text", + "content": "[12] Peter Hedman, Julien Philip, True Price, Jan-Michael Frahm, George Drettakis, and Gabriel Brostow. Deep blending for free-viewpoint image-based rendering. ACM Transactions on Graphics, 37(6):1-15, 2018. 7" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 316, + 73, + 555, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 555, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 555, + 116 + ], + "type": "text", + "content": "[13] Lukas Hollein, Aljaž Božić, Michael Zollhöfer, and Matthias Nießner. 3dgs-lm: Faster gaussian-splatting optimization with levenberg-marquardt. arXiv preprint arXiv:2409.12892, 2024. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 118, + 554, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 118, + 554, + 161 + ], + "spans": [ + { + "bbox": [ + 317, + 118, + 554, + 161 + ], + "type": "text", + "content": "[14] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42 (4):1-14, 2023. 1, 2, 3, 7, 8, 12, 13" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 162, + 555, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 162, + 555, + 226 + ], + "spans": [ + { + "bbox": [ + 317, + 162, + 555, + 226 + ], + "type": "text", + "content": "[15] Shakiba Kheradmand, Daniel Rebain, Gopal Sharma, Weiwei Sun, Yang-Che Tseng, Hossam Isack, Abhishek Kar, Andrea Tagliasacchi, and Kwang Moo Yi. 3d gaussian splatting as markov chain monte carlo. In Advances in Neural Information Processing Systems (NeurIPS), 2024. Spotlight Presentation. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 228, + 554, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 228, + 554, + 249 + ], + "spans": [ + { + "bbox": [ + 316, + 228, + 554, + 249 + ], + "type": "text", + "content": "[16] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. CoRR, abs/1412.6980, 2014. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 251, + 554, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 251, + 554, + 293 + ], + "spans": [ + { + "bbox": [ + 317, + 251, + 554, + 293 + ], + "type": "text", + "content": "[17] Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and temples: Benchmarking large-scale scene reconstruction. ACM Transactions on Graphics, 36(4):1-13, 2017. 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 294, + 554, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 294, + 554, + 336 + ], + "spans": [ + { + "bbox": [ + 316, + 294, + 554, + 336 + ], + "type": "text", + "content": "[18] Joo Chan Lee, Daniel Rho, Xiangyu Sun, Jong Hwan Ko, and Eunbyung Park. Compact 3d gaussian representation for radiance field. arXiv preprint arXiv:2311.13681, 2023. 2, 8, 12" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 338, + 554, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 338, + 554, + 381 + ], + "spans": [ + { + "bbox": [ + 316, + 338, + 554, + 381 + ], + "type": "text", + "content": "[19] Weikai Lin, Yu Feng, and Yuhao Zhu. Rtgs: Enabling real-time gaussian splatting on mobile devices using efficiency-guided pruning and foveated rendering. arXiv preprint arXiv:2407.00435, 2024. 1, 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 383, + 554, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 383, + 554, + 425 + ], + "spans": [ + { + "bbox": [ + 316, + 383, + 554, + 425 + ], + "type": "text", + "content": "[20] Xiangrui Liu, Xinju Wu, Pingping Zhang, Shiqi Wang, Zhu Li, and Sam Kwong. Compgs: Efficient 3d scene representation via compressed gaussian splatting. arXiv preprint arXiv:2404.09458, 2024. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 427, + 555, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 427, + 555, + 482 + ], + "spans": [ + { + "bbox": [ + 316, + 427, + 555, + 482 + ], + "type": "text", + "content": "[21] Tao Lu, Mulin Yu, Linning Xu, Yuanbo Xiangli, Limin Wang, Dahua Lin, and Bo Dai. Scaffold-gs: Structured 3d gaussians for view-adaptive rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20654-20664, 2024. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 483, + 555, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 483, + 555, + 548 + ], + "spans": [ + { + "bbox": [ + 317, + 483, + 555, + 548 + ], + "type": "text", + "content": "[22] Saswat Subhajyoti Mallick, Rahul Goel, Bernhard Kerbl, Markus Steinberger, Francisco Vicente Carrasco, and Fernando De La Torre. Taming 3dgs: High-quality radiance fields with limited resources. In SIGGRAPH Asia 2024 Conference Papers, New York, NY, USA, 2024. Association for Computing Machinery. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 548, + 554, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 548, + 554, + 601 + ], + "spans": [ + { + "bbox": [ + 317, + 548, + 554, + 601 + ], + "type": "text", + "content": "[23] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 603, + 554, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 603, + 554, + 646 + ], + "spans": [ + { + "bbox": [ + 316, + 603, + 554, + 646 + ], + "type": "text", + "content": "[24] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph., 41(4):102:1-102:15, 2022. 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 647, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 554, + 713 + ], + "type": "text", + "content": "[25] Michael Niemeyer, Fabian Manhardt, Marie-Julie Rakoto-saona, Michael Oechsle, Daniel Duckworth, Rama Gosula, Keisuke Tateno, John Bates, Dominik Kaeser, and Federico Tombari. Radsplat: Radiance field-informed gaussian splatting for robust real-time rendering with " + }, + { + "bbox": [ + 316, + 647, + 554, + 713 + ], + "type": "inline_equation", + "content": "900+" + }, + { + "bbox": [ + 316, + 647, + 554, + 713 + ], + "type": "text", + "content": " fps. arXiv preprint arXiv:2403.13806, 2024. 1, 2" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "21545" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 474 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 127 + ], + "type": "text", + "content": "[26] Panagiotis Papantonakis, Georgios Kopanas, Bernhard Kerbl, Alexandre Lanvin, and George Drettakis. Reducing the memory footprint of 3d gaussian splatting. Proceedings of the ACM on Computer Graphics and Interactive Techniques, 7(1):1-17, 2024. 2, 8, 12" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 294, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 294, + 183 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 294, + 183 + ], + "type": "text", + "content": "[27] Lukas Radl, Michael Steiner, Mathias Parger, Alexander Weinrauch, Bernhard Kerbl, and Markus Steinberger. StopThePop: Sorted Gaussian Splatting for View-Consistent Real-time Rendering. ACM Transactions on Graphics, 4 (43), 2024. 2, 13" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 185, + 294, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 185, + 294, + 228 + ], + "spans": [ + { + "bbox": [ + 56, + 185, + 294, + 228 + ], + "type": "text", + "content": "[28] Kerui Ren, Lihan Jiang, Tao Lu, Mulin Yu, Linning Xu, Zhangkai Ni, and Bo Dai. Octree-gs: Towards consistent real-time rendering with lod-structured 3d gaussians. arXiv preprint arXiv:2403.17898, 2024. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 230, + 294, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 230, + 294, + 295 + ], + "spans": [ + { + "bbox": [ + 56, + 230, + 294, + 295 + ], + "type": "text", + "content": "[29] Fernando Rivas-Manzaneque, Jorge Sierra-Acosta, Adrian Penate-Sanchez, Francesc Moreno-Noguer, and Angela Ribeiro. Nerflight: Fast and light neural radiance fields using a shared feature grid. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12417-12427, 2023. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 297, + 294, + 339 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 297, + 294, + 339 + ], + "spans": [ + { + "bbox": [ + 56, + 297, + 294, + 339 + ], + "type": "text", + "content": "[30] Samuel Rota Bulò, Lorenzo Porzi, and Peter Kontschieder. Revising densification in gaussian splatting. In European Conference on Computer Vision, pages 347-362. Springer, 2024. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 341, + 294, + 394 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 341, + 294, + 394 + ], + "spans": [ + { + "bbox": [ + 56, + 341, + 294, + 394 + ], + "type": "text", + "content": "[31] Evangelos Ververas, Rolandos Alexandros Potamias, Jifei Song, Jiankang Deng, and Stefanos Zafeiriou. Sags: structure-aware 3d gaussian splatting. In European Conference on Computer Vision, pages 221-238. Springer, 2024. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 397, + 294, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 397, + 294, + 441 + ], + "spans": [ + { + "bbox": [ + 56, + 397, + 294, + 441 + ], + "type": "text", + "content": "[32] Meng Wei, Qianyi Wu, Jianmin Zheng, Hamid Rezatofighi, and Jianfei Cai. Normal-gs: 3d gaussian splatting with normal-involved rendering. In Advances in Neural Information Processing Systems (NeurIPS), 2024. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 442, + 294, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 442, + 294, + 474 + ], + "spans": [ + { + "bbox": [ + 56, + 442, + 294, + 474 + ], + "type": "text", + "content": "[33] Matthias Zwicker, Hanspeter Pfister, Jeroen Van Baar, and Markus Gross. Ewa splatting. IEEE Transactions on Visualization and Computer Graphics, 8(3):223-238, 2002. 3" + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "21546" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/SphereUFormer_ A U-Shaped Transformer for Spherical 360 Perception/d3dc320a-7961-421e-ae82-56124645a3a0_content_list.json b/2025/SphereUFormer_ A U-Shaped Transformer for Spherical 360 Perception/d3dc320a-7961-421e-ae82-56124645a3a0_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..bf18b22d5e42dc6b3164f82e8411c5d8091ca9fa --- /dev/null +++ b/2025/SphereUFormer_ A U-Shaped Transformer for Spherical 360 Perception/d3dc320a-7961-421e-ae82-56124645a3a0_content_list.json @@ -0,0 +1,1836 @@ +[ + { + "type": "text", + "text": "SphereUFormer: A U-Shaped Transformer for Spherical 360 Perception", + "text_level": 1, + "bbox": [ + 133, + 130, + 864, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yaniv Benny \nTel Aviv University", + "bbox": [ + 305, + 181, + 462, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Lior Wolf \nTel Aviv University", + "bbox": [ + 532, + 181, + 691, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 268, + 326, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "This paper proposes a novel method for omnidirectional $360^{\\circ}$ perception. Most common previous methods relied on equirectangular projection. This representation is easily applicable to 2D operation layers but introduces distortions into the image. Other methods attempted to remove the distortions by maintaining a sphere representation but relied on complicated convolution kernels that failed to show competitive results. In this work, we introduce a transformer-based architecture that, by incorporating a novel \"Spherical Local Self-Attention\" and other spherically-oriented modules, successfully operates in the spherical domain and outperforms the state-of-the-art in $360^{\\circ}$ perception benchmarks for depth estimation and semantic segmentation. Our code is available at https://github.com/yanivbenny/sphere_uformer.", + "bbox": [ + 89, + 300, + 485, + 527 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 556, + 220, + 571 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Monocular omnidirectional $360^{\\circ}$ perception is an important setting, as it allows for a full field-of-view (FOV) receptive field for the underlying model. Consequently, an entire room layout can be inferred, for example. In contrast to regular images, a $360^{\\circ}$ image has no boundaries. Instead, the image is horizontally cyclical, and converges to singular points at the vertical extreme points. This suggests that this domain requires a different treatment, and that best practices that work well with regular images might not have the same effect on $360^{\\circ}$ images. To tackle this task, multiple topics need to be considered.", + "bbox": [ + 89, + 582, + 482, + 747 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "How should the data be represented? What kind of distortions and irregularities are created by this representation? How will a model process this representation? How to deal with these distortions and irregularities?", + "bbox": [ + 89, + 750, + 482, + 809 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "These questions have attracted researchers to construct multiple datasets [3, 4, 9, 53] where depth estimation and semantic segmentation are the main tasks, and to formulate many solutions [20, 23, 28, 32, 35, 36, 38, 42, 43, 47, 48, 51], which propose novel approaches to solving these tasks. A common pattern in previous works is that they rep", + "bbox": [ + 89, + 810, + 482, + 901 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "resent the data by projecting it to a 2D plane in different ways, usually equirectangular projection [30], cube mapping [26], or patch cropping. Each approach has its advantages and disadvantages, but the common motivation is to represent the image in a grid formation. This enables grid architectures, such as Convolutional Neural Networks or Vision Transformers to easily be applied. This comes at a cost of distortions, irregularities, or limited FOV, that are caused by the selected projection. A less common alternative representation that has been proposed is to represent the image in its original sphere domain by discretizing the sphere surface in some manner [6, 14, 17, 24]. The motivation is to maintain an undistorted representation of the $360^{\\circ}$ image. However, this representation formed architectural challenges that require novel solutions. Some previous works [8, 12, 27, 46, 49, 50] addressed these challenges in various manners, but they fail to compete with current state-of-the-art solutions that follow a grid representation.", + "bbox": [ + 511, + 268, + 906, + 542 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this work, we tackle the drawbacks of spherical representations. Specifically, we introduce SphereUFormer, which is a novel U-Shaped Transformer architecture tailored for spherical $360^{\\circ}$ perception. By directly operating on a spherical domain without distortion-inducing projections, SphereUFormer addresses the core challenges of omnidirectional perception. Our approach builds on the strengths of transformer models, incorporating modifications that allow it to efficiently process and understand spherical data. This includes the development of spherical-specific operations such as localized self-attention mechanism that respects the geometry of the sphere, innovative up/downsampling techniques that maintain the integrity of spherical data throughout the model, and positional encoding scheme that matches the domain properties. SphereUFormer outperforms the state of the art in the field of $360^{\\circ}$ perception, providing a robust and efficient tool for a wide range of applications that require comprehensive understanding of spherical environments. In Sec. 3 we describe this representation in detail. We then further describe a new method in Sec. 4, which utilizes this representation and performs $360^{\\circ}$ perception on spherical representations, and in Sec. 5, we analyze and compare its results.", + "bbox": [ + 511, + 553, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "940", + "bbox": [ + 485, + 945, + 511, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 91, + 89, + 232, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Omnidirectional perception. There have been many advances, novelties, and research verticals in omnidirectional perception over the years. Initial works attempted standard convolutional networks on an equirectangular projection [54] (ERP). Although promising, the results were affected by the distortions created by the projection. To better handle these distortions, later works suggested technical adjustments to the convolution kernel such as Spherical Convolutions [37] or Distortion-Aware Convolutions [12, 39]. The latter sample the convolution features from a tangent plane. Other solutions attempted to eliminate distortions by sampling the image into smaller FOV patches [16, 28, 35, 47] followed by a post-processing stage that fuses the independent results. This, however, is at a cost of a restricted receptive field per patch and possible cutoff of crucial information, which either affects results or requires high overlap between patches. Due to the advantages and limitations of each method, [2, 42, 43] suggested a combined approach that fuses the ERP path with the individual patches at some stage of the network. Alternative approaches avoided the use of projection, and instead, operated directly on the sphere by constructing a spherical mesh. Usually either an icosphere [13, 22, 27, 46] or healpix [8, 13], and apply Graph Convolution [52], Mesh Convolution [21], or some custom convolution-like operation.", + "bbox": [ + 93, + 116, + 483, + 493 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Vision Transformers. ViT's [15] have disrupted the field of computer vision, from general-purpose architectures [34, 44] to specific purposes [7, 10, 45]. The attention mechanism [5, 29, 40] allows for dynamic pooling of features based on relevance and structural position. ViT's benefit from an efficient large context window compared to their CNN counterparts, which allows them to more easily learn long-distance relations between features. While earlier omnidirectional perception works proposed solutions based on CNN's, recent contributions replaced the convolution kernels with local attention of various forms [20, 36, 48] with favorable results. This work follows this direction. It utilizes the foundation of UFormer [44] and adapts it to spherical representations with custom attention layers.", + "bbox": [ + 89, + 515, + 483, + 727 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Graph Attention. Graph Convolution Networks [52] (GCN) are neural networks that apply convolution-like operations on data represented as a graph. A graph convolution operation aggregates the results of local operations between a node and each of its neighbors. Over multiple layers, this allows for information to pass between distant nodes in the graph and the update of these nodes. Graph Attention Networks [41] were proposed to enhance Graph Convolution Networks, thereby improving the message-passing capabilities of the GCN. The attention mechanism", + "bbox": [ + 89, + 750, + 483, + 900 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/f6b9bb34f3fa26282bcf785443990c35dec9582ee7f7d4e8cdc5f972ad1f70a0.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 532, + 90, + 604, + 138 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/4554437dc1bffff70ead73229baf292b2c3060de5bfa7e2353e068a0d261d2ac.jpg", + "image_caption": [ + "(b)", + "Figure 1. Sphere Representations. From left to right: uvsphere, cubesphere, icosphere, hexasphere." + ], + "image_footnote": [], + "bbox": [ + 627, + 90, + 700, + 138 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/c29ad352dee20996e1614884e1ccab961625b85196b11fce84aa5140276d2df0.jpg", + "image_caption": [ + "(c)" + ], + "image_footnote": [], + "bbox": [ + 723, + 90, + 795, + 138 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/ce2dc61b351747cfde2dc7c64a8cbd4c193d000a3ab068314f2ed4529853b8a4.jpg", + "image_caption": [ + "(d)" + ], + "image_footnote": [], + "bbox": [ + 820, + 90, + 890, + 138 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "has the benefit of allowing the operations to focus on a few sets of relevant nodes and providing an efficient method to encode positional information in the graph. For Geometric Graphs [31] geometric location is encoded to inform of each node's absolute and relative position, thereby making nodes aware of their global position and their respective neighbors' relative position.", + "bbox": [ + 511, + 226, + 906, + 332 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Sphere Representation", + "text_level": 1, + "bbox": [ + 511, + 345, + 728, + 363 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Arguably the first topic in question when attempting to solve a problem is how to formulate the data. For omnidirectional $360^{\\circ}$ images, the data can be seen as an array of values $v_{\\phi ,\\theta}$ assigned to a ray pointing from a center point-of-view towards $(\\theta ,\\phi)$ for $\\theta \\in [0,2\\pi ]$ being the horizontal angle and $\\phi \\in [0,\\pi ]$ the vertical. The value $v$ can correspond to the RGB value, depth, object category, etc. By selecting to sample on finite such angle pairs, the sphere becomes discretized. The sampling method determines how the sampled data can be processed and how successful the solution will become.", + "bbox": [ + 511, + 372, + 906, + 536 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The most common and straightforward is a uvsphere [17], where points are sampled in a grid with equal spacing for $(\\theta, \\phi)$ . See Fig. 1a. The uvsphere has a very high degree of horizontal symmetry, which is a good property. Another advantage is that it allows an unwrapping of the sphere into a 2D matrix. This unwrapping is called equirectangular projection (ERP). The downside of this representation is that the density of samples on the sphere is higher around the poles, due to the decreasing cross-section. This creates an imbalanced effective resolution over the sphere, and the projected image in 2D is distorted.", + "bbox": [ + 511, + 537, + 905, + 703 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "An alternative that mitigates imbalanced resolution and distortion is cube mapping (or cubosphere) [14] (Fig. 1b), which projects the sphere onto the faces of a cube. Sample resolution is controlled by deciding how many rectangles each face of the cube is subdivided into. This removes the sampling density around the poles and the sphere can be unwrapped into six 2D images. However, two of these images are not vertically oriented and require special treatment. In addition, the transition between faces is not fluent, and complex padding methods [11] can only increase the receptive field of each image to some extent. Since the images are separated, a post-processing stage needs to fuse their predictions to form a unified prediction on the sphere. One op-", + "bbox": [ + 511, + 704, + 906, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "941", + "bbox": [ + 485, + 945, + 509, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/dd82f349e41cfb609088cc68eb7fb79b93e1e3520780d4cf91e4de59c751f6c9.jpg", + "image_caption": [ + "0" + ], + "image_footnote": [], + "bbox": [ + 106, + 94, + 158, + 143 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/1e48d5f8596e033fdd8b70b3a1126c05fd15fa75f99f804c50786b22e6d43c4b.jpg", + "image_caption": [ + "1" + ], + "image_footnote": [], + "bbox": [ + 178, + 95, + 240, + 143 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/83a76e3f3ded829311a02ebe795e95d25d38547ac549c2a431c1a1ebc07808fb.jpg", + "image_caption": [ + "2", + "Figure 2. Icospheres of different ranks. An increase in rank is made by subdividing each triangle into 4 smaller triangles, and results in an increased resolution." + ], + "image_footnote": [], + "bbox": [ + 254, + 94, + 318, + 143 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/6e43b0e4efa2e8f54e574c6f85bed29998c43d45cecf8b01791a1ea603d0c7ba.jpg", + "image_caption": [ + "3" + ], + "image_footnote": [], + "bbox": [ + 331, + 94, + 395, + 143 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/be04b33dde31b4d8a742c4fe36839a736ed0ff06874bb0a1577df740b574458e.jpg", + "image_caption": [ + "4" + ], + "image_footnote": [], + "bbox": [ + 410, + 94, + 473, + 143 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/50ba434a066253c32422efd9e99c0a1568bf1181707ecfd93feb41faa6709e4f.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Rank:01234567
Faces:208032012805120~20K~82K~328K
Vertices:12421626422562~10K~41K~164K
", + "bbox": [ + 109, + 239, + 467, + 292 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Table 1. Icosphere Properties. Number of faces $(20 \\times 4^i)$ and number of vertices $(10 \\times 4^i + 2)$ per subdivision rank $i$ .", + "bbox": [ + 89, + 304, + 482, + 333 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/f7724cd19a4b4836e19e5a61040721a9dfa0e43c0503c8298b902909e9532aea.jpg", + "image_caption": [ + "(a1)", + "Figure 3. Data as Icospheres. Icosphere of rank 3 (a1-c1) and rank 6 (a2-c2). (a) RGB, (b) Depth Map, and (c) Semantic Layout." + ], + "image_footnote": [], + "bbox": [ + 101, + 354, + 161, + 402 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/4e3d73cc128036581d3f3c47d3bbf23711ce7fdc9ca766e22ab5707c309d82c6.jpg", + "image_caption": [ + "(b1)" + ], + "image_footnote": [], + "bbox": [ + 165, + 354, + 223, + 402 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/959764e5059a9c618a84e49fcbd7c0764389ef0da880f2ab4ebfcefdbc1e3215.jpg", + "image_caption": [ + "(c1)" + ], + "image_footnote": [], + "bbox": [ + 227, + 354, + 284, + 402 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/1fff96c2dadca431aea67e0f12be9dc718bc1b5ed18fd284a6d66db575fa2561.jpg", + "image_caption": [ + "(a2)" + ], + "image_footnote": [], + "bbox": [ + 289, + 354, + 346, + 402 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/1d63be6755d21db5f8fcbc6d3052698685442f334b5cb6cdac4c0470503afd61.jpg", + "image_caption": [ + "(b2)" + ], + "image_footnote": [], + "bbox": [ + 349, + 354, + 408, + 402 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/4ce7b793d7f57c13ca3bb2a0c052bb2cdecf884a4cdb2f9f40fb19bfb7e2196c.jpg", + "image_caption": [ + "(c2)" + ], + "image_footnote": [], + "bbox": [ + 413, + 354, + 470, + 402 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "tion is to operate on the cubosphere mesh instead of the 2D projection. Unfortunately, the cubosphere has irregularities at the coordinates of the cube corners and does not have a high horizontal symmetry, which makes it less optimal for this purpose.", + "bbox": [ + 89, + 489, + 482, + 566 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Instead of projecting the data to a grid and then having to deal with its limitations, a more elegant approach can be formulated using the sphere representation directly. Aside from the previously mentioned options, a common way to discretize a sphere is with Geodesic (icosphere, Fig. 1c) and Goldberg (hexasphere, Fig. 1d) polyhedrons [6]. An icosphere is created by repeatedly subdividing an icosahedron into smaller triangles. A hexasphere is the dual polyhedron where vertices are swapped with face normals, and edge connections with face adjacency. These types of sphere representation have a very high degree of symmetry and have a very evenly spaced sampling on the sphere surface. Due to their duality, our work follows the icosphere mesh, and icosphere/hexasphere mode is toggled by whether data points $(\\theta, \\phi)$ describe mesh vertices (hexasphere) or face normals (icosphere). Therefore we will sometimes refer to icospheres/hexaspheres by referring to the \"node type\" (face/vertex respectively). Other notable options for sphere sampling are Fibonacci spheres [24] and HealPix [19].", + "bbox": [ + 89, + 566, + 482, + 853 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Icospheres have the highest order of symmetry, the most uniform distribution of points (perfect uniform distribution is impossible), easy setup, and an inherent up/down", + "bbox": [ + 89, + 854, + 482, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "sampling definition through subdivision. All these reasons make them ideal for discrete spherical representation. As Fig. 2 and Tab. 1 show, in different ranks, icospheres produce sphere discretization in various resolutions. Each increase in sphere rank divides each triangle face into four smaller ones, hereby gradually increasing the sphere resolution. An example of representing data with icospheres can be seen in Fig. 3. In low resolution, the image seems \"pixelated\" as the independent triangle faces are distinguishable. But, with a high enough resolution, the image becomes detailed and visibly smooth. Our method uses high-resolution icospheres to represent the $360^{\\circ}$ image, and through up and downsampling on the sphere structure, uses low-resolution spheres as hidden layers inside the model.", + "bbox": [ + 511, + 90, + 903, + 303 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Method", + "text_level": 1, + "bbox": [ + 513, + 315, + 604, + 330 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This section introduces SphereUFormer, our proposed U-shaped Transformer for Spherical $360^{\\circ}$ Perception. As the name suggests, the model is influenced by UFormer [44], with significant modifications to support spherically structured data. A high-level diagram is depicted in Fig. 4. The model's components are described in this section. For technical implementation details, see the supplementary.", + "bbox": [ + 511, + 340, + 903, + 446 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The network begins with an input projection that applies linear projection from the RGB values to latent embedding vectors. The projection output is then passed into an Encoder-Decoder network with multiple levels of self-attention modules in different sphere resolutions. We call these modules \"Spherical Attention Modules\" (SAM). At the bottom of the network lies a bottleneck SAM module that operates on the lowest resolution. Skip connections pass from the encoder to the decoder and feed high-resolution features. Finally, a linear Output Projection projects the decoder output to the output dimensions.", + "bbox": [ + 511, + 446, + 905, + 612 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As described in Sec. 3, our solution features the icosphere as its spherical representation structure, as we have found it to be the most intuitive and to contain the most satisfying properties. But it is not limited to it, and with minor modifications supports any spherical structure.", + "bbox": [ + 511, + 613, + 903, + 686 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our method contains many local operations on the graph, from the down and up sampling, to the self-attention mechanism. Computing the cosine similarity or finding the neighbors of a node to generate the mapping scheme is slightly costly. This would be a problem if it was performed on each iteration, as the model would be slowed down. Luckily, since the icosphere graph is fixed, these mappings only need to be computed once in advance. Therefore there is no need to recompute it in every message massing operation.", + "bbox": [ + 511, + 688, + 903, + 824 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1. Up/Down Sampling", + "text_level": 1, + "bbox": [ + 511, + 833, + 700, + 849 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the spherical setting, upsampling and downsampling operations are performed by transforming a spherical graph into a new one with more or fewer nodes, respectively. This", + "bbox": [ + 511, + 854, + 903, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "942", + "bbox": [ + 485, + 944, + 511, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/9a998734b9ccfb2e19d2e04bb23a716bb5d948d46206bd12a0d0252d671cb333.jpg", + "image_caption": [ + "Figure 4. The SphereUFormer architecture. A spherical representation is fed into the model. A linear input projection layer encodes the RGB values to latent embedding vectors. A sequence of SAM modules apply local self-attention on the spherical data along with downsampling layers that gradually reduce the resolution of the sphere. A sequence of SAM modules along with an upsampling layer and bypass skip connections decode the data. An output projection converts the latent embeddings to the output channel size." + ], + "image_footnote": [], + "bbox": [ + 106, + 88, + 901, + 313 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "can be difficult and inefficient to compute for two arbitrary spherical polyhedrons. However, icospheres have multiple elegant and desired properties. First, an increase in resolution by one rank is the result of splitting each triangle face into four smaller ones, with one central triangle that shares its center with the large one. Second, for the same reason, each vertex in the low-resolution icosphere also exists in the high-resolution one, with new vertices created at the center of each edge connection. Third, the number of nodes/faces increases at a constant rate.", + "bbox": [ + 88, + 404, + 483, + 553 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "All these properties make constructing up and down operations extremely straightforward and efficient. Every common pooling technique, such as max pooling, average pooling, center pooling, and interpolation, can be implemented. However, there are preferences for each setting. For downsampling, when the data nodes are represented by the faces, we have found max and center pooling to have the same overall results. When the data nodes are on the vertices (equivalent to hexosphere), max and average pooling become a little complicated, since there are both instances of five and six neighbors in the graph. Therefore, we resorted to only center pooling. For upsampling, the faces mode most simply operates with nearest upsampling, where each triangle receives the value according to the lower resolution triangle it was subdivided from. Other interpolation methods require more complicated computations and were therefore not considered. For vertices, however, interpolation is straightforward, since each new node is located precisely at the center of an existing edge. For more upscaled upsampling, Barycentric coordinates [18] interpolation can also be performed quite easily, but we did not find a reason to upscale beyond a factor of 2 at a time.", + "bbox": [ + 91, + 569, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Global Positional Encoding", + "text_level": 1, + "bbox": [ + 511, + 402, + 759, + 420 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Positional encoding is a common and critical technique in Transformers [33, 40] and Vision Transformers [7, 10, 15, 34, 44, 45]. SphereUFormer utilizes positional encoding as well, but in a manner that suits spherical representations. The node position on the unit sphere is expressed using spherical coordinates $(\\theta, \\phi)$ . In our setting of omnidirectional perception, the data is always vertically aligned, therefore the model does not need to be vertically equivariant. However, the horizontal rotation is arbitrary, therefore it is desired to make the model horizontally equivariant to rotation and flipping. We therefore apply absolute positional encoding for the vertical position but not for the horizontal position. Horizontal position is only used in relative to another node as position bias in the Local Self-Attention described in the next section.", + "bbox": [ + 511, + 426, + 906, + 652 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For the vertical position, we consider $\\phi \\in [0,\\pi ]$ . The position was encoded by applying sinusoidal encoding on the position value, then passed through an MLP with an output dimension equal to the embedding dimension. We apply the vertical global positional encoding on the projected input and on the queries and keys in each attention module. More details can be found in the supplement.", + "bbox": [ + 511, + 652, + 908, + 760 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.3. Spherical Local Self-Attention", + "text_level": 1, + "bbox": [ + 511, + 771, + 782, + 787 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The Spherical Local Self-Attention is the main component in our network that drives the SphereUFormer's $360^{\\circ}$ perception. A detailed diagram is shown in Fig. 5. Given an input sphere, each node $x_{i}$ in the sphere is projected into a query $(q_{i})$ , key $(k_{i})$ , and value $(v_{i})$ . These vectors are all of the shape $(N,L,H,D_H)$ , where $H$ is the number of heads and $D_H$ is the head dimension. Similar to standard prac", + "bbox": [ + 511, + 794, + 908, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "943", + "bbox": [ + 485, + 944, + 509, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/8e3578f1b6abdee667ebccf5518ea29ee3fdacd56bbc9bdacb713f51a58614fa.jpg", + "image_caption": [ + "Figure 5. Spherical Local Self Attention. Attention is applied between each data node and its $K$ neighbors. A learned relative position bias encodes information about the neighbors' relative position. In the right corner is a diagram of the enclosing block." + ], + "image_footnote": [], + "bbox": [ + 94, + 87, + 486, + 244 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/3df419116959d1c79b9db42eaf035bb3d4833408380c9242438277b18bbf1a65.jpg", + "image_caption": [ + "Figure 6. Relative Positional Encondig. On the left, a center point (blue) and its first (green) and second (yellow) degree neighbors are defined by the graph structure. The relative position kernel is defined by a $7 \\times 7$ grid, represented by the black nodes in the right image. The center point is positioned at the center of the grid and the neighbors are distributed across the grid according to their relative change in horizontal ( $\\Delta \\theta$ ) and vertical ( $\\Delta \\phi$ ) angles. Bilinear interpolation samples a learned positional encoding." + ], + "image_footnote": [], + "bbox": [ + 112, + 325, + 478, + 474 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "tices, $D_H$ is derived from the input dimension $D$ and the number of heads. However, since the self-attention mechanism is the only spatial operation in the network, we added a \"head dimension coefficient\" ( $C_{head}$ ) hyperparameter to upscale the head dimension $D_H$ as a reverse bottleneck. Thus: $D_H = (D / H) \\cdot C_{head}$ . This coefficient allows an increase in the number of attention heads without reducing the dimension of each head and without increasing the overall size of the entire network, resulting in only a minor increase in parameters. We further discuss the motivation of this in the ablation section (Sec. 5.4).", + "bbox": [ + 89, + 626, + 483, + 792 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For each query, the neighboring $K$ vectors are grouped according to the sphere geometry. Attention for each query is performed on the selected neighboring group. The number of neighbors $K$ depends on a hyperparameter named \"window coefficient\" ( $C_{win}$ ), which dictates the N-order neighbors to collect. A window coefficient of 0 provides no neighbors, a coefficient of 1 provides only first-order", + "bbox": [ + 89, + 795, + 483, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "neighbors, 2 also provides second-order neighbors, and so forth. Note that due to irregularities in the graph structure, the number of keys for each node is not identical as some have slightly fewer neighbors than others. This is homogenized for parallelization purposes by adding null neighbor keys and masking them.", + "bbox": [ + 511, + 90, + 903, + 181 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For relative position information, a learned relative position bias is added to the attention map. Since having a learned parameter per query-key pair is both memory expensive and harmful in terms of inductive bias, we took a shared-weights approach. For each node $i \\in [1, L]$ , we measure the angular difference $(\\Delta \\phi_{i,k}, \\Delta \\theta_{i,k})$ for each key $k \\in [1, K]$ and normalize them by the max deltas. This is only computed once during initialization. During runtime, the normalized deltas (in $[-1, 1]$ ) are used to sample learned parameters from a $7 \\times 7$ window. This sampling from a non-linear 2D function acts as a cheap relative positional encoding for the query-key pairs. Fig. 6 illustrates the relative position implementation. Given the blue center points, its first-order neighbors are depicted in green, and second-order neighbors in yellow, are placed in their relative position. A $7 \\times 7$ grid, represented by the black points, represents the learned weights of the grid. Each neighbor gets its positional encoding by bilinearly interpolating between the four nearby weights.", + "bbox": [ + 511, + 181, + 906, + 470 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 511, + 481, + 648, + 500 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We evaluate our method by comparing it with state-of-the-art methods for depth estimation and semantic segmentation. For this purpose, we used the Stanford2D3D [4] and Structured3D [53] datasets. Both contain RGB, depth, and semantic segmentation data. We trained our model and all baselines using the training configuration of PanoFormer [36]. For more information on the training protocol please refer to the supplementary.", + "bbox": [ + 511, + 507, + 905, + 628 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We compared against ERP methods [20, 36, 48], spherical methods [8, 49], and Elite360D [1], that utilizes a combination of ERP with a low-resolution icosphere. We did not compare against patch-based solutions, as the difference in compute resources makes an unfair comparison.", + "bbox": [ + 511, + 628, + 905, + 704 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. Design Decisions", + "text_level": 1, + "bbox": [ + 511, + 712, + 679, + 728 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Certain design decisions were made for the evaluation section. Some due to fairness considerations when comparing to the baselines, and others after testing various alternatives.", + "bbox": [ + 511, + 734, + 905, + 779 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For fairness reasons, it is important to compare methods of similar size. We configured all models to have 4 down-sampling stages, and a depth (number of blocks per stage) of 2 for each scale. Tab. 2 show the parameter and flop count per model. We gave PanoFormer [36], EGFormer [48], Elite360D [1] and HerRUNet[49] base embedding dim of 32, SFSS [20] 48, HealSWIN [8] 64. The resnet18 backbone was used for Elite360D. This resulted in as equal as", + "bbox": [ + 511, + 780, + 906, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "944", + "bbox": [ + 485, + 944, + 511, + 957 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/7916fd44978b01b6fc6abb7a7dd590c00bc5488937e8192b7ba2d1aab6efc49e.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelRes. (#Pixels)ParamsFlops
PanoFormer [36]256×512(131K)14.5M11.8G
EGFormer [48]256×512(131K)15.2M15.6G
SFSS [20]256×512(131K)15.1M18.9G
Elite360D [1]256×512†(136K)14.7M13.6G
HealSWIN [8]12·47(196K)12.0M39.0G
HexRUnet [49]10·47+2(164K)14.0M12.4G
PanoFormer [36]512×1024(524K)14.5M44.7G
EGFormer [48]512×1024(524K)15.2M65.8G
SFSS [20]512×1024(524K)15.1M80.3G
Elite360D [1]512×1024†(529K)14.7M51.4G
HealSWIN [8]12·48(786K)12.0M156.1G
", + "bbox": [ + 93, + 88, + 480, + 292 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/3114f51dfeab11e30dd10e1e9b65ef5e0da7b17688df7d978e8fce6acffac695.jpg", + "table_caption": [ + "Table 2. Resolution and parameter count of baseline models. ${}^{ \\dagger }$ Elite360D uses additional icosphere of rank 4 (5120 data points)." + ], + "table_footnote": [], + "table_body": "
#RankTypeCheadCwinRes.ParamsFlops
6hex1141K11.2M2.5G
6ico1182K11.2M4.9G
7hex11164K11.2M9.9G
7ico11328K11.2M19.1G
7hex21164K14.9M13.0G
(1)7hex22164K14.9M13.1G
(2)8hex22655K14.9M52.7G
", + "bbox": [ + 96, + 345, + 475, + 483 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 3. Resolution (total data points) and parameter count of various configurations of our model. Selected configurations in green.", + "bbox": [ + 89, + 494, + 482, + 523 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "possible resolution, parameter count, and flops that the different architectures allow. We analyzed input resolutions of $256 \\times 512$ and $512 \\times 1024$ for the grid models. Resolution of the spherical models was chosen to be as close as possible to those pixel counts. This was with subdivision ranks 7 and 8 respectively. Also for fairness, none of the methods used pretrained weights. To evaluate all models fairly and on equal terms, the predictions in the various representations were projected uniformly to the sphere surface. Unlike evaluating directly on ERP, this provides an evaluation that evenly averages over the sphere surface and is not biased to extreme vertical angles.", + "bbox": [ + 88, + 551, + 482, + 733 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We followed these guidelines to select the sphere setting with the closest number of nodes, total parameters and flops to the baselines. These stats are reported in Tab. 2 for the existing methods, while Tab. 3 lists variations of our method concerning: sphere rank, node type, head, and window coefficients $(C_{head}, C_{win})$ . As can be seen, spheres with a rank of 6 have a very low resolution, regardless of the node type. For spheres with rank 7, we have found that the \"vertex\" node type (with $164\\mathrm{K}$ nodes) provides a resolution that is closest to the $256\\times 512$ (total of $131\\mathrm{K}$ pixels) resolution, and rank 8 is the closest to $512\\times 1024$ . The base configura", + "bbox": [ + 88, + 734, + 482, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "tion of our method has fewer parameters and flops than the baselines. This is largely because, unlike most baselines, our method does not contain spatial layers, such as convolution kernels, outside of the attention layer. As a result, the base configuration of our method is less powerful. This is also further addressed in the ablation section (Sec. 5.4). To minimize this difference, we have found that increasing the head coefficient results in a model with slightly more parameters to match the minimum of the baselines. Increasing the window coefficient does not add any parameters and an insignificant amount of flops. Finally, we have selected two configurations for comparison with the baselines. One with rank 7 to be compared to $256 \\times 512$ models, and one in rank 8 for larger resolution. These are highlighted in Tab. 3.", + "bbox": [ + 511, + 90, + 906, + 303 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Depth Estimation", + "text_level": 1, + "bbox": [ + 511, + 314, + 684, + 330 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In depth estimation, the task is a dense prediction of the pixels depths. We evaluated the depth prediction up to 10 meters for Stanford2D3D and 5 meters for Structured3D. Following PanoFormer [36], we used the BerhuLoss [25] as a differentiable criterion to train the models. We then measured the Mean Absolute Error (MAE), Mean Relative Error (MRE), Root Mean Squared Error (RMSE), and $\\delta_{1}$ accuracy, which are standard evaluation metrics for depth estimation. Quantitative results on $256\\times 512$ are reported in Tab. 4. Our method is shown to outperform the baselines on all but one metric, where it was second. In evaluation on Stanford2D3D in $512\\times 1024$ (Tab. 5), the gap is even larger, and our method significantly improved while the baselines did not. Fig. 7 presents sample results. Evidently, our results are considerably sharper than the baselines, especially around the center, and handle distortions around the poles better. We argue that this is due to spherical representation having a better effective resolution at the center, and does not suffer from distortions. We have also found depth predictions of most baselines to have a boundary misalignment effect where the $360^{\\circ}$ and $0^{\\circ}$ meet, which our method has no such effect. This is further explored in the supplement.", + "bbox": [ + 511, + 337, + 906, + 670 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.3. Semantic Segmentation", + "text_level": 1, + "bbox": [ + 511, + 681, + 730, + 698 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In semantic segmentation, the task is to predict a semantic class per image pixel. Stanford2D3D has 13 categories, while Structured3D has 40. We used the standard Categorical Cross Entropy loss, where the background category was ignored. For evaluation, we measured the global pixel accuracy and the mIoU percentage, which are standard metrics for semantic segmentation. Results are depicted in Tab. 4. Tab. 5 shows additional high-resolution setup. In this task as well, our method outperformed the baselines on all metrics. Fig. 8 presents typical samples. Similarly, our semantic segmentation results present the same improvement in terms of distortion handling and improvement around the center of the image we have observed for depth estimation.", + "bbox": [ + 511, + 704, + 908, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "945", + "bbox": [ + 485, + 944, + 509, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/883fec8c51991e9240ced3066678396503e48a9cfdbc5df8145ad7e3fced6395.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelDepth EstimationSemantic Segmentation
Stanford2D3DStructured3DStanford2D3DStructured3D
MAE↓MRE↓RMSE↓δ1↑MAE↓MRE↓RMSE↓δ1↑Acc.↑mIoU↑Acc.↑mIoU↑
PanoFormer [36].174.078.45192.5.154.051.33094.883.160.694.949.7
EGFormer [48].170.075.45693.1.150.049.32895.286.566.495.051.5
SFSS [20].179.081.45592.2.155.051.33095.086.968.295.251.9
HexRUnet [49].201.090.51190.1----81.756.1--
HealSWIN [8].189.084.48292.2----85.563.2--
Elite360D [1].169.069.44893.5.147.046.32195.987.471.495.352.0
OURS (Tab. 3 (1)).165.071.43294.0.142.045.31496.488.672.295.853.0
", + "bbox": [ + 94, + 88, + 903, + 275 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/48a9c203806e696fc9c424112604ae65f6c90c832957c5f96adfd56ba38c2663.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 114, + 313, + 225, + 398 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f61dcfec21972ec5f190feb2b822142c808ace57fa2babdd1d961c3767588707.jpg", + "image_caption": [ + "RGB" + ], + "image_footnote": [], + "bbox": [ + 114, + 407, + 225, + 492 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/61d52f096ca5f453ef288d7fd3c02e9da02d25f290fe055639f3e829c045e66d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 246, + 314, + 357, + 398 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/81f65050a0b7937899498948208c583664f1317e9d70708b687fbc2a35b0f745.jpg", + "image_caption": [ + "PanoFormer" + ], + "image_footnote": [], + "bbox": [ + 246, + 407, + 357, + 492 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/146a8a60d2f4f1ebcb3efa6f903b6c3077998c92f54e62c34cee9be6f32d0205.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 377, + 314, + 488, + 398 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/69caf7a2d787868cc97039069a98990ddb52e3a7b9c06ecfd71af3c664b8f32e.jpg", + "image_caption": [ + "EGFormer" + ], + "image_footnote": [], + "bbox": [ + 377, + 407, + 488, + 492 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e98bf00c09c4957bf43d0a7396a607a9d7bd2c326845d8f7e6341db9001bb2f0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 509, + 314, + 619, + 398 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6d2068fa0fa4383216e986cf4e1b3c60fe0b5bc29963157b73bd353092e89c11.jpg", + "image_caption": [ + "Elite360D" + ], + "image_footnote": [], + "bbox": [ + 509, + 407, + 619, + 492 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3a226aebdd51df120d90aedb810585abb9cb99e792bbd4f315ab59c2eff06e2d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 640, + 314, + 751, + 398 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/847558752fbe02aa5ca306c897621a01790fb993b22ad284ff9af5fbf589ee70.jpg", + "image_caption": [ + "Ours", + "Figure 7. Depth Estimation. Top: Stanford2D3D. Bottom: Structured3D." + ], + "image_footnote": [], + "bbox": [ + 640, + 407, + 751, + 492 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d2b31beac34d5041bc4bb3f6eacea12e739504f4adc94db6f1279b80ad738d1b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 771, + 314, + 880, + 398 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/460b1984ab08eb1df6c2e89296086de7c91eb23bfd9651bd13d245dd30ebbfec.jpg", + "image_caption": [ + "GT" + ], + "image_footnote": [], + "bbox": [ + 771, + 407, + 880, + 492 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/9e512472d86e9f4e85b04c6541776abf3951df4a909873003a9686e1dcb3d024.jpg", + "table_caption": [ + "Table 4. Quantitative comparison on sphere rank 7 $(256\\times 512)$ resolution." + ], + "table_footnote": [], + "table_body": "
ModelMAE↓MRE↓δ1↑Acc.↑IoU↑
PanoFormer.167.07293.782.455.6
Elite360D.181.07793.284.563.3
OURS (Tab. 3 (2)).147.06594.089.171.5
", + "bbox": [ + 102, + 561, + 470, + 650 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 5. Comparison on Stanford2D3D in rank 8 (512 × 1024).", + "bbox": [ + 99, + 659, + 472, + 674 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.4. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 702, + 243, + 718 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To explore the contribution of each one of the elements we have conducted several ablation studies in which multiple variants of our method are tested. These tests were conducted in the context of depth estimation on Stanford2D3D.", + "bbox": [ + 89, + 724, + 483, + 786 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Positional Encoding Two main components of the method are the global absolute and relative bias position encodings. The vertical global position encoding injects vertical position while remaining rotation equivariant horizontally. The relative bias within the self-attention provides relative information. We compared various configurations", + "bbox": [ + 89, + 809, + 483, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "that omit one or both of the encodings. Results can be seen in Tab. 6. Evidently, the absence of positional encoding, either global or relative, drastically impacts the performance of the model. Also, relative positional bias proved more critical than the global one.", + "bbox": [ + 511, + 563, + 906, + 638 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Scaling Coefficients Two key modifiable scaling coefficients, the \"window\" $(C_{win})$ and \"head\" $(C_{head})$ , affect the capacity of the attention module. The former affects the size of the attention window to incorporate further nodes in the graph. The latter increases the capacity of the attention head by increasing the head dimension. We tested the effect of these coefficients by gradually increasing them separately and together. As shown in Tab. 7. Increasing the window up until $C_{win} = 2$ improves the performance. Interestingly, further increase of the window did not show an improvement. This could suggest that such large windows interfere with convergence. As expected, increasing the attention head also improves results, even up to $C_{head} = 4$ . Unsurprisingly, this suggests that the model could benefit from an overall increase in parameters, which is beyond the scope of this work.", + "bbox": [ + 511, + 657, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "946", + "bbox": [ + 485, + 945, + 511, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2829777ed08c2ede78485f2504f1e76af3780e3965f982cc61abaaab25f7d3fd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 114, + 89, + 225, + 176 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/fb935d7211132be200d086b06a041d3db4c6b9b4c9ee1a05288287a065575d9d.jpg", + "image_caption": [ + "RGB" + ], + "image_footnote": [], + "bbox": [ + 114, + 184, + 225, + 268 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/63b7beb973642fd6ec89d2dafe25684b125dc5cf9dc1f78cdf2e6156edc59bca.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 246, + 90, + 357, + 176 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8717697ffd72d63f5b9f354509a58f7c589d06fb860685d8a87af3fac381f338.jpg", + "image_caption": [ + "PanoFormer" + ], + "image_footnote": [], + "bbox": [ + 246, + 184, + 357, + 268 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4854a2a5d9f9979fbcbe044df53129d2313bb102c99e509afa1ba0b5da6fa22b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 375, + 90, + 488, + 176 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e2775a91bebf7c1381a5d8971f199d0e10688445fe225cb4e71c862d58bd80dc.jpg", + "image_caption": [ + "EGFormer" + ], + "image_footnote": [], + "bbox": [ + 377, + 184, + 488, + 268 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ae84e928b8b10506b2f528ffde90bc6d55c515d033fa25673b607b823cd9f552.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 509, + 90, + 620, + 176 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/666c629d97f2b11d794caaf9524bd78e065150812f4275d6581a60339fb03807.jpg", + "image_caption": [ + "Elite360D", + "Figure 8. Semantic Segmentation. Top: Stanford2D3D. Bottom: Structured3D." + ], + "image_footnote": [], + "bbox": [ + 509, + 184, + 619, + 268 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8a33613f0f8ae52c2b5aeed34a043d5a4e627d4d45821277d80eda194492987e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 640, + 90, + 751, + 176 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/cd16e1dfb09ad36aa9b6d8d8ce4afc9cb9aadf0530096c75209a8e27e9104d02.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 640, + 184, + 751, + 268 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ccfda3f569d92a0c3df6dc14594accc46e4d6ba66ac60ef025ea2621ca0b8868.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 772, + 90, + 882, + 176 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5fc1858bd06e5119cb49e5558d2629c46e8789f140b0b339ad742a5ebb119abc.jpg", + "image_caption": [ + "GT" + ], + "image_footnote": [], + "bbox": [ + 771, + 184, + 882, + 268 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/039e914ee827a47782895981781104fd32959861557d30a902452daf4e72ed89.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
VariantAbsPosRelPosMAE↓MRE↓
No Pos. Enc.XX.326.094
No Rel. Pos. Enc.✓X.251.091
No Abs. Pos. Enc.X✓.218.088
With Pos. Enc.✓✓.189.077
", + "bbox": [ + 91, + 338, + 482, + 433 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/ab7a8df83697c709ac52db051e86e3d0131a0ec22019d0cbdcc73f62b55c4047.jpg", + "table_caption": [ + "Table 6. Ablation Study: Positional Encoding." + ], + "table_footnote": [], + "table_body": "
VariantCheadCwinMAE↓MRE↓
Base variant11.189.077
No window10.412.122
Medium window12.175.073
Large window13.180.078
Medium head21.171.075
Large head41.162.067
Conf. (1) Tab. 322.165.071
", + "bbox": [ + 91, + 472, + 480, + 613 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 7. Ablation Study: Head and Window Coefficients.", + "bbox": [ + 112, + 623, + 459, + 637 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Discussion and Limitations", + "text_level": 1, + "bbox": [ + 89, + 662, + 341, + 679 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper introduces a pioneering approach to omnidirectional $360^{\\circ}$ perception, leveraging a geodesic polyhedron and a unique architecture designed for spherical computation. However, there is ample room for optimization and enhancement of its features. The potential for improvement spans various aspects of the model, from its core architecture to its computational efficiency.", + "bbox": [ + 89, + 688, + 482, + 794 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "One of the promising avenues for future work lies in exploring additional domains where this architecture could be applied. Beyond the realms of depth estimation and semantic segmentation in $360^{\\circ}$ imagery, spherical data representation is also prominent in other areas such as medical imaging, where spherical structures are common, and geographical data analysis, which often deals with global-scale data", + "bbox": [ + 89, + 795, + 483, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "that naturally fit a spherical model. The adaptability of our approach to these varied domains could pave the way for groundbreaking applications and methodologies.", + "bbox": [ + 511, + 340, + 903, + 386 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "While the current work has focused on supervised classification and regression tasks, extending this model to generative models presents an exciting frontier. Generative models could leverage the spherical data representation for creating highly realistic and comprehensive 3D environments among other applications. The unique challenges of generative tasks on spherical domains, such as ensuring continuity and avoiding distortions, call for accurate modeling of the type we present here.", + "bbox": [ + 511, + 387, + 905, + 523 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Another area for improvement is the computational efficiency of the model, particularly regarding its GPU usage. Despite the number of parameters and flops being on par with baseline models, our approach exhibits slower performance ( $\\sim 30\\%$ slower than PanoFormer). This is largely because the gathering operations on the sphere graph are not optimized with dedicated CUDA functions in PyTorch, leading to less efficient execution. Addressing this engineering challenge could significantly enhance the model's appeal by reducing both the GPU footprint and runtime.", + "bbox": [ + 511, + 523, + 905, + 675 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 691, + 633, + 707 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We have presented a novel method for omnidirectional perception that successfully utilizes spherically represented data. The method addresses the underlying spherical structure, which comes up in three different places: in the down and upsampling operators, in the positional encoding, and, above all, in the definition geometric representation and that of the neighborhoods for the local attention operator.", + "bbox": [ + 511, + 717, + 905, + 824 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our proposed method relies on the icosphere representation and utilizes its favorable properties to derive sphere-based operators. Using these, we surpass the state of the art in omnidirectional perception in both depth estimation and semantic segmentation.", + "bbox": [ + 511, + 825, + 905, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "947", + "bbox": [ + 485, + 944, + 509, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 91, + 90, + 258, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported by the Tel Aviv University Center for AI and Data Science (TAD). The contribution of the first author is part of a PhD thesis at Tel Aviv University.", + "bbox": [ + 89, + 114, + 485, + 161 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 176, + 187, + 193 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Hao Ai and Lin Wang. Elite360d: Towards efficient 360 depth estimation via semantic-and distance-aware biprojection fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9926-9935, 2024. 5, 6, 7", + "[2] Hao Ai, Zidong Cao, Yan-Pei Cao, Ying Shan, and Lin Wang. Hrdfuse: Monocular 360deg depth estimation by collaboratively learning holistic-with-regional depth distributions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13273-13282, 2023. 2", + "[3] Georgios Albanis, Nikolaos Zioulis, Petros Drakoulis, Vasileios Gkitsas, Vladimiros Sterzentenko, Federico Alvarez, Dimitrios Zarpalas, and Petros Daras. Pano3d: A holistic benchmark and a solid baseline for $360^{\\circ}$ depth estimation. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 3722-3732, 2021. 1", + "[4] Iro Armeni, Sasha Sax, Amir R Zamir, and Silvio Savarese. Joint 2d-3d-semantic data for indoor scene understanding. arXiv preprint arXiv:1702.01105, 2017. 1, 5", + "[5] Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. Neural machine translation by jointly learning to align and translate. arXiv preprint arXiv:1409.0473, 2014. 2", + "[6] Herbert Busemann. The geometry of geodesics. Courier Corporation, 2012. 1, 3", + "[7] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 2, 4", + "[8] Oscar Carlsson, Jan E Gerken, Hampus Linander, Heiner Spieß, Fredrik Ohlsson, Christoffer Petersson, and Daniel Persson. Heal-swin: A vision transformer on the sphere. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6067-6077, 2024. 1, 2, 5, 6, 7", + "[9] Angel Chang, Angela Dai, Thomas Funkhouser, Maciej Halber, Matthias Niessner, Manolis Savva, Shuran Song, Andy Zeng, and Yinda Zhang. Matterport3d: Learning from rgbd data in indoor environments. International Conference on 3D Vision (3DV), 2017. 1", + "[10] Bowen Cheng, Alex Schwing, and Alexander Kirillov. Per-pixel classification is not all you need for semantic segmentation. Advances in Neural Information Processing Systems, 34:17864-17875, 2021. 2, 4", + "[11] Hsien-Tzu Cheng, Chun-Hung Chao, Jin-Dong Dong, Hao-Kai Wen, Tyng-Luh Liu, and Min Sun. Cube padding for weakly-supervised saliency prediction in 360 videos. In Pro" + ], + "bbox": [ + 93, + 202, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ceedings of the IEEE conference on computer vision and pattern recognition, pages 1420-1429, 2018. 2", + "[12] Benjamin Coors, Alexandru Paul Condurache, and Andreas Geiger. Spherenet: Learning spherical representations for detection and classification in omnidirectional images. In Proceedings of the European conference on computer vision (ECCV), pages 518-533, 2018. 1, 2", + "[13] Michael Defferrard, Martino Milani, Frédérique Gusset, and Nathanaël Perraudin. DeepSphere: a graph-based spherical CNN. In International Conference on Learning Representations, 2020. 2", + "[14] Aleksandar M Dimitrijevic, Martin Lambers, and Dejan Rancic. Comparison of spherical cube map projections used in planet-sized terrain rendering. Facta Universitatis, Series: Mathematics and Informatics, 31(2):259-297, 2016. 1, 2", + "[15] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021. 2, 4", + "[16] Marc Eder, Mykhailo Shvets, John Lim, and Jan-Michael Frahm. Tangent images for mitigating spherical distortion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12426-12434, 2020. 2", + "[17] Lance Flavell. Uv mapping. In *Beginning Blender: Open Source 3D Modeling*, Animation, and Game Design, pages 97-122. Springer, 2010. 1, 2", + "[18] Michael S Floater. Generalized barycentric coordinates and applications. Acta Numerica, 24:161-214, 2015. 4", + "[19] Krzysztof M Górski, E Hivon, and BD Wandelt. Analysis issues for large cmb data sets. Proceedings: Evolution of Large Scale Structure-Garching, 1998. 3", + "[20] Suresh Guttikonda and Jason Rambach. Single frame semantic segmentation using multi-modal spherical images. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 3222-3231, 2024. 1, 2, 5, 6, 7", + "[21] Shi-Min Hu, Zheng-Ning Liu, Meng-Hao Guo, Jun-Xiong Cai, Jiahui Huang, Tai-Jiang Mu, and Ralph R Martin. Subdivision-based mesh convolution networks. ACM Transactions on Graphics (TOG), 41(3):1-16, 2022. 2", + "[22] Chiyu Jiang, Jingwei Huang, Karthik Kashinath, Philip Marcus, Matthias Niessner, et al. Spherical cnns on unstructured grids. International Conference on Learning Representations, 2019. 2", + "[23] Hualie Jiang, Zhe Sheng, Siyu Zhu, Zilong Dong, and Rui Huang. Unifuse: Unidirectional fusion for 360 panorama depth estimation. IEEE Robotics and Automation Letters, 6 (2):1519-1526, 2021. 1", + "[24] Benjamin Keinert, Matthias Innmann, Michael Sanger, and Marc Stamminger. Spherical fibonacci mapping. ACM Transactions on Graphics (TOG), 34(6):1-7, 2015. 1, 3", + "[25] Iro Laina, Christian Rupprecht, Vasileios Belagiannis, Federico Tombari, and Nassir Navab. Deeper depth prediction" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "948", + "bbox": [ + 485, + 945, + 511, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "with fully convolutional residual networks. In 2016 Fourth international conference on 3D vision (3DV), pages 239-248. IEEE, 2016. 6", + "[26] Martin Lambers. Survey of cube mapping methods in interactive computer graphics. The Visual Computer, 36(5): 1043-1051, 2020. 1", + "[27] Yeonkun Lee, Jaeseok Jeong, Jongseob Yun, Wonjune Cho, and Kuk-Jin Yoon. Spherephd: Applying cnns on a spherical polyhedron representation of 360deg images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9181-9189, 2019. 1, 2", + "[28] Yuyan Li, Yuliang Guo, Zhixin Yan, Xinyu Huang, Ye Duan, and Liu Ren. Omnifusion: 360 monocular depth estimation via geometry-aware fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2801-2810, 2022. 1, 2", + "[29] Thang Luong, Hieu Pham, and Christopher D. Manning. Effective approaches to attention-based neural machine translation. In Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing. Association for Computational Linguistics, 2015. 2", + "[30] Ronald Miller. An equi-rectangular map projection. Geography, pages 196–201, 1949. 1", + "[31] Hongbin Pei, Bingzhe Wei, Kevin Chen-Chuan Chang, Yu Lei, and Bo Yang. Geom-gcn: Geometric graph convolutional networks. International Conference on Learning Representations, 2020. 2", + "[32] Giovanni Pintore, Marco Agus, Eva Almansa, Jens Schneider, and Enrico Gobbetti. Slicenet: deep dense depth estimation from a single indoor panorama using a slice-based representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11536-11545, 2021. 1", + "[33] Alec Radford, Karthik Narasimhan, Tim Salimans, Ilya Sutskever, et al. Improving language understanding by generative pre-training. 2018. 4", + "[34] René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. In Proceedings of the IEEE/CVF international conference on computer vision, pages 12179-12188, 2021. 2, 4", + "[35] Manuel Rey-Area, Mingze Yuan, and Christian Richardt. 360monodepth: High-resolution 360deg monocular depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3762-3772, 2022. 1, 2", + "[36] Zhijie Shen, Chunyu Lin, Kang Liao, Lang Nie, Zishuo Zheng, and Yao Zhao. Panofformer: Panorama transformer for indoor $360^{\\circ}$ depth estimation. In European Conference on Computer Vision, pages 195-211. Springer, 2022. 1, 2, 5, 6, 7", + "[37] Yu-Chuan Su and Kristen Grauman. Learning spherical convolution for fast features from 360 imagery. Advances in Neural Information Processing Systems, 30, 2017. 2", + "[38] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Hohonet: 360 indoor holistic understanding with latent horizontal features. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2573-2582, 2021. 1" + ], + "bbox": [ + 91, + 90, + 483, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[39] Keisuke Tateno, Nassir Navab, and Federico Tombari. Distortion-aware convolutional filters for dense prediction in panoramic images. In Proceedings of the European Conference on Computer Vision (ECCV), pages 707-722, 2018. 2", + "[40] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 2, 4", + "[41] Petar Velicković, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Liò, and Yoshua Bengio. Graph attention networks. In International Conference on Learning Representations, 2018. 2", + "[42] Fu-En Wang, Yu-Hsuan Yeh, Min Sun, Wei-Chen Chiu, and Yi-Hsuan Tsai. Bifuse: Monocular 360 depth estimation via bi-projection fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 462–471, 2020. 1, 2", + "[43] Fu-En Wang, Yu-Hsuan Yeh, Yi-Hsuan Tsai, Wei-Chen Chiu, and Min Sun. Bifuse++: Self-supervised and efficient bi-projection fusion for 360 depth estimation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(5): 5448-5460, 2022. 1, 2", + "[44] Zhendong Wang, Xiaodong Cun, Jianmin Bao, Wengang Zhou, Jianzhuang Liu, and Houqiang Li. Uformer: A general u-shaped transformer for image restoration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 17683-17693, 2022. 2, 3, 4", + "[45] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. Advances in Neural Information Processing Systems, 34:12077-12090, 2021. 2, 4", + "[46] Qingsong Yan, Qiang Wang, Kaiyong Zhao, Bo Li, Xiaowei Chu, and Fei Deng. Sphereddepth: Panorama depth estimation from spherical domain. In 2022 International Conference on 3D Vision (3DV), pages 1-10. IEEE, 2022. 1, 2", + "[47] Ilwi Yun, Hyuk-Jae Lee, and Chae Eun Rhee. Improving 360 monocular depth estimation via non-local dense prediction transformer and joint supervised and self-supervised learning. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3224-3233, 2022. 1, 2", + "[48] Ilwi Yun, Chanyong Shin, Hyunku Lee, Hyuk-Jae Lee, and Chae Eun Rhee. Egformer: Equirectangular geometry-biased transformer for 360 depth estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 6101-6112, 2023. 1, 2, 5, 6, 7", + "[49] Chao Zhang, Stephan Liwicki, William Smith, and Roberto Cipolla. Orientation-aware semantic segmentation on icosahedron spheres. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3533-3541, 2019. 1, 5, 6, 7", + "[50] Chao Zhang, Sen He, and Stephan Liwicki. A spherical approach to planar semantic segmentation. In BMVC, 2020. 1", + "[51] Jiaming Zhang, Kailun Yang, Chaoxiang Ma, Simon Reiß, Kunyu Peng, and Rainer Stiefelhagen. Bending reality: Distortion-aware transformers for adapting to panoramic se" + ], + "bbox": [ + 516, + 90, + 903, + 901 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "949", + "bbox": [ + 486, + 945, + 511, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "matic segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16917-16927, 2022. 1", + "[52] Si Zhang, Hanghang Tong, Jiejun Xu, and Ross Maciejewski. Graph convolutional networks: a comprehensive review. Computational Social Networks, 6(1):1-23, 2019. 2", + "[53] Jia Zheng, Junfei Zhang, Jing Li, Rui Tang, Shenghua Gao, and Zihan Zhou. Structured3d: A large photo-realistic dataset for structured 3d modeling. In Proceedings of The European Conference on Computer Vision (ECCV), 2020. 1, 5", + "[54] Nikolaos Zioulis, Antonis Karakottas, Dimitrios Zarpalas, and Petros Daras. Omnidepth: Dense depth estimation for indoors spherical panoramicas. In Proceedings of the European Conference on Computer Vision (ECCV), pages 448-465, 2018. 2" + ], + "bbox": [ + 91, + 90, + 480, + 316 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "950", + "bbox": [ + 486, + 945, + 509, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/SphereUFormer_ A U-Shaped Transformer for Spherical 360 Perception/d3dc320a-7961-421e-ae82-56124645a3a0_model.json b/2025/SphereUFormer_ A U-Shaped Transformer for Spherical 360 Perception/d3dc320a-7961-421e-ae82-56124645a3a0_model.json new file mode 100644 index 0000000000000000000000000000000000000000..9cc98deeb3ff15f832aca810a50a81471c4a9070 --- /dev/null +++ b/2025/SphereUFormer_ A U-Shaped Transformer for Spherical 360 Perception/d3dc320a-7961-421e-ae82-56124645a3a0_model.json @@ -0,0 +1,2642 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.135, + 0.131, + 0.865, + 0.154 + ], + "angle": 0, + "content": "SphereUFormer: A U-Shaped Transformer for Spherical 360 Perception" + }, + { + "type": "text", + "bbox": [ + 0.306, + 0.182, + 0.464, + 0.218 + ], + "angle": 0, + "content": "Yaniv Benny \nTel Aviv University" + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.182, + 0.692, + 0.218 + ], + "angle": 0, + "content": "Lior Wolf \nTel Aviv University" + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.269, + 0.327, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.301, + 0.486, + 0.528 + ], + "angle": 0, + "content": "This paper proposes a novel method for omnidirectional \\(360^{\\circ}\\) perception. Most common previous methods relied on equirectangular projection. This representation is easily applicable to 2D operation layers but introduces distortions into the image. Other methods attempted to remove the distortions by maintaining a sphere representation but relied on complicated convolution kernels that failed to show competitive results. In this work, we introduce a transformer-based architecture that, by incorporating a novel \"Spherical Local Self-Attention\" and other spherically-oriented modules, successfully operates in the spherical domain and outperforms the state-of-the-art in \\(360^{\\circ}\\) perception benchmarks for depth estimation and semantic segmentation. Our code is available at https://github.com/yanivbenny/sphere_uformer." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.558, + 0.222, + 0.573 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.583, + 0.483, + 0.748 + ], + "angle": 0, + "content": "Monocular omnidirectional \\(360^{\\circ}\\) perception is an important setting, as it allows for a full field-of-view (FOV) receptive field for the underlying model. Consequently, an entire room layout can be inferred, for example. In contrast to regular images, a \\(360^{\\circ}\\) image has no boundaries. Instead, the image is horizontally cyclical, and converges to singular points at the vertical extreme points. This suggests that this domain requires a different treatment, and that best practices that work well with regular images might not have the same effect on \\(360^{\\circ}\\) images. To tackle this task, multiple topics need to be considered." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.81 + ], + "angle": 0, + "content": "How should the data be represented? What kind of distortions and irregularities are created by this representation? How will a model process this representation? How to deal with these distortions and irregularities?" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.483, + 0.902 + ], + "angle": 0, + "content": "These questions have attracted researchers to construct multiple datasets [3, 4, 9, 53] where depth estimation and semantic segmentation are the main tasks, and to formulate many solutions [20, 23, 28, 32, 35, 36, 38, 42, 43, 47, 48, 51], which propose novel approaches to solving these tasks. A common pattern in previous works is that they rep" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.27, + 0.908, + 0.543 + ], + "angle": 0, + "content": "resent the data by projecting it to a 2D plane in different ways, usually equirectangular projection [30], cube mapping [26], or patch cropping. Each approach has its advantages and disadvantages, but the common motivation is to represent the image in a grid formation. This enables grid architectures, such as Convolutional Neural Networks or Vision Transformers to easily be applied. This comes at a cost of distortions, irregularities, or limited FOV, that are caused by the selected projection. A less common alternative representation that has been proposed is to represent the image in its original sphere domain by discretizing the sphere surface in some manner [6, 14, 17, 24]. The motivation is to maintain an undistorted representation of the \\(360^{\\circ}\\) image. However, this representation formed architectural challenges that require novel solutions. Some previous works [8, 12, 27, 46, 49, 50] addressed these challenges in various manners, but they fail to compete with current state-of-the-art solutions that follow a grid representation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.554, + 0.909, + 0.903 + ], + "angle": 0, + "content": "In this work, we tackle the drawbacks of spherical representations. Specifically, we introduce SphereUFormer, which is a novel U-Shaped Transformer architecture tailored for spherical \\(360^{\\circ}\\) perception. By directly operating on a spherical domain without distortion-inducing projections, SphereUFormer addresses the core challenges of omnidirectional perception. Our approach builds on the strengths of transformer models, incorporating modifications that allow it to efficiently process and understand spherical data. This includes the development of spherical-specific operations such as localized self-attention mechanism that respects the geometry of the sphere, innovative up/downsampling techniques that maintain the integrity of spherical data throughout the model, and positional encoding scheme that matches the domain properties. SphereUFormer outperforms the state of the art in the field of \\(360^{\\circ}\\) perception, providing a robust and efficient tool for a wide range of applications that require comprehensive understanding of spherical environments. In Sec. 3 we describe this representation in detail. We then further describe a new method in Sec. 4, which utilizes this representation and performs \\(360^{\\circ}\\) perception on spherical representations, and in Sec. 5, we analyze and compare its results." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.513, + 0.957 + ], + "angle": 0, + "content": "940" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.09, + 0.233, + 0.106 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.117, + 0.485, + 0.494 + ], + "angle": 0, + "content": "Omnidirectional perception. There have been many advances, novelties, and research verticals in omnidirectional perception over the years. Initial works attempted standard convolutional networks on an equirectangular projection [54] (ERP). Although promising, the results were affected by the distortions created by the projection. To better handle these distortions, later works suggested technical adjustments to the convolution kernel such as Spherical Convolutions [37] or Distortion-Aware Convolutions [12, 39]. The latter sample the convolution features from a tangent plane. Other solutions attempted to eliminate distortions by sampling the image into smaller FOV patches [16, 28, 35, 47] followed by a post-processing stage that fuses the independent results. This, however, is at a cost of a restricted receptive field per patch and possible cutoff of crucial information, which either affects results or requires high overlap between patches. Due to the advantages and limitations of each method, [2, 42, 43] suggested a combined approach that fuses the ERP path with the individual patches at some stage of the network. Alternative approaches avoided the use of projection, and instead, operated directly on the sphere by constructing a spherical mesh. Usually either an icosphere [13, 22, 27, 46] or healpix [8, 13], and apply Graph Convolution [52], Mesh Convolution [21], or some custom convolution-like operation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.516, + 0.484, + 0.728 + ], + "angle": 0, + "content": "Vision Transformers. ViT's [15] have disrupted the field of computer vision, from general-purpose architectures [34, 44] to specific purposes [7, 10, 45]. The attention mechanism [5, 29, 40] allows for dynamic pooling of features based on relevance and structural position. ViT's benefit from an efficient large context window compared to their CNN counterparts, which allows them to more easily learn long-distance relations between features. While earlier omnidirectional perception works proposed solutions based on CNN's, recent contributions replaced the convolution kernels with local attention of various forms [20, 36, 48] with favorable results. This work follows this direction. It utilizes the foundation of UFormer [44] and adapts it to spherical representations with custom attention layers." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Graph Attention. Graph Convolution Networks [52] (GCN) are neural networks that apply convolution-like operations on data represented as a graph. A graph convolution operation aggregates the results of local operations between a node and each of its neighbors. Over multiple layers, this allows for information to pass between distant nodes in the graph and the update of these nodes. Graph Attention Networks [41] were proposed to enhance Graph Convolution Networks, thereby improving the message-passing capabilities of the GCN. The attention mechanism" + }, + { + "type": "image", + "bbox": [ + 0.533, + 0.092, + 0.605, + 0.139 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.558, + 0.149, + 0.574, + 0.159 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.628, + 0.092, + 0.701, + 0.139 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.654, + 0.149, + 0.671, + 0.159 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.725, + 0.092, + 0.796, + 0.139 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.751, + 0.149, + 0.767, + 0.159 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image", + "bbox": [ + 0.821, + 0.092, + 0.892, + 0.139 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.847, + 0.147, + 0.865, + 0.159 + ], + "angle": 0, + "content": "(d)" + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.172, + 0.905, + 0.2 + ], + "angle": 0, + "content": "Figure 1. Sphere Representations. From left to right: uvsphere, cubesphere, icosphere, hexasphere." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.227, + 0.907, + 0.333 + ], + "angle": 0, + "content": "has the benefit of allowing the operations to focus on a few sets of relevant nodes and providing an efficient method to encode positional information in the graph. For Geometric Graphs [31] geometric location is encoded to inform of each node's absolute and relative position, thereby making nodes aware of their global position and their respective neighbors' relative position." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.347, + 0.729, + 0.364 + ], + "angle": 0, + "content": "3. Sphere Representation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.373, + 0.907, + 0.537 + ], + "angle": 0, + "content": "Arguably the first topic in question when attempting to solve a problem is how to formulate the data. For omnidirectional \\(360^{\\circ}\\) images, the data can be seen as an array of values \\(v_{\\phi ,\\theta}\\) assigned to a ray pointing from a center point-of-view towards \\((\\theta ,\\phi)\\) for \\(\\theta \\in [0,2\\pi ]\\) being the horizontal angle and \\(\\phi \\in [0,\\pi ]\\) the vertical. The value \\(v\\) can correspond to the RGB value, depth, object category, etc. By selecting to sample on finite such angle pairs, the sphere becomes discretized. The sampling method determines how the sampled data can be processed and how successful the solution will become." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.539, + 0.906, + 0.704 + ], + "angle": 0, + "content": "The most common and straightforward is a uvsphere [17], where points are sampled in a grid with equal spacing for \\((\\theta, \\phi)\\). See Fig. 1a. The uvsphere has a very high degree of horizontal symmetry, which is a good property. Another advantage is that it allows an unwrapping of the sphere into a 2D matrix. This unwrapping is called equirectangular projection (ERP). The downside of this representation is that the density of samples on the sphere is higher around the poles, due to the decreasing cross-section. This creates an imbalanced effective resolution over the sphere, and the projected image in 2D is distorted." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.907, + 0.901 + ], + "angle": 0, + "content": "An alternative that mitigates imbalanced resolution and distortion is cube mapping (or cubosphere) [14] (Fig. 1b), which projects the sphere onto the faces of a cube. Sample resolution is controlled by deciding how many rectangles each face of the cube is subdivided into. This removes the sampling density around the poles and the sphere can be unwrapped into six 2D images. However, two of these images are not vertically oriented and require special treatment. In addition, the transition between faces is not fluent, and complex padding methods [11] can only increase the receptive field of each image to some extent. Since the images are separated, a post-processing stage needs to fuse their predictions to form a unified prediction on the sphere. One op-" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.511, + 0.957 + ], + "angle": 0, + "content": "941" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.107, + 0.095, + 0.16, + 0.145 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.128, + 0.155, + 0.14, + 0.166 + ], + "angle": 0, + "content": "0" + }, + { + "type": "image", + "bbox": [ + 0.179, + 0.096, + 0.241, + 0.145 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.207, + 0.155, + 0.214, + 0.166 + ], + "angle": 0, + "content": "1" + }, + { + "type": "image", + "bbox": [ + 0.255, + 0.095, + 0.32, + 0.145 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.283, + 0.155, + 0.293, + 0.166 + ], + "angle": 0, + "content": "2" + }, + { + "type": "image", + "bbox": [ + 0.333, + 0.095, + 0.397, + 0.145 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.361, + 0.155, + 0.37, + 0.166 + ], + "angle": 0, + "content": "3" + }, + { + "type": "image", + "bbox": [ + 0.411, + 0.095, + 0.474, + 0.145 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.437, + 0.155, + 0.447, + 0.166 + ], + "angle": 0, + "content": "4" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.181, + 0.483, + 0.223 + ], + "angle": 0, + "content": "Figure 2. Icospheres of different ranks. An increase in rank is made by subdividing each triangle into 4 smaller triangles, and results in an increased resolution." + }, + { + "type": "table", + "bbox": [ + 0.11, + 0.241, + 0.468, + 0.294 + ], + "angle": 0, + "content": "
Rank:01234567
Faces:208032012805120~20K~82K~328K
Vertices:12421626422562~10K~41K~164K
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.305, + 0.483, + 0.334 + ], + "angle": 0, + "content": "Table 1. Icosphere Properties. Number of faces \\( (20 \\times 4^i) \\) and number of vertices \\( (10 \\times 4^i + 2) \\) per subdivision rank \\( i \\)." + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.355, + 0.162, + 0.403 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.121, + 0.41, + 0.145, + 0.421 + ], + "angle": 0, + "content": "(a1)" + }, + { + "type": "image", + "bbox": [ + 0.166, + 0.355, + 0.225, + 0.403 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.184, + 0.41, + 0.207, + 0.421 + ], + "angle": 0, + "content": "(b1)" + }, + { + "type": "image", + "bbox": [ + 0.228, + 0.355, + 0.285, + 0.403 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.245, + 0.41, + 0.268, + 0.421 + ], + "angle": 0, + "content": "(c1)" + }, + { + "type": "image", + "bbox": [ + 0.29, + 0.355, + 0.347, + 0.403 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.308, + 0.41, + 0.33, + 0.421 + ], + "angle": 0, + "content": "(a2)" + }, + { + "type": "image", + "bbox": [ + 0.351, + 0.355, + 0.41, + 0.403 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.369, + 0.41, + 0.393, + 0.421 + ], + "angle": 0, + "content": "(b2)" + }, + { + "type": "image", + "bbox": [ + 0.414, + 0.355, + 0.472, + 0.403 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.432, + 0.41, + 0.453, + 0.421 + ], + "angle": 0, + "content": "(c2)" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.434, + 0.483, + 0.463 + ], + "angle": 0, + "content": "Figure 3. Data as Icospheres. Icosphere of rank 3 (a1-c1) and rank 6 (a2-c2). (a) RGB, (b) Depth Map, and (c) Semantic Layout." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.491, + 0.483, + 0.567 + ], + "angle": 0, + "content": "tion is to operate on the cubosphere mesh instead of the 2D projection. Unfortunately, the cubosphere has irregularities at the coordinates of the cube corners and does not have a high horizontal symmetry, which makes it less optimal for this purpose." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.568, + 0.483, + 0.854 + ], + "angle": 0, + "content": "Instead of projecting the data to a grid and then having to deal with its limitations, a more elegant approach can be formulated using the sphere representation directly. Aside from the previously mentioned options, a common way to discretize a sphere is with Geodesic (icosphere, Fig. 1c) and Goldberg (hexasphere, Fig. 1d) polyhedrons [6]. An icosphere is created by repeatedly subdividing an icosahedron into smaller triangles. A hexasphere is the dual polyhedron where vertices are swapped with face normals, and edge connections with face adjacency. These types of sphere representation have a very high degree of symmetry and have a very evenly spaced sampling on the sphere surface. Due to their duality, our work follows the icosphere mesh, and icosphere/hexasphere mode is toggled by whether data points \\((\\theta, \\phi)\\) describe mesh vertices (hexasphere) or face normals (icosphere). Therefore we will sometimes refer to icospheres/hexaspheres by referring to the \"node type\" (face/vertex respectively). Other notable options for sphere sampling are Fibonacci spheres [24] and HealPix [19]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Icospheres have the highest order of symmetry, the most uniform distribution of points (perfect uniform distribution is impossible), easy setup, and an inherent up/down" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.304 + ], + "angle": 0, + "content": "sampling definition through subdivision. All these reasons make them ideal for discrete spherical representation. As Fig. 2 and Tab. 1 show, in different ranks, icospheres produce sphere discretization in various resolutions. Each increase in sphere rank divides each triangle face into four smaller ones, hereby gradually increasing the sphere resolution. An example of representing data with icospheres can be seen in Fig. 3. In low resolution, the image seems \"pixelated\" as the independent triangle faces are distinguishable. But, with a high enough resolution, the image becomes detailed and visibly smooth. Our method uses high-resolution icospheres to represent the \\(360^{\\circ}\\) image, and through up and downsampling on the sphere structure, uses low-resolution spheres as hidden layers inside the model." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.316, + 0.605, + 0.332 + ], + "angle": 0, + "content": "4. Method" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.342, + 0.905, + 0.447 + ], + "angle": 0, + "content": "This section introduces SphereUFormer, our proposed U-shaped Transformer for Spherical \\(360^{\\circ}\\) Perception. As the name suggests, the model is influenced by UFormer [44], with significant modifications to support spherically structured data. A high-level diagram is depicted in Fig. 4. The model's components are described in this section. For technical implementation details, see the supplementary." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.448, + 0.906, + 0.613 + ], + "angle": 0, + "content": "The network begins with an input projection that applies linear projection from the RGB values to latent embedding vectors. The projection output is then passed into an Encoder-Decoder network with multiple levels of self-attention modules in different sphere resolutions. We call these modules \"Spherical Attention Modules\" (SAM). At the bottom of the network lies a bottleneck SAM module that operates on the lowest resolution. Skip connections pass from the encoder to the decoder and feed high-resolution features. Finally, a linear Output Projection projects the decoder output to the output dimensions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.614, + 0.905, + 0.688 + ], + "angle": 0, + "content": "As described in Sec. 3, our solution features the icosphere as its spherical representation structure, as we have found it to be the most intuitive and to contain the most satisfying properties. But it is not limited to it, and with minor modifications supports any spherical structure." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.905, + 0.825 + ], + "angle": 0, + "content": "Our method contains many local operations on the graph, from the down and up sampling, to the self-attention mechanism. Computing the cosine similarity or finding the neighbors of a node to generate the mapping scheme is slightly costly. This would be a problem if it was performed on each iteration, as the model would be slowed down. Luckily, since the icosphere graph is fixed, these mappings only need to be computed once in advance. Therefore there is no need to recompute it in every message massing operation." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.834, + 0.702, + 0.851 + ], + "angle": 0, + "content": "4.1. Up/Down Sampling" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.905, + 0.901 + ], + "angle": 0, + "content": "In the spherical setting, upsampling and downsampling operations are performed by transforming a spherical graph into a new one with more or fewer nodes, respectively. This" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.945, + 0.512, + 0.957 + ], + "angle": 0, + "content": "942" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.107, + 0.089, + 0.902, + 0.314 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.322, + 0.908, + 0.381 + ], + "angle": 0, + "content": "Figure 4. The SphereUFormer architecture. A spherical representation is fed into the model. A linear input projection layer encodes the RGB values to latent embedding vectors. A sequence of SAM modules apply local self-attention on the spherical data along with downsampling layers that gradually reduce the resolution of the sphere. A sequence of SAM modules along with an upsampling layer and bypass skip connections decode the data. An output projection converts the latent embeddings to the output channel size." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.405, + 0.484, + 0.554 + ], + "angle": 0, + "content": "can be difficult and inefficient to compute for two arbitrary spherical polyhedrons. However, icospheres have multiple elegant and desired properties. First, an increase in resolution by one rank is the result of splitting each triangle face into four smaller ones, with one central triangle that shares its center with the large one. Second, for the same reason, each vertex in the low-resolution icosphere also exists in the high-resolution one, with new vertices created at the center of each edge connection. Third, the number of nodes/faces increases at a constant rate." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.57, + 0.484, + 0.901 + ], + "angle": 0, + "content": "All these properties make constructing up and down operations extremely straightforward and efficient. Every common pooling technique, such as max pooling, average pooling, center pooling, and interpolation, can be implemented. However, there are preferences for each setting. For downsampling, when the data nodes are represented by the faces, we have found max and center pooling to have the same overall results. When the data nodes are on the vertices (equivalent to hexosphere), max and average pooling become a little complicated, since there are both instances of five and six neighbors in the graph. Therefore, we resorted to only center pooling. For upsampling, the faces mode most simply operates with nearest upsampling, where each triangle receives the value according to the lower resolution triangle it was subdivided from. Other interpolation methods require more complicated computations and were therefore not considered. For vertices, however, interpolation is straightforward, since each new node is located precisely at the center of an existing edge. For more upscaled upsampling, Barycentric coordinates [18] interpolation can also be performed quite easily, but we did not find a reason to upscale beyond a factor of 2 at a time." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.404, + 0.761, + 0.421 + ], + "angle": 0, + "content": "4.2. Global Positional Encoding" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.427, + 0.907, + 0.653 + ], + "angle": 0, + "content": "Positional encoding is a common and critical technique in Transformers [33, 40] and Vision Transformers [7, 10, 15, 34, 44, 45]. SphereUFormer utilizes positional encoding as well, but in a manner that suits spherical representations. The node position on the unit sphere is expressed using spherical coordinates \\((\\theta, \\phi)\\). In our setting of omnidirectional perception, the data is always vertically aligned, therefore the model does not need to be vertically equivariant. However, the horizontal rotation is arbitrary, therefore it is desired to make the model horizontally equivariant to rotation and flipping. We therefore apply absolute positional encoding for the vertical position but not for the horizontal position. Horizontal position is only used in relative to another node as position bias in the Local Self-Attention described in the next section." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.654, + 0.909, + 0.761 + ], + "angle": 0, + "content": "For the vertical position, we consider \\(\\phi \\in [0,\\pi ]\\). The position was encoded by applying sinusoidal encoding on the position value, then passed through an MLP with an output dimension equal to the embedding dimension. We apply the vertical global positional encoding on the projected input and on the queries and keys in each attention module. More details can be found in the supplement." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.772, + 0.783, + 0.789 + ], + "angle": 0, + "content": "4.3. Spherical Local Self-Attention" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.909, + 0.903 + ], + "angle": 0, + "content": "The Spherical Local Self-Attention is the main component in our network that drives the SphereUFormer's \\(360^{\\circ}\\) perception. A detailed diagram is shown in Fig. 5. Given an input sphere, each node \\(x_{i}\\) in the sphere is projected into a query \\((q_{i})\\), key \\((k_{i})\\), and value \\((v_{i})\\). These vectors are all of the shape \\((N,L,H,D_H)\\), where \\(H\\) is the number of heads and \\(D_H\\) is the head dimension. Similar to standard prac" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.945, + 0.511, + 0.956 + ], + "angle": 0, + "content": "943" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.088, + 0.488, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.253, + 0.486, + 0.312 + ], + "angle": 0, + "content": "Figure 5. Spherical Local Self Attention. Attention is applied between each data node and its \\( K \\) neighbors. A learned relative position bias encodes information about the neighbors' relative position. In the right corner is a diagram of the enclosing block." + }, + { + "type": "image", + "bbox": [ + 0.114, + 0.327, + 0.48, + 0.475 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.484, + 0.484, + 0.598 + ], + "angle": 0, + "content": "Figure 6. Relative Positional Encondig. On the left, a center point (blue) and its first (green) and second (yellow) degree neighbors are defined by the graph structure. The relative position kernel is defined by a \\(7 \\times 7\\) grid, represented by the black nodes in the right image. The center point is positioned at the center of the grid and the neighbors are distributed across the grid according to their relative change in horizontal (\\(\\Delta \\theta\\)) and vertical (\\(\\Delta \\phi\\)) angles. Bilinear interpolation samples a learned positional encoding." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.627, + 0.484, + 0.793 + ], + "angle": 0, + "content": "tices, \\( D_H \\) is derived from the input dimension \\( D \\) and the number of heads. However, since the self-attention mechanism is the only spatial operation in the network, we added a \"head dimension coefficient\" (\\( C_{head} \\)) hyperparameter to upscale the head dimension \\( D_H \\) as a reverse bottleneck. Thus: \\( D_H = (D / H) \\cdot C_{head} \\). This coefficient allows an increase in the number of attention heads without reducing the dimension of each head and without increasing the overall size of the entire network, resulting in only a minor increase in parameters. We further discuss the motivation of this in the ablation section (Sec. 5.4)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.484, + 0.903 + ], + "angle": 0, + "content": "For each query, the neighboring \\( K \\) vectors are grouped according to the sphere geometry. Attention for each query is performed on the selected neighboring group. The number of neighbors \\( K \\) depends on a hyperparameter named \"window coefficient\" (\\( C_{win} \\)), which dictates the N-order neighbors to collect. A window coefficient of 0 provides no neighbors, a coefficient of 1 provides only first-order" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.182 + ], + "angle": 0, + "content": "neighbors, 2 also provides second-order neighbors, and so forth. Note that due to irregularities in the graph structure, the number of keys for each node is not identical as some have slightly fewer neighbors than others. This is homogenized for parallelization purposes by adding null neighbor keys and masking them." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.183, + 0.907, + 0.472 + ], + "angle": 0, + "content": "For relative position information, a learned relative position bias is added to the attention map. Since having a learned parameter per query-key pair is both memory expensive and harmful in terms of inductive bias, we took a shared-weights approach. For each node \\( i \\in [1, L] \\), we measure the angular difference \\( (\\Delta \\phi_{i,k}, \\Delta \\theta_{i,k}) \\) for each key \\( k \\in [1, K] \\) and normalize them by the max deltas. This is only computed once during initialization. During runtime, the normalized deltas (in \\( [-1, 1] \\)) are used to sample learned parameters from a \\( 7 \\times 7 \\) window. This sampling from a non-linear 2D function acts as a cheap relative positional encoding for the query-key pairs. Fig. 6 illustrates the relative position implementation. Given the blue center points, its first-order neighbors are depicted in green, and second-order neighbors in yellow, are placed in their relative position. A \\( 7 \\times 7 \\) grid, represented by the black points, represents the learned weights of the grid. Each neighbor gets its positional encoding by bilinearly interpolating between the four nearby weights." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.482, + 0.649, + 0.5 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.508, + 0.906, + 0.629 + ], + "angle": 0, + "content": "We evaluate our method by comparing it with state-of-the-art methods for depth estimation and semantic segmentation. For this purpose, we used the Stanford2D3D [4] and Structured3D [53] datasets. Both contain RGB, depth, and semantic segmentation data. We trained our model and all baselines using the training configuration of PanoFormer [36]. For more information on the training protocol please refer to the supplementary." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.629, + 0.906, + 0.705 + ], + "angle": 0, + "content": "We compared against ERP methods [20, 36, 48], spherical methods [8, 49], and Elite360D [1], that utilizes a combination of ERP with a low-resolution icosphere. We did not compare against patch-based solutions, as the difference in compute resources makes an unfair comparison." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.713, + 0.681, + 0.729 + ], + "angle": 0, + "content": "5.1. Design Decisions" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.906, + 0.78 + ], + "angle": 0, + "content": "Certain design decisions were made for the evaluation section. Some due to fairness considerations when comparing to the baselines, and others after testing various alternatives." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.781, + 0.907, + 0.902 + ], + "angle": 0, + "content": "For fairness reasons, it is important to compare methods of similar size. We configured all models to have 4 down-sampling stages, and a depth (number of blocks per stage) of 2 for each scale. Tab. 2 show the parameter and flop count per model. We gave PanoFormer [36], EGFormer [48], Elite360D [1] and HerRUNet[49] base embedding dim of 32, SFSS [20] 48, HealSWIN [8] 64. The resnet18 backbone was used for Elite360D. This resulted in as equal as" + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.945, + 0.513, + 0.958 + ], + "angle": 0, + "content": "944" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.094, + 0.089, + 0.482, + 0.294 + ], + "angle": 0, + "content": "
ModelRes. (#Pixels)ParamsFlops
PanoFormer [36]256×512(131K)14.5M11.8G
EGFormer [48]256×512(131K)15.2M15.6G
SFSS [20]256×512(131K)15.1M18.9G
Elite360D [1]256×512†(136K)14.7M13.6G
HealSWIN [8]12·47(196K)12.0M39.0G
HexRUnet [49]10·47+2(164K)14.0M12.4G
PanoFormer [36]512×1024(524K)14.5M44.7G
EGFormer [48]512×1024(524K)15.2M65.8G
SFSS [20]512×1024(524K)15.1M80.3G
Elite360D [1]512×1024†(529K)14.7M51.4G
HealSWIN [8]12·48(786K)12.0M156.1G
" + }, + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.304, + 0.482, + 0.333 + ], + "angle": 0, + "content": "Table 2. Resolution and parameter count of baseline models. \\( {}^{ \\dagger } \\) Elite360D uses additional icosphere of rank 4 (5120 data points)." + }, + { + "type": "table", + "bbox": [ + 0.098, + 0.347, + 0.476, + 0.484 + ], + "angle": 0, + "content": "
#RankTypeCheadCwinRes.ParamsFlops
6hex1141K11.2M2.5G
6ico1182K11.2M4.9G
7hex11164K11.2M9.9G
7ico11328K11.2M19.1G
7hex21164K14.9M13.0G
(1)7hex22164K14.9M13.1G
(2)8hex22655K14.9M52.7G
" + }, + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.496, + 0.483, + 0.525 + ], + "angle": 0, + "content": "Table 3. Resolution (total data points) and parameter count of various configurations of our model. Selected configurations in green." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.552, + 0.483, + 0.734 + ], + "angle": 0, + "content": "possible resolution, parameter count, and flops that the different architectures allow. We analyzed input resolutions of \\(256 \\times 512\\) and \\(512 \\times 1024\\) for the grid models. Resolution of the spherical models was chosen to be as close as possible to those pixel counts. This was with subdivision ranks 7 and 8 respectively. Also for fairness, none of the methods used pretrained weights. To evaluate all models fairly and on equal terms, the predictions in the various representations were projected uniformly to the sphere surface. Unlike evaluating directly on ERP, this provides an evaluation that evenly averages over the sphere surface and is not biased to extreme vertical angles." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.735, + 0.483, + 0.902 + ], + "angle": 0, + "content": "We followed these guidelines to select the sphere setting with the closest number of nodes, total parameters and flops to the baselines. These stats are reported in Tab. 2 for the existing methods, while Tab. 3 lists variations of our method concerning: sphere rank, node type, head, and window coefficients \\((C_{head}, C_{win})\\). As can be seen, spheres with a rank of 6 have a very low resolution, regardless of the node type. For spheres with rank 7, we have found that the \"vertex\" node type (with \\(164\\mathrm{K}\\) nodes) provides a resolution that is closest to the \\(256\\times 512\\) (total of \\(131\\mathrm{K}\\) pixels) resolution, and rank 8 is the closest to \\(512\\times 1024\\). The base configura" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.304 + ], + "angle": 0, + "content": "tion of our method has fewer parameters and flops than the baselines. This is largely because, unlike most baselines, our method does not contain spatial layers, such as convolution kernels, outside of the attention layer. As a result, the base configuration of our method is less powerful. This is also further addressed in the ablation section (Sec. 5.4). To minimize this difference, we have found that increasing the head coefficient results in a model with slightly more parameters to match the minimum of the baselines. Increasing the window coefficient does not add any parameters and an insignificant amount of flops. Finally, we have selected two configurations for comparison with the baselines. One with rank 7 to be compared to \\(256 \\times 512\\) models, and one in rank 8 for larger resolution. These are highlighted in Tab. 3." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.315, + 0.686, + 0.331 + ], + "angle": 0, + "content": "5.2. Depth Estimation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.338, + 0.908, + 0.671 + ], + "angle": 0, + "content": "In depth estimation, the task is a dense prediction of the pixels depths. We evaluated the depth prediction up to 10 meters for Stanford2D3D and 5 meters for Structured3D. Following PanoFormer [36], we used the BerhuLoss [25] as a differentiable criterion to train the models. We then measured the Mean Absolute Error (MAE), Mean Relative Error (MRE), Root Mean Squared Error (RMSE), and \\(\\delta_{1}\\) accuracy, which are standard evaluation metrics for depth estimation. Quantitative results on \\(256\\times 512\\) are reported in Tab. 4. Our method is shown to outperform the baselines on all but one metric, where it was second. In evaluation on Stanford2D3D in \\(512\\times 1024\\) (Tab. 5), the gap is even larger, and our method significantly improved while the baselines did not. Fig. 7 presents sample results. Evidently, our results are considerably sharper than the baselines, especially around the center, and handle distortions around the poles better. We argue that this is due to spherical representation having a better effective resolution at the center, and does not suffer from distortions. We have also found depth predictions of most baselines to have a boundary misalignment effect where the \\(360^{\\circ}\\) and \\(0^{\\circ}\\) meet, which our method has no such effect. This is further explored in the supplement." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.682, + 0.731, + 0.699 + ], + "angle": 0, + "content": "5.3. Semantic Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.909, + 0.901 + ], + "angle": 0, + "content": "In semantic segmentation, the task is to predict a semantic class per image pixel. Stanford2D3D has 13 categories, while Structured3D has 40. We used the standard Categorical Cross Entropy loss, where the background category was ignored. For evaluation, we measured the global pixel accuracy and the mIoU percentage, which are standard metrics for semantic segmentation. Results are depicted in Tab. 4. Tab. 5 shows additional high-resolution setup. In this task as well, our method outperformed the baselines on all metrics. Fig. 8 presents typical samples. Similarly, our semantic segmentation results present the same improvement in terms of distortion handling and improvement around the center of the image we have observed for depth estimation." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.945, + 0.511, + 0.957 + ], + "angle": 0, + "content": "945" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.089, + 0.904, + 0.276 + ], + "angle": 0, + "content": "
ModelDepth EstimationSemantic Segmentation
Stanford2D3DStructured3DStanford2D3DStructured3D
MAE↓MRE↓RMSE↓δ1↑MAE↓MRE↓RMSE↓δ1↑Acc.↑mIoU↑Acc.↑mIoU↑
PanoFormer [36].174.078.45192.5.154.051.33094.883.160.694.949.7
EGFormer [48].170.075.45693.1.150.049.32895.286.566.495.051.5
SFSS [20].179.081.45592.2.155.051.33095.086.968.295.251.9
HexRUnet [49].201.090.51190.1----81.756.1--
HealSWIN [8].189.084.48292.2----85.563.2--
Elite360D [1].169.069.44893.5.147.046.32195.987.471.495.352.0
OURS (Tab. 3 (1)).165.071.43294.0.142.045.31496.488.672.295.853.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.28, + 0.287, + 0.719, + 0.302 + ], + "angle": 0, + "content": "Table 4. Quantitative comparison on sphere rank 7 \\((256\\times 512)\\) resolution." + }, + { + "type": "image", + "bbox": [ + 0.116, + 0.314, + 0.226, + 0.4 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.116, + 0.409, + 0.226, + 0.493 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.155, + 0.497, + 0.189, + 0.508 + ], + "angle": 0, + "content": "RGB" + }, + { + "type": "image", + "bbox": [ + 0.247, + 0.315, + 0.358, + 0.4 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.247, + 0.409, + 0.358, + 0.493 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.262, + 0.497, + 0.344, + 0.509 + ], + "angle": 0, + "content": "PanoFormer" + }, + { + "type": "image", + "bbox": [ + 0.379, + 0.315, + 0.489, + 0.4 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.378, + 0.409, + 0.489, + 0.493 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.399, + 0.497, + 0.47, + 0.508 + ], + "angle": 0, + "content": "EGFormer" + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.315, + 0.62, + 0.4 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.409, + 0.62, + 0.493 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.53, + 0.497, + 0.599, + 0.508 + ], + "angle": 0, + "content": "Elite360D" + }, + { + "type": "image", + "bbox": [ + 0.642, + 0.315, + 0.752, + 0.4 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.642, + 0.409, + 0.752, + 0.493 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.679, + 0.497, + 0.712, + 0.508 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.772, + 0.315, + 0.882, + 0.4 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.772, + 0.409, + 0.882, + 0.493 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.816, + 0.497, + 0.839, + 0.508 + ], + "angle": 0, + "content": "GT" + }, + { + "type": "image_caption", + "bbox": [ + 0.277, + 0.523, + 0.719, + 0.538 + ], + "angle": 0, + "content": "Figure 7. Depth Estimation. Top: Stanford2D3D. Bottom: Structured3D." + }, + { + "type": "table", + "bbox": [ + 0.104, + 0.562, + 0.472, + 0.651 + ], + "angle": 0, + "content": "
ModelMAE↓MRE↓δ1↑Acc.↑IoU↑
PanoFormer.167.07293.782.455.6
Elite360D.181.07793.284.563.3
OURS (Tab. 3 (2)).147.06594.089.171.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.101, + 0.66, + 0.473, + 0.675 + ], + "angle": 0, + "content": "Table 5. Comparison on Stanford2D3D in rank 8 (512 × 1024)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.703, + 0.244, + 0.719 + ], + "angle": 0, + "content": "5.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.726, + 0.484, + 0.787 + ], + "angle": 0, + "content": "To explore the contribution of each one of the elements we have conducted several ablation studies in which multiple variants of our method are tested. These tests were conducted in the context of depth estimation on Stanford2D3D." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Positional Encoding Two main components of the method are the global absolute and relative bias position encodings. The vertical global position encoding injects vertical position while remaining rotation equivariant horizontally. The relative bias within the self-attention provides relative information. We compared various configurations" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.564, + 0.907, + 0.64 + ], + "angle": 0, + "content": "that omit one or both of the encodings. Results can be seen in Tab. 6. Evidently, the absence of positional encoding, either global or relative, drastically impacts the performance of the model. Also, relative positional bias proved more critical than the global one." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.659, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Scaling Coefficients Two key modifiable scaling coefficients, the \"window\" \\((C_{win})\\) and \"head\" \\((C_{head})\\), affect the capacity of the attention module. The former affects the size of the attention window to incorporate further nodes in the graph. The latter increases the capacity of the attention head by increasing the head dimension. We tested the effect of these coefficients by gradually increasing them separately and together. As shown in Tab. 7. Increasing the window up until \\(C_{win} = 2\\) improves the performance. Interestingly, further increase of the window did not show an improvement. This could suggest that such large windows interfere with convergence. As expected, increasing the attention head also improves results, even up to \\(C_{head} = 4\\). Unsurprisingly, this suggests that the model could benefit from an overall increase in parameters, which is beyond the scope of this work." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "946" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.115, + 0.09, + 0.226, + 0.178 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.116, + 0.185, + 0.226, + 0.269 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.154, + 0.273, + 0.188, + 0.285 + ], + "angle": 0, + "content": "RGB" + }, + { + "type": "image", + "bbox": [ + 0.247, + 0.091, + 0.359, + 0.178 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.247, + 0.185, + 0.358, + 0.269 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.261, + 0.273, + 0.345, + 0.286 + ], + "angle": 0, + "content": "PanoFormer" + }, + { + "type": "image", + "bbox": [ + 0.377, + 0.091, + 0.49, + 0.178 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.378, + 0.185, + 0.489, + 0.269 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.398, + 0.273, + 0.47, + 0.286 + ], + "angle": 0, + "content": "EGFormer" + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.091, + 0.621, + 0.178 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.185, + 0.62, + 0.269 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.53, + 0.273, + 0.6, + 0.286 + ], + "angle": 0, + "content": "Elite360D" + }, + { + "type": "image", + "bbox": [ + 0.641, + 0.091, + 0.752, + 0.178 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.641, + 0.185, + 0.752, + 0.269 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.678, + 0.273, + 0.712, + 0.286 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.773, + 0.091, + 0.883, + 0.178 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.772, + 0.185, + 0.883, + 0.269 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.815, + 0.273, + 0.839, + 0.285 + ], + "angle": 0, + "content": "GT" + }, + { + "type": "image_caption", + "bbox": [ + 0.259, + 0.3, + 0.738, + 0.315 + ], + "angle": 0, + "content": "Figure 8. Semantic Segmentation. Top: Stanford2D3D. Bottom: Structured3D." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.339, + 0.483, + 0.434 + ], + "angle": 0, + "content": "
VariantAbsPosRelPosMAE↓MRE↓
No Pos. Enc.XX.326.094
No Rel. Pos. Enc.✓X.251.091
No Abs. Pos. Enc.X✓.218.088
With Pos. Enc.✓✓.189.077
" + }, + { + "type": "table_caption", + "bbox": [ + 0.145, + 0.444, + 0.428, + 0.459 + ], + "angle": 0, + "content": "Table 6. Ablation Study: Positional Encoding." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.473, + 0.481, + 0.614 + ], + "angle": 0, + "content": "
VariantCheadCwinMAE↓MRE↓
Base variant11.189.077
No window10.412.122
Medium window12.175.073
Large window13.180.078
Medium head21.171.075
Large head41.162.067
Conf. (1) Tab. 322.165.071
" + }, + { + "type": "table_caption", + "bbox": [ + 0.113, + 0.624, + 0.46, + 0.638 + ], + "angle": 0, + "content": "Table 7. Ablation Study: Head and Window Coefficients." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.664, + 0.342, + 0.68 + ], + "angle": 0, + "content": "6. Discussion and Limitations" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.483, + 0.795 + ], + "angle": 0, + "content": "This paper introduces a pioneering approach to omnidirectional \\(360^{\\circ}\\) perception, leveraging a geodesic polyhedron and a unique architecture designed for spherical computation. However, there is ample room for optimization and enhancement of its features. The potential for improvement spans various aspects of the model, from its core architecture to its computational efficiency." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.484, + 0.902 + ], + "angle": 0, + "content": "One of the promising avenues for future work lies in exploring additional domains where this architecture could be applied. Beyond the realms of depth estimation and semantic segmentation in \\(360^{\\circ}\\) imagery, spherical data representation is also prominent in other areas such as medical imaging, where spherical structures are common, and geographical data analysis, which often deals with global-scale data" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.341, + 0.905, + 0.387 + ], + "angle": 0, + "content": "that naturally fit a spherical model. The adaptability of our approach to these varied domains could pave the way for groundbreaking applications and methodologies." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.388, + 0.906, + 0.524 + ], + "angle": 0, + "content": "While the current work has focused on supervised classification and regression tasks, extending this model to generative models presents an exciting frontier. Generative models could leverage the spherical data representation for creating highly realistic and comprehensive 3D environments among other applications. The unique challenges of generative tasks on spherical domains, such as ensuring continuity and avoiding distortions, call for accurate modeling of the type we present here." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.525, + 0.906, + 0.676 + ], + "angle": 0, + "content": "Another area for improvement is the computational efficiency of the model, particularly regarding its GPU usage. Despite the number of parameters and flops being on par with baseline models, our approach exhibits slower performance (\\(\\sim 30\\%\\) slower than PanoFormer). This is largely because the gathering operations on the sphere graph are not optimized with dedicated CUDA functions in PyTorch, leading to less efficient execution. Addressing this engineering challenge could significantly enhance the model's appeal by reducing both the GPU footprint and runtime." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.692, + 0.634, + 0.708 + ], + "angle": 0, + "content": "7. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.718, + 0.906, + 0.825 + ], + "angle": 0, + "content": "We have presented a novel method for omnidirectional perception that successfully utilizes spherically represented data. The method addresses the underlying spherical structure, which comes up in three different places: in the down and upsampling operators, in the positional encoding, and, above all, in the definition geometric representation and that of the neighborhoods for the local attention operator." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Our proposed method relies on the icosphere representation and utilizes its favorable properties to derive sphere-based operators. Using these, we surpass the state of the art in omnidirectional perception in both depth estimation and semantic segmentation." + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.945, + 0.511, + 0.956 + ], + "angle": 0, + "content": "947" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.259, + 0.108 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.116, + 0.486, + 0.162 + ], + "angle": 0, + "content": "This work was supported by the Tel Aviv University Center for AI and Data Science (TAD). The contribution of the first author is part of a PhD thesis at Tel Aviv University." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.178, + 0.188, + 0.194 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.203, + 0.484, + 0.271 + ], + "angle": 0, + "content": "[1] Hao Ai and Lin Wang. Elite360d: Towards efficient 360 depth estimation via semantic-and distance-aware biprojection fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9926-9935, 2024. 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.274, + 0.484, + 0.356 + ], + "angle": 0, + "content": "[2] Hao Ai, Zidong Cao, Yan-Pei Cao, Ying Shan, and Lin Wang. Hrdfuse: Monocular 360deg depth estimation by collaboratively learning holistic-with-regional depth distributions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13273-13282, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.359, + 0.484, + 0.455 + ], + "angle": 0, + "content": "[3] Georgios Albanis, Nikolaos Zioulis, Petros Drakoulis, Vasileios Gkitsas, Vladimiros Sterzentenko, Federico Alvarez, Dimitrios Zarpalas, and Petros Daras. Pano3d: A holistic benchmark and a solid baseline for \\(360^{\\circ}\\) depth estimation. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 3722-3732, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.458, + 0.482, + 0.499 + ], + "angle": 0, + "content": "[4] Iro Armeni, Sasha Sax, Amir R Zamir, and Silvio Savarese. Joint 2d-3d-semantic data for indoor scene understanding. arXiv preprint arXiv:1702.01105, 2017. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.502, + 0.482, + 0.543 + ], + "angle": 0, + "content": "[5] Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. Neural machine translation by jointly learning to align and translate. arXiv preprint arXiv:1409.0473, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.546, + 0.483, + 0.573 + ], + "angle": 0, + "content": "[6] Herbert Busemann. The geometry of geodesics. Courier Corporation, 2012. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.575, + 0.484, + 0.641 + ], + "angle": 0, + "content": "[7] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.646, + 0.484, + 0.728 + ], + "angle": 0, + "content": "[8] Oscar Carlsson, Jan E Gerken, Hampus Linander, Heiner Spieß, Fredrik Ohlsson, Christoffer Petersson, and Daniel Persson. Heal-swin: A vision transformer on the sphere. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6067-6077, 2024. 1, 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.731, + 0.484, + 0.8 + ], + "angle": 0, + "content": "[9] Angel Chang, Angela Dai, Thomas Funkhouser, Maciej Halber, Matthias Niessner, Manolis Savva, Shuran Song, Andy Zeng, and Yinda Zhang. Matterport3d: Learning from rgbd data in indoor environments. International Conference on 3D Vision (3DV), 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.803, + 0.482, + 0.857 + ], + "angle": 0, + "content": "[10] Bowen Cheng, Alex Schwing, and Alexander Kirillov. Per-pixel classification is not all you need for semantic segmentation. Advances in Neural Information Processing Systems, 34:17864-17875, 2021. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.859, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[11] Hsien-Tzu Cheng, Chun-Hung Chao, Jin-Dong Dong, Hao-Kai Wen, Tyng-Luh Liu, and Min Sun. Cube padding for weakly-supervised saliency prediction in 360 videos. In Pro" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.203, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "ceedings of the IEEE conference on computer vision and pattern recognition, pages 1420-1429, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.123, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[12] Benjamin Coors, Alexandru Paul Condurache, and Andreas Geiger. Spherenet: Learning spherical representations for detection and classification in omnidirectional images. In Proceedings of the European conference on computer vision (ECCV), pages 518-533, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.193, + 0.905, + 0.246 + ], + "angle": 0, + "content": "[13] Michael Defferrard, Martino Milani, Frédérique Gusset, and Nathanaël Perraudin. DeepSphere: a graph-based spherical CNN. In International Conference on Learning Representations, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.25, + 0.905, + 0.304 + ], + "angle": 0, + "content": "[14] Aleksandar M Dimitrijevic, Martin Lambers, and Dejan Rancic. Comparison of spherical cube map projections used in planet-sized terrain rendering. Facta Universitatis, Series: Mathematics and Informatics, 31(2):259-297, 2016. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.307, + 0.905, + 0.401 + ], + "angle": 0, + "content": "[15] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.405, + 0.905, + 0.471 + ], + "angle": 0, + "content": "[16] Marc Eder, Mykhailo Shvets, John Lim, and Jan-Michael Frahm. Tangent images for mitigating spherical distortion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12426-12434, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.476, + 0.904, + 0.517 + ], + "angle": 0, + "content": "[17] Lance Flavell. Uv mapping. In *Beginning Blender: Open Source 3D Modeling*, Animation, and Game Design, pages 97-122. Springer, 2010. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.519, + 0.904, + 0.545 + ], + "angle": 0, + "content": "[18] Michael S Floater. Generalized barycentric coordinates and applications. Acta Numerica, 24:161-214, 2015. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.548, + 0.907, + 0.588 + ], + "angle": 0, + "content": "[19] Krzysztof M Górski, E Hivon, and BD Wandelt. Analysis issues for large cmb data sets. Proceedings: Evolution of Large Scale Structure-Garching, 1998. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.591, + 0.905, + 0.657 + ], + "angle": 0, + "content": "[20] Suresh Guttikonda and Jason Rambach. Single frame semantic segmentation using multi-modal spherical images. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 3222-3231, 2024. 1, 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.661, + 0.905, + 0.715 + ], + "angle": 0, + "content": "[21] Shi-Min Hu, Zheng-Ning Liu, Meng-Hao Guo, Jun-Xiong Cai, Jiahui Huang, Tai-Jiang Mu, and Ralph R Martin. Subdivision-based mesh convolution networks. ACM Transactions on Graphics (TOG), 41(3):1-16, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.718, + 0.905, + 0.771 + ], + "angle": 0, + "content": "[22] Chiyu Jiang, Jingwei Huang, Karthik Kashinath, Philip Marcus, Matthias Niessner, et al. Spherical cnns on unstructured grids. International Conference on Learning Representations, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.774, + 0.905, + 0.828 + ], + "angle": 0, + "content": "[23] Hualie Jiang, Zhe Sheng, Siyu Zhu, Zilong Dong, and Rui Huang. Unifuse: Unidirectional fusion for 360 panorama depth estimation. IEEE Robotics and Automation Letters, 6 (2):1519-1526, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.831, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[24] Benjamin Keinert, Matthias Innmann, Michael Sanger, and Marc Stamminger. Spherical fibonacci mapping. ACM Transactions on Graphics (TOG), 34(6):1-7, 2015. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[25] Iro Laina, Christian Rupprecht, Vasileios Belagiannis, Federico Tombari, and Nassir Navab. Deeper depth prediction" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.486, + 0.946, + 0.512, + 0.957 + ], + "angle": 0, + "content": "948" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.484, + 0.134 + ], + "angle": 0, + "content": "with fully convolutional residual networks. In 2016 Fourth international conference on 3D vision (3DV), pages 239-248. IEEE, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.135, + 0.484, + 0.176 + ], + "angle": 0, + "content": "[26] Martin Lambers. Survey of cube mapping methods in interactive computer graphics. The Visual Computer, 36(5): 1043-1051, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.177, + 0.483, + 0.247 + ], + "angle": 0, + "content": "[27] Yeonkun Lee, Jaeseok Jeong, Jongseob Yun, Wonjune Cho, and Kuk-Jin Yoon. Spherephd: Applying cnns on a spherical polyhedron representation of 360deg images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9181-9189, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.247, + 0.483, + 0.316 + ], + "angle": 0, + "content": "[28] Yuyan Li, Yuliang Guo, Zhixin Yan, Xinyu Huang, Ye Duan, and Liu Ren. Omnifusion: 360 monocular depth estimation via geometry-aware fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2801-2810, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.316, + 0.483, + 0.386 + ], + "angle": 0, + "content": "[29] Thang Luong, Hieu Pham, and Christopher D. Manning. Effective approaches to attention-based neural machine translation. In Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing. Association for Computational Linguistics, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.386, + 0.483, + 0.413 + ], + "angle": 0, + "content": "[30] Ronald Miller. An equi-rectangular map projection. Geography, pages 196–201, 1949. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.413, + 0.483, + 0.469 + ], + "angle": 0, + "content": "[31] Hongbin Pei, Bingzhe Wei, Kevin Chen-Chuan Chang, Yu Lei, and Bo Yang. Geom-gcn: Geometric graph convolutional networks. International Conference on Learning Representations, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.469, + 0.483, + 0.552 + ], + "angle": 0, + "content": "[32] Giovanni Pintore, Marco Agus, Eva Almansa, Jens Schneider, and Enrico Gobbetti. Slicenet: deep dense depth estimation from a single indoor panorama using a slice-based representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11536-11545, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.553, + 0.483, + 0.595 + ], + "angle": 0, + "content": "[33] Alec Radford, Karthik Narasimhan, Tim Salimans, Ilya Sutskever, et al. Improving language understanding by generative pre-training. 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.595, + 0.484, + 0.651 + ], + "angle": 0, + "content": "[34] René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. In Proceedings of the IEEE/CVF international conference on computer vision, pages 12179-12188, 2021. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.651, + 0.483, + 0.718 + ], + "angle": 0, + "content": "[35] Manuel Rey-Area, Mingze Yuan, and Christian Richardt. 360monodepth: High-resolution 360deg monocular depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3762-3772, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.72, + 0.483, + 0.789 + ], + "angle": 0, + "content": "[36] Zhijie Shen, Chunyu Lin, Kang Liao, Lang Nie, Zishuo Zheng, and Yao Zhao. Panofformer: Panorama transformer for indoor \\(360^{\\circ}\\) depth estimation. In European Conference on Computer Vision, pages 195-211. Springer, 2022. 1, 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.79, + 0.483, + 0.831 + ], + "angle": 0, + "content": "[37] Yu-Chuan Su and Kristen Grauman. Learning spherical convolution for fast features from 360 imagery. Advances in Neural Information Processing Systems, 30, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.483, + 0.9 + ], + "angle": 0, + "content": "[38] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Hohonet: 360 indoor holistic understanding with latent horizontal features. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2573-2582, 2021. 1" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.149 + ], + "angle": 0, + "content": "[39] Keisuke Tateno, Nassir Navab, and Federico Tombari. Distortion-aware convolutional filters for dense prediction in panoramic images. In Proceedings of the European Conference on Computer Vision (ECCV), pages 707-722, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.149, + 0.905, + 0.206 + ], + "angle": 0, + "content": "[40] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.207, + 0.905, + 0.262 + ], + "angle": 0, + "content": "[41] Petar Velicković, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Liò, and Yoshua Bengio. Graph attention networks. In International Conference on Learning Representations, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.262, + 0.905, + 0.332 + ], + "angle": 0, + "content": "[42] Fu-En Wang, Yu-Hsuan Yeh, Min Sun, Wei-Chen Chiu, and Yi-Hsuan Tsai. Bifuse: Monocular 360 depth estimation via bi-projection fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 462–471, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.334, + 0.905, + 0.403 + ], + "angle": 0, + "content": "[43] Fu-En Wang, Yu-Hsuan Yeh, Yi-Hsuan Tsai, Wei-Chen Chiu, and Min Sun. Bifuse++: Self-supervised and efficient bi-projection fusion for 360 depth estimation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(5): 5448-5460, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.403, + 0.905, + 0.475 + ], + "angle": 0, + "content": "[44] Zhendong Wang, Xiaodong Cun, Jianmin Bao, Wengang Zhou, Jianzhuang Liu, and Houqiang Li. Uformer: A general u-shaped transformer for image restoration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 17683-17693, 2022. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.476, + 0.905, + 0.545 + ], + "angle": 0, + "content": "[45] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. Advances in Neural Information Processing Systems, 34:12077-12090, 2021. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.546, + 0.905, + 0.603 + ], + "angle": 0, + "content": "[46] Qingsong Yan, Qiang Wang, Kaiyong Zhao, Bo Li, Xiaowei Chu, and Fei Deng. Sphereddepth: Panorama depth estimation from spherical domain. In 2022 International Conference on 3D Vision (3DV), pages 1-10. IEEE, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.603, + 0.905, + 0.673 + ], + "angle": 0, + "content": "[47] Ilwi Yun, Hyuk-Jae Lee, and Chae Eun Rhee. Improving 360 monocular depth estimation via non-local dense prediction transformer and joint supervised and self-supervised learning. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3224-3233, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.673, + 0.905, + 0.744 + ], + "angle": 0, + "content": "[48] Ilwi Yun, Chanyong Shin, Hyunku Lee, Hyuk-Jae Lee, and Chae Eun Rhee. Egformer: Equirectangular geometry-biased transformer for 360 depth estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 6101-6112, 2023. 1, 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.744, + 0.905, + 0.814 + ], + "angle": 0, + "content": "[49] Chao Zhang, Stephan Liwicki, William Smith, and Roberto Cipolla. Orientation-aware semantic segmentation on icosahedron spheres. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3533-3541, 2019. 1, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.814, + 0.905, + 0.857 + ], + "angle": 0, + "content": "[50] Chao Zhang, Sen He, and Stephan Liwicki. A spherical approach to planar semantic segmentation. In BMVC, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.857, + 0.905, + 0.902 + ], + "angle": 0, + "content": "[51] Jiaming Zhang, Kailun Yang, Chaoxiang Ma, Simon Reiß, Kunyu Peng, and Rainer Stiefelhagen. Bending reality: Distortion-aware transformers for adapting to panoramic se" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.512, + 0.956 + ], + "angle": 0, + "content": "949" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.123, + 0.092, + 0.482, + 0.134 + ], + "angle": 0, + "content": "matic segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16917-16927, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.482, + 0.178 + ], + "angle": 0, + "content": "[52] Si Zhang, Hanghang Tong, Jiejun Xu, and Ross Maciejewski. Graph convolutional networks: a comprehensive review. Computational Social Networks, 6(1):1-23, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.179, + 0.482, + 0.246 + ], + "angle": 0, + "content": "[53] Jia Zheng, Junfei Zhang, Jing Li, Rui Tang, Shenghua Gao, and Zihan Zhou. Structured3d: A large photo-realistic dataset for structured 3d modeling. In Proceedings of The European Conference on Computer Vision (ECCV), 2020. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.249, + 0.482, + 0.317 + ], + "angle": 0, + "content": "[54] Nikolaos Zioulis, Antonis Karakottas, Dimitrios Zarpalas, and Petros Daras. Omnidepth: Dense depth estimation for indoors spherical panoramicas. In Proceedings of the European Conference on Computer Vision (ECCV), pages 448-465, 2018. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.317 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.487, + 0.946, + 0.511, + 0.956 + ], + "angle": 0, + "content": "950" + } + ] +] \ No newline at end of file diff --git a/2025/SphereUFormer_ A U-Shaped Transformer for Spherical 360 Perception/d3dc320a-7961-421e-ae82-56124645a3a0_origin.pdf b/2025/SphereUFormer_ A U-Shaped Transformer for Spherical 360 Perception/d3dc320a-7961-421e-ae82-56124645a3a0_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..03c32579df527968812084c3d1a4df16ad310c26 --- /dev/null +++ b/2025/SphereUFormer_ A U-Shaped Transformer for Spherical 360 Perception/d3dc320a-7961-421e-ae82-56124645a3a0_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ed8918e57c26f8dd0b9ad99132be06ae0bf5e174d0e8ec0a46916c8411ef2d71 +size 7178785 diff --git a/2025/SphereUFormer_ A U-Shaped Transformer for Spherical 360 Perception/full.md b/2025/SphereUFormer_ A U-Shaped Transformer for Spherical 360 Perception/full.md new file mode 100644 index 0000000000000000000000000000000000000000..3208f16f0ca83885833cc95617002ff1a93676c1 --- /dev/null +++ b/2025/SphereUFormer_ A U-Shaped Transformer for Spherical 360 Perception/full.md @@ -0,0 +1,354 @@ +# SphereUFormer: A U-Shaped Transformer for Spherical 360 Perception + +Yaniv Benny +Tel Aviv University + +Lior Wolf +Tel Aviv University + +# Abstract + +This paper proposes a novel method for omnidirectional $360^{\circ}$ perception. Most common previous methods relied on equirectangular projection. This representation is easily applicable to 2D operation layers but introduces distortions into the image. Other methods attempted to remove the distortions by maintaining a sphere representation but relied on complicated convolution kernels that failed to show competitive results. In this work, we introduce a transformer-based architecture that, by incorporating a novel "Spherical Local Self-Attention" and other spherically-oriented modules, successfully operates in the spherical domain and outperforms the state-of-the-art in $360^{\circ}$ perception benchmarks for depth estimation and semantic segmentation. Our code is available at https://github.com/yanivbenny/sphere_uformer. + +# 1. Introduction + +Monocular omnidirectional $360^{\circ}$ perception is an important setting, as it allows for a full field-of-view (FOV) receptive field for the underlying model. Consequently, an entire room layout can be inferred, for example. In contrast to regular images, a $360^{\circ}$ image has no boundaries. Instead, the image is horizontally cyclical, and converges to singular points at the vertical extreme points. This suggests that this domain requires a different treatment, and that best practices that work well with regular images might not have the same effect on $360^{\circ}$ images. To tackle this task, multiple topics need to be considered. + +How should the data be represented? What kind of distortions and irregularities are created by this representation? How will a model process this representation? How to deal with these distortions and irregularities? + +These questions have attracted researchers to construct multiple datasets [3, 4, 9, 53] where depth estimation and semantic segmentation are the main tasks, and to formulate many solutions [20, 23, 28, 32, 35, 36, 38, 42, 43, 47, 48, 51], which propose novel approaches to solving these tasks. A common pattern in previous works is that they rep + +resent the data by projecting it to a 2D plane in different ways, usually equirectangular projection [30], cube mapping [26], or patch cropping. Each approach has its advantages and disadvantages, but the common motivation is to represent the image in a grid formation. This enables grid architectures, such as Convolutional Neural Networks or Vision Transformers to easily be applied. This comes at a cost of distortions, irregularities, or limited FOV, that are caused by the selected projection. A less common alternative representation that has been proposed is to represent the image in its original sphere domain by discretizing the sphere surface in some manner [6, 14, 17, 24]. The motivation is to maintain an undistorted representation of the $360^{\circ}$ image. However, this representation formed architectural challenges that require novel solutions. Some previous works [8, 12, 27, 46, 49, 50] addressed these challenges in various manners, but they fail to compete with current state-of-the-art solutions that follow a grid representation. + +In this work, we tackle the drawbacks of spherical representations. Specifically, we introduce SphereUFormer, which is a novel U-Shaped Transformer architecture tailored for spherical $360^{\circ}$ perception. By directly operating on a spherical domain without distortion-inducing projections, SphereUFormer addresses the core challenges of omnidirectional perception. Our approach builds on the strengths of transformer models, incorporating modifications that allow it to efficiently process and understand spherical data. This includes the development of spherical-specific operations such as localized self-attention mechanism that respects the geometry of the sphere, innovative up/downsampling techniques that maintain the integrity of spherical data throughout the model, and positional encoding scheme that matches the domain properties. SphereUFormer outperforms the state of the art in the field of $360^{\circ}$ perception, providing a robust and efficient tool for a wide range of applications that require comprehensive understanding of spherical environments. In Sec. 3 we describe this representation in detail. We then further describe a new method in Sec. 4, which utilizes this representation and performs $360^{\circ}$ perception on spherical representations, and in Sec. 5, we analyze and compare its results. + +# 2. Related Work + +Omnidirectional perception. There have been many advances, novelties, and research verticals in omnidirectional perception over the years. Initial works attempted standard convolutional networks on an equirectangular projection [54] (ERP). Although promising, the results were affected by the distortions created by the projection. To better handle these distortions, later works suggested technical adjustments to the convolution kernel such as Spherical Convolutions [37] or Distortion-Aware Convolutions [12, 39]. The latter sample the convolution features from a tangent plane. Other solutions attempted to eliminate distortions by sampling the image into smaller FOV patches [16, 28, 35, 47] followed by a post-processing stage that fuses the independent results. This, however, is at a cost of a restricted receptive field per patch and possible cutoff of crucial information, which either affects results or requires high overlap between patches. Due to the advantages and limitations of each method, [2, 42, 43] suggested a combined approach that fuses the ERP path with the individual patches at some stage of the network. Alternative approaches avoided the use of projection, and instead, operated directly on the sphere by constructing a spherical mesh. Usually either an icosphere [13, 22, 27, 46] or healpix [8, 13], and apply Graph Convolution [52], Mesh Convolution [21], or some custom convolution-like operation. + +Vision Transformers. ViT's [15] have disrupted the field of computer vision, from general-purpose architectures [34, 44] to specific purposes [7, 10, 45]. The attention mechanism [5, 29, 40] allows for dynamic pooling of features based on relevance and structural position. ViT's benefit from an efficient large context window compared to their CNN counterparts, which allows them to more easily learn long-distance relations between features. While earlier omnidirectional perception works proposed solutions based on CNN's, recent contributions replaced the convolution kernels with local attention of various forms [20, 36, 48] with favorable results. This work follows this direction. It utilizes the foundation of UFormer [44] and adapts it to spherical representations with custom attention layers. + +Graph Attention. Graph Convolution Networks [52] (GCN) are neural networks that apply convolution-like operations on data represented as a graph. A graph convolution operation aggregates the results of local operations between a node and each of its neighbors. Over multiple layers, this allows for information to pass between distant nodes in the graph and the update of these nodes. Graph Attention Networks [41] were proposed to enhance Graph Convolution Networks, thereby improving the message-passing capabilities of the GCN. The attention mechanism + +![](images/f6b9bb34f3fa26282bcf785443990c35dec9582ee7f7d4e8cdc5f972ad1f70a0.jpg) +(a) + +![](images/4554437dc1bffff70ead73229baf292b2c3060de5bfa7e2353e068a0d261d2ac.jpg) +(b) +Figure 1. Sphere Representations. From left to right: uvsphere, cubesphere, icosphere, hexasphere. + +![](images/c29ad352dee20996e1614884e1ccab961625b85196b11fce84aa5140276d2df0.jpg) +(c) + +![](images/ce2dc61b351747cfde2dc7c64a8cbd4c193d000a3ab068314f2ed4529853b8a4.jpg) +(d) + +has the benefit of allowing the operations to focus on a few sets of relevant nodes and providing an efficient method to encode positional information in the graph. For Geometric Graphs [31] geometric location is encoded to inform of each node's absolute and relative position, thereby making nodes aware of their global position and their respective neighbors' relative position. + +# 3. Sphere Representation + +Arguably the first topic in question when attempting to solve a problem is how to formulate the data. For omnidirectional $360^{\circ}$ images, the data can be seen as an array of values $v_{\phi ,\theta}$ assigned to a ray pointing from a center point-of-view towards $(\theta ,\phi)$ for $\theta \in [0,2\pi ]$ being the horizontal angle and $\phi \in [0,\pi ]$ the vertical. The value $v$ can correspond to the RGB value, depth, object category, etc. By selecting to sample on finite such angle pairs, the sphere becomes discretized. The sampling method determines how the sampled data can be processed and how successful the solution will become. + +The most common and straightforward is a uvsphere [17], where points are sampled in a grid with equal spacing for $(\theta, \phi)$ . See Fig. 1a. The uvsphere has a very high degree of horizontal symmetry, which is a good property. Another advantage is that it allows an unwrapping of the sphere into a 2D matrix. This unwrapping is called equirectangular projection (ERP). The downside of this representation is that the density of samples on the sphere is higher around the poles, due to the decreasing cross-section. This creates an imbalanced effective resolution over the sphere, and the projected image in 2D is distorted. + +An alternative that mitigates imbalanced resolution and distortion is cube mapping (or cubosphere) [14] (Fig. 1b), which projects the sphere onto the faces of a cube. Sample resolution is controlled by deciding how many rectangles each face of the cube is subdivided into. This removes the sampling density around the poles and the sphere can be unwrapped into six 2D images. However, two of these images are not vertically oriented and require special treatment. In addition, the transition between faces is not fluent, and complex padding methods [11] can only increase the receptive field of each image to some extent. Since the images are separated, a post-processing stage needs to fuse their predictions to form a unified prediction on the sphere. One op- + +![](images/dd82f349e41cfb609088cc68eb7fb79b93e1e3520780d4cf91e4de59c751f6c9.jpg) +0 + +![](images/1e48d5f8596e033fdd8b70b3a1126c05fd15fa75f99f804c50786b22e6d43c4b.jpg) +1 + +![](images/83a76e3f3ded829311a02ebe795e95d25d38547ac549c2a431c1a1ebc07808fb.jpg) +2 +Figure 2. Icospheres of different ranks. An increase in rank is made by subdividing each triangle into 4 smaller triangles, and results in an increased resolution. + +![](images/6e43b0e4efa2e8f54e574c6f85bed29998c43d45cecf8b01791a1ea603d0c7ba.jpg) +3 + +![](images/be04b33dde31b4d8a742c4fe36839a736ed0ff06874bb0a1577df740b574458e.jpg) +4 + +
Rank:01234567
Faces:208032012805120~20K~82K~328K
Vertices:12421626422562~10K~41K~164K
+ +Table 1. Icosphere Properties. Number of faces $(20 \times 4^i)$ and number of vertices $(10 \times 4^i + 2)$ per subdivision rank $i$ . + +![](images/f7724cd19a4b4836e19e5a61040721a9dfa0e43c0503c8298b902909e9532aea.jpg) +(a1) +Figure 3. Data as Icospheres. Icosphere of rank 3 (a1-c1) and rank 6 (a2-c2). (a) RGB, (b) Depth Map, and (c) Semantic Layout. + +![](images/4e3d73cc128036581d3f3c47d3bbf23711ce7fdc9ca766e22ab5707c309d82c6.jpg) +(b1) + +![](images/959764e5059a9c618a84e49fcbd7c0764389ef0da880f2ab4ebfcefdbc1e3215.jpg) +(c1) + +![](images/1fff96c2dadca431aea67e0f12be9dc718bc1b5ed18fd284a6d66db575fa2561.jpg) +(a2) + +![](images/1d63be6755d21db5f8fcbc6d3052698685442f334b5cb6cdac4c0470503afd61.jpg) +(b2) + +![](images/4ce7b793d7f57c13ca3bb2a0c052bb2cdecf884a4cdb2f9f40fb19bfb7e2196c.jpg) +(c2) + +tion is to operate on the cubosphere mesh instead of the 2D projection. Unfortunately, the cubosphere has irregularities at the coordinates of the cube corners and does not have a high horizontal symmetry, which makes it less optimal for this purpose. + +Instead of projecting the data to a grid and then having to deal with its limitations, a more elegant approach can be formulated using the sphere representation directly. Aside from the previously mentioned options, a common way to discretize a sphere is with Geodesic (icosphere, Fig. 1c) and Goldberg (hexasphere, Fig. 1d) polyhedrons [6]. An icosphere is created by repeatedly subdividing an icosahedron into smaller triangles. A hexasphere is the dual polyhedron where vertices are swapped with face normals, and edge connections with face adjacency. These types of sphere representation have a very high degree of symmetry and have a very evenly spaced sampling on the sphere surface. Due to their duality, our work follows the icosphere mesh, and icosphere/hexasphere mode is toggled by whether data points $(\theta, \phi)$ describe mesh vertices (hexasphere) or face normals (icosphere). Therefore we will sometimes refer to icospheres/hexaspheres by referring to the "node type" (face/vertex respectively). Other notable options for sphere sampling are Fibonacci spheres [24] and HealPix [19]. + +Icospheres have the highest order of symmetry, the most uniform distribution of points (perfect uniform distribution is impossible), easy setup, and an inherent up/down + +sampling definition through subdivision. All these reasons make them ideal for discrete spherical representation. As Fig. 2 and Tab. 1 show, in different ranks, icospheres produce sphere discretization in various resolutions. Each increase in sphere rank divides each triangle face into four smaller ones, hereby gradually increasing the sphere resolution. An example of representing data with icospheres can be seen in Fig. 3. In low resolution, the image seems "pixelated" as the independent triangle faces are distinguishable. But, with a high enough resolution, the image becomes detailed and visibly smooth. Our method uses high-resolution icospheres to represent the $360^{\circ}$ image, and through up and downsampling on the sphere structure, uses low-resolution spheres as hidden layers inside the model. + +# 4. Method + +This section introduces SphereUFormer, our proposed U-shaped Transformer for Spherical $360^{\circ}$ Perception. As the name suggests, the model is influenced by UFormer [44], with significant modifications to support spherically structured data. A high-level diagram is depicted in Fig. 4. The model's components are described in this section. For technical implementation details, see the supplementary. + +The network begins with an input projection that applies linear projection from the RGB values to latent embedding vectors. The projection output is then passed into an Encoder-Decoder network with multiple levels of self-attention modules in different sphere resolutions. We call these modules "Spherical Attention Modules" (SAM). At the bottom of the network lies a bottleneck SAM module that operates on the lowest resolution. Skip connections pass from the encoder to the decoder and feed high-resolution features. Finally, a linear Output Projection projects the decoder output to the output dimensions. + +As described in Sec. 3, our solution features the icosphere as its spherical representation structure, as we have found it to be the most intuitive and to contain the most satisfying properties. But it is not limited to it, and with minor modifications supports any spherical structure. + +Our method contains many local operations on the graph, from the down and up sampling, to the self-attention mechanism. Computing the cosine similarity or finding the neighbors of a node to generate the mapping scheme is slightly costly. This would be a problem if it was performed on each iteration, as the model would be slowed down. Luckily, since the icosphere graph is fixed, these mappings only need to be computed once in advance. Therefore there is no need to recompute it in every message massing operation. + +# 4.1. Up/Down Sampling + +In the spherical setting, upsampling and downsampling operations are performed by transforming a spherical graph into a new one with more or fewer nodes, respectively. This + +![](images/9a998734b9ccfb2e19d2e04bb23a716bb5d948d46206bd12a0d0252d671cb333.jpg) +Figure 4. The SphereUFormer architecture. A spherical representation is fed into the model. A linear input projection layer encodes the RGB values to latent embedding vectors. A sequence of SAM modules apply local self-attention on the spherical data along with downsampling layers that gradually reduce the resolution of the sphere. A sequence of SAM modules along with an upsampling layer and bypass skip connections decode the data. An output projection converts the latent embeddings to the output channel size. + +can be difficult and inefficient to compute for two arbitrary spherical polyhedrons. However, icospheres have multiple elegant and desired properties. First, an increase in resolution by one rank is the result of splitting each triangle face into four smaller ones, with one central triangle that shares its center with the large one. Second, for the same reason, each vertex in the low-resolution icosphere also exists in the high-resolution one, with new vertices created at the center of each edge connection. Third, the number of nodes/faces increases at a constant rate. + +All these properties make constructing up and down operations extremely straightforward and efficient. Every common pooling technique, such as max pooling, average pooling, center pooling, and interpolation, can be implemented. However, there are preferences for each setting. For downsampling, when the data nodes are represented by the faces, we have found max and center pooling to have the same overall results. When the data nodes are on the vertices (equivalent to hexosphere), max and average pooling become a little complicated, since there are both instances of five and six neighbors in the graph. Therefore, we resorted to only center pooling. For upsampling, the faces mode most simply operates with nearest upsampling, where each triangle receives the value according to the lower resolution triangle it was subdivided from. Other interpolation methods require more complicated computations and were therefore not considered. For vertices, however, interpolation is straightforward, since each new node is located precisely at the center of an existing edge. For more upscaled upsampling, Barycentric coordinates [18] interpolation can also be performed quite easily, but we did not find a reason to upscale beyond a factor of 2 at a time. + +# 4.2. Global Positional Encoding + +Positional encoding is a common and critical technique in Transformers [33, 40] and Vision Transformers [7, 10, 15, 34, 44, 45]. SphereUFormer utilizes positional encoding as well, but in a manner that suits spherical representations. The node position on the unit sphere is expressed using spherical coordinates $(\theta, \phi)$ . In our setting of omnidirectional perception, the data is always vertically aligned, therefore the model does not need to be vertically equivariant. However, the horizontal rotation is arbitrary, therefore it is desired to make the model horizontally equivariant to rotation and flipping. We therefore apply absolute positional encoding for the vertical position but not for the horizontal position. Horizontal position is only used in relative to another node as position bias in the Local Self-Attention described in the next section. + +For the vertical position, we consider $\phi \in [0,\pi ]$ . The position was encoded by applying sinusoidal encoding on the position value, then passed through an MLP with an output dimension equal to the embedding dimension. We apply the vertical global positional encoding on the projected input and on the queries and keys in each attention module. More details can be found in the supplement. + +# 4.3. Spherical Local Self-Attention + +The Spherical Local Self-Attention is the main component in our network that drives the SphereUFormer's $360^{\circ}$ perception. A detailed diagram is shown in Fig. 5. Given an input sphere, each node $x_{i}$ in the sphere is projected into a query $(q_{i})$ , key $(k_{i})$ , and value $(v_{i})$ . These vectors are all of the shape $(N,L,H,D_H)$ , where $H$ is the number of heads and $D_H$ is the head dimension. Similar to standard prac + +![](images/8e3578f1b6abdee667ebccf5518ea29ee3fdacd56bbc9bdacb713f51a58614fa.jpg) +Figure 5. Spherical Local Self Attention. Attention is applied between each data node and its $K$ neighbors. A learned relative position bias encodes information about the neighbors' relative position. In the right corner is a diagram of the enclosing block. + +![](images/3df419116959d1c79b9db42eaf035bb3d4833408380c9242438277b18bbf1a65.jpg) +Figure 6. Relative Positional Encondig. On the left, a center point (blue) and its first (green) and second (yellow) degree neighbors are defined by the graph structure. The relative position kernel is defined by a $7 \times 7$ grid, represented by the black nodes in the right image. The center point is positioned at the center of the grid and the neighbors are distributed across the grid according to their relative change in horizontal ( $\Delta \theta$ ) and vertical ( $\Delta \phi$ ) angles. Bilinear interpolation samples a learned positional encoding. + +tices, $D_H$ is derived from the input dimension $D$ and the number of heads. However, since the self-attention mechanism is the only spatial operation in the network, we added a "head dimension coefficient" ( $C_{head}$ ) hyperparameter to upscale the head dimension $D_H$ as a reverse bottleneck. Thus: $D_H = (D / H) \cdot C_{head}$ . This coefficient allows an increase in the number of attention heads without reducing the dimension of each head and without increasing the overall size of the entire network, resulting in only a minor increase in parameters. We further discuss the motivation of this in the ablation section (Sec. 5.4). + +For each query, the neighboring $K$ vectors are grouped according to the sphere geometry. Attention for each query is performed on the selected neighboring group. The number of neighbors $K$ depends on a hyperparameter named "window coefficient" ( $C_{win}$ ), which dictates the N-order neighbors to collect. A window coefficient of 0 provides no neighbors, a coefficient of 1 provides only first-order + +neighbors, 2 also provides second-order neighbors, and so forth. Note that due to irregularities in the graph structure, the number of keys for each node is not identical as some have slightly fewer neighbors than others. This is homogenized for parallelization purposes by adding null neighbor keys and masking them. + +For relative position information, a learned relative position bias is added to the attention map. Since having a learned parameter per query-key pair is both memory expensive and harmful in terms of inductive bias, we took a shared-weights approach. For each node $i \in [1, L]$ , we measure the angular difference $(\Delta \phi_{i,k}, \Delta \theta_{i,k})$ for each key $k \in [1, K]$ and normalize them by the max deltas. This is only computed once during initialization. During runtime, the normalized deltas (in $[-1, 1]$ ) are used to sample learned parameters from a $7 \times 7$ window. This sampling from a non-linear 2D function acts as a cheap relative positional encoding for the query-key pairs. Fig. 6 illustrates the relative position implementation. Given the blue center points, its first-order neighbors are depicted in green, and second-order neighbors in yellow, are placed in their relative position. A $7 \times 7$ grid, represented by the black points, represents the learned weights of the grid. Each neighbor gets its positional encoding by bilinearly interpolating between the four nearby weights. + +# 5. Experiments + +We evaluate our method by comparing it with state-of-the-art methods for depth estimation and semantic segmentation. For this purpose, we used the Stanford2D3D [4] and Structured3D [53] datasets. Both contain RGB, depth, and semantic segmentation data. We trained our model and all baselines using the training configuration of PanoFormer [36]. For more information on the training protocol please refer to the supplementary. + +We compared against ERP methods [20, 36, 48], spherical methods [8, 49], and Elite360D [1], that utilizes a combination of ERP with a low-resolution icosphere. We did not compare against patch-based solutions, as the difference in compute resources makes an unfair comparison. + +# 5.1. Design Decisions + +Certain design decisions were made for the evaluation section. Some due to fairness considerations when comparing to the baselines, and others after testing various alternatives. + +For fairness reasons, it is important to compare methods of similar size. We configured all models to have 4 down-sampling stages, and a depth (number of blocks per stage) of 2 for each scale. Tab. 2 show the parameter and flop count per model. We gave PanoFormer [36], EGFormer [48], Elite360D [1] and HerRUNet[49] base embedding dim of 32, SFSS [20] 48, HealSWIN [8] 64. The resnet18 backbone was used for Elite360D. This resulted in as equal as + +
ModelRes. (#Pixels)ParamsFlops
PanoFormer [36]256×512(131K)14.5M11.8G
EGFormer [48]256×512(131K)15.2M15.6G
SFSS [20]256×512(131K)15.1M18.9G
Elite360D [1]256×512†(136K)14.7M13.6G
HealSWIN [8]12·47(196K)12.0M39.0G
HexRUnet [49]10·47+2(164K)14.0M12.4G
PanoFormer [36]512×1024(524K)14.5M44.7G
EGFormer [48]512×1024(524K)15.2M65.8G
SFSS [20]512×1024(524K)15.1M80.3G
Elite360D [1]512×1024†(529K)14.7M51.4G
HealSWIN [8]12·48(786K)12.0M156.1G
+ +Table 2. Resolution and parameter count of baseline models. ${}^{ \dagger }$ Elite360D uses additional icosphere of rank 4 (5120 data points). + +
#RankTypeCheadCwinRes.ParamsFlops
6hex1141K11.2M2.5G
6ico1182K11.2M4.9G
7hex11164K11.2M9.9G
7ico11328K11.2M19.1G
7hex21164K14.9M13.0G
(1)7hex22164K14.9M13.1G
(2)8hex22655K14.9M52.7G
+ +Table 3. Resolution (total data points) and parameter count of various configurations of our model. Selected configurations in green. + +possible resolution, parameter count, and flops that the different architectures allow. We analyzed input resolutions of $256 \times 512$ and $512 \times 1024$ for the grid models. Resolution of the spherical models was chosen to be as close as possible to those pixel counts. This was with subdivision ranks 7 and 8 respectively. Also for fairness, none of the methods used pretrained weights. To evaluate all models fairly and on equal terms, the predictions in the various representations were projected uniformly to the sphere surface. Unlike evaluating directly on ERP, this provides an evaluation that evenly averages over the sphere surface and is not biased to extreme vertical angles. + +We followed these guidelines to select the sphere setting with the closest number of nodes, total parameters and flops to the baselines. These stats are reported in Tab. 2 for the existing methods, while Tab. 3 lists variations of our method concerning: sphere rank, node type, head, and window coefficients $(C_{head}, C_{win})$ . As can be seen, spheres with a rank of 6 have a very low resolution, regardless of the node type. For spheres with rank 7, we have found that the "vertex" node type (with $164\mathrm{K}$ nodes) provides a resolution that is closest to the $256\times 512$ (total of $131\mathrm{K}$ pixels) resolution, and rank 8 is the closest to $512\times 1024$ . The base configura + +tion of our method has fewer parameters and flops than the baselines. This is largely because, unlike most baselines, our method does not contain spatial layers, such as convolution kernels, outside of the attention layer. As a result, the base configuration of our method is less powerful. This is also further addressed in the ablation section (Sec. 5.4). To minimize this difference, we have found that increasing the head coefficient results in a model with slightly more parameters to match the minimum of the baselines. Increasing the window coefficient does not add any parameters and an insignificant amount of flops. Finally, we have selected two configurations for comparison with the baselines. One with rank 7 to be compared to $256 \times 512$ models, and one in rank 8 for larger resolution. These are highlighted in Tab. 3. + +# 5.2. Depth Estimation + +In depth estimation, the task is a dense prediction of the pixels depths. We evaluated the depth prediction up to 10 meters for Stanford2D3D and 5 meters for Structured3D. Following PanoFormer [36], we used the BerhuLoss [25] as a differentiable criterion to train the models. We then measured the Mean Absolute Error (MAE), Mean Relative Error (MRE), Root Mean Squared Error (RMSE), and $\delta_{1}$ accuracy, which are standard evaluation metrics for depth estimation. Quantitative results on $256\times 512$ are reported in Tab. 4. Our method is shown to outperform the baselines on all but one metric, where it was second. In evaluation on Stanford2D3D in $512\times 1024$ (Tab. 5), the gap is even larger, and our method significantly improved while the baselines did not. Fig. 7 presents sample results. Evidently, our results are considerably sharper than the baselines, especially around the center, and handle distortions around the poles better. We argue that this is due to spherical representation having a better effective resolution at the center, and does not suffer from distortions. We have also found depth predictions of most baselines to have a boundary misalignment effect where the $360^{\circ}$ and $0^{\circ}$ meet, which our method has no such effect. This is further explored in the supplement. + +# 5.3. Semantic Segmentation + +In semantic segmentation, the task is to predict a semantic class per image pixel. Stanford2D3D has 13 categories, while Structured3D has 40. We used the standard Categorical Cross Entropy loss, where the background category was ignored. For evaluation, we measured the global pixel accuracy and the mIoU percentage, which are standard metrics for semantic segmentation. Results are depicted in Tab. 4. Tab. 5 shows additional high-resolution setup. In this task as well, our method outperformed the baselines on all metrics. Fig. 8 presents typical samples. Similarly, our semantic segmentation results present the same improvement in terms of distortion handling and improvement around the center of the image we have observed for depth estimation. + +
ModelDepth EstimationSemantic Segmentation
Stanford2D3DStructured3DStanford2D3DStructured3D
MAE↓MRE↓RMSE↓δ1↑MAE↓MRE↓RMSE↓δ1↑Acc.↑mIoU↑Acc.↑mIoU↑
PanoFormer [36].174.078.45192.5.154.051.33094.883.160.694.949.7
EGFormer [48].170.075.45693.1.150.049.32895.286.566.495.051.5
SFSS [20].179.081.45592.2.155.051.33095.086.968.295.251.9
HexRUnet [49].201.090.51190.1----81.756.1--
HealSWIN [8].189.084.48292.2----85.563.2--
Elite360D [1].169.069.44893.5.147.046.32195.987.471.495.352.0
OURS (Tab. 3 (1)).165.071.43294.0.142.045.31496.488.672.295.853.0
+ +![](images/48a9c203806e696fc9c424112604ae65f6c90c832957c5f96adfd56ba38c2663.jpg) + +![](images/f61dcfec21972ec5f190feb2b822142c808ace57fa2babdd1d961c3767588707.jpg) +RGB + +![](images/61d52f096ca5f453ef288d7fd3c02e9da02d25f290fe055639f3e829c045e66d.jpg) + +![](images/81f65050a0b7937899498948208c583664f1317e9d70708b687fbc2a35b0f745.jpg) +PanoFormer + +![](images/146a8a60d2f4f1ebcb3efa6f903b6c3077998c92f54e62c34cee9be6f32d0205.jpg) + +![](images/69caf7a2d787868cc97039069a98990ddb52e3a7b9c06ecfd71af3c664b8f32e.jpg) +EGFormer + +![](images/e98bf00c09c4957bf43d0a7396a607a9d7bd2c326845d8f7e6341db9001bb2f0.jpg) + +![](images/6d2068fa0fa4383216e986cf4e1b3c60fe0b5bc29963157b73bd353092e89c11.jpg) +Elite360D + +![](images/3a226aebdd51df120d90aedb810585abb9cb99e792bbd4f315ab59c2eff06e2d.jpg) + +![](images/847558752fbe02aa5ca306c897621a01790fb993b22ad284ff9af5fbf589ee70.jpg) +Ours +Figure 7. Depth Estimation. Top: Stanford2D3D. Bottom: Structured3D. + +![](images/d2b31beac34d5041bc4bb3f6eacea12e739504f4adc94db6f1279b80ad738d1b.jpg) + +![](images/460b1984ab08eb1df6c2e89296086de7c91eb23bfd9651bd13d245dd30ebbfec.jpg) +GT + +Table 4. Quantitative comparison on sphere rank 7 $(256\times 512)$ resolution. + +
ModelMAE↓MRE↓δ1↑Acc.↑IoU↑
PanoFormer.167.07293.782.455.6
Elite360D.181.07793.284.563.3
OURS (Tab. 3 (2)).147.06594.089.171.5
+ +Table 5. Comparison on Stanford2D3D in rank 8 (512 × 1024). + +# 5.4. Ablation Study + +To explore the contribution of each one of the elements we have conducted several ablation studies in which multiple variants of our method are tested. These tests were conducted in the context of depth estimation on Stanford2D3D. + +Positional Encoding Two main components of the method are the global absolute and relative bias position encodings. The vertical global position encoding injects vertical position while remaining rotation equivariant horizontally. The relative bias within the self-attention provides relative information. We compared various configurations + +that omit one or both of the encodings. Results can be seen in Tab. 6. Evidently, the absence of positional encoding, either global or relative, drastically impacts the performance of the model. Also, relative positional bias proved more critical than the global one. + +Scaling Coefficients Two key modifiable scaling coefficients, the "window" $(C_{win})$ and "head" $(C_{head})$ , affect the capacity of the attention module. The former affects the size of the attention window to incorporate further nodes in the graph. The latter increases the capacity of the attention head by increasing the head dimension. We tested the effect of these coefficients by gradually increasing them separately and together. As shown in Tab. 7. Increasing the window up until $C_{win} = 2$ improves the performance. Interestingly, further increase of the window did not show an improvement. This could suggest that such large windows interfere with convergence. As expected, increasing the attention head also improves results, even up to $C_{head} = 4$ . Unsurprisingly, this suggests that the model could benefit from an overall increase in parameters, which is beyond the scope of this work. + +![](images/2829777ed08c2ede78485f2504f1e76af3780e3965f982cc61abaaab25f7d3fd.jpg) + +![](images/fb935d7211132be200d086b06a041d3db4c6b9b4c9ee1a05288287a065575d9d.jpg) +RGB + +![](images/63b7beb973642fd6ec89d2dafe25684b125dc5cf9dc1f78cdf2e6156edc59bca.jpg) + +![](images/8717697ffd72d63f5b9f354509a58f7c589d06fb860685d8a87af3fac381f338.jpg) +PanoFormer + +![](images/4854a2a5d9f9979fbcbe044df53129d2313bb102c99e509afa1ba0b5da6fa22b.jpg) + +![](images/e2775a91bebf7c1381a5d8971f199d0e10688445fe225cb4e71c862d58bd80dc.jpg) +EGFormer + +![](images/ae84e928b8b10506b2f528ffde90bc6d55c515d033fa25673b607b823cd9f552.jpg) + +![](images/666c629d97f2b11d794caaf9524bd78e065150812f4275d6581a60339fb03807.jpg) +Elite360D +Figure 8. Semantic Segmentation. Top: Stanford2D3D. Bottom: Structured3D. + +![](images/8a33613f0f8ae52c2b5aeed34a043d5a4e627d4d45821277d80eda194492987e.jpg) + +![](images/cd16e1dfb09ad36aa9b6d8d8ce4afc9cb9aadf0530096c75209a8e27e9104d02.jpg) +Ours + +![](images/ccfda3f569d92a0c3df6dc14594accc46e4d6ba66ac60ef025ea2621ca0b8868.jpg) + +![](images/5fc1858bd06e5119cb49e5558d2629c46e8789f140b0b339ad742a5ebb119abc.jpg) +GT + +
VariantAbsPosRelPosMAE↓MRE↓
No Pos. Enc.XX.326.094
No Rel. Pos. Enc.✓X.251.091
No Abs. Pos. Enc.X✓.218.088
With Pos. Enc.✓✓.189.077
+ +Table 6. Ablation Study: Positional Encoding. + +
VariantCheadCwinMAE↓MRE↓
Base variant11.189.077
No window10.412.122
Medium window12.175.073
Large window13.180.078
Medium head21.171.075
Large head41.162.067
Conf. (1) Tab. 322.165.071
+ +Table 7. Ablation Study: Head and Window Coefficients. + +# 6. Discussion and Limitations + +This paper introduces a pioneering approach to omnidirectional $360^{\circ}$ perception, leveraging a geodesic polyhedron and a unique architecture designed for spherical computation. However, there is ample room for optimization and enhancement of its features. The potential for improvement spans various aspects of the model, from its core architecture to its computational efficiency. + +One of the promising avenues for future work lies in exploring additional domains where this architecture could be applied. Beyond the realms of depth estimation and semantic segmentation in $360^{\circ}$ imagery, spherical data representation is also prominent in other areas such as medical imaging, where spherical structures are common, and geographical data analysis, which often deals with global-scale data + +that naturally fit a spherical model. The adaptability of our approach to these varied domains could pave the way for groundbreaking applications and methodologies. + +While the current work has focused on supervised classification and regression tasks, extending this model to generative models presents an exciting frontier. Generative models could leverage the spherical data representation for creating highly realistic and comprehensive 3D environments among other applications. The unique challenges of generative tasks on spherical domains, such as ensuring continuity and avoiding distortions, call for accurate modeling of the type we present here. + +Another area for improvement is the computational efficiency of the model, particularly regarding its GPU usage. Despite the number of parameters and flops being on par with baseline models, our approach exhibits slower performance ( $\sim 30\%$ slower than PanoFormer). This is largely because the gathering operations on the sphere graph are not optimized with dedicated CUDA functions in PyTorch, leading to less efficient execution. Addressing this engineering challenge could significantly enhance the model's appeal by reducing both the GPU footprint and runtime. + +# 7. Conclusion + +We have presented a novel method for omnidirectional perception that successfully utilizes spherically represented data. The method addresses the underlying spherical structure, which comes up in three different places: in the down and upsampling operators, in the positional encoding, and, above all, in the definition geometric representation and that of the neighborhoods for the local attention operator. + +Our proposed method relies on the icosphere representation and utilizes its favorable properties to derive sphere-based operators. Using these, we surpass the state of the art in omnidirectional perception in both depth estimation and semantic segmentation. + +# Acknowledgements + +This work was supported by the Tel Aviv University Center for AI and Data Science (TAD). The contribution of the first author is part of a PhD thesis at Tel Aviv University. + +# References + +[1] Hao Ai and Lin Wang. Elite360d: Towards efficient 360 depth estimation via semantic-and distance-aware biprojection fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9926-9935, 2024. 5, 6, 7 +[2] Hao Ai, Zidong Cao, Yan-Pei Cao, Ying Shan, and Lin Wang. Hrdfuse: Monocular 360deg depth estimation by collaboratively learning holistic-with-regional depth distributions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13273-13282, 2023. 2 +[3] Georgios Albanis, Nikolaos Zioulis, Petros Drakoulis, Vasileios Gkitsas, Vladimiros Sterzentenko, Federico Alvarez, Dimitrios Zarpalas, and Petros Daras. Pano3d: A holistic benchmark and a solid baseline for $360^{\circ}$ depth estimation. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 3722-3732, 2021. 1 +[4] Iro Armeni, Sasha Sax, Amir R Zamir, and Silvio Savarese. Joint 2d-3d-semantic data for indoor scene understanding. arXiv preprint arXiv:1702.01105, 2017. 1, 5 +[5] Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. Neural machine translation by jointly learning to align and translate. arXiv preprint arXiv:1409.0473, 2014. 2 +[6] Herbert Busemann. The geometry of geodesics. Courier Corporation, 2012. 1, 3 +[7] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 2, 4 +[8] Oscar Carlsson, Jan E Gerken, Hampus Linander, Heiner Spieß, Fredrik Ohlsson, Christoffer Petersson, and Daniel Persson. Heal-swin: A vision transformer on the sphere. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6067-6077, 2024. 1, 2, 5, 6, 7 +[9] Angel Chang, Angela Dai, Thomas Funkhouser, Maciej Halber, Matthias Niessner, Manolis Savva, Shuran Song, Andy Zeng, and Yinda Zhang. Matterport3d: Learning from rgbd data in indoor environments. International Conference on 3D Vision (3DV), 2017. 1 +[10] Bowen Cheng, Alex Schwing, and Alexander Kirillov. Per-pixel classification is not all you need for semantic segmentation. Advances in Neural Information Processing Systems, 34:17864-17875, 2021. 2, 4 +[11] Hsien-Tzu Cheng, Chun-Hung Chao, Jin-Dong Dong, Hao-Kai Wen, Tyng-Luh Liu, and Min Sun. Cube padding for weakly-supervised saliency prediction in 360 videos. In Pro + +ceedings of the IEEE conference on computer vision and pattern recognition, pages 1420-1429, 2018. 2 +[12] Benjamin Coors, Alexandru Paul Condurache, and Andreas Geiger. Spherenet: Learning spherical representations for detection and classification in omnidirectional images. In Proceedings of the European conference on computer vision (ECCV), pages 518-533, 2018. 1, 2 +[13] Michael Defferrard, Martino Milani, Frédérique Gusset, and Nathanaël Perraudin. DeepSphere: a graph-based spherical CNN. In International Conference on Learning Representations, 2020. 2 +[14] Aleksandar M Dimitrijevic, Martin Lambers, and Dejan Rancic. Comparison of spherical cube map projections used in planet-sized terrain rendering. Facta Universitatis, Series: Mathematics and Informatics, 31(2):259-297, 2016. 1, 2 +[15] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021. 2, 4 +[16] Marc Eder, Mykhailo Shvets, John Lim, and Jan-Michael Frahm. Tangent images for mitigating spherical distortion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12426-12434, 2020. 2 +[17] Lance Flavell. Uv mapping. In *Beginning Blender: Open Source 3D Modeling*, Animation, and Game Design, pages 97-122. Springer, 2010. 1, 2 +[18] Michael S Floater. Generalized barycentric coordinates and applications. Acta Numerica, 24:161-214, 2015. 4 +[19] Krzysztof M Górski, E Hivon, and BD Wandelt. Analysis issues for large cmb data sets. Proceedings: Evolution of Large Scale Structure-Garching, 1998. 3 +[20] Suresh Guttikonda and Jason Rambach. Single frame semantic segmentation using multi-modal spherical images. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 3222-3231, 2024. 1, 2, 5, 6, 7 +[21] Shi-Min Hu, Zheng-Ning Liu, Meng-Hao Guo, Jun-Xiong Cai, Jiahui Huang, Tai-Jiang Mu, and Ralph R Martin. Subdivision-based mesh convolution networks. ACM Transactions on Graphics (TOG), 41(3):1-16, 2022. 2 +[22] Chiyu Jiang, Jingwei Huang, Karthik Kashinath, Philip Marcus, Matthias Niessner, et al. Spherical cnns on unstructured grids. International Conference on Learning Representations, 2019. 2 +[23] Hualie Jiang, Zhe Sheng, Siyu Zhu, Zilong Dong, and Rui Huang. Unifuse: Unidirectional fusion for 360 panorama depth estimation. IEEE Robotics and Automation Letters, 6 (2):1519-1526, 2021. 1 +[24] Benjamin Keinert, Matthias Innmann, Michael Sanger, and Marc Stamminger. Spherical fibonacci mapping. ACM Transactions on Graphics (TOG), 34(6):1-7, 2015. 1, 3 +[25] Iro Laina, Christian Rupprecht, Vasileios Belagiannis, Federico Tombari, and Nassir Navab. Deeper depth prediction + +with fully convolutional residual networks. In 2016 Fourth international conference on 3D vision (3DV), pages 239-248. IEEE, 2016. 6 +[26] Martin Lambers. Survey of cube mapping methods in interactive computer graphics. The Visual Computer, 36(5): 1043-1051, 2020. 1 +[27] Yeonkun Lee, Jaeseok Jeong, Jongseob Yun, Wonjune Cho, and Kuk-Jin Yoon. Spherephd: Applying cnns on a spherical polyhedron representation of 360deg images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9181-9189, 2019. 1, 2 +[28] Yuyan Li, Yuliang Guo, Zhixin Yan, Xinyu Huang, Ye Duan, and Liu Ren. Omnifusion: 360 monocular depth estimation via geometry-aware fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2801-2810, 2022. 1, 2 +[29] Thang Luong, Hieu Pham, and Christopher D. Manning. Effective approaches to attention-based neural machine translation. In Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing. Association for Computational Linguistics, 2015. 2 +[30] Ronald Miller. An equi-rectangular map projection. Geography, pages 196–201, 1949. 1 +[31] Hongbin Pei, Bingzhe Wei, Kevin Chen-Chuan Chang, Yu Lei, and Bo Yang. Geom-gcn: Geometric graph convolutional networks. International Conference on Learning Representations, 2020. 2 +[32] Giovanni Pintore, Marco Agus, Eva Almansa, Jens Schneider, and Enrico Gobbetti. Slicenet: deep dense depth estimation from a single indoor panorama using a slice-based representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11536-11545, 2021. 1 +[33] Alec Radford, Karthik Narasimhan, Tim Salimans, Ilya Sutskever, et al. Improving language understanding by generative pre-training. 2018. 4 +[34] René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. In Proceedings of the IEEE/CVF international conference on computer vision, pages 12179-12188, 2021. 2, 4 +[35] Manuel Rey-Area, Mingze Yuan, and Christian Richardt. 360monodepth: High-resolution 360deg monocular depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3762-3772, 2022. 1, 2 +[36] Zhijie Shen, Chunyu Lin, Kang Liao, Lang Nie, Zishuo Zheng, and Yao Zhao. Panofformer: Panorama transformer for indoor $360^{\circ}$ depth estimation. In European Conference on Computer Vision, pages 195-211. Springer, 2022. 1, 2, 5, 6, 7 +[37] Yu-Chuan Su and Kristen Grauman. Learning spherical convolution for fast features from 360 imagery. Advances in Neural Information Processing Systems, 30, 2017. 2 +[38] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Hohonet: 360 indoor holistic understanding with latent horizontal features. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2573-2582, 2021. 1 + +[39] Keisuke Tateno, Nassir Navab, and Federico Tombari. Distortion-aware convolutional filters for dense prediction in panoramic images. In Proceedings of the European Conference on Computer Vision (ECCV), pages 707-722, 2018. 2 +[40] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 2, 4 +[41] Petar Velicković, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Liò, and Yoshua Bengio. Graph attention networks. In International Conference on Learning Representations, 2018. 2 +[42] Fu-En Wang, Yu-Hsuan Yeh, Min Sun, Wei-Chen Chiu, and Yi-Hsuan Tsai. Bifuse: Monocular 360 depth estimation via bi-projection fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 462–471, 2020. 1, 2 +[43] Fu-En Wang, Yu-Hsuan Yeh, Yi-Hsuan Tsai, Wei-Chen Chiu, and Min Sun. Bifuse++: Self-supervised and efficient bi-projection fusion for 360 depth estimation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(5): 5448-5460, 2022. 1, 2 +[44] Zhendong Wang, Xiaodong Cun, Jianmin Bao, Wengang Zhou, Jianzhuang Liu, and Houqiang Li. Uformer: A general u-shaped transformer for image restoration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 17683-17693, 2022. 2, 3, 4 +[45] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. Advances in Neural Information Processing Systems, 34:12077-12090, 2021. 2, 4 +[46] Qingsong Yan, Qiang Wang, Kaiyong Zhao, Bo Li, Xiaowei Chu, and Fei Deng. Sphereddepth: Panorama depth estimation from spherical domain. In 2022 International Conference on 3D Vision (3DV), pages 1-10. IEEE, 2022. 1, 2 +[47] Ilwi Yun, Hyuk-Jae Lee, and Chae Eun Rhee. Improving 360 monocular depth estimation via non-local dense prediction transformer and joint supervised and self-supervised learning. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3224-3233, 2022. 1, 2 +[48] Ilwi Yun, Chanyong Shin, Hyunku Lee, Hyuk-Jae Lee, and Chae Eun Rhee. Egformer: Equirectangular geometry-biased transformer for 360 depth estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 6101-6112, 2023. 1, 2, 5, 6, 7 +[49] Chao Zhang, Stephan Liwicki, William Smith, and Roberto Cipolla. Orientation-aware semantic segmentation on icosahedron spheres. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3533-3541, 2019. 1, 5, 6, 7 +[50] Chao Zhang, Sen He, and Stephan Liwicki. A spherical approach to planar semantic segmentation. In BMVC, 2020. 1 +[51] Jiaming Zhang, Kailun Yang, Chaoxiang Ma, Simon Reiß, Kunyu Peng, and Rainer Stiefelhagen. Bending reality: Distortion-aware transformers for adapting to panoramic se + +matic segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16917-16927, 2022. 1 +[52] Si Zhang, Hanghang Tong, Jiejun Xu, and Ross Maciejewski. Graph convolutional networks: a comprehensive review. Computational Social Networks, 6(1):1-23, 2019. 2 +[53] Jia Zheng, Junfei Zhang, Jing Li, Rui Tang, Shenghua Gao, and Zihan Zhou. Structured3d: A large photo-realistic dataset for structured 3d modeling. In Proceedings of The European Conference on Computer Vision (ECCV), 2020. 1, 5 +[54] Nikolaos Zioulis, Antonis Karakottas, Dimitrios Zarpalas, and Petros Daras. Omnidepth: Dense depth estimation for indoors spherical panoramicas. In Proceedings of the European Conference on Computer Vision (ECCV), pages 448-465, 2018. 2 \ No newline at end of file diff --git a/2025/SphereUFormer_ A U-Shaped Transformer for Spherical 360 Perception/images.zip b/2025/SphereUFormer_ A U-Shaped Transformer for Spherical 360 Perception/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..419fc7e9c3719cf4083d89c15c21951f3446e620 --- /dev/null +++ b/2025/SphereUFormer_ A U-Shaped Transformer for Spherical 360 Perception/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fb19c9d1da2abda2ecf8810c74b474f4645ac3db2168418400d6f80bc79aadf0 +size 559927 diff --git a/2025/SphereUFormer_ A U-Shaped Transformer for Spherical 360 Perception/layout.json b/2025/SphereUFormer_ A U-Shaped Transformer for Spherical 360 Perception/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..d1eca4cda27401d59a8d7557aec34846aa0b759a --- /dev/null +++ b/2025/SphereUFormer_ A U-Shaped Transformer for Spherical 360 Perception/layout.json @@ -0,0 +1,9961 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 82, + 103, + 529, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 103, + 529, + 121 + ], + "spans": [ + { + "bbox": [ + 82, + 103, + 529, + 121 + ], + "type": "text", + "content": "SphereUFormer: A U-Shaped Transformer for Spherical 360 Perception" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 187, + 144, + 283, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 187, + 144, + 283, + 172 + ], + "spans": [ + { + "bbox": [ + 187, + 144, + 283, + 172 + ], + "type": "text", + "content": "Yaniv Benny \nTel Aviv University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 326, + 144, + 423, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 144, + 423, + 172 + ], + "spans": [ + { + "bbox": [ + 326, + 144, + 423, + 172 + ], + "type": "text", + "content": "Lior Wolf \nTel Aviv University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 152, + 213, + 200, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 213, + 200, + 225 + ], + "spans": [ + { + "bbox": [ + 152, + 213, + 200, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 238, + 297, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 238, + 297, + 418 + ], + "spans": [ + { + "bbox": [ + 55, + 238, + 297, + 418 + ], + "type": "text", + "content": "This paper proposes a novel method for omnidirectional " + }, + { + "bbox": [ + 55, + 238, + 297, + 418 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 55, + 238, + 297, + 418 + ], + "type": "text", + "content": " perception. Most common previous methods relied on equirectangular projection. This representation is easily applicable to 2D operation layers but introduces distortions into the image. Other methods attempted to remove the distortions by maintaining a sphere representation but relied on complicated convolution kernels that failed to show competitive results. In this work, we introduce a transformer-based architecture that, by incorporating a novel \"Spherical Local Self-Attention\" and other spherically-oriented modules, successfully operates in the spherical domain and outperforms the state-of-the-art in " + }, + { + "bbox": [ + 55, + 238, + 297, + 418 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 55, + 238, + 297, + 418 + ], + "type": "text", + "content": " perception benchmarks for depth estimation and semantic segmentation. Our code is available at https://github.com/yanivbenny/sphere_uformer." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 441, + 135, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 441, + 135, + 453 + ], + "spans": [ + { + "bbox": [ + 55, + 441, + 135, + 453 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 461, + 295, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 461, + 295, + 592 + ], + "spans": [ + { + "bbox": [ + 55, + 461, + 295, + 592 + ], + "type": "text", + "content": "Monocular omnidirectional " + }, + { + "bbox": [ + 55, + 461, + 295, + 592 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 55, + 461, + 295, + 592 + ], + "type": "text", + "content": " perception is an important setting, as it allows for a full field-of-view (FOV) receptive field for the underlying model. Consequently, an entire room layout can be inferred, for example. In contrast to regular images, a " + }, + { + "bbox": [ + 55, + 461, + 295, + 592 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 55, + 461, + 295, + 592 + ], + "type": "text", + "content": " image has no boundaries. Instead, the image is horizontally cyclical, and converges to singular points at the vertical extreme points. This suggests that this domain requires a different treatment, and that best practices that work well with regular images might not have the same effect on " + }, + { + "bbox": [ + 55, + 461, + 295, + 592 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 55, + 461, + 295, + 592 + ], + "type": "text", + "content": " images. To tackle this task, multiple topics need to be considered." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 594, + 295, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 641 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 641 + ], + "type": "text", + "content": "How should the data be represented? What kind of distortions and irregularities are created by this representation? How will a model process this representation? How to deal with these distortions and irregularities?" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "type": "text", + "content": "These questions have attracted researchers to construct multiple datasets [3, 4, 9, 53] where depth estimation and semantic segmentation are the main tasks, and to formulate many solutions [20, 23, 28, 32, 35, 36, 38, 42, 43, 47, 48, 51], which propose novel approaches to solving these tasks. A common pattern in previous works is that they rep" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 213, + 555, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 213, + 555, + 430 + ], + "spans": [ + { + "bbox": [ + 313, + 213, + 555, + 430 + ], + "type": "text", + "content": "resent the data by projecting it to a 2D plane in different ways, usually equirectangular projection [30], cube mapping [26], or patch cropping. Each approach has its advantages and disadvantages, but the common motivation is to represent the image in a grid formation. This enables grid architectures, such as Convolutional Neural Networks or Vision Transformers to easily be applied. This comes at a cost of distortions, irregularities, or limited FOV, that are caused by the selected projection. A less common alternative representation that has been proposed is to represent the image in its original sphere domain by discretizing the sphere surface in some manner [6, 14, 17, 24]. The motivation is to maintain an undistorted representation of the " + }, + { + "bbox": [ + 313, + 213, + 555, + 430 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 313, + 213, + 555, + 430 + ], + "type": "text", + "content": " image. However, this representation formed architectural challenges that require novel solutions. Some previous works [8, 12, 27, 46, 49, 50] addressed these challenges in various manners, but they fail to compete with current state-of-the-art solutions that follow a grid representation." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 438, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 438, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 438, + 556, + 715 + ], + "type": "text", + "content": "In this work, we tackle the drawbacks of spherical representations. Specifically, we introduce SphereUFormer, which is a novel U-Shaped Transformer architecture tailored for spherical " + }, + { + "bbox": [ + 313, + 438, + 556, + 715 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 313, + 438, + 556, + 715 + ], + "type": "text", + "content": " perception. By directly operating on a spherical domain without distortion-inducing projections, SphereUFormer addresses the core challenges of omnidirectional perception. Our approach builds on the strengths of transformer models, incorporating modifications that allow it to efficiently process and understand spherical data. This includes the development of spherical-specific operations such as localized self-attention mechanism that respects the geometry of the sphere, innovative up/downsampling techniques that maintain the integrity of spherical data throughout the model, and positional encoding scheme that matches the domain properties. SphereUFormer outperforms the state of the art in the field of " + }, + { + "bbox": [ + 313, + 438, + 556, + 715 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 313, + 438, + 556, + 715 + ], + "type": "text", + "content": " perception, providing a robust and efficient tool for a wide range of applications that require comprehensive understanding of spherical environments. In Sec. 3 we describe this representation in detail. We then further describe a new method in Sec. 4, which utilizes this representation and performs " + }, + { + "bbox": [ + 313, + 438, + 556, + 715 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 313, + 438, + 556, + 715 + ], + "type": "text", + "content": " perception on spherical representations, and in Sec. 5, we analyze and compare its results." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "text", + "content": "940" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 142, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 142, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 142, + 83 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 92, + 296, + 391 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 92, + 296, + 391 + ], + "spans": [ + { + "bbox": [ + 57, + 92, + 296, + 391 + ], + "type": "text", + "content": "Omnidirectional perception. There have been many advances, novelties, and research verticals in omnidirectional perception over the years. Initial works attempted standard convolutional networks on an equirectangular projection [54] (ERP). Although promising, the results were affected by the distortions created by the projection. To better handle these distortions, later works suggested technical adjustments to the convolution kernel such as Spherical Convolutions [37] or Distortion-Aware Convolutions [12, 39]. The latter sample the convolution features from a tangent plane. Other solutions attempted to eliminate distortions by sampling the image into smaller FOV patches [16, 28, 35, 47] followed by a post-processing stage that fuses the independent results. This, however, is at a cost of a restricted receptive field per patch and possible cutoff of crucial information, which either affects results or requires high overlap between patches. Due to the advantages and limitations of each method, [2, 42, 43] suggested a combined approach that fuses the ERP path with the individual patches at some stage of the network. Alternative approaches avoided the use of projection, and instead, operated directly on the sphere by constructing a spherical mesh. Usually either an icosphere [13, 22, 27, 46] or healpix [8, 13], and apply Graph Convolution [52], Mesh Convolution [21], or some custom convolution-like operation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 408, + 296, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 408, + 296, + 576 + ], + "spans": [ + { + "bbox": [ + 55, + 408, + 296, + 576 + ], + "type": "text", + "content": "Vision Transformers. ViT's [15] have disrupted the field of computer vision, from general-purpose architectures [34, 44] to specific purposes [7, 10, 45]. The attention mechanism [5, 29, 40] allows for dynamic pooling of features based on relevance and structural position. ViT's benefit from an efficient large context window compared to their CNN counterparts, which allows them to more easily learn long-distance relations between features. While earlier omnidirectional perception works proposed solutions based on CNN's, recent contributions replaced the convolution kernels with local attention of various forms [20, 36, 48] with favorable results. This work follows this direction. It utilizes the foundation of UFormer [44] and adapts it to spherical representations with custom attention layers." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "content": "Graph Attention. Graph Convolution Networks [52] (GCN) are neural networks that apply convolution-like operations on data represented as a graph. A graph convolution operation aggregates the results of local operations between a node and each of its neighbors. Over multiple layers, this allows for information to pass between distant nodes in the graph and the update of these nodes. Graph Attention Networks [41] were proposed to enhance Graph Convolution Networks, thereby improving the message-passing capabilities of the GCN. The attention mechanism" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 326, + 72, + 370, + 110 + ], + "blocks": [ + { + "bbox": [ + 326, + 72, + 370, + 110 + ], + "lines": [ + { + "bbox": [ + 326, + 72, + 370, + 110 + ], + "spans": [ + { + "bbox": [ + 326, + 72, + 370, + 110 + ], + "type": "image", + "image_path": "f6b9bb34f3fa26282bcf785443990c35dec9582ee7f7d4e8cdc5f972ad1f70a0.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 341, + 118, + 351, + 125 + ], + "lines": [ + { + "bbox": [ + 341, + 118, + 351, + 125 + ], + "spans": [ + { + "bbox": [ + 341, + 118, + 351, + 125 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 384, + 72, + 429, + 110 + ], + "blocks": [ + { + "bbox": [ + 384, + 72, + 429, + 110 + ], + "lines": [ + { + "bbox": [ + 384, + 72, + 429, + 110 + ], + "spans": [ + { + "bbox": [ + 384, + 72, + 429, + 110 + ], + "type": "image", + "image_path": "4554437dc1bffff70ead73229baf292b2c3060de5bfa7e2353e068a0d261d2ac.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 400, + 118, + 410, + 125 + ], + "lines": [ + { + "bbox": [ + 400, + 118, + 410, + 125 + ], + "spans": [ + { + "bbox": [ + 400, + 118, + 410, + 125 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 136, + 553, + 158 + ], + "lines": [ + { + "bbox": [ + 313, + 136, + 553, + 158 + ], + "spans": [ + { + "bbox": [ + 313, + 136, + 553, + 158 + ], + "type": "text", + "content": "Figure 1. Sphere Representations. From left to right: uvsphere, cubesphere, icosphere, hexasphere." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 443, + 72, + 487, + 110 + ], + "blocks": [ + { + "bbox": [ + 443, + 72, + 487, + 110 + ], + "lines": [ + { + "bbox": [ + 443, + 72, + 487, + 110 + ], + "spans": [ + { + "bbox": [ + 443, + 72, + 487, + 110 + ], + "type": "image", + "image_path": "c29ad352dee20996e1614884e1ccab961625b85196b11fce84aa5140276d2df0.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 459, + 118, + 469, + 125 + ], + "lines": [ + { + "bbox": [ + 459, + 118, + 469, + 125 + ], + "spans": [ + { + "bbox": [ + 459, + 118, + 469, + 125 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 502, + 72, + 545, + 110 + ], + "blocks": [ + { + "bbox": [ + 502, + 72, + 545, + 110 + ], + "lines": [ + { + "bbox": [ + 502, + 72, + 545, + 110 + ], + "spans": [ + { + "bbox": [ + 502, + 72, + 545, + 110 + ], + "type": "image", + "image_path": "ce2dc61b351747cfde2dc7c64a8cbd4c193d000a3ab068314f2ed4529853b8a4.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 518, + 116, + 529, + 125 + ], + "lines": [ + { + "bbox": [ + 518, + 116, + 529, + 125 + ], + "spans": [ + { + "bbox": [ + 518, + 116, + 529, + 125 + ], + "type": "text", + "content": "(d)" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 179, + 555, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 179, + 555, + 263 + ], + "spans": [ + { + "bbox": [ + 313, + 179, + 555, + 263 + ], + "type": "text", + "content": "has the benefit of allowing the operations to focus on a few sets of relevant nodes and providing an efficient method to encode positional information in the graph. For Geometric Graphs [31] geometric location is encoded to inform of each node's absolute and relative position, thereby making nodes aware of their global position and their respective neighbors' relative position." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 274, + 446, + 288 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 274, + 446, + 288 + ], + "spans": [ + { + "bbox": [ + 313, + 274, + 446, + 288 + ], + "type": "text", + "content": "3. Sphere Representation" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 295, + 555, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 295, + 555, + 425 + ], + "spans": [ + { + "bbox": [ + 313, + 295, + 555, + 425 + ], + "type": "text", + "content": "Arguably the first topic in question when attempting to solve a problem is how to formulate the data. For omnidirectional " + }, + { + "bbox": [ + 313, + 295, + 555, + 425 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 313, + 295, + 555, + 425 + ], + "type": "text", + "content": " images, the data can be seen as an array of values " + }, + { + "bbox": [ + 313, + 295, + 555, + 425 + ], + "type": "inline_equation", + "content": "v_{\\phi ,\\theta}" + }, + { + "bbox": [ + 313, + 295, + 555, + 425 + ], + "type": "text", + "content": " assigned to a ray pointing from a center point-of-view towards " + }, + { + "bbox": [ + 313, + 295, + 555, + 425 + ], + "type": "inline_equation", + "content": "(\\theta ,\\phi)" + }, + { + "bbox": [ + 313, + 295, + 555, + 425 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 313, + 295, + 555, + 425 + ], + "type": "inline_equation", + "content": "\\theta \\in [0,2\\pi ]" + }, + { + "bbox": [ + 313, + 295, + 555, + 425 + ], + "type": "text", + "content": " being the horizontal angle and " + }, + { + "bbox": [ + 313, + 295, + 555, + 425 + ], + "type": "inline_equation", + "content": "\\phi \\in [0,\\pi ]" + }, + { + "bbox": [ + 313, + 295, + 555, + 425 + ], + "type": "text", + "content": " the vertical. The value " + }, + { + "bbox": [ + 313, + 295, + 555, + 425 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 313, + 295, + 555, + 425 + ], + "type": "text", + "content": " can correspond to the RGB value, depth, object category, etc. By selecting to sample on finite such angle pairs, the sphere becomes discretized. The sampling method determines how the sampled data can be processed and how successful the solution will become." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 426, + 554, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 554, + 557 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 554, + 557 + ], + "type": "text", + "content": "The most common and straightforward is a uvsphere [17], where points are sampled in a grid with equal spacing for " + }, + { + "bbox": [ + 313, + 426, + 554, + 557 + ], + "type": "inline_equation", + "content": "(\\theta, \\phi)" + }, + { + "bbox": [ + 313, + 426, + 554, + 557 + ], + "type": "text", + "content": ". See Fig. 1a. The uvsphere has a very high degree of horizontal symmetry, which is a good property. Another advantage is that it allows an unwrapping of the sphere into a 2D matrix. This unwrapping is called equirectangular projection (ERP). The downside of this representation is that the density of samples on the sphere is higher around the poles, due to the decreasing cross-section. This creates an imbalanced effective resolution over the sphere, and the projected image in 2D is distorted." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "text", + "content": "An alternative that mitigates imbalanced resolution and distortion is cube mapping (or cubosphere) [14] (Fig. 1b), which projects the sphere onto the faces of a cube. Sample resolution is controlled by deciding how many rectangles each face of the cube is subdivided into. This removes the sampling density around the poles and the sphere can be unwrapped into six 2D images. However, two of these images are not vertically oriented and require special treatment. In addition, the transition between faces is not fluent, and complex padding methods [11] can only increase the receptive field of each image to some extent. Since the images are separated, a post-processing stage needs to fuse their predictions to form a unified prediction on the sphere. One op-" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 312, + 757 + ], + "type": "text", + "content": "941" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 65, + 75, + 97, + 114 + ], + "blocks": [ + { + "bbox": [ + 65, + 75, + 97, + 114 + ], + "lines": [ + { + "bbox": [ + 65, + 75, + 97, + 114 + ], + "spans": [ + { + "bbox": [ + 65, + 75, + 97, + 114 + ], + "type": "image", + "image_path": "dd82f349e41cfb609088cc68eb7fb79b93e1e3520780d4cf91e4de59c751f6c9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 78, + 122, + 85, + 131 + ], + "lines": [ + { + "bbox": [ + 78, + 122, + 85, + 131 + ], + "spans": [ + { + "bbox": [ + 78, + 122, + 85, + 131 + ], + "type": "text", + "content": "0" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 109, + 76, + 147, + 114 + ], + "blocks": [ + { + "bbox": [ + 109, + 76, + 147, + 114 + ], + "lines": [ + { + "bbox": [ + 109, + 76, + 147, + 114 + ], + "spans": [ + { + "bbox": [ + 109, + 76, + 147, + 114 + ], + "type": "image", + "image_path": "1e48d5f8596e033fdd8b70b3a1126c05fd15fa75f99f804c50786b22e6d43c4b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 126, + 122, + 130, + 131 + ], + "lines": [ + { + "bbox": [ + 126, + 122, + 130, + 131 + ], + "spans": [ + { + "bbox": [ + 126, + 122, + 130, + 131 + ], + "type": "text", + "content": "1" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 156, + 75, + 195, + 114 + ], + "blocks": [ + { + "bbox": [ + 156, + 75, + 195, + 114 + ], + "lines": [ + { + "bbox": [ + 156, + 75, + 195, + 114 + ], + "spans": [ + { + "bbox": [ + 156, + 75, + 195, + 114 + ], + "type": "image", + "image_path": "83a76e3f3ded829311a02ebe795e95d25d38547ac549c2a431c1a1ebc07808fb.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 173, + 122, + 179, + 131 + ], + "lines": [ + { + "bbox": [ + 173, + 122, + 179, + 131 + ], + "spans": [ + { + "bbox": [ + 173, + 122, + 179, + 131 + ], + "type": "text", + "content": "2" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 143, + 295, + 176 + ], + "lines": [ + { + "bbox": [ + 55, + 143, + 295, + 176 + ], + "spans": [ + { + "bbox": [ + 55, + 143, + 295, + 176 + ], + "type": "text", + "content": "Figure 2. Icospheres of different ranks. An increase in rank is made by subdividing each triangle into 4 smaller triangles, and results in an increased resolution." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 203, + 75, + 242, + 114 + ], + "blocks": [ + { + "bbox": [ + 203, + 75, + 242, + 114 + ], + "lines": [ + { + "bbox": [ + 203, + 75, + 242, + 114 + ], + "spans": [ + { + "bbox": [ + 203, + 75, + 242, + 114 + ], + "type": "image", + "image_path": "6e43b0e4efa2e8f54e574c6f85bed29998c43d45cecf8b01791a1ea603d0c7ba.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 220, + 122, + 226, + 131 + ], + "lines": [ + { + "bbox": [ + 220, + 122, + 226, + 131 + ], + "spans": [ + { + "bbox": [ + 220, + 122, + 226, + 131 + ], + "type": "text", + "content": "3" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 251, + 75, + 290, + 114 + ], + "blocks": [ + { + "bbox": [ + 251, + 75, + 290, + 114 + ], + "lines": [ + { + "bbox": [ + 251, + 75, + 290, + 114 + ], + "spans": [ + { + "bbox": [ + 251, + 75, + 290, + 114 + ], + "type": "image", + "image_path": "be04b33dde31b4d8a742c4fe36839a736ed0ff06874bb0a1577df740b574458e.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 267, + 122, + 273, + 131 + ], + "lines": [ + { + "bbox": [ + 267, + 122, + 273, + 131 + ], + "spans": [ + { + "bbox": [ + 267, + 122, + 273, + 131 + ], + "type": "text", + "content": "4" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 67, + 190, + 286, + 232 + ], + "blocks": [ + { + "bbox": [ + 67, + 190, + 286, + 232 + ], + "lines": [ + { + "bbox": [ + 67, + 190, + 286, + 232 + ], + "spans": [ + { + "bbox": [ + 67, + 190, + 286, + 232 + ], + "type": "table", + "html": "
Rank:01234567
Faces:208032012805120~20K~82K~328K
Vertices:12421626422562~10K~41K~164K
", + "image_path": "50ba434a066253c32422efd9e99c0a1568bf1181707ecfd93feb41faa6709e4f.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 241, + 295, + 264 + ], + "lines": [ + { + "bbox": [ + 55, + 241, + 295, + 264 + ], + "spans": [ + { + "bbox": [ + 55, + 241, + 295, + 264 + ], + "type": "text", + "content": "Table 1. Icosphere Properties. Number of faces " + }, + { + "bbox": [ + 55, + 241, + 295, + 264 + ], + "type": "inline_equation", + "content": "(20 \\times 4^i)" + }, + { + "bbox": [ + 55, + 241, + 295, + 264 + ], + "type": "text", + "content": " and number of vertices " + }, + { + "bbox": [ + 55, + 241, + 295, + 264 + ], + "type": "inline_equation", + "content": "(10 \\times 4^i + 2)" + }, + { + "bbox": [ + 55, + 241, + 295, + 264 + ], + "type": "text", + "content": " per subdivision rank " + }, + { + "bbox": [ + 55, + 241, + 295, + 264 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 241, + 295, + 264 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 62, + 281, + 99, + 319 + ], + "blocks": [ + { + "bbox": [ + 62, + 281, + 99, + 319 + ], + "lines": [ + { + "bbox": [ + 62, + 281, + 99, + 319 + ], + "spans": [ + { + "bbox": [ + 62, + 281, + 99, + 319 + ], + "type": "image", + "image_path": "f7724cd19a4b4836e19e5a61040721a9dfa0e43c0503c8298b902909e9532aea.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 74, + 324, + 88, + 333 + ], + "lines": [ + { + "bbox": [ + 74, + 324, + 88, + 333 + ], + "spans": [ + { + "bbox": [ + 74, + 324, + 88, + 333 + ], + "type": "text", + "content": "(a1)" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 343, + 295, + 366 + ], + "lines": [ + { + "bbox": [ + 55, + 343, + 295, + 366 + ], + "spans": [ + { + "bbox": [ + 55, + 343, + 295, + 366 + ], + "type": "text", + "content": "Figure 3. Data as Icospheres. Icosphere of rank 3 (a1-c1) and rank 6 (a2-c2). (a) RGB, (b) Depth Map, and (c) Semantic Layout." + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 101, + 281, + 137, + 319 + ], + "blocks": [ + { + "bbox": [ + 101, + 281, + 137, + 319 + ], + "lines": [ + { + "bbox": [ + 101, + 281, + 137, + 319 + ], + "spans": [ + { + "bbox": [ + 101, + 281, + 137, + 319 + ], + "type": "image", + "image_path": "4e3d73cc128036581d3f3c47d3bbf23711ce7fdc9ca766e22ab5707c309d82c6.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 112, + 324, + 126, + 333 + ], + "lines": [ + { + "bbox": [ + 112, + 324, + 126, + 333 + ], + "spans": [ + { + "bbox": [ + 112, + 324, + 126, + 333 + ], + "type": "text", + "content": "(b1)" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 139, + 281, + 174, + 319 + ], + "blocks": [ + { + "bbox": [ + 139, + 281, + 174, + 319 + ], + "lines": [ + { + "bbox": [ + 139, + 281, + 174, + 319 + ], + "spans": [ + { + "bbox": [ + 139, + 281, + 174, + 319 + ], + "type": "image", + "image_path": "959764e5059a9c618a84e49fcbd7c0764389ef0da880f2ab4ebfcefdbc1e3215.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 149, + 324, + 164, + 333 + ], + "lines": [ + { + "bbox": [ + 149, + 324, + 164, + 333 + ], + "spans": [ + { + "bbox": [ + 149, + 324, + 164, + 333 + ], + "type": "text", + "content": "(c1)" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 177, + 281, + 212, + 319 + ], + "blocks": [ + { + "bbox": [ + 177, + 281, + 212, + 319 + ], + "lines": [ + { + "bbox": [ + 177, + 281, + 212, + 319 + ], + "spans": [ + { + "bbox": [ + 177, + 281, + 212, + 319 + ], + "type": "image", + "image_path": "1fff96c2dadca431aea67e0f12be9dc718bc1b5ed18fd284a6d66db575fa2561.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 188, + 324, + 201, + 333 + ], + "lines": [ + { + "bbox": [ + 188, + 324, + 201, + 333 + ], + "spans": [ + { + "bbox": [ + 188, + 324, + 201, + 333 + ], + "type": "text", + "content": "(a2)" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 214, + 281, + 250, + 319 + ], + "blocks": [ + { + "bbox": [ + 214, + 281, + 250, + 319 + ], + "lines": [ + { + "bbox": [ + 214, + 281, + 250, + 319 + ], + "spans": [ + { + "bbox": [ + 214, + 281, + 250, + 319 + ], + "type": "image", + "image_path": "1d63be6755d21db5f8fcbc6d3052698685442f334b5cb6cdac4c0470503afd61.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 225, + 324, + 240, + 333 + ], + "lines": [ + { + "bbox": [ + 225, + 324, + 240, + 333 + ], + "spans": [ + { + "bbox": [ + 225, + 324, + 240, + 333 + ], + "type": "text", + "content": "(b2)" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 253, + 281, + 288, + 319 + ], + "blocks": [ + { + "bbox": [ + 253, + 281, + 288, + 319 + ], + "lines": [ + { + "bbox": [ + 253, + 281, + 288, + 319 + ], + "spans": [ + { + "bbox": [ + 253, + 281, + 288, + 319 + ], + "type": "image", + "image_path": "4ce7b793d7f57c13ca3bb2a0c052bb2cdecf884a4cdb2f9f40fb19bfb7e2196c.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 264, + 324, + 277, + 333 + ], + "lines": [ + { + "bbox": [ + 264, + 324, + 277, + 333 + ], + "spans": [ + { + "bbox": [ + 264, + 324, + 277, + 333 + ], + "type": "text", + "content": "(c2)" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "bbox": [ + 55, + 388, + 295, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 388, + 295, + 449 + ], + "spans": [ + { + "bbox": [ + 55, + 388, + 295, + 449 + ], + "type": "text", + "content": "tion is to operate on the cubosphere mesh instead of the 2D projection. Unfortunately, the cubosphere has irregularities at the coordinates of the cube corners and does not have a high horizontal symmetry, which makes it less optimal for this purpose." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 55, + 449, + 295, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 449, + 295, + 676 + ], + "spans": [ + { + "bbox": [ + 55, + 449, + 295, + 676 + ], + "type": "text", + "content": "Instead of projecting the data to a grid and then having to deal with its limitations, a more elegant approach can be formulated using the sphere representation directly. Aside from the previously mentioned options, a common way to discretize a sphere is with Geodesic (icosphere, Fig. 1c) and Goldberg (hexasphere, Fig. 1d) polyhedrons [6]. An icosphere is created by repeatedly subdividing an icosahedron into smaller triangles. A hexasphere is the dual polyhedron where vertices are swapped with face normals, and edge connections with face adjacency. These types of sphere representation have a very high degree of symmetry and have a very evenly spaced sampling on the sphere surface. Due to their duality, our work follows the icosphere mesh, and icosphere/hexasphere mode is toggled by whether data points " + }, + { + "bbox": [ + 55, + 449, + 295, + 676 + ], + "type": "inline_equation", + "content": "(\\theta, \\phi)" + }, + { + "bbox": [ + 55, + 449, + 295, + 676 + ], + "type": "text", + "content": " describe mesh vertices (hexasphere) or face normals (icosphere). Therefore we will sometimes refer to icospheres/hexaspheres by referring to the \"node type\" (face/vertex respectively). Other notable options for sphere sampling are Fibonacci spheres [24] and HealPix [19]." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "content": "Icospheres have the highest order of symmetry, the most uniform distribution of points (perfect uniform distribution is impossible), easy setup, and an inherent up/down" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 313, + 72, + 553, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 240 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 240 + ], + "type": "text", + "content": "sampling definition through subdivision. All these reasons make them ideal for discrete spherical representation. As Fig. 2 and Tab. 1 show, in different ranks, icospheres produce sphere discretization in various resolutions. Each increase in sphere rank divides each triangle face into four smaller ones, hereby gradually increasing the sphere resolution. An example of representing data with icospheres can be seen in Fig. 3. In low resolution, the image seems \"pixelated\" as the independent triangle faces are distinguishable. But, with a high enough resolution, the image becomes detailed and visibly smooth. Our method uses high-resolution icospheres to represent the " + }, + { + "bbox": [ + 313, + 72, + 553, + 240 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 313, + 72, + 553, + 240 + ], + "type": "text", + "content": " image, and through up and downsampling on the sphere structure, uses low-resolution spheres as hidden layers inside the model." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 314, + 250, + 370, + 262 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 250, + 370, + 262 + ], + "spans": [ + { + "bbox": [ + 314, + 250, + 370, + 262 + ], + "type": "text", + "content": "4. Method" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 313, + 270, + 553, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 270, + 553, + 354 + ], + "spans": [ + { + "bbox": [ + 313, + 270, + 553, + 354 + ], + "type": "text", + "content": "This section introduces SphereUFormer, our proposed U-shaped Transformer for Spherical " + }, + { + "bbox": [ + 313, + 270, + 553, + 354 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 313, + 270, + 553, + 354 + ], + "type": "text", + "content": " Perception. As the name suggests, the model is influenced by UFormer [44], with significant modifications to support spherically structured data. A high-level diagram is depicted in Fig. 4. The model's components are described in this section. For technical implementation details, see the supplementary." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 313, + 354, + 554, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 354, + 554, + 485 + ], + "spans": [ + { + "bbox": [ + 313, + 354, + 554, + 485 + ], + "type": "text", + "content": "The network begins with an input projection that applies linear projection from the RGB values to latent embedding vectors. The projection output is then passed into an Encoder-Decoder network with multiple levels of self-attention modules in different sphere resolutions. We call these modules \"Spherical Attention Modules\" (SAM). At the bottom of the network lies a bottleneck SAM module that operates on the lowest resolution. Skip connections pass from the encoder to the decoder and feed high-resolution features. Finally, a linear Output Projection projects the decoder output to the output dimensions." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 313, + 486, + 553, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 486, + 553, + 544 + ], + "spans": [ + { + "bbox": [ + 313, + 486, + 553, + 544 + ], + "type": "text", + "content": "As described in Sec. 3, our solution features the icosphere as its spherical representation structure, as we have found it to be the most intuitive and to contain the most satisfying properties. But it is not limited to it, and with minor modifications supports any spherical structure." + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 313, + 545, + 553, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 553, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 553, + 653 + ], + "type": "text", + "content": "Our method contains many local operations on the graph, from the down and up sampling, to the self-attention mechanism. Computing the cosine similarity or finding the neighbors of a node to generate the mapping scheme is slightly costly. This would be a problem if it was performed on each iteration, as the model would be slowed down. Luckily, since the icosphere graph is fixed, these mappings only need to be computed once in advance. Therefore there is no need to recompute it in every message massing operation." + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 313, + 660, + 429, + 673 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 660, + 429, + 673 + ], + "spans": [ + { + "bbox": [ + 313, + 660, + 429, + 673 + ], + "type": "text", + "content": "4.1. Up/Down Sampling" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "content": "In the spherical setting, upsampling and downsampling operations are performed by transforming a spherical graph into a new one with more or fewer nodes, respectively. This" + } + ] + } + ], + "index": 36 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 748, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 748, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 748, + 313, + 757 + ], + "type": "text", + "content": "942" + } + ] + } + ], + "index": 37 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 65, + 70, + 552, + 248 + ], + "blocks": [ + { + "bbox": [ + 65, + 70, + 552, + 248 + ], + "lines": [ + { + "bbox": [ + 65, + 70, + 552, + 248 + ], + "spans": [ + { + "bbox": [ + 65, + 70, + 552, + 248 + ], + "type": "image", + "image_path": "9a998734b9ccfb2e19d2e04bb23a716bb5d948d46206bd12a0d0252d671cb333.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 255, + 555, + 301 + ], + "lines": [ + { + "bbox": [ + 54, + 255, + 555, + 301 + ], + "spans": [ + { + "bbox": [ + 54, + 255, + 555, + 301 + ], + "type": "text", + "content": "Figure 4. The SphereUFormer architecture. A spherical representation is fed into the model. A linear input projection layer encodes the RGB values to latent embedding vectors. A sequence of SAM modules apply local self-attention on the spherical data along with downsampling layers that gradually reduce the resolution of the sphere. A sequence of SAM modules along with an upsampling layer and bypass skip connections decode the data. An output projection converts the latent embeddings to the output channel size." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 320, + 296, + 438 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 320, + 296, + 438 + ], + "spans": [ + { + "bbox": [ + 54, + 320, + 296, + 438 + ], + "type": "text", + "content": "can be difficult and inefficient to compute for two arbitrary spherical polyhedrons. However, icospheres have multiple elegant and desired properties. First, an increase in resolution by one rank is the result of splitting each triangle face into four smaller ones, with one central triangle that shares its center with the large one. Second, for the same reason, each vertex in the low-resolution icosphere also exists in the high-resolution one, with new vertices created at the center of each edge connection. Third, the number of nodes/faces increases at a constant rate." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 451, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 451, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 451, + 296, + 713 + ], + "type": "text", + "content": "All these properties make constructing up and down operations extremely straightforward and efficient. Every common pooling technique, such as max pooling, average pooling, center pooling, and interpolation, can be implemented. However, there are preferences for each setting. For downsampling, when the data nodes are represented by the faces, we have found max and center pooling to have the same overall results. When the data nodes are on the vertices (equivalent to hexosphere), max and average pooling become a little complicated, since there are both instances of five and six neighbors in the graph. Therefore, we resorted to only center pooling. For upsampling, the faces mode most simply operates with nearest upsampling, where each triangle receives the value according to the lower resolution triangle it was subdivided from. Other interpolation methods require more complicated computations and were therefore not considered. For vertices, however, interpolation is straightforward, since each new node is located precisely at the center of an existing edge. For more upscaled upsampling, Barycentric coordinates [18] interpolation can also be performed quite easily, but we did not find a reason to upscale beyond a factor of 2 at a time." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 319, + 465, + 333 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 319, + 465, + 333 + ], + "spans": [ + { + "bbox": [ + 313, + 319, + 465, + 333 + ], + "type": "text", + "content": "4.2. Global Positional Encoding" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 338, + 555, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 338, + 555, + 517 + ], + "spans": [ + { + "bbox": [ + 313, + 338, + 555, + 517 + ], + "type": "text", + "content": "Positional encoding is a common and critical technique in Transformers [33, 40] and Vision Transformers [7, 10, 15, 34, 44, 45]. SphereUFormer utilizes positional encoding as well, but in a manner that suits spherical representations. The node position on the unit sphere is expressed using spherical coordinates " + }, + { + "bbox": [ + 313, + 338, + 555, + 517 + ], + "type": "inline_equation", + "content": "(\\theta, \\phi)" + }, + { + "bbox": [ + 313, + 338, + 555, + 517 + ], + "type": "text", + "content": ". In our setting of omnidirectional perception, the data is always vertically aligned, therefore the model does not need to be vertically equivariant. However, the horizontal rotation is arbitrary, therefore it is desired to make the model horizontally equivariant to rotation and flipping. We therefore apply absolute positional encoding for the vertical position but not for the horizontal position. Horizontal position is only used in relative to another node as position bias in the Local Self-Attention described in the next section." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 517, + 556, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 517, + 556, + 602 + ], + "spans": [ + { + "bbox": [ + 313, + 517, + 556, + 602 + ], + "type": "text", + "content": "For the vertical position, we consider " + }, + { + "bbox": [ + 313, + 517, + 556, + 602 + ], + "type": "inline_equation", + "content": "\\phi \\in [0,\\pi ]" + }, + { + "bbox": [ + 313, + 517, + 556, + 602 + ], + "type": "text", + "content": ". The position was encoded by applying sinusoidal encoding on the position value, then passed through an MLP with an output dimension equal to the embedding dimension. We apply the vertical global positional encoding on the projected input and on the queries and keys in each attention module. More details can be found in the supplement." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 611, + 479, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 611, + 479, + 624 + ], + "spans": [ + { + "bbox": [ + 313, + 611, + 479, + 624 + ], + "type": "text", + "content": "4.3. Spherical Local Self-Attention" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "content": "The Spherical Local Self-Attention is the main component in our network that drives the SphereUFormer's " + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "content": " perception. A detailed diagram is shown in Fig. 5. Given an input sphere, each node " + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "content": " in the sphere is projected into a query " + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "inline_equation", + "content": "(q_{i})" + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "content": ", key " + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "inline_equation", + "content": "(k_{i})" + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "content": ", and value " + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "inline_equation", + "content": "(v_{i})" + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "content": ". These vectors are all of the shape " + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "inline_equation", + "content": "(N,L,H,D_H)" + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "content": " is the number of heads and " + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "inline_equation", + "content": "D_H" + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "content": " is the head dimension. Similar to standard prac" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 748, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 748, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 748, + 312, + 757 + ], + "type": "text", + "content": "943" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 69, + 298, + 194 + ], + "blocks": [ + { + "bbox": [ + 58, + 69, + 298, + 194 + ], + "lines": [ + { + "bbox": [ + 58, + 69, + 298, + 194 + ], + "spans": [ + { + "bbox": [ + 58, + 69, + 298, + 194 + ], + "type": "image", + "image_path": "8e3578f1b6abdee667ebccf5518ea29ee3fdacd56bbc9bdacb713f51a58614fa.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 200, + 297, + 247 + ], + "lines": [ + { + "bbox": [ + 55, + 200, + 297, + 247 + ], + "spans": [ + { + "bbox": [ + 55, + 200, + 297, + 247 + ], + "type": "text", + "content": "Figure 5. Spherical Local Self Attention. Attention is applied between each data node and its " + }, + { + "bbox": [ + 55, + 200, + 297, + 247 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 200, + 297, + 247 + ], + "type": "text", + "content": " neighbors. A learned relative position bias encodes information about the neighbors' relative position. In the right corner is a diagram of the enclosing block." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 69, + 258, + 293, + 376 + ], + "blocks": [ + { + "bbox": [ + 69, + 258, + 293, + 376 + ], + "lines": [ + { + "bbox": [ + 69, + 258, + 293, + 376 + ], + "spans": [ + { + "bbox": [ + 69, + 258, + 293, + 376 + ], + "type": "image", + "image_path": "3df419116959d1c79b9db42eaf035bb3d4833408380c9242438277b18bbf1a65.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 383, + 296, + 473 + ], + "lines": [ + { + "bbox": [ + 55, + 383, + 296, + 473 + ], + "spans": [ + { + "bbox": [ + 55, + 383, + 296, + 473 + ], + "type": "text", + "content": "Figure 6. Relative Positional Encondig. On the left, a center point (blue) and its first (green) and second (yellow) degree neighbors are defined by the graph structure. The relative position kernel is defined by a " + }, + { + "bbox": [ + 55, + 383, + 296, + 473 + ], + "type": "inline_equation", + "content": "7 \\times 7" + }, + { + "bbox": [ + 55, + 383, + 296, + 473 + ], + "type": "text", + "content": " grid, represented by the black nodes in the right image. The center point is positioned at the center of the grid and the neighbors are distributed across the grid according to their relative change in horizontal (" + }, + { + "bbox": [ + 55, + 383, + 296, + 473 + ], + "type": "inline_equation", + "content": "\\Delta \\theta" + }, + { + "bbox": [ + 55, + 383, + 296, + 473 + ], + "type": "text", + "content": ") and vertical (" + }, + { + "bbox": [ + 55, + 383, + 296, + 473 + ], + "type": "inline_equation", + "content": "\\Delta \\phi" + }, + { + "bbox": [ + 55, + 383, + 296, + 473 + ], + "type": "text", + "content": ") angles. Bilinear interpolation samples a learned positional encoding." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 496, + 296, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 496, + 296, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 496, + 296, + 628 + ], + "type": "text", + "content": "tices, " + }, + { + "bbox": [ + 55, + 496, + 296, + 628 + ], + "type": "inline_equation", + "content": "D_H" + }, + { + "bbox": [ + 55, + 496, + 296, + 628 + ], + "type": "text", + "content": " is derived from the input dimension " + }, + { + "bbox": [ + 55, + 496, + 296, + 628 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 55, + 496, + 296, + 628 + ], + "type": "text", + "content": " and the number of heads. However, since the self-attention mechanism is the only spatial operation in the network, we added a \"head dimension coefficient\" (" + }, + { + "bbox": [ + 55, + 496, + 296, + 628 + ], + "type": "inline_equation", + "content": "C_{head}" + }, + { + "bbox": [ + 55, + 496, + 296, + 628 + ], + "type": "text", + "content": ") hyperparameter to upscale the head dimension " + }, + { + "bbox": [ + 55, + 496, + 296, + 628 + ], + "type": "inline_equation", + "content": "D_H" + }, + { + "bbox": [ + 55, + 496, + 296, + 628 + ], + "type": "text", + "content": " as a reverse bottleneck. Thus: " + }, + { + "bbox": [ + 55, + 496, + 296, + 628 + ], + "type": "inline_equation", + "content": "D_H = (D / H) \\cdot C_{head}" + }, + { + "bbox": [ + 55, + 496, + 296, + 628 + ], + "type": "text", + "content": ". This coefficient allows an increase in the number of attention heads without reducing the dimension of each head and without increasing the overall size of the entire network, resulting in only a minor increase in parameters. We further discuss the motivation of this in the ablation section (Sec. 5.4)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 630, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 296, + 715 + ], + "type": "text", + "content": "For each query, the neighboring " + }, + { + "bbox": [ + 55, + 630, + 296, + 715 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 630, + 296, + 715 + ], + "type": "text", + "content": " vectors are grouped according to the sphere geometry. Attention for each query is performed on the selected neighboring group. The number of neighbors " + }, + { + "bbox": [ + 55, + 630, + 296, + 715 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 630, + 296, + 715 + ], + "type": "text", + "content": " depends on a hyperparameter named \"window coefficient\" (" + }, + { + "bbox": [ + 55, + 630, + 296, + 715 + ], + "type": "inline_equation", + "content": "C_{win}" + }, + { + "bbox": [ + 55, + 630, + 296, + 715 + ], + "type": "text", + "content": "), which dictates the N-order neighbors to collect. A window coefficient of 0 provides no neighbors, a coefficient of 1 provides only first-order" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": "neighbors, 2 also provides second-order neighbors, and so forth. Note that due to irregularities in the graph structure, the number of keys for each node is not identical as some have slightly fewer neighbors than others. This is homogenized for parallelization purposes by adding null neighbor keys and masking them." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 144, + 555, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 144, + 555, + 373 + ], + "spans": [ + { + "bbox": [ + 313, + 144, + 555, + 373 + ], + "type": "text", + "content": "For relative position information, a learned relative position bias is added to the attention map. Since having a learned parameter per query-key pair is both memory expensive and harmful in terms of inductive bias, we took a shared-weights approach. For each node " + }, + { + "bbox": [ + 313, + 144, + 555, + 373 + ], + "type": "inline_equation", + "content": "i \\in [1, L]" + }, + { + "bbox": [ + 313, + 144, + 555, + 373 + ], + "type": "text", + "content": ", we measure the angular difference " + }, + { + "bbox": [ + 313, + 144, + 555, + 373 + ], + "type": "inline_equation", + "content": "(\\Delta \\phi_{i,k}, \\Delta \\theta_{i,k})" + }, + { + "bbox": [ + 313, + 144, + 555, + 373 + ], + "type": "text", + "content": " for each key " + }, + { + "bbox": [ + 313, + 144, + 555, + 373 + ], + "type": "inline_equation", + "content": "k \\in [1, K]" + }, + { + "bbox": [ + 313, + 144, + 555, + 373 + ], + "type": "text", + "content": " and normalize them by the max deltas. This is only computed once during initialization. During runtime, the normalized deltas (in " + }, + { + "bbox": [ + 313, + 144, + 555, + 373 + ], + "type": "inline_equation", + "content": "[-1, 1]" + }, + { + "bbox": [ + 313, + 144, + 555, + 373 + ], + "type": "text", + "content": ") are used to sample learned parameters from a " + }, + { + "bbox": [ + 313, + 144, + 555, + 373 + ], + "type": "inline_equation", + "content": "7 \\times 7" + }, + { + "bbox": [ + 313, + 144, + 555, + 373 + ], + "type": "text", + "content": " window. This sampling from a non-linear 2D function acts as a cheap relative positional encoding for the query-key pairs. Fig. 6 illustrates the relative position implementation. Given the blue center points, its first-order neighbors are depicted in green, and second-order neighbors in yellow, are placed in their relative position. A " + }, + { + "bbox": [ + 313, + 144, + 555, + 373 + ], + "type": "inline_equation", + "content": "7 \\times 7" + }, + { + "bbox": [ + 313, + 144, + 555, + 373 + ], + "type": "text", + "content": " grid, represented by the black points, represents the learned weights of the grid. Each neighbor gets its positional encoding by bilinearly interpolating between the four nearby weights." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 381, + 397, + 396 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 381, + 397, + 396 + ], + "spans": [ + { + "bbox": [ + 313, + 381, + 397, + 396 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 402, + 554, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 402, + 554, + 498 + ], + "spans": [ + { + "bbox": [ + 313, + 402, + 554, + 498 + ], + "type": "text", + "content": "We evaluate our method by comparing it with state-of-the-art methods for depth estimation and semantic segmentation. For this purpose, we used the Stanford2D3D [4] and Structured3D [53] datasets. Both contain RGB, depth, and semantic segmentation data. We trained our model and all baselines using the training configuration of PanoFormer [36]. For more information on the training protocol please refer to the supplementary." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 498, + 554, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 554, + 558 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 554, + 558 + ], + "type": "text", + "content": "We compared against ERP methods [20, 36, 48], spherical methods [8, 49], and Elite360D [1], that utilizes a combination of ERP with a low-resolution icosphere. We did not compare against patch-based solutions, as the difference in compute resources makes an unfair comparison." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 564, + 416, + 577 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 564, + 416, + 577 + ], + "spans": [ + { + "bbox": [ + 313, + 564, + 416, + 577 + ], + "type": "text", + "content": "5.1. Design Decisions" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 582, + 554, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 554, + 617 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 554, + 617 + ], + "type": "text", + "content": "Certain design decisions were made for the evaluation section. Some due to fairness considerations when comparing to the baselines, and others after testing various alternatives." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 618, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 618, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 618, + 555, + 714 + ], + "type": "text", + "content": "For fairness reasons, it is important to compare methods of similar size. We configured all models to have 4 down-sampling stages, and a depth (number of blocks per stage) of 2 for each scale. Tab. 2 show the parameter and flop count per model. We gave PanoFormer [36], EGFormer [48], Elite360D [1] and HerRUNet[49] base embedding dim of 32, SFSS [20] 48, HealSWIN [8] 64. The resnet18 backbone was used for Elite360D. This resulted in as equal as" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 748, + 313, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 748, + 313, + 758 + ], + "spans": [ + { + "bbox": [ + 297, + 748, + 313, + 758 + ], + "type": "text", + "content": "944" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 57, + 70, + 294, + 232 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 294, + 232 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 294, + 232 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 294, + 232 + ], + "type": "table", + "html": "
ModelRes. (#Pixels)ParamsFlops
PanoFormer [36]256×512(131K)14.5M11.8G
EGFormer [48]256×512(131K)15.2M15.6G
SFSS [20]256×512(131K)15.1M18.9G
Elite360D [1]256×512†(136K)14.7M13.6G
HealSWIN [8]12·47(196K)12.0M39.0G
HexRUnet [49]10·47+2(164K)14.0M12.4G
PanoFormer [36]512×1024(524K)14.5M44.7G
EGFormer [48]512×1024(524K)15.2M65.8G
SFSS [20]512×1024(524K)15.1M80.3G
Elite360D [1]512×1024†(529K)14.7M51.4G
HealSWIN [8]12·48(786K)12.0M156.1G
", + "image_path": "7916fd44978b01b6fc6abb7a7dd590c00bc5488937e8192b7ba2d1aab6efc49e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 59, + 274, + 291, + 383 + ], + "blocks": [ + { + "bbox": [ + 55, + 240, + 294, + 263 + ], + "lines": [ + { + "bbox": [ + 55, + 240, + 294, + 263 + ], + "spans": [ + { + "bbox": [ + 55, + 240, + 294, + 263 + ], + "type": "text", + "content": "Table 2. Resolution and parameter count of baseline models. " + }, + { + "bbox": [ + 55, + 240, + 294, + 263 + ], + "type": "inline_equation", + "content": "{}^{ \\dagger }" + }, + { + "bbox": [ + 55, + 240, + 294, + 263 + ], + "type": "text", + "content": " Elite360D uses additional icosphere of rank 4 (5120 data points)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 59, + 274, + 291, + 383 + ], + "lines": [ + { + "bbox": [ + 59, + 274, + 291, + 383 + ], + "spans": [ + { + "bbox": [ + 59, + 274, + 291, + 383 + ], + "type": "table", + "html": "
#RankTypeCheadCwinRes.ParamsFlops
6hex1141K11.2M2.5G
6ico1182K11.2M4.9G
7hex11164K11.2M9.9G
7ico11328K11.2M19.1G
7hex21164K14.9M13.0G
(1)7hex22164K14.9M13.1G
(2)8hex22655K14.9M52.7G
", + "image_path": "3114f51dfeab11e30dd10e1e9b65ef5e0da7b17688df7d978e8fce6acffac695.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 392, + 295, + 415 + ], + "lines": [ + { + "bbox": [ + 55, + 392, + 295, + 415 + ], + "spans": [ + { + "bbox": [ + 55, + 392, + 295, + 415 + ], + "type": "text", + "content": "Table 3. Resolution (total data points) and parameter count of various configurations of our model. Selected configurations in green." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 437, + 295, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 437, + 295, + 581 + ], + "spans": [ + { + "bbox": [ + 54, + 437, + 295, + 581 + ], + "type": "text", + "content": "possible resolution, parameter count, and flops that the different architectures allow. We analyzed input resolutions of " + }, + { + "bbox": [ + 54, + 437, + 295, + 581 + ], + "type": "inline_equation", + "content": "256 \\times 512" + }, + { + "bbox": [ + 54, + 437, + 295, + 581 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 437, + 295, + 581 + ], + "type": "inline_equation", + "content": "512 \\times 1024" + }, + { + "bbox": [ + 54, + 437, + 295, + 581 + ], + "type": "text", + "content": " for the grid models. Resolution of the spherical models was chosen to be as close as possible to those pixel counts. This was with subdivision ranks 7 and 8 respectively. Also for fairness, none of the methods used pretrained weights. To evaluate all models fairly and on equal terms, the predictions in the various representations were projected uniformly to the sphere surface. Unlike evaluating directly on ERP, this provides an evaluation that evenly averages over the sphere surface and is not biased to extreme vertical angles." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 582, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 582, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 582, + 295, + 714 + ], + "type": "text", + "content": "We followed these guidelines to select the sphere setting with the closest number of nodes, total parameters and flops to the baselines. These stats are reported in Tab. 2 for the existing methods, while Tab. 3 lists variations of our method concerning: sphere rank, node type, head, and window coefficients " + }, + { + "bbox": [ + 54, + 582, + 295, + 714 + ], + "type": "inline_equation", + "content": "(C_{head}, C_{win})" + }, + { + "bbox": [ + 54, + 582, + 295, + 714 + ], + "type": "text", + "content": ". As can be seen, spheres with a rank of 6 have a very low resolution, regardless of the node type. For spheres with rank 7, we have found that the \"vertex\" node type (with " + }, + { + "bbox": [ + 54, + 582, + 295, + 714 + ], + "type": "inline_equation", + "content": "164\\mathrm{K}" + }, + { + "bbox": [ + 54, + 582, + 295, + 714 + ], + "type": "text", + "content": " nodes) provides a resolution that is closest to the " + }, + { + "bbox": [ + 54, + 582, + 295, + 714 + ], + "type": "inline_equation", + "content": "256\\times 512" + }, + { + "bbox": [ + 54, + 582, + 295, + 714 + ], + "type": "text", + "content": " (total of " + }, + { + "bbox": [ + 54, + 582, + 295, + 714 + ], + "type": "inline_equation", + "content": "131\\mathrm{K}" + }, + { + "bbox": [ + 54, + 582, + 295, + 714 + ], + "type": "text", + "content": " pixels) resolution, and rank 8 is the closest to " + }, + { + "bbox": [ + 54, + 582, + 295, + 714 + ], + "type": "inline_equation", + "content": "512\\times 1024" + }, + { + "bbox": [ + 54, + 582, + 295, + 714 + ], + "type": "text", + "content": ". The base configura" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "type": "text", + "content": "tion of our method has fewer parameters and flops than the baselines. This is largely because, unlike most baselines, our method does not contain spatial layers, such as convolution kernels, outside of the attention layer. As a result, the base configuration of our method is less powerful. This is also further addressed in the ablation section (Sec. 5.4). To minimize this difference, we have found that increasing the head coefficient results in a model with slightly more parameters to match the minimum of the baselines. Increasing the window coefficient does not add any parameters and an insignificant amount of flops. Finally, we have selected two configurations for comparison with the baselines. One with rank 7 to be compared to " + }, + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "type": "inline_equation", + "content": "256 \\times 512" + }, + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "type": "text", + "content": " models, and one in rank 8 for larger resolution. These are highlighted in Tab. 3." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 249, + 419, + 262 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 249, + 419, + 262 + ], + "spans": [ + { + "bbox": [ + 313, + 249, + 419, + 262 + ], + "type": "text", + "content": "5.2. Depth Estimation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 267, + 555, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 267, + 555, + 531 + ], + "spans": [ + { + "bbox": [ + 313, + 267, + 555, + 531 + ], + "type": "text", + "content": "In depth estimation, the task is a dense prediction of the pixels depths. We evaluated the depth prediction up to 10 meters for Stanford2D3D and 5 meters for Structured3D. Following PanoFormer [36], we used the BerhuLoss [25] as a differentiable criterion to train the models. We then measured the Mean Absolute Error (MAE), Mean Relative Error (MRE), Root Mean Squared Error (RMSE), and " + }, + { + "bbox": [ + 313, + 267, + 555, + 531 + ], + "type": "inline_equation", + "content": "\\delta_{1}" + }, + { + "bbox": [ + 313, + 267, + 555, + 531 + ], + "type": "text", + "content": " accuracy, which are standard evaluation metrics for depth estimation. Quantitative results on " + }, + { + "bbox": [ + 313, + 267, + 555, + 531 + ], + "type": "inline_equation", + "content": "256\\times 512" + }, + { + "bbox": [ + 313, + 267, + 555, + 531 + ], + "type": "text", + "content": " are reported in Tab. 4. Our method is shown to outperform the baselines on all but one metric, where it was second. In evaluation on Stanford2D3D in " + }, + { + "bbox": [ + 313, + 267, + 555, + 531 + ], + "type": "inline_equation", + "content": "512\\times 1024" + }, + { + "bbox": [ + 313, + 267, + 555, + 531 + ], + "type": "text", + "content": " (Tab. 5), the gap is even larger, and our method significantly improved while the baselines did not. Fig. 7 presents sample results. Evidently, our results are considerably sharper than the baselines, especially around the center, and handle distortions around the poles better. We argue that this is due to spherical representation having a better effective resolution at the center, and does not suffer from distortions. We have also found depth predictions of most baselines to have a boundary misalignment effect where the " + }, + { + "bbox": [ + 313, + 267, + 555, + 531 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 313, + 267, + 555, + 531 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 267, + 555, + 531 + ], + "type": "inline_equation", + "content": "0^{\\circ}" + }, + { + "bbox": [ + 313, + 267, + 555, + 531 + ], + "type": "text", + "content": " meet, which our method has no such effect. This is further explored in the supplement." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 540, + 447, + 553 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 540, + 447, + 553 + ], + "spans": [ + { + "bbox": [ + 313, + 540, + 447, + 553 + ], + "type": "text", + "content": "5.3. Semantic Segmentation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 558, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 556, + 713 + ], + "type": "text", + "content": "In semantic segmentation, the task is to predict a semantic class per image pixel. Stanford2D3D has 13 categories, while Structured3D has 40. We used the standard Categorical Cross Entropy loss, where the background category was ignored. For evaluation, we measured the global pixel accuracy and the mIoU percentage, which are standard metrics for semantic segmentation. Results are depicted in Tab. 4. Tab. 5 shows additional high-resolution setup. In this task as well, our method outperformed the baselines on all metrics. Fig. 8 presents typical samples. Similarly, our semantic segmentation results present the same improvement in terms of distortion handling and improvement around the center of the image we have observed for depth estimation." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 748, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 748, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 748, + 312, + 757 + ], + "type": "text", + "content": "945" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 70, + 553, + 218 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 553, + 218 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 553, + 218 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 553, + 218 + ], + "type": "table", + "html": "
ModelDepth EstimationSemantic Segmentation
Stanford2D3DStructured3DStanford2D3DStructured3D
MAE↓MRE↓RMSE↓δ1↑MAE↓MRE↓RMSE↓δ1↑Acc.↑mIoU↑Acc.↑mIoU↑
PanoFormer [36].174.078.45192.5.154.051.33094.883.160.694.949.7
EGFormer [48].170.075.45693.1.150.049.32895.286.566.495.051.5
SFSS [20].179.081.45592.2.155.051.33095.086.968.295.251.9
HexRUnet [49].201.090.51190.1----81.756.1--
HealSWIN [8].189.084.48292.2----85.563.2--
Elite360D [1].169.069.44893.5.147.046.32195.987.471.495.352.0
OURS (Tab. 3 (1)).165.071.43294.0.142.045.31496.488.672.295.853.0
", + "image_path": "883fec8c51991e9240ced3066678396503e48a9cfdbc5df8145ad7e3fced6395.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 70, + 248, + 138, + 316 + ], + "blocks": [ + { + "bbox": [ + 70, + 248, + 138, + 316 + ], + "lines": [ + { + "bbox": [ + 70, + 248, + 138, + 316 + ], + "spans": [ + { + "bbox": [ + 70, + 248, + 138, + 316 + ], + "type": "image", + "image_path": "48a9c203806e696fc9c424112604ae65f6c90c832957c5f96adfd56ba38c2663.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 70, + 323, + 138, + 390 + ], + "blocks": [ + { + "bbox": [ + 70, + 323, + 138, + 390 + ], + "lines": [ + { + "bbox": [ + 70, + 323, + 138, + 390 + ], + "spans": [ + { + "bbox": [ + 70, + 323, + 138, + 390 + ], + "type": "image", + "image_path": "f61dcfec21972ec5f190feb2b822142c808ace57fa2babdd1d961c3767588707.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 94, + 393, + 115, + 402 + ], + "lines": [ + { + "bbox": [ + 94, + 393, + 115, + 402 + ], + "spans": [ + { + "bbox": [ + 94, + 393, + 115, + 402 + ], + "type": "text", + "content": "RGB" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 151, + 249, + 219, + 316 + ], + "blocks": [ + { + "bbox": [ + 151, + 249, + 219, + 316 + ], + "lines": [ + { + "bbox": [ + 151, + 249, + 219, + 316 + ], + "spans": [ + { + "bbox": [ + 151, + 249, + 219, + 316 + ], + "type": "image", + "image_path": "61d52f096ca5f453ef288d7fd3c02e9da02d25f290fe055639f3e829c045e66d.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 151, + 323, + 219, + 390 + ], + "blocks": [ + { + "bbox": [ + 151, + 323, + 219, + 390 + ], + "lines": [ + { + "bbox": [ + 151, + 323, + 219, + 390 + ], + "spans": [ + { + "bbox": [ + 151, + 323, + 219, + 390 + ], + "type": "image", + "image_path": "81f65050a0b7937899498948208c583664f1317e9d70708b687fbc2a35b0f745.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 160, + 393, + 210, + 403 + ], + "lines": [ + { + "bbox": [ + 160, + 393, + 210, + 403 + ], + "spans": [ + { + "bbox": [ + 160, + 393, + 210, + 403 + ], + "type": "text", + "content": "PanoFormer" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 231, + 249, + 299, + 316 + ], + "blocks": [ + { + "bbox": [ + 231, + 249, + 299, + 316 + ], + "lines": [ + { + "bbox": [ + 231, + 249, + 299, + 316 + ], + "spans": [ + { + "bbox": [ + 231, + 249, + 299, + 316 + ], + "type": "image", + "image_path": "146a8a60d2f4f1ebcb3efa6f903b6c3077998c92f54e62c34cee9be6f32d0205.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 231, + 323, + 299, + 390 + ], + "blocks": [ + { + "bbox": [ + 231, + 323, + 299, + 390 + ], + "lines": [ + { + "bbox": [ + 231, + 323, + 299, + 390 + ], + "spans": [ + { + "bbox": [ + 231, + 323, + 299, + 390 + ], + "type": "image", + "image_path": "69caf7a2d787868cc97039069a98990ddb52e3a7b9c06ecfd71af3c664b8f32e.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 244, + 393, + 287, + 402 + ], + "lines": [ + { + "bbox": [ + 244, + 393, + 287, + 402 + ], + "spans": [ + { + "bbox": [ + 244, + 393, + 287, + 402 + ], + "type": "text", + "content": "EGFormer" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 312, + 249, + 379, + 316 + ], + "blocks": [ + { + "bbox": [ + 312, + 249, + 379, + 316 + ], + "lines": [ + { + "bbox": [ + 312, + 249, + 379, + 316 + ], + "spans": [ + { + "bbox": [ + 312, + 249, + 379, + 316 + ], + "type": "image", + "image_path": "e98bf00c09c4957bf43d0a7396a607a9d7bd2c326845d8f7e6341db9001bb2f0.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 312, + 323, + 379, + 390 + ], + "blocks": [ + { + "bbox": [ + 312, + 323, + 379, + 390 + ], + "lines": [ + { + "bbox": [ + 312, + 323, + 379, + 390 + ], + "spans": [ + { + "bbox": [ + 312, + 323, + 379, + 390 + ], + "type": "image", + "image_path": "6d2068fa0fa4383216e986cf4e1b3c60fe0b5bc29963157b73bd353092e89c11.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 324, + 393, + 366, + 402 + ], + "lines": [ + { + "bbox": [ + 324, + 393, + 366, + 402 + ], + "spans": [ + { + "bbox": [ + 324, + 393, + 366, + 402 + ], + "type": "text", + "content": "Elite360D" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 392, + 249, + 460, + 316 + ], + "blocks": [ + { + "bbox": [ + 392, + 249, + 460, + 316 + ], + "lines": [ + { + "bbox": [ + 392, + 249, + 460, + 316 + ], + "spans": [ + { + "bbox": [ + 392, + 249, + 460, + 316 + ], + "type": "image", + "image_path": "3a226aebdd51df120d90aedb810585abb9cb99e792bbd4f315ab59c2eff06e2d.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 392, + 323, + 460, + 390 + ], + "blocks": [ + { + "bbox": [ + 392, + 323, + 460, + 390 + ], + "lines": [ + { + "bbox": [ + 392, + 323, + 460, + 390 + ], + "spans": [ + { + "bbox": [ + 392, + 323, + 460, + 390 + ], + "type": "image", + "image_path": "847558752fbe02aa5ca306c897621a01790fb993b22ad284ff9af5fbf589ee70.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 415, + 393, + 435, + 402 + ], + "lines": [ + { + "bbox": [ + 415, + 393, + 435, + 402 + ], + "spans": [ + { + "bbox": [ + 415, + 393, + 435, + 402 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 169, + 414, + 440, + 426 + ], + "lines": [ + { + "bbox": [ + 169, + 414, + 440, + 426 + ], + "spans": [ + { + "bbox": [ + 169, + 414, + 440, + 426 + ], + "type": "text", + "content": "Figure 7. Depth Estimation. Top: Stanford2D3D. Bottom: Structured3D." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 472, + 249, + 539, + 316 + ], + "blocks": [ + { + "bbox": [ + 472, + 249, + 539, + 316 + ], + "lines": [ + { + "bbox": [ + 472, + 249, + 539, + 316 + ], + "spans": [ + { + "bbox": [ + 472, + 249, + 539, + 316 + ], + "type": "image", + "image_path": "d2b31beac34d5041bc4bb3f6eacea12e739504f4adc94db6f1279b80ad738d1b.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 472, + 323, + 539, + 390 + ], + "blocks": [ + { + "bbox": [ + 472, + 323, + 539, + 390 + ], + "lines": [ + { + "bbox": [ + 472, + 323, + 539, + 390 + ], + "spans": [ + { + "bbox": [ + 472, + 323, + 539, + 390 + ], + "type": "image", + "image_path": "460b1984ab08eb1df6c2e89296086de7c91eb23bfd9651bd13d245dd30ebbfec.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 499, + 393, + 513, + 402 + ], + "lines": [ + { + "bbox": [ + 499, + 393, + 513, + 402 + ], + "spans": [ + { + "bbox": [ + 499, + 393, + 513, + 402 + ], + "type": "text", + "content": "GT" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "table", + "bbox": [ + 63, + 445, + 288, + 515 + ], + "blocks": [ + { + "bbox": [ + 171, + 227, + 440, + 239 + ], + "lines": [ + { + "bbox": [ + 171, + 227, + 440, + 239 + ], + "spans": [ + { + "bbox": [ + 171, + 227, + 440, + 239 + ], + "type": "text", + "content": "Table 4. Quantitative comparison on sphere rank 7 " + }, + { + "bbox": [ + 171, + 227, + 440, + 239 + ], + "type": "inline_equation", + "content": "(256\\times 512)" + }, + { + "bbox": [ + 171, + 227, + 440, + 239 + ], + "type": "text", + "content": " resolution." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 63, + 445, + 288, + 515 + ], + "lines": [ + { + "bbox": [ + 63, + 445, + 288, + 515 + ], + "spans": [ + { + "bbox": [ + 63, + 445, + 288, + 515 + ], + "type": "table", + "html": "
ModelMAE↓MRE↓δ1↑Acc.↑IoU↑
PanoFormer.167.07293.782.455.6
Elite360D.181.07793.284.563.3
OURS (Tab. 3 (2)).147.06594.089.171.5
", + "image_path": "9e512472d86e9f4e85b04c6541776abf3951df4a909873003a9686e1dcb3d024.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "table_body" + } + ], + "index": 21 + }, + { + "bbox": [ + 61, + 522, + 289, + 534 + ], + "lines": [ + { + "bbox": [ + 61, + 522, + 289, + 534 + ], + "spans": [ + { + "bbox": [ + 61, + 522, + 289, + 534 + ], + "type": "text", + "content": "Table 5. Comparison on Stanford2D3D in rank 8 (512 × 1024)." + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 556, + 149, + 569 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 556, + 149, + 569 + ], + "spans": [ + { + "bbox": [ + 55, + 556, + 149, + 569 + ], + "type": "text", + "content": "5.4. Ablation Study" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 55, + 574, + 296, + 623 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 574, + 296, + 623 + ], + "spans": [ + { + "bbox": [ + 55, + 574, + 296, + 623 + ], + "type": "text", + "content": "To explore the contribution of each one of the elements we have conducted several ablation studies in which multiple variants of our method are tested. These tests were conducted in the context of depth estimation on Stanford2D3D." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": "Positional Encoding Two main components of the method are the global absolute and relative bias position encodings. The vertical global position encoding injects vertical position while remaining rotation equivariant horizontally. The relative bias within the self-attention provides relative information. We compared various configurations" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 446, + 555, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 446, + 555, + 506 + ], + "spans": [ + { + "bbox": [ + 313, + 446, + 555, + 506 + ], + "type": "text", + "content": "that omit one or both of the encodings. Results can be seen in Tab. 6. Evidently, the absence of positional encoding, either global or relative, drastically impacts the performance of the model. Also, relative positional bias proved more critical than the global one." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "type": "text", + "content": "Scaling Coefficients Two key modifiable scaling coefficients, the \"window\" " + }, + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "type": "inline_equation", + "content": "(C_{win})" + }, + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "type": "text", + "content": " and \"head\" " + }, + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "type": "inline_equation", + "content": "(C_{head})" + }, + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "type": "text", + "content": ", affect the capacity of the attention module. The former affects the size of the attention window to incorporate further nodes in the graph. The latter increases the capacity of the attention head by increasing the head dimension. We tested the effect of these coefficients by gradually increasing them separately and together. As shown in Tab. 7. Increasing the window up until " + }, + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "type": "inline_equation", + "content": "C_{win} = 2" + }, + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "type": "text", + "content": " improves the performance. Interestingly, further increase of the window did not show an improvement. This could suggest that such large windows interfere with convergence. As expected, increasing the attention head also improves results, even up to " + }, + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "type": "inline_equation", + "content": "C_{head} = 4" + }, + { + "bbox": [ + 313, + 521, + 555, + 713 + ], + "type": "text", + "content": ". Unsurprisingly, this suggests that the model could benefit from an overall increase in parameters, which is beyond the scope of this work." + } + ] + } + ], + "index": 27 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "text", + "content": "946" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 70, + 71, + 138, + 140 + ], + "blocks": [ + { + "bbox": [ + 70, + 71, + 138, + 140 + ], + "lines": [ + { + "bbox": [ + 70, + 71, + 138, + 140 + ], + "spans": [ + { + "bbox": [ + 70, + 71, + 138, + 140 + ], + "type": "image", + "image_path": "2829777ed08c2ede78485f2504f1e76af3780e3965f982cc61abaaab25f7d3fd.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 70, + 146, + 138, + 213 + ], + "blocks": [ + { + "bbox": [ + 70, + 146, + 138, + 213 + ], + "lines": [ + { + "bbox": [ + 70, + 146, + 138, + 213 + ], + "spans": [ + { + "bbox": [ + 70, + 146, + 138, + 213 + ], + "type": "image", + "image_path": "fb935d7211132be200d086b06a041d3db4c6b9b4c9ee1a05288287a065575d9d.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 94, + 216, + 115, + 225 + ], + "lines": [ + { + "bbox": [ + 94, + 216, + 115, + 225 + ], + "spans": [ + { + "bbox": [ + 94, + 216, + 115, + 225 + ], + "type": "text", + "content": "RGB" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 151, + 72, + 219, + 140 + ], + "blocks": [ + { + "bbox": [ + 151, + 72, + 219, + 140 + ], + "lines": [ + { + "bbox": [ + 151, + 72, + 219, + 140 + ], + "spans": [ + { + "bbox": [ + 151, + 72, + 219, + 140 + ], + "type": "image", + "image_path": "63b7beb973642fd6ec89d2dafe25684b125dc5cf9dc1f78cdf2e6156edc59bca.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 151, + 146, + 219, + 213 + ], + "blocks": [ + { + "bbox": [ + 151, + 146, + 219, + 213 + ], + "lines": [ + { + "bbox": [ + 151, + 146, + 219, + 213 + ], + "spans": [ + { + "bbox": [ + 151, + 146, + 219, + 213 + ], + "type": "image", + "image_path": "8717697ffd72d63f5b9f354509a58f7c589d06fb860685d8a87af3fac381f338.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 159, + 216, + 211, + 226 + ], + "lines": [ + { + "bbox": [ + 159, + 216, + 211, + 226 + ], + "spans": [ + { + "bbox": [ + 159, + 216, + 211, + 226 + ], + "type": "text", + "content": "PanoFormer" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 230, + 72, + 299, + 140 + ], + "blocks": [ + { + "bbox": [ + 230, + 72, + 299, + 140 + ], + "lines": [ + { + "bbox": [ + 230, + 72, + 299, + 140 + ], + "spans": [ + { + "bbox": [ + 230, + 72, + 299, + 140 + ], + "type": "image", + "image_path": "4854a2a5d9f9979fbcbe044df53129d2313bb102c99e509afa1ba0b5da6fa22b.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 231, + 146, + 299, + 213 + ], + "blocks": [ + { + "bbox": [ + 231, + 146, + 299, + 213 + ], + "lines": [ + { + "bbox": [ + 231, + 146, + 299, + 213 + ], + "spans": [ + { + "bbox": [ + 231, + 146, + 299, + 213 + ], + "type": "image", + "image_path": "e2775a91bebf7c1381a5d8971f199d0e10688445fe225cb4e71c862d58bd80dc.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 243, + 216, + 287, + 226 + ], + "lines": [ + { + "bbox": [ + 243, + 216, + 287, + 226 + ], + "spans": [ + { + "bbox": [ + 243, + 216, + 287, + 226 + ], + "type": "text", + "content": "EGFormer" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 312, + 72, + 380, + 140 + ], + "blocks": [ + { + "bbox": [ + 312, + 72, + 380, + 140 + ], + "lines": [ + { + "bbox": [ + 312, + 72, + 380, + 140 + ], + "spans": [ + { + "bbox": [ + 312, + 72, + 380, + 140 + ], + "type": "image", + "image_path": "ae84e928b8b10506b2f528ffde90bc6d55c515d033fa25673b607b823cd9f552.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 312, + 146, + 379, + 213 + ], + "blocks": [ + { + "bbox": [ + 312, + 146, + 379, + 213 + ], + "lines": [ + { + "bbox": [ + 312, + 146, + 379, + 213 + ], + "spans": [ + { + "bbox": [ + 312, + 146, + 379, + 213 + ], + "type": "image", + "image_path": "666c629d97f2b11d794caaf9524bd78e065150812f4275d6581a60339fb03807.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 324, + 216, + 367, + 226 + ], + "lines": [ + { + "bbox": [ + 324, + 216, + 367, + 226 + ], + "spans": [ + { + "bbox": [ + 324, + 216, + 367, + 226 + ], + "type": "text", + "content": "Elite360D" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 158, + 237, + 451, + 249 + ], + "lines": [ + { + "bbox": [ + 158, + 237, + 451, + 249 + ], + "spans": [ + { + "bbox": [ + 158, + 237, + 451, + 249 + ], + "type": "text", + "content": "Figure 8. Semantic Segmentation. Top: Stanford2D3D. Bottom: Structured3D." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 392, + 72, + 460, + 140 + ], + "blocks": [ + { + "bbox": [ + 392, + 72, + 460, + 140 + ], + "lines": [ + { + "bbox": [ + 392, + 72, + 460, + 140 + ], + "spans": [ + { + "bbox": [ + 392, + 72, + 460, + 140 + ], + "type": "image", + "image_path": "8a33613f0f8ae52c2b5aeed34a043d5a4e627d4d45821277d80eda194492987e.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 392, + 146, + 460, + 213 + ], + "blocks": [ + { + "bbox": [ + 392, + 146, + 460, + 213 + ], + "lines": [ + { + "bbox": [ + 392, + 146, + 460, + 213 + ], + "spans": [ + { + "bbox": [ + 392, + 146, + 460, + 213 + ], + "type": "image", + "image_path": "cd16e1dfb09ad36aa9b6d8d8ce4afc9cb9aadf0530096c75209a8e27e9104d02.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 414, + 216, + 435, + 226 + ], + "lines": [ + { + "bbox": [ + 414, + 216, + 435, + 226 + ], + "spans": [ + { + "bbox": [ + 414, + 216, + 435, + 226 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 473, + 72, + 540, + 140 + ], + "blocks": [ + { + "bbox": [ + 473, + 72, + 540, + 140 + ], + "lines": [ + { + "bbox": [ + 473, + 72, + 540, + 140 + ], + "spans": [ + { + "bbox": [ + 473, + 72, + 540, + 140 + ], + "type": "image", + "image_path": "ccfda3f569d92a0c3df6dc14594accc46e4d6ba66ac60ef025ea2621ca0b8868.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 472, + 146, + 540, + 213 + ], + "blocks": [ + { + "bbox": [ + 472, + 146, + 540, + 213 + ], + "lines": [ + { + "bbox": [ + 472, + 146, + 540, + 213 + ], + "spans": [ + { + "bbox": [ + 472, + 146, + 540, + 213 + ], + "type": "image", + "image_path": "5fc1858bd06e5119cb49e5558d2629c46e8789f140b0b339ad742a5ebb119abc.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 498, + 216, + 513, + 225 + ], + "lines": [ + { + "bbox": [ + 498, + 216, + 513, + 225 + ], + "spans": [ + { + "bbox": [ + 498, + 216, + 513, + 225 + ], + "type": "text", + "content": "GT" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "table", + "bbox": [ + 56, + 268, + 295, + 343 + ], + "blocks": [ + { + "bbox": [ + 56, + 268, + 295, + 343 + ], + "lines": [ + { + "bbox": [ + 56, + 268, + 295, + 343 + ], + "spans": [ + { + "bbox": [ + 56, + 268, + 295, + 343 + ], + "type": "table", + "html": "
VariantAbsPosRelPosMAE↓MRE↓
No Pos. Enc.XX.326.094
No Rel. Pos. Enc.✓X.251.091
No Abs. Pos. Enc.X✓.218.088
With Pos. Enc.✓✓.189.077
", + "image_path": "039e914ee827a47782895981781104fd32959861557d30a902452daf4e72ed89.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "table_body" + } + ], + "index": 19 + }, + { + "type": "table", + "bbox": [ + 56, + 374, + 294, + 486 + ], + "blocks": [ + { + "bbox": [ + 88, + 351, + 261, + 363 + ], + "lines": [ + { + "bbox": [ + 88, + 351, + 261, + 363 + ], + "spans": [ + { + "bbox": [ + 88, + 351, + 261, + 363 + ], + "type": "text", + "content": "Table 6. Ablation Study: Positional Encoding." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 374, + 294, + 486 + ], + "lines": [ + { + "bbox": [ + 56, + 374, + 294, + 486 + ], + "spans": [ + { + "bbox": [ + 56, + 374, + 294, + 486 + ], + "type": "table", + "html": "
VariantCheadCwinMAE↓MRE↓
Base variant11.189.077
No window10.412.122
Medium window12.175.073
Large window13.180.078
Medium head21.171.075
Large head41.162.067
Conf. (1) Tab. 322.165.071
", + "image_path": "ab7a8df83697c709ac52db051e86e3d0131a0ec22019d0cbdcc73f62b55c4047.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "table_body" + } + ], + "index": 21 + }, + { + "bbox": [ + 69, + 494, + 281, + 505 + ], + "lines": [ + { + "bbox": [ + 69, + 494, + 281, + 505 + ], + "spans": [ + { + "bbox": [ + 69, + 494, + 281, + 505 + ], + "type": "text", + "content": "Table 7. Ablation Study: Head and Window Coefficients." + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 525, + 209, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 525, + 209, + 538 + ], + "spans": [ + { + "bbox": [ + 55, + 525, + 209, + 538 + ], + "type": "text", + "content": "6. Discussion and Limitations" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 55, + 545, + 295, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 295, + 629 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 295, + 629 + ], + "type": "text", + "content": "This paper introduces a pioneering approach to omnidirectional " + }, + { + "bbox": [ + 55, + 545, + 295, + 629 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 55, + 545, + 295, + 629 + ], + "type": "text", + "content": " perception, leveraging a geodesic polyhedron and a unique architecture designed for spherical computation. However, there is ample room for optimization and enhancement of its features. The potential for improvement spans various aspects of the model, from its core architecture to its computational efficiency." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "text", + "content": "One of the promising avenues for future work lies in exploring additional domains where this architecture could be applied. Beyond the realms of depth estimation and semantic segmentation in " + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "text", + "content": " imagery, spherical data representation is also prominent in other areas such as medical imaging, where spherical structures are common, and geographical data analysis, which often deals with global-scale data" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 270, + 553, + 306 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 270, + 553, + 306 + ], + "spans": [ + { + "bbox": [ + 313, + 270, + 553, + 306 + ], + "type": "text", + "content": "that naturally fit a spherical model. The adaptability of our approach to these varied domains could pave the way for groundbreaking applications and methodologies." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 307, + 554, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 307, + 554, + 415 + ], + "spans": [ + { + "bbox": [ + 313, + 307, + 554, + 415 + ], + "type": "text", + "content": "While the current work has focused on supervised classification and regression tasks, extending this model to generative models presents an exciting frontier. Generative models could leverage the spherical data representation for creating highly realistic and comprehensive 3D environments among other applications. The unique challenges of generative tasks on spherical domains, such as ensuring continuity and avoiding distortions, call for accurate modeling of the type we present here." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 313, + 415, + 554, + 535 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 415, + 554, + 535 + ], + "spans": [ + { + "bbox": [ + 313, + 415, + 554, + 535 + ], + "type": "text", + "content": "Another area for improvement is the computational efficiency of the model, particularly regarding its GPU usage. Despite the number of parameters and flops being on par with baseline models, our approach exhibits slower performance (" + }, + { + "bbox": [ + 313, + 415, + 554, + 535 + ], + "type": "inline_equation", + "content": "\\sim 30\\%" + }, + { + "bbox": [ + 313, + 415, + 554, + 535 + ], + "type": "text", + "content": " slower than PanoFormer). This is largely because the gathering operations on the sphere graph are not optimized with dedicated CUDA functions in PyTorch, leading to less efficient execution. Addressing this engineering challenge could significantly enhance the model's appeal by reducing both the GPU footprint and runtime." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 313, + 548, + 388, + 560 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 548, + 388, + 560 + ], + "spans": [ + { + "bbox": [ + 313, + 548, + 388, + 560 + ], + "type": "text", + "content": "7. Conclusion" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 313, + 568, + 554, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 568, + 554, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 568, + 554, + 653 + ], + "type": "text", + "content": "We have presented a novel method for omnidirectional perception that successfully utilizes spherically represented data. The method addresses the underlying spherical structure, which comes up in three different places: in the down and upsampling operators, in the positional encoding, and, above all, in the definition geometric representation and that of the neighborhoods for the local attention operator." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "text", + "content": "Our proposed method relies on the icosphere representation and utilizes its favorable properties to derive sphere-based operators. Using these, we surpass the state of the art in omnidirectional perception in both depth estimation and semantic segmentation." + } + ] + } + ], + "index": 31 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 748, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 748, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 748, + 312, + 757 + ], + "type": "text", + "content": "947" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 297, + 128 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 297, + 128 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 297, + 128 + ], + "type": "text", + "content": "This work was supported by the Tel Aviv University Center for AI and Data Science (TAD). The contribution of the first author is part of a PhD thesis at Tel Aviv University." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 140, + 115, + 153 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 140, + 115, + 153 + ], + "spans": [ + { + "bbox": [ + 56, + 140, + 115, + 153 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 160, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 160, + 296, + 214 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 160, + 296, + 214 + ], + "spans": [ + { + "bbox": [ + 61, + 160, + 296, + 214 + ], + "type": "text", + "content": "[1] Hao Ai and Lin Wang. Elite360d: Towards efficient 360 depth estimation via semantic-and distance-aware biprojection fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9926-9935, 2024. 5, 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 217, + 296, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 217, + 296, + 281 + ], + "spans": [ + { + "bbox": [ + 61, + 217, + 296, + 281 + ], + "type": "text", + "content": "[2] Hao Ai, Zidong Cao, Yan-Pei Cao, Ying Shan, and Lin Wang. Hrdfuse: Monocular 360deg depth estimation by collaboratively learning holistic-with-regional depth distributions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13273-13282, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 284, + 296, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 284, + 296, + 360 + ], + "spans": [ + { + "bbox": [ + 61, + 284, + 296, + 360 + ], + "type": "text", + "content": "[3] Georgios Albanis, Nikolaos Zioulis, Petros Drakoulis, Vasileios Gkitsas, Vladimiros Sterzentenko, Federico Alvarez, Dimitrios Zarpalas, and Petros Daras. Pano3d: A holistic benchmark and a solid baseline for " + }, + { + "bbox": [ + 61, + 284, + 296, + 360 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 61, + 284, + 296, + 360 + ], + "type": "text", + "content": " depth estimation. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 3722-3732, 2021. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 362, + 294, + 395 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 362, + 294, + 395 + ], + "spans": [ + { + "bbox": [ + 62, + 362, + 294, + 395 + ], + "type": "text", + "content": "[4] Iro Armeni, Sasha Sax, Amir R Zamir, and Silvio Savarese. Joint 2d-3d-semantic data for indoor scene understanding. arXiv preprint arXiv:1702.01105, 2017. 1, 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 397, + 294, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 397, + 294, + 430 + ], + "spans": [ + { + "bbox": [ + 62, + 397, + 294, + 430 + ], + "type": "text", + "content": "[5] Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Bengio. Neural machine translation by jointly learning to align and translate. arXiv preprint arXiv:1409.0473, 2014. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 432, + 295, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 432, + 295, + 453 + ], + "spans": [ + { + "bbox": [ + 62, + 432, + 295, + 453 + ], + "type": "text", + "content": "[6] Herbert Busemann. The geometry of geodesics. Courier Corporation, 2012. 1, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 455, + 296, + 507 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 455, + 296, + 507 + ], + "spans": [ + { + "bbox": [ + 62, + 455, + 296, + 507 + ], + "type": "text", + "content": "[7] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 2, 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 511, + 296, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 511, + 296, + 576 + ], + "spans": [ + { + "bbox": [ + 62, + 511, + 296, + 576 + ], + "type": "text", + "content": "[8] Oscar Carlsson, Jan E Gerken, Hampus Linander, Heiner Spieß, Fredrik Ohlsson, Christoffer Petersson, and Daniel Persson. Heal-swin: A vision transformer on the sphere. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6067-6077, 2024. 1, 2, 5, 6, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 578, + 296, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 578, + 296, + 633 + ], + "spans": [ + { + "bbox": [ + 62, + 578, + 296, + 633 + ], + "type": "text", + "content": "[9] Angel Chang, Angela Dai, Thomas Funkhouser, Maciej Halber, Matthias Niessner, Manolis Savva, Shuran Song, Andy Zeng, and Yinda Zhang. Matterport3d: Learning from rgbd data in indoor environments. International Conference on 3D Vision (3DV), 2017. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 635, + 294, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 635, + 294, + 678 + ], + "spans": [ + { + "bbox": [ + 57, + 635, + 294, + 678 + ], + "type": "text", + "content": "[10] Bowen Cheng, Alex Schwing, and Alexander Kirillov. Per-pixel classification is not all you need for semantic segmentation. Advances in Neural Information Processing Systems, 34:17864-17875, 2021. 2, 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 680, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 680, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 680, + 294, + 713 + ], + "type": "text", + "content": "[11] Hsien-Tzu Cheng, Chun-Hung Chao, Jin-Dong Dong, Hao-Kai Wen, Tyng-Luh Liu, and Min Sun. Cube padding for weakly-supervised saliency prediction in 360 videos. In Pro" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "ceedings of the IEEE conference on computer vision and pattern recognition, pages 1420-1429, 2018. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 97, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 97, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 97, + 553, + 150 + ], + "type": "text", + "content": "[12] Benjamin Coors, Alexandru Paul Condurache, and Andreas Geiger. Spherenet: Learning spherical representations for detection and classification in omnidirectional images. In Proceedings of the European conference on computer vision (ECCV), pages 518-533, 2018. 1, 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 152, + 553, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 553, + 194 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 553, + 194 + ], + "type": "text", + "content": "[13] Michael Defferrard, Martino Milani, Frédérique Gusset, and Nathanaël Perraudin. DeepSphere: a graph-based spherical CNN. In International Conference on Learning Representations, 2020. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 198, + 553, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 198, + 553, + 240 + ], + "spans": [ + { + "bbox": [ + 316, + 198, + 553, + 240 + ], + "type": "text", + "content": "[14] Aleksandar M Dimitrijevic, Martin Lambers, and Dejan Rancic. Comparison of spherical cube map projections used in planet-sized terrain rendering. Facta Universitatis, Series: Mathematics and Informatics, 31(2):259-297, 2016. 1, 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 243, + 553, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 243, + 553, + 317 + ], + "spans": [ + { + "bbox": [ + 317, + 243, + 553, + 317 + ], + "type": "text", + "content": "[15] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021. 2, 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 320, + 553, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 320, + 553, + 373 + ], + "spans": [ + { + "bbox": [ + 317, + 320, + 553, + 373 + ], + "type": "text", + "content": "[16] Marc Eder, Mykhailo Shvets, John Lim, and Jan-Michael Frahm. Tangent images for mitigating spherical distortion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12426-12434, 2020. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 376, + 553, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 376, + 553, + 409 + ], + "spans": [ + { + "bbox": [ + 316, + 376, + 553, + 409 + ], + "type": "text", + "content": "[17] Lance Flavell. Uv mapping. In *Beginning Blender: Open Source 3D Modeling*, Animation, and Game Design, pages 97-122. Springer, 2010. 1, 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 411, + 553, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 411, + 553, + 431 + ], + "spans": [ + { + "bbox": [ + 316, + 411, + 553, + 431 + ], + "type": "text", + "content": "[18] Michael S Floater. Generalized barycentric coordinates and applications. Acta Numerica, 24:161-214, 2015. 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 434, + 555, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 434, + 555, + 465 + ], + "spans": [ + { + "bbox": [ + 316, + 434, + 555, + 465 + ], + "type": "text", + "content": "[19] Krzysztof M Górski, E Hivon, and BD Wandelt. Analysis issues for large cmb data sets. Proceedings: Evolution of Large Scale Structure-Garching, 1998. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 468, + 553, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 468, + 553, + 520 + ], + "spans": [ + { + "bbox": [ + 317, + 468, + 553, + 520 + ], + "type": "text", + "content": "[20] Suresh Guttikonda and Jason Rambach. Single frame semantic segmentation using multi-modal spherical images. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 3222-3231, 2024. 1, 2, 5, 6, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 523, + 553, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 523, + 553, + 566 + ], + "spans": [ + { + "bbox": [ + 316, + 523, + 553, + 566 + ], + "type": "text", + "content": "[21] Shi-Min Hu, Zheng-Ning Liu, Meng-Hao Guo, Jun-Xiong Cai, Jiahui Huang, Tai-Jiang Mu, and Ralph R Martin. Subdivision-based mesh convolution networks. ACM Transactions on Graphics (TOG), 41(3):1-16, 2022. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 568, + 553, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 568, + 553, + 610 + ], + "spans": [ + { + "bbox": [ + 317, + 568, + 553, + 610 + ], + "type": "text", + "content": "[22] Chiyu Jiang, Jingwei Huang, Karthik Kashinath, Philip Marcus, Matthias Niessner, et al. Spherical cnns on unstructured grids. International Conference on Learning Representations, 2019. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 613, + 553, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 613, + 553, + 655 + ], + "spans": [ + { + "bbox": [ + 316, + 613, + 553, + 655 + ], + "type": "text", + "content": "[23] Hualie Jiang, Zhe Sheng, Siyu Zhu, Zilong Dong, and Rui Huang. Unifuse: Unidirectional fusion for 360 panorama depth estimation. IEEE Robotics and Automation Letters, 6 (2):1519-1526, 2021. 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 658, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 553, + 689 + ], + "type": "text", + "content": "[24] Benjamin Keinert, Matthias Innmann, Michael Sanger, and Marc Stamminger. Spherical fibonacci mapping. ACM Transactions on Graphics (TOG), 34(6):1-7, 2015. 1, 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "text", + "content": "[25] Iro Laina, Christian Rupprecht, Vasileios Belagiannis, Federico Tombari, and Nassir Navab. Deeper depth prediction" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 297, + 749, + 313, + 757 + ], + "type": "text", + "content": "948" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 76, + 72, + 296, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 296, + 106 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 296, + 106 + ], + "type": "text", + "content": "with fully convolutional residual networks. In 2016 Fourth international conference on 3D vision (3DV), pages 239-248. IEEE, 2016. 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 106, + 296, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 106, + 296, + 139 + ], + "spans": [ + { + "bbox": [ + 56, + 106, + 296, + 139 + ], + "type": "text", + "content": "[26] Martin Lambers. Survey of cube mapping methods in interactive computer graphics. The Visual Computer, 36(5): 1043-1051, 2020. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 140, + 295, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 140, + 295, + 195 + ], + "spans": [ + { + "bbox": [ + 56, + 140, + 295, + 195 + ], + "type": "text", + "content": "[27] Yeonkun Lee, Jaeseok Jeong, Jongseob Yun, Wonjune Cho, and Kuk-Jin Yoon. Spherephd: Applying cnns on a spherical polyhedron representation of 360deg images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9181-9189, 2019. 1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 195, + 295, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 195, + 295, + 250 + ], + "spans": [ + { + "bbox": [ + 56, + 195, + 295, + 250 + ], + "type": "text", + "content": "[28] Yuyan Li, Yuliang Guo, Zhixin Yan, Xinyu Huang, Ye Duan, and Liu Ren. Omnifusion: 360 monocular depth estimation via geometry-aware fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2801-2810, 2022. 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 250, + 295, + 305 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 250, + 295, + 305 + ], + "spans": [ + { + "bbox": [ + 56, + 250, + 295, + 305 + ], + "type": "text", + "content": "[29] Thang Luong, Hieu Pham, and Christopher D. Manning. Effective approaches to attention-based neural machine translation. In Proceedings of the 2015 Conference on Empirical Methods in Natural Language Processing. Association for Computational Linguistics, 2015. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 305, + 295, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 305, + 295, + 327 + ], + "spans": [ + { + "bbox": [ + 56, + 305, + 295, + 327 + ], + "type": "text", + "content": "[30] Ronald Miller. An equi-rectangular map projection. Geography, pages 196–201, 1949. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 327, + 295, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 327, + 295, + 371 + ], + "spans": [ + { + "bbox": [ + 56, + 327, + 295, + 371 + ], + "type": "text", + "content": "[31] Hongbin Pei, Bingzhe Wei, Kevin Chen-Chuan Chang, Yu Lei, and Bo Yang. Geom-gcn: Geometric graph convolutional networks. International Conference on Learning Representations, 2020. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 371, + 295, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 371, + 295, + 437 + ], + "spans": [ + { + "bbox": [ + 56, + 371, + 295, + 437 + ], + "type": "text", + "content": "[32] Giovanni Pintore, Marco Agus, Eva Almansa, Jens Schneider, and Enrico Gobbetti. Slicenet: deep dense depth estimation from a single indoor panorama using a slice-based representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11536-11545, 2021. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 437, + 295, + 471 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 437, + 295, + 471 + ], + "spans": [ + { + "bbox": [ + 56, + 437, + 295, + 471 + ], + "type": "text", + "content": "[33] Alec Radford, Karthik Narasimhan, Tim Salimans, Ilya Sutskever, et al. Improving language understanding by generative pre-training. 2018. 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 471, + 296, + 515 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 471, + 296, + 515 + ], + "spans": [ + { + "bbox": [ + 56, + 471, + 296, + 515 + ], + "type": "text", + "content": "[34] René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. In Proceedings of the IEEE/CVF international conference on computer vision, pages 12179-12188, 2021. 2, 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 515, + 295, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 515, + 295, + 568 + ], + "spans": [ + { + "bbox": [ + 56, + 515, + 295, + 568 + ], + "type": "text", + "content": "[35] Manuel Rey-Area, Mingze Yuan, and Christian Richardt. 360monodepth: High-resolution 360deg monocular depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3762-3772, 2022. 1, 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 570, + 295, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 570, + 295, + 624 + ], + "spans": [ + { + "bbox": [ + 56, + 570, + 295, + 624 + ], + "type": "text", + "content": "[36] Zhijie Shen, Chunyu Lin, Kang Liao, Lang Nie, Zishuo Zheng, and Yao Zhao. Panofformer: Panorama transformer for indoor " + }, + { + "bbox": [ + 56, + 570, + 295, + 624 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 56, + 570, + 295, + 624 + ], + "type": "text", + "content": " depth estimation. In European Conference on Computer Vision, pages 195-211. Springer, 2022. 1, 2, 5, 6, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 625, + 295, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 625, + 295, + 658 + ], + "spans": [ + { + "bbox": [ + 56, + 625, + 295, + 658 + ], + "type": "text", + "content": "[37] Yu-Chuan Su and Kristen Grauman. Learning spherical convolution for fast features from 360 imagery. Advances in Neural Information Processing Systems, 30, 2017. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 658, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 295, + 712 + ], + "type": "text", + "content": "[38] Cheng Sun, Min Sun, and Hwann-Tzong Chen. Hohonet: 360 indoor holistic understanding with latent horizontal features. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2573-2582, 2021. 1" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 714 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 118 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 118 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 118 + ], + "type": "text", + "content": "[39] Keisuke Tateno, Nassir Navab, and Federico Tombari. Distortion-aware convolutional filters for dense prediction in panoramic images. In Proceedings of the European Conference on Computer Vision (ECCV), pages 707-722, 2018. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 118, + 553, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 553, + 163 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 553, + 163 + ], + "type": "text", + "content": "[40] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 2, 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 163, + 553, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 553, + 207 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 553, + 207 + ], + "type": "text", + "content": "[41] Petar Velicković, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Liò, and Yoshua Bengio. Graph attention networks. In International Conference on Learning Representations, 2018. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 207, + 553, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 207, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 316, + 207, + 553, + 262 + ], + "type": "text", + "content": "[42] Fu-En Wang, Yu-Hsuan Yeh, Min Sun, Wei-Chen Chiu, and Yi-Hsuan Tsai. Bifuse: Monocular 360 depth estimation via bi-projection fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 462–471, 2020. 1, 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 264, + 553, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 264, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 316, + 264, + 553, + 319 + ], + "type": "text", + "content": "[43] Fu-En Wang, Yu-Hsuan Yeh, Yi-Hsuan Tsai, Wei-Chen Chiu, and Min Sun. Bifuse++: Self-supervised and efficient bi-projection fusion for 360 depth estimation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(5): 5448-5460, 2022. 1, 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 319, + 553, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 319, + 553, + 376 + ], + "spans": [ + { + "bbox": [ + 316, + 319, + 553, + 376 + ], + "type": "text", + "content": "[44] Zhendong Wang, Xiaodong Cun, Jianmin Bao, Wengang Zhou, Jianzhuang Liu, and Houqiang Li. Uformer: A general u-shaped transformer for image restoration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 17683-17693, 2022. 2, 3, 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 376, + 553, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 376, + 553, + 431 + ], + "spans": [ + { + "bbox": [ + 316, + 376, + 553, + 431 + ], + "type": "text", + "content": "[45] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. Advances in Neural Information Processing Systems, 34:12077-12090, 2021. 2, 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 432, + 553, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 432, + 553, + 477 + ], + "spans": [ + { + "bbox": [ + 316, + 432, + 553, + 477 + ], + "type": "text", + "content": "[46] Qingsong Yan, Qiang Wang, Kaiyong Zhao, Bo Li, Xiaowei Chu, and Fei Deng. Sphereddepth: Panorama depth estimation from spherical domain. In 2022 International Conference on 3D Vision (3DV), pages 1-10. IEEE, 2022. 1, 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 477, + 553, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 477, + 553, + 533 + ], + "spans": [ + { + "bbox": [ + 316, + 477, + 553, + 533 + ], + "type": "text", + "content": "[47] Ilwi Yun, Hyuk-Jae Lee, and Chae Eun Rhee. Improving 360 monocular depth estimation via non-local dense prediction transformer and joint supervised and self-supervised learning. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3224-3233, 2022. 1, 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 533, + 553, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 533, + 553, + 589 + ], + "spans": [ + { + "bbox": [ + 316, + 533, + 553, + 589 + ], + "type": "text", + "content": "[48] Ilwi Yun, Chanyong Shin, Hyunku Lee, Hyuk-Jae Lee, and Chae Eun Rhee. Egformer: Equirectangular geometry-biased transformer for 360 depth estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 6101-6112, 2023. 1, 2, 5, 6, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 589, + 553, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 589, + 553, + 644 + ], + "spans": [ + { + "bbox": [ + 316, + 589, + 553, + 644 + ], + "type": "text", + "content": "[49] Chao Zhang, Stephan Liwicki, William Smith, and Roberto Cipolla. Orientation-aware semantic segmentation on icosahedron spheres. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3533-3541, 2019. 1, 5, 6, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 644, + 553, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 644, + 553, + 678 + ], + "spans": [ + { + "bbox": [ + 316, + 644, + 553, + 678 + ], + "type": "text", + "content": "[50] Chao Zhang, Sen He, and Stephan Liwicki. A spherical approach to planar semantic segmentation. In BMVC, 2020. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 678, + 553, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 678, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 316, + 678, + 553, + 714 + ], + "type": "text", + "content": "[51] Jiaming Zhang, Kailun Yang, Chaoxiang Ma, Simon Reiß, Kunyu Peng, and Rainer Stiefelhagen. Bending reality: Distortion-aware transformers for adapting to panoramic se" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 313, + 757 + ], + "type": "text", + "content": "949" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 251 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 75, + 72, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 72, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 75, + 72, + 294, + 106 + ], + "type": "text", + "content": "matic segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16917-16927, 2022. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 294, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 294, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 294, + 140 + ], + "type": "text", + "content": "[52] Si Zhang, Hanghang Tong, Jiejun Xu, and Ross Maciejewski. Graph convolutional networks: a comprehensive review. Computational Social Networks, 6(1):1-23, 2019. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 141, + 294, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 141, + 294, + 194 + ], + "spans": [ + { + "bbox": [ + 56, + 141, + 294, + 194 + ], + "type": "text", + "content": "[53] Jia Zheng, Junfei Zhang, Jing Li, Rui Tang, Shenghua Gao, and Zihan Zhou. Structured3d: A large photo-realistic dataset for structured 3d modeling. In Proceedings of The European Conference on Computer Vision (ECCV), 2020. 1, 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 197, + 294, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 197, + 294, + 251 + ], + "spans": [ + { + "bbox": [ + 56, + 197, + 294, + 251 + ], + "type": "text", + "content": "[54] Nikolaos Zioulis, Antonis Karakottas, Dimitrios Zarpalas, and Petros Daras. Omnidepth: Dense depth estimation for indoors spherical panoramicas. In Proceedings of the European Conference on Computer Vision (ECCV), pages 448-465, 2018. 2" + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "spans": [ + { + "bbox": [ + 298, + 749, + 312, + 757 + ], + "type": "text", + "content": "950" + } + ] + } + ], + "index": 5 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Spherical Manifold Guided Diffusion Model for Panoramic Image Generation/25742454-f26d-401b-a2ab-0a30d46da30a_content_list.json b/2025/Spherical Manifold Guided Diffusion Model for Panoramic Image Generation/25742454-f26d-401b-a2ab-0a30d46da30a_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..39b99f0041c7dd0b47ce3b3fbf49f798f321f6e8 --- /dev/null +++ b/2025/Spherical Manifold Guided Diffusion Model for Panoramic Image Generation/25742454-f26d-401b-a2ab-0a30d46da30a_content_list.json @@ -0,0 +1,1397 @@ +[ + { + "type": "text", + "text": "Spherical Manifold Guided Diffusion Model for Panoramic Image Generation", + "text_level": 1, + "bbox": [ + 106, + 130, + 890, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xiancheng Sun $^{1}$ Mai Xu $^{1}$ Shengxi Li $^{1*}$ Senmao Ma $^{1}$ Xin Deng $^{1}$ Lai Jiang $^{1}$ Gang Shen $^{2}$ $^{1}$ Beihang University $^{2}$ China Tower", + "bbox": [ + 99, + 178, + 901, + 217 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/77f9dff45c20491206b720effb9105058a844c3b356c70c6428895139d28aabd.jpg", + "image_caption": [ + "\"A serene room with a large door leading to a balcony and a window showcasing the garden, decorated in soft pastels\"", + "\"A bright and airy staircase with large windows overlooking a manicured garden, enhancing the natural light that floods the space.\"", + "Figure 1. A subjective comparison of zero-shot text-conditioned panoramic image generation between our SMGD model and the state-of-the-art method PanFusion [53]. Our model demonstrates superior performance in generating images with superior overall natural content, enhanced subjective quality, and better alignment with the text description, while benefiting reduced model parameters." + ], + "image_footnote": [], + "bbox": [ + 127, + 224, + 498, + 486 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/825af3b51b3a7cbcc0e3ca75b623a3ac4263186e086c56eb595e77c0c7ed070d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 498, + 226, + 831, + 484 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 566, + 326, + 580 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Panoramic image essentially acts as a pivotal role in emerging virtual reality and augmented reality scenarios; however, the generation of panoramic images are essentially challenging due to the intrinsic spherical geometry and spherical distortions caused by equirectangular projection (ERP). To address this, we start from the very basics of $S^2$ manifold inherent to panoramic images, and propose a novel spherical manifold convolution (SMConv) on $S^2$ manifold. Based on the SMConv operation, we propose a spherical manifold guided diffusion (SMGD) model for text-conditioned panoramic image generation, which can well accommodate the spherical geometry during generation. We further develop a novel evaluation method by calculating grouped Fréchet inception distance (FID) on cube-map projections, which can well reflect the quality of generated panoramic images, compared to existing methods that randomly crop ERP-distorted content. Experiment results demonstrate that our SMGD model achieves the state-of-the-art generation quality and accuracy, whilst retaining", + "bbox": [ + 89, + 587, + 485, + 875 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "the shortest sampling time in the text-conditioned panoramic image generation task. Codes are publicly available at https://github.com/chronos123/SMGD.", + "bbox": [ + 511, + 568, + 908, + 613 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 630, + 645, + 645 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Being able to be controlled by text descriptions, panoramic image generation with a full $360^{\\circ}$ horizontal and $180^{\\circ}$ vertical field-of-view [44] opens a new avenue for various applications in virtual reality [48], augmented reality [47], autonomous driving [46], etc. Although recent developments in deep probabilistic models including generative adversarial networks (GANs) [14, 20, 34, 57] and diffusion models [10, 32, 33] have been acting as the workhorse in generating photo-realistic images, panoramic image generation still remains challenging due to their intrinsic spherical geometry property and distorted scene structures. In particular, the equirectangular projection (ERP) [50] employed to convert the panoramic content into a planar rectangular format, introduces spherical distortions and breaks the continuity at the left and right edges. Within the ERP format, rectangular panoramic images with an aspect ratio of 1:2 exhibit a", + "bbox": [ + 509, + 657, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author.", + "bbox": [ + 109, + 887, + 236, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "5824", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "non-uniform pixel distribution. In other words, the pixel density is notably sparse near the polar regions and significantly denser around the equatorial region. Consequently, existing state-of-the-art deep probabilistic models designed for generating planar images [31, 34, 49, 51] still yet to generalize well for the panoramic image generation.", + "bbox": [ + 89, + 90, + 480, + 181 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address the panoramic image generation task, several methods proposed to generate rectangular images, including autoregressive generation based on the latent codes [11], stitching latent features [3], and directly autoregressive generation on image patches [24]. To retain the continuity between the left and right boundaries, strategies such as circular convolution [5], circular token mechanisms [2], and exploiting the symmetry of panoramic images [15] have been further proposed. However, the above methods fail to effectively mitigate the spherical distortions, thus motivating the usage of deformable convolution [43] and attention mechanisms [53]. Unfortunately, existing methods still suffer from sub-optimal performance, which lacks well-defined global structure and optimal preservation of spherical geometry, as illustrated in Fig. 1, whilst compensating this with heavy computational complexity.", + "bbox": [ + 89, + 186, + 485, + 428 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we start from the intrinsic property of panoramic format, and propose the spherical manifold guided diffusion (SMGD) model operating on $S^2$ manifold, to generate panoramic images from text descriptions. Since panoramic images can be recognized as pixels distributed on $S^2$ manifold, we propose the spherical manifold convolution (SMConv), to effectively address spherical distortions and preserve spherical geometry in panoramic images. To the best of our knowledge, there exists no spherical convolution designed for the panorama generation task, in which our SMConv fills this void by performing the correct convolution on spherical surfaces. Based on SMConv, we establish spherical manifold encoder (SME) and spherical manifold decoder (SMD) blocks, constituting a spherical manifold U-Net (SMUNet). Our SMGD model is subsequently constructed by employing the SMUNet as the denoising network, with the VQGAN [11] serving as its image encoder. Unlike previous spherical convolutions [8, 36, 37, 45, 55] applied in classification and detection tasks, our SMConv is supported by the intrinsic manifold theory. Guided by exponential mapping, our SMConv incorporates the convolution within spherical neighborhoods under the measure of the geodesic distance [40]. Moreover, to further address the spherical geometry of panoramic images, we develop a spherical loss to optimize the proposed SMGD model. We also adopt an enhanced strategy to evaluate the quality of generated panoramic images by calculating the grouped Fréchet inception distance (FID) [17] in cubemap projection (CMP) [52] format. Experimental results demonstrate that our method achieves superior generation performances by effectively addressing the unique challenges associated with panoramic", + "bbox": [ + 89, + 431, + 485, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "image generation, such as complex scene structure, spherical distortions caused by ERP and the sparse pixel distribution near the poles. In summary, our contributions are three-fold:", + "bbox": [ + 511, + 90, + 906, + 136 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We set out the first attempt to propose the spherical convolution, namely, SMConv, operating on $S^2$ manifold, which is able to optimally capture the intrinsic spherical geometry in panoramic images.", + "- We propose to incorporate SMConv into spherical manifold guided blocks, including the SME and SMD blocks, so as to mitigate spherical distortions and preserve spatial coherence across the spherical domain.", + "- Building upon these foundational blocks, we propose the SMGD model as the first generative approach to incorporate spherical convolution and achieve the state-of-the-art performance in generating panoramic images from text descriptions." + ], + "bbox": [ + 513, + 138, + 906, + 334 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 513, + 345, + 663, + 363 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Panoramic Image Generation. Early research on panoramic image generation can be achieved by the image outpainting task, which generates panoramic images from partial view [1, 2, 9, 15, 28, 38, 42]. However, these outpainting methods are limited by the freedom of control over the generated content. Most recently, with the rapid advancements in generative models, text-conditioned panoramic image generation [3, 5, 23, 24, 39, 41, 43, 53, 54] has received increasing research efforts. More specifically, Text2Light [5] first proposed to generate panoramic images from text descriptions, using CLIP [30] to encode text together with VQGAN [11] to encode panoramic images. To generate images at arbitrary sizes, DiffCollage [54], Multi-Diffusion [3] and SyncDiffusion [23] aggregated the intermediate outputs of pre-trained diffusion models. PanoGen [24] proposed to recursively outpaint the images generated by pretrained models to create panoramic images. However, those methods either fail to handle spherical distortions nor to ensure continuity between the left and right boundaries. To address this problem, MVDiffusion [39] generated perspective images and stitches them into panoramic images. However, the stitched panoramic image may witness its limitation in the field of view. PanFusion [53] introduced an attention mechanism for perspective images to enhance the generation of local details, while the main branch focused on generating the overall panoramic image. However, PanFusion utilizes two pretrained diffusion models [18] whereby multiple perspective images are processed in a separate branch; this can lead to unnatural global structures and result in heavy computational complexity.", + "bbox": [ + 511, + 372, + 908, + 825 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Convolution Operation for Panoramic Images. Several attempts have introduced modifications to the convolution operation [7, 8, 12, 21, 29, 36, 37, 45, 55] in order to effectively manage spherical distortions for panoramic image classification and segmentation tasks. The pioneering work [36]", + "bbox": [ + 511, + 825, + 908, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "5825", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "adopted kernels with different shapes at different latitudes in object detection task. Inspired by these works, deformable kernels [12, 55] were adopted to handle the spherical geometry. On the other hand, Coors et al. [8] retained the kernel unchanged on the tangent plane of the sphere and projected it back to the sphere using inverse gnomonic projection in image classification task. For the same task, Xu et al. [45] proposed to enlarge the receptive filed of convolution kernel by defining the kernel area in sphere. Furthermore, the rotation invariance property has been considered for designing spherical convolution in image classification [7] and image segmentation tasks [29]. However, spherical convolution operations have not yet been proposed for generative models that essentially suffer from improper spherical geometry. Therefore, we propose to incorporate the guidance from SMConv by exponential mapping, operating on $S^2$ manifold for high quality panoramic image generation.", + "bbox": [ + 89, + 90, + 483, + 349 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 89, + 366, + 225, + 383 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Spherical Manifold Convolution", + "text_level": 1, + "bbox": [ + 89, + 392, + 374, + 407 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The spherical distortion introduced by ERP leads to object deformation and interrupts the continuity of content on the sphere, which together significantly hinder the correct information aggregation of standard convolution. In contrast, performing convolution operation directly on the spherical domain is of significant potential to preserve the spherical geometry of panoramic images and to mitigate spherical distortions. By inspecting that panoramic images can be viewed as pixels distributed on a spherical surface, i.e., the $S^2$ manifold, we propose the SMConv operation that performs convolution on the $S^2$ manifold. More specifically, we employ exponential mapping [13] to project uniform convolution kernels from the tangent space onto the $S^2$ manifold, enabling SMConv with kernels to uniformly aggregate the content across the $S^2$ manifold. More importantly, due to the length-preserving property of exponential mapping [13], SMConv retains the uniformity of geodesic distances, known as great-circle distances [40], on the $S^2$ manifold. This enables SMConv to better handle spherical distortions, providing an effective solution for panoramic image generation. Compared to previous spherical convolution based on inverse gnomonic projection [8], employing exponential mapping to construct SMConv kernels enables an expanded receptive field, refined kernel patterns, and improved preservation of spherical geometry, as demonstrated in Fig. 2.", + "bbox": [ + 89, + 415, + 483, + 792 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "More specifically, the $S^2$ manifold is defined as a set of points $x \\in \\mathbb{R}^3$ with norm 1, which can be parameterized by spherical coordinates $\\theta \\in [0,2\\pi]$ and $\\varphi \\in [-\\pi/2, \\pi/2]$ . Through a transformation between spherical and Cartesian coordinates, we can obtain 3D Cartesian coordinates and then apply the exponential mapping. For instance, a point $\\mathbf{p} = (\\theta, \\varphi)$ on $S^2$ manifold can be converted to a corre", + "bbox": [ + 89, + 795, + 483, + 902 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/7efba72479e359fdadabea1b821408dd4a8373bbefaba8ec641e63cee73d8852.jpg", + "image_caption": [ + "Figure 2. Comparison of exponential mapping (denoted as $\\mathrm{Exp}_{\\mathbf{p}}(\\cdot))$ and inverse gnomonic projection (denoted as $\\mathrm{IGP}(\\cdot))$ [8] through forward and backward mappings. Left: visualization of a point $\\mathbf{v}$ in the tangent space $T_{\\mathbf{p}}S^{2}$ at point $\\mathbf{p}$ , mapped onto the sphere. Right: projection of a panoramic image, the spherical data, onto the tangent space $T_{\\mathbf{p}}S^{2}$ centered at $\\mathbf{p} = (\\theta, \\varphi)$ whereby $\\theta = 0$ and $\\varphi = \\frac{\\pi}{4}$ ." + ], + "image_footnote": [], + "bbox": [ + 517, + 85, + 903, + 244 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "sponding point $\\mathbf{p} = (x,y,z)$ with Cartesian coordinates. The exponential mapping [13] can be formulated as follows:", + "bbox": [ + 511, + 356, + 908, + 386 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {q} = \\operatorname {E x p} _ {\\mathbf {p}} (\\mathbf {v}) = \\cos (| | \\mathbf {v} | |) \\mathbf {p} + \\sin (| | \\mathbf {v} | |) \\frac {\\mathbf {v}}{| | \\mathbf {v} | |}, \\quad (1)\n$$\n", + "text_format": "latex", + "bbox": [ + 553, + 393, + 906, + 425 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbf{q}$ denotes the mapped point, $\\mathbf{p}$ is the tangent point, and $\\mathbf{v}$ is a vector representing the coordinates of the kernel points $\\mathbf{K}_{T_{\\mathbf{p}}}(m,n)$ in tangent space $T_{\\mathbf{p}}S^{2}$ centered at $\\mathbf{p}$ .", + "bbox": [ + 511, + 431, + 906, + 478 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In practice, given a predefined position $(u_{m},v_{n})$ of a kernel point $\\mathbf{K}_{T_{\\mathbf{p}}}(m,n)$ in the tangent space $T_{\\mathbf{p}}S^{2}$ , the corresponding vector $\\mathbf{v}$ in $T_{\\mathbf{p}}S^{2}$ can be obtained as follows:", + "bbox": [ + 511, + 479, + 908, + 523 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {v} = u _ {m} \\hat {\\mathbf {e}} _ {\\theta} + v _ {n} \\hat {\\mathbf {e}} _ {\\varphi},\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 534, + 712, + 550 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {e}} _ {\\theta} = (- \\sin \\theta , \\cos \\theta , 0), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 571, + 553, + 906, + 569 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {e}} _ {\\varphi} = (- \\sin \\varphi \\cos \\theta , - \\sin \\varphi \\sin \\theta , \\cos \\varphi),\n$$\n", + "text_format": "latex", + "bbox": [ + 571, + 571, + 848, + 588 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbf{p} = (\\theta, \\varphi)$ denotes the center point in $T_{\\mathbf{p}}S^2$ and the vectors $\\hat{\\mathbf{e}}_{\\theta}, \\hat{\\mathbf{e}}_{\\varphi}$ serve as the basis for the coordinate system in $T_{\\mathbf{p}}S^2$ . Note that the shape and the position of the convolution kernel are provided by $(u_m, v_n)$ in $T_{\\mathbf{p}}S^2$ and we can calculate the position $\\mathbf{q} = \\mathrm{Exp}_{\\mathbf{p}}(\\mathbf{v}) = (\\hat{x}_{m,n}, \\hat{y}_{m,n}, \\hat{z}_{m,n})$ of the kernel point $\\mathbf{K}(m,n)$ on $S^2$ manifold in Cartesian coordinates, via (1) and (2). Here, the kernel $\\mathbf{K}$ is centered at the point $\\mathbf{p}$ on $S^2$ manifold. Due to the properties of the exponential mapping [13], the norm of the mapped point $\\mathbf{q} = (\\hat{x}_{m,n}, \\hat{y}_{m,n}, \\hat{z}_{m,n})$ is guaranteed to be 1, ensuring that the point remains on the $S^2$ manifold. Since we need the position $\\mathbf{q} = (\\theta_m, \\varphi_n)$ in spherical coordinates for performing SMConv, we can simple apply the inverse transformation from Cartesian to spherical coordinates.", + "bbox": [ + 511, + 599, + 906, + 809 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Furthermore, suppose $\\widetilde{\\mathbf{F}}_{S^2},\\mathbf{F}_{S^2}$ are features on $S^2$ manifold. The operation of SMConv can be formulated by", + "bbox": [ + 511, + 810, + 908, + 842 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\widetilde {\\mathbf {F}} _ {S ^ {2}} (\\theta , \\varphi) = \\sum_ {m, n} \\mathbf {F} _ {S ^ {2}} (\\theta + \\theta_ {m}, \\varphi + \\varphi_ {n}) \\cdot \\mathbf {K} (m, n), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 529, + 849, + 906, + 883 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\left(\\theta_ {m}, \\varphi_ {n}\\right) = \\mathbf {q} = \\operatorname {E x p} _ {\\mathbf {p}} \\left(\\mathbf {K} _ {T _ {\\mathbf {p}}} (m, n)\\right),\n$$\n", + "text_format": "latex", + "bbox": [ + 537, + 886, + 785, + 904 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "5826", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/efd97cab99608783b86bfab09bc73bdf8999a217e8f906208ba2cb90653d9477.jpg", + "image_caption": [ + "Figure 3. Overall architecture of the proposed SMGD model, which introduces SMUNet as the denoising network and a panoramic VQGAN as the image encoder. The SMUNet is primarily constructed by the SMConv, operating on the $S^2$ manifold, and integrates the SPB along with circular encoder and decoder blocks." + ], + "image_footnote": [], + "bbox": [ + 91, + 85, + 901, + 349 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In (3), $\\mathbf{q}$ is the position of the kernel point $\\mathbf{K}(m,n)$ for spherical features in ERP format, which is obtained by performing exponential mapping for the kernel point $\\mathbf{K}_{T_{\\mathbf{p}}}(m,n)$ in the tangent space $T_{\\mathbf{p}}S^{2}$ . Moreover, $\\mathbf{K}$ represents the kernel centered at point $\\mathbf{p}$ on $S^2$ manifold and the weight of $\\mathbf{K}$ is identical to that of $\\mathbf{K}_{T_{\\mathbf{p}}}$ . By uniformly choosing $u_{m}$ and $v_{m}$ and constructing a regular kernel in the tangent space $T_{\\mathbf{p}}S^{2}$ , we can perform SMConv with kernels uniformly distributed across the $S^2$ manifold using exponential mapping, according to (3).", + "bbox": [ + 89, + 419, + 485, + 570 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. SMConv Inspired Diffusion Model", + "text_level": 1, + "bbox": [ + 89, + 579, + 393, + 595 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Building upon SMConv, we introduce the SMGD model to achieve high-quality panoramic image generation. The overall architecture of our SMGD model is shown in Fig. 3. The diffusion process can be formulated as", + "bbox": [ + 89, + 599, + 486, + 662 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nq (\\mathbf {z} _ {t} | \\mathbf {z} _ {t - 1}) := \\mathcal {N} (\\mathbf {z} _ {t - 1}; \\sqrt {1 - \\beta_ {t}} \\mathbf {z} _ {t - 1}, \\beta_ {t} \\mathbf {I}),\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 670, + 392, + 690 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {z} _ {t - 1} = \\frac {1}{\\sqrt {\\alpha_ {t}}} \\left(\\mathbf {z} _ {t} - \\frac {1 - \\alpha_ {t}}{\\sqrt {1 - \\bar {\\alpha} _ {t}}} \\epsilon_ {\\psi} (\\mathbf {z} _ {t}, t, \\mathbf {T})\\right) + \\sqrt {\\beta_ {t}} \\eta , ^ {(4)}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 691, + 483, + 728 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{z}_{t - 1},\\mathbf{z}_t$ denote the noisy spherical features with the same dimensionality as the data, at time step $t - 1\\in [0,T]$ and $t\\in [0,T]$ respectively. $\\mathbf{T}$ is the text embedding obtained by CLIP model [31], i.e., from the pretrained CLIP ViT/L-14 model. In addition, $q(\\mathbf{z}_t|\\mathbf{z}_{t - 1})$ represents the forward process gradually adding noise to the image, $\\epsilon_{\\psi}(\\mathbf{z}_t,t)$ is the noise predicted by our SMUNet at time step $t$ , $\\psi$ represents the parameters of SMUNet and $\\eta$ denotes a random Gaussian noise. Furthermore, the noise scheduler configuration $\\beta_{t}$ is a hyperparameter that controls the strength of the noise involved at each time step. The parameter $\\alpha_{t}$ is defined as", + "bbox": [ + 89, + 734, + 486, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$\\alpha_{t} = 1 - \\beta_{t}$ , and $\\bar{\\alpha}_{t}$ is given by $\\bar{\\alpha}_{t} = \\prod_{s=1}^{t} \\alpha_{s}$ , the default setting within diffusion models.", + "bbox": [ + 511, + 417, + 906, + 450 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As the core component of the SMGD model, our SMUNet predicts noise at each time step $t$ from the noisy spherical feature $\\mathbf{z}_t$ and facilitates the recovery of the spherical feature $\\mathbf{z}_0$ on $S^2$ manifold by subtracting the predicted noise. We employ spherical processor blocks (SPBs), along with circular encoder and decoder blocks based on circular convolution [2] at shallower layers of SMUNet. Then, we apply SME and SMD blocks at deeper layers with middle blocks consisted of spherical manifold residual blocks (SMRBs) and cross-attention layers to integrate text embeddings $\\mathbf{T}$ . The design of a hybrid architecture to integrate circular convolution at shallower layers with SMConv at deeper layers is motivated by the following considerations. As resolution increases, spherical convolution operation including SMConv, consumes substantially higher GPU memory cost compared to standard convolution [8] due to its bilinear interpolation. Furthermore, in deeper layers of SMUNet where spatial resolution is reduced, the sparse pixel distribution near the poles becomes increasingly pronounced, exacerbating the limitations of standard convolution. Consequently, the proposed hybrid architecture well balances between output quality, computational efficiency, and memory utilization.", + "bbox": [ + 511, + 450, + 908, + 781 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Then, the SME block comprises 2 SMRBs, 2 cross-attention layers, and an SMConv with a stride of 2 for downsampling. For the $l$ -th SME block, the SMRB can be formulated as", + "bbox": [ + 511, + 781, + 908, + 842 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\widetilde {\\mathbf {z}} _ {t} ^ {E _ {l}} = \\mathrm {S M C o n v} [ \\mathrm {S M C o n v} [ \\mathbf {z} _ {t} ^ {E _ {l}} ] + g _ {\\psi} (t) ] + \\mathbf {z} _ {t} ^ {E _ {l}}, \\quad (5)\n$$\n", + "text_format": "latex", + "bbox": [ + 547, + 845, + 906, + 867 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where the calculation of SMConv follows (3) and $g_{\\psi}(t)$ is the learnable time-step embedding function [32]. Moreover,", + "bbox": [ + 511, + 869, + 908, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "5827", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$\\mathbf{z}_t^{E_l}$ denotes the input, while $\\widetilde{\\mathbf{z}}_t^{E_l}$ represents the output for this layer. The structure of the SMD block closely mirrors that of the indexing SME block, with the upsampling layer using nearest-neighbor interpolation. Note that the final SME block and the first SMD block omit upsampling and downsampling operations. Moreover, the SPB employs the SMConv combined with a $1\\times 1$ residual convolutional layer to enhance the transformation between different feature channels. The circular encoder and decoder blocks follow the structure in LDM [32], with all planar convolutions replaced by circular convolution operations [2]. Notably, the circular encoder blocks use average pooling layers for downsampling. In addition to the SMUNet, we propose to employ a panoramic VQGAN [11] with a downsampling factor of 16 as the image encoder and decoder for our SMGD model. In the panoramic VQGAN, all planar convolutions in the layers of decoder are replaced by circular convolutions [2].", + "bbox": [ + 89, + 88, + 483, + 349 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Spherical Guided Training Strategy", + "text_level": 1, + "bbox": [ + 89, + 356, + 401, + 372 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The training process involves two stages: training the panoramic VQGAN, and then training the SMGD model. During the training of the panoramic VQGAN, reconstruction loss $\\mathcal{L}_{\\mathrm{rec}}$ [11], commitment loss $\\mathcal{L}_{\\mathrm{Commit}}$ [11] and perceptual loss $\\mathcal{L}_{\\mathrm{Per}}$ [19] are employed, which can be formulated as", + "bbox": [ + 89, + 378, + 485, + 467 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {V Q G A N}} = \\mathcal {L} _ {\\text {r e c}} + \\lambda_ {\\text {c o m m i t}} \\mathcal {L} _ {\\text {c o m m i t}} + \\lambda_ {\\text {p e r}} \\mathcal {L} _ {\\text {P e r}}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 109, + 479, + 483, + 496 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda_{\\mathrm{commit}}$ and $\\lambda_{\\mathrm{per}}$ are hyperparameters to control the weights of different loss components during training.", + "bbox": [ + 89, + 505, + 482, + 535 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "During the training of the SMGD model, a spherical loss based on the weighted spherically uniform PSNR (WSPSNR) [50] is employed. The total loss can be formulated as", + "bbox": [ + 89, + 535, + 483, + 593 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {D i f f}} = \\lambda_ {\\text {M S E}} \\mathcal {L} _ {\\text {M S E}} + \\lambda_ {\\text {S M S E}} \\mathcal {L} _ {\\text {S M S E}}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 151, + 595, + 483, + 611 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda_{\\mathrm{MSE}}$ and $\\lambda_{\\mathrm{SMSE}}$ are also balancing hyperparameters. Furthermore, the $\\mathcal{L}_{\\mathrm{MSE}}$ and $\\mathcal{L}_{\\mathrm{SMSE}}$ are given by", + "bbox": [ + 89, + 617, + 483, + 647 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {M S E}} = \\mathbb {E} _ {t, \\mathbf {z} _ {0}, \\epsilon} [ | | \\epsilon - \\epsilon_ {\\psi} (\\mathbf {z} _ {t}, t, \\mathbf {T}) | | _ {2} ^ {2} ], \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 137, + 656, + 483, + 680 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {S M S E}} = \\mathbb {E} _ {t, \\mathbf {z} _ {0}, \\epsilon} [ | | \\mathbf {W} \\odot (\\epsilon - \\epsilon_ {\\psi} (\\mathbf {z} _ {t}, t, \\mathbf {T})) | | _ {2} ^ {2} ],\n$$\n", + "text_format": "latex", + "bbox": [ + 133, + 676, + 441, + 694 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{W}$ is given by the WS-PSNR [50] to reflect the panoramic characteristics as", + "bbox": [ + 89, + 702, + 482, + 733 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {W} _ {j} = \\cos \\left(\\left(\\frac {j + 0 . 5}{H} - 0. 5\\right) \\pi\\right), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 181, + 741, + 483, + 771 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $j \\in [0, H]$ is the index of height in $\\mathbf{z}_t$ and $H$ denotes the height of the spherical feature $\\mathbf{z}_t$ . $\\mathcal{L}_{\\mathrm{SMSE}}$ can improve the preservation of spherical geometry, e.g., a sparse pixel distribution near the poles.", + "bbox": [ + 89, + 779, + 483, + 839 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Quality Evaluation for Panoramic Images", + "text_level": 1, + "bbox": [ + 89, + 848, + 447, + 864 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "FID score is the de facto standard for evaluating the quality of generated images, whereas performing a direct calculation", + "bbox": [ + 89, + 869, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/5a7d36517b32d90b987b488e8262b38e73b5d03ad968704942cd54fb3cd45c24.jpg", + "image_caption": [ + "Figure 4. Illustration of the conversion between ERP and CMP formats. In the CMP format, four blue-bordered images represent equatorial regions, while two orange-bordered images represent polar areas. These images are grouped accordingly, with $\\mathrm{FID}_{\\mathrm{equ}}$ and $\\mathrm{FID}_{\\mathrm{pole}}$ calculated separately for each group." + ], + "image_footnote": [], + "bbox": [ + 517, + 90, + 903, + 191 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "of FID on panoramic images in rectangular ERP format introduces severe distortion, leading to inaccuracy. We propose to convert the generated panoramic images from ERP format to CMP format consisted of 6 square images for calculating FID score. The mapping from ERP to CMP format [50, 52] provides a robust solution to evaluate the quality of generated panoramic images. Note that although sharing similar goals with the concurrent work [6], our FID-based metrics are able to reflect detailed quality for different regions when evaluating generated panoramic images. As illustrated in Fig. 4, the conversion reveals the continuity across the left and right boundaries and represents the spherical geometry by obtaining 4 images representing areas near the equator and 2 images representing areas near the poles. Thus, the 6 square images within CMP format can be categorized into 2 groups, allowing for separate FID calculations to assess the quality of generated images. More specifically, $\\mathrm{FID}_{\\mathrm{equ}}$ is calculated using the 4 images near the equator, while $\\mathrm{FID}_{\\mathrm{pole}}$ is derived from the 2 images near the poles. $\\mathrm{FID}_{\\mathrm{equ}}$ primarily assesses the quality and visual fidelity of the panoramic content, whereas $\\mathrm{FID}_{\\mathrm{pole}}$ evaluates the preservation of spherical geometry. Analyzing $\\mathrm{FID}_{\\mathrm{equ}}$ and $\\mathrm{FID}_{\\mathrm{pole}}$ provides a comprehensive measure of the quality of the generated panoramic image.", + "bbox": [ + 511, + 279, + 908, + 643 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiment", + "text_level": 1, + "bbox": [ + 513, + 657, + 637, + 672 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Settings", + "text_level": 1, + "bbox": [ + 511, + 681, + 718, + 698 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Dataset. We employed the widely-used Matterport3D [4] dataset to evaluate the performance for text-conditioned panoramic image generation. More specifically, we obtained 10,912 panoramic images with a resolution of $1024 \\times 512$ according to [26]. We then utilized the BLIP-2 [25] model to generate text descriptions for each panoramic image, forming paired text-image data for training and testing. To enable FID-10K calculation, we allocated 10,000 images for training and testing.", + "bbox": [ + 511, + 703, + 908, + 839 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation Details. We adopted the panoramic VQGAN according to Sec. 3.2. SMUNet, the denoising network for our SMGD model, was configured with 320 and 640 channels for the two circular encoder and decoder blocks,", + "bbox": [ + 511, + 839, + 908, + 898 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5828", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/20c700be3c8ab732b6535f2cffd2c0bd373656b694f65549642224e23339a65f.jpg", + "table_caption": [ + "Table 1. Comparisons in terms of FID and CLIP score (denoted as CS) values for panoramic images generated by our method and existing baseline methods, along with evaluations of parameter number (denoted by Para.) and time required to inference (denoted as $t_{\\mathrm{sample}}$ ). We represent the best numbers by red color and the second best by blue color." + ], + "table_footnote": [], + "table_body": "
MethodsFID ↓FAED ↓OmniFID ↓FIDavg ↓FIDcent ↓FIDbord ↓FIDrand ↓FIDequ ↓FIDpole ↓CS ↑Para. ↓t_sample(s) ↓
Multi-Diffusion [3]69.961.1268.7056.7458.1956.5355.5027.2687.730.1821.3B39.45 ± 0.31
SD2+LoRA [18, 32]59.321.4769.0750.0451.7849.1649.1724.4988.680.1711.4B6.48 ± 0.49
Text2Light [5]48.424.1063.1643.6846.9346.6437.4730.8772.060.1881.0B50.12 ± 11.18
MVDiffusion [39]56.210.9394.5338.1538.8837.6037.9627.58125.480.1811.4B97.14 ± 0.79
PanFusion [53]16.150.3527.6414.3414.9113.9614.1414.8131.990.1822.3B38.51 ± 11.57
SMGD (Ours)9.850.2218.249.709.779.629.717.8821.050.1891.4B2.69 ± 0.39
", + "bbox": [ + 94, + 135, + 903, + 232 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "and 960 and 1280 channels for the two SME and SMD blocks utilizing the $3 \\times 3$ SMConv. At the first training stage, we trained the panoramic VQGAN with the Adam optimizer [22], using a batch size of 1 and a learning rate of $4.5 \\times 10^{-6}$ for 300 epochs, with $\\lambda_{\\mathrm{commit}} = 1$ and $\\lambda_{\\mathrm{per}} = 0.8$ . At the second stage, we trained the SMGD model from the scratch with the AdamW optimizer [27], a batch size of 8, and learning rate of $1 \\times 10^{-5}$ for 800 epochs by setting $\\lambda_{\\mathrm{MSE}} = 0.8$ and $\\lambda_{\\mathrm{SMSE}} = 0.2$ . Moreover, the noise scheduler configuration $\\beta_{t}$ followed that of LDM [32]. For inference, we used a DDIM sampler [35] with 50 steps and a classifier-free guidance scale [32] of 2.5.", + "bbox": [ + 88, + 256, + 485, + 439 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines. We compared our SMGD model with existing text-conditioned panoramic image generation methods, including Multi-Diffusion [3], Text2Light [5], MVDiffusion [39] and PanFusion [53]. We also compared our SMGD model with the widely-used stable diffusion model [32] by performing a rank-16 LoRA [18] (denoted as SD2+LoRA). Different from [53], we directly trained the LoRA model with a learning rate of $1 \\times 10^{-6}$ for 50 epochs on our training data.", + "bbox": [ + 88, + 450, + 486, + 585 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metrics. We computed $\\mathrm{FID}_{\\mathrm{equ}}$ and $\\mathrm{FID}_{\\mathrm{pole}}$ , as described in Sec. 3.4 to effectively assess the quality of panoramic images. Note that $\\mathrm{FID}_{\\mathrm{equ}}$ was calculated against 40K images and $\\mathrm{FID}_{\\mathrm{pole}}$ was calculated against 20K images. To further validate our metrics and assess the quality of the generated images, we calculated an extra FID value similar to that used in previous methods [53]. More specifically, we cropped the panoramic images into 3 representative areas instead of obtaining 20 perspective images in calculating FID values. Specifically, we cropped the panoramic images into the center, cross-border and random patches with size $512 \\times 512$ . Then, we calculated the corresponding FID value of them and obtained the $\\mathrm{FID}_{\\mathrm{cent}}$ , $\\mathrm{FID}_{\\mathrm{bord}}$ and $\\mathrm{FID}_{\\mathrm{rand}}$ . By averaging these three FID values calculated against 10K images, we obtained $\\mathrm{FID}_{\\mathrm{avg}}$ . Moreover, we reported FID, OmniFID [6], and Fréchet Auto-Encoder Distance (FAED) [53] to comprehensively evaluate the quality of the generated full panoramic images. On the other hand, to evaluate the alignment between text and generated panoramic images, we utilized the widely adopted CLIP score [16] (denoted", + "bbox": [ + 91, + 598, + 485, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "as CS).1 Furthermore, we recorded the time for inference $t_{\\mathrm{sample}}$ by averaging the time required to generate 10 images on an NVIDIA 4090 GPU for each method.", + "bbox": [ + 511, + 256, + 906, + 301 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Comparison with Existing Methods", + "text_level": 1, + "bbox": [ + 511, + 310, + 821, + 325 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Quantitative Results. In Table 1, we report the quantitative results of our method and baselines for panoramic image generation from text descriptions. From this table, our method consistently achieves the lowest FID values, the highest CLIP scores, and the least sampling times. More importantly, among the top-performance models, our method maintains a comparatively lower parameter number. More specifically, the lower values of $\\mathrm{FID}_{\\mathrm{cent}}$ , $\\mathrm{FID}_{\\mathrm{bord}}$ , $\\mathrm{FID}_{\\mathrm{rand}}$ , and $\\mathrm{FID}_{\\mathrm{avg}}$ demonstrate that our SMGD model achieves the state-of-the-art performance within text-conditioned panoramic image generation2. Moreover, the even lower $\\mathrm{FID}_{\\mathrm{equu}}$ and $\\mathrm{FID}_{\\mathrm{pole}}$ values indicate that our method well preserves the spherical geometry of panoramic images, compared to other baseline methods. Note that MVDiffusion [39] and SD2+LoRA achieved low $\\mathrm{FID}_{\\mathrm{avg}}$ values but high $\\mathrm{FID}_{\\mathrm{pole}}$ values. This may be due to the fact that images generated by MVDiffusion exhibited a limited field of view, while images produced by SD2+LoRA lacked the preservation of the spherical geometry in panoramic images. Furthermore, as shown in Table 1, our SMGD model achieves the fastest inference time among all comparison models, outperforming PanFusion [53], the second-best method in terms of image quality, by more than a factor of 10 in inference speed.", + "bbox": [ + 511, + 332, + 908, + 679 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Qualitative Results. We randomly chose 2 text descriptions and report the generated panoramic images with corresponding real images from dataset in Fig. 5. As illustrated in the figure, the panoramic images generated by our SMGD model achieved the best subjective quality and the most accurate alignment with the text descriptions. More specifically, improper spherical geometry, including inconsistency across the border and an unnatural pixel distribution near the poles, was observed in panoramic images generated by SD2+LoRA, Text2Light, and Multi-Diffusion, due to their", + "bbox": [ + 511, + 679, + 910, + 830 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "1The CLIP score was computed using the ViT-L/14 CLIP model, consistent with the CLIP model employed in our SMGD model.", + "To further validate the robustness of our model, we conduct additional evaluations on the Structured3D [56] dataset in Supplementary Sec. A. We also provide additional qualitative results in Supplementary Sec. C." + ], + "bbox": [ + 511, + 837, + 906, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "5829", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/2e1faf734cdc582f4c64069394f0b38288bb2aba3b85970453cfe576b5f0379c.jpg", + "image_caption": [ + "Figure 5. Qualitative results for text-conditioned panoramic image generation with real images from dataset displayed on the top row. We also show the copy from the left side of the panoramic image to the right to reveal the consistency across the whole image. Note that images generated by MVDiffusion [39] exhibit a limited field of view, resulting in images with a reduced height. We highlight the improper spherical geometry, artifact regions and distorted objects and lines with corresponding color of boxes, which are mitigated by our SMGD model." + ], + "image_footnote": [], + "bbox": [ + 145, + 89, + 838, + 750 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "neglect of spherical distortions. Panoramic images generated by Panfusion and MVDiffusion suffered from artifact regions and distorted objects and lines as highlighted with corresponding color boxes in Fig. 5. Moreover, images", + "bbox": [ + 89, + 835, + 485, + 897 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "generated by MVDiffusion exhibited a limited field of view. This limitation impeded the visual quality and completeness of the generated images, making them less suitable for fully immersive visual experiences. In contrast, our SMGD", + "bbox": [ + 511, + 835, + 908, + 897 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "5830", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/aa848cbcf9215c6b6a4660b0bcaa3ee54d5d887bd240233a2af47e8ae427b6c7.jpg", + "table_caption": [ + "Table 2. Ablation study for our SMGD model. We compare the performance achieved by adding individual components to the baseline model sequentially, including the spherical loss (denoted by $\\mathcal{L}_{\\mathrm{SMSE}}$ ), the SMD blocks, the SME blocks, and the SPB modules. We represent the best numbers by red color." + ], + "table_footnote": [], + "table_body": "
\\( \\mathcal{L}_{\\text{SMSE}} \\)SMD BlocksSME BlocksSPBFID\\( \\mathrm{FID}_{\\text{avg}} \\downarrow \\)\\( \\mathrm{FID}_{\\text{cent}} \\downarrow \\)\\( \\mathrm{FID}_{\\text{bord}} \\downarrow \\)\\( \\mathrm{FID}_{\\text{rand}} \\downarrow \\)\\( \\mathrm{FID}_{\\text{equ}} \\downarrow \\)\\( \\mathrm{FID}_{\\text{pole}} \\downarrow \\)CS ↑
××××23.1120.5720.3420.9120.4515.3534.870.187
✓×××20.0317.5017.0518.0417.4012.7832.950.187
✓✓××13.3512.8712.7912.9412.889.6120.450.188
✓✓✓×12.3611.4211.4611.6011.199.2121.700.186
✓✓✓✓9.859.709.779.629.717.8821.050.189
", + "bbox": [ + 94, + 125, + 903, + 224 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/67d77b24560738f886b2cb57ef64d73c118993529795ed41f55e20a561613c8f.jpg", + "image_caption": [ + "Comparison in terms of FID and CS" + ], + "image_footnote": [], + "bbox": [ + 94, + 253, + 475, + 411 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c20ba096886dc23931c46ebd905b36949f02da1b3b3c82f77ecd5def95d75dc0.jpg", + "image_caption": [ + "Figure 6. Comparison between LDM+SphereNet and the proposed SMGD model. Top: Bar plot of FID and CLIP score values achieved by these two methods. Bottom: Subjective results of panoramic images generated by the text prompt \"a hallway in a house\". The comparison is conducted in both ERP and CMP formats, with distorted objects and lines highlighted." + ], + "image_footnote": [], + "bbox": [ + 93, + 415, + 292, + 613 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/25dc8097b3ba681eff062dbc2a98e608db193922ccd75e1b2131cc944d1d2f31.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 295, + 417, + 478, + 613 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "model produced visually pleasing panoramic images with a well-defined overall structure and proper spherical geometry.", + "bbox": [ + 89, + 715, + 485, + 747 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 757, + 243, + 772 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this section, we conducted ablation study to systematically evaluate the impact of core components in our SMGD model. More specifically, since the spherical loss and SMConv are important for handling spherical distortions and preserving spherical geometry in panoramic images, we analyzed performance changes by gradually adding the spherical loss $\\mathcal{L}_{\\mathrm{SMSE}}$ , the SMD blocks, the SME blocks and the SPB. As reported in Table 2, the involvement of SMD blocks", + "bbox": [ + 89, + 780, + 485, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "significantly enhanced performance, with SME blocks further contributing to the improvements. Moreover, both the spherical loss function and the SPB effectively improved the quality of generated panoramic images. Therefore, each module plays a vital role in optimizing the performance of the SMGD model.", + "bbox": [ + 511, + 238, + 906, + 327 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Furthermore, to validate the effectiveness of SMConv compared to previous spherical convolution methods, we implemented LDM+SphereNet, which employs the panoramic VQGAN and a U-Net architecture identical to our SMUNet, whilst all SMConv layers were replaced by the convolution operations proposed in SphereNet [8]. The training and inference were kept consistent with those of our SMGD model. As shown in Fig. 6, we can conclude that the quality of panoramic images generated by LDM+SphereNet is limited, whereas our proposed SMGD model produces visually pleasing images with lower FID values and a higher CLIP score3. This demonstrates the effectiveness of using SMConv for advanced panoramic image generation.", + "bbox": [ + 511, + 329, + 908, + 525 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 537, + 633, + 554 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we have proposed a novel method named as spherical manifold guided diffusion (SMGD) model to address the challenges of spherical distortions associated with panoramic image generation. This was achieved by the spherical manifold convolution (SMConv) operation that is able to uniformly sample spherical content within each convolution layer. Different from existing spherical convolution operations applied in other tasks, the proposed SMConv operation was built upon the exponential mapping grounded in manifold theory, which is able to preserve lengths on the tangent plane and exhibit enhanced local properties. The overall training procedure of our SMGD model was also inspired by the spherical loss. Furthermore, we have proposed to calculate $\\mathrm{FID}_{\\mathrm{equ}}$ and $\\mathrm{FID}_{\\mathrm{pole}}$ for panoramic images in cubemap format to improve the evaluation on the quality of generated panoramic images. Experimental results have demonstrated that our method achieves superior performance in text-conditioned image generation, surpassing existing state-of-the-art methods in both the quality of generated images and computational efficiency.", + "bbox": [ + 511, + 564, + 908, + 866 + ], + "page_idx": 7 + }, + { + "type": "page_footnote", + "text": "3More qualitative comparison results between SMGD and LDM+SphereNet, as well as a detailed analysis, are provided in Supplementary Sec. B.", + "bbox": [ + 513, + 875, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "5831", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 91, + 90, + 250, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported by Natural Science Foundation of China (NSFC) under Grants 62450131, 62206011, 62231002 and 62401027, and Beijing Natural Science Foundation under Grant L223021.", + "bbox": [ + 89, + 114, + 485, + 174 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 186, + 187, + 204 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Hao Ai, Zidong Cao, Haonan Lu, Chen Chen, Jian Ma, Pengyuan Zhou, Tae-Kyun Kim, Pan Hui, and Lin Wang. Dream360: Diverse and immersive outdoor virtual scene creation via transformer-based $360^{\\circ}$ image outpainting. IEEE Transactions on Visualization and Computer Graphics, 2024. 2", + "[2] Naofumi Akimoto, Yuhi Matsuo, and Yoshimitsu Aoki. Diverse plausible 360-degree image outpainting for efficient 3dcg background creation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11441-11450, 2022. 2, 4, 5", + "[3] Omer Bar-Tal, Lior Yariv, Yaron Lipman, and Tali Dekel. Multidiffusion: Fusing diffusion paths for controlled image generation. In International Conference on Machine Learning, 2023. 2, 6", + "[4] Angel Chang, Angela Dai, Thomas Funkhouser, Maciej Halber, Matthias Niessner, Manolis Savva, Shuran Song, Andy Zeng, and Yinda Zhang. Matterport3d: Learning from rgb-d data in indoor environments. arXiv preprint arXiv:1709.06158, 2017. 5", + "[5] Zhaoxi Chen, Guangcong Wang, and Ziwei Liu. Text2light: Zero-shot text-driven hdr panorama generation. ACM Transactions on Graphics (TOG), 41(6):1-16, 2022. 2, 6", + "[6] Anders Christensen, Nooshin Mojab, Khushman Patel, Karan Ahuja, Zeynep Akata, Ole Winther, Mar Gonzalez-Franco, and Andrea Colaco. Geometry fidelity for spherical images. In European Conference on Computer Vision, pages 276-292. Springer, 2024. 5, 6", + "[7] Taco S Cohen, Mario Geiger, Jonas Köhler, and Max Welling. Spherical cnns. arXiv preprint arXiv:1801.10130, 2018. 2, 3", + "[8] Benjamin Coors, Alexandru Paul Condurache, and Andreas Geiger. Spherenet: Learning spherical representations for detection and classification in omnidirectional images. In Proceedings of the European conference on computer vision (ECCV), pages 518-533, 2018. 2, 3, 4, 8", + "[9] Mohammad Reza Karimi Dastjerdi, Yannick Hold-Geoffroy, Jonathan Eisenmann, Siavash Khodadadeh, and Jean-François Lalonde. Guided co-modulated gan for 360 field of view extrapolation. In 2022 International Conference on 3D Vision (3DV), pages 475–485. IEEE, 2022. 2", + "[10] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 1", + "[11] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12873-12883, 2021. 2, 5", + "[12] Clara Fernandez-Labrador, Jose M Facil, Alejandro Perez-Yus, Cédric Demonceaux, Javier Civera, and Jose J Guerrero." + ], + "bbox": [ + 93, + 212, + 485, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Corners for layout: End-to-end layout recovery from 360 images. IEEE Robotics and Automation Letters, 5(2):1255-1262, 2020. 2, 3", + "[13] Jean Gallier. Basics of Classical Lie Groups: The Exponential Map, Lie Groups, and Lie Algebras, pages 367-414. Springer New York, New York, NY, 2001. 3", + "[14] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 1", + "[15] Takayuki Hara, Yusuke Mukuta, and Tatsuya Harada. Spherical image generation from a single image by considering scene symmetry. Proceedings of the AAAI Conference on Artificial Intelligence, 35(2):1513-1521, 2021. 2", + "[16] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 7514-7528, 2021. 6", + "[17] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 2", + "[18] Edward J Hu, Phillip Wallis, Zeyuan Allen-Zhu, Yuzhhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. In International Conference on Learning Representations, 2022. 2, 6", + "[19] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In European Conference on Computer Vision, pages 694–711. Springer International Publishing, 2016. 5", + "[20] Minguk Kang, Jun-Yan Zhu, Richard Zhang, Jaesik Park, Eli Shechtman, Sylvain Paris, and Taesung Park. Scaling up gans for text-to-image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10124-10134, 2023. 1", + "[21] Renata Khasanova and Pascal Frossard. Geometry aware convolutional filters for omnidirectional images representation. In International conference on machine learning, pages 3351-3359. PMLR, 2019. 2", + "[22] Diederik P Kingma. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6", + "[23] Yuseung Lee, Kunho Kim, Hyunjin Kim, and Minhyuk Sung. Syncdiffusion: Coherent montage via synchronized joint diffusions. Advances in Neural Information Processing Systems, 36:50648-50660, 2023. 2", + "[24] Jialu Li and Mohit Bansal. Panogen: Text-conditioned panoramic environment generation for vision-and-language navigation. Advances in Neural Information Processing Systems, 36, 2024. 2", + "[25] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730-19742. PMLR, 2023. 5", + "[26] Chieh Hubert Lin, Chia-Che Chang, Yu-Sheng Chen, Da-Cheng Juan, Wei Wei, and Hwann-Tzong Chen. Coco-gan:" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "5832", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Generation by parts via conditional coordinating. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4512-4521, 2019. 5", + "[27] I Loshchilov. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6", + "[28] Zhuqiang Lu, Kun Hu, Chaoyue Wang, Lei Bai, and Zhiyong Wang. Autoregressive omni-aware outpainting for open-vocabulary 360-degree image generation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 14211-14219, 2024. 2", + "[29] Thomas W Mitchel, Noam Aigerman, Vladimir G Kim, and Michael Kazhdan. Möbius convolutions for spherical cnns. In ACM SIGGRAPH 2022 Conference Proceedings, pages 1-9, 2022. 2, 3", + "[30] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2", + "[31] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1(2): 3, 2022. 2, 4", + "[32] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684-10695, 2022. 1, 4, 5, 6", + "[33] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in neural information processing systems, 35:36479-36494, 2022. 1", + "[34] Axel Sauer, Tero Karras, Samuli Laine, Andreas Geiger, and Timo Aila. Stylegan-t: Unlocking the power of gans for fast large-scale text-to-image synthesis. In International conference on machine learning, pages 30105-30118. PMLR, 2023. 1, 2", + "[35] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020.6", + "[36] Yu-Chuan Su and Kristen Grauman. Learning spherical convolution for fast features from 360 imagery. Advances in neural information processing systems, 30, 2017. 2", + "[37] Yu-Chuan Su and Kristen Grauman. Kernel transformer networks for compact spherical convolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9442-9451, 2019. 2", + "[38] Julius Surya Sumantri and In Kyu Park. 360 panorama synthesis from a sparse set of images with unknown field of view. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 2386-2395, 2020. 2", + "[39] Shitao Tang, Fuyang Zhang, Jiacheng Chen, Peng Wang, and Yasutaka Furukawa. Mvdiffusion: Enabling holistic multiview image generation with correspondence-aware diffusion. ArXiv, abs/2307.01097, 2023. 2, 6, 7" + ], + "bbox": [ + 91, + 90, + 483, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[40] T. Vincenty. Direct and inverse solutions of geodesics on the ellipsoid with application of nested equations. Survey Review, 23(176):88-93, 1975. 2, 3", + "[41] Hai Wang, Xiaoyu Xiang, Yuchen Fan, and Jing-Hao Xue. Customizing 360-degree pancreas through text-to-image diffusion models. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 4933-4943, 2024. 2", + "[42] Tianhao Wu, Chuanxia Zheng, and Tat-Jen Cham. Panodiffusion: 360-degree panorama outpainting via diffusion. In The Twelfth International Conference on Learning Representations, 2023. 2", + "[43] Tao Wu, Xuewei Li, Zhongang Qi, Di Hu, Xintao Wang, Ying Shan, and Xi Li. Spherediffusion: Spherical geometry-aware distortion resilient diffusion model. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 6126-6134, 2024. 2", + "[44] Mai Xu, Chen Li, Shanyi Zhang, and Patrick Le Callet. State-of-the-art in 360 video/image processing: Perception, assessment and compression. IEEE Journal of Selected Topics in Signal Processing, 14(1):5-26, 2020. 1", + "[45] Yanyu Xu, Ziheng Zhang, and Shenghua Gao. Spherical dnns and their applications in 360 images and videos. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(10):7235-7252, 2021. 2, 3", + "[46] Jian-Ru Xue, Jian-Wu Fang, and Pu Zhang. A survey of scene understanding by event reasoning in autonomous driving. International Journal of Automation and Computing, 15 (3):249-266, 2018. 1", + "[47] Bangbang Yang, Yinda Zhang, Yijin Li, Zhaopeng Cui, Sean Fanello, Hujun Bao, and Guofeng Zhang. Neural rendering in a room: amodal 3d understanding and free-viewpoint rendering for the closed scene composed of pre-captured objects. ACM Transactions on Graphics (TOG), 41(4):1-10, 2022. 1", + "[48] Bangbang Yang, Wenqi Dong, Lin Ma, Wenbo Hu, Xiao Liu, Zhaopeng Cui, and Yuewen Ma. Dreamspace: Dreaming your room space with text-driven panoramic texture propagation. In 2024 IEEE Conference Virtual Reality and 3D User Interfaces (VR), pages 650-660. IEEE, 2024. 1", + "[49] Michihiro Yasunaga, Armen Aghajanyan, Weijia Shi, Rich James, Jure Leskovec, Percy Liang, Mike Lewis, Luke Zettlemoyer, and Wen-tau Yih. Retrieval-augmented multimodal language modeling. arXiv preprint arXiv:2211.12561, 2022. 2", + "[50] Yan Ye, Elena Alshina, and Jill Boyce. Algorithm descriptions of projection format conversion and video quality metrics in 360lib. In Joint Video Exploration Team of ITU-T SG, 2018. 1, 5", + "[51] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, et al. Scaling autoregressive models for content-rich text-to-image generation. arXiv preprint arXiv:2206.10789, 2(3):5, 2022. 2", + "[52] Li Yu, Yanjun Gao, Farhad Pakdaman, and Moncef Gabbouj. Panoramic image inpainting with gated convolution and contextual reconstruction loss. In ICASSP 2024 - 2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 4255-4259, 2024. 2, 5" + ], + "bbox": [ + 516, + 90, + 906, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "5833", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[53] Cheng Zhang, Qianyi Wu, Camilo Cruz Gambardella, Xiaoshui Huang, Dinh Phung, Wanli Ouyang, and Jianfei Cai. Taming stable diffusion for text to 360 panorama image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6347-6357, 2024. 1, 2, 6", + "[54] Qinsheng Zhang, Jiaming Song, Xun Huang, Yongxin Chen, and Ming-Yu Liu. Diffcollage: Parallel generation of large content with diffusion models. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10188-10198. IEEE, 2023. 2", + "[55] Qiang Zhao, Chen Zhu, Feng Dai, Yike Ma, Guoqing Jin, and Yongdong Zhang. Distortion-aware cnns for spherical images. In IJCAI, pages 1198-1204, 2018. 2, 3", + "[56] Jia Zheng, Junfei Zhang, Jing Li, Rui Tang, Shenghua Gao, and Zihan Zhou. Structured3d: A large photo-realistic dataset for structured 3d modeling. In Proceedings of The European Conference on Computer Vision (ECCV), 2020. 6", + "[57] Yufan Zhou, Ruiyi Zhang, Changyou Chen, Chunyuan Li, Chris Tensmeyer, Tong Yu, Jiuming Gu, Jinhui Xu, and Tong Sun. Towards language-free training for text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17907-17917, 2022. 1" + ], + "bbox": [ + 91, + 90, + 485, + 428 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "5834", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Spherical Manifold Guided Diffusion Model for Panoramic Image Generation/25742454-f26d-401b-a2ab-0a30d46da30a_model.json b/2025/Spherical Manifold Guided Diffusion Model for Panoramic Image Generation/25742454-f26d-401b-a2ab-0a30d46da30a_model.json new file mode 100644 index 0000000000000000000000000000000000000000..77341a96a1db51f3e4a6d527e1d96888e2a82f01 --- /dev/null +++ b/2025/Spherical Manifold Guided Diffusion Model for Panoramic Image Generation/25742454-f26d-401b-a2ab-0a30d46da30a_model.json @@ -0,0 +1,2092 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.107, + 0.131, + 0.892, + 0.154 + ], + "angle": 0, + "content": "Spherical Manifold Guided Diffusion Model for Panoramic Image Generation" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.179, + 0.903, + 0.218 + ], + "angle": 0, + "content": "Xiancheng Sun\\(^{1}\\) Mai Xu\\(^{1}\\) Shengxi Li\\(^{1*}\\) Senmao Ma\\(^{1}\\) Xin Deng\\(^{1}\\) Lai Jiang\\(^{1}\\) Gang Shen\\(^{2}\\) \n\\(^{1}\\)Beihang University \\(^{2}\\)China Tower" + }, + { + "type": "image", + "bbox": [ + 0.129, + 0.226, + 0.499, + 0.487 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.191, + 0.485, + 0.475, + 0.507 + ], + "angle": 0, + "content": "\"A serene room with a large door leading to a balcony and a window showcasing the garden, decorated in soft pastels\"" + }, + { + "type": "image", + "bbox": [ + 0.499, + 0.227, + 0.832, + 0.485 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.499, + 0.485, + 0.829, + 0.508 + ], + "angle": 0, + "content": "\"A bright and airy staircase with large windows overlooking a manicured garden, enhancing the natural light that floods the space.\"" + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.512, + 0.908, + 0.556 + ], + "angle": 0, + "content": "Figure 1. A subjective comparison of zero-shot text-conditioned panoramic image generation between our SMGD model and the state-of-the-art method PanFusion [53]. Our model demonstrates superior performance in generating images with superior overall natural content, enhanced subjective quality, and better alignment with the text description, while benefiting reduced model parameters." + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.567, + 0.327, + 0.582 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.588, + 0.486, + 0.875 + ], + "angle": 0, + "content": "Panoramic image essentially acts as a pivotal role in emerging virtual reality and augmented reality scenarios; however, the generation of panoramic images are essentially challenging due to the intrinsic spherical geometry and spherical distortions caused by equirectangular projection (ERP). To address this, we start from the very basics of \\( S^2 \\) manifold inherent to panoramic images, and propose a novel spherical manifold convolution (SMConv) on \\( S^2 \\) manifold. Based on the SMConv operation, we propose a spherical manifold guided diffusion (SMGD) model for text-conditioned panoramic image generation, which can well accommodate the spherical geometry during generation. We further develop a novel evaluation method by calculating grouped Fréchet inception distance (FID) on cube-map projections, which can well reflect the quality of generated panoramic images, compared to existing methods that randomly crop ERP-distorted content. Experiment results demonstrate that our SMGD model achieves the state-of-the-art generation quality and accuracy, whilst retaining" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.569, + 0.909, + 0.614 + ], + "angle": 0, + "content": "the shortest sampling time in the text-conditioned panoramic image generation task. Codes are publicly available at https://github.com/chronos123/SMGD." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.631, + 0.646, + 0.646 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.659, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Being able to be controlled by text descriptions, panoramic image generation with a full \\(360^{\\circ}\\) horizontal and \\(180^{\\circ}\\) vertical field-of-view [44] opens a new avenue for various applications in virtual reality [48], augmented reality [47], autonomous driving [46], etc. Although recent developments in deep probabilistic models including generative adversarial networks (GANs) [14, 20, 34, 57] and diffusion models [10, 32, 33] have been acting as the workhorse in generating photo-realistic images, panoramic image generation still remains challenging due to their intrinsic spherical geometry property and distorted scene structures. In particular, the equirectangular projection (ERP) [50] employed to convert the panoramic content into a planar rectangular format, introduces spherical distortions and breaks the continuity at the left and right edges. Within the ERP format, rectangular panoramic images with an aspect ratio of 1:2 exhibit a" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.237, + 0.901 + ], + "angle": 0, + "content": "*Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5824" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.182 + ], + "angle": 0, + "content": "non-uniform pixel distribution. In other words, the pixel density is notably sparse near the polar regions and significantly denser around the equatorial region. Consequently, existing state-of-the-art deep probabilistic models designed for generating planar images [31, 34, 49, 51] still yet to generalize well for the panoramic image generation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.187, + 0.486, + 0.429 + ], + "angle": 0, + "content": "To address the panoramic image generation task, several methods proposed to generate rectangular images, including autoregressive generation based on the latent codes [11], stitching latent features [3], and directly autoregressive generation on image patches [24]. To retain the continuity between the left and right boundaries, strategies such as circular convolution [5], circular token mechanisms [2], and exploiting the symmetry of panoramic images [15] have been further proposed. However, the above methods fail to effectively mitigate the spherical distortions, thus motivating the usage of deformable convolution [43] and attention mechanisms [53]. Unfortunately, existing methods still suffer from sub-optimal performance, which lacks well-defined global structure and optimal preservation of spherical geometry, as illustrated in Fig. 1, whilst compensating this with heavy computational complexity." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.433, + 0.486, + 0.902 + ], + "angle": 0, + "content": "In this paper, we start from the intrinsic property of panoramic format, and propose the spherical manifold guided diffusion (SMGD) model operating on \\( S^2 \\) manifold, to generate panoramic images from text descriptions. Since panoramic images can be recognized as pixels distributed on \\( S^2 \\) manifold, we propose the spherical manifold convolution (SMConv), to effectively address spherical distortions and preserve spherical geometry in panoramic images. To the best of our knowledge, there exists no spherical convolution designed for the panorama generation task, in which our SMConv fills this void by performing the correct convolution on spherical surfaces. Based on SMConv, we establish spherical manifold encoder (SME) and spherical manifold decoder (SMD) blocks, constituting a spherical manifold U-Net (SMUNet). Our SMGD model is subsequently constructed by employing the SMUNet as the denoising network, with the VQGAN [11] serving as its image encoder. Unlike previous spherical convolutions [8, 36, 37, 45, 55] applied in classification and detection tasks, our SMConv is supported by the intrinsic manifold theory. Guided by exponential mapping, our SMConv incorporates the convolution within spherical neighborhoods under the measure of the geodesic distance [40]. Moreover, to further address the spherical geometry of panoramic images, we develop a spherical loss to optimize the proposed SMGD model. We also adopt an enhanced strategy to evaluate the quality of generated panoramic images by calculating the grouped Fréchet inception distance (FID) [17] in cubemap projection (CMP) [52] format. Experimental results demonstrate that our method achieves superior generation performances by effectively addressing the unique challenges associated with panoramic" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.907, + 0.137 + ], + "angle": 0, + "content": "image generation, such as complex scene structure, spherical distortions caused by ERP and the sparse pixel distribution near the poles. In summary, our contributions are three-fold:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.139, + 0.907, + 0.199 + ], + "angle": 0, + "content": "- We set out the first attempt to propose the spherical convolution, namely, SMConv, operating on \\( S^2 \\) manifold, which is able to optimally capture the intrinsic spherical geometry in panoramic images." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.2, + 0.907, + 0.259 + ], + "angle": 0, + "content": "- We propose to incorporate SMConv into spherical manifold guided blocks, including the SME and SMD blocks, so as to mitigate spherical distortions and preserve spatial coherence across the spherical domain." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.26, + 0.907, + 0.335 + ], + "angle": 0, + "content": "- Building upon these foundational blocks, we propose the SMGD model as the first generative approach to incorporate spherical convolution and achieve the state-of-the-art performance in generating panoramic images from text descriptions." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.139, + 0.907, + 0.335 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.347, + 0.665, + 0.364 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.373, + 0.909, + 0.827 + ], + "angle": 0, + "content": "Panoramic Image Generation. Early research on panoramic image generation can be achieved by the image outpainting task, which generates panoramic images from partial view [1, 2, 9, 15, 28, 38, 42]. However, these outpainting methods are limited by the freedom of control over the generated content. Most recently, with the rapid advancements in generative models, text-conditioned panoramic image generation [3, 5, 23, 24, 39, 41, 43, 53, 54] has received increasing research efforts. More specifically, Text2Light [5] first proposed to generate panoramic images from text descriptions, using CLIP [30] to encode text together with VQGAN [11] to encode panoramic images. To generate images at arbitrary sizes, DiffCollage [54], Multi-Diffusion [3] and SyncDiffusion [23] aggregated the intermediate outputs of pre-trained diffusion models. PanoGen [24] proposed to recursively outpaint the images generated by pretrained models to create panoramic images. However, those methods either fail to handle spherical distortions nor to ensure continuity between the left and right boundaries. To address this problem, MVDiffusion [39] generated perspective images and stitches them into panoramic images. However, the stitched panoramic image may witness its limitation in the field of view. PanFusion [53] introduced an attention mechanism for perspective images to enhance the generation of local details, while the main branch focused on generating the overall panoramic image. However, PanFusion utilizes two pretrained diffusion models [18] whereby multiple perspective images are processed in a separate branch; this can lead to unnatural global structures and result in heavy computational complexity." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.827, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Convolution Operation for Panoramic Images. Several attempts have introduced modifications to the convolution operation [7, 8, 12, 21, 29, 36, 37, 45, 55] in order to effectively manage spherical distortions for panoramic image classification and segmentation tasks. The pioneering work [36]" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5825" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.351 + ], + "angle": 0, + "content": "adopted kernels with different shapes at different latitudes in object detection task. Inspired by these works, deformable kernels [12, 55] were adopted to handle the spherical geometry. On the other hand, Coors et al. [8] retained the kernel unchanged on the tangent plane of the sphere and projected it back to the sphere using inverse gnomonic projection in image classification task. For the same task, Xu et al. [45] proposed to enlarge the receptive filed of convolution kernel by defining the kernel area in sphere. Furthermore, the rotation invariance property has been considered for designing spherical convolution in image classification [7] and image segmentation tasks [29]. However, spherical convolution operations have not yet been proposed for generative models that essentially suffer from improper spherical geometry. Therefore, we propose to incorporate the guidance from SMConv by exponential mapping, operating on \\( S^2 \\) manifold for high quality panoramic image generation." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.367, + 0.226, + 0.384 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.393, + 0.375, + 0.409 + ], + "angle": 0, + "content": "3.1. Spherical Manifold Convolution" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.416, + 0.485, + 0.794 + ], + "angle": 0, + "content": "The spherical distortion introduced by ERP leads to object deformation and interrupts the continuity of content on the sphere, which together significantly hinder the correct information aggregation of standard convolution. In contrast, performing convolution operation directly on the spherical domain is of significant potential to preserve the spherical geometry of panoramic images and to mitigate spherical distortions. By inspecting that panoramic images can be viewed as pixels distributed on a spherical surface, i.e., the \\( S^2 \\) manifold, we propose the SMConv operation that performs convolution on the \\( S^2 \\) manifold. More specifically, we employ exponential mapping [13] to project uniform convolution kernels from the tangent space onto the \\( S^2 \\) manifold, enabling SMConv with kernels to uniformly aggregate the content across the \\( S^2 \\) manifold. More importantly, due to the length-preserving property of exponential mapping [13], SMConv retains the uniformity of geodesic distances, known as great-circle distances [40], on the \\( S^2 \\) manifold. This enables SMConv to better handle spherical distortions, providing an effective solution for panoramic image generation. Compared to previous spherical convolution based on inverse gnomonic projection [8], employing exponential mapping to construct SMConv kernels enables an expanded receptive field, refined kernel patterns, and improved preservation of spherical geometry, as demonstrated in Fig. 2." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.484, + 0.904 + ], + "angle": 0, + "content": "More specifically, the \\(S^2\\) manifold is defined as a set of points \\(x \\in \\mathbb{R}^3\\) with norm 1, which can be parameterized by spherical coordinates \\(\\theta \\in [0,2\\pi]\\) and \\(\\varphi \\in [-\\pi/2, \\pi/2]\\). Through a transformation between spherical and Cartesian coordinates, we can obtain 3D Cartesian coordinates and then apply the exponential mapping. For instance, a point \\(\\mathbf{p} = (\\theta, \\varphi)\\) on \\(S^2\\) manifold can be converted to a corre" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.087, + 0.905, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.253, + 0.907, + 0.352 + ], + "angle": 0, + "content": "Figure 2. Comparison of exponential mapping (denoted as \\(\\mathrm{Exp}_{\\mathbf{p}}(\\cdot))\\) and inverse gnomonic projection (denoted as \\(\\mathrm{IGP}(\\cdot))\\) [8] through forward and backward mappings. Left: visualization of a point \\(\\mathbf{v}\\) in the tangent space \\(T_{\\mathbf{p}}S^{2}\\) at point \\(\\mathbf{p}\\), mapped onto the sphere. Right: projection of a panoramic image, the spherical data, onto the tangent space \\(T_{\\mathbf{p}}S^{2}\\) centered at \\(\\mathbf{p} = (\\theta, \\varphi)\\) whereby \\(\\theta = 0\\) and \\(\\varphi = \\frac{\\pi}{4}\\)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.357, + 0.909, + 0.387 + ], + "angle": 0, + "content": "sponding point \\(\\mathbf{p} = (x,y,z)\\) with Cartesian coordinates. The exponential mapping [13] can be formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.554, + 0.395, + 0.907, + 0.426 + ], + "angle": 0, + "content": "\\[\n\\mathbf {q} = \\operatorname {E x p} _ {\\mathbf {p}} (\\mathbf {v}) = \\cos (| | \\mathbf {v} | |) \\mathbf {p} + \\sin (| | \\mathbf {v} | |) \\frac {\\mathbf {v}}{| | \\mathbf {v} | |}, \\quad (1)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.433, + 0.907, + 0.479 + ], + "angle": 0, + "content": "where \\(\\mathbf{q}\\) denotes the mapped point, \\(\\mathbf{p}\\) is the tangent point, and \\(\\mathbf{v}\\) is a vector representing the coordinates of the kernel points \\(\\mathbf{K}_{T_{\\mathbf{p}}}(m,n)\\) in tangent space \\(T_{\\mathbf{p}}S^{2}\\) centered at \\(\\mathbf{p}\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.48, + 0.909, + 0.525 + ], + "angle": 0, + "content": "In practice, given a predefined position \\((u_{m},v_{n})\\) of a kernel point \\(\\mathbf{K}_{T_{\\mathbf{p}}}(m,n)\\) in the tangent space \\(T_{\\mathbf{p}}S^{2}\\), the corresponding vector \\(\\mathbf{v}\\) in \\(T_{\\mathbf{p}}S^{2}\\) can be obtained as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.577, + 0.535, + 0.713, + 0.551 + ], + "angle": 0, + "content": "\\[\n\\mathbf {v} = u _ {m} \\hat {\\mathbf {e}} _ {\\theta} + v _ {n} \\hat {\\mathbf {e}} _ {\\varphi},\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.573, + 0.554, + 0.907, + 0.57 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {e}} _ {\\theta} = (- \\sin \\theta , \\cos \\theta , 0), \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.572, + 0.573, + 0.849, + 0.589 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {e}} _ {\\varphi} = (- \\sin \\varphi \\cos \\theta , - \\sin \\varphi \\sin \\theta , \\cos \\varphi),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.6, + 0.907, + 0.81 + ], + "angle": 0, + "content": "where \\(\\mathbf{p} = (\\theta, \\varphi)\\) denotes the center point in \\(T_{\\mathbf{p}}S^2\\) and the vectors \\(\\hat{\\mathbf{e}}_{\\theta}, \\hat{\\mathbf{e}}_{\\varphi}\\) serve as the basis for the coordinate system in \\(T_{\\mathbf{p}}S^2\\). Note that the shape and the position of the convolution kernel are provided by \\((u_m, v_n)\\) in \\(T_{\\mathbf{p}}S^2\\) and we can calculate the position \\(\\mathbf{q} = \\mathrm{Exp}_{\\mathbf{p}}(\\mathbf{v}) = (\\hat{x}_{m,n}, \\hat{y}_{m,n}, \\hat{z}_{m,n})\\) of the kernel point \\(\\mathbf{K}(m,n)\\) on \\(S^2\\) manifold in Cartesian coordinates, via (1) and (2). Here, the kernel \\(\\mathbf{K}\\) is centered at the point \\(\\mathbf{p}\\) on \\(S^2\\) manifold. Due to the properties of the exponential mapping [13], the norm of the mapped point \\(\\mathbf{q} = (\\hat{x}_{m,n}, \\hat{y}_{m,n}, \\hat{z}_{m,n})\\) is guaranteed to be 1, ensuring that the point remains on the \\(S^2\\) manifold. Since we need the position \\(\\mathbf{q} = (\\theta_m, \\varphi_n)\\) in spherical coordinates for performing SMConv, we can simple apply the inverse transformation from Cartesian to spherical coordinates." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.909, + 0.843 + ], + "angle": 0, + "content": "Furthermore, suppose \\(\\widetilde{\\mathbf{F}}_{S^2},\\mathbf{F}_{S^2}\\) are features on \\(S^2\\) manifold. The operation of SMConv can be formulated by" + }, + { + "type": "equation", + "bbox": [ + 0.53, + 0.85, + 0.907, + 0.885 + ], + "angle": 0, + "content": "\\[\n\\widetilde {\\mathbf {F}} _ {S ^ {2}} (\\theta , \\varphi) = \\sum_ {m, n} \\mathbf {F} _ {S ^ {2}} (\\theta + \\theta_ {m}, \\varphi + \\varphi_ {n}) \\cdot \\mathbf {K} (m, n), \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.539, + 0.887, + 0.786, + 0.905 + ], + "angle": 0, + "content": "\\[\n\\left(\\theta_ {m}, \\varphi_ {n}\\right) = \\mathbf {q} = \\operatorname {E x p} _ {\\mathbf {p}} \\left(\\mathbf {K} _ {T _ {\\mathbf {p}}} (m, n)\\right),\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5826" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.087, + 0.902, + 0.351 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.37, + 0.908, + 0.414 + ], + "angle": 0, + "content": "Figure 3. Overall architecture of the proposed SMGD model, which introduces SMUNet as the denoising network and a panoramic VQGAN as the image encoder. The SMUNet is primarily constructed by the SMConv, operating on the \\(S^2\\) manifold, and integrates the SPB along with circular encoder and decoder blocks." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.42, + 0.486, + 0.571 + ], + "angle": 0, + "content": "In (3), \\(\\mathbf{q}\\) is the position of the kernel point \\(\\mathbf{K}(m,n)\\) for spherical features in ERP format, which is obtained by performing exponential mapping for the kernel point \\(\\mathbf{K}_{T_{\\mathbf{p}}}(m,n)\\) in the tangent space \\(T_{\\mathbf{p}}S^{2}\\). Moreover, \\(\\mathbf{K}\\) represents the kernel centered at point \\(\\mathbf{p}\\) on \\(S^2\\) manifold and the weight of \\(\\mathbf{K}\\) is identical to that of \\(\\mathbf{K}_{T_{\\mathbf{p}}}\\). By uniformly choosing \\(u_{m}\\) and \\(v_{m}\\) and constructing a regular kernel in the tangent space \\(T_{\\mathbf{p}}S^{2}\\), we can perform SMConv with kernels uniformly distributed across the \\(S^2\\) manifold using exponential mapping, according to (3)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.58, + 0.395, + 0.596 + ], + "angle": 0, + "content": "3.2. SMConv Inspired Diffusion Model" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.601, + 0.487, + 0.663 + ], + "angle": 0, + "content": "Building upon SMConv, we introduce the SMGD model to achieve high-quality panoramic image generation. The overall architecture of our SMGD model is shown in Fig. 3. The diffusion process can be formulated as" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.671, + 0.393, + 0.691 + ], + "angle": 0, + "content": "\\[\nq (\\mathbf {z} _ {t} | \\mathbf {z} _ {t - 1}) := \\mathcal {N} (\\mathbf {z} _ {t - 1}; \\sqrt {1 - \\beta_ {t}} \\mathbf {z} _ {t - 1}, \\beta_ {t} \\mathbf {I}),\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.692, + 0.484, + 0.729 + ], + "angle": 0, + "content": "\\[\n\\mathbf {z} _ {t - 1} = \\frac {1}{\\sqrt {\\alpha_ {t}}} \\left(\\mathbf {z} _ {t} - \\frac {1 - \\alpha_ {t}}{\\sqrt {1 - \\bar {\\alpha} _ {t}}} \\epsilon_ {\\psi} (\\mathbf {z} _ {t}, t, \\mathbf {T})\\right) + \\sqrt {\\beta_ {t}} \\eta , ^ {(4)}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.487, + 0.902 + ], + "angle": 0, + "content": "where \\(\\mathbf{z}_{t - 1},\\mathbf{z}_t\\) denote the noisy spherical features with the same dimensionality as the data, at time step \\(t - 1\\in [0,T]\\) and \\(t\\in [0,T]\\) respectively. \\(\\mathbf{T}\\) is the text embedding obtained by CLIP model [31], i.e., from the pretrained CLIP ViT/L-14 model. In addition, \\(q(\\mathbf{z}_t|\\mathbf{z}_{t - 1})\\) represents the forward process gradually adding noise to the image, \\(\\epsilon_{\\psi}(\\mathbf{z}_t,t)\\) is the noise predicted by our SMUNet at time step \\(t\\), \\(\\psi\\) represents the parameters of SMUNet and \\(\\eta\\) denotes a random Gaussian noise. Furthermore, the noise scheduler configuration \\(\\beta_{t}\\) is a hyperparameter that controls the strength of the noise involved at each time step. The parameter \\(\\alpha_{t}\\) is defined as" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.419, + 0.907, + 0.451 + ], + "angle": 0, + "content": "\\(\\alpha_{t} = 1 - \\beta_{t}\\), and \\(\\bar{\\alpha}_{t}\\) is given by \\(\\bar{\\alpha}_{t} = \\prod_{s=1}^{t} \\alpha_{s}\\), the default setting within diffusion models." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.451, + 0.909, + 0.782 + ], + "angle": 0, + "content": "As the core component of the SMGD model, our SMUNet predicts noise at each time step \\( t \\) from the noisy spherical feature \\( \\mathbf{z}_t \\) and facilitates the recovery of the spherical feature \\( \\mathbf{z}_0 \\) on \\( S^2 \\) manifold by subtracting the predicted noise. We employ spherical processor blocks (SPBs), along with circular encoder and decoder blocks based on circular convolution [2] at shallower layers of SMUNet. Then, we apply SME and SMD blocks at deeper layers with middle blocks consisted of spherical manifold residual blocks (SMRBs) and cross-attention layers to integrate text embeddings \\( \\mathbf{T} \\). The design of a hybrid architecture to integrate circular convolution at shallower layers with SMConv at deeper layers is motivated by the following considerations. As resolution increases, spherical convolution operation including SMConv, consumes substantially higher GPU memory cost compared to standard convolution [8] due to its bilinear interpolation. Furthermore, in deeper layers of SMUNet where spatial resolution is reduced, the sparse pixel distribution near the poles becomes increasingly pronounced, exacerbating the limitations of standard convolution. Consequently, the proposed hybrid architecture well balances between output quality, computational efficiency, and memory utilization." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.782, + 0.91, + 0.843 + ], + "angle": 0, + "content": "Then, the SME block comprises 2 SMRBs, 2 cross-attention layers, and an SMConv with a stride of 2 for downsampling. For the \\(l\\)-th SME block, the SMRB can be formulated as" + }, + { + "type": "equation", + "bbox": [ + 0.549, + 0.847, + 0.907, + 0.868 + ], + "angle": 0, + "content": "\\[\n\\widetilde {\\mathbf {z}} _ {t} ^ {E _ {l}} = \\mathrm {S M C o n v} [ \\mathrm {S M C o n v} [ \\mathbf {z} _ {t} ^ {E _ {l}} ] + g _ {\\psi} (t) ] + \\mathbf {z} _ {t} ^ {E _ {l}}, \\quad (5)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.91, + 0.902 + ], + "angle": 0, + "content": "where the calculation of SMConv follows (3) and \\(g_{\\psi}(t)\\) is the learnable time-step embedding function [32]. Moreover," + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5827" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.089, + 0.485, + 0.351 + ], + "angle": 0, + "content": "\\(\\mathbf{z}_t^{E_l}\\) denotes the input, while \\(\\widetilde{\\mathbf{z}}_t^{E_l}\\) represents the output for this layer. The structure of the SMD block closely mirrors that of the indexing SME block, with the upsampling layer using nearest-neighbor interpolation. Note that the final SME block and the first SMD block omit upsampling and downsampling operations. Moreover, the SPB employs the SMConv combined with a \\(1\\times 1\\) residual convolutional layer to enhance the transformation between different feature channels. The circular encoder and decoder blocks follow the structure in LDM [32], with all planar convolutions replaced by circular convolution operations [2]. Notably, the circular encoder blocks use average pooling layers for downsampling. In addition to the SMUNet, we propose to employ a panoramic VQGAN [11] with a downsampling factor of 16 as the image encoder and decoder for our SMGD model. In the panoramic VQGAN, all planar convolutions in the layers of decoder are replaced by circular convolutions [2]." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.357, + 0.403, + 0.373 + ], + "angle": 0, + "content": "3.3. Spherical Guided Training Strategy" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.379, + 0.486, + 0.468 + ], + "angle": 0, + "content": "The training process involves two stages: training the panoramic VQGAN, and then training the SMGD model. During the training of the panoramic VQGAN, reconstruction loss \\(\\mathcal{L}_{\\mathrm{rec}}\\) [11], commitment loss \\(\\mathcal{L}_{\\mathrm{Commit}}\\) [11] and perceptual loss \\(\\mathcal{L}_{\\mathrm{Per}}\\) [19] are employed, which can be formulated as" + }, + { + "type": "equation", + "bbox": [ + 0.111, + 0.48, + 0.484, + 0.497 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {V Q G A N}} = \\mathcal {L} _ {\\text {r e c}} + \\lambda_ {\\text {c o m m i t}} \\mathcal {L} _ {\\text {c o m m i t}} + \\lambda_ {\\text {p e r}} \\mathcal {L} _ {\\text {P e r}}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.506, + 0.483, + 0.536 + ], + "angle": 0, + "content": "where \\(\\lambda_{\\mathrm{commit}}\\) and \\(\\lambda_{\\mathrm{per}}\\) are hyperparameters to control the weights of different loss components during training." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.536, + 0.484, + 0.594 + ], + "angle": 0, + "content": "During the training of the SMGD model, a spherical loss based on the weighted spherically uniform PSNR (WSPSNR) [50] is employed. The total loss can be formulated as" + }, + { + "type": "equation", + "bbox": [ + 0.152, + 0.596, + 0.484, + 0.612 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {D i f f}} = \\lambda_ {\\text {M S E}} \\mathcal {L} _ {\\text {M S E}} + \\lambda_ {\\text {S M S E}} \\mathcal {L} _ {\\text {S M S E}}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.618, + 0.484, + 0.648 + ], + "angle": 0, + "content": "where \\(\\lambda_{\\mathrm{MSE}}\\) and \\(\\lambda_{\\mathrm{SMSE}}\\) are also balancing hyperparameters. Furthermore, the \\(\\mathcal{L}_{\\mathrm{MSE}}\\) and \\(\\mathcal{L}_{\\mathrm{SMSE}}\\) are given by" + }, + { + "type": "equation", + "bbox": [ + 0.138, + 0.657, + 0.484, + 0.681 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {M S E}} = \\mathbb {E} _ {t, \\mathbf {z} _ {0}, \\epsilon} [ | | \\epsilon - \\epsilon_ {\\psi} (\\mathbf {z} _ {t}, t, \\mathbf {T}) | | _ {2} ^ {2} ], \\tag {8}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.134, + 0.678, + 0.442, + 0.695 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {S M S E}} = \\mathbb {E} _ {t, \\mathbf {z} _ {0}, \\epsilon} [ | | \\mathbf {W} \\odot (\\epsilon - \\epsilon_ {\\psi} (\\mathbf {z} _ {t}, t, \\mathbf {T})) | | _ {2} ^ {2} ],\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.703, + 0.483, + 0.734 + ], + "angle": 0, + "content": "where \\(\\mathbf{W}\\) is given by the WS-PSNR [50] to reflect the panoramic characteristics as" + }, + { + "type": "equation", + "bbox": [ + 0.183, + 0.742, + 0.484, + 0.772 + ], + "angle": 0, + "content": "\\[\n\\mathbf {W} _ {j} = \\cos \\left(\\left(\\frac {j + 0 . 5}{H} - 0. 5\\right) \\pi\\right), \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.484, + 0.84 + ], + "angle": 0, + "content": "where \\( j \\in [0, H] \\) is the index of height in \\( \\mathbf{z}_t \\) and \\( H \\) denotes the height of the spherical feature \\( \\mathbf{z}_t \\). \\( \\mathcal{L}_{\\mathrm{SMSE}} \\) can improve the preservation of spherical geometry, e.g., a sparse pixel distribution near the poles." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.849, + 0.449, + 0.866 + ], + "angle": 0, + "content": "3.4. Quality Evaluation for Panoramic Images" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.901 + ], + "angle": 0, + "content": "FID score is the de facto standard for evaluating the quality of generated images, whereas performing a direct calculation" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.092, + 0.905, + 0.193 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.203, + 0.907, + 0.273 + ], + "angle": 0, + "content": "Figure 4. Illustration of the conversion between ERP and CMP formats. In the CMP format, four blue-bordered images represent equatorial regions, while two orange-bordered images represent polar areas. These images are grouped accordingly, with \\(\\mathrm{FID}_{\\mathrm{equ}}\\) and \\(\\mathrm{FID}_{\\mathrm{pole}}\\) calculated separately for each group." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.28, + 0.909, + 0.644 + ], + "angle": 0, + "content": "of FID on panoramic images in rectangular ERP format introduces severe distortion, leading to inaccuracy. We propose to convert the generated panoramic images from ERP format to CMP format consisted of 6 square images for calculating FID score. The mapping from ERP to CMP format [50, 52] provides a robust solution to evaluate the quality of generated panoramic images. Note that although sharing similar goals with the concurrent work [6], our FID-based metrics are able to reflect detailed quality for different regions when evaluating generated panoramic images. As illustrated in Fig. 4, the conversion reveals the continuity across the left and right boundaries and represents the spherical geometry by obtaining 4 images representing areas near the equator and 2 images representing areas near the poles. Thus, the 6 square images within CMP format can be categorized into 2 groups, allowing for separate FID calculations to assess the quality of generated images. More specifically, \\(\\mathrm{FID}_{\\mathrm{equ}}\\) is calculated using the 4 images near the equator, while \\(\\mathrm{FID}_{\\mathrm{pole}}\\) is derived from the 2 images near the poles. \\(\\mathrm{FID}_{\\mathrm{equ}}\\) primarily assesses the quality and visual fidelity of the panoramic content, whereas \\(\\mathrm{FID}_{\\mathrm{pole}}\\) evaluates the preservation of spherical geometry. Analyzing \\(\\mathrm{FID}_{\\mathrm{equ}}\\) and \\(\\mathrm{FID}_{\\mathrm{pole}}\\) provides a comprehensive measure of the quality of the generated panoramic image." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.658, + 0.638, + 0.674 + ], + "angle": 0, + "content": "4. Experiment" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.683, + 0.72, + 0.699 + ], + "angle": 0, + "content": "4.1. Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.704, + 0.909, + 0.84 + ], + "angle": 0, + "content": "Dataset. We employed the widely-used Matterport3D [4] dataset to evaluate the performance for text-conditioned panoramic image generation. More specifically, we obtained 10,912 panoramic images with a resolution of \\(1024 \\times 512\\) according to [26]. We then utilized the BLIP-2 [25] model to generate text descriptions for each panoramic image, forming paired text-image data for training and testing. To enable FID-10K calculation, we allocated 10,000 images for training and testing." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.909, + 0.9 + ], + "angle": 0, + "content": "Implementation Details. We adopted the panoramic VQGAN according to Sec. 3.2. SMUNet, the denoising network for our SMGD model, was configured with 320 and 640 channels for the two circular encoder and decoder blocks," + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "5828" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.083, + 0.907, + 0.127 + ], + "angle": 0, + "content": "Table 1. Comparisons in terms of FID and CLIP score (denoted as CS) values for panoramic images generated by our method and existing baseline methods, along with evaluations of parameter number (denoted by Para.) and time required to inference (denoted as \\( t_{\\mathrm{sample}} \\)). We represent the best numbers by red color and the second best by blue color." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.136, + 0.905, + 0.233 + ], + "angle": 0, + "content": "
MethodsFID ↓FAED ↓OmniFID ↓FIDavg ↓FIDcent ↓FIDbord ↓FIDrand ↓FIDequ ↓FIDpole ↓CS ↑Para. ↓t_sample(s) ↓
Multi-Diffusion [3]69.961.1268.7056.7458.1956.5355.5027.2687.730.1821.3B39.45 ± 0.31
SD2+LoRA [18, 32]59.321.4769.0750.0451.7849.1649.1724.4988.680.1711.4B6.48 ± 0.49
Text2Light [5]48.424.1063.1643.6846.9346.6437.4730.8772.060.1881.0B50.12 ± 11.18
MVDiffusion [39]56.210.9394.5338.1538.8837.6037.9627.58125.480.1811.4B97.14 ± 0.79
PanFusion [53]16.150.3527.6414.3414.9113.9614.1414.8131.990.1822.3B38.51 ± 11.57
SMGD (Ours)9.850.2218.249.709.779.629.717.8821.050.1891.4B2.69 ± 0.39
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.257, + 0.486, + 0.44 + ], + "angle": 0, + "content": "and 960 and 1280 channels for the two SME and SMD blocks utilizing the \\(3 \\times 3\\) SMConv. At the first training stage, we trained the panoramic VQGAN with the Adam optimizer [22], using a batch size of 1 and a learning rate of \\(4.5 \\times 10^{-6}\\) for 300 epochs, with \\(\\lambda_{\\mathrm{commit}} = 1\\) and \\(\\lambda_{\\mathrm{per}} = 0.8\\). At the second stage, we trained the SMGD model from the scratch with the AdamW optimizer [27], a batch size of 8, and learning rate of \\(1 \\times 10^{-5}\\) for 800 epochs by setting \\(\\lambda_{\\mathrm{MSE}} = 0.8\\) and \\(\\lambda_{\\mathrm{SMSE}} = 0.2\\). Moreover, the noise scheduler configuration \\(\\beta_{t}\\) followed that of LDM [32]. For inference, we used a DDIM sampler [35] with 50 steps and a classifier-free guidance scale [32] of 2.5." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.451, + 0.487, + 0.587 + ], + "angle": 0, + "content": "Baselines. We compared our SMGD model with existing text-conditioned panoramic image generation methods, including Multi-Diffusion [3], Text2Light [5], MVDiffusion [39] and PanFusion [53]. We also compared our SMGD model with the widely-used stable diffusion model [32] by performing a rank-16 LoRA [18] (denoted as SD2+LoRA). Different from [53], we directly trained the LoRA model with a learning rate of \\( 1 \\times 10^{-6} \\) for 50 epochs on our training data." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.599, + 0.486, + 0.902 + ], + "angle": 0, + "content": "Evaluation Metrics. We computed \\(\\mathrm{FID}_{\\mathrm{equ}}\\) and \\(\\mathrm{FID}_{\\mathrm{pole}}\\), as described in Sec. 3.4 to effectively assess the quality of panoramic images. Note that \\(\\mathrm{FID}_{\\mathrm{equ}}\\) was calculated against 40K images and \\(\\mathrm{FID}_{\\mathrm{pole}}\\) was calculated against 20K images. To further validate our metrics and assess the quality of the generated images, we calculated an extra FID value similar to that used in previous methods [53]. More specifically, we cropped the panoramic images into 3 representative areas instead of obtaining 20 perspective images in calculating FID values. Specifically, we cropped the panoramic images into the center, cross-border and random patches with size \\(512 \\times 512\\). Then, we calculated the corresponding FID value of them and obtained the \\(\\mathrm{FID}_{\\mathrm{cent}}\\), \\(\\mathrm{FID}_{\\mathrm{bord}}\\) and \\(\\mathrm{FID}_{\\mathrm{rand}}\\). By averaging these three FID values calculated against 10K images, we obtained \\(\\mathrm{FID}_{\\mathrm{avg}}\\). Moreover, we reported FID, OmniFID [6], and Fréchet Auto-Encoder Distance (FAED) [53] to comprehensively evaluate the quality of the generated full panoramic images. On the other hand, to evaluate the alignment between text and generated panoramic images, we utilized the widely adopted CLIP score [16] (denoted" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.257, + 0.907, + 0.303 + ], + "angle": 0, + "content": "as CS).1 Furthermore, we recorded the time for inference \\( t_{\\mathrm{sample}} \\) by averaging the time required to generate 10 images on an NVIDIA 4090 GPU for each method." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.311, + 0.822, + 0.327 + ], + "angle": 0, + "content": "4.2. Comparison with Existing Methods" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.333, + 0.909, + 0.68 + ], + "angle": 0, + "content": "Quantitative Results. In Table 1, we report the quantitative results of our method and baselines for panoramic image generation from text descriptions. From this table, our method consistently achieves the lowest FID values, the highest CLIP scores, and the least sampling times. More importantly, among the top-performance models, our method maintains a comparatively lower parameter number. More specifically, the lower values of \\(\\mathrm{FID}_{\\mathrm{cent}}\\), \\(\\mathrm{FID}_{\\mathrm{bord}}\\), \\(\\mathrm{FID}_{\\mathrm{rand}}\\), and \\(\\mathrm{FID}_{\\mathrm{avg}}\\) demonstrate that our SMGD model achieves the state-of-the-art performance within text-conditioned panoramic image generation2. Moreover, the even lower \\(\\mathrm{FID}_{\\mathrm{equu}}\\) and \\(\\mathrm{FID}_{\\mathrm{pole}}\\) values indicate that our method well preserves the spherical geometry of panoramic images, compared to other baseline methods. Note that MVDiffusion [39] and SD2+LoRA achieved low \\(\\mathrm{FID}_{\\mathrm{avg}}\\) values but high \\(\\mathrm{FID}_{\\mathrm{pole}}\\) values. This may be due to the fact that images generated by MVDiffusion exhibited a limited field of view, while images produced by SD2+LoRA lacked the preservation of the spherical geometry in panoramic images. Furthermore, as shown in Table 1, our SMGD model achieves the fastest inference time among all comparison models, outperforming PanFusion [53], the second-best method in terms of image quality, by more than a factor of 10 in inference speed." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.68, + 0.911, + 0.832 + ], + "angle": 0, + "content": "Qualitative Results. We randomly chose 2 text descriptions and report the generated panoramic images with corresponding real images from dataset in Fig. 5. As illustrated in the figure, the panoramic images generated by our SMGD model achieved the best subjective quality and the most accurate alignment with the text descriptions. More specifically, improper spherical geometry, including inconsistency across the border and an unnatural pixel distribution near the poles, was observed in panoramic images generated by SD2+LoRA, Text2Light, and Multi-Diffusion, due to their" + }, + { + "type": "ref_text", + "bbox": [ + 0.513, + 0.838, + 0.908, + 0.864 + ], + "angle": 0, + "content": "1The CLIP score was computed using the ViT-L/14 CLIP model, consistent with the CLIP model employed in our SMGD model." + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.865, + 0.908, + 0.901 + ], + "angle": 0, + "content": "To further validate the robustness of our model, we conduct additional evaluations on the Structured3D [56] dataset in Supplementary Sec. A. We also provide additional qualitative results in Supplementary Sec. C." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.838, + 0.908, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "5829" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.147, + 0.09, + 0.839, + 0.75 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.758, + 0.908, + 0.827 + ], + "angle": 0, + "content": "Figure 5. Qualitative results for text-conditioned panoramic image generation with real images from dataset displayed on the top row. We also show the copy from the left side of the panoramic image to the right to reveal the consistency across the whole image. Note that images generated by MVDiffusion [39] exhibit a limited field of view, resulting in images with a reduced height. We highlight the improper spherical geometry, artifact regions and distorted objects and lines with corresponding color of boxes, which are mitigated by our SMGD model." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.836, + 0.486, + 0.898 + ], + "angle": 0, + "content": "neglect of spherical distortions. Panoramic images generated by Panfusion and MVDiffusion suffered from artifact regions and distorted objects and lines as highlighted with corresponding color boxes in Fig. 5. Moreover, images" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.836, + 0.91, + 0.898 + ], + "angle": 0, + "content": "generated by MVDiffusion exhibited a limited field of view. This limitation impeded the visual quality and completeness of the generated images, making them less suitable for fully immersive visual experiences. In contrast, our SMGD" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "5830" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.071, + 0.908, + 0.113 + ], + "angle": 0, + "content": "Table 2. Ablation study for our SMGD model. We compare the performance achieved by adding individual components to the baseline model sequentially, including the spherical loss (denoted by \\( \\mathcal{L}_{\\mathrm{SMSE}} \\)), the SMD blocks, the SME blocks, and the SPB modules. We represent the best numbers by red color." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.125, + 0.905, + 0.226 + ], + "angle": 0, + "content": "
\\( \\mathcal{L}_{\\text{SMSE}} \\)SMD BlocksSME BlocksSPBFID\\( \\mathrm{FID}_{\\text{avg}} \\downarrow \\)\\( \\mathrm{FID}_{\\text{cent}} \\downarrow \\)\\( \\mathrm{FID}_{\\text{bord}} \\downarrow \\)\\( \\mathrm{FID}_{\\text{rand}} \\downarrow \\)\\( \\mathrm{FID}_{\\text{equ}} \\downarrow \\)\\( \\mathrm{FID}_{\\text{pole}} \\downarrow \\)CS ↑
××××23.1120.5720.3420.9120.4515.3534.870.187
✓×××20.0317.5017.0518.0417.4012.7832.950.187
✓✓××13.3512.8712.7912.9412.889.6120.450.188
✓✓✓×12.3611.4211.4611.6011.199.2121.700.186
✓✓✓✓9.859.709.779.629.717.8821.050.189
" + }, + { + "type": "image_caption", + "bbox": [ + 0.178, + 0.239, + 0.383, + 0.252 + ], + "angle": 0, + "content": "Comparison in terms of FID and CS" + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.254, + 0.476, + 0.412 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.416, + 0.293, + 0.614 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.296, + 0.419, + 0.48, + 0.614 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.624, + 0.486, + 0.707 + ], + "angle": 0, + "content": "Figure 6. Comparison between LDM+SphereNet and the proposed SMGD model. Top: Bar plot of FID and CLIP score values achieved by these two methods. Bottom: Subjective results of panoramic images generated by the text prompt \"a hallway in a house\". The comparison is conducted in both ERP and CMP formats, with distorted objects and lines highlighted." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.716, + 0.486, + 0.748 + ], + "angle": 0, + "content": "model produced visually pleasing panoramic images with a well-defined overall structure and proper spherical geometry." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.758, + 0.245, + 0.773 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.486, + 0.901 + ], + "angle": 0, + "content": "In this section, we conducted ablation study to systematically evaluate the impact of core components in our SMGD model. More specifically, since the spherical loss and SMConv are important for handling spherical distortions and preserving spherical geometry in panoramic images, we analyzed performance changes by gradually adding the spherical loss \\(\\mathcal{L}_{\\mathrm{SMSE}}\\), the SMD blocks, the SME blocks and the SPB. As reported in Table 2, the involvement of SMD blocks" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.239, + 0.908, + 0.328 + ], + "angle": 0, + "content": "significantly enhanced performance, with SME blocks further contributing to the improvements. Moreover, both the spherical loss function and the SPB effectively improved the quality of generated panoramic images. Therefore, each module plays a vital role in optimizing the performance of the SMGD model." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.33, + 0.909, + 0.526 + ], + "angle": 0, + "content": "Furthermore, to validate the effectiveness of SMConv compared to previous spherical convolution methods, we implemented LDM+SphereNet, which employs the panoramic VQGAN and a U-Net architecture identical to our SMUNet, whilst all SMConv layers were replaced by the convolution operations proposed in SphereNet [8]. The training and inference were kept consistent with those of our SMGD model. As shown in Fig. 6, we can conclude that the quality of panoramic images generated by LDM+SphereNet is limited, whereas our proposed SMGD model produces visually pleasing images with lower FID values and a higher CLIP score3. This demonstrates the effectiveness of using SMConv for advanced panoramic image generation." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.539, + 0.634, + 0.555 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.565, + 0.909, + 0.867 + ], + "angle": 0, + "content": "In this paper, we have proposed a novel method named as spherical manifold guided diffusion (SMGD) model to address the challenges of spherical distortions associated with panoramic image generation. This was achieved by the spherical manifold convolution (SMConv) operation that is able to uniformly sample spherical content within each convolution layer. Different from existing spherical convolution operations applied in other tasks, the proposed SMConv operation was built upon the exponential mapping grounded in manifold theory, which is able to preserve lengths on the tangent plane and exhibit enhanced local properties. The overall training procedure of our SMGD model was also inspired by the spherical loss. Furthermore, we have proposed to calculate \\(\\mathrm{FID}_{\\mathrm{equ}}\\) and \\(\\mathrm{FID}_{\\mathrm{pole}}\\) for panoramic images in cubemap format to improve the evaluation on the quality of generated panoramic images. Experimental results have demonstrated that our method achieves superior performance in text-conditioned image generation, surpassing existing state-of-the-art methods in both the quality of generated images and computational efficiency." + }, + { + "type": "page_footnote", + "bbox": [ + 0.514, + 0.875, + 0.908, + 0.901 + ], + "angle": 0, + "content": "3More qualitative comparison results between SMGD and LDM+SphereNet, as well as a detailed analysis, are provided in Supplementary Sec. B." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "5831" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.251, + 0.108 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.486, + 0.175 + ], + "angle": 0, + "content": "This work was supported by Natural Science Foundation of China (NSFC) under Grants 62450131, 62206011, 62231002 and 62401027, and Beijing Natural Science Foundation under Grant L223021." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.188, + 0.188, + 0.205 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.213, + 0.486, + 0.295 + ], + "angle": 0, + "content": "[1] Hao Ai, Zidong Cao, Haonan Lu, Chen Chen, Jian Ma, Pengyuan Zhou, Tae-Kyun Kim, Pan Hui, and Lin Wang. Dream360: Diverse and immersive outdoor virtual scene creation via transformer-based \\(360^{\\circ}\\) image outpainting. IEEE Transactions on Visualization and Computer Graphics, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.297, + 0.485, + 0.367 + ], + "angle": 0, + "content": "[2] Naofumi Akimoto, Yuhi Matsuo, and Yoshimitsu Aoki. Diverse plausible 360-degree image outpainting for efficient 3dcg background creation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11441-11450, 2022. 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.368, + 0.484, + 0.422 + ], + "angle": 0, + "content": "[3] Omer Bar-Tal, Lior Yariv, Yaron Lipman, and Tali Dekel. Multidiffusion: Fusing diffusion paths for controlled image generation. In International Conference on Machine Learning, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.424, + 0.484, + 0.491 + ], + "angle": 0, + "content": "[4] Angel Chang, Angela Dai, Thomas Funkhouser, Maciej Halber, Matthias Niessner, Manolis Savva, Shuran Song, Andy Zeng, and Yinda Zhang. Matterport3d: Learning from rgb-d data in indoor environments. arXiv preprint arXiv:1709.06158, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.493, + 0.485, + 0.535 + ], + "angle": 0, + "content": "[5] Zhaoxi Chen, Guangcong Wang, and Ziwei Liu. Text2light: Zero-shot text-driven hdr panorama generation. ACM Transactions on Graphics (TOG), 41(6):1-16, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.537, + 0.484, + 0.605 + ], + "angle": 0, + "content": "[6] Anders Christensen, Nooshin Mojab, Khushman Patel, Karan Ahuja, Zeynep Akata, Ole Winther, Mar Gonzalez-Franco, and Andrea Colaco. Geometry fidelity for spherical images. In European Conference on Computer Vision, pages 276-292. Springer, 2024. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.607, + 0.484, + 0.634 + ], + "angle": 0, + "content": "[7] Taco S Cohen, Mario Geiger, Jonas Köhler, and Max Welling. Spherical cnns. arXiv preprint arXiv:1801.10130, 2018. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.635, + 0.484, + 0.704 + ], + "angle": 0, + "content": "[8] Benjamin Coors, Alexandru Paul Condurache, and Andreas Geiger. Spherenet: Learning spherical representations for detection and classification in omnidirectional images. In Proceedings of the European conference on computer vision (ECCV), pages 518-533, 2018. 2, 3, 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.705, + 0.484, + 0.773 + ], + "angle": 0, + "content": "[9] Mohammad Reza Karimi Dastjerdi, Yannick Hold-Geoffroy, Jonathan Eisenmann, Siavash Khodadadeh, and Jean-François Lalonde. Guided co-modulated gan for 360 field of view extrapolation. In 2022 International Conference on 3D Vision (3DV), pages 475–485. IEEE, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.775, + 0.482, + 0.816 + ], + "angle": 0, + "content": "[10] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.817, + 0.484, + 0.873 + ], + "angle": 0, + "content": "[11] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12873-12883, 2021. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.874, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[12] Clara Fernandez-Labrador, Jose M Facil, Alejandro Perez-Yus, Cédric Demonceaux, Javier Civera, and Jose J Guerrero." + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.213, + 0.486, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.908, + 0.134 + ], + "angle": 0, + "content": "Corners for layout: End-to-end layout recovery from 360 images. IEEE Robotics and Automation Letters, 5(2):1255-1262, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.135, + 0.907, + 0.175 + ], + "angle": 0, + "content": "[13] Jean Gallier. Basics of Classical Lie Groups: The Exponential Map, Lie Groups, and Lie Algebras, pages 367-414. Springer New York, New York, NY, 2001. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.177, + 0.906, + 0.232 + ], + "angle": 0, + "content": "[14] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.233, + 0.907, + 0.288 + ], + "angle": 0, + "content": "[15] Takayuki Hara, Yusuke Mukuta, and Tatsuya Harada. Spherical image generation from a single image by considering scene symmetry. Proceedings of the AAAI Conference on Artificial Intelligence, 35(2):1513-1521, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.289, + 0.907, + 0.356 + ], + "angle": 0, + "content": "[16] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 7514-7528, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.358, + 0.907, + 0.425 + ], + "angle": 0, + "content": "[17] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.427, + 0.907, + 0.482 + ], + "angle": 0, + "content": "[18] Edward J Hu, Phillip Wallis, Zeyuan Allen-Zhu, Yuzhhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. In International Conference on Learning Representations, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.484, + 0.907, + 0.538 + ], + "angle": 0, + "content": "[19] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In European Conference on Computer Vision, pages 694–711. Springer International Publishing, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.539, + 0.906, + 0.607 + ], + "angle": 0, + "content": "[20] Minguk Kang, Jun-Yan Zhu, Richard Zhang, Jaesik Park, Eli Shechtman, Sylvain Paris, and Taesung Park. Scaling up gans for text-to-image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10124-10134, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.609, + 0.907, + 0.663 + ], + "angle": 0, + "content": "[21] Renata Khasanova and Pascal Frossard. Geometry aware convolutional filters for omnidirectional images representation. In International conference on machine learning, pages 3351-3359. PMLR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.665, + 0.907, + 0.692 + ], + "angle": 0, + "content": "[22] Diederik P Kingma. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.693, + 0.907, + 0.747 + ], + "angle": 0, + "content": "[23] Yuseung Lee, Kunho Kim, Hyunjin Kim, and Minhyuk Sung. Syncdiffusion: Coherent montage via synchronized joint diffusions. Advances in Neural Information Processing Systems, 36:50648-50660, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.748, + 0.907, + 0.802 + ], + "angle": 0, + "content": "[24] Jialu Li and Mohit Bansal. Panogen: Text-conditioned panoramic environment generation for vision-and-language navigation. Advances in Neural Information Processing Systems, 36, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.804, + 0.907, + 0.871 + ], + "angle": 0, + "content": "[25] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730-19742. PMLR, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.907, + 0.901 + ], + "angle": 0, + "content": "[26] Chieh Hubert Lin, Chia-Che Chang, Yu-Sheng Chen, Da-Cheng Juan, Wei Wei, and Hwann-Tzong Chen. Coco-gan:" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "5832" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.484, + 0.134 + ], + "angle": 0, + "content": "Generation by parts via conditional coordinating. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4512-4521, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.135, + 0.484, + 0.162 + ], + "angle": 0, + "content": "[27] I Loshchilov. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.163, + 0.485, + 0.232 + ], + "angle": 0, + "content": "[28] Zhuqiang Lu, Kun Hu, Chaoyue Wang, Lei Bai, and Zhiyong Wang. Autoregressive omni-aware outpainting for open-vocabulary 360-degree image generation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 14211-14219, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.233, + 0.484, + 0.287 + ], + "angle": 0, + "content": "[29] Thomas W Mitchel, Noam Aigerman, Vladimir G Kim, and Michael Kazhdan. Möbius convolutions for spherical cnns. In ACM SIGGRAPH 2022 Conference Proceedings, pages 1-9, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.288, + 0.484, + 0.37 + ], + "angle": 0, + "content": "[30] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.371, + 0.484, + 0.426 + ], + "angle": 0, + "content": "[31] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1(2): 3, 2022. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.428, + 0.484, + 0.497 + ], + "angle": 0, + "content": "[32] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684-10695, 2022. 1, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.498, + 0.484, + 0.58 + ], + "angle": 0, + "content": "[33] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in neural information processing systems, 35:36479-36494, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.581, + 0.484, + 0.649 + ], + "angle": 0, + "content": "[34] Axel Sauer, Tero Karras, Samuli Laine, Andreas Geiger, and Timo Aila. Stylegan-t: Unlocking the power of gans for fast large-scale text-to-image synthesis. In International conference on machine learning, pages 30105-30118. PMLR, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.65, + 0.484, + 0.691 + ], + "angle": 0, + "content": "[35] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.692, + 0.484, + 0.734 + ], + "angle": 0, + "content": "[36] Yu-Chuan Su and Kristen Grauman. Learning spherical convolution for fast features from 360 imagery. Advances in neural information processing systems, 30, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.735, + 0.484, + 0.79 + ], + "angle": 0, + "content": "[37] Yu-Chuan Su and Kristen Grauman. Kernel transformer networks for compact spherical convolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9442-9451, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.79, + 0.484, + 0.846 + ], + "angle": 0, + "content": "[38] Julius Surya Sumantri and In Kyu Park. 360 panorama synthesis from a sparse set of images with unknown field of view. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 2386-2395, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[39] Shitao Tang, Fuyang Zhang, Jiacheng Chen, Peng Wang, and Yasutaka Furukawa. Mvdiffusion: Enabling holistic multiview image generation with correspondence-aware diffusion. ArXiv, abs/2307.01097, 2023. 2, 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.485, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.907, + 0.134 + ], + "angle": 0, + "content": "[40] T. Vincenty. Direct and inverse solutions of geodesics on the ellipsoid with application of nested equations. Survey Review, 23(176):88-93, 1975. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.135, + 0.907, + 0.203 + ], + "angle": 0, + "content": "[41] Hai Wang, Xiaoyu Xiang, Yuchen Fan, and Jing-Hao Xue. Customizing 360-degree pancreas through text-to-image diffusion models. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 4933-4943, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.204, + 0.907, + 0.259 + ], + "angle": 0, + "content": "[42] Tianhao Wu, Chuanxia Zheng, and Tat-Jen Cham. Panodiffusion: 360-degree panorama outpainting via diffusion. In The Twelfth International Conference on Learning Representations, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.26, + 0.906, + 0.328 + ], + "angle": 0, + "content": "[43] Tao Wu, Xuewei Li, Zhongang Qi, Di Hu, Xintao Wang, Ying Shan, and Xi Li. Spherediffusion: Spherical geometry-aware distortion resilient diffusion model. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 6126-6134, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.329, + 0.907, + 0.385 + ], + "angle": 0, + "content": "[44] Mai Xu, Chen Li, Shanyi Zhang, and Patrick Le Callet. State-of-the-art in 360 video/image processing: Perception, assessment and compression. IEEE Journal of Selected Topics in Signal Processing, 14(1):5-26, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.386, + 0.906, + 0.441 + ], + "angle": 0, + "content": "[45] Yanyu Xu, Ziheng Zhang, and Shenghua Gao. Spherical dnns and their applications in 360 images and videos. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(10):7235-7252, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.442, + 0.907, + 0.496 + ], + "angle": 0, + "content": "[46] Jian-Ru Xue, Jian-Wu Fang, and Pu Zhang. A survey of scene understanding by event reasoning in autonomous driving. International Journal of Automation and Computing, 15 (3):249-266, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.497, + 0.907, + 0.567 + ], + "angle": 0, + "content": "[47] Bangbang Yang, Yinda Zhang, Yijin Li, Zhaopeng Cui, Sean Fanello, Hujun Bao, and Guofeng Zhang. Neural rendering in a room: amodal 3d understanding and free-viewpoint rendering for the closed scene composed of pre-captured objects. ACM Transactions on Graphics (TOG), 41(4):1-10, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.568, + 0.907, + 0.637 + ], + "angle": 0, + "content": "[48] Bangbang Yang, Wenqi Dong, Lin Ma, Wenbo Hu, Xiao Liu, Zhaopeng Cui, and Yuewen Ma. Dreamspace: Dreaming your room space with text-driven panoramic texture propagation. In 2024 IEEE Conference Virtual Reality and 3D User Interfaces (VR), pages 650-660. IEEE, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.637, + 0.907, + 0.704 + ], + "angle": 0, + "content": "[49] Michihiro Yasunaga, Armen Aghajanyan, Weijia Shi, Rich James, Jure Leskovec, Percy Liang, Mike Lewis, Luke Zettlemoyer, and Wen-tau Yih. Retrieval-augmented multimodal language modeling. arXiv preprint arXiv:2211.12561, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.706, + 0.907, + 0.761 + ], + "angle": 0, + "content": "[50] Yan Ye, Elena Alshina, and Jill Boyce. Algorithm descriptions of projection format conversion and video quality metrics in 360lib. In Joint Video Exploration Team of ITU-T SG, 2018. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.762, + 0.907, + 0.831 + ], + "angle": 0, + "content": "[51] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, et al. Scaling autoregressive models for content-rich text-to-image generation. arXiv preprint arXiv:2206.10789, 2(3):5, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.832, + 0.907, + 0.901 + ], + "angle": 0, + "content": "[52] Li Yu, Yanjun Gao, Farhad Pakdaman, and Moncef Gabbouj. Panoramic image inpainting with gated convolution and contextual reconstruction loss. In ICASSP 2024 - 2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 4255-4259, 2024. 2, 5" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "5833" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.174 + ], + "angle": 0, + "content": "[53] Cheng Zhang, Qianyi Wu, Camilo Cruz Gambardella, Xiaoshui Huang, Dinh Phung, Wanli Ouyang, and Jianfei Cai. Taming stable diffusion for text to 360 panorama image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6347-6357, 2024. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.178, + 0.483, + 0.245 + ], + "angle": 0, + "content": "[54] Qinsheng Zhang, Jiaming Song, Xun Huang, Yongxin Chen, and Ming-Yu Liu. Diffcollage: Parallel generation of large content with diffusion models. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10188-10198. IEEE, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.248, + 0.483, + 0.288 + ], + "angle": 0, + "content": "[55] Qiang Zhao, Chen Zhu, Feng Dai, Yike Ma, Guoqing Jin, and Yongdong Zhang. Distortion-aware cnns for spherical images. In IJCAI, pages 1198-1204, 2018. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.291, + 0.483, + 0.345 + ], + "angle": 0, + "content": "[56] Jia Zheng, Junfei Zhang, Jing Li, Rui Tang, Shenghua Gao, and Zihan Zhou. Structured3d: A large photo-realistic dataset for structured 3d modeling. In Proceedings of The European Conference on Computer Vision (ECCV), 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.348, + 0.483, + 0.429 + ], + "angle": 0, + "content": "[57] Yufan Zhou, Ruiyi Zhang, Changyou Chen, Chunyuan Li, Chris Tensmeyer, Tong Yu, Jiuming Gu, Jinhui Xu, and Tong Sun. Towards language-free training for text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17907-17917, 2022. 1" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.429 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.515, + 0.956 + ], + "angle": 0, + "content": "5834" + } + ] +] \ No newline at end of file diff --git a/2025/Spherical Manifold Guided Diffusion Model for Panoramic Image Generation/25742454-f26d-401b-a2ab-0a30d46da30a_origin.pdf b/2025/Spherical Manifold Guided Diffusion Model for Panoramic Image Generation/25742454-f26d-401b-a2ab-0a30d46da30a_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..fff38c773285696cf68185809bb534dc97b56313 --- /dev/null +++ b/2025/Spherical Manifold Guided Diffusion Model for Panoramic Image Generation/25742454-f26d-401b-a2ab-0a30d46da30a_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dc0ec7aa2adb9320356b52569d3b073de8dbcd927c8c3f8a5be292f2edbd4857 +size 9591171 diff --git a/2025/Spherical Manifold Guided Diffusion Model for Panoramic Image Generation/full.md b/2025/Spherical Manifold Guided Diffusion Model for Panoramic Image Generation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..c34c4dab6f13a485ce6df3704aa46d23db767377 --- /dev/null +++ b/2025/Spherical Manifold Guided Diffusion Model for Panoramic Image Generation/full.md @@ -0,0 +1,293 @@ +# Spherical Manifold Guided Diffusion Model for Panoramic Image Generation + +Xiancheng Sun $^{1}$ Mai Xu $^{1}$ Shengxi Li $^{1*}$ Senmao Ma $^{1}$ Xin Deng $^{1}$ Lai Jiang $^{1}$ Gang Shen $^{2}$ $^{1}$ Beihang University $^{2}$ China Tower + +![](images/77f9dff45c20491206b720effb9105058a844c3b356c70c6428895139d28aabd.jpg) +"A serene room with a large door leading to a balcony and a window showcasing the garden, decorated in soft pastels" +"A bright and airy staircase with large windows overlooking a manicured garden, enhancing the natural light that floods the space." +Figure 1. A subjective comparison of zero-shot text-conditioned panoramic image generation between our SMGD model and the state-of-the-art method PanFusion [53]. Our model demonstrates superior performance in generating images with superior overall natural content, enhanced subjective quality, and better alignment with the text description, while benefiting reduced model parameters. + +![](images/825af3b51b3a7cbcc0e3ca75b623a3ac4263186e086c56eb595e77c0c7ed070d.jpg) + +# Abstract + +Panoramic image essentially acts as a pivotal role in emerging virtual reality and augmented reality scenarios; however, the generation of panoramic images are essentially challenging due to the intrinsic spherical geometry and spherical distortions caused by equirectangular projection (ERP). To address this, we start from the very basics of $S^2$ manifold inherent to panoramic images, and propose a novel spherical manifold convolution (SMConv) on $S^2$ manifold. Based on the SMConv operation, we propose a spherical manifold guided diffusion (SMGD) model for text-conditioned panoramic image generation, which can well accommodate the spherical geometry during generation. We further develop a novel evaluation method by calculating grouped Fréchet inception distance (FID) on cube-map projections, which can well reflect the quality of generated panoramic images, compared to existing methods that randomly crop ERP-distorted content. Experiment results demonstrate that our SMGD model achieves the state-of-the-art generation quality and accuracy, whilst retaining + +the shortest sampling time in the text-conditioned panoramic image generation task. Codes are publicly available at https://github.com/chronos123/SMGD. + +# 1. Introduction + +Being able to be controlled by text descriptions, panoramic image generation with a full $360^{\circ}$ horizontal and $180^{\circ}$ vertical field-of-view [44] opens a new avenue for various applications in virtual reality [48], augmented reality [47], autonomous driving [46], etc. Although recent developments in deep probabilistic models including generative adversarial networks (GANs) [14, 20, 34, 57] and diffusion models [10, 32, 33] have been acting as the workhorse in generating photo-realistic images, panoramic image generation still remains challenging due to their intrinsic spherical geometry property and distorted scene structures. In particular, the equirectangular projection (ERP) [50] employed to convert the panoramic content into a planar rectangular format, introduces spherical distortions and breaks the continuity at the left and right edges. Within the ERP format, rectangular panoramic images with an aspect ratio of 1:2 exhibit a + +non-uniform pixel distribution. In other words, the pixel density is notably sparse near the polar regions and significantly denser around the equatorial region. Consequently, existing state-of-the-art deep probabilistic models designed for generating planar images [31, 34, 49, 51] still yet to generalize well for the panoramic image generation. + +To address the panoramic image generation task, several methods proposed to generate rectangular images, including autoregressive generation based on the latent codes [11], stitching latent features [3], and directly autoregressive generation on image patches [24]. To retain the continuity between the left and right boundaries, strategies such as circular convolution [5], circular token mechanisms [2], and exploiting the symmetry of panoramic images [15] have been further proposed. However, the above methods fail to effectively mitigate the spherical distortions, thus motivating the usage of deformable convolution [43] and attention mechanisms [53]. Unfortunately, existing methods still suffer from sub-optimal performance, which lacks well-defined global structure and optimal preservation of spherical geometry, as illustrated in Fig. 1, whilst compensating this with heavy computational complexity. + +In this paper, we start from the intrinsic property of panoramic format, and propose the spherical manifold guided diffusion (SMGD) model operating on $S^2$ manifold, to generate panoramic images from text descriptions. Since panoramic images can be recognized as pixels distributed on $S^2$ manifold, we propose the spherical manifold convolution (SMConv), to effectively address spherical distortions and preserve spherical geometry in panoramic images. To the best of our knowledge, there exists no spherical convolution designed for the panorama generation task, in which our SMConv fills this void by performing the correct convolution on spherical surfaces. Based on SMConv, we establish spherical manifold encoder (SME) and spherical manifold decoder (SMD) blocks, constituting a spherical manifold U-Net (SMUNet). Our SMGD model is subsequently constructed by employing the SMUNet as the denoising network, with the VQGAN [11] serving as its image encoder. Unlike previous spherical convolutions [8, 36, 37, 45, 55] applied in classification and detection tasks, our SMConv is supported by the intrinsic manifold theory. Guided by exponential mapping, our SMConv incorporates the convolution within spherical neighborhoods under the measure of the geodesic distance [40]. Moreover, to further address the spherical geometry of panoramic images, we develop a spherical loss to optimize the proposed SMGD model. We also adopt an enhanced strategy to evaluate the quality of generated panoramic images by calculating the grouped Fréchet inception distance (FID) [17] in cubemap projection (CMP) [52] format. Experimental results demonstrate that our method achieves superior generation performances by effectively addressing the unique challenges associated with panoramic + +image generation, such as complex scene structure, spherical distortions caused by ERP and the sparse pixel distribution near the poles. In summary, our contributions are three-fold: + +- We set out the first attempt to propose the spherical convolution, namely, SMConv, operating on $S^2$ manifold, which is able to optimally capture the intrinsic spherical geometry in panoramic images. +- We propose to incorporate SMConv into spherical manifold guided blocks, including the SME and SMD blocks, so as to mitigate spherical distortions and preserve spatial coherence across the spherical domain. +- Building upon these foundational blocks, we propose the SMGD model as the first generative approach to incorporate spherical convolution and achieve the state-of-the-art performance in generating panoramic images from text descriptions. + +# 2. Related Works + +Panoramic Image Generation. Early research on panoramic image generation can be achieved by the image outpainting task, which generates panoramic images from partial view [1, 2, 9, 15, 28, 38, 42]. However, these outpainting methods are limited by the freedom of control over the generated content. Most recently, with the rapid advancements in generative models, text-conditioned panoramic image generation [3, 5, 23, 24, 39, 41, 43, 53, 54] has received increasing research efforts. More specifically, Text2Light [5] first proposed to generate panoramic images from text descriptions, using CLIP [30] to encode text together with VQGAN [11] to encode panoramic images. To generate images at arbitrary sizes, DiffCollage [54], Multi-Diffusion [3] and SyncDiffusion [23] aggregated the intermediate outputs of pre-trained diffusion models. PanoGen [24] proposed to recursively outpaint the images generated by pretrained models to create panoramic images. However, those methods either fail to handle spherical distortions nor to ensure continuity between the left and right boundaries. To address this problem, MVDiffusion [39] generated perspective images and stitches them into panoramic images. However, the stitched panoramic image may witness its limitation in the field of view. PanFusion [53] introduced an attention mechanism for perspective images to enhance the generation of local details, while the main branch focused on generating the overall panoramic image. However, PanFusion utilizes two pretrained diffusion models [18] whereby multiple perspective images are processed in a separate branch; this can lead to unnatural global structures and result in heavy computational complexity. + +Convolution Operation for Panoramic Images. Several attempts have introduced modifications to the convolution operation [7, 8, 12, 21, 29, 36, 37, 45, 55] in order to effectively manage spherical distortions for panoramic image classification and segmentation tasks. The pioneering work [36] + +adopted kernels with different shapes at different latitudes in object detection task. Inspired by these works, deformable kernels [12, 55] were adopted to handle the spherical geometry. On the other hand, Coors et al. [8] retained the kernel unchanged on the tangent plane of the sphere and projected it back to the sphere using inverse gnomonic projection in image classification task. For the same task, Xu et al. [45] proposed to enlarge the receptive filed of convolution kernel by defining the kernel area in sphere. Furthermore, the rotation invariance property has been considered for designing spherical convolution in image classification [7] and image segmentation tasks [29]. However, spherical convolution operations have not yet been proposed for generative models that essentially suffer from improper spherical geometry. Therefore, we propose to incorporate the guidance from SMConv by exponential mapping, operating on $S^2$ manifold for high quality panoramic image generation. + +# 3. Methodology + +# 3.1. Spherical Manifold Convolution + +The spherical distortion introduced by ERP leads to object deformation and interrupts the continuity of content on the sphere, which together significantly hinder the correct information aggregation of standard convolution. In contrast, performing convolution operation directly on the spherical domain is of significant potential to preserve the spherical geometry of panoramic images and to mitigate spherical distortions. By inspecting that panoramic images can be viewed as pixels distributed on a spherical surface, i.e., the $S^2$ manifold, we propose the SMConv operation that performs convolution on the $S^2$ manifold. More specifically, we employ exponential mapping [13] to project uniform convolution kernels from the tangent space onto the $S^2$ manifold, enabling SMConv with kernels to uniformly aggregate the content across the $S^2$ manifold. More importantly, due to the length-preserving property of exponential mapping [13], SMConv retains the uniformity of geodesic distances, known as great-circle distances [40], on the $S^2$ manifold. This enables SMConv to better handle spherical distortions, providing an effective solution for panoramic image generation. Compared to previous spherical convolution based on inverse gnomonic projection [8], employing exponential mapping to construct SMConv kernels enables an expanded receptive field, refined kernel patterns, and improved preservation of spherical geometry, as demonstrated in Fig. 2. + +More specifically, the $S^2$ manifold is defined as a set of points $x \in \mathbb{R}^3$ with norm 1, which can be parameterized by spherical coordinates $\theta \in [0,2\pi]$ and $\varphi \in [-\pi/2, \pi/2]$ . Through a transformation between spherical and Cartesian coordinates, we can obtain 3D Cartesian coordinates and then apply the exponential mapping. For instance, a point $\mathbf{p} = (\theta, \varphi)$ on $S^2$ manifold can be converted to a corre + +![](images/7efba72479e359fdadabea1b821408dd4a8373bbefaba8ec641e63cee73d8852.jpg) +Figure 2. Comparison of exponential mapping (denoted as $\mathrm{Exp}_{\mathbf{p}}(\cdot))$ and inverse gnomonic projection (denoted as $\mathrm{IGP}(\cdot))$ [8] through forward and backward mappings. Left: visualization of a point $\mathbf{v}$ in the tangent space $T_{\mathbf{p}}S^{2}$ at point $\mathbf{p}$ , mapped onto the sphere. Right: projection of a panoramic image, the spherical data, onto the tangent space $T_{\mathbf{p}}S^{2}$ centered at $\mathbf{p} = (\theta, \varphi)$ whereby $\theta = 0$ and $\varphi = \frac{\pi}{4}$ . + +sponding point $\mathbf{p} = (x,y,z)$ with Cartesian coordinates. The exponential mapping [13] can be formulated as follows: + +$$ +\mathbf {q} = \operatorname {E x p} _ {\mathbf {p}} (\mathbf {v}) = \cos (| | \mathbf {v} | |) \mathbf {p} + \sin (| | \mathbf {v} | |) \frac {\mathbf {v}}{| | \mathbf {v} | |}, \quad (1) +$$ + +where $\mathbf{q}$ denotes the mapped point, $\mathbf{p}$ is the tangent point, and $\mathbf{v}$ is a vector representing the coordinates of the kernel points $\mathbf{K}_{T_{\mathbf{p}}}(m,n)$ in tangent space $T_{\mathbf{p}}S^{2}$ centered at $\mathbf{p}$ . + +In practice, given a predefined position $(u_{m},v_{n})$ of a kernel point $\mathbf{K}_{T_{\mathbf{p}}}(m,n)$ in the tangent space $T_{\mathbf{p}}S^{2}$ , the corresponding vector $\mathbf{v}$ in $T_{\mathbf{p}}S^{2}$ can be obtained as follows: + +$$ +\mathbf {v} = u _ {m} \hat {\mathbf {e}} _ {\theta} + v _ {n} \hat {\mathbf {e}} _ {\varphi}, +$$ + +$$ +\hat {\mathbf {e}} _ {\theta} = (- \sin \theta , \cos \theta , 0), \tag {2} +$$ + +$$ +\hat {\mathbf {e}} _ {\varphi} = (- \sin \varphi \cos \theta , - \sin \varphi \sin \theta , \cos \varphi), +$$ + +where $\mathbf{p} = (\theta, \varphi)$ denotes the center point in $T_{\mathbf{p}}S^2$ and the vectors $\hat{\mathbf{e}}_{\theta}, \hat{\mathbf{e}}_{\varphi}$ serve as the basis for the coordinate system in $T_{\mathbf{p}}S^2$ . Note that the shape and the position of the convolution kernel are provided by $(u_m, v_n)$ in $T_{\mathbf{p}}S^2$ and we can calculate the position $\mathbf{q} = \mathrm{Exp}_{\mathbf{p}}(\mathbf{v}) = (\hat{x}_{m,n}, \hat{y}_{m,n}, \hat{z}_{m,n})$ of the kernel point $\mathbf{K}(m,n)$ on $S^2$ manifold in Cartesian coordinates, via (1) and (2). Here, the kernel $\mathbf{K}$ is centered at the point $\mathbf{p}$ on $S^2$ manifold. Due to the properties of the exponential mapping [13], the norm of the mapped point $\mathbf{q} = (\hat{x}_{m,n}, \hat{y}_{m,n}, \hat{z}_{m,n})$ is guaranteed to be 1, ensuring that the point remains on the $S^2$ manifold. Since we need the position $\mathbf{q} = (\theta_m, \varphi_n)$ in spherical coordinates for performing SMConv, we can simple apply the inverse transformation from Cartesian to spherical coordinates. + +Furthermore, suppose $\widetilde{\mathbf{F}}_{S^2},\mathbf{F}_{S^2}$ are features on $S^2$ manifold. The operation of SMConv can be formulated by + +$$ +\widetilde {\mathbf {F}} _ {S ^ {2}} (\theta , \varphi) = \sum_ {m, n} \mathbf {F} _ {S ^ {2}} (\theta + \theta_ {m}, \varphi + \varphi_ {n}) \cdot \mathbf {K} (m, n), \tag {3} +$$ + +$$ +\left(\theta_ {m}, \varphi_ {n}\right) = \mathbf {q} = \operatorname {E x p} _ {\mathbf {p}} \left(\mathbf {K} _ {T _ {\mathbf {p}}} (m, n)\right), +$$ + +![](images/efd97cab99608783b86bfab09bc73bdf8999a217e8f906208ba2cb90653d9477.jpg) +Figure 3. Overall architecture of the proposed SMGD model, which introduces SMUNet as the denoising network and a panoramic VQGAN as the image encoder. The SMUNet is primarily constructed by the SMConv, operating on the $S^2$ manifold, and integrates the SPB along with circular encoder and decoder blocks. + +In (3), $\mathbf{q}$ is the position of the kernel point $\mathbf{K}(m,n)$ for spherical features in ERP format, which is obtained by performing exponential mapping for the kernel point $\mathbf{K}_{T_{\mathbf{p}}}(m,n)$ in the tangent space $T_{\mathbf{p}}S^{2}$ . Moreover, $\mathbf{K}$ represents the kernel centered at point $\mathbf{p}$ on $S^2$ manifold and the weight of $\mathbf{K}$ is identical to that of $\mathbf{K}_{T_{\mathbf{p}}}$ . By uniformly choosing $u_{m}$ and $v_{m}$ and constructing a regular kernel in the tangent space $T_{\mathbf{p}}S^{2}$ , we can perform SMConv with kernels uniformly distributed across the $S^2$ manifold using exponential mapping, according to (3). + +# 3.2. SMConv Inspired Diffusion Model + +Building upon SMConv, we introduce the SMGD model to achieve high-quality panoramic image generation. The overall architecture of our SMGD model is shown in Fig. 3. The diffusion process can be formulated as + +$$ +q (\mathbf {z} _ {t} | \mathbf {z} _ {t - 1}) := \mathcal {N} (\mathbf {z} _ {t - 1}; \sqrt {1 - \beta_ {t}} \mathbf {z} _ {t - 1}, \beta_ {t} \mathbf {I}), +$$ + +$$ +\mathbf {z} _ {t - 1} = \frac {1}{\sqrt {\alpha_ {t}}} \left(\mathbf {z} _ {t} - \frac {1 - \alpha_ {t}}{\sqrt {1 - \bar {\alpha} _ {t}}} \epsilon_ {\psi} (\mathbf {z} _ {t}, t, \mathbf {T})\right) + \sqrt {\beta_ {t}} \eta , ^ {(4)} +$$ + +where $\mathbf{z}_{t - 1},\mathbf{z}_t$ denote the noisy spherical features with the same dimensionality as the data, at time step $t - 1\in [0,T]$ and $t\in [0,T]$ respectively. $\mathbf{T}$ is the text embedding obtained by CLIP model [31], i.e., from the pretrained CLIP ViT/L-14 model. In addition, $q(\mathbf{z}_t|\mathbf{z}_{t - 1})$ represents the forward process gradually adding noise to the image, $\epsilon_{\psi}(\mathbf{z}_t,t)$ is the noise predicted by our SMUNet at time step $t$ , $\psi$ represents the parameters of SMUNet and $\eta$ denotes a random Gaussian noise. Furthermore, the noise scheduler configuration $\beta_{t}$ is a hyperparameter that controls the strength of the noise involved at each time step. The parameter $\alpha_{t}$ is defined as + +$\alpha_{t} = 1 - \beta_{t}$ , and $\bar{\alpha}_{t}$ is given by $\bar{\alpha}_{t} = \prod_{s=1}^{t} \alpha_{s}$ , the default setting within diffusion models. + +As the core component of the SMGD model, our SMUNet predicts noise at each time step $t$ from the noisy spherical feature $\mathbf{z}_t$ and facilitates the recovery of the spherical feature $\mathbf{z}_0$ on $S^2$ manifold by subtracting the predicted noise. We employ spherical processor blocks (SPBs), along with circular encoder and decoder blocks based on circular convolution [2] at shallower layers of SMUNet. Then, we apply SME and SMD blocks at deeper layers with middle blocks consisted of spherical manifold residual blocks (SMRBs) and cross-attention layers to integrate text embeddings $\mathbf{T}$ . The design of a hybrid architecture to integrate circular convolution at shallower layers with SMConv at deeper layers is motivated by the following considerations. As resolution increases, spherical convolution operation including SMConv, consumes substantially higher GPU memory cost compared to standard convolution [8] due to its bilinear interpolation. Furthermore, in deeper layers of SMUNet where spatial resolution is reduced, the sparse pixel distribution near the poles becomes increasingly pronounced, exacerbating the limitations of standard convolution. Consequently, the proposed hybrid architecture well balances between output quality, computational efficiency, and memory utilization. + +Then, the SME block comprises 2 SMRBs, 2 cross-attention layers, and an SMConv with a stride of 2 for downsampling. For the $l$ -th SME block, the SMRB can be formulated as + +$$ +\widetilde {\mathbf {z}} _ {t} ^ {E _ {l}} = \mathrm {S M C o n v} [ \mathrm {S M C o n v} [ \mathbf {z} _ {t} ^ {E _ {l}} ] + g _ {\psi} (t) ] + \mathbf {z} _ {t} ^ {E _ {l}}, \quad (5) +$$ + +where the calculation of SMConv follows (3) and $g_{\psi}(t)$ is the learnable time-step embedding function [32]. Moreover, + +$\mathbf{z}_t^{E_l}$ denotes the input, while $\widetilde{\mathbf{z}}_t^{E_l}$ represents the output for this layer. The structure of the SMD block closely mirrors that of the indexing SME block, with the upsampling layer using nearest-neighbor interpolation. Note that the final SME block and the first SMD block omit upsampling and downsampling operations. Moreover, the SPB employs the SMConv combined with a $1\times 1$ residual convolutional layer to enhance the transformation between different feature channels. The circular encoder and decoder blocks follow the structure in LDM [32], with all planar convolutions replaced by circular convolution operations [2]. Notably, the circular encoder blocks use average pooling layers for downsampling. In addition to the SMUNet, we propose to employ a panoramic VQGAN [11] with a downsampling factor of 16 as the image encoder and decoder for our SMGD model. In the panoramic VQGAN, all planar convolutions in the layers of decoder are replaced by circular convolutions [2]. + +# 3.3. Spherical Guided Training Strategy + +The training process involves two stages: training the panoramic VQGAN, and then training the SMGD model. During the training of the panoramic VQGAN, reconstruction loss $\mathcal{L}_{\mathrm{rec}}$ [11], commitment loss $\mathcal{L}_{\mathrm{Commit}}$ [11] and perceptual loss $\mathcal{L}_{\mathrm{Per}}$ [19] are employed, which can be formulated as + +$$ +\mathcal {L} _ {\mathrm {V Q G A N}} = \mathcal {L} _ {\text {r e c}} + \lambda_ {\text {c o m m i t}} \mathcal {L} _ {\text {c o m m i t}} + \lambda_ {\text {p e r}} \mathcal {L} _ {\text {P e r}}, \tag {6} +$$ + +where $\lambda_{\mathrm{commit}}$ and $\lambda_{\mathrm{per}}$ are hyperparameters to control the weights of different loss components during training. + +During the training of the SMGD model, a spherical loss based on the weighted spherically uniform PSNR (WSPSNR) [50] is employed. The total loss can be formulated as + +$$ +\mathcal {L} _ {\text {D i f f}} = \lambda_ {\text {M S E}} \mathcal {L} _ {\text {M S E}} + \lambda_ {\text {S M S E}} \mathcal {L} _ {\text {S M S E}}, \tag {7} +$$ + +where $\lambda_{\mathrm{MSE}}$ and $\lambda_{\mathrm{SMSE}}$ are also balancing hyperparameters. Furthermore, the $\mathcal{L}_{\mathrm{MSE}}$ and $\mathcal{L}_{\mathrm{SMSE}}$ are given by + +$$ +\mathcal {L} _ {\mathrm {M S E}} = \mathbb {E} _ {t, \mathbf {z} _ {0}, \epsilon} [ | | \epsilon - \epsilon_ {\psi} (\mathbf {z} _ {t}, t, \mathbf {T}) | | _ {2} ^ {2} ], \tag {8} +$$ + +$$ +\mathcal {L} _ {\mathrm {S M S E}} = \mathbb {E} _ {t, \mathbf {z} _ {0}, \epsilon} [ | | \mathbf {W} \odot (\epsilon - \epsilon_ {\psi} (\mathbf {z} _ {t}, t, \mathbf {T})) | | _ {2} ^ {2} ], +$$ + +where $\mathbf{W}$ is given by the WS-PSNR [50] to reflect the panoramic characteristics as + +$$ +\mathbf {W} _ {j} = \cos \left(\left(\frac {j + 0 . 5}{H} - 0. 5\right) \pi\right), \tag {9} +$$ + +where $j \in [0, H]$ is the index of height in $\mathbf{z}_t$ and $H$ denotes the height of the spherical feature $\mathbf{z}_t$ . $\mathcal{L}_{\mathrm{SMSE}}$ can improve the preservation of spherical geometry, e.g., a sparse pixel distribution near the poles. + +# 3.4. Quality Evaluation for Panoramic Images + +FID score is the de facto standard for evaluating the quality of generated images, whereas performing a direct calculation + +![](images/5a7d36517b32d90b987b488e8262b38e73b5d03ad968704942cd54fb3cd45c24.jpg) +Figure 4. Illustration of the conversion between ERP and CMP formats. In the CMP format, four blue-bordered images represent equatorial regions, while two orange-bordered images represent polar areas. These images are grouped accordingly, with $\mathrm{FID}_{\mathrm{equ}}$ and $\mathrm{FID}_{\mathrm{pole}}$ calculated separately for each group. + +of FID on panoramic images in rectangular ERP format introduces severe distortion, leading to inaccuracy. We propose to convert the generated panoramic images from ERP format to CMP format consisted of 6 square images for calculating FID score. The mapping from ERP to CMP format [50, 52] provides a robust solution to evaluate the quality of generated panoramic images. Note that although sharing similar goals with the concurrent work [6], our FID-based metrics are able to reflect detailed quality for different regions when evaluating generated panoramic images. As illustrated in Fig. 4, the conversion reveals the continuity across the left and right boundaries and represents the spherical geometry by obtaining 4 images representing areas near the equator and 2 images representing areas near the poles. Thus, the 6 square images within CMP format can be categorized into 2 groups, allowing for separate FID calculations to assess the quality of generated images. More specifically, $\mathrm{FID}_{\mathrm{equ}}$ is calculated using the 4 images near the equator, while $\mathrm{FID}_{\mathrm{pole}}$ is derived from the 2 images near the poles. $\mathrm{FID}_{\mathrm{equ}}$ primarily assesses the quality and visual fidelity of the panoramic content, whereas $\mathrm{FID}_{\mathrm{pole}}$ evaluates the preservation of spherical geometry. Analyzing $\mathrm{FID}_{\mathrm{equ}}$ and $\mathrm{FID}_{\mathrm{pole}}$ provides a comprehensive measure of the quality of the generated panoramic image. + +# 4. Experiment + +# 4.1. Experimental Settings + +Dataset. We employed the widely-used Matterport3D [4] dataset to evaluate the performance for text-conditioned panoramic image generation. More specifically, we obtained 10,912 panoramic images with a resolution of $1024 \times 512$ according to [26]. We then utilized the BLIP-2 [25] model to generate text descriptions for each panoramic image, forming paired text-image data for training and testing. To enable FID-10K calculation, we allocated 10,000 images for training and testing. + +Implementation Details. We adopted the panoramic VQGAN according to Sec. 3.2. SMUNet, the denoising network for our SMGD model, was configured with 320 and 640 channels for the two circular encoder and decoder blocks, + +Table 1. Comparisons in terms of FID and CLIP score (denoted as CS) values for panoramic images generated by our method and existing baseline methods, along with evaluations of parameter number (denoted by Para.) and time required to inference (denoted as $t_{\mathrm{sample}}$ ). We represent the best numbers by red color and the second best by blue color. + +
MethodsFID ↓FAED ↓OmniFID ↓FIDavg ↓FIDcent ↓FIDbord ↓FIDrand ↓FIDequ ↓FIDpole ↓CS ↑Para. ↓t_sample(s) ↓
Multi-Diffusion [3]69.961.1268.7056.7458.1956.5355.5027.2687.730.1821.3B39.45 ± 0.31
SD2+LoRA [18, 32]59.321.4769.0750.0451.7849.1649.1724.4988.680.1711.4B6.48 ± 0.49
Text2Light [5]48.424.1063.1643.6846.9346.6437.4730.8772.060.1881.0B50.12 ± 11.18
MVDiffusion [39]56.210.9394.5338.1538.8837.6037.9627.58125.480.1811.4B97.14 ± 0.79
PanFusion [53]16.150.3527.6414.3414.9113.9614.1414.8131.990.1822.3B38.51 ± 11.57
SMGD (Ours)9.850.2218.249.709.779.629.717.8821.050.1891.4B2.69 ± 0.39
+ +and 960 and 1280 channels for the two SME and SMD blocks utilizing the $3 \times 3$ SMConv. At the first training stage, we trained the panoramic VQGAN with the Adam optimizer [22], using a batch size of 1 and a learning rate of $4.5 \times 10^{-6}$ for 300 epochs, with $\lambda_{\mathrm{commit}} = 1$ and $\lambda_{\mathrm{per}} = 0.8$ . At the second stage, we trained the SMGD model from the scratch with the AdamW optimizer [27], a batch size of 8, and learning rate of $1 \times 10^{-5}$ for 800 epochs by setting $\lambda_{\mathrm{MSE}} = 0.8$ and $\lambda_{\mathrm{SMSE}} = 0.2$ . Moreover, the noise scheduler configuration $\beta_{t}$ followed that of LDM [32]. For inference, we used a DDIM sampler [35] with 50 steps and a classifier-free guidance scale [32] of 2.5. + +Baselines. We compared our SMGD model with existing text-conditioned panoramic image generation methods, including Multi-Diffusion [3], Text2Light [5], MVDiffusion [39] and PanFusion [53]. We also compared our SMGD model with the widely-used stable diffusion model [32] by performing a rank-16 LoRA [18] (denoted as SD2+LoRA). Different from [53], we directly trained the LoRA model with a learning rate of $1 \times 10^{-6}$ for 50 epochs on our training data. + +Evaluation Metrics. We computed $\mathrm{FID}_{\mathrm{equ}}$ and $\mathrm{FID}_{\mathrm{pole}}$ , as described in Sec. 3.4 to effectively assess the quality of panoramic images. Note that $\mathrm{FID}_{\mathrm{equ}}$ was calculated against 40K images and $\mathrm{FID}_{\mathrm{pole}}$ was calculated against 20K images. To further validate our metrics and assess the quality of the generated images, we calculated an extra FID value similar to that used in previous methods [53]. More specifically, we cropped the panoramic images into 3 representative areas instead of obtaining 20 perspective images in calculating FID values. Specifically, we cropped the panoramic images into the center, cross-border and random patches with size $512 \times 512$ . Then, we calculated the corresponding FID value of them and obtained the $\mathrm{FID}_{\mathrm{cent}}$ , $\mathrm{FID}_{\mathrm{bord}}$ and $\mathrm{FID}_{\mathrm{rand}}$ . By averaging these three FID values calculated against 10K images, we obtained $\mathrm{FID}_{\mathrm{avg}}$ . Moreover, we reported FID, OmniFID [6], and Fréchet Auto-Encoder Distance (FAED) [53] to comprehensively evaluate the quality of the generated full panoramic images. On the other hand, to evaluate the alignment between text and generated panoramic images, we utilized the widely adopted CLIP score [16] (denoted + +as CS).1 Furthermore, we recorded the time for inference $t_{\mathrm{sample}}$ by averaging the time required to generate 10 images on an NVIDIA 4090 GPU for each method. + +# 4.2. Comparison with Existing Methods + +Quantitative Results. In Table 1, we report the quantitative results of our method and baselines for panoramic image generation from text descriptions. From this table, our method consistently achieves the lowest FID values, the highest CLIP scores, and the least sampling times. More importantly, among the top-performance models, our method maintains a comparatively lower parameter number. More specifically, the lower values of $\mathrm{FID}_{\mathrm{cent}}$ , $\mathrm{FID}_{\mathrm{bord}}$ , $\mathrm{FID}_{\mathrm{rand}}$ , and $\mathrm{FID}_{\mathrm{avg}}$ demonstrate that our SMGD model achieves the state-of-the-art performance within text-conditioned panoramic image generation2. Moreover, the even lower $\mathrm{FID}_{\mathrm{equu}}$ and $\mathrm{FID}_{\mathrm{pole}}$ values indicate that our method well preserves the spherical geometry of panoramic images, compared to other baseline methods. Note that MVDiffusion [39] and SD2+LoRA achieved low $\mathrm{FID}_{\mathrm{avg}}$ values but high $\mathrm{FID}_{\mathrm{pole}}$ values. This may be due to the fact that images generated by MVDiffusion exhibited a limited field of view, while images produced by SD2+LoRA lacked the preservation of the spherical geometry in panoramic images. Furthermore, as shown in Table 1, our SMGD model achieves the fastest inference time among all comparison models, outperforming PanFusion [53], the second-best method in terms of image quality, by more than a factor of 10 in inference speed. + +Qualitative Results. We randomly chose 2 text descriptions and report the generated panoramic images with corresponding real images from dataset in Fig. 5. As illustrated in the figure, the panoramic images generated by our SMGD model achieved the best subjective quality and the most accurate alignment with the text descriptions. More specifically, improper spherical geometry, including inconsistency across the border and an unnatural pixel distribution near the poles, was observed in panoramic images generated by SD2+LoRA, Text2Light, and Multi-Diffusion, due to their + +1The CLIP score was computed using the ViT-L/14 CLIP model, consistent with the CLIP model employed in our SMGD model. +To further validate the robustness of our model, we conduct additional evaluations on the Structured3D [56] dataset in Supplementary Sec. A. We also provide additional qualitative results in Supplementary Sec. C. + +![](images/2e1faf734cdc582f4c64069394f0b38288bb2aba3b85970453cfe576b5f0379c.jpg) +Figure 5. Qualitative results for text-conditioned panoramic image generation with real images from dataset displayed on the top row. We also show the copy from the left side of the panoramic image to the right to reveal the consistency across the whole image. Note that images generated by MVDiffusion [39] exhibit a limited field of view, resulting in images with a reduced height. We highlight the improper spherical geometry, artifact regions and distorted objects and lines with corresponding color of boxes, which are mitigated by our SMGD model. + +neglect of spherical distortions. Panoramic images generated by Panfusion and MVDiffusion suffered from artifact regions and distorted objects and lines as highlighted with corresponding color boxes in Fig. 5. Moreover, images + +generated by MVDiffusion exhibited a limited field of view. This limitation impeded the visual quality and completeness of the generated images, making them less suitable for fully immersive visual experiences. In contrast, our SMGD + +Table 2. Ablation study for our SMGD model. We compare the performance achieved by adding individual components to the baseline model sequentially, including the spherical loss (denoted by $\mathcal{L}_{\mathrm{SMSE}}$ ), the SMD blocks, the SME blocks, and the SPB modules. We represent the best numbers by red color. + +
\( \mathcal{L}_{\text{SMSE}} \)SMD BlocksSME BlocksSPBFID\( \mathrm{FID}_{\text{avg}} \downarrow \)\( \mathrm{FID}_{\text{cent}} \downarrow \)\( \mathrm{FID}_{\text{bord}} \downarrow \)\( \mathrm{FID}_{\text{rand}} \downarrow \)\( \mathrm{FID}_{\text{equ}} \downarrow \)\( \mathrm{FID}_{\text{pole}} \downarrow \)CS ↑
××××23.1120.5720.3420.9120.4515.3534.870.187
✓×××20.0317.5017.0518.0417.4012.7832.950.187
✓✓××13.3512.8712.7912.9412.889.6120.450.188
✓✓✓×12.3611.4211.4611.6011.199.2121.700.186
✓✓✓✓9.859.709.779.629.717.8821.050.189
+ +![](images/67d77b24560738f886b2cb57ef64d73c118993529795ed41f55e20a561613c8f.jpg) +Comparison in terms of FID and CS + +![](images/c20ba096886dc23931c46ebd905b36949f02da1b3b3c82f77ecd5def95d75dc0.jpg) +Figure 6. Comparison between LDM+SphereNet and the proposed SMGD model. Top: Bar plot of FID and CLIP score values achieved by these two methods. Bottom: Subjective results of panoramic images generated by the text prompt "a hallway in a house". The comparison is conducted in both ERP and CMP formats, with distorted objects and lines highlighted. + +![](images/25dc8097b3ba681eff062dbc2a98e608db193922ccd75e1b2131cc944d1d2f31.jpg) + +model produced visually pleasing panoramic images with a well-defined overall structure and proper spherical geometry. + +# 4.3. Ablation Study + +In this section, we conducted ablation study to systematically evaluate the impact of core components in our SMGD model. More specifically, since the spherical loss and SMConv are important for handling spherical distortions and preserving spherical geometry in panoramic images, we analyzed performance changes by gradually adding the spherical loss $\mathcal{L}_{\mathrm{SMSE}}$ , the SMD blocks, the SME blocks and the SPB. As reported in Table 2, the involvement of SMD blocks + +significantly enhanced performance, with SME blocks further contributing to the improvements. Moreover, both the spherical loss function and the SPB effectively improved the quality of generated panoramic images. Therefore, each module plays a vital role in optimizing the performance of the SMGD model. + +Furthermore, to validate the effectiveness of SMConv compared to previous spherical convolution methods, we implemented LDM+SphereNet, which employs the panoramic VQGAN and a U-Net architecture identical to our SMUNet, whilst all SMConv layers were replaced by the convolution operations proposed in SphereNet [8]. The training and inference were kept consistent with those of our SMGD model. As shown in Fig. 6, we can conclude that the quality of panoramic images generated by LDM+SphereNet is limited, whereas our proposed SMGD model produces visually pleasing images with lower FID values and a higher CLIP score3. This demonstrates the effectiveness of using SMConv for advanced panoramic image generation. + +# 5. Conclusion + +In this paper, we have proposed a novel method named as spherical manifold guided diffusion (SMGD) model to address the challenges of spherical distortions associated with panoramic image generation. This was achieved by the spherical manifold convolution (SMConv) operation that is able to uniformly sample spherical content within each convolution layer. Different from existing spherical convolution operations applied in other tasks, the proposed SMConv operation was built upon the exponential mapping grounded in manifold theory, which is able to preserve lengths on the tangent plane and exhibit enhanced local properties. The overall training procedure of our SMGD model was also inspired by the spherical loss. Furthermore, we have proposed to calculate $\mathrm{FID}_{\mathrm{equ}}$ and $\mathrm{FID}_{\mathrm{pole}}$ for panoramic images in cubemap format to improve the evaluation on the quality of generated panoramic images. Experimental results have demonstrated that our method achieves superior performance in text-conditioned image generation, surpassing existing state-of-the-art methods in both the quality of generated images and computational efficiency. + +# Acknowledgement + +This work was supported by Natural Science Foundation of China (NSFC) under Grants 62450131, 62206011, 62231002 and 62401027, and Beijing Natural Science Foundation under Grant L223021. + +# References + +[1] Hao Ai, Zidong Cao, Haonan Lu, Chen Chen, Jian Ma, Pengyuan Zhou, Tae-Kyun Kim, Pan Hui, and Lin Wang. Dream360: Diverse and immersive outdoor virtual scene creation via transformer-based $360^{\circ}$ image outpainting. IEEE Transactions on Visualization and Computer Graphics, 2024. 2 +[2] Naofumi Akimoto, Yuhi Matsuo, and Yoshimitsu Aoki. Diverse plausible 360-degree image outpainting for efficient 3dcg background creation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11441-11450, 2022. 2, 4, 5 +[3] Omer Bar-Tal, Lior Yariv, Yaron Lipman, and Tali Dekel. Multidiffusion: Fusing diffusion paths for controlled image generation. In International Conference on Machine Learning, 2023. 2, 6 +[4] Angel Chang, Angela Dai, Thomas Funkhouser, Maciej Halber, Matthias Niessner, Manolis Savva, Shuran Song, Andy Zeng, and Yinda Zhang. Matterport3d: Learning from rgb-d data in indoor environments. arXiv preprint arXiv:1709.06158, 2017. 5 +[5] Zhaoxi Chen, Guangcong Wang, and Ziwei Liu. Text2light: Zero-shot text-driven hdr panorama generation. ACM Transactions on Graphics (TOG), 41(6):1-16, 2022. 2, 6 +[6] Anders Christensen, Nooshin Mojab, Khushman Patel, Karan Ahuja, Zeynep Akata, Ole Winther, Mar Gonzalez-Franco, and Andrea Colaco. Geometry fidelity for spherical images. In European Conference on Computer Vision, pages 276-292. Springer, 2024. 5, 6 +[7] Taco S Cohen, Mario Geiger, Jonas Köhler, and Max Welling. Spherical cnns. arXiv preprint arXiv:1801.10130, 2018. 2, 3 +[8] Benjamin Coors, Alexandru Paul Condurache, and Andreas Geiger. Spherenet: Learning spherical representations for detection and classification in omnidirectional images. In Proceedings of the European conference on computer vision (ECCV), pages 518-533, 2018. 2, 3, 4, 8 +[9] Mohammad Reza Karimi Dastjerdi, Yannick Hold-Geoffroy, Jonathan Eisenmann, Siavash Khodadadeh, and Jean-François Lalonde. Guided co-modulated gan for 360 field of view extrapolation. In 2022 International Conference on 3D Vision (3DV), pages 475–485. IEEE, 2022. 2 +[10] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 1 +[11] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12873-12883, 2021. 2, 5 +[12] Clara Fernandez-Labrador, Jose M Facil, Alejandro Perez-Yus, Cédric Demonceaux, Javier Civera, and Jose J Guerrero. + +Corners for layout: End-to-end layout recovery from 360 images. IEEE Robotics and Automation Letters, 5(2):1255-1262, 2020. 2, 3 +[13] Jean Gallier. Basics of Classical Lie Groups: The Exponential Map, Lie Groups, and Lie Algebras, pages 367-414. Springer New York, New York, NY, 2001. 3 +[14] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 1 +[15] Takayuki Hara, Yusuke Mukuta, and Tatsuya Harada. Spherical image generation from a single image by considering scene symmetry. Proceedings of the AAAI Conference on Artificial Intelligence, 35(2):1513-1521, 2021. 2 +[16] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 7514-7528, 2021. 6 +[17] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 2 +[18] Edward J Hu, Phillip Wallis, Zeyuan Allen-Zhu, Yuzhhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. In International Conference on Learning Representations, 2022. 2, 6 +[19] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In European Conference on Computer Vision, pages 694–711. Springer International Publishing, 2016. 5 +[20] Minguk Kang, Jun-Yan Zhu, Richard Zhang, Jaesik Park, Eli Shechtman, Sylvain Paris, and Taesung Park. Scaling up gans for text-to-image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10124-10134, 2023. 1 +[21] Renata Khasanova and Pascal Frossard. Geometry aware convolutional filters for omnidirectional images representation. In International conference on machine learning, pages 3351-3359. PMLR, 2019. 2 +[22] Diederik P Kingma. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6 +[23] Yuseung Lee, Kunho Kim, Hyunjin Kim, and Minhyuk Sung. Syncdiffusion: Coherent montage via synchronized joint diffusions. Advances in Neural Information Processing Systems, 36:50648-50660, 2023. 2 +[24] Jialu Li and Mohit Bansal. Panogen: Text-conditioned panoramic environment generation for vision-and-language navigation. Advances in Neural Information Processing Systems, 36, 2024. 2 +[25] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730-19742. PMLR, 2023. 5 +[26] Chieh Hubert Lin, Chia-Che Chang, Yu-Sheng Chen, Da-Cheng Juan, Wei Wei, and Hwann-Tzong Chen. Coco-gan: + +Generation by parts via conditional coordinating. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4512-4521, 2019. 5 +[27] I Loshchilov. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6 +[28] Zhuqiang Lu, Kun Hu, Chaoyue Wang, Lei Bai, and Zhiyong Wang. Autoregressive omni-aware outpainting for open-vocabulary 360-degree image generation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 14211-14219, 2024. 2 +[29] Thomas W Mitchel, Noam Aigerman, Vladimir G Kim, and Michael Kazhdan. Möbius convolutions for spherical cnns. In ACM SIGGRAPH 2022 Conference Proceedings, pages 1-9, 2022. 2, 3 +[30] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2 +[31] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1(2): 3, 2022. 2, 4 +[32] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684-10695, 2022. 1, 4, 5, 6 +[33] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in neural information processing systems, 35:36479-36494, 2022. 1 +[34] Axel Sauer, Tero Karras, Samuli Laine, Andreas Geiger, and Timo Aila. Stylegan-t: Unlocking the power of gans for fast large-scale text-to-image synthesis. In International conference on machine learning, pages 30105-30118. PMLR, 2023. 1, 2 +[35] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020.6 +[36] Yu-Chuan Su and Kristen Grauman. Learning spherical convolution for fast features from 360 imagery. Advances in neural information processing systems, 30, 2017. 2 +[37] Yu-Chuan Su and Kristen Grauman. Kernel transformer networks for compact spherical convolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9442-9451, 2019. 2 +[38] Julius Surya Sumantri and In Kyu Park. 360 panorama synthesis from a sparse set of images with unknown field of view. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 2386-2395, 2020. 2 +[39] Shitao Tang, Fuyang Zhang, Jiacheng Chen, Peng Wang, and Yasutaka Furukawa. Mvdiffusion: Enabling holistic multiview image generation with correspondence-aware diffusion. ArXiv, abs/2307.01097, 2023. 2, 6, 7 + +[40] T. Vincenty. Direct and inverse solutions of geodesics on the ellipsoid with application of nested equations. Survey Review, 23(176):88-93, 1975. 2, 3 +[41] Hai Wang, Xiaoyu Xiang, Yuchen Fan, and Jing-Hao Xue. Customizing 360-degree pancreas through text-to-image diffusion models. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 4933-4943, 2024. 2 +[42] Tianhao Wu, Chuanxia Zheng, and Tat-Jen Cham. Panodiffusion: 360-degree panorama outpainting via diffusion. In The Twelfth International Conference on Learning Representations, 2023. 2 +[43] Tao Wu, Xuewei Li, Zhongang Qi, Di Hu, Xintao Wang, Ying Shan, and Xi Li. Spherediffusion: Spherical geometry-aware distortion resilient diffusion model. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 6126-6134, 2024. 2 +[44] Mai Xu, Chen Li, Shanyi Zhang, and Patrick Le Callet. State-of-the-art in 360 video/image processing: Perception, assessment and compression. IEEE Journal of Selected Topics in Signal Processing, 14(1):5-26, 2020. 1 +[45] Yanyu Xu, Ziheng Zhang, and Shenghua Gao. Spherical dnns and their applications in 360 images and videos. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(10):7235-7252, 2021. 2, 3 +[46] Jian-Ru Xue, Jian-Wu Fang, and Pu Zhang. A survey of scene understanding by event reasoning in autonomous driving. International Journal of Automation and Computing, 15 (3):249-266, 2018. 1 +[47] Bangbang Yang, Yinda Zhang, Yijin Li, Zhaopeng Cui, Sean Fanello, Hujun Bao, and Guofeng Zhang. Neural rendering in a room: amodal 3d understanding and free-viewpoint rendering for the closed scene composed of pre-captured objects. ACM Transactions on Graphics (TOG), 41(4):1-10, 2022. 1 +[48] Bangbang Yang, Wenqi Dong, Lin Ma, Wenbo Hu, Xiao Liu, Zhaopeng Cui, and Yuewen Ma. Dreamspace: Dreaming your room space with text-driven panoramic texture propagation. In 2024 IEEE Conference Virtual Reality and 3D User Interfaces (VR), pages 650-660. IEEE, 2024. 1 +[49] Michihiro Yasunaga, Armen Aghajanyan, Weijia Shi, Rich James, Jure Leskovec, Percy Liang, Mike Lewis, Luke Zettlemoyer, and Wen-tau Yih. Retrieval-augmented multimodal language modeling. arXiv preprint arXiv:2211.12561, 2022. 2 +[50] Yan Ye, Elena Alshina, and Jill Boyce. Algorithm descriptions of projection format conversion and video quality metrics in 360lib. In Joint Video Exploration Team of ITU-T SG, 2018. 1, 5 +[51] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, et al. Scaling autoregressive models for content-rich text-to-image generation. arXiv preprint arXiv:2206.10789, 2(3):5, 2022. 2 +[52] Li Yu, Yanjun Gao, Farhad Pakdaman, and Moncef Gabbouj. Panoramic image inpainting with gated convolution and contextual reconstruction loss. In ICASSP 2024 - 2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 4255-4259, 2024. 2, 5 + +[53] Cheng Zhang, Qianyi Wu, Camilo Cruz Gambardella, Xiaoshui Huang, Dinh Phung, Wanli Ouyang, and Jianfei Cai. Taming stable diffusion for text to 360 panorama image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6347-6357, 2024. 1, 2, 6 +[54] Qinsheng Zhang, Jiaming Song, Xun Huang, Yongxin Chen, and Ming-Yu Liu. Diffcollage: Parallel generation of large content with diffusion models. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10188-10198. IEEE, 2023. 2 +[55] Qiang Zhao, Chen Zhu, Feng Dai, Yike Ma, Guoqing Jin, and Yongdong Zhang. Distortion-aware cnns for spherical images. In IJCAI, pages 1198-1204, 2018. 2, 3 +[56] Jia Zheng, Junfei Zhang, Jing Li, Rui Tang, Shenghua Gao, and Zihan Zhou. Structured3d: A large photo-realistic dataset for structured 3d modeling. In Proceedings of The European Conference on Computer Vision (ECCV), 2020. 6 +[57] Yufan Zhou, Ruiyi Zhang, Changyou Chen, Chunyuan Li, Chris Tensmeyer, Tong Yu, Jiuming Gu, Jinhui Xu, and Tong Sun. Towards language-free training for text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17907-17917, 2022. 1 \ No newline at end of file diff --git a/2025/Spherical Manifold Guided Diffusion Model for Panoramic Image Generation/images.zip b/2025/Spherical Manifold Guided Diffusion Model for Panoramic Image Generation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..5d71420b5c347f28783ad142e86b9b0098116077 --- /dev/null +++ b/2025/Spherical Manifold Guided Diffusion Model for Panoramic Image Generation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:57a194f5cec31ce545b884c11a48605ba49cb21b8d4a3a2128b5b567b61a5e2b +size 805737 diff --git a/2025/Spherical Manifold Guided Diffusion Model for Panoramic Image Generation/layout.json b/2025/Spherical Manifold Guided Diffusion Model for Panoramic Image Generation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..becda619c8f8155819d01cc77eb96ec885540000 --- /dev/null +++ b/2025/Spherical Manifold Guided Diffusion Model for Panoramic Image Generation/layout.json @@ -0,0 +1,9536 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 65, + 103, + 545, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 103, + 545, + 121 + ], + "spans": [ + { + "bbox": [ + 65, + 103, + 545, + 121 + ], + "type": "text", + "content": "Spherical Manifold Guided Diffusion Model for Panoramic Image Generation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 141, + 552, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 141, + 552, + 172 + ], + "spans": [ + { + "bbox": [ + 61, + 141, + 552, + 172 + ], + "type": "text", + "content": "Xiancheng Sun" + }, + { + "bbox": [ + 61, + 141, + 552, + 172 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 61, + 141, + 552, + 172 + ], + "type": "text", + "content": " Mai Xu" + }, + { + "bbox": [ + 61, + 141, + 552, + 172 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 61, + 141, + 552, + 172 + ], + "type": "text", + "content": " Shengxi Li" + }, + { + "bbox": [ + 61, + 141, + 552, + 172 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 61, + 141, + 552, + 172 + ], + "type": "text", + "content": " Senmao Ma" + }, + { + "bbox": [ + 61, + 141, + 552, + 172 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 61, + 141, + 552, + 172 + ], + "type": "text", + "content": " Xin Deng" + }, + { + "bbox": [ + 61, + 141, + 552, + 172 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 61, + 141, + 552, + 172 + ], + "type": "text", + "content": " Lai Jiang" + }, + { + "bbox": [ + 61, + 141, + 552, + 172 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 61, + 141, + 552, + 172 + ], + "type": "text", + "content": " Gang Shen" + }, + { + "bbox": [ + 61, + 141, + 552, + 172 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 61, + 141, + 552, + 172 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 61, + 141, + 552, + 172 + ], + "type": "text", + "content": "Beihang University " + }, + { + "bbox": [ + 61, + 141, + 552, + 172 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 61, + 141, + 552, + 172 + ], + "type": "text", + "content": "China Tower" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 78, + 178, + 305, + 385 + ], + "blocks": [ + { + "bbox": [ + 78, + 178, + 305, + 385 + ], + "lines": [ + { + "bbox": [ + 78, + 178, + 305, + 385 + ], + "spans": [ + { + "bbox": [ + 78, + 178, + 305, + 385 + ], + "type": "image", + "image_path": "77f9dff45c20491206b720effb9105058a844c3b356c70c6428895139d28aabd.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 116, + 384, + 290, + 401 + ], + "lines": [ + { + "bbox": [ + 116, + 384, + 290, + 401 + ], + "spans": [ + { + "bbox": [ + 116, + 384, + 290, + 401 + ], + "type": "text", + "content": "\"A serene room with a large door leading to a balcony and a window showcasing the garden, decorated in soft pastels\"" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 305, + 384, + 507, + 402 + ], + "lines": [ + { + "bbox": [ + 305, + 384, + 507, + 402 + ], + "spans": [ + { + "bbox": [ + 305, + 384, + 507, + 402 + ], + "type": "text", + "content": "\"A bright and airy staircase with large windows overlooking a manicured garden, enhancing the natural light that floods the space.\"" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 54, + 405, + 555, + 440 + ], + "lines": [ + { + "bbox": [ + 54, + 405, + 555, + 440 + ], + "spans": [ + { + "bbox": [ + 54, + 405, + 555, + 440 + ], + "type": "text", + "content": "Figure 1. A subjective comparison of zero-shot text-conditioned panoramic image generation between our SMGD model and the state-of-the-art method PanFusion [53]. Our model demonstrates superior performance in generating images with superior overall natural content, enhanced subjective quality, and better alignment with the text description, while benefiting reduced model parameters." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 305, + 179, + 509, + 384 + ], + "blocks": [ + { + "bbox": [ + 305, + 179, + 509, + 384 + ], + "lines": [ + { + "bbox": [ + 305, + 179, + 509, + 384 + ], + "spans": [ + { + "bbox": [ + 305, + 179, + 509, + 384 + ], + "type": "image", + "image_path": "825af3b51b3a7cbcc0e3ca75b623a3ac4263186e086c56eb595e77c0c7ed070d.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 152, + 449, + 200, + 460 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 449, + 200, + 460 + ], + "spans": [ + { + "bbox": [ + 152, + 449, + 200, + 460 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 465, + 297, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 465, + 297, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 465, + 297, + 693 + ], + "type": "text", + "content": "Panoramic image essentially acts as a pivotal role in emerging virtual reality and augmented reality scenarios; however, the generation of panoramic images are essentially challenging due to the intrinsic spherical geometry and spherical distortions caused by equirectangular projection (ERP). To address this, we start from the very basics of " + }, + { + "bbox": [ + 55, + 465, + 297, + 693 + ], + "type": "inline_equation", + "content": "S^2" + }, + { + "bbox": [ + 55, + 465, + 297, + 693 + ], + "type": "text", + "content": " manifold inherent to panoramic images, and propose a novel spherical manifold convolution (SMConv) on " + }, + { + "bbox": [ + 55, + 465, + 297, + 693 + ], + "type": "inline_equation", + "content": "S^2" + }, + { + "bbox": [ + 55, + 465, + 297, + 693 + ], + "type": "text", + "content": " manifold. Based on the SMConv operation, we propose a spherical manifold guided diffusion (SMGD) model for text-conditioned panoramic image generation, which can well accommodate the spherical geometry during generation. We further develop a novel evaluation method by calculating grouped Fréchet inception distance (FID) on cube-map projections, which can well reflect the quality of generated panoramic images, compared to existing methods that randomly crop ERP-distorted content. Experiment results demonstrate that our SMGD model achieves the state-of-the-art generation quality and accuracy, whilst retaining" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 450, + 556, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 450, + 556, + 486 + ], + "spans": [ + { + "bbox": [ + 313, + 450, + 556, + 486 + ], + "type": "text", + "content": "the shortest sampling time in the text-conditioned panoramic image generation task. Codes are publicly available at https://github.com/chronos123/SMGD." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 499, + 395, + 511 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 499, + 395, + 511 + ], + "spans": [ + { + "bbox": [ + 314, + 499, + 395, + 511 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 312, + 521, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 521, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 312, + 521, + 556, + 715 + ], + "type": "text", + "content": "Being able to be controlled by text descriptions, panoramic image generation with a full " + }, + { + "bbox": [ + 312, + 521, + 556, + 715 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 312, + 521, + 556, + 715 + ], + "type": "text", + "content": " horizontal and " + }, + { + "bbox": [ + 312, + 521, + 556, + 715 + ], + "type": "inline_equation", + "content": "180^{\\circ}" + }, + { + "bbox": [ + 312, + 521, + 556, + 715 + ], + "type": "text", + "content": " vertical field-of-view [44] opens a new avenue for various applications in virtual reality [48], augmented reality [47], autonomous driving [46], etc. Although recent developments in deep probabilistic models including generative adversarial networks (GANs) [14, 20, 34, 57] and diffusion models [10, 32, 33] have been acting as the workhorse in generating photo-realistic images, panoramic image generation still remains challenging due to their intrinsic spherical geometry property and distorted scene structures. In particular, the equirectangular projection (ERP) [50] employed to convert the panoramic content into a planar rectangular format, introduces spherical distortions and breaks the continuity at the left and right edges. Within the ERP format, rectangular panoramic images with an aspect ratio of 1:2 exhibit a" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 703, + 145, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 145, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 145, + 713 + ], + "type": "text", + "content": "*Corresponding author." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5824" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "content": "non-uniform pixel distribution. In other words, the pixel density is notably sparse near the polar regions and significantly denser around the equatorial region. Consequently, existing state-of-the-art deep probabilistic models designed for generating planar images [31, 34, 49, 51] still yet to generalize well for the panoramic image generation." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 148, + 297, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 148, + 297, + 339 + ], + "spans": [ + { + "bbox": [ + 55, + 148, + 297, + 339 + ], + "type": "text", + "content": "To address the panoramic image generation task, several methods proposed to generate rectangular images, including autoregressive generation based on the latent codes [11], stitching latent features [3], and directly autoregressive generation on image patches [24]. To retain the continuity between the left and right boundaries, strategies such as circular convolution [5], circular token mechanisms [2], and exploiting the symmetry of panoramic images [15] have been further proposed. However, the above methods fail to effectively mitigate the spherical distortions, thus motivating the usage of deformable convolution [43] and attention mechanisms [53]. Unfortunately, existing methods still suffer from sub-optimal performance, which lacks well-defined global structure and optimal preservation of spherical geometry, as illustrated in Fig. 1, whilst compensating this with heavy computational complexity." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 342, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 342, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 342, + 297, + 714 + ], + "type": "text", + "content": "In this paper, we start from the intrinsic property of panoramic format, and propose the spherical manifold guided diffusion (SMGD) model operating on " + }, + { + "bbox": [ + 55, + 342, + 297, + 714 + ], + "type": "inline_equation", + "content": "S^2" + }, + { + "bbox": [ + 55, + 342, + 297, + 714 + ], + "type": "text", + "content": " manifold, to generate panoramic images from text descriptions. Since panoramic images can be recognized as pixels distributed on " + }, + { + "bbox": [ + 55, + 342, + 297, + 714 + ], + "type": "inline_equation", + "content": "S^2" + }, + { + "bbox": [ + 55, + 342, + 297, + 714 + ], + "type": "text", + "content": " manifold, we propose the spherical manifold convolution (SMConv), to effectively address spherical distortions and preserve spherical geometry in panoramic images. To the best of our knowledge, there exists no spherical convolution designed for the panorama generation task, in which our SMConv fills this void by performing the correct convolution on spherical surfaces. Based on SMConv, we establish spherical manifold encoder (SME) and spherical manifold decoder (SMD) blocks, constituting a spherical manifold U-Net (SMUNet). Our SMGD model is subsequently constructed by employing the SMUNet as the denoising network, with the VQGAN [11] serving as its image encoder. Unlike previous spherical convolutions [8, 36, 37, 45, 55] applied in classification and detection tasks, our SMConv is supported by the intrinsic manifold theory. Guided by exponential mapping, our SMConv incorporates the convolution within spherical neighborhoods under the measure of the geodesic distance [40]. Moreover, to further address the spherical geometry of panoramic images, we develop a spherical loss to optimize the proposed SMGD model. We also adopt an enhanced strategy to evaluate the quality of generated panoramic images by calculating the grouped Fréchet inception distance (FID) [17] in cubemap projection (CMP) [52] format. Experimental results demonstrate that our method achieves superior generation performances by effectively addressing the unique challenges associated with panoramic" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 313, + 72, + 555, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 108 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 108 + ], + "type": "text", + "content": "image generation, such as complex scene structure, spherical distortions caused by ERP and the sparse pixel distribution near the poles. In summary, our contributions are three-fold:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 314, + 110, + 555, + 265 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 314, + 110, + 555, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 110, + 555, + 157 + ], + "spans": [ + { + "bbox": [ + 314, + 110, + 555, + 157 + ], + "type": "text", + "content": "- We set out the first attempt to propose the spherical convolution, namely, SMConv, operating on " + }, + { + "bbox": [ + 314, + 110, + 555, + 157 + ], + "type": "inline_equation", + "content": "S^2" + }, + { + "bbox": [ + 314, + 110, + 555, + 157 + ], + "type": "text", + "content": " manifold, which is able to optimally capture the intrinsic spherical geometry in panoramic images." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 314, + 158, + 555, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 158, + 555, + 205 + ], + "spans": [ + { + "bbox": [ + 314, + 158, + 555, + 205 + ], + "type": "text", + "content": "- We propose to incorporate SMConv into spherical manifold guided blocks, including the SME and SMD blocks, so as to mitigate spherical distortions and preserve spatial coherence across the spherical domain." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 205, + 555, + 265 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 205, + 555, + 265 + ], + "spans": [ + { + "bbox": [ + 314, + 205, + 555, + 265 + ], + "type": "text", + "content": "- Building upon these foundational blocks, we propose the SMGD model as the first generative approach to incorporate spherical convolution and achieve the state-of-the-art performance in generating panoramic images from text descriptions." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 274, + 406, + 288 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 274, + 406, + 288 + ], + "spans": [ + { + "bbox": [ + 314, + 274, + 406, + 288 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 295, + 556, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 295, + 556, + 654 + ], + "spans": [ + { + "bbox": [ + 313, + 295, + 556, + 654 + ], + "type": "text", + "content": "Panoramic Image Generation. Early research on panoramic image generation can be achieved by the image outpainting task, which generates panoramic images from partial view [1, 2, 9, 15, 28, 38, 42]. However, these outpainting methods are limited by the freedom of control over the generated content. Most recently, with the rapid advancements in generative models, text-conditioned panoramic image generation [3, 5, 23, 24, 39, 41, 43, 53, 54] has received increasing research efforts. More specifically, Text2Light [5] first proposed to generate panoramic images from text descriptions, using CLIP [30] to encode text together with VQGAN [11] to encode panoramic images. To generate images at arbitrary sizes, DiffCollage [54], Multi-Diffusion [3] and SyncDiffusion [23] aggregated the intermediate outputs of pre-trained diffusion models. PanoGen [24] proposed to recursively outpaint the images generated by pretrained models to create panoramic images. However, those methods either fail to handle spherical distortions nor to ensure continuity between the left and right boundaries. To address this problem, MVDiffusion [39] generated perspective images and stitches them into panoramic images. However, the stitched panoramic image may witness its limitation in the field of view. PanFusion [53] introduced an attention mechanism for perspective images to enhance the generation of local details, while the main branch focused on generating the overall panoramic image. However, PanFusion utilizes two pretrained diffusion models [18] whereby multiple perspective images are processed in a separate branch; this can lead to unnatural global structures and result in heavy computational complexity." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "text", + "content": "Convolution Operation for Panoramic Images. Several attempts have introduced modifications to the convolution operation [7, 8, 12, 21, 29, 36, 37, 45, 55] in order to effectively manage spherical distortions for panoramic image classification and segmentation tasks. The pioneering work [36]" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5825" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 277 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 277 + ], + "type": "text", + "content": "adopted kernels with different shapes at different latitudes in object detection task. Inspired by these works, deformable kernels [12, 55] were adopted to handle the spherical geometry. On the other hand, Coors et al. [8] retained the kernel unchanged on the tangent plane of the sphere and projected it back to the sphere using inverse gnomonic projection in image classification task. For the same task, Xu et al. [45] proposed to enlarge the receptive filed of convolution kernel by defining the kernel area in sphere. Furthermore, the rotation invariance property has been considered for designing spherical convolution in image classification [7] and image segmentation tasks [29]. However, spherical convolution operations have not yet been proposed for generative models that essentially suffer from improper spherical geometry. Therefore, we propose to incorporate the guidance from SMConv by exponential mapping, operating on " + }, + { + "bbox": [ + 55, + 72, + 296, + 277 + ], + "type": "inline_equation", + "content": "S^2" + }, + { + "bbox": [ + 55, + 72, + 296, + 277 + ], + "type": "text", + "content": " manifold for high quality panoramic image generation." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 290, + 138, + 304 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 290, + 138, + 304 + ], + "spans": [ + { + "bbox": [ + 55, + 290, + 138, + 304 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 311, + 229, + 323 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 311, + 229, + 323 + ], + "spans": [ + { + "bbox": [ + 55, + 311, + 229, + 323 + ], + "type": "text", + "content": "3.1. Spherical Manifold Convolution" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 329, + 296, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 329, + 296, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 329, + 296, + 628 + ], + "type": "text", + "content": "The spherical distortion introduced by ERP leads to object deformation and interrupts the continuity of content on the sphere, which together significantly hinder the correct information aggregation of standard convolution. In contrast, performing convolution operation directly on the spherical domain is of significant potential to preserve the spherical geometry of panoramic images and to mitigate spherical distortions. By inspecting that panoramic images can be viewed as pixels distributed on a spherical surface, i.e., the " + }, + { + "bbox": [ + 55, + 329, + 296, + 628 + ], + "type": "inline_equation", + "content": "S^2" + }, + { + "bbox": [ + 55, + 329, + 296, + 628 + ], + "type": "text", + "content": " manifold, we propose the SMConv operation that performs convolution on the " + }, + { + "bbox": [ + 55, + 329, + 296, + 628 + ], + "type": "inline_equation", + "content": "S^2" + }, + { + "bbox": [ + 55, + 329, + 296, + 628 + ], + "type": "text", + "content": " manifold. More specifically, we employ exponential mapping [13] to project uniform convolution kernels from the tangent space onto the " + }, + { + "bbox": [ + 55, + 329, + 296, + 628 + ], + "type": "inline_equation", + "content": "S^2" + }, + { + "bbox": [ + 55, + 329, + 296, + 628 + ], + "type": "text", + "content": " manifold, enabling SMConv with kernels to uniformly aggregate the content across the " + }, + { + "bbox": [ + 55, + 329, + 296, + 628 + ], + "type": "inline_equation", + "content": "S^2" + }, + { + "bbox": [ + 55, + 329, + 296, + 628 + ], + "type": "text", + "content": " manifold. More importantly, due to the length-preserving property of exponential mapping [13], SMConv retains the uniformity of geodesic distances, known as great-circle distances [40], on the " + }, + { + "bbox": [ + 55, + 329, + 296, + 628 + ], + "type": "inline_equation", + "content": "S^2" + }, + { + "bbox": [ + 55, + 329, + 296, + 628 + ], + "type": "text", + "content": " manifold. This enables SMConv to better handle spherical distortions, providing an effective solution for panoramic image generation. Compared to previous spherical convolution based on inverse gnomonic projection [8], employing exponential mapping to construct SMConv kernels enables an expanded receptive field, refined kernel patterns, and improved preservation of spherical geometry, as demonstrated in Fig. 2." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 630, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 296, + 715 + ], + "type": "text", + "content": "More specifically, the " + }, + { + "bbox": [ + 55, + 630, + 296, + 715 + ], + "type": "inline_equation", + "content": "S^2" + }, + { + "bbox": [ + 55, + 630, + 296, + 715 + ], + "type": "text", + "content": " manifold is defined as a set of points " + }, + { + "bbox": [ + 55, + 630, + 296, + 715 + ], + "type": "inline_equation", + "content": "x \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 55, + 630, + 296, + 715 + ], + "type": "text", + "content": " with norm 1, which can be parameterized by spherical coordinates " + }, + { + "bbox": [ + 55, + 630, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\theta \\in [0,2\\pi]" + }, + { + "bbox": [ + 55, + 630, + 296, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 630, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\varphi \\in [-\\pi/2, \\pi/2]" + }, + { + "bbox": [ + 55, + 630, + 296, + 715 + ], + "type": "text", + "content": ". Through a transformation between spherical and Cartesian coordinates, we can obtain 3D Cartesian coordinates and then apply the exponential mapping. For instance, a point " + }, + { + "bbox": [ + 55, + 630, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{p} = (\\theta, \\varphi)" + }, + { + "bbox": [ + 55, + 630, + 296, + 715 + ], + "type": "text", + "content": " on " + }, + { + "bbox": [ + 55, + 630, + 296, + 715 + ], + "type": "inline_equation", + "content": "S^2" + }, + { + "bbox": [ + 55, + 630, + 296, + 715 + ], + "type": "text", + "content": " manifold can be converted to a corre" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 317, + 68, + 553, + 194 + ], + "blocks": [ + { + "bbox": [ + 317, + 68, + 553, + 194 + ], + "lines": [ + { + "bbox": [ + 317, + 68, + 553, + 194 + ], + "spans": [ + { + "bbox": [ + 317, + 68, + 553, + 194 + ], + "type": "image", + "image_path": "7efba72479e359fdadabea1b821408dd4a8373bbefaba8ec641e63cee73d8852.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 200, + 555, + 278 + ], + "lines": [ + { + "bbox": [ + 313, + 200, + 555, + 278 + ], + "spans": [ + { + "bbox": [ + 313, + 200, + 555, + 278 + ], + "type": "text", + "content": "Figure 2. Comparison of exponential mapping (denoted as " + }, + { + "bbox": [ + 313, + 200, + 555, + 278 + ], + "type": "inline_equation", + "content": "\\mathrm{Exp}_{\\mathbf{p}}(\\cdot))" + }, + { + "bbox": [ + 313, + 200, + 555, + 278 + ], + "type": "text", + "content": " and inverse gnomonic projection (denoted as " + }, + { + "bbox": [ + 313, + 200, + 555, + 278 + ], + "type": "inline_equation", + "content": "\\mathrm{IGP}(\\cdot))" + }, + { + "bbox": [ + 313, + 200, + 555, + 278 + ], + "type": "text", + "content": " [8] through forward and backward mappings. Left: visualization of a point " + }, + { + "bbox": [ + 313, + 200, + 555, + 278 + ], + "type": "inline_equation", + "content": "\\mathbf{v}" + }, + { + "bbox": [ + 313, + 200, + 555, + 278 + ], + "type": "text", + "content": " in the tangent space " + }, + { + "bbox": [ + 313, + 200, + 555, + 278 + ], + "type": "inline_equation", + "content": "T_{\\mathbf{p}}S^{2}" + }, + { + "bbox": [ + 313, + 200, + 555, + 278 + ], + "type": "text", + "content": " at point " + }, + { + "bbox": [ + 313, + 200, + 555, + 278 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 313, + 200, + 555, + 278 + ], + "type": "text", + "content": ", mapped onto the sphere. Right: projection of a panoramic image, the spherical data, onto the tangent space " + }, + { + "bbox": [ + 313, + 200, + 555, + 278 + ], + "type": "inline_equation", + "content": "T_{\\mathbf{p}}S^{2}" + }, + { + "bbox": [ + 313, + 200, + 555, + 278 + ], + "type": "text", + "content": " centered at " + }, + { + "bbox": [ + 313, + 200, + 555, + 278 + ], + "type": "inline_equation", + "content": "\\mathbf{p} = (\\theta, \\varphi)" + }, + { + "bbox": [ + 313, + 200, + 555, + 278 + ], + "type": "text", + "content": " whereby " + }, + { + "bbox": [ + 313, + 200, + 555, + 278 + ], + "type": "inline_equation", + "content": "\\theta = 0" + }, + { + "bbox": [ + 313, + 200, + 555, + 278 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 200, + 555, + 278 + ], + "type": "inline_equation", + "content": "\\varphi = \\frac{\\pi}{4}" + }, + { + "bbox": [ + 313, + 200, + 555, + 278 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 282, + 556, + 306 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 282, + 556, + 306 + ], + "spans": [ + { + "bbox": [ + 313, + 282, + 556, + 306 + ], + "type": "text", + "content": "sponding point " + }, + { + "bbox": [ + 313, + 282, + 556, + 306 + ], + "type": "inline_equation", + "content": "\\mathbf{p} = (x,y,z)" + }, + { + "bbox": [ + 313, + 282, + 556, + 306 + ], + "type": "text", + "content": " with Cartesian coordinates. The exponential mapping [13] can be formulated as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 339, + 312, + 555, + 337 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 312, + 555, + 337 + ], + "spans": [ + { + "bbox": [ + 339, + 312, + 555, + 337 + ], + "type": "interline_equation", + "content": "\\mathbf {q} = \\operatorname {E x p} _ {\\mathbf {p}} (\\mathbf {v}) = \\cos (| | \\mathbf {v} | |) \\mathbf {p} + \\sin (| | \\mathbf {v} | |) \\frac {\\mathbf {v}}{| | \\mathbf {v} | |}, \\quad (1)", + "image_path": "ee3bd6065ac4615b0e1994034513bf41108b31a2bfc7c1fc2875edf79ae62711.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 342, + 555, + 379 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 342, + 555, + 379 + ], + "spans": [ + { + "bbox": [ + 313, + 342, + 555, + 379 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 342, + 555, + 379 + ], + "type": "inline_equation", + "content": "\\mathbf{q}" + }, + { + "bbox": [ + 313, + 342, + 555, + 379 + ], + "type": "text", + "content": " denotes the mapped point, " + }, + { + "bbox": [ + 313, + 342, + 555, + 379 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 313, + 342, + 555, + 379 + ], + "type": "text", + "content": " is the tangent point, and " + }, + { + "bbox": [ + 313, + 342, + 555, + 379 + ], + "type": "inline_equation", + "content": "\\mathbf{v}" + }, + { + "bbox": [ + 313, + 342, + 555, + 379 + ], + "type": "text", + "content": " is a vector representing the coordinates of the kernel points " + }, + { + "bbox": [ + 313, + 342, + 555, + 379 + ], + "type": "inline_equation", + "content": "\\mathbf{K}_{T_{\\mathbf{p}}}(m,n)" + }, + { + "bbox": [ + 313, + 342, + 555, + 379 + ], + "type": "text", + "content": " in tangent space " + }, + { + "bbox": [ + 313, + 342, + 555, + 379 + ], + "type": "inline_equation", + "content": "T_{\\mathbf{p}}S^{2}" + }, + { + "bbox": [ + 313, + 342, + 555, + 379 + ], + "type": "text", + "content": " centered at " + }, + { + "bbox": [ + 313, + 342, + 555, + 379 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 313, + 342, + 555, + 379 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 380, + 556, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 380, + 556, + 415 + ], + "spans": [ + { + "bbox": [ + 313, + 380, + 556, + 415 + ], + "type": "text", + "content": "In practice, given a predefined position " + }, + { + "bbox": [ + 313, + 380, + 556, + 415 + ], + "type": "inline_equation", + "content": "(u_{m},v_{n})" + }, + { + "bbox": [ + 313, + 380, + 556, + 415 + ], + "type": "text", + "content": " of a kernel point " + }, + { + "bbox": [ + 313, + 380, + 556, + 415 + ], + "type": "inline_equation", + "content": "\\mathbf{K}_{T_{\\mathbf{p}}}(m,n)" + }, + { + "bbox": [ + 313, + 380, + 556, + 415 + ], + "type": "text", + "content": " in the tangent space " + }, + { + "bbox": [ + 313, + 380, + 556, + 415 + ], + "type": "inline_equation", + "content": "T_{\\mathbf{p}}S^{2}" + }, + { + "bbox": [ + 313, + 380, + 556, + 415 + ], + "type": "text", + "content": ", the corresponding vector " + }, + { + "bbox": [ + 313, + 380, + 556, + 415 + ], + "type": "inline_equation", + "content": "\\mathbf{v}" + }, + { + "bbox": [ + 313, + 380, + 556, + 415 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 313, + 380, + 556, + 415 + ], + "type": "inline_equation", + "content": "T_{\\mathbf{p}}S^{2}" + }, + { + "bbox": [ + 313, + 380, + 556, + 415 + ], + "type": "text", + "content": " can be obtained as follows:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 353, + 423, + 436, + 436 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 423, + 436, + 436 + ], + "spans": [ + { + "bbox": [ + 353, + 423, + 436, + 436 + ], + "type": "interline_equation", + "content": "\\mathbf {v} = u _ {m} \\hat {\\mathbf {e}} _ {\\theta} + v _ {n} \\hat {\\mathbf {e}} _ {\\varphi},", + "image_path": "0e993985511f644ece885512173d03dbcc840f86e7c70ee6686d1e17bb8e58bd.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 350, + 438, + 555, + 451 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 438, + 555, + 451 + ], + "spans": [ + { + "bbox": [ + 350, + 438, + 555, + 451 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {e}} _ {\\theta} = (- \\sin \\theta , \\cos \\theta , 0), \\tag {2}", + "image_path": "71518ef0aa05f5f53b54a7e12e7ef64aa47e71c7a85e0b58a8dd6a2e7627a6b3.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 350, + 453, + 519, + 466 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 453, + 519, + 466 + ], + "spans": [ + { + "bbox": [ + 350, + 453, + 519, + 466 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {e}} _ {\\varphi} = (- \\sin \\varphi \\cos \\theta , - \\sin \\varphi \\sin \\theta , \\cos \\varphi),", + "image_path": "a78c75128478fbd804eafb699db3d37417c142da3e6dc9c3fe613637b3b96e2f.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 475, + 555, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 475, + 555, + 641 + ], + "spans": [ + { + "bbox": [ + 313, + 475, + 555, + 641 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 475, + 555, + 641 + ], + "type": "inline_equation", + "content": "\\mathbf{p} = (\\theta, \\varphi)" + }, + { + "bbox": [ + 313, + 475, + 555, + 641 + ], + "type": "text", + "content": " denotes the center point in " + }, + { + "bbox": [ + 313, + 475, + 555, + 641 + ], + "type": "inline_equation", + "content": "T_{\\mathbf{p}}S^2" + }, + { + "bbox": [ + 313, + 475, + 555, + 641 + ], + "type": "text", + "content": " and the vectors " + }, + { + "bbox": [ + 313, + 475, + 555, + 641 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{e}}_{\\theta}, \\hat{\\mathbf{e}}_{\\varphi}" + }, + { + "bbox": [ + 313, + 475, + 555, + 641 + ], + "type": "text", + "content": " serve as the basis for the coordinate system in " + }, + { + "bbox": [ + 313, + 475, + 555, + 641 + ], + "type": "inline_equation", + "content": "T_{\\mathbf{p}}S^2" + }, + { + "bbox": [ + 313, + 475, + 555, + 641 + ], + "type": "text", + "content": ". Note that the shape and the position of the convolution kernel are provided by " + }, + { + "bbox": [ + 313, + 475, + 555, + 641 + ], + "type": "inline_equation", + "content": "(u_m, v_n)" + }, + { + "bbox": [ + 313, + 475, + 555, + 641 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 313, + 475, + 555, + 641 + ], + "type": "inline_equation", + "content": "T_{\\mathbf{p}}S^2" + }, + { + "bbox": [ + 313, + 475, + 555, + 641 + ], + "type": "text", + "content": " and we can calculate the position " + }, + { + "bbox": [ + 313, + 475, + 555, + 641 + ], + "type": "inline_equation", + "content": "\\mathbf{q} = \\mathrm{Exp}_{\\mathbf{p}}(\\mathbf{v}) = (\\hat{x}_{m,n}, \\hat{y}_{m,n}, \\hat{z}_{m,n})" + }, + { + "bbox": [ + 313, + 475, + 555, + 641 + ], + "type": "text", + "content": " of the kernel point " + }, + { + "bbox": [ + 313, + 475, + 555, + 641 + ], + "type": "inline_equation", + "content": "\\mathbf{K}(m,n)" + }, + { + "bbox": [ + 313, + 475, + 555, + 641 + ], + "type": "text", + "content": " on " + }, + { + "bbox": [ + 313, + 475, + 555, + 641 + ], + "type": "inline_equation", + "content": "S^2" + }, + { + "bbox": [ + 313, + 475, + 555, + 641 + ], + "type": "text", + "content": " manifold in Cartesian coordinates, via (1) and (2). Here, the kernel " + }, + { + "bbox": [ + 313, + 475, + 555, + 641 + ], + "type": "inline_equation", + "content": "\\mathbf{K}" + }, + { + "bbox": [ + 313, + 475, + 555, + 641 + ], + "type": "text", + "content": " is centered at the point " + }, + { + "bbox": [ + 313, + 475, + 555, + 641 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 313, + 475, + 555, + 641 + ], + "type": "text", + "content": " on " + }, + { + "bbox": [ + 313, + 475, + 555, + 641 + ], + "type": "inline_equation", + "content": "S^2" + }, + { + "bbox": [ + 313, + 475, + 555, + 641 + ], + "type": "text", + "content": " manifold. Due to the properties of the exponential mapping [13], the norm of the mapped point " + }, + { + "bbox": [ + 313, + 475, + 555, + 641 + ], + "type": "inline_equation", + "content": "\\mathbf{q} = (\\hat{x}_{m,n}, \\hat{y}_{m,n}, \\hat{z}_{m,n})" + }, + { + "bbox": [ + 313, + 475, + 555, + 641 + ], + "type": "text", + "content": " is guaranteed to be 1, ensuring that the point remains on the " + }, + { + "bbox": [ + 313, + 475, + 555, + 641 + ], + "type": "inline_equation", + "content": "S^2" + }, + { + "bbox": [ + 313, + 475, + 555, + 641 + ], + "type": "text", + "content": " manifold. Since we need the position " + }, + { + "bbox": [ + 313, + 475, + 555, + 641 + ], + "type": "inline_equation", + "content": "\\mathbf{q} = (\\theta_m, \\varphi_n)" + }, + { + "bbox": [ + 313, + 475, + 555, + 641 + ], + "type": "text", + "content": " in spherical coordinates for performing SMConv, we can simple apply the inverse transformation from Cartesian to spherical coordinates." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 642, + 556, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 556, + 667 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 556, + 667 + ], + "type": "text", + "content": "Furthermore, suppose " + }, + { + "bbox": [ + 313, + 642, + 556, + 667 + ], + "type": "inline_equation", + "content": "\\widetilde{\\mathbf{F}}_{S^2},\\mathbf{F}_{S^2}" + }, + { + "bbox": [ + 313, + 642, + 556, + 667 + ], + "type": "text", + "content": " are features on " + }, + { + "bbox": [ + 313, + 642, + 556, + 667 + ], + "type": "inline_equation", + "content": "S^2" + }, + { + "bbox": [ + 313, + 642, + 556, + 667 + ], + "type": "text", + "content": " manifold. The operation of SMConv can be formulated by" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 324, + 673, + 555, + 700 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 673, + 555, + 700 + ], + "spans": [ + { + "bbox": [ + 324, + 673, + 555, + 700 + ], + "type": "interline_equation", + "content": "\\widetilde {\\mathbf {F}} _ {S ^ {2}} (\\theta , \\varphi) = \\sum_ {m, n} \\mathbf {F} _ {S ^ {2}} (\\theta + \\theta_ {m}, \\varphi + \\varphi_ {n}) \\cdot \\mathbf {K} (m, n), \\tag {3}", + "image_path": "a9aba7bba37817d34f41f042e5a9e36b4dc0689a3e02df3b87cef83e8b5c47fc.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 329, + 702, + 481, + 716 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 329, + 702, + 481, + 716 + ], + "spans": [ + { + "bbox": [ + 329, + 702, + 481, + 716 + ], + "type": "interline_equation", + "content": "\\left(\\theta_ {m}, \\varphi_ {n}\\right) = \\mathbf {q} = \\operatorname {E x p} _ {\\mathbf {p}} \\left(\\mathbf {K} _ {T _ {\\mathbf {p}}} (m, n)\\right),", + "image_path": "53a4abf9cd001e959a5490e2a315f46df507fe23a7bd19adcfd44da0ca3970bb.jpg" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5826" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 68, + 552, + 277 + ], + "blocks": [ + { + "bbox": [ + 56, + 68, + 552, + 277 + ], + "lines": [ + { + "bbox": [ + 56, + 68, + 552, + 277 + ], + "spans": [ + { + "bbox": [ + 56, + 68, + 552, + 277 + ], + "type": "image", + "image_path": "efd97cab99608783b86bfab09bc73bdf8999a217e8f906208ba2cb90653d9477.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 293, + 555, + 327 + ], + "lines": [ + { + "bbox": [ + 54, + 293, + 555, + 327 + ], + "spans": [ + { + "bbox": [ + 54, + 293, + 555, + 327 + ], + "type": "text", + "content": "Figure 3. Overall architecture of the proposed SMGD model, which introduces SMUNet as the denoising network and a panoramic VQGAN as the image encoder. The SMUNet is primarily constructed by the SMConv, operating on the " + }, + { + "bbox": [ + 54, + 293, + 555, + 327 + ], + "type": "inline_equation", + "content": "S^2" + }, + { + "bbox": [ + 54, + 293, + 555, + 327 + ], + "type": "text", + "content": " manifold, and integrates the SPB along with circular encoder and decoder blocks." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 332, + 297, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 332, + 297, + 452 + ], + "spans": [ + { + "bbox": [ + 55, + 332, + 297, + 452 + ], + "type": "text", + "content": "In (3), " + }, + { + "bbox": [ + 55, + 332, + 297, + 452 + ], + "type": "inline_equation", + "content": "\\mathbf{q}" + }, + { + "bbox": [ + 55, + 332, + 297, + 452 + ], + "type": "text", + "content": " is the position of the kernel point " + }, + { + "bbox": [ + 55, + 332, + 297, + 452 + ], + "type": "inline_equation", + "content": "\\mathbf{K}(m,n)" + }, + { + "bbox": [ + 55, + 332, + 297, + 452 + ], + "type": "text", + "content": " for spherical features in ERP format, which is obtained by performing exponential mapping for the kernel point " + }, + { + "bbox": [ + 55, + 332, + 297, + 452 + ], + "type": "inline_equation", + "content": "\\mathbf{K}_{T_{\\mathbf{p}}}(m,n)" + }, + { + "bbox": [ + 55, + 332, + 297, + 452 + ], + "type": "text", + "content": " in the tangent space " + }, + { + "bbox": [ + 55, + 332, + 297, + 452 + ], + "type": "inline_equation", + "content": "T_{\\mathbf{p}}S^{2}" + }, + { + "bbox": [ + 55, + 332, + 297, + 452 + ], + "type": "text", + "content": ". Moreover, " + }, + { + "bbox": [ + 55, + 332, + 297, + 452 + ], + "type": "inline_equation", + "content": "\\mathbf{K}" + }, + { + "bbox": [ + 55, + 332, + 297, + 452 + ], + "type": "text", + "content": " represents the kernel centered at point " + }, + { + "bbox": [ + 55, + 332, + 297, + 452 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 55, + 332, + 297, + 452 + ], + "type": "text", + "content": " on " + }, + { + "bbox": [ + 55, + 332, + 297, + 452 + ], + "type": "inline_equation", + "content": "S^2" + }, + { + "bbox": [ + 55, + 332, + 297, + 452 + ], + "type": "text", + "content": " manifold and the weight of " + }, + { + "bbox": [ + 55, + 332, + 297, + 452 + ], + "type": "inline_equation", + "content": "\\mathbf{K}" + }, + { + "bbox": [ + 55, + 332, + 297, + 452 + ], + "type": "text", + "content": " is identical to that of " + }, + { + "bbox": [ + 55, + 332, + 297, + 452 + ], + "type": "inline_equation", + "content": "\\mathbf{K}_{T_{\\mathbf{p}}}" + }, + { + "bbox": [ + 55, + 332, + 297, + 452 + ], + "type": "text", + "content": ". By uniformly choosing " + }, + { + "bbox": [ + 55, + 332, + 297, + 452 + ], + "type": "inline_equation", + "content": "u_{m}" + }, + { + "bbox": [ + 55, + 332, + 297, + 452 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 332, + 297, + 452 + ], + "type": "inline_equation", + "content": "v_{m}" + }, + { + "bbox": [ + 55, + 332, + 297, + 452 + ], + "type": "text", + "content": " and constructing a regular kernel in the tangent space " + }, + { + "bbox": [ + 55, + 332, + 297, + 452 + ], + "type": "inline_equation", + "content": "T_{\\mathbf{p}}S^{2}" + }, + { + "bbox": [ + 55, + 332, + 297, + 452 + ], + "type": "text", + "content": ", we can perform SMConv with kernels uniformly distributed across the " + }, + { + "bbox": [ + 55, + 332, + 297, + 452 + ], + "type": "inline_equation", + "content": "S^2" + }, + { + "bbox": [ + 55, + 332, + 297, + 452 + ], + "type": "text", + "content": " manifold using exponential mapping, according to (3)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 459, + 241, + 472 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 459, + 241, + 472 + ], + "spans": [ + { + "bbox": [ + 55, + 459, + 241, + 472 + ], + "type": "text", + "content": "3.2. SMConv Inspired Diffusion Model" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 475, + 298, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 475, + 298, + 525 + ], + "spans": [ + { + "bbox": [ + 55, + 475, + 298, + 525 + ], + "type": "text", + "content": "Building upon SMConv, we introduce the SMGD model to achieve high-quality panoramic image generation. The overall architecture of our SMGD model is shown in Fig. 3. The diffusion process can be formulated as" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 531, + 240, + 547 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 531, + 240, + 547 + ], + "spans": [ + { + "bbox": [ + 61, + 531, + 240, + 547 + ], + "type": "interline_equation", + "content": "q (\\mathbf {z} _ {t} | \\mathbf {z} _ {t - 1}) := \\mathcal {N} (\\mathbf {z} _ {t - 1}; \\sqrt {1 - \\beta_ {t}} \\mathbf {z} _ {t - 1}, \\beta_ {t} \\mathbf {I}),", + "image_path": "eedc4ac13ac02adeac9bf73bf2960fa02596da210ff054fc3163cc11be38ab70.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 61, + 548, + 296, + 577 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 548, + 296, + 577 + ], + "spans": [ + { + "bbox": [ + 61, + 548, + 296, + 577 + ], + "type": "interline_equation", + "content": "\\mathbf {z} _ {t - 1} = \\frac {1}{\\sqrt {\\alpha_ {t}}} \\left(\\mathbf {z} _ {t} - \\frac {1 - \\alpha_ {t}}{\\sqrt {1 - \\bar {\\alpha} _ {t}}} \\epsilon_ {\\psi} (\\mathbf {z} _ {t}, t, \\mathbf {T})\\right) + \\sqrt {\\beta_ {t}} \\eta , ^ {(4)}", + "image_path": "1b757f3c8c1e66b766e6aee81457ef9559d85c3d326672b1b787af0638e45c7c.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 582, + 298, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 298, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 298, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 582, + 298, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_{t - 1},\\mathbf{z}_t" + }, + { + "bbox": [ + 55, + 582, + 298, + 714 + ], + "type": "text", + "content": " denote the noisy spherical features with the same dimensionality as the data, at time step " + }, + { + "bbox": [ + 55, + 582, + 298, + 714 + ], + "type": "inline_equation", + "content": "t - 1\\in [0,T]" + }, + { + "bbox": [ + 55, + 582, + 298, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 582, + 298, + 714 + ], + "type": "inline_equation", + "content": "t\\in [0,T]" + }, + { + "bbox": [ + 55, + 582, + 298, + 714 + ], + "type": "text", + "content": " respectively. " + }, + { + "bbox": [ + 55, + 582, + 298, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{T}" + }, + { + "bbox": [ + 55, + 582, + 298, + 714 + ], + "type": "text", + "content": " is the text embedding obtained by CLIP model [31], i.e., from the pretrained CLIP ViT/L-14 model. In addition, " + }, + { + "bbox": [ + 55, + 582, + 298, + 714 + ], + "type": "inline_equation", + "content": "q(\\mathbf{z}_t|\\mathbf{z}_{t - 1})" + }, + { + "bbox": [ + 55, + 582, + 298, + 714 + ], + "type": "text", + "content": " represents the forward process gradually adding noise to the image, " + }, + { + "bbox": [ + 55, + 582, + 298, + 714 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\psi}(\\mathbf{z}_t,t)" + }, + { + "bbox": [ + 55, + 582, + 298, + 714 + ], + "type": "text", + "content": " is the noise predicted by our SMUNet at time step " + }, + { + "bbox": [ + 55, + 582, + 298, + 714 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 582, + 298, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 582, + 298, + 714 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 55, + 582, + 298, + 714 + ], + "type": "text", + "content": " represents the parameters of SMUNet and " + }, + { + "bbox": [ + 55, + 582, + 298, + 714 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 55, + 582, + 298, + 714 + ], + "type": "text", + "content": " denotes a random Gaussian noise. Furthermore, the noise scheduler configuration " + }, + { + "bbox": [ + 55, + 582, + 298, + 714 + ], + "type": "inline_equation", + "content": "\\beta_{t}" + }, + { + "bbox": [ + 55, + 582, + 298, + 714 + ], + "type": "text", + "content": " is a hyperparameter that controls the strength of the noise involved at each time step. The parameter " + }, + { + "bbox": [ + 55, + 582, + 298, + 714 + ], + "type": "inline_equation", + "content": "\\alpha_{t}" + }, + { + "bbox": [ + 55, + 582, + 298, + 714 + ], + "type": "text", + "content": " is defined as" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 331, + 555, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 331, + 555, + 357 + ], + "spans": [ + { + "bbox": [ + 313, + 331, + 555, + 357 + ], + "type": "inline_equation", + "content": "\\alpha_{t} = 1 - \\beta_{t}" + }, + { + "bbox": [ + 313, + 331, + 555, + 357 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 331, + 555, + 357 + ], + "type": "inline_equation", + "content": "\\bar{\\alpha}_{t}" + }, + { + "bbox": [ + 313, + 331, + 555, + 357 + ], + "type": "text", + "content": " is given by " + }, + { + "bbox": [ + 313, + 331, + 555, + 357 + ], + "type": "inline_equation", + "content": "\\bar{\\alpha}_{t} = \\prod_{s=1}^{t} \\alpha_{s}" + }, + { + "bbox": [ + 313, + 331, + 555, + 357 + ], + "type": "text", + "content": ", the default setting within diffusion models." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 357, + 556, + 619 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 357, + 556, + 619 + ], + "spans": [ + { + "bbox": [ + 313, + 357, + 556, + 619 + ], + "type": "text", + "content": "As the core component of the SMGD model, our SMUNet predicts noise at each time step " + }, + { + "bbox": [ + 313, + 357, + 556, + 619 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 357, + 556, + 619 + ], + "type": "text", + "content": " from the noisy spherical feature " + }, + { + "bbox": [ + 313, + 357, + 556, + 619 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_t" + }, + { + "bbox": [ + 313, + 357, + 556, + 619 + ], + "type": "text", + "content": " and facilitates the recovery of the spherical feature " + }, + { + "bbox": [ + 313, + 357, + 556, + 619 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_0" + }, + { + "bbox": [ + 313, + 357, + 556, + 619 + ], + "type": "text", + "content": " on " + }, + { + "bbox": [ + 313, + 357, + 556, + 619 + ], + "type": "inline_equation", + "content": "S^2" + }, + { + "bbox": [ + 313, + 357, + 556, + 619 + ], + "type": "text", + "content": " manifold by subtracting the predicted noise. We employ spherical processor blocks (SPBs), along with circular encoder and decoder blocks based on circular convolution [2] at shallower layers of SMUNet. Then, we apply SME and SMD blocks at deeper layers with middle blocks consisted of spherical manifold residual blocks (SMRBs) and cross-attention layers to integrate text embeddings " + }, + { + "bbox": [ + 313, + 357, + 556, + 619 + ], + "type": "inline_equation", + "content": "\\mathbf{T}" + }, + { + "bbox": [ + 313, + 357, + 556, + 619 + ], + "type": "text", + "content": ". The design of a hybrid architecture to integrate circular convolution at shallower layers with SMConv at deeper layers is motivated by the following considerations. As resolution increases, spherical convolution operation including SMConv, consumes substantially higher GPU memory cost compared to standard convolution [8] due to its bilinear interpolation. Furthermore, in deeper layers of SMUNet where spatial resolution is reduced, the sparse pixel distribution near the poles becomes increasingly pronounced, exacerbating the limitations of standard convolution. Consequently, the proposed hybrid architecture well balances between output quality, computational efficiency, and memory utilization." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 619, + 556, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 619, + 556, + 667 + ], + "spans": [ + { + "bbox": [ + 313, + 619, + 556, + 667 + ], + "type": "text", + "content": "Then, the SME block comprises 2 SMRBs, 2 cross-attention layers, and an SMConv with a stride of 2 for downsampling. For the " + }, + { + "bbox": [ + 313, + 619, + 556, + 667 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 313, + 619, + 556, + 667 + ], + "type": "text", + "content": "-th SME block, the SMRB can be formulated as" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 335, + 670, + 555, + 687 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 670, + 555, + 687 + ], + "spans": [ + { + "bbox": [ + 335, + 670, + 555, + 687 + ], + "type": "interline_equation", + "content": "\\widetilde {\\mathbf {z}} _ {t} ^ {E _ {l}} = \\mathrm {S M C o n v} [ \\mathrm {S M C o n v} [ \\mathbf {z} _ {t} ^ {E _ {l}} ] + g _ {\\psi} (t) ] + \\mathbf {z} _ {t} ^ {E _ {l}}, \\quad (5)", + "image_path": "68267d6c24501a8f8a390ce7ddc6df4fda15614f4fd5b5f30c9d5ebb06fe0c66.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 689, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 556, + 714 + ], + "type": "text", + "content": "where the calculation of SMConv follows (3) and " + }, + { + "bbox": [ + 313, + 689, + 556, + 714 + ], + "type": "inline_equation", + "content": "g_{\\psi}(t)" + }, + { + "bbox": [ + 313, + 689, + 556, + 714 + ], + "type": "text", + "content": " is the learnable time-step embedding function [32]. Moreover," + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5827" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 70, + 296, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 70, + 296, + 277 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 296, + 277 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_t^{E_l}" + }, + { + "bbox": [ + 55, + 70, + 296, + 277 + ], + "type": "text", + "content": " denotes the input, while " + }, + { + "bbox": [ + 55, + 70, + 296, + 277 + ], + "type": "inline_equation", + "content": "\\widetilde{\\mathbf{z}}_t^{E_l}" + }, + { + "bbox": [ + 55, + 70, + 296, + 277 + ], + "type": "text", + "content": " represents the output for this layer. The structure of the SMD block closely mirrors that of the indexing SME block, with the upsampling layer using nearest-neighbor interpolation. Note that the final SME block and the first SMD block omit upsampling and downsampling operations. Moreover, the SPB employs the SMConv combined with a " + }, + { + "bbox": [ + 55, + 70, + 296, + 277 + ], + "type": "inline_equation", + "content": "1\\times 1" + }, + { + "bbox": [ + 55, + 70, + 296, + 277 + ], + "type": "text", + "content": " residual convolutional layer to enhance the transformation between different feature channels. The circular encoder and decoder blocks follow the structure in LDM [32], with all planar convolutions replaced by circular convolution operations [2]. Notably, the circular encoder blocks use average pooling layers for downsampling. In addition to the SMUNet, we propose to employ a panoramic VQGAN [11] with a downsampling factor of 16 as the image encoder and decoder for our SMGD model. In the panoramic VQGAN, all planar convolutions in the layers of decoder are replaced by circular convolutions [2]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 282, + 246, + 295 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 282, + 246, + 295 + ], + "spans": [ + { + "bbox": [ + 55, + 282, + 246, + 295 + ], + "type": "text", + "content": "3.3. Spherical Guided Training Strategy" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 300, + 297, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 300, + 297, + 370 + ], + "spans": [ + { + "bbox": [ + 55, + 300, + 297, + 370 + ], + "type": "text", + "content": "The training process involves two stages: training the panoramic VQGAN, and then training the SMGD model. During the training of the panoramic VQGAN, reconstruction loss " + }, + { + "bbox": [ + 55, + 300, + 297, + 370 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{rec}}" + }, + { + "bbox": [ + 55, + 300, + 297, + 370 + ], + "type": "text", + "content": " [11], commitment loss " + }, + { + "bbox": [ + 55, + 300, + 297, + 370 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{Commit}}" + }, + { + "bbox": [ + 55, + 300, + 297, + 370 + ], + "type": "text", + "content": " [11] and perceptual loss " + }, + { + "bbox": [ + 55, + 300, + 297, + 370 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{Per}}" + }, + { + "bbox": [ + 55, + 300, + 297, + 370 + ], + "type": "text", + "content": " [19] are employed, which can be formulated as" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 380, + 296, + 393 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 380, + 296, + 393 + ], + "spans": [ + { + "bbox": [ + 67, + 380, + 296, + 393 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {V Q G A N}} = \\mathcal {L} _ {\\text {r e c}} + \\lambda_ {\\text {c o m m i t}} \\mathcal {L} _ {\\text {c o m m i t}} + \\lambda_ {\\text {p e r}} \\mathcal {L} _ {\\text {P e r}}, \\tag {6}", + "image_path": "5e624dd7204138325dcacd55b975ac90bc5ac7a5a8aa088e130b18d579996f08.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 400, + 295, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 400, + 295, + 424 + ], + "spans": [ + { + "bbox": [ + 55, + 400, + 295, + 424 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 400, + 295, + 424 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{commit}}" + }, + { + "bbox": [ + 55, + 400, + 295, + 424 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 400, + 295, + 424 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{per}}" + }, + { + "bbox": [ + 55, + 400, + 295, + 424 + ], + "type": "text", + "content": " are hyperparameters to control the weights of different loss components during training." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 424, + 296, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 424, + 296, + 470 + ], + "spans": [ + { + "bbox": [ + 55, + 424, + 296, + 470 + ], + "type": "text", + "content": "During the training of the SMGD model, a spherical loss based on the weighted spherically uniform PSNR (WSPSNR) [50] is employed. The total loss can be formulated as" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 93, + 472, + 296, + 484 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 472, + 296, + 484 + ], + "spans": [ + { + "bbox": [ + 93, + 472, + 296, + 484 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {D i f f}} = \\lambda_ {\\text {M S E}} \\mathcal {L} _ {\\text {M S E}} + \\lambda_ {\\text {S M S E}} \\mathcal {L} _ {\\text {S M S E}}, \\tag {7}", + "image_path": "a5440cf79cbad4328bd9af4c6dd307c6509830ae7391ea5179be16e779b404bc.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 489, + 296, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 489, + 296, + 513 + ], + "spans": [ + { + "bbox": [ + 55, + 489, + 296, + 513 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 489, + 296, + 513 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{MSE}}" + }, + { + "bbox": [ + 55, + 489, + 296, + 513 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 489, + 296, + 513 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{SMSE}}" + }, + { + "bbox": [ + 55, + 489, + 296, + 513 + ], + "type": "text", + "content": " are also balancing hyperparameters. Furthermore, the " + }, + { + "bbox": [ + 55, + 489, + 296, + 513 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{MSE}}" + }, + { + "bbox": [ + 55, + 489, + 296, + 513 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 489, + 296, + 513 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{SMSE}}" + }, + { + "bbox": [ + 55, + 489, + 296, + 513 + ], + "type": "text", + "content": " are given by" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 84, + 520, + 296, + 539 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 84, + 520, + 296, + 539 + ], + "spans": [ + { + "bbox": [ + 84, + 520, + 296, + 539 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {M S E}} = \\mathbb {E} _ {t, \\mathbf {z} _ {0}, \\epsilon} [ | | \\epsilon - \\epsilon_ {\\psi} (\\mathbf {z} _ {t}, t, \\mathbf {T}) | | _ {2} ^ {2} ], \\tag {8}", + "image_path": "a9bc960628fde7e1ccea5971dc1413d44bf3a3920034d839bbefb10282a7be3f.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 82, + 536, + 270, + 550 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 536, + 270, + 550 + ], + "spans": [ + { + "bbox": [ + 82, + 536, + 270, + 550 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {S M S E}} = \\mathbb {E} _ {t, \\mathbf {z} _ {0}, \\epsilon} [ | | \\mathbf {W} \\odot (\\epsilon - \\epsilon_ {\\psi} (\\mathbf {z} _ {t}, t, \\mathbf {T})) | | _ {2} ^ {2} ],", + "image_path": "1021c702cd4963264d18b59c7d402a3c70cde2e8aaf075aa2149cfc1344dd300.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 556, + 295, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 556, + 295, + 581 + ], + "spans": [ + { + "bbox": [ + 55, + 556, + 295, + 581 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 556, + 295, + 581 + ], + "type": "inline_equation", + "content": "\\mathbf{W}" + }, + { + "bbox": [ + 55, + 556, + 295, + 581 + ], + "type": "text", + "content": " is given by the WS-PSNR [50] to reflect the panoramic characteristics as" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 111, + 587, + 296, + 611 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 587, + 296, + 611 + ], + "spans": [ + { + "bbox": [ + 111, + 587, + 296, + 611 + ], + "type": "interline_equation", + "content": "\\mathbf {W} _ {j} = \\cos \\left(\\left(\\frac {j + 0 . 5}{H} - 0. 5\\right) \\pi\\right), \\tag {9}", + "image_path": "217335cca5a91b7b9a9afa96e13bf6febc7062ffdad79d054a2fbac7ce16b442.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 617, + 296, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 296, + 665 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 296, + 665 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 617, + 296, + 665 + ], + "type": "inline_equation", + "content": "j \\in [0, H]" + }, + { + "bbox": [ + 55, + 617, + 296, + 665 + ], + "type": "text", + "content": " is the index of height in " + }, + { + "bbox": [ + 55, + 617, + 296, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_t" + }, + { + "bbox": [ + 55, + 617, + 296, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 617, + 296, + 665 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 55, + 617, + 296, + 665 + ], + "type": "text", + "content": " denotes the height of the spherical feature " + }, + { + "bbox": [ + 55, + 617, + 296, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_t" + }, + { + "bbox": [ + 55, + 617, + 296, + 665 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 55, + 617, + 296, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{SMSE}}" + }, + { + "bbox": [ + 55, + 617, + 296, + 665 + ], + "type": "text", + "content": " can improve the preservation of spherical geometry, e.g., a sparse pixel distribution near the poles." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 672, + 274, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 672, + 274, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 672, + 274, + 685 + ], + "type": "text", + "content": "3.4. Quality Evaluation for Panoramic Images" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": "FID score is the de facto standard for evaluating the quality of generated images, whereas performing a direct calculation" + } + ] + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 317, + 72, + 553, + 152 + ], + "blocks": [ + { + "bbox": [ + 317, + 72, + 553, + 152 + ], + "lines": [ + { + "bbox": [ + 317, + 72, + 553, + 152 + ], + "spans": [ + { + "bbox": [ + 317, + 72, + 553, + 152 + ], + "type": "image", + "image_path": "5a7d36517b32d90b987b488e8262b38e73b5d03ad968704942cd54fb3cd45c24.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 160, + 555, + 216 + ], + "lines": [ + { + "bbox": [ + 313, + 160, + 555, + 216 + ], + "spans": [ + { + "bbox": [ + 313, + 160, + 555, + 216 + ], + "type": "text", + "content": "Figure 4. Illustration of the conversion between ERP and CMP formats. In the CMP format, four blue-bordered images represent equatorial regions, while two orange-bordered images represent polar areas. These images are grouped accordingly, with " + }, + { + "bbox": [ + 313, + 160, + 555, + 216 + ], + "type": "inline_equation", + "content": "\\mathrm{FID}_{\\mathrm{equ}}" + }, + { + "bbox": [ + 313, + 160, + 555, + 216 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 160, + 555, + 216 + ], + "type": "inline_equation", + "content": "\\mathrm{FID}_{\\mathrm{pole}}" + }, + { + "bbox": [ + 313, + 160, + 555, + 216 + ], + "type": "text", + "content": " calculated separately for each group." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 221, + 556, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 221, + 556, + 510 + ], + "spans": [ + { + "bbox": [ + 313, + 221, + 556, + 510 + ], + "type": "text", + "content": "of FID on panoramic images in rectangular ERP format introduces severe distortion, leading to inaccuracy. We propose to convert the generated panoramic images from ERP format to CMP format consisted of 6 square images for calculating FID score. The mapping from ERP to CMP format [50, 52] provides a robust solution to evaluate the quality of generated panoramic images. Note that although sharing similar goals with the concurrent work [6], our FID-based metrics are able to reflect detailed quality for different regions when evaluating generated panoramic images. As illustrated in Fig. 4, the conversion reveals the continuity across the left and right boundaries and represents the spherical geometry by obtaining 4 images representing areas near the equator and 2 images representing areas near the poles. Thus, the 6 square images within CMP format can be categorized into 2 groups, allowing for separate FID calculations to assess the quality of generated images. More specifically, " + }, + { + "bbox": [ + 313, + 221, + 556, + 510 + ], + "type": "inline_equation", + "content": "\\mathrm{FID}_{\\mathrm{equ}}" + }, + { + "bbox": [ + 313, + 221, + 556, + 510 + ], + "type": "text", + "content": " is calculated using the 4 images near the equator, while " + }, + { + "bbox": [ + 313, + 221, + 556, + 510 + ], + "type": "inline_equation", + "content": "\\mathrm{FID}_{\\mathrm{pole}}" + }, + { + "bbox": [ + 313, + 221, + 556, + 510 + ], + "type": "text", + "content": " is derived from the 2 images near the poles. " + }, + { + "bbox": [ + 313, + 221, + 556, + 510 + ], + "type": "inline_equation", + "content": "\\mathrm{FID}_{\\mathrm{equ}}" + }, + { + "bbox": [ + 313, + 221, + 556, + 510 + ], + "type": "text", + "content": " primarily assesses the quality and visual fidelity of the panoramic content, whereas " + }, + { + "bbox": [ + 313, + 221, + 556, + 510 + ], + "type": "inline_equation", + "content": "\\mathrm{FID}_{\\mathrm{pole}}" + }, + { + "bbox": [ + 313, + 221, + 556, + 510 + ], + "type": "text", + "content": " evaluates the preservation of spherical geometry. Analyzing " + }, + { + "bbox": [ + 313, + 221, + 556, + 510 + ], + "type": "inline_equation", + "content": "\\mathrm{FID}_{\\mathrm{equ}}" + }, + { + "bbox": [ + 313, + 221, + 556, + 510 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 221, + 556, + 510 + ], + "type": "inline_equation", + "content": "\\mathrm{FID}_{\\mathrm{pole}}" + }, + { + "bbox": [ + 313, + 221, + 556, + 510 + ], + "type": "text", + "content": " provides a comprehensive measure of the quality of the generated panoramic image." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 314, + 521, + 390, + 533 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 521, + 390, + 533 + ], + "spans": [ + { + "bbox": [ + 314, + 521, + 390, + 533 + ], + "type": "text", + "content": "4. Experiment" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 540, + 440, + 553 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 540, + 440, + 553 + ], + "spans": [ + { + "bbox": [ + 313, + 540, + 440, + 553 + ], + "type": "text", + "content": "4.1. Experimental Settings" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 557, + 556, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 557, + 556, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 557, + 556, + 665 + ], + "type": "text", + "content": "Dataset. We employed the widely-used Matterport3D [4] dataset to evaluate the performance for text-conditioned panoramic image generation. More specifically, we obtained 10,912 panoramic images with a resolution of " + }, + { + "bbox": [ + 313, + 557, + 556, + 665 + ], + "type": "inline_equation", + "content": "1024 \\times 512" + }, + { + "bbox": [ + 313, + 557, + 556, + 665 + ], + "type": "text", + "content": " according to [26]. We then utilized the BLIP-2 [25] model to generate text descriptions for each panoramic image, forming paired text-image data for training and testing. To enable FID-10K calculation, we allocated 10,000 images for training and testing." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 665, + 556, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 556, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 556, + 712 + ], + "type": "text", + "content": "Implementation Details. We adopted the panoramic VQGAN according to Sec. 3.2. SMUNet, the denoising network for our SMGD model, was configured with 320 and 640 channels for the two circular encoder and decoder blocks," + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5828" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 107, + 553, + 184 + ], + "blocks": [ + { + "bbox": [ + 55, + 65, + 555, + 100 + ], + "lines": [ + { + "bbox": [ + 55, + 65, + 555, + 100 + ], + "spans": [ + { + "bbox": [ + 55, + 65, + 555, + 100 + ], + "type": "text", + "content": "Table 1. Comparisons in terms of FID and CLIP score (denoted as CS) values for panoramic images generated by our method and existing baseline methods, along with evaluations of parameter number (denoted by Para.) and time required to inference (denoted as " + }, + { + "bbox": [ + 55, + 65, + 555, + 100 + ], + "type": "inline_equation", + "content": "t_{\\mathrm{sample}}" + }, + { + "bbox": [ + 55, + 65, + 555, + 100 + ], + "type": "text", + "content": "). We represent the best numbers by red color and the second best by blue color." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 107, + 553, + 184 + ], + "lines": [ + { + "bbox": [ + 58, + 107, + 553, + 184 + ], + "spans": [ + { + "bbox": [ + 58, + 107, + 553, + 184 + ], + "type": "table", + "html": "
MethodsFID ↓FAED ↓OmniFID ↓FIDavg ↓FIDcent ↓FIDbord ↓FIDrand ↓FIDequ ↓FIDpole ↓CS ↑Para. ↓t_sample(s) ↓
Multi-Diffusion [3]69.961.1268.7056.7458.1956.5355.5027.2687.730.1821.3B39.45 ± 0.31
SD2+LoRA [18, 32]59.321.4769.0750.0451.7849.1649.1724.4988.680.1711.4B6.48 ± 0.49
Text2Light [5]48.424.1063.1643.6846.9346.6437.4730.8772.060.1881.0B50.12 ± 11.18
MVDiffusion [39]56.210.9394.5338.1538.8837.6037.9627.58125.480.1811.4B97.14 ± 0.79
PanFusion [53]16.150.3527.6414.3414.9113.9614.1414.8131.990.1822.3B38.51 ± 11.57
SMGD (Ours)9.850.2218.249.709.779.629.717.8821.050.1891.4B2.69 ± 0.39
", + "image_path": "20c700be3c8ab732b6535f2cffd2c0bd373656b694f65549642224e23339a65f.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 203, + 297, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 203, + 297, + 348 + ], + "spans": [ + { + "bbox": [ + 54, + 203, + 297, + 348 + ], + "type": "text", + "content": "and 960 and 1280 channels for the two SME and SMD blocks utilizing the " + }, + { + "bbox": [ + 54, + 203, + 297, + 348 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 54, + 203, + 297, + 348 + ], + "type": "text", + "content": " SMConv. At the first training stage, we trained the panoramic VQGAN with the Adam optimizer [22], using a batch size of 1 and a learning rate of " + }, + { + "bbox": [ + 54, + 203, + 297, + 348 + ], + "type": "inline_equation", + "content": "4.5 \\times 10^{-6}" + }, + { + "bbox": [ + 54, + 203, + 297, + 348 + ], + "type": "text", + "content": " for 300 epochs, with " + }, + { + "bbox": [ + 54, + 203, + 297, + 348 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{commit}} = 1" + }, + { + "bbox": [ + 54, + 203, + 297, + 348 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 203, + 297, + 348 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{per}} = 0.8" + }, + { + "bbox": [ + 54, + 203, + 297, + 348 + ], + "type": "text", + "content": ". At the second stage, we trained the SMGD model from the scratch with the AdamW optimizer [27], a batch size of 8, and learning rate of " + }, + { + "bbox": [ + 54, + 203, + 297, + 348 + ], + "type": "inline_equation", + "content": "1 \\times 10^{-5}" + }, + { + "bbox": [ + 54, + 203, + 297, + 348 + ], + "type": "text", + "content": " for 800 epochs by setting " + }, + { + "bbox": [ + 54, + 203, + 297, + 348 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{MSE}} = 0.8" + }, + { + "bbox": [ + 54, + 203, + 297, + 348 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 203, + 297, + 348 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{SMSE}} = 0.2" + }, + { + "bbox": [ + 54, + 203, + 297, + 348 + ], + "type": "text", + "content": ". Moreover, the noise scheduler configuration " + }, + { + "bbox": [ + 54, + 203, + 297, + 348 + ], + "type": "inline_equation", + "content": "\\beta_{t}" + }, + { + "bbox": [ + 54, + 203, + 297, + 348 + ], + "type": "text", + "content": " followed that of LDM [32]. For inference, we used a DDIM sampler [35] with 50 steps and a classifier-free guidance scale [32] of 2.5." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 357, + 298, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 357, + 298, + 464 + ], + "spans": [ + { + "bbox": [ + 54, + 357, + 298, + 464 + ], + "type": "text", + "content": "Baselines. We compared our SMGD model with existing text-conditioned panoramic image generation methods, including Multi-Diffusion [3], Text2Light [5], MVDiffusion [39] and PanFusion [53]. We also compared our SMGD model with the widely-used stable diffusion model [32] by performing a rank-16 LoRA [18] (denoted as SD2+LoRA). Different from [53], we directly trained the LoRA model with a learning rate of " + }, + { + "bbox": [ + 54, + 357, + 298, + 464 + ], + "type": "inline_equation", + "content": "1 \\times 10^{-6}" + }, + { + "bbox": [ + 54, + 357, + 298, + 464 + ], + "type": "text", + "content": " for 50 epochs on our training data." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 474, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 474, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 56, + 474, + 297, + 714 + ], + "type": "text", + "content": "Evaluation Metrics. We computed " + }, + { + "bbox": [ + 56, + 474, + 297, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{FID}_{\\mathrm{equ}}" + }, + { + "bbox": [ + 56, + 474, + 297, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 56, + 474, + 297, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{FID}_{\\mathrm{pole}}" + }, + { + "bbox": [ + 56, + 474, + 297, + 714 + ], + "type": "text", + "content": ", as described in Sec. 3.4 to effectively assess the quality of panoramic images. Note that " + }, + { + "bbox": [ + 56, + 474, + 297, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{FID}_{\\mathrm{equ}}" + }, + { + "bbox": [ + 56, + 474, + 297, + 714 + ], + "type": "text", + "content": " was calculated against 40K images and " + }, + { + "bbox": [ + 56, + 474, + 297, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{FID}_{\\mathrm{pole}}" + }, + { + "bbox": [ + 56, + 474, + 297, + 714 + ], + "type": "text", + "content": " was calculated against 20K images. To further validate our metrics and assess the quality of the generated images, we calculated an extra FID value similar to that used in previous methods [53]. More specifically, we cropped the panoramic images into 3 representative areas instead of obtaining 20 perspective images in calculating FID values. Specifically, we cropped the panoramic images into the center, cross-border and random patches with size " + }, + { + "bbox": [ + 56, + 474, + 297, + 714 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 56, + 474, + 297, + 714 + ], + "type": "text", + "content": ". Then, we calculated the corresponding FID value of them and obtained the " + }, + { + "bbox": [ + 56, + 474, + 297, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{FID}_{\\mathrm{cent}}" + }, + { + "bbox": [ + 56, + 474, + 297, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 56, + 474, + 297, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{FID}_{\\mathrm{bord}}" + }, + { + "bbox": [ + 56, + 474, + 297, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 56, + 474, + 297, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{FID}_{\\mathrm{rand}}" + }, + { + "bbox": [ + 56, + 474, + 297, + 714 + ], + "type": "text", + "content": ". By averaging these three FID values calculated against 10K images, we obtained " + }, + { + "bbox": [ + 56, + 474, + 297, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{FID}_{\\mathrm{avg}}" + }, + { + "bbox": [ + 56, + 474, + 297, + 714 + ], + "type": "text", + "content": ". Moreover, we reported FID, OmniFID [6], and Fréchet Auto-Encoder Distance (FAED) [53] to comprehensively evaluate the quality of the generated full panoramic images. On the other hand, to evaluate the alignment between text and generated panoramic images, we utilized the widely adopted CLIP score [16] (denoted" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 203, + 555, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 203, + 555, + 239 + ], + "spans": [ + { + "bbox": [ + 313, + 203, + 555, + 239 + ], + "type": "text", + "content": "as CS).1 Furthermore, we recorded the time for inference " + }, + { + "bbox": [ + 313, + 203, + 555, + 239 + ], + "type": "inline_equation", + "content": "t_{\\mathrm{sample}}" + }, + { + "bbox": [ + 313, + 203, + 555, + 239 + ], + "type": "text", + "content": " by averaging the time required to generate 10 images on an NVIDIA 4090 GPU for each method." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 246, + 503, + 258 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 246, + 503, + 258 + ], + "spans": [ + { + "bbox": [ + 313, + 246, + 503, + 258 + ], + "type": "text", + "content": "4.2. Comparison with Existing Methods" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 263, + 556, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 263, + 556, + 538 + ], + "spans": [ + { + "bbox": [ + 313, + 263, + 556, + 538 + ], + "type": "text", + "content": "Quantitative Results. In Table 1, we report the quantitative results of our method and baselines for panoramic image generation from text descriptions. From this table, our method consistently achieves the lowest FID values, the highest CLIP scores, and the least sampling times. More importantly, among the top-performance models, our method maintains a comparatively lower parameter number. More specifically, the lower values of " + }, + { + "bbox": [ + 313, + 263, + 556, + 538 + ], + "type": "inline_equation", + "content": "\\mathrm{FID}_{\\mathrm{cent}}" + }, + { + "bbox": [ + 313, + 263, + 556, + 538 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 263, + 556, + 538 + ], + "type": "inline_equation", + "content": "\\mathrm{FID}_{\\mathrm{bord}}" + }, + { + "bbox": [ + 313, + 263, + 556, + 538 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 263, + 556, + 538 + ], + "type": "inline_equation", + "content": "\\mathrm{FID}_{\\mathrm{rand}}" + }, + { + "bbox": [ + 313, + 263, + 556, + 538 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 263, + 556, + 538 + ], + "type": "inline_equation", + "content": "\\mathrm{FID}_{\\mathrm{avg}}" + }, + { + "bbox": [ + 313, + 263, + 556, + 538 + ], + "type": "text", + "content": " demonstrate that our SMGD model achieves the state-of-the-art performance within text-conditioned panoramic image generation2. Moreover, the even lower " + }, + { + "bbox": [ + 313, + 263, + 556, + 538 + ], + "type": "inline_equation", + "content": "\\mathrm{FID}_{\\mathrm{equu}}" + }, + { + "bbox": [ + 313, + 263, + 556, + 538 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 263, + 556, + 538 + ], + "type": "inline_equation", + "content": "\\mathrm{FID}_{\\mathrm{pole}}" + }, + { + "bbox": [ + 313, + 263, + 556, + 538 + ], + "type": "text", + "content": " values indicate that our method well preserves the spherical geometry of panoramic images, compared to other baseline methods. Note that MVDiffusion [39] and SD2+LoRA achieved low " + }, + { + "bbox": [ + 313, + 263, + 556, + 538 + ], + "type": "inline_equation", + "content": "\\mathrm{FID}_{\\mathrm{avg}}" + }, + { + "bbox": [ + 313, + 263, + 556, + 538 + ], + "type": "text", + "content": " values but high " + }, + { + "bbox": [ + 313, + 263, + 556, + 538 + ], + "type": "inline_equation", + "content": "\\mathrm{FID}_{\\mathrm{pole}}" + }, + { + "bbox": [ + 313, + 263, + 556, + 538 + ], + "type": "text", + "content": " values. This may be due to the fact that images generated by MVDiffusion exhibited a limited field of view, while images produced by SD2+LoRA lacked the preservation of the spherical geometry in panoramic images. Furthermore, as shown in Table 1, our SMGD model achieves the fastest inference time among all comparison models, outperforming PanFusion [53], the second-best method in terms of image quality, by more than a factor of 10 in inference speed." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 538, + 557, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 538, + 557, + 658 + ], + "spans": [ + { + "bbox": [ + 313, + 538, + 557, + 658 + ], + "type": "text", + "content": "Qualitative Results. We randomly chose 2 text descriptions and report the generated panoramic images with corresponding real images from dataset in Fig. 5. As illustrated in the figure, the panoramic images generated by our SMGD model achieved the best subjective quality and the most accurate alignment with the text descriptions. More specifically, improper spherical geometry, including inconsistency across the border and an unnatural pixel distribution near the poles, was observed in panoramic images generated by SD2+LoRA, Text2Light, and Multi-Diffusion, due to their" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 663, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 313, + 663, + 555, + 684 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 663, + 555, + 684 + ], + "spans": [ + { + "bbox": [ + 313, + 663, + 555, + 684 + ], + "type": "text", + "content": "1The CLIP score was computed using the ViT-L/14 CLIP model, consistent with the CLIP model employed in our SMGD model." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 315, + 685, + 555, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 685, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 315, + 685, + 555, + 713 + ], + "type": "text", + "content": "To further validate the robustness of our model, we conduct additional evaluations on the Structured3D [56] dataset in Supplementary Sec. A. We also provide additional qualitative results in Supplementary Sec. C." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5829" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 89, + 71, + 513, + 594 + ], + "blocks": [ + { + "bbox": [ + 89, + 71, + 513, + 594 + ], + "lines": [ + { + "bbox": [ + 89, + 71, + 513, + 594 + ], + "spans": [ + { + "bbox": [ + 89, + 71, + 513, + 594 + ], + "type": "image", + "image_path": "2e1faf734cdc582f4c64069394f0b38288bb2aba3b85970453cfe576b5f0379c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 600, + 555, + 654 + ], + "lines": [ + { + "bbox": [ + 55, + 600, + 555, + 654 + ], + "spans": [ + { + "bbox": [ + 55, + 600, + 555, + 654 + ], + "type": "text", + "content": "Figure 5. Qualitative results for text-conditioned panoramic image generation with real images from dataset displayed on the top row. We also show the copy from the left side of the panoramic image to the right to reveal the consistency across the whole image. Note that images generated by MVDiffusion [39] exhibit a limited field of view, resulting in images with a reduced height. We highlight the improper spherical geometry, artifact regions and distorted objects and lines with corresponding color of boxes, which are mitigated by our SMGD model." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 662, + 297, + 711 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 662, + 297, + 711 + ], + "spans": [ + { + "bbox": [ + 55, + 662, + 297, + 711 + ], + "type": "text", + "content": "neglect of spherical distortions. Panoramic images generated by Panfusion and MVDiffusion suffered from artifact regions and distorted objects and lines as highlighted with corresponding color boxes in Fig. 5. Moreover, images" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 313, + 662, + 556, + 711 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 662, + 556, + 711 + ], + "spans": [ + { + "bbox": [ + 313, + 662, + 556, + 711 + ], + "type": "text", + "content": "generated by MVDiffusion exhibited a limited field of view. This limitation impeded the visual quality and completeness of the generated images, making them less suitable for fully immersive visual experiences. In contrast, our SMGD" + } + ] + } + ], + "index": 3 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5830" + } + ] + } + ], + "index": 4 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 99, + 553, + 178 + ], + "blocks": [ + { + "bbox": [ + 55, + 56, + 555, + 89 + ], + "lines": [ + { + "bbox": [ + 55, + 56, + 555, + 89 + ], + "spans": [ + { + "bbox": [ + 55, + 56, + 555, + 89 + ], + "type": "text", + "content": "Table 2. Ablation study for our SMGD model. We compare the performance achieved by adding individual components to the baseline model sequentially, including the spherical loss (denoted by " + }, + { + "bbox": [ + 55, + 56, + 555, + 89 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{SMSE}}" + }, + { + "bbox": [ + 55, + 56, + 555, + 89 + ], + "type": "text", + "content": "), the SMD blocks, the SME blocks, and the SPB modules. We represent the best numbers by red color." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 99, + 553, + 178 + ], + "lines": [ + { + "bbox": [ + 58, + 99, + 553, + 178 + ], + "spans": [ + { + "bbox": [ + 58, + 99, + 553, + 178 + ], + "type": "table", + "html": "
\\( \\mathcal{L}_{\\text{SMSE}} \\)SMD BlocksSME BlocksSPBFID\\( \\mathrm{FID}_{\\text{avg}} \\downarrow \\)\\( \\mathrm{FID}_{\\text{cent}} \\downarrow \\)\\( \\mathrm{FID}_{\\text{bord}} \\downarrow \\)\\( \\mathrm{FID}_{\\text{rand}} \\downarrow \\)\\( \\mathrm{FID}_{\\text{equ}} \\downarrow \\)\\( \\mathrm{FID}_{\\text{pole}} \\downarrow \\)CS ↑
××××23.1120.5720.3420.9120.4515.3534.870.187
✓×××20.0317.5017.0518.0417.4012.7832.950.187
✓✓××13.3512.8712.7912.9412.889.6120.450.188
✓✓✓×12.3611.4211.4611.6011.199.2121.700.186
✓✓✓✓9.859.709.779.629.717.8821.050.189
", + "image_path": "aa848cbcf9215c6b6a4660b0bcaa3ee54d5d887bd240233a2af47e8ae427b6c7.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 58, + 201, + 291, + 326 + ], + "blocks": [ + { + "bbox": [ + 108, + 189, + 234, + 199 + ], + "lines": [ + { + "bbox": [ + 108, + 189, + 234, + 199 + ], + "spans": [ + { + "bbox": [ + 108, + 189, + 234, + 199 + ], + "type": "text", + "content": "Comparison in terms of FID and CS" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 58, + 201, + 291, + 326 + ], + "lines": [ + { + "bbox": [ + 58, + 201, + 291, + 326 + ], + "spans": [ + { + "bbox": [ + 58, + 201, + 291, + 326 + ], + "type": "image", + "image_path": "67d77b24560738f886b2cb57ef64d73c118993529795ed41f55e20a561613c8f.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 57, + 329, + 179, + 486 + ], + "blocks": [ + { + "bbox": [ + 57, + 329, + 179, + 486 + ], + "lines": [ + { + "bbox": [ + 57, + 329, + 179, + 486 + ], + "spans": [ + { + "bbox": [ + 57, + 329, + 179, + 486 + ], + "type": "image", + "image_path": "c20ba096886dc23931c46ebd905b36949f02da1b3b3c82f77ecd5def95d75dc0.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 494, + 297, + 559 + ], + "lines": [ + { + "bbox": [ + 55, + 494, + 297, + 559 + ], + "spans": [ + { + "bbox": [ + 55, + 494, + 297, + 559 + ], + "type": "text", + "content": "Figure 6. Comparison between LDM+SphereNet and the proposed SMGD model. Top: Bar plot of FID and CLIP score values achieved by these two methods. Bottom: Subjective results of panoramic images generated by the text prompt \"a hallway in a house\". The comparison is conducted in both ERP and CMP formats, with distorted objects and lines highlighted." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 181, + 331, + 293, + 486 + ], + "blocks": [ + { + "bbox": [ + 181, + 331, + 293, + 486 + ], + "lines": [ + { + "bbox": [ + 181, + 331, + 293, + 486 + ], + "spans": [ + { + "bbox": [ + 181, + 331, + 293, + 486 + ], + "type": "image", + "image_path": "25dc8097b3ba681eff062dbc2a98e608db193922ccd75e1b2131cc944d1d2f31.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 567, + 297, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 567, + 297, + 592 + ], + "spans": [ + { + "bbox": [ + 55, + 567, + 297, + 592 + ], + "type": "text", + "content": "model produced visually pleasing panoramic images with a well-defined overall structure and proper spherical geometry." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 600, + 149, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 600, + 149, + 612 + ], + "spans": [ + { + "bbox": [ + 55, + 600, + 149, + 612 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 618, + 297, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 297, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 297, + 713 + ], + "type": "text", + "content": "In this section, we conducted ablation study to systematically evaluate the impact of core components in our SMGD model. More specifically, since the spherical loss and SMConv are important for handling spherical distortions and preserving spherical geometry in panoramic images, we analyzed performance changes by gradually adding the spherical loss " + }, + { + "bbox": [ + 55, + 618, + 297, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{SMSE}}" + }, + { + "bbox": [ + 55, + 618, + 297, + 713 + ], + "type": "text", + "content": ", the SMD blocks, the SME blocks and the SPB. As reported in Table 2, the involvement of SMD blocks" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 189, + 555, + 259 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 189, + 555, + 259 + ], + "spans": [ + { + "bbox": [ + 313, + 189, + 555, + 259 + ], + "type": "text", + "content": "significantly enhanced performance, with SME blocks further contributing to the improvements. Moreover, both the spherical loss function and the SPB effectively improved the quality of generated panoramic images. Therefore, each module plays a vital role in optimizing the performance of the SMGD model." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 261, + 556, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 261, + 556, + 416 + ], + "spans": [ + { + "bbox": [ + 313, + 261, + 556, + 416 + ], + "type": "text", + "content": "Furthermore, to validate the effectiveness of SMConv compared to previous spherical convolution methods, we implemented LDM+SphereNet, which employs the panoramic VQGAN and a U-Net architecture identical to our SMUNet, whilst all SMConv layers were replaced by the convolution operations proposed in SphereNet [8]. The training and inference were kept consistent with those of our SMGD model. As shown in Fig. 6, we can conclude that the quality of panoramic images generated by LDM+SphereNet is limited, whereas our proposed SMGD model produces visually pleasing images with lower FID values and a higher CLIP score3. This demonstrates the effectiveness of using SMConv for advanced panoramic image generation." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 426, + 388, + 439 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 426, + 388, + 439 + ], + "spans": [ + { + "bbox": [ + 314, + 426, + 388, + 439 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 447, + 556, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 447, + 556, + 686 + ], + "spans": [ + { + "bbox": [ + 313, + 447, + 556, + 686 + ], + "type": "text", + "content": "In this paper, we have proposed a novel method named as spherical manifold guided diffusion (SMGD) model to address the challenges of spherical distortions associated with panoramic image generation. This was achieved by the spherical manifold convolution (SMConv) operation that is able to uniformly sample spherical content within each convolution layer. Different from existing spherical convolution operations applied in other tasks, the proposed SMConv operation was built upon the exponential mapping grounded in manifold theory, which is able to preserve lengths on the tangent plane and exhibit enhanced local properties. The overall training procedure of our SMGD model was also inspired by the spherical loss. Furthermore, we have proposed to calculate " + }, + { + "bbox": [ + 313, + 447, + 556, + 686 + ], + "type": "inline_equation", + "content": "\\mathrm{FID}_{\\mathrm{equ}}" + }, + { + "bbox": [ + 313, + 447, + 556, + 686 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 447, + 556, + 686 + ], + "type": "inline_equation", + "content": "\\mathrm{FID}_{\\mathrm{pole}}" + }, + { + "bbox": [ + 313, + 447, + 556, + 686 + ], + "type": "text", + "content": " for panoramic images in cubemap format to improve the evaluation on the quality of generated panoramic images. Experimental results have demonstrated that our method achieves superior performance in text-conditioned image generation, surpassing existing state-of-the-art methods in both the quality of generated images and computational efficiency." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 314, + 693, + 555, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 693, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 314, + 693, + 555, + 713 + ], + "type": "text", + "content": "3More qualitative comparison results between SMGD and LDM+SphereNet, as well as a detailed analysis, are provided in Supplementary Sec. B." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "5831" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 297, + 138 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 297, + 138 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 297, + 138 + ], + "type": "text", + "content": "This work was supported by Natural Science Foundation of China (NSFC) under Grants 62450131, 62206011, 62231002 and 62401027, and Beijing Natural Science Foundation under Grant L223021." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 148, + 115, + 162 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 148, + 115, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 148, + 115, + 162 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 168, + 297, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 168, + 297, + 233 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 168, + 297, + 233 + ], + "spans": [ + { + "bbox": [ + 61, + 168, + 297, + 233 + ], + "type": "text", + "content": "[1] Hao Ai, Zidong Cao, Haonan Lu, Chen Chen, Jian Ma, Pengyuan Zhou, Tae-Kyun Kim, Pan Hui, and Lin Wang. Dream360: Diverse and immersive outdoor virtual scene creation via transformer-based " + }, + { + "bbox": [ + 61, + 168, + 297, + 233 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 61, + 168, + 297, + 233 + ], + "type": "text", + "content": " image outpainting. IEEE Transactions on Visualization and Computer Graphics, 2024. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 235, + 296, + 290 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 235, + 296, + 290 + ], + "spans": [ + { + "bbox": [ + 61, + 235, + 296, + 290 + ], + "type": "text", + "content": "[2] Naofumi Akimoto, Yuhi Matsuo, and Yoshimitsu Aoki. Diverse plausible 360-degree image outpainting for efficient 3dcg background creation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11441-11450, 2022. 2, 4, 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 291, + 296, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 291, + 296, + 334 + ], + "spans": [ + { + "bbox": [ + 61, + 291, + 296, + 334 + ], + "type": "text", + "content": "[3] Omer Bar-Tal, Lior Yariv, Yaron Lipman, and Tali Dekel. Multidiffusion: Fusing diffusion paths for controlled image generation. In International Conference on Machine Learning, 2023. 2, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 61, + 335, + 296, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 335, + 296, + 388 + ], + "spans": [ + { + "bbox": [ + 61, + 335, + 296, + 388 + ], + "type": "text", + "content": "[4] Angel Chang, Angela Dai, Thomas Funkhouser, Maciej Halber, Matthias Niessner, Manolis Savva, Shuran Song, Andy Zeng, and Yinda Zhang. Matterport3d: Learning from rgb-d data in indoor environments. arXiv preprint arXiv:1709.06158, 2017. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 390, + 296, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 390, + 296, + 423 + ], + "spans": [ + { + "bbox": [ + 62, + 390, + 296, + 423 + ], + "type": "text", + "content": "[5] Zhaoxi Chen, Guangcong Wang, and Ziwei Liu. Text2light: Zero-shot text-driven hdr panorama generation. ACM Transactions on Graphics (TOG), 41(6):1-16, 2022. 2, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 425, + 296, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 425, + 296, + 479 + ], + "spans": [ + { + "bbox": [ + 62, + 425, + 296, + 479 + ], + "type": "text", + "content": "[6] Anders Christensen, Nooshin Mojab, Khushman Patel, Karan Ahuja, Zeynep Akata, Ole Winther, Mar Gonzalez-Franco, and Andrea Colaco. Geometry fidelity for spherical images. In European Conference on Computer Vision, pages 276-292. Springer, 2024. 5, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 480, + 296, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 480, + 296, + 502 + ], + "spans": [ + { + "bbox": [ + 62, + 480, + 296, + 502 + ], + "type": "text", + "content": "[7] Taco S Cohen, Mario Geiger, Jonas Köhler, and Max Welling. Spherical cnns. arXiv preprint arXiv:1801.10130, 2018. 2, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 502, + 296, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 502, + 296, + 557 + ], + "spans": [ + { + "bbox": [ + 62, + 502, + 296, + 557 + ], + "type": "text", + "content": "[8] Benjamin Coors, Alexandru Paul Condurache, and Andreas Geiger. Spherenet: Learning spherical representations for detection and classification in omnidirectional images. In Proceedings of the European conference on computer vision (ECCV), pages 518-533, 2018. 2, 3, 4, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 558, + 296, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 558, + 296, + 612 + ], + "spans": [ + { + "bbox": [ + 62, + 558, + 296, + 612 + ], + "type": "text", + "content": "[9] Mohammad Reza Karimi Dastjerdi, Yannick Hold-Geoffroy, Jonathan Eisenmann, Siavash Khodadadeh, and Jean-François Lalonde. Guided co-modulated gan for 360 field of view extrapolation. In 2022 International Conference on 3D Vision (3DV), pages 475–485. IEEE, 2022. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 613, + 294, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 613, + 294, + 646 + ], + "spans": [ + { + "bbox": [ + 57, + 613, + 294, + 646 + ], + "type": "text", + "content": "[10] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 647, + 296, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 647, + 296, + 691 + ], + "spans": [ + { + "bbox": [ + 57, + 647, + 296, + 691 + ], + "type": "text", + "content": "[11] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12873-12883, 2021. 2, 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 692, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 692, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 692, + 296, + 713 + ], + "type": "text", + "content": "[12] Clara Fernandez-Labrador, Jose M Facil, Alejandro Perez-Yus, Cédric Demonceaux, Javier Civera, and Jose J Guerrero." + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 333, + 73, + 555, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 555, + 106 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 555, + 106 + ], + "type": "text", + "content": "Corners for layout: End-to-end layout recovery from 360 images. IEEE Robotics and Automation Letters, 5(2):1255-1262, 2020. 2, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 106, + 555, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 106, + 555, + 138 + ], + "spans": [ + { + "bbox": [ + 316, + 106, + 555, + 138 + ], + "type": "text", + "content": "[13] Jean Gallier. Basics of Classical Lie Groups: The Exponential Map, Lie Groups, and Lie Algebras, pages 367-414. Springer New York, New York, NY, 2001. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 140, + 554, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 140, + 554, + 183 + ], + "spans": [ + { + "bbox": [ + 316, + 140, + 554, + 183 + ], + "type": "text", + "content": "[14] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 184, + 555, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 184, + 555, + 228 + ], + "spans": [ + { + "bbox": [ + 316, + 184, + 555, + 228 + ], + "type": "text", + "content": "[15] Takayuki Hara, Yusuke Mukuta, and Tatsuya Harada. Spherical image generation from a single image by considering scene symmetry. Proceedings of the AAAI Conference on Artificial Intelligence, 35(2):1513-1521, 2021. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 228, + 555, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 228, + 555, + 281 + ], + "spans": [ + { + "bbox": [ + 316, + 228, + 555, + 281 + ], + "type": "text", + "content": "[16] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 7514-7528, 2021. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 283, + 555, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 283, + 555, + 336 + ], + "spans": [ + { + "bbox": [ + 316, + 283, + 555, + 336 + ], + "type": "text", + "content": "[17] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 338, + 555, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 338, + 555, + 381 + ], + "spans": [ + { + "bbox": [ + 316, + 338, + 555, + 381 + ], + "type": "text", + "content": "[18] Edward J Hu, Phillip Wallis, Zeyuan Allen-Zhu, Yuzhhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. In International Conference on Learning Representations, 2022. 2, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 383, + 555, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 383, + 555, + 426 + ], + "spans": [ + { + "bbox": [ + 316, + 383, + 555, + 426 + ], + "type": "text", + "content": "[19] Justin Johnson, Alexandre Alahi, and Li Fei-Fei. Perceptual losses for real-time style transfer and super-resolution. In European Conference on Computer Vision, pages 694–711. Springer International Publishing, 2016. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 426, + 554, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 426, + 554, + 480 + ], + "spans": [ + { + "bbox": [ + 316, + 426, + 554, + 480 + ], + "type": "text", + "content": "[20] Minguk Kang, Jun-Yan Zhu, Richard Zhang, Jaesik Park, Eli Shechtman, Sylvain Paris, and Taesung Park. Scaling up gans for text-to-image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10124-10134, 2023. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 482, + 555, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 482, + 555, + 525 + ], + "spans": [ + { + "bbox": [ + 316, + 482, + 555, + 525 + ], + "type": "text", + "content": "[21] Renata Khasanova and Pascal Frossard. Geometry aware convolutional filters for omnidirectional images representation. In International conference on machine learning, pages 3351-3359. PMLR, 2019. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 526, + 555, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 526, + 555, + 548 + ], + "spans": [ + { + "bbox": [ + 316, + 526, + 555, + 548 + ], + "type": "text", + "content": "[22] Diederik P Kingma. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 548, + 555, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 548, + 555, + 591 + ], + "spans": [ + { + "bbox": [ + 317, + 548, + 555, + 591 + ], + "type": "text", + "content": "[23] Yuseung Lee, Kunho Kim, Hyunjin Kim, and Minhyuk Sung. Syncdiffusion: Coherent montage via synchronized joint diffusions. Advances in Neural Information Processing Systems, 36:50648-50660, 2023. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 592, + 555, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 592, + 555, + 635 + ], + "spans": [ + { + "bbox": [ + 316, + 592, + 555, + 635 + ], + "type": "text", + "content": "[24] Jialu Li and Mohit Bansal. Panogen: Text-conditioned panoramic environment generation for vision-and-language navigation. Advances in Neural Information Processing Systems, 36, 2024. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 636, + 555, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 636, + 555, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 636, + 555, + 689 + ], + "type": "text", + "content": "[25] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730-19742. PMLR, 2023. 5" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 691, + 555, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 555, + 713 + ], + "type": "text", + "content": "[26] Chieh Hubert Lin, Chia-Che Chang, Yu-Sheng Chen, Da-Cheng Juan, Wei Wei, and Hwann-Tzong Chen. Coco-gan:" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5832" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 76, + 72, + 296, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 296, + 106 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 296, + 106 + ], + "type": "text", + "content": "Generation by parts via conditional coordinating. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4512-4521, 2019. 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 106, + 296, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 106, + 296, + 128 + ], + "spans": [ + { + "bbox": [ + 56, + 106, + 296, + 128 + ], + "type": "text", + "content": "[27] I Loshchilov. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 129, + 296, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 296, + 183 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 296, + 183 + ], + "type": "text", + "content": "[28] Zhuqiang Lu, Kun Hu, Chaoyue Wang, Lei Bai, and Zhiyong Wang. Autoregressive omni-aware outpainting for open-vocabulary 360-degree image generation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 14211-14219, 2024. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 184, + 296, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 184, + 296, + 227 + ], + "spans": [ + { + "bbox": [ + 56, + 184, + 296, + 227 + ], + "type": "text", + "content": "[29] Thomas W Mitchel, Noam Aigerman, Vladimir G Kim, and Michael Kazhdan. Möbius convolutions for spherical cnns. In ACM SIGGRAPH 2022 Conference Proceedings, pages 1-9, 2022. 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 228, + 296, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 228, + 296, + 293 + ], + "spans": [ + { + "bbox": [ + 56, + 228, + 296, + 293 + ], + "type": "text", + "content": "[30] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 293, + 296, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 293, + 296, + 337 + ], + "spans": [ + { + "bbox": [ + 56, + 293, + 296, + 337 + ], + "type": "text", + "content": "[31] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1(2): 3, 2022. 2, 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 338, + 296, + 393 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 338, + 296, + 393 + ], + "spans": [ + { + "bbox": [ + 56, + 338, + 296, + 393 + ], + "type": "text", + "content": "[32] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684-10695, 2022. 1, 4, 5, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 394, + 296, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 394, + 296, + 459 + ], + "spans": [ + { + "bbox": [ + 56, + 394, + 296, + 459 + ], + "type": "text", + "content": "[33] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in neural information processing systems, 35:36479-36494, 2022. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 460, + 296, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 460, + 296, + 514 + ], + "spans": [ + { + "bbox": [ + 56, + 460, + 296, + 514 + ], + "type": "text", + "content": "[34] Axel Sauer, Tero Karras, Samuli Laine, Andreas Geiger, and Timo Aila. Stylegan-t: Unlocking the power of gans for fast large-scale text-to-image synthesis. In International conference on machine learning, pages 30105-30118. PMLR, 2023. 1, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 514, + 296, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 514, + 296, + 547 + ], + "spans": [ + { + "bbox": [ + 56, + 514, + 296, + 547 + ], + "type": "text", + "content": "[35] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020.6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 548, + 296, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 548, + 296, + 581 + ], + "spans": [ + { + "bbox": [ + 56, + 548, + 296, + 581 + ], + "type": "text", + "content": "[36] Yu-Chuan Su and Kristen Grauman. Learning spherical convolution for fast features from 360 imagery. Advances in neural information processing systems, 30, 2017. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 582, + 296, + 625 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 582, + 296, + 625 + ], + "spans": [ + { + "bbox": [ + 56, + 582, + 296, + 625 + ], + "type": "text", + "content": "[37] Yu-Chuan Su and Kristen Grauman. Kernel transformer networks for compact spherical convolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9442-9451, 2019. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 625, + 296, + 670 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 625, + 296, + 670 + ], + "spans": [ + { + "bbox": [ + 56, + 625, + 296, + 670 + ], + "type": "text", + "content": "[38] Julius Surya Sumantri and In Kyu Park. 360 panorama synthesis from a sparse set of images with unknown field of view. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 2386-2395, 2020. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 670, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 296, + 713 + ], + "type": "text", + "content": "[39] Shitao Tang, Fuyang Zhang, Jiacheng Chen, Peng Wang, and Yasutaka Furukawa. Mvdiffusion: Enabling holistic multiview image generation with correspondence-aware diffusion. ArXiv, abs/2307.01097, 2023. 2, 6, 7" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 316, + 72, + 555, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 555, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 555, + 106 + ], + "type": "text", + "content": "[40] T. Vincenty. Direct and inverse solutions of geodesics on the ellipsoid with application of nested equations. Survey Review, 23(176):88-93, 1975. 2, 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 106, + 555, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 106, + 555, + 160 + ], + "spans": [ + { + "bbox": [ + 316, + 106, + 555, + 160 + ], + "type": "text", + "content": "[41] Hai Wang, Xiaoyu Xiang, Yuchen Fan, and Jing-Hao Xue. Customizing 360-degree pancreas through text-to-image diffusion models. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 4933-4943, 2024. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 161, + 555, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 161, + 555, + 205 + ], + "spans": [ + { + "bbox": [ + 316, + 161, + 555, + 205 + ], + "type": "text", + "content": "[42] Tianhao Wu, Chuanxia Zheng, and Tat-Jen Cham. Panodiffusion: 360-degree panorama outpainting via diffusion. In The Twelfth International Conference on Learning Representations, 2023. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 205, + 554, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 205, + 554, + 259 + ], + "spans": [ + { + "bbox": [ + 316, + 205, + 554, + 259 + ], + "type": "text", + "content": "[43] Tao Wu, Xuewei Li, Zhongang Qi, Di Hu, Xintao Wang, Ying Shan, and Xi Li. Spherediffusion: Spherical geometry-aware distortion resilient diffusion model. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 6126-6134, 2024. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 260, + 555, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 260, + 555, + 304 + ], + "spans": [ + { + "bbox": [ + 316, + 260, + 555, + 304 + ], + "type": "text", + "content": "[44] Mai Xu, Chen Li, Shanyi Zhang, and Patrick Le Callet. State-of-the-art in 360 video/image processing: Perception, assessment and compression. IEEE Journal of Selected Topics in Signal Processing, 14(1):5-26, 2020. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 305, + 554, + 349 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 305, + 554, + 349 + ], + "spans": [ + { + "bbox": [ + 316, + 305, + 554, + 349 + ], + "type": "text", + "content": "[45] Yanyu Xu, Ziheng Zhang, and Shenghua Gao. Spherical dnns and their applications in 360 images and videos. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(10):7235-7252, 2021. 2, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 350, + 555, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 350, + 555, + 392 + ], + "spans": [ + { + "bbox": [ + 316, + 350, + 555, + 392 + ], + "type": "text", + "content": "[46] Jian-Ru Xue, Jian-Wu Fang, and Pu Zhang. A survey of scene understanding by event reasoning in autonomous driving. International Journal of Automation and Computing, 15 (3):249-266, 2018. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 393, + 555, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 393, + 555, + 449 + ], + "spans": [ + { + "bbox": [ + 316, + 393, + 555, + 449 + ], + "type": "text", + "content": "[47] Bangbang Yang, Yinda Zhang, Yijin Li, Zhaopeng Cui, Sean Fanello, Hujun Bao, and Guofeng Zhang. Neural rendering in a room: amodal 3d understanding and free-viewpoint rendering for the closed scene composed of pre-captured objects. ACM Transactions on Graphics (TOG), 41(4):1-10, 2022. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 449, + 555, + 504 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 449, + 555, + 504 + ], + "spans": [ + { + "bbox": [ + 316, + 449, + 555, + 504 + ], + "type": "text", + "content": "[48] Bangbang Yang, Wenqi Dong, Lin Ma, Wenbo Hu, Xiao Liu, Zhaopeng Cui, and Yuewen Ma. Dreamspace: Dreaming your room space with text-driven panoramic texture propagation. In 2024 IEEE Conference Virtual Reality and 3D User Interfaces (VR), pages 650-660. IEEE, 2024. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 504, + 555, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 504, + 555, + 557 + ], + "spans": [ + { + "bbox": [ + 316, + 504, + 555, + 557 + ], + "type": "text", + "content": "[49] Michihiro Yasunaga, Armen Aghajanyan, Weijia Shi, Rich James, Jure Leskovec, Percy Liang, Mike Lewis, Luke Zettlemoyer, and Wen-tau Yih. Retrieval-augmented multimodal language modeling. arXiv preprint arXiv:2211.12561, 2022. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 559, + 555, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 559, + 555, + 602 + ], + "spans": [ + { + "bbox": [ + 316, + 559, + 555, + 602 + ], + "type": "text", + "content": "[50] Yan Ye, Elena Alshina, and Jill Boyce. Algorithm descriptions of projection format conversion and video quality metrics in 360lib. In Joint Video Exploration Team of ITU-T SG, 2018. 1, 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 603, + 555, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 603, + 555, + 658 + ], + "spans": [ + { + "bbox": [ + 316, + 603, + 555, + 658 + ], + "type": "text", + "content": "[51] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, et al. Scaling autoregressive models for content-rich text-to-image generation. arXiv preprint arXiv:2206.10789, 2(3):5, 2022. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 658, + 555, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 555, + 713 + ], + "type": "text", + "content": "[52] Li Yu, Yanjun Gao, Farhad Pakdaman, and Moncef Gabbouj. Panoramic image inpainting with gated convolution and contextual reconstruction loss. In ICASSP 2024 - 2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 4255-4259, 2024. 2, 5" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5833" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 339 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 297, + 137 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 297, + 137 + ], + "type": "text", + "content": "[53] Cheng Zhang, Qianyi Wu, Camilo Cruz Gambardella, Xiaoshui Huang, Dinh Phung, Wanli Ouyang, and Jianfei Cai. Taming stable diffusion for text to 360 panorama image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6347-6357, 2024. 1, 2, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 140, + 295, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 140, + 295, + 194 + ], + "spans": [ + { + "bbox": [ + 56, + 140, + 295, + 194 + ], + "type": "text", + "content": "[54] Qinsheng Zhang, Jiaming Song, Xun Huang, Yongxin Chen, and Ming-Yu Liu. Diffcollage: Parallel generation of large content with diffusion models. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10188-10198. IEEE, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 196, + 295, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 196, + 295, + 228 + ], + "spans": [ + { + "bbox": [ + 56, + 196, + 295, + 228 + ], + "type": "text", + "content": "[55] Qiang Zhao, Chen Zhu, Feng Dai, Yike Ma, Guoqing Jin, and Yongdong Zhang. Distortion-aware cnns for spherical images. In IJCAI, pages 1198-1204, 2018. 2, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 230, + 295, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 230, + 295, + 273 + ], + "spans": [ + { + "bbox": [ + 56, + 230, + 295, + 273 + ], + "type": "text", + "content": "[56] Jia Zheng, Junfei Zhang, Jing Li, Rui Tang, Shenghua Gao, and Zihan Zhou. Structured3d: A large photo-realistic dataset for structured 3d modeling. In Proceedings of The European Conference on Computer Vision (ECCV), 2020. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 275, + 295, + 339 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 275, + 295, + 339 + ], + "spans": [ + { + "bbox": [ + 56, + 275, + 295, + 339 + ], + "type": "text", + "content": "[57] Yufan Zhou, Ruiyi Zhang, Changyou Chen, Chunyuan Li, Chris Tensmeyer, Tong Yu, Jiuming Gu, Jinhui Xu, and Tong Sun. Towards language-free training for text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17907-17917, 2022. 1" + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "5834" + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Spiking Transformer with Spatial-Temporal Attention/de8548d3-fd04-4e18-9d13-559b7231ddb6_content_list.json b/2025/Spiking Transformer with Spatial-Temporal Attention/de8548d3-fd04-4e18-9d13-559b7231ddb6_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..f2125dcba3ed345b4881e37270bc4db6235d5e8c --- /dev/null +++ b/2025/Spiking Transformer with Spatial-Temporal Attention/de8548d3-fd04-4e18-9d13-559b7231ddb6_content_list.json @@ -0,0 +1,1570 @@ +[ + { + "type": "text", + "text": "Spiking Transformer with Spatial-Temporal Attention", + "text_level": 1, + "bbox": [ + 223, + 130, + 772, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Donghyun Lee, Yuhang Li, Youngeun Kim, Shiting Xiao, Priyadarshini Panda Department of Electrical Engineering, Yale University", + "bbox": [ + 187, + 180, + 807, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{donghyun.lee, yuhang.li, youngeun.kim, ginny.xiao, priya.panda}@yale.edu", + "bbox": [ + 176, + 219, + 820, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 268, + 326, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Spike-based Transformer presents a compelling and energy-efficient alternative to traditional Artificial Neural Network (ANN)-based Transformers, achieving impressive results through sparse binary computations. However, existing spike-based transformers predominantly focus on spatial attention while neglecting crucial temporal dependencies inherent in spike-based processing, leading to suboptimal feature representation and limited performance. To address this limitation, we propose Spiking Transformer with Spatial-Temporal Attention (STAtten), a simple and straightforward architecture that efficiently integrates both spatial and temporal information in the self-attention mechanism. STAtten introduces a block-wise computation strategy that processes information in spatial-temporal chunks, enabling comprehensive feature capture while maintaining the same computational complexity as previous spatial-only approaches. Our method can be seamlessly integrated into existing spike-based transformers without architectural overhaul. Extensive experiments demonstrate that STAtten significantly improves the performance of existing spike-based transformers across both static and neuromorphic datasets, including CIFAR10/100, ImageNet, CIFAR10-DVS, and N-Caltech101. The code is available at https://github.com/Intelligent-Computing-Lab-Yale/STAtten.", + "bbox": [ + 88, + 300, + 483, + 679 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 708, + 220, + 724 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Spiking Neural Networks (SNNs) are emerging as a promising alternative to conventional Artificial Neural Networks (ANNs) [35, 40] due to their bio-inspired, energy-efficient computing paradigm. By leveraging sparse binary spike-based computations, SNNs can achieve significant energy savings while enabling deployment across various neuromorphic computing platforms such as TrueNorth [1], Loihi [6], and Tianjic [38]. However, the challenge lies in effectively processing and learning from these discrete, time-dependent spike patterns while maintaining computational efficiency and accuracy. Traditional SNNs have primarily", + "bbox": [ + 89, + 734, + 483, + 901 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c324ea93315e4534ed615b17fe8a2b2544ff9db8591f0479fb70d24ea953bc25.jpg", + "image_caption": [ + "Figure 1. Heatmaps of spatial-only attention versus our STAten on sequential CIFAR100 dataset. Input images are divided column-wise, where each column corresponds to one timestep." + ], + "image_footnote": [], + "bbox": [ + 521, + 271, + 905, + 409 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "relied on convolution-based architectures adapted from successful ANN models like VGGNet [41, 49] and ResNet [20, 60]. While these architectures benefit from computational efficiency, they often struggle with binary spike activations, leading to information loss and degraded performance. This limitation has driven researchers to explore alternative architectures that can better handle sparse, temporal data while maintaining the energy benefits of spike-based computation. Transformer architectures, with their remarkable success across various domains [2, 4, 11, 47], have emerged as a promising direction for enhancing SNN capabilities. Recent works have introduced spike-formed transformers that adapt the transformer's self-attention mechanism to the spike-based domain for diverse tasks, such as, object tracking [58], computer vision [42, 50, 52, 53, 62, 64], depth estimation [59], and speech recognition [45, 48]. These architectures, such as Spikformer [64] and Spike-driven Transformer [53], implement innovative binary representations for query (Q), key (K), and value (V) computations while eliminating the computationally expensive Softmax function. This binary attention design preserves the essential characteristics of spike-based computation while enabling more efficient information processing compared to conventional transformer architectures. However, current spike-formed transformers predominantly rely on spatial-only attention, overlooking the inherently dynamic and temporal nature of spike events. To understand these differ", + "bbox": [ + 511, + 493, + 906, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "13948", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ent attention patterns, we compare spatial-only and spatial-temporal attention on sequential CIFAR100, as shown in Fig. 1. The visualization reveals that spatial-only attention focuses solely on vertical (spatial) relationships within each timestep, missing crucial object features that evolve over time. In contrast, our spatial-temporal attention captures both vertical (spatial) and horizontal (temporal) relationships, enabling comprehensive feature representation of sequential inputs.", + "bbox": [ + 89, + 90, + 480, + 224 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we introduce Spiking Transformer with Spatial-Temporal Attention (STAtten), which integrates both spatial and temporal information within the self-attention mechanism. Through analysis of memory requirements and neuronal activity patterns, we design a blockwise computation strategy with local temporal correlations. By eliminating Softmax operation, we maintain the same computational complexity $(O(TND^{2}))$ as existing spike-based transformers [42, 52, 62, 64]. Furthermore, our approach can be flexibly integrated into existing architectures without structural modifications, enhancing their representational capacity. Experiments on both static and dynamic datasets demonstrate consistent performance improvements over spatial-only attention architectures. The main contributions of our work are as follows:", + "bbox": [ + 91, + 227, + 482, + 450 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We identify limitations in current spiking transformers that rely on spatial-only attention through empirical analysis, demonstrating the importance of capturing temporal dependencies inherent in spike-based processing.", + "- We propose STAtten, a block-wise spatial-temporal attention mechanism for spike-based transformers that maintains the original computational complexity $(O(TND^{2}))$ .", + "- We introduce a flexible plug-and-play design that enables STAtten to be integrated into existing spike-based transformers without compromising their core architecture.", + "- Through extensive experiments across both static (CIFAR10/100, ImageNet) and neuromorphic (CIFAR10-DVS, N-Caltech101) datasets, we demonstrate that STatten consistently improves performance across different backbone architectures." + ], + "bbox": [ + 89, + 454, + 480, + 694 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 89, + 709, + 240, + 724 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Training Strategies for SNNs. Significant progress has been made in developing effective training strategies for SNNs, primarily following two approaches: ANN-to-SNN conversion and direct training. The conversion approach transforms pre-trained ANNs into SNNs, leveraging established ANN architectures to enhance SNN performance. Several studies [8, 10, 17, 30] have formalized this conversion process and demonstrated its effectiveness in transferring ANN knowledge to the spike domain. Direct training through backpropagation, while fundamental to ANNs, presents unique challenges in SNNs due to the", + "bbox": [ + 89, + 734, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "non-differentiable nature of spike events. The introduction of surrogate gradients has been crucial in enabling effective backpropagation-based training in SNNs, as demonstrated in pioneering works [5, 43, 49]. Notably, direct training with surrogate gradients has shown superior performance compared to conversion methods across various tasks, including image classification [14, 20, 31, 55, 57], semantic segmentation [23], and object detection [44], while maintaining the energy efficiency inherent to spike-based computation.", + "bbox": [ + 511, + 90, + 903, + 241 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Spiking Transformer. Despite the energy efficiency of convolution-based SNNs, their performance often lags behind ANNs due to information loss from sparse binary activations. To address this limitation, researchers have adapted transformer architectures to the SNN domain, leveraging their powerful self-attention mechanisms and ability to capture long-range dependencies. This adaptation has led to several innovative spike-formed transformer architectures. Spikformer [64] introduced the first SNN-based transformer architecture, establishing foundational principles for spikeformed self-attention mechanisms. Their key innovation lies in eliminating the Softmax function from self-attention computations, justified by two key observations: (1) the binary nature of spike-based Q, K, and V values, and (2) the redundancy of Softmax scaling for binarized operations. Building upon this foundation, Spike-driven Transformer [53] enhanced the architecture with a novel self-attention block incorporating masking and sparse addition to optimize power efficiency. Additionally, inspired by MS-ResNet [20], Spike-driven Transformer reimagined residual connections to propagate membrane potential rather than spikes. Recent works have further extended these concepts, leading to architectures like SpikingReformer [42], QK-Former [62], and Spike-driven Transformer-V2 [52], each introducing unique optimizations for spike-based processing. However, existing spiking transformer architectures primarily focus on spatial attention, neglecting the temporal dynamics inherent in spike-based computation.", + "bbox": [ + 511, + 243, + 906, + 664 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Preliminary", + "text_level": 1, + "bbox": [ + 513, + 676, + 638, + 694 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Leaky Integrate-and-Fire Neuron. As a fundamental building block of SNNs, the Leaky Integrate-and-Fire (LIF) neuron [3] has emerged as an important component for energy-efficient computation. The LIF neuron is a nonlinear activation function that determines whether neurons fire spikes as follows:", + "bbox": [ + 511, + 702, + 903, + 792 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {u} [ t + 1 ] ^ {l} = \\tau \\mathbf {u} [ t ] ^ {l} + \\mathbf {W} ^ {l} f (\\mathbf {u} [ t ] ^ {l - 1}) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 799, + 903, + 819 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\nf \\left(\\mathbf {u} [ t ] ^ {l}\\right) = \\left\\{ \\begin{array}{l l} 1 & \\text {i f} \\mathbf {u} [ t ] ^ {l} > V _ {t h}, \\\\ 0 & \\text {o t h e r w i s e} \\end{array} , \\right. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 607, + 821, + 903, + 861 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "where, $\\mathbf{u}[t]^l$ is the membrane potential in $l$ -th layer at timestep $t$ , $\\tau \\in (0,1]$ is the leaky factor for membrane po", + "bbox": [ + 511, + 869, + 906, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "13949", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/57a643a56b805f3a858c0f0fca5f6a00b73194fbb45b0993f9a3360bbee9eb6c.jpg", + "image_caption": [ + "Figure 2. Comparison between different self-attention with the CIFAR100 dataset. Analysis of entropy and accuracy for Temporal-only (T), Spatial-only (S), and Spatial-temporal (ST)." + ], + "image_footnote": [], + "bbox": [ + 107, + 89, + 281, + 207 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/5a8ef6d7cfd081febdb07f4e9dbe44f2a3e8984722cca88f84f02537a48cde77.jpg", + "image_caption": [ + "Figure 3. (a) Maximum batch size of block-wise STAtten and full spatial-temporal attention (without block partitioning) for running on A5000 GPU with 24GB VRAM memory. (b) Average number of active neurons after QKV computation at different timestep combinations, where [t] indicates timestep index." + ], + "image_footnote": [], + "bbox": [ + 282, + 89, + 459, + 207 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "tential leakage, $\\mathbf{W}^l$ is the weight of $l$ -th layer, and $f(\\cdot)$ is the LIF function with firing threshold $V_{th}$ . Therefore, when the membrane $\\mathbf{u}[t]^l$ is higher than $V_{th}$ , the LIF function fires a spike and the membrane potential is reset to 0.", + "bbox": [ + 89, + 265, + 483, + 325 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Vanilla Self-attention. Transformers [11, 47] have surpassed convolution-based architectures due to their unique self-attention mechanism, which can capture global features through long-range dependency. For a floating-point input tensor $\\mathbf{X}_f\\in \\mathbb{R}^{N\\times D}$ with $N$ tokens, and $D$ features, we formulate the self-attention as follows:", + "bbox": [ + 89, + 325, + 483, + 416 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {Q} _ {f} = \\mathbf {W} _ {\\mathbf {Q}} \\mathbf {X} _ {f}, \\quad \\mathbf {K} _ {f} = \\mathbf {W} _ {f} \\mathbf {X} _ {f}, \\quad \\mathbf {V} _ {f} = \\mathbf {W} _ {\\mathbf {V}} \\mathbf {X} _ {f} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 104, + 426, + 482, + 444 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {A t t n} = \\operatorname {S o f t m a x} \\left(\\frac {\\mathbf {Q K} ^ {\\top}}{\\sqrt {D}}\\right) \\mathbf {V}, \\quad \\mathbf {Q} _ {f}, \\mathbf {K} _ {f}, \\mathbf {V} _ {f} \\in \\mathbb {R} ^ {N \\times D}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 470, + 482, + 517 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbf{W}_{\\mathbf{Q}}, \\mathbf{W}_{\\mathbf{K}}, \\mathbf{W}_{\\mathbf{V}} \\in \\mathbb{R}^{D \\times D}$ are linear projections for floating-point $\\mathbf{Q}_f, \\mathbf{K}_f, \\mathbf{V}_f$ , respectively. Note that the computational complexity of attention is $\\mathcal{O}(N^2 D)$ due to its matrix multiplication operation.", + "bbox": [ + 89, + 517, + 483, + 578 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Spike-based Self-attention. The first spike-based transformer work, Spikformer [64], proposes Spiking Self Attention (SSA) with binary $\\mathbf{Q}$ , $\\mathbf{K}$ , $\\mathbf{V}$ . For binary input tensor $\\mathbf{X} \\in \\mathbb{R}^{T \\times N \\times D}$ with $T$ timestep, we formulate the SSA as follows:", + "bbox": [ + 89, + 579, + 483, + 654 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathbf {Q}, \\mathbf {K}, \\mathbf {V} = \\operatorname {L I F} (\\mathbf {W} _ {\\mathbf {Q}} \\mathbf {X}), \\operatorname {L I F} (\\mathbf {W} _ {\\mathbf {K}} \\mathbf {X}), \\operatorname {L I F} (\\mathbf {W} _ {\\mathbf {K}} \\mathbf {X}), \\\\ \\operatorname {S S A} (\\mathbf {Q K V}) = \\operatorname {L I F} \\left(\\mathbf {Q K} ^ {\\top} \\mathbf {V} \\cdot \\alpha\\right) \\tag {5} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 104, + 660, + 482, + 710 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbf{W}_{\\mathbf{Q}}, \\mathbf{W}_{\\mathbf{K}}, \\mathbf{W}_{\\mathbf{V}} \\in \\mathbb{R}^{D \\times D}$ are linear projections for binary $\\mathbf{Q}, \\mathbf{K}, \\mathbf{V}$ , respectively, and $\\alpha$ denotes a scaling factor.", + "bbox": [ + 89, + 710, + 482, + 756 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Methodology", + "text_level": 1, + "bbox": [ + 89, + 768, + 225, + 787 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we present our analysis of spike-based transformers and provide the details of our proposed methodology. We first establish the necessity of spatial-temporal attention in spike-based transformers through an information-theoretic analysis based on entropy measurements. Building upon these observations, we introduce our STAtten mechanism.", + "bbox": [ + 89, + 794, + 483, + 898 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/07fccc209c2c29094fca7d2e499594afceafd65519edf3fcda38ffabfffd6219.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 517, + 92, + 723, + 186 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/012624bc4f9fb21e1dc486e9082ba65e1ad012631e1872a46aca2c29ba3a8920.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 733, + 90, + 903, + 170 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1. Motivation of Spatial-Temporal Attention", + "text_level": 1, + "bbox": [ + 511, + 292, + 870, + 309 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The feature maps of spike-based transformers at each timestep carry distinct information, as their activation distributions are dynamically influenced by both the leaky factor and reset mechanisms. This temporal aspect of information flow raises a fundamental question: how do different attention mechanisms effectively capture and utilize this spatiotemporal information?", + "bbox": [ + 511, + 316, + 905, + 421 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "First, we define the three different types of attention: Spatial, Temporal, and Spatial-temporal attention with timestep $t$ , token position $n$ , and feature dimension $d$ as follows:", + "bbox": [ + 511, + 422, + 905, + 481 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathrm {S} _ {-} \\operatorname {A t t n} [ t,,:,: ] = \\operatorname {L I F} (\\mathbf {Q} [ t,:,: ] \\mathbf {K} ^ {\\top} [ t,:,: ] \\mathbf {V} [ t,:,: ] \\cdot \\alpha), \\\\ \\mathrm {T} _ {\\mathbf {\\Lambda}} \\operatorname {A t t n} [:, n,: ] = \\operatorname {L I F} (\\mathbf {Q} [:, n,: ] \\mathbf {K} ^ {\\top} [:, n,: ] \\mathbf {V} [:, n,: ] \\cdot \\alpha), \\\\ \\operatorname {S T} \\text {A t t n} [:,:, d ] = \\operatorname {L I F} (\\mathbf {Q} [:,:, d ] \\mathbf {K} ^ {\\top} [:,:, d ] \\mathbf {V} [:,:, d ] \\cdot \\alpha), \\tag {6} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 522, + 489, + 903, + 564 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where S_Attn, T_Attn, and ST_Attn are spatial, temporal, and spatial-temporal attention, and the Q, K, V dimensions of S_Attn, T_Attn, and ST_Attn are $\\mathbb{R}^{T\\times N\\times D}$ . The key distinction lies in their attention computation scope: spatial attention operates on token relationships within individual timesteps, temporal attention focuses on cross-timestep feature dependencies, and spatial-temporal attention integrates both spatial and temporal feature correlations. To quantify the information distribution in each attention mechanism, we measure Shannon entropy [16] as follows:", + "bbox": [ + 511, + 565, + 905, + 715 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nH (\\mathrm {A t t n}) = - \\sum_ {t = 1} ^ {T} \\sum_ {n = 1} ^ {N} \\sum_ {d = 1} ^ {D} \\hat {\\mathbf {p}} [ t, n, d ] \\log \\hat {\\mathbf {p}} [ t, n, d ], \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 534, + 726, + 906, + 768 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\hat{\\mathbf{p}}[t,n,d] = \\text{Softmax}(\\text{Attn}[t,n,d])$ . Lower entropy values indicate more focused spike patterns, while higher entropy values suggest dispersed patterns [25]. We analyze this relationship using a pretrained Spike-driven Transformer [53] on the CIFAR100 dataset, as shown in Fig. 2. The results reveal an inverse correlation between entropy and accuracy: temporal-only attention shows high entropy (5.81) but lowest accuracy (77.7%), while", + "bbox": [ + 511, + 779, + 906, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "13950", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/cac2d98b2a42e1917ce9c60b58e0008f7e9e9500c4aaa35794f0fa5dcacbe75d.jpg", + "image_caption": [ + "(a)", + "Spatial-Temporal Attention (STAtten)", + "Figure 4. Overview of STAtten architecture. (a) Block-wise temporal attention mechanism. Binary Q, K, V tensors are partitioned into temporal blocks, where black lines indicate paired timestep processing. (b) Computation flow with tensor dimensions, where $T$ is the number of timesteps, $N$ is the number of tokens, $B$ is block size, and $D$ is the feature dimension." + ], + "image_footnote": [], + "bbox": [ + 112, + 119, + 586, + 262 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/d62e9d2b639829e04d5385ee6d827087828fdbbaa9139fdbf1e9797dcc5d4954.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 611, + 102, + 880, + 261 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "spatial-temporal attention achieves lower entropy (4.85) with higher accuracy $(79.9\\%)$ . These findings verify that combining spatial and temporal processing leads to more structured feature representations, providing strong motivation for spatial-temporal attention.", + "bbox": [ + 88, + 323, + 483, + 398 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "However, we argue that fully correlating temporal information across all timesteps like ST_Attn in Eq. 6 introduces two challenges: (1) substantial memory requirements for storing temporal information, and (2) increased dead neurons when correlating distant timesteps due to weak spike similarity. The first challenge arises from matrix multiplication with large temporal dimensions $(TN,D)$ . To address this, we partition the temporal dimension into blocks, reducing the memory footprint for matrix operations. As shown in Fig. 3(a), this block-wise computation strategy leads to around $1.6\\times$ higher batch size compared to full temporal attention. The second challenge stems from binary matrix multiplication in spike-based processing. Our analysis in Fig. 3(b) reveals that QKV computations show higher active neurons when processing same-timestep, while activity decreases with temporal distance.", + "bbox": [ + 89, + 398, + 483, + 641 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. STAtten Mechanism", + "text_level": 1, + "bbox": [ + 89, + 650, + 284, + 666 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Based on these observations, we propose Spatial-Temporal Attention (STAtten). The block-wise STAtten partitions the temporal sequence into blocks of size $B$ , as shown in Fig. 4(a). We define STAtten as follows:", + "bbox": [ + 89, + 672, + 483, + 734 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {S T A t t e n} (\\mathbf {X} [ b ]) = \\operatorname {L I F} (\\mathbf {Q} [ b ] \\mathbf {K} ^ {\\top} [ b ] \\mathbf {V} [ b ] \\cdot \\alpha),\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 744, + 382, + 762 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\text {w h e r e} [ b ] = [ i B: (i + 1) B, :, d ], \\quad i \\in \\{0, 1, \\dots , T / B - 1 \\} \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 763, + 482, + 792 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $[iB:(i + 1)B, :, d]$ denotes the aggregated features from timesteps $iB$ to $(i + 1)B$ . For instance, with $T = 4$ and $B = 2$ , features are grouped in pairs of timesteps [0,1], [2,3]. By processing blocks of timesteps together, we capture local temporal relationships while maintaining manageable computation. This approach requires storing only a subset of temporal information at any given time, reduc", + "bbox": [ + 89, + 794, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/4baa0e3533484639448e6ff5520c6e08a28f030fed2898beb8f901ac0858e9d8.jpg", + "table_caption": [ + "Table 1. Computational complexity and energy consumption of different attention mechanisms. $E_{MAC}$ and $E_{AC}$ are energy costs for MAC and AC operations, $N$ , $D$ , $T$ denote patches, dimension, timesteps, and $S_{Q}$ , $S_{K}$ , $S_{V}$ are firing rates of $Q$ , $K$ , $V$ . SDT represents Spike-driven Transformer." + ], + "table_footnote": [], + "table_body": "
MethodSTComplexityEnergy
ViT [11]✘O(N2D)EMAC·N2D
ViViT [2]✓O(T2N2D)EMAC·T2N2D
Spikformer [64]✘O(TND2)EAC·TND2·(SQ+SK+SV)
SDT [53]✘O(TND)EAC·TND·(SQ+SK)
SDT-V2 [52]✘O(TND2)EAC·TND2·(SQ+SK+SV)
QKFormer [62]✘O(TND2)EAC·TND2·(SQ+SK+SV)
STAtten✓O(TND2)EAC·TND2·(SQ+SK+SV)
", + "bbox": [ + 516, + 402, + 906, + 515 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ing memory overhead. We set block size $B = 2$ with 4 timesteps for static datasets and $B = 4$ with 16 timesteps for neuromorphic datasets.", + "bbox": [ + 511, + 546, + 906, + 590 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Unlike traditional transformers that require Softmax operations, the binary $\\mathbf{Q}$ , $\\mathbf{K}$ , and $\\mathbf{V}$ allow flexible reordering of computations. Generalizing to arbitrary number of timesteps $T$ , the computation follows:", + "bbox": [ + 511, + 595, + 906, + 655 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {Q}, \\mathbf {K}, \\mathbf {V} \\in \\mathbb {R} ^ {\\frac {T}{B} \\times N \\times D} \\xrightarrow {\\text {r e s h a p e}} \\mathbf {Q}, \\mathbf {K}, \\mathbf {V} \\in \\mathbb {R} ^ {\\left(\\frac {T}{B} N\\right) \\times D} \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 522, + 672, + 903, + 700 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {K} ^ {\\top} \\mathbf {V} \\in \\mathbb {R} ^ {D \\times D} \\xrightarrow {\\text {a t t e n t i o n}} \\mathbf {Q} (\\mathbf {K} ^ {\\top} \\mathbf {V}) \\in \\mathbb {R} ^ {\\frac {T}{B} \\times N \\times D}.\n$$\n", + "text_format": "latex", + "bbox": [ + 542, + 699, + 870, + 717 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The computation sequence reduces intermediate memory requirements through early $\\mathbf{K}^{\\top}\\mathbf{V}$ calculation while preserving cross-timestep information flow within blocks. This implementation maintains computational complexity of $\\mathcal{O}(TND^{2})$ , matching spatial-only attention [42, 52, 53, 62, 64] while incorporating temporal dependencies. By avoiding the quadratic complexity in timesteps and the number of tokens ( $T^{2}N^{2}$ ) that would result from full temporal correlation, our method scales efficiently to longer sequences. Fig. 4(b) illustrates the complete computation process with corresponding tensor dimensions.", + "bbox": [ + 511, + 734, + 908, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "13951", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/1e5466c4f7b682260cd1755f79787c5b6547fde135617473001c843f1629522f.jpg", + "table_caption": [ + "Table 2. Performance comparison on sequential CIFAR10/100 datasets, demonstrating the effectiveness of spatial-temporal attention in capturing long-term dependencies." + ], + "table_footnote": [], + "table_body": "
Methods-CIFAR10 (%)s-CIFAR100 (%)
PLIF [14]81.4753.38
Spikformer [64]79.2957.17
STAtten + [64]82.4558.75
Spike-driven Transformer [53]80.3261.08
STAtten + [53]83.4164.30
SpikingReformer [42]79.2557.48
STAtten + [42]81.8458.30
", + "bbox": [ + 91, + 141, + 483, + 268 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. STAtten with Existing Spiking Transformers", + "text_level": 1, + "bbox": [ + 89, + 290, + 470, + 306 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "STAtten is designed to be a plug-and-play module that can enhance various spike-based transformer architectures, including Spikformer [64], Spike-driven Transformer [53], Spike-driven Transformer-V2 [52], SpikingReformer [42], and QKformer [62] while preserving their unique characteristics.", + "bbox": [ + 89, + 313, + 483, + 404 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The key to successful integration lies in preserving each architecture's distinct residual connection strategies while replacing their spatial-only attentions with STAtten. Two primary residual connection mechanisms are prevalent in current architectures: Spike Element-Wise (SEW) shortcut [13] and Membrane-Shortcut (MS) [20]. In architectures employing SEW shortcuts, such as Spikformer and QKFormer, spike activations are directly added to the attention output. In contrast, Spike-driven Transformer, Spikedriven Transformer-V2, and SpikingReformer utilize MSshortcuts that propagate membrane potentials rather than spikes Our STAtten module maintains compatibility with both connection types. We detail the existing spike-based attention mechanisms of each architecture and our integration strategy.", + "bbox": [ + 89, + 405, + 483, + 630 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Spikformer [64]: Spikformer proposes Spiking Self Attention (SSA) module which is inspired by the vanilla self-attention of ViT [11] as follows:", + "bbox": [ + 89, + 631, + 483, + 676 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nS S A (\\mathbf {Q}, \\mathbf {K}, \\mathbf {V}) = \\operatorname {L I F} (\\mathbf {Q} \\odot \\mathbf {K} ^ {\\top} \\odot \\mathbf {V} \\cdot \\alpha), \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 127, + 685, + 480, + 704 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\odot$ is matrix multiplication. This work establishes two key principles for spike-based attention: non-softmax operation and flexible Q, K, V operation ordering. While these principles form the foundation for subsequent spike-based architectures, the SSA module processes features independently at each timestep. We address this limitation by replacing the attention mechanism with STAtten.", + "bbox": [ + 89, + 712, + 483, + 818 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Spike-driven Transformer [53]: Spike-driven Transformer introduces Spike-Driven Self-Attention (SDSA), which reformulates attention computation as:", + "bbox": [ + 89, + 818, + 483, + 864 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nS D S A (\\mathbf {Q}, \\mathbf {K}, \\mathbf {V}) = \\mathbf {Q} \\otimes \\operatorname {L I F} \\left(\\sum_ {c} (\\mathbf {K} \\otimes \\mathbf {V})\\right), \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 119, + 872, + 483, + 904 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\otimes$ denotes element-wise multiplication and $\\sum_{c}$ represents column-wise summation. While SDSA achieves computational efficiency through element-wise operations, this design limits feature correlation even within the spatial domain due to the element-wise multiplication. By integrating STAtten into this architecture, we enable comprehensive spatial-temporal feature correlation.", + "bbox": [ + 511, + 90, + 903, + 196 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "SpikingReformer [61]: SpikingReformer identifies two key challenges in adapting conventional Q, K, V attention to spike-based transformers: (1) floating-point matrix multiplication between Q, K, V and (2) Softmax operations involving exponentiation and division. To address these challenges, the authors propose Dual Spike Self-Attention (DSSA):", + "bbox": [ + 511, + 196, + 905, + 303 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nD S S A (\\mathbf {X}) = \\operatorname {L I F} \\left(\\mathbf {X} \\odot \\mathbf {W} ^ {\\top} \\odot \\mathbf {X} ^ {\\top} \\cdot \\alpha\\right), \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 571, + 313, + 903, + 332 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{X}$ and $\\mathbf{W}$ denote binary input and linear projection, respectively. We try to verify that these concerns can be addressed by replacing DSSA with STAtten, showing that $\\mathbf{Q}$ , $\\mathbf{K}$ , $\\mathbf{V}$ matrix multiplication is still compatible with spike-based processing when properly designed.", + "bbox": [ + 511, + 343, + 905, + 419 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "QKFormer [62]: QKFormer implements dual attention mechanisms: Q-K Channel Attention (QKTA) and SSA. While SSA follows the Spikformer design, QKTA introduces a token-wise attention:", + "bbox": [ + 511, + 419, + 905, + 479 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nQ K T A = \\operatorname {L I F} \\left(\\sum_ {N} (\\mathbf {Q})\\right) \\otimes \\mathbf {K}, \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 604, + 489, + 903, + 520 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\sum_{N}$ represents token-wise summation. QKFormer redesigns the architecture by incorporating embedding layers in each encoder. In this architecture, we only replace SSA with STAtten as SSA is the primary mechanism responsible for capturing feature relationships.", + "bbox": [ + 511, + 532, + 905, + 608 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As we only modify the self-attention module, the original information flow through residual connections remains unchanged, maintaining the architecture's core characteristics. After computing STAtten across temporal blocks, the features are concatenated and processed through subsequent layers: $\\mathbf{F}_{\\mathrm{out}} = \\mathrm{Concat}(\\mathrm{STAtten}(\\mathbf{X}[iB:(i + 1)B])_{i = 0}^{T / B - 1})$ This concatenated feature then flows through convolutional and batch normalization layers: $\\mathbf{Z} = \\mathrm{BN}(\\mathrm{Conv}(\\mathbf{F}_{\\mathrm{out}}))$", + "bbox": [ + 511, + 609, + 905, + 732 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.4. Complexity and Energy of Self-attention", + "text_level": 1, + "bbox": [ + 511, + 741, + 861, + 758 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We analyze both computational complexity and theoretical energy consumption of our proposed architecture, focusing specifically on the self-attention mechanism. A comprehensive evaluation of the overall architecture's energy profile is provided in Supplementary Material.", + "bbox": [ + 511, + 763, + 905, + 840 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As shown in Table 1, despite incorporating spatial-temporal attention capabilities, our STAtten maintains the same computational complexity $\\mathcal{O}(TND^2)$ as previous spatial-only spiking transformers. This efficiency is", + "bbox": [ + 511, + 840, + 905, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "13952", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/23783534b76bd21b1c4f7f097a3a4420659b924a0355a289eee0651bea5530c5.jpg", + "table_caption": [ + "Table 3. Performance comparison between our methods and previous works on CIFAR10 and CIFAR100 datasets. In the architecture column, Architecture- $L - D$ represents $L$ number of encoder blocks and $D$ hidden dimensions. All baseline architectures of spike-based transformer and corresponding STAtten implementations are trained from scratch for fair comparison." + ], + "table_footnote": [], + "table_body": "
MethodArchitectureTimestepCIFAR10 (%)CIFAR100 (%)
tdBN [60]ResNet19492.92-
PLIF [14]ConvNet893.50-
Dspike [31]ResNet18694.2574.24
DSR [36]ResNet182095.4078.50
DIET-SNN [39]VGG16592.7069.67
SNASNet [24]ConvNet593.6473.04
Spikformer [64]Spikformer-4-384493.9975.06
STAtten + [64]Spikformer-4-384494.3675.85
Spike-driven Transformer [53]Spike-driven Transformer-2-512495.6078.40
STAtten + [53]Spike-driven Transformer-2-512496.0379.85
SpikingReformer [42]SpikingReformer-6-384495.0377.16
STAtten + [42]SpikingReformer-6-384495.2677.90
QKFormer [62]QKFormer-4-384495.1279.79
STAtten + [62]QKFormer-4-384495.3580.20
", + "bbox": [ + 181, + 142, + 818, + 358 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "achieved through our non-Softmax design, which enables flexible reordering of Q, K, V operations. In contrast, conventional ANN-based spatial-temporal attention methods like ViViT [2] require substantially higher complexity $(\\mathcal{O}(T^2 N^2 D))$ due to their softmax-constrained computation order. For energy consumption estimation, we follow the floating-point operations (FLOPs) implementation in 45nm technology [18]. The energy costs are categorized into $E_{MAC}$ for multiply-accumulate operations corresponding to 32-/8-bit ANN computations, and $E_{AC}$ for accumulate operations used in binary SNN calculations. The final energy consumption is modulated by the firing rates $(S_Q, S_K, \\text{and} S_V)$ of the respective Q, K, and V. A detailed analysis of these energy measurements and their implications is presented in Section 5.3.", + "bbox": [ + 88, + 375, + 480, + 602 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 89, + 614, + 223, + 633 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We verify our proposed methods on both static and dynamic datasets, e.g., CIFAR10/100 [26], ImageNet [7], CIFAR10-DVS [29], and N-Caltech101 [37] datasets. We use direct coding to convert float pixel values into binary spikes [49]. In all our experiments, the spiking transformer STAtten models are trained from scratch. The detailed training strategy for each dataset is explained in the Supplementary Material.", + "bbox": [ + 89, + 640, + 482, + 760 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Notation: In the experiments, we use the notation STAtten $+\\left[\\cdot \\right]$ to denote STAtten implementation based on the backbone architecture used in $\\left[\\cdot \\right]$ . E.g., STAtten+[64] refers to training a STAtten model with Spikformer backbone from scratch.", + "bbox": [ + 89, + 761, + 482, + 835 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Sequential CIFAR10/100 Classification", + "text_level": 1, + "bbox": [ + 89, + 847, + 424, + 864 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We first evaluate our spatial-temporal attention mechanism's capability to capture long-term temporal dependen", + "bbox": [ + 89, + 869, + 482, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "cies using sequential CIFAR10 and CIFAR100 datasets. In this task, we transform the standard image classification task into a temporal sequence processing task by dividing each input image column-wise, shown in Fig. 1, where each column serves as input for one timestep. This results in a sequence length of 32 timesteps, corresponding to the image width. To accommodate this 1-D sequential input, we modify the convolutional, batch normalization, and pooling layers from 2-D to 1-D operations. As shown in Table 2, STAtten outperforms existing spatial-only attention methods on both datasets. Specifically, in Spike-driven Transformer [53], our approach achieves improvements of $3.09\\%$ and $3.22\\%$ on s-CIFAR10 and s-CIFAR100, respectively, compared to the baseline models. This performance gain demonstrates STAtten's capability in capturing long-range temporal dependencies.", + "bbox": [ + 511, + 375, + 906, + 617 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Performance Analysis", + "text_level": 1, + "bbox": [ + 511, + 622, + 718, + 638 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "CIFAR10/100. The CIFAR10/100 datasets comprise static images of $32 \\times 32$ pixels, with 50,000 training and 10,000 test images. As shown in Table 3, STAtten improves the performance of various spike-based transformers, achieving $95.35\\%$ accuracy on CIFAR10 and $80.20\\%$ on CIFAR100 when integrated with QKFormer. These improvements over spatial-only attention methods are consistent across different architectures, demonstrating the benefits of spatial-temporal processing.", + "bbox": [ + 511, + 642, + 906, + 779 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ImageNet. ImageNet is a large-scale image dataset with images sized $224 \\times 224$ pixels, comprising approximately 1.2 million training images and 50,000 validation images across 1,000 classes. As shown in Table 4, STAtten enhances the performance of spike-based transformers. With Spike-driven Transformer [53], it achieves $78.11\\%$ accuracy at $224 \\times 224$ resolution and $78.39\\%$ at $288 \\times 288$ resolution. Integration with Spike-driven Transformer-V2", + "bbox": [ + 511, + 779, + 908, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "13953", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/1f3839cf5e17c44d3b2af05676cd00a5b946f1fdcaad3a091d743b4064fb8821.jpg", + "table_caption": [ + "Table 4. Performance comparison between our methods and previous works on ImageNet datasets. Note that * represents the inference accuracy with $288 \\times 288$ resolution. In the architecture column, $L-D$ represents $L$ number of encoder blocks and $D$ hidden dimensions. † represents our implementation." + ], + "table_footnote": [], + "table_body": "
MethodTypeArchitectureParam(M)TimestepAccuracy(%)
Vision Transformer [11]ANNViT-B/1686.59177.90
DeiT [46]ANNDeiT-B86.59181.80
Swin Transformer [33]ANNSwin Transformer-B87.77183.50
TET [9]SNNSEW-ResNet3421.79468.00
Spiking ResNet [19]SNNResNet5025.5635072.75
tdBN [60]SNNResNet3421.79663.72
SEW-ResNet [13]SNNSEW-ResNet15260.19469.26
MS-ResNet [20]SNNResNet10478.37574.21 / 76.02*
Att-MS-ResNet [51]SNNResNet10478.37477.08*
Spikformer [64]SNNSpikformer-8-76866.34474.81
SpikingReformer [42]SNNSpikingReformer-6-102466.38478.77 / 79.40*
QKFormer [62]SNNQKFormer-6-102464.96484.22
Spike-driven Transformer [53]SNNSpike-driven Transformer-8-51229.68474.57
SNNSpike-driven Transformer-10-51236.01474.66
SNNSpike-driven Transformer-8-76866.34476.32 / 77.07*
STAtten + [53]SNNSpike-driven Transformer-8-51229.68476.18 / 76.56*
SNNSpike-driven Transformer-8-76866.34478.11 / 78.39*
Spike-driven Transformer-V2 [52]SNNSpike-driven Transformer-V2-8-51255.4479.49† / 79.98*†
STAtten + [52]SNNSpike-driven Transformer-V2-8-51255.4479.85 / 80.67*
", + "bbox": [ + 143, + 141, + 854, + 426 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/7fd3e889e109d6cde778998dd5b344685eca1434b385e8532d7ae19b125e2ed6.jpg", + "table_caption": [ + "Table 5. Performance comparison between our methods and previous works on CIFAR10-DVS and N-Caltech101 datasets. In the architecture column, $L - D$ represents $L$ number of encoder blocks and $D$ hidden dimensions. All baseline architectures of spike-based transformer and corresponding STAtten implementations are trained from scratch for fair comparison." + ], + "table_footnote": [], + "table_body": "
MethodArchitectureTimestepCIFAR10-DVS (%)N-Caltech101 (%)
tdBN [60]ResNet191067.80-
PLIF [14]ConvNet2074.80-
Dspike [31]ResNet181075.40-
DSR [36]ResNet182077.27-
SEW-ResNet [13]ConvNet2074.80-
TT-SNN [27]ResNet346-77.80
NDA [32]VGG111079.678.2
Spikformer [64]Spikformer-2-2561680.9-
Spike-driven Transformer [53]Spike-driven Transformer-2-2561680.081.80
STAtten + [53]Spike-driven Transformer-2-2561681.183.15
SpikingReformer [42]SpikingReformer-4-3841678.8081.29
STAtten + [42]SpikingReformer-4-3841680.6081.95
QKFormer [62]QKFormer-4-3841682.9083.58
STAtten + [62]SpikingReformer-4-3841683.9084.25
", + "bbox": [ + 158, + 492, + 839, + 720 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "[52] further improves performance to $79.85\\%$ and $80.67\\%$ at respective resolutions. It is noteworthy that Spike-driven Transformer [53] achieves comparable or better performance than recent Spikformer models while using only 29.68M parameters. Furthermore, STAtten with Spik-driven Transformer-V2 achieves even higher accuracy than SpikingReformer and Spikformer while using around $20\\%$ fewer parameters.", + "bbox": [ + 89, + 729, + 482, + 849 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Neuromorphic Datasets. We evaluate our approach on two neuromorphic vision datasets: CIFAR10-DVS [29] and N-Caltech101 [37]. CIFAR10-DVS contains 10,000 sam", + "bbox": [ + 89, + 854, + 483, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "plies (9,000 training, 1,000 testing) captured using a Dynamic Vision Sensor (DVS) from static CIFAR10 images. N-Caltech101, derived from the Caltech101 dataset [28], comprises 8,831 DVS image samples. Following standard protocols, we resize CIFAR10-DVS samples to $64 \\times 64$ and apply the augmentation strategy from [53], while N-Caltech101 samples are resized to $64 \\times 64$ with NDA [32] augmentation. As shown in Table 5, integrating STAtten consistently improves performance across different architectures. With Spike-driven Transformer as the backbone, our approach achieves $81.1\\%$ accuracy on CIFAR10-DVS", + "bbox": [ + 511, + 729, + 906, + 897 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "13954", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/90d30f6e5a8ecfe35e807d2aa60bb87137fee3767a8e3d1d0e20e5b375aab964.jpg", + "table_caption": [ + "Table 6. Comparison of memory, energy, and accuracy on ImageNet across different transformer architectures. The precision format weight/activation bit-width and the memory include the size of model parameters and activations. We show the memory consumption for both 32-bit and 8-bit weight spiking models. SDT represents Spike-driven Transformer." + ], + "table_footnote": [], + "table_body": "
MethodPrecisionTimestepMemory (MB)Energy (mJ)
ViT-B/16 [11]32/321351.8254.84
Liu (ViT-B) [34]8/8187.96110.8
Spikformer [64]32(8)/14285.99 (86.48)21.48
SpikingReformer [42]32(8)/14273.57 (85.4)8.76
QKFormer [62]32(8)/14280.28 (84.99)38.91
SDT [53]32(8)/14283.97 (87.42)12.42
STAtten + [53]32(8)/14283.97 (87.42)12.36
SDT-V2 [52]32(8)/14250.22 (84.02)52.40
STAtten + [53]32(8)/14250.22 (84.02)52.38
", + "bbox": [ + 91, + 176, + 483, + 314 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/94e18cb5eafbd04650560b31e5b958270b53d8b65a734d2896aeeb0d1c8b8ba1.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 94, + 330, + 279, + 459 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/58e98b294bc28df5f6974027d45af64fcbfd288272783b8d1e7aff1c76f01f9b.jpg", + "image_caption": [ + "(b)", + "Figure 5. Accuracy comparison with respect to the number of parameters on (a) CIFAR100 and (b) Sequential CIFAR100 datasets. Spike-driven Transformer [53] is used as the baseline of spatial-only architecture." + ], + "image_footnote": [], + "bbox": [ + 287, + 330, + 473, + 460 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "and $83.15\\%$ on N-Caltech101, improving the baseline by $1.1\\%$ and $1.35\\%$ respectively. The enhancement extends to other architectures as well, with SpikingReformer achieving $80.60\\%$ on CIFAR10-DVS and $81.95\\%$ on N-Caltech101, and QKFormer reaching $83.90\\%$ and $84.25\\%$ respectively, which is the state-of-the-art result.", + "bbox": [ + 89, + 547, + 482, + 638 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.3. Memory and Energy Analysis", + "text_level": 1, + "bbox": [ + 89, + 648, + 356, + 665 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We analyze memory and energy consumption during ImageNet inference. As shown in Table 6, traditional transformer, ViT-B/16, with 32-bit precision requires high memory (351.8 MB) and energy consumption (254.84 mJ). Even with 8-bit precision, Liu [34] still consumes $110.8\\mathrm{mJ}$ of energy. In contrast, spike-based transformers with 32-bit weights and binary activations demonstrate significant energy efficiency. Notably, STAtten integrated with SDT and SDT-V2 achieves similar energy consumption to the baseline architectures, showing that our spatial-temporal processing can be realized without additional energy overhead. Correspondingly, STAtten attains comparable performance with fewer parameters, as detailed in the next section. The methodology for energy calculations is provided in the Supplementary Material.", + "bbox": [ + 89, + 674, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.4. Model Capacity", + "text_level": 1, + "bbox": [ + 513, + 90, + 671, + 107 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To further evaluate the capacity of our STAtten in spike-based transformers, we compare the accuracy based on the number of trainable parameters through ablation studies. Fig. 5 shows the performance scaling across different model sizes on both (a) standard CIFAR100 and (b) sequential CIFAR100 datasets. Spike-driven Transformer [53] is used as the baseline of spatial-only architecture. As shown in Fig. 5(a), on CIFAR100, STAtten demonstrates strong performance across various model sizes ranging from 2.56M to 22.97M parameters. The accuracy improvement ( $\\sim$ 0.5-1.0%) is consistently maintained across different architectural scales. For Sequential CIFAR100 (Fig. 5(b)), where temporal dependencies are more crucial, STAtten shows particularly robust performance improvement ( $\\sim$ 3-5%) even with smaller architectures.", + "bbox": [ + 511, + 113, + 906, + 339 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.5. Limitation", + "text_level": 1, + "bbox": [ + 513, + 345, + 633, + 359 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The hardware implementation of STAtten faces challenges despite its improved accuracy. Full spatial-temporal attention without block partitioning is impractical for neuromorphic deployment, as it requires complete temporal information. Our block-wise approach partially addresses this by processing temporal information in steps, where each step corresponds to one block size. However, even with this optimization, deployment remains challenging on traditional neuromorphic chips like TrueNorth [1] and Loihi [6] that process information step-by-step. A potential solution lies in layer-by-layer neuromorphic chips [22, 54, 56, 63] that support layer-wise processing across timesteps. This architectural shift provides an efficient implementation path for STAtten's block-wise processing. Future integration of parallel LIF neurons [15] could further accelerate computation in layer-by-layer architectures.", + "bbox": [ + 511, + 368, + 908, + 611 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 617, + 633, + 632 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper proposes a block-wise spatial-temporal attention mechanism, STAtten, that enhances spike-based transformers. Through block-wise processing and leveraging non-softmax properties, STAtten captures spatial-temporal information without additional overhead. STAtten can be integrated into existing architectures while preserving their energy efficiency, consistently improving performance across different backbones. Experimental results across both static and neuromorphic datasets validate the effectiveness of our approach.", + "bbox": [ + 511, + 642, + 906, + 792 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgment", + "text_level": 1, + "bbox": [ + 513, + 800, + 665, + 816 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work was supported in part by CoCoSys, a JUMP2.0 center sponsored by DARPA and SRC, the National Science Foundation (CAREER Award, Grant #2312366, Grant #2318152), the DARPA Young Faculty Award and the DoEMMICC center SEA-CROGS (Award #DE-SC0023198).", + "bbox": [ + 511, + 824, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "13955", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Filipp Akopyan, Jun Sawada, Andrew Cassidy, Rodrigo Alvarez-Icaza, John Arthur, Paul Merolla, Nabil Imam, Yutaka Nakamura, Pallab Datta, Gi-Joon Nam, et al. Truenorth: Design and tool flow of a 65 mw 1 million neuron programmable neurosynaptic chip. IEEE transactions on computer-aided design of integrated circuits and systems, 34 (10):1537-1557, 2015. 1, 8", + "[2] Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lucic, and Cordelia Schmid. Vivit: A video vision transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6836-6846, 2021. 1, 4, 6", + "[3] Anthony N Burkitt. A review of the integrate-and-fire neuron model: I. homogeneous synaptic input. Biological cybernetics, 95:1-19, 2006. 2", + "[4] Lili Chen, Kevin Lu, Aravind Rajeswaran, Kimin Lee, Aditya Grover, Misha Laskin, Pieter Abbeel, Aravind Srinivas, and Igor Mordatch. Decision transformer: Reinforcement learning via sequence modeling. Advances in neural information processing systems, 34:15084-15097, 2021. 1", + "[5] Sayeed Shafayet Chowdhury, Isha Garg, and Kaushik Roy. Spatio-temporal pruning and quantization for low-latency spiking neural networks. In 2021 International Joint Conference on Neural Networks (IJCNN), pages 1-9. IEEE, 2021. 2", + "[6] Mike Davies, Narayan Srinivasa, Tsung-Han Lin, Gautham Chinya, Yongqiang Cao, Sri Harsha Choday, Georgios Dimou, Prasad Joshi, Nabil Imam, Shweta Jain, et al. Loihi: A neuromorphic manycore processor with on-chip learning. IEEE Micro, 38(1):82-99, 2018. 1, 8", + "[7] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 6", + "[8] Shikuang Deng and Shi Gu. Optimal conversion of conventional artificial neural networks to spiking neural networks. arXiv preprint arXiv:2103.00476, 2021. 2", + "[9] Shikuang Deng, Yuhang Li, Shanghang Zhang, and Shi Gu. Temporal efficient training of spiking neural network via gradient re-weighting. arXiv preprint arXiv:2202.11946, 2022. 7", + "[10] Peter U Diehl et al. Fast-classifying, high-accuracy spiking deep networks through weight and threshold balancing. In 2015 International joint conference on neural networks (IJCNN), pages 1–8. IEEE, 2015. 2", + "[11] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 1, 3, 4, 5, 7, 8", + "[12] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. International journal of computer vision, 88:303-338, 2010. 3" + ], + "bbox": [ + 93, + 114, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Wei Fang, Zhaofei Yu, Yanqi Chen, Tiejun Huang, Timothee Masquelier, and Yonghong Tian. Deep residual learning in spiking neural networks. Advances in Neural Information Processing Systems, 34:21056-21069, 2021. 5, 7", + "[14] Wei Fang, Zhaofei Yu, Yanqi Chen, Timothee Masquelier, Tiejun Huang, and Yonghong Tian. Incorporating learnable membrane time constant to enhance learning of spiking neural networks. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2661-2671, 2021. 2, 5, 6, 7", + "[15] Wei Fang, Zhaofei Yu, Zhaokun Zhou, Ding Chen, Yanqi Chen, Zhengyu Ma, Timothée Masquelier, and Yonghong Tian. Parallel spiking neurons with high efficiency and ability to learn long-term dependencies. Advances in Neural Information Processing Systems, 36, 2024. 8", + "[16] Marylou Gabrie, Andre Manoel, Clément Luneau, Nicolas Macris, Florent Krzakala, Lenka Zdeborova, et al. Entropy and mutual information in models of deep neural networks. Advances in neural information processing systems, 31, 2018. 3", + "[17] Bing Han and Kaushik Roy. Deep spiking neural network: Energy efficiency through time based coding. In European Conference on Computer Vision, pages 388-404. Springer, 2020. 2", + "[18] Mark Horowitz. 1.1 computing's energy problem (and what we can do about it). In 2014 IEEE international solid-state circuits conference digest of technical papers (ISSCC), pages 10-14. IEEE, 2014. 6, 1", + "[19] Yangfan Hu, Huajin Tang, and Gang Pan. Spiking deep residual networks. IEEE Transactions on Neural Networks and Learning Systems, 34(8):5200-5205, 2021. 7", + "[20] Yifan Hu, Lei Deng, Yujie Wu, Man Yao, and Guoqi Li. Advancing spiking neural networks toward deep residual learning. IEEE Transactions on Neural Networks and Learning Systems, 2024. 1, 2, 5, 7", + "[21] Seijoon Kim, Seongsik Park, Byunggook Na, and Sungroh Yoon. Spiking-yolo: spiking neural network for energy-efficient object detection. In Proceedings of the AAAI conference on artificial intelligence, pages 11270-11277, 2020. 3", + "[22] Sangyeob Kim, Soyeon Kim, Seongyon Hong, Sangjin Kim, Donghyeon Han, and Hoi-Jun Yoo. C-dnn: A 24.5-85.8 tops/w complementary-deep-neural-network processor with heterogeneous cnn/snn core architecture and forward-gradient-based sparsity generation. In 2023 IEEE International Solid-State Circuits Conference (ISSCC), pages 334-336. IEEE, 2023. 8", + "[23] Youngeun Kim, Joshua Chough, and Priyadarshini Panda. Beyond classification: Directly training spiking neural networks for semantic segmentation. Neuromorphic Computing and Engineering, 2(4):044015, 2022. 2", + "[24] Youngeun Kim, Yuhang Li, Hyoungseob Park, Yeshwanth Venkatesha, and Priyadarshini Panda. Neural architecture search for spiking neural networks. In European Conference on Computer Vision, pages 36-56. Springer, 2022. 6", + "[25] Youngeun Kim, Yuhang Li, Hyoungseob Park, Yeshwanth Venkatesha, Anna Hambitzer, and Priyadarshini Panda. Ex" + ], + "bbox": [ + 516, + 92, + 905, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "13956", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ploring temporal information dynamics in spiking neural networks. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 8308-8316, 2023. 3", + "[26] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 6", + "[27] Donghyun Lee, Ruokai Yin, Youngeun Kim, Abhishek Moitra, Yuhang Li, and Priyadarshini Panda. Tt-snn: tensor train decomposition for efficient spiking neural network training. In 2024 Design, Automation & Test in Europe Conference & Exhibition (DATE), pages 1-6. IEEE, 2024. 7", + "[28] FF Li, M Andreeto, M Ranzato, and P Perona. Caltech 101 (1.0)[data set]. caltechdata, 2022. 7", + "[29] Hongmin Li, Hanchao Liu, Xiangyang Ji, Guoqi Li, and Luping Shi. Cifar10-dvs: an event-stream dataset for object classification. Frontiers in neuroscience, 11:244131, 2017. 6, 7", + "[30] Yuhang Li, Shikuang Deng, Xin Dong, Ruihao Gong, and Shi Gu. A free lunch from ann: Towards efficient, accurate spiking neural networks calibration. In International conference on machine learning, pages 6316-6325. PMLR, 2021. 2", + "[31] Yuhang Li, Yufei Guo, Shanghang Zhang, Shikuang Deng, Yongqing Hai, and Shi Gu. Differentiable spike: Rethinking gradient-descent for training spiking neural networks. Advances in Neural Information Processing Systems, 34: 23426-23439, 2021. 2, 6, 7", + "[32] Yuhang Li, Youngeun Kim, Hyoungseob Park, Tamar Geller, and Priyadarshini Panda. Neuromorphic data augmentation for training spiking neural networks. In European Conference on Computer Vision, pages 631-649. Springer, 2022. 7, 1", + "[33] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 7", + "[34] Zhenhua Liu, Yunhe Wang, Kai Han, Wei Zhang, Siwei Ma, and Wen Gao. Post-training quantization for vision transformer. Advances in Neural Information Processing Systems, 34:28092-28103, 2021. 8", + "[35] Wolfgang Maass. Networks of spiking neurons: the third generation of neural network models. Neural networks, 10 (9):1659-1671, 1997. 1", + "[36] Qingyan Meng, Mingqing Xiao, Shen Yan, Yisen Wang, Zhouchen Lin, and Zhi-Quan Luo. Training high-performance low-latency spiking neural networks by differentiation on spike representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12444–12453, 2022. 6, 7", + "[37] Garrick Orchard, Ajinkya Jayawant, Gregory K Cohen, and Nitish Thakor. Converting static image datasets to spiking neuromorphic datasets using saccades. Frontiers in neuroscience, 9:159859, 2015. 6, 7", + "[38] Jing Pei, Lei Deng, Sen Song, Mingguo Zhao, Youhui Zhang, Shuang Wu, Guanrui Wang, Zhe Zou, Zhenzhi Wu, Wei He, et al. Towards artificial general intelligence with hybrid tianjic chip architecture. Nature, 572(7767):106-111, 2019. 1" + ], + "bbox": [ + 91, + 92, + 482, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[39] Nitin Rathi and Kaushik Roy. Diet-snn: Direct input encoding with leakage and threshold optimization in deep spiking neural networks. arXiv preprint arXiv:2008.03658, 2020. 6", + "[40] Kaushik Roy, Akhilesh Jaiswal, and Priyadarshini Panda. Towards spike-based machine intelligence with neuromorphic computing. Nature, 575(7784):607-617, 2019. 1", + "[41] Abhronil Sengupta, Yuting Ye, Robert Wang, Chiao Liu, and Kaushik Roy. Going deeper in spiking neural networks: Vgg and residual architectures. Frontiers in neuroscience, 13:95, 2019. 1", + "[42] Xinyu Shi, Zecheng Hao, and Zhaofei Yu. Spikingresformer: Bridging resnet and vision transformer in spiking neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5610-5619, 2024. 1, 2, 4, 5, 6, 7, 8, 3", + "[43] Sumit B Shrestha and Garrick Orchard. Slayer: Spike layer error reassignment in time. Advances in neural information processing systems, 31, 2018. 2", + "[44] Qiaoyi Su, Yuhong Chou, Yifan Hu, Jianing Li, Shijie Mei, Ziyang Zhang, and Guoqi Li. Deep directly-trained spiking neural networks for object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6555-6565, 2023. 2, 3", + "[45] Zhengkun Tian, Jiangyan Yi, Jianhua Tao, Ye Bai, Shuai Zhang, and Zhengqi Wen. Spike-triggered nonautoregressive transformer for end-to-end speech recognition. arXiv preprint arXiv:2005.07903, 2020. 1", + "[46] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In International conference on machine learning, pages 10347-10357. PMLR, 2021. 7", + "[47] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 1, 3", + "[48] Qingyu Wang, Tielin Zhang, Minglun Han, Yi Wang, Duzhen Zhang, and Bo Xu. Complex dynamic neurons improved spiking transformer network for efficient automatic speech recognition. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 102-109, 2023. 1", + "[49] Yujie Wu, Lei Deng, Guoqi Li, Jun Zhu, Yuan Xie, and Luping Shi. Direct training for spiking neural networks: Faster, larger, better. In Proceedings of the AAAI conference on artificial intelligence, pages 1311-1318, 2019. 1, 2, 6", + "[50] Shiting Xiao, Yuhang Li, Youngeun Kim, Donghyun Lee, and Priyadarshini Panda. Respike: residual frames-based hybrid spiking neural networks for efficient action recognition. Neuromorphic Computing and Engineering, 2025. 1", + "[51] Man Yao, Guangshe Zhao, Hengyu Zhang, Yifan Hu, Lei Deng, Yonghong Tian, Bo Xu, and Guoqi Li. Attention spiking neural networks. IEEE transactions on pattern analysis and machine intelligence, 2023. 7", + "[52] Man Yao, Jiakui Hu, Tianxiang Hu, Yifan Xu, Zhaokun Zhou, Yonghong Tian, Bo Xu, and Guoqi Li. Spike-driven transformer v2: Meta spiking neural network architecture inspiring the design of next-generation neuromorphic chips. arXiv preprint arXiv:2404.03663, 2024. 1, 2, 4, 5, 7, 8" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "13957", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[53] Man Yao, Jiakui Hu, Zhaokun Zhou, Li Yuan, Yonghong Tian, Bo Xu, and Guoqi Li. Spike-driven transformer. Advances in Neural Information Processing Systems, 36, 2024. 1, 2, 3, 4, 5, 6, 7, 8", + "[54] Ruokai Yin, Abhishek Moitra, Abhiroop Bhattacharjee, Youngeun Kim, and Priyadarshini Panda. Sata: Sparsity-aware training accelerator for spiking neural networks. IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems, 42(6):1926-1938, 2022. 8", + "[55] Ruokai Yin, Youngeun Kim, Yuhang Li, Abhishek Moitra, Nitin Satpute, Anna Hambitzer, and Priyadarshini Panda. Workload-balanced pruning for sparse spiking neural networks. IEEE Transactions on Emerging Topics in Computational Intelligence, 2024. 2", + "[56] Ruokai Yin, Youngeun Kim, Di Wu, and Priyadarshini Panda. Loas: Fully temporal-parallel dataflow for dualsparse spiking neural networks. In 2024 57th IEEE/ACM International Symposium on Microarchitecture (MICRO), pages 1107–1121. IEEE, 2024. 8", + "[57] Ruokai Yin, Yuhang Li, Abhishek Moitra, and Priyadarshini Panda. Mint: Multiplier-less integer quantization for energy efficient spiking neural networks. In 2024 29th Asia and South Pacific Design Automation Conference (ASP-DAC), pages 830-835. IEEE, 2024. 2", + "[58] Jiqing Zhang, Bo Dong, Haiwei Zhang, Jianchuan Ding, Felix Heide, Baocai Yin, and Xin Yang. Spiking transformers for event-based single object tracking. In Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition, pages 8801-8810, 2022. 1", + "[59] Jiyuan Zhang, Lulu Tang, Zhaofei Yu, Jiwen Lu, and Tiejun Huang. Spike transformer: Monocular depth estimation for spiking camera. In European Conference on Computer Vision, pages 34-52. Springer, 2022. 1", + "[60] Hanle Zheng, Yujie Wu, Lei Deng, Yifan Hu, and Guoqi Li. Going deeper with directly-trained larger spiking neural networks. In Proceedings of the AAAI conference on artificial intelligence, pages 11062-11070, 2021. 1, 6, 7", + "[61] Chenlin Zhou, Liutao Yu, Zhaokun Zhou, Han Zhang, Zhengyu Ma, Huihui Zhou, and Yonghong Tian. Spikingformer: Spike-driven residual learning for transformer-based spiking neural network. arXiv preprint arXiv:2304.11954, 2023. 5", + "[62] Chenlin Zhou, Han Zhang, Zhaokun Zhou, Liutao Yu, Liwei Huang, Xiaopeng Fan, Li Yuan, Zhengyu Ma, Huihui Zhou, and Yonghong Tian. Qkformer: Hierarchical spiking transformer using qk attention. arXiv preprint arXiv:2403.16552, 2024. 1, 2, 4, 5, 6, 7, 8", + "[63] PJ Zhou, Q Yu, M Chen, YC Wang, LW Meng, Y Zuo, N Ning, Y Liu, SG Hu, and GC Qiao. A 0.96 pj/sop, 30.23 k-neuron/mm 2 heterogeneous neuromorphic chip with fullerene-like interconnection topology for edge-ai computing. In 2024 IEEE International Symposium on Circuits and Systems (ISCAS), pages 1–5. IEEE, 2024. 8", + "[64] Zhaokun Zhou, Yuesheng Zhu, Chao He, Yaowei Wang, Shuicheng Yan, Yonghong Tian, and Li Yuan. Spikformer: When spiking neural network meets transformer. arXiv preprint arXiv:2209.15425, 2022. 1, 2, 3, 4, 5, 6, 7, 8" + ], + "bbox": [ + 91, + 92, + 480, + 895 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "13958", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Spiking Transformer with Spatial-Temporal Attention/de8548d3-fd04-4e18-9d13-559b7231ddb6_model.json b/2025/Spiking Transformer with Spatial-Temporal Attention/de8548d3-fd04-4e18-9d13-559b7231ddb6_model.json new file mode 100644 index 0000000000000000000000000000000000000000..8eab2fe6999a4c887a0ec5b9d71779bb37c3e68c --- /dev/null +++ b/2025/Spiking Transformer with Spatial-Temporal Attention/de8548d3-fd04-4e18-9d13-559b7231ddb6_model.json @@ -0,0 +1,2356 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.225, + 0.131, + 0.773, + 0.154 + ], + "angle": 0, + "content": "Spiking Transformer with Spatial-Temporal Attention" + }, + { + "type": "text", + "bbox": [ + 0.189, + 0.181, + 0.808, + 0.218 + ], + "angle": 0, + "content": "Donghyun Lee, Yuhang Li, Youngeun Kim, Shiting Xiao, Priyadarshini Panda Department of Electrical Engineering, Yale University" + }, + { + "type": "text", + "bbox": [ + 0.178, + 0.22, + 0.821, + 0.235 + ], + "angle": 0, + "content": "{donghyun.lee, yuhang.li, youngeun.kim, ginny.xiao, priya.panda}@yale.edu" + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.269, + 0.327, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.301, + 0.485, + 0.68 + ], + "angle": 0, + "content": "Spike-based Transformer presents a compelling and energy-efficient alternative to traditional Artificial Neural Network (ANN)-based Transformers, achieving impressive results through sparse binary computations. However, existing spike-based transformers predominantly focus on spatial attention while neglecting crucial temporal dependencies inherent in spike-based processing, leading to suboptimal feature representation and limited performance. To address this limitation, we propose Spiking Transformer with Spatial-Temporal Attention (STAtten), a simple and straightforward architecture that efficiently integrates both spatial and temporal information in the self-attention mechanism. STAtten introduces a block-wise computation strategy that processes information in spatial-temporal chunks, enabling comprehensive feature capture while maintaining the same computational complexity as previous spatial-only approaches. Our method can be seamlessly integrated into existing spike-based transformers without architectural overhaul. Extensive experiments demonstrate that STAtten significantly improves the performance of existing spike-based transformers across both static and neuromorphic datasets, including CIFAR10/100, ImageNet, CIFAR10-DVS, and N-Caltech101. The code is available at https://github.com/Intelligent-Computing-Lab-Yale/STAtten." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.709, + 0.222, + 0.725 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Spiking Neural Networks (SNNs) are emerging as a promising alternative to conventional Artificial Neural Networks (ANNs) [35, 40] due to their bio-inspired, energy-efficient computing paradigm. By leveraging sparse binary spike-based computations, SNNs can achieve significant energy savings while enabling deployment across various neuromorphic computing platforms such as TrueNorth [1], Loihi [6], and Tianjic [38]. However, the challenge lies in effectively processing and learning from these discrete, time-dependent spike patterns while maintaining computational efficiency and accuracy. Traditional SNNs have primarily" + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.272, + 0.906, + 0.41 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.427, + 0.907, + 0.47 + ], + "angle": 0, + "content": "Figure 1. Heatmaps of spatial-only attention versus our STAten on sequential CIFAR100 dataset. Input images are divided column-wise, where each column corresponds to one timestep." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.494, + 0.907, + 0.903 + ], + "angle": 0, + "content": "relied on convolution-based architectures adapted from successful ANN models like VGGNet [41, 49] and ResNet [20, 60]. While these architectures benefit from computational efficiency, they often struggle with binary spike activations, leading to information loss and degraded performance. This limitation has driven researchers to explore alternative architectures that can better handle sparse, temporal data while maintaining the energy benefits of spike-based computation. Transformer architectures, with their remarkable success across various domains [2, 4, 11, 47], have emerged as a promising direction for enhancing SNN capabilities. Recent works have introduced spike-formed transformers that adapt the transformer's self-attention mechanism to the spike-based domain for diverse tasks, such as, object tracking [58], computer vision [42, 50, 52, 53, 62, 64], depth estimation [59], and speech recognition [45, 48]. These architectures, such as Spikformer [64] and Spike-driven Transformer [53], implement innovative binary representations for query (Q), key (K), and value (V) computations while eliminating the computationally expensive Softmax function. This binary attention design preserves the essential characteristics of spike-based computation while enabling more efficient information processing compared to conventional transformer architectures. However, current spike-formed transformers predominantly rely on spatial-only attention, overlooking the inherently dynamic and temporal nature of spike events. To understand these differ" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13948" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.226 + ], + "angle": 0, + "content": "ent attention patterns, we compare spatial-only and spatial-temporal attention on sequential CIFAR100, as shown in Fig. 1. The visualization reveals that spatial-only attention focuses solely on vertical (spatial) relationships within each timestep, missing crucial object features that evolve over time. In contrast, our spatial-temporal attention captures both vertical (spatial) and horizontal (temporal) relationships, enabling comprehensive feature representation of sequential inputs." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.228, + 0.483, + 0.452 + ], + "angle": 0, + "content": "In this work, we introduce Spiking Transformer with Spatial-Temporal Attention (STAtten), which integrates both spatial and temporal information within the self-attention mechanism. Through analysis of memory requirements and neuronal activity patterns, we design a blockwise computation strategy with local temporal correlations. By eliminating Softmax operation, we maintain the same computational complexity \\((O(TND^{2}))\\) as existing spike-based transformers [42, 52, 62, 64]. Furthermore, our approach can be flexibly integrated into existing architectures without structural modifications, enhancing their representational capacity. Experiments on both static and dynamic datasets demonstrate consistent performance improvements over spatial-only attention architectures. The main contributions of our work are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.455, + 0.482, + 0.514 + ], + "angle": 0, + "content": "- We identify limitations in current spiking transformers that rely on spatial-only attention through empirical analysis, demonstrating the importance of capturing temporal dependencies inherent in spike-based processing." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.515, + 0.482, + 0.575 + ], + "angle": 0, + "content": "- We propose STAtten, a block-wise spatial-temporal attention mechanism for spike-based transformers that maintains the original computational complexity \\((O(TND^{2}))\\)." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.575, + 0.482, + 0.62 + ], + "angle": 0, + "content": "- We introduce a flexible plug-and-play design that enables STAtten to be integrated into existing spike-based transformers without compromising their core architecture." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.62, + 0.482, + 0.695 + ], + "angle": 0, + "content": "- Through extensive experiments across both static (CIFAR10/100, ImageNet) and neuromorphic (CIFAR10-DVS, N-Caltech101) datasets, we demonstrate that STatten consistently improves performance across different backbone architectures." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.455, + 0.482, + 0.695 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.71, + 0.241, + 0.725 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Training Strategies for SNNs. Significant progress has been made in developing effective training strategies for SNNs, primarily following two approaches: ANN-to-SNN conversion and direct training. The conversion approach transforms pre-trained ANNs into SNNs, leveraging established ANN architectures to enhance SNN performance. Several studies [8, 10, 17, 30] have formalized this conversion process and demonstrated its effectiveness in transferring ANN knowledge to the spike domain. Direct training through backpropagation, while fundamental to ANNs, presents unique challenges in SNNs due to the" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.242 + ], + "angle": 0, + "content": "non-differentiable nature of spike events. The introduction of surrogate gradients has been crucial in enabling effective backpropagation-based training in SNNs, as demonstrated in pioneering works [5, 43, 49]. Notably, direct training with surrogate gradients has shown superior performance compared to conversion methods across various tasks, including image classification [14, 20, 31, 55, 57], semantic segmentation [23], and object detection [44], while maintaining the energy efficiency inherent to spike-based computation." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.244, + 0.907, + 0.665 + ], + "angle": 0, + "content": "Spiking Transformer. Despite the energy efficiency of convolution-based SNNs, their performance often lags behind ANNs due to information loss from sparse binary activations. To address this limitation, researchers have adapted transformer architectures to the SNN domain, leveraging their powerful self-attention mechanisms and ability to capture long-range dependencies. This adaptation has led to several innovative spike-formed transformer architectures. Spikformer [64] introduced the first SNN-based transformer architecture, establishing foundational principles for spikeformed self-attention mechanisms. Their key innovation lies in eliminating the Softmax function from self-attention computations, justified by two key observations: (1) the binary nature of spike-based Q, K, and V values, and (2) the redundancy of Softmax scaling for binarized operations. Building upon this foundation, Spike-driven Transformer [53] enhanced the architecture with a novel self-attention block incorporating masking and sparse addition to optimize power efficiency. Additionally, inspired by MS-ResNet [20], Spike-driven Transformer reimagined residual connections to propagate membrane potential rather than spikes. Recent works have further extended these concepts, leading to architectures like SpikingReformer [42], QK-Former [62], and Spike-driven Transformer-V2 [52], each introducing unique optimizations for spike-based processing. However, existing spiking transformer architectures primarily focus on spatial attention, neglecting the temporal dynamics inherent in spike-based computation." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.678, + 0.64, + 0.695 + ], + "angle": 0, + "content": "3. Preliminary" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.703, + 0.905, + 0.793 + ], + "angle": 0, + "content": "Leaky Integrate-and-Fire Neuron. As a fundamental building block of SNNs, the Leaky Integrate-and-Fire (LIF) neuron [3] has emerged as an important component for energy-efficient computation. The LIF neuron is a nonlinear activation function that determines whether neurons fire spikes as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.59, + 0.8, + 0.905, + 0.82 + ], + "angle": 0, + "content": "\\[\n\\mathbf {u} [ t + 1 ] ^ {l} = \\tau \\mathbf {u} [ t ] ^ {l} + \\mathbf {W} ^ {l} f (\\mathbf {u} [ t ] ^ {l - 1}) \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.609, + 0.822, + 0.905, + 0.862 + ], + "angle": 0, + "content": "\\[\nf \\left(\\mathbf {u} [ t ] ^ {l}\\right) = \\left\\{ \\begin{array}{l l} 1 & \\text {i f} \\mathbf {u} [ t ] ^ {l} > V _ {t h}, \\\\ 0 & \\text {o t h e r w i s e} \\end{array} , \\right. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.87, + 0.907, + 0.902 + ], + "angle": 0, + "content": "where, \\(\\mathbf{u}[t]^l\\) is the membrane potential in \\(l\\)-th layer at timestep \\(t\\), \\(\\tau \\in (0,1]\\) is the leaky factor for membrane po" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13949" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.109, + 0.09, + 0.282, + 0.208 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.284, + 0.09, + 0.46, + 0.208 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.216, + 0.484, + 0.259 + ], + "angle": 0, + "content": "Figure 2. Comparison between different self-attention with the CIFAR100 dataset. Analysis of entropy and accuracy for Temporal-only (T), Spatial-only (S), and Spatial-temporal (ST)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.266, + 0.484, + 0.327 + ], + "angle": 0, + "content": "tential leakage, \\(\\mathbf{W}^l\\) is the weight of \\(l\\)-th layer, and \\(f(\\cdot)\\) is the LIF function with firing threshold \\(V_{th}\\). Therefore, when the membrane \\(\\mathbf{u}[t]^l\\) is higher than \\(V_{th}\\), the LIF function fires a spike and the membrane potential is reset to 0." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.327, + 0.484, + 0.417 + ], + "angle": 0, + "content": "Vanilla Self-attention. Transformers [11, 47] have surpassed convolution-based architectures due to their unique self-attention mechanism, which can capture global features through long-range dependency. For a floating-point input tensor \\(\\mathbf{X}_f\\in \\mathbb{R}^{N\\times D}\\) with \\(N\\) tokens, and \\(D\\) features, we formulate the self-attention as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.105, + 0.427, + 0.483, + 0.445 + ], + "angle": 0, + "content": "\\[\n\\mathbf {Q} _ {f} = \\mathbf {W} _ {\\mathbf {Q}} \\mathbf {X} _ {f}, \\quad \\mathbf {K} _ {f} = \\mathbf {W} _ {f} \\mathbf {X} _ {f}, \\quad \\mathbf {V} _ {f} = \\mathbf {W} _ {\\mathbf {V}} \\mathbf {X} _ {f} \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.101, + 0.472, + 0.483, + 0.518 + ], + "angle": 0, + "content": "\\[\n\\operatorname {A t t n} = \\operatorname {S o f t m a x} \\left(\\frac {\\mathbf {Q K} ^ {\\top}}{\\sqrt {D}}\\right) \\mathbf {V}, \\quad \\mathbf {Q} _ {f}, \\mathbf {K} _ {f}, \\mathbf {V} _ {f} \\in \\mathbb {R} ^ {N \\times D}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.518, + 0.484, + 0.579 + ], + "angle": 0, + "content": "where \\(\\mathbf{W}_{\\mathbf{Q}}, \\mathbf{W}_{\\mathbf{K}}, \\mathbf{W}_{\\mathbf{V}} \\in \\mathbb{R}^{D \\times D}\\) are linear projections for floating-point \\(\\mathbf{Q}_f, \\mathbf{K}_f, \\mathbf{V}_f\\), respectively. Note that the computational complexity of attention is \\(\\mathcal{O}(N^2 D)\\) due to its matrix multiplication operation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.58, + 0.484, + 0.655 + ], + "angle": 0, + "content": "Spike-based Self-attention. The first spike-based transformer work, Spikformer [64], proposes Spiking Self Attention (SSA) with binary \\(\\mathbf{Q}\\), \\(\\mathbf{K}\\), \\(\\mathbf{V}\\). For binary input tensor \\(\\mathbf{X} \\in \\mathbb{R}^{T \\times N \\times D}\\) with \\(T\\) timestep, we formulate the SSA as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.106, + 0.661, + 0.483, + 0.712 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathbf {Q}, \\mathbf {K}, \\mathbf {V} = \\operatorname {L I F} (\\mathbf {W} _ {\\mathbf {Q}} \\mathbf {X}), \\operatorname {L I F} (\\mathbf {W} _ {\\mathbf {K}} \\mathbf {X}), \\operatorname {L I F} (\\mathbf {W} _ {\\mathbf {K}} \\mathbf {X}), \\\\ \\operatorname {S S A} (\\mathbf {Q K V}) = \\operatorname {L I F} \\left(\\mathbf {Q K} ^ {\\top} \\mathbf {V} \\cdot \\alpha\\right) \\tag {5} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.712, + 0.483, + 0.757 + ], + "angle": 0, + "content": "where \\(\\mathbf{W}_{\\mathbf{Q}}, \\mathbf{W}_{\\mathbf{K}}, \\mathbf{W}_{\\mathbf{V}} \\in \\mathbb{R}^{D \\times D}\\) are linear projections for binary \\(\\mathbf{Q}, \\mathbf{K}, \\mathbf{V}\\), respectively, and \\(\\alpha\\) denotes a scaling factor." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.77, + 0.226, + 0.788 + ], + "angle": 0, + "content": "4. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.484, + 0.9 + ], + "angle": 0, + "content": "In this section, we present our analysis of spike-based transformers and provide the details of our proposed methodology. We first establish the necessity of spatial-temporal attention in spike-based transformers through an information-theoretic analysis based on entropy measurements. Building upon these observations, we introduce our STAtten mechanism." + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.093, + 0.725, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.636, + 0.188, + 0.653, + 0.2 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.735, + 0.092, + 0.905, + 0.171 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.815, + 0.188, + 0.832, + 0.2 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.209, + 0.907, + 0.279 + ], + "angle": 0, + "content": "Figure 3. (a) Maximum batch size of block-wise STAtten and full spatial-temporal attention (without block partitioning) for running on A5000 GPU with 24GB VRAM memory. (b) Average number of active neurons after QKV computation at different timestep combinations, where [t] indicates timestep index." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.293, + 0.871, + 0.31 + ], + "angle": 0, + "content": "4.1. Motivation of Spatial-Temporal Attention" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.317, + 0.906, + 0.422 + ], + "angle": 0, + "content": "The feature maps of spike-based transformers at each timestep carry distinct information, as their activation distributions are dynamically influenced by both the leaky factor and reset mechanisms. This temporal aspect of information flow raises a fundamental question: how do different attention mechanisms effectively capture and utilize this spatiotemporal information?" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.423, + 0.906, + 0.482 + ], + "angle": 0, + "content": "First, we define the three different types of attention: Spatial, Temporal, and Spatial-temporal attention with timestep \\( t \\), token position \\( n \\), and feature dimension \\( d \\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.523, + 0.491, + 0.905, + 0.565 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathrm {S} _ {-} \\operatorname {A t t n} [ t,,:,: ] = \\operatorname {L I F} (\\mathbf {Q} [ t,:,: ] \\mathbf {K} ^ {\\top} [ t,:,: ] \\mathbf {V} [ t,:,: ] \\cdot \\alpha), \\\\ \\mathrm {T} _ {\\mathbf {\\Lambda}} \\operatorname {A t t n} [:, n,: ] = \\operatorname {L I F} (\\mathbf {Q} [:, n,: ] \\mathbf {K} ^ {\\top} [:, n,: ] \\mathbf {V} [:, n,: ] \\cdot \\alpha), \\\\ \\operatorname {S T} \\text {A t t n} [:,:, d ] = \\operatorname {L I F} (\\mathbf {Q} [:,:, d ] \\mathbf {K} ^ {\\top} [:,:, d ] \\mathbf {V} [:,:, d ] \\cdot \\alpha), \\tag {6} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.566, + 0.906, + 0.716 + ], + "angle": 0, + "content": "where S_Attn, T_Attn, and ST_Attn are spatial, temporal, and spatial-temporal attention, and the Q, K, V dimensions of S_Attn, T_Attn, and ST_Attn are \\(\\mathbb{R}^{T\\times N\\times D}\\). The key distinction lies in their attention computation scope: spatial attention operates on token relationships within individual timesteps, temporal attention focuses on cross-timestep feature dependencies, and spatial-temporal attention integrates both spatial and temporal feature correlations. To quantify the information distribution in each attention mechanism, we measure Shannon entropy [16] as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.535, + 0.727, + 0.907, + 0.769 + ], + "angle": 0, + "content": "\\[\nH (\\mathrm {A t t n}) = - \\sum_ {t = 1} ^ {T} \\sum_ {n = 1} ^ {N} \\sum_ {d = 1} ^ {D} \\hat {\\mathbf {p}} [ t, n, d ] \\log \\hat {\\mathbf {p}} [ t, n, d ], \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.907, + 0.902 + ], + "angle": 0, + "content": "where \\(\\hat{\\mathbf{p}}[t,n,d] = \\text{Softmax}(\\text{Attn}[t,n,d])\\). Lower entropy values indicate more focused spike patterns, while higher entropy values suggest dispersed patterns [25]. We analyze this relationship using a pretrained Spike-driven Transformer [53] on the CIFAR100 dataset, as shown in Fig. 2. The results reveal an inverse correlation between entropy and accuracy: temporal-only attention shows high entropy (5.81) but lowest accuracy (77.7%), while" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13950" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.114, + 0.105, + 0.136, + 0.119 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image_caption", + "bbox": [ + 0.252, + 0.094, + 0.453, + 0.107 + ], + "angle": 0, + "content": "Spatial-Temporal Attention (STAtten)" + }, + { + "type": "image", + "bbox": [ + 0.114, + 0.12, + 0.587, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.612, + 0.103, + 0.882, + 0.262 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.266, + 0.908, + 0.309 + ], + "angle": 0, + "content": "Figure 4. Overview of STAtten architecture. (a) Block-wise temporal attention mechanism. Binary Q, K, V tensors are partitioned into temporal blocks, where black lines indicate paired timestep processing. (b) Computation flow with tensor dimensions, where \\( T \\) is the number of timesteps, \\( N \\) is the number of tokens, \\( B \\) is block size, and \\( D \\) is the feature dimension." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.324, + 0.484, + 0.399 + ], + "angle": 0, + "content": "spatial-temporal attention achieves lower entropy (4.85) with higher accuracy \\((79.9\\%)\\). These findings verify that combining spatial and temporal processing leads to more structured feature representations, providing strong motivation for spatial-temporal attention." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.4, + 0.484, + 0.642 + ], + "angle": 0, + "content": "However, we argue that fully correlating temporal information across all timesteps like ST_Attn in Eq. 6 introduces two challenges: (1) substantial memory requirements for storing temporal information, and (2) increased dead neurons when correlating distant timesteps due to weak spike similarity. The first challenge arises from matrix multiplication with large temporal dimensions \\((TN,D)\\). To address this, we partition the temporal dimension into blocks, reducing the memory footprint for matrix operations. As shown in Fig. 3(a), this block-wise computation strategy leads to around \\(1.6\\times\\) higher batch size compared to full temporal attention. The second challenge stems from binary matrix multiplication in spike-based processing. Our analysis in Fig. 3(b) reveals that QKV computations show higher active neurons when processing same-timestep, while activity decreases with temporal distance." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.651, + 0.285, + 0.667 + ], + "angle": 0, + "content": "4.2. STAtten Mechanism" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.674, + 0.484, + 0.735 + ], + "angle": 0, + "content": "Based on these observations, we propose Spatial-Temporal Attention (STAtten). The block-wise STAtten partitions the temporal sequence into blocks of size \\( B \\), as shown in Fig. 4(a). We define STAtten as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.09, + 0.745, + 0.383, + 0.763 + ], + "angle": 0, + "content": "\\[\n\\operatorname {S T A t t e n} (\\mathbf {X} [ b ]) = \\operatorname {L I F} (\\mathbf {Q} [ b ] \\mathbf {K} ^ {\\top} [ b ] \\mathbf {V} [ b ] \\cdot \\alpha),\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.794 + ], + "angle": 0, + "content": "\\[\n\\text {w h e r e} [ b ] = [ i B: (i + 1) B, :, d ], \\quad i \\in \\{0, 1, \\dots , T / B - 1 \\} \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.484, + 0.901 + ], + "angle": 0, + "content": "where \\([iB:(i + 1)B, :, d]\\) denotes the aggregated features from timesteps \\(iB\\) to \\((i + 1)B\\). For instance, with \\(T = 4\\) and \\(B = 2\\), features are grouped in pairs of timesteps [0,1], [2,3]. By processing blocks of timesteps together, we capture local temporal relationships while maintaining manageable computation. This approach requires storing only a subset of temporal information at any given time, reduc" + }, + { + "type": "table_caption", + "bbox": [ + 0.512, + 0.322, + 0.908, + 0.392 + ], + "angle": 0, + "content": "Table 1. Computational complexity and energy consumption of different attention mechanisms. \\( E_{MAC} \\) and \\( E_{AC} \\) are energy costs for MAC and AC operations, \\( N \\), \\( D \\), \\( T \\) denote patches, dimension, timesteps, and \\( S_{Q} \\), \\( S_{K} \\), \\( S_{V} \\) are firing rates of \\( Q \\), \\( K \\), \\( V \\). SDT represents Spike-driven Transformer." + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.403, + 0.907, + 0.516 + ], + "angle": 0, + "content": "
MethodSTComplexityEnergy
ViT [11]✘O(N2D)EMAC·N2D
ViViT [2]✓O(T2N2D)EMAC·T2N2D
Spikformer [64]✘O(TND2)EAC·TND2·(SQ+SK+SV)
SDT [53]✘O(TND)EAC·TND·(SQ+SK)
SDT-V2 [52]✘O(TND2)EAC·TND2·(SQ+SK+SV)
QKFormer [62]✘O(TND2)EAC·TND2·(SQ+SK+SV)
STAtten✓O(TND2)EAC·TND2·(SQ+SK+SV)
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.547, + 0.907, + 0.592 + ], + "angle": 0, + "content": "ing memory overhead. We set block size \\( B = 2 \\) with 4 timesteps for static datasets and \\( B = 4 \\) with 16 timesteps for neuromorphic datasets." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.596, + 0.908, + 0.656 + ], + "angle": 0, + "content": "Unlike traditional transformers that require Softmax operations, the binary \\(\\mathbf{Q}\\), \\(\\mathbf{K}\\), and \\(\\mathbf{V}\\) allow flexible reordering of computations. Generalizing to arbitrary number of timesteps \\(T\\), the computation follows:" + }, + { + "type": "equation", + "bbox": [ + 0.524, + 0.674, + 0.905, + 0.702 + ], + "angle": 0, + "content": "\\[\n\\mathbf {Q}, \\mathbf {K}, \\mathbf {V} \\in \\mathbb {R} ^ {\\frac {T}{B} \\times N \\times D} \\xrightarrow {\\text {r e s h a p e}} \\mathbf {Q}, \\mathbf {K}, \\mathbf {V} \\in \\mathbb {R} ^ {\\left(\\frac {T}{B} N\\right) \\times D} \\tag {9}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.544, + 0.7, + 0.871, + 0.718 + ], + "angle": 0, + "content": "\\[\n\\mathbf {K} ^ {\\top} \\mathbf {V} \\in \\mathbb {R} ^ {D \\times D} \\xrightarrow {\\text {a t t e n t i o n}} \\mathbf {Q} (\\mathbf {K} ^ {\\top} \\mathbf {V}) \\in \\mathbb {R} ^ {\\frac {T}{B} \\times N \\times D}.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.909, + 0.901 + ], + "angle": 0, + "content": "The computation sequence reduces intermediate memory requirements through early \\(\\mathbf{K}^{\\top}\\mathbf{V}\\) calculation while preserving cross-timestep information flow within blocks. This implementation maintains computational complexity of \\(\\mathcal{O}(TND^{2})\\), matching spatial-only attention [42, 52, 53, 62, 64] while incorporating temporal dependencies. By avoiding the quadratic complexity in timesteps and the number of tokens (\\(T^{2}N^{2}\\)) that would result from full temporal correlation, our method scales efficiently to longer sequences. Fig. 4(b) illustrates the complete computation process with corresponding tensor dimensions." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "13951" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.486, + 0.133 + ], + "angle": 0, + "content": "Table 2. Performance comparison on sequential CIFAR10/100 datasets, demonstrating the effectiveness of spatial-temporal attention in capturing long-term dependencies." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.142, + 0.485, + 0.27 + ], + "angle": 0, + "content": "
Methods-CIFAR10 (%)s-CIFAR100 (%)
PLIF [14]81.4753.38
Spikformer [64]79.2957.17
STAtten + [64]82.4558.75
Spike-driven Transformer [53]80.3261.08
STAtten + [53]83.4164.30
SpikingReformer [42]79.2557.48
STAtten + [42]81.8458.30
" + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.291, + 0.471, + 0.308 + ], + "angle": 0, + "content": "4.3. STAtten with Existing Spiking Transformers" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.314, + 0.484, + 0.405 + ], + "angle": 0, + "content": "STAtten is designed to be a plug-and-play module that can enhance various spike-based transformer architectures, including Spikformer [64], Spike-driven Transformer [53], Spike-driven Transformer-V2 [52], SpikingReformer [42], and QKformer [62] while preserving their unique characteristics." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.406, + 0.484, + 0.631 + ], + "angle": 0, + "content": "The key to successful integration lies in preserving each architecture's distinct residual connection strategies while replacing their spatial-only attentions with STAtten. Two primary residual connection mechanisms are prevalent in current architectures: Spike Element-Wise (SEW) shortcut [13] and Membrane-Shortcut (MS) [20]. In architectures employing SEW shortcuts, such as Spikformer and QKFormer, spike activations are directly added to the attention output. In contrast, Spike-driven Transformer, Spikedriven Transformer-V2, and SpikingReformer utilize MSshortcuts that propagate membrane potentials rather than spikes Our STAtten module maintains compatibility with both connection types. We detail the existing spike-based attention mechanisms of each architecture and our integration strategy." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.632, + 0.484, + 0.678 + ], + "angle": 0, + "content": "Spikformer [64]: Spikformer proposes Spiking Self Attention (SSA) module which is inspired by the vanilla self-attention of ViT [11] as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.129, + 0.686, + 0.482, + 0.705 + ], + "angle": 0, + "content": "\\[\nS S A (\\mathbf {Q}, \\mathbf {K}, \\mathbf {V}) = \\operatorname {L I F} (\\mathbf {Q} \\odot \\mathbf {K} ^ {\\top} \\odot \\mathbf {V} \\cdot \\alpha), \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.713, + 0.484, + 0.819 + ], + "angle": 0, + "content": "where \\(\\odot\\) is matrix multiplication. This work establishes two key principles for spike-based attention: non-softmax operation and flexible Q, K, V operation ordering. While these principles form the foundation for subsequent spike-based architectures, the SSA module processes features independently at each timestep. We address this limitation by replacing the attention mechanism with STAtten." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.819, + 0.484, + 0.865 + ], + "angle": 0, + "content": "Spike-driven Transformer [53]: Spike-driven Transformer introduces Spike-Driven Self-Attention (SDSA), which reformulates attention computation as:" + }, + { + "type": "equation", + "bbox": [ + 0.12, + 0.873, + 0.484, + 0.905 + ], + "angle": 0, + "content": "\\[\nS D S A (\\mathbf {Q}, \\mathbf {K}, \\mathbf {V}) = \\mathbf {Q} \\otimes \\operatorname {L I F} \\left(\\sum_ {c} (\\mathbf {K} \\otimes \\mathbf {V})\\right), \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.198 + ], + "angle": 0, + "content": "where \\(\\otimes\\) denotes element-wise multiplication and \\(\\sum_{c}\\) represents column-wise summation. While SDSA achieves computational efficiency through element-wise operations, this design limits feature correlation even within the spatial domain due to the element-wise multiplication. By integrating STAtten into this architecture, we enable comprehensive spatial-temporal feature correlation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.198, + 0.906, + 0.304 + ], + "angle": 0, + "content": "SpikingReformer [61]: SpikingReformer identifies two key challenges in adapting conventional Q, K, V attention to spike-based transformers: (1) floating-point matrix multiplication between Q, K, V and (2) Softmax operations involving exponentiation and division. To address these challenges, the authors propose Dual Spike Self-Attention (DSSA):" + }, + { + "type": "equation", + "bbox": [ + 0.572, + 0.314, + 0.905, + 0.333 + ], + "angle": 0, + "content": "\\[\nD S S A (\\mathbf {X}) = \\operatorname {L I F} \\left(\\mathbf {X} \\odot \\mathbf {W} ^ {\\top} \\odot \\mathbf {X} ^ {\\top} \\cdot \\alpha\\right), \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.344, + 0.906, + 0.42 + ], + "angle": 0, + "content": "where \\(\\mathbf{X}\\) and \\(\\mathbf{W}\\) denote binary input and linear projection, respectively. We try to verify that these concerns can be addressed by replacing DSSA with STAtten, showing that \\(\\mathbf{Q}\\), \\(\\mathbf{K}\\), \\(\\mathbf{V}\\) matrix multiplication is still compatible with spike-based processing when properly designed." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.42, + 0.906, + 0.48 + ], + "angle": 0, + "content": "QKFormer [62]: QKFormer implements dual attention mechanisms: Q-K Channel Attention (QKTA) and SSA. While SSA follows the Spikformer design, QKTA introduces a token-wise attention:" + }, + { + "type": "equation", + "bbox": [ + 0.606, + 0.49, + 0.905, + 0.521 + ], + "angle": 0, + "content": "\\[\nQ K T A = \\operatorname {L I F} \\left(\\sum_ {N} (\\mathbf {Q})\\right) \\otimes \\mathbf {K}, \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.533, + 0.906, + 0.609 + ], + "angle": 0, + "content": "where \\(\\sum_{N}\\) represents token-wise summation. QKFormer redesigns the architecture by incorporating embedding layers in each encoder. In this architecture, we only replace SSA with STAtten as SSA is the primary mechanism responsible for capturing feature relationships." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.61, + 0.906, + 0.733 + ], + "angle": 0, + "content": "As we only modify the self-attention module, the original information flow through residual connections remains unchanged, maintaining the architecture's core characteristics. After computing STAtten across temporal blocks, the features are concatenated and processed through subsequent layers: \\(\\mathbf{F}_{\\mathrm{out}} = \\mathrm{Concat}(\\mathrm{STAtten}(\\mathbf{X}[iB:(i + 1)B])_{i = 0}^{T / B - 1})\\) This concatenated feature then flows through convolutional and batch normalization layers: \\(\\mathbf{Z} = \\mathrm{BN}(\\mathrm{Conv}(\\mathbf{F}_{\\mathrm{out}}))\\)" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.742, + 0.862, + 0.759 + ], + "angle": 0, + "content": "4.4. Complexity and Energy of Self-attention" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.906, + 0.841 + ], + "angle": 0, + "content": "We analyze both computational complexity and theoretical energy consumption of our proposed architecture, focusing specifically on the self-attention mechanism. A comprehensive evaluation of the overall architecture's energy profile is provided in Supplementary Material." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.841, + 0.906, + 0.902 + ], + "angle": 0, + "content": "As shown in Table 1, despite incorporating spatial-temporal attention capabilities, our STAtten maintains the same computational complexity \\(\\mathcal{O}(TND^2)\\) as previous spatial-only spiking transformers. This efficiency is" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "13952" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.907, + 0.133 + ], + "angle": 0, + "content": "Table 3. Performance comparison between our methods and previous works on CIFAR10 and CIFAR100 datasets. In the architecture column, Architecture- \\( L - D \\) represents \\( L \\) number of encoder blocks and \\( D \\) hidden dimensions. All baseline architectures of spike-based transformer and corresponding STAtten implementations are trained from scratch for fair comparison." + }, + { + "type": "table", + "bbox": [ + 0.182, + 0.143, + 0.819, + 0.359 + ], + "angle": 0, + "content": "
MethodArchitectureTimestepCIFAR10 (%)CIFAR100 (%)
tdBN [60]ResNet19492.92-
PLIF [14]ConvNet893.50-
Dspike [31]ResNet18694.2574.24
DSR [36]ResNet182095.4078.50
DIET-SNN [39]VGG16592.7069.67
SNASNet [24]ConvNet593.6473.04
Spikformer [64]Spikformer-4-384493.9975.06
STAtten + [64]Spikformer-4-384494.3675.85
Spike-driven Transformer [53]Spike-driven Transformer-2-512495.6078.40
STAtten + [53]Spike-driven Transformer-2-512496.0379.85
SpikingReformer [42]SpikingReformer-6-384495.0377.16
STAtten + [42]SpikingReformer-6-384495.2677.90
QKFormer [62]QKFormer-4-384495.1279.79
STAtten + [62]QKFormer-4-384495.3580.20
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.375, + 0.482, + 0.603 + ], + "angle": 0, + "content": "achieved through our non-Softmax design, which enables flexible reordering of Q, K, V operations. In contrast, conventional ANN-based spatial-temporal attention methods like ViViT [2] require substantially higher complexity \\((\\mathcal{O}(T^2 N^2 D))\\) due to their softmax-constrained computation order. For energy consumption estimation, we follow the floating-point operations (FLOPs) implementation in 45nm technology [18]. The energy costs are categorized into \\(E_{MAC}\\) for multiply-accumulate operations corresponding to 32-/8-bit ANN computations, and \\(E_{AC}\\) for accumulate operations used in binary SNN calculations. The final energy consumption is modulated by the firing rates \\((S_Q, S_K, \\text{and} S_V)\\) of the respective Q, K, and V. A detailed analysis of these energy measurements and their implications is presented in Section 5.3." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.616, + 0.225, + 0.634 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.641, + 0.483, + 0.761 + ], + "angle": 0, + "content": "We verify our proposed methods on both static and dynamic datasets, e.g., CIFAR10/100 [26], ImageNet [7], CIFAR10-DVS [29], and N-Caltech101 [37] datasets. We use direct coding to convert float pixel values into binary spikes [49]. In all our experiments, the spiking transformer STAtten models are trained from scratch. The detailed training strategy for each dataset is explained in the Supplementary Material." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.762, + 0.483, + 0.837 + ], + "angle": 0, + "content": "Notation: In the experiments, we use the notation STAtten \\(+\\left[\\cdot \\right]\\) to denote STAtten implementation based on the backbone architecture used in \\(\\left[\\cdot \\right]\\). E.g., STAtten+[64] refers to training a STAtten model with Spikformer backbone from scratch." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.848, + 0.426, + 0.865 + ], + "angle": 0, + "content": "5.1. Sequential CIFAR10/100 Classification" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.903 + ], + "angle": 0, + "content": "We first evaluate our spatial-temporal attention mechanism's capability to capture long-term temporal dependen" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.375, + 0.908, + 0.618 + ], + "angle": 0, + "content": "cies using sequential CIFAR10 and CIFAR100 datasets. In this task, we transform the standard image classification task into a temporal sequence processing task by dividing each input image column-wise, shown in Fig. 1, where each column serves as input for one timestep. This results in a sequence length of 32 timesteps, corresponding to the image width. To accommodate this 1-D sequential input, we modify the convolutional, batch normalization, and pooling layers from 2-D to 1-D operations. As shown in Table 2, STAtten outperforms existing spatial-only attention methods on both datasets. Specifically, in Spike-driven Transformer [53], our approach achieves improvements of \\(3.09\\%\\) and \\(3.22\\%\\) on s-CIFAR10 and s-CIFAR100, respectively, compared to the baseline models. This performance gain demonstrates STAtten's capability in capturing long-range temporal dependencies." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.623, + 0.72, + 0.639 + ], + "angle": 0, + "content": "5.2. Performance Analysis" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.643, + 0.907, + 0.78 + ], + "angle": 0, + "content": "CIFAR10/100. The CIFAR10/100 datasets comprise static images of \\(32 \\times 32\\) pixels, with 50,000 training and 10,000 test images. As shown in Table 3, STAtten improves the performance of various spike-based transformers, achieving \\(95.35\\%\\) accuracy on CIFAR10 and \\(80.20\\%\\) on CIFAR100 when integrated with QKFormer. These improvements over spatial-only attention methods are consistent across different architectures, demonstrating the benefits of spatial-temporal processing." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.909, + 0.902 + ], + "angle": 0, + "content": "ImageNet. ImageNet is a large-scale image dataset with images sized \\(224 \\times 224\\) pixels, comprising approximately 1.2 million training images and 50,000 validation images across 1,000 classes. As shown in Table 4, STAtten enhances the performance of spike-based transformers. With Spike-driven Transformer [53], it achieves \\(78.11\\%\\) accuracy at \\(224 \\times 224\\) resolution and \\(78.39\\%\\) at \\(288 \\times 288\\) resolution. Integration with Spike-driven Transformer-V2" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13953" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.908, + 0.133 + ], + "angle": 0, + "content": "Table 4. Performance comparison between our methods and previous works on ImageNet datasets. Note that * represents the inference accuracy with \\(288 \\times 288\\) resolution. In the architecture column, \\(L-D\\) represents \\(L\\) number of encoder blocks and \\(D\\) hidden dimensions. † represents our implementation." + }, + { + "type": "table", + "bbox": [ + 0.145, + 0.142, + 0.855, + 0.428 + ], + "angle": 0, + "content": "
MethodTypeArchitectureParam(M)TimestepAccuracy(%)
Vision Transformer [11]ANNViT-B/1686.59177.90
DeiT [46]ANNDeiT-B86.59181.80
Swin Transformer [33]ANNSwin Transformer-B87.77183.50
TET [9]SNNSEW-ResNet3421.79468.00
Spiking ResNet [19]SNNResNet5025.5635072.75
tdBN [60]SNNResNet3421.79663.72
SEW-ResNet [13]SNNSEW-ResNet15260.19469.26
MS-ResNet [20]SNNResNet10478.37574.21 / 76.02*
Att-MS-ResNet [51]SNNResNet10478.37477.08*
Spikformer [64]SNNSpikformer-8-76866.34474.81
SpikingReformer [42]SNNSpikingReformer-6-102466.38478.77 / 79.40*
QKFormer [62]SNNQKFormer-6-102464.96484.22
Spike-driven Transformer [53]SNNSpike-driven Transformer-8-51229.68474.57
SNNSpike-driven Transformer-10-51236.01474.66
SNNSpike-driven Transformer-8-76866.34476.32 / 77.07*
STAtten + [53]SNNSpike-driven Transformer-8-51229.68476.18 / 76.56*
SNNSpike-driven Transformer-8-76866.34478.11 / 78.39*
Spike-driven Transformer-V2 [52]SNNSpike-driven Transformer-V2-8-51255.4479.49† / 79.98*†
STAtten + [52]SNNSpike-driven Transformer-V2-8-51255.4479.85 / 80.67*
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.44, + 0.908, + 0.484 + ], + "angle": 0, + "content": "Table 5. Performance comparison between our methods and previous works on CIFAR10-DVS and N-Caltech101 datasets. In the architecture column, \\( L - D \\) represents \\( L \\) number of encoder blocks and \\( D \\) hidden dimensions. All baseline architectures of spike-based transformer and corresponding STAtten implementations are trained from scratch for fair comparison." + }, + { + "type": "table", + "bbox": [ + 0.159, + 0.493, + 0.84, + 0.721 + ], + "angle": 0, + "content": "
MethodArchitectureTimestepCIFAR10-DVS (%)N-Caltech101 (%)
tdBN [60]ResNet191067.80-
PLIF [14]ConvNet2074.80-
Dspike [31]ResNet181075.40-
DSR [36]ResNet182077.27-
SEW-ResNet [13]ConvNet2074.80-
TT-SNN [27]ResNet346-77.80
NDA [32]VGG111079.678.2
Spikformer [64]Spikformer-2-2561680.9-
Spike-driven Transformer [53]Spike-driven Transformer-2-2561680.081.80
STAtten + [53]Spike-driven Transformer-2-2561681.183.15
SpikingReformer [42]SpikingReformer-4-3841678.8081.29
STAtten + [42]SpikingReformer-4-3841680.6081.95
QKFormer [62]QKFormer-4-3841682.9083.58
STAtten + [62]SpikingReformer-4-3841683.9084.25
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.73, + 0.483, + 0.851 + ], + "angle": 0, + "content": "[52] further improves performance to \\(79.85\\%\\) and \\(80.67\\%\\) at respective resolutions. It is noteworthy that Spike-driven Transformer [53] achieves comparable or better performance than recent Spikformer models while using only 29.68M parameters. Furthermore, STAtten with Spik-driven Transformer-V2 achieves even higher accuracy than SpikingReformer and Spikformer while using around \\(20\\%\\) fewer parameters." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Neuromorphic Datasets. We evaluate our approach on two neuromorphic vision datasets: CIFAR10-DVS [29] and N-Caltech101 [37]. CIFAR10-DVS contains 10,000 sam" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.73, + 0.907, + 0.898 + ], + "angle": 0, + "content": "plies (9,000 training, 1,000 testing) captured using a Dynamic Vision Sensor (DVS) from static CIFAR10 images. N-Caltech101, derived from the Caltech101 dataset [28], comprises 8,831 DVS image samples. Following standard protocols, we resize CIFAR10-DVS samples to \\(64 \\times 64\\) and apply the augmentation strategy from [53], while N-Caltech101 samples are resized to \\(64 \\times 64\\) with NDA [32] augmentation. As shown in Table 5, integrating STAtten consistently improves performance across different architectures. With Spike-driven Transformer as the backbone, our approach achieves \\(81.1\\%\\) accuracy on CIFAR10-DVS" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13954" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.09, + 0.483, + 0.173 + ], + "angle": 0, + "content": "Table 6. Comparison of memory, energy, and accuracy on ImageNet across different transformer architectures. The precision format weight/activation bit-width and the memory include the size of model parameters and activations. We show the memory consumption for both 32-bit and 8-bit weight spiking models. SDT represents Spike-driven Transformer." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.177, + 0.485, + 0.315 + ], + "angle": 0, + "content": "
MethodPrecisionTimestepMemory (MB)Energy (mJ)
ViT-B/16 [11]32/321351.8254.84
Liu (ViT-B) [34]8/8187.96110.8
Spikformer [64]32(8)/14285.99 (86.48)21.48
SpikingReformer [42]32(8)/14273.57 (85.4)8.76
QKFormer [62]32(8)/14280.28 (84.99)38.91
SDT [53]32(8)/14283.97 (87.42)12.42
STAtten + [53]32(8)/14283.97 (87.42)12.36
SDT-V2 [52]32(8)/14250.22 (84.02)52.40
STAtten + [53]32(8)/14250.22 (84.02)52.38
" + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.332, + 0.28, + 0.46 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.182, + 0.468, + 0.202, + 0.48 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.289, + 0.332, + 0.475, + 0.461 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.371, + 0.467, + 0.393, + 0.48 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.487, + 0.483, + 0.543 + ], + "angle": 0, + "content": "Figure 5. Accuracy comparison with respect to the number of parameters on (a) CIFAR100 and (b) Sequential CIFAR100 datasets. Spike-driven Transformer [53] is used as the baseline of spatial-only architecture." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.548, + 0.483, + 0.639 + ], + "angle": 0, + "content": "and \\(83.15\\%\\) on N-Caltech101, improving the baseline by \\(1.1\\%\\) and \\(1.35\\%\\) respectively. The enhancement extends to other architectures as well, with SpikingReformer achieving \\(80.60\\%\\) on CIFAR10-DVS and \\(81.95\\%\\) on N-Caltech101, and QKFormer reaching \\(83.90\\%\\) and \\(84.25\\%\\) respectively, which is the state-of-the-art result." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.649, + 0.357, + 0.666 + ], + "angle": 0, + "content": "5.3. Memory and Energy Analysis" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.675, + 0.483, + 0.901 + ], + "angle": 0, + "content": "We analyze memory and energy consumption during ImageNet inference. As shown in Table 6, traditional transformer, ViT-B/16, with 32-bit precision requires high memory (351.8 MB) and energy consumption (254.84 mJ). Even with 8-bit precision, Liu [34] still consumes \\(110.8\\mathrm{mJ}\\) of energy. In contrast, spike-based transformers with 32-bit weights and binary activations demonstrate significant energy efficiency. Notably, STAtten integrated with SDT and SDT-V2 achieves similar energy consumption to the baseline architectures, showing that our spatial-temporal processing can be realized without additional energy overhead. Correspondingly, STAtten attains comparable performance with fewer parameters, as detailed in the next section. The methodology for energy calculations is provided in the Supplementary Material." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.091, + 0.673, + 0.108 + ], + "angle": 0, + "content": "5.4. Model Capacity" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.114, + 0.907, + 0.34 + ], + "angle": 0, + "content": "To further evaluate the capacity of our STAtten in spike-based transformers, we compare the accuracy based on the number of trainable parameters through ablation studies. Fig. 5 shows the performance scaling across different model sizes on both (a) standard CIFAR100 and (b) sequential CIFAR100 datasets. Spike-driven Transformer [53] is used as the baseline of spatial-only architecture. As shown in Fig. 5(a), on CIFAR100, STAtten demonstrates strong performance across various model sizes ranging from 2.56M to 22.97M parameters. The accuracy improvement (\\(\\sim\\) 0.5-1.0%) is consistently maintained across different architectural scales. For Sequential CIFAR100 (Fig. 5(b)), where temporal dependencies are more crucial, STAtten shows particularly robust performance improvement (\\(\\sim\\) 3-5%) even with smaller architectures." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.347, + 0.634, + 0.361 + ], + "angle": 0, + "content": "5.5. Limitation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.369, + 0.909, + 0.612 + ], + "angle": 0, + "content": "The hardware implementation of STAtten faces challenges despite its improved accuracy. Full spatial-temporal attention without block partitioning is impractical for neuromorphic deployment, as it requires complete temporal information. Our block-wise approach partially addresses this by processing temporal information in steps, where each step corresponds to one block size. However, even with this optimization, deployment remains challenging on traditional neuromorphic chips like TrueNorth [1] and Loihi [6] that process information step-by-step. A potential solution lies in layer-by-layer neuromorphic chips [22, 54, 56, 63] that support layer-wise processing across timesteps. This architectural shift provides an efficient implementation path for STAtten's block-wise processing. Future integration of parallel LIF neurons [15] could further accelerate computation in layer-by-layer architectures." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.618, + 0.634, + 0.633 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.643, + 0.907, + 0.794 + ], + "angle": 0, + "content": "This paper proposes a block-wise spatial-temporal attention mechanism, STAtten, that enhances spike-based transformers. Through block-wise processing and leveraging non-softmax properties, STAtten captures spatial-temporal information without additional overhead. STAtten can be integrated into existing architectures while preserving their energy efficiency, consistently improving performance across different backbones. Experimental results across both static and neuromorphic datasets validate the effectiveness of our approach." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.801, + 0.666, + 0.817 + ], + "angle": 0, + "content": "Acknowledgment" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.907, + 0.901 + ], + "angle": 0, + "content": "This work was supported in part by CoCoSys, a JUMP2.0 center sponsored by DARPA and SRC, the National Science Foundation (CAREER Award, Grant #2312366, Grant #2318152), the DARPA Young Faculty Award and the DoEMMICC center SEA-CROGS (Award #DE-SC0023198)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13955" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.212 + ], + "angle": 0, + "content": "[1] Filipp Akopyan, Jun Sawada, Andrew Cassidy, Rodrigo Alvarez-Icaza, John Arthur, Paul Merolla, Nabil Imam, Yutaka Nakamura, Pallab Datta, Gi-Joon Nam, et al. Truenorth: Design and tool flow of a 65 mw 1 million neuron programmable neurosynaptic chip. IEEE transactions on computer-aided design of integrated circuits and systems, 34 (10):1537-1557, 2015. 1, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.215, + 0.484, + 0.284 + ], + "angle": 0, + "content": "[2] Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lucic, and Cordelia Schmid. Vivit: A video vision transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6836-6846, 2021. 1, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.286, + 0.483, + 0.327 + ], + "angle": 0, + "content": "[3] Anthony N Burkitt. A review of the integrate-and-fire neuron model: I. homogeneous synaptic input. Biological cybernetics, 95:1-19, 2006. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.329, + 0.484, + 0.399 + ], + "angle": 0, + "content": "[4] Lili Chen, Kevin Lu, Aravind Rajeswaran, Kimin Lee, Aditya Grover, Misha Laskin, Pieter Abbeel, Aravind Srinivas, and Igor Mordatch. Decision transformer: Reinforcement learning via sequence modeling. Advances in neural information processing systems, 34:15084-15097, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.401, + 0.483, + 0.469 + ], + "angle": 0, + "content": "[5] Sayeed Shafayet Chowdhury, Isha Garg, and Kaushik Roy. Spatio-temporal pruning and quantization for low-latency spiking neural networks. In 2021 International Joint Conference on Neural Networks (IJCNN), pages 1-9. IEEE, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.472, + 0.484, + 0.542 + ], + "angle": 0, + "content": "[6] Mike Davies, Narayan Srinivasa, Tsung-Han Lin, Gautham Chinya, Yongqiang Cao, Sri Harsha Choday, Georgios Dimou, Prasad Joshi, Nabil Imam, Shweta Jain, et al. Loihi: A neuromorphic manycore processor with on-chip learning. IEEE Micro, 38(1):82-99, 2018. 1, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.544, + 0.484, + 0.6 + ], + "angle": 0, + "content": "[7] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.601, + 0.483, + 0.643 + ], + "angle": 0, + "content": "[8] Shikuang Deng and Shi Gu. Optimal conversion of conventional artificial neural networks to spiking neural networks. arXiv preprint arXiv:2103.00476, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.645, + 0.484, + 0.699 + ], + "angle": 0, + "content": "[9] Shikuang Deng, Yuhang Li, Shanghang Zhang, and Shi Gu. Temporal efficient training of spiking neural network via gradient re-weighting. arXiv preprint arXiv:2202.11946, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.703, + 0.483, + 0.758 + ], + "angle": 0, + "content": "[10] Peter U Diehl et al. Fast-classifying, high-accuracy spiking deep networks through weight and threshold balancing. In 2015 International joint conference on neural networks (IJCNN), pages 1–8. IEEE, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.76, + 0.483, + 0.843 + ], + "angle": 0, + "content": "[11] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 1, 3, 4, 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.845, + 0.483, + 0.9 + ], + "angle": 0, + "content": "[12] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. International journal of computer vision, 88:303-338, 2010. 3" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.147 + ], + "angle": 0, + "content": "[13] Wei Fang, Zhaofei Yu, Yanqi Chen, Tiejun Huang, Timothee Masquelier, and Yonghong Tian. Deep residual learning in spiking neural networks. Advances in Neural Information Processing Systems, 34:21056-21069, 2021. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.15, + 0.906, + 0.233 + ], + "angle": 0, + "content": "[14] Wei Fang, Zhaofei Yu, Yanqi Chen, Timothee Masquelier, Tiejun Huang, and Yonghong Tian. Incorporating learnable membrane time constant to enhance learning of spiking neural networks. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2661-2671, 2021. 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.234, + 0.905, + 0.303 + ], + "angle": 0, + "content": "[15] Wei Fang, Zhaofei Yu, Zhaokun Zhou, Ding Chen, Yanqi Chen, Zhengyu Ma, Timothée Masquelier, and Yonghong Tian. Parallel spiking neurons with high efficiency and ability to learn long-term dependencies. Advances in Neural Information Processing Systems, 36, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.305, + 0.905, + 0.374 + ], + "angle": 0, + "content": "[16] Marylou Gabrie, Andre Manoel, Clément Luneau, Nicolas Macris, Florent Krzakala, Lenka Zdeborova, et al. Entropy and mutual information in models of deep neural networks. Advances in neural information processing systems, 31, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.376, + 0.905, + 0.431 + ], + "angle": 0, + "content": "[17] Bing Han and Kaushik Roy. Deep spiking neural network: Energy efficiency through time based coding. In European Conference on Computer Vision, pages 388-404. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.433, + 0.905, + 0.488 + ], + "angle": 0, + "content": "[18] Mark Horowitz. 1.1 computing's energy problem (and what we can do about it). In 2014 IEEE international solid-state circuits conference digest of technical papers (ISSCC), pages 10-14. IEEE, 2014. 6, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.49, + 0.905, + 0.532 + ], + "angle": 0, + "content": "[19] Yangfan Hu, Huajin Tang, and Gang Pan. Spiking deep residual networks. IEEE Transactions on Neural Networks and Learning Systems, 34(8):5200-5205, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.533, + 0.905, + 0.588 + ], + "angle": 0, + "content": "[20] Yifan Hu, Lei Deng, Yujie Wu, Man Yao, and Guoqi Li. Advancing spiking neural networks toward deep residual learning. IEEE Transactions on Neural Networks and Learning Systems, 2024. 1, 2, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.59, + 0.905, + 0.657 + ], + "angle": 0, + "content": "[21] Seijoon Kim, Seongsik Park, Byunggook Na, and Sungroh Yoon. Spiking-yolo: spiking neural network for energy-efficient object detection. In Proceedings of the AAAI conference on artificial intelligence, pages 11270-11277, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.66, + 0.906, + 0.757 + ], + "angle": 0, + "content": "[22] Sangyeob Kim, Soyeon Kim, Seongyon Hong, Sangjin Kim, Donghyeon Han, and Hoi-Jun Yoo. C-dnn: A 24.5-85.8 tops/w complementary-deep-neural-network processor with heterogeneous cnn/snn core architecture and forward-gradient-based sparsity generation. In 2023 IEEE International Solid-State Circuits Conference (ISSCC), pages 334-336. IEEE, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.758, + 0.905, + 0.815 + ], + "angle": 0, + "content": "[23] Youngeun Kim, Joshua Chough, and Priyadarshini Panda. Beyond classification: Directly training spiking neural networks for semantic segmentation. Neuromorphic Computing and Engineering, 2(4):044015, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.817, + 0.905, + 0.872 + ], + "angle": 0, + "content": "[24] Youngeun Kim, Yuhang Li, Hyoungseob Park, Yeshwanth Venkatesha, and Priyadarshini Panda. Neural architecture search for spiking neural networks. In European Conference on Computer Vision, pages 36-56. Springer, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[25] Youngeun Kim, Yuhang Li, Hyoungseob Park, Yeshwanth Venkatesha, Anna Hambitzer, and Priyadarshini Panda. Ex" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13956" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.093, + 0.483, + 0.134 + ], + "angle": 0, + "content": "ploring temporal information dynamics in spiking neural networks. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 8308-8316, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.483, + 0.163 + ], + "angle": 0, + "content": "[26] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.164, + 0.483, + 0.232 + ], + "angle": 0, + "content": "[27] Donghyun Lee, Ruokai Yin, Youngeun Kim, Abhishek Moitra, Yuhang Li, and Priyadarshini Panda. Tt-snn: tensor train decomposition for efficient spiking neural network training. In 2024 Design, Automation & Test in Europe Conference & Exhibition (DATE), pages 1-6. IEEE, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.233, + 0.481, + 0.26 + ], + "angle": 0, + "content": "[28] FF Li, M Andreeto, M Ranzato, and P Perona. Caltech 101 (1.0)[data set]. caltechdata, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.261, + 0.482, + 0.315 + ], + "angle": 0, + "content": "[29] Hongmin Li, Hanchao Liu, Xiangyang Ji, Guoqi Li, and Luping Shi. Cifar10-dvs: an event-stream dataset for object classification. Frontiers in neuroscience, 11:244131, 2017. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.317, + 0.483, + 0.383 + ], + "angle": 0, + "content": "[30] Yuhang Li, Shikuang Deng, Xin Dong, Ruihao Gong, and Shi Gu. A free lunch from ann: Towards efficient, accurate spiking neural networks calibration. In International conference on machine learning, pages 6316-6325. PMLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.387, + 0.482, + 0.455 + ], + "angle": 0, + "content": "[31] Yuhang Li, Yufei Guo, Shanghang Zhang, Shikuang Deng, Yongqing Hai, and Shi Gu. Differentiable spike: Rethinking gradient-descent for training spiking neural networks. Advances in Neural Information Processing Systems, 34: 23426-23439, 2021. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.457, + 0.483, + 0.523 + ], + "angle": 0, + "content": "[32] Yuhang Li, Youngeun Kim, Hyoungseob Park, Tamar Geller, and Priyadarshini Panda. Neuromorphic data augmentation for training spiking neural networks. In European Conference on Computer Vision, pages 631-649. Springer, 2022. 7, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.526, + 0.483, + 0.594 + ], + "angle": 0, + "content": "[33] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.596, + 0.483, + 0.649 + ], + "angle": 0, + "content": "[34] Zhenhua Liu, Yunhe Wang, Kai Han, Wei Zhang, Siwei Ma, and Wen Gao. Post-training quantization for vision transformer. Advances in Neural Information Processing Systems, 34:28092-28103, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.651, + 0.483, + 0.691 + ], + "angle": 0, + "content": "[35] Wolfgang Maass. Networks of spiking neurons: the third generation of neural network models. Neural networks, 10 (9):1659-1671, 1997. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.693, + 0.483, + 0.775 + ], + "angle": 0, + "content": "[36] Qingyan Meng, Mingqing Xiao, Shen Yan, Yisen Wang, Zhouchen Lin, and Zhi-Quan Luo. Training high-performance low-latency spiking neural networks by differentiation on spike representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12444–12453, 2022. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.777, + 0.483, + 0.831 + ], + "angle": 0, + "content": "[37] Garrick Orchard, Ajinkya Jayawant, Gregory K Cohen, and Nitish Thakor. Converting static image datasets to spiking neuromorphic datasets using saccades. Frontiers in neuroscience, 9:159859, 2015. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.483, + 0.899 + ], + "angle": 0, + "content": "[38] Jing Pei, Lei Deng, Sen Song, Mingguo Zhao, Youhui Zhang, Shuang Wu, Guanrui Wang, Zhe Zou, Zhenzhi Wu, Wei He, et al. Towards artificial general intelligence with hybrid tianjic chip architecture. Nature, 572(7767):106-111, 2019. 1" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.135 + ], + "angle": 0, + "content": "[39] Nitin Rathi and Kaushik Roy. Diet-snn: Direct input encoding with leakage and threshold optimization in deep spiking neural networks. arXiv preprint arXiv:2008.03658, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.136, + 0.905, + 0.176 + ], + "angle": 0, + "content": "[40] Kaushik Roy, Akhilesh Jaiswal, and Priyadarshini Panda. Towards spike-based machine intelligence with neuromorphic computing. Nature, 575(7784):607-617, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.178, + 0.905, + 0.231 + ], + "angle": 0, + "content": "[41] Abhronil Sengupta, Yuting Ye, Robert Wang, Chiao Liu, and Kaushik Roy. Going deeper in spiking neural networks: Vgg and residual architectures. Frontiers in neuroscience, 13:95, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.233, + 0.905, + 0.301 + ], + "angle": 0, + "content": "[42] Xinyu Shi, Zecheng Hao, and Zhaofei Yu. Spikingresformer: Bridging resnet and vision transformer in spiking neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5610-5619, 2024. 1, 2, 4, 5, 6, 7, 8, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.303, + 0.905, + 0.344 + ], + "angle": 0, + "content": "[43] Sumit B Shrestha and Garrick Orchard. Slayer: Spike layer error reassignment in time. Advances in neural information processing systems, 31, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.345, + 0.905, + 0.413 + ], + "angle": 0, + "content": "[44] Qiaoyi Su, Yuhong Chou, Yifan Hu, Jianing Li, Shijie Mei, Ziyang Zhang, and Guoqi Li. Deep directly-trained spiking neural networks for object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6555-6565, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.415, + 0.905, + 0.469 + ], + "angle": 0, + "content": "[45] Zhengkun Tian, Jiangyan Yi, Jianhua Tao, Ye Bai, Shuai Zhang, and Zhengqi Wen. Spike-triggered nonautoregressive transformer for end-to-end speech recognition. arXiv preprint arXiv:2005.07903, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.471, + 0.905, + 0.538 + ], + "angle": 0, + "content": "[46] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In International conference on machine learning, pages 10347-10357. PMLR, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.54, + 0.905, + 0.594 + ], + "angle": 0, + "content": "[47] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.596, + 0.905, + 0.663 + ], + "angle": 0, + "content": "[48] Qingyu Wang, Tielin Zhang, Minglun Han, Yi Wang, Duzhen Zhang, and Bo Xu. Complex dynamic neurons improved spiking transformer network for efficient automatic speech recognition. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 102-109, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.665, + 0.905, + 0.719 + ], + "angle": 0, + "content": "[49] Yujie Wu, Lei Deng, Guoqi Li, Jun Zhu, Yuan Xie, and Luping Shi. Direct training for spiking neural networks: Faster, larger, better. In Proceedings of the AAAI conference on artificial intelligence, pages 1311-1318, 2019. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.721, + 0.905, + 0.775 + ], + "angle": 0, + "content": "[50] Shiting Xiao, Yuhang Li, Youngeun Kim, Donghyun Lee, and Priyadarshini Panda. Respike: residual frames-based hybrid spiking neural networks for efficient action recognition. Neuromorphic Computing and Engineering, 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.777, + 0.905, + 0.831 + ], + "angle": 0, + "content": "[51] Man Yao, Guangshe Zhao, Hengyu Zhang, Yifan Hu, Lei Deng, Yonghong Tian, Bo Xu, and Guoqi Li. Attention spiking neural networks. IEEE transactions on pattern analysis and machine intelligence, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.833, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[52] Man Yao, Jiakui Hu, Tianxiang Hu, Yifan Xu, Zhaokun Zhou, Yonghong Tian, Bo Xu, and Guoqi Li. Spike-driven transformer v2: Meta spiking neural network architecture inspiring the design of next-generation neuromorphic chips. arXiv preprint arXiv:2404.03663, 2024. 1, 2, 4, 5, 7, 8" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13957" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[53] Man Yao, Jiakui Hu, Zhaokun Zhou, Li Yuan, Yonghong Tian, Bo Xu, and Guoqi Li. Spike-driven transformer. Advances in Neural Information Processing Systems, 36, 2024. 1, 2, 3, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.219 + ], + "angle": 0, + "content": "[54] Ruokai Yin, Abhishek Moitra, Abhiroop Bhattacharjee, Youngeun Kim, and Priyadarshini Panda. Sata: Sparsity-aware training accelerator for spiking neural networks. IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems, 42(6):1926-1938, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.22, + 0.482, + 0.289 + ], + "angle": 0, + "content": "[55] Ruokai Yin, Youngeun Kim, Yuhang Li, Abhishek Moitra, Nitin Satpute, Anna Hambitzer, and Priyadarshini Panda. Workload-balanced pruning for sparse spiking neural networks. IEEE Transactions on Emerging Topics in Computational Intelligence, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.29, + 0.482, + 0.359 + ], + "angle": 0, + "content": "[56] Ruokai Yin, Youngeun Kim, Di Wu, and Priyadarshini Panda. Loas: Fully temporal-parallel dataflow for dualsparse spiking neural networks. In 2024 57th IEEE/ACM International Symposium on Microarchitecture (MICRO), pages 1107–1121. IEEE, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.36, + 0.482, + 0.43 + ], + "angle": 0, + "content": "[57] Ruokai Yin, Yuhang Li, Abhishek Moitra, and Priyadarshini Panda. Mint: Multiplier-less integer quantization for energy efficient spiking neural networks. In 2024 29th Asia and South Pacific Design Automation Conference (ASP-DAC), pages 830-835. IEEE, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.431, + 0.482, + 0.5 + ], + "angle": 0, + "content": "[58] Jiqing Zhang, Bo Dong, Haiwei Zhang, Jianchuan Ding, Felix Heide, Baocai Yin, and Xin Yang. Spiking transformers for event-based single object tracking. In Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition, pages 8801-8810, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.501, + 0.482, + 0.558 + ], + "angle": 0, + "content": "[59] Jiyuan Zhang, Lulu Tang, Zhaofei Yu, Jiwen Lu, and Tiejun Huang. Spike transformer: Monocular depth estimation for spiking camera. In European Conference on Computer Vision, pages 34-52. Springer, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.559, + 0.482, + 0.613 + ], + "angle": 0, + "content": "[60] Hanle Zheng, Yujie Wu, Lei Deng, Yifan Hu, and Guoqi Li. Going deeper with directly-trained larger spiking neural networks. In Proceedings of the AAAI conference on artificial intelligence, pages 11062-11070, 2021. 1, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.615, + 0.482, + 0.683 + ], + "angle": 0, + "content": "[61] Chenlin Zhou, Liutao Yu, Zhaokun Zhou, Han Zhang, Zhengyu Ma, Huihui Zhou, and Yonghong Tian. Spikingformer: Spike-driven residual learning for transformer-based spiking neural network. arXiv preprint arXiv:2304.11954, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.685, + 0.482, + 0.754 + ], + "angle": 0, + "content": "[62] Chenlin Zhou, Han Zhang, Zhaokun Zhou, Liutao Yu, Liwei Huang, Xiaopeng Fan, Li Yuan, Zhengyu Ma, Huihui Zhou, and Yonghong Tian. Qkformer: Hierarchical spiking transformer using qk attention. arXiv preprint arXiv:2403.16552, 2024. 1, 2, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.755, + 0.482, + 0.839 + ], + "angle": 0, + "content": "[63] PJ Zhou, Q Yu, M Chen, YC Wang, LW Meng, Y Zuo, N Ning, Y Liu, SG Hu, and GC Qiao. A 0.96 pj/sop, 30.23 k-neuron/mm 2 heterogeneous neuromorphic chip with fullerene-like interconnection topology for edge-ai computing. In 2024 IEEE International Symposium on Circuits and Systems (ISCAS), pages 1–5. IEEE, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.84, + 0.482, + 0.896 + ], + "angle": 0, + "content": "[64] Zhaokun Zhou, Yuesheng Zhu, Chao He, Yaowei Wang, Shuicheng Yan, Yonghong Tian, and Li Yuan. Spikformer: When spiking neural network meets transformer. arXiv preprint arXiv:2209.15425, 2022. 1, 2, 3, 4, 5, 6, 7, 8" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.896 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.956 + ], + "angle": 0, + "content": "13958" + } + ] +] \ No newline at end of file diff --git a/2025/Spiking Transformer with Spatial-Temporal Attention/de8548d3-fd04-4e18-9d13-559b7231ddb6_origin.pdf b/2025/Spiking Transformer with Spatial-Temporal Attention/de8548d3-fd04-4e18-9d13-559b7231ddb6_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..8a01bf81dfad8bc10a793b9956178a8bc5559c93 --- /dev/null +++ b/2025/Spiking Transformer with Spatial-Temporal Attention/de8548d3-fd04-4e18-9d13-559b7231ddb6_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3263259779cfaf8630d6d56b4e01a26743f92448f68476730a6c2ec35ee18ca7 +size 1873921 diff --git a/2025/Spiking Transformer with Spatial-Temporal Attention/full.md b/2025/Spiking Transformer with Spatial-Temporal Attention/full.md new file mode 100644 index 0000000000000000000000000000000000000000..6c376d904bb6a7866470e63fe3650dcefa5893f1 --- /dev/null +++ b/2025/Spiking Transformer with Spatial-Temporal Attention/full.md @@ -0,0 +1,341 @@ +# Spiking Transformer with Spatial-Temporal Attention + +Donghyun Lee, Yuhang Li, Youngeun Kim, Shiting Xiao, Priyadarshini Panda Department of Electrical Engineering, Yale University + +{donghyun.lee, yuhang.li, youngeun.kim, ginny.xiao, priya.panda}@yale.edu + +# Abstract + +Spike-based Transformer presents a compelling and energy-efficient alternative to traditional Artificial Neural Network (ANN)-based Transformers, achieving impressive results through sparse binary computations. However, existing spike-based transformers predominantly focus on spatial attention while neglecting crucial temporal dependencies inherent in spike-based processing, leading to suboptimal feature representation and limited performance. To address this limitation, we propose Spiking Transformer with Spatial-Temporal Attention (STAtten), a simple and straightforward architecture that efficiently integrates both spatial and temporal information in the self-attention mechanism. STAtten introduces a block-wise computation strategy that processes information in spatial-temporal chunks, enabling comprehensive feature capture while maintaining the same computational complexity as previous spatial-only approaches. Our method can be seamlessly integrated into existing spike-based transformers without architectural overhaul. Extensive experiments demonstrate that STAtten significantly improves the performance of existing spike-based transformers across both static and neuromorphic datasets, including CIFAR10/100, ImageNet, CIFAR10-DVS, and N-Caltech101. The code is available at https://github.com/Intelligent-Computing-Lab-Yale/STAtten. + +# 1. Introduction + +Spiking Neural Networks (SNNs) are emerging as a promising alternative to conventional Artificial Neural Networks (ANNs) [35, 40] due to their bio-inspired, energy-efficient computing paradigm. By leveraging sparse binary spike-based computations, SNNs can achieve significant energy savings while enabling deployment across various neuromorphic computing platforms such as TrueNorth [1], Loihi [6], and Tianjic [38]. However, the challenge lies in effectively processing and learning from these discrete, time-dependent spike patterns while maintaining computational efficiency and accuracy. Traditional SNNs have primarily + +![](images/c324ea93315e4534ed615b17fe8a2b2544ff9db8591f0479fb70d24ea953bc25.jpg) +Figure 1. Heatmaps of spatial-only attention versus our STAten on sequential CIFAR100 dataset. Input images are divided column-wise, where each column corresponds to one timestep. + +relied on convolution-based architectures adapted from successful ANN models like VGGNet [41, 49] and ResNet [20, 60]. While these architectures benefit from computational efficiency, they often struggle with binary spike activations, leading to information loss and degraded performance. This limitation has driven researchers to explore alternative architectures that can better handle sparse, temporal data while maintaining the energy benefits of spike-based computation. Transformer architectures, with their remarkable success across various domains [2, 4, 11, 47], have emerged as a promising direction for enhancing SNN capabilities. Recent works have introduced spike-formed transformers that adapt the transformer's self-attention mechanism to the spike-based domain for diverse tasks, such as, object tracking [58], computer vision [42, 50, 52, 53, 62, 64], depth estimation [59], and speech recognition [45, 48]. These architectures, such as Spikformer [64] and Spike-driven Transformer [53], implement innovative binary representations for query (Q), key (K), and value (V) computations while eliminating the computationally expensive Softmax function. This binary attention design preserves the essential characteristics of spike-based computation while enabling more efficient information processing compared to conventional transformer architectures. However, current spike-formed transformers predominantly rely on spatial-only attention, overlooking the inherently dynamic and temporal nature of spike events. To understand these differ + +ent attention patterns, we compare spatial-only and spatial-temporal attention on sequential CIFAR100, as shown in Fig. 1. The visualization reveals that spatial-only attention focuses solely on vertical (spatial) relationships within each timestep, missing crucial object features that evolve over time. In contrast, our spatial-temporal attention captures both vertical (spatial) and horizontal (temporal) relationships, enabling comprehensive feature representation of sequential inputs. + +In this work, we introduce Spiking Transformer with Spatial-Temporal Attention (STAtten), which integrates both spatial and temporal information within the self-attention mechanism. Through analysis of memory requirements and neuronal activity patterns, we design a blockwise computation strategy with local temporal correlations. By eliminating Softmax operation, we maintain the same computational complexity $(O(TND^{2}))$ as existing spike-based transformers [42, 52, 62, 64]. Furthermore, our approach can be flexibly integrated into existing architectures without structural modifications, enhancing their representational capacity. Experiments on both static and dynamic datasets demonstrate consistent performance improvements over spatial-only attention architectures. The main contributions of our work are as follows: + +- We identify limitations in current spiking transformers that rely on spatial-only attention through empirical analysis, demonstrating the importance of capturing temporal dependencies inherent in spike-based processing. +- We propose STAtten, a block-wise spatial-temporal attention mechanism for spike-based transformers that maintains the original computational complexity $(O(TND^{2}))$ . +- We introduce a flexible plug-and-play design that enables STAtten to be integrated into existing spike-based transformers without compromising their core architecture. +- Through extensive experiments across both static (CIFAR10/100, ImageNet) and neuromorphic (CIFAR10-DVS, N-Caltech101) datasets, we demonstrate that STatten consistently improves performance across different backbone architectures. + +# 2. Related Works + +Training Strategies for SNNs. Significant progress has been made in developing effective training strategies for SNNs, primarily following two approaches: ANN-to-SNN conversion and direct training. The conversion approach transforms pre-trained ANNs into SNNs, leveraging established ANN architectures to enhance SNN performance. Several studies [8, 10, 17, 30] have formalized this conversion process and demonstrated its effectiveness in transferring ANN knowledge to the spike domain. Direct training through backpropagation, while fundamental to ANNs, presents unique challenges in SNNs due to the + +non-differentiable nature of spike events. The introduction of surrogate gradients has been crucial in enabling effective backpropagation-based training in SNNs, as demonstrated in pioneering works [5, 43, 49]. Notably, direct training with surrogate gradients has shown superior performance compared to conversion methods across various tasks, including image classification [14, 20, 31, 55, 57], semantic segmentation [23], and object detection [44], while maintaining the energy efficiency inherent to spike-based computation. + +Spiking Transformer. Despite the energy efficiency of convolution-based SNNs, their performance often lags behind ANNs due to information loss from sparse binary activations. To address this limitation, researchers have adapted transformer architectures to the SNN domain, leveraging their powerful self-attention mechanisms and ability to capture long-range dependencies. This adaptation has led to several innovative spike-formed transformer architectures. Spikformer [64] introduced the first SNN-based transformer architecture, establishing foundational principles for spikeformed self-attention mechanisms. Their key innovation lies in eliminating the Softmax function from self-attention computations, justified by two key observations: (1) the binary nature of spike-based Q, K, and V values, and (2) the redundancy of Softmax scaling for binarized operations. Building upon this foundation, Spike-driven Transformer [53] enhanced the architecture with a novel self-attention block incorporating masking and sparse addition to optimize power efficiency. Additionally, inspired by MS-ResNet [20], Spike-driven Transformer reimagined residual connections to propagate membrane potential rather than spikes. Recent works have further extended these concepts, leading to architectures like SpikingReformer [42], QK-Former [62], and Spike-driven Transformer-V2 [52], each introducing unique optimizations for spike-based processing. However, existing spiking transformer architectures primarily focus on spatial attention, neglecting the temporal dynamics inherent in spike-based computation. + +# 3. Preliminary + +Leaky Integrate-and-Fire Neuron. As a fundamental building block of SNNs, the Leaky Integrate-and-Fire (LIF) neuron [3] has emerged as an important component for energy-efficient computation. The LIF neuron is a nonlinear activation function that determines whether neurons fire spikes as follows: + +$$ +\mathbf {u} [ t + 1 ] ^ {l} = \tau \mathbf {u} [ t ] ^ {l} + \mathbf {W} ^ {l} f (\mathbf {u} [ t ] ^ {l - 1}) \tag {1} +$$ + +$$ +f \left(\mathbf {u} [ t ] ^ {l}\right) = \left\{ \begin{array}{l l} 1 & \text {i f} \mathbf {u} [ t ] ^ {l} > V _ {t h}, \\ 0 & \text {o t h e r w i s e} \end{array} , \right. \tag {2} +$$ + +where, $\mathbf{u}[t]^l$ is the membrane potential in $l$ -th layer at timestep $t$ , $\tau \in (0,1]$ is the leaky factor for membrane po + +![](images/57a643a56b805f3a858c0f0fca5f6a00b73194fbb45b0993f9a3360bbee9eb6c.jpg) +Figure 2. Comparison between different self-attention with the CIFAR100 dataset. Analysis of entropy and accuracy for Temporal-only (T), Spatial-only (S), and Spatial-temporal (ST). + +![](images/5a8ef6d7cfd081febdb07f4e9dbe44f2a3e8984722cca88f84f02537a48cde77.jpg) +Figure 3. (a) Maximum batch size of block-wise STAtten and full spatial-temporal attention (without block partitioning) for running on A5000 GPU with 24GB VRAM memory. (b) Average number of active neurons after QKV computation at different timestep combinations, where [t] indicates timestep index. + +tential leakage, $\mathbf{W}^l$ is the weight of $l$ -th layer, and $f(\cdot)$ is the LIF function with firing threshold $V_{th}$ . Therefore, when the membrane $\mathbf{u}[t]^l$ is higher than $V_{th}$ , the LIF function fires a spike and the membrane potential is reset to 0. + +Vanilla Self-attention. Transformers [11, 47] have surpassed convolution-based architectures due to their unique self-attention mechanism, which can capture global features through long-range dependency. For a floating-point input tensor $\mathbf{X}_f\in \mathbb{R}^{N\times D}$ with $N$ tokens, and $D$ features, we formulate the self-attention as follows: + +$$ +\mathbf {Q} _ {f} = \mathbf {W} _ {\mathbf {Q}} \mathbf {X} _ {f}, \quad \mathbf {K} _ {f} = \mathbf {W} _ {f} \mathbf {X} _ {f}, \quad \mathbf {V} _ {f} = \mathbf {W} _ {\mathbf {V}} \mathbf {X} _ {f} \tag {3} +$$ + +$$ +\operatorname {A t t n} = \operatorname {S o f t m a x} \left(\frac {\mathbf {Q K} ^ {\top}}{\sqrt {D}}\right) \mathbf {V}, \quad \mathbf {Q} _ {f}, \mathbf {K} _ {f}, \mathbf {V} _ {f} \in \mathbb {R} ^ {N \times D}, \tag {4} +$$ + +where $\mathbf{W}_{\mathbf{Q}}, \mathbf{W}_{\mathbf{K}}, \mathbf{W}_{\mathbf{V}} \in \mathbb{R}^{D \times D}$ are linear projections for floating-point $\mathbf{Q}_f, \mathbf{K}_f, \mathbf{V}_f$ , respectively. Note that the computational complexity of attention is $\mathcal{O}(N^2 D)$ due to its matrix multiplication operation. + +Spike-based Self-attention. The first spike-based transformer work, Spikformer [64], proposes Spiking Self Attention (SSA) with binary $\mathbf{Q}$ , $\mathbf{K}$ , $\mathbf{V}$ . For binary input tensor $\mathbf{X} \in \mathbb{R}^{T \times N \times D}$ with $T$ timestep, we formulate the SSA as follows: + +$$ +\begin{array}{l} \mathbf {Q}, \mathbf {K}, \mathbf {V} = \operatorname {L I F} (\mathbf {W} _ {\mathbf {Q}} \mathbf {X}), \operatorname {L I F} (\mathbf {W} _ {\mathbf {K}} \mathbf {X}), \operatorname {L I F} (\mathbf {W} _ {\mathbf {K}} \mathbf {X}), \\ \operatorname {S S A} (\mathbf {Q K V}) = \operatorname {L I F} \left(\mathbf {Q K} ^ {\top} \mathbf {V} \cdot \alpha\right) \tag {5} \\ \end{array} +$$ + +where $\mathbf{W}_{\mathbf{Q}}, \mathbf{W}_{\mathbf{K}}, \mathbf{W}_{\mathbf{V}} \in \mathbb{R}^{D \times D}$ are linear projections for binary $\mathbf{Q}, \mathbf{K}, \mathbf{V}$ , respectively, and $\alpha$ denotes a scaling factor. + +# 4. Methodology + +In this section, we present our analysis of spike-based transformers and provide the details of our proposed methodology. We first establish the necessity of spatial-temporal attention in spike-based transformers through an information-theoretic analysis based on entropy measurements. Building upon these observations, we introduce our STAtten mechanism. + +![](images/07fccc209c2c29094fca7d2e499594afceafd65519edf3fcda38ffabfffd6219.jpg) +(a) + +![](images/012624bc4f9fb21e1dc486e9082ba65e1ad012631e1872a46aca2c29ba3a8920.jpg) +(b) + +# 4.1. Motivation of Spatial-Temporal Attention + +The feature maps of spike-based transformers at each timestep carry distinct information, as their activation distributions are dynamically influenced by both the leaky factor and reset mechanisms. This temporal aspect of information flow raises a fundamental question: how do different attention mechanisms effectively capture and utilize this spatiotemporal information? + +First, we define the three different types of attention: Spatial, Temporal, and Spatial-temporal attention with timestep $t$ , token position $n$ , and feature dimension $d$ as follows: + +$$ +\begin{array}{l} \mathrm {S} _ {-} \operatorname {A t t n} [ t,,:,: ] = \operatorname {L I F} (\mathbf {Q} [ t,:,: ] \mathbf {K} ^ {\top} [ t,:,: ] \mathbf {V} [ t,:,: ] \cdot \alpha), \\ \mathrm {T} _ {\mathbf {\Lambda}} \operatorname {A t t n} [:, n,: ] = \operatorname {L I F} (\mathbf {Q} [:, n,: ] \mathbf {K} ^ {\top} [:, n,: ] \mathbf {V} [:, n,: ] \cdot \alpha), \\ \operatorname {S T} \text {A t t n} [:,:, d ] = \operatorname {L I F} (\mathbf {Q} [:,:, d ] \mathbf {K} ^ {\top} [:,:, d ] \mathbf {V} [:,:, d ] \cdot \alpha), \tag {6} \\ \end{array} +$$ + +where S_Attn, T_Attn, and ST_Attn are spatial, temporal, and spatial-temporal attention, and the Q, K, V dimensions of S_Attn, T_Attn, and ST_Attn are $\mathbb{R}^{T\times N\times D}$ . The key distinction lies in their attention computation scope: spatial attention operates on token relationships within individual timesteps, temporal attention focuses on cross-timestep feature dependencies, and spatial-temporal attention integrates both spatial and temporal feature correlations. To quantify the information distribution in each attention mechanism, we measure Shannon entropy [16] as follows: + +$$ +H (\mathrm {A t t n}) = - \sum_ {t = 1} ^ {T} \sum_ {n = 1} ^ {N} \sum_ {d = 1} ^ {D} \hat {\mathbf {p}} [ t, n, d ] \log \hat {\mathbf {p}} [ t, n, d ], \tag {7} +$$ + +where $\hat{\mathbf{p}}[t,n,d] = \text{Softmax}(\text{Attn}[t,n,d])$ . Lower entropy values indicate more focused spike patterns, while higher entropy values suggest dispersed patterns [25]. We analyze this relationship using a pretrained Spike-driven Transformer [53] on the CIFAR100 dataset, as shown in Fig. 2. The results reveal an inverse correlation between entropy and accuracy: temporal-only attention shows high entropy (5.81) but lowest accuracy (77.7%), while + +![](images/cac2d98b2a42e1917ce9c60b58e0008f7e9e9500c4aaa35794f0fa5dcacbe75d.jpg) +(a) +Spatial-Temporal Attention (STAtten) +Figure 4. Overview of STAtten architecture. (a) Block-wise temporal attention mechanism. Binary Q, K, V tensors are partitioned into temporal blocks, where black lines indicate paired timestep processing. (b) Computation flow with tensor dimensions, where $T$ is the number of timesteps, $N$ is the number of tokens, $B$ is block size, and $D$ is the feature dimension. + +![](images/d62e9d2b639829e04d5385ee6d827087828fdbbaa9139fdbf1e9797dcc5d4954.jpg) + +spatial-temporal attention achieves lower entropy (4.85) with higher accuracy $(79.9\%)$ . These findings verify that combining spatial and temporal processing leads to more structured feature representations, providing strong motivation for spatial-temporal attention. + +However, we argue that fully correlating temporal information across all timesteps like ST_Attn in Eq. 6 introduces two challenges: (1) substantial memory requirements for storing temporal information, and (2) increased dead neurons when correlating distant timesteps due to weak spike similarity. The first challenge arises from matrix multiplication with large temporal dimensions $(TN,D)$ . To address this, we partition the temporal dimension into blocks, reducing the memory footprint for matrix operations. As shown in Fig. 3(a), this block-wise computation strategy leads to around $1.6\times$ higher batch size compared to full temporal attention. The second challenge stems from binary matrix multiplication in spike-based processing. Our analysis in Fig. 3(b) reveals that QKV computations show higher active neurons when processing same-timestep, while activity decreases with temporal distance. + +# 4.2. STAtten Mechanism + +Based on these observations, we propose Spatial-Temporal Attention (STAtten). The block-wise STAtten partitions the temporal sequence into blocks of size $B$ , as shown in Fig. 4(a). We define STAtten as follows: + +$$ +\operatorname {S T A t t e n} (\mathbf {X} [ b ]) = \operatorname {L I F} (\mathbf {Q} [ b ] \mathbf {K} ^ {\top} [ b ] \mathbf {V} [ b ] \cdot \alpha), +$$ + +$$ +\text {w h e r e} [ b ] = [ i B: (i + 1) B, :, d ], \quad i \in \{0, 1, \dots , T / B - 1 \} \tag {8} +$$ + +where $[iB:(i + 1)B, :, d]$ denotes the aggregated features from timesteps $iB$ to $(i + 1)B$ . For instance, with $T = 4$ and $B = 2$ , features are grouped in pairs of timesteps [0,1], [2,3]. By processing blocks of timesteps together, we capture local temporal relationships while maintaining manageable computation. This approach requires storing only a subset of temporal information at any given time, reduc + +Table 1. Computational complexity and energy consumption of different attention mechanisms. $E_{MAC}$ and $E_{AC}$ are energy costs for MAC and AC operations, $N$ , $D$ , $T$ denote patches, dimension, timesteps, and $S_{Q}$ , $S_{K}$ , $S_{V}$ are firing rates of $Q$ , $K$ , $V$ . SDT represents Spike-driven Transformer. + +
MethodSTComplexityEnergy
ViT [11]✘O(N2D)EMAC·N2D
ViViT [2]✓O(T2N2D)EMAC·T2N2D
Spikformer [64]✘O(TND2)EAC·TND2·(SQ+SK+SV)
SDT [53]✘O(TND)EAC·TND·(SQ+SK)
SDT-V2 [52]✘O(TND2)EAC·TND2·(SQ+SK+SV)
QKFormer [62]✘O(TND2)EAC·TND2·(SQ+SK+SV)
STAtten✓O(TND2)EAC·TND2·(SQ+SK+SV)
+ +ing memory overhead. We set block size $B = 2$ with 4 timesteps for static datasets and $B = 4$ with 16 timesteps for neuromorphic datasets. + +Unlike traditional transformers that require Softmax operations, the binary $\mathbf{Q}$ , $\mathbf{K}$ , and $\mathbf{V}$ allow flexible reordering of computations. Generalizing to arbitrary number of timesteps $T$ , the computation follows: + +$$ +\mathbf {Q}, \mathbf {K}, \mathbf {V} \in \mathbb {R} ^ {\frac {T}{B} \times N \times D} \xrightarrow {\text {r e s h a p e}} \mathbf {Q}, \mathbf {K}, \mathbf {V} \in \mathbb {R} ^ {\left(\frac {T}{B} N\right) \times D} \tag {9} +$$ + +$$ +\mathbf {K} ^ {\top} \mathbf {V} \in \mathbb {R} ^ {D \times D} \xrightarrow {\text {a t t e n t i o n}} \mathbf {Q} (\mathbf {K} ^ {\top} \mathbf {V}) \in \mathbb {R} ^ {\frac {T}{B} \times N \times D}. +$$ + +The computation sequence reduces intermediate memory requirements through early $\mathbf{K}^{\top}\mathbf{V}$ calculation while preserving cross-timestep information flow within blocks. This implementation maintains computational complexity of $\mathcal{O}(TND^{2})$ , matching spatial-only attention [42, 52, 53, 62, 64] while incorporating temporal dependencies. By avoiding the quadratic complexity in timesteps and the number of tokens ( $T^{2}N^{2}$ ) that would result from full temporal correlation, our method scales efficiently to longer sequences. Fig. 4(b) illustrates the complete computation process with corresponding tensor dimensions. + +Table 2. Performance comparison on sequential CIFAR10/100 datasets, demonstrating the effectiveness of spatial-temporal attention in capturing long-term dependencies. + +
Methods-CIFAR10 (%)s-CIFAR100 (%)
PLIF [14]81.4753.38
Spikformer [64]79.2957.17
STAtten + [64]82.4558.75
Spike-driven Transformer [53]80.3261.08
STAtten + [53]83.4164.30
SpikingReformer [42]79.2557.48
STAtten + [42]81.8458.30
+ +# 4.3. STAtten with Existing Spiking Transformers + +STAtten is designed to be a plug-and-play module that can enhance various spike-based transformer architectures, including Spikformer [64], Spike-driven Transformer [53], Spike-driven Transformer-V2 [52], SpikingReformer [42], and QKformer [62] while preserving their unique characteristics. + +The key to successful integration lies in preserving each architecture's distinct residual connection strategies while replacing their spatial-only attentions with STAtten. Two primary residual connection mechanisms are prevalent in current architectures: Spike Element-Wise (SEW) shortcut [13] and Membrane-Shortcut (MS) [20]. In architectures employing SEW shortcuts, such as Spikformer and QKFormer, spike activations are directly added to the attention output. In contrast, Spike-driven Transformer, Spikedriven Transformer-V2, and SpikingReformer utilize MSshortcuts that propagate membrane potentials rather than spikes Our STAtten module maintains compatibility with both connection types. We detail the existing spike-based attention mechanisms of each architecture and our integration strategy. + +Spikformer [64]: Spikformer proposes Spiking Self Attention (SSA) module which is inspired by the vanilla self-attention of ViT [11] as follows: + +$$ +S S A (\mathbf {Q}, \mathbf {K}, \mathbf {V}) = \operatorname {L I F} (\mathbf {Q} \odot \mathbf {K} ^ {\top} \odot \mathbf {V} \cdot \alpha), \tag {10} +$$ + +where $\odot$ is matrix multiplication. This work establishes two key principles for spike-based attention: non-softmax operation and flexible Q, K, V operation ordering. While these principles form the foundation for subsequent spike-based architectures, the SSA module processes features independently at each timestep. We address this limitation by replacing the attention mechanism with STAtten. + +Spike-driven Transformer [53]: Spike-driven Transformer introduces Spike-Driven Self-Attention (SDSA), which reformulates attention computation as: + +$$ +S D S A (\mathbf {Q}, \mathbf {K}, \mathbf {V}) = \mathbf {Q} \otimes \operatorname {L I F} \left(\sum_ {c} (\mathbf {K} \otimes \mathbf {V})\right), \tag {11} +$$ + +where $\otimes$ denotes element-wise multiplication and $\sum_{c}$ represents column-wise summation. While SDSA achieves computational efficiency through element-wise operations, this design limits feature correlation even within the spatial domain due to the element-wise multiplication. By integrating STAtten into this architecture, we enable comprehensive spatial-temporal feature correlation. + +SpikingReformer [61]: SpikingReformer identifies two key challenges in adapting conventional Q, K, V attention to spike-based transformers: (1) floating-point matrix multiplication between Q, K, V and (2) Softmax operations involving exponentiation and division. To address these challenges, the authors propose Dual Spike Self-Attention (DSSA): + +$$ +D S S A (\mathbf {X}) = \operatorname {L I F} \left(\mathbf {X} \odot \mathbf {W} ^ {\top} \odot \mathbf {X} ^ {\top} \cdot \alpha\right), \tag {12} +$$ + +where $\mathbf{X}$ and $\mathbf{W}$ denote binary input and linear projection, respectively. We try to verify that these concerns can be addressed by replacing DSSA with STAtten, showing that $\mathbf{Q}$ , $\mathbf{K}$ , $\mathbf{V}$ matrix multiplication is still compatible with spike-based processing when properly designed. + +QKFormer [62]: QKFormer implements dual attention mechanisms: Q-K Channel Attention (QKTA) and SSA. While SSA follows the Spikformer design, QKTA introduces a token-wise attention: + +$$ +Q K T A = \operatorname {L I F} \left(\sum_ {N} (\mathbf {Q})\right) \otimes \mathbf {K}, \tag {13} +$$ + +where $\sum_{N}$ represents token-wise summation. QKFormer redesigns the architecture by incorporating embedding layers in each encoder. In this architecture, we only replace SSA with STAtten as SSA is the primary mechanism responsible for capturing feature relationships. + +As we only modify the self-attention module, the original information flow through residual connections remains unchanged, maintaining the architecture's core characteristics. After computing STAtten across temporal blocks, the features are concatenated and processed through subsequent layers: $\mathbf{F}_{\mathrm{out}} = \mathrm{Concat}(\mathrm{STAtten}(\mathbf{X}[iB:(i + 1)B])_{i = 0}^{T / B - 1})$ This concatenated feature then flows through convolutional and batch normalization layers: $\mathbf{Z} = \mathrm{BN}(\mathrm{Conv}(\mathbf{F}_{\mathrm{out}}))$ + +# 4.4. Complexity and Energy of Self-attention + +We analyze both computational complexity and theoretical energy consumption of our proposed architecture, focusing specifically on the self-attention mechanism. A comprehensive evaluation of the overall architecture's energy profile is provided in Supplementary Material. + +As shown in Table 1, despite incorporating spatial-temporal attention capabilities, our STAtten maintains the same computational complexity $\mathcal{O}(TND^2)$ as previous spatial-only spiking transformers. This efficiency is + +Table 3. Performance comparison between our methods and previous works on CIFAR10 and CIFAR100 datasets. In the architecture column, Architecture- $L - D$ represents $L$ number of encoder blocks and $D$ hidden dimensions. All baseline architectures of spike-based transformer and corresponding STAtten implementations are trained from scratch for fair comparison. + +
MethodArchitectureTimestepCIFAR10 (%)CIFAR100 (%)
tdBN [60]ResNet19492.92-
PLIF [14]ConvNet893.50-
Dspike [31]ResNet18694.2574.24
DSR [36]ResNet182095.4078.50
DIET-SNN [39]VGG16592.7069.67
SNASNet [24]ConvNet593.6473.04
Spikformer [64]Spikformer-4-384493.9975.06
STAtten + [64]Spikformer-4-384494.3675.85
Spike-driven Transformer [53]Spike-driven Transformer-2-512495.6078.40
STAtten + [53]Spike-driven Transformer-2-512496.0379.85
SpikingReformer [42]SpikingReformer-6-384495.0377.16
STAtten + [42]SpikingReformer-6-384495.2677.90
QKFormer [62]QKFormer-4-384495.1279.79
STAtten + [62]QKFormer-4-384495.3580.20
+ +achieved through our non-Softmax design, which enables flexible reordering of Q, K, V operations. In contrast, conventional ANN-based spatial-temporal attention methods like ViViT [2] require substantially higher complexity $(\mathcal{O}(T^2 N^2 D))$ due to their softmax-constrained computation order. For energy consumption estimation, we follow the floating-point operations (FLOPs) implementation in 45nm technology [18]. The energy costs are categorized into $E_{MAC}$ for multiply-accumulate operations corresponding to 32-/8-bit ANN computations, and $E_{AC}$ for accumulate operations used in binary SNN calculations. The final energy consumption is modulated by the firing rates $(S_Q, S_K, \text{and} S_V)$ of the respective Q, K, and V. A detailed analysis of these energy measurements and their implications is presented in Section 5.3. + +# 5. Experiments + +We verify our proposed methods on both static and dynamic datasets, e.g., CIFAR10/100 [26], ImageNet [7], CIFAR10-DVS [29], and N-Caltech101 [37] datasets. We use direct coding to convert float pixel values into binary spikes [49]. In all our experiments, the spiking transformer STAtten models are trained from scratch. The detailed training strategy for each dataset is explained in the Supplementary Material. + +Notation: In the experiments, we use the notation STAtten $+\left[\cdot \right]$ to denote STAtten implementation based on the backbone architecture used in $\left[\cdot \right]$ . E.g., STAtten+[64] refers to training a STAtten model with Spikformer backbone from scratch. + +# 5.1. Sequential CIFAR10/100 Classification + +We first evaluate our spatial-temporal attention mechanism's capability to capture long-term temporal dependen + +cies using sequential CIFAR10 and CIFAR100 datasets. In this task, we transform the standard image classification task into a temporal sequence processing task by dividing each input image column-wise, shown in Fig. 1, where each column serves as input for one timestep. This results in a sequence length of 32 timesteps, corresponding to the image width. To accommodate this 1-D sequential input, we modify the convolutional, batch normalization, and pooling layers from 2-D to 1-D operations. As shown in Table 2, STAtten outperforms existing spatial-only attention methods on both datasets. Specifically, in Spike-driven Transformer [53], our approach achieves improvements of $3.09\%$ and $3.22\%$ on s-CIFAR10 and s-CIFAR100, respectively, compared to the baseline models. This performance gain demonstrates STAtten's capability in capturing long-range temporal dependencies. + +# 5.2. Performance Analysis + +CIFAR10/100. The CIFAR10/100 datasets comprise static images of $32 \times 32$ pixels, with 50,000 training and 10,000 test images. As shown in Table 3, STAtten improves the performance of various spike-based transformers, achieving $95.35\%$ accuracy on CIFAR10 and $80.20\%$ on CIFAR100 when integrated with QKFormer. These improvements over spatial-only attention methods are consistent across different architectures, demonstrating the benefits of spatial-temporal processing. + +ImageNet. ImageNet is a large-scale image dataset with images sized $224 \times 224$ pixels, comprising approximately 1.2 million training images and 50,000 validation images across 1,000 classes. As shown in Table 4, STAtten enhances the performance of spike-based transformers. With Spike-driven Transformer [53], it achieves $78.11\%$ accuracy at $224 \times 224$ resolution and $78.39\%$ at $288 \times 288$ resolution. Integration with Spike-driven Transformer-V2 + +Table 4. Performance comparison between our methods and previous works on ImageNet datasets. Note that * represents the inference accuracy with $288 \times 288$ resolution. In the architecture column, $L-D$ represents $L$ number of encoder blocks and $D$ hidden dimensions. † represents our implementation. + +
MethodTypeArchitectureParam(M)TimestepAccuracy(%)
Vision Transformer [11]ANNViT-B/1686.59177.90
DeiT [46]ANNDeiT-B86.59181.80
Swin Transformer [33]ANNSwin Transformer-B87.77183.50
TET [9]SNNSEW-ResNet3421.79468.00
Spiking ResNet [19]SNNResNet5025.5635072.75
tdBN [60]SNNResNet3421.79663.72
SEW-ResNet [13]SNNSEW-ResNet15260.19469.26
MS-ResNet [20]SNNResNet10478.37574.21 / 76.02*
Att-MS-ResNet [51]SNNResNet10478.37477.08*
Spikformer [64]SNNSpikformer-8-76866.34474.81
SpikingReformer [42]SNNSpikingReformer-6-102466.38478.77 / 79.40*
QKFormer [62]SNNQKFormer-6-102464.96484.22
Spike-driven Transformer [53]SNNSpike-driven Transformer-8-51229.68474.57
SNNSpike-driven Transformer-10-51236.01474.66
SNNSpike-driven Transformer-8-76866.34476.32 / 77.07*
STAtten + [53]SNNSpike-driven Transformer-8-51229.68476.18 / 76.56*
SNNSpike-driven Transformer-8-76866.34478.11 / 78.39*
Spike-driven Transformer-V2 [52]SNNSpike-driven Transformer-V2-8-51255.4479.49† / 79.98*†
STAtten + [52]SNNSpike-driven Transformer-V2-8-51255.4479.85 / 80.67*
+ +Table 5. Performance comparison between our methods and previous works on CIFAR10-DVS and N-Caltech101 datasets. In the architecture column, $L - D$ represents $L$ number of encoder blocks and $D$ hidden dimensions. All baseline architectures of spike-based transformer and corresponding STAtten implementations are trained from scratch for fair comparison. + +
MethodArchitectureTimestepCIFAR10-DVS (%)N-Caltech101 (%)
tdBN [60]ResNet191067.80-
PLIF [14]ConvNet2074.80-
Dspike [31]ResNet181075.40-
DSR [36]ResNet182077.27-
SEW-ResNet [13]ConvNet2074.80-
TT-SNN [27]ResNet346-77.80
NDA [32]VGG111079.678.2
Spikformer [64]Spikformer-2-2561680.9-
Spike-driven Transformer [53]Spike-driven Transformer-2-2561680.081.80
STAtten + [53]Spike-driven Transformer-2-2561681.183.15
SpikingReformer [42]SpikingReformer-4-3841678.8081.29
STAtten + [42]SpikingReformer-4-3841680.6081.95
QKFormer [62]QKFormer-4-3841682.9083.58
STAtten + [62]SpikingReformer-4-3841683.9084.25
+ +[52] further improves performance to $79.85\%$ and $80.67\%$ at respective resolutions. It is noteworthy that Spike-driven Transformer [53] achieves comparable or better performance than recent Spikformer models while using only 29.68M parameters. Furthermore, STAtten with Spik-driven Transformer-V2 achieves even higher accuracy than SpikingReformer and Spikformer while using around $20\%$ fewer parameters. + +Neuromorphic Datasets. We evaluate our approach on two neuromorphic vision datasets: CIFAR10-DVS [29] and N-Caltech101 [37]. CIFAR10-DVS contains 10,000 sam + +plies (9,000 training, 1,000 testing) captured using a Dynamic Vision Sensor (DVS) from static CIFAR10 images. N-Caltech101, derived from the Caltech101 dataset [28], comprises 8,831 DVS image samples. Following standard protocols, we resize CIFAR10-DVS samples to $64 \times 64$ and apply the augmentation strategy from [53], while N-Caltech101 samples are resized to $64 \times 64$ with NDA [32] augmentation. As shown in Table 5, integrating STAtten consistently improves performance across different architectures. With Spike-driven Transformer as the backbone, our approach achieves $81.1\%$ accuracy on CIFAR10-DVS + +Table 6. Comparison of memory, energy, and accuracy on ImageNet across different transformer architectures. The precision format weight/activation bit-width and the memory include the size of model parameters and activations. We show the memory consumption for both 32-bit and 8-bit weight spiking models. SDT represents Spike-driven Transformer. + +
MethodPrecisionTimestepMemory (MB)Energy (mJ)
ViT-B/16 [11]32/321351.8254.84
Liu (ViT-B) [34]8/8187.96110.8
Spikformer [64]32(8)/14285.99 (86.48)21.48
SpikingReformer [42]32(8)/14273.57 (85.4)8.76
QKFormer [62]32(8)/14280.28 (84.99)38.91
SDT [53]32(8)/14283.97 (87.42)12.42
STAtten + [53]32(8)/14283.97 (87.42)12.36
SDT-V2 [52]32(8)/14250.22 (84.02)52.40
STAtten + [53]32(8)/14250.22 (84.02)52.38
+ +![](images/94e18cb5eafbd04650560b31e5b958270b53d8b65a734d2896aeeb0d1c8b8ba1.jpg) +(a) + +![](images/58e98b294bc28df5f6974027d45af64fcbfd288272783b8d1e7aff1c76f01f9b.jpg) +(b) +Figure 5. Accuracy comparison with respect to the number of parameters on (a) CIFAR100 and (b) Sequential CIFAR100 datasets. Spike-driven Transformer [53] is used as the baseline of spatial-only architecture. + +and $83.15\%$ on N-Caltech101, improving the baseline by $1.1\%$ and $1.35\%$ respectively. The enhancement extends to other architectures as well, with SpikingReformer achieving $80.60\%$ on CIFAR10-DVS and $81.95\%$ on N-Caltech101, and QKFormer reaching $83.90\%$ and $84.25\%$ respectively, which is the state-of-the-art result. + +# 5.3. Memory and Energy Analysis + +We analyze memory and energy consumption during ImageNet inference. As shown in Table 6, traditional transformer, ViT-B/16, with 32-bit precision requires high memory (351.8 MB) and energy consumption (254.84 mJ). Even with 8-bit precision, Liu [34] still consumes $110.8\mathrm{mJ}$ of energy. In contrast, spike-based transformers with 32-bit weights and binary activations demonstrate significant energy efficiency. Notably, STAtten integrated with SDT and SDT-V2 achieves similar energy consumption to the baseline architectures, showing that our spatial-temporal processing can be realized without additional energy overhead. Correspondingly, STAtten attains comparable performance with fewer parameters, as detailed in the next section. The methodology for energy calculations is provided in the Supplementary Material. + +# 5.4. Model Capacity + +To further evaluate the capacity of our STAtten in spike-based transformers, we compare the accuracy based on the number of trainable parameters through ablation studies. Fig. 5 shows the performance scaling across different model sizes on both (a) standard CIFAR100 and (b) sequential CIFAR100 datasets. Spike-driven Transformer [53] is used as the baseline of spatial-only architecture. As shown in Fig. 5(a), on CIFAR100, STAtten demonstrates strong performance across various model sizes ranging from 2.56M to 22.97M parameters. The accuracy improvement ( $\sim$ 0.5-1.0%) is consistently maintained across different architectural scales. For Sequential CIFAR100 (Fig. 5(b)), where temporal dependencies are more crucial, STAtten shows particularly robust performance improvement ( $\sim$ 3-5%) even with smaller architectures. + +# 5.5. Limitation + +The hardware implementation of STAtten faces challenges despite its improved accuracy. Full spatial-temporal attention without block partitioning is impractical for neuromorphic deployment, as it requires complete temporal information. Our block-wise approach partially addresses this by processing temporal information in steps, where each step corresponds to one block size. However, even with this optimization, deployment remains challenging on traditional neuromorphic chips like TrueNorth [1] and Loihi [6] that process information step-by-step. A potential solution lies in layer-by-layer neuromorphic chips [22, 54, 56, 63] that support layer-wise processing across timesteps. This architectural shift provides an efficient implementation path for STAtten's block-wise processing. Future integration of parallel LIF neurons [15] could further accelerate computation in layer-by-layer architectures. + +# 6. Conclusion + +This paper proposes a block-wise spatial-temporal attention mechanism, STAtten, that enhances spike-based transformers. Through block-wise processing and leveraging non-softmax properties, STAtten captures spatial-temporal information without additional overhead. STAtten can be integrated into existing architectures while preserving their energy efficiency, consistently improving performance across different backbones. Experimental results across both static and neuromorphic datasets validate the effectiveness of our approach. + +# Acknowledgment + +This work was supported in part by CoCoSys, a JUMP2.0 center sponsored by DARPA and SRC, the National Science Foundation (CAREER Award, Grant #2312366, Grant #2318152), the DARPA Young Faculty Award and the DoEMMICC center SEA-CROGS (Award #DE-SC0023198). + +# References + +[1] Filipp Akopyan, Jun Sawada, Andrew Cassidy, Rodrigo Alvarez-Icaza, John Arthur, Paul Merolla, Nabil Imam, Yutaka Nakamura, Pallab Datta, Gi-Joon Nam, et al. Truenorth: Design and tool flow of a 65 mw 1 million neuron programmable neurosynaptic chip. IEEE transactions on computer-aided design of integrated circuits and systems, 34 (10):1537-1557, 2015. 1, 8 +[2] Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lucic, and Cordelia Schmid. Vivit: A video vision transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6836-6846, 2021. 1, 4, 6 +[3] Anthony N Burkitt. A review of the integrate-and-fire neuron model: I. homogeneous synaptic input. Biological cybernetics, 95:1-19, 2006. 2 +[4] Lili Chen, Kevin Lu, Aravind Rajeswaran, Kimin Lee, Aditya Grover, Misha Laskin, Pieter Abbeel, Aravind Srinivas, and Igor Mordatch. Decision transformer: Reinforcement learning via sequence modeling. Advances in neural information processing systems, 34:15084-15097, 2021. 1 +[5] Sayeed Shafayet Chowdhury, Isha Garg, and Kaushik Roy. Spatio-temporal pruning and quantization for low-latency spiking neural networks. In 2021 International Joint Conference on Neural Networks (IJCNN), pages 1-9. IEEE, 2021. 2 +[6] Mike Davies, Narayan Srinivasa, Tsung-Han Lin, Gautham Chinya, Yongqiang Cao, Sri Harsha Choday, Georgios Dimou, Prasad Joshi, Nabil Imam, Shweta Jain, et al. Loihi: A neuromorphic manycore processor with on-chip learning. IEEE Micro, 38(1):82-99, 2018. 1, 8 +[7] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 6 +[8] Shikuang Deng and Shi Gu. Optimal conversion of conventional artificial neural networks to spiking neural networks. arXiv preprint arXiv:2103.00476, 2021. 2 +[9] Shikuang Deng, Yuhang Li, Shanghang Zhang, and Shi Gu. Temporal efficient training of spiking neural network via gradient re-weighting. arXiv preprint arXiv:2202.11946, 2022. 7 +[10] Peter U Diehl et al. Fast-classifying, high-accuracy spiking deep networks through weight and threshold balancing. In 2015 International joint conference on neural networks (IJCNN), pages 1–8. IEEE, 2015. 2 +[11] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 1, 3, 4, 5, 7, 8 +[12] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. International journal of computer vision, 88:303-338, 2010. 3 + +[13] Wei Fang, Zhaofei Yu, Yanqi Chen, Tiejun Huang, Timothee Masquelier, and Yonghong Tian. Deep residual learning in spiking neural networks. Advances in Neural Information Processing Systems, 34:21056-21069, 2021. 5, 7 +[14] Wei Fang, Zhaofei Yu, Yanqi Chen, Timothee Masquelier, Tiejun Huang, and Yonghong Tian. Incorporating learnable membrane time constant to enhance learning of spiking neural networks. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2661-2671, 2021. 2, 5, 6, 7 +[15] Wei Fang, Zhaofei Yu, Zhaokun Zhou, Ding Chen, Yanqi Chen, Zhengyu Ma, Timothée Masquelier, and Yonghong Tian. Parallel spiking neurons with high efficiency and ability to learn long-term dependencies. Advances in Neural Information Processing Systems, 36, 2024. 8 +[16] Marylou Gabrie, Andre Manoel, Clément Luneau, Nicolas Macris, Florent Krzakala, Lenka Zdeborova, et al. Entropy and mutual information in models of deep neural networks. Advances in neural information processing systems, 31, 2018. 3 +[17] Bing Han and Kaushik Roy. Deep spiking neural network: Energy efficiency through time based coding. In European Conference on Computer Vision, pages 388-404. Springer, 2020. 2 +[18] Mark Horowitz. 1.1 computing's energy problem (and what we can do about it). In 2014 IEEE international solid-state circuits conference digest of technical papers (ISSCC), pages 10-14. IEEE, 2014. 6, 1 +[19] Yangfan Hu, Huajin Tang, and Gang Pan. Spiking deep residual networks. IEEE Transactions on Neural Networks and Learning Systems, 34(8):5200-5205, 2021. 7 +[20] Yifan Hu, Lei Deng, Yujie Wu, Man Yao, and Guoqi Li. Advancing spiking neural networks toward deep residual learning. IEEE Transactions on Neural Networks and Learning Systems, 2024. 1, 2, 5, 7 +[21] Seijoon Kim, Seongsik Park, Byunggook Na, and Sungroh Yoon. Spiking-yolo: spiking neural network for energy-efficient object detection. In Proceedings of the AAAI conference on artificial intelligence, pages 11270-11277, 2020. 3 +[22] Sangyeob Kim, Soyeon Kim, Seongyon Hong, Sangjin Kim, Donghyeon Han, and Hoi-Jun Yoo. C-dnn: A 24.5-85.8 tops/w complementary-deep-neural-network processor with heterogeneous cnn/snn core architecture and forward-gradient-based sparsity generation. In 2023 IEEE International Solid-State Circuits Conference (ISSCC), pages 334-336. IEEE, 2023. 8 +[23] Youngeun Kim, Joshua Chough, and Priyadarshini Panda. Beyond classification: Directly training spiking neural networks for semantic segmentation. Neuromorphic Computing and Engineering, 2(4):044015, 2022. 2 +[24] Youngeun Kim, Yuhang Li, Hyoungseob Park, Yeshwanth Venkatesha, and Priyadarshini Panda. Neural architecture search for spiking neural networks. In European Conference on Computer Vision, pages 36-56. Springer, 2022. 6 +[25] Youngeun Kim, Yuhang Li, Hyoungseob Park, Yeshwanth Venkatesha, Anna Hambitzer, and Priyadarshini Panda. Ex + +ploring temporal information dynamics in spiking neural networks. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 8308-8316, 2023. 3 +[26] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 6 +[27] Donghyun Lee, Ruokai Yin, Youngeun Kim, Abhishek Moitra, Yuhang Li, and Priyadarshini Panda. Tt-snn: tensor train decomposition for efficient spiking neural network training. In 2024 Design, Automation & Test in Europe Conference & Exhibition (DATE), pages 1-6. IEEE, 2024. 7 +[28] FF Li, M Andreeto, M Ranzato, and P Perona. Caltech 101 (1.0)[data set]. caltechdata, 2022. 7 +[29] Hongmin Li, Hanchao Liu, Xiangyang Ji, Guoqi Li, and Luping Shi. Cifar10-dvs: an event-stream dataset for object classification. Frontiers in neuroscience, 11:244131, 2017. 6, 7 +[30] Yuhang Li, Shikuang Deng, Xin Dong, Ruihao Gong, and Shi Gu. A free lunch from ann: Towards efficient, accurate spiking neural networks calibration. In International conference on machine learning, pages 6316-6325. PMLR, 2021. 2 +[31] Yuhang Li, Yufei Guo, Shanghang Zhang, Shikuang Deng, Yongqing Hai, and Shi Gu. Differentiable spike: Rethinking gradient-descent for training spiking neural networks. Advances in Neural Information Processing Systems, 34: 23426-23439, 2021. 2, 6, 7 +[32] Yuhang Li, Youngeun Kim, Hyoungseob Park, Tamar Geller, and Priyadarshini Panda. Neuromorphic data augmentation for training spiking neural networks. In European Conference on Computer Vision, pages 631-649. Springer, 2022. 7, 1 +[33] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 7 +[34] Zhenhua Liu, Yunhe Wang, Kai Han, Wei Zhang, Siwei Ma, and Wen Gao. Post-training quantization for vision transformer. Advances in Neural Information Processing Systems, 34:28092-28103, 2021. 8 +[35] Wolfgang Maass. Networks of spiking neurons: the third generation of neural network models. Neural networks, 10 (9):1659-1671, 1997. 1 +[36] Qingyan Meng, Mingqing Xiao, Shen Yan, Yisen Wang, Zhouchen Lin, and Zhi-Quan Luo. Training high-performance low-latency spiking neural networks by differentiation on spike representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12444–12453, 2022. 6, 7 +[37] Garrick Orchard, Ajinkya Jayawant, Gregory K Cohen, and Nitish Thakor. Converting static image datasets to spiking neuromorphic datasets using saccades. Frontiers in neuroscience, 9:159859, 2015. 6, 7 +[38] Jing Pei, Lei Deng, Sen Song, Mingguo Zhao, Youhui Zhang, Shuang Wu, Guanrui Wang, Zhe Zou, Zhenzhi Wu, Wei He, et al. Towards artificial general intelligence with hybrid tianjic chip architecture. Nature, 572(7767):106-111, 2019. 1 + +[39] Nitin Rathi and Kaushik Roy. Diet-snn: Direct input encoding with leakage and threshold optimization in deep spiking neural networks. arXiv preprint arXiv:2008.03658, 2020. 6 +[40] Kaushik Roy, Akhilesh Jaiswal, and Priyadarshini Panda. Towards spike-based machine intelligence with neuromorphic computing. Nature, 575(7784):607-617, 2019. 1 +[41] Abhronil Sengupta, Yuting Ye, Robert Wang, Chiao Liu, and Kaushik Roy. Going deeper in spiking neural networks: Vgg and residual architectures. Frontiers in neuroscience, 13:95, 2019. 1 +[42] Xinyu Shi, Zecheng Hao, and Zhaofei Yu. Spikingresformer: Bridging resnet and vision transformer in spiking neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5610-5619, 2024. 1, 2, 4, 5, 6, 7, 8, 3 +[43] Sumit B Shrestha and Garrick Orchard. Slayer: Spike layer error reassignment in time. Advances in neural information processing systems, 31, 2018. 2 +[44] Qiaoyi Su, Yuhong Chou, Yifan Hu, Jianing Li, Shijie Mei, Ziyang Zhang, and Guoqi Li. Deep directly-trained spiking neural networks for object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6555-6565, 2023. 2, 3 +[45] Zhengkun Tian, Jiangyan Yi, Jianhua Tao, Ye Bai, Shuai Zhang, and Zhengqi Wen. Spike-triggered nonautoregressive transformer for end-to-end speech recognition. arXiv preprint arXiv:2005.07903, 2020. 1 +[46] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In International conference on machine learning, pages 10347-10357. PMLR, 2021. 7 +[47] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 1, 3 +[48] Qingyu Wang, Tielin Zhang, Minglun Han, Yi Wang, Duzhen Zhang, and Bo Xu. Complex dynamic neurons improved spiking transformer network for efficient automatic speech recognition. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 102-109, 2023. 1 +[49] Yujie Wu, Lei Deng, Guoqi Li, Jun Zhu, Yuan Xie, and Luping Shi. Direct training for spiking neural networks: Faster, larger, better. In Proceedings of the AAAI conference on artificial intelligence, pages 1311-1318, 2019. 1, 2, 6 +[50] Shiting Xiao, Yuhang Li, Youngeun Kim, Donghyun Lee, and Priyadarshini Panda. Respike: residual frames-based hybrid spiking neural networks for efficient action recognition. Neuromorphic Computing and Engineering, 2025. 1 +[51] Man Yao, Guangshe Zhao, Hengyu Zhang, Yifan Hu, Lei Deng, Yonghong Tian, Bo Xu, and Guoqi Li. Attention spiking neural networks. IEEE transactions on pattern analysis and machine intelligence, 2023. 7 +[52] Man Yao, Jiakui Hu, Tianxiang Hu, Yifan Xu, Zhaokun Zhou, Yonghong Tian, Bo Xu, and Guoqi Li. Spike-driven transformer v2: Meta spiking neural network architecture inspiring the design of next-generation neuromorphic chips. arXiv preprint arXiv:2404.03663, 2024. 1, 2, 4, 5, 7, 8 + +[53] Man Yao, Jiakui Hu, Zhaokun Zhou, Li Yuan, Yonghong Tian, Bo Xu, and Guoqi Li. Spike-driven transformer. Advances in Neural Information Processing Systems, 36, 2024. 1, 2, 3, 4, 5, 6, 7, 8 +[54] Ruokai Yin, Abhishek Moitra, Abhiroop Bhattacharjee, Youngeun Kim, and Priyadarshini Panda. Sata: Sparsity-aware training accelerator for spiking neural networks. IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems, 42(6):1926-1938, 2022. 8 +[55] Ruokai Yin, Youngeun Kim, Yuhang Li, Abhishek Moitra, Nitin Satpute, Anna Hambitzer, and Priyadarshini Panda. Workload-balanced pruning for sparse spiking neural networks. IEEE Transactions on Emerging Topics in Computational Intelligence, 2024. 2 +[56] Ruokai Yin, Youngeun Kim, Di Wu, and Priyadarshini Panda. Loas: Fully temporal-parallel dataflow for dualsparse spiking neural networks. In 2024 57th IEEE/ACM International Symposium on Microarchitecture (MICRO), pages 1107–1121. IEEE, 2024. 8 +[57] Ruokai Yin, Yuhang Li, Abhishek Moitra, and Priyadarshini Panda. Mint: Multiplier-less integer quantization for energy efficient spiking neural networks. In 2024 29th Asia and South Pacific Design Automation Conference (ASP-DAC), pages 830-835. IEEE, 2024. 2 +[58] Jiqing Zhang, Bo Dong, Haiwei Zhang, Jianchuan Ding, Felix Heide, Baocai Yin, and Xin Yang. Spiking transformers for event-based single object tracking. In Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition, pages 8801-8810, 2022. 1 +[59] Jiyuan Zhang, Lulu Tang, Zhaofei Yu, Jiwen Lu, and Tiejun Huang. Spike transformer: Monocular depth estimation for spiking camera. In European Conference on Computer Vision, pages 34-52. Springer, 2022. 1 +[60] Hanle Zheng, Yujie Wu, Lei Deng, Yifan Hu, and Guoqi Li. Going deeper with directly-trained larger spiking neural networks. In Proceedings of the AAAI conference on artificial intelligence, pages 11062-11070, 2021. 1, 6, 7 +[61] Chenlin Zhou, Liutao Yu, Zhaokun Zhou, Han Zhang, Zhengyu Ma, Huihui Zhou, and Yonghong Tian. Spikingformer: Spike-driven residual learning for transformer-based spiking neural network. arXiv preprint arXiv:2304.11954, 2023. 5 +[62] Chenlin Zhou, Han Zhang, Zhaokun Zhou, Liutao Yu, Liwei Huang, Xiaopeng Fan, Li Yuan, Zhengyu Ma, Huihui Zhou, and Yonghong Tian. Qkformer: Hierarchical spiking transformer using qk attention. arXiv preprint arXiv:2403.16552, 2024. 1, 2, 4, 5, 6, 7, 8 +[63] PJ Zhou, Q Yu, M Chen, YC Wang, LW Meng, Y Zuo, N Ning, Y Liu, SG Hu, and GC Qiao. A 0.96 pj/sop, 30.23 k-neuron/mm 2 heterogeneous neuromorphic chip with fullerene-like interconnection topology for edge-ai computing. In 2024 IEEE International Symposium on Circuits and Systems (ISCAS), pages 1–5. IEEE, 2024. 8 +[64] Zhaokun Zhou, Yuesheng Zhu, Chao He, Yaowei Wang, Shuicheng Yan, Yonghong Tian, and Li Yuan. Spikformer: When spiking neural network meets transformer. arXiv preprint arXiv:2209.15425, 2022. 1, 2, 3, 4, 5, 6, 7, 8 \ No newline at end of file diff --git a/2025/Spiking Transformer with Spatial-Temporal Attention/images.zip b/2025/Spiking Transformer with Spatial-Temporal Attention/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..b4b919fa9ed33cd8535661884e165ff178c6bed0 --- /dev/null +++ b/2025/Spiking Transformer with Spatial-Temporal Attention/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d578d2f8ab3c343ace93cf58ba29f254b1e8687bf2fdc14d6616d78211a7f033 +size 643396 diff --git a/2025/Spiking Transformer with Spatial-Temporal Attention/layout.json b/2025/Spiking Transformer with Spatial-Temporal Attention/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..42991aa6b6c730a7a7290838f34cd654bc0f9707 --- /dev/null +++ b/2025/Spiking Transformer with Spatial-Temporal Attention/layout.json @@ -0,0 +1,9560 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 137, + 103, + 473, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 103, + 473, + 121 + ], + "spans": [ + { + "bbox": [ + 137, + 103, + 473, + 121 + ], + "type": "text", + "content": "Spiking Transformer with Spatial-Temporal Attention" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 115, + 143, + 494, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 143, + 494, + 172 + ], + "spans": [ + { + "bbox": [ + 115, + 143, + 494, + 172 + ], + "type": "text", + "content": "Donghyun Lee, Yuhang Li, Youngeun Kim, Shiting Xiao, Priyadarshini Panda Department of Electrical Engineering, Yale University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 108, + 174, + 502, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 174, + 502, + 186 + ], + "spans": [ + { + "bbox": [ + 108, + 174, + 502, + 186 + ], + "type": "text", + "content": "{donghyun.lee, yuhang.li, youngeun.kim, ginny.xiao, priya.panda}@yale.edu" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 152, + 213, + 200, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 213, + 200, + 225 + ], + "spans": [ + { + "bbox": [ + 152, + 213, + 200, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 238, + 296, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 238, + 296, + 538 + ], + "spans": [ + { + "bbox": [ + 54, + 238, + 296, + 538 + ], + "type": "text", + "content": "Spike-based Transformer presents a compelling and energy-efficient alternative to traditional Artificial Neural Network (ANN)-based Transformers, achieving impressive results through sparse binary computations. However, existing spike-based transformers predominantly focus on spatial attention while neglecting crucial temporal dependencies inherent in spike-based processing, leading to suboptimal feature representation and limited performance. To address this limitation, we propose Spiking Transformer with Spatial-Temporal Attention (STAtten), a simple and straightforward architecture that efficiently integrates both spatial and temporal information in the self-attention mechanism. STAtten introduces a block-wise computation strategy that processes information in spatial-temporal chunks, enabling comprehensive feature capture while maintaining the same computational complexity as previous spatial-only approaches. Our method can be seamlessly integrated into existing spike-based transformers without architectural overhaul. Extensive experiments demonstrate that STAtten significantly improves the performance of existing spike-based transformers across both static and neuromorphic datasets, including CIFAR10/100, ImageNet, CIFAR10-DVS, and N-Caltech101. The code is available at https://github.com/Intelligent-Computing-Lab-Yale/STAtten." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 561, + 135, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 561, + 135, + 574 + ], + "spans": [ + { + "bbox": [ + 55, + 561, + 135, + 574 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "content": "Spiking Neural Networks (SNNs) are emerging as a promising alternative to conventional Artificial Neural Networks (ANNs) [35, 40] due to their bio-inspired, energy-efficient computing paradigm. By leveraging sparse binary spike-based computations, SNNs can achieve significant energy savings while enabling deployment across various neuromorphic computing platforms such as TrueNorth [1], Loihi [6], and Tianjic [38]. However, the challenge lies in effectively processing and learning from these discrete, time-dependent spike patterns while maintaining computational efficiency and accuracy. Traditional SNNs have primarily" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 319, + 215, + 554, + 324 + ], + "blocks": [ + { + "bbox": [ + 319, + 215, + 554, + 324 + ], + "lines": [ + { + "bbox": [ + 319, + 215, + 554, + 324 + ], + "spans": [ + { + "bbox": [ + 319, + 215, + 554, + 324 + ], + "type": "image", + "image_path": "c324ea93315e4534ed615b17fe8a2b2544ff9db8591f0479fb70d24ea953bc25.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 338, + 555, + 372 + ], + "lines": [ + { + "bbox": [ + 313, + 338, + 555, + 372 + ], + "spans": [ + { + "bbox": [ + 313, + 338, + 555, + 372 + ], + "type": "text", + "content": "Figure 1. Heatmaps of spatial-only attention versus our STAten on sequential CIFAR100 dataset. Input images are divided column-wise, where each column corresponds to one timestep." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 391, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 391, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 391, + 555, + 715 + ], + "type": "text", + "content": "relied on convolution-based architectures adapted from successful ANN models like VGGNet [41, 49] and ResNet [20, 60]. While these architectures benefit from computational efficiency, they often struggle with binary spike activations, leading to information loss and degraded performance. This limitation has driven researchers to explore alternative architectures that can better handle sparse, temporal data while maintaining the energy benefits of spike-based computation. Transformer architectures, with their remarkable success across various domains [2, 4, 11, 47], have emerged as a promising direction for enhancing SNN capabilities. Recent works have introduced spike-formed transformers that adapt the transformer's self-attention mechanism to the spike-based domain for diverse tasks, such as, object tracking [58], computer vision [42, 50, 52, 53, 62, 64], depth estimation [59], and speech recognition [45, 48]. These architectures, such as Spikformer [64] and Spike-driven Transformer [53], implement innovative binary representations for query (Q), key (K), and value (V) computations while eliminating the computationally expensive Softmax function. This binary attention design preserves the essential characteristics of spike-based computation while enabling more efficient information processing compared to conventional transformer architectures. However, current spike-formed transformers predominantly rely on spatial-only attention, overlooking the inherently dynamic and temporal nature of spike events. To understand these differ" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13948" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 178 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 178 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 178 + ], + "type": "text", + "content": "ent attention patterns, we compare spatial-only and spatial-temporal attention on sequential CIFAR100, as shown in Fig. 1. The visualization reveals that spatial-only attention focuses solely on vertical (spatial) relationships within each timestep, missing crucial object features that evolve over time. In contrast, our spatial-temporal attention captures both vertical (spatial) and horizontal (temporal) relationships, enabling comprehensive feature representation of sequential inputs." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 180, + 295, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 180, + 295, + 357 + ], + "spans": [ + { + "bbox": [ + 56, + 180, + 295, + 357 + ], + "type": "text", + "content": "In this work, we introduce Spiking Transformer with Spatial-Temporal Attention (STAtten), which integrates both spatial and temporal information within the self-attention mechanism. Through analysis of memory requirements and neuronal activity patterns, we design a blockwise computation strategy with local temporal correlations. By eliminating Softmax operation, we maintain the same computational complexity " + }, + { + "bbox": [ + 56, + 180, + 295, + 357 + ], + "type": "inline_equation", + "content": "(O(TND^{2}))" + }, + { + "bbox": [ + 56, + 180, + 295, + 357 + ], + "type": "text", + "content": " as existing spike-based transformers [42, 52, 62, 64]. Furthermore, our approach can be flexibly integrated into existing architectures without structural modifications, enhancing their representational capacity. Experiments on both static and dynamic datasets demonstrate consistent performance improvements over spatial-only attention architectures. The main contributions of our work are as follows:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 360, + 294, + 550 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 55, + 360, + 294, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 360, + 294, + 407 + ], + "spans": [ + { + "bbox": [ + 55, + 360, + 294, + 407 + ], + "type": "text", + "content": "- We identify limitations in current spiking transformers that rely on spatial-only attention through empirical analysis, demonstrating the importance of capturing temporal dependencies inherent in spike-based processing." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 407, + 294, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 407, + 294, + 455 + ], + "spans": [ + { + "bbox": [ + 56, + 407, + 294, + 455 + ], + "type": "text", + "content": "- We propose STAtten, a block-wise spatial-temporal attention mechanism for spike-based transformers that maintains the original computational complexity " + }, + { + "bbox": [ + 56, + 407, + 294, + 455 + ], + "type": "inline_equation", + "content": "(O(TND^{2}))" + }, + { + "bbox": [ + 56, + 407, + 294, + 455 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 455, + 294, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 455, + 294, + 491 + ], + "spans": [ + { + "bbox": [ + 56, + 455, + 294, + 491 + ], + "type": "text", + "content": "- We introduce a flexible plug-and-play design that enables STAtten to be integrated into existing spike-based transformers without compromising their core architecture." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 491, + 294, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 491, + 294, + 550 + ], + "spans": [ + { + "bbox": [ + 56, + 491, + 294, + 550 + ], + "type": "text", + "content": "- Through extensive experiments across both static (CIFAR10/100, ImageNet) and neuromorphic (CIFAR10-DVS, N-Caltech101) datasets, we demonstrate that STatten consistently improves performance across different backbone architectures." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 562, + 147, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 562, + 147, + 574 + ], + "spans": [ + { + "bbox": [ + 55, + 562, + 147, + 574 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "text", + "content": "Training Strategies for SNNs. Significant progress has been made in developing effective training strategies for SNNs, primarily following two approaches: ANN-to-SNN conversion and direct training. The conversion approach transforms pre-trained ANNs into SNNs, leveraging established ANN architectures to enhance SNN performance. Several studies [8, 10, 17, 30] have formalized this conversion process and demonstrated its effectiveness in transferring ANN knowledge to the spike domain. Direct training through backpropagation, while fundamental to ANNs, presents unique challenges in SNNs due to the" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 553, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 191 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 191 + ], + "type": "text", + "content": "non-differentiable nature of spike events. The introduction of surrogate gradients has been crucial in enabling effective backpropagation-based training in SNNs, as demonstrated in pioneering works [5, 43, 49]. Notably, direct training with surrogate gradients has shown superior performance compared to conversion methods across various tasks, including image classification [14, 20, 31, 55, 57], semantic segmentation [23], and object detection [44], while maintaining the energy efficiency inherent to spike-based computation." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 193, + 555, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 193, + 555, + 526 + ], + "spans": [ + { + "bbox": [ + 313, + 193, + 555, + 526 + ], + "type": "text", + "content": "Spiking Transformer. Despite the energy efficiency of convolution-based SNNs, their performance often lags behind ANNs due to information loss from sparse binary activations. To address this limitation, researchers have adapted transformer architectures to the SNN domain, leveraging their powerful self-attention mechanisms and ability to capture long-range dependencies. This adaptation has led to several innovative spike-formed transformer architectures. Spikformer [64] introduced the first SNN-based transformer architecture, establishing foundational principles for spikeformed self-attention mechanisms. Their key innovation lies in eliminating the Softmax function from self-attention computations, justified by two key observations: (1) the binary nature of spike-based Q, K, and V values, and (2) the redundancy of Softmax scaling for binarized operations. Building upon this foundation, Spike-driven Transformer [53] enhanced the architecture with a novel self-attention block incorporating masking and sparse addition to optimize power efficiency. Additionally, inspired by MS-ResNet [20], Spike-driven Transformer reimagined residual connections to propagate membrane potential rather than spikes. Recent works have further extended these concepts, leading to architectures like SpikingReformer [42], QK-Former [62], and Spike-driven Transformer-V2 [52], each introducing unique optimizations for spike-based processing. However, existing spiking transformer architectures primarily focus on spatial attention, neglecting the temporal dynamics inherent in spike-based computation." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 536, + 391, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 536, + 391, + 550 + ], + "spans": [ + { + "bbox": [ + 314, + 536, + 391, + 550 + ], + "type": "text", + "content": "3. Preliminary" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 556, + 553, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 556, + 553, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 556, + 553, + 628 + ], + "type": "text", + "content": "Leaky Integrate-and-Fire Neuron. As a fundamental building block of SNNs, the Leaky Integrate-and-Fire (LIF) neuron [3] has emerged as an important component for energy-efficient computation. The LIF neuron is a nonlinear activation function that determines whether neurons fire spikes as follows:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 361, + 633, + 553, + 649 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 633, + 553, + 649 + ], + "spans": [ + { + "bbox": [ + 361, + 633, + 553, + 649 + ], + "type": "interline_equation", + "content": "\\mathbf {u} [ t + 1 ] ^ {l} = \\tau \\mathbf {u} [ t ] ^ {l} + \\mathbf {W} ^ {l} f (\\mathbf {u} [ t ] ^ {l - 1}) \\tag {1}", + "image_path": "2f36650904553f065e6cf0519c6e640acda3f863f4cf0a88d9152d897008d112.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 372, + 651, + 553, + 682 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 651, + 553, + 682 + ], + "spans": [ + { + "bbox": [ + 372, + 651, + 553, + 682 + ], + "type": "interline_equation", + "content": "f \\left(\\mathbf {u} [ t ] ^ {l}\\right) = \\left\\{ \\begin{array}{l l} 1 & \\text {i f} \\mathbf {u} [ t ] ^ {l} > V _ {t h}, \\\\ 0 & \\text {o t h e r w i s e} \\end{array} , \\right. \\tag {2}", + "image_path": "4becb22aa11e34d66c5c4eaa09899fe56a27d2e61027626b6e2ae0734e833166.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": "where, " + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{u}[t]^l" + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": " is the membrane potential in " + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": "-th layer at timestep " + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\tau \\in (0,1]" + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": " is the leaky factor for membrane po" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "13949" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 66, + 71, + 172, + 164 + ], + "blocks": [ + { + "bbox": [ + 66, + 71, + 172, + 164 + ], + "lines": [ + { + "bbox": [ + 66, + 71, + 172, + 164 + ], + "spans": [ + { + "bbox": [ + 66, + 71, + 172, + 164 + ], + "type": "image", + "image_path": "57a643a56b805f3a858c0f0fca5f6a00b73194fbb45b0993f9a3360bbee9eb6c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 171, + 296, + 205 + ], + "lines": [ + { + "bbox": [ + 55, + 171, + 296, + 205 + ], + "spans": [ + { + "bbox": [ + 55, + 171, + 296, + 205 + ], + "type": "text", + "content": "Figure 2. Comparison between different self-attention with the CIFAR100 dataset. Analysis of entropy and accuracy for Temporal-only (T), Spatial-only (S), and Spatial-temporal (ST)." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 173, + 71, + 281, + 164 + ], + "blocks": [ + { + "bbox": [ + 173, + 71, + 281, + 164 + ], + "lines": [ + { + "bbox": [ + 173, + 71, + 281, + 164 + ], + "spans": [ + { + "bbox": [ + 173, + 71, + 281, + 164 + ], + "type": "image", + "image_path": "5a8ef6d7cfd081febdb07f4e9dbe44f2a3e8984722cca88f84f02537a48cde77.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 165, + 555, + 220 + ], + "lines": [ + { + "bbox": [ + 313, + 165, + 555, + 220 + ], + "spans": [ + { + "bbox": [ + 313, + 165, + 555, + 220 + ], + "type": "text", + "content": "Figure 3. (a) Maximum batch size of block-wise STAtten and full spatial-temporal attention (without block partitioning) for running on A5000 GPU with 24GB VRAM memory. (b) Average number of active neurons after QKV computation at different timestep combinations, where [t] indicates timestep index." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 210, + 296, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 210, + 296, + 258 + ], + "spans": [ + { + "bbox": [ + 55, + 210, + 296, + 258 + ], + "type": "text", + "content": "tential leakage, " + }, + { + "bbox": [ + 55, + 210, + 296, + 258 + ], + "type": "inline_equation", + "content": "\\mathbf{W}^l" + }, + { + "bbox": [ + 55, + 210, + 296, + 258 + ], + "type": "text", + "content": " is the weight of " + }, + { + "bbox": [ + 55, + 210, + 296, + 258 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 55, + 210, + 296, + 258 + ], + "type": "text", + "content": "-th layer, and " + }, + { + "bbox": [ + 55, + 210, + 296, + 258 + ], + "type": "inline_equation", + "content": "f(\\cdot)" + }, + { + "bbox": [ + 55, + 210, + 296, + 258 + ], + "type": "text", + "content": " is the LIF function with firing threshold " + }, + { + "bbox": [ + 55, + 210, + 296, + 258 + ], + "type": "inline_equation", + "content": "V_{th}" + }, + { + "bbox": [ + 55, + 210, + 296, + 258 + ], + "type": "text", + "content": ". Therefore, when the membrane " + }, + { + "bbox": [ + 55, + 210, + 296, + 258 + ], + "type": "inline_equation", + "content": "\\mathbf{u}[t]^l" + }, + { + "bbox": [ + 55, + 210, + 296, + 258 + ], + "type": "text", + "content": " is higher than " + }, + { + "bbox": [ + 55, + 210, + 296, + 258 + ], + "type": "inline_equation", + "content": "V_{th}" + }, + { + "bbox": [ + 55, + 210, + 296, + 258 + ], + "type": "text", + "content": ", the LIF function fires a spike and the membrane potential is reset to 0." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 258, + 296, + 330 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 258, + 296, + 330 + ], + "spans": [ + { + "bbox": [ + 55, + 258, + 296, + 330 + ], + "type": "text", + "content": "Vanilla Self-attention. Transformers [11, 47] have surpassed convolution-based architectures due to their unique self-attention mechanism, which can capture global features through long-range dependency. For a floating-point input tensor " + }, + { + "bbox": [ + 55, + 258, + 296, + 330 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_f\\in \\mathbb{R}^{N\\times D}" + }, + { + "bbox": [ + 55, + 258, + 296, + 330 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 55, + 258, + 296, + 330 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 258, + 296, + 330 + ], + "type": "text", + "content": " tokens, and " + }, + { + "bbox": [ + 55, + 258, + 296, + 330 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 55, + 258, + 296, + 330 + ], + "type": "text", + "content": " features, we formulate the self-attention as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 64, + 338, + 295, + 352 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 338, + 295, + 352 + ], + "spans": [ + { + "bbox": [ + 64, + 338, + 295, + 352 + ], + "type": "interline_equation", + "content": "\\mathbf {Q} _ {f} = \\mathbf {W} _ {\\mathbf {Q}} \\mathbf {X} _ {f}, \\quad \\mathbf {K} _ {f} = \\mathbf {W} _ {f} \\mathbf {X} _ {f}, \\quad \\mathbf {V} _ {f} = \\mathbf {W} _ {\\mathbf {V}} \\mathbf {X} _ {f} \\tag {3}", + "image_path": "473eed06e5299d6686be93b653e1a6d1105624661a69df7f0adb219ae651b398.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 61, + 373, + 295, + 410 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 373, + 295, + 410 + ], + "spans": [ + { + "bbox": [ + 61, + 373, + 295, + 410 + ], + "type": "interline_equation", + "content": "\\operatorname {A t t n} = \\operatorname {S o f t m a x} \\left(\\frac {\\mathbf {Q K} ^ {\\top}}{\\sqrt {D}}\\right) \\mathbf {V}, \\quad \\mathbf {Q} _ {f}, \\mathbf {K} _ {f}, \\mathbf {V} _ {f} \\in \\mathbb {R} ^ {N \\times D}, \\tag {4}", + "image_path": "aa7f98c01340164dfbc297f90fcda23a709bd73a6254fdfe905b0cda2cef500d.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 410, + 296, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 410, + 296, + 458 + ], + "spans": [ + { + "bbox": [ + 55, + 410, + 296, + 458 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 410, + 296, + 458 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_{\\mathbf{Q}}, \\mathbf{W}_{\\mathbf{K}}, \\mathbf{W}_{\\mathbf{V}} \\in \\mathbb{R}^{D \\times D}" + }, + { + "bbox": [ + 55, + 410, + 296, + 458 + ], + "type": "text", + "content": " are linear projections for floating-point " + }, + { + "bbox": [ + 55, + 410, + 296, + 458 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_f, \\mathbf{K}_f, \\mathbf{V}_f" + }, + { + "bbox": [ + 55, + 410, + 296, + 458 + ], + "type": "text", + "content": ", respectively. Note that the computational complexity of attention is " + }, + { + "bbox": [ + 55, + 410, + 296, + 458 + ], + "type": "inline_equation", + "content": "\\mathcal{O}(N^2 D)" + }, + { + "bbox": [ + 55, + 410, + 296, + 458 + ], + "type": "text", + "content": " due to its matrix multiplication operation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 459, + 296, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 459, + 296, + 518 + ], + "spans": [ + { + "bbox": [ + 55, + 459, + 296, + 518 + ], + "type": "text", + "content": "Spike-based Self-attention. The first spike-based transformer work, Spikformer [64], proposes Spiking Self Attention (SSA) with binary " + }, + { + "bbox": [ + 55, + 459, + 296, + 518 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}" + }, + { + "bbox": [ + 55, + 459, + 296, + 518 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 459, + 296, + 518 + ], + "type": "inline_equation", + "content": "\\mathbf{K}" + }, + { + "bbox": [ + 55, + 459, + 296, + 518 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 459, + 296, + 518 + ], + "type": "inline_equation", + "content": "\\mathbf{V}" + }, + { + "bbox": [ + 55, + 459, + 296, + 518 + ], + "type": "text", + "content": ". For binary input tensor " + }, + { + "bbox": [ + 55, + 459, + 296, + 518 + ], + "type": "inline_equation", + "content": "\\mathbf{X} \\in \\mathbb{R}^{T \\times N \\times D}" + }, + { + "bbox": [ + 55, + 459, + 296, + 518 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 55, + 459, + 296, + 518 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 459, + 296, + 518 + ], + "type": "text", + "content": " timestep, we formulate the SSA as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 64, + 523, + 295, + 563 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 523, + 295, + 563 + ], + "spans": [ + { + "bbox": [ + 64, + 523, + 295, + 563 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathbf {Q}, \\mathbf {K}, \\mathbf {V} = \\operatorname {L I F} (\\mathbf {W} _ {\\mathbf {Q}} \\mathbf {X}), \\operatorname {L I F} (\\mathbf {W} _ {\\mathbf {K}} \\mathbf {X}), \\operatorname {L I F} (\\mathbf {W} _ {\\mathbf {K}} \\mathbf {X}), \\\\ \\operatorname {S S A} (\\mathbf {Q K V}) = \\operatorname {L I F} \\left(\\mathbf {Q K} ^ {\\top} \\mathbf {V} \\cdot \\alpha\\right) \\tag {5} \\\\ \\end{array}", + "image_path": "61ce617862dbd6f7cb2cd5f9c54daf82bf6a099fb584919b32e2c24e61178f64.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 563, + 295, + 599 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 563, + 295, + 599 + ], + "spans": [ + { + "bbox": [ + 55, + 563, + 295, + 599 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 563, + 295, + 599 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_{\\mathbf{Q}}, \\mathbf{W}_{\\mathbf{K}}, \\mathbf{W}_{\\mathbf{V}} \\in \\mathbb{R}^{D \\times D}" + }, + { + "bbox": [ + 55, + 563, + 295, + 599 + ], + "type": "text", + "content": " are linear projections for binary " + }, + { + "bbox": [ + 55, + 563, + 295, + 599 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}, \\mathbf{K}, \\mathbf{V}" + }, + { + "bbox": [ + 55, + 563, + 295, + 599 + ], + "type": "text", + "content": ", respectively, and " + }, + { + "bbox": [ + 55, + 563, + 295, + 599 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 563, + 295, + 599 + ], + "type": "text", + "content": " denotes a scaling factor." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 609, + 138, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 609, + 138, + 624 + ], + "spans": [ + { + "bbox": [ + 55, + 609, + 138, + 624 + ], + "type": "text", + "content": "4. Methodology" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 629, + 296, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 296, + 712 + ], + "type": "text", + "content": "In this section, we present our analysis of spike-based transformers and provide the details of our proposed methodology. We first establish the necessity of spatial-temporal attention in spike-based transformers through an information-theoretic analysis based on entropy measurements. Building upon these observations, we introduce our STAtten mechanism." + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 317, + 73, + 443, + 148 + ], + "blocks": [ + { + "bbox": [ + 317, + 73, + 443, + 148 + ], + "lines": [ + { + "bbox": [ + 317, + 73, + 443, + 148 + ], + "spans": [ + { + "bbox": [ + 317, + 73, + 443, + 148 + ], + "type": "image", + "image_path": "07fccc209c2c29094fca7d2e499594afceafd65519edf3fcda38ffabfffd6219.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 389, + 148, + 399, + 158 + ], + "lines": [ + { + "bbox": [ + 389, + 148, + 399, + 158 + ], + "spans": [ + { + "bbox": [ + 389, + 148, + 399, + 158 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 449, + 72, + 553, + 135 + ], + "blocks": [ + { + "bbox": [ + 449, + 72, + 553, + 135 + ], + "lines": [ + { + "bbox": [ + 449, + 72, + 553, + 135 + ], + "spans": [ + { + "bbox": [ + 449, + 72, + 553, + 135 + ], + "type": "image", + "image_path": "012624bc4f9fb21e1dc486e9082ba65e1ad012631e1872a46aca2c29ba3a8920.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 498, + 148, + 509, + 158 + ], + "lines": [ + { + "bbox": [ + 498, + 148, + 509, + 158 + ], + "spans": [ + { + "bbox": [ + 498, + 148, + 509, + 158 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 232, + 533, + 245 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 232, + 533, + 245 + ], + "spans": [ + { + "bbox": [ + 313, + 232, + 533, + 245 + ], + "type": "text", + "content": "4.1. Motivation of Spatial-Temporal Attention" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 251, + 554, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 251, + 554, + 334 + ], + "spans": [ + { + "bbox": [ + 313, + 251, + 554, + 334 + ], + "type": "text", + "content": "The feature maps of spike-based transformers at each timestep carry distinct information, as their activation distributions are dynamically influenced by both the leaky factor and reset mechanisms. This temporal aspect of information flow raises a fundamental question: how do different attention mechanisms effectively capture and utilize this spatiotemporal information?" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 335, + 554, + 381 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 335, + 554, + 381 + ], + "spans": [ + { + "bbox": [ + 313, + 335, + 554, + 381 + ], + "type": "text", + "content": "First, we define the three different types of attention: Spatial, Temporal, and Spatial-temporal attention with timestep " + }, + { + "bbox": [ + 313, + 335, + 554, + 381 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 335, + 554, + 381 + ], + "type": "text", + "content": ", token position " + }, + { + "bbox": [ + 313, + 335, + 554, + 381 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 335, + 554, + 381 + ], + "type": "text", + "content": ", and feature dimension " + }, + { + "bbox": [ + 313, + 335, + 554, + 381 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 313, + 335, + 554, + 381 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 320, + 388, + 553, + 447 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 388, + 553, + 447 + ], + "spans": [ + { + "bbox": [ + 320, + 388, + 553, + 447 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathrm {S} _ {-} \\operatorname {A t t n} [ t,,:,: ] = \\operatorname {L I F} (\\mathbf {Q} [ t,:,: ] \\mathbf {K} ^ {\\top} [ t,:,: ] \\mathbf {V} [ t,:,: ] \\cdot \\alpha), \\\\ \\mathrm {T} _ {\\mathbf {\\Lambda}} \\operatorname {A t t n} [:, n,: ] = \\operatorname {L I F} (\\mathbf {Q} [:, n,: ] \\mathbf {K} ^ {\\top} [:, n,: ] \\mathbf {V} [:, n,: ] \\cdot \\alpha), \\\\ \\operatorname {S T} \\text {A t t n} [:,:, d ] = \\operatorname {L I F} (\\mathbf {Q} [:,:, d ] \\mathbf {K} ^ {\\top} [:,:, d ] \\mathbf {V} [:,:, d ] \\cdot \\alpha), \\tag {6} \\\\ \\end{array}", + "image_path": "19b546933007af272f4b8d687cfeb25d503cdcc60b0724995224d5d3809910b6.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 448, + 554, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 448, + 554, + 567 + ], + "spans": [ + { + "bbox": [ + 313, + 448, + 554, + 567 + ], + "type": "text", + "content": "where S_Attn, T_Attn, and ST_Attn are spatial, temporal, and spatial-temporal attention, and the Q, K, V dimensions of S_Attn, T_Attn, and ST_Attn are " + }, + { + "bbox": [ + 313, + 448, + 554, + 567 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{T\\times N\\times D}" + }, + { + "bbox": [ + 313, + 448, + 554, + 567 + ], + "type": "text", + "content": ". The key distinction lies in their attention computation scope: spatial attention operates on token relationships within individual timesteps, temporal attention focuses on cross-timestep feature dependencies, and spatial-temporal attention integrates both spatial and temporal feature correlations. To quantify the information distribution in each attention mechanism, we measure Shannon entropy [16] as follows:" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 327, + 575, + 555, + 609 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 575, + 555, + 609 + ], + "spans": [ + { + "bbox": [ + 327, + 575, + 555, + 609 + ], + "type": "interline_equation", + "content": "H (\\mathrm {A t t n}) = - \\sum_ {t = 1} ^ {T} \\sum_ {n = 1} ^ {N} \\sum_ {d = 1} ^ {D} \\hat {\\mathbf {p}} [ t, n, d ] \\log \\hat {\\mathbf {p}} [ t, n, d ], \\tag {7}", + "image_path": "23087213c58a64345bcdf0a587e92a9fbdfe4e1e233aa1a8733a8c2280b67391.jpg" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 617, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 555, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 617, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{p}}[t,n,d] = \\text{Softmax}(\\text{Attn}[t,n,d])" + }, + { + "bbox": [ + 313, + 617, + 555, + 714 + ], + "type": "text", + "content": ". Lower entropy values indicate more focused spike patterns, while higher entropy values suggest dispersed patterns [25]. We analyze this relationship using a pretrained Spike-driven Transformer [53] on the CIFAR100 dataset, as shown in Fig. 2. The results reveal an inverse correlation between entropy and accuracy: temporal-only attention shows high entropy (5.81) but lowest accuracy (77.7%), while" + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "13950" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 69, + 95, + 359, + 208 + ], + "blocks": [ + { + "bbox": [ + 69, + 83, + 83, + 94 + ], + "lines": [ + { + "bbox": [ + 69, + 83, + 83, + 94 + ], + "spans": [ + { + "bbox": [ + 69, + 83, + 83, + 94 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 154, + 74, + 277, + 84 + ], + "lines": [ + { + "bbox": [ + 154, + 74, + 277, + 84 + ], + "spans": [ + { + "bbox": [ + 154, + 74, + 277, + 84 + ], + "type": "text", + "content": "Spatial-Temporal Attention (STAtten)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 69, + 95, + 359, + 208 + ], + "lines": [ + { + "bbox": [ + 69, + 95, + 359, + 208 + ], + "spans": [ + { + "bbox": [ + 69, + 95, + 359, + 208 + ], + "type": "image", + "image_path": "cac2d98b2a42e1917ce9c60b58e0008f7e9e9500c4aaa35794f0fa5dcacbe75d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 210, + 555, + 244 + ], + "lines": [ + { + "bbox": [ + 54, + 210, + 555, + 244 + ], + "spans": [ + { + "bbox": [ + 54, + 210, + 555, + 244 + ], + "type": "text", + "content": "Figure 4. Overview of STAtten architecture. (a) Block-wise temporal attention mechanism. Binary Q, K, V tensors are partitioned into temporal blocks, where black lines indicate paired timestep processing. (b) Computation flow with tensor dimensions, where " + }, + { + "bbox": [ + 54, + 210, + 555, + 244 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 54, + 210, + 555, + 244 + ], + "type": "text", + "content": " is the number of timesteps, " + }, + { + "bbox": [ + 54, + 210, + 555, + 244 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 54, + 210, + 555, + 244 + ], + "type": "text", + "content": " is the number of tokens, " + }, + { + "bbox": [ + 54, + 210, + 555, + 244 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 54, + 210, + 555, + 244 + ], + "type": "text", + "content": " is block size, and " + }, + { + "bbox": [ + 54, + 210, + 555, + 244 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 54, + 210, + 555, + 244 + ], + "type": "text", + "content": " is the feature dimension." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 374, + 81, + 539, + 207 + ], + "blocks": [ + { + "bbox": [ + 374, + 81, + 539, + 207 + ], + "lines": [ + { + "bbox": [ + 374, + 81, + 539, + 207 + ], + "spans": [ + { + "bbox": [ + 374, + 81, + 539, + 207 + ], + "type": "image", + "image_path": "d62e9d2b639829e04d5385ee6d827087828fdbbaa9139fdbf1e9797dcc5d4954.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 256, + 296, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 256, + 296, + 316 + ], + "spans": [ + { + "bbox": [ + 54, + 256, + 296, + 316 + ], + "type": "text", + "content": "spatial-temporal attention achieves lower entropy (4.85) with higher accuracy " + }, + { + "bbox": [ + 54, + 256, + 296, + 316 + ], + "type": "inline_equation", + "content": "(79.9\\%)" + }, + { + "bbox": [ + 54, + 256, + 296, + 316 + ], + "type": "text", + "content": ". These findings verify that combining spatial and temporal processing leads to more structured feature representations, providing strong motivation for spatial-temporal attention." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 316, + 296, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 316, + 296, + 508 + ], + "spans": [ + { + "bbox": [ + 55, + 316, + 296, + 508 + ], + "type": "text", + "content": "However, we argue that fully correlating temporal information across all timesteps like ST_Attn in Eq. 6 introduces two challenges: (1) substantial memory requirements for storing temporal information, and (2) increased dead neurons when correlating distant timesteps due to weak spike similarity. The first challenge arises from matrix multiplication with large temporal dimensions " + }, + { + "bbox": [ + 55, + 316, + 296, + 508 + ], + "type": "inline_equation", + "content": "(TN,D)" + }, + { + "bbox": [ + 55, + 316, + 296, + 508 + ], + "type": "text", + "content": ". To address this, we partition the temporal dimension into blocks, reducing the memory footprint for matrix operations. As shown in Fig. 3(a), this block-wise computation strategy leads to around " + }, + { + "bbox": [ + 55, + 316, + 296, + 508 + ], + "type": "inline_equation", + "content": "1.6\\times" + }, + { + "bbox": [ + 55, + 316, + 296, + 508 + ], + "type": "text", + "content": " higher batch size compared to full temporal attention. The second challenge stems from binary matrix multiplication in spike-based processing. Our analysis in Fig. 3(b) reveals that QKV computations show higher active neurons when processing same-timestep, while activity decreases with temporal distance." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 515, + 174, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 515, + 174, + 528 + ], + "spans": [ + { + "bbox": [ + 55, + 515, + 174, + 528 + ], + "type": "text", + "content": "4.2. STAtten Mechanism" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 533, + 296, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 533, + 296, + 582 + ], + "spans": [ + { + "bbox": [ + 55, + 533, + 296, + 582 + ], + "type": "text", + "content": "Based on these observations, we propose Spatial-Temporal Attention (STAtten). The block-wise STAtten partitions the temporal sequence into blocks of size " + }, + { + "bbox": [ + 55, + 533, + 296, + 582 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 55, + 533, + 296, + 582 + ], + "type": "text", + "content": ", as shown in Fig. 4(a). We define STAtten as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 590, + 234, + 604 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 590, + 234, + 604 + ], + "spans": [ + { + "bbox": [ + 55, + 590, + 234, + 604 + ], + "type": "interline_equation", + "content": "\\operatorname {S T A t t e n} (\\mathbf {X} [ b ]) = \\operatorname {L I F} (\\mathbf {Q} [ b ] \\mathbf {K} ^ {\\top} [ b ] \\mathbf {V} [ b ] \\cdot \\alpha),", + "image_path": "a3f90d7a8f3402e55390a71f92eaa83e4268a850b55e916d9d6919e7df84f318.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 605, + 295, + 628 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 628 + ], + "type": "interline_equation", + "content": "\\text {w h e r e} [ b ] = [ i B: (i + 1) B, :, d ], \\quad i \\in \\{0, 1, \\dots , T / B - 1 \\} \\tag {8}", + "image_path": "68117a6ce531df61fbdaf1ccd18e112437e28eac43f436f4bca442ac3faaf9df.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "inline_equation", + "content": "[iB:(i + 1)B, :, d]" + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "content": " denotes the aggregated features from timesteps " + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "inline_equation", + "content": "iB" + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "inline_equation", + "content": "(i + 1)B" + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "content": ". For instance, with " + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "inline_equation", + "content": "T = 4" + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "inline_equation", + "content": "B = 2" + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "content": ", features are grouped in pairs of timesteps [0,1], [2,3]. By processing blocks of timesteps together, we capture local temporal relationships while maintaining manageable computation. This approach requires storing only a subset of temporal information at any given time, reduc" + } + ] + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 316, + 319, + 555, + 408 + ], + "blocks": [ + { + "bbox": [ + 313, + 255, + 555, + 310 + ], + "lines": [ + { + "bbox": [ + 313, + 255, + 555, + 310 + ], + "spans": [ + { + "bbox": [ + 313, + 255, + 555, + 310 + ], + "type": "text", + "content": "Table 1. Computational complexity and energy consumption of different attention mechanisms. " + }, + { + "bbox": [ + 313, + 255, + 555, + 310 + ], + "type": "inline_equation", + "content": "E_{MAC}" + }, + { + "bbox": [ + 313, + 255, + 555, + 310 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 255, + 555, + 310 + ], + "type": "inline_equation", + "content": "E_{AC}" + }, + { + "bbox": [ + 313, + 255, + 555, + 310 + ], + "type": "text", + "content": " are energy costs for MAC and AC operations, " + }, + { + "bbox": [ + 313, + 255, + 555, + 310 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 255, + 555, + 310 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 255, + 555, + 310 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 313, + 255, + 555, + 310 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 255, + 555, + 310 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 255, + 555, + 310 + ], + "type": "text", + "content": " denote patches, dimension, timesteps, and " + }, + { + "bbox": [ + 313, + 255, + 555, + 310 + ], + "type": "inline_equation", + "content": "S_{Q}" + }, + { + "bbox": [ + 313, + 255, + 555, + 310 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 255, + 555, + 310 + ], + "type": "inline_equation", + "content": "S_{K}" + }, + { + "bbox": [ + 313, + 255, + 555, + 310 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 255, + 555, + 310 + ], + "type": "inline_equation", + "content": "S_{V}" + }, + { + "bbox": [ + 313, + 255, + 555, + 310 + ], + "type": "text", + "content": " are firing rates of " + }, + { + "bbox": [ + 313, + 255, + 555, + 310 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 313, + 255, + 555, + 310 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 255, + 555, + 310 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 255, + 555, + 310 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 255, + 555, + 310 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 313, + 255, + 555, + 310 + ], + "type": "text", + "content": ". SDT represents Spike-driven Transformer." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 316, + 319, + 555, + 408 + ], + "lines": [ + { + "bbox": [ + 316, + 319, + 555, + 408 + ], + "spans": [ + { + "bbox": [ + 316, + 319, + 555, + 408 + ], + "type": "table", + "html": "
MethodSTComplexityEnergy
ViT [11]✘O(N2D)EMAC·N2D
ViViT [2]✓O(T2N2D)EMAC·T2N2D
Spikformer [64]✘O(TND2)EAC·TND2·(SQ+SK+SV)
SDT [53]✘O(TND)EAC·TND·(SQ+SK)
SDT-V2 [52]✘O(TND2)EAC·TND2·(SQ+SK+SV)
QKFormer [62]✘O(TND2)EAC·TND2·(SQ+SK+SV)
STAtten✓O(TND2)EAC·TND2·(SQ+SK+SV)
", + "image_path": "4baa0e3533484639448e6ff5520c6e08a28f030fed2898beb8f901ac0858e9d8.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 433, + 555, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 433, + 555, + 468 + ], + "spans": [ + { + "bbox": [ + 313, + 433, + 555, + 468 + ], + "type": "text", + "content": "ing memory overhead. We set block size " + }, + { + "bbox": [ + 313, + 433, + 555, + 468 + ], + "type": "inline_equation", + "content": "B = 2" + }, + { + "bbox": [ + 313, + 433, + 555, + 468 + ], + "type": "text", + "content": " with 4 timesteps for static datasets and " + }, + { + "bbox": [ + 313, + 433, + 555, + 468 + ], + "type": "inline_equation", + "content": "B = 4" + }, + { + "bbox": [ + 313, + 433, + 555, + 468 + ], + "type": "text", + "content": " with 16 timesteps for neuromorphic datasets." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 472, + 555, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 472, + 555, + 519 + ], + "spans": [ + { + "bbox": [ + 313, + 472, + 555, + 519 + ], + "type": "text", + "content": "Unlike traditional transformers that require Softmax operations, the binary " + }, + { + "bbox": [ + 313, + 472, + 555, + 519 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}" + }, + { + "bbox": [ + 313, + 472, + 555, + 519 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 472, + 555, + 519 + ], + "type": "inline_equation", + "content": "\\mathbf{K}" + }, + { + "bbox": [ + 313, + 472, + 555, + 519 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 472, + 555, + 519 + ], + "type": "inline_equation", + "content": "\\mathbf{V}" + }, + { + "bbox": [ + 313, + 472, + 555, + 519 + ], + "type": "text", + "content": " allow flexible reordering of computations. Generalizing to arbitrary number of timesteps " + }, + { + "bbox": [ + 313, + 472, + 555, + 519 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 472, + 555, + 519 + ], + "type": "text", + "content": ", the computation follows:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 320, + 533, + 553, + 555 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 533, + 553, + 555 + ], + "spans": [ + { + "bbox": [ + 320, + 533, + 553, + 555 + ], + "type": "interline_equation", + "content": "\\mathbf {Q}, \\mathbf {K}, \\mathbf {V} \\in \\mathbb {R} ^ {\\frac {T}{B} \\times N \\times D} \\xrightarrow {\\text {r e s h a p e}} \\mathbf {Q}, \\mathbf {K}, \\mathbf {V} \\in \\mathbb {R} ^ {\\left(\\frac {T}{B} N\\right) \\times D} \\tag {9}", + "image_path": "5b03ef7af44401fd8817b7d8e298be0a97b704ce7707f4309d22e3c94cd7deca.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 332, + 554, + 533, + 568 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 332, + 554, + 533, + 568 + ], + "spans": [ + { + "bbox": [ + 332, + 554, + 533, + 568 + ], + "type": "interline_equation", + "content": "\\mathbf {K} ^ {\\top} \\mathbf {V} \\in \\mathbb {R} ^ {D \\times D} \\xrightarrow {\\text {a t t e n t i o n}} \\mathbf {Q} (\\mathbf {K} ^ {\\top} \\mathbf {V}) \\in \\mathbb {R} ^ {\\frac {T}{B} \\times N \\times D}.", + "image_path": "829e750114fd4ed463afb2aea99a3b23f1b2f7bbb6f13c67e4ed78c0b325a3f2.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 582, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 556, + 713 + ], + "type": "text", + "content": "The computation sequence reduces intermediate memory requirements through early " + }, + { + "bbox": [ + 313, + 582, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{K}^{\\top}\\mathbf{V}" + }, + { + "bbox": [ + 313, + 582, + 556, + 713 + ], + "type": "text", + "content": " calculation while preserving cross-timestep information flow within blocks. This implementation maintains computational complexity of " + }, + { + "bbox": [ + 313, + 582, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{O}(TND^{2})" + }, + { + "bbox": [ + 313, + 582, + 556, + 713 + ], + "type": "text", + "content": ", matching spatial-only attention [42, 52, 53, 62, 64] while incorporating temporal dependencies. By avoiding the quadratic complexity in timesteps and the number of tokens (" + }, + { + "bbox": [ + 313, + 582, + 556, + 713 + ], + "type": "inline_equation", + "content": "T^{2}N^{2}" + }, + { + "bbox": [ + 313, + 582, + 556, + 713 + ], + "type": "text", + "content": ") that would result from full temporal correlation, our method scales efficiently to longer sequences. Fig. 4(b) illustrates the complete computation process with corresponding tensor dimensions." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13951" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 112, + 296, + 213 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 297, + 105 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 297, + 105 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 297, + 105 + ], + "type": "text", + "content": "Table 2. Performance comparison on sequential CIFAR10/100 datasets, demonstrating the effectiveness of spatial-temporal attention in capturing long-term dependencies." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 112, + 296, + 213 + ], + "lines": [ + { + "bbox": [ + 56, + 112, + 296, + 213 + ], + "spans": [ + { + "bbox": [ + 56, + 112, + 296, + 213 + ], + "type": "table", + "html": "
Methods-CIFAR10 (%)s-CIFAR100 (%)
PLIF [14]81.4753.38
Spikformer [64]79.2957.17
STAtten + [64]82.4558.75
Spike-driven Transformer [53]80.3261.08
STAtten + [53]83.4164.30
SpikingReformer [42]79.2557.48
STAtten + [42]81.8458.30
", + "image_path": "1e5466c4f7b682260cd1755f79787c5b6547fde135617473001c843f1629522f.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 230, + 288, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 230, + 288, + 243 + ], + "spans": [ + { + "bbox": [ + 55, + 230, + 288, + 243 + ], + "type": "text", + "content": "4.3. STAtten with Existing Spiking Transformers" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 248, + 296, + 320 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 248, + 296, + 320 + ], + "spans": [ + { + "bbox": [ + 55, + 248, + 296, + 320 + ], + "type": "text", + "content": "STAtten is designed to be a plug-and-play module that can enhance various spike-based transformer architectures, including Spikformer [64], Spike-driven Transformer [53], Spike-driven Transformer-V2 [52], SpikingReformer [42], and QKformer [62] while preserving their unique characteristics." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 321, + 296, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 321, + 296, + 499 + ], + "spans": [ + { + "bbox": [ + 55, + 321, + 296, + 499 + ], + "type": "text", + "content": "The key to successful integration lies in preserving each architecture's distinct residual connection strategies while replacing their spatial-only attentions with STAtten. Two primary residual connection mechanisms are prevalent in current architectures: Spike Element-Wise (SEW) shortcut [13] and Membrane-Shortcut (MS) [20]. In architectures employing SEW shortcuts, such as Spikformer and QKFormer, spike activations are directly added to the attention output. In contrast, Spike-driven Transformer, Spikedriven Transformer-V2, and SpikingReformer utilize MSshortcuts that propagate membrane potentials rather than spikes Our STAtten module maintains compatibility with both connection types. We detail the existing spike-based attention mechanisms of each architecture and our integration strategy." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 500, + 296, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 500, + 296, + 536 + ], + "spans": [ + { + "bbox": [ + 55, + 500, + 296, + 536 + ], + "type": "text", + "content": "Spikformer [64]: Spikformer proposes Spiking Self Attention (SSA) module which is inspired by the vanilla self-attention of ViT [11] as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 78, + 543, + 294, + 558 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 543, + 294, + 558 + ], + "spans": [ + { + "bbox": [ + 78, + 543, + 294, + 558 + ], + "type": "interline_equation", + "content": "S S A (\\mathbf {Q}, \\mathbf {K}, \\mathbf {V}) = \\operatorname {L I F} (\\mathbf {Q} \\odot \\mathbf {K} ^ {\\top} \\odot \\mathbf {V} \\cdot \\alpha), \\tag {10}", + "image_path": "98fd9b93c7f32774d545dcdcd70ed6a95ba2353602ae94ca6a4c2c1e3902d677.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 564, + 296, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 564, + 296, + 648 + ], + "spans": [ + { + "bbox": [ + 55, + 564, + 296, + 648 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 564, + 296, + 648 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 55, + 564, + 296, + 648 + ], + "type": "text", + "content": " is matrix multiplication. This work establishes two key principles for spike-based attention: non-softmax operation and flexible Q, K, V operation ordering. While these principles form the foundation for subsequent spike-based architectures, the SSA module processes features independently at each timestep. We address this limitation by replacing the attention mechanism with STAtten." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 648, + 296, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 648, + 296, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 648, + 296, + 685 + ], + "type": "text", + "content": "Spike-driven Transformer [53]: Spike-driven Transformer introduces Spike-Driven Self-Attention (SDSA), which reformulates attention computation as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 73, + 691, + 296, + 716 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 691, + 296, + 716 + ], + "spans": [ + { + "bbox": [ + 73, + 691, + 296, + 716 + ], + "type": "interline_equation", + "content": "S D S A (\\mathbf {Q}, \\mathbf {K}, \\mathbf {V}) = \\mathbf {Q} \\otimes \\operatorname {L I F} \\left(\\sum_ {c} (\\mathbf {K} \\otimes \\mathbf {V})\\right), \\tag {11}", + "image_path": "fdfc668c82176c46d9e0ae2c2472ea36a097c6c0158844b58518046c512ab94a.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "inline_equation", + "content": "\\otimes" + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "content": " denotes element-wise multiplication and " + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "inline_equation", + "content": "\\sum_{c}" + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "content": " represents column-wise summation. While SDSA achieves computational efficiency through element-wise operations, this design limits feature correlation even within the spatial domain due to the element-wise multiplication. By integrating STAtten into this architecture, we enable comprehensive spatial-temporal feature correlation." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 156, + 554, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 156, + 554, + 240 + ], + "spans": [ + { + "bbox": [ + 313, + 156, + 554, + 240 + ], + "type": "text", + "content": "SpikingReformer [61]: SpikingReformer identifies two key challenges in adapting conventional Q, K, V attention to spike-based transformers: (1) floating-point matrix multiplication between Q, K, V and (2) Softmax operations involving exponentiation and division. To address these challenges, the authors propose Dual Spike Self-Attention (DSSA):" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 350, + 248, + 553, + 263 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 248, + 553, + 263 + ], + "spans": [ + { + "bbox": [ + 350, + 248, + 553, + 263 + ], + "type": "interline_equation", + "content": "D S S A (\\mathbf {X}) = \\operatorname {L I F} \\left(\\mathbf {X} \\odot \\mathbf {W} ^ {\\top} \\odot \\mathbf {X} ^ {\\top} \\cdot \\alpha\\right), \\tag {12}", + "image_path": "bf8eecf502cd48839a639fe86a502849dfd6e467e22cf38c89e117a999e35318.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 272, + 554, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 272, + 554, + 332 + ], + "spans": [ + { + "bbox": [ + 313, + 272, + 554, + 332 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 272, + 554, + 332 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 313, + 272, + 554, + 332 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 272, + 554, + 332 + ], + "type": "inline_equation", + "content": "\\mathbf{W}" + }, + { + "bbox": [ + 313, + 272, + 554, + 332 + ], + "type": "text", + "content": " denote binary input and linear projection, respectively. We try to verify that these concerns can be addressed by replacing DSSA with STAtten, showing that " + }, + { + "bbox": [ + 313, + 272, + 554, + 332 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}" + }, + { + "bbox": [ + 313, + 272, + 554, + 332 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 272, + 554, + 332 + ], + "type": "inline_equation", + "content": "\\mathbf{K}" + }, + { + "bbox": [ + 313, + 272, + 554, + 332 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 272, + 554, + 332 + ], + "type": "inline_equation", + "content": "\\mathbf{V}" + }, + { + "bbox": [ + 313, + 272, + 554, + 332 + ], + "type": "text", + "content": " matrix multiplication is still compatible with spike-based processing when properly designed." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 332, + 554, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 332, + 554, + 380 + ], + "spans": [ + { + "bbox": [ + 313, + 332, + 554, + 380 + ], + "type": "text", + "content": "QKFormer [62]: QKFormer implements dual attention mechanisms: Q-K Channel Attention (QKTA) and SSA. While SSA follows the Spikformer design, QKTA introduces a token-wise attention:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 370, + 388, + 553, + 412 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 388, + 553, + 412 + ], + "spans": [ + { + "bbox": [ + 370, + 388, + 553, + 412 + ], + "type": "interline_equation", + "content": "Q K T A = \\operatorname {L I F} \\left(\\sum_ {N} (\\mathbf {Q})\\right) \\otimes \\mathbf {K}, \\tag {13}", + "image_path": "410ee85875bd2ebad841a4623215e7e70afa73dd5d2699186f123fc6694c86ab.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 422, + 554, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 422, + 554, + 482 + ], + "spans": [ + { + "bbox": [ + 313, + 422, + 554, + 482 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 422, + 554, + 482 + ], + "type": "inline_equation", + "content": "\\sum_{N}" + }, + { + "bbox": [ + 313, + 422, + 554, + 482 + ], + "type": "text", + "content": " represents token-wise summation. QKFormer redesigns the architecture by incorporating embedding layers in each encoder. In this architecture, we only replace SSA with STAtten as SSA is the primary mechanism responsible for capturing feature relationships." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 483, + 554, + 580 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 483, + 554, + 580 + ], + "spans": [ + { + "bbox": [ + 313, + 483, + 554, + 580 + ], + "type": "text", + "content": "As we only modify the self-attention module, the original information flow through residual connections remains unchanged, maintaining the architecture's core characteristics. After computing STAtten across temporal blocks, the features are concatenated and processed through subsequent layers: " + }, + { + "bbox": [ + 313, + 483, + 554, + 580 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{\\mathrm{out}} = \\mathrm{Concat}(\\mathrm{STAtten}(\\mathbf{X}[iB:(i + 1)B])_{i = 0}^{T / B - 1})" + }, + { + "bbox": [ + 313, + 483, + 554, + 580 + ], + "type": "text", + "content": " This concatenated feature then flows through convolutional and batch normalization layers: " + }, + { + "bbox": [ + 313, + 483, + 554, + 580 + ], + "type": "inline_equation", + "content": "\\mathbf{Z} = \\mathrm{BN}(\\mathrm{Conv}(\\mathbf{F}_{\\mathrm{out}}))" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 587, + 527, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 587, + 527, + 601 + ], + "spans": [ + { + "bbox": [ + 313, + 587, + 527, + 601 + ], + "type": "text", + "content": "4.4. Complexity and Energy of Self-attention" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 605, + 554, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 554, + 666 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 554, + 666 + ], + "type": "text", + "content": "We analyze both computational complexity and theoretical energy consumption of our proposed architecture, focusing specifically on the self-attention mechanism. A comprehensive evaluation of the overall architecture's energy profile is provided in Supplementary Material." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 666, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 666, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 666, + 554, + 714 + ], + "type": "text", + "content": "As shown in Table 1, despite incorporating spatial-temporal attention capabilities, our STAtten maintains the same computational complexity " + }, + { + "bbox": [ + 313, + 666, + 554, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{O}(TND^2)" + }, + { + "bbox": [ + 313, + 666, + 554, + 714 + ], + "type": "text", + "content": " as previous spatial-only spiking transformers. This efficiency is" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "13952" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 111, + 113, + 501, + 284 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 105 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 105 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 105 + ], + "type": "text", + "content": "Table 3. Performance comparison between our methods and previous works on CIFAR10 and CIFAR100 datasets. In the architecture column, Architecture- " + }, + { + "bbox": [ + 55, + 70, + 555, + 105 + ], + "type": "inline_equation", + "content": "L - D" + }, + { + "bbox": [ + 55, + 70, + 555, + 105 + ], + "type": "text", + "content": " represents " + }, + { + "bbox": [ + 55, + 70, + 555, + 105 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 55, + 70, + 555, + 105 + ], + "type": "text", + "content": " number of encoder blocks and " + }, + { + "bbox": [ + 55, + 70, + 555, + 105 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 55, + 70, + 555, + 105 + ], + "type": "text", + "content": " hidden dimensions. All baseline architectures of spike-based transformer and corresponding STAtten implementations are trained from scratch for fair comparison." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 111, + 113, + 501, + 284 + ], + "lines": [ + { + "bbox": [ + 111, + 113, + 501, + 284 + ], + "spans": [ + { + "bbox": [ + 111, + 113, + 501, + 284 + ], + "type": "table", + "html": "
MethodArchitectureTimestepCIFAR10 (%)CIFAR100 (%)
tdBN [60]ResNet19492.92-
PLIF [14]ConvNet893.50-
Dspike [31]ResNet18694.2574.24
DSR [36]ResNet182095.4078.50
DIET-SNN [39]VGG16592.7069.67
SNASNet [24]ConvNet593.6473.04
Spikformer [64]Spikformer-4-384493.9975.06
STAtten + [64]Spikformer-4-384494.3675.85
Spike-driven Transformer [53]Spike-driven Transformer-2-512495.6078.40
STAtten + [53]Spike-driven Transformer-2-512496.0379.85
SpikingReformer [42]SpikingReformer-6-384495.0377.16
STAtten + [42]SpikingReformer-6-384495.2677.90
QKFormer [62]QKFormer-4-384495.1279.79
STAtten + [62]QKFormer-4-384495.3580.20
", + "image_path": "23783534b76bd21b1c4f7f097a3a4420659b924a0355a289eee0651bea5530c5.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 297, + 294, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 297, + 294, + 477 + ], + "spans": [ + { + "bbox": [ + 54, + 297, + 294, + 477 + ], + "type": "text", + "content": "achieved through our non-Softmax design, which enables flexible reordering of Q, K, V operations. In contrast, conventional ANN-based spatial-temporal attention methods like ViViT [2] require substantially higher complexity " + }, + { + "bbox": [ + 54, + 297, + 294, + 477 + ], + "type": "inline_equation", + "content": "(\\mathcal{O}(T^2 N^2 D))" + }, + { + "bbox": [ + 54, + 297, + 294, + 477 + ], + "type": "text", + "content": " due to their softmax-constrained computation order. For energy consumption estimation, we follow the floating-point operations (FLOPs) implementation in 45nm technology [18]. The energy costs are categorized into " + }, + { + "bbox": [ + 54, + 297, + 294, + 477 + ], + "type": "inline_equation", + "content": "E_{MAC}" + }, + { + "bbox": [ + 54, + 297, + 294, + 477 + ], + "type": "text", + "content": " for multiply-accumulate operations corresponding to 32-/8-bit ANN computations, and " + }, + { + "bbox": [ + 54, + 297, + 294, + 477 + ], + "type": "inline_equation", + "content": "E_{AC}" + }, + { + "bbox": [ + 54, + 297, + 294, + 477 + ], + "type": "text", + "content": " for accumulate operations used in binary SNN calculations. The final energy consumption is modulated by the firing rates " + }, + { + "bbox": [ + 54, + 297, + 294, + 477 + ], + "type": "inline_equation", + "content": "(S_Q, S_K, \\text{and} S_V)" + }, + { + "bbox": [ + 54, + 297, + 294, + 477 + ], + "type": "text", + "content": " of the respective Q, K, and V. A detailed analysis of these energy measurements and their implications is presented in Section 5.3." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 487, + 137, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 487, + 137, + 502 + ], + "spans": [ + { + "bbox": [ + 55, + 487, + 137, + 502 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 507, + 295, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 507, + 295, + 602 + ], + "spans": [ + { + "bbox": [ + 55, + 507, + 295, + 602 + ], + "type": "text", + "content": "We verify our proposed methods on both static and dynamic datasets, e.g., CIFAR10/100 [26], ImageNet [7], CIFAR10-DVS [29], and N-Caltech101 [37] datasets. We use direct coding to convert float pixel values into binary spikes [49]. In all our experiments, the spiking transformer STAtten models are trained from scratch. The detailed training strategy for each dataset is explained in the Supplementary Material." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 603, + 295, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 603, + 295, + 662 + ], + "spans": [ + { + "bbox": [ + 55, + 603, + 295, + 662 + ], + "type": "text", + "content": "Notation: In the experiments, we use the notation STAtten " + }, + { + "bbox": [ + 55, + 603, + 295, + 662 + ], + "type": "inline_equation", + "content": "+\\left[\\cdot \\right]" + }, + { + "bbox": [ + 55, + 603, + 295, + 662 + ], + "type": "text", + "content": " to denote STAtten implementation based on the backbone architecture used in " + }, + { + "bbox": [ + 55, + 603, + 295, + 662 + ], + "type": "inline_equation", + "content": "\\left[\\cdot \\right]" + }, + { + "bbox": [ + 55, + 603, + 295, + 662 + ], + "type": "text", + "content": ". E.g., STAtten+[64] refers to training a STAtten model with Spikformer backbone from scratch." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 671, + 260, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 671, + 260, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 671, + 260, + 685 + ], + "type": "text", + "content": "5.1. Sequential CIFAR10/100 Classification" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "type": "text", + "content": "We first evaluate our spatial-temporal attention mechanism's capability to capture long-term temporal dependen" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 297, + 555, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 297, + 555, + 489 + ], + "spans": [ + { + "bbox": [ + 313, + 297, + 555, + 489 + ], + "type": "text", + "content": "cies using sequential CIFAR10 and CIFAR100 datasets. In this task, we transform the standard image classification task into a temporal sequence processing task by dividing each input image column-wise, shown in Fig. 1, where each column serves as input for one timestep. This results in a sequence length of 32 timesteps, corresponding to the image width. To accommodate this 1-D sequential input, we modify the convolutional, batch normalization, and pooling layers from 2-D to 1-D operations. As shown in Table 2, STAtten outperforms existing spatial-only attention methods on both datasets. Specifically, in Spike-driven Transformer [53], our approach achieves improvements of " + }, + { + "bbox": [ + 313, + 297, + 555, + 489 + ], + "type": "inline_equation", + "content": "3.09\\%" + }, + { + "bbox": [ + 313, + 297, + 555, + 489 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 297, + 555, + 489 + ], + "type": "inline_equation", + "content": "3.22\\%" + }, + { + "bbox": [ + 313, + 297, + 555, + 489 + ], + "type": "text", + "content": " on s-CIFAR10 and s-CIFAR100, respectively, compared to the baseline models. This performance gain demonstrates STAtten's capability in capturing long-range temporal dependencies." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 493, + 440, + 506 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 493, + 440, + 506 + ], + "spans": [ + { + "bbox": [ + 313, + 493, + 440, + 506 + ], + "type": "text", + "content": "5.2. Performance Analysis" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 509, + 555, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 509, + 555, + 617 + ], + "spans": [ + { + "bbox": [ + 313, + 509, + 555, + 617 + ], + "type": "text", + "content": "CIFAR10/100. The CIFAR10/100 datasets comprise static images of " + }, + { + "bbox": [ + 313, + 509, + 555, + 617 + ], + "type": "inline_equation", + "content": "32 \\times 32" + }, + { + "bbox": [ + 313, + 509, + 555, + 617 + ], + "type": "text", + "content": " pixels, with 50,000 training and 10,000 test images. As shown in Table 3, STAtten improves the performance of various spike-based transformers, achieving " + }, + { + "bbox": [ + 313, + 509, + 555, + 617 + ], + "type": "inline_equation", + "content": "95.35\\%" + }, + { + "bbox": [ + 313, + 509, + 555, + 617 + ], + "type": "text", + "content": " accuracy on CIFAR10 and " + }, + { + "bbox": [ + 313, + 509, + 555, + 617 + ], + "type": "inline_equation", + "content": "80.20\\%" + }, + { + "bbox": [ + 313, + 509, + 555, + 617 + ], + "type": "text", + "content": " on CIFAR100 when integrated with QKFormer. These improvements over spatial-only attention methods are consistent across different architectures, demonstrating the benefits of spatial-temporal processing." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "type": "text", + "content": "ImageNet. ImageNet is a large-scale image dataset with images sized " + }, + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "type": "text", + "content": " pixels, comprising approximately 1.2 million training images and 50,000 validation images across 1,000 classes. As shown in Table 4, STAtten enhances the performance of spike-based transformers. With Spike-driven Transformer [53], it achieves " + }, + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "type": "inline_equation", + "content": "78.11\\%" + }, + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "type": "text", + "content": " accuracy at " + }, + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "type": "text", + "content": " resolution and " + }, + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "type": "inline_equation", + "content": "78.39\\%" + }, + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "type": "text", + "content": " at " + }, + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "type": "inline_equation", + "content": "288 \\times 288" + }, + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "type": "text", + "content": " resolution. Integration with Spike-driven Transformer-V2" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13953" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 88, + 112, + 523, + 338 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 105 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 105 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 105 + ], + "type": "text", + "content": "Table 4. Performance comparison between our methods and previous works on ImageNet datasets. Note that * represents the inference accuracy with " + }, + { + "bbox": [ + 55, + 70, + 555, + 105 + ], + "type": "inline_equation", + "content": "288 \\times 288" + }, + { + "bbox": [ + 55, + 70, + 555, + 105 + ], + "type": "text", + "content": " resolution. In the architecture column, " + }, + { + "bbox": [ + 55, + 70, + 555, + 105 + ], + "type": "inline_equation", + "content": "L-D" + }, + { + "bbox": [ + 55, + 70, + 555, + 105 + ], + "type": "text", + "content": " represents " + }, + { + "bbox": [ + 55, + 70, + 555, + 105 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 55, + 70, + 555, + 105 + ], + "type": "text", + "content": " number of encoder blocks and " + }, + { + "bbox": [ + 55, + 70, + 555, + 105 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 55, + 70, + 555, + 105 + ], + "type": "text", + "content": " hidden dimensions. † represents our implementation." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 88, + 112, + 523, + 338 + ], + "lines": [ + { + "bbox": [ + 88, + 112, + 523, + 338 + ], + "spans": [ + { + "bbox": [ + 88, + 112, + 523, + 338 + ], + "type": "table", + "html": "
MethodTypeArchitectureParam(M)TimestepAccuracy(%)
Vision Transformer [11]ANNViT-B/1686.59177.90
DeiT [46]ANNDeiT-B86.59181.80
Swin Transformer [33]ANNSwin Transformer-B87.77183.50
TET [9]SNNSEW-ResNet3421.79468.00
Spiking ResNet [19]SNNResNet5025.5635072.75
tdBN [60]SNNResNet3421.79663.72
SEW-ResNet [13]SNNSEW-ResNet15260.19469.26
MS-ResNet [20]SNNResNet10478.37574.21 / 76.02*
Att-MS-ResNet [51]SNNResNet10478.37477.08*
Spikformer [64]SNNSpikformer-8-76866.34474.81
SpikingReformer [42]SNNSpikingReformer-6-102466.38478.77 / 79.40*
QKFormer [62]SNNQKFormer-6-102464.96484.22
Spike-driven Transformer [53]SNNSpike-driven Transformer-8-51229.68474.57
SNNSpike-driven Transformer-10-51236.01474.66
SNNSpike-driven Transformer-8-76866.34476.32 / 77.07*
STAtten + [53]SNNSpike-driven Transformer-8-51229.68476.18 / 76.56*
SNNSpike-driven Transformer-8-76866.34478.11 / 78.39*
Spike-driven Transformer-V2 [52]SNNSpike-driven Transformer-V2-8-51255.4479.49† / 79.98*†
STAtten + [52]SNNSpike-driven Transformer-V2-8-51255.4479.85 / 80.67*
", + "image_path": "1f3839cf5e17c44d3b2af05676cd00a5b946f1fdcaad3a091d743b4064fb8821.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 97, + 390, + 514, + 571 + ], + "blocks": [ + { + "bbox": [ + 55, + 348, + 555, + 383 + ], + "lines": [ + { + "bbox": [ + 55, + 348, + 555, + 383 + ], + "spans": [ + { + "bbox": [ + 55, + 348, + 555, + 383 + ], + "type": "text", + "content": "Table 5. Performance comparison between our methods and previous works on CIFAR10-DVS and N-Caltech101 datasets. In the architecture column, " + }, + { + "bbox": [ + 55, + 348, + 555, + 383 + ], + "type": "inline_equation", + "content": "L - D" + }, + { + "bbox": [ + 55, + 348, + 555, + 383 + ], + "type": "text", + "content": " represents " + }, + { + "bbox": [ + 55, + 348, + 555, + 383 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 55, + 348, + 555, + 383 + ], + "type": "text", + "content": " number of encoder blocks and " + }, + { + "bbox": [ + 55, + 348, + 555, + 383 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 55, + 348, + 555, + 383 + ], + "type": "text", + "content": " hidden dimensions. All baseline architectures of spike-based transformer and corresponding STAtten implementations are trained from scratch for fair comparison." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 97, + 390, + 514, + 571 + ], + "lines": [ + { + "bbox": [ + 97, + 390, + 514, + 571 + ], + "spans": [ + { + "bbox": [ + 97, + 390, + 514, + 571 + ], + "type": "table", + "html": "
MethodArchitectureTimestepCIFAR10-DVS (%)N-Caltech101 (%)
tdBN [60]ResNet191067.80-
PLIF [14]ConvNet2074.80-
Dspike [31]ResNet181075.40-
DSR [36]ResNet182077.27-
SEW-ResNet [13]ConvNet2074.80-
TT-SNN [27]ResNet346-77.80
NDA [32]VGG111079.678.2
Spikformer [64]Spikformer-2-2561680.9-
Spike-driven Transformer [53]Spike-driven Transformer-2-2561680.081.80
STAtten + [53]Spike-driven Transformer-2-2561681.183.15
SpikingReformer [42]SpikingReformer-4-3841678.8081.29
STAtten + [42]SpikingReformer-4-3841680.6081.95
QKFormer [62]QKFormer-4-3841682.9083.58
STAtten + [62]SpikingReformer-4-3841683.9084.25
", + "image_path": "7fd3e889e109d6cde778998dd5b344685eca1434b385e8532d7ae19b125e2ed6.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 578, + 295, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 578, + 295, + 673 + ], + "spans": [ + { + "bbox": [ + 55, + 578, + 295, + 673 + ], + "type": "text", + "content": "[52] further improves performance to " + }, + { + "bbox": [ + 55, + 578, + 295, + 673 + ], + "type": "inline_equation", + "content": "79.85\\%" + }, + { + "bbox": [ + 55, + 578, + 295, + 673 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 578, + 295, + 673 + ], + "type": "inline_equation", + "content": "80.67\\%" + }, + { + "bbox": [ + 55, + 578, + 295, + 673 + ], + "type": "text", + "content": " at respective resolutions. It is noteworthy that Spike-driven Transformer [53] achieves comparable or better performance than recent Spikformer models while using only 29.68M parameters. Furthermore, STAtten with Spik-driven Transformer-V2 achieves even higher accuracy than SpikingReformer and Spikformer while using around " + }, + { + "bbox": [ + 55, + 578, + 295, + 673 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 55, + 578, + 295, + 673 + ], + "type": "text", + "content": " fewer parameters." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": "Neuromorphic Datasets. We evaluate our approach on two neuromorphic vision datasets: CIFAR10-DVS [29] and N-Caltech101 [37]. CIFAR10-DVS contains 10,000 sam" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 578, + 555, + 711 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 578, + 555, + 711 + ], + "spans": [ + { + "bbox": [ + 313, + 578, + 555, + 711 + ], + "type": "text", + "content": "plies (9,000 training, 1,000 testing) captured using a Dynamic Vision Sensor (DVS) from static CIFAR10 images. N-Caltech101, derived from the Caltech101 dataset [28], comprises 8,831 DVS image samples. Following standard protocols, we resize CIFAR10-DVS samples to " + }, + { + "bbox": [ + 313, + 578, + 555, + 711 + ], + "type": "inline_equation", + "content": "64 \\times 64" + }, + { + "bbox": [ + 313, + 578, + 555, + 711 + ], + "type": "text", + "content": " and apply the augmentation strategy from [53], while N-Caltech101 samples are resized to " + }, + { + "bbox": [ + 313, + 578, + 555, + 711 + ], + "type": "inline_equation", + "content": "64 \\times 64" + }, + { + "bbox": [ + 313, + 578, + 555, + 711 + ], + "type": "text", + "content": " with NDA [32] augmentation. As shown in Table 5, integrating STAtten consistently improves performance across different architectures. With Spike-driven Transformer as the backbone, our approach achieves " + }, + { + "bbox": [ + 313, + 578, + 555, + 711 + ], + "type": "inline_equation", + "content": "81.1\\%" + }, + { + "bbox": [ + 313, + 578, + 555, + 711 + ], + "type": "text", + "content": " accuracy on CIFAR10-DVS" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "13954" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 140, + 296, + 249 + ], + "blocks": [ + { + "bbox": [ + 55, + 71, + 295, + 137 + ], + "lines": [ + { + "bbox": [ + 55, + 71, + 295, + 137 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 295, + 137 + ], + "type": "text", + "content": "Table 6. Comparison of memory, energy, and accuracy on ImageNet across different transformer architectures. The precision format weight/activation bit-width and the memory include the size of model parameters and activations. We show the memory consumption for both 32-bit and 8-bit weight spiking models. SDT represents Spike-driven Transformer." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 140, + 296, + 249 + ], + "lines": [ + { + "bbox": [ + 56, + 140, + 296, + 249 + ], + "spans": [ + { + "bbox": [ + 56, + 140, + 296, + 249 + ], + "type": "table", + "html": "
MethodPrecisionTimestepMemory (MB)Energy (mJ)
ViT-B/16 [11]32/321351.8254.84
Liu (ViT-B) [34]8/8187.96110.8
Spikformer [64]32(8)/14285.99 (86.48)21.48
SpikingReformer [42]32(8)/14273.57 (85.4)8.76
QKFormer [62]32(8)/14280.28 (84.99)38.91
SDT [53]32(8)/14283.97 (87.42)12.42
STAtten + [53]32(8)/14283.97 (87.42)12.36
SDT-V2 [52]32(8)/14250.22 (84.02)52.40
STAtten + [53]32(8)/14250.22 (84.02)52.38
", + "image_path": "90d30f6e5a8ecfe35e807d2aa60bb87137fee3767a8e3d1d0e20e5b375aab964.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 58, + 262, + 171, + 364 + ], + "blocks": [ + { + "bbox": [ + 58, + 262, + 171, + 364 + ], + "lines": [ + { + "bbox": [ + 58, + 262, + 171, + 364 + ], + "spans": [ + { + "bbox": [ + 58, + 262, + 171, + 364 + ], + "type": "image", + "image_path": "94e18cb5eafbd04650560b31e5b958270b53d8b65a734d2896aeeb0d1c8b8ba1.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 111, + 370, + 123, + 380 + ], + "lines": [ + { + "bbox": [ + 111, + 370, + 123, + 380 + ], + "spans": [ + { + "bbox": [ + 111, + 370, + 123, + 380 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 176, + 262, + 290, + 365 + ], + "blocks": [ + { + "bbox": [ + 176, + 262, + 290, + 365 + ], + "lines": [ + { + "bbox": [ + 176, + 262, + 290, + 365 + ], + "spans": [ + { + "bbox": [ + 176, + 262, + 290, + 365 + ], + "type": "image", + "image_path": "58e98b294bc28df5f6974027d45af64fcbfd288272783b8d1e7aff1c76f01f9b.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 227, + 369, + 240, + 380 + ], + "lines": [ + { + "bbox": [ + 227, + 369, + 240, + 380 + ], + "spans": [ + { + "bbox": [ + 227, + 369, + 240, + 380 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 385, + 295, + 430 + ], + "lines": [ + { + "bbox": [ + 55, + 385, + 295, + 430 + ], + "spans": [ + { + "bbox": [ + 55, + 385, + 295, + 430 + ], + "type": "text", + "content": "Figure 5. Accuracy comparison with respect to the number of parameters on (a) CIFAR100 and (b) Sequential CIFAR100 datasets. Spike-driven Transformer [53] is used as the baseline of spatial-only architecture." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 434, + 295, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 434, + 295, + 506 + ], + "spans": [ + { + "bbox": [ + 55, + 434, + 295, + 506 + ], + "type": "text", + "content": "and " + }, + { + "bbox": [ + 55, + 434, + 295, + 506 + ], + "type": "inline_equation", + "content": "83.15\\%" + }, + { + "bbox": [ + 55, + 434, + 295, + 506 + ], + "type": "text", + "content": " on N-Caltech101, improving the baseline by " + }, + { + "bbox": [ + 55, + 434, + 295, + 506 + ], + "type": "inline_equation", + "content": "1.1\\%" + }, + { + "bbox": [ + 55, + 434, + 295, + 506 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 434, + 295, + 506 + ], + "type": "inline_equation", + "content": "1.35\\%" + }, + { + "bbox": [ + 55, + 434, + 295, + 506 + ], + "type": "text", + "content": " respectively. The enhancement extends to other architectures as well, with SpikingReformer achieving " + }, + { + "bbox": [ + 55, + 434, + 295, + 506 + ], + "type": "inline_equation", + "content": "80.60\\%" + }, + { + "bbox": [ + 55, + 434, + 295, + 506 + ], + "type": "text", + "content": " on CIFAR10-DVS and " + }, + { + "bbox": [ + 55, + 434, + 295, + 506 + ], + "type": "inline_equation", + "content": "81.95\\%" + }, + { + "bbox": [ + 55, + 434, + 295, + 506 + ], + "type": "text", + "content": " on N-Caltech101, and QKFormer reaching " + }, + { + "bbox": [ + 55, + 434, + 295, + 506 + ], + "type": "inline_equation", + "content": "83.90\\%" + }, + { + "bbox": [ + 55, + 434, + 295, + 506 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 434, + 295, + 506 + ], + "type": "inline_equation", + "content": "84.25\\%" + }, + { + "bbox": [ + 55, + 434, + 295, + 506 + ], + "type": "text", + "content": " respectively, which is the state-of-the-art result." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 514, + 218, + 527 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 514, + 218, + 527 + ], + "spans": [ + { + "bbox": [ + 55, + 514, + 218, + 527 + ], + "type": "text", + "content": "5.3. Memory and Energy Analysis" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "text", + "content": "We analyze memory and energy consumption during ImageNet inference. As shown in Table 6, traditional transformer, ViT-B/16, with 32-bit precision requires high memory (351.8 MB) and energy consumption (254.84 mJ). Even with 8-bit precision, Liu [34] still consumes " + }, + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "inline_equation", + "content": "110.8\\mathrm{mJ}" + }, + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "text", + "content": " of energy. In contrast, spike-based transformers with 32-bit weights and binary activations demonstrate significant energy efficiency. Notably, STAtten integrated with SDT and SDT-V2 achieves similar energy consumption to the baseline architectures, showing that our spatial-temporal processing can be realized without additional energy overhead. Correspondingly, STAtten attains comparable performance with fewer parameters, as detailed in the next section. The methodology for energy calculations is provided in the Supplementary Material." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 72, + 411, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 411, + 85 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 411, + 85 + ], + "type": "text", + "content": "5.4. Model Capacity" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 90, + 555, + 269 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 90, + 555, + 269 + ], + "spans": [ + { + "bbox": [ + 313, + 90, + 555, + 269 + ], + "type": "text", + "content": "To further evaluate the capacity of our STAtten in spike-based transformers, we compare the accuracy based on the number of trainable parameters through ablation studies. Fig. 5 shows the performance scaling across different model sizes on both (a) standard CIFAR100 and (b) sequential CIFAR100 datasets. Spike-driven Transformer [53] is used as the baseline of spatial-only architecture. As shown in Fig. 5(a), on CIFAR100, STAtten demonstrates strong performance across various model sizes ranging from 2.56M to 22.97M parameters. The accuracy improvement (" + }, + { + "bbox": [ + 313, + 90, + 555, + 269 + ], + "type": "inline_equation", + "content": "\\sim" + }, + { + "bbox": [ + 313, + 90, + 555, + 269 + ], + "type": "text", + "content": " 0.5-1.0%) is consistently maintained across different architectural scales. For Sequential CIFAR100 (Fig. 5(b)), where temporal dependencies are more crucial, STAtten shows particularly robust performance improvement (" + }, + { + "bbox": [ + 313, + 90, + 555, + 269 + ], + "type": "inline_equation", + "content": "\\sim" + }, + { + "bbox": [ + 313, + 90, + 555, + 269 + ], + "type": "text", + "content": " 3-5%) even with smaller architectures." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 274, + 388, + 285 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 274, + 388, + 285 + ], + "spans": [ + { + "bbox": [ + 314, + 274, + 388, + 285 + ], + "type": "text", + "content": "5.5. Limitation" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 292, + 556, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 292, + 556, + 484 + ], + "spans": [ + { + "bbox": [ + 313, + 292, + 556, + 484 + ], + "type": "text", + "content": "The hardware implementation of STAtten faces challenges despite its improved accuracy. Full spatial-temporal attention without block partitioning is impractical for neuromorphic deployment, as it requires complete temporal information. Our block-wise approach partially addresses this by processing temporal information in steps, where each step corresponds to one block size. However, even with this optimization, deployment remains challenging on traditional neuromorphic chips like TrueNorth [1] and Loihi [6] that process information step-by-step. A potential solution lies in layer-by-layer neuromorphic chips [22, 54, 56, 63] that support layer-wise processing across timesteps. This architectural shift provides an efficient implementation path for STAtten's block-wise processing. Future integration of parallel LIF neurons [15] could further accelerate computation in layer-by-layer architectures." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 314, + 489, + 388, + 501 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 489, + 388, + 501 + ], + "spans": [ + { + "bbox": [ + 314, + 489, + 388, + 501 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 509, + 555, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 509, + 555, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 509, + 555, + 628 + ], + "type": "text", + "content": "This paper proposes a block-wise spatial-temporal attention mechanism, STAtten, that enhances spike-based transformers. Through block-wise processing and leveraging non-softmax properties, STAtten captures spatial-temporal information without additional overhead. STAtten can be integrated into existing architectures while preserving their energy efficiency, consistently improving performance across different backbones. Experimental results across both static and neuromorphic datasets validate the effectiveness of our approach." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 314, + 634, + 407, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 634, + 407, + 647 + ], + "spans": [ + { + "bbox": [ + 314, + 634, + 407, + 647 + ], + "type": "text", + "content": "Acknowledgment" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 653, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 555, + 713 + ], + "type": "text", + "content": "This work was supported in part by CoCoSys, a JUMP2.0 center sponsored by DARPA and SRC, the National Science Foundation (CAREER Award, Grant #2312366, Grant #2318152), the DARPA Young Faculty Award and the DoEMMICC center SEA-CROGS (Award #DE-SC0023198)." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13955" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 167 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 167 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 167 + ], + "type": "text", + "content": "[1] Filipp Akopyan, Jun Sawada, Andrew Cassidy, Rodrigo Alvarez-Icaza, John Arthur, Paul Merolla, Nabil Imam, Yutaka Nakamura, Pallab Datta, Gi-Joon Nam, et al. Truenorth: Design and tool flow of a 65 mw 1 million neuron programmable neurosynaptic chip. IEEE transactions on computer-aided design of integrated circuits and systems, 34 (10):1537-1557, 2015. 1, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 170, + 296, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 170, + 296, + 224 + ], + "spans": [ + { + "bbox": [ + 61, + 170, + 296, + 224 + ], + "type": "text", + "content": "[2] Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lucic, and Cordelia Schmid. Vivit: A video vision transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6836-6846, 2021. 1, 4, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 226, + 295, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 226, + 295, + 258 + ], + "spans": [ + { + "bbox": [ + 61, + 226, + 295, + 258 + ], + "type": "text", + "content": "[3] Anthony N Burkitt. A review of the integrate-and-fire neuron model: I. homogeneous synaptic input. Biological cybernetics, 95:1-19, 2006. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 260, + 296, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 260, + 296, + 316 + ], + "spans": [ + { + "bbox": [ + 61, + 260, + 296, + 316 + ], + "type": "text", + "content": "[4] Lili Chen, Kevin Lu, Aravind Rajeswaran, Kimin Lee, Aditya Grover, Misha Laskin, Pieter Abbeel, Aravind Srinivas, and Igor Mordatch. Decision transformer: Reinforcement learning via sequence modeling. Advances in neural information processing systems, 34:15084-15097, 2021. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 317, + 295, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 317, + 295, + 371 + ], + "spans": [ + { + "bbox": [ + 61, + 317, + 295, + 371 + ], + "type": "text", + "content": "[5] Sayeed Shafayet Chowdhury, Isha Garg, and Kaushik Roy. Spatio-temporal pruning and quantization for low-latency spiking neural networks. In 2021 International Joint Conference on Neural Networks (IJCNN), pages 1-9. IEEE, 2021. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 61, + 373, + 296, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 373, + 296, + 429 + ], + "spans": [ + { + "bbox": [ + 61, + 373, + 296, + 429 + ], + "type": "text", + "content": "[6] Mike Davies, Narayan Srinivasa, Tsung-Han Lin, Gautham Chinya, Yongqiang Cao, Sri Harsha Choday, Georgios Dimou, Prasad Joshi, Nabil Imam, Shweta Jain, et al. Loihi: A neuromorphic manycore processor with on-chip learning. IEEE Micro, 38(1):82-99, 2018. 1, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 430, + 296, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 430, + 296, + 475 + ], + "spans": [ + { + "bbox": [ + 62, + 430, + 296, + 475 + ], + "type": "text", + "content": "[7] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 475, + 295, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 475, + 295, + 509 + ], + "spans": [ + { + "bbox": [ + 62, + 475, + 295, + 509 + ], + "type": "text", + "content": "[8] Shikuang Deng and Shi Gu. Optimal conversion of conventional artificial neural networks to spiking neural networks. arXiv preprint arXiv:2103.00476, 2021. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 510, + 296, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 510, + 296, + 553 + ], + "spans": [ + { + "bbox": [ + 62, + 510, + 296, + 553 + ], + "type": "text", + "content": "[9] Shikuang Deng, Yuhang Li, Shanghang Zhang, and Shi Gu. Temporal efficient training of spiking neural network via gradient re-weighting. arXiv preprint arXiv:2202.11946, 2022. 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 556, + 295, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 556, + 295, + 600 + ], + "spans": [ + { + "bbox": [ + 57, + 556, + 295, + 600 + ], + "type": "text", + "content": "[10] Peter U Diehl et al. Fast-classifying, high-accuracy spiking deep networks through weight and threshold balancing. In 2015 International joint conference on neural networks (IJCNN), pages 1–8. IEEE, 2015. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 601, + 295, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 601, + 295, + 667 + ], + "spans": [ + { + "bbox": [ + 57, + 601, + 295, + 667 + ], + "type": "text", + "content": "[11] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 1, 3, 4, 5, 7, 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 669, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 669, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 669, + 295, + 712 + ], + "type": "text", + "content": "[12] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. International journal of computer vision, 88:303-338, 2010. 3" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 554, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 316, + 73, + 554, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 116 + ], + "type": "text", + "content": "[13] Wei Fang, Zhaofei Yu, Yanqi Chen, Tiejun Huang, Timothee Masquelier, and Yonghong Tian. Deep residual learning in spiking neural networks. Advances in Neural Information Processing Systems, 34:21056-21069, 2021. 5, 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 118, + 554, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 118, + 554, + 184 + ], + "spans": [ + { + "bbox": [ + 317, + 118, + 554, + 184 + ], + "type": "text", + "content": "[14] Wei Fang, Zhaofei Yu, Yanqi Chen, Timothee Masquelier, Tiejun Huang, and Yonghong Tian. Incorporating learnable membrane time constant to enhance learning of spiking neural networks. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2661-2671, 2021. 2, 5, 6, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 185, + 553, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 185, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 316, + 185, + 553, + 239 + ], + "type": "text", + "content": "[15] Wei Fang, Zhaofei Yu, Zhaokun Zhou, Ding Chen, Yanqi Chen, Zhengyu Ma, Timothée Masquelier, and Yonghong Tian. Parallel spiking neurons with high efficiency and ability to learn long-term dependencies. Advances in Neural Information Processing Systems, 36, 2024. 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 241, + 553, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 241, + 553, + 296 + ], + "spans": [ + { + "bbox": [ + 316, + 241, + 553, + 296 + ], + "type": "text", + "content": "[16] Marylou Gabrie, Andre Manoel, Clément Luneau, Nicolas Macris, Florent Krzakala, Lenka Zdeborova, et al. Entropy and mutual information in models of deep neural networks. Advances in neural information processing systems, 31, 2018. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 297, + 553, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 297, + 553, + 341 + ], + "spans": [ + { + "bbox": [ + 316, + 297, + 553, + 341 + ], + "type": "text", + "content": "[17] Bing Han and Kaushik Roy. Deep spiking neural network: Energy efficiency through time based coding. In European Conference on Computer Vision, pages 388-404. Springer, 2020. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 342, + 553, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 342, + 553, + 386 + ], + "spans": [ + { + "bbox": [ + 316, + 342, + 553, + 386 + ], + "type": "text", + "content": "[18] Mark Horowitz. 1.1 computing's energy problem (and what we can do about it). In 2014 IEEE international solid-state circuits conference digest of technical papers (ISSCC), pages 10-14. IEEE, 2014. 6, 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 388, + 553, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 388, + 553, + 421 + ], + "spans": [ + { + "bbox": [ + 316, + 388, + 553, + 421 + ], + "type": "text", + "content": "[19] Yangfan Hu, Huajin Tang, and Gang Pan. Spiking deep residual networks. IEEE Transactions on Neural Networks and Learning Systems, 34(8):5200-5205, 2021. 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 422, + 553, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 422, + 553, + 465 + ], + "spans": [ + { + "bbox": [ + 316, + 422, + 553, + 465 + ], + "type": "text", + "content": "[20] Yifan Hu, Lei Deng, Yujie Wu, Man Yao, and Guoqi Li. Advancing spiking neural networks toward deep residual learning. IEEE Transactions on Neural Networks and Learning Systems, 2024. 1, 2, 5, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 467, + 553, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 467, + 553, + 520 + ], + "spans": [ + { + "bbox": [ + 317, + 467, + 553, + 520 + ], + "type": "text", + "content": "[21] Seijoon Kim, Seongsik Park, Byunggook Na, and Sungroh Yoon. Spiking-yolo: spiking neural network for energy-efficient object detection. In Proceedings of the AAAI conference on artificial intelligence, pages 11270-11277, 2020. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 522, + 554, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 522, + 554, + 599 + ], + "spans": [ + { + "bbox": [ + 317, + 522, + 554, + 599 + ], + "type": "text", + "content": "[22] Sangyeob Kim, Soyeon Kim, Seongyon Hong, Sangjin Kim, Donghyeon Han, and Hoi-Jun Yoo. C-dnn: A 24.5-85.8 tops/w complementary-deep-neural-network processor with heterogeneous cnn/snn core architecture and forward-gradient-based sparsity generation. In 2023 IEEE International Solid-State Circuits Conference (ISSCC), pages 334-336. IEEE, 2023. 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 600, + 553, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 600, + 553, + 645 + ], + "spans": [ + { + "bbox": [ + 316, + 600, + 553, + 645 + ], + "type": "text", + "content": "[23] Youngeun Kim, Joshua Chough, and Priyadarshini Panda. Beyond classification: Directly training spiking neural networks for semantic segmentation. Neuromorphic Computing and Engineering, 2(4):044015, 2022. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 647, + 553, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 690 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 690 + ], + "type": "text", + "content": "[24] Youngeun Kim, Yuhang Li, Hyoungseob Park, Yeshwanth Venkatesha, and Priyadarshini Panda. Neural architecture search for spiking neural networks. In European Conference on Computer Vision, pages 36-56. Springer, 2022. 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "text", + "content": "[25] Youngeun Kim, Yuhang Li, Hyoungseob Park, Yeshwanth Venkatesha, Anna Hambitzer, and Priyadarshini Panda. Ex" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "13956" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 76, + 73, + 295, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 73, + 295, + 106 + ], + "spans": [ + { + "bbox": [ + 76, + 73, + 295, + 106 + ], + "type": "text", + "content": "ploring temporal information dynamics in spiking neural networks. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 8308-8316, 2023. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 295, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 295, + 129 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 295, + 129 + ], + "type": "text", + "content": "[26] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 129, + 295, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 295, + 183 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 295, + 183 + ], + "type": "text", + "content": "[27] Donghyun Lee, Ruokai Yin, Youngeun Kim, Abhishek Moitra, Yuhang Li, and Priyadarshini Panda. Tt-snn: tensor train decomposition for efficient spiking neural network training. In 2024 Design, Automation & Test in Europe Conference & Exhibition (DATE), pages 1-6. IEEE, 2024. 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 184, + 294, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 184, + 294, + 205 + ], + "spans": [ + { + "bbox": [ + 57, + 184, + 294, + 205 + ], + "type": "text", + "content": "[28] FF Li, M Andreeto, M Ranzato, and P Perona. Caltech 101 (1.0)[data set]. caltechdata, 2022. 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 57, + 206, + 294, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 206, + 294, + 249 + ], + "spans": [ + { + "bbox": [ + 57, + 206, + 294, + 249 + ], + "type": "text", + "content": "[29] Hongmin Li, Hanchao Liu, Xiangyang Ji, Guoqi Li, and Luping Shi. Cifar10-dvs: an event-stream dataset for object classification. Frontiers in neuroscience, 11:244131, 2017. 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 251, + 295, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 251, + 295, + 303 + ], + "spans": [ + { + "bbox": [ + 56, + 251, + 295, + 303 + ], + "type": "text", + "content": "[30] Yuhang Li, Shikuang Deng, Xin Dong, Ruihao Gong, and Shi Gu. A free lunch from ann: Towards efficient, accurate spiking neural networks calibration. In International conference on machine learning, pages 6316-6325. PMLR, 2021. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 57, + 306, + 294, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 306, + 294, + 360 + ], + "spans": [ + { + "bbox": [ + 57, + 306, + 294, + 360 + ], + "type": "text", + "content": "[31] Yuhang Li, Yufei Guo, Shanghang Zhang, Shikuang Deng, Yongqing Hai, and Shi Gu. Differentiable spike: Rethinking gradient-descent for training spiking neural networks. Advances in Neural Information Processing Systems, 34: 23426-23439, 2021. 2, 6, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 361, + 295, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 361, + 295, + 414 + ], + "spans": [ + { + "bbox": [ + 56, + 361, + 295, + 414 + ], + "type": "text", + "content": "[32] Yuhang Li, Youngeun Kim, Hyoungseob Park, Tamar Geller, and Priyadarshini Panda. Neuromorphic data augmentation for training spiking neural networks. In European Conference on Computer Vision, pages 631-649. Springer, 2022. 7, 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 416, + 295, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 416, + 295, + 470 + ], + "spans": [ + { + "bbox": [ + 56, + 416, + 295, + 470 + ], + "type": "text", + "content": "[33] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 472, + 295, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 472, + 295, + 514 + ], + "spans": [ + { + "bbox": [ + 56, + 472, + 295, + 514 + ], + "type": "text", + "content": "[34] Zhenhua Liu, Yunhe Wang, Kai Han, Wei Zhang, Siwei Ma, and Wen Gao. Post-training quantization for vision transformer. Advances in Neural Information Processing Systems, 34:28092-28103, 2021. 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 515, + 295, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 515, + 295, + 547 + ], + "spans": [ + { + "bbox": [ + 56, + 515, + 295, + 547 + ], + "type": "text", + "content": "[35] Wolfgang Maass. Networks of spiking neurons: the third generation of neural network models. Neural networks, 10 (9):1659-1671, 1997. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 548, + 295, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 548, + 295, + 613 + ], + "spans": [ + { + "bbox": [ + 56, + 548, + 295, + 613 + ], + "type": "text", + "content": "[36] Qingyan Meng, Mingqing Xiao, Shen Yan, Yisen Wang, Zhouchen Lin, and Zhi-Quan Luo. Training high-performance low-latency spiking neural networks by differentiation on spike representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12444–12453, 2022. 6, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 615, + 295, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 615, + 295, + 658 + ], + "spans": [ + { + "bbox": [ + 56, + 615, + 295, + 658 + ], + "type": "text", + "content": "[37] Garrick Orchard, Ajinkya Jayawant, Gregory K Cohen, and Nitish Thakor. Converting static image datasets to spiking neuromorphic datasets using saccades. Frontiers in neuroscience, 9:159859, 2015. 6, 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 658, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 295, + 712 + ], + "type": "text", + "content": "[38] Jing Pei, Lei Deng, Sen Song, Mingguo Zhao, Youhui Zhang, Shuang Wu, Guanrui Wang, Zhe Zou, Zhenzhi Wu, Wei He, et al. Towards artificial general intelligence with hybrid tianjic chip architecture. Nature, 572(7767):106-111, 2019. 1" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "text", + "content": "[39] Nitin Rathi and Kaushik Roy. Diet-snn: Direct input encoding with leakage and threshold optimization in deep spiking neural networks. arXiv preprint arXiv:2008.03658, 2020. 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 107, + 553, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 107, + 553, + 139 + ], + "spans": [ + { + "bbox": [ + 317, + 107, + 553, + 139 + ], + "type": "text", + "content": "[40] Kaushik Roy, Akhilesh Jaiswal, and Priyadarshini Panda. Towards spike-based machine intelligence with neuromorphic computing. Nature, 575(7784):607-617, 2019. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 140, + 553, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 140, + 553, + 182 + ], + "spans": [ + { + "bbox": [ + 316, + 140, + 553, + 182 + ], + "type": "text", + "content": "[41] Abhronil Sengupta, Yuting Ye, Robert Wang, Chiao Liu, and Kaushik Roy. Going deeper in spiking neural networks: Vgg and residual architectures. Frontiers in neuroscience, 13:95, 2019. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 184, + 553, + 238 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 184, + 553, + 238 + ], + "spans": [ + { + "bbox": [ + 316, + 184, + 553, + 238 + ], + "type": "text", + "content": "[42] Xinyu Shi, Zecheng Hao, and Zhaofei Yu. Spikingresformer: Bridging resnet and vision transformer in spiking neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5610-5619, 2024. 1, 2, 4, 5, 6, 7, 8, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 239, + 553, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 239, + 553, + 272 + ], + "spans": [ + { + "bbox": [ + 316, + 239, + 553, + 272 + ], + "type": "text", + "content": "[43] Sumit B Shrestha and Garrick Orchard. Slayer: Spike layer error reassignment in time. Advances in neural information processing systems, 31, 2018. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 273, + 553, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 273, + 553, + 327 + ], + "spans": [ + { + "bbox": [ + 316, + 273, + 553, + 327 + ], + "type": "text", + "content": "[44] Qiaoyi Su, Yuhong Chou, Yifan Hu, Jianing Li, Shijie Mei, Ziyang Zhang, and Guoqi Li. Deep directly-trained spiking neural networks for object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6555-6565, 2023. 2, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 328, + 553, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 328, + 553, + 371 + ], + "spans": [ + { + "bbox": [ + 316, + 328, + 553, + 371 + ], + "type": "text", + "content": "[45] Zhengkun Tian, Jiangyan Yi, Jianhua Tao, Ye Bai, Shuai Zhang, and Zhengqi Wen. Spike-triggered nonautoregressive transformer for end-to-end speech recognition. arXiv preprint arXiv:2005.07903, 2020. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 373, + 553, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 373, + 553, + 426 + ], + "spans": [ + { + "bbox": [ + 316, + 373, + 553, + 426 + ], + "type": "text", + "content": "[46] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In International conference on machine learning, pages 10347-10357. PMLR, 2021. 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 427, + 553, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 427, + 553, + 470 + ], + "spans": [ + { + "bbox": [ + 316, + 427, + 553, + 470 + ], + "type": "text", + "content": "[47] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 1, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 472, + 553, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 472, + 553, + 525 + ], + "spans": [ + { + "bbox": [ + 316, + 472, + 553, + 525 + ], + "type": "text", + "content": "[48] Qingyu Wang, Tielin Zhang, Minglun Han, Yi Wang, Duzhen Zhang, and Bo Xu. Complex dynamic neurons improved spiking transformer network for efficient automatic speech recognition. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 102-109, 2023. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 526, + 553, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 526, + 553, + 569 + ], + "spans": [ + { + "bbox": [ + 316, + 526, + 553, + 569 + ], + "type": "text", + "content": "[49] Yujie Wu, Lei Deng, Guoqi Li, Jun Zhu, Yuan Xie, and Luping Shi. Direct training for spiking neural networks: Faster, larger, better. In Proceedings of the AAAI conference on artificial intelligence, pages 1311-1318, 2019. 1, 2, 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 571, + 553, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 571, + 553, + 613 + ], + "spans": [ + { + "bbox": [ + 316, + 571, + 553, + 613 + ], + "type": "text", + "content": "[50] Shiting Xiao, Yuhang Li, Youngeun Kim, Donghyun Lee, and Priyadarshini Panda. Respike: residual frames-based hybrid spiking neural networks for efficient action recognition. Neuromorphic Computing and Engineering, 2025. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 615, + 553, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 615, + 553, + 658 + ], + "spans": [ + { + "bbox": [ + 316, + 615, + 553, + 658 + ], + "type": "text", + "content": "[51] Man Yao, Guangshe Zhao, Hengyu Zhang, Yifan Hu, Lei Deng, Yonghong Tian, Bo Xu, and Guoqi Li. Attention spiking neural networks. IEEE transactions on pattern analysis and machine intelligence, 2023. 7" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 659, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 659, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 659, + 553, + 713 + ], + "type": "text", + "content": "[52] Man Yao, Jiakui Hu, Tianxiang Hu, Yifan Xu, Zhaokun Zhou, Yonghong Tian, Bo Xu, and Guoqi Li. Spike-driven transformer v2: Meta spiking neural network architecture inspiring the design of next-generation neuromorphic chips. arXiv preprint arXiv:2404.03663, 2024. 1, 2, 4, 5, 7, 8" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13957" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 709 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "text", + "content": "[53] Man Yao, Jiakui Hu, Zhaokun Zhou, Li Yuan, Yonghong Tian, Bo Xu, and Guoqi Li. Spike-driven transformer. Advances in Neural Information Processing Systems, 36, 2024. 1, 2, 3, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 173 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 173 + ], + "type": "text", + "content": "[54] Ruokai Yin, Abhishek Moitra, Abhiroop Bhattacharjee, Youngeun Kim, and Priyadarshini Panda. Sata: Sparsity-aware training accelerator for spiking neural networks. IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems, 42(6):1926-1938, 2022. 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 174, + 294, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 174, + 294, + 228 + ], + "spans": [ + { + "bbox": [ + 56, + 174, + 294, + 228 + ], + "type": "text", + "content": "[55] Ruokai Yin, Youngeun Kim, Yuhang Li, Abhishek Moitra, Nitin Satpute, Anna Hambitzer, and Priyadarshini Panda. Workload-balanced pruning for sparse spiking neural networks. IEEE Transactions on Emerging Topics in Computational Intelligence, 2024. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 229, + 294, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 229, + 294, + 284 + ], + "spans": [ + { + "bbox": [ + 56, + 229, + 294, + 284 + ], + "type": "text", + "content": "[56] Ruokai Yin, Youngeun Kim, Di Wu, and Priyadarshini Panda. Loas: Fully temporal-parallel dataflow for dualsparse spiking neural networks. In 2024 57th IEEE/ACM International Symposium on Microarchitecture (MICRO), pages 1107–1121. IEEE, 2024. 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 285, + 294, + 340 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 285, + 294, + 340 + ], + "spans": [ + { + "bbox": [ + 56, + 285, + 294, + 340 + ], + "type": "text", + "content": "[57] Ruokai Yin, Yuhang Li, Abhishek Moitra, and Priyadarshini Panda. Mint: Multiplier-less integer quantization for energy efficient spiking neural networks. In 2024 29th Asia and South Pacific Design Automation Conference (ASP-DAC), pages 830-835. IEEE, 2024. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 341, + 294, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 341, + 294, + 396 + ], + "spans": [ + { + "bbox": [ + 56, + 341, + 294, + 396 + ], + "type": "text", + "content": "[58] Jiqing Zhang, Bo Dong, Haiwei Zhang, Jianchuan Ding, Felix Heide, Baocai Yin, and Xin Yang. Spiking transformers for event-based single object tracking. In Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition, pages 8801-8810, 2022. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 396, + 294, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 396, + 294, + 441 + ], + "spans": [ + { + "bbox": [ + 56, + 396, + 294, + 441 + ], + "type": "text", + "content": "[59] Jiyuan Zhang, Lulu Tang, Zhaofei Yu, Jiwen Lu, and Tiejun Huang. Spike transformer: Monocular depth estimation for spiking camera. In European Conference on Computer Vision, pages 34-52. Springer, 2022. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 442, + 294, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 442, + 294, + 485 + ], + "spans": [ + { + "bbox": [ + 56, + 442, + 294, + 485 + ], + "type": "text", + "content": "[60] Hanle Zheng, Yujie Wu, Lei Deng, Yifan Hu, and Guoqi Li. Going deeper with directly-trained larger spiking neural networks. In Proceedings of the AAAI conference on artificial intelligence, pages 11062-11070, 2021. 1, 6, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 487, + 294, + 540 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 487, + 294, + 540 + ], + "spans": [ + { + "bbox": [ + 56, + 487, + 294, + 540 + ], + "type": "text", + "content": "[61] Chenlin Zhou, Liutao Yu, Zhaokun Zhou, Han Zhang, Zhengyu Ma, Huihui Zhou, and Yonghong Tian. Spikingformer: Spike-driven residual learning for transformer-based spiking neural network. arXiv preprint arXiv:2304.11954, 2023. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 542, + 294, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 542, + 294, + 597 + ], + "spans": [ + { + "bbox": [ + 56, + 542, + 294, + 597 + ], + "type": "text", + "content": "[62] Chenlin Zhou, Han Zhang, Zhaokun Zhou, Liutao Yu, Liwei Huang, Xiaopeng Fan, Li Yuan, Zhengyu Ma, Huihui Zhou, and Yonghong Tian. Qkformer: Hierarchical spiking transformer using qk attention. arXiv preprint arXiv:2403.16552, 2024. 1, 2, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 597, + 294, + 664 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 597, + 294, + 664 + ], + "spans": [ + { + "bbox": [ + 56, + 597, + 294, + 664 + ], + "type": "text", + "content": "[63] PJ Zhou, Q Yu, M Chen, YC Wang, LW Meng, Y Zuo, N Ning, Y Liu, SG Hu, and GC Qiao. A 0.96 pj/sop, 30.23 k-neuron/mm 2 heterogeneous neuromorphic chip with fullerene-like interconnection topology for edge-ai computing. In 2024 IEEE International Symposium on Circuits and Systems (ISCAS), pages 1–5. IEEE, 2024. 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 665, + 294, + 709 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 665, + 294, + 709 + ], + "spans": [ + { + "bbox": [ + 56, + 665, + 294, + 709 + ], + "type": "text", + "content": "[64] Zhaokun Zhou, Yuesheng Zhu, Chao He, Yaowei Wang, Shuicheng Yan, Yonghong Tian, and Li Yuan. Spikformer: When spiking neural network meets transformer. arXiv preprint arXiv:2209.15425, 2022. 1, 2, 3, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "13958" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Spiking Transformer_ Introducing Accurate Addition-Only Spiking Self-Attention for Transformer/24b0ab2b-a4ff-49fc-865c-b8d7799e49bd_content_list.json b/2025/Spiking Transformer_ Introducing Accurate Addition-Only Spiking Self-Attention for Transformer/24b0ab2b-a4ff-49fc-865c-b8d7799e49bd_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..5aa864b79b15309c7f954d9ee14d3575fe512442 --- /dev/null +++ b/2025/Spiking Transformer_ Introducing Accurate Addition-Only Spiking Self-Attention for Transformer/24b0ab2b-a4ff-49fc-865c-b8d7799e49bd_content_list.json @@ -0,0 +1,1712 @@ +[ + { + "type": "text", + "text": "Spiking Transformer: Introducing Accurate Addition-Only Spiking Self-Attention for Transformer", + "text_level": 1, + "bbox": [ + 112, + 130, + 883, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yufei Guo,\\* Xiaode Liu,\\* Yuanpei Chen, Weihang Peng, Yuhan Zhang, Zhe Ma\\* Intelligent Science & Technology Academy of CASIC, China", + "bbox": [ + 192, + 202, + 810, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "yfguo@pku.edu.cn, lxde@pku.edu.cn, mazhe_thu@163.com", + "bbox": [ + 264, + 242, + 725, + 255 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 291, + 326, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Transformers have demonstrated outstanding performance across a wide range of tasks, owing to their self-attention mechanism, but they are highly energy-consuming. Spiking Neural Networks have emerged as a promising energy-efficient alternative to traditional Artificial Neural Networks, leveraging event-driven computation and binary spikes for information transfer. The combination of Transformers' capabilities with the energy efficiency of SNNs offers a compelling opportunity. This paper addresses the challenge of adapting the self-attention mechanism of Transformers to the spiking paradigm by introducing a novel approach: Accurate Addition-Only Spiking Self-Attention ( $A^2OS^2A$ ). Unlike existing methods that rely solely on binary spiking neurons for all components of the self-attention mechanism, our approach integrates binary, ReLU, and ternary spiking neurons. This hybrid strategy significantly improves accuracy while preserving non-multiplicative computations. Moreover, our method eliminates the need for softmax and scaling operations. Extensive experiments show that the $A^2OS^2A$ -based Spiking Transformer outperforms existing SNN-based Transformers on several datasets, even achieving an accuracy of $78.66\\%$ on ImageNet-1K. Our work represents a significant advancement in SNN-based Transformer models, offering a more accurate and efficient solution for real-world applications.", + "bbox": [ + 91, + 324, + 483, + 700 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 732, + 220, + 748 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Spiking Neural Networks (SNNs) have gained significant traction as an efficient neural network model, finding applications in diverse areas such as object recognition [15, 16, 34, 56], object detection [29, 44], and pose estimation [66]. These networks utilize binary spike signals for information transfer. A spiking neuron generates a spike, denoted by 1, when its membrane potential surpasses a cer-", + "bbox": [ + 89, + 757, + 485, + 864 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tain threshold; conversely, it does not spike, represented by 0, when the threshold is not met. This distinctive approach to information processing is notably energy-efficient, as it substitutes the complex multiplications involved in weight and activation interactions with simpler addition operations. Furthermore, SNNs can be effectively implemented using event-driven computations on neuromorphic hardware [1, 7, 39, 43, 52]. In this setup, the computational units are activated solely upon the occurrence of a spike, conserving energy by remaining inactive in the absence of spikes. Research indicates that SNNs can achieve significantly greater energy savings compared to their Artificial Neural Network (ANN) counterparts [1, 7].", + "bbox": [ + 511, + 292, + 906, + 489 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite the clear advantages of SNNs in terms of energy efficiency, their practical application is often constrained by limited task accuracy. Meanwhile, Transformers have demonstrated exceptional performance across a wide range of tasks, largely due to their self-attention mechanism [30, 42, 50]. Combining the strengths of Transformers with the energy efficiency of SNNs presents a promising opportunity. However, adapting the self-attention mechanism to SNNs is non-trivial. In the vanilla self-attention framework (VSA), three key components are involved: Query $(Q)$ , Key $(K)$ , and Value $(V)$ . As illustrated in Figure 1(a), the VSA process begins by computing the dot product of the floating-point representations of $Q$ and $K$ , resulting in a matrix. This matrix is then normalized using a softmax function, which involves exponential and division operations, to produce an attention map that determines the weighting of $V$ . These operations in the VSA are incompatible with the operational principles of SNNs, which aim to minimize multiplication. Consequently, to implement a Transformer architecture in SNNs, it is essential to develop a novel, efficient self-attention mechanism that avoids multiplication. Ongoing research efforts are exploring solutions in this domain. Approaches such as Spikformer [64, 65], Spikingformer [63], and Spike-driven Transformer [58] have focused on transforming $Q$ , $K$ , and $V$ into spike representations prior to performing matrix operations. This transformation enables the replacement of matrix multiplications with addition, aligning", + "bbox": [ + 511, + 493, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution.", + "bbox": [ + 107, + 875, + 220, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author.", + "bbox": [ + 109, + 887, + 235, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "24398", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "better with the principles of SNNs.", + "bbox": [ + 89, + 90, + 321, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "While previous approaches utilize spike-form $Q$ , $K$ , and $V$ to avoid multiplications, simply introducing spiking neurons to convert $Q$ , $K$ , and $V$ into spike form prior to performing matrix multiplications does not fully exploit the advantages of the addition-only operation inherent in SNNs. It is unnecessary to convert all of $Q$ , $K$ , and $V$ into spike form in order to transition from matrix multiplications to matrix additions. However, this method can lead to significant information loss, as demonstrated in Section 4. To address this issue, we propose the Accurate Addition-Only Spiking Self-Attention (A $^2$ OS $^2$ A). The distinctions between A $^2$ OS $^2$ A and the vanilla Spiking Self-Attention (VSSA) are illustrated in Figure 1(b) and Figure 1(c). Unlike VSSA, which applies the same binary spiking neuron to generate $Q$ , $K$ , and $V$ , our approach employs a binary spiking neuron for $Q$ , a ReLU activation for $K$ , and a ternary spiking neuron for $V$ . This design allows A $^2$ OS $^2$ A to retain the multiplication-addition transformation while reducing information loss. Moreover, it eliminates the need for both the scaling and softmax functions, enabling the output of floating-point values similar to the VSA, whereas VSSA produces only integer outputs. The key contributions of this paper are as follows:", + "bbox": [ + 89, + 107, + 485, + 454 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We present a theoretical framework for understanding the information loss associated with vanilla Spiking Self-Attention mechanisms. To the best of our knowledge, this is the first detailed examination of this issue within the context of Spiking Self-Attention, which opens the door for future advancements in SNN-based Transformers.", + "- We introduce $\\mathrm{A}^2\\mathrm{OS}^2\\mathrm{A}$ , an innovative and efficient approach that incorporates binary spiking neurons, ReLU activations, and ternary spiking neurons, moving beyond the traditional binary-only spiking neuron model. This method reduces information loss while retaining the advantages of addition-only processing in SNNs. Additionally, it eliminates the need for scaling and softmax functions.", + "- Comprehensive experiments demonstrate that our proposed architecture either surpasses or matches the performance of State-of-the-Art (SoTA) SNN-based Transformer models on various datasets. Notably, we achieved an accuracy of $78.66\\%$ on ImageNet-1K, setting a new benchmark in the SNN domain." + ], + "bbox": [ + 89, + 458, + 482, + 744 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 761, + 232, + 777 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Spiking Neural Networks", + "text_level": 1, + "bbox": [ + 89, + 787, + 321, + 803 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "SNNs are a class of neural networks that more closely emulate the behavior of biological neurons compared to ANNs. The primary distinction lies in their use of discrete spikes for information transmission, as opposed to the continuous signals employed by ANNs. Foundational work in this area can be traced back to Hodgkin and Huxley [26], who", + "bbox": [ + 89, + 809, + 485, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "developed models to simulate the action potentials of biological neurons. Since then, numerous frameworks and models have been introduced to enhance the capabilities and learning efficiency of SNNs. A notable example is the Leaky Integrate-and-Fire (LIF) model [40], which provides a simple yet effective means of simulating neuronal behavior. More recently, the incorporation of Spike-Timing-Dependent Plasticity (STDP) as a learning rule has enabled SNNs to learn temporal patterns effectively [4].", + "bbox": [ + 511, + 90, + 906, + 227 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In the past decade, there has been a significant increase in interest regarding the training of SNNs using techniques adapted from ANNs, as well as their application to practical tasks. Two prevalent learning paradigms in current SNN research are the conversion from ANN to SNN (ANN2SNN) [2, 3, 22-25, 32, 34, 48] and supervised learning [12-14, 18-21, 35, 47, 59]. The ANN-SNN conversion method involves initially training an ANN and subsequently converting it into a homogeneous SNN by transferring the trained weights and substituting ReLU neurons with temporal spiking neurons. However, this method often fails for neuromorphic datasets, as ReLU neurons do not adequately capture the complex temporal dynamics required for processing sequential information. In contrast, supervised learning [11, 17, 53, 61] employs alternative functions during backpropagation to approximate the firing process, enabling the direct training of SNNs as if they were ANNs. This approach leverages the advantages of gradient-based optimization, achieving impressive performance with just a few time steps, even on large-scale datasets. Furthermore, supervised learning has proven to be effective in handling temporal data, establishing it as a preferred methodology in SNN research.", + "bbox": [ + 511, + 229, + 908, + 575 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Vision Transformers", + "text_level": 1, + "bbox": [ + 513, + 590, + 707, + 604 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Vision Transformers (ViTs) have recently emerged as a powerful alternative to convolutional neural networks (CNNs) for a wide range of visual tasks, demonstrating remarkable efficacy and flexibility. Introduced by Dosovitskiy et al. [10], ViTs leverage the transformer architecture—originally designed for natural language processing—to process images in a novel manner. The architecture innovatively splits images into patches and treats these patches as sequences, enabling the model to capture long-range dependencies and relationships within the visual data, a limitation often encountered in traditional CNNs. The introduction of ViTs has spurred significant advancements in various computer vision tasks, including image classification [41], object detection [38], and segmentation [28]. These advancements highlight the potential of ViTs to tackle complex visual recognition problems with unprecedented accuracy. Notably, it has been shown that with appropriate training techniques and larger datasets, ViTs can not only match but surpass the performance of state-of-the-art CNNs, even in data-scarce scenarios, presenting a", + "bbox": [ + 511, + 613, + 908, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "24399", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/450144b6bf9cce2e7f345cdf281ed193bc9ac75e1450daeb4316cc0d6fe35738.jpg", + "image_caption": [ + "Figure 1. The difference between our Spiking Self-Attention and the vanilla Spiking Self-Attention. Our Spiking Self-Attention differs significantly from the vanilla Spiking Self-Attention. In the vanilla version, only binary spikes are employed, which can result in considerable information loss. In contrast, our approach integrates a combination of binary spiking neurons, ReLU, and ternary spiking neurons. This hybrid structure effectively reduces information loss, while preserving the advantages of addition-only processing. Moreover, our method eliminates the need for both scaling and the softmax function, streamlining the computation." + ], + "image_footnote": [], + "bbox": [ + 99, + 89, + 357, + 358 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/e2ca4f51cb57896d5beadc3da775bafdaaf9468bfe96318c9c0ef58f2c8dad44.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 359, + 88, + 635, + 357 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/e4ab02036ee1026ef21566b118003223f9804a95c5e45d0939ea639bf050daf0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 640, + 95, + 901, + 357 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "significant advantage in real-world applications [49].", + "bbox": [ + 89, + 467, + 439, + 483 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Furthermore, the adaptation of transformers for visual tasks has inspired numerous variations and improvements across the field. Researchers have explored strategies such as integrating convolutional layers into the architecture [6] to retain spatial information more effectively. Other approaches include the use of hierarchical representations [36], which facilitate multi-scale feature extraction, and the incorporation of attention mechanisms tailored specifically for vision tasks [5], improving the model's ability to focus on relevant parts of the input data. These developments underscore the versatility and robustness of the transformer architecture in addressing complex visual challenges, paving the way for further innovations in computer vision.", + "bbox": [ + 89, + 484, + 483, + 680 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Spiking Neural Networks for Transformers", + "text_level": 1, + "bbox": [ + 89, + 695, + 457, + 710 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The integration of SNNs into transformer architectures represents an exciting and rapidly developing area of research. This field aims to leverage the distinct advantages of both paradigms: the energy-efficient characteristics of SNNs and the robust contextual representation capabilities of transformers. Recent studies [33, 37, 51] have explored replacing certain neurons within transformers with spiking neurons, marking a significant step towards this integration. While these efforts contribute to mitigating the accuracy loss associated with introducing spiking neurons into transformers, they still face challenges in fully realizing the low energy consumption benefits of SNNs. This is primarily due to", + "bbox": [ + 89, + 719, + 485, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "their reliance on hybrid computing methodologies, which continue to require traditional Multiply-Accumulate (MAC) operations, such as dot products, softmax calculations, and scaling operations.", + "bbox": [ + 511, + 467, + 906, + 527 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To address this issue, Spikformer [64, 65] proposes converting the $Q$ , $K$ , and $V$ components into spike form before performing matrix multiplications similar to those used in VSA. This innovative approach allows spike matrix multiplications to be transformed into addition operations, eliminating the need for softmax computations. However, when considering the residual connections in these architectures, Spikformer [64, 65] still incorporates non-spiking computations within the ConvBN layers. To overcome this limitation, Spikingformer [63] and Spike-driven Transformer [58] have restructured the residual connections in standard transformer architectures, placing them before activation functions to ensure that all neurons convey binary spike signals.", + "bbox": [ + 511, + 547, + 908, + 744 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The computation of spike-form $Q$ , $K$ , and $V$ in these models circumvent multiplications, relying solely on additions. However, simply placing spiking neurons in front of $Q$ , $K$ , and $V$ to convert them into spike form before performing matrix multiplications does not fully exploit the addition-only advantage of SNNs. In this work, we propose a novel, addition-only spiking self-attention mechanism for transformers, aiming to push the boundaries of what can be achieved through the integration of these two technologies.", + "bbox": [ + 511, + 763, + 908, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "24400", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Preliminary", + "text_level": 1, + "bbox": [ + 89, + 89, + 217, + 107 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Spiking Neuron Layer", + "text_level": 1, + "bbox": [ + 89, + 114, + 299, + 131 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The spiking neuron layer plays a crucial role in integrating both spatial and temporal information, which is then encoded into the membrane potential and subsequently transformed into binary spikes. These spikes drive further computations in subsequent layers of the network. We model the dynamics of the spiking neuron using the LIF model [40]. The evolution of the membrane potential and the spike generation mechanism are described by the following equations:", + "bbox": [ + 89, + 136, + 485, + 257 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nU [ t ] = H [ t - 1 ] + X [ t ], \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 163, + 267, + 483, + 285 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nS [ t ] = \\operatorname {H e a} (U [ t ] - V _ {\\mathrm {t h}}), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 166, + 286, + 483, + 303 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nH [ t ] = V _ {\\text {r e s e t}} S [ t ] + \\beta U [ t ] (1 - S [ t ]), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 163, + 305, + 483, + 321 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where: $X[t]$ is the spatial input current at time $t$ , $U[t]$ is the membrane potential at time $t$ , combining the spatial input $X[t]$ and the temporal input $H[t - 1]$ from the previous time step, $\\mathrm{Hea}(\\cdot)$ is the Heaviside step function, which outputs 1 if its argument is non-negative, and 0 otherwise, $S[t]$ is the spike output at time $t$ , $V_{\\mathrm{th}}$ is the spike threshold, $V_{\\mathrm{reset}}$ is the reset potential, and $\\beta$ is a decay factor controlling the membrane potential's decay.", + "bbox": [ + 89, + 332, + 483, + 453 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "When the membrane potential $U[t]$ exceeds a predefined threshold $V_{\\mathrm{th}}$ , the neuron emits a spike, $S[t] = 1$ , and the internal state $H[t]$ is reset to $V_{\\mathrm{reset}}$ . If $U[t]$ does not surpass the threshold, the membrane potential decays towards the previous state $H[t - 1]$ at a rate governed by $\\beta$ . To simplify notation, we represent the spiking neuron layer as $\\mathcal{SN}(\\cdot)$ , where the input is the membrane potential tensor $U$ , and the output is the spike tensor $S$ .", + "bbox": [ + 89, + 454, + 483, + 575 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Vanilla Self-Attention Mechanism (VSA)", + "text_level": 1, + "bbox": [ + 89, + 583, + 439, + 599 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The Vanilla Self-Attention (VSA) mechanism enables a model to focus on different parts of the input sequence while constructing its output representations. The self-attention operation can be mathematically expressed as:", + "bbox": [ + 89, + 604, + 483, + 666 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {V S A} (Q, K, V) = \\operatorname {s o f t m a x} \\left(\\frac {Q K ^ {T}}{\\sqrt {d _ {k}}}\\right) V \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 153, + 675, + 483, + 710 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $Q$ , $K$ , and $V$ represent the query, key, and value matrices, respectively, and $d_k$ is the dimension of the key vectors. The softmax function is applied to normalize the attention scores, ensuring they are non-negative. The resulting weighted sum of the values reflects the importance of each value, based on the similarity between the query and its corresponding key.", + "bbox": [ + 89, + 718, + 483, + 824 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The input to the self-attention mechanism consists of a sequence of embeddings $x_{1}, x_{2}, \\ldots, x_{n}$ , which are linearly projected into $Q, K$ , and $V$ using learned weight matrices:", + "bbox": [ + 89, + 824, + 483, + 869 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nQ = X W _ {Q}, \\quad K = X W _ {K}, \\quad V = X W _ {V}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 140, + 880, + 483, + 897 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $W_{Q}$ , $W_{K}$ , and $W_{V}$ are the learned weight matrices corresponding to the $Q$ , $K$ , and $V$ , respectively.", + "bbox": [ + 511, + 90, + 903, + 122 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "However, the standard VSA is not directly compatible with SNNs due to the conflict between the floating-point operations required for matrix multiplication of $Q$ , $K$ , and $V$ , and the energy efficiency of SNN computations. What's more, the softmax operation, which involves exponentiation and division, does not align with the computation paradigm of SNNs too.", + "bbox": [ + 511, + 123, + 905, + 229 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Vanilla Spiking Self-Attention Mechanism (VSSA)", + "text_level": 1, + "bbox": [ + 511, + 247, + 905, + 277 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To address the incompatibility between the VSA and SNNs, the Vanilla Spiking Self-Attention (VSSA) mechanism is proposed in [64, 65], which is more suitable for SNNs, as shown in Figure 1(b). The $Q$ , $K$ , and $V$ are initially derived through learnable weight matrices. These are then transformed into spiking sequences by applying binary spiking neuron layers:", + "bbox": [ + 511, + 286, + 906, + 393 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nQ = \\mathcal {S N} _ {Q} (\\mathrm {B N} (X W _ {Q})), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 620, + 410, + 906, + 426 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nK = \\mathcal {S N} _ {K} (\\mathrm {B N} (X W _ {K})), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 620, + 429, + 906, + 445 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nV = \\mathcal {S N} _ {V} (\\mathrm {B N} (X W _ {V})). \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 624, + 448, + 906, + 464 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this approach, the computation of the attention matrix is performed using binary spike-form queries and keys (which contain only 0s and 1s), thus replacing traditional matrix multiplication with addition-based operations.", + "bbox": [ + 511, + 481, + 905, + 542 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To further address the challenges of large values resulting from the matrix multiplication, a scaling factor $s$ is introduced to regulate the magnitude of the result. The spike-efficient VSSA is formulated as follows:", + "bbox": [ + 511, + 544, + 906, + 604 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {V S S A} (Q, K, V) = \\mathcal {S N} \\left(Q K ^ {T} V \\cdot s\\right). \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 575, + 621, + 906, + 640 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this formulation, the scaling factor $s$ adjusts the magnitude of the result from the matrix product, while all other operations in VSSA—such as the attention calculation—are performed using addition, in line with the spike-based computation model.", + "bbox": [ + 511, + 656, + 906, + 731 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "However, simply converting the queries, keys, and values into spike form prior to performing matrix multiplication does not fully leverage the advantages of SNNs, which thrive on addition-only and event-driven computations. It is unnecessary to convert all of $Q$ , $K$ , and $V$ into binary spike sequences to transition from matrix multiplication to addition. Furthermore, converting all of the values to binary spikes can result in significant information loss. Thus, the challenge remains to strike a balance between maintaining the richness of the data while taking full advantage of the addition operation-based nature of SNNs.", + "bbox": [ + 511, + 734, + 908, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "24401", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/2cfcd6b4c360199ad8437603374939eff8452fe837c2b75f5c0449d0acc9189a.jpg", + "image_caption": [ + "Figure 2. The overview of Spiking Transformer." + ], + "image_footnote": [], + "bbox": [ + 98, + 88, + 898, + 383 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Method", + "text_level": 1, + "bbox": [ + 89, + 431, + 181, + 448 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Overall Architecture", + "text_level": 1, + "bbox": [ + 89, + 457, + 287, + 472 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Figure 2 illustrates the architecture of our Spiking Transformer, which comprises four primary components: Spiking Patch Splitting (SPS), Accurate Addition-Only Spiking Self-Attention $(\\mathrm{A}^2\\mathrm{OS}^2\\mathrm{A})$ , Multi-Layer Perceptron (MLP), and a linear classification head. The design follows the approach outlined in [58]. For the SPS module, we adopt the structure from [63, 64]. Given a 2D image sequence $I \\in \\mathbb{R}^{T \\times C \\times H \\times W}$ , the Patch Splitting Module (PSM) performs a linear projection and splits the input into a sequence of $N$ flattened spike patches with $D$ -dimensional channels, where $T$ denotes the number of timesteps (images are repeated $T$ times in the dataset), $C$ is the number of channels, and $H$ and $W$ represent the height and width of the image sequence. Additionally, a convolutional layer generates Relative Position Embeddings (RPE) as described in [58]. Together, the SPS part is written as:", + "bbox": [ + 88, + 479, + 485, + 720 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nu = \\operatorname {P S M} (I), I \\in \\mathbb {R} ^ {T \\times C \\times H \\times W}, x \\in \\mathbb {R} ^ {T \\times N \\times D}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 109, + 729, + 483, + 748 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\ns = \\mathcal {S N} (u), s \\in \\mathbb {R} ^ {T \\times N \\times D}, \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 191, + 767, + 483, + 785 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {R P E} = \\mathrm {B N} ((\\operatorname {C o n v 2 d} (s))), \\mathrm {R P E} \\in \\mathbb {R} ^ {T \\times N \\times D}, \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 114, + 804, + 483, + 824 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nU _ {0} = u + \\operatorname {R P E}, U _ {0} \\in \\mathbb {R} ^ {T \\times N \\times D}, \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 174, + 842, + 483, + 861 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $u$ and $U_{0}$ are the output membrane potential tensors from the PSM and SPS modules, respectively, and $\\mathcal{SN}(\\cdot)$", + "bbox": [ + 89, + 869, + 483, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "represents the spike neuron layer. The output $U_{0}$ is then passed to the $L$ -block Spiking Transformer encoder, which consists of the $\\mathrm{A}^{2}\\mathrm{OS}^{2}\\mathrm{A}$ and MLP blocks. Residual connections are applied to the membrane potentials in both the $\\mathrm{A}^{2}\\mathrm{OS}^{2}\\mathrm{A}$ and MLP blocks. To avoid multi-bit spike outputs, we adopt the design from [58, 63], ensuring that the values before the convolutional layers are binary, thus allowing the spike and weight matrix multiplication to be simplified to addition operations.", + "bbox": [ + 511, + 434, + 906, + 569 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The $\\mathrm{A}^2\\mathrm{OS}^2\\mathrm{A}$ mechanism models both local and global information of the input sequence using binary spike $Q$ , full-precision $K$ , and ternary spike $V$ . This approach reduces information loss and eliminates the need for scaling and softmax operations, thereby improving task accuracy. A Global Average Pooling (GAP) is applied to the processed feature from the Spiking Transformer encoder, and the resulting $D$ -dimensional channel is passed to a fully connected Classification Head (CH) for final prediction. The $\\mathrm{A}^2\\mathrm{OS}^2\\mathrm{A}$ , MLP, and CH components are as follows:", + "bbox": [ + 511, + 570, + 908, + 720 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nS _ {0} = \\mathcal {S N} (U _ {0}), S _ {0} \\in \\mathbb {R} ^ {T \\times N \\times D}, \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 599, + 726, + 906, + 744 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nU _ {l} ^ {\\prime} = \\mathrm {A} ^ {2} \\mathrm {O S} ^ {2} \\mathrm {A} \\left(S _ {l - 1}\\right) + U _ {l - 1}, U _ {l} ^ {\\prime} \\in \\mathbb {R} ^ {T \\times N \\times D}, l = 1 \\dots L, \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 517, + 758, + 906, + 790 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nS _ {l} ^ {\\prime} = \\mathcal {S N} \\left(U _ {l} ^ {\\prime}\\right), S _ {l} ^ {\\prime} \\in \\mathbb {R} ^ {T \\times N \\times D}, \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 602, + 805, + 906, + 824 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nS _ {l} = \\mathcal {S N} \\left(\\operatorname {M L P} \\left(S _ {l} ^ {\\prime}\\right) + U _ {l} ^ {\\prime}\\right), S _ {l} \\in \\mathbb {R} ^ {T \\times N \\times D}, l = 1 \\dots L, \\tag {17}\n$$\n", + "text_format": "latex", + "bbox": [ + 527, + 837, + 906, + 869 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nY = \\mathrm {C H} (\\mathrm {G A P} (S _ {L})). \\tag {18}\n$$\n", + "text_format": "latex", + "bbox": [ + 633, + 883, + 906, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "24402", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Information Loss in Spiking Transformers", + "text_level": 1, + "bbox": [ + 89, + 90, + 452, + 107 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we analyze the limitations of binary spikes in SNNs with respect to information representation, which motivates the focus of this paper. While binary spikes offer energy efficiency, they inherently suffer from reduced representational capacity, leading to information loss when used for all $Q$ , $K$ , and $V$ in VSSA, which results in accuracy degradation.", + "bbox": [ + 89, + 112, + 483, + 217 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To support this claim, we first perform a theoretical analysis using the concept of entropy. The representational capability $\\mathcal{C}(\\mathbf{X})$ of a set $\\mathbf{X}$ is determined by the maximum entropy of $\\mathbf{X}$ , expressed as:", + "bbox": [ + 89, + 218, + 483, + 279 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {C} (\\mathbf {X}) = \\max \\mathcal {H} (\\mathbf {X}) = - \\sum_ {x \\in \\mathbf {X}} p _ {\\mathbf {X}} (x) \\log p _ {\\mathbf {X}} (x), \\tag {19}\n$$\n", + "text_format": "latex", + "bbox": [ + 112, + 285, + 483, + 316 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $p_{\\mathbf{X}}(x)$ is the probability of a sample $x$ from $\\mathbf{X}$ . We now present the following proposition:", + "bbox": [ + 89, + 324, + 483, + 356 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Proposition 1 For a set $\\mathbf{X}$ , its representational capacity is $\\mathcal{C}(\\mathbf{X}) = \\max \\mathcal{H}(\\mathbf{X})$ . When the probability distribution of $\\mathbf{X}$ is uniform, i.e., $p_{\\mathbf{X}}(x) = \\frac{1}{N}$ , where $N$ is the total number of samples in $\\mathbf{X}$ , the entropy $\\mathcal{H}(\\mathbf{X})$ reaches its maximum value of $\\log(N)$ . Hence, we conclude that $\\mathcal{C}(\\mathbf{X}) = \\log(N)$ .", + "bbox": [ + 89, + 363, + 483, + 441 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Using Proposition 1, we can evaluate the representational capacity of binary spike layers in SNNs and compare them with real-valued layers in ANNs. Let $\\mathbf{F}_{\\mathrm{B}} \\in \\mathbb{B}^{C \\times H \\times W}$ denote the binary spike layer of the SNN, and $\\mathbf{F}_{\\mathrm{R}} \\in \\mathbb{R}^{C \\times H \\times W}$ denote the real-valued feature map in the corresponding ANN. For a binary spike output $s$ , it requires 1 bit, and thus the number of possible samples from $s$ is 2. Therefore, the number of samples for $\\mathbf{F}_{\\mathrm{B}}$ is $2^{(C \\times H \\times W)}$ , leading to:", + "bbox": [ + 89, + 446, + 483, + 569 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {C} \\left(\\mathbf {F} _ {\\mathrm {B}}\\right) = \\log \\left(2 ^ {(C \\times H \\times W)}\\right) = C \\times H \\times W. \\tag {20}\n$$\n", + "text_format": "latex", + "bbox": [ + 122, + 575, + 483, + 601 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In contrast, a real-valued output requires 32 bits per sample, yielding $2^{32}$ possible values. Hence, the representational capacity for real-valued layers is:", + "bbox": [ + 89, + 606, + 483, + 652 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {C} \\left(\\mathbf {F} _ {\\mathrm {R}}\\right) = \\log \\left(2 ^ {3 2 \\times (C \\times H \\times W)}\\right) = 3 2 \\times C \\times H \\times W. \\tag {21}\n$$\n", + "text_format": "latex", + "bbox": [ + 96, + 657, + 483, + 684 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "This highlights the limited representational capacity of the binary spike layer. By transforming the real-valued $Q_{\\mathbb{R}}$ , $K_{\\mathbb{R}}$ , $V_{\\mathbb{R}}$ into binary $Q_{\\mathbb{B}}$ , $K_{\\mathbb{B}}$ , $V_{\\mathbb{B}}$ , significant information loss occurs, reducing accuracy. Therefore, we propose increasing the bit precision of $Q$ , $K$ , and $V$ in spiking attention to minimize information loss, while preserving the advantages of multiplication-addition transformations.", + "bbox": [ + 89, + 689, + 483, + 796 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Accurate Addition-Only Spiking Self-Attention Mechanism (A $^2$ OS $^2$ A)", + "text_level": 1, + "bbox": [ + 89, + 803, + 483, + 834 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To maintain the integrity of information within spiking attention while leveraging the advantages of multiplication-addition transformations, we propose the Accurate Addition-Only Spiking Self-Attention $(\\mathrm{A}^2\\mathrm{OS}^2\\mathrm{A})$ . This mechanism", + "bbox": [ + 89, + 839, + 485, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "employs binary $Q$ , full-precision $K$ , and ternary $V$ in the self-attention process. Specifically, we utilize a binary spiking neuron after the linear layer to produce $Q$ , a ReLU function for generating $K$ , and a ternary spiking neuron for producing $V$ , as illustrated in Figure 1(c). The formulations for $Q, K, V$ in $\\mathrm{A}^2\\mathrm{OS}^2\\mathrm{A}$ are defined as follows:", + "bbox": [ + 511, + 90, + 906, + 181 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nQ = \\mathcal {S N} _ {Q} ^ {b} (\\mathrm {B N} (X W _ {Q})), \\tag {22}\n$$\n", + "text_format": "latex", + "bbox": [ + 614, + 189, + 906, + 208 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nK = \\operatorname {R e L U} _ {K} (\\operatorname {B N} (X W _ {K})), \\tag {23}\n$$\n", + "text_format": "latex", + "bbox": [ + 614, + 210, + 906, + 228 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nV = \\mathcal {S N} _ {V} ^ {t} (\\mathrm {B N} (X W _ {V})), \\tag {24}\n$$\n", + "text_format": "latex", + "bbox": [ + 616, + 231, + 906, + 250 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathcal{SN}^b (\\cdot)$ denotes the binary spiking neuron as used in standard SNNs, $\\mathrm{ReLU}_K(\\cdot)$ is the ReLU activation function typical in ANNs, and $\\mathcal{SN}^t (\\cdot)$ represents a ternary spiking neuron that outputs values in the set $\\{-1,0,1\\}$ .", + "bbox": [ + 511, + 257, + 905, + 318 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The evolution of the membrane potential and the spike generation mechanism within $\\mathcal{SN}^t(\\cdot)$ can be described by the following equations:", + "bbox": [ + 511, + 319, + 905, + 364 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nU [ t ] = H [ t - 1 ] + X [ t ], \\tag {25}\n$$\n", + "text_format": "latex", + "bbox": [ + 581, + 373, + 906, + 390 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nS [ t ] = \\operatorname {H e a} (| U [ t ] | - V _ {\\mathrm {t h}}), \\tag {26}\n$$\n", + "text_format": "latex", + "bbox": [ + 584, + 392, + 906, + 409 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nH [ t ] = V _ {\\text {r e s e t}} S [ t ] + \\beta U [ t ] (1 - | S [ t ] |). \\tag {27}\n$$\n", + "text_format": "latex", + "bbox": [ + 581, + 411, + 906, + 428 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The formulation of the $\\mathrm{A}^2\\mathrm{OS}^2\\mathrm{A}$ is expressed as:", + "bbox": [ + 511, + 436, + 834, + 450 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {A} ^ {2} \\mathrm {O S} ^ {2} \\mathrm {A} (Q, K, V) = \\mathcal {S N} \\left(Q \\cdot K ^ {T} \\cdot V\\right). \\tag {28}\n$$\n", + "text_format": "latex", + "bbox": [ + 568, + 459, + 906, + 478 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In $\\mathrm{A}^2\\mathrm{OS}^2\\mathrm{A}$ , since $Q$ takes values in $\\{0,1\\}$ and $K$ is in $\\mathbb{R}$ , the matrix multiplication $Q\\cdot K^{T}$ can be transformed into equivalent addition operations, resulting in $Q\\cdot K^{T}\\in \\mathbb{R}$ . Moreover, because both $Q$ and $K$ are naturally non-negative, this preserves a non-negative attention map $Q\\cdot K^T$ . Consequently, the mechanism eliminates the need for a softmax operation to enforce non-negativity within the attention map. Furthermore, since $K$ is a real-valued vector without boundaries, there is no necessity for a scaling factor $s$ to manage large values resulting from the matrix multiplication, unlike the approaches in [64, 65]. Additionally, given that $V$ can also take values in $\\{-1,0,1\\}$ , the matrix product $Q\\cdot K^{T}\\cdot V$ can similarly be expressed through addition operations. At the same time, unlike other matrix multiplication methods outlined in [64, 65], our framework allows for the result to include negative values, akin to what is seen in traditional ANNs.", + "bbox": [ + 511, + 484, + 908, + 739 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In summary, the $\\mathrm{A}^2\\mathrm{OS}^2\\mathrm{A}$ mechanism utilizes full-precision $K$ and ternary spike $V$ to enhance the representational capability of self-attention and minimize information loss. Additionally, all attention calculations are performed using addition operations, which align with the characteristics of SNNs.", + "bbox": [ + 511, + 741, + 908, + 832 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiment", + "text_level": 1, + "bbox": [ + 511, + 844, + 638, + 862 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We evaluate our method on various datasets, including CIFAR-10/100 [31] and ImageNet-1K [8]. The network", + "bbox": [ + 511, + 869, + 906, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "24403", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/7b2922acaeb01d1eb432e88162c8c5d8ca7a4647cc5ee30b928f217683d8e7b5.jpg", + "table_caption": [ + "Table 1. Ablation study for Spiking Transformer on CIFAR10/100. Param refers to the number of parameters. Spiking Transformer- $L - D$ is a Spiking Transformer model with $L$ Spiking Transformer encoder blocks and $D$ feature embedding dimensions." + ], + "table_footnote": [], + "table_body": "
MethodsParam (M)Time StepCIFAR10 AccCIFAR100 Acc
Baseline-2-2562.59494.3976.00
Spiking Transformer-2-2562.59494.9176.96
Baseline-2-51210.23495.5178.83
Spiking Transformer-2-51210.23496.4279.90
", + "bbox": [ + 91, + 169, + 490, + 258 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "architecture and experimental setup follow the baseline from Spike-driven Transformer [58]. Further details regarding the experimental settings can be found in the Spike-driven Transformer [58].", + "bbox": [ + 89, + 295, + 483, + 357 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.1. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 364, + 243, + 380 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We conduct a series of ablation experiments to assess the effectiveness of the proposed Spiking Transformer, comparing it with the baseline model from [58] on the CIFAR-10 and CIFAR-100 datasets. The results are summarized in Table 1.", + "bbox": [ + 89, + 386, + 483, + 446 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The baseline models used are Spike-driven Transformer-2-256 and Spike-driven Transformer-2-512. For the Transformer-2-256 with 4 timesteps, the baseline accuracy is $94.39\\%$ on CIFAR-10 and $76.00\\%$ on CIFAR-100, consistent with previous results. Our Spiking Transformer, using the same architecture, achieves notable improvements, with accuracies of $94.91\\%$ and $76.96\\%$ on CIFAR-10 and CIFAR-100, respectively—resulting in performance gains of approximately $0.5\\%$ and $1.0\\%$ . For the larger model, Transformer-2-512 with 4 timesteps achieves baseline accuracies of $95.51\\%$ and $78.43\\%$ on CIFAR-10 and CIFAR-100. Our model with the same architecture shows substantial improvements, reaching accuracy scores of $96.42\\%$ and $79.90\\%$ , corresponding to gains of approximately $0.9\\%$ and $1.1\\%$ on CIFAR-10 and CIFAR-100, respectively.", + "bbox": [ + 89, + 446, + 485, + 674 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. CIFAR", + "text_level": 1, + "bbox": [ + 89, + 681, + 184, + 696 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To further evaluate the performance of our method, we compare it with recent SNN-based Transformer approaches, including Spikformer [64], Spikingformer [63] and so on, using a range of models with varying embedding dimensions and numbers of transformer blocks on the CIFAR dataset. The CIFAR dataset consists of 50,000 training images and 10,000 test images, all with a resolution of $32 \\times 32$ . We maintain the experimental setup from [58, 63, 64], including network architecture, training configurations, etc. Table 2 presents the accuracy of our method compared to other models on the CIFAR dataset. As shown in Table 2, our Spiking Transformer-2-384 achieves an accuracy of $95.70\\%$ on the CIFAR-10 dataset, surpassing TET $(94.44\\%)$ and even", + "bbox": [ + 89, + 704, + 485, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/9266591a827ac6208d6d971c4037cf5a1e8af6006766d77204cdbf1c1244f458.jpg", + "table_caption": [ + "Table 2. Comparison of the performance between Spiking Transformer and existing approaches on CIFAR10/100." + ], + "table_footnote": [], + "table_body": "
MethodsParam (M)Time StepCIFAR10 AccCIFAR100 Acc
Hybrid training[46]9.2712592.2267.87
Diet-SNN[45]0.2710/592.5464.07
STBP[55]17.541289.83-
STBP NeuNorm[54]17.541290.53-
TSSL-BP[60]17.54591.41-
STBP-tdBN[62]12.63492.9270.86
TET[9]12.63494.4474.47
Spikformer-4-256[64]4.15493.9475.96
Spikformer-2-384[64]5.76494.8076.95
Spikformer-4-384[64]9.32495.1977.86
Spikingformer-4-256[63]4.15494.7777.43
Spikingformer-2-384[63]5.76495.2278.34
Spikingformer-4-384[63]9.32495.6179.09
Spiking Transformer-4-2564.15494.9677.49
Spiking Transformer-2-3845.76495.7078.59
Spiking Transformer-4-3849.32496.3279.69
Spiking Transformer-2-51210.23496.4279.90
", + "bbox": [ + 514, + 142, + 915, + 392 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Spikingformer-4-384 (94.97%). Notably, performance improves further with higher embedding dimensions and more transformer blocks. In particular, Spiking Transformer-4-384 shows a 1.13% improvement over Spikformer-4-384 and a 0.71% improvement over Spikingformer-4-384. The performance gain of our Spiking Transformer is even more pronounced on more complex datasets such as CIFAR-100. Specifically, Spiking Transformer-4-384 achieves a notable improvement of 1.83% over Spikformer-4-384.", + "bbox": [ + 511, + 430, + 908, + 566 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. ImageNet", + "text_level": 1, + "bbox": [ + 511, + 575, + 625, + 590 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We further evaluated our approach on the challenging ImageNet-1K dataset. ImageNet-1K [8] is a widely-used benchmark for image classification, consisting of 1,000 categories. The dataset includes approximately 1.28 million training images and 50,000 test images. For both training and evaluation, the images are resized to a default resolution of $224 \\times 224$ pixels. To ensure consistency with previous work, we maintain the experimental setup from [58, 63, 64], including network architecture, training configurations, and other experimental details.", + "bbox": [ + 511, + 598, + 908, + 748 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3 presents a comparative analysis of recent SoTA methods evaluated on the ImageNet dataset. Notable models in this domain include Spikformer [64], Spikingformer [63], and Spike-driven Transformer [58], which achieve accuracies of $73.38\\%$ , $74.79\\%$ , and $74.57\\%$ , respectively, with Transformer-8-512 as the baseline configuration. In comparison, our approach demonstrates a significant improvement, achieving an accuracy of $76.28\\%$ , which represents a $1.49\\%$ increase over the best-performing transformer-based SNN model, Spikingformer. This improvement highlights the", + "bbox": [ + 511, + 750, + 910, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "24404", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/abbd8dac46e774fae487ce791a0b65c1d9dd9f89ef0d95f2bcb01ac94b4a0f50.jpg", + "table_caption": [ + "Table 3. Comparison of the performance between Spiking Transformer and existing approaches on ImageNet-1k." + ], + "table_footnote": [], + "table_body": "
CategoryMethodsArchitectureParam (M)Time StepAcc
ANN-to-SNNHybrid training[46]ResNet-3421.7925061.48
Spiking ResNet[27]ResNet-3421.7935071.61
ResNet-5025.5635072.75
QCFS[2]VGG-16138.426472.85
COS[24]ResNet-3421.79874.17
Directly LearningTET[9]Spiking-ResNet-3421.79664.79
SEW-ResNet-3421.79468.00
STBP-tdBN[62]Spiking-ResNet-3421.79663.72
SEW-ResNet-3421.79467.04
SEW ResNet[11]SEW-ResNet-5025.56467.78
SEW-ResNet-10144.55468.76
SEW-ResNet-15260.19469.26
Attention-SNN[57]ResNet-10478.37477.08
Directly LearningSpikformer[64]Spikformer-8-38416.81470.24
Spikformer-6-51223.37472.46
Spikformer-8-51229.68473.38
Spikformer-10-51236.01473.68
Spikformer-8-76866.34474.81
Directly LearningSpikingformer[63]Spikingformer-8-38416.81472.45
Spikingformer-8-51229.68474.79
Spikingformer-8-76866.34475.85
Directly LearningSpike-driven Transformer[58]Spike-driven Transformer-8-38416.81472.28
Spike-driven Transformer-6-51223.37474.11
Spike-driven Transformer-8-51229.68474.57
Spike-driven Transformer-10-51236.01474.66
Spike-driven Transformer-8-76866.34477.07
Directly LearningSpiking TransformerSpiking Transformer-8-38416.81474.04
Spiking Transformer-6-51223.37476.22
Spiking Transformer-8-51229.68476.28
Spiking Transformer-10-51236.01478.66
", + "bbox": [ + 155, + 126, + 848, + 604 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "effectiveness of our method in leveraging the potential of spiking neuron models within the transformer architecture, resulting in enhanced performance on large-scale datasets like ImageNet-1K.", + "bbox": [ + 89, + 640, + 482, + 700 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Furthermore, even with a smaller network architecture, our Spiking Transformer-10-512 model outperforms the other methods, including those using Transformer-8-768 as the baseline. This demonstrates that our approach is not only more accurate but also more efficient, as it achieves superior results with fewer parameters, highlighting its scalability and robustness.", + "bbox": [ + 89, + 702, + 482, + 808 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 89, + 828, + 209, + 843 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we have introduced the Accurate Addition-Only Spiking Self-Attention $(\\mathrm{A}^2\\mathrm{OS}^2\\mathrm{A})$ mechanism as a novel method for integrating Spiking Neural Networks with", + "bbox": [ + 89, + 854, + 483, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Transformer models. Our approach tackles the inherent challenge of adapting the self-attention mechanism to the spiking paradigm by proposing a hybrid solution that combines binary, ReLU, and ternary spiking neurons. This innovative hybrid strategy enables the elimination of both softmax and scaling operations, which are traditionally essential in Transformer models, while preserving the fundamental advantages of addition-only computations in SNNs.", + "bbox": [ + 511, + 640, + 906, + 762 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Through extensive experiments, we have demonstrated that $\\mathrm{A}^2\\mathrm{OS}^2\\mathrm{A}$ outperforms existing SNN-based Transformer models in terms of both accuracy and efficiency. Specifically, our method achieves a SoTA accuracy of $78.66\\%$ on the ImageNet-1K dataset. This performance, combined with the energy-efficient nature of SNNs, makes $\\mathrm{A}^2\\mathrm{OS}^2\\mathrm{A}$ a promising candidate for real-world applications, especially those requiring high computational performance with low energy consumption.", + "bbox": [ + 511, + 763, + 908, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "24405", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgment", + "text_level": 1, + "bbox": [ + 91, + 90, + 240, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work is supported by grants from the National Natural Science Foundation of China under contracts No.12202412 and No.12202413.", + "bbox": [ + 89, + 113, + 485, + 156 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 184, + 187, + 200 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Filipp Akopyan, Jun Sawada, Andrew Cassidy, Rodrigo Alvarez-Icaza, John Arthur, Paul Merolla, Nabil Imam, Yutaka Nakamura, Pallab Datta, Gi-Joon Nam, et al. Truenorth: Design and tool flow of a 65 mw 1 million neuron programmable neurosynaptic chip. IEEE transactions on computer-aided design of integrated circuits and systems, 34(10):1537-1557, 2015. 1", + "[2] Tong Bu, Wei Fang, Jianhao Ding, PengLin Dai, Zhaofei Yu, and Tiejun Huang. Optimal ann-snn conversion for high-accuracy and ultra-low-latency spiking neural networks. In International Conference on Learning Representations, 2021. 2, 8", + "[3] Tong Bu, Jianhao Ding, Zhaofei Yu, and Tiejun Huang. Optimized potential initialization for low-latency spiking neural networks. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 11-20, 2022. 2", + "[4] Natalia Caporale and Yang Dan. Spike timing-dependent plasticity: a hebbian learning rule. Annu. Rev. Neurosci., 31 (1):25-46, 2008. 2", + "[5] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 3", + "[6] Chun-Fu Richard Chen, Quanfu Fan, and Rameswar Panda. Crossvit: Cross-attention multi-scale vision transformer for image classification. In Proceedings of the IEEE/CVF international conference on computer vision, pages 357–366, 2021. 3", + "[7] Mike Davies, Narayan Srinivasa, Tsung-Han Lin, Gautham Chinya, Yongqiang Cao, Sri Harsha Choday, Georgios Dimou, Prasad Joshi, Nabil Imam, Shweta Jain, et al. Loihi: A neuromorphic manycore processor with on-chip learning. *IEEE Micro*, 38(1):82-99, 2018. 1", + "[8] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 6, 7", + "[9] Shikuang Deng, Yuhang Li, Shanghang Zhang, and Shi Gu. Temporal efficient training of spiking neural network via gradient re-weighting. arXiv preprint arXiv:2202.11946, 2022.7, 8", + "[10] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2", + "[11] Wei Fang, Zhaofei Yu, Yanqi Chen, Tiejun Huang, Timothee Masquelier, and Yonghong Tian. Deep residual learning in spiking neural networks. Advances in Neural Information Processing Systems, 34:21056-21069, 2021. 2, 8" + ], + "bbox": [ + 93, + 209, + 485, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[12] Yufei Guo, Weihang Peng, Xiaode Liu, Yuanpei Chen, Yuhan Zhang, Xin Tong, Zhou Jie, and Zhe Ma. Enof-snn: Training accurate spiking neural networks via enhancing the output feature. In The Thirty-eighth Annual Conference on Neural Information Processing Systems. 2", + "[13] Yufei Guo, Yuanpei Chen, Liwen Zhang, Xiaode Liu, Yinglei Wang, Xuhui Huang, and Zhe Ma. IM-loss: Information maximization loss for spiking neural networks. In Advances in Neural Information Processing Systems, 2022.", + "[14] Yufei Guo, Yuanpei Chen, Liwen Zhang, YingLei Wang, Xiaode Liu, Xinyi Tong, Yuanyuan Ou, Xuhui Huang, and Zhe Ma. Reducing information loss for spiking neural networks. In Computer Vision - ECCV 2022, pages 36-52, Cham, 2022. Springer Nature Switzerland. 2", + "[15] Yufei Guo, Xinyi Tong, Yuanpei Chen, Liwen Zhang, Xiaode Liu, Zhe Ma, and Xuhui Huang. Recdis-snn: Rectifying membrane potential distribution for directly training spiking neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 326-335, 2022. 1", + "[16] Yufei Guo, Liwen Zhang, Yuanpei Chen, Xinyi Tong, Xiaode Liu, YingLei Wang, Xuhui Huang, and Zhe Ma. Real spike: Learning real-valued spikes for spiking neural networks. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XII, pages 52-68. Springer, 2022. 1", + "[17] Yufei Guo, Xiaode Liu, Yuanpei Chen, Liwen Zhang, Weihang Peng, Yuhan Zhang, Xuhui Huang, and Zhe Ma. Rmploss: Regularizing membrane potential distribution for spiking neural networks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17391-17401, 2023. 2", + "[18] Yufei Guo, Weihang Peng, Yuanpei Chen, Liwen Zhang, Xiaode Liu, Xuhui Huang, and Zhe Ma. Joint a-snn: Joint training of artificial and spiking neural networks via self-distillation and weight factorization. Pattern Recognition, 142:109639, 2023. 2", + "[19] Yufei Guo, Yuhan Zhang, Yuanpei Chen, Weihang Peng, Xiaode Liu, Liwen Zhang, Xuhui Huang, and Zhe Ma. Membrane potential batch normalization for spiking neural networks. arXiv preprint arXiv:2308.08359, 2023.", + "[20] Yufei Guo, Yuanpei Chen, Xiaode Liu, Weihang Peng, Yuhan Zhang, Xuhui Huang, and Zhe Ma. Ternary spike: Learning ternary spikes for spiking neural networks. In Proceedings of the AAAI conference on artificial intelligence, pages 12244-12252, 2024.", + "[21] Yufei Guo, Yuanpei Chen, Zecheng Hao, Weihang Peng, Zhou Jie, Yuhan Zhang, Xiaode Liu, and Zhe Ma. Take a shortcut back: Mitigating the gradient vanishing for training spiking neural networks. Advances in Neural Information Processing Systems, 37:24849-24867, 2025. 2", + "[22] Bing Han and Kaushik Roy. Deep spiking neural network: Energy efficiency through time based coding. In European Conference on Computer Vision, pages 388-404. Springer, 2020. 2", + "[23] Zecheng Hao, Tong Bu, Jianhao Ding, Tiejun Huang, and Zhaofei Yu. Reducing ann-snn conversion error through resid" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "24406", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ual membrane potential. arXiv preprint arXiv:2302.02091, 2023.", + "[24] Zecheng Hao, Jianhao Ding, Tong Bu, Tiejun Huang, and Zhaofei Yu. Bridging the gap between anns and snns by calibrating offset spikes. arXiv preprint arXiv:2302.10685, 2023. 8", + "[25] Nguyen-Dong Ho and Ik-Joon Chang. Tcl: an ann-to-snn conversion with trainable clipping layers. In 2021 58th ACM/IEEE Design Automation Conference (DAC), pages 793-798. IEEE, 2021. 2", + "[26] Allan L Hodgkin and Andrew F Huxley. Currents carried by sodium and potassium ions through the membrane of the giant axon of loligo. The Journal of physiology, 116(4):449, 1952. 2", + "[27] Yangfan Hu, Huajin Tang, and Gang Pan. Spiking deep residual networks. IEEE Transactions on Neural Networks and Learning Systems, 34(8):5200-5205, 2021. 8", + "[28] Asifullah Khan, Zunaira Rauf, Abdul Rehman Khan, Saima Rathore, Saddam Hussain Khan, Najmus Saher Shah, Umair Farooq, Hifsa Asif, Aqsa Asif, Umme Zahoora, et al. A recent survey of vision transformers for medical image segmentation. arXiv preprint arXiv:2312.00634, 2023. 2", + "[29] Seijoon Kim, Seongsik Park, Byunggook Na, and Sungroh Yoon. Spiking-yolo: Spiking neural network for energy-efficient object detection, 2019. 1", + "[30] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 1", + "[31] Alex Krizhevsky, Vinod Nair, and Geoffrey Hinton. Cifar-10 (canadian institute for advanced research). URL http://www.cs.toronto.edu/kriz/cifar.html, 5(4):1, 2010. 6", + "[32] Yuxiang Lan, Yachao Zhang, Xu Ma, Yanyun Qu, and Yun Fu. Efficient converted spiking neural network for 3d and 2d classification. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9211-9220, 2023. 2", + "[33] Nathan Leroux, Jan Finkbeiner, and Emre Neftci. Online transformers with spiking neurons for fast prosthetic hand control. In 2023 IEEE Biomedical Circuits and Systems Conference (BioCAS), pages 1-6. IEEE, 2023. 3", + "[34] Yuhang Li, Shikuang Deng, Xin Dong, Ruihao Gong, and Shi Gu. A free lunch from ann: Towards efficient, accurate spiking neural networks calibration. In International Conference on Machine Learning, pages 6316-6325. PMLR, 2021. 1, 2", + "[35] Yuhang Li, Yufei Guo, Shanghang Zhang, Shikuang Deng, Yongqing Hai, and Shi Gu. Differentiable spike: Rethinking gradient-descent for training spiking neural networks. Advances in Neural Information Processing Systems, 34:23426-23439, 2021. 2", + "[36] Yawei Li, Kai Zhang, Jiezhang Cao, Radu Timofte, and Luc Van Gool. Localvit: Bringing locality to vision transformers. arXiv preprint arXiv:2104.05707, 2021. 3", + "[37] Yudong Li, Yunlin Lei, and Xu Yang. Spikeformer: a novel architecture for training high-performance low-latency spiking neural network. arXiv preprint arXiv:2211.10686, 2022.3" + ], + "bbox": [ + 91, + 90, + 483, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[38] Yanghao Li, Hanzi Mao, Ross Girshick, and Kaiming He. Exploring plain vision transformer backbones for object detection. In European conference on computer vision, pages 280-296. Springer, 2022. 2", + "[39] De Ma, Juncheng Shen, Zonghua Gu, Ming Zhang, Xiaolei Zhu, Xiaoqiang Xu, Qi Xu, Yangjing Shen, and Gang Pan. Darwin: A neuromorphic hardware co-processor based on spiking neural networks. Journal of Systems Architecture, 77: 43-51, 2017. 1", + "[40] Wolfgang Maass. Networks of spiking neurons: the third generation of neural network models. Neural networks, 10 (9):1659-1671, 1997. 2, 4", + "[41] ZHI Min and LU Jingfang. A review of vision transformer for image classification. Journal of Zhengzhou University: Engineering Science, 45(4), 2024. 2", + "[42] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4195-4205, 2023. 1", + "[43] Jing Pei, Lei Deng, Sen Song, Mingguo Zhao, Youhui Zhang, Shuang Wu, Guanrui Wang, Zhe Zou, Zhenzhi Wu, Wei He, et al. Towards artificial general intelligence with hybrid tianjin chip architecture. Nature, 572(7767):106-111, 2019. 1", + "[44] Jinye Qu, Zeyu Gao, Tielin Zhang, Yanfeng Lu, Huajin Tang, and Hong Qiao. Spiking neural network for ultra-low-latency and high-accurate object detection, 2023. 1", + "[45] Nitin Rathi and Kaushik Roy. Diet-snn: Direct input encoding with leakage and threshold optimization in deep spiking neural networks. arXiv preprint arXiv:2008.03658, 2020. 7", + "[46] Nitin Rathi, Gopalakrishnan Srinivasan, Priyadarshini Panda, and Kaushik Roy. Enabling deep spiking neural networks with hybrid conversion and spike timing dependent backpropagation. arXiv preprint arXiv:2005.01807, 2020. 7, 8", + "[47] Dayong Ren, Zhe Ma, Yuanpei Chen, Weihang Peng, Xiaode Liu, Yuhan Zhang, and Yufei Guo. Spiking pointnet: Spiking neural networks for point clouds. arXiv preprint arXiv:2310.06232, 2023. 2", + "[48] Abhronil Sengupta, Yuting Ye, Robert Wang, Chiao Liu, and Kaushik Roy. Going deeper in spiking neural networks: Vgg and residual architectures. Frontiers in neuroscience, 13:95, 2019. 2", + "[49] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In International conference on machine learning, pages 10347-10357. PMLR, 2021. 3", + "[50] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 1", + "[51] Shuai Wang, Malu Zhang, Dehao Zhang, Ammar Belatreche, Yichen Xiao, Yu Liang, Yimeng Shan, Qian Sun, Enqi Zhang, and Yang Yang. Spiking vision transformer with saccadic attention. arXiv preprint arXiv:2502.12677, 2025. 3", + "[52] Wenjie Wei, Malu Zhang, Jilin Zhang, Ammar Belatreche, Jibin Wu, Zijing Xu, Xuerui Qiu, Hong Chen, Yang Yang, and Haizhou Li. Event-driven learning for spiking neural networks. arXiv preprint arXiv:2403.00270, 2024.1" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "24407", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[53] Yujie Wu, Lei Deng, Guoqi Li, Jun Zhu, and Luping Shi. Spatio-temporal backpropagation for training high-performance spiking neural networks. Frontiers in neuroscience, 12:331, 2018. 2", + "[54] Yujie Wu, Lei Deng, Guoqi Li, Jun Zhu, Yuan Xie, and Luping Shi. Direct training for spiking neural networks: Faster, larger, better. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1311-1318, 2019. 7", + "[55] Y. et al. Wu. Spatio-temporal learning with spiking neural networks. Nature Communications, 9, 2018. 7", + "[56] Mingqing Xiao, Qingyan Meng, Zongpeng Zhang, Yisen Wang, and Zhouchen Lin. Training feedback spiking neural networks by implicit differentiation on the equilibrium state. Advances in Neural Information Processing Systems, 34:14516-14528, 2021. 1", + "[57] Man Yao, Guangshe Zhao, Hengyu Zhang, Yifan Hu, Lei Deng, Yonghong Tian, Bo Xu, and Guoqi Li. Attention spiking neural networks. IEEE transactions on pattern analysis and machine intelligence, 45(8):9393-9410, 2023. 8", + "[58] Man Yao, Jiakui Hu, Zhaokun Zhou, Li Yuan, Yonghong Tian, Bo Xu, and Guoqi Li. Spike-driven transformer. Advances in neural information processing systems, 36, 2024. 1, 3, 5, 7, 8", + "[59] Malu Zhang, Jiadong Wang, Jibin Wu, Ammar Belatreche, Burin Amornpaisannon, Zhixuan Zhang, Venkata Pavan Kumar Miriyala, Hong Qu, Yansong Chua, Trevor E Carlson, et al. Rectified linear postsynaptic potential function for backpropagation in deep spiking neural networks. IEEE transactions on neural networks and learning systems, 33 (5):1947-1958, 2021. 2", + "[60] Wenrui Zhang and Peng Li. Temporal spike sequence learning via backpropagation for deep spiking neural networks. Advances in Neural Information Processing Systems, 33:12022-12033, 2020. 7", + "[61] Yuhan Zhang, Xiaode Liu, Yuanpei Chen, Weihang Peng, Yufei Guo, Xuhui Huang, and Zhe Ma. Enhancing representation of spiking neural networks via similarity-sensitive contrastive learning. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 16926-16934, 2024. 2", + "[62] Hanle Zheng, Yujie Wu, Lei Deng, Yifan Hu, and Guoqi Li. Going deeper with directly-trained larger spiking neural networks. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 11062-11070, 2021. 7, 8", + "[63] Chenlin Zhou, Liutao Yu, Zhaokun Zhou, Zhengyu Ma, Han Zhang, Huihui Zhou, and Yonghong Tian. Spikingformer: Spike-driven residual learning for transformer-based spiking neural network. arXiv preprint arXiv:2304.11954, 2023. 1, 3, 5, 7, 8", + "[64] Zhaokun Zhou, Yuesheng Zhu, Chao He, Yaowei Wang, Shuicheng Yan, Yonghong Tian, and Li Yuan. Spikformer: When spiking neural network meets transformer. arXiv preprint arXiv:2209.15425, 2022. 1, 3, 4, 5, 6, 7, 8", + "[65] Zhaokun Zhou, Kaiwei Che, Wei Fang, Keyu Tian, Yuesheng Zhu, Shuicheng Yan, Yonghong Tian, and Li Yuan. Spikformer v2: Join the high accuracy club on imagenet with an snn ticket. arXiv preprint arXiv:2401.02020, 2024. 1, 3, 4, 6", + "[66] Shihao Zou, Yuxuan Mu, Xinxin Zuo, Sen Wang, and Li Cheng. Event-based human pose tracking by spiking spatiotemporal transformer, 2023. 1" + ], + "bbox": [ + 91, + 92, + 485, + 900 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "24408", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Spiking Transformer_ Introducing Accurate Addition-Only Spiking Self-Attention for Transformer/24b0ab2b-a4ff-49fc-865c-b8d7799e49bd_model.json b/2025/Spiking Transformer_ Introducing Accurate Addition-Only Spiking Self-Attention for Transformer/24b0ab2b-a4ff-49fc-865c-b8d7799e49bd_model.json new file mode 100644 index 0000000000000000000000000000000000000000..40fd3088931a4edc15c383c9ee3a69f66b354577 --- /dev/null +++ b/2025/Spiking Transformer_ Introducing Accurate Addition-Only Spiking Self-Attention for Transformer/24b0ab2b-a4ff-49fc-865c-b8d7799e49bd_model.json @@ -0,0 +1,2400 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.131, + 0.885, + 0.177 + ], + "angle": 0, + "content": "Spiking Transformer: Introducing Accurate Addition-Only Spiking Self-Attention for Transformer" + }, + { + "type": "text", + "bbox": [ + 0.193, + 0.203, + 0.812, + 0.24 + ], + "angle": 0, + "content": "Yufei Guo,\\* Xiaode Liu,\\* Yuanpei Chen, Weihang Peng, Yuhan Zhang, Zhe Ma\\* Intelligent Science & Technology Academy of CASIC, China" + }, + { + "type": "text", + "bbox": [ + 0.266, + 0.243, + 0.726, + 0.256 + ], + "angle": 0, + "content": "yfguo@pku.edu.cn, lxde@pku.edu.cn, mazhe_thu@163.com" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.292, + 0.327, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.325, + 0.485, + 0.702 + ], + "angle": 0, + "content": "Transformers have demonstrated outstanding performance across a wide range of tasks, owing to their self-attention mechanism, but they are highly energy-consuming. Spiking Neural Networks have emerged as a promising energy-efficient alternative to traditional Artificial Neural Networks, leveraging event-driven computation and binary spikes for information transfer. The combination of Transformers' capabilities with the energy efficiency of SNNs offers a compelling opportunity. This paper addresses the challenge of adapting the self-attention mechanism of Transformers to the spiking paradigm by introducing a novel approach: Accurate Addition-Only Spiking Self-Attention (\\(A^2OS^2A\\)). Unlike existing methods that rely solely on binary spiking neurons for all components of the self-attention mechanism, our approach integrates binary, ReLU, and ternary spiking neurons. This hybrid strategy significantly improves accuracy while preserving non-multiplicative computations. Moreover, our method eliminates the need for softmax and scaling operations. Extensive experiments show that the \\(A^2OS^2A\\)-based Spiking Transformer outperforms existing SNN-based Transformers on several datasets, even achieving an accuracy of \\(78.66\\%\\) on ImageNet-1K. Our work represents a significant advancement in SNN-based Transformer models, offering a more accurate and efficient solution for real-world applications." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.733, + 0.222, + 0.749 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.758, + 0.486, + 0.866 + ], + "angle": 0, + "content": "Spiking Neural Networks (SNNs) have gained significant traction as an efficient neural network model, finding applications in diverse areas such as object recognition [15, 16, 34, 56], object detection [29, 44], and pose estimation [66]. These networks utilize binary spike signals for information transfer. A spiking neuron generates a spike, denoted by 1, when its membrane potential surpasses a cer-" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.293, + 0.908, + 0.49 + ], + "angle": 0, + "content": "tain threshold; conversely, it does not spike, represented by 0, when the threshold is not met. This distinctive approach to information processing is notably energy-efficient, as it substitutes the complex multiplications involved in weight and activation interactions with simpler addition operations. Furthermore, SNNs can be effectively implemented using event-driven computations on neuromorphic hardware [1, 7, 39, 43, 52]. In this setup, the computational units are activated solely upon the occurrence of a spike, conserving energy by remaining inactive in the absence of spikes. Research indicates that SNNs can achieve significantly greater energy savings compared to their Artificial Neural Network (ANN) counterparts [1, 7]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.494, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Despite the clear advantages of SNNs in terms of energy efficiency, their practical application is often constrained by limited task accuracy. Meanwhile, Transformers have demonstrated exceptional performance across a wide range of tasks, largely due to their self-attention mechanism [30, 42, 50]. Combining the strengths of Transformers with the energy efficiency of SNNs presents a promising opportunity. However, adapting the self-attention mechanism to SNNs is non-trivial. In the vanilla self-attention framework (VSA), three key components are involved: Query \\((Q)\\), Key \\((K)\\), and Value \\((V)\\). As illustrated in Figure 1(a), the VSA process begins by computing the dot product of the floating-point representations of \\(Q\\) and \\(K\\), resulting in a matrix. This matrix is then normalized using a softmax function, which involves exponential and division operations, to produce an attention map that determines the weighting of \\(V\\). These operations in the VSA are incompatible with the operational principles of SNNs, which aim to minimize multiplication. Consequently, to implement a Transformer architecture in SNNs, it is essential to develop a novel, efficient self-attention mechanism that avoids multiplication. Ongoing research efforts are exploring solutions in this domain. Approaches such as Spikformer [64, 65], Spikingformer [63], and Spike-driven Transformer [58] have focused on transforming \\(Q\\), \\(K\\), and \\(V\\) into spike representations prior to performing matrix operations. This transformation enables the replacement of matrix multiplications with addition, aligning" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.875, + 0.221, + 0.888 + ], + "angle": 0, + "content": "*Equal contribution." + }, + { + "type": "page_footnote", + "bbox": [ + 0.111, + 0.888, + 0.236, + 0.9 + ], + "angle": 0, + "content": "† Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "24398" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.323, + 0.106 + ], + "angle": 0, + "content": "better with the principles of SNNs." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.108, + 0.486, + 0.455 + ], + "angle": 0, + "content": "While previous approaches utilize spike-form \\( Q \\), \\( K \\), and \\( V \\) to avoid multiplications, simply introducing spiking neurons to convert \\( Q \\), \\( K \\), and \\( V \\) into spike form prior to performing matrix multiplications does not fully exploit the advantages of the addition-only operation inherent in SNNs. It is unnecessary to convert all of \\( Q \\), \\( K \\), and \\( V \\) into spike form in order to transition from matrix multiplications to matrix additions. However, this method can lead to significant information loss, as demonstrated in Section 4. To address this issue, we propose the Accurate Addition-Only Spiking Self-Attention (A\\(^2\\)OS\\(^2\\)A). The distinctions between A\\(^2\\)OS\\(^2\\)A and the vanilla Spiking Self-Attention (VSSA) are illustrated in Figure 1(b) and Figure 1(c). Unlike VSSA, which applies the same binary spiking neuron to generate \\( Q \\), \\( K \\), and \\( V \\), our approach employs a binary spiking neuron for \\( Q \\), a ReLU activation for \\( K \\), and a ternary spiking neuron for \\( V \\). This design allows A\\(^2\\)OS\\(^2\\)A to retain the multiplication-addition transformation while reducing information loss. Moreover, it eliminates the need for both the scaling and softmax functions, enabling the output of floating-point values similar to the VSA, whereas VSSA produces only integer outputs. The key contributions of this paper are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.459, + 0.483, + 0.548 + ], + "angle": 0, + "content": "- We present a theoretical framework for understanding the information loss associated with vanilla Spiking Self-Attention mechanisms. To the best of our knowledge, this is the first detailed examination of this issue within the context of Spiking Self-Attention, which opens the door for future advancements in SNN-based Transformers." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.55, + 0.483, + 0.655 + ], + "angle": 0, + "content": "- We introduce \\(\\mathrm{A}^2\\mathrm{OS}^2\\mathrm{A}\\), an innovative and efficient approach that incorporates binary spiking neurons, ReLU activations, and ternary spiking neurons, moving beyond the traditional binary-only spiking neuron model. This method reduces information loss while retaining the advantages of addition-only processing in SNNs. Additionally, it eliminates the need for scaling and softmax functions." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.657, + 0.483, + 0.745 + ], + "angle": 0, + "content": "- Comprehensive experiments demonstrate that our proposed architecture either surpasses or matches the performance of State-of-the-Art (SoTA) SNN-based Transformer models on various datasets. Notably, we achieved an accuracy of \\(78.66\\%\\) on ImageNet-1K, setting a new benchmark in the SNN domain." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.459, + 0.483, + 0.745 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.762, + 0.233, + 0.779 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.788, + 0.322, + 0.804 + ], + "angle": 0, + "content": "2.1. Spiking Neural Networks" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.486, + 0.902 + ], + "angle": 0, + "content": "SNNs are a class of neural networks that more closely emulate the behavior of biological neurons compared to ANNs. The primary distinction lies in their use of discrete spikes for information transmission, as opposed to the continuous signals employed by ANNs. Foundational work in this area can be traced back to Hodgkin and Huxley [26], who" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.908, + 0.228 + ], + "angle": 0, + "content": "developed models to simulate the action potentials of biological neurons. Since then, numerous frameworks and models have been introduced to enhance the capabilities and learning efficiency of SNNs. A notable example is the Leaky Integrate-and-Fire (LIF) model [40], which provides a simple yet effective means of simulating neuronal behavior. More recently, the incorporation of Spike-Timing-Dependent Plasticity (STDP) as a learning rule has enabled SNNs to learn temporal patterns effectively [4]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.23, + 0.909, + 0.577 + ], + "angle": 0, + "content": "In the past decade, there has been a significant increase in interest regarding the training of SNNs using techniques adapted from ANNs, as well as their application to practical tasks. Two prevalent learning paradigms in current SNN research are the conversion from ANN to SNN (ANN2SNN) [2, 3, 22-25, 32, 34, 48] and supervised learning [12-14, 18-21, 35, 47, 59]. The ANN-SNN conversion method involves initially training an ANN and subsequently converting it into a homogeneous SNN by transferring the trained weights and substituting ReLU neurons with temporal spiking neurons. However, this method often fails for neuromorphic datasets, as ReLU neurons do not adequately capture the complex temporal dynamics required for processing sequential information. In contrast, supervised learning [11, 17, 53, 61] employs alternative functions during backpropagation to approximate the firing process, enabling the direct training of SNNs as if they were ANNs. This approach leverages the advantages of gradient-based optimization, achieving impressive performance with just a few time steps, even on large-scale datasets. Furthermore, supervised learning has proven to be effective in handling temporal data, establishing it as a preferred methodology in SNN research." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.591, + 0.708, + 0.605 + ], + "angle": 0, + "content": "2.2. Vision Transformers" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.614, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Vision Transformers (ViTs) have recently emerged as a powerful alternative to convolutional neural networks (CNNs) for a wide range of visual tasks, demonstrating remarkable efficacy and flexibility. Introduced by Dosovitskiy et al. [10], ViTs leverage the transformer architecture—originally designed for natural language processing—to process images in a novel manner. The architecture innovatively splits images into patches and treats these patches as sequences, enabling the model to capture long-range dependencies and relationships within the visual data, a limitation often encountered in traditional CNNs. The introduction of ViTs has spurred significant advancements in various computer vision tasks, including image classification [41], object detection [38], and segmentation [28]. These advancements highlight the potential of ViTs to tackle complex visual recognition problems with unprecedented accuracy. Notably, it has been shown that with appropriate training techniques and larger datasets, ViTs can not only match but surpass the performance of state-of-the-art CNNs, even in data-scarce scenarios, presenting a" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24399" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.1, + 0.09, + 0.359, + 0.359 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.361, + 0.089, + 0.637, + 0.358 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.641, + 0.097, + 0.902, + 0.358 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.371, + 0.908, + 0.442 + ], + "angle": 0, + "content": "Figure 1. The difference between our Spiking Self-Attention and the vanilla Spiking Self-Attention. Our Spiking Self-Attention differs significantly from the vanilla Spiking Self-Attention. In the vanilla version, only binary spikes are employed, which can result in considerable information loss. In contrast, our approach integrates a combination of binary spiking neurons, ReLU, and ternary spiking neurons. This hybrid structure effectively reduces information loss, while preserving the advantages of addition-only processing. Moreover, our method eliminates the need for both scaling and the softmax function, streamlining the computation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.468, + 0.441, + 0.484 + ], + "angle": 0, + "content": "significant advantage in real-world applications [49]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.485, + 0.484, + 0.681 + ], + "angle": 0, + "content": "Furthermore, the adaptation of transformers for visual tasks has inspired numerous variations and improvements across the field. Researchers have explored strategies such as integrating convolutional layers into the architecture [6] to retain spatial information more effectively. Other approaches include the use of hierarchical representations [36], which facilitate multi-scale feature extraction, and the incorporation of attention mechanisms tailored specifically for vision tasks [5], improving the model's ability to focus on relevant parts of the input data. These developments underscore the versatility and robustness of the transformer architecture in addressing complex visual challenges, paving the way for further innovations in computer vision." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.696, + 0.458, + 0.712 + ], + "angle": 0, + "content": "2.3. Spiking Neural Networks for Transformers" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.486, + 0.901 + ], + "angle": 0, + "content": "The integration of SNNs into transformer architectures represents an exciting and rapidly developing area of research. This field aims to leverage the distinct advantages of both paradigms: the energy-efficient characteristics of SNNs and the robust contextual representation capabilities of transformers. Recent studies [33, 37, 51] have explored replacing certain neurons within transformers with spiking neurons, marking a significant step towards this integration. While these efforts contribute to mitigating the accuracy loss associated with introducing spiking neurons into transformers, they still face challenges in fully realizing the low energy consumption benefits of SNNs. This is primarily due to" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.468, + 0.907, + 0.529 + ], + "angle": 0, + "content": "their reliance on hybrid computing methodologies, which continue to require traditional Multiply-Accumulate (MAC) operations, such as dot products, softmax calculations, and scaling operations." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.548, + 0.909, + 0.746 + ], + "angle": 0, + "content": "To address this issue, Spikformer [64, 65] proposes converting the \\(Q\\), \\(K\\), and \\(V\\) components into spike form before performing matrix multiplications similar to those used in VSA. This innovative approach allows spike matrix multiplications to be transformed into addition operations, eliminating the need for softmax computations. However, when considering the residual connections in these architectures, Spikformer [64, 65] still incorporates non-spiking computations within the ConvBN layers. To overcome this limitation, Spikingformer [63] and Spike-driven Transformer [58] have restructured the residual connections in standard transformer architectures, placing them before activation functions to ensure that all neurons convey binary spike signals." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.909, + 0.903 + ], + "angle": 0, + "content": "The computation of spike-form \\( Q \\), \\( K \\), and \\( V \\) in these models circumvent multiplications, relying solely on additions. However, simply placing spiking neurons in front of \\( Q \\), \\( K \\), and \\( V \\) to convert them into spike form before performing matrix multiplications does not fully exploit the addition-only advantage of SNNs. In this work, we propose a novel, addition-only spiking self-attention mechanism for transformers, aiming to push the boundaries of what can be achieved through the integration of these two technologies." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "24400" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.09, + 0.218, + 0.108 + ], + "angle": 0, + "content": "3. Preliminary" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.115, + 0.3, + 0.132 + ], + "angle": 0, + "content": "3.1. Spiking Neuron Layer" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.137, + 0.486, + 0.258 + ], + "angle": 0, + "content": "The spiking neuron layer plays a crucial role in integrating both spatial and temporal information, which is then encoded into the membrane potential and subsequently transformed into binary spikes. These spikes drive further computations in subsequent layers of the network. We model the dynamics of the spiking neuron using the LIF model [40]. The evolution of the membrane potential and the spike generation mechanism are described by the following equations:" + }, + { + "type": "equation", + "bbox": [ + 0.164, + 0.268, + 0.484, + 0.286 + ], + "angle": 0, + "content": "\\[\nU [ t ] = H [ t - 1 ] + X [ t ], \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.168, + 0.287, + 0.484, + 0.304 + ], + "angle": 0, + "content": "\\[\nS [ t ] = \\operatorname {H e a} (U [ t ] - V _ {\\mathrm {t h}}), \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.165, + 0.306, + 0.484, + 0.323 + ], + "angle": 0, + "content": "\\[\nH [ t ] = V _ {\\text {r e s e t}} S [ t ] + \\beta U [ t ] (1 - S [ t ]), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.333, + 0.484, + 0.454 + ], + "angle": 0, + "content": "where: \\( X[t] \\) is the spatial input current at time \\( t \\), \\( U[t] \\) is the membrane potential at time \\( t \\), combining the spatial input \\( X[t] \\) and the temporal input \\( H[t - 1] \\) from the previous time step, \\( \\mathrm{Hea}(\\cdot) \\) is the Heaviside step function, which outputs 1 if its argument is non-negative, and 0 otherwise, \\( S[t] \\) is the spike output at time \\( t \\), \\( V_{\\mathrm{th}} \\) is the spike threshold, \\( V_{\\mathrm{reset}} \\) is the reset potential, and \\( \\beta \\) is a decay factor controlling the membrane potential's decay." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.455, + 0.484, + 0.576 + ], + "angle": 0, + "content": "When the membrane potential \\( U[t] \\) exceeds a predefined threshold \\( V_{\\mathrm{th}} \\), the neuron emits a spike, \\( S[t] = 1 \\), and the internal state \\( H[t] \\) is reset to \\( V_{\\mathrm{reset}} \\). If \\( U[t] \\) does not surpass the threshold, the membrane potential decays towards the previous state \\( H[t - 1] \\) at a rate governed by \\( \\beta \\). To simplify notation, we represent the spiking neuron layer as \\( \\mathcal{SN}(\\cdot) \\), where the input is the membrane potential tensor \\( U \\), and the output is the spike tensor \\( S \\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.584, + 0.44, + 0.6 + ], + "angle": 0, + "content": "3.2. Vanilla Self-Attention Mechanism (VSA)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.606, + 0.484, + 0.667 + ], + "angle": 0, + "content": "The Vanilla Self-Attention (VSA) mechanism enables a model to focus on different parts of the input sequence while constructing its output representations. The self-attention operation can be mathematically expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.154, + 0.676, + 0.484, + 0.711 + ], + "angle": 0, + "content": "\\[\n\\operatorname {V S A} (Q, K, V) = \\operatorname {s o f t m a x} \\left(\\frac {Q K ^ {T}}{\\sqrt {d _ {k}}}\\right) V \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.719, + 0.484, + 0.825 + ], + "angle": 0, + "content": "where \\( Q \\), \\( K \\), and \\( V \\) represent the query, key, and value matrices, respectively, and \\( d_k \\) is the dimension of the key vectors. The softmax function is applied to normalize the attention scores, ensuring they are non-negative. The resulting weighted sum of the values reflects the importance of each value, based on the similarity between the query and its corresponding key." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.484, + 0.871 + ], + "angle": 0, + "content": "The input to the self-attention mechanism consists of a sequence of embeddings \\( x_{1}, x_{2}, \\ldots, x_{n} \\), which are linearly projected into \\( Q, K \\), and \\( V \\) using learned weight matrices:" + }, + { + "type": "equation", + "bbox": [ + 0.142, + 0.881, + 0.484, + 0.898 + ], + "angle": 0, + "content": "\\[\nQ = X W _ {Q}, \\quad K = X W _ {K}, \\quad V = X W _ {V}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.123 + ], + "angle": 0, + "content": "where \\(W_{Q}\\), \\(W_{K}\\), and \\(W_{V}\\) are the learned weight matrices corresponding to the \\(Q\\), \\(K\\), and \\(V\\), respectively." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.124, + 0.906, + 0.23 + ], + "angle": 0, + "content": "However, the standard VSA is not directly compatible with SNNs due to the conflict between the floating-point operations required for matrix multiplication of \\( Q \\), \\( K \\), and \\( V \\), and the energy efficiency of SNN computations. What's more, the softmax operation, which involves exponentiation and division, does not align with the computation paradigm of SNNs too." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.248, + 0.906, + 0.278 + ], + "angle": 0, + "content": "3.3. Vanilla Spiking Self-Attention Mechanism (VSSA)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.287, + 0.908, + 0.394 + ], + "angle": 0, + "content": "To address the incompatibility between the VSA and SNNs, the Vanilla Spiking Self-Attention (VSSA) mechanism is proposed in [64, 65], which is more suitable for SNNs, as shown in Figure 1(b). The \\( Q \\), \\( K \\), and \\( V \\) are initially derived through learnable weight matrices. These are then transformed into spiking sequences by applying binary spiking neuron layers:" + }, + { + "type": "equation", + "bbox": [ + 0.622, + 0.411, + 0.907, + 0.428 + ], + "angle": 0, + "content": "\\[\nQ = \\mathcal {S N} _ {Q} (\\mathrm {B N} (X W _ {Q})), \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.622, + 0.43, + 0.907, + 0.446 + ], + "angle": 0, + "content": "\\[\nK = \\mathcal {S N} _ {K} (\\mathrm {B N} (X W _ {K})), \\tag {7}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.625, + 0.449, + 0.907, + 0.465 + ], + "angle": 0, + "content": "\\[\nV = \\mathcal {S N} _ {V} (\\mathrm {B N} (X W _ {V})). \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.482, + 0.906, + 0.543 + ], + "angle": 0, + "content": "In this approach, the computation of the attention matrix is performed using binary spike-form queries and keys (which contain only 0s and 1s), thus replacing traditional matrix multiplication with addition-based operations." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.545, + 0.908, + 0.605 + ], + "angle": 0, + "content": "To further address the challenges of large values resulting from the matrix multiplication, a scaling factor \\( s \\) is introduced to regulate the magnitude of the result. The spike-efficient VSSA is formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.576, + 0.622, + 0.907, + 0.641 + ], + "angle": 0, + "content": "\\[\n\\operatorname {V S S A} (Q, K, V) = \\mathcal {S N} \\left(Q K ^ {T} V \\cdot s\\right). \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.657, + 0.908, + 0.732 + ], + "angle": 0, + "content": "In this formulation, the scaling factor \\( s \\) adjusts the magnitude of the result from the matrix product, while all other operations in VSSA—such as the attention calculation—are performed using addition, in line with the spike-based computation model." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.909, + 0.901 + ], + "angle": 0, + "content": "However, simply converting the queries, keys, and values into spike form prior to performing matrix multiplication does not fully leverage the advantages of SNNs, which thrive on addition-only and event-driven computations. It is unnecessary to convert all of \\( Q \\), \\( K \\), and \\( V \\) into binary spike sequences to transition from matrix multiplication to addition. Furthermore, converting all of the values to binary spikes can result in significant information loss. Thus, the challenge remains to strike a balance between maintaining the richness of the data while taking full advantage of the addition operation-based nature of SNNs." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "24401" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.089, + 0.9, + 0.384 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.354, + 0.394, + 0.644, + 0.409 + ], + "angle": 0, + "content": "Figure 2. The overview of Spiking Transformer." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.433, + 0.183, + 0.449 + ], + "angle": 0, + "content": "4. Method" + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.458, + 0.288, + 0.473 + ], + "angle": 0, + "content": "4.1. Overall Architecture" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.48, + 0.486, + 0.722 + ], + "angle": 0, + "content": "Figure 2 illustrates the architecture of our Spiking Transformer, which comprises four primary components: Spiking Patch Splitting (SPS), Accurate Addition-Only Spiking Self-Attention \\((\\mathrm{A}^2\\mathrm{OS}^2\\mathrm{A})\\), Multi-Layer Perceptron (MLP), and a linear classification head. The design follows the approach outlined in [58]. For the SPS module, we adopt the structure from [63, 64]. Given a 2D image sequence \\(I \\in \\mathbb{R}^{T \\times C \\times H \\times W}\\), the Patch Splitting Module (PSM) performs a linear projection and splits the input into a sequence of \\(N\\) flattened spike patches with \\(D\\)-dimensional channels, where \\(T\\) denotes the number of timesteps (images are repeated \\(T\\) times in the dataset), \\(C\\) is the number of channels, and \\(H\\) and \\(W\\) represent the height and width of the image sequence. Additionally, a convolutional layer generates Relative Position Embeddings (RPE) as described in [58]. Together, the SPS part is written as:" + }, + { + "type": "equation", + "bbox": [ + 0.11, + 0.73, + 0.484, + 0.749 + ], + "angle": 0, + "content": "\\[\nu = \\operatorname {P S M} (I), I \\in \\mathbb {R} ^ {T \\times C \\times H \\times W}, x \\in \\mathbb {R} ^ {T \\times N \\times D}, \\tag {10}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.192, + 0.768, + 0.484, + 0.786 + ], + "angle": 0, + "content": "\\[\ns = \\mathcal {S N} (u), s \\in \\mathbb {R} ^ {T \\times N \\times D}, \\tag {11}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.116, + 0.805, + 0.484, + 0.825 + ], + "angle": 0, + "content": "\\[\n\\mathrm {R P E} = \\mathrm {B N} ((\\operatorname {C o n v 2 d} (s))), \\mathrm {R P E} \\in \\mathbb {R} ^ {T \\times N \\times D}, \\tag {12}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.175, + 0.843, + 0.484, + 0.862 + ], + "angle": 0, + "content": "\\[\nU _ {0} = u + \\operatorname {R P E}, U _ {0} \\in \\mathbb {R} ^ {T \\times N \\times D}, \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.485, + 0.903 + ], + "angle": 0, + "content": "where \\( u \\) and \\( U_{0} \\) are the output membrane potential tensors from the PSM and SPS modules, respectively, and \\( \\mathcal{SN}(\\cdot) \\)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.435, + 0.907, + 0.57 + ], + "angle": 0, + "content": "represents the spike neuron layer. The output \\( U_{0} \\) is then passed to the \\( L \\)-block Spiking Transformer encoder, which consists of the \\( \\mathrm{A}^{2}\\mathrm{OS}^{2}\\mathrm{A} \\) and MLP blocks. Residual connections are applied to the membrane potentials in both the \\( \\mathrm{A}^{2}\\mathrm{OS}^{2}\\mathrm{A} \\) and MLP blocks. To avoid multi-bit spike outputs, we adopt the design from [58, 63], ensuring that the values before the convolutional layers are binary, thus allowing the spike and weight matrix multiplication to be simplified to addition operations." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.571, + 0.909, + 0.722 + ], + "angle": 0, + "content": "The \\(\\mathrm{A}^2\\mathrm{OS}^2\\mathrm{A}\\) mechanism models both local and global information of the input sequence using binary spike \\(Q\\), full-precision \\(K\\), and ternary spike \\(V\\). This approach reduces information loss and eliminates the need for scaling and softmax operations, thereby improving task accuracy. A Global Average Pooling (GAP) is applied to the processed feature from the Spiking Transformer encoder, and the resulting \\(D\\)-dimensional channel is passed to a fully connected Classification Head (CH) for final prediction. The \\(\\mathrm{A}^2\\mathrm{OS}^2\\mathrm{A}\\), MLP, and CH components are as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.601, + 0.727, + 0.907, + 0.745 + ], + "angle": 0, + "content": "\\[\nS _ {0} = \\mathcal {S N} (U _ {0}), S _ {0} \\in \\mathbb {R} ^ {T \\times N \\times D}, \\tag {14}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.518, + 0.759, + 0.907, + 0.791 + ], + "angle": 0, + "content": "\\[\nU _ {l} ^ {\\prime} = \\mathrm {A} ^ {2} \\mathrm {O S} ^ {2} \\mathrm {A} \\left(S _ {l - 1}\\right) + U _ {l - 1}, U _ {l} ^ {\\prime} \\in \\mathbb {R} ^ {T \\times N \\times D}, l = 1 \\dots L, \\tag {15}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.604, + 0.806, + 0.907, + 0.825 + ], + "angle": 0, + "content": "\\[\nS _ {l} ^ {\\prime} = \\mathcal {S N} \\left(U _ {l} ^ {\\prime}\\right), S _ {l} ^ {\\prime} \\in \\mathbb {R} ^ {T \\times N \\times D}, \\tag {16}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.529, + 0.838, + 0.907, + 0.87 + ], + "angle": 0, + "content": "\\[\nS _ {l} = \\mathcal {S N} \\left(\\operatorname {M L P} \\left(S _ {l} ^ {\\prime}\\right) + U _ {l} ^ {\\prime}\\right), S _ {l} \\in \\mathbb {R} ^ {T \\times N \\times D}, l = 1 \\dots L, \\tag {17}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.635, + 0.885, + 0.907, + 0.903 + ], + "angle": 0, + "content": "\\[\nY = \\mathrm {C H} (\\mathrm {G A P} (S _ {L})). \\tag {18}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "24402" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.453, + 0.108 + ], + "angle": 0, + "content": "4.2. Information Loss in Spiking Transformers" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.113, + 0.484, + 0.218 + ], + "angle": 0, + "content": "In this section, we analyze the limitations of binary spikes in SNNs with respect to information representation, which motivates the focus of this paper. While binary spikes offer energy efficiency, they inherently suffer from reduced representational capacity, leading to information loss when used for all \\( Q \\), \\( K \\), and \\( V \\) in VSSA, which results in accuracy degradation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.219, + 0.485, + 0.28 + ], + "angle": 0, + "content": "To support this claim, we first perform a theoretical analysis using the concept of entropy. The representational capability \\(\\mathcal{C}(\\mathbf{X})\\) of a set \\(\\mathbf{X}\\) is determined by the maximum entropy of \\(\\mathbf{X}\\), expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.114, + 0.286, + 0.484, + 0.318 + ], + "angle": 0, + "content": "\\[\n\\mathcal {C} (\\mathbf {X}) = \\max \\mathcal {H} (\\mathbf {X}) = - \\sum_ {x \\in \\mathbf {X}} p _ {\\mathbf {X}} (x) \\log p _ {\\mathbf {X}} (x), \\tag {19}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.325, + 0.484, + 0.357 + ], + "angle": 0, + "content": "where \\( p_{\\mathbf{X}}(x) \\) is the probability of a sample \\( x \\) from \\( \\mathbf{X} \\). We now present the following proposition:" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.364, + 0.485, + 0.442 + ], + "angle": 0, + "content": "Proposition 1 For a set \\(\\mathbf{X}\\), its representational capacity is \\(\\mathcal{C}(\\mathbf{X}) = \\max \\mathcal{H}(\\mathbf{X})\\). When the probability distribution of \\(\\mathbf{X}\\) is uniform, i.e., \\(p_{\\mathbf{X}}(x) = \\frac{1}{N}\\), where \\(N\\) is the total number of samples in \\(\\mathbf{X}\\), the entropy \\(\\mathcal{H}(\\mathbf{X})\\) reaches its maximum value of \\(\\log(N)\\). Hence, we conclude that \\(\\mathcal{C}(\\mathbf{X}) = \\log(N)\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.448, + 0.484, + 0.57 + ], + "angle": 0, + "content": "Using Proposition 1, we can evaluate the representational capacity of binary spike layers in SNNs and compare them with real-valued layers in ANNs. Let \\(\\mathbf{F}_{\\mathrm{B}} \\in \\mathbb{B}^{C \\times H \\times W}\\) denote the binary spike layer of the SNN, and \\(\\mathbf{F}_{\\mathrm{R}} \\in \\mathbb{R}^{C \\times H \\times W}\\) denote the real-valued feature map in the corresponding ANN. For a binary spike output \\(s\\), it requires 1 bit, and thus the number of possible samples from \\(s\\) is 2. Therefore, the number of samples for \\(\\mathbf{F}_{\\mathrm{B}}\\) is \\(2^{(C \\times H \\times W)}\\), leading to:" + }, + { + "type": "equation", + "bbox": [ + 0.124, + 0.576, + 0.484, + 0.602 + ], + "angle": 0, + "content": "\\[\n\\mathcal {C} \\left(\\mathbf {F} _ {\\mathrm {B}}\\right) = \\log \\left(2 ^ {(C \\times H \\times W)}\\right) = C \\times H \\times W. \\tag {20}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.607, + 0.484, + 0.653 + ], + "angle": 0, + "content": "In contrast, a real-valued output requires 32 bits per sample, yielding \\(2^{32}\\) possible values. Hence, the representational capacity for real-valued layers is:" + }, + { + "type": "equation", + "bbox": [ + 0.098, + 0.659, + 0.484, + 0.685 + ], + "angle": 0, + "content": "\\[\n\\mathcal {C} \\left(\\mathbf {F} _ {\\mathrm {R}}\\right) = \\log \\left(2 ^ {3 2 \\times (C \\times H \\times W)}\\right) = 3 2 \\times C \\times H \\times W. \\tag {21}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.69, + 0.484, + 0.797 + ], + "angle": 0, + "content": "This highlights the limited representational capacity of the binary spike layer. By transforming the real-valued \\( Q_{\\mathbb{R}} \\), \\( K_{\\mathbb{R}} \\), \\( V_{\\mathbb{R}} \\) into binary \\( Q_{\\mathbb{B}} \\), \\( K_{\\mathbb{B}} \\), \\( V_{\\mathbb{B}} \\), significant information loss occurs, reducing accuracy. Therefore, we propose increasing the bit precision of \\( Q \\), \\( K \\), and \\( V \\) in spiking attention to minimize information loss, while preserving the advantages of multiplication-addition transformations." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.804, + 0.484, + 0.835 + ], + "angle": 0, + "content": "4.3. Accurate Addition-Only Spiking Self-Attention Mechanism (A\\(^2\\)OS\\(^2\\)A)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.486, + 0.902 + ], + "angle": 0, + "content": "To maintain the integrity of information within spiking attention while leveraging the advantages of multiplication-addition transformations, we propose the Accurate Addition-Only Spiking Self-Attention \\((\\mathrm{A}^2\\mathrm{OS}^2\\mathrm{A})\\). This mechanism" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.183 + ], + "angle": 0, + "content": "employs binary \\(Q\\), full-precision \\(K\\), and ternary \\(V\\) in the self-attention process. Specifically, we utilize a binary spiking neuron after the linear layer to produce \\(Q\\), a ReLU function for generating \\(K\\), and a ternary spiking neuron for producing \\(V\\), as illustrated in Figure 1(c). The formulations for \\(Q, K, V\\) in \\(\\mathrm{A}^2\\mathrm{OS}^2\\mathrm{A}\\) are defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.615, + 0.19, + 0.907, + 0.209 + ], + "angle": 0, + "content": "\\[\nQ = \\mathcal {S N} _ {Q} ^ {b} (\\mathrm {B N} (X W _ {Q})), \\tag {22}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.615, + 0.212, + 0.907, + 0.229 + ], + "angle": 0, + "content": "\\[\nK = \\operatorname {R e L U} _ {K} (\\operatorname {B N} (X W _ {K})), \\tag {23}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.617, + 0.232, + 0.907, + 0.25 + ], + "angle": 0, + "content": "\\[\nV = \\mathcal {S N} _ {V} ^ {t} (\\mathrm {B N} (X W _ {V})), \\tag {24}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.258, + 0.906, + 0.319 + ], + "angle": 0, + "content": "where \\(\\mathcal{SN}^b (\\cdot)\\) denotes the binary spiking neuron as used in standard SNNs, \\(\\mathrm{ReLU}_K(\\cdot)\\) is the ReLU activation function typical in ANNs, and \\(\\mathcal{SN}^t (\\cdot)\\) represents a ternary spiking neuron that outputs values in the set \\(\\{-1,0,1\\}\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.32, + 0.906, + 0.366 + ], + "angle": 0, + "content": "The evolution of the membrane potential and the spike generation mechanism within \\( \\mathcal{SN}^t(\\cdot) \\) can be described by the following equations:" + }, + { + "type": "equation", + "bbox": [ + 0.582, + 0.374, + 0.907, + 0.391 + ], + "angle": 0, + "content": "\\[\nU [ t ] = H [ t - 1 ] + X [ t ], \\tag {25}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.585, + 0.393, + 0.907, + 0.41 + ], + "angle": 0, + "content": "\\[\nS [ t ] = \\operatorname {H e a} (| U [ t ] | - V _ {\\mathrm {t h}}), \\tag {26}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.582, + 0.412, + 0.907, + 0.429 + ], + "angle": 0, + "content": "\\[\nH [ t ] = V _ {\\text {r e s e t}} S [ t ] + \\beta U [ t ] (1 - | S [ t ] |). \\tag {27}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.437, + 0.836, + 0.452 + ], + "angle": 0, + "content": "The formulation of the \\(\\mathrm{A}^2\\mathrm{OS}^2\\mathrm{A}\\) is expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.57, + 0.46, + 0.907, + 0.479 + ], + "angle": 0, + "content": "\\[\n\\mathrm {A} ^ {2} \\mathrm {O S} ^ {2} \\mathrm {A} (Q, K, V) = \\mathcal {S N} \\left(Q \\cdot K ^ {T} \\cdot V\\right). \\tag {28}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.485, + 0.909, + 0.741 + ], + "angle": 0, + "content": "In \\(\\mathrm{A}^2\\mathrm{OS}^2\\mathrm{A}\\), since \\(Q\\) takes values in \\(\\{0,1\\}\\) and \\(K\\) is in \\(\\mathbb{R}\\), the matrix multiplication \\(Q\\cdot K^{T}\\) can be transformed into equivalent addition operations, resulting in \\(Q\\cdot K^{T}\\in \\mathbb{R}\\). Moreover, because both \\(Q\\) and \\(K\\) are naturally non-negative, this preserves a non-negative attention map \\(Q\\cdot K^T\\). Consequently, the mechanism eliminates the need for a softmax operation to enforce non-negativity within the attention map. Furthermore, since \\(K\\) is a real-valued vector without boundaries, there is no necessity for a scaling factor \\(s\\) to manage large values resulting from the matrix multiplication, unlike the approaches in [64, 65]. Additionally, given that \\(V\\) can also take values in \\(\\{-1,0,1\\}\\), the matrix product \\(Q\\cdot K^{T}\\cdot V\\) can similarly be expressed through addition operations. At the same time, unlike other matrix multiplication methods outlined in [64, 65], our framework allows for the result to include negative values, akin to what is seen in traditional ANNs." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.742, + 0.909, + 0.833 + ], + "angle": 0, + "content": "In summary, the \\(\\mathrm{A}^2\\mathrm{OS}^2\\mathrm{A}\\) mechanism utilizes full-precision \\(K\\) and ternary spike \\(V\\) to enhance the representational capability of self-attention and minimize information loss. Additionally, all attention calculations are performed using addition operations, which align with the characteristics of SNNs." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.845, + 0.64, + 0.863 + ], + "angle": 0, + "content": "5. Experiment" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.907, + 0.902 + ], + "angle": 0, + "content": "We evaluate our method on various datasets, including CIFAR-10/100 [31] and ImageNet-1K [8]. The network" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "24403" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.089, + 0.486, + 0.147 + ], + "angle": 0, + "content": "Table 1. Ablation study for Spiking Transformer on CIFAR10/100. Param refers to the number of parameters. Spiking Transformer- \\( L - D \\) is a Spiking Transformer model with \\( L \\) Spiking Transformer encoder blocks and \\( D \\) feature embedding dimensions." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.17, + 0.491, + 0.26 + ], + "angle": 0, + "content": "
MethodsParam (M)Time StepCIFAR10 AccCIFAR100 Acc
Baseline-2-2562.59494.3976.00
Spiking Transformer-2-2562.59494.9176.96
Baseline-2-51210.23495.5178.83
Spiking Transformer-2-51210.23496.4279.90
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.296, + 0.485, + 0.358 + ], + "angle": 0, + "content": "architecture and experimental setup follow the baseline from Spike-driven Transformer [58]. Further details regarding the experimental settings can be found in the Spike-driven Transformer [58]." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.365, + 0.245, + 0.381 + ], + "angle": 0, + "content": "5.1. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.387, + 0.485, + 0.447 + ], + "angle": 0, + "content": "We conduct a series of ablation experiments to assess the effectiveness of the proposed Spiking Transformer, comparing it with the baseline model from [58] on the CIFAR-10 and CIFAR-100 datasets. The results are summarized in Table 1." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.448, + 0.486, + 0.675 + ], + "angle": 0, + "content": "The baseline models used are Spike-driven Transformer-2-256 and Spike-driven Transformer-2-512. For the Transformer-2-256 with 4 timesteps, the baseline accuracy is \\(94.39\\%\\) on CIFAR-10 and \\(76.00\\%\\) on CIFAR-100, consistent with previous results. Our Spiking Transformer, using the same architecture, achieves notable improvements, with accuracies of \\(94.91\\%\\) and \\(76.96\\%\\) on CIFAR-10 and CIFAR-100, respectively—resulting in performance gains of approximately \\(0.5\\%\\) and \\(1.0\\%\\). For the larger model, Transformer-2-512 with 4 timesteps achieves baseline accuracies of \\(95.51\\%\\) and \\(78.43\\%\\) on CIFAR-10 and CIFAR-100. Our model with the same architecture shows substantial improvements, reaching accuracy scores of \\(96.42\\%\\) and \\(79.90\\%\\), corresponding to gains of approximately \\(0.9\\%\\) and \\(1.1\\%\\) on CIFAR-10 and CIFAR-100, respectively." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.682, + 0.185, + 0.697 + ], + "angle": 0, + "content": "5.2. CIFAR" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.486, + 0.901 + ], + "angle": 0, + "content": "To further evaluate the performance of our method, we compare it with recent SNN-based Transformer approaches, including Spikformer [64], Spikingformer [63] and so on, using a range of models with varying embedding dimensions and numbers of transformer blocks on the CIFAR dataset. The CIFAR dataset consists of 50,000 training images and 10,000 test images, all with a resolution of \\(32 \\times 32\\). We maintain the experimental setup from [58, 63, 64], including network architecture, training configurations, etc. Table 2 presents the accuracy of our method compared to other models on the CIFAR dataset. As shown in Table 2, our Spiking Transformer-2-384 achieves an accuracy of \\(95.70\\%\\) on the CIFAR-10 dataset, surpassing TET \\((94.44\\%)\\) and even" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.089, + 0.909, + 0.119 + ], + "angle": 0, + "content": "Table 2. Comparison of the performance between Spiking Transformer and existing approaches on CIFAR10/100." + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.143, + 0.916, + 0.393 + ], + "angle": 0, + "content": "
MethodsParam (M)Time StepCIFAR10 AccCIFAR100 Acc
Hybrid training[46]9.2712592.2267.87
Diet-SNN[45]0.2710/592.5464.07
STBP[55]17.541289.83-
STBP NeuNorm[54]17.541290.53-
TSSL-BP[60]17.54591.41-
STBP-tdBN[62]12.63492.9270.86
TET[9]12.63494.4474.47
Spikformer-4-256[64]4.15493.9475.96
Spikformer-2-384[64]5.76494.8076.95
Spikformer-4-384[64]9.32495.1977.86
Spikingformer-4-256[63]4.15494.7777.43
Spikingformer-2-384[63]5.76495.2278.34
Spikingformer-4-384[63]9.32495.6179.09
Spiking Transformer-4-2564.15494.9677.49
Spiking Transformer-2-3845.76495.7078.59
Spiking Transformer-4-3849.32496.3279.69
Spiking Transformer-2-51210.23496.4279.90
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.431, + 0.909, + 0.568 + ], + "angle": 0, + "content": "Spikingformer-4-384 (94.97%). Notably, performance improves further with higher embedding dimensions and more transformer blocks. In particular, Spiking Transformer-4-384 shows a 1.13% improvement over Spikformer-4-384 and a 0.71% improvement over Spikingformer-4-384. The performance gain of our Spiking Transformer is even more pronounced on more complex datasets such as CIFAR-100. Specifically, Spiking Transformer-4-384 achieves a notable improvement of 1.83% over Spikformer-4-384." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.577, + 0.627, + 0.592 + ], + "angle": 0, + "content": "5.3. ImageNet" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.599, + 0.909, + 0.749 + ], + "angle": 0, + "content": "We further evaluated our approach on the challenging ImageNet-1K dataset. ImageNet-1K [8] is a widely-used benchmark for image classification, consisting of 1,000 categories. The dataset includes approximately 1.28 million training images and 50,000 test images. For both training and evaluation, the images are resized to a default resolution of \\(224 \\times 224\\) pixels. To ensure consistency with previous work, we maintain the experimental setup from [58, 63, 64], including network architecture, training configurations, and other experimental details." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.911, + 0.903 + ], + "angle": 0, + "content": "Table 3 presents a comparative analysis of recent SoTA methods evaluated on the ImageNet dataset. Notable models in this domain include Spikformer [64], Spikingformer [63], and Spike-driven Transformer [58], which achieve accuracies of \\(73.38\\%\\), \\(74.79\\%\\), and \\(74.57\\%\\), respectively, with Transformer-8-512 as the baseline configuration. In comparison, our approach demonstrates a significant improvement, achieving an accuracy of \\(76.28\\%\\), which represents a \\(1.49\\%\\) increase over the best-performing transformer-based SNN model, Spikingformer. This improvement highlights the" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "24404" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.165, + 0.089, + 0.832, + 0.105 + ], + "angle": 0, + "content": "Table 3. Comparison of the performance between Spiking Transformer and existing approaches on ImageNet-1k." + }, + { + "type": "table", + "bbox": [ + 0.156, + 0.127, + 0.849, + 0.606 + ], + "angle": 0, + "content": "
CategoryMethodsArchitectureParam (M)Time StepAcc
ANN-to-SNNHybrid training[46]ResNet-3421.7925061.48
Spiking ResNet[27]ResNet-3421.7935071.61
ResNet-5025.5635072.75
QCFS[2]VGG-16138.426472.85
COS[24]ResNet-3421.79874.17
Directly LearningTET[9]Spiking-ResNet-3421.79664.79
SEW-ResNet-3421.79468.00
STBP-tdBN[62]Spiking-ResNet-3421.79663.72
SEW-ResNet-3421.79467.04
SEW ResNet[11]SEW-ResNet-5025.56467.78
SEW-ResNet-10144.55468.76
SEW-ResNet-15260.19469.26
Attention-SNN[57]ResNet-10478.37477.08
Directly LearningSpikformer[64]Spikformer-8-38416.81470.24
Spikformer-6-51223.37472.46
Spikformer-8-51229.68473.38
Spikformer-10-51236.01473.68
Spikformer-8-76866.34474.81
Directly LearningSpikingformer[63]Spikingformer-8-38416.81472.45
Spikingformer-8-51229.68474.79
Spikingformer-8-76866.34475.85
Directly LearningSpike-driven Transformer[58]Spike-driven Transformer-8-38416.81472.28
Spike-driven Transformer-6-51223.37474.11
Spike-driven Transformer-8-51229.68474.57
Spike-driven Transformer-10-51236.01474.66
Spike-driven Transformer-8-76866.34477.07
Directly LearningSpiking TransformerSpiking Transformer-8-38416.81474.04
Spiking Transformer-6-51223.37476.22
Spiking Transformer-8-51229.68476.28
Spiking Transformer-10-51236.01478.66
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.641, + 0.483, + 0.702 + ], + "angle": 0, + "content": "effectiveness of our method in leveraging the potential of spiking neuron models within the transformer architecture, resulting in enhanced performance on large-scale datasets like ImageNet-1K." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.703, + 0.483, + 0.809 + ], + "angle": 0, + "content": "Furthermore, even with a smaller network architecture, our Spiking Transformer-10-512 model outperforms the other methods, including those using Transformer-8-768 as the baseline. This demonstrates that our approach is not only more accurate but also more efficient, as it achieves superior results with fewer parameters, highlighting its scalability and robustness." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.829, + 0.21, + 0.844 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.485, + 0.903 + ], + "angle": 0, + "content": "In this paper, we have introduced the Accurate Addition-Only Spiking Self-Attention \\((\\mathrm{A}^2\\mathrm{OS}^2\\mathrm{A})\\) mechanism as a novel method for integrating Spiking Neural Networks with" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.641, + 0.908, + 0.763 + ], + "angle": 0, + "content": "Transformer models. Our approach tackles the inherent challenge of adapting the self-attention mechanism to the spiking paradigm by proposing a hybrid solution that combines binary, ReLU, and ternary spiking neurons. This innovative hybrid strategy enables the elimination of both softmax and scaling operations, which are traditionally essential in Transformer models, while preserving the fundamental advantages of addition-only computations in SNNs." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Through extensive experiments, we have demonstrated that \\(\\mathrm{A}^2\\mathrm{OS}^2\\mathrm{A}\\) outperforms existing SNN-based Transformer models in terms of both accuracy and efficiency. Specifically, our method achieves a SoTA accuracy of \\(78.66\\%\\) on the ImageNet-1K dataset. This performance, combined with the energy-efficient nature of SNNs, makes \\(\\mathrm{A}^2\\mathrm{OS}^2\\mathrm{A}\\) a promising candidate for real-world applications, especially those requiring high computational performance with low energy consumption." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "24405" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.241, + 0.108 + ], + "angle": 0, + "content": "Acknowledgment" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.114, + 0.486, + 0.157 + ], + "angle": 0, + "content": "This work is supported by grants from the National Natural Science Foundation of China under contracts No.12202412 and No.12202413." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.185, + 0.188, + 0.201 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.21, + 0.486, + 0.307 + ], + "angle": 0, + "content": "[1] Filipp Akopyan, Jun Sawada, Andrew Cassidy, Rodrigo Alvarez-Icaza, John Arthur, Paul Merolla, Nabil Imam, Yutaka Nakamura, Pallab Datta, Gi-Joon Nam, et al. Truenorth: Design and tool flow of a 65 mw 1 million neuron programmable neurosynaptic chip. IEEE transactions on computer-aided design of integrated circuits and systems, 34(10):1537-1557, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.308, + 0.485, + 0.376 + ], + "angle": 0, + "content": "[2] Tong Bu, Wei Fang, Jianhao Ding, PengLin Dai, Zhaofei Yu, and Tiejun Huang. Optimal ann-snn conversion for high-accuracy and ultra-low-latency spiking neural networks. In International Conference on Learning Representations, 2021. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.379, + 0.484, + 0.434 + ], + "angle": 0, + "content": "[3] Tong Bu, Jianhao Ding, Zhaofei Yu, and Tiejun Huang. Optimized potential initialization for low-latency spiking neural networks. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 11-20, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.436, + 0.483, + 0.476 + ], + "angle": 0, + "content": "[4] Natalia Caporale and Yang Dan. Spike timing-dependent plasticity: a hebbian learning rule. Annu. Rev. Neurosci., 31 (1):25-46, 2008. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.478, + 0.485, + 0.547 + ], + "angle": 0, + "content": "[5] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.549, + 0.484, + 0.616 + ], + "angle": 0, + "content": "[6] Chun-Fu Richard Chen, Quanfu Fan, and Rameswar Panda. Crossvit: Cross-attention multi-scale vision transformer for image classification. In Proceedings of the IEEE/CVF international conference on computer vision, pages 357–366, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.619, + 0.484, + 0.687 + ], + "angle": 0, + "content": "[7] Mike Davies, Narayan Srinivasa, Tsung-Han Lin, Gautham Chinya, Yongqiang Cao, Sri Harsha Choday, Georgios Dimou, Prasad Joshi, Nabil Imam, Shweta Jain, et al. Loihi: A neuromorphic manycore processor with on-chip learning. *IEEE Micro*, 38(1):82-99, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.689, + 0.484, + 0.745 + ], + "angle": 0, + "content": "[8] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.747, + 0.484, + 0.8 + ], + "angle": 0, + "content": "[9] Shikuang Deng, Yuhang Li, Shanghang Zhang, and Shi Gu. Temporal efficient training of spiking neural network via gradient re-weighting. arXiv preprint arXiv:2202.11946, 2022.7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.803, + 0.484, + 0.843 + ], + "angle": 0, + "content": "[10] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[11] Wei Fang, Zhaofei Yu, Yanqi Chen, Tiejun Huang, Timothee Masquelier, and Yonghong Tian. Deep residual learning in spiking neural networks. Advances in Neural Information Processing Systems, 34:21056-21069, 2021. 2, 8" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.21, + 0.486, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.162 + ], + "angle": 0, + "content": "[12] Yufei Guo, Weihang Peng, Xiaode Liu, Yuanpei Chen, Yuhan Zhang, Xin Tong, Zhou Jie, and Zhe Ma. Enof-snn: Training accurate spiking neural networks via enhancing the output feature. In The Thirty-eighth Annual Conference on Neural Information Processing Systems. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.164, + 0.905, + 0.219 + ], + "angle": 0, + "content": "[13] Yufei Guo, Yuanpei Chen, Liwen Zhang, Xiaode Liu, Yinglei Wang, Xuhui Huang, and Zhe Ma. IM-loss: Information maximization loss for spiking neural networks. In Advances in Neural Information Processing Systems, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.221, + 0.907, + 0.29 + ], + "angle": 0, + "content": "[14] Yufei Guo, Yuanpei Chen, Liwen Zhang, YingLei Wang, Xiaode Liu, Xinyi Tong, Yuanyuan Ou, Xuhui Huang, and Zhe Ma. Reducing information loss for spiking neural networks. In Computer Vision - ECCV 2022, pages 36-52, Cham, 2022. Springer Nature Switzerland. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.292, + 0.906, + 0.374 + ], + "angle": 0, + "content": "[15] Yufei Guo, Xinyi Tong, Yuanpei Chen, Liwen Zhang, Xiaode Liu, Zhe Ma, and Xuhui Huang. Recdis-snn: Rectifying membrane potential distribution for directly training spiking neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 326-335, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.377, + 0.907, + 0.459 + ], + "angle": 0, + "content": "[16] Yufei Guo, Liwen Zhang, Yuanpei Chen, Xinyi Tong, Xiaode Liu, YingLei Wang, Xuhui Huang, and Zhe Ma. Real spike: Learning real-valued spikes for spiking neural networks. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XII, pages 52-68. Springer, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.462, + 0.907, + 0.543 + ], + "angle": 0, + "content": "[17] Yufei Guo, Xiaode Liu, Yuanpei Chen, Liwen Zhang, Weihang Peng, Yuhan Zhang, Xuhui Huang, and Zhe Ma. Rmploss: Regularizing membrane potential distribution for spiking neural networks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17391-17401, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.546, + 0.906, + 0.614 + ], + "angle": 0, + "content": "[18] Yufei Guo, Weihang Peng, Yuanpei Chen, Liwen Zhang, Xiaode Liu, Xuhui Huang, and Zhe Ma. Joint a-snn: Joint training of artificial and spiking neural networks via self-distillation and weight factorization. Pattern Recognition, 142:109639, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.617, + 0.907, + 0.672 + ], + "angle": 0, + "content": "[19] Yufei Guo, Yuhan Zhang, Yuanpei Chen, Weihang Peng, Xiaode Liu, Liwen Zhang, Xuhui Huang, and Zhe Ma. Membrane potential batch normalization for spiking neural networks. arXiv preprint arXiv:2308.08359, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.675, + 0.907, + 0.743 + ], + "angle": 0, + "content": "[20] Yufei Guo, Yuanpei Chen, Xiaode Liu, Weihang Peng, Yuhan Zhang, Xuhui Huang, and Zhe Ma. Ternary spike: Learning ternary spikes for spiking neural networks. In Proceedings of the AAAI conference on artificial intelligence, pages 12244-12252, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.746, + 0.906, + 0.814 + ], + "angle": 0, + "content": "[21] Yufei Guo, Yuanpei Chen, Zecheng Hao, Weihang Peng, Zhou Jie, Yuhan Zhang, Xiaode Liu, and Zhe Ma. Take a shortcut back: Mitigating the gradient vanishing for training spiking neural networks. Advances in Neural Information Processing Systems, 37:24849-24867, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.817, + 0.906, + 0.87 + ], + "angle": 0, + "content": "[22] Bing Han and Kaushik Roy. Deep spiking neural network: Energy efficiency through time based coding. In European Conference on Computer Vision, pages 388-404. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[23] Zecheng Hao, Tong Bu, Jianhao Ding, Tiejun Huang, and Zhaofei Yu. Reducing ann-snn conversion error through resid" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "24406" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.126, + 0.092, + 0.484, + 0.119 + ], + "angle": 0, + "content": "ual membrane potential. arXiv preprint arXiv:2302.02091, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.484, + 0.176 + ], + "angle": 0, + "content": "[24] Zecheng Hao, Jianhao Ding, Tong Bu, Tiejun Huang, and Zhaofei Yu. Bridging the gap between anns and snns by calibrating offset spikes. arXiv preprint arXiv:2302.10685, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.179, + 0.484, + 0.233 + ], + "angle": 0, + "content": "[25] Nguyen-Dong Ho and Ik-Joon Chang. Tcl: an ann-to-snn conversion with trainable clipping layers. In 2021 58th ACM/IEEE Design Automation Conference (DAC), pages 793-798. IEEE, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.236, + 0.483, + 0.289 + ], + "angle": 0, + "content": "[26] Allan L Hodgkin and Andrew F Huxley. Currents carried by sodium and potassium ions through the membrane of the giant axon of loligo. The Journal of physiology, 116(4):449, 1952. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.293, + 0.483, + 0.333 + ], + "angle": 0, + "content": "[27] Yangfan Hu, Huajin Tang, and Gang Pan. Spiking deep residual networks. IEEE Transactions on Neural Networks and Learning Systems, 34(8):5200-5205, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.335, + 0.483, + 0.404 + ], + "angle": 0, + "content": "[28] Asifullah Khan, Zunaira Rauf, Abdul Rehman Khan, Saima Rathore, Saddam Hussain Khan, Najmus Saher Shah, Umair Farooq, Hifsa Asif, Aqsa Asif, Umme Zahoora, et al. A recent survey of vision transformers for medical image segmentation. arXiv preprint arXiv:2312.00634, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.406, + 0.483, + 0.447 + ], + "angle": 0, + "content": "[29] Seijoon Kim, Seongsik Park, Byunggook Na, and Sungroh Yoon. Spiking-yolo: Spiking neural network for energy-efficient object detection, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.449, + 0.484, + 0.518 + ], + "angle": 0, + "content": "[30] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.519, + 0.483, + 0.56 + ], + "angle": 0, + "content": "[31] Alex Krizhevsky, Vinod Nair, and Geoffrey Hinton. Cifar-10 (canadian institute for advanced research). URL http://www.cs.toronto.edu/kriz/cifar.html, 5(4):1, 2010. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.563, + 0.483, + 0.617 + ], + "angle": 0, + "content": "[32] Yuxiang Lan, Yachao Zhang, Xu Ma, Yanyun Qu, and Yun Fu. Efficient converted spiking neural network for 3d and 2d classification. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9211-9220, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.619, + 0.483, + 0.674 + ], + "angle": 0, + "content": "[33] Nathan Leroux, Jan Finkbeiner, and Emre Neftci. Online transformers with spiking neurons for fast prosthetic hand control. In 2023 IEEE Biomedical Circuits and Systems Conference (BioCAS), pages 1-6. IEEE, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.676, + 0.483, + 0.731 + ], + "angle": 0, + "content": "[34] Yuhang Li, Shikuang Deng, Xin Dong, Ruihao Gong, and Shi Gu. A free lunch from ann: Towards efficient, accurate spiking neural networks calibration. In International Conference on Machine Learning, pages 6316-6325. PMLR, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.733, + 0.483, + 0.8 + ], + "angle": 0, + "content": "[35] Yuhang Li, Yufei Guo, Shanghang Zhang, Shikuang Deng, Yongqing Hai, and Shi Gu. Differentiable spike: Rethinking gradient-descent for training spiking neural networks. Advances in Neural Information Processing Systems, 34:23426-23439, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.803, + 0.483, + 0.844 + ], + "angle": 0, + "content": "[36] Yawei Li, Kai Zhang, Jiezhang Cao, Radu Timofte, and Luc Van Gool. Localvit: Bringing locality to vision transformers. arXiv preprint arXiv:2104.05707, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.483, + 0.899 + ], + "angle": 0, + "content": "[37] Yudong Li, Yunlin Lei, and Xu Yang. Spikeformer: a novel architecture for training high-performance low-latency spiking neural network. arXiv preprint arXiv:2211.10686, 2022.3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.148 + ], + "angle": 0, + "content": "[38] Yanghao Li, Hanzi Mao, Ross Girshick, and Kaiming He. Exploring plain vision transformer backbones for object detection. In European conference on computer vision, pages 280-296. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.151, + 0.908, + 0.219 + ], + "angle": 0, + "content": "[39] De Ma, Juncheng Shen, Zonghua Gu, Ming Zhang, Xiaolei Zhu, Xiaoqiang Xu, Qi Xu, Yangjing Shen, and Gang Pan. Darwin: A neuromorphic hardware co-processor based on spiking neural networks. Journal of Systems Architecture, 77: 43-51, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.222, + 0.905, + 0.263 + ], + "angle": 0, + "content": "[40] Wolfgang Maass. Networks of spiking neurons: the third generation of neural network models. Neural networks, 10 (9):1659-1671, 1997. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.266, + 0.908, + 0.307 + ], + "angle": 0, + "content": "[41] ZHI Min and LU Jingfang. A review of vision transformer for image classification. Journal of Zhengzhou University: Engineering Science, 45(4), 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.31, + 0.908, + 0.364 + ], + "angle": 0, + "content": "[42] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4195-4205, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.368, + 0.905, + 0.422 + ], + "angle": 0, + "content": "[43] Jing Pei, Lei Deng, Sen Song, Mingguo Zhao, Youhui Zhang, Shuang Wu, Guanrui Wang, Zhe Zou, Zhenzhi Wu, Wei He, et al. Towards artificial general intelligence with hybrid tianjin chip architecture. Nature, 572(7767):106-111, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.425, + 0.905, + 0.466 + ], + "angle": 0, + "content": "[44] Jinye Qu, Zeyu Gao, Tielin Zhang, Yanfeng Lu, Huajin Tang, and Hong Qiao. Spiking neural network for ultra-low-latency and high-accurate object detection, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.469, + 0.905, + 0.511 + ], + "angle": 0, + "content": "[45] Nitin Rathi and Kaushik Roy. Diet-snn: Direct input encoding with leakage and threshold optimization in deep spiking neural networks. arXiv preprint arXiv:2008.03658, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.513, + 0.905, + 0.568 + ], + "angle": 0, + "content": "[46] Nitin Rathi, Gopalakrishnan Srinivasan, Priyadarshini Panda, and Kaushik Roy. Enabling deep spiking neural networks with hybrid conversion and spike timing dependent backpropagation. arXiv preprint arXiv:2005.01807, 2020. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.571, + 0.905, + 0.625 + ], + "angle": 0, + "content": "[47] Dayong Ren, Zhe Ma, Yuanpei Chen, Weihang Peng, Xiaode Liu, Yuhan Zhang, and Yufei Guo. Spiking pointnet: Spiking neural networks for point clouds. arXiv preprint arXiv:2310.06232, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.628, + 0.905, + 0.682 + ], + "angle": 0, + "content": "[48] Abhronil Sengupta, Yuting Ye, Robert Wang, Chiao Liu, and Kaushik Roy. Going deeper in spiking neural networks: Vgg and residual architectures. Frontiers in neuroscience, 13:95, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.686, + 0.905, + 0.755 + ], + "angle": 0, + "content": "[49] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In International conference on machine learning, pages 10347-10357. PMLR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.758, + 0.905, + 0.785 + ], + "angle": 0, + "content": "[50] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.788, + 0.905, + 0.843 + ], + "angle": 0, + "content": "[51] Shuai Wang, Malu Zhang, Dehao Zhang, Ammar Belatreche, Yichen Xiao, Yu Liang, Yimeng Shan, Qian Sun, Enqi Zhang, and Yang Yang. Spiking vision transformer with saccadic attention. arXiv preprint arXiv:2502.12677, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[52] Wenjie Wei, Malu Zhang, Jilin Zhang, Ammar Belatreche, Jibin Wu, Zijing Xu, Xuerui Qiu, Hong Chen, Yang Yang, and Haizhou Li. Event-driven learning for spiking neural networks. arXiv preprint arXiv:2403.00270, 2024.1" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24407" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.486, + 0.147 + ], + "angle": 0, + "content": "[53] Yujie Wu, Lei Deng, Guoqi Li, Jun Zhu, and Luping Shi. Spatio-temporal backpropagation for training high-performance spiking neural networks. Frontiers in neuroscience, 12:331, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.483, + 0.204 + ], + "angle": 0, + "content": "[54] Yujie Wu, Lei Deng, Guoqi Li, Jun Zhu, Yuan Xie, and Luping Shi. Direct training for spiking neural networks: Faster, larger, better. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1311-1318, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.205, + 0.482, + 0.232 + ], + "angle": 0, + "content": "[55] Y. et al. Wu. Spatio-temporal learning with spiking neural networks. Nature Communications, 9, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.233, + 0.485, + 0.3 + ], + "angle": 0, + "content": "[56] Mingqing Xiao, Qingyan Meng, Zongpeng Zhang, Yisen Wang, and Zhouchen Lin. Training feedback spiking neural networks by implicit differentiation on the equilibrium state. Advances in Neural Information Processing Systems, 34:14516-14528, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.302, + 0.483, + 0.357 + ], + "angle": 0, + "content": "[57] Man Yao, Guangshe Zhao, Hengyu Zhang, Yifan Hu, Lei Deng, Yonghong Tian, Bo Xu, and Guoqi Li. Attention spiking neural networks. IEEE transactions on pattern analysis and machine intelligence, 45(8):9393-9410, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.359, + 0.483, + 0.399 + ], + "angle": 0, + "content": "[58] Man Yao, Jiakui Hu, Zhaokun Zhou, Li Yuan, Yonghong Tian, Bo Xu, and Guoqi Li. Spike-driven transformer. Advances in neural information processing systems, 36, 2024. 1, 3, 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.401, + 0.483, + 0.496 + ], + "angle": 0, + "content": "[59] Malu Zhang, Jiadong Wang, Jibin Wu, Ammar Belatreche, Burin Amornpaisannon, Zhixuan Zhang, Venkata Pavan Kumar Miriyala, Hong Qu, Yansong Chua, Trevor E Carlson, et al. Rectified linear postsynaptic potential function for backpropagation in deep spiking neural networks. IEEE transactions on neural networks and learning systems, 33 (5):1947-1958, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.498, + 0.483, + 0.551 + ], + "angle": 0, + "content": "[60] Wenrui Zhang and Peng Li. Temporal spike sequence learning via backpropagation for deep spiking neural networks. Advances in Neural Information Processing Systems, 33:12022-12033, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.553, + 0.483, + 0.621 + ], + "angle": 0, + "content": "[61] Yuhan Zhang, Xiaode Liu, Yuanpei Chen, Weihang Peng, Yufei Guo, Xuhui Huang, and Zhe Ma. Enhancing representation of spiking neural networks via similarity-sensitive contrastive learning. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 16926-16934, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.623, + 0.483, + 0.678 + ], + "angle": 0, + "content": "[62] Hanle Zheng, Yujie Wu, Lei Deng, Yifan Hu, and Guoqi Li. Going deeper with directly-trained larger spiking neural networks. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 11062-11070, 2021. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.679, + 0.483, + 0.746 + ], + "angle": 0, + "content": "[63] Chenlin Zhou, Liutao Yu, Zhaokun Zhou, Zhengyu Ma, Han Zhang, Huihui Zhou, and Yonghong Tian. Spikingformer: Spike-driven residual learning for transformer-based spiking neural network. arXiv preprint arXiv:2304.11954, 2023. 1, 3, 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.748, + 0.483, + 0.803 + ], + "angle": 0, + "content": "[64] Zhaokun Zhou, Yuesheng Zhu, Chao He, Yaowei Wang, Shuicheng Yan, Yonghong Tian, and Li Yuan. Spikformer: When spiking neural network meets transformer. arXiv preprint arXiv:2209.15425, 2022. 1, 3, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.805, + 0.483, + 0.859 + ], + "angle": 0, + "content": "[65] Zhaokun Zhou, Kaiwei Che, Wei Fang, Keyu Tian, Yuesheng Zhu, Shuicheng Yan, Yonghong Tian, and Li Yuan. Spikformer v2: Join the high accuracy club on imagenet with an snn ticket. arXiv preprint arXiv:2401.02020, 2024. 1, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.86, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[66] Shihao Zou, Yuxuan Mu, Xinxin Zuo, Sen Wang, and Li Cheng. Event-based human pose tracking by spiking spatiotemporal transformer, 2023. 1" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.486, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.956 + ], + "angle": 0, + "content": "24408" + } + ] +] \ No newline at end of file diff --git a/2025/Spiking Transformer_ Introducing Accurate Addition-Only Spiking Self-Attention for Transformer/24b0ab2b-a4ff-49fc-865c-b8d7799e49bd_origin.pdf b/2025/Spiking Transformer_ Introducing Accurate Addition-Only Spiking Self-Attention for Transformer/24b0ab2b-a4ff-49fc-865c-b8d7799e49bd_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..93c5d7897fad283c4f9da93c5260cb13bddec2f1 --- /dev/null +++ b/2025/Spiking Transformer_ Introducing Accurate Addition-Only Spiking Self-Attention for Transformer/24b0ab2b-a4ff-49fc-865c-b8d7799e49bd_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f136f96a82cc5425b8c4d56f17d2208039a6a3b13edffa48a3126a3269ac4658 +size 641694 diff --git a/2025/Spiking Transformer_ Introducing Accurate Addition-Only Spiking Self-Attention for Transformer/full.md b/2025/Spiking Transformer_ Introducing Accurate Addition-Only Spiking Self-Attention for Transformer/full.md new file mode 100644 index 0000000000000000000000000000000000000000..0720c7cda6f2c05f647626d70ccef5929301914d --- /dev/null +++ b/2025/Spiking Transformer_ Introducing Accurate Addition-Only Spiking Self-Attention for Transformer/full.md @@ -0,0 +1,379 @@ +# Spiking Transformer: Introducing Accurate Addition-Only Spiking Self-Attention for Transformer + +Yufei Guo,\* Xiaode Liu,\* Yuanpei Chen, Weihang Peng, Yuhan Zhang, Zhe Ma\* Intelligent Science & Technology Academy of CASIC, China + +yfguo@pku.edu.cn, lxde@pku.edu.cn, mazhe_thu@163.com + +# Abstract + +Transformers have demonstrated outstanding performance across a wide range of tasks, owing to their self-attention mechanism, but they are highly energy-consuming. Spiking Neural Networks have emerged as a promising energy-efficient alternative to traditional Artificial Neural Networks, leveraging event-driven computation and binary spikes for information transfer. The combination of Transformers' capabilities with the energy efficiency of SNNs offers a compelling opportunity. This paper addresses the challenge of adapting the self-attention mechanism of Transformers to the spiking paradigm by introducing a novel approach: Accurate Addition-Only Spiking Self-Attention ( $A^2OS^2A$ ). Unlike existing methods that rely solely on binary spiking neurons for all components of the self-attention mechanism, our approach integrates binary, ReLU, and ternary spiking neurons. This hybrid strategy significantly improves accuracy while preserving non-multiplicative computations. Moreover, our method eliminates the need for softmax and scaling operations. Extensive experiments show that the $A^2OS^2A$ -based Spiking Transformer outperforms existing SNN-based Transformers on several datasets, even achieving an accuracy of $78.66\%$ on ImageNet-1K. Our work represents a significant advancement in SNN-based Transformer models, offering a more accurate and efficient solution for real-world applications. + +# 1. Introduction + +Spiking Neural Networks (SNNs) have gained significant traction as an efficient neural network model, finding applications in diverse areas such as object recognition [15, 16, 34, 56], object detection [29, 44], and pose estimation [66]. These networks utilize binary spike signals for information transfer. A spiking neuron generates a spike, denoted by 1, when its membrane potential surpasses a cer- + +tain threshold; conversely, it does not spike, represented by 0, when the threshold is not met. This distinctive approach to information processing is notably energy-efficient, as it substitutes the complex multiplications involved in weight and activation interactions with simpler addition operations. Furthermore, SNNs can be effectively implemented using event-driven computations on neuromorphic hardware [1, 7, 39, 43, 52]. In this setup, the computational units are activated solely upon the occurrence of a spike, conserving energy by remaining inactive in the absence of spikes. Research indicates that SNNs can achieve significantly greater energy savings compared to their Artificial Neural Network (ANN) counterparts [1, 7]. + +Despite the clear advantages of SNNs in terms of energy efficiency, their practical application is often constrained by limited task accuracy. Meanwhile, Transformers have demonstrated exceptional performance across a wide range of tasks, largely due to their self-attention mechanism [30, 42, 50]. Combining the strengths of Transformers with the energy efficiency of SNNs presents a promising opportunity. However, adapting the self-attention mechanism to SNNs is non-trivial. In the vanilla self-attention framework (VSA), three key components are involved: Query $(Q)$ , Key $(K)$ , and Value $(V)$ . As illustrated in Figure 1(a), the VSA process begins by computing the dot product of the floating-point representations of $Q$ and $K$ , resulting in a matrix. This matrix is then normalized using a softmax function, which involves exponential and division operations, to produce an attention map that determines the weighting of $V$ . These operations in the VSA are incompatible with the operational principles of SNNs, which aim to minimize multiplication. Consequently, to implement a Transformer architecture in SNNs, it is essential to develop a novel, efficient self-attention mechanism that avoids multiplication. Ongoing research efforts are exploring solutions in this domain. Approaches such as Spikformer [64, 65], Spikingformer [63], and Spike-driven Transformer [58] have focused on transforming $Q$ , $K$ , and $V$ into spike representations prior to performing matrix operations. This transformation enables the replacement of matrix multiplications with addition, aligning + +better with the principles of SNNs. + +While previous approaches utilize spike-form $Q$ , $K$ , and $V$ to avoid multiplications, simply introducing spiking neurons to convert $Q$ , $K$ , and $V$ into spike form prior to performing matrix multiplications does not fully exploit the advantages of the addition-only operation inherent in SNNs. It is unnecessary to convert all of $Q$ , $K$ , and $V$ into spike form in order to transition from matrix multiplications to matrix additions. However, this method can lead to significant information loss, as demonstrated in Section 4. To address this issue, we propose the Accurate Addition-Only Spiking Self-Attention (A $^2$ OS $^2$ A). The distinctions between A $^2$ OS $^2$ A and the vanilla Spiking Self-Attention (VSSA) are illustrated in Figure 1(b) and Figure 1(c). Unlike VSSA, which applies the same binary spiking neuron to generate $Q$ , $K$ , and $V$ , our approach employs a binary spiking neuron for $Q$ , a ReLU activation for $K$ , and a ternary spiking neuron for $V$ . This design allows A $^2$ OS $^2$ A to retain the multiplication-addition transformation while reducing information loss. Moreover, it eliminates the need for both the scaling and softmax functions, enabling the output of floating-point values similar to the VSA, whereas VSSA produces only integer outputs. The key contributions of this paper are as follows: + +- We present a theoretical framework for understanding the information loss associated with vanilla Spiking Self-Attention mechanisms. To the best of our knowledge, this is the first detailed examination of this issue within the context of Spiking Self-Attention, which opens the door for future advancements in SNN-based Transformers. +- We introduce $\mathrm{A}^2\mathrm{OS}^2\mathrm{A}$ , an innovative and efficient approach that incorporates binary spiking neurons, ReLU activations, and ternary spiking neurons, moving beyond the traditional binary-only spiking neuron model. This method reduces information loss while retaining the advantages of addition-only processing in SNNs. Additionally, it eliminates the need for scaling and softmax functions. +- Comprehensive experiments demonstrate that our proposed architecture either surpasses or matches the performance of State-of-the-Art (SoTA) SNN-based Transformer models on various datasets. Notably, we achieved an accuracy of $78.66\%$ on ImageNet-1K, setting a new benchmark in the SNN domain. + +# 2. Related Work + +# 2.1. Spiking Neural Networks + +SNNs are a class of neural networks that more closely emulate the behavior of biological neurons compared to ANNs. The primary distinction lies in their use of discrete spikes for information transmission, as opposed to the continuous signals employed by ANNs. Foundational work in this area can be traced back to Hodgkin and Huxley [26], who + +developed models to simulate the action potentials of biological neurons. Since then, numerous frameworks and models have been introduced to enhance the capabilities and learning efficiency of SNNs. A notable example is the Leaky Integrate-and-Fire (LIF) model [40], which provides a simple yet effective means of simulating neuronal behavior. More recently, the incorporation of Spike-Timing-Dependent Plasticity (STDP) as a learning rule has enabled SNNs to learn temporal patterns effectively [4]. + +In the past decade, there has been a significant increase in interest regarding the training of SNNs using techniques adapted from ANNs, as well as their application to practical tasks. Two prevalent learning paradigms in current SNN research are the conversion from ANN to SNN (ANN2SNN) [2, 3, 22-25, 32, 34, 48] and supervised learning [12-14, 18-21, 35, 47, 59]. The ANN-SNN conversion method involves initially training an ANN and subsequently converting it into a homogeneous SNN by transferring the trained weights and substituting ReLU neurons with temporal spiking neurons. However, this method often fails for neuromorphic datasets, as ReLU neurons do not adequately capture the complex temporal dynamics required for processing sequential information. In contrast, supervised learning [11, 17, 53, 61] employs alternative functions during backpropagation to approximate the firing process, enabling the direct training of SNNs as if they were ANNs. This approach leverages the advantages of gradient-based optimization, achieving impressive performance with just a few time steps, even on large-scale datasets. Furthermore, supervised learning has proven to be effective in handling temporal data, establishing it as a preferred methodology in SNN research. + +# 2.2. Vision Transformers + +Vision Transformers (ViTs) have recently emerged as a powerful alternative to convolutional neural networks (CNNs) for a wide range of visual tasks, demonstrating remarkable efficacy and flexibility. Introduced by Dosovitskiy et al. [10], ViTs leverage the transformer architecture—originally designed for natural language processing—to process images in a novel manner. The architecture innovatively splits images into patches and treats these patches as sequences, enabling the model to capture long-range dependencies and relationships within the visual data, a limitation often encountered in traditional CNNs. The introduction of ViTs has spurred significant advancements in various computer vision tasks, including image classification [41], object detection [38], and segmentation [28]. These advancements highlight the potential of ViTs to tackle complex visual recognition problems with unprecedented accuracy. Notably, it has been shown that with appropriate training techniques and larger datasets, ViTs can not only match but surpass the performance of state-of-the-art CNNs, even in data-scarce scenarios, presenting a + +![](images/450144b6bf9cce2e7f345cdf281ed193bc9ac75e1450daeb4316cc0d6fe35738.jpg) +Figure 1. The difference between our Spiking Self-Attention and the vanilla Spiking Self-Attention. Our Spiking Self-Attention differs significantly from the vanilla Spiking Self-Attention. In the vanilla version, only binary spikes are employed, which can result in considerable information loss. In contrast, our approach integrates a combination of binary spiking neurons, ReLU, and ternary spiking neurons. This hybrid structure effectively reduces information loss, while preserving the advantages of addition-only processing. Moreover, our method eliminates the need for both scaling and the softmax function, streamlining the computation. + +![](images/e2ca4f51cb57896d5beadc3da775bafdaaf9468bfe96318c9c0ef58f2c8dad44.jpg) + +![](images/e4ab02036ee1026ef21566b118003223f9804a95c5e45d0939ea639bf050daf0.jpg) + +significant advantage in real-world applications [49]. + +Furthermore, the adaptation of transformers for visual tasks has inspired numerous variations and improvements across the field. Researchers have explored strategies such as integrating convolutional layers into the architecture [6] to retain spatial information more effectively. Other approaches include the use of hierarchical representations [36], which facilitate multi-scale feature extraction, and the incorporation of attention mechanisms tailored specifically for vision tasks [5], improving the model's ability to focus on relevant parts of the input data. These developments underscore the versatility and robustness of the transformer architecture in addressing complex visual challenges, paving the way for further innovations in computer vision. + +# 2.3. Spiking Neural Networks for Transformers + +The integration of SNNs into transformer architectures represents an exciting and rapidly developing area of research. This field aims to leverage the distinct advantages of both paradigms: the energy-efficient characteristics of SNNs and the robust contextual representation capabilities of transformers. Recent studies [33, 37, 51] have explored replacing certain neurons within transformers with spiking neurons, marking a significant step towards this integration. While these efforts contribute to mitigating the accuracy loss associated with introducing spiking neurons into transformers, they still face challenges in fully realizing the low energy consumption benefits of SNNs. This is primarily due to + +their reliance on hybrid computing methodologies, which continue to require traditional Multiply-Accumulate (MAC) operations, such as dot products, softmax calculations, and scaling operations. + +To address this issue, Spikformer [64, 65] proposes converting the $Q$ , $K$ , and $V$ components into spike form before performing matrix multiplications similar to those used in VSA. This innovative approach allows spike matrix multiplications to be transformed into addition operations, eliminating the need for softmax computations. However, when considering the residual connections in these architectures, Spikformer [64, 65] still incorporates non-spiking computations within the ConvBN layers. To overcome this limitation, Spikingformer [63] and Spike-driven Transformer [58] have restructured the residual connections in standard transformer architectures, placing them before activation functions to ensure that all neurons convey binary spike signals. + +The computation of spike-form $Q$ , $K$ , and $V$ in these models circumvent multiplications, relying solely on additions. However, simply placing spiking neurons in front of $Q$ , $K$ , and $V$ to convert them into spike form before performing matrix multiplications does not fully exploit the addition-only advantage of SNNs. In this work, we propose a novel, addition-only spiking self-attention mechanism for transformers, aiming to push the boundaries of what can be achieved through the integration of these two technologies. + +# 3. Preliminary + +# 3.1. Spiking Neuron Layer + +The spiking neuron layer plays a crucial role in integrating both spatial and temporal information, which is then encoded into the membrane potential and subsequently transformed into binary spikes. These spikes drive further computations in subsequent layers of the network. We model the dynamics of the spiking neuron using the LIF model [40]. The evolution of the membrane potential and the spike generation mechanism are described by the following equations: + +$$ +U [ t ] = H [ t - 1 ] + X [ t ], \tag {1} +$$ + +$$ +S [ t ] = \operatorname {H e a} (U [ t ] - V _ {\mathrm {t h}}), \tag {2} +$$ + +$$ +H [ t ] = V _ {\text {r e s e t}} S [ t ] + \beta U [ t ] (1 - S [ t ]), \tag {3} +$$ + +where: $X[t]$ is the spatial input current at time $t$ , $U[t]$ is the membrane potential at time $t$ , combining the spatial input $X[t]$ and the temporal input $H[t - 1]$ from the previous time step, $\mathrm{Hea}(\cdot)$ is the Heaviside step function, which outputs 1 if its argument is non-negative, and 0 otherwise, $S[t]$ is the spike output at time $t$ , $V_{\mathrm{th}}$ is the spike threshold, $V_{\mathrm{reset}}$ is the reset potential, and $\beta$ is a decay factor controlling the membrane potential's decay. + +When the membrane potential $U[t]$ exceeds a predefined threshold $V_{\mathrm{th}}$ , the neuron emits a spike, $S[t] = 1$ , and the internal state $H[t]$ is reset to $V_{\mathrm{reset}}$ . If $U[t]$ does not surpass the threshold, the membrane potential decays towards the previous state $H[t - 1]$ at a rate governed by $\beta$ . To simplify notation, we represent the spiking neuron layer as $\mathcal{SN}(\cdot)$ , where the input is the membrane potential tensor $U$ , and the output is the spike tensor $S$ . + +# 3.2. Vanilla Self-Attention Mechanism (VSA) + +The Vanilla Self-Attention (VSA) mechanism enables a model to focus on different parts of the input sequence while constructing its output representations. The self-attention operation can be mathematically expressed as: + +$$ +\operatorname {V S A} (Q, K, V) = \operatorname {s o f t m a x} \left(\frac {Q K ^ {T}}{\sqrt {d _ {k}}}\right) V \tag {4} +$$ + +where $Q$ , $K$ , and $V$ represent the query, key, and value matrices, respectively, and $d_k$ is the dimension of the key vectors. The softmax function is applied to normalize the attention scores, ensuring they are non-negative. The resulting weighted sum of the values reflects the importance of each value, based on the similarity between the query and its corresponding key. + +The input to the self-attention mechanism consists of a sequence of embeddings $x_{1}, x_{2}, \ldots, x_{n}$ , which are linearly projected into $Q, K$ , and $V$ using learned weight matrices: + +$$ +Q = X W _ {Q}, \quad K = X W _ {K}, \quad V = X W _ {V}, \tag {5} +$$ + +where $W_{Q}$ , $W_{K}$ , and $W_{V}$ are the learned weight matrices corresponding to the $Q$ , $K$ , and $V$ , respectively. + +However, the standard VSA is not directly compatible with SNNs due to the conflict between the floating-point operations required for matrix multiplication of $Q$ , $K$ , and $V$ , and the energy efficiency of SNN computations. What's more, the softmax operation, which involves exponentiation and division, does not align with the computation paradigm of SNNs too. + +# 3.3. Vanilla Spiking Self-Attention Mechanism (VSSA) + +To address the incompatibility between the VSA and SNNs, the Vanilla Spiking Self-Attention (VSSA) mechanism is proposed in [64, 65], which is more suitable for SNNs, as shown in Figure 1(b). The $Q$ , $K$ , and $V$ are initially derived through learnable weight matrices. These are then transformed into spiking sequences by applying binary spiking neuron layers: + +$$ +Q = \mathcal {S N} _ {Q} (\mathrm {B N} (X W _ {Q})), \tag {6} +$$ + +$$ +K = \mathcal {S N} _ {K} (\mathrm {B N} (X W _ {K})), \tag {7} +$$ + +$$ +V = \mathcal {S N} _ {V} (\mathrm {B N} (X W _ {V})). \tag {8} +$$ + +In this approach, the computation of the attention matrix is performed using binary spike-form queries and keys (which contain only 0s and 1s), thus replacing traditional matrix multiplication with addition-based operations. + +To further address the challenges of large values resulting from the matrix multiplication, a scaling factor $s$ is introduced to regulate the magnitude of the result. The spike-efficient VSSA is formulated as follows: + +$$ +\operatorname {V S S A} (Q, K, V) = \mathcal {S N} \left(Q K ^ {T} V \cdot s\right). \tag {9} +$$ + +In this formulation, the scaling factor $s$ adjusts the magnitude of the result from the matrix product, while all other operations in VSSA—such as the attention calculation—are performed using addition, in line with the spike-based computation model. + +However, simply converting the queries, keys, and values into spike form prior to performing matrix multiplication does not fully leverage the advantages of SNNs, which thrive on addition-only and event-driven computations. It is unnecessary to convert all of $Q$ , $K$ , and $V$ into binary spike sequences to transition from matrix multiplication to addition. Furthermore, converting all of the values to binary spikes can result in significant information loss. Thus, the challenge remains to strike a balance between maintaining the richness of the data while taking full advantage of the addition operation-based nature of SNNs. + +![](images/2cfcd6b4c360199ad8437603374939eff8452fe837c2b75f5c0449d0acc9189a.jpg) +Figure 2. The overview of Spiking Transformer. + +# 4. Method + +# 4.1. Overall Architecture + +Figure 2 illustrates the architecture of our Spiking Transformer, which comprises four primary components: Spiking Patch Splitting (SPS), Accurate Addition-Only Spiking Self-Attention $(\mathrm{A}^2\mathrm{OS}^2\mathrm{A})$ , Multi-Layer Perceptron (MLP), and a linear classification head. The design follows the approach outlined in [58]. For the SPS module, we adopt the structure from [63, 64]. Given a 2D image sequence $I \in \mathbb{R}^{T \times C \times H \times W}$ , the Patch Splitting Module (PSM) performs a linear projection and splits the input into a sequence of $N$ flattened spike patches with $D$ -dimensional channels, where $T$ denotes the number of timesteps (images are repeated $T$ times in the dataset), $C$ is the number of channels, and $H$ and $W$ represent the height and width of the image sequence. Additionally, a convolutional layer generates Relative Position Embeddings (RPE) as described in [58]. Together, the SPS part is written as: + +$$ +u = \operatorname {P S M} (I), I \in \mathbb {R} ^ {T \times C \times H \times W}, x \in \mathbb {R} ^ {T \times N \times D}, \tag {10} +$$ + +$$ +s = \mathcal {S N} (u), s \in \mathbb {R} ^ {T \times N \times D}, \tag {11} +$$ + +$$ +\mathrm {R P E} = \mathrm {B N} ((\operatorname {C o n v 2 d} (s))), \mathrm {R P E} \in \mathbb {R} ^ {T \times N \times D}, \tag {12} +$$ + +$$ +U _ {0} = u + \operatorname {R P E}, U _ {0} \in \mathbb {R} ^ {T \times N \times D}, \tag {13} +$$ + +where $u$ and $U_{0}$ are the output membrane potential tensors from the PSM and SPS modules, respectively, and $\mathcal{SN}(\cdot)$ + +represents the spike neuron layer. The output $U_{0}$ is then passed to the $L$ -block Spiking Transformer encoder, which consists of the $\mathrm{A}^{2}\mathrm{OS}^{2}\mathrm{A}$ and MLP blocks. Residual connections are applied to the membrane potentials in both the $\mathrm{A}^{2}\mathrm{OS}^{2}\mathrm{A}$ and MLP blocks. To avoid multi-bit spike outputs, we adopt the design from [58, 63], ensuring that the values before the convolutional layers are binary, thus allowing the spike and weight matrix multiplication to be simplified to addition operations. + +The $\mathrm{A}^2\mathrm{OS}^2\mathrm{A}$ mechanism models both local and global information of the input sequence using binary spike $Q$ , full-precision $K$ , and ternary spike $V$ . This approach reduces information loss and eliminates the need for scaling and softmax operations, thereby improving task accuracy. A Global Average Pooling (GAP) is applied to the processed feature from the Spiking Transformer encoder, and the resulting $D$ -dimensional channel is passed to a fully connected Classification Head (CH) for final prediction. The $\mathrm{A}^2\mathrm{OS}^2\mathrm{A}$ , MLP, and CH components are as follows: + +$$ +S _ {0} = \mathcal {S N} (U _ {0}), S _ {0} \in \mathbb {R} ^ {T \times N \times D}, \tag {14} +$$ + +$$ +U _ {l} ^ {\prime} = \mathrm {A} ^ {2} \mathrm {O S} ^ {2} \mathrm {A} \left(S _ {l - 1}\right) + U _ {l - 1}, U _ {l} ^ {\prime} \in \mathbb {R} ^ {T \times N \times D}, l = 1 \dots L, \tag {15} +$$ + +$$ +S _ {l} ^ {\prime} = \mathcal {S N} \left(U _ {l} ^ {\prime}\right), S _ {l} ^ {\prime} \in \mathbb {R} ^ {T \times N \times D}, \tag {16} +$$ + +$$ +S _ {l} = \mathcal {S N} \left(\operatorname {M L P} \left(S _ {l} ^ {\prime}\right) + U _ {l} ^ {\prime}\right), S _ {l} \in \mathbb {R} ^ {T \times N \times D}, l = 1 \dots L, \tag {17} +$$ + +$$ +Y = \mathrm {C H} (\mathrm {G A P} (S _ {L})). \tag {18} +$$ + +# 4.2. Information Loss in Spiking Transformers + +In this section, we analyze the limitations of binary spikes in SNNs with respect to information representation, which motivates the focus of this paper. While binary spikes offer energy efficiency, they inherently suffer from reduced representational capacity, leading to information loss when used for all $Q$ , $K$ , and $V$ in VSSA, which results in accuracy degradation. + +To support this claim, we first perform a theoretical analysis using the concept of entropy. The representational capability $\mathcal{C}(\mathbf{X})$ of a set $\mathbf{X}$ is determined by the maximum entropy of $\mathbf{X}$ , expressed as: + +$$ +\mathcal {C} (\mathbf {X}) = \max \mathcal {H} (\mathbf {X}) = - \sum_ {x \in \mathbf {X}} p _ {\mathbf {X}} (x) \log p _ {\mathbf {X}} (x), \tag {19} +$$ + +where $p_{\mathbf{X}}(x)$ is the probability of a sample $x$ from $\mathbf{X}$ . We now present the following proposition: + +Proposition 1 For a set $\mathbf{X}$ , its representational capacity is $\mathcal{C}(\mathbf{X}) = \max \mathcal{H}(\mathbf{X})$ . When the probability distribution of $\mathbf{X}$ is uniform, i.e., $p_{\mathbf{X}}(x) = \frac{1}{N}$ , where $N$ is the total number of samples in $\mathbf{X}$ , the entropy $\mathcal{H}(\mathbf{X})$ reaches its maximum value of $\log(N)$ . Hence, we conclude that $\mathcal{C}(\mathbf{X}) = \log(N)$ . + +Using Proposition 1, we can evaluate the representational capacity of binary spike layers in SNNs and compare them with real-valued layers in ANNs. Let $\mathbf{F}_{\mathrm{B}} \in \mathbb{B}^{C \times H \times W}$ denote the binary spike layer of the SNN, and $\mathbf{F}_{\mathrm{R}} \in \mathbb{R}^{C \times H \times W}$ denote the real-valued feature map in the corresponding ANN. For a binary spike output $s$ , it requires 1 bit, and thus the number of possible samples from $s$ is 2. Therefore, the number of samples for $\mathbf{F}_{\mathrm{B}}$ is $2^{(C \times H \times W)}$ , leading to: + +$$ +\mathcal {C} \left(\mathbf {F} _ {\mathrm {B}}\right) = \log \left(2 ^ {(C \times H \times W)}\right) = C \times H \times W. \tag {20} +$$ + +In contrast, a real-valued output requires 32 bits per sample, yielding $2^{32}$ possible values. Hence, the representational capacity for real-valued layers is: + +$$ +\mathcal {C} \left(\mathbf {F} _ {\mathrm {R}}\right) = \log \left(2 ^ {3 2 \times (C \times H \times W)}\right) = 3 2 \times C \times H \times W. \tag {21} +$$ + +This highlights the limited representational capacity of the binary spike layer. By transforming the real-valued $Q_{\mathbb{R}}$ , $K_{\mathbb{R}}$ , $V_{\mathbb{R}}$ into binary $Q_{\mathbb{B}}$ , $K_{\mathbb{B}}$ , $V_{\mathbb{B}}$ , significant information loss occurs, reducing accuracy. Therefore, we propose increasing the bit precision of $Q$ , $K$ , and $V$ in spiking attention to minimize information loss, while preserving the advantages of multiplication-addition transformations. + +# 4.3. Accurate Addition-Only Spiking Self-Attention Mechanism (A $^2$ OS $^2$ A) + +To maintain the integrity of information within spiking attention while leveraging the advantages of multiplication-addition transformations, we propose the Accurate Addition-Only Spiking Self-Attention $(\mathrm{A}^2\mathrm{OS}^2\mathrm{A})$ . This mechanism + +employs binary $Q$ , full-precision $K$ , and ternary $V$ in the self-attention process. Specifically, we utilize a binary spiking neuron after the linear layer to produce $Q$ , a ReLU function for generating $K$ , and a ternary spiking neuron for producing $V$ , as illustrated in Figure 1(c). The formulations for $Q, K, V$ in $\mathrm{A}^2\mathrm{OS}^2\mathrm{A}$ are defined as follows: + +$$ +Q = \mathcal {S N} _ {Q} ^ {b} (\mathrm {B N} (X W _ {Q})), \tag {22} +$$ + +$$ +K = \operatorname {R e L U} _ {K} (\operatorname {B N} (X W _ {K})), \tag {23} +$$ + +$$ +V = \mathcal {S N} _ {V} ^ {t} (\mathrm {B N} (X W _ {V})), \tag {24} +$$ + +where $\mathcal{SN}^b (\cdot)$ denotes the binary spiking neuron as used in standard SNNs, $\mathrm{ReLU}_K(\cdot)$ is the ReLU activation function typical in ANNs, and $\mathcal{SN}^t (\cdot)$ represents a ternary spiking neuron that outputs values in the set $\{-1,0,1\}$ . + +The evolution of the membrane potential and the spike generation mechanism within $\mathcal{SN}^t(\cdot)$ can be described by the following equations: + +$$ +U [ t ] = H [ t - 1 ] + X [ t ], \tag {25} +$$ + +$$ +S [ t ] = \operatorname {H e a} (| U [ t ] | - V _ {\mathrm {t h}}), \tag {26} +$$ + +$$ +H [ t ] = V _ {\text {r e s e t}} S [ t ] + \beta U [ t ] (1 - | S [ t ] |). \tag {27} +$$ + +The formulation of the $\mathrm{A}^2\mathrm{OS}^2\mathrm{A}$ is expressed as: + +$$ +\mathrm {A} ^ {2} \mathrm {O S} ^ {2} \mathrm {A} (Q, K, V) = \mathcal {S N} \left(Q \cdot K ^ {T} \cdot V\right). \tag {28} +$$ + +In $\mathrm{A}^2\mathrm{OS}^2\mathrm{A}$ , since $Q$ takes values in $\{0,1\}$ and $K$ is in $\mathbb{R}$ , the matrix multiplication $Q\cdot K^{T}$ can be transformed into equivalent addition operations, resulting in $Q\cdot K^{T}\in \mathbb{R}$ . Moreover, because both $Q$ and $K$ are naturally non-negative, this preserves a non-negative attention map $Q\cdot K^T$ . Consequently, the mechanism eliminates the need for a softmax operation to enforce non-negativity within the attention map. Furthermore, since $K$ is a real-valued vector without boundaries, there is no necessity for a scaling factor $s$ to manage large values resulting from the matrix multiplication, unlike the approaches in [64, 65]. Additionally, given that $V$ can also take values in $\{-1,0,1\}$ , the matrix product $Q\cdot K^{T}\cdot V$ can similarly be expressed through addition operations. At the same time, unlike other matrix multiplication methods outlined in [64, 65], our framework allows for the result to include negative values, akin to what is seen in traditional ANNs. + +In summary, the $\mathrm{A}^2\mathrm{OS}^2\mathrm{A}$ mechanism utilizes full-precision $K$ and ternary spike $V$ to enhance the representational capability of self-attention and minimize information loss. Additionally, all attention calculations are performed using addition operations, which align with the characteristics of SNNs. + +# 5. Experiment + +We evaluate our method on various datasets, including CIFAR-10/100 [31] and ImageNet-1K [8]. The network + +Table 1. Ablation study for Spiking Transformer on CIFAR10/100. Param refers to the number of parameters. Spiking Transformer- $L - D$ is a Spiking Transformer model with $L$ Spiking Transformer encoder blocks and $D$ feature embedding dimensions. + +
MethodsParam (M)Time StepCIFAR10 AccCIFAR100 Acc
Baseline-2-2562.59494.3976.00
Spiking Transformer-2-2562.59494.9176.96
Baseline-2-51210.23495.5178.83
Spiking Transformer-2-51210.23496.4279.90
+ +architecture and experimental setup follow the baseline from Spike-driven Transformer [58]. Further details regarding the experimental settings can be found in the Spike-driven Transformer [58]. + +# 5.1. Ablation Study + +We conduct a series of ablation experiments to assess the effectiveness of the proposed Spiking Transformer, comparing it with the baseline model from [58] on the CIFAR-10 and CIFAR-100 datasets. The results are summarized in Table 1. + +The baseline models used are Spike-driven Transformer-2-256 and Spike-driven Transformer-2-512. For the Transformer-2-256 with 4 timesteps, the baseline accuracy is $94.39\%$ on CIFAR-10 and $76.00\%$ on CIFAR-100, consistent with previous results. Our Spiking Transformer, using the same architecture, achieves notable improvements, with accuracies of $94.91\%$ and $76.96\%$ on CIFAR-10 and CIFAR-100, respectively—resulting in performance gains of approximately $0.5\%$ and $1.0\%$ . For the larger model, Transformer-2-512 with 4 timesteps achieves baseline accuracies of $95.51\%$ and $78.43\%$ on CIFAR-10 and CIFAR-100. Our model with the same architecture shows substantial improvements, reaching accuracy scores of $96.42\%$ and $79.90\%$ , corresponding to gains of approximately $0.9\%$ and $1.1\%$ on CIFAR-10 and CIFAR-100, respectively. + +# 5.2. CIFAR + +To further evaluate the performance of our method, we compare it with recent SNN-based Transformer approaches, including Spikformer [64], Spikingformer [63] and so on, using a range of models with varying embedding dimensions and numbers of transformer blocks on the CIFAR dataset. The CIFAR dataset consists of 50,000 training images and 10,000 test images, all with a resolution of $32 \times 32$ . We maintain the experimental setup from [58, 63, 64], including network architecture, training configurations, etc. Table 2 presents the accuracy of our method compared to other models on the CIFAR dataset. As shown in Table 2, our Spiking Transformer-2-384 achieves an accuracy of $95.70\%$ on the CIFAR-10 dataset, surpassing TET $(94.44\%)$ and even + +Table 2. Comparison of the performance between Spiking Transformer and existing approaches on CIFAR10/100. + +
MethodsParam (M)Time StepCIFAR10 AccCIFAR100 Acc
Hybrid training[46]9.2712592.2267.87
Diet-SNN[45]0.2710/592.5464.07
STBP[55]17.541289.83-
STBP NeuNorm[54]17.541290.53-
TSSL-BP[60]17.54591.41-
STBP-tdBN[62]12.63492.9270.86
TET[9]12.63494.4474.47
Spikformer-4-256[64]4.15493.9475.96
Spikformer-2-384[64]5.76494.8076.95
Spikformer-4-384[64]9.32495.1977.86
Spikingformer-4-256[63]4.15494.7777.43
Spikingformer-2-384[63]5.76495.2278.34
Spikingformer-4-384[63]9.32495.6179.09
Spiking Transformer-4-2564.15494.9677.49
Spiking Transformer-2-3845.76495.7078.59
Spiking Transformer-4-3849.32496.3279.69
Spiking Transformer-2-51210.23496.4279.90
+ +Spikingformer-4-384 (94.97%). Notably, performance improves further with higher embedding dimensions and more transformer blocks. In particular, Spiking Transformer-4-384 shows a 1.13% improvement over Spikformer-4-384 and a 0.71% improvement over Spikingformer-4-384. The performance gain of our Spiking Transformer is even more pronounced on more complex datasets such as CIFAR-100. Specifically, Spiking Transformer-4-384 achieves a notable improvement of 1.83% over Spikformer-4-384. + +# 5.3. ImageNet + +We further evaluated our approach on the challenging ImageNet-1K dataset. ImageNet-1K [8] is a widely-used benchmark for image classification, consisting of 1,000 categories. The dataset includes approximately 1.28 million training images and 50,000 test images. For both training and evaluation, the images are resized to a default resolution of $224 \times 224$ pixels. To ensure consistency with previous work, we maintain the experimental setup from [58, 63, 64], including network architecture, training configurations, and other experimental details. + +Table 3 presents a comparative analysis of recent SoTA methods evaluated on the ImageNet dataset. Notable models in this domain include Spikformer [64], Spikingformer [63], and Spike-driven Transformer [58], which achieve accuracies of $73.38\%$ , $74.79\%$ , and $74.57\%$ , respectively, with Transformer-8-512 as the baseline configuration. In comparison, our approach demonstrates a significant improvement, achieving an accuracy of $76.28\%$ , which represents a $1.49\%$ increase over the best-performing transformer-based SNN model, Spikingformer. This improvement highlights the + +Table 3. Comparison of the performance between Spiking Transformer and existing approaches on ImageNet-1k. + +
CategoryMethodsArchitectureParam (M)Time StepAcc
ANN-to-SNNHybrid training[46]ResNet-3421.7925061.48
Spiking ResNet[27]ResNet-3421.7935071.61
ResNet-5025.5635072.75
QCFS[2]VGG-16138.426472.85
COS[24]ResNet-3421.79874.17
Directly LearningTET[9]Spiking-ResNet-3421.79664.79
SEW-ResNet-3421.79468.00
STBP-tdBN[62]Spiking-ResNet-3421.79663.72
SEW-ResNet-3421.79467.04
SEW ResNet[11]SEW-ResNet-5025.56467.78
SEW-ResNet-10144.55468.76
SEW-ResNet-15260.19469.26
Attention-SNN[57]ResNet-10478.37477.08
Directly LearningSpikformer[64]Spikformer-8-38416.81470.24
Spikformer-6-51223.37472.46
Spikformer-8-51229.68473.38
Spikformer-10-51236.01473.68
Spikformer-8-76866.34474.81
Directly LearningSpikingformer[63]Spikingformer-8-38416.81472.45
Spikingformer-8-51229.68474.79
Spikingformer-8-76866.34475.85
Directly LearningSpike-driven Transformer[58]Spike-driven Transformer-8-38416.81472.28
Spike-driven Transformer-6-51223.37474.11
Spike-driven Transformer-8-51229.68474.57
Spike-driven Transformer-10-51236.01474.66
Spike-driven Transformer-8-76866.34477.07
Directly LearningSpiking TransformerSpiking Transformer-8-38416.81474.04
Spiking Transformer-6-51223.37476.22
Spiking Transformer-8-51229.68476.28
Spiking Transformer-10-51236.01478.66
+ +effectiveness of our method in leveraging the potential of spiking neuron models within the transformer architecture, resulting in enhanced performance on large-scale datasets like ImageNet-1K. + +Furthermore, even with a smaller network architecture, our Spiking Transformer-10-512 model outperforms the other methods, including those using Transformer-8-768 as the baseline. This demonstrates that our approach is not only more accurate but also more efficient, as it achieves superior results with fewer parameters, highlighting its scalability and robustness. + +# 6. Conclusion + +In this paper, we have introduced the Accurate Addition-Only Spiking Self-Attention $(\mathrm{A}^2\mathrm{OS}^2\mathrm{A})$ mechanism as a novel method for integrating Spiking Neural Networks with + +Transformer models. Our approach tackles the inherent challenge of adapting the self-attention mechanism to the spiking paradigm by proposing a hybrid solution that combines binary, ReLU, and ternary spiking neurons. This innovative hybrid strategy enables the elimination of both softmax and scaling operations, which are traditionally essential in Transformer models, while preserving the fundamental advantages of addition-only computations in SNNs. + +Through extensive experiments, we have demonstrated that $\mathrm{A}^2\mathrm{OS}^2\mathrm{A}$ outperforms existing SNN-based Transformer models in terms of both accuracy and efficiency. Specifically, our method achieves a SoTA accuracy of $78.66\%$ on the ImageNet-1K dataset. This performance, combined with the energy-efficient nature of SNNs, makes $\mathrm{A}^2\mathrm{OS}^2\mathrm{A}$ a promising candidate for real-world applications, especially those requiring high computational performance with low energy consumption. + +# Acknowledgment + +This work is supported by grants from the National Natural Science Foundation of China under contracts No.12202412 and No.12202413. + +# References + +[1] Filipp Akopyan, Jun Sawada, Andrew Cassidy, Rodrigo Alvarez-Icaza, John Arthur, Paul Merolla, Nabil Imam, Yutaka Nakamura, Pallab Datta, Gi-Joon Nam, et al. Truenorth: Design and tool flow of a 65 mw 1 million neuron programmable neurosynaptic chip. IEEE transactions on computer-aided design of integrated circuits and systems, 34(10):1537-1557, 2015. 1 +[2] Tong Bu, Wei Fang, Jianhao Ding, PengLin Dai, Zhaofei Yu, and Tiejun Huang. Optimal ann-snn conversion for high-accuracy and ultra-low-latency spiking neural networks. In International Conference on Learning Representations, 2021. 2, 8 +[3] Tong Bu, Jianhao Ding, Zhaofei Yu, and Tiejun Huang. Optimized potential initialization for low-latency spiking neural networks. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 11-20, 2022. 2 +[4] Natalia Caporale and Yang Dan. Spike timing-dependent plasticity: a hebbian learning rule. Annu. Rev. Neurosci., 31 (1):25-46, 2008. 2 +[5] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 3 +[6] Chun-Fu Richard Chen, Quanfu Fan, and Rameswar Panda. Crossvit: Cross-attention multi-scale vision transformer for image classification. In Proceedings of the IEEE/CVF international conference on computer vision, pages 357–366, 2021. 3 +[7] Mike Davies, Narayan Srinivasa, Tsung-Han Lin, Gautham Chinya, Yongqiang Cao, Sri Harsha Choday, Georgios Dimou, Prasad Joshi, Nabil Imam, Shweta Jain, et al. Loihi: A neuromorphic manycore processor with on-chip learning. *IEEE Micro*, 38(1):82-99, 2018. 1 +[8] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 6, 7 +[9] Shikuang Deng, Yuhang Li, Shanghang Zhang, and Shi Gu. Temporal efficient training of spiking neural network via gradient re-weighting. arXiv preprint arXiv:2202.11946, 2022.7, 8 +[10] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2 +[11] Wei Fang, Zhaofei Yu, Yanqi Chen, Tiejun Huang, Timothee Masquelier, and Yonghong Tian. Deep residual learning in spiking neural networks. Advances in Neural Information Processing Systems, 34:21056-21069, 2021. 2, 8 + +[12] Yufei Guo, Weihang Peng, Xiaode Liu, Yuanpei Chen, Yuhan Zhang, Xin Tong, Zhou Jie, and Zhe Ma. Enof-snn: Training accurate spiking neural networks via enhancing the output feature. In The Thirty-eighth Annual Conference on Neural Information Processing Systems. 2 +[13] Yufei Guo, Yuanpei Chen, Liwen Zhang, Xiaode Liu, Yinglei Wang, Xuhui Huang, and Zhe Ma. IM-loss: Information maximization loss for spiking neural networks. In Advances in Neural Information Processing Systems, 2022. +[14] Yufei Guo, Yuanpei Chen, Liwen Zhang, YingLei Wang, Xiaode Liu, Xinyi Tong, Yuanyuan Ou, Xuhui Huang, and Zhe Ma. Reducing information loss for spiking neural networks. In Computer Vision - ECCV 2022, pages 36-52, Cham, 2022. Springer Nature Switzerland. 2 +[15] Yufei Guo, Xinyi Tong, Yuanpei Chen, Liwen Zhang, Xiaode Liu, Zhe Ma, and Xuhui Huang. Recdis-snn: Rectifying membrane potential distribution for directly training spiking neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 326-335, 2022. 1 +[16] Yufei Guo, Liwen Zhang, Yuanpei Chen, Xinyi Tong, Xiaode Liu, YingLei Wang, Xuhui Huang, and Zhe Ma. Real spike: Learning real-valued spikes for spiking neural networks. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XII, pages 52-68. Springer, 2022. 1 +[17] Yufei Guo, Xiaode Liu, Yuanpei Chen, Liwen Zhang, Weihang Peng, Yuhan Zhang, Xuhui Huang, and Zhe Ma. Rmploss: Regularizing membrane potential distribution for spiking neural networks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17391-17401, 2023. 2 +[18] Yufei Guo, Weihang Peng, Yuanpei Chen, Liwen Zhang, Xiaode Liu, Xuhui Huang, and Zhe Ma. Joint a-snn: Joint training of artificial and spiking neural networks via self-distillation and weight factorization. Pattern Recognition, 142:109639, 2023. 2 +[19] Yufei Guo, Yuhan Zhang, Yuanpei Chen, Weihang Peng, Xiaode Liu, Liwen Zhang, Xuhui Huang, and Zhe Ma. Membrane potential batch normalization for spiking neural networks. arXiv preprint arXiv:2308.08359, 2023. +[20] Yufei Guo, Yuanpei Chen, Xiaode Liu, Weihang Peng, Yuhan Zhang, Xuhui Huang, and Zhe Ma. Ternary spike: Learning ternary spikes for spiking neural networks. In Proceedings of the AAAI conference on artificial intelligence, pages 12244-12252, 2024. +[21] Yufei Guo, Yuanpei Chen, Zecheng Hao, Weihang Peng, Zhou Jie, Yuhan Zhang, Xiaode Liu, and Zhe Ma. Take a shortcut back: Mitigating the gradient vanishing for training spiking neural networks. Advances in Neural Information Processing Systems, 37:24849-24867, 2025. 2 +[22] Bing Han and Kaushik Roy. Deep spiking neural network: Energy efficiency through time based coding. In European Conference on Computer Vision, pages 388-404. Springer, 2020. 2 +[23] Zecheng Hao, Tong Bu, Jianhao Ding, Tiejun Huang, and Zhaofei Yu. Reducing ann-snn conversion error through resid + +ual membrane potential. arXiv preprint arXiv:2302.02091, 2023. +[24] Zecheng Hao, Jianhao Ding, Tong Bu, Tiejun Huang, and Zhaofei Yu. Bridging the gap between anns and snns by calibrating offset spikes. arXiv preprint arXiv:2302.10685, 2023. 8 +[25] Nguyen-Dong Ho and Ik-Joon Chang. Tcl: an ann-to-snn conversion with trainable clipping layers. In 2021 58th ACM/IEEE Design Automation Conference (DAC), pages 793-798. IEEE, 2021. 2 +[26] Allan L Hodgkin and Andrew F Huxley. Currents carried by sodium and potassium ions through the membrane of the giant axon of loligo. The Journal of physiology, 116(4):449, 1952. 2 +[27] Yangfan Hu, Huajin Tang, and Gang Pan. Spiking deep residual networks. IEEE Transactions on Neural Networks and Learning Systems, 34(8):5200-5205, 2021. 8 +[28] Asifullah Khan, Zunaira Rauf, Abdul Rehman Khan, Saima Rathore, Saddam Hussain Khan, Najmus Saher Shah, Umair Farooq, Hifsa Asif, Aqsa Asif, Umme Zahoora, et al. A recent survey of vision transformers for medical image segmentation. arXiv preprint arXiv:2312.00634, 2023. 2 +[29] Seijoon Kim, Seongsik Park, Byunggook Na, and Sungroh Yoon. Spiking-yolo: Spiking neural network for energy-efficient object detection, 2019. 1 +[30] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 1 +[31] Alex Krizhevsky, Vinod Nair, and Geoffrey Hinton. Cifar-10 (canadian institute for advanced research). URL http://www.cs.toronto.edu/kriz/cifar.html, 5(4):1, 2010. 6 +[32] Yuxiang Lan, Yachao Zhang, Xu Ma, Yanyun Qu, and Yun Fu. Efficient converted spiking neural network for 3d and 2d classification. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9211-9220, 2023. 2 +[33] Nathan Leroux, Jan Finkbeiner, and Emre Neftci. Online transformers with spiking neurons for fast prosthetic hand control. In 2023 IEEE Biomedical Circuits and Systems Conference (BioCAS), pages 1-6. IEEE, 2023. 3 +[34] Yuhang Li, Shikuang Deng, Xin Dong, Ruihao Gong, and Shi Gu. A free lunch from ann: Towards efficient, accurate spiking neural networks calibration. In International Conference on Machine Learning, pages 6316-6325. PMLR, 2021. 1, 2 +[35] Yuhang Li, Yufei Guo, Shanghang Zhang, Shikuang Deng, Yongqing Hai, and Shi Gu. Differentiable spike: Rethinking gradient-descent for training spiking neural networks. Advances in Neural Information Processing Systems, 34:23426-23439, 2021. 2 +[36] Yawei Li, Kai Zhang, Jiezhang Cao, Radu Timofte, and Luc Van Gool. Localvit: Bringing locality to vision transformers. arXiv preprint arXiv:2104.05707, 2021. 3 +[37] Yudong Li, Yunlin Lei, and Xu Yang. Spikeformer: a novel architecture for training high-performance low-latency spiking neural network. arXiv preprint arXiv:2211.10686, 2022.3 + +[38] Yanghao Li, Hanzi Mao, Ross Girshick, and Kaiming He. Exploring plain vision transformer backbones for object detection. In European conference on computer vision, pages 280-296. Springer, 2022. 2 +[39] De Ma, Juncheng Shen, Zonghua Gu, Ming Zhang, Xiaolei Zhu, Xiaoqiang Xu, Qi Xu, Yangjing Shen, and Gang Pan. Darwin: A neuromorphic hardware co-processor based on spiking neural networks. Journal of Systems Architecture, 77: 43-51, 2017. 1 +[40] Wolfgang Maass. Networks of spiking neurons: the third generation of neural network models. Neural networks, 10 (9):1659-1671, 1997. 2, 4 +[41] ZHI Min and LU Jingfang. A review of vision transformer for image classification. Journal of Zhengzhou University: Engineering Science, 45(4), 2024. 2 +[42] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4195-4205, 2023. 1 +[43] Jing Pei, Lei Deng, Sen Song, Mingguo Zhao, Youhui Zhang, Shuang Wu, Guanrui Wang, Zhe Zou, Zhenzhi Wu, Wei He, et al. Towards artificial general intelligence with hybrid tianjin chip architecture. Nature, 572(7767):106-111, 2019. 1 +[44] Jinye Qu, Zeyu Gao, Tielin Zhang, Yanfeng Lu, Huajin Tang, and Hong Qiao. Spiking neural network for ultra-low-latency and high-accurate object detection, 2023. 1 +[45] Nitin Rathi and Kaushik Roy. Diet-snn: Direct input encoding with leakage and threshold optimization in deep spiking neural networks. arXiv preprint arXiv:2008.03658, 2020. 7 +[46] Nitin Rathi, Gopalakrishnan Srinivasan, Priyadarshini Panda, and Kaushik Roy. Enabling deep spiking neural networks with hybrid conversion and spike timing dependent backpropagation. arXiv preprint arXiv:2005.01807, 2020. 7, 8 +[47] Dayong Ren, Zhe Ma, Yuanpei Chen, Weihang Peng, Xiaode Liu, Yuhan Zhang, and Yufei Guo. Spiking pointnet: Spiking neural networks for point clouds. arXiv preprint arXiv:2310.06232, 2023. 2 +[48] Abhronil Sengupta, Yuting Ye, Robert Wang, Chiao Liu, and Kaushik Roy. Going deeper in spiking neural networks: Vgg and residual architectures. Frontiers in neuroscience, 13:95, 2019. 2 +[49] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In International conference on machine learning, pages 10347-10357. PMLR, 2021. 3 +[50] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 1 +[51] Shuai Wang, Malu Zhang, Dehao Zhang, Ammar Belatreche, Yichen Xiao, Yu Liang, Yimeng Shan, Qian Sun, Enqi Zhang, and Yang Yang. Spiking vision transformer with saccadic attention. arXiv preprint arXiv:2502.12677, 2025. 3 +[52] Wenjie Wei, Malu Zhang, Jilin Zhang, Ammar Belatreche, Jibin Wu, Zijing Xu, Xuerui Qiu, Hong Chen, Yang Yang, and Haizhou Li. Event-driven learning for spiking neural networks. arXiv preprint arXiv:2403.00270, 2024.1 + +[53] Yujie Wu, Lei Deng, Guoqi Li, Jun Zhu, and Luping Shi. Spatio-temporal backpropagation for training high-performance spiking neural networks. Frontiers in neuroscience, 12:331, 2018. 2 +[54] Yujie Wu, Lei Deng, Guoqi Li, Jun Zhu, Yuan Xie, and Luping Shi. Direct training for spiking neural networks: Faster, larger, better. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1311-1318, 2019. 7 +[55] Y. et al. Wu. Spatio-temporal learning with spiking neural networks. Nature Communications, 9, 2018. 7 +[56] Mingqing Xiao, Qingyan Meng, Zongpeng Zhang, Yisen Wang, and Zhouchen Lin. Training feedback spiking neural networks by implicit differentiation on the equilibrium state. Advances in Neural Information Processing Systems, 34:14516-14528, 2021. 1 +[57] Man Yao, Guangshe Zhao, Hengyu Zhang, Yifan Hu, Lei Deng, Yonghong Tian, Bo Xu, and Guoqi Li. Attention spiking neural networks. IEEE transactions on pattern analysis and machine intelligence, 45(8):9393-9410, 2023. 8 +[58] Man Yao, Jiakui Hu, Zhaokun Zhou, Li Yuan, Yonghong Tian, Bo Xu, and Guoqi Li. Spike-driven transformer. Advances in neural information processing systems, 36, 2024. 1, 3, 5, 7, 8 +[59] Malu Zhang, Jiadong Wang, Jibin Wu, Ammar Belatreche, Burin Amornpaisannon, Zhixuan Zhang, Venkata Pavan Kumar Miriyala, Hong Qu, Yansong Chua, Trevor E Carlson, et al. Rectified linear postsynaptic potential function for backpropagation in deep spiking neural networks. IEEE transactions on neural networks and learning systems, 33 (5):1947-1958, 2021. 2 +[60] Wenrui Zhang and Peng Li. Temporal spike sequence learning via backpropagation for deep spiking neural networks. Advances in Neural Information Processing Systems, 33:12022-12033, 2020. 7 +[61] Yuhan Zhang, Xiaode Liu, Yuanpei Chen, Weihang Peng, Yufei Guo, Xuhui Huang, and Zhe Ma. Enhancing representation of spiking neural networks via similarity-sensitive contrastive learning. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 16926-16934, 2024. 2 +[62] Hanle Zheng, Yujie Wu, Lei Deng, Yifan Hu, and Guoqi Li. Going deeper with directly-trained larger spiking neural networks. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 11062-11070, 2021. 7, 8 +[63] Chenlin Zhou, Liutao Yu, Zhaokun Zhou, Zhengyu Ma, Han Zhang, Huihui Zhou, and Yonghong Tian. Spikingformer: Spike-driven residual learning for transformer-based spiking neural network. arXiv preprint arXiv:2304.11954, 2023. 1, 3, 5, 7, 8 +[64] Zhaokun Zhou, Yuesheng Zhu, Chao He, Yaowei Wang, Shuicheng Yan, Yonghong Tian, and Li Yuan. Spikformer: When spiking neural network meets transformer. arXiv preprint arXiv:2209.15425, 2022. 1, 3, 4, 5, 6, 7, 8 +[65] Zhaokun Zhou, Kaiwei Che, Wei Fang, Keyu Tian, Yuesheng Zhu, Shuicheng Yan, Yonghong Tian, and Li Yuan. Spikformer v2: Join the high accuracy club on imagenet with an snn ticket. arXiv preprint arXiv:2401.02020, 2024. 1, 3, 4, 6 +[66] Shihao Zou, Yuxuan Mu, Xinxin Zuo, Sen Wang, and Li Cheng. Event-based human pose tracking by spiking spatiotemporal transformer, 2023. 1 \ No newline at end of file diff --git a/2025/Spiking Transformer_ Introducing Accurate Addition-Only Spiking Self-Attention for Transformer/images.zip b/2025/Spiking Transformer_ Introducing Accurate Addition-Only Spiking Self-Attention for Transformer/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..bf1cf179bd912c33e0004e410c0641cef959a918 --- /dev/null +++ b/2025/Spiking Transformer_ Introducing Accurate Addition-Only Spiking Self-Attention for Transformer/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fceed628642eb6fc45f688578b63b1d40daf3e76a0c67761bea23259d3585bb0 +size 542408 diff --git a/2025/Spiking Transformer_ Introducing Accurate Addition-Only Spiking Self-Attention for Transformer/layout.json b/2025/Spiking Transformer_ Introducing Accurate Addition-Only Spiking Self-Attention for Transformer/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..d7e6f33c8b6ac5347934e830c996aebd83184637 --- /dev/null +++ b/2025/Spiking Transformer_ Introducing Accurate Addition-Only Spiking Self-Attention for Transformer/layout.json @@ -0,0 +1,11446 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 69, + 103, + 541, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 103, + 541, + 140 + ], + "spans": [ + { + "bbox": [ + 69, + 103, + 541, + 140 + ], + "type": "text", + "content": "Spiking Transformer: Introducing Accurate Addition-Only Spiking Self-Attention for Transformer" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 118, + 160, + 496, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 160, + 496, + 190 + ], + "spans": [ + { + "bbox": [ + 118, + 160, + 496, + 190 + ], + "type": "text", + "content": "Yufei Guo,\\* Xiaode Liu,\\* Yuanpei Chen, Weihang Peng, Yuhan Zhang, Zhe Ma\\* Intelligent Science & Technology Academy of CASIC, China" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 162, + 192, + 444, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 162, + 192, + 444, + 202 + ], + "spans": [ + { + "bbox": [ + 162, + 192, + 444, + 202 + ], + "type": "text", + "content": "yfguo@pku.edu.cn, lxde@pku.edu.cn, mazhe_thu@163.com" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "spans": [ + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 257, + 296, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 257, + 296, + 555 + ], + "spans": [ + { + "bbox": [ + 56, + 257, + 296, + 555 + ], + "type": "text", + "content": "Transformers have demonstrated outstanding performance across a wide range of tasks, owing to their self-attention mechanism, but they are highly energy-consuming. Spiking Neural Networks have emerged as a promising energy-efficient alternative to traditional Artificial Neural Networks, leveraging event-driven computation and binary spikes for information transfer. The combination of Transformers' capabilities with the energy efficiency of SNNs offers a compelling opportunity. This paper addresses the challenge of adapting the self-attention mechanism of Transformers to the spiking paradigm by introducing a novel approach: Accurate Addition-Only Spiking Self-Attention (" + }, + { + "bbox": [ + 56, + 257, + 296, + 555 + ], + "type": "inline_equation", + "content": "A^2OS^2A" + }, + { + "bbox": [ + 56, + 257, + 296, + 555 + ], + "type": "text", + "content": "). Unlike existing methods that rely solely on binary spiking neurons for all components of the self-attention mechanism, our approach integrates binary, ReLU, and ternary spiking neurons. This hybrid strategy significantly improves accuracy while preserving non-multiplicative computations. Moreover, our method eliminates the need for softmax and scaling operations. Extensive experiments show that the " + }, + { + "bbox": [ + 56, + 257, + 296, + 555 + ], + "type": "inline_equation", + "content": "A^2OS^2A" + }, + { + "bbox": [ + 56, + 257, + 296, + 555 + ], + "type": "text", + "content": "-based Spiking Transformer outperforms existing SNN-based Transformers on several datasets, even achieving an accuracy of " + }, + { + "bbox": [ + 56, + 257, + 296, + 555 + ], + "type": "inline_equation", + "content": "78.66\\%" + }, + { + "bbox": [ + 56, + 257, + 296, + 555 + ], + "type": "text", + "content": " on ImageNet-1K. Our work represents a significant advancement in SNN-based Transformer models, offering a more accurate and efficient solution for real-world applications." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 580, + 135, + 593 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 580, + 135, + 593 + ], + "spans": [ + { + "bbox": [ + 56, + 580, + 135, + 593 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 600, + 297, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 600, + 297, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 600, + 297, + 685 + ], + "type": "text", + "content": "Spiking Neural Networks (SNNs) have gained significant traction as an efficient neural network model, finding applications in diverse areas such as object recognition [15, 16, 34, 56], object detection [29, 44], and pose estimation [66]. These networks utilize binary spike signals for information transfer. A spiking neuron generates a spike, denoted by 1, when its membrane potential surpasses a cer-" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 232, + 555, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 232, + 555, + 388 + ], + "spans": [ + { + "bbox": [ + 313, + 232, + 555, + 388 + ], + "type": "text", + "content": "tain threshold; conversely, it does not spike, represented by 0, when the threshold is not met. This distinctive approach to information processing is notably energy-efficient, as it substitutes the complex multiplications involved in weight and activation interactions with simpler addition operations. Furthermore, SNNs can be effectively implemented using event-driven computations on neuromorphic hardware [1, 7, 39, 43, 52]. In this setup, the computational units are activated solely upon the occurrence of a spike, conserving energy by remaining inactive in the absence of spikes. Research indicates that SNNs can achieve significantly greater energy savings compared to their Artificial Neural Network (ANN) counterparts [1, 7]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 391, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 391, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 391, + 556, + 715 + ], + "type": "text", + "content": "Despite the clear advantages of SNNs in terms of energy efficiency, their practical application is often constrained by limited task accuracy. Meanwhile, Transformers have demonstrated exceptional performance across a wide range of tasks, largely due to their self-attention mechanism [30, 42, 50]. Combining the strengths of Transformers with the energy efficiency of SNNs presents a promising opportunity. However, adapting the self-attention mechanism to SNNs is non-trivial. In the vanilla self-attention framework (VSA), three key components are involved: Query " + }, + { + "bbox": [ + 313, + 391, + 556, + 715 + ], + "type": "inline_equation", + "content": "(Q)" + }, + { + "bbox": [ + 313, + 391, + 556, + 715 + ], + "type": "text", + "content": ", Key " + }, + { + "bbox": [ + 313, + 391, + 556, + 715 + ], + "type": "inline_equation", + "content": "(K)" + }, + { + "bbox": [ + 313, + 391, + 556, + 715 + ], + "type": "text", + "content": ", and Value " + }, + { + "bbox": [ + 313, + 391, + 556, + 715 + ], + "type": "inline_equation", + "content": "(V)" + }, + { + "bbox": [ + 313, + 391, + 556, + 715 + ], + "type": "text", + "content": ". As illustrated in Figure 1(a), the VSA process begins by computing the dot product of the floating-point representations of " + }, + { + "bbox": [ + 313, + 391, + 556, + 715 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 313, + 391, + 556, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 391, + 556, + 715 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 391, + 556, + 715 + ], + "type": "text", + "content": ", resulting in a matrix. This matrix is then normalized using a softmax function, which involves exponential and division operations, to produce an attention map that determines the weighting of " + }, + { + "bbox": [ + 313, + 391, + 556, + 715 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 313, + 391, + 556, + 715 + ], + "type": "text", + "content": ". These operations in the VSA are incompatible with the operational principles of SNNs, which aim to minimize multiplication. Consequently, to implement a Transformer architecture in SNNs, it is essential to develop a novel, efficient self-attention mechanism that avoids multiplication. Ongoing research efforts are exploring solutions in this domain. Approaches such as Spikformer [64, 65], Spikingformer [63], and Spike-driven Transformer [58] have focused on transforming " + }, + { + "bbox": [ + 313, + 391, + 556, + 715 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 313, + 391, + 556, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 391, + 556, + 715 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 391, + 556, + 715 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 391, + 556, + 715 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 313, + 391, + 556, + 715 + ], + "type": "text", + "content": " into spike representations prior to performing matrix operations. This transformation enables the replacement of matrix multiplications with addition, aligning" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 693, + 135, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 693, + 135, + 703 + ], + "spans": [ + { + "bbox": [ + 66, + 693, + 135, + 703 + ], + "type": "text", + "content": "*Equal contribution." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "type": "text", + "content": "† Corresponding author." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "24398" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 197, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 197, + 83 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 197, + 83 + ], + "type": "text", + "content": "better with the principles of SNNs." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 85, + 297, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 85, + 297, + 360 + ], + "spans": [ + { + "bbox": [ + 55, + 85, + 297, + 360 + ], + "type": "text", + "content": "While previous approaches utilize spike-form " + }, + { + "bbox": [ + 55, + 85, + 297, + 360 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 55, + 85, + 297, + 360 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 85, + 297, + 360 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 85, + 297, + 360 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 85, + 297, + 360 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 55, + 85, + 297, + 360 + ], + "type": "text", + "content": " to avoid multiplications, simply introducing spiking neurons to convert " + }, + { + "bbox": [ + 55, + 85, + 297, + 360 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 55, + 85, + 297, + 360 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 85, + 297, + 360 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 85, + 297, + 360 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 85, + 297, + 360 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 55, + 85, + 297, + 360 + ], + "type": "text", + "content": " into spike form prior to performing matrix multiplications does not fully exploit the advantages of the addition-only operation inherent in SNNs. It is unnecessary to convert all of " + }, + { + "bbox": [ + 55, + 85, + 297, + 360 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 55, + 85, + 297, + 360 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 85, + 297, + 360 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 85, + 297, + 360 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 85, + 297, + 360 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 55, + 85, + 297, + 360 + ], + "type": "text", + "content": " into spike form in order to transition from matrix multiplications to matrix additions. However, this method can lead to significant information loss, as demonstrated in Section 4. To address this issue, we propose the Accurate Addition-Only Spiking Self-Attention (A" + }, + { + "bbox": [ + 55, + 85, + 297, + 360 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 55, + 85, + 297, + 360 + ], + "type": "text", + "content": "OS" + }, + { + "bbox": [ + 55, + 85, + 297, + 360 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 55, + 85, + 297, + 360 + ], + "type": "text", + "content": "A). The distinctions between A" + }, + { + "bbox": [ + 55, + 85, + 297, + 360 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 55, + 85, + 297, + 360 + ], + "type": "text", + "content": "OS" + }, + { + "bbox": [ + 55, + 85, + 297, + 360 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 55, + 85, + 297, + 360 + ], + "type": "text", + "content": "A and the vanilla Spiking Self-Attention (VSSA) are illustrated in Figure 1(b) and Figure 1(c). Unlike VSSA, which applies the same binary spiking neuron to generate " + }, + { + "bbox": [ + 55, + 85, + 297, + 360 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 55, + 85, + 297, + 360 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 85, + 297, + 360 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 85, + 297, + 360 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 85, + 297, + 360 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 55, + 85, + 297, + 360 + ], + "type": "text", + "content": ", our approach employs a binary spiking neuron for " + }, + { + "bbox": [ + 55, + 85, + 297, + 360 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 55, + 85, + 297, + 360 + ], + "type": "text", + "content": ", a ReLU activation for " + }, + { + "bbox": [ + 55, + 85, + 297, + 360 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 85, + 297, + 360 + ], + "type": "text", + "content": ", and a ternary spiking neuron for " + }, + { + "bbox": [ + 55, + 85, + 297, + 360 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 55, + 85, + 297, + 360 + ], + "type": "text", + "content": ". This design allows A" + }, + { + "bbox": [ + 55, + 85, + 297, + 360 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 55, + 85, + 297, + 360 + ], + "type": "text", + "content": "OS" + }, + { + "bbox": [ + 55, + 85, + 297, + 360 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 55, + 85, + 297, + 360 + ], + "type": "text", + "content": "A to retain the multiplication-addition transformation while reducing information loss. Moreover, it eliminates the need for both the scaling and softmax functions, enabling the output of floating-point values similar to the VSA, whereas VSSA produces only integer outputs. The key contributions of this paper are as follows:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 363, + 295, + 590 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 55, + 363, + 295, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 363, + 295, + 434 + ], + "spans": [ + { + "bbox": [ + 55, + 363, + 295, + 434 + ], + "type": "text", + "content": "- We present a theoretical framework for understanding the information loss associated with vanilla Spiking Self-Attention mechanisms. To the best of our knowledge, this is the first detailed examination of this issue within the context of Spiking Self-Attention, which opens the door for future advancements in SNN-based Transformers." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 435, + 295, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 435, + 295, + 518 + ], + "spans": [ + { + "bbox": [ + 55, + 435, + 295, + 518 + ], + "type": "text", + "content": "- We introduce " + }, + { + "bbox": [ + 55, + 435, + 295, + 518 + ], + "type": "inline_equation", + "content": "\\mathrm{A}^2\\mathrm{OS}^2\\mathrm{A}" + }, + { + "bbox": [ + 55, + 435, + 295, + 518 + ], + "type": "text", + "content": ", an innovative and efficient approach that incorporates binary spiking neurons, ReLU activations, and ternary spiking neurons, moving beyond the traditional binary-only spiking neuron model. This method reduces information loss while retaining the advantages of addition-only processing in SNNs. Additionally, it eliminates the need for scaling and softmax functions." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 520, + 295, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 520, + 295, + 590 + ], + "spans": [ + { + "bbox": [ + 55, + 520, + 295, + 590 + ], + "type": "text", + "content": "- Comprehensive experiments demonstrate that our proposed architecture either surpasses or matches the performance of State-of-the-Art (SoTA) SNN-based Transformer models on various datasets. Notably, we achieved an accuracy of " + }, + { + "bbox": [ + 55, + 520, + 295, + 590 + ], + "type": "inline_equation", + "content": "78.66\\%" + }, + { + "bbox": [ + 55, + 520, + 295, + 590 + ], + "type": "text", + "content": " on ImageNet-1K, setting a new benchmark in the SNN domain." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 603, + 142, + 616 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 603, + 142, + 616 + ], + "spans": [ + { + "bbox": [ + 55, + 603, + 142, + 616 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 624, + 197, + 636 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 624, + 197, + 636 + ], + "spans": [ + { + "bbox": [ + 55, + 624, + 197, + 636 + ], + "type": "text", + "content": "2.1. Spiking Neural Networks" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 641, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 297, + 714 + ], + "type": "text", + "content": "SNNs are a class of neural networks that more closely emulate the behavior of biological neurons compared to ANNs. The primary distinction lies in their use of discrete spikes for information transmission, as opposed to the continuous signals employed by ANNs. Foundational work in this area can be traced back to Hodgkin and Huxley [26], who" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 180 + ], + "type": "text", + "content": "developed models to simulate the action potentials of biological neurons. Since then, numerous frameworks and models have been introduced to enhance the capabilities and learning efficiency of SNNs. A notable example is the Leaky Integrate-and-Fire (LIF) model [40], which provides a simple yet effective means of simulating neuronal behavior. More recently, the incorporation of Spike-Timing-Dependent Plasticity (STDP) as a learning rule has enabled SNNs to learn temporal patterns effectively [4]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 182, + 556, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 182, + 556, + 456 + ], + "spans": [ + { + "bbox": [ + 313, + 182, + 556, + 456 + ], + "type": "text", + "content": "In the past decade, there has been a significant increase in interest regarding the training of SNNs using techniques adapted from ANNs, as well as their application to practical tasks. Two prevalent learning paradigms in current SNN research are the conversion from ANN to SNN (ANN2SNN) [2, 3, 22-25, 32, 34, 48] and supervised learning [12-14, 18-21, 35, 47, 59]. The ANN-SNN conversion method involves initially training an ANN and subsequently converting it into a homogeneous SNN by transferring the trained weights and substituting ReLU neurons with temporal spiking neurons. However, this method often fails for neuromorphic datasets, as ReLU neurons do not adequately capture the complex temporal dynamics required for processing sequential information. In contrast, supervised learning [11, 17, 53, 61] employs alternative functions during backpropagation to approximate the firing process, enabling the direct training of SNNs as if they were ANNs. This approach leverages the advantages of gradient-based optimization, achieving impressive performance with just a few time steps, even on large-scale datasets. Furthermore, supervised learning has proven to be effective in handling temporal data, establishing it as a preferred methodology in SNN research." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 468, + 433, + 479 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 468, + 433, + 479 + ], + "spans": [ + { + "bbox": [ + 314, + 468, + 433, + 479 + ], + "type": "text", + "content": "2.2. Vision Transformers" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 486, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 486, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 486, + 556, + 713 + ], + "type": "text", + "content": "Vision Transformers (ViTs) have recently emerged as a powerful alternative to convolutional neural networks (CNNs) for a wide range of visual tasks, demonstrating remarkable efficacy and flexibility. Introduced by Dosovitskiy et al. [10], ViTs leverage the transformer architecture—originally designed for natural language processing—to process images in a novel manner. The architecture innovatively splits images into patches and treats these patches as sequences, enabling the model to capture long-range dependencies and relationships within the visual data, a limitation often encountered in traditional CNNs. The introduction of ViTs has spurred significant advancements in various computer vision tasks, including image classification [41], object detection [38], and segmentation [28]. These advancements highlight the potential of ViTs to tackle complex visual recognition problems with unprecedented accuracy. Notably, it has been shown that with appropriate training techniques and larger datasets, ViTs can not only match but surpass the performance of state-of-the-art CNNs, even in data-scarce scenarios, presenting a" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24399" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 71, + 219, + 284 + ], + "blocks": [ + { + "bbox": [ + 61, + 71, + 219, + 284 + ], + "lines": [ + { + "bbox": [ + 61, + 71, + 219, + 284 + ], + "spans": [ + { + "bbox": [ + 61, + 71, + 219, + 284 + ], + "type": "image", + "image_path": "450144b6bf9cce2e7f345cdf281ed193bc9ac75e1450daeb4316cc0d6fe35738.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 293, + 555, + 350 + ], + "lines": [ + { + "bbox": [ + 54, + 293, + 555, + 350 + ], + "spans": [ + { + "bbox": [ + 54, + 293, + 555, + 350 + ], + "type": "text", + "content": "Figure 1. The difference between our Spiking Self-Attention and the vanilla Spiking Self-Attention. Our Spiking Self-Attention differs significantly from the vanilla Spiking Self-Attention. In the vanilla version, only binary spikes are employed, which can result in considerable information loss. In contrast, our approach integrates a combination of binary spiking neurons, ReLU, and ternary spiking neurons. This hybrid structure effectively reduces information loss, while preserving the advantages of addition-only processing. Moreover, our method eliminates the need for both scaling and the softmax function, streamlining the computation." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 220, + 70, + 389, + 283 + ], + "blocks": [ + { + "bbox": [ + 220, + 70, + 389, + 283 + ], + "lines": [ + { + "bbox": [ + 220, + 70, + 389, + 283 + ], + "spans": [ + { + "bbox": [ + 220, + 70, + 389, + 283 + ], + "type": "image", + "image_path": "e2ca4f51cb57896d5beadc3da775bafdaaf9468bfe96318c9c0ef58f2c8dad44.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 392, + 76, + 552, + 283 + ], + "blocks": [ + { + "bbox": [ + 392, + 76, + 552, + 283 + ], + "lines": [ + { + "bbox": [ + 392, + 76, + 552, + 283 + ], + "spans": [ + { + "bbox": [ + 392, + 76, + 552, + 283 + ], + "type": "image", + "image_path": "e4ab02036ee1026ef21566b118003223f9804a95c5e45d0939ea639bf050daf0.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 370, + 269, + 383 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 370, + 269, + 383 + ], + "spans": [ + { + "bbox": [ + 55, + 370, + 269, + 383 + ], + "type": "text", + "content": "significant advantage in real-world applications [49]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 384, + 296, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 384, + 296, + 539 + ], + "spans": [ + { + "bbox": [ + 55, + 384, + 296, + 539 + ], + "type": "text", + "content": "Furthermore, the adaptation of transformers for visual tasks has inspired numerous variations and improvements across the field. Researchers have explored strategies such as integrating convolutional layers into the architecture [6] to retain spatial information more effectively. Other approaches include the use of hierarchical representations [36], which facilitate multi-scale feature extraction, and the incorporation of attention mechanisms tailored specifically for vision tasks [5], improving the model's ability to focus on relevant parts of the input data. These developments underscore the versatility and robustness of the transformer architecture in addressing complex visual challenges, paving the way for further innovations in computer vision." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 551, + 280, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 551, + 280, + 563 + ], + "spans": [ + { + "bbox": [ + 55, + 551, + 280, + 563 + ], + "type": "text", + "content": "2.3. Spiking Neural Networks for Transformers" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 570, + 297, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 297, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 297, + 713 + ], + "type": "text", + "content": "The integration of SNNs into transformer architectures represents an exciting and rapidly developing area of research. This field aims to leverage the distinct advantages of both paradigms: the energy-efficient characteristics of SNNs and the robust contextual representation capabilities of transformers. Recent studies [33, 37, 51] have explored replacing certain neurons within transformers with spiking neurons, marking a significant step towards this integration. While these efforts contribute to mitigating the accuracy loss associated with introducing spiking neurons into transformers, they still face challenges in fully realizing the low energy consumption benefits of SNNs. This is primarily due to" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 370, + 555, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 370, + 555, + 418 + ], + "spans": [ + { + "bbox": [ + 313, + 370, + 555, + 418 + ], + "type": "text", + "content": "their reliance on hybrid computing methodologies, which continue to require traditional Multiply-Accumulate (MAC) operations, such as dot products, softmax calculations, and scaling operations." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 434, + 556, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 434, + 556, + 590 + ], + "spans": [ + { + "bbox": [ + 313, + 434, + 556, + 590 + ], + "type": "text", + "content": "To address this issue, Spikformer [64, 65] proposes converting the " + }, + { + "bbox": [ + 313, + 434, + 556, + 590 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 313, + 434, + 556, + 590 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 434, + 556, + 590 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 434, + 556, + 590 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 434, + 556, + 590 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 313, + 434, + 556, + 590 + ], + "type": "text", + "content": " components into spike form before performing matrix multiplications similar to those used in VSA. This innovative approach allows spike matrix multiplications to be transformed into addition operations, eliminating the need for softmax computations. However, when considering the residual connections in these architectures, Spikformer [64, 65] still incorporates non-spiking computations within the ConvBN layers. To overcome this limitation, Spikingformer [63] and Spike-driven Transformer [58] have restructured the residual connections in standard transformer architectures, placing them before activation functions to ensure that all neurons convey binary spike signals." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "content": "The computation of spike-form " + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "content": " in these models circumvent multiplications, relying solely on additions. However, simply placing spiking neurons in front of " + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "content": " to convert them into spike form before performing matrix multiplications does not fully exploit the addition-only advantage of SNNs. In this work, we propose a novel, addition-only spiking self-attention mechanism for transformers, aiming to push the boundaries of what can be achieved through the integration of these two technologies." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "24400" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 71, + 133, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 71, + 133, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 133, + 85 + ], + "type": "text", + "content": "3. Preliminary" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 183, + 104 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 183, + 104 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 183, + 104 + ], + "type": "text", + "content": "3.1. Spiking Neuron Layer" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 108, + 297, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 108, + 297, + 204 + ], + "spans": [ + { + "bbox": [ + 55, + 108, + 297, + 204 + ], + "type": "text", + "content": "The spiking neuron layer plays a crucial role in integrating both spatial and temporal information, which is then encoded into the membrane potential and subsequently transformed into binary spikes. These spikes drive further computations in subsequent layers of the network. We model the dynamics of the spiking neuron using the LIF model [40]. The evolution of the membrane potential and the spike generation mechanism are described by the following equations:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 100, + 212, + 296, + 226 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 212, + 296, + 226 + ], + "spans": [ + { + "bbox": [ + 100, + 212, + 296, + 226 + ], + "type": "interline_equation", + "content": "U [ t ] = H [ t - 1 ] + X [ t ], \\tag {1}", + "image_path": "fb539dd731cdadbe3923ff7a37bad25cbf8477d2185f1d5826019e920fdd84e8.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 102, + 227, + 296, + 240 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 227, + 296, + 240 + ], + "spans": [ + { + "bbox": [ + 102, + 227, + 296, + 240 + ], + "type": "interline_equation", + "content": "S [ t ] = \\operatorname {H e a} (U [ t ] - V _ {\\mathrm {t h}}), \\tag {2}", + "image_path": "eb620d6a7b9a0fde9ddbc1bae82692f4f63d765a16adb8b2c390eae42c05b09d.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 100, + 242, + 296, + 255 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 242, + 296, + 255 + ], + "spans": [ + { + "bbox": [ + 100, + 242, + 296, + 255 + ], + "type": "interline_equation", + "content": "H [ t ] = V _ {\\text {r e s e t}} S [ t ] + \\beta U [ t ] (1 - S [ t ]), \\tag {3}", + "image_path": "ad6810049d591961111c61efeeb85f9784d4707271f524375369980ee0e1045c.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 263, + 296, + 359 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 263, + 296, + 359 + ], + "spans": [ + { + "bbox": [ + 55, + 263, + 296, + 359 + ], + "type": "text", + "content": "where: " + }, + { + "bbox": [ + 55, + 263, + 296, + 359 + ], + "type": "inline_equation", + "content": "X[t]" + }, + { + "bbox": [ + 55, + 263, + 296, + 359 + ], + "type": "text", + "content": " is the spatial input current at time " + }, + { + "bbox": [ + 55, + 263, + 296, + 359 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 263, + 296, + 359 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 263, + 296, + 359 + ], + "type": "inline_equation", + "content": "U[t]" + }, + { + "bbox": [ + 55, + 263, + 296, + 359 + ], + "type": "text", + "content": " is the membrane potential at time " + }, + { + "bbox": [ + 55, + 263, + 296, + 359 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 263, + 296, + 359 + ], + "type": "text", + "content": ", combining the spatial input " + }, + { + "bbox": [ + 55, + 263, + 296, + 359 + ], + "type": "inline_equation", + "content": "X[t]" + }, + { + "bbox": [ + 55, + 263, + 296, + 359 + ], + "type": "text", + "content": " and the temporal input " + }, + { + "bbox": [ + 55, + 263, + 296, + 359 + ], + "type": "inline_equation", + "content": "H[t - 1]" + }, + { + "bbox": [ + 55, + 263, + 296, + 359 + ], + "type": "text", + "content": " from the previous time step, " + }, + { + "bbox": [ + 55, + 263, + 296, + 359 + ], + "type": "inline_equation", + "content": "\\mathrm{Hea}(\\cdot)" + }, + { + "bbox": [ + 55, + 263, + 296, + 359 + ], + "type": "text", + "content": " is the Heaviside step function, which outputs 1 if its argument is non-negative, and 0 otherwise, " + }, + { + "bbox": [ + 55, + 263, + 296, + 359 + ], + "type": "inline_equation", + "content": "S[t]" + }, + { + "bbox": [ + 55, + 263, + 296, + 359 + ], + "type": "text", + "content": " is the spike output at time " + }, + { + "bbox": [ + 55, + 263, + 296, + 359 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 263, + 296, + 359 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 263, + 296, + 359 + ], + "type": "inline_equation", + "content": "V_{\\mathrm{th}}" + }, + { + "bbox": [ + 55, + 263, + 296, + 359 + ], + "type": "text", + "content": " is the spike threshold, " + }, + { + "bbox": [ + 55, + 263, + 296, + 359 + ], + "type": "inline_equation", + "content": "V_{\\mathrm{reset}}" + }, + { + "bbox": [ + 55, + 263, + 296, + 359 + ], + "type": "text", + "content": " is the reset potential, and " + }, + { + "bbox": [ + 55, + 263, + 296, + 359 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 55, + 263, + 296, + 359 + ], + "type": "text", + "content": " is a decay factor controlling the membrane potential's decay." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 360, + 296, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 360, + 296, + 456 + ], + "spans": [ + { + "bbox": [ + 55, + 360, + 296, + 456 + ], + "type": "text", + "content": "When the membrane potential " + }, + { + "bbox": [ + 55, + 360, + 296, + 456 + ], + "type": "inline_equation", + "content": "U[t]" + }, + { + "bbox": [ + 55, + 360, + 296, + 456 + ], + "type": "text", + "content": " exceeds a predefined threshold " + }, + { + "bbox": [ + 55, + 360, + 296, + 456 + ], + "type": "inline_equation", + "content": "V_{\\mathrm{th}}" + }, + { + "bbox": [ + 55, + 360, + 296, + 456 + ], + "type": "text", + "content": ", the neuron emits a spike, " + }, + { + "bbox": [ + 55, + 360, + 296, + 456 + ], + "type": "inline_equation", + "content": "S[t] = 1" + }, + { + "bbox": [ + 55, + 360, + 296, + 456 + ], + "type": "text", + "content": ", and the internal state " + }, + { + "bbox": [ + 55, + 360, + 296, + 456 + ], + "type": "inline_equation", + "content": "H[t]" + }, + { + "bbox": [ + 55, + 360, + 296, + 456 + ], + "type": "text", + "content": " is reset to " + }, + { + "bbox": [ + 55, + 360, + 296, + 456 + ], + "type": "inline_equation", + "content": "V_{\\mathrm{reset}}" + }, + { + "bbox": [ + 55, + 360, + 296, + 456 + ], + "type": "text", + "content": ". If " + }, + { + "bbox": [ + 55, + 360, + 296, + 456 + ], + "type": "inline_equation", + "content": "U[t]" + }, + { + "bbox": [ + 55, + 360, + 296, + 456 + ], + "type": "text", + "content": " does not surpass the threshold, the membrane potential decays towards the previous state " + }, + { + "bbox": [ + 55, + 360, + 296, + 456 + ], + "type": "inline_equation", + "content": "H[t - 1]" + }, + { + "bbox": [ + 55, + 360, + 296, + 456 + ], + "type": "text", + "content": " at a rate governed by " + }, + { + "bbox": [ + 55, + 360, + 296, + 456 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 55, + 360, + 296, + 456 + ], + "type": "text", + "content": ". To simplify notation, we represent the spiking neuron layer as " + }, + { + "bbox": [ + 55, + 360, + 296, + 456 + ], + "type": "inline_equation", + "content": "\\mathcal{SN}(\\cdot)" + }, + { + "bbox": [ + 55, + 360, + 296, + 456 + ], + "type": "text", + "content": ", where the input is the membrane potential tensor " + }, + { + "bbox": [ + 55, + 360, + 296, + 456 + ], + "type": "inline_equation", + "content": "U" + }, + { + "bbox": [ + 55, + 360, + 296, + 456 + ], + "type": "text", + "content": ", and the output is the spike tensor " + }, + { + "bbox": [ + 55, + 360, + 296, + 456 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 55, + 360, + 296, + 456 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 462, + 269, + 475 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 462, + 269, + 475 + ], + "spans": [ + { + "bbox": [ + 55, + 462, + 269, + 475 + ], + "type": "text", + "content": "3.2. Vanilla Self-Attention Mechanism (VSA)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 479, + 296, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 479, + 296, + 528 + ], + "spans": [ + { + "bbox": [ + 55, + 479, + 296, + 528 + ], + "type": "text", + "content": "The Vanilla Self-Attention (VSA) mechanism enables a model to focus on different parts of the input sequence while constructing its output representations. The self-attention operation can be mathematically expressed as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 94, + 535, + 296, + 563 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 535, + 296, + 563 + ], + "spans": [ + { + "bbox": [ + 94, + 535, + 296, + 563 + ], + "type": "interline_equation", + "content": "\\operatorname {V S A} (Q, K, V) = \\operatorname {s o f t m a x} \\left(\\frac {Q K ^ {T}}{\\sqrt {d _ {k}}}\\right) V \\tag {4}", + "image_path": "9bc2e6fb3026dfc2b4d8332435830eb889a470c97aedf9fdcb4ea3a33709941e.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 569, + 296, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 569, + 296, + 653 + ], + "spans": [ + { + "bbox": [ + 55, + 569, + 296, + 653 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 569, + 296, + 653 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 55, + 569, + 296, + 653 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 569, + 296, + 653 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 569, + 296, + 653 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 569, + 296, + 653 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 55, + 569, + 296, + 653 + ], + "type": "text", + "content": " represent the query, key, and value matrices, respectively, and " + }, + { + "bbox": [ + 55, + 569, + 296, + 653 + ], + "type": "inline_equation", + "content": "d_k" + }, + { + "bbox": [ + 55, + 569, + 296, + 653 + ], + "type": "text", + "content": " is the dimension of the key vectors. The softmax function is applied to normalize the attention scores, ensuring they are non-negative. The resulting weighted sum of the values reflects the importance of each value, based on the similarity between the query and its corresponding key." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 653, + 296, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 689 + ], + "type": "text", + "content": "The input to the self-attention mechanism consists of a sequence of embeddings " + }, + { + "bbox": [ + 55, + 653, + 296, + 689 + ], + "type": "inline_equation", + "content": "x_{1}, x_{2}, \\ldots, x_{n}" + }, + { + "bbox": [ + 55, + 653, + 296, + 689 + ], + "type": "text", + "content": ", which are linearly projected into " + }, + { + "bbox": [ + 55, + 653, + 296, + 689 + ], + "type": "inline_equation", + "content": "Q, K" + }, + { + "bbox": [ + 55, + 653, + 296, + 689 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 653, + 296, + 689 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 55, + 653, + 296, + 689 + ], + "type": "text", + "content": " using learned weight matrices:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 86, + 697, + 296, + 711 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 697, + 296, + 711 + ], + "spans": [ + { + "bbox": [ + 86, + 697, + 296, + 711 + ], + "type": "interline_equation", + "content": "Q = X W _ {Q}, \\quad K = X W _ {K}, \\quad V = X W _ {V}, \\tag {5}", + "image_path": "36a0027600148b23ca8f4e4b97deff706e87ff800bb25a4305e679fcd94a4a72.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 72, + 553, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 97 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 97 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 72, + 553, + 97 + ], + "type": "inline_equation", + "content": "W_{Q}" + }, + { + "bbox": [ + 313, + 72, + 553, + 97 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 72, + 553, + 97 + ], + "type": "inline_equation", + "content": "W_{K}" + }, + { + "bbox": [ + 313, + 72, + 553, + 97 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 72, + 553, + 97 + ], + "type": "inline_equation", + "content": "W_{V}" + }, + { + "bbox": [ + 313, + 72, + 553, + 97 + ], + "type": "text", + "content": " are the learned weight matrices corresponding to the " + }, + { + "bbox": [ + 313, + 72, + 553, + 97 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 313, + 72, + 553, + 97 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 72, + 553, + 97 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 72, + 553, + 97 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 72, + 553, + 97 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 313, + 72, + 553, + 97 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 98, + 554, + 182 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 98, + 554, + 182 + ], + "spans": [ + { + "bbox": [ + 313, + 98, + 554, + 182 + ], + "type": "text", + "content": "However, the standard VSA is not directly compatible with SNNs due to the conflict between the floating-point operations required for matrix multiplication of " + }, + { + "bbox": [ + 313, + 98, + 554, + 182 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 313, + 98, + 554, + 182 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 98, + 554, + 182 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 98, + 554, + 182 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 98, + 554, + 182 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 313, + 98, + 554, + 182 + ], + "type": "text", + "content": ", and the energy efficiency of SNN computations. What's more, the softmax operation, which involves exponentiation and division, does not align with the computation paradigm of SNNs too." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 196, + 554, + 220 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 196, + 554, + 220 + ], + "spans": [ + { + "bbox": [ + 313, + 196, + 554, + 220 + ], + "type": "text", + "content": "3.3. Vanilla Spiking Self-Attention Mechanism (VSSA)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 227, + 555, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 227, + 555, + 312 + ], + "spans": [ + { + "bbox": [ + 313, + 227, + 555, + 312 + ], + "type": "text", + "content": "To address the incompatibility between the VSA and SNNs, the Vanilla Spiking Self-Attention (VSSA) mechanism is proposed in [64, 65], which is more suitable for SNNs, as shown in Figure 1(b). The " + }, + { + "bbox": [ + 313, + 227, + 555, + 312 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 313, + 227, + 555, + 312 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 227, + 555, + 312 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 227, + 555, + 312 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 227, + 555, + 312 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 313, + 227, + 555, + 312 + ], + "type": "text", + "content": " are initially derived through learnable weight matrices. These are then transformed into spiking sequences by applying binary spiking neuron layers:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 380, + 325, + 555, + 338 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 380, + 325, + 555, + 338 + ], + "spans": [ + { + "bbox": [ + 380, + 325, + 555, + 338 + ], + "type": "interline_equation", + "content": "Q = \\mathcal {S N} _ {Q} (\\mathrm {B N} (X W _ {Q})), \\tag {6}", + "image_path": "af41e35331bc0ecd2e0199bb8c673a2e47f62226ec082d2bbd05d7adf4c0abbf.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 380, + 340, + 555, + 353 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 380, + 340, + 555, + 353 + ], + "spans": [ + { + "bbox": [ + 380, + 340, + 555, + 353 + ], + "type": "interline_equation", + "content": "K = \\mathcal {S N} _ {K} (\\mathrm {B N} (X W _ {K})), \\tag {7}", + "image_path": "e671ace4a8d3fe21dfd933a2b445d80fd2e49eaa23cfb8a7d79e7eadd76c433a.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 382, + 355, + 555, + 368 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 382, + 355, + 555, + 368 + ], + "spans": [ + { + "bbox": [ + 382, + 355, + 555, + 368 + ], + "type": "interline_equation", + "content": "V = \\mathcal {S N} _ {V} (\\mathrm {B N} (X W _ {V})). \\tag {8}", + "image_path": "4820228d6c291d8f53357f2a911f8c9d6eb277aa29bb4168700eb6640cd9451d.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 381, + 554, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 381, + 554, + 430 + ], + "spans": [ + { + "bbox": [ + 313, + 381, + 554, + 430 + ], + "type": "text", + "content": "In this approach, the computation of the attention matrix is performed using binary spike-form queries and keys (which contain only 0s and 1s), thus replacing traditional matrix multiplication with addition-based operations." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 431, + 555, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 431, + 555, + 479 + ], + "spans": [ + { + "bbox": [ + 313, + 431, + 555, + 479 + ], + "type": "text", + "content": "To further address the challenges of large values resulting from the matrix multiplication, a scaling factor " + }, + { + "bbox": [ + 313, + 431, + 555, + 479 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 313, + 431, + 555, + 479 + ], + "type": "text", + "content": " is introduced to regulate the magnitude of the result. The spike-efficient VSSA is formulated as follows:" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 352, + 492, + 555, + 507 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 352, + 492, + 555, + 507 + ], + "spans": [ + { + "bbox": [ + 352, + 492, + 555, + 507 + ], + "type": "interline_equation", + "content": "\\operatorname {V S S A} (Q, K, V) = \\mathcal {S N} \\left(Q K ^ {T} V \\cdot s\\right). \\tag {9}", + "image_path": "7558b7b06be9d57490be09eb1a4edcd408d2d6996099504c12b64247181c8257.jpg" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 520, + 555, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 520, + 555, + 579 + ], + "spans": [ + { + "bbox": [ + 313, + 520, + 555, + 579 + ], + "type": "text", + "content": "In this formulation, the scaling factor " + }, + { + "bbox": [ + 313, + 520, + 555, + 579 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 313, + 520, + 555, + 579 + ], + "type": "text", + "content": " adjusts the magnitude of the result from the matrix product, while all other operations in VSSA—such as the attention calculation—are performed using addition, in line with the spike-based computation model." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 582, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 556, + 713 + ], + "type": "text", + "content": "However, simply converting the queries, keys, and values into spike form prior to performing matrix multiplication does not fully leverage the advantages of SNNs, which thrive on addition-only and event-driven computations. It is unnecessary to convert all of " + }, + { + "bbox": [ + 313, + 582, + 556, + 713 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 313, + 582, + 556, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 582, + 556, + 713 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 582, + 556, + 713 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 582, + 556, + 713 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 313, + 582, + 556, + 713 + ], + "type": "text", + "content": " into binary spike sequences to transition from matrix multiplication to addition. Furthermore, converting all of the values to binary spikes can result in significant information loss. Thus, the challenge remains to strike a balance between maintaining the richness of the data while taking full advantage of the addition operation-based nature of SNNs." + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24401" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 70, + 550, + 304 + ], + "blocks": [ + { + "bbox": [ + 60, + 70, + 550, + 304 + ], + "lines": [ + { + "bbox": [ + 60, + 70, + 550, + 304 + ], + "spans": [ + { + "bbox": [ + 60, + 70, + 550, + 304 + ], + "type": "image", + "image_path": "2cfcd6b4c360199ad8437603374939eff8452fe837c2b75f5c0449d0acc9189a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 216, + 312, + 394, + 323 + ], + "lines": [ + { + "bbox": [ + 216, + 312, + 394, + 323 + ], + "spans": [ + { + "bbox": [ + 216, + 312, + 394, + 323 + ], + "type": "text", + "content": "Figure 2. The overview of Spiking Transformer." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 342, + 111, + 355 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 342, + 111, + 355 + ], + "spans": [ + { + "bbox": [ + 55, + 342, + 111, + 355 + ], + "type": "text", + "content": "4. Method" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 362, + 176, + 374 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 362, + 176, + 374 + ], + "spans": [ + { + "bbox": [ + 55, + 362, + 176, + 374 + ], + "type": "text", + "content": "4.1. Overall Architecture" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 380, + 297, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 380, + 297, + 571 + ], + "spans": [ + { + "bbox": [ + 54, + 380, + 297, + 571 + ], + "type": "text", + "content": "Figure 2 illustrates the architecture of our Spiking Transformer, which comprises four primary components: Spiking Patch Splitting (SPS), Accurate Addition-Only Spiking Self-Attention " + }, + { + "bbox": [ + 54, + 380, + 297, + 571 + ], + "type": "inline_equation", + "content": "(\\mathrm{A}^2\\mathrm{OS}^2\\mathrm{A})" + }, + { + "bbox": [ + 54, + 380, + 297, + 571 + ], + "type": "text", + "content": ", Multi-Layer Perceptron (MLP), and a linear classification head. The design follows the approach outlined in [58]. For the SPS module, we adopt the structure from [63, 64]. Given a 2D image sequence " + }, + { + "bbox": [ + 54, + 380, + 297, + 571 + ], + "type": "inline_equation", + "content": "I \\in \\mathbb{R}^{T \\times C \\times H \\times W}" + }, + { + "bbox": [ + 54, + 380, + 297, + 571 + ], + "type": "text", + "content": ", the Patch Splitting Module (PSM) performs a linear projection and splits the input into a sequence of " + }, + { + "bbox": [ + 54, + 380, + 297, + 571 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 54, + 380, + 297, + 571 + ], + "type": "text", + "content": " flattened spike patches with " + }, + { + "bbox": [ + 54, + 380, + 297, + 571 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 54, + 380, + 297, + 571 + ], + "type": "text", + "content": "-dimensional channels, where " + }, + { + "bbox": [ + 54, + 380, + 297, + 571 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 54, + 380, + 297, + 571 + ], + "type": "text", + "content": " denotes the number of timesteps (images are repeated " + }, + { + "bbox": [ + 54, + 380, + 297, + 571 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 54, + 380, + 297, + 571 + ], + "type": "text", + "content": " times in the dataset), " + }, + { + "bbox": [ + 54, + 380, + 297, + 571 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 54, + 380, + 297, + 571 + ], + "type": "text", + "content": " is the number of channels, and " + }, + { + "bbox": [ + 54, + 380, + 297, + 571 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 54, + 380, + 297, + 571 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 380, + 297, + 571 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 54, + 380, + 297, + 571 + ], + "type": "text", + "content": " represent the height and width of the image sequence. Additionally, a convolutional layer generates Relative Position Embeddings (RPE) as described in [58]. Together, the SPS part is written as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 578, + 296, + 593 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 578, + 296, + 593 + ], + "spans": [ + { + "bbox": [ + 67, + 578, + 296, + 593 + ], + "type": "interline_equation", + "content": "u = \\operatorname {P S M} (I), I \\in \\mathbb {R} ^ {T \\times C \\times H \\times W}, x \\in \\mathbb {R} ^ {T \\times N \\times D}, \\tag {10}", + "image_path": "761e667fb4eb9a010bf885c4f637eea2896212049a0fe345ff21e3f1b902e074.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 117, + 608, + 296, + 622 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 117, + 608, + 296, + 622 + ], + "spans": [ + { + "bbox": [ + 117, + 608, + 296, + 622 + ], + "type": "interline_equation", + "content": "s = \\mathcal {S N} (u), s \\in \\mathbb {R} ^ {T \\times N \\times D}, \\tag {11}", + "image_path": "a0db62a56cf6617588157e804adba7fdb4dff9713cc901a11c0b172631a7e772.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 70, + 637, + 296, + 653 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 637, + 296, + 653 + ], + "spans": [ + { + "bbox": [ + 70, + 637, + 296, + 653 + ], + "type": "interline_equation", + "content": "\\mathrm {R P E} = \\mathrm {B N} ((\\operatorname {C o n v 2 d} (s))), \\mathrm {R P E} \\in \\mathbb {R} ^ {T \\times N \\times D}, \\tag {12}", + "image_path": "4515f678034a2ecbea2d3ff59a7ed3d6a4db65296170698839aeef5827f4bfd0.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 107, + 667, + 296, + 682 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 667, + 296, + 682 + ], + "spans": [ + { + "bbox": [ + 107, + 667, + 296, + 682 + ], + "type": "interline_equation", + "content": "U _ {0} = u + \\operatorname {R P E}, U _ {0} \\in \\mathbb {R} ^ {T \\times N \\times D}, \\tag {13}", + "image_path": "dfdd8460ac126eed93875ecd86c7c9083325423d46e3479f3900dc78d1a12a6c.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "inline_equation", + "content": "U_{0}" + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": " are the output membrane potential tensors from the PSM and SPS modules, respectively, and " + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{SN}(\\cdot)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 344, + 555, + 451 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 344, + 555, + 451 + ], + "spans": [ + { + "bbox": [ + 313, + 344, + 555, + 451 + ], + "type": "text", + "content": "represents the spike neuron layer. The output " + }, + { + "bbox": [ + 313, + 344, + 555, + 451 + ], + "type": "inline_equation", + "content": "U_{0}" + }, + { + "bbox": [ + 313, + 344, + 555, + 451 + ], + "type": "text", + "content": " is then passed to the " + }, + { + "bbox": [ + 313, + 344, + 555, + 451 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 313, + 344, + 555, + 451 + ], + "type": "text", + "content": "-block Spiking Transformer encoder, which consists of the " + }, + { + "bbox": [ + 313, + 344, + 555, + 451 + ], + "type": "inline_equation", + "content": "\\mathrm{A}^{2}\\mathrm{OS}^{2}\\mathrm{A}" + }, + { + "bbox": [ + 313, + 344, + 555, + 451 + ], + "type": "text", + "content": " and MLP blocks. Residual connections are applied to the membrane potentials in both the " + }, + { + "bbox": [ + 313, + 344, + 555, + 451 + ], + "type": "inline_equation", + "content": "\\mathrm{A}^{2}\\mathrm{OS}^{2}\\mathrm{A}" + }, + { + "bbox": [ + 313, + 344, + 555, + 451 + ], + "type": "text", + "content": " and MLP blocks. To avoid multi-bit spike outputs, we adopt the design from [58, 63], ensuring that the values before the convolutional layers are binary, thus allowing the spike and weight matrix multiplication to be simplified to addition operations." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 452, + 556, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 452, + 556, + 571 + ], + "spans": [ + { + "bbox": [ + 313, + 452, + 556, + 571 + ], + "type": "text", + "content": "The " + }, + { + "bbox": [ + 313, + 452, + 556, + 571 + ], + "type": "inline_equation", + "content": "\\mathrm{A}^2\\mathrm{OS}^2\\mathrm{A}" + }, + { + "bbox": [ + 313, + 452, + 556, + 571 + ], + "type": "text", + "content": " mechanism models both local and global information of the input sequence using binary spike " + }, + { + "bbox": [ + 313, + 452, + 556, + 571 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 313, + 452, + 556, + 571 + ], + "type": "text", + "content": ", full-precision " + }, + { + "bbox": [ + 313, + 452, + 556, + 571 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 452, + 556, + 571 + ], + "type": "text", + "content": ", and ternary spike " + }, + { + "bbox": [ + 313, + 452, + 556, + 571 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 313, + 452, + 556, + 571 + ], + "type": "text", + "content": ". This approach reduces information loss and eliminates the need for scaling and softmax operations, thereby improving task accuracy. A Global Average Pooling (GAP) is applied to the processed feature from the Spiking Transformer encoder, and the resulting " + }, + { + "bbox": [ + 313, + 452, + 556, + 571 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 313, + 452, + 556, + 571 + ], + "type": "text", + "content": "-dimensional channel is passed to a fully connected Classification Head (CH) for final prediction. The " + }, + { + "bbox": [ + 313, + 452, + 556, + 571 + ], + "type": "inline_equation", + "content": "\\mathrm{A}^2\\mathrm{OS}^2\\mathrm{A}" + }, + { + "bbox": [ + 313, + 452, + 556, + 571 + ], + "type": "text", + "content": ", MLP, and CH components are as follows:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 367, + 575, + 555, + 590 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 575, + 555, + 590 + ], + "spans": [ + { + "bbox": [ + 367, + 575, + 555, + 590 + ], + "type": "interline_equation", + "content": "S _ {0} = \\mathcal {S N} (U _ {0}), S _ {0} \\in \\mathbb {R} ^ {T \\times N \\times D}, \\tag {14}", + "image_path": "66d1ffc888e971eea9a309bf4c50ecd0bb1c6a74eb973ea7919873c5647d31a6.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 317, + 601, + 555, + 626 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 601, + 555, + 626 + ], + "spans": [ + { + "bbox": [ + 317, + 601, + 555, + 626 + ], + "type": "interline_equation", + "content": "U _ {l} ^ {\\prime} = \\mathrm {A} ^ {2} \\mathrm {O S} ^ {2} \\mathrm {A} \\left(S _ {l - 1}\\right) + U _ {l - 1}, U _ {l} ^ {\\prime} \\in \\mathbb {R} ^ {T \\times N \\times D}, l = 1 \\dots L, \\tag {15}", + "image_path": "db3c4f1aa739fac90df5d77c2829be4f0c824d2c407a31a39ce42d8dd087bf47.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 369, + 638, + 555, + 653 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 369, + 638, + 555, + 653 + ], + "spans": [ + { + "bbox": [ + 369, + 638, + 555, + 653 + ], + "type": "interline_equation", + "content": "S _ {l} ^ {\\prime} = \\mathcal {S N} \\left(U _ {l} ^ {\\prime}\\right), S _ {l} ^ {\\prime} \\in \\mathbb {R} ^ {T \\times N \\times D}, \\tag {16}", + "image_path": "e8870e2f65c3639a2151deee6de4c148d5ba86611db19fd35286b53c6558987e.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 323, + 663, + 555, + 689 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 663, + 555, + 689 + ], + "spans": [ + { + "bbox": [ + 323, + 663, + 555, + 689 + ], + "type": "interline_equation", + "content": "S _ {l} = \\mathcal {S N} \\left(\\operatorname {M L P} \\left(S _ {l} ^ {\\prime}\\right) + U _ {l} ^ {\\prime}\\right), S _ {l} \\in \\mathbb {R} ^ {T \\times N \\times D}, l = 1 \\dots L, \\tag {17}", + "image_path": "f62daeee35ea275a366158f35a8ee6e41a581fc09110935d8170e67de68a02ba.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 388, + 700, + 555, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 388, + 700, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 388, + 700, + 555, + 715 + ], + "type": "interline_equation", + "content": "Y = \\mathrm {C H} (\\mathrm {G A P} (S _ {L})). \\tag {18}", + "image_path": "cab77ec2654338533cd4edfa69d2b32f6d6459ee21f55ae9271ca2e749bd4656.jpg" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "24402" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 277, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 277, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 277, + 85 + ], + "type": "text", + "content": "4.2. Information Loss in Spiking Transformers" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 89, + 296, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 89, + 296, + 172 + ], + "spans": [ + { + "bbox": [ + 55, + 89, + 296, + 172 + ], + "type": "text", + "content": "In this section, we analyze the limitations of binary spikes in SNNs with respect to information representation, which motivates the focus of this paper. While binary spikes offer energy efficiency, they inherently suffer from reduced representational capacity, leading to information loss when used for all " + }, + { + "bbox": [ + 55, + 89, + 296, + 172 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 55, + 89, + 296, + 172 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 89, + 296, + 172 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 89, + 296, + 172 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 89, + 296, + 172 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 55, + 89, + 296, + 172 + ], + "type": "text", + "content": " in VSSA, which results in accuracy degradation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 173, + 296, + 221 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 173, + 296, + 221 + ], + "spans": [ + { + "bbox": [ + 55, + 173, + 296, + 221 + ], + "type": "text", + "content": "To support this claim, we first perform a theoretical analysis using the concept of entropy. The representational capability " + }, + { + "bbox": [ + 55, + 173, + 296, + 221 + ], + "type": "inline_equation", + "content": "\\mathcal{C}(\\mathbf{X})" + }, + { + "bbox": [ + 55, + 173, + 296, + 221 + ], + "type": "text", + "content": " of a set " + }, + { + "bbox": [ + 55, + 173, + 296, + 221 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 55, + 173, + 296, + 221 + ], + "type": "text", + "content": " is determined by the maximum entropy of " + }, + { + "bbox": [ + 55, + 173, + 296, + 221 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 55, + 173, + 296, + 221 + ], + "type": "text", + "content": ", expressed as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 69, + 226, + 296, + 251 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 226, + 296, + 251 + ], + "spans": [ + { + "bbox": [ + 69, + 226, + 296, + 251 + ], + "type": "interline_equation", + "content": "\\mathcal {C} (\\mathbf {X}) = \\max \\mathcal {H} (\\mathbf {X}) = - \\sum_ {x \\in \\mathbf {X}} p _ {\\mathbf {X}} (x) \\log p _ {\\mathbf {X}} (x), \\tag {19}", + "image_path": "67c3d7f68d9c38a5cc662fa20586d822cfb9d9e26c1aefbd7a43c3c49fbfad07.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 257, + 296, + 282 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 257, + 296, + 282 + ], + "spans": [ + { + "bbox": [ + 55, + 257, + 296, + 282 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 257, + 296, + 282 + ], + "type": "inline_equation", + "content": "p_{\\mathbf{X}}(x)" + }, + { + "bbox": [ + 55, + 257, + 296, + 282 + ], + "type": "text", + "content": " is the probability of a sample " + }, + { + "bbox": [ + 55, + 257, + 296, + 282 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 257, + 296, + 282 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 55, + 257, + 296, + 282 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 55, + 257, + 296, + 282 + ], + "type": "text", + "content": ". We now present the following proposition:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 288, + 296, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 288, + 296, + 350 + ], + "spans": [ + { + "bbox": [ + 55, + 288, + 296, + 350 + ], + "type": "text", + "content": "Proposition 1 For a set " + }, + { + "bbox": [ + 55, + 288, + 296, + 350 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 55, + 288, + 296, + 350 + ], + "type": "text", + "content": ", its representational capacity is " + }, + { + "bbox": [ + 55, + 288, + 296, + 350 + ], + "type": "inline_equation", + "content": "\\mathcal{C}(\\mathbf{X}) = \\max \\mathcal{H}(\\mathbf{X})" + }, + { + "bbox": [ + 55, + 288, + 296, + 350 + ], + "type": "text", + "content": ". When the probability distribution of " + }, + { + "bbox": [ + 55, + 288, + 296, + 350 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 55, + 288, + 296, + 350 + ], + "type": "text", + "content": " is uniform, i.e., " + }, + { + "bbox": [ + 55, + 288, + 296, + 350 + ], + "type": "inline_equation", + "content": "p_{\\mathbf{X}}(x) = \\frac{1}{N}" + }, + { + "bbox": [ + 55, + 288, + 296, + 350 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 288, + 296, + 350 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 288, + 296, + 350 + ], + "type": "text", + "content": " is the total number of samples in " + }, + { + "bbox": [ + 55, + 288, + 296, + 350 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 55, + 288, + 296, + 350 + ], + "type": "text", + "content": ", the entropy " + }, + { + "bbox": [ + 55, + 288, + 296, + 350 + ], + "type": "inline_equation", + "content": "\\mathcal{H}(\\mathbf{X})" + }, + { + "bbox": [ + 55, + 288, + 296, + 350 + ], + "type": "text", + "content": " reaches its maximum value of " + }, + { + "bbox": [ + 55, + 288, + 296, + 350 + ], + "type": "inline_equation", + "content": "\\log(N)" + }, + { + "bbox": [ + 55, + 288, + 296, + 350 + ], + "type": "text", + "content": ". Hence, we conclude that " + }, + { + "bbox": [ + 55, + 288, + 296, + 350 + ], + "type": "inline_equation", + "content": "\\mathcal{C}(\\mathbf{X}) = \\log(N)" + }, + { + "bbox": [ + 55, + 288, + 296, + 350 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 354, + 296, + 451 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 354, + 296, + 451 + ], + "spans": [ + { + "bbox": [ + 55, + 354, + 296, + 451 + ], + "type": "text", + "content": "Using Proposition 1, we can evaluate the representational capacity of binary spike layers in SNNs and compare them with real-valued layers in ANNs. Let " + }, + { + "bbox": [ + 55, + 354, + 296, + 451 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{\\mathrm{B}} \\in \\mathbb{B}^{C \\times H \\times W}" + }, + { + "bbox": [ + 55, + 354, + 296, + 451 + ], + "type": "text", + "content": " denote the binary spike layer of the SNN, and " + }, + { + "bbox": [ + 55, + 354, + 296, + 451 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{\\mathrm{R}} \\in \\mathbb{R}^{C \\times H \\times W}" + }, + { + "bbox": [ + 55, + 354, + 296, + 451 + ], + "type": "text", + "content": " denote the real-valued feature map in the corresponding ANN. For a binary spike output " + }, + { + "bbox": [ + 55, + 354, + 296, + 451 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 55, + 354, + 296, + 451 + ], + "type": "text", + "content": ", it requires 1 bit, and thus the number of possible samples from " + }, + { + "bbox": [ + 55, + 354, + 296, + 451 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 55, + 354, + 296, + 451 + ], + "type": "text", + "content": " is 2. Therefore, the number of samples for " + }, + { + "bbox": [ + 55, + 354, + 296, + 451 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{\\mathrm{B}}" + }, + { + "bbox": [ + 55, + 354, + 296, + 451 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 55, + 354, + 296, + 451 + ], + "type": "inline_equation", + "content": "2^{(C \\times H \\times W)}" + }, + { + "bbox": [ + 55, + 354, + 296, + 451 + ], + "type": "text", + "content": ", leading to:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 75, + 456, + 296, + 476 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 456, + 296, + 476 + ], + "spans": [ + { + "bbox": [ + 75, + 456, + 296, + 476 + ], + "type": "interline_equation", + "content": "\\mathcal {C} \\left(\\mathbf {F} _ {\\mathrm {B}}\\right) = \\log \\left(2 ^ {(C \\times H \\times W)}\\right) = C \\times H \\times W. \\tag {20}", + "image_path": "1bc19fd01ede4d242c928c326192092557057224c9b94a90755ecd62da09f870.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 480, + 296, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 480, + 296, + 517 + ], + "spans": [ + { + "bbox": [ + 55, + 480, + 296, + 517 + ], + "type": "text", + "content": "In contrast, a real-valued output requires 32 bits per sample, yielding " + }, + { + "bbox": [ + 55, + 480, + 296, + 517 + ], + "type": "inline_equation", + "content": "2^{32}" + }, + { + "bbox": [ + 55, + 480, + 296, + 517 + ], + "type": "text", + "content": " possible values. Hence, the representational capacity for real-valued layers is:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 59, + 521, + 296, + 542 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 521, + 296, + 542 + ], + "spans": [ + { + "bbox": [ + 59, + 521, + 296, + 542 + ], + "type": "interline_equation", + "content": "\\mathcal {C} \\left(\\mathbf {F} _ {\\mathrm {R}}\\right) = \\log \\left(2 ^ {3 2 \\times (C \\times H \\times W)}\\right) = 3 2 \\times C \\times H \\times W. \\tag {21}", + "image_path": "d90ca47ca919906a3e073ec004e5d28a0c5b99c4ff01d0a621165233cc77362a.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 546, + 296, + 631 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 546, + 296, + 631 + ], + "spans": [ + { + "bbox": [ + 55, + 546, + 296, + 631 + ], + "type": "text", + "content": "This highlights the limited representational capacity of the binary spike layer. By transforming the real-valued " + }, + { + "bbox": [ + 55, + 546, + 296, + 631 + ], + "type": "inline_equation", + "content": "Q_{\\mathbb{R}}" + }, + { + "bbox": [ + 55, + 546, + 296, + 631 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 546, + 296, + 631 + ], + "type": "inline_equation", + "content": "K_{\\mathbb{R}}" + }, + { + "bbox": [ + 55, + 546, + 296, + 631 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 546, + 296, + 631 + ], + "type": "inline_equation", + "content": "V_{\\mathbb{R}}" + }, + { + "bbox": [ + 55, + 546, + 296, + 631 + ], + "type": "text", + "content": " into binary " + }, + { + "bbox": [ + 55, + 546, + 296, + 631 + ], + "type": "inline_equation", + "content": "Q_{\\mathbb{B}}" + }, + { + "bbox": [ + 55, + 546, + 296, + 631 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 546, + 296, + 631 + ], + "type": "inline_equation", + "content": "K_{\\mathbb{B}}" + }, + { + "bbox": [ + 55, + 546, + 296, + 631 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 546, + 296, + 631 + ], + "type": "inline_equation", + "content": "V_{\\mathbb{B}}" + }, + { + "bbox": [ + 55, + 546, + 296, + 631 + ], + "type": "text", + "content": ", significant information loss occurs, reducing accuracy. Therefore, we propose increasing the bit precision of " + }, + { + "bbox": [ + 55, + 546, + 296, + 631 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 55, + 546, + 296, + 631 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 546, + 296, + 631 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 546, + 296, + 631 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 546, + 296, + 631 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 55, + 546, + 296, + 631 + ], + "type": "text", + "content": " in spiking attention to minimize information loss, while preserving the advantages of multiplication-addition transformations." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 636, + 296, + 661 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 636, + 296, + 661 + ], + "spans": [ + { + "bbox": [ + 55, + 636, + 296, + 661 + ], + "type": "text", + "content": "4.3. Accurate Addition-Only Spiking Self-Attention Mechanism (A" + }, + { + "bbox": [ + 55, + 636, + 296, + 661 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 55, + 636, + 296, + 661 + ], + "type": "text", + "content": "OS" + }, + { + "bbox": [ + 55, + 636, + 296, + 661 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 55, + 636, + 296, + 661 + ], + "type": "text", + "content": "A)" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 665, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 297, + 714 + ], + "type": "text", + "content": "To maintain the integrity of information within spiking attention while leveraging the advantages of multiplication-addition transformations, we propose the Accurate Addition-Only Spiking Self-Attention " + }, + { + "bbox": [ + 55, + 665, + 297, + 714 + ], + "type": "inline_equation", + "content": "(\\mathrm{A}^2\\mathrm{OS}^2\\mathrm{A})" + }, + { + "bbox": [ + 55, + 665, + 297, + 714 + ], + "type": "text", + "content": ". This mechanism" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 72, + 555, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 144 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 144 + ], + "type": "text", + "content": "employs binary " + }, + { + "bbox": [ + 313, + 72, + 555, + 144 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 313, + 72, + 555, + 144 + ], + "type": "text", + "content": ", full-precision " + }, + { + "bbox": [ + 313, + 72, + 555, + 144 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 72, + 555, + 144 + ], + "type": "text", + "content": ", and ternary " + }, + { + "bbox": [ + 313, + 72, + 555, + 144 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 313, + 72, + 555, + 144 + ], + "type": "text", + "content": " in the self-attention process. Specifically, we utilize a binary spiking neuron after the linear layer to produce " + }, + { + "bbox": [ + 313, + 72, + 555, + 144 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 313, + 72, + 555, + 144 + ], + "type": "text", + "content": ", a ReLU function for generating " + }, + { + "bbox": [ + 313, + 72, + 555, + 144 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 72, + 555, + 144 + ], + "type": "text", + "content": ", and a ternary spiking neuron for producing " + }, + { + "bbox": [ + 313, + 72, + 555, + 144 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 313, + 72, + 555, + 144 + ], + "type": "text", + "content": ", as illustrated in Figure 1(c). The formulations for " + }, + { + "bbox": [ + 313, + 72, + 555, + 144 + ], + "type": "inline_equation", + "content": "Q, K, V" + }, + { + "bbox": [ + 313, + 72, + 555, + 144 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 313, + 72, + 555, + 144 + ], + "type": "inline_equation", + "content": "\\mathrm{A}^2\\mathrm{OS}^2\\mathrm{A}" + }, + { + "bbox": [ + 313, + 72, + 555, + 144 + ], + "type": "text", + "content": " are defined as follows:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 376, + 150, + 555, + 165 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 376, + 150, + 555, + 165 + ], + "spans": [ + { + "bbox": [ + 376, + 150, + 555, + 165 + ], + "type": "interline_equation", + "content": "Q = \\mathcal {S N} _ {Q} ^ {b} (\\mathrm {B N} (X W _ {Q})), \\tag {22}", + "image_path": "fdd5e6d86f66365f2ea8616f6de140ea60ad8640b56241e5dd4149b6daab41d3.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 376, + 167, + 555, + 181 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 376, + 167, + 555, + 181 + ], + "spans": [ + { + "bbox": [ + 376, + 167, + 555, + 181 + ], + "type": "interline_equation", + "content": "K = \\operatorname {R e L U} _ {K} (\\operatorname {B N} (X W _ {K})), \\tag {23}", + "image_path": "f797e047ceb4f0b946bd4223d30d2ed7b6d6a68ed65737835ed1842063ea82e9.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 377, + 183, + 555, + 198 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 377, + 183, + 555, + 198 + ], + "spans": [ + { + "bbox": [ + 377, + 183, + 555, + 198 + ], + "type": "interline_equation", + "content": "V = \\mathcal {S N} _ {V} ^ {t} (\\mathrm {B N} (X W _ {V})), \\tag {24}", + "image_path": "b3bc5ebb75e5e20f3f257f1ac6a1de83d92bb6bb97b2666adc93f243c7661ad7.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 204, + 554, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 204, + 554, + 252 + ], + "spans": [ + { + "bbox": [ + 313, + 204, + 554, + 252 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 204, + 554, + 252 + ], + "type": "inline_equation", + "content": "\\mathcal{SN}^b (\\cdot)" + }, + { + "bbox": [ + 313, + 204, + 554, + 252 + ], + "type": "text", + "content": " denotes the binary spiking neuron as used in standard SNNs, " + }, + { + "bbox": [ + 313, + 204, + 554, + 252 + ], + "type": "inline_equation", + "content": "\\mathrm{ReLU}_K(\\cdot)" + }, + { + "bbox": [ + 313, + 204, + 554, + 252 + ], + "type": "text", + "content": " is the ReLU activation function typical in ANNs, and " + }, + { + "bbox": [ + 313, + 204, + 554, + 252 + ], + "type": "inline_equation", + "content": "\\mathcal{SN}^t (\\cdot)" + }, + { + "bbox": [ + 313, + 204, + 554, + 252 + ], + "type": "text", + "content": " represents a ternary spiking neuron that outputs values in the set " + }, + { + "bbox": [ + 313, + 204, + 554, + 252 + ], + "type": "inline_equation", + "content": "\\{-1,0,1\\}" + }, + { + "bbox": [ + 313, + 204, + 554, + 252 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 253, + 554, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 253, + 554, + 289 + ], + "spans": [ + { + "bbox": [ + 313, + 253, + 554, + 289 + ], + "type": "text", + "content": "The evolution of the membrane potential and the spike generation mechanism within " + }, + { + "bbox": [ + 313, + 253, + 554, + 289 + ], + "type": "inline_equation", + "content": "\\mathcal{SN}^t(\\cdot)" + }, + { + "bbox": [ + 313, + 253, + 554, + 289 + ], + "type": "text", + "content": " can be described by the following equations:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 356, + 296, + 555, + 309 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 356, + 296, + 555, + 309 + ], + "spans": [ + { + "bbox": [ + 356, + 296, + 555, + 309 + ], + "type": "interline_equation", + "content": "U [ t ] = H [ t - 1 ] + X [ t ], \\tag {25}", + "image_path": "51df8af2a2e7678800336d7cb61bbd1d77c8ae92723a9645d07af2324c381dee.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 358, + 311, + 555, + 324 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 311, + 555, + 324 + ], + "spans": [ + { + "bbox": [ + 358, + 311, + 555, + 324 + ], + "type": "interline_equation", + "content": "S [ t ] = \\operatorname {H e a} (| U [ t ] | - V _ {\\mathrm {t h}}), \\tag {26}", + "image_path": "fb54a5f65af9ed1d8ba0777b61dd30d6f2f3522d6e18edd14fbaa3ced22530e5.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 356, + 326, + 555, + 339 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 356, + 326, + 555, + 339 + ], + "spans": [ + { + "bbox": [ + 356, + 326, + 555, + 339 + ], + "type": "interline_equation", + "content": "H [ t ] = V _ {\\text {r e s e t}} S [ t ] + \\beta U [ t ] (1 - | S [ t ] |). \\tag {27}", + "image_path": "c06b8b453ab3ace1afaf25f1c2fe1cbd37e603b1151d2b067d81e69504cfbd9b.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 346, + 511, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 346, + 511, + 357 + ], + "spans": [ + { + "bbox": [ + 313, + 346, + 511, + 357 + ], + "type": "text", + "content": "The formulation of the " + }, + { + "bbox": [ + 313, + 346, + 511, + 357 + ], + "type": "inline_equation", + "content": "\\mathrm{A}^2\\mathrm{OS}^2\\mathrm{A}" + }, + { + "bbox": [ + 313, + 346, + 511, + 357 + ], + "type": "text", + "content": " is expressed as:" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 348, + 364, + 555, + 379 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 348, + 364, + 555, + 379 + ], + "spans": [ + { + "bbox": [ + 348, + 364, + 555, + 379 + ], + "type": "interline_equation", + "content": "\\mathrm {A} ^ {2} \\mathrm {O S} ^ {2} \\mathrm {A} (Q, K, V) = \\mathcal {S N} \\left(Q \\cdot K ^ {T} \\cdot V\\right). \\tag {28}", + "image_path": "736cf192bd2c4de80eaf2f78ccb6badf95ed187e437e9593d09ce8c5202f8f1c.jpg" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 384, + 556, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 384, + 556, + 586 + ], + "spans": [ + { + "bbox": [ + 313, + 384, + 556, + 586 + ], + "type": "text", + "content": "In " + }, + { + "bbox": [ + 313, + 384, + 556, + 586 + ], + "type": "inline_equation", + "content": "\\mathrm{A}^2\\mathrm{OS}^2\\mathrm{A}" + }, + { + "bbox": [ + 313, + 384, + 556, + 586 + ], + "type": "text", + "content": ", since " + }, + { + "bbox": [ + 313, + 384, + 556, + 586 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 313, + 384, + 556, + 586 + ], + "type": "text", + "content": " takes values in " + }, + { + "bbox": [ + 313, + 384, + 556, + 586 + ], + "type": "inline_equation", + "content": "\\{0,1\\}" + }, + { + "bbox": [ + 313, + 384, + 556, + 586 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 384, + 556, + 586 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 384, + 556, + 586 + ], + "type": "text", + "content": " is in " + }, + { + "bbox": [ + 313, + 384, + 556, + 586 + ], + "type": "inline_equation", + "content": "\\mathbb{R}" + }, + { + "bbox": [ + 313, + 384, + 556, + 586 + ], + "type": "text", + "content": ", the matrix multiplication " + }, + { + "bbox": [ + 313, + 384, + 556, + 586 + ], + "type": "inline_equation", + "content": "Q\\cdot K^{T}" + }, + { + "bbox": [ + 313, + 384, + 556, + 586 + ], + "type": "text", + "content": " can be transformed into equivalent addition operations, resulting in " + }, + { + "bbox": [ + 313, + 384, + 556, + 586 + ], + "type": "inline_equation", + "content": "Q\\cdot K^{T}\\in \\mathbb{R}" + }, + { + "bbox": [ + 313, + 384, + 556, + 586 + ], + "type": "text", + "content": ". Moreover, because both " + }, + { + "bbox": [ + 313, + 384, + 556, + 586 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 313, + 384, + 556, + 586 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 384, + 556, + 586 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 384, + 556, + 586 + ], + "type": "text", + "content": " are naturally non-negative, this preserves a non-negative attention map " + }, + { + "bbox": [ + 313, + 384, + 556, + 586 + ], + "type": "inline_equation", + "content": "Q\\cdot K^T" + }, + { + "bbox": [ + 313, + 384, + 556, + 586 + ], + "type": "text", + "content": ". Consequently, the mechanism eliminates the need for a softmax operation to enforce non-negativity within the attention map. Furthermore, since " + }, + { + "bbox": [ + 313, + 384, + 556, + 586 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 384, + 556, + 586 + ], + "type": "text", + "content": " is a real-valued vector without boundaries, there is no necessity for a scaling factor " + }, + { + "bbox": [ + 313, + 384, + 556, + 586 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 313, + 384, + 556, + 586 + ], + "type": "text", + "content": " to manage large values resulting from the matrix multiplication, unlike the approaches in [64, 65]. Additionally, given that " + }, + { + "bbox": [ + 313, + 384, + 556, + 586 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 313, + 384, + 556, + 586 + ], + "type": "text", + "content": " can also take values in " + }, + { + "bbox": [ + 313, + 384, + 556, + 586 + ], + "type": "inline_equation", + "content": "\\{-1,0,1\\}" + }, + { + "bbox": [ + 313, + 384, + 556, + 586 + ], + "type": "text", + "content": ", the matrix product " + }, + { + "bbox": [ + 313, + 384, + 556, + 586 + ], + "type": "inline_equation", + "content": "Q\\cdot K^{T}\\cdot V" + }, + { + "bbox": [ + 313, + 384, + 556, + 586 + ], + "type": "text", + "content": " can similarly be expressed through addition operations. At the same time, unlike other matrix multiplication methods outlined in [64, 65], our framework allows for the result to include negative values, akin to what is seen in traditional ANNs." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 587, + 556, + 659 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 587, + 556, + 659 + ], + "spans": [ + { + "bbox": [ + 313, + 587, + 556, + 659 + ], + "type": "text", + "content": "In summary, the " + }, + { + "bbox": [ + 313, + 587, + 556, + 659 + ], + "type": "inline_equation", + "content": "\\mathrm{A}^2\\mathrm{OS}^2\\mathrm{A}" + }, + { + "bbox": [ + 313, + 587, + 556, + 659 + ], + "type": "text", + "content": " mechanism utilizes full-precision " + }, + { + "bbox": [ + 313, + 587, + 556, + 659 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 587, + 556, + 659 + ], + "type": "text", + "content": " and ternary spike " + }, + { + "bbox": [ + 313, + 587, + 556, + 659 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 313, + 587, + 556, + 659 + ], + "type": "text", + "content": " to enhance the representational capability of self-attention and minimize information loss. Additionally, all attention calculations are performed using addition operations, which align with the characteristics of SNNs." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 669, + 391, + 683 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 669, + 391, + 683 + ], + "spans": [ + { + "bbox": [ + 313, + 669, + 391, + 683 + ], + "type": "text", + "content": "5. Experiment" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": "We evaluate our method on various datasets, including CIFAR-10/100 [31] and ImageNet-1K [8]. The network" + } + ] + } + ], + "index": 27 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "24403" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 134, + 300, + 205 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 297, + 116 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 297, + 116 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 297, + 116 + ], + "type": "text", + "content": "Table 1. Ablation study for Spiking Transformer on CIFAR10/100. Param refers to the number of parameters. Spiking Transformer- " + }, + { + "bbox": [ + 55, + 70, + 297, + 116 + ], + "type": "inline_equation", + "content": "L - D" + }, + { + "bbox": [ + 55, + 70, + 297, + 116 + ], + "type": "text", + "content": " is a Spiking Transformer model with " + }, + { + "bbox": [ + 55, + 70, + 297, + 116 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 55, + 70, + 297, + 116 + ], + "type": "text", + "content": " Spiking Transformer encoder blocks and " + }, + { + "bbox": [ + 55, + 70, + 297, + 116 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 55, + 70, + 297, + 116 + ], + "type": "text", + "content": " feature embedding dimensions." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 134, + 300, + 205 + ], + "lines": [ + { + "bbox": [ + 56, + 134, + 300, + 205 + ], + "spans": [ + { + "bbox": [ + 56, + 134, + 300, + 205 + ], + "type": "table", + "html": "
MethodsParam (M)Time StepCIFAR10 AccCIFAR100 Acc
Baseline-2-2562.59494.3976.00
Spiking Transformer-2-2562.59494.9176.96
Baseline-2-51210.23495.5178.83
Spiking Transformer-2-51210.23496.4279.90
", + "image_path": "7b2922acaeb01d1eb432e88162c8c5d8ca7a4647cc5ee30b928f217683d8e7b5.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 234, + 296, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 234, + 296, + 283 + ], + "spans": [ + { + "bbox": [ + 55, + 234, + 296, + 283 + ], + "type": "text", + "content": "architecture and experimental setup follow the baseline from Spike-driven Transformer [58]. Further details regarding the experimental settings can be found in the Spike-driven Transformer [58]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 289, + 149, + 301 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 289, + 149, + 301 + ], + "spans": [ + { + "bbox": [ + 55, + 289, + 149, + 301 + ], + "type": "text", + "content": "5.1. Ablation Study" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 306, + 296, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 306, + 296, + 354 + ], + "spans": [ + { + "bbox": [ + 55, + 306, + 296, + 354 + ], + "type": "text", + "content": "We conduct a series of ablation experiments to assess the effectiveness of the proposed Spiking Transformer, comparing it with the baseline model from [58] on the CIFAR-10 and CIFAR-100 datasets. The results are summarized in Table 1." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 354, + 297, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 354, + 297, + 534 + ], + "spans": [ + { + "bbox": [ + 55, + 354, + 297, + 534 + ], + "type": "text", + "content": "The baseline models used are Spike-driven Transformer-2-256 and Spike-driven Transformer-2-512. For the Transformer-2-256 with 4 timesteps, the baseline accuracy is " + }, + { + "bbox": [ + 55, + 354, + 297, + 534 + ], + "type": "inline_equation", + "content": "94.39\\%" + }, + { + "bbox": [ + 55, + 354, + 297, + 534 + ], + "type": "text", + "content": " on CIFAR-10 and " + }, + { + "bbox": [ + 55, + 354, + 297, + 534 + ], + "type": "inline_equation", + "content": "76.00\\%" + }, + { + "bbox": [ + 55, + 354, + 297, + 534 + ], + "type": "text", + "content": " on CIFAR-100, consistent with previous results. Our Spiking Transformer, using the same architecture, achieves notable improvements, with accuracies of " + }, + { + "bbox": [ + 55, + 354, + 297, + 534 + ], + "type": "inline_equation", + "content": "94.91\\%" + }, + { + "bbox": [ + 55, + 354, + 297, + 534 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 354, + 297, + 534 + ], + "type": "inline_equation", + "content": "76.96\\%" + }, + { + "bbox": [ + 55, + 354, + 297, + 534 + ], + "type": "text", + "content": " on CIFAR-10 and CIFAR-100, respectively—resulting in performance gains of approximately " + }, + { + "bbox": [ + 55, + 354, + 297, + 534 + ], + "type": "inline_equation", + "content": "0.5\\%" + }, + { + "bbox": [ + 55, + 354, + 297, + 534 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 354, + 297, + 534 + ], + "type": "inline_equation", + "content": "1.0\\%" + }, + { + "bbox": [ + 55, + 354, + 297, + 534 + ], + "type": "text", + "content": ". For the larger model, Transformer-2-512 with 4 timesteps achieves baseline accuracies of " + }, + { + "bbox": [ + 55, + 354, + 297, + 534 + ], + "type": "inline_equation", + "content": "95.51\\%" + }, + { + "bbox": [ + 55, + 354, + 297, + 534 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 354, + 297, + 534 + ], + "type": "inline_equation", + "content": "78.43\\%" + }, + { + "bbox": [ + 55, + 354, + 297, + 534 + ], + "type": "text", + "content": " on CIFAR-10 and CIFAR-100. Our model with the same architecture shows substantial improvements, reaching accuracy scores of " + }, + { + "bbox": [ + 55, + 354, + 297, + 534 + ], + "type": "inline_equation", + "content": "96.42\\%" + }, + { + "bbox": [ + 55, + 354, + 297, + 534 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 354, + 297, + 534 + ], + "type": "inline_equation", + "content": "79.90\\%" + }, + { + "bbox": [ + 55, + 354, + 297, + 534 + ], + "type": "text", + "content": ", corresponding to gains of approximately " + }, + { + "bbox": [ + 55, + 354, + 297, + 534 + ], + "type": "inline_equation", + "content": "0.9\\%" + }, + { + "bbox": [ + 55, + 354, + 297, + 534 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 354, + 297, + 534 + ], + "type": "inline_equation", + "content": "1.1\\%" + }, + { + "bbox": [ + 55, + 354, + 297, + 534 + ], + "type": "text", + "content": " on CIFAR-10 and CIFAR-100, respectively." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 540, + 113, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 540, + 113, + 552 + ], + "spans": [ + { + "bbox": [ + 55, + 540, + 113, + 552 + ], + "type": "text", + "content": "5.2. CIFAR" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 558, + 297, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 297, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 297, + 713 + ], + "type": "text", + "content": "To further evaluate the performance of our method, we compare it with recent SNN-based Transformer approaches, including Spikformer [64], Spikingformer [63] and so on, using a range of models with varying embedding dimensions and numbers of transformer blocks on the CIFAR dataset. The CIFAR dataset consists of 50,000 training images and 10,000 test images, all with a resolution of " + }, + { + "bbox": [ + 55, + 558, + 297, + 713 + ], + "type": "inline_equation", + "content": "32 \\times 32" + }, + { + "bbox": [ + 55, + 558, + 297, + 713 + ], + "type": "text", + "content": ". We maintain the experimental setup from [58, 63, 64], including network architecture, training configurations, etc. Table 2 presents the accuracy of our method compared to other models on the CIFAR dataset. As shown in Table 2, our Spiking Transformer-2-384 achieves an accuracy of " + }, + { + "bbox": [ + 55, + 558, + 297, + 713 + ], + "type": "inline_equation", + "content": "95.70\\%" + }, + { + "bbox": [ + 55, + 558, + 297, + 713 + ], + "type": "text", + "content": " on the CIFAR-10 dataset, surpassing TET " + }, + { + "bbox": [ + 55, + 558, + 297, + 713 + ], + "type": "inline_equation", + "content": "(94.44\\%)" + }, + { + "bbox": [ + 55, + 558, + 297, + 713 + ], + "type": "text", + "content": " and even" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 315, + 113, + 560, + 311 + ], + "blocks": [ + { + "bbox": [ + 313, + 70, + 556, + 94 + ], + "lines": [ + { + "bbox": [ + 313, + 70, + 556, + 94 + ], + "spans": [ + { + "bbox": [ + 313, + 70, + 556, + 94 + ], + "type": "text", + "content": "Table 2. Comparison of the performance between Spiking Transformer and existing approaches on CIFAR10/100." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 113, + 560, + 311 + ], + "lines": [ + { + "bbox": [ + 315, + 113, + 560, + 311 + ], + "spans": [ + { + "bbox": [ + 315, + 113, + 560, + 311 + ], + "type": "table", + "html": "
MethodsParam (M)Time StepCIFAR10 AccCIFAR100 Acc
Hybrid training[46]9.2712592.2267.87
Diet-SNN[45]0.2710/592.5464.07
STBP[55]17.541289.83-
STBP NeuNorm[54]17.541290.53-
TSSL-BP[60]17.54591.41-
STBP-tdBN[62]12.63492.9270.86
TET[9]12.63494.4474.47
Spikformer-4-256[64]4.15493.9475.96
Spikformer-2-384[64]5.76494.8076.95
Spikformer-4-384[64]9.32495.1977.86
Spikingformer-4-256[63]4.15494.7777.43
Spikingformer-2-384[63]5.76495.2278.34
Spikingformer-4-384[63]9.32495.6179.09
Spiking Transformer-4-2564.15494.9677.49
Spiking Transformer-2-3845.76495.7078.59
Spiking Transformer-4-3849.32496.3279.69
Spiking Transformer-2-51210.23496.4279.90
", + "image_path": "9266591a827ac6208d6d971c4037cf5a1e8af6006766d77204cdbf1c1244f458.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 341, + 556, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 341, + 556, + 449 + ], + "spans": [ + { + "bbox": [ + 313, + 341, + 556, + 449 + ], + "type": "text", + "content": "Spikingformer-4-384 (94.97%). Notably, performance improves further with higher embedding dimensions and more transformer blocks. In particular, Spiking Transformer-4-384 shows a 1.13% improvement over Spikformer-4-384 and a 0.71% improvement over Spikingformer-4-384. The performance gain of our Spiking Transformer is even more pronounced on more complex datasets such as CIFAR-100. Specifically, Spiking Transformer-4-384 achieves a notable improvement of 1.83% over Spikformer-4-384." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 456, + 383, + 468 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 456, + 383, + 468 + ], + "spans": [ + { + "bbox": [ + 313, + 456, + 383, + 468 + ], + "type": "text", + "content": "5.3. ImageNet" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 474, + 556, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 474, + 556, + 593 + ], + "spans": [ + { + "bbox": [ + 313, + 474, + 556, + 593 + ], + "type": "text", + "content": "We further evaluated our approach on the challenging ImageNet-1K dataset. ImageNet-1K [8] is a widely-used benchmark for image classification, consisting of 1,000 categories. The dataset includes approximately 1.28 million training images and 50,000 test images. For both training and evaluation, the images are resized to a default resolution of " + }, + { + "bbox": [ + 313, + 474, + 556, + 593 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 313, + 474, + 556, + 593 + ], + "type": "text", + "content": " pixels. To ensure consistency with previous work, we maintain the experimental setup from [58, 63, 64], including network architecture, training configurations, and other experimental details." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 594, + 557, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 557, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 557, + 715 + ], + "type": "text", + "content": "Table 3 presents a comparative analysis of recent SoTA methods evaluated on the ImageNet dataset. Notable models in this domain include Spikformer [64], Spikingformer [63], and Spike-driven Transformer [58], which achieve accuracies of " + }, + { + "bbox": [ + 313, + 594, + 557, + 715 + ], + "type": "inline_equation", + "content": "73.38\\%" + }, + { + "bbox": [ + 313, + 594, + 557, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 594, + 557, + 715 + ], + "type": "inline_equation", + "content": "74.79\\%" + }, + { + "bbox": [ + 313, + 594, + 557, + 715 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 594, + 557, + 715 + ], + "type": "inline_equation", + "content": "74.57\\%" + }, + { + "bbox": [ + 313, + 594, + 557, + 715 + ], + "type": "text", + "content": ", respectively, with Transformer-8-512 as the baseline configuration. In comparison, our approach demonstrates a significant improvement, achieving an accuracy of " + }, + { + "bbox": [ + 313, + 594, + 557, + 715 + ], + "type": "inline_equation", + "content": "76.28\\%" + }, + { + "bbox": [ + 313, + 594, + 557, + 715 + ], + "type": "text", + "content": ", which represents a " + }, + { + "bbox": [ + 313, + 594, + 557, + 715 + ], + "type": "inline_equation", + "content": "1.49\\%" + }, + { + "bbox": [ + 313, + 594, + 557, + 715 + ], + "type": "text", + "content": " increase over the best-performing transformer-based SNN model, Spikingformer. This improvement highlights the" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "24404" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 95, + 100, + 519, + 479 + ], + "blocks": [ + { + "bbox": [ + 100, + 70, + 509, + 83 + ], + "lines": [ + { + "bbox": [ + 100, + 70, + 509, + 83 + ], + "spans": [ + { + "bbox": [ + 100, + 70, + 509, + 83 + ], + "type": "text", + "content": "Table 3. Comparison of the performance between Spiking Transformer and existing approaches on ImageNet-1k." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 95, + 100, + 519, + 479 + ], + "lines": [ + { + "bbox": [ + 95, + 100, + 519, + 479 + ], + "spans": [ + { + "bbox": [ + 95, + 100, + 519, + 479 + ], + "type": "table", + "html": "
CategoryMethodsArchitectureParam (M)Time StepAcc
ANN-to-SNNHybrid training[46]ResNet-3421.7925061.48
Spiking ResNet[27]ResNet-3421.7935071.61
ResNet-5025.5635072.75
QCFS[2]VGG-16138.426472.85
COS[24]ResNet-3421.79874.17
Directly LearningTET[9]Spiking-ResNet-3421.79664.79
SEW-ResNet-3421.79468.00
STBP-tdBN[62]Spiking-ResNet-3421.79663.72
SEW-ResNet-3421.79467.04
SEW ResNet[11]SEW-ResNet-5025.56467.78
SEW-ResNet-10144.55468.76
SEW-ResNet-15260.19469.26
Attention-SNN[57]ResNet-10478.37477.08
Directly LearningSpikformer[64]Spikformer-8-38416.81470.24
Spikformer-6-51223.37472.46
Spikformer-8-51229.68473.38
Spikformer-10-51236.01473.68
Spikformer-8-76866.34474.81
Directly LearningSpikingformer[63]Spikingformer-8-38416.81472.45
Spikingformer-8-51229.68474.79
Spikingformer-8-76866.34475.85
Directly LearningSpike-driven Transformer[58]Spike-driven Transformer-8-38416.81472.28
Spike-driven Transformer-6-51223.37474.11
Spike-driven Transformer-8-51229.68474.57
Spike-driven Transformer-10-51236.01474.66
Spike-driven Transformer-8-76866.34477.07
Directly LearningSpiking TransformerSpiking Transformer-8-38416.81474.04
Spiking Transformer-6-51223.37476.22
Spiking Transformer-8-51229.68476.28
Spiking Transformer-10-51236.01478.66
", + "image_path": "abbd8dac46e774fae487ce791a0b65c1d9dd9f89ef0d95f2bcb01ac94b4a0f50.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 507, + 295, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 507, + 295, + 555 + ], + "spans": [ + { + "bbox": [ + 55, + 507, + 295, + 555 + ], + "type": "text", + "content": "effectiveness of our method in leveraging the potential of spiking neuron models within the transformer architecture, resulting in enhanced performance on large-scale datasets like ImageNet-1K." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 556, + 295, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 556, + 295, + 640 + ], + "spans": [ + { + "bbox": [ + 55, + 556, + 295, + 640 + ], + "type": "text", + "content": "Furthermore, even with a smaller network architecture, our Spiking Transformer-10-512 model outperforms the other methods, including those using Transformer-8-768 as the baseline. This demonstrates that our approach is not only more accurate but also more efficient, as it achieves superior results with fewer parameters, highlighting its scalability and robustness." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 656, + 128, + 668 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 656, + 128, + 668 + ], + "spans": [ + { + "bbox": [ + 55, + 656, + 128, + 668 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": "In this paper, we have introduced the Accurate Addition-Only Spiking Self-Attention " + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "inline_equation", + "content": "(\\mathrm{A}^2\\mathrm{OS}^2\\mathrm{A})" + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": " mechanism as a novel method for integrating Spiking Neural Networks with" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 507, + 555, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 507, + 555, + 604 + ], + "spans": [ + { + "bbox": [ + 313, + 507, + 555, + 604 + ], + "type": "text", + "content": "Transformer models. Our approach tackles the inherent challenge of adapting the self-attention mechanism to the spiking paradigm by proposing a hybrid solution that combines binary, ReLU, and ternary spiking neurons. This innovative hybrid strategy enables the elimination of both softmax and scaling operations, which are traditionally essential in Transformer models, while preserving the fundamental advantages of addition-only computations in SNNs." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "content": "Through extensive experiments, we have demonstrated that " + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{A}^2\\mathrm{OS}^2\\mathrm{A}" + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "content": " outperforms existing SNN-based Transformer models in terms of both accuracy and efficiency. Specifically, our method achieves a SoTA accuracy of " + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "inline_equation", + "content": "78.66\\%" + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "content": " on the ImageNet-1K dataset. This performance, combined with the energy-efficient nature of SNNs, makes " + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{A}^2\\mathrm{OS}^2\\mathrm{A}" + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "content": " a promising candidate for real-world applications, especially those requiring high computational performance with low energy consumption." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "24405" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "type": "text", + "content": "Acknowledgment" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 90, + 297, + 124 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 90, + 297, + 124 + ], + "spans": [ + { + "bbox": [ + 55, + 90, + 297, + 124 + ], + "type": "text", + "content": "This work is supported by grants from the National Natural Science Foundation of China under contracts No.12202412 and No.12202413." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 146, + 115, + 159 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 146, + 115, + 159 + ], + "spans": [ + { + "bbox": [ + 56, + 146, + 115, + 159 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 166, + 297, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 166, + 297, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 166, + 297, + 243 + ], + "spans": [ + { + "bbox": [ + 61, + 166, + 297, + 243 + ], + "type": "text", + "content": "[1] Filipp Akopyan, Jun Sawada, Andrew Cassidy, Rodrigo Alvarez-Icaza, John Arthur, Paul Merolla, Nabil Imam, Yutaka Nakamura, Pallab Datta, Gi-Joon Nam, et al. Truenorth: Design and tool flow of a 65 mw 1 million neuron programmable neurosynaptic chip. IEEE transactions on computer-aided design of integrated circuits and systems, 34(10):1537-1557, 2015. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 243, + 296, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 243, + 296, + 297 + ], + "spans": [ + { + "bbox": [ + 61, + 243, + 296, + 297 + ], + "type": "text", + "content": "[2] Tong Bu, Wei Fang, Jianhao Ding, PengLin Dai, Zhaofei Yu, and Tiejun Huang. Optimal ann-snn conversion for high-accuracy and ultra-low-latency spiking neural networks. In International Conference on Learning Representations, 2021. 2, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 300, + 296, + 343 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 300, + 296, + 343 + ], + "spans": [ + { + "bbox": [ + 62, + 300, + 296, + 343 + ], + "type": "text", + "content": "[3] Tong Bu, Jianhao Ding, Zhaofei Yu, and Tiejun Huang. Optimized potential initialization for low-latency spiking neural networks. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 11-20, 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 345, + 295, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 345, + 295, + 376 + ], + "spans": [ + { + "bbox": [ + 62, + 345, + 295, + 376 + ], + "type": "text", + "content": "[4] Natalia Caporale and Yang Dan. Spike timing-dependent plasticity: a hebbian learning rule. Annu. Rev. Neurosci., 31 (1):25-46, 2008. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 378, + 296, + 433 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 378, + 296, + 433 + ], + "spans": [ + { + "bbox": [ + 62, + 378, + 296, + 433 + ], + "type": "text", + "content": "[5] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 434, + 296, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 434, + 296, + 487 + ], + "spans": [ + { + "bbox": [ + 62, + 434, + 296, + 487 + ], + "type": "text", + "content": "[6] Chun-Fu Richard Chen, Quanfu Fan, and Rameswar Panda. Crossvit: Cross-attention multi-scale vision transformer for image classification. In Proceedings of the IEEE/CVF international conference on computer vision, pages 357–366, 2021. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 490, + 296, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 490, + 296, + 544 + ], + "spans": [ + { + "bbox": [ + 62, + 490, + 296, + 544 + ], + "type": "text", + "content": "[7] Mike Davies, Narayan Srinivasa, Tsung-Han Lin, Gautham Chinya, Yongqiang Cao, Sri Harsha Choday, Georgios Dimou, Prasad Joshi, Nabil Imam, Shweta Jain, et al. Loihi: A neuromorphic manycore processor with on-chip learning. *IEEE Micro*, 38(1):82-99, 2018. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 545, + 296, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 545, + 296, + 590 + ], + "spans": [ + { + "bbox": [ + 62, + 545, + 296, + 590 + ], + "type": "text", + "content": "[8] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 6, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 591, + 296, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 591, + 296, + 633 + ], + "spans": [ + { + "bbox": [ + 62, + 591, + 296, + 633 + ], + "type": "text", + "content": "[9] Shikuang Deng, Yuhang Li, Shanghang Zhang, and Shi Gu. Temporal efficient training of spiking neural network via gradient re-weighting. arXiv preprint arXiv:2202.11946, 2022.7, 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 635, + 296, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 635, + 296, + 667 + ], + "spans": [ + { + "bbox": [ + 57, + 635, + 296, + 667 + ], + "type": "text", + "content": "[10] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 670, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 294, + 712 + ], + "type": "text", + "content": "[11] Wei Fang, Zhaofei Yu, Yanqi Chen, Tiejun Huang, Timothee Masquelier, and Yonghong Tian. Deep residual learning in spiking neural networks. Advances in Neural Information Processing Systems, 34:21056-21069, 2021. 2, 8" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 316, + 73, + 554, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 128 + ], + "type": "text", + "content": "[12] Yufei Guo, Weihang Peng, Xiaode Liu, Yuanpei Chen, Yuhan Zhang, Xin Tong, Zhou Jie, and Zhe Ma. Enof-snn: Training accurate spiking neural networks via enhancing the output feature. In The Thirty-eighth Annual Conference on Neural Information Processing Systems. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 129, + 553, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 553, + 173 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 553, + 173 + ], + "type": "text", + "content": "[13] Yufei Guo, Yuanpei Chen, Liwen Zhang, Xiaode Liu, Yinglei Wang, Xuhui Huang, and Zhe Ma. IM-loss: Information maximization loss for spiking neural networks. In Advances in Neural Information Processing Systems, 2022." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 175, + 555, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 175, + 555, + 229 + ], + "spans": [ + { + "bbox": [ + 317, + 175, + 555, + 229 + ], + "type": "text", + "content": "[14] Yufei Guo, Yuanpei Chen, Liwen Zhang, YingLei Wang, Xiaode Liu, Xinyi Tong, Yuanyuan Ou, Xuhui Huang, and Zhe Ma. Reducing information loss for spiking neural networks. In Computer Vision - ECCV 2022, pages 36-52, Cham, 2022. Springer Nature Switzerland. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 231, + 554, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 231, + 554, + 296 + ], + "spans": [ + { + "bbox": [ + 317, + 231, + 554, + 296 + ], + "type": "text", + "content": "[15] Yufei Guo, Xinyi Tong, Yuanpei Chen, Liwen Zhang, Xiaode Liu, Zhe Ma, and Xuhui Huang. Recdis-snn: Rectifying membrane potential distribution for directly training spiking neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 326-335, 2022. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 298, + 555, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 298, + 555, + 363 + ], + "spans": [ + { + "bbox": [ + 317, + 298, + 555, + 363 + ], + "type": "text", + "content": "[16] Yufei Guo, Liwen Zhang, Yuanpei Chen, Xinyi Tong, Xiaode Liu, YingLei Wang, Xuhui Huang, and Zhe Ma. Real spike: Learning real-valued spikes for spiking neural networks. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part XII, pages 52-68. Springer, 2022. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 365, + 555, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 365, + 555, + 430 + ], + "spans": [ + { + "bbox": [ + 317, + 365, + 555, + 430 + ], + "type": "text", + "content": "[17] Yufei Guo, Xiaode Liu, Yuanpei Chen, Liwen Zhang, Weihang Peng, Yuhan Zhang, Xuhui Huang, and Zhe Ma. Rmploss: Regularizing membrane potential distribution for spiking neural networks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17391-17401, 2023. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 432, + 554, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 432, + 554, + 486 + ], + "spans": [ + { + "bbox": [ + 316, + 432, + 554, + 486 + ], + "type": "text", + "content": "[18] Yufei Guo, Weihang Peng, Yuanpei Chen, Liwen Zhang, Xiaode Liu, Xuhui Huang, and Zhe Ma. Joint a-snn: Joint training of artificial and spiking neural networks via self-distillation and weight factorization. Pattern Recognition, 142:109639, 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 488, + 555, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 488, + 555, + 532 + ], + "spans": [ + { + "bbox": [ + 317, + 488, + 555, + 532 + ], + "type": "text", + "content": "[19] Yufei Guo, Yuhan Zhang, Yuanpei Chen, Weihang Peng, Xiaode Liu, Liwen Zhang, Xuhui Huang, and Zhe Ma. Membrane potential batch normalization for spiking neural networks. arXiv preprint arXiv:2308.08359, 2023." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 534, + 555, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 534, + 555, + 588 + ], + "spans": [ + { + "bbox": [ + 317, + 534, + 555, + 588 + ], + "type": "text", + "content": "[20] Yufei Guo, Yuanpei Chen, Xiaode Liu, Weihang Peng, Yuhan Zhang, Xuhui Huang, and Zhe Ma. Ternary spike: Learning ternary spikes for spiking neural networks. In Proceedings of the AAAI conference on artificial intelligence, pages 12244-12252, 2024." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 590, + 554, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 590, + 554, + 644 + ], + "spans": [ + { + "bbox": [ + 317, + 590, + 554, + 644 + ], + "type": "text", + "content": "[21] Yufei Guo, Yuanpei Chen, Zecheng Hao, Weihang Peng, Zhou Jie, Yuhan Zhang, Xiaode Liu, and Zhe Ma. Take a shortcut back: Mitigating the gradient vanishing for training spiking neural networks. Advances in Neural Information Processing Systems, 37:24849-24867, 2025. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 647, + 554, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 554, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 554, + 689 + ], + "type": "text", + "content": "[22] Bing Han and Kaushik Roy. Deep spiking neural network: Energy efficiency through time based coding. In European Conference on Computer Vision, pages 388-404. Springer, 2020. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 691, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 554, + 713 + ], + "type": "text", + "content": "[23] Zecheng Hao, Tong Bu, Jianhao Ding, Tiejun Huang, and Zhaofei Yu. Reducing ann-snn conversion error through resid" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "24406" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 77, + 72, + 296, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 72, + 296, + 94 + ], + "spans": [ + { + "bbox": [ + 77, + 72, + 296, + 94 + ], + "type": "text", + "content": "ual membrane potential. arXiv preprint arXiv:2302.02091, 2023." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 296, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 296, + 139 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 296, + 139 + ], + "type": "text", + "content": "[24] Zecheng Hao, Jianhao Ding, Tong Bu, Tiejun Huang, and Zhaofei Yu. Bridging the gap between anns and snns by calibrating offset spikes. arXiv preprint arXiv:2302.10685, 2023. 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 141, + 296, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 141, + 296, + 184 + ], + "spans": [ + { + "bbox": [ + 56, + 141, + 296, + 184 + ], + "type": "text", + "content": "[25] Nguyen-Dong Ho and Ik-Joon Chang. Tcl: an ann-to-snn conversion with trainable clipping layers. In 2021 58th ACM/IEEE Design Automation Conference (DAC), pages 793-798. IEEE, 2021. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 186, + 295, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 186, + 295, + 228 + ], + "spans": [ + { + "bbox": [ + 56, + 186, + 295, + 228 + ], + "type": "text", + "content": "[26] Allan L Hodgkin and Andrew F Huxley. Currents carried by sodium and potassium ions through the membrane of the giant axon of loligo. The Journal of physiology, 116(4):449, 1952. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 232, + 295, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 232, + 295, + 263 + ], + "spans": [ + { + "bbox": [ + 56, + 232, + 295, + 263 + ], + "type": "text", + "content": "[27] Yangfan Hu, Huajin Tang, and Gang Pan. Spiking deep residual networks. IEEE Transactions on Neural Networks and Learning Systems, 34(8):5200-5205, 2021. 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 265, + 295, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 265, + 295, + 319 + ], + "spans": [ + { + "bbox": [ + 56, + 265, + 295, + 319 + ], + "type": "text", + "content": "[28] Asifullah Khan, Zunaira Rauf, Abdul Rehman Khan, Saima Rathore, Saddam Hussain Khan, Najmus Saher Shah, Umair Farooq, Hifsa Asif, Aqsa Asif, Umme Zahoora, et al. A recent survey of vision transformers for medical image segmentation. arXiv preprint arXiv:2312.00634, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 321, + 295, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 321, + 295, + 354 + ], + "spans": [ + { + "bbox": [ + 56, + 321, + 295, + 354 + ], + "type": "text", + "content": "[29] Seijoon Kim, Seongsik Park, Byunggook Na, and Sungroh Yoon. Spiking-yolo: Spiking neural network for energy-efficient object detection, 2019. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 355, + 296, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 355, + 296, + 410 + ], + "spans": [ + { + "bbox": [ + 56, + 355, + 296, + 410 + ], + "type": "text", + "content": "[30] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 411, + 295, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 411, + 295, + 443 + ], + "spans": [ + { + "bbox": [ + 56, + 411, + 295, + 443 + ], + "type": "text", + "content": "[31] Alex Krizhevsky, Vinod Nair, and Geoffrey Hinton. Cifar-10 (canadian institute for advanced research). URL http://www.cs.toronto.edu/kriz/cifar.html, 5(4):1, 2010. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 445, + 295, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 445, + 295, + 488 + ], + "spans": [ + { + "bbox": [ + 56, + 445, + 295, + 488 + ], + "type": "text", + "content": "[32] Yuxiang Lan, Yachao Zhang, Xu Ma, Yanyun Qu, and Yun Fu. Efficient converted spiking neural network for 3d and 2d classification. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9211-9220, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 490, + 295, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 490, + 295, + 533 + ], + "spans": [ + { + "bbox": [ + 56, + 490, + 295, + 533 + ], + "type": "text", + "content": "[33] Nathan Leroux, Jan Finkbeiner, and Emre Neftci. Online transformers with spiking neurons for fast prosthetic hand control. In 2023 IEEE Biomedical Circuits and Systems Conference (BioCAS), pages 1-6. IEEE, 2023. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 535, + 295, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 535, + 295, + 578 + ], + "spans": [ + { + "bbox": [ + 56, + 535, + 295, + 578 + ], + "type": "text", + "content": "[34] Yuhang Li, Shikuang Deng, Xin Dong, Ruihao Gong, and Shi Gu. A free lunch from ann: Towards efficient, accurate spiking neural networks calibration. In International Conference on Machine Learning, pages 6316-6325. PMLR, 2021. 1, 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 580, + 295, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 580, + 295, + 633 + ], + "spans": [ + { + "bbox": [ + 56, + 580, + 295, + 633 + ], + "type": "text", + "content": "[35] Yuhang Li, Yufei Guo, Shanghang Zhang, Shikuang Deng, Yongqing Hai, and Shi Gu. Differentiable spike: Rethinking gradient-descent for training spiking neural networks. Advances in Neural Information Processing Systems, 34:23426-23439, 2021. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 635, + 295, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 295, + 668 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 295, + 668 + ], + "type": "text", + "content": "[36] Yawei Li, Kai Zhang, Jiezhang Cao, Radu Timofte, and Luc Van Gool. Localvit: Bringing locality to vision transformers. arXiv preprint arXiv:2104.05707, 2021. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 670, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 295, + 712 + ], + "type": "text", + "content": "[37] Yudong Li, Yunlin Lei, and Xu Yang. Spikeformer: a novel architecture for training high-performance low-latency spiking neural network. arXiv preprint arXiv:2211.10686, 2022.3" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 316, + 73, + 555, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 555, + 117 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 555, + 117 + ], + "type": "text", + "content": "[38] Yanghao Li, Hanzi Mao, Ross Girshick, and Kaiming He. Exploring plain vision transformer backbones for object detection. In European conference on computer vision, pages 280-296. Springer, 2022. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 119, + 555, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 119, + 555, + 173 + ], + "spans": [ + { + "bbox": [ + 316, + 119, + 555, + 173 + ], + "type": "text", + "content": "[39] De Ma, Juncheng Shen, Zonghua Gu, Ming Zhang, Xiaolei Zhu, Xiaoqiang Xu, Qi Xu, Yangjing Shen, and Gang Pan. Darwin: A neuromorphic hardware co-processor based on spiking neural networks. Journal of Systems Architecture, 77: 43-51, 2017. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 175, + 553, + 208 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 553, + 208 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 553, + 208 + ], + "type": "text", + "content": "[40] Wolfgang Maass. Networks of spiking neurons: the third generation of neural network models. Neural networks, 10 (9):1659-1671, 1997. 2, 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 210, + 555, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 210, + 555, + 243 + ], + "spans": [ + { + "bbox": [ + 316, + 210, + 555, + 243 + ], + "type": "text", + "content": "[41] ZHI Min and LU Jingfang. A review of vision transformer for image classification. Journal of Zhengzhou University: Engineering Science, 45(4), 2024. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 245, + 555, + 288 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 245, + 555, + 288 + ], + "spans": [ + { + "bbox": [ + 316, + 245, + 555, + 288 + ], + "type": "text", + "content": "[42] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4195-4205, 2023. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 291, + 553, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 291, + 553, + 334 + ], + "spans": [ + { + "bbox": [ + 316, + 291, + 553, + 334 + ], + "type": "text", + "content": "[43] Jing Pei, Lei Deng, Sen Song, Mingguo Zhao, Youhui Zhang, Shuang Wu, Guanrui Wang, Zhe Zou, Zhenzhi Wu, Wei He, et al. Towards artificial general intelligence with hybrid tianjin chip architecture. Nature, 572(7767):106-111, 2019. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 336, + 553, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 336, + 553, + 369 + ], + "spans": [ + { + "bbox": [ + 316, + 336, + 553, + 369 + ], + "type": "text", + "content": "[44] Jinye Qu, Zeyu Gao, Tielin Zhang, Yanfeng Lu, Huajin Tang, and Hong Qiao. Spiking neural network for ultra-low-latency and high-accurate object detection, 2023. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 371, + 553, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 371, + 553, + 404 + ], + "spans": [ + { + "bbox": [ + 316, + 371, + 553, + 404 + ], + "type": "text", + "content": "[45] Nitin Rathi and Kaushik Roy. Diet-snn: Direct input encoding with leakage and threshold optimization in deep spiking neural networks. arXiv preprint arXiv:2008.03658, 2020. 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 406, + 553, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 406, + 553, + 449 + ], + "spans": [ + { + "bbox": [ + 316, + 406, + 553, + 449 + ], + "type": "text", + "content": "[46] Nitin Rathi, Gopalakrishnan Srinivasan, Priyadarshini Panda, and Kaushik Roy. Enabling deep spiking neural networks with hybrid conversion and spike timing dependent backpropagation. arXiv preprint arXiv:2005.01807, 2020. 7, 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 452, + 553, + 495 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 452, + 553, + 495 + ], + "spans": [ + { + "bbox": [ + 316, + 452, + 553, + 495 + ], + "type": "text", + "content": "[47] Dayong Ren, Zhe Ma, Yuanpei Chen, Weihang Peng, Xiaode Liu, Yuhan Zhang, and Yufei Guo. Spiking pointnet: Spiking neural networks for point clouds. arXiv preprint arXiv:2310.06232, 2023. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 497, + 553, + 540 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 497, + 553, + 540 + ], + "spans": [ + { + "bbox": [ + 316, + 497, + 553, + 540 + ], + "type": "text", + "content": "[48] Abhronil Sengupta, Yuting Ye, Robert Wang, Chiao Liu, and Kaushik Roy. Going deeper in spiking neural networks: Vgg and residual architectures. Frontiers in neuroscience, 13:95, 2019. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 543, + 553, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 543, + 553, + 597 + ], + "spans": [ + { + "bbox": [ + 316, + 543, + 553, + 597 + ], + "type": "text", + "content": "[49] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In International conference on machine learning, pages 10347-10357. PMLR, 2021. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 600, + 553, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 600, + 553, + 621 + ], + "spans": [ + { + "bbox": [ + 316, + 600, + 553, + 621 + ], + "type": "text", + "content": "[50] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 1" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 624, + 553, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 553, + 667 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 553, + 667 + ], + "type": "text", + "content": "[51] Shuai Wang, Malu Zhang, Dehao Zhang, Ammar Belatreche, Yichen Xiao, Yu Liang, Yimeng Shan, Qian Sun, Enqi Zhang, and Yang Yang. Spiking vision transformer with saccadic attention. arXiv preprint arXiv:2502.12677, 2025. 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "text", + "content": "[52] Wenjie Wei, Malu Zhang, Jilin Zhang, Ammar Belatreche, Jibin Wu, Zijing Xu, Xuerui Qiu, Hong Chen, Yang Yang, and Haizhou Li. Event-driven learning for spiking neural networks. arXiv preprint arXiv:2403.00270, 2024.1" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24407" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 297, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 56, + 73, + 297, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 297, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 297, + 116 + ], + "type": "text", + "content": "[53] Yujie Wu, Lei Deng, Guoqi Li, Jun Zhu, and Luping Shi. Spatio-temporal backpropagation for training high-performance spiking neural networks. Frontiers in neuroscience, 12:331, 2018. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 295, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 295, + 161 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 295, + 161 + ], + "type": "text", + "content": "[54] Yujie Wu, Lei Deng, Guoqi Li, Jun Zhu, Yuan Xie, and Luping Shi. Direct training for spiking neural networks: Faster, larger, better. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1311-1318, 2019. 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 162, + 294, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 162, + 294, + 183 + ], + "spans": [ + { + "bbox": [ + 57, + 162, + 294, + 183 + ], + "type": "text", + "content": "[55] Y. et al. Wu. Spatio-temporal learning with spiking neural networks. Nature Communications, 9, 2018. 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 184, + 296, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 184, + 296, + 237 + ], + "spans": [ + { + "bbox": [ + 56, + 184, + 296, + 237 + ], + "type": "text", + "content": "[56] Mingqing Xiao, Qingyan Meng, Zongpeng Zhang, Yisen Wang, and Zhouchen Lin. Training feedback spiking neural networks by implicit differentiation on the equilibrium state. Advances in Neural Information Processing Systems, 34:14516-14528, 2021. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 239, + 295, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 239, + 295, + 282 + ], + "spans": [ + { + "bbox": [ + 56, + 239, + 295, + 282 + ], + "type": "text", + "content": "[57] Man Yao, Guangshe Zhao, Hengyu Zhang, Yifan Hu, Lei Deng, Yonghong Tian, Bo Xu, and Guoqi Li. Attention spiking neural networks. IEEE transactions on pattern analysis and machine intelligence, 45(8):9393-9410, 2023. 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 57, + 284, + 295, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 284, + 295, + 316 + ], + "spans": [ + { + "bbox": [ + 57, + 284, + 295, + 316 + ], + "type": "text", + "content": "[58] Man Yao, Jiakui Hu, Zhaokun Zhou, Li Yuan, Yonghong Tian, Bo Xu, and Guoqi Li. Spike-driven transformer. Advances in neural information processing systems, 36, 2024. 1, 3, 5, 7, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 317, + 295, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 317, + 295, + 392 + ], + "spans": [ + { + "bbox": [ + 56, + 317, + 295, + 392 + ], + "type": "text", + "content": "[59] Malu Zhang, Jiadong Wang, Jibin Wu, Ammar Belatreche, Burin Amornpaisannon, Zhixuan Zhang, Venkata Pavan Kumar Miriyala, Hong Qu, Yansong Chua, Trevor E Carlson, et al. Rectified linear postsynaptic potential function for backpropagation in deep spiking neural networks. IEEE transactions on neural networks and learning systems, 33 (5):1947-1958, 2021. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 57, + 394, + 295, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 394, + 295, + 436 + ], + "spans": [ + { + "bbox": [ + 57, + 394, + 295, + 436 + ], + "type": "text", + "content": "[60] Wenrui Zhang and Peng Li. Temporal spike sequence learning via backpropagation for deep spiking neural networks. Advances in Neural Information Processing Systems, 33:12022-12033, 2020. 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 57, + 437, + 295, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 437, + 295, + 491 + ], + "spans": [ + { + "bbox": [ + 57, + 437, + 295, + 491 + ], + "type": "text", + "content": "[61] Yuhan Zhang, Xiaode Liu, Yuanpei Chen, Weihang Peng, Yufei Guo, Xuhui Huang, and Zhe Ma. Enhancing representation of spiking neural networks via similarity-sensitive contrastive learning. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 16926-16934, 2024. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 57, + 493, + 295, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 493, + 295, + 536 + ], + "spans": [ + { + "bbox": [ + 57, + 493, + 295, + 536 + ], + "type": "text", + "content": "[62] Hanle Zheng, Yujie Wu, Lei Deng, Yifan Hu, and Guoqi Li. Going deeper with directly-trained larger spiking neural networks. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 11062-11070, 2021. 7, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 537, + 295, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 537, + 295, + 590 + ], + "spans": [ + { + "bbox": [ + 57, + 537, + 295, + 590 + ], + "type": "text", + "content": "[63] Chenlin Zhou, Liutao Yu, Zhaokun Zhou, Zhengyu Ma, Han Zhang, Huihui Zhou, and Yonghong Tian. Spikingformer: Spike-driven residual learning for transformer-based spiking neural network. arXiv preprint arXiv:2304.11954, 2023. 1, 3, 5, 7, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 592, + 295, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 592, + 295, + 635 + ], + "spans": [ + { + "bbox": [ + 57, + 592, + 295, + 635 + ], + "type": "text", + "content": "[64] Zhaokun Zhou, Yuesheng Zhu, Chao He, Yaowei Wang, Shuicheng Yan, Yonghong Tian, and Li Yuan. Spikformer: When spiking neural network meets transformer. arXiv preprint arXiv:2209.15425, 2022. 1, 3, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 637, + 295, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 637, + 295, + 680 + ], + "spans": [ + { + "bbox": [ + 57, + 637, + 295, + 680 + ], + "type": "text", + "content": "[65] Zhaokun Zhou, Kaiwei Che, Wei Fang, Keyu Tian, Yuesheng Zhu, Shuicheng Yan, Yonghong Tian, and Li Yuan. Spikformer v2: Join the high accuracy club on imagenet with an snn ticket. arXiv preprint arXiv:2401.02020, 2024. 1, 3, 4, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 681, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 681, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 681, + 295, + 713 + ], + "type": "text", + "content": "[66] Shihao Zou, Yuxuan Mu, Xinxin Zuo, Sen Wang, and Li Cheng. Event-based human pose tracking by spiking spatiotemporal transformer, 2023. 1" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "24408" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/SpiritSight Agent_ Advanced GUI Agent with One Look/c38ba9da-d4d0-4e2a-a186-fdd0cc0687b1_content_list.json b/2025/SpiritSight Agent_ Advanced GUI Agent with One Look/c38ba9da-d4d0-4e2a-a186-fdd0cc0687b1_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..4841a334156b6ba21d454c92ba52698585c53f9a --- /dev/null +++ b/2025/SpiritSight Agent_ Advanced GUI Agent with One Look/c38ba9da-d4d0-4e2a-a186-fdd0cc0687b1_content_list.json @@ -0,0 +1,1775 @@ +[ + { + "type": "image", + "img_path": "images/f6471b3df2f89a2cd7883fa6ed05731ecd52be93cc5193613bdebf5724d4fd70.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 200, + 126, + 233, + 147 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "SpiritSight Agent: Advanced GUI Agent with One Look", + "text_level": 1, + "bbox": [ + 232, + 130, + 797, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zhiyuan Huang $^{1}$ Ziming Cheng $^{1,2*}$ Junting Pan $^{3}$ Zhaohui Hou $^{1}$ Mingjie Zhan $^{1}$", + "bbox": [ + 122, + 186, + 872, + 207 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ SenseTime Research, $^{2}$ Beijing University of Posts and Telecommunications, $^{3}$ MMLab, CUHK", + "bbox": [ + 124, + 210, + 874, + 229 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{huangzhiyuan, houzhaohui, zhanmingjie}@sensetime.com", + "bbox": [ + 263, + 250, + 733, + 265 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "harrytheorange@bupt.edu.cn junting.pa@gmail.com", + "bbox": [ + 271, + 271, + 718, + 285 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 321, + 326, + 338 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Graphical User Interface (GUI) agents demonstrate promising potential in assisting human-computer interaction, automating human user's navigation on digital devices. An ideal GUI agent is expected to achieve high accuracy, low latency, and compatibility for different GUI platforms. Recent vision-based approaches have shown promise by leveraging advanced Vision Language Models (VLMs). While they generally meet the requirements of compatibility and low latency, these vision-based GUI agents tend to have low accuracy due to their limitations in element grounding. To address this issue, we propose SpiritSight, a vision-based, end-to-end GUI agent that excels in GUI navigation tasks across various GUI platforms. First, we create a multi-level, large-scale, high-quality GUI dataset called GUI-Lasagne using scalable methods, empowering SpiritSight with robust GUI understanding and grounding capabilities. Second, we introduce the Universal Block Parsing (UBP) method to resolve the ambiguity problem inherited from the dynamic resolution strategy, further enhancing SpiritSight's ability to ground GUI objects. Through these efforts, SpiritSight agent outperforms other advanced methods on diverse GUI benchmarks, demonstrating its superior capability and compatibility in GUI navigation tasks. The models and code will be made available upon publication.", + "bbox": [ + 89, + 353, + 483, + 731 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 760, + 220, + 776 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Graphical User Interface (GUI) automation has long been pursued by people along with the development of modern digital devices. Thanks to recent advances in Large Language Models (LLMs), GUI agents are constructed to assist users in interacting with graphical interfaces, automatically making action decisions based on observations of environ-", + "bbox": [ + 89, + 785, + 483, + 876 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b75cfff39780575a3f479d6e58c42d4b8a4ff6db0da1f119a00fe244d844c1dc.jpg", + "image_caption": [ + "Figure 1. Our SpiritSight agent achieves new state-of-the-art (SOTA) performance across various benchmarks in web, mobile, and desktop scenarios." + ], + "image_footnote": [], + "bbox": [ + 529, + 321, + 890, + 551 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ment and user's instruction.", + "bbox": [ + 511, + 635, + 697, + 648 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Current approaches can be divided into three categories based on their input modalities. Language-based and vision-based approaches use Hyper Text Markup Language (HTML) or Extensible Markup Language (XML) and screenshots as input [7, 11, 13, 19, 22, 26, 28, 53, 68, 75], respectively, while vision-language-based approaches combine both HTML and screenshots as input [15, 24, 52, 74].", + "bbox": [ + 511, + 652, + 906, + 758 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The language-based and vision-language-based methods typically applied only in the web domain, and are often limited by the excessive length of HTML or security concerns [34, 61, 71]. The vision-based approaches exhibit compatibility across various GUI platforms, as acquiring screenshots is generally easier than obtaining hierarchical data from non-web platforms.", + "bbox": [ + 511, + 761, + 908, + 867 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "However, vision-based approaches struggle to ground GUI elements (e.g. buttons, text boxes) from the visual", + "bbox": [ + 511, + 869, + 908, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution.", + "bbox": [ + 109, + 887, + 218, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "29490", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "input [74]. Some works solve this problem by introducing additional tools, such as optical character recognition (OCR) and icon recognition models, to convert the visual grounding task into a language QA task [55, 56]. This may increase the complexity and inference latency of the agent system. Others attempt to collect large-scale training data through manual synthesis [27, 36, 51] or human annotation [5, 6, 13, 38, 39, 46, 65], while these data are either unrealistic or costly. Additionally, recent advanced Vision Language Models (VLMs) [9, 67] adapt dynamic high-resolution [9, 67] strategy to better suit high-resolution inputs, which we find may introduce ambiguity to the model learning process in GUI scenarios.", + "bbox": [ + 89, + 90, + 483, + 287 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address the aforementioned challenges, we propose an end-to-end, vision-based GUI agent—SpiritSight, which has strong ability for GUI navigation task. Our contributions are summarized as follows:", + "bbox": [ + 89, + 287, + 483, + 348 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Firstly, we propose GUI-Lasagne, a multi-level, large-scale and high-quality GUI dataset composed of 5.73 million samples to enhance our model's GUI understanding and grounding capabilities. The dataset is collected from the real world and filtered through carefully designed rules to ensure data quality. It is also constructed hierarchically and consists of three levels of components. The first two parts constitute $90\\%$ of the total dataset and are collected for free, thus significantly reducing the data collection cost.", + "bbox": [ + 89, + 349, + 483, + 500 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Secondly, we introduce a Universal Block Parsing (UBP) method to resolve the ambiguity problem inherited from the dynamic resolution strategy. With the effectiveness of UBP, SpiritSight gains an improved elements grounding capability and achieves enhanced performance in GUI navigation tasks.", + "bbox": [ + 89, + 501, + 483, + 590 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Thirdly, we release and evaluate SpiritSight agent in various GUI benchmarks and it exhibits remarkable performance among them. SpiritSight agent is pre-trained on GUI-Lasagne and employs UBP to ground elements, as shown in Fig. 3. We release three versions of SpiritSight with different model size: SpiritSight-26B, SpiritSight-8B, and SpiritSight-2B. The standard SpiritSight-8B consistently outperforms previous state-of-the-art methods across various GUI benchmarks, as shown in Fig. 1. On the Multimodal-Mind2Web [13] benchmark, the SpiritSight series outperforms all kinds of methods—including language-based, vision-based and even vision-language-based methods—in the non-candidate setting, as shown in Fig. 2.", + "bbox": [ + 89, + 592, + 483, + 789 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 89, + 805, + 228, + 821 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Language-based and vision-language-based GUI agent", + "text_level": 1, + "bbox": [ + 89, + 830, + 482, + 862 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Several works leverage the capabilities of large-scale language models (LLMs) to construct GUI agents. It is", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/1c61129ecf89b1b09352b18720a18fa0f35f80a7644c09b4ab3d4bf137012067.jpg", + "image_caption": [ + "Figure 2. Comparison of the average step success rate on Multimodal-Mind2Web benchmark of our SpiritSight agent of three sizes (2B, 8B, 26B) with various previous methods." + ], + "image_footnote": [], + "bbox": [ + 531, + 90, + 879, + 282 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/702443052d2c731f186e652a64491d001cbb7d776c03265819be8e16b6a5d74d.jpg", + "image_caption": [ + "Figure 3. The overview of our SpiritSight agent. Our proposed GUI-Lasagne dataset equips SpiritSight with three levels of comprehensive GUI knowledge. Additionally, we introduce a Universal Block Parsing (UBP) method to enhance SpiritSight's grounding capabilities." + ], + "image_footnote": [], + "bbox": [ + 527, + 354, + 895, + 571 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "noticed that they are mostly multi-stage architectures. Mind2Web [13] utilizes a lightweight language model to extract candidate elements from HTML, followed by a ranking model that sorts these elements based on task descriptions and historical actions. Finally, a large language model predicts the current actions and selects the target elements to operate on. WebAgent [17] first uses an encoder-decoder model to generate low-level instructions and relevant HTML code snippets, then uses another decoder to produce executable Python code. AutoWebGLM [26] simplifies HTML code through manually designed rules before predicting the executable action codes.", + "bbox": [ + 511, + 688, + 906, + 869 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Other vision-language-based works leverage both GUI screenshots and hierarchical HTML/XML to enhance the", + "bbox": [ + 511, + 869, + 906, + 898 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "29491", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "table", + "img_path": "images/cf5c1cffe79d20257a05972072800f189ee3c2eb099bb88c2900d108980b3648.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Model SizeInput ModalitySelect From TopCross-TaskCross-WebsiteCross-Domain
Ele.AccOp.F1Step SREle.AccOp.F1Step SREle.AccOp.F1Step SR
AutoWebGLM [26]6BText✓--66.4%--56.4%--55.8%
LLaMA2-7B [26]7BText✓--52.7%--47.1%--50.3%
CogAgent [19]18BImage✓--62.3%--54.0%--59.4%
HTML-T5-XL [17]3BText✓76.4%78.8%71.5%68.4%71.0%62.2%73.0%75.6%67.1%
SeeAct* [74]-Text+Image×46.4%73.4%40.2%38.0%67.8%32.4%42.4%69.3%36.8%
ReadAgent-P* [28]340BText×33.7%72.5%29.2%37.4%75.1%31.1%37.2%76.3%33.4%
MiniCPM-GUI [7]3BImage×23.8%86.8%20.8%20.3%81.7%17.3%17.9%74.5%14.6%
Fuyu-GUI [4]8BImage×19.1%86.1%15.6%13.9%80.7%12.2%14.2%83.1%11.7%
SeeClick [11]9.6BImage×28.3%87.0%25.5%21.4%80.6%16.4%23.2%84.8%20.8%
OmniParser* [53]-Image×42.4%87.6%39.4%41.0%84.8%36.5%45.5%85.7%42.0%
SpiritSight-2B2BImage×51.7%87.2%44.9%44.0%83.6%37.8%42.4%83.5%36.9%
SpiritSight-8B8BImage×59.2%88.9%52.7%52.2%84.7%44.0%50.1%86.0%44.4%
SpiritSight-26B26BImage×60.5%89.7%54.7%57.0%85.7%48.1%54.1%87.2%49.2%
", + "bbox": [ + 93, + 87, + 903, + 300 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Table 1. Comparison of SpiritSight agent with various previous methods on Multimodal-Mind2Web benchmark. SpiritSight significantly outperforms all methods that do not rely on any candidate element, including vision-based methods, language-based methods, and even vision-language-based methods. The asterisk (*) in the upper-right indicates that the model is not fine-tuned on the downstream tasks.", + "bbox": [ + 89, + 323, + 906, + 367 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "robustness of GUI agents. WebGUM [15] and CC-Net [52] use ResNet and ViT to extract features from screenshots, respectively. The image embeddings are then combined with text embeddings and fed into a multi-modal transformer. SeeAct [74] and AppAgent [64] identify all interactive elements using HTML or XML data. They then assign each interactive element a unique identifier in the screenshot before feeding the screenshot into the model.", + "bbox": [ + 88, + 391, + 480, + 513 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "These language-based and vision-language-based methods, which rely on hierarchical information, exhibit several limitations: (1) Hierarchical representations, such as HTML or XML, are not consistently available across different platforms. And even this information is available, differences in their internal rules make language-based GUI agents less compatible; (2) HTML often contains redundant information, requiring additional models or extensive manually crafted rules for effective filtering. (3) Text-based GUI agents are susceptible to injection attacks [34, 61, 71], where malicious instructions hidden in HTML can easily lead to erroneous or unsafe actions.", + "bbox": [ + 89, + 513, + 482, + 695 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Vision-based GUI agent", + "text_level": 1, + "bbox": [ + 89, + 709, + 310, + 726 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Recently, some vision-based approaches have been proposed to overcome the drawbacks of language-based methods. Some of them [2, 11, 19, 50] are single-stage methods that only use GUI screenshots as input for VLMs and output the next action in an end-to-end manner. However, these agents perform worse on GUI benchmarks compared to other methods. Two primary reasons may be the relatively low input resolution of their base models and the insufficient scale and suboptimal quality of the training data. SeeAct [74], MobileAgent [56] and MobileAgent-v2 [55] are two-stage methods. They use GPT-4V instead of open", + "bbox": [ + 89, + 734, + 482, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "source VLMs as base models and find that even the top models struggle with element grounding. Consequently, they introduce additional OCR and icon recognition models to assist with element grounding, which may increase the complexity and inference latency of the agent system. Overall, the current VLMs exhibit limited ability in handling GUI grounding tasks, thus constrains the navigation capabilities of single-stage vision-based GUI agents. Related works on base models are discussed in Sec. 9.", + "bbox": [ + 511, + 392, + 906, + 527 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Data collection", + "text_level": 1, + "bbox": [ + 511, + 544, + 663, + 559 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we present a cost-effective data collection strategy designed to construct a multi-level, large-scale and high-quality GUI dataset, called GUI-Lasagne, as shown in Fig. 4. This dataset helps equip our models with robust abilities in GUI understanding, grounding, and navigation. More details of the collection can be seen in Tab. 8.", + "bbox": [ + 511, + 571, + 906, + 661 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Level one: visual-text alignment", + "text_level": 1, + "bbox": [ + 511, + 672, + 797, + 690 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Visual-text alignment refers to the foundational ability of a GUI agent to perceive the visual appearance of the GUI, specifically the ability to recognize or locate text/icon elements. We collect visual-text data directly from the GUI source data. In the web scenario, we obtain web URLs from CommonCrawl [16] and website ranking sources12. We then develop a data collection tool using Playwright3 to gather real-world web data from the collected URLs. With this tool, we collect 755K web-page screenshots alone with their DOM data, featuring diversity in both resolution", + "bbox": [ + 511, + 698, + 906, + 849 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "1 https://www.semrush.com/trending-sites/global/all", + "bbox": [ + 531, + 862, + 826, + 875 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "2https://top.chinaz.com/all/", + "bbox": [ + 532, + 875, + 679, + 887 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "3https://playwright.dev/", + "bbox": [ + 532, + 887, + 661, + 898 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "29492", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/0927b7c078b46c951d58911de6fd71473ddd4b354468d91ba93747d2cced5ce5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 91, + 88, + 331, + 224 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/2564866784a69ce115cdfd5f7a88f4490320c4b703aaf837dfd8169147627e27.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 341, + 88, + 519, + 224 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/774c4ba630de293c2b7f03ca4b633358f6d9101666ec5a6833fd574cd95b8a47.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 527, + 88, + 700, + 224 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/bae001d24c286e962856c041e62532013e97a9e312903b28958e301856b9f47a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 707, + 88, + 898, + 224 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/87b3fde8e17abff1c0705f8b1dda6c3b9ca815b7e2d1b2ab380d1840b7fd35f7.jpg", + "image_caption": [ + "Figure 4. The collection pipeline of our GUILasagne dataset. The left, middle and right parts show the construction of level-1, level-2, and level-3 data, respectively." + ], + "image_footnote": [], + "bbox": [ + 91, + 232, + 338, + 328 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/0b0aa35a26767ccecb81fe588da766a58912b3e01bb0615d80aae061d27feb8d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 341, + 232, + 517, + 328 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/e545eec66bfd52478f2e4021d4e397d473917c19ef7317eb5798852a8a267c77.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 527, + 232, + 696, + 328 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/df77e095b408b7bd199e9f9e8e2f178bbca0cffc95922f945c97cbc56c1963df.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 709, + 232, + 898, + 328 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "and language. Additionally, we develop an icon recognition model called InternVL-Icon as an icon annotation tool, by constructing a dataset of 30K icon-caption pairs from Alibaba Icon Library $^4$ and fine-tuning InternVL1.5-26B using this dataset. We then use InternVL-Icon to annotate all icons on web-pages with generated captions. As for the mobile scenario, we collect data from AitW [46], a large-scale GUI dataset on the mobile platform.", + "bbox": [ + 88, + 398, + 482, + 518 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Inspired by Lee et al. [27] and Cheng et al. [11], we construct three tasks based the collected GUI source data: text2bbox,_bbox2text, and_bbox2dom. The text2bbox task prompts the model to ground the element based on the given text or icon caption. To avoid ambiguity, we provide additional contextual information for elements that appear multiple times in the screenshots. The text2bbox data is the most abundant among the three tasks, to help the model develop robust grounding capabilities. The_bbox2text task is the inverse of the text2bbox task, training the model to recognize text and icons. The_bbox2dom task asks the model to generate a DOM-tree corresponding to the given bounding box, as show in Sec. 12.7. This task helps the model learn not only to recognize text/icon elements but also to understand the GUI layout. To make sufficient use of the context length of the model, we pack multiple data pairs in each training sample for text2bbox and_bbox2text task, and select the box that includes as many elements as possible for_bbox2dom task. Overall, we construct a total of 1.9M and 1.1M training samples for the web and mobile scenarios, respectively. The data significantly enhance the GUI foundational abilities, especially the GUI grounding ability, of our SpiritSight agent. See Sec. 14 for more details on", + "bbox": [ + 89, + 521, + 483, + 869 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "sample construction.", + "bbox": [ + 513, + 398, + 653, + 412 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Level two: vision-function alignment", + "text_level": 1, + "bbox": [ + 511, + 422, + 831, + 438 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Vision-function alignment refers to a model's ability to locate an element based on its function. This type of data is not directly accessible from the raw GUI data. Inspired by the back-translation method [48] for dataset collection in translation tasks, we leverage InternVL's image understanding capabilities to collect function grounding data. Specifically, we divide each screenshot into a 3x3 grid and describe the approximate location of the target element in text (e.g. in the top-left corner of the image). Additionally, we place a bounding box around the target element in the screenshot to specify its precise location. We provide InternVL2-26B with the screenshot, the element's text content or icon caption, and the location description to prompt it to generate the element's function. Additionally, we utilize InternLM2.5-20B to enhance the quality and diversity of the generated function descriptions. These function descriptions achieve an approximate $90.9\\%$ acceptance rate based on human judgment, which we consider sufficient for constructing the function grounding pre-training data.", + "bbox": [ + 511, + 445, + 906, + 733 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Based on the strategy above, we collect function2bbox pairs for all interactive elements in the web scenario. In the mobile scenario, we construct the function grounding data from the GUI navigation dataset, which is be described in Sec. 3.3. We use the same packing method as in the text2bbox and_bbox2text tasks for efficient training, ultimately obtaining 1.5M training samples.", + "bbox": [ + 511, + 733, + 906, + 839 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Level three: visual GUI navigation", + "text_level": 1, + "bbox": [ + 511, + 847, + 815, + 863 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We utilize the public available AitW [46] dataset to construct our GUI navigation training data. AitW is a large-", + "bbox": [ + 511, + 869, + 906, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "4https://www.iconfont.cn/?v=20230914", + "bbox": [ + 107, + 886, + 320, + 898 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "29493", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "scale mobile navigation dataset where each screenshot is labeled with the corresponding goal, the current step, etc. However, as noted in AitZ [72] and AMEX [5], the AitW dataset contains a number of incorrectly labeled samples. We choose to clean it with GPT-4o and employ Chain-of-Thought (CoT) [60] to make the judgment more accurate. Specifically, we prompt GPT-4o with the task description, the current action annotation, and screenshots from both the current and next steps. GPT-4o is then instructed to first summarize the two screenshots and identify the differences between them, then describe the current step based on these differences, and finally assess reasonability of the current action annotation.", + "bbox": [ + 89, + 90, + 483, + 286 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We filter out steps identified as unreasonable by GPT-4o, resulting a final dataset of 0.64M CoT-style GUI navigation training samples. With the collected CoT-style data, we are able to construct additional function grounding data for the mobile scenario, as mentioned in Sec. 3.2, by treating the collected step descriptions as functions of the corresponding elements.", + "bbox": [ + 89, + 287, + 483, + 393 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Other training data", + "text_level": 1, + "bbox": [ + 89, + 407, + 277, + 424 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To enhance the model's understanding of GUI content, we further collected some public datasets as a supplement, including doc/web/mobile VQA [7, 8, 20, 40], image captioning [12, 54], and mobile grounding datasets [12, 33], resulting in 0.59M QA pairs for model training.", + "bbox": [ + 89, + 431, + 483, + 508 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Universal block parsing", + "text_level": 1, + "bbox": [ + 89, + 525, + 313, + 542 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Problem statement", + "text_level": 1, + "bbox": [ + 89, + 551, + 272, + 566 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We use InternVL2 [9] as the base model of SpiritSight, which uses a dynamic high-resolution training approach to largely preserves the details of input images. The approach first matches the optimal aspect ratio from a pre-defined set of aspect ratios to each input image, then resize and divide each image into blocks of $448 \\times 448$ pixels.", + "bbox": [ + 89, + 575, + 483, + 667 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "However, this approach may introduce ambiguity in grounding GUI elements. A typical example is represented in Fig. 5a, with two screenshots having aspect ratios of 1:2 and 2:1, respectively. According to the dynamic high-resolution strategy, the screenshots are divided into two blocks, one vertically and the other horizontally. Suppose the target elements in each sample are located in the same relative position within the second block (block-1) after cropping. The cropped screenshots are then fed into InternVL and the model is expected to predict different locations from these two identical inputs. We refer to this situation as the location ambiguity problem.", + "bbox": [ + 89, + 669, + 483, + 849 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Generally, a point $\\mathbf{p}$ is expressed in the global coordinate system as", + "bbox": [ + 89, + 851, + 482, + 881 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {p} = [ x, y ] \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 250, + 883, + 482, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $x$ and $y$ represent the horizontal and vertical coordinate values of the point in the original image, respectively. Assume that the image is resized and divided into $n_w \\times n_h$ blocks. We can express the point $\\mathbf{p}$ in the corresponding block as $\\mathbf{p}'$ :", + "bbox": [ + 511, + 90, + 905, + 167 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{ \\begin{array}{l} b _ {x} = \\left\\lfloor \\frac {x}{w _ {\\text {b l o c k}}} \\right\\rfloor \\\\ b _ {y} = \\left\\lfloor \\frac {y}{h _ {\\text {b l o c k}}} \\right\\rfloor \\end{array} \\quad \\left\\{ \\begin{array}{l} x ^ {\\prime} = x \\bmod w _ {\\text {b l o c k}} \\\\ y ^ {\\prime} = y \\bmod h _ {\\text {b l o c k}} \\end{array} \\right. \\right. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 568, + 175, + 905, + 217 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {p} ^ {\\prime} = \\left[ b _ {x}, b _ {y}, x ^ {\\prime}, y ^ {\\prime} \\right] \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 643, + 224, + 905, + 243 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $w_{block}$ and $h_{block}$ (both set to 448 in our experiments) represent the width and height of each block, respectively. $b_{x}$ and $b_{y}$ represent the horizontal and vertical indices of the block containing $\\mathbf{p}'$ , and $x'$ and $y'$ represent the coordinates of $\\mathbf{p}'$ within this block. The blocks are flattened into a sequence before being fed into the model, thus we have", + "bbox": [ + 511, + 250, + 906, + 339 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {p} ^ {\\prime \\prime} = \\left[ b _ {i}, x ^ {\\prime}, y ^ {\\prime} \\right] \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 653, + 349, + 905, + 367 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $b_{i}$ represents the index of the block within the sequence, satisfying:", + "bbox": [ + 511, + 378, + 905, + 407 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nb _ {i} = b _ {y} \\cdot n _ {w} + b _ {x} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 648, + 420, + 903, + 436 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The model is trained to approximate a mapping $f: \\mathbf{p}'' \\to \\mathbf{p}$ , which is inherently a multi-valued function. For example, when the input $\\mathbf{p}''$ is [1, 168, 245], possible values for $p$ include [168, 693] when $n_w = 1$ , or [616, 245] when $n_w = 2$ , as shown in Fig. 5a.", + "bbox": [ + 511, + 446, + 905, + 523 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Method", + "text_level": 1, + "bbox": [ + 513, + 531, + 609, + 545 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The ambiguity primarily arises from the flattening operation, as shown in Eq. (5), which results in the loss of the spatial relationship between blocks. One solution is to input an additional thumbnail, but this may lead to extra computational and memory overhead. We propose to solve this positional ambiguity with two steps. Firstly, we introduce a 2D Block-wise Position Embedding (2D-BPE) [67] by adding two position embeddings to each block to capture spatial information. Secondly, we introduce a Universal Block Parsing (UBP) method, which replaces the global coordinate representation with a block-specific coordinate representation. Specifically, we express a point as Eq. (4). In this case, the model is trained to approximate a injective mapping $f: \\mathbf{p}'' \\rightarrow \\mathbf{p}''$ , thereby resolving the ambiguity problem. During the model inference, the global coordinate of this point can be computed in post-processing as follows:", + "bbox": [ + 511, + 554, + 906, + 796 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{ \\begin{array}{l} x = x ^ {\\prime} + \\left(b _ {i} \\mod n _ {w}\\right) \\cdot w _ {\\text {b l o c k}} \\\\ y = y ^ {\\prime} + \\left\\lfloor \\frac {b _ {i}}{n _ {w}} \\right\\rfloor \\cdot h _ {\\text {b l o c k}} \\end{array} \\right. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 584, + 806, + 903, + 848 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Overall, our UBP method ensures a clear mapping of positional information between the model's inputs and outputs, thereby improving the model's grounding capability.", + "bbox": [ + 511, + 854, + 906, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "29494", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/6cbf1f56b2c8862f6675f16b75ec849427e015e45df37a8cbb3645b5929d639e.jpg", + "image_caption": [ + "(a) Baseline Block Parsing vs Universal Block Parsing." + ], + "image_footnote": [], + "bbox": [ + 96, + 88, + 584, + 248 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/01113d42e69228e6e352f1806bb26775fd8ff3a752d69a993076a66ed6e57181.jpg", + "image_caption": [ + "(b) Performance gain with UBP.", + "Figure 5. (a) Comparison between baseline block parsing and our proposed UBP. (b) The results of baseline block parsing and our proposed UBP methods on Multimodal-Mind2Web benchmark. UBP improves the performance of our model. The combination of UBP and 2D-BPE achieves the best results." + ], + "image_footnote": [], + "bbox": [ + 616, + 89, + 898, + 250 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Settings", + "text_level": 1, + "bbox": [ + 89, + 342, + 181, + 359 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Implementation details", + "text_level": 1, + "bbox": [ + 89, + 378, + 303, + 395 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We use InternVL2 (2B, 8B and 26B) [9] as the base models and train them with two stages: continual pre-training and fine-tuning. During the pre-training stage, we use all the GUI-Lasagne dataset mentioned in Sec. 3. Different prompts are designed for different training tasks to avoid task confusion. See Sec. 14 for the prompts. We unfreeze the vision encoder, decoder, and MLP layer of InternVL. The learning rate is set to $1\\mathrm{e} - 4 / 1\\mathrm{e} - 4 / 5\\mathrm{e} - 5$ for InternVL-2B/8B/26B, respectively, and the batch size is set to 1024. We get SpiritSight-Base model after pre-training and then fine-tune it on multiple downstream tasks individually. More details are shown in Sec. 12.2 and Sec. 12.3.", + "bbox": [ + 89, + 412, + 483, + 595 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Benchmarks & metrics", + "text_level": 1, + "bbox": [ + 89, + 638, + 305, + 652 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To assess SpiritSight's capability in diverse real-world environments, we evaluate SpiritSight on six benchmarks covering various GUI platforms and tasks. Multimodal-Mind2Web [13], ScreenSpot [11], AMEX [5], GUIAct [7], AndroidControl [32], and GUI-Odyssey [38]. For Multimodal-Mind2Web and ScreenSpot, we use the same data pre-processing and metrics as SeeClick [11] uses. For GUIAct, we evaluate SpiritSight on the web-single and web-multi sub-sets and report step success rate (Step SR). For AndroidControl, we evaluate SpiritSight on the High-Level (HL) and Low-Level (LL) tasks and report step accuracy. For AMEX and GUI-Odyssey, we report the action matching score (AMS) defined in AitW [46]. These metrics are similar in that they all measure the single-step accuracy. Refer to Sec. 12.4 for more information about the metrics.", + "bbox": [ + 89, + 672, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6. Experiment", + "text_level": 1, + "bbox": [ + 513, + 342, + 637, + 359 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6.1. Advanced vision-based GUI agent", + "text_level": 1, + "bbox": [ + 511, + 368, + 810, + 385 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we compare the performance of SpiritSight with other advanced methods across various input modalities and test configurations on Multimodal-Mind2Web [13], a classic and high-quality benchmark for GUI navigation. The results are shown in Tab. 1. Methods that use top-50 candidates as input perform the best. This is evident, as the assistance of candidate elements can significantly reduce the decision space. However, such methods are not particularly feasible in practice.", + "bbox": [ + 511, + 392, + 906, + 527 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The results indicate that SpiritSight significantly outperforms all methods that do not rely on any candidate element, including vision-based methods, language-based methods, and even vision-language-based methods. This demonstrates strong capabilities of SpiritSight in Web GUI navigation tasks. It should be noted that SpiritSight achieves a significant advantage in the Ele.Acc metric compared to other vision-based methods, which can be attributed to the specially constructed visual grounding training data and the proposed UBP approach.", + "bbox": [ + 511, + 529, + 908, + 681 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6.2. Strong cross-platform compatibility", + "text_level": 1, + "bbox": [ + 511, + 694, + 823, + 710 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We evaluated SpiritSight on other GUI navigation benchmarks across various platforms and compare it with state-of-the-art (SOTA) GUI agents as shown in Tab. 2. SpiritSight demonstrated the best performance on these benchmarks, showing its strong capabilities across various platforms. OdysseyAgent uses additional historical screenshot images as input, yet SpiritSight achieves comparable results.", + "bbox": [ + 511, + 717, + 905, + 837 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We also evaluated SpiritSight on ScreenSpot, a function grounding benchmark. As shown in Tab. 3, SpiritSight performs well across all three platforms, showcasing its cross-platform capability and robust grounding ability. Notably,", + "bbox": [ + 511, + 839, + 905, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "29495", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/b21c1a4c11910f953d1c28d865588f2e33acd71851290ef8fed6c506a1044270.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
AgentOdysseyAMEXAndroidCtrlGUIAct
HighHighHighLowMultiSingle
GPT-4o[42]20.4%-21.2%28.4%-41.8%
Previous SOTA74.3%70.7%64.8%80.0%45.4%74.9%
SpiritSight-2B72.3%74.5%64.9%86.3%45.5%76.0%
SpiritSight-8B75.8%80.7%68.1%87.6%49.3%78.2%
SpiritSight-26B75.9%82.1%69.6%88.2%52.6%78.3%
", + "bbox": [ + 91, + 89, + 482, + 203 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "SpiritSight models are trained specifically for GUI navigation tasks, and no training data aligned with ScreenSpot is used. There remains room for further improvement in SpiritSight's performance on the ScreenSpot benchmark. We further assess the foundational ability of SpiritSight, the visual grounding ability, on our custom-constructed testing dataset. See Sec. 12.7 for more details.", + "bbox": [ + 89, + 337, + 482, + 444 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6.3. Recognition and grounding as priors for GUI navigation", + "text_level": 1, + "bbox": [ + 89, + 453, + 482, + 486 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To verify the significance of the three levels of GUI-Lasagne data, we progressively removed level-3, level-2, and level-1 data from the training set during the pretraining stage and evaluate SpiritSight-8B on MultimodalMind2Web. The results are shown by the blue line in Fig. 6a. It can be seen that each level of data contributes to improving Step SR. While the tasks in level-1 data differ the most from web navigation task compared to the other two levels, they provide an effective initialization for the pre-trained model. Although the level-3 data is constructed from the mobile scenario, it also assists in web-based GUI navigation tasks. This indicates that the joint learning strategy helps SpiritSight develop strong navigation abilities across different GUI environments with limited resources.", + "bbox": [ + 89, + 492, + 482, + 702 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We also conducted ablation experiments to evaluate the effectiveness of data cleaning and CoT construction strategies on the level-3 data, as shown by the orange line in Fig. 6a. It can be observed that training in a CoT manner effectively improves the model's GUI navigation capabilities, while the impact of data cleaning strategy is relatively small. This may be due to the fact that erroneous samples are relatively tolerated during the pre-training phase of large models. The difference in results of \"w/ CoT\" and \"w/ Level $1 + 2 + 3$ \" is due to the different implementation for fine-tuning stage. See Sec. 12.3 for more details.", + "bbox": [ + 89, + 704, + 482, + 869 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To evaluate the effectiveness of our GUI-Lasagne dataset, we train InternVL-2 (8B) on SeeClick train", + "bbox": [ + 89, + 869, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/bca23e087cb4a97f93000c8cb832a340996d071fd12aab186b51f873a4b5138f.jpg", + "table_caption": [ + "Table 2. Results of SpiritSight on four recently published GUI navigation benchmarks. 'High' and 'Multi' indicates high-level tasks, 'Low' and 'Single' indicates single-stepped tasks. The SOTA results of GUI-Odyssey, AMEX, AndroidControl and GUIAct come from OdysseyAgent [38], SPHINX-GUI-Agent [5], fine-tuned PaLM-2S [32] and MiniCPM-GUI [7], respectively." + ], + "table_footnote": [], + "table_body": "
AgentModel SizeScreenSpot
WebMobileDesktop
GPT4V[1]-5.0%7.5%4.6%
Qwen-VL[3]9.6B3.0%7.2%5.4%
Fuyu[4]8B19.2%21.2%18.3%
CogAgent[19]18B49.5%45.5%47.1%
SeeClick[11]9.6B44.1%65.0%51.1%
SpiritSight-2B2B63.6%62.5%61.8%
SpiritSight-8B8B68.3%68.4%62.9%
SpiritSight-26B26B74.4%73.6%69.9%
", + "bbox": [ + 527, + 88, + 893, + 250 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Results of SpiritSight and other vision-based methods on ScreenSpot Benchmark.", + "bbox": [ + 511, + 271, + 903, + 299 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ing data [11] using full parameter training and LoRA training, resulting in SeeClick(InternVL-Full) and SeeClick(InternVL-LoRA), respectively. We then fine-tune and evaluate the two models on the Multimodal-Mind2Web benchmark. The results are shown by the dashed line in Fig. 6a, where we also present the result of the original SeeClick model. SeeClick(InternVL-LoRA) performs better than the original SeeClick model, which may be attributed to the dynamic high-resolution approach used by InternVL that preserves the details of input images. SeeClick(InternVL-Full) performs the worst, as the scale of the SeeClick training data is not large enough for the model to converge during full parameter training. SpiritSight agent that trained on only GUI-Lasagne level-1 data outperforms all three models trained on SeeClick data, demonstrating the effectiveness of GUI-Lasagne dataset. More ablation studies can be found in Sec. 12.5.", + "bbox": [ + 511, + 323, + 906, + 580 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6.4. Better grounding ability from UBP", + "text_level": 1, + "bbox": [ + 511, + 589, + 818, + 604 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To verify the effectiveness of UBP on grounding task, we use LoRA for resource efficiency to fine-tune InternVL in 4 different settings (original as baseline, 2D-BPE, UBP, and 2D-BPE&UBP), respectively. We then evaluate these models on Multimodal-Mind2Web benchmark, as shown in Fig. 5b. It can be seen that UBP shows a significant advantage in Ele.Acc compared to baseline, while the results of Op.F1 show little variation across the 4 settings. This indicates that UBP improves the performance of GUI agent primarily by enhancing the grounding ability. Finally, the combination of UBP and 2D-BPE achieves the best results. This indicates that UBP is compatible with 2D-BPE, leading to better performance. More ablation studies can be found in Sec. 12.6.", + "bbox": [ + 511, + 612, + 906, + 823 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6.5. Scaling effects on dataset and model size", + "text_level": 1, + "bbox": [ + 511, + 832, + 859, + 848 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We explore the impact of pre-training dataset and model size on SpiritSight using Multimodal-Mind2Web benchmark. The results are shown in Fig. 6b. SpiritSight-2B out", + "bbox": [ + 511, + 854, + 903, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "29496", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/68cb0cf0cbde3674e8ff4c66a96a4808e3d7365dcca842f4fc647ed1a9af8a2f.jpg", + "image_caption": [ + "(a) Effect of three data levels and data augmentation for level-3 data", + "Figure 6. Ablation studies on GUILasagne dataset and scaling effects in the Multimodal-Mind2Web benchmark." + ], + "image_footnote": [], + "bbox": [ + 94, + 89, + 336, + 253 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/75a9adfdde148c1f0f53837684c4722adfd282f03f634c1ff30599f5b936d3a5.jpg", + "image_caption": [ + "(b) Ablation Study: Effect of training data percentages on model performance." + ], + "image_footnote": [], + "bbox": [ + 377, + 92, + 620, + 252 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/94a23edebfe12bbac891c5fae7d6da1e2d962594bc8f6112469c5f217008e905.jpg", + "image_caption": [ + "(c) Ablation Study: Effect of data percentages on downstream task." + ], + "image_footnote": [], + "bbox": [ + 656, + 92, + 901, + 252 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "performs SeeClick [11] by using just 1/8 of the pre-training dataset. This remarkable performance comes from the high quality and grounding-focus nature of GUI-Lasagne data. The performance improves as the size of dataset increases, demonstrating the significance of collecting large-scale data. SpiritSight-2B reaches saturation with a smaller amount of pre-training data, while SpiritSight-26B appears to have further potential for improvement, which aligns with the scaling law of LLMs and VLMs.", + "bbox": [ + 88, + 330, + 480, + 465 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We also evaluate the ability of SpiritSight to transferring to downstream GUI agent tasks. We fine-tune SpiritSight-Base (8B) on various proportions of the Multimodal-Mind2Web training data and show the results in Fig. 6c. It is noticed that SpiritSight achieves $36.6\\%$ Step SR with only 1/8 of the fine-tuning data, showing strong transferability to GUI navigation tasks.", + "bbox": [ + 89, + 468, + 482, + 574 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6.6. Effective transfer to other languages", + "text_level": 1, + "bbox": [ + 89, + 587, + 405, + 603 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Exploring the cross-lingual capabilities of GUI agents is highly beneficial for their application in non-English environments. We split the training and testing sets of GUIAct (web-multi) data into English and Chinese partitions, respectively. We then fine-tune SpiritSight-Base (8B) on two sets: the entire training set (English&Chinese) and the English-only training set. The results are shown in Tab. 4.", + "bbox": [ + 89, + 611, + 482, + 715 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Under the English&Chinese configuration, SpiritSight achieves comparable results on both the English and Chinese testing sets, despite having fewer Chinese samples in the pre-training data than English ones. Under the English-only configuration, SpiritSight achieves $24.5\\%$ Step SR on the Chinese testing set, reaching half of the English&Chinese performance. The zero-shot capability of SpiritSight in Chinese results from the small but effective foundational Chinese data included in the pre-training stage.", + "bbox": [ + 89, + 718, + 482, + 868 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This experiment offers a paradigm for applying GUI agents to non-English environments: by collecting (1) free", + "bbox": [ + 89, + 869, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/8b3064ee7455fe2f035587fe2b32c9b2ab81f40eac9b2cf673817a3f39491450.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
SFT DataOverallChineseEnglish
English+Chinese49.3%49.3%49.2%
English35.0%24.5%48.6%
", + "bbox": [ + 550, + 327, + 870, + 386 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4. Results of SpiritSight on GUIAct (web-multi), trained with different language datasets.", + "bbox": [ + 511, + 407, + 903, + 435 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "web and mobile GUI information from the target language environment (level 1 & level 2 data), and (2) a small amount of high-quality GUI navigation data at low cost (level 3 data). With this language transferring strategy, the same capabilities as in the English environment can be achieved in non-English environments with minimal extra costs.", + "bbox": [ + 511, + 460, + 905, + 551 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Limitation", + "text_level": 1, + "bbox": [ + 513, + 565, + 627, + 580 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As SpiritSight is a vision-based GUI agent, it constantly requires access to screenshots which may contain personal information or other sensitive data. Users and system providers should carefully manage the system privileges granted to SpiritSight agent to mitigate potential privacy and security risks.", + "bbox": [ + 511, + 590, + 905, + 679 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "8. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 694, + 633, + 709 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we propose an advanced vision-based end-to-end GUI agent, SpiritSight, with high generalization across multiple GUI platforms. We construct an efficient multilevel, large-scale, high-quality GUI pre-training dataset to equip SpiritSight with robust GUI perception, grounding and understanding capabilities. Additionally, we introduce a UBP method to resolve ambiguity in dynamic high-resolution inputs during model training, further enhancing the ability of SpiritSight to ground GUI objects. As a result, SpiritSight achieves strong performance in numerous GUI navigation benchmarks, demonstrating significant potential for practical deployment in real-world applications.", + "bbox": [ + 511, + 719, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "29497", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 90, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 7", + "[2] Gilles Baechler, Srinivas Sunkara, Maria Wang, Fedir Zubach, Hassan Mansoor, Vincent Etter, Victor Carbune, Jason Lin, Jindong Chen, and Abhanshu Sharma. Screenai: A vision-language model for ui and infographics understanding. arXiv preprint arXiv:2402.04615, 2024. 3", + "[3] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966, 2023.7, 1,4", + "[4] Rohan Bavishi, Erich Elsen, Curtis Hawthorne, Maxwell Nye, Augustus Odena, Arushi Somani, and Săgnak Taşürlar. Introducing our multimodal models, 2023. 3, 7", + "[5] Yuxiang Chai, Siyuan Huang, Yazhe Niu, Han Xiao, Liang Liu, Dingyu Zhang, Peng Gao, Shuai Ren, and Hongsheng Li. Amex: Android multi-annotation expo dataset for mobile gui agents. arXiv preprint arXiv:2407.17490, 2024. 2, 5, 6, 7, 1, 4", + "[6] Dongping Chen, Yue Huang, Siyuan Wu, Jingyu Tang, Liuyi Chen, Yilin Bai, Zhigang He, Chenlong Wang, Huichi Zhou, Yiqiang Li, et al. Gui-world: A dataset for gui-oriented multimodal llm-based agents. arXiv preprint arXiv:2406.10819, 2024. 2", + "[7] Wentong Chen, Junbo Cui, Jinyi Hu, Yujia Qin, Junjie Fang, Yue Zhao, Chongyi Wang, Jun Liu, Guirong Chen, Yupeng Huo, et al. Guicourse: From general vision language models to versatile gui agents. arXiv preprint arXiv:2406.11317, 2024. 1, 3, 5, 6, 7, 4", + "[8] Xingyu Chen, Zihan Zhao, Lu Chen, Danyang Zhang, Jiabao Ji, Ao Luo, Yuxuan Xiong, and Kai Yu. Websrc: A dataset for web-based structural reading comprehension. arXiv preprint arXiv:2101.09465, 2021. 5", + "[9] Zhe Chen, Weiyun Wang, Hao Tian, Shenglong Ye, Zhangwei Gao, Erfei Cui, Wenwen Tong, Kongzhi Hu, Jiapeng Luo, Zheng Ma, et al. How far are we to gpt-4v? closing the gap to commercial multimodal models with open-source suites. arXiv preprint arXiv:2404.16821, 2024. 2, 5, 6, 1", + "[10] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24185–24198, 2024. 1", + "[11] Kanzhi Cheng, Qiushi Sun, Yougang Chu, Fangzhi Xu, Yantao Li, Jianbing Zhang, and Zhiyong Wu. Seeclick: Harnessing gui grounding for advanced visual gui agents. arXiv preprint arXiv:2401.10935, 2024. 1, 3, 4, 6, 7, 8", + "[12] Biplab Deka, Zifeng Huang, Chad Franzen, Joshua Hibschman, Daniel Afergan, Yang Li, Jeffrey Nichols, and Ranjitha Kumar. Rico: A mobile app dataset for building data-driven design applications. In Proceedings of the 30th annual" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ACM symposium on user interface software and technology, pages 845-854, 2017. 5, 3", + "[13] Xiang Deng, Yu Gu, Boyuan Zheng, Shijie Chen, Sam Stevens, Boshi Wang, Huan Sun, and Yu Su. Mind2web: Towards a generalist agent for the web. Advances in Neural Information Processing Systems, 36, 2024. 1, 2, 6, 4", + "[14] Jacob Devlin. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 1", + "[15] Hiroki Furuta, Kuang-Huei Lee, Ofir Nachum, Yutaka Matsuo, Aleksandra Faust, Shixiang Shane Gu, and Izzeddin Gur. Multimodal web navigation with instruction-finetuned foundation models. arXiv preprint arXiv:2305.11854, 2023. 1, 3", + "[16] Common crawl Group. Common crawl - open repository of web crawl data, 2024. 3, 5", + "[17] Izzeddin Gur, Hiroki Furuta, Austin Huang, Mustafa Safdari, Yutaka Matsuo, Douglas Eck, and Aleksandra Faust. A real-world webagent with planning, long context understanding, and program synthesis. arXiv preprint arXiv:2307.12856, 2023. 2, 3", + "[18] Markus Hafner, Maria Katsantoni, Tino Köster, James Marks, Joyita Mukherjee, Dorothee Staiger, Jernej Ule, and Mihaela Zavolan. Clip and complementary methods. Nature Reviews Methods Primers, 1(1):1-23, 2021. 1", + "[19] Wenyi Hong, Weihan Wang, Qingsong Lv, Jiazheng Xu, Wenmeng Yu, Junhui Ji, Yan Wang, Zihan Wang, Yuxiao Dong, Ming Ding, et al. Cogagent: A visual language model for gui agents. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14281-14290, 2024. 1, 3, 7", + "[20] Yu-Chung Hsiao, Fedir Zubach, Maria Wang, et al. Screenqa: Large-scale question-answer pairs over mobile app screenshots. arXiv preprint arXiv:2209.08199, 2022. 5", + "[21] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 3", + "[22] Faria Huq, Jeffrey P Bigham, and Nikolas Martelaro. “what’s important here?”: Opportunities and challenges of using llms in retrieving information from web interfaces. arXiv preprint arXiv:2312.06147, 2023. 1", + "[23] Yang Jin, Kun Xu, Kun Xu, Liwei Chen, Chao Liao, Jianchao Tan, Quzhe Huang, Bin Chen, Chenyi Lei, An Liu, et al. Unified language-vision pretraining in llm with dynamic discrete visual tokenization. arxiv 2024. arXiv preprint arXiv:2309.04669, 2023. 1", + "[24] Jihyung Kil, Chan Hee Song, Boyuan Zheng, Xiang Deng, Yu Su, and Wei-Lun Chao. Dual-view visual contextualization for web navigation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14445-14454, 2024. 1", + "[25] Geewook Kim, Teakgyu Hong, Moonbin Yim, Jeong Yeon Nam, Jinyoung Park, Jinyeong Yim, Wonseok Hwang, Sang-doo Yun, Dongyoon Han, and Seunghyun Park. Ocr-free document understanding transformer. In European Conference on Computer Vision, pages 498-517. Springer, 2022. 1, 2" + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "29498", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[26] Hanyu Lai, Xiao Liu, Iat Long Iong, Shuntian Yao, Yuxuan Chen, Pengbo Shen, Hao Yu, Hanchen Zhang, Xiaohan Zhang, Yuxiao Dong, et al. Autowebglm: Bootstrap and reinforce a large language model-based web navigating agent. arXiv preprint arXiv:2404.03648, 2024. 1, 2, 3", + "[27] Kenton Lee, Mandar Joshi, Iulia Raluca Turc, Hexiang Hu, Fangyu Liu, Julian Martin Eisenschlos, Urvashi Khandelwal, Peter Shaw, Ming-Wei Chang, and Kristina Toutanova. Pix2struct: Screenshot parsing as pretraining for visual language understanding. In International Conference on Machine Learning, pages 18893–18912. PMLR, 2023. 2, 4", + "[28] Kuang-Huei Lee, Xinyun Chen, Hiroki Furuta, John Canny, and Ian Fischer. A human-inspired reading agent with gist memory of very long contexts. arXiv preprint arXiv:2402.09727, 2024. 1, 3", + "[29] Feng Li, Renrui Zhang, Hao Zhang, Yuanhan Zhang, Bo Li, Wei Li, Zejun Ma, and Chunyuan Li. Llava next-interleave: Tackling multi-image, video, and 3d in large multimodal models. arXiv preprint arXiv:2407.07895, 2024. 1", + "[30] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International conference on machine learning, pages 12888-12900. PMLR, 2022. 1", + "[31] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730-19742. PMLR, 2023. 1", + "[32] Wei Li, William Bishop, Alice Li, Chris Rawles, Folawiyo Campbell-Ajala, Divya Tyamagundlu, and Oriana Riva. On the effects of data scale on computer control agents. arXiv preprint arXiv:2406.03679, 2024. 6, 7, 1, 4", + "[33] Yang Li, Gang Li, Luheng He, Jingjie Zheng, Hong Li, and Zhiwei Guan. Widget captioning: Generating natural language description for mobile user interface elements. arXiv preprint arXiv:2010.04295, 2020. 5, 3", + "[34] Zeyi Liao, Lingbo Mo, Chejian Xu, Mintong Kang, Jiawei Zhang, Chaowei Xiao, Yuan Tian, Bo Li, and Huan Sun. Eia: Environmental injection attack on generalist web agents for privacy leakage. arXiv preprint arXiv:2409.11295, 2024. 1, 3", + "[35] Ziyi Lin, Chris Liu, Renrui Zhang, Peng Gao, Longtian Qiu, Han Xiao, Han Qiu, Chen Lin, Wenqi Shao, Keqin Chen, et al. Sphinx: The joint mixing of weights, tasks, and visual embeddings for multi-modal large language models. arXiv preprint arXiv:2311.07575, 2023. 1", + "[36] Evan Zheran Liu, Kelvin Guu, Panupong Pasupat, Tianlin Shi, and Percy Liang. Reinforcement learning on web interfaces using workflow-guided exploration. arXiv preprint arXiv:1802.08802, 2018. 2, 1", + "[37] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning, 2023. 1", + "[38] Quanfeng Lu, Wenqi Shao, Zitao Liu, Fanqing Meng, Boxuan Li, Botong Chen, Siyuan Huang, Kaipeng Zhang, Yu Qiao, and Ping Luo. Gui odyssey: A comprehensive dataset for cross-app gui navigation on mobile devices. arXiv preprint arXiv:2406.08451, 2024. 2, 6, 7, 1, 4" + ], + "bbox": [ + 91, + 90, + 480, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[39] Xing Han Lu, Zdenek Kasner, and Siva Reddy. Weblinx: Real-world website navigation with multi-turn dialogue. arXiv preprint arXiv:2402.05930, 2024. 2", + "[40] Minesh Mathew, Dimosthenis Karatzas, and CV Jawahar. Docvqa: A dataset for vqa on document images. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 2200-2209, 2021. 5", + "[41] Erik Nijkamp, Bo Pang, Hiroaki Hayashi, Lifu Tu, Huan Wang, Yingbo Zhou, Silvio Savarese, and Caiming Xiong. Codegen: An open large language model for code with multi-turn program synthesis. arXiv preprint arXiv:2203.13474, 2022. 1", + "[42] OpenAI. Hello gpt-4o, 2024. 7", + "[43] Liangming Pan, Alon Albalak, Xinyi Wang, and William Yang Wang. Logic-lm: Empowering large language models with symbolic solvers for faithful logical reasoning. arXiv preprint arXiv:2305.12295, 2023. 1", + "[44] Alec Radford, Karthik Narasimhan, Tim Salimans, Ilya Sutskever, et al. Improving language understanding by generative pre-training. OpenAI, 2018.", + "[45] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of machine learning research, 21(140):1-67, 2020. 1", + "[46] Christopher Rawles, Alice Li, Daniel Rodriguez, Oriana Riva, and Timothy Lillicrap. Androidinthewild: A large-scale dataset for android device control. Advances in Neural Information Processing Systems, 36, 2024. 2, 4, 6, 1, 7", + "[47] Baptiste Roziere, Jonas Gehring, Fabian Gloeckle, Sten Sootla, Itai Gat, Xiaqing Ellen Tan, Yossi Adi, Jingyu Liu, Romain Sauvestre, Tal Remez, et al. Code llama: Open foundation models for code. arXiv preprint arXiv:2308.12950, 2023. 1", + "[48] Rico Sennrich. Improving neural machine translation models with monolingual data. arXiv preprint arXiv:1511.06709, 2015. 4", + "[49] Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Mingchuan Zhang, YK Li, Yu Wu, and Daya Guo. Deepseekmath: Pushing the limits of mathematical reasoning in open language models. arXiv preprint arXiv:2402.03300, 2024.1", + "[50] Peter Shaw, Mandar Joshi, James Cohan, Jonathan Berant, Panupong Pasupat, Hexiang Hu, Urvashi Khandelwal, Kenton Lee, and Kristina N Toutanova. From pixels to ui actions: Learning to follow instructions via graphical user interfaces. Advances in Neural Information Processing Systems, 36:34354-34370, 2023. 3", + "[51] Tianlin Shi, Andrej Karpathy, Linxi Fan, Jonathan Hernandez, and Percy Liang. World of bits: An open-domain platform for web-based agents. In International Conference on Machine Learning, pages 3135-3144. PMLR, 2017. 2, 1", + "[52] Lucas-Andrei Thil, Mirela Popa, and Gerasimos Spanakis. Navigating webai: Training agents to complete web tasks with large language models and reinforcement learning. In Proceedings of the 39th ACM/SIGAPP Symposium on Applied Computing, pages 866–874, 2024. 1, 3" + ], + "bbox": [ + 516, + 90, + 906, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "29499", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[53] Jianqiang Wan, Sibo Song, Wenwen Yu, Yuliang Liu, Wenqing Cheng, Fei Huang, Xiang Bai, Cong Yao, and Zhibo Yang. Omniparser: A unified framework for text spotting key information extraction and table recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15641-15653, 2024. 1, 3", + "[54] Bryan Wang, Gang Li, Xin Zhou, Zhourong Chen, Tovi Grossman, and Yang Li. Screen2words: Automatic mobile ui summarization with multimodal learning. In The 34th Annual ACM Symposium on User Interface Software and Technology, pages 498-510, 2021. 5, 3", + "[55] Junyang Wang, Haiyang Xu, Haitao Jia, Xi Zhang, Ming Yan, Weizhou Shen, Ji Zhang, Fei Huang, and Jitao Sang. Mobile-agent-v2: Mobile device operation assistant with effective navigation via multi-agent collaboration. arXiv preprint arXiv:2406.01014, 2024. 2, 3", + "[56] Junyang Wang, Haiyang Xu, Jiabo Ye, Ming Yan, Weizhou Shen, Ji Zhang, Fei Huang, and Jitao Sang. Mobile-agent: Autonomous multi-modal mobile device agent with visual perception. arXiv preprint arXiv:2401.16158, 2024. 2, 3", + "[57] Ke Wang, Houxing Ren, Aojun Zhou, Zimu Lu, Sichun Luo, Weikang Shi, Renrui Zhang, Linqi Song, Mingjie Zhan, and Hongsheng Li. Mathcoder: Seamless code integration in llms for enhanced mathematical reasoning. arXiv preprint arXiv:2310.03731, 2023. 1", + "[58] Weihan Wang, Qingsong Lv, Wenmeng Yu, Wenyi Hong, Ji Qi, Yan Wang, Junhui Ji, Zhuoyi Yang, Lei Zhao, Xixuan Song, et al. Cogvlm: Visual expert for pretrained language models. arXiv preprint arXiv:2311.03079, 2023. 1", + "[59] Jason Wei, Yi Tay, Rishi Bommasani, Colin Raffel, Barret Zoph, Sebastian Borgeaud, Dani Yogatama, Maarten Bosma, Denny Zhou, Donald Metzler, et al. Emergent abilities of large language models. arXiv preprint arXiv:2206.07682, 2022. 1", + "[60] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824-24837, 2022. 5, 1, 7", + "[61] Chen Henry Wu, Jing Yu Koh, Ruslan Salakhutdinov, Daniel Fried, and Aditi Raghunathan. Adversarial attacks on multimodal agents. arXiv preprint arXiv:2406.12814, 2024. 1, 3", + "[62] Shijie Wu, Ozan Irsoy, Steven Lu, Vadim Dabravolski, Mark Dredze, Sebastian Gehrmann, Prabhanjan Kambadur, David Rosenberg, and Gideon Mann. Bloomberg: A large language model for finance. arXiv preprint arXiv:2303.17564, 2023. 1", + "[63] Yichong Xu, Chenguang Zhu, Shuohang Wang, Siqi Sun, Hao Cheng, Xiaodong Liu, Jianfeng Gao, Pengcheng He, Michael Zeng, and Xuedong Huang. Human parity on commonsenseqa: Augmenting self-attention with external attention. arXiv preprint arXiv:2112.03254, 2021. 1", + "[64] Zhao Yang, Jiaxuan Liu, Yucheng Han, Xin Chen, Zebiao Huang, Bin Fu, and Gang Yu. Appagent: Multimodal agents as smartphone users. arXiv preprint arXiv:2312.13771, 2023. 3" + ], + "bbox": [ + 91, + 90, + 480, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[65] Shunyu Yao, Howard Chen, John Yang, and Karthik Narasimhan. Webshop: Towards scalable real-world web interaction with grounded language agents. Advances in Neural Information Processing Systems, 35:20744-20757, 2022. 2, 1", + "[66] Yuan Yao, Tianyu Yu, Ao Zhang, Chongyi Wang, Junbo Cui, Hongji Zhu, Tianchi Cai, Haoyu Li, Weilin Zhao, Zhihui He, et al. Minicpm-v: A gpt-4v level mllm on your phone. arXiv preprint arXiv:2408.01800, 2024. 1", + "[67] Jiabo Ye, Anwen Hu, Haiyang Xu, Qinghao Ye, Ming Yan, Guohai Xu, Chenliang Li, Junfeng Tian, Qi Qian, Ji Zhang, et al. Ureader: UniversalOCR-free visually-situated language understanding with multimodal large language model. arXiv preprint arXiv:2310.05126, 2023. 2, 5", + "[68] Da Yin, Faeze Brahman, Abhilasha Ravichander, Khyathi Chandu, Kai-Wei Chang, Yejin Choi, and Bill Yuchen Lin. Agent lumos: Unified and modular training for open-source language agents. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 12380–12403, 2024. 1", + "[69] Huaiyuan Ying, Shuo Zhang, Linyang Li, Zhejian Zhou, Yunfan Shao, Zhaoye Fei, Yichuan Ma, Jiawei Hong, Kuikun Liu, Ziyi Wang, et al. Internlm-math: Open math large language models toward verifiable reasoning. arXiv preprint arXiv:2402.06332, 2024. 1", + "[70] Longhui Yu, Weisen Jiang, Han Shi, Jincheng Yu, Zhengying Liu, Yu Zhang, James T Kwok, Zhenguo Li, Adrian Weller, and Weiyang Liu. Metamath: Bootstrap your own mathematical questions for large language models. arXiv preprint arXiv:2309.12284, 2023. 1", + "[71] Qiusi Zhan, Zhixiang Liang, Zifan Ying, and Daniel Kang. Injecagent: Benchmarking indirect prompt injections in tool-integrated large language model agents. arXiv preprint arXiv:2403.02691, 2024. 1, 3", + "[72] Jiwen Zhang, Jihao Wu, Yihua Teng, Minghui Liao, Nuo Xu, Xiao Xiao, Zhongyu Wei, and Duyu Tang. Android in the zoo: Chain-of-action-thought for gui agents. arXiv preprint arXiv:2403.02713, 2024. 5, 7", + "[73] Pan Zhang, Xiaoyi Dong, Yuhang Zang, Yuhang Cao, Rui Qian, Lin Chen, Qipeng Guo, Haodong Duan, Bin Wang, Linke Ouyang, Songyang Zhang, Wenwei Zhang, Yining Li, Yang Gao, Peng Sun, Xinyue Zhang, Wei Li, Jingwen Li, Wenhai Wang, Hang Yan, Conghui He, Xingcheng Zhang, Kai Chen, Jifeng Dai, Yu Qiao, Dahua Lin, and Jiaqi Wang. Internlm-xcomposer-2.5: A versatile large vision language model supporting long-contextual input and output, 2024. 1", + "[74] Boyuan Zheng, Boyu Gou, Jihyung Kil, Huan Sun, and Yu Su. Gpt-4v (isdiction) is a generalist web agent, if grounded. arXiv preprint arXiv:2401.01614, 2024. 1, 2, 3", + "[75] Longtao Zheng, Rundong Wang, Xinrun Wang, and Bo An. Synapse: Trajectory-as-exemplar prompting with memory for computer control. In The Twelfth International Conference on Learning Representations, 2023. 1" + ], + "bbox": [ + 516, + 90, + 903, + 838 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "29500", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/SpiritSight Agent_ Advanced GUI Agent with One Look/c38ba9da-d4d0-4e2a-a186-fdd0cc0687b1_model.json b/2025/SpiritSight Agent_ Advanced GUI Agent with One Look/c38ba9da-d4d0-4e2a-a186-fdd0cc0687b1_model.json new file mode 100644 index 0000000000000000000000000000000000000000..61d3df854116174b81a5dfd1777ec9222ff7b693 --- /dev/null +++ b/2025/SpiritSight Agent_ Advanced GUI Agent with One Look/c38ba9da-d4d0-4e2a-a186-fdd0cc0687b1_model.json @@ -0,0 +1,2587 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "image", + "bbox": [ + 0.201, + 0.127, + 0.235, + 0.148 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.233, + 0.131, + 0.798, + 0.153 + ], + "angle": 0, + "content": "SpiritSight Agent: Advanced GUI Agent with One Look" + }, + { + "type": "text", + "bbox": [ + 0.123, + 0.188, + 0.874, + 0.208 + ], + "angle": 0, + "content": "Zhiyuan Huang\\(^{1}\\) Ziming Cheng\\(^{1,2*}\\) Junting Pan\\(^{3}\\) Zhaohui Hou\\(^{1}\\) Mingjie Zhan\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.125, + 0.211, + 0.875, + 0.23 + ], + "angle": 0, + "content": "\\(^{1}\\)SenseTime Research, \\(^{2}\\)Beijing University of Posts and Telecommunications, \\(^{3}\\)MMLab, CUHK" + }, + { + "type": "text", + "bbox": [ + 0.264, + 0.25, + 0.734, + 0.266 + ], + "angle": 0, + "content": "{huangzhiyuan, houzhaohui, zhanmingjie}@sensetime.com" + }, + { + "type": "text", + "bbox": [ + 0.272, + 0.272, + 0.72, + 0.286 + ], + "angle": 0, + "content": "harrytheorange@bupt.edu.cn junting.pa@gmail.com" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.322, + 0.327, + 0.339 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.354, + 0.485, + 0.732 + ], + "angle": 0, + "content": "Graphical User Interface (GUI) agents demonstrate promising potential in assisting human-computer interaction, automating human user's navigation on digital devices. An ideal GUI agent is expected to achieve high accuracy, low latency, and compatibility for different GUI platforms. Recent vision-based approaches have shown promise by leveraging advanced Vision Language Models (VLMs). While they generally meet the requirements of compatibility and low latency, these vision-based GUI agents tend to have low accuracy due to their limitations in element grounding. To address this issue, we propose SpiritSight, a vision-based, end-to-end GUI agent that excels in GUI navigation tasks across various GUI platforms. First, we create a multi-level, large-scale, high-quality GUI dataset called GUI-Lasagne using scalable methods, empowering SpiritSight with robust GUI understanding and grounding capabilities. Second, we introduce the Universal Block Parsing (UBP) method to resolve the ambiguity problem inherited from the dynamic resolution strategy, further enhancing SpiritSight's ability to ground GUI objects. Through these efforts, SpiritSight agent outperforms other advanced methods on diverse GUI benchmarks, demonstrating its superior capability and compatibility in GUI navigation tasks. The models and code will be made available upon publication." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.761, + 0.222, + 0.777 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.786, + 0.484, + 0.877 + ], + "angle": 0, + "content": "Graphical User Interface (GUI) automation has long been pursued by people along with the development of modern digital devices. Thanks to recent advances in Large Language Models (LLMs), GUI agents are constructed to assist users in interacting with graphical interfaces, automatically making action decisions based on observations of environ-" + }, + { + "type": "image", + "bbox": [ + 0.53, + 0.322, + 0.891, + 0.552 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.561, + 0.908, + 0.604 + ], + "angle": 0, + "content": "Figure 1. Our SpiritSight agent achieves new state-of-the-art (SOTA) performance across various benchmarks in web, mobile, and desktop scenarios." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.636, + 0.698, + 0.65 + ], + "angle": 0, + "content": "ment and user's instruction." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.654, + 0.907, + 0.759 + ], + "angle": 0, + "content": "Current approaches can be divided into three categories based on their input modalities. Language-based and vision-based approaches use Hyper Text Markup Language (HTML) or Extensible Markup Language (XML) and screenshots as input [7, 11, 13, 19, 22, 26, 28, 53, 68, 75], respectively, while vision-language-based approaches combine both HTML and screenshots as input [15, 24, 52, 74]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.762, + 0.909, + 0.868 + ], + "angle": 0, + "content": "The language-based and vision-language-based methods typically applied only in the web domain, and are often limited by the excessive length of HTML or security concerns [34, 61, 71]. The vision-based approaches exhibit compatibility across various GUI platforms, as acquiring screenshots is generally easier than obtaining hierarchical data from non-web platforms." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.909, + 0.901 + ], + "angle": 0, + "content": "However, vision-based approaches struggle to ground GUI elements (e.g. buttons, text boxes) from the visual" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.22, + 0.9 + ], + "angle": 0, + "content": "*Equal contribution." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "29490" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.288 + ], + "angle": 0, + "content": "input [74]. Some works solve this problem by introducing additional tools, such as optical character recognition (OCR) and icon recognition models, to convert the visual grounding task into a language QA task [55, 56]. This may increase the complexity and inference latency of the agent system. Others attempt to collect large-scale training data through manual synthesis [27, 36, 51] or human annotation [5, 6, 13, 38, 39, 46, 65], while these data are either unrealistic or costly. Additionally, recent advanced Vision Language Models (VLMs) [9, 67] adapt dynamic high-resolution [9, 67] strategy to better suit high-resolution inputs, which we find may introduce ambiguity to the model learning process in GUI scenarios." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.289, + 0.484, + 0.349 + ], + "angle": 0, + "content": "To address the aforementioned challenges, we propose an end-to-end, vision-based GUI agent—SpiritSight, which has strong ability for GUI navigation task. Our contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.35, + 0.485, + 0.501 + ], + "angle": 0, + "content": "Firstly, we propose GUI-Lasagne, a multi-level, large-scale and high-quality GUI dataset composed of 5.73 million samples to enhance our model's GUI understanding and grounding capabilities. The dataset is collected from the real world and filtered through carefully designed rules to ensure data quality. It is also constructed hierarchically and consists of three levels of components. The first two parts constitute \\(90\\%\\) of the total dataset and are collected for free, thus significantly reducing the data collection cost." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.502, + 0.484, + 0.592 + ], + "angle": 0, + "content": "Secondly, we introduce a Universal Block Parsing (UBP) method to resolve the ambiguity problem inherited from the dynamic resolution strategy. With the effectiveness of UBP, SpiritSight gains an improved elements grounding capability and achieves enhanced performance in GUI navigation tasks." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.593, + 0.484, + 0.79 + ], + "angle": 0, + "content": "Thirdly, we release and evaluate SpiritSight agent in various GUI benchmarks and it exhibits remarkable performance among them. SpiritSight agent is pre-trained on GUI-Lasagne and employs UBP to ground elements, as shown in Fig. 3. We release three versions of SpiritSight with different model size: SpiritSight-26B, SpiritSight-8B, and SpiritSight-2B. The standard SpiritSight-8B consistently outperforms previous state-of-the-art methods across various GUI benchmarks, as shown in Fig. 1. On the Multimodal-Mind2Web [13] benchmark, the SpiritSight series outperforms all kinds of methods—including language-based, vision-based and even vision-language-based methods—in the non-candidate setting, as shown in Fig. 2." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.806, + 0.229, + 0.822 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.832, + 0.483, + 0.863 + ], + "angle": 0, + "content": "2.1. Language-based and vision-language-based GUI agent" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Several works leverage the capabilities of large-scale language models (LLMs) to construct GUI agents. It is" + }, + { + "type": "image", + "bbox": [ + 0.532, + 0.092, + 0.88, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.296, + 0.908, + 0.339 + ], + "angle": 0, + "content": "Figure 2. Comparison of the average step success rate on Multimodal-Mind2Web benchmark of our SpiritSight agent of three sizes (2B, 8B, 26B) with various previous methods." + }, + { + "type": "image", + "bbox": [ + 0.529, + 0.356, + 0.897, + 0.573 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.592, + 0.907, + 0.662 + ], + "angle": 0, + "content": "Figure 3. The overview of our SpiritSight agent. Our proposed GUI-Lasagne dataset equips SpiritSight with three levels of comprehensive GUI knowledge. Additionally, we introduce a Universal Block Parsing (UBP) method to enhance SpiritSight's grounding capabilities." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.907, + 0.87 + ], + "angle": 0, + "content": "noticed that they are mostly multi-stage architectures. Mind2Web [13] utilizes a lightweight language model to extract candidate elements from HTML, followed by a ranking model that sorts these elements based on task descriptions and historical actions. Finally, a large language model predicts the current actions and selects the target elements to operate on. WebAgent [17] first uses an encoder-decoder model to generate low-level instructions and relevant HTML code snippets, then uses another decoder to produce executable Python code. AutoWebGLM [26] simplifies HTML code through manually designed rules before predicting the executable action codes." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.907, + 0.9 + ], + "angle": 0, + "content": "Other vision-language-based works leverage both GUI screenshots and hierarchical HTML/XML to enhance the" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "29491" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.094, + 0.088, + 0.905, + 0.301 + ], + "angle": 0, + "content": "
Model SizeInput ModalitySelect From TopCross-TaskCross-WebsiteCross-Domain
Ele.AccOp.F1Step SREle.AccOp.F1Step SREle.AccOp.F1Step SR
AutoWebGLM [26]6BText✓--66.4%--56.4%--55.8%
LLaMA2-7B [26]7BText✓--52.7%--47.1%--50.3%
CogAgent [19]18BImage✓--62.3%--54.0%--59.4%
HTML-T5-XL [17]3BText✓76.4%78.8%71.5%68.4%71.0%62.2%73.0%75.6%67.1%
SeeAct* [74]-Text+Image×46.4%73.4%40.2%38.0%67.8%32.4%42.4%69.3%36.8%
ReadAgent-P* [28]340BText×33.7%72.5%29.2%37.4%75.1%31.1%37.2%76.3%33.4%
MiniCPM-GUI [7]3BImage×23.8%86.8%20.8%20.3%81.7%17.3%17.9%74.5%14.6%
Fuyu-GUI [4]8BImage×19.1%86.1%15.6%13.9%80.7%12.2%14.2%83.1%11.7%
SeeClick [11]9.6BImage×28.3%87.0%25.5%21.4%80.6%16.4%23.2%84.8%20.8%
OmniParser* [53]-Image×42.4%87.6%39.4%41.0%84.8%36.5%45.5%85.7%42.0%
SpiritSight-2B2BImage×51.7%87.2%44.9%44.0%83.6%37.8%42.4%83.5%36.9%
SpiritSight-8B8BImage×59.2%88.9%52.7%52.2%84.7%44.0%50.1%86.0%44.4%
SpiritSight-26B26BImage×60.5%89.7%54.7%57.0%85.7%48.1%54.1%87.2%49.2%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.324, + 0.908, + 0.368 + ], + "angle": 0, + "content": "Table 1. Comparison of SpiritSight agent with various previous methods on Multimodal-Mind2Web benchmark. SpiritSight significantly outperforms all methods that do not rely on any candidate element, including vision-based methods, language-based methods, and even vision-language-based methods. The asterisk (*) in the upper-right indicates that the model is not fine-tuned on the downstream tasks." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.392, + 0.482, + 0.514 + ], + "angle": 0, + "content": "robustness of GUI agents. WebGUM [15] and CC-Net [52] use ResNet and ViT to extract features from screenshots, respectively. The image embeddings are then combined with text embeddings and fed into a multi-modal transformer. SeeAct [74] and AppAgent [64] identify all interactive elements using HTML or XML data. They then assign each interactive element a unique identifier in the screenshot before feeding the screenshot into the model." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.515, + 0.483, + 0.696 + ], + "angle": 0, + "content": "These language-based and vision-language-based methods, which rely on hierarchical information, exhibit several limitations: (1) Hierarchical representations, such as HTML or XML, are not consistently available across different platforms. And even this information is available, differences in their internal rules make language-based GUI agents less compatible; (2) HTML often contains redundant information, requiring additional models or extensive manually crafted rules for effective filtering. (3) Text-based GUI agents are susceptible to injection attacks [34, 61, 71], where malicious instructions hidden in HTML can easily lead to erroneous or unsafe actions." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.71, + 0.312, + 0.727 + ], + "angle": 0, + "content": "2.2. Vision-based GUI agent" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Recently, some vision-based approaches have been proposed to overcome the drawbacks of language-based methods. Some of them [2, 11, 19, 50] are single-stage methods that only use GUI screenshots as input for VLMs and output the next action in an end-to-end manner. However, these agents perform worse on GUI benchmarks compared to other methods. Two primary reasons may be the relatively low input resolution of their base models and the insufficient scale and suboptimal quality of the training data. SeeAct [74], MobileAgent [56] and MobileAgent-v2 [55] are two-stage methods. They use GPT-4V instead of open" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.393, + 0.907, + 0.528 + ], + "angle": 0, + "content": "source VLMs as base models and find that even the top models struggle with element grounding. Consequently, they introduce additional OCR and icon recognition models to assist with element grounding, which may increase the complexity and inference latency of the agent system. Overall, the current VLMs exhibit limited ability in handling GUI grounding tasks, thus constrains the navigation capabilities of single-stage vision-based GUI agents. Related works on base models are discussed in Sec. 9." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.545, + 0.665, + 0.56 + ], + "angle": 0, + "content": "3. Data collection" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.572, + 0.907, + 0.662 + ], + "angle": 0, + "content": "In this section, we present a cost-effective data collection strategy designed to construct a multi-level, large-scale and high-quality GUI dataset, called GUI-Lasagne, as shown in Fig. 4. This dataset helps equip our models with robust abilities in GUI understanding, grounding, and navigation. More details of the collection can be seen in Tab. 8." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.674, + 0.798, + 0.691 + ], + "angle": 0, + "content": "3.1. Level one: visual-text alignment" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.699, + 0.907, + 0.851 + ], + "angle": 0, + "content": "Visual-text alignment refers to the foundational ability of a GUI agent to perceive the visual appearance of the GUI, specifically the ability to recognize or locate text/icon elements. We collect visual-text data directly from the GUI source data. In the web scenario, we obtain web URLs from CommonCrawl [16] and website ranking sources12. We then develop a data collection tool using Playwright3 to gather real-world web data from the collected URLs. With this tool, we collect 755K web-page screenshots alone with their DOM data, featuring diversity in both resolution" + }, + { + "type": "page_footnote", + "bbox": [ + 0.532, + 0.863, + 0.828, + 0.876 + ], + "angle": 0, + "content": "1 https://www.semrush.com/trending-sites/global/all" + }, + { + "type": "page_footnote", + "bbox": [ + 0.534, + 0.876, + 0.68, + 0.888 + ], + "angle": 0, + "content": "2https://top.chinaz.com/all/" + }, + { + "type": "page_footnote", + "bbox": [ + 0.534, + 0.888, + 0.662, + 0.9 + ], + "angle": 0, + "content": "3https://playwright.dev/" + }, + { + "type": "list", + "bbox": [ + 0.532, + 0.863, + 0.828, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "29492" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.089, + 0.333, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.342, + 0.089, + 0.521, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.528, + 0.089, + 0.701, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.708, + 0.089, + 0.899, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.092, + 0.233, + 0.339, + 0.329 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.342, + 0.233, + 0.518, + 0.329 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.529, + 0.233, + 0.697, + 0.329 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.71, + 0.233, + 0.899, + 0.329 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.344, + 0.908, + 0.373 + ], + "angle": 0, + "content": "Figure 4. The collection pipeline of our GUILasagne dataset. The left, middle and right parts show the construction of level-1, level-2, and level-3 data, respectively." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.399, + 0.483, + 0.519 + ], + "angle": 0, + "content": "and language. Additionally, we develop an icon recognition model called InternVL-Icon as an icon annotation tool, by constructing a dataset of 30K icon-caption pairs from Alibaba Icon Library\\(^4\\) and fine-tuning InternVL1.5-26B using this dataset. We then use InternVL-Icon to annotate all icons on web-pages with generated captions. As for the mobile scenario, we collect data from AitW [46], a large-scale GUI dataset on the mobile platform." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.522, + 0.484, + 0.87 + ], + "angle": 0, + "content": "Inspired by Lee et al. [27] and Cheng et al. [11], we construct three tasks based the collected GUI source data: text2bbox,_bbox2text, and_bbox2dom. The text2bbox task prompts the model to ground the element based on the given text or icon caption. To avoid ambiguity, we provide additional contextual information for elements that appear multiple times in the screenshots. The text2bbox data is the most abundant among the three tasks, to help the model develop robust grounding capabilities. The_bbox2text task is the inverse of the text2bbox task, training the model to recognize text and icons. The_bbox2dom task asks the model to generate a DOM-tree corresponding to the given bounding box, as show in Sec. 12.7. This task helps the model learn not only to recognize text/icon elements but also to understand the GUI layout. To make sufficient use of the context length of the model, we pack multiple data pairs in each training sample for text2bbox and_bbox2text task, and select the box that includes as many elements as possible for_bbox2dom task. Overall, we construct a total of 1.9M and 1.1M training samples for the web and mobile scenarios, respectively. The data significantly enhance the GUI foundational abilities, especially the GUI grounding ability, of our SpiritSight agent. See Sec. 14 for more details on" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.4, + 0.654, + 0.414 + ], + "angle": 0, + "content": "sample construction." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.423, + 0.833, + 0.439 + ], + "angle": 0, + "content": "3.2. Level two: vision-function alignment" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.446, + 0.907, + 0.734 + ], + "angle": 0, + "content": "Vision-function alignment refers to a model's ability to locate an element based on its function. This type of data is not directly accessible from the raw GUI data. Inspired by the back-translation method [48] for dataset collection in translation tasks, we leverage InternVL's image understanding capabilities to collect function grounding data. Specifically, we divide each screenshot into a 3x3 grid and describe the approximate location of the target element in text (e.g. in the top-left corner of the image). Additionally, we place a bounding box around the target element in the screenshot to specify its precise location. We provide InternVL2-26B with the screenshot, the element's text content or icon caption, and the location description to prompt it to generate the element's function. Additionally, we utilize InternLM2.5-20B to enhance the quality and diversity of the generated function descriptions. These function descriptions achieve an approximate \\(90.9\\%\\) acceptance rate based on human judgment, which we consider sufficient for constructing the function grounding pre-training data." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.734, + 0.907, + 0.84 + ], + "angle": 0, + "content": "Based on the strategy above, we collect function2bbox pairs for all interactive elements in the web scenario. In the mobile scenario, we construct the function grounding data from the GUI navigation dataset, which is be described in Sec. 3.3. We use the same packing method as in the text2bbox and_bbox2text tasks for efficient training, ultimately obtaining 1.5M training samples." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.848, + 0.816, + 0.864 + ], + "angle": 0, + "content": "3.3. Level three: visual GUI navigation" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.907, + 0.901 + ], + "angle": 0, + "content": "We utilize the public available AitW [46] dataset to construct our GUI navigation training data. AitW is a large-" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.887, + 0.321, + 0.9 + ], + "angle": 0, + "content": "4https://www.iconfont.cn/?v=20230914" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "29493" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.484, + 0.287 + ], + "angle": 0, + "content": "scale mobile navigation dataset where each screenshot is labeled with the corresponding goal, the current step, etc. However, as noted in AitZ [72] and AMEX [5], the AitW dataset contains a number of incorrectly labeled samples. We choose to clean it with GPT-4o and employ Chain-of-Thought (CoT) [60] to make the judgment more accurate. Specifically, we prompt GPT-4o with the task description, the current action annotation, and screenshots from both the current and next steps. GPT-4o is then instructed to first summarize the two screenshots and identify the differences between them, then describe the current step based on these differences, and finally assess reasonability of the current action annotation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.289, + 0.484, + 0.395 + ], + "angle": 0, + "content": "We filter out steps identified as unreasonable by GPT-4o, resulting a final dataset of 0.64M CoT-style GUI navigation training samples. With the collected CoT-style data, we are able to construct additional function grounding data for the mobile scenario, as mentioned in Sec. 3.2, by treating the collected step descriptions as functions of the corresponding elements." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.408, + 0.279, + 0.425 + ], + "angle": 0, + "content": "3.4. Other training data" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.433, + 0.484, + 0.51 + ], + "angle": 0, + "content": "To enhance the model's understanding of GUI content, we further collected some public datasets as a supplement, including doc/web/mobile VQA [7, 8, 20, 40], image captioning [12, 54], and mobile grounding datasets [12, 33], resulting in 0.59M QA pairs for model training." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.526, + 0.315, + 0.544 + ], + "angle": 0, + "content": "4. Universal block parsing" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.552, + 0.274, + 0.567 + ], + "angle": 0, + "content": "4.1. Problem statement" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.577, + 0.484, + 0.668 + ], + "angle": 0, + "content": "We use InternVL2 [9] as the base model of SpiritSight, which uses a dynamic high-resolution training approach to largely preserves the details of input images. The approach first matches the optimal aspect ratio from a pre-defined set of aspect ratios to each input image, then resize and divide each image into blocks of \\(448 \\times 448\\) pixels." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.67, + 0.484, + 0.85 + ], + "angle": 0, + "content": "However, this approach may introduce ambiguity in grounding GUI elements. A typical example is represented in Fig. 5a, with two screenshots having aspect ratios of 1:2 and 2:1, respectively. According to the dynamic high-resolution strategy, the screenshots are divided into two blocks, one vertically and the other horizontally. Suppose the target elements in each sample are located in the same relative position within the second block (block-1) after cropping. The cropped screenshots are then fed into InternVL and the model is expected to predict different locations from these two identical inputs. We refer to this situation as the location ambiguity problem." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.852, + 0.483, + 0.882 + ], + "angle": 0, + "content": "Generally, a point \\(\\mathbf{p}\\) is expressed in the global coordinate system as" + }, + { + "type": "equation", + "bbox": [ + 0.251, + 0.885, + 0.483, + 0.903 + ], + "angle": 0, + "content": "\\[\n\\mathbf {p} = [ x, y ] \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.168 + ], + "angle": 0, + "content": "where \\( x \\) and \\( y \\) represent the horizontal and vertical coordinate values of the point in the original image, respectively. Assume that the image is resized and divided into \\( n_w \\times n_h \\) blocks. We can express the point \\( \\mathbf{p} \\) in the corresponding block as \\( \\mathbf{p}' \\):" + }, + { + "type": "equation", + "bbox": [ + 0.57, + 0.176, + 0.906, + 0.218 + ], + "angle": 0, + "content": "\\[\n\\left\\{ \\begin{array}{l} b _ {x} = \\left\\lfloor \\frac {x}{w _ {\\text {b l o c k}}} \\right\\rfloor \\\\ b _ {y} = \\left\\lfloor \\frac {y}{h _ {\\text {b l o c k}}} \\right\\rfloor \\end{array} \\quad \\left\\{ \\begin{array}{l} x ^ {\\prime} = x \\bmod w _ {\\text {b l o c k}} \\\\ y ^ {\\prime} = y \\bmod h _ {\\text {b l o c k}} \\end{array} \\right. \\right. \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.644, + 0.226, + 0.906, + 0.244 + ], + "angle": 0, + "content": "\\[\n\\mathbf {p} ^ {\\prime} = \\left[ b _ {x}, b _ {y}, x ^ {\\prime}, y ^ {\\prime} \\right] \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.25, + 0.907, + 0.34 + ], + "angle": 0, + "content": "where \\(w_{block}\\) and \\(h_{block}\\) (both set to 448 in our experiments) represent the width and height of each block, respectively. \\(b_{x}\\) and \\(b_{y}\\) represent the horizontal and vertical indices of the block containing \\(\\mathbf{p}'\\), and \\(x'\\) and \\(y'\\) represent the coordinates of \\(\\mathbf{p}'\\) within this block. The blocks are flattened into a sequence before being fed into the model, thus we have" + }, + { + "type": "equation", + "bbox": [ + 0.655, + 0.351, + 0.906, + 0.368 + ], + "angle": 0, + "content": "\\[\n\\mathbf {p} ^ {\\prime \\prime} = \\left[ b _ {i}, x ^ {\\prime}, y ^ {\\prime} \\right] \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.379, + 0.906, + 0.409 + ], + "angle": 0, + "content": "where \\( b_{i} \\) represents the index of the block within the sequence, satisfying:" + }, + { + "type": "equation", + "bbox": [ + 0.649, + 0.421, + 0.905, + 0.437 + ], + "angle": 0, + "content": "\\[\nb _ {i} = b _ {y} \\cdot n _ {w} + b _ {x} \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.447, + 0.906, + 0.524 + ], + "angle": 0, + "content": "The model is trained to approximate a mapping \\( f: \\mathbf{p}'' \\to \\mathbf{p} \\), which is inherently a multi-valued function. For example, when the input \\( \\mathbf{p}'' \\) is [1, 168, 245], possible values for \\( p \\) include [168, 693] when \\( n_w = 1 \\), or [616, 245] when \\( n_w = 2 \\), as shown in Fig. 5a." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.532, + 0.611, + 0.546 + ], + "angle": 0, + "content": "4.2. Method" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.555, + 0.907, + 0.797 + ], + "angle": 0, + "content": "The ambiguity primarily arises from the flattening operation, as shown in Eq. (5), which results in the loss of the spatial relationship between blocks. One solution is to input an additional thumbnail, but this may lead to extra computational and memory overhead. We propose to solve this positional ambiguity with two steps. Firstly, we introduce a 2D Block-wise Position Embedding (2D-BPE) [67] by adding two position embeddings to each block to capture spatial information. Secondly, we introduce a Universal Block Parsing (UBP) method, which replaces the global coordinate representation with a block-specific coordinate representation. Specifically, we express a point as Eq. (4). In this case, the model is trained to approximate a injective mapping \\( f: \\mathbf{p}'' \\rightarrow \\mathbf{p}'' \\), thereby resolving the ambiguity problem. During the model inference, the global coordinate of this point can be computed in post-processing as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.585, + 0.807, + 0.905, + 0.849 + ], + "angle": 0, + "content": "\\[\n\\left\\{ \\begin{array}{l} x = x ^ {\\prime} + \\left(b _ {i} \\mod n _ {w}\\right) \\cdot w _ {\\text {b l o c k}} \\\\ y = y ^ {\\prime} + \\left\\lfloor \\frac {b _ {i}}{n _ {w}} \\right\\rfloor \\cdot h _ {\\text {b l o c k}} \\end{array} \\right. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Overall, our UBP method ensures a clear mapping of positional information between the model's inputs and outputs, thereby improving the model's grounding capability." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "29494" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.089, + 0.585, + 0.249 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.192, + 0.251, + 0.483, + 0.265 + ], + "angle": 0, + "content": "(a) Baseline Block Parsing vs Universal Block Parsing." + }, + { + "type": "image", + "bbox": [ + 0.617, + 0.09, + 0.9, + 0.25 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.677, + 0.252, + 0.847, + 0.264 + ], + "angle": 0, + "content": "(b) Performance gain with UBP." + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.276, + 0.908, + 0.318 + ], + "angle": 0, + "content": "Figure 5. (a) Comparison between baseline block parsing and our proposed UBP. (b) The results of baseline block parsing and our proposed UBP methods on Multimodal-Mind2Web benchmark. UBP improves the performance of our model. The combination of UBP and 2D-BPE achieves the best results." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.343, + 0.183, + 0.361 + ], + "angle": 0, + "content": "5. Settings" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.379, + 0.305, + 0.396 + ], + "angle": 0, + "content": "5.1. Implementation details" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.414, + 0.484, + 0.596 + ], + "angle": 0, + "content": "We use InternVL2 (2B, 8B and 26B) [9] as the base models and train them with two stages: continual pre-training and fine-tuning. During the pre-training stage, we use all the GUI-Lasagne dataset mentioned in Sec. 3. Different prompts are designed for different training tasks to avoid task confusion. See Sec. 14 for the prompts. We unfreeze the vision encoder, decoder, and MLP layer of InternVL. The learning rate is set to \\(1\\mathrm{e} - 4 / 1\\mathrm{e} - 4 / 5\\mathrm{e} - 5\\) for InternVL-2B/8B/26B, respectively, and the batch size is set to 1024. We get SpiritSight-Base model after pre-training and then fine-tune it on multiple downstream tasks individually. More details are shown in Sec. 12.2 and Sec. 12.3." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.64, + 0.306, + 0.654 + ], + "angle": 0, + "content": "5.2. Benchmarks & metrics" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.674, + 0.484, + 0.901 + ], + "angle": 0, + "content": "To assess SpiritSight's capability in diverse real-world environments, we evaluate SpiritSight on six benchmarks covering various GUI platforms and tasks. Multimodal-Mind2Web [13], ScreenSpot [11], AMEX [5], GUIAct [7], AndroidControl [32], and GUI-Odyssey [38]. For Multimodal-Mind2Web and ScreenSpot, we use the same data pre-processing and metrics as SeeClick [11] uses. For GUIAct, we evaluate SpiritSight on the web-single and web-multi sub-sets and report step success rate (Step SR). For AndroidControl, we evaluate SpiritSight on the High-Level (HL) and Low-Level (LL) tasks and report step accuracy. For AMEX and GUI-Odyssey, we report the action matching score (AMS) defined in AitW [46]. These metrics are similar in that they all measure the single-step accuracy. Refer to Sec. 12.4 for more information about the metrics." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.343, + 0.638, + 0.361 + ], + "angle": 0, + "content": "6. Experiment" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.369, + 0.811, + 0.386 + ], + "angle": 0, + "content": "6.1. Advanced vision-based GUI agent" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.393, + 0.907, + 0.529 + ], + "angle": 0, + "content": "In this section, we compare the performance of SpiritSight with other advanced methods across various input modalities and test configurations on Multimodal-Mind2Web [13], a classic and high-quality benchmark for GUI navigation. The results are shown in Tab. 1. Methods that use top-50 candidates as input perform the best. This is evident, as the assistance of candidate elements can significantly reduce the decision space. However, such methods are not particularly feasible in practice." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.53, + 0.909, + 0.682 + ], + "angle": 0, + "content": "The results indicate that SpiritSight significantly outperforms all methods that do not rely on any candidate element, including vision-based methods, language-based methods, and even vision-language-based methods. This demonstrates strong capabilities of SpiritSight in Web GUI navigation tasks. It should be noted that SpiritSight achieves a significant advantage in the Ele.Acc metric compared to other vision-based methods, which can be attributed to the specially constructed visual grounding training data and the proposed UBP approach." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.695, + 0.825, + 0.712 + ], + "angle": 0, + "content": "6.2. Strong cross-platform compatibility" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.718, + 0.906, + 0.838 + ], + "angle": 0, + "content": "We evaluated SpiritSight on other GUI navigation benchmarks across various platforms and compare it with state-of-the-art (SOTA) GUI agents as shown in Tab. 2. SpiritSight demonstrated the best performance on these benchmarks, showing its strong capabilities across various platforms. OdysseyAgent uses additional historical screenshot images as input, yet SpiritSight achieves comparable results." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.906, + 0.903 + ], + "angle": 0, + "content": "We also evaluated SpiritSight on ScreenSpot, a function grounding benchmark. As shown in Tab. 3, SpiritSight performs well across all three platforms, showcasing its cross-platform capability and robust grounding ability. Notably," + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "29495" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.093, + 0.09, + 0.483, + 0.204 + ], + "angle": 0, + "content": "
AgentOdysseyAMEXAndroidCtrlGUIAct
HighHighHighLowMultiSingle
GPT-4o[42]20.4%-21.2%28.4%-41.8%
Previous SOTA74.3%70.7%64.8%80.0%45.4%74.9%
SpiritSight-2B72.3%74.5%64.9%86.3%45.5%76.0%
SpiritSight-8B75.8%80.7%68.1%87.6%49.3%78.2%
SpiritSight-26B75.9%82.1%69.6%88.2%52.6%78.3%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.227, + 0.483, + 0.312 + ], + "angle": 0, + "content": "Table 2. Results of SpiritSight on four recently published GUI navigation benchmarks. 'High' and 'Multi' indicates high-level tasks, 'Low' and 'Single' indicates single-stepped tasks. The SOTA results of GUI-Odyssey, AMEX, AndroidControl and GUIAct come from OdysseyAgent [38], SPHINX-GUI-Agent [5], fine-tuned PaLM-2S [32] and MiniCPM-GUI [7], respectively." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.338, + 0.483, + 0.445 + ], + "angle": 0, + "content": "SpiritSight models are trained specifically for GUI navigation tasks, and no training data aligned with ScreenSpot is used. There remains room for further improvement in SpiritSight's performance on the ScreenSpot benchmark. We further assess the foundational ability of SpiritSight, the visual grounding ability, on our custom-constructed testing dataset. See Sec. 12.7 for more details." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.454, + 0.483, + 0.487 + ], + "angle": 0, + "content": "6.3. Recognition and grounding as priors for GUI navigation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.493, + 0.483, + 0.703 + ], + "angle": 0, + "content": "To verify the significance of the three levels of GUI-Lasagne data, we progressively removed level-3, level-2, and level-1 data from the training set during the pretraining stage and evaluate SpiritSight-8B on MultimodalMind2Web. The results are shown by the blue line in Fig. 6a. It can be seen that each level of data contributes to improving Step SR. While the tasks in level-1 data differ the most from web navigation task compared to the other two levels, they provide an effective initialization for the pre-trained model. Although the level-3 data is constructed from the mobile scenario, it also assists in web-based GUI navigation tasks. This indicates that the joint learning strategy helps SpiritSight develop strong navigation abilities across different GUI environments with limited resources." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.483, + 0.87 + ], + "angle": 0, + "content": "We also conducted ablation experiments to evaluate the effectiveness of data cleaning and CoT construction strategies on the level-3 data, as shown by the orange line in Fig. 6a. It can be observed that training in a CoT manner effectively improves the model's GUI navigation capabilities, while the impact of data cleaning strategy is relatively small. This may be due to the fact that erroneous samples are relatively tolerated during the pre-training phase of large models. The difference in results of \"w/ CoT\" and \"w/ Level \\(1 + 2 + 3\\) \" is due to the different implementation for fine-tuning stage. See Sec. 12.3 for more details." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.901 + ], + "angle": 0, + "content": "To evaluate the effectiveness of our GUI-Lasagne dataset, we train InternVL-2 (8B) on SeeClick train" + }, + { + "type": "table", + "bbox": [ + 0.529, + 0.089, + 0.895, + 0.25 + ], + "angle": 0, + "content": "
AgentModel SizeScreenSpot
WebMobileDesktop
GPT4V[1]-5.0%7.5%4.6%
Qwen-VL[3]9.6B3.0%7.2%5.4%
Fuyu[4]8B19.2%21.2%18.3%
CogAgent[19]18B49.5%45.5%47.1%
SeeClick[11]9.6B44.1%65.0%51.1%
SpiritSight-2B2B63.6%62.5%61.8%
SpiritSight-8B8B68.3%68.4%62.9%
SpiritSight-26B26B74.4%73.6%69.9%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.272, + 0.905, + 0.3 + ], + "angle": 0, + "content": "Table 3. Results of SpiritSight and other vision-based methods on ScreenSpot Benchmark." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.324, + 0.907, + 0.581 + ], + "angle": 0, + "content": "ing data [11] using full parameter training and LoRA training, resulting in SeeClick(InternVL-Full) and SeeClick(InternVL-LoRA), respectively. We then fine-tune and evaluate the two models on the Multimodal-Mind2Web benchmark. The results are shown by the dashed line in Fig. 6a, where we also present the result of the original SeeClick model. SeeClick(InternVL-LoRA) performs better than the original SeeClick model, which may be attributed to the dynamic high-resolution approach used by InternVL that preserves the details of input images. SeeClick(InternVL-Full) performs the worst, as the scale of the SeeClick training data is not large enough for the model to converge during full parameter training. SpiritSight agent that trained on only GUI-Lasagne level-1 data outperforms all three models trained on SeeClick data, demonstrating the effectiveness of GUI-Lasagne dataset. More ablation studies can be found in Sec. 12.5." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.59, + 0.819, + 0.606 + ], + "angle": 0, + "content": "6.4. Better grounding ability from UBP" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.613, + 0.907, + 0.824 + ], + "angle": 0, + "content": "To verify the effectiveness of UBP on grounding task, we use LoRA for resource efficiency to fine-tune InternVL in 4 different settings (original as baseline, 2D-BPE, UBP, and 2D-BPE&UBP), respectively. We then evaluate these models on Multimodal-Mind2Web benchmark, as shown in Fig. 5b. It can be seen that UBP shows a significant advantage in Ele.Acc compared to baseline, while the results of Op.F1 show little variation across the 4 settings. This indicates that UBP improves the performance of GUI agent primarily by enhancing the grounding ability. Finally, the combination of UBP and 2D-BPE achieves the best results. This indicates that UBP is compatible with 2D-BPE, leading to better performance. More ablation studies can be found in Sec. 12.6." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.833, + 0.86, + 0.849 + ], + "angle": 0, + "content": "6.5. Scaling effects on dataset and model size" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.905, + 0.901 + ], + "angle": 0, + "content": "We explore the impact of pre-training dataset and model size on SpiritSight using Multimodal-Mind2Web benchmark. The results are shown in Fig. 6b. SpiritSight-2B out" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "29496" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.09, + 0.337, + 0.254 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.093, + 0.255, + 0.338, + 0.277 + ], + "angle": 0, + "content": "(a) Effect of three data levels and data augmentation for level-3 data" + }, + { + "type": "image", + "bbox": [ + 0.379, + 0.093, + 0.621, + 0.253 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.378, + 0.255, + 0.623, + 0.278 + ], + "angle": 0, + "content": "(b) Ablation Study: Effect of training data percentages on model performance." + }, + { + "type": "image", + "bbox": [ + 0.658, + 0.093, + 0.903, + 0.253 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.658, + 0.255, + 0.905, + 0.277 + ], + "angle": 0, + "content": "(c) Ablation Study: Effect of data percentages on downstream task." + }, + { + "type": "image_caption", + "bbox": [ + 0.165, + 0.291, + 0.83, + 0.305 + ], + "angle": 0, + "content": "Figure 6. Ablation studies on GUILasagne dataset and scaling effects in the Multimodal-Mind2Web benchmark." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.332, + 0.482, + 0.466 + ], + "angle": 0, + "content": "performs SeeClick [11] by using just 1/8 of the pre-training dataset. This remarkable performance comes from the high quality and grounding-focus nature of GUI-Lasagne data. The performance improves as the size of dataset increases, demonstrating the significance of collecting large-scale data. SpiritSight-2B reaches saturation with a smaller amount of pre-training data, while SpiritSight-26B appears to have further potential for improvement, which aligns with the scaling law of LLMs and VLMs." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.469, + 0.483, + 0.575 + ], + "angle": 0, + "content": "We also evaluate the ability of SpiritSight to transferring to downstream GUI agent tasks. We fine-tune SpiritSight-Base (8B) on various proportions of the Multimodal-Mind2Web training data and show the results in Fig. 6c. It is noticed that SpiritSight achieves \\(36.6\\%\\) Step SR with only 1/8 of the fine-tuning data, showing strong transferability to GUI navigation tasks." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.588, + 0.406, + 0.604 + ], + "angle": 0, + "content": "6.6. Effective transfer to other languages" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.612, + 0.483, + 0.717 + ], + "angle": 0, + "content": "Exploring the cross-lingual capabilities of GUI agents is highly beneficial for their application in non-English environments. We split the training and testing sets of GUIAct (web-multi) data into English and Chinese partitions, respectively. We then fine-tune SpiritSight-Base (8B) on two sets: the entire training set (English&Chinese) and the English-only training set. The results are shown in Tab. 4." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.719, + 0.483, + 0.869 + ], + "angle": 0, + "content": "Under the English&Chinese configuration, SpiritSight achieves comparable results on both the English and Chinese testing sets, despite having fewer Chinese samples in the pre-training data than English ones. Under the English-only configuration, SpiritSight achieves \\(24.5\\%\\) Step SR on the Chinese testing set, reaching half of the English&Chinese performance. The zero-shot capability of SpiritSight in Chinese results from the small but effective foundational Chinese data included in the pre-training stage." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.483, + 0.901 + ], + "angle": 0, + "content": "This experiment offers a paradigm for applying GUI agents to non-English environments: by collecting (1) free" + }, + { + "type": "table", + "bbox": [ + 0.552, + 0.328, + 0.871, + 0.387 + ], + "angle": 0, + "content": "
SFT DataOverallChineseEnglish
English+Chinese49.3%49.3%49.2%
English35.0%24.5%48.6%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.409, + 0.905, + 0.436 + ], + "angle": 0, + "content": "Table 4. Results of SpiritSight on GUIAct (web-multi), trained with different language datasets." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.462, + 0.906, + 0.553 + ], + "angle": 0, + "content": "web and mobile GUI information from the target language environment (level 1 & level 2 data), and (2) a small amount of high-quality GUI navigation data at low cost (level 3 data). With this language transferring strategy, the same capabilities as in the English environment can be achieved in non-English environments with minimal extra costs." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.566, + 0.628, + 0.581 + ], + "angle": 0, + "content": "7. Limitation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.591, + 0.906, + 0.68 + ], + "angle": 0, + "content": "As SpiritSight is a vision-based GUI agent, it constantly requires access to screenshots which may contain personal information or other sensitive data. Users and system providers should carefully manage the system privileges granted to SpiritSight agent to mitigate potential privacy and security risks." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.695, + 0.634, + 0.71 + ], + "angle": 0, + "content": "8. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In this paper, we propose an advanced vision-based end-to-end GUI agent, SpiritSight, with high generalization across multiple GUI platforms. We construct an efficient multilevel, large-scale, high-quality GUI pre-training dataset to equip SpiritSight with robust GUI perception, grounding and understanding capabilities. Additionally, we introduce a UBP method to resolve ambiguity in dynamic high-resolution inputs during model training, further enhancing the ability of SpiritSight to ground GUI objects. As a result, SpiritSight achieves strong performance in numerous GUI navigation benchmarks, demonstrating significant potential for practical deployment in real-world applications." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "29497" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.091, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.184 + ], + "angle": 0, + "content": "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.185, + 0.484, + 0.255 + ], + "angle": 0, + "content": "[2] Gilles Baechler, Srinivas Sunkara, Maria Wang, Fedir Zubach, Hassan Mansoor, Vincent Etter, Victor Carbune, Jason Lin, Jindong Chen, and Abhanshu Sharma. Screenai: A vision-language model for ui and infographics understanding. arXiv preprint arXiv:2402.04615, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.256, + 0.483, + 0.324 + ], + "angle": 0, + "content": "[3] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966, 2023.7, 1,4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.326, + 0.483, + 0.368 + ], + "angle": 0, + "content": "[4] Rohan Bavishi, Erich Elsen, Curtis Hawthorne, Maxwell Nye, Augustus Odena, Arushi Somani, and Săgnak Taşürlar. Introducing our multimodal models, 2023. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.369, + 0.483, + 0.436 + ], + "angle": 0, + "content": "[5] Yuxiang Chai, Siyuan Huang, Yazhe Niu, Han Xiao, Liang Liu, Dingyu Zhang, Peng Gao, Shuai Ren, and Hongsheng Li. Amex: Android multi-annotation expo dataset for mobile gui agents. arXiv preprint arXiv:2407.17490, 2024. 2, 5, 6, 7, 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.439, + 0.483, + 0.506 + ], + "angle": 0, + "content": "[6] Dongping Chen, Yue Huang, Siyuan Wu, Jingyu Tang, Liuyi Chen, Yilin Bai, Zhigang He, Chenlong Wang, Huichi Zhou, Yiqiang Li, et al. Gui-world: A dataset for gui-oriented multimodal llm-based agents. arXiv preprint arXiv:2406.10819, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.509, + 0.483, + 0.577 + ], + "angle": 0, + "content": "[7] Wentong Chen, Junbo Cui, Jinyi Hu, Yujia Qin, Junjie Fang, Yue Zhao, Chongyi Wang, Jun Liu, Guirong Chen, Yupeng Huo, et al. Guicourse: From general vision language models to versatile gui agents. arXiv preprint arXiv:2406.11317, 2024. 1, 3, 5, 6, 7, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.579, + 0.483, + 0.634 + ], + "angle": 0, + "content": "[8] Xingyu Chen, Zihan Zhao, Lu Chen, Danyang Zhang, Jiabao Ji, Ao Luo, Yuxuan Xiong, and Kai Yu. Websrc: A dataset for web-based structural reading comprehension. arXiv preprint arXiv:2101.09465, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.636, + 0.483, + 0.704 + ], + "angle": 0, + "content": "[9] Zhe Chen, Weiyun Wang, Hao Tian, Shenglong Ye, Zhangwei Gao, Erfei Cui, Wenwen Tong, Kongzhi Hu, Jiapeng Luo, Zheng Ma, et al. How far are we to gpt-4v? closing the gap to commercial multimodal models with open-source suites. arXiv preprint arXiv:2404.16821, 2024. 2, 5, 6, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.705, + 0.483, + 0.789 + ], + "angle": 0, + "content": "[10] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24185–24198, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.79, + 0.483, + 0.845 + ], + "angle": 0, + "content": "[11] Kanzhi Cheng, Qiushi Sun, Yougang Chu, Fangzhi Xu, Yantao Li, Jianbing Zhang, and Zhiyong Wu. Seeclick: Harnessing gui grounding for advanced visual gui agents. arXiv preprint arXiv:2401.10935, 2024. 1, 3, 4, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[12] Biplab Deka, Zifeng Huang, Chad Franzen, Joshua Hibschman, Daniel Afergan, Yang Li, Jeffrey Nichols, and Ranjitha Kumar. Rico: A mobile app dataset for building data-driven design applications. In Proceedings of the 30th annual" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "ACM symposium on user interface software and technology, pages 845-854, 2017. 5, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.176 + ], + "angle": 0, + "content": "[13] Xiang Deng, Yu Gu, Boyuan Zheng, Shijie Chen, Sam Stevens, Boshi Wang, Huan Sun, and Yu Su. Mind2web: Towards a generalist agent for the web. Advances in Neural Information Processing Systems, 36, 2024. 1, 2, 6, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.178, + 0.905, + 0.217 + ], + "angle": 0, + "content": "[14] Jacob Devlin. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.219, + 0.905, + 0.286 + ], + "angle": 0, + "content": "[15] Hiroki Furuta, Kuang-Huei Lee, Ofir Nachum, Yutaka Matsuo, Aleksandra Faust, Shixiang Shane Gu, and Izzeddin Gur. Multimodal web navigation with instruction-finetuned foundation models. arXiv preprint arXiv:2305.11854, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.289, + 0.905, + 0.316 + ], + "angle": 0, + "content": "[16] Common crawl Group. Common crawl - open repository of web crawl data, 2024. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.317, + 0.905, + 0.384 + ], + "angle": 0, + "content": "[17] Izzeddin Gur, Hiroki Furuta, Austin Huang, Mustafa Safdari, Yutaka Matsuo, Douglas Eck, and Aleksandra Faust. A real-world webagent with planning, long context understanding, and program synthesis. arXiv preprint arXiv:2307.12856, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.386, + 0.905, + 0.442 + ], + "angle": 0, + "content": "[18] Markus Hafner, Maria Katsantoni, Tino Köster, James Marks, Joyita Mukherjee, Dorothee Staiger, Jernej Ule, and Mihaela Zavolan. Clip and complementary methods. Nature Reviews Methods Primers, 1(1):1-23, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.442, + 0.905, + 0.524 + ], + "angle": 0, + "content": "[19] Wenyi Hong, Weihan Wang, Qingsong Lv, Jiazheng Xu, Wenmeng Yu, Junhui Ji, Yan Wang, Zihan Wang, Yuxiao Dong, Ming Ding, et al. Cogagent: A visual language model for gui agents. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14281-14290, 2024. 1, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.526, + 0.905, + 0.567 + ], + "angle": 0, + "content": "[20] Yu-Chung Hsiao, Fedir Zubach, Maria Wang, et al. Screenqa: Large-scale question-answer pairs over mobile app screenshots. arXiv preprint arXiv:2209.08199, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.568, + 0.905, + 0.622 + ], + "angle": 0, + "content": "[21] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.623, + 0.905, + 0.678 + ], + "angle": 0, + "content": "[22] Faria Huq, Jeffrey P Bigham, and Nikolas Martelaro. “what’s important here?”: Opportunities and challenges of using llms in retrieving information from web interfaces. arXiv preprint arXiv:2312.06147, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.679, + 0.905, + 0.747 + ], + "angle": 0, + "content": "[23] Yang Jin, Kun Xu, Kun Xu, Liwei Chen, Chao Liao, Jianchao Tan, Quzhe Huang, Bin Chen, Chenyi Lei, An Liu, et al. Unified language-vision pretraining in llm with dynamic discrete visual tokenization. arxiv 2024. arXiv preprint arXiv:2309.04669, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.749, + 0.905, + 0.817 + ], + "angle": 0, + "content": "[24] Jihyung Kil, Chan Hee Song, Boyuan Zheng, Xiang Deng, Yu Su, and Wei-Lun Chao. Dual-view visual contextualization for web navigation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14445-14454, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.818, + 0.905, + 0.899 + ], + "angle": 0, + "content": "[25] Geewook Kim, Teakgyu Hong, Moonbin Yim, Jeong Yeon Nam, Jinyoung Park, Jinyeong Yim, Wonseok Hwang, Sang-doo Yun, Dongyoon Han, and Seunghyun Park. Ocr-free document understanding transformer. In European Conference on Computer Vision, pages 498-517. Springer, 2022. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "29498" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.161 + ], + "angle": 0, + "content": "[26] Hanyu Lai, Xiao Liu, Iat Long Iong, Shuntian Yao, Yuxuan Chen, Pengbo Shen, Hao Yu, Hanchen Zhang, Xiaohan Zhang, Yuxiao Dong, et al. Autowebglm: Bootstrap and reinforce a large language model-based web navigating agent. arXiv preprint arXiv:2404.03648, 2024. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.162, + 0.482, + 0.245 + ], + "angle": 0, + "content": "[27] Kenton Lee, Mandar Joshi, Iulia Raluca Turc, Hexiang Hu, Fangyu Liu, Julian Martin Eisenschlos, Urvashi Khandelwal, Peter Shaw, Ming-Wei Chang, and Kristina Toutanova. Pix2struct: Screenshot parsing as pretraining for visual language understanding. In International Conference on Machine Learning, pages 18893–18912. PMLR, 2023. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.246, + 0.482, + 0.3 + ], + "angle": 0, + "content": "[28] Kuang-Huei Lee, Xinyun Chen, Hiroki Furuta, John Canny, and Ian Fischer. A human-inspired reading agent with gist memory of very long contexts. arXiv preprint arXiv:2402.09727, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.301, + 0.482, + 0.357 + ], + "angle": 0, + "content": "[29] Feng Li, Renrui Zhang, Hao Zhang, Yuanhan Zhang, Bo Li, Wei Li, Zejun Ma, and Chunyuan Li. Llava next-interleave: Tackling multi-image, video, and 3d in large multimodal models. arXiv preprint arXiv:2407.07895, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.358, + 0.482, + 0.426 + ], + "angle": 0, + "content": "[30] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International conference on machine learning, pages 12888-12900. PMLR, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.427, + 0.482, + 0.495 + ], + "angle": 0, + "content": "[31] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730-19742. PMLR, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.497, + 0.482, + 0.552 + ], + "angle": 0, + "content": "[32] Wei Li, William Bishop, Alice Li, Chris Rawles, Folawiyo Campbell-Ajala, Divya Tyamagundlu, and Oriana Riva. On the effects of data scale on computer control agents. arXiv preprint arXiv:2406.03679, 2024. 6, 7, 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.553, + 0.482, + 0.608 + ], + "angle": 0, + "content": "[33] Yang Li, Gang Li, Luheng He, Jingjie Zheng, Hong Li, and Zhiwei Guan. Widget captioning: Generating natural language description for mobile user interface elements. arXiv preprint arXiv:2010.04295, 2020. 5, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.608, + 0.482, + 0.676 + ], + "angle": 0, + "content": "[34] Zeyi Liao, Lingbo Mo, Chejian Xu, Mintong Kang, Jiawei Zhang, Chaowei Xiao, Yuan Tian, Bo Li, and Huan Sun. Eia: Environmental injection attack on generalist web agents for privacy leakage. arXiv preprint arXiv:2409.11295, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.678, + 0.482, + 0.747 + ], + "angle": 0, + "content": "[35] Ziyi Lin, Chris Liu, Renrui Zhang, Peng Gao, Longtian Qiu, Han Xiao, Han Qiu, Chen Lin, Wenqi Shao, Keqin Chen, et al. Sphinx: The joint mixing of weights, tasks, and visual embeddings for multi-modal large language models. arXiv preprint arXiv:2311.07575, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.748, + 0.482, + 0.803 + ], + "angle": 0, + "content": "[36] Evan Zheran Liu, Kelvin Guu, Panupong Pasupat, Tianlin Shi, and Percy Liang. Reinforcement learning on web interfaces using workflow-guided exploration. arXiv preprint arXiv:1802.08802, 2018. 2, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.804, + 0.482, + 0.831 + ], + "angle": 0, + "content": "[37] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[38] Quanfeng Lu, Wenqi Shao, Zitao Liu, Fanqing Meng, Boxuan Li, Botong Chen, Siyuan Huang, Kaipeng Zhang, Yu Qiao, and Ping Luo. Gui odyssey: A comprehensive dataset for cross-app gui navigation on mobile devices. arXiv preprint arXiv:2406.08451, 2024. 2, 6, 7, 1, 4" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.134 + ], + "angle": 0, + "content": "[39] Xing Han Lu, Zdenek Kasner, and Siva Reddy. Weblinx: Real-world website navigation with multi-turn dialogue. arXiv preprint arXiv:2402.05930, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.907, + 0.191 + ], + "angle": 0, + "content": "[40] Minesh Mathew, Dimosthenis Karatzas, and CV Jawahar. Docvqa: A dataset for vqa on document images. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 2200-2209, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.192, + 0.905, + 0.261 + ], + "angle": 0, + "content": "[41] Erik Nijkamp, Bo Pang, Hiroaki Hayashi, Lifu Tu, Huan Wang, Yingbo Zhou, Silvio Savarese, and Caiming Xiong. Codegen: An open large language model for code with multi-turn program synthesis. arXiv preprint arXiv:2203.13474, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.263, + 0.735, + 0.277 + ], + "angle": 0, + "content": "[42] OpenAI. Hello gpt-4o, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.279, + 0.905, + 0.333 + ], + "angle": 0, + "content": "[43] Liangming Pan, Alon Albalak, Xinyi Wang, and William Yang Wang. Logic-lm: Empowering large language models with symbolic solvers for faithful logical reasoning. arXiv preprint arXiv:2305.12295, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.335, + 0.905, + 0.377 + ], + "angle": 0, + "content": "[44] Alec Radford, Karthik Narasimhan, Tim Salimans, Ilya Sutskever, et al. Improving language understanding by generative pre-training. OpenAI, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.378, + 0.905, + 0.447 + ], + "angle": 0, + "content": "[45] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of machine learning research, 21(140):1-67, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.449, + 0.905, + 0.505 + ], + "angle": 0, + "content": "[46] Christopher Rawles, Alice Li, Daniel Rodriguez, Oriana Riva, and Timothy Lillicrap. Androidinthewild: A large-scale dataset for android device control. Advances in Neural Information Processing Systems, 36, 2024. 2, 4, 6, 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.506, + 0.905, + 0.573 + ], + "angle": 0, + "content": "[47] Baptiste Roziere, Jonas Gehring, Fabian Gloeckle, Sten Sootla, Itai Gat, Xiaqing Ellen Tan, Yossi Adi, Jingyu Liu, Romain Sauvestre, Tal Remez, et al. Code llama: Open foundation models for code. arXiv preprint arXiv:2308.12950, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.576, + 0.905, + 0.616 + ], + "angle": 0, + "content": "[48] Rico Sennrich. Improving neural machine translation models with monolingual data. arXiv preprint arXiv:1511.06709, 2015. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.619, + 0.905, + 0.688 + ], + "angle": 0, + "content": "[49] Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Mingchuan Zhang, YK Li, Yu Wu, and Daya Guo. Deepseekmath: Pushing the limits of mathematical reasoning in open language models. arXiv preprint arXiv:2402.03300, 2024.1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.689, + 0.905, + 0.772 + ], + "angle": 0, + "content": "[50] Peter Shaw, Mandar Joshi, James Cohan, Jonathan Berant, Panupong Pasupat, Hexiang Hu, Urvashi Khandelwal, Kenton Lee, and Kristina N Toutanova. From pixels to ui actions: Learning to follow instructions via graphical user interfaces. Advances in Neural Information Processing Systems, 36:34354-34370, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.774, + 0.905, + 0.831 + ], + "angle": 0, + "content": "[51] Tianlin Shi, Andrej Karpathy, Linxi Fan, Jonathan Hernandez, and Percy Liang. World of bits: An open-domain platform for web-based agents. In International Conference on Machine Learning, pages 3135-3144. PMLR, 2017. 2, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.832, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[52] Lucas-Andrei Thil, Mirela Popa, and Gerasimos Spanakis. Navigating webai: Training agents to complete web tasks with large language models and reinforcement learning. In Proceedings of the 39th ACM/SIGAPP Symposium on Applied Computing, pages 866–874, 2024. 1, 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "29499" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.176 + ], + "angle": 0, + "content": "[53] Jianqiang Wan, Sibo Song, Wenwen Yu, Yuliang Liu, Wenqing Cheng, Fei Huang, Xiang Bai, Cong Yao, and Zhibo Yang. Omniparser: A unified framework for text spotting key information extraction and table recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15641-15653, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.178, + 0.482, + 0.247 + ], + "angle": 0, + "content": "[54] Bryan Wang, Gang Li, Xin Zhou, Zhourong Chen, Tovi Grossman, and Yang Li. Screen2words: Automatic mobile ui summarization with multimodal learning. In The 34th Annual ACM Symposium on User Interface Software and Technology, pages 498-510, 2021. 5, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.249, + 0.482, + 0.318 + ], + "angle": 0, + "content": "[55] Junyang Wang, Haiyang Xu, Haitao Jia, Xi Zhang, Ming Yan, Weizhou Shen, Ji Zhang, Fei Huang, and Jitao Sang. Mobile-agent-v2: Mobile device operation assistant with effective navigation via multi-agent collaboration. arXiv preprint arXiv:2406.01014, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.32, + 0.482, + 0.375 + ], + "angle": 0, + "content": "[56] Junyang Wang, Haiyang Xu, Jiabo Ye, Ming Yan, Weizhou Shen, Ji Zhang, Fei Huang, and Jitao Sang. Mobile-agent: Autonomous multi-modal mobile device agent with visual perception. arXiv preprint arXiv:2401.16158, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.377, + 0.482, + 0.445 + ], + "angle": 0, + "content": "[57] Ke Wang, Houxing Ren, Aojun Zhou, Zimu Lu, Sichun Luo, Weikang Shi, Renrui Zhang, Linqi Song, Mingjie Zhan, and Hongsheng Li. Mathcoder: Seamless code integration in llms for enhanced mathematical reasoning. arXiv preprint arXiv:2310.03731, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.447, + 0.482, + 0.503 + ], + "angle": 0, + "content": "[58] Weihan Wang, Qingsong Lv, Wenmeng Yu, Wenyi Hong, Ji Qi, Yan Wang, Junhui Ji, Zhuoyi Yang, Lei Zhao, Xixuan Song, et al. Cogvlm: Visual expert for pretrained language models. arXiv preprint arXiv:2311.03079, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.505, + 0.482, + 0.572 + ], + "angle": 0, + "content": "[59] Jason Wei, Yi Tay, Rishi Bommasani, Colin Raffel, Barret Zoph, Sebastian Borgeaud, Dani Yogatama, Maarten Bosma, Denny Zhou, Donald Metzler, et al. Emergent abilities of large language models. arXiv preprint arXiv:2206.07682, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.576, + 0.482, + 0.645 + ], + "angle": 0, + "content": "[60] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824-24837, 2022. 5, 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.647, + 0.482, + 0.7 + ], + "angle": 0, + "content": "[61] Chen Henry Wu, Jing Yu Koh, Ruslan Salakhutdinov, Daniel Fried, and Aditi Raghunathan. Adversarial attacks on multimodal agents. arXiv preprint arXiv:2406.12814, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.704, + 0.482, + 0.771 + ], + "angle": 0, + "content": "[62] Shijie Wu, Ozan Irsoy, Steven Lu, Vadim Dabravolski, Mark Dredze, Sebastian Gehrmann, Prabhanjan Kambadur, David Rosenberg, and Gideon Mann. Bloomberg: A large language model for finance. arXiv preprint arXiv:2303.17564, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.775, + 0.482, + 0.843 + ], + "angle": 0, + "content": "[63] Yichong Xu, Chenguang Zhu, Shuohang Wang, Siqi Sun, Hao Cheng, Xiaodong Liu, Jianfeng Gao, Pengcheng He, Michael Zeng, and Xuedong Huang. Human parity on commonsenseqa: Augmenting self-attention with external attention. arXiv preprint arXiv:2112.03254, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.482, + 0.899 + ], + "angle": 0, + "content": "[64] Zhao Yang, Jiaxuan Liu, Yucheng Han, Xin Chen, Zebiao Huang, Bin Fu, and Gang Yu. Appagent: Multimodal agents as smartphone users. arXiv preprint arXiv:2312.13771, 2023. 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.161 + ], + "angle": 0, + "content": "[65] Shunyu Yao, Howard Chen, John Yang, and Karthik Narasimhan. Webshop: Towards scalable real-world web interaction with grounded language agents. Advances in Neural Information Processing Systems, 35:20744-20757, 2022. 2, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.164, + 0.905, + 0.219 + ], + "angle": 0, + "content": "[66] Yuan Yao, Tianyu Yu, Ao Zhang, Chongyi Wang, Junbo Cui, Hongji Zhu, Tianchi Cai, Haoyu Li, Weilin Zhao, Zhihui He, et al. Minicpm-v: A gpt-4v level mllm on your phone. arXiv preprint arXiv:2408.01800, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.221, + 0.905, + 0.289 + ], + "angle": 0, + "content": "[67] Jiabo Ye, Anwen Hu, Haiyang Xu, Qinghao Ye, Ming Yan, Guohai Xu, Chenliang Li, Junfeng Tian, Qi Qian, Ji Zhang, et al. Ureader: UniversalOCR-free visually-situated language understanding with multimodal large language model. arXiv preprint arXiv:2310.05126, 2023. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.291, + 0.905, + 0.373 + ], + "angle": 0, + "content": "[68] Da Yin, Faeze Brahman, Abhilasha Ravichander, Khyathi Chandu, Kai-Wei Chang, Yejin Choi, and Bill Yuchen Lin. Agent lumos: Unified and modular training for open-source language agents. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 12380–12403, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.375, + 0.905, + 0.443 + ], + "angle": 0, + "content": "[69] Huaiyuan Ying, Shuo Zhang, Linyang Li, Zhejian Zhou, Yunfan Shao, Zhaoye Fei, Yichuan Ma, Jiawei Hong, Kuikun Liu, Ziyi Wang, et al. Internlm-math: Open math large language models toward verifiable reasoning. arXiv preprint arXiv:2402.06332, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.446, + 0.905, + 0.513 + ], + "angle": 0, + "content": "[70] Longhui Yu, Weisen Jiang, Han Shi, Jincheng Yu, Zhengying Liu, Yu Zhang, James T Kwok, Zhenguo Li, Adrian Weller, and Weiyang Liu. Metamath: Bootstrap your own mathematical questions for large language models. arXiv preprint arXiv:2309.12284, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.516, + 0.905, + 0.571 + ], + "angle": 0, + "content": "[71] Qiusi Zhan, Zhixiang Liang, Zifan Ying, and Daniel Kang. Injecagent: Benchmarking indirect prompt injections in tool-integrated large language model agents. arXiv preprint arXiv:2403.02691, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.573, + 0.905, + 0.627 + ], + "angle": 0, + "content": "[72] Jiwen Zhang, Jihao Wu, Yihua Teng, Minghui Liao, Nuo Xu, Xiao Xiao, Zhongyu Wei, and Duyu Tang. Android in the zoo: Chain-of-action-thought for gui agents. arXiv preprint arXiv:2403.02713, 2024. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.63, + 0.905, + 0.74 + ], + "angle": 0, + "content": "[73] Pan Zhang, Xiaoyi Dong, Yuhang Zang, Yuhang Cao, Rui Qian, Lin Chen, Qipeng Guo, Haodong Duan, Bin Wang, Linke Ouyang, Songyang Zhang, Wenwei Zhang, Yining Li, Yang Gao, Peng Sun, Xinyue Zhang, Wei Li, Jingwen Li, Wenhai Wang, Hang Yan, Conghui He, Xingcheng Zhang, Kai Chen, Jifeng Dai, Yu Qiao, Dahua Lin, and Jiaqi Wang. Internlm-xcomposer-2.5: A versatile large vision language model supporting long-contextual input and output, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.742, + 0.905, + 0.782 + ], + "angle": 0, + "content": "[74] Boyuan Zheng, Boyu Gou, Jihyung Kil, Huan Sun, and Yu Su. Gpt-4v (isdiction) is a generalist web agent, if grounded. arXiv preprint arXiv:2401.01614, 2024. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.784, + 0.905, + 0.839 + ], + "angle": 0, + "content": "[75] Longtao Zheng, Rundong Wang, Xinrun Wang, and Bo An. Synapse: Trajectory-as-exemplar prompting with memory for computer control. In The Twelfth International Conference on Learning Representations, 2023. 1" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.839 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "29500" + } + ] +] \ No newline at end of file diff --git a/2025/SpiritSight Agent_ Advanced GUI Agent with One Look/c38ba9da-d4d0-4e2a-a186-fdd0cc0687b1_origin.pdf b/2025/SpiritSight Agent_ Advanced GUI Agent with One Look/c38ba9da-d4d0-4e2a-a186-fdd0cc0687b1_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..1a868c0e949259a7b7211a226bd38fa29721bb4a --- /dev/null +++ b/2025/SpiritSight Agent_ Advanced GUI Agent with One Look/c38ba9da-d4d0-4e2a-a186-fdd0cc0687b1_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a9d39e3537b3341a496cf9706899228d18b5f726bcd6c2f5fdef14c7ae2ac54b +size 7432823 diff --git a/2025/SpiritSight Agent_ Advanced GUI Agent with One Look/full.md b/2025/SpiritSight Agent_ Advanced GUI Agent with One Look/full.md new file mode 100644 index 0000000000000000000000000000000000000000..6e67a7e55bb48ea802574b82ae10d64f1a353844 --- /dev/null +++ b/2025/SpiritSight Agent_ Advanced GUI Agent with One Look/full.md @@ -0,0 +1,346 @@ +![](images/f6471b3df2f89a2cd7883fa6ed05731ecd52be93cc5193613bdebf5724d4fd70.jpg) + +# SpiritSight Agent: Advanced GUI Agent with One Look + +Zhiyuan Huang $^{1}$ Ziming Cheng $^{1,2*}$ Junting Pan $^{3}$ Zhaohui Hou $^{1}$ Mingjie Zhan $^{1}$ + +$^{1}$ SenseTime Research, $^{2}$ Beijing University of Posts and Telecommunications, $^{3}$ MMLab, CUHK + +{huangzhiyuan, houzhaohui, zhanmingjie}@sensetime.com + +harrytheorange@bupt.edu.cn junting.pa@gmail.com + +# Abstract + +Graphical User Interface (GUI) agents demonstrate promising potential in assisting human-computer interaction, automating human user's navigation on digital devices. An ideal GUI agent is expected to achieve high accuracy, low latency, and compatibility for different GUI platforms. Recent vision-based approaches have shown promise by leveraging advanced Vision Language Models (VLMs). While they generally meet the requirements of compatibility and low latency, these vision-based GUI agents tend to have low accuracy due to their limitations in element grounding. To address this issue, we propose SpiritSight, a vision-based, end-to-end GUI agent that excels in GUI navigation tasks across various GUI platforms. First, we create a multi-level, large-scale, high-quality GUI dataset called GUI-Lasagne using scalable methods, empowering SpiritSight with robust GUI understanding and grounding capabilities. Second, we introduce the Universal Block Parsing (UBP) method to resolve the ambiguity problem inherited from the dynamic resolution strategy, further enhancing SpiritSight's ability to ground GUI objects. Through these efforts, SpiritSight agent outperforms other advanced methods on diverse GUI benchmarks, demonstrating its superior capability and compatibility in GUI navigation tasks. The models and code will be made available upon publication. + +# 1. Introduction + +Graphical User Interface (GUI) automation has long been pursued by people along with the development of modern digital devices. Thanks to recent advances in Large Language Models (LLMs), GUI agents are constructed to assist users in interacting with graphical interfaces, automatically making action decisions based on observations of environ- + +![](images/b75cfff39780575a3f479d6e58c42d4b8a4ff6db0da1f119a00fe244d844c1dc.jpg) +Figure 1. Our SpiritSight agent achieves new state-of-the-art (SOTA) performance across various benchmarks in web, mobile, and desktop scenarios. + +ment and user's instruction. + +Current approaches can be divided into three categories based on their input modalities. Language-based and vision-based approaches use Hyper Text Markup Language (HTML) or Extensible Markup Language (XML) and screenshots as input [7, 11, 13, 19, 22, 26, 28, 53, 68, 75], respectively, while vision-language-based approaches combine both HTML and screenshots as input [15, 24, 52, 74]. + +The language-based and vision-language-based methods typically applied only in the web domain, and are often limited by the excessive length of HTML or security concerns [34, 61, 71]. The vision-based approaches exhibit compatibility across various GUI platforms, as acquiring screenshots is generally easier than obtaining hierarchical data from non-web platforms. + +However, vision-based approaches struggle to ground GUI elements (e.g. buttons, text boxes) from the visual + +input [74]. Some works solve this problem by introducing additional tools, such as optical character recognition (OCR) and icon recognition models, to convert the visual grounding task into a language QA task [55, 56]. This may increase the complexity and inference latency of the agent system. Others attempt to collect large-scale training data through manual synthesis [27, 36, 51] or human annotation [5, 6, 13, 38, 39, 46, 65], while these data are either unrealistic or costly. Additionally, recent advanced Vision Language Models (VLMs) [9, 67] adapt dynamic high-resolution [9, 67] strategy to better suit high-resolution inputs, which we find may introduce ambiguity to the model learning process in GUI scenarios. + +To address the aforementioned challenges, we propose an end-to-end, vision-based GUI agent—SpiritSight, which has strong ability for GUI navigation task. Our contributions are summarized as follows: + +Firstly, we propose GUI-Lasagne, a multi-level, large-scale and high-quality GUI dataset composed of 5.73 million samples to enhance our model's GUI understanding and grounding capabilities. The dataset is collected from the real world and filtered through carefully designed rules to ensure data quality. It is also constructed hierarchically and consists of three levels of components. The first two parts constitute $90\%$ of the total dataset and are collected for free, thus significantly reducing the data collection cost. + +Secondly, we introduce a Universal Block Parsing (UBP) method to resolve the ambiguity problem inherited from the dynamic resolution strategy. With the effectiveness of UBP, SpiritSight gains an improved elements grounding capability and achieves enhanced performance in GUI navigation tasks. + +Thirdly, we release and evaluate SpiritSight agent in various GUI benchmarks and it exhibits remarkable performance among them. SpiritSight agent is pre-trained on GUI-Lasagne and employs UBP to ground elements, as shown in Fig. 3. We release three versions of SpiritSight with different model size: SpiritSight-26B, SpiritSight-8B, and SpiritSight-2B. The standard SpiritSight-8B consistently outperforms previous state-of-the-art methods across various GUI benchmarks, as shown in Fig. 1. On the Multimodal-Mind2Web [13] benchmark, the SpiritSight series outperforms all kinds of methods—including language-based, vision-based and even vision-language-based methods—in the non-candidate setting, as shown in Fig. 2. + +# 2. Related work + +# 2.1. Language-based and vision-language-based GUI agent + +Several works leverage the capabilities of large-scale language models (LLMs) to construct GUI agents. It is + +![](images/1c61129ecf89b1b09352b18720a18fa0f35f80a7644c09b4ab3d4bf137012067.jpg) +Figure 2. Comparison of the average step success rate on Multimodal-Mind2Web benchmark of our SpiritSight agent of three sizes (2B, 8B, 26B) with various previous methods. + +![](images/702443052d2c731f186e652a64491d001cbb7d776c03265819be8e16b6a5d74d.jpg) +Figure 3. The overview of our SpiritSight agent. Our proposed GUI-Lasagne dataset equips SpiritSight with three levels of comprehensive GUI knowledge. Additionally, we introduce a Universal Block Parsing (UBP) method to enhance SpiritSight's grounding capabilities. + +noticed that they are mostly multi-stage architectures. Mind2Web [13] utilizes a lightweight language model to extract candidate elements from HTML, followed by a ranking model that sorts these elements based on task descriptions and historical actions. Finally, a large language model predicts the current actions and selects the target elements to operate on. WebAgent [17] first uses an encoder-decoder model to generate low-level instructions and relevant HTML code snippets, then uses another decoder to produce executable Python code. AutoWebGLM [26] simplifies HTML code through manually designed rules before predicting the executable action codes. + +Other vision-language-based works leverage both GUI screenshots and hierarchical HTML/XML to enhance the + +
Model SizeInput ModalitySelect From TopCross-TaskCross-WebsiteCross-Domain
Ele.AccOp.F1Step SREle.AccOp.F1Step SREle.AccOp.F1Step SR
AutoWebGLM [26]6BText✓--66.4%--56.4%--55.8%
LLaMA2-7B [26]7BText✓--52.7%--47.1%--50.3%
CogAgent [19]18BImage✓--62.3%--54.0%--59.4%
HTML-T5-XL [17]3BText✓76.4%78.8%71.5%68.4%71.0%62.2%73.0%75.6%67.1%
SeeAct* [74]-Text+Image×46.4%73.4%40.2%38.0%67.8%32.4%42.4%69.3%36.8%
ReadAgent-P* [28]340BText×33.7%72.5%29.2%37.4%75.1%31.1%37.2%76.3%33.4%
MiniCPM-GUI [7]3BImage×23.8%86.8%20.8%20.3%81.7%17.3%17.9%74.5%14.6%
Fuyu-GUI [4]8BImage×19.1%86.1%15.6%13.9%80.7%12.2%14.2%83.1%11.7%
SeeClick [11]9.6BImage×28.3%87.0%25.5%21.4%80.6%16.4%23.2%84.8%20.8%
OmniParser* [53]-Image×42.4%87.6%39.4%41.0%84.8%36.5%45.5%85.7%42.0%
SpiritSight-2B2BImage×51.7%87.2%44.9%44.0%83.6%37.8%42.4%83.5%36.9%
SpiritSight-8B8BImage×59.2%88.9%52.7%52.2%84.7%44.0%50.1%86.0%44.4%
SpiritSight-26B26BImage×60.5%89.7%54.7%57.0%85.7%48.1%54.1%87.2%49.2%
+ +Table 1. Comparison of SpiritSight agent with various previous methods on Multimodal-Mind2Web benchmark. SpiritSight significantly outperforms all methods that do not rely on any candidate element, including vision-based methods, language-based methods, and even vision-language-based methods. The asterisk (*) in the upper-right indicates that the model is not fine-tuned on the downstream tasks. + +robustness of GUI agents. WebGUM [15] and CC-Net [52] use ResNet and ViT to extract features from screenshots, respectively. The image embeddings are then combined with text embeddings and fed into a multi-modal transformer. SeeAct [74] and AppAgent [64] identify all interactive elements using HTML or XML data. They then assign each interactive element a unique identifier in the screenshot before feeding the screenshot into the model. + +These language-based and vision-language-based methods, which rely on hierarchical information, exhibit several limitations: (1) Hierarchical representations, such as HTML or XML, are not consistently available across different platforms. And even this information is available, differences in their internal rules make language-based GUI agents less compatible; (2) HTML often contains redundant information, requiring additional models or extensive manually crafted rules for effective filtering. (3) Text-based GUI agents are susceptible to injection attacks [34, 61, 71], where malicious instructions hidden in HTML can easily lead to erroneous or unsafe actions. + +# 2.2. Vision-based GUI agent + +Recently, some vision-based approaches have been proposed to overcome the drawbacks of language-based methods. Some of them [2, 11, 19, 50] are single-stage methods that only use GUI screenshots as input for VLMs and output the next action in an end-to-end manner. However, these agents perform worse on GUI benchmarks compared to other methods. Two primary reasons may be the relatively low input resolution of their base models and the insufficient scale and suboptimal quality of the training data. SeeAct [74], MobileAgent [56] and MobileAgent-v2 [55] are two-stage methods. They use GPT-4V instead of open + +source VLMs as base models and find that even the top models struggle with element grounding. Consequently, they introduce additional OCR and icon recognition models to assist with element grounding, which may increase the complexity and inference latency of the agent system. Overall, the current VLMs exhibit limited ability in handling GUI grounding tasks, thus constrains the navigation capabilities of single-stage vision-based GUI agents. Related works on base models are discussed in Sec. 9. + +# 3. Data collection + +In this section, we present a cost-effective data collection strategy designed to construct a multi-level, large-scale and high-quality GUI dataset, called GUI-Lasagne, as shown in Fig. 4. This dataset helps equip our models with robust abilities in GUI understanding, grounding, and navigation. More details of the collection can be seen in Tab. 8. + +# 3.1. Level one: visual-text alignment + +Visual-text alignment refers to the foundational ability of a GUI agent to perceive the visual appearance of the GUI, specifically the ability to recognize or locate text/icon elements. We collect visual-text data directly from the GUI source data. In the web scenario, we obtain web URLs from CommonCrawl [16] and website ranking sources12. We then develop a data collection tool using Playwright3 to gather real-world web data from the collected URLs. With this tool, we collect 755K web-page screenshots alone with their DOM data, featuring diversity in both resolution + +![](images/0927b7c078b46c951d58911de6fd71473ddd4b354468d91ba93747d2cced5ce5.jpg) + +![](images/2564866784a69ce115cdfd5f7a88f4490320c4b703aaf837dfd8169147627e27.jpg) + +![](images/774c4ba630de293c2b7f03ca4b633358f6d9101666ec5a6833fd574cd95b8a47.jpg) + +![](images/bae001d24c286e962856c041e62532013e97a9e312903b28958e301856b9f47a.jpg) + +![](images/87b3fde8e17abff1c0705f8b1dda6c3b9ca815b7e2d1b2ab380d1840b7fd35f7.jpg) +Figure 4. The collection pipeline of our GUILasagne dataset. The left, middle and right parts show the construction of level-1, level-2, and level-3 data, respectively. + +![](images/0b0aa35a26767ccecb81fe588da766a58912b3e01bb0615d80aae061d27feb8d.jpg) + +![](images/e545eec66bfd52478f2e4021d4e397d473917c19ef7317eb5798852a8a267c77.jpg) + +![](images/df77e095b408b7bd199e9f9e8e2f178bbca0cffc95922f945c97cbc56c1963df.jpg) + +and language. Additionally, we develop an icon recognition model called InternVL-Icon as an icon annotation tool, by constructing a dataset of 30K icon-caption pairs from Alibaba Icon Library $^4$ and fine-tuning InternVL1.5-26B using this dataset. We then use InternVL-Icon to annotate all icons on web-pages with generated captions. As for the mobile scenario, we collect data from AitW [46], a large-scale GUI dataset on the mobile platform. + +Inspired by Lee et al. [27] and Cheng et al. [11], we construct three tasks based the collected GUI source data: text2bbox,_bbox2text, and_bbox2dom. The text2bbox task prompts the model to ground the element based on the given text or icon caption. To avoid ambiguity, we provide additional contextual information for elements that appear multiple times in the screenshots. The text2bbox data is the most abundant among the three tasks, to help the model develop robust grounding capabilities. The_bbox2text task is the inverse of the text2bbox task, training the model to recognize text and icons. The_bbox2dom task asks the model to generate a DOM-tree corresponding to the given bounding box, as show in Sec. 12.7. This task helps the model learn not only to recognize text/icon elements but also to understand the GUI layout. To make sufficient use of the context length of the model, we pack multiple data pairs in each training sample for text2bbox and_bbox2text task, and select the box that includes as many elements as possible for_bbox2dom task. Overall, we construct a total of 1.9M and 1.1M training samples for the web and mobile scenarios, respectively. The data significantly enhance the GUI foundational abilities, especially the GUI grounding ability, of our SpiritSight agent. See Sec. 14 for more details on + +sample construction. + +# 3.2. Level two: vision-function alignment + +Vision-function alignment refers to a model's ability to locate an element based on its function. This type of data is not directly accessible from the raw GUI data. Inspired by the back-translation method [48] for dataset collection in translation tasks, we leverage InternVL's image understanding capabilities to collect function grounding data. Specifically, we divide each screenshot into a 3x3 grid and describe the approximate location of the target element in text (e.g. in the top-left corner of the image). Additionally, we place a bounding box around the target element in the screenshot to specify its precise location. We provide InternVL2-26B with the screenshot, the element's text content or icon caption, and the location description to prompt it to generate the element's function. Additionally, we utilize InternLM2.5-20B to enhance the quality and diversity of the generated function descriptions. These function descriptions achieve an approximate $90.9\%$ acceptance rate based on human judgment, which we consider sufficient for constructing the function grounding pre-training data. + +Based on the strategy above, we collect function2bbox pairs for all interactive elements in the web scenario. In the mobile scenario, we construct the function grounding data from the GUI navigation dataset, which is be described in Sec. 3.3. We use the same packing method as in the text2bbox and_bbox2text tasks for efficient training, ultimately obtaining 1.5M training samples. + +# 3.3. Level three: visual GUI navigation + +We utilize the public available AitW [46] dataset to construct our GUI navigation training data. AitW is a large- + +scale mobile navigation dataset where each screenshot is labeled with the corresponding goal, the current step, etc. However, as noted in AitZ [72] and AMEX [5], the AitW dataset contains a number of incorrectly labeled samples. We choose to clean it with GPT-4o and employ Chain-of-Thought (CoT) [60] to make the judgment more accurate. Specifically, we prompt GPT-4o with the task description, the current action annotation, and screenshots from both the current and next steps. GPT-4o is then instructed to first summarize the two screenshots and identify the differences between them, then describe the current step based on these differences, and finally assess reasonability of the current action annotation. + +We filter out steps identified as unreasonable by GPT-4o, resulting a final dataset of 0.64M CoT-style GUI navigation training samples. With the collected CoT-style data, we are able to construct additional function grounding data for the mobile scenario, as mentioned in Sec. 3.2, by treating the collected step descriptions as functions of the corresponding elements. + +# 3.4. Other training data + +To enhance the model's understanding of GUI content, we further collected some public datasets as a supplement, including doc/web/mobile VQA [7, 8, 20, 40], image captioning [12, 54], and mobile grounding datasets [12, 33], resulting in 0.59M QA pairs for model training. + +# 4. Universal block parsing + +# 4.1. Problem statement + +We use InternVL2 [9] as the base model of SpiritSight, which uses a dynamic high-resolution training approach to largely preserves the details of input images. The approach first matches the optimal aspect ratio from a pre-defined set of aspect ratios to each input image, then resize and divide each image into blocks of $448 \times 448$ pixels. + +However, this approach may introduce ambiguity in grounding GUI elements. A typical example is represented in Fig. 5a, with two screenshots having aspect ratios of 1:2 and 2:1, respectively. According to the dynamic high-resolution strategy, the screenshots are divided into two blocks, one vertically and the other horizontally. Suppose the target elements in each sample are located in the same relative position within the second block (block-1) after cropping. The cropped screenshots are then fed into InternVL and the model is expected to predict different locations from these two identical inputs. We refer to this situation as the location ambiguity problem. + +Generally, a point $\mathbf{p}$ is expressed in the global coordinate system as + +$$ +\mathbf {p} = [ x, y ] \tag {1} +$$ + +where $x$ and $y$ represent the horizontal and vertical coordinate values of the point in the original image, respectively. Assume that the image is resized and divided into $n_w \times n_h$ blocks. We can express the point $\mathbf{p}$ in the corresponding block as $\mathbf{p}'$ : + +$$ +\left\{ \begin{array}{l} b _ {x} = \left\lfloor \frac {x}{w _ {\text {b l o c k}}} \right\rfloor \\ b _ {y} = \left\lfloor \frac {y}{h _ {\text {b l o c k}}} \right\rfloor \end{array} \quad \left\{ \begin{array}{l} x ^ {\prime} = x \bmod w _ {\text {b l o c k}} \\ y ^ {\prime} = y \bmod h _ {\text {b l o c k}} \end{array} \right. \right. \tag {2} +$$ + +$$ +\mathbf {p} ^ {\prime} = \left[ b _ {x}, b _ {y}, x ^ {\prime}, y ^ {\prime} \right] \tag {3} +$$ + +where $w_{block}$ and $h_{block}$ (both set to 448 in our experiments) represent the width and height of each block, respectively. $b_{x}$ and $b_{y}$ represent the horizontal and vertical indices of the block containing $\mathbf{p}'$ , and $x'$ and $y'$ represent the coordinates of $\mathbf{p}'$ within this block. The blocks are flattened into a sequence before being fed into the model, thus we have + +$$ +\mathbf {p} ^ {\prime \prime} = \left[ b _ {i}, x ^ {\prime}, y ^ {\prime} \right] \tag {4} +$$ + +where $b_{i}$ represents the index of the block within the sequence, satisfying: + +$$ +b _ {i} = b _ {y} \cdot n _ {w} + b _ {x} \tag {5} +$$ + +The model is trained to approximate a mapping $f: \mathbf{p}'' \to \mathbf{p}$ , which is inherently a multi-valued function. For example, when the input $\mathbf{p}''$ is [1, 168, 245], possible values for $p$ include [168, 693] when $n_w = 1$ , or [616, 245] when $n_w = 2$ , as shown in Fig. 5a. + +# 4.2. Method + +The ambiguity primarily arises from the flattening operation, as shown in Eq. (5), which results in the loss of the spatial relationship between blocks. One solution is to input an additional thumbnail, but this may lead to extra computational and memory overhead. We propose to solve this positional ambiguity with two steps. Firstly, we introduce a 2D Block-wise Position Embedding (2D-BPE) [67] by adding two position embeddings to each block to capture spatial information. Secondly, we introduce a Universal Block Parsing (UBP) method, which replaces the global coordinate representation with a block-specific coordinate representation. Specifically, we express a point as Eq. (4). In this case, the model is trained to approximate a injective mapping $f: \mathbf{p}'' \rightarrow \mathbf{p}''$ , thereby resolving the ambiguity problem. During the model inference, the global coordinate of this point can be computed in post-processing as follows: + +$$ +\left\{ \begin{array}{l} x = x ^ {\prime} + \left(b _ {i} \mod n _ {w}\right) \cdot w _ {\text {b l o c k}} \\ y = y ^ {\prime} + \left\lfloor \frac {b _ {i}}{n _ {w}} \right\rfloor \cdot h _ {\text {b l o c k}} \end{array} \right. \tag {6} +$$ + +Overall, our UBP method ensures a clear mapping of positional information between the model's inputs and outputs, thereby improving the model's grounding capability. + +![](images/6cbf1f56b2c8862f6675f16b75ec849427e015e45df37a8cbb3645b5929d639e.jpg) +(a) Baseline Block Parsing vs Universal Block Parsing. + +![](images/01113d42e69228e6e352f1806bb26775fd8ff3a752d69a993076a66ed6e57181.jpg) +(b) Performance gain with UBP. +Figure 5. (a) Comparison between baseline block parsing and our proposed UBP. (b) The results of baseline block parsing and our proposed UBP methods on Multimodal-Mind2Web benchmark. UBP improves the performance of our model. The combination of UBP and 2D-BPE achieves the best results. + +# 5. Settings + +# 5.1. Implementation details + +We use InternVL2 (2B, 8B and 26B) [9] as the base models and train them with two stages: continual pre-training and fine-tuning. During the pre-training stage, we use all the GUI-Lasagne dataset mentioned in Sec. 3. Different prompts are designed for different training tasks to avoid task confusion. See Sec. 14 for the prompts. We unfreeze the vision encoder, decoder, and MLP layer of InternVL. The learning rate is set to $1\mathrm{e} - 4 / 1\mathrm{e} - 4 / 5\mathrm{e} - 5$ for InternVL-2B/8B/26B, respectively, and the batch size is set to 1024. We get SpiritSight-Base model after pre-training and then fine-tune it on multiple downstream tasks individually. More details are shown in Sec. 12.2 and Sec. 12.3. + +# 5.2. Benchmarks & metrics + +To assess SpiritSight's capability in diverse real-world environments, we evaluate SpiritSight on six benchmarks covering various GUI platforms and tasks. Multimodal-Mind2Web [13], ScreenSpot [11], AMEX [5], GUIAct [7], AndroidControl [32], and GUI-Odyssey [38]. For Multimodal-Mind2Web and ScreenSpot, we use the same data pre-processing and metrics as SeeClick [11] uses. For GUIAct, we evaluate SpiritSight on the web-single and web-multi sub-sets and report step success rate (Step SR). For AndroidControl, we evaluate SpiritSight on the High-Level (HL) and Low-Level (LL) tasks and report step accuracy. For AMEX and GUI-Odyssey, we report the action matching score (AMS) defined in AitW [46]. These metrics are similar in that they all measure the single-step accuracy. Refer to Sec. 12.4 for more information about the metrics. + +# 6. Experiment + +# 6.1. Advanced vision-based GUI agent + +In this section, we compare the performance of SpiritSight with other advanced methods across various input modalities and test configurations on Multimodal-Mind2Web [13], a classic and high-quality benchmark for GUI navigation. The results are shown in Tab. 1. Methods that use top-50 candidates as input perform the best. This is evident, as the assistance of candidate elements can significantly reduce the decision space. However, such methods are not particularly feasible in practice. + +The results indicate that SpiritSight significantly outperforms all methods that do not rely on any candidate element, including vision-based methods, language-based methods, and even vision-language-based methods. This demonstrates strong capabilities of SpiritSight in Web GUI navigation tasks. It should be noted that SpiritSight achieves a significant advantage in the Ele.Acc metric compared to other vision-based methods, which can be attributed to the specially constructed visual grounding training data and the proposed UBP approach. + +# 6.2. Strong cross-platform compatibility + +We evaluated SpiritSight on other GUI navigation benchmarks across various platforms and compare it with state-of-the-art (SOTA) GUI agents as shown in Tab. 2. SpiritSight demonstrated the best performance on these benchmarks, showing its strong capabilities across various platforms. OdysseyAgent uses additional historical screenshot images as input, yet SpiritSight achieves comparable results. + +We also evaluated SpiritSight on ScreenSpot, a function grounding benchmark. As shown in Tab. 3, SpiritSight performs well across all three platforms, showcasing its cross-platform capability and robust grounding ability. Notably, + +
AgentOdysseyAMEXAndroidCtrlGUIAct
HighHighHighLowMultiSingle
GPT-4o[42]20.4%-21.2%28.4%-41.8%
Previous SOTA74.3%70.7%64.8%80.0%45.4%74.9%
SpiritSight-2B72.3%74.5%64.9%86.3%45.5%76.0%
SpiritSight-8B75.8%80.7%68.1%87.6%49.3%78.2%
SpiritSight-26B75.9%82.1%69.6%88.2%52.6%78.3%
+ +SpiritSight models are trained specifically for GUI navigation tasks, and no training data aligned with ScreenSpot is used. There remains room for further improvement in SpiritSight's performance on the ScreenSpot benchmark. We further assess the foundational ability of SpiritSight, the visual grounding ability, on our custom-constructed testing dataset. See Sec. 12.7 for more details. + +# 6.3. Recognition and grounding as priors for GUI navigation + +To verify the significance of the three levels of GUI-Lasagne data, we progressively removed level-3, level-2, and level-1 data from the training set during the pretraining stage and evaluate SpiritSight-8B on MultimodalMind2Web. The results are shown by the blue line in Fig. 6a. It can be seen that each level of data contributes to improving Step SR. While the tasks in level-1 data differ the most from web navigation task compared to the other two levels, they provide an effective initialization for the pre-trained model. Although the level-3 data is constructed from the mobile scenario, it also assists in web-based GUI navigation tasks. This indicates that the joint learning strategy helps SpiritSight develop strong navigation abilities across different GUI environments with limited resources. + +We also conducted ablation experiments to evaluate the effectiveness of data cleaning and CoT construction strategies on the level-3 data, as shown by the orange line in Fig. 6a. It can be observed that training in a CoT manner effectively improves the model's GUI navigation capabilities, while the impact of data cleaning strategy is relatively small. This may be due to the fact that erroneous samples are relatively tolerated during the pre-training phase of large models. The difference in results of "w/ CoT" and "w/ Level $1 + 2 + 3$ " is due to the different implementation for fine-tuning stage. See Sec. 12.3 for more details. + +To evaluate the effectiveness of our GUI-Lasagne dataset, we train InternVL-2 (8B) on SeeClick train + +Table 2. Results of SpiritSight on four recently published GUI navigation benchmarks. 'High' and 'Multi' indicates high-level tasks, 'Low' and 'Single' indicates single-stepped tasks. The SOTA results of GUI-Odyssey, AMEX, AndroidControl and GUIAct come from OdysseyAgent [38], SPHINX-GUI-Agent [5], fine-tuned PaLM-2S [32] and MiniCPM-GUI [7], respectively. + +
AgentModel SizeScreenSpot
WebMobileDesktop
GPT4V[1]-5.0%7.5%4.6%
Qwen-VL[3]9.6B3.0%7.2%5.4%
Fuyu[4]8B19.2%21.2%18.3%
CogAgent[19]18B49.5%45.5%47.1%
SeeClick[11]9.6B44.1%65.0%51.1%
SpiritSight-2B2B63.6%62.5%61.8%
SpiritSight-8B8B68.3%68.4%62.9%
SpiritSight-26B26B74.4%73.6%69.9%
+ +Table 3. Results of SpiritSight and other vision-based methods on ScreenSpot Benchmark. + +ing data [11] using full parameter training and LoRA training, resulting in SeeClick(InternVL-Full) and SeeClick(InternVL-LoRA), respectively. We then fine-tune and evaluate the two models on the Multimodal-Mind2Web benchmark. The results are shown by the dashed line in Fig. 6a, where we also present the result of the original SeeClick model. SeeClick(InternVL-LoRA) performs better than the original SeeClick model, which may be attributed to the dynamic high-resolution approach used by InternVL that preserves the details of input images. SeeClick(InternVL-Full) performs the worst, as the scale of the SeeClick training data is not large enough for the model to converge during full parameter training. SpiritSight agent that trained on only GUI-Lasagne level-1 data outperforms all three models trained on SeeClick data, demonstrating the effectiveness of GUI-Lasagne dataset. More ablation studies can be found in Sec. 12.5. + +# 6.4. Better grounding ability from UBP + +To verify the effectiveness of UBP on grounding task, we use LoRA for resource efficiency to fine-tune InternVL in 4 different settings (original as baseline, 2D-BPE, UBP, and 2D-BPE&UBP), respectively. We then evaluate these models on Multimodal-Mind2Web benchmark, as shown in Fig. 5b. It can be seen that UBP shows a significant advantage in Ele.Acc compared to baseline, while the results of Op.F1 show little variation across the 4 settings. This indicates that UBP improves the performance of GUI agent primarily by enhancing the grounding ability. Finally, the combination of UBP and 2D-BPE achieves the best results. This indicates that UBP is compatible with 2D-BPE, leading to better performance. More ablation studies can be found in Sec. 12.6. + +# 6.5. Scaling effects on dataset and model size + +We explore the impact of pre-training dataset and model size on SpiritSight using Multimodal-Mind2Web benchmark. The results are shown in Fig. 6b. SpiritSight-2B out + +![](images/68cb0cf0cbde3674e8ff4c66a96a4808e3d7365dcca842f4fc647ed1a9af8a2f.jpg) +(a) Effect of three data levels and data augmentation for level-3 data +Figure 6. Ablation studies on GUILasagne dataset and scaling effects in the Multimodal-Mind2Web benchmark. + +![](images/75a9adfdde148c1f0f53837684c4722adfd282f03f634c1ff30599f5b936d3a5.jpg) +(b) Ablation Study: Effect of training data percentages on model performance. + +![](images/94a23edebfe12bbac891c5fae7d6da1e2d962594bc8f6112469c5f217008e905.jpg) +(c) Ablation Study: Effect of data percentages on downstream task. + +performs SeeClick [11] by using just 1/8 of the pre-training dataset. This remarkable performance comes from the high quality and grounding-focus nature of GUI-Lasagne data. The performance improves as the size of dataset increases, demonstrating the significance of collecting large-scale data. SpiritSight-2B reaches saturation with a smaller amount of pre-training data, while SpiritSight-26B appears to have further potential for improvement, which aligns with the scaling law of LLMs and VLMs. + +We also evaluate the ability of SpiritSight to transferring to downstream GUI agent tasks. We fine-tune SpiritSight-Base (8B) on various proportions of the Multimodal-Mind2Web training data and show the results in Fig. 6c. It is noticed that SpiritSight achieves $36.6\%$ Step SR with only 1/8 of the fine-tuning data, showing strong transferability to GUI navigation tasks. + +# 6.6. Effective transfer to other languages + +Exploring the cross-lingual capabilities of GUI agents is highly beneficial for their application in non-English environments. We split the training and testing sets of GUIAct (web-multi) data into English and Chinese partitions, respectively. We then fine-tune SpiritSight-Base (8B) on two sets: the entire training set (English&Chinese) and the English-only training set. The results are shown in Tab. 4. + +Under the English&Chinese configuration, SpiritSight achieves comparable results on both the English and Chinese testing sets, despite having fewer Chinese samples in the pre-training data than English ones. Under the English-only configuration, SpiritSight achieves $24.5\%$ Step SR on the Chinese testing set, reaching half of the English&Chinese performance. The zero-shot capability of SpiritSight in Chinese results from the small but effective foundational Chinese data included in the pre-training stage. + +This experiment offers a paradigm for applying GUI agents to non-English environments: by collecting (1) free + +
SFT DataOverallChineseEnglish
English+Chinese49.3%49.3%49.2%
English35.0%24.5%48.6%
+ +Table 4. Results of SpiritSight on GUIAct (web-multi), trained with different language datasets. + +web and mobile GUI information from the target language environment (level 1 & level 2 data), and (2) a small amount of high-quality GUI navigation data at low cost (level 3 data). With this language transferring strategy, the same capabilities as in the English environment can be achieved in non-English environments with minimal extra costs. + +# 7. Limitation + +As SpiritSight is a vision-based GUI agent, it constantly requires access to screenshots which may contain personal information or other sensitive data. Users and system providers should carefully manage the system privileges granted to SpiritSight agent to mitigate potential privacy and security risks. + +# 8. Conclusion + +In this paper, we propose an advanced vision-based end-to-end GUI agent, SpiritSight, with high generalization across multiple GUI platforms. We construct an efficient multilevel, large-scale, high-quality GUI pre-training dataset to equip SpiritSight with robust GUI perception, grounding and understanding capabilities. Additionally, we introduce a UBP method to resolve ambiguity in dynamic high-resolution inputs during model training, further enhancing the ability of SpiritSight to ground GUI objects. As a result, SpiritSight achieves strong performance in numerous GUI navigation benchmarks, demonstrating significant potential for practical deployment in real-world applications. + +# References + +[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 7 +[2] Gilles Baechler, Srinivas Sunkara, Maria Wang, Fedir Zubach, Hassan Mansoor, Vincent Etter, Victor Carbune, Jason Lin, Jindong Chen, and Abhanshu Sharma. Screenai: A vision-language model for ui and infographics understanding. arXiv preprint arXiv:2402.04615, 2024. 3 +[3] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966, 2023.7, 1,4 +[4] Rohan Bavishi, Erich Elsen, Curtis Hawthorne, Maxwell Nye, Augustus Odena, Arushi Somani, and Săgnak Taşürlar. Introducing our multimodal models, 2023. 3, 7 +[5] Yuxiang Chai, Siyuan Huang, Yazhe Niu, Han Xiao, Liang Liu, Dingyu Zhang, Peng Gao, Shuai Ren, and Hongsheng Li. Amex: Android multi-annotation expo dataset for mobile gui agents. arXiv preprint arXiv:2407.17490, 2024. 2, 5, 6, 7, 1, 4 +[6] Dongping Chen, Yue Huang, Siyuan Wu, Jingyu Tang, Liuyi Chen, Yilin Bai, Zhigang He, Chenlong Wang, Huichi Zhou, Yiqiang Li, et al. Gui-world: A dataset for gui-oriented multimodal llm-based agents. arXiv preprint arXiv:2406.10819, 2024. 2 +[7] Wentong Chen, Junbo Cui, Jinyi Hu, Yujia Qin, Junjie Fang, Yue Zhao, Chongyi Wang, Jun Liu, Guirong Chen, Yupeng Huo, et al. Guicourse: From general vision language models to versatile gui agents. arXiv preprint arXiv:2406.11317, 2024. 1, 3, 5, 6, 7, 4 +[8] Xingyu Chen, Zihan Zhao, Lu Chen, Danyang Zhang, Jiabao Ji, Ao Luo, Yuxuan Xiong, and Kai Yu. Websrc: A dataset for web-based structural reading comprehension. arXiv preprint arXiv:2101.09465, 2021. 5 +[9] Zhe Chen, Weiyun Wang, Hao Tian, Shenglong Ye, Zhangwei Gao, Erfei Cui, Wenwen Tong, Kongzhi Hu, Jiapeng Luo, Zheng Ma, et al. How far are we to gpt-4v? closing the gap to commercial multimodal models with open-source suites. arXiv preprint arXiv:2404.16821, 2024. 2, 5, 6, 1 +[10] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24185–24198, 2024. 1 +[11] Kanzhi Cheng, Qiushi Sun, Yougang Chu, Fangzhi Xu, Yantao Li, Jianbing Zhang, and Zhiyong Wu. Seeclick: Harnessing gui grounding for advanced visual gui agents. arXiv preprint arXiv:2401.10935, 2024. 1, 3, 4, 6, 7, 8 +[12] Biplab Deka, Zifeng Huang, Chad Franzen, Joshua Hibschman, Daniel Afergan, Yang Li, Jeffrey Nichols, and Ranjitha Kumar. Rico: A mobile app dataset for building data-driven design applications. In Proceedings of the 30th annual + +ACM symposium on user interface software and technology, pages 845-854, 2017. 5, 3 +[13] Xiang Deng, Yu Gu, Boyuan Zheng, Shijie Chen, Sam Stevens, Boshi Wang, Huan Sun, and Yu Su. Mind2web: Towards a generalist agent for the web. Advances in Neural Information Processing Systems, 36, 2024. 1, 2, 6, 4 +[14] Jacob Devlin. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 1 +[15] Hiroki Furuta, Kuang-Huei Lee, Ofir Nachum, Yutaka Matsuo, Aleksandra Faust, Shixiang Shane Gu, and Izzeddin Gur. Multimodal web navigation with instruction-finetuned foundation models. arXiv preprint arXiv:2305.11854, 2023. 1, 3 +[16] Common crawl Group. Common crawl - open repository of web crawl data, 2024. 3, 5 +[17] Izzeddin Gur, Hiroki Furuta, Austin Huang, Mustafa Safdari, Yutaka Matsuo, Douglas Eck, and Aleksandra Faust. A real-world webagent with planning, long context understanding, and program synthesis. arXiv preprint arXiv:2307.12856, 2023. 2, 3 +[18] Markus Hafner, Maria Katsantoni, Tino Köster, James Marks, Joyita Mukherjee, Dorothee Staiger, Jernej Ule, and Mihaela Zavolan. Clip and complementary methods. Nature Reviews Methods Primers, 1(1):1-23, 2021. 1 +[19] Wenyi Hong, Weihan Wang, Qingsong Lv, Jiazheng Xu, Wenmeng Yu, Junhui Ji, Yan Wang, Zihan Wang, Yuxiao Dong, Ming Ding, et al. Cogagent: A visual language model for gui agents. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14281-14290, 2024. 1, 3, 7 +[20] Yu-Chung Hsiao, Fedir Zubach, Maria Wang, et al. Screenqa: Large-scale question-answer pairs over mobile app screenshots. arXiv preprint arXiv:2209.08199, 2022. 5 +[21] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 3 +[22] Faria Huq, Jeffrey P Bigham, and Nikolas Martelaro. “what’s important here?”: Opportunities and challenges of using llms in retrieving information from web interfaces. arXiv preprint arXiv:2312.06147, 2023. 1 +[23] Yang Jin, Kun Xu, Kun Xu, Liwei Chen, Chao Liao, Jianchao Tan, Quzhe Huang, Bin Chen, Chenyi Lei, An Liu, et al. Unified language-vision pretraining in llm with dynamic discrete visual tokenization. arxiv 2024. arXiv preprint arXiv:2309.04669, 2023. 1 +[24] Jihyung Kil, Chan Hee Song, Boyuan Zheng, Xiang Deng, Yu Su, and Wei-Lun Chao. Dual-view visual contextualization for web navigation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14445-14454, 2024. 1 +[25] Geewook Kim, Teakgyu Hong, Moonbin Yim, Jeong Yeon Nam, Jinyoung Park, Jinyeong Yim, Wonseok Hwang, Sang-doo Yun, Dongyoon Han, and Seunghyun Park. Ocr-free document understanding transformer. In European Conference on Computer Vision, pages 498-517. Springer, 2022. 1, 2 + +[26] Hanyu Lai, Xiao Liu, Iat Long Iong, Shuntian Yao, Yuxuan Chen, Pengbo Shen, Hao Yu, Hanchen Zhang, Xiaohan Zhang, Yuxiao Dong, et al. Autowebglm: Bootstrap and reinforce a large language model-based web navigating agent. arXiv preprint arXiv:2404.03648, 2024. 1, 2, 3 +[27] Kenton Lee, Mandar Joshi, Iulia Raluca Turc, Hexiang Hu, Fangyu Liu, Julian Martin Eisenschlos, Urvashi Khandelwal, Peter Shaw, Ming-Wei Chang, and Kristina Toutanova. Pix2struct: Screenshot parsing as pretraining for visual language understanding. In International Conference on Machine Learning, pages 18893–18912. PMLR, 2023. 2, 4 +[28] Kuang-Huei Lee, Xinyun Chen, Hiroki Furuta, John Canny, and Ian Fischer. A human-inspired reading agent with gist memory of very long contexts. arXiv preprint arXiv:2402.09727, 2024. 1, 3 +[29] Feng Li, Renrui Zhang, Hao Zhang, Yuanhan Zhang, Bo Li, Wei Li, Zejun Ma, and Chunyuan Li. Llava next-interleave: Tackling multi-image, video, and 3d in large multimodal models. arXiv preprint arXiv:2407.07895, 2024. 1 +[30] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International conference on machine learning, pages 12888-12900. PMLR, 2022. 1 +[31] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730-19742. PMLR, 2023. 1 +[32] Wei Li, William Bishop, Alice Li, Chris Rawles, Folawiyo Campbell-Ajala, Divya Tyamagundlu, and Oriana Riva. On the effects of data scale on computer control agents. arXiv preprint arXiv:2406.03679, 2024. 6, 7, 1, 4 +[33] Yang Li, Gang Li, Luheng He, Jingjie Zheng, Hong Li, and Zhiwei Guan. Widget captioning: Generating natural language description for mobile user interface elements. arXiv preprint arXiv:2010.04295, 2020. 5, 3 +[34] Zeyi Liao, Lingbo Mo, Chejian Xu, Mintong Kang, Jiawei Zhang, Chaowei Xiao, Yuan Tian, Bo Li, and Huan Sun. Eia: Environmental injection attack on generalist web agents for privacy leakage. arXiv preprint arXiv:2409.11295, 2024. 1, 3 +[35] Ziyi Lin, Chris Liu, Renrui Zhang, Peng Gao, Longtian Qiu, Han Xiao, Han Qiu, Chen Lin, Wenqi Shao, Keqin Chen, et al. Sphinx: The joint mixing of weights, tasks, and visual embeddings for multi-modal large language models. arXiv preprint arXiv:2311.07575, 2023. 1 +[36] Evan Zheran Liu, Kelvin Guu, Panupong Pasupat, Tianlin Shi, and Percy Liang. Reinforcement learning on web interfaces using workflow-guided exploration. arXiv preprint arXiv:1802.08802, 2018. 2, 1 +[37] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning, 2023. 1 +[38] Quanfeng Lu, Wenqi Shao, Zitao Liu, Fanqing Meng, Boxuan Li, Botong Chen, Siyuan Huang, Kaipeng Zhang, Yu Qiao, and Ping Luo. Gui odyssey: A comprehensive dataset for cross-app gui navigation on mobile devices. arXiv preprint arXiv:2406.08451, 2024. 2, 6, 7, 1, 4 + +[39] Xing Han Lu, Zdenek Kasner, and Siva Reddy. Weblinx: Real-world website navigation with multi-turn dialogue. arXiv preprint arXiv:2402.05930, 2024. 2 +[40] Minesh Mathew, Dimosthenis Karatzas, and CV Jawahar. Docvqa: A dataset for vqa on document images. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 2200-2209, 2021. 5 +[41] Erik Nijkamp, Bo Pang, Hiroaki Hayashi, Lifu Tu, Huan Wang, Yingbo Zhou, Silvio Savarese, and Caiming Xiong. Codegen: An open large language model for code with multi-turn program synthesis. arXiv preprint arXiv:2203.13474, 2022. 1 +[42] OpenAI. Hello gpt-4o, 2024. 7 +[43] Liangming Pan, Alon Albalak, Xinyi Wang, and William Yang Wang. Logic-lm: Empowering large language models with symbolic solvers for faithful logical reasoning. arXiv preprint arXiv:2305.12295, 2023. 1 +[44] Alec Radford, Karthik Narasimhan, Tim Salimans, Ilya Sutskever, et al. Improving language understanding by generative pre-training. OpenAI, 2018. +[45] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of machine learning research, 21(140):1-67, 2020. 1 +[46] Christopher Rawles, Alice Li, Daniel Rodriguez, Oriana Riva, and Timothy Lillicrap. Androidinthewild: A large-scale dataset for android device control. Advances in Neural Information Processing Systems, 36, 2024. 2, 4, 6, 1, 7 +[47] Baptiste Roziere, Jonas Gehring, Fabian Gloeckle, Sten Sootla, Itai Gat, Xiaqing Ellen Tan, Yossi Adi, Jingyu Liu, Romain Sauvestre, Tal Remez, et al. Code llama: Open foundation models for code. arXiv preprint arXiv:2308.12950, 2023. 1 +[48] Rico Sennrich. Improving neural machine translation models with monolingual data. arXiv preprint arXiv:1511.06709, 2015. 4 +[49] Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Mingchuan Zhang, YK Li, Yu Wu, and Daya Guo. Deepseekmath: Pushing the limits of mathematical reasoning in open language models. arXiv preprint arXiv:2402.03300, 2024.1 +[50] Peter Shaw, Mandar Joshi, James Cohan, Jonathan Berant, Panupong Pasupat, Hexiang Hu, Urvashi Khandelwal, Kenton Lee, and Kristina N Toutanova. From pixels to ui actions: Learning to follow instructions via graphical user interfaces. Advances in Neural Information Processing Systems, 36:34354-34370, 2023. 3 +[51] Tianlin Shi, Andrej Karpathy, Linxi Fan, Jonathan Hernandez, and Percy Liang. World of bits: An open-domain platform for web-based agents. In International Conference on Machine Learning, pages 3135-3144. PMLR, 2017. 2, 1 +[52] Lucas-Andrei Thil, Mirela Popa, and Gerasimos Spanakis. Navigating webai: Training agents to complete web tasks with large language models and reinforcement learning. In Proceedings of the 39th ACM/SIGAPP Symposium on Applied Computing, pages 866–874, 2024. 1, 3 + +[53] Jianqiang Wan, Sibo Song, Wenwen Yu, Yuliang Liu, Wenqing Cheng, Fei Huang, Xiang Bai, Cong Yao, and Zhibo Yang. Omniparser: A unified framework for text spotting key information extraction and table recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15641-15653, 2024. 1, 3 +[54] Bryan Wang, Gang Li, Xin Zhou, Zhourong Chen, Tovi Grossman, and Yang Li. Screen2words: Automatic mobile ui summarization with multimodal learning. In The 34th Annual ACM Symposium on User Interface Software and Technology, pages 498-510, 2021. 5, 3 +[55] Junyang Wang, Haiyang Xu, Haitao Jia, Xi Zhang, Ming Yan, Weizhou Shen, Ji Zhang, Fei Huang, and Jitao Sang. Mobile-agent-v2: Mobile device operation assistant with effective navigation via multi-agent collaboration. arXiv preprint arXiv:2406.01014, 2024. 2, 3 +[56] Junyang Wang, Haiyang Xu, Jiabo Ye, Ming Yan, Weizhou Shen, Ji Zhang, Fei Huang, and Jitao Sang. Mobile-agent: Autonomous multi-modal mobile device agent with visual perception. arXiv preprint arXiv:2401.16158, 2024. 2, 3 +[57] Ke Wang, Houxing Ren, Aojun Zhou, Zimu Lu, Sichun Luo, Weikang Shi, Renrui Zhang, Linqi Song, Mingjie Zhan, and Hongsheng Li. Mathcoder: Seamless code integration in llms for enhanced mathematical reasoning. arXiv preprint arXiv:2310.03731, 2023. 1 +[58] Weihan Wang, Qingsong Lv, Wenmeng Yu, Wenyi Hong, Ji Qi, Yan Wang, Junhui Ji, Zhuoyi Yang, Lei Zhao, Xixuan Song, et al. Cogvlm: Visual expert for pretrained language models. arXiv preprint arXiv:2311.03079, 2023. 1 +[59] Jason Wei, Yi Tay, Rishi Bommasani, Colin Raffel, Barret Zoph, Sebastian Borgeaud, Dani Yogatama, Maarten Bosma, Denny Zhou, Donald Metzler, et al. Emergent abilities of large language models. arXiv preprint arXiv:2206.07682, 2022. 1 +[60] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824-24837, 2022. 5, 1, 7 +[61] Chen Henry Wu, Jing Yu Koh, Ruslan Salakhutdinov, Daniel Fried, and Aditi Raghunathan. Adversarial attacks on multimodal agents. arXiv preprint arXiv:2406.12814, 2024. 1, 3 +[62] Shijie Wu, Ozan Irsoy, Steven Lu, Vadim Dabravolski, Mark Dredze, Sebastian Gehrmann, Prabhanjan Kambadur, David Rosenberg, and Gideon Mann. Bloomberg: A large language model for finance. arXiv preprint arXiv:2303.17564, 2023. 1 +[63] Yichong Xu, Chenguang Zhu, Shuohang Wang, Siqi Sun, Hao Cheng, Xiaodong Liu, Jianfeng Gao, Pengcheng He, Michael Zeng, and Xuedong Huang. Human parity on commonsenseqa: Augmenting self-attention with external attention. arXiv preprint arXiv:2112.03254, 2021. 1 +[64] Zhao Yang, Jiaxuan Liu, Yucheng Han, Xin Chen, Zebiao Huang, Bin Fu, and Gang Yu. Appagent: Multimodal agents as smartphone users. arXiv preprint arXiv:2312.13771, 2023. 3 + +[65] Shunyu Yao, Howard Chen, John Yang, and Karthik Narasimhan. Webshop: Towards scalable real-world web interaction with grounded language agents. Advances in Neural Information Processing Systems, 35:20744-20757, 2022. 2, 1 +[66] Yuan Yao, Tianyu Yu, Ao Zhang, Chongyi Wang, Junbo Cui, Hongji Zhu, Tianchi Cai, Haoyu Li, Weilin Zhao, Zhihui He, et al. Minicpm-v: A gpt-4v level mllm on your phone. arXiv preprint arXiv:2408.01800, 2024. 1 +[67] Jiabo Ye, Anwen Hu, Haiyang Xu, Qinghao Ye, Ming Yan, Guohai Xu, Chenliang Li, Junfeng Tian, Qi Qian, Ji Zhang, et al. Ureader: UniversalOCR-free visually-situated language understanding with multimodal large language model. arXiv preprint arXiv:2310.05126, 2023. 2, 5 +[68] Da Yin, Faeze Brahman, Abhilasha Ravichander, Khyathi Chandu, Kai-Wei Chang, Yejin Choi, and Bill Yuchen Lin. Agent lumos: Unified and modular training for open-source language agents. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 12380–12403, 2024. 1 +[69] Huaiyuan Ying, Shuo Zhang, Linyang Li, Zhejian Zhou, Yunfan Shao, Zhaoye Fei, Yichuan Ma, Jiawei Hong, Kuikun Liu, Ziyi Wang, et al. Internlm-math: Open math large language models toward verifiable reasoning. arXiv preprint arXiv:2402.06332, 2024. 1 +[70] Longhui Yu, Weisen Jiang, Han Shi, Jincheng Yu, Zhengying Liu, Yu Zhang, James T Kwok, Zhenguo Li, Adrian Weller, and Weiyang Liu. Metamath: Bootstrap your own mathematical questions for large language models. arXiv preprint arXiv:2309.12284, 2023. 1 +[71] Qiusi Zhan, Zhixiang Liang, Zifan Ying, and Daniel Kang. Injecagent: Benchmarking indirect prompt injections in tool-integrated large language model agents. arXiv preprint arXiv:2403.02691, 2024. 1, 3 +[72] Jiwen Zhang, Jihao Wu, Yihua Teng, Minghui Liao, Nuo Xu, Xiao Xiao, Zhongyu Wei, and Duyu Tang. Android in the zoo: Chain-of-action-thought for gui agents. arXiv preprint arXiv:2403.02713, 2024. 5, 7 +[73] Pan Zhang, Xiaoyi Dong, Yuhang Zang, Yuhang Cao, Rui Qian, Lin Chen, Qipeng Guo, Haodong Duan, Bin Wang, Linke Ouyang, Songyang Zhang, Wenwei Zhang, Yining Li, Yang Gao, Peng Sun, Xinyue Zhang, Wei Li, Jingwen Li, Wenhai Wang, Hang Yan, Conghui He, Xingcheng Zhang, Kai Chen, Jifeng Dai, Yu Qiao, Dahua Lin, and Jiaqi Wang. Internlm-xcomposer-2.5: A versatile large vision language model supporting long-contextual input and output, 2024. 1 +[74] Boyuan Zheng, Boyu Gou, Jihyung Kil, Huan Sun, and Yu Su. Gpt-4v (isdiction) is a generalist web agent, if grounded. arXiv preprint arXiv:2401.01614, 2024. 1, 2, 3 +[75] Longtao Zheng, Rundong Wang, Xinrun Wang, and Bo An. Synapse: Trajectory-as-exemplar prompting with memory for computer control. In The Twelfth International Conference on Learning Representations, 2023. 1 \ No newline at end of file diff --git a/2025/SpiritSight Agent_ Advanced GUI Agent with One Look/images.zip b/2025/SpiritSight Agent_ Advanced GUI Agent with One Look/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..5ecceccc221882db96d6e210d2eab678aab36aa7 --- /dev/null +++ b/2025/SpiritSight Agent_ Advanced GUI Agent with One Look/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3f2be017738912ac3f7c61ba222b13c6bc9086b842cce15b4fa8b8ad0d0af842 +size 646763 diff --git a/2025/SpiritSight Agent_ Advanced GUI Agent with One Look/layout.json b/2025/SpiritSight Agent_ Advanced GUI Agent with One Look/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..f1842450ef6d7f47c2eb12e7c0f5d246f24ab908 --- /dev/null +++ b/2025/SpiritSight Agent_ Advanced GUI Agent with One Look/layout.json @@ -0,0 +1,8651 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 123, + 100, + 143, + 117 + ], + "blocks": [ + { + "bbox": [ + 123, + 100, + 143, + 117 + ], + "lines": [ + { + "bbox": [ + 123, + 100, + 143, + 117 + ], + "spans": [ + { + "bbox": [ + 123, + 100, + 143, + 117 + ], + "type": "image", + "image_path": "f6471b3df2f89a2cd7883fa6ed05731ecd52be93cc5193613bdebf5724d4fd70.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 142, + 103, + 488, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 142, + 103, + 488, + 121 + ], + "spans": [ + { + "bbox": [ + 142, + 103, + 488, + 121 + ], + "type": "text", + "content": "SpiritSight Agent: Advanced GUI Agent with One Look" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 75, + 148, + 534, + 164 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 148, + 534, + 164 + ], + "spans": [ + { + "bbox": [ + 75, + 148, + 534, + 164 + ], + "type": "text", + "content": "Zhiyuan Huang" + }, + { + "bbox": [ + 75, + 148, + 534, + 164 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 75, + 148, + 534, + 164 + ], + "type": "text", + "content": " Ziming Cheng" + }, + { + "bbox": [ + 75, + 148, + 534, + 164 + ], + "type": "inline_equation", + "content": "^{1,2*}" + }, + { + "bbox": [ + 75, + 148, + 534, + 164 + ], + "type": "text", + "content": " Junting Pan" + }, + { + "bbox": [ + 75, + 148, + 534, + 164 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 75, + 148, + 534, + 164 + ], + "type": "text", + "content": " Zhaohui Hou" + }, + { + "bbox": [ + 75, + 148, + 534, + 164 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 75, + 148, + 534, + 164 + ], + "type": "text", + "content": " Mingjie Zhan" + }, + { + "bbox": [ + 75, + 148, + 534, + 164 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 76, + 167, + 535, + 182 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 167, + 535, + 182 + ], + "spans": [ + { + "bbox": [ + 76, + 167, + 535, + 182 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 76, + 167, + 535, + 182 + ], + "type": "text", + "content": "SenseTime Research, " + }, + { + "bbox": [ + 76, + 167, + 535, + 182 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 76, + 167, + 535, + 182 + ], + "type": "text", + "content": "Beijing University of Posts and Telecommunications, " + }, + { + "bbox": [ + 76, + 167, + 535, + 182 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 76, + 167, + 535, + 182 + ], + "type": "text", + "content": "MMLab, CUHK" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 161, + 198, + 449, + 210 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 161, + 198, + 449, + 210 + ], + "spans": [ + { + "bbox": [ + 161, + 198, + 449, + 210 + ], + "type": "text", + "content": "{huangzhiyuan, houzhaohui, zhanmingjie}@sensetime.com" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 166, + 215, + 440, + 226 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 166, + 215, + 440, + 226 + ], + "spans": [ + { + "bbox": [ + 166, + 215, + 440, + 226 + ], + "type": "text", + "content": "harrytheorange@bupt.edu.cn junting.pa@gmail.com" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 151, + 255, + 200, + 268 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 255, + 200, + 268 + ], + "spans": [ + { + "bbox": [ + 151, + 255, + 200, + 268 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 280, + 296, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 280, + 296, + 579 + ], + "spans": [ + { + "bbox": [ + 55, + 280, + 296, + 579 + ], + "type": "text", + "content": "Graphical User Interface (GUI) agents demonstrate promising potential in assisting human-computer interaction, automating human user's navigation on digital devices. An ideal GUI agent is expected to achieve high accuracy, low latency, and compatibility for different GUI platforms. Recent vision-based approaches have shown promise by leveraging advanced Vision Language Models (VLMs). While they generally meet the requirements of compatibility and low latency, these vision-based GUI agents tend to have low accuracy due to their limitations in element grounding. To address this issue, we propose SpiritSight, a vision-based, end-to-end GUI agent that excels in GUI navigation tasks across various GUI platforms. First, we create a multi-level, large-scale, high-quality GUI dataset called GUI-Lasagne using scalable methods, empowering SpiritSight with robust GUI understanding and grounding capabilities. Second, we introduce the Universal Block Parsing (UBP) method to resolve the ambiguity problem inherited from the dynamic resolution strategy, further enhancing SpiritSight's ability to ground GUI objects. Through these efforts, SpiritSight agent outperforms other advanced methods on diverse GUI benchmarks, demonstrating its superior capability and compatibility in GUI navigation tasks. The models and code will be made available upon publication." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 602, + 135, + 615 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 602, + 135, + 615 + ], + "spans": [ + { + "bbox": [ + 56, + 602, + 135, + 615 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 622, + 296, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 622, + 296, + 694 + ], + "spans": [ + { + "bbox": [ + 55, + 622, + 296, + 694 + ], + "type": "text", + "content": "Graphical User Interface (GUI) automation has long been pursued by people along with the development of modern digital devices. Thanks to recent advances in Large Language Models (LLMs), GUI agents are constructed to assist users in interacting with graphical interfaces, automatically making action decisions based on observations of environ-" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 324, + 255, + 545, + 437 + ], + "blocks": [ + { + "bbox": [ + 324, + 255, + 545, + 437 + ], + "lines": [ + { + "bbox": [ + 324, + 255, + 545, + 437 + ], + "spans": [ + { + "bbox": [ + 324, + 255, + 545, + 437 + ], + "type": "image", + "image_path": "b75cfff39780575a3f479d6e58c42d4b8a4ff6db0da1f119a00fe244d844c1dc.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 444, + 555, + 478 + ], + "lines": [ + { + "bbox": [ + 313, + 444, + 555, + 478 + ], + "spans": [ + { + "bbox": [ + 313, + 444, + 555, + 478 + ], + "type": "text", + "content": "Figure 1. Our SpiritSight agent achieves new state-of-the-art (SOTA) performance across various benchmarks in web, mobile, and desktop scenarios." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 503, + 427, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 503, + 427, + 514 + ], + "spans": [ + { + "bbox": [ + 313, + 503, + 427, + 514 + ], + "type": "text", + "content": "ment and user's instruction." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 517, + 555, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 517, + 555, + 601 + ], + "spans": [ + { + "bbox": [ + 313, + 517, + 555, + 601 + ], + "type": "text", + "content": "Current approaches can be divided into three categories based on their input modalities. Language-based and vision-based approaches use Hyper Text Markup Language (HTML) or Extensible Markup Language (XML) and screenshots as input [7, 11, 13, 19, 22, 26, 28, 53, 68, 75], respectively, while vision-language-based approaches combine both HTML and screenshots as input [15, 24, 52, 74]." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 603, + 556, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 603, + 556, + 687 + ], + "spans": [ + { + "bbox": [ + 313, + 603, + 556, + 687 + ], + "type": "text", + "content": "The language-based and vision-language-based methods typically applied only in the web domain, and are often limited by the excessive length of HTML or security concerns [34, 61, 71]. The vision-based approaches exhibit compatibility across various GUI platforms, as acquiring screenshots is generally easier than obtaining hierarchical data from non-web platforms." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 689, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 556, + 713 + ], + "type": "text", + "content": "However, vision-based approaches struggle to ground GUI elements (e.g. buttons, text boxes) from the visual" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 703, + 134, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 134, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 134, + 712 + ], + "type": "text", + "content": "*Equal contribution." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "29490" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "type": "text", + "content": "input [74]. Some works solve this problem by introducing additional tools, such as optical character recognition (OCR) and icon recognition models, to convert the visual grounding task into a language QA task [55, 56]. This may increase the complexity and inference latency of the agent system. Others attempt to collect large-scale training data through manual synthesis [27, 36, 51] or human annotation [5, 6, 13, 38, 39, 46, 65], while these data are either unrealistic or costly. Additionally, recent advanced Vision Language Models (VLMs) [9, 67] adapt dynamic high-resolution [9, 67] strategy to better suit high-resolution inputs, which we find may introduce ambiguity to the model learning process in GUI scenarios." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 228, + 296, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 228, + 296, + 276 + ], + "spans": [ + { + "bbox": [ + 55, + 228, + 296, + 276 + ], + "type": "text", + "content": "To address the aforementioned challenges, we propose an end-to-end, vision-based GUI agent—SpiritSight, which has strong ability for GUI navigation task. Our contributions are summarized as follows:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 277, + 296, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 277, + 296, + 396 + ], + "spans": [ + { + "bbox": [ + 55, + 277, + 296, + 396 + ], + "type": "text", + "content": "Firstly, we propose GUI-Lasagne, a multi-level, large-scale and high-quality GUI dataset composed of 5.73 million samples to enhance our model's GUI understanding and grounding capabilities. The dataset is collected from the real world and filtered through carefully designed rules to ensure data quality. It is also constructed hierarchically and consists of three levels of components. The first two parts constitute " + }, + { + "bbox": [ + 55, + 277, + 296, + 396 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 55, + 277, + 296, + 396 + ], + "type": "text", + "content": " of the total dataset and are collected for free, thus significantly reducing the data collection cost." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 397, + 296, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 397, + 296, + 468 + ], + "spans": [ + { + "bbox": [ + 55, + 397, + 296, + 468 + ], + "type": "text", + "content": "Secondly, we introduce a Universal Block Parsing (UBP) method to resolve the ambiguity problem inherited from the dynamic resolution strategy. With the effectiveness of UBP, SpiritSight gains an improved elements grounding capability and achieves enhanced performance in GUI navigation tasks." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 469, + 296, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 469, + 296, + 625 + ], + "spans": [ + { + "bbox": [ + 55, + 469, + 296, + 625 + ], + "type": "text", + "content": "Thirdly, we release and evaluate SpiritSight agent in various GUI benchmarks and it exhibits remarkable performance among them. SpiritSight agent is pre-trained on GUI-Lasagne and employs UBP to ground elements, as shown in Fig. 3. We release three versions of SpiritSight with different model size: SpiritSight-26B, SpiritSight-8B, and SpiritSight-2B. The standard SpiritSight-8B consistently outperforms previous state-of-the-art methods across various GUI benchmarks, as shown in Fig. 1. On the Multimodal-Mind2Web [13] benchmark, the SpiritSight series outperforms all kinds of methods—including language-based, vision-based and even vision-language-based methods—in the non-candidate setting, as shown in Fig. 2." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 638, + 140, + 651 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 638, + 140, + 651 + ], + "spans": [ + { + "bbox": [ + 55, + 638, + 140, + 651 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 658, + 295, + 683 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 658, + 295, + 683 + ], + "spans": [ + { + "bbox": [ + 55, + 658, + 295, + 683 + ], + "type": "text", + "content": "2.1. Language-based and vision-language-based GUI agent" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "Several works leverage the capabilities of large-scale language models (LLMs) to construct GUI agents. It is" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 325, + 72, + 538, + 224 + ], + "blocks": [ + { + "bbox": [ + 325, + 72, + 538, + 224 + ], + "lines": [ + { + "bbox": [ + 325, + 72, + 538, + 224 + ], + "spans": [ + { + "bbox": [ + 325, + 72, + 538, + 224 + ], + "type": "image", + "image_path": "1c61129ecf89b1b09352b18720a18fa0f35f80a7644c09b4ab3d4bf137012067.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 234, + 555, + 268 + ], + "lines": [ + { + "bbox": [ + 313, + 234, + 555, + 268 + ], + "spans": [ + { + "bbox": [ + 313, + 234, + 555, + 268 + ], + "type": "text", + "content": "Figure 2. Comparison of the average step success rate on Multimodal-Mind2Web benchmark of our SpiritSight agent of three sizes (2B, 8B, 26B) with various previous methods." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 323, + 281, + 548, + 453 + ], + "blocks": [ + { + "bbox": [ + 323, + 281, + 548, + 453 + ], + "lines": [ + { + "bbox": [ + 323, + 281, + 548, + 453 + ], + "spans": [ + { + "bbox": [ + 323, + 281, + 548, + 453 + ], + "type": "image", + "image_path": "702443052d2c731f186e652a64491d001cbb7d776c03265819be8e16b6a5d74d.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 468, + 555, + 524 + ], + "lines": [ + { + "bbox": [ + 313, + 468, + 555, + 524 + ], + "spans": [ + { + "bbox": [ + 313, + 468, + 555, + 524 + ], + "type": "text", + "content": "Figure 3. The overview of our SpiritSight agent. Our proposed GUI-Lasagne dataset equips SpiritSight with three levels of comprehensive GUI knowledge. Additionally, we introduce a Universal Block Parsing (UBP) method to enhance SpiritSight's grounding capabilities." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 545, + 555, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 555, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 555, + 689 + ], + "type": "text", + "content": "noticed that they are mostly multi-stage architectures. Mind2Web [13] utilizes a lightweight language model to extract candidate elements from HTML, followed by a ranking model that sorts these elements based on task descriptions and historical actions. Finally, a large language model predicts the current actions and selects the target elements to operate on. WebAgent [17] first uses an encoder-decoder model to generate low-level instructions and relevant HTML code snippets, then uses another decoder to produce executable Python code. AutoWebGLM [26] simplifies HTML code through manually designed rules before predicting the executable action codes." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 689, + 555, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 712 + ], + "type": "text", + "content": "Other vision-language-based works leverage both GUI screenshots and hierarchical HTML/XML to enhance the" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "29491" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 57, + 69, + 553, + 238 + ], + "blocks": [ + { + "bbox": [ + 57, + 69, + 553, + 238 + ], + "lines": [ + { + "bbox": [ + 57, + 69, + 553, + 238 + ], + "spans": [ + { + "bbox": [ + 57, + 69, + 553, + 238 + ], + "type": "table", + "html": "
Model SizeInput ModalitySelect From TopCross-TaskCross-WebsiteCross-Domain
Ele.AccOp.F1Step SREle.AccOp.F1Step SREle.AccOp.F1Step SR
AutoWebGLM [26]6BText✓--66.4%--56.4%--55.8%
LLaMA2-7B [26]7BText✓--52.7%--47.1%--50.3%
CogAgent [19]18BImage✓--62.3%--54.0%--59.4%
HTML-T5-XL [17]3BText✓76.4%78.8%71.5%68.4%71.0%62.2%73.0%75.6%67.1%
SeeAct* [74]-Text+Image×46.4%73.4%40.2%38.0%67.8%32.4%42.4%69.3%36.8%
ReadAgent-P* [28]340BText×33.7%72.5%29.2%37.4%75.1%31.1%37.2%76.3%33.4%
MiniCPM-GUI [7]3BImage×23.8%86.8%20.8%20.3%81.7%17.3%17.9%74.5%14.6%
Fuyu-GUI [4]8BImage×19.1%86.1%15.6%13.9%80.7%12.2%14.2%83.1%11.7%
SeeClick [11]9.6BImage×28.3%87.0%25.5%21.4%80.6%16.4%23.2%84.8%20.8%
OmniParser* [53]-Image×42.4%87.6%39.4%41.0%84.8%36.5%45.5%85.7%42.0%
SpiritSight-2B2BImage×51.7%87.2%44.9%44.0%83.6%37.8%42.4%83.5%36.9%
SpiritSight-8B8BImage×59.2%88.9%52.7%52.2%84.7%44.0%50.1%86.0%44.4%
SpiritSight-26B26BImage×60.5%89.7%54.7%57.0%85.7%48.1%54.1%87.2%49.2%
", + "image_path": "cf5c1cffe79d20257a05972072800f189ee3c2eb099bb88c2900d108980b3648.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 256, + 555, + 291 + ], + "lines": [ + { + "bbox": [ + 55, + 256, + 555, + 291 + ], + "spans": [ + { + "bbox": [ + 55, + 256, + 555, + 291 + ], + "type": "text", + "content": "Table 1. Comparison of SpiritSight agent with various previous methods on Multimodal-Mind2Web benchmark. SpiritSight significantly outperforms all methods that do not rely on any candidate element, including vision-based methods, language-based methods, and even vision-language-based methods. The asterisk (*) in the upper-right indicates that the model is not fine-tuned on the downstream tasks." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 310, + 294, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 310, + 294, + 407 + ], + "spans": [ + { + "bbox": [ + 54, + 310, + 294, + 407 + ], + "type": "text", + "content": "robustness of GUI agents. WebGUM [15] and CC-Net [52] use ResNet and ViT to extract features from screenshots, respectively. The image embeddings are then combined with text embeddings and fed into a multi-modal transformer. SeeAct [74] and AppAgent [64] identify all interactive elements using HTML or XML data. They then assign each interactive element a unique identifier in the screenshot before feeding the screenshot into the model." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 407, + 295, + 551 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 407, + 295, + 551 + ], + "spans": [ + { + "bbox": [ + 55, + 407, + 295, + 551 + ], + "type": "text", + "content": "These language-based and vision-language-based methods, which rely on hierarchical information, exhibit several limitations: (1) Hierarchical representations, such as HTML or XML, are not consistently available across different platforms. And even this information is available, differences in their internal rules make language-based GUI agents less compatible; (2) HTML often contains redundant information, requiring additional models or extensive manually crafted rules for effective filtering. (3) Text-based GUI agents are susceptible to injection attacks [34, 61, 71], where malicious instructions hidden in HTML can easily lead to erroneous or unsafe actions." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 562, + 190, + 575 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 562, + 190, + 575 + ], + "spans": [ + { + "bbox": [ + 55, + 562, + 190, + 575 + ], + "type": "text", + "content": "2.2. Vision-based GUI agent" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "type": "text", + "content": "Recently, some vision-based approaches have been proposed to overcome the drawbacks of language-based methods. Some of them [2, 11, 19, 50] are single-stage methods that only use GUI screenshots as input for VLMs and output the next action in an end-to-end manner. However, these agents perform worse on GUI benchmarks compared to other methods. Two primary reasons may be the relatively low input resolution of their base models and the insufficient scale and suboptimal quality of the training data. SeeAct [74], MobileAgent [56] and MobileAgent-v2 [55] are two-stage methods. They use GPT-4V instead of open" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 311, + 555, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 311, + 555, + 418 + ], + "spans": [ + { + "bbox": [ + 313, + 311, + 555, + 418 + ], + "type": "text", + "content": "source VLMs as base models and find that even the top models struggle with element grounding. Consequently, they introduce additional OCR and icon recognition models to assist with element grounding, which may increase the complexity and inference latency of the agent system. Overall, the current VLMs exhibit limited ability in handling GUI grounding tasks, thus constrains the navigation capabilities of single-stage vision-based GUI agents. Related works on base models are discussed in Sec. 9." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 431, + 406, + 443 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 431, + 406, + 443 + ], + "spans": [ + { + "bbox": [ + 313, + 431, + 406, + 443 + ], + "type": "text", + "content": "3. Data collection" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 453, + 555, + 524 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 453, + 555, + 524 + ], + "spans": [ + { + "bbox": [ + 313, + 453, + 555, + 524 + ], + "type": "text", + "content": "In this section, we present a cost-effective data collection strategy designed to construct a multi-level, large-scale and high-quality GUI dataset, called GUI-Lasagne, as shown in Fig. 4. This dataset helps equip our models with robust abilities in GUI understanding, grounding, and navigation. More details of the collection can be seen in Tab. 8." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 533, + 488, + 547 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 488, + 547 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 488, + 547 + ], + "type": "text", + "content": "3.1. Level one: visual-text alignment" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 553, + 555, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 553, + 555, + 673 + ], + "spans": [ + { + "bbox": [ + 313, + 553, + 555, + 673 + ], + "type": "text", + "content": "Visual-text alignment refers to the foundational ability of a GUI agent to perceive the visual appearance of the GUI, specifically the ability to recognize or locate text/icon elements. We collect visual-text data directly from the GUI source data. In the web scenario, we obtain web URLs from CommonCrawl [16] and website ranking sources12. We then develop a data collection tool using Playwright3 to gather real-world web data from the collected URLs. With this tool, we collect 755K web-page screenshots alone with their DOM data, featuring diversity in both resolution" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 325, + 683, + 506, + 693 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 683, + 506, + 693 + ], + "spans": [ + { + "bbox": [ + 325, + 683, + 506, + 693 + ], + "type": "text", + "content": "1 https://www.semrush.com/trending-sites/global/all" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 326, + 693, + 416, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 693, + 416, + 703 + ], + "spans": [ + { + "bbox": [ + 326, + 693, + 416, + 703 + ], + "type": "text", + "content": "2https://top.chinaz.com/all/" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 326, + 703, + 405, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 703, + 405, + 712 + ], + "spans": [ + { + "bbox": [ + 326, + 703, + 405, + 712 + ], + "type": "text", + "content": "3https://playwright.dev/" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "29492" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 70, + 203, + 178 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 203, + 178 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 203, + 178 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 203, + 178 + ], + "type": "image", + "image_path": "0927b7c078b46c951d58911de6fd71473ddd4b354468d91ba93747d2cced5ce5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 209, + 70, + 318, + 178 + ], + "blocks": [ + { + "bbox": [ + 209, + 70, + 318, + 178 + ], + "lines": [ + { + "bbox": [ + 209, + 70, + 318, + 178 + ], + "spans": [ + { + "bbox": [ + 209, + 70, + 318, + 178 + ], + "type": "image", + "image_path": "2564866784a69ce115cdfd5f7a88f4490320c4b703aaf837dfd8169147627e27.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 323, + 70, + 429, + 178 + ], + "blocks": [ + { + "bbox": [ + 323, + 70, + 429, + 178 + ], + "lines": [ + { + "bbox": [ + 323, + 70, + 429, + 178 + ], + "spans": [ + { + "bbox": [ + 323, + 70, + 429, + 178 + ], + "type": "image", + "image_path": "774c4ba630de293c2b7f03ca4b633358f6d9101666ec5a6833fd574cd95b8a47.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 433, + 70, + 550, + 178 + ], + "blocks": [ + { + "bbox": [ + 433, + 70, + 550, + 178 + ], + "lines": [ + { + "bbox": [ + 433, + 70, + 550, + 178 + ], + "spans": [ + { + "bbox": [ + 433, + 70, + 550, + 178 + ], + "type": "image", + "image_path": "bae001d24c286e962856c041e62532013e97a9e312903b28958e301856b9f47a.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 56, + 184, + 207, + 260 + ], + "blocks": [ + { + "bbox": [ + 56, + 184, + 207, + 260 + ], + "lines": [ + { + "bbox": [ + 56, + 184, + 207, + 260 + ], + "spans": [ + { + "bbox": [ + 56, + 184, + 207, + 260 + ], + "type": "image", + "image_path": "87b3fde8e17abff1c0705f8b1dda6c3b9ca815b7e2d1b2ab380d1840b7fd35f7.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 272, + 555, + 295 + ], + "lines": [ + { + "bbox": [ + 55, + 272, + 555, + 295 + ], + "spans": [ + { + "bbox": [ + 55, + 272, + 555, + 295 + ], + "type": "text", + "content": "Figure 4. The collection pipeline of our GUILasagne dataset. The left, middle and right parts show the construction of level-1, level-2, and level-3 data, respectively." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 209, + 184, + 317, + 260 + ], + "blocks": [ + { + "bbox": [ + 209, + 184, + 317, + 260 + ], + "lines": [ + { + "bbox": [ + 209, + 184, + 317, + 260 + ], + "spans": [ + { + "bbox": [ + 209, + 184, + 317, + 260 + ], + "type": "image", + "image_path": "0b0aa35a26767ccecb81fe588da766a58912b3e01bb0615d80aae061d27feb8d.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 323, + 184, + 426, + 260 + ], + "blocks": [ + { + "bbox": [ + 323, + 184, + 426, + 260 + ], + "lines": [ + { + "bbox": [ + 323, + 184, + 426, + 260 + ], + "spans": [ + { + "bbox": [ + 323, + 184, + 426, + 260 + ], + "type": "image", + "image_path": "e545eec66bfd52478f2e4021d4e397d473917c19ef7317eb5798852a8a267c77.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 434, + 184, + 550, + 260 + ], + "blocks": [ + { + "bbox": [ + 434, + 184, + 550, + 260 + ], + "lines": [ + { + "bbox": [ + 434, + 184, + 550, + 260 + ], + "spans": [ + { + "bbox": [ + 434, + 184, + 550, + 260 + ], + "type": "image", + "image_path": "df77e095b408b7bd199e9f9e8e2f178bbca0cffc95922f945c97cbc56c1963df.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 316, + 295, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 316, + 295, + 411 + ], + "spans": [ + { + "bbox": [ + 54, + 316, + 295, + 411 + ], + "type": "text", + "content": "and language. Additionally, we develop an icon recognition model called InternVL-Icon as an icon annotation tool, by constructing a dataset of 30K icon-caption pairs from Alibaba Icon Library" + }, + { + "bbox": [ + 54, + 316, + 295, + 411 + ], + "type": "inline_equation", + "content": "^4" + }, + { + "bbox": [ + 54, + 316, + 295, + 411 + ], + "type": "text", + "content": " and fine-tuning InternVL1.5-26B using this dataset. We then use InternVL-Icon to annotate all icons on web-pages with generated captions. As for the mobile scenario, we collect data from AitW [46], a large-scale GUI dataset on the mobile platform." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 413, + 296, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 413, + 296, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 413, + 296, + 689 + ], + "type": "text", + "content": "Inspired by Lee et al. [27] and Cheng et al. [11], we construct three tasks based the collected GUI source data: text2bbox,_bbox2text, and_bbox2dom. The text2bbox task prompts the model to ground the element based on the given text or icon caption. To avoid ambiguity, we provide additional contextual information for elements that appear multiple times in the screenshots. The text2bbox data is the most abundant among the three tasks, to help the model develop robust grounding capabilities. The_bbox2text task is the inverse of the text2bbox task, training the model to recognize text and icons. The_bbox2dom task asks the model to generate a DOM-tree corresponding to the given bounding box, as show in Sec. 12.7. This task helps the model learn not only to recognize text/icon elements but also to understand the GUI layout. To make sufficient use of the context length of the model, we pack multiple data pairs in each training sample for text2bbox and_bbox2text task, and select the box that includes as many elements as possible for_bbox2dom task. Overall, we construct a total of 1.9M and 1.1M training samples for the web and mobile scenarios, respectively. The data significantly enhance the GUI foundational abilities, especially the GUI grounding ability, of our SpiritSight agent. See Sec. 14 for more details on" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 316, + 400, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 316, + 400, + 327 + ], + "spans": [ + { + "bbox": [ + 314, + 316, + 400, + 327 + ], + "type": "text", + "content": "sample construction." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 335, + 509, + 347 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 335, + 509, + 347 + ], + "spans": [ + { + "bbox": [ + 313, + 335, + 509, + 347 + ], + "type": "text", + "content": "3.2. Level two: vision-function alignment" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 353, + 555, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 353, + 555, + 581 + ], + "spans": [ + { + "bbox": [ + 313, + 353, + 555, + 581 + ], + "type": "text", + "content": "Vision-function alignment refers to a model's ability to locate an element based on its function. This type of data is not directly accessible from the raw GUI data. Inspired by the back-translation method [48] for dataset collection in translation tasks, we leverage InternVL's image understanding capabilities to collect function grounding data. Specifically, we divide each screenshot into a 3x3 grid and describe the approximate location of the target element in text (e.g. in the top-left corner of the image). Additionally, we place a bounding box around the target element in the screenshot to specify its precise location. We provide InternVL2-26B with the screenshot, the element's text content or icon caption, and the location description to prompt it to generate the element's function. Additionally, we utilize InternLM2.5-20B to enhance the quality and diversity of the generated function descriptions. These function descriptions achieve an approximate " + }, + { + "bbox": [ + 313, + 353, + 555, + 581 + ], + "type": "inline_equation", + "content": "90.9\\%" + }, + { + "bbox": [ + 313, + 353, + 555, + 581 + ], + "type": "text", + "content": " acceptance rate based on human judgment, which we consider sufficient for constructing the function grounding pre-training data." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 581, + 555, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 581, + 555, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 581, + 555, + 665 + ], + "type": "text", + "content": "Based on the strategy above, we collect function2bbox pairs for all interactive elements in the web scenario. In the mobile scenario, we construct the function grounding data from the GUI navigation dataset, which is be described in Sec. 3.3. We use the same packing method as in the text2bbox and_bbox2text tasks for efficient training, ultimately obtaining 1.5M training samples." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 671, + 499, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 671, + 499, + 684 + ], + "spans": [ + { + "bbox": [ + 313, + 671, + 499, + 684 + ], + "type": "text", + "content": "3.3. Level three: visual GUI navigation" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "content": "We utilize the public available AitW [46] dataset to construct our GUI navigation training data. AitW is a large-" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 66, + 702, + 196, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 702, + 196, + 712 + ], + "spans": [ + { + "bbox": [ + 66, + 702, + 196, + 712 + ], + "type": "text", + "content": "4https://www.iconfont.cn/?v=20230914" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "29493" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 227 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 227 + ], + "type": "text", + "content": "scale mobile navigation dataset where each screenshot is labeled with the corresponding goal, the current step, etc. However, as noted in AitZ [72] and AMEX [5], the AitW dataset contains a number of incorrectly labeled samples. We choose to clean it with GPT-4o and employ Chain-of-Thought (CoT) [60] to make the judgment more accurate. Specifically, we prompt GPT-4o with the task description, the current action annotation, and screenshots from both the current and next steps. GPT-4o is then instructed to first summarize the two screenshots and identify the differences between them, then describe the current step based on these differences, and finally assess reasonability of the current action annotation." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 228, + 296, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 228, + 296, + 312 + ], + "spans": [ + { + "bbox": [ + 55, + 228, + 296, + 312 + ], + "type": "text", + "content": "We filter out steps identified as unreasonable by GPT-4o, resulting a final dataset of 0.64M CoT-style GUI navigation training samples. With the collected CoT-style data, we are able to construct additional function grounding data for the mobile scenario, as mentioned in Sec. 3.2, by treating the collected step descriptions as functions of the corresponding elements." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 323, + 170, + 336 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 323, + 170, + 336 + ], + "spans": [ + { + "bbox": [ + 55, + 323, + 170, + 336 + ], + "type": "text", + "content": "3.4. Other training data" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 342, + 296, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 342, + 296, + 403 + ], + "spans": [ + { + "bbox": [ + 55, + 342, + 296, + 403 + ], + "type": "text", + "content": "To enhance the model's understanding of GUI content, we further collected some public datasets as a supplement, including doc/web/mobile VQA [7, 8, 20, 40], image captioning [12, 54], and mobile grounding datasets [12, 33], resulting in 0.59M QA pairs for model training." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 416, + 192, + 430 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 416, + 192, + 430 + ], + "spans": [ + { + "bbox": [ + 55, + 416, + 192, + 430 + ], + "type": "text", + "content": "4. Universal block parsing" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 437, + 167, + 449 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 437, + 167, + 449 + ], + "spans": [ + { + "bbox": [ + 55, + 437, + 167, + 449 + ], + "type": "text", + "content": "4.1. Problem statement" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 456, + 296, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 456, + 296, + 529 + ], + "spans": [ + { + "bbox": [ + 55, + 456, + 296, + 529 + ], + "type": "text", + "content": "We use InternVL2 [9] as the base model of SpiritSight, which uses a dynamic high-resolution training approach to largely preserves the details of input images. The approach first matches the optimal aspect ratio from a pre-defined set of aspect ratios to each input image, then resize and divide each image into blocks of " + }, + { + "bbox": [ + 55, + 456, + 296, + 529 + ], + "type": "inline_equation", + "content": "448 \\times 448" + }, + { + "bbox": [ + 55, + 456, + 296, + 529 + ], + "type": "text", + "content": " pixels." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 530, + 296, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 530, + 296, + 673 + ], + "spans": [ + { + "bbox": [ + 55, + 530, + 296, + 673 + ], + "type": "text", + "content": "However, this approach may introduce ambiguity in grounding GUI elements. A typical example is represented in Fig. 5a, with two screenshots having aspect ratios of 1:2 and 2:1, respectively. According to the dynamic high-resolution strategy, the screenshots are divided into two blocks, one vertically and the other horizontally. Suppose the target elements in each sample are located in the same relative position within the second block (block-1) after cropping. The cropped screenshots are then fed into InternVL and the model is expected to predict different locations from these two identical inputs. We refer to this situation as the location ambiguity problem." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 674, + 295, + 698 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 674, + 295, + 698 + ], + "spans": [ + { + "bbox": [ + 55, + 674, + 295, + 698 + ], + "type": "text", + "content": "Generally, a point " + }, + { + "bbox": [ + 55, + 674, + 295, + 698 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 55, + 674, + 295, + 698 + ], + "type": "text", + "content": " is expressed in the global coordinate system as" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 153, + 700, + 295, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 153, + 700, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 153, + 700, + 295, + 715 + ], + "type": "interline_equation", + "content": "\\mathbf {p} = [ x, y ] \\tag {1}", + "image_path": "f52968cce278343089cc397441cad4614d4e7789085e7e9e2a722aed24796f44.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "text", + "content": " represent the horizontal and vertical coordinate values of the point in the original image, respectively. Assume that the image is resized and divided into " + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "inline_equation", + "content": "n_w \\times n_h" + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "text", + "content": " blocks. We can express the point " + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "text", + "content": " in the corresponding block as " + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "inline_equation", + "content": "\\mathbf{p}'" + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 348, + 139, + 554, + 172 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 348, + 139, + 554, + 172 + ], + "spans": [ + { + "bbox": [ + 348, + 139, + 554, + 172 + ], + "type": "interline_equation", + "content": "\\left\\{ \\begin{array}{l} b _ {x} = \\left\\lfloor \\frac {x}{w _ {\\text {b l o c k}}} \\right\\rfloor \\\\ b _ {y} = \\left\\lfloor \\frac {y}{h _ {\\text {b l o c k}}} \\right\\rfloor \\end{array} \\quad \\left\\{ \\begin{array}{l} x ^ {\\prime} = x \\bmod w _ {\\text {b l o c k}} \\\\ y ^ {\\prime} = y \\bmod h _ {\\text {b l o c k}} \\end{array} \\right. \\right. \\tag {2}", + "image_path": "9d8ec01a93529d5faa4297b6fd7ef01deabe8f24d6dc41ba6493e2e6d38a0630.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 394, + 178, + 554, + 193 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 394, + 178, + 554, + 193 + ], + "spans": [ + { + "bbox": [ + 394, + 178, + 554, + 193 + ], + "type": "interline_equation", + "content": "\\mathbf {p} ^ {\\prime} = \\left[ b _ {x}, b _ {y}, x ^ {\\prime}, y ^ {\\prime} \\right] \\tag {3}", + "image_path": "720b13a9ba1999b7a78a996b52ee3d93c4e25d35b450413ab97a1fdfbf349d83.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 198, + 555, + 269 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 198, + 555, + 269 + ], + "spans": [ + { + "bbox": [ + 313, + 198, + 555, + 269 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 198, + 555, + 269 + ], + "type": "inline_equation", + "content": "w_{block}" + }, + { + "bbox": [ + 313, + 198, + 555, + 269 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 198, + 555, + 269 + ], + "type": "inline_equation", + "content": "h_{block}" + }, + { + "bbox": [ + 313, + 198, + 555, + 269 + ], + "type": "text", + "content": " (both set to 448 in our experiments) represent the width and height of each block, respectively. " + }, + { + "bbox": [ + 313, + 198, + 555, + 269 + ], + "type": "inline_equation", + "content": "b_{x}" + }, + { + "bbox": [ + 313, + 198, + 555, + 269 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 198, + 555, + 269 + ], + "type": "inline_equation", + "content": "b_{y}" + }, + { + "bbox": [ + 313, + 198, + 555, + 269 + ], + "type": "text", + "content": " represent the horizontal and vertical indices of the block containing " + }, + { + "bbox": [ + 313, + 198, + 555, + 269 + ], + "type": "inline_equation", + "content": "\\mathbf{p}'" + }, + { + "bbox": [ + 313, + 198, + 555, + 269 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 198, + 555, + 269 + ], + "type": "inline_equation", + "content": "x'" + }, + { + "bbox": [ + 313, + 198, + 555, + 269 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 198, + 555, + 269 + ], + "type": "inline_equation", + "content": "y'" + }, + { + "bbox": [ + 313, + 198, + 555, + 269 + ], + "type": "text", + "content": " represent the coordinates of " + }, + { + "bbox": [ + 313, + 198, + 555, + 269 + ], + "type": "inline_equation", + "content": "\\mathbf{p}'" + }, + { + "bbox": [ + 313, + 198, + 555, + 269 + ], + "type": "text", + "content": " within this block. The blocks are flattened into a sequence before being fed into the model, thus we have" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 400, + 277, + 554, + 291 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 400, + 277, + 554, + 291 + ], + "spans": [ + { + "bbox": [ + 400, + 277, + 554, + 291 + ], + "type": "interline_equation", + "content": "\\mathbf {p} ^ {\\prime \\prime} = \\left[ b _ {i}, x ^ {\\prime}, y ^ {\\prime} \\right] \\tag {4}", + "image_path": "acdd176369a2c1b22b5e78ad85f16fb2f5ebbf75b910cb0bb15e45ca51cfc4e4.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 300, + 554, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 300, + 554, + 323 + ], + "spans": [ + { + "bbox": [ + 313, + 300, + 554, + 323 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 300, + 554, + 323 + ], + "type": "inline_equation", + "content": "b_{i}" + }, + { + "bbox": [ + 313, + 300, + 554, + 323 + ], + "type": "text", + "content": " represents the index of the block within the sequence, satisfying:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 397, + 333, + 553, + 346 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 397, + 333, + 553, + 346 + ], + "spans": [ + { + "bbox": [ + 397, + 333, + 553, + 346 + ], + "type": "interline_equation", + "content": "b _ {i} = b _ {y} \\cdot n _ {w} + b _ {x} \\tag {5}", + "image_path": "86b4fdcc2e46640ef1d399a75d13ffca4d418b27fab523ebfcdca3c6cd90676a.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 354, + 554, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 354, + 554, + 415 + ], + "spans": [ + { + "bbox": [ + 313, + 354, + 554, + 415 + ], + "type": "text", + "content": "The model is trained to approximate a mapping " + }, + { + "bbox": [ + 313, + 354, + 554, + 415 + ], + "type": "inline_equation", + "content": "f: \\mathbf{p}'' \\to \\mathbf{p}" + }, + { + "bbox": [ + 313, + 354, + 554, + 415 + ], + "type": "text", + "content": ", which is inherently a multi-valued function. For example, when the input " + }, + { + "bbox": [ + 313, + 354, + 554, + 415 + ], + "type": "inline_equation", + "content": "\\mathbf{p}''" + }, + { + "bbox": [ + 313, + 354, + 554, + 415 + ], + "type": "text", + "content": " is [1, 168, 245], possible values for " + }, + { + "bbox": [ + 313, + 354, + 554, + 415 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 313, + 354, + 554, + 415 + ], + "type": "text", + "content": " include [168, 693] when " + }, + { + "bbox": [ + 313, + 354, + 554, + 415 + ], + "type": "inline_equation", + "content": "n_w = 1" + }, + { + "bbox": [ + 313, + 354, + 554, + 415 + ], + "type": "text", + "content": ", or [616, 245] when " + }, + { + "bbox": [ + 313, + 354, + 554, + 415 + ], + "type": "inline_equation", + "content": "n_w = 2" + }, + { + "bbox": [ + 313, + 354, + 554, + 415 + ], + "type": "text", + "content": ", as shown in Fig. 5a." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 314, + 421, + 373, + 432 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 421, + 373, + 432 + ], + "spans": [ + { + "bbox": [ + 314, + 421, + 373, + 432 + ], + "type": "text", + "content": "4.2. Method" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 439, + 555, + 631 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 439, + 555, + 631 + ], + "spans": [ + { + "bbox": [ + 313, + 439, + 555, + 631 + ], + "type": "text", + "content": "The ambiguity primarily arises from the flattening operation, as shown in Eq. (5), which results in the loss of the spatial relationship between blocks. One solution is to input an additional thumbnail, but this may lead to extra computational and memory overhead. We propose to solve this positional ambiguity with two steps. Firstly, we introduce a 2D Block-wise Position Embedding (2D-BPE) [67] by adding two position embeddings to each block to capture spatial information. Secondly, we introduce a Universal Block Parsing (UBP) method, which replaces the global coordinate representation with a block-specific coordinate representation. Specifically, we express a point as Eq. (4). In this case, the model is trained to approximate a injective mapping " + }, + { + "bbox": [ + 313, + 439, + 555, + 631 + ], + "type": "inline_equation", + "content": "f: \\mathbf{p}'' \\rightarrow \\mathbf{p}''" + }, + { + "bbox": [ + 313, + 439, + 555, + 631 + ], + "type": "text", + "content": ", thereby resolving the ambiguity problem. During the model inference, the global coordinate of this point can be computed in post-processing as follows:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 358, + 639, + 553, + 672 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 639, + 553, + 672 + ], + "spans": [ + { + "bbox": [ + 358, + 639, + 553, + 672 + ], + "type": "interline_equation", + "content": "\\left\\{ \\begin{array}{l} x = x ^ {\\prime} + \\left(b _ {i} \\mod n _ {w}\\right) \\cdot w _ {\\text {b l o c k}} \\\\ y = y ^ {\\prime} + \\left\\lfloor \\frac {b _ {i}}{n _ {w}} \\right\\rfloor \\cdot h _ {\\text {b l o c k}} \\end{array} \\right. \\tag {6}", + "image_path": "b7d375ff6c6b88b22381c8ed327be26c6749db4e1e387e6a3e5282abeb03178e.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "text", + "content": "Overall, our UBP method ensures a clear mapping of positional information between the model's inputs and outputs, thereby improving the model's grounding capability." + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "29494" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 70, + 358, + 197 + ], + "blocks": [ + { + "bbox": [ + 59, + 70, + 358, + 197 + ], + "lines": [ + { + "bbox": [ + 59, + 70, + 358, + 197 + ], + "spans": [ + { + "bbox": [ + 59, + 70, + 358, + 197 + ], + "type": "image", + "image_path": "6cbf1f56b2c8862f6675f16b75ec849427e015e45df37a8cbb3645b5929d639e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 117, + 198, + 295, + 209 + ], + "lines": [ + { + "bbox": [ + 117, + 198, + 295, + 209 + ], + "spans": [ + { + "bbox": [ + 117, + 198, + 295, + 209 + ], + "type": "text", + "content": "(a) Baseline Block Parsing vs Universal Block Parsing." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 377, + 71, + 550, + 198 + ], + "blocks": [ + { + "bbox": [ + 377, + 71, + 550, + 198 + ], + "lines": [ + { + "bbox": [ + 377, + 71, + 550, + 198 + ], + "spans": [ + { + "bbox": [ + 377, + 71, + 550, + 198 + ], + "type": "image", + "image_path": "01113d42e69228e6e352f1806bb26775fd8ff3a752d69a993076a66ed6e57181.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 414, + 199, + 518, + 209 + ], + "lines": [ + { + "bbox": [ + 414, + 199, + 518, + 209 + ], + "spans": [ + { + "bbox": [ + 414, + 199, + 518, + 209 + ], + "type": "text", + "content": "(b) Performance gain with UBP." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 218, + 555, + 251 + ], + "lines": [ + { + "bbox": [ + 55, + 218, + 555, + 251 + ], + "spans": [ + { + "bbox": [ + 55, + 218, + 555, + 251 + ], + "type": "text", + "content": "Figure 5. (a) Comparison between baseline block parsing and our proposed UBP. (b) The results of baseline block parsing and our proposed UBP methods on Multimodal-Mind2Web benchmark. UBP improves the performance of our model. The combination of UBP and 2D-BPE achieves the best results." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 271, + 111, + 285 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 271, + 111, + 285 + ], + "spans": [ + { + "bbox": [ + 55, + 271, + 111, + 285 + ], + "type": "text", + "content": "5. Settings" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 300, + 186, + 313 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 300, + 186, + 313 + ], + "spans": [ + { + "bbox": [ + 55, + 300, + 186, + 313 + ], + "type": "text", + "content": "5.1. Implementation details" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 327, + 296, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 327, + 296, + 472 + ], + "spans": [ + { + "bbox": [ + 55, + 327, + 296, + 472 + ], + "type": "text", + "content": "We use InternVL2 (2B, 8B and 26B) [9] as the base models and train them with two stages: continual pre-training and fine-tuning. During the pre-training stage, we use all the GUI-Lasagne dataset mentioned in Sec. 3. Different prompts are designed for different training tasks to avoid task confusion. See Sec. 14 for the prompts. We unfreeze the vision encoder, decoder, and MLP layer of InternVL. The learning rate is set to " + }, + { + "bbox": [ + 55, + 327, + 296, + 472 + ], + "type": "inline_equation", + "content": "1\\mathrm{e} - 4 / 1\\mathrm{e} - 4 / 5\\mathrm{e} - 5" + }, + { + "bbox": [ + 55, + 327, + 296, + 472 + ], + "type": "text", + "content": " for InternVL-2B/8B/26B, respectively, and the batch size is set to 1024. We get SpiritSight-Base model after pre-training and then fine-tune it on multiple downstream tasks individually. More details are shown in Sec. 12.2 and Sec. 12.3." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 506, + 187, + 517 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 506, + 187, + 517 + ], + "spans": [ + { + "bbox": [ + 55, + 506, + 187, + 517 + ], + "type": "text", + "content": "5.2. Benchmarks & metrics" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 533, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 533, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 533, + 296, + 713 + ], + "type": "text", + "content": "To assess SpiritSight's capability in diverse real-world environments, we evaluate SpiritSight on six benchmarks covering various GUI platforms and tasks. Multimodal-Mind2Web [13], ScreenSpot [11], AMEX [5], GUIAct [7], AndroidControl [32], and GUI-Odyssey [38]. For Multimodal-Mind2Web and ScreenSpot, we use the same data pre-processing and metrics as SeeClick [11] uses. For GUIAct, we evaluate SpiritSight on the web-single and web-multi sub-sets and report step success rate (Step SR). For AndroidControl, we evaluate SpiritSight on the High-Level (HL) and Low-Level (LL) tasks and report step accuracy. For AMEX and GUI-Odyssey, we report the action matching score (AMS) defined in AitW [46]. These metrics are similar in that they all measure the single-step accuracy. Refer to Sec. 12.4 for more information about the metrics." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 271, + 390, + 285 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 271, + 390, + 285 + ], + "spans": [ + { + "bbox": [ + 314, + 271, + 390, + 285 + ], + "type": "text", + "content": "6. Experiment" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 292, + 496, + 305 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 292, + 496, + 305 + ], + "spans": [ + { + "bbox": [ + 313, + 292, + 496, + 305 + ], + "type": "text", + "content": "6.1. Advanced vision-based GUI agent" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 311, + 555, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 311, + 555, + 418 + ], + "spans": [ + { + "bbox": [ + 313, + 311, + 555, + 418 + ], + "type": "text", + "content": "In this section, we compare the performance of SpiritSight with other advanced methods across various input modalities and test configurations on Multimodal-Mind2Web [13], a classic and high-quality benchmark for GUI navigation. The results are shown in Tab. 1. Methods that use top-50 candidates as input perform the best. This is evident, as the assistance of candidate elements can significantly reduce the decision space. However, such methods are not particularly feasible in practice." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 419, + 556, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 419, + 556, + 540 + ], + "spans": [ + { + "bbox": [ + 313, + 419, + 556, + 540 + ], + "type": "text", + "content": "The results indicate that SpiritSight significantly outperforms all methods that do not rely on any candidate element, including vision-based methods, language-based methods, and even vision-language-based methods. This demonstrates strong capabilities of SpiritSight in Web GUI navigation tasks. It should be noted that SpiritSight achieves a significant advantage in the Ele.Acc metric compared to other vision-based methods, which can be attributed to the specially constructed visual grounding training data and the proposed UBP approach." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 550, + 504, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 550, + 504, + 563 + ], + "spans": [ + { + "bbox": [ + 313, + 550, + 504, + 563 + ], + "type": "text", + "content": "6.2. Strong cross-platform compatibility" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 568, + 554, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 568, + 554, + 663 + ], + "spans": [ + { + "bbox": [ + 313, + 568, + 554, + 663 + ], + "type": "text", + "content": "We evaluated SpiritSight on other GUI navigation benchmarks across various platforms and compare it with state-of-the-art (SOTA) GUI agents as shown in Tab. 2. SpiritSight demonstrated the best performance on these benchmarks, showing its strong capabilities across various platforms. OdysseyAgent uses additional historical screenshot images as input, yet SpiritSight achieves comparable results." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 665, + 554, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 554, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 554, + 715 + ], + "type": "text", + "content": "We also evaluated SpiritSight on ScreenSpot, a function grounding benchmark. As shown in Tab. 3, SpiritSight performs well across all three platforms, showcasing its cross-platform capability and robust grounding ability. Notably," + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "29495" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 71, + 295, + 161 + ], + "blocks": [ + { + "bbox": [ + 56, + 71, + 295, + 161 + ], + "lines": [ + { + "bbox": [ + 56, + 71, + 295, + 161 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 295, + 161 + ], + "type": "table", + "html": "
AgentOdysseyAMEXAndroidCtrlGUIAct
HighHighHighLowMultiSingle
GPT-4o[42]20.4%-21.2%28.4%-41.8%
Previous SOTA74.3%70.7%64.8%80.0%45.4%74.9%
SpiritSight-2B72.3%74.5%64.9%86.3%45.5%76.0%
SpiritSight-8B75.8%80.7%68.1%87.6%49.3%78.2%
SpiritSight-26B75.9%82.1%69.6%88.2%52.6%78.3%
", + "image_path": "b21c1a4c11910f953d1c28d865588f2e33acd71851290ef8fed6c506a1044270.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 267, + 295, + 352 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 267, + 295, + 352 + ], + "spans": [ + { + "bbox": [ + 55, + 267, + 295, + 352 + ], + "type": "text", + "content": "SpiritSight models are trained specifically for GUI navigation tasks, and no training data aligned with ScreenSpot is used. There remains room for further improvement in SpiritSight's performance on the ScreenSpot benchmark. We further assess the foundational ability of SpiritSight, the visual grounding ability, on our custom-constructed testing dataset. See Sec. 12.7 for more details." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 359, + 295, + 385 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 359, + 295, + 385 + ], + "spans": [ + { + "bbox": [ + 55, + 359, + 295, + 385 + ], + "type": "text", + "content": "6.3. Recognition and grounding as priors for GUI navigation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 390, + 295, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 390, + 295, + 556 + ], + "spans": [ + { + "bbox": [ + 55, + 390, + 295, + 556 + ], + "type": "text", + "content": "To verify the significance of the three levels of GUI-Lasagne data, we progressively removed level-3, level-2, and level-1 data from the training set during the pretraining stage and evaluate SpiritSight-8B on MultimodalMind2Web. The results are shown by the blue line in Fig. 6a. It can be seen that each level of data contributes to improving Step SR. While the tasks in level-1 data differ the most from web navigation task compared to the other two levels, they provide an effective initialization for the pre-trained model. Although the level-3 data is constructed from the mobile scenario, it also assists in web-based GUI navigation tasks. This indicates that the joint learning strategy helps SpiritSight develop strong navigation abilities across different GUI environments with limited resources." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 558, + 295, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 295, + 689 + ], + "type": "text", + "content": "We also conducted ablation experiments to evaluate the effectiveness of data cleaning and CoT construction strategies on the level-3 data, as shown by the orange line in Fig. 6a. It can be observed that training in a CoT manner effectively improves the model's GUI navigation capabilities, while the impact of data cleaning strategy is relatively small. This may be due to the fact that erroneous samples are relatively tolerated during the pre-training phase of large models. The difference in results of \"w/ CoT\" and \"w/ Level " + }, + { + "bbox": [ + 55, + 558, + 295, + 689 + ], + "type": "inline_equation", + "content": "1 + 2 + 3" + }, + { + "bbox": [ + 55, + 558, + 295, + 689 + ], + "type": "text", + "content": " \" is due to the different implementation for fine-tuning stage. See Sec. 12.3 for more details." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "content": "To evaluate the effectiveness of our GUI-Lasagne dataset, we train InternVL-2 (8B) on SeeClick train" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 323, + 70, + 547, + 198 + ], + "blocks": [ + { + "bbox": [ + 55, + 179, + 295, + 247 + ], + "lines": [ + { + "bbox": [ + 55, + 179, + 295, + 247 + ], + "spans": [ + { + "bbox": [ + 55, + 179, + 295, + 247 + ], + "type": "text", + "content": "Table 2. Results of SpiritSight on four recently published GUI navigation benchmarks. 'High' and 'Multi' indicates high-level tasks, 'Low' and 'Single' indicates single-stepped tasks. The SOTA results of GUI-Odyssey, AMEX, AndroidControl and GUIAct come from OdysseyAgent [38], SPHINX-GUI-Agent [5], fine-tuned PaLM-2S [32] and MiniCPM-GUI [7], respectively." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 323, + 70, + 547, + 198 + ], + "lines": [ + { + "bbox": [ + 323, + 70, + 547, + 198 + ], + "spans": [ + { + "bbox": [ + 323, + 70, + 547, + 198 + ], + "type": "table", + "html": "
AgentModel SizeScreenSpot
WebMobileDesktop
GPT4V[1]-5.0%7.5%4.6%
Qwen-VL[3]9.6B3.0%7.2%5.4%
Fuyu[4]8B19.2%21.2%18.3%
CogAgent[19]18B49.5%45.5%47.1%
SeeClick[11]9.6B44.1%65.0%51.1%
SpiritSight-2B2B63.6%62.5%61.8%
SpiritSight-8B8B68.3%68.4%62.9%
SpiritSight-26B26B74.4%73.6%69.9%
", + "image_path": "bca23e087cb4a97f93000c8cb832a340996d071fd12aab186b51f873a4b5138f.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 215, + 553, + 237 + ], + "lines": [ + { + "bbox": [ + 313, + 215, + 553, + 237 + ], + "spans": [ + { + "bbox": [ + 313, + 215, + 553, + 237 + ], + "type": "text", + "content": "Table 3. Results of SpiritSight and other vision-based methods on ScreenSpot Benchmark." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 256, + 555, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 256, + 555, + 460 + ], + "spans": [ + { + "bbox": [ + 313, + 256, + 555, + 460 + ], + "type": "text", + "content": "ing data [11] using full parameter training and LoRA training, resulting in SeeClick(InternVL-Full) and SeeClick(InternVL-LoRA), respectively. We then fine-tune and evaluate the two models on the Multimodal-Mind2Web benchmark. The results are shown by the dashed line in Fig. 6a, where we also present the result of the original SeeClick model. SeeClick(InternVL-LoRA) performs better than the original SeeClick model, which may be attributed to the dynamic high-resolution approach used by InternVL that preserves the details of input images. SeeClick(InternVL-Full) performs the worst, as the scale of the SeeClick training data is not large enough for the model to converge during full parameter training. SpiritSight agent that trained on only GUI-Lasagne level-1 data outperforms all three models trained on SeeClick data, demonstrating the effectiveness of GUI-Lasagne dataset. More ablation studies can be found in Sec. 12.5." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 467, + 501, + 479 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 467, + 501, + 479 + ], + "spans": [ + { + "bbox": [ + 313, + 467, + 501, + 479 + ], + "type": "text", + "content": "6.4. Better grounding ability from UBP" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 485, + 555, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 485, + 555, + 652 + ], + "spans": [ + { + "bbox": [ + 313, + 485, + 555, + 652 + ], + "type": "text", + "content": "To verify the effectiveness of UBP on grounding task, we use LoRA for resource efficiency to fine-tune InternVL in 4 different settings (original as baseline, 2D-BPE, UBP, and 2D-BPE&UBP), respectively. We then evaluate these models on Multimodal-Mind2Web benchmark, as shown in Fig. 5b. It can be seen that UBP shows a significant advantage in Ele.Acc compared to baseline, while the results of Op.F1 show little variation across the 4 settings. This indicates that UBP improves the performance of GUI agent primarily by enhancing the grounding ability. Finally, the combination of UBP and 2D-BPE achieves the best results. This indicates that UBP is compatible with 2D-BPE, leading to better performance. More ablation studies can be found in Sec. 12.6." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 659, + 526, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 659, + 526, + 672 + ], + "spans": [ + { + "bbox": [ + 313, + 659, + 526, + 672 + ], + "type": "text", + "content": "6.5. Scaling effects on dataset and model size" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "content": "We explore the impact of pre-training dataset and model size on SpiritSight using Multimodal-Mind2Web benchmark. The results are shown in Fig. 6b. SpiritSight-2B out" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "29496" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 71, + 206, + 201 + ], + "blocks": [ + { + "bbox": [ + 58, + 71, + 206, + 201 + ], + "lines": [ + { + "bbox": [ + 58, + 71, + 206, + 201 + ], + "spans": [ + { + "bbox": [ + 58, + 71, + 206, + 201 + ], + "type": "image", + "image_path": "68cb0cf0cbde3674e8ff4c66a96a4808e3d7365dcca842f4fc647ed1a9af8a2f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 56, + 201, + 206, + 219 + ], + "lines": [ + { + "bbox": [ + 56, + 201, + 206, + 219 + ], + "spans": [ + { + "bbox": [ + 56, + 201, + 206, + 219 + ], + "type": "text", + "content": "(a) Effect of three data levels and data augmentation for level-3 data" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 100, + 230, + 507, + 241 + ], + "lines": [ + { + "bbox": [ + 100, + 230, + 507, + 241 + ], + "spans": [ + { + "bbox": [ + 100, + 230, + 507, + 241 + ], + "type": "text", + "content": "Figure 6. Ablation studies on GUILasagne dataset and scaling effects in the Multimodal-Mind2Web benchmark." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 231, + 73, + 380, + 200 + ], + "blocks": [ + { + "bbox": [ + 231, + 73, + 380, + 200 + ], + "lines": [ + { + "bbox": [ + 231, + 73, + 380, + 200 + ], + "spans": [ + { + "bbox": [ + 231, + 73, + 380, + 200 + ], + "type": "image", + "image_path": "75a9adfdde148c1f0f53837684c4722adfd282f03f634c1ff30599f5b936d3a5.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 231, + 201, + 381, + 220 + ], + "lines": [ + { + "bbox": [ + 231, + 201, + 381, + 220 + ], + "spans": [ + { + "bbox": [ + 231, + 201, + 381, + 220 + ], + "type": "text", + "content": "(b) Ablation Study: Effect of training data percentages on model performance." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 402, + 73, + 552, + 200 + ], + "blocks": [ + { + "bbox": [ + 402, + 73, + 552, + 200 + ], + "lines": [ + { + "bbox": [ + 402, + 73, + 552, + 200 + ], + "spans": [ + { + "bbox": [ + 402, + 73, + 552, + 200 + ], + "type": "image", + "image_path": "94a23edebfe12bbac891c5fae7d6da1e2d962594bc8f6112469c5f217008e905.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 402, + 201, + 553, + 219 + ], + "lines": [ + { + "bbox": [ + 402, + 201, + 553, + 219 + ], + "spans": [ + { + "bbox": [ + 402, + 201, + 553, + 219 + ], + "type": "text", + "content": "(c) Ablation Study: Effect of data percentages on downstream task." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 262, + 294, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 262, + 294, + 369 + ], + "spans": [ + { + "bbox": [ + 54, + 262, + 294, + 369 + ], + "type": "text", + "content": "performs SeeClick [11] by using just 1/8 of the pre-training dataset. This remarkable performance comes from the high quality and grounding-focus nature of GUI-Lasagne data. The performance improves as the size of dataset increases, demonstrating the significance of collecting large-scale data. SpiritSight-2B reaches saturation with a smaller amount of pre-training data, while SpiritSight-26B appears to have further potential for improvement, which aligns with the scaling law of LLMs and VLMs." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 371, + 295, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 371, + 295, + 455 + ], + "spans": [ + { + "bbox": [ + 55, + 371, + 295, + 455 + ], + "type": "text", + "content": "We also evaluate the ability of SpiritSight to transferring to downstream GUI agent tasks. We fine-tune SpiritSight-Base (8B) on various proportions of the Multimodal-Mind2Web training data and show the results in Fig. 6c. It is noticed that SpiritSight achieves " + }, + { + "bbox": [ + 55, + 371, + 295, + 455 + ], + "type": "inline_equation", + "content": "36.6\\%" + }, + { + "bbox": [ + 55, + 371, + 295, + 455 + ], + "type": "text", + "content": " Step SR with only 1/8 of the fine-tuning data, showing strong transferability to GUI navigation tasks." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 465, + 248, + 478 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 465, + 248, + 478 + ], + "spans": [ + { + "bbox": [ + 55, + 465, + 248, + 478 + ], + "type": "text", + "content": "6.6. Effective transfer to other languages" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 484, + 295, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 484, + 295, + 567 + ], + "spans": [ + { + "bbox": [ + 55, + 484, + 295, + 567 + ], + "type": "text", + "content": "Exploring the cross-lingual capabilities of GUI agents is highly beneficial for their application in non-English environments. We split the training and testing sets of GUIAct (web-multi) data into English and Chinese partitions, respectively. We then fine-tune SpiritSight-Base (8B) on two sets: the entire training set (English&Chinese) and the English-only training set. The results are shown in Tab. 4." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 569, + 295, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 569, + 295, + 688 + ], + "spans": [ + { + "bbox": [ + 55, + 569, + 295, + 688 + ], + "type": "text", + "content": "Under the English&Chinese configuration, SpiritSight achieves comparable results on both the English and Chinese testing sets, despite having fewer Chinese samples in the pre-training data than English ones. Under the English-only configuration, SpiritSight achieves " + }, + { + "bbox": [ + 55, + 569, + 295, + 688 + ], + "type": "inline_equation", + "content": "24.5\\%" + }, + { + "bbox": [ + 55, + 569, + 295, + 688 + ], + "type": "text", + "content": " Step SR on the Chinese testing set, reaching half of the English&Chinese performance. The zero-shot capability of SpiritSight in Chinese results from the small but effective foundational Chinese data included in the pre-training stage." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "content": "This experiment offers a paradigm for applying GUI agents to non-English environments: by collecting (1) free" + } + ] + } + ], + "index": 12 + }, + { + "type": "table", + "bbox": [ + 337, + 259, + 533, + 306 + ], + "blocks": [ + { + "bbox": [ + 337, + 259, + 533, + 306 + ], + "lines": [ + { + "bbox": [ + 337, + 259, + 533, + 306 + ], + "spans": [ + { + "bbox": [ + 337, + 259, + 533, + 306 + ], + "type": "table", + "html": "
SFT DataOverallChineseEnglish
English+Chinese49.3%49.3%49.2%
English35.0%24.5%48.6%
", + "image_path": "8b3064ee7455fe2f035587fe2b32c9b2ab81f40eac9b2cf673817a3f39491450.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 323, + 553, + 345 + ], + "lines": [ + { + "bbox": [ + 313, + 323, + 553, + 345 + ], + "spans": [ + { + "bbox": [ + 313, + 323, + 553, + 345 + ], + "type": "text", + "content": "Table 4. Results of SpiritSight on GUIAct (web-multi), trained with different language datasets." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 365, + 554, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 365, + 554, + 437 + ], + "spans": [ + { + "bbox": [ + 313, + 365, + 554, + 437 + ], + "type": "text", + "content": "web and mobile GUI information from the target language environment (level 1 & level 2 data), and (2) a small amount of high-quality GUI navigation data at low cost (level 3 data). With this language transferring strategy, the same capabilities as in the English environment can be achieved in non-English environments with minimal extra costs." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 314, + 448, + 384, + 460 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 448, + 384, + 460 + ], + "spans": [ + { + "bbox": [ + 314, + 448, + 384, + 460 + ], + "type": "text", + "content": "7. Limitation" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 468, + 554, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 468, + 554, + 538 + ], + "spans": [ + { + "bbox": [ + 313, + 468, + 554, + 538 + ], + "type": "text", + "content": "As SpiritSight is a vision-based GUI agent, it constantly requires access to screenshots which may contain personal information or other sensitive data. Users and system providers should carefully manage the system privileges granted to SpiritSight agent to mitigate potential privacy and security risks." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 550, + 388, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 550, + 388, + 562 + ], + "spans": [ + { + "bbox": [ + 313, + 550, + 388, + 562 + ], + "type": "text", + "content": "8. Conclusion" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "content": "In this paper, we propose an advanced vision-based end-to-end GUI agent, SpiritSight, with high generalization across multiple GUI platforms. We construct an efficient multilevel, large-scale, high-quality GUI pre-training dataset to equip SpiritSight with robust GUI perception, grounding and understanding capabilities. Additionally, we introduce a UBP method to resolve ambiguity in dynamic high-resolution inputs during model training, further enhancing the ability of SpiritSight to ground GUI objects. As a result, SpiritSight achieves strong performance in numerous GUI navigation benchmarks, demonstrating significant potential for practical deployment in real-world applications." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "29497" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "text", + "content": "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 146, + 296, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 146, + 296, + 201 + ], + "spans": [ + { + "bbox": [ + 61, + 146, + 296, + 201 + ], + "type": "text", + "content": "[2] Gilles Baechler, Srinivas Sunkara, Maria Wang, Fedir Zubach, Hassan Mansoor, Vincent Etter, Victor Carbune, Jason Lin, Jindong Chen, and Abhanshu Sharma. Screenai: A vision-language model for ui and infographics understanding. arXiv preprint arXiv:2402.04615, 2024. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 202, + 295, + 256 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 202, + 295, + 256 + ], + "spans": [ + { + "bbox": [ + 62, + 202, + 295, + 256 + ], + "type": "text", + "content": "[3] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. arXiv preprint arXiv:2308.12966, 2023.7, 1,4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 258, + 295, + 291 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 258, + 295, + 291 + ], + "spans": [ + { + "bbox": [ + 62, + 258, + 295, + 291 + ], + "type": "text", + "content": "[4] Rohan Bavishi, Erich Elsen, Curtis Hawthorne, Maxwell Nye, Augustus Odena, Arushi Somani, and Săgnak Taşürlar. Introducing our multimodal models, 2023. 3, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 292, + 295, + 345 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 292, + 295, + 345 + ], + "spans": [ + { + "bbox": [ + 62, + 292, + 295, + 345 + ], + "type": "text", + "content": "[5] Yuxiang Chai, Siyuan Huang, Yazhe Niu, Han Xiao, Liang Liu, Dingyu Zhang, Peng Gao, Shuai Ren, and Hongsheng Li. Amex: Android multi-annotation expo dataset for mobile gui agents. arXiv preprint arXiv:2407.17490, 2024. 2, 5, 6, 7, 1, 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 347, + 295, + 400 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 347, + 295, + 400 + ], + "spans": [ + { + "bbox": [ + 62, + 347, + 295, + 400 + ], + "type": "text", + "content": "[6] Dongping Chen, Yue Huang, Siyuan Wu, Jingyu Tang, Liuyi Chen, Yilin Bai, Zhigang He, Chenlong Wang, Huichi Zhou, Yiqiang Li, et al. Gui-world: A dataset for gui-oriented multimodal llm-based agents. arXiv preprint arXiv:2406.10819, 2024. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 403, + 295, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 403, + 295, + 456 + ], + "spans": [ + { + "bbox": [ + 62, + 403, + 295, + 456 + ], + "type": "text", + "content": "[7] Wentong Chen, Junbo Cui, Jinyi Hu, Yujia Qin, Junjie Fang, Yue Zhao, Chongyi Wang, Jun Liu, Guirong Chen, Yupeng Huo, et al. Guicourse: From general vision language models to versatile gui agents. arXiv preprint arXiv:2406.11317, 2024. 1, 3, 5, 6, 7, 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 458, + 295, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 458, + 295, + 502 + ], + "spans": [ + { + "bbox": [ + 62, + 458, + 295, + 502 + ], + "type": "text", + "content": "[8] Xingyu Chen, Zihan Zhao, Lu Chen, Danyang Zhang, Jiabao Ji, Ao Luo, Yuxuan Xiong, and Kai Yu. Websrc: A dataset for web-based structural reading comprehension. arXiv preprint arXiv:2101.09465, 2021. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 503, + 295, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 503, + 295, + 557 + ], + "spans": [ + { + "bbox": [ + 62, + 503, + 295, + 557 + ], + "type": "text", + "content": "[9] Zhe Chen, Weiyun Wang, Hao Tian, Shenglong Ye, Zhangwei Gao, Erfei Cui, Wenwen Tong, Kongzhi Hu, Jiapeng Luo, Zheng Ma, et al. How far are we to gpt-4v? closing the gap to commercial multimodal models with open-source suites. arXiv preprint arXiv:2404.16821, 2024. 2, 5, 6, 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 558, + 295, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 558, + 295, + 624 + ], + "spans": [ + { + "bbox": [ + 57, + 558, + 295, + 624 + ], + "type": "text", + "content": "[10] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24185–24198, 2024. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 625, + 295, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 625, + 295, + 669 + ], + "spans": [ + { + "bbox": [ + 57, + 625, + 295, + 669 + ], + "type": "text", + "content": "[11] Kanzhi Cheng, Qiushi Sun, Yougang Chu, Fangzhi Xu, Yantao Li, Jianbing Zhang, and Zhiyong Wu. Seeclick: Harnessing gui grounding for advanced visual gui agents. arXiv preprint arXiv:2401.10935, 2024. 1, 3, 4, 6, 7, 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "type": "text", + "content": "[12] Biplab Deka, Zifeng Huang, Chad Franzen, Joshua Hibschman, Daniel Afergan, Yang Li, Jeffrey Nichols, and Ranjitha Kumar. Rico: A mobile app dataset for building data-driven design applications. In Proceedings of the 30th annual" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "ACM symposium on user interface software and technology, pages 845-854, 2017. 5, 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 96, + 553, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 139 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 139 + ], + "type": "text", + "content": "[13] Xiang Deng, Yu Gu, Boyuan Zheng, Shijie Chen, Sam Stevens, Boshi Wang, Huan Sun, and Yu Su. Mind2web: Towards a generalist agent for the web. Advances in Neural Information Processing Systems, 36, 2024. 1, 2, 6, 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 140, + 553, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 140, + 553, + 171 + ], + "spans": [ + { + "bbox": [ + 317, + 140, + 553, + 171 + ], + "type": "text", + "content": "[14] Jacob Devlin. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 173, + 553, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 173, + 553, + 226 + ], + "spans": [ + { + "bbox": [ + 317, + 173, + 553, + 226 + ], + "type": "text", + "content": "[15] Hiroki Furuta, Kuang-Huei Lee, Ofir Nachum, Yutaka Matsuo, Aleksandra Faust, Shixiang Shane Gu, and Izzeddin Gur. Multimodal web navigation with instruction-finetuned foundation models. arXiv preprint arXiv:2305.11854, 2023. 1, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 228, + 553, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 228, + 553, + 250 + ], + "spans": [ + { + "bbox": [ + 316, + 228, + 553, + 250 + ], + "type": "text", + "content": "[16] Common crawl Group. Common crawl - open repository of web crawl data, 2024. 3, 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 251, + 553, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 251, + 553, + 304 + ], + "spans": [ + { + "bbox": [ + 317, + 251, + 553, + 304 + ], + "type": "text", + "content": "[17] Izzeddin Gur, Hiroki Furuta, Austin Huang, Mustafa Safdari, Yutaka Matsuo, Douglas Eck, and Aleksandra Faust. A real-world webagent with planning, long context understanding, and program synthesis. arXiv preprint arXiv:2307.12856, 2023. 2, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 305, + 553, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 305, + 553, + 350 + ], + "spans": [ + { + "bbox": [ + 316, + 305, + 553, + 350 + ], + "type": "text", + "content": "[18] Markus Hafner, Maria Katsantoni, Tino Köster, James Marks, Joyita Mukherjee, Dorothee Staiger, Jernej Ule, and Mihaela Zavolan. Clip and complementary methods. Nature Reviews Methods Primers, 1(1):1-23, 2021. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 350, + 553, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 350, + 553, + 415 + ], + "spans": [ + { + "bbox": [ + 317, + 350, + 553, + 415 + ], + "type": "text", + "content": "[19] Wenyi Hong, Weihan Wang, Qingsong Lv, Jiazheng Xu, Wenmeng Yu, Junhui Ji, Yan Wang, Zihan Wang, Yuxiao Dong, Ming Ding, et al. Cogagent: A visual language model for gui agents. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14281-14290, 2024. 1, 3, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 416, + 553, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 416, + 553, + 449 + ], + "spans": [ + { + "bbox": [ + 316, + 416, + 553, + 449 + ], + "type": "text", + "content": "[20] Yu-Chung Hsiao, Fedir Zubach, Maria Wang, et al. Screenqa: Large-scale question-answer pairs over mobile app screenshots. arXiv preprint arXiv:2209.08199, 2022. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 449, + 553, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 449, + 553, + 492 + ], + "spans": [ + { + "bbox": [ + 317, + 449, + 553, + 492 + ], + "type": "text", + "content": "[21] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 493, + 553, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 493, + 553, + 536 + ], + "spans": [ + { + "bbox": [ + 317, + 493, + 553, + 536 + ], + "type": "text", + "content": "[22] Faria Huq, Jeffrey P Bigham, and Nikolas Martelaro. “what’s important here?”: Opportunities and challenges of using llms in retrieving information from web interfaces. arXiv preprint arXiv:2312.06147, 2023. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 537, + 553, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 537, + 553, + 591 + ], + "spans": [ + { + "bbox": [ + 317, + 537, + 553, + 591 + ], + "type": "text", + "content": "[23] Yang Jin, Kun Xu, Kun Xu, Liwei Chen, Chao Liao, Jianchao Tan, Quzhe Huang, Bin Chen, Chenyi Lei, An Liu, et al. Unified language-vision pretraining in llm with dynamic discrete visual tokenization. arxiv 2024. arXiv preprint arXiv:2309.04669, 2023. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 593, + 553, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 593, + 553, + 647 + ], + "spans": [ + { + "bbox": [ + 317, + 593, + 553, + 647 + ], + "type": "text", + "content": "[24] Jihyung Kil, Chan Hee Song, Boyuan Zheng, Xiang Deng, Yu Su, and Wei-Lun Chao. Dual-view visual contextualization for web navigation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14445-14454, 2024. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 647, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 647, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 317, + 647, + 553, + 712 + ], + "type": "text", + "content": "[25] Geewook Kim, Teakgyu Hong, Moonbin Yim, Jeong Yeon Nam, Jinyoung Park, Jinyeong Yim, Wonseok Hwang, Sang-doo Yun, Dongyoon Han, and Seunghyun Park. Ocr-free document understanding transformer. In European Conference on Computer Vision, pages 498-517. Springer, 2022. 1, 2" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "29498" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "text", + "content": "[26] Hanyu Lai, Xiao Liu, Iat Long Iong, Shuntian Yao, Yuxuan Chen, Pengbo Shen, Hao Yu, Hanchen Zhang, Xiaohan Zhang, Yuxiao Dong, et al. Autowebglm: Bootstrap and reinforce a large language model-based web navigating agent. arXiv preprint arXiv:2404.03648, 2024. 1, 2, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 128, + 294, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 128, + 294, + 194 + ], + "spans": [ + { + "bbox": [ + 56, + 128, + 294, + 194 + ], + "type": "text", + "content": "[27] Kenton Lee, Mandar Joshi, Iulia Raluca Turc, Hexiang Hu, Fangyu Liu, Julian Martin Eisenschlos, Urvashi Khandelwal, Peter Shaw, Ming-Wei Chang, and Kristina Toutanova. Pix2struct: Screenshot parsing as pretraining for visual language understanding. In International Conference on Machine Learning, pages 18893–18912. PMLR, 2023. 2, 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 194, + 294, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 194, + 294, + 237 + ], + "spans": [ + { + "bbox": [ + 56, + 194, + 294, + 237 + ], + "type": "text", + "content": "[28] Kuang-Huei Lee, Xinyun Chen, Hiroki Furuta, John Canny, and Ian Fischer. A human-inspired reading agent with gist memory of very long contexts. arXiv preprint arXiv:2402.09727, 2024. 1, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 238, + 294, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 238, + 294, + 282 + ], + "spans": [ + { + "bbox": [ + 56, + 238, + 294, + 282 + ], + "type": "text", + "content": "[29] Feng Li, Renrui Zhang, Hao Zhang, Yuanhan Zhang, Bo Li, Wei Li, Zejun Ma, and Chunyuan Li. Llava next-interleave: Tackling multi-image, video, and 3d in large multimodal models. arXiv preprint arXiv:2407.07895, 2024. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 283, + 294, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 283, + 294, + 337 + ], + "spans": [ + { + "bbox": [ + 56, + 283, + 294, + 337 + ], + "type": "text", + "content": "[30] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International conference on machine learning, pages 12888-12900. PMLR, 2022. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 338, + 294, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 338, + 294, + 392 + ], + "spans": [ + { + "bbox": [ + 56, + 338, + 294, + 392 + ], + "type": "text", + "content": "[31] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730-19742. PMLR, 2023. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 393, + 294, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 393, + 294, + 437 + ], + "spans": [ + { + "bbox": [ + 56, + 393, + 294, + 437 + ], + "type": "text", + "content": "[32] Wei Li, William Bishop, Alice Li, Chris Rawles, Folawiyo Campbell-Ajala, Divya Tyamagundlu, and Oriana Riva. On the effects of data scale on computer control agents. arXiv preprint arXiv:2406.03679, 2024. 6, 7, 1, 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 437, + 294, + 481 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 437, + 294, + 481 + ], + "spans": [ + { + "bbox": [ + 56, + 437, + 294, + 481 + ], + "type": "text", + "content": "[33] Yang Li, Gang Li, Luheng He, Jingjie Zheng, Hong Li, and Zhiwei Guan. Widget captioning: Generating natural language description for mobile user interface elements. arXiv preprint arXiv:2010.04295, 2020. 5, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 481, + 294, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 481, + 294, + 535 + ], + "spans": [ + { + "bbox": [ + 56, + 481, + 294, + 535 + ], + "type": "text", + "content": "[34] Zeyi Liao, Lingbo Mo, Chejian Xu, Mintong Kang, Jiawei Zhang, Chaowei Xiao, Yuan Tian, Bo Li, and Huan Sun. Eia: Environmental injection attack on generalist web agents for privacy leakage. arXiv preprint arXiv:2409.11295, 2024. 1, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 536, + 294, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 536, + 294, + 591 + ], + "spans": [ + { + "bbox": [ + 56, + 536, + 294, + 591 + ], + "type": "text", + "content": "[35] Ziyi Lin, Chris Liu, Renrui Zhang, Peng Gao, Longtian Qiu, Han Xiao, Han Qiu, Chen Lin, Wenqi Shao, Keqin Chen, et al. Sphinx: The joint mixing of weights, tasks, and visual embeddings for multi-modal large language models. arXiv preprint arXiv:2311.07575, 2023. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 592, + 294, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 592, + 294, + 635 + ], + "spans": [ + { + "bbox": [ + 56, + 592, + 294, + 635 + ], + "type": "text", + "content": "[36] Evan Zheran Liu, Kelvin Guu, Panupong Pasupat, Tianlin Shi, and Percy Liang. Reinforcement learning on web interfaces using workflow-guided exploration. arXiv preprint arXiv:1802.08802, 2018. 2, 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 636, + 294, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 636, + 294, + 658 + ], + "spans": [ + { + "bbox": [ + 56, + 636, + 294, + 658 + ], + "type": "text", + "content": "[37] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning, 2023. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 658, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 294, + 713 + ], + "type": "text", + "content": "[38] Quanfeng Lu, Wenqi Shao, Zitao Liu, Fanqing Meng, Boxuan Li, Botong Chen, Siyuan Huang, Kaipeng Zhang, Yu Qiao, and Ping Luo. Gui odyssey: A comprehensive dataset for cross-app gui navigation on mobile devices. arXiv preprint arXiv:2406.08451, 2024. 2, 6, 7, 1, 4" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 106 + ], + "type": "text", + "content": "[39] Xing Han Lu, Zdenek Kasner, and Siva Reddy. Weblinx: Real-world website navigation with multi-turn dialogue. arXiv preprint arXiv:2402.05930, 2024. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 107, + 555, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 555, + 151 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 555, + 151 + ], + "type": "text", + "content": "[40] Minesh Mathew, Dimosthenis Karatzas, and CV Jawahar. Docvqa: A dataset for vqa on document images. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 2200-2209, 2021. 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 152, + 553, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 553, + 206 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 553, + 206 + ], + "type": "text", + "content": "[41] Erik Nijkamp, Bo Pang, Hiroaki Hayashi, Lifu Tu, Huan Wang, Yingbo Zhou, Silvio Savarese, and Caiming Xiong. Codegen: An open large language model for code with multi-turn program synthesis. arXiv preprint arXiv:2203.13474, 2022. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 208, + 449, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 208, + 449, + 219 + ], + "spans": [ + { + "bbox": [ + 316, + 208, + 449, + 219 + ], + "type": "text", + "content": "[42] OpenAI. Hello gpt-4o, 2024. 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 220, + 553, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 220, + 553, + 263 + ], + "spans": [ + { + "bbox": [ + 316, + 220, + 553, + 263 + ], + "type": "text", + "content": "[43] Liangming Pan, Alon Albalak, Xinyi Wang, and William Yang Wang. Logic-lm: Empowering large language models with symbolic solvers for faithful logical reasoning. arXiv preprint arXiv:2305.12295, 2023. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 265, + 553, + 298 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 265, + 553, + 298 + ], + "spans": [ + { + "bbox": [ + 316, + 265, + 553, + 298 + ], + "type": "text", + "content": "[44] Alec Radford, Karthik Narasimhan, Tim Salimans, Ilya Sutskever, et al. Improving language understanding by generative pre-training. OpenAI, 2018." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 299, + 553, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 299, + 553, + 354 + ], + "spans": [ + { + "bbox": [ + 316, + 299, + 553, + 354 + ], + "type": "text", + "content": "[45] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of machine learning research, 21(140):1-67, 2020. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 355, + 553, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 355, + 553, + 399 + ], + "spans": [ + { + "bbox": [ + 316, + 355, + 553, + 399 + ], + "type": "text", + "content": "[46] Christopher Rawles, Alice Li, Daniel Rodriguez, Oriana Riva, and Timothy Lillicrap. Androidinthewild: A large-scale dataset for android device control. Advances in Neural Information Processing Systems, 36, 2024. 2, 4, 6, 1, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 400, + 553, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 400, + 553, + 453 + ], + "spans": [ + { + "bbox": [ + 316, + 400, + 553, + 453 + ], + "type": "text", + "content": "[47] Baptiste Roziere, Jonas Gehring, Fabian Gloeckle, Sten Sootla, Itai Gat, Xiaqing Ellen Tan, Yossi Adi, Jingyu Liu, Romain Sauvestre, Tal Remez, et al. Code llama: Open foundation models for code. arXiv preprint arXiv:2308.12950, 2023. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 456, + 553, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 456, + 553, + 487 + ], + "spans": [ + { + "bbox": [ + 316, + 456, + 553, + 487 + ], + "type": "text", + "content": "[48] Rico Sennrich. Improving neural machine translation models with monolingual data. arXiv preprint arXiv:1511.06709, 2015. 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 490, + 553, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 490, + 553, + 544 + ], + "spans": [ + { + "bbox": [ + 316, + 490, + 553, + 544 + ], + "type": "text", + "content": "[49] Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Mingchuan Zhang, YK Li, Yu Wu, and Daya Guo. Deepseekmath: Pushing the limits of mathematical reasoning in open language models. arXiv preprint arXiv:2402.03300, 2024.1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 545, + 553, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 545, + 553, + 611 + ], + "spans": [ + { + "bbox": [ + 316, + 545, + 553, + 611 + ], + "type": "text", + "content": "[50] Peter Shaw, Mandar Joshi, James Cohan, Jonathan Berant, Panupong Pasupat, Hexiang Hu, Urvashi Khandelwal, Kenton Lee, and Kristina N Toutanova. From pixels to ui actions: Learning to follow instructions via graphical user interfaces. Advances in Neural Information Processing Systems, 36:34354-34370, 2023. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 613, + 553, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 613, + 553, + 658 + ], + "spans": [ + { + "bbox": [ + 316, + 613, + 553, + 658 + ], + "type": "text", + "content": "[51] Tianlin Shi, Andrej Karpathy, Linxi Fan, Jonathan Hernandez, and Percy Liang. World of bits: An open-domain platform for web-based agents. In International Conference on Machine Learning, pages 3135-3144. PMLR, 2017. 2, 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 658, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 553, + 713 + ], + "type": "text", + "content": "[52] Lucas-Andrei Thil, Mirela Popa, and Gerasimos Spanakis. Navigating webai: Training agents to complete web tasks with large language models and reinforcement learning. In Proceedings of the 39th ACM/SIGAPP Symposium on Applied Computing, pages 866–874, 2024. 1, 3" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "29499" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 139 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 139 + ], + "type": "text", + "content": "[53] Jianqiang Wan, Sibo Song, Wenwen Yu, Yuliang Liu, Wenqing Cheng, Fei Huang, Xiang Bai, Cong Yao, and Zhibo Yang. Omniparser: A unified framework for text spotting key information extraction and table recognition. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15641-15653, 2024. 1, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 140, + 294, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 140, + 294, + 195 + ], + "spans": [ + { + "bbox": [ + 56, + 140, + 294, + 195 + ], + "type": "text", + "content": "[54] Bryan Wang, Gang Li, Xin Zhou, Zhourong Chen, Tovi Grossman, and Yang Li. Screen2words: Automatic mobile ui summarization with multimodal learning. In The 34th Annual ACM Symposium on User Interface Software and Technology, pages 498-510, 2021. 5, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 197, + 294, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 197, + 294, + 251 + ], + "spans": [ + { + "bbox": [ + 56, + 197, + 294, + 251 + ], + "type": "text", + "content": "[55] Junyang Wang, Haiyang Xu, Haitao Jia, Xi Zhang, Ming Yan, Weizhou Shen, Ji Zhang, Fei Huang, and Jitao Sang. Mobile-agent-v2: Mobile device operation assistant with effective navigation via multi-agent collaboration. arXiv preprint arXiv:2406.01014, 2024. 2, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 253, + 294, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 253, + 294, + 297 + ], + "spans": [ + { + "bbox": [ + 56, + 253, + 294, + 297 + ], + "type": "text", + "content": "[56] Junyang Wang, Haiyang Xu, Jiabo Ye, Ming Yan, Weizhou Shen, Ji Zhang, Fei Huang, and Jitao Sang. Mobile-agent: Autonomous multi-modal mobile device agent with visual perception. arXiv preprint arXiv:2401.16158, 2024. 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 298, + 294, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 298, + 294, + 352 + ], + "spans": [ + { + "bbox": [ + 56, + 298, + 294, + 352 + ], + "type": "text", + "content": "[57] Ke Wang, Houxing Ren, Aojun Zhou, Zimu Lu, Sichun Luo, Weikang Shi, Renrui Zhang, Linqi Song, Mingjie Zhan, and Hongsheng Li. Mathcoder: Seamless code integration in llms for enhanced mathematical reasoning. arXiv preprint arXiv:2310.03731, 2023. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 354, + 294, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 354, + 294, + 398 + ], + "spans": [ + { + "bbox": [ + 56, + 354, + 294, + 398 + ], + "type": "text", + "content": "[58] Weihan Wang, Qingsong Lv, Wenmeng Yu, Wenyi Hong, Ji Qi, Yan Wang, Junhui Ji, Zhuoyi Yang, Lei Zhao, Xixuan Song, et al. Cogvlm: Visual expert for pretrained language models. arXiv preprint arXiv:2311.03079, 2023. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 399, + 294, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 399, + 294, + 453 + ], + "spans": [ + { + "bbox": [ + 56, + 399, + 294, + 453 + ], + "type": "text", + "content": "[59] Jason Wei, Yi Tay, Rishi Bommasani, Colin Raffel, Barret Zoph, Sebastian Borgeaud, Dani Yogatama, Maarten Bosma, Denny Zhou, Donald Metzler, et al. Emergent abilities of large language models. arXiv preprint arXiv:2206.07682, 2022. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 456, + 294, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 456, + 294, + 510 + ], + "spans": [ + { + "bbox": [ + 56, + 456, + 294, + 510 + ], + "type": "text", + "content": "[60] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824-24837, 2022. 5, 1, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 512, + 294, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 512, + 294, + 554 + ], + "spans": [ + { + "bbox": [ + 56, + 512, + 294, + 554 + ], + "type": "text", + "content": "[61] Chen Henry Wu, Jing Yu Koh, Ruslan Salakhutdinov, Daniel Fried, and Aditi Raghunathan. Adversarial attacks on multimodal agents. arXiv preprint arXiv:2406.12814, 2024. 1, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 557, + 294, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 557, + 294, + 610 + ], + "spans": [ + { + "bbox": [ + 56, + 557, + 294, + 610 + ], + "type": "text", + "content": "[62] Shijie Wu, Ozan Irsoy, Steven Lu, Vadim Dabravolski, Mark Dredze, Sebastian Gehrmann, Prabhanjan Kambadur, David Rosenberg, and Gideon Mann. Bloomberg: A large language model for finance. arXiv preprint arXiv:2303.17564, 2023. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 613, + 294, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 613, + 294, + 667 + ], + "spans": [ + { + "bbox": [ + 56, + 613, + 294, + 667 + ], + "type": "text", + "content": "[63] Yichong Xu, Chenguang Zhu, Shuohang Wang, Siqi Sun, Hao Cheng, Xiaodong Liu, Jianfeng Gao, Pengcheng He, Michael Zeng, and Xuedong Huang. Human parity on commonsenseqa: Augmenting self-attention with external attention. arXiv preprint arXiv:2112.03254, 2021. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "type": "text", + "content": "[64] Zhao Yang, Jiaxuan Liu, Yucheng Han, Xin Chen, Zebiao Huang, Bin Fu, and Gang Yu. Appagent: Multimodal agents as smartphone users. arXiv preprint arXiv:2312.13771, 2023. 3" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 664 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 127 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 127 + ], + "type": "text", + "content": "[65] Shunyu Yao, Howard Chen, John Yang, and Karthik Narasimhan. Webshop: Towards scalable real-world web interaction with grounded language agents. Advances in Neural Information Processing Systems, 35:20744-20757, 2022. 2, 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 129, + 553, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 553, + 173 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 553, + 173 + ], + "type": "text", + "content": "[66] Yuan Yao, Tianyu Yu, Ao Zhang, Chongyi Wang, Junbo Cui, Hongji Zhu, Tianchi Cai, Haoyu Li, Weilin Zhao, Zhihui He, et al. Minicpm-v: A gpt-4v level mllm on your phone. arXiv preprint arXiv:2408.01800, 2024. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 175, + 553, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 553, + 228 + ], + "type": "text", + "content": "[67] Jiabo Ye, Anwen Hu, Haiyang Xu, Qinghao Ye, Ming Yan, Guohai Xu, Chenliang Li, Junfeng Tian, Qi Qian, Ji Zhang, et al. Ureader: UniversalOCR-free visually-situated language understanding with multimodal large language model. arXiv preprint arXiv:2310.05126, 2023. 2, 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 230, + 553, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 230, + 553, + 295 + ], + "spans": [ + { + "bbox": [ + 316, + 230, + 553, + 295 + ], + "type": "text", + "content": "[68] Da Yin, Faeze Brahman, Abhilasha Ravichander, Khyathi Chandu, Kai-Wei Chang, Yejin Choi, and Bill Yuchen Lin. Agent lumos: Unified and modular training for open-source language agents. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 12380–12403, 2024. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 297, + 553, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 297, + 553, + 350 + ], + "spans": [ + { + "bbox": [ + 316, + 297, + 553, + 350 + ], + "type": "text", + "content": "[69] Huaiyuan Ying, Shuo Zhang, Linyang Li, Zhejian Zhou, Yunfan Shao, Zhaoye Fei, Yichuan Ma, Jiawei Hong, Kuikun Liu, Ziyi Wang, et al. Internlm-math: Open math large language models toward verifiable reasoning. arXiv preprint arXiv:2402.06332, 2024. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 353, + 553, + 406 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 353, + 553, + 406 + ], + "spans": [ + { + "bbox": [ + 316, + 353, + 553, + 406 + ], + "type": "text", + "content": "[70] Longhui Yu, Weisen Jiang, Han Shi, Jincheng Yu, Zhengying Liu, Yu Zhang, James T Kwok, Zhenguo Li, Adrian Weller, and Weiyang Liu. Metamath: Bootstrap your own mathematical questions for large language models. arXiv preprint arXiv:2309.12284, 2023. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 408, + 553, + 452 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 408, + 553, + 452 + ], + "spans": [ + { + "bbox": [ + 316, + 408, + 553, + 452 + ], + "type": "text", + "content": "[71] Qiusi Zhan, Zhixiang Liang, Zifan Ying, and Daniel Kang. Injecagent: Benchmarking indirect prompt injections in tool-integrated large language model agents. arXiv preprint arXiv:2403.02691, 2024. 1, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 453, + 553, + 496 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 453, + 553, + 496 + ], + "spans": [ + { + "bbox": [ + 316, + 453, + 553, + 496 + ], + "type": "text", + "content": "[72] Jiwen Zhang, Jihao Wu, Yihua Teng, Minghui Liao, Nuo Xu, Xiao Xiao, Zhongyu Wei, and Duyu Tang. Android in the zoo: Chain-of-action-thought for gui agents. arXiv preprint arXiv:2403.02713, 2024. 5, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 498, + 553, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 498, + 553, + 586 + ], + "spans": [ + { + "bbox": [ + 316, + 498, + 553, + 586 + ], + "type": "text", + "content": "[73] Pan Zhang, Xiaoyi Dong, Yuhang Zang, Yuhang Cao, Rui Qian, Lin Chen, Qipeng Guo, Haodong Duan, Bin Wang, Linke Ouyang, Songyang Zhang, Wenwei Zhang, Yining Li, Yang Gao, Peng Sun, Xinyue Zhang, Wei Li, Jingwen Li, Wenhai Wang, Hang Yan, Conghui He, Xingcheng Zhang, Kai Chen, Jifeng Dai, Yu Qiao, Dahua Lin, and Jiaqi Wang. Internlm-xcomposer-2.5: A versatile large vision language model supporting long-contextual input and output, 2024. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 587, + 553, + 619 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 587, + 553, + 619 + ], + "spans": [ + { + "bbox": [ + 316, + 587, + 553, + 619 + ], + "type": "text", + "content": "[74] Boyuan Zheng, Boyu Gou, Jihyung Kil, Huan Sun, and Yu Su. Gpt-4v (isdiction) is a generalist web agent, if grounded. arXiv preprint arXiv:2401.01614, 2024. 1, 2, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 620, + 553, + 664 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 620, + 553, + 664 + ], + "spans": [ + { + "bbox": [ + 316, + 620, + 553, + 664 + ], + "type": "text", + "content": "[75] Longtao Zheng, Rundong Wang, Xinrun Wang, and Bo An. Synapse: Trajectory-as-exemplar prompting with memory for computer control. In The Twelfth International Conference on Learning Representations, 2023. 1" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "29500" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Spk2SRImgNet_ Super-Resolve Dynamic Scene from Spike Stream via Motion Aligned Collaborative Filtering/e0bcc965-cf40-409b-a203-db12626ad68b_content_list.json b/2025/Spk2SRImgNet_ Super-Resolve Dynamic Scene from Spike Stream via Motion Aligned Collaborative Filtering/e0bcc965-cf40-409b-a203-db12626ad68b_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..4f871073a190cae940b5f3e02cf147c5feee039e --- /dev/null +++ b/2025/Spk2SRImgNet_ Super-Resolve Dynamic Scene from Spike Stream via Motion Aligned Collaborative Filtering/e0bcc965-cf40-409b-a203-db12626ad68b_content_list.json @@ -0,0 +1,1869 @@ +[ + { + "type": "text", + "text": "Spk2SRImgNet: Super-Resolve Dynamic Scene from Spike Stream via Motion Aligned Collaborative Filtering", + "text_level": 1, + "bbox": [ + 161, + 128, + 836, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yuanlin Wang $^{1,2}$ , Yiyang Zhang $^{1,2}$ , Ruiqin Xiong $^{1,2,*}$ , Jing Zhao $^{1,2}$ , Jian Zhang $^{3}$ , Xiaopeng Fan $^{4}$ , Tiejun Huang $^{1,2}$", + "bbox": [ + 236, + 202, + 761, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ School of Computer Science, Peking University", + "bbox": [ + 305, + 239, + 691, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ National Key Laboratory for Multimedia Information Processing, Peking University", + "bbox": [ + 163, + 256, + 833, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{3}$ School of Electronic and Computer Engineering, Shenzhen Graduate School, Peking University", + "bbox": [ + 119, + 273, + 877, + 292 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{4}$ School of Computer Science and Technology, Harbin Institute of Technology", + "bbox": [ + 191, + 292, + 805, + 310 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "wangyuanlin@stu.pku.edu.cn", + "bbox": [ + 380, + 313, + 614, + 325 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{yiyangz, rqxiong, jzhaopku, zhangjian.sz, tjhuang}@pku.edu.cn fxp@hit.edu.cn", + "bbox": [ + 148, + 329, + 841, + 344 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 378, + 326, + 395 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Spike camera is a kind of neuromorphic camera that records dynamic scenes by firing a stream of binary spikes with extremely high temporal resolution. It demonstrates great potential for vision tasks in high-speed scenarios. One limitation in its current implementation is the relatively low spatial resolution. This paper develops a network called Spk2SRImgNet to super-resolve high resolution images from low resolution spike stream. However, fluctuations in spike stream hinder the performance of spike camera super resolution. To address this issue, we propose a motion aligned collaborative filtering (MACF) module, which is motivated by key ideas in classic image restoration schemes to mitigate fluctuations in spike data. MACF leverages the temporal similarity of spike stream to acquire similar features from neighboring moments via motion alignment. To separate disturbances from features, MACF filters these similar features jointly in transform domain to exploit representation sparsity, and generates refinement features that will be used to update initial fluctuated features. Specifically, MACF designs an inverse motion alignment operation to map these refinement features back to their original positions. The initial features are aggregated with the repositioned refinement features to enhance reliability. Experimental results demonstrate that the proposed method achieves state-of-the-art performance compared with existing methods.", + "bbox": [ + 89, + 411, + 485, + 806 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 818, + 220, + 832 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "High-speed imaging attracts increasing attention in emerging computer vision applications such as autonomous driv-", + "bbox": [ + 89, + 843, + 483, + 875 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/482e27539178c625a69df88b96f02f63c321c9eddf755e8369557e0ec804f353.jpg", + "image_caption": [ + "Figure 1. Illustration of spike camera super resolution (SCSR) and visual comparison $(\\times 4)$ with existing SCSR methods: VidarSR [42] and SpikeSR-Net [50]. Please zoom in for better comparison." + ], + "image_footnote": [], + "bbox": [ + 517, + 378, + 903, + 452 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ing [15, 17] and robotics [16, 29]. Conventional camera employs an exposure-based imaging model, which limits its ability to capture high-speed scenes. Specifically, conventional camera accumulates photoelectric signals over an exposure time window to form a snapshot image, which requires the scene to be still during the exposure period. Otherwise, a single point on a moving object may be projected onto different sensor pixels, resulting in motion blur.", + "bbox": [ + 511, + 518, + 906, + 638 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Neuromorphic camera is an emerging camera that abandons the concept of exposure window and mimics the working mechanism of biological retina. It records intensity information with high temporal resolution, showing great potential for high speed vision tasks. One kind of neuromorphic camera is event camera [2, 25, 26, 31]. It detects light intensity changes of each pixel in the logarithmic domain and fires an event when the change exceeds a specified threshold. Another kind of neuromorphic camera is spike camera [8, 9, 19]. Each pixel of spike camera independently and continuously accumulates photons, and fires a spike when the accumulation exceeds a predefined threshold. Unlike event camera that only records relative changes of light intensity, spike camera captures absolute intensity by firing a stream of binary spikes, providing clearer information for reconstructing image details.", + "bbox": [ + 511, + 641, + 908, + 883 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Benefiting from its extremely high temporal resolution", + "bbox": [ + 532, + 885, + 906, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author.", + "bbox": [ + 109, + 887, + 235, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "11416", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "(e.g. $40\\mathrm{kHz}$ ), spike camera could sample a large amount of intensity information in a short period. Besides, these samples cover different locations of the scene due to the relative motion between camera and scene. Therefore, there is promising potential to reconstruct high resolution (HR) images from spike stream by converting the dense sampling in temporal domain to spatial domain.", + "bbox": [ + 89, + 90, + 480, + 196 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This paper aims to super-resolve high resolution images from low resolution spike stream. However, due to the Poisson effect of photon arrival, the quantization effect of spike readout, and the thermal noises introduced by circuits, spikes exhibit fluctuations, i.e., the length of spike accumulation period varies over time, even when light intensity remains unchanged. Thus, the features extracted from fluctuated spikes tend to be noisy and temporally incoherent. These unreliable features provide low-quality information source for super resolution (SR) reconstruction. Moreover, they affect the alignment accuracy when leveraging temporal information, further degrading the quality of SR reconstruction.", + "bbox": [ + 89, + 198, + 480, + 391 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we propose a spike-to-super-resolution-image network Spk2SRImgNet. To handle spike fluctuations, we take inspiration from several classic image restoration methods, such as BM3D [6], BAS [43], LSSC [27], and WNNM [18]. These methods incorporate the idea of leveraging non-local similarity within natural images to reduce the unknown random disturbance. We embed these concepts into the design of Spk2SRImgNet and develop a motion aligned collaborative filtering (MACF) module to reduce disturbances in spike information. Firstly, MACF obtains similar features from neighboring moments via motion alignment, exploiting the temporal similarity of spike stream. To separate disturbances from features, MACF performs collaborative filtering on grouped similar features in transform domain for leveraging representation sparsity, and generates refinement features for each moment. These refinement features will be utilized to update initial fluctuated features. Specifically, MACF introduces an inverse motion alignment operation to map these multi-moment refinement features to their original unaligned positions. The initial features are integrated with the repositioned refinement features to improve reliability. Experimental results show that the proposed Spk2SRImgNet achieves state-of-the-art performance on synthesized and real-world spike data.", + "bbox": [ + 91, + 393, + 482, + 768 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 784, + 232, + 800 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Video Super Resolution. Video super resolution (VSR) produces high resolution images from low resolution (LR) image sequences. Temporal alignment is a crucial step to model temporal correspondences of misaligned frames. Early learning-based methods [3, 21] employ traditional optical flow estimation [40] for motion compensation. With", + "bbox": [ + 89, + 809, + 480, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "the diverse development of network structures, TDAN [35] and EDVR [37] utilize deformable convolution [7, 56] for feature alignment. BasicVSR++ [4] leverages optical flow to guide deformable alignment for stable training. VRT [24] proposes a temporal mutual attention block to facilitate motion estimation and feature alignment.", + "bbox": [ + 511, + 90, + 903, + 181 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Event Camera Reconstruction and Super Resolution. Event camera monitors relative changes in light intensity with high temporal resolution. Reconstructing intensity images from events has been studied in the past decade. Kim et al. [22] propose to use the extended Kalman filter and Poisson reconstruction for event reconstruction. Bardow et al. [1] estimate both intensity and optical flow using the primal-dual algorithm. With the development of deep learning, E2VID [32], FireNet [20], ET-Net [41], EVSNN [55] have been proposed successively with good performance. To reconstruct HR images from LR events, Wang et al. [36] propose an unsupervised GAN EventSR. Mostafavi et al. [28] propose an end-to-end recurrent network E2SRI.", + "bbox": [ + 511, + 181, + 903, + 377 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Spike Camera Reconstruction and Super Resolution. Image reconstruction is a popular topic for spike camera [38, 48, 51-54]. TFI [53] reconstructs images based on the interval between two temporally neighboring spikes. TFP [53] utilizes the number of spikes within a temporal window to estimate the intensity. For better performance, CNN-based [48, 52] and spike neural network (SNN)-based [51, 54] frameworks have been proposed successively. To reconstruct HR images from LR spike stream, Zhao et al. [47] derive the constraint relationship between pixel intensities and each spike period, and obtain HR images based on gradient descent algorithm. Xiang et al. [42] design a multi-level feature learning network to achieve spike camera super resolution. Zhao et al. [50] propose SpikeSR-Net based on the observation model of spike camera. Furthermore, Dong et al. [11] propose a super resolution network for the newly developed color spike camera [10, 12-14].", + "bbox": [ + 511, + 378, + 903, + 633 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Spike Camera", + "text_level": 1, + "bbox": [ + 513, + 648, + 656, + 665 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Working Mechanism. Spike camera contains a pixel array of size $H \\times W$ and each pixel works independently. Each pixel comprises three main components: photon receptor, integrator, and comparator, as shown in Fig. 2. The photoreceptor captures incident photons from external scenes and converts them into electric charges through photoelectric conversion. The integrator accumulates the electric charges continuously. Meanwhile, the comparator checks the accumulated voltage persistently. Once the voltage reaches a predefined threshold $\\theta$ , the pixel sets up a spike flag and resets the integrator, immediately restarting a new \"accumulation-and-set\" cycle. Spike readout circuit periodically checks the spike flag at a frequency of $40\\mathrm{kHz}$ . When detecting the flag, the circuit outputs a spike \"1\", otherwise, it outputs \"0\". The voltage accumulated by the in", + "bbox": [ + 511, + 674, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "11417", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/b1caeff65e5a9539fd7868965e875e8773b7f426097fe61f95efb0a5ce78069a.jpg", + "image_caption": [ + "Figure 2. Working mechanism of spike camera." + ], + "image_footnote": [], + "bbox": [ + 93, + 79, + 482, + 233 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "tegrator at an arbitrary time $t$ can be formulated as [49]:", + "bbox": [ + 89, + 281, + 460, + 296 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nA (x, y, t) = \\int_ {0} ^ {t} \\eta I (x, y, u) d u \\quad \\mathrm {m o d} \\theta , \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 303, + 482, + 338 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $I(x,y,u)$ denotes the arrival rate of photons received by pixel $(x,y)$ at time $u$ , $\\eta$ denotes the photoelectric conversion rate. The output of spike camera is a 3D binary spike matrix $\\{S(u)\\}_{u=1}^{T}$ with size $H\\times W\\times T$ , where $T$ is the temporal length, $S(u)$ is a spike frame at time $u$ .", + "bbox": [ + 89, + 345, + 483, + 420 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Spike Interval Based Reconstruction. According to the working mechanism of spike camera, the stronger the light intensity, the shorter the interval between two neighboring spikes, i.e., the spike interval is inversely proportional to the light intensity. This property is exploited by spike-interval-based reconstruction method (SIR) [53]. Specifically, the spike interval of two temporally neighboring spikes that cover the position $(x,y,u)$ can be formulated as:", + "bbox": [ + 89, + 421, + 483, + 542 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\Phi (x, y, u) = \\min \\left\\{t \\mid S (x, y, t) = 1, t \\geq u \\right\\} \\tag {2} \\\\ - \\max \\{t | S (x, y, t) = 1, t < u \\}. \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 138, + 551, + 482, + 585 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "SIR estimates the light intensity of pixel $(x,y)$ at time $u$ by", + "bbox": [ + 89, + 595, + 482, + 611 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nY (x, y, u) = \\frac {\\delta}{\\Phi (x , y , u)}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 200, + 619, + 480, + 651 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\delta$ denotes the maximum dynamic range of reconstruction.", + "bbox": [ + 89, + 659, + 482, + 686 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Spike Fluctuations. The imaging of spike camera is affected by multiple kinds of noise. First, the Poisson process of photon arrival causes the time taken for the number of photons to reach the threshold to exhibit randomness, which is denoted as Poisson noise. Second, the spike readout circuit is controlled by a periodic clock signal, i.e., it only outputs spikes at discrete times, which results in a time difference between spike flag set and spike readout (see Fig. 2). This difference leads to quantization noise. Third, the thermal noise introduced by circuits brings randomness to spike fire. For the above reasons, the intervals of neighboring spikes exhibit fluctuations and struggle to accurately reflect light intensity, leading to noisy reconstruction results from SIR.", + "bbox": [ + 89, + 688, + 483, + 898 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Motivation and Idea", + "text_level": 1, + "bbox": [ + 513, + 89, + 709, + 104 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The key problem we aim to address is how to super-resolve the original dynamic scene from captured binary spike data to obtain high quality visual reconstruction with clean image details. According to the working mechanism of spike camera, each individual spike does not reveal the light intensity directly. Instead, it only tells the time point at which a spike firing occurs. Essentially, the instantaneous light intensity can be inferred only by considering the length of intervals between consecutive spikes. Unfortunately, for the reasons discussed above, the recorded spike intervals are not stable, and they fluctuate around an expectation value determined by actual light intensity.", + "bbox": [ + 511, + 114, + 903, + 296 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To handle such fluctuations, we draw inspiration from several classic image restoration methods, such as BM3D [6], BAS [43], LSSC [27], and WNNM [18]. Although they are quite diversified in algorithmic design, all these methods incorporate the idea of leveraging non-local similarity within natural image signals to reduce the unknown random disturbance. They utilize a group of similar patches and establish a prior model for them so that a regularization term is formed to separate the true signal from the noise.", + "bbox": [ + 511, + 296, + 903, + 431 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Suppose $\\mathbf{y} = \\mathbf{x} + \\mathbf{n}$ is an image corrupted by noise, where $\\mathbf{x}$ is the clean image and $\\mathbf{n}$ is the noise. The idea of these methods can be summarized as follows:", + "bbox": [ + 511, + 431, + 903, + 476 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "(1) For each pixel location $\\mathbf{p}$ , search for a group of similar patches from the noisy image $\\mathbf{y}$ :", + "bbox": [ + 511, + 477, + 903, + 508 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {y} _ {1}, \\mathbf {y} _ {2}, \\dots , \\mathbf {y} _ {N} = \\Lambda (\\mathbf {y} | \\mathbf {p}), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 635, + 518, + 903, + 536 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $N$ is the number of similar patches, $\\Lambda (\\cdot |\\cdot)$ is the operation of searching for and extracting similar patches. Note that the notation $\\mathbf{y}_i$ should depend on $\\mathbf{p}$ but we omit it for conciseness.", + "bbox": [ + 537, + 545, + 903, + 604 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "(2) Exploit the correlation within the patch group $\\{\\mathbf{y}_i\\}_{i = 1}^N$ to generate an estimate for each patch:", + "bbox": [ + 513, + 604, + 903, + 636 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\mathbf {x}} _ {1}, \\tilde {\\mathbf {x}} _ {2}, \\dots , \\tilde {\\mathbf {x}} _ {N} = \\Pi \\left(\\left\\{\\mathbf {y} _ {i} \\right\\} _ {i = 1} ^ {N}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 620, + 645, + 903, + 664 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\tilde{\\mathbf{x}}_i$ is the estimate of the true signal $\\mathbf{x}_i$ corresponding to $\\mathbf{y}_i$ . The actual operation of $\\Pi(\\cdot)$ depends on the prior model adopted to characterize the consistency between patches $\\{\\mathbf{y}_i\\}_{i=1}^N$ . Generally, $\\Pi(\\cdot)$ can be regarded as an operator to solve the following optimization objective:", + "bbox": [ + 537, + 672, + 903, + 763 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{\\tilde {\\mathbf {x}} _ {i} \\right\\} = \\arg \\min \\sum_ {i = 1} ^ {N} \\left| \\left| \\mathbf {x} _ {i} - \\mathbf {y} _ {i} \\right| \\right| _ {2} ^ {2} + \\Phi \\left(\\left\\{\\mathbf {x} _ {i} \\right\\} _ {i = 1} ^ {N}\\right) \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 553, + 773, + 903, + 816 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "BM3D [6] solves Eq. (6) by applying a 3D transform to the stacked similar patches and constructing a sparsity regularization term in the transform domain. It utilizes hard-thresholding, soft-thresholding, and Wiener filtering in different steps, corresponding to $l_{0}$ , $l_{1}$ , and $l_{2}$", + "bbox": [ + 537, + 824, + 903, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "11418", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/82f650853a543bc0384bc7025ac8ff79d805a954f33dbf8e841bc620bf268102.jpg", + "image_caption": [ + "Figure 3. The main idea of motion aligned collaborative filtering. First, we collect temporal similar features via motion alignment. Then, we stack these aligned features and perform transform-domain collaborative filtering to generate refinement features for each moment. Subsequently, the refinement features are unstacked and mapped back to their original positions via inverse warping. These repositioned refinement features are aggregated with input features to obtain refined features with improved consistency." + ], + "image_footnote": [], + "bbox": [ + 93, + 88, + 903, + 188 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "sparsity norm, respectively. BAS [43] applies a content adaptive 2D decorrelation transform to the similar patches via principle component analysis (PCA), and builds a separate distribution model and regularization term for each transform band, using the transform coefficients of these patches as samples. It leads to a PCA-domain adaptive soft-thresholding to solve Eq. (6). LLSC [27] adopts another prior model that emphasizes that similar patches tend to share the location pattern of non-zero dictionary elements in their sparse decomposition. WNNM [18], on the other hand, utilizes the low-rank property of similar patches so that the problem (6) reduces to singular value decomposition (SVD) and soft-thresholding on the singular values.", + "bbox": [ + 114, + 265, + 483, + 489 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "(3) Place each estimate $\\tilde{\\mathbf{x}}_i$ back to the position where $\\mathbf{y}_i$ is taken from, and aggregate the overlapping estimates from all position $\\mathbf{p}$ to form a whole estimate for $\\mathbf{x}$ :", + "bbox": [ + 89, + 491, + 482, + 536 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\mathbf {x}} = \\Lambda^ {- 1} \\left(\\left\\{\\tilde {\\mathbf {x}} _ {1}, \\tilde {\\mathbf {x}} _ {2}, \\dots , \\tilde {\\mathbf {x}} _ {N} \\right\\} _ {\\mathbf {p}}\\right). \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 196, + 542, + 482, + 560 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here $\\Lambda^{-1}(\\cdot)$ denotes patch reposition and aggregation operation. The reposition of $\\tilde{\\mathbf{x}}_i$ to $\\tilde{\\mathbf{x}}$ is the inverse operation of taking $\\mathbf{y}_i$ from $\\mathbf{y}$ in stage (1).", + "bbox": [ + 114, + 568, + 482, + 613 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Taking these ideas into account, we propose a novel framework tailored for spike data to reduce disturbances in spike information. The main idea is illustrated in Fig. 3. The similarity exploited in this framework is temporal similarity, and it may be extended to include spatial non-local similarity in the future. First, we align the spike features at neighboring moments with the spike feature at the reconstruction moment to obtain a set of similar features. Then, we stack the aligned features and perform a neural-network-based transform-domain collaborative filtering, to separate the disturbances from clean features by exploiting representation sparsity and generate refinement features for each moment. These refinement features are used to update the initial input features. Specifically, the refinement features are unstacked and mapped back to their original positions via inverse warping, and then aggregated with input features. The purpose of this framework is to improve the consistency of multi-moment features along the motion trajectory.", + "bbox": [ + 89, + 614, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/2060476049fe074415973e3ab9de0778e54ed28cf5da0f89029d87f22b26f2f4.jpg", + "image_caption": [ + "Figure 4. Overall architecture of the proposed spike camera super resolution network Spk2SRImgNet." + ], + "image_footnote": [], + "bbox": [ + 513, + 262, + 906, + 372 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "5. Method", + "text_level": 1, + "bbox": [ + 513, + 426, + 604, + 441 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "5.1. Overall Architecture", + "text_level": 1, + "bbox": [ + 511, + 455, + 709, + 470 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To reconstruct HR dynamic scenes from LR binary spike stream, we develop an end-to-end trainable spike camera super resolution network, Spk2SRImgNet. The overall architecture is shown in Fig. 4. To reconstruct an HR image $\\mathcal{I}(t_0) \\in \\mathbb{R}^{rH \\times rW}$ at time $t_0$ , we take a set of consecutive spike frames $S(t_0 + i), i = 0, \\pm 1, \\pm 2, \\ldots$ around $S(t_0)$ as input, where $S(t_0 + i) \\in \\mathbb{B}^{H \\times W \\times 1}$ , $r$ is SR scale, $\\mathbb{B}$ represents the binary domain.", + "bbox": [ + 511, + 483, + 906, + 604 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Firstly, the input spike stream is partitioned into $K$ overlapping short-term spike blocks $\\{B_i\\}_{i = -n}^n (K = 2n + 1)$ , where $B_{i}\\in \\mathbb{B}^{H\\times W\\times L}$ , $L$ is the temporal length of each block, $B_{0}$ is the key block centered at time $t_0$ , $\\{B_i\\}, i = \\pm 1,\\pm 2,\\ldots$ are neighboring blocks distributed symmetrically on both sides of $B_{0}$ , $B_{i}$ is centered at time $t_i$ . To bridge the domain gap between binary spikes and real pixel values, we first convert spike frames in $B_{i} = \\{S(u)\\}_{u = t_{i} - w_{h}}^{t_{i} + w_{h}}$ to intensity frames in $Y_{i} = \\{Y(u)\\}_{u = t_{i} - w_{h}}^{t_{i} + w_{h}}\\in \\mathbb{R}^{H\\times W\\times L}$ based on the spike interval based reconstruction (SIR) method [53], where $L = 2w_{h} + 1$ . Then, we extract features $\\{F_i\\}_{i = -n}^n$ from $\\{Y_i\\}_{i = -n}^n$ through an encoder, which consists of two convolutional layers and three residual blocks. To reduce fluctuations, $\\{F_i\\}_{i = -n}^n$ are input to motion aligned collaborative filtering (MACF) module, which generates improved features $\\{F_i\\}_{i = -n}^n$ . Finally, the multi-moment features $\\{\\hat{F}_i\\}_{i = -n}^n$ are fed into motion aligned super resolved reconstruction module to generate an HR image $\\mathcal{I}(t_0)$ .", + "bbox": [ + 511, + 609, + 908, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "11419", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/3318ce1e321fdff7b9477e3808f00cc75f152ab598022ff514495d56f02ecaa8.jpg", + "image_caption": [ + "Figure 5. The overall framework of motion aligned collaborative filtering (MACF) module." + ], + "image_footnote": [], + "bbox": [ + 91, + 87, + 903, + 301 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/b7624ecf54f11fd719d43f561df9dee72268361eb7507230fd596663c32dfe1f.jpg", + "image_caption": [ + "Figure 6. Illustration of channel rearrangement (CR) and inverse channel rearrangement (ICR). The left part is $K$ features $\\{F_i\\}_{i = -n}^n$ at $K$ different moments, where $F_{i}\\in \\mathbb{R}^{H\\times W\\times C}$ . The right part is $C$ features $\\{G_j\\}_{j = 0}^{C - 1}$ , where $G_{j}\\in \\mathbb{R}^{H\\times W\\times K}$ , each $G_{j}$ contains features at $K$ different moments." + ], + "image_footnote": [], + "bbox": [ + 91, + 332, + 482, + 429 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.2. Motion Aligned Collaborative Filtering", + "text_level": 1, + "bbox": [ + 89, + 516, + 428, + 532 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We propose a motion aligned collaborative filtering (MACF) module to implement the idea proposed in Sec. 4. The overall framework of MACF is shown in Fig. 5, which includes six steps. We provide a detailed description of each step below.", + "bbox": [ + 89, + 537, + 482, + 612 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Step 1: Offset Estimation. This step generates the forward $(F_{i}\\rightarrow F_{0})$ and backward $(F_0\\to F_i)$ motion offsets between neighboring features $\\{F_i\\} ,i = \\pm 1,\\dots ,\\pm n$ and key feature $F_{0}$ . These offsets will be utilized in the subsequent feature warping (step 3) and inverse feature warping (step 5), respectively. Specifically, the forward and backward motion offsets: $\\mathcal{O}_i^{\\mathbf{F}}$ and $\\mathcal{O}_i^{\\mathbf{B}}$ are both derived from $F_{i}$ and $F_{0}$ , which are formulated as follows ( $\\forall i\\in \\{\\pm 1,\\dots ,\\pm n\\}$ ):", + "bbox": [ + 89, + 613, + 482, + 734 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {O} _ {i} ^ {\\mathbf {F}} = m _ {1} \\left(\\left\\{F _ {i}, F _ {0} \\right\\}\\right), \\\\ \\mathcal {O} _ {i} ^ {\\mathbf {B}} = m _ {2} \\left(\\left\\{F _ {i}, F _ {0} \\right\\}\\right), \\tag {8} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 212, + 743, + 480, + 782 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $m_{1}(\\cdot)$ and $m_{2}(\\cdot)$ denote two independent convolutional blocks that consist of several convolutional layers, $\\{,\\}$ denotes channel concatenation operation.", + "bbox": [ + 89, + 792, + 482, + 837 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Step 2: Channel Rearrange. The $\\{F_i\\}_{i = -n}^n$ at $K$ different moments are features extracted from the fluctuated reconstruction $\\{Y_{i}\\}_{i = -n}^{n}$ , where $F_{i}\\in \\mathbb{R}^{H\\times W\\times C}$ , $F_{i} = \\{F_{i}^{0},F_{i}^{1},\\dots,F_{i}^{C - 1}\\}$ , the superscript of $F_{i}$ denotes channel", + "bbox": [ + 89, + 837, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "index. Each feature $F_{i}$ consists of $C$ channels, with each channel encoding different information about the scene. For example, some channels capture edges, some channels focus on textures, and so on. We perform channel rearrangement (CR) to collect data at the same channel index for each feature and concatenate them, as shown in Fig. 6. We form $\\{G_j\\}_{j = 0}^{C - 1}$ from $\\{F_i\\}_{i = -n}^n$ via CR operation, where $G_{j}\\in \\mathbb{R}^{H\\times W\\times K}$ , $G_{j} = \\{G_{j}^{-n},\\dots,G_{j}^{0},\\dots,G_{j}^{n}\\} = \\{F_{-n}^{j},\\dots,F_{0}^{j},\\dots,F_{n}^{j}\\}$ , the subscript of $G$ denotes channel index, the superscript of $G$ denotes time index. $G_{j}$ contains $K$ features with channel index $j$ at $K$ different moments. To offer an intuitive understanding of this operation, we provide a visualization of CR in supplementary material. After CR, steps 3, 4 and 5 are performed sequentially for each $G_{j}$ . Step 3: Feature Warping. Step 2 collects temporally unaligned features that encode the same information. We perform feature warping based on deformable convolution $\\mathcal{D}$ to align neighboring features $\\{G_j^i\\}$ , $i = \\pm 1,\\ldots ,\\pm n$ with key feature $G_{j}^{0}$ . Specifically, given the forward offsets from neighboring features to key feature $\\{\\mathcal{O}_i^{\\mathbf{F}}\\}$ , $i = \\pm 1,\\ldots ,\\pm n$ , the aligned features could be obtained by:", + "bbox": [ + 511, + 334, + 906, + 661 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {G} _ {j} ^ {i} = \\mathcal {D} \\left(G _ {j} ^ {i}, \\mathcal {O} _ {i} ^ {\\mathbf {F}}\\right) = \\mathcal {D} \\left(F _ {i} ^ {j}, \\mathcal {O} _ {i} ^ {\\mathbf {F}}\\right), \\forall i \\in \\{\\pm 1, \\dots , \\pm n \\}. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 521, + 667, + 906, + 688 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "These aligned features $\\{\\tilde{G}_j^i\\}_{i = -n}^n$ are then concatenated to derive $\\tilde{G}_j = \\{\\tilde{G}_j^{-n},\\dots,\\tilde{G}_j^n\\}$ , where $\\tilde{G}_j^0 = G_j^0 = F_0^j$ . $\\tilde{G}_j$ could be regarded as stacked similar features.", + "bbox": [ + 511, + 696, + 905, + 744 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Step 4: Transform Domain Stacked Feature Filtering. We design a Transform Domain stacked feature Filtering (TDF) step to generate refinement features $\\Delta \\tilde{G}_j$ from stacked similar features $\\tilde{G}_j$ by leveraging representation sparsity in transform domain. TDF is designed with four submodules: transform, noise estimation, filtering, and inverse transform, as depicted in step 4 of Fig. 5. Specifically, the input $\\tilde{G}_j$ is first transformed into feature $\\dot{G}_j$ . Simultaneously, noise feature $N_j$ are extracted from $\\tilde{G}_j$ via noise estimation submodule. $\\dot{G}_j$ and $N_j$ are passed through filter", + "bbox": [ + 511, + 744, + 906, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "11420", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/8bfd45319a78c07051618e16cb04cf02ba648aae7b83bd7b9c3f4d2045f14d25.jpg", + "image_caption": [ + "Figure 7. Illustration of submodules in step 4 of MACF. (a) Transform and inverse transform submodules; (b) noise estimation submodule; (c) filtering submodule." + ], + "image_footnote": [], + "bbox": [ + 93, + 88, + 482, + 218 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ing submodule to generate $\\ddot{G}_j$ . $\\ddot{G}_j$ are subsequently fed into inverse transform submodule. Additionally, a residual connection is applied in TDF step, producing refinement features $\\Delta \\tilde{G}_j$ at last.", + "bbox": [ + 89, + 277, + 482, + 340 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The above four submodules are illustrated in detail in Fig. 7. Transform and inverse transform submodules are independently realized via several convolutional layers (Fig. 7 (a)), involving both spatial and inter-channel transform. For noise estimation (NE) (Fig. 7 (b)), we first average $\\tilde{G}_j$ along the channel dimension to obtain a rough temporal average result $\\mathcal{E}$ . Then we calculate the residual between $\\tilde{G}_j$ and $\\mathcal{E}$ to represent fluctuations in features. Lastly, we extract deeper features from the residual to derive noise feature $N_j$ . We concatenate the transformed feature $\\dot{G}_j$ and noise feature $N_j$ as input to the filtering submodule (Fig. 7 (c)), which contains a U-Net structure [33] and outputs $\\ddot{G}_j$ .", + "bbox": [ + 89, + 340, + 483, + 523 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Step 5: Inverse Feature Warping and Aggregation. We propose an inverse motion alignment operation to map these multi-moment refinement features $\\Delta \\tilde{G}_j$ to their original positions. Specifically, given the backward offsets from key feature to neighboring features $\\{\\mathcal{O}_i^{\\mathbf{B}}\\}$ , $i = \\pm 1,\\dots,\\pm n$ , we achieve inverse feature warping by:", + "bbox": [ + 89, + 523, + 483, + 614 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\Delta G _ {j} ^ {i} = \\mathcal {D} \\left(\\Delta \\tilde {G} _ {j} ^ {i}, \\mathcal {O} _ {i} ^ {\\mathbf {B}}\\right), \\forall i \\in \\{\\pm 1, \\dots , \\pm n \\}. \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 130, + 625, + 482, + 646 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "$\\Delta G_{j}^{i}$ is repositioned refinement feature and corresponds to initial unaligned feature $G_{j}^{i}$ . We concatenate these repositioned refinement features and obtain $\\Delta G_{j} = \\{\\Delta G_{j}^{-n}, \\dots, \\Delta G_{j}^{n}\\}$ , where $\\Delta G_{j}^{0} = \\Delta \\tilde{G}_{j}^{0}$ . $\\Delta G_{j}$ are then aggregated with initial features $\\tilde{G}_{j}$ to derive refined features with improved consistency:", + "bbox": [ + 89, + 656, + 483, + 751 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\acute {G} _ {j} = G _ {j} + \\Delta G _ {j}. \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 223, + 760, + 482, + 781 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Step 6: Inverse Channel Rearrange. Finally, the features $\\{\\hat{G}_j\\}_{j=0}^{C-1}$ are restored to $\\{\\hat{F}_i\\}_{i=-n}^n$ via inverse channel rearrangement (ICR), i.e., the inverse operation of CR, as shown in Fig. 6.", + "bbox": [ + 89, + 792, + 482, + 851 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "MACF alleviates disturbances in $\\{F_i\\}_{i = -n}^n$ and generate improved features $\\{\\dot{F}_i\\}_{i = -n}^n$ , which are beneficial for enhancing SR reconstruction quality.", + "bbox": [ + 89, + 852, + 483, + 901 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e4afe44b1bf93eb80fd5222ea4d8074623b4a93178ad135d3fe086f41774c847.jpg", + "image_caption": [ + "Figure 8. Illustration of motion aligned super resolved reconstruction module." + ], + "image_footnote": [], + "bbox": [ + 517, + 88, + 903, + 232 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.3. Motion Aligned Super Resolved Reconstruction", + "text_level": 1, + "bbox": [ + 511, + 280, + 903, + 310 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The improved features $\\{\\hat{F}_i\\}_{i = -n}^n$ are then input to motion aligned super resolved reconstruction module to generate the high resolution image $\\mathcal{I}(t_0)$ , as illustrated in Fig. 8. Firstly, features at neighboring moments $\\{\\hat{F}_i\\}, i = \\pm 1,\\dots,\\pm n$ are aligned with the key feature $\\hat{F}_0$ by deformable convolution [7, 56], where the features $\\{\\hat{F}_i\\}_{i = -n}^n$ with improved consistency help enhance the accuracy of motion alignment, benefiting the utilization of temporal information. Subsequently, key feature and aligned neighboring features are concatenated and fused through several convolutional layers. The fused feature is upsampled to the HR image by pixel shuffle operation [3].", + "bbox": [ + 511, + 316, + 906, + 501 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6. Experiments", + "text_level": 1, + "bbox": [ + 511, + 515, + 645, + 532 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6.1. Implementation Details", + "text_level": 1, + "bbox": [ + 511, + 541, + 730, + 556 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We set the number of input spike frames to 101. The number of partitioned short-term blocks is 9 ( $K = 9$ , $n = 4$ ). The temporal length $L$ of each block $B_{i}$ is 21. The channel $C$ of $F_{i}$ and $\\hat{F}_i$ is 64. During training, we randomly crop the spike stream to a spatial size of $64\\times 64$ . In addition, we randomly perform horizontal and vertical flips as well as rotation for data augmentation. The batch size is set to 8. We use Adam optimizer [23] with $\\beta_{1} = 0.9$ and $\\beta_{2} = 0.999$ . The learning rate is initialized as 2e-4 and decayed by 0.7 every 100 epochs. The model is trained with $\\mathcal{L}_1$ loss for 800 epochs. All models are trained and tested on NVIDIA RTX 3090 GPU.", + "bbox": [ + 511, + 564, + 906, + 744 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6.2. Datasets", + "text_level": 1, + "bbox": [ + 511, + 757, + 614, + 771 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Synthesized Dataset. We need a large number of \"LR spike stream and HR image\" pairs to train the model. However, it is challenging to obtain corresponding high-quality HR images, especially in high dynamic scenes. Inspired by the experiments conducted in [48, 50], we use a spike camera simulator (detailed in supplementary material) to mimic the working mechanism of spike camera and generate synthesized SCSR datasets for training and evaluation, where", + "bbox": [ + 511, + 779, + 906, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "11421", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/1dc90f30285c37a3b27226fbf30eaa283fda1627b90193e5494562ee82e36d88.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ScaleMethodREDS-based DatasetAdobe240-based DatasetParams(M)Runtime(s)
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
×2TFP+DAT25.550.59540.445627.660.67820.367714.654.40
Spk2ImgNet+DAT31.770.90240.190734.920.96040.094918.564.63
WGSE+DAT31.540.89420.205534.750.95710.102218.604.50
BSF+DAT31.920.90440.186735.030.96100.093417.135.01
VidarSR31.970.89640.204035.130.95580.110510.162.09
SpikeSR-Net34.290.93300.148137.040.97040.08383.262.14
Ours34.500.93770.140337.220.97250.08163.720.81
×3TFP+DAT24.880.57540.446426.170.65770.372114.841.88
Spk2ImgNet+DAT28.900.82930.284930.500.90640.170818.752.10
WGSE+DAT28.760.82170.296130.330.90130.177918.782.02
BSF+DAT28.990.83150.280730.440.90540.169917.322.02
VidarSR29.550.83730.279132.010.91970.168511.380.85
SpikeSR-Net31.090.87980.230433.740.94390.13553.301.35
Ours31.410.88500.223134.110.94700.13273.900.37
×4TFP+DAT24.350.56170.473425.090.64070.399514.801.06
TFP+BasicVSR++23.690.50230.485024.210.56100.44157.320.08
Spk2ImgNet+DAT27.230.76780.352127.970.84590.237818.711.15
Spk2ImgNet+BasicVSR++27.280.77300.352727.770.84210.252611.230.17
WGSE+DAT27.150.76180.359727.910.84160.242718.751.08
WGSE+BasicVSR++27.160.76990.356127.720.84040.255811.080.12
BSF+DAT27.300.77010.347327.900.84510.235817.281.10
BSF+BasicVSR++27.240.78180.337427.430.83890.24819.800.18
VidarSR28.300.79600.324930.180.88340.209712.790.75
SpikeSR-Net29.220.82660.294631.310.90730.18853.341.07
Ours29.500.83690.281631.490.91190.18223.860.21
", + "bbox": [ + 106, + 88, + 893, + 497 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/3d138d00e430ff818b2df16dfa58991cf491f268873a6ed248f5033bd58050d2.jpg", + "table_caption": [ + "Table 1. Quantitative results on REDS-based and Adobe240-based SCSR evaluation dataset. Red and blue indicate the best and the second-best performance, respectively. It is noted that BasicVSR++ [4] only support $\\times 4$ SR scale. The runtime is tested on input spike stream of sizes $360 \\times 640 \\times 101 (\\times 2)$ , $240 \\times 426 \\times 101 (\\times 3)$ and $180 \\times 320 \\times 101 (\\times 4)$ , respectively." + ], + "table_footnote": [], + "table_body": "
CaseSIRMACFPSNR↑SSIM↑
step 3step 4step 5
(A)29.070.8207
(B)✓29.150.8237
(C)✓29.170.8244
(D)✓29.320.8306
(E)✓✓29.340.8316
(F)✓✓29.350.8322
(G)✓✓29.330.8304
(H)✓✓✓29.460.8344
(I)✓✓✓✓29.500.8369
", + "bbox": [ + 94, + 556, + 480, + 738 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Ablation study of the proposed Spk2SRImgNet on REDS-based SCSR evaluation dataset $(\\times 4)$", + "bbox": [ + 89, + 747, + 482, + 775 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "each sample consists of a spike stream-image pair.", + "bbox": [ + 89, + 779, + 424, + 794 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Specifically, we use the 120 fps REDS dataset [30] with a spatial resolution of $720 \\times 1280$ to generate the synthesized dataset. REDS includes 240 scenes for training and 30 scenes for evaluation, and we follow its original partition. For each scene, we choose 5 segments, resulting in $240 \\times 5 = 1200$ and $30 \\times 5 = 150$ pairs for training and evaluation, respectively. Each pair consists of spike", + "bbox": [ + 89, + 795, + 483, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "stream of size $360 \\times 640 \\times 101 \\times 2) / 240 \\times 426 \\times 101 \\times 3) / 180 \\times 320 \\times 101 \\times 4$ (101 is the temporal length) and a ground truth image of size $720 \\times 1280 \\times 2$ and $\\times 4$ / $720 \\times 1278 \\times 3$ . To demonstrate the generalization of the proposed model, we generate another SCSR evaluation dataset from Adobe240 dataset [34]. Adobe240 dataset includes 133 videos of 240 fps with resolution of $720 \\times 1280$ taken by hand-held cameras. We follow the test set partition of Adobe240 in [44], which includes 17 scenes. For each scene, we choose 10 segments, generating $17 \\times 10 = 170$ pairs for evaluation.", + "bbox": [ + 511, + 558, + 906, + 724 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Real-World Captured Data. In addition to the synthesized data, we also evaluate the model performance on real-world spike data captured by the spike camera with a spatial resolution of $250 \\times 400$ . Due to the absence of ground truth, we perform qualitative analysis for these data.", + "bbox": [ + 511, + 724, + 908, + 801 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6.3. Comparison with Existing Methods", + "text_level": 1, + "bbox": [ + 511, + 816, + 823, + 832 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We compare our proposed method with existing SCSR method: VidarSR [42] and SpikeSR-Net [50]. Besides, we combine four spike camera image reconstruction methods: TFP [53], Spk2ImgNet [48], WGSE [45] and BSF [52] with", + "bbox": [ + 511, + 839, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "11422", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9c2ca347e4b924fca87f0eeaabd76516ef0640d65a5e9f9f3102e02bcf853b86.jpg", + "image_caption": [ + "Spike" + ], + "image_footnote": [], + "bbox": [ + 93, + 89, + 228, + 183 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/edf5a0058cc08da47aeaf2a5c4cdc9beb9f2e63022ade05c3be3909e08bd26ad.jpg", + "image_caption": [ + "TFP+DAT" + ], + "image_footnote": [], + "bbox": [ + 228, + 90, + 362, + 183 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3ec66d28c1f792a036be8c30fbd3e32e1676b6c2e90a002ad23aa9cd4f63f6cb.jpg", + "image_caption": [ + "TFP+BasicVSR++" + ], + "image_footnote": [], + "bbox": [ + 362, + 90, + 496, + 183 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2aa6d8f6d3fc9e99e14ed5877f7677e051a8d803e97d6a5029507202e1c787a7.jpg", + "image_caption": [ + "Spk2ImgNet+DAT" + ], + "image_footnote": [], + "bbox": [ + 496, + 90, + 630, + 183 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/67a356828c4b89629382dec22ba463b7d0eeefaf6981262e297c0ee9ddc8094c.jpg", + "image_caption": [ + "Spk2ImgNet+BasicVSR++" + ], + "image_footnote": [], + "bbox": [ + 630, + 90, + 766, + 183 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2dd1c94efa4edc0d926bd0a1066ab9708513b19e822e38487a01b362a0c01d21.jpg", + "image_caption": [ + "WGSE+DAT" + ], + "image_footnote": [], + "bbox": [ + 767, + 90, + 901, + 183 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e6b86425aa247392e9321f625813d9c450f5108d7ad88fb201ccd7c76d2f9c8a.jpg", + "image_caption": [ + "WGSE+BasicVSR++" + ], + "image_footnote": [], + "bbox": [ + 94, + 194, + 227, + 286 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3fe5b20d437dbeb07174671b2f94f5f017f72b7fe115a856e8e4b986c784ea72.jpg", + "image_caption": [ + "BSF+DAT" + ], + "image_footnote": [], + "bbox": [ + 228, + 194, + 362, + 286 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ea64bfaa7344eed273c514d5f555b6575630351c4c9d05e2ad3f1032897ae654.jpg", + "image_caption": [ + "BSF+BasicVSR++" + ], + "image_footnote": [], + "bbox": [ + 362, + 194, + 496, + 286 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d46f6e6f139f61284ace99c9a38d148fb565f9f911d84c598bb019b3e5272db6.jpg", + "image_caption": [ + "VidarSR" + ], + "image_footnote": [], + "bbox": [ + 496, + 194, + 630, + 286 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/51f15c14bf48fedb3509deb57bdeafd18590766bf48080f3fab9d3d11b8d974e.jpg", + "image_caption": [ + "SpikeSR-Net" + ], + "image_footnote": [], + "bbox": [ + 630, + 194, + 766, + 286 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c4a67c6d2ddc6fbc50b1a32b34f39f69d02d81f38590a86a13b501bd509fc11b.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 767, + 194, + 901, + 286 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/88a6b2f1bf8d21314180e59a00823a09b6c24d8fed1d8137943ac84578d01027.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 94, + 299, + 227, + 392 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/82e5a4ae2802552d2a6191889da7c10db672c987f8526a61e7f1939cdd98734a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 228, + 299, + 362, + 392 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/fee3930eac04be89d9a12881237818cedd0d6709309351dcddd25976f8611d87.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 362, + 299, + 496, + 392 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1f8415383d0f7909ea2829b17a6828c08d7db9ab6037c015dd702cbbaff4552d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 496, + 299, + 630, + 392 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ffc0b050aa9dd4169f8e6e65b542be65edf8758d46e7320c43d8647747d083fb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 630, + 299, + 766, + 392 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/943720ade2f9e54c09b579f45416bd23b83dedfd061211ea3d881e27e640fcfc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 767, + 299, + 901, + 392 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/392c45de6f37c59b29fb1b8c240885fc207d57331002d8d36ad2aeb132141268.jpg", + "image_caption": [ + "Spike", + "WGSE+BasicVSR++", + "Figure 9. Visual comparison $(\\times 4)$ on real-world spike data. The first scene is a stationary clock recorded by a fast-moving spike camera. The second scene is a rotating fan. Please enlarge the figure for better comparison." + ], + "image_footnote": [], + "bbox": [ + 94, + 402, + 227, + 496 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/bab5e91e3ea16b40b013c32ef4b377f8a8e1d1762327adb18577e1df2dee4cef.jpg", + "image_caption": [ + "TFP+DAT", + "BSF+DAT" + ], + "image_footnote": [], + "bbox": [ + 228, + 402, + 362, + 496 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/fa1139c06a0f2928144c25a979fe5dcb7aeb69ceeda0cf82107b826098caca39.jpg", + "image_caption": [ + "TFP+BasicVSR++", + "BSF+BasicVSR++" + ], + "image_footnote": [], + "bbox": [ + 362, + 402, + 496, + 496 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3937c6da29913f837123c7962252d6e7d5e0a787ba184346e859b18099e9c37f.jpg", + "image_caption": [ + "Spk2ImgNet+DAT", + "VidarSR" + ], + "image_footnote": [], + "bbox": [ + 496, + 402, + 630, + 496 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a32869622dd3b7ae426783078b3ec66e3402d53361ffa21e661e49375bd6259a.jpg", + "image_caption": [ + "Spk2ImgNet+BasicVSR++", + "SpikeSR-Net" + ], + "image_footnote": [], + "bbox": [ + 630, + 402, + 766, + 496 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/28c2d0a3f7e3ac01de22cc6354db7632c128e8f2f6abef6c2c8b9eff92e6895a.jpg", + "image_caption": [ + "WGSE+DAT", + "Ours" + ], + "image_footnote": [], + "bbox": [ + 767, + 402, + 901, + 496 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "image / video SR method: DAT [5] / BasicVSR++ [4], resulting in another eight comparison methods, as presented in Tab. 1. We conduct experiments on $\\times 2$ , $\\times 3$ , and $\\times 4$ SR scales. We employ PSNR, SSIM [39] and LPIPS [46] as quantitative metrics. All compared methods except TFP [53] are retrained on the REDS-based SCSR training dataset with the same settings as ours. As shown in Tab. 1, our method achieves the best performance on the three metrics. Besides, we conduct experiments on real-world captured spike stream, as shown in Fig. 9, our method can restore better textures and details. More evaluation and visual results are provided in supplementary material.", + "bbox": [ + 89, + 556, + 483, + 738 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6.4. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 755, + 243, + 771 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To verify the effectiveness of our methodology, we perform ablation study with REDS-based evaluation dataset on $\\times 4$ SR scale, as presented in Tab. 2. Since ablation study only involves steps 3, 4, and 5 in MACF, the remaining steps are omitted in Tab. 2 (Steps 2 and 6 do not involve network parameters, and step 1 is present only when step 3 and/or step 5 are present). Case (I) is the final model. Case (A) is formed by removing SIR and MACF modules from", + "bbox": [ + 89, + 779, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Spk2SRImgNet. Comparisons between (A) and (B), (H) and (I) illustrate the advantages of converting information from spike domain to image domain (SIR module). Comparisons between (A) and (H), (B) and (I) demonstrate the effectiveness of MACF module, showing gains of 0.39dB and 0.35dB, respectively. Comparisons between (C-G) and (H) validate the effectiveness of each step in MACF. More ablation study is provided in supplementary material.", + "bbox": [ + 511, + 556, + 906, + 679 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 693, + 633, + 709 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper develops an end-to-end neural network named Spk2SRImgNet to reconstruct high resolution images from low resolution spike stream. To handle spike fluctuations, we propose a motion aligned collaborative filtering (MACF) module motivated by key ideas in classic image restoration schemes. Specifically, MACF leverages the temporal similarity of spike stream to enhance the consistency of multi-moment features along the motion trajectory, mitigating fluctuations and improving the quality of super resolution reconstruction. Experimental results demonstrate that the proposed Spk2SRImgNet achieves state-of-the-art performance on synthesized and real-captured spike data.", + "bbox": [ + 511, + 719, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "11423", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 91, + 90, + 258, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported by the National Natural Science Foundation of China under Grants 62072009 and U22B2035.", + "bbox": [ + 89, + 114, + 485, + 160 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 172, + 187, + 188 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Patrick Bardow, Andrew J. Davison, and Stefan Leutenegger. Simultaneous optical flow and intensity estimation from an event camera. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 884-892, 2016. 2", + "[2] Christian Brandli, Raphael Berner, Minhao Yang, Shih-Chii Liu, and Tobi Delbruck. A $240 \\times 180$ 130 db $3\\mu s$ latency global shutter spatiotemporal vision sensor. IEEE Journal of Solid-State Circuits, 49(10):2333-2341, 2014. 1", + "[3] Jose Caballero, Christian Ledig, Andrew Aitken, Alejandro Acosta, Johannes Totz, Zehan Wang, and Wenzhe Shi. Real-time video super-resolution with spatio-temporal networks and motion compensation. In 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2848-2857, 2017. 2, 6", + "[4] Kelvin C.K. Chan, Shangchen Zhou, Xiangyu Xu, and Chen Change Loy. Basicvsr++: Improving video superresolution with enhanced propagation and alignment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5972-5981, 2022. 2, 7, 8", + "[5] Zheng Chen, Yulun Zhang, Jinjin Gu, Linghe Kong, Xiaokang Yang, and Fisher Yu. Dual aggregation transformer for image super-resolution. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 12312-12321, 2023. 8", + "[6] Kostadin Dabov, Alessandro Foi, Vladimir Katkovnik, and Karen Egiazarian. Image denoising by sparse 3-d transform-domain collaborative filtering. IEEE Transactions on Image Processing, 16(8):2080-2095, 2007. 2, 3", + "[7] Jifeng Dai, Haozhi Qi, Yuwen Xiong, Yi Li, Guodong Zhang, Han Hu, and Yichen Wei. Deformable convolutional networks. In 2017 IEEE International Conference on Computer Vision (ICCV), pages 764-773, 2017. 2, 6", + "[8] Siwei Dong, Tiejun Huang, and Yonghong Tian. Spike camera and its coding methods. In 2017 Data Compression Conference (DCC), pages 437-437, 2017. 1", + "[9] Siwei Dong, Lin Zhu, Daoyuan Xu, Yonghong Tian, and Tiejun Huang. An efficient coding method for spike camera using inter-spike intervals. In 2019 Data Compression Conference (DCC), pages 568-568, 2019. 1", + "[10] Yanchen Dong, Jing Zhao, Ruiqin Xiong, and Tiejun Huang. 3D residual interpolation for spike camera demosaicing. In IEEE International Conference on Image Processing (ICIP), pages 1461-1465. IEEE, 2022. 2", + "[11] Yanchen Dong, Ruiqin Xiong, Jian Zhang, Zhaofei Yu, Xiaopeng Fan, Shuyuan Zhu, and Tiejun Huang. Superresolution reconstruction from bayer-pattern spike streams. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 24871-24880, 2024. 2" + ], + "bbox": [ + 93, + 198, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[12] Yanchen Dong, Ruiqin Xiong, Jing Zhao, Jian Zhang, Xiaopeng Fan, Shuyuan Zhu, and Tiejun Huang. Joint demosaicing and denoising for spike camera. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1582-1590, 2024. 2", + "[13] Yanchen Dong, Ruiqin Xiong, Jing Zhao, Jian Zhang, Xiaopeng Fan, Shuyuan Zhu, and Tiejun Huang. Learning a deep demosaicing network for spike camera with color filter array. IEEE Transactions on Image Processing, 33:3634-3647, 2024.", + "[14] Yanchen Dong, Ruiqin Xiong, Xiaopeng Fan, Shuyuan Zhu, Jin Wang, and Tiejun Huang. Dynamic scene reconstruction for color spike camera via zero-shot learning. IEEE Transactions on Computational Imaging, 11:129-141, 2025. 2", + "[15] Daniel Gehrig and Davide Scaramuzza. Low-latency automotive vision with event cameras. Nature, 629(8014):1034-1040, 2024. 1", + "[16] Arren Glover, Valentina Vasco, and Chiara Bartolozzi. A controlled-delay event camera framework for on-line robotics. In 2018 IEEE International Conference on Robotics and Automation (ICRA), pages 2178-2183. IEEE, 2018. 1", + "[17] Sorin Grigorescu, Bogdan Trasnea, Tiberiu Cocias, and Gigel Macesanu. A survey of deep learning techniques for autonomous driving. Journal of field robotics, 37(3):362-386, 2020. 1", + "[18] Shuhang Gu, Lei Zhang, Wangmeng Zuo, and Xiangchu Feng. Weighted nuclear norm minimization with application to image denoising. In 2014 IEEE Conference on Computer Vision and Pattern Recognition, pages 2862-2869, 2014. 2, 3, 4", + "[19] Tiejun Huang, Yajing Zheng, Zhaofei Yu, Rui Chen, Yuan Li, Ruiqin Xiong, Lei Ma, Junwei Zhao, Siwei Dong, Lin Zhu, Jianing Li, Shanshan Jia, Yihua Fu, Boxin Shi, Si Wu, and Yonghong Tian. $1000\\times$ faster camera and machine vision with ordinary devices. Engineering, 25:110-119, 2023. 1", + "[20] Arpit Jadon, Mohd Omama, Akshay Varshney, Mohammad Samar Ansari, and Rishabh Sharma. Firenet: A specialized lightweight fire & smoke detection model for real-time IoT applications. arXiv preprint arXiv:1905.11922, 2019. 2", + "[21] Armin Kappeler, Seunghwan Yoo, Qiqin Dai, and Aggelos K. Katsaggelos. Video super-resolution with convolutional neural networks. IEEE Transactions on Computational Imaging, 2(2):109-122, 2016. 2", + "[22] Hanme Kim, Ankur Handa, Ryad Benosman, Sio-Hoi Ieng, and Andrew Davison. Simultaneous mosaicing and tracking with an event camera. In Proceedings of the British Machine Vision Conference. BMVA Press, 2014. 2", + "[23] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6", + "[24] Jingyun Liang, Jiezhang Cao, Yuchen Fan, Kai Zhang, Rakesh Ranjan, Yawei Li, Radu Timofte, and Luc Van Gool. Vrt: A video restoration transformer. IEEE Transactions on Image Processing, 33:2171-2182, 2024. 2", + "[25] Patrick Lichtsteiner, Christoph Posch, and Tobi Delbruck. A $128 \\times 128$ 120 db $15\\mu s$ latency asynchronous temporal con" + ], + "bbox": [ + 516, + 92, + 906, + 902 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "11424", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "trast vision sensor. IEEE journal of solid-state circuits, 43 (2):566-576, 2008. 1", + "[26] Martin Litzenberger, Christoph Posch, D Bauer, Ahmed Nabil Belbachir, P Schon, B Kohn, and H Garn. Embedded vision system for real-time object tracking using an asynchronous transient vision sensor. In 2006 IEEE 12th Digital Signal Processing Workshop & 4th IEEE Signal Processing Education Workshop, pages 173-178. IEEE, 2006. 1", + "[27] Julien Mairal, Francis Bach, Jean Ponce, Guillermo Sapiro, and Andrew Zisserman. Non-local sparse models for image restoration. In 2009 IEEE 12th International Conference on Computer Vision, pages 2272-2279, 2009. 2, 3, 4", + "[28] Mohammad Mostafavi, Yeongwoo Nam, Jonghyun Choi, and Kuk-Jin Yoon. E2sri: Learning to super-resolve intensity images from events. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(10):6890-6909, 2022. 2", + "[29] Elias Mueggler. Event-based vision for high-speed robotics. PhD thesis, University of Zurich, 2017. 1", + "[30] Seungjun Nah, Sungyong Baik, Seokil Hong, Gyeongsik Moon, Sanghyun Son, Radu Timofte, and Kyoung Mu Lee. Ntire 2019 challenge on video deblurring and superresolution: Dataset and study. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2019. 7", + "[31] Christoph Posch, Daniel Matolin, and Rainer Wohlgenannt. A qvga 143 db dynamic range frame-free pwm image sensor with lossless pixel-level video compression and time-domain cds. IEEE Journal of Solid-State Circuits, 46(1):259-275, 2010. 1", + "[32] Henri Rebecq, René Ranftl, Vladlen Koltun, and Davide Scaramuzza. Events-to-video: Bringing modern computer vision to event cameras. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3857-3866, 2019. 2", + "[33] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention-MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, pages 234-241. Springer, 2015. 6", + "[34] Shuochen Su, Maurizio Delbrazio, Jue Wang, Guillermo Sapiro, Wolfgang Heidrich, and Oliver Wang. Deep video deblurring for hand-held cameras. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 1279-1288, 2017. 7", + "[35] Yapeng Tian, Yulun Zhang, Yun Fu, and Chenliang Xu. Tdan: Temporally-deformable alignment network for video super-resolution. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2", + "[36] Lin Wang, Tae-Kyun Kim, and Kuk-Jin Yoon. Eventsr: From asynchronous events to image reconstruction, restoration, and super-resolution via end-to-end adversarial learning. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8312-8322, 2020. 2", + "[37] Xintao Wang, Kelvin C.K. Chan, Ke Yu, Chao Dong, and Chen Change Loy. Edvr: Video restoration with enhanced" + ], + "bbox": [ + 91, + 92, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "deformable convolutional networks. In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 1954-1963, 2019. 2", + "[38] Yuanlin Wang, Ruiqin Xiong, Jing Zhao, and Tiejun Huang. Reconstruct dynamic scene for spike camera based on 3d space time similarity. In 2024 IEEE International Conference on Image Processing (ICIP), pages 1595-1601, 2024. 2", + "[39] Zhou Wang, A.C. Bovik, H.R. Sheikh, and E.P. Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing, 13(4): 600-612, 2004. 8", + "[40] Philippe Weinzaepfel, Jerome Revaud, Zaid Harchaoui, and Cordelia Schmid. Deepflow: Large displacement optical flow with deep matching. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), pages 1385-1392, 2013. 2", + "[41] Wenming Weng, Yueyi Zhang, and Zhiwei Xiong. Event-based video reconstruction using transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 2", + "[42] Xijie Xiang, Lin Zhu, Jianing Li, Yixuan Wang, Tiejun Huang, and Yonghong Tian. Learning super-resolution reconstruction for high temporal resolution spike stream. IEEE Transactions on Circuits and Systems for Video Technology, 33(1):16-29, 2023. 1, 2, 7", + "[43] Ruiqin Xiong, Hangfan Liu, Xinfeng Zhang, Jian Zhang, Siwei Ma, Feng Wu, and Wen Gao. Image denoising via bandwise adaptive modeling and regularization exploiting nonlocal similarity. IEEE Transactions on Image Processing, 25 (12):5793-5805, 2016. 2, 3, 4", + "[44] Gang Xu, Jun Xu, Zhen Li, Liang Wang, Xing Sun, and Ming-Ming Cheng. Temporal modulation network for controllable space-time video super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6388-6397, 2021. 7", + "[45] Jiyuan Zhang, Shanshan Jia, Zhaofei Yu, and Tiejun Huang. Learning temporal-ordered representation for spike streams based on discrete wavelet transforms. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 137–147, 2023. 7", + "[46] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 586-595, 2018. 8", + "[47] Jing Zhao, Jiyu Xie, Ruiqin Xiong, Jian Zhang, Zhaofei Yu, and Tiejun Huang. Super resolve dynamic scene from continuous spike streams. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 2533-2542, 2021. 2", + "[48] Jing Zhao, Ruiqin Xiong, Hangfan Liu, Jian Zhang, and Tiejun Huang. Spk2imgnet: Learning to reconstruct dynamic scene from continuous spike stream. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11996-12005, 2021. 2, 6, 7" + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "11425", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[49] Jing Zhao, Ruiqin Xiong, Jiyu Xie, Boxin Shi, Zhaofei Yu, Wen Gao, and Tiejun Huang. Reconstructing clear image for high-speed motion scene with a retina-inspired spike camera. IEEE Transactions on Computational Imaging, 8:12-27, 2021. 3", + "[50] Jing Zhao, Ruiqin Xiong, Jian Zhang, Rui Zhao, Hangfan Liu, and Tiejun Huang. Learning to super-resolve dynamic scenes for neuromorphic spike camera. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3579-3587, 2023. 1, 2, 6, 7", + "[51] Rui Zhao, Ruiqin Xiong, Jian Zhang, Zhaofei Yu, Shuyuan Zhu, Lei Ma, and Tiejun Huang. Spike camera image reconstruction using deep spiking neural networks. IEEE Transactions on Circuits and Systems for Video Technology, 34(6): 5207-5212, 2024. 2", + "[52] Rui Zhao, Ruiqin Xiong, Jing Zhao, Jian Zhang, Xiaopeng Fan, Zhaofei Yu, and Tiejun Huang. Boosting spike camera image reconstruction from a perspective of dealing with spike fluctuations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 24955-24965, 2024. 2, 7", + "[53] Lin Zhu, Siwei Dong, Tiejun Huang, and Yonghong Tian. A retina-inspired sampling method for visual texture reconstruction. In 2019 IEEE International Conference on Multimedia and Expo (ICME), pages 1432-1437. IEEE, 2019. 2, 3, 4, 7, 8", + "[54] Lin Zhu, Siwei Dong, Jianing Li, Tiejun Huang, and Yonghong Tian. Ultra-high temporal resolution visual reconstruction from a fovea-like spike camera via spiking neuron model. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(1):1233-1249, 2022. 2", + "[55] Lin Zhu, Xiao Wang, Yi Chang, Jianing Li, Tiejun Huang, and Yonghong Tian. Event-based video reconstruction via potential-assisted spiking neural network. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3584-3594, 2022. 2", + "[56] Xizhou Zhu, Han Hu, Stephen Lin, and Jifeng Dai. Deformable convnets v2: More deformable, better results. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9308-9316, 2019. 2, 6" + ], + "bbox": [ + 89, + 92, + 482, + 667 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11426", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Spk2SRImgNet_ Super-Resolve Dynamic Scene from Spike Stream via Motion Aligned Collaborative Filtering/e0bcc965-cf40-409b-a203-db12626ad68b_model.json b/2025/Spk2SRImgNet_ Super-Resolve Dynamic Scene from Spike Stream via Motion Aligned Collaborative Filtering/e0bcc965-cf40-409b-a203-db12626ad68b_model.json new file mode 100644 index 0000000000000000000000000000000000000000..e54ac435ddcea2540a74e02d77626fa9f146df65 --- /dev/null +++ b/2025/Spk2SRImgNet_ Super-Resolve Dynamic Scene from Spike Stream via Motion Aligned Collaborative Filtering/e0bcc965-cf40-409b-a203-db12626ad68b_model.json @@ -0,0 +1,2675 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.163, + 0.13, + 0.838, + 0.177 + ], + "angle": 0, + "content": "Spk2SRImgNet: Super-Resolve Dynamic Scene from Spike Stream via Motion Aligned Collaborative Filtering" + }, + { + "type": "text", + "bbox": [ + 0.237, + 0.203, + 0.762, + 0.24 + ], + "angle": 0, + "content": "Yuanlin Wang\\(^{1,2}\\), Yiyang Zhang\\(^{1,2}\\), Ruiqin Xiong\\(^{1,2,*}\\), Jing Zhao\\(^{1,2}\\), Jian Zhang\\(^{3}\\), Xiaopeng Fan\\(^{4}\\), Tiejun Huang\\(^{1,2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.307, + 0.24, + 0.692, + 0.257 + ], + "angle": 0, + "content": "\\(^{1}\\)School of Computer Science, Peking University" + }, + { + "type": "text", + "bbox": [ + 0.164, + 0.257, + 0.834, + 0.275 + ], + "angle": 0, + "content": "\\(^{2}\\)National Key Laboratory for Multimedia Information Processing, Peking University" + }, + { + "type": "text", + "bbox": [ + 0.12, + 0.275, + 0.879, + 0.293 + ], + "angle": 0, + "content": "\\(^{3}\\)School of Electronic and Computer Engineering, Shenzhen Graduate School, Peking University" + }, + { + "type": "text", + "bbox": [ + 0.192, + 0.293, + 0.806, + 0.311 + ], + "angle": 0, + "content": "\\(^{4}\\)School of Computer Science and Technology, Harbin Institute of Technology" + }, + { + "type": "text", + "bbox": [ + 0.382, + 0.314, + 0.616, + 0.327 + ], + "angle": 0, + "content": "wangyuanlin@stu.pku.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.149, + 0.33, + 0.843, + 0.345 + ], + "angle": 0, + "content": "{yiyangz, rqxiong, jzhaopku, zhangjian.sz, tjhuang}@pku.edu.cn fxp@hit.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.38, + 0.327, + 0.396 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.412, + 0.486, + 0.807 + ], + "angle": 0, + "content": "Spike camera is a kind of neuromorphic camera that records dynamic scenes by firing a stream of binary spikes with extremely high temporal resolution. It demonstrates great potential for vision tasks in high-speed scenarios. One limitation in its current implementation is the relatively low spatial resolution. This paper develops a network called Spk2SRImgNet to super-resolve high resolution images from low resolution spike stream. However, fluctuations in spike stream hinder the performance of spike camera super resolution. To address this issue, we propose a motion aligned collaborative filtering (MACF) module, which is motivated by key ideas in classic image restoration schemes to mitigate fluctuations in spike data. MACF leverages the temporal similarity of spike stream to acquire similar features from neighboring moments via motion alignment. To separate disturbances from features, MACF filters these similar features jointly in transform domain to exploit representation sparsity, and generates refinement features that will be used to update initial fluctuated features. Specifically, MACF designs an inverse motion alignment operation to map these refinement features back to their original positions. The initial features are aggregated with the repositioned refinement features to enhance reliability. Experimental results demonstrate that the proposed method achieves state-of-the-art performance compared with existing methods." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.819, + 0.222, + 0.833 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.844, + 0.484, + 0.875 + ], + "angle": 0, + "content": "High-speed imaging attracts increasing attention in emerging computer vision applications such as autonomous driv-" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.38, + 0.905, + 0.453 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.463, + 0.907, + 0.507 + ], + "angle": 0, + "content": "Figure 1. Illustration of spike camera super resolution (SCSR) and visual comparison \\((\\times 4)\\) with existing SCSR methods: VidarSR [42] and SpikeSR-Net [50]. Please zoom in for better comparison." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.519, + 0.907, + 0.64 + ], + "angle": 0, + "content": "ing [15, 17] and robotics [16, 29]. Conventional camera employs an exposure-based imaging model, which limits its ability to capture high-speed scenes. Specifically, conventional camera accumulates photoelectric signals over an exposure time window to form a snapshot image, which requires the scene to be still during the exposure period. Otherwise, a single point on a moving object may be projected onto different sensor pixels, resulting in motion blur." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.642, + 0.909, + 0.884 + ], + "angle": 0, + "content": "Neuromorphic camera is an emerging camera that abandons the concept of exposure window and mimics the working mechanism of biological retina. It records intensity information with high temporal resolution, showing great potential for high speed vision tasks. One kind of neuromorphic camera is event camera [2, 25, 26, 31]. It detects light intensity changes of each pixel in the logarithmic domain and fires an event when the change exceeds a specified threshold. Another kind of neuromorphic camera is spike camera [8, 9, 19]. Each pixel of spike camera independently and continuously accumulates photons, and fires a spike when the accumulation exceeds a predefined threshold. Unlike event camera that only records relative changes of light intensity, spike camera captures absolute intensity by firing a stream of binary spikes, providing clearer information for reconstructing image details." + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.886, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Benefiting from its extremely high temporal resolution" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.236, + 0.9 + ], + "angle": 0, + "content": "*Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11416" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.198 + ], + "angle": 0, + "content": "(e.g. \\(40\\mathrm{kHz}\\)), spike camera could sample a large amount of intensity information in a short period. Besides, these samples cover different locations of the scene due to the relative motion between camera and scene. Therefore, there is promising potential to reconstruct high resolution (HR) images from spike stream by converting the dense sampling in temporal domain to spatial domain." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.199, + 0.482, + 0.392 + ], + "angle": 0, + "content": "This paper aims to super-resolve high resolution images from low resolution spike stream. However, due to the Poisson effect of photon arrival, the quantization effect of spike readout, and the thermal noises introduced by circuits, spikes exhibit fluctuations, i.e., the length of spike accumulation period varies over time, even when light intensity remains unchanged. Thus, the features extracted from fluctuated spikes tend to be noisy and temporally incoherent. These unreliable features provide low-quality information source for super resolution (SR) reconstruction. Moreover, they affect the alignment accuracy when leveraging temporal information, further degrading the quality of SR reconstruction." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.395, + 0.483, + 0.77 + ], + "angle": 0, + "content": "In this paper, we propose a spike-to-super-resolution-image network Spk2SRImgNet. To handle spike fluctuations, we take inspiration from several classic image restoration methods, such as BM3D [6], BAS [43], LSSC [27], and WNNM [18]. These methods incorporate the idea of leveraging non-local similarity within natural images to reduce the unknown random disturbance. We embed these concepts into the design of Spk2SRImgNet and develop a motion aligned collaborative filtering (MACF) module to reduce disturbances in spike information. Firstly, MACF obtains similar features from neighboring moments via motion alignment, exploiting the temporal similarity of spike stream. To separate disturbances from features, MACF performs collaborative filtering on grouped similar features in transform domain for leveraging representation sparsity, and generates refinement features for each moment. These refinement features will be utilized to update initial fluctuated features. Specifically, MACF introduces an inverse motion alignment operation to map these multi-moment refinement features to their original unaligned positions. The initial features are integrated with the repositioned refinement features to improve reliability. Experimental results show that the proposed Spk2SRImgNet achieves state-of-the-art performance on synthesized and real-world spike data." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.785, + 0.233, + 0.801 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.482, + 0.901 + ], + "angle": 0, + "content": "Video Super Resolution. Video super resolution (VSR) produces high resolution images from low resolution (LR) image sequences. Temporal alignment is a crucial step to model temporal correspondences of misaligned frames. Early learning-based methods [3, 21] employ traditional optical flow estimation [40] for motion compensation. With" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.182 + ], + "angle": 0, + "content": "the diverse development of network structures, TDAN [35] and EDVR [37] utilize deformable convolution [7, 56] for feature alignment. BasicVSR++ [4] leverages optical flow to guide deformable alignment for stable training. VRT [24] proposes a temporal mutual attention block to facilitate motion estimation and feature alignment." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.183, + 0.905, + 0.378 + ], + "angle": 0, + "content": "Event Camera Reconstruction and Super Resolution. Event camera monitors relative changes in light intensity with high temporal resolution. Reconstructing intensity images from events has been studied in the past decade. Kim et al. [22] propose to use the extended Kalman filter and Poisson reconstruction for event reconstruction. Bardow et al. [1] estimate both intensity and optical flow using the primal-dual algorithm. With the development of deep learning, E2VID [32], FireNet [20], ET-Net [41], EVSNN [55] have been proposed successively with good performance. To reconstruct HR images from LR events, Wang et al. [36] propose an unsupervised GAN EventSR. Mostafavi et al. [28] propose an end-to-end recurrent network E2SRI." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.379, + 0.905, + 0.635 + ], + "angle": 0, + "content": "Spike Camera Reconstruction and Super Resolution. Image reconstruction is a popular topic for spike camera [38, 48, 51-54]. TFI [53] reconstructs images based on the interval between two temporally neighboring spikes. TFP [53] utilizes the number of spikes within a temporal window to estimate the intensity. For better performance, CNN-based [48, 52] and spike neural network (SNN)-based [51, 54] frameworks have been proposed successively. To reconstruct HR images from LR spike stream, Zhao et al. [47] derive the constraint relationship between pixel intensities and each spike period, and obtain HR images based on gradient descent algorithm. Xiang et al. [42] design a multi-level feature learning network to achieve spike camera super resolution. Zhao et al. [50] propose SpikeSR-Net based on the observation model of spike camera. Furthermore, Dong et al. [11] propose a super resolution network for the newly developed color spike camera [10, 12-14]." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.649, + 0.658, + 0.666 + ], + "angle": 0, + "content": "3. Spike Camera" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Working Mechanism. Spike camera contains a pixel array of size \\( H \\times W \\) and each pixel works independently. Each pixel comprises three main components: photon receptor, integrator, and comparator, as shown in Fig. 2. The photoreceptor captures incident photons from external scenes and converts them into electric charges through photoelectric conversion. The integrator accumulates the electric charges continuously. Meanwhile, the comparator checks the accumulated voltage persistently. Once the voltage reaches a predefined threshold \\( \\theta \\), the pixel sets up a spike flag and resets the integrator, immediately restarting a new \"accumulation-and-set\" cycle. Spike readout circuit periodically checks the spike flag at a frequency of \\( 40\\mathrm{kHz} \\). When detecting the flag, the circuit outputs a spike \"1\", otherwise, it outputs \"0\". The voltage accumulated by the in" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.956 + ], + "angle": 0, + "content": "11417" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.08, + 0.483, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.145, + 0.242, + 0.428, + 0.256 + ], + "angle": 0, + "content": "Figure 2. Working mechanism of spike camera." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.282, + 0.461, + 0.297 + ], + "angle": 0, + "content": "tegrator at an arbitrary time \\( t \\) can be formulated as [49]:" + }, + { + "type": "equation", + "bbox": [ + 0.15, + 0.304, + 0.483, + 0.339 + ], + "angle": 0, + "content": "\\[\nA (x, y, t) = \\int_ {0} ^ {t} \\eta I (x, y, u) d u \\quad \\mathrm {m o d} \\theta , \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.346, + 0.484, + 0.421 + ], + "angle": 0, + "content": "where \\(I(x,y,u)\\) denotes the arrival rate of photons received by pixel \\((x,y)\\) at time \\(u\\), \\(\\eta\\) denotes the photoelectric conversion rate. The output of spike camera is a 3D binary spike matrix \\(\\{S(u)\\}_{u=1}^{T}\\) with size \\(H\\times W\\times T\\), where \\(T\\) is the temporal length, \\(S(u)\\) is a spike frame at time \\(u\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.422, + 0.484, + 0.543 + ], + "angle": 0, + "content": "Spike Interval Based Reconstruction. According to the working mechanism of spike camera, the stronger the light intensity, the shorter the interval between two neighboring spikes, i.e., the spike interval is inversely proportional to the light intensity. This property is exploited by spike-interval-based reconstruction method (SIR) [53]. Specifically, the spike interval of two temporally neighboring spikes that cover the position \\((x,y,u)\\) can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.14, + 0.552, + 0.483, + 0.587 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\Phi (x, y, u) = \\min \\left\\{t \\mid S (x, y, t) = 1, t \\geq u \\right\\} \\tag {2} \\\\ - \\max \\{t | S (x, y, t) = 1, t < u \\}. \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.596, + 0.483, + 0.612 + ], + "angle": 0, + "content": "SIR estimates the light intensity of pixel \\((x,y)\\) at time \\(u\\) by" + }, + { + "type": "equation", + "bbox": [ + 0.202, + 0.62, + 0.482, + 0.652 + ], + "angle": 0, + "content": "\\[\nY (x, y, u) = \\frac {\\delta}{\\Phi (x , y , u)}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.66, + 0.483, + 0.688 + ], + "angle": 0, + "content": "where \\(\\delta\\) denotes the maximum dynamic range of reconstruction." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.484, + 0.9 + ], + "angle": 0, + "content": "Spike Fluctuations. The imaging of spike camera is affected by multiple kinds of noise. First, the Poisson process of photon arrival causes the time taken for the number of photons to reach the threshold to exhibit randomness, which is denoted as Poisson noise. Second, the spike readout circuit is controlled by a periodic clock signal, i.e., it only outputs spikes at discrete times, which results in a time difference between spike flag set and spike readout (see Fig. 2). This difference leads to quantization noise. Third, the thermal noise introduced by circuits brings randomness to spike fire. For the above reasons, the intervals of neighboring spikes exhibit fluctuations and struggle to accurately reflect light intensity, leading to noisy reconstruction results from SIR." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.09, + 0.71, + 0.105 + ], + "angle": 0, + "content": "4. Motivation and Idea" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.115, + 0.905, + 0.297 + ], + "angle": 0, + "content": "The key problem we aim to address is how to super-resolve the original dynamic scene from captured binary spike data to obtain high quality visual reconstruction with clean image details. According to the working mechanism of spike camera, each individual spike does not reveal the light intensity directly. Instead, it only tells the time point at which a spike firing occurs. Essentially, the instantaneous light intensity can be inferred only by considering the length of intervals between consecutive spikes. Unfortunately, for the reasons discussed above, the recorded spike intervals are not stable, and they fluctuate around an expectation value determined by actual light intensity." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.297, + 0.905, + 0.432 + ], + "angle": 0, + "content": "To handle such fluctuations, we draw inspiration from several classic image restoration methods, such as BM3D [6], BAS [43], LSSC [27], and WNNM [18]. Although they are quite diversified in algorithmic design, all these methods incorporate the idea of leveraging non-local similarity within natural image signals to reduce the unknown random disturbance. They utilize a group of similar patches and establish a prior model for them so that a regularization term is formed to separate the true signal from the noise." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.433, + 0.905, + 0.477 + ], + "angle": 0, + "content": "Suppose \\(\\mathbf{y} = \\mathbf{x} + \\mathbf{n}\\) is an image corrupted by noise, where \\(\\mathbf{x}\\) is the clean image and \\(\\mathbf{n}\\) is the noise. The idea of these methods can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.478, + 0.905, + 0.509 + ], + "angle": 0, + "content": "(1) For each pixel location \\(\\mathbf{p}\\), search for a group of similar patches from the noisy image \\(\\mathbf{y}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.636, + 0.52, + 0.905, + 0.537 + ], + "angle": 0, + "content": "\\[\n\\mathbf {y} _ {1}, \\mathbf {y} _ {2}, \\dots , \\mathbf {y} _ {N} = \\Lambda (\\mathbf {y} | \\mathbf {p}), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.539, + 0.546, + 0.905, + 0.605 + ], + "angle": 0, + "content": "where \\(N\\) is the number of similar patches, \\(\\Lambda (\\cdot |\\cdot)\\) is the operation of searching for and extracting similar patches. Note that the notation \\(\\mathbf{y}_i\\) should depend on \\(\\mathbf{p}\\) but we omit it for conciseness." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.606, + 0.904, + 0.637 + ], + "angle": 0, + "content": "(2) Exploit the correlation within the patch group \\(\\{\\mathbf{y}_i\\}_{i = 1}^N\\) to generate an estimate for each patch:" + }, + { + "type": "equation", + "bbox": [ + 0.621, + 0.646, + 0.905, + 0.665 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\mathbf {x}} _ {1}, \\tilde {\\mathbf {x}} _ {2}, \\dots , \\tilde {\\mathbf {x}} _ {N} = \\Pi \\left(\\left\\{\\mathbf {y} _ {i} \\right\\} _ {i = 1} ^ {N}\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.539, + 0.674, + 0.905, + 0.765 + ], + "angle": 0, + "content": "where \\(\\tilde{\\mathbf{x}}_i\\) is the estimate of the true signal \\(\\mathbf{x}_i\\) corresponding to \\(\\mathbf{y}_i\\). The actual operation of \\(\\Pi(\\cdot)\\) depends on the prior model adopted to characterize the consistency between patches \\(\\{\\mathbf{y}_i\\}_{i=1}^N\\). Generally, \\(\\Pi(\\cdot)\\) can be regarded as an operator to solve the following optimization objective:" + }, + { + "type": "equation", + "bbox": [ + 0.555, + 0.774, + 0.905, + 0.817 + ], + "angle": 0, + "content": "\\[\n\\left\\{\\tilde {\\mathbf {x}} _ {i} \\right\\} = \\arg \\min \\sum_ {i = 1} ^ {N} \\left| \\left| \\mathbf {x} _ {i} - \\mathbf {y} _ {i} \\right| \\right| _ {2} ^ {2} + \\Phi \\left(\\left\\{\\mathbf {x} _ {i} \\right\\} _ {i = 1} ^ {N}\\right) \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.539, + 0.825, + 0.905, + 0.902 + ], + "angle": 0, + "content": "BM3D [6] solves Eq. (6) by applying a 3D transform to the stacked similar patches and constructing a sparsity regularization term in the transform domain. It utilizes hard-thresholding, soft-thresholding, and Wiener filtering in different steps, corresponding to \\( l_{0} \\), \\( l_{1} \\), and \\( l_{2} \\)" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11418" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.089, + 0.905, + 0.189 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.2, + 0.908, + 0.258 + ], + "angle": 0, + "content": "Figure 3. The main idea of motion aligned collaborative filtering. First, we collect temporal similar features via motion alignment. Then, we stack these aligned features and perform transform-domain collaborative filtering to generate refinement features for each moment. Subsequently, the refinement features are unstacked and mapped back to their original positions via inverse warping. These repositioned refinement features are aggregated with input features to obtain refined features with improved consistency." + }, + { + "type": "text", + "bbox": [ + 0.116, + 0.266, + 0.485, + 0.49 + ], + "angle": 0, + "content": "sparsity norm, respectively. BAS [43] applies a content adaptive 2D decorrelation transform to the similar patches via principle component analysis (PCA), and builds a separate distribution model and regularization term for each transform band, using the transform coefficients of these patches as samples. It leads to a PCA-domain adaptive soft-thresholding to solve Eq. (6). LLSC [27] adopts another prior model that emphasizes that similar patches tend to share the location pattern of non-zero dictionary elements in their sparse decomposition. WNNM [18], on the other hand, utilizes the low-rank property of similar patches so that the problem (6) reduces to singular value decomposition (SVD) and soft-thresholding on the singular values." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.492, + 0.483, + 0.537 + ], + "angle": 0, + "content": "(3) Place each estimate \\(\\tilde{\\mathbf{x}}_i\\) back to the position where \\(\\mathbf{y}_i\\) is taken from, and aggregate the overlapping estimates from all position \\(\\mathbf{p}\\) to form a whole estimate for \\(\\mathbf{x}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.197, + 0.544, + 0.483, + 0.561 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\mathbf {x}} = \\Lambda^ {- 1} \\left(\\left\\{\\tilde {\\mathbf {x}} _ {1}, \\tilde {\\mathbf {x}} _ {2}, \\dots , \\tilde {\\mathbf {x}} _ {N} \\right\\} _ {\\mathbf {p}}\\right). \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.116, + 0.569, + 0.483, + 0.614 + ], + "angle": 0, + "content": "Here \\(\\Lambda^{-1}(\\cdot)\\) denotes patch reposition and aggregation operation. The reposition of \\(\\tilde{\\mathbf{x}}_i\\) to \\(\\tilde{\\mathbf{x}}\\) is the inverse operation of taking \\(\\mathbf{y}_i\\) from \\(\\mathbf{y}\\) in stage (1)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.615, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Taking these ideas into account, we propose a novel framework tailored for spike data to reduce disturbances in spike information. The main idea is illustrated in Fig. 3. The similarity exploited in this framework is temporal similarity, and it may be extended to include spatial non-local similarity in the future. First, we align the spike features at neighboring moments with the spike feature at the reconstruction moment to obtain a set of similar features. Then, we stack the aligned features and perform a neural-network-based transform-domain collaborative filtering, to separate the disturbances from clean features by exploiting representation sparsity and generate refinement features for each moment. These refinement features are used to update the initial input features. Specifically, the refinement features are unstacked and mapped back to their original positions via inverse warping, and then aggregated with input features. The purpose of this framework is to improve the consistency of multi-moment features along the motion trajectory." + }, + { + "type": "image", + "bbox": [ + 0.514, + 0.263, + 0.907, + 0.373 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.383, + 0.908, + 0.412 + ], + "angle": 0, + "content": "Figure 4. Overall architecture of the proposed spike camera super resolution network Spk2SRImgNet." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.428, + 0.605, + 0.443 + ], + "angle": 0, + "content": "5. Method" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.457, + 0.71, + 0.472 + ], + "angle": 0, + "content": "5.1. Overall Architecture" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.484, + 0.907, + 0.605 + ], + "angle": 0, + "content": "To reconstruct HR dynamic scenes from LR binary spike stream, we develop an end-to-end trainable spike camera super resolution network, Spk2SRImgNet. The overall architecture is shown in Fig. 4. To reconstruct an HR image \\(\\mathcal{I}(t_0) \\in \\mathbb{R}^{rH \\times rW}\\) at time \\(t_0\\), we take a set of consecutive spike frames \\(S(t_0 + i), i = 0, \\pm 1, \\pm 2, \\ldots\\) around \\(S(t_0)\\) as input, where \\(S(t_0 + i) \\in \\mathbb{B}^{H \\times W \\times 1}\\), \\(r\\) is SR scale, \\(\\mathbb{B}\\) represents the binary domain." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.61, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Firstly, the input spike stream is partitioned into \\(K\\) overlapping short-term spike blocks \\(\\{B_i\\}_{i = -n}^n (K = 2n + 1)\\), where \\(B_{i}\\in \\mathbb{B}^{H\\times W\\times L}\\), \\(L\\) is the temporal length of each block, \\(B_{0}\\) is the key block centered at time \\(t_0\\), \\(\\{B_i\\}, i = \\pm 1,\\pm 2,\\ldots\\) are neighboring blocks distributed symmetrically on both sides of \\(B_{0}\\), \\(B_{i}\\) is centered at time \\(t_i\\). To bridge the domain gap between binary spikes and real pixel values, we first convert spike frames in \\(B_{i} = \\{S(u)\\}_{u = t_{i} - w_{h}}^{t_{i} + w_{h}}\\) to intensity frames in \\(Y_{i} = \\{Y(u)\\}_{u = t_{i} - w_{h}}^{t_{i} + w_{h}}\\in \\mathbb{R}^{H\\times W\\times L}\\) based on the spike interval based reconstruction (SIR) method [53], where \\(L = 2w_{h} + 1\\). Then, we extract features \\(\\{F_i\\}_{i = -n}^n\\) from \\(\\{Y_i\\}_{i = -n}^n\\) through an encoder, which consists of two convolutional layers and three residual blocks. To reduce fluctuations, \\(\\{F_i\\}_{i = -n}^n\\) are input to motion aligned collaborative filtering (MACF) module, which generates improved features \\(\\{F_i\\}_{i = -n}^n\\). Finally, the multi-moment features \\(\\{\\hat{F}_i\\}_{i = -n}^n\\) are fed into motion aligned super resolved reconstruction module to generate an HR image \\(\\mathcal{I}(t_0)\\)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11419" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.088, + 0.905, + 0.302 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.227, + 0.313, + 0.768, + 0.327 + ], + "angle": 0, + "content": "Figure 5. The overall framework of motion aligned collaborative filtering (MACF) module." + }, + { + "type": "image", + "bbox": [ + 0.092, + 0.333, + 0.483, + 0.43 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.44, + 0.483, + 0.51 + ], + "angle": 0, + "content": "Figure 6. Illustration of channel rearrangement (CR) and inverse channel rearrangement (ICR). The left part is \\( K \\) features \\( \\{F_i\\}_{i = -n}^n \\) at \\( K \\) different moments, where \\( F_{i}\\in \\mathbb{R}^{H\\times W\\times C} \\). The right part is \\( C \\) features \\( \\{G_j\\}_{j = 0}^{C - 1} \\), where \\( G_{j}\\in \\mathbb{R}^{H\\times W\\times K} \\), each \\( G_{j} \\) contains features at \\( K \\) different moments." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.517, + 0.429, + 0.533 + ], + "angle": 0, + "content": "5.2. Motion Aligned Collaborative Filtering" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.539, + 0.483, + 0.613 + ], + "angle": 0, + "content": "We propose a motion aligned collaborative filtering (MACF) module to implement the idea proposed in Sec. 4. The overall framework of MACF is shown in Fig. 5, which includes six steps. We provide a detailed description of each step below." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.614, + 0.483, + 0.736 + ], + "angle": 0, + "content": "Step 1: Offset Estimation. This step generates the forward \\((F_{i}\\rightarrow F_{0})\\) and backward \\((F_0\\to F_i)\\) motion offsets between neighboring features \\(\\{F_i\\} ,i = \\pm 1,\\dots ,\\pm n\\) and key feature \\(F_{0}\\). These offsets will be utilized in the subsequent feature warping (step 3) and inverse feature warping (step 5), respectively. Specifically, the forward and backward motion offsets: \\(\\mathcal{O}_i^{\\mathbf{F}}\\) and \\(\\mathcal{O}_i^{\\mathbf{B}}\\) are both derived from \\(F_{i}\\) and \\(F_{0}\\), which are formulated as follows (\\(\\forall i\\in \\{\\pm 1,\\dots ,\\pm n\\}\\)):" + }, + { + "type": "equation", + "bbox": [ + 0.214, + 0.744, + 0.482, + 0.784 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {O} _ {i} ^ {\\mathbf {F}} = m _ {1} \\left(\\left\\{F _ {i}, F _ {0} \\right\\}\\right), \\\\ \\mathcal {O} _ {i} ^ {\\mathbf {B}} = m _ {2} \\left(\\left\\{F _ {i}, F _ {0} \\right\\}\\right), \\tag {8} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.793, + 0.483, + 0.838 + ], + "angle": 0, + "content": "where \\(m_{1}(\\cdot)\\) and \\(m_{2}(\\cdot)\\) denote two independent convolutional blocks that consist of several convolutional layers, \\(\\{,\\}\\) denotes channel concatenation operation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.838, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Step 2: Channel Rearrange. The \\(\\{F_i\\}_{i = -n}^n\\) at \\(K\\) different moments are features extracted from the fluctuated reconstruction \\(\\{Y_{i}\\}_{i = -n}^{n}\\), where \\(F_{i}\\in \\mathbb{R}^{H\\times W\\times C}\\), \\(F_{i} = \\{F_{i}^{0},F_{i}^{1},\\dots,F_{i}^{C - 1}\\}\\), the superscript of \\(F_{i}\\) denotes channel" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.335, + 0.907, + 0.662 + ], + "angle": 0, + "content": "index. Each feature \\(F_{i}\\) consists of \\(C\\) channels, with each channel encoding different information about the scene. For example, some channels capture edges, some channels focus on textures, and so on. We perform channel rearrangement (CR) to collect data at the same channel index for each feature and concatenate them, as shown in Fig. 6. We form \\(\\{G_j\\}_{j = 0}^{C - 1}\\) from \\(\\{F_i\\}_{i = -n}^n\\) via CR operation, where \\(G_{j}\\in \\mathbb{R}^{H\\times W\\times K}\\), \\(G_{j} = \\{G_{j}^{-n},\\dots,G_{j}^{0},\\dots,G_{j}^{n}\\} = \\{F_{-n}^{j},\\dots,F_{0}^{j},\\dots,F_{n}^{j}\\}\\), the subscript of \\(G\\) denotes channel index, the superscript of \\(G\\) denotes time index. \\(G_{j}\\) contains \\(K\\) features with channel index \\(j\\) at \\(K\\) different moments. To offer an intuitive understanding of this operation, we provide a visualization of CR in supplementary material. After CR, steps 3, 4 and 5 are performed sequentially for each \\(G_{j}\\). Step 3: Feature Warping. Step 2 collects temporally unaligned features that encode the same information. We perform feature warping based on deformable convolution \\(\\mathcal{D}\\) to align neighboring features \\(\\{G_j^i\\}\\), \\(i = \\pm 1,\\ldots ,\\pm n\\) with key feature \\(G_{j}^{0}\\). Specifically, given the forward offsets from neighboring features to key feature \\(\\{\\mathcal{O}_i^{\\mathbf{F}}\\}\\), \\(i = \\pm 1,\\ldots ,\\pm n\\), the aligned features could be obtained by:" + }, + { + "type": "equation", + "bbox": [ + 0.522, + 0.668, + 0.907, + 0.689 + ], + "angle": 0, + "content": "\\[\n\\tilde {G} _ {j} ^ {i} = \\mathcal {D} \\left(G _ {j} ^ {i}, \\mathcal {O} _ {i} ^ {\\mathbf {F}}\\right) = \\mathcal {D} \\left(F _ {i} ^ {j}, \\mathcal {O} _ {i} ^ {\\mathbf {F}}\\right), \\forall i \\in \\{\\pm 1, \\dots , \\pm n \\}. \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.697, + 0.906, + 0.746 + ], + "angle": 0, + "content": "These aligned features \\(\\{\\tilde{G}_j^i\\}_{i = -n}^n\\) are then concatenated to derive \\(\\tilde{G}_j = \\{\\tilde{G}_j^{-n},\\dots,\\tilde{G}_j^n\\}\\), where \\(\\tilde{G}_j^0 = G_j^0 = F_0^j\\). \\(\\tilde{G}_j\\) could be regarded as stacked similar features." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.746, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Step 4: Transform Domain Stacked Feature Filtering. We design a Transform Domain stacked feature Filtering (TDF) step to generate refinement features \\(\\Delta \\tilde{G}_j\\) from stacked similar features \\(\\tilde{G}_j\\) by leveraging representation sparsity in transform domain. TDF is designed with four submodules: transform, noise estimation, filtering, and inverse transform, as depicted in step 4 of Fig. 5. Specifically, the input \\(\\tilde{G}_j\\) is first transformed into feature \\(\\dot{G}_j\\). Simultaneously, noise feature \\(N_j\\) are extracted from \\(\\tilde{G}_j\\) via noise estimation submodule. \\(\\dot{G}_j\\) and \\(N_j\\) are passed through filter" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11420" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.089, + 0.483, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.229, + 0.483, + 0.271 + ], + "angle": 0, + "content": "Figure 7. Illustration of submodules in step 4 of MACF. (a) Transform and inverse transform submodules; (b) noise estimation submodule; (c) filtering submodule." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.279, + 0.483, + 0.342 + ], + "angle": 0, + "content": "ing submodule to generate \\(\\ddot{G}_j\\). \\(\\ddot{G}_j\\) are subsequently fed into inverse transform submodule. Additionally, a residual connection is applied in TDF step, producing refinement features \\(\\Delta \\tilde{G}_j\\) at last." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.342, + 0.484, + 0.524 + ], + "angle": 0, + "content": "The above four submodules are illustrated in detail in Fig. 7. Transform and inverse transform submodules are independently realized via several convolutional layers (Fig. 7 (a)), involving both spatial and inter-channel transform. For noise estimation (NE) (Fig. 7 (b)), we first average \\(\\tilde{G}_j\\) along the channel dimension to obtain a rough temporal average result \\(\\mathcal{E}\\). Then we calculate the residual between \\(\\tilde{G}_j\\) and \\(\\mathcal{E}\\) to represent fluctuations in features. Lastly, we extract deeper features from the residual to derive noise feature \\(N_j\\). We concatenate the transformed feature \\(\\dot{G}_j\\) and noise feature \\(N_j\\) as input to the filtering submodule (Fig. 7 (c)), which contains a U-Net structure [33] and outputs \\(\\ddot{G}_j\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.524, + 0.484, + 0.615 + ], + "angle": 0, + "content": "Step 5: Inverse Feature Warping and Aggregation. We propose an inverse motion alignment operation to map these multi-moment refinement features \\(\\Delta \\tilde{G}_j\\) to their original positions. Specifically, given the backward offsets from key feature to neighboring features \\(\\{\\mathcal{O}_i^{\\mathbf{B}}\\}\\), \\(i = \\pm 1,\\dots,\\pm n\\), we achieve inverse feature warping by:" + }, + { + "type": "equation", + "bbox": [ + 0.131, + 0.625, + 0.483, + 0.647 + ], + "angle": 0, + "content": "\\[\n\\Delta G _ {j} ^ {i} = \\mathcal {D} \\left(\\Delta \\tilde {G} _ {j} ^ {i}, \\mathcal {O} _ {i} ^ {\\mathbf {B}}\\right), \\forall i \\in \\{\\pm 1, \\dots , \\pm n \\}. \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.657, + 0.484, + 0.752 + ], + "angle": 0, + "content": "\\(\\Delta G_{j}^{i}\\) is repositioned refinement feature and corresponds to initial unaligned feature \\(G_{j}^{i}\\). We concatenate these repositioned refinement features and obtain \\(\\Delta G_{j} = \\{\\Delta G_{j}^{-n}, \\dots, \\Delta G_{j}^{n}\\}\\), where \\(\\Delta G_{j}^{0} = \\Delta \\tilde{G}_{j}^{0}\\). \\(\\Delta G_{j}\\) are then aggregated with initial features \\(\\tilde{G}_{j}\\) to derive refined features with improved consistency:" + }, + { + "type": "equation", + "bbox": [ + 0.225, + 0.761, + 0.483, + 0.782 + ], + "angle": 0, + "content": "\\[\n\\acute {G} _ {j} = G _ {j} + \\Delta G _ {j}. \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.793, + 0.483, + 0.852 + ], + "angle": 0, + "content": "Step 6: Inverse Channel Rearrange. Finally, the features \\(\\{\\hat{G}_j\\}_{j=0}^{C-1}\\) are restored to \\(\\{\\hat{F}_i\\}_{i=-n}^n\\) via inverse channel rearrangement (ICR), i.e., the inverse operation of CR, as shown in Fig. 6." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.853, + 0.484, + 0.902 + ], + "angle": 0, + "content": "MACF alleviates disturbances in \\(\\{F_i\\}_{i = -n}^n\\) and generate improved features \\(\\{\\dot{F}_i\\}_{i = -n}^n\\), which are beneficial for enhancing SR reconstruction quality." + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.089, + 0.905, + 0.233 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.243, + 0.905, + 0.271 + ], + "angle": 0, + "content": "Figure 8. Illustration of motion aligned super resolved reconstruction module." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.281, + 0.905, + 0.311 + ], + "angle": 0, + "content": "5.3. Motion Aligned Super Resolved Reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.318, + 0.907, + 0.502 + ], + "angle": 0, + "content": "The improved features \\(\\{\\hat{F}_i\\}_{i = -n}^n\\) are then input to motion aligned super resolved reconstruction module to generate the high resolution image \\(\\mathcal{I}(t_0)\\), as illustrated in Fig. 8. Firstly, features at neighboring moments \\(\\{\\hat{F}_i\\}, i = \\pm 1,\\dots,\\pm n\\) are aligned with the key feature \\(\\hat{F}_0\\) by deformable convolution [7, 56], where the features \\(\\{\\hat{F}_i\\}_{i = -n}^n\\) with improved consistency help enhance the accuracy of motion alignment, benefiting the utilization of temporal information. Subsequently, key feature and aligned neighboring features are concatenated and fused through several convolutional layers. The fused feature is upsampled to the HR image by pixel shuffle operation [3]." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.516, + 0.646, + 0.533 + ], + "angle": 0, + "content": "6. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.542, + 0.731, + 0.558 + ], + "angle": 0, + "content": "6.1. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.565, + 0.907, + 0.745 + ], + "angle": 0, + "content": "We set the number of input spike frames to 101. The number of partitioned short-term blocks is 9 (\\(K = 9\\), \\(n = 4\\)). The temporal length \\(L\\) of each block \\(B_{i}\\) is 21. The channel \\(C\\) of \\(F_{i}\\) and \\(\\hat{F}_i\\) is 64. During training, we randomly crop the spike stream to a spatial size of \\(64\\times 64\\). In addition, we randomly perform horizontal and vertical flips as well as rotation for data augmentation. The batch size is set to 8. We use Adam optimizer [23] with \\(\\beta_{1} = 0.9\\) and \\(\\beta_{2} = 0.999\\). The learning rate is initialized as 2e-4 and decayed by 0.7 every 100 epochs. The model is trained with \\(\\mathcal{L}_1\\) loss for 800 epochs. All models are trained and tested on NVIDIA RTX 3090 GPU." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.758, + 0.615, + 0.772 + ], + "angle": 0, + "content": "6.2. Datasets" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Synthesized Dataset. We need a large number of \"LR spike stream and HR image\" pairs to train the model. However, it is challenging to obtain corresponding high-quality HR images, especially in high dynamic scenes. Inspired by the experiments conducted in [48, 50], we use a spike camera simulator (detailed in supplementary material) to mimic the working mechanism of spike camera and generate synthesized SCSR datasets for training and evaluation, where" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "11421" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.107, + 0.089, + 0.895, + 0.498 + ], + "angle": 0, + "content": "
ScaleMethodREDS-based DatasetAdobe240-based DatasetParams(M)Runtime(s)
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
×2TFP+DAT25.550.59540.445627.660.67820.367714.654.40
Spk2ImgNet+DAT31.770.90240.190734.920.96040.094918.564.63
WGSE+DAT31.540.89420.205534.750.95710.102218.604.50
BSF+DAT31.920.90440.186735.030.96100.093417.135.01
VidarSR31.970.89640.204035.130.95580.110510.162.09
SpikeSR-Net34.290.93300.148137.040.97040.08383.262.14
Ours34.500.93770.140337.220.97250.08163.720.81
×3TFP+DAT24.880.57540.446426.170.65770.372114.841.88
Spk2ImgNet+DAT28.900.82930.284930.500.90640.170818.752.10
WGSE+DAT28.760.82170.296130.330.90130.177918.782.02
BSF+DAT28.990.83150.280730.440.90540.169917.322.02
VidarSR29.550.83730.279132.010.91970.168511.380.85
SpikeSR-Net31.090.87980.230433.740.94390.13553.301.35
Ours31.410.88500.223134.110.94700.13273.900.37
×4TFP+DAT24.350.56170.473425.090.64070.399514.801.06
TFP+BasicVSR++23.690.50230.485024.210.56100.44157.320.08
Spk2ImgNet+DAT27.230.76780.352127.970.84590.237818.711.15
Spk2ImgNet+BasicVSR++27.280.77300.352727.770.84210.252611.230.17
WGSE+DAT27.150.76180.359727.910.84160.242718.751.08
WGSE+BasicVSR++27.160.76990.356127.720.84040.255811.080.12
BSF+DAT27.300.77010.347327.900.84510.235817.281.10
BSF+BasicVSR++27.240.78180.337427.430.83890.24819.800.18
VidarSR28.300.79600.324930.180.88340.209712.790.75
SpikeSR-Net29.220.82660.294631.310.90730.18853.341.07
Ours29.500.83690.281631.490.91190.18223.860.21
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.508, + 0.908, + 0.552 + ], + "angle": 0, + "content": "Table 1. Quantitative results on REDS-based and Adobe240-based SCSR evaluation dataset. Red and blue indicate the best and the second-best performance, respectively. It is noted that BasicVSR++ [4] only support \\(\\times 4\\) SR scale. The runtime is tested on input spike stream of sizes \\(360 \\times 640 \\times 101 (\\times 2)\\), \\(240 \\times 426 \\times 101 (\\times 3)\\) and \\(180 \\times 320 \\times 101 (\\times 4)\\), respectively." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.557, + 0.481, + 0.739 + ], + "angle": 0, + "content": "
CaseSIRMACFPSNR↑SSIM↑
step 3step 4step 5
(A)29.070.8207
(B)✓29.150.8237
(C)✓29.170.8244
(D)✓29.320.8306
(E)✓✓29.340.8316
(F)✓✓29.350.8322
(G)✓✓29.330.8304
(H)✓✓✓29.460.8344
(I)✓✓✓✓29.500.8369
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.748, + 0.483, + 0.776 + ], + "angle": 0, + "content": "Table 2. Ablation study of the proposed Spk2SRImgNet on REDS-based SCSR evaluation dataset \\((\\times 4)\\)" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.78, + 0.425, + 0.795 + ], + "angle": 0, + "content": "each sample consists of a spike stream-image pair." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Specifically, we use the 120 fps REDS dataset [30] with a spatial resolution of \\(720 \\times 1280\\) to generate the synthesized dataset. REDS includes 240 scenes for training and 30 scenes for evaluation, and we follow its original partition. For each scene, we choose 5 segments, resulting in \\(240 \\times 5 = 1200\\) and \\(30 \\times 5 = 150\\) pairs for training and evaluation, respectively. Each pair consists of spike" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.559, + 0.907, + 0.725 + ], + "angle": 0, + "content": "stream of size \\(360 \\times 640 \\times 101 \\times 2) / 240 \\times 426 \\times 101 \\times 3) / 180 \\times 320 \\times 101 \\times 4\\) (101 is the temporal length) and a ground truth image of size \\(720 \\times 1280 \\times 2\\) and \\(\\times 4\\) / \\(720 \\times 1278 \\times 3\\). To demonstrate the generalization of the proposed model, we generate another SCSR evaluation dataset from Adobe240 dataset [34]. Adobe240 dataset includes 133 videos of 240 fps with resolution of \\(720 \\times 1280\\) taken by hand-held cameras. We follow the test set partition of Adobe240 in [44], which includes 17 scenes. For each scene, we choose 10 segments, generating \\(17 \\times 10 = 170\\) pairs for evaluation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.726, + 0.909, + 0.803 + ], + "angle": 0, + "content": "Real-World Captured Data. In addition to the synthesized data, we also evaluate the model performance on real-world spike data captured by the spike camera with a spatial resolution of \\(250 \\times 400\\). Due to the absence of ground truth, we perform qualitative analysis for these data." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.817, + 0.824, + 0.833 + ], + "angle": 0, + "content": "6.3. Comparison with Existing Methods" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.909, + 0.903 + ], + "angle": 0, + "content": "We compare our proposed method with existing SCSR method: VidarSR [42] and SpikeSR-Net [50]. Besides, we combine four spike camera image reconstruction methods: TFP [53], Spk2ImgNet [48], WGSE [45] and BSF [52] with" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "11422" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.09, + 0.229, + 0.184 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.147, + 0.184, + 0.176, + 0.194 + ], + "angle": 0, + "content": "Spike" + }, + { + "type": "image", + "bbox": [ + 0.229, + 0.091, + 0.363, + 0.184 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.271, + 0.184, + 0.321, + 0.194 + ], + "angle": 0, + "content": "TFP+DAT" + }, + { + "type": "image", + "bbox": [ + 0.364, + 0.091, + 0.498, + 0.184 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.387, + 0.184, + 0.476, + 0.193 + ], + "angle": 0, + "content": "TFP+BasicVSR++" + }, + { + "type": "image", + "bbox": [ + 0.498, + 0.091, + 0.632, + 0.184 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.521, + 0.184, + 0.61, + 0.194 + ], + "angle": 0, + "content": "Spk2ImgNet+DAT" + }, + { + "type": "image", + "bbox": [ + 0.632, + 0.091, + 0.767, + 0.184 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.637, + 0.184, + 0.765, + 0.194 + ], + "angle": 0, + "content": "Spk2ImgNet+BasicVSR++" + }, + { + "type": "image", + "bbox": [ + 0.768, + 0.091, + 0.903, + 0.184 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.805, + 0.184, + 0.866, + 0.194 + ], + "angle": 0, + "content": "WGSE+DAT" + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.195, + 0.228, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.11, + 0.289, + 0.212, + 0.299 + ], + "angle": 0, + "content": "WGSE+BasicVSR++" + }, + { + "type": "image", + "bbox": [ + 0.229, + 0.195, + 0.363, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.271, + 0.289, + 0.322, + 0.299 + ], + "angle": 0, + "content": "BSF+DAT" + }, + { + "type": "image", + "bbox": [ + 0.364, + 0.195, + 0.498, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.386, + 0.289, + 0.476, + 0.299 + ], + "angle": 0, + "content": "BSF+BasicVSR++" + }, + { + "type": "image", + "bbox": [ + 0.498, + 0.195, + 0.632, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.545, + 0.289, + 0.588, + 0.299 + ], + "angle": 0, + "content": "VidarSR" + }, + { + "type": "image", + "bbox": [ + 0.632, + 0.195, + 0.767, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.669, + 0.289, + 0.733, + 0.299 + ], + "angle": 0, + "content": "SpikeSR-Net" + }, + { + "type": "image", + "bbox": [ + 0.768, + 0.195, + 0.903, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.823, + 0.289, + 0.848, + 0.299 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.3, + 0.228, + 0.393 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.147, + 0.394, + 0.176, + 0.403 + ], + "angle": 0, + "content": "Spike" + }, + { + "type": "image", + "bbox": [ + 0.229, + 0.3, + 0.363, + 0.393 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.271, + 0.394, + 0.322, + 0.403 + ], + "angle": 0, + "content": "TFP+DAT" + }, + { + "type": "image", + "bbox": [ + 0.364, + 0.3, + 0.498, + 0.393 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.386, + 0.394, + 0.476, + 0.403 + ], + "angle": 0, + "content": "TFP+BasicVSR++" + }, + { + "type": "image", + "bbox": [ + 0.498, + 0.3, + 0.632, + 0.393 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.522, + 0.394, + 0.61, + 0.404 + ], + "angle": 0, + "content": "Spk2ImgNet+DAT" + }, + { + "type": "image", + "bbox": [ + 0.632, + 0.3, + 0.767, + 0.393 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.637, + 0.394, + 0.765, + 0.404 + ], + "angle": 0, + "content": "Spk2ImgNet+BasicVSR++" + }, + { + "type": "image", + "bbox": [ + 0.768, + 0.3, + 0.903, + 0.393 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.804, + 0.394, + 0.867, + 0.403 + ], + "angle": 0, + "content": "WGSE+DAT" + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.404, + 0.228, + 0.497 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.11, + 0.498, + 0.213, + 0.508 + ], + "angle": 0, + "content": "WGSE+BasicVSR++" + }, + { + "type": "image", + "bbox": [ + 0.229, + 0.404, + 0.363, + 0.497 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.271, + 0.498, + 0.322, + 0.507 + ], + "angle": 0, + "content": "BSF+DAT" + }, + { + "type": "image", + "bbox": [ + 0.364, + 0.404, + 0.498, + 0.497 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.386, + 0.498, + 0.476, + 0.508 + ], + "angle": 0, + "content": "BSF+BasicVSR++" + }, + { + "type": "image", + "bbox": [ + 0.498, + 0.404, + 0.632, + 0.497 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.545, + 0.498, + 0.588, + 0.508 + ], + "angle": 0, + "content": "VidarSR" + }, + { + "type": "image", + "bbox": [ + 0.632, + 0.404, + 0.767, + 0.497 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.669, + 0.498, + 0.733, + 0.508 + ], + "angle": 0, + "content": "SpikeSR-Net" + }, + { + "type": "image", + "bbox": [ + 0.768, + 0.404, + 0.903, + 0.497 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.823, + 0.498, + 0.848, + 0.508 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.521, + 0.904, + 0.549 + ], + "angle": 0, + "content": "Figure 9. Visual comparison \\((\\times 4)\\) on real-world spike data. The first scene is a stationary clock recorded by a fast-moving spike camera. The second scene is a rotating fan. Please enlarge the figure for better comparison." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.557, + 0.484, + 0.739 + ], + "angle": 0, + "content": "image / video SR method: DAT [5] / BasicVSR++ [4], resulting in another eight comparison methods, as presented in Tab. 1. We conduct experiments on \\(\\times 2\\), \\(\\times 3\\), and \\(\\times 4\\) SR scales. We employ PSNR, SSIM [39] and LPIPS [46] as quantitative metrics. All compared methods except TFP [53] are retrained on the REDS-based SCSR training dataset with the same settings as ours. As shown in Tab. 1, our method achieves the best performance on the three metrics. Besides, we conduct experiments on real-world captured spike stream, as shown in Fig. 9, our method can restore better textures and details. More evaluation and visual results are provided in supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.756, + 0.244, + 0.772 + ], + "angle": 0, + "content": "6.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.484, + 0.901 + ], + "angle": 0, + "content": "To verify the effectiveness of our methodology, we perform ablation study with REDS-based evaluation dataset on \\(\\times 4\\) SR scale, as presented in Tab. 2. Since ablation study only involves steps 3, 4, and 5 in MACF, the remaining steps are omitted in Tab. 2 (Steps 2 and 6 do not involve network parameters, and step 1 is present only when step 3 and/or step 5 are present). Case (I) is the final model. Case (A) is formed by removing SIR and MACF modules from" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.557, + 0.907, + 0.68 + ], + "angle": 0, + "content": "Spk2SRImgNet. Comparisons between (A) and (B), (H) and (I) illustrate the advantages of converting information from spike domain to image domain (SIR module). Comparisons between (A) and (H), (B) and (I) demonstrate the effectiveness of MACF module, showing gains of 0.39dB and 0.35dB, respectively. Comparisons between (C-G) and (H) validate the effectiveness of each step in MACF. More ablation study is provided in supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.694, + 0.634, + 0.71 + ], + "angle": 0, + "content": "7. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.907, + 0.901 + ], + "angle": 0, + "content": "This paper develops an end-to-end neural network named Spk2SRImgNet to reconstruct high resolution images from low resolution spike stream. To handle spike fluctuations, we propose a motion aligned collaborative filtering (MACF) module motivated by key ideas in classic image restoration schemes. Specifically, MACF leverages the temporal similarity of spike stream to enhance the consistency of multi-moment features along the motion trajectory, mitigating fluctuations and improving the quality of super resolution reconstruction. Experimental results demonstrate that the proposed Spk2SRImgNet achieves state-of-the-art performance on synthesized and real-captured spike data." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11423" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.259, + 0.108 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.115, + 0.486, + 0.161 + ], + "angle": 0, + "content": "This work was supported by the National Natural Science Foundation of China under Grants 62072009 and U22B2035." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.173, + 0.188, + 0.189 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.199, + 0.484, + 0.255 + ], + "angle": 0, + "content": "[1] Patrick Bardow, Andrew J. Davison, and Stefan Leutenegger. Simultaneous optical flow and intensity estimation from an event camera. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 884-892, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.256, + 0.484, + 0.311 + ], + "angle": 0, + "content": "[2] Christian Brandli, Raphael Berner, Minhao Yang, Shih-Chii Liu, and Tobi Delbruck. A \\(240 \\times 180\\) 130 db \\(3\\mu s\\) latency global shutter spatiotemporal vision sensor. IEEE Journal of Solid-State Circuits, 49(10):2333-2341, 2014. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.312, + 0.484, + 0.395 + ], + "angle": 0, + "content": "[3] Jose Caballero, Christian Ledig, Andrew Aitken, Alejandro Acosta, Johannes Totz, Zehan Wang, and Wenzhe Shi. Real-time video super-resolution with spatio-temporal networks and motion compensation. In 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2848-2857, 2017. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.396, + 0.484, + 0.478 + ], + "angle": 0, + "content": "[4] Kelvin C.K. Chan, Shangchen Zhou, Xiangyu Xu, and Chen Change Loy. Basicvsr++: Improving video superresolution with enhanced propagation and alignment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5972-5981, 2022. 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.48, + 0.484, + 0.548 + ], + "angle": 0, + "content": "[5] Zheng Chen, Yulun Zhang, Jinjin Gu, Linghe Kong, Xiaokang Yang, and Fisher Yu. Dual aggregation transformer for image super-resolution. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 12312-12321, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.55, + 0.484, + 0.605 + ], + "angle": 0, + "content": "[6] Kostadin Dabov, Alessandro Foi, Vladimir Katkovnik, and Karen Egiazarian. Image denoising by sparse 3-d transform-domain collaborative filtering. IEEE Transactions on Image Processing, 16(8):2080-2095, 2007. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.607, + 0.484, + 0.662 + ], + "angle": 0, + "content": "[7] Jifeng Dai, Haozhi Qi, Yuwen Xiong, Yi Li, Guodong Zhang, Han Hu, and Yichen Wei. Deformable convolutional networks. In 2017 IEEE International Conference on Computer Vision (ICCV), pages 764-773, 2017. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.663, + 0.484, + 0.704 + ], + "angle": 0, + "content": "[8] Siwei Dong, Tiejun Huang, and Yonghong Tian. Spike camera and its coding methods. In 2017 Data Compression Conference (DCC), pages 437-437, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.705, + 0.484, + 0.761 + ], + "angle": 0, + "content": "[9] Siwei Dong, Lin Zhu, Daoyuan Xu, Yonghong Tian, and Tiejun Huang. An efficient coding method for spike camera using inter-spike intervals. In 2019 Data Compression Conference (DCC), pages 568-568, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.762, + 0.484, + 0.817 + ], + "angle": 0, + "content": "[10] Yanchen Dong, Jing Zhao, Ruiqin Xiong, and Tiejun Huang. 3D residual interpolation for spike camera demosaicing. In IEEE International Conference on Image Processing (ICIP), pages 1461-1465. IEEE, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.818, + 0.484, + 0.9 + ], + "angle": 0, + "content": "[11] Yanchen Dong, Ruiqin Xiong, Jian Zhang, Zhaofei Yu, Xiaopeng Fan, Shuyuan Zhu, and Tiejun Huang. Superresolution reconstruction from bayer-pattern spike streams. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 24871-24880, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.199, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.162 + ], + "angle": 0, + "content": "[12] Yanchen Dong, Ruiqin Xiong, Jing Zhao, Jian Zhang, Xiaopeng Fan, Shuyuan Zhu, and Tiejun Huang. Joint demosaicing and denoising for spike camera. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1582-1590, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.163, + 0.907, + 0.233 + ], + "angle": 0, + "content": "[13] Yanchen Dong, Ruiqin Xiong, Jing Zhao, Jian Zhang, Xiaopeng Fan, Shuyuan Zhu, and Tiejun Huang. Learning a deep demosaicing network for spike camera with color filter array. IEEE Transactions on Image Processing, 33:3634-3647, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.234, + 0.906, + 0.29 + ], + "angle": 0, + "content": "[14] Yanchen Dong, Ruiqin Xiong, Xiaopeng Fan, Shuyuan Zhu, Jin Wang, and Tiejun Huang. Dynamic scene reconstruction for color spike camera via zero-shot learning. IEEE Transactions on Computational Imaging, 11:129-141, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.291, + 0.906, + 0.332 + ], + "angle": 0, + "content": "[15] Daniel Gehrig and Davide Scaramuzza. Low-latency automotive vision with event cameras. Nature, 629(8014):1034-1040, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.334, + 0.906, + 0.403 + ], + "angle": 0, + "content": "[16] Arren Glover, Valentina Vasco, and Chiara Bartolozzi. A controlled-delay event camera framework for on-line robotics. In 2018 IEEE International Conference on Robotics and Automation (ICRA), pages 2178-2183. IEEE, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.404, + 0.906, + 0.46 + ], + "angle": 0, + "content": "[17] Sorin Grigorescu, Bogdan Trasnea, Tiberiu Cocias, and Gigel Macesanu. A survey of deep learning techniques for autonomous driving. Journal of field robotics, 37(3):362-386, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.461, + 0.906, + 0.531 + ], + "angle": 0, + "content": "[18] Shuhang Gu, Lei Zhang, Wangmeng Zuo, and Xiangchu Feng. Weighted nuclear norm minimization with application to image denoising. In 2014 IEEE Conference on Computer Vision and Pattern Recognition, pages 2862-2869, 2014. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.532, + 0.906, + 0.602 + ], + "angle": 0, + "content": "[19] Tiejun Huang, Yajing Zheng, Zhaofei Yu, Rui Chen, Yuan Li, Ruiqin Xiong, Lei Ma, Junwei Zhao, Siwei Dong, Lin Zhu, Jianing Li, Shanshan Jia, Yihua Fu, Boxin Shi, Si Wu, and Yonghong Tian. \\(1000\\times\\) faster camera and machine vision with ordinary devices. Engineering, 25:110-119, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.602, + 0.906, + 0.659 + ], + "angle": 0, + "content": "[20] Arpit Jadon, Mohd Omama, Akshay Varshney, Mohammad Samar Ansari, and Rishabh Sharma. Firenet: A specialized lightweight fire & smoke detection model for real-time IoT applications. arXiv preprint arXiv:1905.11922, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.66, + 0.906, + 0.716 + ], + "angle": 0, + "content": "[21] Armin Kappeler, Seunghwan Yoo, Qiqin Dai, and Aggelos K. Katsaggelos. Video super-resolution with convolutional neural networks. IEEE Transactions on Computational Imaging, 2(2):109-122, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.717, + 0.906, + 0.772 + ], + "angle": 0, + "content": "[22] Hanme Kim, Ankur Handa, Ryad Benosman, Sio-Hoi Ieng, and Andrew Davison. Simultaneous mosaicing and tracking with an event camera. In Proceedings of the British Machine Vision Conference. BMVA Press, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.773, + 0.906, + 0.815 + ], + "angle": 0, + "content": "[23] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.816, + 0.906, + 0.872 + ], + "angle": 0, + "content": "[24] Jingyun Liang, Jiezhang Cao, Yuchen Fan, Kai Zhang, Rakesh Ranjan, Yawei Li, Radu Timofte, and Luc Van Gool. Vrt: A video restoration transformer. IEEE Transactions on Image Processing, 33:2171-2182, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.906, + 0.903 + ], + "angle": 0, + "content": "[25] Patrick Lichtsteiner, Christoph Posch, and Tobi Delbruck. A \\(128 \\times 128\\) 120 db \\(15\\mu s\\) latency asynchronous temporal con" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11424" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.126, + 0.093, + 0.482, + 0.12 + ], + "angle": 0, + "content": "trast vision sensor. IEEE journal of solid-state circuits, 43 (2):566-576, 2008. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.219 + ], + "angle": 0, + "content": "[26] Martin Litzenberger, Christoph Posch, D Bauer, Ahmed Nabil Belbachir, P Schon, B Kohn, and H Garn. Embedded vision system for real-time object tracking using an asynchronous transient vision sensor. In 2006 IEEE 12th Digital Signal Processing Workshop & 4th IEEE Signal Processing Education Workshop, pages 173-178. IEEE, 2006. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.221, + 0.483, + 0.276 + ], + "angle": 0, + "content": "[27] Julien Mairal, Francis Bach, Jean Ponce, Guillermo Sapiro, and Andrew Zisserman. Non-local sparse models for image restoration. In 2009 IEEE 12th International Conference on Computer Vision, pages 2272-2279, 2009. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.278, + 0.482, + 0.332 + ], + "angle": 0, + "content": "[28] Mohammad Mostafavi, Yeongwoo Nam, Jonghyun Choi, and Kuk-Jin Yoon. E2sri: Learning to super-resolve intensity images from events. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(10):6890-6909, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.334, + 0.482, + 0.362 + ], + "angle": 0, + "content": "[29] Elias Mueggler. Event-based vision for high-speed robotics. PhD thesis, University of Zurich, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.364, + 0.482, + 0.446 + ], + "angle": 0, + "content": "[30] Seungjun Nah, Sungyong Baik, Seokil Hong, Gyeongsik Moon, Sanghyun Son, Radu Timofte, and Kyoung Mu Lee. Ntire 2019 challenge on video deblurring and superresolution: Dataset and study. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.449, + 0.482, + 0.516 + ], + "angle": 0, + "content": "[31] Christoph Posch, Daniel Matolin, and Rainer Wohlgenannt. A qvga 143 db dynamic range frame-free pwm image sensor with lossless pixel-level video compression and time-domain cds. IEEE Journal of Solid-State Circuits, 46(1):259-275, 2010. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.519, + 0.482, + 0.588 + ], + "angle": 0, + "content": "[32] Henri Rebecq, René Ranftl, Vladlen Koltun, and Davide Scaramuzza. Events-to-video: Bringing modern computer vision to event cameras. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3857-3866, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.59, + 0.482, + 0.673 + ], + "angle": 0, + "content": "[33] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention-MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, pages 234-241. Springer, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.675, + 0.482, + 0.743 + ], + "angle": 0, + "content": "[34] Shuochen Su, Maurizio Delbrazio, Jue Wang, Guillermo Sapiro, Wolfgang Heidrich, and Oliver Wang. Deep video deblurring for hand-held cameras. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 1279-1288, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.746, + 0.482, + 0.801 + ], + "angle": 0, + "content": "[35] Yapeng Tian, Yulun Zhang, Yun Fu, and Chenliang Xu. Tdan: Temporally-deformable alignment network for video super-resolution. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.803, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[36] Lin Wang, Tae-Kyun Kim, and Kuk-Jin Yoon. Eventsr: From asynchronous events to image reconstruction, restoration, and super-resolution via end-to-end adversarial learning. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8312-8322, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[37] Xintao Wang, Kelvin C.K. Chan, Ke Yu, Chao Dong, and Chen Change Loy. Edvr: Video restoration with enhanced" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "deformable convolutional networks. In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 1954-1963, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.137, + 0.905, + 0.205 + ], + "angle": 0, + "content": "[38] Yuanlin Wang, Ruiqin Xiong, Jing Zhao, and Tiejun Huang. Reconstruct dynamic scene for spike camera based on 3d space time similarity. In 2024 IEEE International Conference on Image Processing (ICIP), pages 1595-1601, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.207, + 0.905, + 0.261 + ], + "angle": 0, + "content": "[39] Zhou Wang, A.C. Bovik, H.R. Sheikh, and E.P. Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing, 13(4): 600-612, 2004. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.264, + 0.905, + 0.332 + ], + "angle": 0, + "content": "[40] Philippe Weinzaepfel, Jerome Revaud, Zaid Harchaoui, and Cordelia Schmid. Deepflow: Large displacement optical flow with deep matching. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), pages 1385-1392, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.334, + 0.905, + 0.39 + ], + "angle": 0, + "content": "[41] Wenming Weng, Yueyi Zhang, and Zhiwei Xiong. Event-based video reconstruction using transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.392, + 0.905, + 0.461 + ], + "angle": 0, + "content": "[42] Xijie Xiang, Lin Zhu, Jianing Li, Yixuan Wang, Tiejun Huang, and Yonghong Tian. Learning super-resolution reconstruction for high temporal resolution spike stream. IEEE Transactions on Circuits and Systems for Video Technology, 33(1):16-29, 2023. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.463, + 0.905, + 0.531 + ], + "angle": 0, + "content": "[43] Ruiqin Xiong, Hangfan Liu, Xinfeng Zhang, Jian Zhang, Siwei Ma, Feng Wu, and Wen Gao. Image denoising via bandwise adaptive modeling and regularization exploiting nonlocal similarity. IEEE Transactions on Image Processing, 25 (12):5793-5805, 2016. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.533, + 0.905, + 0.603 + ], + "angle": 0, + "content": "[44] Gang Xu, Jun Xu, Zhen Li, Liang Wang, Xing Sun, and Ming-Ming Cheng. Temporal modulation network for controllable space-time video super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6388-6397, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.605, + 0.905, + 0.673 + ], + "angle": 0, + "content": "[45] Jiyuan Zhang, Shanshan Jia, Zhaofei Yu, and Tiejun Huang. Learning temporal-ordered representation for spike streams based on discrete wavelet transforms. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 137–147, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.676, + 0.905, + 0.745 + ], + "angle": 0, + "content": "[46] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 586-595, 2018. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.747, + 0.905, + 0.815 + ], + "angle": 0, + "content": "[47] Jing Zhao, Jiyu Xie, Ruiqin Xiong, Jian Zhang, Zhaofei Yu, and Tiejun Huang. Super resolve dynamic scene from continuous spike streams. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 2533-2542, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.818, + 0.905, + 0.899 + ], + "angle": 0, + "content": "[48] Jing Zhao, Ruiqin Xiong, Hangfan Liu, Jian Zhang, and Tiejun Huang. Spk2imgnet: Learning to reconstruct dynamic scene from continuous spike stream. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11996-12005, 2021. 2, 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11425" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.093, + 0.482, + 0.16 + ], + "angle": 0, + "content": "[49] Jing Zhao, Ruiqin Xiong, Jiyu Xie, Boxin Shi, Zhaofei Yu, Wen Gao, and Tiejun Huang. Reconstructing clear image for high-speed motion scene with a retina-inspired spike camera. IEEE Transactions on Computational Imaging, 8:12-27, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.091, + 0.163, + 0.483, + 0.232 + ], + "angle": 0, + "content": "[50] Jing Zhao, Ruiqin Xiong, Jian Zhang, Rui Zhao, Hangfan Liu, and Tiejun Huang. Learning to super-resolve dynamic scenes for neuromorphic spike camera. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3579-3587, 2023. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.234, + 0.483, + 0.302 + ], + "angle": 0, + "content": "[51] Rui Zhao, Ruiqin Xiong, Jian Zhang, Zhaofei Yu, Shuyuan Zhu, Lei Ma, and Tiejun Huang. Spike camera image reconstruction using deep spiking neural networks. IEEE Transactions on Circuits and Systems for Video Technology, 34(6): 5207-5212, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.304, + 0.482, + 0.387 + ], + "angle": 0, + "content": "[52] Rui Zhao, Ruiqin Xiong, Jing Zhao, Jian Zhang, Xiaopeng Fan, Zhaofei Yu, and Tiejun Huang. Boosting spike camera image reconstruction from a perspective of dealing with spike fluctuations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 24955-24965, 2024. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.389, + 0.482, + 0.457 + ], + "angle": 0, + "content": "[53] Lin Zhu, Siwei Dong, Tiejun Huang, and Yonghong Tian. A retina-inspired sampling method for visual texture reconstruction. In 2019 IEEE International Conference on Multimedia and Expo (ICME), pages 1432-1437. IEEE, 2019. 2, 3, 4, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.459, + 0.482, + 0.528 + ], + "angle": 0, + "content": "[54] Lin Zhu, Siwei Dong, Jianing Li, Tiejun Huang, and Yonghong Tian. Ultra-high temporal resolution visual reconstruction from a fovea-like spike camera via spiking neuron model. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(1):1233-1249, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.53, + 0.482, + 0.599 + ], + "angle": 0, + "content": "[55] Lin Zhu, Xiao Wang, Yi Chang, Jianing Li, Tiejun Huang, and Yonghong Tian. Event-based video reconstruction via potential-assisted spiking neural network. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3584-3594, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.601, + 0.482, + 0.668 + ], + "angle": 0, + "content": "[56] Xizhou Zhu, Han Hu, Stephen Lin, and Jifeng Dai. Deformable convnets v2: More deformable, better results. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9308-9316, 2019. 2, 6" + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.093, + 0.483, + 0.668 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.956 + ], + "angle": 0, + "content": "11426" + } + ] +] \ No newline at end of file diff --git a/2025/Spk2SRImgNet_ Super-Resolve Dynamic Scene from Spike Stream via Motion Aligned Collaborative Filtering/e0bcc965-cf40-409b-a203-db12626ad68b_origin.pdf b/2025/Spk2SRImgNet_ Super-Resolve Dynamic Scene from Spike Stream via Motion Aligned Collaborative Filtering/e0bcc965-cf40-409b-a203-db12626ad68b_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..b0c49f5301951aac7e57f1f40512814607e5d09b --- /dev/null +++ b/2025/Spk2SRImgNet_ Super-Resolve Dynamic Scene from Spike Stream via Motion Aligned Collaborative Filtering/e0bcc965-cf40-409b-a203-db12626ad68b_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7a9e4252af8237115da6638ffd7243d14ace3f85df2fa2058aa496a212c2f3d1 +size 1822593 diff --git a/2025/Spk2SRImgNet_ Super-Resolve Dynamic Scene from Spike Stream via Motion Aligned Collaborative Filtering/full.md b/2025/Spk2SRImgNet_ Super-Resolve Dynamic Scene from Spike Stream via Motion Aligned Collaborative Filtering/full.md new file mode 100644 index 0000000000000000000000000000000000000000..03b2b96a0cca5a79d1a25e161af1677372f88036 --- /dev/null +++ b/2025/Spk2SRImgNet_ Super-Resolve Dynamic Scene from Spike Stream via Motion Aligned Collaborative Filtering/full.md @@ -0,0 +1,379 @@ +# Spk2SRImgNet: Super-Resolve Dynamic Scene from Spike Stream via Motion Aligned Collaborative Filtering + +Yuanlin Wang $^{1,2}$ , Yiyang Zhang $^{1,2}$ , Ruiqin Xiong $^{1,2,*}$ , Jing Zhao $^{1,2}$ , Jian Zhang $^{3}$ , Xiaopeng Fan $^{4}$ , Tiejun Huang $^{1,2}$ + +$^{1}$ School of Computer Science, Peking University + +$^{2}$ National Key Laboratory for Multimedia Information Processing, Peking University + +$^{3}$ School of Electronic and Computer Engineering, Shenzhen Graduate School, Peking University + +$^{4}$ School of Computer Science and Technology, Harbin Institute of Technology + +wangyuanlin@stu.pku.edu.cn + +{yiyangz, rqxiong, jzhaopku, zhangjian.sz, tjhuang}@pku.edu.cn fxp@hit.edu.cn + +# Abstract + +Spike camera is a kind of neuromorphic camera that records dynamic scenes by firing a stream of binary spikes with extremely high temporal resolution. It demonstrates great potential for vision tasks in high-speed scenarios. One limitation in its current implementation is the relatively low spatial resolution. This paper develops a network called Spk2SRImgNet to super-resolve high resolution images from low resolution spike stream. However, fluctuations in spike stream hinder the performance of spike camera super resolution. To address this issue, we propose a motion aligned collaborative filtering (MACF) module, which is motivated by key ideas in classic image restoration schemes to mitigate fluctuations in spike data. MACF leverages the temporal similarity of spike stream to acquire similar features from neighboring moments via motion alignment. To separate disturbances from features, MACF filters these similar features jointly in transform domain to exploit representation sparsity, and generates refinement features that will be used to update initial fluctuated features. Specifically, MACF designs an inverse motion alignment operation to map these refinement features back to their original positions. The initial features are aggregated with the repositioned refinement features to enhance reliability. Experimental results demonstrate that the proposed method achieves state-of-the-art performance compared with existing methods. + +# 1. Introduction + +High-speed imaging attracts increasing attention in emerging computer vision applications such as autonomous driv- + +![](images/482e27539178c625a69df88b96f02f63c321c9eddf755e8369557e0ec804f353.jpg) +Figure 1. Illustration of spike camera super resolution (SCSR) and visual comparison $(\times 4)$ with existing SCSR methods: VidarSR [42] and SpikeSR-Net [50]. Please zoom in for better comparison. + +ing [15, 17] and robotics [16, 29]. Conventional camera employs an exposure-based imaging model, which limits its ability to capture high-speed scenes. Specifically, conventional camera accumulates photoelectric signals over an exposure time window to form a snapshot image, which requires the scene to be still during the exposure period. Otherwise, a single point on a moving object may be projected onto different sensor pixels, resulting in motion blur. + +Neuromorphic camera is an emerging camera that abandons the concept of exposure window and mimics the working mechanism of biological retina. It records intensity information with high temporal resolution, showing great potential for high speed vision tasks. One kind of neuromorphic camera is event camera [2, 25, 26, 31]. It detects light intensity changes of each pixel in the logarithmic domain and fires an event when the change exceeds a specified threshold. Another kind of neuromorphic camera is spike camera [8, 9, 19]. Each pixel of spike camera independently and continuously accumulates photons, and fires a spike when the accumulation exceeds a predefined threshold. Unlike event camera that only records relative changes of light intensity, spike camera captures absolute intensity by firing a stream of binary spikes, providing clearer information for reconstructing image details. + +Benefiting from its extremely high temporal resolution + +(e.g. $40\mathrm{kHz}$ ), spike camera could sample a large amount of intensity information in a short period. Besides, these samples cover different locations of the scene due to the relative motion between camera and scene. Therefore, there is promising potential to reconstruct high resolution (HR) images from spike stream by converting the dense sampling in temporal domain to spatial domain. + +This paper aims to super-resolve high resolution images from low resolution spike stream. However, due to the Poisson effect of photon arrival, the quantization effect of spike readout, and the thermal noises introduced by circuits, spikes exhibit fluctuations, i.e., the length of spike accumulation period varies over time, even when light intensity remains unchanged. Thus, the features extracted from fluctuated spikes tend to be noisy and temporally incoherent. These unreliable features provide low-quality information source for super resolution (SR) reconstruction. Moreover, they affect the alignment accuracy when leveraging temporal information, further degrading the quality of SR reconstruction. + +In this paper, we propose a spike-to-super-resolution-image network Spk2SRImgNet. To handle spike fluctuations, we take inspiration from several classic image restoration methods, such as BM3D [6], BAS [43], LSSC [27], and WNNM [18]. These methods incorporate the idea of leveraging non-local similarity within natural images to reduce the unknown random disturbance. We embed these concepts into the design of Spk2SRImgNet and develop a motion aligned collaborative filtering (MACF) module to reduce disturbances in spike information. Firstly, MACF obtains similar features from neighboring moments via motion alignment, exploiting the temporal similarity of spike stream. To separate disturbances from features, MACF performs collaborative filtering on grouped similar features in transform domain for leveraging representation sparsity, and generates refinement features for each moment. These refinement features will be utilized to update initial fluctuated features. Specifically, MACF introduces an inverse motion alignment operation to map these multi-moment refinement features to their original unaligned positions. The initial features are integrated with the repositioned refinement features to improve reliability. Experimental results show that the proposed Spk2SRImgNet achieves state-of-the-art performance on synthesized and real-world spike data. + +# 2. Related Work + +Video Super Resolution. Video super resolution (VSR) produces high resolution images from low resolution (LR) image sequences. Temporal alignment is a crucial step to model temporal correspondences of misaligned frames. Early learning-based methods [3, 21] employ traditional optical flow estimation [40] for motion compensation. With + +the diverse development of network structures, TDAN [35] and EDVR [37] utilize deformable convolution [7, 56] for feature alignment. BasicVSR++ [4] leverages optical flow to guide deformable alignment for stable training. VRT [24] proposes a temporal mutual attention block to facilitate motion estimation and feature alignment. + +Event Camera Reconstruction and Super Resolution. Event camera monitors relative changes in light intensity with high temporal resolution. Reconstructing intensity images from events has been studied in the past decade. Kim et al. [22] propose to use the extended Kalman filter and Poisson reconstruction for event reconstruction. Bardow et al. [1] estimate both intensity and optical flow using the primal-dual algorithm. With the development of deep learning, E2VID [32], FireNet [20], ET-Net [41], EVSNN [55] have been proposed successively with good performance. To reconstruct HR images from LR events, Wang et al. [36] propose an unsupervised GAN EventSR. Mostafavi et al. [28] propose an end-to-end recurrent network E2SRI. + +Spike Camera Reconstruction and Super Resolution. Image reconstruction is a popular topic for spike camera [38, 48, 51-54]. TFI [53] reconstructs images based on the interval between two temporally neighboring spikes. TFP [53] utilizes the number of spikes within a temporal window to estimate the intensity. For better performance, CNN-based [48, 52] and spike neural network (SNN)-based [51, 54] frameworks have been proposed successively. To reconstruct HR images from LR spike stream, Zhao et al. [47] derive the constraint relationship between pixel intensities and each spike period, and obtain HR images based on gradient descent algorithm. Xiang et al. [42] design a multi-level feature learning network to achieve spike camera super resolution. Zhao et al. [50] propose SpikeSR-Net based on the observation model of spike camera. Furthermore, Dong et al. [11] propose a super resolution network for the newly developed color spike camera [10, 12-14]. + +# 3. Spike Camera + +Working Mechanism. Spike camera contains a pixel array of size $H \times W$ and each pixel works independently. Each pixel comprises three main components: photon receptor, integrator, and comparator, as shown in Fig. 2. The photoreceptor captures incident photons from external scenes and converts them into electric charges through photoelectric conversion. The integrator accumulates the electric charges continuously. Meanwhile, the comparator checks the accumulated voltage persistently. Once the voltage reaches a predefined threshold $\theta$ , the pixel sets up a spike flag and resets the integrator, immediately restarting a new "accumulation-and-set" cycle. Spike readout circuit periodically checks the spike flag at a frequency of $40\mathrm{kHz}$ . When detecting the flag, the circuit outputs a spike "1", otherwise, it outputs "0". The voltage accumulated by the in + +![](images/b1caeff65e5a9539fd7868965e875e8773b7f426097fe61f95efb0a5ce78069a.jpg) +Figure 2. Working mechanism of spike camera. + +tegrator at an arbitrary time $t$ can be formulated as [49]: + +$$ +A (x, y, t) = \int_ {0} ^ {t} \eta I (x, y, u) d u \quad \mathrm {m o d} \theta , \tag {1} +$$ + +where $I(x,y,u)$ denotes the arrival rate of photons received by pixel $(x,y)$ at time $u$ , $\eta$ denotes the photoelectric conversion rate. The output of spike camera is a 3D binary spike matrix $\{S(u)\}_{u=1}^{T}$ with size $H\times W\times T$ , where $T$ is the temporal length, $S(u)$ is a spike frame at time $u$ . + +Spike Interval Based Reconstruction. According to the working mechanism of spike camera, the stronger the light intensity, the shorter the interval between two neighboring spikes, i.e., the spike interval is inversely proportional to the light intensity. This property is exploited by spike-interval-based reconstruction method (SIR) [53]. Specifically, the spike interval of two temporally neighboring spikes that cover the position $(x,y,u)$ can be formulated as: + +$$ +\begin{array}{l} \Phi (x, y, u) = \min \left\{t \mid S (x, y, t) = 1, t \geq u \right\} \tag {2} \\ - \max \{t | S (x, y, t) = 1, t < u \}. \\ \end{array} +$$ + +SIR estimates the light intensity of pixel $(x,y)$ at time $u$ by + +$$ +Y (x, y, u) = \frac {\delta}{\Phi (x , y , u)}, \tag {3} +$$ + +where $\delta$ denotes the maximum dynamic range of reconstruction. + +Spike Fluctuations. The imaging of spike camera is affected by multiple kinds of noise. First, the Poisson process of photon arrival causes the time taken for the number of photons to reach the threshold to exhibit randomness, which is denoted as Poisson noise. Second, the spike readout circuit is controlled by a periodic clock signal, i.e., it only outputs spikes at discrete times, which results in a time difference between spike flag set and spike readout (see Fig. 2). This difference leads to quantization noise. Third, the thermal noise introduced by circuits brings randomness to spike fire. For the above reasons, the intervals of neighboring spikes exhibit fluctuations and struggle to accurately reflect light intensity, leading to noisy reconstruction results from SIR. + +# 4. Motivation and Idea + +The key problem we aim to address is how to super-resolve the original dynamic scene from captured binary spike data to obtain high quality visual reconstruction with clean image details. According to the working mechanism of spike camera, each individual spike does not reveal the light intensity directly. Instead, it only tells the time point at which a spike firing occurs. Essentially, the instantaneous light intensity can be inferred only by considering the length of intervals between consecutive spikes. Unfortunately, for the reasons discussed above, the recorded spike intervals are not stable, and they fluctuate around an expectation value determined by actual light intensity. + +To handle such fluctuations, we draw inspiration from several classic image restoration methods, such as BM3D [6], BAS [43], LSSC [27], and WNNM [18]. Although they are quite diversified in algorithmic design, all these methods incorporate the idea of leveraging non-local similarity within natural image signals to reduce the unknown random disturbance. They utilize a group of similar patches and establish a prior model for them so that a regularization term is formed to separate the true signal from the noise. + +Suppose $\mathbf{y} = \mathbf{x} + \mathbf{n}$ is an image corrupted by noise, where $\mathbf{x}$ is the clean image and $\mathbf{n}$ is the noise. The idea of these methods can be summarized as follows: + +(1) For each pixel location $\mathbf{p}$ , search for a group of similar patches from the noisy image $\mathbf{y}$ : + +$$ +\mathbf {y} _ {1}, \mathbf {y} _ {2}, \dots , \mathbf {y} _ {N} = \Lambda (\mathbf {y} | \mathbf {p}), \tag {4} +$$ + +where $N$ is the number of similar patches, $\Lambda (\cdot |\cdot)$ is the operation of searching for and extracting similar patches. Note that the notation $\mathbf{y}_i$ should depend on $\mathbf{p}$ but we omit it for conciseness. + +(2) Exploit the correlation within the patch group $\{\mathbf{y}_i\}_{i = 1}^N$ to generate an estimate for each patch: + +$$ +\tilde {\mathbf {x}} _ {1}, \tilde {\mathbf {x}} _ {2}, \dots , \tilde {\mathbf {x}} _ {N} = \Pi \left(\left\{\mathbf {y} _ {i} \right\} _ {i = 1} ^ {N}\right), \tag {5} +$$ + +where $\tilde{\mathbf{x}}_i$ is the estimate of the true signal $\mathbf{x}_i$ corresponding to $\mathbf{y}_i$ . The actual operation of $\Pi(\cdot)$ depends on the prior model adopted to characterize the consistency between patches $\{\mathbf{y}_i\}_{i=1}^N$ . Generally, $\Pi(\cdot)$ can be regarded as an operator to solve the following optimization objective: + +$$ +\left\{\tilde {\mathbf {x}} _ {i} \right\} = \arg \min \sum_ {i = 1} ^ {N} \left| \left| \mathbf {x} _ {i} - \mathbf {y} _ {i} \right| \right| _ {2} ^ {2} + \Phi \left(\left\{\mathbf {x} _ {i} \right\} _ {i = 1} ^ {N}\right) \tag {6} +$$ + +BM3D [6] solves Eq. (6) by applying a 3D transform to the stacked similar patches and constructing a sparsity regularization term in the transform domain. It utilizes hard-thresholding, soft-thresholding, and Wiener filtering in different steps, corresponding to $l_{0}$ , $l_{1}$ , and $l_{2}$ + +![](images/82f650853a543bc0384bc7025ac8ff79d805a954f33dbf8e841bc620bf268102.jpg) +Figure 3. The main idea of motion aligned collaborative filtering. First, we collect temporal similar features via motion alignment. Then, we stack these aligned features and perform transform-domain collaborative filtering to generate refinement features for each moment. Subsequently, the refinement features are unstacked and mapped back to their original positions via inverse warping. These repositioned refinement features are aggregated with input features to obtain refined features with improved consistency. + +sparsity norm, respectively. BAS [43] applies a content adaptive 2D decorrelation transform to the similar patches via principle component analysis (PCA), and builds a separate distribution model and regularization term for each transform band, using the transform coefficients of these patches as samples. It leads to a PCA-domain adaptive soft-thresholding to solve Eq. (6). LLSC [27] adopts another prior model that emphasizes that similar patches tend to share the location pattern of non-zero dictionary elements in their sparse decomposition. WNNM [18], on the other hand, utilizes the low-rank property of similar patches so that the problem (6) reduces to singular value decomposition (SVD) and soft-thresholding on the singular values. + +(3) Place each estimate $\tilde{\mathbf{x}}_i$ back to the position where $\mathbf{y}_i$ is taken from, and aggregate the overlapping estimates from all position $\mathbf{p}$ to form a whole estimate for $\mathbf{x}$ : + +$$ +\tilde {\mathbf {x}} = \Lambda^ {- 1} \left(\left\{\tilde {\mathbf {x}} _ {1}, \tilde {\mathbf {x}} _ {2}, \dots , \tilde {\mathbf {x}} _ {N} \right\} _ {\mathbf {p}}\right). \tag {7} +$$ + +Here $\Lambda^{-1}(\cdot)$ denotes patch reposition and aggregation operation. The reposition of $\tilde{\mathbf{x}}_i$ to $\tilde{\mathbf{x}}$ is the inverse operation of taking $\mathbf{y}_i$ from $\mathbf{y}$ in stage (1). + +Taking these ideas into account, we propose a novel framework tailored for spike data to reduce disturbances in spike information. The main idea is illustrated in Fig. 3. The similarity exploited in this framework is temporal similarity, and it may be extended to include spatial non-local similarity in the future. First, we align the spike features at neighboring moments with the spike feature at the reconstruction moment to obtain a set of similar features. Then, we stack the aligned features and perform a neural-network-based transform-domain collaborative filtering, to separate the disturbances from clean features by exploiting representation sparsity and generate refinement features for each moment. These refinement features are used to update the initial input features. Specifically, the refinement features are unstacked and mapped back to their original positions via inverse warping, and then aggregated with input features. The purpose of this framework is to improve the consistency of multi-moment features along the motion trajectory. + +![](images/2060476049fe074415973e3ab9de0778e54ed28cf5da0f89029d87f22b26f2f4.jpg) +Figure 4. Overall architecture of the proposed spike camera super resolution network Spk2SRImgNet. + +# 5. Method + +# 5.1. Overall Architecture + +To reconstruct HR dynamic scenes from LR binary spike stream, we develop an end-to-end trainable spike camera super resolution network, Spk2SRImgNet. The overall architecture is shown in Fig. 4. To reconstruct an HR image $\mathcal{I}(t_0) \in \mathbb{R}^{rH \times rW}$ at time $t_0$ , we take a set of consecutive spike frames $S(t_0 + i), i = 0, \pm 1, \pm 2, \ldots$ around $S(t_0)$ as input, where $S(t_0 + i) \in \mathbb{B}^{H \times W \times 1}$ , $r$ is SR scale, $\mathbb{B}$ represents the binary domain. + +Firstly, the input spike stream is partitioned into $K$ overlapping short-term spike blocks $\{B_i\}_{i = -n}^n (K = 2n + 1)$ , where $B_{i}\in \mathbb{B}^{H\times W\times L}$ , $L$ is the temporal length of each block, $B_{0}$ is the key block centered at time $t_0$ , $\{B_i\}, i = \pm 1,\pm 2,\ldots$ are neighboring blocks distributed symmetrically on both sides of $B_{0}$ , $B_{i}$ is centered at time $t_i$ . To bridge the domain gap between binary spikes and real pixel values, we first convert spike frames in $B_{i} = \{S(u)\}_{u = t_{i} - w_{h}}^{t_{i} + w_{h}}$ to intensity frames in $Y_{i} = \{Y(u)\}_{u = t_{i} - w_{h}}^{t_{i} + w_{h}}\in \mathbb{R}^{H\times W\times L}$ based on the spike interval based reconstruction (SIR) method [53], where $L = 2w_{h} + 1$ . Then, we extract features $\{F_i\}_{i = -n}^n$ from $\{Y_i\}_{i = -n}^n$ through an encoder, which consists of two convolutional layers and three residual blocks. To reduce fluctuations, $\{F_i\}_{i = -n}^n$ are input to motion aligned collaborative filtering (MACF) module, which generates improved features $\{F_i\}_{i = -n}^n$ . Finally, the multi-moment features $\{\hat{F}_i\}_{i = -n}^n$ are fed into motion aligned super resolved reconstruction module to generate an HR image $\mathcal{I}(t_0)$ . + +![](images/3318ce1e321fdff7b9477e3808f00cc75f152ab598022ff514495d56f02ecaa8.jpg) +Figure 5. The overall framework of motion aligned collaborative filtering (MACF) module. + +![](images/b7624ecf54f11fd719d43f561df9dee72268361eb7507230fd596663c32dfe1f.jpg) +Figure 6. Illustration of channel rearrangement (CR) and inverse channel rearrangement (ICR). The left part is $K$ features $\{F_i\}_{i = -n}^n$ at $K$ different moments, where $F_{i}\in \mathbb{R}^{H\times W\times C}$ . The right part is $C$ features $\{G_j\}_{j = 0}^{C - 1}$ , where $G_{j}\in \mathbb{R}^{H\times W\times K}$ , each $G_{j}$ contains features at $K$ different moments. + +# 5.2. Motion Aligned Collaborative Filtering + +We propose a motion aligned collaborative filtering (MACF) module to implement the idea proposed in Sec. 4. The overall framework of MACF is shown in Fig. 5, which includes six steps. We provide a detailed description of each step below. + +Step 1: Offset Estimation. This step generates the forward $(F_{i}\rightarrow F_{0})$ and backward $(F_0\to F_i)$ motion offsets between neighboring features $\{F_i\} ,i = \pm 1,\dots ,\pm n$ and key feature $F_{0}$ . These offsets will be utilized in the subsequent feature warping (step 3) and inverse feature warping (step 5), respectively. Specifically, the forward and backward motion offsets: $\mathcal{O}_i^{\mathbf{F}}$ and $\mathcal{O}_i^{\mathbf{B}}$ are both derived from $F_{i}$ and $F_{0}$ , which are formulated as follows ( $\forall i\in \{\pm 1,\dots ,\pm n\}$ ): + +$$ +\begin{array}{l} \mathcal {O} _ {i} ^ {\mathbf {F}} = m _ {1} \left(\left\{F _ {i}, F _ {0} \right\}\right), \\ \mathcal {O} _ {i} ^ {\mathbf {B}} = m _ {2} \left(\left\{F _ {i}, F _ {0} \right\}\right), \tag {8} \\ \end{array} +$$ + +where $m_{1}(\cdot)$ and $m_{2}(\cdot)$ denote two independent convolutional blocks that consist of several convolutional layers, $\{,\}$ denotes channel concatenation operation. + +Step 2: Channel Rearrange. The $\{F_i\}_{i = -n}^n$ at $K$ different moments are features extracted from the fluctuated reconstruction $\{Y_{i}\}_{i = -n}^{n}$ , where $F_{i}\in \mathbb{R}^{H\times W\times C}$ , $F_{i} = \{F_{i}^{0},F_{i}^{1},\dots,F_{i}^{C - 1}\}$ , the superscript of $F_{i}$ denotes channel + +index. Each feature $F_{i}$ consists of $C$ channels, with each channel encoding different information about the scene. For example, some channels capture edges, some channels focus on textures, and so on. We perform channel rearrangement (CR) to collect data at the same channel index for each feature and concatenate them, as shown in Fig. 6. We form $\{G_j\}_{j = 0}^{C - 1}$ from $\{F_i\}_{i = -n}^n$ via CR operation, where $G_{j}\in \mathbb{R}^{H\times W\times K}$ , $G_{j} = \{G_{j}^{-n},\dots,G_{j}^{0},\dots,G_{j}^{n}\} = \{F_{-n}^{j},\dots,F_{0}^{j},\dots,F_{n}^{j}\}$ , the subscript of $G$ denotes channel index, the superscript of $G$ denotes time index. $G_{j}$ contains $K$ features with channel index $j$ at $K$ different moments. To offer an intuitive understanding of this operation, we provide a visualization of CR in supplementary material. After CR, steps 3, 4 and 5 are performed sequentially for each $G_{j}$ . Step 3: Feature Warping. Step 2 collects temporally unaligned features that encode the same information. We perform feature warping based on deformable convolution $\mathcal{D}$ to align neighboring features $\{G_j^i\}$ , $i = \pm 1,\ldots ,\pm n$ with key feature $G_{j}^{0}$ . Specifically, given the forward offsets from neighboring features to key feature $\{\mathcal{O}_i^{\mathbf{F}}\}$ , $i = \pm 1,\ldots ,\pm n$ , the aligned features could be obtained by: + +$$ +\tilde {G} _ {j} ^ {i} = \mathcal {D} \left(G _ {j} ^ {i}, \mathcal {O} _ {i} ^ {\mathbf {F}}\right) = \mathcal {D} \left(F _ {i} ^ {j}, \mathcal {O} _ {i} ^ {\mathbf {F}}\right), \forall i \in \{\pm 1, \dots , \pm n \}. \tag {9} +$$ + +These aligned features $\{\tilde{G}_j^i\}_{i = -n}^n$ are then concatenated to derive $\tilde{G}_j = \{\tilde{G}_j^{-n},\dots,\tilde{G}_j^n\}$ , where $\tilde{G}_j^0 = G_j^0 = F_0^j$ . $\tilde{G}_j$ could be regarded as stacked similar features. + +Step 4: Transform Domain Stacked Feature Filtering. We design a Transform Domain stacked feature Filtering (TDF) step to generate refinement features $\Delta \tilde{G}_j$ from stacked similar features $\tilde{G}_j$ by leveraging representation sparsity in transform domain. TDF is designed with four submodules: transform, noise estimation, filtering, and inverse transform, as depicted in step 4 of Fig. 5. Specifically, the input $\tilde{G}_j$ is first transformed into feature $\dot{G}_j$ . Simultaneously, noise feature $N_j$ are extracted from $\tilde{G}_j$ via noise estimation submodule. $\dot{G}_j$ and $N_j$ are passed through filter + +![](images/8bfd45319a78c07051618e16cb04cf02ba648aae7b83bd7b9c3f4d2045f14d25.jpg) +Figure 7. Illustration of submodules in step 4 of MACF. (a) Transform and inverse transform submodules; (b) noise estimation submodule; (c) filtering submodule. + +ing submodule to generate $\ddot{G}_j$ . $\ddot{G}_j$ are subsequently fed into inverse transform submodule. Additionally, a residual connection is applied in TDF step, producing refinement features $\Delta \tilde{G}_j$ at last. + +The above four submodules are illustrated in detail in Fig. 7. Transform and inverse transform submodules are independently realized via several convolutional layers (Fig. 7 (a)), involving both spatial and inter-channel transform. For noise estimation (NE) (Fig. 7 (b)), we first average $\tilde{G}_j$ along the channel dimension to obtain a rough temporal average result $\mathcal{E}$ . Then we calculate the residual between $\tilde{G}_j$ and $\mathcal{E}$ to represent fluctuations in features. Lastly, we extract deeper features from the residual to derive noise feature $N_j$ . We concatenate the transformed feature $\dot{G}_j$ and noise feature $N_j$ as input to the filtering submodule (Fig. 7 (c)), which contains a U-Net structure [33] and outputs $\ddot{G}_j$ . + +Step 5: Inverse Feature Warping and Aggregation. We propose an inverse motion alignment operation to map these multi-moment refinement features $\Delta \tilde{G}_j$ to their original positions. Specifically, given the backward offsets from key feature to neighboring features $\{\mathcal{O}_i^{\mathbf{B}}\}$ , $i = \pm 1,\dots,\pm n$ , we achieve inverse feature warping by: + +$$ +\Delta G _ {j} ^ {i} = \mathcal {D} \left(\Delta \tilde {G} _ {j} ^ {i}, \mathcal {O} _ {i} ^ {\mathbf {B}}\right), \forall i \in \{\pm 1, \dots , \pm n \}. \tag {10} +$$ + +$\Delta G_{j}^{i}$ is repositioned refinement feature and corresponds to initial unaligned feature $G_{j}^{i}$ . We concatenate these repositioned refinement features and obtain $\Delta G_{j} = \{\Delta G_{j}^{-n}, \dots, \Delta G_{j}^{n}\}$ , where $\Delta G_{j}^{0} = \Delta \tilde{G}_{j}^{0}$ . $\Delta G_{j}$ are then aggregated with initial features $\tilde{G}_{j}$ to derive refined features with improved consistency: + +$$ +\acute {G} _ {j} = G _ {j} + \Delta G _ {j}. \tag {11} +$$ + +Step 6: Inverse Channel Rearrange. Finally, the features $\{\hat{G}_j\}_{j=0}^{C-1}$ are restored to $\{\hat{F}_i\}_{i=-n}^n$ via inverse channel rearrangement (ICR), i.e., the inverse operation of CR, as shown in Fig. 6. + +MACF alleviates disturbances in $\{F_i\}_{i = -n}^n$ and generate improved features $\{\dot{F}_i\}_{i = -n}^n$ , which are beneficial for enhancing SR reconstruction quality. + +![](images/e4afe44b1bf93eb80fd5222ea4d8074623b4a93178ad135d3fe086f41774c847.jpg) +Figure 8. Illustration of motion aligned super resolved reconstruction module. + +# 5.3. Motion Aligned Super Resolved Reconstruction + +The improved features $\{\hat{F}_i\}_{i = -n}^n$ are then input to motion aligned super resolved reconstruction module to generate the high resolution image $\mathcal{I}(t_0)$ , as illustrated in Fig. 8. Firstly, features at neighboring moments $\{\hat{F}_i\}, i = \pm 1,\dots,\pm n$ are aligned with the key feature $\hat{F}_0$ by deformable convolution [7, 56], where the features $\{\hat{F}_i\}_{i = -n}^n$ with improved consistency help enhance the accuracy of motion alignment, benefiting the utilization of temporal information. Subsequently, key feature and aligned neighboring features are concatenated and fused through several convolutional layers. The fused feature is upsampled to the HR image by pixel shuffle operation [3]. + +# 6. Experiments + +# 6.1. Implementation Details + +We set the number of input spike frames to 101. The number of partitioned short-term blocks is 9 ( $K = 9$ , $n = 4$ ). The temporal length $L$ of each block $B_{i}$ is 21. The channel $C$ of $F_{i}$ and $\hat{F}_i$ is 64. During training, we randomly crop the spike stream to a spatial size of $64\times 64$ . In addition, we randomly perform horizontal and vertical flips as well as rotation for data augmentation. The batch size is set to 8. We use Adam optimizer [23] with $\beta_{1} = 0.9$ and $\beta_{2} = 0.999$ . The learning rate is initialized as 2e-4 and decayed by 0.7 every 100 epochs. The model is trained with $\mathcal{L}_1$ loss for 800 epochs. All models are trained and tested on NVIDIA RTX 3090 GPU. + +# 6.2. Datasets + +Synthesized Dataset. We need a large number of "LR spike stream and HR image" pairs to train the model. However, it is challenging to obtain corresponding high-quality HR images, especially in high dynamic scenes. Inspired by the experiments conducted in [48, 50], we use a spike camera simulator (detailed in supplementary material) to mimic the working mechanism of spike camera and generate synthesized SCSR datasets for training and evaluation, where + +
ScaleMethodREDS-based DatasetAdobe240-based DatasetParams(M)Runtime(s)
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
×2TFP+DAT25.550.59540.445627.660.67820.367714.654.40
Spk2ImgNet+DAT31.770.90240.190734.920.96040.094918.564.63
WGSE+DAT31.540.89420.205534.750.95710.102218.604.50
BSF+DAT31.920.90440.186735.030.96100.093417.135.01
VidarSR31.970.89640.204035.130.95580.110510.162.09
SpikeSR-Net34.290.93300.148137.040.97040.08383.262.14
Ours34.500.93770.140337.220.97250.08163.720.81
×3TFP+DAT24.880.57540.446426.170.65770.372114.841.88
Spk2ImgNet+DAT28.900.82930.284930.500.90640.170818.752.10
WGSE+DAT28.760.82170.296130.330.90130.177918.782.02
BSF+DAT28.990.83150.280730.440.90540.169917.322.02
VidarSR29.550.83730.279132.010.91970.168511.380.85
SpikeSR-Net31.090.87980.230433.740.94390.13553.301.35
Ours31.410.88500.223134.110.94700.13273.900.37
×4TFP+DAT24.350.56170.473425.090.64070.399514.801.06
TFP+BasicVSR++23.690.50230.485024.210.56100.44157.320.08
Spk2ImgNet+DAT27.230.76780.352127.970.84590.237818.711.15
Spk2ImgNet+BasicVSR++27.280.77300.352727.770.84210.252611.230.17
WGSE+DAT27.150.76180.359727.910.84160.242718.751.08
WGSE+BasicVSR++27.160.76990.356127.720.84040.255811.080.12
BSF+DAT27.300.77010.347327.900.84510.235817.281.10
BSF+BasicVSR++27.240.78180.337427.430.83890.24819.800.18
VidarSR28.300.79600.324930.180.88340.209712.790.75
SpikeSR-Net29.220.82660.294631.310.90730.18853.341.07
Ours29.500.83690.281631.490.91190.18223.860.21
+ +Table 1. Quantitative results on REDS-based and Adobe240-based SCSR evaluation dataset. Red and blue indicate the best and the second-best performance, respectively. It is noted that BasicVSR++ [4] only support $\times 4$ SR scale. The runtime is tested on input spike stream of sizes $360 \times 640 \times 101 (\times 2)$ , $240 \times 426 \times 101 (\times 3)$ and $180 \times 320 \times 101 (\times 4)$ , respectively. + +
CaseSIRMACFPSNR↑SSIM↑
step 3step 4step 5
(A)29.070.8207
(B)✓29.150.8237
(C)✓29.170.8244
(D)✓29.320.8306
(E)✓✓29.340.8316
(F)✓✓29.350.8322
(G)✓✓29.330.8304
(H)✓✓✓29.460.8344
(I)✓✓✓✓29.500.8369
+ +Table 2. Ablation study of the proposed Spk2SRImgNet on REDS-based SCSR evaluation dataset $(\times 4)$ + +each sample consists of a spike stream-image pair. + +Specifically, we use the 120 fps REDS dataset [30] with a spatial resolution of $720 \times 1280$ to generate the synthesized dataset. REDS includes 240 scenes for training and 30 scenes for evaluation, and we follow its original partition. For each scene, we choose 5 segments, resulting in $240 \times 5 = 1200$ and $30 \times 5 = 150$ pairs for training and evaluation, respectively. Each pair consists of spike + +stream of size $360 \times 640 \times 101 \times 2) / 240 \times 426 \times 101 \times 3) / 180 \times 320 \times 101 \times 4$ (101 is the temporal length) and a ground truth image of size $720 \times 1280 \times 2$ and $\times 4$ / $720 \times 1278 \times 3$ . To demonstrate the generalization of the proposed model, we generate another SCSR evaluation dataset from Adobe240 dataset [34]. Adobe240 dataset includes 133 videos of 240 fps with resolution of $720 \times 1280$ taken by hand-held cameras. We follow the test set partition of Adobe240 in [44], which includes 17 scenes. For each scene, we choose 10 segments, generating $17 \times 10 = 170$ pairs for evaluation. + +Real-World Captured Data. In addition to the synthesized data, we also evaluate the model performance on real-world spike data captured by the spike camera with a spatial resolution of $250 \times 400$ . Due to the absence of ground truth, we perform qualitative analysis for these data. + +# 6.3. Comparison with Existing Methods + +We compare our proposed method with existing SCSR method: VidarSR [42] and SpikeSR-Net [50]. Besides, we combine four spike camera image reconstruction methods: TFP [53], Spk2ImgNet [48], WGSE [45] and BSF [52] with + +![](images/9c2ca347e4b924fca87f0eeaabd76516ef0640d65a5e9f9f3102e02bcf853b86.jpg) +Spike + +![](images/edf5a0058cc08da47aeaf2a5c4cdc9beb9f2e63022ade05c3be3909e08bd26ad.jpg) +TFP+DAT + +![](images/3ec66d28c1f792a036be8c30fbd3e32e1676b6c2e90a002ad23aa9cd4f63f6cb.jpg) +TFP+BasicVSR++ + +![](images/2aa6d8f6d3fc9e99e14ed5877f7677e051a8d803e97d6a5029507202e1c787a7.jpg) +Spk2ImgNet+DAT + +![](images/67a356828c4b89629382dec22ba463b7d0eeefaf6981262e297c0ee9ddc8094c.jpg) +Spk2ImgNet+BasicVSR++ + +![](images/2dd1c94efa4edc0d926bd0a1066ab9708513b19e822e38487a01b362a0c01d21.jpg) +WGSE+DAT + +![](images/e6b86425aa247392e9321f625813d9c450f5108d7ad88fb201ccd7c76d2f9c8a.jpg) +WGSE+BasicVSR++ + +![](images/3fe5b20d437dbeb07174671b2f94f5f017f72b7fe115a856e8e4b986c784ea72.jpg) +BSF+DAT + +![](images/ea64bfaa7344eed273c514d5f555b6575630351c4c9d05e2ad3f1032897ae654.jpg) +BSF+BasicVSR++ + +![](images/d46f6e6f139f61284ace99c9a38d148fb565f9f911d84c598bb019b3e5272db6.jpg) +VidarSR + +![](images/51f15c14bf48fedb3509deb57bdeafd18590766bf48080f3fab9d3d11b8d974e.jpg) +SpikeSR-Net + +![](images/c4a67c6d2ddc6fbc50b1a32b34f39f69d02d81f38590a86a13b501bd509fc11b.jpg) +Ours + +![](images/88a6b2f1bf8d21314180e59a00823a09b6c24d8fed1d8137943ac84578d01027.jpg) + +![](images/82e5a4ae2802552d2a6191889da7c10db672c987f8526a61e7f1939cdd98734a.jpg) + +![](images/fee3930eac04be89d9a12881237818cedd0d6709309351dcddd25976f8611d87.jpg) + +![](images/1f8415383d0f7909ea2829b17a6828c08d7db9ab6037c015dd702cbbaff4552d.jpg) + +![](images/ffc0b050aa9dd4169f8e6e65b542be65edf8758d46e7320c43d8647747d083fb.jpg) + +![](images/943720ade2f9e54c09b579f45416bd23b83dedfd061211ea3d881e27e640fcfc.jpg) + +![](images/392c45de6f37c59b29fb1b8c240885fc207d57331002d8d36ad2aeb132141268.jpg) +Spike +WGSE+BasicVSR++ +Figure 9. Visual comparison $(\times 4)$ on real-world spike data. The first scene is a stationary clock recorded by a fast-moving spike camera. The second scene is a rotating fan. Please enlarge the figure for better comparison. + +![](images/bab5e91e3ea16b40b013c32ef4b377f8a8e1d1762327adb18577e1df2dee4cef.jpg) +TFP+DAT +BSF+DAT + +![](images/fa1139c06a0f2928144c25a979fe5dcb7aeb69ceeda0cf82107b826098caca39.jpg) +TFP+BasicVSR++ +BSF+BasicVSR++ + +![](images/3937c6da29913f837123c7962252d6e7d5e0a787ba184346e859b18099e9c37f.jpg) +Spk2ImgNet+DAT +VidarSR + +![](images/a32869622dd3b7ae426783078b3ec66e3402d53361ffa21e661e49375bd6259a.jpg) +Spk2ImgNet+BasicVSR++ +SpikeSR-Net + +![](images/28c2d0a3f7e3ac01de22cc6354db7632c128e8f2f6abef6c2c8b9eff92e6895a.jpg) +WGSE+DAT +Ours + +image / video SR method: DAT [5] / BasicVSR++ [4], resulting in another eight comparison methods, as presented in Tab. 1. We conduct experiments on $\times 2$ , $\times 3$ , and $\times 4$ SR scales. We employ PSNR, SSIM [39] and LPIPS [46] as quantitative metrics. All compared methods except TFP [53] are retrained on the REDS-based SCSR training dataset with the same settings as ours. As shown in Tab. 1, our method achieves the best performance on the three metrics. Besides, we conduct experiments on real-world captured spike stream, as shown in Fig. 9, our method can restore better textures and details. More evaluation and visual results are provided in supplementary material. + +# 6.4. Ablation Study + +To verify the effectiveness of our methodology, we perform ablation study with REDS-based evaluation dataset on $\times 4$ SR scale, as presented in Tab. 2. Since ablation study only involves steps 3, 4, and 5 in MACF, the remaining steps are omitted in Tab. 2 (Steps 2 and 6 do not involve network parameters, and step 1 is present only when step 3 and/or step 5 are present). Case (I) is the final model. Case (A) is formed by removing SIR and MACF modules from + +Spk2SRImgNet. Comparisons between (A) and (B), (H) and (I) illustrate the advantages of converting information from spike domain to image domain (SIR module). Comparisons between (A) and (H), (B) and (I) demonstrate the effectiveness of MACF module, showing gains of 0.39dB and 0.35dB, respectively. Comparisons between (C-G) and (H) validate the effectiveness of each step in MACF. More ablation study is provided in supplementary material. + +# 7. Conclusion + +This paper develops an end-to-end neural network named Spk2SRImgNet to reconstruct high resolution images from low resolution spike stream. To handle spike fluctuations, we propose a motion aligned collaborative filtering (MACF) module motivated by key ideas in classic image restoration schemes. Specifically, MACF leverages the temporal similarity of spike stream to enhance the consistency of multi-moment features along the motion trajectory, mitigating fluctuations and improving the quality of super resolution reconstruction. Experimental results demonstrate that the proposed Spk2SRImgNet achieves state-of-the-art performance on synthesized and real-captured spike data. + +# Acknowledgements + +This work was supported by the National Natural Science Foundation of China under Grants 62072009 and U22B2035. + +# References + +[1] Patrick Bardow, Andrew J. Davison, and Stefan Leutenegger. Simultaneous optical flow and intensity estimation from an event camera. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 884-892, 2016. 2 +[2] Christian Brandli, Raphael Berner, Minhao Yang, Shih-Chii Liu, and Tobi Delbruck. A $240 \times 180$ 130 db $3\mu s$ latency global shutter spatiotemporal vision sensor. IEEE Journal of Solid-State Circuits, 49(10):2333-2341, 2014. 1 +[3] Jose Caballero, Christian Ledig, Andrew Aitken, Alejandro Acosta, Johannes Totz, Zehan Wang, and Wenzhe Shi. Real-time video super-resolution with spatio-temporal networks and motion compensation. In 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2848-2857, 2017. 2, 6 +[4] Kelvin C.K. Chan, Shangchen Zhou, Xiangyu Xu, and Chen Change Loy. Basicvsr++: Improving video superresolution with enhanced propagation and alignment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5972-5981, 2022. 2, 7, 8 +[5] Zheng Chen, Yulun Zhang, Jinjin Gu, Linghe Kong, Xiaokang Yang, and Fisher Yu. Dual aggregation transformer for image super-resolution. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 12312-12321, 2023. 8 +[6] Kostadin Dabov, Alessandro Foi, Vladimir Katkovnik, and Karen Egiazarian. Image denoising by sparse 3-d transform-domain collaborative filtering. IEEE Transactions on Image Processing, 16(8):2080-2095, 2007. 2, 3 +[7] Jifeng Dai, Haozhi Qi, Yuwen Xiong, Yi Li, Guodong Zhang, Han Hu, and Yichen Wei. Deformable convolutional networks. In 2017 IEEE International Conference on Computer Vision (ICCV), pages 764-773, 2017. 2, 6 +[8] Siwei Dong, Tiejun Huang, and Yonghong Tian. Spike camera and its coding methods. In 2017 Data Compression Conference (DCC), pages 437-437, 2017. 1 +[9] Siwei Dong, Lin Zhu, Daoyuan Xu, Yonghong Tian, and Tiejun Huang. An efficient coding method for spike camera using inter-spike intervals. In 2019 Data Compression Conference (DCC), pages 568-568, 2019. 1 +[10] Yanchen Dong, Jing Zhao, Ruiqin Xiong, and Tiejun Huang. 3D residual interpolation for spike camera demosaicing. In IEEE International Conference on Image Processing (ICIP), pages 1461-1465. IEEE, 2022. 2 +[11] Yanchen Dong, Ruiqin Xiong, Jian Zhang, Zhaofei Yu, Xiaopeng Fan, Shuyuan Zhu, and Tiejun Huang. Superresolution reconstruction from bayer-pattern spike streams. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 24871-24880, 2024. 2 + +[12] Yanchen Dong, Ruiqin Xiong, Jing Zhao, Jian Zhang, Xiaopeng Fan, Shuyuan Zhu, and Tiejun Huang. Joint demosaicing and denoising for spike camera. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1582-1590, 2024. 2 +[13] Yanchen Dong, Ruiqin Xiong, Jing Zhao, Jian Zhang, Xiaopeng Fan, Shuyuan Zhu, and Tiejun Huang. Learning a deep demosaicing network for spike camera with color filter array. IEEE Transactions on Image Processing, 33:3634-3647, 2024. +[14] Yanchen Dong, Ruiqin Xiong, Xiaopeng Fan, Shuyuan Zhu, Jin Wang, and Tiejun Huang. Dynamic scene reconstruction for color spike camera via zero-shot learning. IEEE Transactions on Computational Imaging, 11:129-141, 2025. 2 +[15] Daniel Gehrig and Davide Scaramuzza. Low-latency automotive vision with event cameras. Nature, 629(8014):1034-1040, 2024. 1 +[16] Arren Glover, Valentina Vasco, and Chiara Bartolozzi. A controlled-delay event camera framework for on-line robotics. In 2018 IEEE International Conference on Robotics and Automation (ICRA), pages 2178-2183. IEEE, 2018. 1 +[17] Sorin Grigorescu, Bogdan Trasnea, Tiberiu Cocias, and Gigel Macesanu. A survey of deep learning techniques for autonomous driving. Journal of field robotics, 37(3):362-386, 2020. 1 +[18] Shuhang Gu, Lei Zhang, Wangmeng Zuo, and Xiangchu Feng. Weighted nuclear norm minimization with application to image denoising. In 2014 IEEE Conference on Computer Vision and Pattern Recognition, pages 2862-2869, 2014. 2, 3, 4 +[19] Tiejun Huang, Yajing Zheng, Zhaofei Yu, Rui Chen, Yuan Li, Ruiqin Xiong, Lei Ma, Junwei Zhao, Siwei Dong, Lin Zhu, Jianing Li, Shanshan Jia, Yihua Fu, Boxin Shi, Si Wu, and Yonghong Tian. $1000\times$ faster camera and machine vision with ordinary devices. Engineering, 25:110-119, 2023. 1 +[20] Arpit Jadon, Mohd Omama, Akshay Varshney, Mohammad Samar Ansari, and Rishabh Sharma. Firenet: A specialized lightweight fire & smoke detection model for real-time IoT applications. arXiv preprint arXiv:1905.11922, 2019. 2 +[21] Armin Kappeler, Seunghwan Yoo, Qiqin Dai, and Aggelos K. Katsaggelos. Video super-resolution with convolutional neural networks. IEEE Transactions on Computational Imaging, 2(2):109-122, 2016. 2 +[22] Hanme Kim, Ankur Handa, Ryad Benosman, Sio-Hoi Ieng, and Andrew Davison. Simultaneous mosaicing and tracking with an event camera. In Proceedings of the British Machine Vision Conference. BMVA Press, 2014. 2 +[23] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6 +[24] Jingyun Liang, Jiezhang Cao, Yuchen Fan, Kai Zhang, Rakesh Ranjan, Yawei Li, Radu Timofte, and Luc Van Gool. Vrt: A video restoration transformer. IEEE Transactions on Image Processing, 33:2171-2182, 2024. 2 +[25] Patrick Lichtsteiner, Christoph Posch, and Tobi Delbruck. A $128 \times 128$ 120 db $15\mu s$ latency asynchronous temporal con + +trast vision sensor. IEEE journal of solid-state circuits, 43 (2):566-576, 2008. 1 +[26] Martin Litzenberger, Christoph Posch, D Bauer, Ahmed Nabil Belbachir, P Schon, B Kohn, and H Garn. Embedded vision system for real-time object tracking using an asynchronous transient vision sensor. In 2006 IEEE 12th Digital Signal Processing Workshop & 4th IEEE Signal Processing Education Workshop, pages 173-178. IEEE, 2006. 1 +[27] Julien Mairal, Francis Bach, Jean Ponce, Guillermo Sapiro, and Andrew Zisserman. Non-local sparse models for image restoration. In 2009 IEEE 12th International Conference on Computer Vision, pages 2272-2279, 2009. 2, 3, 4 +[28] Mohammad Mostafavi, Yeongwoo Nam, Jonghyun Choi, and Kuk-Jin Yoon. E2sri: Learning to super-resolve intensity images from events. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(10):6890-6909, 2022. 2 +[29] Elias Mueggler. Event-based vision for high-speed robotics. PhD thesis, University of Zurich, 2017. 1 +[30] Seungjun Nah, Sungyong Baik, Seokil Hong, Gyeongsik Moon, Sanghyun Son, Radu Timofte, and Kyoung Mu Lee. Ntire 2019 challenge on video deblurring and superresolution: Dataset and study. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2019. 7 +[31] Christoph Posch, Daniel Matolin, and Rainer Wohlgenannt. A qvga 143 db dynamic range frame-free pwm image sensor with lossless pixel-level video compression and time-domain cds. IEEE Journal of Solid-State Circuits, 46(1):259-275, 2010. 1 +[32] Henri Rebecq, René Ranftl, Vladlen Koltun, and Davide Scaramuzza. Events-to-video: Bringing modern computer vision to event cameras. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3857-3866, 2019. 2 +[33] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention-MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, pages 234-241. Springer, 2015. 6 +[34] Shuochen Su, Maurizio Delbrazio, Jue Wang, Guillermo Sapiro, Wolfgang Heidrich, and Oliver Wang. Deep video deblurring for hand-held cameras. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 1279-1288, 2017. 7 +[35] Yapeng Tian, Yulun Zhang, Yun Fu, and Chenliang Xu. Tdan: Temporally-deformable alignment network for video super-resolution. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2 +[36] Lin Wang, Tae-Kyun Kim, and Kuk-Jin Yoon. Eventsr: From asynchronous events to image reconstruction, restoration, and super-resolution via end-to-end adversarial learning. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8312-8322, 2020. 2 +[37] Xintao Wang, Kelvin C.K. Chan, Ke Yu, Chao Dong, and Chen Change Loy. Edvr: Video restoration with enhanced + +deformable convolutional networks. In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 1954-1963, 2019. 2 +[38] Yuanlin Wang, Ruiqin Xiong, Jing Zhao, and Tiejun Huang. Reconstruct dynamic scene for spike camera based on 3d space time similarity. In 2024 IEEE International Conference on Image Processing (ICIP), pages 1595-1601, 2024. 2 +[39] Zhou Wang, A.C. Bovik, H.R. Sheikh, and E.P. Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing, 13(4): 600-612, 2004. 8 +[40] Philippe Weinzaepfel, Jerome Revaud, Zaid Harchaoui, and Cordelia Schmid. Deepflow: Large displacement optical flow with deep matching. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), pages 1385-1392, 2013. 2 +[41] Wenming Weng, Yueyi Zhang, and Zhiwei Xiong. Event-based video reconstruction using transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 2 +[42] Xijie Xiang, Lin Zhu, Jianing Li, Yixuan Wang, Tiejun Huang, and Yonghong Tian. Learning super-resolution reconstruction for high temporal resolution spike stream. IEEE Transactions on Circuits and Systems for Video Technology, 33(1):16-29, 2023. 1, 2, 7 +[43] Ruiqin Xiong, Hangfan Liu, Xinfeng Zhang, Jian Zhang, Siwei Ma, Feng Wu, and Wen Gao. Image denoising via bandwise adaptive modeling and regularization exploiting nonlocal similarity. IEEE Transactions on Image Processing, 25 (12):5793-5805, 2016. 2, 3, 4 +[44] Gang Xu, Jun Xu, Zhen Li, Liang Wang, Xing Sun, and Ming-Ming Cheng. Temporal modulation network for controllable space-time video super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6388-6397, 2021. 7 +[45] Jiyuan Zhang, Shanshan Jia, Zhaofei Yu, and Tiejun Huang. Learning temporal-ordered representation for spike streams based on discrete wavelet transforms. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 137–147, 2023. 7 +[46] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 586-595, 2018. 8 +[47] Jing Zhao, Jiyu Xie, Ruiqin Xiong, Jian Zhang, Zhaofei Yu, and Tiejun Huang. Super resolve dynamic scene from continuous spike streams. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 2533-2542, 2021. 2 +[48] Jing Zhao, Ruiqin Xiong, Hangfan Liu, Jian Zhang, and Tiejun Huang. Spk2imgnet: Learning to reconstruct dynamic scene from continuous spike stream. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11996-12005, 2021. 2, 6, 7 + +[49] Jing Zhao, Ruiqin Xiong, Jiyu Xie, Boxin Shi, Zhaofei Yu, Wen Gao, and Tiejun Huang. Reconstructing clear image for high-speed motion scene with a retina-inspired spike camera. IEEE Transactions on Computational Imaging, 8:12-27, 2021. 3 +[50] Jing Zhao, Ruiqin Xiong, Jian Zhang, Rui Zhao, Hangfan Liu, and Tiejun Huang. Learning to super-resolve dynamic scenes for neuromorphic spike camera. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3579-3587, 2023. 1, 2, 6, 7 +[51] Rui Zhao, Ruiqin Xiong, Jian Zhang, Zhaofei Yu, Shuyuan Zhu, Lei Ma, and Tiejun Huang. Spike camera image reconstruction using deep spiking neural networks. IEEE Transactions on Circuits and Systems for Video Technology, 34(6): 5207-5212, 2024. 2 +[52] Rui Zhao, Ruiqin Xiong, Jing Zhao, Jian Zhang, Xiaopeng Fan, Zhaofei Yu, and Tiejun Huang. Boosting spike camera image reconstruction from a perspective of dealing with spike fluctuations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 24955-24965, 2024. 2, 7 +[53] Lin Zhu, Siwei Dong, Tiejun Huang, and Yonghong Tian. A retina-inspired sampling method for visual texture reconstruction. In 2019 IEEE International Conference on Multimedia and Expo (ICME), pages 1432-1437. IEEE, 2019. 2, 3, 4, 7, 8 +[54] Lin Zhu, Siwei Dong, Jianing Li, Tiejun Huang, and Yonghong Tian. Ultra-high temporal resolution visual reconstruction from a fovea-like spike camera via spiking neuron model. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(1):1233-1249, 2022. 2 +[55] Lin Zhu, Xiao Wang, Yi Chang, Jianing Li, Tiejun Huang, and Yonghong Tian. Event-based video reconstruction via potential-assisted spiking neural network. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3584-3594, 2022. 2 +[56] Xizhou Zhu, Han Hu, Stephen Lin, and Jifeng Dai. Deformable convnets v2: More deformable, better results. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9308-9316, 2019. 2, 6 \ No newline at end of file diff --git a/2025/Spk2SRImgNet_ Super-Resolve Dynamic Scene from Spike Stream via Motion Aligned Collaborative Filtering/images.zip b/2025/Spk2SRImgNet_ Super-Resolve Dynamic Scene from Spike Stream via Motion Aligned Collaborative Filtering/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..b16250f81f51d92a87a1423c59c98e524e0d1b93 --- /dev/null +++ b/2025/Spk2SRImgNet_ Super-Resolve Dynamic Scene from Spike Stream via Motion Aligned Collaborative Filtering/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:792cce6738ad392585a456d0d17be7d1ef5da9ba5601a3e10289e162ce539541 +size 876502 diff --git a/2025/Spk2SRImgNet_ Super-Resolve Dynamic Scene from Spike Stream via Motion Aligned Collaborative Filtering/layout.json b/2025/Spk2SRImgNet_ Super-Resolve Dynamic Scene from Spike Stream via Motion Aligned Collaborative Filtering/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..a37a4e28cc28ec399788e3cd27e18ef0e02ae2d3 --- /dev/null +++ b/2025/Spk2SRImgNet_ Super-Resolve Dynamic Scene from Spike Stream via Motion Aligned Collaborative Filtering/layout.json @@ -0,0 +1,12466 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 99, + 102, + 512, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 102, + 512, + 140 + ], + "spans": [ + { + "bbox": [ + 99, + 102, + 512, + 140 + ], + "type": "text", + "content": "Spk2SRImgNet: Super-Resolve Dynamic Scene from Spike Stream via Motion Aligned Collaborative Filtering" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 145, + 160, + 466, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 160, + 466, + 190 + ], + "spans": [ + { + "bbox": [ + 145, + 160, + 466, + 190 + ], + "type": "text", + "content": "Yuanlin Wang" + }, + { + "bbox": [ + 145, + 160, + 466, + 190 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 145, + 160, + 466, + 190 + ], + "type": "text", + "content": ", Yiyang Zhang" + }, + { + "bbox": [ + 145, + 160, + 466, + 190 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 145, + 160, + 466, + 190 + ], + "type": "text", + "content": ", Ruiqin Xiong" + }, + { + "bbox": [ + 145, + 160, + 466, + 190 + ], + "type": "inline_equation", + "content": "^{1,2,*}" + }, + { + "bbox": [ + 145, + 160, + 466, + 190 + ], + "type": "text", + "content": ", Jing Zhao" + }, + { + "bbox": [ + 145, + 160, + 466, + 190 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 145, + 160, + 466, + 190 + ], + "type": "text", + "content": ", Jian Zhang" + }, + { + "bbox": [ + 145, + 160, + 466, + 190 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 145, + 160, + 466, + 190 + ], + "type": "text", + "content": ", Xiaopeng Fan" + }, + { + "bbox": [ + 145, + 160, + 466, + 190 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 145, + 160, + 466, + 190 + ], + "type": "text", + "content": ", Tiejun Huang" + }, + { + "bbox": [ + 145, + 160, + 466, + 190 + ], + "type": "inline_equation", + "content": "^{1,2}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 187, + 190, + 423, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 187, + 190, + 423, + 203 + ], + "spans": [ + { + "bbox": [ + 187, + 190, + 423, + 203 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 187, + 190, + 423, + 203 + ], + "type": "text", + "content": "School of Computer Science, Peking University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 100, + 203, + 510, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 203, + 510, + 217 + ], + "spans": [ + { + "bbox": [ + 100, + 203, + 510, + 217 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 100, + 203, + 510, + 217 + ], + "type": "text", + "content": "National Key Laboratory for Multimedia Information Processing, Peking University" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 73, + 217, + 537, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 217, + 537, + 232 + ], + "spans": [ + { + "bbox": [ + 73, + 217, + 537, + 232 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 73, + 217, + 537, + 232 + ], + "type": "text", + "content": "School of Electronic and Computer Engineering, Shenzhen Graduate School, Peking University" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 117, + 232, + 493, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 117, + 232, + 493, + 246 + ], + "spans": [ + { + "bbox": [ + 117, + 232, + 493, + 246 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 117, + 232, + 493, + 246 + ], + "type": "text", + "content": "School of Computer Science and Technology, Harbin Institute of Technology" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 233, + 248, + 376, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 233, + 248, + 376, + 258 + ], + "spans": [ + { + "bbox": [ + 233, + 248, + 376, + 258 + ], + "type": "text", + "content": "wangyuanlin@stu.pku.edu.cn" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 91, + 261, + 515, + 273 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 261, + 515, + 273 + ], + "spans": [ + { + "bbox": [ + 91, + 261, + 515, + 273 + ], + "type": "text", + "content": "{yiyangz, rqxiong, jzhaopku, zhangjian.sz, tjhuang}@pku.edu.cn fxp@hit.edu.cn" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 152, + 300, + 200, + 313 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 300, + 200, + 313 + ], + "spans": [ + { + "bbox": [ + 152, + 300, + 200, + 313 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 326, + 297, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 326, + 297, + 639 + ], + "spans": [ + { + "bbox": [ + 55, + 326, + 297, + 639 + ], + "type": "text", + "content": "Spike camera is a kind of neuromorphic camera that records dynamic scenes by firing a stream of binary spikes with extremely high temporal resolution. It demonstrates great potential for vision tasks in high-speed scenarios. One limitation in its current implementation is the relatively low spatial resolution. This paper develops a network called Spk2SRImgNet to super-resolve high resolution images from low resolution spike stream. However, fluctuations in spike stream hinder the performance of spike camera super resolution. To address this issue, we propose a motion aligned collaborative filtering (MACF) module, which is motivated by key ideas in classic image restoration schemes to mitigate fluctuations in spike data. MACF leverages the temporal similarity of spike stream to acquire similar features from neighboring moments via motion alignment. To separate disturbances from features, MACF filters these similar features jointly in transform domain to exploit representation sparsity, and generates refinement features that will be used to update initial fluctuated features. Specifically, MACF designs an inverse motion alignment operation to map these refinement features back to their original positions. The initial features are aggregated with the repositioned refinement features to enhance reliability. Experimental results demonstrate that the proposed method achieves state-of-the-art performance compared with existing methods." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 648, + 135, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 648, + 135, + 659 + ], + "spans": [ + { + "bbox": [ + 56, + 648, + 135, + 659 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 668, + 296, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 668, + 296, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 668, + 296, + 693 + ], + "type": "text", + "content": "High-speed imaging attracts increasing attention in emerging computer vision applications such as autonomous driv-" + } + ] + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 317, + 300, + 553, + 358 + ], + "blocks": [ + { + "bbox": [ + 317, + 300, + 553, + 358 + ], + "lines": [ + { + "bbox": [ + 317, + 300, + 553, + 358 + ], + "spans": [ + { + "bbox": [ + 317, + 300, + 553, + 358 + ], + "type": "image", + "image_path": "482e27539178c625a69df88b96f02f63c321c9eddf755e8369557e0ec804f353.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 366, + 555, + 401 + ], + "lines": [ + { + "bbox": [ + 313, + 366, + 555, + 401 + ], + "spans": [ + { + "bbox": [ + 313, + 366, + 555, + 401 + ], + "type": "text", + "content": "Figure 1. Illustration of spike camera super resolution (SCSR) and visual comparison " + }, + { + "bbox": [ + 313, + 366, + 555, + 401 + ], + "type": "inline_equation", + "content": "(\\times 4)" + }, + { + "bbox": [ + 313, + 366, + 555, + 401 + ], + "type": "text", + "content": " with existing SCSR methods: VidarSR [42] and SpikeSR-Net [50]. Please zoom in for better comparison." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 411, + 555, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 411, + 555, + 506 + ], + "spans": [ + { + "bbox": [ + 313, + 411, + 555, + 506 + ], + "type": "text", + "content": "ing [15, 17] and robotics [16, 29]. Conventional camera employs an exposure-based imaging model, which limits its ability to capture high-speed scenes. Specifically, conventional camera accumulates photoelectric signals over an exposure time window to form a snapshot image, which requires the scene to be still during the exposure period. Otherwise, a single point on a moving object may be projected onto different sensor pixels, resulting in motion blur." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 508, + 556, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 508, + 556, + 700 + ], + "spans": [ + { + "bbox": [ + 313, + 508, + 556, + 700 + ], + "type": "text", + "content": "Neuromorphic camera is an emerging camera that abandons the concept of exposure window and mimics the working mechanism of biological retina. It records intensity information with high temporal resolution, showing great potential for high speed vision tasks. One kind of neuromorphic camera is event camera [2, 25, 26, 31]. It detects light intensity changes of each pixel in the logarithmic domain and fires an event when the change exceeds a specified threshold. Another kind of neuromorphic camera is spike camera [8, 9, 19]. Each pixel of spike camera independently and continuously accumulates photons, and fires a spike when the accumulation exceeds a predefined threshold. Unlike event camera that only records relative changes of light intensity, spike camera captures absolute intensity by firing a stream of binary spikes, providing clearer information for reconstructing image details." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 326, + 701, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 701, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 326, + 701, + 555, + 714 + ], + "type": "text", + "content": "Benefiting from its extremely high temporal resolution" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "type": "text", + "content": "*Corresponding author." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "11416" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": "(e.g. " + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "inline_equation", + "content": "40\\mathrm{kHz}" + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": "), spike camera could sample a large amount of intensity information in a short period. Besides, these samples cover different locations of the scene due to the relative motion between camera and scene. Therefore, there is promising potential to reconstruct high resolution (HR) images from spike stream by converting the dense sampling in temporal domain to spatial domain." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 157, + 294, + 310 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 157, + 294, + 310 + ], + "spans": [ + { + "bbox": [ + 55, + 157, + 294, + 310 + ], + "type": "text", + "content": "This paper aims to super-resolve high resolution images from low resolution spike stream. However, due to the Poisson effect of photon arrival, the quantization effect of spike readout, and the thermal noises introduced by circuits, spikes exhibit fluctuations, i.e., the length of spike accumulation period varies over time, even when light intensity remains unchanged. Thus, the features extracted from fluctuated spikes tend to be noisy and temporally incoherent. These unreliable features provide low-quality information source for super resolution (SR) reconstruction. Moreover, they affect the alignment accuracy when leveraging temporal information, further degrading the quality of SR reconstruction." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 312, + 295, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 312, + 295, + 609 + ], + "spans": [ + { + "bbox": [ + 56, + 312, + 295, + 609 + ], + "type": "text", + "content": "In this paper, we propose a spike-to-super-resolution-image network Spk2SRImgNet. To handle spike fluctuations, we take inspiration from several classic image restoration methods, such as BM3D [6], BAS [43], LSSC [27], and WNNM [18]. These methods incorporate the idea of leveraging non-local similarity within natural images to reduce the unknown random disturbance. We embed these concepts into the design of Spk2SRImgNet and develop a motion aligned collaborative filtering (MACF) module to reduce disturbances in spike information. Firstly, MACF obtains similar features from neighboring moments via motion alignment, exploiting the temporal similarity of spike stream. To separate disturbances from features, MACF performs collaborative filtering on grouped similar features in transform domain for leveraging representation sparsity, and generates refinement features for each moment. These refinement features will be utilized to update initial fluctuated features. Specifically, MACF introduces an inverse motion alignment operation to map these multi-moment refinement features to their original unaligned positions. The initial features are integrated with the repositioned refinement features to improve reliability. Experimental results show that the proposed Spk2SRImgNet achieves state-of-the-art performance on synthesized and real-world spike data." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 621, + 142, + 634 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 621, + 142, + 634 + ], + "spans": [ + { + "bbox": [ + 55, + 621, + 142, + 634 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 641, + 294, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 294, + 713 + ], + "type": "text", + "content": "Video Super Resolution. Video super resolution (VSR) produces high resolution images from low resolution (LR) image sequences. Temporal alignment is a crucial step to model temporal correspondences of misaligned frames. Early learning-based methods [3, 21] employ traditional optical flow estimation [40] for motion compensation. With" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": "the diverse development of network structures, TDAN [35] and EDVR [37] utilize deformable convolution [7, 56] for feature alignment. BasicVSR++ [4] leverages optical flow to guide deformable alignment for stable training. VRT [24] proposes a temporal mutual attention block to facilitate motion estimation and feature alignment." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 144, + 553, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 144, + 553, + 299 + ], + "spans": [ + { + "bbox": [ + 313, + 144, + 553, + 299 + ], + "type": "text", + "content": "Event Camera Reconstruction and Super Resolution. Event camera monitors relative changes in light intensity with high temporal resolution. Reconstructing intensity images from events has been studied in the past decade. Kim et al. [22] propose to use the extended Kalman filter and Poisson reconstruction for event reconstruction. Bardow et al. [1] estimate both intensity and optical flow using the primal-dual algorithm. With the development of deep learning, E2VID [32], FireNet [20], ET-Net [41], EVSNN [55] have been proposed successively with good performance. To reconstruct HR images from LR events, Wang et al. [36] propose an unsupervised GAN EventSR. Mostafavi et al. [28] propose an end-to-end recurrent network E2SRI." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 300, + 553, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 300, + 553, + 502 + ], + "spans": [ + { + "bbox": [ + 313, + 300, + 553, + 502 + ], + "type": "text", + "content": "Spike Camera Reconstruction and Super Resolution. Image reconstruction is a popular topic for spike camera [38, 48, 51-54]. TFI [53] reconstructs images based on the interval between two temporally neighboring spikes. TFP [53] utilizes the number of spikes within a temporal window to estimate the intensity. For better performance, CNN-based [48, 52] and spike neural network (SNN)-based [51, 54] frameworks have been proposed successively. To reconstruct HR images from LR spike stream, Zhao et al. [47] derive the constraint relationship between pixel intensities and each spike period, and obtain HR images based on gradient descent algorithm. Xiang et al. [42] design a multi-level feature learning network to achieve spike camera super resolution. Zhao et al. [50] propose SpikeSR-Net based on the observation model of spike camera. Furthermore, Dong et al. [11] propose a super resolution network for the newly developed color spike camera [10, 12-14]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 514, + 402, + 527 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 514, + 402, + 527 + ], + "spans": [ + { + "bbox": [ + 314, + 514, + 402, + 527 + ], + "type": "text", + "content": "3. Spike Camera" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 534, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 553, + 713 + ], + "type": "text", + "content": "Working Mechanism. Spike camera contains a pixel array of size " + }, + { + "bbox": [ + 313, + 534, + 553, + 713 + ], + "type": "inline_equation", + "content": "H \\times W" + }, + { + "bbox": [ + 313, + 534, + 553, + 713 + ], + "type": "text", + "content": " and each pixel works independently. Each pixel comprises three main components: photon receptor, integrator, and comparator, as shown in Fig. 2. The photoreceptor captures incident photons from external scenes and converts them into electric charges through photoelectric conversion. The integrator accumulates the electric charges continuously. Meanwhile, the comparator checks the accumulated voltage persistently. Once the voltage reaches a predefined threshold " + }, + { + "bbox": [ + 313, + 534, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 534, + 553, + 713 + ], + "type": "text", + "content": ", the pixel sets up a spike flag and resets the integrator, immediately restarting a new \"accumulation-and-set\" cycle. Spike readout circuit periodically checks the spike flag at a frequency of " + }, + { + "bbox": [ + 313, + 534, + 553, + 713 + ], + "type": "inline_equation", + "content": "40\\mathrm{kHz}" + }, + { + "bbox": [ + 313, + 534, + 553, + 713 + ], + "type": "text", + "content": ". When detecting the flag, the circuit outputs a spike \"1\", otherwise, it outputs \"0\". The voltage accumulated by the in" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11417" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 63, + 295, + 185 + ], + "blocks": [ + { + "bbox": [ + 57, + 63, + 295, + 185 + ], + "lines": [ + { + "bbox": [ + 57, + 63, + 295, + 185 + ], + "spans": [ + { + "bbox": [ + 57, + 63, + 295, + 185 + ], + "type": "image", + "image_path": "b1caeff65e5a9539fd7868965e875e8773b7f426097fe61f95efb0a5ce78069a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 88, + 191, + 261, + 202 + ], + "lines": [ + { + "bbox": [ + 88, + 191, + 261, + 202 + ], + "spans": [ + { + "bbox": [ + 88, + 191, + 261, + 202 + ], + "type": "text", + "content": "Figure 2. Working mechanism of spike camera." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 223, + 282, + 235 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 223, + 282, + 235 + ], + "spans": [ + { + "bbox": [ + 55, + 223, + 282, + 235 + ], + "type": "text", + "content": "tegrator at an arbitrary time " + }, + { + "bbox": [ + 55, + 223, + 282, + 235 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 223, + 282, + 235 + ], + "type": "text", + "content": " can be formulated as [49]:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 91, + 240, + 295, + 268 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 240, + 295, + 268 + ], + "spans": [ + { + "bbox": [ + 91, + 240, + 295, + 268 + ], + "type": "interline_equation", + "content": "A (x, y, t) = \\int_ {0} ^ {t} \\eta I (x, y, u) d u \\quad \\mathrm {m o d} \\theta , \\tag {1}", + "image_path": "b08a1bd508b806ef7e0269084eb8920dbe106c8d679809339206eb8c05a5eea7.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 274, + 296, + 333 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 274, + 296, + 333 + ], + "spans": [ + { + "bbox": [ + 55, + 274, + 296, + 333 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 274, + 296, + 333 + ], + "type": "inline_equation", + "content": "I(x,y,u)" + }, + { + "bbox": [ + 55, + 274, + 296, + 333 + ], + "type": "text", + "content": " denotes the arrival rate of photons received by pixel " + }, + { + "bbox": [ + 55, + 274, + 296, + 333 + ], + "type": "inline_equation", + "content": "(x,y)" + }, + { + "bbox": [ + 55, + 274, + 296, + 333 + ], + "type": "text", + "content": " at time " + }, + { + "bbox": [ + 55, + 274, + 296, + 333 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 55, + 274, + 296, + 333 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 274, + 296, + 333 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 55, + 274, + 296, + 333 + ], + "type": "text", + "content": " denotes the photoelectric conversion rate. The output of spike camera is a 3D binary spike matrix " + }, + { + "bbox": [ + 55, + 274, + 296, + 333 + ], + "type": "inline_equation", + "content": "\\{S(u)\\}_{u=1}^{T}" + }, + { + "bbox": [ + 55, + 274, + 296, + 333 + ], + "type": "text", + "content": " with size " + }, + { + "bbox": [ + 55, + 274, + 296, + 333 + ], + "type": "inline_equation", + "content": "H\\times W\\times T" + }, + { + "bbox": [ + 55, + 274, + 296, + 333 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 274, + 296, + 333 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 274, + 296, + 333 + ], + "type": "text", + "content": " is the temporal length, " + }, + { + "bbox": [ + 55, + 274, + 296, + 333 + ], + "type": "inline_equation", + "content": "S(u)" + }, + { + "bbox": [ + 55, + 274, + 296, + 333 + ], + "type": "text", + "content": " is a spike frame at time " + }, + { + "bbox": [ + 55, + 274, + 296, + 333 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 55, + 274, + 296, + 333 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 334, + 296, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 334, + 296, + 430 + ], + "spans": [ + { + "bbox": [ + 55, + 334, + 296, + 430 + ], + "type": "text", + "content": "Spike Interval Based Reconstruction. According to the working mechanism of spike camera, the stronger the light intensity, the shorter the interval between two neighboring spikes, i.e., the spike interval is inversely proportional to the light intensity. This property is exploited by spike-interval-based reconstruction method (SIR) [53]. Specifically, the spike interval of two temporally neighboring spikes that cover the position " + }, + { + "bbox": [ + 55, + 334, + 296, + 430 + ], + "type": "inline_equation", + "content": "(x,y,u)" + }, + { + "bbox": [ + 55, + 334, + 296, + 430 + ], + "type": "text", + "content": " can be formulated as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 85, + 437, + 295, + 464 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 437, + 295, + 464 + ], + "spans": [ + { + "bbox": [ + 85, + 437, + 295, + 464 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\Phi (x, y, u) = \\min \\left\\{t \\mid S (x, y, t) = 1, t \\geq u \\right\\} \\tag {2} \\\\ - \\max \\{t | S (x, y, t) = 1, t < u \\}. \\\\ \\end{array}", + "image_path": "0e682319eb0f4a4bc6cb6fe105408b71bad5c2926f1d306b976c4bce4f342df2.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 472, + 295, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 472, + 295, + 484 + ], + "spans": [ + { + "bbox": [ + 55, + 472, + 295, + 484 + ], + "type": "text", + "content": "SIR estimates the light intensity of pixel " + }, + { + "bbox": [ + 55, + 472, + 295, + 484 + ], + "type": "inline_equation", + "content": "(x,y)" + }, + { + "bbox": [ + 55, + 472, + 295, + 484 + ], + "type": "text", + "content": " at time " + }, + { + "bbox": [ + 55, + 472, + 295, + 484 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 55, + 472, + 295, + 484 + ], + "type": "text", + "content": " by" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 123, + 491, + 294, + 516 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 123, + 491, + 294, + 516 + ], + "spans": [ + { + "bbox": [ + 123, + 491, + 294, + 516 + ], + "type": "interline_equation", + "content": "Y (x, y, u) = \\frac {\\delta}{\\Phi (x , y , u)}, \\tag {3}", + "image_path": "59c079aa0a37b8c9a4c56708cce1258c61c2974da59074e7581bd905a84e691b.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 522, + 295, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 522, + 295, + 544 + ], + "spans": [ + { + "bbox": [ + 55, + 522, + 295, + 544 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 522, + 295, + 544 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 55, + 522, + 295, + 544 + ], + "type": "text", + "content": " denotes the maximum dynamic range of reconstruction." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 545, + 296, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 296, + 712 + ], + "type": "text", + "content": "Spike Fluctuations. The imaging of spike camera is affected by multiple kinds of noise. First, the Poisson process of photon arrival causes the time taken for the number of photons to reach the threshold to exhibit randomness, which is denoted as Poisson noise. Second, the spike readout circuit is controlled by a periodic clock signal, i.e., it only outputs spikes at discrete times, which results in a time difference between spike flag set and spike readout (see Fig. 2). This difference leads to quantization noise. Third, the thermal noise introduced by circuits brings randomness to spike fire. For the above reasons, the intervals of neighboring spikes exhibit fluctuations and struggle to accurately reflect light intensity, leading to noisy reconstruction results from SIR." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 71, + 434, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 71, + 434, + 83 + ], + "spans": [ + { + "bbox": [ + 314, + 71, + 434, + 83 + ], + "type": "text", + "content": "4. Motivation and Idea" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 91, + 553, + 235 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 91, + 553, + 235 + ], + "spans": [ + { + "bbox": [ + 313, + 91, + 553, + 235 + ], + "type": "text", + "content": "The key problem we aim to address is how to super-resolve the original dynamic scene from captured binary spike data to obtain high quality visual reconstruction with clean image details. According to the working mechanism of spike camera, each individual spike does not reveal the light intensity directly. Instead, it only tells the time point at which a spike firing occurs. Essentially, the instantaneous light intensity can be inferred only by considering the length of intervals between consecutive spikes. Unfortunately, for the reasons discussed above, the recorded spike intervals are not stable, and they fluctuate around an expectation value determined by actual light intensity." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 235, + 553, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 235, + 553, + 342 + ], + "spans": [ + { + "bbox": [ + 313, + 235, + 553, + 342 + ], + "type": "text", + "content": "To handle such fluctuations, we draw inspiration from several classic image restoration methods, such as BM3D [6], BAS [43], LSSC [27], and WNNM [18]. Although they are quite diversified in algorithmic design, all these methods incorporate the idea of leveraging non-local similarity within natural image signals to reduce the unknown random disturbance. They utilize a group of similar patches and establish a prior model for them so that a regularization term is formed to separate the true signal from the noise." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 342, + 553, + 377 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 342, + 553, + 377 + ], + "spans": [ + { + "bbox": [ + 313, + 342, + 553, + 377 + ], + "type": "text", + "content": "Suppose " + }, + { + "bbox": [ + 313, + 342, + 553, + 377 + ], + "type": "inline_equation", + "content": "\\mathbf{y} = \\mathbf{x} + \\mathbf{n}" + }, + { + "bbox": [ + 313, + 342, + 553, + 377 + ], + "type": "text", + "content": " is an image corrupted by noise, where " + }, + { + "bbox": [ + 313, + 342, + 553, + 377 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 313, + 342, + 553, + 377 + ], + "type": "text", + "content": " is the clean image and " + }, + { + "bbox": [ + 313, + 342, + 553, + 377 + ], + "type": "inline_equation", + "content": "\\mathbf{n}" + }, + { + "bbox": [ + 313, + 342, + 553, + 377 + ], + "type": "text", + "content": " is the noise. The idea of these methods can be summarized as follows:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 378, + 553, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 378, + 553, + 403 + ], + "spans": [ + { + "bbox": [ + 313, + 378, + 553, + 403 + ], + "type": "text", + "content": "(1) For each pixel location " + }, + { + "bbox": [ + 313, + 378, + 553, + 403 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 313, + 378, + 553, + 403 + ], + "type": "text", + "content": ", search for a group of similar patches from the noisy image " + }, + { + "bbox": [ + 313, + 378, + 553, + 403 + ], + "type": "inline_equation", + "content": "\\mathbf{y}" + }, + { + "bbox": [ + 313, + 378, + 553, + 403 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 389, + 411, + 553, + 425 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 389, + 411, + 553, + 425 + ], + "spans": [ + { + "bbox": [ + 389, + 411, + 553, + 425 + ], + "type": "interline_equation", + "content": "\\mathbf {y} _ {1}, \\mathbf {y} _ {2}, \\dots , \\mathbf {y} _ {N} = \\Lambda (\\mathbf {y} | \\mathbf {p}), \\tag {4}", + "image_path": "37f8bffc631ad459638651d9098497ac2f45c3023db7197cb564ff37cbd84d7e.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 329, + 432, + 553, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 329, + 432, + 553, + 479 + ], + "spans": [ + { + "bbox": [ + 329, + 432, + 553, + 479 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 329, + 432, + 553, + 479 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 329, + 432, + 553, + 479 + ], + "type": "text", + "content": " is the number of similar patches, " + }, + { + "bbox": [ + 329, + 432, + 553, + 479 + ], + "type": "inline_equation", + "content": "\\Lambda (\\cdot |\\cdot)" + }, + { + "bbox": [ + 329, + 432, + 553, + 479 + ], + "type": "text", + "content": " is the operation of searching for and extracting similar patches. Note that the notation " + }, + { + "bbox": [ + 329, + 432, + 553, + 479 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_i" + }, + { + "bbox": [ + 329, + 432, + 553, + 479 + ], + "type": "text", + "content": " should depend on " + }, + { + "bbox": [ + 329, + 432, + 553, + 479 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 329, + 432, + 553, + 479 + ], + "type": "text", + "content": " but we omit it for conciseness." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 314, + 479, + 553, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 479, + 553, + 504 + ], + "spans": [ + { + "bbox": [ + 314, + 479, + 553, + 504 + ], + "type": "text", + "content": "(2) Exploit the correlation within the patch group " + }, + { + "bbox": [ + 314, + 479, + 553, + 504 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{y}_i\\}_{i = 1}^N" + }, + { + "bbox": [ + 314, + 479, + 553, + 504 + ], + "type": "text", + "content": " to generate an estimate for each patch:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 380, + 511, + 553, + 526 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 380, + 511, + 553, + 526 + ], + "spans": [ + { + "bbox": [ + 380, + 511, + 553, + 526 + ], + "type": "interline_equation", + "content": "\\tilde {\\mathbf {x}} _ {1}, \\tilde {\\mathbf {x}} _ {2}, \\dots , \\tilde {\\mathbf {x}} _ {N} = \\Pi \\left(\\left\\{\\mathbf {y} _ {i} \\right\\} _ {i = 1} ^ {N}\\right), \\tag {5}", + "image_path": "1f29173e66f00771acd852d1a87a3d7deab4a327a3c7df7bdedd16a87ba3b2cb.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 329, + 533, + 553, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 329, + 533, + 553, + 605 + ], + "spans": [ + { + "bbox": [ + 329, + 533, + 553, + 605 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 329, + 533, + 553, + 605 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{x}}_i" + }, + { + "bbox": [ + 329, + 533, + 553, + 605 + ], + "type": "text", + "content": " is the estimate of the true signal " + }, + { + "bbox": [ + 329, + 533, + 553, + 605 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i" + }, + { + "bbox": [ + 329, + 533, + 553, + 605 + ], + "type": "text", + "content": " corresponding to " + }, + { + "bbox": [ + 329, + 533, + 553, + 605 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_i" + }, + { + "bbox": [ + 329, + 533, + 553, + 605 + ], + "type": "text", + "content": ". The actual operation of " + }, + { + "bbox": [ + 329, + 533, + 553, + 605 + ], + "type": "inline_equation", + "content": "\\Pi(\\cdot)" + }, + { + "bbox": [ + 329, + 533, + 553, + 605 + ], + "type": "text", + "content": " depends on the prior model adopted to characterize the consistency between patches " + }, + { + "bbox": [ + 329, + 533, + 553, + 605 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{y}_i\\}_{i=1}^N" + }, + { + "bbox": [ + 329, + 533, + 553, + 605 + ], + "type": "text", + "content": ". Generally, " + }, + { + "bbox": [ + 329, + 533, + 553, + 605 + ], + "type": "inline_equation", + "content": "\\Pi(\\cdot)" + }, + { + "bbox": [ + 329, + 533, + 553, + 605 + ], + "type": "text", + "content": " can be regarded as an operator to solve the following optimization objective:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 339, + 613, + 553, + 647 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 613, + 553, + 647 + ], + "spans": [ + { + "bbox": [ + 339, + 613, + 553, + 647 + ], + "type": "interline_equation", + "content": "\\left\\{\\tilde {\\mathbf {x}} _ {i} \\right\\} = \\arg \\min \\sum_ {i = 1} ^ {N} \\left| \\left| \\mathbf {x} _ {i} - \\mathbf {y} _ {i} \\right| \\right| _ {2} ^ {2} + \\Phi \\left(\\left\\{\\mathbf {x} _ {i} \\right\\} _ {i = 1} ^ {N}\\right) \\tag {6}", + "image_path": "9b4c4cae1a3647c80ab09828efb33d81ae6956b9df1f72af2577f3ffa7754345.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 329, + 653, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 329, + 653, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 329, + 653, + 553, + 714 + ], + "type": "text", + "content": "BM3D [6] solves Eq. (6) by applying a 3D transform to the stacked similar patches and constructing a sparsity regularization term in the transform domain. It utilizes hard-thresholding, soft-thresholding, and Wiener filtering in different steps, corresponding to " + }, + { + "bbox": [ + 329, + 653, + 553, + 714 + ], + "type": "inline_equation", + "content": "l_{0}" + }, + { + "bbox": [ + 329, + 653, + 553, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 329, + 653, + 553, + 714 + ], + "type": "inline_equation", + "content": "l_{1}" + }, + { + "bbox": [ + 329, + 653, + 553, + 714 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 329, + 653, + 553, + 714 + ], + "type": "inline_equation", + "content": "l_{2}" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11418" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 70, + 553, + 149 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 553, + 149 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 553, + 149 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 553, + 149 + ], + "type": "image", + "image_path": "82f650853a543bc0384bc7025ac8ff79d805a954f33dbf8e841bc620bf268102.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 158, + 555, + 204 + ], + "lines": [ + { + "bbox": [ + 54, + 158, + 555, + 204 + ], + "spans": [ + { + "bbox": [ + 54, + 158, + 555, + 204 + ], + "type": "text", + "content": "Figure 3. The main idea of motion aligned collaborative filtering. First, we collect temporal similar features via motion alignment. Then, we stack these aligned features and perform transform-domain collaborative filtering to generate refinement features for each moment. Subsequently, the refinement features are unstacked and mapped back to their original positions via inverse warping. These repositioned refinement features are aggregated with input features to obtain refined features with improved consistency." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 70, + 210, + 296, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 210, + 296, + 388 + ], + "spans": [ + { + "bbox": [ + 70, + 210, + 296, + 388 + ], + "type": "text", + "content": "sparsity norm, respectively. BAS [43] applies a content adaptive 2D decorrelation transform to the similar patches via principle component analysis (PCA), and builds a separate distribution model and regularization term for each transform band, using the transform coefficients of these patches as samples. It leads to a PCA-domain adaptive soft-thresholding to solve Eq. (6). LLSC [27] adopts another prior model that emphasizes that similar patches tend to share the location pattern of non-zero dictionary elements in their sparse decomposition. WNNM [18], on the other hand, utilizes the low-rank property of similar patches so that the problem (6) reduces to singular value decomposition (SVD) and soft-thresholding on the singular values." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 389, + 295, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 389, + 295, + 425 + ], + "spans": [ + { + "bbox": [ + 55, + 389, + 295, + 425 + ], + "type": "text", + "content": "(3) Place each estimate " + }, + { + "bbox": [ + 55, + 389, + 295, + 425 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{x}}_i" + }, + { + "bbox": [ + 55, + 389, + 295, + 425 + ], + "type": "text", + "content": " back to the position where " + }, + { + "bbox": [ + 55, + 389, + 295, + 425 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_i" + }, + { + "bbox": [ + 55, + 389, + 295, + 425 + ], + "type": "text", + "content": " is taken from, and aggregate the overlapping estimates from all position " + }, + { + "bbox": [ + 55, + 389, + 295, + 425 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 55, + 389, + 295, + 425 + ], + "type": "text", + "content": " to form a whole estimate for " + }, + { + "bbox": [ + 55, + 389, + 295, + 425 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 55, + 389, + 295, + 425 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 120, + 430, + 295, + 444 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 120, + 430, + 295, + 444 + ], + "spans": [ + { + "bbox": [ + 120, + 430, + 295, + 444 + ], + "type": "interline_equation", + "content": "\\tilde {\\mathbf {x}} = \\Lambda^ {- 1} \\left(\\left\\{\\tilde {\\mathbf {x}} _ {1}, \\tilde {\\mathbf {x}} _ {2}, \\dots , \\tilde {\\mathbf {x}} _ {N} \\right\\} _ {\\mathbf {p}}\\right). \\tag {7}", + "image_path": "41164001df11b783b52e7f3f5d989383953428f4c23c22f736bd739cf6f2ec5b.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 70, + 450, + 295, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 450, + 295, + 486 + ], + "spans": [ + { + "bbox": [ + 70, + 450, + 295, + 486 + ], + "type": "text", + "content": "Here " + }, + { + "bbox": [ + 70, + 450, + 295, + 486 + ], + "type": "inline_equation", + "content": "\\Lambda^{-1}(\\cdot)" + }, + { + "bbox": [ + 70, + 450, + 295, + 486 + ], + "type": "text", + "content": " denotes patch reposition and aggregation operation. The reposition of " + }, + { + "bbox": [ + 70, + 450, + 295, + 486 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{x}}_i" + }, + { + "bbox": [ + 70, + 450, + 295, + 486 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 70, + 450, + 295, + 486 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{x}}" + }, + { + "bbox": [ + 70, + 450, + 295, + 486 + ], + "type": "text", + "content": " is the inverse operation of taking " + }, + { + "bbox": [ + 70, + 450, + 295, + 486 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_i" + }, + { + "bbox": [ + 70, + 450, + 295, + 486 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 70, + 450, + 295, + 486 + ], + "type": "inline_equation", + "content": "\\mathbf{y}" + }, + { + "bbox": [ + 70, + 450, + 295, + 486 + ], + "type": "text", + "content": " in stage (1)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 487, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 487, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 487, + 296, + 714 + ], + "type": "text", + "content": "Taking these ideas into account, we propose a novel framework tailored for spike data to reduce disturbances in spike information. The main idea is illustrated in Fig. 3. The similarity exploited in this framework is temporal similarity, and it may be extended to include spatial non-local similarity in the future. First, we align the spike features at neighboring moments with the spike feature at the reconstruction moment to obtain a set of similar features. Then, we stack the aligned features and perform a neural-network-based transform-domain collaborative filtering, to separate the disturbances from clean features by exploiting representation sparsity and generate refinement features for each moment. These refinement features are used to update the initial input features. Specifically, the refinement features are unstacked and mapped back to their original positions via inverse warping, and then aggregated with input features. The purpose of this framework is to improve the consistency of multi-moment features along the motion trajectory." + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 314, + 208, + 555, + 295 + ], + "blocks": [ + { + "bbox": [ + 314, + 208, + 555, + 295 + ], + "lines": [ + { + "bbox": [ + 314, + 208, + 555, + 295 + ], + "spans": [ + { + "bbox": [ + 314, + 208, + 555, + 295 + ], + "type": "image", + "image_path": "2060476049fe074415973e3ab9de0778e54ed28cf5da0f89029d87f22b26f2f4.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 303, + 555, + 326 + ], + "lines": [ + { + "bbox": [ + 313, + 303, + 555, + 326 + ], + "spans": [ + { + "bbox": [ + 313, + 303, + 555, + 326 + ], + "type": "text", + "content": "Figure 4. Overall architecture of the proposed spike camera super resolution network Spk2SRImgNet." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 338, + 370, + 350 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 338, + 370, + 350 + ], + "spans": [ + { + "bbox": [ + 314, + 338, + 370, + 350 + ], + "type": "text", + "content": "5. Method" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 361, + 434, + 373 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 361, + 434, + 373 + ], + "spans": [ + { + "bbox": [ + 313, + 361, + 434, + 373 + ], + "type": "text", + "content": "5.1. Overall Architecture" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 383, + 555, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 383, + 555, + 479 + ], + "spans": [ + { + "bbox": [ + 313, + 383, + 555, + 479 + ], + "type": "text", + "content": "To reconstruct HR dynamic scenes from LR binary spike stream, we develop an end-to-end trainable spike camera super resolution network, Spk2SRImgNet. The overall architecture is shown in Fig. 4. To reconstruct an HR image " + }, + { + "bbox": [ + 313, + 383, + 555, + 479 + ], + "type": "inline_equation", + "content": "\\mathcal{I}(t_0) \\in \\mathbb{R}^{rH \\times rW}" + }, + { + "bbox": [ + 313, + 383, + 555, + 479 + ], + "type": "text", + "content": " at time " + }, + { + "bbox": [ + 313, + 383, + 555, + 479 + ], + "type": "inline_equation", + "content": "t_0" + }, + { + "bbox": [ + 313, + 383, + 555, + 479 + ], + "type": "text", + "content": ", we take a set of consecutive spike frames " + }, + { + "bbox": [ + 313, + 383, + 555, + 479 + ], + "type": "inline_equation", + "content": "S(t_0 + i), i = 0, \\pm 1, \\pm 2, \\ldots" + }, + { + "bbox": [ + 313, + 383, + 555, + 479 + ], + "type": "text", + "content": " around " + }, + { + "bbox": [ + 313, + 383, + 555, + 479 + ], + "type": "inline_equation", + "content": "S(t_0)" + }, + { + "bbox": [ + 313, + 383, + 555, + 479 + ], + "type": "text", + "content": " as input, where " + }, + { + "bbox": [ + 313, + 383, + 555, + 479 + ], + "type": "inline_equation", + "content": "S(t_0 + i) \\in \\mathbb{B}^{H \\times W \\times 1}" + }, + { + "bbox": [ + 313, + 383, + 555, + 479 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 383, + 555, + 479 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 313, + 383, + 555, + 479 + ], + "type": "text", + "content": " is SR scale, " + }, + { + "bbox": [ + 313, + 383, + 555, + 479 + ], + "type": "inline_equation", + "content": "\\mathbb{B}" + }, + { + "bbox": [ + 313, + 383, + 555, + 479 + ], + "type": "text", + "content": " represents the binary domain." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 483, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 483, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 483, + 556, + 714 + ], + "type": "text", + "content": "Firstly, the input spike stream is partitioned into " + }, + { + "bbox": [ + 313, + 483, + 556, + 714 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 483, + 556, + 714 + ], + "type": "text", + "content": " overlapping short-term spike blocks " + }, + { + "bbox": [ + 313, + 483, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\{B_i\\}_{i = -n}^n (K = 2n + 1)" + }, + { + "bbox": [ + 313, + 483, + 556, + 714 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 483, + 556, + 714 + ], + "type": "inline_equation", + "content": "B_{i}\\in \\mathbb{B}^{H\\times W\\times L}" + }, + { + "bbox": [ + 313, + 483, + 556, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 483, + 556, + 714 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 313, + 483, + 556, + 714 + ], + "type": "text", + "content": " is the temporal length of each block, " + }, + { + "bbox": [ + 313, + 483, + 556, + 714 + ], + "type": "inline_equation", + "content": "B_{0}" + }, + { + "bbox": [ + 313, + 483, + 556, + 714 + ], + "type": "text", + "content": " is the key block centered at time " + }, + { + "bbox": [ + 313, + 483, + 556, + 714 + ], + "type": "inline_equation", + "content": "t_0" + }, + { + "bbox": [ + 313, + 483, + 556, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 483, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\{B_i\\}, i = \\pm 1,\\pm 2,\\ldots" + }, + { + "bbox": [ + 313, + 483, + 556, + 714 + ], + "type": "text", + "content": " are neighboring blocks distributed symmetrically on both sides of " + }, + { + "bbox": [ + 313, + 483, + 556, + 714 + ], + "type": "inline_equation", + "content": "B_{0}" + }, + { + "bbox": [ + 313, + 483, + 556, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 483, + 556, + 714 + ], + "type": "inline_equation", + "content": "B_{i}" + }, + { + "bbox": [ + 313, + 483, + 556, + 714 + ], + "type": "text", + "content": " is centered at time " + }, + { + "bbox": [ + 313, + 483, + 556, + 714 + ], + "type": "inline_equation", + "content": "t_i" + }, + { + "bbox": [ + 313, + 483, + 556, + 714 + ], + "type": "text", + "content": ". To bridge the domain gap between binary spikes and real pixel values, we first convert spike frames in " + }, + { + "bbox": [ + 313, + 483, + 556, + 714 + ], + "type": "inline_equation", + "content": "B_{i} = \\{S(u)\\}_{u = t_{i} - w_{h}}^{t_{i} + w_{h}}" + }, + { + "bbox": [ + 313, + 483, + 556, + 714 + ], + "type": "text", + "content": " to intensity frames in " + }, + { + "bbox": [ + 313, + 483, + 556, + 714 + ], + "type": "inline_equation", + "content": "Y_{i} = \\{Y(u)\\}_{u = t_{i} - w_{h}}^{t_{i} + w_{h}}\\in \\mathbb{R}^{H\\times W\\times L}" + }, + { + "bbox": [ + 313, + 483, + 556, + 714 + ], + "type": "text", + "content": " based on the spike interval based reconstruction (SIR) method [53], where " + }, + { + "bbox": [ + 313, + 483, + 556, + 714 + ], + "type": "inline_equation", + "content": "L = 2w_{h} + 1" + }, + { + "bbox": [ + 313, + 483, + 556, + 714 + ], + "type": "text", + "content": ". Then, we extract features " + }, + { + "bbox": [ + 313, + 483, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\{F_i\\}_{i = -n}^n" + }, + { + "bbox": [ + 313, + 483, + 556, + 714 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 313, + 483, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\{Y_i\\}_{i = -n}^n" + }, + { + "bbox": [ + 313, + 483, + 556, + 714 + ], + "type": "text", + "content": " through an encoder, which consists of two convolutional layers and three residual blocks. To reduce fluctuations, " + }, + { + "bbox": [ + 313, + 483, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\{F_i\\}_{i = -n}^n" + }, + { + "bbox": [ + 313, + 483, + 556, + 714 + ], + "type": "text", + "content": " are input to motion aligned collaborative filtering (MACF) module, which generates improved features " + }, + { + "bbox": [ + 313, + 483, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\{F_i\\}_{i = -n}^n" + }, + { + "bbox": [ + 313, + 483, + 556, + 714 + ], + "type": "text", + "content": ". Finally, the multi-moment features " + }, + { + "bbox": [ + 313, + 483, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\{\\hat{F}_i\\}_{i = -n}^n" + }, + { + "bbox": [ + 313, + 483, + 556, + 714 + ], + "type": "text", + "content": " are fed into motion aligned super resolved reconstruction module to generate an HR image " + }, + { + "bbox": [ + 313, + 483, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{I}(t_0)" + }, + { + "bbox": [ + 313, + 483, + 556, + 714 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11419" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 69, + 553, + 239 + ], + "blocks": [ + { + "bbox": [ + 56, + 69, + 553, + 239 + ], + "lines": [ + { + "bbox": [ + 56, + 69, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 56, + 69, + 553, + 239 + ], + "type": "image", + "image_path": "3318ce1e321fdff7b9477e3808f00cc75f152ab598022ff514495d56f02ecaa8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 138, + 247, + 470, + 258 + ], + "lines": [ + { + "bbox": [ + 138, + 247, + 470, + 258 + ], + "spans": [ + { + "bbox": [ + 138, + 247, + 470, + 258 + ], + "type": "text", + "content": "Figure 5. The overall framework of motion aligned collaborative filtering (MACF) module." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 56, + 263, + 295, + 340 + ], + "blocks": [ + { + "bbox": [ + 56, + 263, + 295, + 340 + ], + "lines": [ + { + "bbox": [ + 56, + 263, + 295, + 340 + ], + "spans": [ + { + "bbox": [ + 56, + 263, + 295, + 340 + ], + "type": "image", + "image_path": "b7624ecf54f11fd719d43f561df9dee72268361eb7507230fd596663c32dfe1f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 348, + 295, + 403 + ], + "lines": [ + { + "bbox": [ + 55, + 348, + 295, + 403 + ], + "spans": [ + { + "bbox": [ + 55, + 348, + 295, + 403 + ], + "type": "text", + "content": "Figure 6. Illustration of channel rearrangement (CR) and inverse channel rearrangement (ICR). The left part is " + }, + { + "bbox": [ + 55, + 348, + 295, + 403 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 348, + 295, + 403 + ], + "type": "text", + "content": " features " + }, + { + "bbox": [ + 55, + 348, + 295, + 403 + ], + "type": "inline_equation", + "content": "\\{F_i\\}_{i = -n}^n" + }, + { + "bbox": [ + 55, + 348, + 295, + 403 + ], + "type": "text", + "content": " at " + }, + { + "bbox": [ + 55, + 348, + 295, + 403 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 348, + 295, + 403 + ], + "type": "text", + "content": " different moments, where " + }, + { + "bbox": [ + 55, + 348, + 295, + 403 + ], + "type": "inline_equation", + "content": "F_{i}\\in \\mathbb{R}^{H\\times W\\times C}" + }, + { + "bbox": [ + 55, + 348, + 295, + 403 + ], + "type": "text", + "content": ". The right part is " + }, + { + "bbox": [ + 55, + 348, + 295, + 403 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 55, + 348, + 295, + 403 + ], + "type": "text", + "content": " features " + }, + { + "bbox": [ + 55, + 348, + 295, + 403 + ], + "type": "inline_equation", + "content": "\\{G_j\\}_{j = 0}^{C - 1}" + }, + { + "bbox": [ + 55, + 348, + 295, + 403 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 348, + 295, + 403 + ], + "type": "inline_equation", + "content": "G_{j}\\in \\mathbb{R}^{H\\times W\\times K}" + }, + { + "bbox": [ + 55, + 348, + 295, + 403 + ], + "type": "text", + "content": ", each " + }, + { + "bbox": [ + 55, + 348, + 295, + 403 + ], + "type": "inline_equation", + "content": "G_{j}" + }, + { + "bbox": [ + 55, + 348, + 295, + 403 + ], + "type": "text", + "content": " contains features at " + }, + { + "bbox": [ + 55, + 348, + 295, + 403 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 348, + 295, + 403 + ], + "type": "text", + "content": " different moments." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 409, + 262, + 422 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 409, + 262, + 422 + ], + "spans": [ + { + "bbox": [ + 55, + 409, + 262, + 422 + ], + "type": "text", + "content": "5.2. Motion Aligned Collaborative Filtering" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 426, + 295, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 426, + 295, + 485 + ], + "spans": [ + { + "bbox": [ + 55, + 426, + 295, + 485 + ], + "type": "text", + "content": "We propose a motion aligned collaborative filtering (MACF) module to implement the idea proposed in Sec. 4. The overall framework of MACF is shown in Fig. 5, which includes six steps. We provide a detailed description of each step below." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 486, + 295, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 486, + 295, + 582 + ], + "spans": [ + { + "bbox": [ + 55, + 486, + 295, + 582 + ], + "type": "text", + "content": "Step 1: Offset Estimation. This step generates the forward " + }, + { + "bbox": [ + 55, + 486, + 295, + 582 + ], + "type": "inline_equation", + "content": "(F_{i}\\rightarrow F_{0})" + }, + { + "bbox": [ + 55, + 486, + 295, + 582 + ], + "type": "text", + "content": " and backward " + }, + { + "bbox": [ + 55, + 486, + 295, + 582 + ], + "type": "inline_equation", + "content": "(F_0\\to F_i)" + }, + { + "bbox": [ + 55, + 486, + 295, + 582 + ], + "type": "text", + "content": " motion offsets between neighboring features " + }, + { + "bbox": [ + 55, + 486, + 295, + 582 + ], + "type": "inline_equation", + "content": "\\{F_i\\} ,i = \\pm 1,\\dots ,\\pm n" + }, + { + "bbox": [ + 55, + 486, + 295, + 582 + ], + "type": "text", + "content": " and key feature " + }, + { + "bbox": [ + 55, + 486, + 295, + 582 + ], + "type": "inline_equation", + "content": "F_{0}" + }, + { + "bbox": [ + 55, + 486, + 295, + 582 + ], + "type": "text", + "content": ". These offsets will be utilized in the subsequent feature warping (step 3) and inverse feature warping (step 5), respectively. Specifically, the forward and backward motion offsets: " + }, + { + "bbox": [ + 55, + 486, + 295, + 582 + ], + "type": "inline_equation", + "content": "\\mathcal{O}_i^{\\mathbf{F}}" + }, + { + "bbox": [ + 55, + 486, + 295, + 582 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 486, + 295, + 582 + ], + "type": "inline_equation", + "content": "\\mathcal{O}_i^{\\mathbf{B}}" + }, + { + "bbox": [ + 55, + 486, + 295, + 582 + ], + "type": "text", + "content": " are both derived from " + }, + { + "bbox": [ + 55, + 486, + 295, + 582 + ], + "type": "inline_equation", + "content": "F_{i}" + }, + { + "bbox": [ + 55, + 486, + 295, + 582 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 486, + 295, + 582 + ], + "type": "inline_equation", + "content": "F_{0}" + }, + { + "bbox": [ + 55, + 486, + 295, + 582 + ], + "type": "text", + "content": ", which are formulated as follows (" + }, + { + "bbox": [ + 55, + 486, + 295, + 582 + ], + "type": "inline_equation", + "content": "\\forall i\\in \\{\\pm 1,\\dots ,\\pm n\\}" + }, + { + "bbox": [ + 55, + 486, + 295, + 582 + ], + "type": "text", + "content": "):" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 130, + 589, + 294, + 620 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 589, + 294, + 620 + ], + "spans": [ + { + "bbox": [ + 130, + 589, + 294, + 620 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {O} _ {i} ^ {\\mathbf {F}} = m _ {1} \\left(\\left\\{F _ {i}, F _ {0} \\right\\}\\right), \\\\ \\mathcal {O} _ {i} ^ {\\mathbf {B}} = m _ {2} \\left(\\left\\{F _ {i}, F _ {0} \\right\\}\\right), \\tag {8} \\\\ \\end{array}", + "image_path": "1b7eacc313118758539095df35305a2fa731ae476fa14c6936d603a74d4290b9.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 628, + 295, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 628, + 295, + 663 + ], + "spans": [ + { + "bbox": [ + 55, + 628, + 295, + 663 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 628, + 295, + 663 + ], + "type": "inline_equation", + "content": "m_{1}(\\cdot)" + }, + { + "bbox": [ + 55, + 628, + 295, + 663 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 628, + 295, + 663 + ], + "type": "inline_equation", + "content": "m_{2}(\\cdot)" + }, + { + "bbox": [ + 55, + 628, + 295, + 663 + ], + "type": "text", + "content": " denote two independent convolutional blocks that consist of several convolutional layers, " + }, + { + "bbox": [ + 55, + 628, + 295, + 663 + ], + "type": "inline_equation", + "content": "\\{,\\}" + }, + { + "bbox": [ + 55, + 628, + 295, + 663 + ], + "type": "text", + "content": " denotes channel concatenation operation." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 663, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 663, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 663, + 296, + 714 + ], + "type": "text", + "content": "Step 2: Channel Rearrange. The " + }, + { + "bbox": [ + 55, + 663, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\{F_i\\}_{i = -n}^n" + }, + { + "bbox": [ + 55, + 663, + 296, + 714 + ], + "type": "text", + "content": " at " + }, + { + "bbox": [ + 55, + 663, + 296, + 714 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 663, + 296, + 714 + ], + "type": "text", + "content": " different moments are features extracted from the fluctuated reconstruction " + }, + { + "bbox": [ + 55, + 663, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\{Y_{i}\\}_{i = -n}^{n}" + }, + { + "bbox": [ + 55, + 663, + 296, + 714 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 663, + 296, + 714 + ], + "type": "inline_equation", + "content": "F_{i}\\in \\mathbb{R}^{H\\times W\\times C}" + }, + { + "bbox": [ + 55, + 663, + 296, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 663, + 296, + 714 + ], + "type": "inline_equation", + "content": "F_{i} = \\{F_{i}^{0},F_{i}^{1},\\dots,F_{i}^{C - 1}\\}" + }, + { + "bbox": [ + 55, + 663, + 296, + 714 + ], + "type": "text", + "content": ", the superscript of " + }, + { + "bbox": [ + 55, + 663, + 296, + 714 + ], + "type": "inline_equation", + "content": "F_{i}" + }, + { + "bbox": [ + 55, + 663, + 296, + 714 + ], + "type": "text", + "content": " denotes channel" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 265, + 555, + 524 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 265, + 555, + 524 + ], + "spans": [ + { + "bbox": [ + 313, + 265, + 555, + 524 + ], + "type": "text", + "content": "index. Each feature " + }, + { + "bbox": [ + 313, + 265, + 555, + 524 + ], + "type": "inline_equation", + "content": "F_{i}" + }, + { + "bbox": [ + 313, + 265, + 555, + 524 + ], + "type": "text", + "content": " consists of " + }, + { + "bbox": [ + 313, + 265, + 555, + 524 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 313, + 265, + 555, + 524 + ], + "type": "text", + "content": " channels, with each channel encoding different information about the scene. For example, some channels capture edges, some channels focus on textures, and so on. We perform channel rearrangement (CR) to collect data at the same channel index for each feature and concatenate them, as shown in Fig. 6. We form " + }, + { + "bbox": [ + 313, + 265, + 555, + 524 + ], + "type": "inline_equation", + "content": "\\{G_j\\}_{j = 0}^{C - 1}" + }, + { + "bbox": [ + 313, + 265, + 555, + 524 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 313, + 265, + 555, + 524 + ], + "type": "inline_equation", + "content": "\\{F_i\\}_{i = -n}^n" + }, + { + "bbox": [ + 313, + 265, + 555, + 524 + ], + "type": "text", + "content": " via CR operation, where " + }, + { + "bbox": [ + 313, + 265, + 555, + 524 + ], + "type": "inline_equation", + "content": "G_{j}\\in \\mathbb{R}^{H\\times W\\times K}" + }, + { + "bbox": [ + 313, + 265, + 555, + 524 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 265, + 555, + 524 + ], + "type": "inline_equation", + "content": "G_{j} = \\{G_{j}^{-n},\\dots,G_{j}^{0},\\dots,G_{j}^{n}\\} = \\{F_{-n}^{j},\\dots,F_{0}^{j},\\dots,F_{n}^{j}\\}" + }, + { + "bbox": [ + 313, + 265, + 555, + 524 + ], + "type": "text", + "content": ", the subscript of " + }, + { + "bbox": [ + 313, + 265, + 555, + 524 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 313, + 265, + 555, + 524 + ], + "type": "text", + "content": " denotes channel index, the superscript of " + }, + { + "bbox": [ + 313, + 265, + 555, + 524 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 313, + 265, + 555, + 524 + ], + "type": "text", + "content": " denotes time index. " + }, + { + "bbox": [ + 313, + 265, + 555, + 524 + ], + "type": "inline_equation", + "content": "G_{j}" + }, + { + "bbox": [ + 313, + 265, + 555, + 524 + ], + "type": "text", + "content": " contains " + }, + { + "bbox": [ + 313, + 265, + 555, + 524 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 265, + 555, + 524 + ], + "type": "text", + "content": " features with channel index " + }, + { + "bbox": [ + 313, + 265, + 555, + 524 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 265, + 555, + 524 + ], + "type": "text", + "content": " at " + }, + { + "bbox": [ + 313, + 265, + 555, + 524 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 265, + 555, + 524 + ], + "type": "text", + "content": " different moments. To offer an intuitive understanding of this operation, we provide a visualization of CR in supplementary material. After CR, steps 3, 4 and 5 are performed sequentially for each " + }, + { + "bbox": [ + 313, + 265, + 555, + 524 + ], + "type": "inline_equation", + "content": "G_{j}" + }, + { + "bbox": [ + 313, + 265, + 555, + 524 + ], + "type": "text", + "content": ". Step 3: Feature Warping. Step 2 collects temporally unaligned features that encode the same information. We perform feature warping based on deformable convolution " + }, + { + "bbox": [ + 313, + 265, + 555, + 524 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 313, + 265, + 555, + 524 + ], + "type": "text", + "content": " to align neighboring features " + }, + { + "bbox": [ + 313, + 265, + 555, + 524 + ], + "type": "inline_equation", + "content": "\\{G_j^i\\}" + }, + { + "bbox": [ + 313, + 265, + 555, + 524 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 265, + 555, + 524 + ], + "type": "inline_equation", + "content": "i = \\pm 1,\\ldots ,\\pm n" + }, + { + "bbox": [ + 313, + 265, + 555, + 524 + ], + "type": "text", + "content": " with key feature " + }, + { + "bbox": [ + 313, + 265, + 555, + 524 + ], + "type": "inline_equation", + "content": "G_{j}^{0}" + }, + { + "bbox": [ + 313, + 265, + 555, + 524 + ], + "type": "text", + "content": ". Specifically, given the forward offsets from neighboring features to key feature " + }, + { + "bbox": [ + 313, + 265, + 555, + 524 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{O}_i^{\\mathbf{F}}\\}" + }, + { + "bbox": [ + 313, + 265, + 555, + 524 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 265, + 555, + 524 + ], + "type": "inline_equation", + "content": "i = \\pm 1,\\ldots ,\\pm n" + }, + { + "bbox": [ + 313, + 265, + 555, + 524 + ], + "type": "text", + "content": ", the aligned features could be obtained by:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 319, + 529, + 555, + 545 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 529, + 555, + 545 + ], + "spans": [ + { + "bbox": [ + 319, + 529, + 555, + 545 + ], + "type": "interline_equation", + "content": "\\tilde {G} _ {j} ^ {i} = \\mathcal {D} \\left(G _ {j} ^ {i}, \\mathcal {O} _ {i} ^ {\\mathbf {F}}\\right) = \\mathcal {D} \\left(F _ {i} ^ {j}, \\mathcal {O} _ {i} ^ {\\mathbf {F}}\\right), \\forall i \\in \\{\\pm 1, \\dots , \\pm n \\}. \\tag {9}", + "image_path": "2b3f12b72ded622c9764d0e837bc48651b7274c64c98d59c8923a64ba09146cd.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 552, + 554, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 552, + 554, + 590 + ], + "spans": [ + { + "bbox": [ + 313, + 552, + 554, + 590 + ], + "type": "text", + "content": "These aligned features " + }, + { + "bbox": [ + 313, + 552, + 554, + 590 + ], + "type": "inline_equation", + "content": "\\{\\tilde{G}_j^i\\}_{i = -n}^n" + }, + { + "bbox": [ + 313, + 552, + 554, + 590 + ], + "type": "text", + "content": " are then concatenated to derive " + }, + { + "bbox": [ + 313, + 552, + 554, + 590 + ], + "type": "inline_equation", + "content": "\\tilde{G}_j = \\{\\tilde{G}_j^{-n},\\dots,\\tilde{G}_j^n\\}" + }, + { + "bbox": [ + 313, + 552, + 554, + 590 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 552, + 554, + 590 + ], + "type": "inline_equation", + "content": "\\tilde{G}_j^0 = G_j^0 = F_0^j" + }, + { + "bbox": [ + 313, + 552, + 554, + 590 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 313, + 552, + 554, + 590 + ], + "type": "inline_equation", + "content": "\\tilde{G}_j" + }, + { + "bbox": [ + 313, + 552, + 554, + 590 + ], + "type": "text", + "content": " could be regarded as stacked similar features." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 590, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 590, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 590, + 555, + 714 + ], + "type": "text", + "content": "Step 4: Transform Domain Stacked Feature Filtering. We design a Transform Domain stacked feature Filtering (TDF) step to generate refinement features " + }, + { + "bbox": [ + 313, + 590, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\Delta \\tilde{G}_j" + }, + { + "bbox": [ + 313, + 590, + 555, + 714 + ], + "type": "text", + "content": " from stacked similar features " + }, + { + "bbox": [ + 313, + 590, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\tilde{G}_j" + }, + { + "bbox": [ + 313, + 590, + 555, + 714 + ], + "type": "text", + "content": " by leveraging representation sparsity in transform domain. TDF is designed with four submodules: transform, noise estimation, filtering, and inverse transform, as depicted in step 4 of Fig. 5. Specifically, the input " + }, + { + "bbox": [ + 313, + 590, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\tilde{G}_j" + }, + { + "bbox": [ + 313, + 590, + 555, + 714 + ], + "type": "text", + "content": " is first transformed into feature " + }, + { + "bbox": [ + 313, + 590, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\dot{G}_j" + }, + { + "bbox": [ + 313, + 590, + 555, + 714 + ], + "type": "text", + "content": ". Simultaneously, noise feature " + }, + { + "bbox": [ + 313, + 590, + 555, + 714 + ], + "type": "inline_equation", + "content": "N_j" + }, + { + "bbox": [ + 313, + 590, + 555, + 714 + ], + "type": "text", + "content": " are extracted from " + }, + { + "bbox": [ + 313, + 590, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\tilde{G}_j" + }, + { + "bbox": [ + 313, + 590, + 555, + 714 + ], + "type": "text", + "content": " via noise estimation submodule. " + }, + { + "bbox": [ + 313, + 590, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\dot{G}_j" + }, + { + "bbox": [ + 313, + 590, + 555, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 590, + 555, + 714 + ], + "type": "inline_equation", + "content": "N_j" + }, + { + "bbox": [ + 313, + 590, + 555, + 714 + ], + "type": "text", + "content": " are passed through filter" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "11420" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 70, + 295, + 173 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 295, + 173 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 295, + 173 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 295, + 173 + ], + "type": "image", + "image_path": "8bfd45319a78c07051618e16cb04cf02ba648aae7b83bd7b9c3f4d2045f14d25.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 181, + 295, + 214 + ], + "lines": [ + { + "bbox": [ + 55, + 181, + 295, + 214 + ], + "spans": [ + { + "bbox": [ + 55, + 181, + 295, + 214 + ], + "type": "text", + "content": "Figure 7. Illustration of submodules in step 4 of MACF. (a) Transform and inverse transform submodules; (b) noise estimation submodule; (c) filtering submodule." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 220, + 295, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 220, + 295, + 270 + ], + "spans": [ + { + "bbox": [ + 55, + 220, + 295, + 270 + ], + "type": "text", + "content": "ing submodule to generate " + }, + { + "bbox": [ + 55, + 220, + 295, + 270 + ], + "type": "inline_equation", + "content": "\\ddot{G}_j" + }, + { + "bbox": [ + 55, + 220, + 295, + 270 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 55, + 220, + 295, + 270 + ], + "type": "inline_equation", + "content": "\\ddot{G}_j" + }, + { + "bbox": [ + 55, + 220, + 295, + 270 + ], + "type": "text", + "content": " are subsequently fed into inverse transform submodule. Additionally, a residual connection is applied in TDF step, producing refinement features " + }, + { + "bbox": [ + 55, + 220, + 295, + 270 + ], + "type": "inline_equation", + "content": "\\Delta \\tilde{G}_j" + }, + { + "bbox": [ + 55, + 220, + 295, + 270 + ], + "type": "text", + "content": " at last." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 270, + 296, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 270, + 296, + 415 + ], + "spans": [ + { + "bbox": [ + 55, + 270, + 296, + 415 + ], + "type": "text", + "content": "The above four submodules are illustrated in detail in Fig. 7. Transform and inverse transform submodules are independently realized via several convolutional layers (Fig. 7 (a)), involving both spatial and inter-channel transform. For noise estimation (NE) (Fig. 7 (b)), we first average " + }, + { + "bbox": [ + 55, + 270, + 296, + 415 + ], + "type": "inline_equation", + "content": "\\tilde{G}_j" + }, + { + "bbox": [ + 55, + 270, + 296, + 415 + ], + "type": "text", + "content": " along the channel dimension to obtain a rough temporal average result " + }, + { + "bbox": [ + 55, + 270, + 296, + 415 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 55, + 270, + 296, + 415 + ], + "type": "text", + "content": ". Then we calculate the residual between " + }, + { + "bbox": [ + 55, + 270, + 296, + 415 + ], + "type": "inline_equation", + "content": "\\tilde{G}_j" + }, + { + "bbox": [ + 55, + 270, + 296, + 415 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 270, + 296, + 415 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 55, + 270, + 296, + 415 + ], + "type": "text", + "content": " to represent fluctuations in features. Lastly, we extract deeper features from the residual to derive noise feature " + }, + { + "bbox": [ + 55, + 270, + 296, + 415 + ], + "type": "inline_equation", + "content": "N_j" + }, + { + "bbox": [ + 55, + 270, + 296, + 415 + ], + "type": "text", + "content": ". We concatenate the transformed feature " + }, + { + "bbox": [ + 55, + 270, + 296, + 415 + ], + "type": "inline_equation", + "content": "\\dot{G}_j" + }, + { + "bbox": [ + 55, + 270, + 296, + 415 + ], + "type": "text", + "content": " and noise feature " + }, + { + "bbox": [ + 55, + 270, + 296, + 415 + ], + "type": "inline_equation", + "content": "N_j" + }, + { + "bbox": [ + 55, + 270, + 296, + 415 + ], + "type": "text", + "content": " as input to the filtering submodule (Fig. 7 (c)), which contains a U-Net structure [33] and outputs " + }, + { + "bbox": [ + 55, + 270, + 296, + 415 + ], + "type": "inline_equation", + "content": "\\ddot{G}_j" + }, + { + "bbox": [ + 55, + 270, + 296, + 415 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 415, + 296, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 415, + 296, + 487 + ], + "spans": [ + { + "bbox": [ + 55, + 415, + 296, + 487 + ], + "type": "text", + "content": "Step 5: Inverse Feature Warping and Aggregation. We propose an inverse motion alignment operation to map these multi-moment refinement features " + }, + { + "bbox": [ + 55, + 415, + 296, + 487 + ], + "type": "inline_equation", + "content": "\\Delta \\tilde{G}_j" + }, + { + "bbox": [ + 55, + 415, + 296, + 487 + ], + "type": "text", + "content": " to their original positions. Specifically, given the backward offsets from key feature to neighboring features " + }, + { + "bbox": [ + 55, + 415, + 296, + 487 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{O}_i^{\\mathbf{B}}\\}" + }, + { + "bbox": [ + 55, + 415, + 296, + 487 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 415, + 296, + 487 + ], + "type": "inline_equation", + "content": "i = \\pm 1,\\dots,\\pm n" + }, + { + "bbox": [ + 55, + 415, + 296, + 487 + ], + "type": "text", + "content": ", we achieve inverse feature warping by:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 80, + 495, + 295, + 512 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 495, + 295, + 512 + ], + "spans": [ + { + "bbox": [ + 80, + 495, + 295, + 512 + ], + "type": "interline_equation", + "content": "\\Delta G _ {j} ^ {i} = \\mathcal {D} \\left(\\Delta \\tilde {G} _ {j} ^ {i}, \\mathcal {O} _ {i} ^ {\\mathbf {B}}\\right), \\forall i \\in \\{\\pm 1, \\dots , \\pm n \\}. \\tag {10}", + "image_path": "408a19d858ba26b9d7491709793ab9ef40bc6eb5e0690b9612b2d09912924ffd.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 520, + 296, + 595 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 520, + 296, + 595 + ], + "spans": [ + { + "bbox": [ + 55, + 520, + 296, + 595 + ], + "type": "inline_equation", + "content": "\\Delta G_{j}^{i}" + }, + { + "bbox": [ + 55, + 520, + 296, + 595 + ], + "type": "text", + "content": " is repositioned refinement feature and corresponds to initial unaligned feature " + }, + { + "bbox": [ + 55, + 520, + 296, + 595 + ], + "type": "inline_equation", + "content": "G_{j}^{i}" + }, + { + "bbox": [ + 55, + 520, + 296, + 595 + ], + "type": "text", + "content": ". We concatenate these repositioned refinement features and obtain " + }, + { + "bbox": [ + 55, + 520, + 296, + 595 + ], + "type": "inline_equation", + "content": "\\Delta G_{j} = \\{\\Delta G_{j}^{-n}, \\dots, \\Delta G_{j}^{n}\\}" + }, + { + "bbox": [ + 55, + 520, + 296, + 595 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 520, + 296, + 595 + ], + "type": "inline_equation", + "content": "\\Delta G_{j}^{0} = \\Delta \\tilde{G}_{j}^{0}" + }, + { + "bbox": [ + 55, + 520, + 296, + 595 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 55, + 520, + 296, + 595 + ], + "type": "inline_equation", + "content": "\\Delta G_{j}" + }, + { + "bbox": [ + 55, + 520, + 296, + 595 + ], + "type": "text", + "content": " are then aggregated with initial features " + }, + { + "bbox": [ + 55, + 520, + 296, + 595 + ], + "type": "inline_equation", + "content": "\\tilde{G}_{j}" + }, + { + "bbox": [ + 55, + 520, + 296, + 595 + ], + "type": "text", + "content": " to derive refined features with improved consistency:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 137, + 602, + 295, + 619 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 602, + 295, + 619 + ], + "spans": [ + { + "bbox": [ + 137, + 602, + 295, + 619 + ], + "type": "interline_equation", + "content": "\\acute {G} _ {j} = G _ {j} + \\Delta G _ {j}. \\tag {11}", + "image_path": "412eec90f6accd8448c626231da17f6858b4c8cc6dbe2db9bca7be4cb950ac01.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 628, + 295, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 628, + 295, + 674 + ], + "spans": [ + { + "bbox": [ + 55, + 628, + 295, + 674 + ], + "type": "text", + "content": "Step 6: Inverse Channel Rearrange. Finally, the features " + }, + { + "bbox": [ + 55, + 628, + 295, + 674 + ], + "type": "inline_equation", + "content": "\\{\\hat{G}_j\\}_{j=0}^{C-1}" + }, + { + "bbox": [ + 55, + 628, + 295, + 674 + ], + "type": "text", + "content": " are restored to " + }, + { + "bbox": [ + 55, + 628, + 295, + 674 + ], + "type": "inline_equation", + "content": "\\{\\hat{F}_i\\}_{i=-n}^n" + }, + { + "bbox": [ + 55, + 628, + 295, + 674 + ], + "type": "text", + "content": " via inverse channel rearrangement (ICR), i.e., the inverse operation of CR, as shown in Fig. 6." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 675, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 675, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 675, + 296, + 714 + ], + "type": "text", + "content": "MACF alleviates disturbances in " + }, + { + "bbox": [ + 55, + 675, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\{F_i\\}_{i = -n}^n" + }, + { + "bbox": [ + 55, + 675, + 296, + 714 + ], + "type": "text", + "content": " and generate improved features " + }, + { + "bbox": [ + 55, + 675, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\{\\dot{F}_i\\}_{i = -n}^n" + }, + { + "bbox": [ + 55, + 675, + 296, + 714 + ], + "type": "text", + "content": ", which are beneficial for enhancing SR reconstruction quality." + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 317, + 70, + 553, + 184 + ], + "blocks": [ + { + "bbox": [ + 317, + 70, + 553, + 184 + ], + "lines": [ + { + "bbox": [ + 317, + 70, + 553, + 184 + ], + "spans": [ + { + "bbox": [ + 317, + 70, + 553, + 184 + ], + "type": "image", + "image_path": "e4afe44b1bf93eb80fd5222ea4d8074623b4a93178ad135d3fe086f41774c847.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 192, + 553, + 214 + ], + "lines": [ + { + "bbox": [ + 313, + 192, + 553, + 214 + ], + "spans": [ + { + "bbox": [ + 313, + 192, + 553, + 214 + ], + "type": "text", + "content": "Figure 8. Illustration of motion aligned super resolved reconstruction module." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 222, + 553, + 246 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 222, + 553, + 246 + ], + "spans": [ + { + "bbox": [ + 313, + 222, + 553, + 246 + ], + "type": "text", + "content": "5.3. Motion Aligned Super Resolved Reconstruction" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 251, + 555, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 251, + 555, + 397 + ], + "spans": [ + { + "bbox": [ + 313, + 251, + 555, + 397 + ], + "type": "text", + "content": "The improved features " + }, + { + "bbox": [ + 313, + 251, + 555, + 397 + ], + "type": "inline_equation", + "content": "\\{\\hat{F}_i\\}_{i = -n}^n" + }, + { + "bbox": [ + 313, + 251, + 555, + 397 + ], + "type": "text", + "content": " are then input to motion aligned super resolved reconstruction module to generate the high resolution image " + }, + { + "bbox": [ + 313, + 251, + 555, + 397 + ], + "type": "inline_equation", + "content": "\\mathcal{I}(t_0)" + }, + { + "bbox": [ + 313, + 251, + 555, + 397 + ], + "type": "text", + "content": ", as illustrated in Fig. 8. Firstly, features at neighboring moments " + }, + { + "bbox": [ + 313, + 251, + 555, + 397 + ], + "type": "inline_equation", + "content": "\\{\\hat{F}_i\\}, i = \\pm 1,\\dots,\\pm n" + }, + { + "bbox": [ + 313, + 251, + 555, + 397 + ], + "type": "text", + "content": " are aligned with the key feature " + }, + { + "bbox": [ + 313, + 251, + 555, + 397 + ], + "type": "inline_equation", + "content": "\\hat{F}_0" + }, + { + "bbox": [ + 313, + 251, + 555, + 397 + ], + "type": "text", + "content": " by deformable convolution [7, 56], where the features " + }, + { + "bbox": [ + 313, + 251, + 555, + 397 + ], + "type": "inline_equation", + "content": "\\{\\hat{F}_i\\}_{i = -n}^n" + }, + { + "bbox": [ + 313, + 251, + 555, + 397 + ], + "type": "text", + "content": " with improved consistency help enhance the accuracy of motion alignment, benefiting the utilization of temporal information. Subsequently, key feature and aligned neighboring features are concatenated and fused through several convolutional layers. The fused feature is upsampled to the HR image by pixel shuffle operation [3]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 408, + 395, + 422 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 408, + 395, + 422 + ], + "spans": [ + { + "bbox": [ + 313, + 408, + 395, + 422 + ], + "type": "text", + "content": "6. Experiments" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 429, + 447, + 441 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 429, + 447, + 441 + ], + "spans": [ + { + "bbox": [ + 313, + 429, + 447, + 441 + ], + "type": "text", + "content": "6.1. Implementation Details" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 447, + 555, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 447, + 555, + 590 + ], + "spans": [ + { + "bbox": [ + 313, + 447, + 555, + 590 + ], + "type": "text", + "content": "We set the number of input spike frames to 101. The number of partitioned short-term blocks is 9 (" + }, + { + "bbox": [ + 313, + 447, + 555, + 590 + ], + "type": "inline_equation", + "content": "K = 9" + }, + { + "bbox": [ + 313, + 447, + 555, + 590 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 447, + 555, + 590 + ], + "type": "inline_equation", + "content": "n = 4" + }, + { + "bbox": [ + 313, + 447, + 555, + 590 + ], + "type": "text", + "content": "). The temporal length " + }, + { + "bbox": [ + 313, + 447, + 555, + 590 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 313, + 447, + 555, + 590 + ], + "type": "text", + "content": " of each block " + }, + { + "bbox": [ + 313, + 447, + 555, + 590 + ], + "type": "inline_equation", + "content": "B_{i}" + }, + { + "bbox": [ + 313, + 447, + 555, + 590 + ], + "type": "text", + "content": " is 21. The channel " + }, + { + "bbox": [ + 313, + 447, + 555, + 590 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 313, + 447, + 555, + 590 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 313, + 447, + 555, + 590 + ], + "type": "inline_equation", + "content": "F_{i}" + }, + { + "bbox": [ + 313, + 447, + 555, + 590 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 447, + 555, + 590 + ], + "type": "inline_equation", + "content": "\\hat{F}_i" + }, + { + "bbox": [ + 313, + 447, + 555, + 590 + ], + "type": "text", + "content": " is 64. During training, we randomly crop the spike stream to a spatial size of " + }, + { + "bbox": [ + 313, + 447, + 555, + 590 + ], + "type": "inline_equation", + "content": "64\\times 64" + }, + { + "bbox": [ + 313, + 447, + 555, + 590 + ], + "type": "text", + "content": ". In addition, we randomly perform horizontal and vertical flips as well as rotation for data augmentation. The batch size is set to 8. We use Adam optimizer [23] with " + }, + { + "bbox": [ + 313, + 447, + 555, + 590 + ], + "type": "inline_equation", + "content": "\\beta_{1} = 0.9" + }, + { + "bbox": [ + 313, + 447, + 555, + 590 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 447, + 555, + 590 + ], + "type": "inline_equation", + "content": "\\beta_{2} = 0.999" + }, + { + "bbox": [ + 313, + 447, + 555, + 590 + ], + "type": "text", + "content": ". The learning rate is initialized as 2e-4 and decayed by 0.7 every 100 epochs. The model is trained with " + }, + { + "bbox": [ + 313, + 447, + 555, + 590 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_1" + }, + { + "bbox": [ + 313, + 447, + 555, + 590 + ], + "type": "text", + "content": " loss for 800 epochs. All models are trained and tested on NVIDIA RTX 3090 GPU." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 600, + 376, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 600, + 376, + 611 + ], + "spans": [ + { + "bbox": [ + 313, + 600, + 376, + 611 + ], + "type": "text", + "content": "6.2. Datasets" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "type": "text", + "content": "Synthesized Dataset. We need a large number of \"LR spike stream and HR image\" pairs to train the model. However, it is challenging to obtain corresponding high-quality HR images, especially in high dynamic scenes. Inspired by the experiments conducted in [48, 50], we use a spike camera simulator (detailed in supplementary material) to mimic the working mechanism of spike camera and generate synthesized SCSR datasets for training and evaluation, where" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11421" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 65, + 70, + 547, + 394 + ], + "blocks": [ + { + "bbox": [ + 65, + 70, + 547, + 394 + ], + "lines": [ + { + "bbox": [ + 65, + 70, + 547, + 394 + ], + "spans": [ + { + "bbox": [ + 65, + 70, + 547, + 394 + ], + "type": "table", + "html": "
ScaleMethodREDS-based DatasetAdobe240-based DatasetParams(M)Runtime(s)
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
×2TFP+DAT25.550.59540.445627.660.67820.367714.654.40
Spk2ImgNet+DAT31.770.90240.190734.920.96040.094918.564.63
WGSE+DAT31.540.89420.205534.750.95710.102218.604.50
BSF+DAT31.920.90440.186735.030.96100.093417.135.01
VidarSR31.970.89640.204035.130.95580.110510.162.09
SpikeSR-Net34.290.93300.148137.040.97040.08383.262.14
Ours34.500.93770.140337.220.97250.08163.720.81
×3TFP+DAT24.880.57540.446426.170.65770.372114.841.88
Spk2ImgNet+DAT28.900.82930.284930.500.90640.170818.752.10
WGSE+DAT28.760.82170.296130.330.90130.177918.782.02
BSF+DAT28.990.83150.280730.440.90540.169917.322.02
VidarSR29.550.83730.279132.010.91970.168511.380.85
SpikeSR-Net31.090.87980.230433.740.94390.13553.301.35
Ours31.410.88500.223134.110.94700.13273.900.37
×4TFP+DAT24.350.56170.473425.090.64070.399514.801.06
TFP+BasicVSR++23.690.50230.485024.210.56100.44157.320.08
Spk2ImgNet+DAT27.230.76780.352127.970.84590.237818.711.15
Spk2ImgNet+BasicVSR++27.280.77300.352727.770.84210.252611.230.17
WGSE+DAT27.150.76180.359727.910.84160.242718.751.08
WGSE+BasicVSR++27.160.76990.356127.720.84040.255811.080.12
BSF+DAT27.300.77010.347327.900.84510.235817.281.10
BSF+BasicVSR++27.240.78180.337427.430.83890.24819.800.18
VidarSR28.300.79600.324930.180.88340.209712.790.75
SpikeSR-Net29.220.82660.294631.310.90730.18853.341.07
Ours29.500.83690.281631.490.91190.18223.860.21
", + "image_path": "1dc90f30285c37a3b27226fbf30eaa283fda1627b90193e5494562ee82e36d88.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 58, + 441, + 294, + 585 + ], + "blocks": [ + { + "bbox": [ + 55, + 402, + 555, + 437 + ], + "lines": [ + { + "bbox": [ + 55, + 402, + 555, + 437 + ], + "spans": [ + { + "bbox": [ + 55, + 402, + 555, + 437 + ], + "type": "text", + "content": "Table 1. Quantitative results on REDS-based and Adobe240-based SCSR evaluation dataset. Red and blue indicate the best and the second-best performance, respectively. It is noted that BasicVSR++ [4] only support " + }, + { + "bbox": [ + 55, + 402, + 555, + 437 + ], + "type": "inline_equation", + "content": "\\times 4" + }, + { + "bbox": [ + 55, + 402, + 555, + 437 + ], + "type": "text", + "content": " SR scale. The runtime is tested on input spike stream of sizes " + }, + { + "bbox": [ + 55, + 402, + 555, + 437 + ], + "type": "inline_equation", + "content": "360 \\times 640 \\times 101 (\\times 2)" + }, + { + "bbox": [ + 55, + 402, + 555, + 437 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 402, + 555, + 437 + ], + "type": "inline_equation", + "content": "240 \\times 426 \\times 101 (\\times 3)" + }, + { + "bbox": [ + 55, + 402, + 555, + 437 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 402, + 555, + 437 + ], + "type": "inline_equation", + "content": "180 \\times 320 \\times 101 (\\times 4)" + }, + { + "bbox": [ + 55, + 402, + 555, + 437 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 441, + 294, + 585 + ], + "lines": [ + { + "bbox": [ + 58, + 441, + 294, + 585 + ], + "spans": [ + { + "bbox": [ + 58, + 441, + 294, + 585 + ], + "type": "table", + "html": "
CaseSIRMACFPSNR↑SSIM↑
step 3step 4step 5
(A)29.070.8207
(B)✓29.150.8237
(C)✓29.170.8244
(D)✓29.320.8306
(E)✓✓29.340.8316
(F)✓✓29.350.8322
(G)✓✓29.330.8304
(H)✓✓✓29.460.8344
(I)✓✓✓✓29.500.8369
", + "image_path": "3d138d00e430ff818b2df16dfa58991cf491f268873a6ed248f5033bd58050d2.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 592, + 295, + 614 + ], + "lines": [ + { + "bbox": [ + 55, + 592, + 295, + 614 + ], + "spans": [ + { + "bbox": [ + 55, + 592, + 295, + 614 + ], + "type": "text", + "content": "Table 2. Ablation study of the proposed Spk2SRImgNet on REDS-based SCSR evaluation dataset " + }, + { + "bbox": [ + 55, + 592, + 295, + 614 + ], + "type": "inline_equation", + "content": "(\\times 4)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 617, + 260, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 260, + 629 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 260, + 629 + ], + "type": "text", + "content": "each sample consists of a spike stream-image pair." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "text", + "content": "Specifically, we use the 120 fps REDS dataset [30] with a spatial resolution of " + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "inline_equation", + "content": "720 \\times 1280" + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "text", + "content": " to generate the synthesized dataset. REDS includes 240 scenes for training and 30 scenes for evaluation, and we follow its original partition. For each scene, we choose 5 segments, resulting in " + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "inline_equation", + "content": "240 \\times 5 = 1200" + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "inline_equation", + "content": "30 \\times 5 = 150" + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "text", + "content": " pairs for training and evaluation, respectively. Each pair consists of spike" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 442, + 555, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 442, + 555, + 574 + ], + "spans": [ + { + "bbox": [ + 313, + 442, + 555, + 574 + ], + "type": "text", + "content": "stream of size " + }, + { + "bbox": [ + 313, + 442, + 555, + 574 + ], + "type": "inline_equation", + "content": "360 \\times 640 \\times 101 \\times 2) / 240 \\times 426 \\times 101 \\times 3) / 180 \\times 320 \\times 101 \\times 4" + }, + { + "bbox": [ + 313, + 442, + 555, + 574 + ], + "type": "text", + "content": " (101 is the temporal length) and a ground truth image of size " + }, + { + "bbox": [ + 313, + 442, + 555, + 574 + ], + "type": "inline_equation", + "content": "720 \\times 1280 \\times 2" + }, + { + "bbox": [ + 313, + 442, + 555, + 574 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 442, + 555, + 574 + ], + "type": "inline_equation", + "content": "\\times 4" + }, + { + "bbox": [ + 313, + 442, + 555, + 574 + ], + "type": "text", + "content": " / " + }, + { + "bbox": [ + 313, + 442, + 555, + 574 + ], + "type": "inline_equation", + "content": "720 \\times 1278 \\times 3" + }, + { + "bbox": [ + 313, + 442, + 555, + 574 + ], + "type": "text", + "content": ". To demonstrate the generalization of the proposed model, we generate another SCSR evaluation dataset from Adobe240 dataset [34]. Adobe240 dataset includes 133 videos of 240 fps with resolution of " + }, + { + "bbox": [ + 313, + 442, + 555, + 574 + ], + "type": "inline_equation", + "content": "720 \\times 1280" + }, + { + "bbox": [ + 313, + 442, + 555, + 574 + ], + "type": "text", + "content": " taken by hand-held cameras. We follow the test set partition of Adobe240 in [44], which includes 17 scenes. For each scene, we choose 10 segments, generating " + }, + { + "bbox": [ + 313, + 442, + 555, + 574 + ], + "type": "inline_equation", + "content": "17 \\times 10 = 170" + }, + { + "bbox": [ + 313, + 442, + 555, + 574 + ], + "type": "text", + "content": " pairs for evaluation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 574, + 556, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 574, + 556, + 635 + ], + "spans": [ + { + "bbox": [ + 313, + 574, + 556, + 635 + ], + "type": "text", + "content": "Real-World Captured Data. In addition to the synthesized data, we also evaluate the model performance on real-world spike data captured by the spike camera with a spatial resolution of " + }, + { + "bbox": [ + 313, + 574, + 556, + 635 + ], + "type": "inline_equation", + "content": "250 \\times 400" + }, + { + "bbox": [ + 313, + 574, + 556, + 635 + ], + "type": "text", + "content": ". Due to the absence of ground truth, we perform qualitative analysis for these data." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 647, + 504, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 647, + 504, + 659 + ], + "spans": [ + { + "bbox": [ + 313, + 647, + 504, + 659 + ], + "type": "text", + "content": "6.3. Comparison with Existing Methods" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "content": "We compare our proposed method with existing SCSR method: VidarSR [42] and SpikeSR-Net [50]. Besides, we combine four spike camera image reconstruction methods: TFP [53], Spk2ImgNet [48], WGSE [45] and BSF [52] with" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "11422" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 71, + 140, + 145 + ], + "blocks": [ + { + "bbox": [ + 57, + 71, + 140, + 145 + ], + "lines": [ + { + "bbox": [ + 57, + 71, + 140, + 145 + ], + "spans": [ + { + "bbox": [ + 57, + 71, + 140, + 145 + ], + "type": "image", + "image_path": "9c2ca347e4b924fca87f0eeaabd76516ef0640d65a5e9f9f3102e02bcf853b86.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 89, + 145, + 107, + 153 + ], + "lines": [ + { + "bbox": [ + 89, + 145, + 107, + 153 + ], + "spans": [ + { + "bbox": [ + 89, + 145, + 107, + 153 + ], + "type": "text", + "content": "Spike" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 140, + 72, + 222, + 145 + ], + "blocks": [ + { + "bbox": [ + 140, + 72, + 222, + 145 + ], + "lines": [ + { + "bbox": [ + 140, + 72, + 222, + 145 + ], + "spans": [ + { + "bbox": [ + 140, + 72, + 222, + 145 + ], + "type": "image", + "image_path": "edf5a0058cc08da47aeaf2a5c4cdc9beb9f2e63022ade05c3be3909e08bd26ad.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 165, + 145, + 196, + 153 + ], + "lines": [ + { + "bbox": [ + 165, + 145, + 196, + 153 + ], + "spans": [ + { + "bbox": [ + 165, + 145, + 196, + 153 + ], + "type": "text", + "content": "TFP+DAT" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 222, + 72, + 304, + 145 + ], + "blocks": [ + { + "bbox": [ + 222, + 72, + 304, + 145 + ], + "lines": [ + { + "bbox": [ + 222, + 72, + 304, + 145 + ], + "spans": [ + { + "bbox": [ + 222, + 72, + 304, + 145 + ], + "type": "image", + "image_path": "3ec66d28c1f792a036be8c30fbd3e32e1676b6c2e90a002ad23aa9cd4f63f6cb.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 236, + 145, + 291, + 152 + ], + "lines": [ + { + "bbox": [ + 236, + 145, + 291, + 152 + ], + "spans": [ + { + "bbox": [ + 236, + 145, + 291, + 152 + ], + "type": "text", + "content": "TFP+BasicVSR++" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 304, + 72, + 386, + 145 + ], + "blocks": [ + { + "bbox": [ + 304, + 72, + 386, + 145 + ], + "lines": [ + { + "bbox": [ + 304, + 72, + 386, + 145 + ], + "spans": [ + { + "bbox": [ + 304, + 72, + 386, + 145 + ], + "type": "image", + "image_path": "2aa6d8f6d3fc9e99e14ed5877f7677e051a8d803e97d6a5029507202e1c787a7.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 318, + 145, + 373, + 153 + ], + "lines": [ + { + "bbox": [ + 318, + 145, + 373, + 153 + ], + "spans": [ + { + "bbox": [ + 318, + 145, + 373, + 153 + ], + "type": "text", + "content": "Spk2ImgNet+DAT" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 386, + 72, + 469, + 145 + ], + "blocks": [ + { + "bbox": [ + 386, + 72, + 469, + 145 + ], + "lines": [ + { + "bbox": [ + 386, + 72, + 469, + 145 + ], + "spans": [ + { + "bbox": [ + 386, + 72, + 469, + 145 + ], + "type": "image", + "image_path": "67a356828c4b89629382dec22ba463b7d0eeefaf6981262e297c0ee9ddc8094c.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 389, + 145, + 468, + 153 + ], + "lines": [ + { + "bbox": [ + 389, + 145, + 468, + 153 + ], + "spans": [ + { + "bbox": [ + 389, + 145, + 468, + 153 + ], + "type": "text", + "content": "Spk2ImgNet+BasicVSR++" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 470, + 72, + 552, + 145 + ], + "blocks": [ + { + "bbox": [ + 470, + 72, + 552, + 145 + ], + "lines": [ + { + "bbox": [ + 470, + 72, + 552, + 145 + ], + "spans": [ + { + "bbox": [ + 470, + 72, + 552, + 145 + ], + "type": "image", + "image_path": "2dd1c94efa4edc0d926bd0a1066ab9708513b19e822e38487a01b362a0c01d21.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 492, + 145, + 529, + 153 + ], + "lines": [ + { + "bbox": [ + 492, + 145, + 529, + 153 + ], + "spans": [ + { + "bbox": [ + 492, + 145, + 529, + 153 + ], + "type": "text", + "content": "WGSE+DAT" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 58, + 154, + 139, + 227 + ], + "blocks": [ + { + "bbox": [ + 58, + 154, + 139, + 227 + ], + "lines": [ + { + "bbox": [ + 58, + 154, + 139, + 227 + ], + "spans": [ + { + "bbox": [ + 58, + 154, + 139, + 227 + ], + "type": "image", + "image_path": "e6b86425aa247392e9321f625813d9c450f5108d7ad88fb201ccd7c76d2f9c8a.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 67, + 228, + 129, + 236 + ], + "lines": [ + { + "bbox": [ + 67, + 228, + 129, + 236 + ], + "spans": [ + { + "bbox": [ + 67, + 228, + 129, + 236 + ], + "type": "text", + "content": "WGSE+BasicVSR++" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 140, + 154, + 222, + 227 + ], + "blocks": [ + { + "bbox": [ + 140, + 154, + 222, + 227 + ], + "lines": [ + { + "bbox": [ + 140, + 154, + 222, + 227 + ], + "spans": [ + { + "bbox": [ + 140, + 154, + 222, + 227 + ], + "type": "image", + "image_path": "3fe5b20d437dbeb07174671b2f94f5f017f72b7fe115a856e8e4b986c784ea72.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 165, + 228, + 197, + 236 + ], + "lines": [ + { + "bbox": [ + 165, + 228, + 197, + 236 + ], + "spans": [ + { + "bbox": [ + 165, + 228, + 197, + 236 + ], + "type": "text", + "content": "BSF+DAT" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 222, + 154, + 304, + 227 + ], + "blocks": [ + { + "bbox": [ + 222, + 154, + 304, + 227 + ], + "lines": [ + { + "bbox": [ + 222, + 154, + 304, + 227 + ], + "spans": [ + { + "bbox": [ + 222, + 154, + 304, + 227 + ], + "type": "image", + "image_path": "ea64bfaa7344eed273c514d5f555b6575630351c4c9d05e2ad3f1032897ae654.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 236, + 228, + 291, + 236 + ], + "lines": [ + { + "bbox": [ + 236, + 228, + 291, + 236 + ], + "spans": [ + { + "bbox": [ + 236, + 228, + 291, + 236 + ], + "type": "text", + "content": "BSF+BasicVSR++" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 304, + 154, + 386, + 227 + ], + "blocks": [ + { + "bbox": [ + 304, + 154, + 386, + 227 + ], + "lines": [ + { + "bbox": [ + 304, + 154, + 386, + 227 + ], + "spans": [ + { + "bbox": [ + 304, + 154, + 386, + 227 + ], + "type": "image", + "image_path": "d46f6e6f139f61284ace99c9a38d148fb565f9f911d84c598bb019b3e5272db6.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 333, + 228, + 359, + 236 + ], + "lines": [ + { + "bbox": [ + 333, + 228, + 359, + 236 + ], + "spans": [ + { + "bbox": [ + 333, + 228, + 359, + 236 + ], + "type": "text", + "content": "VidarSR" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 386, + 154, + 469, + 227 + ], + "blocks": [ + { + "bbox": [ + 386, + 154, + 469, + 227 + ], + "lines": [ + { + "bbox": [ + 386, + 154, + 469, + 227 + ], + "spans": [ + { + "bbox": [ + 386, + 154, + 469, + 227 + ], + "type": "image", + "image_path": "51f15c14bf48fedb3509deb57bdeafd18590766bf48080f3fab9d3d11b8d974e.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 409, + 228, + 448, + 236 + ], + "lines": [ + { + "bbox": [ + 409, + 228, + 448, + 236 + ], + "spans": [ + { + "bbox": [ + 409, + 228, + 448, + 236 + ], + "type": "text", + "content": "SpikeSR-Net" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 470, + 154, + 552, + 227 + ], + "blocks": [ + { + "bbox": [ + 470, + 154, + 552, + 227 + ], + "lines": [ + { + "bbox": [ + 470, + 154, + 552, + 227 + ], + "spans": [ + { + "bbox": [ + 470, + 154, + 552, + 227 + ], + "type": "image", + "image_path": "c4a67c6d2ddc6fbc50b1a32b34f39f69d02d81f38590a86a13b501bd509fc11b.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 503, + 228, + 518, + 236 + ], + "lines": [ + { + "bbox": [ + 503, + 228, + 518, + 236 + ], + "spans": [ + { + "bbox": [ + 503, + 228, + 518, + 236 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 58, + 237, + 139, + 311 + ], + "blocks": [ + { + "bbox": [ + 58, + 237, + 139, + 311 + ], + "lines": [ + { + "bbox": [ + 58, + 237, + 139, + 311 + ], + "spans": [ + { + "bbox": [ + 58, + 237, + 139, + 311 + ], + "type": "image", + "image_path": "88a6b2f1bf8d21314180e59a00823a09b6c24d8fed1d8137943ac84578d01027.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 140, + 237, + 222, + 311 + ], + "blocks": [ + { + "bbox": [ + 140, + 237, + 222, + 311 + ], + "lines": [ + { + "bbox": [ + 140, + 237, + 222, + 311 + ], + "spans": [ + { + "bbox": [ + 140, + 237, + 222, + 311 + ], + "type": "image", + "image_path": "82e5a4ae2802552d2a6191889da7c10db672c987f8526a61e7f1939cdd98734a.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 222, + 237, + 304, + 311 + ], + "blocks": [ + { + "bbox": [ + 222, + 237, + 304, + 311 + ], + "lines": [ + { + "bbox": [ + 222, + 237, + 304, + 311 + ], + "spans": [ + { + "bbox": [ + 222, + 237, + 304, + 311 + ], + "type": "image", + "image_path": "fee3930eac04be89d9a12881237818cedd0d6709309351dcddd25976f8611d87.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 304, + 237, + 386, + 311 + ], + "blocks": [ + { + "bbox": [ + 304, + 237, + 386, + 311 + ], + "lines": [ + { + "bbox": [ + 304, + 237, + 386, + 311 + ], + "spans": [ + { + "bbox": [ + 304, + 237, + 386, + 311 + ], + "type": "image", + "image_path": "1f8415383d0f7909ea2829b17a6828c08d7db9ab6037c015dd702cbbaff4552d.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 386, + 237, + 469, + 311 + ], + "blocks": [ + { + "bbox": [ + 386, + 237, + 469, + 311 + ], + "lines": [ + { + "bbox": [ + 386, + 237, + 469, + 311 + ], + "spans": [ + { + "bbox": [ + 386, + 237, + 469, + 311 + ], + "type": "image", + "image_path": "ffc0b050aa9dd4169f8e6e65b542be65edf8758d46e7320c43d8647747d083fb.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 470, + 237, + 552, + 311 + ], + "blocks": [ + { + "bbox": [ + 470, + 237, + 552, + 311 + ], + "lines": [ + { + "bbox": [ + 470, + 237, + 552, + 311 + ], + "spans": [ + { + "bbox": [ + 470, + 237, + 552, + 311 + ], + "type": "image", + "image_path": "943720ade2f9e54c09b579f45416bd23b83dedfd061211ea3d881e27e640fcfc.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 58, + 319, + 139, + 393 + ], + "blocks": [ + { + "bbox": [ + 89, + 312, + 107, + 319 + ], + "lines": [ + { + "bbox": [ + 89, + 312, + 107, + 319 + ], + "spans": [ + { + "bbox": [ + 89, + 312, + 107, + 319 + ], + "type": "text", + "content": "Spike" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 58, + 319, + 139, + 393 + ], + "lines": [ + { + "bbox": [ + 58, + 319, + 139, + 393 + ], + "spans": [ + { + "bbox": [ + 58, + 319, + 139, + 393 + ], + "type": "image", + "image_path": "392c45de6f37c59b29fb1b8c240885fc207d57331002d8d36ad2aeb132141268.jpg" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 67, + 394, + 130, + 402 + ], + "lines": [ + { + "bbox": [ + 67, + 394, + 130, + 402 + ], + "spans": [ + { + "bbox": [ + 67, + 394, + 130, + 402 + ], + "type": "text", + "content": "WGSE+BasicVSR++" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 412, + 553, + 434 + ], + "lines": [ + { + "bbox": [ + 55, + 412, + 553, + 434 + ], + "spans": [ + { + "bbox": [ + 55, + 412, + 553, + 434 + ], + "type": "text", + "content": "Figure 9. Visual comparison " + }, + { + "bbox": [ + 55, + 412, + 553, + 434 + ], + "type": "inline_equation", + "content": "(\\times 4)" + }, + { + "bbox": [ + 55, + 412, + 553, + 434 + ], + "type": "text", + "content": " on real-world spike data. The first scene is a stationary clock recorded by a fast-moving spike camera. The second scene is a rotating fan. Please enlarge the figure for better comparison." + } + ] + } + ], + "index": 48, + "angle": 0, + "type": "image_caption" + } + ], + "index": 36 + }, + { + "type": "image", + "bbox": [ + 140, + 319, + 222, + 393 + ], + "blocks": [ + { + "bbox": [ + 165, + 312, + 197, + 319 + ], + "lines": [ + { + "bbox": [ + 165, + 312, + 197, + 319 + ], + "spans": [ + { + "bbox": [ + 165, + 312, + 197, + 319 + ], + "type": "text", + "content": "TFP+DAT" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 140, + 319, + 222, + 393 + ], + "lines": [ + { + "bbox": [ + 140, + 319, + 222, + 393 + ], + "spans": [ + { + "bbox": [ + 140, + 319, + 222, + 393 + ], + "type": "image", + "image_path": "bab5e91e3ea16b40b013c32ef4b377f8a8e1d1762327adb18577e1df2dee4cef.jpg" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 165, + 394, + 197, + 401 + ], + "lines": [ + { + "bbox": [ + 165, + 394, + 197, + 401 + ], + "spans": [ + { + "bbox": [ + 165, + 394, + 197, + 401 + ], + "type": "text", + "content": "BSF+DAT" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_caption" + } + ], + "index": 38 + }, + { + "type": "image", + "bbox": [ + 222, + 319, + 304, + 393 + ], + "blocks": [ + { + "bbox": [ + 236, + 312, + 291, + 319 + ], + "lines": [ + { + "bbox": [ + 236, + 312, + 291, + 319 + ], + "spans": [ + { + "bbox": [ + 236, + 312, + 291, + 319 + ], + "type": "text", + "content": "TFP+BasicVSR++" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 222, + 319, + 304, + 393 + ], + "lines": [ + { + "bbox": [ + 222, + 319, + 304, + 393 + ], + "spans": [ + { + "bbox": [ + 222, + 319, + 304, + 393 + ], + "type": "image", + "image_path": "fa1139c06a0f2928144c25a979fe5dcb7aeb69ceeda0cf82107b826098caca39.jpg" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 236, + 394, + 291, + 402 + ], + "lines": [ + { + "bbox": [ + 236, + 394, + 291, + 402 + ], + "spans": [ + { + "bbox": [ + 236, + 394, + 291, + 402 + ], + "type": "text", + "content": "BSF+BasicVSR++" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_caption" + } + ], + "index": 40 + }, + { + "type": "image", + "bbox": [ + 304, + 319, + 386, + 393 + ], + "blocks": [ + { + "bbox": [ + 319, + 312, + 373, + 319 + ], + "lines": [ + { + "bbox": [ + 319, + 312, + 373, + 319 + ], + "spans": [ + { + "bbox": [ + 319, + 312, + 373, + 319 + ], + "type": "text", + "content": "Spk2ImgNet+DAT" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 319, + 386, + 393 + ], + "lines": [ + { + "bbox": [ + 304, + 319, + 386, + 393 + ], + "spans": [ + { + "bbox": [ + 304, + 319, + 386, + 393 + ], + "type": "image", + "image_path": "3937c6da29913f837123c7962252d6e7d5e0a787ba184346e859b18099e9c37f.jpg" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 333, + 394, + 359, + 402 + ], + "lines": [ + { + "bbox": [ + 333, + 394, + 359, + 402 + ], + "spans": [ + { + "bbox": [ + 333, + 394, + 359, + 402 + ], + "type": "text", + "content": "VidarSR" + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_caption" + } + ], + "index": 42 + }, + { + "type": "image", + "bbox": [ + 386, + 319, + 469, + 393 + ], + "blocks": [ + { + "bbox": [ + 389, + 312, + 468, + 319 + ], + "lines": [ + { + "bbox": [ + 389, + 312, + 468, + 319 + ], + "spans": [ + { + "bbox": [ + 389, + 312, + 468, + 319 + ], + "type": "text", + "content": "Spk2ImgNet+BasicVSR++" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 386, + 319, + 469, + 393 + ], + "lines": [ + { + "bbox": [ + 386, + 319, + 469, + 393 + ], + "spans": [ + { + "bbox": [ + 386, + 319, + 469, + 393 + ], + "type": "image", + "image_path": "a32869622dd3b7ae426783078b3ec66e3402d53361ffa21e661e49375bd6259a.jpg" + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 409, + 394, + 448, + 402 + ], + "lines": [ + { + "bbox": [ + 409, + 394, + 448, + 402 + ], + "spans": [ + { + "bbox": [ + 409, + 394, + 448, + 402 + ], + "type": "text", + "content": "SpikeSR-Net" + } + ] + } + ], + "index": 45, + "angle": 0, + "type": "image_caption" + } + ], + "index": 44 + }, + { + "type": "image", + "bbox": [ + 470, + 319, + 552, + 393 + ], + "blocks": [ + { + "bbox": [ + 492, + 312, + 530, + 319 + ], + "lines": [ + { + "bbox": [ + 492, + 312, + 530, + 319 + ], + "spans": [ + { + "bbox": [ + 492, + 312, + 530, + 319 + ], + "type": "text", + "content": "WGSE+DAT" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 470, + 319, + 552, + 393 + ], + "lines": [ + { + "bbox": [ + 470, + 319, + 552, + 393 + ], + "spans": [ + { + "bbox": [ + 470, + 319, + 552, + 393 + ], + "type": "image", + "image_path": "28c2d0a3f7e3ac01de22cc6354db7632c128e8f2f6abef6c2c8b9eff92e6895a.jpg" + } + ] + } + ], + "index": 46, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 503, + 394, + 518, + 402 + ], + "lines": [ + { + "bbox": [ + 503, + 394, + 518, + 402 + ], + "spans": [ + { + "bbox": [ + 503, + 394, + 518, + 402 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 47, + "angle": 0, + "type": "image_caption" + } + ], + "index": 46 + }, + { + "bbox": [ + 55, + 441, + 296, + 585 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 441, + 296, + 585 + ], + "spans": [ + { + "bbox": [ + 55, + 441, + 296, + 585 + ], + "type": "text", + "content": "image / video SR method: DAT [5] / BasicVSR++ [4], resulting in another eight comparison methods, as presented in Tab. 1. We conduct experiments on " + }, + { + "bbox": [ + 55, + 441, + 296, + 585 + ], + "type": "inline_equation", + "content": "\\times 2" + }, + { + "bbox": [ + 55, + 441, + 296, + 585 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 441, + 296, + 585 + ], + "type": "inline_equation", + "content": "\\times 3" + }, + { + "bbox": [ + 55, + 441, + 296, + 585 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 441, + 296, + 585 + ], + "type": "inline_equation", + "content": "\\times 4" + }, + { + "bbox": [ + 55, + 441, + 296, + 585 + ], + "type": "text", + "content": " SR scales. We employ PSNR, SSIM [39] and LPIPS [46] as quantitative metrics. All compared methods except TFP [53] are retrained on the REDS-based SCSR training dataset with the same settings as ours. As shown in Tab. 1, our method achieves the best performance on the three metrics. Besides, we conduct experiments on real-world captured spike stream, as shown in Fig. 9, our method can restore better textures and details. More evaluation and visual results are provided in supplementary material." + } + ] + } + ], + "index": 49 + }, + { + "bbox": [ + 55, + 598, + 149, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 598, + 149, + 611 + ], + "spans": [ + { + "bbox": [ + 55, + 598, + 149, + 611 + ], + "type": "text", + "content": "6.4. Ablation Study" + } + ] + } + ], + "index": 50 + }, + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "text", + "content": "To verify the effectiveness of our methodology, we perform ablation study with REDS-based evaluation dataset on " + }, + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\times 4" + }, + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "text", + "content": " SR scale, as presented in Tab. 2. Since ablation study only involves steps 3, 4, and 5 in MACF, the remaining steps are omitted in Tab. 2 (Steps 2 and 6 do not involve network parameters, and step 1 is present only when step 3 and/or step 5 are present). Case (I) is the final model. Case (A) is formed by removing SIR and MACF modules from" + } + ] + } + ], + "index": 51 + }, + { + "bbox": [ + 313, + 441, + 555, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 441, + 555, + 538 + ], + "spans": [ + { + "bbox": [ + 313, + 441, + 555, + 538 + ], + "type": "text", + "content": "Spk2SRImgNet. Comparisons between (A) and (B), (H) and (I) illustrate the advantages of converting information from spike domain to image domain (SIR module). Comparisons between (A) and (H), (B) and (I) demonstrate the effectiveness of MACF module, showing gains of 0.39dB and 0.35dB, respectively. Comparisons between (C-G) and (H) validate the effectiveness of each step in MACF. More ablation study is provided in supplementary material." + } + ] + } + ], + "index": 52 + }, + { + "bbox": [ + 313, + 549, + 388, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 549, + 388, + 562 + ], + "spans": [ + { + "bbox": [ + 313, + 549, + 388, + 562 + ], + "type": "text", + "content": "7. Conclusion" + } + ] + } + ], + "index": 53 + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "content": "This paper develops an end-to-end neural network named Spk2SRImgNet to reconstruct high resolution images from low resolution spike stream. To handle spike fluctuations, we propose a motion aligned collaborative filtering (MACF) module motivated by key ideas in classic image restoration schemes. Specifically, MACF leverages the temporal similarity of spike stream to enhance the consistency of multi-moment features along the motion trajectory, mitigating fluctuations and improving the quality of super resolution reconstruction. Experimental results demonstrate that the proposed Spk2SRImgNet achieves state-of-the-art performance on synthesized and real-captured spike data." + } + ] + } + ], + "index": 54 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11423" + } + ] + } + ], + "index": 55 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 297, + 127 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 297, + 127 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 297, + 127 + ], + "type": "text", + "content": "This work was supported by the National Natural Science Foundation of China under Grants 62072009 and U22B2035." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 137, + 115, + 149 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 137, + 115, + 149 + ], + "spans": [ + { + "bbox": [ + 56, + 137, + 115, + 149 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 157, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 157, + 296, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 157, + 296, + 201 + ], + "spans": [ + { + "bbox": [ + 61, + 157, + 296, + 201 + ], + "type": "text", + "content": "[1] Patrick Bardow, Andrew J. Davison, and Stefan Leutenegger. Simultaneous optical flow and intensity estimation from an event camera. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 884-892, 2016. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 202, + 296, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 202, + 296, + 246 + ], + "spans": [ + { + "bbox": [ + 62, + 202, + 296, + 246 + ], + "type": "text", + "content": "[2] Christian Brandli, Raphael Berner, Minhao Yang, Shih-Chii Liu, and Tobi Delbruck. A " + }, + { + "bbox": [ + 62, + 202, + 296, + 246 + ], + "type": "inline_equation", + "content": "240 \\times 180" + }, + { + "bbox": [ + 62, + 202, + 296, + 246 + ], + "type": "text", + "content": " 130 db " + }, + { + "bbox": [ + 62, + 202, + 296, + 246 + ], + "type": "inline_equation", + "content": "3\\mu s" + }, + { + "bbox": [ + 62, + 202, + 296, + 246 + ], + "type": "text", + "content": " latency global shutter spatiotemporal vision sensor. IEEE Journal of Solid-State Circuits, 49(10):2333-2341, 2014. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 247, + 296, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 247, + 296, + 312 + ], + "spans": [ + { + "bbox": [ + 62, + 247, + 296, + 312 + ], + "type": "text", + "content": "[3] Jose Caballero, Christian Ledig, Andrew Aitken, Alejandro Acosta, Johannes Totz, Zehan Wang, and Wenzhe Shi. Real-time video super-resolution with spatio-temporal networks and motion compensation. In 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2848-2857, 2017. 2, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 313, + 296, + 378 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 313, + 296, + 378 + ], + "spans": [ + { + "bbox": [ + 62, + 313, + 296, + 378 + ], + "type": "text", + "content": "[4] Kelvin C.K. Chan, Shangchen Zhou, Xiangyu Xu, and Chen Change Loy. Basicvsr++: Improving video superresolution with enhanced propagation and alignment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5972-5981, 2022. 2, 7, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 380, + 296, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 380, + 296, + 434 + ], + "spans": [ + { + "bbox": [ + 62, + 380, + 296, + 434 + ], + "type": "text", + "content": "[5] Zheng Chen, Yulun Zhang, Jinjin Gu, Linghe Kong, Xiaokang Yang, and Fisher Yu. Dual aggregation transformer for image super-resolution. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 12312-12321, 2023. 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 435, + 296, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 435, + 296, + 479 + ], + "spans": [ + { + "bbox": [ + 62, + 435, + 296, + 479 + ], + "type": "text", + "content": "[6] Kostadin Dabov, Alessandro Foi, Vladimir Katkovnik, and Karen Egiazarian. Image denoising by sparse 3-d transform-domain collaborative filtering. IEEE Transactions on Image Processing, 16(8):2080-2095, 2007. 2, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 480, + 296, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 480, + 296, + 524 + ], + "spans": [ + { + "bbox": [ + 62, + 480, + 296, + 524 + ], + "type": "text", + "content": "[7] Jifeng Dai, Haozhi Qi, Yuwen Xiong, Yi Li, Guodong Zhang, Han Hu, and Yichen Wei. Deformable convolutional networks. In 2017 IEEE International Conference on Computer Vision (ICCV), pages 764-773, 2017. 2, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 525, + 296, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 525, + 296, + 557 + ], + "spans": [ + { + "bbox": [ + 62, + 525, + 296, + 557 + ], + "type": "text", + "content": "[8] Siwei Dong, Tiejun Huang, and Yonghong Tian. Spike camera and its coding methods. In 2017 Data Compression Conference (DCC), pages 437-437, 2017. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 558, + 296, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 558, + 296, + 602 + ], + "spans": [ + { + "bbox": [ + 62, + 558, + 296, + 602 + ], + "type": "text", + "content": "[9] Siwei Dong, Lin Zhu, Daoyuan Xu, Yonghong Tian, and Tiejun Huang. An efficient coding method for spike camera using inter-spike intervals. In 2019 Data Compression Conference (DCC), pages 568-568, 2019. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 603, + 296, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 603, + 296, + 647 + ], + "spans": [ + { + "bbox": [ + 57, + 603, + 296, + 647 + ], + "type": "text", + "content": "[10] Yanchen Dong, Jing Zhao, Ruiqin Xiong, and Tiejun Huang. 3D residual interpolation for spike camera demosaicing. In IEEE International Conference on Image Processing (ICIP), pages 1461-1465. IEEE, 2022. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 647, + 296, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 647, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 647, + 296, + 712 + ], + "type": "text", + "content": "[11] Yanchen Dong, Ruiqin Xiong, Jian Zhang, Zhaofei Yu, Xiaopeng Fan, Shuyuan Zhu, and Tiejun Huang. Superresolution reconstruction from bayer-pattern spike streams. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 24871-24880, 2024. 2" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 715 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 316, + 73, + 555, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 555, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 555, + 128 + ], + "type": "text", + "content": "[12] Yanchen Dong, Ruiqin Xiong, Jing Zhao, Jian Zhang, Xiaopeng Fan, Shuyuan Zhu, and Tiejun Huang. Joint demosaicing and denoising for spike camera. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1582-1590, 2024. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 129, + 555, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 129, + 555, + 184 + ], + "spans": [ + { + "bbox": [ + 317, + 129, + 555, + 184 + ], + "type": "text", + "content": "[13] Yanchen Dong, Ruiqin Xiong, Jing Zhao, Jian Zhang, Xiaopeng Fan, Shuyuan Zhu, and Tiejun Huang. Learning a deep demosaicing network for spike camera with color filter array. IEEE Transactions on Image Processing, 33:3634-3647, 2024." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 185, + 554, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 185, + 554, + 229 + ], + "spans": [ + { + "bbox": [ + 316, + 185, + 554, + 229 + ], + "type": "text", + "content": "[14] Yanchen Dong, Ruiqin Xiong, Xiaopeng Fan, Shuyuan Zhu, Jin Wang, and Tiejun Huang. Dynamic scene reconstruction for color spike camera via zero-shot learning. IEEE Transactions on Computational Imaging, 11:129-141, 2025. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 230, + 554, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 230, + 554, + 262 + ], + "spans": [ + { + "bbox": [ + 316, + 230, + 554, + 262 + ], + "type": "text", + "content": "[15] Daniel Gehrig and Davide Scaramuzza. Low-latency automotive vision with event cameras. Nature, 629(8014):1034-1040, 2024. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 264, + 554, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 264, + 554, + 319 + ], + "spans": [ + { + "bbox": [ + 316, + 264, + 554, + 319 + ], + "type": "text", + "content": "[16] Arren Glover, Valentina Vasco, and Chiara Bartolozzi. A controlled-delay event camera framework for on-line robotics. In 2018 IEEE International Conference on Robotics and Automation (ICRA), pages 2178-2183. IEEE, 2018. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 319, + 554, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 319, + 554, + 364 + ], + "spans": [ + { + "bbox": [ + 316, + 319, + 554, + 364 + ], + "type": "text", + "content": "[17] Sorin Grigorescu, Bogdan Trasnea, Tiberiu Cocias, and Gigel Macesanu. A survey of deep learning techniques for autonomous driving. Journal of field robotics, 37(3):362-386, 2020. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 365, + 554, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 365, + 554, + 420 + ], + "spans": [ + { + "bbox": [ + 316, + 365, + 554, + 420 + ], + "type": "text", + "content": "[18] Shuhang Gu, Lei Zhang, Wangmeng Zuo, and Xiangchu Feng. Weighted nuclear norm minimization with application to image denoising. In 2014 IEEE Conference on Computer Vision and Pattern Recognition, pages 2862-2869, 2014. 2, 3, 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 421, + 554, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 421, + 554, + 476 + ], + "spans": [ + { + "bbox": [ + 316, + 421, + 554, + 476 + ], + "type": "text", + "content": "[19] Tiejun Huang, Yajing Zheng, Zhaofei Yu, Rui Chen, Yuan Li, Ruiqin Xiong, Lei Ma, Junwei Zhao, Siwei Dong, Lin Zhu, Jianing Li, Shanshan Jia, Yihua Fu, Boxin Shi, Si Wu, and Yonghong Tian. " + }, + { + "bbox": [ + 316, + 421, + 554, + 476 + ], + "type": "inline_equation", + "content": "1000\\times" + }, + { + "bbox": [ + 316, + 421, + 554, + 476 + ], + "type": "text", + "content": " faster camera and machine vision with ordinary devices. Engineering, 25:110-119, 2023. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 476, + 554, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 476, + 554, + 521 + ], + "spans": [ + { + "bbox": [ + 316, + 476, + 554, + 521 + ], + "type": "text", + "content": "[20] Arpit Jadon, Mohd Omama, Akshay Varshney, Mohammad Samar Ansari, and Rishabh Sharma. Firenet: A specialized lightweight fire & smoke detection model for real-time IoT applications. arXiv preprint arXiv:1905.11922, 2019. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 522, + 554, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 522, + 554, + 567 + ], + "spans": [ + { + "bbox": [ + 316, + 522, + 554, + 567 + ], + "type": "text", + "content": "[21] Armin Kappeler, Seunghwan Yoo, Qiqin Dai, and Aggelos K. Katsaggelos. Video super-resolution with convolutional neural networks. IEEE Transactions on Computational Imaging, 2(2):109-122, 2016. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 567, + 554, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 567, + 554, + 611 + ], + "spans": [ + { + "bbox": [ + 316, + 567, + 554, + 611 + ], + "type": "text", + "content": "[22] Hanme Kim, Ankur Handa, Ryad Benosman, Sio-Hoi Ieng, and Andrew Davison. Simultaneous mosaicing and tracking with an event camera. In Proceedings of the British Machine Vision Conference. BMVA Press, 2014. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 612, + 554, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 612, + 554, + 645 + ], + "spans": [ + { + "bbox": [ + 316, + 612, + 554, + 645 + ], + "type": "text", + "content": "[23] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 646, + 554, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 646, + 554, + 690 + ], + "spans": [ + { + "bbox": [ + 316, + 646, + 554, + 690 + ], + "type": "text", + "content": "[24] Jingyun Liang, Jiezhang Cao, Yuchen Fan, Kai Zhang, Rakesh Ranjan, Yawei Li, Radu Timofte, and Luc Van Gool. Vrt: A video restoration transformer. IEEE Transactions on Image Processing, 33:2171-2182, 2024. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 691, + 554, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 554, + 715 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 554, + 715 + ], + "type": "text", + "content": "[25] Patrick Lichtsteiner, Christoph Posch, and Tobi Delbruck. A " + }, + { + "bbox": [ + 316, + 691, + 554, + 715 + ], + "type": "inline_equation", + "content": "128 \\times 128" + }, + { + "bbox": [ + 316, + 691, + 554, + 715 + ], + "type": "text", + "content": " 120 db " + }, + { + "bbox": [ + 316, + 691, + 554, + 715 + ], + "type": "inline_equation", + "content": "15\\mu s" + }, + { + "bbox": [ + 316, + 691, + 554, + 715 + ], + "type": "text", + "content": " latency asynchronous temporal con" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "11424" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 77, + 73, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 73, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 77, + 73, + 294, + 95 + ], + "type": "text", + "content": "trast vision sensor. IEEE journal of solid-state circuits, 43 (2):566-576, 2008. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 173 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 173 + ], + "type": "text", + "content": "[26] Martin Litzenberger, Christoph Posch, D Bauer, Ahmed Nabil Belbachir, P Schon, B Kohn, and H Garn. Embedded vision system for real-time object tracking using an asynchronous transient vision sensor. In 2006 IEEE 12th Digital Signal Processing Workshop & 4th IEEE Signal Processing Education Workshop, pages 173-178. IEEE, 2006. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 175, + 295, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 295, + 218 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 295, + 218 + ], + "type": "text", + "content": "[27] Julien Mairal, Francis Bach, Jean Ponce, Guillermo Sapiro, and Andrew Zisserman. Non-local sparse models for image restoration. In 2009 IEEE 12th International Conference on Computer Vision, pages 2272-2279, 2009. 2, 3, 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 220, + 294, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 220, + 294, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 220, + 294, + 262 + ], + "type": "text", + "content": "[28] Mohammad Mostafavi, Yeongwoo Nam, Jonghyun Choi, and Kuk-Jin Yoon. E2sri: Learning to super-resolve intensity images from events. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(10):6890-6909, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 264, + 294, + 286 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 264, + 294, + 286 + ], + "spans": [ + { + "bbox": [ + 56, + 264, + 294, + 286 + ], + "type": "text", + "content": "[29] Elias Mueggler. Event-based vision for high-speed robotics. PhD thesis, University of Zurich, 2017. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 288, + 294, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 288, + 294, + 353 + ], + "spans": [ + { + "bbox": [ + 56, + 288, + 294, + 353 + ], + "type": "text", + "content": "[30] Seungjun Nah, Sungyong Baik, Seokil Hong, Gyeongsik Moon, Sanghyun Son, Radu Timofte, and Kyoung Mu Lee. Ntire 2019 challenge on video deblurring and superresolution: Dataset and study. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2019. 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 355, + 294, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 355, + 294, + 408 + ], + "spans": [ + { + "bbox": [ + 56, + 355, + 294, + 408 + ], + "type": "text", + "content": "[31] Christoph Posch, Daniel Matolin, and Rainer Wohlgenannt. A qvga 143 db dynamic range frame-free pwm image sensor with lossless pixel-level video compression and time-domain cds. IEEE Journal of Solid-State Circuits, 46(1):259-275, 2010. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 411, + 294, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 411, + 294, + 465 + ], + "spans": [ + { + "bbox": [ + 56, + 411, + 294, + 465 + ], + "type": "text", + "content": "[32] Henri Rebecq, René Ranftl, Vladlen Koltun, and Davide Scaramuzza. Events-to-video: Bringing modern computer vision to event cameras. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3857-3866, 2019. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 467, + 294, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 467, + 294, + 533 + ], + "spans": [ + { + "bbox": [ + 56, + 467, + 294, + 533 + ], + "type": "text", + "content": "[33] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention-MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, pages 234-241. Springer, 2015. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 534, + 294, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 534, + 294, + 588 + ], + "spans": [ + { + "bbox": [ + 56, + 534, + 294, + 588 + ], + "type": "text", + "content": "[34] Shuochen Su, Maurizio Delbrazio, Jue Wang, Guillermo Sapiro, Wolfgang Heidrich, and Oliver Wang. Deep video deblurring for hand-held cameras. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 1279-1288, 2017. 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 590, + 294, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 590, + 294, + 634 + ], + "spans": [ + { + "bbox": [ + 56, + 590, + 294, + 634 + ], + "type": "text", + "content": "[35] Yapeng Tian, Yulun Zhang, Yun Fu, and Chenliang Xu. Tdan: Temporally-deformable alignment network for video super-resolution. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 635, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 294, + 689 + ], + "type": "text", + "content": "[36] Lin Wang, Tae-Kyun Kim, and Kuk-Jin Yoon. Eventsr: From asynchronous events to image reconstruction, restoration, and super-resolution via end-to-end adversarial learning. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8312-8322, 2020. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "text", + "content": "[37] Xintao Wang, Kelvin C.K. Chan, Ke Yu, Chao Dong, and Chen Change Loy. Edvr: Video restoration with enhanced" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "text", + "content": "deformable convolutional networks. In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 1954-1963, 2019. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 108, + 553, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 108, + 553, + 162 + ], + "spans": [ + { + "bbox": [ + 316, + 108, + 553, + 162 + ], + "type": "text", + "content": "[38] Yuanlin Wang, Ruiqin Xiong, Jing Zhao, and Tiejun Huang. Reconstruct dynamic scene for spike camera based on 3d space time similarity. In 2024 IEEE International Conference on Image Processing (ICIP), pages 1595-1601, 2024. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 163, + 553, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 553, + 206 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 553, + 206 + ], + "type": "text", + "content": "[39] Zhou Wang, A.C. Bovik, H.R. Sheikh, and E.P. Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing, 13(4): 600-612, 2004. 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 209, + 553, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 209, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 316, + 209, + 553, + 262 + ], + "type": "text", + "content": "[40] Philippe Weinzaepfel, Jerome Revaud, Zaid Harchaoui, and Cordelia Schmid. Deepflow: Large displacement optical flow with deep matching. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), pages 1385-1392, 2013. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 264, + 553, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 264, + 553, + 308 + ], + "spans": [ + { + "bbox": [ + 316, + 264, + 553, + 308 + ], + "type": "text", + "content": "[41] Wenming Weng, Yueyi Zhang, and Zhiwei Xiong. Event-based video reconstruction using transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 310, + 553, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 310, + 553, + 365 + ], + "spans": [ + { + "bbox": [ + 316, + 310, + 553, + 365 + ], + "type": "text", + "content": "[42] Xijie Xiang, Lin Zhu, Jianing Li, Yixuan Wang, Tiejun Huang, and Yonghong Tian. Learning super-resolution reconstruction for high temporal resolution spike stream. IEEE Transactions on Circuits and Systems for Video Technology, 33(1):16-29, 2023. 1, 2, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 366, + 553, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 366, + 553, + 420 + ], + "spans": [ + { + "bbox": [ + 316, + 366, + 553, + 420 + ], + "type": "text", + "content": "[43] Ruiqin Xiong, Hangfan Liu, Xinfeng Zhang, Jian Zhang, Siwei Ma, Feng Wu, and Wen Gao. Image denoising via bandwise adaptive modeling and regularization exploiting nonlocal similarity. IEEE Transactions on Image Processing, 25 (12):5793-5805, 2016. 2, 3, 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 422, + 553, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 422, + 553, + 477 + ], + "spans": [ + { + "bbox": [ + 316, + 422, + 553, + 477 + ], + "type": "text", + "content": "[44] Gang Xu, Jun Xu, Zhen Li, Liang Wang, Xing Sun, and Ming-Ming Cheng. Temporal modulation network for controllable space-time video super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6388-6397, 2021. 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 479, + 553, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 479, + 553, + 533 + ], + "spans": [ + { + "bbox": [ + 316, + 479, + 553, + 533 + ], + "type": "text", + "content": "[45] Jiyuan Zhang, Shanshan Jia, Zhaofei Yu, and Tiejun Huang. Learning temporal-ordered representation for spike streams based on discrete wavelet transforms. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 137–147, 2023. 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 535, + 553, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 535, + 553, + 590 + ], + "spans": [ + { + "bbox": [ + 316, + 535, + 553, + 590 + ], + "type": "text", + "content": "[46] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 586-595, 2018. 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 591, + 553, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 591, + 553, + 645 + ], + "spans": [ + { + "bbox": [ + 316, + 591, + 553, + 645 + ], + "type": "text", + "content": "[47] Jing Zhao, Jiyu Xie, Ruiqin Xiong, Jian Zhang, Zhaofei Yu, and Tiejun Huang. Super resolve dynamic scene from continuous spike streams. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 2533-2542, 2021. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 647, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 712 + ], + "type": "text", + "content": "[48] Jing Zhao, Ruiqin Xiong, Hangfan Liu, Jian Zhang, and Tiejun Huang. Spk2imgnet: Learning to reconstruct dynamic scene from continuous spike stream. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11996-12005, 2021. 2, 6, 7" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "11425" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 73, + 295, + 529 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 126 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 126 + ], + "type": "text", + "content": "[49] Jing Zhao, Ruiqin Xiong, Jiyu Xie, Boxin Shi, Zhaofei Yu, Wen Gao, and Tiejun Huang. Reconstructing clear image for high-speed motion scene with a retina-inspired spike camera. IEEE Transactions on Computational Imaging, 8:12-27, 2021. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 129, + 295, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 129, + 295, + 183 + ], + "spans": [ + { + "bbox": [ + 55, + 129, + 295, + 183 + ], + "type": "text", + "content": "[50] Jing Zhao, Ruiqin Xiong, Jian Zhang, Rui Zhao, Hangfan Liu, and Tiejun Huang. Learning to super-resolve dynamic scenes for neuromorphic spike camera. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3579-3587, 2023. 1, 2, 6, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 185, + 295, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 185, + 295, + 239 + ], + "spans": [ + { + "bbox": [ + 57, + 185, + 295, + 239 + ], + "type": "text", + "content": "[51] Rui Zhao, Ruiqin Xiong, Jian Zhang, Zhaofei Yu, Shuyuan Zhu, Lei Ma, and Tiejun Huang. Spike camera image reconstruction using deep spiking neural networks. IEEE Transactions on Circuits and Systems for Video Technology, 34(6): 5207-5212, 2024. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 240, + 294, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 240, + 294, + 306 + ], + "spans": [ + { + "bbox": [ + 57, + 240, + 294, + 306 + ], + "type": "text", + "content": "[52] Rui Zhao, Ruiqin Xiong, Jing Zhao, Jian Zhang, Xiaopeng Fan, Zhaofei Yu, and Tiejun Huang. Boosting spike camera image reconstruction from a perspective of dealing with spike fluctuations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 24955-24965, 2024. 2, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 57, + 308, + 294, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 308, + 294, + 361 + ], + "spans": [ + { + "bbox": [ + 57, + 308, + 294, + 361 + ], + "type": "text", + "content": "[53] Lin Zhu, Siwei Dong, Tiejun Huang, and Yonghong Tian. A retina-inspired sampling method for visual texture reconstruction. In 2019 IEEE International Conference on Multimedia and Expo (ICME), pages 1432-1437. IEEE, 2019. 2, 3, 4, 7, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 57, + 363, + 294, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 363, + 294, + 418 + ], + "spans": [ + { + "bbox": [ + 57, + 363, + 294, + 418 + ], + "type": "text", + "content": "[54] Lin Zhu, Siwei Dong, Jianing Li, Tiejun Huang, and Yonghong Tian. Ultra-high temporal resolution visual reconstruction from a fovea-like spike camera via spiking neuron model. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(1):1233-1249, 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 57, + 419, + 294, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 419, + 294, + 474 + ], + "spans": [ + { + "bbox": [ + 57, + 419, + 294, + 474 + ], + "type": "text", + "content": "[55] Lin Zhu, Xiao Wang, Yi Chang, Jianing Li, Tiejun Huang, and Yonghong Tian. Event-based video reconstruction via potential-assisted spiking neural network. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3584-3594, 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 57, + 475, + 294, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 475, + 294, + 529 + ], + "spans": [ + { + "bbox": [ + 57, + 475, + 294, + 529 + ], + "type": "text", + "content": "[56] Xizhou Zhu, Han Hu, Stephen Lin, and Jifeng Dai. Deformable convnets v2: More deformable, better results. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9308-9316, 2019. 2, 6" + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "11426" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/SplatAD_ Real-Time Lidar and Camera Rendering with 3D Gaussian Splatting for Autonomous Driving/58aed1a8-44d5-4314-9f37-1456921ac406_content_list.json b/2025/SplatAD_ Real-Time Lidar and Camera Rendering with 3D Gaussian Splatting for Autonomous Driving/58aed1a8-44d5-4314-9f37-1456921ac406_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..8cc6bba06e74f301292955faa29b935cbd02e3f3 --- /dev/null +++ b/2025/SplatAD_ Real-Time Lidar and Camera Rendering with 3D Gaussian Splatting for Autonomous Driving/58aed1a8-44d5-4314-9f37-1456921ac406_content_list.json @@ -0,0 +1,1593 @@ +[ + { + "type": "text", + "text": "SplatAD: Real-Time Lidar and Camera Rendering with 3D Gaussian Splatting for Autonomous Driving", + "text_level": 1, + "bbox": [ + 91, + 128, + 908, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Georg Hess $^{\\dagger,1,2}$ Carl Lindström $^{\\dagger,1,2}$ Maryam Fatemi $^{1}$ Christoffer Petersson $^{1,2}$ Lennart Svensson $^{2}$", + "bbox": [ + 99, + 200, + 937, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Zenseact $^{2}$ Chalmers University of Technology", + "bbox": [ + 339, + 222, + 694, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{firstname_lastname}@{zenseact.com, chalmers.se}", + "bbox": [ + 307, + 241, + 725, + 256 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/19714502aa07182b3219d434c13f7483494374467b078db7a0a7adf3d8faa014.jpg", + "image_caption": [ + "Figure 1. SplatAD is the first method capable of realistic camera and lidar rendering using 3D Gaussian Splatting. Whereas previous methods are either fast or multi-modal, SplatAD enables real-time, high-quality rendering for both camera and lidar. In addition, SplatAD can reach competitive performance, e.g., PSNR for image and Chamfer distance for point cloud, within minutes." + ], + "image_footnote": [], + "bbox": [ + 93, + 292, + 295, + 396 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/391b65fc110a83970041c0db396b49d439b7bd8ff76cf9715a9ed4edfb8efdcd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 297, + 292, + 496, + 396 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8fb6c77975566fdd0f4afdeb0be7cbcb33e13040dcf3f5679d9b54e1034459ea.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 292, + 700, + 396 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e3b1848f205a306c38b098b596956599caf495e37462f7d9e6386292040a36da.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 702, + 292, + 903, + 396 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 464, + 326, + 479 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ensuring the safety of autonomous robots, such as self-driving vehicles, requires extensive testing across diverse driving scenarios. Simulation is a key ingredient for conducting such testing in a cost-effective and scalable way. Neural rendering methods have gained popularity, as they can build simulation environments from collected logs in a data-driven manner. However, existing neural radiance field (NeRF) methods for sensor-realistic rendering of camera and lidar data suffer from low rendering speeds, limiting their applicability for large-scale testing. While 3D Gaussian Splatting (3DGS) enables real-time rendering, current methods are limited to camera data and are unable to render lidar data essential for autonomous driving. To address these limitations, we propose SplatAD, the first 3DGS-based method for realistic, real-time rendering of dynamic scenes for both camera and lidar data. SplatAD accurately models key sensor-specific phenomena such as rolling shutter effects, lidar intensity, and lidar ray dropouts, using purpose-built algorithms to optimize rendering efficiency. Evaluation across three autonomous driving datasets demonstrates that SplatAD achieves state-of-the-art rendering quality with up to $+2$ PSNR for NVS and $+3$ PSNR for reconstruction while increasing rendering speed over NeRF-based methods by an order of magnitude. See here for our project page.", + "bbox": [ + 86, + 498, + 485, + 861 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 514, + 464, + 643, + 479 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Large-scale testing is essential to ensure the safety of autonomous robots, such as self-driving vehicles (SDVs), before real-world deployment. Data-driven methods that generate digital twins from collected logs offer a scalable way to build diverse, realistic simulation environments for testing. Unlike real-world testing, which can be costly, time-intensive, and limited by physical constraints, simulation enables rapid, low-cost exploration of multiple scenarios, helping to optimize SDVs for safety, comfort, and efficiency. Motivated by this, a multitude of methods have emerged based on neural radiance fields (NeRFs) [25, 35, 36, 45] and 3D Gaussian Splatting (3DGS) [14, 43, 51].", + "bbox": [ + 509, + 491, + 908, + 672 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent NeRF-based methods [35, 45] provide high-fidelity sensor simulation jointly for both camera and lidar, matching the most common sensor rigs in popular autonomous driving (AD) datasets [1, 5, 38]. However, the slow rendering speed of NeRF-based methods makes them costly and challenging to use for large-scale testing. 3DGS provides an attractive alternative to NeRFs, as its accelerated rendering achieves comparable image realism while increasing inference speeds by an order of magnitude. Nevertheless, 3DGS-based methods for the AD setting [7, 43, 51] inherit the limitation of only being able to render camera data, overlooking the lidar modality. The lidar's ability to directly perceive the 3D environment makes it a powerful tool in modern AD stacks and, consequently, an important modality to simulate.", + "bbox": [ + 509, + 674, + 910, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "These authors contributed equally to this work.", + "bbox": [ + 107, + 893, + 364, + 907 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "11982", + "bbox": [ + 478, + 946, + 519, + 959 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we aim to solve efficient, differentiable, and realistic rendering of camera and lidar data using 3D Gaussian Splatting. Applying 3DGS to lidar sensors presents unique challenges due to their distinct characteristics. First, unlike cameras, which capture dense, regularly spaced pixels, lidars record sparse, non-linear point clouds with large empty gaps from non-returning rays. Second, most lidars capture a full $360^{\\circ}$ view of the scene, where existing methods often project data into multiple depth images [7]—an inefficient approach that overlooks the sparse structure of the lidar. Last, lidars suffer from a rolling shutter effect, with each sweep taking up to $100\\mathrm{ms}$ , during which the ego vehicle may move several meters, violating 3DGS's single-origin assumption.", + "bbox": [ + 89, + 90, + 480, + 287 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To overcome these challenges, we introduce SplatAD, a novel view synthesis method that unifies camera and lidar rendering and is designed for real-time rendering of large-scale dynamic traffic scenes. Our method replaces spherical harmonics with per-Gaussian learnable features to jointly model sensor-specific phenomena—from lidar ray drop and intensity variations to camera-specific appearance changes. By introducing purpose-built rasterization algorithms for rendering lidar in spherical coordinates, we achieve both superior efficiency and realism. Further, we demonstrate efficient methods for modeling rolling shutter effects across both sensor modalities. We verify the effectiveness and generalizability of our method across three popular automotive datasets, achieving state-of-the-art results on all of them. To summarize, our contributions are as follows:", + "bbox": [ + 89, + 287, + 482, + 513 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose the first method for efficient lidar rendering using 3D Gaussians, introducing custom CUDA-accelerated algorithms for rasterizing sparse point clouds in spherical coordinates.", + "- We introduce the first 3DGS method capable of rendering both camera and lidar from a unified representation, enabling accelerated scaling of novel view synthesis for automotive applications.", + "- We present effective techniques for realistic sensor modeling using 3D Gaussians, enabling accurate handling of rolling shutter, lidar intensity, ray dropping, and variations in sensor appearance.", + "- Through extensive evaluation on three popular automotive datasets, we demonstrate state-of-the-art results across all benchmarks, validating our method's effectiveness and generalizability." + ], + "bbox": [ + 91, + 513, + 482, + 755 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 91, + 768, + 227, + 784 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "NeRFs for automotive data: Since the introduction of NeRFs, neural representations have become central to 3D reconstruction and novel view synthesis [2-4, 25]. Many works have applied NeRF-based methods to automotive data [35, 36, 44, 45], enabling sensor-realistic renderings of novel views in large and dynamic scenes. Recent advances provide realism sufficient for downstream applica", + "bbox": [ + 89, + 794, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tions [22, 23]. While early methods focused on either camera [10, 19, 29, 42] or lidar [13, 39, 49], newer methods aim to handle them jointly. UniSim [45] showcased realistic renderings for front camera and $360^{\\circ}$ lidar on data from PandaSet [41]. The method uses a hash grid representation [27], with separate features for the sky, static background, and dynamic actors, and decodes the color and lidar intensity from volume-rendered features. NeuRAD [35] proposes a simplified network architecture and improved sensor modeling, achieving state-of-the-art results for the full $360^{\\circ}$ camera and lidar rig. Even so, the slow rendering speeds of NeRF-based methods make scaling costly and challenging. SplatAD aims to overcome this limitation by using rasterization-based CUDA-accelerated algorithms to render both camera and lidar data while, inspired by NeuRAD, emphasizing the modeling of important sensor characteristics for increased realism.", + "bbox": [ + 511, + 90, + 906, + 330 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3DGS for automotive data: 3DGS [14] uses an explicit representation paired with rasterization techniques and custom hardware-accelerated algorithms to achieve real-time rendering. Additionally, 3DGS has been shown to create faithful scene reconstructions with rendering quality close to recent NeRF-based methods. As a result, several works have applied 3DGS for camera rendering on automotive data [7, 15, 43, 50, 51]. Periodic Vibration Gaussian (PVG) [6] applies 3DGS to dynamic scenes by learning the 3D flow of all Gaussians. Yet, PVG's lack of an explicit actor representation limits the method's controllability and, hence, applicability for simulation. Street Gaussians [43], like Unisim and NeuRAD, instead decomposes the scene into static background and rigid dynamic actors using 3D bounding boxes. The authors further extend 3DGS by adding Fourier coefficients to capture time-varying characteristics of dynamic actors. Moreover, to handle the sky regions, their method uses semantic masks and a view-dependent cube map. OmniRe [7] enhances the scene graph composition with non-rigid nodes for better modeling of pedestrians and cyclists. However, this requires pre-processing to track and align human body poses [12] for initializing SMPL [24] pose parameters.", + "bbox": [ + 511, + 332, + 908, + 665 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "PVG, Street Gaussians, and OmniRe all use lidar points for initialization and depth supervision. However, to generate new point clouds, these methods predict depth images and rely on existing lidar points to query said image at the right locations. This is both inefficient and inaccurate, as $360^{\\circ}$ lidars are approximated as multiple cameras. Further, neither of these methods model important lidar characteristics such as intensity variations, ray drop, and rolling shutter. In contrast, our method supports efficient rendering of both camera and lidar from 3D Gaussians in novel views, and models important sensor characteristics for both modalities.", + "bbox": [ + 511, + 665, + 908, + 830 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 511, + 844, + 604, + 859 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our aim is to learn a scene representation from collected vehicle logs that enables rendering of realistic camera and lidar", + "bbox": [ + 511, + 869, + 906, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "11983", + "bbox": [ + 478, + 946, + 517, + 958 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/3536cb69060a185c55a3bdfeba2bb4a0a85014e2298f4ddbd805a9979bdf1cc2.jpg", + "image_caption": [ + "Figure 2. Overview of our proposed method. Given the composition of static and dynamic 3D Gaussians, SplatAD is capable of differentiable rendering of both lidar and camera data. Our proposed lidar rendering matches the image rendering on a high level, but modifies each component to accurately model sensor characteristics. Our method projects 3D Gaussians with associated feature vectors onto the corresponding sensor modalities (camera and lidar) and employs sensor-specific tiling to match their distinct characteristics. During rasterization, the projected Gaussians are corrected for rolling shutter effects caused by the movement of sensors and, potentially, their own velocity. Finally, the rasterized features are decoded into the respective image and lidar point cloud representations." + ], + "image_footnote": [], + "bbox": [ + 96, + 89, + 911, + 327 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "data, with the ability to alter the locations of both the ego vehicle and other actors. To scale effectively, the rendering process must be fast, as faster inference speeds enhance its practicality for applications. In the following, we describe our scene representation (Sec. 3.1), rendering algorithms (Secs. 3.2 and 3.3), and implementation and optimization strategy (Sec. 3.4), see Fig. 2 for an overview.", + "bbox": [ + 88, + 446, + 480, + 553 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Scene representation", + "text_level": 1, + "bbox": [ + 89, + 561, + 287, + 575 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our scene representation builds upon 3DGS [14], but with key changes to handle the specifics of AD scenes and to enable both camera and lidar rendering from the same representation. Like 3DGS, each scene is represented by a set of translucent 3D Gaussians with learnable occupancy $o \\in (0,1)$ , mean $\\mu \\in \\mathbb{R}^3$ , and covariance matrix $\\Sigma \\in \\mathbb{R}^{3\\times 3}$ which in turn is parameterized by scale $\\mathbf{S} \\in \\mathbb{R}^3$ and quaternion $\\mathbf{q} \\in \\mathbb{R}^4$ parameters. Instead of spherical harmonics [14], we assign each Gaussian a learnable base color $\\mathbf{f}_{\\mathrm{rgb}}^{\\mathrm{b}} \\in \\mathbb{R}^3$ and feature vector $\\mathbf{f} \\in \\mathbb{R}^{D_{\\mathrm{f}}}$ , where $\\mathbf{f}$ is used for both view-dependent effects and lidar properties. Last, our representation contains a learnable embedding per sensor to model their specific appearance characteristics.", + "bbox": [ + 88, + 583, + 482, + 779 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To handle dynamics, we follow the commonly used scene graph decomposition [29, 35, 43, 51] and divide the scene into a static background and a set of dynamic actors. Each dynamic actor is described by a 3D bounding box and a sequence of SE(3) poses, obtained either from an off-the-shelf object detector and tracker, or annotations. Each Gaussian has a non-learnable ID, indicating whether it is assigned to the static world, or to which actor it belongs. For Gaussians", + "bbox": [ + 88, + 780, + 482, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "assigned to actors, their mean and covariance are expressed in the local coordinate system of the corresponding axis-aligned bounding box. To compose the scene at a given time $t$ , Gaussians assigned to bounding boxes are transformed to world coordinates based on their actor's pose. Since pose estimates for actors may contain inaccuracies, we adjust these with learnable offsets. Furthermore, each actor has a velocity initialized from pose differences and a learnable velocity offset.", + "bbox": [ + 511, + 446, + 906, + 583 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Camera rendering", + "text_level": 1, + "bbox": [ + 511, + 590, + 692, + 607 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given a posed camera, we compose the set of Gaussians at the corresponding capture time $t$ and render an image $I$ using the efficient tile-based rendering from 3DGS [14]. While we retain 3DGS's high-level steps—projection and view frustum culling, tile-assignment, depth sorting, and tile-based rasterization—we introduce key adaptations to better model the unique characteristics of AD data.", + "bbox": [ + 511, + 613, + 906, + 718 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Projection, tiling, and sorting: Each mean and covariance is transformed from world to camera coordinates, giving $\\pmb{\\mu}^{C}$ and $\\pmb{\\Sigma}^{C}$ . The means are then converted to image space $\\pmb{\\mu}^{I} \\in \\mathbb{R}^{2}$ using the perspective projection, while the covariances are transformed using the two upper rows of the Jacobian of the projection, $\\pmb{\\Sigma}^{I} = \\mathbf{J}^{I}\\pmb{\\Sigma}^{C}(\\mathbf{J}^{I})^{\\mathrm{T}} \\in \\mathbb{R}^{2 \\times 2}$ . Gaussians outside the view frustum are culled, where, for efficiency, 3DGS approximates the extent of the Gaussians using a square axis-aligned bounding box (AABB) that covers their $99\\%$ confidence level. Further, 3DGS divides the image into tiles, each measuring $16 \\times 16$ pixels, and assigns Gaussians to all tiles they intersect, duplicating them if necessary. This", + "bbox": [ + 511, + 719, + 908, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "11984", + "bbox": [ + 478, + 946, + 517, + 958 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "way, each pixel only has to process a subset of all Gaussians during rasterization. Lastly, the Gaussians are sorted based on the $z$ -depth of their mean $\\pmb{\\mu}^{C}$ .", + "bbox": [ + 89, + 90, + 480, + 136 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Rolling shutter: Many cameras use a rolling shutter, where the capture of an image is not instantaneous, but is done row-by-row, allowing the camera to move during the exposure. Previous work [35] has highlighted the importance of modeling rolling shutter effects due to the high sensor velocities present in AD data, where their ray-tracing-based method can easily account for this by shifting the origins of each ray. For 3DGS, the equivalent would require projecting all 3D Gaussians to all camera poses encountered during an exposure, as the Gaussians' positions relative to the camera change with time. As this would be prohibitively expensive, we draw inspiration from [31] and instead approximate the effects directly in 2D image space. Specifically, each Gaussian's velocity relative to the camera is projected to image space, and their pixel mean $\\mu^I$ is adjusted during rasterization based on the capture time of the pixel. However, [31] considered static scenes only, hence we adjust the formulation to account for the dynamics as well.", + "bbox": [ + 89, + 136, + 483, + 407 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For each Gaussian, its pixel velocity is approximated as", + "bbox": [ + 109, + 407, + 478, + 422 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {v} ^ {I} = \\mathbf {J} ^ {I} \\left(- \\boldsymbol {\\omega} _ {C} \\times \\boldsymbol {\\mu} ^ {C} - \\mathbf {v} _ {C} + \\mathbf {v} _ {\\mathrm {d y n}}\\right) \\in \\mathbb {R} ^ {2}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 143, + 431, + 482, + 450 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\omega_{C}$ and $\\mathbf{v}_C$ denote the angular and linear velocity of the camera expressed in the camera coordinate system, and", + "bbox": [ + 89, + 458, + 482, + 488 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {v} _ {\\mathrm {d y n}} = T ^ {\\mathrm {a c t} \\rightarrow C} \\left(\\boldsymbol {\\omega} _ {\\mathrm {a c t}} \\times \\boldsymbol {\\mu} ^ {\\mathrm {a c t}} + \\mathbf {v} _ {\\mathrm {a c t}}\\right) \\in \\mathbb {R} ^ {3}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 496, + 482, + 513 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "denotes the velocity of the Gaussian induced by the angular $\\omega^{\\mathrm{dyn}}$ and linear velocity $\\mathbf{v}^{\\mathrm{dyn}}$ of its associated actor (expressed in actor coordinate system) and $\\mu^{\\mathrm{act}}$ denotes the Gaussian mean in the actor coordinate system. To transform velocities from actor to camera coordinates, $T^{\\mathrm{act}\\rightarrow C}$ is the composition of the actor-to-world and world-to-camera transforms. For Gaussians that are part of the static background, $\\mathbf{v}_{\\mathrm{dyn}}$ is zero. See Appendix D for further details.", + "bbox": [ + 89, + 523, + 483, + 643 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We account for the pixel velocity when culling Gaussians and checking intersections between Gaussians and tiles by increasing the approximated Gaussian extent. Instead of the square AABB used by 3DGS and [31], we use a rectangular AABB covering $\\boldsymbol{\\Sigma}^{I}$ at three standard deviations and add $|\\mathbf{v}_I| \\cdot t_{\\mathrm{rs}} / 2$ to its extent. Here, $t_{\\mathrm{rs}}$ denotes the rolling shutter duration, i.e., the time difference between the last and first row of pixels. The increase in extent corresponds to the area covered by the Gaussian mean during the rolling shutter time, assuming the sensor timestamp is in the middle of exposure. As this is not always the case, we further include a learnable time offset, describing the difference between the sensor timestamp and the middle of the exposure. Using rectangular AABBs reduces unnecessary intersections compared to [14, 31], especially for narrow Gaussians and Gaussians with axis-aligned velocities, the latter often being the case for side-facing cameras.", + "bbox": [ + 89, + 645, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Rasterization: 3DGS rasterizes pixels in parallel by launching a thread block per tile and assigning each thread to a pixel within the tile. For each pixel, its coordinates $\\mathbf{p} = [p_u,p_v]^{\\mathrm{T}}$ are inferred from the block and the thread index. From this, we find the time difference between the pixel's capture time and the image's middle row as", + "bbox": [ + 511, + 90, + 906, + 181 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nt _ {\\mathrm {p i x}} = \\left(p _ {v} / H - 0. 5\\right) \\cdot t _ {\\mathrm {r s}}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 622, + 194, + 906, + 210 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To rasterize a pixel, we $\\alpha$ -blend RGB values $\\mathbf{f}_i^{\\mathrm{rgb}}$ and features $\\mathbf{f}_i$ of the depth-sorted Gaussians intersecting the current tile", + "bbox": [ + 511, + 222, + 903, + 253 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\left[ \\mathbf {F} _ {\\mathbf {p}} ^ {\\mathrm {r g b}}, \\mathbf {F} _ {\\mathbf {p}} \\right] = \\sum_ {i} \\left[ \\mathbf {f} _ {i} ^ {\\mathrm {r g b}}, \\mathbf {f} _ {i} \\right] \\alpha_ {i} \\prod_ {j = 1} ^ {i - 1} (1 - \\alpha_ {j}), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 573, + 263, + 906, + 306 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "with $\\alpha$ computed as", + "bbox": [ + 511, + 316, + 648, + 332 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\alpha_ {i} = \\sqrt {\\frac {\\left| \\boldsymbol {\\Sigma} _ {i} ^ {I} \\right|}{\\left| \\boldsymbol {\\Sigma} _ {i} ^ {I} + s \\mathbf {I} \\right|}} o _ {i} \\exp \\left(- \\frac {1}{2} \\boldsymbol {\\Delta} _ {i} ^ {\\mathrm {T}} \\left(\\boldsymbol {\\Sigma} _ {i} ^ {I} + s \\mathbf {I}\\right) ^ {- 1} \\boldsymbol {\\Delta} _ {i}\\right) \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 344, + 906, + 383 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "and", + "bbox": [ + 513, + 393, + 542, + 406 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\Delta} _ {i} = \\mathbf {p} - \\left(\\boldsymbol {\\mu} _ {i} ^ {I} + \\mathbf {v} _ {i} ^ {I} t _ {\\mathrm {p i x}}\\right). \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 622, + 407, + 906, + 426 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As in [31], we calculate the distance $\\Delta_{i}$ between pixels and Gaussians using the means compensated by rolling shutter, effectively shifting Gaussians to the correct location in the image. Compared to 3DGS, in Eq. (5) we use the 3DGS+EWA [52] formulation from [47], where $s = 0.3$ .", + "bbox": [ + 511, + 431, + 906, + 506 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "View-dependent effects are modeled using a small CNN; given the feature map $\\mathbf{F} \\in \\mathbb{R}^{H \\times W \\times D_{\\mathrm{f}}}$ , the corresponding ray directions $\\mathbf{d} \\in \\mathbb{R}^{H \\times W \\times 3}$ , and a camera-specific learned embedding $\\mathbf{e}$ , it predicts a pixel-wise affine mapping $M \\in \\mathbb{R}^{H \\times W \\times 3}$ and $\\mathbf{b} \\in \\mathbb{R}^{H \\times W \\times 3}$ applied to $\\mathbf{F}^{\\mathrm{rgb}}$", + "bbox": [ + 511, + 507, + 905, + 580 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nM, \\mathbf {b} = \\operatorname {C N N} (\\operatorname {c o n c a t} (\\mathbf {F}, \\mathbf {d}, \\mathbf {e})), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 602, + 594, + 906, + 611 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nI = (1 + M) \\odot \\mathbf {F} ^ {\\mathrm {r g b}} + \\mathbf {b}. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 627, + 614, + 906, + 631 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We found that using a small CNN instead of an MLP is essential for efficient texture modeling, e.g., for high-frequency road surfaces. In addition, inspired by [30], the camera-specific learned embeddings enable modeling differences in exposure.", + "bbox": [ + 511, + 642, + 906, + 717 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Lidar rendering", + "text_level": 1, + "bbox": [ + 511, + 727, + 671, + 743 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Lidar sensors enable self-driving vehicles to measure the distance and reflectivity (intensity) of a discrete set of points. They do so by emitting laser beam pulses and measuring the time of flight to determine distance and returning power for reflectivity. Most AD datasets deploy lidars that use several laser diodes (16-128) mounted in a vertical array, where the array of diodes is rotated to capture $360^{\\circ}$ data and a single scan often takes $100\\mathrm{ms}$ to capture. Therefore, we focus on this type of lidar. However, we note that modeling other types of lidars, such as solid-state lidars [21], can be easily", + "bbox": [ + 511, + 750, + 906, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "11985", + "bbox": [ + 478, + 946, + 517, + 958 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "done in our method by modifying the projection accordingly. Our proposed lidar rendering follows the high-level steps of the image rendering, but we modify each component to accurately model lidar characteristics.", + "bbox": [ + 89, + 90, + 483, + 150 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Projection: For a posed lidar and Gaussians composed at its capture time $t$ , we transform each mean and covariance from world to lidar coordinates, yielding $\\boldsymbol{\\mu}^{L} = [x,y,z]^{\\mathrm{T}}$ and $\\boldsymbol{\\Sigma}^{L}$ . Raw lidar data are generally captured in spherical coordinates azimuth $\\phi$ , elevation $\\omega$ , and range $r$ . Thus, we transform Gaussian means from lidar coordinates to spherical coordinates", + "bbox": [ + 89, + 151, + 482, + 261 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\mu} ^ {S} = \\left[ \\begin{array}{l} \\phi \\\\ \\omega \\\\ r \\end{array} \\right] = \\left[ \\begin{array}{c} \\arctan 2 (y, x) \\\\ \\arcsin (z / r) \\\\ \\sqrt {x ^ {2} + y ^ {2} + z ^ {2}} \\end{array} \\right], \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 166, + 268, + 483, + 321 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "and convert the covariance $\\pmb{\\Sigma}^{L}$ with the Jacobian of the transform Eq. (9) as $\\pmb{\\Sigma}^{S} = \\mathbf{J}^{S}\\pmb{\\Sigma}^{L}(\\mathbf{J}^{S})^{\\mathrm{T}}$ , where", + "bbox": [ + 89, + 332, + 482, + 364 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {J} ^ {S} = \\left[ \\begin{array}{c c c} - \\frac {y}{x ^ {2} + y ^ {2}} & \\frac {x}{x ^ {2} + y ^ {2}} & 0 \\\\ - \\frac {x z}{r ^ {2} \\sqrt {x ^ {2} + y ^ {2}}} & - \\frac {y z}{r ^ {2} \\sqrt {x ^ {2} + y ^ {2}}} & \\frac {\\sqrt {x ^ {2} + y ^ {2}}}{r ^ {2}} \\\\ \\frac {x}{r} & \\frac {y}{r} & \\frac {z}{r} \\end{array} \\right], \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 111, + 375, + 483, + 441 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "and $\\pmb{\\mu}^{S}$ and $\\pmb{\\Sigma}^{S}$ denote mean and covariance in spherical coordinates.", + "bbox": [ + 89, + 450, + 482, + 479 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Same as for the camera, we account for the rolling shutter effect by approximating its impact in sensor space", + "bbox": [ + 89, + 482, + 483, + 513 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {v} ^ {S} = \\mathbf {J} ^ {S} \\left(- \\boldsymbol {\\omega} _ {L} \\times \\boldsymbol {\\mu} ^ {L} - \\mathbf {v} _ {L} + \\mathbf {v} _ {\\mathrm {d y n}}\\right) \\in \\mathbb {R} ^ {3}. \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 143, + 522, + 482, + 542 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Note that $\\mathbf{v}^S$ is three-dimensional as opposed to $\\mathbf{v}^I$ in Eq. (1), as we need this for accurate range rasterization. Last, we use the first two elements of $\\mathbf{v}^S t_{\\mathrm{rs}} / 2$ to increase the extent of the rectangular AABB approximating the size of the Gaussian, and cull Gaussians outside the lidar's field-of-view.", + "bbox": [ + 89, + 551, + 483, + 627 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Tiling and sorting: At the core of 3DGS's efficiency is its tile-based rasterization. While lidars often have a fixed azimuth resolution, many use non-equidistant spacing between diodes to get higher resolution in areas of interest, resulting in nonlinear elevation resolution, as seen in Fig. 1. In line with this, we design tiles with a fixed number of diodes vertically and a fixed azimuth resolution horizontally. Horizontally, intersecting tiles are found by dividing azimuth coordinates by the common tile azimuth size and wrapping the results within $360^{\\circ}$ . Vertically, we sort the elevation tile boundaries and find intersections with a short for-loop, see Appendix A for details. After tile assignment, Gaussians are sorted based on the range of their mean $\\mu^S$ .", + "bbox": [ + 89, + 628, + 483, + 824 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Comparing our strategy to using tiles of equal size, similar to the depth image-based rendering in [7], we avoid many unnecessary computations. There, the tile size must be determined from the finest resolution to ensure having at least as many threads as lidar points. Consequently, tiles", + "bbox": [ + 89, + 825, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "in areas where the lidar is sparse would have a superficially high resolution and might even lack the corresponding lidar points to render.", + "bbox": [ + 511, + 90, + 905, + 135 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Rasterization: Lidar data are rendered using our proposed differentiable tile-based rasterizer. Again, each tile launches one thread block, where each thread is responsible for rasterizing one lidar point. Each thread is provided the azimuth, elevation, and capture time for it to rasterize. These are either determined from lidar specifications or, for training, calculated from existing point clouds, see Appendix A for details. The reason for this is that for training, one cannot expect the data to follow perfectly straight scan lines. Therefore, we cannot follow the image strategy and rely on the block and thread index to directly infer the pixel coordinates. Further, the azimuth and elevation of a lidar ray do not necessarily determine its time offset to the center of the lidar scan.", + "bbox": [ + 511, + 137, + 906, + 332 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For each lidar point, we $\\alpha$ -blend the Gaussians' features $\\mathbf{f}_i$ using Eqs. (4) to (6). Inspired by [47] arguing for $s$ in Eq. (5) to correspond to one pixel, we set it to correspond to the geometric mean of the lidar's vertical and horizontal beam divergence. The rendered features are concatenated with their corresponding ray direction and decoded to intensity and ray drop probability using a small MLP.", + "bbox": [ + 511, + 334, + 905, + 439 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To render the expected range of a point, we $\\alpha$ -blend the rolling shutter corrected ranges for each Gaussian", + "bbox": [ + 511, + 440, + 905, + 470 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nr _ {i, \\mathrm {r s}} = r _ {i} + \\mathbf {v} _ {r} ^ {S} t _ {l}, \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 648, + 482, + 906, + 501 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{v}_r^S$ is the range-component of the relative speed to the lidar and $t_l$ is the time between the capture of the current lidar point, and the middle of the lidar scan. Further, we render the median range as the rolling shutter corrected range of the first Gaussian that satisfies $\\prod_{j=1}^{i}(1 - \\alpha_j) < 0.5$ during the $\\alpha$ -blending in Eq. (4). While we use the expected range for training, the median range is used during inference as it, in contrast to the expected range, does not yield depths between Gaussians.", + "bbox": [ + 511, + 513, + 905, + 648 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Optimization and implementation", + "text_level": 1, + "bbox": [ + 511, + 661, + 810, + 678 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "All model components are optimized jointly using the loss", + "bbox": [ + 511, + 684, + 905, + 700 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} = \\lambda_ {r} \\mathcal {L} _ {1} + (1 - \\lambda_ {r}) \\mathcal {L} _ {\\text {S S I M}} + \\lambda_ {\\text {d e p t h}} \\mathcal {L} _ {\\text {d e p t h}} + \\lambda_ {\\text {l o s}} \\mathcal {L} _ {\\text {l o s}} + \\\\ \\lambda_ {\\text {i n t e n s}} \\mathcal {L} _ {\\text {i n t e n}} + \\lambda_ {\\text {r a y d r o p}} \\mathcal {L} _ {\\text {B C E}} + \\lambda_ {\\text {M C M C}} \\mathcal {L} _ {\\text {M C M C}}, \\tag {13} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 529, + 731, + 906, + 768 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_1$ and $\\mathcal{L}_{\\mathrm{SSIM}}$ are L1 and SSIM losses on the rendered images. $\\mathcal{L}_{\\mathrm{depth}}$ and $\\mathcal{L}_{\\mathrm{inten}}$ are L2 losses on the rendered expected lidar range and intensity. $\\mathcal{L}_{\\mathrm{los}}$ is a line-of-sight loss, penalizing opacity before the ground truth lidar range. $\\mathcal{L}_{\\mathrm{BCE}}$ is a binary cross-entropy loss on the predicted ray drop probability, where ground-truth is generated in the same way as for NeuRAD. $\\mathcal{L}_{\\mathrm{MCMC}}$ is the opacity and scale regularization used in [16].", + "bbox": [ + 511, + 779, + 906, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "11986", + "bbox": [ + 478, + 946, + 519, + 958 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/9b7e4808099cb50ebee8698abe578747e00b562dbcd15da935fdbc3734773916.jpg", + "image_caption": [ + "Figure 3. Compared to the baselines, SplatAD produces sharp images with a high level of detail. Further, the bottom row highlights the superiority of our lidar rendering. Projecting lidar points into images for depth supervision, as used by previous 3DGS methods, causes line-of-sight errors and incorrect volume carving due to the pose differences between camera and lidar." + ], + "image_footnote": [], + "bbox": [ + 93, + 88, + 903, + 325 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We initialize the set of Gaussians using both lidar points and random points. Lidar points within bounding boxes are assigned to the corresponding actor. Further, all lidar points are projected into their closest image to set their initial color. The random points are divided into (1) uniformly sampled points within the lidar range and (2) points sampled linearly in disparity beyond the lidar range, both using random colors at initialization. Furthermore, instead of the 3DGS [14] splitting and densification strategy, we rely on the MCMC strategy introduced in [16]. We chose MCMC partially because we find it to improve far-field rendering quality and partially because it has predictable computation requirements, as it permits choosing the maximum number of Gaussian. We use [16] as is, without any special treatment of Gaussians assigned to dynamic actors. See Appendix B for further training details.", + "bbox": [ + 88, + 402, + 485, + 645 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation: For efficiency, we implement the forward and backward passes of the rolling shutter compensation and lidar projection and rasterization using custom CUDA kernels. SplatAD builds upon the collaborative open-source frameworks gsplat [46] and neurad-studio [34], and we hope that our contributions can benefit future research on multi-modal simulation, not only within the AD community. We train SplatAD for 30,000 iterations using the Adam optimizer, which takes an hour on a single NVIDIA A100. See Appendix B for further details.", + "bbox": [ + 89, + 646, + 485, + 797 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 814, + 223, + 830 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To validate the robustness of our method, we evaluate it across multiple popular AD datasets, using the same set of hyperparameters. For each dataset, we compare SplatAD to the best performing NeRF and 3DGS-based methods tailored", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/4bd13918f7d480ebe31f42ddf0c2f51d5415f113c397b8fddd30a1420ac62b0f.jpg", + "table_caption": [ + "Table 1. NVS results for image, over three datasets. First, second, third." + ], + "table_footnote": [], + "table_body": "
PSNR ↑SSIM ↑LPIPS ↓MP/s ↑Train (H) ↓
PandaSetUniSim23.120.6820.36011.50.9
NeuRAD25.800.7530.2509.71.3
PVG24.010.7120.45222.61.6
Street-GS24.730.7450.31463.12.3
OmniRE24.710.7450.31553.62.6
SplatAD (ours)26.700.8140.193104.11.0
Argo2UniSim22.350.6550.45810.01.1
NeuRAD26.180.7210.3109.71.3
PVG24.470.7120.50219.72.5
Street-GS25.520.7540.37492.82.6
OmniRE25.610.7530.37575.63.0
SplatAD (ours)28.400.8260.271133.81.2
nuScenesUniSim22.660.7430.41413.30.7
NeuRAD26.170.7900.31211.81.1
PVG24.940.7680.49718.81.4
Street-GS25.540.7990.37557.91.6
OmniRE25.500.7980.38646.91.9
SplatAD (ours)27.530.8490.302104.31.2
", + "bbox": [ + 542, + 445, + 874, + 638 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "to AD data on novel view synthesis for both camera and lidar data. Further, we ablate key elements of SplatAD and measure the impact they have on sensor realism and rendering speed.", + "bbox": [ + 511, + 670, + 906, + 733 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets: We perform experiments on PandaSet [41], Argoverse2 [38] and nuScenes [5]. These datasets vary in lidars, camera appearance, and resolution. For PandaSet and nuScenes, we use the six available cameras. For Argoverse2, we use the seven ring cameras, leaving out the black-and-white stereo cameras. Further, some cameras are cropped slightly to remove views of the ego-vehicle, such as the hood and the trunk. We use the same 10 challenging sequences per dataset as in [35], containing different types of illumination, dynamic actors, and velocities. We also follow the evaluation protocol of [35] i.e., using full resolution data,", + "bbox": [ + 511, + 734, + 910, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "11987", + "bbox": [ + 478, + 946, + 519, + 959 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/aa1652697f1d36c050ba5ec07f6d612a2cebaff2fe0920073068f804c0ede250.jpg", + "table_caption": [ + "Table 2. NVS results for lidar, over three datasets. ${}^{8}$ without missing points. First ,second ,third ." + ], + "table_footnote": [], + "table_body": "
Depth ↓Intensity ↓Drop acc. ↑CD ↓MR/s ↑
PandaSetUniSim0.080.086-10.3$0.9
NeuRAD0.010.06396.21.91.1
PVG38.74--125.2$0.3
Street-GS6.18--37.3$0.8
OmniRE2.88--29.8$0.6
SplatAD (ours)0.010.05996.71.619.5
Argo2UniSim0.180.081-29.2$0.7
NeuRAD0.020.05892.22.60.9
SplatAD (ours)0.020.05292.62.89.5
nuScenesUniSim0.060.063-52.2$1.0
NeuRAD0.010.04293.16.31.1
PVG10.34--77.9$0.05
Street-GS1.56--4.5$0.2
OmniRE1.71--4.6$0.1
SplatAD (ours)0.020.03693.81.75.7
", + "bbox": [ + 120, + 128, + 450, + 292 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "and, for novel view synthesis (NVS), use every other frame for training and the remaining frames for hold-out validation. We believe that this setting is more challenging than what is often encountered in previous work [6, 7, 43], where fewer datasets are considered, images are downsampled, and a larger fraction of the data is used for training.", + "bbox": [ + 88, + 318, + 482, + 409 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Baselines: We compare our results with the best performing NeRF and 3DGS methods on AD data, namely, UniSim [45], NeuRAD [35], PVG [6], Street Gaussians [43] and OmniRe [7]. Note that all methods are designed for dynamic scenes and use lidar data for supervision. We use NeuRAD's official implementation neurad-studio [34] and their version of UniSim, as well as OmniRe's official implementation drivestudio [8] and their implementations of SVG and Street Gaussians. We use drivestudio to generate the required sky masks for Street Gaussians and OmniRe, as well as extract human pose estimates, which takes about 10 and 30 minutes per sequence on an A100.", + "bbox": [ + 88, + 410, + 482, + 590 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Although point cloud rendering is shown in [7], there exists no feature for it in drivestudio. To render point clouds in training and validation views, we render depth images by placing 6 virtual cameras in the lidar origin, each with a horizontal FOV of $60^{\\circ}$ . Their vertical FOV is set to capture all lidar points in the sweep, and the focal length is the median focal length for each dataset. We project ground truth lidar points into these images, fetch each point's depth, and project them back into 3D. As drivestudio lacks functionality for inferring missing lidar points, these baselines are evaluated only for existing points. In contrast, SplatAD is evaluated on the more challenging task of predicting the range also for missing rays and must filter them based on the predicted ray drop probability. See Appendix C for further baseline details.", + "bbox": [ + 89, + 593, + 485, + 819 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.1. Image rendering", + "text_level": 1, + "bbox": [ + 89, + 832, + 256, + 848 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Novel view synthesis: We assess the rendering quality of images using the standard NVS metrics PSNR, SSIM [37], and LPIPS [48] on hold-out validation images in Tab. 1.", + "bbox": [ + 89, + 854, + 485, + 902 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/adcb9d405382fb770cf80cf64294c796be9a07363c0de4e12795b9de42d707f8.jpg", + "image_caption": [ + "Figure 4. Removing our rolling shutter modeling compensation leads to inaccurate geometries and inconsistencies in the learning." + ], + "image_footnote": [], + "bbox": [ + 516, + 89, + 905, + 252 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We measure speed using resolution-agnostic megapixels per second. SplatAD consistently outperforms existing 3DGS methods in terms of image quality with a large margin for the three considered datasets. Further, our method improves upon NeuRAD's SOTA results in all NVS metrics, while rendering images an order of magnitude faster. We provide a qualitative comparison in Fig. 3.", + "bbox": [ + 511, + 321, + 906, + 426 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Reconstruction: To measure the upper limit of the modeling capacity of the methods, we report reconstruction metrics on PandaSet in Tab. 3. Here, methods train on all data in a sequence and evaluate on the same views. Further, we enable sensor pose optimization for all methods to account for any pose inaccuracies. Comparing to the NVS setting, all methods except UniSim show improvements with the additional training data. SplatAD achieves SOTA results while rendering $\\times 10$ faster than the previous best method. Further, we note that SplatAD's performance on hold-out validation images in Tab. 1 is on par with previous 3DGS methods' results for reconstruction in Tab. 3.", + "bbox": [ + 511, + 428, + 908, + 608 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Extrapolation: We investigate SplatAD's ability to generalize to views that differ significantly from what is encountered during training. We use the models trained on every other frame and follow the three settings in [35]: shift the ego-vehicle horizontally, shift the ego-vehicle vertically, and apply shifts and rotation to all dynamic actors. We report the Fréchet distance using DINOv2 [28] features, as these have been shown [32] to align better with human perception than using Inception-v3 [33]. However, we note that using Inception-v3 features instead does not change the model ranking or our conclusions. Tab. 4 shows SplatAD's ability to learn meaningful representations for generalization, clearly outperforming other 3DGS methods.", + "bbox": [ + 511, + 609, + 910, + 806 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Lidar rendering", + "text_level": 1, + "bbox": [ + 511, + 816, + 674, + 833 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We measure the quality of our lidar point clouds using the same metrics as in [35], i.e., median squared depth error, RMSE intensity error, ray drop accuracy, and chamfer distance, see Tab. 2 for NVS and Tab. 3 for reconstruction. Note", + "bbox": [ + 511, + 839, + 908, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "11988", + "bbox": [ + 480, + 946, + 519, + 959 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/28b59ed19280b1cf5b9c3ec962a2735ac443decbb16912dccce2450a83f0a4f1.jpg", + "table_caption": [ + "Table 3. Reconstruction results for image and lidar point clouds on PandaSet. $\\S$ without missing points. First, second, third." + ], + "table_footnote": [], + "table_body": "
ImageLidarEfficiency
PSNR ↑SSIM ↑LPIPS ↓Depth ↓Intensity ↓Drop acc. ↑CD ↓Camera MP/s ↑Lidar MR/s ↑Train time (H) ↓
PandasetUniSim23.180.6840.3620.060.087-10.2$11.10.80.8
NeuRAD26.760.7780.2330.030.07496.11.79.81.11.3
PVG25.350.7520.42353.18--150.8$23.60.31.6
Street-GS26.680.8080.2933.27--27.9$65.70.82.2
OmniRE26.680.8080.2953.80--33.5$56.70.62.4
SplatAD (ours)29.740.8930.1750.020.05397.41.6104.117.41.0
", + "bbox": [ + 153, + 116, + 841, + 209 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/204cac16f9789402aaf4ce5f7336e69f24ec5ee6c401880234bcec5953b6f355.jpg", + "table_caption": [ + "Table 4. ${\\mathrm{{FD}}}_{\\mathrm{{DINOV}}2}$ scores when shifting pose of ego vehicle or actors. First,second,third ." + ], + "table_footnote": [], + "table_body": "
Ego lane shiftEgo vert. shiftActor shift
0m2m3m1mRot.Trans.
PandayetNeuRAD377.4540.8629.2571.0424.8420.2
PVG618.8782.7885.6913.9618.8618.8
Street-GS497.1718.6824.2868.7554.7558.0
OmniRE504.5723.2833.5872.8569.6573.2
SplatAD (ours)276.3520.7647.2613.2355.1345.1
", + "bbox": [ + 120, + 272, + 450, + 347 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/50a4eaa74700d6d0ce39347c8ae2473f1f3871726edf7e530b98079972d32cd2.jpg", + "table_caption": [ + "Table 5. NVS results averaged over 10 sequences from PandaSet, nuScenes, and Argoverse2 when removing model components." + ], + "table_footnote": [], + "table_body": "
Full modelPSNR ↑SSIM ↑LPIPS ↓Depth ↓CD ↓Camera MP/s ↑Lidar MR/s ↑
27.540.8300.2550.022.0114.111.4
a)Camera RS27.480.8280.2580.022.1116.311.5
b)CNN decoder27.230.8240.2640.022.2178.011.6
c)EWA antialias27.600.8280.2680.021.5114.511.0
d)Appearance emb.27.280.8240.2640.022.1121.011.5
e)Velocity opt.27.550.8300.2570.022.0114.111.4
f)Lidar RS27.420.8270.2600.032.1114.212.2
g)MCMC27.350.8260.2590.012.1130.614.1
h)Median depth27.540.8300.2550.024.9114.111.4
", + "bbox": [ + 94, + 393, + 478, + 513 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c448484c33faf286ff729a8f41f4439392ed19dd276f7245c1e5eb3ce057bbfd.jpg", + "image_caption": [ + "Figure 5. The CNN decoder improves sharpness and is more true to color than the MLP decoder." + ], + "image_footnote": [], + "bbox": [ + 93, + 521, + 256, + 604 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a30645caa572704cc825a30a068dd3d158e63ef9efbdf29017da8cae571f8b09.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 256, + 521, + 480, + 604 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "that we do not report intensity or ray drop accuracy for the 3DGS baselines, as they have no way of inferring these. Further, it was not possible to produce Argoverse2 point cloud results for these baselines due to the memory requirements of drivestudio not scaling well with the increase in number of images and higher image resolution. We find SplatAD to be on par with the accurate ray-tracing formulation of NeuRAD in terms of point cloud quality, while running up to $\\times 18$ faster. Furthermore, our lidar rendering approach is superior to the naive depth image-based method used for 3DGS baselines, both in terms of speed and quality.", + "bbox": [ + 89, + 652, + 483, + 819 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Ablations", + "text_level": 1, + "bbox": [ + 89, + 832, + 202, + 845 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We validate the effectiveness of key components of our method by measuring their impact on NVS metrics in Tab. 5. Metrics are averaged over all three datasets to avoid any", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "dataset-specific biases. We see that removing the rolling shutter modeling of the camera (a) or lidar (f) reduces NVS performance, and that leaving out optimization of sensor velocities (e) negatively impacts our depth predictions. While the impact of rolling shutter modeling on quantitative metrics is somewhat modest, we qualitatively find that it is important for capturing thin structures and edges, see Fig. 4. Removing these components leads to line-of-sight errors, which erroneously can cut through objects. Note that when removing our lidar rolling shutter compensation, we use ego-motion compensated point clouds for supervision. Moreover, we note the efficiency of our rolling shutter algorithms, as removing them barely improves rendering speeds.", + "bbox": [ + 511, + 234, + 906, + 431 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We also find that using a CNN instead of an MLP for image decoding (c) clearly improves image realism, but at the cost of runtime, see Fig. 5 for an example. Further, we verify that using median depth instead of expected depth during inference is important for realistic lidar data in terms of Chamfer distance (h). Last, we note that MCMC [16] and EWA antialIASing [47, 52] both improve our performance, with the antialIASing having the largest impact on perceptual quality in terms of LPIPS.", + "bbox": [ + 511, + 439, + 908, + 575 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 614, + 633, + 631 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we have proposed SplatAD, the first method for rendering camera and lidar data from 3D Gaussians. We combined accurate sensor modeling with efficient algorithms to achieve state-of-the-art novel view synthesis results while improving run-time by more than an order of magnitude. We hope that SplatAD will advance the scalability and fidelity of simulation environments, facilitating safer and more cost-effective testing of AD systems.", + "bbox": [ + 511, + 650, + 906, + 771 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations and future work: SplatAD is currently limited to modeling all dynamic actors as rigid. Drawing inspiration from recent advances in human reconstruction [18, 20, 26] can provide inspiration how to overcome this limitation in future research. In addition, we believe that incorporating data-driven priors such as diffusion models [9, 11, 40] is a promising direction to improve the extrapolation performance, and for changing appearance attributes.", + "bbox": [ + 511, + 779, + 908, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "11989", + "bbox": [ + 480, + 946, + 517, + 958 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 91, + 90, + 243, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We thank Adam Lilja, William Ljungbergh, and Adam Tenderski for valuable feedback. This work was partially supported by the Wallenberg AI, Autonomous Systems and Software Program (WASP) funded by the Knut and Alice Wallenberg Foundation. Computational resources were provided by NAISS at NSC Berzelius, partially funded by the Swedish Research Council, grant agreement no. 2022-06725.", + "bbox": [ + 89, + 111, + 485, + 210 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 234, + 187, + 250 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Mina Alibeigi, William Ljungbergh, Adam Tonderski, Georg Hess, Adam Lilja, Carl Lindstrom, Daria Motorniuk, Junsheng Fu, Jenny Widahl, and Christoffer Petersson. Zenseact open dataset: A large-scale and diverse multimodal dataset for autonomous driving. In ICCV, pages 20178-20188, 2023. 1", + "[2] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In ICCV, pages 5855-5864, 2021. 2", + "[3] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In CVPR, pages 5470–5479, 2022.", + "[4] Jonathan T. Barron, Ben Mildenhall, Dor Verbin, Pratul P. Srinivasan, and Peter Hedman. Zip-nerf: Anti-aliased grid-based neural radiance fields. In ICCV, pages 19697-19705, 2023. 2", + "[5] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In CVPR, pages 11621-11631, 2020. 1, 6", + "[6] Yurui Chen, Chun Gu, Junzhe Jiang, Xiatian Zhu, and Li Zhang. Periodic vibration gaussian: Dynamic urban scene reconstruction and real-time rendering. arXiv preprint arXiv:2311.18561, 2023. 2, 7", + "[7] Ziyu Chen, Jiawei Yang, Jiahui Huang, Riccardo de Lutio, Janick Martinez Esturo, Boris Ivanovic, Or Litany, Zan Gojcic, Sanja Fidler, Marco Pavone, et al. Omnire: Omni urban scene reconstruction. arXiv preprint arXiv:2408.16760, 2024. 1, 2, 5, 7", + "[8] Ziyu Chen, Jiawei Yang, Jiahui Huang, Riccardo de Lutio, Janick Martinez Esturo, Boris Ivanovic, Or Litany, Zan Gojcic, Sanja Fidler, Marco Pavone, Li Song, and Yue Wang. *drivestudio*. https://github.com/ziyc/drivestudio, 2024. 7, 2", + "[9] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In ICML, 2024. 8", + "[10] Xiao Fu, Shangzhan Zhang, Tianrun Chen, Yichong Lu, Lanyun Zhu, Xiaowei Zhou, Andreas Geiger, and Yiyi Liao. Panoptic nerf: 3d-to-2d label transfer for panoptic urban scene segmentation. In 3DV, pages 1-11. IEEE, 2022. 2" + ], + "bbox": [ + 93, + 258, + 485, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[11] Ruiqi Gao, Aleksander Holynski, Philipp Henzler, Arthur Brussee, Ricardo Martin Brualla, Pratul P. Srinivasan, Jonathan T. Barron, and Ben Poole. CAT3d: Create anything in 3d with multi-view diffusion models. In NeurIPS, 2024. 8", + "[12] Shubham Goel, Georgios Pavlakos, Jathushan Rajasegaran, Angjoo Kanazawa, and Jitendra Malik. Humans in 4d: Reconstructing and tracking humans with transformers. In ICCV, pages 14783-14794, 2023. 2", + "[13] Shengyu Huang, Zan Gojcic, Zian Wang, Francis Williams, Yoni Kasten, Sanja Fidler, Konrad Schindler, and Or Litany. Neural lidar fields for novel view synthesis. In ICCV, pages 18236-18246, 2023. 2", + "[14] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. TOG, 42(4):139-1, 2023. 1, 2, 3, 4, 6", + "[15] Mustafa Khan, Hamidreza Fazlali, Dhruv Sharma, Tongtong Cao, Dongfeng Bai, Yuan Ren, and Bingbing Liu. Autosplat: Constrained gaussian splatting for autonomous driving scene reconstruction, 2024. 2", + "[16] Shakiba Kheradmand, Daniel Rebain, Gopal Sharma, Weiwei Sun, Yang-Che Tseng, Hossam Isack, Abhishek Kar, Andrea Tagliasacchi, and Kwang Moo Yi. 3d gaussian splatting as markov chain monte carlo. In NeurIPS, 2024. 5, 6, 8, 2", + "[17] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR, 2015. 2", + "[18] Muhammed Kocabas, Jen-Hao Rick Chang, James Gabriel, Oncel Tuzel, and Anurag Ranjan. Hugs: Human gaussian splats. In CVPR, pages 505-515, 2024. 8", + "[19] Abhijit Kundu, Kyle Genova, Xiaqi Yin, Alireza Fathi, Caroline Pantofaru, Leonidas J Guibas, Andrea Tagliasacchi, Frank Dellaert, and Thomas Funkhouser. Panoptic neural fields: A semantic object-aware neural scene representation. In CVPR, pages 12871-12881, 2022. 2", + "[20] Youngjoong Kwon, Baole Fang, Yixing Lu, Haoye Dong, Cheng Zhang, Francisco Vicente Carrasco, Albert Mosella-Montoro, Jianjin Xu, Shingo Takagi, Daeil Kim, Aayush Prakash, and Fernando De la Torre. Generalizable human gaussians for sparse view synthesis. In ECCV, pages 451-468, Cham, 2025. Springer Nature Switzerland. 8", + "[21] Nanxi Li, Chong Pei Ho, Jin Xue, Leh Woon Lim, Guanyu Chen, Yuan Hsing Fu, and Lennon Yao Ting Lee. A progress review on solid-state lidar and nanophotonics-based lidar sensors. *Laser & Photonics Reviews*, 16(11):2100511, 2022. 4", + "[22] Carl Lindström, Georg Hess, Adam Lilja, Maryam Fatemi, Lars Hammarstrand, Christoffer Petersson, and Lennart Svensson. Are nerfs ready for autonomous driving? towards closing the real-to-simulation gap. In CVPRW, pages 4461-4471, 2024. 2", + "[23] William Ljungbergh, Adam Tonderski, Joakim Johnander, Holger Caesar, Kalle Åström, Michael Felsberg, and Christoffer Petersson. Neuroncap: Photorealistic closed-loop safety testing for autonomous driving. In ECCV, pages 161-177. Springer, 2025. 2", + "[24] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J. Black. SMPL: A skinned multiperson linear model. TOG, 34(6):248:1-248:16, 2015. 2" + ], + "bbox": [ + 516, + 92, + 906, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "11990", + "bbox": [ + 478, + 946, + 519, + 959 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[25] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, pages 405-421, Cham, 2020. Springer International Publishing. 1, 2", + "[26] Arthur Moreau, Jifei Song, Helisa Dhamo, Richard Shaw, Yiren Zhou, and Eduardo Pérez-Pellitero. Human gaussian splatting: Real-time rendering of animatable avatars. In CVPR, pages 788-798, 2024. 8", + "[27] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. TOG, 41(4):1-15, 2022. 2", + "[28] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel HAZIZA, Francisco Massa, Alaaeldin El-Nouby, Mido Assran, Nicolas Ballas, Wojciech Galuba, Russell Howes, Po-Yao Huang, Shang-Wen Li, Ishan Misra, Michael Rabbat, Vasu Sharma, Gabriel Synnaeve, Hu Xu, Herve Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. DINov2: Learning robust visual features without supervision. TMLR, 2024. 7", + "[29] Julian Ost, Fahim Mannan, Nils Thuerey, Julian Knodt, and Felix Heide. Neural scene graphs for dynamic scenes. In CVPR, pages 2856-2865, 2021. 2, 3", + "[30] Konstantinos Rematas, Andrew Liu, Pratul P Srinivasan, Jonathan T Barron, Andrea Tagliasacchi, Thomas Funkhouser, and Vittorio Ferrari. Urban radiance fields. In CVPR, pages 12932-12942, 2022. 4", + "[31] Otto Seiskari, Jerry Yilammi, Valtteri Kaatrasalo, Pekka Rantalankila, Matias Turkulainen, Juho Kannala, Esa Rahtu, and Arno Solin. Gaussian splattering on the move: Blur and rolling shutter compensation for natural camera motion. In ECCV, pages 160-177, Cham, 2025. Springer Nature Switzerland. 4, 3", + "[32] George Stein, Jesse Cresswell, Rasa Hosseinzadeh, Yi Sui, Brendan Ross, Valentin Villecloze, Zhaoyan Liu, Anthony L Caterini, Eric Taylor, and Gabriel Loaiza-Ganem. Exposing flaws of generative model evaluation metrics and their unfair treatment of diffusion models. NeurIPS, 36, 2024. 7", + "[33] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. In CVPR, pages 2818-2826, 2016. 7", + "[34] Adam Tondersski, Carl Lindstrom, Georg Hess, and William Ljungbergh. neurad-studio. https://github.com/georghess/neurad-studio, 2024. 6, 7", + "[35] Adam Tondersski, Carl Lindström, Georg Hess, William Ljungbergh, Lennart Svensson, and Christoffer Petersson. Neurad: Neural rendering for autonomous driving. In CVPR, pages 14895-14904, 2024. 1, 2, 3, 4, 6, 7", + "[36] Haithem Turki, Jason Y Zhang, Francesco Ferroni, and Deva Ramanan. Suds: Scalable urban dynamic scenes. In CVPR, pages 12375-12385, 2023. 1, 2", + "[37] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE TIP, 13(4):600-612, 2004. 7" + ], + "bbox": [ + 91, + 90, + 485, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[38] Benjamin Wilson, William Qi, Tanmay Agarwal, John Lambert, Jagjeet Singh, Siddhesh Khandelwal, Bowen Pan, Ratnesh Kumar, Andrew Hartnett, Jhony Kaesemodel Pontes, Deva Ramanan, Peter Carr, and James Hays. Argoverse 2: Next generation datasets for self-driving perception and forecasting. In Proceedings of the Neural Information Processing Systems Track on Datasets and Benchmarks (NeurIPS Datasets and Benchmarks 2021), 2021. 1, 6", + "[39] Hanfeng Wu, Xingxing Zuo, Stefan Leutenegger, Or Litany, Konrad Schindler, and Shengyu Huang. Dynamic lidar resimulation using compositional neural fields. In CVPR, pages 1998-19998, 2024. 2", + "[40] Rundi Wu, Ben Mildenhall, Philipp Henzler, Keunhong Park, Ruiqi Gao, Daniel Watson, Pratul P Srinivasan, Dor Verbin, Jonathan T Barron, Ben Poole, et al. Reconfusion: 3d reconstruction with diffusion priors. In CVPR, pages 21551-21561, 2024. 8", + "[41] Pengchuan Xiao, Zhenlei Shao, Steven Hao, Zishuo Zhang, Xiaolin Chai, Judy Jiao, Zesong Li, Jian Wu, Kai Sun, Kun Jiang, Yunlong Wang, and Diange Yang. Pandaset: Advanced sensor suite dataset for autonomous driving. In 2021 IEEE International Intelligent Transportation Systems Conference (ITSC), pages 3095-3101, 2021. 2, 6", + "[42] Ziyang Xie, Junge Zhang, Wenye Li, Feihu Zhang, and Li Zhang. S-neRF: Neural radiance fields for street views. In ICLR, 2023. 2", + "[43] Yunzhi Yan, Haotong Lin, Chenxu Zhou, Weijie Wang, Haiyang Sun, Kun Zhan, Xianpeng Lang, Xiaowei Zhou, and Sida Peng. Street gaussians for modeling dynamic urban scenes. In ECCV, 2024. 1, 2, 3, 7", + "[44] Jiawei Yang, Boris Ivanovic, Or Litany, Xinshuo Weng, Seung Wook Kim, Boyi Li, Tong Che, Danfei Xu, Sanja Fidler, Marco Pavone, and Yue Wang. EmerneRF: Emergent spatial-temporal scene decomposition via self-supervision. In ICLR, 2024. 2", + "[45] Ze Yang, Yun Chen, Jingkang Wang, Sivabalan Manivasagam, Wei-Chiu Ma, Anqi Joyce Yang, and Raquel Urtasun. Unisim: A neural closed-loop sensor simulator. In CVPR, pages 1389-1399, 2023. 1, 2, 7, 3", + "[46] Vickie Ye, Ruilong Li, Justin Kerr, Matias Turkulainen, Brent Yi, Zhuoyang Pan, Otto Seiskari, Jianbo Ye, Jeffrey Hu, Matthew Tancik, et al. gsplat: An open-source library for gaussian splatting. arXiv preprint arXiv:2409.06765, 2024. 6", + "[47] Zehao Yu, Anpei Chen, Binbin Huang, Torsten Sattler, and Andreas Geiger. Mip-splatting: Alias-free 3d gaussian splatting. In CVPR, pages 19447-19456, 2024. 4, 5, 8", + "[48] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, pages 586-595, 2018. 7", + "[49] Zehan Zheng, Fan Lu, Weiyi Xue, Guang Chen, and Changjun Jiang. Lidar4d: Dynamic neural fields for novel space-time view lidar synthesis. In CVPR, pages 5145-5154, 2024. 2", + "[50] Hongyu Zhou, Jiahao Shao, Lu Xu, Dongfeng Bai, Weichao Qiu, Bingbing Liu, Yue Wang, Andreas Geiger, and Yiyi Liao. Hugs: Holistic urban 3d scene understanding via gaussian splatting. In CVPR, pages 21336-21345, 2024. 2" + ], + "bbox": [ + 516, + 90, + 906, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "11991", + "bbox": [ + 478, + 946, + 517, + 958 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[51] Xiaoyu Zhou, Zhiwei Lin, Xiaojun Shan, Yongtao Wang, Deqing Sun, and Ming-Hsuan Yang. Drivinggaussian: Composite gaussian splatting for surrounding dynamic autonomous driving scenes. In CVPR, pages 21634-21643, 2024. 1, 2, 3", + "[52] Matthias Zwicker, Hanspeter Pfister, Jeroen Van Baar, and Markus Gross. Ewa volume splatting. In Proceedings Visualization, 2001. VIS'01., pages 29-538. IEEE, 2001. 4, 8" + ], + "bbox": [ + 91, + 92, + 486, + 202 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11992", + "bbox": [ + 480, + 946, + 519, + 959 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/SplatAD_ Real-Time Lidar and Camera Rendering with 3D Gaussian Splatting for Autonomous Driving/58aed1a8-44d5-4314-9f37-1456921ac406_model.json b/2025/SplatAD_ Real-Time Lidar and Camera Rendering with 3D Gaussian Splatting for Autonomous Driving/58aed1a8-44d5-4314-9f37-1456921ac406_model.json new file mode 100644 index 0000000000000000000000000000000000000000..59748e0e4c3f180ccafa9767a10f4aa05a53ad67 --- /dev/null +++ b/2025/SplatAD_ Real-Time Lidar and Camera Rendering with 3D Gaussian Splatting for Autonomous Driving/58aed1a8-44d5-4314-9f37-1456921ac406_model.json @@ -0,0 +1,2191 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.13, + 0.909, + 0.178 + ], + "angle": 0, + "content": "SplatAD: Real-Time Lidar and Camera Rendering with 3D Gaussian Splatting for Autonomous Driving" + }, + { + "type": "text", + "bbox": [ + 0.1, + 0.202, + 0.938, + 0.221 + ], + "angle": 0, + "content": "Georg Hess\\(^{\\dagger,1,2}\\) Carl Lindström\\(^{\\dagger,1,2}\\) Maryam Fatemi\\(^{1}\\) Christoffer Petersson\\(^{1,2}\\) Lennart Svensson\\(^{2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.341, + 0.223, + 0.695, + 0.24 + ], + "angle": 0, + "content": "\\(^{1}\\)Zenseact \\(^{2}\\)Chalmers University of Technology" + }, + { + "type": "text", + "bbox": [ + 0.308, + 0.242, + 0.726, + 0.257 + ], + "angle": 0, + "content": "{firstname_lastname}@{zenseact.com, chalmers.se}" + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.293, + 0.296, + 0.397 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.298, + 0.293, + 0.498, + 0.397 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.5, + 0.293, + 0.702, + 0.397 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.704, + 0.293, + 0.905, + 0.397 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.41, + 0.908, + 0.452 + ], + "angle": 0, + "content": "Figure 1. SplatAD is the first method capable of realistic camera and lidar rendering using 3D Gaussian Splatting. Whereas previous methods are either fast or multi-modal, SplatAD enables real-time, high-quality rendering for both camera and lidar. In addition, SplatAD can reach competitive performance, e.g., PSNR for image and Chamfer distance for point cloud, within minutes." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.465, + 0.327, + 0.48 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.087, + 0.499, + 0.486, + 0.862 + ], + "angle": 0, + "content": "Ensuring the safety of autonomous robots, such as self-driving vehicles, requires extensive testing across diverse driving scenarios. Simulation is a key ingredient for conducting such testing in a cost-effective and scalable way. Neural rendering methods have gained popularity, as they can build simulation environments from collected logs in a data-driven manner. However, existing neural radiance field (NeRF) methods for sensor-realistic rendering of camera and lidar data suffer from low rendering speeds, limiting their applicability for large-scale testing. While 3D Gaussian Splatting (3DGS) enables real-time rendering, current methods are limited to camera data and are unable to render lidar data essential for autonomous driving. To address these limitations, we propose SplatAD, the first 3DGS-based method for realistic, real-time rendering of dynamic scenes for both camera and lidar data. SplatAD accurately models key sensor-specific phenomena such as rolling shutter effects, lidar intensity, and lidar ray dropouts, using purpose-built algorithms to optimize rendering efficiency. Evaluation across three autonomous driving datasets demonstrates that SplatAD achieves state-of-the-art rendering quality with up to \\(+2\\) PSNR for NVS and \\(+3\\) PSNR for reconstruction while increasing rendering speed over NeRF-based methods by an order of magnitude. See here for our project page." + }, + { + "type": "title", + "bbox": [ + 0.515, + 0.465, + 0.645, + 0.48 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.492, + 0.909, + 0.673 + ], + "angle": 0, + "content": "Large-scale testing is essential to ensure the safety of autonomous robots, such as self-driving vehicles (SDVs), before real-world deployment. Data-driven methods that generate digital twins from collected logs offer a scalable way to build diverse, realistic simulation environments for testing. Unlike real-world testing, which can be costly, time-intensive, and limited by physical constraints, simulation enables rapid, low-cost exploration of multiple scenarios, helping to optimize SDVs for safety, comfort, and efficiency. Motivated by this, a multitude of methods have emerged based on neural radiance fields (NeRFs) [25, 35, 36, 45] and 3D Gaussian Splatting (3DGS) [14, 43, 51]." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.675, + 0.911, + 0.901 + ], + "angle": 0, + "content": "Recent NeRF-based methods [35, 45] provide high-fidelity sensor simulation jointly for both camera and lidar, matching the most common sensor rigs in popular autonomous driving (AD) datasets [1, 5, 38]. However, the slow rendering speed of NeRF-based methods makes them costly and challenging to use for large-scale testing. 3DGS provides an attractive alternative to NeRFs, as its accelerated rendering achieves comparable image realism while increasing inference speeds by an order of magnitude. Nevertheless, 3DGS-based methods for the AD setting [7, 43, 51] inherit the limitation of only being able to render camera data, overlooking the lidar modality. The lidar's ability to directly perceive the 3D environment makes it a powerful tool in modern AD stacks and, consequently, an important modality to simulate." + }, + { + "type": "page_footnote", + "bbox": [ + 0.108, + 0.894, + 0.366, + 0.908 + ], + "angle": 0, + "content": "These authors contributed equally to this work." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.948, + 0.52, + 0.96 + ], + "angle": 0, + "content": "11982" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.288 + ], + "angle": 0, + "content": "In this paper, we aim to solve efficient, differentiable, and realistic rendering of camera and lidar data using 3D Gaussian Splatting. Applying 3DGS to lidar sensors presents unique challenges due to their distinct characteristics. First, unlike cameras, which capture dense, regularly spaced pixels, lidars record sparse, non-linear point clouds with large empty gaps from non-returning rays. Second, most lidars capture a full \\(360^{\\circ}\\) view of the scene, where existing methods often project data into multiple depth images [7]—an inefficient approach that overlooks the sparse structure of the lidar. Last, lidars suffer from a rolling shutter effect, with each sweep taking up to \\(100\\mathrm{ms}\\), during which the ego vehicle may move several meters, violating 3DGS's single-origin assumption." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.289, + 0.483, + 0.514 + ], + "angle": 0, + "content": "To overcome these challenges, we introduce SplatAD, a novel view synthesis method that unifies camera and lidar rendering and is designed for real-time rendering of large-scale dynamic traffic scenes. Our method replaces spherical harmonics with per-Gaussian learnable features to jointly model sensor-specific phenomena—from lidar ray drop and intensity variations to camera-specific appearance changes. By introducing purpose-built rasterization algorithms for rendering lidar in spherical coordinates, we achieve both superior efficiency and realism. Further, we demonstrate efficient methods for modeling rolling shutter effects across both sensor modalities. We verify the effectiveness and generalizability of our method across three popular automotive datasets, achieving state-of-the-art results on all of them. To summarize, our contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.515, + 0.483, + 0.573 + ], + "angle": 0, + "content": "- We propose the first method for efficient lidar rendering using 3D Gaussians, introducing custom CUDA-accelerated algorithms for rasterizing sparse point clouds in spherical coordinates." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.575, + 0.483, + 0.635 + ], + "angle": 0, + "content": "- We introduce the first 3DGS method capable of rendering both camera and lidar from a unified representation, enabling accelerated scaling of novel view synthesis for automotive applications." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.636, + 0.483, + 0.695 + ], + "angle": 0, + "content": "- We present effective techniques for realistic sensor modeling using 3D Gaussians, enabling accurate handling of rolling shutter, lidar intensity, ray dropping, and variations in sensor appearance." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.696, + 0.483, + 0.756 + ], + "angle": 0, + "content": "- Through extensive evaluation on three popular automotive datasets, we demonstrate state-of-the-art results across all benchmarks, validating our method's effectiveness and generalizability." + }, + { + "type": "list", + "bbox": [ + 0.092, + 0.515, + 0.483, + 0.756 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.77, + 0.228, + 0.785 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.483, + 0.901 + ], + "angle": 0, + "content": "NeRFs for automotive data: Since the introduction of NeRFs, neural representations have become central to 3D reconstruction and novel view synthesis [2-4, 25]. Many works have applied NeRF-based methods to automotive data [35, 36, 44, 45], enabling sensor-realistic renderings of novel views in large and dynamic scenes. Recent advances provide realism sufficient for downstream applica" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.908, + 0.332 + ], + "angle": 0, + "content": "tions [22, 23]. While early methods focused on either camera [10, 19, 29, 42] or lidar [13, 39, 49], newer methods aim to handle them jointly. UniSim [45] showcased realistic renderings for front camera and \\(360^{\\circ}\\) lidar on data from PandaSet [41]. The method uses a hash grid representation [27], with separate features for the sky, static background, and dynamic actors, and decodes the color and lidar intensity from volume-rendered features. NeuRAD [35] proposes a simplified network architecture and improved sensor modeling, achieving state-of-the-art results for the full \\(360^{\\circ}\\) camera and lidar rig. Even so, the slow rendering speeds of NeRF-based methods make scaling costly and challenging. SplatAD aims to overcome this limitation by using rasterization-based CUDA-accelerated algorithms to render both camera and lidar data while, inspired by NeuRAD, emphasizing the modeling of important sensor characteristics for increased realism." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.333, + 0.909, + 0.666 + ], + "angle": 0, + "content": "3DGS for automotive data: 3DGS [14] uses an explicit representation paired with rasterization techniques and custom hardware-accelerated algorithms to achieve real-time rendering. Additionally, 3DGS has been shown to create faithful scene reconstructions with rendering quality close to recent NeRF-based methods. As a result, several works have applied 3DGS for camera rendering on automotive data [7, 15, 43, 50, 51]. Periodic Vibration Gaussian (PVG) [6] applies 3DGS to dynamic scenes by learning the 3D flow of all Gaussians. Yet, PVG's lack of an explicit actor representation limits the method's controllability and, hence, applicability for simulation. Street Gaussians [43], like Unisim and NeuRAD, instead decomposes the scene into static background and rigid dynamic actors using 3D bounding boxes. The authors further extend 3DGS by adding Fourier coefficients to capture time-varying characteristics of dynamic actors. Moreover, to handle the sky regions, their method uses semantic masks and a view-dependent cube map. OmniRe [7] enhances the scene graph composition with non-rigid nodes for better modeling of pedestrians and cyclists. However, this requires pre-processing to track and align human body poses [12] for initializing SMPL [24] pose parameters." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.666, + 0.909, + 0.832 + ], + "angle": 0, + "content": "PVG, Street Gaussians, and OmniRe all use lidar points for initialization and depth supervision. However, to generate new point clouds, these methods predict depth images and rely on existing lidar points to query said image at the right locations. This is both inefficient and inaccurate, as \\(360^{\\circ}\\) lidars are approximated as multiple cameras. Further, neither of these methods model important lidar characteristics such as intensity variations, ray drop, and rolling shutter. In contrast, our method supports efficient rendering of both camera and lidar from 3D Gaussians in novel views, and models important sensor characteristics for both modalities." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.845, + 0.605, + 0.86 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.908, + 0.901 + ], + "angle": 0, + "content": "Our aim is to learn a scene representation from collected vehicle logs that enables rendering of realistic camera and lidar" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.948, + 0.518, + 0.959 + ], + "angle": 0, + "content": "11983" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.098, + 0.09, + 0.913, + 0.328 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.337, + 0.908, + 0.425 + ], + "angle": 0, + "content": "Figure 2. Overview of our proposed method. Given the composition of static and dynamic 3D Gaussians, SplatAD is capable of differentiable rendering of both lidar and camera data. Our proposed lidar rendering matches the image rendering on a high level, but modifies each component to accurately model sensor characteristics. Our method projects 3D Gaussians with associated feature vectors onto the corresponding sensor modalities (camera and lidar) and employs sensor-specific tiling to match their distinct characteristics. During rasterization, the projected Gaussians are corrected for rolling shutter effects caused by the movement of sensors and, potentially, their own velocity. Finally, the rasterized features are decoded into the respective image and lidar point cloud representations." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.448, + 0.482, + 0.554 + ], + "angle": 0, + "content": "data, with the ability to alter the locations of both the ego vehicle and other actors. To scale effectively, the rendering process must be fast, as faster inference speeds enhance its practicality for applications. In the following, we describe our scene representation (Sec. 3.1), rendering algorithms (Secs. 3.2 and 3.3), and implementation and optimization strategy (Sec. 3.4), see Fig. 2 for an overview." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.562, + 0.288, + 0.577 + ], + "angle": 0, + "content": "3.1. Scene representation" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.584, + 0.483, + 0.78 + ], + "angle": 0, + "content": "Our scene representation builds upon 3DGS [14], but with key changes to handle the specifics of AD scenes and to enable both camera and lidar rendering from the same representation. Like 3DGS, each scene is represented by a set of translucent 3D Gaussians with learnable occupancy \\(o \\in (0,1)\\), mean \\(\\mu \\in \\mathbb{R}^3\\), and covariance matrix \\(\\Sigma \\in \\mathbb{R}^{3\\times 3}\\) which in turn is parameterized by scale \\(\\mathbf{S} \\in \\mathbb{R}^3\\) and quaternion \\(\\mathbf{q} \\in \\mathbb{R}^4\\) parameters. Instead of spherical harmonics [14], we assign each Gaussian a learnable base color \\(\\mathbf{f}_{\\mathrm{rgb}}^{\\mathrm{b}} \\in \\mathbb{R}^3\\) and feature vector \\(\\mathbf{f} \\in \\mathbb{R}^{D_{\\mathrm{f}}}\\), where \\(\\mathbf{f}\\) is used for both view-dependent effects and lidar properties. Last, our representation contains a learnable embedding per sensor to model their specific appearance characteristics." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.781, + 0.483, + 0.901 + ], + "angle": 0, + "content": "To handle dynamics, we follow the commonly used scene graph decomposition [29, 35, 43, 51] and divide the scene into a static background and a set of dynamic actors. Each dynamic actor is described by a 3D bounding box and a sequence of SE(3) poses, obtained either from an off-the-shelf object detector and tracker, or annotations. Each Gaussian has a non-learnable ID, indicating whether it is assigned to the static world, or to which actor it belongs. For Gaussians" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.448, + 0.907, + 0.584 + ], + "angle": 0, + "content": "assigned to actors, their mean and covariance are expressed in the local coordinate system of the corresponding axis-aligned bounding box. To compose the scene at a given time \\( t \\), Gaussians assigned to bounding boxes are transformed to world coordinates based on their actor's pose. Since pose estimates for actors may contain inaccuracies, we adjust these with learnable offsets. Furthermore, each actor has a velocity initialized from pose differences and a learnable velocity offset." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.592, + 0.693, + 0.608 + ], + "angle": 0, + "content": "3.2. Camera rendering" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.614, + 0.907, + 0.719 + ], + "angle": 0, + "content": "Given a posed camera, we compose the set of Gaussians at the corresponding capture time \\( t \\) and render an image \\( I \\) using the efficient tile-based rendering from 3DGS [14]. While we retain 3DGS's high-level steps—projection and view frustum culling, tile-assignment, depth sorting, and tile-based rasterization—we introduce key adaptations to better model the unique characteristics of AD data." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Projection, tiling, and sorting: Each mean and covariance is transformed from world to camera coordinates, giving \\(\\pmb{\\mu}^{C}\\) and \\(\\pmb{\\Sigma}^{C}\\). The means are then converted to image space \\(\\pmb{\\mu}^{I} \\in \\mathbb{R}^{2}\\) using the perspective projection, while the covariances are transformed using the two upper rows of the Jacobian of the projection, \\(\\pmb{\\Sigma}^{I} = \\mathbf{J}^{I}\\pmb{\\Sigma}^{C}(\\mathbf{J}^{I})^{\\mathrm{T}} \\in \\mathbb{R}^{2 \\times 2}\\). Gaussians outside the view frustum are culled, where, for efficiency, 3DGS approximates the extent of the Gaussians using a square axis-aligned bounding box (AABB) that covers their \\(99\\%\\) confidence level. Further, 3DGS divides the image into tiles, each measuring \\(16 \\times 16\\) pixels, and assigns Gaussians to all tiles they intersect, duplicating them if necessary. This" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.948, + 0.518, + 0.959 + ], + "angle": 0, + "content": "11984" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.137 + ], + "angle": 0, + "content": "way, each pixel only has to process a subset of all Gaussians during rasterization. Lastly, the Gaussians are sorted based on the \\(z\\)-depth of their mean \\(\\pmb{\\mu}^{C}\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.137, + 0.484, + 0.408 + ], + "angle": 0, + "content": "Rolling shutter: Many cameras use a rolling shutter, where the capture of an image is not instantaneous, but is done row-by-row, allowing the camera to move during the exposure. Previous work [35] has highlighted the importance of modeling rolling shutter effects due to the high sensor velocities present in AD data, where their ray-tracing-based method can easily account for this by shifting the origins of each ray. For 3DGS, the equivalent would require projecting all 3D Gaussians to all camera poses encountered during an exposure, as the Gaussians' positions relative to the camera change with time. As this would be prohibitively expensive, we draw inspiration from [31] and instead approximate the effects directly in 2D image space. Specifically, each Gaussian's velocity relative to the camera is projected to image space, and their pixel mean \\(\\mu^I\\) is adjusted during rasterization based on the capture time of the pixel. However, [31] considered static scenes only, hence we adjust the formulation to account for the dynamics as well." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.409, + 0.479, + 0.424 + ], + "angle": 0, + "content": "For each Gaussian, its pixel velocity is approximated as" + }, + { + "type": "equation", + "bbox": [ + 0.144, + 0.432, + 0.483, + 0.451 + ], + "angle": 0, + "content": "\\[\n\\mathbf {v} ^ {I} = \\mathbf {J} ^ {I} \\left(- \\boldsymbol {\\omega} _ {C} \\times \\boldsymbol {\\mu} ^ {C} - \\mathbf {v} _ {C} + \\mathbf {v} _ {\\mathrm {d y n}}\\right) \\in \\mathbb {R} ^ {2}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.459, + 0.483, + 0.489 + ], + "angle": 0, + "content": "where \\(\\omega_{C}\\) and \\(\\mathbf{v}_C\\) denote the angular and linear velocity of the camera expressed in the camera coordinate system, and" + }, + { + "type": "equation", + "bbox": [ + 0.149, + 0.497, + 0.483, + 0.515 + ], + "angle": 0, + "content": "\\[\n\\mathbf {v} _ {\\mathrm {d y n}} = T ^ {\\mathrm {a c t} \\rightarrow C} \\left(\\boldsymbol {\\omega} _ {\\mathrm {a c t}} \\times \\boldsymbol {\\mu} ^ {\\mathrm {a c t}} + \\mathbf {v} _ {\\mathrm {a c t}}\\right) \\in \\mathbb {R} ^ {3}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.524, + 0.484, + 0.644 + ], + "angle": 0, + "content": "denotes the velocity of the Gaussian induced by the angular \\(\\omega^{\\mathrm{dyn}}\\) and linear velocity \\(\\mathbf{v}^{\\mathrm{dyn}}\\) of its associated actor (expressed in actor coordinate system) and \\(\\mu^{\\mathrm{act}}\\) denotes the Gaussian mean in the actor coordinate system. To transform velocities from actor to camera coordinates, \\(T^{\\mathrm{act}\\rightarrow C}\\) is the composition of the actor-to-world and world-to-camera transforms. For Gaussians that are part of the static background, \\(\\mathbf{v}_{\\mathrm{dyn}}\\) is zero. See Appendix D for further details." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.646, + 0.484, + 0.901 + ], + "angle": 0, + "content": "We account for the pixel velocity when culling Gaussians and checking intersections between Gaussians and tiles by increasing the approximated Gaussian extent. Instead of the square AABB used by 3DGS and [31], we use a rectangular AABB covering \\(\\boldsymbol{\\Sigma}^{I}\\) at three standard deviations and add \\(|\\mathbf{v}_I| \\cdot t_{\\mathrm{rs}} / 2\\) to its extent. Here, \\(t_{\\mathrm{rs}}\\) denotes the rolling shutter duration, i.e., the time difference between the last and first row of pixels. The increase in extent corresponds to the area covered by the Gaussian mean during the rolling shutter time, assuming the sensor timestamp is in the middle of exposure. As this is not always the case, we further include a learnable time offset, describing the difference between the sensor timestamp and the middle of the exposure. Using rectangular AABBs reduces unnecessary intersections compared to [14, 31], especially for narrow Gaussians and Gaussians with axis-aligned velocities, the latter often being the case for side-facing cameras." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.182 + ], + "angle": 0, + "content": "Rasterization: 3DGS rasterizes pixels in parallel by launching a thread block per tile and assigning each thread to a pixel within the tile. For each pixel, its coordinates \\(\\mathbf{p} = [p_u,p_v]^{\\mathrm{T}}\\) are inferred from the block and the thread index. From this, we find the time difference between the pixel's capture time and the image's middle row as" + }, + { + "type": "equation", + "bbox": [ + 0.623, + 0.195, + 0.907, + 0.212 + ], + "angle": 0, + "content": "\\[\nt _ {\\mathrm {p i x}} = \\left(p _ {v} / H - 0. 5\\right) \\cdot t _ {\\mathrm {r s}}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.223, + 0.905, + 0.255 + ], + "angle": 0, + "content": "To rasterize a pixel, we \\(\\alpha\\)-blend RGB values \\(\\mathbf{f}_i^{\\mathrm{rgb}}\\) and features \\(\\mathbf{f}_i\\) of the depth-sorted Gaussians intersecting the current tile" + }, + { + "type": "equation", + "bbox": [ + 0.575, + 0.265, + 0.907, + 0.307 + ], + "angle": 0, + "content": "\\[\n\\left[ \\mathbf {F} _ {\\mathbf {p}} ^ {\\mathrm {r g b}}, \\mathbf {F} _ {\\mathbf {p}} \\right] = \\sum_ {i} \\left[ \\mathbf {f} _ {i} ^ {\\mathrm {r g b}}, \\mathbf {f} _ {i} \\right] \\alpha_ {i} \\prod_ {j = 1} ^ {i - 1} (1 - \\alpha_ {j}), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.318, + 0.649, + 0.333 + ], + "angle": 0, + "content": "with \\(\\alpha\\) computed as" + }, + { + "type": "equation", + "bbox": [ + 0.525, + 0.345, + 0.907, + 0.385 + ], + "angle": 0, + "content": "\\[\n\\alpha_ {i} = \\sqrt {\\frac {\\left| \\boldsymbol {\\Sigma} _ {i} ^ {I} \\right|}{\\left| \\boldsymbol {\\Sigma} _ {i} ^ {I} + s \\mathbf {I} \\right|}} o _ {i} \\exp \\left(- \\frac {1}{2} \\boldsymbol {\\Delta} _ {i} ^ {\\mathrm {T}} \\left(\\boldsymbol {\\Sigma} _ {i} ^ {I} + s \\mathbf {I}\\right) ^ {- 1} \\boldsymbol {\\Delta} _ {i}\\right) \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.395, + 0.543, + 0.407 + ], + "angle": 0, + "content": "and" + }, + { + "type": "equation", + "bbox": [ + 0.623, + 0.408, + 0.907, + 0.427 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\Delta} _ {i} = \\mathbf {p} - \\left(\\boldsymbol {\\mu} _ {i} ^ {I} + \\mathbf {v} _ {i} ^ {I} t _ {\\mathrm {p i x}}\\right). \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.432, + 0.907, + 0.507 + ], + "angle": 0, + "content": "As in [31], we calculate the distance \\(\\Delta_{i}\\) between pixels and Gaussians using the means compensated by rolling shutter, effectively shifting Gaussians to the correct location in the image. Compared to 3DGS, in Eq. (5) we use the 3DGS+EWA [52] formulation from [47], where \\(s = 0.3\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.508, + 0.906, + 0.582 + ], + "angle": 0, + "content": "View-dependent effects are modeled using a small CNN; given the feature map \\(\\mathbf{F} \\in \\mathbb{R}^{H \\times W \\times D_{\\mathrm{f}}}\\), the corresponding ray directions \\(\\mathbf{d} \\in \\mathbb{R}^{H \\times W \\times 3}\\), and a camera-specific learned embedding \\(\\mathbf{e}\\), it predicts a pixel-wise affine mapping \\(M \\in \\mathbb{R}^{H \\times W \\times 3}\\) and \\(\\mathbf{b} \\in \\mathbb{R}^{H \\times W \\times 3}\\) applied to \\(\\mathbf{F}^{\\mathrm{rgb}}\\)" + }, + { + "type": "equation", + "bbox": [ + 0.603, + 0.595, + 0.907, + 0.612 + ], + "angle": 0, + "content": "\\[\nM, \\mathbf {b} = \\operatorname {C N N} (\\operatorname {c o n c a t} (\\mathbf {F}, \\mathbf {d}, \\mathbf {e})), \\tag {7}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.629, + 0.615, + 0.907, + 0.632 + ], + "angle": 0, + "content": "\\[\nI = (1 + M) \\odot \\mathbf {F} ^ {\\mathrm {r g b}} + \\mathbf {b}. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.643, + 0.908, + 0.718 + ], + "angle": 0, + "content": "We found that using a small CNN instead of an MLP is essential for efficient texture modeling, e.g., for high-frequency road surfaces. In addition, inspired by [30], the camera-specific learned embeddings enable modeling differences in exposure." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.728, + 0.673, + 0.744 + ], + "angle": 0, + "content": "3.3. Lidar rendering" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Lidar sensors enable self-driving vehicles to measure the distance and reflectivity (intensity) of a discrete set of points. They do so by emitting laser beam pulses and measuring the time of flight to determine distance and returning power for reflectivity. Most AD datasets deploy lidars that use several laser diodes (16-128) mounted in a vertical array, where the array of diodes is rotated to capture \\(360^{\\circ}\\) data and a single scan often takes \\(100\\mathrm{ms}\\) to capture. Therefore, we focus on this type of lidar. However, we note that modeling other types of lidars, such as solid-state lidars [21], can be easily" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.948, + 0.518, + 0.959 + ], + "angle": 0, + "content": "11985" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.151 + ], + "angle": 0, + "content": "done in our method by modifying the projection accordingly. Our proposed lidar rendering follows the high-level steps of the image rendering, but we modify each component to accurately model lidar characteristics." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.152, + 0.483, + 0.262 + ], + "angle": 0, + "content": "Projection: For a posed lidar and Gaussians composed at its capture time \\( t \\), we transform each mean and covariance from world to lidar coordinates, yielding \\( \\boldsymbol{\\mu}^{L} = [x,y,z]^{\\mathrm{T}} \\) and \\( \\boldsymbol{\\Sigma}^{L} \\). Raw lidar data are generally captured in spherical coordinates azimuth \\( \\phi \\), elevation \\( \\omega \\), and range \\( r \\). Thus, we transform Gaussian means from lidar coordinates to spherical coordinates" + }, + { + "type": "equation", + "bbox": [ + 0.168, + 0.27, + 0.484, + 0.322 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\mu} ^ {S} = \\left[ \\begin{array}{l} \\phi \\\\ \\omega \\\\ r \\end{array} \\right] = \\left[ \\begin{array}{c} \\arctan 2 (y, x) \\\\ \\arcsin (z / r) \\\\ \\sqrt {x ^ {2} + y ^ {2} + z ^ {2}} \\end{array} \\right], \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.333, + 0.483, + 0.366 + ], + "angle": 0, + "content": "and convert the covariance \\(\\pmb{\\Sigma}^{L}\\) with the Jacobian of the transform Eq. (9) as \\(\\pmb{\\Sigma}^{S} = \\mathbf{J}^{S}\\pmb{\\Sigma}^{L}(\\mathbf{J}^{S})^{\\mathrm{T}}\\), where" + }, + { + "type": "equation", + "bbox": [ + 0.112, + 0.375, + 0.484, + 0.442 + ], + "angle": 0, + "content": "\\[\n\\mathbf {J} ^ {S} = \\left[ \\begin{array}{c c c} - \\frac {y}{x ^ {2} + y ^ {2}} & \\frac {x}{x ^ {2} + y ^ {2}} & 0 \\\\ - \\frac {x z}{r ^ {2} \\sqrt {x ^ {2} + y ^ {2}}} & - \\frac {y z}{r ^ {2} \\sqrt {x ^ {2} + y ^ {2}}} & \\frac {\\sqrt {x ^ {2} + y ^ {2}}}{r ^ {2}} \\\\ \\frac {x}{r} & \\frac {y}{r} & \\frac {z}{r} \\end{array} \\right], \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.451, + 0.483, + 0.481 + ], + "angle": 0, + "content": "and \\(\\pmb{\\mu}^{S}\\) and \\(\\pmb{\\Sigma}^{S}\\) denote mean and covariance in spherical coordinates." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.483, + 0.484, + 0.514 + ], + "angle": 0, + "content": "Same as for the camera, we account for the rolling shutter effect by approximating its impact in sensor space" + }, + { + "type": "equation", + "bbox": [ + 0.144, + 0.523, + 0.483, + 0.543 + ], + "angle": 0, + "content": "\\[\n\\mathbf {v} ^ {S} = \\mathbf {J} ^ {S} \\left(- \\boldsymbol {\\omega} _ {L} \\times \\boldsymbol {\\mu} ^ {L} - \\mathbf {v} _ {L} + \\mathbf {v} _ {\\mathrm {d y n}}\\right) \\in \\mathbb {R} ^ {3}. \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.553, + 0.484, + 0.628 + ], + "angle": 0, + "content": "Note that \\(\\mathbf{v}^S\\) is three-dimensional as opposed to \\(\\mathbf{v}^I\\) in Eq. (1), as we need this for accurate range rasterization. Last, we use the first two elements of \\(\\mathbf{v}^S t_{\\mathrm{rs}} / 2\\) to increase the extent of the rectangular AABB approximating the size of the Gaussian, and cull Gaussians outside the lidar's field-of-view." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.63, + 0.484, + 0.825 + ], + "angle": 0, + "content": "Tiling and sorting: At the core of 3DGS's efficiency is its tile-based rasterization. While lidars often have a fixed azimuth resolution, many use non-equidistant spacing between diodes to get higher resolution in areas of interest, resulting in nonlinear elevation resolution, as seen in Fig. 1. In line with this, we design tiles with a fixed number of diodes vertically and a fixed azimuth resolution horizontally. Horizontally, intersecting tiles are found by dividing azimuth coordinates by the common tile azimuth size and wrapping the results within \\(360^{\\circ}\\). Vertically, we sort the elevation tile boundaries and find intersections with a short for-loop, see Appendix A for details. After tile assignment, Gaussians are sorted based on the range of their mean \\(\\mu^S\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Comparing our strategy to using tiles of equal size, similar to the depth image-based rendering in [7], we avoid many unnecessary computations. There, the tile size must be determined from the finest resolution to ensure having at least as many threads as lidar points. Consequently, tiles" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.136 + ], + "angle": 0, + "content": "in areas where the lidar is sparse would have a superficially high resolution and might even lack the corresponding lidar points to render." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.138, + 0.907, + 0.333 + ], + "angle": 0, + "content": "Rasterization: Lidar data are rendered using our proposed differentiable tile-based rasterizer. Again, each tile launches one thread block, where each thread is responsible for rasterizing one lidar point. Each thread is provided the azimuth, elevation, and capture time for it to rasterize. These are either determined from lidar specifications or, for training, calculated from existing point clouds, see Appendix A for details. The reason for this is that for training, one cannot expect the data to follow perfectly straight scan lines. Therefore, we cannot follow the image strategy and rely on the block and thread index to directly infer the pixel coordinates. Further, the azimuth and elevation of a lidar ray do not necessarily determine its time offset to the center of the lidar scan." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.335, + 0.906, + 0.44 + ], + "angle": 0, + "content": "For each lidar point, we \\(\\alpha\\)-blend the Gaussians' features \\(\\mathbf{f}_i\\) using Eqs. (4) to (6). Inspired by [47] arguing for \\(s\\) in Eq. (5) to correspond to one pixel, we set it to correspond to the geometric mean of the lidar's vertical and horizontal beam divergence. The rendered features are concatenated with their corresponding ray direction and decoded to intensity and ray drop probability using a small MLP." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.441, + 0.906, + 0.471 + ], + "angle": 0, + "content": "To render the expected range of a point, we \\(\\alpha\\)-blend the rolling shutter corrected ranges for each Gaussian" + }, + { + "type": "equation", + "bbox": [ + 0.649, + 0.483, + 0.907, + 0.502 + ], + "angle": 0, + "content": "\\[\nr _ {i, \\mathrm {r s}} = r _ {i} + \\mathbf {v} _ {r} ^ {S} t _ {l}, \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.514, + 0.906, + 0.65 + ], + "angle": 0, + "content": "where \\(\\mathbf{v}_r^S\\) is the range-component of the relative speed to the lidar and \\(t_l\\) is the time between the capture of the current lidar point, and the middle of the lidar scan. Further, we render the median range as the rolling shutter corrected range of the first Gaussian that satisfies \\(\\prod_{j=1}^{i}(1 - \\alpha_j) < 0.5\\) during the \\(\\alpha\\)-blending in Eq. (4). While we use the expected range for training, the median range is used during inference as it, in contrast to the expected range, does not yield depths between Gaussians." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.662, + 0.812, + 0.679 + ], + "angle": 0, + "content": "3.4. Optimization and implementation" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.685, + 0.906, + 0.701 + ], + "angle": 0, + "content": "All model components are optimized jointly using the loss" + }, + { + "type": "equation", + "bbox": [ + 0.53, + 0.732, + 0.907, + 0.769 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} = \\lambda_ {r} \\mathcal {L} _ {1} + (1 - \\lambda_ {r}) \\mathcal {L} _ {\\text {S S I M}} + \\lambda_ {\\text {d e p t h}} \\mathcal {L} _ {\\text {d e p t h}} + \\lambda_ {\\text {l o s}} \\mathcal {L} _ {\\text {l o s}} + \\\\ \\lambda_ {\\text {i n t e n s}} \\mathcal {L} _ {\\text {i n t e n}} + \\lambda_ {\\text {r a y d r o p}} \\mathcal {L} _ {\\text {B C E}} + \\lambda_ {\\text {M C M C}} \\mathcal {L} _ {\\text {M C M C}}, \\tag {13} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.907, + 0.901 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_1\\) and \\(\\mathcal{L}_{\\mathrm{SSIM}}\\) are L1 and SSIM losses on the rendered images. \\(\\mathcal{L}_{\\mathrm{depth}}\\) and \\(\\mathcal{L}_{\\mathrm{inten}}\\) are L2 losses on the rendered expected lidar range and intensity. \\(\\mathcal{L}_{\\mathrm{los}}\\) is a line-of-sight loss, penalizing opacity before the ground truth lidar range. \\(\\mathcal{L}_{\\mathrm{BCE}}\\) is a binary cross-entropy loss on the predicted ray drop probability, where ground-truth is generated in the same way as for NeuRAD. \\(\\mathcal{L}_{\\mathrm{MCMC}}\\) is the opacity and scale regularization used in [16]." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.948, + 0.52, + 0.959 + ], + "angle": 0, + "content": "11986" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.089, + 0.905, + 0.326 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.335, + 0.908, + 0.381 + ], + "angle": 0, + "content": "Figure 3. Compared to the baselines, SplatAD produces sharp images with a high level of detail. Further, the bottom row highlights the superiority of our lidar rendering. Projecting lidar points into images for depth supervision, as used by previous 3DGS methods, causes line-of-sight errors and incorrect volume carving due to the pose differences between camera and lidar." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.404, + 0.486, + 0.646 + ], + "angle": 0, + "content": "We initialize the set of Gaussians using both lidar points and random points. Lidar points within bounding boxes are assigned to the corresponding actor. Further, all lidar points are projected into their closest image to set their initial color. The random points are divided into (1) uniformly sampled points within the lidar range and (2) points sampled linearly in disparity beyond the lidar range, both using random colors at initialization. Furthermore, instead of the 3DGS [14] splitting and densification strategy, we rely on the MCMC strategy introduced in [16]. We chose MCMC partially because we find it to improve far-field rendering quality and partially because it has predictable computation requirements, as it permits choosing the maximum number of Gaussian. We use [16] as is, without any special treatment of Gaussians assigned to dynamic actors. See Appendix B for further training details." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.647, + 0.486, + 0.799 + ], + "angle": 0, + "content": "Implementation: For efficiency, we implement the forward and backward passes of the rolling shutter compensation and lidar projection and rasterization using custom CUDA kernels. SplatAD builds upon the collaborative open-source frameworks gsplat [46] and neurad-studio [34], and we hope that our contributions can benefit future research on multi-modal simulation, not only within the AD community. We train SplatAD for 30,000 iterations using the Adam optimizer, which takes an hour on a single NVIDIA A100. See Appendix B for further details." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.815, + 0.224, + 0.832 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.485, + 0.902 + ], + "angle": 0, + "content": "To validate the robustness of our method, we evaluate it across multiple popular AD datasets, using the same set of hyperparameters. For each dataset, we compare SplatAD to the best performing NeRF and 3DGS-based methods tailored" + }, + { + "type": "table_caption", + "bbox": [ + 0.514, + 0.404, + 0.906, + 0.437 + ], + "angle": 0, + "content": "Table 1. NVS results for image, over three datasets. First, second, third." + }, + { + "type": "table", + "bbox": [ + 0.544, + 0.446, + 0.875, + 0.64 + ], + "angle": 0, + "content": "
PSNR ↑SSIM ↑LPIPS ↓MP/s ↑Train (H) ↓
PandaSetUniSim23.120.6820.36011.50.9
NeuRAD25.800.7530.2509.71.3
PVG24.010.7120.45222.61.6
Street-GS24.730.7450.31463.12.3
OmniRE24.710.7450.31553.62.6
SplatAD (ours)26.700.8140.193104.11.0
Argo2UniSim22.350.6550.45810.01.1
NeuRAD26.180.7210.3109.71.3
PVG24.470.7120.50219.72.5
Street-GS25.520.7540.37492.82.6
OmniRE25.610.7530.37575.63.0
SplatAD (ours)28.400.8260.271133.81.2
nuScenesUniSim22.660.7430.41413.30.7
NeuRAD26.170.7900.31211.81.1
PVG24.940.7680.49718.81.4
Street-GS25.540.7990.37557.91.6
OmniRE25.500.7980.38646.91.9
SplatAD (ours)27.530.8490.302104.31.2
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.671, + 0.907, + 0.734 + ], + "angle": 0, + "content": "to AD data on novel view synthesis for both camera and lidar data. Further, we ablate key elements of SplatAD and measure the impact they have on sensor realism and rendering speed." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.911, + 0.903 + ], + "angle": 0, + "content": "Datasets: We perform experiments on PandaSet [41], Argoverse2 [38] and nuScenes [5]. These datasets vary in lidars, camera appearance, and resolution. For PandaSet and nuScenes, we use the six available cameras. For Argoverse2, we use the seven ring cameras, leaving out the black-and-white stereo cameras. Further, some cameras are cropped slightly to remove views of the ego-vehicle, such as the hood and the trunk. We use the same 10 challenging sequences per dataset as in [35], containing different types of illumination, dynamic actors, and velocities. We also follow the evaluation protocol of [35] i.e., using full resolution data," + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.948, + 0.52, + 0.96 + ], + "angle": 0, + "content": "11987" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.486, + 0.121 + ], + "angle": 0, + "content": "Table 2. NVS results for lidar, over three datasets. \\( {}^{8} \\) without missing points. First ,second ,third ." + }, + { + "type": "table", + "bbox": [ + 0.122, + 0.13, + 0.452, + 0.293 + ], + "angle": 0, + "content": "
Depth ↓Intensity ↓Drop acc. ↑CD ↓MR/s ↑
PandaSetUniSim0.080.086-10.3$0.9
NeuRAD0.010.06396.21.91.1
PVG38.74--125.2$0.3
Street-GS6.18--37.3$0.8
OmniRE2.88--29.8$0.6
SplatAD (ours)0.010.05996.71.619.5
Argo2UniSim0.180.081-29.2$0.7
NeuRAD0.020.05892.22.60.9
SplatAD (ours)0.020.05292.62.89.5
nuScenesUniSim0.060.063-52.2$1.0
NeuRAD0.010.04293.16.31.1
PVG10.34--77.9$0.05
Street-GS1.56--4.5$0.2
OmniRE1.71--4.6$0.1
SplatAD (ours)0.020.03693.81.75.7
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.319, + 0.483, + 0.41 + ], + "angle": 0, + "content": "and, for novel view synthesis (NVS), use every other frame for training and the remaining frames for hold-out validation. We believe that this setting is more challenging than what is often encountered in previous work [6, 7, 43], where fewer datasets are considered, images are downsampled, and a larger fraction of the data is used for training." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.411, + 0.483, + 0.592 + ], + "angle": 0, + "content": "Baselines: We compare our results with the best performing NeRF and 3DGS methods on AD data, namely, UniSim [45], NeuRAD [35], PVG [6], Street Gaussians [43] and OmniRe [7]. Note that all methods are designed for dynamic scenes and use lidar data for supervision. We use NeuRAD's official implementation neurad-studio [34] and their version of UniSim, as well as OmniRe's official implementation drivestudio [8] and their implementations of SVG and Street Gaussians. We use drivestudio to generate the required sky masks for Street Gaussians and OmniRe, as well as extract human pose estimates, which takes about 10 and 30 minutes per sequence on an A100." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.594, + 0.486, + 0.82 + ], + "angle": 0, + "content": "Although point cloud rendering is shown in [7], there exists no feature for it in drivestudio. To render point clouds in training and validation views, we render depth images by placing 6 virtual cameras in the lidar origin, each with a horizontal FOV of \\(60^{\\circ}\\). Their vertical FOV is set to capture all lidar points in the sweep, and the focal length is the median focal length for each dataset. We project ground truth lidar points into these images, fetch each point's depth, and project them back into 3D. As drivestudio lacks functionality for inferring missing lidar points, these baselines are evaluated only for existing points. In contrast, SplatAD is evaluated on the more challenging task of predicting the range also for missing rays and must filter them based on the predicted ray drop probability. See Appendix C for further baseline details." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.833, + 0.258, + 0.849 + ], + "angle": 0, + "content": "4.1. Image rendering" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.486, + 0.903 + ], + "angle": 0, + "content": "Novel view synthesis: We assess the rendering quality of images using the standard NVS metrics PSNR, SSIM [37], and LPIPS [48] on hold-out validation images in Tab. 1." + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.09, + 0.906, + 0.253 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.265, + 0.907, + 0.294 + ], + "angle": 0, + "content": "Figure 4. Removing our rolling shutter modeling compensation leads to inaccurate geometries and inconsistencies in the learning." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.322, + 0.907, + 0.428 + ], + "angle": 0, + "content": "We measure speed using resolution-agnostic megapixels per second. SplatAD consistently outperforms existing 3DGS methods in terms of image quality with a large margin for the three considered datasets. Further, our method improves upon NeuRAD's SOTA results in all NVS metrics, while rendering images an order of magnitude faster. We provide a qualitative comparison in Fig. 3." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.429, + 0.909, + 0.609 + ], + "angle": 0, + "content": "Reconstruction: To measure the upper limit of the modeling capacity of the methods, we report reconstruction metrics on PandaSet in Tab. 3. Here, methods train on all data in a sequence and evaluate on the same views. Further, we enable sensor pose optimization for all methods to account for any pose inaccuracies. Comparing to the NVS setting, all methods except UniSim show improvements with the additional training data. SplatAD achieves SOTA results while rendering \\(\\times 10\\) faster than the previous best method. Further, we note that SplatAD's performance on hold-out validation images in Tab. 1 is on par with previous 3DGS methods' results for reconstruction in Tab. 3." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.61, + 0.911, + 0.807 + ], + "angle": 0, + "content": "Extrapolation: We investigate SplatAD's ability to generalize to views that differ significantly from what is encountered during training. We use the models trained on every other frame and follow the three settings in [35]: shift the ego-vehicle horizontally, shift the ego-vehicle vertically, and apply shifts and rotation to all dynamic actors. We report the Fréchet distance using DINOv2 [28] features, as these have been shown [32] to align better with human perception than using Inception-v3 [33]. However, we note that using Inception-v3 features instead does not change the model ranking or our conclusions. Tab. 4 shows SplatAD's ability to learn meaningful representations for generalization, clearly outperforming other 3DGS methods." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.818, + 0.676, + 0.834 + ], + "angle": 0, + "content": "4.2. Lidar rendering" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.909, + 0.901 + ], + "angle": 0, + "content": "We measure the quality of our lidar point clouds using the same metrics as in [35], i.e., median squared depth error, RMSE intensity error, ray drop accuracy, and chamfer distance, see Tab. 2 for NVS and Tab. 3 for reconstruction. Note" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.948, + 0.52, + 0.96 + ], + "angle": 0, + "content": "11988" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.121, + 0.09, + 0.874, + 0.107 + ], + "angle": 0, + "content": "Table 3. Reconstruction results for image and lidar point clouds on PandaSet. \\(\\S\\) without missing points. First, second, third." + }, + { + "type": "table", + "bbox": [ + 0.154, + 0.117, + 0.842, + 0.21 + ], + "angle": 0, + "content": "
ImageLidarEfficiency
PSNR ↑SSIM ↑LPIPS ↓Depth ↓Intensity ↓Drop acc. ↑CD ↓Camera MP/s ↑Lidar MR/s ↑Train time (H) ↓
PandasetUniSim23.180.6840.3620.060.087-10.2$11.10.80.8
NeuRAD26.760.7780.2330.030.07496.11.79.81.11.3
PVG25.350.7520.42353.18--150.8$23.60.31.6
Street-GS26.680.8080.2933.27--27.9$65.70.82.2
OmniRE26.680.8080.2953.80--33.5$56.70.62.4
SplatAD (ours)29.740.8930.1750.020.05397.41.6104.117.41.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.233, + 0.484, + 0.263 + ], + "angle": 0, + "content": "Table 4. \\( {\\mathrm{{FD}}}_{\\mathrm{{DINOV}}2} \\) scores when shifting pose of ego vehicle or actors. First,second,third ." + }, + { + "type": "table", + "bbox": [ + 0.122, + 0.273, + 0.451, + 0.348 + ], + "angle": 0, + "content": "
Ego lane shiftEgo vert. shiftActor shift
0m2m3m1mRot.Trans.
PandayetNeuRAD377.4540.8629.2571.0424.8420.2
PVG618.8782.7885.6913.9618.8618.8
Street-GS497.1718.6824.2868.7554.7558.0
OmniRE504.5723.2833.5872.8569.6573.2
SplatAD (ours)276.3520.7647.2613.2355.1345.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.363, + 0.484, + 0.391 + ], + "angle": 0, + "content": "Table 5. NVS results averaged over 10 sequences from PandaSet, nuScenes, and Argoverse2 when removing model components." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.395, + 0.48, + 0.514 + ], + "angle": 0, + "content": "
Full modelPSNR ↑SSIM ↑LPIPS ↓Depth ↓CD ↓Camera MP/s ↑Lidar MR/s ↑
27.540.8300.2550.022.0114.111.4
a)Camera RS27.480.8280.2580.022.1116.311.5
b)CNN decoder27.230.8240.2640.022.2178.011.6
c)EWA antialias27.600.8280.2680.021.5114.511.0
d)Appearance emb.27.280.8240.2640.022.1121.011.5
e)Velocity opt.27.550.8300.2570.022.0114.111.4
f)Lidar RS27.420.8270.2600.032.1114.212.2
g)MCMC27.350.8260.2590.012.1130.614.1
h)Median depth27.540.8300.2550.024.9114.111.4
" + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.522, + 0.258, + 0.606 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.258, + 0.522, + 0.481, + 0.606 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.618, + 0.483, + 0.645 + ], + "angle": 0, + "content": "Figure 5. The CNN decoder improves sharpness and is more true to color than the MLP decoder." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.654, + 0.484, + 0.82 + ], + "angle": 0, + "content": "that we do not report intensity or ray drop accuracy for the 3DGS baselines, as they have no way of inferring these. Further, it was not possible to produce Argoverse2 point cloud results for these baselines due to the memory requirements of drivestudio not scaling well with the increase in number of images and higher image resolution. We find SplatAD to be on par with the accurate ray-tracing formulation of NeuRAD in terms of point cloud quality, while running up to \\(\\times 18\\) faster. Furthermore, our lidar rendering approach is superior to the naive depth image-based method used for 3DGS baselines, both in terms of speed and quality." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.833, + 0.203, + 0.847 + ], + "angle": 0, + "content": "4.3. Ablations" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "We validate the effectiveness of key components of our method by measuring their impact on NVS metrics in Tab. 5. Metrics are averaged over all three datasets to avoid any" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.235, + 0.908, + 0.432 + ], + "angle": 0, + "content": "dataset-specific biases. We see that removing the rolling shutter modeling of the camera (a) or lidar (f) reduces NVS performance, and that leaving out optimization of sensor velocities (e) negatively impacts our depth predictions. While the impact of rolling shutter modeling on quantitative metrics is somewhat modest, we qualitatively find that it is important for capturing thin structures and edges, see Fig. 4. Removing these components leads to line-of-sight errors, which erroneously can cut through objects. Note that when removing our lidar rolling shutter compensation, we use ego-motion compensated point clouds for supervision. Moreover, we note the efficiency of our rolling shutter algorithms, as removing them barely improves rendering speeds." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.44, + 0.909, + 0.577 + ], + "angle": 0, + "content": "We also find that using a CNN instead of an MLP for image decoding (c) clearly improves image realism, but at the cost of runtime, see Fig. 5 for an example. Further, we verify that using median depth instead of expected depth during inference is important for realistic lidar data in terms of Chamfer distance (h). Last, we note that MCMC [16] and EWA antialIASing [47, 52] both improve our performance, with the antialIASing having the largest impact on perceptual quality in terms of LPIPS." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.616, + 0.634, + 0.632 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.651, + 0.907, + 0.772 + ], + "angle": 0, + "content": "In this work, we have proposed SplatAD, the first method for rendering camera and lidar data from 3D Gaussians. We combined accurate sensor modeling with efficient algorithms to achieve state-of-the-art novel view synthesis results while improving run-time by more than an order of magnitude. We hope that SplatAD will advance the scalability and fidelity of simulation environments, facilitating safer and more cost-effective testing of AD systems." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Limitations and future work: SplatAD is currently limited to modeling all dynamic actors as rigid. Drawing inspiration from recent advances in human reconstruction [18, 20, 26] can provide inspiration how to overcome this limitation in future research. In addition, we believe that incorporating data-driven priors such as diffusion models [9, 11, 40] is a promising direction to improve the extrapolation performance, and for changing appearance attributes." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.948, + 0.519, + 0.959 + ], + "angle": 0, + "content": "11989" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.092, + 0.245, + 0.108 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.112, + 0.486, + 0.211 + ], + "angle": 0, + "content": "We thank Adam Lilja, William Ljungbergh, and Adam Tenderski for valuable feedback. This work was partially supported by the Wallenberg AI, Autonomous Systems and Software Program (WASP) funded by the Knut and Alice Wallenberg Foundation. Computational resources were provided by NAISS at NSC Berzelius, partially funded by the Swedish Research Council, grant agreement no. 2022-06725." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.236, + 0.188, + 0.251 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.26, + 0.486, + 0.344 + ], + "angle": 0, + "content": "[1] Mina Alibeigi, William Ljungbergh, Adam Tonderski, Georg Hess, Adam Lilja, Carl Lindstrom, Daria Motorniuk, Junsheng Fu, Jenny Widahl, and Christoffer Petersson. Zenseact open dataset: A large-scale and diverse multimodal dataset for autonomous driving. In ICCV, pages 20178-20188, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.344, + 0.484, + 0.4 + ], + "angle": 0, + "content": "[2] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In ICCV, pages 5855-5864, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.401, + 0.485, + 0.455 + ], + "angle": 0, + "content": "[3] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In CVPR, pages 5470–5479, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.457, + 0.485, + 0.51 + ], + "angle": 0, + "content": "[4] Jonathan T. Barron, Ben Mildenhall, Dor Verbin, Pratul P. Srinivasan, and Peter Hedman. Zip-nerf: Anti-aliased grid-based neural radiance fields. In ICCV, pages 19697-19705, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.512, + 0.485, + 0.58 + ], + "angle": 0, + "content": "[5] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In CVPR, pages 11621-11631, 2020. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.581, + 0.485, + 0.636 + ], + "angle": 0, + "content": "[6] Yurui Chen, Chun Gu, Junzhe Jiang, Xiatian Zhu, and Li Zhang. Periodic vibration gaussian: Dynamic urban scene reconstruction and real-time rendering. arXiv preprint arXiv:2311.18561, 2023. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.637, + 0.485, + 0.704 + ], + "angle": 0, + "content": "[7] Ziyu Chen, Jiawei Yang, Jiahui Huang, Riccardo de Lutio, Janick Martinez Esturo, Boris Ivanovic, Or Litany, Zan Gojcic, Sanja Fidler, Marco Pavone, et al. Omnire: Omni urban scene reconstruction. arXiv preprint arXiv:2408.16760, 2024. 1, 2, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.707, + 0.485, + 0.775 + ], + "angle": 0, + "content": "[8] Ziyu Chen, Jiawei Yang, Jiahui Huang, Riccardo de Lutio, Janick Martinez Esturo, Boris Ivanovic, Or Litany, Zan Gojcic, Sanja Fidler, Marco Pavone, Li Song, and Yue Wang. *drivestudio*. https://github.com/ziyc/drivestudio, 2024. 7, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.776, + 0.485, + 0.844 + ], + "angle": 0, + "content": "[9] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In ICML, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.485, + 0.902 + ], + "angle": 0, + "content": "[10] Xiao Fu, Shangzhan Zhang, Tianrun Chen, Yichong Lu, Lanyun Zhu, Xiaowei Zhou, Andreas Geiger, and Yiyi Liao. Panoptic nerf: 3d-to-2d label transfer for panoptic urban scene segmentation. In 3DV, pages 1-11. IEEE, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.26, + 0.486, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.161 + ], + "angle": 0, + "content": "[11] Ruiqi Gao, Aleksander Holynski, Philipp Henzler, Arthur Brussee, Ricardo Martin Brualla, Pratul P. Srinivasan, Jonathan T. Barron, and Ben Poole. CAT3d: Create anything in 3d with multi-view diffusion models. In NeurIPS, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.163, + 0.908, + 0.218 + ], + "angle": 0, + "content": "[12] Shubham Goel, Georgios Pavlakos, Jathushan Rajasegaran, Angjoo Kanazawa, and Jitendra Malik. Humans in 4d: Reconstructing and tracking humans with transformers. In ICCV, pages 14783-14794, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.219, + 0.907, + 0.272 + ], + "angle": 0, + "content": "[13] Shengyu Huang, Zan Gojcic, Zian Wang, Francis Williams, Yoni Kasten, Sanja Fidler, Konrad Schindler, and Or Litany. Neural lidar fields for novel view synthesis. In ICCV, pages 18236-18246, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.274, + 0.906, + 0.316 + ], + "angle": 0, + "content": "[14] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. TOG, 42(4):139-1, 2023. 1, 2, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.317, + 0.907, + 0.371 + ], + "angle": 0, + "content": "[15] Mustafa Khan, Hamidreza Fazlali, Dhruv Sharma, Tongtong Cao, Dongfeng Bai, Yuan Ren, and Bingbing Liu. Autosplat: Constrained gaussian splatting for autonomous driving scene reconstruction, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.372, + 0.906, + 0.427 + ], + "angle": 0, + "content": "[16] Shakiba Kheradmand, Daniel Rebain, Gopal Sharma, Weiwei Sun, Yang-Che Tseng, Hossam Isack, Abhishek Kar, Andrea Tagliasacchi, and Kwang Moo Yi. 3d gaussian splatting as markov chain monte carlo. In NeurIPS, 2024. 5, 6, 8, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.428, + 0.906, + 0.455 + ], + "angle": 0, + "content": "[17] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.456, + 0.906, + 0.497 + ], + "angle": 0, + "content": "[18] Muhammed Kocabas, Jen-Hao Rick Chang, James Gabriel, Oncel Tuzel, and Anurag Ranjan. Hugs: Human gaussian splats. In CVPR, pages 505-515, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.498, + 0.908, + 0.567 + ], + "angle": 0, + "content": "[19] Abhijit Kundu, Kyle Genova, Xiaqi Yin, Alireza Fathi, Caroline Pantofaru, Leonidas J Guibas, Andrea Tagliasacchi, Frank Dellaert, and Thomas Funkhouser. Panoptic neural fields: A semantic object-aware neural scene representation. In CVPR, pages 12871-12881, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.568, + 0.908, + 0.65 + ], + "angle": 0, + "content": "[20] Youngjoong Kwon, Baole Fang, Yixing Lu, Haoye Dong, Cheng Zhang, Francisco Vicente Carrasco, Albert Mosella-Montoro, Jianjin Xu, Shingo Takagi, Daeil Kim, Aayush Prakash, and Fernando De la Torre. Generalizable human gaussians for sparse view synthesis. In ECCV, pages 451-468, Cham, 2025. Springer Nature Switzerland. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.651, + 0.908, + 0.718 + ], + "angle": 0, + "content": "[21] Nanxi Li, Chong Pei Ho, Jin Xue, Leh Woon Lim, Guanyu Chen, Yuan Hsing Fu, and Lennon Yao Ting Lee. A progress review on solid-state lidar and nanophotonics-based lidar sensors. *Laser & Photonics Reviews*, 16(11):2100511, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.72, + 0.908, + 0.788 + ], + "angle": 0, + "content": "[22] Carl Lindström, Georg Hess, Adam Lilja, Maryam Fatemi, Lars Hammarstrand, Christoffer Petersson, and Lennart Svensson. Are nerfs ready for autonomous driving? towards closing the real-to-simulation gap. In CVPRW, pages 4461-4471, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.79, + 0.908, + 0.859 + ], + "angle": 0, + "content": "[23] William Ljungbergh, Adam Tonderski, Joakim Johnander, Holger Caesar, Kalle Åström, Michael Felsberg, and Christoffer Petersson. Neuroncap: Photorealistic closed-loop safety testing for autonomous driving. In ECCV, pages 161-177. Springer, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.908, + 0.902 + ], + "angle": 0, + "content": "[24] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J. Black. SMPL: A skinned multiperson linear model. TOG, 34(6):248:1-248:16, 2015. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.948, + 0.521, + 0.96 + ], + "angle": 0, + "content": "11990" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.162 + ], + "angle": 0, + "content": "[25] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, pages 405-421, Cham, 2020. Springer International Publishing. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.165, + 0.483, + 0.22 + ], + "angle": 0, + "content": "[26] Arthur Moreau, Jifei Song, Helisa Dhamo, Richard Shaw, Yiren Zhou, and Eduardo Pérez-Pellitero. Human gaussian splatting: Real-time rendering of animatable avatars. In CVPR, pages 788-798, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.223, + 0.485, + 0.264 + ], + "angle": 0, + "content": "[27] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. TOG, 41(4):1-15, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.267, + 0.485, + 0.392 + ], + "angle": 0, + "content": "[28] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel HAZIZA, Francisco Massa, Alaaeldin El-Nouby, Mido Assran, Nicolas Ballas, Wojciech Galuba, Russell Howes, Po-Yao Huang, Shang-Wen Li, Ishan Misra, Michael Rabbat, Vasu Sharma, Gabriel Synnaeve, Hu Xu, Herve Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. DINov2: Learning robust visual features without supervision. TMLR, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.394, + 0.483, + 0.436 + ], + "angle": 0, + "content": "[29] Julian Ost, Fahim Mannan, Nils Thuerey, Julian Knodt, and Felix Heide. Neural scene graphs for dynamic scenes. In CVPR, pages 2856-2865, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.439, + 0.483, + 0.493 + ], + "angle": 0, + "content": "[30] Konstantinos Rematas, Andrew Liu, Pratul P Srinivasan, Jonathan T Barron, Andrea Tagliasacchi, Thomas Funkhouser, and Vittorio Ferrari. Urban radiance fields. In CVPR, pages 12932-12942, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.497, + 0.483, + 0.579 + ], + "angle": 0, + "content": "[31] Otto Seiskari, Jerry Yilammi, Valtteri Kaatrasalo, Pekka Rantalankila, Matias Turkulainen, Juho Kannala, Esa Rahtu, and Arno Solin. Gaussian splattering on the move: Blur and rolling shutter compensation for natural camera motion. In ECCV, pages 160-177, Cham, 2025. Springer Nature Switzerland. 4, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.582, + 0.483, + 0.651 + ], + "angle": 0, + "content": "[32] George Stein, Jesse Cresswell, Rasa Hosseinzadeh, Yi Sui, Brendan Ross, Valentin Villecloze, Zhaoyan Liu, Anthony L Caterini, Eric Taylor, and Gabriel Loaiza-Ganem. Exposing flaws of generative model evaluation metrics and their unfair treatment of diffusion models. NeurIPS, 36, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.655, + 0.483, + 0.709 + ], + "angle": 0, + "content": "[33] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. In CVPR, pages 2818-2826, 2016. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.713, + 0.483, + 0.754 + ], + "angle": 0, + "content": "[34] Adam Tondersski, Carl Lindstrom, Georg Hess, and William Ljungbergh. neurad-studio. https://github.com/georghess/neurad-studio, 2024. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.757, + 0.483, + 0.813 + ], + "angle": 0, + "content": "[35] Adam Tondersski, Carl Lindström, Georg Hess, William Ljungbergh, Lennart Svensson, and Christoffer Petersson. Neurad: Neural rendering for autonomous driving. In CVPR, pages 14895-14904, 2024. 1, 2, 3, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.816, + 0.483, + 0.857 + ], + "angle": 0, + "content": "[36] Haithem Turki, Jason Y Zhang, Francesco Ferroni, and Deva Ramanan. Suds: Scalable urban dynamic scenes. In CVPR, pages 12375-12385, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.859, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[37] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE TIP, 13(4):600-612, 2004. 7" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.908, + 0.203 + ], + "angle": 0, + "content": "[38] Benjamin Wilson, William Qi, Tanmay Agarwal, John Lambert, Jagjeet Singh, Siddhesh Khandelwal, Bowen Pan, Ratnesh Kumar, Andrew Hartnett, Jhony Kaesemodel Pontes, Deva Ramanan, Peter Carr, and James Hays. Argoverse 2: Next generation datasets for self-driving perception and forecasting. In Proceedings of the Neural Information Processing Systems Track on Datasets and Benchmarks (NeurIPS Datasets and Benchmarks 2021), 2021. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.206, + 0.908, + 0.26 + ], + "angle": 0, + "content": "[39] Hanfeng Wu, Xingxing Zuo, Stefan Leutenegger, Or Litany, Konrad Schindler, and Shengyu Huang. Dynamic lidar resimulation using compositional neural fields. In CVPR, pages 1998-19998, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.262, + 0.908, + 0.33 + ], + "angle": 0, + "content": "[40] Rundi Wu, Ben Mildenhall, Philipp Henzler, Keunhong Park, Ruiqi Gao, Daniel Watson, Pratul P Srinivasan, Dor Verbin, Jonathan T Barron, Ben Poole, et al. Reconfusion: 3d reconstruction with diffusion priors. In CVPR, pages 21551-21561, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.333, + 0.906, + 0.416 + ], + "angle": 0, + "content": "[41] Pengchuan Xiao, Zhenlei Shao, Steven Hao, Zishuo Zhang, Xiaolin Chai, Judy Jiao, Zesong Li, Jian Wu, Kai Sun, Kun Jiang, Yunlong Wang, and Diange Yang. Pandaset: Advanced sensor suite dataset for autonomous driving. In 2021 IEEE International Intelligent Transportation Systems Conference (ITSC), pages 3095-3101, 2021. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.418, + 0.906, + 0.459 + ], + "angle": 0, + "content": "[42] Ziyang Xie, Junge Zhang, Wenye Li, Feihu Zhang, and Li Zhang. S-neRF: Neural radiance fields for street views. In ICLR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.461, + 0.906, + 0.515 + ], + "angle": 0, + "content": "[43] Yunzhi Yan, Haotong Lin, Chenxu Zhou, Weijie Wang, Haiyang Sun, Kun Zhan, Xianpeng Lang, Xiaowei Zhou, and Sida Peng. Street gaussians for modeling dynamic urban scenes. In ECCV, 2024. 1, 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.518, + 0.908, + 0.585 + ], + "angle": 0, + "content": "[44] Jiawei Yang, Boris Ivanovic, Or Litany, Xinshuo Weng, Seung Wook Kim, Boyi Li, Tong Che, Danfei Xu, Sanja Fidler, Marco Pavone, and Yue Wang. EmerneRF: Emergent spatial-temporal scene decomposition via self-supervision. In ICLR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.588, + 0.908, + 0.643 + ], + "angle": 0, + "content": "[45] Ze Yang, Yun Chen, Jingkang Wang, Sivabalan Manivasagam, Wei-Chiu Ma, Anqi Joyce Yang, and Raquel Urtasun. Unisim: A neural closed-loop sensor simulator. In CVPR, pages 1389-1399, 2023. 1, 2, 7, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.646, + 0.908, + 0.701 + ], + "angle": 0, + "content": "[46] Vickie Ye, Ruilong Li, Justin Kerr, Matias Turkulainen, Brent Yi, Zhuoyang Pan, Otto Seiskari, Jianbo Ye, Jeffrey Hu, Matthew Tancik, et al. gsplat: An open-source library for gaussian splatting. arXiv preprint arXiv:2409.06765, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.703, + 0.908, + 0.744 + ], + "angle": 0, + "content": "[47] Zehao Yu, Anpei Chen, Binbin Huang, Torsten Sattler, and Andreas Geiger. Mip-splatting: Alias-free 3d gaussian splatting. In CVPR, pages 19447-19456, 2024. 4, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.746, + 0.906, + 0.8 + ], + "angle": 0, + "content": "[48] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, pages 586-595, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.803, + 0.906, + 0.845 + ], + "angle": 0, + "content": "[49] Zehan Zheng, Fan Lu, Weiyi Xue, Guang Chen, and Changjun Jiang. Lidar4d: Dynamic neural fields for novel space-time view lidar synthesis. In CVPR, pages 5145-5154, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.847, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[50] Hongyu Zhou, Jiahao Shao, Lu Xu, Dongfeng Bai, Weichao Qiu, Bingbing Liu, Yue Wang, Andreas Geiger, and Yiyi Liao. Hugs: Holistic urban 3d scene understanding via gaussian splatting. In CVPR, pages 21336-21345, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.908, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.948, + 0.518, + 0.959 + ], + "angle": 0, + "content": "11991" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.487, + 0.149 + ], + "angle": 0, + "content": "[51] Xiaoyu Zhou, Zhiwei Lin, Xiaojun Shan, Yongtao Wang, Deqing Sun, and Ming-Hsuan Yang. Drivinggaussian: Composite gaussian splatting for surrounding dynamic autonomous driving scenes. In CVPR, pages 21634-21643, 2024. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.486, + 0.203 + ], + "angle": 0, + "content": "[52] Matthias Zwicker, Hanspeter Pfister, Jeroen Van Baar, and Markus Gross. Ewa volume splatting. In Proceedings Visualization, 2001. VIS'01., pages 29-538. IEEE, 2001. 4, 8" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.487, + 0.203 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.948, + 0.52, + 0.96 + ], + "angle": 0, + "content": "11992" + } + ] +] \ No newline at end of file diff --git a/2025/SplatAD_ Real-Time Lidar and Camera Rendering with 3D Gaussian Splatting for Autonomous Driving/58aed1a8-44d5-4314-9f37-1456921ac406_origin.pdf b/2025/SplatAD_ Real-Time Lidar and Camera Rendering with 3D Gaussian Splatting for Autonomous Driving/58aed1a8-44d5-4314-9f37-1456921ac406_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..736bf06068b38eeedaae8e22b7067dc365e8a0aa --- /dev/null +++ b/2025/SplatAD_ Real-Time Lidar and Camera Rendering with 3D Gaussian Splatting for Autonomous Driving/58aed1a8-44d5-4314-9f37-1456921ac406_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:44fb96113163a536b53463659622db2b922b8d8d9ca5baf6d3fabc0fc8ed1b40 +size 9498595 diff --git a/2025/SplatAD_ Real-Time Lidar and Camera Rendering with 3D Gaussian Splatting for Autonomous Driving/full.md b/2025/SplatAD_ Real-Time Lidar and Camera Rendering with 3D Gaussian Splatting for Autonomous Driving/full.md new file mode 100644 index 0000000000000000000000000000000000000000..7f97e9bc106e0c3fccd7ec537a0c9f7f034d0d64 --- /dev/null +++ b/2025/SplatAD_ Real-Time Lidar and Camera Rendering with 3D Gaussian Splatting for Autonomous Driving/full.md @@ -0,0 +1,324 @@ +# SplatAD: Real-Time Lidar and Camera Rendering with 3D Gaussian Splatting for Autonomous Driving + +Georg Hess $^{\dagger,1,2}$ Carl Lindström $^{\dagger,1,2}$ Maryam Fatemi $^{1}$ Christoffer Petersson $^{1,2}$ Lennart Svensson $^{2}$ + +$^{1}$ Zenseact $^{2}$ Chalmers University of Technology + +{firstname_lastname}@{zenseact.com, chalmers.se} + +![](images/19714502aa07182b3219d434c13f7483494374467b078db7a0a7adf3d8faa014.jpg) +Figure 1. SplatAD is the first method capable of realistic camera and lidar rendering using 3D Gaussian Splatting. Whereas previous methods are either fast or multi-modal, SplatAD enables real-time, high-quality rendering for both camera and lidar. In addition, SplatAD can reach competitive performance, e.g., PSNR for image and Chamfer distance for point cloud, within minutes. + +![](images/391b65fc110a83970041c0db396b49d439b7bd8ff76cf9715a9ed4edfb8efdcd.jpg) + +![](images/8fb6c77975566fdd0f4afdeb0be7cbcb33e13040dcf3f5679d9b54e1034459ea.jpg) + +![](images/e3b1848f205a306c38b098b596956599caf495e37462f7d9e6386292040a36da.jpg) + +# Abstract + +Ensuring the safety of autonomous robots, such as self-driving vehicles, requires extensive testing across diverse driving scenarios. Simulation is a key ingredient for conducting such testing in a cost-effective and scalable way. Neural rendering methods have gained popularity, as they can build simulation environments from collected logs in a data-driven manner. However, existing neural radiance field (NeRF) methods for sensor-realistic rendering of camera and lidar data suffer from low rendering speeds, limiting their applicability for large-scale testing. While 3D Gaussian Splatting (3DGS) enables real-time rendering, current methods are limited to camera data and are unable to render lidar data essential for autonomous driving. To address these limitations, we propose SplatAD, the first 3DGS-based method for realistic, real-time rendering of dynamic scenes for both camera and lidar data. SplatAD accurately models key sensor-specific phenomena such as rolling shutter effects, lidar intensity, and lidar ray dropouts, using purpose-built algorithms to optimize rendering efficiency. Evaluation across three autonomous driving datasets demonstrates that SplatAD achieves state-of-the-art rendering quality with up to $+2$ PSNR for NVS and $+3$ PSNR for reconstruction while increasing rendering speed over NeRF-based methods by an order of magnitude. See here for our project page. + +# 1. Introduction + +Large-scale testing is essential to ensure the safety of autonomous robots, such as self-driving vehicles (SDVs), before real-world deployment. Data-driven methods that generate digital twins from collected logs offer a scalable way to build diverse, realistic simulation environments for testing. Unlike real-world testing, which can be costly, time-intensive, and limited by physical constraints, simulation enables rapid, low-cost exploration of multiple scenarios, helping to optimize SDVs for safety, comfort, and efficiency. Motivated by this, a multitude of methods have emerged based on neural radiance fields (NeRFs) [25, 35, 36, 45] and 3D Gaussian Splatting (3DGS) [14, 43, 51]. + +Recent NeRF-based methods [35, 45] provide high-fidelity sensor simulation jointly for both camera and lidar, matching the most common sensor rigs in popular autonomous driving (AD) datasets [1, 5, 38]. However, the slow rendering speed of NeRF-based methods makes them costly and challenging to use for large-scale testing. 3DGS provides an attractive alternative to NeRFs, as its accelerated rendering achieves comparable image realism while increasing inference speeds by an order of magnitude. Nevertheless, 3DGS-based methods for the AD setting [7, 43, 51] inherit the limitation of only being able to render camera data, overlooking the lidar modality. The lidar's ability to directly perceive the 3D environment makes it a powerful tool in modern AD stacks and, consequently, an important modality to simulate. + +In this paper, we aim to solve efficient, differentiable, and realistic rendering of camera and lidar data using 3D Gaussian Splatting. Applying 3DGS to lidar sensors presents unique challenges due to their distinct characteristics. First, unlike cameras, which capture dense, regularly spaced pixels, lidars record sparse, non-linear point clouds with large empty gaps from non-returning rays. Second, most lidars capture a full $360^{\circ}$ view of the scene, where existing methods often project data into multiple depth images [7]—an inefficient approach that overlooks the sparse structure of the lidar. Last, lidars suffer from a rolling shutter effect, with each sweep taking up to $100\mathrm{ms}$ , during which the ego vehicle may move several meters, violating 3DGS's single-origin assumption. + +To overcome these challenges, we introduce SplatAD, a novel view synthesis method that unifies camera and lidar rendering and is designed for real-time rendering of large-scale dynamic traffic scenes. Our method replaces spherical harmonics with per-Gaussian learnable features to jointly model sensor-specific phenomena—from lidar ray drop and intensity variations to camera-specific appearance changes. By introducing purpose-built rasterization algorithms for rendering lidar in spherical coordinates, we achieve both superior efficiency and realism. Further, we demonstrate efficient methods for modeling rolling shutter effects across both sensor modalities. We verify the effectiveness and generalizability of our method across three popular automotive datasets, achieving state-of-the-art results on all of them. To summarize, our contributions are as follows: + +- We propose the first method for efficient lidar rendering using 3D Gaussians, introducing custom CUDA-accelerated algorithms for rasterizing sparse point clouds in spherical coordinates. +- We introduce the first 3DGS method capable of rendering both camera and lidar from a unified representation, enabling accelerated scaling of novel view synthesis for automotive applications. +- We present effective techniques for realistic sensor modeling using 3D Gaussians, enabling accurate handling of rolling shutter, lidar intensity, ray dropping, and variations in sensor appearance. +- Through extensive evaluation on three popular automotive datasets, we demonstrate state-of-the-art results across all benchmarks, validating our method's effectiveness and generalizability. + +# 2. Related work + +NeRFs for automotive data: Since the introduction of NeRFs, neural representations have become central to 3D reconstruction and novel view synthesis [2-4, 25]. Many works have applied NeRF-based methods to automotive data [35, 36, 44, 45], enabling sensor-realistic renderings of novel views in large and dynamic scenes. Recent advances provide realism sufficient for downstream applica + +tions [22, 23]. While early methods focused on either camera [10, 19, 29, 42] or lidar [13, 39, 49], newer methods aim to handle them jointly. UniSim [45] showcased realistic renderings for front camera and $360^{\circ}$ lidar on data from PandaSet [41]. The method uses a hash grid representation [27], with separate features for the sky, static background, and dynamic actors, and decodes the color and lidar intensity from volume-rendered features. NeuRAD [35] proposes a simplified network architecture and improved sensor modeling, achieving state-of-the-art results for the full $360^{\circ}$ camera and lidar rig. Even so, the slow rendering speeds of NeRF-based methods make scaling costly and challenging. SplatAD aims to overcome this limitation by using rasterization-based CUDA-accelerated algorithms to render both camera and lidar data while, inspired by NeuRAD, emphasizing the modeling of important sensor characteristics for increased realism. + +3DGS for automotive data: 3DGS [14] uses an explicit representation paired with rasterization techniques and custom hardware-accelerated algorithms to achieve real-time rendering. Additionally, 3DGS has been shown to create faithful scene reconstructions with rendering quality close to recent NeRF-based methods. As a result, several works have applied 3DGS for camera rendering on automotive data [7, 15, 43, 50, 51]. Periodic Vibration Gaussian (PVG) [6] applies 3DGS to dynamic scenes by learning the 3D flow of all Gaussians. Yet, PVG's lack of an explicit actor representation limits the method's controllability and, hence, applicability for simulation. Street Gaussians [43], like Unisim and NeuRAD, instead decomposes the scene into static background and rigid dynamic actors using 3D bounding boxes. The authors further extend 3DGS by adding Fourier coefficients to capture time-varying characteristics of dynamic actors. Moreover, to handle the sky regions, their method uses semantic masks and a view-dependent cube map. OmniRe [7] enhances the scene graph composition with non-rigid nodes for better modeling of pedestrians and cyclists. However, this requires pre-processing to track and align human body poses [12] for initializing SMPL [24] pose parameters. + +PVG, Street Gaussians, and OmniRe all use lidar points for initialization and depth supervision. However, to generate new point clouds, these methods predict depth images and rely on existing lidar points to query said image at the right locations. This is both inefficient and inaccurate, as $360^{\circ}$ lidars are approximated as multiple cameras. Further, neither of these methods model important lidar characteristics such as intensity variations, ray drop, and rolling shutter. In contrast, our method supports efficient rendering of both camera and lidar from 3D Gaussians in novel views, and models important sensor characteristics for both modalities. + +# 3. Method + +Our aim is to learn a scene representation from collected vehicle logs that enables rendering of realistic camera and lidar + +![](images/3536cb69060a185c55a3bdfeba2bb4a0a85014e2298f4ddbd805a9979bdf1cc2.jpg) +Figure 2. Overview of our proposed method. Given the composition of static and dynamic 3D Gaussians, SplatAD is capable of differentiable rendering of both lidar and camera data. Our proposed lidar rendering matches the image rendering on a high level, but modifies each component to accurately model sensor characteristics. Our method projects 3D Gaussians with associated feature vectors onto the corresponding sensor modalities (camera and lidar) and employs sensor-specific tiling to match their distinct characteristics. During rasterization, the projected Gaussians are corrected for rolling shutter effects caused by the movement of sensors and, potentially, their own velocity. Finally, the rasterized features are decoded into the respective image and lidar point cloud representations. + +data, with the ability to alter the locations of both the ego vehicle and other actors. To scale effectively, the rendering process must be fast, as faster inference speeds enhance its practicality for applications. In the following, we describe our scene representation (Sec. 3.1), rendering algorithms (Secs. 3.2 and 3.3), and implementation and optimization strategy (Sec. 3.4), see Fig. 2 for an overview. + +# 3.1. Scene representation + +Our scene representation builds upon 3DGS [14], but with key changes to handle the specifics of AD scenes and to enable both camera and lidar rendering from the same representation. Like 3DGS, each scene is represented by a set of translucent 3D Gaussians with learnable occupancy $o \in (0,1)$ , mean $\mu \in \mathbb{R}^3$ , and covariance matrix $\Sigma \in \mathbb{R}^{3\times 3}$ which in turn is parameterized by scale $\mathbf{S} \in \mathbb{R}^3$ and quaternion $\mathbf{q} \in \mathbb{R}^4$ parameters. Instead of spherical harmonics [14], we assign each Gaussian a learnable base color $\mathbf{f}_{\mathrm{rgb}}^{\mathrm{b}} \in \mathbb{R}^3$ and feature vector $\mathbf{f} \in \mathbb{R}^{D_{\mathrm{f}}}$ , where $\mathbf{f}$ is used for both view-dependent effects and lidar properties. Last, our representation contains a learnable embedding per sensor to model their specific appearance characteristics. + +To handle dynamics, we follow the commonly used scene graph decomposition [29, 35, 43, 51] and divide the scene into a static background and a set of dynamic actors. Each dynamic actor is described by a 3D bounding box and a sequence of SE(3) poses, obtained either from an off-the-shelf object detector and tracker, or annotations. Each Gaussian has a non-learnable ID, indicating whether it is assigned to the static world, or to which actor it belongs. For Gaussians + +assigned to actors, their mean and covariance are expressed in the local coordinate system of the corresponding axis-aligned bounding box. To compose the scene at a given time $t$ , Gaussians assigned to bounding boxes are transformed to world coordinates based on their actor's pose. Since pose estimates for actors may contain inaccuracies, we adjust these with learnable offsets. Furthermore, each actor has a velocity initialized from pose differences and a learnable velocity offset. + +# 3.2. Camera rendering + +Given a posed camera, we compose the set of Gaussians at the corresponding capture time $t$ and render an image $I$ using the efficient tile-based rendering from 3DGS [14]. While we retain 3DGS's high-level steps—projection and view frustum culling, tile-assignment, depth sorting, and tile-based rasterization—we introduce key adaptations to better model the unique characteristics of AD data. + +Projection, tiling, and sorting: Each mean and covariance is transformed from world to camera coordinates, giving $\pmb{\mu}^{C}$ and $\pmb{\Sigma}^{C}$ . The means are then converted to image space $\pmb{\mu}^{I} \in \mathbb{R}^{2}$ using the perspective projection, while the covariances are transformed using the two upper rows of the Jacobian of the projection, $\pmb{\Sigma}^{I} = \mathbf{J}^{I}\pmb{\Sigma}^{C}(\mathbf{J}^{I})^{\mathrm{T}} \in \mathbb{R}^{2 \times 2}$ . Gaussians outside the view frustum are culled, where, for efficiency, 3DGS approximates the extent of the Gaussians using a square axis-aligned bounding box (AABB) that covers their $99\%$ confidence level. Further, 3DGS divides the image into tiles, each measuring $16 \times 16$ pixels, and assigns Gaussians to all tiles they intersect, duplicating them if necessary. This + +way, each pixel only has to process a subset of all Gaussians during rasterization. Lastly, the Gaussians are sorted based on the $z$ -depth of their mean $\pmb{\mu}^{C}$ . + +Rolling shutter: Many cameras use a rolling shutter, where the capture of an image is not instantaneous, but is done row-by-row, allowing the camera to move during the exposure. Previous work [35] has highlighted the importance of modeling rolling shutter effects due to the high sensor velocities present in AD data, where their ray-tracing-based method can easily account for this by shifting the origins of each ray. For 3DGS, the equivalent would require projecting all 3D Gaussians to all camera poses encountered during an exposure, as the Gaussians' positions relative to the camera change with time. As this would be prohibitively expensive, we draw inspiration from [31] and instead approximate the effects directly in 2D image space. Specifically, each Gaussian's velocity relative to the camera is projected to image space, and their pixel mean $\mu^I$ is adjusted during rasterization based on the capture time of the pixel. However, [31] considered static scenes only, hence we adjust the formulation to account for the dynamics as well. + +For each Gaussian, its pixel velocity is approximated as + +$$ +\mathbf {v} ^ {I} = \mathbf {J} ^ {I} \left(- \boldsymbol {\omega} _ {C} \times \boldsymbol {\mu} ^ {C} - \mathbf {v} _ {C} + \mathbf {v} _ {\mathrm {d y n}}\right) \in \mathbb {R} ^ {2}, \tag {1} +$$ + +where $\omega_{C}$ and $\mathbf{v}_C$ denote the angular and linear velocity of the camera expressed in the camera coordinate system, and + +$$ +\mathbf {v} _ {\mathrm {d y n}} = T ^ {\mathrm {a c t} \rightarrow C} \left(\boldsymbol {\omega} _ {\mathrm {a c t}} \times \boldsymbol {\mu} ^ {\mathrm {a c t}} + \mathbf {v} _ {\mathrm {a c t}}\right) \in \mathbb {R} ^ {3}, \tag {2} +$$ + +denotes the velocity of the Gaussian induced by the angular $\omega^{\mathrm{dyn}}$ and linear velocity $\mathbf{v}^{\mathrm{dyn}}$ of its associated actor (expressed in actor coordinate system) and $\mu^{\mathrm{act}}$ denotes the Gaussian mean in the actor coordinate system. To transform velocities from actor to camera coordinates, $T^{\mathrm{act}\rightarrow C}$ is the composition of the actor-to-world and world-to-camera transforms. For Gaussians that are part of the static background, $\mathbf{v}_{\mathrm{dyn}}$ is zero. See Appendix D for further details. + +We account for the pixel velocity when culling Gaussians and checking intersections between Gaussians and tiles by increasing the approximated Gaussian extent. Instead of the square AABB used by 3DGS and [31], we use a rectangular AABB covering $\boldsymbol{\Sigma}^{I}$ at three standard deviations and add $|\mathbf{v}_I| \cdot t_{\mathrm{rs}} / 2$ to its extent. Here, $t_{\mathrm{rs}}$ denotes the rolling shutter duration, i.e., the time difference between the last and first row of pixels. The increase in extent corresponds to the area covered by the Gaussian mean during the rolling shutter time, assuming the sensor timestamp is in the middle of exposure. As this is not always the case, we further include a learnable time offset, describing the difference between the sensor timestamp and the middle of the exposure. Using rectangular AABBs reduces unnecessary intersections compared to [14, 31], especially for narrow Gaussians and Gaussians with axis-aligned velocities, the latter often being the case for side-facing cameras. + +Rasterization: 3DGS rasterizes pixels in parallel by launching a thread block per tile and assigning each thread to a pixel within the tile. For each pixel, its coordinates $\mathbf{p} = [p_u,p_v]^{\mathrm{T}}$ are inferred from the block and the thread index. From this, we find the time difference between the pixel's capture time and the image's middle row as + +$$ +t _ {\mathrm {p i x}} = \left(p _ {v} / H - 0. 5\right) \cdot t _ {\mathrm {r s}}. \tag {3} +$$ + +To rasterize a pixel, we $\alpha$ -blend RGB values $\mathbf{f}_i^{\mathrm{rgb}}$ and features $\mathbf{f}_i$ of the depth-sorted Gaussians intersecting the current tile + +$$ +\left[ \mathbf {F} _ {\mathbf {p}} ^ {\mathrm {r g b}}, \mathbf {F} _ {\mathbf {p}} \right] = \sum_ {i} \left[ \mathbf {f} _ {i} ^ {\mathrm {r g b}}, \mathbf {f} _ {i} \right] \alpha_ {i} \prod_ {j = 1} ^ {i - 1} (1 - \alpha_ {j}), \tag {4} +$$ + +with $\alpha$ computed as + +$$ +\alpha_ {i} = \sqrt {\frac {\left| \boldsymbol {\Sigma} _ {i} ^ {I} \right|}{\left| \boldsymbol {\Sigma} _ {i} ^ {I} + s \mathbf {I} \right|}} o _ {i} \exp \left(- \frac {1}{2} \boldsymbol {\Delta} _ {i} ^ {\mathrm {T}} \left(\boldsymbol {\Sigma} _ {i} ^ {I} + s \mathbf {I}\right) ^ {- 1} \boldsymbol {\Delta} _ {i}\right) \tag {5} +$$ + +and + +$$ +\boldsymbol {\Delta} _ {i} = \mathbf {p} - \left(\boldsymbol {\mu} _ {i} ^ {I} + \mathbf {v} _ {i} ^ {I} t _ {\mathrm {p i x}}\right). \tag {6} +$$ + +As in [31], we calculate the distance $\Delta_{i}$ between pixels and Gaussians using the means compensated by rolling shutter, effectively shifting Gaussians to the correct location in the image. Compared to 3DGS, in Eq. (5) we use the 3DGS+EWA [52] formulation from [47], where $s = 0.3$ . + +View-dependent effects are modeled using a small CNN; given the feature map $\mathbf{F} \in \mathbb{R}^{H \times W \times D_{\mathrm{f}}}$ , the corresponding ray directions $\mathbf{d} \in \mathbb{R}^{H \times W \times 3}$ , and a camera-specific learned embedding $\mathbf{e}$ , it predicts a pixel-wise affine mapping $M \in \mathbb{R}^{H \times W \times 3}$ and $\mathbf{b} \in \mathbb{R}^{H \times W \times 3}$ applied to $\mathbf{F}^{\mathrm{rgb}}$ + +$$ +M, \mathbf {b} = \operatorname {C N N} (\operatorname {c o n c a t} (\mathbf {F}, \mathbf {d}, \mathbf {e})), \tag {7} +$$ + +$$ +I = (1 + M) \odot \mathbf {F} ^ {\mathrm {r g b}} + \mathbf {b}. \tag {8} +$$ + +We found that using a small CNN instead of an MLP is essential for efficient texture modeling, e.g., for high-frequency road surfaces. In addition, inspired by [30], the camera-specific learned embeddings enable modeling differences in exposure. + +# 3.3. Lidar rendering + +Lidar sensors enable self-driving vehicles to measure the distance and reflectivity (intensity) of a discrete set of points. They do so by emitting laser beam pulses and measuring the time of flight to determine distance and returning power for reflectivity. Most AD datasets deploy lidars that use several laser diodes (16-128) mounted in a vertical array, where the array of diodes is rotated to capture $360^{\circ}$ data and a single scan often takes $100\mathrm{ms}$ to capture. Therefore, we focus on this type of lidar. However, we note that modeling other types of lidars, such as solid-state lidars [21], can be easily + +done in our method by modifying the projection accordingly. Our proposed lidar rendering follows the high-level steps of the image rendering, but we modify each component to accurately model lidar characteristics. + +Projection: For a posed lidar and Gaussians composed at its capture time $t$ , we transform each mean and covariance from world to lidar coordinates, yielding $\boldsymbol{\mu}^{L} = [x,y,z]^{\mathrm{T}}$ and $\boldsymbol{\Sigma}^{L}$ . Raw lidar data are generally captured in spherical coordinates azimuth $\phi$ , elevation $\omega$ , and range $r$ . Thus, we transform Gaussian means from lidar coordinates to spherical coordinates + +$$ +\boldsymbol {\mu} ^ {S} = \left[ \begin{array}{l} \phi \\ \omega \\ r \end{array} \right] = \left[ \begin{array}{c} \arctan 2 (y, x) \\ \arcsin (z / r) \\ \sqrt {x ^ {2} + y ^ {2} + z ^ {2}} \end{array} \right], \tag {9} +$$ + +and convert the covariance $\pmb{\Sigma}^{L}$ with the Jacobian of the transform Eq. (9) as $\pmb{\Sigma}^{S} = \mathbf{J}^{S}\pmb{\Sigma}^{L}(\mathbf{J}^{S})^{\mathrm{T}}$ , where + +$$ +\mathbf {J} ^ {S} = \left[ \begin{array}{c c c} - \frac {y}{x ^ {2} + y ^ {2}} & \frac {x}{x ^ {2} + y ^ {2}} & 0 \\ - \frac {x z}{r ^ {2} \sqrt {x ^ {2} + y ^ {2}}} & - \frac {y z}{r ^ {2} \sqrt {x ^ {2} + y ^ {2}}} & \frac {\sqrt {x ^ {2} + y ^ {2}}}{r ^ {2}} \\ \frac {x}{r} & \frac {y}{r} & \frac {z}{r} \end{array} \right], \tag {10} +$$ + +and $\pmb{\mu}^{S}$ and $\pmb{\Sigma}^{S}$ denote mean and covariance in spherical coordinates. + +Same as for the camera, we account for the rolling shutter effect by approximating its impact in sensor space + +$$ +\mathbf {v} ^ {S} = \mathbf {J} ^ {S} \left(- \boldsymbol {\omega} _ {L} \times \boldsymbol {\mu} ^ {L} - \mathbf {v} _ {L} + \mathbf {v} _ {\mathrm {d y n}}\right) \in \mathbb {R} ^ {3}. \tag {11} +$$ + +Note that $\mathbf{v}^S$ is three-dimensional as opposed to $\mathbf{v}^I$ in Eq. (1), as we need this for accurate range rasterization. Last, we use the first two elements of $\mathbf{v}^S t_{\mathrm{rs}} / 2$ to increase the extent of the rectangular AABB approximating the size of the Gaussian, and cull Gaussians outside the lidar's field-of-view. + +Tiling and sorting: At the core of 3DGS's efficiency is its tile-based rasterization. While lidars often have a fixed azimuth resolution, many use non-equidistant spacing between diodes to get higher resolution in areas of interest, resulting in nonlinear elevation resolution, as seen in Fig. 1. In line with this, we design tiles with a fixed number of diodes vertically and a fixed azimuth resolution horizontally. Horizontally, intersecting tiles are found by dividing azimuth coordinates by the common tile azimuth size and wrapping the results within $360^{\circ}$ . Vertically, we sort the elevation tile boundaries and find intersections with a short for-loop, see Appendix A for details. After tile assignment, Gaussians are sorted based on the range of their mean $\mu^S$ . + +Comparing our strategy to using tiles of equal size, similar to the depth image-based rendering in [7], we avoid many unnecessary computations. There, the tile size must be determined from the finest resolution to ensure having at least as many threads as lidar points. Consequently, tiles + +in areas where the lidar is sparse would have a superficially high resolution and might even lack the corresponding lidar points to render. + +Rasterization: Lidar data are rendered using our proposed differentiable tile-based rasterizer. Again, each tile launches one thread block, where each thread is responsible for rasterizing one lidar point. Each thread is provided the azimuth, elevation, and capture time for it to rasterize. These are either determined from lidar specifications or, for training, calculated from existing point clouds, see Appendix A for details. The reason for this is that for training, one cannot expect the data to follow perfectly straight scan lines. Therefore, we cannot follow the image strategy and rely on the block and thread index to directly infer the pixel coordinates. Further, the azimuth and elevation of a lidar ray do not necessarily determine its time offset to the center of the lidar scan. + +For each lidar point, we $\alpha$ -blend the Gaussians' features $\mathbf{f}_i$ using Eqs. (4) to (6). Inspired by [47] arguing for $s$ in Eq. (5) to correspond to one pixel, we set it to correspond to the geometric mean of the lidar's vertical and horizontal beam divergence. The rendered features are concatenated with their corresponding ray direction and decoded to intensity and ray drop probability using a small MLP. + +To render the expected range of a point, we $\alpha$ -blend the rolling shutter corrected ranges for each Gaussian + +$$ +r _ {i, \mathrm {r s}} = r _ {i} + \mathbf {v} _ {r} ^ {S} t _ {l}, \tag {12} +$$ + +where $\mathbf{v}_r^S$ is the range-component of the relative speed to the lidar and $t_l$ is the time between the capture of the current lidar point, and the middle of the lidar scan. Further, we render the median range as the rolling shutter corrected range of the first Gaussian that satisfies $\prod_{j=1}^{i}(1 - \alpha_j) < 0.5$ during the $\alpha$ -blending in Eq. (4). While we use the expected range for training, the median range is used during inference as it, in contrast to the expected range, does not yield depths between Gaussians. + +# 3.4. Optimization and implementation + +All model components are optimized jointly using the loss + +$$ +\begin{array}{l} \mathcal {L} = \lambda_ {r} \mathcal {L} _ {1} + (1 - \lambda_ {r}) \mathcal {L} _ {\text {S S I M}} + \lambda_ {\text {d e p t h}} \mathcal {L} _ {\text {d e p t h}} + \lambda_ {\text {l o s}} \mathcal {L} _ {\text {l o s}} + \\ \lambda_ {\text {i n t e n s}} \mathcal {L} _ {\text {i n t e n}} + \lambda_ {\text {r a y d r o p}} \mathcal {L} _ {\text {B C E}} + \lambda_ {\text {M C M C}} \mathcal {L} _ {\text {M C M C}}, \tag {13} \\ \end{array} +$$ + +where $\mathcal{L}_1$ and $\mathcal{L}_{\mathrm{SSIM}}$ are L1 and SSIM losses on the rendered images. $\mathcal{L}_{\mathrm{depth}}$ and $\mathcal{L}_{\mathrm{inten}}$ are L2 losses on the rendered expected lidar range and intensity. $\mathcal{L}_{\mathrm{los}}$ is a line-of-sight loss, penalizing opacity before the ground truth lidar range. $\mathcal{L}_{\mathrm{BCE}}$ is a binary cross-entropy loss on the predicted ray drop probability, where ground-truth is generated in the same way as for NeuRAD. $\mathcal{L}_{\mathrm{MCMC}}$ is the opacity and scale regularization used in [16]. + +![](images/9b7e4808099cb50ebee8698abe578747e00b562dbcd15da935fdbc3734773916.jpg) +Figure 3. Compared to the baselines, SplatAD produces sharp images with a high level of detail. Further, the bottom row highlights the superiority of our lidar rendering. Projecting lidar points into images for depth supervision, as used by previous 3DGS methods, causes line-of-sight errors and incorrect volume carving due to the pose differences between camera and lidar. + +We initialize the set of Gaussians using both lidar points and random points. Lidar points within bounding boxes are assigned to the corresponding actor. Further, all lidar points are projected into their closest image to set their initial color. The random points are divided into (1) uniformly sampled points within the lidar range and (2) points sampled linearly in disparity beyond the lidar range, both using random colors at initialization. Furthermore, instead of the 3DGS [14] splitting and densification strategy, we rely on the MCMC strategy introduced in [16]. We chose MCMC partially because we find it to improve far-field rendering quality and partially because it has predictable computation requirements, as it permits choosing the maximum number of Gaussian. We use [16] as is, without any special treatment of Gaussians assigned to dynamic actors. See Appendix B for further training details. + +Implementation: For efficiency, we implement the forward and backward passes of the rolling shutter compensation and lidar projection and rasterization using custom CUDA kernels. SplatAD builds upon the collaborative open-source frameworks gsplat [46] and neurad-studio [34], and we hope that our contributions can benefit future research on multi-modal simulation, not only within the AD community. We train SplatAD for 30,000 iterations using the Adam optimizer, which takes an hour on a single NVIDIA A100. See Appendix B for further details. + +# 4. Experiments + +To validate the robustness of our method, we evaluate it across multiple popular AD datasets, using the same set of hyperparameters. For each dataset, we compare SplatAD to the best performing NeRF and 3DGS-based methods tailored + +Table 1. NVS results for image, over three datasets. First, second, third. + +
PSNR ↑SSIM ↑LPIPS ↓MP/s ↑Train (H) ↓
PandaSetUniSim23.120.6820.36011.50.9
NeuRAD25.800.7530.2509.71.3
PVG24.010.7120.45222.61.6
Street-GS24.730.7450.31463.12.3
OmniRE24.710.7450.31553.62.6
SplatAD (ours)26.700.8140.193104.11.0
Argo2UniSim22.350.6550.45810.01.1
NeuRAD26.180.7210.3109.71.3
PVG24.470.7120.50219.72.5
Street-GS25.520.7540.37492.82.6
OmniRE25.610.7530.37575.63.0
SplatAD (ours)28.400.8260.271133.81.2
nuScenesUniSim22.660.7430.41413.30.7
NeuRAD26.170.7900.31211.81.1
PVG24.940.7680.49718.81.4
Street-GS25.540.7990.37557.91.6
OmniRE25.500.7980.38646.91.9
SplatAD (ours)27.530.8490.302104.31.2
+ +to AD data on novel view synthesis for both camera and lidar data. Further, we ablate key elements of SplatAD and measure the impact they have on sensor realism and rendering speed. + +Datasets: We perform experiments on PandaSet [41], Argoverse2 [38] and nuScenes [5]. These datasets vary in lidars, camera appearance, and resolution. For PandaSet and nuScenes, we use the six available cameras. For Argoverse2, we use the seven ring cameras, leaving out the black-and-white stereo cameras. Further, some cameras are cropped slightly to remove views of the ego-vehicle, such as the hood and the trunk. We use the same 10 challenging sequences per dataset as in [35], containing different types of illumination, dynamic actors, and velocities. We also follow the evaluation protocol of [35] i.e., using full resolution data, + +Table 2. NVS results for lidar, over three datasets. ${}^{8}$ without missing points. First ,second ,third . + +
Depth ↓Intensity ↓Drop acc. ↑CD ↓MR/s ↑
PandaSetUniSim0.080.086-10.3$0.9
NeuRAD0.010.06396.21.91.1
PVG38.74--125.2$0.3
Street-GS6.18--37.3$0.8
OmniRE2.88--29.8$0.6
SplatAD (ours)0.010.05996.71.619.5
Argo2UniSim0.180.081-29.2$0.7
NeuRAD0.020.05892.22.60.9
SplatAD (ours)0.020.05292.62.89.5
nuScenesUniSim0.060.063-52.2$1.0
NeuRAD0.010.04293.16.31.1
PVG10.34--77.9$0.05
Street-GS1.56--4.5$0.2
OmniRE1.71--4.6$0.1
SplatAD (ours)0.020.03693.81.75.7
+ +and, for novel view synthesis (NVS), use every other frame for training and the remaining frames for hold-out validation. We believe that this setting is more challenging than what is often encountered in previous work [6, 7, 43], where fewer datasets are considered, images are downsampled, and a larger fraction of the data is used for training. + +Baselines: We compare our results with the best performing NeRF and 3DGS methods on AD data, namely, UniSim [45], NeuRAD [35], PVG [6], Street Gaussians [43] and OmniRe [7]. Note that all methods are designed for dynamic scenes and use lidar data for supervision. We use NeuRAD's official implementation neurad-studio [34] and their version of UniSim, as well as OmniRe's official implementation drivestudio [8] and their implementations of SVG and Street Gaussians. We use drivestudio to generate the required sky masks for Street Gaussians and OmniRe, as well as extract human pose estimates, which takes about 10 and 30 minutes per sequence on an A100. + +Although point cloud rendering is shown in [7], there exists no feature for it in drivestudio. To render point clouds in training and validation views, we render depth images by placing 6 virtual cameras in the lidar origin, each with a horizontal FOV of $60^{\circ}$ . Their vertical FOV is set to capture all lidar points in the sweep, and the focal length is the median focal length for each dataset. We project ground truth lidar points into these images, fetch each point's depth, and project them back into 3D. As drivestudio lacks functionality for inferring missing lidar points, these baselines are evaluated only for existing points. In contrast, SplatAD is evaluated on the more challenging task of predicting the range also for missing rays and must filter them based on the predicted ray drop probability. See Appendix C for further baseline details. + +# 4.1. Image rendering + +Novel view synthesis: We assess the rendering quality of images using the standard NVS metrics PSNR, SSIM [37], and LPIPS [48] on hold-out validation images in Tab. 1. + +![](images/adcb9d405382fb770cf80cf64294c796be9a07363c0de4e12795b9de42d707f8.jpg) +Figure 4. Removing our rolling shutter modeling compensation leads to inaccurate geometries and inconsistencies in the learning. + +We measure speed using resolution-agnostic megapixels per second. SplatAD consistently outperforms existing 3DGS methods in terms of image quality with a large margin for the three considered datasets. Further, our method improves upon NeuRAD's SOTA results in all NVS metrics, while rendering images an order of magnitude faster. We provide a qualitative comparison in Fig. 3. + +Reconstruction: To measure the upper limit of the modeling capacity of the methods, we report reconstruction metrics on PandaSet in Tab. 3. Here, methods train on all data in a sequence and evaluate on the same views. Further, we enable sensor pose optimization for all methods to account for any pose inaccuracies. Comparing to the NVS setting, all methods except UniSim show improvements with the additional training data. SplatAD achieves SOTA results while rendering $\times 10$ faster than the previous best method. Further, we note that SplatAD's performance on hold-out validation images in Tab. 1 is on par with previous 3DGS methods' results for reconstruction in Tab. 3. + +Extrapolation: We investigate SplatAD's ability to generalize to views that differ significantly from what is encountered during training. We use the models trained on every other frame and follow the three settings in [35]: shift the ego-vehicle horizontally, shift the ego-vehicle vertically, and apply shifts and rotation to all dynamic actors. We report the Fréchet distance using DINOv2 [28] features, as these have been shown [32] to align better with human perception than using Inception-v3 [33]. However, we note that using Inception-v3 features instead does not change the model ranking or our conclusions. Tab. 4 shows SplatAD's ability to learn meaningful representations for generalization, clearly outperforming other 3DGS methods. + +# 4.2. Lidar rendering + +We measure the quality of our lidar point clouds using the same metrics as in [35], i.e., median squared depth error, RMSE intensity error, ray drop accuracy, and chamfer distance, see Tab. 2 for NVS and Tab. 3 for reconstruction. Note + +Table 3. Reconstruction results for image and lidar point clouds on PandaSet. $\S$ without missing points. First, second, third. + +
ImageLidarEfficiency
PSNR ↑SSIM ↑LPIPS ↓Depth ↓Intensity ↓Drop acc. ↑CD ↓Camera MP/s ↑Lidar MR/s ↑Train time (H) ↓
PandasetUniSim23.180.6840.3620.060.087-10.2$11.10.80.8
NeuRAD26.760.7780.2330.030.07496.11.79.81.11.3
PVG25.350.7520.42353.18--150.8$23.60.31.6
Street-GS26.680.8080.2933.27--27.9$65.70.82.2
OmniRE26.680.8080.2953.80--33.5$56.70.62.4
SplatAD (ours)29.740.8930.1750.020.05397.41.6104.117.41.0
+ +Table 4. ${\mathrm{{FD}}}_{\mathrm{{DINOV}}2}$ scores when shifting pose of ego vehicle or actors. First,second,third . + +
Ego lane shiftEgo vert. shiftActor shift
0m2m3m1mRot.Trans.
PandayetNeuRAD377.4540.8629.2571.0424.8420.2
PVG618.8782.7885.6913.9618.8618.8
Street-GS497.1718.6824.2868.7554.7558.0
OmniRE504.5723.2833.5872.8569.6573.2
SplatAD (ours)276.3520.7647.2613.2355.1345.1
+ +Table 5. NVS results averaged over 10 sequences from PandaSet, nuScenes, and Argoverse2 when removing model components. + +
Full modelPSNR ↑SSIM ↑LPIPS ↓Depth ↓CD ↓Camera MP/s ↑Lidar MR/s ↑
27.540.8300.2550.022.0114.111.4
a)Camera RS27.480.8280.2580.022.1116.311.5
b)CNN decoder27.230.8240.2640.022.2178.011.6
c)EWA antialias27.600.8280.2680.021.5114.511.0
d)Appearance emb.27.280.8240.2640.022.1121.011.5
e)Velocity opt.27.550.8300.2570.022.0114.111.4
f)Lidar RS27.420.8270.2600.032.1114.212.2
g)MCMC27.350.8260.2590.012.1130.614.1
h)Median depth27.540.8300.2550.024.9114.111.4
+ +![](images/c448484c33faf286ff729a8f41f4439392ed19dd276f7245c1e5eb3ce057bbfd.jpg) +Figure 5. The CNN decoder improves sharpness and is more true to color than the MLP decoder. + +![](images/a30645caa572704cc825a30a068dd3d158e63ef9efbdf29017da8cae571f8b09.jpg) + +that we do not report intensity or ray drop accuracy for the 3DGS baselines, as they have no way of inferring these. Further, it was not possible to produce Argoverse2 point cloud results for these baselines due to the memory requirements of drivestudio not scaling well with the increase in number of images and higher image resolution. We find SplatAD to be on par with the accurate ray-tracing formulation of NeuRAD in terms of point cloud quality, while running up to $\times 18$ faster. Furthermore, our lidar rendering approach is superior to the naive depth image-based method used for 3DGS baselines, both in terms of speed and quality. + +# 4.3. Ablations + +We validate the effectiveness of key components of our method by measuring their impact on NVS metrics in Tab. 5. Metrics are averaged over all three datasets to avoid any + +dataset-specific biases. We see that removing the rolling shutter modeling of the camera (a) or lidar (f) reduces NVS performance, and that leaving out optimization of sensor velocities (e) negatively impacts our depth predictions. While the impact of rolling shutter modeling on quantitative metrics is somewhat modest, we qualitatively find that it is important for capturing thin structures and edges, see Fig. 4. Removing these components leads to line-of-sight errors, which erroneously can cut through objects. Note that when removing our lidar rolling shutter compensation, we use ego-motion compensated point clouds for supervision. Moreover, we note the efficiency of our rolling shutter algorithms, as removing them barely improves rendering speeds. + +We also find that using a CNN instead of an MLP for image decoding (c) clearly improves image realism, but at the cost of runtime, see Fig. 5 for an example. Further, we verify that using median depth instead of expected depth during inference is important for realistic lidar data in terms of Chamfer distance (h). Last, we note that MCMC [16] and EWA antialIASing [47, 52] both improve our performance, with the antialIASing having the largest impact on perceptual quality in terms of LPIPS. + +# 5. Conclusion + +In this work, we have proposed SplatAD, the first method for rendering camera and lidar data from 3D Gaussians. We combined accurate sensor modeling with efficient algorithms to achieve state-of-the-art novel view synthesis results while improving run-time by more than an order of magnitude. We hope that SplatAD will advance the scalability and fidelity of simulation environments, facilitating safer and more cost-effective testing of AD systems. + +Limitations and future work: SplatAD is currently limited to modeling all dynamic actors as rigid. Drawing inspiration from recent advances in human reconstruction [18, 20, 26] can provide inspiration how to overcome this limitation in future research. In addition, we believe that incorporating data-driven priors such as diffusion models [9, 11, 40] is a promising direction to improve the extrapolation performance, and for changing appearance attributes. + +# Acknowledgements + +We thank Adam Lilja, William Ljungbergh, and Adam Tenderski for valuable feedback. This work was partially supported by the Wallenberg AI, Autonomous Systems and Software Program (WASP) funded by the Knut and Alice Wallenberg Foundation. Computational resources were provided by NAISS at NSC Berzelius, partially funded by the Swedish Research Council, grant agreement no. 2022-06725. + +# References + +[1] Mina Alibeigi, William Ljungbergh, Adam Tonderski, Georg Hess, Adam Lilja, Carl Lindstrom, Daria Motorniuk, Junsheng Fu, Jenny Widahl, and Christoffer Petersson. Zenseact open dataset: A large-scale and diverse multimodal dataset for autonomous driving. In ICCV, pages 20178-20188, 2023. 1 +[2] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In ICCV, pages 5855-5864, 2021. 2 +[3] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In CVPR, pages 5470–5479, 2022. +[4] Jonathan T. Barron, Ben Mildenhall, Dor Verbin, Pratul P. Srinivasan, and Peter Hedman. Zip-nerf: Anti-aliased grid-based neural radiance fields. In ICCV, pages 19697-19705, 2023. 2 +[5] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In CVPR, pages 11621-11631, 2020. 1, 6 +[6] Yurui Chen, Chun Gu, Junzhe Jiang, Xiatian Zhu, and Li Zhang. Periodic vibration gaussian: Dynamic urban scene reconstruction and real-time rendering. arXiv preprint arXiv:2311.18561, 2023. 2, 7 +[7] Ziyu Chen, Jiawei Yang, Jiahui Huang, Riccardo de Lutio, Janick Martinez Esturo, Boris Ivanovic, Or Litany, Zan Gojcic, Sanja Fidler, Marco Pavone, et al. Omnire: Omni urban scene reconstruction. arXiv preprint arXiv:2408.16760, 2024. 1, 2, 5, 7 +[8] Ziyu Chen, Jiawei Yang, Jiahui Huang, Riccardo de Lutio, Janick Martinez Esturo, Boris Ivanovic, Or Litany, Zan Gojcic, Sanja Fidler, Marco Pavone, Li Song, and Yue Wang. *drivestudio*. https://github.com/ziyc/drivestudio, 2024. 7, 2 +[9] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In ICML, 2024. 8 +[10] Xiao Fu, Shangzhan Zhang, Tianrun Chen, Yichong Lu, Lanyun Zhu, Xiaowei Zhou, Andreas Geiger, and Yiyi Liao. Panoptic nerf: 3d-to-2d label transfer for panoptic urban scene segmentation. In 3DV, pages 1-11. IEEE, 2022. 2 + +[11] Ruiqi Gao, Aleksander Holynski, Philipp Henzler, Arthur Brussee, Ricardo Martin Brualla, Pratul P. Srinivasan, Jonathan T. Barron, and Ben Poole. CAT3d: Create anything in 3d with multi-view diffusion models. In NeurIPS, 2024. 8 +[12] Shubham Goel, Georgios Pavlakos, Jathushan Rajasegaran, Angjoo Kanazawa, and Jitendra Malik. Humans in 4d: Reconstructing and tracking humans with transformers. In ICCV, pages 14783-14794, 2023. 2 +[13] Shengyu Huang, Zan Gojcic, Zian Wang, Francis Williams, Yoni Kasten, Sanja Fidler, Konrad Schindler, and Or Litany. Neural lidar fields for novel view synthesis. In ICCV, pages 18236-18246, 2023. 2 +[14] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. TOG, 42(4):139-1, 2023. 1, 2, 3, 4, 6 +[15] Mustafa Khan, Hamidreza Fazlali, Dhruv Sharma, Tongtong Cao, Dongfeng Bai, Yuan Ren, and Bingbing Liu. Autosplat: Constrained gaussian splatting for autonomous driving scene reconstruction, 2024. 2 +[16] Shakiba Kheradmand, Daniel Rebain, Gopal Sharma, Weiwei Sun, Yang-Che Tseng, Hossam Isack, Abhishek Kar, Andrea Tagliasacchi, and Kwang Moo Yi. 3d gaussian splatting as markov chain monte carlo. In NeurIPS, 2024. 5, 6, 8, 2 +[17] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR, 2015. 2 +[18] Muhammed Kocabas, Jen-Hao Rick Chang, James Gabriel, Oncel Tuzel, and Anurag Ranjan. Hugs: Human gaussian splats. In CVPR, pages 505-515, 2024. 8 +[19] Abhijit Kundu, Kyle Genova, Xiaqi Yin, Alireza Fathi, Caroline Pantofaru, Leonidas J Guibas, Andrea Tagliasacchi, Frank Dellaert, and Thomas Funkhouser. Panoptic neural fields: A semantic object-aware neural scene representation. In CVPR, pages 12871-12881, 2022. 2 +[20] Youngjoong Kwon, Baole Fang, Yixing Lu, Haoye Dong, Cheng Zhang, Francisco Vicente Carrasco, Albert Mosella-Montoro, Jianjin Xu, Shingo Takagi, Daeil Kim, Aayush Prakash, and Fernando De la Torre. Generalizable human gaussians for sparse view synthesis. In ECCV, pages 451-468, Cham, 2025. Springer Nature Switzerland. 8 +[21] Nanxi Li, Chong Pei Ho, Jin Xue, Leh Woon Lim, Guanyu Chen, Yuan Hsing Fu, and Lennon Yao Ting Lee. A progress review on solid-state lidar and nanophotonics-based lidar sensors. *Laser & Photonics Reviews*, 16(11):2100511, 2022. 4 +[22] Carl Lindström, Georg Hess, Adam Lilja, Maryam Fatemi, Lars Hammarstrand, Christoffer Petersson, and Lennart Svensson. Are nerfs ready for autonomous driving? towards closing the real-to-simulation gap. In CVPRW, pages 4461-4471, 2024. 2 +[23] William Ljungbergh, Adam Tonderski, Joakim Johnander, Holger Caesar, Kalle Åström, Michael Felsberg, and Christoffer Petersson. Neuroncap: Photorealistic closed-loop safety testing for autonomous driving. In ECCV, pages 161-177. Springer, 2025. 2 +[24] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J. Black. SMPL: A skinned multiperson linear model. TOG, 34(6):248:1-248:16, 2015. 2 + +[25] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, pages 405-421, Cham, 2020. Springer International Publishing. 1, 2 +[26] Arthur Moreau, Jifei Song, Helisa Dhamo, Richard Shaw, Yiren Zhou, and Eduardo Pérez-Pellitero. Human gaussian splatting: Real-time rendering of animatable avatars. In CVPR, pages 788-798, 2024. 8 +[27] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. TOG, 41(4):1-15, 2022. 2 +[28] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel HAZIZA, Francisco Massa, Alaaeldin El-Nouby, Mido Assran, Nicolas Ballas, Wojciech Galuba, Russell Howes, Po-Yao Huang, Shang-Wen Li, Ishan Misra, Michael Rabbat, Vasu Sharma, Gabriel Synnaeve, Hu Xu, Herve Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. DINov2: Learning robust visual features without supervision. TMLR, 2024. 7 +[29] Julian Ost, Fahim Mannan, Nils Thuerey, Julian Knodt, and Felix Heide. Neural scene graphs for dynamic scenes. In CVPR, pages 2856-2865, 2021. 2, 3 +[30] Konstantinos Rematas, Andrew Liu, Pratul P Srinivasan, Jonathan T Barron, Andrea Tagliasacchi, Thomas Funkhouser, and Vittorio Ferrari. Urban radiance fields. In CVPR, pages 12932-12942, 2022. 4 +[31] Otto Seiskari, Jerry Yilammi, Valtteri Kaatrasalo, Pekka Rantalankila, Matias Turkulainen, Juho Kannala, Esa Rahtu, and Arno Solin. Gaussian splattering on the move: Blur and rolling shutter compensation for natural camera motion. In ECCV, pages 160-177, Cham, 2025. Springer Nature Switzerland. 4, 3 +[32] George Stein, Jesse Cresswell, Rasa Hosseinzadeh, Yi Sui, Brendan Ross, Valentin Villecloze, Zhaoyan Liu, Anthony L Caterini, Eric Taylor, and Gabriel Loaiza-Ganem. Exposing flaws of generative model evaluation metrics and their unfair treatment of diffusion models. NeurIPS, 36, 2024. 7 +[33] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. In CVPR, pages 2818-2826, 2016. 7 +[34] Adam Tondersski, Carl Lindstrom, Georg Hess, and William Ljungbergh. neurad-studio. https://github.com/georghess/neurad-studio, 2024. 6, 7 +[35] Adam Tondersski, Carl Lindström, Georg Hess, William Ljungbergh, Lennart Svensson, and Christoffer Petersson. Neurad: Neural rendering for autonomous driving. In CVPR, pages 14895-14904, 2024. 1, 2, 3, 4, 6, 7 +[36] Haithem Turki, Jason Y Zhang, Francesco Ferroni, and Deva Ramanan. Suds: Scalable urban dynamic scenes. In CVPR, pages 12375-12385, 2023. 1, 2 +[37] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE TIP, 13(4):600-612, 2004. 7 + +[38] Benjamin Wilson, William Qi, Tanmay Agarwal, John Lambert, Jagjeet Singh, Siddhesh Khandelwal, Bowen Pan, Ratnesh Kumar, Andrew Hartnett, Jhony Kaesemodel Pontes, Deva Ramanan, Peter Carr, and James Hays. Argoverse 2: Next generation datasets for self-driving perception and forecasting. In Proceedings of the Neural Information Processing Systems Track on Datasets and Benchmarks (NeurIPS Datasets and Benchmarks 2021), 2021. 1, 6 +[39] Hanfeng Wu, Xingxing Zuo, Stefan Leutenegger, Or Litany, Konrad Schindler, and Shengyu Huang. Dynamic lidar resimulation using compositional neural fields. In CVPR, pages 1998-19998, 2024. 2 +[40] Rundi Wu, Ben Mildenhall, Philipp Henzler, Keunhong Park, Ruiqi Gao, Daniel Watson, Pratul P Srinivasan, Dor Verbin, Jonathan T Barron, Ben Poole, et al. Reconfusion: 3d reconstruction with diffusion priors. In CVPR, pages 21551-21561, 2024. 8 +[41] Pengchuan Xiao, Zhenlei Shao, Steven Hao, Zishuo Zhang, Xiaolin Chai, Judy Jiao, Zesong Li, Jian Wu, Kai Sun, Kun Jiang, Yunlong Wang, and Diange Yang. Pandaset: Advanced sensor suite dataset for autonomous driving. In 2021 IEEE International Intelligent Transportation Systems Conference (ITSC), pages 3095-3101, 2021. 2, 6 +[42] Ziyang Xie, Junge Zhang, Wenye Li, Feihu Zhang, and Li Zhang. S-neRF: Neural radiance fields for street views. In ICLR, 2023. 2 +[43] Yunzhi Yan, Haotong Lin, Chenxu Zhou, Weijie Wang, Haiyang Sun, Kun Zhan, Xianpeng Lang, Xiaowei Zhou, and Sida Peng. Street gaussians for modeling dynamic urban scenes. In ECCV, 2024. 1, 2, 3, 7 +[44] Jiawei Yang, Boris Ivanovic, Or Litany, Xinshuo Weng, Seung Wook Kim, Boyi Li, Tong Che, Danfei Xu, Sanja Fidler, Marco Pavone, and Yue Wang. EmerneRF: Emergent spatial-temporal scene decomposition via self-supervision. In ICLR, 2024. 2 +[45] Ze Yang, Yun Chen, Jingkang Wang, Sivabalan Manivasagam, Wei-Chiu Ma, Anqi Joyce Yang, and Raquel Urtasun. Unisim: A neural closed-loop sensor simulator. In CVPR, pages 1389-1399, 2023. 1, 2, 7, 3 +[46] Vickie Ye, Ruilong Li, Justin Kerr, Matias Turkulainen, Brent Yi, Zhuoyang Pan, Otto Seiskari, Jianbo Ye, Jeffrey Hu, Matthew Tancik, et al. gsplat: An open-source library for gaussian splatting. arXiv preprint arXiv:2409.06765, 2024. 6 +[47] Zehao Yu, Anpei Chen, Binbin Huang, Torsten Sattler, and Andreas Geiger. Mip-splatting: Alias-free 3d gaussian splatting. In CVPR, pages 19447-19456, 2024. 4, 5, 8 +[48] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, pages 586-595, 2018. 7 +[49] Zehan Zheng, Fan Lu, Weiyi Xue, Guang Chen, and Changjun Jiang. Lidar4d: Dynamic neural fields for novel space-time view lidar synthesis. In CVPR, pages 5145-5154, 2024. 2 +[50] Hongyu Zhou, Jiahao Shao, Lu Xu, Dongfeng Bai, Weichao Qiu, Bingbing Liu, Yue Wang, Andreas Geiger, and Yiyi Liao. Hugs: Holistic urban 3d scene understanding via gaussian splatting. In CVPR, pages 21336-21345, 2024. 2 + +[51] Xiaoyu Zhou, Zhiwei Lin, Xiaojun Shan, Yongtao Wang, Deqing Sun, and Ming-Hsuan Yang. Drivinggaussian: Composite gaussian splatting for surrounding dynamic autonomous driving scenes. In CVPR, pages 21634-21643, 2024. 1, 2, 3 +[52] Matthias Zwicker, Hanspeter Pfister, Jeroen Van Baar, and Markus Gross. Ewa volume splatting. In Proceedings Visualization, 2001. VIS'01., pages 29-538. IEEE, 2001. 4, 8 \ No newline at end of file diff --git a/2025/SplatAD_ Real-Time Lidar and Camera Rendering with 3D Gaussian Splatting for Autonomous Driving/images.zip b/2025/SplatAD_ Real-Time Lidar and Camera Rendering with 3D Gaussian Splatting for Autonomous Driving/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..c63b24188f1475d8c7e66390958bb36173e0b787 --- /dev/null +++ b/2025/SplatAD_ Real-Time Lidar and Camera Rendering with 3D Gaussian Splatting for Autonomous Driving/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e38038adc64738a0ce49ff72ac65c5a1d89ed9f8865d0032fcb928af344e33d9 +size 624435 diff --git a/2025/SplatAD_ Real-Time Lidar and Camera Rendering with 3D Gaussian Splatting for Autonomous Driving/layout.json b/2025/SplatAD_ Real-Time Lidar and Camera Rendering with 3D Gaussian Splatting for Autonomous Driving/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..5fa092004e9c5713438ea630c31a2bcd4d94ed53 --- /dev/null +++ b/2025/SplatAD_ Real-Time Lidar and Camera Rendering with 3D Gaussian Splatting for Autonomous Driving/layout.json @@ -0,0 +1,8570 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 56, + 102, + 556, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 102, + 556, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 102, + 556, + 140 + ], + "type": "text", + "content": "SplatAD: Real-Time Lidar and Camera Rendering with 3D Gaussian Splatting for Autonomous Driving" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 159, + 574, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 159, + 574, + 175 + ], + "spans": [ + { + "bbox": [ + 61, + 159, + 574, + 175 + ], + "type": "text", + "content": "Georg Hess" + }, + { + "bbox": [ + 61, + 159, + 574, + 175 + ], + "type": "inline_equation", + "content": "^{\\dagger,1,2}" + }, + { + "bbox": [ + 61, + 159, + 574, + 175 + ], + "type": "text", + "content": " Carl Lindström" + }, + { + "bbox": [ + 61, + 159, + 574, + 175 + ], + "type": "inline_equation", + "content": "^{\\dagger,1,2}" + }, + { + "bbox": [ + 61, + 159, + 574, + 175 + ], + "type": "text", + "content": " Maryam Fatemi" + }, + { + "bbox": [ + 61, + 159, + 574, + 175 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 61, + 159, + 574, + 175 + ], + "type": "text", + "content": " Christoffer Petersson" + }, + { + "bbox": [ + 61, + 159, + 574, + 175 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 61, + 159, + 574, + 175 + ], + "type": "text", + "content": " Lennart Svensson" + }, + { + "bbox": [ + 61, + 159, + 574, + 175 + ], + "type": "inline_equation", + "content": "^{2}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 208, + 176, + 425, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 208, + 176, + 425, + 190 + ], + "spans": [ + { + "bbox": [ + 208, + 176, + 425, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 208, + 176, + 425, + 190 + ], + "type": "text", + "content": "Zenseact " + }, + { + "bbox": [ + 208, + 176, + 425, + 190 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 208, + 176, + 425, + 190 + ], + "type": "text", + "content": "Chalmers University of Technology" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 188, + 191, + 444, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 188, + 191, + 444, + 203 + ], + "spans": [ + { + "bbox": [ + 188, + 191, + 444, + 203 + ], + "type": "text", + "content": "{firstname_lastname}@{zenseact.com, chalmers.se}" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 57, + 232, + 181, + 314 + ], + "blocks": [ + { + "bbox": [ + 57, + 232, + 181, + 314 + ], + "lines": [ + { + "bbox": [ + 57, + 232, + 181, + 314 + ], + "spans": [ + { + "bbox": [ + 57, + 232, + 181, + 314 + ], + "type": "image", + "image_path": "19714502aa07182b3219d434c13f7483494374467b078db7a0a7adf3d8faa014.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 324, + 555, + 357 + ], + "lines": [ + { + "bbox": [ + 54, + 324, + 555, + 357 + ], + "spans": [ + { + "bbox": [ + 54, + 324, + 555, + 357 + ], + "type": "text", + "content": "Figure 1. SplatAD is the first method capable of realistic camera and lidar rendering using 3D Gaussian Splatting. Whereas previous methods are either fast or multi-modal, SplatAD enables real-time, high-quality rendering for both camera and lidar. In addition, SplatAD can reach competitive performance, e.g., PSNR for image and Chamfer distance for point cloud, within minutes." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 182, + 232, + 304, + 314 + ], + "blocks": [ + { + "bbox": [ + 182, + 232, + 304, + 314 + ], + "lines": [ + { + "bbox": [ + 182, + 232, + 304, + 314 + ], + "spans": [ + { + "bbox": [ + 182, + 232, + 304, + 314 + ], + "type": "image", + "image_path": "391b65fc110a83970041c0db396b49d439b7bd8ff76cf9715a9ed4edfb8efdcd.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 306, + 232, + 429, + 314 + ], + "blocks": [ + { + "bbox": [ + 306, + 232, + 429, + 314 + ], + "lines": [ + { + "bbox": [ + 306, + 232, + 429, + 314 + ], + "spans": [ + { + "bbox": [ + 306, + 232, + 429, + 314 + ], + "type": "image", + "image_path": "8fb6c77975566fdd0f4afdeb0be7cbcb33e13040dcf3f5679d9b54e1034459ea.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 430, + 232, + 553, + 314 + ], + "blocks": [ + { + "bbox": [ + 430, + 232, + 553, + 314 + ], + "lines": [ + { + "bbox": [ + 430, + 232, + 553, + 314 + ], + "spans": [ + { + "bbox": [ + 430, + 232, + 553, + 314 + ], + "type": "image", + "image_path": "e3b1848f205a306c38b098b596956599caf495e37462f7d9e6386292040a36da.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 151, + 368, + 200, + 380 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 368, + 200, + 380 + ], + "spans": [ + { + "bbox": [ + 151, + 368, + 200, + 380 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 53, + 395, + 297, + 682 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 395, + 297, + 682 + ], + "spans": [ + { + "bbox": [ + 53, + 395, + 297, + 682 + ], + "type": "text", + "content": "Ensuring the safety of autonomous robots, such as self-driving vehicles, requires extensive testing across diverse driving scenarios. Simulation is a key ingredient for conducting such testing in a cost-effective and scalable way. Neural rendering methods have gained popularity, as they can build simulation environments from collected logs in a data-driven manner. However, existing neural radiance field (NeRF) methods for sensor-realistic rendering of camera and lidar data suffer from low rendering speeds, limiting their applicability for large-scale testing. While 3D Gaussian Splatting (3DGS) enables real-time rendering, current methods are limited to camera data and are unable to render lidar data essential for autonomous driving. To address these limitations, we propose SplatAD, the first 3DGS-based method for realistic, real-time rendering of dynamic scenes for both camera and lidar data. SplatAD accurately models key sensor-specific phenomena such as rolling shutter effects, lidar intensity, and lidar ray dropouts, using purpose-built algorithms to optimize rendering efficiency. Evaluation across three autonomous driving datasets demonstrates that SplatAD achieves state-of-the-art rendering quality with up to " + }, + { + "bbox": [ + 53, + 395, + 297, + 682 + ], + "type": "inline_equation", + "content": "+2" + }, + { + "bbox": [ + 53, + 395, + 297, + 682 + ], + "type": "text", + "content": " PSNR for NVS and " + }, + { + "bbox": [ + 53, + 395, + 297, + 682 + ], + "type": "inline_equation", + "content": "+3" + }, + { + "bbox": [ + 53, + 395, + 297, + 682 + ], + "type": "text", + "content": " PSNR for reconstruction while increasing rendering speed over NeRF-based methods by an order of magnitude. See here for our project page." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 315, + 368, + 394, + 380 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 368, + 394, + 380 + ], + "spans": [ + { + "bbox": [ + 315, + 368, + 394, + 380 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 312, + 389, + 556, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 389, + 556, + 533 + ], + "spans": [ + { + "bbox": [ + 312, + 389, + 556, + 533 + ], + "type": "text", + "content": "Large-scale testing is essential to ensure the safety of autonomous robots, such as self-driving vehicles (SDVs), before real-world deployment. Data-driven methods that generate digital twins from collected logs offer a scalable way to build diverse, realistic simulation environments for testing. Unlike real-world testing, which can be costly, time-intensive, and limited by physical constraints, simulation enables rapid, low-cost exploration of multiple scenarios, helping to optimize SDVs for safety, comfort, and efficiency. Motivated by this, a multitude of methods have emerged based on neural radiance fields (NeRFs) [25, 35, 36, 45] and 3D Gaussian Splatting (3DGS) [14, 43, 51]." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 312, + 534, + 557, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 534, + 557, + 713 + ], + "spans": [ + { + "bbox": [ + 312, + 534, + 557, + 713 + ], + "type": "text", + "content": "Recent NeRF-based methods [35, 45] provide high-fidelity sensor simulation jointly for both camera and lidar, matching the most common sensor rigs in popular autonomous driving (AD) datasets [1, 5, 38]. However, the slow rendering speed of NeRF-based methods makes them costly and challenging to use for large-scale testing. 3DGS provides an attractive alternative to NeRFs, as its accelerated rendering achieves comparable image realism while increasing inference speeds by an order of magnitude. Nevertheless, 3DGS-based methods for the AD setting [7, 43, 51] inherit the limitation of only being able to render camera data, overlooking the lidar modality. The lidar's ability to directly perceive the 3D environment makes it a powerful tool in modern AD stacks and, consequently, an important modality to simulate." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 708, + 223, + 719 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 708, + 223, + 719 + ], + "spans": [ + { + "bbox": [ + 66, + 708, + 223, + 719 + ], + "type": "text", + "content": "These authors contributed equally to this work." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 293, + 750, + 318, + 760 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 750, + 318, + 760 + ], + "spans": [ + { + "bbox": [ + 293, + 750, + 318, + 760 + ], + "type": "text", + "content": "11982" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 228 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 228 + ], + "type": "text", + "content": "In this paper, we aim to solve efficient, differentiable, and realistic rendering of camera and lidar data using 3D Gaussian Splatting. Applying 3DGS to lidar sensors presents unique challenges due to their distinct characteristics. First, unlike cameras, which capture dense, regularly spaced pixels, lidars record sparse, non-linear point clouds with large empty gaps from non-returning rays. Second, most lidars capture a full " + }, + { + "bbox": [ + 55, + 72, + 294, + 228 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 55, + 72, + 294, + 228 + ], + "type": "text", + "content": " view of the scene, where existing methods often project data into multiple depth images [7]—an inefficient approach that overlooks the sparse structure of the lidar. Last, lidars suffer from a rolling shutter effect, with each sweep taking up to " + }, + { + "bbox": [ + 55, + 72, + 294, + 228 + ], + "type": "inline_equation", + "content": "100\\mathrm{ms}" + }, + { + "bbox": [ + 55, + 72, + 294, + 228 + ], + "type": "text", + "content": ", during which the ego vehicle may move several meters, violating 3DGS's single-origin assumption." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 228, + 295, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 228, + 295, + 407 + ], + "spans": [ + { + "bbox": [ + 55, + 228, + 295, + 407 + ], + "type": "text", + "content": "To overcome these challenges, we introduce SplatAD, a novel view synthesis method that unifies camera and lidar rendering and is designed for real-time rendering of large-scale dynamic traffic scenes. Our method replaces spherical harmonics with per-Gaussian learnable features to jointly model sensor-specific phenomena—from lidar ray drop and intensity variations to camera-specific appearance changes. By introducing purpose-built rasterization algorithms for rendering lidar in spherical coordinates, we achieve both superior efficiency and realism. Further, we demonstrate efficient methods for modeling rolling shutter effects across both sensor modalities. We verify the effectiveness and generalizability of our method across three popular automotive datasets, achieving state-of-the-art results on all of them. To summarize, our contributions are as follows:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 407, + 295, + 598 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 56, + 407, + 295, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 407, + 295, + 453 + ], + "spans": [ + { + "bbox": [ + 56, + 407, + 295, + 453 + ], + "type": "text", + "content": "- We propose the first method for efficient lidar rendering using 3D Gaussians, introducing custom CUDA-accelerated algorithms for rasterizing sparse point clouds in spherical coordinates." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 455, + 295, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 455, + 295, + 502 + ], + "spans": [ + { + "bbox": [ + 56, + 455, + 295, + 502 + ], + "type": "text", + "content": "- We introduce the first 3DGS method capable of rendering both camera and lidar from a unified representation, enabling accelerated scaling of novel view synthesis for automotive applications." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 503, + 295, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 503, + 295, + 550 + ], + "spans": [ + { + "bbox": [ + 56, + 503, + 295, + 550 + ], + "type": "text", + "content": "- We present effective techniques for realistic sensor modeling using 3D Gaussians, enabling accurate handling of rolling shutter, lidar intensity, ray dropping, and variations in sensor appearance." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 551, + 295, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 551, + 295, + 598 + ], + "spans": [ + { + "bbox": [ + 56, + 551, + 295, + 598 + ], + "type": "text", + "content": "- Through extensive evaluation on three popular automotive datasets, we demonstrate state-of-the-art results across all benchmarks, validating our method's effectiveness and generalizability." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 56, + 609, + 139, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 609, + 139, + 621 + ], + "spans": [ + { + "bbox": [ + 56, + 609, + 139, + 621 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 629, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 295, + 713 + ], + "type": "text", + "content": "NeRFs for automotive data: Since the introduction of NeRFs, neural representations have become central to 3D reconstruction and novel view synthesis [2-4, 25]. Many works have applied NeRF-based methods to automotive data [35, 36, 44, 45], enabling sensor-realistic renderings of novel views in large and dynamic scenes. Recent advances provide realism sufficient for downstream applica" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 555, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 262 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 262 + ], + "type": "text", + "content": "tions [22, 23]. While early methods focused on either camera [10, 19, 29, 42] or lidar [13, 39, 49], newer methods aim to handle them jointly. UniSim [45] showcased realistic renderings for front camera and " + }, + { + "bbox": [ + 313, + 72, + 555, + 262 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 313, + 72, + 555, + 262 + ], + "type": "text", + "content": " lidar on data from PandaSet [41]. The method uses a hash grid representation [27], with separate features for the sky, static background, and dynamic actors, and decodes the color and lidar intensity from volume-rendered features. NeuRAD [35] proposes a simplified network architecture and improved sensor modeling, achieving state-of-the-art results for the full " + }, + { + "bbox": [ + 313, + 72, + 555, + 262 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 313, + 72, + 555, + 262 + ], + "type": "text", + "content": " camera and lidar rig. Even so, the slow rendering speeds of NeRF-based methods make scaling costly and challenging. SplatAD aims to overcome this limitation by using rasterization-based CUDA-accelerated algorithms to render both camera and lidar data while, inspired by NeuRAD, emphasizing the modeling of important sensor characteristics for increased realism." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 263, + 556, + 527 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 263, + 556, + 527 + ], + "spans": [ + { + "bbox": [ + 313, + 263, + 556, + 527 + ], + "type": "text", + "content": "3DGS for automotive data: 3DGS [14] uses an explicit representation paired with rasterization techniques and custom hardware-accelerated algorithms to achieve real-time rendering. Additionally, 3DGS has been shown to create faithful scene reconstructions with rendering quality close to recent NeRF-based methods. As a result, several works have applied 3DGS for camera rendering on automotive data [7, 15, 43, 50, 51]. Periodic Vibration Gaussian (PVG) [6] applies 3DGS to dynamic scenes by learning the 3D flow of all Gaussians. Yet, PVG's lack of an explicit actor representation limits the method's controllability and, hence, applicability for simulation. Street Gaussians [43], like Unisim and NeuRAD, instead decomposes the scene into static background and rigid dynamic actors using 3D bounding boxes. The authors further extend 3DGS by adding Fourier coefficients to capture time-varying characteristics of dynamic actors. Moreover, to handle the sky regions, their method uses semantic masks and a view-dependent cube map. OmniRe [7] enhances the scene graph composition with non-rigid nodes for better modeling of pedestrians and cyclists. However, this requires pre-processing to track and align human body poses [12] for initializing SMPL [24] pose parameters." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 527, + 556, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 527, + 556, + 658 + ], + "spans": [ + { + "bbox": [ + 313, + 527, + 556, + 658 + ], + "type": "text", + "content": "PVG, Street Gaussians, and OmniRe all use lidar points for initialization and depth supervision. However, to generate new point clouds, these methods predict depth images and rely on existing lidar points to query said image at the right locations. This is both inefficient and inaccurate, as " + }, + { + "bbox": [ + 313, + 527, + 556, + 658 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 313, + 527, + 556, + 658 + ], + "type": "text", + "content": " lidars are approximated as multiple cameras. Further, neither of these methods model important lidar characteristics such as intensity variations, ray drop, and rolling shutter. In contrast, our method supports efficient rendering of both camera and lidar from 3D Gaussians in novel views, and models important sensor characteristics for both modalities." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 669, + 370, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 669, + 370, + 681 + ], + "spans": [ + { + "bbox": [ + 313, + 669, + 370, + 681 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "content": "Our aim is to learn a scene representation from collected vehicle logs that enables rendering of realistic camera and lidar" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 750, + 317, + 759 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 750, + 317, + 759 + ], + "spans": [ + { + "bbox": [ + 293, + 750, + 317, + 759 + ], + "type": "text", + "content": "11983" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 71, + 558, + 259 + ], + "blocks": [ + { + "bbox": [ + 59, + 71, + 558, + 259 + ], + "lines": [ + { + "bbox": [ + 59, + 71, + 558, + 259 + ], + "spans": [ + { + "bbox": [ + 59, + 71, + 558, + 259 + ], + "type": "image", + "image_path": "3536cb69060a185c55a3bdfeba2bb4a0a85014e2298f4ddbd805a9979bdf1cc2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 266, + 555, + 336 + ], + "lines": [ + { + "bbox": [ + 54, + 266, + 555, + 336 + ], + "spans": [ + { + "bbox": [ + 54, + 266, + 555, + 336 + ], + "type": "text", + "content": "Figure 2. Overview of our proposed method. Given the composition of static and dynamic 3D Gaussians, SplatAD is capable of differentiable rendering of both lidar and camera data. Our proposed lidar rendering matches the image rendering on a high level, but modifies each component to accurately model sensor characteristics. Our method projects 3D Gaussians with associated feature vectors onto the corresponding sensor modalities (camera and lidar) and employs sensor-specific tiling to match their distinct characteristics. During rasterization, the projected Gaussians are corrected for rolling shutter effects caused by the movement of sensors and, potentially, their own velocity. Finally, the rasterized features are decoded into the respective image and lidar point cloud representations." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 354, + 294, + 438 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 354, + 294, + 438 + ], + "spans": [ + { + "bbox": [ + 54, + 354, + 294, + 438 + ], + "type": "text", + "content": "data, with the ability to alter the locations of both the ego vehicle and other actors. To scale effectively, the rendering process must be fast, as faster inference speeds enhance its practicality for applications. In the following, we describe our scene representation (Sec. 3.1), rendering algorithms (Secs. 3.2 and 3.3), and implementation and optimization strategy (Sec. 3.4), see Fig. 2 for an overview." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 445, + 176, + 456 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 445, + 176, + 456 + ], + "spans": [ + { + "bbox": [ + 55, + 445, + 176, + 456 + ], + "type": "text", + "content": "3.1. Scene representation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 462, + 295, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 462, + 295, + 617 + ], + "spans": [ + { + "bbox": [ + 54, + 462, + 295, + 617 + ], + "type": "text", + "content": "Our scene representation builds upon 3DGS [14], but with key changes to handle the specifics of AD scenes and to enable both camera and lidar rendering from the same representation. Like 3DGS, each scene is represented by a set of translucent 3D Gaussians with learnable occupancy " + }, + { + "bbox": [ + 54, + 462, + 295, + 617 + ], + "type": "inline_equation", + "content": "o \\in (0,1)" + }, + { + "bbox": [ + 54, + 462, + 295, + 617 + ], + "type": "text", + "content": ", mean " + }, + { + "bbox": [ + 54, + 462, + 295, + 617 + ], + "type": "inline_equation", + "content": "\\mu \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 54, + 462, + 295, + 617 + ], + "type": "text", + "content": ", and covariance matrix " + }, + { + "bbox": [ + 54, + 462, + 295, + 617 + ], + "type": "inline_equation", + "content": "\\Sigma \\in \\mathbb{R}^{3\\times 3}" + }, + { + "bbox": [ + 54, + 462, + 295, + 617 + ], + "type": "text", + "content": " which in turn is parameterized by scale " + }, + { + "bbox": [ + 54, + 462, + 295, + 617 + ], + "type": "inline_equation", + "content": "\\mathbf{S} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 54, + 462, + 295, + 617 + ], + "type": "text", + "content": " and quaternion " + }, + { + "bbox": [ + 54, + 462, + 295, + 617 + ], + "type": "inline_equation", + "content": "\\mathbf{q} \\in \\mathbb{R}^4" + }, + { + "bbox": [ + 54, + 462, + 295, + 617 + ], + "type": "text", + "content": " parameters. Instead of spherical harmonics [14], we assign each Gaussian a learnable base color " + }, + { + "bbox": [ + 54, + 462, + 295, + 617 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{\\mathrm{rgb}}^{\\mathrm{b}} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 54, + 462, + 295, + 617 + ], + "type": "text", + "content": " and feature vector " + }, + { + "bbox": [ + 54, + 462, + 295, + 617 + ], + "type": "inline_equation", + "content": "\\mathbf{f} \\in \\mathbb{R}^{D_{\\mathrm{f}}}" + }, + { + "bbox": [ + 54, + 462, + 295, + 617 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 54, + 462, + 295, + 617 + ], + "type": "inline_equation", + "content": "\\mathbf{f}" + }, + { + "bbox": [ + 54, + 462, + 295, + 617 + ], + "type": "text", + "content": " is used for both view-dependent effects and lidar properties. Last, our representation contains a learnable embedding per sensor to model their specific appearance characteristics." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 618, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 618, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 54, + 618, + 295, + 713 + ], + "type": "text", + "content": "To handle dynamics, we follow the commonly used scene graph decomposition [29, 35, 43, 51] and divide the scene into a static background and a set of dynamic actors. Each dynamic actor is described by a 3D bounding box and a sequence of SE(3) poses, obtained either from an off-the-shelf object detector and tracker, or annotations. Each Gaussian has a non-learnable ID, indicating whether it is assigned to the static world, or to which actor it belongs. For Gaussians" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 354, + 555, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 354, + 555, + 462 + ], + "spans": [ + { + "bbox": [ + 313, + 354, + 555, + 462 + ], + "type": "text", + "content": "assigned to actors, their mean and covariance are expressed in the local coordinate system of the corresponding axis-aligned bounding box. To compose the scene at a given time " + }, + { + "bbox": [ + 313, + 354, + 555, + 462 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 354, + 555, + 462 + ], + "type": "text", + "content": ", Gaussians assigned to bounding boxes are transformed to world coordinates based on their actor's pose. Since pose estimates for actors may contain inaccuracies, we adjust these with learnable offsets. Furthermore, each actor has a velocity initialized from pose differences and a learnable velocity offset." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 468, + 424, + 481 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 468, + 424, + 481 + ], + "spans": [ + { + "bbox": [ + 313, + 468, + 424, + 481 + ], + "type": "text", + "content": "3.2. Camera rendering" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 486, + 555, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 486, + 555, + 569 + ], + "spans": [ + { + "bbox": [ + 313, + 486, + 555, + 569 + ], + "type": "text", + "content": "Given a posed camera, we compose the set of Gaussians at the corresponding capture time " + }, + { + "bbox": [ + 313, + 486, + 555, + 569 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 486, + 555, + 569 + ], + "type": "text", + "content": " and render an image " + }, + { + "bbox": [ + 313, + 486, + 555, + 569 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 313, + 486, + 555, + 569 + ], + "type": "text", + "content": " using the efficient tile-based rendering from 3DGS [14]. While we retain 3DGS's high-level steps—projection and view frustum culling, tile-assignment, depth sorting, and tile-based rasterization—we introduce key adaptations to better model the unique characteristics of AD data." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "text", + "content": "Projection, tiling, and sorting: Each mean and covariance is transformed from world to camera coordinates, giving " + }, + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\pmb{\\mu}^{C}" + }, + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\pmb{\\Sigma}^{C}" + }, + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "text", + "content": ". The means are then converted to image space " + }, + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\pmb{\\mu}^{I} \\in \\mathbb{R}^{2}" + }, + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "text", + "content": " using the perspective projection, while the covariances are transformed using the two upper rows of the Jacobian of the projection, " + }, + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\pmb{\\Sigma}^{I} = \\mathbf{J}^{I}\\pmb{\\Sigma}^{C}(\\mathbf{J}^{I})^{\\mathrm{T}} \\in \\mathbb{R}^{2 \\times 2}" + }, + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "text", + "content": ". Gaussians outside the view frustum are culled, where, for efficiency, 3DGS approximates the extent of the Gaussians using a square axis-aligned bounding box (AABB) that covers their " + }, + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "inline_equation", + "content": "99\\%" + }, + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "text", + "content": " confidence level. Further, 3DGS divides the image into tiles, each measuring " + }, + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "inline_equation", + "content": "16 \\times 16" + }, + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "text", + "content": " pixels, and assigns Gaussians to all tiles they intersect, duplicating them if necessary. This" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 750, + 317, + 759 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 750, + 317, + 759 + ], + "spans": [ + { + "bbox": [ + 293, + 750, + 317, + 759 + ], + "type": "text", + "content": "11984" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "type": "text", + "content": "way, each pixel only has to process a subset of all Gaussians during rasterization. Lastly, the Gaussians are sorted based on the " + }, + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "type": "text", + "content": "-depth of their mean " + }, + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "type": "inline_equation", + "content": "\\pmb{\\mu}^{C}" + }, + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 108, + 296, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 108, + 296, + 323 + ], + "spans": [ + { + "bbox": [ + 55, + 108, + 296, + 323 + ], + "type": "text", + "content": "Rolling shutter: Many cameras use a rolling shutter, where the capture of an image is not instantaneous, but is done row-by-row, allowing the camera to move during the exposure. Previous work [35] has highlighted the importance of modeling rolling shutter effects due to the high sensor velocities present in AD data, where their ray-tracing-based method can easily account for this by shifting the origins of each ray. For 3DGS, the equivalent would require projecting all 3D Gaussians to all camera poses encountered during an exposure, as the Gaussians' positions relative to the camera change with time. As this would be prohibitively expensive, we draw inspiration from [31] and instead approximate the effects directly in 2D image space. Specifically, each Gaussian's velocity relative to the camera is projected to image space, and their pixel mean " + }, + { + "bbox": [ + 55, + 108, + 296, + 323 + ], + "type": "inline_equation", + "content": "\\mu^I" + }, + { + "bbox": [ + 55, + 108, + 296, + 323 + ], + "type": "text", + "content": " is adjusted during rasterization based on the capture time of the pixel. However, [31] considered static scenes only, hence we adjust the formulation to account for the dynamics as well." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 323, + 293, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 323, + 293, + 335 + ], + "spans": [ + { + "bbox": [ + 67, + 323, + 293, + 335 + ], + "type": "text", + "content": "For each Gaussian, its pixel velocity is approximated as" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 88, + 342, + 295, + 357 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 342, + 295, + 357 + ], + "spans": [ + { + "bbox": [ + 88, + 342, + 295, + 357 + ], + "type": "interline_equation", + "content": "\\mathbf {v} ^ {I} = \\mathbf {J} ^ {I} \\left(- \\boldsymbol {\\omega} _ {C} \\times \\boldsymbol {\\mu} ^ {C} - \\mathbf {v} _ {C} + \\mathbf {v} _ {\\mathrm {d y n}}\\right) \\in \\mathbb {R} ^ {2}, \\tag {1}", + "image_path": "ccfcaafd1bf05eaa20129f46339caa1e3d4c8a4fb1b62649a7c32cdbeadb839f.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 363, + 295, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 363, + 295, + 387 + ], + "spans": [ + { + "bbox": [ + 55, + 363, + 295, + 387 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 363, + 295, + 387 + ], + "type": "inline_equation", + "content": "\\omega_{C}" + }, + { + "bbox": [ + 55, + 363, + 295, + 387 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 363, + 295, + 387 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_C" + }, + { + "bbox": [ + 55, + 363, + 295, + 387 + ], + "type": "text", + "content": " denote the angular and linear velocity of the camera expressed in the camera coordinate system, and" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 91, + 393, + 295, + 407 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 393, + 295, + 407 + ], + "spans": [ + { + "bbox": [ + 91, + 393, + 295, + 407 + ], + "type": "interline_equation", + "content": "\\mathbf {v} _ {\\mathrm {d y n}} = T ^ {\\mathrm {a c t} \\rightarrow C} \\left(\\boldsymbol {\\omega} _ {\\mathrm {a c t}} \\times \\boldsymbol {\\mu} ^ {\\mathrm {a c t}} + \\mathbf {v} _ {\\mathrm {a c t}}\\right) \\in \\mathbb {R} ^ {3}, \\tag {2}", + "image_path": "889eee2c3f74ec64e0f66ccb9910b0fff4ffc72b266f432f7a2502a1154e6c59.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 415, + 296, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 415, + 296, + 510 + ], + "spans": [ + { + "bbox": [ + 55, + 415, + 296, + 510 + ], + "type": "text", + "content": "denotes the velocity of the Gaussian induced by the angular " + }, + { + "bbox": [ + 55, + 415, + 296, + 510 + ], + "type": "inline_equation", + "content": "\\omega^{\\mathrm{dyn}}" + }, + { + "bbox": [ + 55, + 415, + 296, + 510 + ], + "type": "text", + "content": " and linear velocity " + }, + { + "bbox": [ + 55, + 415, + 296, + 510 + ], + "type": "inline_equation", + "content": "\\mathbf{v}^{\\mathrm{dyn}}" + }, + { + "bbox": [ + 55, + 415, + 296, + 510 + ], + "type": "text", + "content": " of its associated actor (expressed in actor coordinate system) and " + }, + { + "bbox": [ + 55, + 415, + 296, + 510 + ], + "type": "inline_equation", + "content": "\\mu^{\\mathrm{act}}" + }, + { + "bbox": [ + 55, + 415, + 296, + 510 + ], + "type": "text", + "content": " denotes the Gaussian mean in the actor coordinate system. To transform velocities from actor to camera coordinates, " + }, + { + "bbox": [ + 55, + 415, + 296, + 510 + ], + "type": "inline_equation", + "content": "T^{\\mathrm{act}\\rightarrow C}" + }, + { + "bbox": [ + 55, + 415, + 296, + 510 + ], + "type": "text", + "content": " is the composition of the actor-to-world and world-to-camera transforms. For Gaussians that are part of the static background, " + }, + { + "bbox": [ + 55, + 415, + 296, + 510 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_{\\mathrm{dyn}}" + }, + { + "bbox": [ + 55, + 415, + 296, + 510 + ], + "type": "text", + "content": " is zero. See Appendix D for further details." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 511, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 511, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 511, + 296, + 713 + ], + "type": "text", + "content": "We account for the pixel velocity when culling Gaussians and checking intersections between Gaussians and tiles by increasing the approximated Gaussian extent. Instead of the square AABB used by 3DGS and [31], we use a rectangular AABB covering " + }, + { + "bbox": [ + 55, + 511, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\boldsymbol{\\Sigma}^{I}" + }, + { + "bbox": [ + 55, + 511, + 296, + 713 + ], + "type": "text", + "content": " at three standard deviations and add " + }, + { + "bbox": [ + 55, + 511, + 296, + 713 + ], + "type": "inline_equation", + "content": "|\\mathbf{v}_I| \\cdot t_{\\mathrm{rs}} / 2" + }, + { + "bbox": [ + 55, + 511, + 296, + 713 + ], + "type": "text", + "content": " to its extent. Here, " + }, + { + "bbox": [ + 55, + 511, + 296, + 713 + ], + "type": "inline_equation", + "content": "t_{\\mathrm{rs}}" + }, + { + "bbox": [ + 55, + 511, + 296, + 713 + ], + "type": "text", + "content": " denotes the rolling shutter duration, i.e., the time difference between the last and first row of pixels. The increase in extent corresponds to the area covered by the Gaussian mean during the rolling shutter time, assuming the sensor timestamp is in the middle of exposure. As this is not always the case, we further include a learnable time offset, describing the difference between the sensor timestamp and the middle of the exposure. Using rectangular AABBs reduces unnecessary intersections compared to [14, 31], especially for narrow Gaussians and Gaussians with axis-aligned velocities, the latter often being the case for side-facing cameras." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 555, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 144 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 144 + ], + "type": "text", + "content": "Rasterization: 3DGS rasterizes pixels in parallel by launching a thread block per tile and assigning each thread to a pixel within the tile. For each pixel, its coordinates " + }, + { + "bbox": [ + 313, + 72, + 555, + 144 + ], + "type": "inline_equation", + "content": "\\mathbf{p} = [p_u,p_v]^{\\mathrm{T}}" + }, + { + "bbox": [ + 313, + 72, + 555, + 144 + ], + "type": "text", + "content": " are inferred from the block and the thread index. From this, we find the time difference between the pixel's capture time and the image's middle row as" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 381, + 154, + 555, + 167 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 381, + 154, + 555, + 167 + ], + "spans": [ + { + "bbox": [ + 381, + 154, + 555, + 167 + ], + "type": "interline_equation", + "content": "t _ {\\mathrm {p i x}} = \\left(p _ {v} / H - 0. 5\\right) \\cdot t _ {\\mathrm {r s}}. \\tag {3}", + "image_path": "a37b891c3e1c2cb9347554a738ec1ec2ae085a15aaac2b2726611e203de3b419.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 176, + 553, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 176, + 553, + 201 + ], + "spans": [ + { + "bbox": [ + 313, + 176, + 553, + 201 + ], + "type": "text", + "content": "To rasterize a pixel, we " + }, + { + "bbox": [ + 313, + 176, + 553, + 201 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 176, + 553, + 201 + ], + "type": "text", + "content": "-blend RGB values " + }, + { + "bbox": [ + 313, + 176, + 553, + 201 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_i^{\\mathrm{rgb}}" + }, + { + "bbox": [ + 313, + 176, + 553, + 201 + ], + "type": "text", + "content": " and features " + }, + { + "bbox": [ + 313, + 176, + 553, + 201 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_i" + }, + { + "bbox": [ + 313, + 176, + 553, + 201 + ], + "type": "text", + "content": " of the depth-sorted Gaussians intersecting the current tile" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 351, + 209, + 555, + 243 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 209, + 555, + 243 + ], + "spans": [ + { + "bbox": [ + 351, + 209, + 555, + 243 + ], + "type": "interline_equation", + "content": "\\left[ \\mathbf {F} _ {\\mathbf {p}} ^ {\\mathrm {r g b}}, \\mathbf {F} _ {\\mathbf {p}} \\right] = \\sum_ {i} \\left[ \\mathbf {f} _ {i} ^ {\\mathrm {r g b}}, \\mathbf {f} _ {i} \\right] \\alpha_ {i} \\prod_ {j = 1} ^ {i - 1} (1 - \\alpha_ {j}), \\tag {4}", + "image_path": "472b42f78829dc035a66efcb7e0fd9ba64777faf1840a384c3d0897db1b94686.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 251, + 397, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 251, + 397, + 263 + ], + "spans": [ + { + "bbox": [ + 313, + 251, + 397, + 263 + ], + "type": "text", + "content": "with " + }, + { + "bbox": [ + 313, + 251, + 397, + 263 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 251, + 397, + 263 + ], + "type": "text", + "content": " computed as" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 321, + 273, + 555, + 304 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 273, + 555, + 304 + ], + "spans": [ + { + "bbox": [ + 321, + 273, + 555, + 304 + ], + "type": "interline_equation", + "content": "\\alpha_ {i} = \\sqrt {\\frac {\\left| \\boldsymbol {\\Sigma} _ {i} ^ {I} \\right|}{\\left| \\boldsymbol {\\Sigma} _ {i} ^ {I} + s \\mathbf {I} \\right|}} o _ {i} \\exp \\left(- \\frac {1}{2} \\boldsymbol {\\Delta} _ {i} ^ {\\mathrm {T}} \\left(\\boldsymbol {\\Sigma} _ {i} ^ {I} + s \\mathbf {I}\\right) ^ {- 1} \\boldsymbol {\\Delta} _ {i}\\right) \\tag {5}", + "image_path": "d8154069a3f38db3699829b19d385bfd035dbcf588947a66acf5b2428dd395b0.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 314, + 312, + 332, + 322 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 312, + 332, + 322 + ], + "spans": [ + { + "bbox": [ + 314, + 312, + 332, + 322 + ], + "type": "text", + "content": "and" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 381, + 323, + 555, + 338 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 381, + 323, + 555, + 338 + ], + "spans": [ + { + "bbox": [ + 381, + 323, + 555, + 338 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\Delta} _ {i} = \\mathbf {p} - \\left(\\boldsymbol {\\mu} _ {i} ^ {I} + \\mathbf {v} _ {i} ^ {I} t _ {\\mathrm {p i x}}\\right). \\tag {6}", + "image_path": "201edb834c20332264a64b25aed14694ff9a692d7c951935a14c4993fcbf94d6.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 342, + 555, + 401 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 342, + 555, + 401 + ], + "spans": [ + { + "bbox": [ + 313, + 342, + 555, + 401 + ], + "type": "text", + "content": "As in [31], we calculate the distance " + }, + { + "bbox": [ + 313, + 342, + 555, + 401 + ], + "type": "inline_equation", + "content": "\\Delta_{i}" + }, + { + "bbox": [ + 313, + 342, + 555, + 401 + ], + "type": "text", + "content": " between pixels and Gaussians using the means compensated by rolling shutter, effectively shifting Gaussians to the correct location in the image. Compared to 3DGS, in Eq. (5) we use the 3DGS+EWA [52] formulation from [47], where " + }, + { + "bbox": [ + 313, + 342, + 555, + 401 + ], + "type": "inline_equation", + "content": "s = 0.3" + }, + { + "bbox": [ + 313, + 342, + 555, + 401 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 402, + 554, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 402, + 554, + 460 + ], + "spans": [ + { + "bbox": [ + 313, + 402, + 554, + 460 + ], + "type": "text", + "content": "View-dependent effects are modeled using a small CNN; given the feature map " + }, + { + "bbox": [ + 313, + 402, + 554, + 460 + ], + "type": "inline_equation", + "content": "\\mathbf{F} \\in \\mathbb{R}^{H \\times W \\times D_{\\mathrm{f}}}" + }, + { + "bbox": [ + 313, + 402, + 554, + 460 + ], + "type": "text", + "content": ", the corresponding ray directions " + }, + { + "bbox": [ + 313, + 402, + 554, + 460 + ], + "type": "inline_equation", + "content": "\\mathbf{d} \\in \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 313, + 402, + 554, + 460 + ], + "type": "text", + "content": ", and a camera-specific learned embedding " + }, + { + "bbox": [ + 313, + 402, + 554, + 460 + ], + "type": "inline_equation", + "content": "\\mathbf{e}" + }, + { + "bbox": [ + 313, + 402, + 554, + 460 + ], + "type": "text", + "content": ", it predicts a pixel-wise affine mapping " + }, + { + "bbox": [ + 313, + 402, + 554, + 460 + ], + "type": "inline_equation", + "content": "M \\in \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 313, + 402, + 554, + 460 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 402, + 554, + 460 + ], + "type": "inline_equation", + "content": "\\mathbf{b} \\in \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 313, + 402, + 554, + 460 + ], + "type": "text", + "content": " applied to " + }, + { + "bbox": [ + 313, + 402, + 554, + 460 + ], + "type": "inline_equation", + "content": "\\mathbf{F}^{\\mathrm{rgb}}" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 369, + 471, + 555, + 484 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 369, + 471, + 555, + 484 + ], + "spans": [ + { + "bbox": [ + 369, + 471, + 555, + 484 + ], + "type": "interline_equation", + "content": "M, \\mathbf {b} = \\operatorname {C N N} (\\operatorname {c o n c a t} (\\mathbf {F}, \\mathbf {d}, \\mathbf {e})), \\tag {7}", + "image_path": "257ebe048ebff24157141c587ffb292578a460e9d240a387bd75cb33a8e461cd.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 384, + 487, + 555, + 500 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 384, + 487, + 555, + 500 + ], + "spans": [ + { + "bbox": [ + 384, + 487, + 555, + 500 + ], + "type": "interline_equation", + "content": "I = (1 + M) \\odot \\mathbf {F} ^ {\\mathrm {r g b}} + \\mathbf {b}. \\tag {8}", + "image_path": "e0550c74549389453f6fae07fe6a0da53d5c8ce606dec639ba2c40ea5135a327.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 509, + 555, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 509, + 555, + 568 + ], + "spans": [ + { + "bbox": [ + 313, + 509, + 555, + 568 + ], + "type": "text", + "content": "We found that using a small CNN instead of an MLP is essential for efficient texture modeling, e.g., for high-frequency road surfaces. In addition, inspired by [30], the camera-specific learned embeddings enable modeling differences in exposure." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 576, + 411, + 589 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 576, + 411, + 589 + ], + "spans": [ + { + "bbox": [ + 313, + 576, + 411, + 589 + ], + "type": "text", + "content": "3.3. Lidar rendering" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "content": "Lidar sensors enable self-driving vehicles to measure the distance and reflectivity (intensity) of a discrete set of points. They do so by emitting laser beam pulses and measuring the time of flight to determine distance and returning power for reflectivity. Most AD datasets deploy lidars that use several laser diodes (16-128) mounted in a vertical array, where the array of diodes is rotated to capture " + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "content": " data and a single scan often takes " + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "inline_equation", + "content": "100\\mathrm{ms}" + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "content": " to capture. Therefore, we focus on this type of lidar. However, we note that modeling other types of lidars, such as solid-state lidars [21], can be easily" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 750, + 317, + 759 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 750, + 317, + 759 + ], + "spans": [ + { + "bbox": [ + 293, + 750, + 317, + 759 + ], + "type": "text", + "content": "11985" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 119 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 119 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 119 + ], + "type": "text", + "content": "done in our method by modifying the projection accordingly. Our proposed lidar rendering follows the high-level steps of the image rendering, but we modify each component to accurately model lidar characteristics." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 120, + 295, + 207 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 120, + 295, + 207 + ], + "spans": [ + { + "bbox": [ + 55, + 120, + 295, + 207 + ], + "type": "text", + "content": "Projection: For a posed lidar and Gaussians composed at its capture time " + }, + { + "bbox": [ + 55, + 120, + 295, + 207 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 120, + 295, + 207 + ], + "type": "text", + "content": ", we transform each mean and covariance from world to lidar coordinates, yielding " + }, + { + "bbox": [ + 55, + 120, + 295, + 207 + ], + "type": "inline_equation", + "content": "\\boldsymbol{\\mu}^{L} = [x,y,z]^{\\mathrm{T}}" + }, + { + "bbox": [ + 55, + 120, + 295, + 207 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 120, + 295, + 207 + ], + "type": "inline_equation", + "content": "\\boldsymbol{\\Sigma}^{L}" + }, + { + "bbox": [ + 55, + 120, + 295, + 207 + ], + "type": "text", + "content": ". Raw lidar data are generally captured in spherical coordinates azimuth " + }, + { + "bbox": [ + 55, + 120, + 295, + 207 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 55, + 120, + 295, + 207 + ], + "type": "text", + "content": ", elevation " + }, + { + "bbox": [ + 55, + 120, + 295, + 207 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 55, + 120, + 295, + 207 + ], + "type": "text", + "content": ", and range " + }, + { + "bbox": [ + 55, + 120, + 295, + 207 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 55, + 120, + 295, + 207 + ], + "type": "text", + "content": ". Thus, we transform Gaussian means from lidar coordinates to spherical coordinates" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 102, + 213, + 296, + 255 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 213, + 296, + 255 + ], + "spans": [ + { + "bbox": [ + 102, + 213, + 296, + 255 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\mu} ^ {S} = \\left[ \\begin{array}{l} \\phi \\\\ \\omega \\\\ r \\end{array} \\right] = \\left[ \\begin{array}{c} \\arctan 2 (y, x) \\\\ \\arcsin (z / r) \\\\ \\sqrt {x ^ {2} + y ^ {2} + z ^ {2}} \\end{array} \\right], \\tag {9}", + "image_path": "5c31e2274602f8955b23c8596696d203f3b47d03e9bd4d8d7b2633968836f449.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 263, + 295, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 263, + 295, + 289 + ], + "spans": [ + { + "bbox": [ + 55, + 263, + 295, + 289 + ], + "type": "text", + "content": "and convert the covariance " + }, + { + "bbox": [ + 55, + 263, + 295, + 289 + ], + "type": "inline_equation", + "content": "\\pmb{\\Sigma}^{L}" + }, + { + "bbox": [ + 55, + 263, + 295, + 289 + ], + "type": "text", + "content": " with the Jacobian of the transform Eq. (9) as " + }, + { + "bbox": [ + 55, + 263, + 295, + 289 + ], + "type": "inline_equation", + "content": "\\pmb{\\Sigma}^{S} = \\mathbf{J}^{S}\\pmb{\\Sigma}^{L}(\\mathbf{J}^{S})^{\\mathrm{T}}" + }, + { + "bbox": [ + 55, + 263, + 295, + 289 + ], + "type": "text", + "content": ", where" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 68, + 297, + 296, + 350 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 297, + 296, + 350 + ], + "spans": [ + { + "bbox": [ + 68, + 297, + 296, + 350 + ], + "type": "interline_equation", + "content": "\\mathbf {J} ^ {S} = \\left[ \\begin{array}{c c c} - \\frac {y}{x ^ {2} + y ^ {2}} & \\frac {x}{x ^ {2} + y ^ {2}} & 0 \\\\ - \\frac {x z}{r ^ {2} \\sqrt {x ^ {2} + y ^ {2}}} & - \\frac {y z}{r ^ {2} \\sqrt {x ^ {2} + y ^ {2}}} & \\frac {\\sqrt {x ^ {2} + y ^ {2}}}{r ^ {2}} \\\\ \\frac {x}{r} & \\frac {y}{r} & \\frac {z}{r} \\end{array} \\right], \\tag {10}", + "image_path": "4ac6e333c5f392344023cea08d87edb77d1555f81ce4ed0ff7cc1d0ce544de7f.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 357, + 295, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 357, + 295, + 380 + ], + "spans": [ + { + "bbox": [ + 55, + 357, + 295, + 380 + ], + "type": "text", + "content": "and " + }, + { + "bbox": [ + 55, + 357, + 295, + 380 + ], + "type": "inline_equation", + "content": "\\pmb{\\mu}^{S}" + }, + { + "bbox": [ + 55, + 357, + 295, + 380 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 357, + 295, + 380 + ], + "type": "inline_equation", + "content": "\\pmb{\\Sigma}^{S}" + }, + { + "bbox": [ + 55, + 357, + 295, + 380 + ], + "type": "text", + "content": " denote mean and covariance in spherical coordinates." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 382, + 296, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 382, + 296, + 407 + ], + "spans": [ + { + "bbox": [ + 55, + 382, + 296, + 407 + ], + "type": "text", + "content": "Same as for the camera, we account for the rolling shutter effect by approximating its impact in sensor space" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 88, + 414, + 295, + 430 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 414, + 295, + 430 + ], + "spans": [ + { + "bbox": [ + 88, + 414, + 295, + 430 + ], + "type": "interline_equation", + "content": "\\mathbf {v} ^ {S} = \\mathbf {J} ^ {S} \\left(- \\boldsymbol {\\omega} _ {L} \\times \\boldsymbol {\\mu} ^ {L} - \\mathbf {v} _ {L} + \\mathbf {v} _ {\\mathrm {d y n}}\\right) \\in \\mathbb {R} ^ {3}. \\tag {11}", + "image_path": "3c171fbafe43f91644e91cd94c351082b5d86717807c75dd4215a702ec65a7a8.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 437, + 296, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 437, + 296, + 497 + ], + "spans": [ + { + "bbox": [ + 55, + 437, + 296, + 497 + ], + "type": "text", + "content": "Note that " + }, + { + "bbox": [ + 55, + 437, + 296, + 497 + ], + "type": "inline_equation", + "content": "\\mathbf{v}^S" + }, + { + "bbox": [ + 55, + 437, + 296, + 497 + ], + "type": "text", + "content": " is three-dimensional as opposed to " + }, + { + "bbox": [ + 55, + 437, + 296, + 497 + ], + "type": "inline_equation", + "content": "\\mathbf{v}^I" + }, + { + "bbox": [ + 55, + 437, + 296, + 497 + ], + "type": "text", + "content": " in Eq. (1), as we need this for accurate range rasterization. Last, we use the first two elements of " + }, + { + "bbox": [ + 55, + 437, + 296, + 497 + ], + "type": "inline_equation", + "content": "\\mathbf{v}^S t_{\\mathrm{rs}} / 2" + }, + { + "bbox": [ + 55, + 437, + 296, + 497 + ], + "type": "text", + "content": " to increase the extent of the rectangular AABB approximating the size of the Gaussian, and cull Gaussians outside the lidar's field-of-view." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 498, + 296, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 296, + 653 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 296, + 653 + ], + "type": "text", + "content": "Tiling and sorting: At the core of 3DGS's efficiency is its tile-based rasterization. While lidars often have a fixed azimuth resolution, many use non-equidistant spacing between diodes to get higher resolution in areas of interest, resulting in nonlinear elevation resolution, as seen in Fig. 1. In line with this, we design tiles with a fixed number of diodes vertically and a fixed azimuth resolution horizontally. Horizontally, intersecting tiles are found by dividing azimuth coordinates by the common tile azimuth size and wrapping the results within " + }, + { + "bbox": [ + 55, + 498, + 296, + 653 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 55, + 498, + 296, + 653 + ], + "type": "text", + "content": ". Vertically, we sort the elevation tile boundaries and find intersections with a short for-loop, see Appendix A for details. After tile assignment, Gaussians are sorted based on the range of their mean " + }, + { + "bbox": [ + 55, + 498, + 296, + 653 + ], + "type": "inline_equation", + "content": "\\mu^S" + }, + { + "bbox": [ + 55, + 498, + 296, + 653 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "content": "Comparing our strategy to using tiles of equal size, similar to the depth image-based rendering in [7], we avoid many unnecessary computations. There, the tile size must be determined from the finest resolution to ensure having at least as many threads as lidar points. Consequently, tiles" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 72, + 554, + 107 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 107 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 107 + ], + "type": "text", + "content": "in areas where the lidar is sparse would have a superficially high resolution and might even lack the corresponding lidar points to render." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 109, + 555, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 109, + 555, + 263 + ], + "spans": [ + { + "bbox": [ + 313, + 109, + 555, + 263 + ], + "type": "text", + "content": "Rasterization: Lidar data are rendered using our proposed differentiable tile-based rasterizer. Again, each tile launches one thread block, where each thread is responsible for rasterizing one lidar point. Each thread is provided the azimuth, elevation, and capture time for it to rasterize. These are either determined from lidar specifications or, for training, calculated from existing point clouds, see Appendix A for details. The reason for this is that for training, one cannot expect the data to follow perfectly straight scan lines. Therefore, we cannot follow the image strategy and rely on the block and thread index to directly infer the pixel coordinates. Further, the azimuth and elevation of a lidar ray do not necessarily determine its time offset to the center of the lidar scan." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 265, + 554, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 265, + 554, + 348 + ], + "spans": [ + { + "bbox": [ + 313, + 265, + 554, + 348 + ], + "type": "text", + "content": "For each lidar point, we " + }, + { + "bbox": [ + 313, + 265, + 554, + 348 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 265, + 554, + 348 + ], + "type": "text", + "content": "-blend the Gaussians' features " + }, + { + "bbox": [ + 313, + 265, + 554, + 348 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_i" + }, + { + "bbox": [ + 313, + 265, + 554, + 348 + ], + "type": "text", + "content": " using Eqs. (4) to (6). Inspired by [47] arguing for " + }, + { + "bbox": [ + 313, + 265, + 554, + 348 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 313, + 265, + 554, + 348 + ], + "type": "text", + "content": " in Eq. (5) to correspond to one pixel, we set it to correspond to the geometric mean of the lidar's vertical and horizontal beam divergence. The rendered features are concatenated with their corresponding ray direction and decoded to intensity and ray drop probability using a small MLP." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 349, + 554, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 349, + 554, + 373 + ], + "spans": [ + { + "bbox": [ + 313, + 349, + 554, + 373 + ], + "type": "text", + "content": "To render the expected range of a point, we " + }, + { + "bbox": [ + 313, + 349, + 554, + 373 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 349, + 554, + 373 + ], + "type": "text", + "content": "-blend the rolling shutter corrected ranges for each Gaussian" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 397, + 382, + 555, + 397 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 397, + 382, + 555, + 397 + ], + "spans": [ + { + "bbox": [ + 397, + 382, + 555, + 397 + ], + "type": "interline_equation", + "content": "r _ {i, \\mathrm {r s}} = r _ {i} + \\mathbf {v} _ {r} ^ {S} t _ {l}, \\tag {12}", + "image_path": "8b613942f92fdae7d592543496c2e05069fd425fbaf55af8debd8eb04719e344.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 407, + 554, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 407, + 554, + 514 + ], + "spans": [ + { + "bbox": [ + 313, + 407, + 554, + 514 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 407, + 554, + 514 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_r^S" + }, + { + "bbox": [ + 313, + 407, + 554, + 514 + ], + "type": "text", + "content": " is the range-component of the relative speed to the lidar and " + }, + { + "bbox": [ + 313, + 407, + 554, + 514 + ], + "type": "inline_equation", + "content": "t_l" + }, + { + "bbox": [ + 313, + 407, + 554, + 514 + ], + "type": "text", + "content": " is the time between the capture of the current lidar point, and the middle of the lidar scan. Further, we render the median range as the rolling shutter corrected range of the first Gaussian that satisfies " + }, + { + "bbox": [ + 313, + 407, + 554, + 514 + ], + "type": "inline_equation", + "content": "\\prod_{j=1}^{i}(1 - \\alpha_j) < 0.5" + }, + { + "bbox": [ + 313, + 407, + 554, + 514 + ], + "type": "text", + "content": " during the " + }, + { + "bbox": [ + 313, + 407, + 554, + 514 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 407, + 554, + 514 + ], + "type": "text", + "content": "-blending in Eq. (4). While we use the expected range for training, the median range is used during inference as it, in contrast to the expected range, does not yield depths between Gaussians." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 524, + 496, + 537 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 524, + 496, + 537 + ], + "spans": [ + { + "bbox": [ + 313, + 524, + 496, + 537 + ], + "type": "text", + "content": "3.4. Optimization and implementation" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 542, + 554, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 542, + 554, + 555 + ], + "spans": [ + { + "bbox": [ + 313, + 542, + 554, + 555 + ], + "type": "text", + "content": "All model components are optimized jointly using the loss" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 324, + 579, + 555, + 609 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 579, + 555, + 609 + ], + "spans": [ + { + "bbox": [ + 324, + 579, + 555, + 609 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} = \\lambda_ {r} \\mathcal {L} _ {1} + (1 - \\lambda_ {r}) \\mathcal {L} _ {\\text {S S I M}} + \\lambda_ {\\text {d e p t h}} \\mathcal {L} _ {\\text {d e p t h}} + \\lambda_ {\\text {l o s}} \\mathcal {L} _ {\\text {l o s}} + \\\\ \\lambda_ {\\text {i n t e n s}} \\mathcal {L} _ {\\text {i n t e n}} + \\lambda_ {\\text {r a y d r o p}} \\mathcal {L} _ {\\text {B C E}} + \\lambda_ {\\text {M C M C}} \\mathcal {L} _ {\\text {M C M C}}, \\tag {13} \\\\ \\end{array}", + "image_path": "927a6cafb5aabddea4514beccb33723698644b1bee3fd912c2bc1930d124243b.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_1" + }, + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{SSIM}}" + }, + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "type": "text", + "content": " are L1 and SSIM losses on the rendered images. " + }, + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{depth}}" + }, + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{inten}}" + }, + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "type": "text", + "content": " are L2 losses on the rendered expected lidar range and intensity. " + }, + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{los}}" + }, + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "type": "text", + "content": " is a line-of-sight loss, penalizing opacity before the ground truth lidar range. " + }, + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{BCE}}" + }, + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "type": "text", + "content": " is a binary cross-entropy loss on the predicted ray drop probability, where ground-truth is generated in the same way as for NeuRAD. " + }, + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{MCMC}}" + }, + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "type": "text", + "content": " is the opacity and scale regularization used in [16]." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 750, + 318, + 759 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 750, + 318, + 759 + ], + "spans": [ + { + "bbox": [ + 293, + 750, + 318, + 759 + ], + "type": "text", + "content": "11986" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 70, + 553, + 258 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 553, + 258 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 553, + 258 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 553, + 258 + ], + "type": "image", + "image_path": "9b7e4808099cb50ebee8698abe578747e00b562dbcd15da935fdbc3734773916.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 265, + 555, + 301 + ], + "lines": [ + { + "bbox": [ + 55, + 265, + 555, + 301 + ], + "spans": [ + { + "bbox": [ + 55, + 265, + 555, + 301 + ], + "type": "text", + "content": "Figure 3. Compared to the baselines, SplatAD produces sharp images with a high level of detail. Further, the bottom row highlights the superiority of our lidar rendering. Projecting lidar points into images for depth supervision, as used by previous 3DGS methods, causes line-of-sight errors and incorrect volume carving due to the pose differences between camera and lidar." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 319, + 297, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 319, + 297, + 511 + ], + "spans": [ + { + "bbox": [ + 54, + 319, + 297, + 511 + ], + "type": "text", + "content": "We initialize the set of Gaussians using both lidar points and random points. Lidar points within bounding boxes are assigned to the corresponding actor. Further, all lidar points are projected into their closest image to set their initial color. The random points are divided into (1) uniformly sampled points within the lidar range and (2) points sampled linearly in disparity beyond the lidar range, both using random colors at initialization. Furthermore, instead of the 3DGS [14] splitting and densification strategy, we rely on the MCMC strategy introduced in [16]. We chose MCMC partially because we find it to improve far-field rendering quality and partially because it has predictable computation requirements, as it permits choosing the maximum number of Gaussian. We use [16] as is, without any special treatment of Gaussians assigned to dynamic actors. See Appendix B for further training details." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 512, + 297, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 512, + 297, + 632 + ], + "spans": [ + { + "bbox": [ + 55, + 512, + 297, + 632 + ], + "type": "text", + "content": "Implementation: For efficiency, we implement the forward and backward passes of the rolling shutter compensation and lidar projection and rasterization using custom CUDA kernels. SplatAD builds upon the collaborative open-source frameworks gsplat [46] and neurad-studio [34], and we hope that our contributions can benefit future research on multi-modal simulation, not only within the AD community. We train SplatAD for 30,000 iterations using the Adam optimizer, which takes an hour on a single NVIDIA A100. See Appendix B for further details." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 645, + 137, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 645, + 137, + 658 + ], + "spans": [ + { + "bbox": [ + 55, + 645, + 137, + 658 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "To validate the robustness of our method, we evaluate it across multiple popular AD datasets, using the same set of hyperparameters. For each dataset, we compare SplatAD to the best performing NeRF and 3DGS-based methods tailored" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 332, + 353, + 535, + 506 + ], + "blocks": [ + { + "bbox": [ + 314, + 319, + 554, + 346 + ], + "lines": [ + { + "bbox": [ + 314, + 319, + 554, + 346 + ], + "spans": [ + { + "bbox": [ + 314, + 319, + 554, + 346 + ], + "type": "text", + "content": "Table 1. NVS results for image, over three datasets. First, second, third." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 332, + 353, + 535, + 506 + ], + "lines": [ + { + "bbox": [ + 332, + 353, + 535, + 506 + ], + "spans": [ + { + "bbox": [ + 332, + 353, + 535, + 506 + ], + "type": "table", + "html": "
PSNR ↑SSIM ↑LPIPS ↓MP/s ↑Train (H) ↓
PandaSetUniSim23.120.6820.36011.50.9
NeuRAD25.800.7530.2509.71.3
PVG24.010.7120.45222.61.6
Street-GS24.730.7450.31463.12.3
OmniRE24.710.7450.31553.62.6
SplatAD (ours)26.700.8140.193104.11.0
Argo2UniSim22.350.6550.45810.01.1
NeuRAD26.180.7210.3109.71.3
PVG24.470.7120.50219.72.5
Street-GS25.520.7540.37492.82.6
OmniRE25.610.7530.37575.63.0
SplatAD (ours)28.400.8260.271133.81.2
nuScenesUniSim22.660.7430.41413.30.7
NeuRAD26.170.7900.31211.81.1
PVG24.940.7680.49718.81.4
Street-GS25.540.7990.37557.91.6
OmniRE25.500.7980.38646.91.9
SplatAD (ours)27.530.8490.302104.31.2
", + "image_path": "4bd13918f7d480ebe31f42ddf0c2f51d5415f113c397b8fddd30a1420ac62b0f.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 531, + 555, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 531, + 555, + 581 + ], + "spans": [ + { + "bbox": [ + 313, + 531, + 555, + 581 + ], + "type": "text", + "content": "to AD data on novel view synthesis for both camera and lidar data. Further, we ablate key elements of SplatAD and measure the impact they have on sensor realism and rendering speed." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 582, + 557, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 557, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 557, + 715 + ], + "type": "text", + "content": "Datasets: We perform experiments on PandaSet [41], Argoverse2 [38] and nuScenes [5]. These datasets vary in lidars, camera appearance, and resolution. For PandaSet and nuScenes, we use the six available cameras. For Argoverse2, we use the seven ring cameras, leaving out the black-and-white stereo cameras. Further, some cameras are cropped slightly to remove views of the ego-vehicle, such as the hood and the trunk. We use the same 10 challenging sequences per dataset as in [35], containing different types of illumination, dynamic actors, and velocities. We also follow the evaluation protocol of [35] i.e., using full resolution data," + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 750, + 318, + 760 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 750, + 318, + 760 + ], + "spans": [ + { + "bbox": [ + 293, + 750, + 318, + 760 + ], + "type": "text", + "content": "11987" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 74, + 102, + 276, + 232 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 297, + 95 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 297, + 95 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 297, + 95 + ], + "type": "text", + "content": "Table 2. NVS results for lidar, over three datasets. " + }, + { + "bbox": [ + 55, + 70, + 297, + 95 + ], + "type": "inline_equation", + "content": "{}^{8}" + }, + { + "bbox": [ + 55, + 70, + 297, + 95 + ], + "type": "text", + "content": " without missing points. First ,second ,third ." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 74, + 102, + 276, + 232 + ], + "lines": [ + { + "bbox": [ + 74, + 102, + 276, + 232 + ], + "spans": [ + { + "bbox": [ + 74, + 102, + 276, + 232 + ], + "type": "table", + "html": "
Depth ↓Intensity ↓Drop acc. ↑CD ↓MR/s ↑
PandaSetUniSim0.080.086-10.3$0.9
NeuRAD0.010.06396.21.91.1
PVG38.74--125.2$0.3
Street-GS6.18--37.3$0.8
OmniRE2.88--29.8$0.6
SplatAD (ours)0.010.05996.71.619.5
Argo2UniSim0.180.081-29.2$0.7
NeuRAD0.020.05892.22.60.9
SplatAD (ours)0.020.05292.62.89.5
nuScenesUniSim0.060.063-52.2$1.0
NeuRAD0.010.04293.16.31.1
PVG10.34--77.9$0.05
Street-GS1.56--4.5$0.2
OmniRE1.71--4.6$0.1
SplatAD (ours)0.020.03693.81.75.7
", + "image_path": "aa1652697f1d36c050ba5ec07f6d612a2cebaff2fe0920073068f804c0ede250.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 252, + 295, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 252, + 295, + 324 + ], + "spans": [ + { + "bbox": [ + 54, + 252, + 295, + 324 + ], + "type": "text", + "content": "and, for novel view synthesis (NVS), use every other frame for training and the remaining frames for hold-out validation. We believe that this setting is more challenging than what is often encountered in previous work [6, 7, 43], where fewer datasets are considered, images are downsampled, and a larger fraction of the data is used for training." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 325, + 295, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 325, + 295, + 468 + ], + "spans": [ + { + "bbox": [ + 54, + 325, + 295, + 468 + ], + "type": "text", + "content": "Baselines: We compare our results with the best performing NeRF and 3DGS methods on AD data, namely, UniSim [45], NeuRAD [35], PVG [6], Street Gaussians [43] and OmniRe [7]. Note that all methods are designed for dynamic scenes and use lidar data for supervision. We use NeuRAD's official implementation neurad-studio [34] and their version of UniSim, as well as OmniRe's official implementation drivestudio [8] and their implementations of SVG and Street Gaussians. We use drivestudio to generate the required sky masks for Street Gaussians and OmniRe, as well as extract human pose estimates, which takes about 10 and 30 minutes per sequence on an A100." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 470, + 297, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 470, + 297, + 649 + ], + "spans": [ + { + "bbox": [ + 55, + 470, + 297, + 649 + ], + "type": "text", + "content": "Although point cloud rendering is shown in [7], there exists no feature for it in drivestudio. To render point clouds in training and validation views, we render depth images by placing 6 virtual cameras in the lidar origin, each with a horizontal FOV of " + }, + { + "bbox": [ + 55, + 470, + 297, + 649 + ], + "type": "inline_equation", + "content": "60^{\\circ}" + }, + { + "bbox": [ + 55, + 470, + 297, + 649 + ], + "type": "text", + "content": ". Their vertical FOV is set to capture all lidar points in the sweep, and the focal length is the median focal length for each dataset. We project ground truth lidar points into these images, fetch each point's depth, and project them back into 3D. As drivestudio lacks functionality for inferring missing lidar points, these baselines are evaluated only for existing points. In contrast, SplatAD is evaluated on the more challenging task of predicting the range also for missing rays and must filter them based on the predicted ray drop probability. See Appendix C for further baseline details." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 659, + 157, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 659, + 157, + 672 + ], + "spans": [ + { + "bbox": [ + 55, + 659, + 157, + 672 + ], + "type": "text", + "content": "4.1. Image rendering" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 677, + 297, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 297, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 297, + 715 + ], + "type": "text", + "content": "Novel view synthesis: We assess the rendering quality of images using the standard NVS metrics PSNR, SSIM [37], and LPIPS [48] on hold-out validation images in Tab. 1." + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 316, + 71, + 554, + 200 + ], + "blocks": [ + { + "bbox": [ + 316, + 71, + 554, + 200 + ], + "lines": [ + { + "bbox": [ + 316, + 71, + 554, + 200 + ], + "spans": [ + { + "bbox": [ + 316, + 71, + 554, + 200 + ], + "type": "image", + "image_path": "adcb9d405382fb770cf80cf64294c796be9a07363c0de4e12795b9de42d707f8.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 209, + 555, + 232 + ], + "lines": [ + { + "bbox": [ + 313, + 209, + 555, + 232 + ], + "spans": [ + { + "bbox": [ + 313, + 209, + 555, + 232 + ], + "type": "text", + "content": "Figure 4. Removing our rolling shutter modeling compensation leads to inaccurate geometries and inconsistencies in the learning." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 255, + 555, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 255, + 555, + 338 + ], + "spans": [ + { + "bbox": [ + 313, + 255, + 555, + 338 + ], + "type": "text", + "content": "We measure speed using resolution-agnostic megapixels per second. SplatAD consistently outperforms existing 3DGS methods in terms of image quality with a large margin for the three considered datasets. Further, our method improves upon NeuRAD's SOTA results in all NVS metrics, while rendering images an order of magnitude faster. We provide a qualitative comparison in Fig. 3." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 339, + 556, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 339, + 556, + 482 + ], + "spans": [ + { + "bbox": [ + 313, + 339, + 556, + 482 + ], + "type": "text", + "content": "Reconstruction: To measure the upper limit of the modeling capacity of the methods, we report reconstruction metrics on PandaSet in Tab. 3. Here, methods train on all data in a sequence and evaluate on the same views. Further, we enable sensor pose optimization for all methods to account for any pose inaccuracies. Comparing to the NVS setting, all methods except UniSim show improvements with the additional training data. SplatAD achieves SOTA results while rendering " + }, + { + "bbox": [ + 313, + 339, + 556, + 482 + ], + "type": "inline_equation", + "content": "\\times 10" + }, + { + "bbox": [ + 313, + 339, + 556, + 482 + ], + "type": "text", + "content": " faster than the previous best method. Further, we note that SplatAD's performance on hold-out validation images in Tab. 1 is on par with previous 3DGS methods' results for reconstruction in Tab. 3." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 483, + 557, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 483, + 557, + 639 + ], + "spans": [ + { + "bbox": [ + 313, + 483, + 557, + 639 + ], + "type": "text", + "content": "Extrapolation: We investigate SplatAD's ability to generalize to views that differ significantly from what is encountered during training. We use the models trained on every other frame and follow the three settings in [35]: shift the ego-vehicle horizontally, shift the ego-vehicle vertically, and apply shifts and rotation to all dynamic actors. We report the Fréchet distance using DINOv2 [28] features, as these have been shown [32] to align better with human perception than using Inception-v3 [33]. However, we note that using Inception-v3 features instead does not change the model ranking or our conclusions. Tab. 4 shows SplatAD's ability to learn meaningful representations for generalization, clearly outperforming other 3DGS methods." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 647, + 413, + 660 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 647, + 413, + 660 + ], + "spans": [ + { + "bbox": [ + 313, + 647, + 413, + 660 + ], + "type": "text", + "content": "4.2. Lidar rendering" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 665, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 556, + 713 + ], + "type": "text", + "content": "We measure the quality of our lidar point clouds using the same metrics as in [35], i.e., median squared depth error, RMSE intensity error, ray drop accuracy, and chamfer distance, see Tab. 2 for NVS and Tab. 3 for reconstruction. Note" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 750, + 318, + 760 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 750, + 318, + 760 + ], + "spans": [ + { + "bbox": [ + 294, + 750, + 318, + 760 + ], + "type": "text", + "content": "11988" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 94, + 92, + 515, + 166 + ], + "blocks": [ + { + "bbox": [ + 74, + 71, + 534, + 84 + ], + "lines": [ + { + "bbox": [ + 74, + 71, + 534, + 84 + ], + "spans": [ + { + "bbox": [ + 74, + 71, + 534, + 84 + ], + "type": "text", + "content": "Table 3. Reconstruction results for image and lidar point clouds on PandaSet. " + }, + { + "bbox": [ + 74, + 71, + 534, + 84 + ], + "type": "inline_equation", + "content": "\\S" + }, + { + "bbox": [ + 74, + 71, + 534, + 84 + ], + "type": "text", + "content": " without missing points. First, second, third." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 94, + 92, + 515, + 166 + ], + "lines": [ + { + "bbox": [ + 94, + 92, + 515, + 166 + ], + "spans": [ + { + "bbox": [ + 94, + 92, + 515, + 166 + ], + "type": "table", + "html": "
ImageLidarEfficiency
PSNR ↑SSIM ↑LPIPS ↓Depth ↓Intensity ↓Drop acc. ↑CD ↓Camera MP/s ↑Lidar MR/s ↑Train time (H) ↓
PandasetUniSim23.180.6840.3620.060.087-10.2$11.10.80.8
NeuRAD26.760.7780.2330.030.07496.11.79.81.11.3
PVG25.350.7520.42353.18--150.8$23.60.31.6
Street-GS26.680.8080.2933.27--27.9$65.70.82.2
OmniRE26.680.8080.2953.80--33.5$56.70.62.4
SplatAD (ours)29.740.8930.1750.020.05397.41.6104.117.41.0
", + "image_path": "28b59ed19280b1cf5b9c3ec962a2735ac443decbb16912dccce2450a83f0a4f1.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 74, + 216, + 276, + 275 + ], + "blocks": [ + { + "bbox": [ + 55, + 184, + 296, + 208 + ], + "lines": [ + { + "bbox": [ + 55, + 184, + 296, + 208 + ], + "spans": [ + { + "bbox": [ + 55, + 184, + 296, + 208 + ], + "type": "text", + "content": "Table 4. " + }, + { + "bbox": [ + 55, + 184, + 296, + 208 + ], + "type": "inline_equation", + "content": "{\\mathrm{{FD}}}_{\\mathrm{{DINOV}}2}" + }, + { + "bbox": [ + 55, + 184, + 296, + 208 + ], + "type": "text", + "content": " scores when shifting pose of ego vehicle or actors. First,second,third ." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 74, + 216, + 276, + 275 + ], + "lines": [ + { + "bbox": [ + 74, + 216, + 276, + 275 + ], + "spans": [ + { + "bbox": [ + 74, + 216, + 276, + 275 + ], + "type": "table", + "html": "
Ego lane shiftEgo vert. shiftActor shift
0m2m3m1mRot.Trans.
PandayetNeuRAD377.4540.8629.2571.0424.8420.2
PVG618.8782.7885.6913.9618.8618.8
Street-GS497.1718.6824.2868.7554.7558.0
OmniRE504.5723.2833.5872.8569.6573.2
SplatAD (ours)276.3520.7647.2613.2355.1345.1
", + "image_path": "204cac16f9789402aaf4ce5f7336e69f24ec5ee6c401880234bcec5953b6f355.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 58, + 312, + 293, + 407 + ], + "blocks": [ + { + "bbox": [ + 55, + 287, + 296, + 309 + ], + "lines": [ + { + "bbox": [ + 55, + 287, + 296, + 309 + ], + "spans": [ + { + "bbox": [ + 55, + 287, + 296, + 309 + ], + "type": "text", + "content": "Table 5. NVS results averaged over 10 sequences from PandaSet, nuScenes, and Argoverse2 when removing model components." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 312, + 293, + 407 + ], + "lines": [ + { + "bbox": [ + 58, + 312, + 293, + 407 + ], + "spans": [ + { + "bbox": [ + 58, + 312, + 293, + 407 + ], + "type": "table", + "html": "
Full modelPSNR ↑SSIM ↑LPIPS ↓Depth ↓CD ↓Camera MP/s ↑Lidar MR/s ↑
27.540.8300.2550.022.0114.111.4
a)Camera RS27.480.8280.2580.022.1116.311.5
b)CNN decoder27.230.8240.2640.022.2178.011.6
c)EWA antialias27.600.8280.2680.021.5114.511.0
d)Appearance emb.27.280.8240.2640.022.1121.011.5
e)Velocity opt.27.550.8300.2570.022.0114.111.4
f)Lidar RS27.420.8270.2600.032.1114.212.2
g)MCMC27.350.8260.2590.012.1130.614.1
h)Median depth27.540.8300.2550.024.9114.111.4
", + "image_path": "50a4eaa74700d6d0ce39347c8ae2473f1f3871726edf7e530b98079972d32cd2.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 57, + 413, + 157, + 479 + ], + "blocks": [ + { + "bbox": [ + 57, + 413, + 157, + 479 + ], + "lines": [ + { + "bbox": [ + 57, + 413, + 157, + 479 + ], + "spans": [ + { + "bbox": [ + 57, + 413, + 157, + 479 + ], + "type": "image", + "image_path": "c448484c33faf286ff729a8f41f4439392ed19dd276f7245c1e5eb3ce057bbfd.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 489, + 295, + 510 + ], + "lines": [ + { + "bbox": [ + 55, + 489, + 295, + 510 + ], + "spans": [ + { + "bbox": [ + 55, + 489, + 295, + 510 + ], + "type": "text", + "content": "Figure 5. The CNN decoder improves sharpness and is more true to color than the MLP decoder." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 157, + 413, + 294, + 479 + ], + "blocks": [ + { + "bbox": [ + 157, + 413, + 294, + 479 + ], + "lines": [ + { + "bbox": [ + 157, + 413, + 294, + 479 + ], + "spans": [ + { + "bbox": [ + 157, + 413, + 294, + 479 + ], + "type": "image", + "image_path": "a30645caa572704cc825a30a068dd3d158e63ef9efbdf29017da8cae571f8b09.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 517, + 296, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 517, + 296, + 649 + ], + "spans": [ + { + "bbox": [ + 55, + 517, + 296, + 649 + ], + "type": "text", + "content": "that we do not report intensity or ray drop accuracy for the 3DGS baselines, as they have no way of inferring these. Further, it was not possible to produce Argoverse2 point cloud results for these baselines due to the memory requirements of drivestudio not scaling well with the increase in number of images and higher image resolution. We find SplatAD to be on par with the accurate ray-tracing formulation of NeuRAD in terms of point cloud quality, while running up to " + }, + { + "bbox": [ + 55, + 517, + 296, + 649 + ], + "type": "inline_equation", + "content": "\\times 18" + }, + { + "bbox": [ + 55, + 517, + 296, + 649 + ], + "type": "text", + "content": " faster. Furthermore, our lidar rendering approach is superior to the naive depth image-based method used for 3DGS baselines, both in terms of speed and quality." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 659, + 124, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 659, + 124, + 670 + ], + "spans": [ + { + "bbox": [ + 55, + 659, + 124, + 670 + ], + "type": "text", + "content": "4.3. Ablations" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "We validate the effectiveness of key components of our method by measuring their impact on NVS metrics in Tab. 5. Metrics are averaged over all three datasets to avoid any" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 186, + 555, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 186, + 555, + 342 + ], + "spans": [ + { + "bbox": [ + 313, + 186, + 555, + 342 + ], + "type": "text", + "content": "dataset-specific biases. We see that removing the rolling shutter modeling of the camera (a) or lidar (f) reduces NVS performance, and that leaving out optimization of sensor velocities (e) negatively impacts our depth predictions. While the impact of rolling shutter modeling on quantitative metrics is somewhat modest, we qualitatively find that it is important for capturing thin structures and edges, see Fig. 4. Removing these components leads to line-of-sight errors, which erroneously can cut through objects. Note that when removing our lidar rolling shutter compensation, we use ego-motion compensated point clouds for supervision. Moreover, we note the efficiency of our rolling shutter algorithms, as removing them barely improves rendering speeds." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 348, + 556, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 348, + 556, + 456 + ], + "spans": [ + { + "bbox": [ + 313, + 348, + 556, + 456 + ], + "type": "text", + "content": "We also find that using a CNN instead of an MLP for image decoding (c) clearly improves image realism, but at the cost of runtime, see Fig. 5 for an example. Further, we verify that using median depth instead of expected depth during inference is important for realistic lidar data in terms of Chamfer distance (h). Last, we note that MCMC [16] and EWA antialIASing [47, 52] both improve our performance, with the antialIASing having the largest impact on perceptual quality in terms of LPIPS." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 487, + 388, + 500 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 487, + 388, + 500 + ], + "spans": [ + { + "bbox": [ + 313, + 487, + 388, + 500 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 515, + 555, + 611 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 515, + 555, + 611 + ], + "spans": [ + { + "bbox": [ + 313, + 515, + 555, + 611 + ], + "type": "text", + "content": "In this work, we have proposed SplatAD, the first method for rendering camera and lidar data from 3D Gaussians. We combined accurate sensor modeling with efficient algorithms to achieve state-of-the-art novel view synthesis results while improving run-time by more than an order of magnitude. We hope that SplatAD will advance the scalability and fidelity of simulation environments, facilitating safer and more cost-effective testing of AD systems." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 617, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 556, + 713 + ], + "type": "text", + "content": "Limitations and future work: SplatAD is currently limited to modeling all dynamic actors as rigid. Drawing inspiration from recent advances in human reconstruction [18, 20, 26] can provide inspiration how to overcome this limitation in future research. In addition, we believe that incorporating data-driven priors such as diffusion models [9, 11, 40] is a promising direction to improve the extrapolation performance, and for changing appearance attributes." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 750, + 317, + 759 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 750, + 317, + 759 + ], + "spans": [ + { + "bbox": [ + 294, + 750, + 317, + 759 + ], + "type": "text", + "content": "11989" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 149, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 149, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 149, + 85 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 88, + 297, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 88, + 297, + 167 + ], + "spans": [ + { + "bbox": [ + 55, + 88, + 297, + 167 + ], + "type": "text", + "content": "We thank Adam Lilja, William Ljungbergh, and Adam Tenderski for valuable feedback. This work was partially supported by the Wallenberg AI, Autonomous Systems and Software Program (WASP) funded by the Knut and Alice Wallenberg Foundation. Computational resources were provided by NAISS at NSC Berzelius, partially funded by the Swedish Research Council, grant agreement no. 2022-06725." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 186, + 115, + 198 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 186, + 115, + 198 + ], + "spans": [ + { + "bbox": [ + 56, + 186, + 115, + 198 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 205, + 297, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 205, + 297, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 205, + 297, + 272 + ], + "spans": [ + { + "bbox": [ + 61, + 205, + 297, + 272 + ], + "type": "text", + "content": "[1] Mina Alibeigi, William Ljungbergh, Adam Tonderski, Georg Hess, Adam Lilja, Carl Lindstrom, Daria Motorniuk, Junsheng Fu, Jenny Widahl, and Christoffer Petersson. Zenseact open dataset: A large-scale and diverse multimodal dataset for autonomous driving. In ICCV, pages 20178-20188, 2023. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 272, + 296, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 272, + 296, + 316 + ], + "spans": [ + { + "bbox": [ + 61, + 272, + 296, + 316 + ], + "type": "text", + "content": "[2] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In ICCV, pages 5855-5864, 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 317, + 296, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 317, + 296, + 360 + ], + "spans": [ + { + "bbox": [ + 62, + 317, + 296, + 360 + ], + "type": "text", + "content": "[3] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In CVPR, pages 5470–5479, 2022." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 361, + 296, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 361, + 296, + 403 + ], + "spans": [ + { + "bbox": [ + 62, + 361, + 296, + 403 + ], + "type": "text", + "content": "[4] Jonathan T. Barron, Ben Mildenhall, Dor Verbin, Pratul P. Srinivasan, and Peter Hedman. Zip-nerf: Anti-aliased grid-based neural radiance fields. In ICCV, pages 19697-19705, 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 405, + 296, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 405, + 296, + 459 + ], + "spans": [ + { + "bbox": [ + 62, + 405, + 296, + 459 + ], + "type": "text", + "content": "[5] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In CVPR, pages 11621-11631, 2020. 1, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 460, + 296, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 460, + 296, + 503 + ], + "spans": [ + { + "bbox": [ + 62, + 460, + 296, + 503 + ], + "type": "text", + "content": "[6] Yurui Chen, Chun Gu, Junzhe Jiang, Xiatian Zhu, and Li Zhang. Periodic vibration gaussian: Dynamic urban scene reconstruction and real-time rendering. arXiv preprint arXiv:2311.18561, 2023. 2, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 504, + 296, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 504, + 296, + 557 + ], + "spans": [ + { + "bbox": [ + 62, + 504, + 296, + 557 + ], + "type": "text", + "content": "[7] Ziyu Chen, Jiawei Yang, Jiahui Huang, Riccardo de Lutio, Janick Martinez Esturo, Boris Ivanovic, Or Litany, Zan Gojcic, Sanja Fidler, Marco Pavone, et al. Omnire: Omni urban scene reconstruction. arXiv preprint arXiv:2408.16760, 2024. 1, 2, 5, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 559, + 296, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 559, + 296, + 613 + ], + "spans": [ + { + "bbox": [ + 62, + 559, + 296, + 613 + ], + "type": "text", + "content": "[8] Ziyu Chen, Jiawei Yang, Jiahui Huang, Riccardo de Lutio, Janick Martinez Esturo, Boris Ivanovic, Or Litany, Zan Gojcic, Sanja Fidler, Marco Pavone, Li Song, and Yue Wang. *drivestudio*. https://github.com/ziyc/drivestudio, 2024. 7, 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 614, + 296, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 614, + 296, + 668 + ], + "spans": [ + { + "bbox": [ + 62, + 614, + 296, + 668 + ], + "type": "text", + "content": "[9] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In ICML, 2024. 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 670, + 296, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 296, + 714 + ], + "type": "text", + "content": "[10] Xiao Fu, Shangzhan Zhang, Tianrun Chen, Yichong Lu, Lanyun Zhu, Xiaowei Zhou, Andreas Geiger, and Yiyi Liao. Panoptic nerf: 3d-to-2d label transfer for panoptic urban scene segmentation. In 3DV, pages 1-11. IEEE, 2022. 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 714 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 316, + 73, + 555, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 555, + 127 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 555, + 127 + ], + "type": "text", + "content": "[11] Ruiqi Gao, Aleksander Holynski, Philipp Henzler, Arthur Brussee, Ricardo Martin Brualla, Pratul P. Srinivasan, Jonathan T. Barron, and Ben Poole. CAT3d: Create anything in 3d with multi-view diffusion models. In NeurIPS, 2024. 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 129, + 555, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 555, + 172 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 555, + 172 + ], + "type": "text", + "content": "[12] Shubham Goel, Georgios Pavlakos, Jathushan Rajasegaran, Angjoo Kanazawa, and Jitendra Malik. Humans in 4d: Reconstructing and tracking humans with transformers. In ICCV, pages 14783-14794, 2023. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 173, + 555, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 173, + 555, + 215 + ], + "spans": [ + { + "bbox": [ + 316, + 173, + 555, + 215 + ], + "type": "text", + "content": "[13] Shengyu Huang, Zan Gojcic, Zian Wang, Francis Williams, Yoni Kasten, Sanja Fidler, Konrad Schindler, and Or Litany. Neural lidar fields for novel view synthesis. In ICCV, pages 18236-18246, 2023. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 217, + 554, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 217, + 554, + 250 + ], + "spans": [ + { + "bbox": [ + 316, + 217, + 554, + 250 + ], + "type": "text", + "content": "[14] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. TOG, 42(4):139-1, 2023. 1, 2, 3, 4, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 251, + 555, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 251, + 555, + 293 + ], + "spans": [ + { + "bbox": [ + 316, + 251, + 555, + 293 + ], + "type": "text", + "content": "[15] Mustafa Khan, Hamidreza Fazlali, Dhruv Sharma, Tongtong Cao, Dongfeng Bai, Yuan Ren, and Bingbing Liu. Autosplat: Constrained gaussian splatting for autonomous driving scene reconstruction, 2024. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 294, + 554, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 294, + 554, + 338 + ], + "spans": [ + { + "bbox": [ + 316, + 294, + 554, + 338 + ], + "type": "text", + "content": "[16] Shakiba Kheradmand, Daniel Rebain, Gopal Sharma, Weiwei Sun, Yang-Che Tseng, Hossam Isack, Abhishek Kar, Andrea Tagliasacchi, and Kwang Moo Yi. 3d gaussian splatting as markov chain monte carlo. In NeurIPS, 2024. 5, 6, 8, 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 338, + 554, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 338, + 554, + 360 + ], + "spans": [ + { + "bbox": [ + 316, + 338, + 554, + 360 + ], + "type": "text", + "content": "[17] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR, 2015. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 361, + 554, + 393 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 361, + 554, + 393 + ], + "spans": [ + { + "bbox": [ + 316, + 361, + 554, + 393 + ], + "type": "text", + "content": "[18] Muhammed Kocabas, Jen-Hao Rick Chang, James Gabriel, Oncel Tuzel, and Anurag Ranjan. Hugs: Human gaussian splats. In CVPR, pages 505-515, 2024. 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 394, + 555, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 394, + 555, + 449 + ], + "spans": [ + { + "bbox": [ + 316, + 394, + 555, + 449 + ], + "type": "text", + "content": "[19] Abhijit Kundu, Kyle Genova, Xiaqi Yin, Alireza Fathi, Caroline Pantofaru, Leonidas J Guibas, Andrea Tagliasacchi, Frank Dellaert, and Thomas Funkhouser. Panoptic neural fields: A semantic object-aware neural scene representation. In CVPR, pages 12871-12881, 2022. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 449, + 555, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 449, + 555, + 514 + ], + "spans": [ + { + "bbox": [ + 316, + 449, + 555, + 514 + ], + "type": "text", + "content": "[20] Youngjoong Kwon, Baole Fang, Yixing Lu, Haoye Dong, Cheng Zhang, Francisco Vicente Carrasco, Albert Mosella-Montoro, Jianjin Xu, Shingo Takagi, Daeil Kim, Aayush Prakash, and Fernando De la Torre. Generalizable human gaussians for sparse view synthesis. In ECCV, pages 451-468, Cham, 2025. Springer Nature Switzerland. 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 515, + 555, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 515, + 555, + 568 + ], + "spans": [ + { + "bbox": [ + 316, + 515, + 555, + 568 + ], + "type": "text", + "content": "[21] Nanxi Li, Chong Pei Ho, Jin Xue, Leh Woon Lim, Guanyu Chen, Yuan Hsing Fu, and Lennon Yao Ting Lee. A progress review on solid-state lidar and nanophotonics-based lidar sensors. *Laser & Photonics Reviews*, 16(11):2100511, 2022. 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 570, + 555, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 570, + 555, + 624 + ], + "spans": [ + { + "bbox": [ + 316, + 570, + 555, + 624 + ], + "type": "text", + "content": "[22] Carl Lindström, Georg Hess, Adam Lilja, Maryam Fatemi, Lars Hammarstrand, Christoffer Petersson, and Lennart Svensson. Are nerfs ready for autonomous driving? towards closing the real-to-simulation gap. In CVPRW, pages 4461-4471, 2024. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 625, + 555, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 625, + 555, + 680 + ], + "spans": [ + { + "bbox": [ + 316, + 625, + 555, + 680 + ], + "type": "text", + "content": "[23] William Ljungbergh, Adam Tonderski, Joakim Johnander, Holger Caesar, Kalle Åström, Michael Felsberg, and Christoffer Petersson. Neuroncap: Photorealistic closed-loop safety testing for autonomous driving. In ECCV, pages 161-177. Springer, 2025. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 681, + 555, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 555, + 714 + ], + "type": "text", + "content": "[24] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J. Black. SMPL: A skinned multiperson linear model. TOG, 34(6):248:1-248:16, 2015. 2" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 750, + 318, + 760 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 750, + 318, + 760 + ], + "spans": [ + { + "bbox": [ + 293, + 750, + 318, + 760 + ], + "type": "text", + "content": "11990" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 297, + 128 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 297, + 128 + ], + "type": "text", + "content": "[25] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, pages 405-421, Cham, 2020. Springer International Publishing. 1, 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 130, + 295, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 130, + 295, + 174 + ], + "spans": [ + { + "bbox": [ + 56, + 130, + 295, + 174 + ], + "type": "text", + "content": "[26] Arthur Moreau, Jifei Song, Helisa Dhamo, Richard Shaw, Yiren Zhou, and Eduardo Pérez-Pellitero. Human gaussian splatting: Real-time rendering of animatable avatars. In CVPR, pages 788-798, 2024. 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 176, + 296, + 209 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 176, + 296, + 209 + ], + "spans": [ + { + "bbox": [ + 56, + 176, + 296, + 209 + ], + "type": "text", + "content": "[27] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. TOG, 41(4):1-15, 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 211, + 296, + 310 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 211, + 296, + 310 + ], + "spans": [ + { + "bbox": [ + 56, + 211, + 296, + 310 + ], + "type": "text", + "content": "[28] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel HAZIZA, Francisco Massa, Alaaeldin El-Nouby, Mido Assran, Nicolas Ballas, Wojciech Galuba, Russell Howes, Po-Yao Huang, Shang-Wen Li, Ishan Misra, Michael Rabbat, Vasu Sharma, Gabriel Synnaeve, Hu Xu, Herve Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. DINov2: Learning robust visual features without supervision. TMLR, 2024. 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 312, + 295, + 345 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 312, + 295, + 345 + ], + "spans": [ + { + "bbox": [ + 56, + 312, + 295, + 345 + ], + "type": "text", + "content": "[29] Julian Ost, Fahim Mannan, Nils Thuerey, Julian Knodt, and Felix Heide. Neural scene graphs for dynamic scenes. In CVPR, pages 2856-2865, 2021. 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 347, + 295, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 347, + 295, + 390 + ], + "spans": [ + { + "bbox": [ + 56, + 347, + 295, + 390 + ], + "type": "text", + "content": "[30] Konstantinos Rematas, Andrew Liu, Pratul P Srinivasan, Jonathan T Barron, Andrea Tagliasacchi, Thomas Funkhouser, and Vittorio Ferrari. Urban radiance fields. In CVPR, pages 12932-12942, 2022. 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 393, + 295, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 393, + 295, + 458 + ], + "spans": [ + { + "bbox": [ + 56, + 393, + 295, + 458 + ], + "type": "text", + "content": "[31] Otto Seiskari, Jerry Yilammi, Valtteri Kaatrasalo, Pekka Rantalankila, Matias Turkulainen, Juho Kannala, Esa Rahtu, and Arno Solin. Gaussian splattering on the move: Blur and rolling shutter compensation for natural camera motion. In ECCV, pages 160-177, Cham, 2025. Springer Nature Switzerland. 4, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 460, + 295, + 515 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 460, + 295, + 515 + ], + "spans": [ + { + "bbox": [ + 56, + 460, + 295, + 515 + ], + "type": "text", + "content": "[32] George Stein, Jesse Cresswell, Rasa Hosseinzadeh, Yi Sui, Brendan Ross, Valentin Villecloze, Zhaoyan Liu, Anthony L Caterini, Eric Taylor, and Gabriel Loaiza-Ganem. Exposing flaws of generative model evaluation metrics and their unfair treatment of diffusion models. NeurIPS, 36, 2024. 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 518, + 295, + 561 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 518, + 295, + 561 + ], + "spans": [ + { + "bbox": [ + 56, + 518, + 295, + 561 + ], + "type": "text", + "content": "[33] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. In CVPR, pages 2818-2826, 2016. 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 564, + 295, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 564, + 295, + 597 + ], + "spans": [ + { + "bbox": [ + 56, + 564, + 295, + 597 + ], + "type": "text", + "content": "[34] Adam Tondersski, Carl Lindstrom, Georg Hess, and William Ljungbergh. neurad-studio. https://github.com/georghess/neurad-studio, 2024. 6, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 599, + 295, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 599, + 295, + 643 + ], + "spans": [ + { + "bbox": [ + 56, + 599, + 295, + 643 + ], + "type": "text", + "content": "[35] Adam Tondersski, Carl Lindström, Georg Hess, William Ljungbergh, Lennart Svensson, and Christoffer Petersson. Neurad: Neural rendering for autonomous driving. In CVPR, pages 14895-14904, 2024. 1, 2, 3, 4, 6, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 646, + 295, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 646, + 295, + 678 + ], + "spans": [ + { + "bbox": [ + 56, + 646, + 295, + 678 + ], + "type": "text", + "content": "[36] Haithem Turki, Jason Y Zhang, Francesco Ferroni, and Deva Ramanan. Suds: Scalable urban dynamic scenes. In CVPR, pages 12375-12385, 2023. 1, 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 680, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 680, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 680, + 295, + 713 + ], + "type": "text", + "content": "[37] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE TIP, 13(4):600-612, 2004. 7" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 316, + 72, + 555, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 555, + 160 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 555, + 160 + ], + "type": "text", + "content": "[38] Benjamin Wilson, William Qi, Tanmay Agarwal, John Lambert, Jagjeet Singh, Siddhesh Khandelwal, Bowen Pan, Ratnesh Kumar, Andrew Hartnett, Jhony Kaesemodel Pontes, Deva Ramanan, Peter Carr, and James Hays. Argoverse 2: Next generation datasets for self-driving perception and forecasting. In Proceedings of the Neural Information Processing Systems Track on Datasets and Benchmarks (NeurIPS Datasets and Benchmarks 2021), 2021. 1, 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 163, + 555, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 555, + 205 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 555, + 205 + ], + "type": "text", + "content": "[39] Hanfeng Wu, Xingxing Zuo, Stefan Leutenegger, Or Litany, Konrad Schindler, and Shengyu Huang. Dynamic lidar resimulation using compositional neural fields. In CVPR, pages 1998-19998, 2024. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 207, + 555, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 207, + 555, + 261 + ], + "spans": [ + { + "bbox": [ + 316, + 207, + 555, + 261 + ], + "type": "text", + "content": "[40] Rundi Wu, Ben Mildenhall, Philipp Henzler, Keunhong Park, Ruiqi Gao, Daniel Watson, Pratul P Srinivasan, Dor Verbin, Jonathan T Barron, Ben Poole, et al. Reconfusion: 3d reconstruction with diffusion priors. In CVPR, pages 21551-21561, 2024. 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 263, + 554, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 263, + 554, + 329 + ], + "spans": [ + { + "bbox": [ + 316, + 263, + 554, + 329 + ], + "type": "text", + "content": "[41] Pengchuan Xiao, Zhenlei Shao, Steven Hao, Zishuo Zhang, Xiaolin Chai, Judy Jiao, Zesong Li, Jian Wu, Kai Sun, Kun Jiang, Yunlong Wang, and Diange Yang. Pandaset: Advanced sensor suite dataset for autonomous driving. In 2021 IEEE International Intelligent Transportation Systems Conference (ITSC), pages 3095-3101, 2021. 2, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 331, + 554, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 331, + 554, + 363 + ], + "spans": [ + { + "bbox": [ + 316, + 331, + 554, + 363 + ], + "type": "text", + "content": "[42] Ziyang Xie, Junge Zhang, Wenye Li, Feihu Zhang, and Li Zhang. S-neRF: Neural radiance fields for street views. In ICLR, 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 365, + 554, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 365, + 554, + 407 + ], + "spans": [ + { + "bbox": [ + 316, + 365, + 554, + 407 + ], + "type": "text", + "content": "[43] Yunzhi Yan, Haotong Lin, Chenxu Zhou, Weijie Wang, Haiyang Sun, Kun Zhan, Xianpeng Lang, Xiaowei Zhou, and Sida Peng. Street gaussians for modeling dynamic urban scenes. In ECCV, 2024. 1, 2, 3, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 410, + 555, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 410, + 555, + 463 + ], + "spans": [ + { + "bbox": [ + 316, + 410, + 555, + 463 + ], + "type": "text", + "content": "[44] Jiawei Yang, Boris Ivanovic, Or Litany, Xinshuo Weng, Seung Wook Kim, Boyi Li, Tong Che, Danfei Xu, Sanja Fidler, Marco Pavone, and Yue Wang. EmerneRF: Emergent spatial-temporal scene decomposition via self-supervision. In ICLR, 2024. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 465, + 555, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 465, + 555, + 509 + ], + "spans": [ + { + "bbox": [ + 316, + 465, + 555, + 509 + ], + "type": "text", + "content": "[45] Ze Yang, Yun Chen, Jingkang Wang, Sivabalan Manivasagam, Wei-Chiu Ma, Anqi Joyce Yang, and Raquel Urtasun. Unisim: A neural closed-loop sensor simulator. In CVPR, pages 1389-1399, 2023. 1, 2, 7, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 511, + 555, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 511, + 555, + 555 + ], + "spans": [ + { + "bbox": [ + 316, + 511, + 555, + 555 + ], + "type": "text", + "content": "[46] Vickie Ye, Ruilong Li, Justin Kerr, Matias Turkulainen, Brent Yi, Zhuoyang Pan, Otto Seiskari, Jianbo Ye, Jeffrey Hu, Matthew Tancik, et al. gsplat: An open-source library for gaussian splatting. arXiv preprint arXiv:2409.06765, 2024. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 556, + 555, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 556, + 555, + 589 + ], + "spans": [ + { + "bbox": [ + 316, + 556, + 555, + 589 + ], + "type": "text", + "content": "[47] Zehao Yu, Anpei Chen, Binbin Huang, Torsten Sattler, and Andreas Geiger. Mip-splatting: Alias-free 3d gaussian splatting. In CVPR, pages 19447-19456, 2024. 4, 5, 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 590, + 554, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 590, + 554, + 633 + ], + "spans": [ + { + "bbox": [ + 316, + 590, + 554, + 633 + ], + "type": "text", + "content": "[48] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, pages 586-595, 2018. 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 635, + 554, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 554, + 669 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 554, + 669 + ], + "type": "text", + "content": "[49] Zehan Zheng, Fan Lu, Weiyi Xue, Guang Chen, and Changjun Jiang. Lidar4d: Dynamic neural fields for novel space-time view lidar synthesis. In CVPR, pages 5145-5154, 2024. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 670, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 554, + 713 + ], + "type": "text", + "content": "[50] Hongyu Zhou, Jiahao Shao, Lu Xu, Dongfeng Bai, Weichao Qiu, Bingbing Liu, Yue Wang, Andreas Geiger, and Yiyi Liao. Hugs: Holistic urban 3d scene understanding via gaussian splatting. In CVPR, pages 21336-21345, 2024. 2" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 750, + 317, + 759 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 750, + 317, + 759 + ], + "spans": [ + { + "bbox": [ + 293, + 750, + 317, + 759 + ], + "type": "text", + "content": "11991" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 298, + 160 + ], + "type": "list", + "angle": 0, + "index": 2, + "blocks": [ + { + "bbox": [ + 56, + 73, + 298, + 118 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 298, + 118 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 298, + 118 + ], + "type": "text", + "content": "[51] Xiaoyu Zhou, Zhiwei Lin, Xiaojun Shan, Yongtao Wang, Deqing Sun, and Ming-Hsuan Yang. Drivinggaussian: Composite gaussian splatting for surrounding dynamic autonomous driving scenes. In CVPR, pages 21634-21643, 2024. 1, 2, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 297, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 297, + 160 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 297, + 160 + ], + "type": "text", + "content": "[52] Matthias Zwicker, Hanspeter Pfister, Jeroen Van Baar, and Markus Gross. Ewa volume splatting. In Proceedings Visualization, 2001. VIS'01., pages 29-538. IEEE, 2001. 4, 8" + } + ] + } + ], + "index": 1 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 750, + 318, + 760 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 750, + 318, + 760 + ], + "spans": [ + { + "bbox": [ + 294, + 750, + 318, + 760 + ], + "type": "text", + "content": "11992" + } + ] + } + ], + "index": 3 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/SplatFlow_ Multi-View Rectified Flow Model for 3D Gaussian Splatting Synthesis/a1da8541-9ef2-4901-8219-e22e7fdfc1fc_content_list.json b/2025/SplatFlow_ Multi-View Rectified Flow Model for 3D Gaussian Splatting Synthesis/a1da8541-9ef2-4901-8219-e22e7fdfc1fc_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..d1cff51d984db9119fc83736535c341529ed7630 --- /dev/null +++ b/2025/SplatFlow_ Multi-View Rectified Flow Model for 3D Gaussian Splatting Synthesis/a1da8541-9ef2-4901-8219-e22e7fdfc1fc_content_list.json @@ -0,0 +1,2010 @@ +[ + { + "type": "text", + "text": "SplatFlow: Multi-View Rectified Flow Model for 3D Gaussian Splitting Synthesis", + "text_level": 1, + "bbox": [ + 89, + 130, + 906, + 154 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/064a97fe37b4b7c8d294be09866d237d0ce22ebffe1b816eceade2541a0823e8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 210, + 178, + 781, + 248 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c818755f2e286d6b4dbfb9b84c92e072c6c88bdd7bd578f66dc09abf9fa4ede6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 96, + 267, + 235, + 309 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/376b9af1fd76759b9f5241ed8e9a204b443a0854db268c54225001534e0dabc2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 96, + 315, + 235, + 359 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f71e8b7573d64979aa4dc12bf96889f72cc6cdcc07a34be67142f9cc7f6309c9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 96, + 361, + 235, + 412 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6f8a8e4d6f4321e2852f254fd033c19e129b3e0dc1ee5d4d1cf706760fc01130.jpg", + "image_caption": [ + "A library with a lot of books" + ], + "image_footnote": [], + "bbox": [ + 96, + 412, + 235, + 450 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c2982e5c1a3b23e30470f0d3878d3c0257f7a8f3677789ec62e75728e02484d3.jpg", + "image_caption": [ + "Figure 1. SplatFlow for 3D Gaussian Splitting synthesis and its training-free applications. (a) Examples of direct 3D Gaussian Splitting (3DGS) generation only from text prompts, (b) Training-free applications, including 3DGS object editing, camera pose estimation, and novel view synthesis. SplatFlow seamlessly integrates these capabilities, showcasing its versatility in generating and editing complex 3D content." + ], + "image_footnote": [], + "bbox": [ + 96, + 459, + 235, + 503 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e3425c2d94e5378e44eb4bda7e499b7e0ba88b81f87061a5fcbef324ece2f5b8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 238, + 267, + 375, + 309 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/da867f973dfe18f63cb0de40727f306d2b21a6448849fb83ba8b30c0e690529d.jpg", + "image_caption": [ + "A forest with tall trees and yellow leaves" + ], + "image_footnote": [], + "bbox": [ + 238, + 315, + 375, + 352 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f348c71eea6be672c093a5c9ce101abc5edca11b02984c37ee394ffbb9acf3f3.jpg", + "image_caption": [ + "Blue car on a clean street with buildings in the background." + ], + "image_footnote": [], + "bbox": [ + 238, + 361, + 375, + 398 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f426a58a2d365faca2fcd43e57b9f5d9b1aed98931a9fcc7389bd614e4c2a1f8.jpg", + "image_caption": [ + "Half avocado" + ], + "image_footnote": [], + "bbox": [ + 238, + 412, + 375, + 449 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a1cb932aff8f3b6017b9a97d752540c969d5c08fe28c81794f99bd4cddb64911.jpg", + "image_caption": [ + "slicedzucchini", + "(a) 3D Gaussian Splitting Synthesis" + ], + "image_footnote": [], + "bbox": [ + 238, + 459, + 375, + 497 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c89704bb8cca0486ed6f9faaca339c1fb0b7517139ae2d6969f9404c61cea928.jpg", + "image_caption": [ + "A lighthouse on a rocky shore" + ], + "image_footnote": [], + "bbox": [ + 380, + 267, + 650, + 367 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/902b10d04f2786fc83d5c96b0918ae15114974c46f0d7145a4e07a495a5be346.jpg", + "image_caption": [ + "A vibrant bush with pink flowers stands in front of a concrete wall, ..." + ], + "image_footnote": [], + "bbox": [ + 380, + 380, + 650, + 484 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1358ff03ebe3bed73134b5ab8e566f41f675fa0c589f08776e9995e744b68f35.jpg", + "image_caption": [ + "3DGS object editing (replacement)" + ], + "image_footnote": [], + "bbox": [ + 663, + 268, + 717, + 352 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6458c8efd497d95c7d4843e6f151d69dbf12531a2bebedaed111ced4fc55e817.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 720, + 268, + 803, + 352 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8fd0d6c19f961b167132882ee6fd41109fc3fb6900a2583bd0df23a82ea11a5d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 805, + 268, + 857, + 352 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ebbbe4a5a18b58e9bc493fce0d2fb3063d97c55375c072502ea8fcd57762a3c7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 862, + 268, + 887, + 352 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6b46d5028c10cd12cc29d8cf1ec9fc092f357a23c7ac29ef734984289c1bfb61.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 663, + 369, + 717, + 424 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3ce1b93006ec113de40bda8dec6865b2604a8b5e8d075c8b5f7c79b4dfcba96a.jpg", + "image_caption": [ + "\"Camera Pose Estimation\"" + ], + "image_footnote": [], + "bbox": [ + 718, + 369, + 808, + 424 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7ec74f2631bb999bf9412660f65b05af585fb47d14b19fb9f7898c598e11eff8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 813, + 369, + 880, + 422 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3b0f77f8518208b80ecf34b7c05a8d692ee584d5e26218b419695991d8736551.jpg", + "image_caption": [ + "(b) Training-Free Applications" + ], + "image_footnote": [], + "bbox": [ + 663, + 445, + 717, + 489 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/092483e1230096e6b5a42dc843f9cf6657e7d5b3334930482438f0117b60eeb0.jpg", + "image_caption": [ + "\"Novel View Synthesis\"" + ], + "image_footnote": [], + "bbox": [ + 720, + 445, + 803, + 489 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/44a82faf846b356c7b18da5c281a23891930af98e26b4577cc0fa179d9df8bdf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 805, + 445, + 885, + 489 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 590, + 326, + 606 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Text-based generation and editing of 3D scenes hold significant potential for streamlining content creation through intuitive user interactions. While recent advances leverage 3D Gaussian Splatting (3DGS) for high-fidelity and real-time rendering, existing methods are often specialized and task-focused, lacking a unified framework for both generation and editing. In this paper, we introduce SplatFlow, a comprehensive framework that addresses this gap by enabling direct 3DGS generation and editing. SplatFlow comprises two main components: a multi-view rectified flow (RF) model and a Gaussian Splatting Decoder (GSDecoder). The multi-view RF model operates in latent space, generating multi-view images, depths, and camera poses simultaneously, conditioned on text prompts—thus addressing challenges like diverse scene scales and complex camera trajectories in real-world settings. Then, the", + "bbox": [ + 88, + 622, + 485, + 864 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "GSDecoder efficiently translates these latent outputs into 3DGS representations through a feed-forward 3DGS method. Leveraging training-free inversion and inpainting techniques, SplatFlow enables seamless 3DGS editing and supports a broad range of 3D tasks—including object editing, novel view synthesis, and camera pose estimation—within a unified framework without requiring additional complex pipelines. We validate SplatFlow's capabilities on the MVImgNet and DL3DV-7K datasets, demonstrating its versatility and effectiveness in various 3D generation, editing, and inpainting-based tasks. Our project page is available at https://gohyojun15.github.io/SplatFlow/.", + "bbox": [ + 511, + 592, + 908, + 773 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 800, + 645, + 816 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The demand for realistic 3D scene generation and editing from text has surged in applications like VR/AR, gaming, and robotics. Early works primarily utilized NeRF [53, 55] for scene generation [100] and editing [2, 29, 30, 33, 63, 96], but NeRF's volumetric rendering is computationally", + "bbox": [ + 511, + 825, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution", + "bbox": [ + 107, + 876, + 215, + 888 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author", + "bbox": [ + 109, + 888, + 233, + 901 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "21524", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "intensive, resulting in slower rendering speeds. Recently, 3D Gaussian Splitting (3DGS) [32] has emerged as a promising alternative, offering real-time and high-fidelity rendering implementations. Building on this, recent advances in 3D generation [8, 34, 58, 81, 87, 88, 90, 92, 99, 102, 108] and editing [8, 34, 58, 87, 88, 90, 108] increasingly leverage 3DGS to achieve both speed and quality.", + "bbox": [ + 89, + 90, + 480, + 196 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Despite these advancements, current methods for generating and editing 3DGS are specialized and task-focused, lacking a unified framework. For 3DGS generation, several works [10, 42, 48, 82, 84, 94] leverage 2D diffusion models [26, 69, 76, 79] with Score Distillation Sampling (SDS) [66], which requires time-intensive per-scene optimization. To address this, recent studies have shifted towards direct 3DGS generation, combining diffusion models with reconstruction models [81, 92, 102] or utilizing 3D generative models [99]. However, most of these methods are restricted to object-level 3D generation using synthetic datasets [15, 16] with controlled, bounded environments. In contrast, real-world scenes, varying in scene scales and camera trajectories, present distinctive challenges.", + "bbox": [ + 89, + 199, + 485, + 411 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In 3DGS editing, several works [8, 34, 58, 87, 88, 90, 108] adopt 2D diffusion models [26, 69, 76, 79] to guide the editing process. The main challenge is lifting 2D editing guidance to 3D while maintaining 3D-consistent modifications. One approach focuses on SDS techniques to update 3DGS [58, 91, 108], but these methods require additional stages like texture adjustments [58, 91] and fine-grained refinements [108]. Another approach aims to edit multiview images with 3D consistency [34, 88, 90], introducing attention-based modules [88, 90] and autoregressive editing [34], adding complexity to the editing pipeline.", + "bbox": [ + 89, + 412, + 483, + 580 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we bridge the gap between 3DGS generation and editing by designing a direct 3DGS generative model. Inspired by the success of 2D diffusion models that enable training-free editing through inversion [13, 52, 54], we argue that a direct 3DGS model can similarly enable training-free editing. This approach integrates editing without requiring complex pipelines to lift 2D priors, making generation and editing adaptable within a unified framework.", + "bbox": [ + 89, + 580, + 483, + 702 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We introduce SplatFlow, which consists of two main components: (1) a multi-view rectified flow (RF) model and (2) a Gaussian Splitting Decoder (GSDecoder). Similar to Latent Diffusion Models [69], our multi-view RF model operates in the latent space and is trained to simultaneously generate multi-view images, depths, and camera poses conditioned on text prompts. This joint modeling enables SplatFlow to effectively handle the challenges of real-world scene generation, such as various scene scales and camera trajectories, unlike synthetic object datasets [15, 16]. The GSDecoder translates the latent outputs into 3DGS, based on efficient feed-forward methods [4, 9]. To ensure compatibility with other generative models, particularly Stable Diffusion 3 [19],", + "bbox": [ + 89, + 704, + 485, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "we incorporate a fixed pre-trained encoder, allowing flexible cross-model usage during sampling.", + "bbox": [ + 511, + 90, + 906, + 121 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Finally, SplatFlow utilizes training-free inversion [13, 52, 54] and inpainting techniques [11, 51, 70, 71, 77] for 3DGS editing and various 3D tasks. For example, it enables object editing [7, 58], novel view synthesis [61, 68, 97], and camera pose estimation [101] by jointly modeling images, depths, and poses to infer missing elements, highlighting its versatility. We validated its capabilities using the MVImgNet [98] and DL3DV-7K [45] datasets, assessing its performance in 3D generation, editing, and inpainting-based tasks.", + "bbox": [ + 511, + 122, + 908, + 258 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 511, + 271, + 663, + 287 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Diffusion Models and Rectified Flows", + "text_level": 1, + "bbox": [ + 511, + 296, + 836, + 311 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Diffusion Models Diffusion models [26, 76, 79] generate data through iterative denoising and have become standard in generative modeling across images [17, 21, 37, 60, 62], videos [27], and text [40]. Trained on large-scale text-to-image datasets, they capture complex semantic relationships, producing high-quality images [6, 56, 65, 69, 72]. An advantage is their adaptability to various tasks in a plug-and-play manner [20, 22, 78], especially inpainting [51, 70] and editing [23, 35]. In painting can be framed as a linear inverse problem [14], for which effective posterior sampling methods have been proposed [11, 51, 70, 71, 77]. For editing, inversion methods [13, 52, 54] produce structured noise by inverting images, allowing new prompts to be effectively processed. Building on these training-free techniques, we extend them to our 3DGS model, adapting RePaint [51] and SDEdit [52] within rectified flow models.", + "bbox": [ + 511, + 318, + 908, + 559 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Rectified Flow Models Rectified Flow (RF) models [1, 46, 47] represent an alternative approach to traditional diffusion models by utilizing a straight-line path from data to noise instead of the typical curved forward process, aiming to simplify sampling and reduce computational costs. This linear path theoretically enables faster sampling by minimizing error accumulation across steps, and recent advances, such as Stable Diffusion 3 (SD3) [19], have successfully scaled RF in text-to-image generation, demonstrating performance that surpasses traditional diffusion models. Our SplatFlow framework builds on the capabilities of SD3, integrating RF to achieve efficient 3DGS generation.", + "bbox": [ + 511, + 566, + 910, + 750 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. 3D Generative Models", + "text_level": 1, + "bbox": [ + 511, + 757, + 720, + 772 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Neural Scene Representation Various neural scene representations have been proposed for 3D generation tasks. Early works use explicit representations, enabling synthesis of 3D point clouds [57] and shapes [83, 106], but struggle with realistic scene generation due to their rendering mechanisms. To address this, NeRF [53] is employed for a realistic 3D generation with per-scene optimization [66], but its inefficient rendering is unsuitable for a fast generation.", + "bbox": [ + 511, + 780, + 908, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "21525", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recently, 3DGS [32] enables real-time rendering, accelerating optimization-based 3D generation [10, 44, 82]. Building on this progress, we leverage 3DGS's efficient rendering for fast training of a generative model on large datasets.", + "bbox": [ + 89, + 90, + 483, + 151 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Lifting Multi-View Generation Advances in image generation models [6, 65] and large-scale 3D object datasets [15, 16] make multi-view generation compelling compared to optimization-based generation [82]. The key idea is to learn 3D priors as conditions [49, 50, 89] or attentions [74, 85] from 3D datasets while leveraging text-to-image diffusion models [69] trained on billions of text-image pairs [73]. This enables the synthesis of multi-view consistent images from an image or text prompt, which is then fed into reconstruction models [5, 81] to directly generate neural scene representations like 3DGS. One critical success factor in these methods is that the training dataset provides rendered images from consistent, canonical viewpoints, allowing instant access to desired views regardless of the 3D object. In this work, we focus on more challenging unbounded 3D real-world scene generation, where each scene varies in scale and camera trajectory. To handle diverse trajectories, our model learns a joint distribution of camera poses and 3D scenes.", + "bbox": [ + 91, + 155, + 485, + 428 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Text-to-3D Scene Generation Generating 3D scenes from text prompts is challenging due to the unbounded scale of scenes, complicating scale determination. One straightforward approach is to create a scene covered by a single image [75, 107] generated from text-to-image diffusion models [6, 65]. Another line of work [12, 28] generates wider scenes using user-provided camera trajectories, but these methods are time-consuming and produce subpar quality due to reliance on 2D generative models. Recently, Director3D [41], the work most similar to ours, generates camera poses from text and then creates multi-view images decoded into 3DGS. In contrast, SplatFlow jointly learns the distribution of camera poses and multi-view images from text, enhancing generation quality and enabling flexible 3D scene editing in a unified framework.", + "bbox": [ + 89, + 431, + 485, + 657 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. 3D Editing", + "text_level": 1, + "bbox": [ + 89, + 666, + 212, + 683 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In recent years, NeRF-based methods [2, 29, 30, 33, 59, 63, 67, 95, 96] have gained popularity for 3D editing, surpassing point clouds and meshes [86, 103, 105]. However, due to the inefficient rendering, these methods are time-consuming and lack precise control. Conversely, incorporating 3DGS into 3D editing [8, 87] improves effectiveness, speed, and control.", + "bbox": [ + 89, + 688, + 483, + 794 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For guiding 3D editing, several methods have leveraged 2D diffusion models [26, 69, 76, 79] for structured 3D editing, but it often requires additional refinement stages. For instance, GSEdit [58] and Xiao et al. [91] apply texture enhancement, while TIP-Editor [108] adds fine-grained adjustments to reduce SDS artifacts. Another line of research in 3D editing emphasizes achieving multi-view consistency", + "bbox": [ + 89, + 795, + 485, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "by sequentially updating each view. For example, Gauss-Ctrl [90] and 3DEdit [88] utilize attention-based modules to propagate edits made from one perspective across multiple views, ensuring that changes are consistently reflected throughout the 3D scene. Similarly, 3D-Ego [34] adopts an autoregressive editing framework, updating each view to enhance multi-view consistency.", + "bbox": [ + 511, + 90, + 906, + 196 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Unlike prior methods that require complex procedures to lift 2D priors for 3D consistency, SplatFlow directly models multi-view consistency within the latent space, efficiently decoding edits through our Gaussian Splitting Decoder.", + "bbox": [ + 511, + 198, + 906, + 258 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Preliminary: Rectified Flows", + "text_level": 1, + "bbox": [ + 511, + 268, + 779, + 286 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Rectified Flows (RFs) [1, 46, 47] consider generative models that construct a noise distribution $q_{0}$ and a time-varying vector field $v_{t}(\\boldsymbol{x}_{t})$ to sample data $x_{0}$ from a data distribution $p_{0}$ via an ordinary differential equation (ODE) as:", + "bbox": [ + 511, + 294, + 906, + 354 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {d} X _ {t} = v \\left(X _ {t}\\right) \\mathrm {d} t, \\quad X _ {0} \\sim q _ {0}, \\quad t \\in [ 0, 1 ], \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 568, + 362, + 906, + 378 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $q_{0}$ typically follows a Gaussian distribution $\\mathcal{N}(0,I)$ . RF employs a neural network with parameters $\\theta$ for modeling the vector field as $v_{t}(X_{t}) = -u_{\\theta}(X_{t},1 - t)$ .", + "bbox": [ + 511, + 386, + 906, + 431 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To train a neural network as the vector field, RFs utilize paired samples from $p_0$ and $q_0$ (here simplified as $p_1$ ), along a straight line path $Y_t = tY_1 + (1 - t)Y_0$ . This setup produces the marginal distribution of $Y_t$ as $p_t(\\pmb{y}_t) = \\int p_t(\\pmb{y}_t|\\pmb{y}_1)p_1(\\pmb{y}_1)\\mathrm{d}\\pmb{y}_1$ . Starting from an initial $Y_0 = \\pmb{y}_0$ and moving to a final state $Y_1 = \\pmb{y}_1$ , a straight line path induces an ODE expressed as $\\mathrm{d}Y_t = u_t(Y_t|Y_1)\\mathrm{d}t$ where the vector field $u_t(Y_t|Y_1)$ represents the conditional change from $\\pmb{y}_0$ to $\\pmb{y}_1$ , given by $u_t(Y_t|Y_1) = \\pmb{y}_1 - \\pmb{y}_0$ . The marginal vector field $u_t(\\pmb{y}_t)$ is computed by averaging over the conditional vector field as [46]:", + "bbox": [ + 511, + 431, + 908, + 598 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nu _ {t} (\\boldsymbol {y} _ {t}) = \\int u _ {t} (\\boldsymbol {y} _ {t} | \\boldsymbol {y} _ {1}) \\frac {p _ {t} (\\boldsymbol {y} _ {t} | \\boldsymbol {y} _ {1})}{p _ {t} (\\boldsymbol {y} _ {t})} p _ {1} (\\boldsymbol {y} _ {1}) \\mathrm {d} \\boldsymbol {y} _ {1}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 560, + 604, + 906, + 638 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To approximate $u_{t}(\\boldsymbol{y}_{t})$ , the neural network is trained with a flow-matching object, defined by:", + "bbox": [ + 511, + 642, + 906, + 672 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {F M}} := \\mathbb {E} _ {t, Y _ {t}} \\left[ \\left| \\left| u _ {t} \\left(Y _ {t}\\right) - u _ {\\theta} \\left(Y _ {t}, t\\right) \\right| \\right| _ {2} ^ {2} \\right]. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 680, + 906, + 705 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Since directly computing this objective can be computationally challenging, an alternative approach, conditional flow matching, is used to simplify training:", + "bbox": [ + 511, + 710, + 906, + 757 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {C F M}} := \\mathbb {E} _ {t, Y _ {t}, Y _ {1} \\sim p _ {1}} \\left[ \\| u _ {t} \\left(Y _ {t} \\mid Y _ {1}\\right) - u _ {\\theta} \\left(Y _ {t}, t\\right) \\| _ {2} ^ {2} \\right]. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 535, + 763, + 906, + 789 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The gradients of $L_{\\mathrm{CFM}}$ and $L_{\\mathrm{FM}}$ are theoretically identical [46], making them equivalent in training. However, $L_{\\mathrm{CFM}}$ offers better computational efficiency and is therefore preferred. Ultimately, the learned vector field for the ODE in Eq. 1 is given by $v_{t}(X_{t}) = -u_{\\theta}(X_{t},1 - t)$ , allowing RFs to generate samples that follow the data distribution using the learned vector field.", + "bbox": [ + 511, + 794, + 908, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "21526", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/55dfe3ec3802dd8d637c1ff9e161605a4829c36665aaa8234007875857183f75.jpg", + "image_caption": [ + "Figure 2. Overview of SplatFlow. SplatFlow consists of two main components: a multi-view Rectified Flow (Section 4.3) model and a Gaussian Splat Decoder (Section 4.2). Conditioned on text, RF model generates multi-view latents—including image, depth, and Plücker ray coordinates. After an optimization process to estimate camera poses, the GSDecoder decodes these latents into pixel-aligned 3DGS." + ], + "image_footnote": [], + "bbox": [ + 112, + 88, + 880, + 311 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Methodology", + "text_level": 1, + "bbox": [ + 89, + 364, + 225, + 383 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Overview of SplatFlow", + "text_level": 1, + "bbox": [ + 89, + 390, + 302, + 406 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We introduce SplatFlow, a 3DGS generation model designed to perform various 3D editing and 3D-related tasks in a training-free manner, despite being trained solely for generation. As shown in Fig. 2, our model consists of two main components: 1) a multi-view Rectified Flow (RF) model and 2) a Gaussian Splitting Decoder (GSDecoder). First, the multi-view RF model generates multi-view camera poses, depths, and images conditioned on text prompts. Then, the GSDecoder translates these latent representations into pixel-aligned 3DGS, constructing a real-world scene structure.", + "bbox": [ + 89, + 412, + 483, + 564 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Also, both the multi-view RF model and GSDecoder operate within a latent space, adopting a Latent Diffusion Model (LDM) [69]. By freezing and sharing the encoder of Stable Diffusion 3 (SD3) [19], our design enables the RF model to share the latent space with SD3, allowing us to leverage its knowledge for enhanced generation and editing capabilities.", + "bbox": [ + 89, + 564, + 483, + 654 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Finally, our model leverages a training-free approach [51, 52], enabling a broad range of editing applications, such as object replacement, as well as 3D-related tasks including camera pose estimation and novel view synthesis.", + "bbox": [ + 89, + 655, + 483, + 715 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the following sections, we present each component in detail. Section 4.2 introduces the GSDecoder, explaining details in constructing pixel-aligned 3DGS. Section 4.3 describes the multi-view RF model, detailing how it jointly generates camera poses, depths, and images. Finally, Section 4.4 outlines the sampling process and demonstrates how our model applies training-free techniques.", + "bbox": [ + 89, + 715, + 483, + 821 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Gaussian Splitting Decoder (GSDecoder)", + "text_level": 1, + "bbox": [ + 89, + 832, + 444, + 849 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Recently, feed-forward 3DGS methods [4, 9, 80, 81] have enabled fast 3DGS reconstruction from sparse views by training on large datasets, achieving much faster reconstruction", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "than per-scene optimization methods [32]. Leveraging this advantage, our GSDecoder $G_{\\phi}$ (parameterized by $\\phi$ ) is designed to decode pixel-aligned 3DGS from latent representations of $K$ sparse views, given their corresponding camera poses $\\mathcal{P} = \\{\\pmb{P}_i\\}_{i=1}^K$ with $\\pmb{P}_i = \\pmb{K}_i[\\pmb{R}_i|\\pmb{T}_i]$ , where $\\pmb{K}_i, \\pmb{R}_i$ , and $\\pmb{T}_i$ denote the intrinsic matrix, rotation matrix, and translation vector of the $i$ -th view, respectively.", + "bbox": [ + 511, + 366, + 906, + 472 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "A straightforward approach is to design $G_{\\phi}$ to directly output the 3D Gaussian parameters from the image latents obtained through the encoder $\\mathcal{E}$ as $G_{\\phi}(\\{\\langle \\mathcal{E}(I_i), P_i \\rangle \\}_{i=1}^K) = \\{(\\pmb{\\mu}_j, \\alpha_j, \\pmb{\\Sigma}_i, c_j)\\}_{j=1}^{H \\times W \\times K}$ , where each 3D Gaussian parameter includes position $\\pmb{\\mu}_j$ , opacity $\\alpha_j$ , covariance $\\pmb{\\Sigma}_j$ , and color $c_j$ in a pixel-aligned manner with each $i$ -th image $I_i \\in \\mathbb{R}^{H \\times W}$ . However, as the image latents are produced through a shared and frozen encoder (to leverage compatibility with 2D generative models during sampling), it might lose fine-grained spatial details, leading to abstract representations. This can limit the preservation of specific 3D structural information. To address these limitations, we propose two improvements to enhance GSDecoder's performance.", + "bbox": [ + 511, + 473, + 908, + 670 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Depth latent integration As demonstrated in [31], the encoder $\\mathcal{E}$ can effectively encode depth maps. Building on this, we incorporate the depth latents $\\{\\mathcal{E}(D_i)\\}_{i=1}^K$ of each depth map $D_i$ as an additional input to the GSDecoder to further enhance 3D structural information. We use DepthAnythingV2 [93] to extract the depth maps and normalize them to the range [-1, 1], stacking them as three channels similar to RGB for constructing $D_i$ . We observe that this improves the convergence speed and performance of the GSDecoder.", + "bbox": [ + 511, + 679, + 908, + 816 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Adversarial loss Applying adversarial losses has demonstrated improvements in the visual quality of decoded output [18, 69]. However, using adversarial loss during training can cause instability, particularly in the early stages. We observe that applying the vision-aided loss [36] after the GSDe", + "bbox": [ + 511, + 824, + 908, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "21527", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "coder has reached a certain level of convergence significantly improves visual quality without destabilizing training.", + "bbox": [ + 89, + 90, + 482, + 122 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training In summary, our GSDecoder predicts 3DGS $\\{(\\pmb{\\mu}_j,\\alpha_j,\\pmb{\\Sigma}_j,\\pmb{c}_j)\\}_{j = 1}^{H\\times W\\times K}$ from $\\{(I_i,D_i,P_i)\\}_{i = 1}^K$ . The model is trained with a combined loss function consisting of LPIPS [104], Mean Squared Error loss, and vision-aided loss [36], computed between the target images of novel views and rendered views from 3DGS. Specifically, the vision-aided loss is activated only after a certain threshold iteration to improve stability and perceptual quality. Additionally, the architecture is initialized with the parameters of the Stable Diffusion 3 decoder, incorporating cross-view attention mechanisms, and the channel dimensions are increased to meet the specific requirements of 3D Gaussian Splatting. Further details on the GSDecoder are provided in Appendix A.", + "bbox": [ + 89, + 125, + 483, + 323 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Multi-View Rectified Flow Model", + "text_level": 1, + "bbox": [ + 89, + 330, + 382, + 345 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "With the GSDecoder in place, the 3D Gaussian Splitting (3DGS) generation process becomes feasible if we can produce consistent multi-view latents of images $\\mathcal{E}(I_i)$ , depth maps $\\mathcal{E}(D_i)$ , and camera poses $P_{i}$ for $i = 1,\\dots ,K$ . Our multi-view rectified flow (RF) model is trained to sample from the joint distribution of these latents, conditioned on text $C$ as $p(\\{I_i\\}_{i = 1}^K,\\{D_i\\}_{i = 1}^K,\\{P_i\\}_{i = 1}^K |C)$ , allowing for simultaneous generation of multi-view images, depths, and poses. The rationale for modeling a joint distribution instead of each component separately is twofold. First, this approach allows the model to handle various tasks by reformulating them as inpainting problems, as discussed in Section 4.4. Second, joint modeling is crucial for real-world scenes, which require adaptive camera poses tailored to each scene [41].", + "bbox": [ + 88, + 352, + 482, + 564 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To achieve this, our multi-view RF model treats the concatenation of multi-view image latents, depth latents, and Plücker ray coordinates [64, 101] along the channel dimension as input data. Formally, each ray $\\pmb{r}_i$ is represented as $\\langle \\pmb{d}_i, \\pmb{m}_i \\rangle$ , where $\\pmb{d}_i = \\pmb{R}_i^\\top \\pmb{K}_i^{-1} \\pmb{w}_i$ denotes the direction vector and $\\pmb{m}_i = (-\\pmb{R}_i^\\top \\pmb{T}_i) \\times \\pmb{d}_i$ represents the moment vector, with $\\pmb{w}_i$ referring to the 2D pixel coordinates. By matching the spatial resolution of $\\pmb{w}_i$ with the image and depth latents, each represented as $\\mathbb{R}^{n \\times h \\times w}$ , $\\pmb{r}_i$ can be expressed as $\\mathbb{R}^{6 \\times h \\times w}$ , enabling concatenation along the channel axis.", + "bbox": [ + 89, + 565, + 483, + 715 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Through this setup, we define $Y_{0}$ as the multi-view latents $\\mathbf{X}$ , where each element $\\mathbf{X}_i = \\langle \\mathcal{E}(\\mathbf{I}_i),\\mathcal{E}(\\mathbf{D}_i),\\mathbf{r}_i\\rangle$ results in $Y_{0} = (X_{1},X_{2},\\ldots ,X_{K})\\in \\mathbb{R}^{K\\times (2n + 6)\\times h\\times w}$ . We then train the multi-view RF model $u_{\\theta}$ with the conditional flow matching loss defined in Eq. 4. Specifically, we fine-tune Stable Diffusion 3 [19] by adjusting the channel dimensions in the input and output layers and incorporating cross-view attention. Further details are provided in Appendix A.", + "bbox": [ + 89, + 715, + 482, + 837 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Sampling process To recover the camera pose from ray $\\boldsymbol{r}_i$ , we follow the procedure in [101] and further refine the pose parameters to ensure that multiple views share the same intrinsic. However, we find that naively solving the ODE", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "can lead to degraded camera pose accuracy. We hypothesize this degradation results from deviations from the camera pose manifold, which introduces errors in camera trajectory estimation. To address this, at each sampling step $t = t_k$ , we predict the sampling destination at $t = 0$ , recover the camera poses from this prediction, and reconstruct the associated camera rays. These rays are then forwarded to $t = t_k$ to maintain the solution within the valid ray manifold at $t = t_k$ .", + "bbox": [ + 511, + 90, + 906, + 212 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Stable Diffusion 3 guidance To improve multi-view image generation quality, we integrate vector fields from Stable Diffusion 3, known for robust generalization, with our multiview RF model to solve the ODE for randomly selected view's image latents at specific sampling steps.", + "bbox": [ + 511, + 215, + 908, + 292 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.4. Inference Given Partial Data", + "text_level": 1, + "bbox": [ + 511, + 301, + 769, + 316 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we demonstrate how SplatFlow can be effectively utilized for 3D editing and applications by applying training-free inversion techniques [13, 52, 54] and inpainting methods [11, 51, 70, 71, 77]. We modify SDEdit [52] inversion for rectified flow-based inversion and adopt the RePaint [51] to support rectified flow inpainting. Additional technical details are provided in Appendix A.", + "bbox": [ + 511, + 323, + 906, + 430 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3DGS editing Our GSDecoder is designed to directly decode 3DGS, enabling efficient editing by modifying only the multi-view latents. Additionally, with the RF model's joint modeling of multi-view image latents, edits can be achieved without extra modules (e.g., cross-view attention mechanisms [88, 90]), streamlining the process through simple inversion and sampling. Specifically, given a target text and input multi-view latents, we apply a modified SDEdit for inversion at a chosen timestep $t_k$ . Conditioned on the target text, we then generate the edited multi-view latents by solving the ODE in Eq. 1 following our sampling procedure.", + "bbox": [ + 511, + 433, + 908, + 601 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Inpainting application Our model is also applicable to various 3D tasks, as it jointly models multi-view images, depths, and camera poses. This joint modeling allows known parts of the data (e.g., some views or camera parameters) to act as constraints, enabling the prediction of unknown parts through inpainting. In this work, we focus on two specific tasks: 1) camera pose estimation from multi-view images and depths, and 2) novel view synthesis using a subset of the $K$ multi-view latents and camera poses for novel viewpoints.", + "bbox": [ + 511, + 604, + 908, + 741 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Experimental Results", + "text_level": 1, + "bbox": [ + 511, + 753, + 718, + 771 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we demonstrate the capabilities of our proposed SplatFlow across 3DGS generation, 3DGS editing, and various 3D tasks. We begin by detailing the implementation of SplatFlow in Section 5.1. Following this, we present comparative results on 3DGS generation in real-world datasets in Section 5.2. In Section 5.3, we evaluate the 3DGS editing performance of SplatFlow against other methods. Finally, Section 5.4 showcases the versatility of our", + "bbox": [ + 511, + 779, + 908, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "21528", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/42ac89b4832e303bdded132ff403ebf887638a38b7f79b795357cc9b61a81e8b.jpg", + "image_caption": [ + "Figure 3. Qualitative results in text-to-3DGS generation on MVImgNet and DL3DV validation sets. The first two rows are rendered scenes from the MVImgNet dataset, while the last two rows are from the DL3DV dataset. Our SplatFlow produces cohesive and realistic scenes with sharp details, accurately capturing the intricacies of real-world environments and accommodating diverse camera trajectories." + ], + "image_footnote": [], + "bbox": [ + 99, + 85, + 898, + 659 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "multi-view RF model in handling diverse 3D tasks through training-free inpainting techniques. Extensive information on all our experiments is provided in Appendix B with additional results that could not be included here in the Appendix.", + "bbox": [ + 89, + 714, + 486, + 777 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Experimental Setups", + "text_level": 1, + "bbox": [ + 89, + 787, + 290, + 803 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Due to space constraints, we provide a concise overview of our experimental setups here. We used the MVImgNet [98] and DL3DV-7K [45] datasets, both containing real-world scenes with images and corresponding camera poses. For evaluation, we utilized 1.25K scenes from MVImgNet for each task and 300 sequences from DL3DV-7K, using the", + "bbox": [ + 89, + 810, + 485, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "remaining for training. The multi-view RF model and GSDecoder were trained with $K = 8$ -view setup. Text descriptions for scenes were extracted using the Llava-One Vision Qwen 7B [38]. We used SDS++ [41] for further improvements.", + "bbox": [ + 511, + 715, + 908, + 777 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Results on Text-to-3DGS Generation", + "text_level": 1, + "bbox": [ + 511, + 787, + 831, + 801 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We evaluated our model against various baselines on the DL3DV and MVImgNet validation sets for real-world scene text-to-3DGS generation. Our evaluation protocol utilized the rendered image of the generated 3DGS. We used the FID score [25] to evaluate image quality and the CLIP score [24] to measure alignment with text prompts.", + "bbox": [ + 511, + 810, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "21529", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/b849d636d8a49b6f5bcde5e6958e6858095cbf417a553e950f4e9957852fb868.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodMVImgNet [98]DL3DV [45]
FID-10K↓CLIPScore↑FID-2.4K↓CLIPScore↑
Director3D [41]39.5530.4888.4430.04
Director3D (w/ SDS++) [41]41.8031.0095.8831.68
SplatFlow34.8531.4379.9130.06
SplatFlow (w/ SDS++)35.4632.3085.3131.90
", + "bbox": [ + 94, + 88, + 480, + 175 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For quantitative comparison, we adopt Director3D [41], which leverages the full sequences of MVImgNet, DL3DV-10K, and LAION [73] datasets. In contrast, our model was trained on a notably smaller dataset, utilizing only a portion of MVImgNet and DL3DV-7K, without access to the full DL3DV-10K or LAION data. For qualitative results, we also compared against other scene generation methods, including LucidDreamer [12], DreamScene [39], and Director3D [41].", + "bbox": [ + 89, + 234, + 483, + 354 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Quantitative results Table 1 presents the quantitative comparison of SplatFlow and the baseline methods on both MVImgNet and DL3DV datasets. SplatFlow consistently outperforms Director3D [41] in terms of FID and CLIP score, demonstrating its effectiveness in generating high-quality images and maintaining strong alignment with text prompts, even with a smaller training dataset. Specifically, on MVImgNet, SplatFlow achieves a significantly lower FID (34.85 vs. 39.55 for Director3D), indicating better image quality, and a higher CLIP score (31.43 vs. 30.48), showing improved text-to-image alignment. Similarly, on DL3DV, SplatFlow achieves an FID of 79.91, outperforming Director3D's 88.44, with a comparable CLIP score. Notably, with the addition of $\\mathrm{SDS}++$ for refinement, both SplatFlow and Director3D show increased CLIP scores, further enhancing text alignment, though at a minor trade-off in FID. This highlights the effectiveness of $\\mathrm{SDS}++$ for boosting alignment with text prompts across different 3DGS generation models.", + "bbox": [ + 89, + 363, + 483, + 636 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Qualitative results Figure 3 shows the qualitative comparisons between our method and baselines on DL3DV and MVImgNet validation sets. Overall, SplatFlow demonstrates clear advantages over the baseline methods. First, DreamScene [39], which relies on Score Distillation Sampling (SDS), often struggles to generate realistic real-world scenes, resulting in outputs that lack natural textures and environmental details. Second, LucidDreamer [12], which is based on single-view inpainting, faces challenges when handling scenes with large camera trajectory variations, leading to inconsistent or incomplete views. Third, Director3D [41] tends to produce slightly blurred outputs, lacking the sharpness needed for intricate details. In contrast, our SplatFlow can generate coherent and realistic scenes with clear details, effectively capturing the complexities of real-world environments and various camera trajectories. Additional qualitative results of SplatFlow can be found in Appendix C.", + "bbox": [ + 89, + 643, + 485, + 901 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/86838318c008a85c46e9f3f96907c998be65ac0cf1a0c5c8b518c2da0e4624f0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 516, + 95, + 565, + 148 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/272d339f885460e47bbda6a32fdf95328d7eb5fb647912a24ba172449f99bd78.jpg", + "image_caption": [ + "\"green inflatable ring\" -> \"colorful red rubber ball\"" + ], + "image_footnote": [], + "bbox": [ + 576, + 95, + 676, + 148 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/82751cfae236038e957eab30d907c57afb1fed744ea0777599412dcc7c535ef3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 689, + 95, + 790, + 148 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e02dd886852cc769955fbeddd251757b7a7f068bd889bf9b172ca157513b4e70.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 803, + 95, + 903, + 148 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5c1d203bd60c12d13f5695c95edfa6ff5f6aad279626ea24f5182e5696a256f8.jpg", + "image_caption": [ + "Figure 4. Qualitative results in 3D editing with MVInpainter [3] and DGE [7]. We show rendered scenes except for MVInpainter." + ], + "image_footnote": [], + "bbox": [ + 516, + 170, + 565, + 210 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b114b2b479f68a9e32df423ebdd098815a890778831b32afbd835dd157b7e334.jpg", + "image_caption": [ + "\"collection of red chili peppers\" -> \"cluster of bananas\"" + ], + "image_footnote": [], + "bbox": [ + 576, + 170, + 676, + 210 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2bf6cfd51c36531d7f5b42e4dddb964d0fb707e7642b4cd7c6034ac139ad2f58.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 689, + 170, + 790, + 210 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c1ce82817da1b479833e2b63d43141c79b95d1ba35e1eea62ab6d56fb3cb0aca.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 803, + 170, + 903, + 210 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Result on 3DGS Editing", + "text_level": 1, + "bbox": [ + 511, + 275, + 733, + 291 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We demonstrate the effectiveness of SplatFlow in the 3DGS editing task by creating a benchmark with 100 sampled scenes from the MVImgNet validation set.", + "bbox": [ + 511, + 297, + 697, + 387 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/eb2e4e10b9ca2342d765ea50a3f1742a0eca084d3ae937aa02678807d6670ef9.jpg", + "table_caption": [ + "Table 1. Quantitative results in text-to-3DGS generation on the MVImgNet and DL3DV datasets. We compared our SplatFlow with and without the SDS++ [41], against Director3D [41]." + ], + "table_footnote": [], + "table_body": "
MethodCLIPScore↑CLIP D-sim↑
DGE [7]27.430.102
SplatFlow28.470.169
+) SDS++31.300.224
", + "bbox": [ + 707, + 299, + 901, + 362 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. 3D object replacement.", + "bbox": [ + 707, + 367, + 903, + 381 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Images and texts were provided to GPT-4 to identify main objects and generate text captions for edited images, which were then used for 3D object replacement. For editing, we used the same mask in [3] for masking out the object, then inpainted this area based on the provided editing captions. We compared this pipeline with an optimization-based 3DGS editing method, DGE [7], and conducted qualitative comparisons with MVInpainter [3], which propagates an edited image from pretrained generative models across multiple views. Following standard protocols [8, 29], we measured the CLIP score and directional similarity (CLIP D-sim).", + "bbox": [ + 511, + 387, + 906, + 554 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Quantitative results As shown in Table 2, SplatFlow achieves higher CLIPScore and CLIP D-sim than DGE [7], indicating more accurate and effective editing. SDS++ refinement further boosts performance, demonstrating enhanced alignment with target edits.", + "bbox": [ + 511, + 558, + 908, + 633 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Qualitative results Figure 4 demonstrates that SplatFlow performs comparably to MVInpainter, a specialized 2D inpainting method that leverages an already edited high-quality image. On the other hand, DGE struggles to replace objects fully, often just altering styles instead of achieving complete transformations. This underscores SplatFlow's superior ability in precise 3D object replacement. Please refer to Appendix C for further qualitative results.", + "bbox": [ + 511, + 638, + 908, + 758 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.4. Result on Inpainting Application", + "text_level": 1, + "bbox": [ + 511, + 768, + 802, + 784 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Here, we present the adaptability of our multi-view RF model to camera pose estimation and novel-view synthesis via the inpainting technique on the MVImgNet dataset [98].", + "bbox": [ + 511, + 790, + 905, + 835 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Camera pose estimation To demonstrate the effectiveness of our multi-view RF model in camera pose estimation, we primarily compare our approach to RayDiffusion [101] and RelPose++ [43]. We conducted comparisons using an", + "bbox": [ + 511, + 840, + 908, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "21530", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/6bed44d30484d48b349ec29d57bdabfd5c3916b38a4cab1e0b987a8c8f13ddbd.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodRotation↑Camera Center↑
@5@10@15@0.05@0.1@0.2
RelPose++ [43]19.437.751.40.612.555.0
Ray Regression [101]10.425.650.115.347.982.9
Ray Diffusion [101]17.538.759.624.160.987.6
SplatFlow (w/ depth)26.852.659.362.391.699.4
SplatFlow (w/o depth)28.854.563.964.994.099.7
", + "bbox": [ + 94, + 88, + 478, + 199 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b57d6b733081eee53271290b47a7c88ef5fc3eb25e1114becb7e0f343a4d9ffd.jpg", + "image_caption": [ + "Figure 5. Qualitative results for camera pose estimation. Camera poses are estimated from multi-view images. Image border colors match each camera, with black cameras indicating GT poses." + ], + "image_footnote": [], + "bbox": [ + 102, + 253, + 336, + 455 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e3e88792debe0a2d7dd9bfdf15666e4d8a8b5a66aef70d55d1659f4a22fde5f1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 354, + 257, + 467, + 455 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "8-view setup, inpainting the ray latent with and without the depth latent from DepthAnythingV2 [93]. We used rotation accuracy@ $Q_{R}$ and translation accuracy@ $Q_{T}$ as evaluation metrics, measuring the proportion of relative camera rotations within $Q_{R}$ degrees of ground truth and camera centers within $Q_{T}$ units of scene scale, respectively.", + "bbox": [ + 89, + 513, + 483, + 604 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3 and Fig. 5 present the quantitative and qualitative results. Overall, the multi-view RF model achieves superior performance across most scenarios, outperforming the baseline methods. Interestingly, the multi-view RF model achieves higher accuracy without depth latents than with them. We hypothesize that this outcome is due to the joint generation of depth and camera poses, which may yield more detailed and contextually enriched depth estimates.", + "bbox": [ + 89, + 606, + 483, + 727 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Novel view-synthesis We evaluated the performance of our multi-view RF model in novel view synthesis under two scenarios: 1) interpolation, with the uniformly sampled input views, and 2) extrapolation, with the input views positioned in the center. This comparison validates our multi-view RF model's capability of 3D reasoning based on relative viewpoints and evaluates its effectiveness in generating novel views. We evaluated PSNR, SSIM, and LPIPS [104] against the ground-truth images for reconstruction quality, as well as absolute relative difference and $\\delta_{1}$ accuracy using depth maps produced by DepthAnythingV2 [93].", + "bbox": [ + 89, + 734, + 483, + 902 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/fbb580335ddfb40b515c1c128e557a87c31be1a8ec6eda39691e4669dd81a243.jpg", + "table_caption": [ + "Table 3. Results in camera pose estimation on MVImgNet validation set. @Q represents the accuracy threshold for rotations (degrees) and camera centers (units)." + ], + "table_footnote": [], + "table_body": "
TypeRGBDepth
PSNR↑SSIM↑LPIPS↓AbsRel↓δ1↑
Interpolation (N=2)14.730.5710.6480.5880.731
Interpolation (N=4)17.050.5900.5510.4980.761
Interpolation (N=6)18.820.6260.4830.4150.775
Extrapolation (N=2)15.150.5770.6270.7710.715
Extrapolation (N=4)16.800.5950.5540.6790.727
Extrapolation (N=6)17.960.6130.5030.6020.747
", + "bbox": [ + 517, + 88, + 903, + 215 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4. Novel view synthesis results on MVImgNet. We use $N$ input views to synthesize $K - N$ novel views, with uniformly sampled views for interpolation and central views for extrapolation.", + "bbox": [ + 513, + 215, + 903, + 258 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/26e343442881543e49df324d7c91b6ae61ee85277f90e733dbeef9ad7a21dfcf.jpg", + "image_caption": [ + "Figure 6. Qualitative results for novel view synthesis. Novel view synthesis is performed from the red-box images and depths." + ], + "image_footnote": [], + "bbox": [ + 516, + 271, + 903, + 511 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As shown in Fig. 6, our multi-view RF model successfully generates novel views for both scenarios. Also, Table 4 shows that interpolation outperforms extrapolation in both RGB reconstruction and depth estimation as the number of input views increases, indicating that interpolation leverages 3D reasoning for accurate novel view generation, whereas extrapolation focuses on generating diverse novel views.", + "bbox": [ + 511, + 553, + 906, + 659 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 676, + 633, + 691 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we have introduced SplatFlow, a novel framework that unifies 3D Gaussian Splitting (3DGS) generation and editing within a single, efficient model. By designing a multi-view Rectified Flow (RF) model alongside a Gaussian Splitting Decoder (GSDecoder), we achieved direct 3DGS generation capable of addressing real-world challenges, such as varying scales and complex camera trajectories. Our approach utilizes joint modeling of multi-view images, depths, and camera poses, enabling seamless integration of training-free inversion and inpainting techniques for a diverse set of 3D tasks. As we present the versatility of our SplatFlow, we believe SplatFlow represents an important step toward a versatile foundational model for 3D content.", + "bbox": [ + 511, + 704, + 908, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "21531", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Michael S Albergo and Eric Vanden-Eijnden. Building normalizing flows with stochastic interpolants. arXiv preprint arXiv:2209.15571, 2022. 2, 3", + "[2] Chong Bao, Yinda Zhang, Bangbang Yang, Tianxing Fan, Zesong Yang, Hujun Bao, Guofeng Zhang, and Zhaopeng Cui. Sine: Semantic-driven image-based nerf editing with prior-guided editing field. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20919-20929, 2023. 1, 3", + "[3] Chenjie Cao, Chaohui Yu, Yanwei Fu, Fan Wang, and Xi-angyang Xue. Mvinpainter: Learning multi-view consistent inpainting to bridge 2d and 3d editing. arXiv preprint arXiv:2408.08000, 2024.7", + "[4] David Charatan, Sizhe Lester Li, Andrea Tagliasacchi, and Vincent Sitzmann. pixelsplat: 3d gaussian splats from image pairs for scalable generalizable 3d reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19457-19467, 2024. 2, 4", + "[5] Anpei Chen, Haofei Xu, Stefano Esposito, Siyu Tang, and Andreas Geiger. Lara: Efficient large-baseline radiance fields. In European Conference on Computer Vision, pages 338-355. Springer, 2025. 3", + "[6] Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, et al. Pixart-α: Fast training of diffusion transformer for photorealistic text-to-image synthesis. arXiv preprint arXiv:2310.00426, 2023. 2, 3", + "[7] Minghao Chen, Iro Laina, and Andrea Vedaldi. Dge: Direct gaussian 3d editing by consistent multi-view editing. arXiv preprint arXiv:2404.18929, 2024. 2, 7", + "[8] Yiwen Chen, Zilong Chen, Chi Zhang, Feng Wang, Xiaofeng Yang, Yikai Wang, Zhongang Cai, Lei Yang, Huaping Liu, and Guosheng Lin. Gaussianeditor: Swift and controllable 3d editing with gaussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21476-21485, 2024. 2, 3, 7", + "[9] Yuedong Chen, Haofei Xu, Chuanxia Zheng, Bohan Zhuang, Marc Pollefeys, Andreas Geiger, Tat-Jen Cham, and Jianfei Cai. Mvsplat: Efficient 3d gaussian splattering from sparse multi-view images. In European Conference on Computer Vision, pages 370-386. Springer, 2024. 2, 4", + "[10] Zilong Chen, Feng Wang, Yikai Wang, and Huaping Liu. Text-to-3d using gaussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21401-21412, 2024. 2, 3", + "[11] Hyungjin Chung, Jeongsol Kim, Michael T McCann, Marc L Klasky, and Jong Chul Ye. Diffusion posterior sampling for general noisy inverse problems. arXiv preprint arXiv:2209.14687, 2022. 2, 5", + "[12] Jaeyoung Chung, Suyoung Lee, Hyeongjin Nam, Jaerin Lee, and Kyoung Mu Lee. Luciddreamer: Domain-free generation of 3d gaussian splatting scenes. arXiv preprint arXiv:2311.13384, 2023. 3, 6, 7", + "[13] Guillaume Couairon, Jakob Verbeek, Holger Schwenk, and Matthieu Cord. Diffedit: Diffusion-based seman" + ], + "bbox": [ + 101, + 114, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "tic image editing with mask guidance. arXiv preprint arXiv:2210.11427, 2022. 2, 5", + "[14] Giannis Daras, Hyungjin Chung, Chieh-Hsin Lai, Yuki Mitsufuji, Jong Chul Ye, Peyman Milanfar, Alexandros G Dimakis, and Mauricio Delbracio. A survey on diffusion models for inverse problems. arXiv preprint arXiv:2410.00083, 2024. 2", + "[15] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3d objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13142-13153, 2023. 2, 3", + "[16] Matt Deitke, Ruoshi Liu, Matthew Wallingford, Huong Ngo, Oscar Michel, Aditya Kusupati, Alan Fan, Christian Laforte, Vikram Voleti, Samir Yitzhak Gadre, et al. Objverse-xl: A universe of $10\\mathrm{m} + 3\\mathrm{d}$ objects. Advances in Neural Information Processing Systems, 36, 2024. 2, 3", + "[17] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 2", + "[18] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12873-12883, 2021. 4", + "[19] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In *Forty-first International Conference on Machine Learning*, 2024. 2, 4, 5", + "[20] Hyojun Go, Yunsung Lee, Jin-Young Kim, Seunghyun Lee, Myeongho Jeong, Hyun Seung Lee, and Seungtaek Choi. Towards practical plug-and-play diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1962-1971, 2023. 2", + "[21] Hyojun Go, Yunsung Lee, Seunghyun Lee, Shinhyeok Oh, Hyeongdon Moon, and Seungtaek Choi. Addressing negative transfer in diffusion models. Advances in Neural Information Processing Systems, 36, 2024. 2", + "[22] Alexandros Graikos, Nikolay Malkin, Nebojsa Jojic, and Dimitris Samaras. Diffusion models as plug-and-play priors. Advances in Neural Information Processing Systems, 35: 14715-14728, 2022. 2", + "[23] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 2", + "[24] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning. arXiv preprint arXiv:2104.08718, 2021.6", + "[25] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 6" + ], + "bbox": [ + 524, + 93, + 906, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "21532", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[26] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2, 3", + "[27] Jonathan Ho, William Chan, Chitwan Sahara, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. 2", + "[28] Lukas Hollein, Ang Cao, Andrew Owens, Justin Johnson, and Matthias Nießner. Text2room: Extracting textured 3d meshes from 2d text-to-image models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7909-7920, 2023. 3", + "[29] Hiromichi Kamata, Yuiko Sakuma, Akio Hayakawa, Masato Ishii, and Takuya Narihira. Instruct 3d-to-3d: Text instruction guided 3d-to-3d conversion. arXiv preprint arXiv:2303.15780, 2023. 1, 3, 7", + "[30] Nazmul Karim, Hasan Iqbal, Umar Khalid, Chen Chen, and Jing Hua. Free-editor: zero-shot text-driven 3d scene editing. In European Conference on Computer Vision, pages 436-453. Springer, 2025. 1, 3", + "[31] Bingxin Ke, Anton Obukhov, Shengyu Huang, Nando Metzger, Rodrigo Caye Daudt, and Konrad Schindler. Repurposing diffusion-based image generators for monocular depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9492-9502, 2024. 4", + "[32] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splattering for real-time radiance field rendering. ACM Trans. Graph., 42(4):139-1, 2023. 2, 3, 4", + "[33] Umar Khalid, Hasan Iqbal, Nazmul Karim, Jing Hua, and Chen Chen. Latenteditor: Text driven local editing of 3d scenes. arXiv preprint arXiv:2312.09313, 2023. 1, 3", + "[34] Umar Khalid, Hasan Iqbal, Azib Farooq, Jing Hua, and Chen Chen. 3dego: 3d editing on the go! In European Conference on Computer Vision, pages 73-89. Springer, 2024. 2, 3", + "[35] Gwanghyun Kim, Taesung Kwon, and Jong Chul Ye. Diffusionclip: Text-guided diffusion models for robust image manipulation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2426-2435, 2022. 2", + "[36] Nupur Kumari, Richard Zhang, Eli Shechtman, and JunYan Zhu. Ensembling off-the-shelf models for gan training. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10651-10662, 2022. 4, 5", + "[37] Yunsung Lee, JinYoung Kim, Hyojun Go, Myeongho Jeong, Shinhyeok Oh, and Seungtaek Choi. Multi-architecture multi-expert diffusion models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 13427-13436, 2024. 2", + "[38] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 6", + "[39] Haoran Li, Haolin Shi, Wenli Zhang, Wenjun Wu, Yong Liao, Lin Wang, Lik-hang Lee, and Pengyuan Zhou. Dreamscene:" + ], + "bbox": [ + 99, + 92, + 483, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "3d gaussian-based text-to-3d scene generation via formation pattern sampling. arXiv preprint arXiv:2404.03575, 2024.6, 7", + "[40] Xiang Li, John Thickstun, Ishaan Gulrajani, Percy S Liang, and Tatsunori B Hashimoto. Diffusion-lm improves controllable text generation. Advances in Neural Information Processing Systems, 35:4328-4343, 2022. 2", + "[41] Xinyang Li, Zhangyu Lai, Linning Xu, Yansong Qu, Liujuan Cao, Shengchuan Zhang, Bo Dai, and Rongrong Ji. Director3d: Real-world camera trajectory and 3d scene generation from text. Advances in neural information processing systems, 2024. 3, 5, 6, 7", + "[42] Zongrui Li, Minghui Hu, Qian Zheng, and Xudong Jiang. Connecting consistency distillation to score distillation for text-to-3d generation. In European Conference on Computer Vision, pages 274–291. Springer, 2024. 2", + "[43] Amy Lin, Jason Y Zhang, Deva Ramanan, and Shubham Tulsiani. Relpose++: Recovering 6d poses from sparse-view observations. arXiv preprint arXiv:2305.04926, 2023. 7, 8", + "[44] Huan Ling, Seung Wook Kim, Antonio Torralba, Sanja Fidler, and Karsten Kreis. Align your gaussians: Text-to-4d with dynamic 3d gaussians and composed diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8576–8588, 2024. 3", + "[45] Lu Ling, Yichen Sheng, Zhi Tu, Wentian Zhao, Cheng Xin, Kun Wan, Lantao Yu, Qianyu Guo, Zixun Yu, Yawen Lu, et al. D13dv-10k: A large-scale scene dataset for deep learning-based 3d vision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22160–22169, 2024. 2, 6, 7", + "[46] Yaron Lipman, Ricky TQ Chen, Heli Ben-Hamu, Maximilian Nickel, and Matt Le. Flow matching for generative modeling. arXiv preprint arXiv:2210.02747, 2022. 2, 3", + "[47] Xingchao Liu, Chengyue Gong, and Qiang Liu. Flow straight and fast: Learning to generate and transfer data with rectified flow. arXiv preprint arXiv:2209.03003, 2022. 2, 3", + "[48] Xian Liu, Xiaohang Zhan, Jiaxiang Tang, Ying Shan, Gang Zeng, Dahua Lin, Xihui Liu, and Ziwei Liu. Humangaussian: Text-driven 3d human generation with gaussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6646-6657, 2024. 2", + "[49] Yuan Liu, Cheng Lin, Zijiao Zeng, Xiaoxiao Long, Lingjie Liu, Taku Komura, and Wenping Wang. Syncdreamer: Generating multiview-consistent images from a single-view image. arXiv preprint arXiv:2309.03453, 2023. 3", + "[50] Xiaoxiao Long, Yuan-Chen Guo, Cheng Lin, Yuan Liu, Zhiyang Dou, Lingjie Liu, Yuexin Ma, Song-Hai Zhang, Marc Habermann, Christian Theobalt, et al. Wonder3d: Single image to 3d using cross-domain diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9970-9980, 2024. 3", + "[51] Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11461-11471, 2022. 2, 4, 5" + ], + "bbox": [ + 522, + 92, + 906, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "21533", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[52] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jiajun Wu, Jun-Yan Zhu, and Stefano Ermon. Sdedit: Guided image synthesis and editing with stochastic differential equations. arXiv preprint arXiv:2108.01073, 2021. 2, 4, 5", + "[53] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 1, 2", + "[54] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6038–6047, 2023. 2, 5", + "[55] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM transactions on graphics (TOG), 41(4):1-15, 2022. 1", + "[56] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 2", + "[57] Alex Nichol, Heewoo Jun, Prafulla Dhariwal, Pamela Mishkin, and Mark Chen. Point-e: A system for generating 3d point clouds from complex prompts. arXiv preprint arXiv:2212.08751, 2022. 2", + "[58] Francesco Palandra, Andrea Sanchietti, Daniele Baieri, and Emanuele Rodola. Gsedit: Efficient text-guided editing of 3d objects via gaussian splatting. arXiv preprint arXiv:2403.05154, 2024. 2, 3", + "[59] Byeongjun Park and Changick Kim. Point-dynrf: Point-based dynamic radiance fields from a monocular video. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 3171-3181, 2024. 3", + "[60] Byeongjun Park, Sangmin Woo, Hyojun Go, Jin-Young Kim, and Changick Kim. Denoising task routing for diffusion models. arXiv preprint arXiv:2310.07138, 2023. 2", + "[61] Byeongjun Park, Hyojun Go, and Changick Kim. Bridging implicit and explicit geometric transformation for single-image view synthesis. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 2", + "[62] Byeongjun Park, Hyojun Go, Jin-Young Kim, Sangmin Woo, Seokil Ham, and Changick Kim. Switch diffusion transformer: Synergizing denoising tasks with sparse mixture-of-experts. arXiv preprint arXiv:2403.09176, 2024. 2", + "[63] Jangho Park, Gihyun Kwon, and Jong Chul Ye. Ed-nerf: Efficient text-guided editing of 3d scene using latent space nerf. arXiv preprint arXiv:2310.02712, 2023. 1, 3", + "[64] Julius Plücker. Analytisch-geometrische Entwicklungen. GD Baedeker, 1828. 5", + "[65] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 2, 3" + ], + "bbox": [ + 99, + 90, + 485, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[66] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 2", + "[67] Elad Richardson, Gal Metzer, Yuval Alaluf, Raja Giryes, and Daniel Cohen-Or. Texture: Text-guided texturing of 3d shapes. In ACM SIGGRAPH 2023 conference proceedings, pages 1–11, 2023. 3", + "[68] Robin Rombach, Patrick Esser, and Björn Ommer. Geometry-free view synthesis: Transformers and no 3d priors. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14356-14366, 2021. 2", + "[69] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2, 3, 4", + "[70] Litu Rout, Advait Parulekar, Constantine Caramanis, and Sanjay Shakkottai. A theoretical justification for image inpainting using denoising diffusion probabilistic models. arXiv preprint arXiv:2302.01217, 2023. 2, 5", + "[71] Litu Rout, Negin Raoof, Giannis Daras, Constantine Caramanis, Alex Dimakis, and Sanjay Shakkottai. Solving linear inverse problems provably via posterior sampling with latent diffusion models. Advances in Neural Information Processing Systems, 36, 2023. 2, 5", + "[72] Axel Sauer, Dominik Lorenz, Andreas Blattmann, and Robin Rombach. Adversarial diffusion distillation. In European Conference on Computer Vision, pages 87-103. Springer, 2024. 2", + "[73] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278-25294, 2022. 3, 7", + "[74] Yichun Shi, Peng Wang, Jianglong Ye, Mai Long, Kejie Li, and Xiao Yang. Mvdream: Multi-view diffusion for 3d generation. arXiv preprint arXiv:2308.16512, 2023. 3", + "[75] Jaidev Shriram, Alex Trevithick, Lingjie Liu, and Ravi Ramamoorthi. Realmdreamer: Text-driven 3d scene generation with inpainting and depth diffusion. arXiv preprint arXiv:2404.07199, 2024. 3", + "[76] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International conference on machine learning, pages 2256-2265. PMLR, 2015. 2,3", + "[77] Jiaming Song, Arash Vahdat, Morteza Mardani, and Jan Kautz. Pseudoinverse-guided diffusion models for inverse problems. In International Conference on Learning Representations, 2023. 2, 5", + "[78] Jiaming Song, Qinsheng Zhang, Hongxu Yin, Morteza Mardani, Ming-Yu Liu, Jan Kautz, Yongxin Chen, and Arash Vahdat. Loss-guided diffusion models for plug-and-play controllable generation. In International Conference on Machine Learning, pages 32483-32498. PMLR, 2023. 2" + ], + "bbox": [ + 524, + 92, + 906, + 898 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "21534", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[79] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 2, 3", + "[80] Stanislaw Szymanowicz, Chrisitian Rupprecht, and Andrea Vedaldi. Splatter image: Ultra-fast single-view 3d reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10208-10217, 2024. 4", + "[81] Jiaxiang Tang, Zhaoxi Chen, Xiaokang Chen, Tengfei Wang, Gang Zeng, and Ziwei Liu. Lgm: Large multi-view gaussian model for high-resolution 3d content creation. In European Conference on Computer Vision, pages 1-18. Springer, 2024. 2, 3, 4", + "[82] Jiaxiang Tang, Jiawei Ren, Hang Zhou, Ziwei Liu, and Gang Zeng. Dreamgaussian: Generative gaussian splatting for efficient 3d content creation. In The Twelfth International Conference on Learning Representations, 2024. 2, 3", + "[83] Arash Vahdat, Francis Williams, Zan Gojcic, Or Litany, Sanja Fidler, Karsten Kreis, et al. Lion: Latent point diffusion models for 3d shape generation. Advances in Neural Information Processing Systems, 35:10021-10039, 2022. 2", + "[84] Alexander Vilesov, Pradyumna Chari, and Achuta Kadambi. Cg3d: Compositional generation for text-to-3d via gaussian splatting. arXiv preprint arXiv:2311.17907, 2023. 2", + "[85] Vikram Voleti, Chun-Han Yao, Mark Boss, Adam Letts, David Pankratz, Dmitry Tochilkin, Christian Laforte, Robin Rombach, and Varun Jampani. Sv3d: Novel multi-view synthesis and 3d generation from a single image using latent video diffusion. In European Conference on Computer Vision, pages 439-457. Springer, 2025. 3", + "[86] Can Wang, Mingming He, Mengei Chai, Dongdong Chen, and Jing Liao. Mesh-guided neural implicit field editing. arXiv preprint arXiv:2312.02157, 2023. 3", + "[87] Junjie Wang, Jiemin Fang, Xiaopeng Zhang, Lingxi Xie, and Qi Tian. Gaussianeditor: Editing 3d gaussians delicately with text instructions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20902-20911, 2024. 2, 3", + "[88] Yuxuan Wang, Xuanyu Yi, Zike Wu, Na Zhao, Long Chen, and Hanwang Zhang. View-consistent 3d editing with gaussian splatting. In European Conference on Computer Vision, pages 404-420. Springer, 2024. 2, 3, 5", + "[89] Sangmin Woo, Byeongjun Park, Hyojun Go, Jin-Young Kim, and Changick Kim. Harmonyview: Harmonizing consistency and diversity in one-image-to-3d. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10574–10584, 2024. 3", + "[90] Jing Wu, Jia-Wang Bian, Xinghui Li, Guangrun Wang, Ian Reid, Philip Torr, and Victor Adrian Prisacariu. Gaussctrl: multi-view consistent text-driven 3d gaussian splattering editing. arXiv preprint arXiv:2403.08733, 2024. 2, 3, 5", + "[91] Hanyuan Xiao, Yingshu Chen, Huajian Huang, Haolin Xiong, Jing Yang, Pratusha Prasad, and Yajie Zhao. Localized gaussian splatting editing with contextual awareness. arXiv preprint arXiv:2408.00083, 2024. 2, 3", + "[92] Yinghao Xu, Hao Tan, Fujun Luan, Sai Bi, Peng Wang, Jiahao Li, Zifan Shi, Kalyan Sunkavalli, Gordon Wetzstein," + ], + "bbox": [ + 99, + 92, + 485, + 900 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Zexiang Xu, and Kai Zhang. DMV3d: Denoising multiview diffusion using 3d large reconstruction model. In The Twelfth International Conference on Learning Representations, 2024. 2", + "[93] Lihe Yang, Bingyi Kang, Zilong Huang, Zhen Zhao, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything v2. arXiv preprint arXiv:2406.09414, 2024. 4, 8", + "[94] Taoran Yi, Jiemin Fang, Junjie Wang, Guanjun Wu, Lingxi Xie, Xiaopeng Zhang, Wenyu Liu, Qi Tian, and Xinggang Wang. Gaussian dreamer: Fast generation from text to 3d gaussians by bridging 2d and 3d diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6796-6807, 2024. 2", + "[95] Xuanyu Yi, Zike Wu, Qingshan Xu, Pan Zhou, Joo-Hwee Lim, and Hanwang Zhang. Diffusion time-step curriculum for one image to 3d generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9948-9958, 2024. 3", + "[96] Lu Yu, Wei Xiang, and Kang Han. Edit-diffnerf: Editing 3d neural radiance fields using 2d diffusion model. arXiv preprint arXiv:2306.09551, 2023. 1, 3", + "[97] Wangbo Yu, Jinbo Xing, Li Yuan, Wenbo Hu, Xiaoyu Li, Zhipeng Huang, Xiangjun Gao, Tien-Tsin Wong, Ying Shan, and Yonghong Tian. Viewcrafter: Taming video diffusion models for high-fidelity novel view synthesis. arXiv preprint arXiv:2409.02048, 2024. 2", + "[98] Xianggang Yu, Mutian Xu, Yidan Zhang, Haolin Liu, Chongjie Ye, Yushuang Wu, Zizheng Yan, Chenming Zhu, Zhangyang Xiong, Tianyou Liang, et al. Mvimgnet: A large-scale dataset of multi-view images. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9150-9161, 2023. 2, 6, 7", + "[99] Bowen Zhang, Yiji Cheng, Jiaolong Yang, Chunyu Wang, Feng Zhao, Yansong Tang, Dong Chen, and Baining Guo. Gaussiancube: Structuring gaussian splatting using optimal transport for 3d generative modeling. arXiv preprint arXiv:2403.19655, 2024. 2", + "[100] Jingbo Zhang, Xiaoyu Li, Ziyu Wan, Can Wang, and Jing Liao. Text2nerf: Text-driven 3d scene generation with neural radiance fields. IEEE Transactions on Visualization & Computer Graphics, 30(12):7749-7762, 2024. 1", + "[101] Jason Y. Zhang, Amy Lin, Moneish Kumar, Tzu-Hsuan Yang, Deva Ramanan, and Shubham Tulsiani. Cameras as rays: Pose estimation via ray diffusion. In The Twelfth International Conference on Learning Representations, 2024. 2, 5, 7, 8", + "[102] Kai Zhang, Sai Bi, Hao Tan, Yuanbo Xiangli, Nanxuan Zhao, Kalyan Sunkavalli, and Zexiang Xu. Gs-lrm: Large reconstruction model for 3d gaussian splatting. In European Conference on Computer Vision, pages 1-19. Springer, 2024. 2", + "[103] Kaiyi Zhang, Yang Chen, Ximing Yang, Weizhong Zhang, and Cheng Jin. Point cloud part editing: Segmentation, generation, assembly, and selection. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 7187-7195, 2024. 3", + "[104] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "21535", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 5, 8", + "[105] Yudi Zhang, Qi Xu, and Lei Zhang. Dragtex: Generative point-based texture editing on 3d mesh. arXiv preprint arXiv:2403.02217, 2024. 3", + "[106] Linqi Zhou, Yilun Du, and Jiajun Wu. 3d shape generation and completion through point-voxel diffusion. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5826-5835, 2021. 2", + "[107] Shijie Zhou, Zhiwen Fan, Dejia Xu, Haoran Chang, Pradyumna Chari, Tejas Bharadwaj, Suya You, Zhangyang Wang, and Achuta Kadambi. Dreamscene360: Unconstrained text-to-3d scene generation with panoramic gaussian splatting. In European Conference on Computer Vision, pages 324–342. Springer, 2025. 3", + "[108] Jingyu Zhuang, Di Kang, Yan-Pei Cao, Guanbin Li, Liang Lin, and Ying Shan. Tip-editor: An accurate 3d editor following both text-prompts and image-prompts. ACM Transactions on Graphics (TOG), 43(4):1-12, 2024. 2, 3" + ], + "bbox": [ + 91, + 92, + 483, + 373 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "21536", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 12 + } +] \ No newline at end of file diff --git a/2025/SplatFlow_ Multi-View Rectified Flow Model for 3D Gaussian Splatting Synthesis/a1da8541-9ef2-4901-8219-e22e7fdfc1fc_model.json b/2025/SplatFlow_ Multi-View Rectified Flow Model for 3D Gaussian Splatting Synthesis/a1da8541-9ef2-4901-8219-e22e7fdfc1fc_model.json new file mode 100644 index 0000000000000000000000000000000000000000..bc1cc81640eb76dec14256464bc53b1862c37bb6 --- /dev/null +++ b/2025/SplatFlow_ Multi-View Rectified Flow Model for 3D Gaussian Splatting Synthesis/a1da8541-9ef2-4901-8219-e22e7fdfc1fc_model.json @@ -0,0 +1,3240 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.0, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.131, + 0.908, + 0.155 + ], + "angle": 0, + "content": "SplatFlow: Multi-View Rectified Flow Model for 3D Gaussian Splitting Synthesis" + }, + { + "type": "image", + "bbox": [ + 0.212, + 0.179, + 0.782, + 0.249 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.268, + 0.236, + 0.31 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.316, + 0.236, + 0.36 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.362, + 0.236, + 0.413 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.413, + 0.236, + 0.451 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.121, + 0.45, + 0.204, + 0.457 + ], + "angle": 0, + "content": "A library with a lot of books" + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.46, + 0.236, + 0.505 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.24, + 0.268, + 0.376, + 0.31 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.24, + 0.316, + 0.376, + 0.353 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.25, + 0.353, + 0.368, + 0.359 + ], + "angle": 0, + "content": "A forest with tall trees and yellow leaves" + }, + { + "type": "image", + "bbox": [ + 0.24, + 0.362, + 0.376, + 0.399 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.24, + 0.4, + 0.371, + 0.413 + ], + "angle": 0, + "content": "Blue car on a clean street with buildings in the background." + }, + { + "type": "image", + "bbox": [ + 0.24, + 0.413, + 0.376, + 0.45 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.283, + 0.45, + 0.325, + 0.457 + ], + "angle": 0, + "content": "Half avocado" + }, + { + "type": "image", + "bbox": [ + 0.24, + 0.46, + 0.376, + 0.498 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.283, + 0.498, + 0.328, + 0.504 + ], + "angle": 0, + "content": "slicedzucchini" + }, + { + "type": "image", + "bbox": [ + 0.382, + 0.268, + 0.651, + 0.368 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.455, + 0.368, + 0.574, + 0.377 + ], + "angle": 0, + "content": "A lighthouse on a rocky shore" + }, + { + "type": "image", + "bbox": [ + 0.382, + 0.381, + 0.651, + 0.485 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.405, + 0.486, + 0.61, + 0.504 + ], + "angle": 0, + "content": "A vibrant bush with pink flowers stands in front of a concrete wall, ..." + }, + { + "type": "image_caption", + "bbox": [ + 0.27, + 0.509, + 0.463, + 0.522 + ], + "angle": 0, + "content": "(a) 3D Gaussian Splitting Synthesis" + }, + { + "type": "image", + "bbox": [ + 0.665, + 0.269, + 0.718, + 0.353 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.721, + 0.269, + 0.804, + 0.353 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.807, + 0.269, + 0.859, + 0.353 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.863, + 0.269, + 0.888, + 0.353 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.718, + 0.353, + 0.848, + 0.363 + ], + "angle": 0, + "content": "3DGS object editing (replacement)" + }, + { + "type": "image", + "bbox": [ + 0.665, + 0.37, + 0.718, + 0.425 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.719, + 0.37, + 0.809, + 0.425 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.815, + 0.37, + 0.882, + 0.423 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.724, + 0.426, + 0.822, + 0.434 + ], + "angle": 0, + "content": "\"Camera Pose Estimation\"" + }, + { + "type": "image", + "bbox": [ + 0.665, + 0.446, + 0.718, + 0.49 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.721, + 0.446, + 0.805, + 0.49 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.807, + 0.446, + 0.886, + 0.49 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.724, + 0.491, + 0.812, + 0.5 + ], + "angle": 0, + "content": "\"Novel View Synthesis\"" + }, + { + "type": "image_caption", + "bbox": [ + 0.692, + 0.509, + 0.859, + 0.522 + ], + "angle": 0, + "content": "(b) Training-Free Applications" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.529, + 0.907, + 0.572 + ], + "angle": 0, + "content": "Figure 1. SplatFlow for 3D Gaussian Splitting synthesis and its training-free applications. (a) Examples of direct 3D Gaussian Splitting (3DGS) generation only from text prompts, (b) Training-free applications, including 3DGS object editing, camera pose estimation, and novel view synthesis. SplatFlow seamlessly integrates these capabilities, showcasing its versatility in generating and editing complex 3D content." + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.591, + 0.327, + 0.607 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.623, + 0.486, + 0.865 + ], + "angle": 0, + "content": "Text-based generation and editing of 3D scenes hold significant potential for streamlining content creation through intuitive user interactions. While recent advances leverage 3D Gaussian Splatting (3DGS) for high-fidelity and real-time rendering, existing methods are often specialized and task-focused, lacking a unified framework for both generation and editing. In this paper, we introduce SplatFlow, a comprehensive framework that addresses this gap by enabling direct 3DGS generation and editing. SplatFlow comprises two main components: a multi-view rectified flow (RF) model and a Gaussian Splatting Decoder (GSDecoder). The multi-view RF model operates in latent space, generating multi-view images, depths, and camera poses simultaneously, conditioned on text prompts—thus addressing challenges like diverse scene scales and complex camera trajectories in real-world settings. Then, the" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.593, + 0.909, + 0.775 + ], + "angle": 0, + "content": "GSDecoder efficiently translates these latent outputs into 3DGS representations through a feed-forward 3DGS method. Leveraging training-free inversion and inpainting techniques, SplatFlow enables seamless 3DGS editing and supports a broad range of 3D tasks—including object editing, novel view synthesis, and camera pose estimation—within a unified framework without requiring additional complex pipelines. We validate SplatFlow's capabilities on the MVImgNet and DL3DV-7K datasets, demonstrating its versatility and effectiveness in various 3D generation, editing, and inpainting-based tasks. Our project page is available at https://gohyojun15.github.io/SplatFlow/." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.801, + 0.646, + 0.817 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.827, + 0.909, + 0.904 + ], + "angle": 0, + "content": "The demand for realistic 3D scene generation and editing from text has surged in applications like VR/AR, gaming, and robotics. Early works primarily utilized NeRF [53, 55] for scene generation [100] and editing [2, 29, 30, 33, 63, 96], but NeRF's volumetric rendering is computationally" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.877, + 0.217, + 0.889 + ], + "angle": 0, + "content": "*Equal contribution" + }, + { + "type": "page_footnote", + "bbox": [ + 0.111, + 0.889, + 0.234, + 0.902 + ], + "angle": 0, + "content": "† Corresponding author" + }, + { + "type": "list", + "bbox": [ + 0.109, + 0.877, + 0.234, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "21524" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.198 + ], + "angle": 0, + "content": "intensive, resulting in slower rendering speeds. Recently, 3D Gaussian Splitting (3DGS) [32] has emerged as a promising alternative, offering real-time and high-fidelity rendering implementations. Building on this, recent advances in 3D generation [8, 34, 58, 81, 87, 88, 90, 92, 99, 102, 108] and editing [8, 34, 58, 87, 88, 90, 108] increasingly leverage 3DGS to achieve both speed and quality." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.2, + 0.486, + 0.412 + ], + "angle": 0, + "content": "Despite these advancements, current methods for generating and editing 3DGS are specialized and task-focused, lacking a unified framework. For 3DGS generation, several works [10, 42, 48, 82, 84, 94] leverage 2D diffusion models [26, 69, 76, 79] with Score Distillation Sampling (SDS) [66], which requires time-intensive per-scene optimization. To address this, recent studies have shifted towards direct 3DGS generation, combining diffusion models with reconstruction models [81, 92, 102] or utilizing 3D generative models [99]. However, most of these methods are restricted to object-level 3D generation using synthetic datasets [15, 16] with controlled, bounded environments. In contrast, real-world scenes, varying in scene scales and camera trajectories, present distinctive challenges." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.413, + 0.485, + 0.581 + ], + "angle": 0, + "content": "In 3DGS editing, several works [8, 34, 58, 87, 88, 90, 108] adopt 2D diffusion models [26, 69, 76, 79] to guide the editing process. The main challenge is lifting 2D editing guidance to 3D while maintaining 3D-consistent modifications. One approach focuses on SDS techniques to update 3DGS [58, 91, 108], but these methods require additional stages like texture adjustments [58, 91] and fine-grained refinements [108]. Another approach aims to edit multiview images with 3D consistency [34, 88, 90], introducing attention-based modules [88, 90] and autoregressive editing [34], adding complexity to the editing pipeline." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.582, + 0.485, + 0.703 + ], + "angle": 0, + "content": "In this paper, we bridge the gap between 3DGS generation and editing by designing a direct 3DGS generative model. Inspired by the success of 2D diffusion models that enable training-free editing through inversion [13, 52, 54], we argue that a direct 3DGS model can similarly enable training-free editing. This approach integrates editing without requiring complex pipelines to lift 2D priors, making generation and editing adaptable within a unified framework." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.486, + 0.903 + ], + "angle": 0, + "content": "We introduce SplatFlow, which consists of two main components: (1) a multi-view rectified flow (RF) model and (2) a Gaussian Splitting Decoder (GSDecoder). Similar to Latent Diffusion Models [69], our multi-view RF model operates in the latent space and is trained to simultaneously generate multi-view images, depths, and camera poses conditioned on text prompts. This joint modeling enables SplatFlow to effectively handle the challenges of real-world scene generation, such as various scene scales and camera trajectories, unlike synthetic object datasets [15, 16]. The GSDecoder translates the latent outputs into 3DGS, based on efficient feed-forward methods [4, 9]. To ensure compatibility with other generative models, particularly Stable Diffusion 3 [19]," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.122 + ], + "angle": 0, + "content": "we incorporate a fixed pre-trained encoder, allowing flexible cross-model usage during sampling." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.123, + 0.909, + 0.259 + ], + "angle": 0, + "content": "Finally, SplatFlow utilizes training-free inversion [13, 52, 54] and inpainting techniques [11, 51, 70, 71, 77] for 3DGS editing and various 3D tasks. For example, it enables object editing [7, 58], novel view synthesis [61, 68, 97], and camera pose estimation [101] by jointly modeling images, depths, and poses to infer missing elements, highlighting its versatility. We validated its capabilities using the MVImgNet [98] and DL3DV-7K [45] datasets, assessing its performance in 3D generation, editing, and inpainting-based tasks." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.272, + 0.665, + 0.288 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.297, + 0.837, + 0.312 + ], + "angle": 0, + "content": "2.1. Diffusion Models and Rectified Flows" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.319, + 0.909, + 0.56 + ], + "angle": 0, + "content": "Diffusion Models Diffusion models [26, 76, 79] generate data through iterative denoising and have become standard in generative modeling across images [17, 21, 37, 60, 62], videos [27], and text [40]. Trained on large-scale text-to-image datasets, they capture complex semantic relationships, producing high-quality images [6, 56, 65, 69, 72]. An advantage is their adaptability to various tasks in a plug-and-play manner [20, 22, 78], especially inpainting [51, 70] and editing [23, 35]. In painting can be framed as a linear inverse problem [14], for which effective posterior sampling methods have been proposed [11, 51, 70, 71, 77]. For editing, inversion methods [13, 52, 54] produce structured noise by inverting images, allowing new prompts to be effectively processed. Building on these training-free techniques, we extend them to our 3DGS model, adapting RePaint [51] and SDEdit [52] within rectified flow models." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.567, + 0.911, + 0.75 + ], + "angle": 0, + "content": "Rectified Flow Models Rectified Flow (RF) models [1, 46, 47] represent an alternative approach to traditional diffusion models by utilizing a straight-line path from data to noise instead of the typical curved forward process, aiming to simplify sampling and reduce computational costs. This linear path theoretically enables faster sampling by minimizing error accumulation across steps, and recent advances, such as Stable Diffusion 3 (SD3) [19], have successfully scaled RF in text-to-image generation, demonstrating performance that surpasses traditional diffusion models. Our SplatFlow framework builds on the capabilities of SD3, integrating RF to achieve efficient 3DGS generation." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.758, + 0.722, + 0.773 + ], + "angle": 0, + "content": "2.2. 3D Generative Models" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.781, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Neural Scene Representation Various neural scene representations have been proposed for 3D generation tasks. Early works use explicit representations, enabling synthesis of 3D point clouds [57] and shapes [83, 106], but struggle with realistic scene generation due to their rendering mechanisms. To address this, NeRF [53] is employed for a realistic 3D generation with per-scene optimization [66], but its inefficient rendering is unsuitable for a fast generation." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "21525" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.152 + ], + "angle": 0, + "content": "Recently, 3DGS [32] enables real-time rendering, accelerating optimization-based 3D generation [10, 44, 82]. Building on this progress, we leverage 3DGS's efficient rendering for fast training of a generative model on large datasets." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.156, + 0.486, + 0.429 + ], + "angle": 0, + "content": "Lifting Multi-View Generation Advances in image generation models [6, 65] and large-scale 3D object datasets [15, 16] make multi-view generation compelling compared to optimization-based generation [82]. The key idea is to learn 3D priors as conditions [49, 50, 89] or attentions [74, 85] from 3D datasets while leveraging text-to-image diffusion models [69] trained on billions of text-image pairs [73]. This enables the synthesis of multi-view consistent images from an image or text prompt, which is then fed into reconstruction models [5, 81] to directly generate neural scene representations like 3DGS. One critical success factor in these methods is that the training dataset provides rendered images from consistent, canonical viewpoints, allowing instant access to desired views regardless of the 3D object. In this work, we focus on more challenging unbounded 3D real-world scene generation, where each scene varies in scale and camera trajectory. To handle diverse trajectories, our model learns a joint distribution of camera poses and 3D scenes." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.432, + 0.486, + 0.658 + ], + "angle": 0, + "content": "Text-to-3D Scene Generation Generating 3D scenes from text prompts is challenging due to the unbounded scale of scenes, complicating scale determination. One straightforward approach is to create a scene covered by a single image [75, 107] generated from text-to-image diffusion models [6, 65]. Another line of work [12, 28] generates wider scenes using user-provided camera trajectories, but these methods are time-consuming and produce subpar quality due to reliance on 2D generative models. Recently, Director3D [41], the work most similar to ours, generates camera poses from text and then creates multi-view images decoded into 3DGS. In contrast, SplatFlow jointly learns the distribution of camera poses and multi-view images from text, enhancing generation quality and enabling flexible 3D scene editing in a unified framework." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.667, + 0.214, + 0.684 + ], + "angle": 0, + "content": "2.3. 3D Editing" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.485, + 0.795 + ], + "angle": 0, + "content": "In recent years, NeRF-based methods [2, 29, 30, 33, 59, 63, 67, 95, 96] have gained popularity for 3D editing, surpassing point clouds and meshes [86, 103, 105]. However, due to the inefficient rendering, these methods are time-consuming and lack precise control. Conversely, incorporating 3DGS into 3D editing [8, 87] improves effectiveness, speed, and control." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.486, + 0.902 + ], + "angle": 0, + "content": "For guiding 3D editing, several methods have leveraged 2D diffusion models [26, 69, 76, 79] for structured 3D editing, but it often requires additional refinement stages. For instance, GSEdit [58] and Xiao et al. [91] apply texture enhancement, while TIP-Editor [108] adds fine-grained adjustments to reduce SDS artifacts. Another line of research in 3D editing emphasizes achieving multi-view consistency" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.908, + 0.198 + ], + "angle": 0, + "content": "by sequentially updating each view. For example, Gauss-Ctrl [90] and 3DEdit [88] utilize attention-based modules to propagate edits made from one perspective across multiple views, ensuring that changes are consistently reflected throughout the 3D scene. Similarly, 3D-Ego [34] adopts an autoregressive editing framework, updating each view to enhance multi-view consistency." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.199, + 0.908, + 0.259 + ], + "angle": 0, + "content": "Unlike prior methods that require complex procedures to lift 2D priors for 3D consistency, SplatFlow directly models multi-view consistency within the latent space, efficiently decoding edits through our Gaussian Splitting Decoder." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.269, + 0.78, + 0.287 + ], + "angle": 0, + "content": "3. Preliminary: Rectified Flows" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.295, + 0.907, + 0.356 + ], + "angle": 0, + "content": "Rectified Flows (RFs) [1, 46, 47] consider generative models that construct a noise distribution \\( q_{0} \\) and a time-varying vector field \\( v_{t}(\\boldsymbol{x}_{t}) \\) to sample data \\( x_{0} \\) from a data distribution \\( p_{0} \\) via an ordinary differential equation (ODE) as:" + }, + { + "type": "equation", + "bbox": [ + 0.57, + 0.363, + 0.907, + 0.38 + ], + "angle": 0, + "content": "\\[\n\\mathrm {d} X _ {t} = v \\left(X _ {t}\\right) \\mathrm {d} t, \\quad X _ {0} \\sim q _ {0}, \\quad t \\in [ 0, 1 ], \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.387, + 0.908, + 0.432 + ], + "angle": 0, + "content": "where \\(q_{0}\\) typically follows a Gaussian distribution \\(\\mathcal{N}(0,I)\\). RF employs a neural network with parameters \\(\\theta\\) for modeling the vector field as \\(v_{t}(X_{t}) = -u_{\\theta}(X_{t},1 - t)\\)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.433, + 0.909, + 0.599 + ], + "angle": 0, + "content": "To train a neural network as the vector field, RFs utilize paired samples from \\( p_0 \\) and \\( q_0 \\) (here simplified as \\( p_1 \\)), along a straight line path \\( Y_t = tY_1 + (1 - t)Y_0 \\). This setup produces the marginal distribution of \\( Y_t \\) as \\( p_t(\\pmb{y}_t) = \\int p_t(\\pmb{y}_t|\\pmb{y}_1)p_1(\\pmb{y}_1)\\mathrm{d}\\pmb{y}_1 \\). Starting from an initial \\( Y_0 = \\pmb{y}_0 \\) and moving to a final state \\( Y_1 = \\pmb{y}_1 \\), a straight line path induces an ODE expressed as \\( \\mathrm{d}Y_t = u_t(Y_t|Y_1)\\mathrm{d}t \\) where the vector field \\( u_t(Y_t|Y_1) \\) represents the conditional change from \\( \\pmb{y}_0 \\) to \\( \\pmb{y}_1 \\), given by \\( u_t(Y_t|Y_1) = \\pmb{y}_1 - \\pmb{y}_0 \\). The marginal vector field \\( u_t(\\pmb{y}_t) \\) is computed by averaging over the conditional vector field as [46]:" + }, + { + "type": "equation", + "bbox": [ + 0.561, + 0.605, + 0.907, + 0.639 + ], + "angle": 0, + "content": "\\[\nu _ {t} (\\boldsymbol {y} _ {t}) = \\int u _ {t} (\\boldsymbol {y} _ {t} | \\boldsymbol {y} _ {1}) \\frac {p _ {t} (\\boldsymbol {y} _ {t} | \\boldsymbol {y} _ {1})}{p _ {t} (\\boldsymbol {y} _ {t})} p _ {1} (\\boldsymbol {y} _ {1}) \\mathrm {d} \\boldsymbol {y} _ {1}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.643, + 0.908, + 0.674 + ], + "angle": 0, + "content": "To approximate \\( u_{t}(\\boldsymbol{y}_{t}) \\), the neural network is trained with a flow-matching object, defined by:" + }, + { + "type": "equation", + "bbox": [ + 0.577, + 0.681, + 0.907, + 0.707 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {F M}} := \\mathbb {E} _ {t, Y _ {t}} \\left[ \\left| \\left| u _ {t} \\left(Y _ {t}\\right) - u _ {\\theta} \\left(Y _ {t}, t\\right) \\right| \\right| _ {2} ^ {2} \\right]. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.712, + 0.908, + 0.758 + ], + "angle": 0, + "content": "Since directly computing this objective can be computationally challenging, an alternative approach, conditional flow matching, is used to simplify training:" + }, + { + "type": "equation", + "bbox": [ + 0.536, + 0.764, + 0.907, + 0.79 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {C F M}} := \\mathbb {E} _ {t, Y _ {t}, Y _ {1} \\sim p _ {1}} \\left[ \\| u _ {t} \\left(Y _ {t} \\mid Y _ {1}\\right) - u _ {\\theta} \\left(Y _ {t}, t\\right) \\| _ {2} ^ {2} \\right]. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.909, + 0.901 + ], + "angle": 0, + "content": "The gradients of \\( L_{\\mathrm{CFM}} \\) and \\( L_{\\mathrm{FM}} \\) are theoretically identical [46], making them equivalent in training. However, \\( L_{\\mathrm{CFM}} \\) offers better computational efficiency and is therefore preferred. Ultimately, the learned vector field for the ODE in Eq. 1 is given by \\( v_{t}(X_{t}) = -u_{\\theta}(X_{t},1 - t) \\), allowing RFs to generate samples that follow the data distribution using the learned vector field." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "21526" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.114, + 0.089, + 0.882, + 0.312 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.313, + 0.908, + 0.357 + ], + "angle": 0, + "content": "Figure 2. Overview of SplatFlow. SplatFlow consists of two main components: a multi-view Rectified Flow (Section 4.3) model and a Gaussian Splat Decoder (Section 4.2). Conditioned on text, RF model generates multi-view latents—including image, depth, and Plücker ray coordinates. After an optimization process to estimate camera poses, the GSDecoder decodes these latents into pixel-aligned 3DGS." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.366, + 0.226, + 0.384 + ], + "angle": 0, + "content": "4. Methodology" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.391, + 0.303, + 0.407 + ], + "angle": 0, + "content": "4.1. Overview of SplatFlow" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.413, + 0.484, + 0.565 + ], + "angle": 0, + "content": "We introduce SplatFlow, a 3DGS generation model designed to perform various 3D editing and 3D-related tasks in a training-free manner, despite being trained solely for generation. As shown in Fig. 2, our model consists of two main components: 1) a multi-view Rectified Flow (RF) model and 2) a Gaussian Splitting Decoder (GSDecoder). First, the multi-view RF model generates multi-view camera poses, depths, and images conditioned on text prompts. Then, the GSDecoder translates these latent representations into pixel-aligned 3DGS, constructing a real-world scene structure." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.565, + 0.484, + 0.655 + ], + "angle": 0, + "content": "Also, both the multi-view RF model and GSDecoder operate within a latent space, adopting a Latent Diffusion Model (LDM) [69]. By freezing and sharing the encoder of Stable Diffusion 3 (SD3) [19], our design enables the RF model to share the latent space with SD3, allowing us to leverage its knowledge for enhanced generation and editing capabilities." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.656, + 0.484, + 0.716 + ], + "angle": 0, + "content": "Finally, our model leverages a training-free approach [51, 52], enabling a broad range of editing applications, such as object replacement, as well as 3D-related tasks including camera pose estimation and novel view synthesis." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.717, + 0.484, + 0.822 + ], + "angle": 0, + "content": "In the following sections, we present each component in detail. Section 4.2 introduces the GSDecoder, explaining details in constructing pixel-aligned 3DGS. Section 4.3 describes the multi-view RF model, detailing how it jointly generates camera poses, depths, and images. Finally, Section 4.4 outlines the sampling process and demonstrates how our model applies training-free techniques." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.833, + 0.446, + 0.85 + ], + "angle": 0, + "content": "4.2. Gaussian Splitting Decoder (GSDecoder)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Recently, feed-forward 3DGS methods [4, 9, 80, 81] have enabled fast 3DGS reconstruction from sparse views by training on large datasets, achieving much faster reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.367, + 0.907, + 0.473 + ], + "angle": 0, + "content": "than per-scene optimization methods [32]. Leveraging this advantage, our GSDecoder \\(G_{\\phi}\\) (parameterized by \\(\\phi\\)) is designed to decode pixel-aligned 3DGS from latent representations of \\(K\\) sparse views, given their corresponding camera poses \\(\\mathcal{P} = \\{\\pmb{P}_i\\}_{i=1}^K\\) with \\(\\pmb{P}_i = \\pmb{K}_i[\\pmb{R}_i|\\pmb{T}_i]\\), where \\(\\pmb{K}_i, \\pmb{R}_i\\), and \\(\\pmb{T}_i\\) denote the intrinsic matrix, rotation matrix, and translation vector of the \\(i\\)-th view, respectively." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.474, + 0.909, + 0.671 + ], + "angle": 0, + "content": "A straightforward approach is to design \\( G_{\\phi} \\) to directly output the 3D Gaussian parameters from the image latents obtained through the encoder \\( \\mathcal{E} \\) as \\( G_{\\phi}(\\{\\langle \\mathcal{E}(I_i), P_i \\rangle \\}_{i=1}^K) = \\{(\\pmb{\\mu}_j, \\alpha_j, \\pmb{\\Sigma}_i, c_j)\\}_{j=1}^{H \\times W \\times K} \\), where each 3D Gaussian parameter includes position \\( \\pmb{\\mu}_j \\), opacity \\( \\alpha_j \\), covariance \\( \\pmb{\\Sigma}_j \\), and color \\( c_j \\) in a pixel-aligned manner with each \\( i \\)-th image \\( I_i \\in \\mathbb{R}^{H \\times W} \\). However, as the image latents are produced through a shared and frozen encoder (to leverage compatibility with 2D generative models during sampling), it might lose fine-grained spatial details, leading to abstract representations. This can limit the preservation of specific 3D structural information. To address these limitations, we propose two improvements to enhance GSDecoder's performance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.68, + 0.909, + 0.817 + ], + "angle": 0, + "content": "Depth latent integration As demonstrated in [31], the encoder \\(\\mathcal{E}\\) can effectively encode depth maps. Building on this, we incorporate the depth latents \\(\\{\\mathcal{E}(D_i)\\}_{i=1}^K\\) of each depth map \\(D_i\\) as an additional input to the GSDecoder to further enhance 3D structural information. We use DepthAnythingV2 [93] to extract the depth maps and normalize them to the range [-1, 1], stacking them as three channels similar to RGB for constructing \\(D_i\\). We observe that this improves the convergence speed and performance of the GSDecoder." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Adversarial loss Applying adversarial losses has demonstrated improvements in the visual quality of decoded output [18, 69]. However, using adversarial loss during training can cause instability, particularly in the early stages. We observe that applying the vision-aided loss [36] after the GSDe" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21527" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.123 + ], + "angle": 0, + "content": "coder has reached a certain level of convergence significantly improves visual quality without destabilizing training." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.126, + 0.484, + 0.324 + ], + "angle": 0, + "content": "Training In summary, our GSDecoder predicts 3DGS \\(\\{(\\pmb{\\mu}_j,\\alpha_j,\\pmb{\\Sigma}_j,\\pmb{c}_j)\\}_{j = 1}^{H\\times W\\times K}\\) from \\(\\{(I_i,D_i,P_i)\\}_{i = 1}^K\\). The model is trained with a combined loss function consisting of LPIPS [104], Mean Squared Error loss, and vision-aided loss [36], computed between the target images of novel views and rendered views from 3DGS. Specifically, the vision-aided loss is activated only after a certain threshold iteration to improve stability and perceptual quality. Additionally, the architecture is initialized with the parameters of the Stable Diffusion 3 decoder, incorporating cross-view attention mechanisms, and the channel dimensions are increased to meet the specific requirements of 3D Gaussian Splatting. Further details on the GSDecoder are provided in Appendix A." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.332, + 0.383, + 0.346 + ], + "angle": 0, + "content": "4.3. Multi-View Rectified Flow Model" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.353, + 0.483, + 0.565 + ], + "angle": 0, + "content": "With the GSDecoder in place, the 3D Gaussian Splitting (3DGS) generation process becomes feasible if we can produce consistent multi-view latents of images \\(\\mathcal{E}(I_i)\\), depth maps \\(\\mathcal{E}(D_i)\\), and camera poses \\(P_{i}\\) for \\(i = 1,\\dots ,K\\). Our multi-view rectified flow (RF) model is trained to sample from the joint distribution of these latents, conditioned on text \\(C\\) as \\(p(\\{I_i\\}_{i = 1}^K,\\{D_i\\}_{i = 1}^K,\\{P_i\\}_{i = 1}^K |C)\\), allowing for simultaneous generation of multi-view images, depths, and poses. The rationale for modeling a joint distribution instead of each component separately is twofold. First, this approach allows the model to handle various tasks by reformulating them as inpainting problems, as discussed in Section 4.4. Second, joint modeling is crucial for real-world scenes, which require adaptive camera poses tailored to each scene [41]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.566, + 0.484, + 0.716 + ], + "angle": 0, + "content": "To achieve this, our multi-view RF model treats the concatenation of multi-view image latents, depth latents, and Plücker ray coordinates [64, 101] along the channel dimension as input data. Formally, each ray \\(\\pmb{r}_i\\) is represented as \\(\\langle \\pmb{d}_i, \\pmb{m}_i \\rangle\\), where \\(\\pmb{d}_i = \\pmb{R}_i^\\top \\pmb{K}_i^{-1} \\pmb{w}_i\\) denotes the direction vector and \\(\\pmb{m}_i = (-\\pmb{R}_i^\\top \\pmb{T}_i) \\times \\pmb{d}_i\\) represents the moment vector, with \\(\\pmb{w}_i\\) referring to the 2D pixel coordinates. By matching the spatial resolution of \\(\\pmb{w}_i\\) with the image and depth latents, each represented as \\(\\mathbb{R}^{n \\times h \\times w}\\), \\(\\pmb{r}_i\\) can be expressed as \\(\\mathbb{R}^{6 \\times h \\times w}\\), enabling concatenation along the channel axis." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.717, + 0.483, + 0.838 + ], + "angle": 0, + "content": "Through this setup, we define \\(Y_{0}\\) as the multi-view latents \\(\\mathbf{X}\\), where each element \\(\\mathbf{X}_i = \\langle \\mathcal{E}(\\mathbf{I}_i),\\mathcal{E}(\\mathbf{D}_i),\\mathbf{r}_i\\rangle\\) results in \\(Y_{0} = (X_{1},X_{2},\\ldots ,X_{K})\\in \\mathbb{R}^{K\\times (2n + 6)\\times h\\times w}\\). We then train the multi-view RF model \\(u_{\\theta}\\) with the conditional flow matching loss defined in Eq. 4. Specifically, we fine-tune Stable Diffusion 3 [19] by adjusting the channel dimensions in the input and output layers and incorporating cross-view attention. Further details are provided in Appendix A." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Sampling process To recover the camera pose from ray \\( \\boldsymbol{r}_i \\), we follow the procedure in [101] and further refine the pose parameters to ensure that multiple views share the same intrinsic. However, we find that naively solving the ODE" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.213 + ], + "angle": 0, + "content": "can lead to degraded camera pose accuracy. We hypothesize this degradation results from deviations from the camera pose manifold, which introduces errors in camera trajectory estimation. To address this, at each sampling step \\( t = t_k \\), we predict the sampling destination at \\( t = 0 \\), recover the camera poses from this prediction, and reconstruct the associated camera rays. These rays are then forwarded to \\( t = t_k \\) to maintain the solution within the valid ray manifold at \\( t = t_k \\)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.217, + 0.909, + 0.294 + ], + "angle": 0, + "content": "Stable Diffusion 3 guidance To improve multi-view image generation quality, we integrate vector fields from Stable Diffusion 3, known for robust generalization, with our multiview RF model to solve the ODE for randomly selected view's image latents at specific sampling steps." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.303, + 0.771, + 0.317 + ], + "angle": 0, + "content": "4.4. Inference Given Partial Data" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.324, + 0.907, + 0.431 + ], + "angle": 0, + "content": "In this section, we demonstrate how SplatFlow can be effectively utilized for 3D editing and applications by applying training-free inversion techniques [13, 52, 54] and inpainting methods [11, 51, 70, 71, 77]. We modify SDEdit [52] inversion for rectified flow-based inversion and adopt the RePaint [51] to support rectified flow inpainting. Additional technical details are provided in Appendix A." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.434, + 0.909, + 0.602 + ], + "angle": 0, + "content": "3DGS editing Our GSDecoder is designed to directly decode 3DGS, enabling efficient editing by modifying only the multi-view latents. Additionally, with the RF model's joint modeling of multi-view image latents, edits can be achieved without extra modules (e.g., cross-view attention mechanisms [88, 90]), streamlining the process through simple inversion and sampling. Specifically, given a target text and input multi-view latents, we apply a modified SDEdit for inversion at a chosen timestep \\( t_k \\). Conditioned on the target text, we then generate the edited multi-view latents by solving the ODE in Eq. 1 following our sampling procedure." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.605, + 0.909, + 0.742 + ], + "angle": 0, + "content": "Inpainting application Our model is also applicable to various 3D tasks, as it jointly models multi-view images, depths, and camera poses. This joint modeling allows known parts of the data (e.g., some views or camera parameters) to act as constraints, enabling the prediction of unknown parts through inpainting. In this work, we focus on two specific tasks: 1) camera pose estimation from multi-view images and depths, and 2) novel view synthesis using a subset of the \\(K\\) multi-view latents and camera poses for novel viewpoints." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.755, + 0.719, + 0.772 + ], + "angle": 0, + "content": "5. Experimental Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.909, + 0.903 + ], + "angle": 0, + "content": "In this section, we demonstrate the capabilities of our proposed SplatFlow across 3DGS generation, 3DGS editing, and various 3D tasks. We begin by detailing the implementation of SplatFlow in Section 5.1. Following this, we present comparative results on 3DGS generation in real-world datasets in Section 5.2. In Section 5.3, we evaluate the 3DGS editing performance of SplatFlow against other methods. Finally, Section 5.4 showcases the versatility of our" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "21528" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.1, + 0.086, + 0.899, + 0.66 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.665, + 0.908, + 0.709 + ], + "angle": 0, + "content": "Figure 3. Qualitative results in text-to-3DGS generation on MVImgNet and DL3DV validation sets. The first two rows are rendered scenes from the MVImgNet dataset, while the last two rows are from the DL3DV dataset. Our SplatFlow produces cohesive and realistic scenes with sharp details, accurately capturing the intricacies of real-world environments and accommodating diverse camera trajectories." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.715, + 0.488, + 0.778 + ], + "angle": 0, + "content": "multi-view RF model in handling diverse 3D tasks through training-free inpainting techniques. Extensive information on all our experiments is provided in Appendix B with additional results that could not be included here in the Appendix." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.788, + 0.291, + 0.804 + ], + "angle": 0, + "content": "5.1. Experimental Setups" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.486, + 0.903 + ], + "angle": 0, + "content": "Due to space constraints, we provide a concise overview of our experimental setups here. We used the MVImgNet [98] and DL3DV-7K [45] datasets, both containing real-world scenes with images and corresponding camera poses. For evaluation, we utilized 1.25K scenes from MVImgNet for each task and 300 sequences from DL3DV-7K, using the" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.716, + 0.909, + 0.778 + ], + "angle": 0, + "content": "remaining for training. The multi-view RF model and GSDecoder were trained with \\( K = 8 \\)-view setup. Text descriptions for scenes were extracted using the Llava-One Vision Qwen 7B [38]. We used SDS++ [41] for further improvements." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.788, + 0.833, + 0.803 + ], + "angle": 0, + "content": "5.2. Results on Text-to-3DGS Generation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.909, + 0.903 + ], + "angle": 0, + "content": "We evaluated our model against various baselines on the DL3DV and MVImgNet validation sets for real-world scene text-to-3DGS generation. Our evaluation protocol utilized the rendered image of the generated 3DGS. We used the FID score [25] to evaluate image quality and the CLIP score [24] to measure alignment with text prompts." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "21529" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.089, + 0.482, + 0.176 + ], + "angle": 0, + "content": "
MethodMVImgNet [98]DL3DV [45]
FID-10K↓CLIPScore↑FID-2.4K↓CLIPScore↑
Director3D [41]39.5530.4888.4430.04
Director3D (w/ SDS++) [41]41.8031.0095.8831.68
SplatFlow34.8531.4379.9130.06
SplatFlow (w/ SDS++)35.4632.3085.3131.90
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.18, + 0.483, + 0.222 + ], + "angle": 0, + "content": "Table 1. Quantitative results in text-to-3DGS generation on the MVImgNet and DL3DV datasets. We compared our SplatFlow with and without the SDS++ [41], against Director3D [41]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.235, + 0.485, + 0.356 + ], + "angle": 0, + "content": "For quantitative comparison, we adopt Director3D [41], which leverages the full sequences of MVImgNet, DL3DV-10K, and LAION [73] datasets. In contrast, our model was trained on a notably smaller dataset, utilizing only a portion of MVImgNet and DL3DV-7K, without access to the full DL3DV-10K or LAION data. For qualitative results, we also compared against other scene generation methods, including LucidDreamer [12], DreamScene [39], and Director3D [41]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.364, + 0.485, + 0.637 + ], + "angle": 0, + "content": "Quantitative results Table 1 presents the quantitative comparison of SplatFlow and the baseline methods on both MVImgNet and DL3DV datasets. SplatFlow consistently outperforms Director3D [41] in terms of FID and CLIP score, demonstrating its effectiveness in generating high-quality images and maintaining strong alignment with text prompts, even with a smaller training dataset. Specifically, on MVImgNet, SplatFlow achieves a significantly lower FID (34.85 vs. 39.55 for Director3D), indicating better image quality, and a higher CLIP score (31.43 vs. 30.48), showing improved text-to-image alignment. Similarly, on DL3DV, SplatFlow achieves an FID of 79.91, outperforming Director3D's 88.44, with a comparable CLIP score. Notably, with the addition of \\(\\mathrm{SDS}++\\) for refinement, both SplatFlow and Director3D show increased CLIP scores, further enhancing text alignment, though at a minor trade-off in FID. This highlights the effectiveness of \\(\\mathrm{SDS}++\\) for boosting alignment with text prompts across different 3DGS generation models." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.645, + 0.486, + 0.902 + ], + "angle": 0, + "content": "Qualitative results Figure 3 shows the qualitative comparisons between our method and baselines on DL3DV and MVImgNet validation sets. Overall, SplatFlow demonstrates clear advantages over the baseline methods. First, DreamScene [39], which relies on Score Distillation Sampling (SDS), often struggles to generate realistic real-world scenes, resulting in outputs that lack natural textures and environmental details. Second, LucidDreamer [12], which is based on single-view inpainting, faces challenges when handling scenes with large camera trajectory variations, leading to inconsistent or incomplete views. Third, Director3D [41] tends to produce slightly blurred outputs, lacking the sharpness needed for intricate details. In contrast, our SplatFlow can generate coherent and realistic scenes with clear details, effectively capturing the complexities of real-world environments and various camera trajectories. Additional qualitative results of SplatFlow can be found in Appendix C." + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.096, + 0.566, + 0.149 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.577, + 0.096, + 0.678, + 0.15 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.691, + 0.096, + 0.791, + 0.15 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.804, + 0.096, + 0.905, + 0.15 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.604, + 0.152, + 0.848, + 0.164 + ], + "angle": 0, + "content": "\"green inflatable ring\" -> \"colorful red rubber ball\"" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.171, + 0.566, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.577, + 0.171, + 0.678, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.691, + 0.171, + 0.792, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.804, + 0.171, + 0.905, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.602, + 0.214, + 0.862, + 0.227 + ], + "angle": 0, + "content": "\"collection of red chili peppers\" -> \"cluster of bananas\"" + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.23, + 0.905, + 0.259 + ], + "angle": 0, + "content": "Figure 4. Qualitative results in 3D editing with MVInpainter [3] and DGE [7]. We show rendered scenes except for MVInpainter." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.276, + 0.735, + 0.292 + ], + "angle": 0, + "content": "5.3. Result on 3DGS Editing" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.298, + 0.699, + 0.388 + ], + "angle": 0, + "content": "We demonstrate the effectiveness of SplatFlow in the 3DGS editing task by creating a benchmark with 100 sampled scenes from the MVImgNet validation set." + }, + { + "type": "table", + "bbox": [ + 0.708, + 0.3, + 0.903, + 0.363 + ], + "angle": 0, + "content": "
MethodCLIPScore↑CLIP D-sim↑
DGE [7]27.430.102
SplatFlow28.470.169
+) SDS++31.300.224
" + }, + { + "type": "table_caption", + "bbox": [ + 0.708, + 0.368, + 0.905, + 0.382 + ], + "angle": 0, + "content": "Table 2. 3D object replacement." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.388, + 0.907, + 0.555 + ], + "angle": 0, + "content": "Images and texts were provided to GPT-4 to identify main objects and generate text captions for edited images, which were then used for 3D object replacement. For editing, we used the same mask in [3] for masking out the object, then inpainted this area based on the provided editing captions. We compared this pipeline with an optimization-based 3DGS editing method, DGE [7], and conducted qualitative comparisons with MVInpainter [3], which propagates an edited image from pretrained generative models across multiple views. Following standard protocols [8, 29], we measured the CLIP score and directional similarity (CLIP D-sim)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.559, + 0.909, + 0.635 + ], + "angle": 0, + "content": "Quantitative results As shown in Table 2, SplatFlow achieves higher CLIPScore and CLIP D-sim than DGE [7], indicating more accurate and effective editing. SDS++ refinement further boosts performance, demonstrating enhanced alignment with target edits." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.639, + 0.909, + 0.76 + ], + "angle": 0, + "content": "Qualitative results Figure 4 demonstrates that SplatFlow performs comparably to MVInpainter, a specialized 2D inpainting method that leverages an already edited high-quality image. On the other hand, DGE struggles to replace objects fully, often just altering styles instead of achieving complete transformations. This underscores SplatFlow's superior ability in precise 3D object replacement. Please refer to Appendix C for further qualitative results." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.769, + 0.803, + 0.785 + ], + "angle": 0, + "content": "5.4. Result on Inpainting Application" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.791, + 0.906, + 0.837 + ], + "angle": 0, + "content": "Here, we present the adaptability of our multi-view RF model to camera pose estimation and novel-view synthesis via the inpainting technique on the MVImgNet dataset [98]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.841, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Camera pose estimation To demonstrate the effectiveness of our multi-view RF model in camera pose estimation, we primarily compare our approach to RayDiffusion [101] and RelPose++ [43]. We conducted comparisons using an" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "21530" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.089, + 0.48, + 0.2 + ], + "angle": 0, + "content": "
MethodRotation↑Camera Center↑
@5@10@15@0.05@0.1@0.2
RelPose++ [43]19.437.751.40.612.555.0
Ray Regression [101]10.425.650.115.347.982.9
Ray Diffusion [101]17.538.759.624.160.987.6
SplatFlow (w/ depth)26.852.659.362.391.699.4
SplatFlow (w/o depth)28.854.563.964.994.099.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.202, + 0.484, + 0.245 + ], + "angle": 0, + "content": "Table 3. Results in camera pose estimation on MVImgNet validation set. @Q represents the accuracy threshold for rotations (degrees) and camera centers (units)." + }, + { + "type": "image", + "bbox": [ + 0.104, + 0.254, + 0.338, + 0.457 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.356, + 0.258, + 0.468, + 0.457 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.46, + 0.483, + 0.503 + ], + "angle": 0, + "content": "Figure 5. Qualitative results for camera pose estimation. Camera poses are estimated from multi-view images. Image border colors match each camera, with black cameras indicating GT poses." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.515, + 0.484, + 0.606 + ], + "angle": 0, + "content": "8-view setup, inpainting the ray latent with and without the depth latent from DepthAnythingV2 [93]. We used rotation accuracy@ \\( Q_{R} \\) and translation accuracy@ \\( Q_{T} \\) as evaluation metrics, measuring the proportion of relative camera rotations within \\( Q_{R} \\) degrees of ground truth and camera centers within \\( Q_{T} \\) units of scene scale, respectively." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.607, + 0.484, + 0.728 + ], + "angle": 0, + "content": "Table 3 and Fig. 5 present the quantitative and qualitative results. Overall, the multi-view RF model achieves superior performance across most scenarios, outperforming the baseline methods. Interestingly, the multi-view RF model achieves higher accuracy without depth latents than with them. We hypothesize that this outcome is due to the joint generation of depth and camera poses, which may yield more detailed and contextually enriched depth estimates." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Novel view-synthesis We evaluated the performance of our multi-view RF model in novel view synthesis under two scenarios: 1) interpolation, with the uniformly sampled input views, and 2) extrapolation, with the input views positioned in the center. This comparison validates our multi-view RF model's capability of 3D reasoning based on relative viewpoints and evaluates its effectiveness in generating novel views. We evaluated PSNR, SSIM, and LPIPS [104] against the ground-truth images for reconstruction quality, as well as absolute relative difference and \\(\\delta_{1}\\) accuracy using depth maps produced by DepthAnythingV2 [93]." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.089, + 0.905, + 0.216 + ], + "angle": 0, + "content": "
TypeRGBDepth
PSNR↑SSIM↑LPIPS↓AbsRel↓δ1↑
Interpolation (N=2)14.730.5710.6480.5880.731
Interpolation (N=4)17.050.5900.5510.4980.761
Interpolation (N=6)18.820.6260.4830.4150.775
Extrapolation (N=2)15.150.5770.6270.7710.715
Extrapolation (N=4)16.800.5950.5540.6790.727
Extrapolation (N=6)17.960.6130.5030.6020.747
" + }, + { + "type": "table_caption", + "bbox": [ + 0.514, + 0.217, + 0.905, + 0.26 + ], + "angle": 0, + "content": "Table 4. Novel view synthesis results on MVImgNet. We use \\(N\\) input views to synthesize \\(K - N\\) novel views, with uniformly sampled views for interpolation and central views for extrapolation." + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.272, + 0.905, + 0.512 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.514, + 0.907, + 0.543 + ], + "angle": 0, + "content": "Figure 6. Qualitative results for novel view synthesis. Novel view synthesis is performed from the red-box images and depths." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.554, + 0.907, + 0.66 + ], + "angle": 0, + "content": "As shown in Fig. 6, our multi-view RF model successfully generates novel views for both scenarios. Also, Table 4 shows that interpolation outperforms extrapolation in both RGB reconstruction and depth estimation as the number of input views increases, indicating that interpolation leverages 3D reasoning for accurate novel view generation, whereas extrapolation focuses on generating diverse novel views." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.678, + 0.634, + 0.693 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.909, + 0.901 + ], + "angle": 0, + "content": "In this paper, we have introduced SplatFlow, a novel framework that unifies 3D Gaussian Splitting (3DGS) generation and editing within a single, efficient model. By designing a multi-view Rectified Flow (RF) model alongside a Gaussian Splitting Decoder (GSDecoder), we achieved direct 3DGS generation capable of addressing real-world challenges, such as varying scales and complex camera trajectories. Our approach utilizes joint modeling of multi-view images, depths, and camera poses, enabling seamless integration of training-free inversion and inpainting techniques for a diverse set of 3D tasks. As we present the versatility of our SplatFlow, we believe SplatFlow represents an important step toward a versatile foundational model for 3D content." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "21531" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.115, + 0.484, + 0.156 + ], + "angle": 0, + "content": "[1] Michael S Albergo and Eric Vanden-Eijnden. Building normalizing flows with stochastic interpolants. arXiv preprint arXiv:2209.15571, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.159, + 0.484, + 0.241 + ], + "angle": 0, + "content": "[2] Chong Bao, Yinda Zhang, Bangbang Yang, Tianxing Fan, Zesong Yang, Hujun Bao, Guofeng Zhang, and Zhaopeng Cui. Sine: Semantic-driven image-based nerf editing with prior-guided editing field. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20919-20929, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.245, + 0.484, + 0.299 + ], + "angle": 0, + "content": "[3] Chenjie Cao, Chaohui Yu, Yanwei Fu, Fan Wang, and Xi-angyang Xue. Mvinpainter: Learning multi-view consistent inpainting to bridge 2d and 3d editing. arXiv preprint arXiv:2408.08000, 2024.7" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.302, + 0.483, + 0.37 + ], + "angle": 0, + "content": "[4] David Charatan, Sizhe Lester Li, Andrea Tagliasacchi, and Vincent Sitzmann. pixelsplat: 3d gaussian splats from image pairs for scalable generalizable 3d reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19457-19467, 2024. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.373, + 0.482, + 0.428 + ], + "angle": 0, + "content": "[5] Anpei Chen, Haofei Xu, Stefano Esposito, Siyu Tang, and Andreas Geiger. Lara: Efficient large-baseline radiance fields. In European Conference on Computer Vision, pages 338-355. Springer, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.431, + 0.483, + 0.499 + ], + "angle": 0, + "content": "[6] Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, et al. Pixart-α: Fast training of diffusion transformer for photorealistic text-to-image synthesis. arXiv preprint arXiv:2310.00426, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.502, + 0.482, + 0.543 + ], + "angle": 0, + "content": "[7] Minghao Chen, Iro Laina, and Andrea Vedaldi. Dge: Direct gaussian 3d editing by consistent multi-view editing. arXiv preprint arXiv:2404.18929, 2024. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.545, + 0.483, + 0.628 + ], + "angle": 0, + "content": "[8] Yiwen Chen, Zilong Chen, Chi Zhang, Feng Wang, Xiaofeng Yang, Yikai Wang, Zhongang Cai, Lei Yang, Huaping Liu, and Guosheng Lin. Gaussianeditor: Swift and controllable 3d editing with gaussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21476-21485, 2024. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.631, + 0.483, + 0.699 + ], + "angle": 0, + "content": "[9] Yuedong Chen, Haofei Xu, Chuanxia Zheng, Bohan Zhuang, Marc Pollefeys, Andreas Geiger, Tat-Jen Cham, and Jianfei Cai. Mvsplat: Efficient 3d gaussian splattering from sparse multi-view images. In European Conference on Computer Vision, pages 370-386. Springer, 2024. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.702, + 0.483, + 0.757 + ], + "angle": 0, + "content": "[10] Zilong Chen, Feng Wang, Yikai Wang, and Huaping Liu. Text-to-3d using gaussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21401-21412, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.759, + 0.483, + 0.813 + ], + "angle": 0, + "content": "[11] Hyungjin Chung, Jeongsol Kim, Michael T McCann, Marc L Klasky, and Jong Chul Ye. Diffusion posterior sampling for general noisy inverse problems. arXiv preprint arXiv:2209.14687, 2022. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.816, + 0.483, + 0.87 + ], + "angle": 0, + "content": "[12] Jaeyoung Chung, Suyoung Lee, Hyeongjin Nam, Jaerin Lee, and Kyoung Mu Lee. Luciddreamer: Domain-free generation of 3d gaussian splatting scenes. arXiv preprint arXiv:2311.13384, 2023. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.873, + 0.483, + 0.9 + ], + "angle": 0, + "content": "[13] Guillaume Couairon, Jakob Verbeek, Holger Schwenk, and Matthieu Cord. Diffedit: Diffusion-based seman" + }, + { + "type": "list", + "bbox": [ + 0.102, + 0.115, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.554, + 0.094, + 0.907, + 0.119 + ], + "angle": 0, + "content": "tic image editing with mask guidance. arXiv preprint arXiv:2210.11427, 2022. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.122, + 0.908, + 0.19 + ], + "angle": 0, + "content": "[14] Giannis Daras, Hyungjin Chung, Chieh-Hsin Lai, Yuki Mitsufuji, Jong Chul Ye, Peyman Milanfar, Alexandros G Dimakis, and Mauricio Delbracio. A survey on diffusion models for inverse problems. arXiv preprint arXiv:2410.00083, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.193, + 0.907, + 0.276 + ], + "angle": 0, + "content": "[15] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3d objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13142-13153, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.278, + 0.907, + 0.346 + ], + "angle": 0, + "content": "[16] Matt Deitke, Ruoshi Liu, Matthew Wallingford, Huong Ngo, Oscar Michel, Aditya Kusupati, Alan Fan, Christian Laforte, Vikram Voleti, Samir Yitzhak Gadre, et al. Objverse-xl: A universe of \\(10\\mathrm{m} + 3\\mathrm{d}\\) objects. Advances in Neural Information Processing Systems, 36, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.349, + 0.907, + 0.39 + ], + "angle": 0, + "content": "[17] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.392, + 0.907, + 0.447 + ], + "angle": 0, + "content": "[18] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12873-12883, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.449, + 0.906, + 0.53 + ], + "angle": 0, + "content": "[19] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In *Forty-first International Conference on Machine Learning*, 2024. 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.533, + 0.907, + 0.602 + ], + "angle": 0, + "content": "[20] Hyojun Go, Yunsung Lee, Jin-Young Kim, Seunghyun Lee, Myeongho Jeong, Hyun Seung Lee, and Seungtaek Choi. Towards practical plug-and-play diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1962-1971, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.604, + 0.907, + 0.659 + ], + "angle": 0, + "content": "[21] Hyojun Go, Yunsung Lee, Seunghyun Lee, Shinhyeok Oh, Hyeongdon Moon, and Seungtaek Choi. Addressing negative transfer in diffusion models. Advances in Neural Information Processing Systems, 36, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.661, + 0.907, + 0.715 + ], + "angle": 0, + "content": "[22] Alexandros Graikos, Nikolay Malkin, Nebojsa Jojic, and Dimitris Samaras. Diffusion models as plug-and-play priors. Advances in Neural Information Processing Systems, 35: 14715-14728, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.718, + 0.907, + 0.772 + ], + "angle": 0, + "content": "[23] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.775, + 0.907, + 0.829 + ], + "angle": 0, + "content": "[24] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning. arXiv preprint arXiv:2104.08718, 2021.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.832, + 0.907, + 0.9 + ], + "angle": 0, + "content": "[25] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 6" + }, + { + "type": "list", + "bbox": [ + 0.525, + 0.094, + 0.908, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21532" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.093, + 0.484, + 0.134 + ], + "angle": 0, + "content": "[26] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.134, + 0.484, + 0.204 + ], + "angle": 0, + "content": "[27] Jonathan Ho, William Chan, Chitwan Sahara, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.205, + 0.484, + 0.273 + ], + "angle": 0, + "content": "[28] Lukas Hollein, Ang Cao, Andrew Owens, Justin Johnson, and Matthias Nießner. Text2room: Extracting textured 3d meshes from 2d text-to-image models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7909-7920, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.274, + 0.484, + 0.328 + ], + "angle": 0, + "content": "[29] Hiromichi Kamata, Yuiko Sakuma, Akio Hayakawa, Masato Ishii, and Takuya Narihira. Instruct 3d-to-3d: Text instruction guided 3d-to-3d conversion. arXiv preprint arXiv:2303.15780, 2023. 1, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.33, + 0.484, + 0.385 + ], + "angle": 0, + "content": "[30] Nazmul Karim, Hasan Iqbal, Umar Khalid, Chen Chen, and Jing Hua. Free-editor: zero-shot text-driven 3d scene editing. In European Conference on Computer Vision, pages 436-453. Springer, 2025. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.386, + 0.484, + 0.468 + ], + "angle": 0, + "content": "[31] Bingxin Ke, Anton Obukhov, Shengyu Huang, Nando Metzger, Rodrigo Caye Daudt, and Konrad Schindler. Repurposing diffusion-based image generators for monocular depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9492-9502, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.469, + 0.484, + 0.523 + ], + "angle": 0, + "content": "[32] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splattering for real-time radiance field rendering. ACM Trans. Graph., 42(4):139-1, 2023. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.525, + 0.484, + 0.566 + ], + "angle": 0, + "content": "[33] Umar Khalid, Hasan Iqbal, Nazmul Karim, Jing Hua, and Chen Chen. Latenteditor: Text driven local editing of 3d scenes. arXiv preprint arXiv:2312.09313, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.567, + 0.484, + 0.608 + ], + "angle": 0, + "content": "[34] Umar Khalid, Hasan Iqbal, Azib Farooq, Jing Hua, and Chen Chen. 3dego: 3d editing on the go! In European Conference on Computer Vision, pages 73-89. Springer, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.609, + 0.484, + 0.677 + ], + "angle": 0, + "content": "[35] Gwanghyun Kim, Taesung Kwon, and Jong Chul Ye. Diffusionclip: Text-guided diffusion models for robust image manipulation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2426-2435, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.679, + 0.484, + 0.746 + ], + "angle": 0, + "content": "[36] Nupur Kumari, Richard Zhang, Eli Shechtman, and JunYan Zhu. Ensembling off-the-shelf models for gan training. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10651-10662, 2022. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.748, + 0.484, + 0.815 + ], + "angle": 0, + "content": "[37] Yunsung Lee, JinYoung Kim, Hyojun Go, Myeongho Jeong, Shinhyeok Oh, and Seungtaek Choi. Multi-architecture multi-expert diffusion models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 13427-13436, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.818, + 0.484, + 0.872 + ], + "angle": 0, + "content": "[38] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.873, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[39] Haoran Li, Haolin Shi, Wenli Zhang, Wenjun Wu, Yong Liao, Lin Wang, Lik-hang Lee, and Pengyuan Zhou. Dreamscene:" + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.093, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.553, + 0.093, + 0.907, + 0.134 + ], + "angle": 0, + "content": "3d gaussian-based text-to-3d scene generation via formation pattern sampling. arXiv preprint arXiv:2404.03575, 2024.6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.136, + 0.907, + 0.191 + ], + "angle": 0, + "content": "[40] Xiang Li, John Thickstun, Ishaan Gulrajani, Percy S Liang, and Tatsunori B Hashimoto. Diffusion-lm improves controllable text generation. Advances in Neural Information Processing Systems, 35:4328-4343, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.193, + 0.907, + 0.262 + ], + "angle": 0, + "content": "[41] Xinyang Li, Zhangyu Lai, Linning Xu, Yansong Qu, Liujuan Cao, Shengchuan Zhang, Bo Dai, and Rongrong Ji. Director3d: Real-world camera trajectory and 3d scene generation from text. Advances in neural information processing systems, 2024. 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.264, + 0.907, + 0.319 + ], + "angle": 0, + "content": "[42] Zongrui Li, Minghui Hu, Qian Zheng, and Xudong Jiang. Connecting consistency distillation to score distillation for text-to-3d generation. In European Conference on Computer Vision, pages 274–291. Springer, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.321, + 0.907, + 0.362 + ], + "angle": 0, + "content": "[43] Amy Lin, Jason Y Zhang, Deva Ramanan, and Shubham Tulsiani. Relpose++: Recovering 6d poses from sparse-view observations. arXiv preprint arXiv:2305.04926, 2023. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.364, + 0.907, + 0.433 + ], + "angle": 0, + "content": "[44] Huan Ling, Seung Wook Kim, Antonio Torralba, Sanja Fidler, and Karsten Kreis. Align your gaussians: Text-to-4d with dynamic 3d gaussians and composed diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8576–8588, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.435, + 0.907, + 0.517 + ], + "angle": 0, + "content": "[45] Lu Ling, Yichen Sheng, Zhi Tu, Wentian Zhao, Cheng Xin, Kun Wan, Lantao Yu, Qianyu Guo, Zixun Yu, Yawen Lu, et al. D13dv-10k: A large-scale scene dataset for deep learning-based 3d vision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22160–22169, 2024. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.519, + 0.907, + 0.561 + ], + "angle": 0, + "content": "[46] Yaron Lipman, Ricky TQ Chen, Heli Ben-Hamu, Maximilian Nickel, and Matt Le. Flow matching for generative modeling. arXiv preprint arXiv:2210.02747, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.563, + 0.907, + 0.616 + ], + "angle": 0, + "content": "[47] Xingchao Liu, Chengyue Gong, and Qiang Liu. Flow straight and fast: Learning to generate and transfer data with rectified flow. arXiv preprint arXiv:2209.03003, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.619, + 0.907, + 0.688 + ], + "angle": 0, + "content": "[48] Xian Liu, Xiaohang Zhan, Jiaxiang Tang, Ying Shan, Gang Zeng, Dahua Lin, Xihui Liu, and Ziwei Liu. Humangaussian: Text-driven 3d human generation with gaussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6646-6657, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.69, + 0.907, + 0.746 + ], + "angle": 0, + "content": "[49] Yuan Liu, Cheng Lin, Zijiao Zeng, Xiaoxiao Long, Lingjie Liu, Taku Komura, and Wenping Wang. Syncdreamer: Generating multiview-consistent images from a single-view image. arXiv preprint arXiv:2309.03453, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.747, + 0.907, + 0.83 + ], + "angle": 0, + "content": "[50] Xiaoxiao Long, Yuan-Chen Guo, Cheng Lin, Yuan Liu, Zhiyang Dou, Lingjie Liu, Yuexin Ma, Song-Hai Zhang, Marc Habermann, Christian Theobalt, et al. Wonder3d: Single image to 3d using cross-domain diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9970-9980, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.832, + 0.907, + 0.901 + ], + "angle": 0, + "content": "[51] Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11461-11471, 2022. 2, 4, 5" + }, + { + "type": "list", + "bbox": [ + 0.524, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21533" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.092, + 0.486, + 0.148 + ], + "angle": 0, + "content": "[52] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jiajun Wu, Jun-Yan Zhu, and Stefano Ermon. Sdedit: Guided image synthesis and editing with stochastic differential equations. arXiv preprint arXiv:2108.01073, 2021. 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.151, + 0.485, + 0.219 + ], + "angle": 0, + "content": "[53] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.222, + 0.486, + 0.292 + ], + "angle": 0, + "content": "[54] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6038–6047, 2023. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.294, + 0.485, + 0.349 + ], + "angle": 0, + "content": "[55] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM transactions on graphics (TOG), 41(4):1-15, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.352, + 0.484, + 0.421 + ], + "angle": 0, + "content": "[56] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.424, + 0.484, + 0.478 + ], + "angle": 0, + "content": "[57] Alex Nichol, Heewoo Jun, Prafulla Dhariwal, Pamela Mishkin, and Mark Chen. Point-e: A system for generating 3d point clouds from complex prompts. arXiv preprint arXiv:2212.08751, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.481, + 0.484, + 0.536 + ], + "angle": 0, + "content": "[58] Francesco Palandra, Andrea Sanchietti, Daniele Baieri, and Emanuele Rodola. Gsedit: Efficient text-guided editing of 3d objects via gaussian splatting. arXiv preprint arXiv:2403.05154, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.539, + 0.484, + 0.595 + ], + "angle": 0, + "content": "[59] Byeongjun Park and Changick Kim. Point-dynrf: Point-based dynamic radiance fields from a monocular video. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 3171-3181, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.598, + 0.484, + 0.639 + ], + "angle": 0, + "content": "[60] Byeongjun Park, Sangmin Woo, Hyojun Go, Jin-Young Kim, and Changick Kim. Denoising task routing for diffusion models. arXiv preprint arXiv:2310.07138, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.641, + 0.484, + 0.696 + ], + "angle": 0, + "content": "[61] Byeongjun Park, Hyojun Go, and Changick Kim. Bridging implicit and explicit geometric transformation for single-image view synthesis. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.699, + 0.484, + 0.755 + ], + "angle": 0, + "content": "[62] Byeongjun Park, Hyojun Go, Jin-Young Kim, Sangmin Woo, Seokil Ham, and Changick Kim. Switch diffusion transformer: Synergizing denoising tasks with sparse mixture-of-experts. arXiv preprint arXiv:2403.09176, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.758, + 0.484, + 0.799 + ], + "angle": 0, + "content": "[63] Jangho Park, Gihyun Kwon, and Jong Chul Ye. Ed-nerf: Efficient text-guided editing of 3d scene using latent space nerf. arXiv preprint arXiv:2310.02712, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.802, + 0.484, + 0.828 + ], + "angle": 0, + "content": "[64] Julius Plücker. Analytisch-geometrische Entwicklungen. GD Baedeker, 1828. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.831, + 0.484, + 0.9 + ], + "angle": 0, + "content": "[65] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.092, + 0.486, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.093, + 0.908, + 0.135 + ], + "angle": 0, + "content": "[66] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.137, + 0.908, + 0.191 + ], + "angle": 0, + "content": "[67] Elad Richardson, Gal Metzer, Yuval Alaluf, Raja Giryes, and Daniel Cohen-Or. Texture: Text-guided texturing of 3d shapes. In ACM SIGGRAPH 2023 conference proceedings, pages 1–11, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.193, + 0.908, + 0.248 + ], + "angle": 0, + "content": "[68] Robin Rombach, Patrick Esser, and Björn Ommer. Geometry-free view synthesis: Transformers and no 3d priors. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14356-14366, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.25, + 0.908, + 0.319 + ], + "angle": 0, + "content": "[69] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.321, + 0.908, + 0.376 + ], + "angle": 0, + "content": "[70] Litu Rout, Advait Parulekar, Constantine Caramanis, and Sanjay Shakkottai. A theoretical justification for image inpainting using denoising diffusion probabilistic models. arXiv preprint arXiv:2302.01217, 2023. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.378, + 0.908, + 0.446 + ], + "angle": 0, + "content": "[71] Litu Rout, Negin Raoof, Giannis Daras, Constantine Caramanis, Alex Dimakis, and Sanjay Shakkottai. Solving linear inverse problems provably via posterior sampling with latent diffusion models. Advances in Neural Information Processing Systems, 36, 2023. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.449, + 0.908, + 0.502 + ], + "angle": 0, + "content": "[72] Axel Sauer, Dominik Lorenz, Andreas Blattmann, and Robin Rombach. Adversarial diffusion distillation. In European Conference on Computer Vision, pages 87-103. Springer, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.505, + 0.908, + 0.599 + ], + "angle": 0, + "content": "[73] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278-25294, 2022. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.604, + 0.905, + 0.646 + ], + "angle": 0, + "content": "[74] Yichun Shi, Peng Wang, Jianglong Ye, Mai Long, Kejie Li, and Xiao Yang. Mvdream: Multi-view diffusion for 3d generation. arXiv preprint arXiv:2308.16512, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.648, + 0.908, + 0.701 + ], + "angle": 0, + "content": "[75] Jaidev Shriram, Alex Trevithick, Lingjie Liu, and Ravi Ramamoorthi. Realmdreamer: Text-driven 3d scene generation with inpainting and depth diffusion. arXiv preprint arXiv:2404.07199, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.704, + 0.908, + 0.771 + ], + "angle": 0, + "content": "[76] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International conference on machine learning, pages 2256-2265. PMLR, 2015. 2,3" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.775, + 0.908, + 0.829 + ], + "angle": 0, + "content": "[77] Jiaming Song, Arash Vahdat, Morteza Mardani, and Jan Kautz. Pseudoinverse-guided diffusion models for inverse problems. In International Conference on Learning Representations, 2023. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.832, + 0.908, + 0.9 + ], + "angle": 0, + "content": "[78] Jiaming Song, Qinsheng Zhang, Hongxu Yin, Morteza Mardani, Ming-Yu Liu, Jan Kautz, Yongxin Chen, and Arash Vahdat. Loss-guided diffusion models for plug-and-play controllable generation. In International Conference on Machine Learning, pages 32483-32498. PMLR, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.525, + 0.093, + 0.908, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "21534" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.093, + 0.486, + 0.147 + ], + "angle": 0, + "content": "[79] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.149, + 0.485, + 0.217 + ], + "angle": 0, + "content": "[80] Stanislaw Szymanowicz, Chrisitian Rupprecht, and Andrea Vedaldi. Splatter image: Ultra-fast single-view 3d reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10208-10217, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.219, + 0.484, + 0.285 + ], + "angle": 0, + "content": "[81] Jiaxiang Tang, Zhaoxi Chen, Xiaokang Chen, Tengfei Wang, Gang Zeng, and Ziwei Liu. Lgm: Large multi-view gaussian model for high-resolution 3d content creation. In European Conference on Computer Vision, pages 1-18. Springer, 2024. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.288, + 0.483, + 0.342 + ], + "angle": 0, + "content": "[82] Jiaxiang Tang, Jiawei Ren, Hang Zhou, Ziwei Liu, and Gang Zeng. Dreamgaussian: Generative gaussian splatting for efficient 3d content creation. In The Twelfth International Conference on Learning Representations, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.344, + 0.484, + 0.399 + ], + "angle": 0, + "content": "[83] Arash Vahdat, Francis Williams, Zan Gojcic, Or Litany, Sanja Fidler, Karsten Kreis, et al. Lion: Latent point diffusion models for 3d shape generation. Advances in Neural Information Processing Systems, 35:10021-10039, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.401, + 0.483, + 0.44 + ], + "angle": 0, + "content": "[84] Alexander Vilesov, Pradyumna Chari, and Achuta Kadambi. Cg3d: Compositional generation for text-to-3d via gaussian splatting. arXiv preprint arXiv:2311.17907, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.442, + 0.484, + 0.524 + ], + "angle": 0, + "content": "[85] Vikram Voleti, Chun-Han Yao, Mark Boss, Adam Letts, David Pankratz, Dmitry Tochilkin, Christian Laforte, Robin Rombach, and Varun Jampani. Sv3d: Novel multi-view synthesis and 3d generation from a single image using latent video diffusion. In European Conference on Computer Vision, pages 439-457. Springer, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.526, + 0.484, + 0.565 + ], + "angle": 0, + "content": "[86] Can Wang, Mingming He, Mengei Chai, Dongdong Chen, and Jing Liao. Mesh-guided neural implicit field editing. arXiv preprint arXiv:2312.02157, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.567, + 0.484, + 0.635 + ], + "angle": 0, + "content": "[87] Junjie Wang, Jiemin Fang, Xiaopeng Zhang, Lingxi Xie, and Qi Tian. Gaussianeditor: Editing 3d gaussians delicately with text instructions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20902-20911, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.637, + 0.484, + 0.691 + ], + "angle": 0, + "content": "[88] Yuxuan Wang, Xuanyu Yi, Zike Wu, Na Zhao, Long Chen, and Hanwang Zhang. View-consistent 3d editing with gaussian splatting. In European Conference on Computer Vision, pages 404-420. Springer, 2024. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.693, + 0.484, + 0.761 + ], + "angle": 0, + "content": "[89] Sangmin Woo, Byeongjun Park, Hyojun Go, Jin-Young Kim, and Changick Kim. Harmonyview: Harmonizing consistency and diversity in one-image-to-3d. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10574–10584, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.763, + 0.484, + 0.816 + ], + "angle": 0, + "content": "[90] Jing Wu, Jia-Wang Bian, Xinghui Li, Guangrun Wang, Ian Reid, Philip Torr, and Victor Adrian Prisacariu. Gaussctrl: multi-view consistent text-driven 3d gaussian splattering editing. arXiv preprint arXiv:2403.08733, 2024. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.818, + 0.484, + 0.872 + ], + "angle": 0, + "content": "[91] Hanyuan Xiao, Yingshu Chen, Huajian Huang, Haolin Xiong, Jing Yang, Pratusha Prasad, and Yajie Zhao. Localized gaussian splatting editing with contextual awareness. arXiv preprint arXiv:2408.00083, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.874, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[92] Yinghao Xu, Hao Tan, Fujun Luan, Sai Bi, Peng Wang, Jiahao Li, Zifan Shi, Kalyan Sunkavalli, Gordon Wetzstein," + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.093, + 0.486, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.553, + 0.093, + 0.908, + 0.147 + ], + "angle": 0, + "content": "Zexiang Xu, and Kai Zhang. DMV3d: Denoising multiview diffusion using 3d large reconstruction model. In The Twelfth International Conference on Learning Representations, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.149, + 0.907, + 0.19 + ], + "angle": 0, + "content": "[93] Lihe Yang, Bingyi Kang, Zilong Huang, Zhen Zhao, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything v2. arXiv preprint arXiv:2406.09414, 2024. 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.192, + 0.907, + 0.273 + ], + "angle": 0, + "content": "[94] Taoran Yi, Jiemin Fang, Junjie Wang, Guanjun Wu, Lingxi Xie, Xiaopeng Zhang, Wenyu Liu, Qi Tian, and Xinggang Wang. Gaussian dreamer: Fast generation from text to 3d gaussians by bridging 2d and 3d diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6796-6807, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.275, + 0.907, + 0.343 + ], + "angle": 0, + "content": "[95] Xuanyu Yi, Zike Wu, Qingshan Xu, Pan Zhou, Joo-Hwee Lim, and Hanwang Zhang. Diffusion time-step curriculum for one image to 3d generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9948-9958, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.345, + 0.906, + 0.385 + ], + "angle": 0, + "content": "[96] Lu Yu, Wei Xiang, and Kang Han. Edit-diffnerf: Editing 3d neural radiance fields using 2d diffusion model. arXiv preprint arXiv:2306.09551, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.387, + 0.907, + 0.453 + ], + "angle": 0, + "content": "[97] Wangbo Yu, Jinbo Xing, Li Yuan, Wenbo Hu, Xiaoyu Li, Zhipeng Huang, Xiangjun Gao, Tien-Tsin Wong, Ying Shan, and Yonghong Tian. Viewcrafter: Taming video diffusion models for high-fidelity novel view synthesis. arXiv preprint arXiv:2409.02048, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.456, + 0.907, + 0.537 + ], + "angle": 0, + "content": "[98] Xianggang Yu, Mutian Xu, Yidan Zhang, Haolin Liu, Chongjie Ye, Yushuang Wu, Zizheng Yan, Chenming Zhu, Zhangyang Xiong, Tianyou Liang, et al. Mvimgnet: A large-scale dataset of multi-view images. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9150-9161, 2023. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.539, + 0.907, + 0.607 + ], + "angle": 0, + "content": "[99] Bowen Zhang, Yiji Cheng, Jiaolong Yang, Chunyu Wang, Feng Zhao, Yansong Tang, Dong Chen, and Baining Guo. Gaussiancube: Structuring gaussian splatting using optimal transport for 3d generative modeling. arXiv preprint arXiv:2403.19655, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.609, + 0.907, + 0.663 + ], + "angle": 0, + "content": "[100] Jingbo Zhang, Xiaoyu Li, Ziyu Wan, Can Wang, and Jing Liao. Text2nerf: Text-driven 3d scene generation with neural radiance fields. IEEE Transactions on Visualization & Computer Graphics, 30(12):7749-7762, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.665, + 0.907, + 0.733 + ], + "angle": 0, + "content": "[101] Jason Y. Zhang, Amy Lin, Moneish Kumar, Tzu-Hsuan Yang, Deva Ramanan, and Shubham Tulsiani. Cameras as rays: Pose estimation via ray diffusion. In The Twelfth International Conference on Learning Representations, 2024. 2, 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.735, + 0.907, + 0.801 + ], + "angle": 0, + "content": "[102] Kai Zhang, Sai Bi, Hao Tan, Yuanbo Xiangli, Nanxuan Zhao, Kalyan Sunkavalli, and Zexiang Xu. Gs-lrm: Large reconstruction model for 3d gaussian splatting. In European Conference on Computer Vision, pages 1-19. Springer, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.804, + 0.907, + 0.871 + ], + "angle": 0, + "content": "[103] Kaiyi Zhang, Yang Chen, Ximing Yang, Weizhong Zhang, and Cheng Jin. Point cloud part editing: Segmentation, generation, assembly, and selection. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 7187-7195, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.907, + 0.901 + ], + "angle": 0, + "content": "[104] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21535" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.133, + 0.093, + 0.484, + 0.134 + ], + "angle": 0, + "content": "deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.484, + 0.175 + ], + "angle": 0, + "content": "[105] Yudi Zhang, Qi Xu, and Lei Zhang. Dragtex: Generative point-based texture editing on 3d mesh. arXiv preprint arXiv:2403.02217, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.179, + 0.484, + 0.233 + ], + "angle": 0, + "content": "[106] Linqi Zhou, Yilun Du, and Jiajun Wu. 3d shape generation and completion through point-voxel diffusion. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5826-5835, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.235, + 0.484, + 0.318 + ], + "angle": 0, + "content": "[107] Shijie Zhou, Zhiwen Fan, Dejia Xu, Haoran Chang, Pradyumna Chari, Tejas Bharadwaj, Suya You, Zhangyang Wang, and Achuta Kadambi. Dreamscene360: Unconstrained text-to-3d scene generation with panoramic gaussian splatting. In European Conference on Computer Vision, pages 324–342. Springer, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.32, + 0.484, + 0.374 + ], + "angle": 0, + "content": "[108] Jingyu Zhuang, Di Kang, Yan-Pei Cao, Guanbin Li, Liang Lin, and Ying Shan. Tip-editor: An accurate 3d editor following both text-prompts and image-prompts. ACM Transactions on Graphics (TOG), 43(4):1-12, 2024. 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.484, + 0.374 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.956 + ], + "angle": 0, + "content": "21536" + } + ] +] \ No newline at end of file diff --git a/2025/SplatFlow_ Multi-View Rectified Flow Model for 3D Gaussian Splatting Synthesis/a1da8541-9ef2-4901-8219-e22e7fdfc1fc_origin.pdf b/2025/SplatFlow_ Multi-View Rectified Flow Model for 3D Gaussian Splatting Synthesis/a1da8541-9ef2-4901-8219-e22e7fdfc1fc_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..5331178daa0e69c42aa39ac8c95ef58647041c68 --- /dev/null +++ b/2025/SplatFlow_ Multi-View Rectified Flow Model for 3D Gaussian Splatting Synthesis/a1da8541-9ef2-4901-8219-e22e7fdfc1fc_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:65049bc74b4ec407bb027d63380866e7072608d7387aced3630e5307bbde051a +size 9782490 diff --git a/2025/SplatFlow_ Multi-View Rectified Flow Model for 3D Gaussian Splatting Synthesis/full.md b/2025/SplatFlow_ Multi-View Rectified Flow Model for 3D Gaussian Splatting Synthesis/full.md new file mode 100644 index 0000000000000000000000000000000000000000..5b11de745f9aba28946aa4586225d87fda18d5b3 --- /dev/null +++ b/2025/SplatFlow_ Multi-View Rectified Flow Model for 3D Gaussian Splatting Synthesis/full.md @@ -0,0 +1,414 @@ +# SplatFlow: Multi-View Rectified Flow Model for 3D Gaussian Splitting Synthesis + +![](images/064a97fe37b4b7c8d294be09866d237d0ce22ebffe1b816eceade2541a0823e8.jpg) + +![](images/c818755f2e286d6b4dbfb9b84c92e072c6c88bdd7bd578f66dc09abf9fa4ede6.jpg) + +![](images/376b9af1fd76759b9f5241ed8e9a204b443a0854db268c54225001534e0dabc2.jpg) + +![](images/f71e8b7573d64979aa4dc12bf96889f72cc6cdcc07a34be67142f9cc7f6309c9.jpg) + +![](images/6f8a8e4d6f4321e2852f254fd033c19e129b3e0dc1ee5d4d1cf706760fc01130.jpg) +A library with a lot of books + +![](images/c2982e5c1a3b23e30470f0d3878d3c0257f7a8f3677789ec62e75728e02484d3.jpg) +Figure 1. SplatFlow for 3D Gaussian Splitting synthesis and its training-free applications. (a) Examples of direct 3D Gaussian Splitting (3DGS) generation only from text prompts, (b) Training-free applications, including 3DGS object editing, camera pose estimation, and novel view synthesis. SplatFlow seamlessly integrates these capabilities, showcasing its versatility in generating and editing complex 3D content. + +![](images/e3425c2d94e5378e44eb4bda7e499b7e0ba88b81f87061a5fcbef324ece2f5b8.jpg) + +![](images/da867f973dfe18f63cb0de40727f306d2b21a6448849fb83ba8b30c0e690529d.jpg) +A forest with tall trees and yellow leaves + +![](images/f348c71eea6be672c093a5c9ce101abc5edca11b02984c37ee394ffbb9acf3f3.jpg) +Blue car on a clean street with buildings in the background. + +![](images/f426a58a2d365faca2fcd43e57b9f5d9b1aed98931a9fcc7389bd614e4c2a1f8.jpg) +Half avocado + +![](images/a1cb932aff8f3b6017b9a97d752540c969d5c08fe28c81794f99bd4cddb64911.jpg) +slicedzucchini +(a) 3D Gaussian Splitting Synthesis + +![](images/c89704bb8cca0486ed6f9faaca339c1fb0b7517139ae2d6969f9404c61cea928.jpg) +A lighthouse on a rocky shore + +![](images/902b10d04f2786fc83d5c96b0918ae15114974c46f0d7145a4e07a495a5be346.jpg) +A vibrant bush with pink flowers stands in front of a concrete wall, ... + +![](images/1358ff03ebe3bed73134b5ab8e566f41f675fa0c589f08776e9995e744b68f35.jpg) +3DGS object editing (replacement) + +![](images/6458c8efd497d95c7d4843e6f151d69dbf12531a2bebedaed111ced4fc55e817.jpg) + +![](images/8fd0d6c19f961b167132882ee6fd41109fc3fb6900a2583bd0df23a82ea11a5d.jpg) + +![](images/ebbbe4a5a18b58e9bc493fce0d2fb3063d97c55375c072502ea8fcd57762a3c7.jpg) + +![](images/6b46d5028c10cd12cc29d8cf1ec9fc092f357a23c7ac29ef734984289c1bfb61.jpg) + +![](images/3ce1b93006ec113de40bda8dec6865b2604a8b5e8d075c8b5f7c79b4dfcba96a.jpg) +"Camera Pose Estimation" + +![](images/7ec74f2631bb999bf9412660f65b05af585fb47d14b19fb9f7898c598e11eff8.jpg) + +![](images/3b0f77f8518208b80ecf34b7c05a8d692ee584d5e26218b419695991d8736551.jpg) +(b) Training-Free Applications + +![](images/092483e1230096e6b5a42dc843f9cf6657e7d5b3334930482438f0117b60eeb0.jpg) +"Novel View Synthesis" + +![](images/44a82faf846b356c7b18da5c281a23891930af98e26b4577cc0fa179d9df8bdf.jpg) + +# Abstract + +Text-based generation and editing of 3D scenes hold significant potential for streamlining content creation through intuitive user interactions. While recent advances leverage 3D Gaussian Splatting (3DGS) for high-fidelity and real-time rendering, existing methods are often specialized and task-focused, lacking a unified framework for both generation and editing. In this paper, we introduce SplatFlow, a comprehensive framework that addresses this gap by enabling direct 3DGS generation and editing. SplatFlow comprises two main components: a multi-view rectified flow (RF) model and a Gaussian Splatting Decoder (GSDecoder). The multi-view RF model operates in latent space, generating multi-view images, depths, and camera poses simultaneously, conditioned on text prompts—thus addressing challenges like diverse scene scales and complex camera trajectories in real-world settings. Then, the + +GSDecoder efficiently translates these latent outputs into 3DGS representations through a feed-forward 3DGS method. Leveraging training-free inversion and inpainting techniques, SplatFlow enables seamless 3DGS editing and supports a broad range of 3D tasks—including object editing, novel view synthesis, and camera pose estimation—within a unified framework without requiring additional complex pipelines. We validate SplatFlow's capabilities on the MVImgNet and DL3DV-7K datasets, demonstrating its versatility and effectiveness in various 3D generation, editing, and inpainting-based tasks. Our project page is available at https://gohyojun15.github.io/SplatFlow/. + +# 1. Introduction + +The demand for realistic 3D scene generation and editing from text has surged in applications like VR/AR, gaming, and robotics. Early works primarily utilized NeRF [53, 55] for scene generation [100] and editing [2, 29, 30, 33, 63, 96], but NeRF's volumetric rendering is computationally + +intensive, resulting in slower rendering speeds. Recently, 3D Gaussian Splitting (3DGS) [32] has emerged as a promising alternative, offering real-time and high-fidelity rendering implementations. Building on this, recent advances in 3D generation [8, 34, 58, 81, 87, 88, 90, 92, 99, 102, 108] and editing [8, 34, 58, 87, 88, 90, 108] increasingly leverage 3DGS to achieve both speed and quality. + +Despite these advancements, current methods for generating and editing 3DGS are specialized and task-focused, lacking a unified framework. For 3DGS generation, several works [10, 42, 48, 82, 84, 94] leverage 2D diffusion models [26, 69, 76, 79] with Score Distillation Sampling (SDS) [66], which requires time-intensive per-scene optimization. To address this, recent studies have shifted towards direct 3DGS generation, combining diffusion models with reconstruction models [81, 92, 102] or utilizing 3D generative models [99]. However, most of these methods are restricted to object-level 3D generation using synthetic datasets [15, 16] with controlled, bounded environments. In contrast, real-world scenes, varying in scene scales and camera trajectories, present distinctive challenges. + +In 3DGS editing, several works [8, 34, 58, 87, 88, 90, 108] adopt 2D diffusion models [26, 69, 76, 79] to guide the editing process. The main challenge is lifting 2D editing guidance to 3D while maintaining 3D-consistent modifications. One approach focuses on SDS techniques to update 3DGS [58, 91, 108], but these methods require additional stages like texture adjustments [58, 91] and fine-grained refinements [108]. Another approach aims to edit multiview images with 3D consistency [34, 88, 90], introducing attention-based modules [88, 90] and autoregressive editing [34], adding complexity to the editing pipeline. + +In this paper, we bridge the gap between 3DGS generation and editing by designing a direct 3DGS generative model. Inspired by the success of 2D diffusion models that enable training-free editing through inversion [13, 52, 54], we argue that a direct 3DGS model can similarly enable training-free editing. This approach integrates editing without requiring complex pipelines to lift 2D priors, making generation and editing adaptable within a unified framework. + +We introduce SplatFlow, which consists of two main components: (1) a multi-view rectified flow (RF) model and (2) a Gaussian Splitting Decoder (GSDecoder). Similar to Latent Diffusion Models [69], our multi-view RF model operates in the latent space and is trained to simultaneously generate multi-view images, depths, and camera poses conditioned on text prompts. This joint modeling enables SplatFlow to effectively handle the challenges of real-world scene generation, such as various scene scales and camera trajectories, unlike synthetic object datasets [15, 16]. The GSDecoder translates the latent outputs into 3DGS, based on efficient feed-forward methods [4, 9]. To ensure compatibility with other generative models, particularly Stable Diffusion 3 [19], + +we incorporate a fixed pre-trained encoder, allowing flexible cross-model usage during sampling. + +Finally, SplatFlow utilizes training-free inversion [13, 52, 54] and inpainting techniques [11, 51, 70, 71, 77] for 3DGS editing and various 3D tasks. For example, it enables object editing [7, 58], novel view synthesis [61, 68, 97], and camera pose estimation [101] by jointly modeling images, depths, and poses to infer missing elements, highlighting its versatility. We validated its capabilities using the MVImgNet [98] and DL3DV-7K [45] datasets, assessing its performance in 3D generation, editing, and inpainting-based tasks. + +# 2. Related Works + +# 2.1. Diffusion Models and Rectified Flows + +Diffusion Models Diffusion models [26, 76, 79] generate data through iterative denoising and have become standard in generative modeling across images [17, 21, 37, 60, 62], videos [27], and text [40]. Trained on large-scale text-to-image datasets, they capture complex semantic relationships, producing high-quality images [6, 56, 65, 69, 72]. An advantage is their adaptability to various tasks in a plug-and-play manner [20, 22, 78], especially inpainting [51, 70] and editing [23, 35]. In painting can be framed as a linear inverse problem [14], for which effective posterior sampling methods have been proposed [11, 51, 70, 71, 77]. For editing, inversion methods [13, 52, 54] produce structured noise by inverting images, allowing new prompts to be effectively processed. Building on these training-free techniques, we extend them to our 3DGS model, adapting RePaint [51] and SDEdit [52] within rectified flow models. + +Rectified Flow Models Rectified Flow (RF) models [1, 46, 47] represent an alternative approach to traditional diffusion models by utilizing a straight-line path from data to noise instead of the typical curved forward process, aiming to simplify sampling and reduce computational costs. This linear path theoretically enables faster sampling by minimizing error accumulation across steps, and recent advances, such as Stable Diffusion 3 (SD3) [19], have successfully scaled RF in text-to-image generation, demonstrating performance that surpasses traditional diffusion models. Our SplatFlow framework builds on the capabilities of SD3, integrating RF to achieve efficient 3DGS generation. + +# 2.2. 3D Generative Models + +Neural Scene Representation Various neural scene representations have been proposed for 3D generation tasks. Early works use explicit representations, enabling synthesis of 3D point clouds [57] and shapes [83, 106], but struggle with realistic scene generation due to their rendering mechanisms. To address this, NeRF [53] is employed for a realistic 3D generation with per-scene optimization [66], but its inefficient rendering is unsuitable for a fast generation. + +Recently, 3DGS [32] enables real-time rendering, accelerating optimization-based 3D generation [10, 44, 82]. Building on this progress, we leverage 3DGS's efficient rendering for fast training of a generative model on large datasets. + +Lifting Multi-View Generation Advances in image generation models [6, 65] and large-scale 3D object datasets [15, 16] make multi-view generation compelling compared to optimization-based generation [82]. The key idea is to learn 3D priors as conditions [49, 50, 89] or attentions [74, 85] from 3D datasets while leveraging text-to-image diffusion models [69] trained on billions of text-image pairs [73]. This enables the synthesis of multi-view consistent images from an image or text prompt, which is then fed into reconstruction models [5, 81] to directly generate neural scene representations like 3DGS. One critical success factor in these methods is that the training dataset provides rendered images from consistent, canonical viewpoints, allowing instant access to desired views regardless of the 3D object. In this work, we focus on more challenging unbounded 3D real-world scene generation, where each scene varies in scale and camera trajectory. To handle diverse trajectories, our model learns a joint distribution of camera poses and 3D scenes. + +Text-to-3D Scene Generation Generating 3D scenes from text prompts is challenging due to the unbounded scale of scenes, complicating scale determination. One straightforward approach is to create a scene covered by a single image [75, 107] generated from text-to-image diffusion models [6, 65]. Another line of work [12, 28] generates wider scenes using user-provided camera trajectories, but these methods are time-consuming and produce subpar quality due to reliance on 2D generative models. Recently, Director3D [41], the work most similar to ours, generates camera poses from text and then creates multi-view images decoded into 3DGS. In contrast, SplatFlow jointly learns the distribution of camera poses and multi-view images from text, enhancing generation quality and enabling flexible 3D scene editing in a unified framework. + +# 2.3. 3D Editing + +In recent years, NeRF-based methods [2, 29, 30, 33, 59, 63, 67, 95, 96] have gained popularity for 3D editing, surpassing point clouds and meshes [86, 103, 105]. However, due to the inefficient rendering, these methods are time-consuming and lack precise control. Conversely, incorporating 3DGS into 3D editing [8, 87] improves effectiveness, speed, and control. + +For guiding 3D editing, several methods have leveraged 2D diffusion models [26, 69, 76, 79] for structured 3D editing, but it often requires additional refinement stages. For instance, GSEdit [58] and Xiao et al. [91] apply texture enhancement, while TIP-Editor [108] adds fine-grained adjustments to reduce SDS artifacts. Another line of research in 3D editing emphasizes achieving multi-view consistency + +by sequentially updating each view. For example, Gauss-Ctrl [90] and 3DEdit [88] utilize attention-based modules to propagate edits made from one perspective across multiple views, ensuring that changes are consistently reflected throughout the 3D scene. Similarly, 3D-Ego [34] adopts an autoregressive editing framework, updating each view to enhance multi-view consistency. + +Unlike prior methods that require complex procedures to lift 2D priors for 3D consistency, SplatFlow directly models multi-view consistency within the latent space, efficiently decoding edits through our Gaussian Splitting Decoder. + +# 3. Preliminary: Rectified Flows + +Rectified Flows (RFs) [1, 46, 47] consider generative models that construct a noise distribution $q_{0}$ and a time-varying vector field $v_{t}(\boldsymbol{x}_{t})$ to sample data $x_{0}$ from a data distribution $p_{0}$ via an ordinary differential equation (ODE) as: + +$$ +\mathrm {d} X _ {t} = v \left(X _ {t}\right) \mathrm {d} t, \quad X _ {0} \sim q _ {0}, \quad t \in [ 0, 1 ], \tag {1} +$$ + +where $q_{0}$ typically follows a Gaussian distribution $\mathcal{N}(0,I)$ . RF employs a neural network with parameters $\theta$ for modeling the vector field as $v_{t}(X_{t}) = -u_{\theta}(X_{t},1 - t)$ . + +To train a neural network as the vector field, RFs utilize paired samples from $p_0$ and $q_0$ (here simplified as $p_1$ ), along a straight line path $Y_t = tY_1 + (1 - t)Y_0$ . This setup produces the marginal distribution of $Y_t$ as $p_t(\pmb{y}_t) = \int p_t(\pmb{y}_t|\pmb{y}_1)p_1(\pmb{y}_1)\mathrm{d}\pmb{y}_1$ . Starting from an initial $Y_0 = \pmb{y}_0$ and moving to a final state $Y_1 = \pmb{y}_1$ , a straight line path induces an ODE expressed as $\mathrm{d}Y_t = u_t(Y_t|Y_1)\mathrm{d}t$ where the vector field $u_t(Y_t|Y_1)$ represents the conditional change from $\pmb{y}_0$ to $\pmb{y}_1$ , given by $u_t(Y_t|Y_1) = \pmb{y}_1 - \pmb{y}_0$ . The marginal vector field $u_t(\pmb{y}_t)$ is computed by averaging over the conditional vector field as [46]: + +$$ +u _ {t} (\boldsymbol {y} _ {t}) = \int u _ {t} (\boldsymbol {y} _ {t} | \boldsymbol {y} _ {1}) \frac {p _ {t} (\boldsymbol {y} _ {t} | \boldsymbol {y} _ {1})}{p _ {t} (\boldsymbol {y} _ {t})} p _ {1} (\boldsymbol {y} _ {1}) \mathrm {d} \boldsymbol {y} _ {1}. \tag {2} +$$ + +To approximate $u_{t}(\boldsymbol{y}_{t})$ , the neural network is trained with a flow-matching object, defined by: + +$$ +\mathcal {L} _ {\mathrm {F M}} := \mathbb {E} _ {t, Y _ {t}} \left[ \left| \left| u _ {t} \left(Y _ {t}\right) - u _ {\theta} \left(Y _ {t}, t\right) \right| \right| _ {2} ^ {2} \right]. \tag {3} +$$ + +Since directly computing this objective can be computationally challenging, an alternative approach, conditional flow matching, is used to simplify training: + +$$ +\mathcal {L} _ {\mathrm {C F M}} := \mathbb {E} _ {t, Y _ {t}, Y _ {1} \sim p _ {1}} \left[ \| u _ {t} \left(Y _ {t} \mid Y _ {1}\right) - u _ {\theta} \left(Y _ {t}, t\right) \| _ {2} ^ {2} \right]. \tag {4} +$$ + +The gradients of $L_{\mathrm{CFM}}$ and $L_{\mathrm{FM}}$ are theoretically identical [46], making them equivalent in training. However, $L_{\mathrm{CFM}}$ offers better computational efficiency and is therefore preferred. Ultimately, the learned vector field for the ODE in Eq. 1 is given by $v_{t}(X_{t}) = -u_{\theta}(X_{t},1 - t)$ , allowing RFs to generate samples that follow the data distribution using the learned vector field. + +![](images/55dfe3ec3802dd8d637c1ff9e161605a4829c36665aaa8234007875857183f75.jpg) +Figure 2. Overview of SplatFlow. SplatFlow consists of two main components: a multi-view Rectified Flow (Section 4.3) model and a Gaussian Splat Decoder (Section 4.2). Conditioned on text, RF model generates multi-view latents—including image, depth, and Plücker ray coordinates. After an optimization process to estimate camera poses, the GSDecoder decodes these latents into pixel-aligned 3DGS. + +# 4. Methodology + +# 4.1. Overview of SplatFlow + +We introduce SplatFlow, a 3DGS generation model designed to perform various 3D editing and 3D-related tasks in a training-free manner, despite being trained solely for generation. As shown in Fig. 2, our model consists of two main components: 1) a multi-view Rectified Flow (RF) model and 2) a Gaussian Splitting Decoder (GSDecoder). First, the multi-view RF model generates multi-view camera poses, depths, and images conditioned on text prompts. Then, the GSDecoder translates these latent representations into pixel-aligned 3DGS, constructing a real-world scene structure. + +Also, both the multi-view RF model and GSDecoder operate within a latent space, adopting a Latent Diffusion Model (LDM) [69]. By freezing and sharing the encoder of Stable Diffusion 3 (SD3) [19], our design enables the RF model to share the latent space with SD3, allowing us to leverage its knowledge for enhanced generation and editing capabilities. + +Finally, our model leverages a training-free approach [51, 52], enabling a broad range of editing applications, such as object replacement, as well as 3D-related tasks including camera pose estimation and novel view synthesis. + +In the following sections, we present each component in detail. Section 4.2 introduces the GSDecoder, explaining details in constructing pixel-aligned 3DGS. Section 4.3 describes the multi-view RF model, detailing how it jointly generates camera poses, depths, and images. Finally, Section 4.4 outlines the sampling process and demonstrates how our model applies training-free techniques. + +# 4.2. Gaussian Splitting Decoder (GSDecoder) + +Recently, feed-forward 3DGS methods [4, 9, 80, 81] have enabled fast 3DGS reconstruction from sparse views by training on large datasets, achieving much faster reconstruction + +than per-scene optimization methods [32]. Leveraging this advantage, our GSDecoder $G_{\phi}$ (parameterized by $\phi$ ) is designed to decode pixel-aligned 3DGS from latent representations of $K$ sparse views, given their corresponding camera poses $\mathcal{P} = \{\pmb{P}_i\}_{i=1}^K$ with $\pmb{P}_i = \pmb{K}_i[\pmb{R}_i|\pmb{T}_i]$ , where $\pmb{K}_i, \pmb{R}_i$ , and $\pmb{T}_i$ denote the intrinsic matrix, rotation matrix, and translation vector of the $i$ -th view, respectively. + +A straightforward approach is to design $G_{\phi}$ to directly output the 3D Gaussian parameters from the image latents obtained through the encoder $\mathcal{E}$ as $G_{\phi}(\{\langle \mathcal{E}(I_i), P_i \rangle \}_{i=1}^K) = \{(\pmb{\mu}_j, \alpha_j, \pmb{\Sigma}_i, c_j)\}_{j=1}^{H \times W \times K}$ , where each 3D Gaussian parameter includes position $\pmb{\mu}_j$ , opacity $\alpha_j$ , covariance $\pmb{\Sigma}_j$ , and color $c_j$ in a pixel-aligned manner with each $i$ -th image $I_i \in \mathbb{R}^{H \times W}$ . However, as the image latents are produced through a shared and frozen encoder (to leverage compatibility with 2D generative models during sampling), it might lose fine-grained spatial details, leading to abstract representations. This can limit the preservation of specific 3D structural information. To address these limitations, we propose two improvements to enhance GSDecoder's performance. + +Depth latent integration As demonstrated in [31], the encoder $\mathcal{E}$ can effectively encode depth maps. Building on this, we incorporate the depth latents $\{\mathcal{E}(D_i)\}_{i=1}^K$ of each depth map $D_i$ as an additional input to the GSDecoder to further enhance 3D structural information. We use DepthAnythingV2 [93] to extract the depth maps and normalize them to the range [-1, 1], stacking them as three channels similar to RGB for constructing $D_i$ . We observe that this improves the convergence speed and performance of the GSDecoder. + +Adversarial loss Applying adversarial losses has demonstrated improvements in the visual quality of decoded output [18, 69]. However, using adversarial loss during training can cause instability, particularly in the early stages. We observe that applying the vision-aided loss [36] after the GSDe + +coder has reached a certain level of convergence significantly improves visual quality without destabilizing training. + +Training In summary, our GSDecoder predicts 3DGS $\{(\pmb{\mu}_j,\alpha_j,\pmb{\Sigma}_j,\pmb{c}_j)\}_{j = 1}^{H\times W\times K}$ from $\{(I_i,D_i,P_i)\}_{i = 1}^K$ . The model is trained with a combined loss function consisting of LPIPS [104], Mean Squared Error loss, and vision-aided loss [36], computed between the target images of novel views and rendered views from 3DGS. Specifically, the vision-aided loss is activated only after a certain threshold iteration to improve stability and perceptual quality. Additionally, the architecture is initialized with the parameters of the Stable Diffusion 3 decoder, incorporating cross-view attention mechanisms, and the channel dimensions are increased to meet the specific requirements of 3D Gaussian Splatting. Further details on the GSDecoder are provided in Appendix A. + +# 4.3. Multi-View Rectified Flow Model + +With the GSDecoder in place, the 3D Gaussian Splitting (3DGS) generation process becomes feasible if we can produce consistent multi-view latents of images $\mathcal{E}(I_i)$ , depth maps $\mathcal{E}(D_i)$ , and camera poses $P_{i}$ for $i = 1,\dots ,K$ . Our multi-view rectified flow (RF) model is trained to sample from the joint distribution of these latents, conditioned on text $C$ as $p(\{I_i\}_{i = 1}^K,\{D_i\}_{i = 1}^K,\{P_i\}_{i = 1}^K |C)$ , allowing for simultaneous generation of multi-view images, depths, and poses. The rationale for modeling a joint distribution instead of each component separately is twofold. First, this approach allows the model to handle various tasks by reformulating them as inpainting problems, as discussed in Section 4.4. Second, joint modeling is crucial for real-world scenes, which require adaptive camera poses tailored to each scene [41]. + +To achieve this, our multi-view RF model treats the concatenation of multi-view image latents, depth latents, and Plücker ray coordinates [64, 101] along the channel dimension as input data. Formally, each ray $\pmb{r}_i$ is represented as $\langle \pmb{d}_i, \pmb{m}_i \rangle$ , where $\pmb{d}_i = \pmb{R}_i^\top \pmb{K}_i^{-1} \pmb{w}_i$ denotes the direction vector and $\pmb{m}_i = (-\pmb{R}_i^\top \pmb{T}_i) \times \pmb{d}_i$ represents the moment vector, with $\pmb{w}_i$ referring to the 2D pixel coordinates. By matching the spatial resolution of $\pmb{w}_i$ with the image and depth latents, each represented as $\mathbb{R}^{n \times h \times w}$ , $\pmb{r}_i$ can be expressed as $\mathbb{R}^{6 \times h \times w}$ , enabling concatenation along the channel axis. + +Through this setup, we define $Y_{0}$ as the multi-view latents $\mathbf{X}$ , where each element $\mathbf{X}_i = \langle \mathcal{E}(\mathbf{I}_i),\mathcal{E}(\mathbf{D}_i),\mathbf{r}_i\rangle$ results in $Y_{0} = (X_{1},X_{2},\ldots ,X_{K})\in \mathbb{R}^{K\times (2n + 6)\times h\times w}$ . We then train the multi-view RF model $u_{\theta}$ with the conditional flow matching loss defined in Eq. 4. Specifically, we fine-tune Stable Diffusion 3 [19] by adjusting the channel dimensions in the input and output layers and incorporating cross-view attention. Further details are provided in Appendix A. + +Sampling process To recover the camera pose from ray $\boldsymbol{r}_i$ , we follow the procedure in [101] and further refine the pose parameters to ensure that multiple views share the same intrinsic. However, we find that naively solving the ODE + +can lead to degraded camera pose accuracy. We hypothesize this degradation results from deviations from the camera pose manifold, which introduces errors in camera trajectory estimation. To address this, at each sampling step $t = t_k$ , we predict the sampling destination at $t = 0$ , recover the camera poses from this prediction, and reconstruct the associated camera rays. These rays are then forwarded to $t = t_k$ to maintain the solution within the valid ray manifold at $t = t_k$ . + +Stable Diffusion 3 guidance To improve multi-view image generation quality, we integrate vector fields from Stable Diffusion 3, known for robust generalization, with our multiview RF model to solve the ODE for randomly selected view's image latents at specific sampling steps. + +# 4.4. Inference Given Partial Data + +In this section, we demonstrate how SplatFlow can be effectively utilized for 3D editing and applications by applying training-free inversion techniques [13, 52, 54] and inpainting methods [11, 51, 70, 71, 77]. We modify SDEdit [52] inversion for rectified flow-based inversion and adopt the RePaint [51] to support rectified flow inpainting. Additional technical details are provided in Appendix A. + +3DGS editing Our GSDecoder is designed to directly decode 3DGS, enabling efficient editing by modifying only the multi-view latents. Additionally, with the RF model's joint modeling of multi-view image latents, edits can be achieved without extra modules (e.g., cross-view attention mechanisms [88, 90]), streamlining the process through simple inversion and sampling. Specifically, given a target text and input multi-view latents, we apply a modified SDEdit for inversion at a chosen timestep $t_k$ . Conditioned on the target text, we then generate the edited multi-view latents by solving the ODE in Eq. 1 following our sampling procedure. + +Inpainting application Our model is also applicable to various 3D tasks, as it jointly models multi-view images, depths, and camera poses. This joint modeling allows known parts of the data (e.g., some views or camera parameters) to act as constraints, enabling the prediction of unknown parts through inpainting. In this work, we focus on two specific tasks: 1) camera pose estimation from multi-view images and depths, and 2) novel view synthesis using a subset of the $K$ multi-view latents and camera poses for novel viewpoints. + +# 5. Experimental Results + +In this section, we demonstrate the capabilities of our proposed SplatFlow across 3DGS generation, 3DGS editing, and various 3D tasks. We begin by detailing the implementation of SplatFlow in Section 5.1. Following this, we present comparative results on 3DGS generation in real-world datasets in Section 5.2. In Section 5.3, we evaluate the 3DGS editing performance of SplatFlow against other methods. Finally, Section 5.4 showcases the versatility of our + +![](images/42ac89b4832e303bdded132ff403ebf887638a38b7f79b795357cc9b61a81e8b.jpg) +Figure 3. Qualitative results in text-to-3DGS generation on MVImgNet and DL3DV validation sets. The first two rows are rendered scenes from the MVImgNet dataset, while the last two rows are from the DL3DV dataset. Our SplatFlow produces cohesive and realistic scenes with sharp details, accurately capturing the intricacies of real-world environments and accommodating diverse camera trajectories. + +multi-view RF model in handling diverse 3D tasks through training-free inpainting techniques. Extensive information on all our experiments is provided in Appendix B with additional results that could not be included here in the Appendix. + +# 5.1. Experimental Setups + +Due to space constraints, we provide a concise overview of our experimental setups here. We used the MVImgNet [98] and DL3DV-7K [45] datasets, both containing real-world scenes with images and corresponding camera poses. For evaluation, we utilized 1.25K scenes from MVImgNet for each task and 300 sequences from DL3DV-7K, using the + +remaining for training. The multi-view RF model and GSDecoder were trained with $K = 8$ -view setup. Text descriptions for scenes were extracted using the Llava-One Vision Qwen 7B [38]. We used SDS++ [41] for further improvements. + +# 5.2. Results on Text-to-3DGS Generation + +We evaluated our model against various baselines on the DL3DV and MVImgNet validation sets for real-world scene text-to-3DGS generation. Our evaluation protocol utilized the rendered image of the generated 3DGS. We used the FID score [25] to evaluate image quality and the CLIP score [24] to measure alignment with text prompts. + +
MethodMVImgNet [98]DL3DV [45]
FID-10K↓CLIPScore↑FID-2.4K↓CLIPScore↑
Director3D [41]39.5530.4888.4430.04
Director3D (w/ SDS++) [41]41.8031.0095.8831.68
SplatFlow34.8531.4379.9130.06
SplatFlow (w/ SDS++)35.4632.3085.3131.90
+ +For quantitative comparison, we adopt Director3D [41], which leverages the full sequences of MVImgNet, DL3DV-10K, and LAION [73] datasets. In contrast, our model was trained on a notably smaller dataset, utilizing only a portion of MVImgNet and DL3DV-7K, without access to the full DL3DV-10K or LAION data. For qualitative results, we also compared against other scene generation methods, including LucidDreamer [12], DreamScene [39], and Director3D [41]. + +Quantitative results Table 1 presents the quantitative comparison of SplatFlow and the baseline methods on both MVImgNet and DL3DV datasets. SplatFlow consistently outperforms Director3D [41] in terms of FID and CLIP score, demonstrating its effectiveness in generating high-quality images and maintaining strong alignment with text prompts, even with a smaller training dataset. Specifically, on MVImgNet, SplatFlow achieves a significantly lower FID (34.85 vs. 39.55 for Director3D), indicating better image quality, and a higher CLIP score (31.43 vs. 30.48), showing improved text-to-image alignment. Similarly, on DL3DV, SplatFlow achieves an FID of 79.91, outperforming Director3D's 88.44, with a comparable CLIP score. Notably, with the addition of $\mathrm{SDS}++$ for refinement, both SplatFlow and Director3D show increased CLIP scores, further enhancing text alignment, though at a minor trade-off in FID. This highlights the effectiveness of $\mathrm{SDS}++$ for boosting alignment with text prompts across different 3DGS generation models. + +Qualitative results Figure 3 shows the qualitative comparisons between our method and baselines on DL3DV and MVImgNet validation sets. Overall, SplatFlow demonstrates clear advantages over the baseline methods. First, DreamScene [39], which relies on Score Distillation Sampling (SDS), often struggles to generate realistic real-world scenes, resulting in outputs that lack natural textures and environmental details. Second, LucidDreamer [12], which is based on single-view inpainting, faces challenges when handling scenes with large camera trajectory variations, leading to inconsistent or incomplete views. Third, Director3D [41] tends to produce slightly blurred outputs, lacking the sharpness needed for intricate details. In contrast, our SplatFlow can generate coherent and realistic scenes with clear details, effectively capturing the complexities of real-world environments and various camera trajectories. Additional qualitative results of SplatFlow can be found in Appendix C. + +![](images/86838318c008a85c46e9f3f96907c998be65ac0cf1a0c5c8b518c2da0e4624f0.jpg) + +![](images/272d339f885460e47bbda6a32fdf95328d7eb5fb647912a24ba172449f99bd78.jpg) +"green inflatable ring" -> "colorful red rubber ball" + +![](images/82751cfae236038e957eab30d907c57afb1fed744ea0777599412dcc7c535ef3.jpg) + +![](images/e02dd886852cc769955fbeddd251757b7a7f068bd889bf9b172ca157513b4e70.jpg) + +![](images/5c1d203bd60c12d13f5695c95edfa6ff5f6aad279626ea24f5182e5696a256f8.jpg) +Figure 4. Qualitative results in 3D editing with MVInpainter [3] and DGE [7]. We show rendered scenes except for MVInpainter. + +![](images/b114b2b479f68a9e32df423ebdd098815a890778831b32afbd835dd157b7e334.jpg) +"collection of red chili peppers" -> "cluster of bananas" + +![](images/2bf6cfd51c36531d7f5b42e4dddb964d0fb707e7642b4cd7c6034ac139ad2f58.jpg) + +![](images/c1ce82817da1b479833e2b63d43141c79b95d1ba35e1eea62ab6d56fb3cb0aca.jpg) + +# 5.3. Result on 3DGS Editing + +We demonstrate the effectiveness of SplatFlow in the 3DGS editing task by creating a benchmark with 100 sampled scenes from the MVImgNet validation set. + +Table 1. Quantitative results in text-to-3DGS generation on the MVImgNet and DL3DV datasets. We compared our SplatFlow with and without the SDS++ [41], against Director3D [41]. + +
MethodCLIPScore↑CLIP D-sim↑
DGE [7]27.430.102
SplatFlow28.470.169
+) SDS++31.300.224
+ +Table 2. 3D object replacement. + +Images and texts were provided to GPT-4 to identify main objects and generate text captions for edited images, which were then used for 3D object replacement. For editing, we used the same mask in [3] for masking out the object, then inpainted this area based on the provided editing captions. We compared this pipeline with an optimization-based 3DGS editing method, DGE [7], and conducted qualitative comparisons with MVInpainter [3], which propagates an edited image from pretrained generative models across multiple views. Following standard protocols [8, 29], we measured the CLIP score and directional similarity (CLIP D-sim). + +Quantitative results As shown in Table 2, SplatFlow achieves higher CLIPScore and CLIP D-sim than DGE [7], indicating more accurate and effective editing. SDS++ refinement further boosts performance, demonstrating enhanced alignment with target edits. + +Qualitative results Figure 4 demonstrates that SplatFlow performs comparably to MVInpainter, a specialized 2D inpainting method that leverages an already edited high-quality image. On the other hand, DGE struggles to replace objects fully, often just altering styles instead of achieving complete transformations. This underscores SplatFlow's superior ability in precise 3D object replacement. Please refer to Appendix C for further qualitative results. + +# 5.4. Result on Inpainting Application + +Here, we present the adaptability of our multi-view RF model to camera pose estimation and novel-view synthesis via the inpainting technique on the MVImgNet dataset [98]. + +Camera pose estimation To demonstrate the effectiveness of our multi-view RF model in camera pose estimation, we primarily compare our approach to RayDiffusion [101] and RelPose++ [43]. We conducted comparisons using an + +
MethodRotation↑Camera Center↑
@5@10@15@0.05@0.1@0.2
RelPose++ [43]19.437.751.40.612.555.0
Ray Regression [101]10.425.650.115.347.982.9
Ray Diffusion [101]17.538.759.624.160.987.6
SplatFlow (w/ depth)26.852.659.362.391.699.4
SplatFlow (w/o depth)28.854.563.964.994.099.7
+ +![](images/b57d6b733081eee53271290b47a7c88ef5fc3eb25e1114becb7e0f343a4d9ffd.jpg) +Figure 5. Qualitative results for camera pose estimation. Camera poses are estimated from multi-view images. Image border colors match each camera, with black cameras indicating GT poses. + +![](images/e3e88792debe0a2d7dd9bfdf15666e4d8a8b5a66aef70d55d1659f4a22fde5f1.jpg) + +8-view setup, inpainting the ray latent with and without the depth latent from DepthAnythingV2 [93]. We used rotation accuracy@ $Q_{R}$ and translation accuracy@ $Q_{T}$ as evaluation metrics, measuring the proportion of relative camera rotations within $Q_{R}$ degrees of ground truth and camera centers within $Q_{T}$ units of scene scale, respectively. + +Table 3 and Fig. 5 present the quantitative and qualitative results. Overall, the multi-view RF model achieves superior performance across most scenarios, outperforming the baseline methods. Interestingly, the multi-view RF model achieves higher accuracy without depth latents than with them. We hypothesize that this outcome is due to the joint generation of depth and camera poses, which may yield more detailed and contextually enriched depth estimates. + +Novel view-synthesis We evaluated the performance of our multi-view RF model in novel view synthesis under two scenarios: 1) interpolation, with the uniformly sampled input views, and 2) extrapolation, with the input views positioned in the center. This comparison validates our multi-view RF model's capability of 3D reasoning based on relative viewpoints and evaluates its effectiveness in generating novel views. We evaluated PSNR, SSIM, and LPIPS [104] against the ground-truth images for reconstruction quality, as well as absolute relative difference and $\delta_{1}$ accuracy using depth maps produced by DepthAnythingV2 [93]. + +Table 3. Results in camera pose estimation on MVImgNet validation set. @Q represents the accuracy threshold for rotations (degrees) and camera centers (units). + +
TypeRGBDepth
PSNR↑SSIM↑LPIPS↓AbsRel↓δ1↑
Interpolation (N=2)14.730.5710.6480.5880.731
Interpolation (N=4)17.050.5900.5510.4980.761
Interpolation (N=6)18.820.6260.4830.4150.775
Extrapolation (N=2)15.150.5770.6270.7710.715
Extrapolation (N=4)16.800.5950.5540.6790.727
Extrapolation (N=6)17.960.6130.5030.6020.747
+ +Table 4. Novel view synthesis results on MVImgNet. We use $N$ input views to synthesize $K - N$ novel views, with uniformly sampled views for interpolation and central views for extrapolation. + +![](images/26e343442881543e49df324d7c91b6ae61ee85277f90e733dbeef9ad7a21dfcf.jpg) +Figure 6. Qualitative results for novel view synthesis. Novel view synthesis is performed from the red-box images and depths. + +As shown in Fig. 6, our multi-view RF model successfully generates novel views for both scenarios. Also, Table 4 shows that interpolation outperforms extrapolation in both RGB reconstruction and depth estimation as the number of input views increases, indicating that interpolation leverages 3D reasoning for accurate novel view generation, whereas extrapolation focuses on generating diverse novel views. + +# 6. Conclusion + +In this paper, we have introduced SplatFlow, a novel framework that unifies 3D Gaussian Splitting (3DGS) generation and editing within a single, efficient model. By designing a multi-view Rectified Flow (RF) model alongside a Gaussian Splitting Decoder (GSDecoder), we achieved direct 3DGS generation capable of addressing real-world challenges, such as varying scales and complex camera trajectories. Our approach utilizes joint modeling of multi-view images, depths, and camera poses, enabling seamless integration of training-free inversion and inpainting techniques for a diverse set of 3D tasks. As we present the versatility of our SplatFlow, we believe SplatFlow represents an important step toward a versatile foundational model for 3D content. + +# References + +[1] Michael S Albergo and Eric Vanden-Eijnden. Building normalizing flows with stochastic interpolants. arXiv preprint arXiv:2209.15571, 2022. 2, 3 +[2] Chong Bao, Yinda Zhang, Bangbang Yang, Tianxing Fan, Zesong Yang, Hujun Bao, Guofeng Zhang, and Zhaopeng Cui. Sine: Semantic-driven image-based nerf editing with prior-guided editing field. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20919-20929, 2023. 1, 3 +[3] Chenjie Cao, Chaohui Yu, Yanwei Fu, Fan Wang, and Xi-angyang Xue. Mvinpainter: Learning multi-view consistent inpainting to bridge 2d and 3d editing. arXiv preprint arXiv:2408.08000, 2024.7 +[4] David Charatan, Sizhe Lester Li, Andrea Tagliasacchi, and Vincent Sitzmann. pixelsplat: 3d gaussian splats from image pairs for scalable generalizable 3d reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19457-19467, 2024. 2, 4 +[5] Anpei Chen, Haofei Xu, Stefano Esposito, Siyu Tang, and Andreas Geiger. Lara: Efficient large-baseline radiance fields. In European Conference on Computer Vision, pages 338-355. Springer, 2025. 3 +[6] Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, et al. Pixart-α: Fast training of diffusion transformer for photorealistic text-to-image synthesis. arXiv preprint arXiv:2310.00426, 2023. 2, 3 +[7] Minghao Chen, Iro Laina, and Andrea Vedaldi. Dge: Direct gaussian 3d editing by consistent multi-view editing. arXiv preprint arXiv:2404.18929, 2024. 2, 7 +[8] Yiwen Chen, Zilong Chen, Chi Zhang, Feng Wang, Xiaofeng Yang, Yikai Wang, Zhongang Cai, Lei Yang, Huaping Liu, and Guosheng Lin. Gaussianeditor: Swift and controllable 3d editing with gaussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21476-21485, 2024. 2, 3, 7 +[9] Yuedong Chen, Haofei Xu, Chuanxia Zheng, Bohan Zhuang, Marc Pollefeys, Andreas Geiger, Tat-Jen Cham, and Jianfei Cai. Mvsplat: Efficient 3d gaussian splattering from sparse multi-view images. In European Conference on Computer Vision, pages 370-386. Springer, 2024. 2, 4 +[10] Zilong Chen, Feng Wang, Yikai Wang, and Huaping Liu. Text-to-3d using gaussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21401-21412, 2024. 2, 3 +[11] Hyungjin Chung, Jeongsol Kim, Michael T McCann, Marc L Klasky, and Jong Chul Ye. Diffusion posterior sampling for general noisy inverse problems. arXiv preprint arXiv:2209.14687, 2022. 2, 5 +[12] Jaeyoung Chung, Suyoung Lee, Hyeongjin Nam, Jaerin Lee, and Kyoung Mu Lee. Luciddreamer: Domain-free generation of 3d gaussian splatting scenes. arXiv preprint arXiv:2311.13384, 2023. 3, 6, 7 +[13] Guillaume Couairon, Jakob Verbeek, Holger Schwenk, and Matthieu Cord. Diffedit: Diffusion-based seman + +tic image editing with mask guidance. arXiv preprint arXiv:2210.11427, 2022. 2, 5 +[14] Giannis Daras, Hyungjin Chung, Chieh-Hsin Lai, Yuki Mitsufuji, Jong Chul Ye, Peyman Milanfar, Alexandros G Dimakis, and Mauricio Delbracio. A survey on diffusion models for inverse problems. arXiv preprint arXiv:2410.00083, 2024. 2 +[15] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3d objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13142-13153, 2023. 2, 3 +[16] Matt Deitke, Ruoshi Liu, Matthew Wallingford, Huong Ngo, Oscar Michel, Aditya Kusupati, Alan Fan, Christian Laforte, Vikram Voleti, Samir Yitzhak Gadre, et al. Objverse-xl: A universe of $10\mathrm{m} + 3\mathrm{d}$ objects. Advances in Neural Information Processing Systems, 36, 2024. 2, 3 +[17] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 2 +[18] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12873-12883, 2021. 4 +[19] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In *Forty-first International Conference on Machine Learning*, 2024. 2, 4, 5 +[20] Hyojun Go, Yunsung Lee, Jin-Young Kim, Seunghyun Lee, Myeongho Jeong, Hyun Seung Lee, and Seungtaek Choi. Towards practical plug-and-play diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1962-1971, 2023. 2 +[21] Hyojun Go, Yunsung Lee, Seunghyun Lee, Shinhyeok Oh, Hyeongdon Moon, and Seungtaek Choi. Addressing negative transfer in diffusion models. Advances in Neural Information Processing Systems, 36, 2024. 2 +[22] Alexandros Graikos, Nikolay Malkin, Nebojsa Jojic, and Dimitris Samaras. Diffusion models as plug-and-play priors. Advances in Neural Information Processing Systems, 35: 14715-14728, 2022. 2 +[23] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 2 +[24] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning. arXiv preprint arXiv:2104.08718, 2021.6 +[25] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 6 + +[26] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2, 3 +[27] Jonathan Ho, William Chan, Chitwan Sahara, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. 2 +[28] Lukas Hollein, Ang Cao, Andrew Owens, Justin Johnson, and Matthias Nießner. Text2room: Extracting textured 3d meshes from 2d text-to-image models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7909-7920, 2023. 3 +[29] Hiromichi Kamata, Yuiko Sakuma, Akio Hayakawa, Masato Ishii, and Takuya Narihira. Instruct 3d-to-3d: Text instruction guided 3d-to-3d conversion. arXiv preprint arXiv:2303.15780, 2023. 1, 3, 7 +[30] Nazmul Karim, Hasan Iqbal, Umar Khalid, Chen Chen, and Jing Hua. Free-editor: zero-shot text-driven 3d scene editing. In European Conference on Computer Vision, pages 436-453. Springer, 2025. 1, 3 +[31] Bingxin Ke, Anton Obukhov, Shengyu Huang, Nando Metzger, Rodrigo Caye Daudt, and Konrad Schindler. Repurposing diffusion-based image generators for monocular depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9492-9502, 2024. 4 +[32] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splattering for real-time radiance field rendering. ACM Trans. Graph., 42(4):139-1, 2023. 2, 3, 4 +[33] Umar Khalid, Hasan Iqbal, Nazmul Karim, Jing Hua, and Chen Chen. Latenteditor: Text driven local editing of 3d scenes. arXiv preprint arXiv:2312.09313, 2023. 1, 3 +[34] Umar Khalid, Hasan Iqbal, Azib Farooq, Jing Hua, and Chen Chen. 3dego: 3d editing on the go! In European Conference on Computer Vision, pages 73-89. Springer, 2024. 2, 3 +[35] Gwanghyun Kim, Taesung Kwon, and Jong Chul Ye. Diffusionclip: Text-guided diffusion models for robust image manipulation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2426-2435, 2022. 2 +[36] Nupur Kumari, Richard Zhang, Eli Shechtman, and JunYan Zhu. Ensembling off-the-shelf models for gan training. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10651-10662, 2022. 4, 5 +[37] Yunsung Lee, JinYoung Kim, Hyojun Go, Myeongho Jeong, Shinhyeok Oh, and Seungtaek Choi. Multi-architecture multi-expert diffusion models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 13427-13436, 2024. 2 +[38] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 6 +[39] Haoran Li, Haolin Shi, Wenli Zhang, Wenjun Wu, Yong Liao, Lin Wang, Lik-hang Lee, and Pengyuan Zhou. Dreamscene: + +3d gaussian-based text-to-3d scene generation via formation pattern sampling. arXiv preprint arXiv:2404.03575, 2024.6, 7 +[40] Xiang Li, John Thickstun, Ishaan Gulrajani, Percy S Liang, and Tatsunori B Hashimoto. Diffusion-lm improves controllable text generation. Advances in Neural Information Processing Systems, 35:4328-4343, 2022. 2 +[41] Xinyang Li, Zhangyu Lai, Linning Xu, Yansong Qu, Liujuan Cao, Shengchuan Zhang, Bo Dai, and Rongrong Ji. Director3d: Real-world camera trajectory and 3d scene generation from text. Advances in neural information processing systems, 2024. 3, 5, 6, 7 +[42] Zongrui Li, Minghui Hu, Qian Zheng, and Xudong Jiang. Connecting consistency distillation to score distillation for text-to-3d generation. In European Conference on Computer Vision, pages 274–291. Springer, 2024. 2 +[43] Amy Lin, Jason Y Zhang, Deva Ramanan, and Shubham Tulsiani. Relpose++: Recovering 6d poses from sparse-view observations. arXiv preprint arXiv:2305.04926, 2023. 7, 8 +[44] Huan Ling, Seung Wook Kim, Antonio Torralba, Sanja Fidler, and Karsten Kreis. Align your gaussians: Text-to-4d with dynamic 3d gaussians and composed diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8576–8588, 2024. 3 +[45] Lu Ling, Yichen Sheng, Zhi Tu, Wentian Zhao, Cheng Xin, Kun Wan, Lantao Yu, Qianyu Guo, Zixun Yu, Yawen Lu, et al. D13dv-10k: A large-scale scene dataset for deep learning-based 3d vision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22160–22169, 2024. 2, 6, 7 +[46] Yaron Lipman, Ricky TQ Chen, Heli Ben-Hamu, Maximilian Nickel, and Matt Le. Flow matching for generative modeling. arXiv preprint arXiv:2210.02747, 2022. 2, 3 +[47] Xingchao Liu, Chengyue Gong, and Qiang Liu. Flow straight and fast: Learning to generate and transfer data with rectified flow. arXiv preprint arXiv:2209.03003, 2022. 2, 3 +[48] Xian Liu, Xiaohang Zhan, Jiaxiang Tang, Ying Shan, Gang Zeng, Dahua Lin, Xihui Liu, and Ziwei Liu. Humangaussian: Text-driven 3d human generation with gaussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6646-6657, 2024. 2 +[49] Yuan Liu, Cheng Lin, Zijiao Zeng, Xiaoxiao Long, Lingjie Liu, Taku Komura, and Wenping Wang. Syncdreamer: Generating multiview-consistent images from a single-view image. arXiv preprint arXiv:2309.03453, 2023. 3 +[50] Xiaoxiao Long, Yuan-Chen Guo, Cheng Lin, Yuan Liu, Zhiyang Dou, Lingjie Liu, Yuexin Ma, Song-Hai Zhang, Marc Habermann, Christian Theobalt, et al. Wonder3d: Single image to 3d using cross-domain diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9970-9980, 2024. 3 +[51] Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11461-11471, 2022. 2, 4, 5 + +[52] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jiajun Wu, Jun-Yan Zhu, and Stefano Ermon. Sdedit: Guided image synthesis and editing with stochastic differential equations. arXiv preprint arXiv:2108.01073, 2021. 2, 4, 5 +[53] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 1, 2 +[54] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6038–6047, 2023. 2, 5 +[55] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM transactions on graphics (TOG), 41(4):1-15, 2022. 1 +[56] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 2 +[57] Alex Nichol, Heewoo Jun, Prafulla Dhariwal, Pamela Mishkin, and Mark Chen. Point-e: A system for generating 3d point clouds from complex prompts. arXiv preprint arXiv:2212.08751, 2022. 2 +[58] Francesco Palandra, Andrea Sanchietti, Daniele Baieri, and Emanuele Rodola. Gsedit: Efficient text-guided editing of 3d objects via gaussian splatting. arXiv preprint arXiv:2403.05154, 2024. 2, 3 +[59] Byeongjun Park and Changick Kim. Point-dynrf: Point-based dynamic radiance fields from a monocular video. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 3171-3181, 2024. 3 +[60] Byeongjun Park, Sangmin Woo, Hyojun Go, Jin-Young Kim, and Changick Kim. Denoising task routing for diffusion models. arXiv preprint arXiv:2310.07138, 2023. 2 +[61] Byeongjun Park, Hyojun Go, and Changick Kim. Bridging implicit and explicit geometric transformation for single-image view synthesis. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 2 +[62] Byeongjun Park, Hyojun Go, Jin-Young Kim, Sangmin Woo, Seokil Ham, and Changick Kim. Switch diffusion transformer: Synergizing denoising tasks with sparse mixture-of-experts. arXiv preprint arXiv:2403.09176, 2024. 2 +[63] Jangho Park, Gihyun Kwon, and Jong Chul Ye. Ed-nerf: Efficient text-guided editing of 3d scene using latent space nerf. arXiv preprint arXiv:2310.02712, 2023. 1, 3 +[64] Julius Plücker. Analytisch-geometrische Entwicklungen. GD Baedeker, 1828. 5 +[65] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 2, 3 + +[66] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 2 +[67] Elad Richardson, Gal Metzer, Yuval Alaluf, Raja Giryes, and Daniel Cohen-Or. Texture: Text-guided texturing of 3d shapes. In ACM SIGGRAPH 2023 conference proceedings, pages 1–11, 2023. 3 +[68] Robin Rombach, Patrick Esser, and Björn Ommer. Geometry-free view synthesis: Transformers and no 3d priors. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14356-14366, 2021. 2 +[69] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2, 3, 4 +[70] Litu Rout, Advait Parulekar, Constantine Caramanis, and Sanjay Shakkottai. A theoretical justification for image inpainting using denoising diffusion probabilistic models. arXiv preprint arXiv:2302.01217, 2023. 2, 5 +[71] Litu Rout, Negin Raoof, Giannis Daras, Constantine Caramanis, Alex Dimakis, and Sanjay Shakkottai. Solving linear inverse problems provably via posterior sampling with latent diffusion models. Advances in Neural Information Processing Systems, 36, 2023. 2, 5 +[72] Axel Sauer, Dominik Lorenz, Andreas Blattmann, and Robin Rombach. Adversarial diffusion distillation. In European Conference on Computer Vision, pages 87-103. Springer, 2024. 2 +[73] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278-25294, 2022. 3, 7 +[74] Yichun Shi, Peng Wang, Jianglong Ye, Mai Long, Kejie Li, and Xiao Yang. Mvdream: Multi-view diffusion for 3d generation. arXiv preprint arXiv:2308.16512, 2023. 3 +[75] Jaidev Shriram, Alex Trevithick, Lingjie Liu, and Ravi Ramamoorthi. Realmdreamer: Text-driven 3d scene generation with inpainting and depth diffusion. arXiv preprint arXiv:2404.07199, 2024. 3 +[76] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International conference on machine learning, pages 2256-2265. PMLR, 2015. 2,3 +[77] Jiaming Song, Arash Vahdat, Morteza Mardani, and Jan Kautz. Pseudoinverse-guided diffusion models for inverse problems. In International Conference on Learning Representations, 2023. 2, 5 +[78] Jiaming Song, Qinsheng Zhang, Hongxu Yin, Morteza Mardani, Ming-Yu Liu, Jan Kautz, Yongxin Chen, and Arash Vahdat. Loss-guided diffusion models for plug-and-play controllable generation. In International Conference on Machine Learning, pages 32483-32498. PMLR, 2023. 2 + +[79] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 2, 3 +[80] Stanislaw Szymanowicz, Chrisitian Rupprecht, and Andrea Vedaldi. Splatter image: Ultra-fast single-view 3d reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10208-10217, 2024. 4 +[81] Jiaxiang Tang, Zhaoxi Chen, Xiaokang Chen, Tengfei Wang, Gang Zeng, and Ziwei Liu. Lgm: Large multi-view gaussian model for high-resolution 3d content creation. In European Conference on Computer Vision, pages 1-18. Springer, 2024. 2, 3, 4 +[82] Jiaxiang Tang, Jiawei Ren, Hang Zhou, Ziwei Liu, and Gang Zeng. Dreamgaussian: Generative gaussian splatting for efficient 3d content creation. In The Twelfth International Conference on Learning Representations, 2024. 2, 3 +[83] Arash Vahdat, Francis Williams, Zan Gojcic, Or Litany, Sanja Fidler, Karsten Kreis, et al. Lion: Latent point diffusion models for 3d shape generation. Advances in Neural Information Processing Systems, 35:10021-10039, 2022. 2 +[84] Alexander Vilesov, Pradyumna Chari, and Achuta Kadambi. Cg3d: Compositional generation for text-to-3d via gaussian splatting. arXiv preprint arXiv:2311.17907, 2023. 2 +[85] Vikram Voleti, Chun-Han Yao, Mark Boss, Adam Letts, David Pankratz, Dmitry Tochilkin, Christian Laforte, Robin Rombach, and Varun Jampani. Sv3d: Novel multi-view synthesis and 3d generation from a single image using latent video diffusion. In European Conference on Computer Vision, pages 439-457. Springer, 2025. 3 +[86] Can Wang, Mingming He, Mengei Chai, Dongdong Chen, and Jing Liao. Mesh-guided neural implicit field editing. arXiv preprint arXiv:2312.02157, 2023. 3 +[87] Junjie Wang, Jiemin Fang, Xiaopeng Zhang, Lingxi Xie, and Qi Tian. Gaussianeditor: Editing 3d gaussians delicately with text instructions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20902-20911, 2024. 2, 3 +[88] Yuxuan Wang, Xuanyu Yi, Zike Wu, Na Zhao, Long Chen, and Hanwang Zhang. View-consistent 3d editing with gaussian splatting. In European Conference on Computer Vision, pages 404-420. Springer, 2024. 2, 3, 5 +[89] Sangmin Woo, Byeongjun Park, Hyojun Go, Jin-Young Kim, and Changick Kim. Harmonyview: Harmonizing consistency and diversity in one-image-to-3d. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10574–10584, 2024. 3 +[90] Jing Wu, Jia-Wang Bian, Xinghui Li, Guangrun Wang, Ian Reid, Philip Torr, and Victor Adrian Prisacariu. Gaussctrl: multi-view consistent text-driven 3d gaussian splattering editing. arXiv preprint arXiv:2403.08733, 2024. 2, 3, 5 +[91] Hanyuan Xiao, Yingshu Chen, Huajian Huang, Haolin Xiong, Jing Yang, Pratusha Prasad, and Yajie Zhao. Localized gaussian splatting editing with contextual awareness. arXiv preprint arXiv:2408.00083, 2024. 2, 3 +[92] Yinghao Xu, Hao Tan, Fujun Luan, Sai Bi, Peng Wang, Jiahao Li, Zifan Shi, Kalyan Sunkavalli, Gordon Wetzstein, + +Zexiang Xu, and Kai Zhang. DMV3d: Denoising multiview diffusion using 3d large reconstruction model. In The Twelfth International Conference on Learning Representations, 2024. 2 +[93] Lihe Yang, Bingyi Kang, Zilong Huang, Zhen Zhao, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything v2. arXiv preprint arXiv:2406.09414, 2024. 4, 8 +[94] Taoran Yi, Jiemin Fang, Junjie Wang, Guanjun Wu, Lingxi Xie, Xiaopeng Zhang, Wenyu Liu, Qi Tian, and Xinggang Wang. Gaussian dreamer: Fast generation from text to 3d gaussians by bridging 2d and 3d diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6796-6807, 2024. 2 +[95] Xuanyu Yi, Zike Wu, Qingshan Xu, Pan Zhou, Joo-Hwee Lim, and Hanwang Zhang. Diffusion time-step curriculum for one image to 3d generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9948-9958, 2024. 3 +[96] Lu Yu, Wei Xiang, and Kang Han. Edit-diffnerf: Editing 3d neural radiance fields using 2d diffusion model. arXiv preprint arXiv:2306.09551, 2023. 1, 3 +[97] Wangbo Yu, Jinbo Xing, Li Yuan, Wenbo Hu, Xiaoyu Li, Zhipeng Huang, Xiangjun Gao, Tien-Tsin Wong, Ying Shan, and Yonghong Tian. Viewcrafter: Taming video diffusion models for high-fidelity novel view synthesis. arXiv preprint arXiv:2409.02048, 2024. 2 +[98] Xianggang Yu, Mutian Xu, Yidan Zhang, Haolin Liu, Chongjie Ye, Yushuang Wu, Zizheng Yan, Chenming Zhu, Zhangyang Xiong, Tianyou Liang, et al. Mvimgnet: A large-scale dataset of multi-view images. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9150-9161, 2023. 2, 6, 7 +[99] Bowen Zhang, Yiji Cheng, Jiaolong Yang, Chunyu Wang, Feng Zhao, Yansong Tang, Dong Chen, and Baining Guo. Gaussiancube: Structuring gaussian splatting using optimal transport for 3d generative modeling. arXiv preprint arXiv:2403.19655, 2024. 2 +[100] Jingbo Zhang, Xiaoyu Li, Ziyu Wan, Can Wang, and Jing Liao. Text2nerf: Text-driven 3d scene generation with neural radiance fields. IEEE Transactions on Visualization & Computer Graphics, 30(12):7749-7762, 2024. 1 +[101] Jason Y. Zhang, Amy Lin, Moneish Kumar, Tzu-Hsuan Yang, Deva Ramanan, and Shubham Tulsiani. Cameras as rays: Pose estimation via ray diffusion. In The Twelfth International Conference on Learning Representations, 2024. 2, 5, 7, 8 +[102] Kai Zhang, Sai Bi, Hao Tan, Yuanbo Xiangli, Nanxuan Zhao, Kalyan Sunkavalli, and Zexiang Xu. Gs-lrm: Large reconstruction model for 3d gaussian splatting. In European Conference on Computer Vision, pages 1-19. Springer, 2024. 2 +[103] Kaiyi Zhang, Yang Chen, Ximing Yang, Weizhong Zhang, and Cheng Jin. Point cloud part editing: Segmentation, generation, assembly, and selection. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 7187-7195, 2024. 3 +[104] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of + +deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 5, 8 +[105] Yudi Zhang, Qi Xu, and Lei Zhang. Dragtex: Generative point-based texture editing on 3d mesh. arXiv preprint arXiv:2403.02217, 2024. 3 +[106] Linqi Zhou, Yilun Du, and Jiajun Wu. 3d shape generation and completion through point-voxel diffusion. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5826-5835, 2021. 2 +[107] Shijie Zhou, Zhiwen Fan, Dejia Xu, Haoran Chang, Pradyumna Chari, Tejas Bharadwaj, Suya You, Zhangyang Wang, and Achuta Kadambi. Dreamscene360: Unconstrained text-to-3d scene generation with panoramic gaussian splatting. In European Conference on Computer Vision, pages 324–342. Springer, 2025. 3 +[108] Jingyu Zhuang, Di Kang, Yan-Pei Cao, Guanbin Li, Liang Lin, and Ying Shan. Tip-editor: An accurate 3d editor following both text-prompts and image-prompts. ACM Transactions on Graphics (TOG), 43(4):1-12, 2024. 2, 3 \ No newline at end of file diff --git a/2025/SplatFlow_ Multi-View Rectified Flow Model for 3D Gaussian Splatting Synthesis/images.zip b/2025/SplatFlow_ Multi-View Rectified Flow Model for 3D Gaussian Splatting Synthesis/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..02b4f957e032640fda7cf47fba1ff0c74d81fc6c --- /dev/null +++ b/2025/SplatFlow_ Multi-View Rectified Flow Model for 3D Gaussian Splatting Synthesis/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:26f589ca6277c35c9000ad98bc5bd4adb48e980a3a0bb4923487f29f0930939a +size 820786 diff --git a/2025/SplatFlow_ Multi-View Rectified Flow Model for 3D Gaussian Splatting Synthesis/layout.json b/2025/SplatFlow_ Multi-View Rectified Flow Model for 3D Gaussian Splatting Synthesis/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..35c712c78c9b7ed132fac0a3e28f02a8b00171c9 --- /dev/null +++ b/2025/SplatFlow_ Multi-View Rectified Flow Model for 3D Gaussian Splatting Synthesis/layout.json @@ -0,0 +1,11829 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 55, + 103, + 555, + 122 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 103, + 555, + 122 + ], + "spans": [ + { + "bbox": [ + 55, + 103, + 555, + 122 + ], + "type": "text", + "content": "SplatFlow: Multi-View Rectified Flow Model for 3D Gaussian Splitting Synthesis" + } + ] + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 129, + 141, + 478, + 197 + ], + "blocks": [ + { + "bbox": [ + 129, + 141, + 478, + 197 + ], + "lines": [ + { + "bbox": [ + 129, + 141, + 478, + 197 + ], + "spans": [ + { + "bbox": [ + 129, + 141, + 478, + 197 + ], + "type": "image", + "image_path": "064a97fe37b4b7c8d294be09866d237d0ce22ebffe1b816eceade2541a0823e8.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 59, + 212, + 144, + 245 + ], + "blocks": [ + { + "bbox": [ + 59, + 212, + 144, + 245 + ], + "lines": [ + { + "bbox": [ + 59, + 212, + 144, + 245 + ], + "spans": [ + { + "bbox": [ + 59, + 212, + 144, + 245 + ], + "type": "image", + "image_path": "c818755f2e286d6b4dbfb9b84c92e072c6c88bdd7bd578f66dc09abf9fa4ede6.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 59, + 250, + 144, + 285 + ], + "blocks": [ + { + "bbox": [ + 59, + 250, + 144, + 285 + ], + "lines": [ + { + "bbox": [ + 59, + 250, + 144, + 285 + ], + "spans": [ + { + "bbox": [ + 59, + 250, + 144, + 285 + ], + "type": "image", + "image_path": "376b9af1fd76759b9f5241ed8e9a204b443a0854db268c54225001534e0dabc2.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 59, + 286, + 144, + 327 + ], + "blocks": [ + { + "bbox": [ + 59, + 286, + 144, + 327 + ], + "lines": [ + { + "bbox": [ + 59, + 286, + 144, + 327 + ], + "spans": [ + { + "bbox": [ + 59, + 286, + 144, + 327 + ], + "type": "image", + "image_path": "f71e8b7573d64979aa4dc12bf96889f72cc6cdcc07a34be67142f9cc7f6309c9.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 59, + 327, + 144, + 357 + ], + "blocks": [ + { + "bbox": [ + 59, + 327, + 144, + 357 + ], + "lines": [ + { + "bbox": [ + 59, + 327, + 144, + 357 + ], + "spans": [ + { + "bbox": [ + 59, + 327, + 144, + 357 + ], + "type": "image", + "image_path": "6f8a8e4d6f4321e2852f254fd033c19e129b3e0dc1ee5d4d1cf706760fc01130.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 74, + 356, + 124, + 361 + ], + "lines": [ + { + "bbox": [ + 74, + 356, + 124, + 361 + ], + "spans": [ + { + "bbox": [ + 74, + 356, + 124, + 361 + ], + "type": "text", + "content": "A library with a lot of books" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 59, + 364, + 144, + 399 + ], + "blocks": [ + { + "bbox": [ + 59, + 364, + 144, + 399 + ], + "lines": [ + { + "bbox": [ + 59, + 364, + 144, + 399 + ], + "spans": [ + { + "bbox": [ + 59, + 364, + 144, + 399 + ], + "type": "image", + "image_path": "c2982e5c1a3b23e30470f0d3878d3c0257f7a8f3677789ec62e75728e02484d3.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 418, + 555, + 453 + ], + "lines": [ + { + "bbox": [ + 55, + 418, + 555, + 453 + ], + "spans": [ + { + "bbox": [ + 55, + 418, + 555, + 453 + ], + "type": "text", + "content": "Figure 1. SplatFlow for 3D Gaussian Splitting synthesis and its training-free applications. (a) Examples of direct 3D Gaussian Splitting (3DGS) generation only from text prompts, (b) Training-free applications, including 3DGS object editing, camera pose estimation, and novel view synthesis. SplatFlow seamlessly integrates these capabilities, showcasing its versatility in generating and editing complex 3D content." + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 146, + 212, + 230, + 245 + ], + "blocks": [ + { + "bbox": [ + 146, + 212, + 230, + 245 + ], + "lines": [ + { + "bbox": [ + 146, + 212, + 230, + 245 + ], + "spans": [ + { + "bbox": [ + 146, + 212, + 230, + 245 + ], + "type": "image", + "image_path": "e3425c2d94e5378e44eb4bda7e499b7e0ba88b81f87061a5fcbef324ece2f5b8.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 146, + 250, + 230, + 279 + ], + "blocks": [ + { + "bbox": [ + 146, + 250, + 230, + 279 + ], + "lines": [ + { + "bbox": [ + 146, + 250, + 230, + 279 + ], + "spans": [ + { + "bbox": [ + 146, + 250, + 230, + 279 + ], + "type": "image", + "image_path": "da867f973dfe18f63cb0de40727f306d2b21a6448849fb83ba8b30c0e690529d.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 153, + 279, + 225, + 284 + ], + "lines": [ + { + "bbox": [ + 153, + 279, + 225, + 284 + ], + "spans": [ + { + "bbox": [ + 153, + 279, + 225, + 284 + ], + "type": "text", + "content": "A forest with tall trees and yellow leaves" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 146, + 286, + 230, + 316 + ], + "blocks": [ + { + "bbox": [ + 146, + 286, + 230, + 316 + ], + "lines": [ + { + "bbox": [ + 146, + 286, + 230, + 316 + ], + "spans": [ + { + "bbox": [ + 146, + 286, + 230, + 316 + ], + "type": "image", + "image_path": "f348c71eea6be672c093a5c9ce101abc5edca11b02984c37ee394ffbb9acf3f3.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 146, + 316, + 227, + 327 + ], + "lines": [ + { + "bbox": [ + 146, + 316, + 227, + 327 + ], + "spans": [ + { + "bbox": [ + 146, + 316, + 227, + 327 + ], + "type": "text", + "content": "Blue car on a clean street with buildings in the background." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 146, + 327, + 230, + 356 + ], + "blocks": [ + { + "bbox": [ + 146, + 327, + 230, + 356 + ], + "lines": [ + { + "bbox": [ + 146, + 327, + 230, + 356 + ], + "spans": [ + { + "bbox": [ + 146, + 327, + 230, + 356 + ], + "type": "image", + "image_path": "f426a58a2d365faca2fcd43e57b9f5d9b1aed98931a9fcc7389bd614e4c2a1f8.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 173, + 356, + 198, + 361 + ], + "lines": [ + { + "bbox": [ + 173, + 356, + 198, + 361 + ], + "spans": [ + { + "bbox": [ + 173, + 356, + 198, + 361 + ], + "type": "text", + "content": "Half avocado" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 146, + 364, + 230, + 394 + ], + "blocks": [ + { + "bbox": [ + 146, + 364, + 230, + 394 + ], + "lines": [ + { + "bbox": [ + 146, + 364, + 230, + 394 + ], + "spans": [ + { + "bbox": [ + 146, + 364, + 230, + 394 + ], + "type": "image", + "image_path": "a1cb932aff8f3b6017b9a97d752540c969d5c08fe28c81794f99bd4cddb64911.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 173, + 394, + 200, + 399 + ], + "lines": [ + { + "bbox": [ + 173, + 394, + 200, + 399 + ], + "spans": [ + { + "bbox": [ + 173, + 394, + 200, + 399 + ], + "type": "text", + "content": "slicedzucchini" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 165, + 403, + 283, + 413 + ], + "lines": [ + { + "bbox": [ + 165, + 403, + 283, + 413 + ], + "spans": [ + { + "bbox": [ + 165, + 403, + 283, + 413 + ], + "type": "text", + "content": "(a) 3D Gaussian Splitting Synthesis" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 233, + 212, + 398, + 291 + ], + "blocks": [ + { + "bbox": [ + 233, + 212, + 398, + 291 + ], + "lines": [ + { + "bbox": [ + 233, + 212, + 398, + 291 + ], + "spans": [ + { + "bbox": [ + 233, + 212, + 398, + 291 + ], + "type": "image", + "image_path": "c89704bb8cca0486ed6f9faaca339c1fb0b7517139ae2d6969f9404c61cea928.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 278, + 291, + 351, + 298 + ], + "lines": [ + { + "bbox": [ + 278, + 291, + 351, + 298 + ], + "spans": [ + { + "bbox": [ + 278, + 291, + 351, + 298 + ], + "type": "text", + "content": "A lighthouse on a rocky shore" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 233, + 301, + 398, + 384 + ], + "blocks": [ + { + "bbox": [ + 233, + 301, + 398, + 384 + ], + "lines": [ + { + "bbox": [ + 233, + 301, + 398, + 384 + ], + "spans": [ + { + "bbox": [ + 233, + 301, + 398, + 384 + ], + "type": "image", + "image_path": "902b10d04f2786fc83d5c96b0918ae15114974c46f0d7145a4e07a495a5be346.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 247, + 384, + 373, + 399 + ], + "lines": [ + { + "bbox": [ + 247, + 384, + 373, + 399 + ], + "spans": [ + { + "bbox": [ + 247, + 384, + 373, + 399 + ], + "type": "text", + "content": "A vibrant bush with pink flowers stands in front of a concrete wall, ..." + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 406, + 213, + 439, + 279 + ], + "blocks": [ + { + "bbox": [ + 406, + 213, + 439, + 279 + ], + "lines": [ + { + "bbox": [ + 406, + 213, + 439, + 279 + ], + "spans": [ + { + "bbox": [ + 406, + 213, + 439, + 279 + ], + "type": "image", + "image_path": "1358ff03ebe3bed73134b5ab8e566f41f675fa0c589f08776e9995e744b68f35.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 439, + 279, + 518, + 287 + ], + "lines": [ + { + "bbox": [ + 439, + 279, + 518, + 287 + ], + "spans": [ + { + "bbox": [ + 439, + 279, + 518, + 287 + ], + "type": "text", + "content": "3DGS object editing (replacement)" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 441, + 213, + 492, + 279 + ], + "blocks": [ + { + "bbox": [ + 441, + 213, + 492, + 279 + ], + "lines": [ + { + "bbox": [ + 441, + 213, + 492, + 279 + ], + "spans": [ + { + "bbox": [ + 441, + 213, + 492, + 279 + ], + "type": "image", + "image_path": "6458c8efd497d95c7d4843e6f151d69dbf12531a2bebedaed111ced4fc55e817.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 493, + 213, + 525, + 279 + ], + "blocks": [ + { + "bbox": [ + 493, + 213, + 525, + 279 + ], + "lines": [ + { + "bbox": [ + 493, + 213, + 525, + 279 + ], + "spans": [ + { + "bbox": [ + 493, + 213, + 525, + 279 + ], + "type": "image", + "image_path": "8fd0d6c19f961b167132882ee6fd41109fc3fb6900a2583bd0df23a82ea11a5d.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 528, + 213, + 543, + 279 + ], + "blocks": [ + { + "bbox": [ + 528, + 213, + 543, + 279 + ], + "lines": [ + { + "bbox": [ + 528, + 213, + 543, + 279 + ], + "spans": [ + { + "bbox": [ + 528, + 213, + 543, + 279 + ], + "type": "image", + "image_path": "ebbbe4a5a18b58e9bc493fce0d2fb3063d97c55375c072502ea8fcd57762a3c7.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 406, + 293, + 439, + 336 + ], + "blocks": [ + { + "bbox": [ + 406, + 293, + 439, + 336 + ], + "lines": [ + { + "bbox": [ + 406, + 293, + 439, + 336 + ], + "spans": [ + { + "bbox": [ + 406, + 293, + 439, + 336 + ], + "type": "image", + "image_path": "6b46d5028c10cd12cc29d8cf1ec9fc092f357a23c7ac29ef734984289c1bfb61.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 440, + 293, + 495, + 336 + ], + "blocks": [ + { + "bbox": [ + 440, + 293, + 495, + 336 + ], + "lines": [ + { + "bbox": [ + 440, + 293, + 495, + 336 + ], + "spans": [ + { + "bbox": [ + 440, + 293, + 495, + 336 + ], + "type": "image", + "image_path": "3ce1b93006ec113de40bda8dec6865b2604a8b5e8d075c8b5f7c79b4dfcba96a.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 443, + 337, + 503, + 343 + ], + "lines": [ + { + "bbox": [ + 443, + 337, + 503, + 343 + ], + "spans": [ + { + "bbox": [ + 443, + 337, + 503, + 343 + ], + "type": "text", + "content": "\"Camera Pose Estimation\"" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_caption" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 498, + 293, + 539, + 335 + ], + "blocks": [ + { + "bbox": [ + 498, + 293, + 539, + 335 + ], + "lines": [ + { + "bbox": [ + 498, + 293, + 539, + 335 + ], + "spans": [ + { + "bbox": [ + 498, + 293, + 539, + 335 + ], + "type": "image", + "image_path": "7ec74f2631bb999bf9412660f65b05af585fb47d14b19fb9f7898c598e11eff8.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 406, + 353, + 439, + 388 + ], + "blocks": [ + { + "bbox": [ + 406, + 353, + 439, + 388 + ], + "lines": [ + { + "bbox": [ + 406, + 353, + 439, + 388 + ], + "spans": [ + { + "bbox": [ + 406, + 353, + 439, + 388 + ], + "type": "image", + "image_path": "3b0f77f8518208b80ecf34b7c05a8d692ee584d5e26218b419695991d8736551.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 423, + 403, + 525, + 413 + ], + "lines": [ + { + "bbox": [ + 423, + 403, + 525, + 413 + ], + "spans": [ + { + "bbox": [ + 423, + 403, + 525, + 413 + ], + "type": "text", + "content": "(b) Training-Free Applications" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_caption" + } + ], + "index": 33 + }, + { + "type": "image", + "bbox": [ + 441, + 353, + 492, + 388 + ], + "blocks": [ + { + "bbox": [ + 441, + 353, + 492, + 388 + ], + "lines": [ + { + "bbox": [ + 441, + 353, + 492, + 388 + ], + "spans": [ + { + "bbox": [ + 441, + 353, + 492, + 388 + ], + "type": "image", + "image_path": "092483e1230096e6b5a42dc843f9cf6657e7d5b3334930482438f0117b60eeb0.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 443, + 388, + 496, + 396 + ], + "lines": [ + { + "bbox": [ + 443, + 388, + 496, + 396 + ], + "spans": [ + { + "bbox": [ + 443, + 388, + 496, + 396 + ], + "type": "text", + "content": "\"Novel View Synthesis\"" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_caption" + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 493, + 353, + 542, + 388 + ], + "blocks": [ + { + "bbox": [ + 493, + 353, + 542, + 388 + ], + "lines": [ + { + "bbox": [ + 493, + 353, + 542, + 388 + ], + "spans": [ + { + "bbox": [ + 493, + 353, + 542, + 388 + ], + "type": "image", + "image_path": "44a82faf846b356c7b18da5c281a23891930af98e26b4577cc0fa179d9df8bdf.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + } + ], + "index": 35 + }, + { + "bbox": [ + 152, + 468, + 200, + 480 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 468, + 200, + 480 + ], + "spans": [ + { + "bbox": [ + 152, + 468, + 200, + 480 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 39 + }, + { + "bbox": [ + 54, + 493, + 297, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 493, + 297, + 685 + ], + "spans": [ + { + "bbox": [ + 54, + 493, + 297, + 685 + ], + "type": "text", + "content": "Text-based generation and editing of 3D scenes hold significant potential for streamlining content creation through intuitive user interactions. While recent advances leverage 3D Gaussian Splatting (3DGS) for high-fidelity and real-time rendering, existing methods are often specialized and task-focused, lacking a unified framework for both generation and editing. In this paper, we introduce SplatFlow, a comprehensive framework that addresses this gap by enabling direct 3DGS generation and editing. SplatFlow comprises two main components: a multi-view rectified flow (RF) model and a Gaussian Splatting Decoder (GSDecoder). The multi-view RF model operates in latent space, generating multi-view images, depths, and camera poses simultaneously, conditioned on text prompts—thus addressing challenges like diverse scene scales and complex camera trajectories in real-world settings. Then, the" + } + ] + } + ], + "index": 40 + }, + { + "bbox": [ + 313, + 469, + 556, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 469, + 556, + 613 + ], + "spans": [ + { + "bbox": [ + 313, + 469, + 556, + 613 + ], + "type": "text", + "content": "GSDecoder efficiently translates these latent outputs into 3DGS representations through a feed-forward 3DGS method. Leveraging training-free inversion and inpainting techniques, SplatFlow enables seamless 3DGS editing and supports a broad range of 3D tasks—including object editing, novel view synthesis, and camera pose estimation—within a unified framework without requiring additional complex pipelines. We validate SplatFlow's capabilities on the MVImgNet and DL3DV-7K datasets, demonstrating its versatility and effectiveness in various 3D generation, editing, and inpainting-based tasks. Our project page is available at https://gohyojun15.github.io/SplatFlow/." + } + ] + } + ], + "index": 41 + }, + { + "bbox": [ + 314, + 634, + 395, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 634, + 395, + 647 + ], + "spans": [ + { + "bbox": [ + 314, + 634, + 395, + 647 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 42 + }, + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "text", + "content": "The demand for realistic 3D scene generation and editing from text has surged in applications like VR/AR, gaming, and robotics. Early works primarily utilized NeRF [53, 55] for scene generation [100] and editing [2, 29, 30, 33, 63, 96], but NeRF's volumetric rendering is computationally" + } + ] + } + ], + "index": 43 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 694, + 132, + 704 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 694, + 132, + 704 + ], + "spans": [ + { + "bbox": [ + 66, + 694, + 132, + 704 + ], + "type": "text", + "content": "*Equal contribution" + } + ] + } + ], + "index": 44 + }, + { + "bbox": [ + 67, + 704, + 143, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 704, + 143, + 714 + ], + "spans": [ + { + "bbox": [ + 67, + 704, + 143, + 714 + ], + "type": "text", + "content": "† Corresponding author" + } + ] + } + ], + "index": 45 + }, + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21524" + } + ] + } + ], + "index": 47 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": "intensive, resulting in slower rendering speeds. Recently, 3D Gaussian Splitting (3DGS) [32] has emerged as a promising alternative, offering real-time and high-fidelity rendering implementations. Building on this, recent advances in 3D generation [8, 34, 58, 81, 87, 88, 90, 92, 99, 102, 108] and editing [8, 34, 58, 87, 88, 90, 108] increasingly leverage 3DGS to achieve both speed and quality." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 158, + 297, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 158, + 297, + 326 + ], + "spans": [ + { + "bbox": [ + 55, + 158, + 297, + 326 + ], + "type": "text", + "content": "Despite these advancements, current methods for generating and editing 3DGS are specialized and task-focused, lacking a unified framework. For 3DGS generation, several works [10, 42, 48, 82, 84, 94] leverage 2D diffusion models [26, 69, 76, 79] with Score Distillation Sampling (SDS) [66], which requires time-intensive per-scene optimization. To address this, recent studies have shifted towards direct 3DGS generation, combining diffusion models with reconstruction models [81, 92, 102] or utilizing 3D generative models [99]. However, most of these methods are restricted to object-level 3D generation using synthetic datasets [15, 16] with controlled, bounded environments. In contrast, real-world scenes, varying in scene scales and camera trajectories, present distinctive challenges." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 327, + 296, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 327, + 296, + 460 + ], + "spans": [ + { + "bbox": [ + 55, + 327, + 296, + 460 + ], + "type": "text", + "content": "In 3DGS editing, several works [8, 34, 58, 87, 88, 90, 108] adopt 2D diffusion models [26, 69, 76, 79] to guide the editing process. The main challenge is lifting 2D editing guidance to 3D while maintaining 3D-consistent modifications. One approach focuses on SDS techniques to update 3DGS [58, 91, 108], but these methods require additional stages like texture adjustments [58, 91] and fine-grained refinements [108]. Another approach aims to edit multiview images with 3D consistency [34, 88, 90], introducing attention-based modules [88, 90] and autoregressive editing [34], adding complexity to the editing pipeline." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 460, + 296, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 460, + 296, + 556 + ], + "spans": [ + { + "bbox": [ + 55, + 460, + 296, + 556 + ], + "type": "text", + "content": "In this paper, we bridge the gap between 3DGS generation and editing by designing a direct 3DGS generative model. Inspired by the success of 2D diffusion models that enable training-free editing through inversion [13, 52, 54], we argue that a direct 3DGS model can similarly enable training-free editing. This approach integrates editing without requiring complex pipelines to lift 2D priors, making generation and editing adaptable within a unified framework." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 558, + 297, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 297, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 297, + 715 + ], + "type": "text", + "content": "We introduce SplatFlow, which consists of two main components: (1) a multi-view rectified flow (RF) model and (2) a Gaussian Splitting Decoder (GSDecoder). Similar to Latent Diffusion Models [69], our multi-view RF model operates in the latent space and is trained to simultaneously generate multi-view images, depths, and camera poses conditioned on text prompts. This joint modeling enables SplatFlow to effectively handle the challenges of real-world scene generation, such as various scene scales and camera trajectories, unlike synthetic object datasets [15, 16]. The GSDecoder translates the latent outputs into 3DGS, based on efficient feed-forward methods [4, 9]. To ensure compatibility with other generative models, particularly Stable Diffusion 3 [19]," + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "type": "text", + "content": "we incorporate a fixed pre-trained encoder, allowing flexible cross-model usage during sampling." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 97, + 556, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 97, + 556, + 205 + ], + "spans": [ + { + "bbox": [ + 313, + 97, + 556, + 205 + ], + "type": "text", + "content": "Finally, SplatFlow utilizes training-free inversion [13, 52, 54] and inpainting techniques [11, 51, 70, 71, 77] for 3DGS editing and various 3D tasks. For example, it enables object editing [7, 58], novel view synthesis [61, 68, 97], and camera pose estimation [101] by jointly modeling images, depths, and poses to infer missing elements, highlighting its versatility. We validated its capabilities using the MVImgNet [98] and DL3DV-7K [45] datasets, assessing its performance in 3D generation, editing, and inpainting-based tasks." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 215, + 406, + 228 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 215, + 406, + 228 + ], + "spans": [ + { + "bbox": [ + 313, + 215, + 406, + 228 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 235, + 512, + 247 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 235, + 512, + 247 + ], + "spans": [ + { + "bbox": [ + 313, + 235, + 512, + 247 + ], + "type": "text", + "content": "2.1. Diffusion Models and Rectified Flows" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 252, + 556, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 252, + 556, + 443 + ], + "spans": [ + { + "bbox": [ + 313, + 252, + 556, + 443 + ], + "type": "text", + "content": "Diffusion Models Diffusion models [26, 76, 79] generate data through iterative denoising and have become standard in generative modeling across images [17, 21, 37, 60, 62], videos [27], and text [40]. Trained on large-scale text-to-image datasets, they capture complex semantic relationships, producing high-quality images [6, 56, 65, 69, 72]. An advantage is their adaptability to various tasks in a plug-and-play manner [20, 22, 78], especially inpainting [51, 70] and editing [23, 35]. In painting can be framed as a linear inverse problem [14], for which effective posterior sampling methods have been proposed [11, 51, 70, 71, 77]. For editing, inversion methods [13, 52, 54] produce structured noise by inverting images, allowing new prompts to be effectively processed. Building on these training-free techniques, we extend them to our 3DGS model, adapting RePaint [51] and SDEdit [52] within rectified flow models." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 449, + 557, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 449, + 557, + 594 + ], + "spans": [ + { + "bbox": [ + 313, + 449, + 557, + 594 + ], + "type": "text", + "content": "Rectified Flow Models Rectified Flow (RF) models [1, 46, 47] represent an alternative approach to traditional diffusion models by utilizing a straight-line path from data to noise instead of the typical curved forward process, aiming to simplify sampling and reduce computational costs. This linear path theoretically enables faster sampling by minimizing error accumulation across steps, and recent advances, such as Stable Diffusion 3 (SD3) [19], have successfully scaled RF in text-to-image generation, demonstrating performance that surpasses traditional diffusion models. Our SplatFlow framework builds on the capabilities of SD3, integrating RF to achieve efficient 3DGS generation." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 600, + 441, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 600, + 441, + 612 + ], + "spans": [ + { + "bbox": [ + 313, + 600, + 441, + 612 + ], + "type": "text", + "content": "2.2. 3D Generative Models" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 618, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 618, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 618, + 556, + 715 + ], + "type": "text", + "content": "Neural Scene Representation Various neural scene representations have been proposed for 3D generation tasks. Early works use explicit representations, enabling synthesis of 3D point clouds [57] and shapes [83, 106], but struggle with realistic scene generation due to their rendering mechanisms. To address this, NeRF [53] is employed for a realistic 3D generation with per-scene optimization [66], but its inefficient rendering is unsuitable for a fast generation." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "21525" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "text", + "content": "Recently, 3DGS [32] enables real-time rendering, accelerating optimization-based 3D generation [10, 44, 82]. Building on this progress, we leverage 3DGS's efficient rendering for fast training of a generative model on large datasets." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 123, + 297, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 123, + 297, + 339 + ], + "spans": [ + { + "bbox": [ + 56, + 123, + 297, + 339 + ], + "type": "text", + "content": "Lifting Multi-View Generation Advances in image generation models [6, 65] and large-scale 3D object datasets [15, 16] make multi-view generation compelling compared to optimization-based generation [82]. The key idea is to learn 3D priors as conditions [49, 50, 89] or attentions [74, 85] from 3D datasets while leveraging text-to-image diffusion models [69] trained on billions of text-image pairs [73]. This enables the synthesis of multi-view consistent images from an image or text prompt, which is then fed into reconstruction models [5, 81] to directly generate neural scene representations like 3DGS. One critical success factor in these methods is that the training dataset provides rendered images from consistent, canonical viewpoints, allowing instant access to desired views regardless of the 3D object. In this work, we focus on more challenging unbounded 3D real-world scene generation, where each scene varies in scale and camera trajectory. To handle diverse trajectories, our model learns a joint distribution of camera poses and 3D scenes." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 342, + 297, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 342, + 297, + 521 + ], + "spans": [ + { + "bbox": [ + 55, + 342, + 297, + 521 + ], + "type": "text", + "content": "Text-to-3D Scene Generation Generating 3D scenes from text prompts is challenging due to the unbounded scale of scenes, complicating scale determination. One straightforward approach is to create a scene covered by a single image [75, 107] generated from text-to-image diffusion models [6, 65]. Another line of work [12, 28] generates wider scenes using user-provided camera trajectories, but these methods are time-consuming and produce subpar quality due to reliance on 2D generative models. Recently, Director3D [41], the work most similar to ours, generates camera poses from text and then creates multi-view images decoded into 3DGS. In contrast, SplatFlow jointly learns the distribution of camera poses and multi-view images from text, enhancing generation quality and enabling flexible 3D scene editing in a unified framework." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 528, + 130, + 541 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 528, + 130, + 541 + ], + "spans": [ + { + "bbox": [ + 55, + 528, + 130, + 541 + ], + "type": "text", + "content": "2.3. 3D Editing" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 545, + 296, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 296, + 629 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 296, + 629 + ], + "type": "text", + "content": "In recent years, NeRF-based methods [2, 29, 30, 33, 59, 63, 67, 95, 96] have gained popularity for 3D editing, surpassing point clouds and meshes [86, 103, 105]. However, due to the inefficient rendering, these methods are time-consuming and lack precise control. Conversely, incorporating 3DGS into 3D editing [8, 87] improves effectiveness, speed, and control." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 630, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 297, + 714 + ], + "type": "text", + "content": "For guiding 3D editing, several methods have leveraged 2D diffusion models [26, 69, 76, 79] for structured 3D editing, but it often requires additional refinement stages. For instance, GSEdit [58] and Xiao et al. [91] apply texture enhancement, while TIP-Editor [108] adds fine-grained adjustments to reduce SDS artifacts. Another line of research in 3D editing emphasizes achieving multi-view consistency" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 156 + ], + "type": "text", + "content": "by sequentially updating each view. For example, Gauss-Ctrl [90] and 3DEdit [88] utilize attention-based modules to propagate edits made from one perspective across multiple views, ensuring that changes are consistently reflected throughout the 3D scene. Similarly, 3D-Ego [34] adopts an autoregressive editing framework, updating each view to enhance multi-view consistency." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 157, + 555, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 157, + 555, + 205 + ], + "spans": [ + { + "bbox": [ + 313, + 157, + 555, + 205 + ], + "type": "text", + "content": "Unlike prior methods that require complex procedures to lift 2D priors for 3D consistency, SplatFlow directly models multi-view consistency within the latent space, efficiently decoding edits through our Gaussian Splitting Decoder." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 213, + 477, + 227 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 213, + 477, + 227 + ], + "spans": [ + { + "bbox": [ + 313, + 213, + 477, + 227 + ], + "type": "text", + "content": "3. Preliminary: Rectified Flows" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 233, + 555, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 233, + 555, + 281 + ], + "spans": [ + { + "bbox": [ + 313, + 233, + 555, + 281 + ], + "type": "text", + "content": "Rectified Flows (RFs) [1, 46, 47] consider generative models that construct a noise distribution " + }, + { + "bbox": [ + 313, + 233, + 555, + 281 + ], + "type": "inline_equation", + "content": "q_{0}" + }, + { + "bbox": [ + 313, + 233, + 555, + 281 + ], + "type": "text", + "content": " and a time-varying vector field " + }, + { + "bbox": [ + 313, + 233, + 555, + 281 + ], + "type": "inline_equation", + "content": "v_{t}(\\boldsymbol{x}_{t})" + }, + { + "bbox": [ + 313, + 233, + 555, + 281 + ], + "type": "text", + "content": " to sample data " + }, + { + "bbox": [ + 313, + 233, + 555, + 281 + ], + "type": "inline_equation", + "content": "x_{0}" + }, + { + "bbox": [ + 313, + 233, + 555, + 281 + ], + "type": "text", + "content": " from a data distribution " + }, + { + "bbox": [ + 313, + 233, + 555, + 281 + ], + "type": "inline_equation", + "content": "p_{0}" + }, + { + "bbox": [ + 313, + 233, + 555, + 281 + ], + "type": "text", + "content": " via an ordinary differential equation (ODE) as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 348, + 287, + 555, + 300 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 348, + 287, + 555, + 300 + ], + "spans": [ + { + "bbox": [ + 348, + 287, + 555, + 300 + ], + "type": "interline_equation", + "content": "\\mathrm {d} X _ {t} = v \\left(X _ {t}\\right) \\mathrm {d} t, \\quad X _ {0} \\sim q _ {0}, \\quad t \\in [ 0, 1 ], \\tag {1}", + "image_path": "7f1dc55f8469c3185e9451b662f16a14ec8097581f8d4ae435baa4489915081a.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 306, + 555, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 306, + 555, + 342 + ], + "spans": [ + { + "bbox": [ + 313, + 306, + 555, + 342 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 306, + 555, + 342 + ], + "type": "inline_equation", + "content": "q_{0}" + }, + { + "bbox": [ + 313, + 306, + 555, + 342 + ], + "type": "text", + "content": " typically follows a Gaussian distribution " + }, + { + "bbox": [ + 313, + 306, + 555, + 342 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(0,I)" + }, + { + "bbox": [ + 313, + 306, + 555, + 342 + ], + "type": "text", + "content": ". RF employs a neural network with parameters " + }, + { + "bbox": [ + 313, + 306, + 555, + 342 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 306, + 555, + 342 + ], + "type": "text", + "content": " for modeling the vector field as " + }, + { + "bbox": [ + 313, + 306, + 555, + 342 + ], + "type": "inline_equation", + "content": "v_{t}(X_{t}) = -u_{\\theta}(X_{t},1 - t)" + }, + { + "bbox": [ + 313, + 306, + 555, + 342 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 342, + 556, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 342, + 556, + 474 + ], + "spans": [ + { + "bbox": [ + 313, + 342, + 556, + 474 + ], + "type": "text", + "content": "To train a neural network as the vector field, RFs utilize paired samples from " + }, + { + "bbox": [ + 313, + 342, + 556, + 474 + ], + "type": "inline_equation", + "content": "p_0" + }, + { + "bbox": [ + 313, + 342, + 556, + 474 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 342, + 556, + 474 + ], + "type": "inline_equation", + "content": "q_0" + }, + { + "bbox": [ + 313, + 342, + 556, + 474 + ], + "type": "text", + "content": " (here simplified as " + }, + { + "bbox": [ + 313, + 342, + 556, + 474 + ], + "type": "inline_equation", + "content": "p_1" + }, + { + "bbox": [ + 313, + 342, + 556, + 474 + ], + "type": "text", + "content": "), along a straight line path " + }, + { + "bbox": [ + 313, + 342, + 556, + 474 + ], + "type": "inline_equation", + "content": "Y_t = tY_1 + (1 - t)Y_0" + }, + { + "bbox": [ + 313, + 342, + 556, + 474 + ], + "type": "text", + "content": ". This setup produces the marginal distribution of " + }, + { + "bbox": [ + 313, + 342, + 556, + 474 + ], + "type": "inline_equation", + "content": "Y_t" + }, + { + "bbox": [ + 313, + 342, + 556, + 474 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 313, + 342, + 556, + 474 + ], + "type": "inline_equation", + "content": "p_t(\\pmb{y}_t) = \\int p_t(\\pmb{y}_t|\\pmb{y}_1)p_1(\\pmb{y}_1)\\mathrm{d}\\pmb{y}_1" + }, + { + "bbox": [ + 313, + 342, + 556, + 474 + ], + "type": "text", + "content": ". Starting from an initial " + }, + { + "bbox": [ + 313, + 342, + 556, + 474 + ], + "type": "inline_equation", + "content": "Y_0 = \\pmb{y}_0" + }, + { + "bbox": [ + 313, + 342, + 556, + 474 + ], + "type": "text", + "content": " and moving to a final state " + }, + { + "bbox": [ + 313, + 342, + 556, + 474 + ], + "type": "inline_equation", + "content": "Y_1 = \\pmb{y}_1" + }, + { + "bbox": [ + 313, + 342, + 556, + 474 + ], + "type": "text", + "content": ", a straight line path induces an ODE expressed as " + }, + { + "bbox": [ + 313, + 342, + 556, + 474 + ], + "type": "inline_equation", + "content": "\\mathrm{d}Y_t = u_t(Y_t|Y_1)\\mathrm{d}t" + }, + { + "bbox": [ + 313, + 342, + 556, + 474 + ], + "type": "text", + "content": " where the vector field " + }, + { + "bbox": [ + 313, + 342, + 556, + 474 + ], + "type": "inline_equation", + "content": "u_t(Y_t|Y_1)" + }, + { + "bbox": [ + 313, + 342, + 556, + 474 + ], + "type": "text", + "content": " represents the conditional change from " + }, + { + "bbox": [ + 313, + 342, + 556, + 474 + ], + "type": "inline_equation", + "content": "\\pmb{y}_0" + }, + { + "bbox": [ + 313, + 342, + 556, + 474 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 342, + 556, + 474 + ], + "type": "inline_equation", + "content": "\\pmb{y}_1" + }, + { + "bbox": [ + 313, + 342, + 556, + 474 + ], + "type": "text", + "content": ", given by " + }, + { + "bbox": [ + 313, + 342, + 556, + 474 + ], + "type": "inline_equation", + "content": "u_t(Y_t|Y_1) = \\pmb{y}_1 - \\pmb{y}_0" + }, + { + "bbox": [ + 313, + 342, + 556, + 474 + ], + "type": "text", + "content": ". The marginal vector field " + }, + { + "bbox": [ + 313, + 342, + 556, + 474 + ], + "type": "inline_equation", + "content": "u_t(\\pmb{y}_t)" + }, + { + "bbox": [ + 313, + 342, + 556, + 474 + ], + "type": "text", + "content": " is computed by averaging over the conditional vector field as [46]:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 343, + 479, + 555, + 506 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 343, + 479, + 555, + 506 + ], + "spans": [ + { + "bbox": [ + 343, + 479, + 555, + 506 + ], + "type": "interline_equation", + "content": "u _ {t} (\\boldsymbol {y} _ {t}) = \\int u _ {t} (\\boldsymbol {y} _ {t} | \\boldsymbol {y} _ {1}) \\frac {p _ {t} (\\boldsymbol {y} _ {t} | \\boldsymbol {y} _ {1})}{p _ {t} (\\boldsymbol {y} _ {t})} p _ {1} (\\boldsymbol {y} _ {1}) \\mathrm {d} \\boldsymbol {y} _ {1}. \\tag {2}", + "image_path": "393e2b497b21d050da604463972f2ea4133e3963632aac1605c829d9dd56e9b9.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 509, + 555, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 509, + 555, + 533 + ], + "spans": [ + { + "bbox": [ + 313, + 509, + 555, + 533 + ], + "type": "text", + "content": "To approximate " + }, + { + "bbox": [ + 313, + 509, + 555, + 533 + ], + "type": "inline_equation", + "content": "u_{t}(\\boldsymbol{y}_{t})" + }, + { + "bbox": [ + 313, + 509, + 555, + 533 + ], + "type": "text", + "content": ", the neural network is trained with a flow-matching object, defined by:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 353, + 539, + 555, + 559 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 539, + 555, + 559 + ], + "spans": [ + { + "bbox": [ + 353, + 539, + 555, + 559 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {F M}} := \\mathbb {E} _ {t, Y _ {t}} \\left[ \\left| \\left| u _ {t} \\left(Y _ {t}\\right) - u _ {\\theta} \\left(Y _ {t}, t\\right) \\right| \\right| _ {2} ^ {2} \\right]. \\tag {3}", + "image_path": "2c210a868306104980c63f8dd028646422ebebd5e094e81d1dc66fc3f62a961d.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 563, + 555, + 600 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 563, + 555, + 600 + ], + "spans": [ + { + "bbox": [ + 313, + 563, + 555, + 600 + ], + "type": "text", + "content": "Since directly computing this objective can be computationally challenging, an alternative approach, conditional flow matching, is used to simplify training:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 328, + 605, + 555, + 625 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 605, + 555, + 625 + ], + "spans": [ + { + "bbox": [ + 328, + 605, + 555, + 625 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {C F M}} := \\mathbb {E} _ {t, Y _ {t}, Y _ {1} \\sim p _ {1}} \\left[ \\| u _ {t} \\left(Y _ {t} \\mid Y _ {1}\\right) - u _ {\\theta} \\left(Y _ {t}, t\\right) \\| _ {2} ^ {2} \\right]. \\tag {4}", + "image_path": "c88aca2ef844ecbcfa819f86e02fc642fb604b01dc123a87f959db30058d0b3d.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 629, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 556, + 713 + ], + "type": "text", + "content": "The gradients of " + }, + { + "bbox": [ + 313, + 629, + 556, + 713 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{CFM}}" + }, + { + "bbox": [ + 313, + 629, + 556, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 629, + 556, + 713 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{FM}}" + }, + { + "bbox": [ + 313, + 629, + 556, + 713 + ], + "type": "text", + "content": " are theoretically identical [46], making them equivalent in training. However, " + }, + { + "bbox": [ + 313, + 629, + 556, + 713 + ], + "type": "inline_equation", + "content": "L_{\\mathrm{CFM}}" + }, + { + "bbox": [ + 313, + 629, + 556, + 713 + ], + "type": "text", + "content": " offers better computational efficiency and is therefore preferred. Ultimately, the learned vector field for the ODE in Eq. 1 is given by " + }, + { + "bbox": [ + 313, + 629, + 556, + 713 + ], + "type": "inline_equation", + "content": "v_{t}(X_{t}) = -u_{\\theta}(X_{t},1 - t)" + }, + { + "bbox": [ + 313, + 629, + 556, + 713 + ], + "type": "text", + "content": ", allowing RFs to generate samples that follow the data distribution using the learned vector field." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21526" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 69, + 70, + 539, + 247 + ], + "blocks": [ + { + "bbox": [ + 69, + 70, + 539, + 247 + ], + "lines": [ + { + "bbox": [ + 69, + 70, + 539, + 247 + ], + "spans": [ + { + "bbox": [ + 69, + 70, + 539, + 247 + ], + "type": "image", + "image_path": "55dfe3ec3802dd8d637c1ff9e161605a4829c36665aaa8234007875857183f75.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 247, + 555, + 282 + ], + "lines": [ + { + "bbox": [ + 54, + 247, + 555, + 282 + ], + "spans": [ + { + "bbox": [ + 54, + 247, + 555, + 282 + ], + "type": "text", + "content": "Figure 2. Overview of SplatFlow. SplatFlow consists of two main components: a multi-view Rectified Flow (Section 4.3) model and a Gaussian Splat Decoder (Section 4.2). Conditioned on text, RF model generates multi-view latents—including image, depth, and Plücker ray coordinates. After an optimization process to estimate camera poses, the GSDecoder decodes these latents into pixel-aligned 3DGS." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 289, + 138, + 304 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 289, + 138, + 304 + ], + "spans": [ + { + "bbox": [ + 55, + 289, + 138, + 304 + ], + "type": "text", + "content": "4. Methodology" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 309, + 185, + 322 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 309, + 185, + 322 + ], + "spans": [ + { + "bbox": [ + 55, + 309, + 185, + 322 + ], + "type": "text", + "content": "4.1. Overview of SplatFlow" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 327, + 296, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 327, + 296, + 447 + ], + "spans": [ + { + "bbox": [ + 55, + 327, + 296, + 447 + ], + "type": "text", + "content": "We introduce SplatFlow, a 3DGS generation model designed to perform various 3D editing and 3D-related tasks in a training-free manner, despite being trained solely for generation. As shown in Fig. 2, our model consists of two main components: 1) a multi-view Rectified Flow (RF) model and 2) a Gaussian Splitting Decoder (GSDecoder). First, the multi-view RF model generates multi-view camera poses, depths, and images conditioned on text prompts. Then, the GSDecoder translates these latent representations into pixel-aligned 3DGS, constructing a real-world scene structure." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 447, + 296, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 447, + 296, + 518 + ], + "spans": [ + { + "bbox": [ + 55, + 447, + 296, + 518 + ], + "type": "text", + "content": "Also, both the multi-view RF model and GSDecoder operate within a latent space, adopting a Latent Diffusion Model (LDM) [69]. By freezing and sharing the encoder of Stable Diffusion 3 (SD3) [19], our design enables the RF model to share the latent space with SD3, allowing us to leverage its knowledge for enhanced generation and editing capabilities." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 519, + 296, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 519, + 296, + 567 + ], + "spans": [ + { + "bbox": [ + 55, + 519, + 296, + 567 + ], + "type": "text", + "content": "Finally, our model leverages a training-free approach [51, 52], enabling a broad range of editing applications, such as object replacement, as well as 3D-related tasks including camera pose estimation and novel view synthesis." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 567, + 296, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 567, + 296, + 651 + ], + "spans": [ + { + "bbox": [ + 55, + 567, + 296, + 651 + ], + "type": "text", + "content": "In the following sections, we present each component in detail. Section 4.2 introduces the GSDecoder, explaining details in constructing pixel-aligned 3DGS. Section 4.3 describes the multi-view RF model, detailing how it jointly generates camera poses, depths, and images. Finally, Section 4.4 outlines the sampling process and demonstrates how our model applies training-free techniques." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 659, + 272, + 673 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 659, + 272, + 673 + ], + "spans": [ + { + "bbox": [ + 55, + 659, + 272, + 673 + ], + "type": "text", + "content": "4.2. Gaussian Splitting Decoder (GSDecoder)" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "Recently, feed-forward 3DGS methods [4, 9, 80, 81] have enabled fast 3DGS reconstruction from sparse views by training on large datasets, achieving much faster reconstruction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 290, + 555, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 290, + 555, + 374 + ], + "spans": [ + { + "bbox": [ + 313, + 290, + 555, + 374 + ], + "type": "text", + "content": "than per-scene optimization methods [32]. Leveraging this advantage, our GSDecoder " + }, + { + "bbox": [ + 313, + 290, + 555, + 374 + ], + "type": "inline_equation", + "content": "G_{\\phi}" + }, + { + "bbox": [ + 313, + 290, + 555, + 374 + ], + "type": "text", + "content": " (parameterized by " + }, + { + "bbox": [ + 313, + 290, + 555, + 374 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 313, + 290, + 555, + 374 + ], + "type": "text", + "content": ") is designed to decode pixel-aligned 3DGS from latent representations of " + }, + { + "bbox": [ + 313, + 290, + 555, + 374 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 290, + 555, + 374 + ], + "type": "text", + "content": " sparse views, given their corresponding camera poses " + }, + { + "bbox": [ + 313, + 290, + 555, + 374 + ], + "type": "inline_equation", + "content": "\\mathcal{P} = \\{\\pmb{P}_i\\}_{i=1}^K" + }, + { + "bbox": [ + 313, + 290, + 555, + 374 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 313, + 290, + 555, + 374 + ], + "type": "inline_equation", + "content": "\\pmb{P}_i = \\pmb{K}_i[\\pmb{R}_i|\\pmb{T}_i]" + }, + { + "bbox": [ + 313, + 290, + 555, + 374 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 290, + 555, + 374 + ], + "type": "inline_equation", + "content": "\\pmb{K}_i, \\pmb{R}_i" + }, + { + "bbox": [ + 313, + 290, + 555, + 374 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 290, + 555, + 374 + ], + "type": "inline_equation", + "content": "\\pmb{T}_i" + }, + { + "bbox": [ + 313, + 290, + 555, + 374 + ], + "type": "text", + "content": " denote the intrinsic matrix, rotation matrix, and translation vector of the " + }, + { + "bbox": [ + 313, + 290, + 555, + 374 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 290, + 555, + 374 + ], + "type": "text", + "content": "-th view, respectively." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 375, + 556, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 375, + 556, + 531 + ], + "spans": [ + { + "bbox": [ + 313, + 375, + 556, + 531 + ], + "type": "text", + "content": "A straightforward approach is to design " + }, + { + "bbox": [ + 313, + 375, + 556, + 531 + ], + "type": "inline_equation", + "content": "G_{\\phi}" + }, + { + "bbox": [ + 313, + 375, + 556, + 531 + ], + "type": "text", + "content": " to directly output the 3D Gaussian parameters from the image latents obtained through the encoder " + }, + { + "bbox": [ + 313, + 375, + 556, + 531 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 313, + 375, + 556, + 531 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 313, + 375, + 556, + 531 + ], + "type": "inline_equation", + "content": "G_{\\phi}(\\{\\langle \\mathcal{E}(I_i), P_i \\rangle \\}_{i=1}^K) = \\{(\\pmb{\\mu}_j, \\alpha_j, \\pmb{\\Sigma}_i, c_j)\\}_{j=1}^{H \\times W \\times K}" + }, + { + "bbox": [ + 313, + 375, + 556, + 531 + ], + "type": "text", + "content": ", where each 3D Gaussian parameter includes position " + }, + { + "bbox": [ + 313, + 375, + 556, + 531 + ], + "type": "inline_equation", + "content": "\\pmb{\\mu}_j" + }, + { + "bbox": [ + 313, + 375, + 556, + 531 + ], + "type": "text", + "content": ", opacity " + }, + { + "bbox": [ + 313, + 375, + 556, + 531 + ], + "type": "inline_equation", + "content": "\\alpha_j" + }, + { + "bbox": [ + 313, + 375, + 556, + 531 + ], + "type": "text", + "content": ", covariance " + }, + { + "bbox": [ + 313, + 375, + 556, + 531 + ], + "type": "inline_equation", + "content": "\\pmb{\\Sigma}_j" + }, + { + "bbox": [ + 313, + 375, + 556, + 531 + ], + "type": "text", + "content": ", and color " + }, + { + "bbox": [ + 313, + 375, + 556, + 531 + ], + "type": "inline_equation", + "content": "c_j" + }, + { + "bbox": [ + 313, + 375, + 556, + 531 + ], + "type": "text", + "content": " in a pixel-aligned manner with each " + }, + { + "bbox": [ + 313, + 375, + 556, + 531 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 375, + 556, + 531 + ], + "type": "text", + "content": "-th image " + }, + { + "bbox": [ + 313, + 375, + 556, + 531 + ], + "type": "inline_equation", + "content": "I_i \\in \\mathbb{R}^{H \\times W}" + }, + { + "bbox": [ + 313, + 375, + 556, + 531 + ], + "type": "text", + "content": ". However, as the image latents are produced through a shared and frozen encoder (to leverage compatibility with 2D generative models during sampling), it might lose fine-grained spatial details, leading to abstract representations. This can limit the preservation of specific 3D structural information. To address these limitations, we propose two improvements to enhance GSDecoder's performance." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 538, + 556, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 538, + 556, + 647 + ], + "spans": [ + { + "bbox": [ + 313, + 538, + 556, + 647 + ], + "type": "text", + "content": "Depth latent integration As demonstrated in [31], the encoder " + }, + { + "bbox": [ + 313, + 538, + 556, + 647 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 313, + 538, + 556, + 647 + ], + "type": "text", + "content": " can effectively encode depth maps. Building on this, we incorporate the depth latents " + }, + { + "bbox": [ + 313, + 538, + 556, + 647 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{E}(D_i)\\}_{i=1}^K" + }, + { + "bbox": [ + 313, + 538, + 556, + 647 + ], + "type": "text", + "content": " of each depth map " + }, + { + "bbox": [ + 313, + 538, + 556, + 647 + ], + "type": "inline_equation", + "content": "D_i" + }, + { + "bbox": [ + 313, + 538, + 556, + 647 + ], + "type": "text", + "content": " as an additional input to the GSDecoder to further enhance 3D structural information. We use DepthAnythingV2 [93] to extract the depth maps and normalize them to the range [-1, 1], stacking them as three channels similar to RGB for constructing " + }, + { + "bbox": [ + 313, + 538, + 556, + 647 + ], + "type": "inline_equation", + "content": "D_i" + }, + { + "bbox": [ + 313, + 538, + 556, + 647 + ], + "type": "text", + "content": ". We observe that this improves the convergence speed and performance of the GSDecoder." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 653, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 556, + 713 + ], + "type": "text", + "content": "Adversarial loss Applying adversarial losses has demonstrated improvements in the visual quality of decoded output [18, 69]. However, using adversarial loss during training can cause instability, particularly in the early stages. We observe that applying the vision-aided loss [36] after the GSDe" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21527" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "text", + "content": "coder has reached a certain level of convergence significantly improves visual quality without destabilizing training." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 99, + 296, + 256 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 99, + 296, + 256 + ], + "spans": [ + { + "bbox": [ + 55, + 99, + 296, + 256 + ], + "type": "text", + "content": "Training In summary, our GSDecoder predicts 3DGS " + }, + { + "bbox": [ + 55, + 99, + 296, + 256 + ], + "type": "inline_equation", + "content": "\\{(\\pmb{\\mu}_j,\\alpha_j,\\pmb{\\Sigma}_j,\\pmb{c}_j)\\}_{j = 1}^{H\\times W\\times K}" + }, + { + "bbox": [ + 55, + 99, + 296, + 256 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 55, + 99, + 296, + 256 + ], + "type": "inline_equation", + "content": "\\{(I_i,D_i,P_i)\\}_{i = 1}^K" + }, + { + "bbox": [ + 55, + 99, + 296, + 256 + ], + "type": "text", + "content": ". The model is trained with a combined loss function consisting of LPIPS [104], Mean Squared Error loss, and vision-aided loss [36], computed between the target images of novel views and rendered views from 3DGS. Specifically, the vision-aided loss is activated only after a certain threshold iteration to improve stability and perceptual quality. Additionally, the architecture is initialized with the parameters of the Stable Diffusion 3 decoder, incorporating cross-view attention mechanisms, and the channel dimensions are increased to meet the specific requirements of 3D Gaussian Splatting. Further details on the GSDecoder are provided in Appendix A." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 262, + 234, + 274 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 262, + 234, + 274 + ], + "spans": [ + { + "bbox": [ + 55, + 262, + 234, + 274 + ], + "type": "text", + "content": "4.3. Multi-View Rectified Flow Model" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 279, + 295, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 279, + 295, + 447 + ], + "spans": [ + { + "bbox": [ + 54, + 279, + 295, + 447 + ], + "type": "text", + "content": "With the GSDecoder in place, the 3D Gaussian Splitting (3DGS) generation process becomes feasible if we can produce consistent multi-view latents of images " + }, + { + "bbox": [ + 54, + 279, + 295, + 447 + ], + "type": "inline_equation", + "content": "\\mathcal{E}(I_i)" + }, + { + "bbox": [ + 54, + 279, + 295, + 447 + ], + "type": "text", + "content": ", depth maps " + }, + { + "bbox": [ + 54, + 279, + 295, + 447 + ], + "type": "inline_equation", + "content": "\\mathcal{E}(D_i)" + }, + { + "bbox": [ + 54, + 279, + 295, + 447 + ], + "type": "text", + "content": ", and camera poses " + }, + { + "bbox": [ + 54, + 279, + 295, + 447 + ], + "type": "inline_equation", + "content": "P_{i}" + }, + { + "bbox": [ + 54, + 279, + 295, + 447 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 54, + 279, + 295, + 447 + ], + "type": "inline_equation", + "content": "i = 1,\\dots ,K" + }, + { + "bbox": [ + 54, + 279, + 295, + 447 + ], + "type": "text", + "content": ". Our multi-view rectified flow (RF) model is trained to sample from the joint distribution of these latents, conditioned on text " + }, + { + "bbox": [ + 54, + 279, + 295, + 447 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 54, + 279, + 295, + 447 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 54, + 279, + 295, + 447 + ], + "type": "inline_equation", + "content": "p(\\{I_i\\}_{i = 1}^K,\\{D_i\\}_{i = 1}^K,\\{P_i\\}_{i = 1}^K |C)" + }, + { + "bbox": [ + 54, + 279, + 295, + 447 + ], + "type": "text", + "content": ", allowing for simultaneous generation of multi-view images, depths, and poses. The rationale for modeling a joint distribution instead of each component separately is twofold. First, this approach allows the model to handle various tasks by reformulating them as inpainting problems, as discussed in Section 4.4. Second, joint modeling is crucial for real-world scenes, which require adaptive camera poses tailored to each scene [41]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 448, + 296, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 448, + 296, + 567 + ], + "spans": [ + { + "bbox": [ + 55, + 448, + 296, + 567 + ], + "type": "text", + "content": "To achieve this, our multi-view RF model treats the concatenation of multi-view image latents, depth latents, and Plücker ray coordinates [64, 101] along the channel dimension as input data. Formally, each ray " + }, + { + "bbox": [ + 55, + 448, + 296, + 567 + ], + "type": "inline_equation", + "content": "\\pmb{r}_i" + }, + { + "bbox": [ + 55, + 448, + 296, + 567 + ], + "type": "text", + "content": " is represented as " + }, + { + "bbox": [ + 55, + 448, + 296, + 567 + ], + "type": "inline_equation", + "content": "\\langle \\pmb{d}_i, \\pmb{m}_i \\rangle" + }, + { + "bbox": [ + 55, + 448, + 296, + 567 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 448, + 296, + 567 + ], + "type": "inline_equation", + "content": "\\pmb{d}_i = \\pmb{R}_i^\\top \\pmb{K}_i^{-1} \\pmb{w}_i" + }, + { + "bbox": [ + 55, + 448, + 296, + 567 + ], + "type": "text", + "content": " denotes the direction vector and " + }, + { + "bbox": [ + 55, + 448, + 296, + 567 + ], + "type": "inline_equation", + "content": "\\pmb{m}_i = (-\\pmb{R}_i^\\top \\pmb{T}_i) \\times \\pmb{d}_i" + }, + { + "bbox": [ + 55, + 448, + 296, + 567 + ], + "type": "text", + "content": " represents the moment vector, with " + }, + { + "bbox": [ + 55, + 448, + 296, + 567 + ], + "type": "inline_equation", + "content": "\\pmb{w}_i" + }, + { + "bbox": [ + 55, + 448, + 296, + 567 + ], + "type": "text", + "content": " referring to the 2D pixel coordinates. By matching the spatial resolution of " + }, + { + "bbox": [ + 55, + 448, + 296, + 567 + ], + "type": "inline_equation", + "content": "\\pmb{w}_i" + }, + { + "bbox": [ + 55, + 448, + 296, + 567 + ], + "type": "text", + "content": " with the image and depth latents, each represented as " + }, + { + "bbox": [ + 55, + 448, + 296, + 567 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{n \\times h \\times w}" + }, + { + "bbox": [ + 55, + 448, + 296, + 567 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 448, + 296, + 567 + ], + "type": "inline_equation", + "content": "\\pmb{r}_i" + }, + { + "bbox": [ + 55, + 448, + 296, + 567 + ], + "type": "text", + "content": " can be expressed as " + }, + { + "bbox": [ + 55, + 448, + 296, + 567 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{6 \\times h \\times w}" + }, + { + "bbox": [ + 55, + 448, + 296, + 567 + ], + "type": "text", + "content": ", enabling concatenation along the channel axis." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 567, + 295, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 567, + 295, + 663 + ], + "spans": [ + { + "bbox": [ + 55, + 567, + 295, + 663 + ], + "type": "text", + "content": "Through this setup, we define " + }, + { + "bbox": [ + 55, + 567, + 295, + 663 + ], + "type": "inline_equation", + "content": "Y_{0}" + }, + { + "bbox": [ + 55, + 567, + 295, + 663 + ], + "type": "text", + "content": " as the multi-view latents " + }, + { + "bbox": [ + 55, + 567, + 295, + 663 + ], + "type": "inline_equation", + "content": "\\mathbf{X}" + }, + { + "bbox": [ + 55, + 567, + 295, + 663 + ], + "type": "text", + "content": ", where each element " + }, + { + "bbox": [ + 55, + 567, + 295, + 663 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_i = \\langle \\mathcal{E}(\\mathbf{I}_i),\\mathcal{E}(\\mathbf{D}_i),\\mathbf{r}_i\\rangle" + }, + { + "bbox": [ + 55, + 567, + 295, + 663 + ], + "type": "text", + "content": " results in " + }, + { + "bbox": [ + 55, + 567, + 295, + 663 + ], + "type": "inline_equation", + "content": "Y_{0} = (X_{1},X_{2},\\ldots ,X_{K})\\in \\mathbb{R}^{K\\times (2n + 6)\\times h\\times w}" + }, + { + "bbox": [ + 55, + 567, + 295, + 663 + ], + "type": "text", + "content": ". We then train the multi-view RF model " + }, + { + "bbox": [ + 55, + 567, + 295, + 663 + ], + "type": "inline_equation", + "content": "u_{\\theta}" + }, + { + "bbox": [ + 55, + 567, + 295, + 663 + ], + "type": "text", + "content": " with the conditional flow matching loss defined in Eq. 4. Specifically, we fine-tune Stable Diffusion 3 [19] by adjusting the channel dimensions in the input and output layers and incorporating cross-view attention. Further details are provided in Appendix A." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "Sampling process To recover the camera pose from ray " + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\boldsymbol{r}_i" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": ", we follow the procedure in [101] and further refine the pose parameters to ensure that multiple views share the same intrinsic. However, we find that naively solving the ODE" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "type": "text", + "content": "can lead to degraded camera pose accuracy. We hypothesize this degradation results from deviations from the camera pose manifold, which introduces errors in camera trajectory estimation. To address this, at each sampling step " + }, + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "type": "inline_equation", + "content": "t = t_k" + }, + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "type": "text", + "content": ", we predict the sampling destination at " + }, + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "type": "inline_equation", + "content": "t = 0" + }, + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "type": "text", + "content": ", recover the camera poses from this prediction, and reconstruct the associated camera rays. These rays are then forwarded to " + }, + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "type": "inline_equation", + "content": "t = t_k" + }, + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "type": "text", + "content": " to maintain the solution within the valid ray manifold at " + }, + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "type": "inline_equation", + "content": "t = t_k" + }, + { + "bbox": [ + 313, + 72, + 555, + 168 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 171, + 556, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 171, + 556, + 232 + ], + "spans": [ + { + "bbox": [ + 313, + 171, + 556, + 232 + ], + "type": "text", + "content": "Stable Diffusion 3 guidance To improve multi-view image generation quality, we integrate vector fields from Stable Diffusion 3, known for robust generalization, with our multiview RF model to solve the ODE for randomly selected view's image latents at specific sampling steps." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 239, + 471, + 251 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 239, + 471, + 251 + ], + "spans": [ + { + "bbox": [ + 313, + 239, + 471, + 251 + ], + "type": "text", + "content": "4.4. Inference Given Partial Data" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 256, + 555, + 341 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 256, + 555, + 341 + ], + "spans": [ + { + "bbox": [ + 313, + 256, + 555, + 341 + ], + "type": "text", + "content": "In this section, we demonstrate how SplatFlow can be effectively utilized for 3D editing and applications by applying training-free inversion techniques [13, 52, 54] and inpainting methods [11, 51, 70, 71, 77]. We modify SDEdit [52] inversion for rectified flow-based inversion and adopt the RePaint [51] to support rectified flow inpainting. Additional technical details are provided in Appendix A." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 343, + 556, + 476 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 343, + 556, + 476 + ], + "spans": [ + { + "bbox": [ + 313, + 343, + 556, + 476 + ], + "type": "text", + "content": "3DGS editing Our GSDecoder is designed to directly decode 3DGS, enabling efficient editing by modifying only the multi-view latents. Additionally, with the RF model's joint modeling of multi-view image latents, edits can be achieved without extra modules (e.g., cross-view attention mechanisms [88, 90]), streamlining the process through simple inversion and sampling. Specifically, given a target text and input multi-view latents, we apply a modified SDEdit for inversion at a chosen timestep " + }, + { + "bbox": [ + 313, + 343, + 556, + 476 + ], + "type": "inline_equation", + "content": "t_k" + }, + { + "bbox": [ + 313, + 343, + 556, + 476 + ], + "type": "text", + "content": ". Conditioned on the target text, we then generate the edited multi-view latents by solving the ODE in Eq. 1 following our sampling procedure." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 479, + 556, + 587 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 479, + 556, + 587 + ], + "spans": [ + { + "bbox": [ + 313, + 479, + 556, + 587 + ], + "type": "text", + "content": "Inpainting application Our model is also applicable to various 3D tasks, as it jointly models multi-view images, depths, and camera poses. This joint modeling allows known parts of the data (e.g., some views or camera parameters) to act as constraints, enabling the prediction of unknown parts through inpainting. In this work, we focus on two specific tasks: 1) camera pose estimation from multi-view images and depths, and 2) novel view synthesis using a subset of the " + }, + { + "bbox": [ + 313, + 479, + 556, + 587 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 479, + 556, + 587 + ], + "type": "text", + "content": " multi-view latents and camera poses for novel viewpoints." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 597, + 440, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 597, + 440, + 611 + ], + "spans": [ + { + "bbox": [ + 313, + 597, + 440, + 611 + ], + "type": "text", + "content": "5. Experimental Results" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "text", + "content": "In this section, we demonstrate the capabilities of our proposed SplatFlow across 3DGS generation, 3DGS editing, and various 3D tasks. We begin by detailing the implementation of SplatFlow in Section 5.1. Following this, we present comparative results on 3DGS generation in real-world datasets in Section 5.2. In Section 5.3, we evaluate the 3DGS editing performance of SplatFlow against other methods. Finally, Section 5.4 showcases the versatility of our" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "21528" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 68, + 550, + 522 + ], + "blocks": [ + { + "bbox": [ + 61, + 68, + 550, + 522 + ], + "lines": [ + { + "bbox": [ + 61, + 68, + 550, + 522 + ], + "spans": [ + { + "bbox": [ + 61, + 68, + 550, + 522 + ], + "type": "image", + "image_path": "42ac89b4832e303bdded132ff403ebf887638a38b7f79b795357cc9b61a81e8b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 526, + 555, + 561 + ], + "lines": [ + { + "bbox": [ + 55, + 526, + 555, + 561 + ], + "spans": [ + { + "bbox": [ + 55, + 526, + 555, + 561 + ], + "type": "text", + "content": "Figure 3. Qualitative results in text-to-3DGS generation on MVImgNet and DL3DV validation sets. The first two rows are rendered scenes from the MVImgNet dataset, while the last two rows are from the DL3DV dataset. Our SplatFlow produces cohesive and realistic scenes with sharp details, accurately capturing the intricacies of real-world environments and accommodating diverse camera trajectories." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 566, + 298, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 566, + 298, + 616 + ], + "spans": [ + { + "bbox": [ + 55, + 566, + 298, + 616 + ], + "type": "text", + "content": "multi-view RF model in handling diverse 3D tasks through training-free inpainting techniques. Extensive information on all our experiments is provided in Appendix B with additional results that could not be included here in the Appendix." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 624, + 178, + 636 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 624, + 178, + 636 + ], + "spans": [ + { + "bbox": [ + 55, + 624, + 178, + 636 + ], + "type": "text", + "content": "5.1. Experimental Setups" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 642, + 297, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 297, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 297, + 715 + ], + "type": "text", + "content": "Due to space constraints, we provide a concise overview of our experimental setups here. We used the MVImgNet [98] and DL3DV-7K [45] datasets, both containing real-world scenes with images and corresponding camera poses. For evaluation, we utilized 1.25K scenes from MVImgNet for each task and 300 sequences from DL3DV-7K, using the" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 567, + 556, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 567, + 556, + 616 + ], + "spans": [ + { + "bbox": [ + 313, + 567, + 556, + 616 + ], + "type": "text", + "content": "remaining for training. The multi-view RF model and GSDecoder were trained with " + }, + { + "bbox": [ + 313, + 567, + 556, + 616 + ], + "type": "inline_equation", + "content": "K = 8" + }, + { + "bbox": [ + 313, + 567, + 556, + 616 + ], + "type": "text", + "content": "-view setup. Text descriptions for scenes were extracted using the Llava-One Vision Qwen 7B [38]. We used SDS++ [41] for further improvements." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 624, + 509, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 624, + 509, + 635 + ], + "spans": [ + { + "bbox": [ + 313, + 624, + 509, + 635 + ], + "type": "text", + "content": "5.2. Results on Text-to-3DGS Generation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "text", + "content": "We evaluated our model against various baselines on the DL3DV and MVImgNet validation sets for real-world scene text-to-3DGS generation. Our evaluation protocol utilized the rendered image of the generated 3DGS. We used the FID score [25] to evaluate image quality and the CLIP score [24] to measure alignment with text prompts." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "21529" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 70, + 294, + 139 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 294, + 139 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 294, + 139 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 294, + 139 + ], + "type": "table", + "html": "
MethodMVImgNet [98]DL3DV [45]
FID-10K↓CLIPScore↑FID-2.4K↓CLIPScore↑
Director3D [41]39.5530.4888.4430.04
Director3D (w/ SDS++) [41]41.8031.0095.8831.68
SplatFlow34.8531.4379.9130.06
SplatFlow (w/ SDS++)35.4632.3085.3131.90
", + "image_path": "b849d636d8a49b6f5bcde5e6958e6858095cbf417a553e950f4e9957852fb868.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 186, + 296, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 186, + 296, + 281 + ], + "spans": [ + { + "bbox": [ + 55, + 186, + 296, + 281 + ], + "type": "text", + "content": "For quantitative comparison, we adopt Director3D [41], which leverages the full sequences of MVImgNet, DL3DV-10K, and LAION [73] datasets. In contrast, our model was trained on a notably smaller dataset, utilizing only a portion of MVImgNet and DL3DV-7K, without access to the full DL3DV-10K or LAION data. For qualitative results, we also compared against other scene generation methods, including LucidDreamer [12], DreamScene [39], and Director3D [41]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 288, + 296, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 288, + 296, + 504 + ], + "spans": [ + { + "bbox": [ + 55, + 288, + 296, + 504 + ], + "type": "text", + "content": "Quantitative results Table 1 presents the quantitative comparison of SplatFlow and the baseline methods on both MVImgNet and DL3DV datasets. SplatFlow consistently outperforms Director3D [41] in terms of FID and CLIP score, demonstrating its effectiveness in generating high-quality images and maintaining strong alignment with text prompts, even with a smaller training dataset. Specifically, on MVImgNet, SplatFlow achieves a significantly lower FID (34.85 vs. 39.55 for Director3D), indicating better image quality, and a higher CLIP score (31.43 vs. 30.48), showing improved text-to-image alignment. Similarly, on DL3DV, SplatFlow achieves an FID of 79.91, outperforming Director3D's 88.44, with a comparable CLIP score. Notably, with the addition of " + }, + { + "bbox": [ + 55, + 288, + 296, + 504 + ], + "type": "inline_equation", + "content": "\\mathrm{SDS}++" + }, + { + "bbox": [ + 55, + 288, + 296, + 504 + ], + "type": "text", + "content": " for refinement, both SplatFlow and Director3D show increased CLIP scores, further enhancing text alignment, though at a minor trade-off in FID. This highlights the effectiveness of " + }, + { + "bbox": [ + 55, + 288, + 296, + 504 + ], + "type": "inline_equation", + "content": "\\mathrm{SDS}++" + }, + { + "bbox": [ + 55, + 288, + 296, + 504 + ], + "type": "text", + "content": " for boosting alignment with text prompts across different 3DGS generation models." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 510, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 297, + 714 + ], + "type": "text", + "content": "Qualitative results Figure 3 shows the qualitative comparisons between our method and baselines on DL3DV and MVImgNet validation sets. Overall, SplatFlow demonstrates clear advantages over the baseline methods. First, DreamScene [39], which relies on Score Distillation Sampling (SDS), often struggles to generate realistic real-world scenes, resulting in outputs that lack natural textures and environmental details. Second, LucidDreamer [12], which is based on single-view inpainting, faces challenges when handling scenes with large camera trajectory variations, leading to inconsistent or incomplete views. Third, Director3D [41] tends to produce slightly blurred outputs, lacking the sharpness needed for intricate details. In contrast, our SplatFlow can generate coherent and realistic scenes with clear details, effectively capturing the complexities of real-world environments and various camera trajectories. Additional qualitative results of SplatFlow can be found in Appendix C." + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 316, + 76, + 346, + 118 + ], + "blocks": [ + { + "bbox": [ + 316, + 76, + 346, + 118 + ], + "lines": [ + { + "bbox": [ + 316, + 76, + 346, + 118 + ], + "spans": [ + { + "bbox": [ + 316, + 76, + 346, + 118 + ], + "type": "image", + "image_path": "86838318c008a85c46e9f3f96907c998be65ac0cf1a0c5c8b518c2da0e4624f0.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 353, + 76, + 414, + 118 + ], + "blocks": [ + { + "bbox": [ + 353, + 76, + 414, + 118 + ], + "lines": [ + { + "bbox": [ + 353, + 76, + 414, + 118 + ], + "spans": [ + { + "bbox": [ + 353, + 76, + 414, + 118 + ], + "type": "image", + "image_path": "272d339f885460e47bbda6a32fdf95328d7eb5fb647912a24ba172449f99bd78.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 369, + 120, + 518, + 129 + ], + "lines": [ + { + "bbox": [ + 369, + 120, + 518, + 129 + ], + "spans": [ + { + "bbox": [ + 369, + 120, + 518, + 129 + ], + "type": "text", + "content": "\"green inflatable ring\" -> \"colorful red rubber ball\"" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 422, + 76, + 484, + 118 + ], + "blocks": [ + { + "bbox": [ + 422, + 76, + 484, + 118 + ], + "lines": [ + { + "bbox": [ + 422, + 76, + 484, + 118 + ], + "spans": [ + { + "bbox": [ + 422, + 76, + 484, + 118 + ], + "type": "image", + "image_path": "82751cfae236038e957eab30d907c57afb1fed744ea0777599412dcc7c535ef3.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 492, + 76, + 553, + 118 + ], + "blocks": [ + { + "bbox": [ + 492, + 76, + 553, + 118 + ], + "lines": [ + { + "bbox": [ + 492, + 76, + 553, + 118 + ], + "spans": [ + { + "bbox": [ + 492, + 76, + 553, + 118 + ], + "type": "image", + "image_path": "e02dd886852cc769955fbeddd251757b7a7f068bd889bf9b172ca157513b4e70.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 316, + 135, + 346, + 167 + ], + "blocks": [ + { + "bbox": [ + 316, + 135, + 346, + 167 + ], + "lines": [ + { + "bbox": [ + 316, + 135, + 346, + 167 + ], + "spans": [ + { + "bbox": [ + 316, + 135, + 346, + 167 + ], + "type": "image", + "image_path": "5c1d203bd60c12d13f5695c95edfa6ff5f6aad279626ea24f5182e5696a256f8.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 182, + 553, + 205 + ], + "lines": [ + { + "bbox": [ + 313, + 182, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 313, + 182, + 553, + 205 + ], + "type": "text", + "content": "Figure 4. Qualitative results in 3D editing with MVInpainter [3] and DGE [7]. We show rendered scenes except for MVInpainter." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 353, + 135, + 414, + 167 + ], + "blocks": [ + { + "bbox": [ + 353, + 135, + 414, + 167 + ], + "lines": [ + { + "bbox": [ + 353, + 135, + 414, + 167 + ], + "spans": [ + { + "bbox": [ + 353, + 135, + 414, + 167 + ], + "type": "image", + "image_path": "b114b2b479f68a9e32df423ebdd098815a890778831b32afbd835dd157b7e334.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 368, + 169, + 527, + 179 + ], + "lines": [ + { + "bbox": [ + 368, + 169, + 527, + 179 + ], + "spans": [ + { + "bbox": [ + 368, + 169, + 527, + 179 + ], + "type": "text", + "content": "\"collection of red chili peppers\" -> \"cluster of bananas\"" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 422, + 135, + 484, + 167 + ], + "blocks": [ + { + "bbox": [ + 422, + 135, + 484, + 167 + ], + "lines": [ + { + "bbox": [ + 422, + 135, + 484, + 167 + ], + "spans": [ + { + "bbox": [ + 422, + 135, + 484, + 167 + ], + "type": "image", + "image_path": "2bf6cfd51c36531d7f5b42e4dddb964d0fb707e7642b4cd7c6034ac139ad2f58.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 492, + 135, + 553, + 167 + ], + "blocks": [ + { + "bbox": [ + 492, + 135, + 553, + 167 + ], + "lines": [ + { + "bbox": [ + 492, + 135, + 553, + 167 + ], + "spans": [ + { + "bbox": [ + 492, + 135, + 553, + 167 + ], + "type": "image", + "image_path": "c1ce82817da1b479833e2b63d43141c79b95d1ba35e1eea62ab6d56fb3cb0aca.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 218, + 449, + 231 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 218, + 449, + 231 + ], + "spans": [ + { + "bbox": [ + 313, + 218, + 449, + 231 + ], + "type": "text", + "content": "5.3. Result on 3DGS Editing" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 236, + 427, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 236, + 427, + 307 + ], + "spans": [ + { + "bbox": [ + 313, + 236, + 427, + 307 + ], + "type": "text", + "content": "We demonstrate the effectiveness of SplatFlow in the 3DGS editing task by creating a benchmark with 100 sampled scenes from the MVImgNet validation set." + } + ] + } + ], + "index": 17 + }, + { + "type": "table", + "bbox": [ + 433, + 237, + 552, + 287 + ], + "blocks": [ + { + "bbox": [ + 55, + 142, + 295, + 175 + ], + "lines": [ + { + "bbox": [ + 55, + 142, + 295, + 175 + ], + "spans": [ + { + "bbox": [ + 55, + 142, + 295, + 175 + ], + "type": "text", + "content": "Table 1. Quantitative results in text-to-3DGS generation on the MVImgNet and DL3DV datasets. We compared our SplatFlow with and without the SDS++ [41], against Director3D [41]." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 433, + 237, + 552, + 287 + ], + "lines": [ + { + "bbox": [ + 433, + 237, + 552, + 287 + ], + "spans": [ + { + "bbox": [ + 433, + 237, + 552, + 287 + ], + "type": "table", + "html": "
MethodCLIPScore↑CLIP D-sim↑
DGE [7]27.430.102
SplatFlow28.470.169
+) SDS++31.300.224
", + "image_path": "eb2e4e10b9ca2342d765ea50a3f1742a0eca084d3ae937aa02678807d6670ef9.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "table_body" + } + ], + "index": 18 + }, + { + "bbox": [ + 433, + 291, + 553, + 302 + ], + "lines": [ + { + "bbox": [ + 433, + 291, + 553, + 302 + ], + "spans": [ + { + "bbox": [ + 433, + 291, + 553, + 302 + ], + "type": "text", + "content": "Table 2. 3D object replacement." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 307, + 555, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 307, + 555, + 439 + ], + "spans": [ + { + "bbox": [ + 313, + 307, + 555, + 439 + ], + "type": "text", + "content": "Images and texts were provided to GPT-4 to identify main objects and generate text captions for edited images, which were then used for 3D object replacement. For editing, we used the same mask in [3] for masking out the object, then inpainted this area based on the provided editing captions. We compared this pipeline with an optimization-based 3DGS editing method, DGE [7], and conducted qualitative comparisons with MVInpainter [3], which propagates an edited image from pretrained generative models across multiple views. Following standard protocols [8, 29], we measured the CLIP score and directional similarity (CLIP D-sim)." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 442, + 556, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 442, + 556, + 502 + ], + "spans": [ + { + "bbox": [ + 313, + 442, + 556, + 502 + ], + "type": "text", + "content": "Quantitative results As shown in Table 2, SplatFlow achieves higher CLIPScore and CLIP D-sim than DGE [7], indicating more accurate and effective editing. SDS++ refinement further boosts performance, demonstrating enhanced alignment with target edits." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 506, + 556, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 506, + 556, + 601 + ], + "spans": [ + { + "bbox": [ + 313, + 506, + 556, + 601 + ], + "type": "text", + "content": "Qualitative results Figure 4 demonstrates that SplatFlow performs comparably to MVInpainter, a specialized 2D inpainting method that leverages an already edited high-quality image. On the other hand, DGE struggles to replace objects fully, often just altering styles instead of achieving complete transformations. This underscores SplatFlow's superior ability in precise 3D object replacement. Please refer to Appendix C for further qualitative results." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 609, + 491, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 609, + 491, + 621 + ], + "spans": [ + { + "bbox": [ + 313, + 609, + 491, + 621 + ], + "type": "text", + "content": "5.4. Result on Inpainting Application" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 626, + 554, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 626, + 554, + 662 + ], + "spans": [ + { + "bbox": [ + 313, + 626, + 554, + 662 + ], + "type": "text", + "content": "Here, we present the adaptability of our multi-view RF model to camera pose estimation and novel-view synthesis via the inpainting technique on the MVImgNet dataset [98]." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 666, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 666, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 666, + 556, + 714 + ], + "type": "text", + "content": "Camera pose estimation To demonstrate the effectiveness of our multi-view RF model in camera pose estimation, we primarily compare our approach to RayDiffusion [101] and RelPose++ [43]. We conducted comparisons using an" + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21530" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 70, + 293, + 158 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 293, + 158 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 293, + 158 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 293, + 158 + ], + "type": "table", + "html": "
MethodRotation↑Camera Center↑
@5@10@15@0.05@0.1@0.2
RelPose++ [43]19.437.751.40.612.555.0
Ray Regression [101]10.425.650.115.347.982.9
Ray Diffusion [101]17.538.759.624.160.987.6
SplatFlow (w/ depth)26.852.659.362.391.699.4
SplatFlow (w/o depth)28.854.563.964.994.099.7
", + "image_path": "6bed44d30484d48b349ec29d57bdabfd5c3916b38a4cab1e0b987a8c8f13ddbd.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 63, + 201, + 206, + 361 + ], + "blocks": [ + { + "bbox": [ + 63, + 201, + 206, + 361 + ], + "lines": [ + { + "bbox": [ + 63, + 201, + 206, + 361 + ], + "spans": [ + { + "bbox": [ + 63, + 201, + 206, + 361 + ], + "type": "image", + "image_path": "b57d6b733081eee53271290b47a7c88ef5fc3eb25e1114becb7e0f343a4d9ffd.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 364, + 295, + 398 + ], + "lines": [ + { + "bbox": [ + 55, + 364, + 295, + 398 + ], + "spans": [ + { + "bbox": [ + 55, + 364, + 295, + 398 + ], + "type": "text", + "content": "Figure 5. Qualitative results for camera pose estimation. Camera poses are estimated from multi-view images. Image border colors match each camera, with black cameras indicating GT poses." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 217, + 204, + 286, + 361 + ], + "blocks": [ + { + "bbox": [ + 217, + 204, + 286, + 361 + ], + "lines": [ + { + "bbox": [ + 217, + 204, + 286, + 361 + ], + "spans": [ + { + "bbox": [ + 217, + 204, + 286, + 361 + ], + "type": "image", + "image_path": "e3e88792debe0a2d7dd9bfdf15666e4d8a8b5a66aef70d55d1659f4a22fde5f1.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 407, + 296, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 407, + 296, + 479 + ], + "spans": [ + { + "bbox": [ + 55, + 407, + 296, + 479 + ], + "type": "text", + "content": "8-view setup, inpainting the ray latent with and without the depth latent from DepthAnythingV2 [93]. We used rotation accuracy@ " + }, + { + "bbox": [ + 55, + 407, + 296, + 479 + ], + "type": "inline_equation", + "content": "Q_{R}" + }, + { + "bbox": [ + 55, + 407, + 296, + 479 + ], + "type": "text", + "content": " and translation accuracy@ " + }, + { + "bbox": [ + 55, + 407, + 296, + 479 + ], + "type": "inline_equation", + "content": "Q_{T}" + }, + { + "bbox": [ + 55, + 407, + 296, + 479 + ], + "type": "text", + "content": " as evaluation metrics, measuring the proportion of relative camera rotations within " + }, + { + "bbox": [ + 55, + 407, + 296, + 479 + ], + "type": "inline_equation", + "content": "Q_{R}" + }, + { + "bbox": [ + 55, + 407, + 296, + 479 + ], + "type": "text", + "content": " degrees of ground truth and camera centers within " + }, + { + "bbox": [ + 55, + 407, + 296, + 479 + ], + "type": "inline_equation", + "content": "Q_{T}" + }, + { + "bbox": [ + 55, + 407, + 296, + 479 + ], + "type": "text", + "content": " units of scene scale, respectively." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 480, + 296, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 480, + 296, + 576 + ], + "spans": [ + { + "bbox": [ + 55, + 480, + 296, + 576 + ], + "type": "text", + "content": "Table 3 and Fig. 5 present the quantitative and qualitative results. Overall, the multi-view RF model achieves superior performance across most scenarios, outperforming the baseline methods. Interestingly, the multi-view RF model achieves higher accuracy without depth latents than with them. We hypothesize that this outcome is due to the joint generation of depth and camera poses, which may yield more detailed and contextually enriched depth estimates." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "type": "text", + "content": "Novel view-synthesis We evaluated the performance of our multi-view RF model in novel view synthesis under two scenarios: 1) interpolation, with the uniformly sampled input views, and 2) extrapolation, with the input views positioned in the center. This comparison validates our multi-view RF model's capability of 3D reasoning based on relative viewpoints and evaluates its effectiveness in generating novel views. We evaluated PSNR, SSIM, and LPIPS [104] against the ground-truth images for reconstruction quality, as well as absolute relative difference and " + }, + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\delta_{1}" + }, + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "type": "text", + "content": " accuracy using depth maps produced by DepthAnythingV2 [93]." + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 317, + 70, + 553, + 171 + ], + "blocks": [ + { + "bbox": [ + 55, + 159, + 296, + 194 + ], + "lines": [ + { + "bbox": [ + 55, + 159, + 296, + 194 + ], + "spans": [ + { + "bbox": [ + 55, + 159, + 296, + 194 + ], + "type": "text", + "content": "Table 3. Results in camera pose estimation on MVImgNet validation set. @Q represents the accuracy threshold for rotations (degrees) and camera centers (units)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 70, + 553, + 171 + ], + "lines": [ + { + "bbox": [ + 317, + 70, + 553, + 171 + ], + "spans": [ + { + "bbox": [ + 317, + 70, + 553, + 171 + ], + "type": "table", + "html": "
TypeRGBDepth
PSNR↑SSIM↑LPIPS↓AbsRel↓δ1↑
Interpolation (N=2)14.730.5710.6480.5880.731
Interpolation (N=4)17.050.5900.5510.4980.761
Interpolation (N=6)18.820.6260.4830.4150.775
Extrapolation (N=2)15.150.5770.6270.7710.715
Extrapolation (N=4)16.800.5950.5540.6790.727
Extrapolation (N=6)17.960.6130.5030.6020.747
", + "image_path": "fbb580335ddfb40b515c1c128e557a87c31be1a8ec6eda39691e4669dd81a243.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 171, + 553, + 205 + ], + "lines": [ + { + "bbox": [ + 314, + 171, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 314, + 171, + 553, + 205 + ], + "type": "text", + "content": "Table 4. Novel view synthesis results on MVImgNet. We use " + }, + { + "bbox": [ + 314, + 171, + 553, + 205 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 314, + 171, + 553, + 205 + ], + "type": "text", + "content": " input views to synthesize " + }, + { + "bbox": [ + 314, + 171, + 553, + 205 + ], + "type": "inline_equation", + "content": "K - N" + }, + { + "bbox": [ + 314, + 171, + 553, + 205 + ], + "type": "text", + "content": " novel views, with uniformly sampled views for interpolation and central views for extrapolation." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 316, + 215, + 553, + 405 + ], + "blocks": [ + { + "bbox": [ + 316, + 215, + 553, + 405 + ], + "lines": [ + { + "bbox": [ + 316, + 215, + 553, + 405 + ], + "spans": [ + { + "bbox": [ + 316, + 215, + 553, + 405 + ], + "type": "image", + "image_path": "26e343442881543e49df324d7c91b6ae61ee85277f90e733dbeef9ad7a21dfcf.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 407, + 555, + 430 + ], + "lines": [ + { + "bbox": [ + 313, + 407, + 555, + 430 + ], + "spans": [ + { + "bbox": [ + 313, + 407, + 555, + 430 + ], + "type": "text", + "content": "Figure 6. Qualitative results for novel view synthesis. Novel view synthesis is performed from the red-box images and depths." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 438, + 555, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 438, + 555, + 522 + ], + "spans": [ + { + "bbox": [ + 313, + 438, + 555, + 522 + ], + "type": "text", + "content": "As shown in Fig. 6, our multi-view RF model successfully generates novel views for both scenarios. Also, Table 4 shows that interpolation outperforms extrapolation in both RGB reconstruction and depth estimation as the number of input views increases, indicating that interpolation leverages 3D reasoning for accurate novel view generation, whereas extrapolation focuses on generating diverse novel views." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 536, + 388, + 548 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 536, + 388, + 548 + ], + "spans": [ + { + "bbox": [ + 313, + 536, + 388, + 548 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 558, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 556, + 713 + ], + "type": "text", + "content": "In this paper, we have introduced SplatFlow, a novel framework that unifies 3D Gaussian Splitting (3DGS) generation and editing within a single, efficient model. By designing a multi-view Rectified Flow (RF) model alongside a Gaussian Splitting Decoder (GSDecoder), we achieved direct 3DGS generation capable of addressing real-world challenges, such as varying scales and complex camera trajectories. Our approach utilizes joint modeling of multi-view images, depths, and camera poses, enabling seamless integration of training-free inversion and inpainting techniques for a diverse set of 3D tasks. As we present the versatility of our SplatFlow, we believe SplatFlow represents an important step toward a versatile foundational model for 3D content." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21531" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 62, + 91, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 66, + 91, + 296, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 91, + 296, + 123 + ], + "spans": [ + { + "bbox": [ + 66, + 91, + 296, + 123 + ], + "type": "text", + "content": "[1] Michael S Albergo and Eric Vanden-Eijnden. Building normalizing flows with stochastic interpolants. arXiv preprint arXiv:2209.15571, 2022. 2, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 125, + 296, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 125, + 296, + 190 + ], + "spans": [ + { + "bbox": [ + 66, + 125, + 296, + 190 + ], + "type": "text", + "content": "[2] Chong Bao, Yinda Zhang, Bangbang Yang, Tianxing Fan, Zesong Yang, Hujun Bao, Guofeng Zhang, and Zhaopeng Cui. Sine: Semantic-driven image-based nerf editing with prior-guided editing field. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20919-20929, 2023. 1, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 66, + 194, + 296, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 194, + 296, + 236 + ], + "spans": [ + { + "bbox": [ + 66, + 194, + 296, + 236 + ], + "type": "text", + "content": "[3] Chenjie Cao, Chaohui Yu, Yanwei Fu, Fan Wang, and Xi-angyang Xue. Mvinpainter: Learning multi-view consistent inpainting to bridge 2d and 3d editing. arXiv preprint arXiv:2408.08000, 2024.7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 66, + 239, + 295, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 239, + 295, + 293 + ], + "spans": [ + { + "bbox": [ + 66, + 239, + 295, + 293 + ], + "type": "text", + "content": "[4] David Charatan, Sizhe Lester Li, Andrea Tagliasacchi, and Vincent Sitzmann. pixelsplat: 3d gaussian splats from image pairs for scalable generalizable 3d reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19457-19467, 2024. 2, 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 66, + 295, + 294, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 295, + 294, + 338 + ], + "spans": [ + { + "bbox": [ + 66, + 295, + 294, + 338 + ], + "type": "text", + "content": "[5] Anpei Chen, Haofei Xu, Stefano Esposito, Siyu Tang, and Andreas Geiger. Lara: Efficient large-baseline radiance fields. In European Conference on Computer Vision, pages 338-355. Springer, 2025. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 66, + 341, + 295, + 395 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 341, + 295, + 395 + ], + "spans": [ + { + "bbox": [ + 66, + 341, + 295, + 395 + ], + "type": "text", + "content": "[6] Junsong Chen, Jincheng Yu, Chongjian Ge, Lewei Yao, Enze Xie, Yue Wu, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, et al. Pixart-α: Fast training of diffusion transformer for photorealistic text-to-image synthesis. arXiv preprint arXiv:2310.00426, 2023. 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 66, + 397, + 294, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 397, + 294, + 430 + ], + "spans": [ + { + "bbox": [ + 66, + 397, + 294, + 430 + ], + "type": "text", + "content": "[7] Minghao Chen, Iro Laina, and Andrea Vedaldi. Dge: Direct gaussian 3d editing by consistent multi-view editing. arXiv preprint arXiv:2404.18929, 2024. 2, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 66, + 431, + 295, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 431, + 295, + 497 + ], + "spans": [ + { + "bbox": [ + 66, + 431, + 295, + 497 + ], + "type": "text", + "content": "[8] Yiwen Chen, Zilong Chen, Chi Zhang, Feng Wang, Xiaofeng Yang, Yikai Wang, Zhongang Cai, Lei Yang, Huaping Liu, and Guosheng Lin. Gaussianeditor: Swift and controllable 3d editing with gaussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21476-21485, 2024. 2, 3, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 66, + 499, + 295, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 499, + 295, + 553 + ], + "spans": [ + { + "bbox": [ + 66, + 499, + 295, + 553 + ], + "type": "text", + "content": "[9] Yuedong Chen, Haofei Xu, Chuanxia Zheng, Bohan Zhuang, Marc Pollefeys, Andreas Geiger, Tat-Jen Cham, and Jianfei Cai. Mvsplat: Efficient 3d gaussian splattering from sparse multi-view images. In European Conference on Computer Vision, pages 370-386. Springer, 2024. 2, 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 555, + 295, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 555, + 295, + 599 + ], + "spans": [ + { + "bbox": [ + 62, + 555, + 295, + 599 + ], + "type": "text", + "content": "[10] Zilong Chen, Feng Wang, Yikai Wang, and Huaping Liu. Text-to-3d using gaussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21401-21412, 2024. 2, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 601, + 295, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 601, + 295, + 643 + ], + "spans": [ + { + "bbox": [ + 62, + 601, + 295, + 643 + ], + "type": "text", + "content": "[11] Hyungjin Chung, Jeongsol Kim, Michael T McCann, Marc L Klasky, and Jong Chul Ye. Diffusion posterior sampling for general noisy inverse problems. arXiv preprint arXiv:2209.14687, 2022. 2, 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 62, + 646, + 295, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 646, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 62, + 646, + 295, + 689 + ], + "type": "text", + "content": "[12] Jaeyoung Chung, Suyoung Lee, Hyeongjin Nam, Jaerin Lee, and Kyoung Mu Lee. Luciddreamer: Domain-free generation of 3d gaussian splatting scenes. arXiv preprint arXiv:2311.13384, 2023. 3, 6, 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 62, + 691, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 691, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 62, + 691, + 295, + 712 + ], + "type": "text", + "content": "[13] Guillaume Couairon, Jakob Verbeek, Holger Schwenk, and Matthieu Cord. Diffedit: Diffusion-based seman" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 321, + 74, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 339, + 74, + 555, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 74, + 555, + 94 + ], + "spans": [ + { + "bbox": [ + 339, + 74, + 555, + 94 + ], + "type": "text", + "content": "tic image editing with mask guidance. arXiv preprint arXiv:2210.11427, 2022. 2, 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 321, + 96, + 555, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 96, + 555, + 150 + ], + "spans": [ + { + "bbox": [ + 321, + 96, + 555, + 150 + ], + "type": "text", + "content": "[14] Giannis Daras, Hyungjin Chung, Chieh-Hsin Lai, Yuki Mitsufuji, Jong Chul Ye, Peyman Milanfar, Alexandros G Dimakis, and Mauricio Delbracio. A survey on diffusion models for inverse problems. arXiv preprint arXiv:2410.00083, 2024. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 321, + 152, + 555, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 152, + 555, + 218 + ], + "spans": [ + { + "bbox": [ + 321, + 152, + 555, + 218 + ], + "type": "text", + "content": "[15] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3d objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13142-13153, 2023. 2, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 321, + 220, + 555, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 220, + 555, + 274 + ], + "spans": [ + { + "bbox": [ + 321, + 220, + 555, + 274 + ], + "type": "text", + "content": "[16] Matt Deitke, Ruoshi Liu, Matthew Wallingford, Huong Ngo, Oscar Michel, Aditya Kusupati, Alan Fan, Christian Laforte, Vikram Voleti, Samir Yitzhak Gadre, et al. Objverse-xl: A universe of " + }, + { + "bbox": [ + 321, + 220, + 555, + 274 + ], + "type": "inline_equation", + "content": "10\\mathrm{m} + 3\\mathrm{d}" + }, + { + "bbox": [ + 321, + 220, + 555, + 274 + ], + "type": "text", + "content": " objects. Advances in Neural Information Processing Systems, 36, 2024. 2, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 321, + 276, + 555, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 276, + 555, + 308 + ], + "spans": [ + { + "bbox": [ + 321, + 276, + 555, + 308 + ], + "type": "text", + "content": "[17] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 321, + 310, + 555, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 310, + 555, + 354 + ], + "spans": [ + { + "bbox": [ + 321, + 310, + 555, + 354 + ], + "type": "text", + "content": "[18] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12873-12883, 2021. 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 321, + 355, + 554, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 355, + 554, + 419 + ], + "spans": [ + { + "bbox": [ + 321, + 355, + 554, + 419 + ], + "type": "text", + "content": "[19] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In *Forty-first International Conference on Machine Learning*, 2024. 2, 4, 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 321, + 422, + 555, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 422, + 555, + 476 + ], + "spans": [ + { + "bbox": [ + 321, + 422, + 555, + 476 + ], + "type": "text", + "content": "[20] Hyojun Go, Yunsung Lee, Jin-Young Kim, Seunghyun Lee, Myeongho Jeong, Hyun Seung Lee, and Seungtaek Choi. Towards practical plug-and-play diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1962-1971, 2023. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 321, + 478, + 555, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 478, + 555, + 521 + ], + "spans": [ + { + "bbox": [ + 321, + 478, + 555, + 521 + ], + "type": "text", + "content": "[21] Hyojun Go, Yunsung Lee, Seunghyun Lee, Shinhyeok Oh, Hyeongdon Moon, and Seungtaek Choi. Addressing negative transfer in diffusion models. Advances in Neural Information Processing Systems, 36, 2024. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 321, + 523, + 555, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 523, + 555, + 566 + ], + "spans": [ + { + "bbox": [ + 321, + 523, + 555, + 566 + ], + "type": "text", + "content": "[22] Alexandros Graikos, Nikolay Malkin, Nebojsa Jojic, and Dimitris Samaras. Diffusion models as plug-and-play priors. Advances in Neural Information Processing Systems, 35: 14715-14728, 2022. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 321, + 568, + 555, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 568, + 555, + 611 + ], + "spans": [ + { + "bbox": [ + 321, + 568, + 555, + 611 + ], + "type": "text", + "content": "[23] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Prompt-to-prompt image editing with cross attention control. arXiv preprint arXiv:2208.01626, 2022. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 321, + 613, + 555, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 613, + 555, + 656 + ], + "spans": [ + { + "bbox": [ + 321, + 613, + 555, + 656 + ], + "type": "text", + "content": "[24] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning. arXiv preprint arXiv:2104.08718, 2021.6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 321, + 658, + 555, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 658, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 321, + 658, + 555, + 712 + ], + "type": "text", + "content": "[25] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 6" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21532" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 61, + 73, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 73, + 296, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 73, + 296, + 106 + ], + "spans": [ + { + "bbox": [ + 61, + 73, + 296, + 106 + ], + "type": "text", + "content": "[26] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 106, + 296, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 106, + 296, + 161 + ], + "spans": [ + { + "bbox": [ + 61, + 106, + 296, + 161 + ], + "type": "text", + "content": "[27] Jonathan Ho, William Chan, Chitwan Sahara, Jay Whang, Ruiqi Gao, Alexey Gritsenko, Diederik P Kingma, Ben Poole, Mohammad Norouzi, David J Fleet, et al. Imagen video: High definition video generation with diffusion models. arXiv preprint arXiv:2210.02303, 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 162, + 296, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 162, + 296, + 216 + ], + "spans": [ + { + "bbox": [ + 62, + 162, + 296, + 216 + ], + "type": "text", + "content": "[28] Lukas Hollein, Ang Cao, Andrew Owens, Justin Johnson, and Matthias Nießner. Text2room: Extracting textured 3d meshes from 2d text-to-image models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 7909-7920, 2023. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 217, + 296, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 217, + 296, + 259 + ], + "spans": [ + { + "bbox": [ + 62, + 217, + 296, + 259 + ], + "type": "text", + "content": "[29] Hiromichi Kamata, Yuiko Sakuma, Akio Hayakawa, Masato Ishii, and Takuya Narihira. Instruct 3d-to-3d: Text instruction guided 3d-to-3d conversion. arXiv preprint arXiv:2303.15780, 2023. 1, 3, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 261, + 296, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 261, + 296, + 304 + ], + "spans": [ + { + "bbox": [ + 62, + 261, + 296, + 304 + ], + "type": "text", + "content": "[30] Nazmul Karim, Hasan Iqbal, Umar Khalid, Chen Chen, and Jing Hua. Free-editor: zero-shot text-driven 3d scene editing. In European Conference on Computer Vision, pages 436-453. Springer, 2025. 1, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 305, + 296, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 305, + 296, + 370 + ], + "spans": [ + { + "bbox": [ + 62, + 305, + 296, + 370 + ], + "type": "text", + "content": "[31] Bingxin Ke, Anton Obukhov, Shengyu Huang, Nando Metzger, Rodrigo Caye Daudt, and Konrad Schindler. Repurposing diffusion-based image generators for monocular depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9492-9502, 2024. 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 371, + 296, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 371, + 296, + 414 + ], + "spans": [ + { + "bbox": [ + 62, + 371, + 296, + 414 + ], + "type": "text", + "content": "[32] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splattering for real-time radiance field rendering. ACM Trans. Graph., 42(4):139-1, 2023. 2, 3, 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 415, + 296, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 415, + 296, + 448 + ], + "spans": [ + { + "bbox": [ + 62, + 415, + 296, + 448 + ], + "type": "text", + "content": "[33] Umar Khalid, Hasan Iqbal, Nazmul Karim, Jing Hua, and Chen Chen. Latenteditor: Text driven local editing of 3d scenes. arXiv preprint arXiv:2312.09313, 2023. 1, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 449, + 296, + 481 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 449, + 296, + 481 + ], + "spans": [ + { + "bbox": [ + 62, + 449, + 296, + 481 + ], + "type": "text", + "content": "[34] Umar Khalid, Hasan Iqbal, Azib Farooq, Jing Hua, and Chen Chen. 3dego: 3d editing on the go! In European Conference on Computer Vision, pages 73-89. Springer, 2024. 2, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 482, + 296, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 482, + 296, + 536 + ], + "spans": [ + { + "bbox": [ + 62, + 482, + 296, + 536 + ], + "type": "text", + "content": "[35] Gwanghyun Kim, Taesung Kwon, and Jong Chul Ye. Diffusionclip: Text-guided diffusion models for robust image manipulation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2426-2435, 2022. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 537, + 296, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 537, + 296, + 590 + ], + "spans": [ + { + "bbox": [ + 62, + 537, + 296, + 590 + ], + "type": "text", + "content": "[36] Nupur Kumari, Richard Zhang, Eli Shechtman, and JunYan Zhu. Ensembling off-the-shelf models for gan training. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10651-10662, 2022. 4, 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 592, + 296, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 592, + 296, + 645 + ], + "spans": [ + { + "bbox": [ + 62, + 592, + 296, + 645 + ], + "type": "text", + "content": "[37] Yunsung Lee, JinYoung Kim, Hyojun Go, Myeongho Jeong, Shinhyeok Oh, and Seungtaek Choi. Multi-architecture multi-expert diffusion models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 13427-13436, 2024. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 62, + 647, + 296, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 647, + 296, + 690 + ], + "spans": [ + { + "bbox": [ + 62, + 647, + 296, + 690 + ], + "type": "text", + "content": "[38] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 62, + 691, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 691, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 62, + 691, + 296, + 713 + ], + "type": "text", + "content": "[39] Haoran Li, Haolin Shi, Wenli Zhang, Wenjun Wu, Yong Liao, Lin Wang, Lik-hang Lee, and Pengyuan Zhou. Dreamscene:" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 320, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 338, + 73, + 555, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 338, + 73, + 555, + 106 + ], + "spans": [ + { + "bbox": [ + 338, + 73, + 555, + 106 + ], + "type": "text", + "content": "3d gaussian-based text-to-3d scene generation via formation pattern sampling. arXiv preprint arXiv:2404.03575, 2024.6, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 321, + 107, + 555, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 107, + 555, + 151 + ], + "spans": [ + { + "bbox": [ + 321, + 107, + 555, + 151 + ], + "type": "text", + "content": "[40] Xiang Li, John Thickstun, Ishaan Gulrajani, Percy S Liang, and Tatsunori B Hashimoto. Diffusion-lm improves controllable text generation. Advances in Neural Information Processing Systems, 35:4328-4343, 2022. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 320, + 152, + 555, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 152, + 555, + 207 + ], + "spans": [ + { + "bbox": [ + 320, + 152, + 555, + 207 + ], + "type": "text", + "content": "[41] Xinyang Li, Zhangyu Lai, Linning Xu, Yansong Qu, Liujuan Cao, Shengchuan Zhang, Bo Dai, and Rongrong Ji. Director3d: Real-world camera trajectory and 3d scene generation from text. Advances in neural information processing systems, 2024. 3, 5, 6, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 320, + 209, + 555, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 209, + 555, + 252 + ], + "spans": [ + { + "bbox": [ + 320, + 209, + 555, + 252 + ], + "type": "text", + "content": "[42] Zongrui Li, Minghui Hu, Qian Zheng, and Xudong Jiang. Connecting consistency distillation to score distillation for text-to-3d generation. In European Conference on Computer Vision, pages 274–291. Springer, 2024. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 320, + 254, + 555, + 286 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 254, + 555, + 286 + ], + "spans": [ + { + "bbox": [ + 320, + 254, + 555, + 286 + ], + "type": "text", + "content": "[43] Amy Lin, Jason Y Zhang, Deva Ramanan, and Shubham Tulsiani. Relpose++: Recovering 6d poses from sparse-view observations. arXiv preprint arXiv:2305.04926, 2023. 7, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 320, + 288, + 555, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 288, + 555, + 342 + ], + "spans": [ + { + "bbox": [ + 320, + 288, + 555, + 342 + ], + "type": "text", + "content": "[44] Huan Ling, Seung Wook Kim, Antonio Torralba, Sanja Fidler, and Karsten Kreis. Align your gaussians: Text-to-4d with dynamic 3d gaussians and composed diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8576–8588, 2024. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 320, + 344, + 555, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 344, + 555, + 409 + ], + "spans": [ + { + "bbox": [ + 320, + 344, + 555, + 409 + ], + "type": "text", + "content": "[45] Lu Ling, Yichen Sheng, Zhi Tu, Wentian Zhao, Cheng Xin, Kun Wan, Lantao Yu, Qianyu Guo, Zixun Yu, Yawen Lu, et al. D13dv-10k: A large-scale scene dataset for deep learning-based 3d vision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22160–22169, 2024. 2, 6, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 320, + 411, + 555, + 444 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 411, + 555, + 444 + ], + "spans": [ + { + "bbox": [ + 320, + 411, + 555, + 444 + ], + "type": "text", + "content": "[46] Yaron Lipman, Ricky TQ Chen, Heli Ben-Hamu, Maximilian Nickel, and Matt Le. Flow matching for generative modeling. arXiv preprint arXiv:2210.02747, 2022. 2, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 320, + 445, + 555, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 445, + 555, + 487 + ], + "spans": [ + { + "bbox": [ + 320, + 445, + 555, + 487 + ], + "type": "text", + "content": "[47] Xingchao Liu, Chengyue Gong, and Qiang Liu. Flow straight and fast: Learning to generate and transfer data with rectified flow. arXiv preprint arXiv:2209.03003, 2022. 2, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 320, + 490, + 555, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 490, + 555, + 544 + ], + "spans": [ + { + "bbox": [ + 320, + 490, + 555, + 544 + ], + "type": "text", + "content": "[48] Xian Liu, Xiaohang Zhan, Jiaxiang Tang, Ying Shan, Gang Zeng, Dahua Lin, Xihui Liu, and Ziwei Liu. Humangaussian: Text-driven 3d human generation with gaussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6646-6657, 2024. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 320, + 546, + 555, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 546, + 555, + 590 + ], + "spans": [ + { + "bbox": [ + 320, + 546, + 555, + 590 + ], + "type": "text", + "content": "[49] Yuan Liu, Cheng Lin, Zijiao Zeng, Xiaoxiao Long, Lingjie Liu, Taku Komura, and Wenping Wang. Syncdreamer: Generating multiview-consistent images from a single-view image. arXiv preprint arXiv:2309.03453, 2023. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 320, + 591, + 555, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 591, + 555, + 657 + ], + "spans": [ + { + "bbox": [ + 320, + 591, + 555, + 657 + ], + "type": "text", + "content": "[50] Xiaoxiao Long, Yuan-Chen Guo, Cheng Lin, Yuan Liu, Zhiyang Dou, Lingjie Liu, Yuexin Ma, Song-Hai Zhang, Marc Habermann, Christian Theobalt, et al. Wonder3d: Single image to 3d using cross-domain diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9970-9980, 2024. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 320, + 658, + 555, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 658, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 320, + 658, + 555, + 713 + ], + "type": "text", + "content": "[51] Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11461-11471, 2022. 2, 4, 5" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "21533" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 61, + 72, + 297, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 72, + 297, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 72, + 297, + 117 + ], + "spans": [ + { + "bbox": [ + 61, + 72, + 297, + 117 + ], + "type": "text", + "content": "[52] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jiajun Wu, Jun-Yan Zhu, and Stefano Ermon. Sdedit: Guided image synthesis and editing with stochastic differential equations. arXiv preprint arXiv:2108.01073, 2021. 2, 4, 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 119, + 296, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 119, + 296, + 173 + ], + "spans": [ + { + "bbox": [ + 61, + 119, + 296, + 173 + ], + "type": "text", + "content": "[53] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 1, 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 175, + 297, + 231 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 175, + 297, + 231 + ], + "spans": [ + { + "bbox": [ + 62, + 175, + 297, + 231 + ], + "type": "text", + "content": "[54] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6038–6047, 2023. 2, 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 232, + 296, + 276 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 232, + 296, + 276 + ], + "spans": [ + { + "bbox": [ + 62, + 232, + 296, + 276 + ], + "type": "text", + "content": "[55] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM transactions on graphics (TOG), 41(4):1-15, 2022. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 278, + 296, + 333 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 278, + 296, + 333 + ], + "spans": [ + { + "bbox": [ + 62, + 278, + 296, + 333 + ], + "type": "text", + "content": "[56] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 335, + 296, + 378 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 335, + 296, + 378 + ], + "spans": [ + { + "bbox": [ + 62, + 335, + 296, + 378 + ], + "type": "text", + "content": "[57] Alex Nichol, Heewoo Jun, Prafulla Dhariwal, Pamela Mishkin, and Mark Chen. Point-e: A system for generating 3d point clouds from complex prompts. arXiv preprint arXiv:2212.08751, 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 380, + 296, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 380, + 296, + 424 + ], + "spans": [ + { + "bbox": [ + 62, + 380, + 296, + 424 + ], + "type": "text", + "content": "[58] Francesco Palandra, Andrea Sanchietti, Daniele Baieri, and Emanuele Rodola. Gsedit: Efficient text-guided editing of 3d objects via gaussian splatting. arXiv preprint arXiv:2403.05154, 2024. 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 426, + 296, + 471 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 426, + 296, + 471 + ], + "spans": [ + { + "bbox": [ + 62, + 426, + 296, + 471 + ], + "type": "text", + "content": "[59] Byeongjun Park and Changick Kim. Point-dynrf: Point-based dynamic radiance fields from a monocular video. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 3171-3181, 2024. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 473, + 296, + 506 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 473, + 296, + 506 + ], + "spans": [ + { + "bbox": [ + 62, + 473, + 296, + 506 + ], + "type": "text", + "content": "[60] Byeongjun Park, Sangmin Woo, Hyojun Go, Jin-Young Kim, and Changick Kim. Denoising task routing for diffusion models. arXiv preprint arXiv:2310.07138, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 507, + 296, + 551 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 507, + 296, + 551 + ], + "spans": [ + { + "bbox": [ + 62, + 507, + 296, + 551 + ], + "type": "text", + "content": "[61] Byeongjun Park, Hyojun Go, and Changick Kim. Bridging implicit and explicit geometric transformation for single-image view synthesis. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 553, + 296, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 553, + 296, + 597 + ], + "spans": [ + { + "bbox": [ + 62, + 553, + 296, + 597 + ], + "type": "text", + "content": "[62] Byeongjun Park, Hyojun Go, Jin-Young Kim, Sangmin Woo, Seokil Ham, and Changick Kim. Switch diffusion transformer: Synergizing denoising tasks with sparse mixture-of-experts. arXiv preprint arXiv:2403.09176, 2024. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 600, + 296, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 600, + 296, + 632 + ], + "spans": [ + { + "bbox": [ + 62, + 600, + 296, + 632 + ], + "type": "text", + "content": "[63] Jangho Park, Gihyun Kwon, and Jong Chul Ye. Ed-nerf: Efficient text-guided editing of 3d scene using latent space nerf. arXiv preprint arXiv:2310.02712, 2023. 1, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 62, + 635, + 296, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 635, + 296, + 655 + ], + "spans": [ + { + "bbox": [ + 62, + 635, + 296, + 655 + ], + "type": "text", + "content": "[64] Julius Plücker. Analytisch-geometrische Entwicklungen. GD Baedeker, 1828. 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 62, + 658, + 296, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 658, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 62, + 658, + 296, + 712 + ], + "type": "text", + "content": "[65] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 2, 3" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 321, + 73, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 321, + 73, + 555, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 73, + 555, + 106 + ], + "spans": [ + { + "bbox": [ + 321, + 73, + 555, + 106 + ], + "type": "text", + "content": "[66] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 321, + 108, + 555, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 108, + 555, + 151 + ], + "spans": [ + { + "bbox": [ + 321, + 108, + 555, + 151 + ], + "type": "text", + "content": "[67] Elad Richardson, Gal Metzer, Yuval Alaluf, Raja Giryes, and Daniel Cohen-Or. Texture: Text-guided texturing of 3d shapes. In ACM SIGGRAPH 2023 conference proceedings, pages 1–11, 2023. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 321, + 152, + 555, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 152, + 555, + 196 + ], + "spans": [ + { + "bbox": [ + 321, + 152, + 555, + 196 + ], + "type": "text", + "content": "[68] Robin Rombach, Patrick Esser, and Björn Ommer. Geometry-free view synthesis: Transformers and no 3d priors. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 14356-14366, 2021. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 321, + 198, + 555, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 198, + 555, + 252 + ], + "spans": [ + { + "bbox": [ + 321, + 198, + 555, + 252 + ], + "type": "text", + "content": "[69] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2, 3, 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 321, + 254, + 555, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 254, + 555, + 297 + ], + "spans": [ + { + "bbox": [ + 321, + 254, + 555, + 297 + ], + "type": "text", + "content": "[70] Litu Rout, Advait Parulekar, Constantine Caramanis, and Sanjay Shakkottai. A theoretical justification for image inpainting using denoising diffusion probabilistic models. arXiv preprint arXiv:2302.01217, 2023. 2, 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 321, + 299, + 555, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 299, + 555, + 353 + ], + "spans": [ + { + "bbox": [ + 321, + 299, + 555, + 353 + ], + "type": "text", + "content": "[71] Litu Rout, Negin Raoof, Giannis Daras, Constantine Caramanis, Alex Dimakis, and Sanjay Shakkottai. Solving linear inverse problems provably via posterior sampling with latent diffusion models. Advances in Neural Information Processing Systems, 36, 2023. 2, 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 321, + 355, + 555, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 355, + 555, + 397 + ], + "spans": [ + { + "bbox": [ + 321, + 355, + 555, + 397 + ], + "type": "text", + "content": "[72] Axel Sauer, Dominik Lorenz, Andreas Blattmann, and Robin Rombach. Adversarial diffusion distillation. In European Conference on Computer Vision, pages 87-103. Springer, 2024. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 321, + 399, + 555, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 399, + 555, + 474 + ], + "spans": [ + { + "bbox": [ + 321, + 399, + 555, + 474 + ], + "type": "text", + "content": "[73] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278-25294, 2022. 3, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 321, + 478, + 553, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 478, + 553, + 511 + ], + "spans": [ + { + "bbox": [ + 321, + 478, + 553, + 511 + ], + "type": "text", + "content": "[74] Yichun Shi, Peng Wang, Jianglong Ye, Mai Long, Kejie Li, and Xiao Yang. Mvdream: Multi-view diffusion for 3d generation. arXiv preprint arXiv:2308.16512, 2023. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 321, + 513, + 555, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 513, + 555, + 555 + ], + "spans": [ + { + "bbox": [ + 321, + 513, + 555, + 555 + ], + "type": "text", + "content": "[75] Jaidev Shriram, Alex Trevithick, Lingjie Liu, and Ravi Ramamoorthi. Realmdreamer: Text-driven 3d scene generation with inpainting and depth diffusion. arXiv preprint arXiv:2404.07199, 2024. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 321, + 557, + 555, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 557, + 555, + 610 + ], + "spans": [ + { + "bbox": [ + 321, + 557, + 555, + 610 + ], + "type": "text", + "content": "[76] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International conference on machine learning, pages 2256-2265. PMLR, 2015. 2,3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 321, + 613, + 555, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 613, + 555, + 656 + ], + "spans": [ + { + "bbox": [ + 321, + 613, + 555, + 656 + ], + "type": "text", + "content": "[77] Jiaming Song, Arash Vahdat, Morteza Mardani, and Jan Kautz. Pseudoinverse-guided diffusion models for inverse problems. In International Conference on Learning Representations, 2023. 2, 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 321, + 658, + 555, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 658, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 321, + 658, + 555, + 712 + ], + "type": "text", + "content": "[78] Jiaming Song, Qinsheng Zhang, Hongxu Yin, Morteza Mardani, Ming-Yu Liu, Jan Kautz, Yongxin Chen, and Arash Vahdat. Loss-guided diffusion models for plug-and-play controllable generation. In International Conference on Machine Learning, pages 32483-32498. PMLR, 2023. 2" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "21534" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 61, + 73, + 297, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 73, + 297, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 73, + 297, + 116 + ], + "spans": [ + { + "bbox": [ + 61, + 73, + 297, + 116 + ], + "type": "text", + "content": "[79] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 2, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 118, + 296, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 118, + 296, + 171 + ], + "spans": [ + { + "bbox": [ + 61, + 118, + 296, + 171 + ], + "type": "text", + "content": "[80] Stanislaw Szymanowicz, Chrisitian Rupprecht, and Andrea Vedaldi. Splatter image: Ultra-fast single-view 3d reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10208-10217, 2024. 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 173, + 296, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 173, + 296, + 225 + ], + "spans": [ + { + "bbox": [ + 62, + 173, + 296, + 225 + ], + "type": "text", + "content": "[81] Jiaxiang Tang, Zhaoxi Chen, Xiaokang Chen, Tengfei Wang, Gang Zeng, and Ziwei Liu. Lgm: Large multi-view gaussian model for high-resolution 3d content creation. In European Conference on Computer Vision, pages 1-18. Springer, 2024. 2, 3, 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 228, + 295, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 228, + 295, + 270 + ], + "spans": [ + { + "bbox": [ + 62, + 228, + 295, + 270 + ], + "type": "text", + "content": "[82] Jiaxiang Tang, Jiawei Ren, Hang Zhou, Ziwei Liu, and Gang Zeng. Dreamgaussian: Generative gaussian splatting for efficient 3d content creation. In The Twelfth International Conference on Learning Representations, 2024. 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 272, + 296, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 272, + 296, + 316 + ], + "spans": [ + { + "bbox": [ + 62, + 272, + 296, + 316 + ], + "type": "text", + "content": "[83] Arash Vahdat, Francis Williams, Zan Gojcic, Or Litany, Sanja Fidler, Karsten Kreis, et al. Lion: Latent point diffusion models for 3d shape generation. Advances in Neural Information Processing Systems, 35:10021-10039, 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 317, + 295, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 317, + 295, + 348 + ], + "spans": [ + { + "bbox": [ + 62, + 317, + 295, + 348 + ], + "type": "text", + "content": "[84] Alexander Vilesov, Pradyumna Chari, and Achuta Kadambi. Cg3d: Compositional generation for text-to-3d via gaussian splatting. arXiv preprint arXiv:2311.17907, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 350, + 296, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 350, + 296, + 415 + ], + "spans": [ + { + "bbox": [ + 62, + 350, + 296, + 415 + ], + "type": "text", + "content": "[85] Vikram Voleti, Chun-Han Yao, Mark Boss, Adam Letts, David Pankratz, Dmitry Tochilkin, Christian Laforte, Robin Rombach, and Varun Jampani. Sv3d: Novel multi-view synthesis and 3d generation from a single image using latent video diffusion. In European Conference on Computer Vision, pages 439-457. Springer, 2025. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 416, + 296, + 447 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 416, + 296, + 447 + ], + "spans": [ + { + "bbox": [ + 62, + 416, + 296, + 447 + ], + "type": "text", + "content": "[86] Can Wang, Mingming He, Mengei Chai, Dongdong Chen, and Jing Liao. Mesh-guided neural implicit field editing. arXiv preprint arXiv:2312.02157, 2023. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 449, + 296, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 449, + 296, + 502 + ], + "spans": [ + { + "bbox": [ + 62, + 449, + 296, + 502 + ], + "type": "text", + "content": "[87] Junjie Wang, Jiemin Fang, Xiaopeng Zhang, Lingxi Xie, and Qi Tian. Gaussianeditor: Editing 3d gaussians delicately with text instructions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20902-20911, 2024. 2, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 504, + 296, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 504, + 296, + 547 + ], + "spans": [ + { + "bbox": [ + 62, + 504, + 296, + 547 + ], + "type": "text", + "content": "[88] Yuxuan Wang, Xuanyu Yi, Zike Wu, Na Zhao, Long Chen, and Hanwang Zhang. View-consistent 3d editing with gaussian splatting. In European Conference on Computer Vision, pages 404-420. Springer, 2024. 2, 3, 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 548, + 296, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 548, + 296, + 602 + ], + "spans": [ + { + "bbox": [ + 62, + 548, + 296, + 602 + ], + "type": "text", + "content": "[89] Sangmin Woo, Byeongjun Park, Hyojun Go, Jin-Young Kim, and Changick Kim. Harmonyview: Harmonizing consistency and diversity in one-image-to-3d. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10574–10584, 2024. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 604, + 296, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 604, + 296, + 646 + ], + "spans": [ + { + "bbox": [ + 62, + 604, + 296, + 646 + ], + "type": "text", + "content": "[90] Jing Wu, Jia-Wang Bian, Xinghui Li, Guangrun Wang, Ian Reid, Philip Torr, and Victor Adrian Prisacariu. Gaussctrl: multi-view consistent text-driven 3d gaussian splattering editing. arXiv preprint arXiv:2403.08733, 2024. 2, 3, 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 62, + 647, + 296, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 647, + 296, + 690 + ], + "spans": [ + { + "bbox": [ + 62, + 647, + 296, + 690 + ], + "type": "text", + "content": "[91] Hanyuan Xiao, Yingshu Chen, Huajian Huang, Haolin Xiong, Jing Yang, Pratusha Prasad, and Yajie Zhao. Localized gaussian splatting editing with contextual awareness. arXiv preprint arXiv:2408.00083, 2024. 2, 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 62, + 692, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 692, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 62, + 692, + 296, + 713 + ], + "type": "text", + "content": "[92] Yinghao Xu, Hao Tan, Fujun Luan, Sai Bi, Peng Wang, Jiahao Li, Zifan Shi, Kalyan Sunkavalli, Gordon Wetzstein," + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 338, + 73, + 555, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 338, + 73, + 555, + 116 + ], + "spans": [ + { + "bbox": [ + 338, + 73, + 555, + 116 + ], + "type": "text", + "content": "Zexiang Xu, and Kai Zhang. DMV3d: Denoising multiview diffusion using 3d large reconstruction model. In The Twelfth International Conference on Learning Representations, 2024. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 321, + 118, + 555, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 118, + 555, + 150 + ], + "spans": [ + { + "bbox": [ + 321, + 118, + 555, + 150 + ], + "type": "text", + "content": "[93] Lihe Yang, Bingyi Kang, Zilong Huang, Zhen Zhao, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything v2. arXiv preprint arXiv:2406.09414, 2024. 4, 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 321, + 152, + 555, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 152, + 555, + 216 + ], + "spans": [ + { + "bbox": [ + 321, + 152, + 555, + 216 + ], + "type": "text", + "content": "[94] Taoran Yi, Jiemin Fang, Junjie Wang, Guanjun Wu, Lingxi Xie, Xiaopeng Zhang, Wenyu Liu, Qi Tian, and Xinggang Wang. Gaussian dreamer: Fast generation from text to 3d gaussians by bridging 2d and 3d diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6796-6807, 2024. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 321, + 217, + 555, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 217, + 555, + 271 + ], + "spans": [ + { + "bbox": [ + 321, + 217, + 555, + 271 + ], + "type": "text", + "content": "[95] Xuanyu Yi, Zike Wu, Qingshan Xu, Pan Zhou, Joo-Hwee Lim, and Hanwang Zhang. Diffusion time-step curriculum for one image to 3d generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9948-9958, 2024. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 321, + 273, + 554, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 273, + 554, + 304 + ], + "spans": [ + { + "bbox": [ + 321, + 273, + 554, + 304 + ], + "type": "text", + "content": "[96] Lu Yu, Wei Xiang, and Kang Han. Edit-diffnerf: Editing 3d neural radiance fields using 2d diffusion model. arXiv preprint arXiv:2306.09551, 2023. 1, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 321, + 306, + 555, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 306, + 555, + 358 + ], + "spans": [ + { + "bbox": [ + 321, + 306, + 555, + 358 + ], + "type": "text", + "content": "[97] Wangbo Yu, Jinbo Xing, Li Yuan, Wenbo Hu, Xiaoyu Li, Zhipeng Huang, Xiangjun Gao, Tien-Tsin Wong, Ying Shan, and Yonghong Tian. Viewcrafter: Taming video diffusion models for high-fidelity novel view synthesis. arXiv preprint arXiv:2409.02048, 2024. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 321, + 361, + 555, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 361, + 555, + 425 + ], + "spans": [ + { + "bbox": [ + 321, + 361, + 555, + 425 + ], + "type": "text", + "content": "[98] Xianggang Yu, Mutian Xu, Yidan Zhang, Haolin Liu, Chongjie Ye, Yushuang Wu, Zizheng Yan, Chenming Zhu, Zhangyang Xiong, Tianyou Liang, et al. Mvimgnet: A large-scale dataset of multi-view images. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9150-9161, 2023. 2, 6, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 321, + 426, + 555, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 426, + 555, + 480 + ], + "spans": [ + { + "bbox": [ + 321, + 426, + 555, + 480 + ], + "type": "text", + "content": "[99] Bowen Zhang, Yiji Cheng, Jiaolong Yang, Chunyu Wang, Feng Zhao, Yansong Tang, Dong Chen, and Baining Guo. Gaussiancube: Structuring gaussian splatting using optimal transport for 3d generative modeling. arXiv preprint arXiv:2403.19655, 2024. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 482, + 555, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 482, + 555, + 525 + ], + "spans": [ + { + "bbox": [ + 316, + 482, + 555, + 525 + ], + "type": "text", + "content": "[100] Jingbo Zhang, Xiaoyu Li, Ziyu Wan, Can Wang, and Jing Liao. Text2nerf: Text-driven 3d scene generation with neural radiance fields. IEEE Transactions on Visualization & Computer Graphics, 30(12):7749-7762, 2024. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 526, + 555, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 526, + 555, + 580 + ], + "spans": [ + { + "bbox": [ + 317, + 526, + 555, + 580 + ], + "type": "text", + "content": "[101] Jason Y. Zhang, Amy Lin, Moneish Kumar, Tzu-Hsuan Yang, Deva Ramanan, and Shubham Tulsiani. Cameras as rays: Pose estimation via ray diffusion. In The Twelfth International Conference on Learning Representations, 2024. 2, 5, 7, 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 582, + 555, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 582, + 555, + 634 + ], + "spans": [ + { + "bbox": [ + 317, + 582, + 555, + 634 + ], + "type": "text", + "content": "[102] Kai Zhang, Sai Bi, Hao Tan, Yuanbo Xiangli, Nanxuan Zhao, Kalyan Sunkavalli, and Zexiang Xu. Gs-lrm: Large reconstruction model for 3d gaussian splatting. In European Conference on Computer Vision, pages 1-19. Springer, 2024. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 636, + 555, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 636, + 555, + 689 + ], + "spans": [ + { + "bbox": [ + 317, + 636, + 555, + 689 + ], + "type": "text", + "content": "[103] Kaiyi Zhang, Yang Chen, Ximing Yang, Weizhong Zhang, and Cheng Jin. Point cloud part editing: Segmentation, generation, assembly, and selection. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 7187-7195, 2024. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 691, + 555, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 555, + 713 + ], + "type": "text", + "content": "[104] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "21535" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 296, + 296 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 81, + 73, + 296, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 73, + 296, + 106 + ], + "spans": [ + { + "bbox": [ + 81, + 73, + 296, + 106 + ], + "type": "text", + "content": "deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 5, 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 296, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 296, + 138 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 296, + 138 + ], + "type": "text", + "content": "[105] Yudi Zhang, Qi Xu, and Lei Zhang. Dragtex: Generative point-based texture editing on 3d mesh. arXiv preprint arXiv:2403.02217, 2024. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 141, + 296, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 141, + 296, + 184 + ], + "spans": [ + { + "bbox": [ + 56, + 141, + 296, + 184 + ], + "type": "text", + "content": "[106] Linqi Zhou, Yilun Du, and Jiajun Wu. 3d shape generation and completion through point-voxel diffusion. In Proceedings of the IEEE/CVF international conference on computer vision, pages 5826-5835, 2021. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 186, + 296, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 186, + 296, + 251 + ], + "spans": [ + { + "bbox": [ + 56, + 186, + 296, + 251 + ], + "type": "text", + "content": "[107] Shijie Zhou, Zhiwen Fan, Dejia Xu, Haoran Chang, Pradyumna Chari, Tejas Bharadwaj, Suya You, Zhangyang Wang, and Achuta Kadambi. Dreamscene360: Unconstrained text-to-3d scene generation with panoramic gaussian splatting. In European Conference on Computer Vision, pages 324–342. Springer, 2025. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 253, + 296, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 253, + 296, + 296 + ], + "spans": [ + { + "bbox": [ + 56, + 253, + 296, + 296 + ], + "type": "text", + "content": "[108] Jingyu Zhuang, Di Kang, Yan-Pei Cao, Guanbin Li, Liang Lin, and Ying Shan. Tip-editor: An accurate 3d editor following both text-prompts and image-prompts. ACM Transactions on Graphics (TOG), 43(4):1-12, 2024. 2, 3" + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "21536" + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/SplatFlow_ Self-Supervised Dynamic Gaussian Splatting in Neural Motion Flow Field for Autonomous Driving/8f50ebde-225c-4574-a537-44bd7c2cfecd_content_list.json b/2025/SplatFlow_ Self-Supervised Dynamic Gaussian Splatting in Neural Motion Flow Field for Autonomous Driving/8f50ebde-225c-4574-a537-44bd7c2cfecd_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..caa37dba7b8c215158ff56567d23ef1d3a84d8c8 --- /dev/null +++ b/2025/SplatFlow_ Self-Supervised Dynamic Gaussian Splatting in Neural Motion Flow Field for Autonomous Driving/8f50ebde-225c-4574-a537-44bd7c2cfecd_content_list.json @@ -0,0 +1,1648 @@ +[ + { + "type": "text", + "text": "SplatFlow: Self-Supervised Dynamic Gaussian Splatting in Neural Motion Flow Field for Autonomous Driving", + "text_level": 1, + "bbox": [ + 96, + 128, + 898, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Su Sun $^{1}$ , Cheng Zhao $^{2}$ , Zhuoyang Sun $^{1}$ , Yingjie Victor Chen $^{1}$ , Mei Chen $^{2}$ , Purdue University, Microsoft", + "bbox": [ + 194, + 202, + 799, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 273, + 326, + 289 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Most existing Dynamic Gaussian Splatting methods for complex dynamic urban scenarios rely on accurate object-level supervision from expensive manual labeling, limiting their scalability in real-world applications. In this paper, we introduce SplatFlow, a Self-Supervised Dynamic Gaussian Splatting within Neural Motion Flow Fields (NMFF) to learn 4D space-time representations without requiring tracked 3D bounding boxes, enabling accurate dynamic scene reconstruction and novel view RGB/depth/flow synthesis. SplatFlow designs a unified framework to seamlessly integrate time-dependent 4D Gaussian representation within NMFF, where NMFF is a set of implicit functions to model temporal motions of both LiDAR points and Gaussians as continuous motion flow fields. Leveraging NMFF, SplatFlow effectively decomposes static background and dynamic objects, representing them with 3D and 4D Gaussian primitives, respectively. NMFF also models the correspondences of each 4D Gaussian across time, which aggregates temporal features to enhance cross-view consistency of dynamic components. SplatFlow further improves dynamic object identification by distilling features from 2D foundation models into 4D space-time representation. Comprehensive evaluations conducted on the Waymo and KITTI Datasets validate SplatFlow's state-of-the-art (SOTA) performance for both image reconstruction and novel view synthesis in dynamic urban scenarios.", + "bbox": [ + 89, + 305, + 483, + 698 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 703, + 220, + 719 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "As autonomous driving systems increasingly shift toward end-to-end models, there is a growing need for scalable simulation environments without domain gaps where these systems can undergo closed-loop evaluations. A promising approach is real-world closed-loop evaluation, which demands controllable sensor inputs, thereby driving the development of advanced scene reconstruction techniques. In this context, Neural Radiance Fields (NeRFs) [10] and 3D Gaussian Splitting (3DGS) [7] have proven to be effective in creating high-quality 3D scene reconstructions with", + "bbox": [ + 89, + 728, + 482, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ad85828fbf5e5c4793ef06b7561c407af95dd2a356c48f6d5a0183a8d7a13a44.jpg", + "image_caption": [ + "Figure 1. Top: Street GS [24]; Middle: PVG [1]; Bottom: Our SplatFlow. SplatFlow eliminates the need for 3D Bboxes required by Street GS, and enhances rendering quality compared to PVG." + ], + "image_footnote": [], + "bbox": [ + 516, + 273, + 903, + 498 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "excellent visual and geometric accuracy. However, accurately and comprehensively reconstructing dynamic driving scenes remains a major challenge, given the complexity of real-world scenarios without dynamic object annotations.", + "bbox": [ + 511, + 547, + 906, + 608 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent approaches advance NeRF-based representations by incorporating object detection and tracking to model dynamic elements, enabling photo-realistic view generation of dynamic urban street environments. Approaches [3, 8, 12, 18, 21, 23] create a scene graph, where both dynamic objects and the static background are represented as nodes, reconstructed within their canonical frames. Self-supervised methods [19, 25] model dynamic driving scenes by combining static neural field and time-dependent neural field to handle the static background and moving foreground objects separately. The SUDS [19] uses optical flow to ease the strict requirement for object labeling, while EmerNerf [25] learns temporal attributes and features in a self-supervision manner to avoid reliance on optical flow. However, these NeRF style methods face efficiency challenges in both training and inference, which becomes a major bottleneck for large-scale scene reconstruction and rendering.", + "bbox": [ + 509, + 611, + 906, + 868 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "While NeRF has proven effective for driving scenes, 3DGS offers a promising alternative due to faster train", + "bbox": [ + 511, + 869, + 906, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equally contributed as co-first author.", + "bbox": [ + 107, + 887, + 315, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "27487", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ing and rendering speed with more explicit representations. However, the original 3DGS faces notable challenges in modeling dynamic urban environments due to its limited representation capabilities in the temporal dimension. To address this, DrivingGaussian [29] introduces composite dynamic Gaussian graphs to handle multiple moving objects and incremental static Gaussians for background representation. StreetGaussian [24] optimizes the tracked bounding boxes of dynamic Gaussians along with 4D spherical harmonics to capture changing vehicle appearances. Both methods, however, require accurate object-level supervision, such as 3D object boxes and trackers, to decompose static and dynamic elements, limiting their scalability in real-world applications. In response, PVG [1] proposes a Periodic Vibration Gaussian mechanism to model dynamic driving scenes without relying on manually labeled 3D bounding boxes. However, the scene decomposition in PVG relies on time-dependent Gaussian attributes optimized by rendering losses, but ignores motion cues within the input point cloud used for Gaussian initialization, leading to suboptimal performance in challenging scenarios with rapid movements.", + "bbox": [ + 89, + 90, + 480, + 422 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To facilitate accurate scene reconstruction and real-time rendering without expensive annotations, we propose SplatFlow, a self-supervised Dynamic Gaussian Splatting method in Neural Motion Flow Fields (NMFF) for dynamic urban scenarios. In contrast to existing methods, as illustrated in Fig. 1, SplatFlow decomposes dynamic objects and the static background in a self-supervised way without requiring expensive 3D bounding box annotations. The key idea of SplatFlow is to seamlessly integrate 4D Gaussian representations within NMFF in a unified framework. The NMFF is a set of implicit functions to model temporal motions of both LiDAR points and Gaussians as continuous motion flow fields. Leveraging NMFF, SplatFlow not only enables the decomposition of dynamic and static elements from 3D LiDAR points, but also enables the status conversion of each 4D Gaussian across time. During Gaussian splatting, we represent dynamic objects using aggregated 4D Gaussians at various viewpoints and timestamps, and the static background using 3D Gaussians, respectively. In SplatFlow, NMFF's capabilities are enhanced through three components: 1) learning 3D motion priors by pretraining on 3D LiDAR data, 2) optimizing temporal status transitions of 4D Gaussians on image data, and 3) distilling knowledge from 2D foundation models.", + "bbox": [ + 91, + 426, + 482, + 787 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The novel features of SplatFlow are summarized as:", + "bbox": [ + 109, + 791, + 454, + 806 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- SplatFlow introduces a unified framework that seamlessly integrates time-varying 4D Gaussian representations into NMFF, enabling self-supervised dynamic scene reconstruction and rendering.", + "- NMFF enables scene decomposition, modeling static elements with 3D Gaussians and dynamic elements with 4D" + ], + "bbox": [ + 91, + 810, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Gaussians.", + "bbox": [ + 527, + 92, + 601, + 104 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- NMFF models correspondences of each 4D Gaussian over time, aggregating temporal features to enhance the cross-view consistency of dynamic components.", + "- SplatFlow enhances the dynamic object identification by uplifting 2D features distilled from foundation models to 4D space-time through optical flow rendering." + ], + "bbox": [ + 511, + 106, + 903, + 196 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Comprehensive experiments on the Waymo [16] and KITTI [4] benchmarks demonstrate that SplatFlow outperforms the state-of-the-art (SOTA) methods in both image reconstruction and novel view synthesis for dynamic urban scenes. Notably, SplatFlow achieves this without relying on tracked 3D bounding boxes of dynamic objects, allowing the proposed model to learn from extensive, in-the-wild data sources.", + "bbox": [ + 511, + 198, + 903, + 316 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 513, + 330, + 650, + 347 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Real-world simulation is essential for generating data to support end-to-end autonomous driving solutions. However, current simulation engines like AirSim [15] and CARLA [2] encounter difficulties due to the expensive manual work needed to create virtual environments and the domain gap between real and simulated data. The rapid progress in Novel View Synthesis (NVS) technologies, such as NeRF [10] and 3DGS [7], and following work [6, 11, 17, 26] enables 3D reconstruction and photorealistic image generation, significantly improving real-world simulation for autonomous driving.", + "bbox": [ + 511, + 356, + 905, + 521 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "For 3D reconstruction and neural rendering in autonomous driving scenes, sensor fusion solution combining surrounding cameras and LiDAR is commonly used in existing work [5, 9, 13, 14, 27]. Urban Radiance Field [14] improves NeRF training by incorporating 3D data from LiDAR to enhance 3D geometry learning. DNMP [9] employs a pre-trained deformable mesh primitive to represent the 3D scene, enhancing neural rendering quality. NPLF [13] uses explicit 3D reconstructions from LiDAR data to efficiently model the radiance field. TCLC-GS [27] design a hybrid explicit and implicit 3D representation derived from LiDAR-camera data, to enrich the properties of 3D Gaussians for splatting. Though these methods show impressive results for static urban scene rendering, they struggle with dynamic objects common in driving scenarios.", + "bbox": [ + 511, + 523, + 905, + 750 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To model dynamic urban scenes, recent approaches [8, 12, 19, 21, 25] decompose the entire scene into static background and dynamic objects, learning their representations separately. PNF [8] uses monocular 3D bounding box predictions to isolate dynamic objects and further jointly optimizes their poses during the reconstruction process. NSG [12] uses neural graphs to represent entire scenes, decomposing dynamic multi-object environments. MARS [21] employs distinct sub-networks to model backgrounds and dynamic objects, creating an instance-aware", + "bbox": [ + 511, + 750, + 905, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "27488", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/07b548e04dad42f58d9a831337019b0e497aa79b209a8c43d4f276e6f969b4ac.jpg", + "image_caption": [ + "Figure 2. The pipeline of SplatFlow." + ], + "image_footnote": [], + "bbox": [ + 94, + 99, + 883, + 313 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "simulation framework. These methods typically rely on manually annotated or predicted 3D bounding boxes. In order to avoid the need of 3D bounding boxes, SUDS [19] introduces a scalable hash table to represent large-scale dynamic urban scenes, using an off-the-shelf 2D optical flow estimator to track dynamic objects. EmerNerf [25] addresses this challenge by learning scene flow to associate points across time in the 4D neural radiance field, allowing the separation of static and dynamic objects without 3D bounding boxes. However, these methods, which rely on implicit representations, still suffer from inefficiency in both reconstruction and rendering. Recently, 3DGS [7] introduced a novel explicit 3D scene representation, combining high-quality volume rendering with a fast speed. However, 3DGS is designed for static scenes and fails when modeling dynamic moving objects. To address this, DrivingGaussian [29] and StreetGaussian [24] partition the driving scene into static and dynamic components by separate sets of Gaussians based on the 3D bounding boxes of vehicles. DrivingGaussian [29] uses a composite dynamic Gaussian graph to manage multiple moving objects while incrementally employing incremental static 3D Gaussians to represent the static background. StreetGaussian [24] models foreground vehicles with estimizable poses and a 4D spherical harmonics appearance model, and background with 3D Gaussians separately. HUGS [28] also adopts 3D object boxes to identify dynamic elements, enabling optimization of geometry, appearance, semantics, and motion attributes for dynamic Gaussians. However, all these methods rely on expensive annotated or predicted 3D bounding boxes to learn the time-dependent representations of dynamic objects. Most recent method PVG [1] introduces periodic vibration based Gaussian attributes, optimized through self-supervision without the need of 3D bounding boxes. Each Gaussian models dynamic changes over time through estimizable attributes including vibration directions, life span, and life peak.", + "bbox": [ + 89, + 333, + 485, + 895 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 513, + 332, + 650, + 349 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Overview", + "text_level": 1, + "bbox": [ + 511, + 358, + 625, + 373 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "By collecting images and point clouds with timestamps from surrounding cameras and LiDAR, we aim to learn a space-time 4D Gaussian representation of a dynamic scene without any human annotations, enabling fast and high-quality novel view rendering. As the pipeline of SplatFlow shown in Fig. 2, NMFF models the temporal motions of both 3D points and Gaussians as continuous motion flow fields. It serves two key roles in SplatFlow: 1) decomposing static and dynamic elements from 3D LiDAR points; 2) modeling the temporal correspondences of each 4D Gaussian across time. We first decompose the time-series LiDAR points into static and dynamic points by NMFF, which are then separately merged to initialize the static and dynamic Gaussians. Dynamic objects are represented by aggregated 4D Gaussians associated with NMFF over time, while the static background is modeled using 3D Gaussians. NMFF learns the correspondence of each 4D Gaussian in conjunction with optimizing 4D Gaussians' attributes along time. We also distill optical flow knowledge from a 2D foundation model into 4D space-time representation. Finally, images, depths and optical flows are rendered from novel viewpoints at different timesteps in dynamic driving scenes.", + "bbox": [ + 511, + 381, + 908, + 715 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Problem Definition:", + "text_level": 1, + "bbox": [ + 511, + 726, + 702, + 742 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In a dynamic urban scenario, we collect time-sequential data from a vehicle equipped with surrounding cameras and a LiDAR sensor. A set of surrounding images $\\mathcal{I}$ is captured by multiple surrounding cameras with corresponding intrinsic matrices $I$ and extrinsic matrices $E$ . Meanwhile, a set of 3D points $\\mathcal{P}$ is obtained from LiDAR, along with corresponding extrinsic matrices $E'$ . We denote the synchronized and calibrated multi-sensor data sequence, including timestamps $t$ , as $\\{\\mathcal{I}_i, \\mathcal{P}_i, E_i, I_i, E_i', t_i | i = 1, 2, 3, \\ldots, n\\}$ , where $\\mathcal{P}_i = \\{x_i, y_i, z_i\\}$ and $n$ is the number of frames.", + "bbox": [ + 511, + 750, + 908, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "27489", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/51b09e7a7704d37a50c9be91290854ae0eecb71dc2a453d00b3ed3c7cc3aa6cf.jpg", + "image_caption": [ + "Figure 3. Visualization of 3D LiDAR points within NMFF on Waymo dataset." + ], + "image_footnote": [], + "bbox": [ + 96, + 89, + 903, + 238 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The vehicle trajectory is either provided or estimated using multi-sensor-based odometry. Our model $\\mathcal{SF}$ aims to perform accurate 3D reconstruction and synthesize novel viewpoints for any given timestamp $t$ and camera pose $[E_t, I_t]$ by rendering $\\hat{\\mathcal{I}} = \\mathcal{SF}(E_t, I_t, t)$ .", + "bbox": [ + 89, + 258, + 483, + 335 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. 4D Gaussian Representation", + "text_level": 1, + "bbox": [ + 89, + 340, + 346, + 357 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given images with associated camera poses, 3DGS optimizes a set of anisotropic 3D Gaussians through differentiable rasterization to represent a static 3D scene. We extend 3DGS to 4DGS of spatial-temporal representations for dynamic objects in an urban scene. Each 4D Gaussian primitives $\\mathcal{G}(t)$ is represented by time-varying attributes: 3D center $\\mu(t) = [x(t), y(t), z(t)]^T$ and covariance matrix $\\Sigma(t)$ along with time-invariant attributes: opacity $\\sigma$ and color $c$ . The density of a Gaussian at point $(x,t)$ is defined as,", + "bbox": [ + 89, + 362, + 483, + 498 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\alpha (t) = \\sigma \\cdot e x p \\left(- \\frac {1}{2} (x - \\mu (t)) ^ {T} \\Sigma (t) ^ {- 1} (x - \\mu (t))\\right). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 106, + 498, + 480, + 523 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The covariance matrix $\\Sigma(t) = R(t)SS^T R(t)^T$ is composed of a scaling matrix $S = \\text{diag}(s_x, s_y, s_z)$ and rotation matrix $R(t) = (q_x(t), q_y(t), q_z(t), q_w(t))$ , constrained to be a positive semi-definite matrix during optimization.", + "bbox": [ + 89, + 523, + 482, + 584 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To render an image from a specific viewpoint at times-tamp $t$ , we first transform all the 4D Gaussians from other times to the target time $t$ , according to the learned correspondence from NMFF. Subsequently, the aggregated 4D Gaussians are splatted onto the image plane, resulting in a collection of 2D Gaussians. The 3D covariance matrix $\\Sigma$ is projected to a 2D covariance matrix $\\Sigma'$ by,", + "bbox": [ + 89, + 585, + 483, + 689 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\Sigma^ {\\prime} (t) = J E \\Sigma (t) E ^ {T} J ^ {T}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 204, + 689, + 480, + 709 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $E$ refers to the world-to-camera matrix and $J$ refers to Jacobian of the perspective transformation.", + "bbox": [ + 89, + 710, + 482, + 741 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "By sorting the Gaussians according to their depth within the camera space, we use $\\alpha$ -blending to estimate the color $C$ and depth $D$ at each pixel $p$ as,", + "bbox": [ + 89, + 741, + 483, + 786 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nC = \\sum_ {i = 1} ^ {N} c _ {i} \\alpha_ {i} \\prod_ {j = 1} ^ {i - 1} (1 - \\alpha_ {j}), D = \\sum_ {i = 1} ^ {N} z _ {i} \\alpha_ {i} \\prod_ {j = 1} ^ {i - 1} (1 - \\alpha_ {j}). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 98, + 785, + 482, + 825 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here $c_{i}$ represents the color of Gaussian $\\mathcal{G}_{i}$ , which is computed using spherical harmonic coefficients. $z_{i}$ represents the distance from the image plane to the $i$ -th Gaussian center. $\\alpha_{i}$ denotes the density computed from the 2D projection of the $i$ -th Gaussian and its learned opacity.", + "bbox": [ + 89, + 825, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Neural Motion Flow Field", + "text_level": 1, + "bbox": [ + 511, + 257, + 750, + 272 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To establish the correspondence of dynamic objects, we design NMFF to model 3D motions as a continuous motion flow field, enabling the temporal transition of both 3D points and Gaussians. The NMFF $\\Phi$ contains a temporal sequence of motion flow field, defined as $\\Phi = \\{\\phi_{t_i:t_{i + 1}}\\} ,i = 0,1,2,\\ldots ,n - 1$ , which predicts 3D motion flow of an arbitrary query point between two consecutive frames as,", + "bbox": [ + 511, + 280, + 906, + 386 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\phi_ {t _ {1}: t _ {2}} \\left(x _ {t _ {1}}, y _ {t _ {1}}, z _ {t _ {1}}\\right) = \\Delta x _ {t _ {1}: t _ {2}}, \\Delta y _ {t _ {1}: t _ {2}}, \\Delta z _ {t _ {1}: t _ {2}}, \\Delta R _ {t _ {1}: t _ {2}} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 521, + 390, + 903, + 421 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\Delta R_{t_1:t_2}$ denotes motion angle between two adjacent timestamps. Each field $\\phi$ is eight ReLU-MLP stacks.", + "bbox": [ + 511, + 421, + 903, + 450 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Point Cloud within NMFF: Although NMFF can be learned during Gaussian Splatting optimization, we observed that the rendering loss alone is insufficient to effectively constrain the motion representation of dynamic scene components. Therefore, we leverage the rich geometric structure inherent in temporally consecutive 3D LiDAR points to derive a robust NMFF prior, as shown in Fig. 3.", + "bbox": [ + 511, + 452, + 905, + 556 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We first pre-train $\\Phi$ on a sequence of 3D LiDAR points by forward and backward 3D geometry consistency, enabling NMFF to learn the 3D motion of each query 3D point over time. Given the source point clouds $\\mathcal{P}$ and target point clouds $\\mathcal{Q}$ from different timestamps, we minimize the point distance between the source and target point clouds by a bidirectional Chamfer Distance $(CD)$ ,", + "bbox": [ + 511, + 558, + 905, + 662 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nd _ {\\mathrm {C D}} (\\mathcal {P}, \\mathcal {Q}) = \\sum_ {p \\in \\mathcal {P}} \\min _ {q \\in \\mathcal {Q}} \\| p - q \\| ^ {2} + \\sum_ {q \\in \\mathcal {Q}} \\min _ {p \\in \\mathcal {P}} \\| q - p \\| ^ {2}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 667, + 903, + 700 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The bidirectional $CD$ is a point-based distance that establishes correspondences in both directions—source to target and target to source—by searching for the nearest neighbors in each point cloud.", + "bbox": [ + 511, + 704, + 905, + 763 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We compensate for ego-car motion in LiDAR points to let NMFF learn only 3D motion flow during pre-training. As a result, the 3D motion flow predicted by NMFF excludes the ego-car motion. We generate a 3D dynamic mask to separate static and dynamic points by applying a threshold on each point's 3D motion flow. We then merge static point clouds using ego-motion alone, and dynamic point clouds using both ego-motion and motion flow. For the static background, we initialize the static 3D Gaus", + "bbox": [ + 511, + 765, + 906, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "27490", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "sians using the merged 3D static points. For dynamic objects, we initialize 4D Gaussians by aggregating points from different timestamps into a common reference frame (the mid-timestamp reference frame), which increases density and completeness by integrating observations from various viewpoints over time.", + "bbox": [ + 88, + 90, + 480, + 180 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4D Gaussian within NMFF: The pre-trained $\\Phi$ is subsequently integrated with 4D Gaussian representations of dynamic objects, enabling joint optimization during Gaussian splatting. Given consecutive time $t_1$ and $t_2$ where $t_1 < t_2$ with respective states $\\{\\mathcal{G}_i(t_1)\\}$ and $\\{\\mathcal{G}_i(t_2)\\}$ , these states are connected by NMFF for each Gaussian as,", + "bbox": [ + 89, + 181, + 482, + 271 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {G} _ {i} \\left(t _ {1}\\right) = \\left\\{\\mu \\left(t _ {1}\\right), R \\left(t _ {1}\\right), S, \\alpha , c \\right\\}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 176, + 272, + 480, + 289 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\Delta \\mu_ {t _ {1}: t _ {2}}, \\Delta R _ {t _ {1}: t _ {2}} = \\phi_ {t _ {1}: t _ {2}} (\\mu (t _ {1})), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 171, + 292, + 482, + 309 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathcal {G}} _ {i} \\left(t _ {2}\\right) = \\left\\{\\mu \\left(t _ {1}\\right) + \\Delta \\mu_ {t _ {1}: t _ {2}}, R \\left(t _ {1}\\right) \\cdot \\Delta R _ {t _ {1}: t _ {2}}, S, \\alpha , c \\right\\}. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 314, + 482, + 332 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We employ NMFF $\\Phi$ correspondence to warp the center of all aggregated 4D Gaussians across time to the desired timestamp, as $\\{\\mathcal{G}_i(t)\\} = \\{\\mathcal{G}(t),\\hat{\\mathcal{G}} (t_1),\\hat{\\mathcal{G}} (t_2),\\dots ,\\hat{\\mathcal{G}} (t_n)\\}$ . For the transition of Gaussians across multiple timestamps, we apply the corresponding NMFF through step-by-step propagation. In practice, we propagate the aggregated 4D Gaussian at the mid-timestamp to the target timestamp. Here NMFF learns a motion transition of the 4D Gaussian among training frames in a self-supervised manner, promoting a multi-view consistent representation.", + "bbox": [ + 89, + 335, + 482, + 484 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For static background, we represent it by static 3D Gaussians $\\mathcal{G}_{\\text{static}}$ following the strategy [1]. For sky area, we adopt a separate environmental map with static sky Gaussians $\\mathcal{G}_{sky}$ following [1, 25]. Given a specific viewpoint $[E_t, I_t, t]$ , the image $\\hat{\\mathcal{I}}$ and depth $\\hat{\\mathcal{D}}$ at time $t$ are obtained by the differentiable rendering using a set of Gaussians,", + "bbox": [ + 89, + 486, + 482, + 579 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathcal {L}}, \\hat {\\mathcal {D}} = \\operatorname {R e n d e r} \\left(\\left\\{\\mathcal {G} _ {i} (t) \\right\\}, \\mathcal {G} _ {\\text {s t a t i c}}, \\mathcal {G} _ {\\text {s k y}} \\mid E _ {t}, I _ {t}, t\\right). \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 114, + 584, + 482, + 602 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Optical Flow Distillation", + "text_level": 1, + "bbox": [ + 89, + 611, + 316, + 627 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The 3D motion flow provided by NMFF enables 2D optical flow rendering. Given two consecutive time $t_1$ and $t_2$ , the optical flow $f_{t_1:t_2}$ of each Gaussian is computed by projecting 3D center $\\mu$ to the image plane using the camera's intrinsic matrix $I$ and extrinsic matrix $E$ ,", + "bbox": [ + 89, + 633, + 482, + 708 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mu^ {\\prime} \\left(t _ {1}\\right) = I [ E ] \\mu \\left(t _ {1}\\right), \\quad \\mu^ {\\prime} \\left(t _ {2}\\right) = I [ E ] \\mu \\left(t _ {2}\\right), \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 129, + 710, + 480, + 729 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nf _ {t _ {1}: t _ {2}} = \\mu^ {\\prime} (t _ {2}) - \\mu^ {\\prime} (t _ {1}). \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 202, + 731, + 480, + 748 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Then we estimate the optical flow $F$ at pixel $p$ via by point-based $\\alpha$ -blending as,", + "bbox": [ + 89, + 751, + 482, + 780 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nF = \\sum_ {i = 1} ^ {N} f _ {i} \\alpha_ {i} \\prod_ {j = 1} ^ {i - 1} (1 - \\alpha_ {j}). \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 194, + 780, + 480, + 823 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The optical flow $\\hat{\\mathcal{F}}$ at time $t$ are rendered by the differentiable rendering similar as Equation 9. In order to distill optical flow knowledge from 2D foundation model into 4D space, we adopt SEA-RAFT[20] to extract optical flow $\\mathcal{F}$ as pseudo ground truth for distillation by optical flow loss,", + "bbox": [ + 89, + 824, + 482, + 900 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {F} = \\lambda_ {f} \\mathcal {L} _ {1} (\\mathcal {F}, \\hat {\\mathcal {F}}) + (1 - \\lambda_ {f}) \\mathcal {L} _ {1} \\left(\\mathcal {I} _ {\\text {n e x t}}, \\mathcal {T} (\\hat {\\mathcal {I}} | \\hat {\\mathcal {F}})\\right), \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 89, + 903, + 108 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda_{f}$ is a scale factor, and $\\mathcal{I}_{\\text{next}}$ denotes the image at the next time step. $\\mathcal{T}$ is used to warp rendered image to next time step according to optical flow. The optical flow $\\mathcal{F}$ extracted from foundation model is only required during training and is not needed during inference.", + "bbox": [ + 511, + 112, + 903, + 188 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.6. Optimization", + "text_level": 1, + "bbox": [ + 511, + 196, + 651, + 212 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "All Gaussian attributes, along with the parameters of NMFF, are optimized end-to-end in a self-supervised manner. Meanwhile, adaptive densification and pruning strategies are introduced to enhance the fitting of the 3D scene. The overall training loss is given as,", + "bbox": [ + 511, + 218, + 906, + 294 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {I} + \\lambda_ {1} \\mathcal {L} _ {D} + \\lambda_ {2} \\mathcal {L} _ {F} + \\lambda_ {3} \\mathcal {L} _ {s k y} + \\lambda_ {4} \\mathcal {L} _ {r e g}, \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 537, + 297, + 903, + 316 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda_{1,2,3,4}$ are scale factors and $\\mathcal{L}_{reg}$ is the regularization term. The image loss $\\mathcal{L}_I$ combines L1 and SSIM losses between rendered and observed images,", + "bbox": [ + 511, + 320, + 903, + 366 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {I} = \\left(1 - \\lambda_ {s s i m}\\right) \\mathcal {L} _ {1} (\\mathcal {I}, \\hat {\\mathcal {I}}) + \\lambda_ {s s i m} \\mathcal {L} _ {s s i m} (\\mathcal {I}, \\hat {\\mathcal {I}}), \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 531, + 369, + 903, + 388 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda_{ssim}$ is a scale factor. The depth loss $\\mathcal{L}_D$ is a L1 loss between inverse of rendered depth and generated depth by projecting sparse LiDAR points onto camera plane as,", + "bbox": [ + 511, + 393, + 903, + 438 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {D} = \\mathcal {L} _ {1} (\\mathcal {D}, \\hat {\\mathcal {D}}). \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 650, + 441, + 903, + 459 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The sky opacity loss $\\mathcal{L}_{sky}$ is a binary cross entropy loss for sky refining using sky mask $M_{sky}$ from SegFormer [22],", + "bbox": [ + 511, + 464, + 903, + 496 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {s k y} = \\mathcal {L} _ {b c e} (O, 1 - M _ {s k y}), \\tag {17}\n$$\n", + "text_format": "latex", + "bbox": [ + 612, + 501, + 903, + 518 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where the accumulated opacity denotes as $O = \\sum_{i=1}^{N} \\alpha_i \\prod_{j=1}^{i-1} (1 - \\alpha_j)$ . Note the sky mask is only required during training and is not needed during inference.", + "bbox": [ + 511, + 523, + 903, + 570 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 582, + 645, + 599 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Datasets, Metrics and Baselines", + "text_level": 1, + "bbox": [ + 511, + 606, + 790, + 621 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Following PVG [1], we validate our approach on two widely-used datasets in autonomous driving: Waymo Open Dataset [16] and KITTI Dataset [4]. Both offer multi-sensor data clips including synchronized and calibrated camera and LiDAR data from urban driving scenarios. We train and test our methods, along with all baselines, using full-resolution images: $1920 \\times 1280$ for the Waymo dataset and $1242 \\times 375$ for the KITTI dataset.", + "bbox": [ + 511, + 628, + 906, + 748 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Following prior research [1, 19, 24, 25, 29], we evaluate image reconstruction and novel view synthesis using three widely-adopted benchmark metrics: peak signal-to-noise ratio (PSNR), structural similarity index measure (SSIM), and learned perceptual image patch similarity (LPIPS).", + "bbox": [ + 511, + 750, + 903, + 825 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We compare our method with extensive baselines including S-NeRF [23], StreetSurf [5], 3DGS [7], NSG [12], Mars [21], SUDS [19], EmerNerf [25], PVG [1], Street-Gaussian [24] on the Waymo and KITTI benchmarks. See supplementary material for more details and results.", + "bbox": [ + 511, + 825, + 903, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "27491", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/fd1fd9d56f6988cf70fa3339459f5f08f8fd50ec5bac41f7f5ef742b84cf96cf.jpg", + "image_caption": [ + "Figure 4. Visual comparison of novel view synthesis on Waymo dataset. Bounding boxes indicate the zoomed-in dynamic areas." + ], + "image_footnote": [], + "bbox": [ + 91, + 87, + 883, + 592 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/9ea1161dc35b3efefc018eb221f77ade5fd1ae45ac1eef9996c8d66463a9a8c8.jpg", + "image_caption": [ + "Figure 5. Dynamic object decomposition comparison on Waymo." + ], + "image_footnote": [], + "bbox": [ + 91, + 612, + 480, + 765 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Evaluation on Waymo Open Dataset", + "text_level": 1, + "bbox": [ + 89, + 784, + 406, + 800 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Following PVG [1], we evaluated our SplatFlow against baselines using the Waymo Open dataset. Table 1 summarizes the average metrics for both image reconstruction and novel view synthesis tasks on selected dynamic scenes. SplatFlow consistently outperforms all baselines across all evaluated metrics for both tasks. In image reconstruction,", + "bbox": [ + 89, + 809, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/ea5e20e232bf5067f76b6951596e10379a0290fdd3c15f6a0dd666648a0cd67d.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Image ReconstructionNovel View Synthesis
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
S-NeRF [23]19.670.5280.38719.220.5150.400
StreetSurf [5]26.700.8460.38723.780.8220.401
3DGS [7]27.990.8660.37225.080.8220.319
NSG [12]24.080.6560.29321.010.5710.487
Mars [21]21.810.6810.44120.690.6360.453
SUDS [19]28.830.8050.43021.830.6560.405
EmerNerf [25]28.110.7860.28925.920.7630.384
PVG [1]32.460.9100.37328.110.8490.279
SplatFlow33.640.9510.19828.710.8740.239
", + "bbox": [ + 517, + 612, + 921, + 741 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Performance comparison on Waymo dataset.", + "bbox": [ + 549, + 744, + 867, + 757 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "SplatFlow achieves high image quality with the highest scores: PSNR at 33.64, SSIM at 0.951, and LPIPS at 0.198. For novel view synthesis, SplatFlow produces high-quality renderings of unseen timestamps, reaching PSNR of 28.71, SSIM of 0.874 and LPIPS of 0.239. These improvements are visually validated in Fig. 4 and Fig. 9 (dynamic only), showing exceptional clarity in details for both static and dynamic regions. The dynamic objects in Fig. 9 are selected from a distant background car in an extremely zoomed-in", + "bbox": [ + 511, + 763, + 906, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "27492", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/428c433cf61a8afe8c8e3d527ba4d11a397f89fb8cd48d1d0f2620400fb0b2bf.jpg", + "image_caption": [ + "Figure 6. Visual comparison of novel view synthesis on KITTI dataset. Bounding boxes indicate the zoomed-in dynamic areas." + ], + "image_footnote": [], + "bbox": [ + 93, + 87, + 906, + 390 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/6507cc75f14291c35ede8eee2ff57164abad78aad9bf81e589335a565189e3d4.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
PVG [1]SplatFlow
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
Seg. 1058...26.860.8400.27327.460.8800.253
Seg. 2259...26.450.8240.30929.300.8560.290
Seg. 7670...26.580.8200.30728.890.8550.283
Seg. 5083...28.350.8930.21330.290.9280.169
PSNR*↑SSIM*↑LPIPS*↓PSNR*↑SSIM*↑LPIPS*↓
Seg. 1058...27.420.8350.26728.500.8770.234
Seg. 2259...22.550.8030.35031.610.8990.191
Seg. 7670...23.560.7560.33028.760.8200.289
Seg. 5083...25.580.8140.29026.710.8550.211
", + "bbox": [ + 102, + 407, + 465, + 551 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/6bd6bd1e7ea16c36494d09c08b8e3ac4e96b431b7496b32e50237515bf80413e.jpg", + "table_caption": [ + "Table 2. Novel view synthesis results on Waymo dataset (*denotes dynamic elements only)." + ], + "table_footnote": [], + "table_body": "
Image ReconstructionNovel View Synthesis
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
S-NeRF [23]19.230.6640.19318.710.6060.352
StreetSurf [5]24.140.8190.25722.480.7630.304
3DGS [7]21.020.8110.20219.540.7760.224
NSG [12]26.660.8060.18621.530.6730.254
Mars [21]27.960.9000.18524.230.8450.160
SUDS [19]28.310.8760.18522.770.7970.171
EmerNeRF [25]26.950.8280.21825.240.8010.237
PVG [1]32.830.9370.07027.430.8960.114
SplatFlow33.370.9430.05728.320.9320.089
", + "bbox": [ + 94, + 585, + 500, + 713 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Performance comparison on KITTI dataset.", + "bbox": [ + 130, + 715, + 441, + 729 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "view. Compared with baselines which show artifacts such as ghosting and blur in regions of dynamic elements, our method preserves high-quality textures and fine details, producing more accurate renderings at novel viewpoints.", + "bbox": [ + 89, + 733, + 482, + 792 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To further evaluate the rendering quality of dynamic objects against the primary baseline PVG, we selected additional four scenes featuring a higher presence of dynamic objects for further comparison. Table 2 displays the metrics for novel view synthesis, evaluated both on entire scenes and exclusively on dynamic objects (denoted as *). Splat-Flow consistently surpasses PVG in all selected scenes,", + "bbox": [ + 89, + 795, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/336bbec7d5ff23eea71f5723ec3dff7eab8e59b8e18489509cb111d4e75b7a55.jpg", + "image_caption": [ + "Figure 7. Visual comparison of novel view synthesis on KITTI-25% (row1), 50% (row2), and 75% (row3) dataset." + ], + "image_footnote": [], + "bbox": [ + 516, + 407, + 905, + 650 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "demonstrating higher PSNR and SSIM values and lower LPIPS scores, underscoring its effectiveness in handling complex, dynamic content. The dynamic objects masks are obtained from 2D ground truth bounding boxes following [25]. The PSNR* of dynamic regions on novel view synthesis significantly outperforms SVG both quantitatively and qualitatively, especially in scenes with fast ego-motion and moving objects in Seg. 2259 and Seg. 7670. We also present the visual comparison of dynamic object decomposition with SVG in Fig. 5, which illustrates that SplatFlow renders sharper and clearer image of dynamic objects.", + "bbox": [ + 511, + 679, + 906, + 845 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Evaluation on KITTI Dataset", + "text_level": 1, + "bbox": [ + 511, + 849, + 774, + 864 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Following PVG [1], we further evaluated our SplatFlow method in comparison to the baselines on KITTI dataset.", + "bbox": [ + 511, + 869, + 905, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "27493", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/fabde9e27eb24c2c0923f35c501fdbc33f2afb1ad7ff45d9e405fc76d92f969d.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
KITTI - 75%KITTI - 50%KITTI - 25%
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
NeRF [10]18.560.5570.55419.120.5870.49718.610.5700.510
NSG [12]21.530.6730.25421.260.6590.26620.000.6320.281
SUDS [19]22.770.7970.17123.120.8210.13520.760.7470.198
MARS [21]24.230.8450.16024.000.8010.16423.230.7560.177
3DGS [7]19.190.7370.17219.230.7390.17419.060.7300.180
PVG [1]27.430.8960.11425.920.8820.11422.550.8330.151
StreetGS [24]25.790.8440.08125.520.8410.08424.530.8240.090
SplatFlow28.320.9320.08927.900.9270.09326.100.8900.120
", + "bbox": [ + 109, + 88, + 596, + 202 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/fc82824f760c923b49de30b58f1b8bdcdaec7b2d7c5e26c92b196aaadb347fad.jpg", + "table_caption": [ + "Table 4. Performance comparison of novel view synthesis on KITTI-75%, 50% and 25% dataset." + ], + "table_footnote": [], + "table_body": "
PSNR↑SSIM↑LPIPS↓
w/o NMFF prior27.690.8630.282
w/o NMFF optimization28.140.8740.269
w/o optical flow distillation28.280.8770.252
Full28.990.8800.249
PSNR*↑SSIM*↑LPIPS*↓
w/o NMFF prior27.280.8270.317
w/o NMFF optimization27.970.8430.254
w/o optical flow distillation28.510.8610.232
Full28.900.8630.231
", + "bbox": [ + 598, + 95, + 885, + 198 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. Ablation study on Waymo dataset (*denotes dynamic elements only).", + "bbox": [ + 596, + 199, + 883, + 226 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b23fb79d066066217788bf69edd4a526f759dd80acd773c4f34f9c6f933bdf6f.jpg", + "image_caption": [ + "Ours (Full)" + ], + "image_footnote": [], + "bbox": [ + 94, + 237, + 287, + 338 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c10f81fdc6b6a701e91d9ab9769e18015d37c53429188620372c08403c115e9d.jpg", + "image_caption": [ + "w/o Opt. flow" + ], + "image_footnote": [], + "bbox": [ + 287, + 237, + 480, + 338 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7984c352b65409864afbe33329dfdda80a0112e3397b395d7df45cc9af2194ac.jpg", + "image_caption": [ + "w/o NMFF optim.", + "Figure 8. Visual comparison of ablation study on Waymo dataset." + ], + "image_footnote": [], + "bbox": [ + 94, + 348, + 287, + 446 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9d2312c8543933d1ba7f9195ecfb5ffbcdf8a45b638585d4ff25448a60f82742.jpg", + "image_caption": [ + "w/o NMFF prior" + ], + "image_footnote": [], + "bbox": [ + 287, + 348, + 480, + 446 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3 presents the average metrics for both image reconstruction and novel view synthesis tasks on selected dynamic scenes characterized by extensive movement. SplatFlow outperforms all other methods across all metrics in both tasks, with particularly high scores in PSNR. SSIM and LPIPS, achieving 33.37, 0.943 and 0.057 for image reconstruction and 28.32, 0.932 and 0.089 for novel view synthesis. SplatFlow achieves exceptional results in novel view synthesis, as shown in Fig. 6. Our approach generates high-fidelity renderings that preserve fine details while accurately reconstructing dynamic areas where existing approaches struggle to maintain stability.", + "bbox": [ + 88, + 489, + 482, + 669 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To evaluate robustness to training data size, we follow the settings of [21, 24] to evaluate our method using different train/test split configurations on the KITTI dataset. Table 4 and Fig. 7 provide the performance comparisons and visualizations. SplatFlow consistently outperforms all other methods across all dataset subsets, achieving the highest scores. Even with reduced training data, SplatFlow robustly delivers high-quality novel view synthesis across varying levels of data availability, highlighting its efficient use of available information and strong generalization capabilities.", + "bbox": [ + 88, + 670, + 482, + 821 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 832, + 243, + 848 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To demonstrate the effectiveness of each component in SplatFlow, we performed ablation studies on the Waymo dataset. Our main contributions, NMFF prior, NMFF opti", + "bbox": [ + 89, + 854, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0243cf949d0320044abb1b88aef349e7e5ef7d6cec9901131ee8bd565d40b8f8.jpg", + "image_caption": [ + "EmerNeRF", + "Figure 9. Detail comparison of novel view synthesis on Waymo." + ], + "image_footnote": [], + "bbox": [ + 516, + 237, + 612, + 474 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/60cc694cc2abedebf79f2a9db38753c85c8a798657e9360f71b41e1244ec7422.jpg", + "image_caption": [ + "PVG" + ], + "image_footnote": [], + "bbox": [ + 614, + 237, + 709, + 474 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b4e74a1558ae485f8e44c8c1daad2b64cbfd1e718bdd456c26411786770e1179.jpg", + "image_caption": [ + "SplatFlow" + ], + "image_footnote": [], + "bbox": [ + 710, + 237, + 805, + 474 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b9d300c49a381b8f94c7574a230ecfab7f3b7ed1e1c7a80e8023b8fc493ea20f.jpg", + "image_caption": [ + "G.T." + ], + "image_footnote": [], + "bbox": [ + 807, + 237, + 903, + 474 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "mization, and optical flow distillation are analyzed in this study to assess their individual impact. We trained four different variations: 1) SplatFlow without pretraining on LiDAR data; 2) SplatFlow without optimization with 4D Gaussians on image data; 3) SplatFlow without optical flow distillation from foundational model; 4) SplatFlow full method. Table 5 presents the average evaluation metrics across test scenes used in Table 2, showing results for both entire scenes and dynamic objects (indicated as *). Ablation study validates the contribution of each component to the overall performance.", + "bbox": [ + 511, + 512, + 906, + 679 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 694, + 633, + 709 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we propose SplatFlow, a novel self-supervised Dynamic Gaussian Splitting within Neural Motion Flow Fields (NMFF) for accurate reconstruction and real-time rendering in dynamic urban scenarios. The core idea of SplatFlow is to seamlessly integrate time-dependent 4D Gaussian representations within NMFF in a unified framework, where NMFF implicitly models the motions of dynamic components across time through self-supervision. Experimental evaluations show that SplatFlow outperforms state-of-the-art across all standard metrics on the Waymo Open and KITTI datasets, without requiring expensive annotations of dynamic object detection and tracking.", + "bbox": [ + 511, + 719, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "27494", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 91, + 90, + 250, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was partially supported by the intramural research program of the U.S. Department of Agriculture Natural Resources Conservation Service, Climate-Smart Commodities program, EFFICACI grant, #NR233A750004G044. The findings and conclusions in this work have not been formally disseminated by the U. S. Department of Agriculture and should not be construed to represent any agency determination or policy.", + "bbox": [ + 89, + 114, + 483, + 237 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 247, + 187, + 263 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Yurui Chen, Chun Gu, Junzhe Jiang, Xiatian Zhu, and Li Zhang. Periodic vibration gaussian: Dynamic urban scene reconstruction and real-time rendering. arXiv preprint arXiv:2311.18561, 20.1, 2, 3, 5, 6, 7, 8", + "[2] Alexey Dosovitskiy, German Ros, Felipe Codevilla, Antonio Lopez, and Vladlen Koltun. Carla: An open urban driving simulator. In Conference on robot learning, pages 1-16. PMLR, 2017. 2", + "[3] Tobias Fischer, Lorenzo Porzi, Samuel Rota Bulo, Marc Pollefeys, and Peter Kontschieder. Multi-level neural scene graphs for dynamic urban environments. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21125-21135, 2024. 1", + "[4] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In 2012 IEEE conference on computer vision and pattern recognition, pages 3354-3361. IEEE, 2012. 2, 5", + "[5] Jianfei Guo, Nianchen Deng, Xinyang Li, Yeqi Bai, Botian Shi, Chiyu Wang, Chenjing Ding, Dongliang Wang, and Yikang Li. Streetsurf: Extending multi-view implicit surface reconstruction to street views. arXiv preprint arXiv:2306.04988, 2023. 2, 5, 6, 7", + "[6] Antoine Guédon and Vincent Lepetit. Sugar: Surface-aligned gaussian splatting for efficient 3d mesh reconstruction and high-quality mesh rendering. In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5354-5363, 2024. 2", + "[7] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42 (4), 2023. 1, 2, 3, 5, 6, 7, 8", + "[8] Abhijit Kundu, Kyle Genova, Xiaqi Yin, Alireza Fathi, Caroline Pantofaru, Leonidas J Guibas, Andrea Tagliasacchi, Frank Dellaert, and Thomas Funkhouser. Panoptic neural fields: A semantic object-aware neural scene representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12871-12881, 2022. 1, 2", + "[9] Fan Lu, Yan Xu, Guang Chen, Hongsheng Li, Kwan-Yee Lin, and Changjun Jiang. Urban radiance field representation with deformable neural mesh primitives. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 465-476, 2023. 2", + "[10] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf:" + ], + "bbox": [ + 93, + 272, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 1, 2, 8", + "[11] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph., 41(4):102:1-102:15, 2022. 2", + "[12] Julian Ost, Fahim Mannan, Nils Thuerey, Julian Knodt, and Felix Heide. Neural scene graphs for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2856-2865, 2021. 1, 2, 5, 6, 7, 8", + "[13] Julian Ost, Issam Laradji, Alejandro Newell, Yuval Bahat, and Felix Heide. Neural point light fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18419-18429, 2022. 2", + "[14] Konstantinos Rematas, Andrew Liu, Pratul P Srinivasan, Jonathan T Barron, Andrea Tagliasacchi, Thomas Funkhouser, and Vittorio Ferrari. Urban radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12932-12942, 2022. 2", + "[15] Shital Shah, Debadeepta Dey, Chris Lovett, and Ashish Kapoor. Airsim: High-fidelity visual and physical simulation for autonomous vehicles. In Field and Service Robotics: Results of the 11th International Conference, pages 621-635. Springer, 2018. 2", + "[16] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, et al. Scalability in perception for autonomous driving: Waymo open dataset. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2446-2454, 2020. 2, 5", + "[17] Su Sun, Cheng Zhao, Yuliang Guo, Ruoyu Wang, Xinyu Huang, Yingjie Victor Chen, and Liu Ren. Behind the veil: Enhanced indoor 3d scene reconstruction with occluded surfaces completion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12744-12753, 2024. 2", + "[18] Adam Tonderski, Carl Lindström, Georg Hess, William Ljungbergh, Lennart Svensson, and Christoffer Petersson. Neurad: Neural rendering for autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14895-14904, 2024. 1", + "[19] Haithem Turki, Jason Y Zhang, Francesco Ferroni, and Deva Ramanan. Suds: Scalable urban dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12375-12385, 2023. 1, 2, 3, 5, 6, 7, 8", + "[20] Yihan Wang, Lahav Lipson, and Jia Deng. Sea-raft: Simple, efficient, accurate raft for optical flow. In European Conference on Computer Vision, pages 36-54. Springer, 2025. 5", + "[21] Zirui Wu, Tianyu Liu, Liyi Luo, Zhide Zhong, Jianteng Chen, Hongmin Xiao, Chao Hou, Haozhe Lou, Yuantao Chen, Runyi Yang, et al. Mars: An instance-aware, modular and realistic simulator for autonomous driving. In CAAI International Conference on Artificial Intelligence, pages 3-15. Springer, 2023. 1, 2, 5, 6, 7, 8" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "27495", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[22] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. In Neural Information Processing Systems (NeurIPS), 2021. 5", + "[23] Ziyang Xie, Junge Zhang, Wenye Li, Feihu Zhang, and Li Zhang. S-nerf: Neural radiance fields for street views. In International Conference on Learning Representations (ICLR), 2023. 1, 5, 6, 7", + "[24] Yunzhi Yan, Haotong Lin, Chenxu Zhou, Weijie Wang, Haiyang Sun, Kun Zhan, Xianpeng Lang, Xiaowei Zhou, and Sida Peng. Street gaussians for modeling dynamic urban scenes. arXiv preprint arXiv:2401.01339, 2024. 1, 2, 3, 5, 8", + "[25] Jiawei Yang, Boris Ivanovic, Or Litany, Xinshuo Weng, Seung Wook Kim, Boyi Li, Tong Che, Danfei Xu, Sanja Fidler, Marco Pavone, et al. Emernerf: Emergent spatial-temporal scene decomposition via self-supervision. arXiv preprint arXiv:2311.02077, 2023. 1, 2, 3, 5, 6, 7", + "[26] Xiaoshuai Zhang, Sai Bi, Kalyan Sunkavalli, Hao Su, and Zexiang Xu. Nerfusion: Fusing radiance fields for large-scale scene reconstruction. CVPR, 2022. 2", + "[27] Cheng Zhao, Su Sun, Ruoyu Wang, Yuliang Guo, Jun-Jun Wan, Zhou Huang, Xinyu Huang, Yingjie Victor Chen, and Liu Ren. Tclc-gs: Tightly coupled lidar-camera gaussian splatting for autonomous driving: Supplementary materials. In European Conference on Computer Vision, pages 91-106. Springer, 2024. 2", + "[28] Hongyu Zhou, Jiahao Shao, Lu Xu, Dongfeng Bai, Weichao Qiu, Bingbing Liu, Yue Wang, Andreas Geiger, and Yiyi Liao. Hugs: Holistic urban 3d scene understanding via gaussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21336-21345, 2024. 3", + "[29] Xiaoyu Zhou, Zhiwei Lin, Xiaojun Shan, Yongtao Wang, Deqing Sun, and Ming-Hsuan Yang. Drivinggaussian: Composite gaussian splatting for surrounding dynamic autonomous driving scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21634-21643, 2024. 2, 3, 5" + ], + "bbox": [ + 89, + 92, + 483, + 640 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "27496", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/SplatFlow_ Self-Supervised Dynamic Gaussian Splatting in Neural Motion Flow Field for Autonomous Driving/8f50ebde-225c-4574-a537-44bd7c2cfecd_model.json b/2025/SplatFlow_ Self-Supervised Dynamic Gaussian Splatting in Neural Motion Flow Field for Autonomous Driving/8f50ebde-225c-4574-a537-44bd7c2cfecd_model.json new file mode 100644 index 0000000000000000000000000000000000000000..cb6ca972b1b22593fa01df053cad4b52244a1353 --- /dev/null +++ b/2025/SplatFlow_ Self-Supervised Dynamic Gaussian Splatting in Neural Motion Flow Field for Autonomous Driving/8f50ebde-225c-4574-a537-44bd7c2cfecd_model.json @@ -0,0 +1,2090 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.098, + 0.13, + 0.9, + 0.177 + ], + "angle": 0, + "content": "SplatFlow: Self-Supervised Dynamic Gaussian Splatting in Neural Motion Flow Field for Autonomous Driving" + }, + { + "type": "text", + "bbox": [ + 0.196, + 0.203, + 0.8, + 0.24 + ], + "angle": 0, + "content": "Su Sun\\(^{1}\\), Cheng Zhao\\(^{2}\\), Zhuoyang Sun\\(^{1}\\), Yingjie Victor Chen\\(^{1}\\), Mei Chen\\(^{2}\\), Purdue University, Microsoft" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.274, + 0.327, + 0.29 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.306, + 0.485, + 0.699 + ], + "angle": 0, + "content": "Most existing Dynamic Gaussian Splatting methods for complex dynamic urban scenarios rely on accurate object-level supervision from expensive manual labeling, limiting their scalability in real-world applications. In this paper, we introduce SplatFlow, a Self-Supervised Dynamic Gaussian Splatting within Neural Motion Flow Fields (NMFF) to learn 4D space-time representations without requiring tracked 3D bounding boxes, enabling accurate dynamic scene reconstruction and novel view RGB/depth/flow synthesis. SplatFlow designs a unified framework to seamlessly integrate time-dependent 4D Gaussian representation within NMFF, where NMFF is a set of implicit functions to model temporal motions of both LiDAR points and Gaussians as continuous motion flow fields. Leveraging NMFF, SplatFlow effectively decomposes static background and dynamic objects, representing them with 3D and 4D Gaussian primitives, respectively. NMFF also models the correspondences of each 4D Gaussian across time, which aggregates temporal features to enhance cross-view consistency of dynamic components. SplatFlow further improves dynamic object identification by distilling features from 2D foundation models into 4D space-time representation. Comprehensive evaluations conducted on the Waymo and KITTI Datasets validate SplatFlow's state-of-the-art (SOTA) performance for both image reconstruction and novel view synthesis in dynamic urban scenarios." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.704, + 0.222, + 0.72 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.729, + 0.483, + 0.88 + ], + "angle": 0, + "content": "As autonomous driving systems increasingly shift toward end-to-end models, there is a growing need for scalable simulation environments without domain gaps where these systems can undergo closed-loop evaluations. A promising approach is real-world closed-loop evaluation, which demands controllable sensor inputs, thereby driving the development of advanced scene reconstruction techniques. In this context, Neural Radiance Fields (NeRFs) [10] and 3D Gaussian Splitting (3DGS) [7] have proven to be effective in creating high-quality 3D scene reconstructions with" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.275, + 0.905, + 0.499 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.501, + 0.907, + 0.543 + ], + "angle": 0, + "content": "Figure 1. Top: Street GS [24]; Middle: PVG [1]; Bottom: Our SplatFlow. SplatFlow eliminates the need for 3D Bboxes required by Street GS, and enhances rendering quality compared to PVG." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.549, + 0.907, + 0.609 + ], + "angle": 0, + "content": "excellent visual and geometric accuracy. However, accurately and comprehensively reconstructing dynamic driving scenes remains a major challenge, given the complexity of real-world scenarios without dynamic object annotations." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.612, + 0.907, + 0.869 + ], + "angle": 0, + "content": "Recent approaches advance NeRF-based representations by incorporating object detection and tracking to model dynamic elements, enabling photo-realistic view generation of dynamic urban street environments. Approaches [3, 8, 12, 18, 21, 23] create a scene graph, where both dynamic objects and the static background are represented as nodes, reconstructed within their canonical frames. Self-supervised methods [19, 25] model dynamic driving scenes by combining static neural field and time-dependent neural field to handle the static background and moving foreground objects separately. The SUDS [19] uses optical flow to ease the strict requirement for object labeling, while EmerNerf [25] learns temporal attributes and features in a self-supervision manner to avoid reliance on optical flow. However, these NeRF style methods face efficiency challenges in both training and inference, which becomes a major bottleneck for large-scale scene reconstruction and rendering." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.907, + 0.901 + ], + "angle": 0, + "content": "While NeRF has proven effective for driving scenes, 3DGS offers a promising alternative due to faster train" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.888, + 0.316, + 0.9 + ], + "angle": 0, + "content": "*Equally contributed as co-first author." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "27487" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.424 + ], + "angle": 0, + "content": "ing and rendering speed with more explicit representations. However, the original 3DGS faces notable challenges in modeling dynamic urban environments due to its limited representation capabilities in the temporal dimension. To address this, DrivingGaussian [29] introduces composite dynamic Gaussian graphs to handle multiple moving objects and incremental static Gaussians for background representation. StreetGaussian [24] optimizes the tracked bounding boxes of dynamic Gaussians along with 4D spherical harmonics to capture changing vehicle appearances. Both methods, however, require accurate object-level supervision, such as 3D object boxes and trackers, to decompose static and dynamic elements, limiting their scalability in real-world applications. In response, PVG [1] proposes a Periodic Vibration Gaussian mechanism to model dynamic driving scenes without relying on manually labeled 3D bounding boxes. However, the scene decomposition in PVG relies on time-dependent Gaussian attributes optimized by rendering losses, but ignores motion cues within the input point cloud used for Gaussian initialization, leading to suboptimal performance in challenging scenarios with rapid movements." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.427, + 0.483, + 0.789 + ], + "angle": 0, + "content": "To facilitate accurate scene reconstruction and real-time rendering without expensive annotations, we propose SplatFlow, a self-supervised Dynamic Gaussian Splatting method in Neural Motion Flow Fields (NMFF) for dynamic urban scenarios. In contrast to existing methods, as illustrated in Fig. 1, SplatFlow decomposes dynamic objects and the static background in a self-supervised way without requiring expensive 3D bounding box annotations. The key idea of SplatFlow is to seamlessly integrate 4D Gaussian representations within NMFF in a unified framework. The NMFF is a set of implicit functions to model temporal motions of both LiDAR points and Gaussians as continuous motion flow fields. Leveraging NMFF, SplatFlow not only enables the decomposition of dynamic and static elements from 3D LiDAR points, but also enables the status conversion of each 4D Gaussian across time. During Gaussian splatting, we represent dynamic objects using aggregated 4D Gaussians at various viewpoints and timestamps, and the static background using 3D Gaussians, respectively. In SplatFlow, NMFF's capabilities are enhanced through three components: 1) learning 3D motion priors by pretraining on 3D LiDAR data, 2) optimizing temporal status transitions of 4D Gaussians on image data, and 3) distilling knowledge from 2D foundation models." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.792, + 0.455, + 0.807 + ], + "angle": 0, + "content": "The novel features of SplatFlow are summarized as:" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.811, + 0.483, + 0.87 + ], + "angle": 0, + "content": "- SplatFlow introduces a unified framework that seamlessly integrates time-varying 4D Gaussian representations into NMFF, enabling self-supervised dynamic scene reconstruction and rendering." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.871, + 0.483, + 0.901 + ], + "angle": 0, + "content": "- NMFF enables scene decomposition, modeling static elements with 3D Gaussians and dynamic elements with 4D" + }, + { + "type": "list", + "bbox": [ + 0.092, + 0.811, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.528, + 0.093, + 0.602, + 0.105 + ], + "angle": 0, + "content": "Gaussians." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.107, + 0.905, + 0.151 + ], + "angle": 0, + "content": "- NMFF models correspondences of each 4D Gaussian over time, aggregating temporal features to enhance the cross-view consistency of dynamic components." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.153, + 0.905, + 0.197 + ], + "angle": 0, + "content": "- SplatFlow enhances the dynamic object identification by uplifting 2D features distilled from foundation models to 4D space-time through optical flow rendering." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.107, + 0.905, + 0.197 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.199, + 0.905, + 0.317 + ], + "angle": 0, + "content": "Comprehensive experiments on the Waymo [16] and KITTI [4] benchmarks demonstrate that SplatFlow outperforms the state-of-the-art (SOTA) methods in both image reconstruction and novel view synthesis for dynamic urban scenes. Notably, SplatFlow achieves this without relying on tracked 3D bounding boxes of dynamic objects, allowing the proposed model to learn from extensive, in-the-wild data sources." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.332, + 0.651, + 0.348 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.357, + 0.906, + 0.522 + ], + "angle": 0, + "content": "Real-world simulation is essential for generating data to support end-to-end autonomous driving solutions. However, current simulation engines like AirSim [15] and CARLA [2] encounter difficulties due to the expensive manual work needed to create virtual environments and the domain gap between real and simulated data. The rapid progress in Novel View Synthesis (NVS) technologies, such as NeRF [10] and 3DGS [7], and following work [6, 11, 17, 26] enables 3D reconstruction and photorealistic image generation, significantly improving real-world simulation for autonomous driving." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.524, + 0.906, + 0.75 + ], + "angle": 0, + "content": "For 3D reconstruction and neural rendering in autonomous driving scenes, sensor fusion solution combining surrounding cameras and LiDAR is commonly used in existing work [5, 9, 13, 14, 27]. Urban Radiance Field [14] improves NeRF training by incorporating 3D data from LiDAR to enhance 3D geometry learning. DNMP [9] employs a pre-trained deformable mesh primitive to represent the 3D scene, enhancing neural rendering quality. NPLF [13] uses explicit 3D reconstructions from LiDAR data to efficiently model the radiance field. TCLC-GS [27] design a hybrid explicit and implicit 3D representation derived from LiDAR-camera data, to enrich the properties of 3D Gaussians for splatting. Though these methods show impressive results for static urban scene rendering, they struggle with dynamic objects common in driving scenarios." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.906, + 0.901 + ], + "angle": 0, + "content": "To model dynamic urban scenes, recent approaches [8, 12, 19, 21, 25] decompose the entire scene into static background and dynamic objects, learning their representations separately. PNF [8] uses monocular 3D bounding box predictions to isolate dynamic objects and further jointly optimizes their poses during the reconstruction process. NSG [12] uses neural graphs to represent entire scenes, decomposing dynamic multi-object environments. MARS [21] employs distinct sub-networks to model backgrounds and dynamic objects, creating an instance-aware" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "27488" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.096, + 0.1, + 0.885, + 0.314 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.389, + 0.315, + 0.608, + 0.33 + ], + "angle": 0, + "content": "Figure 2. The pipeline of SplatFlow." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.334, + 0.486, + 0.896 + ], + "angle": 0, + "content": "simulation framework. These methods typically rely on manually annotated or predicted 3D bounding boxes. In order to avoid the need of 3D bounding boxes, SUDS [19] introduces a scalable hash table to represent large-scale dynamic urban scenes, using an off-the-shelf 2D optical flow estimator to track dynamic objects. EmerNerf [25] addresses this challenge by learning scene flow to associate points across time in the 4D neural radiance field, allowing the separation of static and dynamic objects without 3D bounding boxes. However, these methods, which rely on implicit representations, still suffer from inefficiency in both reconstruction and rendering. Recently, 3DGS [7] introduced a novel explicit 3D scene representation, combining high-quality volume rendering with a fast speed. However, 3DGS is designed for static scenes and fails when modeling dynamic moving objects. To address this, DrivingGaussian [29] and StreetGaussian [24] partition the driving scene into static and dynamic components by separate sets of Gaussians based on the 3D bounding boxes of vehicles. DrivingGaussian [29] uses a composite dynamic Gaussian graph to manage multiple moving objects while incrementally employing incremental static 3D Gaussians to represent the static background. StreetGaussian [24] models foreground vehicles with estimizable poses and a 4D spherical harmonics appearance model, and background with 3D Gaussians separately. HUGS [28] also adopts 3D object boxes to identify dynamic elements, enabling optimization of geometry, appearance, semantics, and motion attributes for dynamic Gaussians. However, all these methods rely on expensive annotated or predicted 3D bounding boxes to learn the time-dependent representations of dynamic objects. Most recent method PVG [1] introduces periodic vibration based Gaussian attributes, optimized through self-supervision without the need of 3D bounding boxes. Each Gaussian models dynamic changes over time through estimizable attributes including vibration directions, life span, and life peak." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.333, + 0.651, + 0.351 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.359, + 0.627, + 0.374 + ], + "angle": 0, + "content": "3.1. Overview" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.382, + 0.909, + 0.716 + ], + "angle": 0, + "content": "By collecting images and point clouds with timestamps from surrounding cameras and LiDAR, we aim to learn a space-time 4D Gaussian representation of a dynamic scene without any human annotations, enabling fast and high-quality novel view rendering. As the pipeline of SplatFlow shown in Fig. 2, NMFF models the temporal motions of both 3D points and Gaussians as continuous motion flow fields. It serves two key roles in SplatFlow: 1) decomposing static and dynamic elements from 3D LiDAR points; 2) modeling the temporal correspondences of each 4D Gaussian across time. We first decompose the time-series LiDAR points into static and dynamic points by NMFF, which are then separately merged to initialize the static and dynamic Gaussians. Dynamic objects are represented by aggregated 4D Gaussians associated with NMFF over time, while the static background is modeled using 3D Gaussians. NMFF learns the correspondence of each 4D Gaussian in conjunction with optimizing 4D Gaussians' attributes along time. We also distill optical flow knowledge from a 2D foundation model into 4D space-time representation. Finally, images, depths and optical flows are rendered from novel viewpoints at different timesteps in dynamic driving scenes." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.727, + 0.704, + 0.743 + ], + "angle": 0, + "content": "3.2. Problem Definition:" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.909, + 0.903 + ], + "angle": 0, + "content": "In a dynamic urban scenario, we collect time-sequential data from a vehicle equipped with surrounding cameras and a LiDAR sensor. A set of surrounding images \\(\\mathcal{I}\\) is captured by multiple surrounding cameras with corresponding intrinsic matrices \\(I\\) and extrinsic matrices \\(E\\). Meanwhile, a set of 3D points \\(\\mathcal{P}\\) is obtained from LiDAR, along with corresponding extrinsic matrices \\(E'\\). We denote the synchronized and calibrated multi-sensor data sequence, including timestamps \\(t\\), as \\(\\{\\mathcal{I}_i, \\mathcal{P}_i, E_i, I_i, E_i', t_i | i = 1, 2, 3, \\ldots, n\\}\\), where \\(\\mathcal{P}_i = \\{x_i, y_i, z_i\\}\\) and \\(n\\) is the number of frames." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "27489" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.09, + 0.904, + 0.239 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.265, + 0.239, + 0.731, + 0.253 + ], + "angle": 0, + "content": "Figure 3. Visualization of 3D LiDAR points within NMFF on Waymo dataset." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.259, + 0.484, + 0.336 + ], + "angle": 0, + "content": "The vehicle trajectory is either provided or estimated using multi-sensor-based odometry. Our model \\( \\mathcal{SF} \\) aims to perform accurate 3D reconstruction and synthesize novel viewpoints for any given timestamp \\( t \\) and camera pose \\( [E_t, I_t] \\) by rendering \\( \\hat{\\mathcal{I}} = \\mathcal{SF}(E_t, I_t, t) \\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.341, + 0.347, + 0.358 + ], + "angle": 0, + "content": "3.3. 4D Gaussian Representation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.363, + 0.484, + 0.499 + ], + "angle": 0, + "content": "Given images with associated camera poses, 3DGS optimizes a set of anisotropic 3D Gaussians through differentiable rasterization to represent a static 3D scene. We extend 3DGS to 4DGS of spatial-temporal representations for dynamic objects in an urban scene. Each 4D Gaussian primitives \\(\\mathcal{G}(t)\\) is represented by time-varying attributes: 3D center \\(\\mu(t) = [x(t), y(t), z(t)]^T\\) and covariance matrix \\(\\Sigma(t)\\) along with time-invariant attributes: opacity \\(\\sigma\\) and color \\(c\\). The density of a Gaussian at point \\((x,t)\\) is defined as," + }, + { + "type": "equation", + "bbox": [ + 0.107, + 0.499, + 0.482, + 0.525 + ], + "angle": 0, + "content": "\\[\n\\alpha (t) = \\sigma \\cdot e x p \\left(- \\frac {1}{2} (x - \\mu (t)) ^ {T} \\Sigma (t) ^ {- 1} (x - \\mu (t))\\right). \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.524, + 0.483, + 0.585 + ], + "angle": 0, + "content": "The covariance matrix \\(\\Sigma(t) = R(t)SS^T R(t)^T\\) is composed of a scaling matrix \\(S = \\text{diag}(s_x, s_y, s_z)\\) and rotation matrix \\(R(t) = (q_x(t), q_y(t), q_z(t), q_w(t))\\), constrained to be a positive semi-definite matrix during optimization." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.586, + 0.484, + 0.69 + ], + "angle": 0, + "content": "To render an image from a specific viewpoint at times-tamp \\( t \\), we first transform all the 4D Gaussians from other times to the target time \\( t \\), according to the learned correspondence from NMFF. Subsequently, the aggregated 4D Gaussians are splatted onto the image plane, resulting in a collection of 2D Gaussians. The 3D covariance matrix \\( \\Sigma \\) is projected to a 2D covariance matrix \\( \\Sigma' \\) by," + }, + { + "type": "equation", + "bbox": [ + 0.205, + 0.69, + 0.482, + 0.71 + ], + "angle": 0, + "content": "\\[\n\\Sigma^ {\\prime} (t) = J E \\Sigma (t) E ^ {T} J ^ {T}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.712, + 0.483, + 0.742 + ], + "angle": 0, + "content": "where \\( E \\) refers to the world-to-camera matrix and \\( J \\) refers to Jacobian of the perspective transformation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.742, + 0.484, + 0.787 + ], + "angle": 0, + "content": "By sorting the Gaussians according to their depth within the camera space, we use \\(\\alpha\\)-blending to estimate the color \\(C\\) and depth \\(D\\) at each pixel \\(p\\) as," + }, + { + "type": "equation", + "bbox": [ + 0.099, + 0.786, + 0.483, + 0.827 + ], + "angle": 0, + "content": "\\[\nC = \\sum_ {i = 1} ^ {N} c _ {i} \\alpha_ {i} \\prod_ {j = 1} ^ {i - 1} (1 - \\alpha_ {j}), D = \\sum_ {i = 1} ^ {N} z _ {i} \\alpha_ {i} \\prod_ {j = 1} ^ {i - 1} (1 - \\alpha_ {j}). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.827, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Here \\( c_{i} \\) represents the color of Gaussian \\( \\mathcal{G}_{i} \\), which is computed using spherical harmonic coefficients. \\( z_{i} \\) represents the distance from the image plane to the \\( i \\)-th Gaussian center. \\( \\alpha_{i} \\) denotes the density computed from the 2D projection of the \\( i \\)-th Gaussian and its learned opacity." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.258, + 0.75, + 0.273 + ], + "angle": 0, + "content": "3.4. Neural Motion Flow Field" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.281, + 0.907, + 0.387 + ], + "angle": 0, + "content": "To establish the correspondence of dynamic objects, we design NMFF to model 3D motions as a continuous motion flow field, enabling the temporal transition of both 3D points and Gaussians. The NMFF \\(\\Phi\\) contains a temporal sequence of motion flow field, defined as \\(\\Phi = \\{\\phi_{t_i:t_{i + 1}}\\} ,i = 0,1,2,\\ldots ,n - 1\\) , which predicts 3D motion flow of an arbitrary query point between two consecutive frames as," + }, + { + "type": "equation", + "bbox": [ + 0.522, + 0.391, + 0.905, + 0.422 + ], + "angle": 0, + "content": "\\[\n\\phi_ {t _ {1}: t _ {2}} \\left(x _ {t _ {1}}, y _ {t _ {1}}, z _ {t _ {1}}\\right) = \\Delta x _ {t _ {1}: t _ {2}}, \\Delta y _ {t _ {1}: t _ {2}}, \\Delta z _ {t _ {1}: t _ {2}}, \\Delta R _ {t _ {1}: t _ {2}} \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.422, + 0.905, + 0.451 + ], + "angle": 0, + "content": "where \\(\\Delta R_{t_1:t_2}\\) denotes motion angle between two adjacent timestamps. Each field \\(\\phi\\) is eight ReLU-MLP stacks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.453, + 0.906, + 0.558 + ], + "angle": 0, + "content": "Point Cloud within NMFF: Although NMFF can be learned during Gaussian Splatting optimization, we observed that the rendering loss alone is insufficient to effectively constrain the motion representation of dynamic scene components. Therefore, we leverage the rich geometric structure inherent in temporally consecutive 3D LiDAR points to derive a robust NMFF prior, as shown in Fig. 3." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.559, + 0.906, + 0.664 + ], + "angle": 0, + "content": "We first pre-train \\(\\Phi\\) on a sequence of 3D LiDAR points by forward and backward 3D geometry consistency, enabling NMFF to learn the 3D motion of each query 3D point over time. Given the source point clouds \\(\\mathcal{P}\\) and target point clouds \\(\\mathcal{Q}\\) from different timestamps, we minimize the point distance between the source and target point clouds by a bidirectional Chamfer Distance \\((CD)\\)," + }, + { + "type": "equation", + "bbox": [ + 0.521, + 0.668, + 0.905, + 0.701 + ], + "angle": 0, + "content": "\\[\nd _ {\\mathrm {C D}} (\\mathcal {P}, \\mathcal {Q}) = \\sum_ {p \\in \\mathcal {P}} \\min _ {q \\in \\mathcal {Q}} \\| p - q \\| ^ {2} + \\sum_ {q \\in \\mathcal {Q}} \\min _ {p \\in \\mathcal {P}} \\| q - p \\| ^ {2}. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.906, + 0.764 + ], + "angle": 0, + "content": "The bidirectional \\(CD\\) is a point-based distance that establishes correspondences in both directions—source to target and target to source—by searching for the nearest neighbors in each point cloud." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.766, + 0.907, + 0.901 + ], + "angle": 0, + "content": "We compensate for ego-car motion in LiDAR points to let NMFF learn only 3D motion flow during pre-training. As a result, the 3D motion flow predicted by NMFF excludes the ego-car motion. We generate a 3D dynamic mask to separate static and dynamic points by applying a threshold on each point's 3D motion flow. We then merge static point clouds using ego-motion alone, and dynamic point clouds using both ego-motion and motion flow. For the static background, we initialize the static 3D Gaus" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "27490" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.089, + 0.092, + 0.482, + 0.181 + ], + "angle": 0, + "content": "sians using the merged 3D static points. For dynamic objects, we initialize 4D Gaussians by aggregating points from different timestamps into a common reference frame (the mid-timestamp reference frame), which increases density and completeness by integrating observations from various viewpoints over time." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.182, + 0.483, + 0.272 + ], + "angle": 0, + "content": "4D Gaussian within NMFF: The pre-trained \\(\\Phi\\) is subsequently integrated with 4D Gaussian representations of dynamic objects, enabling joint optimization during Gaussian splatting. Given consecutive time \\(t_1\\) and \\(t_2\\) where \\(t_1 < t_2\\) with respective states \\(\\{\\mathcal{G}_i(t_1)\\}\\) and \\(\\{\\mathcal{G}_i(t_2)\\}\\), these states are connected by NMFF for each Gaussian as," + }, + { + "type": "equation", + "bbox": [ + 0.177, + 0.273, + 0.482, + 0.29 + ], + "angle": 0, + "content": "\\[\n\\mathcal {G} _ {i} \\left(t _ {1}\\right) = \\left\\{\\mu \\left(t _ {1}\\right), R \\left(t _ {1}\\right), S, \\alpha , c \\right\\}, \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.172, + 0.293, + 0.483, + 0.31 + ], + "angle": 0, + "content": "\\[\n\\Delta \\mu_ {t _ {1}: t _ {2}}, \\Delta R _ {t _ {1}: t _ {2}} = \\phi_ {t _ {1}: t _ {2}} (\\mu (t _ {1})), \\tag {7}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.315, + 0.483, + 0.333 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathcal {G}} _ {i} \\left(t _ {2}\\right) = \\left\\{\\mu \\left(t _ {1}\\right) + \\Delta \\mu_ {t _ {1}: t _ {2}}, R \\left(t _ {1}\\right) \\cdot \\Delta R _ {t _ {1}: t _ {2}}, S, \\alpha , c \\right\\}. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.336, + 0.483, + 0.486 + ], + "angle": 0, + "content": "We employ NMFF \\(\\Phi\\) correspondence to warp the center of all aggregated 4D Gaussians across time to the desired timestamp, as \\(\\{\\mathcal{G}_i(t)\\} = \\{\\mathcal{G}(t),\\hat{\\mathcal{G}} (t_1),\\hat{\\mathcal{G}} (t_2),\\dots ,\\hat{\\mathcal{G}} (t_n)\\}\\). For the transition of Gaussians across multiple timestamps, we apply the corresponding NMFF through step-by-step propagation. In practice, we propagate the aggregated 4D Gaussian at the mid-timestamp to the target timestamp. Here NMFF learns a motion transition of the 4D Gaussian among training frames in a self-supervised manner, promoting a multi-view consistent representation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.487, + 0.483, + 0.58 + ], + "angle": 0, + "content": "For static background, we represent it by static 3D Gaussians \\(\\mathcal{G}_{\\text{static}}\\) following the strategy [1]. For sky area, we adopt a separate environmental map with static sky Gaussians \\(\\mathcal{G}_{sky}\\) following [1, 25]. Given a specific viewpoint \\([E_t, I_t, t]\\), the image \\(\\hat{\\mathcal{I}}\\) and depth \\(\\hat{\\mathcal{D}}\\) at time \\(t\\) are obtained by the differentiable rendering using a set of Gaussians," + }, + { + "type": "equation", + "bbox": [ + 0.115, + 0.585, + 0.483, + 0.603 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathcal {L}}, \\hat {\\mathcal {D}} = \\operatorname {R e n d e r} \\left(\\left\\{\\mathcal {G} _ {i} (t) \\right\\}, \\mathcal {G} _ {\\text {s t a t i c}}, \\mathcal {G} _ {\\text {s k y}} \\mid E _ {t}, I _ {t}, t\\right). \\tag {9}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.612, + 0.317, + 0.628 + ], + "angle": 0, + "content": "3.5. Optical Flow Distillation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.634, + 0.483, + 0.709 + ], + "angle": 0, + "content": "The 3D motion flow provided by NMFF enables 2D optical flow rendering. Given two consecutive time \\( t_1 \\) and \\( t_2 \\), the optical flow \\( f_{t_1:t_2} \\) of each Gaussian is computed by projecting 3D center \\( \\mu \\) to the image plane using the camera's intrinsic matrix \\( I \\) and extrinsic matrix \\( E \\)," + }, + { + "type": "equation", + "bbox": [ + 0.13, + 0.712, + 0.482, + 0.73 + ], + "angle": 0, + "content": "\\[\n\\mu^ {\\prime} \\left(t _ {1}\\right) = I [ E ] \\mu \\left(t _ {1}\\right), \\quad \\mu^ {\\prime} \\left(t _ {2}\\right) = I [ E ] \\mu \\left(t _ {2}\\right), \\tag {10}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.204, + 0.732, + 0.482, + 0.749 + ], + "angle": 0, + "content": "\\[\nf _ {t _ {1}: t _ {2}} = \\mu^ {\\prime} (t _ {2}) - \\mu^ {\\prime} (t _ {1}). \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.752, + 0.483, + 0.781 + ], + "angle": 0, + "content": "Then we estimate the optical flow \\(F\\) at pixel \\(p\\) via by point-based \\(\\alpha\\)-blending as," + }, + { + "type": "equation", + "bbox": [ + 0.195, + 0.781, + 0.482, + 0.824 + ], + "angle": 0, + "content": "\\[\nF = \\sum_ {i = 1} ^ {N} f _ {i} \\alpha_ {i} \\prod_ {j = 1} ^ {i - 1} (1 - \\alpha_ {j}). \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.483, + 0.901 + ], + "angle": 0, + "content": "The optical flow \\(\\hat{\\mathcal{F}}\\) at time \\(t\\) are rendered by the differentiable rendering similar as Equation 9. In order to distill optical flow knowledge from 2D foundation model into 4D space, we adopt SEA-RAFT[20] to extract optical flow \\(\\mathcal{F}\\) as pseudo ground truth for distillation by optical flow loss," + }, + { + "type": "equation", + "bbox": [ + 0.525, + 0.09, + 0.905, + 0.109 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {F} = \\lambda_ {f} \\mathcal {L} _ {1} (\\mathcal {F}, \\hat {\\mathcal {F}}) + (1 - \\lambda_ {f}) \\mathcal {L} _ {1} \\left(\\mathcal {I} _ {\\text {n e x t}}, \\mathcal {T} (\\hat {\\mathcal {I}} | \\hat {\\mathcal {F}})\\right), \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.113, + 0.905, + 0.189 + ], + "angle": 0, + "content": "where \\(\\lambda_{f}\\) is a scale factor, and \\(\\mathcal{I}_{\\text{next}}\\) denotes the image at the next time step. \\(\\mathcal{T}\\) is used to warp rendered image to next time step according to optical flow. The optical flow \\(\\mathcal{F}\\) extracted from foundation model is only required during training and is not needed during inference." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.197, + 0.653, + 0.213 + ], + "angle": 0, + "content": "3.6. Optimization" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.219, + 0.907, + 0.295 + ], + "angle": 0, + "content": "All Gaussian attributes, along with the parameters of NMFF, are optimized end-to-end in a self-supervised manner. Meanwhile, adaptive densification and pruning strategies are introduced to enhance the fitting of the 3D scene. The overall training loss is given as," + }, + { + "type": "equation", + "bbox": [ + 0.538, + 0.299, + 0.905, + 0.317 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {I} + \\lambda_ {1} \\mathcal {L} _ {D} + \\lambda_ {2} \\mathcal {L} _ {F} + \\lambda_ {3} \\mathcal {L} _ {s k y} + \\lambda_ {4} \\mathcal {L} _ {r e g}, \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.321, + 0.905, + 0.367 + ], + "angle": 0, + "content": "where \\(\\lambda_{1,2,3,4}\\) are scale factors and \\(\\mathcal{L}_{reg}\\) is the regularization term. The image loss \\(\\mathcal{L}_I\\) combines L1 and SSIM losses between rendered and observed images," + }, + { + "type": "equation", + "bbox": [ + 0.532, + 0.37, + 0.905, + 0.389 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {I} = \\left(1 - \\lambda_ {s s i m}\\right) \\mathcal {L} _ {1} (\\mathcal {I}, \\hat {\\mathcal {I}}) + \\lambda_ {s s i m} \\mathcal {L} _ {s s i m} (\\mathcal {I}, \\hat {\\mathcal {I}}), \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.394, + 0.905, + 0.439 + ], + "angle": 0, + "content": "where \\(\\lambda_{ssim}\\) is a scale factor. The depth loss \\(\\mathcal{L}_D\\) is a L1 loss between inverse of rendered depth and generated depth by projecting sparse LiDAR points onto camera plane as," + }, + { + "type": "equation", + "bbox": [ + 0.651, + 0.443, + 0.905, + 0.46 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {D} = \\mathcal {L} _ {1} (\\mathcal {D}, \\hat {\\mathcal {D}}). \\tag {16}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.465, + 0.905, + 0.497 + ], + "angle": 0, + "content": "The sky opacity loss \\(\\mathcal{L}_{sky}\\) is a binary cross entropy loss for sky refining using sky mask \\(M_{sky}\\) from SegFormer [22]," + }, + { + "type": "equation", + "bbox": [ + 0.614, + 0.502, + 0.905, + 0.519 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {s k y} = \\mathcal {L} _ {b c e} (O, 1 - M _ {s k y}), \\tag {17}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.524, + 0.905, + 0.571 + ], + "angle": 0, + "content": "where the accumulated opacity denotes as \\( O = \\sum_{i=1}^{N} \\alpha_i \\prod_{j=1}^{i-1} (1 - \\alpha_j) \\). Note the sky mask is only required during training and is not needed during inference." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.583, + 0.646, + 0.6 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.607, + 0.792, + 0.622 + ], + "angle": 0, + "content": "4.1. Datasets, Metrics and Baselines" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.629, + 0.907, + 0.749 + ], + "angle": 0, + "content": "Following PVG [1], we validate our approach on two widely-used datasets in autonomous driving: Waymo Open Dataset [16] and KITTI Dataset [4]. Both offer multi-sensor data clips including synchronized and calibrated camera and LiDAR data from urban driving scenarios. We train and test our methods, along with all baselines, using full-resolution images: \\( 1920 \\times 1280 \\) for the Waymo dataset and \\( 1242 \\times 375 \\) for the KITTI dataset." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.905, + 0.826 + ], + "angle": 0, + "content": "Following prior research [1, 19, 24, 25, 29], we evaluate image reconstruction and novel view synthesis using three widely-adopted benchmark metrics: peak signal-to-noise ratio (PSNR), structural similarity index measure (SSIM), and learned perceptual image patch similarity (LPIPS)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.905, + 0.901 + ], + "angle": 0, + "content": "We compare our method with extensive baselines including S-NeRF [23], StreetSurf [5], 3DGS [7], NSG [12], Mars [21], SUDS [19], EmerNerf [25], PVG [1], Street-Gaussian [24] on the Waymo and KITTI benchmarks. See supplementary material for more details and results." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "27491" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.093, + 0.088, + 0.885, + 0.593 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.119, + 0.594, + 0.877, + 0.608 + ], + "angle": 0, + "content": "Figure 4. Visual comparison of novel view synthesis on Waymo dataset. Bounding boxes indicate the zoomed-in dynamic areas." + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.613, + 0.482, + 0.766 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.092, + 0.767, + 0.48, + 0.78 + ], + "angle": 0, + "content": "Figure 5. Dynamic object decomposition comparison on Waymo." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.785, + 0.407, + 0.801 + ], + "angle": 0, + "content": "4.2. Evaluation on Waymo Open Dataset" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Following PVG [1], we evaluated our SplatFlow against baselines using the Waymo Open dataset. Table 1 summarizes the average metrics for both image reconstruction and novel view synthesis tasks on selected dynamic scenes. SplatFlow consistently outperforms all baselines across all evaluated metrics for both tasks. In image reconstruction," + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.613, + 0.922, + 0.742 + ], + "angle": 0, + "content": "
Image ReconstructionNovel View Synthesis
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
S-NeRF [23]19.670.5280.38719.220.5150.400
StreetSurf [5]26.700.8460.38723.780.8220.401
3DGS [7]27.990.8660.37225.080.8220.319
NSG [12]24.080.6560.29321.010.5710.487
Mars [21]21.810.6810.44120.690.6360.453
SUDS [19]28.830.8050.43021.830.6560.405
EmerNerf [25]28.110.7860.28925.920.7630.384
PVG [1]32.460.9100.37328.110.8490.279
SplatFlow33.640.9510.19828.710.8740.239
" + }, + { + "type": "table_caption", + "bbox": [ + 0.55, + 0.745, + 0.868, + 0.758 + ], + "angle": 0, + "content": "Table 1. Performance comparison on Waymo dataset." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.907, + 0.901 + ], + "angle": 0, + "content": "SplatFlow achieves high image quality with the highest scores: PSNR at 33.64, SSIM at 0.951, and LPIPS at 0.198. For novel view synthesis, SplatFlow produces high-quality renderings of unseen timestamps, reaching PSNR of 28.71, SSIM of 0.874 and LPIPS of 0.239. These improvements are visually validated in Fig. 4 and Fig. 9 (dynamic only), showing exceptional clarity in details for both static and dynamic regions. The dynamic objects in Fig. 9 are selected from a distant background car in an extremely zoomed-in" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "27492" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.088, + 0.907, + 0.391 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.123, + 0.393, + 0.873, + 0.406 + ], + "angle": 0, + "content": "Figure 6. Visual comparison of novel view synthesis on KITTI dataset. Bounding boxes indicate the zoomed-in dynamic areas." + }, + { + "type": "table", + "bbox": [ + 0.104, + 0.409, + 0.467, + 0.552 + ], + "angle": 0, + "content": "
PVG [1]SplatFlow
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
Seg. 1058...26.860.8400.27327.460.8800.253
Seg. 2259...26.450.8240.30929.300.8560.290
Seg. 7670...26.580.8200.30728.890.8550.283
Seg. 5083...28.350.8930.21330.290.9280.169
PSNR*↑SSIM*↑LPIPS*↓PSNR*↑SSIM*↑LPIPS*↓
Seg. 1058...27.420.8350.26728.500.8770.234
Seg. 2259...22.550.8030.35031.610.8990.191
Seg. 7670...23.560.7560.33028.760.8200.289
Seg. 5083...25.580.8140.29026.710.8550.211
" + }, + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.553, + 0.483, + 0.581 + ], + "angle": 0, + "content": "Table 2. Novel view synthesis results on Waymo dataset (*denotes dynamic elements only)." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.586, + 0.5, + 0.714 + ], + "angle": 0, + "content": "
Image ReconstructionNovel View Synthesis
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
S-NeRF [23]19.230.6640.19318.710.6060.352
StreetSurf [5]24.140.8190.25722.480.7630.304
3DGS [7]21.020.8110.20219.540.7760.224
NSG [12]26.660.8060.18621.530.6730.254
Mars [21]27.960.9000.18524.230.8450.160
SUDS [19]28.310.8760.18522.770.7970.171
EmerNeRF [25]26.950.8280.21825.240.8010.237
PVG [1]32.830.9370.07027.430.8960.114
SplatFlow33.370.9430.05728.320.9320.089
" + }, + { + "type": "table_caption", + "bbox": [ + 0.131, + 0.716, + 0.442, + 0.73 + ], + "angle": 0, + "content": "Table 3. Performance comparison on KITTI dataset." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.734, + 0.483, + 0.794 + ], + "angle": 0, + "content": "view. Compared with baselines which show artifacts such as ghosting and blur in regions of dynamic elements, our method preserves high-quality textures and fine details, producing more accurate renderings at novel viewpoints." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.483, + 0.901 + ], + "angle": 0, + "content": "To further evaluate the rendering quality of dynamic objects against the primary baseline PVG, we selected additional four scenes featuring a higher presence of dynamic objects for further comparison. Table 2 displays the metrics for novel view synthesis, evaluated both on entire scenes and exclusively on dynamic objects (denoted as *). Splat-Flow consistently surpasses PVG in all selected scenes," + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.409, + 0.906, + 0.651 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.653, + 0.907, + 0.679 + ], + "angle": 0, + "content": "Figure 7. Visual comparison of novel view synthesis on KITTI-25% (row1), 50% (row2), and 75% (row3) dataset." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.68, + 0.907, + 0.846 + ], + "angle": 0, + "content": "demonstrating higher PSNR and SSIM values and lower LPIPS scores, underscoring its effectiveness in handling complex, dynamic content. The dynamic objects masks are obtained from 2D ground truth bounding boxes following [25]. The PSNR* of dynamic regions on novel view synthesis significantly outperforms SVG both quantitatively and qualitatively, especially in scenes with fast ego-motion and moving objects in Seg. 2259 and Seg. 7670. We also present the visual comparison of dynamic object decomposition with SVG in Fig. 5, which illustrates that SplatFlow renders sharper and clearer image of dynamic objects." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.851, + 0.776, + 0.866 + ], + "angle": 0, + "content": "4.3. Evaluation on KITTI Dataset" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Following PVG [1], we further evaluated our SplatFlow method in comparison to the baselines on KITTI dataset." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "27493" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.111, + 0.089, + 0.598, + 0.203 + ], + "angle": 0, + "content": "
KITTI - 75%KITTI - 50%KITTI - 25%
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
NeRF [10]18.560.5570.55419.120.5870.49718.610.5700.510
NSG [12]21.530.6730.25421.260.6590.26620.000.6320.281
SUDS [19]22.770.7970.17123.120.8210.13520.760.7470.198
MARS [21]24.230.8450.16024.000.8010.16423.230.7560.177
3DGS [7]19.190.7370.17219.230.7390.17419.060.7300.180
PVG [1]27.430.8960.11425.920.8820.11422.550.8330.151
StreetGS [24]25.790.8440.08125.520.8410.08424.530.8240.090
SplatFlow28.320.9320.08927.900.9270.09326.100.8900.120
" + }, + { + "type": "table_caption", + "bbox": [ + 0.108, + 0.207, + 0.598, + 0.233 + ], + "angle": 0, + "content": "Table 4. Performance comparison of novel view synthesis on KITTI-75%, 50% and 25% dataset." + }, + { + "type": "table", + "bbox": [ + 0.599, + 0.096, + 0.887, + 0.199 + ], + "angle": 0, + "content": "
PSNR↑SSIM↑LPIPS↓
w/o NMFF prior27.690.8630.282
w/o NMFF optimization28.140.8740.269
w/o optical flow distillation28.280.8770.252
Full28.990.8800.249
PSNR*↑SSIM*↑LPIPS*↓
w/o NMFF prior27.280.8270.317
w/o NMFF optimization27.970.8430.254
w/o optical flow distillation28.510.8610.232
Full28.900.8630.231
" + }, + { + "type": "table_caption", + "bbox": [ + 0.598, + 0.2, + 0.884, + 0.227 + ], + "angle": 0, + "content": "Table 5. Ablation study on Waymo dataset (*denotes dynamic elements only)." + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.238, + 0.288, + 0.339 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.155, + 0.339, + 0.22, + 0.348 + ], + "angle": 0, + "content": "Ours (Full)" + }, + { + "type": "image", + "bbox": [ + 0.288, + 0.238, + 0.482, + 0.339 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.351, + 0.339, + 0.426, + 0.349 + ], + "angle": 0, + "content": "w/o Opt. flow" + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.349, + 0.288, + 0.448 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.138, + 0.45, + 0.236, + 0.461 + ], + "angle": 0, + "content": "w/o NMFF optim." + }, + { + "type": "image", + "bbox": [ + 0.288, + 0.349, + 0.482, + 0.448 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.344, + 0.45, + 0.434, + 0.461 + ], + "angle": 0, + "content": "w/o NMFF prior" + }, + { + "type": "image_caption", + "bbox": [ + 0.092, + 0.463, + 0.481, + 0.476 + ], + "angle": 0, + "content": "Figure 8. Visual comparison of ablation study on Waymo dataset." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.49, + 0.483, + 0.67 + ], + "angle": 0, + "content": "Table 3 presents the average metrics for both image reconstruction and novel view synthesis tasks on selected dynamic scenes characterized by extensive movement. SplatFlow outperforms all other methods across all metrics in both tasks, with particularly high scores in PSNR. SSIM and LPIPS, achieving 33.37, 0.943 and 0.057 for image reconstruction and 28.32, 0.932 and 0.089 for novel view synthesis. SplatFlow achieves exceptional results in novel view synthesis, as shown in Fig. 6. Our approach generates high-fidelity renderings that preserve fine details while accurately reconstructing dynamic areas where existing approaches struggle to maintain stability." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.671, + 0.483, + 0.822 + ], + "angle": 0, + "content": "To evaluate robustness to training data size, we follow the settings of [21, 24] to evaluate our method using different train/test split configurations on the KITTI dataset. Table 4 and Fig. 7 provide the performance comparisons and visualizations. SplatFlow consistently outperforms all other methods across all dataset subsets, achieving the highest scores. Even with reduced training data, SplatFlow robustly delivers high-quality novel view synthesis across varying levels of data availability, highlighting its efficient use of available information and strong generalization capabilities." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.833, + 0.245, + 0.849 + ], + "angle": 0, + "content": "4.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.483, + 0.901 + ], + "angle": 0, + "content": "To demonstrate the effectiveness of each component in SplatFlow, we performed ablation studies on the Waymo dataset. Our main contributions, NMFF prior, NMFF opti" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.238, + 0.614, + 0.475 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.536, + 0.475, + 0.599, + 0.485 + ], + "angle": 0, + "content": "EmerNeRF" + }, + { + "type": "image", + "bbox": [ + 0.615, + 0.238, + 0.71, + 0.475 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.648, + 0.475, + 0.676, + 0.485 + ], + "angle": 0, + "content": "PVG" + }, + { + "type": "image", + "bbox": [ + 0.712, + 0.238, + 0.807, + 0.475 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.73, + 0.475, + 0.788, + 0.485 + ], + "angle": 0, + "content": "SplatFlow" + }, + { + "type": "image", + "bbox": [ + 0.808, + 0.238, + 0.904, + 0.475 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.843, + 0.475, + 0.869, + 0.485 + ], + "angle": 0, + "content": "G.T." + }, + { + "type": "image_caption", + "bbox": [ + 0.518, + 0.487, + 0.9, + 0.502 + ], + "angle": 0, + "content": "Figure 9. Detail comparison of novel view synthesis on Waymo." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.513, + 0.907, + 0.68 + ], + "angle": 0, + "content": "mization, and optical flow distillation are analyzed in this study to assess their individual impact. We trained four different variations: 1) SplatFlow without pretraining on LiDAR data; 2) SplatFlow without optimization with 4D Gaussians on image data; 3) SplatFlow without optical flow distillation from foundational model; 4) SplatFlow full method. Table 5 presents the average evaluation metrics across test scenes used in Table 2, showing results for both entire scenes and dynamic objects (indicated as *). Ablation study validates the contribution of each component to the overall performance." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.695, + 0.634, + 0.71 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In this paper, we propose SplatFlow, a novel self-supervised Dynamic Gaussian Splitting within Neural Motion Flow Fields (NMFF) for accurate reconstruction and real-time rendering in dynamic urban scenarios. The core idea of SplatFlow is to seamlessly integrate time-dependent 4D Gaussian representations within NMFF in a unified framework, where NMFF implicitly models the motions of dynamic components across time through self-supervision. Experimental evaluations show that SplatFlow outperforms state-of-the-art across all standard metrics on the Waymo Open and KITTI datasets, without requiring expensive annotations of dynamic object detection and tracking." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "27494" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.251, + 0.108 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.484, + 0.238 + ], + "angle": 0, + "content": "This work was partially supported by the intramural research program of the U.S. Department of Agriculture Natural Resources Conservation Service, Climate-Smart Commodities program, EFFICACI grant, #NR233A750004G044. The findings and conclusions in this work have not been formally disseminated by the U. S. Department of Agriculture and should not be construed to represent any agency determination or policy." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.248, + 0.188, + 0.264 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.273, + 0.484, + 0.328 + ], + "angle": 0, + "content": "[1] Yurui Chen, Chun Gu, Junzhe Jiang, Xiatian Zhu, and Li Zhang. Periodic vibration gaussian: Dynamic urban scene reconstruction and real-time rendering. arXiv preprint arXiv:2311.18561, 20.1, 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.33, + 0.484, + 0.384 + ], + "angle": 0, + "content": "[2] Alexey Dosovitskiy, German Ros, Felipe Codevilla, Antonio Lopez, and Vladlen Koltun. Carla: An open urban driving simulator. In Conference on robot learning, pages 1-16. PMLR, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.385, + 0.484, + 0.455 + ], + "angle": 0, + "content": "[3] Tobias Fischer, Lorenzo Porzi, Samuel Rota Bulo, Marc Pollefeys, and Peter Kontschieder. Multi-level neural scene graphs for dynamic urban environments. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21125-21135, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.456, + 0.483, + 0.511 + ], + "angle": 0, + "content": "[4] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In 2012 IEEE conference on computer vision and pattern recognition, pages 3354-3361. IEEE, 2012. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.511, + 0.484, + 0.58 + ], + "angle": 0, + "content": "[5] Jianfei Guo, Nianchen Deng, Xinyang Li, Yeqi Bai, Botian Shi, Chiyu Wang, Chenjing Ding, Dongliang Wang, and Yikang Li. Streetsurf: Extending multi-view implicit surface reconstruction to street views. arXiv preprint arXiv:2306.04988, 2023. 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.58, + 0.484, + 0.65 + ], + "angle": 0, + "content": "[6] Antoine Guédon and Vincent Lepetit. Sugar: Surface-aligned gaussian splatting for efficient 3d mesh reconstruction and high-quality mesh rendering. In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5354-5363, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.65, + 0.484, + 0.705 + ], + "angle": 0, + "content": "[7] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42 (4), 2023. 1, 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.706, + 0.484, + 0.802 + ], + "angle": 0, + "content": "[8] Abhijit Kundu, Kyle Genova, Xiaqi Yin, Alireza Fathi, Caroline Pantofaru, Leonidas J Guibas, Andrea Tagliasacchi, Frank Dellaert, and Thomas Funkhouser. Panoptic neural fields: A semantic object-aware neural scene representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12871-12881, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.803, + 0.484, + 0.873 + ], + "angle": 0, + "content": "[9] Fan Lu, Yan Xu, Guang Chen, Hongsheng Li, Kwan-Yee Lin, and Changjun Jiang. Urban radiance field representation with deformable neural mesh primitives. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 465-476, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[10] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf:" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.273, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.907, + 0.134 + ], + "angle": 0, + "content": "Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 1, 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.907, + 0.19 + ], + "angle": 0, + "content": "[11] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph., 41(4):102:1-102:15, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.192, + 0.906, + 0.262 + ], + "angle": 0, + "content": "[12] Julian Ost, Fahim Mannan, Nils Thuerey, Julian Knodt, and Felix Heide. Neural scene graphs for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2856-2865, 2021. 1, 2, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.264, + 0.907, + 0.32 + ], + "angle": 0, + "content": "[13] Julian Ost, Issam Laradji, Alejandro Newell, Yuval Bahat, and Felix Heide. Neural point light fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18419-18429, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.321, + 0.906, + 0.39 + ], + "angle": 0, + "content": "[14] Konstantinos Rematas, Andrew Liu, Pratul P Srinivasan, Jonathan T Barron, Andrea Tagliasacchi, Thomas Funkhouser, and Vittorio Ferrari. Urban radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12932-12942, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.392, + 0.906, + 0.462 + ], + "angle": 0, + "content": "[15] Shital Shah, Debadeepta Dey, Chris Lovett, and Ashish Kapoor. Airsim: High-fidelity visual and physical simulation for autonomous vehicles. In Field and Service Robotics: Results of the 11th International Conference, pages 621-635. Springer, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.463, + 0.906, + 0.547 + ], + "angle": 0, + "content": "[16] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, et al. Scalability in perception for autonomous driving: Waymo open dataset. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2446-2454, 2020. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.548, + 0.906, + 0.631 + ], + "angle": 0, + "content": "[17] Su Sun, Cheng Zhao, Yuliang Guo, Ruoyu Wang, Xinyu Huang, Yingjie Victor Chen, and Liu Ren. Behind the veil: Enhanced indoor 3d scene reconstruction with occluded surfaces completion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12744-12753, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.632, + 0.906, + 0.702 + ], + "angle": 0, + "content": "[18] Adam Tonderski, Carl Lindström, Georg Hess, William Ljungbergh, Lennart Svensson, and Christoffer Petersson. Neurad: Neural rendering for autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14895-14904, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.704, + 0.906, + 0.773 + ], + "angle": 0, + "content": "[19] Haithem Turki, Jason Y Zhang, Francesco Ferroni, and Deva Ramanan. Suds: Scalable urban dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12375-12385, 2023. 1, 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.774, + 0.906, + 0.816 + ], + "angle": 0, + "content": "[20] Yihan Wang, Lahav Lipson, and Jia Deng. Sea-raft: Simple, efficient, accurate raft for optical flow. In European Conference on Computer Vision, pages 36-54. Springer, 2025. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.818, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[21] Zirui Wu, Tianyu Liu, Liyi Luo, Zhide Zhong, Jianteng Chen, Hongmin Xiao, Chao Hou, Haozhe Lou, Yuantao Chen, Runyi Yang, et al. Mars: An instance-aware, modular and realistic simulator for autonomous driving. In CAAI International Conference on Artificial Intelligence, pages 3-15. Springer, 2023. 1, 2, 5, 6, 7, 8" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "27495" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.093, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[22] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. In Neural Information Processing Systems (NeurIPS), 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.091, + 0.149, + 0.483, + 0.205 + ], + "angle": 0, + "content": "[23] Ziyang Xie, Junge Zhang, Wenye Li, Feihu Zhang, and Li Zhang. S-nerf: Neural radiance fields for street views. In International Conference on Learning Representations (ICLR), 2023. 1, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.207, + 0.484, + 0.274 + ], + "angle": 0, + "content": "[24] Yunzhi Yan, Haotong Lin, Chenxu Zhou, Weijie Wang, Haiyang Sun, Kun Zhan, Xianpeng Lang, Xiaowei Zhou, and Sida Peng. Street gaussians for modeling dynamic urban scenes. arXiv preprint arXiv:2401.01339, 2024. 1, 2, 3, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.277, + 0.482, + 0.345 + ], + "angle": 0, + "content": "[25] Jiawei Yang, Boris Ivanovic, Or Litany, Xinshuo Weng, Seung Wook Kim, Boyi Li, Tong Che, Danfei Xu, Sanja Fidler, Marco Pavone, et al. Emernerf: Emergent spatial-temporal scene decomposition via self-supervision. arXiv preprint arXiv:2311.02077, 2023. 1, 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.347, + 0.482, + 0.387 + ], + "angle": 0, + "content": "[26] Xiaoshuai Zhang, Sai Bi, Kalyan Sunkavalli, Hao Su, and Zexiang Xu. Nerfusion: Fusing radiance fields for large-scale scene reconstruction. CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.39, + 0.482, + 0.471 + ], + "angle": 0, + "content": "[27] Cheng Zhao, Su Sun, Ruoyu Wang, Yuliang Guo, Jun-Jun Wan, Zhou Huang, Xinyu Huang, Yingjie Victor Chen, and Liu Ren. Tclc-gs: Tightly coupled lidar-camera gaussian splatting for autonomous driving: Supplementary materials. In European Conference on Computer Vision, pages 91-106. Springer, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.474, + 0.482, + 0.555 + ], + "angle": 0, + "content": "[28] Hongyu Zhou, Jiahao Shao, Lu Xu, Dongfeng Bai, Weichao Qiu, Bingbing Liu, Yue Wang, Andreas Geiger, and Yiyi Liao. Hugs: Holistic urban 3d scene understanding via gaussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21336-21345, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.095, + 0.558, + 0.482, + 0.641 + ], + "angle": 0, + "content": "[29] Xiaoyu Zhou, Zhiwei Lin, Xiaojun Shan, Yongtao Wang, Deqing Sun, and Ming-Hsuan Yang. Drivinggaussian: Composite gaussian splatting for surrounding dynamic autonomous driving scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21634-21643, 2024. 2, 3, 5" + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.093, + 0.484, + 0.641 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.956 + ], + "angle": 0, + "content": "27496" + } + ] +] \ No newline at end of file diff --git a/2025/SplatFlow_ Self-Supervised Dynamic Gaussian Splatting in Neural Motion Flow Field for Autonomous Driving/8f50ebde-225c-4574-a537-44bd7c2cfecd_origin.pdf b/2025/SplatFlow_ Self-Supervised Dynamic Gaussian Splatting in Neural Motion Flow Field for Autonomous Driving/8f50ebde-225c-4574-a537-44bd7c2cfecd_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..fc18d043e29994b5278f38fd47f1abc5fe404aad --- /dev/null +++ b/2025/SplatFlow_ Self-Supervised Dynamic Gaussian Splatting in Neural Motion Flow Field for Autonomous Driving/8f50ebde-225c-4574-a537-44bd7c2cfecd_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8cc852fcb419f5623640e69cc1f7141343c4c8f19c42d43c6574f6414df75dd6 +size 9239227 diff --git a/2025/SplatFlow_ Self-Supervised Dynamic Gaussian Splatting in Neural Motion Flow Field for Autonomous Driving/full.md b/2025/SplatFlow_ Self-Supervised Dynamic Gaussian Splatting in Neural Motion Flow Field for Autonomous Driving/full.md new file mode 100644 index 0000000000000000000000000000000000000000..47d376c647d61a980c131860694fec77f59453cc --- /dev/null +++ b/2025/SplatFlow_ Self-Supervised Dynamic Gaussian Splatting in Neural Motion Flow Field for Autonomous Driving/full.md @@ -0,0 +1,327 @@ +# SplatFlow: Self-Supervised Dynamic Gaussian Splatting in Neural Motion Flow Field for Autonomous Driving + +Su Sun $^{1}$ , Cheng Zhao $^{2}$ , Zhuoyang Sun $^{1}$ , Yingjie Victor Chen $^{1}$ , Mei Chen $^{2}$ , Purdue University, Microsoft + +# Abstract + +Most existing Dynamic Gaussian Splatting methods for complex dynamic urban scenarios rely on accurate object-level supervision from expensive manual labeling, limiting their scalability in real-world applications. In this paper, we introduce SplatFlow, a Self-Supervised Dynamic Gaussian Splatting within Neural Motion Flow Fields (NMFF) to learn 4D space-time representations without requiring tracked 3D bounding boxes, enabling accurate dynamic scene reconstruction and novel view RGB/depth/flow synthesis. SplatFlow designs a unified framework to seamlessly integrate time-dependent 4D Gaussian representation within NMFF, where NMFF is a set of implicit functions to model temporal motions of both LiDAR points and Gaussians as continuous motion flow fields. Leveraging NMFF, SplatFlow effectively decomposes static background and dynamic objects, representing them with 3D and 4D Gaussian primitives, respectively. NMFF also models the correspondences of each 4D Gaussian across time, which aggregates temporal features to enhance cross-view consistency of dynamic components. SplatFlow further improves dynamic object identification by distilling features from 2D foundation models into 4D space-time representation. Comprehensive evaluations conducted on the Waymo and KITTI Datasets validate SplatFlow's state-of-the-art (SOTA) performance for both image reconstruction and novel view synthesis in dynamic urban scenarios. + +# 1. Introduction + +As autonomous driving systems increasingly shift toward end-to-end models, there is a growing need for scalable simulation environments without domain gaps where these systems can undergo closed-loop evaluations. A promising approach is real-world closed-loop evaluation, which demands controllable sensor inputs, thereby driving the development of advanced scene reconstruction techniques. In this context, Neural Radiance Fields (NeRFs) [10] and 3D Gaussian Splitting (3DGS) [7] have proven to be effective in creating high-quality 3D scene reconstructions with + +![](images/ad85828fbf5e5c4793ef06b7561c407af95dd2a356c48f6d5a0183a8d7a13a44.jpg) +Figure 1. Top: Street GS [24]; Middle: PVG [1]; Bottom: Our SplatFlow. SplatFlow eliminates the need for 3D Bboxes required by Street GS, and enhances rendering quality compared to PVG. + +excellent visual and geometric accuracy. However, accurately and comprehensively reconstructing dynamic driving scenes remains a major challenge, given the complexity of real-world scenarios without dynamic object annotations. + +Recent approaches advance NeRF-based representations by incorporating object detection and tracking to model dynamic elements, enabling photo-realistic view generation of dynamic urban street environments. Approaches [3, 8, 12, 18, 21, 23] create a scene graph, where both dynamic objects and the static background are represented as nodes, reconstructed within their canonical frames. Self-supervised methods [19, 25] model dynamic driving scenes by combining static neural field and time-dependent neural field to handle the static background and moving foreground objects separately. The SUDS [19] uses optical flow to ease the strict requirement for object labeling, while EmerNerf [25] learns temporal attributes and features in a self-supervision manner to avoid reliance on optical flow. However, these NeRF style methods face efficiency challenges in both training and inference, which becomes a major bottleneck for large-scale scene reconstruction and rendering. + +While NeRF has proven effective for driving scenes, 3DGS offers a promising alternative due to faster train + +ing and rendering speed with more explicit representations. However, the original 3DGS faces notable challenges in modeling dynamic urban environments due to its limited representation capabilities in the temporal dimension. To address this, DrivingGaussian [29] introduces composite dynamic Gaussian graphs to handle multiple moving objects and incremental static Gaussians for background representation. StreetGaussian [24] optimizes the tracked bounding boxes of dynamic Gaussians along with 4D spherical harmonics to capture changing vehicle appearances. Both methods, however, require accurate object-level supervision, such as 3D object boxes and trackers, to decompose static and dynamic elements, limiting their scalability in real-world applications. In response, PVG [1] proposes a Periodic Vibration Gaussian mechanism to model dynamic driving scenes without relying on manually labeled 3D bounding boxes. However, the scene decomposition in PVG relies on time-dependent Gaussian attributes optimized by rendering losses, but ignores motion cues within the input point cloud used for Gaussian initialization, leading to suboptimal performance in challenging scenarios with rapid movements. + +To facilitate accurate scene reconstruction and real-time rendering without expensive annotations, we propose SplatFlow, a self-supervised Dynamic Gaussian Splatting method in Neural Motion Flow Fields (NMFF) for dynamic urban scenarios. In contrast to existing methods, as illustrated in Fig. 1, SplatFlow decomposes dynamic objects and the static background in a self-supervised way without requiring expensive 3D bounding box annotations. The key idea of SplatFlow is to seamlessly integrate 4D Gaussian representations within NMFF in a unified framework. The NMFF is a set of implicit functions to model temporal motions of both LiDAR points and Gaussians as continuous motion flow fields. Leveraging NMFF, SplatFlow not only enables the decomposition of dynamic and static elements from 3D LiDAR points, but also enables the status conversion of each 4D Gaussian across time. During Gaussian splatting, we represent dynamic objects using aggregated 4D Gaussians at various viewpoints and timestamps, and the static background using 3D Gaussians, respectively. In SplatFlow, NMFF's capabilities are enhanced through three components: 1) learning 3D motion priors by pretraining on 3D LiDAR data, 2) optimizing temporal status transitions of 4D Gaussians on image data, and 3) distilling knowledge from 2D foundation models. + +The novel features of SplatFlow are summarized as: + +- SplatFlow introduces a unified framework that seamlessly integrates time-varying 4D Gaussian representations into NMFF, enabling self-supervised dynamic scene reconstruction and rendering. +- NMFF enables scene decomposition, modeling static elements with 3D Gaussians and dynamic elements with 4D + +Gaussians. + +- NMFF models correspondences of each 4D Gaussian over time, aggregating temporal features to enhance the cross-view consistency of dynamic components. +- SplatFlow enhances the dynamic object identification by uplifting 2D features distilled from foundation models to 4D space-time through optical flow rendering. + +Comprehensive experiments on the Waymo [16] and KITTI [4] benchmarks demonstrate that SplatFlow outperforms the state-of-the-art (SOTA) methods in both image reconstruction and novel view synthesis for dynamic urban scenes. Notably, SplatFlow achieves this without relying on tracked 3D bounding boxes of dynamic objects, allowing the proposed model to learn from extensive, in-the-wild data sources. + +# 2. Related work + +Real-world simulation is essential for generating data to support end-to-end autonomous driving solutions. However, current simulation engines like AirSim [15] and CARLA [2] encounter difficulties due to the expensive manual work needed to create virtual environments and the domain gap between real and simulated data. The rapid progress in Novel View Synthesis (NVS) technologies, such as NeRF [10] and 3DGS [7], and following work [6, 11, 17, 26] enables 3D reconstruction and photorealistic image generation, significantly improving real-world simulation for autonomous driving. + +For 3D reconstruction and neural rendering in autonomous driving scenes, sensor fusion solution combining surrounding cameras and LiDAR is commonly used in existing work [5, 9, 13, 14, 27]. Urban Radiance Field [14] improves NeRF training by incorporating 3D data from LiDAR to enhance 3D geometry learning. DNMP [9] employs a pre-trained deformable mesh primitive to represent the 3D scene, enhancing neural rendering quality. NPLF [13] uses explicit 3D reconstructions from LiDAR data to efficiently model the radiance field. TCLC-GS [27] design a hybrid explicit and implicit 3D representation derived from LiDAR-camera data, to enrich the properties of 3D Gaussians for splatting. Though these methods show impressive results for static urban scene rendering, they struggle with dynamic objects common in driving scenarios. + +To model dynamic urban scenes, recent approaches [8, 12, 19, 21, 25] decompose the entire scene into static background and dynamic objects, learning their representations separately. PNF [8] uses monocular 3D bounding box predictions to isolate dynamic objects and further jointly optimizes their poses during the reconstruction process. NSG [12] uses neural graphs to represent entire scenes, decomposing dynamic multi-object environments. MARS [21] employs distinct sub-networks to model backgrounds and dynamic objects, creating an instance-aware + +![](images/07b548e04dad42f58d9a831337019b0e497aa79b209a8c43d4f276e6f969b4ac.jpg) +Figure 2. The pipeline of SplatFlow. + +simulation framework. These methods typically rely on manually annotated or predicted 3D bounding boxes. In order to avoid the need of 3D bounding boxes, SUDS [19] introduces a scalable hash table to represent large-scale dynamic urban scenes, using an off-the-shelf 2D optical flow estimator to track dynamic objects. EmerNerf [25] addresses this challenge by learning scene flow to associate points across time in the 4D neural radiance field, allowing the separation of static and dynamic objects without 3D bounding boxes. However, these methods, which rely on implicit representations, still suffer from inefficiency in both reconstruction and rendering. Recently, 3DGS [7] introduced a novel explicit 3D scene representation, combining high-quality volume rendering with a fast speed. However, 3DGS is designed for static scenes and fails when modeling dynamic moving objects. To address this, DrivingGaussian [29] and StreetGaussian [24] partition the driving scene into static and dynamic components by separate sets of Gaussians based on the 3D bounding boxes of vehicles. DrivingGaussian [29] uses a composite dynamic Gaussian graph to manage multiple moving objects while incrementally employing incremental static 3D Gaussians to represent the static background. StreetGaussian [24] models foreground vehicles with estimizable poses and a 4D spherical harmonics appearance model, and background with 3D Gaussians separately. HUGS [28] also adopts 3D object boxes to identify dynamic elements, enabling optimization of geometry, appearance, semantics, and motion attributes for dynamic Gaussians. However, all these methods rely on expensive annotated or predicted 3D bounding boxes to learn the time-dependent representations of dynamic objects. Most recent method PVG [1] introduces periodic vibration based Gaussian attributes, optimized through self-supervision without the need of 3D bounding boxes. Each Gaussian models dynamic changes over time through estimizable attributes including vibration directions, life span, and life peak. + +# 3. Methodology + +# 3.1. Overview + +By collecting images and point clouds with timestamps from surrounding cameras and LiDAR, we aim to learn a space-time 4D Gaussian representation of a dynamic scene without any human annotations, enabling fast and high-quality novel view rendering. As the pipeline of SplatFlow shown in Fig. 2, NMFF models the temporal motions of both 3D points and Gaussians as continuous motion flow fields. It serves two key roles in SplatFlow: 1) decomposing static and dynamic elements from 3D LiDAR points; 2) modeling the temporal correspondences of each 4D Gaussian across time. We first decompose the time-series LiDAR points into static and dynamic points by NMFF, which are then separately merged to initialize the static and dynamic Gaussians. Dynamic objects are represented by aggregated 4D Gaussians associated with NMFF over time, while the static background is modeled using 3D Gaussians. NMFF learns the correspondence of each 4D Gaussian in conjunction with optimizing 4D Gaussians' attributes along time. We also distill optical flow knowledge from a 2D foundation model into 4D space-time representation. Finally, images, depths and optical flows are rendered from novel viewpoints at different timesteps in dynamic driving scenes. + +# 3.2. Problem Definition: + +In a dynamic urban scenario, we collect time-sequential data from a vehicle equipped with surrounding cameras and a LiDAR sensor. A set of surrounding images $\mathcal{I}$ is captured by multiple surrounding cameras with corresponding intrinsic matrices $I$ and extrinsic matrices $E$ . Meanwhile, a set of 3D points $\mathcal{P}$ is obtained from LiDAR, along with corresponding extrinsic matrices $E'$ . We denote the synchronized and calibrated multi-sensor data sequence, including timestamps $t$ , as $\{\mathcal{I}_i, \mathcal{P}_i, E_i, I_i, E_i', t_i | i = 1, 2, 3, \ldots, n\}$ , where $\mathcal{P}_i = \{x_i, y_i, z_i\}$ and $n$ is the number of frames. + +![](images/51b09e7a7704d37a50c9be91290854ae0eecb71dc2a453d00b3ed3c7cc3aa6cf.jpg) +Figure 3. Visualization of 3D LiDAR points within NMFF on Waymo dataset. + +The vehicle trajectory is either provided or estimated using multi-sensor-based odometry. Our model $\mathcal{SF}$ aims to perform accurate 3D reconstruction and synthesize novel viewpoints for any given timestamp $t$ and camera pose $[E_t, I_t]$ by rendering $\hat{\mathcal{I}} = \mathcal{SF}(E_t, I_t, t)$ . + +# 3.3. 4D Gaussian Representation + +Given images with associated camera poses, 3DGS optimizes a set of anisotropic 3D Gaussians through differentiable rasterization to represent a static 3D scene. We extend 3DGS to 4DGS of spatial-temporal representations for dynamic objects in an urban scene. Each 4D Gaussian primitives $\mathcal{G}(t)$ is represented by time-varying attributes: 3D center $\mu(t) = [x(t), y(t), z(t)]^T$ and covariance matrix $\Sigma(t)$ along with time-invariant attributes: opacity $\sigma$ and color $c$ . The density of a Gaussian at point $(x,t)$ is defined as, + +$$ +\alpha (t) = \sigma \cdot e x p \left(- \frac {1}{2} (x - \mu (t)) ^ {T} \Sigma (t) ^ {- 1} (x - \mu (t))\right). \tag {1} +$$ + +The covariance matrix $\Sigma(t) = R(t)SS^T R(t)^T$ is composed of a scaling matrix $S = \text{diag}(s_x, s_y, s_z)$ and rotation matrix $R(t) = (q_x(t), q_y(t), q_z(t), q_w(t))$ , constrained to be a positive semi-definite matrix during optimization. + +To render an image from a specific viewpoint at times-tamp $t$ , we first transform all the 4D Gaussians from other times to the target time $t$ , according to the learned correspondence from NMFF. Subsequently, the aggregated 4D Gaussians are splatted onto the image plane, resulting in a collection of 2D Gaussians. The 3D covariance matrix $\Sigma$ is projected to a 2D covariance matrix $\Sigma'$ by, + +$$ +\Sigma^ {\prime} (t) = J E \Sigma (t) E ^ {T} J ^ {T}, \tag {2} +$$ + +where $E$ refers to the world-to-camera matrix and $J$ refers to Jacobian of the perspective transformation. + +By sorting the Gaussians according to their depth within the camera space, we use $\alpha$ -blending to estimate the color $C$ and depth $D$ at each pixel $p$ as, + +$$ +C = \sum_ {i = 1} ^ {N} c _ {i} \alpha_ {i} \prod_ {j = 1} ^ {i - 1} (1 - \alpha_ {j}), D = \sum_ {i = 1} ^ {N} z _ {i} \alpha_ {i} \prod_ {j = 1} ^ {i - 1} (1 - \alpha_ {j}). \tag {3} +$$ + +Here $c_{i}$ represents the color of Gaussian $\mathcal{G}_{i}$ , which is computed using spherical harmonic coefficients. $z_{i}$ represents the distance from the image plane to the $i$ -th Gaussian center. $\alpha_{i}$ denotes the density computed from the 2D projection of the $i$ -th Gaussian and its learned opacity. + +# 3.4. Neural Motion Flow Field + +To establish the correspondence of dynamic objects, we design NMFF to model 3D motions as a continuous motion flow field, enabling the temporal transition of both 3D points and Gaussians. The NMFF $\Phi$ contains a temporal sequence of motion flow field, defined as $\Phi = \{\phi_{t_i:t_{i + 1}}\} ,i = 0,1,2,\ldots ,n - 1$ , which predicts 3D motion flow of an arbitrary query point between two consecutive frames as, + +$$ +\phi_ {t _ {1}: t _ {2}} \left(x _ {t _ {1}}, y _ {t _ {1}}, z _ {t _ {1}}\right) = \Delta x _ {t _ {1}: t _ {2}}, \Delta y _ {t _ {1}: t _ {2}}, \Delta z _ {t _ {1}: t _ {2}}, \Delta R _ {t _ {1}: t _ {2}} \tag {4} +$$ + +where $\Delta R_{t_1:t_2}$ denotes motion angle between two adjacent timestamps. Each field $\phi$ is eight ReLU-MLP stacks. + +Point Cloud within NMFF: Although NMFF can be learned during Gaussian Splatting optimization, we observed that the rendering loss alone is insufficient to effectively constrain the motion representation of dynamic scene components. Therefore, we leverage the rich geometric structure inherent in temporally consecutive 3D LiDAR points to derive a robust NMFF prior, as shown in Fig. 3. + +We first pre-train $\Phi$ on a sequence of 3D LiDAR points by forward and backward 3D geometry consistency, enabling NMFF to learn the 3D motion of each query 3D point over time. Given the source point clouds $\mathcal{P}$ and target point clouds $\mathcal{Q}$ from different timestamps, we minimize the point distance between the source and target point clouds by a bidirectional Chamfer Distance $(CD)$ , + +$$ +d _ {\mathrm {C D}} (\mathcal {P}, \mathcal {Q}) = \sum_ {p \in \mathcal {P}} \min _ {q \in \mathcal {Q}} \| p - q \| ^ {2} + \sum_ {q \in \mathcal {Q}} \min _ {p \in \mathcal {P}} \| q - p \| ^ {2}. \tag {5} +$$ + +The bidirectional $CD$ is a point-based distance that establishes correspondences in both directions—source to target and target to source—by searching for the nearest neighbors in each point cloud. + +We compensate for ego-car motion in LiDAR points to let NMFF learn only 3D motion flow during pre-training. As a result, the 3D motion flow predicted by NMFF excludes the ego-car motion. We generate a 3D dynamic mask to separate static and dynamic points by applying a threshold on each point's 3D motion flow. We then merge static point clouds using ego-motion alone, and dynamic point clouds using both ego-motion and motion flow. For the static background, we initialize the static 3D Gaus + +sians using the merged 3D static points. For dynamic objects, we initialize 4D Gaussians by aggregating points from different timestamps into a common reference frame (the mid-timestamp reference frame), which increases density and completeness by integrating observations from various viewpoints over time. + +4D Gaussian within NMFF: The pre-trained $\Phi$ is subsequently integrated with 4D Gaussian representations of dynamic objects, enabling joint optimization during Gaussian splatting. Given consecutive time $t_1$ and $t_2$ where $t_1 < t_2$ with respective states $\{\mathcal{G}_i(t_1)\}$ and $\{\mathcal{G}_i(t_2)\}$ , these states are connected by NMFF for each Gaussian as, + +$$ +\mathcal {G} _ {i} \left(t _ {1}\right) = \left\{\mu \left(t _ {1}\right), R \left(t _ {1}\right), S, \alpha , c \right\}, \tag {6} +$$ + +$$ +\Delta \mu_ {t _ {1}: t _ {2}}, \Delta R _ {t _ {1}: t _ {2}} = \phi_ {t _ {1}: t _ {2}} (\mu (t _ {1})), \tag {7} +$$ + +$$ +\hat {\mathcal {G}} _ {i} \left(t _ {2}\right) = \left\{\mu \left(t _ {1}\right) + \Delta \mu_ {t _ {1}: t _ {2}}, R \left(t _ {1}\right) \cdot \Delta R _ {t _ {1}: t _ {2}}, S, \alpha , c \right\}. \tag {8} +$$ + +We employ NMFF $\Phi$ correspondence to warp the center of all aggregated 4D Gaussians across time to the desired timestamp, as $\{\mathcal{G}_i(t)\} = \{\mathcal{G}(t),\hat{\mathcal{G}} (t_1),\hat{\mathcal{G}} (t_2),\dots ,\hat{\mathcal{G}} (t_n)\}$ . For the transition of Gaussians across multiple timestamps, we apply the corresponding NMFF through step-by-step propagation. In practice, we propagate the aggregated 4D Gaussian at the mid-timestamp to the target timestamp. Here NMFF learns a motion transition of the 4D Gaussian among training frames in a self-supervised manner, promoting a multi-view consistent representation. + +For static background, we represent it by static 3D Gaussians $\mathcal{G}_{\text{static}}$ following the strategy [1]. For sky area, we adopt a separate environmental map with static sky Gaussians $\mathcal{G}_{sky}$ following [1, 25]. Given a specific viewpoint $[E_t, I_t, t]$ , the image $\hat{\mathcal{I}}$ and depth $\hat{\mathcal{D}}$ at time $t$ are obtained by the differentiable rendering using a set of Gaussians, + +$$ +\hat {\mathcal {L}}, \hat {\mathcal {D}} = \operatorname {R e n d e r} \left(\left\{\mathcal {G} _ {i} (t) \right\}, \mathcal {G} _ {\text {s t a t i c}}, \mathcal {G} _ {\text {s k y}} \mid E _ {t}, I _ {t}, t\right). \tag {9} +$$ + +# 3.5. Optical Flow Distillation + +The 3D motion flow provided by NMFF enables 2D optical flow rendering. Given two consecutive time $t_1$ and $t_2$ , the optical flow $f_{t_1:t_2}$ of each Gaussian is computed by projecting 3D center $\mu$ to the image plane using the camera's intrinsic matrix $I$ and extrinsic matrix $E$ , + +$$ +\mu^ {\prime} \left(t _ {1}\right) = I [ E ] \mu \left(t _ {1}\right), \quad \mu^ {\prime} \left(t _ {2}\right) = I [ E ] \mu \left(t _ {2}\right), \tag {10} +$$ + +$$ +f _ {t _ {1}: t _ {2}} = \mu^ {\prime} (t _ {2}) - \mu^ {\prime} (t _ {1}). \tag {11} +$$ + +Then we estimate the optical flow $F$ at pixel $p$ via by point-based $\alpha$ -blending as, + +$$ +F = \sum_ {i = 1} ^ {N} f _ {i} \alpha_ {i} \prod_ {j = 1} ^ {i - 1} (1 - \alpha_ {j}). \tag {12} +$$ + +The optical flow $\hat{\mathcal{F}}$ at time $t$ are rendered by the differentiable rendering similar as Equation 9. In order to distill optical flow knowledge from 2D foundation model into 4D space, we adopt SEA-RAFT[20] to extract optical flow $\mathcal{F}$ as pseudo ground truth for distillation by optical flow loss, + +$$ +\mathcal {L} _ {F} = \lambda_ {f} \mathcal {L} _ {1} (\mathcal {F}, \hat {\mathcal {F}}) + (1 - \lambda_ {f}) \mathcal {L} _ {1} \left(\mathcal {I} _ {\text {n e x t}}, \mathcal {T} (\hat {\mathcal {I}} | \hat {\mathcal {F}})\right), \tag {13} +$$ + +where $\lambda_{f}$ is a scale factor, and $\mathcal{I}_{\text{next}}$ denotes the image at the next time step. $\mathcal{T}$ is used to warp rendered image to next time step according to optical flow. The optical flow $\mathcal{F}$ extracted from foundation model is only required during training and is not needed during inference. + +# 3.6. Optimization + +All Gaussian attributes, along with the parameters of NMFF, are optimized end-to-end in a self-supervised manner. Meanwhile, adaptive densification and pruning strategies are introduced to enhance the fitting of the 3D scene. The overall training loss is given as, + +$$ +\mathcal {L} = \mathcal {L} _ {I} + \lambda_ {1} \mathcal {L} _ {D} + \lambda_ {2} \mathcal {L} _ {F} + \lambda_ {3} \mathcal {L} _ {s k y} + \lambda_ {4} \mathcal {L} _ {r e g}, \tag {14} +$$ + +where $\lambda_{1,2,3,4}$ are scale factors and $\mathcal{L}_{reg}$ is the regularization term. The image loss $\mathcal{L}_I$ combines L1 and SSIM losses between rendered and observed images, + +$$ +\mathcal {L} _ {I} = \left(1 - \lambda_ {s s i m}\right) \mathcal {L} _ {1} (\mathcal {I}, \hat {\mathcal {I}}) + \lambda_ {s s i m} \mathcal {L} _ {s s i m} (\mathcal {I}, \hat {\mathcal {I}}), \tag {15} +$$ + +where $\lambda_{ssim}$ is a scale factor. The depth loss $\mathcal{L}_D$ is a L1 loss between inverse of rendered depth and generated depth by projecting sparse LiDAR points onto camera plane as, + +$$ +\mathcal {L} _ {D} = \mathcal {L} _ {1} (\mathcal {D}, \hat {\mathcal {D}}). \tag {16} +$$ + +The sky opacity loss $\mathcal{L}_{sky}$ is a binary cross entropy loss for sky refining using sky mask $M_{sky}$ from SegFormer [22], + +$$ +\mathcal {L} _ {s k y} = \mathcal {L} _ {b c e} (O, 1 - M _ {s k y}), \tag {17} +$$ + +where the accumulated opacity denotes as $O = \sum_{i=1}^{N} \alpha_i \prod_{j=1}^{i-1} (1 - \alpha_j)$ . Note the sky mask is only required during training and is not needed during inference. + +# 4. Experiments + +# 4.1. Datasets, Metrics and Baselines + +Following PVG [1], we validate our approach on two widely-used datasets in autonomous driving: Waymo Open Dataset [16] and KITTI Dataset [4]. Both offer multi-sensor data clips including synchronized and calibrated camera and LiDAR data from urban driving scenarios. We train and test our methods, along with all baselines, using full-resolution images: $1920 \times 1280$ for the Waymo dataset and $1242 \times 375$ for the KITTI dataset. + +Following prior research [1, 19, 24, 25, 29], we evaluate image reconstruction and novel view synthesis using three widely-adopted benchmark metrics: peak signal-to-noise ratio (PSNR), structural similarity index measure (SSIM), and learned perceptual image patch similarity (LPIPS). + +We compare our method with extensive baselines including S-NeRF [23], StreetSurf [5], 3DGS [7], NSG [12], Mars [21], SUDS [19], EmerNerf [25], PVG [1], Street-Gaussian [24] on the Waymo and KITTI benchmarks. See supplementary material for more details and results. + +![](images/fd1fd9d56f6988cf70fa3339459f5f08f8fd50ec5bac41f7f5ef742b84cf96cf.jpg) +Figure 4. Visual comparison of novel view synthesis on Waymo dataset. Bounding boxes indicate the zoomed-in dynamic areas. + +![](images/9ea1161dc35b3efefc018eb221f77ade5fd1ae45ac1eef9996c8d66463a9a8c8.jpg) +Figure 5. Dynamic object decomposition comparison on Waymo. + +# 4.2. Evaluation on Waymo Open Dataset + +Following PVG [1], we evaluated our SplatFlow against baselines using the Waymo Open dataset. Table 1 summarizes the average metrics for both image reconstruction and novel view synthesis tasks on selected dynamic scenes. SplatFlow consistently outperforms all baselines across all evaluated metrics for both tasks. In image reconstruction, + +
Image ReconstructionNovel View Synthesis
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
S-NeRF [23]19.670.5280.38719.220.5150.400
StreetSurf [5]26.700.8460.38723.780.8220.401
3DGS [7]27.990.8660.37225.080.8220.319
NSG [12]24.080.6560.29321.010.5710.487
Mars [21]21.810.6810.44120.690.6360.453
SUDS [19]28.830.8050.43021.830.6560.405
EmerNerf [25]28.110.7860.28925.920.7630.384
PVG [1]32.460.9100.37328.110.8490.279
SplatFlow33.640.9510.19828.710.8740.239
+ +Table 1. Performance comparison on Waymo dataset. + +SplatFlow achieves high image quality with the highest scores: PSNR at 33.64, SSIM at 0.951, and LPIPS at 0.198. For novel view synthesis, SplatFlow produces high-quality renderings of unseen timestamps, reaching PSNR of 28.71, SSIM of 0.874 and LPIPS of 0.239. These improvements are visually validated in Fig. 4 and Fig. 9 (dynamic only), showing exceptional clarity in details for both static and dynamic regions. The dynamic objects in Fig. 9 are selected from a distant background car in an extremely zoomed-in + +![](images/428c433cf61a8afe8c8e3d527ba4d11a397f89fb8cd48d1d0f2620400fb0b2bf.jpg) +Figure 6. Visual comparison of novel view synthesis on KITTI dataset. Bounding boxes indicate the zoomed-in dynamic areas. + +
PVG [1]SplatFlow
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
Seg. 1058...26.860.8400.27327.460.8800.253
Seg. 2259...26.450.8240.30929.300.8560.290
Seg. 7670...26.580.8200.30728.890.8550.283
Seg. 5083...28.350.8930.21330.290.9280.169
PSNR*↑SSIM*↑LPIPS*↓PSNR*↑SSIM*↑LPIPS*↓
Seg. 1058...27.420.8350.26728.500.8770.234
Seg. 2259...22.550.8030.35031.610.8990.191
Seg. 7670...23.560.7560.33028.760.8200.289
Seg. 5083...25.580.8140.29026.710.8550.211
+ +Table 2. Novel view synthesis results on Waymo dataset (*denotes dynamic elements only). + +
Image ReconstructionNovel View Synthesis
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
S-NeRF [23]19.230.6640.19318.710.6060.352
StreetSurf [5]24.140.8190.25722.480.7630.304
3DGS [7]21.020.8110.20219.540.7760.224
NSG [12]26.660.8060.18621.530.6730.254
Mars [21]27.960.9000.18524.230.8450.160
SUDS [19]28.310.8760.18522.770.7970.171
EmerNeRF [25]26.950.8280.21825.240.8010.237
PVG [1]32.830.9370.07027.430.8960.114
SplatFlow33.370.9430.05728.320.9320.089
+ +Table 3. Performance comparison on KITTI dataset. + +view. Compared with baselines which show artifacts such as ghosting and blur in regions of dynamic elements, our method preserves high-quality textures and fine details, producing more accurate renderings at novel viewpoints. + +To further evaluate the rendering quality of dynamic objects against the primary baseline PVG, we selected additional four scenes featuring a higher presence of dynamic objects for further comparison. Table 2 displays the metrics for novel view synthesis, evaluated both on entire scenes and exclusively on dynamic objects (denoted as *). Splat-Flow consistently surpasses PVG in all selected scenes, + +![](images/336bbec7d5ff23eea71f5723ec3dff7eab8e59b8e18489509cb111d4e75b7a55.jpg) +Figure 7. Visual comparison of novel view synthesis on KITTI-25% (row1), 50% (row2), and 75% (row3) dataset. + +demonstrating higher PSNR and SSIM values and lower LPIPS scores, underscoring its effectiveness in handling complex, dynamic content. The dynamic objects masks are obtained from 2D ground truth bounding boxes following [25]. The PSNR* of dynamic regions on novel view synthesis significantly outperforms SVG both quantitatively and qualitatively, especially in scenes with fast ego-motion and moving objects in Seg. 2259 and Seg. 7670. We also present the visual comparison of dynamic object decomposition with SVG in Fig. 5, which illustrates that SplatFlow renders sharper and clearer image of dynamic objects. + +# 4.3. Evaluation on KITTI Dataset + +Following PVG [1], we further evaluated our SplatFlow method in comparison to the baselines on KITTI dataset. + +
KITTI - 75%KITTI - 50%KITTI - 25%
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
NeRF [10]18.560.5570.55419.120.5870.49718.610.5700.510
NSG [12]21.530.6730.25421.260.6590.26620.000.6320.281
SUDS [19]22.770.7970.17123.120.8210.13520.760.7470.198
MARS [21]24.230.8450.16024.000.8010.16423.230.7560.177
3DGS [7]19.190.7370.17219.230.7390.17419.060.7300.180
PVG [1]27.430.8960.11425.920.8820.11422.550.8330.151
StreetGS [24]25.790.8440.08125.520.8410.08424.530.8240.090
SplatFlow28.320.9320.08927.900.9270.09326.100.8900.120
+ +Table 4. Performance comparison of novel view synthesis on KITTI-75%, 50% and 25% dataset. + +
PSNR↑SSIM↑LPIPS↓
w/o NMFF prior27.690.8630.282
w/o NMFF optimization28.140.8740.269
w/o optical flow distillation28.280.8770.252
Full28.990.8800.249
PSNR*↑SSIM*↑LPIPS*↓
w/o NMFF prior27.280.8270.317
w/o NMFF optimization27.970.8430.254
w/o optical flow distillation28.510.8610.232
Full28.900.8630.231
+ +Table 5. Ablation study on Waymo dataset (*denotes dynamic elements only). + +![](images/b23fb79d066066217788bf69edd4a526f759dd80acd773c4f34f9c6f933bdf6f.jpg) +Ours (Full) + +![](images/c10f81fdc6b6a701e91d9ab9769e18015d37c53429188620372c08403c115e9d.jpg) +w/o Opt. flow + +![](images/7984c352b65409864afbe33329dfdda80a0112e3397b395d7df45cc9af2194ac.jpg) +w/o NMFF optim. +Figure 8. Visual comparison of ablation study on Waymo dataset. + +![](images/9d2312c8543933d1ba7f9195ecfb5ffbcdf8a45b638585d4ff25448a60f82742.jpg) +w/o NMFF prior + +Table 3 presents the average metrics for both image reconstruction and novel view synthesis tasks on selected dynamic scenes characterized by extensive movement. SplatFlow outperforms all other methods across all metrics in both tasks, with particularly high scores in PSNR. SSIM and LPIPS, achieving 33.37, 0.943 and 0.057 for image reconstruction and 28.32, 0.932 and 0.089 for novel view synthesis. SplatFlow achieves exceptional results in novel view synthesis, as shown in Fig. 6. Our approach generates high-fidelity renderings that preserve fine details while accurately reconstructing dynamic areas where existing approaches struggle to maintain stability. + +To evaluate robustness to training data size, we follow the settings of [21, 24] to evaluate our method using different train/test split configurations on the KITTI dataset. Table 4 and Fig. 7 provide the performance comparisons and visualizations. SplatFlow consistently outperforms all other methods across all dataset subsets, achieving the highest scores. Even with reduced training data, SplatFlow robustly delivers high-quality novel view synthesis across varying levels of data availability, highlighting its efficient use of available information and strong generalization capabilities. + +# 4.4. Ablation Study + +To demonstrate the effectiveness of each component in SplatFlow, we performed ablation studies on the Waymo dataset. Our main contributions, NMFF prior, NMFF opti + +![](images/0243cf949d0320044abb1b88aef349e7e5ef7d6cec9901131ee8bd565d40b8f8.jpg) +EmerNeRF +Figure 9. Detail comparison of novel view synthesis on Waymo. + +![](images/60cc694cc2abedebf79f2a9db38753c85c8a798657e9360f71b41e1244ec7422.jpg) +PVG + +![](images/b4e74a1558ae485f8e44c8c1daad2b64cbfd1e718bdd456c26411786770e1179.jpg) +SplatFlow + +![](images/b9d300c49a381b8f94c7574a230ecfab7f3b7ed1e1c7a80e8023b8fc493ea20f.jpg) +G.T. + +mization, and optical flow distillation are analyzed in this study to assess their individual impact. We trained four different variations: 1) SplatFlow without pretraining on LiDAR data; 2) SplatFlow without optimization with 4D Gaussians on image data; 3) SplatFlow without optical flow distillation from foundational model; 4) SplatFlow full method. Table 5 presents the average evaluation metrics across test scenes used in Table 2, showing results for both entire scenes and dynamic objects (indicated as *). Ablation study validates the contribution of each component to the overall performance. + +# 5. Conclusion + +In this paper, we propose SplatFlow, a novel self-supervised Dynamic Gaussian Splitting within Neural Motion Flow Fields (NMFF) for accurate reconstruction and real-time rendering in dynamic urban scenarios. The core idea of SplatFlow is to seamlessly integrate time-dependent 4D Gaussian representations within NMFF in a unified framework, where NMFF implicitly models the motions of dynamic components across time through self-supervision. Experimental evaluations show that SplatFlow outperforms state-of-the-art across all standard metrics on the Waymo Open and KITTI datasets, without requiring expensive annotations of dynamic object detection and tracking. + +# Acknowledgement + +This work was partially supported by the intramural research program of the U.S. Department of Agriculture Natural Resources Conservation Service, Climate-Smart Commodities program, EFFICACI grant, #NR233A750004G044. The findings and conclusions in this work have not been formally disseminated by the U. S. Department of Agriculture and should not be construed to represent any agency determination or policy. + +# References + +[1] Yurui Chen, Chun Gu, Junzhe Jiang, Xiatian Zhu, and Li Zhang. Periodic vibration gaussian: Dynamic urban scene reconstruction and real-time rendering. arXiv preprint arXiv:2311.18561, 20.1, 2, 3, 5, 6, 7, 8 +[2] Alexey Dosovitskiy, German Ros, Felipe Codevilla, Antonio Lopez, and Vladlen Koltun. Carla: An open urban driving simulator. In Conference on robot learning, pages 1-16. PMLR, 2017. 2 +[3] Tobias Fischer, Lorenzo Porzi, Samuel Rota Bulo, Marc Pollefeys, and Peter Kontschieder. Multi-level neural scene graphs for dynamic urban environments. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21125-21135, 2024. 1 +[4] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In 2012 IEEE conference on computer vision and pattern recognition, pages 3354-3361. IEEE, 2012. 2, 5 +[5] Jianfei Guo, Nianchen Deng, Xinyang Li, Yeqi Bai, Botian Shi, Chiyu Wang, Chenjing Ding, Dongliang Wang, and Yikang Li. Streetsurf: Extending multi-view implicit surface reconstruction to street views. arXiv preprint arXiv:2306.04988, 2023. 2, 5, 6, 7 +[6] Antoine Guédon and Vincent Lepetit. Sugar: Surface-aligned gaussian splatting for efficient 3d mesh reconstruction and high-quality mesh rendering. In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5354-5363, 2024. 2 +[7] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42 (4), 2023. 1, 2, 3, 5, 6, 7, 8 +[8] Abhijit Kundu, Kyle Genova, Xiaqi Yin, Alireza Fathi, Caroline Pantofaru, Leonidas J Guibas, Andrea Tagliasacchi, Frank Dellaert, and Thomas Funkhouser. Panoptic neural fields: A semantic object-aware neural scene representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12871-12881, 2022. 1, 2 +[9] Fan Lu, Yan Xu, Guang Chen, Hongsheng Li, Kwan-Yee Lin, and Changjun Jiang. Urban radiance field representation with deformable neural mesh primitives. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 465-476, 2023. 2 +[10] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: + +Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 1, 2, 8 +[11] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph., 41(4):102:1-102:15, 2022. 2 +[12] Julian Ost, Fahim Mannan, Nils Thuerey, Julian Knodt, and Felix Heide. Neural scene graphs for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2856-2865, 2021. 1, 2, 5, 6, 7, 8 +[13] Julian Ost, Issam Laradji, Alejandro Newell, Yuval Bahat, and Felix Heide. Neural point light fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18419-18429, 2022. 2 +[14] Konstantinos Rematas, Andrew Liu, Pratul P Srinivasan, Jonathan T Barron, Andrea Tagliasacchi, Thomas Funkhouser, and Vittorio Ferrari. Urban radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12932-12942, 2022. 2 +[15] Shital Shah, Debadeepta Dey, Chris Lovett, and Ashish Kapoor. Airsim: High-fidelity visual and physical simulation for autonomous vehicles. In Field and Service Robotics: Results of the 11th International Conference, pages 621-635. Springer, 2018. 2 +[16] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, et al. Scalability in perception for autonomous driving: Waymo open dataset. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2446-2454, 2020. 2, 5 +[17] Su Sun, Cheng Zhao, Yuliang Guo, Ruoyu Wang, Xinyu Huang, Yingjie Victor Chen, and Liu Ren. Behind the veil: Enhanced indoor 3d scene reconstruction with occluded surfaces completion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12744-12753, 2024. 2 +[18] Adam Tonderski, Carl Lindström, Georg Hess, William Ljungbergh, Lennart Svensson, and Christoffer Petersson. Neurad: Neural rendering for autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14895-14904, 2024. 1 +[19] Haithem Turki, Jason Y Zhang, Francesco Ferroni, and Deva Ramanan. Suds: Scalable urban dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12375-12385, 2023. 1, 2, 3, 5, 6, 7, 8 +[20] Yihan Wang, Lahav Lipson, and Jia Deng. Sea-raft: Simple, efficient, accurate raft for optical flow. In European Conference on Computer Vision, pages 36-54. Springer, 2025. 5 +[21] Zirui Wu, Tianyu Liu, Liyi Luo, Zhide Zhong, Jianteng Chen, Hongmin Xiao, Chao Hou, Haozhe Lou, Yuantao Chen, Runyi Yang, et al. Mars: An instance-aware, modular and realistic simulator for autonomous driving. In CAAI International Conference on Artificial Intelligence, pages 3-15. Springer, 2023. 1, 2, 5, 6, 7, 8 + +[22] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. In Neural Information Processing Systems (NeurIPS), 2021. 5 +[23] Ziyang Xie, Junge Zhang, Wenye Li, Feihu Zhang, and Li Zhang. S-nerf: Neural radiance fields for street views. In International Conference on Learning Representations (ICLR), 2023. 1, 5, 6, 7 +[24] Yunzhi Yan, Haotong Lin, Chenxu Zhou, Weijie Wang, Haiyang Sun, Kun Zhan, Xianpeng Lang, Xiaowei Zhou, and Sida Peng. Street gaussians for modeling dynamic urban scenes. arXiv preprint arXiv:2401.01339, 2024. 1, 2, 3, 5, 8 +[25] Jiawei Yang, Boris Ivanovic, Or Litany, Xinshuo Weng, Seung Wook Kim, Boyi Li, Tong Che, Danfei Xu, Sanja Fidler, Marco Pavone, et al. Emernerf: Emergent spatial-temporal scene decomposition via self-supervision. arXiv preprint arXiv:2311.02077, 2023. 1, 2, 3, 5, 6, 7 +[26] Xiaoshuai Zhang, Sai Bi, Kalyan Sunkavalli, Hao Su, and Zexiang Xu. Nerfusion: Fusing radiance fields for large-scale scene reconstruction. CVPR, 2022. 2 +[27] Cheng Zhao, Su Sun, Ruoyu Wang, Yuliang Guo, Jun-Jun Wan, Zhou Huang, Xinyu Huang, Yingjie Victor Chen, and Liu Ren. Tclc-gs: Tightly coupled lidar-camera gaussian splatting for autonomous driving: Supplementary materials. In European Conference on Computer Vision, pages 91-106. Springer, 2024. 2 +[28] Hongyu Zhou, Jiahao Shao, Lu Xu, Dongfeng Bai, Weichao Qiu, Bingbing Liu, Yue Wang, Andreas Geiger, and Yiyi Liao. Hugs: Holistic urban 3d scene understanding via gaussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21336-21345, 2024. 3 +[29] Xiaoyu Zhou, Zhiwei Lin, Xiaojun Shan, Yongtao Wang, Deqing Sun, and Ming-Hsuan Yang. Drivinggaussian: Composite gaussian splatting for surrounding dynamic autonomous driving scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21634-21643, 2024. 2, 3, 5 \ No newline at end of file diff --git a/2025/SplatFlow_ Self-Supervised Dynamic Gaussian Splatting in Neural Motion Flow Field for Autonomous Driving/images.zip b/2025/SplatFlow_ Self-Supervised Dynamic Gaussian Splatting in Neural Motion Flow Field for Autonomous Driving/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..e6f6048d0f4a1130d20bc2ea8ffee63cb1bc84d4 --- /dev/null +++ b/2025/SplatFlow_ Self-Supervised Dynamic Gaussian Splatting in Neural Motion Flow Field for Autonomous Driving/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4d7fc005d616216113e23114c70dfa0a0a28442214fcccb8d09419938d909aff +size 1093254 diff --git a/2025/SplatFlow_ Self-Supervised Dynamic Gaussian Splatting in Neural Motion Flow Field for Autonomous Driving/layout.json b/2025/SplatFlow_ Self-Supervised Dynamic Gaussian Splatting in Neural Motion Flow Field for Autonomous Driving/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..30d4924a57f452e6997325334f00ccdf80e0dfb5 --- /dev/null +++ b/2025/SplatFlow_ Self-Supervised Dynamic Gaussian Splatting in Neural Motion Flow Field for Autonomous Driving/layout.json @@ -0,0 +1,8327 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 59, + 102, + 550, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 102, + 550, + 140 + ], + "spans": [ + { + "bbox": [ + 59, + 102, + 550, + 140 + ], + "type": "text", + "content": "SplatFlow: Self-Supervised Dynamic Gaussian Splatting in Neural Motion Flow Field for Autonomous Driving" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 119, + 160, + 489, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 119, + 160, + 489, + 190 + ], + "spans": [ + { + "bbox": [ + 119, + 160, + 489, + 190 + ], + "type": "text", + "content": "Su Sun" + }, + { + "bbox": [ + 119, + 160, + 489, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 119, + 160, + 489, + 190 + ], + "type": "text", + "content": ", Cheng Zhao" + }, + { + "bbox": [ + 119, + 160, + 489, + 190 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 119, + 160, + 489, + 190 + ], + "type": "text", + "content": ", Zhuoyang Sun" + }, + { + "bbox": [ + 119, + 160, + 489, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 119, + 160, + 489, + 190 + ], + "type": "text", + "content": ", Yingjie Victor Chen" + }, + { + "bbox": [ + 119, + 160, + 489, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 119, + 160, + 489, + 190 + ], + "type": "text", + "content": ", Mei Chen" + }, + { + "bbox": [ + 119, + 160, + 489, + 190 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 119, + 160, + 489, + 190 + ], + "type": "text", + "content": ", Purdue University, Microsoft" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 151, + 217, + 200, + 229 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 217, + 200, + 229 + ], + "spans": [ + { + "bbox": [ + 151, + 217, + 200, + 229 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 242, + 296, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 242, + 296, + 553 + ], + "spans": [ + { + "bbox": [ + 55, + 242, + 296, + 553 + ], + "type": "text", + "content": "Most existing Dynamic Gaussian Splatting methods for complex dynamic urban scenarios rely on accurate object-level supervision from expensive manual labeling, limiting their scalability in real-world applications. In this paper, we introduce SplatFlow, a Self-Supervised Dynamic Gaussian Splatting within Neural Motion Flow Fields (NMFF) to learn 4D space-time representations without requiring tracked 3D bounding boxes, enabling accurate dynamic scene reconstruction and novel view RGB/depth/flow synthesis. SplatFlow designs a unified framework to seamlessly integrate time-dependent 4D Gaussian representation within NMFF, where NMFF is a set of implicit functions to model temporal motions of both LiDAR points and Gaussians as continuous motion flow fields. Leveraging NMFF, SplatFlow effectively decomposes static background and dynamic objects, representing them with 3D and 4D Gaussian primitives, respectively. NMFF also models the correspondences of each 4D Gaussian across time, which aggregates temporal features to enhance cross-view consistency of dynamic components. SplatFlow further improves dynamic object identification by distilling features from 2D foundation models into 4D space-time representation. Comprehensive evaluations conducted on the Waymo and KITTI Datasets validate SplatFlow's state-of-the-art (SOTA) performance for both image reconstruction and novel view synthesis in dynamic urban scenarios." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 557, + 135, + 570 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 557, + 135, + 570 + ], + "spans": [ + { + "bbox": [ + 56, + 557, + 135, + 570 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 577, + 295, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 577, + 295, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 577, + 295, + 696 + ], + "type": "text", + "content": "As autonomous driving systems increasingly shift toward end-to-end models, there is a growing need for scalable simulation environments without domain gaps where these systems can undergo closed-loop evaluations. A promising approach is real-world closed-loop evaluation, which demands controllable sensor inputs, thereby driving the development of advanced scene reconstruction techniques. In this context, Neural Radiance Fields (NeRFs) [10] and 3D Gaussian Splitting (3DGS) [7] have proven to be effective in creating high-quality 3D scene reconstructions with" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 316, + 217, + 553, + 395 + ], + "blocks": [ + { + "bbox": [ + 316, + 217, + 553, + 395 + ], + "lines": [ + { + "bbox": [ + 316, + 217, + 553, + 395 + ], + "spans": [ + { + "bbox": [ + 316, + 217, + 553, + 395 + ], + "type": "image", + "image_path": "ad85828fbf5e5c4793ef06b7561c407af95dd2a356c48f6d5a0183a8d7a13a44.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 396, + 555, + 430 + ], + "lines": [ + { + "bbox": [ + 313, + 396, + 555, + 430 + ], + "spans": [ + { + "bbox": [ + 313, + 396, + 555, + 430 + ], + "type": "text", + "content": "Figure 1. Top: Street GS [24]; Middle: PVG [1]; Bottom: Our SplatFlow. SplatFlow eliminates the need for 3D Bboxes required by Street GS, and enhances rendering quality compared to PVG." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 434, + 555, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 434, + 555, + 482 + ], + "spans": [ + { + "bbox": [ + 313, + 434, + 555, + 482 + ], + "type": "text", + "content": "excellent visual and geometric accuracy. However, accurately and comprehensively reconstructing dynamic driving scenes remains a major challenge, given the complexity of real-world scenarios without dynamic object annotations." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 312, + 484, + 555, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 484, + 555, + 688 + ], + "spans": [ + { + "bbox": [ + 312, + 484, + 555, + 688 + ], + "type": "text", + "content": "Recent approaches advance NeRF-based representations by incorporating object detection and tracking to model dynamic elements, enabling photo-realistic view generation of dynamic urban street environments. Approaches [3, 8, 12, 18, 21, 23] create a scene graph, where both dynamic objects and the static background are represented as nodes, reconstructed within their canonical frames. Self-supervised methods [19, 25] model dynamic driving scenes by combining static neural field and time-dependent neural field to handle the static background and moving foreground objects separately. The SUDS [19] uses optical flow to ease the strict requirement for object labeling, while EmerNerf [25] learns temporal attributes and features in a self-supervision manner to avoid reliance on optical flow. However, these NeRF style methods face efficiency challenges in both training and inference, which becomes a major bottleneck for large-scale scene reconstruction and rendering." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "content": "While NeRF has proven effective for driving scenes, 3DGS offers a promising alternative due to faster train" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 703, + 193, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 703, + 193, + 712 + ], + "spans": [ + { + "bbox": [ + 66, + 703, + 193, + 712 + ], + "type": "text", + "content": "*Equally contributed as co-first author." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27487" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 335 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 335 + ], + "type": "text", + "content": "ing and rendering speed with more explicit representations. However, the original 3DGS faces notable challenges in modeling dynamic urban environments due to its limited representation capabilities in the temporal dimension. To address this, DrivingGaussian [29] introduces composite dynamic Gaussian graphs to handle multiple moving objects and incremental static Gaussians for background representation. StreetGaussian [24] optimizes the tracked bounding boxes of dynamic Gaussians along with 4D spherical harmonics to capture changing vehicle appearances. Both methods, however, require accurate object-level supervision, such as 3D object boxes and trackers, to decompose static and dynamic elements, limiting their scalability in real-world applications. In response, PVG [1] proposes a Periodic Vibration Gaussian mechanism to model dynamic driving scenes without relying on manually labeled 3D bounding boxes. However, the scene decomposition in PVG relies on time-dependent Gaussian attributes optimized by rendering losses, but ignores motion cues within the input point cloud used for Gaussian initialization, leading to suboptimal performance in challenging scenarios with rapid movements." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 338, + 295, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 338, + 295, + 624 + ], + "spans": [ + { + "bbox": [ + 56, + 338, + 295, + 624 + ], + "type": "text", + "content": "To facilitate accurate scene reconstruction and real-time rendering without expensive annotations, we propose SplatFlow, a self-supervised Dynamic Gaussian Splatting method in Neural Motion Flow Fields (NMFF) for dynamic urban scenarios. In contrast to existing methods, as illustrated in Fig. 1, SplatFlow decomposes dynamic objects and the static background in a self-supervised way without requiring expensive 3D bounding box annotations. The key idea of SplatFlow is to seamlessly integrate 4D Gaussian representations within NMFF in a unified framework. The NMFF is a set of implicit functions to model temporal motions of both LiDAR points and Gaussians as continuous motion flow fields. Leveraging NMFF, SplatFlow not only enables the decomposition of dynamic and static elements from 3D LiDAR points, but also enables the status conversion of each 4D Gaussian across time. During Gaussian splatting, we represent dynamic objects using aggregated 4D Gaussians at various viewpoints and timestamps, and the static background using 3D Gaussians, respectively. In SplatFlow, NMFF's capabilities are enhanced through three components: 1) learning 3D motion priors by pretraining on 3D LiDAR data, 2) optimizing temporal status transitions of 4D Gaussians on image data, and 3) distilling knowledge from 2D foundation models." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 627, + 278, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 627, + 278, + 639 + ], + "spans": [ + { + "bbox": [ + 67, + 627, + 278, + 639 + ], + "type": "text", + "content": "The novel features of SplatFlow are summarized as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 642, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 56, + 642, + 295, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 642, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 642, + 295, + 689 + ], + "type": "text", + "content": "- SplatFlow introduces a unified framework that seamlessly integrates time-varying 4D Gaussian representations into NMFF, enabling self-supervised dynamic scene reconstruction and rendering." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 689, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 689, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 689, + 295, + 713 + ], + "type": "text", + "content": "- NMFF enables scene decomposition, modeling static elements with 3D Gaussians and dynamic elements with 4D" + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 323, + 73, + 368, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 73, + 368, + 83 + ], + "spans": [ + { + "bbox": [ + 323, + 73, + 368, + 83 + ], + "type": "text", + "content": "Gaussians." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 84, + 553, + 156 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 314, + 84, + 553, + 119 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 84, + 553, + 119 + ], + "spans": [ + { + "bbox": [ + 314, + 84, + 553, + 119 + ], + "type": "text", + "content": "- NMFF models correspondences of each 4D Gaussian over time, aggregating temporal features to enhance the cross-view consistency of dynamic components." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 121, + 553, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 121, + 553, + 156 + ], + "spans": [ + { + "bbox": [ + 313, + 121, + 553, + 156 + ], + "type": "text", + "content": "- SplatFlow enhances the dynamic object identification by uplifting 2D features distilled from foundation models to 4D space-time through optical flow rendering." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 157, + 553, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 157, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 313, + 157, + 553, + 251 + ], + "type": "text", + "content": "Comprehensive experiments on the Waymo [16] and KITTI [4] benchmarks demonstrate that SplatFlow outperforms the state-of-the-art (SOTA) methods in both image reconstruction and novel view synthesis for dynamic urban scenes. Notably, SplatFlow achieves this without relying on tracked 3D bounding boxes of dynamic objects, allowing the proposed model to learn from extensive, in-the-wild data sources." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 262, + 398, + 275 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 262, + 398, + 275 + ], + "spans": [ + { + "bbox": [ + 314, + 262, + 398, + 275 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 282, + 554, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 282, + 554, + 413 + ], + "spans": [ + { + "bbox": [ + 313, + 282, + 554, + 413 + ], + "type": "text", + "content": "Real-world simulation is essential for generating data to support end-to-end autonomous driving solutions. However, current simulation engines like AirSim [15] and CARLA [2] encounter difficulties due to the expensive manual work needed to create virtual environments and the domain gap between real and simulated data. The rapid progress in Novel View Synthesis (NVS) technologies, such as NeRF [10] and 3DGS [7], and following work [6, 11, 17, 26] enables 3D reconstruction and photorealistic image generation, significantly improving real-world simulation for autonomous driving." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 415, + 554, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 415, + 554, + 594 + ], + "spans": [ + { + "bbox": [ + 313, + 415, + 554, + 594 + ], + "type": "text", + "content": "For 3D reconstruction and neural rendering in autonomous driving scenes, sensor fusion solution combining surrounding cameras and LiDAR is commonly used in existing work [5, 9, 13, 14, 27]. Urban Radiance Field [14] improves NeRF training by incorporating 3D data from LiDAR to enhance 3D geometry learning. DNMP [9] employs a pre-trained deformable mesh primitive to represent the 3D scene, enhancing neural rendering quality. NPLF [13] uses explicit 3D reconstructions from LiDAR data to efficiently model the radiance field. TCLC-GS [27] design a hybrid explicit and implicit 3D representation derived from LiDAR-camera data, to enrich the properties of 3D Gaussians for splatting. Though these methods show impressive results for static urban scene rendering, they struggle with dynamic objects common in driving scenarios." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 594, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 554, + 713 + ], + "type": "text", + "content": "To model dynamic urban scenes, recent approaches [8, 12, 19, 21, 25] decompose the entire scene into static background and dynamic objects, learning their representations separately. PNF [8] uses monocular 3D bounding box predictions to isolate dynamic objects and further jointly optimizes their poses during the reconstruction process. NSG [12] uses neural graphs to represent entire scenes, decomposing dynamic multi-object environments. MARS [21] employs distinct sub-networks to model backgrounds and dynamic objects, creating an instance-aware" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27488" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 79, + 541, + 248 + ], + "blocks": [ + { + "bbox": [ + 58, + 79, + 541, + 248 + ], + "lines": [ + { + "bbox": [ + 58, + 79, + 541, + 248 + ], + "spans": [ + { + "bbox": [ + 58, + 79, + 541, + 248 + ], + "type": "image", + "image_path": "07b548e04dad42f58d9a831337019b0e497aa79b209a8c43d4f276e6f969b4ac.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 238, + 249, + 372, + 261 + ], + "lines": [ + { + "bbox": [ + 238, + 249, + 372, + 261 + ], + "spans": [ + { + "bbox": [ + 238, + 249, + 372, + 261 + ], + "type": "text", + "content": "Figure 2. The pipeline of SplatFlow." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 264, + 297, + 709 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 264, + 297, + 709 + ], + "spans": [ + { + "bbox": [ + 55, + 264, + 297, + 709 + ], + "type": "text", + "content": "simulation framework. These methods typically rely on manually annotated or predicted 3D bounding boxes. In order to avoid the need of 3D bounding boxes, SUDS [19] introduces a scalable hash table to represent large-scale dynamic urban scenes, using an off-the-shelf 2D optical flow estimator to track dynamic objects. EmerNerf [25] addresses this challenge by learning scene flow to associate points across time in the 4D neural radiance field, allowing the separation of static and dynamic objects without 3D bounding boxes. However, these methods, which rely on implicit representations, still suffer from inefficiency in both reconstruction and rendering. Recently, 3DGS [7] introduced a novel explicit 3D scene representation, combining high-quality volume rendering with a fast speed. However, 3DGS is designed for static scenes and fails when modeling dynamic moving objects. To address this, DrivingGaussian [29] and StreetGaussian [24] partition the driving scene into static and dynamic components by separate sets of Gaussians based on the 3D bounding boxes of vehicles. DrivingGaussian [29] uses a composite dynamic Gaussian graph to manage multiple moving objects while incrementally employing incremental static 3D Gaussians to represent the static background. StreetGaussian [24] models foreground vehicles with estimizable poses and a 4D spherical harmonics appearance model, and background with 3D Gaussians separately. HUGS [28] also adopts 3D object boxes to identify dynamic elements, enabling optimization of geometry, appearance, semantics, and motion attributes for dynamic Gaussians. However, all these methods rely on expensive annotated or predicted 3D bounding boxes to learn the time-dependent representations of dynamic objects. Most recent method PVG [1] introduces periodic vibration based Gaussian attributes, optimized through self-supervision without the need of 3D bounding boxes. Each Gaussian models dynamic changes over time through estimizable attributes including vibration directions, life span, and life peak." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 314, + 263, + 398, + 277 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 263, + 398, + 277 + ], + "spans": [ + { + "bbox": [ + 314, + 263, + 398, + 277 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 284, + 383, + 296 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 284, + 383, + 296 + ], + "spans": [ + { + "bbox": [ + 313, + 284, + 383, + 296 + ], + "type": "text", + "content": "3.1. Overview" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 302, + 556, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 302, + 556, + 567 + ], + "spans": [ + { + "bbox": [ + 313, + 302, + 556, + 567 + ], + "type": "text", + "content": "By collecting images and point clouds with timestamps from surrounding cameras and LiDAR, we aim to learn a space-time 4D Gaussian representation of a dynamic scene without any human annotations, enabling fast and high-quality novel view rendering. As the pipeline of SplatFlow shown in Fig. 2, NMFF models the temporal motions of both 3D points and Gaussians as continuous motion flow fields. It serves two key roles in SplatFlow: 1) decomposing static and dynamic elements from 3D LiDAR points; 2) modeling the temporal correspondences of each 4D Gaussian across time. We first decompose the time-series LiDAR points into static and dynamic points by NMFF, which are then separately merged to initialize the static and dynamic Gaussians. Dynamic objects are represented by aggregated 4D Gaussians associated with NMFF over time, while the static background is modeled using 3D Gaussians. NMFF learns the correspondence of each 4D Gaussian in conjunction with optimizing 4D Gaussians' attributes along time. We also distill optical flow knowledge from a 2D foundation model into 4D space-time representation. Finally, images, depths and optical flows are rendered from novel viewpoints at different timesteps in dynamic driving scenes." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 575, + 430, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 575, + 430, + 588 + ], + "spans": [ + { + "bbox": [ + 313, + 575, + 430, + 588 + ], + "type": "text", + "content": "3.2. Problem Definition:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "content": "In a dynamic urban scenario, we collect time-sequential data from a vehicle equipped with surrounding cameras and a LiDAR sensor. A set of surrounding images " + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{I}" + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "content": " is captured by multiple surrounding cameras with corresponding intrinsic matrices " + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "content": " and extrinsic matrices " + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "content": ". Meanwhile, a set of 3D points " + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "content": " is obtained from LiDAR, along with corresponding extrinsic matrices " + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "inline_equation", + "content": "E'" + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "content": ". We denote the synchronized and calibrated multi-sensor data sequence, including timestamps " + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "content": ", as " + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{I}_i, \\mathcal{P}_i, E_i, I_i, E_i', t_i | i = 1, 2, 3, \\ldots, n\\}" + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_i = \\{x_i, y_i, z_i\\}" + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "content": " is the number of frames." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "27489" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 71, + 553, + 189 + ], + "blocks": [ + { + "bbox": [ + 59, + 71, + 553, + 189 + ], + "lines": [ + { + "bbox": [ + 59, + 71, + 553, + 189 + ], + "spans": [ + { + "bbox": [ + 59, + 71, + 553, + 189 + ], + "type": "image", + "image_path": "51b09e7a7704d37a50c9be91290854ae0eecb71dc2a453d00b3ed3c7cc3aa6cf.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 162, + 189, + 447, + 200 + ], + "lines": [ + { + "bbox": [ + 162, + 189, + 447, + 200 + ], + "spans": [ + { + "bbox": [ + 162, + 189, + 447, + 200 + ], + "type": "text", + "content": "Figure 3. Visualization of 3D LiDAR points within NMFF on Waymo dataset." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 205, + 296, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 205, + 296, + 266 + ], + "spans": [ + { + "bbox": [ + 55, + 205, + 296, + 266 + ], + "type": "text", + "content": "The vehicle trajectory is either provided or estimated using multi-sensor-based odometry. Our model " + }, + { + "bbox": [ + 55, + 205, + 296, + 266 + ], + "type": "inline_equation", + "content": "\\mathcal{SF}" + }, + { + "bbox": [ + 55, + 205, + 296, + 266 + ], + "type": "text", + "content": " aims to perform accurate 3D reconstruction and synthesize novel viewpoints for any given timestamp " + }, + { + "bbox": [ + 55, + 205, + 296, + 266 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 205, + 296, + 266 + ], + "type": "text", + "content": " and camera pose " + }, + { + "bbox": [ + 55, + 205, + 296, + 266 + ], + "type": "inline_equation", + "content": "[E_t, I_t]" + }, + { + "bbox": [ + 55, + 205, + 296, + 266 + ], + "type": "text", + "content": " by rendering " + }, + { + "bbox": [ + 55, + 205, + 296, + 266 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{I}} = \\mathcal{SF}(E_t, I_t, t)" + }, + { + "bbox": [ + 55, + 205, + 296, + 266 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 270, + 212, + 283 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 270, + 212, + 283 + ], + "spans": [ + { + "bbox": [ + 55, + 270, + 212, + 283 + ], + "type": "text", + "content": "3.3. 4D Gaussian Representation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 287, + 296, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 287, + 296, + 395 + ], + "spans": [ + { + "bbox": [ + 55, + 287, + 296, + 395 + ], + "type": "text", + "content": "Given images with associated camera poses, 3DGS optimizes a set of anisotropic 3D Gaussians through differentiable rasterization to represent a static 3D scene. We extend 3DGS to 4DGS of spatial-temporal representations for dynamic objects in an urban scene. Each 4D Gaussian primitives " + }, + { + "bbox": [ + 55, + 287, + 296, + 395 + ], + "type": "inline_equation", + "content": "\\mathcal{G}(t)" + }, + { + "bbox": [ + 55, + 287, + 296, + 395 + ], + "type": "text", + "content": " is represented by time-varying attributes: 3D center " + }, + { + "bbox": [ + 55, + 287, + 296, + 395 + ], + "type": "inline_equation", + "content": "\\mu(t) = [x(t), y(t), z(t)]^T" + }, + { + "bbox": [ + 55, + 287, + 296, + 395 + ], + "type": "text", + "content": " and covariance matrix " + }, + { + "bbox": [ + 55, + 287, + 296, + 395 + ], + "type": "inline_equation", + "content": "\\Sigma(t)" + }, + { + "bbox": [ + 55, + 287, + 296, + 395 + ], + "type": "text", + "content": " along with time-invariant attributes: opacity " + }, + { + "bbox": [ + 55, + 287, + 296, + 395 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 55, + 287, + 296, + 395 + ], + "type": "text", + "content": " and color " + }, + { + "bbox": [ + 55, + 287, + 296, + 395 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 287, + 296, + 395 + ], + "type": "text", + "content": ". The density of a Gaussian at point " + }, + { + "bbox": [ + 55, + 287, + 296, + 395 + ], + "type": "inline_equation", + "content": "(x,t)" + }, + { + "bbox": [ + 55, + 287, + 296, + 395 + ], + "type": "text", + "content": " is defined as," + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 65, + 395, + 294, + 415 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 395, + 294, + 415 + ], + "spans": [ + { + "bbox": [ + 65, + 395, + 294, + 415 + ], + "type": "interline_equation", + "content": "\\alpha (t) = \\sigma \\cdot e x p \\left(- \\frac {1}{2} (x - \\mu (t)) ^ {T} \\Sigma (t) ^ {- 1} (x - \\mu (t))\\right). \\tag {1}", + "image_path": "bd6e3bed47c957dd4293c5e308c0d8cb6e39b342c8760a29ade6581da06bb7a1.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 415, + 295, + 463 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 415, + 295, + 463 + ], + "spans": [ + { + "bbox": [ + 55, + 415, + 295, + 463 + ], + "type": "text", + "content": "The covariance matrix " + }, + { + "bbox": [ + 55, + 415, + 295, + 463 + ], + "type": "inline_equation", + "content": "\\Sigma(t) = R(t)SS^T R(t)^T" + }, + { + "bbox": [ + 55, + 415, + 295, + 463 + ], + "type": "text", + "content": " is composed of a scaling matrix " + }, + { + "bbox": [ + 55, + 415, + 295, + 463 + ], + "type": "inline_equation", + "content": "S = \\text{diag}(s_x, s_y, s_z)" + }, + { + "bbox": [ + 55, + 415, + 295, + 463 + ], + "type": "text", + "content": " and rotation matrix " + }, + { + "bbox": [ + 55, + 415, + 295, + 463 + ], + "type": "inline_equation", + "content": "R(t) = (q_x(t), q_y(t), q_z(t), q_w(t))" + }, + { + "bbox": [ + 55, + 415, + 295, + 463 + ], + "type": "text", + "content": ", constrained to be a positive semi-definite matrix during optimization." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 464, + 296, + 546 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 464, + 296, + 546 + ], + "spans": [ + { + "bbox": [ + 55, + 464, + 296, + 546 + ], + "type": "text", + "content": "To render an image from a specific viewpoint at times-tamp " + }, + { + "bbox": [ + 55, + 464, + 296, + 546 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 464, + 296, + 546 + ], + "type": "text", + "content": ", we first transform all the 4D Gaussians from other times to the target time " + }, + { + "bbox": [ + 55, + 464, + 296, + 546 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 464, + 296, + 546 + ], + "type": "text", + "content": ", according to the learned correspondence from NMFF. Subsequently, the aggregated 4D Gaussians are splatted onto the image plane, resulting in a collection of 2D Gaussians. The 3D covariance matrix " + }, + { + "bbox": [ + 55, + 464, + 296, + 546 + ], + "type": "inline_equation", + "content": "\\Sigma" + }, + { + "bbox": [ + 55, + 464, + 296, + 546 + ], + "type": "text", + "content": " is projected to a 2D covariance matrix " + }, + { + "bbox": [ + 55, + 464, + 296, + 546 + ], + "type": "inline_equation", + "content": "\\Sigma'" + }, + { + "bbox": [ + 55, + 464, + 296, + 546 + ], + "type": "text", + "content": " by," + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 125, + 546, + 294, + 562 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 125, + 546, + 294, + 562 + ], + "spans": [ + { + "bbox": [ + 125, + 546, + 294, + 562 + ], + "type": "interline_equation", + "content": "\\Sigma^ {\\prime} (t) = J E \\Sigma (t) E ^ {T} J ^ {T}, \\tag {2}", + "image_path": "13c57cc185a8134d2e93b748010be42c4be2296224b9e8be77f4dfb1f630187b.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 563, + 295, + 587 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 563, + 295, + 587 + ], + "spans": [ + { + "bbox": [ + 55, + 563, + 295, + 587 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 563, + 295, + 587 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 55, + 563, + 295, + 587 + ], + "type": "text", + "content": " refers to the world-to-camera matrix and " + }, + { + "bbox": [ + 55, + 563, + 295, + 587 + ], + "type": "inline_equation", + "content": "J" + }, + { + "bbox": [ + 55, + 563, + 295, + 587 + ], + "type": "text", + "content": " refers to Jacobian of the perspective transformation." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 587, + 296, + 623 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 587, + 296, + 623 + ], + "spans": [ + { + "bbox": [ + 55, + 587, + 296, + 623 + ], + "type": "text", + "content": "By sorting the Gaussians according to their depth within the camera space, we use " + }, + { + "bbox": [ + 55, + 587, + 296, + 623 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 587, + 296, + 623 + ], + "type": "text", + "content": "-blending to estimate the color " + }, + { + "bbox": [ + 55, + 587, + 296, + 623 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 55, + 587, + 296, + 623 + ], + "type": "text", + "content": " and depth " + }, + { + "bbox": [ + 55, + 587, + 296, + 623 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 55, + 587, + 296, + 623 + ], + "type": "text", + "content": " at each pixel " + }, + { + "bbox": [ + 55, + 587, + 296, + 623 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 55, + 587, + 296, + 623 + ], + "type": "text", + "content": " as," + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 60, + 622, + 295, + 654 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 622, + 295, + 654 + ], + "spans": [ + { + "bbox": [ + 60, + 622, + 295, + 654 + ], + "type": "interline_equation", + "content": "C = \\sum_ {i = 1} ^ {N} c _ {i} \\alpha_ {i} \\prod_ {j = 1} ^ {i - 1} (1 - \\alpha_ {j}), D = \\sum_ {i = 1} ^ {N} z _ {i} \\alpha_ {i} \\prod_ {j = 1} ^ {i - 1} (1 - \\alpha_ {j}). \\tag {3}", + "image_path": "41c98dd42efbe70d0db5881486718c7774eb0d980524f117e8f2812d9a20ca53.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": "Here " + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "inline_equation", + "content": "c_{i}" + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": " represents the color of Gaussian " + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{i}" + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": ", which is computed using spherical harmonic coefficients. " + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "inline_equation", + "content": "z_{i}" + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": " represents the distance from the image plane to the " + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": "-th Gaussian center. " + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\alpha_{i}" + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": " denotes the density computed from the 2D projection of the " + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": "-th Gaussian and its learned opacity." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 204, + 459, + 216 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 204, + 459, + 216 + ], + "spans": [ + { + "bbox": [ + 313, + 204, + 459, + 216 + ], + "type": "text", + "content": "3.4. Neural Motion Flow Field" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 222, + 555, + 306 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 222, + 555, + 306 + ], + "spans": [ + { + "bbox": [ + 313, + 222, + 555, + 306 + ], + "type": "text", + "content": "To establish the correspondence of dynamic objects, we design NMFF to model 3D motions as a continuous motion flow field, enabling the temporal transition of both 3D points and Gaussians. The NMFF " + }, + { + "bbox": [ + 313, + 222, + 555, + 306 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 313, + 222, + 555, + 306 + ], + "type": "text", + "content": " contains a temporal sequence of motion flow field, defined as " + }, + { + "bbox": [ + 313, + 222, + 555, + 306 + ], + "type": "inline_equation", + "content": "\\Phi = \\{\\phi_{t_i:t_{i + 1}}\\} ,i = 0,1,2,\\ldots ,n - 1" + }, + { + "bbox": [ + 313, + 222, + 555, + 306 + ], + "type": "text", + "content": " , which predicts 3D motion flow of an arbitrary query point between two consecutive frames as," + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 319, + 309, + 553, + 334 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 309, + 553, + 334 + ], + "spans": [ + { + "bbox": [ + 319, + 309, + 553, + 334 + ], + "type": "interline_equation", + "content": "\\phi_ {t _ {1}: t _ {2}} \\left(x _ {t _ {1}}, y _ {t _ {1}}, z _ {t _ {1}}\\right) = \\Delta x _ {t _ {1}: t _ {2}}, \\Delta y _ {t _ {1}: t _ {2}}, \\Delta z _ {t _ {1}: t _ {2}}, \\Delta R _ {t _ {1}: t _ {2}} \\tag {4}", + "image_path": "f9d0f80f4aab8f151adaff59c5b8848d9a85725664f9eb7b2d4c6d51cf30b630.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 334, + 553, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 334, + 553, + 357 + ], + "spans": [ + { + "bbox": [ + 313, + 334, + 553, + 357 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 334, + 553, + 357 + ], + "type": "inline_equation", + "content": "\\Delta R_{t_1:t_2}" + }, + { + "bbox": [ + 313, + 334, + 553, + 357 + ], + "type": "text", + "content": " denotes motion angle between two adjacent timestamps. Each field " + }, + { + "bbox": [ + 313, + 334, + 553, + 357 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 313, + 334, + 553, + 357 + ], + "type": "text", + "content": " is eight ReLU-MLP stacks." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 358, + 554, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 358, + 554, + 441 + ], + "spans": [ + { + "bbox": [ + 313, + 358, + 554, + 441 + ], + "type": "text", + "content": "Point Cloud within NMFF: Although NMFF can be learned during Gaussian Splatting optimization, we observed that the rendering loss alone is insufficient to effectively constrain the motion representation of dynamic scene components. Therefore, we leverage the rich geometric structure inherent in temporally consecutive 3D LiDAR points to derive a robust NMFF prior, as shown in Fig. 3." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 442, + 554, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 442, + 554, + 525 + ], + "spans": [ + { + "bbox": [ + 313, + 442, + 554, + 525 + ], + "type": "text", + "content": "We first pre-train " + }, + { + "bbox": [ + 313, + 442, + 554, + 525 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 313, + 442, + 554, + 525 + ], + "type": "text", + "content": " on a sequence of 3D LiDAR points by forward and backward 3D geometry consistency, enabling NMFF to learn the 3D motion of each query 3D point over time. Given the source point clouds " + }, + { + "bbox": [ + 313, + 442, + 554, + 525 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 313, + 442, + 554, + 525 + ], + "type": "text", + "content": " and target point clouds " + }, + { + "bbox": [ + 313, + 442, + 554, + 525 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}" + }, + { + "bbox": [ + 313, + 442, + 554, + 525 + ], + "type": "text", + "content": " from different timestamps, we minimize the point distance between the source and target point clouds by a bidirectional Chamfer Distance " + }, + { + "bbox": [ + 313, + 442, + 554, + 525 + ], + "type": "inline_equation", + "content": "(CD)" + }, + { + "bbox": [ + 313, + 442, + 554, + 525 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 318, + 529, + 553, + 555 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 529, + 553, + 555 + ], + "spans": [ + { + "bbox": [ + 318, + 529, + 553, + 555 + ], + "type": "interline_equation", + "content": "d _ {\\mathrm {C D}} (\\mathcal {P}, \\mathcal {Q}) = \\sum_ {p \\in \\mathcal {P}} \\min _ {q \\in \\mathcal {Q}} \\| p - q \\| ^ {2} + \\sum_ {q \\in \\mathcal {Q}} \\min _ {p \\in \\mathcal {P}} \\| q - p \\| ^ {2}. \\tag {5}", + "image_path": "e1ed0cb2add6b1ac24e43ef53819037f47208a47ae471fc93b3498a3477f66e3.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 558, + 554, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 554, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 554, + 605 + ], + "type": "text", + "content": "The bidirectional " + }, + { + "bbox": [ + 313, + 558, + 554, + 605 + ], + "type": "inline_equation", + "content": "CD" + }, + { + "bbox": [ + 313, + 558, + 554, + 605 + ], + "type": "text", + "content": " is a point-based distance that establishes correspondences in both directions—source to target and target to source—by searching for the nearest neighbors in each point cloud." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 606, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 606, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 606, + 555, + 713 + ], + "type": "text", + "content": "We compensate for ego-car motion in LiDAR points to let NMFF learn only 3D motion flow during pre-training. As a result, the 3D motion flow predicted by NMFF excludes the ego-car motion. We generate a 3D dynamic mask to separate static and dynamic points by applying a threshold on each point's 3D motion flow. We then merge static point clouds using ego-motion alone, and dynamic point clouds using both ego-motion and motion flow. For the static background, we initialize the static 3D Gaus" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "27490" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 54, + 72, + 294, + 143 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 72, + 294, + 143 + ], + "spans": [ + { + "bbox": [ + 54, + 72, + 294, + 143 + ], + "type": "text", + "content": "sians using the merged 3D static points. For dynamic objects, we initialize 4D Gaussians by aggregating points from different timestamps into a common reference frame (the mid-timestamp reference frame), which increases density and completeness by integrating observations from various viewpoints over time." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 144, + 295, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 144, + 295, + 215 + ], + "spans": [ + { + "bbox": [ + 55, + 144, + 295, + 215 + ], + "type": "text", + "content": "4D Gaussian within NMFF: The pre-trained " + }, + { + "bbox": [ + 55, + 144, + 295, + 215 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 55, + 144, + 295, + 215 + ], + "type": "text", + "content": " is subsequently integrated with 4D Gaussian representations of dynamic objects, enabling joint optimization during Gaussian splatting. Given consecutive time " + }, + { + "bbox": [ + 55, + 144, + 295, + 215 + ], + "type": "inline_equation", + "content": "t_1" + }, + { + "bbox": [ + 55, + 144, + 295, + 215 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 144, + 295, + 215 + ], + "type": "inline_equation", + "content": "t_2" + }, + { + "bbox": [ + 55, + 144, + 295, + 215 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 55, + 144, + 295, + 215 + ], + "type": "inline_equation", + "content": "t_1 < t_2" + }, + { + "bbox": [ + 55, + 144, + 295, + 215 + ], + "type": "text", + "content": " with respective states " + }, + { + "bbox": [ + 55, + 144, + 295, + 215 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{G}_i(t_1)\\}" + }, + { + "bbox": [ + 55, + 144, + 295, + 215 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 144, + 295, + 215 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{G}_i(t_2)\\}" + }, + { + "bbox": [ + 55, + 144, + 295, + 215 + ], + "type": "text", + "content": ", these states are connected by NMFF for each Gaussian as," + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 108, + 216, + 294, + 229 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 216, + 294, + 229 + ], + "spans": [ + { + "bbox": [ + 108, + 216, + 294, + 229 + ], + "type": "interline_equation", + "content": "\\mathcal {G} _ {i} \\left(t _ {1}\\right) = \\left\\{\\mu \\left(t _ {1}\\right), R \\left(t _ {1}\\right), S, \\alpha , c \\right\\}, \\tag {6}", + "image_path": "73a8fefc90fa4e0473cecd8ed9ba791213fcf3e7e956c54b5afc97646dbc65c1.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 105, + 232, + 295, + 245 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 232, + 295, + 245 + ], + "spans": [ + { + "bbox": [ + 105, + 232, + 295, + 245 + ], + "type": "interline_equation", + "content": "\\Delta \\mu_ {t _ {1}: t _ {2}}, \\Delta R _ {t _ {1}: t _ {2}} = \\phi_ {t _ {1}: t _ {2}} (\\mu (t _ {1})), \\tag {7}", + "image_path": "887eaf626fa15a9e9350a9b670b3a4da1739f1241ed34ccaa87ef2cb94b5e96e.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 249, + 295, + 263 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 249, + 295, + 263 + ], + "spans": [ + { + "bbox": [ + 61, + 249, + 295, + 263 + ], + "type": "interline_equation", + "content": "\\hat {\\mathcal {G}} _ {i} \\left(t _ {2}\\right) = \\left\\{\\mu \\left(t _ {1}\\right) + \\Delta \\mu_ {t _ {1}: t _ {2}}, R \\left(t _ {1}\\right) \\cdot \\Delta R _ {t _ {1}: t _ {2}}, S, \\alpha , c \\right\\}. \\tag {8}", + "image_path": "844d9e9d795b6a53a78a8776e8ede31517945d0ecb140449b1e579f8e5eb4c70.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 266, + 295, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 266, + 295, + 384 + ], + "spans": [ + { + "bbox": [ + 55, + 266, + 295, + 384 + ], + "type": "text", + "content": "We employ NMFF " + }, + { + "bbox": [ + 55, + 266, + 295, + 384 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 55, + 266, + 295, + 384 + ], + "type": "text", + "content": " correspondence to warp the center of all aggregated 4D Gaussians across time to the desired timestamp, as " + }, + { + "bbox": [ + 55, + 266, + 295, + 384 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{G}_i(t)\\} = \\{\\mathcal{G}(t),\\hat{\\mathcal{G}} (t_1),\\hat{\\mathcal{G}} (t_2),\\dots ,\\hat{\\mathcal{G}} (t_n)\\}" + }, + { + "bbox": [ + 55, + 266, + 295, + 384 + ], + "type": "text", + "content": ". For the transition of Gaussians across multiple timestamps, we apply the corresponding NMFF through step-by-step propagation. In practice, we propagate the aggregated 4D Gaussian at the mid-timestamp to the target timestamp. Here NMFF learns a motion transition of the 4D Gaussian among training frames in a self-supervised manner, promoting a multi-view consistent representation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 385, + 295, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 385, + 295, + 459 + ], + "spans": [ + { + "bbox": [ + 55, + 385, + 295, + 459 + ], + "type": "text", + "content": "For static background, we represent it by static 3D Gaussians " + }, + { + "bbox": [ + 55, + 385, + 295, + 459 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{\\text{static}}" + }, + { + "bbox": [ + 55, + 385, + 295, + 459 + ], + "type": "text", + "content": " following the strategy [1]. For sky area, we adopt a separate environmental map with static sky Gaussians " + }, + { + "bbox": [ + 55, + 385, + 295, + 459 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{sky}" + }, + { + "bbox": [ + 55, + 385, + 295, + 459 + ], + "type": "text", + "content": " following [1, 25]. Given a specific viewpoint " + }, + { + "bbox": [ + 55, + 385, + 295, + 459 + ], + "type": "inline_equation", + "content": "[E_t, I_t, t]" + }, + { + "bbox": [ + 55, + 385, + 295, + 459 + ], + "type": "text", + "content": ", the image " + }, + { + "bbox": [ + 55, + 385, + 295, + 459 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{I}}" + }, + { + "bbox": [ + 55, + 385, + 295, + 459 + ], + "type": "text", + "content": " and depth " + }, + { + "bbox": [ + 55, + 385, + 295, + 459 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{D}}" + }, + { + "bbox": [ + 55, + 385, + 295, + 459 + ], + "type": "text", + "content": " at time " + }, + { + "bbox": [ + 55, + 385, + 295, + 459 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 385, + 295, + 459 + ], + "type": "text", + "content": " are obtained by the differentiable rendering using a set of Gaussians," + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 70, + 463, + 295, + 477 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 463, + 295, + 477 + ], + "spans": [ + { + "bbox": [ + 70, + 463, + 295, + 477 + ], + "type": "interline_equation", + "content": "\\hat {\\mathcal {L}}, \\hat {\\mathcal {D}} = \\operatorname {R e n d e r} \\left(\\left\\{\\mathcal {G} _ {i} (t) \\right\\}, \\mathcal {G} _ {\\text {s t a t i c}}, \\mathcal {G} _ {\\text {s k y}} \\mid E _ {t}, I _ {t}, t\\right). \\tag {9}", + "image_path": "7bf74bf96599ebd408efa2653e0eecd732ffb28b8b91e7226b46ff5a762cd333.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 484, + 194, + 497 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 484, + 194, + 497 + ], + "spans": [ + { + "bbox": [ + 55, + 484, + 194, + 497 + ], + "type": "text", + "content": "3.5. Optical Flow Distillation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 502, + 295, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 502, + 295, + 561 + ], + "spans": [ + { + "bbox": [ + 55, + 502, + 295, + 561 + ], + "type": "text", + "content": "The 3D motion flow provided by NMFF enables 2D optical flow rendering. Given two consecutive time " + }, + { + "bbox": [ + 55, + 502, + 295, + 561 + ], + "type": "inline_equation", + "content": "t_1" + }, + { + "bbox": [ + 55, + 502, + 295, + 561 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 502, + 295, + 561 + ], + "type": "inline_equation", + "content": "t_2" + }, + { + "bbox": [ + 55, + 502, + 295, + 561 + ], + "type": "text", + "content": ", the optical flow " + }, + { + "bbox": [ + 55, + 502, + 295, + 561 + ], + "type": "inline_equation", + "content": "f_{t_1:t_2}" + }, + { + "bbox": [ + 55, + 502, + 295, + 561 + ], + "type": "text", + "content": " of each Gaussian is computed by projecting 3D center " + }, + { + "bbox": [ + 55, + 502, + 295, + 561 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 55, + 502, + 295, + 561 + ], + "type": "text", + "content": " to the image plane using the camera's intrinsic matrix " + }, + { + "bbox": [ + 55, + 502, + 295, + 561 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 55, + 502, + 295, + 561 + ], + "type": "text", + "content": " and extrinsic matrix " + }, + { + "bbox": [ + 55, + 502, + 295, + 561 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 55, + 502, + 295, + 561 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 79, + 563, + 294, + 578 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 563, + 294, + 578 + ], + "spans": [ + { + "bbox": [ + 79, + 563, + 294, + 578 + ], + "type": "interline_equation", + "content": "\\mu^ {\\prime} \\left(t _ {1}\\right) = I [ E ] \\mu \\left(t _ {1}\\right), \\quad \\mu^ {\\prime} \\left(t _ {2}\\right) = I [ E ] \\mu \\left(t _ {2}\\right), \\tag {10}", + "image_path": "aebe8456a178ee94ad40c5ae140e328fc8fe8d782febaefabd339a08e475d338.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 124, + 579, + 294, + 593 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 124, + 579, + 294, + 593 + ], + "spans": [ + { + "bbox": [ + 124, + 579, + 294, + 593 + ], + "type": "interline_equation", + "content": "f _ {t _ {1}: t _ {2}} = \\mu^ {\\prime} (t _ {2}) - \\mu^ {\\prime} (t _ {1}). \\tag {11}", + "image_path": "c238ec1f5c045abaa513dab1b55f09eff660746338c828355499611ac851bca1.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 595, + 295, + 618 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 595, + 295, + 618 + ], + "spans": [ + { + "bbox": [ + 55, + 595, + 295, + 618 + ], + "type": "text", + "content": "Then we estimate the optical flow " + }, + { + "bbox": [ + 55, + 595, + 295, + 618 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 55, + 595, + 295, + 618 + ], + "type": "text", + "content": " at pixel " + }, + { + "bbox": [ + 55, + 595, + 295, + 618 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 55, + 595, + 295, + 618 + ], + "type": "text", + "content": " via by point-based " + }, + { + "bbox": [ + 55, + 595, + 295, + 618 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 595, + 295, + 618 + ], + "type": "text", + "content": "-blending as," + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 119, + 618, + 294, + 652 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 119, + 618, + 294, + 652 + ], + "spans": [ + { + "bbox": [ + 119, + 618, + 294, + 652 + ], + "type": "interline_equation", + "content": "F = \\sum_ {i = 1} ^ {N} f _ {i} \\alpha_ {i} \\prod_ {j = 1} ^ {i - 1} (1 - \\alpha_ {j}). \\tag {12}", + "image_path": "76dda06c053bdc365b9b76d94b201ac2d8b5f3335f29d659c04a818ff663db15.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 653, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 295, + 713 + ], + "type": "text", + "content": "The optical flow " + }, + { + "bbox": [ + 55, + 653, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{F}}" + }, + { + "bbox": [ + 55, + 653, + 295, + 713 + ], + "type": "text", + "content": " at time " + }, + { + "bbox": [ + 55, + 653, + 295, + 713 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 653, + 295, + 713 + ], + "type": "text", + "content": " are rendered by the differentiable rendering similar as Equation 9. In order to distill optical flow knowledge from 2D foundation model into 4D space, we adopt SEA-RAFT[20] to extract optical flow " + }, + { + "bbox": [ + 55, + 653, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 55, + 653, + 295, + 713 + ], + "type": "text", + "content": " as pseudo ground truth for distillation by optical flow loss," + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 321, + 71, + 553, + 86 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 71, + 553, + 86 + ], + "spans": [ + { + "bbox": [ + 321, + 71, + 553, + 86 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {F} = \\lambda_ {f} \\mathcal {L} _ {1} (\\mathcal {F}, \\hat {\\mathcal {F}}) + (1 - \\lambda_ {f}) \\mathcal {L} _ {1} \\left(\\mathcal {I} _ {\\text {n e x t}}, \\mathcal {T} (\\hat {\\mathcal {I}} | \\hat {\\mathcal {F}})\\right), \\tag {13}", + "image_path": "08dda4cee6a51d09a305294b05a566962fad0407ee597983e26e1cd20af2a505.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 89, + 553, + 149 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 89, + 553, + 149 + ], + "spans": [ + { + "bbox": [ + 313, + 89, + 553, + 149 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 89, + 553, + 149 + ], + "type": "inline_equation", + "content": "\\lambda_{f}" + }, + { + "bbox": [ + 313, + 89, + 553, + 149 + ], + "type": "text", + "content": " is a scale factor, and " + }, + { + "bbox": [ + 313, + 89, + 553, + 149 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_{\\text{next}}" + }, + { + "bbox": [ + 313, + 89, + 553, + 149 + ], + "type": "text", + "content": " denotes the image at the next time step. " + }, + { + "bbox": [ + 313, + 89, + 553, + 149 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 313, + 89, + 553, + 149 + ], + "type": "text", + "content": " is used to warp rendered image to next time step according to optical flow. The optical flow " + }, + { + "bbox": [ + 313, + 89, + 553, + 149 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 313, + 89, + 553, + 149 + ], + "type": "text", + "content": " extracted from foundation model is only required during training and is not needed during inference." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 156, + 399, + 168 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 156, + 399, + 168 + ], + "spans": [ + { + "bbox": [ + 313, + 156, + 399, + 168 + ], + "type": "text", + "content": "3.6. Optimization" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 173, + 555, + 233 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 173, + 555, + 233 + ], + "spans": [ + { + "bbox": [ + 313, + 173, + 555, + 233 + ], + "type": "text", + "content": "All Gaussian attributes, along with the parameters of NMFF, are optimized end-to-end in a self-supervised manner. Meanwhile, adaptive densification and pruning strategies are introduced to enhance the fitting of the 3D scene. The overall training loss is given as," + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 329, + 236, + 553, + 251 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 329, + 236, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 329, + 236, + 553, + 251 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {I} + \\lambda_ {1} \\mathcal {L} _ {D} + \\lambda_ {2} \\mathcal {L} _ {F} + \\lambda_ {3} \\mathcal {L} _ {s k y} + \\lambda_ {4} \\mathcal {L} _ {r e g}, \\tag {14}", + "image_path": "d654d89176e1fe1a963f3e2ab832df1f257f5f75eead31185ecf39ef254507db.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 254, + 553, + 290 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 254, + 553, + 290 + ], + "spans": [ + { + "bbox": [ + 313, + 254, + 553, + 290 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 254, + 553, + 290 + ], + "type": "inline_equation", + "content": "\\lambda_{1,2,3,4}" + }, + { + "bbox": [ + 313, + 254, + 553, + 290 + ], + "type": "text", + "content": " are scale factors and " + }, + { + "bbox": [ + 313, + 254, + 553, + 290 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{reg}" + }, + { + "bbox": [ + 313, + 254, + 553, + 290 + ], + "type": "text", + "content": " is the regularization term. The image loss " + }, + { + "bbox": [ + 313, + 254, + 553, + 290 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_I" + }, + { + "bbox": [ + 313, + 254, + 553, + 290 + ], + "type": "text", + "content": " combines L1 and SSIM losses between rendered and observed images," + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 325, + 293, + 553, + 308 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 293, + 553, + 308 + ], + "spans": [ + { + "bbox": [ + 325, + 293, + 553, + 308 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {I} = \\left(1 - \\lambda_ {s s i m}\\right) \\mathcal {L} _ {1} (\\mathcal {I}, \\hat {\\mathcal {I}}) + \\lambda_ {s s i m} \\mathcal {L} _ {s s i m} (\\mathcal {I}, \\hat {\\mathcal {I}}), \\tag {15}", + "image_path": "6b0c0096e6980a6b5cb1ab3e37a75365bf810c33c4a9ca019620b99e3fbaca80.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 312, + 553, + 347 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 312, + 553, + 347 + ], + "spans": [ + { + "bbox": [ + 313, + 312, + 553, + 347 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 312, + 553, + 347 + ], + "type": "inline_equation", + "content": "\\lambda_{ssim}" + }, + { + "bbox": [ + 313, + 312, + 553, + 347 + ], + "type": "text", + "content": " is a scale factor. The depth loss " + }, + { + "bbox": [ + 313, + 312, + 553, + 347 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_D" + }, + { + "bbox": [ + 313, + 312, + 553, + 347 + ], + "type": "text", + "content": " is a L1 loss between inverse of rendered depth and generated depth by projecting sparse LiDAR points onto camera plane as," + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 398, + 350, + 553, + 364 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 398, + 350, + 553, + 364 + ], + "spans": [ + { + "bbox": [ + 398, + 350, + 553, + 364 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {D} = \\mathcal {L} _ {1} (\\mathcal {D}, \\hat {\\mathcal {D}}). \\tag {16}", + "image_path": "c0f5171125b0216606fb4d819d25f04bd333b7fe144bab3b7f70d3e27dd0d857.jpg" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 368, + 553, + 393 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 368, + 553, + 393 + ], + "spans": [ + { + "bbox": [ + 313, + 368, + 553, + 393 + ], + "type": "text", + "content": "The sky opacity loss " + }, + { + "bbox": [ + 313, + 368, + 553, + 393 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{sky}" + }, + { + "bbox": [ + 313, + 368, + 553, + 393 + ], + "type": "text", + "content": " is a binary cross entropy loss for sky refining using sky mask " + }, + { + "bbox": [ + 313, + 368, + 553, + 393 + ], + "type": "inline_equation", + "content": "M_{sky}" + }, + { + "bbox": [ + 313, + 368, + 553, + 393 + ], + "type": "text", + "content": " from SegFormer [22]," + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 375, + 397, + 553, + 411 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 375, + 397, + 553, + 411 + ], + "spans": [ + { + "bbox": [ + 375, + 397, + 553, + 411 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {s k y} = \\mathcal {L} _ {b c e} (O, 1 - M _ {s k y}), \\tag {17}", + "image_path": "8f6a4995b0d88feaa46aebd1bac6356469f691e5007e2255f38a93eb2f03516e.jpg" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 415, + 553, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 415, + 553, + 452 + ], + "spans": [ + { + "bbox": [ + 313, + 415, + 553, + 452 + ], + "type": "text", + "content": "where the accumulated opacity denotes as " + }, + { + "bbox": [ + 313, + 415, + 553, + 452 + ], + "type": "inline_equation", + "content": "O = \\sum_{i=1}^{N} \\alpha_i \\prod_{j=1}^{i-1} (1 - \\alpha_j)" + }, + { + "bbox": [ + 313, + 415, + 553, + 452 + ], + "type": "text", + "content": ". Note the sky mask is only required during training and is not needed during inference." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 461, + 395, + 475 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 461, + 395, + 475 + ], + "spans": [ + { + "bbox": [ + 313, + 461, + 395, + 475 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 313, + 480, + 484, + 492 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 480, + 484, + 492 + ], + "spans": [ + { + "bbox": [ + 313, + 480, + 484, + 492 + ], + "type": "text", + "content": "4.1. Datasets, Metrics and Baselines" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 313, + 498, + 555, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 555, + 593 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 555, + 593 + ], + "type": "text", + "content": "Following PVG [1], we validate our approach on two widely-used datasets in autonomous driving: Waymo Open Dataset [16] and KITTI Dataset [4]. Both offer multi-sensor data clips including synchronized and calibrated camera and LiDAR data from urban driving scenarios. We train and test our methods, along with all baselines, using full-resolution images: " + }, + { + "bbox": [ + 313, + 498, + 555, + 593 + ], + "type": "inline_equation", + "content": "1920 \\times 1280" + }, + { + "bbox": [ + 313, + 498, + 555, + 593 + ], + "type": "text", + "content": " for the Waymo dataset and " + }, + { + "bbox": [ + 313, + 498, + 555, + 593 + ], + "type": "inline_equation", + "content": "1242 \\times 375" + }, + { + "bbox": [ + 313, + 498, + 555, + 593 + ], + "type": "text", + "content": " for the KITTI dataset." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 313, + 594, + 553, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 553, + 654 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 553, + 654 + ], + "type": "text", + "content": "Following prior research [1, 19, 24, 25, 29], we evaluate image reconstruction and novel view synthesis using three widely-adopted benchmark metrics: peak signal-to-noise ratio (PSNR), structural similarity index measure (SSIM), and learned perceptual image patch similarity (LPIPS)." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "text", + "content": "We compare our method with extensive baselines including S-NeRF [23], StreetSurf [5], 3DGS [7], NSG [12], Mars [21], SUDS [19], EmerNerf [25], PVG [1], Street-Gaussian [24] on the Waymo and KITTI benchmarks. See supplementary material for more details and results." + } + ] + } + ], + "index": 31 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27491" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 69, + 541, + 469 + ], + "blocks": [ + { + "bbox": [ + 56, + 69, + 541, + 469 + ], + "lines": [ + { + "bbox": [ + 56, + 69, + 541, + 469 + ], + "spans": [ + { + "bbox": [ + 56, + 69, + 541, + 469 + ], + "type": "image", + "image_path": "fd1fd9d56f6988cf70fa3339459f5f08f8fd50ec5bac41f7f5ef742b84cf96cf.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 72, + 470, + 536, + 481 + ], + "lines": [ + { + "bbox": [ + 72, + 470, + 536, + 481 + ], + "spans": [ + { + "bbox": [ + 72, + 470, + 536, + 481 + ], + "type": "text", + "content": "Figure 4. Visual comparison of novel view synthesis on Waymo dataset. Bounding boxes indicate the zoomed-in dynamic areas." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 56, + 485, + 294, + 606 + ], + "blocks": [ + { + "bbox": [ + 56, + 485, + 294, + 606 + ], + "lines": [ + { + "bbox": [ + 56, + 485, + 294, + 606 + ], + "spans": [ + { + "bbox": [ + 56, + 485, + 294, + 606 + ], + "type": "image", + "image_path": "9ea1161dc35b3efefc018eb221f77ade5fd1ae45ac1eef9996c8d66463a9a8c8.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 56, + 607, + 293, + 617 + ], + "lines": [ + { + "bbox": [ + 56, + 607, + 293, + 617 + ], + "spans": [ + { + "bbox": [ + 56, + 607, + 293, + 617 + ], + "type": "text", + "content": "Figure 5. Dynamic object decomposition comparison on Waymo." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 621, + 249, + 634 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 621, + 249, + 634 + ], + "spans": [ + { + "bbox": [ + 55, + 621, + 249, + 634 + ], + "type": "text", + "content": "4.2. Evaluation on Waymo Open Dataset" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 641, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 296, + 713 + ], + "type": "text", + "content": "Following PVG [1], we evaluated our SplatFlow against baselines using the Waymo Open dataset. Table 1 summarizes the average metrics for both image reconstruction and novel view synthesis tasks on selected dynamic scenes. SplatFlow consistently outperforms all baselines across all evaluated metrics for both tasks. In image reconstruction," + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 317, + 485, + 564, + 587 + ], + "blocks": [ + { + "bbox": [ + 317, + 485, + 564, + 587 + ], + "lines": [ + { + "bbox": [ + 317, + 485, + 564, + 587 + ], + "spans": [ + { + "bbox": [ + 317, + 485, + 564, + 587 + ], + "type": "table", + "html": "
Image ReconstructionNovel View Synthesis
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
S-NeRF [23]19.670.5280.38719.220.5150.400
StreetSurf [5]26.700.8460.38723.780.8220.401
3DGS [7]27.990.8660.37225.080.8220.319
NSG [12]24.080.6560.29321.010.5710.487
Mars [21]21.810.6810.44120.690.6360.453
SUDS [19]28.830.8050.43021.830.6560.405
EmerNerf [25]28.110.7860.28925.920.7630.384
PVG [1]32.460.9100.37328.110.8490.279
SplatFlow33.640.9510.19828.710.8740.239
", + "image_path": "ea5e20e232bf5067f76b6951596e10379a0290fdd3c15f6a0dd666648a0cd67d.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 336, + 590, + 531, + 600 + ], + "lines": [ + { + "bbox": [ + 336, + 590, + 531, + 600 + ], + "spans": [ + { + "bbox": [ + 336, + 590, + 531, + 600 + ], + "type": "text", + "content": "Table 1. Performance comparison on Waymo dataset." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "content": "SplatFlow achieves high image quality with the highest scores: PSNR at 33.64, SSIM at 0.951, and LPIPS at 0.198. For novel view synthesis, SplatFlow produces high-quality renderings of unseen timestamps, reaching PSNR of 28.71, SSIM of 0.874 and LPIPS of 0.239. These improvements are visually validated in Fig. 4 and Fig. 9 (dynamic only), showing exceptional clarity in details for both static and dynamic regions. The dynamic objects in Fig. 9 are selected from a distant background car in an extremely zoomed-in" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27492" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 69, + 555, + 309 + ], + "blocks": [ + { + "bbox": [ + 57, + 69, + 555, + 309 + ], + "lines": [ + { + "bbox": [ + 57, + 69, + 555, + 309 + ], + "spans": [ + { + "bbox": [ + 57, + 69, + 555, + 309 + ], + "type": "image", + "image_path": "428c433cf61a8afe8c8e3d527ba4d11a397f89fb8cd48d1d0f2620400fb0b2bf.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 75, + 311, + 534, + 321 + ], + "lines": [ + { + "bbox": [ + 75, + 311, + 534, + 321 + ], + "spans": [ + { + "bbox": [ + 75, + 311, + 534, + 321 + ], + "type": "text", + "content": "Figure 6. Visual comparison of novel view synthesis on KITTI dataset. Bounding boxes indicate the zoomed-in dynamic areas." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 63, + 323, + 285, + 437 + ], + "blocks": [ + { + "bbox": [ + 63, + 323, + 285, + 437 + ], + "lines": [ + { + "bbox": [ + 63, + 323, + 285, + 437 + ], + "spans": [ + { + "bbox": [ + 63, + 323, + 285, + 437 + ], + "type": "table", + "html": "
PVG [1]SplatFlow
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
Seg. 1058...26.860.8400.27327.460.8800.253
Seg. 2259...26.450.8240.30929.300.8560.290
Seg. 7670...26.580.8200.30728.890.8550.283
Seg. 5083...28.350.8930.21330.290.9280.169
PSNR*↑SSIM*↑LPIPS*↓PSNR*↑SSIM*↑LPIPS*↓
Seg. 1058...27.420.8350.26728.500.8770.234
Seg. 2259...22.550.8030.35031.610.8990.191
Seg. 7670...23.560.7560.33028.760.8200.289
Seg. 5083...25.580.8140.29026.710.8550.211
", + "image_path": "6507cc75f14291c35ede8eee2ff57164abad78aad9bf81e589335a565189e3d4.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 58, + 464, + 306, + 565 + ], + "blocks": [ + { + "bbox": [ + 55, + 437, + 295, + 460 + ], + "lines": [ + { + "bbox": [ + 55, + 437, + 295, + 460 + ], + "spans": [ + { + "bbox": [ + 55, + 437, + 295, + 460 + ], + "type": "text", + "content": "Table 2. Novel view synthesis results on Waymo dataset (*denotes dynamic elements only)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 464, + 306, + 565 + ], + "lines": [ + { + "bbox": [ + 58, + 464, + 306, + 565 + ], + "spans": [ + { + "bbox": [ + 58, + 464, + 306, + 565 + ], + "type": "table", + "html": "
Image ReconstructionNovel View Synthesis
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
S-NeRF [23]19.230.6640.19318.710.6060.352
StreetSurf [5]24.140.8190.25722.480.7630.304
3DGS [7]21.020.8110.20219.540.7760.224
NSG [12]26.660.8060.18621.530.6730.254
Mars [21]27.960.9000.18524.230.8450.160
SUDS [19]28.310.8760.18522.770.7970.171
EmerNeRF [25]26.950.8280.21825.240.8010.237
PVG [1]32.830.9370.07027.430.8960.114
SplatFlow33.370.9430.05728.320.9320.089
", + "image_path": "6bd6bd1e7ea16c36494d09c08b8e3ac4e96b431b7496b32e50237515bf80413e.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 80, + 567, + 270, + 578 + ], + "lines": [ + { + "bbox": [ + 80, + 567, + 270, + 578 + ], + "spans": [ + { + "bbox": [ + 80, + 567, + 270, + 578 + ], + "type": "text", + "content": "Table 3. Performance comparison on KITTI dataset." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 581, + 295, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 581, + 295, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 581, + 295, + 628 + ], + "type": "text", + "content": "view. Compared with baselines which show artifacts such as ghosting and blur in regions of dynamic elements, our method preserves high-quality textures and fine details, producing more accurate renderings at novel viewpoints." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "type": "text", + "content": "To further evaluate the rendering quality of dynamic objects against the primary baseline PVG, we selected additional four scenes featuring a higher presence of dynamic objects for further comparison. Table 2 displays the metrics for novel view synthesis, evaluated both on entire scenes and exclusively on dynamic objects (denoted as *). Splat-Flow consistently surpasses PVG in all selected scenes," + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 316, + 323, + 554, + 515 + ], + "blocks": [ + { + "bbox": [ + 316, + 323, + 554, + 515 + ], + "lines": [ + { + "bbox": [ + 316, + 323, + 554, + 515 + ], + "spans": [ + { + "bbox": [ + 316, + 323, + 554, + 515 + ], + "type": "image", + "image_path": "336bbec7d5ff23eea71f5723ec3dff7eab8e59b8e18489509cb111d4e75b7a55.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 517, + 555, + 537 + ], + "lines": [ + { + "bbox": [ + 313, + 517, + 555, + 537 + ], + "spans": [ + { + "bbox": [ + 313, + 517, + 555, + 537 + ], + "type": "text", + "content": "Figure 7. Visual comparison of novel view synthesis on KITTI-25% (row1), 50% (row2), and 75% (row3) dataset." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 538, + 555, + 670 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 538, + 555, + 670 + ], + "spans": [ + { + "bbox": [ + 313, + 538, + 555, + 670 + ], + "type": "text", + "content": "demonstrating higher PSNR and SSIM values and lower LPIPS scores, underscoring its effectiveness in handling complex, dynamic content. The dynamic objects masks are obtained from 2D ground truth bounding boxes following [25]. The PSNR* of dynamic regions on novel view synthesis significantly outperforms SVG both quantitatively and qualitatively, especially in scenes with fast ego-motion and moving objects in Seg. 2259 and Seg. 7670. We also present the visual comparison of dynamic object decomposition with SVG in Fig. 5, which illustrates that SplatFlow renders sharper and clearer image of dynamic objects." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 673, + 474, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 673, + 474, + 685 + ], + "spans": [ + { + "bbox": [ + 313, + 673, + 474, + 685 + ], + "type": "text", + "content": "4.3. Evaluation on KITTI Dataset" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "content": "Following PVG [1], we further evaluated our SplatFlow method in comparison to the baselines on KITTI dataset." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27493" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 67, + 70, + 365, + 160 + ], + "blocks": [ + { + "bbox": [ + 67, + 70, + 365, + 160 + ], + "lines": [ + { + "bbox": [ + 67, + 70, + 365, + 160 + ], + "spans": [ + { + "bbox": [ + 67, + 70, + 365, + 160 + ], + "type": "table", + "html": "
KITTI - 75%KITTI - 50%KITTI - 25%
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
NeRF [10]18.560.5570.55419.120.5870.49718.610.5700.510
NSG [12]21.530.6730.25421.260.6590.26620.000.6320.281
SUDS [19]22.770.7970.17123.120.8210.13520.760.7470.198
MARS [21]24.230.8450.16024.000.8010.16423.230.7560.177
3DGS [7]19.190.7370.17219.230.7390.17419.060.7300.180
PVG [1]27.430.8960.11425.920.8820.11422.550.8330.151
StreetGS [24]25.790.8440.08125.520.8410.08424.530.8240.090
SplatFlow28.320.9320.08927.900.9270.09326.100.8900.120
", + "image_path": "fabde9e27eb24c2c0923f35c501fdbc33f2afb1ad7ff45d9e405fc76d92f969d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 366, + 76, + 542, + 157 + ], + "blocks": [ + { + "bbox": [ + 66, + 163, + 365, + 184 + ], + "lines": [ + { + "bbox": [ + 66, + 163, + 365, + 184 + ], + "spans": [ + { + "bbox": [ + 66, + 163, + 365, + 184 + ], + "type": "text", + "content": "Table 4. Performance comparison of novel view synthesis on KITTI-75%, 50% and 25% dataset." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 366, + 76, + 542, + 157 + ], + "lines": [ + { + "bbox": [ + 366, + 76, + 542, + 157 + ], + "spans": [ + { + "bbox": [ + 366, + 76, + 542, + 157 + ], + "type": "table", + "html": "
PSNR↑SSIM↑LPIPS↓
w/o NMFF prior27.690.8630.282
w/o NMFF optimization28.140.8740.269
w/o optical flow distillation28.280.8770.252
Full28.990.8800.249
PSNR*↑SSIM*↑LPIPS*↓
w/o NMFF prior27.280.8270.317
w/o NMFF optimization27.970.8430.254
w/o optical flow distillation28.510.8610.232
Full28.900.8630.231
", + "image_path": "fc82824f760c923b49de30b58f1b8bdcdaec7b2d7c5e26c92b196aaadb347fad.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 365, + 158, + 541, + 179 + ], + "lines": [ + { + "bbox": [ + 365, + 158, + 541, + 179 + ], + "spans": [ + { + "bbox": [ + 365, + 158, + 541, + 179 + ], + "type": "text", + "content": "Table 5. Ablation study on Waymo dataset (*denotes dynamic elements only)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 58, + 188, + 176, + 268 + ], + "blocks": [ + { + "bbox": [ + 58, + 188, + 176, + 268 + ], + "lines": [ + { + "bbox": [ + 58, + 188, + 176, + 268 + ], + "spans": [ + { + "bbox": [ + 58, + 188, + 176, + 268 + ], + "type": "image", + "image_path": "b23fb79d066066217788bf69edd4a526f759dd80acd773c4f34f9c6f933bdf6f.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 94, + 268, + 134, + 275 + ], + "lines": [ + { + "bbox": [ + 94, + 268, + 134, + 275 + ], + "spans": [ + { + "bbox": [ + 94, + 268, + 134, + 275 + ], + "type": "text", + "content": "Ours (Full)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 176, + 188, + 294, + 268 + ], + "blocks": [ + { + "bbox": [ + 176, + 188, + 294, + 268 + ], + "lines": [ + { + "bbox": [ + 176, + 188, + 294, + 268 + ], + "spans": [ + { + "bbox": [ + 176, + 188, + 294, + 268 + ], + "type": "image", + "image_path": "c10f81fdc6b6a701e91d9ab9769e18015d37c53429188620372c08403c115e9d.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 214, + 268, + 260, + 276 + ], + "lines": [ + { + "bbox": [ + 214, + 268, + 260, + 276 + ], + "spans": [ + { + "bbox": [ + 214, + 268, + 260, + 276 + ], + "type": "text", + "content": "w/o Opt. flow" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 58, + 276, + 176, + 354 + ], + "blocks": [ + { + "bbox": [ + 58, + 276, + 176, + 354 + ], + "lines": [ + { + "bbox": [ + 58, + 276, + 176, + 354 + ], + "spans": [ + { + "bbox": [ + 58, + 276, + 176, + 354 + ], + "type": "image", + "image_path": "7984c352b65409864afbe33329dfdda80a0112e3397b395d7df45cc9af2194ac.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 84, + 356, + 144, + 365 + ], + "lines": [ + { + "bbox": [ + 84, + 356, + 144, + 365 + ], + "spans": [ + { + "bbox": [ + 84, + 356, + 144, + 365 + ], + "type": "text", + "content": "w/o NMFF optim." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 56, + 366, + 294, + 376 + ], + "lines": [ + { + "bbox": [ + 56, + 366, + 294, + 376 + ], + "spans": [ + { + "bbox": [ + 56, + 366, + 294, + 376 + ], + "type": "text", + "content": "Figure 8. Visual comparison of ablation study on Waymo dataset." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 176, + 276, + 294, + 354 + ], + "blocks": [ + { + "bbox": [ + 176, + 276, + 294, + 354 + ], + "lines": [ + { + "bbox": [ + 176, + 276, + 294, + 354 + ], + "spans": [ + { + "bbox": [ + 176, + 276, + 294, + 354 + ], + "type": "image", + "image_path": "9d2312c8543933d1ba7f9195ecfb5ffbcdf8a45b638585d4ff25448a60f82742.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 210, + 356, + 265, + 365 + ], + "lines": [ + { + "bbox": [ + 210, + 356, + 265, + 365 + ], + "spans": [ + { + "bbox": [ + 210, + 356, + 265, + 365 + ], + "type": "text", + "content": "w/o NMFF prior" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 54, + 388, + 295, + 530 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 388, + 295, + 530 + ], + "spans": [ + { + "bbox": [ + 54, + 388, + 295, + 530 + ], + "type": "text", + "content": "Table 3 presents the average metrics for both image reconstruction and novel view synthesis tasks on selected dynamic scenes characterized by extensive movement. SplatFlow outperforms all other methods across all metrics in both tasks, with particularly high scores in PSNR. SSIM and LPIPS, achieving 33.37, 0.943 and 0.057 for image reconstruction and 28.32, 0.932 and 0.089 for novel view synthesis. SplatFlow achieves exceptional results in novel view synthesis, as shown in Fig. 6. Our approach generates high-fidelity renderings that preserve fine details while accurately reconstructing dynamic areas where existing approaches struggle to maintain stability." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 54, + 531, + 295, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 531, + 295, + 651 + ], + "spans": [ + { + "bbox": [ + 54, + 531, + 295, + 651 + ], + "type": "text", + "content": "To evaluate robustness to training data size, we follow the settings of [21, 24] to evaluate our method using different train/test split configurations on the KITTI dataset. Table 4 and Fig. 7 provide the performance comparisons and visualizations. SplatFlow consistently outperforms all other methods across all dataset subsets, achieving the highest scores. Even with reduced training data, SplatFlow robustly delivers high-quality novel view synthesis across varying levels of data availability, highlighting its efficient use of available information and strong generalization capabilities." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 659, + 149, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 659, + 149, + 672 + ], + "spans": [ + { + "bbox": [ + 55, + 659, + 149, + 672 + ], + "type": "text", + "content": "4.4. Ablation Study" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "content": "To demonstrate the effectiveness of each component in SplatFlow, we performed ablation studies on the Waymo dataset. Our main contributions, NMFF prior, NMFF opti" + } + ] + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 316, + 188, + 375, + 376 + ], + "blocks": [ + { + "bbox": [ + 316, + 188, + 375, + 376 + ], + "lines": [ + { + "bbox": [ + 316, + 188, + 375, + 376 + ], + "spans": [ + { + "bbox": [ + 316, + 188, + 375, + 376 + ], + "type": "image", + "image_path": "0243cf949d0320044abb1b88aef349e7e5ef7d6cec9901131ee8bd565d40b8f8.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 328, + 376, + 366, + 384 + ], + "lines": [ + { + "bbox": [ + 328, + 376, + 366, + 384 + ], + "spans": [ + { + "bbox": [ + 328, + 376, + 366, + 384 + ], + "type": "text", + "content": "EmerNeRF" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 317, + 385, + 550, + 397 + ], + "lines": [ + { + "bbox": [ + 317, + 385, + 550, + 397 + ], + "spans": [ + { + "bbox": [ + 317, + 385, + 550, + 397 + ], + "type": "text", + "content": "Figure 9. Detail comparison of novel view synthesis on Waymo." + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 376, + 188, + 434, + 376 + ], + "blocks": [ + { + "bbox": [ + 376, + 188, + 434, + 376 + ], + "lines": [ + { + "bbox": [ + 376, + 188, + 434, + 376 + ], + "spans": [ + { + "bbox": [ + 376, + 188, + 434, + 376 + ], + "type": "image", + "image_path": "60cc694cc2abedebf79f2a9db38753c85c8a798657e9360f71b41e1244ec7422.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 396, + 376, + 413, + 384 + ], + "lines": [ + { + "bbox": [ + 396, + 376, + 413, + 384 + ], + "spans": [ + { + "bbox": [ + 396, + 376, + 413, + 384 + ], + "type": "text", + "content": "PVG" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 435, + 188, + 493, + 376 + ], + "blocks": [ + { + "bbox": [ + 435, + 188, + 493, + 376 + ], + "lines": [ + { + "bbox": [ + 435, + 188, + 493, + 376 + ], + "spans": [ + { + "bbox": [ + 435, + 188, + 493, + 376 + ], + "type": "image", + "image_path": "b4e74a1558ae485f8e44c8c1daad2b64cbfd1e718bdd456c26411786770e1179.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 446, + 376, + 482, + 384 + ], + "lines": [ + { + "bbox": [ + 446, + 376, + 482, + 384 + ], + "spans": [ + { + "bbox": [ + 446, + 376, + 482, + 384 + ], + "type": "text", + "content": "SplatFlow" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 494, + 188, + 553, + 376 + ], + "blocks": [ + { + "bbox": [ + 494, + 188, + 553, + 376 + ], + "lines": [ + { + "bbox": [ + 494, + 188, + 553, + 376 + ], + "spans": [ + { + "bbox": [ + 494, + 188, + 553, + 376 + ], + "type": "image", + "image_path": "b9d300c49a381b8f94c7574a230ecfab7f3b7ed1e1c7a80e8023b8fc493ea20f.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 515, + 376, + 531, + 384 + ], + "lines": [ + { + "bbox": [ + 515, + 376, + 531, + 384 + ], + "spans": [ + { + "bbox": [ + 515, + 376, + 531, + 384 + ], + "type": "text", + "content": "G.T." + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 406, + 555, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 406, + 555, + 538 + ], + "spans": [ + { + "bbox": [ + 313, + 406, + 555, + 538 + ], + "type": "text", + "content": "mization, and optical flow distillation are analyzed in this study to assess their individual impact. We trained four different variations: 1) SplatFlow without pretraining on LiDAR data; 2) SplatFlow without optimization with 4D Gaussians on image data; 3) SplatFlow without optical flow distillation from foundational model; 4) SplatFlow full method. Table 5 presents the average evaluation metrics across test scenes used in Table 2, showing results for both entire scenes and dynamic objects (indicated as *). Ablation study validates the contribution of each component to the overall performance." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 550, + 388, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 550, + 388, + 562 + ], + "spans": [ + { + "bbox": [ + 313, + 550, + 388, + 562 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "content": "In this paper, we propose SplatFlow, a novel self-supervised Dynamic Gaussian Splitting within Neural Motion Flow Fields (NMFF) for accurate reconstruction and real-time rendering in dynamic urban scenarios. The core idea of SplatFlow is to seamlessly integrate time-dependent 4D Gaussian representations within NMFF in a unified framework, where NMFF implicitly models the motions of dynamic components across time through self-supervision. Experimental evaluations show that SplatFlow outperforms state-of-the-art across all standard metrics on the Waymo Open and KITTI datasets, without requiring expensive annotations of dynamic object detection and tracking." + } + ] + } + ], + "index": 28 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "27494" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 188 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 188 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 188 + ], + "type": "text", + "content": "This work was partially supported by the intramural research program of the U.S. Department of Agriculture Natural Resources Conservation Service, Climate-Smart Commodities program, EFFICACI grant, #NR233A750004G044. The findings and conclusions in this work have not been formally disseminated by the U. S. Department of Agriculture and should not be construed to represent any agency determination or policy." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 196, + 115, + 209 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 196, + 115, + 209 + ], + "spans": [ + { + "bbox": [ + 56, + 196, + 115, + 209 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 216, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 216, + 296, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 216, + 296, + 259 + ], + "spans": [ + { + "bbox": [ + 61, + 216, + 296, + 259 + ], + "type": "text", + "content": "[1] Yurui Chen, Chun Gu, Junzhe Jiang, Xiatian Zhu, and Li Zhang. Periodic vibration gaussian: Dynamic urban scene reconstruction and real-time rendering. arXiv preprint arXiv:2311.18561, 20.1, 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 261, + 296, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 261, + 296, + 304 + ], + "spans": [ + { + "bbox": [ + 61, + 261, + 296, + 304 + ], + "type": "text", + "content": "[2] Alexey Dosovitskiy, German Ros, Felipe Codevilla, Antonio Lopez, and Vladlen Koltun. Carla: An open urban driving simulator. In Conference on robot learning, pages 1-16. PMLR, 2017. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 304, + 296, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 304, + 296, + 360 + ], + "spans": [ + { + "bbox": [ + 62, + 304, + 296, + 360 + ], + "type": "text", + "content": "[3] Tobias Fischer, Lorenzo Porzi, Samuel Rota Bulo, Marc Pollefeys, and Peter Kontschieder. Multi-level neural scene graphs for dynamic urban environments. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21125-21135, 2024. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 361, + 295, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 361, + 295, + 404 + ], + "spans": [ + { + "bbox": [ + 62, + 361, + 295, + 404 + ], + "type": "text", + "content": "[4] Andreas Geiger, Philip Lenz, and Raquel Urtasun. Are we ready for autonomous driving? the kitti vision benchmark suite. In 2012 IEEE conference on computer vision and pattern recognition, pages 3354-3361. IEEE, 2012. 2, 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 404, + 296, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 404, + 296, + 459 + ], + "spans": [ + { + "bbox": [ + 62, + 404, + 296, + 459 + ], + "type": "text", + "content": "[5] Jianfei Guo, Nianchen Deng, Xinyang Li, Yeqi Bai, Botian Shi, Chiyu Wang, Chenjing Ding, Dongliang Wang, and Yikang Li. Streetsurf: Extending multi-view implicit surface reconstruction to street views. arXiv preprint arXiv:2306.04988, 2023. 2, 5, 6, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 459, + 296, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 459, + 296, + 514 + ], + "spans": [ + { + "bbox": [ + 62, + 459, + 296, + 514 + ], + "type": "text", + "content": "[6] Antoine Guédon and Vincent Lepetit. Sugar: Surface-aligned gaussian splatting for efficient 3d mesh reconstruction and high-quality mesh rendering. In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5354-5363, 2024. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 514, + 296, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 514, + 296, + 558 + ], + "spans": [ + { + "bbox": [ + 62, + 514, + 296, + 558 + ], + "type": "text", + "content": "[7] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42 (4), 2023. 1, 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 559, + 296, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 559, + 296, + 635 + ], + "spans": [ + { + "bbox": [ + 62, + 559, + 296, + 635 + ], + "type": "text", + "content": "[8] Abhijit Kundu, Kyle Genova, Xiaqi Yin, Alireza Fathi, Caroline Pantofaru, Leonidas J Guibas, Andrea Tagliasacchi, Frank Dellaert, and Thomas Funkhouser. Panoptic neural fields: A semantic object-aware neural scene representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12871-12881, 2022. 1, 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 635, + 296, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 635, + 296, + 691 + ], + "spans": [ + { + "bbox": [ + 62, + 635, + 296, + 691 + ], + "type": "text", + "content": "[9] Fan Lu, Yan Xu, Guang Chen, Hongsheng Li, Kwan-Yee Lin, and Changjun Jiang. Urban radiance field representation with deformable neural mesh primitives. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 465-476, 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "type": "text", + "content": "[10] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf:" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 333, + 73, + 555, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 555, + 106 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 555, + 106 + ], + "type": "text", + "content": "Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 1, 2, 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 107, + 555, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 555, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 555, + 150 + ], + "type": "text", + "content": "[11] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph., 41(4):102:1-102:15, 2022. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 152, + 554, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 554, + 207 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 554, + 207 + ], + "type": "text", + "content": "[12] Julian Ost, Fahim Mannan, Nils Thuerey, Julian Knodt, and Felix Heide. Neural scene graphs for dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2856-2865, 2021. 1, 2, 5, 6, 7, 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 209, + 555, + 253 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 209, + 555, + 253 + ], + "spans": [ + { + "bbox": [ + 316, + 209, + 555, + 253 + ], + "type": "text", + "content": "[13] Julian Ost, Issam Laradji, Alejandro Newell, Yuval Bahat, and Felix Heide. Neural point light fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18419-18429, 2022. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 254, + 554, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 254, + 554, + 308 + ], + "spans": [ + { + "bbox": [ + 316, + 254, + 554, + 308 + ], + "type": "text", + "content": "[14] Konstantinos Rematas, Andrew Liu, Pratul P Srinivasan, Jonathan T Barron, Andrea Tagliasacchi, Thomas Funkhouser, and Vittorio Ferrari. Urban radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12932-12942, 2022. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 310, + 554, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 310, + 554, + 365 + ], + "spans": [ + { + "bbox": [ + 316, + 310, + 554, + 365 + ], + "type": "text", + "content": "[15] Shital Shah, Debadeepta Dey, Chris Lovett, and Ashish Kapoor. Airsim: High-fidelity visual and physical simulation for autonomous vehicles. In Field and Service Robotics: Results of the 11th International Conference, pages 621-635. Springer, 2018. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 366, + 554, + 433 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 366, + 554, + 433 + ], + "spans": [ + { + "bbox": [ + 316, + 366, + 554, + 433 + ], + "type": "text", + "content": "[16] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, et al. Scalability in perception for autonomous driving: Waymo open dataset. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2446-2454, 2020. 2, 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 434, + 554, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 434, + 554, + 499 + ], + "spans": [ + { + "bbox": [ + 316, + 434, + 554, + 499 + ], + "type": "text", + "content": "[17] Su Sun, Cheng Zhao, Yuliang Guo, Ruoyu Wang, Xinyu Huang, Yingjie Victor Chen, and Liu Ren. Behind the veil: Enhanced indoor 3d scene reconstruction with occluded surfaces completion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12744-12753, 2024. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 500, + 554, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 500, + 554, + 555 + ], + "spans": [ + { + "bbox": [ + 316, + 500, + 554, + 555 + ], + "type": "text", + "content": "[18] Adam Tonderski, Carl Lindström, Georg Hess, William Ljungbergh, Lennart Svensson, and Christoffer Petersson. Neurad: Neural rendering for autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14895-14904, 2024. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 557, + 554, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 557, + 554, + 612 + ], + "spans": [ + { + "bbox": [ + 316, + 557, + 554, + 612 + ], + "type": "text", + "content": "[19] Haithem Turki, Jason Y Zhang, Francesco Ferroni, and Deva Ramanan. Suds: Scalable urban dynamic scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12375-12385, 2023. 1, 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 613, + 554, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 613, + 554, + 646 + ], + "spans": [ + { + "bbox": [ + 316, + 613, + 554, + 646 + ], + "type": "text", + "content": "[20] Yihan Wang, Lahav Lipson, and Jia Deng. Sea-raft: Simple, efficient, accurate raft for optical flow. In European Conference on Computer Vision, pages 36-54. Springer, 2025. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 647, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 554, + 713 + ], + "type": "text", + "content": "[21] Zirui Wu, Tianyu Liu, Liyi Luo, Zhide Zhong, Jianteng Chen, Hongmin Xiao, Chao Hou, Haozhe Lou, Yuantao Chen, Runyi Yang, et al. Mars: An instance-aware, modular and realistic simulator for autonomous driving. In CAAI International Conference on Artificial Intelligence, pages 3-15. Springer, 2023. 1, 2, 5, 6, 7, 8" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27495" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 73, + 296, + 507 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "text", + "content": "[22] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. In Neural Information Processing Systems (NeurIPS), 2021. 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 118, + 295, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 118, + 295, + 162 + ], + "spans": [ + { + "bbox": [ + 55, + 118, + 295, + 162 + ], + "type": "text", + "content": "[23] Ziyang Xie, Junge Zhang, Wenye Li, Feihu Zhang, and Li Zhang. S-nerf: Neural radiance fields for street views. In International Conference on Learning Representations (ICLR), 2023. 1, 5, 6, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 163, + 296, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 163, + 296, + 217 + ], + "spans": [ + { + "bbox": [ + 57, + 163, + 296, + 217 + ], + "type": "text", + "content": "[24] Yunzhi Yan, Haotong Lin, Chenxu Zhou, Weijie Wang, Haiyang Sun, Kun Zhan, Xianpeng Lang, Xiaowei Zhou, and Sida Peng. Street gaussians for modeling dynamic urban scenes. arXiv preprint arXiv:2401.01339, 2024. 1, 2, 3, 5, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 219, + 294, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 219, + 294, + 273 + ], + "spans": [ + { + "bbox": [ + 57, + 219, + 294, + 273 + ], + "type": "text", + "content": "[25] Jiawei Yang, Boris Ivanovic, Or Litany, Xinshuo Weng, Seung Wook Kim, Boyi Li, Tong Che, Danfei Xu, Sanja Fidler, Marco Pavone, et al. Emernerf: Emergent spatial-temporal scene decomposition via self-supervision. arXiv preprint arXiv:2311.02077, 2023. 1, 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 57, + 274, + 294, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 274, + 294, + 306 + ], + "spans": [ + { + "bbox": [ + 57, + 274, + 294, + 306 + ], + "type": "text", + "content": "[26] Xiaoshuai Zhang, Sai Bi, Kalyan Sunkavalli, Hao Su, and Zexiang Xu. Nerfusion: Fusing radiance fields for large-scale scene reconstruction. CVPR, 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 57, + 308, + 294, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 308, + 294, + 373 + ], + "spans": [ + { + "bbox": [ + 57, + 308, + 294, + 373 + ], + "type": "text", + "content": "[27] Cheng Zhao, Su Sun, Ruoyu Wang, Yuliang Guo, Jun-Jun Wan, Zhou Huang, Xinyu Huang, Yingjie Victor Chen, and Liu Ren. Tclc-gs: Tightly coupled lidar-camera gaussian splatting for autonomous driving: Supplementary materials. In European Conference on Computer Vision, pages 91-106. Springer, 2024. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 57, + 375, + 294, + 439 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 375, + 294, + 439 + ], + "spans": [ + { + "bbox": [ + 57, + 375, + 294, + 439 + ], + "type": "text", + "content": "[28] Hongyu Zhou, Jiahao Shao, Lu Xu, Dongfeng Bai, Weichao Qiu, Bingbing Liu, Yue Wang, Andreas Geiger, and Yiyi Liao. Hugs: Holistic urban 3d scene understanding via gaussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21336-21345, 2024. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 58, + 441, + 294, + 507 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 441, + 294, + 507 + ], + "spans": [ + { + "bbox": [ + 58, + 441, + 294, + 507 + ], + "type": "text", + "content": "[29] Xiaoyu Zhou, Zhiwei Lin, Xiaojun Shan, Yongtao Wang, Deqing Sun, and Ming-Hsuan Yang. Drivinggaussian: Composite gaussian splatting for surrounding dynamic autonomous driving scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21634-21643, 2024. 2, 3, 5" + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "27496" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Splatter-360_ Generalizable 360 Gaussian Splatting for Wide-baseline Panoramic Images/695c9e66-9d36-4a85-a2d0-e99f9431d57c_content_list.json b/2025/Splatter-360_ Generalizable 360 Gaussian Splatting for Wide-baseline Panoramic Images/695c9e66-9d36-4a85-a2d0-e99f9431d57c_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..281d7598f7f38759b5a8e72ef15226d515aaac11 --- /dev/null +++ b/2025/Splatter-360_ Generalizable 360 Gaussian Splatting for Wide-baseline Panoramic Images/695c9e66-9d36-4a85-a2d0-e99f9431d57c_content_list.json @@ -0,0 +1,1423 @@ +[ + { + "type": "text", + "text": "Splatter-360: Generalizable $360^{\\circ}$ Gaussian Splating for Wide-baseline Panoramic Images", + "text_level": 1, + "bbox": [ + 142, + 128, + 854, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zheng Chen $^{1*†}$ , Chenming Wu $^{2*}$ , Zhelun Shen $^{2}$ , Chen Zhao $^{2}$ , Weicai Ye $^{3}$ , Haocheng Feng $^{2}$ , Errui Ding $^{2}$ , Song-Hai Zhang $^{1‡}$", + "bbox": [ + 207, + 202, + 787, + 241 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Tsinghua University", + "bbox": [ + 411, + 244, + 584, + 263 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Baidu VIS", + "bbox": [ + 450, + 263, + 545, + 280 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{3}$ Zhejiang University", + "bbox": [ + 413, + 281, + 581, + 299 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 333, + 326, + 349 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Wide-baseline panoramic images are frequently used in applications like VR and simulations to minimize capturing labor costs and storage needs. However, synthesizing novel views from these panoramic images in real time remains a significant challenge, especially due to panoramic imagery's high resolution and inherent distortions. Although existing 3D Gaussian splattering (3DGS) methods can produce photo-realistic views under narrow baselines, they often overfit the training views when dealing with wide-baseline panoramic images due to the difficulty in learning precise geometry from sparse $360^{\\circ}$ views. This paper presents Splatter-360, a novel end-to-end generalizable 3DGS framework designed to handle wide-baseline panoramic images. Unlike previous approaches, Splatter-360 performs multi-view matching directly in the spherical domain by constructing a spherical cost volume through a spherical sweep algorithm, enhancing the network's depth perception and geometry estimation. Additionally, we introduce a 3D-aware bi-projection encoder to mitigate the distortions inherent in panoramic images and integrate cross-view attention to improve feature interactions across multiple viewpoints. This enables robust 3D-aware feature representations and real-time rendering capabilities. Experimental results on the HM3D [26] and Replica [27] demonstrate that Splatter-360 significantly outperforms state-of-the-art NeRF and 3DGS methods (e.g., PanoGRF, MVSplat, DepthSplat, and HiSplat) in both synthesis quality and generalization performance for wide-baseline panoramic images. Code and trained models are available at https://3d-aigc.github.io/Splatter-360/.", + "bbox": [ + 89, + 368, + 483, + 821 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 514, + 333, + 643, + 349 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In recent years, 360-degree cameras and VR/AR headsets have gained widespread popularity, enabling seamless scene capture and delivering immersive experiences by presenting visuals directly to users. However, due to the labor-intensive nature of data acquisition and the high storage cost, panoramic images in industry settings often exhibit wide baselines. Generating novel views from these wide-baseline panoramicas is essential to provide users full freedom of movement within virtual environments.", + "bbox": [ + 511, + 363, + 906, + 500 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The emergence of neural radiance fields (NeRF) has demonstrated impressive performance in synthesizing novel views from perspective images. However, NeRF typically requires dense images captured from various angles and positions to address the well-known shape-radiance ambiguity. To reduce the need for dense input, generalizable NeRF appears [31, 36, 44]. Building on the success of these approaches, previous work such as PanoGRF [8] aims to mitigate overfitting in spherical radiance fields by leveraging 360-degree scene priors. Despite their strengths, NeRF-based methods rely on implicit representations, and the rendering process demands extensive network evaluations. Consequently, these methods require powerful GPUs, making them unsuitable for real-time applications on lightweight mobile devices like VR headsets, where low-power processing is essential. In contrast to NeRF, 3D Gaussian Splatting (3DGS) [17] moves away from implicit representation, instead adopting an explicit ellipsoid primitive representation. This representation is particularly well-suited for real-time rendering on traditional rasterization-based devices. However, like NeRF, 3DGS faces challenges with overfitting training views, especially when handling wide-baseline panoramic images.", + "bbox": [ + 511, + 503, + 908, + 851 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "This paper addresses the challenges of learning from wide-baseline panoramic images for real-time novel view rendering of 3DGS. Unlike previous generalizable 3DGS", + "bbox": [ + 511, + 854, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Denotes equal contribution.", + "bbox": [ + 107, + 849, + 263, + 863 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "†This work was done when Zheng Chen interned in Baidu VIS. E-mail: zhengchenecho@gmail.com", + "bbox": [ + 91, + 863, + 482, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$\\ddagger$ Corresponding author. E-mail: shz@tsinghua.edu.cn", + "bbox": [ + 109, + 887, + 393, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "21590", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "methods such as [4, 5, 19], our approach, dubbed as Splatter-360, operates end-to-end with panoramic image inputs and outputs. The design of our method is motivated by the following key factors:1) Panoramic images provide a full $360^{\\circ}$ field-of-view (FoV), whereas cubemap perspective images split the scene into six independent views, causing the issue of projecting points behind the camera in the source view when sampling point in the planar cost volume of the reference view — an issue that panoramic images naturally circumvent; 2) Our method eliminates the explicit back-and-forth conversion between panoramic and cube map representations, where backward conversions often introduce seam artifacts during stitching [2].", + "bbox": [ + 89, + 90, + 480, + 287 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The end-to-end approach for generalizable 3D Gaussian Splitting (3DGS) presents two major challenges. First, panoramic images have significantly higher resolution than perspective images, requiring an extremely efficient network to avoid excessive memory consumption. To address this, we propose performing multi-view matching directly in the spherical domain by constructing a spherical cost volume through a spherical sweep algorithm. Second, the uneven distortion inherent in panoramic images complicates accurate depth estimation. We introduce a 3D-aware bi-projection encoding that leverages monocular depth, equirectangular projection, and cube-map branch information to tackle this issue. Additionally, by incorporating a cross-view attention mechanism to enhance feature interaction across different viewpoints, we obtain a robust 3D-aware feature representation.", + "bbox": [ + 89, + 292, + 482, + 535 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions can be summarized as follows:", + "bbox": [ + 109, + 541, + 434, + 555 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose Splatter-360, a generalizable 3D Gaussian Splatting method for wide-baseline panoramic images, which is an end-to-end trainable network that can generate 3DGS primitives for novel panoramic view synthesis and enables real-time rendering experience.", + "- We introduce a spherical cost volume based on a spherical sweep algorithm in the feed-forward Gaussian splattering framework, which improves the geometry estimation capabilities of the network.", + "- Extensive experiments conducted on the HM3D [26] and Replica [27] datasets demonstrate that our Splatter-360 significantly outperforms state-of-the-art methods in handling wide-baseline panoramic images, both in terms of synthesis quality and generalization performance." + ], + "bbox": [ + 89, + 561, + 482, + 773 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In parallel with our research, PanSplat [47] introduces a hierarchical architecture for high-resolution rendering. Our approach diverges by centering on a distortion-aware feature encoding mechanism. Furthermore, Splatter-360 adopts training under randomized camera baselines combined with a larger-scale dataset [26]. This strategic design choice enhances generalization capabilities by exposing the model to broader view variation patterns.", + "bbox": [ + 89, + 780, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 89, + 653, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Generalizable Novel View Synthesis", + "text_level": 1, + "bbox": [ + 511, + 114, + 823, + 131 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recent advancements in novel view synthesis have been largely driven by NeRF and 3D Gaussian Splatting (3DGS). These methods can be broadly classified into two categories: per-scene optimization approaches [9, 32, 38] and generalizable models [4, 5, 10, 28, 40, 44]. Generalizable methods [4, 5, 10, 28, 40, 44] enable rapid reconstruction by learning priors from large-scale datasets, allowing for efficient feedforward inference. Below, we provide an overview of generalizable NeRF and 3DGS techniques:", + "bbox": [ + 511, + 138, + 903, + 273 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Generalizable NeRF: Early attempts in generalizable NeRF models aim to reconstruct objects and scenes by generating pixel-aligned features for radiance field prediction, pioneered by [31, 36, 44]. Subsequent work such as NeuRay [24] predicts the visibility of 3D points relative to input images, focusing more on visible features. [10] and [35] further propose a multi-view transformer encoder with epipolar line sampling to capture multi-view geometric priors efficiently. MuRF [40] uses a target view frustum volume aligned with the target image plane to enable sharp, high-quality rendering. More recently, LRM [13] and Instant3D [20] designed a large transformer model to regress triplane NeRF feature for single- or sparse-view reconstruction.", + "bbox": [ + 511, + 275, + 903, + 484 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Generalizable 3DGS: To address NeRF's high computational demands, generalizable 3DGS has emerged to simplify view synthesis through rasterization-based splatting, bypassing the need for expensive volume sampling. Methods like Splatter Image [29] demonstrate efficient object-level reconstruction by learning Gaussian parameters from single views. In contrast, pixelSplat [4] and Flash3D [28] have pushed this technique to scene-level reconstruction. More recently, [48] extends the paradigm of LRM [13] to scene-level with 3DGS, MVSplat [5] introduces a cost volume representation using plane sweeping to boost performance further. HiSplat [30] proposes a hierarchical structure for generalizable 3DGS. DepthSplat [41] leverages depth estimation to address multi-view depth methods' limitations and failure cases. However, most of these methods are designed for perspective images and struggle with panoramic inputs due to the large field of view and wide baselines. In response, we propose Splatter-360, extending the strengths of 3DGS to panoramic images.", + "bbox": [ + 511, + 487, + 903, + 776 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Panoramic View Synthesis and Generation", + "text_level": 1, + "bbox": [ + 511, + 787, + 877, + 801 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Unlike perspective, novel view synthesis, panoramic novel view synthesis introduces unique challenges due to the distortions caused by equirectangular projection. Early works [1, 12] attempted to synthesize panoramic views using multi-sphere images. More recent approaches have focused on reconstructing radiance fields or 3D Gaussian", + "bbox": [ + 511, + 809, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "21591", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "scenes from dense panoramic inputs [2, 11]. The challenge intensifies with sparse $360^{\\circ}$ inputs, as depth estimation becomes significantly more difficult. A few methods seek to address this by leveraging the predicted depth and geometric warping to construct neural radiance fields for panoramic novel view synthesis [3, 18]. 360Roam [15] proposes to adapt neural rendering methods for panoramic inputs, struggling in wide-baseline scenarios. PERF [34], on the other hand, explores an inpainting-based approach to generate neural radiance fields from a single panoramic image. Most relevant to our work is PanoGRF [8], which introduces Generalizable Spherical Radiance Fields for wide-baseline panoramicas. PanoGRF achieves state-of-the-art performance by directly aggregating geometry and appearance features of 3D sample points from each panoramic view. However, its inference and rendering speed are limited, making it unsuitable for real-time applications. In contrast, our work proposes a generalizable panoramic view synthesis pipeline designed to generate 3D Gaussian primitives, carefully considering network design and inherent real-time performance during rendering.", + "bbox": [ + 89, + 90, + 485, + 409 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In contrast to the generalizable setting, several works have explored text-based panoramic view generation, albeit with distinct objectives. Nonetheless, these approaches often share similar network design philosophies. For example, Text2Light [7] focuses on high dynamic range panorama generation, while FastScene [25] employs coarse view synthesis followed by progressive inpainting-based generation. DiffPano [43] introduces a spherical epipolar-aware diffusion model for panorama synthesis. Additionally, Panfusion [46], Dreamscene360 [49], and SceneDreamer360 [21] propose novel architectures built on 2D diffusion models. Although these works are orthogonal to ours, they share several challenges in handling panoramic view synthesis.", + "bbox": [ + 89, + 440, + 483, + 652 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Concurrent work. While preparing our work, we found several concurrent works addressing similar challenges using diffusion priors. For instance, ViewCrafter [45], GaussianEnhancer [23], ReconX [22], and FreeVS [37] all use video diffusion models to improve the novel view synthesis. However, these works require computationally intensive de-noising and are not designed for panoramic view synthesis. A very recent work, MVSplat-360 [6], focuses on synthesizing novel perspective images that enable arbitrary position and viewpoint exploration (i.e., $360^{\\circ}$ navigation) using video diffusion priors, aligning with the goals of aforementioned methods but in generalizable 3DGS setting. In contrast, our work specifically tackles generating novel views for panoramic images (i.e., $360^{\\circ}$ images).", + "bbox": [ + 89, + 688, + 483, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Proposed Method", + "text_level": 1, + "bbox": [ + 513, + 89, + 687, + 107 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our proposed framework, Splatter-360, is designed to synthesize novel views from wide-baseline $360^{\\circ}$ panoramic images by leveraging the strengths of 3DGS. Unlike conventional 3DGS methods [5, 41], which are trained on perspective images, Splatter-360 operates directly on panoramic images, thereby mitigating the information loss associated with perspective transformation. The overall pipeline is depicted in Fig. 1. A detailed explanation of the 3D-aware biprojection encoding is provided in Sec. 3.1, spherical depth estimation is discussed in Sec. 3.2, and the pixel-aligned Gaussian decoding is outlined in Sec. 3.3. Additionally, preliminary on 3DGS is included in the supplementary material to ensure this paper is self-contained.", + "bbox": [ + 511, + 114, + 906, + 313 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. 3D-Aware Bi-Projection Encoding", + "text_level": 1, + "bbox": [ + 511, + 319, + 812, + 335 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Although equirectangular projection (ERP) provides a wide field of view and seamless image continuity, the significant distortions, especially near the poles, hinder the network's ability to learn meaningful features effectively. To address this, we propose a bi-projection encoder that leverages both the advantages of ERP and the complementary strengths of cube-map projection (CP). By extracting auxiliary features through CP, our approach enhances the network's capacity to handle these distortions better.", + "bbox": [ + 511, + 340, + 908, + 476 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The overall structure of the bi-projection encoder is illustrated on the left side of Fig. 1. In each branch of the bi-projection network, we begin by applying the convolutional neural network from Unimatch [39] to independently extract local features from both the ERP and CP representations. Unlike prior panorama feature extraction methods [16, 33], which primarily focus on 2D features, our network is designed to learn 3D-aware features. To achieve this, we employ a cross-view attention mechanism to facilitate feature interactions between different viewpoints in the ERP and across the various views in the CP. We obtained $F_{ERP}$ and $F_{CP}$ after the cross-view attention module. $F_{CP}$ is then stitched up into ERP feature $F_{C2E}$ .", + "bbox": [ + 511, + 478, + 908, + 674 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For wide-baseline panoramas, naive $360^{\\circ}$ multi-view matching methods are hard to handle occlusions and textureless areas. To enhance the 3D-aware capability of the encoder in reasoning about geometric details, we introduce the CP single-view depth encoder. This is accomplished by incorporating the single-view geometric priors from a pre-trained monocular depth network into the CP branch at the feature level. Specifically, we extract monocular depth features from each perspective view using a pre-trained depth estimation network DepthAnythingV2 [42]. Once the monocular depth feature maps $F_{CP}^{mono}$ for all perspective views are obtained, we stitch these CP features into the ERP view, denoted as $F_{C2E}^{mono}$ . Next, we concatenate $F_{C2E}^{mono}$ with the transformed CP feature $F_{C2E}$ , and use a two-layer MLP to get the final CP branch feature:", + "bbox": [ + 511, + 674, + 910, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "21592", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/b02674da8fbb225841a82ba6e3fd2f9b5f5850a83392af0434f33072f74cd0df.jpg", + "image_caption": [ + "Figure 1. Our Splatter-360 processes $360^{\\circ}$ panoramic images using a bi-projection encoder that extracts features from both equirectangular projection (ERP) and cube-map projection (CP) through multi-view transformers. These features are used for spherical cost volume construction, and multi-view matching is performed between the reference and source views in spherical space. Next, a refinement U-Net is applied to enhance the spherical cost volume, yielding refined cost volumes and more accurate spherical depth estimations. These refined outputs are then fed into the Gaussian decoder, which produces pixel-aligned Gaussian primitives for synthesizing novel views." + ], + "image_footnote": [], + "bbox": [ + 91, + 89, + 901, + 281 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {F} _ {C 2 E} ^ {\\prime} = \\mathcal {F} _ {1} \\left(\\left[ \\boldsymbol {F} _ {C 2 E} ^ {m o n o}, \\boldsymbol {F} _ {C 2 E} \\right]\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 181, + 401, + 480, + 419 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $[.,.]$ denotes the concatenation operation and $\\mathcal{F}_1$ is a two-layer MLP. Then, we propose a fusion module, $\\mathcal{F}_2$ , to fuse the CP branch feature $F_{C2E}^{\\prime}$ and ERP branch feature $F_{ERP}$ . The final output feature of our encoder is represented as:", + "bbox": [ + 89, + 424, + 483, + 498 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {F} = \\mathcal {F} _ {2} \\left(\\boldsymbol {F} _ {C 2 E} ^ {\\prime}, \\boldsymbol {F} _ {E R P}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 200, + 500, + 482, + 516 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{F}_2$ is the proposed fusion module, which consists of convolutional layers augmented with squeeze-and-excitation blocks [14, 16].", + "bbox": [ + 89, + 522, + 483, + 568 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Spherical Depth Estimation", + "text_level": 1, + "bbox": [ + 89, + 575, + 341, + 593 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "After obtaining a robust feature representation for panoramic images, the next step is cost volume construction. One straightforward approach is to convert the panorama into cube-map-like perspective images and then follow the perspective-based methods such as MVSplat [5] and DepthSplat [41] to construct cost volume. However, such a method has an unavoidable shortcoming: for each 3D sampling point in the planar cost volume of the reference view, these points may project behind the camera in the source view (i.e., $z$ -depth $< 0$ ). When this occurs, incorrect local image features from the source view are sampled, leading to inaccurate local similarity computations and, ultimately, erroneous depth estimates. This issue is a consequence of the narrow field-of-view plane sweep algorithm traditionally used for the perspective-based method.", + "bbox": [ + 89, + 598, + 482, + 824 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To address this, we leverage the 360-degree field of view inherent to panoramas and apply the spherical projection formula derived from [43] to construct a spherical cost volume. More concretely, the spherical sweep algorithm samples depth candidates within the spherical domain, comput-", + "bbox": [ + 89, + 825, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ing feature correlations across multiple views to construct the cost volume. For each reference view, we sample $D$ depth candidates in the spherical space, where the depth values $r_m \\in [r_{\\mathrm{near}}, r_{\\mathrm{far}}]$ are sampled in the logarithmic space between the near range $r_{\\mathrm{near}}$ and the far range $r_{\\mathrm{far}}$ .", + "bbox": [ + 511, + 387, + 906, + 463 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To transform the equirectangular coordinates to spherical coordinates, we use the following equations:", + "bbox": [ + 511, + 463, + 906, + 494 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{ \\begin{array}{l} \\theta = (0. 5 - \\frac {u}{W}) \\cdot 2 \\pi \\\\ \\phi = (0. 5 - \\frac {v}{H}) \\cdot \\pi , \\end{array} \\right. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 632, + 505, + 905, + 560 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $u$ and $v$ represent the pixel coordinates in the equirectangular grid, while $\\theta$ and $\\phi$ denote the longitude and latitude, respectively. Additionally, $W$ and $H$ refer to the width and height of the ERP feature maps on which multiview matching is performed.", + "bbox": [ + 511, + 566, + 905, + 641 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Next, we transform the spherical polar coordinates $(r,\\theta ,\\phi)$ into Cartesian coordinates $(x,y,z)$ to obtain the camera coordinates $p_{\\mathrm{camera}}^i$", + "bbox": [ + 511, + 642, + 906, + 689 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{ \\begin{array}{l} x _ {c a m} = r \\cos (\\phi) \\cdot \\sin (\\theta) \\\\ y _ {c a m} = r \\sin (\\phi) \\\\ z _ {c a m} = r \\cos (\\phi) \\cdot \\cos (\\theta), \\end{array} \\right. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 614, + 699, + 905, + 755 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To obtain the pixel coordinates $(u,v)$ at the source view $j$ , we back-project the camera coordinates using the relative camera pose $W^{i\\rightarrow j}$ between the reference view and the source view:", + "bbox": [ + 511, + 767, + 906, + 828 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {p} _ {\\text {c a m e r a}} ^ {j} = W ^ {i \\rightarrow j} \\boldsymbol {p} _ {\\text {c a m e r a}} ^ {i} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 617, + 839, + 903, + 859 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "With known corresponding pixel coordinates, we sample the local image feature $F^{j\\rightarrow i}$ from the source view feature", + "bbox": [ + 511, + 869, + 906, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "21593", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "and compute the feature similarity between $F^{j\\rightarrow i}$ and $F^i$ as follows:", + "bbox": [ + 89, + 90, + 485, + 122 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nC _ {r _ {m}} ^ {i} = \\frac {\\boldsymbol {F} ^ {i} \\cdot \\boldsymbol {F} _ {r _ {m}} ^ {j \\rightarrow i}}{\\sqrt {C}} \\in \\mathbb {R} ^ {H \\times W}, \\quad m = 1, 2, \\dots , D, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 106, + 130, + 482, + 167 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $C$ denotes the total feature channel and $C_{r_m}^i$ denotes the feature similarity between $F^{j\\rightarrow i}$ and $F^i$ at the $m^{th}$ depth candidate. Then, we can concatenate each depth candidates' feature similarity to get the spherical cost volume.", + "bbox": [ + 89, + 179, + 483, + 241 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {C} ^ {i} = \\left[ \\boldsymbol {C} _ {r _ {1}} ^ {i}, \\boldsymbol {C} _ {r _ {2}} ^ {i}, \\dots , \\boldsymbol {C} _ {r _ {D}} ^ {i} \\right] \\in \\mathbb {R} ^ {H \\times W \\times D}. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 143, + 251, + 482, + 272 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As the initial spherical cost volume only considers the pixel-level similarity and is hard to handle occlusions and textureless areas. We further use a U-Net [5] to refine it. Specifically, the proposed U-Net mainly learns a residual volume value $\\Delta C^i$ , which can be added to the initial spherical cost volume to get the final one:", + "bbox": [ + 89, + 282, + 483, + 375 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\boldsymbol {C}} ^ {i} = \\boldsymbol {C} ^ {i} + \\Delta \\boldsymbol {C} ^ {i} \\in \\mathbb {R} ^ {H \\times W \\times D}. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 179, + 387, + 482, + 407 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Finally, we can use the softmax operation to get the final spherical depth estimation:", + "bbox": [ + 89, + 416, + 483, + 446 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {D} ^ {i} = \\operatorname {s o f t m a x} \\left(\\tilde {\\boldsymbol {C}} ^ {i}\\right) \\boldsymbol {G} \\in \\mathbb {R} ^ {H \\times W}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 174, + 459, + 482, + 479 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\pmb{G} = [r_1, r_2, \\dots, r_D] \\in \\mathbb{R}^D$ is the predefined depth candidates and softmax denotes the softmax function which can transform spherical cost volume to the probability distribution for each depth candidates.", + "bbox": [ + 89, + 492, + 483, + 554 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Pixel-Aligned Gaussian in Equirect-Coordinate", + "text_level": 1, + "bbox": [ + 89, + 564, + 482, + 580 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Once the 3D-aware feature representation and spherical cost volume are obtained, we proceed to predict pixel-aligned Gaussians on the equirectangular coordinate grid. Specifically, for each pixel, we estimate Gaussian centers, quaternions, spherical harmonics (SH), and scaling factors, respectively.", + "bbox": [ + 89, + 587, + 483, + 679 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Gaussian centers $\\mu$ . We use the estimated spherical depth $D$ to calculate the gaussian centers. Specifically, for each pixel, we first compute the camera cartesian coordinates $p_{\\mathrm{camera}}$ of the gaussian center using its spherical depth $D$ via Eq. 3 and Eq. 4. Next, the world coordinates of the Gaussian center are calculated as $p_{\\mathrm{world}} = W p_{\\mathrm{camera}}$ , where $W$ represents the camera-to-world transformation matrix.", + "bbox": [ + 89, + 683, + 483, + 789 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Opacity $\\alpha$ . The opacity is calculated according to the matching confidence. Specifically, we first use the probability distribution of spherical cost volume(i.e., the softmax output of Eq. 9) to get the matching confidence. Then, we feed the matching confidence along with the image features into a two-layer convolutional network to predict the opacity.", + "bbox": [ + 89, + 795, + 483, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Covariance $\\Sigma$ and SH $c$ . We predict these parameters using two convolutional layers, which take as input the concatenation of image features, the refined cost volume, and the original ERP images. The covariance matrix $\\Sigma$ is computed as:", + "bbox": [ + 511, + 90, + 905, + 165 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\Sigma = R (\\theta) ^ {\\top} \\operatorname {d i a g} (s) R (\\theta), \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 620, + 166, + 906, + 185 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "following the formulation in [4]. Here, $R(\\theta)$ is the rotation matrix parameterized by quaternions, $s$ represents the scaling matrix. The $c$ (spherical harmonics, SH) is generated directly by our network for direction-dependent color encoding.", + "bbox": [ + 511, + 191, + 906, + 268 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 282, + 645, + 300 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Datasets, Baselines, and Metrics", + "text_level": 1, + "bbox": [ + 511, + 309, + 795, + 324 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. We evaluate Splatter-360 on two large-scale panoramic datasets: HM3D [26] and Replica [27].", + "bbox": [ + 511, + 332, + 905, + 362 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Metrics. We evaluate the performance of Splatter-360 using standard metrics for novel view synthesis, including PSNR, SSIM, and LPIPS.", + "bbox": [ + 511, + 367, + 905, + 412 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Quantitative Results", + "text_level": 1, + "bbox": [ + 511, + 422, + 705, + 439 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our results demonstrate that Splatter-360 outperforms existing methods in both synthesis quality and generalization to novel views. As for the perspective-based method, we convert panoramicas to cube maps (12 perspective images) and input these cube maps into HiSplat, DepthSplat, and MVSplat. We did not compare with PixelSplat [4] due to out-of-memory errors during inference with 12 perspective image views. Furthermore, HiSplat and DepthSplat cannot be trained on the HM3D dataset with 12 perspective images due to GPU memory constraints; therefore, we tested their official pre-trained models on Re10K [50] for evaluation on both HM3D and Replica. MVSplat and Splatter-360 are trained on HM3D using 8 Tesla V100 GPUs.", + "bbox": [ + 511, + 446, + 906, + 642 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1 shows that existing perspective-based methods, including HiSplat, DepthSplat, and MVSplat, struggle with generalization to cube-map inputs and cannot be directly applied to panoramic datasets. We tested the fine-tuned version of MVSplat, but it still lags behind Splatter-360 in terms of generalization, both on HM3D and Replica. Specifically, the PSNR of MVSplat is 1.114 dB lower than Splatter-360 on HM3D and 1.489 dB lower on Replica. We also compared it with PanoGRF [8], specifically designed for panoramic inputs. Splatter-360 consistently outperforms PanoGRF across all metrics on Replica and HM3D. On HM3D, the PSNR of Splatter-360 is 2.662 dB higher than PanoGRF, and on Replica, it is 1.968 dB higher.", + "bbox": [ + 511, + 643, + 906, + 839 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Moreover, Table 2 provides a quantitative comparison of novel-view depth estimation between MVSplat and Splatter-360. Splatter-360 consistently outperforms MVSplat across all depth metrics on HM3D and Replica, verify", + "bbox": [ + 511, + 839, + 905, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "21594", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/b8471b2e018f46ba8891aebcc73896ca31df52aae5298e5c6934b85291e410ec.jpg", + "image_caption": [ + "Figure 2. Qualitative comparison between our Splatter-360 and PanoGRF, MVSplat on the Replica dataset. Regions with notable differences are highlighted using red and blue rectangles. Please zoom in for a clearer view." + ], + "image_footnote": [], + "bbox": [ + 96, + 88, + 903, + 375 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/8c4412d0bb2eecd4fef5a2bfe7bb16da483ad1f92bf611b99a7b565c6651a4d6.jpg", + "image_caption": [ + "Figure 3. Qualitative comparison between our Splatter-360 and PanoGRF, MVSplat on the HM3D dataset. Regions with notable differences are highlighted using red and blue rectangles. Please zoom in for a clearer view." + ], + "image_footnote": [], + "bbox": [ + 94, + 426, + 906, + 718 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ing that the proposed method can better capture geometry for panoramic inputs.", + "bbox": [ + 89, + 782, + 480, + 811 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Qualitative Results", + "text_level": 1, + "bbox": [ + 89, + 830, + 272, + 845 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Qualitative comparisons are provided in Fig. 3 and Fig. 2, with key differences highlighted using red and blue rectangles. As shown in the first sample of Fig. 2, Splatter-360", + "bbox": [ + 89, + 854, + 483, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "produces much sharper edges and more accurate textures, particularly on objects like the table and chair. In the second and third samples, PanoGRF demonstrates less defined edges, particularly on the floor, while MAVSplat exhibits noticeable artifacts in the form of floaters near the white wall.", + "bbox": [ + 511, + 782, + 906, + 857 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "A comparison of learned geometry between Splatter-360 and MVSplat is presented in Fig. 4. MVSplat shows signifi", + "bbox": [ + 511, + 869, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "21595", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e096dbcad5c710364b29ebab755d905e0d04da1d24c694d53453566251895d21.jpg", + "image_caption": [ + "Figure 4. Novel view depth comparison between Splatter-360 and PanoGRF on the Replica dataset. \"Pano.\" denotes panoramic view and \"Perspec.\" denotes perspective view." + ], + "image_footnote": [], + "bbox": [ + 99, + 88, + 906, + 696 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "cantly poorer depth estimation, as evidenced by its incorrect reconstruction of surfaces such as the lamp and table. In contrast, Splatter-360 delivers more accurate and detailed depth predictions with clearer geometry and well-defined surfaces.", + "bbox": [ + 89, + 761, + 480, + 835 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 848, + 243, + 864 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We validate the effectiveness of different modules proposed in Splatter-360 in this section. Due to the large number", + "bbox": [ + 89, + 869, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "of ablation studies and limited resources, we utilized 2 NVIDIA Tesla V100 GPUs for each group of experiments in this section.", + "bbox": [ + 511, + 761, + 906, + 806 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Spherical cost volume. Eliminating the spherical cost volume leads to a notable decline in performance across all metrics, highlighting its crucial role in the model's ability to generate high-quality reconstructions. Specifically, as shown in the first row of Table 3 , compared to the full", + "bbox": [ + 511, + 824, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "21596", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/801fc2b9bcd777b50f9a575515b84113e949b44e6252256b47022883fc2f72b8.jpg", + "table_caption": [ + "Table 1. Quantitative comparison with baseline methods on the HM3D and Replica datasets. ${}^{ \\dagger }$ indicates models that were trained by us on the panoramic dataset,whereas for all other methods,we used the pre-trained models provided by the original authors." + ], + "table_footnote": [], + "table_body": "
MethodHM3D [26]Replica [27]
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
HiSplat [30]17.2680.6240.48817.1570.6420.417
MVSplat [5]17.5740.6360.44118.0050.6310.512
DepthSplat [41]18.4950.6570.45719.3690.7320.334
PanoGRF [8]25.6310.8130.26827.9200.8920.171
MVSplat† [5]27.1790.8510.17628.3990.9080.115
Splatter-360†28.2930.8750.15529.8880.9240.097
", + "bbox": [ + 135, + 128, + 857, + 280 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/1e5bb0f14d3754ce0728224d47527544d9261ef50acbee4b9a7b3d3218666446.jpg", + "table_caption": [ + "Table 2. Estimated depth comparison between MVSplat and Splatter-360 on the Replica and HM3D datasets." + ], + "table_footnote": [], + "table_body": "
DatasetMetricMVSplatSplatter-360
Replica [27]Abs Diff↓0.1320.102
Abs Rel↓0.0880.063
RMSE↓0.2470.197
δ < 1.25↑89.91394.572
HM3D [26]Abs Diff↓0.1300.106
Abs Rel↓0.0940.076
RMSE↓0.2710.223
δ < 1.25↑90.46993.851
", + "bbox": [ + 91, + 329, + 475, + 488 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "model, the PSNR drops by 5.271 dB and 2.263 dB on the Replica and HM3D datasets, respectively.", + "bbox": [ + 89, + 503, + 482, + 532 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ERP encoder vs. CP encoder. We separately removed the ERP and CP encoders to evaluate their contributions. The results of these ablations are presented in the third and fourth rows of Table 3. The ERP encoder is crucial for maintaining high PSNR and SSIM values and achieving low LPIPS values, highlighting its vital role in panoramic feature extraction and encoding. In contrast, the removal of the CP encoder results in a smaller performance drop compared to the removal of the ERP encoder. However, removing the CP encoder on the Replica dataset still leads to a PSNR decrease of approximately $0.448\\mathrm{dB}$ , suggesting that the CP encoder provides auxiliary support in the Gaussian reconstruction of panoramic images.", + "bbox": [ + 88, + 540, + 482, + 736 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Cross-view attention. Corss-view attention is essential to extract the 3D-awareness feature. Here, we test the influence of removing this module. As shown in the second row of Table 3, such a removing resulted in a noticeable decline in PSNR and SSIM, alongside an increase in LPIPS, indicating the effectiveness of cross-view attention.", + "bbox": [ + 88, + 743, + 482, + 834 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Monocular depth network feature. Removing the monocular depth encoder led to a PSNR drop of 0.467 dB in the Replica dataset. This suggests that the pre-trained monocular depth features offer a robust single-view geometry-", + "bbox": [ + 89, + 839, + 483, + 902 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/198363ceb65cfcd0ec3a1c3b23585e9bdcfc7d5e2f693e1ba36091673efc0f98.jpg", + "table_caption": [ + "Table 3. Ablation studies were conducted on the HM3D and Replica datasets. For simplicity, we use the following abbreviations: 'SCV' for spherical cost volume and 'CVA' for cross-view attention." + ], + "table_footnote": [], + "table_body": "
Ablated moduleReplica [27]HM3D [26]
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
× SCV23.8500.8180.21025.2240.8020.223
× CVA28.2170.9050.12426.9180.8510.182
× ERP26.9850.8870.14225.9050.8270.202
× CP28.6730.9090.11727.2770.8570.174
× Mono Feat.28.6540.9110.11627.3800.8580.173
Full29.1210.9140.11127.4870.8600.171
", + "bbox": [ + 514, + 358, + 916, + 500 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "aware prior, which is especially advantageous for reconstructing 360-degree sparse views in textureless indoor scenes.", + "bbox": [ + 511, + 526, + 906, + 571 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 585, + 633, + 602 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper introduced Splatter-360, a novel framework for generalizable 3D Gaussian Splatting designed for wide-baseline panoramic images. By leveraging 3D-Aware BiProjection Encoding and spherical cost volume, our method significantly improves the quality of novel view synthesis and geometry estimation from panoramic inputs. Experimental results on the HM3D and Replica datasets demonstrate that Splatter-360 sets a new state-of-the-art performance in the panoramic setting, making it a promising solution for applications in VR, simulation rendering, and $360^{\\circ}$ video streaming.", + "bbox": [ + 511, + 609, + 906, + 777 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Acknowledgement", + "text_level": 1, + "bbox": [ + 511, + 791, + 692, + 808 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work was supported by the Natural Science Foundation of China (No. 62132012, 62361146854) and Tsinghua-Tencent Joint Laboratory for Internet Innovation Technology.", + "bbox": [ + 511, + 816, + 905, + 878 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "21597", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 90, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Benjamin Attal, Selena Ling, Aaron Gokaslan, Christian Richardt, and James Tompkin. Matryodshka: Real-time 6dof video view synthesis using multi-sphere images. In European Conference on Computer Vision, pages 441-459. Springer, 2020. 2", + "[2] Jiayang Bai, Letian Huang, Jie Guo, Wen Gong, Yuanqi Li, and Yanwen Guo. 360-gs: Layout-guided panoramic gaussian splatting for indoor roaming. arXiv preprint arXiv:2402.00763, 2024.2, 3", + "[3] Wenjie Chang, Yueyi Zhang, and Zhiwei Xiong. Depth estimation from indoor panoramas with neural scene representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 899-908, 2023. 3", + "[4] David Charatan, Sizhe Lester Li, Andrea Tagliasacchi, and Vincent Sitzmann. pixelsplat: 3d gaussian splats from image pairs for scalable generalizable 3d reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19457-19467, 2024. 2, 5", + "[5] Yuedong Chen, Haofei Xu, Chuanxia Zheng, Bohan Zhuang, Marc Pollefeys, Andreas Geiger, Tat-Jen Cham, and Jianfei Cai. Mvsplat: Efficient 3d gaussian splatting from sparse multi-view images. CoRR, abs/2403.14627, 2024. 2, 3, 4, 5, 8", + "[6] Yuedong Chen, Chuanxia Zheng, Haofei Xu, Bohan Zhuang, Andrea Vedaldi, Tat-Jen Cham, and Jianfei Cai. Mvsplat360: Feed-forward 360 scene synthesis from sparse views. In Advances in Neural Information Processing Systems (NeurIPS), 2024. 3", + "[7] Zhaoxi Chen, Guangcong Wang, and Ziwei Liu. Text2light: Zero-shot text-driven hdr panorama generation. ACM Transactions on Graphics (TOG), 41(6):1-16, 2022. 3", + "[8] Zheng Chen, Yan-Pei Cao, Yuan-Chen Guo, Chen Wang, Ying Shan, and Song-Hai Zhang. Panogrf: generalizable spherical radiance fields for wide-baseline pororamas. Advances in Neural Information Processing Systems, 36:6961-6985, 2023. 1, 3, 5, 8", + "[9] Kangle Deng, Andrew Liu, Jun-Yan Zhu, and Deva Ramanan. Depth-supervised nerf: Fewer views and faster training for free. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12882-12891, 2022. 2", + "[10] Yilun Du, Cameron Smith, Ayush Tewari, and Vincent Sitzmann. Learning to render novel views from wide-baseline stereo pairs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4970-4980, 2023. 2", + "[11] Kai Gu, Thomas Maugey, Sebastian Knorr, and Christine Guillemot. Omni-nerf: neural radiance field from 360 image captures. In 2022 IEEE International Conference on Multimedia and Expo (ICME), pages 1-6. IEEE, 2022. 3", + "[12] Tewodros Habtegebrial, Christiano Gava, Marcel Rogge, Didier Stricker, and Varun Jampani. Somsi: Spherical novel view synthesis with soft occlusion multi-sphere images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15725-15734, 2022. 2" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Yicong Hong, Kai Zhang, Juixiang Gu, Sai Bi, Yang Zhou, Difan Liu, Feng Liu, Kalyan Sunkavalli, Trung Bui, and Hao Tan. Lrm: Large reconstruction model for single image to 3d. arXiv preprint arXiv:2311.04400, 2023. 2", + "[14] Jie Hu, Li Shen, and Gang Sun. Squeeze-and-excitation networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7132-7141, 2018. 4", + "[15] Huajian Huang, Yingshu Chen, Tianjian Zhang, and SaiKit Yeung. 360Roam: Real-Time Indoor Roaming Using Geometry-Aware $360^{\\circ}$ Radiance Fields. arXiv preprint arXiv:2208.02705, 2022. 3", + "[16] Hualie Jiang, Zhe Sheng, Siyu Zhu, Zilong Dong, and Rui Huang. Unifuse: Unidirectional fusion for 360 panorama depth estimation. IEEE Robotics and Automation Letters, 6 (2):1519-1526, 2021. 3, 4", + "[17] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4):139-1, 2023. 1", + "[18] Shreyas Kulkarni, Peng Yin, and Sebastian Scherer. 360fusionnerf: Panoramic neural radiance fields with joint guidance. In 2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 7202-7209. IEEE, 2023. 3", + "[19] Hao Li, Yuanyuan Gao, Dingwen Zhang, Chenming Wu, Yalun Dai, Chen Zhao, Haocheng Feng, Errui Ding, Jingdong Wang, and Junwei Han. Ggrt: Towards generalizable 3d gaussians without pose priors in real-time. arXiv preprint arXiv:2403.10147, 2024. 2", + "[20] Jiahao Li, Hao Tan, Kai Zhang, Zexiang Xu, Fujun Luan, Yinghao Xu, Yicong Hong, Kalyan Sunkavalli, Greg Shakhnarovich, and Sai Bi. Instant3d: Fast text-to-3d with sparse-view generation and large reconstruction model. arXiv preprint arXiv:2311.06214, 2023. 2", + "[21] Wenrui Li, Yapeng Mi, Fucheng Cai, Zhe Yang, Wangmeng Zuo, Xingtao Wang, and Xiaopeng Fan. Scenedreamer360: Text-driven 3d-consistent scene generation with panoramic gaussian splatting. arXiv preprint arXiv:2408.13711, 2024. 3", + "[22] Fangfu Liu, Wenqiang Sun, Hanyang Wang, Yikai Wang, Haowen Sun, Junliang Ye, Jun Zhang, and Yueqi Duan. Reconx: Reconstruct any scene from sparse views with video diffusion model. arXiv preprint arXiv:2408.16767, 2024. 3", + "[23] Xi Liu, Chaoyi Zhou, and Siyu Huang. 3dgssenhancer: Enhancing unbounded 3d gaussian splattering with view-consistent 2d diffusion priors. arXiv preprint arXiv:2410.16266, 2024. 3", + "[24] Yuan Liu, Sida Peng, Lingjie Liu, Qianqian Wang, Peng Wang, Christian Theobalt, Xiaowei Zhou, and Wenping Wang. Neural rays for occlusion-aware image-based rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7824-7833, 2022. 2", + "[25] Yikun Ma, Dandan Zhan, and Zhi Jin. Fastscene: Text-driven fast 3d indoor scene generation via panoramic gaussian splatt-ing. arXiv preprint arXiv:2405.05768, 2024. 3" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "21598", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[26] Santhosh K Ramakrishnan, Aaron Gokaslan, Erik Wijmans, Oleksandr Maksymets, Alex Clegg, John Turner, Eric Undersander, Wojciech Galuba, Andrew Westbury, Angel X Chang, et al. Habitat-matterport 3d dataset (hm3d): 1000 large-scale 3d environments for embodied ai. arXiv preprint arXiv:2109.08238, 2021. 1, 2, 5, 8", + "[27] Julian Straub, Thomas Whelan, Lingni Ma, Yufan Chen, Erik Wijmans, Simon Green, Jakob J Engel, Raul Mur-Artal, Carl Ren, Shobhit Verma, et al. The replica dataset: A digital replica of indoor spaces. arXiv preprint arXiv:1906.05797, 2019. 1, 2, 5, 8", + "[28] Stanislaw Szymanowicz, Eldar Insafutdinov, Chuanxia Zheng, Dylan Campbell, João F Henriques, Christian Rupprecht, and Andrea Vedaldi. Flash3d: Feed-forward generalisable 3d scene reconstruction from a single image. arXiv preprint arXiv:2406.04343, 2024. 2", + "[29] Stanislaw Szymanowicz, Chrisitian Rupprecht, and Andrea Vedaldi. Splatter image: Ultra-fast single-view 3d reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10208-10217, 2024. 2", + "[30] Shengji Tang, Weicai Ye, Peng Ye, Weihao Lin, Yang Zhou, Tao Chen, and Wanli Ouyang. Hisplat: Hierarchical 3d gaussian splatting for generalizable sparse-view reconstruction. arXiv preprint arXiv:2410.06245, 2024. 2, 8", + "[31] Alex Trevithick and Bo Yang. Grf: Learning a general radiance field for 3d representation and rendering. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15182-15192, 2021. 1, 2", + "[32] Prune Truong, Marie-Julie Rakotosaona, Fabian Manhardt, and Federico Tombari. Sparf: Neural radiance fields from sparse and noisy poses. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4190-4200, 2023. 2", + "[33] Fu-En Wang, Yu-Hsuan Yeh, Min Sun, Wei-Chen Chiu, and Yi-Hsuan Tsai. Bifuse: Monocular 360 depth estimation via bi-projection fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 462–471, 2020. 3", + "[34] Guangcong Wang, Peng Wang, Zhaoxi Chen, Wenping Wang, Chen Change Loy, and Ziwei Liu. Perf: Panoramic neural radiance field from a single panorama. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(10): 6905-6918, 2024. 3", + "[35] Peihao Wang, Xuxi Chen, Tianlong Chen, Subhashini Venugopalan, Zhangyang Wang, et al. Is attention all that nef needs? arXiv preprint arXiv:2207.13298, 2022. 2", + "[36] Qianqian Wang, Zhicheng Wang, Kyle Genova, Pratul P Srinivasan, Howard Zhou, Jonathan T Barron, Ricardo Martin-Brualla, Noah Snavely, and Thomas Funkhouser. Ibrnet: Learning multi-view image-based rendering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4690-4699, 2021. 1, 2", + "[37] Qitai Wang, Lue Fan, Yuqi Wang, Yuntao Chen, and Zhaoxiang Zhang. Freevs: Generative view synthesis on free driving trajectory. arXiv preprint arXiv:2410.18079, 2024. 3", + "[38] Rundi Wu, Ben Mildenhall, Philipp Henzler, Keunhong Park, Ruiqi Gao, Daniel Watson, Pratul P Srinivasan, Dor" + ], + "bbox": [ + 91, + 92, + 480, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Verbin, Jonathan T Barron, Ben Poole, et al. Reconfusion: 3d reconstruction with diffusion priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21551-21561, 2024. 2", + "[39] Haofei Xu, Jing Zhang, Jianfei Cai, Hamid Rezatofighi, Fisher Yu, Dacheng Tao, and Andreas Geiger. Unifying flow, stereo and depth estimation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 3", + "[40] Haofei Xu, Anpei Chen, Yuedong Chen, Christos Sakaridis, Yulun Zhang, Marc Pollefeys, Andreas Geiger, and Fisher Yu. Murf: Multi-baseline radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20041-20050, 2024. 2", + "[41] Haofei Xu, Songyou Peng, Fangjinhua Wang, Hermann Blum, Daniel Barath, Andreas Geiger, and Marc Pollefeys. Depthsplat: Connecting gaussian splatting and depth. arXiv preprint arXiv:2410.13862, 2024. 2, 3, 4, 8", + "[42] Lihe Yang, Bingyi Kang, Zilong Huang, Zhen Zhao, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything v2. arXiv preprint arXiv:2406.09414, 2024. 3", + "[43] Weicai Ye, Chenhao Ji, Zheng Chen, Junyao Gao, Xiaoshui Huang, Song-Hai Zhang, Wanli Ouyang, Tong He, Cairong Zhao, and Guofeng Zhang. Diffpano: Scalable and consistent text to panorama generation with spherical epipolar-aware diffusion. arXiv preprint arXiv:2410.24203, 2024. 3, 4", + "[44] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelnerf: Neural radiance fields from one or few images. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4578-4587, 2021. 1, 2", + "[45] Wangbo Yu, Jinbo Xing, Li Yuan, Wenbo Hu, Xiaoyu Li, Zhipeng Huang, Xiangjun Gao, Tien-Tsin Wong, Ying Shan, and Yonghong Tian. Viewcrafter: Taming video diffusion models for high-fidelity novel view synthesis. arXiv preprint arXiv:2409.02048, 2024.3", + "[46] Cheng Zhang, Qianyi Wu, Camilo Cruz Gambardella, Xiaoshui Huang, Dinh Phung, Wanli Ouyang, and Jianfei Cai. Taming stable diffusion for text to 360 $\\{\\backslash \\mathrm{deg}\\}$ panorama image generation. arXiv preprint arXiv:2404.07949, 2024. 3", + "[47] Cheng Zhang, Haofei Xu, Qianyi Wu, Camilo Cruz Gambardella, Dinh Phung, and Jianfei Cai. Pansplat: 4k panorama synthesis with feed-forward gaussian splatting. arXiv preprint arXiv:2412.12096, 2024. 2", + "[48] Kai Zhang, Sai Bi, Hao Tan, Yuanbo Xiangli, Nanxuan Zhao, Kalyan Sunkavalli, and Zexiang Xu. Gs-lrm: Large reconstruction model for 3d gaussian splatting. In European Conference on Computer Vision, pages 1-19. Springer, 2025. 2", + "[49] Shijie Zhou, Zhiwen Fan, Dejia Xu, Haoran Chang, Pradyumna Chari, Tejas Bharadwaj, Suya You, Zhangyang Wang, and Achuta Kadambi. Dreamscene360: Unconstrained text-to-3d scene generation with panoramic gaussian splatting. In European Conference on Computer Vision, pages 324–342. Springer, 2025. 3", + "[50] Tinghui Zhou, Richard Tucker, John Flynn, Graham Fyffe, and Noah Snavely. Stereo magnification: learning view synthesis using multiplane images. ACM Trans. Graph., 37(4): 65, 2018. 5" + ], + "bbox": [ + 516, + 92, + 906, + 893 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "21599", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Splatter-360_ Generalizable 360 Gaussian Splatting for Wide-baseline Panoramic Images/695c9e66-9d36-4a85-a2d0-e99f9431d57c_model.json b/2025/Splatter-360_ Generalizable 360 Gaussian Splatting for Wide-baseline Panoramic Images/695c9e66-9d36-4a85-a2d0-e99f9431d57c_model.json new file mode 100644 index 0000000000000000000000000000000000000000..f16a98ab2e0b7f742fa81b4cbdd6c1b6df899906 --- /dev/null +++ b/2025/Splatter-360_ Generalizable 360 Gaussian Splatting for Wide-baseline Panoramic Images/695c9e66-9d36-4a85-a2d0-e99f9431d57c_model.json @@ -0,0 +1,1991 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.143, + 0.13, + 0.856, + 0.178 + ], + "angle": 0, + "content": "Splatter-360: Generalizable \\(360^{\\circ}\\) Gaussian Splating for Wide-baseline Panoramic Images" + }, + { + "type": "text", + "bbox": [ + 0.209, + 0.203, + 0.788, + 0.242 + ], + "angle": 0, + "content": "Zheng Chen\\(^{1*†}\\), Chenming Wu\\(^{2*}\\), Zhelun Shen\\(^{2}\\), Chen Zhao\\(^{2}\\), Weicai Ye\\(^{3}\\), Haocheng Feng\\(^{2}\\), Errui Ding\\(^{2}\\), Song-Hai Zhang\\(^{1‡}\\)" + }, + { + "type": "text", + "bbox": [ + 0.413, + 0.246, + 0.585, + 0.265 + ], + "angle": 0, + "content": "\\(^{1}\\)Tsinghua University" + }, + { + "type": "text", + "bbox": [ + 0.451, + 0.265, + 0.546, + 0.281 + ], + "angle": 0, + "content": "\\(^{2}\\)Baidu VIS" + }, + { + "type": "text", + "bbox": [ + 0.415, + 0.282, + 0.583, + 0.3 + ], + "angle": 0, + "content": "\\(^{3}\\)Zhejiang University" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.334, + 0.327, + 0.35 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.369, + 0.485, + 0.822 + ], + "angle": 0, + "content": "Wide-baseline panoramic images are frequently used in applications like VR and simulations to minimize capturing labor costs and storage needs. However, synthesizing novel views from these panoramic images in real time remains a significant challenge, especially due to panoramic imagery's high resolution and inherent distortions. Although existing 3D Gaussian splattering (3DGS) methods can produce photo-realistic views under narrow baselines, they often overfit the training views when dealing with wide-baseline panoramic images due to the difficulty in learning precise geometry from sparse \\(360^{\\circ}\\) views. This paper presents Splatter-360, a novel end-to-end generalizable 3DGS framework designed to handle wide-baseline panoramic images. Unlike previous approaches, Splatter-360 performs multi-view matching directly in the spherical domain by constructing a spherical cost volume through a spherical sweep algorithm, enhancing the network's depth perception and geometry estimation. Additionally, we introduce a 3D-aware bi-projection encoder to mitigate the distortions inherent in panoramic images and integrate cross-view attention to improve feature interactions across multiple viewpoints. This enables robust 3D-aware feature representations and real-time rendering capabilities. Experimental results on the HM3D [26] and Replica [27] demonstrate that Splatter-360 significantly outperforms state-of-the-art NeRF and 3DGS methods (e.g., PanoGRF, MVSplat, DepthSplat, and HiSplat) in both synthesis quality and generalization performance for wide-baseline panoramic images. Code and trained models are available at https://3d-aigc.github.io/Splatter-360/." + }, + { + "type": "title", + "bbox": [ + 0.515, + 0.334, + 0.645, + 0.35 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.364, + 0.907, + 0.501 + ], + "angle": 0, + "content": "In recent years, 360-degree cameras and VR/AR headsets have gained widespread popularity, enabling seamless scene capture and delivering immersive experiences by presenting visuals directly to users. However, due to the labor-intensive nature of data acquisition and the high storage cost, panoramic images in industry settings often exhibit wide baselines. Generating novel views from these wide-baseline panoramicas is essential to provide users full freedom of movement within virtual environments." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.504, + 0.909, + 0.852 + ], + "angle": 0, + "content": "The emergence of neural radiance fields (NeRF) has demonstrated impressive performance in synthesizing novel views from perspective images. However, NeRF typically requires dense images captured from various angles and positions to address the well-known shape-radiance ambiguity. To reduce the need for dense input, generalizable NeRF appears [31, 36, 44]. Building on the success of these approaches, previous work such as PanoGRF [8] aims to mitigate overfitting in spherical radiance fields by leveraging 360-degree scene priors. Despite their strengths, NeRF-based methods rely on implicit representations, and the rendering process demands extensive network evaluations. Consequently, these methods require powerful GPUs, making them unsuitable for real-time applications on lightweight mobile devices like VR headsets, where low-power processing is essential. In contrast to NeRF, 3D Gaussian Splatting (3DGS) [17] moves away from implicit representation, instead adopting an explicit ellipsoid primitive representation. This representation is particularly well-suited for real-time rendering on traditional rasterization-based devices. However, like NeRF, 3DGS faces challenges with overfitting training views, especially when handling wide-baseline panoramic images." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.856, + 0.909, + 0.903 + ], + "angle": 0, + "content": "This paper addresses the challenges of learning from wide-baseline panoramic images for real-time novel view rendering of 3DGS. Unlike previous generalizable 3DGS" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.851, + 0.264, + 0.864 + ], + "angle": 0, + "content": "*Denotes equal contribution." + }, + { + "type": "page_footnote", + "bbox": [ + 0.093, + 0.864, + 0.483, + 0.888 + ], + "angle": 0, + "content": "†This work was done when Zheng Chen interned in Baidu VIS. E-mail: zhengchenecho@gmail.com" + }, + { + "type": "page_footnote", + "bbox": [ + 0.111, + 0.888, + 0.394, + 0.9 + ], + "angle": 0, + "content": "\\(\\ddagger\\) Corresponding author. E-mail: shz@tsinghua.edu.cn" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "21590" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.289 + ], + "angle": 0, + "content": "methods such as [4, 5, 19], our approach, dubbed as Splatter-360, operates end-to-end with panoramic image inputs and outputs. The design of our method is motivated by the following key factors:1) Panoramic images provide a full \\(360^{\\circ}\\) field-of-view (FoV), whereas cubemap perspective images split the scene into six independent views, causing the issue of projecting points behind the camera in the source view when sampling point in the planar cost volume of the reference view — an issue that panoramic images naturally circumvent; 2) Our method eliminates the explicit back-and-forth conversion between panoramic and cube map representations, where backward conversions often introduce seam artifacts during stitching [2]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.294, + 0.483, + 0.536 + ], + "angle": 0, + "content": "The end-to-end approach for generalizable 3D Gaussian Splitting (3DGS) presents two major challenges. First, panoramic images have significantly higher resolution than perspective images, requiring an extremely efficient network to avoid excessive memory consumption. To address this, we propose performing multi-view matching directly in the spherical domain by constructing a spherical cost volume through a spherical sweep algorithm. Second, the uneven distortion inherent in panoramic images complicates accurate depth estimation. We introduce a 3D-aware bi-projection encoding that leverages monocular depth, equirectangular projection, and cube-map branch information to tackle this issue. Additionally, by incorporating a cross-view attention mechanism to enhance feature interaction across different viewpoints, we obtain a robust 3D-aware feature representation." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.542, + 0.435, + 0.556 + ], + "angle": 0, + "content": "Our contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.563, + 0.483, + 0.639 + ], + "angle": 0, + "content": "- We propose Splatter-360, a generalizable 3D Gaussian Splatting method for wide-baseline panoramic images, which is an end-to-end trainable network that can generate 3DGS primitives for novel panoramic view synthesis and enables real-time rendering experience." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.639, + 0.483, + 0.699 + ], + "angle": 0, + "content": "- We introduce a spherical cost volume based on a spherical sweep algorithm in the feed-forward Gaussian splattering framework, which improves the geometry estimation capabilities of the network." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.699, + 0.483, + 0.775 + ], + "angle": 0, + "content": "- Extensive experiments conducted on the HM3D [26] and Replica [27] datasets demonstrate that our Splatter-360 significantly outperforms state-of-the-art methods in handling wide-baseline panoramic images, both in terms of synthesis quality and generalization performance." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.563, + 0.483, + 0.775 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.483, + 0.902 + ], + "angle": 0, + "content": "In parallel with our research, PanSplat [47] introduces a hierarchical architecture for high-resolution rendering. Our approach diverges by centering on a distortion-aware feature encoding mechanism. Furthermore, Splatter-360 adopts training under randomized camera baselines combined with a larger-scale dataset [26]. This strategic design choice enhances generalization capabilities by exposing the model to broader view variation patterns." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.09, + 0.655, + 0.105 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.115, + 0.825, + 0.132 + ], + "angle": 0, + "content": "2.1. Generalizable Novel View Synthesis" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.139, + 0.905, + 0.274 + ], + "angle": 0, + "content": "Recent advancements in novel view synthesis have been largely driven by NeRF and 3D Gaussian Splatting (3DGS). These methods can be broadly classified into two categories: per-scene optimization approaches [9, 32, 38] and generalizable models [4, 5, 10, 28, 40, 44]. Generalizable methods [4, 5, 10, 28, 40, 44] enable rapid reconstruction by learning priors from large-scale datasets, allowing for efficient feedforward inference. Below, we provide an overview of generalizable NeRF and 3DGS techniques:" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.276, + 0.905, + 0.486 + ], + "angle": 0, + "content": "Generalizable NeRF: Early attempts in generalizable NeRF models aim to reconstruct objects and scenes by generating pixel-aligned features for radiance field prediction, pioneered by [31, 36, 44]. Subsequent work such as NeuRay [24] predicts the visibility of 3D points relative to input images, focusing more on visible features. [10] and [35] further propose a multi-view transformer encoder with epipolar line sampling to capture multi-view geometric priors efficiently. MuRF [40] uses a target view frustum volume aligned with the target image plane to enable sharp, high-quality rendering. More recently, LRM [13] and Instant3D [20] designed a large transformer model to regress triplane NeRF feature for single- or sparse-view reconstruction." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.488, + 0.905, + 0.777 + ], + "angle": 0, + "content": "Generalizable 3DGS: To address NeRF's high computational demands, generalizable 3DGS has emerged to simplify view synthesis through rasterization-based splatting, bypassing the need for expensive volume sampling. Methods like Splatter Image [29] demonstrate efficient object-level reconstruction by learning Gaussian parameters from single views. In contrast, pixelSplat [4] and Flash3D [28] have pushed this technique to scene-level reconstruction. More recently, [48] extends the paradigm of LRM [13] to scene-level with 3DGS, MVSplat [5] introduces a cost volume representation using plane sweeping to boost performance further. HiSplat [30] proposes a hierarchical structure for generalizable 3DGS. DepthSplat [41] leverages depth estimation to address multi-view depth methods' limitations and failure cases. However, most of these methods are designed for perspective images and struggle with panoramic inputs due to the large field of view and wide baselines. In response, we propose Splatter-360, extending the strengths of 3DGS to panoramic images." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.788, + 0.878, + 0.803 + ], + "angle": 0, + "content": "2.2. Panoramic View Synthesis and Generation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Unlike perspective, novel view synthesis, panoramic novel view synthesis introduces unique challenges due to the distortions caused by equirectangular projection. Early works [1, 12] attempted to synthesize panoramic views using multi-sphere images. More recent approaches have focused on reconstructing radiance fields or 3D Gaussian" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "21591" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.41 + ], + "angle": 0, + "content": "scenes from dense panoramic inputs [2, 11]. The challenge intensifies with sparse \\(360^{\\circ}\\) inputs, as depth estimation becomes significantly more difficult. A few methods seek to address this by leveraging the predicted depth and geometric warping to construct neural radiance fields for panoramic novel view synthesis [3, 18]. 360Roam [15] proposes to adapt neural rendering methods for panoramic inputs, struggling in wide-baseline scenarios. PERF [34], on the other hand, explores an inpainting-based approach to generate neural radiance fields from a single panoramic image. Most relevant to our work is PanoGRF [8], which introduces Generalizable Spherical Radiance Fields for wide-baseline panoramicas. PanoGRF achieves state-of-the-art performance by directly aggregating geometry and appearance features of 3D sample points from each panoramic view. However, its inference and rendering speed are limited, making it unsuitable for real-time applications. In contrast, our work proposes a generalizable panoramic view synthesis pipeline designed to generate 3D Gaussian primitives, carefully considering network design and inherent real-time performance during rendering." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.441, + 0.485, + 0.653 + ], + "angle": 0, + "content": "In contrast to the generalizable setting, several works have explored text-based panoramic view generation, albeit with distinct objectives. Nonetheless, these approaches often share similar network design philosophies. For example, Text2Light [7] focuses on high dynamic range panorama generation, while FastScene [25] employs coarse view synthesis followed by progressive inpainting-based generation. DiffPano [43] introduces a spherical epipolar-aware diffusion model for panorama synthesis. Additionally, Panfusion [46], Dreamscene360 [49], and SceneDreamer360 [21] propose novel architectures built on 2D diffusion models. Although these works are orthogonal to ours, they share several challenges in handling panoramic view synthesis." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.485, + 0.903 + ], + "angle": 0, + "content": "Concurrent work. While preparing our work, we found several concurrent works addressing similar challenges using diffusion priors. For instance, ViewCrafter [45], GaussianEnhancer [23], ReconX [22], and FreeVS [37] all use video diffusion models to improve the novel view synthesis. However, these works require computationally intensive de-noising and are not designed for panoramic view synthesis. A very recent work, MVSplat-360 [6], focuses on synthesizing novel perspective images that enable arbitrary position and viewpoint exploration (i.e., \\(360^{\\circ}\\) navigation) using video diffusion priors, aligning with the goals of aforementioned methods but in generalizable 3DGS setting. In contrast, our work specifically tackles generating novel views for panoramic images (i.e., \\(360^{\\circ}\\) images)." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.09, + 0.689, + 0.108 + ], + "angle": 0, + "content": "3. Proposed Method" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.115, + 0.908, + 0.314 + ], + "angle": 0, + "content": "Our proposed framework, Splatter-360, is designed to synthesize novel views from wide-baseline \\(360^{\\circ}\\) panoramic images by leveraging the strengths of 3DGS. Unlike conventional 3DGS methods [5, 41], which are trained on perspective images, Splatter-360 operates directly on panoramic images, thereby mitigating the information loss associated with perspective transformation. The overall pipeline is depicted in Fig. 1. A detailed explanation of the 3D-aware biprojection encoding is provided in Sec. 3.1, spherical depth estimation is discussed in Sec. 3.2, and the pixel-aligned Gaussian decoding is outlined in Sec. 3.3. Additionally, preliminary on 3DGS is included in the supplementary material to ensure this paper is self-contained." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.32, + 0.813, + 0.337 + ], + "angle": 0, + "content": "3.1. 3D-Aware Bi-Projection Encoding" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.342, + 0.909, + 0.477 + ], + "angle": 0, + "content": "Although equirectangular projection (ERP) provides a wide field of view and seamless image continuity, the significant distortions, especially near the poles, hinder the network's ability to learn meaningful features effectively. To address this, we propose a bi-projection encoder that leverages both the advantages of ERP and the complementary strengths of cube-map projection (CP). By extracting auxiliary features through CP, our approach enhances the network's capacity to handle these distortions better." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.479, + 0.909, + 0.675 + ], + "angle": 0, + "content": "The overall structure of the bi-projection encoder is illustrated on the left side of Fig. 1. In each branch of the bi-projection network, we begin by applying the convolutional neural network from Unimatch [39] to independently extract local features from both the ERP and CP representations. Unlike prior panorama feature extraction methods [16, 33], which primarily focus on 2D features, our network is designed to learn 3D-aware features. To achieve this, we employ a cross-view attention mechanism to facilitate feature interactions between different viewpoints in the ERP and across the various views in the CP. We obtained \\( F_{ERP} \\) and \\( F_{CP} \\) after the cross-view attention module. \\( F_{CP} \\) is then stitched up into ERP feature \\( F_{C2E} \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.911, + 0.902 + ], + "angle": 0, + "content": "For wide-baseline panoramas, naive \\(360^{\\circ}\\) multi-view matching methods are hard to handle occlusions and textureless areas. To enhance the 3D-aware capability of the encoder in reasoning about geometric details, we introduce the CP single-view depth encoder. This is accomplished by incorporating the single-view geometric priors from a pre-trained monocular depth network into the CP branch at the feature level. Specifically, we extract monocular depth features from each perspective view using a pre-trained depth estimation network DepthAnythingV2 [42]. Once the monocular depth feature maps \\(F_{CP}^{mono}\\) for all perspective views are obtained, we stitch these CP features into the ERP view, denoted as \\(F_{C2E}^{mono}\\). Next, we concatenate \\(F_{C2E}^{mono}\\) with the transformed CP feature \\(F_{C2E}\\), and use a two-layer MLP to get the final CP branch feature:" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "21592" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.093, + 0.09, + 0.903, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.291, + 0.907, + 0.366 + ], + "angle": 0, + "content": "Figure 1. Our Splatter-360 processes \\(360^{\\circ}\\) panoramic images using a bi-projection encoder that extracts features from both equirectangular projection (ERP) and cube-map projection (CP) through multi-view transformers. These features are used for spherical cost volume construction, and multi-view matching is performed between the reference and source views in spherical space. Next, a refinement U-Net is applied to enhance the spherical cost volume, yielding refined cost volumes and more accurate spherical depth estimations. These refined outputs are then fed into the Gaussian decoder, which produces pixel-aligned Gaussian primitives for synthesizing novel views." + }, + { + "type": "equation", + "bbox": [ + 0.182, + 0.402, + 0.482, + 0.42 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {F} _ {C 2 E} ^ {\\prime} = \\mathcal {F} _ {1} \\left(\\left[ \\boldsymbol {F} _ {C 2 E} ^ {m o n o}, \\boldsymbol {F} _ {C 2 E} \\right]\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.425, + 0.484, + 0.499 + ], + "angle": 0, + "content": "where \\([.,.]\\) denotes the concatenation operation and \\(\\mathcal{F}_1\\) is a two-layer MLP. Then, we propose a fusion module, \\(\\mathcal{F}_2\\), to fuse the CP branch feature \\(F_{C2E}^{\\prime}\\) and ERP branch feature \\(F_{ERP}\\). The final output feature of our encoder is represented as:" + }, + { + "type": "equation", + "bbox": [ + 0.202, + 0.5, + 0.483, + 0.517 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {F} = \\mathcal {F} _ {2} \\left(\\boldsymbol {F} _ {C 2 E} ^ {\\prime}, \\boldsymbol {F} _ {E R P}\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.523, + 0.484, + 0.569 + ], + "angle": 0, + "content": "where \\(\\mathcal{F}_2\\) is the proposed fusion module, which consists of convolutional layers augmented with squeeze-and-excitation blocks [14, 16]." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.577, + 0.342, + 0.594 + ], + "angle": 0, + "content": "3.2. Spherical Depth Estimation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.599, + 0.483, + 0.825 + ], + "angle": 0, + "content": "After obtaining a robust feature representation for panoramic images, the next step is cost volume construction. One straightforward approach is to convert the panorama into cube-map-like perspective images and then follow the perspective-based methods such as MVSplat [5] and DepthSplat [41] to construct cost volume. However, such a method has an unavoidable shortcoming: for each 3D sampling point in the planar cost volume of the reference view, these points may project behind the camera in the source view (i.e., \\( z \\)-depth \\( < 0 \\)). When this occurs, incorrect local image features from the source view are sampled, leading to inaccurate local similarity computations and, ultimately, erroneous depth estimates. This issue is a consequence of the narrow field-of-view plane sweep algorithm traditionally used for the perspective-based method." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.484, + 0.902 + ], + "angle": 0, + "content": "To address this, we leverage the 360-degree field of view inherent to panoramas and apply the spherical projection formula derived from [43] to construct a spherical cost volume. More concretely, the spherical sweep algorithm samples depth candidates within the spherical domain, comput-" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.388, + 0.907, + 0.464 + ], + "angle": 0, + "content": "ing feature correlations across multiple views to construct the cost volume. For each reference view, we sample \\(D\\) depth candidates in the spherical space, where the depth values \\(r_m \\in [r_{\\mathrm{near}}, r_{\\mathrm{far}}]\\) are sampled in the logarithmic space between the near range \\(r_{\\mathrm{near}}\\) and the far range \\(r_{\\mathrm{far}}\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.464, + 0.907, + 0.495 + ], + "angle": 0, + "content": "To transform the equirectangular coordinates to spherical coordinates, we use the following equations:" + }, + { + "type": "equation", + "bbox": [ + 0.633, + 0.506, + 0.906, + 0.561 + ], + "angle": 0, + "content": "\\[\n\\left\\{ \\begin{array}{l} \\theta = (0. 5 - \\frac {u}{W}) \\cdot 2 \\pi \\\\ \\phi = (0. 5 - \\frac {v}{H}) \\cdot \\pi , \\end{array} \\right. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.567, + 0.906, + 0.642 + ], + "angle": 0, + "content": "where \\( u \\) and \\( v \\) represent the pixel coordinates in the equirectangular grid, while \\( \\theta \\) and \\( \\phi \\) denote the longitude and latitude, respectively. Additionally, \\( W \\) and \\( H \\) refer to the width and height of the ERP feature maps on which multiview matching is performed." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.643, + 0.907, + 0.69 + ], + "angle": 0, + "content": "Next, we transform the spherical polar coordinates \\((r,\\theta ,\\phi)\\) into Cartesian coordinates \\((x,y,z)\\) to obtain the camera coordinates \\(p_{\\mathrm{camera}}^i\\)" + }, + { + "type": "equation", + "bbox": [ + 0.616, + 0.7, + 0.906, + 0.756 + ], + "angle": 0, + "content": "\\[\n\\left\\{ \\begin{array}{l} x _ {c a m} = r \\cos (\\phi) \\cdot \\sin (\\theta) \\\\ y _ {c a m} = r \\sin (\\phi) \\\\ z _ {c a m} = r \\cos (\\phi) \\cdot \\cos (\\theta), \\end{array} \\right. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.768, + 0.907, + 0.829 + ], + "angle": 0, + "content": "To obtain the pixel coordinates \\((u,v)\\) at the source view \\(j\\), we back-project the camera coordinates using the relative camera pose \\(W^{i\\rightarrow j}\\) between the reference view and the source view:" + }, + { + "type": "equation", + "bbox": [ + 0.619, + 0.84, + 0.905, + 0.86 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {p} _ {\\text {c a m e r a}} ^ {j} = W ^ {i \\rightarrow j} \\boldsymbol {p} _ {\\text {c a m e r a}} ^ {i} \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.907, + 0.902 + ], + "angle": 0, + "content": "With known corresponding pixel coordinates, we sample the local image feature \\( F^{j\\rightarrow i} \\) from the source view feature" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "21593" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.091, + 0.486, + 0.123 + ], + "angle": 0, + "content": "and compute the feature similarity between \\( F^{j\\rightarrow i} \\) and \\( F^i \\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.107, + 0.131, + 0.483, + 0.168 + ], + "angle": 0, + "content": "\\[\nC _ {r _ {m}} ^ {i} = \\frac {\\boldsymbol {F} ^ {i} \\cdot \\boldsymbol {F} _ {r _ {m}} ^ {j \\rightarrow i}}{\\sqrt {C}} \\in \\mathbb {R} ^ {H \\times W}, \\quad m = 1, 2, \\dots , D, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.18, + 0.485, + 0.242 + ], + "angle": 0, + "content": "where \\(C\\) denotes the total feature channel and \\(C_{r_m}^i\\) denotes the feature similarity between \\(F^{j\\rightarrow i}\\) and \\(F^i\\) at the \\(m^{th}\\) depth candidate. Then, we can concatenate each depth candidates' feature similarity to get the spherical cost volume." + }, + { + "type": "equation", + "bbox": [ + 0.145, + 0.252, + 0.483, + 0.273 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {C} ^ {i} = \\left[ \\boldsymbol {C} _ {r _ {1}} ^ {i}, \\boldsymbol {C} _ {r _ {2}} ^ {i}, \\dots , \\boldsymbol {C} _ {r _ {D}} ^ {i} \\right] \\in \\mathbb {R} ^ {H \\times W \\times D}. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.284, + 0.485, + 0.375 + ], + "angle": 0, + "content": "As the initial spherical cost volume only considers the pixel-level similarity and is hard to handle occlusions and textureless areas. We further use a U-Net [5] to refine it. Specifically, the proposed U-Net mainly learns a residual volume value \\(\\Delta C^i\\), which can be added to the initial spherical cost volume to get the final one:" + }, + { + "type": "equation", + "bbox": [ + 0.18, + 0.388, + 0.483, + 0.408 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\boldsymbol {C}} ^ {i} = \\boldsymbol {C} ^ {i} + \\Delta \\boldsymbol {C} ^ {i} \\in \\mathbb {R} ^ {H \\times W \\times D}. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.417, + 0.484, + 0.448 + ], + "angle": 0, + "content": "Finally, we can use the softmax operation to get the final spherical depth estimation:" + }, + { + "type": "equation", + "bbox": [ + 0.175, + 0.46, + 0.483, + 0.481 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {D} ^ {i} = \\operatorname {s o f t m a x} \\left(\\tilde {\\boldsymbol {C}} ^ {i}\\right) \\boldsymbol {G} \\in \\mathbb {R} ^ {H \\times W}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.493, + 0.484, + 0.555 + ], + "angle": 0, + "content": "where \\( \\pmb{G} = [r_1, r_2, \\dots, r_D] \\in \\mathbb{R}^D \\) is the predefined depth candidates and softmax denotes the softmax function which can transform spherical cost volume to the probability distribution for each depth candidates." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.565, + 0.483, + 0.582 + ], + "angle": 0, + "content": "3.3. Pixel-Aligned Gaussian in Equirect-Coordinate" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.588, + 0.484, + 0.68 + ], + "angle": 0, + "content": "Once the 3D-aware feature representation and spherical cost volume are obtained, we proceed to predict pixel-aligned Gaussians on the equirectangular coordinate grid. Specifically, for each pixel, we estimate Gaussian centers, quaternions, spherical harmonics (SH), and scaling factors, respectively." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.684, + 0.484, + 0.79 + ], + "angle": 0, + "content": "Gaussian centers \\(\\mu\\). We use the estimated spherical depth \\(D\\) to calculate the gaussian centers. Specifically, for each pixel, we first compute the camera cartesian coordinates \\(p_{\\mathrm{camera}}\\) of the gaussian center using its spherical depth \\(D\\) via Eq. 3 and Eq. 4. Next, the world coordinates of the Gaussian center are calculated as \\(p_{\\mathrm{world}} = W p_{\\mathrm{camera}}\\), where \\(W\\) represents the camera-to-world transformation matrix." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Opacity \\(\\alpha\\). The opacity is calculated according to the matching confidence. Specifically, we first use the probability distribution of spherical cost volume(i.e., the softmax output of Eq. 9) to get the matching confidence. Then, we feed the matching confidence along with the image features into a two-layer convolutional network to predict the opacity." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.166 + ], + "angle": 0, + "content": "Covariance \\(\\Sigma\\) and SH \\(c\\). We predict these parameters using two convolutional layers, which take as input the concatenation of image features, the refined cost volume, and the original ERP images. The covariance matrix \\(\\Sigma\\) is computed as:" + }, + { + "type": "equation", + "bbox": [ + 0.622, + 0.167, + 0.907, + 0.186 + ], + "angle": 0, + "content": "\\[\n\\Sigma = R (\\theta) ^ {\\top} \\operatorname {d i a g} (s) R (\\theta), \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.193, + 0.907, + 0.269 + ], + "angle": 0, + "content": "following the formulation in [4]. Here, \\( R(\\theta) \\) is the rotation matrix parameterized by quaternions, \\( s \\) represents the scaling matrix. The \\( c \\) (spherical harmonics, SH) is generated directly by our network for direction-dependent color encoding." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.284, + 0.646, + 0.301 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.31, + 0.797, + 0.325 + ], + "angle": 0, + "content": "4.1. Datasets, Baselines, and Metrics" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.333, + 0.906, + 0.363 + ], + "angle": 0, + "content": "Datasets. We evaluate Splatter-360 on two large-scale panoramic datasets: HM3D [26] and Replica [27]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.368, + 0.906, + 0.413 + ], + "angle": 0, + "content": "Metrics. We evaluate the performance of Splatter-360 using standard metrics for novel view synthesis, including PSNR, SSIM, and LPIPS." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.424, + 0.707, + 0.44 + ], + "angle": 0, + "content": "4.2. Quantitative Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.447, + 0.907, + 0.643 + ], + "angle": 0, + "content": "Our results demonstrate that Splatter-360 outperforms existing methods in both synthesis quality and generalization to novel views. As for the perspective-based method, we convert panoramicas to cube maps (12 perspective images) and input these cube maps into HiSplat, DepthSplat, and MVSplat. We did not compare with PixelSplat [4] due to out-of-memory errors during inference with 12 perspective image views. Furthermore, HiSplat and DepthSplat cannot be trained on the HM3D dataset with 12 perspective images due to GPU memory constraints; therefore, we tested their official pre-trained models on Re10K [50] for evaluation on both HM3D and Replica. MVSplat and Splatter-360 are trained on HM3D using 8 Tesla V100 GPUs." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.644, + 0.907, + 0.84 + ], + "angle": 0, + "content": "Table 1 shows that existing perspective-based methods, including HiSplat, DepthSplat, and MVSplat, struggle with generalization to cube-map inputs and cannot be directly applied to panoramic datasets. We tested the fine-tuned version of MVSplat, but it still lags behind Splatter-360 in terms of generalization, both on HM3D and Replica. Specifically, the PSNR of MVSplat is 1.114 dB lower than Splatter-360 on HM3D and 1.489 dB lower on Replica. We also compared it with PanoGRF [8], specifically designed for panoramic inputs. Splatter-360 consistently outperforms PanoGRF across all metrics on Replica and HM3D. On HM3D, the PSNR of Splatter-360 is 2.662 dB higher than PanoGRF, and on Replica, it is 1.968 dB higher." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.906, + 0.902 + ], + "angle": 0, + "content": "Moreover, Table 2 provides a quantitative comparison of novel-view depth estimation between MVSplat and Splatter-360. Splatter-360 consistently outperforms MVSplat across all depth metrics on HM3D and Replica, verify" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "21594" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.089, + 0.905, + 0.376 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.387, + 0.908, + 0.416 + ], + "angle": 0, + "content": "Figure 2. Qualitative comparison between our Splatter-360 and PanoGRF, MVSplat on the Replica dataset. Regions with notable differences are highlighted using red and blue rectangles. Please zoom in for a clearer view." + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.428, + 0.907, + 0.719 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.727, + 0.908, + 0.756 + ], + "angle": 0, + "content": "Figure 3. Qualitative comparison between our Splatter-360 and PanoGRF, MVSplat on the HM3D dataset. Regions with notable differences are highlighted using red and blue rectangles. Please zoom in for a clearer view." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.783, + 0.482, + 0.813 + ], + "angle": 0, + "content": "ing that the proposed method can better capture geometry for panoramic inputs." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.831, + 0.274, + 0.847 + ], + "angle": 0, + "content": "4.3. Qualitative Results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Qualitative comparisons are provided in Fig. 3 and Fig. 2, with key differences highlighted using red and blue rectangles. As shown in the first sample of Fig. 2, Splatter-360" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.783, + 0.907, + 0.858 + ], + "angle": 0, + "content": "produces much sharper edges and more accurate textures, particularly on objects like the table and chair. In the second and third samples, PanoGRF demonstrates less defined edges, particularly on the floor, while MAVSplat exhibits noticeable artifacts in the form of floaters near the white wall." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.909, + 0.903 + ], + "angle": 0, + "content": "A comparison of learned geometry between Splatter-360 and MVSplat is presented in Fig. 4. MVSplat shows signifi" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21595" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.1, + 0.089, + 0.907, + 0.697 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.707, + 0.908, + 0.737 + ], + "angle": 0, + "content": "Figure 4. Novel view depth comparison between Splatter-360 and PanoGRF on the Replica dataset. \"Pano.\" denotes panoramic view and \"Perspec.\" denotes perspective view." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.762, + 0.482, + 0.837 + ], + "angle": 0, + "content": "cantly poorer depth estimation, as evidenced by its incorrect reconstruction of surfaces such as the lamp and table. In contrast, Splatter-360 delivers more accurate and detailed depth predictions with clearer geometry and well-defined surfaces." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.849, + 0.245, + 0.866 + ], + "angle": 0, + "content": "4.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.901 + ], + "angle": 0, + "content": "We validate the effectiveness of different modules proposed in Splatter-360 in this section. Due to the large number" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.762, + 0.907, + 0.807 + ], + "angle": 0, + "content": "of ablation studies and limited resources, we utilized 2 NVIDIA Tesla V100 GPUs for each group of experiments in this section." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Spherical cost volume. Eliminating the spherical cost volume leads to a notable decline in performance across all metrics, highlighting its crucial role in the model's ability to generate high-quality reconstructions. Specifically, as shown in the first row of Table 3 , compared to the full" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "21596" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.908, + 0.12 + ], + "angle": 0, + "content": "Table 1. Quantitative comparison with baseline methods on the HM3D and Replica datasets. \\( {}^{ \\dagger } \\) indicates models that were trained by us on the panoramic dataset,whereas for all other methods,we used the pre-trained models provided by the original authors." + }, + { + "type": "table", + "bbox": [ + 0.137, + 0.13, + 0.858, + 0.281 + ], + "angle": 0, + "content": "
MethodHM3D [26]Replica [27]
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
HiSplat [30]17.2680.6240.48817.1570.6420.417
MVSplat [5]17.5740.6360.44118.0050.6310.512
DepthSplat [41]18.4950.6570.45719.3690.7320.334
PanoGRF [8]25.6310.8130.26827.9200.8920.171
MVSplat† [5]27.1790.8510.17628.3990.9080.115
Splatter-360†28.2930.8750.15529.8880.9240.097
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.29, + 0.484, + 0.32 + ], + "angle": 0, + "content": "Table 2. Estimated depth comparison between MVSplat and Splatter-360 on the Replica and HM3D datasets." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.33, + 0.476, + 0.489 + ], + "angle": 0, + "content": "
DatasetMetricMVSplatSplatter-360
Replica [27]Abs Diff↓0.1320.102
Abs Rel↓0.0880.063
RMSE↓0.2470.197
δ < 1.25↑89.91394.572
HM3D [26]Abs Diff↓0.1300.106
Abs Rel↓0.0940.076
RMSE↓0.2710.223
δ < 1.25↑90.46993.851
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.504, + 0.483, + 0.534 + ], + "angle": 0, + "content": "model, the PSNR drops by 5.271 dB and 2.263 dB on the Replica and HM3D datasets, respectively." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.541, + 0.483, + 0.737 + ], + "angle": 0, + "content": "ERP encoder vs. CP encoder. We separately removed the ERP and CP encoders to evaluate their contributions. The results of these ablations are presented in the third and fourth rows of Table 3. The ERP encoder is crucial for maintaining high PSNR and SSIM values and achieving low LPIPS values, highlighting its vital role in panoramic feature extraction and encoding. In contrast, the removal of the CP encoder results in a smaller performance drop compared to the removal of the ERP encoder. However, removing the CP encoder on the Replica dataset still leads to a PSNR decrease of approximately \\(0.448\\mathrm{dB}\\), suggesting that the CP encoder provides auxiliary support in the Gaussian reconstruction of panoramic images." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.744, + 0.483, + 0.835 + ], + "angle": 0, + "content": "Cross-view attention. Corss-view attention is essential to extract the 3D-awareness feature. Here, we test the influence of removing this module. As shown in the second row of Table 3, such a removing resulted in a noticeable decline in PSNR and SSIM, alongside an increase in LPIPS, indicating the effectiveness of cross-view attention." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Monocular depth network feature. Removing the monocular depth encoder led to a PSNR drop of 0.467 dB in the Replica dataset. This suggests that the pre-trained monocular depth features offer a robust single-view geometry-" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.29, + 0.907, + 0.346 + ], + "angle": 0, + "content": "Table 3. Ablation studies were conducted on the HM3D and Replica datasets. For simplicity, we use the following abbreviations: 'SCV' for spherical cost volume and 'CVA' for cross-view attention." + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.359, + 0.918, + 0.501 + ], + "angle": 0, + "content": "
Ablated moduleReplica [27]HM3D [26]
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
× SCV23.8500.8180.21025.2240.8020.223
× CVA28.2170.9050.12426.9180.8510.182
× ERP26.9850.8870.14225.9050.8270.202
× CP28.6730.9090.11727.2770.8570.174
× Mono Feat.28.6540.9110.11627.3800.8580.173
Full29.1210.9140.11127.4870.8600.171
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.527, + 0.907, + 0.573 + ], + "angle": 0, + "content": "aware prior, which is especially advantageous for reconstructing 360-degree sparse views in textureless indoor scenes." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.587, + 0.634, + 0.603 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.611, + 0.907, + 0.779 + ], + "angle": 0, + "content": "This paper introduced Splatter-360, a novel framework for generalizable 3D Gaussian Splatting designed for wide-baseline panoramic images. By leveraging 3D-Aware BiProjection Encoding and spherical cost volume, our method significantly improves the quality of novel view synthesis and geometry estimation from panoramic inputs. Experimental results on the HM3D and Replica datasets demonstrate that Splatter-360 sets a new state-of-the-art performance in the panoramic setting, making it a promising solution for applications in VR, simulation rendering, and \\(360^{\\circ}\\) video streaming." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.792, + 0.694, + 0.809 + ], + "angle": 0, + "content": "6. Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.817, + 0.906, + 0.879 + ], + "angle": 0, + "content": "This work was supported by the Natural Science Foundation of China (No. 62132012, 62361146854) and Tsinghua-Tencent Joint Laboratory for Internet Innovation Technology." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "21597" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.091, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.184 + ], + "angle": 0, + "content": "[1] Benjamin Attal, Selena Ling, Aaron Gokaslan, Christian Richardt, and James Tompkin. Matryodshka: Real-time 6dof video view synthesis using multi-sphere images. In European Conference on Computer Vision, pages 441-459. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.185, + 0.484, + 0.24 + ], + "angle": 0, + "content": "[2] Jiayang Bai, Letian Huang, Jie Guo, Wen Gong, Yuanqi Li, and Yanwen Guo. 360-gs: Layout-guided panoramic gaussian splatting for indoor roaming. arXiv preprint arXiv:2402.00763, 2024.2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.242, + 0.483, + 0.31 + ], + "angle": 0, + "content": "[3] Wenjie Chang, Yueyi Zhang, and Zhiwei Xiong. Depth estimation from indoor panoramas with neural scene representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 899-908, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.311, + 0.483, + 0.382 + ], + "angle": 0, + "content": "[4] David Charatan, Sizhe Lester Li, Andrea Tagliasacchi, and Vincent Sitzmann. pixelsplat: 3d gaussian splats from image pairs for scalable generalizable 3d reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19457-19467, 2024. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.382, + 0.482, + 0.45 + ], + "angle": 0, + "content": "[5] Yuedong Chen, Haofei Xu, Chuanxia Zheng, Bohan Zhuang, Marc Pollefeys, Andreas Geiger, Tat-Jen Cham, and Jianfei Cai. Mvsplat: Efficient 3d gaussian splatting from sparse multi-view images. CoRR, abs/2403.14627, 2024. 2, 3, 4, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.452, + 0.483, + 0.52 + ], + "angle": 0, + "content": "[6] Yuedong Chen, Chuanxia Zheng, Haofei Xu, Bohan Zhuang, Andrea Vedaldi, Tat-Jen Cham, and Jianfei Cai. Mvsplat360: Feed-forward 360 scene synthesis from sparse views. In Advances in Neural Information Processing Systems (NeurIPS), 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.522, + 0.483, + 0.564 + ], + "angle": 0, + "content": "[7] Zhaoxi Chen, Guangcong Wang, and Ziwei Liu. Text2light: Zero-shot text-driven hdr panorama generation. ACM Transactions on Graphics (TOG), 41(6):1-16, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.565, + 0.483, + 0.633 + ], + "angle": 0, + "content": "[8] Zheng Chen, Yan-Pei Cao, Yuan-Chen Guo, Chen Wang, Ying Shan, and Song-Hai Zhang. Panogrf: generalizable spherical radiance fields for wide-baseline pororamas. Advances in Neural Information Processing Systems, 36:6961-6985, 2023. 1, 3, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.635, + 0.483, + 0.703 + ], + "angle": 0, + "content": "[9] Kangle Deng, Andrew Liu, Jun-Yan Zhu, and Deva Ramanan. Depth-supervised nerf: Fewer views and faster training for free. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12882-12891, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.705, + 0.483, + 0.773 + ], + "angle": 0, + "content": "[10] Yilun Du, Cameron Smith, Ayush Tewari, and Vincent Sitzmann. Learning to render novel views from wide-baseline stereo pairs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4970-4980, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.775, + 0.483, + 0.831 + ], + "angle": 0, + "content": "[11] Kai Gu, Thomas Maugey, Sebastian Knorr, and Christine Guillemot. Omni-nerf: neural radiance field from 360 image captures. In 2022 IEEE International Conference on Multimedia and Expo (ICME), pages 1-6. IEEE, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.832, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[12] Tewodros Habtegebrial, Christiano Gava, Marcel Rogge, Didier Stricker, and Varun Jampani. Somsi: Spherical novel view synthesis with soft occlusion multi-sphere images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15725-15734, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.148 + ], + "angle": 0, + "content": "[13] Yicong Hong, Kai Zhang, Juixiang Gu, Sai Bi, Yang Zhou, Difan Liu, Feng Liu, Kalyan Sunkavalli, Trung Bui, and Hao Tan. Lrm: Large reconstruction model for single image to 3d. arXiv preprint arXiv:2311.04400, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.151, + 0.905, + 0.192 + ], + "angle": 0, + "content": "[14] Jie Hu, Li Shen, and Gang Sun. Squeeze-and-excitation networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7132-7141, 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.194, + 0.905, + 0.249 + ], + "angle": 0, + "content": "[15] Huajian Huang, Yingshu Chen, Tianjian Zhang, and SaiKit Yeung. 360Roam: Real-Time Indoor Roaming Using Geometry-Aware \\(360^{\\circ}\\) Radiance Fields. arXiv preprint arXiv:2208.02705, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.253, + 0.905, + 0.308 + ], + "angle": 0, + "content": "[16] Hualie Jiang, Zhe Sheng, Siyu Zhu, Zilong Dong, and Rui Huang. Unifuse: Unidirectional fusion for 360 panorama depth estimation. IEEE Robotics and Automation Letters, 6 (2):1519-1526, 2021. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.311, + 0.905, + 0.366 + ], + "angle": 0, + "content": "[17] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4):139-1, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.369, + 0.905, + 0.437 + ], + "angle": 0, + "content": "[18] Shreyas Kulkarni, Peng Yin, and Sebastian Scherer. 360fusionnerf: Panoramic neural radiance fields with joint guidance. In 2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 7202-7209. IEEE, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.441, + 0.905, + 0.51 + ], + "angle": 0, + "content": "[19] Hao Li, Yuanyuan Gao, Dingwen Zhang, Chenming Wu, Yalun Dai, Chen Zhao, Haocheng Feng, Errui Ding, Jingdong Wang, and Junwei Han. Ggrt: Towards generalizable 3d gaussians without pose priors in real-time. arXiv preprint arXiv:2403.10147, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.513, + 0.905, + 0.582 + ], + "angle": 0, + "content": "[20] Jiahao Li, Hao Tan, Kai Zhang, Zexiang Xu, Fujun Luan, Yinghao Xu, Yicong Hong, Kalyan Sunkavalli, Greg Shakhnarovich, and Sai Bi. Instant3d: Fast text-to-3d with sparse-view generation and large reconstruction model. arXiv preprint arXiv:2311.06214, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.584, + 0.905, + 0.653 + ], + "angle": 0, + "content": "[21] Wenrui Li, Yapeng Mi, Fucheng Cai, Zhe Yang, Wangmeng Zuo, Xingtao Wang, and Xiaopeng Fan. Scenedreamer360: Text-driven 3d-consistent scene generation with panoramic gaussian splatting. arXiv preprint arXiv:2408.13711, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.657, + 0.905, + 0.713 + ], + "angle": 0, + "content": "[22] Fangfu Liu, Wenqiang Sun, Hanyang Wang, Yikai Wang, Haowen Sun, Junliang Ye, Jun Zhang, and Yueqi Duan. Reconx: Reconstruct any scene from sparse views with video diffusion model. arXiv preprint arXiv:2408.16767, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.715, + 0.905, + 0.77 + ], + "angle": 0, + "content": "[23] Xi Liu, Chaoyi Zhou, and Siyu Huang. 3dgssenhancer: Enhancing unbounded 3d gaussian splattering with view-consistent 2d diffusion priors. arXiv preprint arXiv:2410.16266, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.773, + 0.905, + 0.855 + ], + "angle": 0, + "content": "[24] Yuan Liu, Sida Peng, Lingjie Liu, Qianqian Wang, Peng Wang, Christian Theobalt, Xiaowei Zhou, and Wenping Wang. Neural rays for occlusion-aware image-based rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7824-7833, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.858, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[25] Yikun Ma, Dandan Zhan, and Zhi Jin. Fastscene: Text-driven fast 3d indoor scene generation via panoramic gaussian splatt-ing. arXiv preprint arXiv:2405.05768, 2024. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21598" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.175 + ], + "angle": 0, + "content": "[26] Santhosh K Ramakrishnan, Aaron Gokaslan, Erik Wijmans, Oleksandr Maksymets, Alex Clegg, John Turner, Eric Undersander, Wojciech Galuba, Andrew Westbury, Angel X Chang, et al. Habitat-matterport 3d dataset (hm3d): 1000 large-scale 3d environments for embodied ai. arXiv preprint arXiv:2109.08238, 2021. 1, 2, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.177, + 0.482, + 0.245 + ], + "angle": 0, + "content": "[27] Julian Straub, Thomas Whelan, Lingni Ma, Yufan Chen, Erik Wijmans, Simon Green, Jakob J Engel, Raul Mur-Artal, Carl Ren, Shobhit Verma, et al. The replica dataset: A digital replica of indoor spaces. arXiv preprint arXiv:1906.05797, 2019. 1, 2, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.247, + 0.482, + 0.315 + ], + "angle": 0, + "content": "[28] Stanislaw Szymanowicz, Eldar Insafutdinov, Chuanxia Zheng, Dylan Campbell, João F Henriques, Christian Rupprecht, and Andrea Vedaldi. Flash3d: Feed-forward generalisable 3d scene reconstruction from a single image. arXiv preprint arXiv:2406.04343, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.317, + 0.482, + 0.383 + ], + "angle": 0, + "content": "[29] Stanislaw Szymanowicz, Chrisitian Rupprecht, and Andrea Vedaldi. Splatter image: Ultra-fast single-view 3d reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10208-10217, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.385, + 0.482, + 0.44 + ], + "angle": 0, + "content": "[30] Shengji Tang, Weicai Ye, Peng Ye, Weihao Lin, Yang Zhou, Tao Chen, and Wanli Ouyang. Hisplat: Hierarchical 3d gaussian splatting for generalizable sparse-view reconstruction. arXiv preprint arXiv:2410.06245, 2024. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.442, + 0.482, + 0.496 + ], + "angle": 0, + "content": "[31] Alex Trevithick and Bo Yang. Grf: Learning a general radiance field for 3d representation and rendering. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15182-15192, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.498, + 0.482, + 0.566 + ], + "angle": 0, + "content": "[32] Prune Truong, Marie-Julie Rakotosaona, Fabian Manhardt, and Federico Tombari. Sparf: Neural radiance fields from sparse and noisy poses. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4190-4200, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.567, + 0.482, + 0.635 + ], + "angle": 0, + "content": "[33] Fu-En Wang, Yu-Hsuan Yeh, Min Sun, Wei-Chen Chiu, and Yi-Hsuan Tsai. Bifuse: Monocular 360 depth estimation via bi-projection fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 462–471, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.637, + 0.482, + 0.704 + ], + "angle": 0, + "content": "[34] Guangcong Wang, Peng Wang, Zhaoxi Chen, Wenping Wang, Chen Change Loy, and Ziwei Liu. Perf: Panoramic neural radiance field from a single panorama. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(10): 6905-6918, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.706, + 0.482, + 0.747 + ], + "angle": 0, + "content": "[35] Peihao Wang, Xuxi Chen, Tianlong Chen, Subhashini Venugopalan, Zhangyang Wang, et al. Is attention all that nef needs? arXiv preprint arXiv:2207.13298, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.749, + 0.482, + 0.83 + ], + "angle": 0, + "content": "[36] Qianqian Wang, Zhicheng Wang, Kyle Genova, Pratul P Srinivasan, Howard Zhou, Jonathan T Barron, Ricardo Martin-Brualla, Noah Snavely, and Thomas Funkhouser. Ibrnet: Learning multi-view image-based rendering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4690-4699, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.482, + 0.873 + ], + "angle": 0, + "content": "[37] Qitai Wang, Lue Fan, Yuqi Wang, Yuntao Chen, and Zhaoxiang Zhang. Freevs: Generative view synthesis on free driving trajectory. arXiv preprint arXiv:2410.18079, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.874, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[38] Rundi Wu, Ben Mildenhall, Philipp Henzler, Keunhong Park, Ruiqi Gao, Daniel Watson, Pratul P Srinivasan, Dor" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.908, + 0.148 + ], + "angle": 0, + "content": "Verbin, Jonathan T Barron, Ben Poole, et al. Reconfusion: 3d reconstruction with diffusion priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21551-21561, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.15, + 0.906, + 0.205 + ], + "angle": 0, + "content": "[39] Haofei Xu, Jing Zhang, Jianfei Cai, Hamid Rezatofighi, Fisher Yu, Dacheng Tao, and Andreas Geiger. Unifying flow, stereo and depth estimation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.207, + 0.908, + 0.275 + ], + "angle": 0, + "content": "[40] Haofei Xu, Anpei Chen, Yuedong Chen, Christos Sakaridis, Yulun Zhang, Marc Pollefeys, Andreas Geiger, and Fisher Yu. Murf: Multi-baseline radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20041-20050, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.277, + 0.906, + 0.331 + ], + "angle": 0, + "content": "[41] Haofei Xu, Songyou Peng, Fangjinhua Wang, Hermann Blum, Daniel Barath, Andreas Geiger, and Marc Pollefeys. Depthsplat: Connecting gaussian splatting and depth. arXiv preprint arXiv:2410.13862, 2024. 2, 3, 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.333, + 0.906, + 0.374 + ], + "angle": 0, + "content": "[42] Lihe Yang, Bingyi Kang, Zilong Huang, Zhen Zhao, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything v2. arXiv preprint arXiv:2406.09414, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.376, + 0.906, + 0.457 + ], + "angle": 0, + "content": "[43] Weicai Ye, Chenhao Ji, Zheng Chen, Junyao Gao, Xiaoshui Huang, Song-Hai Zhang, Wanli Ouyang, Tong He, Cairong Zhao, and Guofeng Zhang. Diffpano: Scalable and consistent text to panorama generation with spherical epipolar-aware diffusion. arXiv preprint arXiv:2410.24203, 2024. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.46, + 0.906, + 0.515 + ], + "angle": 0, + "content": "[44] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelnerf: Neural radiance fields from one or few images. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4578-4587, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.517, + 0.906, + 0.584 + ], + "angle": 0, + "content": "[45] Wangbo Yu, Jinbo Xing, Li Yuan, Wenbo Hu, Xiaoyu Li, Zhipeng Huang, Xiangjun Gao, Tien-Tsin Wong, Ying Shan, and Yonghong Tian. Viewcrafter: Taming video diffusion models for high-fidelity novel view synthesis. arXiv preprint arXiv:2409.02048, 2024.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.587, + 0.906, + 0.642 + ], + "angle": 0, + "content": "[46] Cheng Zhang, Qianyi Wu, Camilo Cruz Gambardella, Xiaoshui Huang, Dinh Phung, Wanli Ouyang, and Jianfei Cai. Taming stable diffusion for text to 360 \\(\\{\\backslash \\mathrm{deg}\\}\\) panorama image generation. arXiv preprint arXiv:2404.07949, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.644, + 0.906, + 0.699 + ], + "angle": 0, + "content": "[47] Cheng Zhang, Haofei Xu, Qianyi Wu, Camilo Cruz Gambardella, Dinh Phung, and Jianfei Cai. Pansplat: 4k panorama synthesis with feed-forward gaussian splatting. arXiv preprint arXiv:2412.12096, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.701, + 0.906, + 0.756 + ], + "angle": 0, + "content": "[48] Kai Zhang, Sai Bi, Hao Tan, Yuanbo Xiangli, Nanxuan Zhao, Kalyan Sunkavalli, and Zexiang Xu. Gs-lrm: Large reconstruction model for 3d gaussian splatting. In European Conference on Computer Vision, pages 1-19. Springer, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.758, + 0.906, + 0.84 + ], + "angle": 0, + "content": "[49] Shijie Zhou, Zhiwen Fan, Dejia Xu, Haoran Chang, Pradyumna Chari, Tejas Bharadwaj, Suya You, Zhangyang Wang, and Achuta Kadambi. Dreamscene360: Unconstrained text-to-3d scene generation with panoramic gaussian splatting. In European Conference on Computer Vision, pages 324–342. Springer, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.842, + 0.906, + 0.895 + ], + "angle": 0, + "content": "[50] Tinghui Zhou, Richard Tucker, John Flynn, Graham Fyffe, and Noah Snavely. Stereo magnification: learning view synthesis using multiplane images. ACM Trans. Graph., 37(4): 65, 2018. 5" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.895 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "21599" + } + ] +] \ No newline at end of file diff --git a/2025/Splatter-360_ Generalizable 360 Gaussian Splatting for Wide-baseline Panoramic Images/695c9e66-9d36-4a85-a2d0-e99f9431d57c_origin.pdf b/2025/Splatter-360_ Generalizable 360 Gaussian Splatting for Wide-baseline Panoramic Images/695c9e66-9d36-4a85-a2d0-e99f9431d57c_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..82e264df8af3c5998ba45d55003ff5e90996e7b5 --- /dev/null +++ b/2025/Splatter-360_ Generalizable 360 Gaussian Splatting for Wide-baseline Panoramic Images/695c9e66-9d36-4a85-a2d0-e99f9431d57c_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2696da5e5e812f09c506b6c33e9b4adcd6f1e26c72d2db8effcdf683d12c6b41 +size 1747174 diff --git a/2025/Splatter-360_ Generalizable 360 Gaussian Splatting for Wide-baseline Panoramic Images/full.md b/2025/Splatter-360_ Generalizable 360 Gaussian Splatting for Wide-baseline Panoramic Images/full.md new file mode 100644 index 0000000000000000000000000000000000000000..b493e06e748b1e2f31c8cb970120f4af3fad4cc7 --- /dev/null +++ b/2025/Splatter-360_ Generalizable 360 Gaussian Splatting for Wide-baseline Panoramic Images/full.md @@ -0,0 +1,284 @@ +# Splatter-360: Generalizable $360^{\circ}$ Gaussian Splating for Wide-baseline Panoramic Images + +Zheng Chen $^{1*†}$ , Chenming Wu $^{2*}$ , Zhelun Shen $^{2}$ , Chen Zhao $^{2}$ , Weicai Ye $^{3}$ , Haocheng Feng $^{2}$ , Errui Ding $^{2}$ , Song-Hai Zhang $^{1‡}$ + +$^{1}$ Tsinghua University + +$^{2}$ Baidu VIS + +$^{3}$ Zhejiang University + +# Abstract + +Wide-baseline panoramic images are frequently used in applications like VR and simulations to minimize capturing labor costs and storage needs. However, synthesizing novel views from these panoramic images in real time remains a significant challenge, especially due to panoramic imagery's high resolution and inherent distortions. Although existing 3D Gaussian splattering (3DGS) methods can produce photo-realistic views under narrow baselines, they often overfit the training views when dealing with wide-baseline panoramic images due to the difficulty in learning precise geometry from sparse $360^{\circ}$ views. This paper presents Splatter-360, a novel end-to-end generalizable 3DGS framework designed to handle wide-baseline panoramic images. Unlike previous approaches, Splatter-360 performs multi-view matching directly in the spherical domain by constructing a spherical cost volume through a spherical sweep algorithm, enhancing the network's depth perception and geometry estimation. Additionally, we introduce a 3D-aware bi-projection encoder to mitigate the distortions inherent in panoramic images and integrate cross-view attention to improve feature interactions across multiple viewpoints. This enables robust 3D-aware feature representations and real-time rendering capabilities. Experimental results on the HM3D [26] and Replica [27] demonstrate that Splatter-360 significantly outperforms state-of-the-art NeRF and 3DGS methods (e.g., PanoGRF, MVSplat, DepthSplat, and HiSplat) in both synthesis quality and generalization performance for wide-baseline panoramic images. Code and trained models are available at https://3d-aigc.github.io/Splatter-360/. + +# 1. Introduction + +In recent years, 360-degree cameras and VR/AR headsets have gained widespread popularity, enabling seamless scene capture and delivering immersive experiences by presenting visuals directly to users. However, due to the labor-intensive nature of data acquisition and the high storage cost, panoramic images in industry settings often exhibit wide baselines. Generating novel views from these wide-baseline panoramicas is essential to provide users full freedom of movement within virtual environments. + +The emergence of neural radiance fields (NeRF) has demonstrated impressive performance in synthesizing novel views from perspective images. However, NeRF typically requires dense images captured from various angles and positions to address the well-known shape-radiance ambiguity. To reduce the need for dense input, generalizable NeRF appears [31, 36, 44]. Building on the success of these approaches, previous work such as PanoGRF [8] aims to mitigate overfitting in spherical radiance fields by leveraging 360-degree scene priors. Despite their strengths, NeRF-based methods rely on implicit representations, and the rendering process demands extensive network evaluations. Consequently, these methods require powerful GPUs, making them unsuitable for real-time applications on lightweight mobile devices like VR headsets, where low-power processing is essential. In contrast to NeRF, 3D Gaussian Splatting (3DGS) [17] moves away from implicit representation, instead adopting an explicit ellipsoid primitive representation. This representation is particularly well-suited for real-time rendering on traditional rasterization-based devices. However, like NeRF, 3DGS faces challenges with overfitting training views, especially when handling wide-baseline panoramic images. + +This paper addresses the challenges of learning from wide-baseline panoramic images for real-time novel view rendering of 3DGS. Unlike previous generalizable 3DGS + +methods such as [4, 5, 19], our approach, dubbed as Splatter-360, operates end-to-end with panoramic image inputs and outputs. The design of our method is motivated by the following key factors:1) Panoramic images provide a full $360^{\circ}$ field-of-view (FoV), whereas cubemap perspective images split the scene into six independent views, causing the issue of projecting points behind the camera in the source view when sampling point in the planar cost volume of the reference view — an issue that panoramic images naturally circumvent; 2) Our method eliminates the explicit back-and-forth conversion between panoramic and cube map representations, where backward conversions often introduce seam artifacts during stitching [2]. + +The end-to-end approach for generalizable 3D Gaussian Splitting (3DGS) presents two major challenges. First, panoramic images have significantly higher resolution than perspective images, requiring an extremely efficient network to avoid excessive memory consumption. To address this, we propose performing multi-view matching directly in the spherical domain by constructing a spherical cost volume through a spherical sweep algorithm. Second, the uneven distortion inherent in panoramic images complicates accurate depth estimation. We introduce a 3D-aware bi-projection encoding that leverages monocular depth, equirectangular projection, and cube-map branch information to tackle this issue. Additionally, by incorporating a cross-view attention mechanism to enhance feature interaction across different viewpoints, we obtain a robust 3D-aware feature representation. + +Our contributions can be summarized as follows: + +- We propose Splatter-360, a generalizable 3D Gaussian Splatting method for wide-baseline panoramic images, which is an end-to-end trainable network that can generate 3DGS primitives for novel panoramic view synthesis and enables real-time rendering experience. +- We introduce a spherical cost volume based on a spherical sweep algorithm in the feed-forward Gaussian splattering framework, which improves the geometry estimation capabilities of the network. +- Extensive experiments conducted on the HM3D [26] and Replica [27] datasets demonstrate that our Splatter-360 significantly outperforms state-of-the-art methods in handling wide-baseline panoramic images, both in terms of synthesis quality and generalization performance. + +In parallel with our research, PanSplat [47] introduces a hierarchical architecture for high-resolution rendering. Our approach diverges by centering on a distortion-aware feature encoding mechanism. Furthermore, Splatter-360 adopts training under randomized camera baselines combined with a larger-scale dataset [26]. This strategic design choice enhances generalization capabilities by exposing the model to broader view variation patterns. + +# 2. Related Work + +# 2.1. Generalizable Novel View Synthesis + +Recent advancements in novel view synthesis have been largely driven by NeRF and 3D Gaussian Splatting (3DGS). These methods can be broadly classified into two categories: per-scene optimization approaches [9, 32, 38] and generalizable models [4, 5, 10, 28, 40, 44]. Generalizable methods [4, 5, 10, 28, 40, 44] enable rapid reconstruction by learning priors from large-scale datasets, allowing for efficient feedforward inference. Below, we provide an overview of generalizable NeRF and 3DGS techniques: + +Generalizable NeRF: Early attempts in generalizable NeRF models aim to reconstruct objects and scenes by generating pixel-aligned features for radiance field prediction, pioneered by [31, 36, 44]. Subsequent work such as NeuRay [24] predicts the visibility of 3D points relative to input images, focusing more on visible features. [10] and [35] further propose a multi-view transformer encoder with epipolar line sampling to capture multi-view geometric priors efficiently. MuRF [40] uses a target view frustum volume aligned with the target image plane to enable sharp, high-quality rendering. More recently, LRM [13] and Instant3D [20] designed a large transformer model to regress triplane NeRF feature for single- or sparse-view reconstruction. + +Generalizable 3DGS: To address NeRF's high computational demands, generalizable 3DGS has emerged to simplify view synthesis through rasterization-based splatting, bypassing the need for expensive volume sampling. Methods like Splatter Image [29] demonstrate efficient object-level reconstruction by learning Gaussian parameters from single views. In contrast, pixelSplat [4] and Flash3D [28] have pushed this technique to scene-level reconstruction. More recently, [48] extends the paradigm of LRM [13] to scene-level with 3DGS, MVSplat [5] introduces a cost volume representation using plane sweeping to boost performance further. HiSplat [30] proposes a hierarchical structure for generalizable 3DGS. DepthSplat [41] leverages depth estimation to address multi-view depth methods' limitations and failure cases. However, most of these methods are designed for perspective images and struggle with panoramic inputs due to the large field of view and wide baselines. In response, we propose Splatter-360, extending the strengths of 3DGS to panoramic images. + +# 2.2. Panoramic View Synthesis and Generation + +Unlike perspective, novel view synthesis, panoramic novel view synthesis introduces unique challenges due to the distortions caused by equirectangular projection. Early works [1, 12] attempted to synthesize panoramic views using multi-sphere images. More recent approaches have focused on reconstructing radiance fields or 3D Gaussian + +scenes from dense panoramic inputs [2, 11]. The challenge intensifies with sparse $360^{\circ}$ inputs, as depth estimation becomes significantly more difficult. A few methods seek to address this by leveraging the predicted depth and geometric warping to construct neural radiance fields for panoramic novel view synthesis [3, 18]. 360Roam [15] proposes to adapt neural rendering methods for panoramic inputs, struggling in wide-baseline scenarios. PERF [34], on the other hand, explores an inpainting-based approach to generate neural radiance fields from a single panoramic image. Most relevant to our work is PanoGRF [8], which introduces Generalizable Spherical Radiance Fields for wide-baseline panoramicas. PanoGRF achieves state-of-the-art performance by directly aggregating geometry and appearance features of 3D sample points from each panoramic view. However, its inference and rendering speed are limited, making it unsuitable for real-time applications. In contrast, our work proposes a generalizable panoramic view synthesis pipeline designed to generate 3D Gaussian primitives, carefully considering network design and inherent real-time performance during rendering. + +In contrast to the generalizable setting, several works have explored text-based panoramic view generation, albeit with distinct objectives. Nonetheless, these approaches often share similar network design philosophies. For example, Text2Light [7] focuses on high dynamic range panorama generation, while FastScene [25] employs coarse view synthesis followed by progressive inpainting-based generation. DiffPano [43] introduces a spherical epipolar-aware diffusion model for panorama synthesis. Additionally, Panfusion [46], Dreamscene360 [49], and SceneDreamer360 [21] propose novel architectures built on 2D diffusion models. Although these works are orthogonal to ours, they share several challenges in handling panoramic view synthesis. + +Concurrent work. While preparing our work, we found several concurrent works addressing similar challenges using diffusion priors. For instance, ViewCrafter [45], GaussianEnhancer [23], ReconX [22], and FreeVS [37] all use video diffusion models to improve the novel view synthesis. However, these works require computationally intensive de-noising and are not designed for panoramic view synthesis. A very recent work, MVSplat-360 [6], focuses on synthesizing novel perspective images that enable arbitrary position and viewpoint exploration (i.e., $360^{\circ}$ navigation) using video diffusion priors, aligning with the goals of aforementioned methods but in generalizable 3DGS setting. In contrast, our work specifically tackles generating novel views for panoramic images (i.e., $360^{\circ}$ images). + +# 3. Proposed Method + +Our proposed framework, Splatter-360, is designed to synthesize novel views from wide-baseline $360^{\circ}$ panoramic images by leveraging the strengths of 3DGS. Unlike conventional 3DGS methods [5, 41], which are trained on perspective images, Splatter-360 operates directly on panoramic images, thereby mitigating the information loss associated with perspective transformation. The overall pipeline is depicted in Fig. 1. A detailed explanation of the 3D-aware biprojection encoding is provided in Sec. 3.1, spherical depth estimation is discussed in Sec. 3.2, and the pixel-aligned Gaussian decoding is outlined in Sec. 3.3. Additionally, preliminary on 3DGS is included in the supplementary material to ensure this paper is self-contained. + +# 3.1. 3D-Aware Bi-Projection Encoding + +Although equirectangular projection (ERP) provides a wide field of view and seamless image continuity, the significant distortions, especially near the poles, hinder the network's ability to learn meaningful features effectively. To address this, we propose a bi-projection encoder that leverages both the advantages of ERP and the complementary strengths of cube-map projection (CP). By extracting auxiliary features through CP, our approach enhances the network's capacity to handle these distortions better. + +The overall structure of the bi-projection encoder is illustrated on the left side of Fig. 1. In each branch of the bi-projection network, we begin by applying the convolutional neural network from Unimatch [39] to independently extract local features from both the ERP and CP representations. Unlike prior panorama feature extraction methods [16, 33], which primarily focus on 2D features, our network is designed to learn 3D-aware features. To achieve this, we employ a cross-view attention mechanism to facilitate feature interactions between different viewpoints in the ERP and across the various views in the CP. We obtained $F_{ERP}$ and $F_{CP}$ after the cross-view attention module. $F_{CP}$ is then stitched up into ERP feature $F_{C2E}$ . + +For wide-baseline panoramas, naive $360^{\circ}$ multi-view matching methods are hard to handle occlusions and textureless areas. To enhance the 3D-aware capability of the encoder in reasoning about geometric details, we introduce the CP single-view depth encoder. This is accomplished by incorporating the single-view geometric priors from a pre-trained monocular depth network into the CP branch at the feature level. Specifically, we extract monocular depth features from each perspective view using a pre-trained depth estimation network DepthAnythingV2 [42]. Once the monocular depth feature maps $F_{CP}^{mono}$ for all perspective views are obtained, we stitch these CP features into the ERP view, denoted as $F_{C2E}^{mono}$ . Next, we concatenate $F_{C2E}^{mono}$ with the transformed CP feature $F_{C2E}$ , and use a two-layer MLP to get the final CP branch feature: + +![](images/b02674da8fbb225841a82ba6e3fd2f9b5f5850a83392af0434f33072f74cd0df.jpg) +Figure 1. Our Splatter-360 processes $360^{\circ}$ panoramic images using a bi-projection encoder that extracts features from both equirectangular projection (ERP) and cube-map projection (CP) through multi-view transformers. These features are used for spherical cost volume construction, and multi-view matching is performed between the reference and source views in spherical space. Next, a refinement U-Net is applied to enhance the spherical cost volume, yielding refined cost volumes and more accurate spherical depth estimations. These refined outputs are then fed into the Gaussian decoder, which produces pixel-aligned Gaussian primitives for synthesizing novel views. + +$$ +\boldsymbol {F} _ {C 2 E} ^ {\prime} = \mathcal {F} _ {1} \left(\left[ \boldsymbol {F} _ {C 2 E} ^ {m o n o}, \boldsymbol {F} _ {C 2 E} \right]\right), \tag {1} +$$ + +where $[.,.]$ denotes the concatenation operation and $\mathcal{F}_1$ is a two-layer MLP. Then, we propose a fusion module, $\mathcal{F}_2$ , to fuse the CP branch feature $F_{C2E}^{\prime}$ and ERP branch feature $F_{ERP}$ . The final output feature of our encoder is represented as: + +$$ +\boldsymbol {F} = \mathcal {F} _ {2} \left(\boldsymbol {F} _ {C 2 E} ^ {\prime}, \boldsymbol {F} _ {E R P}\right), \tag {2} +$$ + +where $\mathcal{F}_2$ is the proposed fusion module, which consists of convolutional layers augmented with squeeze-and-excitation blocks [14, 16]. + +# 3.2. Spherical Depth Estimation + +After obtaining a robust feature representation for panoramic images, the next step is cost volume construction. One straightforward approach is to convert the panorama into cube-map-like perspective images and then follow the perspective-based methods such as MVSplat [5] and DepthSplat [41] to construct cost volume. However, such a method has an unavoidable shortcoming: for each 3D sampling point in the planar cost volume of the reference view, these points may project behind the camera in the source view (i.e., $z$ -depth $< 0$ ). When this occurs, incorrect local image features from the source view are sampled, leading to inaccurate local similarity computations and, ultimately, erroneous depth estimates. This issue is a consequence of the narrow field-of-view plane sweep algorithm traditionally used for the perspective-based method. + +To address this, we leverage the 360-degree field of view inherent to panoramas and apply the spherical projection formula derived from [43] to construct a spherical cost volume. More concretely, the spherical sweep algorithm samples depth candidates within the spherical domain, comput- + +ing feature correlations across multiple views to construct the cost volume. For each reference view, we sample $D$ depth candidates in the spherical space, where the depth values $r_m \in [r_{\mathrm{near}}, r_{\mathrm{far}}]$ are sampled in the logarithmic space between the near range $r_{\mathrm{near}}$ and the far range $r_{\mathrm{far}}$ . + +To transform the equirectangular coordinates to spherical coordinates, we use the following equations: + +$$ +\left\{ \begin{array}{l} \theta = (0. 5 - \frac {u}{W}) \cdot 2 \pi \\ \phi = (0. 5 - \frac {v}{H}) \cdot \pi , \end{array} \right. \tag {3} +$$ + +where $u$ and $v$ represent the pixel coordinates in the equirectangular grid, while $\theta$ and $\phi$ denote the longitude and latitude, respectively. Additionally, $W$ and $H$ refer to the width and height of the ERP feature maps on which multiview matching is performed. + +Next, we transform the spherical polar coordinates $(r,\theta ,\phi)$ into Cartesian coordinates $(x,y,z)$ to obtain the camera coordinates $p_{\mathrm{camera}}^i$ + +$$ +\left\{ \begin{array}{l} x _ {c a m} = r \cos (\phi) \cdot \sin (\theta) \\ y _ {c a m} = r \sin (\phi) \\ z _ {c a m} = r \cos (\phi) \cdot \cos (\theta), \end{array} \right. \tag {4} +$$ + +To obtain the pixel coordinates $(u,v)$ at the source view $j$ , we back-project the camera coordinates using the relative camera pose $W^{i\rightarrow j}$ between the reference view and the source view: + +$$ +\boldsymbol {p} _ {\text {c a m e r a}} ^ {j} = W ^ {i \rightarrow j} \boldsymbol {p} _ {\text {c a m e r a}} ^ {i} \tag {5} +$$ + +With known corresponding pixel coordinates, we sample the local image feature $F^{j\rightarrow i}$ from the source view feature + +and compute the feature similarity between $F^{j\rightarrow i}$ and $F^i$ as follows: + +$$ +C _ {r _ {m}} ^ {i} = \frac {\boldsymbol {F} ^ {i} \cdot \boldsymbol {F} _ {r _ {m}} ^ {j \rightarrow i}}{\sqrt {C}} \in \mathbb {R} ^ {H \times W}, \quad m = 1, 2, \dots , D, \tag {6} +$$ + +where $C$ denotes the total feature channel and $C_{r_m}^i$ denotes the feature similarity between $F^{j\rightarrow i}$ and $F^i$ at the $m^{th}$ depth candidate. Then, we can concatenate each depth candidates' feature similarity to get the spherical cost volume. + +$$ +\boldsymbol {C} ^ {i} = \left[ \boldsymbol {C} _ {r _ {1}} ^ {i}, \boldsymbol {C} _ {r _ {2}} ^ {i}, \dots , \boldsymbol {C} _ {r _ {D}} ^ {i} \right] \in \mathbb {R} ^ {H \times W \times D}. \tag {7} +$$ + +As the initial spherical cost volume only considers the pixel-level similarity and is hard to handle occlusions and textureless areas. We further use a U-Net [5] to refine it. Specifically, the proposed U-Net mainly learns a residual volume value $\Delta C^i$ , which can be added to the initial spherical cost volume to get the final one: + +$$ +\tilde {\boldsymbol {C}} ^ {i} = \boldsymbol {C} ^ {i} + \Delta \boldsymbol {C} ^ {i} \in \mathbb {R} ^ {H \times W \times D}. \tag {8} +$$ + +Finally, we can use the softmax operation to get the final spherical depth estimation: + +$$ +\boldsymbol {D} ^ {i} = \operatorname {s o f t m a x} \left(\tilde {\boldsymbol {C}} ^ {i}\right) \boldsymbol {G} \in \mathbb {R} ^ {H \times W}, \tag {9} +$$ + +where $\pmb{G} = [r_1, r_2, \dots, r_D] \in \mathbb{R}^D$ is the predefined depth candidates and softmax denotes the softmax function which can transform spherical cost volume to the probability distribution for each depth candidates. + +# 3.3. Pixel-Aligned Gaussian in Equirect-Coordinate + +Once the 3D-aware feature representation and spherical cost volume are obtained, we proceed to predict pixel-aligned Gaussians on the equirectangular coordinate grid. Specifically, for each pixel, we estimate Gaussian centers, quaternions, spherical harmonics (SH), and scaling factors, respectively. + +Gaussian centers $\mu$ . We use the estimated spherical depth $D$ to calculate the gaussian centers. Specifically, for each pixel, we first compute the camera cartesian coordinates $p_{\mathrm{camera}}$ of the gaussian center using its spherical depth $D$ via Eq. 3 and Eq. 4. Next, the world coordinates of the Gaussian center are calculated as $p_{\mathrm{world}} = W p_{\mathrm{camera}}$ , where $W$ represents the camera-to-world transformation matrix. + +Opacity $\alpha$ . The opacity is calculated according to the matching confidence. Specifically, we first use the probability distribution of spherical cost volume(i.e., the softmax output of Eq. 9) to get the matching confidence. Then, we feed the matching confidence along with the image features into a two-layer convolutional network to predict the opacity. + +Covariance $\Sigma$ and SH $c$ . We predict these parameters using two convolutional layers, which take as input the concatenation of image features, the refined cost volume, and the original ERP images. The covariance matrix $\Sigma$ is computed as: + +$$ +\Sigma = R (\theta) ^ {\top} \operatorname {d i a g} (s) R (\theta), \tag {10} +$$ + +following the formulation in [4]. Here, $R(\theta)$ is the rotation matrix parameterized by quaternions, $s$ represents the scaling matrix. The $c$ (spherical harmonics, SH) is generated directly by our network for direction-dependent color encoding. + +# 4. Experiments + +# 4.1. Datasets, Baselines, and Metrics + +Datasets. We evaluate Splatter-360 on two large-scale panoramic datasets: HM3D [26] and Replica [27]. + +Metrics. We evaluate the performance of Splatter-360 using standard metrics for novel view synthesis, including PSNR, SSIM, and LPIPS. + +# 4.2. Quantitative Results + +Our results demonstrate that Splatter-360 outperforms existing methods in both synthesis quality and generalization to novel views. As for the perspective-based method, we convert panoramicas to cube maps (12 perspective images) and input these cube maps into HiSplat, DepthSplat, and MVSplat. We did not compare with PixelSplat [4] due to out-of-memory errors during inference with 12 perspective image views. Furthermore, HiSplat and DepthSplat cannot be trained on the HM3D dataset with 12 perspective images due to GPU memory constraints; therefore, we tested their official pre-trained models on Re10K [50] for evaluation on both HM3D and Replica. MVSplat and Splatter-360 are trained on HM3D using 8 Tesla V100 GPUs. + +Table 1 shows that existing perspective-based methods, including HiSplat, DepthSplat, and MVSplat, struggle with generalization to cube-map inputs and cannot be directly applied to panoramic datasets. We tested the fine-tuned version of MVSplat, but it still lags behind Splatter-360 in terms of generalization, both on HM3D and Replica. Specifically, the PSNR of MVSplat is 1.114 dB lower than Splatter-360 on HM3D and 1.489 dB lower on Replica. We also compared it with PanoGRF [8], specifically designed for panoramic inputs. Splatter-360 consistently outperforms PanoGRF across all metrics on Replica and HM3D. On HM3D, the PSNR of Splatter-360 is 2.662 dB higher than PanoGRF, and on Replica, it is 1.968 dB higher. + +Moreover, Table 2 provides a quantitative comparison of novel-view depth estimation between MVSplat and Splatter-360. Splatter-360 consistently outperforms MVSplat across all depth metrics on HM3D and Replica, verify + +![](images/b8471b2e018f46ba8891aebcc73896ca31df52aae5298e5c6934b85291e410ec.jpg) +Figure 2. Qualitative comparison between our Splatter-360 and PanoGRF, MVSplat on the Replica dataset. Regions with notable differences are highlighted using red and blue rectangles. Please zoom in for a clearer view. + +![](images/8c4412d0bb2eecd4fef5a2bfe7bb16da483ad1f92bf611b99a7b565c6651a4d6.jpg) +Figure 3. Qualitative comparison between our Splatter-360 and PanoGRF, MVSplat on the HM3D dataset. Regions with notable differences are highlighted using red and blue rectangles. Please zoom in for a clearer view. + +ing that the proposed method can better capture geometry for panoramic inputs. + +# 4.3. Qualitative Results + +Qualitative comparisons are provided in Fig. 3 and Fig. 2, with key differences highlighted using red and blue rectangles. As shown in the first sample of Fig. 2, Splatter-360 + +produces much sharper edges and more accurate textures, particularly on objects like the table and chair. In the second and third samples, PanoGRF demonstrates less defined edges, particularly on the floor, while MAVSplat exhibits noticeable artifacts in the form of floaters near the white wall. + +A comparison of learned geometry between Splatter-360 and MVSplat is presented in Fig. 4. MVSplat shows signifi + +![](images/e096dbcad5c710364b29ebab755d905e0d04da1d24c694d53453566251895d21.jpg) +Figure 4. Novel view depth comparison between Splatter-360 and PanoGRF on the Replica dataset. "Pano." denotes panoramic view and "Perspec." denotes perspective view. + +cantly poorer depth estimation, as evidenced by its incorrect reconstruction of surfaces such as the lamp and table. In contrast, Splatter-360 delivers more accurate and detailed depth predictions with clearer geometry and well-defined surfaces. + +# 4.4. Ablation Study + +We validate the effectiveness of different modules proposed in Splatter-360 in this section. Due to the large number + +of ablation studies and limited resources, we utilized 2 NVIDIA Tesla V100 GPUs for each group of experiments in this section. + +Spherical cost volume. Eliminating the spherical cost volume leads to a notable decline in performance across all metrics, highlighting its crucial role in the model's ability to generate high-quality reconstructions. Specifically, as shown in the first row of Table 3 , compared to the full + +Table 1. Quantitative comparison with baseline methods on the HM3D and Replica datasets. ${}^{ \dagger }$ indicates models that were trained by us on the panoramic dataset,whereas for all other methods,we used the pre-trained models provided by the original authors. + +
MethodHM3D [26]Replica [27]
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
HiSplat [30]17.2680.6240.48817.1570.6420.417
MVSplat [5]17.5740.6360.44118.0050.6310.512
DepthSplat [41]18.4950.6570.45719.3690.7320.334
PanoGRF [8]25.6310.8130.26827.9200.8920.171
MVSplat† [5]27.1790.8510.17628.3990.9080.115
Splatter-360†28.2930.8750.15529.8880.9240.097
+ +Table 2. Estimated depth comparison between MVSplat and Splatter-360 on the Replica and HM3D datasets. + +
DatasetMetricMVSplatSplatter-360
Replica [27]Abs Diff↓0.1320.102
Abs Rel↓0.0880.063
RMSE↓0.2470.197
δ < 1.25↑89.91394.572
HM3D [26]Abs Diff↓0.1300.106
Abs Rel↓0.0940.076
RMSE↓0.2710.223
δ < 1.25↑90.46993.851
+ +model, the PSNR drops by 5.271 dB and 2.263 dB on the Replica and HM3D datasets, respectively. + +ERP encoder vs. CP encoder. We separately removed the ERP and CP encoders to evaluate their contributions. The results of these ablations are presented in the third and fourth rows of Table 3. The ERP encoder is crucial for maintaining high PSNR and SSIM values and achieving low LPIPS values, highlighting its vital role in panoramic feature extraction and encoding. In contrast, the removal of the CP encoder results in a smaller performance drop compared to the removal of the ERP encoder. However, removing the CP encoder on the Replica dataset still leads to a PSNR decrease of approximately $0.448\mathrm{dB}$ , suggesting that the CP encoder provides auxiliary support in the Gaussian reconstruction of panoramic images. + +Cross-view attention. Corss-view attention is essential to extract the 3D-awareness feature. Here, we test the influence of removing this module. As shown in the second row of Table 3, such a removing resulted in a noticeable decline in PSNR and SSIM, alongside an increase in LPIPS, indicating the effectiveness of cross-view attention. + +Monocular depth network feature. Removing the monocular depth encoder led to a PSNR drop of 0.467 dB in the Replica dataset. This suggests that the pre-trained monocular depth features offer a robust single-view geometry- + +Table 3. Ablation studies were conducted on the HM3D and Replica datasets. For simplicity, we use the following abbreviations: 'SCV' for spherical cost volume and 'CVA' for cross-view attention. + +
Ablated moduleReplica [27]HM3D [26]
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
× SCV23.8500.8180.21025.2240.8020.223
× CVA28.2170.9050.12426.9180.8510.182
× ERP26.9850.8870.14225.9050.8270.202
× CP28.6730.9090.11727.2770.8570.174
× Mono Feat.28.6540.9110.11627.3800.8580.173
Full29.1210.9140.11127.4870.8600.171
+ +aware prior, which is especially advantageous for reconstructing 360-degree sparse views in textureless indoor scenes. + +# 5. Conclusion + +This paper introduced Splatter-360, a novel framework for generalizable 3D Gaussian Splatting designed for wide-baseline panoramic images. By leveraging 3D-Aware BiProjection Encoding and spherical cost volume, our method significantly improves the quality of novel view synthesis and geometry estimation from panoramic inputs. Experimental results on the HM3D and Replica datasets demonstrate that Splatter-360 sets a new state-of-the-art performance in the panoramic setting, making it a promising solution for applications in VR, simulation rendering, and $360^{\circ}$ video streaming. + +# 6. Acknowledgement + +This work was supported by the Natural Science Foundation of China (No. 62132012, 62361146854) and Tsinghua-Tencent Joint Laboratory for Internet Innovation Technology. + +# References + +[1] Benjamin Attal, Selena Ling, Aaron Gokaslan, Christian Richardt, and James Tompkin. Matryodshka: Real-time 6dof video view synthesis using multi-sphere images. In European Conference on Computer Vision, pages 441-459. Springer, 2020. 2 +[2] Jiayang Bai, Letian Huang, Jie Guo, Wen Gong, Yuanqi Li, and Yanwen Guo. 360-gs: Layout-guided panoramic gaussian splatting for indoor roaming. arXiv preprint arXiv:2402.00763, 2024.2, 3 +[3] Wenjie Chang, Yueyi Zhang, and Zhiwei Xiong. Depth estimation from indoor panoramas with neural scene representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 899-908, 2023. 3 +[4] David Charatan, Sizhe Lester Li, Andrea Tagliasacchi, and Vincent Sitzmann. pixelsplat: 3d gaussian splats from image pairs for scalable generalizable 3d reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19457-19467, 2024. 2, 5 +[5] Yuedong Chen, Haofei Xu, Chuanxia Zheng, Bohan Zhuang, Marc Pollefeys, Andreas Geiger, Tat-Jen Cham, and Jianfei Cai. Mvsplat: Efficient 3d gaussian splatting from sparse multi-view images. CoRR, abs/2403.14627, 2024. 2, 3, 4, 5, 8 +[6] Yuedong Chen, Chuanxia Zheng, Haofei Xu, Bohan Zhuang, Andrea Vedaldi, Tat-Jen Cham, and Jianfei Cai. Mvsplat360: Feed-forward 360 scene synthesis from sparse views. In Advances in Neural Information Processing Systems (NeurIPS), 2024. 3 +[7] Zhaoxi Chen, Guangcong Wang, and Ziwei Liu. Text2light: Zero-shot text-driven hdr panorama generation. ACM Transactions on Graphics (TOG), 41(6):1-16, 2022. 3 +[8] Zheng Chen, Yan-Pei Cao, Yuan-Chen Guo, Chen Wang, Ying Shan, and Song-Hai Zhang. Panogrf: generalizable spherical radiance fields for wide-baseline pororamas. Advances in Neural Information Processing Systems, 36:6961-6985, 2023. 1, 3, 5, 8 +[9] Kangle Deng, Andrew Liu, Jun-Yan Zhu, and Deva Ramanan. Depth-supervised nerf: Fewer views and faster training for free. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12882-12891, 2022. 2 +[10] Yilun Du, Cameron Smith, Ayush Tewari, and Vincent Sitzmann. Learning to render novel views from wide-baseline stereo pairs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4970-4980, 2023. 2 +[11] Kai Gu, Thomas Maugey, Sebastian Knorr, and Christine Guillemot. Omni-nerf: neural radiance field from 360 image captures. In 2022 IEEE International Conference on Multimedia and Expo (ICME), pages 1-6. IEEE, 2022. 3 +[12] Tewodros Habtegebrial, Christiano Gava, Marcel Rogge, Didier Stricker, and Varun Jampani. Somsi: Spherical novel view synthesis with soft occlusion multi-sphere images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15725-15734, 2022. 2 + +[13] Yicong Hong, Kai Zhang, Juixiang Gu, Sai Bi, Yang Zhou, Difan Liu, Feng Liu, Kalyan Sunkavalli, Trung Bui, and Hao Tan. Lrm: Large reconstruction model for single image to 3d. arXiv preprint arXiv:2311.04400, 2023. 2 +[14] Jie Hu, Li Shen, and Gang Sun. Squeeze-and-excitation networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7132-7141, 2018. 4 +[15] Huajian Huang, Yingshu Chen, Tianjian Zhang, and SaiKit Yeung. 360Roam: Real-Time Indoor Roaming Using Geometry-Aware $360^{\circ}$ Radiance Fields. arXiv preprint arXiv:2208.02705, 2022. 3 +[16] Hualie Jiang, Zhe Sheng, Siyu Zhu, Zilong Dong, and Rui Huang. Unifuse: Unidirectional fusion for 360 panorama depth estimation. IEEE Robotics and Automation Letters, 6 (2):1519-1526, 2021. 3, 4 +[17] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4):139-1, 2023. 1 +[18] Shreyas Kulkarni, Peng Yin, and Sebastian Scherer. 360fusionnerf: Panoramic neural radiance fields with joint guidance. In 2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 7202-7209. IEEE, 2023. 3 +[19] Hao Li, Yuanyuan Gao, Dingwen Zhang, Chenming Wu, Yalun Dai, Chen Zhao, Haocheng Feng, Errui Ding, Jingdong Wang, and Junwei Han. Ggrt: Towards generalizable 3d gaussians without pose priors in real-time. arXiv preprint arXiv:2403.10147, 2024. 2 +[20] Jiahao Li, Hao Tan, Kai Zhang, Zexiang Xu, Fujun Luan, Yinghao Xu, Yicong Hong, Kalyan Sunkavalli, Greg Shakhnarovich, and Sai Bi. Instant3d: Fast text-to-3d with sparse-view generation and large reconstruction model. arXiv preprint arXiv:2311.06214, 2023. 2 +[21] Wenrui Li, Yapeng Mi, Fucheng Cai, Zhe Yang, Wangmeng Zuo, Xingtao Wang, and Xiaopeng Fan. Scenedreamer360: Text-driven 3d-consistent scene generation with panoramic gaussian splatting. arXiv preprint arXiv:2408.13711, 2024. 3 +[22] Fangfu Liu, Wenqiang Sun, Hanyang Wang, Yikai Wang, Haowen Sun, Junliang Ye, Jun Zhang, and Yueqi Duan. Reconx: Reconstruct any scene from sparse views with video diffusion model. arXiv preprint arXiv:2408.16767, 2024. 3 +[23] Xi Liu, Chaoyi Zhou, and Siyu Huang. 3dgssenhancer: Enhancing unbounded 3d gaussian splattering with view-consistent 2d diffusion priors. arXiv preprint arXiv:2410.16266, 2024. 3 +[24] Yuan Liu, Sida Peng, Lingjie Liu, Qianqian Wang, Peng Wang, Christian Theobalt, Xiaowei Zhou, and Wenping Wang. Neural rays for occlusion-aware image-based rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7824-7833, 2022. 2 +[25] Yikun Ma, Dandan Zhan, and Zhi Jin. Fastscene: Text-driven fast 3d indoor scene generation via panoramic gaussian splatt-ing. arXiv preprint arXiv:2405.05768, 2024. 3 + +[26] Santhosh K Ramakrishnan, Aaron Gokaslan, Erik Wijmans, Oleksandr Maksymets, Alex Clegg, John Turner, Eric Undersander, Wojciech Galuba, Andrew Westbury, Angel X Chang, et al. Habitat-matterport 3d dataset (hm3d): 1000 large-scale 3d environments for embodied ai. arXiv preprint arXiv:2109.08238, 2021. 1, 2, 5, 8 +[27] Julian Straub, Thomas Whelan, Lingni Ma, Yufan Chen, Erik Wijmans, Simon Green, Jakob J Engel, Raul Mur-Artal, Carl Ren, Shobhit Verma, et al. The replica dataset: A digital replica of indoor spaces. arXiv preprint arXiv:1906.05797, 2019. 1, 2, 5, 8 +[28] Stanislaw Szymanowicz, Eldar Insafutdinov, Chuanxia Zheng, Dylan Campbell, João F Henriques, Christian Rupprecht, and Andrea Vedaldi. Flash3d: Feed-forward generalisable 3d scene reconstruction from a single image. arXiv preprint arXiv:2406.04343, 2024. 2 +[29] Stanislaw Szymanowicz, Chrisitian Rupprecht, and Andrea Vedaldi. Splatter image: Ultra-fast single-view 3d reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10208-10217, 2024. 2 +[30] Shengji Tang, Weicai Ye, Peng Ye, Weihao Lin, Yang Zhou, Tao Chen, and Wanli Ouyang. Hisplat: Hierarchical 3d gaussian splatting for generalizable sparse-view reconstruction. arXiv preprint arXiv:2410.06245, 2024. 2, 8 +[31] Alex Trevithick and Bo Yang. Grf: Learning a general radiance field for 3d representation and rendering. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15182-15192, 2021. 1, 2 +[32] Prune Truong, Marie-Julie Rakotosaona, Fabian Manhardt, and Federico Tombari. Sparf: Neural radiance fields from sparse and noisy poses. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4190-4200, 2023. 2 +[33] Fu-En Wang, Yu-Hsuan Yeh, Min Sun, Wei-Chen Chiu, and Yi-Hsuan Tsai. Bifuse: Monocular 360 depth estimation via bi-projection fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 462–471, 2020. 3 +[34] Guangcong Wang, Peng Wang, Zhaoxi Chen, Wenping Wang, Chen Change Loy, and Ziwei Liu. Perf: Panoramic neural radiance field from a single panorama. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(10): 6905-6918, 2024. 3 +[35] Peihao Wang, Xuxi Chen, Tianlong Chen, Subhashini Venugopalan, Zhangyang Wang, et al. Is attention all that nef needs? arXiv preprint arXiv:2207.13298, 2022. 2 +[36] Qianqian Wang, Zhicheng Wang, Kyle Genova, Pratul P Srinivasan, Howard Zhou, Jonathan T Barron, Ricardo Martin-Brualla, Noah Snavely, and Thomas Funkhouser. Ibrnet: Learning multi-view image-based rendering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4690-4699, 2021. 1, 2 +[37] Qitai Wang, Lue Fan, Yuqi Wang, Yuntao Chen, and Zhaoxiang Zhang. Freevs: Generative view synthesis on free driving trajectory. arXiv preprint arXiv:2410.18079, 2024. 3 +[38] Rundi Wu, Ben Mildenhall, Philipp Henzler, Keunhong Park, Ruiqi Gao, Daniel Watson, Pratul P Srinivasan, Dor + +Verbin, Jonathan T Barron, Ben Poole, et al. Reconfusion: 3d reconstruction with diffusion priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21551-21561, 2024. 2 +[39] Haofei Xu, Jing Zhang, Jianfei Cai, Hamid Rezatofighi, Fisher Yu, Dacheng Tao, and Andreas Geiger. Unifying flow, stereo and depth estimation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 3 +[40] Haofei Xu, Anpei Chen, Yuedong Chen, Christos Sakaridis, Yulun Zhang, Marc Pollefeys, Andreas Geiger, and Fisher Yu. Murf: Multi-baseline radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20041-20050, 2024. 2 +[41] Haofei Xu, Songyou Peng, Fangjinhua Wang, Hermann Blum, Daniel Barath, Andreas Geiger, and Marc Pollefeys. Depthsplat: Connecting gaussian splatting and depth. arXiv preprint arXiv:2410.13862, 2024. 2, 3, 4, 8 +[42] Lihe Yang, Bingyi Kang, Zilong Huang, Zhen Zhao, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything v2. arXiv preprint arXiv:2406.09414, 2024. 3 +[43] Weicai Ye, Chenhao Ji, Zheng Chen, Junyao Gao, Xiaoshui Huang, Song-Hai Zhang, Wanli Ouyang, Tong He, Cairong Zhao, and Guofeng Zhang. Diffpano: Scalable and consistent text to panorama generation with spherical epipolar-aware diffusion. arXiv preprint arXiv:2410.24203, 2024. 3, 4 +[44] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelnerf: Neural radiance fields from one or few images. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4578-4587, 2021. 1, 2 +[45] Wangbo Yu, Jinbo Xing, Li Yuan, Wenbo Hu, Xiaoyu Li, Zhipeng Huang, Xiangjun Gao, Tien-Tsin Wong, Ying Shan, and Yonghong Tian. Viewcrafter: Taming video diffusion models for high-fidelity novel view synthesis. arXiv preprint arXiv:2409.02048, 2024.3 +[46] Cheng Zhang, Qianyi Wu, Camilo Cruz Gambardella, Xiaoshui Huang, Dinh Phung, Wanli Ouyang, and Jianfei Cai. Taming stable diffusion for text to 360 $\{\backslash \mathrm{deg}\}$ panorama image generation. arXiv preprint arXiv:2404.07949, 2024. 3 +[47] Cheng Zhang, Haofei Xu, Qianyi Wu, Camilo Cruz Gambardella, Dinh Phung, and Jianfei Cai. Pansplat: 4k panorama synthesis with feed-forward gaussian splatting. arXiv preprint arXiv:2412.12096, 2024. 2 +[48] Kai Zhang, Sai Bi, Hao Tan, Yuanbo Xiangli, Nanxuan Zhao, Kalyan Sunkavalli, and Zexiang Xu. Gs-lrm: Large reconstruction model for 3d gaussian splatting. In European Conference on Computer Vision, pages 1-19. Springer, 2025. 2 +[49] Shijie Zhou, Zhiwen Fan, Dejia Xu, Haoran Chang, Pradyumna Chari, Tejas Bharadwaj, Suya You, Zhangyang Wang, and Achuta Kadambi. Dreamscene360: Unconstrained text-to-3d scene generation with panoramic gaussian splatting. In European Conference on Computer Vision, pages 324–342. Springer, 2025. 3 +[50] Tinghui Zhou, Richard Tucker, John Flynn, Graham Fyffe, and Noah Snavely. Stereo magnification: learning view synthesis using multiplane images. ACM Trans. Graph., 37(4): 65, 2018. 5 \ No newline at end of file diff --git a/2025/Splatter-360_ Generalizable 360 Gaussian Splatting for Wide-baseline Panoramic Images/images.zip b/2025/Splatter-360_ Generalizable 360 Gaussian Splatting for Wide-baseline Panoramic Images/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..da68eac2a2b54f44e8a20c638ecf61e4e398c287 --- /dev/null +++ b/2025/Splatter-360_ Generalizable 360 Gaussian Splatting for Wide-baseline Panoramic Images/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bf933f60a17db318b8f79f46e89890906744661c2dc77f8ad0646c7dd563f881 +size 693007 diff --git a/2025/Splatter-360_ Generalizable 360 Gaussian Splatting for Wide-baseline Panoramic Images/layout.json b/2025/Splatter-360_ Generalizable 360 Gaussian Splatting for Wide-baseline Panoramic Images/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..46ccb32744d1233eee0506bec35c17f485c8b4f9 --- /dev/null +++ b/2025/Splatter-360_ Generalizable 360 Gaussian Splatting for Wide-baseline Panoramic Images/layout.json @@ -0,0 +1,7585 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 87, + 102, + 523, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 87, + 102, + 523, + 140 + ], + "spans": [ + { + "bbox": [ + 87, + 102, + 523, + 140 + ], + "type": "text", + "content": "Splatter-360: Generalizable " + }, + { + "bbox": [ + 87, + 102, + 523, + 140 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 87, + 102, + 523, + 140 + ], + "type": "text", + "content": " Gaussian Splating for Wide-baseline Panoramic Images" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 127, + 160, + 482, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 160, + 482, + 191 + ], + "spans": [ + { + "bbox": [ + 127, + 160, + 482, + 191 + ], + "type": "text", + "content": "Zheng Chen" + }, + { + "bbox": [ + 127, + 160, + 482, + 191 + ], + "type": "inline_equation", + "content": "^{1*†}" + }, + { + "bbox": [ + 127, + 160, + 482, + 191 + ], + "type": "text", + "content": ", Chenming Wu" + }, + { + "bbox": [ + 127, + 160, + 482, + 191 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 127, + 160, + 482, + 191 + ], + "type": "text", + "content": ", Zhelun Shen" + }, + { + "bbox": [ + 127, + 160, + 482, + 191 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 127, + 160, + 482, + 191 + ], + "type": "text", + "content": ", Chen Zhao" + }, + { + "bbox": [ + 127, + 160, + 482, + 191 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 127, + 160, + 482, + 191 + ], + "type": "text", + "content": ", Weicai Ye" + }, + { + "bbox": [ + 127, + 160, + 482, + 191 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 127, + 160, + 482, + 191 + ], + "type": "text", + "content": ", Haocheng Feng" + }, + { + "bbox": [ + 127, + 160, + 482, + 191 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 127, + 160, + 482, + 191 + ], + "type": "text", + "content": ", Errui Ding" + }, + { + "bbox": [ + 127, + 160, + 482, + 191 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 127, + 160, + 482, + 191 + ], + "type": "text", + "content": ", Song-Hai Zhang" + }, + { + "bbox": [ + 127, + 160, + 482, + 191 + ], + "type": "inline_equation", + "content": "^{1‡}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 252, + 194, + 358, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 252, + 194, + 358, + 209 + ], + "spans": [ + { + "bbox": [ + 252, + 194, + 358, + 209 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 252, + 194, + 358, + 209 + ], + "type": "text", + "content": "Tsinghua University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 276, + 209, + 334, + 222 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 276, + 209, + 334, + 222 + ], + "spans": [ + { + "bbox": [ + 276, + 209, + 334, + 222 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 276, + 209, + 334, + 222 + ], + "type": "text", + "content": "Baidu VIS" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 253, + 223, + 356, + 237 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 253, + 223, + 356, + 237 + ], + "spans": [ + { + "bbox": [ + 253, + 223, + 356, + 237 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 253, + 223, + 356, + 237 + ], + "type": "text", + "content": "Zhejiang University" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 151, + 264, + 200, + 277 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 264, + 200, + 277 + ], + "spans": [ + { + "bbox": [ + 151, + 264, + 200, + 277 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 292, + 296, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 292, + 296, + 651 + ], + "spans": [ + { + "bbox": [ + 55, + 292, + 296, + 651 + ], + "type": "text", + "content": "Wide-baseline panoramic images are frequently used in applications like VR and simulations to minimize capturing labor costs and storage needs. However, synthesizing novel views from these panoramic images in real time remains a significant challenge, especially due to panoramic imagery's high resolution and inherent distortions. Although existing 3D Gaussian splattering (3DGS) methods can produce photo-realistic views under narrow baselines, they often overfit the training views when dealing with wide-baseline panoramic images due to the difficulty in learning precise geometry from sparse " + }, + { + "bbox": [ + 55, + 292, + 296, + 651 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 55, + 292, + 296, + 651 + ], + "type": "text", + "content": " views. This paper presents Splatter-360, a novel end-to-end generalizable 3DGS framework designed to handle wide-baseline panoramic images. Unlike previous approaches, Splatter-360 performs multi-view matching directly in the spherical domain by constructing a spherical cost volume through a spherical sweep algorithm, enhancing the network's depth perception and geometry estimation. Additionally, we introduce a 3D-aware bi-projection encoder to mitigate the distortions inherent in panoramic images and integrate cross-view attention to improve feature interactions across multiple viewpoints. This enables robust 3D-aware feature representations and real-time rendering capabilities. Experimental results on the HM3D [26] and Replica [27] demonstrate that Splatter-360 significantly outperforms state-of-the-art NeRF and 3DGS methods (e.g., PanoGRF, MVSplat, DepthSplat, and HiSplat) in both synthesis quality and generalization performance for wide-baseline panoramic images. Code and trained models are available at https://3d-aigc.github.io/Splatter-360/." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 315, + 264, + 394, + 277 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 264, + 394, + 277 + ], + "spans": [ + { + "bbox": [ + 315, + 264, + 394, + 277 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 288, + 555, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 288, + 555, + 396 + ], + "spans": [ + { + "bbox": [ + 313, + 288, + 555, + 396 + ], + "type": "text", + "content": "In recent years, 360-degree cameras and VR/AR headsets have gained widespread popularity, enabling seamless scene capture and delivering immersive experiences by presenting visuals directly to users. However, due to the labor-intensive nature of data acquisition and the high storage cost, panoramic images in industry settings often exhibit wide baselines. Generating novel views from these wide-baseline panoramicas is essential to provide users full freedom of movement within virtual environments." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 399, + 556, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 399, + 556, + 674 + ], + "spans": [ + { + "bbox": [ + 313, + 399, + 556, + 674 + ], + "type": "text", + "content": "The emergence of neural radiance fields (NeRF) has demonstrated impressive performance in synthesizing novel views from perspective images. However, NeRF typically requires dense images captured from various angles and positions to address the well-known shape-radiance ambiguity. To reduce the need for dense input, generalizable NeRF appears [31, 36, 44]. Building on the success of these approaches, previous work such as PanoGRF [8] aims to mitigate overfitting in spherical radiance fields by leveraging 360-degree scene priors. Despite their strengths, NeRF-based methods rely on implicit representations, and the rendering process demands extensive network evaluations. Consequently, these methods require powerful GPUs, making them unsuitable for real-time applications on lightweight mobile devices like VR headsets, where low-power processing is essential. In contrast to NeRF, 3D Gaussian Splatting (3DGS) [17] moves away from implicit representation, instead adopting an explicit ellipsoid primitive representation. This representation is particularly well-suited for real-time rendering on traditional rasterization-based devices. However, like NeRF, 3DGS faces challenges with overfitting training views, especially when handling wide-baseline panoramic images." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "text", + "content": "This paper addresses the challenges of learning from wide-baseline panoramic images for real-time novel view rendering of 3DGS. Unlike previous generalizable 3DGS" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 673, + 161, + 684 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 673, + 161, + 684 + ], + "spans": [ + { + "bbox": [ + 66, + 673, + 161, + 684 + ], + "type": "text", + "content": "*Denotes equal contribution." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 684, + 295, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 684, + 295, + 703 + ], + "spans": [ + { + "bbox": [ + 56, + 684, + 295, + 703 + ], + "type": "text", + "content": "†This work was done when Zheng Chen interned in Baidu VIS. E-mail: zhengchenecho@gmail.com" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 67, + 703, + 241, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 241, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 241, + 712 + ], + "type": "inline_equation", + "content": "\\ddagger" + }, + { + "bbox": [ + 67, + 703, + 241, + 712 + ], + "type": "text", + "content": " Corresponding author. E-mail: shz@tsinghua.edu.cn" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "21590" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 228 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 228 + ], + "type": "text", + "content": "methods such as [4, 5, 19], our approach, dubbed as Splatter-360, operates end-to-end with panoramic image inputs and outputs. The design of our method is motivated by the following key factors:1) Panoramic images provide a full " + }, + { + "bbox": [ + 55, + 72, + 294, + 228 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 55, + 72, + 294, + 228 + ], + "type": "text", + "content": " field-of-view (FoV), whereas cubemap perspective images split the scene into six independent views, causing the issue of projecting points behind the camera in the source view when sampling point in the planar cost volume of the reference view — an issue that panoramic images naturally circumvent; 2) Our method eliminates the explicit back-and-forth conversion between panoramic and cube map representations, where backward conversions often introduce seam artifacts during stitching [2]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 232, + 295, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 232, + 295, + 424 + ], + "spans": [ + { + "bbox": [ + 55, + 232, + 295, + 424 + ], + "type": "text", + "content": "The end-to-end approach for generalizable 3D Gaussian Splitting (3DGS) presents two major challenges. First, panoramic images have significantly higher resolution than perspective images, requiring an extremely efficient network to avoid excessive memory consumption. To address this, we propose performing multi-view matching directly in the spherical domain by constructing a spherical cost volume through a spherical sweep algorithm. Second, the uneven distortion inherent in panoramic images complicates accurate depth estimation. We introduce a 3D-aware bi-projection encoding that leverages monocular depth, equirectangular projection, and cube-map branch information to tackle this issue. Additionally, by incorporating a cross-view attention mechanism to enhance feature interaction across different viewpoints, we obtain a robust 3D-aware feature representation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 429, + 266, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 429, + 266, + 440 + ], + "spans": [ + { + "bbox": [ + 67, + 429, + 266, + 440 + ], + "type": "text", + "content": "Our contributions can be summarized as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 445, + 295, + 613 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 55, + 445, + 295, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 445, + 295, + 506 + ], + "spans": [ + { + "bbox": [ + 55, + 445, + 295, + 506 + ], + "type": "text", + "content": "- We propose Splatter-360, a generalizable 3D Gaussian Splatting method for wide-baseline panoramic images, which is an end-to-end trainable network that can generate 3DGS primitives for novel panoramic view synthesis and enables real-time rendering experience." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 506, + 295, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 506, + 295, + 553 + ], + "spans": [ + { + "bbox": [ + 55, + 506, + 295, + 553 + ], + "type": "text", + "content": "- We introduce a spherical cost volume based on a spherical sweep algorithm in the feed-forward Gaussian splattering framework, which improves the geometry estimation capabilities of the network." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 553, + 295, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 553, + 295, + 613 + ], + "spans": [ + { + "bbox": [ + 55, + 553, + 295, + 613 + ], + "type": "text", + "content": "- Extensive experiments conducted on the HM3D [26] and Replica [27] datasets demonstrate that our Splatter-360 significantly outperforms state-of-the-art methods in handling wide-baseline panoramic images, both in terms of synthesis quality and generalization performance." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 618, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 295, + 714 + ], + "type": "text", + "content": "In parallel with our research, PanSplat [47] introduces a hierarchical architecture for high-resolution rendering. Our approach diverges by centering on a distortion-aware feature encoding mechanism. Furthermore, Splatter-360 adopts training under randomized camera baselines combined with a larger-scale dataset [26]. This strategic design choice enhances generalization capabilities by exposing the model to broader view variation patterns." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 71, + 400, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 71, + 400, + 83 + ], + "spans": [ + { + "bbox": [ + 314, + 71, + 400, + 83 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 91, + 504, + 104 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 91, + 504, + 104 + ], + "spans": [ + { + "bbox": [ + 313, + 91, + 504, + 104 + ], + "type": "text", + "content": "2.1. Generalizable Novel View Synthesis" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 110, + 553, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 110, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 313, + 110, + 553, + 217 + ], + "type": "text", + "content": "Recent advancements in novel view synthesis have been largely driven by NeRF and 3D Gaussian Splatting (3DGS). These methods can be broadly classified into two categories: per-scene optimization approaches [9, 32, 38] and generalizable models [4, 5, 10, 28, 40, 44]. Generalizable methods [4, 5, 10, 28, 40, 44] enable rapid reconstruction by learning priors from large-scale datasets, allowing for efficient feedforward inference. Below, we provide an overview of generalizable NeRF and 3DGS techniques:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 218, + 553, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 218, + 553, + 384 + ], + "spans": [ + { + "bbox": [ + 313, + 218, + 553, + 384 + ], + "type": "text", + "content": "Generalizable NeRF: Early attempts in generalizable NeRF models aim to reconstruct objects and scenes by generating pixel-aligned features for radiance field prediction, pioneered by [31, 36, 44]. Subsequent work such as NeuRay [24] predicts the visibility of 3D points relative to input images, focusing more on visible features. [10] and [35] further propose a multi-view transformer encoder with epipolar line sampling to capture multi-view geometric priors efficiently. MuRF [40] uses a target view frustum volume aligned with the target image plane to enable sharp, high-quality rendering. More recently, LRM [13] and Instant3D [20] designed a large transformer model to regress triplane NeRF feature for single- or sparse-view reconstruction." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 386, + 553, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 386, + 553, + 615 + ], + "spans": [ + { + "bbox": [ + 313, + 386, + 553, + 615 + ], + "type": "text", + "content": "Generalizable 3DGS: To address NeRF's high computational demands, generalizable 3DGS has emerged to simplify view synthesis through rasterization-based splatting, bypassing the need for expensive volume sampling. Methods like Splatter Image [29] demonstrate efficient object-level reconstruction by learning Gaussian parameters from single views. In contrast, pixelSplat [4] and Flash3D [28] have pushed this technique to scene-level reconstruction. More recently, [48] extends the paradigm of LRM [13] to scene-level with 3DGS, MVSplat [5] introduces a cost volume representation using plane sweeping to boost performance further. HiSplat [30] proposes a hierarchical structure for generalizable 3DGS. DepthSplat [41] leverages depth estimation to address multi-view depth methods' limitations and failure cases. However, most of these methods are designed for perspective images and struggle with panoramic inputs due to the large field of view and wide baselines. In response, we propose Splatter-360, extending the strengths of 3DGS to panoramic images." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 624, + 537, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 624, + 537, + 635 + ], + "spans": [ + { + "bbox": [ + 313, + 624, + 537, + 635 + ], + "type": "text", + "content": "2.2. Panoramic View Synthesis and Generation" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 641, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 553, + 713 + ], + "type": "text", + "content": "Unlike perspective, novel view synthesis, panoramic novel view synthesis introduces unique challenges due to the distortions caused by equirectangular projection. Early works [1, 12] attempted to synthesize panoramic views using multi-sphere images. More recent approaches have focused on reconstructing radiance fields or 3D Gaussian" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21591" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 324 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 324 + ], + "type": "text", + "content": "scenes from dense panoramic inputs [2, 11]. The challenge intensifies with sparse " + }, + { + "bbox": [ + 55, + 72, + 297, + 324 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 55, + 72, + 297, + 324 + ], + "type": "text", + "content": " inputs, as depth estimation becomes significantly more difficult. A few methods seek to address this by leveraging the predicted depth and geometric warping to construct neural radiance fields for panoramic novel view synthesis [3, 18]. 360Roam [15] proposes to adapt neural rendering methods for panoramic inputs, struggling in wide-baseline scenarios. PERF [34], on the other hand, explores an inpainting-based approach to generate neural radiance fields from a single panoramic image. Most relevant to our work is PanoGRF [8], which introduces Generalizable Spherical Radiance Fields for wide-baseline panoramicas. PanoGRF achieves state-of-the-art performance by directly aggregating geometry and appearance features of 3D sample points from each panoramic view. However, its inference and rendering speed are limited, making it unsuitable for real-time applications. In contrast, our work proposes a generalizable panoramic view synthesis pipeline designed to generate 3D Gaussian primitives, carefully considering network design and inherent real-time performance during rendering." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 349, + 296, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 349, + 296, + 517 + ], + "spans": [ + { + "bbox": [ + 55, + 349, + 296, + 517 + ], + "type": "text", + "content": "In contrast to the generalizable setting, several works have explored text-based panoramic view generation, albeit with distinct objectives. Nonetheless, these approaches often share similar network design philosophies. For example, Text2Light [7] focuses on high dynamic range panorama generation, while FastScene [25] employs coarse view synthesis followed by progressive inpainting-based generation. DiffPano [43] introduces a spherical epipolar-aware diffusion model for panorama synthesis. Additionally, Panfusion [46], Dreamscene360 [49], and SceneDreamer360 [21] propose novel architectures built on 2D diffusion models. Although these works are orthogonal to ours, they share several challenges in handling panoramic view synthesis." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 545, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 296, + 715 + ], + "type": "text", + "content": "Concurrent work. While preparing our work, we found several concurrent works addressing similar challenges using diffusion priors. For instance, ViewCrafter [45], GaussianEnhancer [23], ReconX [22], and FreeVS [37] all use video diffusion models to improve the novel view synthesis. However, these works require computationally intensive de-noising and are not designed for panoramic view synthesis. A very recent work, MVSplat-360 [6], focuses on synthesizing novel perspective images that enable arbitrary position and viewpoint exploration (i.e., " + }, + { + "bbox": [ + 55, + 545, + 296, + 715 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 55, + 545, + 296, + 715 + ], + "type": "text", + "content": " navigation) using video diffusion priors, aligning with the goals of aforementioned methods but in generalizable 3DGS setting. In contrast, our work specifically tackles generating novel views for panoramic images (i.e., " + }, + { + "bbox": [ + 55, + 545, + 296, + 715 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 55, + 545, + 296, + 715 + ], + "type": "text", + "content": " images)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 314, + 71, + 421, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 71, + 421, + 85 + ], + "spans": [ + { + "bbox": [ + 314, + 71, + 421, + 85 + ], + "type": "text", + "content": "3. Proposed Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 91, + 555, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 91, + 555, + 248 + ], + "spans": [ + { + "bbox": [ + 313, + 91, + 555, + 248 + ], + "type": "text", + "content": "Our proposed framework, Splatter-360, is designed to synthesize novel views from wide-baseline " + }, + { + "bbox": [ + 313, + 91, + 555, + 248 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 313, + 91, + 555, + 248 + ], + "type": "text", + "content": " panoramic images by leveraging the strengths of 3DGS. Unlike conventional 3DGS methods [5, 41], which are trained on perspective images, Splatter-360 operates directly on panoramic images, thereby mitigating the information loss associated with perspective transformation. The overall pipeline is depicted in Fig. 1. A detailed explanation of the 3D-aware biprojection encoding is provided in Sec. 3.1, spherical depth estimation is discussed in Sec. 3.2, and the pixel-aligned Gaussian decoding is outlined in Sec. 3.3. Additionally, preliminary on 3DGS is included in the supplementary material to ensure this paper is self-contained." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 253, + 497, + 266 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 253, + 497, + 266 + ], + "spans": [ + { + "bbox": [ + 313, + 253, + 497, + 266 + ], + "type": "text", + "content": "3.1. 3D-Aware Bi-Projection Encoding" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 270, + 556, + 377 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 270, + 556, + 377 + ], + "spans": [ + { + "bbox": [ + 313, + 270, + 556, + 377 + ], + "type": "text", + "content": "Although equirectangular projection (ERP) provides a wide field of view and seamless image continuity, the significant distortions, especially near the poles, hinder the network's ability to learn meaningful features effectively. To address this, we propose a bi-projection encoder that leverages both the advantages of ERP and the complementary strengths of cube-map projection (CP). By extracting auxiliary features through CP, our approach enhances the network's capacity to handle these distortions better." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 379, + 556, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 379, + 556, + 534 + ], + "spans": [ + { + "bbox": [ + 313, + 379, + 556, + 534 + ], + "type": "text", + "content": "The overall structure of the bi-projection encoder is illustrated on the left side of Fig. 1. In each branch of the bi-projection network, we begin by applying the convolutional neural network from Unimatch [39] to independently extract local features from both the ERP and CP representations. Unlike prior panorama feature extraction methods [16, 33], which primarily focus on 2D features, our network is designed to learn 3D-aware features. To achieve this, we employ a cross-view attention mechanism to facilitate feature interactions between different viewpoints in the ERP and across the various views in the CP. We obtained " + }, + { + "bbox": [ + 313, + 379, + 556, + 534 + ], + "type": "inline_equation", + "content": "F_{ERP}" + }, + { + "bbox": [ + 313, + 379, + 556, + 534 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 379, + 556, + 534 + ], + "type": "inline_equation", + "content": "F_{CP}" + }, + { + "bbox": [ + 313, + 379, + 556, + 534 + ], + "type": "text", + "content": " after the cross-view attention module. " + }, + { + "bbox": [ + 313, + 379, + 556, + 534 + ], + "type": "inline_equation", + "content": "F_{CP}" + }, + { + "bbox": [ + 313, + 379, + 556, + 534 + ], + "type": "text", + "content": " is then stitched up into ERP feature " + }, + { + "bbox": [ + 313, + 379, + 556, + 534 + ], + "type": "inline_equation", + "content": "F_{C2E}" + }, + { + "bbox": [ + 313, + 379, + 556, + 534 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 534, + 557, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 557, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 557, + 714 + ], + "type": "text", + "content": "For wide-baseline panoramas, naive " + }, + { + "bbox": [ + 313, + 534, + 557, + 714 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 313, + 534, + 557, + 714 + ], + "type": "text", + "content": " multi-view matching methods are hard to handle occlusions and textureless areas. To enhance the 3D-aware capability of the encoder in reasoning about geometric details, we introduce the CP single-view depth encoder. This is accomplished by incorporating the single-view geometric priors from a pre-trained monocular depth network into the CP branch at the feature level. Specifically, we extract monocular depth features from each perspective view using a pre-trained depth estimation network DepthAnythingV2 [42]. Once the monocular depth feature maps " + }, + { + "bbox": [ + 313, + 534, + 557, + 714 + ], + "type": "inline_equation", + "content": "F_{CP}^{mono}" + }, + { + "bbox": [ + 313, + 534, + 557, + 714 + ], + "type": "text", + "content": " for all perspective views are obtained, we stitch these CP features into the ERP view, denoted as " + }, + { + "bbox": [ + 313, + 534, + 557, + 714 + ], + "type": "inline_equation", + "content": "F_{C2E}^{mono}" + }, + { + "bbox": [ + 313, + 534, + 557, + 714 + ], + "type": "text", + "content": ". Next, we concatenate " + }, + { + "bbox": [ + 313, + 534, + 557, + 714 + ], + "type": "inline_equation", + "content": "F_{C2E}^{mono}" + }, + { + "bbox": [ + 313, + 534, + 557, + 714 + ], + "type": "text", + "content": " with the transformed CP feature " + }, + { + "bbox": [ + 313, + 534, + 557, + 714 + ], + "type": "inline_equation", + "content": "F_{C2E}" + }, + { + "bbox": [ + 313, + 534, + 557, + 714 + ], + "type": "text", + "content": ", and use a two-layer MLP to get the final CP branch feature:" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "21592" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 71, + 552, + 223 + ], + "blocks": [ + { + "bbox": [ + 56, + 71, + 552, + 223 + ], + "lines": [ + { + "bbox": [ + 56, + 71, + 552, + 223 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 552, + 223 + ], + "type": "image", + "image_path": "b02674da8fbb225841a82ba6e3fd2f9b5f5850a83392af0434f33072f74cd0df.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 230, + 555, + 289 + ], + "lines": [ + { + "bbox": [ + 54, + 230, + 555, + 289 + ], + "spans": [ + { + "bbox": [ + 54, + 230, + 555, + 289 + ], + "type": "text", + "content": "Figure 1. Our Splatter-360 processes " + }, + { + "bbox": [ + 54, + 230, + 555, + 289 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 54, + 230, + 555, + 289 + ], + "type": "text", + "content": " panoramic images using a bi-projection encoder that extracts features from both equirectangular projection (ERP) and cube-map projection (CP) through multi-view transformers. These features are used for spherical cost volume construction, and multi-view matching is performed between the reference and source views in spherical space. Next, a refinement U-Net is applied to enhance the spherical cost volume, yielding refined cost volumes and more accurate spherical depth estimations. These refined outputs are then fed into the Gaussian decoder, which produces pixel-aligned Gaussian primitives for synthesizing novel views." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 111, + 318, + 294, + 332 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 318, + 294, + 332 + ], + "spans": [ + { + "bbox": [ + 111, + 318, + 294, + 332 + ], + "type": "interline_equation", + "content": "\\boldsymbol {F} _ {C 2 E} ^ {\\prime} = \\mathcal {F} _ {1} \\left(\\left[ \\boldsymbol {F} _ {C 2 E} ^ {m o n o}, \\boldsymbol {F} _ {C 2 E} \\right]\\right), \\tag {1}", + "image_path": "1b9c271280c5982d0cfa352f4e049ce289d49421e18d12337b213e0c216d8b21.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 336, + 296, + 395 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 336, + 296, + 395 + ], + "spans": [ + { + "bbox": [ + 55, + 336, + 296, + 395 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 336, + 296, + 395 + ], + "type": "inline_equation", + "content": "[.,.]" + }, + { + "bbox": [ + 55, + 336, + 296, + 395 + ], + "type": "text", + "content": " denotes the concatenation operation and " + }, + { + "bbox": [ + 55, + 336, + 296, + 395 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_1" + }, + { + "bbox": [ + 55, + 336, + 296, + 395 + ], + "type": "text", + "content": " is a two-layer MLP. Then, we propose a fusion module, " + }, + { + "bbox": [ + 55, + 336, + 296, + 395 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_2" + }, + { + "bbox": [ + 55, + 336, + 296, + 395 + ], + "type": "text", + "content": ", to fuse the CP branch feature " + }, + { + "bbox": [ + 55, + 336, + 296, + 395 + ], + "type": "inline_equation", + "content": "F_{C2E}^{\\prime}" + }, + { + "bbox": [ + 55, + 336, + 296, + 395 + ], + "type": "text", + "content": " and ERP branch feature " + }, + { + "bbox": [ + 55, + 336, + 296, + 395 + ], + "type": "inline_equation", + "content": "F_{ERP}" + }, + { + "bbox": [ + 55, + 336, + 296, + 395 + ], + "type": "text", + "content": ". The final output feature of our encoder is represented as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 123, + 396, + 295, + 409 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 123, + 396, + 295, + 409 + ], + "spans": [ + { + "bbox": [ + 123, + 396, + 295, + 409 + ], + "type": "interline_equation", + "content": "\\boldsymbol {F} = \\mathcal {F} _ {2} \\left(\\boldsymbol {F} _ {C 2 E} ^ {\\prime}, \\boldsymbol {F} _ {E R P}\\right), \\tag {2}", + "image_path": "3c472eea9e9662e152d1131a9c11b1d400e49c624688927da21c6794235940c5.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 414, + 296, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 414, + 296, + 450 + ], + "spans": [ + { + "bbox": [ + 55, + 414, + 296, + 450 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 414, + 296, + 450 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_2" + }, + { + "bbox": [ + 55, + 414, + 296, + 450 + ], + "type": "text", + "content": " is the proposed fusion module, which consists of convolutional layers augmented with squeeze-and-excitation blocks [14, 16]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 456, + 209, + 470 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 456, + 209, + 470 + ], + "spans": [ + { + "bbox": [ + 55, + 456, + 209, + 470 + ], + "type": "text", + "content": "3.2. Spherical Depth Estimation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 474, + 295, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 474, + 295, + 653 + ], + "spans": [ + { + "bbox": [ + 55, + 474, + 295, + 653 + ], + "type": "text", + "content": "After obtaining a robust feature representation for panoramic images, the next step is cost volume construction. One straightforward approach is to convert the panorama into cube-map-like perspective images and then follow the perspective-based methods such as MVSplat [5] and DepthSplat [41] to construct cost volume. However, such a method has an unavoidable shortcoming: for each 3D sampling point in the planar cost volume of the reference view, these points may project behind the camera in the source view (i.e., " + }, + { + "bbox": [ + 55, + 474, + 295, + 653 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 55, + 474, + 295, + 653 + ], + "type": "text", + "content": "-depth " + }, + { + "bbox": [ + 55, + 474, + 295, + 653 + ], + "type": "inline_equation", + "content": "< 0" + }, + { + "bbox": [ + 55, + 474, + 295, + 653 + ], + "type": "text", + "content": "). When this occurs, incorrect local image features from the source view are sampled, leading to inaccurate local similarity computations and, ultimately, erroneous depth estimates. This issue is a consequence of the narrow field-of-view plane sweep algorithm traditionally used for the perspective-based method." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": "To address this, we leverage the 360-degree field of view inherent to panoramas and apply the spherical projection formula derived from [43] to construct a spherical cost volume. More concretely, the spherical sweep algorithm samples depth candidates within the spherical domain, comput-" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 307, + 555, + 367 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 307, + 555, + 367 + ], + "spans": [ + { + "bbox": [ + 313, + 307, + 555, + 367 + ], + "type": "text", + "content": "ing feature correlations across multiple views to construct the cost volume. For each reference view, we sample " + }, + { + "bbox": [ + 313, + 307, + 555, + 367 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 313, + 307, + 555, + 367 + ], + "type": "text", + "content": " depth candidates in the spherical space, where the depth values " + }, + { + "bbox": [ + 313, + 307, + 555, + 367 + ], + "type": "inline_equation", + "content": "r_m \\in [r_{\\mathrm{near}}, r_{\\mathrm{far}}]" + }, + { + "bbox": [ + 313, + 307, + 555, + 367 + ], + "type": "text", + "content": " are sampled in the logarithmic space between the near range " + }, + { + "bbox": [ + 313, + 307, + 555, + 367 + ], + "type": "inline_equation", + "content": "r_{\\mathrm{near}}" + }, + { + "bbox": [ + 313, + 307, + 555, + 367 + ], + "type": "text", + "content": " and the far range " + }, + { + "bbox": [ + 313, + 307, + 555, + 367 + ], + "type": "inline_equation", + "content": "r_{\\mathrm{far}}" + }, + { + "bbox": [ + 313, + 307, + 555, + 367 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 367, + 555, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 367, + 555, + 392 + ], + "spans": [ + { + "bbox": [ + 313, + 367, + 555, + 392 + ], + "type": "text", + "content": "To transform the equirectangular coordinates to spherical coordinates, we use the following equations:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 387, + 400, + 554, + 444 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 387, + 400, + 554, + 444 + ], + "spans": [ + { + "bbox": [ + 387, + 400, + 554, + 444 + ], + "type": "interline_equation", + "content": "\\left\\{ \\begin{array}{l} \\theta = (0. 5 - \\frac {u}{W}) \\cdot 2 \\pi \\\\ \\phi = (0. 5 - \\frac {v}{H}) \\cdot \\pi , \\end{array} \\right. \\tag {3}", + "image_path": "86884ae38a6eeffdba1b1de78c05564d92ad58d67db3e44119945bfb212d5ed7.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 449, + 554, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 449, + 554, + 508 + ], + "spans": [ + { + "bbox": [ + 313, + 449, + 554, + 508 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 449, + 554, + 508 + ], + "type": "inline_equation", + "content": "u" + }, + { + "bbox": [ + 313, + 449, + 554, + 508 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 449, + 554, + 508 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 313, + 449, + 554, + 508 + ], + "type": "text", + "content": " represent the pixel coordinates in the equirectangular grid, while " + }, + { + "bbox": [ + 313, + 449, + 554, + 508 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 449, + 554, + 508 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 449, + 554, + 508 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 313, + 449, + 554, + 508 + ], + "type": "text", + "content": " denote the longitude and latitude, respectively. Additionally, " + }, + { + "bbox": [ + 313, + 449, + 554, + 508 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 313, + 449, + 554, + 508 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 449, + 554, + 508 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 313, + 449, + 554, + 508 + ], + "type": "text", + "content": " refer to the width and height of the ERP feature maps on which multiview matching is performed." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 509, + 555, + 546 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 509, + 555, + 546 + ], + "spans": [ + { + "bbox": [ + 313, + 509, + 555, + 546 + ], + "type": "text", + "content": "Next, we transform the spherical polar coordinates " + }, + { + "bbox": [ + 313, + 509, + 555, + 546 + ], + "type": "inline_equation", + "content": "(r,\\theta ,\\phi)" + }, + { + "bbox": [ + 313, + 509, + 555, + 546 + ], + "type": "text", + "content": " into Cartesian coordinates " + }, + { + "bbox": [ + 313, + 509, + 555, + 546 + ], + "type": "inline_equation", + "content": "(x,y,z)" + }, + { + "bbox": [ + 313, + 509, + 555, + 546 + ], + "type": "text", + "content": " to obtain the camera coordinates " + }, + { + "bbox": [ + 313, + 509, + 555, + 546 + ], + "type": "inline_equation", + "content": "p_{\\mathrm{camera}}^i" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 376, + 554, + 554, + 598 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 376, + 554, + 554, + 598 + ], + "spans": [ + { + "bbox": [ + 376, + 554, + 554, + 598 + ], + "type": "interline_equation", + "content": "\\left\\{ \\begin{array}{l} x _ {c a m} = r \\cos (\\phi) \\cdot \\sin (\\theta) \\\\ y _ {c a m} = r \\sin (\\phi) \\\\ z _ {c a m} = r \\cos (\\phi) \\cdot \\cos (\\theta), \\end{array} \\right. \\tag {4}", + "image_path": "e8de0ba919dfee2f42bca5a445c9b5732c1a56f8b47d46f2f0137c01a2823576.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 608, + 555, + 656 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 608, + 555, + 656 + ], + "spans": [ + { + "bbox": [ + 313, + 608, + 555, + 656 + ], + "type": "text", + "content": "To obtain the pixel coordinates " + }, + { + "bbox": [ + 313, + 608, + 555, + 656 + ], + "type": "inline_equation", + "content": "(u,v)" + }, + { + "bbox": [ + 313, + 608, + 555, + 656 + ], + "type": "text", + "content": " at the source view " + }, + { + "bbox": [ + 313, + 608, + 555, + 656 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 608, + 555, + 656 + ], + "type": "text", + "content": ", we back-project the camera coordinates using the relative camera pose " + }, + { + "bbox": [ + 313, + 608, + 555, + 656 + ], + "type": "inline_equation", + "content": "W^{i\\rightarrow j}" + }, + { + "bbox": [ + 313, + 608, + 555, + 656 + ], + "type": "text", + "content": " between the reference view and the source view:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 378, + 665, + 553, + 681 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 378, + 665, + 553, + 681 + ], + "spans": [ + { + "bbox": [ + 378, + 665, + 553, + 681 + ], + "type": "interline_equation", + "content": "\\boldsymbol {p} _ {\\text {c a m e r a}} ^ {j} = W ^ {i \\rightarrow j} \\boldsymbol {p} _ {\\text {c a m e r a}} ^ {i} \\tag {5}", + "image_path": "a8fa3cdd09eabd66691a7ca735e5a631ae20cf58e19fdf5d4fc9049a69f3c8a9.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": "With known corresponding pixel coordinates, we sample the local image feature " + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "inline_equation", + "content": "F^{j\\rightarrow i}" + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": " from the source view feature" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "21593" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 97 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 97 + ], + "type": "text", + "content": "and compute the feature similarity between " + }, + { + "bbox": [ + 55, + 72, + 297, + 97 + ], + "type": "inline_equation", + "content": "F^{j\\rightarrow i}" + }, + { + "bbox": [ + 55, + 72, + 297, + 97 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 72, + 297, + 97 + ], + "type": "inline_equation", + "content": "F^i" + }, + { + "bbox": [ + 55, + 72, + 297, + 97 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 65, + 103, + 295, + 133 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 103, + 295, + 133 + ], + "spans": [ + { + "bbox": [ + 65, + 103, + 295, + 133 + ], + "type": "interline_equation", + "content": "C _ {r _ {m}} ^ {i} = \\frac {\\boldsymbol {F} ^ {i} \\cdot \\boldsymbol {F} _ {r _ {m}} ^ {j \\rightarrow i}}{\\sqrt {C}} \\in \\mathbb {R} ^ {H \\times W}, \\quad m = 1, 2, \\dots , D, \\tag {6}", + "image_path": "c9d06cd77b7b57264473316ff676fb6880ee813a95d2fd9053be0831400baf25.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 142, + 296, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 142, + 296, + 191 + ], + "spans": [ + { + "bbox": [ + 55, + 142, + 296, + 191 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 142, + 296, + 191 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 55, + 142, + 296, + 191 + ], + "type": "text", + "content": " denotes the total feature channel and " + }, + { + "bbox": [ + 55, + 142, + 296, + 191 + ], + "type": "inline_equation", + "content": "C_{r_m}^i" + }, + { + "bbox": [ + 55, + 142, + 296, + 191 + ], + "type": "text", + "content": " denotes the feature similarity between " + }, + { + "bbox": [ + 55, + 142, + 296, + 191 + ], + "type": "inline_equation", + "content": "F^{j\\rightarrow i}" + }, + { + "bbox": [ + 55, + 142, + 296, + 191 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 142, + 296, + 191 + ], + "type": "inline_equation", + "content": "F^i" + }, + { + "bbox": [ + 55, + 142, + 296, + 191 + ], + "type": "text", + "content": " at the " + }, + { + "bbox": [ + 55, + 142, + 296, + 191 + ], + "type": "inline_equation", + "content": "m^{th}" + }, + { + "bbox": [ + 55, + 142, + 296, + 191 + ], + "type": "text", + "content": " depth candidate. Then, we can concatenate each depth candidates' feature similarity to get the spherical cost volume." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 88, + 199, + 295, + 216 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 199, + 295, + 216 + ], + "spans": [ + { + "bbox": [ + 88, + 199, + 295, + 216 + ], + "type": "interline_equation", + "content": "\\boldsymbol {C} ^ {i} = \\left[ \\boldsymbol {C} _ {r _ {1}} ^ {i}, \\boldsymbol {C} _ {r _ {2}} ^ {i}, \\dots , \\boldsymbol {C} _ {r _ {D}} ^ {i} \\right] \\in \\mathbb {R} ^ {H \\times W \\times D}. \\tag {7}", + "image_path": "06520dcb0e66248b208babef4f52ee505ca34854a2a4523632d580024716a91a.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 224, + 296, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 224, + 296, + 297 + ], + "spans": [ + { + "bbox": [ + 55, + 224, + 296, + 297 + ], + "type": "text", + "content": "As the initial spherical cost volume only considers the pixel-level similarity and is hard to handle occlusions and textureless areas. We further use a U-Net [5] to refine it. Specifically, the proposed U-Net mainly learns a residual volume value " + }, + { + "bbox": [ + 55, + 224, + 296, + 297 + ], + "type": "inline_equation", + "content": "\\Delta C^i" + }, + { + "bbox": [ + 55, + 224, + 296, + 297 + ], + "type": "text", + "content": ", which can be added to the initial spherical cost volume to get the final one:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 110, + 307, + 295, + 323 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 307, + 295, + 323 + ], + "spans": [ + { + "bbox": [ + 110, + 307, + 295, + 323 + ], + "type": "interline_equation", + "content": "\\tilde {\\boldsymbol {C}} ^ {i} = \\boldsymbol {C} ^ {i} + \\Delta \\boldsymbol {C} ^ {i} \\in \\mathbb {R} ^ {H \\times W \\times D}. \\tag {8}", + "image_path": "ce82908dfe34e3c30bb017cbe70890f94705429839b9eaaf0bb2f5d0d7366934.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 330, + 296, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 330, + 296, + 354 + ], + "spans": [ + { + "bbox": [ + 55, + 330, + 296, + 354 + ], + "type": "text", + "content": "Finally, we can use the softmax operation to get the final spherical depth estimation:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 107, + 364, + 295, + 380 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 364, + 295, + 380 + ], + "spans": [ + { + "bbox": [ + 107, + 364, + 295, + 380 + ], + "type": "interline_equation", + "content": "\\boldsymbol {D} ^ {i} = \\operatorname {s o f t m a x} \\left(\\tilde {\\boldsymbol {C}} ^ {i}\\right) \\boldsymbol {G} \\in \\mathbb {R} ^ {H \\times W}, \\tag {9}", + "image_path": "27ae84b562475f92abf4fd58469630d6d6447b1e214a9a0d19fdff39fc3747c5.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 390, + 296, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 390, + 296, + 439 + ], + "spans": [ + { + "bbox": [ + 55, + 390, + 296, + 439 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 390, + 296, + 439 + ], + "type": "inline_equation", + "content": "\\pmb{G} = [r_1, r_2, \\dots, r_D] \\in \\mathbb{R}^D" + }, + { + "bbox": [ + 55, + 390, + 296, + 439 + ], + "type": "text", + "content": " is the predefined depth candidates and softmax denotes the softmax function which can transform spherical cost volume to the probability distribution for each depth candidates." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 447, + 295, + 460 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 447, + 295, + 460 + ], + "spans": [ + { + "bbox": [ + 55, + 447, + 295, + 460 + ], + "type": "text", + "content": "3.3. Pixel-Aligned Gaussian in Equirect-Coordinate" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 465, + 296, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 465, + 296, + 538 + ], + "spans": [ + { + "bbox": [ + 55, + 465, + 296, + 538 + ], + "type": "text", + "content": "Once the 3D-aware feature representation and spherical cost volume are obtained, we proceed to predict pixel-aligned Gaussians on the equirectangular coordinate grid. Specifically, for each pixel, we estimate Gaussian centers, quaternions, spherical harmonics (SH), and scaling factors, respectively." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 541, + 296, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 541, + 296, + 625 + ], + "spans": [ + { + "bbox": [ + 55, + 541, + 296, + 625 + ], + "type": "text", + "content": "Gaussian centers " + }, + { + "bbox": [ + 55, + 541, + 296, + 625 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 55, + 541, + 296, + 625 + ], + "type": "text", + "content": ". We use the estimated spherical depth " + }, + { + "bbox": [ + 55, + 541, + 296, + 625 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 55, + 541, + 296, + 625 + ], + "type": "text", + "content": " to calculate the gaussian centers. Specifically, for each pixel, we first compute the camera cartesian coordinates " + }, + { + "bbox": [ + 55, + 541, + 296, + 625 + ], + "type": "inline_equation", + "content": "p_{\\mathrm{camera}}" + }, + { + "bbox": [ + 55, + 541, + 296, + 625 + ], + "type": "text", + "content": " of the gaussian center using its spherical depth " + }, + { + "bbox": [ + 55, + 541, + 296, + 625 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 55, + 541, + 296, + 625 + ], + "type": "text", + "content": " via Eq. 3 and Eq. 4. Next, the world coordinates of the Gaussian center are calculated as " + }, + { + "bbox": [ + 55, + 541, + 296, + 625 + ], + "type": "inline_equation", + "content": "p_{\\mathrm{world}} = W p_{\\mathrm{camera}}" + }, + { + "bbox": [ + 55, + 541, + 296, + 625 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 541, + 296, + 625 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 55, + 541, + 296, + 625 + ], + "type": "text", + "content": " represents the camera-to-world transformation matrix." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 630, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 296, + 715 + ], + "type": "text", + "content": "Opacity " + }, + { + "bbox": [ + 55, + 630, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 630, + 296, + 715 + ], + "type": "text", + "content": ". The opacity is calculated according to the matching confidence. Specifically, we first use the probability distribution of spherical cost volume(i.e., the softmax output of Eq. 9) to get the matching confidence. Then, we feed the matching confidence along with the image features into a two-layer convolutional network to predict the opacity." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 72, + 554, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 131 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 131 + ], + "type": "text", + "content": "Covariance " + }, + { + "bbox": [ + 313, + 72, + 554, + 131 + ], + "type": "inline_equation", + "content": "\\Sigma" + }, + { + "bbox": [ + 313, + 72, + 554, + 131 + ], + "type": "text", + "content": " and SH " + }, + { + "bbox": [ + 313, + 72, + 554, + 131 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 72, + 554, + 131 + ], + "type": "text", + "content": ". We predict these parameters using two convolutional layers, which take as input the concatenation of image features, the refined cost volume, and the original ERP images. The covariance matrix " + }, + { + "bbox": [ + 313, + 72, + 554, + 131 + ], + "type": "inline_equation", + "content": "\\Sigma" + }, + { + "bbox": [ + 313, + 72, + 554, + 131 + ], + "type": "text", + "content": " is computed as:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 380, + 132, + 555, + 147 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 380, + 132, + 555, + 147 + ], + "spans": [ + { + "bbox": [ + 380, + 132, + 555, + 147 + ], + "type": "interline_equation", + "content": "\\Sigma = R (\\theta) ^ {\\top} \\operatorname {d i a g} (s) R (\\theta), \\tag {10}", + "image_path": "a7ad8d5f8cce1fd087ec504b297d1ab562b52a5bb9c1d6368b15f42b16367db8.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 152, + 555, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 152, + 555, + 213 + ], + "spans": [ + { + "bbox": [ + 313, + 152, + 555, + 213 + ], + "type": "text", + "content": "following the formulation in [4]. Here, " + }, + { + "bbox": [ + 313, + 152, + 555, + 213 + ], + "type": "inline_equation", + "content": "R(\\theta)" + }, + { + "bbox": [ + 313, + 152, + 555, + 213 + ], + "type": "text", + "content": " is the rotation matrix parameterized by quaternions, " + }, + { + "bbox": [ + 313, + 152, + 555, + 213 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 313, + 152, + 555, + 213 + ], + "type": "text", + "content": " represents the scaling matrix. The " + }, + { + "bbox": [ + 313, + 152, + 555, + 213 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 152, + 555, + 213 + ], + "type": "text", + "content": " (spherical harmonics, SH) is generated directly by our network for direction-dependent color encoding." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 224, + 395, + 238 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 224, + 395, + 238 + ], + "spans": [ + { + "bbox": [ + 313, + 224, + 395, + 238 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 245, + 487, + 257 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 245, + 487, + 257 + ], + "spans": [ + { + "bbox": [ + 313, + 245, + 487, + 257 + ], + "type": "text", + "content": "4.1. Datasets, Baselines, and Metrics" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 263, + 554, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 263, + 554, + 287 + ], + "spans": [ + { + "bbox": [ + 313, + 263, + 554, + 287 + ], + "type": "text", + "content": "Datasets. We evaluate Splatter-360 on two large-scale panoramic datasets: HM3D [26] and Replica [27]." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 291, + 554, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 291, + 554, + 327 + ], + "spans": [ + { + "bbox": [ + 313, + 291, + 554, + 327 + ], + "type": "text", + "content": "Metrics. We evaluate the performance of Splatter-360 using standard metrics for novel view synthesis, including PSNR, SSIM, and LPIPS." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 335, + 432, + 348 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 335, + 432, + 348 + ], + "spans": [ + { + "bbox": [ + 313, + 335, + 432, + 348 + ], + "type": "text", + "content": "4.2. Quantitative Results" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 354, + 555, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 354, + 555, + 509 + ], + "spans": [ + { + "bbox": [ + 313, + 354, + 555, + 509 + ], + "type": "text", + "content": "Our results demonstrate that Splatter-360 outperforms existing methods in both synthesis quality and generalization to novel views. As for the perspective-based method, we convert panoramicas to cube maps (12 perspective images) and input these cube maps into HiSplat, DepthSplat, and MVSplat. We did not compare with PixelSplat [4] due to out-of-memory errors during inference with 12 perspective image views. Furthermore, HiSplat and DepthSplat cannot be trained on the HM3D dataset with 12 perspective images due to GPU memory constraints; therefore, we tested their official pre-trained models on Re10K [50] for evaluation on both HM3D and Replica. MVSplat and Splatter-360 are trained on HM3D using 8 Tesla V100 GPUs." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 510, + 555, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 555, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 555, + 665 + ], + "type": "text", + "content": "Table 1 shows that existing perspective-based methods, including HiSplat, DepthSplat, and MVSplat, struggle with generalization to cube-map inputs and cannot be directly applied to panoramic datasets. We tested the fine-tuned version of MVSplat, but it still lags behind Splatter-360 in terms of generalization, both on HM3D and Replica. Specifically, the PSNR of MVSplat is 1.114 dB lower than Splatter-360 on HM3D and 1.489 dB lower on Replica. We also compared it with PanoGRF [8], specifically designed for panoramic inputs. Splatter-360 consistently outperforms PanoGRF across all metrics on Replica and HM3D. On HM3D, the PSNR of Splatter-360 is 2.662 dB higher than PanoGRF, and on Replica, it is 1.968 dB higher." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "type": "text", + "content": "Moreover, Table 2 provides a quantitative comparison of novel-view depth estimation between MVSplat and Splatter-360. Splatter-360 consistently outperforms MVSplat across all depth metrics on HM3D and Replica, verify" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "21594" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 70, + 553, + 297 + ], + "blocks": [ + { + "bbox": [ + 59, + 70, + 553, + 297 + ], + "lines": [ + { + "bbox": [ + 59, + 70, + 553, + 297 + ], + "spans": [ + { + "bbox": [ + 59, + 70, + 553, + 297 + ], + "type": "image", + "image_path": "b8471b2e018f46ba8891aebcc73896ca31df52aae5298e5c6934b85291e410ec.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 306, + 555, + 329 + ], + "lines": [ + { + "bbox": [ + 55, + 306, + 555, + 329 + ], + "spans": [ + { + "bbox": [ + 55, + 306, + 555, + 329 + ], + "type": "text", + "content": "Figure 2. Qualitative comparison between our Splatter-360 and PanoGRF, MVSplat on the Replica dataset. Regions with notable differences are highlighted using red and blue rectangles. Please zoom in for a clearer view." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 58, + 338, + 555, + 569 + ], + "blocks": [ + { + "bbox": [ + 58, + 338, + 555, + 569 + ], + "lines": [ + { + "bbox": [ + 58, + 338, + 555, + 569 + ], + "spans": [ + { + "bbox": [ + 58, + 338, + 555, + 569 + ], + "type": "image", + "image_path": "8c4412d0bb2eecd4fef5a2bfe7bb16da483ad1f92bf611b99a7b565c6651a4d6.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 575, + 555, + 598 + ], + "lines": [ + { + "bbox": [ + 55, + 575, + 555, + 598 + ], + "spans": [ + { + "bbox": [ + 55, + 575, + 555, + 598 + ], + "type": "text", + "content": "Figure 3. Qualitative comparison between our Splatter-360 and PanoGRF, MVSplat on the HM3D dataset. Regions with notable differences are highlighted using red and blue rectangles. Please zoom in for a clearer view." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 620, + 294, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 620, + 294, + 643 + ], + "spans": [ + { + "bbox": [ + 55, + 620, + 294, + 643 + ], + "type": "text", + "content": "ing that the proposed method can better capture geometry for panoramic inputs." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 658, + 167, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 658, + 167, + 670 + ], + "spans": [ + { + "bbox": [ + 55, + 658, + 167, + 670 + ], + "type": "text", + "content": "4.3. Qualitative Results" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": "Qualitative comparisons are provided in Fig. 3 and Fig. 2, with key differences highlighted using red and blue rectangles. As shown in the first sample of Fig. 2, Splatter-360" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 620, + 555, + 679 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 620, + 555, + 679 + ], + "spans": [ + { + "bbox": [ + 313, + 620, + 555, + 679 + ], + "type": "text", + "content": "produces much sharper edges and more accurate textures, particularly on objects like the table and chair. In the second and third samples, PanoGRF demonstrates less defined edges, particularly on the floor, while MAVSplat exhibits noticeable artifacts in the form of floaters near the white wall." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 689, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 556, + 715 + ], + "type": "text", + "content": "A comparison of learned geometry between Splatter-360 and MVSplat is presented in Fig. 4. MVSplat shows signifi" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21595" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 70, + 555, + 552 + ], + "blocks": [ + { + "bbox": [ + 61, + 70, + 555, + 552 + ], + "lines": [ + { + "bbox": [ + 61, + 70, + 555, + 552 + ], + "spans": [ + { + "bbox": [ + 61, + 70, + 555, + 552 + ], + "type": "image", + "image_path": "e096dbcad5c710364b29ebab755d905e0d04da1d24c694d53453566251895d21.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 559, + 555, + 583 + ], + "lines": [ + { + "bbox": [ + 55, + 559, + 555, + 583 + ], + "spans": [ + { + "bbox": [ + 55, + 559, + 555, + 583 + ], + "type": "text", + "content": "Figure 4. Novel view depth comparison between Splatter-360 and PanoGRF on the Replica dataset. \"Pano.\" denotes panoramic view and \"Perspec.\" denotes perspective view." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 603, + 294, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 603, + 294, + 662 + ], + "spans": [ + { + "bbox": [ + 55, + 603, + 294, + 662 + ], + "type": "text", + "content": "cantly poorer depth estimation, as evidenced by its incorrect reconstruction of surfaces such as the lamp and table. In contrast, Splatter-360 delivers more accurate and detailed depth predictions with clearer geometry and well-defined surfaces." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 672, + 149, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 672, + 149, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 672, + 149, + 685 + ], + "type": "text", + "content": "4.4. Ablation Study" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "content": "We validate the effectiveness of different modules proposed in Splatter-360 in this section. Due to the large number" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 603, + 555, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 603, + 555, + 639 + ], + "spans": [ + { + "bbox": [ + 313, + 603, + 555, + 639 + ], + "type": "text", + "content": "of ablation studies and limited resources, we utilized 2 NVIDIA Tesla V100 GPUs for each group of experiments in this section." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 653, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 556, + 715 + ], + "type": "text", + "content": "Spherical cost volume. Eliminating the spherical cost volume leads to a notable decline in performance across all metrics, highlighting its crucial role in the model's ability to generate high-quality reconstructions. Specifically, as shown in the first row of Table 3 , compared to the full" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "21596" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 83, + 102, + 525, + 222 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 95 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 95 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 95 + ], + "type": "text", + "content": "Table 1. Quantitative comparison with baseline methods on the HM3D and Replica datasets. " + }, + { + "bbox": [ + 55, + 70, + 555, + 95 + ], + "type": "inline_equation", + "content": "{}^{ \\dagger }" + }, + { + "bbox": [ + 55, + 70, + 555, + 95 + ], + "type": "text", + "content": " indicates models that were trained by us on the panoramic dataset,whereas for all other methods,we used the pre-trained models provided by the original authors." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 83, + 102, + 525, + 222 + ], + "lines": [ + { + "bbox": [ + 83, + 102, + 525, + 222 + ], + "spans": [ + { + "bbox": [ + 83, + 102, + 525, + 222 + ], + "type": "table", + "html": "
MethodHM3D [26]Replica [27]
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
HiSplat [30]17.2680.6240.48817.1570.6420.417
MVSplat [5]17.5740.6360.44118.0050.6310.512
DepthSplat [41]18.4950.6570.45719.3690.7320.334
PanoGRF [8]25.6310.8130.26827.9200.8920.171
MVSplat† [5]27.1790.8510.17628.3990.9080.115
Splatter-360†28.2930.8750.15529.8880.9240.097
", + "image_path": "801fc2b9bcd777b50f9a575515b84113e949b44e6252256b47022883fc2f72b8.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 56, + 261, + 291, + 387 + ], + "blocks": [ + { + "bbox": [ + 55, + 229, + 296, + 253 + ], + "lines": [ + { + "bbox": [ + 55, + 229, + 296, + 253 + ], + "spans": [ + { + "bbox": [ + 55, + 229, + 296, + 253 + ], + "type": "text", + "content": "Table 2. Estimated depth comparison between MVSplat and Splatter-360 on the Replica and HM3D datasets." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 261, + 291, + 387 + ], + "lines": [ + { + "bbox": [ + 56, + 261, + 291, + 387 + ], + "spans": [ + { + "bbox": [ + 56, + 261, + 291, + 387 + ], + "type": "table", + "html": "
DatasetMetricMVSplatSplatter-360
Replica [27]Abs Diff↓0.1320.102
Abs Rel↓0.0880.063
RMSE↓0.2470.197
δ < 1.25↑89.91394.572
HM3D [26]Abs Diff↓0.1300.106
Abs Rel↓0.0940.076
RMSE↓0.2710.223
δ < 1.25↑90.46993.851
", + "image_path": "1e5bb0f14d3754ce0728224d47527544d9261ef50acbee4b9a7b3d3218666446.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 399, + 295, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 399, + 295, + 422 + ], + "spans": [ + { + "bbox": [ + 55, + 399, + 295, + 422 + ], + "type": "text", + "content": "model, the PSNR drops by 5.271 dB and 2.263 dB on the Replica and HM3D datasets, respectively." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 428, + 295, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 428, + 295, + 583 + ], + "spans": [ + { + "bbox": [ + 54, + 428, + 295, + 583 + ], + "type": "text", + "content": "ERP encoder vs. CP encoder. We separately removed the ERP and CP encoders to evaluate their contributions. The results of these ablations are presented in the third and fourth rows of Table 3. The ERP encoder is crucial for maintaining high PSNR and SSIM values and achieving low LPIPS values, highlighting its vital role in panoramic feature extraction and encoding. In contrast, the removal of the CP encoder results in a smaller performance drop compared to the removal of the ERP encoder. However, removing the CP encoder on the Replica dataset still leads to a PSNR decrease of approximately " + }, + { + "bbox": [ + 54, + 428, + 295, + 583 + ], + "type": "inline_equation", + "content": "0.448\\mathrm{dB}" + }, + { + "bbox": [ + 54, + 428, + 295, + 583 + ], + "type": "text", + "content": ", suggesting that the CP encoder provides auxiliary support in the Gaussian reconstruction of panoramic images." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 589, + 295, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 589, + 295, + 661 + ], + "spans": [ + { + "bbox": [ + 54, + 589, + 295, + 661 + ], + "type": "text", + "content": "Cross-view attention. Corss-view attention is essential to extract the 3D-awareness feature. Here, we test the influence of removing this module. As shown in the second row of Table 3, such a removing resulted in a noticeable decline in PSNR and SSIM, alongside an increase in LPIPS, indicating the effectiveness of cross-view attention." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "content": "Monocular depth network feature. Removing the monocular depth encoder led to a PSNR drop of 0.467 dB in the Replica dataset. This suggests that the pre-trained monocular depth features offer a robust single-view geometry-" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 315, + 284, + 561, + 396 + ], + "blocks": [ + { + "bbox": [ + 313, + 229, + 555, + 274 + ], + "lines": [ + { + "bbox": [ + 313, + 229, + 555, + 274 + ], + "spans": [ + { + "bbox": [ + 313, + 229, + 555, + 274 + ], + "type": "text", + "content": "Table 3. Ablation studies were conducted on the HM3D and Replica datasets. For simplicity, we use the following abbreviations: 'SCV' for spherical cost volume and 'CVA' for cross-view attention." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 284, + 561, + 396 + ], + "lines": [ + { + "bbox": [ + 315, + 284, + 561, + 396 + ], + "spans": [ + { + "bbox": [ + 315, + 284, + 561, + 396 + ], + "type": "table", + "html": "
Ablated moduleReplica [27]HM3D [26]
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
× SCV23.8500.8180.21025.2240.8020.223
× CVA28.2170.9050.12426.9180.8510.182
× ERP26.9850.8870.14225.9050.8270.202
× CP28.6730.9090.11727.2770.8570.174
× Mono Feat.28.6540.9110.11627.3800.8580.173
Full29.1210.9140.11127.4870.8600.171
", + "image_path": "198363ceb65cfcd0ec3a1c3b23585e9bdcfc7d5e2f693e1ba36091673efc0f98.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 417, + 555, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 417, + 555, + 453 + ], + "spans": [ + { + "bbox": [ + 313, + 417, + 555, + 453 + ], + "type": "text", + "content": "aware prior, which is especially advantageous for reconstructing 360-degree sparse views in textureless indoor scenes." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 464, + 388, + 477 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 464, + 388, + 477 + ], + "spans": [ + { + "bbox": [ + 313, + 464, + 388, + 477 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 483, + 555, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 483, + 555, + 616 + ], + "spans": [ + { + "bbox": [ + 313, + 483, + 555, + 616 + ], + "type": "text", + "content": "This paper introduced Splatter-360, a novel framework for generalizable 3D Gaussian Splatting designed for wide-baseline panoramic images. By leveraging 3D-Aware BiProjection Encoding and spherical cost volume, our method significantly improves the quality of novel view synthesis and geometry estimation from panoramic inputs. Experimental results on the HM3D and Replica datasets demonstrate that Splatter-360 sets a new state-of-the-art performance in the panoramic setting, making it a promising solution for applications in VR, simulation rendering, and " + }, + { + "bbox": [ + 313, + 483, + 555, + 616 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 313, + 483, + 555, + 616 + ], + "type": "text", + "content": " video streaming." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 627, + 424, + 640 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 627, + 424, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 627, + 424, + 640 + ], + "type": "text", + "content": "6. Acknowledgement" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 647, + 554, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 647, + 554, + 696 + ], + "spans": [ + { + "bbox": [ + 313, + 647, + 554, + 696 + ], + "type": "text", + "content": "This work was supported by the Natural Science Foundation of China (No. 62132012, 62361146854) and Tsinghua-Tencent Joint Laboratory for Internet Innovation Technology." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "21597" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "text", + "content": "[1] Benjamin Attal, Selena Ling, Aaron Gokaslan, Christian Richardt, and James Tompkin. Matryodshka: Real-time 6dof video view synthesis using multi-sphere images. In European Conference on Computer Vision, pages 441-459. Springer, 2020. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 146, + 296, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 146, + 296, + 190 + ], + "spans": [ + { + "bbox": [ + 61, + 146, + 296, + 190 + ], + "type": "text", + "content": "[2] Jiayang Bai, Letian Huang, Jie Guo, Wen Gong, Yuanqi Li, and Yanwen Guo. 360-gs: Layout-guided panoramic gaussian splatting for indoor roaming. arXiv preprint arXiv:2402.00763, 2024.2, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 191, + 295, + 245 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 191, + 295, + 245 + ], + "spans": [ + { + "bbox": [ + 62, + 191, + 295, + 245 + ], + "type": "text", + "content": "[3] Wenjie Chang, Yueyi Zhang, and Zhiwei Xiong. Depth estimation from indoor panoramas with neural scene representation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 899-908, 2023. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 246, + 295, + 302 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 246, + 295, + 302 + ], + "spans": [ + { + "bbox": [ + 62, + 246, + 295, + 302 + ], + "type": "text", + "content": "[4] David Charatan, Sizhe Lester Li, Andrea Tagliasacchi, and Vincent Sitzmann. pixelsplat: 3d gaussian splats from image pairs for scalable generalizable 3d reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19457-19467, 2024. 2, 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 302, + 294, + 356 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 302, + 294, + 356 + ], + "spans": [ + { + "bbox": [ + 62, + 302, + 294, + 356 + ], + "type": "text", + "content": "[5] Yuedong Chen, Haofei Xu, Chuanxia Zheng, Bohan Zhuang, Marc Pollefeys, Andreas Geiger, Tat-Jen Cham, and Jianfei Cai. Mvsplat: Efficient 3d gaussian splatting from sparse multi-view images. CoRR, abs/2403.14627, 2024. 2, 3, 4, 5, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 357, + 295, + 411 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 357, + 295, + 411 + ], + "spans": [ + { + "bbox": [ + 62, + 357, + 295, + 411 + ], + "type": "text", + "content": "[6] Yuedong Chen, Chuanxia Zheng, Haofei Xu, Bohan Zhuang, Andrea Vedaldi, Tat-Jen Cham, and Jianfei Cai. Mvsplat360: Feed-forward 360 scene synthesis from sparse views. In Advances in Neural Information Processing Systems (NeurIPS), 2024. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 413, + 295, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 413, + 295, + 446 + ], + "spans": [ + { + "bbox": [ + 62, + 413, + 295, + 446 + ], + "type": "text", + "content": "[7] Zhaoxi Chen, Guangcong Wang, and Ziwei Liu. Text2light: Zero-shot text-driven hdr panorama generation. ACM Transactions on Graphics (TOG), 41(6):1-16, 2022. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 447, + 295, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 447, + 295, + 501 + ], + "spans": [ + { + "bbox": [ + 62, + 447, + 295, + 501 + ], + "type": "text", + "content": "[8] Zheng Chen, Yan-Pei Cao, Yuan-Chen Guo, Chen Wang, Ying Shan, and Song-Hai Zhang. Panogrf: generalizable spherical radiance fields for wide-baseline pororamas. Advances in Neural Information Processing Systems, 36:6961-6985, 2023. 1, 3, 5, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 502, + 295, + 556 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 502, + 295, + 556 + ], + "spans": [ + { + "bbox": [ + 62, + 502, + 295, + 556 + ], + "type": "text", + "content": "[9] Kangle Deng, Andrew Liu, Jun-Yan Zhu, and Deva Ramanan. Depth-supervised nerf: Fewer views and faster training for free. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12882-12891, 2022. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 558, + 295, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 558, + 295, + 612 + ], + "spans": [ + { + "bbox": [ + 57, + 558, + 295, + 612 + ], + "type": "text", + "content": "[10] Yilun Du, Cameron Smith, Ayush Tewari, and Vincent Sitzmann. Learning to render novel views from wide-baseline stereo pairs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4970-4980, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 613, + 295, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 613, + 295, + 658 + ], + "spans": [ + { + "bbox": [ + 57, + 613, + 295, + 658 + ], + "type": "text", + "content": "[11] Kai Gu, Thomas Maugey, Sebastian Knorr, and Christine Guillemot. Omni-nerf: neural radiance field from 360 image captures. In 2022 IEEE International Conference on Multimedia and Expo (ICME), pages 1-6. IEEE, 2022. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 658, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 658, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 658, + 295, + 713 + ], + "type": "text", + "content": "[12] Tewodros Habtegebrial, Christiano Gava, Marcel Rogge, Didier Stricker, and Varun Jampani. Somsi: Spherical novel view synthesis with soft occlusion multi-sphere images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15725-15734, 2022. 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 117 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 117 + ], + "type": "text", + "content": "[13] Yicong Hong, Kai Zhang, Juixiang Gu, Sai Bi, Yang Zhou, Difan Liu, Feng Liu, Kalyan Sunkavalli, Trung Bui, and Hao Tan. Lrm: Large reconstruction model for single image to 3d. arXiv preprint arXiv:2311.04400, 2023. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 119, + 553, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 119, + 553, + 152 + ], + "spans": [ + { + "bbox": [ + 317, + 119, + 553, + 152 + ], + "type": "text", + "content": "[14] Jie Hu, Li Shen, and Gang Sun. Squeeze-and-excitation networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7132-7141, 2018. 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 153, + 553, + 197 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 153, + 553, + 197 + ], + "spans": [ + { + "bbox": [ + 317, + 153, + 553, + 197 + ], + "type": "text", + "content": "[15] Huajian Huang, Yingshu Chen, Tianjian Zhang, and SaiKit Yeung. 360Roam: Real-Time Indoor Roaming Using Geometry-Aware " + }, + { + "bbox": [ + 317, + 153, + 553, + 197 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 317, + 153, + 553, + 197 + ], + "type": "text", + "content": " Radiance Fields. arXiv preprint arXiv:2208.02705, 2022. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 200, + 553, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 200, + 553, + 243 + ], + "spans": [ + { + "bbox": [ + 316, + 200, + 553, + 243 + ], + "type": "text", + "content": "[16] Hualie Jiang, Zhe Sheng, Siyu Zhu, Zilong Dong, and Rui Huang. Unifuse: Unidirectional fusion for 360 panorama depth estimation. IEEE Robotics and Automation Letters, 6 (2):1519-1526, 2021. 3, 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 246, + 553, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 246, + 553, + 289 + ], + "spans": [ + { + "bbox": [ + 316, + 246, + 553, + 289 + ], + "type": "text", + "content": "[17] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4):139-1, 2023. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 292, + 553, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 292, + 553, + 346 + ], + "spans": [ + { + "bbox": [ + 316, + 292, + 553, + 346 + ], + "type": "text", + "content": "[18] Shreyas Kulkarni, Peng Yin, and Sebastian Scherer. 360fusionnerf: Panoramic neural radiance fields with joint guidance. In 2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 7202-7209. IEEE, 2023. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 349, + 553, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 349, + 553, + 403 + ], + "spans": [ + { + "bbox": [ + 316, + 349, + 553, + 403 + ], + "type": "text", + "content": "[19] Hao Li, Yuanyuan Gao, Dingwen Zhang, Chenming Wu, Yalun Dai, Chen Zhao, Haocheng Feng, Errui Ding, Jingdong Wang, and Junwei Han. Ggrt: Towards generalizable 3d gaussians without pose priors in real-time. arXiv preprint arXiv:2403.10147, 2024. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 406, + 553, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 406, + 553, + 460 + ], + "spans": [ + { + "bbox": [ + 316, + 406, + 553, + 460 + ], + "type": "text", + "content": "[20] Jiahao Li, Hao Tan, Kai Zhang, Zexiang Xu, Fujun Luan, Yinghao Xu, Yicong Hong, Kalyan Sunkavalli, Greg Shakhnarovich, and Sai Bi. Instant3d: Fast text-to-3d with sparse-view generation and large reconstruction model. arXiv preprint arXiv:2311.06214, 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 462, + 553, + 517 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 462, + 553, + 517 + ], + "spans": [ + { + "bbox": [ + 316, + 462, + 553, + 517 + ], + "type": "text", + "content": "[21] Wenrui Li, Yapeng Mi, Fucheng Cai, Zhe Yang, Wangmeng Zuo, Xingtao Wang, and Xiaopeng Fan. Scenedreamer360: Text-driven 3d-consistent scene generation with panoramic gaussian splatting. arXiv preprint arXiv:2408.13711, 2024. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 520, + 553, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 520, + 553, + 564 + ], + "spans": [ + { + "bbox": [ + 316, + 520, + 553, + 564 + ], + "type": "text", + "content": "[22] Fangfu Liu, Wenqiang Sun, Hanyang Wang, Yikai Wang, Haowen Sun, Junliang Ye, Jun Zhang, and Yueqi Duan. Reconx: Reconstruct any scene from sparse views with video diffusion model. arXiv preprint arXiv:2408.16767, 2024. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 566, + 553, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 566, + 553, + 609 + ], + "spans": [ + { + "bbox": [ + 316, + 566, + 553, + 609 + ], + "type": "text", + "content": "[23] Xi Liu, Chaoyi Zhou, and Siyu Huang. 3dgssenhancer: Enhancing unbounded 3d gaussian splattering with view-consistent 2d diffusion priors. arXiv preprint arXiv:2410.16266, 2024. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 612, + 553, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 612, + 553, + 677 + ], + "spans": [ + { + "bbox": [ + 316, + 612, + 553, + 677 + ], + "type": "text", + "content": "[24] Yuan Liu, Sida Peng, Lingjie Liu, Qianqian Wang, Peng Wang, Christian Theobalt, Xiaowei Zhou, and Wenping Wang. Neural rays for occlusion-aware image-based rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7824-7833, 2022. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 679, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 679, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 679, + 553, + 713 + ], + "type": "text", + "content": "[25] Yikun Ma, Dandan Zhan, and Zhi Jin. Fastscene: Text-driven fast 3d indoor scene generation via panoramic gaussian splatt-ing. arXiv preprint arXiv:2405.05768, 2024. 3" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "21598" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 138 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 138 + ], + "type": "text", + "content": "[26] Santhosh K Ramakrishnan, Aaron Gokaslan, Erik Wijmans, Oleksandr Maksymets, Alex Clegg, John Turner, Eric Undersander, Wojciech Galuba, Andrew Westbury, Angel X Chang, et al. Habitat-matterport 3d dataset (hm3d): 1000 large-scale 3d environments for embodied ai. arXiv preprint arXiv:2109.08238, 2021. 1, 2, 5, 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 140, + 294, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 140, + 294, + 194 + ], + "spans": [ + { + "bbox": [ + 56, + 140, + 294, + 194 + ], + "type": "text", + "content": "[27] Julian Straub, Thomas Whelan, Lingni Ma, Yufan Chen, Erik Wijmans, Simon Green, Jakob J Engel, Raul Mur-Artal, Carl Ren, Shobhit Verma, et al. The replica dataset: A digital replica of indoor spaces. arXiv preprint arXiv:1906.05797, 2019. 1, 2, 5, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 195, + 294, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 195, + 294, + 249 + ], + "spans": [ + { + "bbox": [ + 56, + 195, + 294, + 249 + ], + "type": "text", + "content": "[28] Stanislaw Szymanowicz, Eldar Insafutdinov, Chuanxia Zheng, Dylan Campbell, João F Henriques, Christian Rupprecht, and Andrea Vedaldi. Flash3d: Feed-forward generalisable 3d scene reconstruction from a single image. arXiv preprint arXiv:2406.04343, 2024. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 251, + 294, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 251, + 294, + 303 + ], + "spans": [ + { + "bbox": [ + 56, + 251, + 294, + 303 + ], + "type": "text", + "content": "[29] Stanislaw Szymanowicz, Chrisitian Rupprecht, and Andrea Vedaldi. Splatter image: Ultra-fast single-view 3d reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10208-10217, 2024. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 304, + 294, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 304, + 294, + 348 + ], + "spans": [ + { + "bbox": [ + 56, + 304, + 294, + 348 + ], + "type": "text", + "content": "[30] Shengji Tang, Weicai Ye, Peng Ye, Weihao Lin, Yang Zhou, Tao Chen, and Wanli Ouyang. Hisplat: Hierarchical 3d gaussian splatting for generalizable sparse-view reconstruction. arXiv preprint arXiv:2410.06245, 2024. 2, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 350, + 294, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 350, + 294, + 392 + ], + "spans": [ + { + "bbox": [ + 56, + 350, + 294, + 392 + ], + "type": "text", + "content": "[31] Alex Trevithick and Bo Yang. Grf: Learning a general radiance field for 3d representation and rendering. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15182-15192, 2021. 1, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 394, + 294, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 394, + 294, + 448 + ], + "spans": [ + { + "bbox": [ + 56, + 394, + 294, + 448 + ], + "type": "text", + "content": "[32] Prune Truong, Marie-Julie Rakotosaona, Fabian Manhardt, and Federico Tombari. Sparf: Neural radiance fields from sparse and noisy poses. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4190-4200, 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 449, + 294, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 449, + 294, + 502 + ], + "spans": [ + { + "bbox": [ + 56, + 449, + 294, + 502 + ], + "type": "text", + "content": "[33] Fu-En Wang, Yu-Hsuan Yeh, Min Sun, Wei-Chen Chiu, and Yi-Hsuan Tsai. Bifuse: Monocular 360 depth estimation via bi-projection fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 462–471, 2020. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 504, + 294, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 504, + 294, + 557 + ], + "spans": [ + { + "bbox": [ + 56, + 504, + 294, + 557 + ], + "type": "text", + "content": "[34] Guangcong Wang, Peng Wang, Zhaoxi Chen, Wenping Wang, Chen Change Loy, and Ziwei Liu. Perf: Panoramic neural radiance field from a single panorama. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(10): 6905-6918, 2024. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 559, + 294, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 559, + 294, + 591 + ], + "spans": [ + { + "bbox": [ + 56, + 559, + 294, + 591 + ], + "type": "text", + "content": "[35] Peihao Wang, Xuxi Chen, Tianlong Chen, Subhashini Venugopalan, Zhangyang Wang, et al. Is attention all that nef needs? arXiv preprint arXiv:2207.13298, 2022. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 593, + 294, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 593, + 294, + 657 + ], + "spans": [ + { + "bbox": [ + 56, + 593, + 294, + 657 + ], + "type": "text", + "content": "[36] Qianqian Wang, Zhicheng Wang, Kyle Genova, Pratul P Srinivasan, Howard Zhou, Jonathan T Barron, Ricardo Martin-Brualla, Noah Snavely, and Thomas Funkhouser. Ibrnet: Learning multi-view image-based rendering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4690-4699, 2021. 1, 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 658, + 294, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 294, + 691 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 294, + 691 + ], + "type": "text", + "content": "[37] Qitai Wang, Lue Fan, Yuqi Wang, Yuntao Chen, and Zhaoxiang Zhang. Freevs: Generative view synthesis on free driving trajectory. arXiv preprint arXiv:2410.18079, 2024. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 692, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 692, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 692, + 294, + 713 + ], + "type": "text", + "content": "[38] Rundi Wu, Ben Mildenhall, Philipp Henzler, Keunhong Park, Ruiqi Gao, Daniel Watson, Pratul P Srinivasan, Dor" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 708 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 333, + 73, + 555, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 555, + 117 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 555, + 117 + ], + "type": "text", + "content": "Verbin, Jonathan T Barron, Ben Poole, et al. Reconfusion: 3d reconstruction with diffusion priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21551-21561, 2024. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 118, + 554, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 554, + 162 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 554, + 162 + ], + "type": "text", + "content": "[39] Haofei Xu, Jing Zhang, Jianfei Cai, Hamid Rezatofighi, Fisher Yu, Dacheng Tao, and Andreas Geiger. Unifying flow, stereo and depth estimation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 163, + 555, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 555, + 217 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 555, + 217 + ], + "type": "text", + "content": "[40] Haofei Xu, Anpei Chen, Yuedong Chen, Christos Sakaridis, Yulun Zhang, Marc Pollefeys, Andreas Geiger, and Fisher Yu. Murf: Multi-baseline radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20041-20050, 2024. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 219, + 554, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 219, + 554, + 262 + ], + "spans": [ + { + "bbox": [ + 316, + 219, + 554, + 262 + ], + "type": "text", + "content": "[41] Haofei Xu, Songyou Peng, Fangjinhua Wang, Hermann Blum, Daniel Barath, Andreas Geiger, and Marc Pollefeys. Depthsplat: Connecting gaussian splatting and depth. arXiv preprint arXiv:2410.13862, 2024. 2, 3, 4, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 263, + 554, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 263, + 554, + 296 + ], + "spans": [ + { + "bbox": [ + 316, + 263, + 554, + 296 + ], + "type": "text", + "content": "[42] Lihe Yang, Bingyi Kang, Zilong Huang, Zhen Zhao, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything v2. arXiv preprint arXiv:2406.09414, 2024. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 297, + 554, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 297, + 554, + 361 + ], + "spans": [ + { + "bbox": [ + 316, + 297, + 554, + 361 + ], + "type": "text", + "content": "[43] Weicai Ye, Chenhao Ji, Zheng Chen, Junyao Gao, Xiaoshui Huang, Song-Hai Zhang, Wanli Ouyang, Tong He, Cairong Zhao, and Guofeng Zhang. Diffpano: Scalable and consistent text to panorama generation with spherical epipolar-aware diffusion. arXiv preprint arXiv:2410.24203, 2024. 3, 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 364, + 554, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 364, + 554, + 407 + ], + "spans": [ + { + "bbox": [ + 316, + 364, + 554, + 407 + ], + "type": "text", + "content": "[44] Alex Yu, Vickie Ye, Matthew Tancik, and Angjoo Kanazawa. pixelnerf: Neural radiance fields from one or few images. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4578-4587, 2021. 1, 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 409, + 554, + 462 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 409, + 554, + 462 + ], + "spans": [ + { + "bbox": [ + 316, + 409, + 554, + 462 + ], + "type": "text", + "content": "[45] Wangbo Yu, Jinbo Xing, Li Yuan, Wenbo Hu, Xiaoyu Li, Zhipeng Huang, Xiangjun Gao, Tien-Tsin Wong, Ying Shan, and Yonghong Tian. Viewcrafter: Taming video diffusion models for high-fidelity novel view synthesis. arXiv preprint arXiv:2409.02048, 2024.3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 464, + 554, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 464, + 554, + 508 + ], + "spans": [ + { + "bbox": [ + 316, + 464, + 554, + 508 + ], + "type": "text", + "content": "[46] Cheng Zhang, Qianyi Wu, Camilo Cruz Gambardella, Xiaoshui Huang, Dinh Phung, Wanli Ouyang, and Jianfei Cai. Taming stable diffusion for text to 360 " + }, + { + "bbox": [ + 316, + 464, + 554, + 508 + ], + "type": "inline_equation", + "content": "\\{\\backslash \\mathrm{deg}\\}" + }, + { + "bbox": [ + 316, + 464, + 554, + 508 + ], + "type": "text", + "content": " panorama image generation. arXiv preprint arXiv:2404.07949, 2024. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 510, + 554, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 510, + 554, + 553 + ], + "spans": [ + { + "bbox": [ + 316, + 510, + 554, + 553 + ], + "type": "text", + "content": "[47] Cheng Zhang, Haofei Xu, Qianyi Wu, Camilo Cruz Gambardella, Dinh Phung, and Jianfei Cai. Pansplat: 4k panorama synthesis with feed-forward gaussian splatting. arXiv preprint arXiv:2412.12096, 2024. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 555, + 554, + 598 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 555, + 554, + 598 + ], + "spans": [ + { + "bbox": [ + 316, + 555, + 554, + 598 + ], + "type": "text", + "content": "[48] Kai Zhang, Sai Bi, Hao Tan, Yuanbo Xiangli, Nanxuan Zhao, Kalyan Sunkavalli, and Zexiang Xu. Gs-lrm: Large reconstruction model for 3d gaussian splatting. In European Conference on Computer Vision, pages 1-19. Springer, 2025. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 600, + 554, + 665 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 600, + 554, + 665 + ], + "spans": [ + { + "bbox": [ + 316, + 600, + 554, + 665 + ], + "type": "text", + "content": "[49] Shijie Zhou, Zhiwen Fan, Dejia Xu, Haoran Chang, Pradyumna Chari, Tejas Bharadwaj, Suya You, Zhangyang Wang, and Achuta Kadambi. Dreamscene360: Unconstrained text-to-3d scene generation with panoramic gaussian splatting. In European Conference on Computer Vision, pages 324–342. Springer, 2025. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 666, + 554, + 708 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 666, + 554, + 708 + ], + "spans": [ + { + "bbox": [ + 316, + 666, + 554, + 708 + ], + "type": "text", + "content": "[50] Tinghui Zhou, Richard Tucker, John Flynn, Graham Fyffe, and Noah Snavely. Stereo magnification: learning view synthesis using multiplane images. ACM Trans. Graph., 37(4): 65, 2018. 5" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "21599" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/SplineGS_ Robust Motion-Adaptive Spline for Real-Time Dynamic 3D Gaussians from Monocular Video/c20079e6-f30b-47d1-a98c-7f27a87d2799_content_list.json b/2025/SplineGS_ Robust Motion-Adaptive Spline for Real-Time Dynamic 3D Gaussians from Monocular Video/c20079e6-f30b-47d1-a98c-7f27a87d2799_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..c96696d9c6df51141c9955d664b337006bced31c --- /dev/null +++ b/2025/SplineGS_ Robust Motion-Adaptive Spline for Real-Time Dynamic 3D Gaussians from Monocular Video/c20079e6-f30b-47d1-a98c-7f27a87d2799_content_list.json @@ -0,0 +1,1845 @@ +[ + { + "type": "text", + "text": "SplineGS: Robust Motion-Adaptive Spline for Real-Time Dynamic 3D Gaussians from Monocular Video", + "text_level": 1, + "bbox": [ + 189, + 128, + 807, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jongmin Park1* Minh-Quan Viet Bui1* Juan Luis Gonzalez Bello1 Jaeho Moon1 \nJihyong Oh2† Munchurl Kim1†", + "bbox": [ + 165, + 202, + 828, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "\\(^{1}\\)Korea Advanced Institute of Science and Technology (KAIST) \\(^{2}\\)Chung-Ang University \\(\\{\\mathrm{jm.park, bvmquan, juanluisgb, jaehomoon, mkimee}\\} @kaist.ac.kr jihyongoh@cau.ac.kr https://kaist-viclab.github.io/splinegs-site/", + "bbox": [ + 132, + 255, + 864, + 309 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2d814176e456e030dc547810d5c45437fb60e893cbbfc71dc4c73b79c3961383.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 91, + 344, + 313, + 443 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b6a70072525f4339034e7b7eac3563aa9a2ea4bacb05fd1919f16215e7522bd3.jpg", + "image_caption": [ + "horsejump-low, $t = 20$ (Ours)", + "train, $t = 37$ (Ours)" + ], + "image_footnote": [ + "(a) Visual comparison for novel view synthesis on the DAVIS dataset" + ], + "bbox": [ + 91, + 455, + 313, + 554 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d9352c8f024540b3c8bf6e9266b5858c9053440c7132ec15b88a96de2b7587f1.jpg", + "image_caption": [ + "STGS [24]" + ], + "image_footnote": [], + "bbox": [ + 316, + 345, + 434, + 386 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/17e67f379b6de6572a4764d79a54bd0f29a90694ab22571c5dc19c1cf204e5c2.jpg", + "image_caption": [ + "D3DGS [55]" + ], + "image_footnote": [], + "bbox": [ + 434, + 345, + 553, + 386 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9dd5a5d3ec9936b0ecd139efda7215dd2aade2231be5f563f77f6cbb3e4d8850.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 316, + 402, + 433, + 443 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/54aa8c97d8f5c60eae6a136c42e4d286c5ddf29d725cb583ba9e6547e452f9bd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 434, + 402, + 552, + 443 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b3fd31d14da4013979908b2ad4f7ef4da7c19e4b5fa7739d9a8030a35b30c842.jpg", + "image_caption": [ + "RoDynRF [30]", + "STGS [24]" + ], + "image_footnote": [], + "bbox": [ + 315, + 455, + 433, + 494 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8f48f9a6140fcffe04e1e9a96352139ff158accf5eed32af915b36f17b9e3506.jpg", + "image_caption": [ + "SplineGS (Ours)", + "D3DGS [55]" + ], + "image_footnote": [], + "bbox": [ + 434, + 455, + 552, + 494 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/fa809b097842eada0d1a61b00139df7b8e5367c2fa50bfc82606e5e685f5a7af.jpg", + "image_caption": [ + "RoDynRF [30]" + ], + "image_footnote": [], + "bbox": [ + 316, + 512, + 433, + 554 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7e260132ffbf6ac6a9c233c420f0aa50b81213865d280c623e8a7939c24a9302.jpg", + "image_caption": [ + "SplineGS (Ours)" + ], + "image_footnote": [], + "bbox": [ + 434, + 512, + 553, + 554 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0ff377d5310212799fa39bf4e7c5a8018446762d1f41155fc3086c69d69c5d05.jpg", + "image_caption": [ + "Figure 1. Our SplineGS achieves state-of-the-art rendering quality with fast rendering speed for novel spatio-temporal view synthesis from monocular videos without relying on pre-computed camera parameters. (a) We use our predicted camera parameters for [24, 55] since COLMAP [42] is unable to provide reasonable camera parameters for most scenes in the DAVIS dataset [39]. (b) SplineGS achieves the best rendering quality while maintaining a rendering speed of 400 FPS on the NVIDIA dataset [56]." + ], + "image_footnote": [ + "(b) Performance gain on the NVIDIA dataset" + ], + "bbox": [ + 560, + 345, + 897, + 564 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 656, + 326, + 672 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Synthesizing novel views from in-the-wild monocular videos is challenging due to scene dynamics and the lack of multi-view cues. To address this, we propose SplineGS, a COLMAP-free dynamic 3D Gaussian Splatting (3DGS) framework for high-quality reconstruction and fast rendering from monocular videos. At its core is a novel Motion-Adaptive Spline (MAS) method, which represents continuous dynamic 3D Gaussian trajectories using cubic Hermite splines with a small number of control points. For MAS, we introduce a Motion-Adaptive Control points Pruning (MACP) method to model the deformation of each dynamic 3D Gaussian across varying motions, progressively pruning control points while maintaining dynamic modeling integrity. Additionally, we present a joint optimization strat", + "bbox": [ + 86, + 688, + 485, + 901 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "egy for camera parameter estimation and 3D Gaussian attributes, leveraging photometric and geometric consistency. This eliminates the need for Structure-from-Motion preprocessing and enhances SplineGS's robustness in real-world conditions. Experiments show that SplineGS significantly outperforms state-of-the-art methods in novel view synthesis quality for dynamic scenes from monocular videos, achieving thousands times faster rendering speed.", + "bbox": [ + 511, + 657, + 906, + 780 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 791, + 645, + 806 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Novel View Synthesis (NVS) is fundamental to 3D vision, supporting applications like virtual reality (VR), augmented reality (AR), and film production. NFS aims to generate", + "bbox": [ + 511, + 816, + 906, + 864 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Co-first authors (equal contribution).", + "bbox": [ + 529, + 875, + 733, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Co-corresponding authors.", + "bbox": [ + 532, + 887, + 679, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "26866", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "images from any viewpoint in a scene, requiring accurate reconstruction from multiple 2D images. NeRF [34] has advanced the field of NVs by utilizing learned implicit functions to represent static scenes, though it requires considerable training and rendering time. Recently, 3D Gaussian Splatting (3DGS) [20] has revolutionized this process by replacing implicit volumetric rendering with differentiable rasterization of 3D Gaussians, enabling real-time rendering and providing a more explicit scene representation.", + "bbox": [ + 89, + 90, + 480, + 227 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "For NVs of dynamic scenes, prior works have extended the static scene representations [20, 34] by incorporating deformation models in canonical space using implicit representations [12, 30, 35, 36, 55], grid-based models that decompose the 4D space-time domain into multiple 2D planes [4, 6, 10, 43, 52], and polynomial trajectories [24]. However, these methods face several challenges: implicit representations significantly increase computational overhead and reduce rendering speed [6, 52]; grid-based models struggle to fully capture the dynamic nature of scene structures, hindering their ability to model fine details accurately [24]; and although polynomial trajectories improve efficiency, their fixed degree restricts flexibility for representing complex motions. To the best of our knowledge, none of the dynamic 3DGS-based methods provide experimental evidence to reliably support their novel spatio-temporal view synthesis capabilities for rendering unseen intermediate time indices. Additionally, most existing methods for dynamic NVs rely heavily on external camera parameter estimators like COLMAP [42], which often produce imprecise results for challenging in-the-wild monocular videos.", + "bbox": [ + 91, + 231, + 483, + 547 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address these issues for modeling scene dynamics, we exploit a spline-based model to dynamic 3D Gaussian trajectories, inspired by classic 3D-curve modeling in computer graphics [9]. Splines are widely used in geometric modeling for graphical applications, providing smooth and continuous representations of complex shapes with a minimal number of control points [2, 8]. This efficiency makes them ideal for modeling intricate geometric structures while maintaining both flexibility and precision. In this paper, we propose SplineGS, a framework for high-quality dynamic scene reconstruction and real-time neural rendering from inthe-wild monocular videos without relying on external camera estimators like COLMAP [42]. SplineGS introduces Motion-Adaptive Spline (MAS), based on cubic Hermite splines [2, 8], to effectively represent dynamic 3D Gaussian deformations. Our MAS consists of piecewise cubic functions defined by control points that dictate each segment's curvature and direction. Each control point is adjustable and optimized as a learnable parameter, enabling faster and more precise modeling of complex motion trajectories. Additionally, to adaptively model the trajectory of each dynamic 3D Gaussian based on motion complexity during training, we introduce a Motion-Adaptive Control", + "bbox": [ + 89, + 553, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "points Pruning (MACP) method that adjusts the number of control points to improve rendering quality and efficiency. Furthermore, we incorporate a camera parameter estimation method jointly optimized with 3D Gaussian attributes and MAS, leveraging photometric and geometric consistency, thus eliminating the need for external estimators. Experiments show that SplineGS significantly outperforms state-of-the-art (SOTA) dynamic NVs methods both qualitatively and quantitatively, offering faster rendering speed, as shown in Fig. 1. Our contributions are as follows:", + "bbox": [ + 511, + 90, + 903, + 242 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a novel Spline-based dynamic 3D Gaussian Splatting framework, SplineGS, which is (i) COLMAP-free, (ii) very fast and (iii) of high quality in reconstructing the dynamic scenes from in-the-wild monocular videos;", + "- A novel Motion-Adaptive Spline (MAS) is introduced, which can accurately and effectively represent the continuous trajectory of each dynamic 3D Gaussian;", + "- A Motion-Adaptive Control points Pruning (MACP) method is presented, which can efficiently adjust the number of control points for each spline function, optimizing both rendering quality and the efficiency of MAS." + ], + "bbox": [ + 511, + 246, + 903, + 420 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 435, + 653, + 450 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Dynamic NeRF. Recent advancements in video view synthesis have extended the static NeRF model [34] to represent scene dynamics. These include dynamic NeRFs using scene flow-based frameworks [12, 25, 26], deformation estimation in canonical fields [1, 3, 16, 30, 35, 36, 40, 44, 47, 51, 53], and 4D grid-based spatio-temporal radiance fields [4, 6, 10, 43]. Techniques such as NSFF [25], DynNeRF [12], and DynIBaR [26] combine time-independent and time-dependent radiance fields to synthesize novel spatiotemporal perspectives from monocular videos. Despite these advancements, current dynamic NeRFs still fall short of recent dynamic 3D Gaussian Splatting (3DGS) techniques in rendering quality and efficiency.", + "bbox": [ + 511, + 460, + 906, + 657 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Dynamic 3DGS. The improvements in rendering quality and speed achieved by 3DGS [20] have inspired further studies [14, 24, 27, 31, 52, 55] to extend the static 3DGS framework to dynamic scenes by enabling the deformation of 3D Gaussian attributes. The pioneering work on dynamic 3DGS [31] introduces time-dependent offsets for the positions and rotations of dynamic 3D Gaussians via an MLP; however, this approach slows down the rendering. D3DGS [55] builds on this concept with an annealing smoothing training mechanism to improve temporal smoothness and rendering quality. 4DGS [52] replaces the MLP network with a grid-based structure to boost efficiency, though this change requires quality trade-offs due to the resolution limits of grid-based methods. In contrast, SC-GS [14] combines an MLP deformation network with sparse spatial deformation, reducing the computational", + "bbox": [ + 511, + 659, + 908, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "26867", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "cost of MLP while maintaining quality. STGS [24] introduces polynomial trajectories for motion modeling, improving speed and quality over implicit representations. CasualFVS [22] warps dynamic content from neighboring frames, and MoSca [23] proposes a graph-based motion modeling approach to handle sparse control deformation. Very recently, a concurrent work [21] proposes modeling dynamic 3D Gaussian trajectories using polynomial interpolation between fixed keyframes for NVS from multi-view videos with COLMAP [42] assistance. Unlike [21], our SplineGS adaptively optimizes the deformation of dynamic 3D Gaussians, accounting for varying motion degrees and types in in-the-wild monocular videos, without requiring preprocessed camera parameters (COLMAP-free approach). Neural Rendering without SfM Preprocessing. Accurate camera parameters, including extrinsic and intrinsic, are essential for neural rendering approaches to capture fine details [17, 28]. However, in real-world settings, camera parameters derived from Structure-from-Motion (SfM) algorithms such as COLMAP [42] often exhibit pixel-level inaccuracies [29, 41], compromising the structural details of rendered scenes [30]. To address this, several NeRF methods [28, 32, 37, 50] jointly optimize NeRF architectures and camera parameters. Recently, a local-to-global training approach [11] is introduced to optimize both camera parameters and 3D Gaussians. However, these methods are limited to static scenes. For dynamic scene reconstruction, RoDynRF [30] and MoSca [23] use motion masks to gather multi-view cues from static regions, allowing robust rendering without pre-computed camera parameters. Our SplineGS is also COLMAP-free, significantly outperforming RoDynRF [30] and MoSca [23] in rendering quality and efficiency, enabled by our novel spline-based architecture.", + "bbox": [ + 89, + 90, + 483, + 590 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Preliminary: 3D Gaussian Splitting", + "text_level": 1, + "bbox": [ + 89, + 601, + 416, + 619 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3DGS [20] represents the radiance field of a scene using anisotropic 3D Gaussians, each of which is formulated as", + "bbox": [ + 89, + 626, + 482, + 656 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nG (\\boldsymbol {x}) = e ^ {- \\frac {1}{2} (\\boldsymbol {x} - \\boldsymbol {\\mu}) ^ {\\top} \\boldsymbol {\\Sigma} ^ {- 1} (\\boldsymbol {x} - \\boldsymbol {\\mu})} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 184, + 659, + 480, + 679 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\pmb{x} \\in \\mathbb{R}^3$ denotes a 3D position, and $\\pmb{\\mu} \\in \\mathbb{R}^3$ and $\\pmb{\\Sigma} \\in \\mathbb{R}^{3 \\times 3}$ represent the mean (center) and the covariance matrix of the 3D Gaussian, respectively. To ensure that $\\pmb{\\Sigma}$ is positive semi-definite and contains physical meaning, it is decomposed into a diagonal scaling matrix $S \\in \\mathbb{R}^{3 \\times 3}$ of a scale vector $s \\in \\mathbb{R}^3$ and a rotation matrix $R \\in \\mathbb{R}^{3 \\times 3}$ as $\\pmb{\\Sigma} = R\\pmb{S}\\pmb{S}^\\top\\pmb{R}^\\top$ , where $\\pmb{R}$ is parameterized by a learnable unit-length quaternion $q \\in \\mathbb{R}^4$ . In addition, each 3D Gaussian is parameterized by an opacity $\\sigma \\in \\mathbb{R}$ and a color $c \\in \\mathbb{R}^3$ . To render the color of each pixel, the color and the opacity of each 3D Gaussian are computed using Eq. 1, and the rendered color $C$ is computed by the alpha-blending of the $\\mathcal{N}$ ordered 3D Gaussians overlapping the pixel as", + "bbox": [ + 89, + 683, + 482, + 880 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {C} = \\sum_ {i \\in \\mathcal {N}} \\boldsymbol {c} _ {i} \\alpha_ {i} \\prod_ {j = 1} ^ {i - 1} (1 - \\alpha_ {j}), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 174, + 883, + 482, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $c_{i}$ is the color of the $i^{\\mathrm{th}}$ 3D Gaussian and $\\alpha_{i}$ is a density of the $i^{\\mathrm{th}}$ 3D Gaussian which is given by evaluating the 2D covariance $\\pmb{\\Sigma}' \\in \\mathbb{R}^{2 \\times 2}$ . Here, $\\pmb{\\Sigma}'$ is formulated as $\\pmb{\\Sigma}' = J\\pmb{W}\\pmb{\\Sigma}\\pmb{W}^\\top J^\\top$ , where $J$ is the Jacobian of the affine approximation of the projective transformation and $W$ is a viewing transformation matrix. Similar to prior works [23, 27], we extend 3DGS [20] to a union of static 3D Gaussians $\\{G_i^{\\mathrm{st}} | i = 1, 2, \\dots, n^{\\mathrm{st}}\\}$ and dynamic 3D Gaussians $\\{G_i^{\\mathrm{dy}} | i = 1, 2, \\dots, n^{\\mathrm{dy}}\\}$ to represent static backgrounds and moving objects, respectively, in dynamic scenes.", + "bbox": [ + 511, + 90, + 906, + 244 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Proposed Method: SplineGS", + "text_level": 1, + "bbox": [ + 511, + 262, + 777, + 280 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1. Overview of SplineGS", + "text_level": 1, + "bbox": [ + 511, + 287, + 720, + 305 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We describe the overall architecture of SplineGS in Fig. 2. Given a monocular video $\\{I_t|t = 0,1,\\dots,N_f - 1\\}$ where $N_{f}$ is the total number of frames, SplineGS is designed to synthesize high-quality novel spatio-temporal views with fast rendering speed, and to estimate the camera parameters, including the extrinsic $[\\hat{R}_t|\\hat{T}_t]\\in \\mathbb{R}^{3\\times 4}$ for each time $t$ , and the shared intrinsic $\\hat{\\pmb{K}}\\in \\mathbb{R}^{3\\times 3}$ . As shown in Fig. 2, to stabilize joint optimization of the 3D Gaussian attributes and camera parameter estimation, we adopt a two-stage optimization process consisting of a warm-up stage and a main training stage for our SplineGS architecture. In the warm-up stage, we coarsely optimize the camera parameters $[\\hat{R}_t|\\hat{T}_t]$ and $\\hat{\\pmb{K}}$ using photometric and geometric consistency. In the main training stage, we initialize the 3D Gaussians based on the estimated camera poses and jointly optimize the 3D Gaussian attributes with the camera parameters. Specifically, for each dynamic 3D Gaussian, we propose a novel spline-based deformation modeling method, called Motion-Adaptive Spline (MAS), which utilizes a cubic Hermite spline function [2, 8] to accurately model the continuous trajectory with a small number of control points. For MAS, we introduce a Motion-Adaptive Control points Pruning (MACP) method that effectively adjusts the number of control points for each dynamic 3D Gaussian, considering the object's motion types and degrees.", + "bbox": [ + 511, + 313, + 906, + 691 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the following sections, we detail the process of our MAS method in Sec. 4.2, followed by the camera parameter estimation process in Sec. 4.3. Finally, we describe the overall optimization process in Sec. 4.4.", + "bbox": [ + 511, + 694, + 905, + 753 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.2. Motion-Adaptive Spline for 3D Gaussians", + "text_level": 1, + "bbox": [ + 511, + 768, + 870, + 784 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To represent the continuous trajectory of each dynamic 3D Gaussian for moving objects over time, we propose MAS which modifies the mean parameter to a set of learnable control points. For this, we use the cubic Hermite spline function [2, 8] to model the mean of each dynamic 3D Gaussian at time $t$ as", + "bbox": [ + 511, + 792, + 905, + 882 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\mu} (t) = S (t, \\mathbf {P}), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 653, + 885, + 906, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "26868", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/a6d6d9f6497404325d227490123d5c8dce1c9fbb7b4e83b61409e0591cf28809.jpg", + "image_caption": [ + "Figure 2. Overview of SplineGS. Our SplineGS leverages spline-based functions to model the deformation of dynamic 3D Gaussians with a novel Motion-Adaptive Spline (MAS) architecture. It is composed of sets of learnable control points based on a cubic Hermite spline function [2, 8] to accurately model the trajectory of each dynamic 3D Gaussian and to achieve faster rendering speed. To avoid any preprocessing of camera parameters, i.e., COLMAP-free, we adopt a two-stage optimization: warm-up and main training stages." + ], + "image_footnote": [], + "bbox": [ + 89, + 85, + 908, + 404 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{P} = \\{\\mathbf{p}_k|\\mathbf{p}_k\\in \\mathbb{R}^3,k\\in 0,\\dots ,N_c - 1\\}$ is a set of $N_{c}$ learnable control points, serving as an additional attribute for each dynamic 3D Gaussian, and $S(t,\\mathbf{P})$ is defined as", + "bbox": [ + 89, + 473, + 482, + 520 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} S (t, \\mathbf {P}) = \\left(2 t _ {r} ^ {3} - 3 t _ {r} ^ {2} + 1\\right) \\mathbf {p} _ {\\lfloor t _ {s} \\rfloor} + \\left(t _ {r} ^ {3} - 2 t _ {r} ^ {2} + t _ {r}\\right) \\mathbf {m} _ {\\lfloor t _ {s} \\rfloor} \\\\ + \\left(- 2 t _ {r} ^ {3} + 3 t _ {r} ^ {2}\\right) \\mathbf {p} _ {[ t _ {s} ] + 1} + \\left(t _ {r} ^ {3} - t _ {r} ^ {2}\\right) \\mathbf {m} _ {[ t _ {s} ] + 1}, \\tag {4} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 104, + 521, + 480, + 560 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nt _ {r} = t _ {s} - \\lfloor t _ {s} \\rfloor , \\quad t _ {s} = t _ {n} (N _ {c} - 1), \\quad t _ {n} = t / (N _ {f} - 1),\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 563, + 454, + 579 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{m}_k$ is the approximated tangent of the control point $\\mathbf{p}_k$ , formulated as $\\mathbf{m}_k = (\\mathbf{p}_{k + 1} - \\mathbf{p}_{k - 1}) / 2$ . Note that $S(t,\\mathbf{P})$ indicates the piecewise cubic function to represent the whole continuous trajectory of each dynamic 3D Gaussian, and $N_{c}$ can be different from $N_{f}$ . The optimal $N_{c}$ can be estimated by Motion-Adaptive Control points Pruning (MACP) in the following section.", + "bbox": [ + 89, + 580, + 482, + 686 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Control Points Initialization. To stably optimize $S(t, \\mathbf{P})$ , it is essential to initialize the control points with appropriate geometric considerations that ensure temporal consistency. For this, we leverage the long-range 2D tracking [19] and the metric depth [38] priors. Let $\\mathcal{T} = \\{\\varphi_t^{\\mathrm{tr}} | \\varphi_t^{\\mathrm{tr}} \\in \\mathbb{R}^2, t = 0, \\dots, N_f - 1\\}$ be a 2D track, $\\pi_K(\\cdot)$ be a projection function from the camera space to the image space with the camera intrinsic $K$ , and $\\varphi_t^{\\mathrm{tr}}$ be a pixel coordinate corresponding to the 2D track $\\mathcal{T}$ at time $t$ . We unproject each 2D track $\\mathcal{T}$ to the world space to compute the 3D track aided by the frame's metric depth $d_t$ and camera extrinsic $[\\hat{R}_t | \\hat{T}_t]$ at time $t$ . The unprojection function $\\mathcal{W}_t(\\cdot)$ from the image space to the world space is formulated as", + "bbox": [ + 89, + 686, + 482, + 882 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {W} _ {t} \\left(\\boldsymbol {\\varphi} _ {t} ^ {\\mathrm {t r}}\\right) = \\hat {\\boldsymbol {R}} _ {t} ^ {\\top} \\pi_ {\\hat {\\boldsymbol {K}}} ^ {- 1} \\left(\\boldsymbol {\\varphi} _ {t} ^ {\\mathrm {t r}}, \\boldsymbol {d} _ {t} \\left(\\boldsymbol {\\varphi} _ {t} ^ {\\mathrm {t r}}\\right)\\right) - \\hat {\\boldsymbol {R}} _ {t} ^ {\\top} \\hat {\\boldsymbol {T}} _ {t}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 140, + 883, + 482, + 906 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "It should be noted that we estimate $[\\hat{\\pmb{R}}_t|\\hat{\\pmb{T}}_t]$ and $\\hat{\\pmb{K}}$ from a sequence of frames only, without using any given ground truth values, as mentioned in Sec. 4.3. Then, we initialize the per-Gaussian control points set $\\mathbf{P}$ , by a least-square (LS) approximation, such that the spline-described curve fits the initial curve by the tracker, given by", + "bbox": [ + 511, + 473, + 906, + 565 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\mathbf {P}} \\sum_ {t = 0} ^ {N _ {f} - 1} \\left\\| \\mathcal {W} _ {t} \\left(\\varphi_ {t} ^ {\\mathrm {t r}}\\right) - S (t, \\mathbf {P}) \\right\\| _ {2} ^ {2}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 569, + 906, + 594 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Motion-Adaptive Control Points Pruning (MACP). An excessive number of control points may cause over-fitting and decrease the processing speed of our MAS module, leading to poorer rendering qualities with reduced rendering speed (see Table 3). Furthermore, as each scene exhibits various types and degrees of motion for moving objects, the number of control points required for each dynamic 3D Gaussian trajectory should be adaptively adjusted to accommodate the scene dynamics. To achieve this, we propose the MACP method for MAS, which can generate sparser control points while ensuring no dynamic modeling degradation. MACP is designed on top of 3D Gaussian densification [20], but focuses on optimizing the number of control points in MAS. Our MACP computes a new spline function $S(t,\\mathbf{P}^{\\prime})$ after every 3D Gaussian densification step, where $\\mathbf{P}^{\\prime} = \\{\\mathbf{p}_{l}^{\\prime}|\\mathbf{p}_{l}^{\\prime}\\in \\mathbb{R}^{3}, l = 0,\\dots ,N_{c} - 2\\}$ is a set of $N_{c} - 1$ control points, which contains one-fewer control points than the current set $\\mathbf{P}$ of control points. We compute the LS approximation to find the reduced set $\\mathbf{P}^{\\prime}$ of control points as", + "bbox": [ + 511, + 598, + 908, + 886 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "26869", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\mathbf {P} ^ {\\prime}} \\sum_ {t = 0} ^ {N _ {f} - 1} \\left\\| S (t, \\mathbf {P}) - S (t, \\mathbf {P} ^ {\\prime}) \\right\\| _ {2} ^ {2}. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 166, + 87, + 482, + 113 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Then, the updated optimal set $\\mathbf{P}$ of control points is assigned with the values of $\\mathbf{P}'$ if the error $E$ between $S(t,\\mathbf{P})$ and $S(t,\\mathbf{P}')$ is smaller than a threshold $\\epsilon$ , as given by", + "bbox": [ + 89, + 116, + 482, + 162 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {P} = \\left\\{ \\begin{array}{l l} \\mathbf {P} ^ {\\prime}, & \\text {i f} E < \\epsilon \\\\ \\mathbf {P}, & \\text {o t h e r w i s e} \\end{array} , \\right. \\text {w h e r e} \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 179, + 167, + 482, + 208 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nE = \\frac {1}{N _ {f}} \\sum_ {t = 0} ^ {N _ {f} - 1} \\| \\pi_ {\\hat {\\mathbf {K}}} (\\hat {\\mathbf {R}} _ {t} S (t, \\mathbf {P}) + \\hat {\\mathbf {T}} _ {t}) - \\pi_ {\\hat {\\mathbf {K}}} (\\hat {\\mathbf {R}} _ {t} S (t, \\mathbf {P} ^ {\\prime}) + \\hat {\\mathbf {T}} _ {t}) \\| _ {2} ^ {2}.\n$$\n", + "text_format": "latex", + "bbox": [ + 91, + 203, + 452, + 239 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "By following MACP, each dynamic 3D Gaussian is allowed to have a different number of control points. Therefore, the MACP can guide our MAS to have a minimal number of control points when modeling simple motion, while having more control points for more complex motions (see Fig. 8).", + "bbox": [ + 89, + 250, + 483, + 325 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Camera Parameter Estimation", + "text_level": 1, + "bbox": [ + 89, + 335, + 362, + 349 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The traditional SfM methods, such as COLMAP [42], fail to reliably estimate camera parameters in dynamic scenes from in-the-wild monocular videos [30]. For this reason, we propose to estimate camera parameters for joint optimization with the 3D Gaussian attributes. For each frame at time $t$ , we predict a rotation matrix $\\hat{R}_t \\in \\mathbb{R}^{3 \\times 3}$ and a translation vector $\\hat{T}_t \\in \\mathbb{R}^3$ , representing the extrinsic parameters of a monocular camera, using a shallow MLP $F_{\\theta}$ as", + "bbox": [ + 89, + 359, + 483, + 479 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left[ \\boldsymbol {\\hat {R}} _ {t} \\mid \\boldsymbol {\\hat {T}} _ {t} \\right] = F _ {\\theta} (\\gamma (t)), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 217, + 484, + 480, + 502 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\gamma (\\cdot)$ is a positional encoding [34]. We also predict the focal length $\\hat{f}$ of the camera intrinsic $\\hat{K}$ as a learnable parameter that is shared across all frames in the monocular video. To accurately optimize the camera parameters, we enforce two types of consistency—photometric and geometric—for the static background between random reference cameras $[\\hat{R}_{t_{\\mathrm{ref}}}|\\hat{T}_{t_{\\mathrm{ref}}}]$ and the target camera $[\\hat{R}_t|\\hat{T}_t]$ , encouraging alignment both visually and structurally.", + "bbox": [ + 89, + 507, + 483, + 630 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Photometric Consistency. Given the pre-computed metric depth [38], the camera intrinsics and extrinsics under optimization will converge as long as the projected reference frame's color $I_{t_{\\mathrm{ref}}}(\\varphi_{t \\to t_{\\mathrm{ref}}})$ is aligned to the target frame's color $I_t(\\varphi_t)$ . $\\varphi_{t \\to t_{\\mathrm{ref}}}$ is the reference pixel coordinate corresponding to the target frame's pixel coordinate $\\varphi_t$ as", + "bbox": [ + 89, + 630, + 483, + 720 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\varphi_ {t \\rightarrow t _ {\\text {r e f}}} = \\pi_ {\\hat {\\mathbf {K}}} \\left(\\hat {\\mathbf {R}} _ {t _ {\\text {r e f}}} \\left(\\hat {\\mathbf {R}} _ {t} ^ {\\top} \\pi_ {\\hat {\\mathbf {K}}} ^ {- 1} \\left(\\varphi_ {t}, \\boldsymbol {d} _ {t} (\\varphi_ {t})\\right) - \\hat {\\mathbf {R}} _ {t} ^ {\\top} \\hat {\\boldsymbol {T}} _ {t}\\right) + \\hat {\\boldsymbol {T}} _ {t _ {\\text {r e f}}}\\right). \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 101, + 726, + 482, + 744 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We refer to such projection alignment as photometric consistency, which is encouraged by the loss $\\mathcal{L}_{\\mathrm{pc}}$ given by", + "bbox": [ + 89, + 750, + 482, + 779 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {p c}} = \\sum_ {\\boldsymbol {\\varphi} _ {t}} \\left\\| \\boldsymbol {M} _ {t, t _ {\\text {r e f}}} \\left(\\boldsymbol {\\varphi} _ {t}\\right) \\odot \\left(\\boldsymbol {I} _ {t} \\left(\\boldsymbol {\\varphi} _ {t}\\right) - \\boldsymbol {I} _ {t _ {\\text {r e f}}} \\left(\\boldsymbol {\\varphi} _ {t \\rightarrow t _ {\\text {r e f}}}\\right)\\right)\\right\\| _ {2} ^ {2}, \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 107, + 784, + 482, + 805 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\odot$ is the Hadamard product [13], and $M_{t,t_{\\mathrm{ref}}}$ is a union motion mask that excludes dynamic objects in $I_{t}$ and $I_{t_{\\mathrm{ref}}}$ for removing the inconsistencies due to dynamic regions, which is given by", + "bbox": [ + 89, + 815, + 483, + 876 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {M} _ {t, t _ {\\text {r e f}}} \\left(\\varphi_ {t}\\right) = \\boldsymbol {M} _ {t} \\left(\\varphi_ {t}\\right) \\boldsymbol {M} _ {t _ {\\text {r e f}}} \\left(\\varphi_ {t \\rightarrow t _ {\\text {r e f}}}\\right), \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 160, + 883, + 482, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $M_{t}$ and $M_{t_{\\mathrm{ref}}}$ are pre-computed motion masks [54] of $\\pmb{I}_t$ and $I_{t_{\\mathrm{ref}}}$ , respectively. Note that motion mask's value is 0 for static regions, and 1 otherwise.", + "bbox": [ + 511, + 90, + 903, + 135 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Geometric Consistency. Along with the photometric consistency, we compute the geometric consistency of unprojected pixels in 3D space to make our optimization more stable. The geometric consistency loss is formulated as", + "bbox": [ + 511, + 136, + 903, + 196 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {g c}} = \\sum_ {\\boldsymbol {\\varphi} _ {t}} \\| M _ {t, t _ {\\text {r e f}}} (\\boldsymbol {\\varphi} _ {t}) \\odot \\left(\\mathcal {W} _ {t} (\\boldsymbol {\\varphi} _ {t}) - \\mathcal {W} _ {t _ {\\text {r e f}}} \\left(\\boldsymbol {\\varphi} _ {t \\rightarrow t _ {\\text {r e f}}}\\right)\\right) \\| _ {2} ^ {2}, \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 521, + 200, + 903, + 220 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{W}_t(\\varphi_t)$ and $\\mathcal{W}_{t_{\\mathrm{ref}}}(\\varphi_{t\\to t_{\\mathrm{ref}}})$ are the corresponding 3D locations of $\\varphi_t$ and $\\varphi_{t\\rightarrow t_{\\mathrm{ref}}}$ , respectively (see Eq. 5).", + "bbox": [ + 511, + 223, + 903, + 253 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.4. Optimization", + "text_level": 1, + "bbox": [ + 511, + 261, + 653, + 277 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To stabilize joint training of the MAS and the camera parameter estimation, we adopt a two-stage optimization process consisting of the warm-up stage and the main training stage. In the warm-up stage, we optimize only the camera parameters using the photometric and geometric consistency. The total loss in the warm-up stage is given by", + "bbox": [ + 511, + 284, + 905, + 375 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {t o t a l}} ^ {\\text {w a r m}} = \\lambda_ {\\mathrm {p c}} \\mathcal {L} _ {\\mathrm {p c}} + \\lambda_ {\\mathrm {g c}} \\mathcal {L} _ {\\mathrm {g c}}. \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 620, + 381, + 903, + 398 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After the warm-up stage, we obtain the coarsely predicted camera intrinsic $\\hat{\\pmb{K}}$ and the set of extrinsics $[\\hat{\\pmb{R}}|\\hat{\\pmb{T}}]$ for all frames, which are then used to initialize the set $\\mathbf{P}$ of control points for each dynamic 3D Gaussian, as described in Sec. 4.2. In the main training stage, we jointly optimize the static and dynamic 3D Gaussians along with the camera parameter estimation based on the total loss function as", + "bbox": [ + 511, + 405, + 905, + 510 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {\\text {t o t a l}} ^ {\\text {m a i n}} = \\lambda_ {\\mathrm {r g b}} \\mathcal {L} _ {\\mathrm {r g b}} + \\lambda_ {\\mathrm {d}} \\mathcal {L} _ {\\mathrm {d}} + \\lambda_ {\\mathrm {M}} \\mathcal {L} _ {\\mathrm {M}} \\\\ + \\lambda_ {\\mathrm {p c}} \\mathcal {L} _ {\\mathrm {p c}} + \\lambda_ {\\mathrm {d - p c}} \\mathcal {L} _ {\\mathrm {d - p c}} + \\lambda_ {\\mathrm {g c}} \\mathcal {L} _ {\\mathrm {g c}}, \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 581, + 516, + 836, + 553 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "(15)", + "bbox": [ + 874, + 527, + 903, + 541 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_{\\mathrm{rgb}}$ and $\\mathcal{L}_{\\mathrm{d}}$ are the L1 losses between the rendered frame and GT frame, and between the rendered depth and GT depth, respectively. In the main training stage, we compute an additional photometric consistency loss $\\mathcal{L}_{\\mathrm{d - pc}}$ that utilizes the rendered depth $\\hat{\\pmb{d}}$ of the 3D Gaussians instead of the metric depth [38] prior $\\pmb{d}$ as $\\hat{\\pmb{d}}$ allows the estimated 3D Gaussian geometry to guide the joint optimization of the camera parameters and 3D Gaussian attributes.", + "bbox": [ + 511, + 558, + 905, + 679 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In addition to the camera parameter estimation and imagery reconstruction losses, we adopt a binary dice loss [45] $\\mathcal{L}_{\\mathrm{M}}$ between the pre-computed motion mask $M_{t}$ [54] and the rendered motion mask $\\hat{M}_t$ that can be derived from the dynamic 3D Gaussians. The binary dice loss initially proposed in [45] for highly imbalanced segmentation of medical imagery helps encouraging better separation between our dynamic and static 3D Gaussians as described by", + "bbox": [ + 511, + 681, + 905, + 801 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {M}} = 1 - \\frac {2 \\sum_ {\\varphi_ {t}} M _ {t} \\left(\\varphi_ {t}\\right) \\hat {M} _ {t} \\left(\\varphi_ {t}\\right) + \\varepsilon}{\\sum_ {\\varphi_ {t}} \\left(M _ {t} \\left(\\varphi_ {t}\\right) + \\hat {M} _ {t} \\left(\\varphi_ {t}\\right)\\right) + \\varepsilon}, \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 545, + 806, + 903, + 852 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\varepsilon$ is a smooth term to avoid numerical issues. $\\hat{M}_t(\\varphi_t)$ is computed by the alpha-blending of the 3D Gaussians overlapping $\\varphi_t$ (similar to Eq. 2) as", + "bbox": [ + 511, + 854, + 903, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "26870", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/27e9de0ea01ad7c7617e03d4515dcf7ce2e20a8e1184c8ac2db648f58748d2da.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsJumpingSkatingTruckUmbrellaBalloon1Balloon2PlaygroundAverageFPS↑Tr. (h)↓
STM preprocessingDynNeRF (ICCV'21) [12]24.68 / 0.09032.66 / 0.03528.56 / 0.08223.26 / 0.13722.36 / 0.10427.06 / 0.04924.15 / 0.08026.10 / 0.0820.05>24
MonoNeRF (ICCV'23) [46]24.26 / 0.09132.06 / 0.04427.56 / 0.11523.62 / 0.18021.89 / 0.12927.36 / 0.05222.61 / 0.13025.62 / 0.1060.05N/A
STGS (CVPR'24) [24]20.82 / 0.18724.80 / 0.10925.01 / 0.10321.88 / 0.19520.36 / 0.19623.12 / 0.12419.23 / 0.15122.17 / 0.1529000.2
SCGS (CVPR'24) [14]15.68 / 0.92014.88 / 0.90823.81 / 0.14021.84 / 0.16020.17 / 0.17921.07 / 0.14920.71 / 0.11519.74 / 0.3671101.1
D3DGS (CVPR'24) [55]22.02 / 0.26624.06 / 0.22723.04 / 0.24722.67 / 0.19221.22 / 0.20225.86 / 0.11822.30 / 0.11123.02 / 0.195251.0
4DGS (CVPR'24) [52]22.37 / 0.17826.72 / 0.08425.93 / 0.09722.36 / 0.17821.89 / 0.15324.85 / 0.08121.36 / 0.08923.64 / 0.123950.25
SP-GS (ICML'24) [48]22.13 / 0.46829.21 / 0.23627.38 / 0.19024.88 / 0.32324.36 / 0.18029.65 / 0.09722.29 / 0.23425.70 / 0.2471806.0
Casual-FVS (ECCV'24) [22]23.45 / 0.10029.98 / 0.04525.22 / 0.09023.24 / 0.09623.76 / 0.07924.15 / 0.08122.19 / 0.07424.57 / 0.081480.25
Kinematic Fields (ECCV'24) [15]23.74 / 0.10231.89 / 0.03528.34 / 0.07425.46 / 0.09823.72 / 0.07926.49 / 0.05524.64 / 0.05326.33 / 0.0710.37.0
E-D3DGS (ECCV'24) [5]22.39 / 0.18024.62 / 0.13726.85 / 0.08123.60 / 0.14520.87 / 0.16124.39 / 0.07320.79 / 0.12423.36 / 0.129452.6
CTNeRF (Pattern Recognition'24) [33]24.35 / 0.09433.51 / 0.03428.27 / 0.08423.48 / 0.12922.19 / 0.11126.86 / 0.04824.28 / 0.07726.13 / 0.0820.1>24
D-NPC (Eurographics'25) [18]24.51 / 0.11630.22 / 0.06128.92 / 0.08424.15 / 0.13622.26 / 0.16125.84 / 0.10723.59 / 0.09925.64 / 0.109700.27
Shape-of-Motion (arXiv'24) [49]22.74 / 0.12829.12 / 0.04824.63 / 0.12623.79 / 0.08922.74 / 0.12025.83 / 0.07220.61 / 0.13024.21 / 0.1022551.1
Ex4DGS (NeurIPS'24) [21]18.93 / 0.32121.92 / 0.23319.04 / 0.30819.03 / 0.34014.69 / 0.50316.29 / 0.45714.16 / 0.43717.72 / 0.371840.65
DynPoint (NeurIPS'23) [58]24.69 / 0.09731.34 / 0.04529.30 / 0.06124.59 / 0.08622.77 / 0.09927.63 / 0.04925.37 / 0.03926.53 / 0.0680.80.5
RoDynRF (CVPR'23) [30] w/25.66 / 0.07128.68 / 0.04029.13 / 0.06324.26 / 0.08922.37 / 0.10326.19 / 0.05424.96 / 0.04825.89 / 0.0670.45>24
MoSca (arXiv'24) [23] w/25.21 / 0.08332.77 / 0.03328.22 / 0.09024.41 / 0.09223.26 / 0.09228.90 / 0.04223.05 / 0.06026.55 / 0.070400.8
SplineGS (Ours) w/25.62 / 0.06434.10 / 0.03427.83 / 0.04824.60 / 0.07622.67 / 0.12029.04 / 0.03024.47 / 0.04626.90 / 0.0604000.5
STM FreeRoDynRF (CVPR'23) [30]24.27 / 0.10028.71 / 0.04628.85 / 0.06623.25 / 0.10421.81 / 0.12225.58 / 0.06425.20 / 0.05225.38 / 0.0790.45>24
MoSca (arXiv'24) [23]25.43 / 0.08032.62 / 0.03328.29 / 0.08624.40 / 0.09123.27 / 0.09129.01 / 0.04223.23 / 0.05826.61 / 0.069401.0
SplineGS (Ours)25.50 / 0.06833.72 / 0.03128.66 / 0.05625.61 / 0.07124.43 / 0.06828.37 / 0.03224.19 / 0.04727.21 / 0.0534001.5
", + "bbox": [ + 94, + 89, + 911, + 335 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Novel view synthesis evaluation on the NVIDIA dataset (PSNR↑ / LPIPS↓). Red and Blue denote the best and second-best performances, respectively. For Casual-FVS [22], we directly use the results from their paper, as the official code is unavailable. 'w/' denotes the use of camera poses obtained by COLMAP [42] for each SfM-free method. 'Tr' denotes the training time. For MonoNeRF [46], the training time is omitted as it is not a per-scene optimized model. We use an RTX 3090 Ti GPU for each method.", + "bbox": [ + 89, + 340, + 906, + 397 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/3b07c8fae3cafce0df17a4fade182ac546c315331fc6996fa4469cefdc9215cf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 93, + 412, + 228, + 546 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/38617b43d9f060febfdb6fa5867c65025a47bdfb2c0ff7de68b745b51f5672de.jpg", + "image_caption": [ + "Figure 3. Visual comparisons for novel view synthesis on the NVIDIA dataset." + ], + "image_footnote": [], + "bbox": [ + 228, + 412, + 364, + 545 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/6ba8827c50382e4b70c4d1fdd83446629df95241eb1b996c99a610fa6cc9bc7a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 364, + 412, + 500, + 545 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/16689226469fb74279850fc2dccd0a6dd21d2163f52f5250576e68bc67c877da.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 412, + 633, + 545 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e14840e2fa731fc70224196f9043d0bea774865c2273b7735c1c49219e5b5699.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 633, + 412, + 767, + 544 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/0bf5619da990411a9a9b77e7ce10578fb9d15d42e695b295083021deb49e1bf8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 767, + 412, + 905, + 544 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {M} _ {t} (\\boldsymbol {\\varphi} _ {t}) = \\sum_ {i \\in \\mathcal {N}} m _ {i} \\alpha_ {i} \\prod_ {j = 1} ^ {i - 1} (1 - \\alpha_ {j}), \\tag {17}\n$$\n", + "text_format": "latex", + "bbox": [ + 151, + 583, + 482, + 603 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $m_{i} = 0$ if the $i^{\\mathrm{th}}$ 3D Gaussian is the static 3D Gaussian and $m_{i} = 1$ otherwise.", + "bbox": [ + 89, + 606, + 482, + 636 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In conjunction, all terms in $\\mathcal{L}_{\\mathrm{total}}^{\\mathrm{main}}$ guide our SplineGS to effectively and efficiently model dynamic 3D scenes from pure monocular videos, achieving more structural details and better temporal consistency than previous works [12, 14, 21-24, 30, 46, 52, 55], without relying on camera parameters obtained from external estimators [42].", + "bbox": [ + 89, + 637, + 483, + 728 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 89, + 739, + 223, + 757 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details. We refer readers to the Supplementary material for the details of our implementation.", + "bbox": [ + 89, + 765, + 482, + 795 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. We evaluate both the quantitative and qualitative performance of novel view and time synthesis on the widely used NVIDIA dataset [56] which features challenging monocular videos. Additionally, we assess novel view synthesis performance on in-the-wild monocular videos from the DAVIS dataset [39] which contains an average of 69.7 frames per video sequence.", + "bbox": [ + 89, + 796, + 483, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Comparison with State-of-the-Art Methods", + "text_level": 1, + "bbox": [ + 511, + 569, + 880, + 585 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Novel View Synthesis. Table 1 presents a quantitative comparison of NVs between our SplineGS and existing SfM-based [5, 12, 14, 15, 18, 21, 22, 24, 33, 46, 48, 49, 52, 55, 58] and SfM-free [23, 30] methods on the NVIDIA dataset [56]. For this comparison, we follow the evaluation configuration in [30]. The results demonstrate that our SplineGS significantly outperforms SOTA methods in both the PSNR and LPIPS [57] metrics. Notably, SplineGS achieves $890 \\times$ and $8,000 \\times$ faster rendering speed compared to RoDynRF [30] and DynNeRF [12], respectively. Ex4DGS [21] and STGS [24], which are designed for multi-view settings, face challenges with inconsistent geometry alignment over time when trained on monocular videos. Fig. 3 shows the qualitative comparison between our SplineGS with the existing methods in [12, 24, 30, 55]. As highlighted by the red boxes, our method yields not only higher rendering quality but also dynamic objects more aligned and closer to the ground truth. In Fig. 4, we show our superior NVs results for in-the-wild monocular videos from the DAVIS dataset [39] compared to the existing methods in [24, 30, 55].", + "bbox": [ + 511, + 598, + 906, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "26871", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ef1c7b7d4cfa7a812fe38d3b2ec250d706f34704c0ec3604d2071aa27dd325dd.jpg", + "image_caption": [ + "Figure 4. Visual comparisons for novel view synthesis on the DAVIS dataset." + ], + "image_footnote": [], + "bbox": [ + 93, + 88, + 287, + 284 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d8e053f2754a6f9dc97640ef988db60489fa8cb7466334a822cc06fe7d895420.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 285, + 88, + 482, + 284 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Compared to [30] that is also COLMAP-free, SplineGS yields considerably more detailed novel views (red boxes) as shown in Fig. 4. For the other methods [24, 55], we observe that COLMAP [42] fails to recover camera parameters and initial point clouds on the DAVIS dataset [39], as also claimed in [30]. On the other hand, our COLMAP-free SplineGS reconstructs accurate camera parameters that are the ones actually used to train [24, 55], which are shown in Fig. 4 for comparison. More results are in the Supplementary material.", + "bbox": [ + 88, + 339, + 482, + 491 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Novel View and Time Synthesis. To evaluate the capability of SplineGS to model continuous trajectories of moving objects in a scene, we compare the novel view and time synthesis results of SplineGS with those of NeRF-based [12, 30] and 3DGS-based [24, 52, 55] methods. For this evaluation, we follow the dataset sampling strategy in [25], which samples 24 timestamps from the NVIDIA dataset [56]. In addition, to simulate a larger motion, we exclude frames with odd time indices in the training sets. To ensure all test timestamps are not seen during training, and thus, to create a more challenging novel view and time synthesis validation, we exclude frames with even time indices in the test sets. Table 2 and Fig. 5 show the quantitative and qualitative comparisons for this challenging experiment. We observe that the NeRF-based methods, RoDynRF [30] and DynNeRF [12], generate inconsistent artifacts and blurriness for unseen times. Furthermore, the 3DGS-based methods [24, 52, 55] yield even more significant degradation when predicting unseen time indices. In contrast, SplineGS, built upon 3DGS [20] and equipped with our novel spline-based deformation, provides SOTA novel view rendering at unseen intermediate timestamps. Thanks to our MAS, SplineGS naturally and precisely captures the continuous trajectories of moving objects over time, enhancing the temporal consistency of rendered scenes that can be checked in tOF scores [7] of Table 2.", + "bbox": [ + 91, + 493, + 482, + 898 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/9fb0400f985f0cbc98130b6744709d2c7551767e3c75827f7d583fda1f7c40ab.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsPSNR↑LPIPS↓tOF↓
SfM preprocessingDynNeRF (ICCV'21) [12]23.360.2190.921
4DGS (CVPR'24) [52]17.070.4596.314
D3DGS (CVPR'24) [55]19.630.3433.225
STGS (CVPR'24) [24]15.720.4742.105
SfM-FreeRoDynRF (CVPR'23) [30]21.580.2212.138
SplineGS (Ours)25.920.0980.703
", + "bbox": [ + 517, + 88, + 898, + 179 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Novel view and time synthesis evaluation on the NVIDIA dataset.", + "bbox": [ + 513, + 183, + 903, + 210 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d5104621fc11f3df0a7975772d82ae75ab57e7fa3b8cebc5800f69a13c77acfc.jpg", + "image_caption": [ + "Figure 5. Visual comparisons for novel view and time synthesis on the NVIDIA dataset." + ], + "image_footnote": [], + "bbox": [ + 514, + 224, + 619, + 321 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ac5c8eb84a529459c8cb420935f2c1478dab1eb4920cdbd17b3c25d108accdcf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 620, + 224, + 702, + 321 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5a130a8da916272055ea4e6bf9a03ca29408682b2a171bb3df995cabc0538a57.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 704, + 224, + 789, + 321 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1d91ca2faf8697bb02bd4a12b80da090c4edff8049dd9529384b2bb290d4e787.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 790, + 224, + 844, + 321 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a7a933d4bd704964aecbad34995d477aa98a24f5d2c7fb44b2ce353e04afd924.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 846, + 224, + 901, + 321 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To further analyze the SplineGS's ability to model continuous trajectories of dynamic 3D Gaussians, we visualize the projected 2D motion tracking of dynamic objects in pixel space, comparing it with D3DGS [55] and STGS [24], as shown in Fig. 6. We observe that D3DGS [55] and STGS [24] cannot provide reliable motion tracking for moving objects, underscoring their limitations in modeling continuous trajectories of dynamic 3D Gaussians. In contrast, SplineGS provides accurate motion tracking, demonstrating the effectiveness of MAS for deforming dynamic 3D Gaussians. Further analysis is in the Supplementary material.", + "bbox": [ + 511, + 363, + 906, + 529 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9401d0320810a7442011e77ff7e49989879b328c1ece6439499d28b7b43aa4b0.jpg", + "image_caption": [ + "Figure 6. Visual comparisons for motion tracking. We visualize 2D pixel tracks to analyze motions of dynamic 3D Gaussians." + ], + "image_footnote": [], + "bbox": [ + 516, + 532, + 645, + 623 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/140c61669249bd5e0f14e8957127219e4dd2c1c41eb0934d5aac4584855a0708.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 645, + 532, + 774, + 625 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7a466f6f93ee90b8328922bb4b7ed4dfbc2d5cc42d26580ed99fe67c8c639ede.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 774, + 532, + 903, + 625 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Ablation Study", + "text_level": 1, + "bbox": [ + 511, + 665, + 666, + 681 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Motion-Adaptive Spline (MAS). To demonstrate the effectiveness of MAS, we replace the MAS model with various deformation models, including an MLP, a grid-based model, polynomial functions of third degree (denoted as 'Poly $(3^{\\mathrm{rd}})$ ') and tenth degree (denoted as 'Poly $(10^{\\mathrm{th}})$ '), and a Bezier curve [9]. For the MLP, grid-based model, and polynomial functions, we apply to them the structures similar to those in prior works, including D3DGS [55], 4DGS [52], and STGS [24], respectively. Additionally, we implement the Bezier curve [9], a commonly used method for curve modeling in computer graphics. Table 3-(a) presents quantitative comparisons of each 3D Gaussian trajectory model, focusing on rendering quality (PSNR, LPIPS) and deformation latency per Gaussian, denoted as $g_{\\mathrm{def}}$ . This latency re", + "bbox": [ + 511, + 688, + 906, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "26872", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/dae64395a9f2a2f04571df543f94a4015fef14c8bba9911c8c798a497b5e2a13.jpg", + "image_caption": [ + "Figure 7. Visual comparisons for MACP ablation study." + ], + "image_footnote": [], + "bbox": [ + 93, + 87, + 482, + 188 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "flects the computational time required to estimate the deformation of a single dynamic 3D Gaussian. As shown in Table 3-(a), our MAS model achieves superior rendering quality compared to all other deformation models. Consistent with analyses in previous works [24, 52, 55], we observe that the MLP and grid-based architectures require substantial computational costs for rendering. Among these methods, 'Poly $(3^{\\mathrm{rd}})$ ', as implemented in [24], demonstrates the best latency. However, fixed-degree polynomial functions have limited flexibility across varying motion complexities which adversely impacts rendering performance. To explore this further, we experiment with 'Poly $(10^{\\mathrm{th}})$ ' to assess changes in modeling capability. This adjustment, however, leads to noisier optimization and reduced efficiency, as variables under high exponents in 'Poly $(10^{\\mathrm{th}})$ ' lead to numerical instability. The Bézier curve [9] offers the second-best rendering quality, but its latency remains higher than our MAS due to its recursive nature of computation.", + "bbox": [ + 89, + 220, + 483, + 491 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Motion-Adaptive Control Points Pruning (MACP). To assess the effectiveness of our MACP technique for MAS, we compare our full model with MACP against other versions of our model with two fixed numbers of control points $N_{c} = 4$ and $N_{c} = N_{f}$ . As shown in Table 3-(c) and Fig. 7, our SplineGS with MACP achieves a good trade-off between rendering quality and $g_{\\mathrm{def}}$ compared to the ablated models with fixed $N_{c}$ . Using $N_{c} = 4$ for every dynamic 3D Gaussian limits the motion modeling capacity of MAS, resulting in significantly lower metrics and visible artifacts in the dynamic regions. Moreover, an excessive $N_{c} = N_{f}$ decreases the rendering speed of our MAS module and still falls short of the quality achieved by our full model with MACP, potentially due to motion overfitting. Fig. 8 shows the distribution of $N_{c}$ values after optimization with MACP across scenes of varying motion complexities. In Fig. 8-(a), we visualize ' $N_{c}$ Heatmap' that contains the pixel-wise averaged $N_{c}$ values of the dynamic 3D Gaussians needed to render the 2D pixels (red higher, blue lower number of control points). As shown, the simpler motions in these scenes, such as those of the human bodies, can be modeled by the dynamic 3D Gaussians with smaller averages of $N_{c}$ values, whereas the objects with complex and extensive motions (e.g., balloon) require higher averaged $N_{c}$ values. Fig. 8-(b) presents the corresponding histogram of $N_{c}$ values for the scenes' dynamic 3D Gaussians. For sequences with simple motion, such as 'Skating', the trajectories of most", + "bbox": [ + 89, + 493, + 483, + 901 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/fa54ab5df1327c9b623b6dfcb7124c3c38b50a47e5fab954d4b70e1e5cb0240e.jpg", + "table_caption": [ + "(a) Motion-Adaptive Spline" + ], + "table_footnote": [], + "table_body": "
PSNR↑LPIPS↓gDef (ns)↓
MLP23.510.125149.41
Grid25.480.09098.89
Poly (3rd)25.140.1111.80
Poly (10th)24.380.1207.71
Bézier27.190.0608.78
Ours27.210.0535.63
", + "bbox": [ + 517, + 104, + 710, + 183 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/42a70a77ee571120e41223d0fa28c060c5c6c39d8319ff099c991d72fd53782e.jpg", + "table_caption": [ + "(b) Loss function" + ], + "table_footnote": [], + "table_body": "
PSNR↑LPIPS↓
w/o Lpc17.490.853
w/o Lgc26.330.067
w/o Ld-pc26.180.066
w/o Lm26.340.088
Ours27.210.053
", + "bbox": [ + 730, + 104, + 885, + 183 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/ac47030047729578435ea17158f8379253577063c2c8778a803c66146486abf8.jpg", + "table_caption": [ + "(c) Motion-Adaptive Control points Pruning" + ], + "table_footnote": [], + "table_body": "
PSNR↑LPIPS↓gdef (ns)↓
w/o MACP (Nc=4)26.620.0655.34
w/o MACP (Nc=Nf)27.080.0546.11
Ours27.210.0535.63
", + "bbox": [ + 575, + 202, + 844, + 260 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3. Ablation studies. We ablate our framework and report the average results on the NVIDIA dataset with the same setting as Novel View Synthesis experiment in Sec. 5.1.", + "bbox": [ + 511, + 263, + 906, + 306 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ac44e70d252a6783e3f0d3fcd541de3c038b9583bf1107457910d3964c2cfe6a.jpg", + "image_caption": [ + "Figure 8. Analysis of MACP's Efficacy. (a) $N_{c}$ Heatmaps as the averaged $N_{c}$ values of dynamic 3D Gaussians and their corresponding rendered frames $\\hat{I}_{t}$ for 'Balloon2' and 'Skating' scenes. (b) Histograms of the number of control points $(N_{c})$ in percentages (\\%) of dynamic 3D Gaussians in two scenes." + ], + "image_footnote": [], + "bbox": [ + 522, + 320, + 712, + 493 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/90c6520ad34e6f0d3370f97583f2b095dffd15915aacf4e9f5e80889ae5f0261.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 715, + 320, + 901, + 493 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "dynamic 3D Gaussians can be represented using a minimal $N_{c}$ , thanks to our MACP. While 'Balloon2' has more evenly distributed $N_{c}$ due to more complex and diverse motion.", + "bbox": [ + 511, + 580, + 905, + 625 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Loss Functions. Table 3-(b) shows the effectiveness of each loss for our overall SplineGS architecture. As noted, no consistent camera parameters can be learned without $\\mathcal{L}_{\\mathrm{pc}}$ , drastically impacting the rendering quality of the dynamic 3D Gaussians. Moreover, our $\\mathcal{L}_{\\mathrm{gc}}$ , $\\mathcal{L}_{\\mathrm{d - pc}}$ and $\\mathcal{L}_{\\mathrm{M}}$ can considerably impact the overall rendering quality.", + "bbox": [ + 511, + 627, + 906, + 719 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 737, + 633, + 752 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We present SplineGS, a COLMAP-free dynamic 3DGS framework designed for novel spatio-temporal view synthesis from monocular videos. Leveraging our innovative Motion-Adaptive Spline (MAS) for dynamic motion modeling, SplineGS efficiently renders high-quality novel views from complex in-the-wild videos. The effectiveness of our approach is validated through extensive quantitative and qualitative comparisons, significantly outperforming the existing SOTA methods with very fast rendering speed.", + "bbox": [ + 511, + 763, + 906, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "26873", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 91, + 90, + 250, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported by Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korean Government [Ministry of Science and ICT (Information and Communications Technology)] (Project Number: RS-2022-00144444, Project Title: Deep Learning Based Visual Representational Learning and Rendering of Static and Dynamic Scenes, 100%).", + "bbox": [ + 89, + 113, + 483, + 212 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 238, + 187, + 255 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Fast dynamic radiance fields with time-aware neural voxels. In SIGGRAPH Asia 2022 Conference Papers, 2022. 2", + "[2] J Harold Ahlberg, Edwin Norman Nilson, and Joseph Leonard Walsh. The Theory of Splines and Their Applications: Mathematics in Science and Engineering: A Series of Monographs and Textbooks, Vol. 38. Elsevier, 2016. 2, 3, 4", + "[3] ShahRukh Athar, Zexiang Xu, Kalyan Sunkavalli, Eli Shechtman, and Zhixin Shu. Rignerf: Fully controllable neural 3d portraits. In CVPR, 2022. 2", + "[4] Benjamin Attal, Jia-Bin Huang, Christian Richardt, Michael Zollhoefer, Johannes Kopf, Matthew O'Toole, and Changil Kim. Hyperreel: High-fidelity 6-dof video with ray-conditioned sampling. In CVPR, 2023. 2", + "[5] Jeongmin Bae, Seoha Kim, Youngsik Yun, Hahyun Lee, Gun Bang, and Youngjung Uh. Per-gaussian embedding-based deformation for deformable 3d gaussian splatting. In European Conference on Computer Vision, pages 321-335. Springer, 2024. 6", + "[6] Ang Cao and Justin Johnson. Hexplane: A fast representation for dynamic scenes. In CVPR, 2023. 2", + "[7] Mengyu Chu, You Xie, Jonas Mayer, Laura Leal-Taixe, and Nils Thuerey. Learning temporal coherence via self-supervision for gan-based video generation. ACM Transactions on Graphics (TOG), 2020. 7", + "[8] C De Boor. A practical guide to splines. Springer-Verlag google schola, 1978. 2, 3, 4", + "[9] Gerald Farin. Curves and surfaces for CAGD: a practical guide. Elsevier, 2001. 2, 7, 8", + "[10] Sara Fridovich-Keil, Giacomo Meanti, Frederik Rahbæk Warburg, Benjamin Recht, and Angjoo Kanazawa. K-planes: Explicit radiance fields in space, time, and appearance. In CVPR, 2023. 2", + "[11] Yang Fu, Sifei Liu, Amey Kulkarni, Jan Kautz, Alexei A. Efros, and Xiaolong Wang. Colmap-free 3d gaussian splatting. In CVPR, 2024. 3", + "[12] Chen Gao, Ayush Saraf, Johannes Kopf, and Jia-Bin Huang. Dynamic view synthesis from dynamic monocular video. In ICCV, 2021. 2, 6, 7", + "[13] Roger A Horn and Charles R Johnson. Matrix analysis. Cambridge university press, 2012. 5", + "[14] Yi-Hua Huang, Yang-Tian Sun, Ziyi Yang, Xiaoyang Lyu, Yan-Pei Cao, and Xiaojuan Qi. Sc-gs: Sparse-controlled gaussian splatting for editable dynamic scenes. CVPR, 2024. 2, 6" + ], + "bbox": [ + 93, + 263, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] Woobin Im, Geonho Cha, Sebin Lee, Jumin Lee, Juhyeong Seon, Dongyoon Wee, and Sung-Eui Yoon. Regularizing dynamic radiance fields with kinematic fields. In European Conference on Computer Vision, pages 312-328. Springer, 2024. 6", + "[16] Wei Jiang, Kwang Moo Yi, Golnoosh Samei, Oncel Tuzel, and Anurag Ranjan. Neuman: Neural human radiance field from a single video. In ECCV, 2022. 2", + "[17] Yifan Jiang, Peter Hedman, Ben Mildenhall, Dejia Xu, Jonathan T Barron, Zhangyang Wang, and Tianfan Xue. Alignerf: High-fidelity neural radiance fields via alignment-aware training. In CVPR, 2023. 3", + "[18] Moritz Kappel, Florian Hahlbohm, Timon Scholz, Susana Castillo, Christian Theobalt, Martin Eisemann, Vladislav Golyanik, and Marcus Magnor. D-npc: Dynamic neural point clouds for non-rigid view synthesis from monocular video. arXiv preprint arXiv:2406.10078, 2024. 6", + "[19] Nikita Karaev, Ignacio Rocco, Benjamin Graham, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. Co-tracker: It is better to track together. arXiv, 2023. 4", + "[20] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 2023. 2, 3, 4, 7", + "[21] Junoh Lee, Chang-Yeon Won, Hyunjun Jung, Inhwan Bae, and Hae-Gon Jeon. Fully explicit dynamic gaussian splatting. In NeurIPS, 2024. 3, 6", + "[22] Yao-Chih Lee, Zhoutong Zhang, Kevin Blackburn-Matzen, Simon Niklaus, Jianming Zhang, Jia-Bin Huang, and Feng Liu. Fast view synthesis of casual videos with soup-of-planes. In ECCV, 2024. 3, 6", + "[23] Jiahui Lei, Yijia Weng, Adam Harley, Leonidas Guibas, and Kostas Daniilidis. Mosca: Dynamic gaussian fusion from casual videos via 4d motion scaffolds. arXiv, 2024. 3, 6", + "[24] Zhan Li, Zhang Chen, Zhong Li, and Yi Xu. Spacetime gaussian feature splatting for real-time dynamic view synthesis. In CVPR. 1, 2, 3, 6, 7, 8", + "[25] Zhengqi Li, Simon Niklaus, Noah Snavely, and Oliver Wang. Neural scene flow fields for space-time view synthesis of dynamic scenes. In CVPR, 2021. 2, 7", + "[26] Zhengqi Li, Qianqian Wang, Forrester Cole, Richard Tucker, and Noah Snavely. Dynibar: Neural dynamic image-based rendering. In CVPR, 2023. 2", + "[27] Yiqing Liang, Numair Khan, Zhengqin Li, Thu Nguyen-Phuoc, Douglas Lanman, James Tompkin, and Lei Xiao. Gaufre: Gaussian deformation fields for real-time dynamic novel view synthesis. arXiv, 2023. 2, 3", + "[28] Chen-Hsuan Lin, Wei-Chiu Ma, Antonio Torralba, and Simon Lucey. Barf: Bundle-adjusting neural radiance fields. In ICCV, 2021. 3", + "[29] Philipp Lindenberger, Paul-Edouard Sarlin, Viktor Larsson, and Marc Pollefeys. Pixel-perfect structure-from-motion with featuremetric refinement. In ICCV, 2021. 3", + "[30] Yu-Lun Liu, Chen Gao, Andreas Meuleman, Hung-Yu Tseng, Ayush Saraf, Changil Kim, Yung-Yu Chuang, Johannes Kopf, and Jia-Bin Huang. Robust dynamic radiance fields. In CVPR, 2023. 2, 3, 5, 6, 7" + ], + "bbox": [ + 516, + 92, + 905, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "26874", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[31] Jonathon Luiten, Georgios Kopanas, Bastian Leibe, and Deva Ramanan. Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis. In 3DV, 2024. 2", + "[32] Quan Meng, Anpei Chen, Haimin Luo, Minye Wu, Hao Su, Lan Xu, Xuming He, and Jingyi Yu. Gnerf: Gan-based neural radiance field without posed camera. In ICCV, 2021. 3", + "[33] Xingyu Miao, Yang Bai, Haoran Duan, Fan Wan, Yawen Huang, Yang Long, and Yefeng Zheng. Ctnerf: Cross-time transformer for dynamic neural radiance field from monocular video. Pattern Recognition, 156:110729, 2024. 6", + "[34] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 2, 5", + "[35] Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In ICCV, 2021. 2", + "[36] Keunhong Park, Utkarsh Sinha, Peter Hedman, Jonathan T. Barron, Sofien Bouaziz, Dan B Goldman, Ricardo Martin-Brualla, and Steven M. Seitz. Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields. ACM Trans. Graph., 2021. 2", + "[37] Keunhong Park, Philipp Henzler, Ben Mildenhall, Jonathan T. Barron, and Ricardo Martin-Brualla. Camp: Camera preconditioning for neural radiance fields. ACM Trans. Graph., 2023. 3", + "[38] Luigi Piccinelli, Yung-Hsu Yang, Christos Sakaridis, Mattia Segu, Siyuan Li, Luc Van Gool, and Fisher Yu. Unidepth: Universal monocular metric depth estimation. In CVPR, 2024. 4, 5", + "[39] Jordi Pont-Tuset, Federico Peruzzi, Sergi Caelles, Pablo Arbeláez, Alex Sorkine-Hornung, and Luc Van Gool. The 2017 davis challenge on video object segmentation. arXiv, 2018. 1, 6, 7", + "[40] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-nerf: Neural radiance fields for dynamic scenes. In CVPR, 2021. 2", + "[41] Vincent Raoult, Sarah Reid-Anderson, Andreas Ferri, and Jane E Williamson. How reliable is structure from motion (sfm) over time and between observers? a case study using coral reef bommie. Remote Sensing, 2017. 3", + "[42] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In CVPR, 2016. 1, 2, 3, 5, 6, 7", + "[43] Ruizhi Shao, Zerong Zheng, Hanzhang Tu, Boning Liu, Hongwen Zhang, and Yebin Liu. Tensor4d: Efficient neural 4d decomposition for high-fidelity dynamic reconstruction and rendering. In CVPR, 2023. 2", + "[44] Liangchen Song, Anpei Chen, Zhong Li, Zhang Chen, Lele Chen, Junsong Yuan, Yi Xu, and Andreas Geiger. Nerfplayer: A streamable dynamic scene representation with decomposed neural radiance fields. IEEE Transactions on Visualization and Computer Graphics, 2023. 2", + "[45] Carole H. Sudre, Wenqi Li, Tom Vercauteren, Sebastien Ourselin, and M. Jorge Cardoso. Generalised dice overlap as a deep learning loss function for highly unbalanced segmentations. In Deep Learning in Medical Image Analysis and Multimodal Learning for Clinical Decision Support, 2017. 5" + ], + "bbox": [ + 91, + 92, + 480, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[46] Fengrui Tian, Shaoyi Du, and Yueqi Duan. MonoN-eRF: Learning a generalizable dynamic radiance field from monocular videos. In ICCV, 2023. 6", + "[47] Edgar Tretschk, Ayush Tewari, Vladislav Golyanik, Michael Zollhöfer, Christoph Lassner, and Christian Theobalt. Non-rigid neural radiance fields: Reconstruction and novel view synthesis of a dynamic scene from monocular video. In ICCV, 2021. 2", + "[48] Diwen Wan, Ruijie Lu, and Gang Zeng. Superpoint gaussian splatting for real-time high-fidelity dynamic scene reconstruction. arXiv preprint arXiv:2406.03697, 2024. 6", + "[49] Qianqian Wang, Vickie Ye, Hang Gao, Jake Austin, Zhengqi Li, and Angjoo Kanazawa. Shape of motion: 4d reconstruction from a single video. arXiv preprint arXiv:2407.13764, 2024. 6", + "[50] Zirui Wang, Shangzhe Wu, Weidi Xie, Min Chen, and Victor Adrian Prisacariu. Nerf: Neural radiance fields without known camera parameters. CoRR, 2021. 3", + "[51] Chung-Yi Weng, Brian Curless, Pratul P. Srinivasan, Jonathan T. Barron, and Ira Kemelmacher-Shlizerman. HumanNeRF: Free-viewpoint rendering of moving people from monocular video. In CVPR, 2022. 2", + "[52] Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Xinggang Wang. 4d gaussian splatting for real-time dynamic scene rendering. In CVPR. 2, 6, 7, 8", + "[53] Gengshan Yang, Minh Vo, Neverova Natalia, Deva Ramanan, Vedaldi Andrea, and Joo Hanbyul. Banmo: Building animatable 3d neural models from many casual videos. In CVPR, 2022. 2", + "[54] Jinyu Yang, Mingqi Gao, Zhe Li, Shang Gao, Fangjing Wang, and Feng Zheng. Track anything: Segment anything meets videos. arXiv, 2023. 5", + "[55] Ziyi Yang, Xinyu Gao, Wen Zhou, Shaohui Jiao, Yuqing Zhang, and Xiaogang Jin. Deformable 3d gaussians for high-fidelity monocular dynamic scene reconstruction. In CVPR, 2024. 1, 2, 6, 7, 8", + "[56] Jae Shin Yoon, Kihwan Kim, Orazio Gallo, Hyun Soo Park, and Jan Kautz. Novel view synthesis of dynamic scenes with globally coherent depths from a monocular camera. In CVPR, 2020. 1, 6, 7", + "[57] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 6", + "[58] Kaichen Zhou, Jia-Xing Zhong, Sangyun Shin, Kai Lu, Yiyuan Yang, Andrew Markham, and Niki Trigoni. Dynpoint: Dynamic neural point for view synthesis. Advances in Neural Information Processing Systems, 36:69532-69545, 2023. 6" + ], + "bbox": [ + 516, + 92, + 903, + 784 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "26875", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/SplineGS_ Robust Motion-Adaptive Spline for Real-Time Dynamic 3D Gaussians from Monocular Video/c20079e6-f30b-47d1-a98c-7f27a87d2799_model.json b/2025/SplineGS_ Robust Motion-Adaptive Spline for Real-Time Dynamic 3D Gaussians from Monocular Video/c20079e6-f30b-47d1-a98c-7f27a87d2799_model.json new file mode 100644 index 0000000000000000000000000000000000000000..382ba6f12cafa256fcfe6cc5eceec38f870156a3 --- /dev/null +++ b/2025/SplineGS_ Robust Motion-Adaptive Spline for Real-Time Dynamic 3D Gaussians from Monocular Video/c20079e6-f30b-47d1-a98c-7f27a87d2799_model.json @@ -0,0 +1,2574 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.19, + 0.13, + 0.808, + 0.177 + ], + "angle": 0, + "content": "SplineGS: Robust Motion-Adaptive Spline for Real-Time Dynamic 3D Gaussians from Monocular Video" + }, + { + "type": "text", + "bbox": [ + 0.166, + 0.203, + 0.83, + 0.24 + ], + "angle": 0, + "content": "Jongmin Park1* Minh-Quan Viet Bui1* Juan Luis Gonzalez Bello1 Jaeho Moon1 \nJihyong Oh2† Munchurl Kim1†" + }, + { + "type": "text", + "bbox": [ + 0.133, + 0.256, + 0.865, + 0.31 + ], + "angle": 0, + "content": "\\(^{1}\\)Korea Advanced Institute of Science and Technology (KAIST) \\(^{2}\\)Chung-Ang University \\(\\{\\mathrm{jm.park, bvmquan, juanluisgb, jaehomoon, mkimee}\\} @kaist.ac.kr jihyongoh@cau.ac.kr https://kaist-viclab.github.io/splinegs-site/" + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.345, + 0.315, + 0.444 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.128, + 0.445, + 0.277, + 0.456 + ], + "angle": 0, + "content": "horsejump-low, \\(t = 20\\) (Ours)" + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.456, + 0.315, + 0.555 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.154, + 0.556, + 0.251, + 0.566 + ], + "angle": 0, + "content": "train, \\(t = 37\\) (Ours)" + }, + { + "type": "image", + "bbox": [ + 0.317, + 0.346, + 0.435, + 0.387 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.346, + 0.387, + 0.405, + 0.398 + ], + "angle": 0, + "content": "STGS [24]" + }, + { + "type": "image", + "bbox": [ + 0.436, + 0.346, + 0.554, + 0.387 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.46, + 0.388, + 0.529, + 0.4 + ], + "angle": 0, + "content": "D3DGS [55]" + }, + { + "type": "image", + "bbox": [ + 0.317, + 0.404, + 0.434, + 0.444 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.335, + 0.445, + 0.415, + 0.456 + ], + "angle": 0, + "content": "RoDynRF [30]" + }, + { + "type": "image", + "bbox": [ + 0.436, + 0.404, + 0.553, + 0.444 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.447, + 0.445, + 0.541, + 0.456 + ], + "angle": 0, + "content": "SplineGS (Ours)" + }, + { + "type": "image", + "bbox": [ + 0.316, + 0.456, + 0.434, + 0.496 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.346, + 0.497, + 0.405, + 0.509 + ], + "angle": 0, + "content": "STGS [24]" + }, + { + "type": "image", + "bbox": [ + 0.436, + 0.456, + 0.553, + 0.496 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.46, + 0.498, + 0.529, + 0.511 + ], + "angle": 0, + "content": "D3DGS [55]" + }, + { + "type": "image", + "bbox": [ + 0.317, + 0.513, + 0.434, + 0.555 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.335, + 0.555, + 0.415, + 0.567 + ], + "angle": 0, + "content": "RoDynRF [30]" + }, + { + "type": "image", + "bbox": [ + 0.436, + 0.513, + 0.554, + 0.555 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.447, + 0.555, + 0.542, + 0.567 + ], + "angle": 0, + "content": "SplineGS (Ours)" + }, + { + "type": "image_footnote", + "bbox": [ + 0.124, + 0.572, + 0.522, + 0.585 + ], + "angle": 0, + "content": "(a) Visual comparison for novel view synthesis on the DAVIS dataset" + }, + { + "type": "image", + "bbox": [ + 0.562, + 0.347, + 0.898, + 0.565 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.601, + 0.571, + 0.861, + 0.585 + ], + "angle": 0, + "content": "(b) Performance gain on the NVIDIA dataset" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.589, + 0.907, + 0.645 + ], + "angle": 0, + "content": "Figure 1. Our SplineGS achieves state-of-the-art rendering quality with fast rendering speed for novel spatio-temporal view synthesis from monocular videos without relying on pre-computed camera parameters. (a) We use our predicted camera parameters for [24, 55] since COLMAP [42] is unable to provide reasonable camera parameters for most scenes in the DAVIS dataset [39]. (b) SplineGS achieves the best rendering quality while maintaining a rendering speed of 400 FPS on the NVIDIA dataset [56]." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.657, + 0.327, + 0.673 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.689, + 0.486, + 0.902 + ], + "angle": 0, + "content": "Synthesizing novel views from in-the-wild monocular videos is challenging due to scene dynamics and the lack of multi-view cues. To address this, we propose SplineGS, a COLMAP-free dynamic 3D Gaussian Splatting (3DGS) framework for high-quality reconstruction and fast rendering from monocular videos. At its core is a novel Motion-Adaptive Spline (MAS) method, which represents continuous dynamic 3D Gaussian trajectories using cubic Hermite splines with a small number of control points. For MAS, we introduce a Motion-Adaptive Control points Pruning (MACP) method to model the deformation of each dynamic 3D Gaussian across varying motions, progressively pruning control points while maintaining dynamic modeling integrity. Additionally, we present a joint optimization strat" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.659, + 0.907, + 0.781 + ], + "angle": 0, + "content": "egy for camera parameter estimation and 3D Gaussian attributes, leveraging photometric and geometric consistency. This eliminates the need for Structure-from-Motion preprocessing and enhances SplineGS's robustness in real-world conditions. Experiments show that SplineGS significantly outperforms state-of-the-art methods in novel view synthesis quality for dynamic scenes from monocular videos, achieving thousands times faster rendering speed." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.792, + 0.646, + 0.807 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.818, + 0.907, + 0.865 + ], + "angle": 0, + "content": "Novel View Synthesis (NVS) is fundamental to 3D vision, supporting applications like virtual reality (VR), augmented reality (AR), and film production. NFS aims to generate" + }, + { + "type": "page_footnote", + "bbox": [ + 0.531, + 0.875, + 0.734, + 0.888 + ], + "angle": 0, + "content": "*Co-first authors (equal contribution)." + }, + { + "type": "page_footnote", + "bbox": [ + 0.534, + 0.888, + 0.681, + 0.901 + ], + "angle": 0, + "content": "† Co-corresponding authors." + }, + { + "type": "list", + "bbox": [ + 0.531, + 0.875, + 0.734, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "26866" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.228 + ], + "angle": 0, + "content": "images from any viewpoint in a scene, requiring accurate reconstruction from multiple 2D images. NeRF [34] has advanced the field of NVs by utilizing learned implicit functions to represent static scenes, though it requires considerable training and rendering time. Recently, 3D Gaussian Splatting (3DGS) [20] has revolutionized this process by replacing implicit volumetric rendering with differentiable rasterization of 3D Gaussians, enabling real-time rendering and providing a more explicit scene representation." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.232, + 0.485, + 0.549 + ], + "angle": 0, + "content": "For NVs of dynamic scenes, prior works have extended the static scene representations [20, 34] by incorporating deformation models in canonical space using implicit representations [12, 30, 35, 36, 55], grid-based models that decompose the 4D space-time domain into multiple 2D planes [4, 6, 10, 43, 52], and polynomial trajectories [24]. However, these methods face several challenges: implicit representations significantly increase computational overhead and reduce rendering speed [6, 52]; grid-based models struggle to fully capture the dynamic nature of scene structures, hindering their ability to model fine details accurately [24]; and although polynomial trajectories improve efficiency, their fixed degree restricts flexibility for representing complex motions. To the best of our knowledge, none of the dynamic 3DGS-based methods provide experimental evidence to reliably support their novel spatio-temporal view synthesis capabilities for rendering unseen intermediate time indices. Additionally, most existing methods for dynamic NVs rely heavily on external camera parameter estimators like COLMAP [42], which often produce imprecise results for challenging in-the-wild monocular videos." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.554, + 0.483, + 0.901 + ], + "angle": 0, + "content": "To address these issues for modeling scene dynamics, we exploit a spline-based model to dynamic 3D Gaussian trajectories, inspired by classic 3D-curve modeling in computer graphics [9]. Splines are widely used in geometric modeling for graphical applications, providing smooth and continuous representations of complex shapes with a minimal number of control points [2, 8]. This efficiency makes them ideal for modeling intricate geometric structures while maintaining both flexibility and precision. In this paper, we propose SplineGS, a framework for high-quality dynamic scene reconstruction and real-time neural rendering from inthe-wild monocular videos without relying on external camera estimators like COLMAP [42]. SplineGS introduces Motion-Adaptive Spline (MAS), based on cubic Hermite splines [2, 8], to effectively represent dynamic 3D Gaussian deformations. Our MAS consists of piecewise cubic functions defined by control points that dictate each segment's curvature and direction. Each control point is adjustable and optimized as a learnable parameter, enabling faster and more precise modeling of complex motion trajectories. Additionally, to adaptively model the trajectory of each dynamic 3D Gaussian based on motion complexity during training, we introduce a Motion-Adaptive Control" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.243 + ], + "angle": 0, + "content": "points Pruning (MACP) method that adjusts the number of control points to improve rendering quality and efficiency. Furthermore, we incorporate a camera parameter estimation method jointly optimized with 3D Gaussian attributes and MAS, leveraging photometric and geometric consistency, thus eliminating the need for external estimators. Experiments show that SplineGS significantly outperforms state-of-the-art (SOTA) dynamic NVs methods both qualitatively and quantitatively, offering faster rendering speed, as shown in Fig. 1. Our contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.247, + 0.905, + 0.308 + ], + "angle": 0, + "content": "- We propose a novel Spline-based dynamic 3D Gaussian Splatting framework, SplineGS, which is (i) COLMAP-free, (ii) very fast and (iii) of high quality in reconstructing the dynamic scenes from in-the-wild monocular videos;" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.311, + 0.904, + 0.357 + ], + "angle": 0, + "content": "- A novel Motion-Adaptive Spline (MAS) is introduced, which can accurately and effectively represent the continuous trajectory of each dynamic 3D Gaussian;" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.36, + 0.905, + 0.421 + ], + "angle": 0, + "content": "- A Motion-Adaptive Control points Pruning (MACP) method is presented, which can efficiently adjust the number of control points for each spline function, optimizing both rendering quality and the efficiency of MAS." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.247, + 0.905, + 0.421 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.436, + 0.655, + 0.451 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.462, + 0.907, + 0.658 + ], + "angle": 0, + "content": "Dynamic NeRF. Recent advancements in video view synthesis have extended the static NeRF model [34] to represent scene dynamics. These include dynamic NeRFs using scene flow-based frameworks [12, 25, 26], deformation estimation in canonical fields [1, 3, 16, 30, 35, 36, 40, 44, 47, 51, 53], and 4D grid-based spatio-temporal radiance fields [4, 6, 10, 43]. Techniques such as NSFF [25], DynNeRF [12], and DynIBaR [26] combine time-independent and time-dependent radiance fields to synthesize novel spatiotemporal perspectives from monocular videos. Despite these advancements, current dynamic NeRFs still fall short of recent dynamic 3D Gaussian Splatting (3DGS) techniques in rendering quality and efficiency." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.66, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Dynamic 3DGS. The improvements in rendering quality and speed achieved by 3DGS [20] have inspired further studies [14, 24, 27, 31, 52, 55] to extend the static 3DGS framework to dynamic scenes by enabling the deformation of 3D Gaussian attributes. The pioneering work on dynamic 3DGS [31] introduces time-dependent offsets for the positions and rotations of dynamic 3D Gaussians via an MLP; however, this approach slows down the rendering. D3DGS [55] builds on this concept with an annealing smoothing training mechanism to improve temporal smoothness and rendering quality. 4DGS [52] replaces the MLP network with a grid-based structure to boost efficiency, though this change requires quality trade-offs due to the resolution limits of grid-based methods. In contrast, SC-GS [14] combines an MLP deformation network with sparse spatial deformation, reducing the computational" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "26867" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.091, + 0.485, + 0.592 + ], + "angle": 0, + "content": "cost of MLP while maintaining quality. STGS [24] introduces polynomial trajectories for motion modeling, improving speed and quality over implicit representations. CasualFVS [22] warps dynamic content from neighboring frames, and MoSca [23] proposes a graph-based motion modeling approach to handle sparse control deformation. Very recently, a concurrent work [21] proposes modeling dynamic 3D Gaussian trajectories using polynomial interpolation between fixed keyframes for NVS from multi-view videos with COLMAP [42] assistance. Unlike [21], our SplineGS adaptively optimizes the deformation of dynamic 3D Gaussians, accounting for varying motion degrees and types in in-the-wild monocular videos, without requiring preprocessed camera parameters (COLMAP-free approach). Neural Rendering without SfM Preprocessing. Accurate camera parameters, including extrinsic and intrinsic, are essential for neural rendering approaches to capture fine details [17, 28]. However, in real-world settings, camera parameters derived from Structure-from-Motion (SfM) algorithms such as COLMAP [42] often exhibit pixel-level inaccuracies [29, 41], compromising the structural details of rendered scenes [30]. To address this, several NeRF methods [28, 32, 37, 50] jointly optimize NeRF architectures and camera parameters. Recently, a local-to-global training approach [11] is introduced to optimize both camera parameters and 3D Gaussians. However, these methods are limited to static scenes. For dynamic scene reconstruction, RoDynRF [30] and MoSca [23] use motion masks to gather multi-view cues from static regions, allowing robust rendering without pre-computed camera parameters. Our SplineGS is also COLMAP-free, significantly outperforming RoDynRF [30] and MoSca [23] in rendering quality and efficiency, enabled by our novel spline-based architecture." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.602, + 0.418, + 0.62 + ], + "angle": 0, + "content": "3. Preliminary: 3D Gaussian Splitting" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.627, + 0.483, + 0.657 + ], + "angle": 0, + "content": "3DGS [20] represents the radiance field of a scene using anisotropic 3D Gaussians, each of which is formulated as" + }, + { + "type": "equation", + "bbox": [ + 0.185, + 0.66, + 0.482, + 0.68 + ], + "angle": 0, + "content": "\\[\nG (\\boldsymbol {x}) = e ^ {- \\frac {1}{2} (\\boldsymbol {x} - \\boldsymbol {\\mu}) ^ {\\top} \\boldsymbol {\\Sigma} ^ {- 1} (\\boldsymbol {x} - \\boldsymbol {\\mu})} \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.684, + 0.483, + 0.881 + ], + "angle": 0, + "content": "where \\(\\pmb{x} \\in \\mathbb{R}^3\\) denotes a 3D position, and \\(\\pmb{\\mu} \\in \\mathbb{R}^3\\) and \\(\\pmb{\\Sigma} \\in \\mathbb{R}^{3 \\times 3}\\) represent the mean (center) and the covariance matrix of the 3D Gaussian, respectively. To ensure that \\(\\pmb{\\Sigma}\\) is positive semi-definite and contains physical meaning, it is decomposed into a diagonal scaling matrix \\(S \\in \\mathbb{R}^{3 \\times 3}\\) of a scale vector \\(s \\in \\mathbb{R}^3\\) and a rotation matrix \\(R \\in \\mathbb{R}^{3 \\times 3}\\) as \\(\\pmb{\\Sigma} = R\\pmb{S}\\pmb{S}^\\top\\pmb{R}^\\top\\), where \\(\\pmb{R}\\) is parameterized by a learnable unit-length quaternion \\(q \\in \\mathbb{R}^4\\). In addition, each 3D Gaussian is parameterized by an opacity \\(\\sigma \\in \\mathbb{R}\\) and a color \\(c \\in \\mathbb{R}^3\\). To render the color of each pixel, the color and the opacity of each 3D Gaussian are computed using Eq. 1, and the rendered color \\(C\\) is computed by the alpha-blending of the \\(\\mathcal{N}\\) ordered 3D Gaussians overlapping the pixel as" + }, + { + "type": "equation", + "bbox": [ + 0.176, + 0.884, + 0.483, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {C} = \\sum_ {i \\in \\mathcal {N}} \\boldsymbol {c} _ {i} \\alpha_ {i} \\prod_ {j = 1} ^ {i - 1} (1 - \\alpha_ {j}), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.091, + 0.907, + 0.245 + ], + "angle": 0, + "content": "where \\( c_{i} \\) is the color of the \\( i^{\\mathrm{th}} \\) 3D Gaussian and \\( \\alpha_{i} \\) is a density of the \\( i^{\\mathrm{th}} \\) 3D Gaussian which is given by evaluating the 2D covariance \\( \\pmb{\\Sigma}' \\in \\mathbb{R}^{2 \\times 2} \\). Here, \\( \\pmb{\\Sigma}' \\) is formulated as \\( \\pmb{\\Sigma}' = J\\pmb{W}\\pmb{\\Sigma}\\pmb{W}^\\top J^\\top \\), where \\( J \\) is the Jacobian of the affine approximation of the projective transformation and \\( W \\) is a viewing transformation matrix. Similar to prior works [23, 27], we extend 3DGS [20] to a union of static 3D Gaussians \\( \\{G_i^{\\mathrm{st}} | i = 1, 2, \\dots, n^{\\mathrm{st}}\\} \\) and dynamic 3D Gaussians \\( \\{G_i^{\\mathrm{dy}} | i = 1, 2, \\dots, n^{\\mathrm{dy}}\\} \\) to represent static backgrounds and moving objects, respectively, in dynamic scenes." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.263, + 0.779, + 0.281 + ], + "angle": 0, + "content": "4. Proposed Method: SplineGS" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.289, + 0.722, + 0.306 + ], + "angle": 0, + "content": "4.1. Overview of SplineGS" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.314, + 0.907, + 0.693 + ], + "angle": 0, + "content": "We describe the overall architecture of SplineGS in Fig. 2. Given a monocular video \\(\\{I_t|t = 0,1,\\dots,N_f - 1\\}\\) where \\(N_{f}\\) is the total number of frames, SplineGS is designed to synthesize high-quality novel spatio-temporal views with fast rendering speed, and to estimate the camera parameters, including the extrinsic \\([\\hat{R}_t|\\hat{T}_t]\\in \\mathbb{R}^{3\\times 4}\\) for each time \\(t\\), and the shared intrinsic \\(\\hat{\\pmb{K}}\\in \\mathbb{R}^{3\\times 3}\\). As shown in Fig. 2, to stabilize joint optimization of the 3D Gaussian attributes and camera parameter estimation, we adopt a two-stage optimization process consisting of a warm-up stage and a main training stage for our SplineGS architecture. In the warm-up stage, we coarsely optimize the camera parameters \\([\\hat{R}_t|\\hat{T}_t]\\) and \\(\\hat{\\pmb{K}}\\) using photometric and geometric consistency. In the main training stage, we initialize the 3D Gaussians based on the estimated camera poses and jointly optimize the 3D Gaussian attributes with the camera parameters. Specifically, for each dynamic 3D Gaussian, we propose a novel spline-based deformation modeling method, called Motion-Adaptive Spline (MAS), which utilizes a cubic Hermite spline function [2, 8] to accurately model the continuous trajectory with a small number of control points. For MAS, we introduce a Motion-Adaptive Control points Pruning (MACP) method that effectively adjusts the number of control points for each dynamic 3D Gaussian, considering the object's motion types and degrees." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.695, + 0.906, + 0.755 + ], + "angle": 0, + "content": "In the following sections, we detail the process of our MAS method in Sec. 4.2, followed by the camera parameter estimation process in Sec. 4.3. Finally, we describe the overall optimization process in Sec. 4.4." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.769, + 0.871, + 0.785 + ], + "angle": 0, + "content": "4.2. Motion-Adaptive Spline for 3D Gaussians" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.794, + 0.906, + 0.883 + ], + "angle": 0, + "content": "To represent the continuous trajectory of each dynamic 3D Gaussian for moving objects over time, we propose MAS which modifies the mean parameter to a set of learnable control points. For this, we use the cubic Hermite spline function [2, 8] to model the mean of each dynamic 3D Gaussian at time \\( t \\) as" + }, + { + "type": "equation", + "bbox": [ + 0.655, + 0.886, + 0.907, + 0.902 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\mu} (t) = S (t, \\mathbf {P}), \\tag {3}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "26868" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.091, + 0.087, + 0.91, + 0.405 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.41, + 0.908, + 0.468 + ], + "angle": 0, + "content": "Figure 2. Overview of SplineGS. Our SplineGS leverages spline-based functions to model the deformation of dynamic 3D Gaussians with a novel Motion-Adaptive Spline (MAS) architecture. It is composed of sets of learnable control points based on a cubic Hermite spline function [2, 8] to accurately model the trajectory of each dynamic 3D Gaussian and to achieve faster rendering speed. To avoid any preprocessing of camera parameters, i.e., COLMAP-free, we adopt a two-stage optimization: warm-up and main training stages." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.474, + 0.483, + 0.521 + ], + "angle": 0, + "content": "where \\(\\mathbf{P} = \\{\\mathbf{p}_k|\\mathbf{p}_k\\in \\mathbb{R}^3,k\\in 0,\\dots ,N_c - 1\\}\\) is a set of \\(N_{c}\\) learnable control points, serving as an additional attribute for each dynamic 3D Gaussian, and \\(S(t,\\mathbf{P})\\) is defined as" + }, + { + "type": "equation", + "bbox": [ + 0.106, + 0.522, + 0.482, + 0.561 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} S (t, \\mathbf {P}) = \\left(2 t _ {r} ^ {3} - 3 t _ {r} ^ {2} + 1\\right) \\mathbf {p} _ {\\lfloor t _ {s} \\rfloor} + \\left(t _ {r} ^ {3} - 2 t _ {r} ^ {2} + t _ {r}\\right) \\mathbf {m} _ {\\lfloor t _ {s} \\rfloor} \\\\ + \\left(- 2 t _ {r} ^ {3} + 3 t _ {r} ^ {2}\\right) \\mathbf {p} _ {[ t _ {s} ] + 1} + \\left(t _ {r} ^ {3} - t _ {r} ^ {2}\\right) \\mathbf {m} _ {[ t _ {s} ] + 1}, \\tag {4} \\\\ \\end{array}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.564, + 0.455, + 0.58 + ], + "angle": 0, + "content": "\\[\nt _ {r} = t _ {s} - \\lfloor t _ {s} \\rfloor , \\quad t _ {s} = t _ {n} (N _ {c} - 1), \\quad t _ {n} = t / (N _ {f} - 1),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.582, + 0.483, + 0.687 + ], + "angle": 0, + "content": "where \\(\\mathbf{m}_k\\) is the approximated tangent of the control point \\(\\mathbf{p}_k\\), formulated as \\(\\mathbf{m}_k = (\\mathbf{p}_{k + 1} - \\mathbf{p}_{k - 1}) / 2\\). Note that \\(S(t,\\mathbf{P})\\) indicates the piecewise cubic function to represent the whole continuous trajectory of each dynamic 3D Gaussian, and \\(N_{c}\\) can be different from \\(N_{f}\\). The optimal \\(N_{c}\\) can be estimated by Motion-Adaptive Control points Pruning (MACP) in the following section." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.687, + 0.483, + 0.883 + ], + "angle": 0, + "content": "Control Points Initialization. To stably optimize \\( S(t, \\mathbf{P}) \\), it is essential to initialize the control points with appropriate geometric considerations that ensure temporal consistency. For this, we leverage the long-range 2D tracking [19] and the metric depth [38] priors. Let \\( \\mathcal{T} = \\{\\varphi_t^{\\mathrm{tr}} | \\varphi_t^{\\mathrm{tr}} \\in \\mathbb{R}^2, t = 0, \\dots, N_f - 1\\} \\) be a 2D track, \\( \\pi_K(\\cdot) \\) be a projection function from the camera space to the image space with the camera intrinsic \\( K \\), and \\( \\varphi_t^{\\mathrm{tr}} \\) be a pixel coordinate corresponding to the 2D track \\( \\mathcal{T} \\) at time \\( t \\). We unproject each 2D track \\( \\mathcal{T} \\) to the world space to compute the 3D track aided by the frame's metric depth \\( d_t \\) and camera extrinsic \\( [\\hat{R}_t | \\hat{T}_t] \\) at time \\( t \\). The unprojection function \\( \\mathcal{W}_t(\\cdot) \\) from the image space to the world space is formulated as" + }, + { + "type": "equation", + "bbox": [ + 0.142, + 0.885, + 0.483, + 0.907 + ], + "angle": 0, + "content": "\\[\n\\mathcal {W} _ {t} \\left(\\boldsymbol {\\varphi} _ {t} ^ {\\mathrm {t r}}\\right) = \\hat {\\boldsymbol {R}} _ {t} ^ {\\top} \\pi_ {\\hat {\\boldsymbol {K}}} ^ {- 1} \\left(\\boldsymbol {\\varphi} _ {t} ^ {\\mathrm {t r}}, \\boldsymbol {d} _ {t} \\left(\\boldsymbol {\\varphi} _ {t} ^ {\\mathrm {t r}}\\right)\\right) - \\hat {\\boldsymbol {R}} _ {t} ^ {\\top} \\hat {\\boldsymbol {T}} _ {t}. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.474, + 0.907, + 0.566 + ], + "angle": 0, + "content": "It should be noted that we estimate \\([\\hat{\\pmb{R}}_t|\\hat{\\pmb{T}}_t]\\) and \\(\\hat{\\pmb{K}}\\) from a sequence of frames only, without using any given ground truth values, as mentioned in Sec. 4.3. Then, we initialize the per-Gaussian control points set \\(\\mathbf{P}\\), by a least-square (LS) approximation, such that the spline-described curve fits the initial curve by the tracker, given by" + }, + { + "type": "equation", + "bbox": [ + 0.588, + 0.57, + 0.907, + 0.595 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\mathbf {P}} \\sum_ {t = 0} ^ {N _ {f} - 1} \\left\\| \\mathcal {W} _ {t} \\left(\\varphi_ {t} ^ {\\mathrm {t r}}\\right) - S (t, \\mathbf {P}) \\right\\| _ {2} ^ {2}. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.599, + 0.909, + 0.887 + ], + "angle": 0, + "content": "Motion-Adaptive Control Points Pruning (MACP). An excessive number of control points may cause over-fitting and decrease the processing speed of our MAS module, leading to poorer rendering qualities with reduced rendering speed (see Table 3). Furthermore, as each scene exhibits various types and degrees of motion for moving objects, the number of control points required for each dynamic 3D Gaussian trajectory should be adaptively adjusted to accommodate the scene dynamics. To achieve this, we propose the MACP method for MAS, which can generate sparser control points while ensuring no dynamic modeling degradation. MACP is designed on top of 3D Gaussian densification [20], but focuses on optimizing the number of control points in MAS. Our MACP computes a new spline function \\( S(t,\\mathbf{P}^{\\prime}) \\) after every 3D Gaussian densification step, where \\( \\mathbf{P}^{\\prime} = \\{\\mathbf{p}_{l}^{\\prime}|\\mathbf{p}_{l}^{\\prime}\\in \\mathbb{R}^{3}, l = 0,\\dots ,N_{c} - 2\\} \\) is a set of \\( N_{c} - 1 \\) control points, which contains one-fewer control points than the current set \\( \\mathbf{P} \\) of control points. We compute the LS approximation to find the reduced set \\( \\mathbf{P}^{\\prime} \\) of control points as" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "26869" + } + ], + [ + { + "type": "equation", + "bbox": [ + 0.168, + 0.088, + 0.483, + 0.114 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\mathbf {P} ^ {\\prime}} \\sum_ {t = 0} ^ {N _ {f} - 1} \\left\\| S (t, \\mathbf {P}) - S (t, \\mathbf {P} ^ {\\prime}) \\right\\| _ {2} ^ {2}. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.117, + 0.483, + 0.163 + ], + "angle": 0, + "content": "Then, the updated optimal set \\(\\mathbf{P}\\) of control points is assigned with the values of \\(\\mathbf{P}'\\) if the error \\(E\\) between \\(S(t,\\mathbf{P})\\) and \\(S(t,\\mathbf{P}')\\) is smaller than a threshold \\(\\epsilon\\), as given by" + }, + { + "type": "equation", + "bbox": [ + 0.18, + 0.168, + 0.483, + 0.209 + ], + "angle": 0, + "content": "\\[\n\\mathbf {P} = \\left\\{ \\begin{array}{l l} \\mathbf {P} ^ {\\prime}, & \\text {i f} E < \\epsilon \\\\ \\mathbf {P}, & \\text {o t h e r w i s e} \\end{array} , \\right. \\text {w h e r e} \\tag {8}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.093, + 0.204, + 0.454, + 0.24 + ], + "angle": 0, + "content": "\\[\nE = \\frac {1}{N _ {f}} \\sum_ {t = 0} ^ {N _ {f} - 1} \\| \\pi_ {\\hat {\\mathbf {K}}} (\\hat {\\mathbf {R}} _ {t} S (t, \\mathbf {P}) + \\hat {\\mathbf {T}} _ {t}) - \\pi_ {\\hat {\\mathbf {K}}} (\\hat {\\mathbf {R}} _ {t} S (t, \\mathbf {P} ^ {\\prime}) + \\hat {\\mathbf {T}} _ {t}) \\| _ {2} ^ {2}.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.25, + 0.485, + 0.327 + ], + "angle": 0, + "content": "By following MACP, each dynamic 3D Gaussian is allowed to have a different number of control points. Therefore, the MACP can guide our MAS to have a minimal number of control points when modeling simple motion, while having more control points for more complex motions (see Fig. 8)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.336, + 0.364, + 0.351 + ], + "angle": 0, + "content": "4.3. Camera Parameter Estimation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.36, + 0.484, + 0.481 + ], + "angle": 0, + "content": "The traditional SfM methods, such as COLMAP [42], fail to reliably estimate camera parameters in dynamic scenes from in-the-wild monocular videos [30]. For this reason, we propose to estimate camera parameters for joint optimization with the 3D Gaussian attributes. For each frame at time \\( t \\), we predict a rotation matrix \\( \\hat{R}_t \\in \\mathbb{R}^{3 \\times 3} \\) and a translation vector \\( \\hat{T}_t \\in \\mathbb{R}^3 \\), representing the extrinsic parameters of a monocular camera, using a shallow MLP \\( F_{\\theta} \\) as" + }, + { + "type": "equation", + "bbox": [ + 0.218, + 0.486, + 0.482, + 0.503 + ], + "angle": 0, + "content": "\\[\n\\left[ \\boldsymbol {\\hat {R}} _ {t} \\mid \\boldsymbol {\\hat {T}} _ {t} \\right] = F _ {\\theta} (\\gamma (t)), \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.508, + 0.484, + 0.631 + ], + "angle": 0, + "content": "where \\(\\gamma (\\cdot)\\) is a positional encoding [34]. We also predict the focal length \\(\\hat{f}\\) of the camera intrinsic \\(\\hat{K}\\) as a learnable parameter that is shared across all frames in the monocular video. To accurately optimize the camera parameters, we enforce two types of consistency—photometric and geometric—for the static background between random reference cameras \\([\\hat{R}_{t_{\\mathrm{ref}}}|\\hat{T}_{t_{\\mathrm{ref}}}]\\) and the target camera \\([\\hat{R}_t|\\hat{T}_t]\\), encouraging alignment both visually and structurally." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.631, + 0.484, + 0.722 + ], + "angle": 0, + "content": "Photometric Consistency. Given the pre-computed metric depth [38], the camera intrinsics and extrinsics under optimization will converge as long as the projected reference frame's color \\( I_{t_{\\mathrm{ref}}}(\\varphi_{t \\to t_{\\mathrm{ref}}}) \\) is aligned to the target frame's color \\( I_t(\\varphi_t) \\). \\( \\varphi_{t \\to t_{\\mathrm{ref}}} \\) is the reference pixel coordinate corresponding to the target frame's pixel coordinate \\( \\varphi_t \\) as" + }, + { + "type": "equation", + "bbox": [ + 0.102, + 0.727, + 0.483, + 0.746 + ], + "angle": 0, + "content": "\\[\n\\varphi_ {t \\rightarrow t _ {\\text {r e f}}} = \\pi_ {\\hat {\\mathbf {K}}} \\left(\\hat {\\mathbf {R}} _ {t _ {\\text {r e f}}} \\left(\\hat {\\mathbf {R}} _ {t} ^ {\\top} \\pi_ {\\hat {\\mathbf {K}}} ^ {- 1} \\left(\\varphi_ {t}, \\boldsymbol {d} _ {t} (\\varphi_ {t})\\right) - \\hat {\\mathbf {R}} _ {t} ^ {\\top} \\hat {\\boldsymbol {T}} _ {t}\\right) + \\hat {\\boldsymbol {T}} _ {t _ {\\text {r e f}}}\\right). \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.78 + ], + "angle": 0, + "content": "We refer to such projection alignment as photometric consistency, which is encouraged by the loss \\(\\mathcal{L}_{\\mathrm{pc}}\\) given by" + }, + { + "type": "equation", + "bbox": [ + 0.108, + 0.785, + 0.483, + 0.806 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {p c}} = \\sum_ {\\boldsymbol {\\varphi} _ {t}} \\left\\| \\boldsymbol {M} _ {t, t _ {\\text {r e f}}} \\left(\\boldsymbol {\\varphi} _ {t}\\right) \\odot \\left(\\boldsymbol {I} _ {t} \\left(\\boldsymbol {\\varphi} _ {t}\\right) - \\boldsymbol {I} _ {t _ {\\text {r e f}}} \\left(\\boldsymbol {\\varphi} _ {t \\rightarrow t _ {\\text {r e f}}}\\right)\\right)\\right\\| _ {2} ^ {2}, \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.816, + 0.484, + 0.877 + ], + "angle": 0, + "content": "where \\(\\odot\\) is the Hadamard product [13], and \\(M_{t,t_{\\mathrm{ref}}}\\) is a union motion mask that excludes dynamic objects in \\(I_{t}\\) and \\(I_{t_{\\mathrm{ref}}}\\) for removing the inconsistencies due to dynamic regions, which is given by" + }, + { + "type": "equation", + "bbox": [ + 0.161, + 0.885, + 0.483, + 0.903 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {M} _ {t, t _ {\\text {r e f}}} \\left(\\varphi_ {t}\\right) = \\boldsymbol {M} _ {t} \\left(\\varphi_ {t}\\right) \\boldsymbol {M} _ {t _ {\\text {r e f}}} \\left(\\varphi_ {t \\rightarrow t _ {\\text {r e f}}}\\right), \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.136 + ], + "angle": 0, + "content": "where \\(M_{t}\\) and \\(M_{t_{\\mathrm{ref}}}\\) are pre-computed motion masks [54] of \\(\\pmb{I}_t\\) and \\(I_{t_{\\mathrm{ref}}}\\), respectively. Note that motion mask's value is 0 for static regions, and 1 otherwise." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.137, + 0.905, + 0.198 + ], + "angle": 0, + "content": "Geometric Consistency. Along with the photometric consistency, we compute the geometric consistency of unprojected pixels in 3D space to make our optimization more stable. The geometric consistency loss is formulated as" + }, + { + "type": "equation", + "bbox": [ + 0.522, + 0.202, + 0.905, + 0.221 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {g c}} = \\sum_ {\\boldsymbol {\\varphi} _ {t}} \\| M _ {t, t _ {\\text {r e f}}} (\\boldsymbol {\\varphi} _ {t}) \\odot \\left(\\mathcal {W} _ {t} (\\boldsymbol {\\varphi} _ {t}) - \\mathcal {W} _ {t _ {\\text {r e f}}} \\left(\\boldsymbol {\\varphi} _ {t \\rightarrow t _ {\\text {r e f}}}\\right)\\right) \\| _ {2} ^ {2}, \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.224, + 0.905, + 0.255 + ], + "angle": 0, + "content": "where \\(\\mathcal{W}_t(\\varphi_t)\\) and \\(\\mathcal{W}_{t_{\\mathrm{ref}}}(\\varphi_{t\\to t_{\\mathrm{ref}}})\\) are the corresponding 3D locations of \\(\\varphi_t\\) and \\(\\varphi_{t\\rightarrow t_{\\mathrm{ref}}}\\), respectively (see Eq. 5)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.262, + 0.654, + 0.278 + ], + "angle": 0, + "content": "4.4. Optimization" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.285, + 0.906, + 0.376 + ], + "angle": 0, + "content": "To stabilize joint training of the MAS and the camera parameter estimation, we adopt a two-stage optimization process consisting of the warm-up stage and the main training stage. In the warm-up stage, we optimize only the camera parameters using the photometric and geometric consistency. The total loss in the warm-up stage is given by" + }, + { + "type": "equation", + "bbox": [ + 0.621, + 0.382, + 0.905, + 0.4 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {t o t a l}} ^ {\\text {w a r m}} = \\lambda_ {\\mathrm {p c}} \\mathcal {L} _ {\\mathrm {p c}} + \\lambda_ {\\mathrm {g c}} \\mathcal {L} _ {\\mathrm {g c}}. \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.406, + 0.906, + 0.511 + ], + "angle": 0, + "content": "After the warm-up stage, we obtain the coarsely predicted camera intrinsic \\(\\hat{\\pmb{K}}\\) and the set of extrinsics \\([\\hat{\\pmb{R}}|\\hat{\\pmb{T}}]\\) for all frames, which are then used to initialize the set \\(\\mathbf{P}\\) of control points for each dynamic 3D Gaussian, as described in Sec. 4.2. In the main training stage, we jointly optimize the static and dynamic 3D Gaussians along with the camera parameter estimation based on the total loss function as" + }, + { + "type": "equation", + "bbox": [ + 0.582, + 0.517, + 0.837, + 0.554 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {\\text {t o t a l}} ^ {\\text {m a i n}} = \\lambda_ {\\mathrm {r g b}} \\mathcal {L} _ {\\mathrm {r g b}} + \\lambda_ {\\mathrm {d}} \\mathcal {L} _ {\\mathrm {d}} + \\lambda_ {\\mathrm {M}} \\mathcal {L} _ {\\mathrm {M}} \\\\ + \\lambda_ {\\mathrm {p c}} \\mathcal {L} _ {\\mathrm {p c}} + \\lambda_ {\\mathrm {d - p c}} \\mathcal {L} _ {\\mathrm {d - p c}} + \\lambda_ {\\mathrm {g c}} \\mathcal {L} _ {\\mathrm {g c}}, \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.875, + 0.529, + 0.905, + 0.542 + ], + "angle": 0, + "content": "(15)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.559, + 0.906, + 0.68 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_{\\mathrm{rgb}}\\) and \\(\\mathcal{L}_{\\mathrm{d}}\\) are the L1 losses between the rendered frame and GT frame, and between the rendered depth and GT depth, respectively. In the main training stage, we compute an additional photometric consistency loss \\(\\mathcal{L}_{\\mathrm{d - pc}}\\) that utilizes the rendered depth \\(\\hat{\\pmb{d}}\\) of the 3D Gaussians instead of the metric depth [38] prior \\(\\pmb{d}\\) as \\(\\hat{\\pmb{d}}\\) allows the estimated 3D Gaussian geometry to guide the joint optimization of the camera parameters and 3D Gaussian attributes." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.682, + 0.906, + 0.803 + ], + "angle": 0, + "content": "In addition to the camera parameter estimation and imagery reconstruction losses, we adopt a binary dice loss [45] \\(\\mathcal{L}_{\\mathrm{M}}\\) between the pre-computed motion mask \\(M_{t}\\) [54] and the rendered motion mask \\(\\hat{M}_t\\) that can be derived from the dynamic 3D Gaussians. The binary dice loss initially proposed in [45] for highly imbalanced segmentation of medical imagery helps encouraging better separation between our dynamic and static 3D Gaussians as described by" + }, + { + "type": "equation", + "bbox": [ + 0.547, + 0.807, + 0.905, + 0.853 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {M}} = 1 - \\frac {2 \\sum_ {\\varphi_ {t}} M _ {t} \\left(\\varphi_ {t}\\right) \\hat {M} _ {t} \\left(\\varphi_ {t}\\right) + \\varepsilon}{\\sum_ {\\varphi_ {t}} \\left(M _ {t} \\left(\\varphi_ {t}\\right) + \\hat {M} _ {t} \\left(\\varphi_ {t}\\right)\\right) + \\varepsilon}, \\tag {16}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.856, + 0.905, + 0.902 + ], + "angle": 0, + "content": "where \\(\\varepsilon\\) is a smooth term to avoid numerical issues. \\(\\hat{M}_t(\\varphi_t)\\) is computed by the alpha-blending of the 3D Gaussians overlapping \\(\\varphi_t\\) (similar to Eq. 2) as" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "26870" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.09, + 0.912, + 0.337 + ], + "angle": 0, + "content": "
MethodsJumpingSkatingTruckUmbrellaBalloon1Balloon2PlaygroundAverageFPS↑Tr. (h)↓
STM preprocessingDynNeRF (ICCV'21) [12]24.68 / 0.09032.66 / 0.03528.56 / 0.08223.26 / 0.13722.36 / 0.10427.06 / 0.04924.15 / 0.08026.10 / 0.0820.05>24
MonoNeRF (ICCV'23) [46]24.26 / 0.09132.06 / 0.04427.56 / 0.11523.62 / 0.18021.89 / 0.12927.36 / 0.05222.61 / 0.13025.62 / 0.1060.05N/A
STGS (CVPR'24) [24]20.82 / 0.18724.80 / 0.10925.01 / 0.10321.88 / 0.19520.36 / 0.19623.12 / 0.12419.23 / 0.15122.17 / 0.1529000.2
SCGS (CVPR'24) [14]15.68 / 0.92014.88 / 0.90823.81 / 0.14021.84 / 0.16020.17 / 0.17921.07 / 0.14920.71 / 0.11519.74 / 0.3671101.1
D3DGS (CVPR'24) [55]22.02 / 0.26624.06 / 0.22723.04 / 0.24722.67 / 0.19221.22 / 0.20225.86 / 0.11822.30 / 0.11123.02 / 0.195251.0
4DGS (CVPR'24) [52]22.37 / 0.17826.72 / 0.08425.93 / 0.09722.36 / 0.17821.89 / 0.15324.85 / 0.08121.36 / 0.08923.64 / 0.123950.25
SP-GS (ICML'24) [48]22.13 / 0.46829.21 / 0.23627.38 / 0.19024.88 / 0.32324.36 / 0.18029.65 / 0.09722.29 / 0.23425.70 / 0.2471806.0
Casual-FVS (ECCV'24) [22]23.45 / 0.10029.98 / 0.04525.22 / 0.09023.24 / 0.09623.76 / 0.07924.15 / 0.08122.19 / 0.07424.57 / 0.081480.25
Kinematic Fields (ECCV'24) [15]23.74 / 0.10231.89 / 0.03528.34 / 0.07425.46 / 0.09823.72 / 0.07926.49 / 0.05524.64 / 0.05326.33 / 0.0710.37.0
E-D3DGS (ECCV'24) [5]22.39 / 0.18024.62 / 0.13726.85 / 0.08123.60 / 0.14520.87 / 0.16124.39 / 0.07320.79 / 0.12423.36 / 0.129452.6
CTNeRF (Pattern Recognition'24) [33]24.35 / 0.09433.51 / 0.03428.27 / 0.08423.48 / 0.12922.19 / 0.11126.86 / 0.04824.28 / 0.07726.13 / 0.0820.1>24
D-NPC (Eurographics'25) [18]24.51 / 0.11630.22 / 0.06128.92 / 0.08424.15 / 0.13622.26 / 0.16125.84 / 0.10723.59 / 0.09925.64 / 0.109700.27
Shape-of-Motion (arXiv'24) [49]22.74 / 0.12829.12 / 0.04824.63 / 0.12623.79 / 0.08922.74 / 0.12025.83 / 0.07220.61 / 0.13024.21 / 0.1022551.1
Ex4DGS (NeurIPS'24) [21]18.93 / 0.32121.92 / 0.23319.04 / 0.30819.03 / 0.34014.69 / 0.50316.29 / 0.45714.16 / 0.43717.72 / 0.371840.65
DynPoint (NeurIPS'23) [58]24.69 / 0.09731.34 / 0.04529.30 / 0.06124.59 / 0.08622.77 / 0.09927.63 / 0.04925.37 / 0.03926.53 / 0.0680.80.5
RoDynRF (CVPR'23) [30] w/25.66 / 0.07128.68 / 0.04029.13 / 0.06324.26 / 0.08922.37 / 0.10326.19 / 0.05424.96 / 0.04825.89 / 0.0670.45>24
MoSca (arXiv'24) [23] w/25.21 / 0.08332.77 / 0.03328.22 / 0.09024.41 / 0.09223.26 / 0.09228.90 / 0.04223.05 / 0.06026.55 / 0.070400.8
SplineGS (Ours) w/25.62 / 0.06434.10 / 0.03427.83 / 0.04824.60 / 0.07622.67 / 0.12029.04 / 0.03024.47 / 0.04626.90 / 0.0604000.5
STM FreeRoDynRF (CVPR'23) [30]24.27 / 0.10028.71 / 0.04628.85 / 0.06623.25 / 0.10421.81 / 0.12225.58 / 0.06425.20 / 0.05225.38 / 0.0790.45>24
MoSca (arXiv'24) [23]25.43 / 0.08032.62 / 0.03328.29 / 0.08624.40 / 0.09123.27 / 0.09129.01 / 0.04223.23 / 0.05826.61 / 0.069401.0
SplineGS (Ours)25.50 / 0.06833.72 / 0.03128.66 / 0.05625.61 / 0.07124.43 / 0.06828.37 / 0.03224.19 / 0.04727.21 / 0.0534001.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.341, + 0.907, + 0.398 + ], + "angle": 0, + "content": "Table 1. Novel view synthesis evaluation on the NVIDIA dataset (PSNR↑ / LPIPS↓). Red and Blue denote the best and second-best performances, respectively. For Casual-FVS [22], we directly use the results from their paper, as the official code is unavailable. 'w/' denotes the use of camera poses obtained by COLMAP [42] for each SfM-free method. 'Tr' denotes the training time. For MonoNeRF [46], the training time is omitted as it is not a per-scene optimized model. We use an RTX 3090 Ti GPU for each method." + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.413, + 0.229, + 0.547 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.23, + 0.414, + 0.365, + 0.546 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.365, + 0.414, + 0.5, + 0.546 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.5, + 0.414, + 0.635, + 0.546 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.635, + 0.414, + 0.769, + 0.545 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.769, + 0.414, + 0.906, + 0.545 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.254, + 0.548, + 0.743, + 0.562 + ], + "angle": 0, + "content": "Figure 3. Visual comparisons for novel view synthesis on the NVIDIA dataset." + }, + { + "type": "equation", + "bbox": [ + 0.153, + 0.584, + 0.483, + 0.604 + ], + "angle": 0, + "content": "\\[\n\\hat {M} _ {t} (\\boldsymbol {\\varphi} _ {t}) = \\sum_ {i \\in \\mathcal {N}} m _ {i} \\alpha_ {i} \\prod_ {j = 1} ^ {i - 1} (1 - \\alpha_ {j}), \\tag {17}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.607, + 0.483, + 0.637 + ], + "angle": 0, + "content": "where \\( m_{i} = 0 \\) if the \\( i^{\\mathrm{th}} \\) 3D Gaussian is the static 3D Gaussian and \\( m_{i} = 1 \\) otherwise." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.638, + 0.484, + 0.729 + ], + "angle": 0, + "content": "In conjunction, all terms in \\(\\mathcal{L}_{\\mathrm{total}}^{\\mathrm{main}}\\) guide our SplineGS to effectively and efficiently model dynamic 3D scenes from pure monocular videos, achieving more structural details and better temporal consistency than previous works [12, 14, 21-24, 30, 46, 52, 55], without relying on camera parameters obtained from external estimators [42]." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.741, + 0.224, + 0.758 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.766, + 0.483, + 0.796 + ], + "angle": 0, + "content": "Implementation Details. We refer readers to the Supplementary material for the details of our implementation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.797, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Datasets. We evaluate both the quantitative and qualitative performance of novel view and time synthesis on the widely used NVIDIA dataset [56] which features challenging monocular videos. Additionally, we assess novel view synthesis performance on in-the-wild monocular videos from the DAVIS dataset [39] which contains an average of 69.7 frames per video sequence." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.57, + 0.882, + 0.587 + ], + "angle": 0, + "content": "5.1. Comparison with State-of-the-Art Methods" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.599, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Novel View Synthesis. Table 1 presents a quantitative comparison of NVs between our SplineGS and existing SfM-based [5, 12, 14, 15, 18, 21, 22, 24, 33, 46, 48, 49, 52, 55, 58] and SfM-free [23, 30] methods on the NVIDIA dataset [56]. For this comparison, we follow the evaluation configuration in [30]. The results demonstrate that our SplineGS significantly outperforms SOTA methods in both the PSNR and LPIPS [57] metrics. Notably, SplineGS achieves \\(890 \\times\\) and \\(8,000 \\times\\) faster rendering speed compared to RoDynRF [30] and DynNeRF [12], respectively. Ex4DGS [21] and STGS [24], which are designed for multi-view settings, face challenges with inconsistent geometry alignment over time when trained on monocular videos. Fig. 3 shows the qualitative comparison between our SplineGS with the existing methods in [12, 24, 30, 55]. As highlighted by the red boxes, our method yields not only higher rendering quality but also dynamic objects more aligned and closer to the ground truth. In Fig. 4, we show our superior NVs results for in-the-wild monocular videos from the DAVIS dataset [39] compared to the existing methods in [24, 30, 55]." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "26871" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.089, + 0.288, + 0.285 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.287, + 0.089, + 0.483, + 0.285 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.091, + 0.288, + 0.483, + 0.316 + ], + "angle": 0, + "content": "Figure 4. Visual comparisons for novel view synthesis on the DAVIS dataset." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.34, + 0.483, + 0.492 + ], + "angle": 0, + "content": "Compared to [30] that is also COLMAP-free, SplineGS yields considerably more detailed novel views (red boxes) as shown in Fig. 4. For the other methods [24, 55], we observe that COLMAP [42] fails to recover camera parameters and initial point clouds on the DAVIS dataset [39], as also claimed in [30]. On the other hand, our COLMAP-free SplineGS reconstructs accurate camera parameters that are the ones actually used to train [24, 55], which are shown in Fig. 4 for comparison. More results are in the Supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.494, + 0.483, + 0.9 + ], + "angle": 0, + "content": "Novel View and Time Synthesis. To evaluate the capability of SplineGS to model continuous trajectories of moving objects in a scene, we compare the novel view and time synthesis results of SplineGS with those of NeRF-based [12, 30] and 3DGS-based [24, 52, 55] methods. For this evaluation, we follow the dataset sampling strategy in [25], which samples 24 timestamps from the NVIDIA dataset [56]. In addition, to simulate a larger motion, we exclude frames with odd time indices in the training sets. To ensure all test timestamps are not seen during training, and thus, to create a more challenging novel view and time synthesis validation, we exclude frames with even time indices in the test sets. Table 2 and Fig. 5 show the quantitative and qualitative comparisons for this challenging experiment. We observe that the NeRF-based methods, RoDynRF [30] and DynNeRF [12], generate inconsistent artifacts and blurriness for unseen times. Furthermore, the 3DGS-based methods [24, 52, 55] yield even more significant degradation when predicting unseen time indices. In contrast, SplineGS, built upon 3DGS [20] and equipped with our novel spline-based deformation, provides SOTA novel view rendering at unseen intermediate timestamps. Thanks to our MAS, SplineGS naturally and precisely captures the continuous trajectories of moving objects over time, enhancing the temporal consistency of rendered scenes that can be checked in tOF scores [7] of Table 2." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.089, + 0.899, + 0.18 + ], + "angle": 0, + "content": "
MethodsPSNR↑LPIPS↓tOF↓
SfM preprocessingDynNeRF (ICCV'21) [12]23.360.2190.921
4DGS (CVPR'24) [52]17.070.4596.314
D3DGS (CVPR'24) [55]19.630.3433.225
STGS (CVPR'24) [24]15.720.4742.105
SfM-FreeRoDynRF (CVPR'23) [30]21.580.2212.138
SplineGS (Ours)25.920.0980.703
" + }, + { + "type": "table_caption", + "bbox": [ + 0.514, + 0.184, + 0.905, + 0.211 + ], + "angle": 0, + "content": "Table 2. Novel view and time synthesis evaluation on the NVIDIA dataset." + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.226, + 0.62, + 0.322 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.622, + 0.226, + 0.704, + 0.322 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.705, + 0.226, + 0.79, + 0.322 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.791, + 0.226, + 0.846, + 0.322 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.847, + 0.226, + 0.902, + 0.322 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.326, + 0.905, + 0.353 + ], + "angle": 0, + "content": "Figure 5. Visual comparisons for novel view and time synthesis on the NVIDIA dataset." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.364, + 0.907, + 0.53 + ], + "angle": 0, + "content": "To further analyze the SplineGS's ability to model continuous trajectories of dynamic 3D Gaussians, we visualize the projected 2D motion tracking of dynamic objects in pixel space, comparing it with D3DGS [55] and STGS [24], as shown in Fig. 6. We observe that D3DGS [55] and STGS [24] cannot provide reliable motion tracking for moving objects, underscoring their limitations in modeling continuous trajectories of dynamic 3D Gaussians. In contrast, SplineGS provides accurate motion tracking, demonstrating the effectiveness of MAS for deforming dynamic 3D Gaussians. Further analysis is in the Supplementary material." + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.534, + 0.646, + 0.624 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.647, + 0.534, + 0.775, + 0.625 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.776, + 0.534, + 0.905, + 0.625 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.628, + 0.905, + 0.656 + ], + "angle": 0, + "content": "Figure 6. Visual comparisons for motion tracking. We visualize 2D pixel tracks to analyze motions of dynamic 3D Gaussians." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.666, + 0.667, + 0.682 + ], + "angle": 0, + "content": "5.2. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Motion-Adaptive Spline (MAS). To demonstrate the effectiveness of MAS, we replace the MAS model with various deformation models, including an MLP, a grid-based model, polynomial functions of third degree (denoted as 'Poly \\((3^{\\mathrm{rd}})\\)') and tenth degree (denoted as 'Poly \\((10^{\\mathrm{th}})\\)'), and a Bezier curve [9]. For the MLP, grid-based model, and polynomial functions, we apply to them the structures similar to those in prior works, including D3DGS [55], 4DGS [52], and STGS [24], respectively. Additionally, we implement the Bezier curve [9], a commonly used method for curve modeling in computer graphics. Table 3-(a) presents quantitative comparisons of each 3D Gaussian trajectory model, focusing on rendering quality (PSNR, LPIPS) and deformation latency per Gaussian, denoted as \\(g_{\\mathrm{def}}\\). This latency re" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26872" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.088, + 0.483, + 0.189 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.112, + 0.193, + 0.462, + 0.206 + ], + "angle": 0, + "content": "Figure 7. Visual comparisons for MACP ablation study." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.221, + 0.484, + 0.492 + ], + "angle": 0, + "content": "flects the computational time required to estimate the deformation of a single dynamic 3D Gaussian. As shown in Table 3-(a), our MAS model achieves superior rendering quality compared to all other deformation models. Consistent with analyses in previous works [24, 52, 55], we observe that the MLP and grid-based architectures require substantial computational costs for rendering. Among these methods, 'Poly \\((3^{\\mathrm{rd}})\\)', as implemented in [24], demonstrates the best latency. However, fixed-degree polynomial functions have limited flexibility across varying motion complexities which adversely impacts rendering performance. To explore this further, we experiment with 'Poly \\((10^{\\mathrm{th}})\\)' to assess changes in modeling capability. This adjustment, however, leads to noisier optimization and reduced efficiency, as variables under high exponents in 'Poly \\((10^{\\mathrm{th}})\\)' lead to numerical instability. The Bézier curve [9] offers the second-best rendering quality, but its latency remains higher than our MAS due to its recursive nature of computation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.494, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Motion-Adaptive Control Points Pruning (MACP). To assess the effectiveness of our MACP technique for MAS, we compare our full model with MACP against other versions of our model with two fixed numbers of control points \\( N_{c} = 4 \\) and \\( N_{c} = N_{f} \\). As shown in Table 3-(c) and Fig. 7, our SplineGS with MACP achieves a good trade-off between rendering quality and \\( g_{\\mathrm{def}} \\) compared to the ablated models with fixed \\( N_{c} \\). Using \\( N_{c} = 4 \\) for every dynamic 3D Gaussian limits the motion modeling capacity of MAS, resulting in significantly lower metrics and visible artifacts in the dynamic regions. Moreover, an excessive \\( N_{c} = N_{f} \\) decreases the rendering speed of our MAS module and still falls short of the quality achieved by our full model with MACP, potentially due to motion overfitting. Fig. 8 shows the distribution of \\( N_{c} \\) values after optimization with MACP across scenes of varying motion complexities. In Fig. 8-(a), we visualize ' \\( N_{c} \\) Heatmap' that contains the pixel-wise averaged \\( N_{c} \\) values of the dynamic 3D Gaussians needed to render the 2D pixels (red higher, blue lower number of control points). As shown, the simpler motions in these scenes, such as those of the human bodies, can be modeled by the dynamic 3D Gaussians with smaller averages of \\( N_{c} \\) values, whereas the objects with complex and extensive motions (e.g., balloon) require higher averaged \\( N_{c} \\) values. Fig. 8-(b) presents the corresponding histogram of \\( N_{c} \\) values for the scenes' dynamic 3D Gaussians. For sequences with simple motion, such as 'Skating', the trajectories of most" + }, + { + "type": "table_caption", + "bbox": [ + 0.532, + 0.09, + 0.697, + 0.104 + ], + "angle": 0, + "content": "(a) Motion-Adaptive Spline" + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.106, + 0.712, + 0.184 + ], + "angle": 0, + "content": "
PSNR↑LPIPS↓gDef (ns)↓
MLP23.510.125149.41
Grid25.480.09098.89
Poly (3rd)25.140.1111.80
Poly (10th)24.380.1207.71
Bézier27.190.0608.78
Ours27.210.0535.63
" + }, + { + "type": "table_caption", + "bbox": [ + 0.755, + 0.09, + 0.859, + 0.103 + ], + "angle": 0, + "content": "(b) Loss function" + }, + { + "type": "table", + "bbox": [ + 0.731, + 0.106, + 0.886, + 0.184 + ], + "angle": 0, + "content": "
PSNR↑LPIPS↓
w/o Lpc17.490.853
w/o Lgc26.330.067
w/o Ld-pc26.180.066
w/o Lm26.340.088
Ours27.210.053
" + }, + { + "type": "table_caption", + "bbox": [ + 0.58, + 0.187, + 0.842, + 0.2 + ], + "angle": 0, + "content": "(c) Motion-Adaptive Control points Pruning" + }, + { + "type": "table", + "bbox": [ + 0.576, + 0.203, + 0.846, + 0.261 + ], + "angle": 0, + "content": "
PSNR↑LPIPS↓gdef (ns)↓
w/o MACP (Nc=4)26.620.0655.34
w/o MACP (Nc=Nf)27.080.0546.11
Ours27.210.0535.63
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.265, + 0.907, + 0.307 + ], + "angle": 0, + "content": "Table 3. Ablation studies. We ablate our framework and report the average results on the NVIDIA dataset with the same setting as Novel View Synthesis experiment in Sec. 5.1." + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.321, + 0.714, + 0.494 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.716, + 0.321, + 0.902, + 0.494 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.499, + 0.907, + 0.569 + ], + "angle": 0, + "content": "Figure 8. Analysis of MACP's Efficacy. (a) \\( N_{c} \\) Heatmaps as the averaged \\( N_{c} \\) values of dynamic 3D Gaussians and their corresponding rendered frames \\( \\hat{I}_{t} \\) for 'Balloon2' and 'Skating' scenes. (b) Histograms of the number of control points \\( (N_{c}) \\) in percentages (\\%) of dynamic 3D Gaussians in two scenes." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.581, + 0.906, + 0.626 + ], + "angle": 0, + "content": "dynamic 3D Gaussians can be represented using a minimal \\( N_{c} \\), thanks to our MACP. While 'Balloon2' has more evenly distributed \\( N_{c} \\) due to more complex and diverse motion." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.628, + 0.907, + 0.72 + ], + "angle": 0, + "content": "Loss Functions. Table 3-(b) shows the effectiveness of each loss for our overall SplineGS architecture. As noted, no consistent camera parameters can be learned without \\(\\mathcal{L}_{\\mathrm{pc}}\\), drastically impacting the rendering quality of the dynamic 3D Gaussians. Moreover, our \\(\\mathcal{L}_{\\mathrm{gc}}\\), \\(\\mathcal{L}_{\\mathrm{d - pc}}\\) and \\(\\mathcal{L}_{\\mathrm{M}}\\) can considerably impact the overall rendering quality." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.738, + 0.634, + 0.753 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.907, + 0.903 + ], + "angle": 0, + "content": "We present SplineGS, a COLMAP-free dynamic 3DGS framework designed for novel spatio-temporal view synthesis from monocular videos. Leveraging our innovative Motion-Adaptive Spline (MAS) for dynamic motion modeling, SplineGS efficiently renders high-quality novel views from complex in-the-wild videos. The effectiveness of our approach is validated through extensive quantitative and qualitative comparisons, significantly outperforming the existing SOTA methods with very fast rendering speed." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26873" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.251, + 0.108 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.114, + 0.484, + 0.213 + ], + "angle": 0, + "content": "This work was supported by Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korean Government [Ministry of Science and ICT (Information and Communications Technology)] (Project Number: RS-2022-00144444, Project Title: Deep Learning Based Visual Representational Learning and Rendering of Static and Dynamic Scenes, 100%)." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.239, + 0.188, + 0.256 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.264, + 0.483, + 0.292 + ], + "angle": 0, + "content": "[1] Fast dynamic radiance fields with time-aware neural voxels. In SIGGRAPH Asia 2022 Conference Papers, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.294, + 0.484, + 0.361 + ], + "angle": 0, + "content": "[2] J Harold Ahlberg, Edwin Norman Nilson, and Joseph Leonard Walsh. The Theory of Splines and Their Applications: Mathematics in Science and Engineering: A Series of Monographs and Textbooks, Vol. 38. Elsevier, 2016. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.364, + 0.483, + 0.404 + ], + "angle": 0, + "content": "[3] ShahRukh Athar, Zexiang Xu, Kalyan Sunkavalli, Eli Shechtman, and Zhixin Shu. Rignerf: Fully controllable neural 3d portraits. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.407, + 0.483, + 0.461 + ], + "angle": 0, + "content": "[4] Benjamin Attal, Jia-Bin Huang, Christian Richardt, Michael Zollhoefer, Johannes Kopf, Matthew O'Toole, and Changil Kim. Hyperreel: High-fidelity 6-dof video with ray-conditioned sampling. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.463, + 0.484, + 0.532 + ], + "angle": 0, + "content": "[5] Jeongmin Bae, Seoha Kim, Youngsik Yun, Hahyun Lee, Gun Bang, and Youngjung Uh. Per-gaussian embedding-based deformation for deformable 3d gaussian splatting. In European Conference on Computer Vision, pages 321-335. Springer, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.534, + 0.483, + 0.56 + ], + "angle": 0, + "content": "[6] Ang Cao and Justin Johnson. Hexplane: A fast representation for dynamic scenes. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.562, + 0.484, + 0.616 + ], + "angle": 0, + "content": "[7] Mengyu Chu, You Xie, Jonas Mayer, Laura Leal-Taixe, and Nils Thuerey. Learning temporal coherence via self-supervision for gan-based video generation. ACM Transactions on Graphics (TOG), 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.618, + 0.483, + 0.645 + ], + "angle": 0, + "content": "[8] C De Boor. A practical guide to splines. Springer-Verlag google schola, 1978. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.647, + 0.483, + 0.674 + ], + "angle": 0, + "content": "[9] Gerald Farin. Curves and surfaces for CAGD: a practical guide. Elsevier, 2001. 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.676, + 0.483, + 0.731 + ], + "angle": 0, + "content": "[10] Sara Fridovich-Keil, Giacomo Meanti, Frederik Rahbæk Warburg, Benjamin Recht, and Angjoo Kanazawa. K-planes: Explicit radiance fields in space, time, and appearance. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.733, + 0.483, + 0.773 + ], + "angle": 0, + "content": "[11] Yang Fu, Sifei Liu, Amey Kulkarni, Jan Kautz, Alexei A. Efros, and Xiaolong Wang. Colmap-free 3d gaussian splatting. In CVPR, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.776, + 0.483, + 0.815 + ], + "angle": 0, + "content": "[12] Chen Gao, Ayush Saraf, Johannes Kopf, and Jia-Bin Huang. Dynamic view synthesis from dynamic monocular video. In ICCV, 2021. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.818, + 0.483, + 0.845 + ], + "angle": 0, + "content": "[13] Roger A Horn and Charles R Johnson. Matrix analysis. Cambridge university press, 2012. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.847, + 0.483, + 0.899 + ], + "angle": 0, + "content": "[14] Yi-Hua Huang, Yang-Tian Sun, Ziyi Yang, Xiaoyang Lyu, Yan-Pei Cao, and Xiaojuan Qi. Sc-gs: Sparse-controlled gaussian splatting for editable dynamic scenes. CVPR, 2024. 2, 6" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.264, + 0.484, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.161 + ], + "angle": 0, + "content": "[15] Woobin Im, Geonho Cha, Sebin Lee, Jumin Lee, Juhyeong Seon, Dongyoon Wee, and Sung-Eui Yoon. Regularizing dynamic radiance fields with kinematic fields. In European Conference on Computer Vision, pages 312-328. Springer, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.164, + 0.905, + 0.205 + ], + "angle": 0, + "content": "[16] Wei Jiang, Kwang Moo Yi, Golnoosh Samei, Oncel Tuzel, and Anurag Ranjan. Neuman: Neural human radiance field from a single video. In ECCV, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.207, + 0.905, + 0.261 + ], + "angle": 0, + "content": "[17] Yifan Jiang, Peter Hedman, Ben Mildenhall, Dejia Xu, Jonathan T Barron, Zhangyang Wang, and Tianfan Xue. Alignerf: High-fidelity neural radiance fields via alignment-aware training. In CVPR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.263, + 0.905, + 0.331 + ], + "angle": 0, + "content": "[18] Moritz Kappel, Florian Hahlbohm, Timon Scholz, Susana Castillo, Christian Theobalt, Martin Eisemann, Vladislav Golyanik, and Marcus Magnor. D-npc: Dynamic neural point clouds for non-rigid view synthesis from monocular video. arXiv preprint arXiv:2406.10078, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.334, + 0.905, + 0.374 + ], + "angle": 0, + "content": "[19] Nikita Karaev, Ignacio Rocco, Benjamin Graham, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. Co-tracker: It is better to track together. arXiv, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.377, + 0.905, + 0.43 + ], + "angle": 0, + "content": "[20] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 2023. 2, 3, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.433, + 0.905, + 0.473 + ], + "angle": 0, + "content": "[21] Junoh Lee, Chang-Yeon Won, Hyunjun Jung, Inhwan Bae, and Hae-Gon Jeon. Fully explicit dynamic gaussian splatting. In NeurIPS, 2024. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.476, + 0.905, + 0.531 + ], + "angle": 0, + "content": "[22] Yao-Chih Lee, Zhoutong Zhang, Kevin Blackburn-Matzen, Simon Niklaus, Jianming Zhang, Jia-Bin Huang, and Feng Liu. Fast view synthesis of casual videos with soup-of-planes. In ECCV, 2024. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.533, + 0.905, + 0.573 + ], + "angle": 0, + "content": "[23] Jiahui Lei, Yijia Weng, Adam Harley, Leonidas Guibas, and Kostas Daniilidis. Mosca: Dynamic gaussian fusion from casual videos via 4d motion scaffolds. arXiv, 2024. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.576, + 0.905, + 0.616 + ], + "angle": 0, + "content": "[24] Zhan Li, Zhang Chen, Zhong Li, and Yi Xu. Spacetime gaussian feature splatting for real-time dynamic view synthesis. In CVPR. 1, 2, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.618, + 0.905, + 0.658 + ], + "angle": 0, + "content": "[25] Zhengqi Li, Simon Niklaus, Noah Snavely, and Oliver Wang. Neural scene flow fields for space-time view synthesis of dynamic scenes. In CVPR, 2021. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.661, + 0.905, + 0.702 + ], + "angle": 0, + "content": "[26] Zhengqi Li, Qianqian Wang, Forrester Cole, Richard Tucker, and Noah Snavely. Dynibar: Neural dynamic image-based rendering. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.704, + 0.905, + 0.758 + ], + "angle": 0, + "content": "[27] Yiqing Liang, Numair Khan, Zhengqin Li, Thu Nguyen-Phuoc, Douglas Lanman, James Tompkin, and Lei Xiao. Gaufre: Gaussian deformation fields for real-time dynamic novel view synthesis. arXiv, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.761, + 0.905, + 0.8 + ], + "angle": 0, + "content": "[28] Chen-Hsuan Lin, Wei-Chiu Ma, Antonio Torralba, and Simon Lucey. Barf: Bundle-adjusting neural radiance fields. In ICCV, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.803, + 0.905, + 0.843 + ], + "angle": 0, + "content": "[29] Philipp Lindenberger, Paul-Edouard Sarlin, Viktor Larsson, and Marc Pollefeys. Pixel-perfect structure-from-motion with featuremetric refinement. In ICCV, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[30] Yu-Lun Liu, Chen Gao, Andreas Meuleman, Hung-Yu Tseng, Ayush Saraf, Changil Kim, Yung-Yu Chuang, Johannes Kopf, and Jia-Bin Huang. Robust dynamic radiance fields. In CVPR, 2023. 2, 3, 5, 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "26874" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.134 + ], + "angle": 0, + "content": "[31] Jonathon Luiten, Georgios Kopanas, Bastian Leibe, and Deva Ramanan. Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis. In 3DV, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.134, + 0.482, + 0.176 + ], + "angle": 0, + "content": "[32] Quan Meng, Anpei Chen, Haimin Luo, Minye Wu, Hao Su, Lan Xu, Xuming He, and Jingyi Yu. Gnerf: Gan-based neural radiance field without posed camera. In ICCV, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.177, + 0.482, + 0.232 + ], + "angle": 0, + "content": "[33] Xingyu Miao, Yang Bai, Haoran Duan, Fan Wan, Yawen Huang, Yang Long, and Yefeng Zheng. Ctnerf: Cross-time transformer for dynamic neural radiance field from monocular video. Pattern Recognition, 156:110729, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.233, + 0.482, + 0.287 + ], + "angle": 0, + "content": "[34] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.288, + 0.482, + 0.342 + ], + "angle": 0, + "content": "[35] Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In ICCV, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.343, + 0.482, + 0.413 + ], + "angle": 0, + "content": "[36] Keunhong Park, Utkarsh Sinha, Peter Hedman, Jonathan T. Barron, Sofien Bouaziz, Dan B Goldman, Ricardo Martin-Brualla, and Steven M. Seitz. Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields. ACM Trans. Graph., 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.414, + 0.482, + 0.468 + ], + "angle": 0, + "content": "[37] Keunhong Park, Philipp Henzler, Ben Mildenhall, Jonathan T. Barron, and Ricardo Martin-Brualla. Camp: Camera preconditioning for neural radiance fields. ACM Trans. Graph., 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.469, + 0.482, + 0.523 + ], + "angle": 0, + "content": "[38] Luigi Piccinelli, Yung-Hsu Yang, Christos Sakaridis, Mattia Segu, Siyuan Li, Luc Van Gool, and Fisher Yu. Unidepth: Universal monocular metric depth estimation. In CVPR, 2024. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.525, + 0.482, + 0.579 + ], + "angle": 0, + "content": "[39] Jordi Pont-Tuset, Federico Peruzzi, Sergi Caelles, Pablo Arbeláez, Alex Sorkine-Hornung, and Luc Van Gool. The 2017 davis challenge on video object segmentation. arXiv, 2018. 1, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.581, + 0.482, + 0.622 + ], + "angle": 0, + "content": "[40] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-nerf: Neural radiance fields for dynamic scenes. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.623, + 0.482, + 0.678 + ], + "angle": 0, + "content": "[41] Vincent Raoult, Sarah Reid-Anderson, Andreas Ferri, and Jane E Williamson. How reliable is structure from motion (sfm) over time and between observers? a case study using coral reef bommie. Remote Sensing, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.679, + 0.482, + 0.705 + ], + "angle": 0, + "content": "[42] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In CVPR, 2016. 1, 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.707, + 0.482, + 0.761 + ], + "angle": 0, + "content": "[43] Ruizhi Shao, Zerong Zheng, Hanzhang Tu, Boning Liu, Hongwen Zhang, and Yebin Liu. Tensor4d: Efficient neural 4d decomposition for high-fidelity dynamic reconstruction and rendering. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.762, + 0.482, + 0.831 + ], + "angle": 0, + "content": "[44] Liangchen Song, Anpei Chen, Zhong Li, Zhang Chen, Lele Chen, Junsong Yuan, Yi Xu, and Andreas Geiger. Nerfplayer: A streamable dynamic scene representation with decomposed neural radiance fields. IEEE Transactions on Visualization and Computer Graphics, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[45] Carole H. Sudre, Wenqi Li, Tom Vercauteren, Sebastien Ourselin, and M. Jorge Cardoso. Generalised dice overlap as a deep learning loss function for highly unbalanced segmentations. In Deep Learning in Medical Image Analysis and Multimodal Learning for Clinical Decision Support, 2017. 5" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "[46] Fengrui Tian, Shaoyi Du, and Yueqi Duan. MonoN-eRF: Learning a generalizable dynamic radiance field from monocular videos. In ICCV, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.204 + ], + "angle": 0, + "content": "[47] Edgar Tretschk, Ayush Tewari, Vladislav Golyanik, Michael Zollhöfer, Christoph Lassner, and Christian Theobalt. Non-rigid neural radiance fields: Reconstruction and novel view synthesis of a dynamic scene from monocular video. In ICCV, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.206, + 0.905, + 0.248 + ], + "angle": 0, + "content": "[48] Diwen Wan, Ruijie Lu, and Gang Zeng. Superpoint gaussian splatting for real-time high-fidelity dynamic scene reconstruction. arXiv preprint arXiv:2406.03697, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.249, + 0.905, + 0.303 + ], + "angle": 0, + "content": "[49] Qianqian Wang, Vickie Ye, Hang Gao, Jake Austin, Zhengqi Li, and Angjoo Kanazawa. Shape of motion: 4d reconstruction from a single video. arXiv preprint arXiv:2407.13764, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.305, + 0.905, + 0.347 + ], + "angle": 0, + "content": "[50] Zirui Wang, Shangzhe Wu, Weidi Xie, Min Chen, and Victor Adrian Prisacariu. Nerf: Neural radiance fields without known camera parameters. CoRR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.348, + 0.905, + 0.403 + ], + "angle": 0, + "content": "[51] Chung-Yi Weng, Brian Curless, Pratul P. Srinivasan, Jonathan T. Barron, and Ira Kemelmacher-Shlizerman. HumanNeRF: Free-viewpoint rendering of moving people from monocular video. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.405, + 0.905, + 0.459 + ], + "angle": 0, + "content": "[52] Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Xinggang Wang. 4d gaussian splatting for real-time dynamic scene rendering. In CVPR. 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.461, + 0.905, + 0.516 + ], + "angle": 0, + "content": "[53] Gengshan Yang, Minh Vo, Neverova Natalia, Deva Ramanan, Vedaldi Andrea, and Joo Hanbyul. Banmo: Building animatable 3d neural models from many casual videos. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.518, + 0.905, + 0.559 + ], + "angle": 0, + "content": "[54] Jinyu Yang, Mingqi Gao, Zhe Li, Shang Gao, Fangjing Wang, and Feng Zheng. Track anything: Segment anything meets videos. arXiv, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.561, + 0.905, + 0.615 + ], + "angle": 0, + "content": "[55] Ziyi Yang, Xinyu Gao, Wen Zhou, Shaohui Jiao, Yuqing Zhang, and Xiaogang Jin. Deformable 3d gaussians for high-fidelity monocular dynamic scene reconstruction. In CVPR, 2024. 1, 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.617, + 0.905, + 0.672 + ], + "angle": 0, + "content": "[56] Jae Shin Yoon, Kihwan Kim, Orazio Gallo, Hyun Soo Park, and Jan Kautz. Novel view synthesis of dynamic scenes with globally coherent depths from a monocular camera. In CVPR, 2020. 1, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.674, + 0.905, + 0.715 + ], + "angle": 0, + "content": "[57] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.717, + 0.905, + 0.785 + ], + "angle": 0, + "content": "[58] Kaichen Zhou, Jia-Xing Zhong, Sangyun Shin, Kai Lu, Yiyuan Yang, Andrew Markham, and Niki Trigoni. Dynpoint: Dynamic neural point for view synthesis. Advances in Neural Information Processing Systems, 36:69532-69545, 2023. 6" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.785 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26875" + } + ] +] \ No newline at end of file diff --git a/2025/SplineGS_ Robust Motion-Adaptive Spline for Real-Time Dynamic 3D Gaussians from Monocular Video/c20079e6-f30b-47d1-a98c-7f27a87d2799_origin.pdf b/2025/SplineGS_ Robust Motion-Adaptive Spline for Real-Time Dynamic 3D Gaussians from Monocular Video/c20079e6-f30b-47d1-a98c-7f27a87d2799_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..0844f3716c0941b9e5f650cd96181112b4acf7ee --- /dev/null +++ b/2025/SplineGS_ Robust Motion-Adaptive Spline for Real-Time Dynamic 3D Gaussians from Monocular Video/c20079e6-f30b-47d1-a98c-7f27a87d2799_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c19869a90ab0ec63d6902c500ff57747efe5993439a92d5d86006eb51a540a9f +size 5696738 diff --git a/2025/SplineGS_ Robust Motion-Adaptive Spline for Real-Time Dynamic 3D Gaussians from Monocular Video/full.md b/2025/SplineGS_ Robust Motion-Adaptive Spline for Real-Time Dynamic 3D Gaussians from Monocular Video/full.md new file mode 100644 index 0000000000000000000000000000000000000000..5e0387e64e3dec6fce2c76bcfacc1e90d257f994 --- /dev/null +++ b/2025/SplineGS_ Robust Motion-Adaptive Spline for Real-Time Dynamic 3D Gaussians from Monocular Video/full.md @@ -0,0 +1,386 @@ +# SplineGS: Robust Motion-Adaptive Spline for Real-Time Dynamic 3D Gaussians from Monocular Video + +Jongmin Park1* Minh-Quan Viet Bui1* Juan Luis Gonzalez Bello1 Jaeho Moon1 +Jihyong Oh2† Munchurl Kim1† + +\(^{1}\)Korea Advanced Institute of Science and Technology (KAIST) \(^{2}\)Chung-Ang University \(\{\mathrm{jm.park, bvmquan, juanluisgb, jaehomoon, mkimee}\} @kaist.ac.kr jihyongoh@cau.ac.kr https://kaist-viclab.github.io/splinegs-site/ + +![](images/2d814176e456e030dc547810d5c45437fb60e893cbbfc71dc4c73b79c3961383.jpg) + +horsejump-low, $t = 20$ (Ours) +train, $t = 37$ (Ours) +![](images/b6a70072525f4339034e7b7eac3563aa9a2ea4bacb05fd1919f16215e7522bd3.jpg) +(a) Visual comparison for novel view synthesis on the DAVIS dataset + +![](images/d9352c8f024540b3c8bf6e9266b5858c9053440c7132ec15b88a96de2b7587f1.jpg) +STGS [24] + +![](images/17e67f379b6de6572a4764d79a54bd0f29a90694ab22571c5dc19c1cf204e5c2.jpg) +D3DGS [55] + +![](images/9dd5a5d3ec9936b0ecd139efda7215dd2aade2231be5f563f77f6cbb3e4d8850.jpg) + +![](images/54aa8c97d8f5c60eae6a136c42e4d286c5ddf29d725cb583ba9e6547e452f9bd.jpg) + +![](images/b3fd31d14da4013979908b2ad4f7ef4da7c19e4b5fa7739d9a8030a35b30c842.jpg) +RoDynRF [30] +STGS [24] + +![](images/8f48f9a6140fcffe04e1e9a96352139ff158accf5eed32af915b36f17b9e3506.jpg) +SplineGS (Ours) +D3DGS [55] + +![](images/fa809b097842eada0d1a61b00139df7b8e5367c2fa50bfc82606e5e685f5a7af.jpg) +RoDynRF [30] + +![](images/7e260132ffbf6ac6a9c233c420f0aa50b81213865d280c623e8a7939c24a9302.jpg) +SplineGS (Ours) + +Figure 1. Our SplineGS achieves state-of-the-art rendering quality with fast rendering speed for novel spatio-temporal view synthesis from monocular videos without relying on pre-computed camera parameters. (a) We use our predicted camera parameters for [24, 55] since COLMAP [42] is unable to provide reasonable camera parameters for most scenes in the DAVIS dataset [39]. (b) SplineGS achieves the best rendering quality while maintaining a rendering speed of 400 FPS on the NVIDIA dataset [56]. +![](images/0ff377d5310212799fa39bf4e7c5a8018446762d1f41155fc3086c69d69c5d05.jpg) +(b) Performance gain on the NVIDIA dataset + +# Abstract + +Synthesizing novel views from in-the-wild monocular videos is challenging due to scene dynamics and the lack of multi-view cues. To address this, we propose SplineGS, a COLMAP-free dynamic 3D Gaussian Splatting (3DGS) framework for high-quality reconstruction and fast rendering from monocular videos. At its core is a novel Motion-Adaptive Spline (MAS) method, which represents continuous dynamic 3D Gaussian trajectories using cubic Hermite splines with a small number of control points. For MAS, we introduce a Motion-Adaptive Control points Pruning (MACP) method to model the deformation of each dynamic 3D Gaussian across varying motions, progressively pruning control points while maintaining dynamic modeling integrity. Additionally, we present a joint optimization strat + +egy for camera parameter estimation and 3D Gaussian attributes, leveraging photometric and geometric consistency. This eliminates the need for Structure-from-Motion preprocessing and enhances SplineGS's robustness in real-world conditions. Experiments show that SplineGS significantly outperforms state-of-the-art methods in novel view synthesis quality for dynamic scenes from monocular videos, achieving thousands times faster rendering speed. + +# 1. Introduction + +Novel View Synthesis (NVS) is fundamental to 3D vision, supporting applications like virtual reality (VR), augmented reality (AR), and film production. NFS aims to generate + +images from any viewpoint in a scene, requiring accurate reconstruction from multiple 2D images. NeRF [34] has advanced the field of NVs by utilizing learned implicit functions to represent static scenes, though it requires considerable training and rendering time. Recently, 3D Gaussian Splatting (3DGS) [20] has revolutionized this process by replacing implicit volumetric rendering with differentiable rasterization of 3D Gaussians, enabling real-time rendering and providing a more explicit scene representation. + +For NVs of dynamic scenes, prior works have extended the static scene representations [20, 34] by incorporating deformation models in canonical space using implicit representations [12, 30, 35, 36, 55], grid-based models that decompose the 4D space-time domain into multiple 2D planes [4, 6, 10, 43, 52], and polynomial trajectories [24]. However, these methods face several challenges: implicit representations significantly increase computational overhead and reduce rendering speed [6, 52]; grid-based models struggle to fully capture the dynamic nature of scene structures, hindering their ability to model fine details accurately [24]; and although polynomial trajectories improve efficiency, their fixed degree restricts flexibility for representing complex motions. To the best of our knowledge, none of the dynamic 3DGS-based methods provide experimental evidence to reliably support their novel spatio-temporal view synthesis capabilities for rendering unseen intermediate time indices. Additionally, most existing methods for dynamic NVs rely heavily on external camera parameter estimators like COLMAP [42], which often produce imprecise results for challenging in-the-wild monocular videos. + +To address these issues for modeling scene dynamics, we exploit a spline-based model to dynamic 3D Gaussian trajectories, inspired by classic 3D-curve modeling in computer graphics [9]. Splines are widely used in geometric modeling for graphical applications, providing smooth and continuous representations of complex shapes with a minimal number of control points [2, 8]. This efficiency makes them ideal for modeling intricate geometric structures while maintaining both flexibility and precision. In this paper, we propose SplineGS, a framework for high-quality dynamic scene reconstruction and real-time neural rendering from inthe-wild monocular videos without relying on external camera estimators like COLMAP [42]. SplineGS introduces Motion-Adaptive Spline (MAS), based on cubic Hermite splines [2, 8], to effectively represent dynamic 3D Gaussian deformations. Our MAS consists of piecewise cubic functions defined by control points that dictate each segment's curvature and direction. Each control point is adjustable and optimized as a learnable parameter, enabling faster and more precise modeling of complex motion trajectories. Additionally, to adaptively model the trajectory of each dynamic 3D Gaussian based on motion complexity during training, we introduce a Motion-Adaptive Control + +points Pruning (MACP) method that adjusts the number of control points to improve rendering quality and efficiency. Furthermore, we incorporate a camera parameter estimation method jointly optimized with 3D Gaussian attributes and MAS, leveraging photometric and geometric consistency, thus eliminating the need for external estimators. Experiments show that SplineGS significantly outperforms state-of-the-art (SOTA) dynamic NVs methods both qualitatively and quantitatively, offering faster rendering speed, as shown in Fig. 1. Our contributions are as follows: + +- We propose a novel Spline-based dynamic 3D Gaussian Splatting framework, SplineGS, which is (i) COLMAP-free, (ii) very fast and (iii) of high quality in reconstructing the dynamic scenes from in-the-wild monocular videos; +- A novel Motion-Adaptive Spline (MAS) is introduced, which can accurately and effectively represent the continuous trajectory of each dynamic 3D Gaussian; +- A Motion-Adaptive Control points Pruning (MACP) method is presented, which can efficiently adjust the number of control points for each spline function, optimizing both rendering quality and the efficiency of MAS. + +# 2. Related Work + +Dynamic NeRF. Recent advancements in video view synthesis have extended the static NeRF model [34] to represent scene dynamics. These include dynamic NeRFs using scene flow-based frameworks [12, 25, 26], deformation estimation in canonical fields [1, 3, 16, 30, 35, 36, 40, 44, 47, 51, 53], and 4D grid-based spatio-temporal radiance fields [4, 6, 10, 43]. Techniques such as NSFF [25], DynNeRF [12], and DynIBaR [26] combine time-independent and time-dependent radiance fields to synthesize novel spatiotemporal perspectives from monocular videos. Despite these advancements, current dynamic NeRFs still fall short of recent dynamic 3D Gaussian Splatting (3DGS) techniques in rendering quality and efficiency. + +Dynamic 3DGS. The improvements in rendering quality and speed achieved by 3DGS [20] have inspired further studies [14, 24, 27, 31, 52, 55] to extend the static 3DGS framework to dynamic scenes by enabling the deformation of 3D Gaussian attributes. The pioneering work on dynamic 3DGS [31] introduces time-dependent offsets for the positions and rotations of dynamic 3D Gaussians via an MLP; however, this approach slows down the rendering. D3DGS [55] builds on this concept with an annealing smoothing training mechanism to improve temporal smoothness and rendering quality. 4DGS [52] replaces the MLP network with a grid-based structure to boost efficiency, though this change requires quality trade-offs due to the resolution limits of grid-based methods. In contrast, SC-GS [14] combines an MLP deformation network with sparse spatial deformation, reducing the computational + +cost of MLP while maintaining quality. STGS [24] introduces polynomial trajectories for motion modeling, improving speed and quality over implicit representations. CasualFVS [22] warps dynamic content from neighboring frames, and MoSca [23] proposes a graph-based motion modeling approach to handle sparse control deformation. Very recently, a concurrent work [21] proposes modeling dynamic 3D Gaussian trajectories using polynomial interpolation between fixed keyframes for NVS from multi-view videos with COLMAP [42] assistance. Unlike [21], our SplineGS adaptively optimizes the deformation of dynamic 3D Gaussians, accounting for varying motion degrees and types in in-the-wild monocular videos, without requiring preprocessed camera parameters (COLMAP-free approach). Neural Rendering without SfM Preprocessing. Accurate camera parameters, including extrinsic and intrinsic, are essential for neural rendering approaches to capture fine details [17, 28]. However, in real-world settings, camera parameters derived from Structure-from-Motion (SfM) algorithms such as COLMAP [42] often exhibit pixel-level inaccuracies [29, 41], compromising the structural details of rendered scenes [30]. To address this, several NeRF methods [28, 32, 37, 50] jointly optimize NeRF architectures and camera parameters. Recently, a local-to-global training approach [11] is introduced to optimize both camera parameters and 3D Gaussians. However, these methods are limited to static scenes. For dynamic scene reconstruction, RoDynRF [30] and MoSca [23] use motion masks to gather multi-view cues from static regions, allowing robust rendering without pre-computed camera parameters. Our SplineGS is also COLMAP-free, significantly outperforming RoDynRF [30] and MoSca [23] in rendering quality and efficiency, enabled by our novel spline-based architecture. + +# 3. Preliminary: 3D Gaussian Splitting + +3DGS [20] represents the radiance field of a scene using anisotropic 3D Gaussians, each of which is formulated as + +$$ +G (\boldsymbol {x}) = e ^ {- \frac {1}{2} (\boldsymbol {x} - \boldsymbol {\mu}) ^ {\top} \boldsymbol {\Sigma} ^ {- 1} (\boldsymbol {x} - \boldsymbol {\mu})} \tag {1} +$$ + +where $\pmb{x} \in \mathbb{R}^3$ denotes a 3D position, and $\pmb{\mu} \in \mathbb{R}^3$ and $\pmb{\Sigma} \in \mathbb{R}^{3 \times 3}$ represent the mean (center) and the covariance matrix of the 3D Gaussian, respectively. To ensure that $\pmb{\Sigma}$ is positive semi-definite and contains physical meaning, it is decomposed into a diagonal scaling matrix $S \in \mathbb{R}^{3 \times 3}$ of a scale vector $s \in \mathbb{R}^3$ and a rotation matrix $R \in \mathbb{R}^{3 \times 3}$ as $\pmb{\Sigma} = R\pmb{S}\pmb{S}^\top\pmb{R}^\top$ , where $\pmb{R}$ is parameterized by a learnable unit-length quaternion $q \in \mathbb{R}^4$ . In addition, each 3D Gaussian is parameterized by an opacity $\sigma \in \mathbb{R}$ and a color $c \in \mathbb{R}^3$ . To render the color of each pixel, the color and the opacity of each 3D Gaussian are computed using Eq. 1, and the rendered color $C$ is computed by the alpha-blending of the $\mathcal{N}$ ordered 3D Gaussians overlapping the pixel as + +$$ +\boldsymbol {C} = \sum_ {i \in \mathcal {N}} \boldsymbol {c} _ {i} \alpha_ {i} \prod_ {j = 1} ^ {i - 1} (1 - \alpha_ {j}), \tag {2} +$$ + +where $c_{i}$ is the color of the $i^{\mathrm{th}}$ 3D Gaussian and $\alpha_{i}$ is a density of the $i^{\mathrm{th}}$ 3D Gaussian which is given by evaluating the 2D covariance $\pmb{\Sigma}' \in \mathbb{R}^{2 \times 2}$ . Here, $\pmb{\Sigma}'$ is formulated as $\pmb{\Sigma}' = J\pmb{W}\pmb{\Sigma}\pmb{W}^\top J^\top$ , where $J$ is the Jacobian of the affine approximation of the projective transformation and $W$ is a viewing transformation matrix. Similar to prior works [23, 27], we extend 3DGS [20] to a union of static 3D Gaussians $\{G_i^{\mathrm{st}} | i = 1, 2, \dots, n^{\mathrm{st}}\}$ and dynamic 3D Gaussians $\{G_i^{\mathrm{dy}} | i = 1, 2, \dots, n^{\mathrm{dy}}\}$ to represent static backgrounds and moving objects, respectively, in dynamic scenes. + +# 4. Proposed Method: SplineGS + +# 4.1. Overview of SplineGS + +We describe the overall architecture of SplineGS in Fig. 2. Given a monocular video $\{I_t|t = 0,1,\dots,N_f - 1\}$ where $N_{f}$ is the total number of frames, SplineGS is designed to synthesize high-quality novel spatio-temporal views with fast rendering speed, and to estimate the camera parameters, including the extrinsic $[\hat{R}_t|\hat{T}_t]\in \mathbb{R}^{3\times 4}$ for each time $t$ , and the shared intrinsic $\hat{\pmb{K}}\in \mathbb{R}^{3\times 3}$ . As shown in Fig. 2, to stabilize joint optimization of the 3D Gaussian attributes and camera parameter estimation, we adopt a two-stage optimization process consisting of a warm-up stage and a main training stage for our SplineGS architecture. In the warm-up stage, we coarsely optimize the camera parameters $[\hat{R}_t|\hat{T}_t]$ and $\hat{\pmb{K}}$ using photometric and geometric consistency. In the main training stage, we initialize the 3D Gaussians based on the estimated camera poses and jointly optimize the 3D Gaussian attributes with the camera parameters. Specifically, for each dynamic 3D Gaussian, we propose a novel spline-based deformation modeling method, called Motion-Adaptive Spline (MAS), which utilizes a cubic Hermite spline function [2, 8] to accurately model the continuous trajectory with a small number of control points. For MAS, we introduce a Motion-Adaptive Control points Pruning (MACP) method that effectively adjusts the number of control points for each dynamic 3D Gaussian, considering the object's motion types and degrees. + +In the following sections, we detail the process of our MAS method in Sec. 4.2, followed by the camera parameter estimation process in Sec. 4.3. Finally, we describe the overall optimization process in Sec. 4.4. + +# 4.2. Motion-Adaptive Spline for 3D Gaussians + +To represent the continuous trajectory of each dynamic 3D Gaussian for moving objects over time, we propose MAS which modifies the mean parameter to a set of learnable control points. For this, we use the cubic Hermite spline function [2, 8] to model the mean of each dynamic 3D Gaussian at time $t$ as + +$$ +\boldsymbol {\mu} (t) = S (t, \mathbf {P}), \tag {3} +$$ + +![](images/a6d6d9f6497404325d227490123d5c8dce1c9fbb7b4e83b61409e0591cf28809.jpg) +Figure 2. Overview of SplineGS. Our SplineGS leverages spline-based functions to model the deformation of dynamic 3D Gaussians with a novel Motion-Adaptive Spline (MAS) architecture. It is composed of sets of learnable control points based on a cubic Hermite spline function [2, 8] to accurately model the trajectory of each dynamic 3D Gaussian and to achieve faster rendering speed. To avoid any preprocessing of camera parameters, i.e., COLMAP-free, we adopt a two-stage optimization: warm-up and main training stages. + +where $\mathbf{P} = \{\mathbf{p}_k|\mathbf{p}_k\in \mathbb{R}^3,k\in 0,\dots ,N_c - 1\}$ is a set of $N_{c}$ learnable control points, serving as an additional attribute for each dynamic 3D Gaussian, and $S(t,\mathbf{P})$ is defined as + +$$ +\begin{array}{l} S (t, \mathbf {P}) = \left(2 t _ {r} ^ {3} - 3 t _ {r} ^ {2} + 1\right) \mathbf {p} _ {\lfloor t _ {s} \rfloor} + \left(t _ {r} ^ {3} - 2 t _ {r} ^ {2} + t _ {r}\right) \mathbf {m} _ {\lfloor t _ {s} \rfloor} \\ + \left(- 2 t _ {r} ^ {3} + 3 t _ {r} ^ {2}\right) \mathbf {p} _ {[ t _ {s} ] + 1} + \left(t _ {r} ^ {3} - t _ {r} ^ {2}\right) \mathbf {m} _ {[ t _ {s} ] + 1}, \tag {4} \\ \end{array} +$$ + +$$ +t _ {r} = t _ {s} - \lfloor t _ {s} \rfloor , \quad t _ {s} = t _ {n} (N _ {c} - 1), \quad t _ {n} = t / (N _ {f} - 1), +$$ + +where $\mathbf{m}_k$ is the approximated tangent of the control point $\mathbf{p}_k$ , formulated as $\mathbf{m}_k = (\mathbf{p}_{k + 1} - \mathbf{p}_{k - 1}) / 2$ . Note that $S(t,\mathbf{P})$ indicates the piecewise cubic function to represent the whole continuous trajectory of each dynamic 3D Gaussian, and $N_{c}$ can be different from $N_{f}$ . The optimal $N_{c}$ can be estimated by Motion-Adaptive Control points Pruning (MACP) in the following section. + +Control Points Initialization. To stably optimize $S(t, \mathbf{P})$ , it is essential to initialize the control points with appropriate geometric considerations that ensure temporal consistency. For this, we leverage the long-range 2D tracking [19] and the metric depth [38] priors. Let $\mathcal{T} = \{\varphi_t^{\mathrm{tr}} | \varphi_t^{\mathrm{tr}} \in \mathbb{R}^2, t = 0, \dots, N_f - 1\}$ be a 2D track, $\pi_K(\cdot)$ be a projection function from the camera space to the image space with the camera intrinsic $K$ , and $\varphi_t^{\mathrm{tr}}$ be a pixel coordinate corresponding to the 2D track $\mathcal{T}$ at time $t$ . We unproject each 2D track $\mathcal{T}$ to the world space to compute the 3D track aided by the frame's metric depth $d_t$ and camera extrinsic $[\hat{R}_t | \hat{T}_t]$ at time $t$ . The unprojection function $\mathcal{W}_t(\cdot)$ from the image space to the world space is formulated as + +$$ +\mathcal {W} _ {t} \left(\boldsymbol {\varphi} _ {t} ^ {\mathrm {t r}}\right) = \hat {\boldsymbol {R}} _ {t} ^ {\top} \pi_ {\hat {\boldsymbol {K}}} ^ {- 1} \left(\boldsymbol {\varphi} _ {t} ^ {\mathrm {t r}}, \boldsymbol {d} _ {t} \left(\boldsymbol {\varphi} _ {t} ^ {\mathrm {t r}}\right)\right) - \hat {\boldsymbol {R}} _ {t} ^ {\top} \hat {\boldsymbol {T}} _ {t}. \tag {5} +$$ + +It should be noted that we estimate $[\hat{\pmb{R}}_t|\hat{\pmb{T}}_t]$ and $\hat{\pmb{K}}$ from a sequence of frames only, without using any given ground truth values, as mentioned in Sec. 4.3. Then, we initialize the per-Gaussian control points set $\mathbf{P}$ , by a least-square (LS) approximation, such that the spline-described curve fits the initial curve by the tracker, given by + +$$ +\min _ {\mathbf {P}} \sum_ {t = 0} ^ {N _ {f} - 1} \left\| \mathcal {W} _ {t} \left(\varphi_ {t} ^ {\mathrm {t r}}\right) - S (t, \mathbf {P}) \right\| _ {2} ^ {2}. \tag {6} +$$ + +Motion-Adaptive Control Points Pruning (MACP). An excessive number of control points may cause over-fitting and decrease the processing speed of our MAS module, leading to poorer rendering qualities with reduced rendering speed (see Table 3). Furthermore, as each scene exhibits various types and degrees of motion for moving objects, the number of control points required for each dynamic 3D Gaussian trajectory should be adaptively adjusted to accommodate the scene dynamics. To achieve this, we propose the MACP method for MAS, which can generate sparser control points while ensuring no dynamic modeling degradation. MACP is designed on top of 3D Gaussian densification [20], but focuses on optimizing the number of control points in MAS. Our MACP computes a new spline function $S(t,\mathbf{P}^{\prime})$ after every 3D Gaussian densification step, where $\mathbf{P}^{\prime} = \{\mathbf{p}_{l}^{\prime}|\mathbf{p}_{l}^{\prime}\in \mathbb{R}^{3}, l = 0,\dots ,N_{c} - 2\}$ is a set of $N_{c} - 1$ control points, which contains one-fewer control points than the current set $\mathbf{P}$ of control points. We compute the LS approximation to find the reduced set $\mathbf{P}^{\prime}$ of control points as + +$$ +\min _ {\mathbf {P} ^ {\prime}} \sum_ {t = 0} ^ {N _ {f} - 1} \left\| S (t, \mathbf {P}) - S (t, \mathbf {P} ^ {\prime}) \right\| _ {2} ^ {2}. \tag {7} +$$ + +Then, the updated optimal set $\mathbf{P}$ of control points is assigned with the values of $\mathbf{P}'$ if the error $E$ between $S(t,\mathbf{P})$ and $S(t,\mathbf{P}')$ is smaller than a threshold $\epsilon$ , as given by + +$$ +\mathbf {P} = \left\{ \begin{array}{l l} \mathbf {P} ^ {\prime}, & \text {i f} E < \epsilon \\ \mathbf {P}, & \text {o t h e r w i s e} \end{array} , \right. \text {w h e r e} \tag {8} +$$ + +$$ +E = \frac {1}{N _ {f}} \sum_ {t = 0} ^ {N _ {f} - 1} \| \pi_ {\hat {\mathbf {K}}} (\hat {\mathbf {R}} _ {t} S (t, \mathbf {P}) + \hat {\mathbf {T}} _ {t}) - \pi_ {\hat {\mathbf {K}}} (\hat {\mathbf {R}} _ {t} S (t, \mathbf {P} ^ {\prime}) + \hat {\mathbf {T}} _ {t}) \| _ {2} ^ {2}. +$$ + +By following MACP, each dynamic 3D Gaussian is allowed to have a different number of control points. Therefore, the MACP can guide our MAS to have a minimal number of control points when modeling simple motion, while having more control points for more complex motions (see Fig. 8). + +# 4.3. Camera Parameter Estimation + +The traditional SfM methods, such as COLMAP [42], fail to reliably estimate camera parameters in dynamic scenes from in-the-wild monocular videos [30]. For this reason, we propose to estimate camera parameters for joint optimization with the 3D Gaussian attributes. For each frame at time $t$ , we predict a rotation matrix $\hat{R}_t \in \mathbb{R}^{3 \times 3}$ and a translation vector $\hat{T}_t \in \mathbb{R}^3$ , representing the extrinsic parameters of a monocular camera, using a shallow MLP $F_{\theta}$ as + +$$ +\left[ \boldsymbol {\hat {R}} _ {t} \mid \boldsymbol {\hat {T}} _ {t} \right] = F _ {\theta} (\gamma (t)), \tag {9} +$$ + +where $\gamma (\cdot)$ is a positional encoding [34]. We also predict the focal length $\hat{f}$ of the camera intrinsic $\hat{K}$ as a learnable parameter that is shared across all frames in the monocular video. To accurately optimize the camera parameters, we enforce two types of consistency—photometric and geometric—for the static background between random reference cameras $[\hat{R}_{t_{\mathrm{ref}}}|\hat{T}_{t_{\mathrm{ref}}}]$ and the target camera $[\hat{R}_t|\hat{T}_t]$ , encouraging alignment both visually and structurally. + +Photometric Consistency. Given the pre-computed metric depth [38], the camera intrinsics and extrinsics under optimization will converge as long as the projected reference frame's color $I_{t_{\mathrm{ref}}}(\varphi_{t \to t_{\mathrm{ref}}})$ is aligned to the target frame's color $I_t(\varphi_t)$ . $\varphi_{t \to t_{\mathrm{ref}}}$ is the reference pixel coordinate corresponding to the target frame's pixel coordinate $\varphi_t$ as + +$$ +\varphi_ {t \rightarrow t _ {\text {r e f}}} = \pi_ {\hat {\mathbf {K}}} \left(\hat {\mathbf {R}} _ {t _ {\text {r e f}}} \left(\hat {\mathbf {R}} _ {t} ^ {\top} \pi_ {\hat {\mathbf {K}}} ^ {- 1} \left(\varphi_ {t}, \boldsymbol {d} _ {t} (\varphi_ {t})\right) - \hat {\mathbf {R}} _ {t} ^ {\top} \hat {\boldsymbol {T}} _ {t}\right) + \hat {\boldsymbol {T}} _ {t _ {\text {r e f}}}\right). \tag {10} +$$ + +We refer to such projection alignment as photometric consistency, which is encouraged by the loss $\mathcal{L}_{\mathrm{pc}}$ given by + +$$ +\mathcal {L} _ {\mathrm {p c}} = \sum_ {\boldsymbol {\varphi} _ {t}} \left\| \boldsymbol {M} _ {t, t _ {\text {r e f}}} \left(\boldsymbol {\varphi} _ {t}\right) \odot \left(\boldsymbol {I} _ {t} \left(\boldsymbol {\varphi} _ {t}\right) - \boldsymbol {I} _ {t _ {\text {r e f}}} \left(\boldsymbol {\varphi} _ {t \rightarrow t _ {\text {r e f}}}\right)\right)\right\| _ {2} ^ {2}, \tag {11} +$$ + +where $\odot$ is the Hadamard product [13], and $M_{t,t_{\mathrm{ref}}}$ is a union motion mask that excludes dynamic objects in $I_{t}$ and $I_{t_{\mathrm{ref}}}$ for removing the inconsistencies due to dynamic regions, which is given by + +$$ +\boldsymbol {M} _ {t, t _ {\text {r e f}}} \left(\varphi_ {t}\right) = \boldsymbol {M} _ {t} \left(\varphi_ {t}\right) \boldsymbol {M} _ {t _ {\text {r e f}}} \left(\varphi_ {t \rightarrow t _ {\text {r e f}}}\right), \tag {12} +$$ + +where $M_{t}$ and $M_{t_{\mathrm{ref}}}$ are pre-computed motion masks [54] of $\pmb{I}_t$ and $I_{t_{\mathrm{ref}}}$ , respectively. Note that motion mask's value is 0 for static regions, and 1 otherwise. + +Geometric Consistency. Along with the photometric consistency, we compute the geometric consistency of unprojected pixels in 3D space to make our optimization more stable. The geometric consistency loss is formulated as + +$$ +\mathcal {L} _ {\mathrm {g c}} = \sum_ {\boldsymbol {\varphi} _ {t}} \| M _ {t, t _ {\text {r e f}}} (\boldsymbol {\varphi} _ {t}) \odot \left(\mathcal {W} _ {t} (\boldsymbol {\varphi} _ {t}) - \mathcal {W} _ {t _ {\text {r e f}}} \left(\boldsymbol {\varphi} _ {t \rightarrow t _ {\text {r e f}}}\right)\right) \| _ {2} ^ {2}, \tag {13} +$$ + +where $\mathcal{W}_t(\varphi_t)$ and $\mathcal{W}_{t_{\mathrm{ref}}}(\varphi_{t\to t_{\mathrm{ref}}})$ are the corresponding 3D locations of $\varphi_t$ and $\varphi_{t\rightarrow t_{\mathrm{ref}}}$ , respectively (see Eq. 5). + +# 4.4. Optimization + +To stabilize joint training of the MAS and the camera parameter estimation, we adopt a two-stage optimization process consisting of the warm-up stage and the main training stage. In the warm-up stage, we optimize only the camera parameters using the photometric and geometric consistency. The total loss in the warm-up stage is given by + +$$ +\mathcal {L} _ {\text {t o t a l}} ^ {\text {w a r m}} = \lambda_ {\mathrm {p c}} \mathcal {L} _ {\mathrm {p c}} + \lambda_ {\mathrm {g c}} \mathcal {L} _ {\mathrm {g c}}. \tag {14} +$$ + +After the warm-up stage, we obtain the coarsely predicted camera intrinsic $\hat{\pmb{K}}$ and the set of extrinsics $[\hat{\pmb{R}}|\hat{\pmb{T}}]$ for all frames, which are then used to initialize the set $\mathbf{P}$ of control points for each dynamic 3D Gaussian, as described in Sec. 4.2. In the main training stage, we jointly optimize the static and dynamic 3D Gaussians along with the camera parameter estimation based on the total loss function as + +$$ +\begin{array}{l} \mathcal {L} _ {\text {t o t a l}} ^ {\text {m a i n}} = \lambda_ {\mathrm {r g b}} \mathcal {L} _ {\mathrm {r g b}} + \lambda_ {\mathrm {d}} \mathcal {L} _ {\mathrm {d}} + \lambda_ {\mathrm {M}} \mathcal {L} _ {\mathrm {M}} \\ + \lambda_ {\mathrm {p c}} \mathcal {L} _ {\mathrm {p c}} + \lambda_ {\mathrm {d - p c}} \mathcal {L} _ {\mathrm {d - p c}} + \lambda_ {\mathrm {g c}} \mathcal {L} _ {\mathrm {g c}}, \\ \end{array} +$$ + +(15) + +where $\mathcal{L}_{\mathrm{rgb}}$ and $\mathcal{L}_{\mathrm{d}}$ are the L1 losses between the rendered frame and GT frame, and between the rendered depth and GT depth, respectively. In the main training stage, we compute an additional photometric consistency loss $\mathcal{L}_{\mathrm{d - pc}}$ that utilizes the rendered depth $\hat{\pmb{d}}$ of the 3D Gaussians instead of the metric depth [38] prior $\pmb{d}$ as $\hat{\pmb{d}}$ allows the estimated 3D Gaussian geometry to guide the joint optimization of the camera parameters and 3D Gaussian attributes. + +In addition to the camera parameter estimation and imagery reconstruction losses, we adopt a binary dice loss [45] $\mathcal{L}_{\mathrm{M}}$ between the pre-computed motion mask $M_{t}$ [54] and the rendered motion mask $\hat{M}_t$ that can be derived from the dynamic 3D Gaussians. The binary dice loss initially proposed in [45] for highly imbalanced segmentation of medical imagery helps encouraging better separation between our dynamic and static 3D Gaussians as described by + +$$ +\mathcal {L} _ {\mathrm {M}} = 1 - \frac {2 \sum_ {\varphi_ {t}} M _ {t} \left(\varphi_ {t}\right) \hat {M} _ {t} \left(\varphi_ {t}\right) + \varepsilon}{\sum_ {\varphi_ {t}} \left(M _ {t} \left(\varphi_ {t}\right) + \hat {M} _ {t} \left(\varphi_ {t}\right)\right) + \varepsilon}, \tag {16} +$$ + +where $\varepsilon$ is a smooth term to avoid numerical issues. $\hat{M}_t(\varphi_t)$ is computed by the alpha-blending of the 3D Gaussians overlapping $\varphi_t$ (similar to Eq. 2) as + +
MethodsJumpingSkatingTruckUmbrellaBalloon1Balloon2PlaygroundAverageFPS↑Tr. (h)↓
STM preprocessingDynNeRF (ICCV'21) [12]24.68 / 0.09032.66 / 0.03528.56 / 0.08223.26 / 0.13722.36 / 0.10427.06 / 0.04924.15 / 0.08026.10 / 0.0820.05>24
MonoNeRF (ICCV'23) [46]24.26 / 0.09132.06 / 0.04427.56 / 0.11523.62 / 0.18021.89 / 0.12927.36 / 0.05222.61 / 0.13025.62 / 0.1060.05N/A
STGS (CVPR'24) [24]20.82 / 0.18724.80 / 0.10925.01 / 0.10321.88 / 0.19520.36 / 0.19623.12 / 0.12419.23 / 0.15122.17 / 0.1529000.2
SCGS (CVPR'24) [14]15.68 / 0.92014.88 / 0.90823.81 / 0.14021.84 / 0.16020.17 / 0.17921.07 / 0.14920.71 / 0.11519.74 / 0.3671101.1
D3DGS (CVPR'24) [55]22.02 / 0.26624.06 / 0.22723.04 / 0.24722.67 / 0.19221.22 / 0.20225.86 / 0.11822.30 / 0.11123.02 / 0.195251.0
4DGS (CVPR'24) [52]22.37 / 0.17826.72 / 0.08425.93 / 0.09722.36 / 0.17821.89 / 0.15324.85 / 0.08121.36 / 0.08923.64 / 0.123950.25
SP-GS (ICML'24) [48]22.13 / 0.46829.21 / 0.23627.38 / 0.19024.88 / 0.32324.36 / 0.18029.65 / 0.09722.29 / 0.23425.70 / 0.2471806.0
Casual-FVS (ECCV'24) [22]23.45 / 0.10029.98 / 0.04525.22 / 0.09023.24 / 0.09623.76 / 0.07924.15 / 0.08122.19 / 0.07424.57 / 0.081480.25
Kinematic Fields (ECCV'24) [15]23.74 / 0.10231.89 / 0.03528.34 / 0.07425.46 / 0.09823.72 / 0.07926.49 / 0.05524.64 / 0.05326.33 / 0.0710.37.0
E-D3DGS (ECCV'24) [5]22.39 / 0.18024.62 / 0.13726.85 / 0.08123.60 / 0.14520.87 / 0.16124.39 / 0.07320.79 / 0.12423.36 / 0.129452.6
CTNeRF (Pattern Recognition'24) [33]24.35 / 0.09433.51 / 0.03428.27 / 0.08423.48 / 0.12922.19 / 0.11126.86 / 0.04824.28 / 0.07726.13 / 0.0820.1>24
D-NPC (Eurographics'25) [18]24.51 / 0.11630.22 / 0.06128.92 / 0.08424.15 / 0.13622.26 / 0.16125.84 / 0.10723.59 / 0.09925.64 / 0.109700.27
Shape-of-Motion (arXiv'24) [49]22.74 / 0.12829.12 / 0.04824.63 / 0.12623.79 / 0.08922.74 / 0.12025.83 / 0.07220.61 / 0.13024.21 / 0.1022551.1
Ex4DGS (NeurIPS'24) [21]18.93 / 0.32121.92 / 0.23319.04 / 0.30819.03 / 0.34014.69 / 0.50316.29 / 0.45714.16 / 0.43717.72 / 0.371840.65
DynPoint (NeurIPS'23) [58]24.69 / 0.09731.34 / 0.04529.30 / 0.06124.59 / 0.08622.77 / 0.09927.63 / 0.04925.37 / 0.03926.53 / 0.0680.80.5
RoDynRF (CVPR'23) [30] w/25.66 / 0.07128.68 / 0.04029.13 / 0.06324.26 / 0.08922.37 / 0.10326.19 / 0.05424.96 / 0.04825.89 / 0.0670.45>24
MoSca (arXiv'24) [23] w/25.21 / 0.08332.77 / 0.03328.22 / 0.09024.41 / 0.09223.26 / 0.09228.90 / 0.04223.05 / 0.06026.55 / 0.070400.8
SplineGS (Ours) w/25.62 / 0.06434.10 / 0.03427.83 / 0.04824.60 / 0.07622.67 / 0.12029.04 / 0.03024.47 / 0.04626.90 / 0.0604000.5
STM FreeRoDynRF (CVPR'23) [30]24.27 / 0.10028.71 / 0.04628.85 / 0.06623.25 / 0.10421.81 / 0.12225.58 / 0.06425.20 / 0.05225.38 / 0.0790.45>24
MoSca (arXiv'24) [23]25.43 / 0.08032.62 / 0.03328.29 / 0.08624.40 / 0.09123.27 / 0.09129.01 / 0.04223.23 / 0.05826.61 / 0.069401.0
SplineGS (Ours)25.50 / 0.06833.72 / 0.03128.66 / 0.05625.61 / 0.07124.43 / 0.06828.37 / 0.03224.19 / 0.04727.21 / 0.0534001.5
+ +Table 1. Novel view synthesis evaluation on the NVIDIA dataset (PSNR↑ / LPIPS↓). Red and Blue denote the best and second-best performances, respectively. For Casual-FVS [22], we directly use the results from their paper, as the official code is unavailable. 'w/' denotes the use of camera poses obtained by COLMAP [42] for each SfM-free method. 'Tr' denotes the training time. For MonoNeRF [46], the training time is omitted as it is not a per-scene optimized model. We use an RTX 3090 Ti GPU for each method. + +![](images/3b07c8fae3cafce0df17a4fade182ac546c315331fc6996fa4469cefdc9215cf.jpg) + +![](images/38617b43d9f060febfdb6fa5867c65025a47bdfb2c0ff7de68b745b51f5672de.jpg) +Figure 3. Visual comparisons for novel view synthesis on the NVIDIA dataset. + +![](images/6ba8827c50382e4b70c4d1fdd83446629df95241eb1b996c99a610fa6cc9bc7a.jpg) + +![](images/16689226469fb74279850fc2dccd0a6dd21d2163f52f5250576e68bc67c877da.jpg) + +![](images/e14840e2fa731fc70224196f9043d0bea774865c2273b7735c1c49219e5b5699.jpg) + +![](images/0bf5619da990411a9a9b77e7ce10578fb9d15d42e695b295083021deb49e1bf8.jpg) + +$$ +\hat {M} _ {t} (\boldsymbol {\varphi} _ {t}) = \sum_ {i \in \mathcal {N}} m _ {i} \alpha_ {i} \prod_ {j = 1} ^ {i - 1} (1 - \alpha_ {j}), \tag {17} +$$ + +where $m_{i} = 0$ if the $i^{\mathrm{th}}$ 3D Gaussian is the static 3D Gaussian and $m_{i} = 1$ otherwise. + +In conjunction, all terms in $\mathcal{L}_{\mathrm{total}}^{\mathrm{main}}$ guide our SplineGS to effectively and efficiently model dynamic 3D scenes from pure monocular videos, achieving more structural details and better temporal consistency than previous works [12, 14, 21-24, 30, 46, 52, 55], without relying on camera parameters obtained from external estimators [42]. + +# 5. Experiments + +Implementation Details. We refer readers to the Supplementary material for the details of our implementation. + +Datasets. We evaluate both the quantitative and qualitative performance of novel view and time synthesis on the widely used NVIDIA dataset [56] which features challenging monocular videos. Additionally, we assess novel view synthesis performance on in-the-wild monocular videos from the DAVIS dataset [39] which contains an average of 69.7 frames per video sequence. + +# 5.1. Comparison with State-of-the-Art Methods + +Novel View Synthesis. Table 1 presents a quantitative comparison of NVs between our SplineGS and existing SfM-based [5, 12, 14, 15, 18, 21, 22, 24, 33, 46, 48, 49, 52, 55, 58] and SfM-free [23, 30] methods on the NVIDIA dataset [56]. For this comparison, we follow the evaluation configuration in [30]. The results demonstrate that our SplineGS significantly outperforms SOTA methods in both the PSNR and LPIPS [57] metrics. Notably, SplineGS achieves $890 \times$ and $8,000 \times$ faster rendering speed compared to RoDynRF [30] and DynNeRF [12], respectively. Ex4DGS [21] and STGS [24], which are designed for multi-view settings, face challenges with inconsistent geometry alignment over time when trained on monocular videos. Fig. 3 shows the qualitative comparison between our SplineGS with the existing methods in [12, 24, 30, 55]. As highlighted by the red boxes, our method yields not only higher rendering quality but also dynamic objects more aligned and closer to the ground truth. In Fig. 4, we show our superior NVs results for in-the-wild monocular videos from the DAVIS dataset [39] compared to the existing methods in [24, 30, 55]. + +![](images/ef1c7b7d4cfa7a812fe38d3b2ec250d706f34704c0ec3604d2071aa27dd325dd.jpg) +Figure 4. Visual comparisons for novel view synthesis on the DAVIS dataset. + +![](images/d8e053f2754a6f9dc97640ef988db60489fa8cb7466334a822cc06fe7d895420.jpg) + +Compared to [30] that is also COLMAP-free, SplineGS yields considerably more detailed novel views (red boxes) as shown in Fig. 4. For the other methods [24, 55], we observe that COLMAP [42] fails to recover camera parameters and initial point clouds on the DAVIS dataset [39], as also claimed in [30]. On the other hand, our COLMAP-free SplineGS reconstructs accurate camera parameters that are the ones actually used to train [24, 55], which are shown in Fig. 4 for comparison. More results are in the Supplementary material. + +Novel View and Time Synthesis. To evaluate the capability of SplineGS to model continuous trajectories of moving objects in a scene, we compare the novel view and time synthesis results of SplineGS with those of NeRF-based [12, 30] and 3DGS-based [24, 52, 55] methods. For this evaluation, we follow the dataset sampling strategy in [25], which samples 24 timestamps from the NVIDIA dataset [56]. In addition, to simulate a larger motion, we exclude frames with odd time indices in the training sets. To ensure all test timestamps are not seen during training, and thus, to create a more challenging novel view and time synthesis validation, we exclude frames with even time indices in the test sets. Table 2 and Fig. 5 show the quantitative and qualitative comparisons for this challenging experiment. We observe that the NeRF-based methods, RoDynRF [30] and DynNeRF [12], generate inconsistent artifacts and blurriness for unseen times. Furthermore, the 3DGS-based methods [24, 52, 55] yield even more significant degradation when predicting unseen time indices. In contrast, SplineGS, built upon 3DGS [20] and equipped with our novel spline-based deformation, provides SOTA novel view rendering at unseen intermediate timestamps. Thanks to our MAS, SplineGS naturally and precisely captures the continuous trajectories of moving objects over time, enhancing the temporal consistency of rendered scenes that can be checked in tOF scores [7] of Table 2. + +
MethodsPSNR↑LPIPS↓tOF↓
SfM preprocessingDynNeRF (ICCV'21) [12]23.360.2190.921
4DGS (CVPR'24) [52]17.070.4596.314
D3DGS (CVPR'24) [55]19.630.3433.225
STGS (CVPR'24) [24]15.720.4742.105
SfM-FreeRoDynRF (CVPR'23) [30]21.580.2212.138
SplineGS (Ours)25.920.0980.703
+ +Table 2. Novel view and time synthesis evaluation on the NVIDIA dataset. + +![](images/d5104621fc11f3df0a7975772d82ae75ab57e7fa3b8cebc5800f69a13c77acfc.jpg) +Figure 5. Visual comparisons for novel view and time synthesis on the NVIDIA dataset. + +![](images/ac5c8eb84a529459c8cb420935f2c1478dab1eb4920cdbd17b3c25d108accdcf.jpg) + +![](images/5a130a8da916272055ea4e6bf9a03ca29408682b2a171bb3df995cabc0538a57.jpg) + +![](images/1d91ca2faf8697bb02bd4a12b80da090c4edff8049dd9529384b2bb290d4e787.jpg) + +![](images/a7a933d4bd704964aecbad34995d477aa98a24f5d2c7fb44b2ce353e04afd924.jpg) + +To further analyze the SplineGS's ability to model continuous trajectories of dynamic 3D Gaussians, we visualize the projected 2D motion tracking of dynamic objects in pixel space, comparing it with D3DGS [55] and STGS [24], as shown in Fig. 6. We observe that D3DGS [55] and STGS [24] cannot provide reliable motion tracking for moving objects, underscoring their limitations in modeling continuous trajectories of dynamic 3D Gaussians. In contrast, SplineGS provides accurate motion tracking, demonstrating the effectiveness of MAS for deforming dynamic 3D Gaussians. Further analysis is in the Supplementary material. + +![](images/9401d0320810a7442011e77ff7e49989879b328c1ece6439499d28b7b43aa4b0.jpg) +Figure 6. Visual comparisons for motion tracking. We visualize 2D pixel tracks to analyze motions of dynamic 3D Gaussians. + +![](images/140c61669249bd5e0f14e8957127219e4dd2c1c41eb0934d5aac4584855a0708.jpg) + +![](images/7a466f6f93ee90b8328922bb4b7ed4dfbc2d5cc42d26580ed99fe67c8c639ede.jpg) + +# 5.2. Ablation Study + +Motion-Adaptive Spline (MAS). To demonstrate the effectiveness of MAS, we replace the MAS model with various deformation models, including an MLP, a grid-based model, polynomial functions of third degree (denoted as 'Poly $(3^{\mathrm{rd}})$ ') and tenth degree (denoted as 'Poly $(10^{\mathrm{th}})$ '), and a Bezier curve [9]. For the MLP, grid-based model, and polynomial functions, we apply to them the structures similar to those in prior works, including D3DGS [55], 4DGS [52], and STGS [24], respectively. Additionally, we implement the Bezier curve [9], a commonly used method for curve modeling in computer graphics. Table 3-(a) presents quantitative comparisons of each 3D Gaussian trajectory model, focusing on rendering quality (PSNR, LPIPS) and deformation latency per Gaussian, denoted as $g_{\mathrm{def}}$ . This latency re + +![](images/dae64395a9f2a2f04571df543f94a4015fef14c8bba9911c8c798a497b5e2a13.jpg) +Figure 7. Visual comparisons for MACP ablation study. + +flects the computational time required to estimate the deformation of a single dynamic 3D Gaussian. As shown in Table 3-(a), our MAS model achieves superior rendering quality compared to all other deformation models. Consistent with analyses in previous works [24, 52, 55], we observe that the MLP and grid-based architectures require substantial computational costs for rendering. Among these methods, 'Poly $(3^{\mathrm{rd}})$ ', as implemented in [24], demonstrates the best latency. However, fixed-degree polynomial functions have limited flexibility across varying motion complexities which adversely impacts rendering performance. To explore this further, we experiment with 'Poly $(10^{\mathrm{th}})$ ' to assess changes in modeling capability. This adjustment, however, leads to noisier optimization and reduced efficiency, as variables under high exponents in 'Poly $(10^{\mathrm{th}})$ ' lead to numerical instability. The Bézier curve [9] offers the second-best rendering quality, but its latency remains higher than our MAS due to its recursive nature of computation. + +Motion-Adaptive Control Points Pruning (MACP). To assess the effectiveness of our MACP technique for MAS, we compare our full model with MACP against other versions of our model with two fixed numbers of control points $N_{c} = 4$ and $N_{c} = N_{f}$ . As shown in Table 3-(c) and Fig. 7, our SplineGS with MACP achieves a good trade-off between rendering quality and $g_{\mathrm{def}}$ compared to the ablated models with fixed $N_{c}$ . Using $N_{c} = 4$ for every dynamic 3D Gaussian limits the motion modeling capacity of MAS, resulting in significantly lower metrics and visible artifacts in the dynamic regions. Moreover, an excessive $N_{c} = N_{f}$ decreases the rendering speed of our MAS module and still falls short of the quality achieved by our full model with MACP, potentially due to motion overfitting. Fig. 8 shows the distribution of $N_{c}$ values after optimization with MACP across scenes of varying motion complexities. In Fig. 8-(a), we visualize ' $N_{c}$ Heatmap' that contains the pixel-wise averaged $N_{c}$ values of the dynamic 3D Gaussians needed to render the 2D pixels (red higher, blue lower number of control points). As shown, the simpler motions in these scenes, such as those of the human bodies, can be modeled by the dynamic 3D Gaussians with smaller averages of $N_{c}$ values, whereas the objects with complex and extensive motions (e.g., balloon) require higher averaged $N_{c}$ values. Fig. 8-(b) presents the corresponding histogram of $N_{c}$ values for the scenes' dynamic 3D Gaussians. For sequences with simple motion, such as 'Skating', the trajectories of most + +(a) Motion-Adaptive Spline + +
PSNR↑LPIPS↓gDef (ns)↓
MLP23.510.125149.41
Grid25.480.09098.89
Poly (3rd)25.140.1111.80
Poly (10th)24.380.1207.71
Bézier27.190.0608.78
Ours27.210.0535.63
+ +(b) Loss function + +
PSNR↑LPIPS↓
w/o Lpc17.490.853
w/o Lgc26.330.067
w/o Ld-pc26.180.066
w/o Lm26.340.088
Ours27.210.053
+ +(c) Motion-Adaptive Control points Pruning + +
PSNR↑LPIPS↓gdef (ns)↓
w/o MACP (Nc=4)26.620.0655.34
w/o MACP (Nc=Nf)27.080.0546.11
Ours27.210.0535.63
+ +Table 3. Ablation studies. We ablate our framework and report the average results on the NVIDIA dataset with the same setting as Novel View Synthesis experiment in Sec. 5.1. + +![](images/ac44e70d252a6783e3f0d3fcd541de3c038b9583bf1107457910d3964c2cfe6a.jpg) +Figure 8. Analysis of MACP's Efficacy. (a) $N_{c}$ Heatmaps as the averaged $N_{c}$ values of dynamic 3D Gaussians and their corresponding rendered frames $\hat{I}_{t}$ for 'Balloon2' and 'Skating' scenes. (b) Histograms of the number of control points $(N_{c})$ in percentages (\%) of dynamic 3D Gaussians in two scenes. + +![](images/90c6520ad34e6f0d3370f97583f2b095dffd15915aacf4e9f5e80889ae5f0261.jpg) + +dynamic 3D Gaussians can be represented using a minimal $N_{c}$ , thanks to our MACP. While 'Balloon2' has more evenly distributed $N_{c}$ due to more complex and diverse motion. + +Loss Functions. Table 3-(b) shows the effectiveness of each loss for our overall SplineGS architecture. As noted, no consistent camera parameters can be learned without $\mathcal{L}_{\mathrm{pc}}$ , drastically impacting the rendering quality of the dynamic 3D Gaussians. Moreover, our $\mathcal{L}_{\mathrm{gc}}$ , $\mathcal{L}_{\mathrm{d - pc}}$ and $\mathcal{L}_{\mathrm{M}}$ can considerably impact the overall rendering quality. + +# 6. Conclusion + +We present SplineGS, a COLMAP-free dynamic 3DGS framework designed for novel spatio-temporal view synthesis from monocular videos. Leveraging our innovative Motion-Adaptive Spline (MAS) for dynamic motion modeling, SplineGS efficiently renders high-quality novel views from complex in-the-wild videos. The effectiveness of our approach is validated through extensive quantitative and qualitative comparisons, significantly outperforming the existing SOTA methods with very fast rendering speed. + +# Acknowledgement + +This work was supported by Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korean Government [Ministry of Science and ICT (Information and Communications Technology)] (Project Number: RS-2022-00144444, Project Title: Deep Learning Based Visual Representational Learning and Rendering of Static and Dynamic Scenes, 100%). + +# References + +[1] Fast dynamic radiance fields with time-aware neural voxels. In SIGGRAPH Asia 2022 Conference Papers, 2022. 2 +[2] J Harold Ahlberg, Edwin Norman Nilson, and Joseph Leonard Walsh. The Theory of Splines and Their Applications: Mathematics in Science and Engineering: A Series of Monographs and Textbooks, Vol. 38. Elsevier, 2016. 2, 3, 4 +[3] ShahRukh Athar, Zexiang Xu, Kalyan Sunkavalli, Eli Shechtman, and Zhixin Shu. Rignerf: Fully controllable neural 3d portraits. In CVPR, 2022. 2 +[4] Benjamin Attal, Jia-Bin Huang, Christian Richardt, Michael Zollhoefer, Johannes Kopf, Matthew O'Toole, and Changil Kim. Hyperreel: High-fidelity 6-dof video with ray-conditioned sampling. In CVPR, 2023. 2 +[5] Jeongmin Bae, Seoha Kim, Youngsik Yun, Hahyun Lee, Gun Bang, and Youngjung Uh. Per-gaussian embedding-based deformation for deformable 3d gaussian splatting. In European Conference on Computer Vision, pages 321-335. Springer, 2024. 6 +[6] Ang Cao and Justin Johnson. Hexplane: A fast representation for dynamic scenes. In CVPR, 2023. 2 +[7] Mengyu Chu, You Xie, Jonas Mayer, Laura Leal-Taixe, and Nils Thuerey. Learning temporal coherence via self-supervision for gan-based video generation. ACM Transactions on Graphics (TOG), 2020. 7 +[8] C De Boor. A practical guide to splines. Springer-Verlag google schola, 1978. 2, 3, 4 +[9] Gerald Farin. Curves and surfaces for CAGD: a practical guide. Elsevier, 2001. 2, 7, 8 +[10] Sara Fridovich-Keil, Giacomo Meanti, Frederik Rahbæk Warburg, Benjamin Recht, and Angjoo Kanazawa. K-planes: Explicit radiance fields in space, time, and appearance. In CVPR, 2023. 2 +[11] Yang Fu, Sifei Liu, Amey Kulkarni, Jan Kautz, Alexei A. Efros, and Xiaolong Wang. Colmap-free 3d gaussian splatting. In CVPR, 2024. 3 +[12] Chen Gao, Ayush Saraf, Johannes Kopf, and Jia-Bin Huang. Dynamic view synthesis from dynamic monocular video. In ICCV, 2021. 2, 6, 7 +[13] Roger A Horn and Charles R Johnson. Matrix analysis. Cambridge university press, 2012. 5 +[14] Yi-Hua Huang, Yang-Tian Sun, Ziyi Yang, Xiaoyang Lyu, Yan-Pei Cao, and Xiaojuan Qi. Sc-gs: Sparse-controlled gaussian splatting for editable dynamic scenes. CVPR, 2024. 2, 6 + +[15] Woobin Im, Geonho Cha, Sebin Lee, Jumin Lee, Juhyeong Seon, Dongyoon Wee, and Sung-Eui Yoon. Regularizing dynamic radiance fields with kinematic fields. In European Conference on Computer Vision, pages 312-328. Springer, 2024. 6 +[16] Wei Jiang, Kwang Moo Yi, Golnoosh Samei, Oncel Tuzel, and Anurag Ranjan. Neuman: Neural human radiance field from a single video. In ECCV, 2022. 2 +[17] Yifan Jiang, Peter Hedman, Ben Mildenhall, Dejia Xu, Jonathan T Barron, Zhangyang Wang, and Tianfan Xue. Alignerf: High-fidelity neural radiance fields via alignment-aware training. In CVPR, 2023. 3 +[18] Moritz Kappel, Florian Hahlbohm, Timon Scholz, Susana Castillo, Christian Theobalt, Martin Eisemann, Vladislav Golyanik, and Marcus Magnor. D-npc: Dynamic neural point clouds for non-rigid view synthesis from monocular video. arXiv preprint arXiv:2406.10078, 2024. 6 +[19] Nikita Karaev, Ignacio Rocco, Benjamin Graham, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. Co-tracker: It is better to track together. arXiv, 2023. 4 +[20] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 2023. 2, 3, 4, 7 +[21] Junoh Lee, Chang-Yeon Won, Hyunjun Jung, Inhwan Bae, and Hae-Gon Jeon. Fully explicit dynamic gaussian splatting. In NeurIPS, 2024. 3, 6 +[22] Yao-Chih Lee, Zhoutong Zhang, Kevin Blackburn-Matzen, Simon Niklaus, Jianming Zhang, Jia-Bin Huang, and Feng Liu. Fast view synthesis of casual videos with soup-of-planes. In ECCV, 2024. 3, 6 +[23] Jiahui Lei, Yijia Weng, Adam Harley, Leonidas Guibas, and Kostas Daniilidis. Mosca: Dynamic gaussian fusion from casual videos via 4d motion scaffolds. arXiv, 2024. 3, 6 +[24] Zhan Li, Zhang Chen, Zhong Li, and Yi Xu. Spacetime gaussian feature splatting for real-time dynamic view synthesis. In CVPR. 1, 2, 3, 6, 7, 8 +[25] Zhengqi Li, Simon Niklaus, Noah Snavely, and Oliver Wang. Neural scene flow fields for space-time view synthesis of dynamic scenes. In CVPR, 2021. 2, 7 +[26] Zhengqi Li, Qianqian Wang, Forrester Cole, Richard Tucker, and Noah Snavely. Dynibar: Neural dynamic image-based rendering. In CVPR, 2023. 2 +[27] Yiqing Liang, Numair Khan, Zhengqin Li, Thu Nguyen-Phuoc, Douglas Lanman, James Tompkin, and Lei Xiao. Gaufre: Gaussian deformation fields for real-time dynamic novel view synthesis. arXiv, 2023. 2, 3 +[28] Chen-Hsuan Lin, Wei-Chiu Ma, Antonio Torralba, and Simon Lucey. Barf: Bundle-adjusting neural radiance fields. In ICCV, 2021. 3 +[29] Philipp Lindenberger, Paul-Edouard Sarlin, Viktor Larsson, and Marc Pollefeys. Pixel-perfect structure-from-motion with featuremetric refinement. In ICCV, 2021. 3 +[30] Yu-Lun Liu, Chen Gao, Andreas Meuleman, Hung-Yu Tseng, Ayush Saraf, Changil Kim, Yung-Yu Chuang, Johannes Kopf, and Jia-Bin Huang. Robust dynamic radiance fields. In CVPR, 2023. 2, 3, 5, 6, 7 + +[31] Jonathon Luiten, Georgios Kopanas, Bastian Leibe, and Deva Ramanan. Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis. In 3DV, 2024. 2 +[32] Quan Meng, Anpei Chen, Haimin Luo, Minye Wu, Hao Su, Lan Xu, Xuming He, and Jingyi Yu. Gnerf: Gan-based neural radiance field without posed camera. In ICCV, 2021. 3 +[33] Xingyu Miao, Yang Bai, Haoran Duan, Fan Wan, Yawen Huang, Yang Long, and Yefeng Zheng. Ctnerf: Cross-time transformer for dynamic neural radiance field from monocular video. Pattern Recognition, 156:110729, 2024. 6 +[34] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 2, 5 +[35] Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In ICCV, 2021. 2 +[36] Keunhong Park, Utkarsh Sinha, Peter Hedman, Jonathan T. Barron, Sofien Bouaziz, Dan B Goldman, Ricardo Martin-Brualla, and Steven M. Seitz. Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields. ACM Trans. Graph., 2021. 2 +[37] Keunhong Park, Philipp Henzler, Ben Mildenhall, Jonathan T. Barron, and Ricardo Martin-Brualla. Camp: Camera preconditioning for neural radiance fields. ACM Trans. Graph., 2023. 3 +[38] Luigi Piccinelli, Yung-Hsu Yang, Christos Sakaridis, Mattia Segu, Siyuan Li, Luc Van Gool, and Fisher Yu. Unidepth: Universal monocular metric depth estimation. In CVPR, 2024. 4, 5 +[39] Jordi Pont-Tuset, Federico Peruzzi, Sergi Caelles, Pablo Arbeláez, Alex Sorkine-Hornung, and Luc Van Gool. The 2017 davis challenge on video object segmentation. arXiv, 2018. 1, 6, 7 +[40] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-nerf: Neural radiance fields for dynamic scenes. In CVPR, 2021. 2 +[41] Vincent Raoult, Sarah Reid-Anderson, Andreas Ferri, and Jane E Williamson. How reliable is structure from motion (sfm) over time and between observers? a case study using coral reef bommie. Remote Sensing, 2017. 3 +[42] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In CVPR, 2016. 1, 2, 3, 5, 6, 7 +[43] Ruizhi Shao, Zerong Zheng, Hanzhang Tu, Boning Liu, Hongwen Zhang, and Yebin Liu. Tensor4d: Efficient neural 4d decomposition for high-fidelity dynamic reconstruction and rendering. In CVPR, 2023. 2 +[44] Liangchen Song, Anpei Chen, Zhong Li, Zhang Chen, Lele Chen, Junsong Yuan, Yi Xu, and Andreas Geiger. Nerfplayer: A streamable dynamic scene representation with decomposed neural radiance fields. IEEE Transactions on Visualization and Computer Graphics, 2023. 2 +[45] Carole H. Sudre, Wenqi Li, Tom Vercauteren, Sebastien Ourselin, and M. Jorge Cardoso. Generalised dice overlap as a deep learning loss function for highly unbalanced segmentations. In Deep Learning in Medical Image Analysis and Multimodal Learning for Clinical Decision Support, 2017. 5 + +[46] Fengrui Tian, Shaoyi Du, and Yueqi Duan. MonoN-eRF: Learning a generalizable dynamic radiance field from monocular videos. In ICCV, 2023. 6 +[47] Edgar Tretschk, Ayush Tewari, Vladislav Golyanik, Michael Zollhöfer, Christoph Lassner, and Christian Theobalt. Non-rigid neural radiance fields: Reconstruction and novel view synthesis of a dynamic scene from monocular video. In ICCV, 2021. 2 +[48] Diwen Wan, Ruijie Lu, and Gang Zeng. Superpoint gaussian splatting for real-time high-fidelity dynamic scene reconstruction. arXiv preprint arXiv:2406.03697, 2024. 6 +[49] Qianqian Wang, Vickie Ye, Hang Gao, Jake Austin, Zhengqi Li, and Angjoo Kanazawa. Shape of motion: 4d reconstruction from a single video. arXiv preprint arXiv:2407.13764, 2024. 6 +[50] Zirui Wang, Shangzhe Wu, Weidi Xie, Min Chen, and Victor Adrian Prisacariu. Nerf: Neural radiance fields without known camera parameters. CoRR, 2021. 3 +[51] Chung-Yi Weng, Brian Curless, Pratul P. Srinivasan, Jonathan T. Barron, and Ira Kemelmacher-Shlizerman. HumanNeRF: Free-viewpoint rendering of moving people from monocular video. In CVPR, 2022. 2 +[52] Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Xinggang Wang. 4d gaussian splatting for real-time dynamic scene rendering. In CVPR. 2, 6, 7, 8 +[53] Gengshan Yang, Minh Vo, Neverova Natalia, Deva Ramanan, Vedaldi Andrea, and Joo Hanbyul. Banmo: Building animatable 3d neural models from many casual videos. In CVPR, 2022. 2 +[54] Jinyu Yang, Mingqi Gao, Zhe Li, Shang Gao, Fangjing Wang, and Feng Zheng. Track anything: Segment anything meets videos. arXiv, 2023. 5 +[55] Ziyi Yang, Xinyu Gao, Wen Zhou, Shaohui Jiao, Yuqing Zhang, and Xiaogang Jin. Deformable 3d gaussians for high-fidelity monocular dynamic scene reconstruction. In CVPR, 2024. 1, 2, 6, 7, 8 +[56] Jae Shin Yoon, Kihwan Kim, Orazio Gallo, Hyun Soo Park, and Jan Kautz. Novel view synthesis of dynamic scenes with globally coherent depths from a monocular camera. In CVPR, 2020. 1, 6, 7 +[57] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 6 +[58] Kaichen Zhou, Jia-Xing Zhong, Sangyun Shin, Kai Lu, Yiyuan Yang, Andrew Markham, and Niki Trigoni. Dynpoint: Dynamic neural point for view synthesis. Advances in Neural Information Processing Systems, 36:69532-69545, 2023. 6 \ No newline at end of file diff --git a/2025/SplineGS_ Robust Motion-Adaptive Spline for Real-Time Dynamic 3D Gaussians from Monocular Video/images.zip b/2025/SplineGS_ Robust Motion-Adaptive Spline for Real-Time Dynamic 3D Gaussians from Monocular Video/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..90d0c5fa49e3707deda6efd19f7362d2d11b89ad --- /dev/null +++ b/2025/SplineGS_ Robust Motion-Adaptive Spline for Real-Time Dynamic 3D Gaussians from Monocular Video/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:24afbff14369f611a69d6aa6f7dbbe3d266ab3b40c0eab3abb7f468b9cffa63b +size 914942 diff --git a/2025/SplineGS_ Robust Motion-Adaptive Spline for Real-Time Dynamic 3D Gaussians from Monocular Video/layout.json b/2025/SplineGS_ Robust Motion-Adaptive Spline for Real-Time Dynamic 3D Gaussians from Monocular Video/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..2d7452d15bc2405fcf4c82606a37b8bbe739152d --- /dev/null +++ b/2025/SplineGS_ Robust Motion-Adaptive Spline for Real-Time Dynamic 3D Gaussians from Monocular Video/layout.json @@ -0,0 +1,11020 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 116, + 102, + 494, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 102, + 494, + 140 + ], + "spans": [ + { + "bbox": [ + 116, + 102, + 494, + 140 + ], + "type": "text", + "content": "SplineGS: Robust Motion-Adaptive Spline for Real-Time Dynamic 3D Gaussians from Monocular Video" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 101, + 160, + 507, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 101, + 160, + 507, + 190 + ], + "spans": [ + { + "bbox": [ + 101, + 160, + 507, + 190 + ], + "type": "text", + "content": "Jongmin Park1* Minh-Quan Viet Bui1* Juan Luis Gonzalez Bello1 Jaeho Moon1 \nJihyong Oh2† Munchurl Kim1†" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 81, + 202, + 529, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 202, + 529, + 245 + ], + "spans": [ + { + "bbox": [ + 81, + 202, + 529, + 245 + ], + "type": "text", + "content": "\\(^{1}\\)Korea Advanced Institute of Science and Technology (KAIST) \\(^{2}\\)Chung-Ang University \\(\\{\\mathrm{jm.park, bvmquan, juanluisgb, jaehomoon, mkimee}\\} @kaist.ac.kr jihyongoh@cau.ac.kr https://kaist-viclab.github.io/splinegs-site/" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 56, + 273, + 192, + 351 + ], + "blocks": [ + { + "bbox": [ + 56, + 273, + 192, + 351 + ], + "lines": [ + { + "bbox": [ + 56, + 273, + 192, + 351 + ], + "spans": [ + { + "bbox": [ + 56, + 273, + 192, + 351 + ], + "type": "image", + "image_path": "2d814176e456e030dc547810d5c45437fb60e893cbbfc71dc4c73b79c3961383.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 56, + 361, + 192, + 439 + ], + "blocks": [ + { + "bbox": [ + 78, + 352, + 169, + 361 + ], + "lines": [ + { + "bbox": [ + 78, + 352, + 169, + 361 + ], + "spans": [ + { + "bbox": [ + 78, + 352, + 169, + 361 + ], + "type": "text", + "content": "horsejump-low, " + }, + { + "bbox": [ + 78, + 352, + 169, + 361 + ], + "type": "inline_equation", + "content": "t = 20" + }, + { + "bbox": [ + 78, + 352, + 169, + 361 + ], + "type": "text", + "content": " (Ours)" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 56, + 361, + 192, + 439 + ], + "lines": [ + { + "bbox": [ + 56, + 361, + 192, + 439 + ], + "spans": [ + { + "bbox": [ + 56, + 361, + 192, + 439 + ], + "type": "image", + "image_path": "b6a70072525f4339034e7b7eac3563aa9a2ea4bacb05fd1919f16215e7522bd3.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 94, + 440, + 153, + 448 + ], + "lines": [ + { + "bbox": [ + 94, + 440, + 153, + 448 + ], + "spans": [ + { + "bbox": [ + 94, + 440, + 153, + 448 + ], + "type": "text", + "content": "train, " + }, + { + "bbox": [ + 94, + 440, + 153, + 448 + ], + "type": "inline_equation", + "content": "t = 37" + }, + { + "bbox": [ + 94, + 440, + 153, + 448 + ], + "type": "text", + "content": " (Ours)" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 75, + 453, + 319, + 463 + ], + "lines": [ + { + "bbox": [ + 75, + 453, + 319, + 463 + ], + "spans": [ + { + "bbox": [ + 75, + 453, + 319, + 463 + ], + "type": "text", + "content": "(a) Visual comparison for novel view synthesis on the DAVIS dataset" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 194, + 274, + 266, + 306 + ], + "blocks": [ + { + "bbox": [ + 194, + 274, + 266, + 306 + ], + "lines": [ + { + "bbox": [ + 194, + 274, + 266, + 306 + ], + "spans": [ + { + "bbox": [ + 194, + 274, + 266, + 306 + ], + "type": "image", + "image_path": "d9352c8f024540b3c8bf6e9266b5858c9053440c7132ec15b88a96de2b7587f1.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 211, + 306, + 247, + 315 + ], + "lines": [ + { + "bbox": [ + 211, + 306, + 247, + 315 + ], + "spans": [ + { + "bbox": [ + 211, + 306, + 247, + 315 + ], + "type": "text", + "content": "STGS [24]" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 266, + 274, + 339, + 306 + ], + "blocks": [ + { + "bbox": [ + 266, + 274, + 339, + 306 + ], + "lines": [ + { + "bbox": [ + 266, + 274, + 339, + 306 + ], + "spans": [ + { + "bbox": [ + 266, + 274, + 339, + 306 + ], + "type": "image", + "image_path": "17e67f379b6de6572a4764d79a54bd0f29a90694ab22571c5dc19c1cf204e5c2.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 281, + 307, + 323, + 316 + ], + "lines": [ + { + "bbox": [ + 281, + 307, + 323, + 316 + ], + "spans": [ + { + "bbox": [ + 281, + 307, + 323, + 316 + ], + "type": "text", + "content": "D3DGS [55]" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 194, + 319, + 265, + 351 + ], + "blocks": [ + { + "bbox": [ + 194, + 319, + 265, + 351 + ], + "lines": [ + { + "bbox": [ + 194, + 319, + 265, + 351 + ], + "spans": [ + { + "bbox": [ + 194, + 319, + 265, + 351 + ], + "type": "image", + "image_path": "9dd5a5d3ec9936b0ecd139efda7215dd2aade2231be5f563f77f6cbb3e4d8850.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 266, + 319, + 338, + 351 + ], + "blocks": [ + { + "bbox": [ + 266, + 319, + 338, + 351 + ], + "lines": [ + { + "bbox": [ + 266, + 319, + 338, + 351 + ], + "spans": [ + { + "bbox": [ + 266, + 319, + 338, + 351 + ], + "type": "image", + "image_path": "54aa8c97d8f5c60eae6a136c42e4d286c5ddf29d725cb583ba9e6547e452f9bd.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 193, + 361, + 265, + 392 + ], + "blocks": [ + { + "bbox": [ + 205, + 352, + 253, + 361 + ], + "lines": [ + { + "bbox": [ + 205, + 352, + 253, + 361 + ], + "spans": [ + { + "bbox": [ + 205, + 352, + 253, + 361 + ], + "type": "text", + "content": "RoDynRF [30]" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 193, + 361, + 265, + 392 + ], + "lines": [ + { + "bbox": [ + 193, + 361, + 265, + 392 + ], + "spans": [ + { + "bbox": [ + 193, + 361, + 265, + 392 + ], + "type": "image", + "image_path": "b3fd31d14da4013979908b2ad4f7ef4da7c19e4b5fa7739d9a8030a35b30c842.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 211, + 393, + 247, + 403 + ], + "lines": [ + { + "bbox": [ + 211, + 393, + 247, + 403 + ], + "spans": [ + { + "bbox": [ + 211, + 393, + 247, + 403 + ], + "type": "text", + "content": "STGS [24]" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 266, + 361, + 338, + 392 + ], + "blocks": [ + { + "bbox": [ + 273, + 352, + 331, + 361 + ], + "lines": [ + { + "bbox": [ + 273, + 352, + 331, + 361 + ], + "spans": [ + { + "bbox": [ + 273, + 352, + 331, + 361 + ], + "type": "text", + "content": "SplineGS (Ours)" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 266, + 361, + 338, + 392 + ], + "lines": [ + { + "bbox": [ + 266, + 361, + 338, + 392 + ], + "spans": [ + { + "bbox": [ + 266, + 361, + 338, + 392 + ], + "type": "image", + "image_path": "8f48f9a6140fcffe04e1e9a96352139ff158accf5eed32af915b36f17b9e3506.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 281, + 394, + 323, + 404 + ], + "lines": [ + { + "bbox": [ + 281, + 394, + 323, + 404 + ], + "spans": [ + { + "bbox": [ + 281, + 394, + 323, + 404 + ], + "type": "text", + "content": "D3DGS [55]" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 194, + 406, + 265, + 439 + ], + "blocks": [ + { + "bbox": [ + 194, + 406, + 265, + 439 + ], + "lines": [ + { + "bbox": [ + 194, + 406, + 265, + 439 + ], + "spans": [ + { + "bbox": [ + 194, + 406, + 265, + 439 + ], + "type": "image", + "image_path": "fa809b097842eada0d1a61b00139df7b8e5367c2fa50bfc82606e5e685f5a7af.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 205, + 439, + 253, + 449 + ], + "lines": [ + { + "bbox": [ + 205, + 439, + 253, + 449 + ], + "spans": [ + { + "bbox": [ + 205, + 439, + 253, + 449 + ], + "type": "text", + "content": "RoDynRF [30]" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 266, + 406, + 339, + 439 + ], + "blocks": [ + { + "bbox": [ + 266, + 406, + 339, + 439 + ], + "lines": [ + { + "bbox": [ + 266, + 406, + 339, + 439 + ], + "spans": [ + { + "bbox": [ + 266, + 406, + 339, + 439 + ], + "type": "image", + "image_path": "7e260132ffbf6ac6a9c233c420f0aa50b81213865d280c623e8a7939c24a9302.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 273, + 439, + 331, + 449 + ], + "lines": [ + { + "bbox": [ + 273, + 439, + 331, + 449 + ], + "spans": [ + { + "bbox": [ + 273, + 439, + 331, + 449 + ], + "type": "text", + "content": "SplineGS (Ours)" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 343, + 274, + 549, + 447 + ], + "blocks": [ + { + "bbox": [ + 343, + 274, + 549, + 447 + ], + "lines": [ + { + "bbox": [ + 343, + 274, + 549, + 447 + ], + "spans": [ + { + "bbox": [ + 343, + 274, + 549, + 447 + ], + "type": "image", + "image_path": "0ff377d5310212799fa39bf4e7c5a8018446762d1f41155fc3086c69d69c5d05.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 367, + 452, + 526, + 463 + ], + "lines": [ + { + "bbox": [ + 367, + 452, + 526, + 463 + ], + "spans": [ + { + "bbox": [ + 367, + 452, + 526, + 463 + ], + "type": "text", + "content": "(b) Performance gain on the NVIDIA dataset" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 55, + 466, + 555, + 510 + ], + "lines": [ + { + "bbox": [ + 55, + 466, + 555, + 510 + ], + "spans": [ + { + "bbox": [ + 55, + 466, + 555, + 510 + ], + "type": "text", + "content": "Figure 1. Our SplineGS achieves state-of-the-art rendering quality with fast rendering speed for novel spatio-temporal view synthesis from monocular videos without relying on pre-computed camera parameters. (a) We use our predicted camera parameters for [24, 55] since COLMAP [42] is unable to provide reasonable camera parameters for most scenes in the DAVIS dataset [39]. (b) SplineGS achieves the best rendering quality while maintaining a rendering speed of 400 FPS on the NVIDIA dataset [56]." + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + } + ], + "index": 26 + }, + { + "bbox": [ + 151, + 520, + 200, + 533 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 520, + 200, + 533 + ], + "spans": [ + { + "bbox": [ + 151, + 520, + 200, + 533 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 53, + 545, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 545, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 53, + 545, + 297, + 714 + ], + "type": "text", + "content": "Synthesizing novel views from in-the-wild monocular videos is challenging due to scene dynamics and the lack of multi-view cues. To address this, we propose SplineGS, a COLMAP-free dynamic 3D Gaussian Splatting (3DGS) framework for high-quality reconstruction and fast rendering from monocular videos. At its core is a novel Motion-Adaptive Spline (MAS) method, which represents continuous dynamic 3D Gaussian trajectories using cubic Hermite splines with a small number of control points. For MAS, we introduce a Motion-Adaptive Control points Pruning (MACP) method to model the deformation of each dynamic 3D Gaussian across varying motions, progressively pruning control points while maintaining dynamic modeling integrity. Additionally, we present a joint optimization strat" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 313, + 521, + 555, + 618 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 555, + 618 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 555, + 618 + ], + "type": "text", + "content": "egy for camera parameter estimation and 3D Gaussian attributes, leveraging photometric and geometric consistency. This eliminates the need for Structure-from-Motion preprocessing and enhances SplineGS's robustness in real-world conditions. Experiments show that SplineGS significantly outperforms state-of-the-art methods in novel view synthesis quality for dynamic scenes from monocular videos, achieving thousands times faster rendering speed." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 314, + 627, + 395, + 639 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 627, + 395, + 639 + ], + "spans": [ + { + "bbox": [ + 314, + 627, + 395, + 639 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 313, + 647, + 555, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 647, + 555, + 685 + ], + "spans": [ + { + "bbox": [ + 313, + 647, + 555, + 685 + ], + "type": "text", + "content": "Novel View Synthesis (NVS) is fundamental to 3D vision, supporting applications like virtual reality (VR), augmented reality (AR), and film production. NFS aims to generate" + } + ] + } + ], + "index": 33 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 324, + 693, + 449, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 693, + 449, + 703 + ], + "spans": [ + { + "bbox": [ + 324, + 693, + 449, + 703 + ], + "type": "text", + "content": "*Co-first authors (equal contribution)." + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 326, + 703, + 416, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 703, + 416, + 713 + ], + "spans": [ + { + "bbox": [ + 326, + 703, + 416, + 713 + ], + "type": "text", + "content": "† Co-corresponding authors." + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "26866" + } + ] + } + ], + "index": 37 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 180 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 180 + ], + "type": "text", + "content": "images from any viewpoint in a scene, requiring accurate reconstruction from multiple 2D images. NeRF [34] has advanced the field of NVs by utilizing learned implicit functions to represent static scenes, though it requires considerable training and rendering time. Recently, 3D Gaussian Splatting (3DGS) [20] has revolutionized this process by replacing implicit volumetric rendering with differentiable rasterization of 3D Gaussians, enabling real-time rendering and providing a more explicit scene representation." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 183, + 296, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 183, + 296, + 434 + ], + "spans": [ + { + "bbox": [ + 56, + 183, + 296, + 434 + ], + "type": "text", + "content": "For NVs of dynamic scenes, prior works have extended the static scene representations [20, 34] by incorporating deformation models in canonical space using implicit representations [12, 30, 35, 36, 55], grid-based models that decompose the 4D space-time domain into multiple 2D planes [4, 6, 10, 43, 52], and polynomial trajectories [24]. However, these methods face several challenges: implicit representations significantly increase computational overhead and reduce rendering speed [6, 52]; grid-based models struggle to fully capture the dynamic nature of scene structures, hindering their ability to model fine details accurately [24]; and although polynomial trajectories improve efficiency, their fixed degree restricts flexibility for representing complex motions. To the best of our knowledge, none of the dynamic 3DGS-based methods provide experimental evidence to reliably support their novel spatio-temporal view synthesis capabilities for rendering unseen intermediate time indices. Additionally, most existing methods for dynamic NVs rely heavily on external camera parameter estimators like COLMAP [42], which often produce imprecise results for challenging in-the-wild monocular videos." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 438, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 438, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 438, + 295, + 713 + ], + "type": "text", + "content": "To address these issues for modeling scene dynamics, we exploit a spline-based model to dynamic 3D Gaussian trajectories, inspired by classic 3D-curve modeling in computer graphics [9]. Splines are widely used in geometric modeling for graphical applications, providing smooth and continuous representations of complex shapes with a minimal number of control points [2, 8]. This efficiency makes them ideal for modeling intricate geometric structures while maintaining both flexibility and precision. In this paper, we propose SplineGS, a framework for high-quality dynamic scene reconstruction and real-time neural rendering from inthe-wild monocular videos without relying on external camera estimators like COLMAP [42]. SplineGS introduces Motion-Adaptive Spline (MAS), based on cubic Hermite splines [2, 8], to effectively represent dynamic 3D Gaussian deformations. Our MAS consists of piecewise cubic functions defined by control points that dictate each segment's curvature and direction. Each control point is adjustable and optimized as a learnable parameter, enabling faster and more precise modeling of complex motion trajectories. Additionally, to adaptively model the trajectory of each dynamic 3D Gaussian based on motion complexity during training, we introduce a Motion-Adaptive Control" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 313, + 72, + 553, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 192 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 192 + ], + "type": "text", + "content": "points Pruning (MACP) method that adjusts the number of control points to improve rendering quality and efficiency. Furthermore, we incorporate a camera parameter estimation method jointly optimized with 3D Gaussian attributes and MAS, leveraging photometric and geometric consistency, thus eliminating the need for external estimators. Experiments show that SplineGS significantly outperforms state-of-the-art (SOTA) dynamic NVs methods both qualitatively and quantitatively, offering faster rendering speed, as shown in Fig. 1. Our contributions are as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 195, + 553, + 333 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 314, + 195, + 553, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 195, + 553, + 243 + ], + "spans": [ + { + "bbox": [ + 314, + 195, + 553, + 243 + ], + "type": "text", + "content": "- We propose a novel Spline-based dynamic 3D Gaussian Splatting framework, SplineGS, which is (i) COLMAP-free, (ii) very fast and (iii) of high quality in reconstructing the dynamic scenes from in-the-wild monocular videos;" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 246, + 553, + 282 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 246, + 553, + 282 + ], + "spans": [ + { + "bbox": [ + 313, + 246, + 553, + 282 + ], + "type": "text", + "content": "- A novel Motion-Adaptive Spline (MAS) is introduced, which can accurately and effectively represent the continuous trajectory of each dynamic 3D Gaussian;" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 285, + 553, + 333 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 285, + 553, + 333 + ], + "spans": [ + { + "bbox": [ + 313, + 285, + 553, + 333 + ], + "type": "text", + "content": "- A Motion-Adaptive Control points Pruning (MACP) method is presented, which can efficiently adjust the number of control points for each spline function, optimizing both rendering quality and the efficiency of MAS." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 345, + 400, + 357 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 345, + 400, + 357 + ], + "spans": [ + { + "bbox": [ + 314, + 345, + 400, + 357 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 365, + 555, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 365, + 555, + 521 + ], + "spans": [ + { + "bbox": [ + 313, + 365, + 555, + 521 + ], + "type": "text", + "content": "Dynamic NeRF. Recent advancements in video view synthesis have extended the static NeRF model [34] to represent scene dynamics. These include dynamic NeRFs using scene flow-based frameworks [12, 25, 26], deformation estimation in canonical fields [1, 3, 16, 30, 35, 36, 40, 44, 47, 51, 53], and 4D grid-based spatio-temporal radiance fields [4, 6, 10, 43]. Techniques such as NSFF [25], DynNeRF [12], and DynIBaR [26] combine time-independent and time-dependent radiance fields to synthesize novel spatiotemporal perspectives from monocular videos. Despite these advancements, current dynamic NeRFs still fall short of recent dynamic 3D Gaussian Splatting (3DGS) techniques in rendering quality and efficiency." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 522, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 522, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 522, + 556, + 714 + ], + "type": "text", + "content": "Dynamic 3DGS. The improvements in rendering quality and speed achieved by 3DGS [20] have inspired further studies [14, 24, 27, 31, 52, 55] to extend the static 3DGS framework to dynamic scenes by enabling the deformation of 3D Gaussian attributes. The pioneering work on dynamic 3DGS [31] introduces time-dependent offsets for the positions and rotations of dynamic 3D Gaussians via an MLP; however, this approach slows down the rendering. D3DGS [55] builds on this concept with an annealing smoothing training mechanism to improve temporal smoothness and rendering quality. 4DGS [52] replaces the MLP network with a grid-based structure to boost efficiency, though this change requires quality trade-offs due to the resolution limits of grid-based methods. In contrast, SC-GS [14] combines an MLP deformation network with sparse spatial deformation, reducing the computational" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26867" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 468 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 468 + ], + "type": "text", + "content": "cost of MLP while maintaining quality. STGS [24] introduces polynomial trajectories for motion modeling, improving speed and quality over implicit representations. CasualFVS [22] warps dynamic content from neighboring frames, and MoSca [23] proposes a graph-based motion modeling approach to handle sparse control deformation. Very recently, a concurrent work [21] proposes modeling dynamic 3D Gaussian trajectories using polynomial interpolation between fixed keyframes for NVS from multi-view videos with COLMAP [42] assistance. Unlike [21], our SplineGS adaptively optimizes the deformation of dynamic 3D Gaussians, accounting for varying motion degrees and types in in-the-wild monocular videos, without requiring preprocessed camera parameters (COLMAP-free approach). Neural Rendering without SfM Preprocessing. Accurate camera parameters, including extrinsic and intrinsic, are essential for neural rendering approaches to capture fine details [17, 28]. However, in real-world settings, camera parameters derived from Structure-from-Motion (SfM) algorithms such as COLMAP [42] often exhibit pixel-level inaccuracies [29, 41], compromising the structural details of rendered scenes [30]. To address this, several NeRF methods [28, 32, 37, 50] jointly optimize NeRF architectures and camera parameters. Recently, a local-to-global training approach [11] is introduced to optimize both camera parameters and 3D Gaussians. However, these methods are limited to static scenes. For dynamic scene reconstruction, RoDynRF [30] and MoSca [23] use motion masks to gather multi-view cues from static regions, allowing robust rendering without pre-computed camera parameters. Our SplineGS is also COLMAP-free, significantly outperforming RoDynRF [30] and MoSca [23] in rendering quality and efficiency, enabled by our novel spline-based architecture." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 476, + 255, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 476, + 255, + 491 + ], + "spans": [ + { + "bbox": [ + 55, + 476, + 255, + 491 + ], + "type": "text", + "content": "3. Preliminary: 3D Gaussian Splitting" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 496, + 295, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 496, + 295, + 520 + ], + "spans": [ + { + "bbox": [ + 55, + 496, + 295, + 520 + ], + "type": "text", + "content": "3DGS [20] represents the radiance field of a scene using anisotropic 3D Gaussians, each of which is formulated as" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 113, + 522, + 294, + 538 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 522, + 294, + 538 + ], + "spans": [ + { + "bbox": [ + 113, + 522, + 294, + 538 + ], + "type": "interline_equation", + "content": "G (\\boldsymbol {x}) = e ^ {- \\frac {1}{2} (\\boldsymbol {x} - \\boldsymbol {\\mu}) ^ {\\top} \\boldsymbol {\\Sigma} ^ {- 1} (\\boldsymbol {x} - \\boldsymbol {\\mu})} \\tag {1}", + "image_path": "c67dd3eb7fed4c34af98f2b44f94da0bf9708df383e316cebac6e0c0f8a153e7.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 541, + 295, + 697 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 541, + 295, + 697 + ], + "spans": [ + { + "bbox": [ + 55, + 541, + 295, + 697 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 541, + 295, + 697 + ], + "type": "inline_equation", + "content": "\\pmb{x} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 55, + 541, + 295, + 697 + ], + "type": "text", + "content": " denotes a 3D position, and " + }, + { + "bbox": [ + 55, + 541, + 295, + 697 + ], + "type": "inline_equation", + "content": "\\pmb{\\mu} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 55, + 541, + 295, + 697 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 541, + 295, + 697 + ], + "type": "inline_equation", + "content": "\\pmb{\\Sigma} \\in \\mathbb{R}^{3 \\times 3}" + }, + { + "bbox": [ + 55, + 541, + 295, + 697 + ], + "type": "text", + "content": " represent the mean (center) and the covariance matrix of the 3D Gaussian, respectively. To ensure that " + }, + { + "bbox": [ + 55, + 541, + 295, + 697 + ], + "type": "inline_equation", + "content": "\\pmb{\\Sigma}" + }, + { + "bbox": [ + 55, + 541, + 295, + 697 + ], + "type": "text", + "content": " is positive semi-definite and contains physical meaning, it is decomposed into a diagonal scaling matrix " + }, + { + "bbox": [ + 55, + 541, + 295, + 697 + ], + "type": "inline_equation", + "content": "S \\in \\mathbb{R}^{3 \\times 3}" + }, + { + "bbox": [ + 55, + 541, + 295, + 697 + ], + "type": "text", + "content": " of a scale vector " + }, + { + "bbox": [ + 55, + 541, + 295, + 697 + ], + "type": "inline_equation", + "content": "s \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 55, + 541, + 295, + 697 + ], + "type": "text", + "content": " and a rotation matrix " + }, + { + "bbox": [ + 55, + 541, + 295, + 697 + ], + "type": "inline_equation", + "content": "R \\in \\mathbb{R}^{3 \\times 3}" + }, + { + "bbox": [ + 55, + 541, + 295, + 697 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 55, + 541, + 295, + 697 + ], + "type": "inline_equation", + "content": "\\pmb{\\Sigma} = R\\pmb{S}\\pmb{S}^\\top\\pmb{R}^\\top" + }, + { + "bbox": [ + 55, + 541, + 295, + 697 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 541, + 295, + 697 + ], + "type": "inline_equation", + "content": "\\pmb{R}" + }, + { + "bbox": [ + 55, + 541, + 295, + 697 + ], + "type": "text", + "content": " is parameterized by a learnable unit-length quaternion " + }, + { + "bbox": [ + 55, + 541, + 295, + 697 + ], + "type": "inline_equation", + "content": "q \\in \\mathbb{R}^4" + }, + { + "bbox": [ + 55, + 541, + 295, + 697 + ], + "type": "text", + "content": ". In addition, each 3D Gaussian is parameterized by an opacity " + }, + { + "bbox": [ + 55, + 541, + 295, + 697 + ], + "type": "inline_equation", + "content": "\\sigma \\in \\mathbb{R}" + }, + { + "bbox": [ + 55, + 541, + 295, + 697 + ], + "type": "text", + "content": " and a color " + }, + { + "bbox": [ + 55, + 541, + 295, + 697 + ], + "type": "inline_equation", + "content": "c \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 55, + 541, + 295, + 697 + ], + "type": "text", + "content": ". To render the color of each pixel, the color and the opacity of each 3D Gaussian are computed using Eq. 1, and the rendered color " + }, + { + "bbox": [ + 55, + 541, + 295, + 697 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 55, + 541, + 295, + 697 + ], + "type": "text", + "content": " is computed by the alpha-blending of the " + }, + { + "bbox": [ + 55, + 541, + 295, + 697 + ], + "type": "inline_equation", + "content": "\\mathcal{N}" + }, + { + "bbox": [ + 55, + 541, + 295, + 697 + ], + "type": "text", + "content": " ordered 3D Gaussians overlapping the pixel as" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 107, + 700, + 295, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 700, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 107, + 700, + 295, + 715 + ], + "type": "interline_equation", + "content": "\\boldsymbol {C} = \\sum_ {i \\in \\mathcal {N}} \\boldsymbol {c} _ {i} \\alpha_ {i} \\prod_ {j = 1} ^ {i - 1} (1 - \\alpha_ {j}), \\tag {2}", + "image_path": "a935c4f4753c677c899cbb28f3085cf6b70e98236c942cc8faa1c32cf7db8a71.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 555, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 194 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 194 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 72, + 555, + 194 + ], + "type": "inline_equation", + "content": "c_{i}" + }, + { + "bbox": [ + 313, + 72, + 555, + 194 + ], + "type": "text", + "content": " is the color of the " + }, + { + "bbox": [ + 313, + 72, + 555, + 194 + ], + "type": "inline_equation", + "content": "i^{\\mathrm{th}}" + }, + { + "bbox": [ + 313, + 72, + 555, + 194 + ], + "type": "text", + "content": " 3D Gaussian and " + }, + { + "bbox": [ + 313, + 72, + 555, + 194 + ], + "type": "inline_equation", + "content": "\\alpha_{i}" + }, + { + "bbox": [ + 313, + 72, + 555, + 194 + ], + "type": "text", + "content": " is a density of the " + }, + { + "bbox": [ + 313, + 72, + 555, + 194 + ], + "type": "inline_equation", + "content": "i^{\\mathrm{th}}" + }, + { + "bbox": [ + 313, + 72, + 555, + 194 + ], + "type": "text", + "content": " 3D Gaussian which is given by evaluating the 2D covariance " + }, + { + "bbox": [ + 313, + 72, + 555, + 194 + ], + "type": "inline_equation", + "content": "\\pmb{\\Sigma}' \\in \\mathbb{R}^{2 \\times 2}" + }, + { + "bbox": [ + 313, + 72, + 555, + 194 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 313, + 72, + 555, + 194 + ], + "type": "inline_equation", + "content": "\\pmb{\\Sigma}'" + }, + { + "bbox": [ + 313, + 72, + 555, + 194 + ], + "type": "text", + "content": " is formulated as " + }, + { + "bbox": [ + 313, + 72, + 555, + 194 + ], + "type": "inline_equation", + "content": "\\pmb{\\Sigma}' = J\\pmb{W}\\pmb{\\Sigma}\\pmb{W}^\\top J^\\top" + }, + { + "bbox": [ + 313, + 72, + 555, + 194 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 72, + 555, + 194 + ], + "type": "inline_equation", + "content": "J" + }, + { + "bbox": [ + 313, + 72, + 555, + 194 + ], + "type": "text", + "content": " is the Jacobian of the affine approximation of the projective transformation and " + }, + { + "bbox": [ + 313, + 72, + 555, + 194 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 313, + 72, + 555, + 194 + ], + "type": "text", + "content": " is a viewing transformation matrix. Similar to prior works [23, 27], we extend 3DGS [20] to a union of static 3D Gaussians " + }, + { + "bbox": [ + 313, + 72, + 555, + 194 + ], + "type": "inline_equation", + "content": "\\{G_i^{\\mathrm{st}} | i = 1, 2, \\dots, n^{\\mathrm{st}}\\}" + }, + { + "bbox": [ + 313, + 72, + 555, + 194 + ], + "type": "text", + "content": " and dynamic 3D Gaussians " + }, + { + "bbox": [ + 313, + 72, + 555, + 194 + ], + "type": "inline_equation", + "content": "\\{G_i^{\\mathrm{dy}} | i = 1, 2, \\dots, n^{\\mathrm{dy}}\\}" + }, + { + "bbox": [ + 313, + 72, + 555, + 194 + ], + "type": "text", + "content": " to represent static backgrounds and moving objects, respectively, in dynamic scenes." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 208, + 476, + 222 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 208, + 476, + 222 + ], + "spans": [ + { + "bbox": [ + 313, + 208, + 476, + 222 + ], + "type": "text", + "content": "4. Proposed Method: SplineGS" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 228, + 441, + 242 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 228, + 441, + 242 + ], + "spans": [ + { + "bbox": [ + 313, + 228, + 441, + 242 + ], + "type": "text", + "content": "4.1. Overview of SplineGS" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 248, + 555, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 248, + 555, + 548 + ], + "spans": [ + { + "bbox": [ + 313, + 248, + 555, + 548 + ], + "type": "text", + "content": "We describe the overall architecture of SplineGS in Fig. 2. Given a monocular video " + }, + { + "bbox": [ + 313, + 248, + 555, + 548 + ], + "type": "inline_equation", + "content": "\\{I_t|t = 0,1,\\dots,N_f - 1\\}" + }, + { + "bbox": [ + 313, + 248, + 555, + 548 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 313, + 248, + 555, + 548 + ], + "type": "inline_equation", + "content": "N_{f}" + }, + { + "bbox": [ + 313, + 248, + 555, + 548 + ], + "type": "text", + "content": " is the total number of frames, SplineGS is designed to synthesize high-quality novel spatio-temporal views with fast rendering speed, and to estimate the camera parameters, including the extrinsic " + }, + { + "bbox": [ + 313, + 248, + 555, + 548 + ], + "type": "inline_equation", + "content": "[\\hat{R}_t|\\hat{T}_t]\\in \\mathbb{R}^{3\\times 4}" + }, + { + "bbox": [ + 313, + 248, + 555, + 548 + ], + "type": "text", + "content": " for each time " + }, + { + "bbox": [ + 313, + 248, + 555, + 548 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 248, + 555, + 548 + ], + "type": "text", + "content": ", and the shared intrinsic " + }, + { + "bbox": [ + 313, + 248, + 555, + 548 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{K}}\\in \\mathbb{R}^{3\\times 3}" + }, + { + "bbox": [ + 313, + 248, + 555, + 548 + ], + "type": "text", + "content": ". As shown in Fig. 2, to stabilize joint optimization of the 3D Gaussian attributes and camera parameter estimation, we adopt a two-stage optimization process consisting of a warm-up stage and a main training stage for our SplineGS architecture. In the warm-up stage, we coarsely optimize the camera parameters " + }, + { + "bbox": [ + 313, + 248, + 555, + 548 + ], + "type": "inline_equation", + "content": "[\\hat{R}_t|\\hat{T}_t]" + }, + { + "bbox": [ + 313, + 248, + 555, + 548 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 248, + 555, + 548 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{K}}" + }, + { + "bbox": [ + 313, + 248, + 555, + 548 + ], + "type": "text", + "content": " using photometric and geometric consistency. In the main training stage, we initialize the 3D Gaussians based on the estimated camera poses and jointly optimize the 3D Gaussian attributes with the camera parameters. Specifically, for each dynamic 3D Gaussian, we propose a novel spline-based deformation modeling method, called Motion-Adaptive Spline (MAS), which utilizes a cubic Hermite spline function [2, 8] to accurately model the continuous trajectory with a small number of control points. For MAS, we introduce a Motion-Adaptive Control points Pruning (MACP) method that effectively adjusts the number of control points for each dynamic 3D Gaussian, considering the object's motion types and degrees." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 550, + 554, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 550, + 554, + 597 + ], + "spans": [ + { + "bbox": [ + 313, + 550, + 554, + 597 + ], + "type": "text", + "content": "In the following sections, we detail the process of our MAS method in Sec. 4.2, followed by the camera parameter estimation process in Sec. 4.3. Finally, we describe the overall optimization process in Sec. 4.4." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 609, + 533, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 609, + 533, + 621 + ], + "spans": [ + { + "bbox": [ + 313, + 609, + 533, + 621 + ], + "type": "text", + "content": "4.2. Motion-Adaptive Spline for 3D Gaussians" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 628, + 554, + 699 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 628, + 554, + 699 + ], + "spans": [ + { + "bbox": [ + 313, + 628, + 554, + 699 + ], + "type": "text", + "content": "To represent the continuous trajectory of each dynamic 3D Gaussian for moving objects over time, we propose MAS which modifies the mean parameter to a set of learnable control points. For this, we use the cubic Hermite spline function [2, 8] to model the mean of each dynamic 3D Gaussian at time " + }, + { + "bbox": [ + 313, + 628, + 554, + 699 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 628, + 554, + 699 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 400, + 701, + 555, + 714 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 400, + 701, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 400, + 701, + 555, + 714 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\mu} (t) = S (t, \\mathbf {P}), \\tag {3}", + "image_path": "f6a199c37b50929830bdde29611fdf042d6050dc5be7d4f0095622f52f293a6d.jpg" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "26868" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 55, + 68, + 556, + 320 + ], + "blocks": [ + { + "bbox": [ + 55, + 68, + 556, + 320 + ], + "lines": [ + { + "bbox": [ + 55, + 68, + 556, + 320 + ], + "spans": [ + { + "bbox": [ + 55, + 68, + 556, + 320 + ], + "type": "image", + "image_path": "a6d6d9f6497404325d227490123d5c8dce1c9fbb7b4e83b61409e0591cf28809.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 324, + 555, + 370 + ], + "lines": [ + { + "bbox": [ + 54, + 324, + 555, + 370 + ], + "spans": [ + { + "bbox": [ + 54, + 324, + 555, + 370 + ], + "type": "text", + "content": "Figure 2. Overview of SplineGS. Our SplineGS leverages spline-based functions to model the deformation of dynamic 3D Gaussians with a novel Motion-Adaptive Spline (MAS) architecture. It is composed of sets of learnable control points based on a cubic Hermite spline function [2, 8] to accurately model the trajectory of each dynamic 3D Gaussian and to achieve faster rendering speed. To avoid any preprocessing of camera parameters, i.e., COLMAP-free, we adopt a two-stage optimization: warm-up and main training stages." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 375, + 295, + 412 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 375, + 295, + 412 + ], + "spans": [ + { + "bbox": [ + 55, + 375, + 295, + 412 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 375, + 295, + 412 + ], + "type": "inline_equation", + "content": "\\mathbf{P} = \\{\\mathbf{p}_k|\\mathbf{p}_k\\in \\mathbb{R}^3,k\\in 0,\\dots ,N_c - 1\\}" + }, + { + "bbox": [ + 55, + 375, + 295, + 412 + ], + "type": "text", + "content": " is a set of " + }, + { + "bbox": [ + 55, + 375, + 295, + 412 + ], + "type": "inline_equation", + "content": "N_{c}" + }, + { + "bbox": [ + 55, + 375, + 295, + 412 + ], + "type": "text", + "content": " learnable control points, serving as an additional attribute for each dynamic 3D Gaussian, and " + }, + { + "bbox": [ + 55, + 375, + 295, + 412 + ], + "type": "inline_equation", + "content": "S(t,\\mathbf{P})" + }, + { + "bbox": [ + 55, + 375, + 295, + 412 + ], + "type": "text", + "content": " is defined as" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 64, + 413, + 294, + 444 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 413, + 294, + 444 + ], + "spans": [ + { + "bbox": [ + 64, + 413, + 294, + 444 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} S (t, \\mathbf {P}) = \\left(2 t _ {r} ^ {3} - 3 t _ {r} ^ {2} + 1\\right) \\mathbf {p} _ {\\lfloor t _ {s} \\rfloor} + \\left(t _ {r} ^ {3} - 2 t _ {r} ^ {2} + t _ {r}\\right) \\mathbf {m} _ {\\lfloor t _ {s} \\rfloor} \\\\ + \\left(- 2 t _ {r} ^ {3} + 3 t _ {r} ^ {2}\\right) \\mathbf {p} _ {[ t _ {s} ] + 1} + \\left(t _ {r} ^ {3} - t _ {r} ^ {2}\\right) \\mathbf {m} _ {[ t _ {s} ] + 1}, \\tag {4} \\\\ \\end{array}", + "image_path": "0487339f7277b636cd74db56dc4c054b6abde55cc7202bfdc9f3c709133d138b.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 446, + 278, + 459 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 446, + 278, + 459 + ], + "spans": [ + { + "bbox": [ + 61, + 446, + 278, + 459 + ], + "type": "interline_equation", + "content": "t _ {r} = t _ {s} - \\lfloor t _ {s} \\rfloor , \\quad t _ {s} = t _ {n} (N _ {c} - 1), \\quad t _ {n} = t / (N _ {f} - 1),", + "image_path": "e737807fbebed5594aa1c26e20bc7b7d5c0258f646dce02fb971a4560243d5bd.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 460, + 295, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 460, + 295, + 544 + ], + "spans": [ + { + "bbox": [ + 55, + 460, + 295, + 544 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 460, + 295, + 544 + ], + "type": "inline_equation", + "content": "\\mathbf{m}_k" + }, + { + "bbox": [ + 55, + 460, + 295, + 544 + ], + "type": "text", + "content": " is the approximated tangent of the control point " + }, + { + "bbox": [ + 55, + 460, + 295, + 544 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_k" + }, + { + "bbox": [ + 55, + 460, + 295, + 544 + ], + "type": "text", + "content": ", formulated as " + }, + { + "bbox": [ + 55, + 460, + 295, + 544 + ], + "type": "inline_equation", + "content": "\\mathbf{m}_k = (\\mathbf{p}_{k + 1} - \\mathbf{p}_{k - 1}) / 2" + }, + { + "bbox": [ + 55, + 460, + 295, + 544 + ], + "type": "text", + "content": ". Note that " + }, + { + "bbox": [ + 55, + 460, + 295, + 544 + ], + "type": "inline_equation", + "content": "S(t,\\mathbf{P})" + }, + { + "bbox": [ + 55, + 460, + 295, + 544 + ], + "type": "text", + "content": " indicates the piecewise cubic function to represent the whole continuous trajectory of each dynamic 3D Gaussian, and " + }, + { + "bbox": [ + 55, + 460, + 295, + 544 + ], + "type": "inline_equation", + "content": "N_{c}" + }, + { + "bbox": [ + 55, + 460, + 295, + 544 + ], + "type": "text", + "content": " can be different from " + }, + { + "bbox": [ + 55, + 460, + 295, + 544 + ], + "type": "inline_equation", + "content": "N_{f}" + }, + { + "bbox": [ + 55, + 460, + 295, + 544 + ], + "type": "text", + "content": ". The optimal " + }, + { + "bbox": [ + 55, + 460, + 295, + 544 + ], + "type": "inline_equation", + "content": "N_{c}" + }, + { + "bbox": [ + 55, + 460, + 295, + 544 + ], + "type": "text", + "content": " can be estimated by Motion-Adaptive Control points Pruning (MACP) in the following section." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 544, + 295, + 699 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 544, + 295, + 699 + ], + "spans": [ + { + "bbox": [ + 55, + 544, + 295, + 699 + ], + "type": "text", + "content": "Control Points Initialization. To stably optimize " + }, + { + "bbox": [ + 55, + 544, + 295, + 699 + ], + "type": "inline_equation", + "content": "S(t, \\mathbf{P})" + }, + { + "bbox": [ + 55, + 544, + 295, + 699 + ], + "type": "text", + "content": ", it is essential to initialize the control points with appropriate geometric considerations that ensure temporal consistency. For this, we leverage the long-range 2D tracking [19] and the metric depth [38] priors. Let " + }, + { + "bbox": [ + 55, + 544, + 295, + 699 + ], + "type": "inline_equation", + "content": "\\mathcal{T} = \\{\\varphi_t^{\\mathrm{tr}} | \\varphi_t^{\\mathrm{tr}} \\in \\mathbb{R}^2, t = 0, \\dots, N_f - 1\\}" + }, + { + "bbox": [ + 55, + 544, + 295, + 699 + ], + "type": "text", + "content": " be a 2D track, " + }, + { + "bbox": [ + 55, + 544, + 295, + 699 + ], + "type": "inline_equation", + "content": "\\pi_K(\\cdot)" + }, + { + "bbox": [ + 55, + 544, + 295, + 699 + ], + "type": "text", + "content": " be a projection function from the camera space to the image space with the camera intrinsic " + }, + { + "bbox": [ + 55, + 544, + 295, + 699 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 544, + 295, + 699 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 544, + 295, + 699 + ], + "type": "inline_equation", + "content": "\\varphi_t^{\\mathrm{tr}}" + }, + { + "bbox": [ + 55, + 544, + 295, + 699 + ], + "type": "text", + "content": " be a pixel coordinate corresponding to the 2D track " + }, + { + "bbox": [ + 55, + 544, + 295, + 699 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 55, + 544, + 295, + 699 + ], + "type": "text", + "content": " at time " + }, + { + "bbox": [ + 55, + 544, + 295, + 699 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 544, + 295, + 699 + ], + "type": "text", + "content": ". We unproject each 2D track " + }, + { + "bbox": [ + 55, + 544, + 295, + 699 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 55, + 544, + 295, + 699 + ], + "type": "text", + "content": " to the world space to compute the 3D track aided by the frame's metric depth " + }, + { + "bbox": [ + 55, + 544, + 295, + 699 + ], + "type": "inline_equation", + "content": "d_t" + }, + { + "bbox": [ + 55, + 544, + 295, + 699 + ], + "type": "text", + "content": " and camera extrinsic " + }, + { + "bbox": [ + 55, + 544, + 295, + 699 + ], + "type": "inline_equation", + "content": "[\\hat{R}_t | \\hat{T}_t]" + }, + { + "bbox": [ + 55, + 544, + 295, + 699 + ], + "type": "text", + "content": " at time " + }, + { + "bbox": [ + 55, + 544, + 295, + 699 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 544, + 295, + 699 + ], + "type": "text", + "content": ". The unprojection function " + }, + { + "bbox": [ + 55, + 544, + 295, + 699 + ], + "type": "inline_equation", + "content": "\\mathcal{W}_t(\\cdot)" + }, + { + "bbox": [ + 55, + 544, + 295, + 699 + ], + "type": "text", + "content": " from the image space to the world space is formulated as" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 86, + 700, + 295, + 718 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 700, + 295, + 718 + ], + "spans": [ + { + "bbox": [ + 86, + 700, + 295, + 718 + ], + "type": "interline_equation", + "content": "\\mathcal {W} _ {t} \\left(\\boldsymbol {\\varphi} _ {t} ^ {\\mathrm {t r}}\\right) = \\hat {\\boldsymbol {R}} _ {t} ^ {\\top} \\pi_ {\\hat {\\boldsymbol {K}}} ^ {- 1} \\left(\\boldsymbol {\\varphi} _ {t} ^ {\\mathrm {t r}}, \\boldsymbol {d} _ {t} \\left(\\boldsymbol {\\varphi} _ {t} ^ {\\mathrm {t r}}\\right)\\right) - \\hat {\\boldsymbol {R}} _ {t} ^ {\\top} \\hat {\\boldsymbol {T}} _ {t}. \\tag {5}", + "image_path": "90a5ca46e9e1976b9604a596f648ea155c5274b77cff856aff54a31e4f3bc960.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 375, + 555, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 375, + 555, + 448 + ], + "spans": [ + { + "bbox": [ + 313, + 375, + 555, + 448 + ], + "type": "text", + "content": "It should be noted that we estimate " + }, + { + "bbox": [ + 313, + 375, + 555, + 448 + ], + "type": "inline_equation", + "content": "[\\hat{\\pmb{R}}_t|\\hat{\\pmb{T}}_t]" + }, + { + "bbox": [ + 313, + 375, + 555, + 448 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 375, + 555, + 448 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{K}}" + }, + { + "bbox": [ + 313, + 375, + 555, + 448 + ], + "type": "text", + "content": " from a sequence of frames only, without using any given ground truth values, as mentioned in Sec. 4.3. Then, we initialize the per-Gaussian control points set " + }, + { + "bbox": [ + 313, + 375, + 555, + 448 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 313, + 375, + 555, + 448 + ], + "type": "text", + "content": ", by a least-square (LS) approximation, such that the spline-described curve fits the initial curve by the tracker, given by" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 359, + 451, + 555, + 471 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 451, + 555, + 471 + ], + "spans": [ + { + "bbox": [ + 359, + 451, + 555, + 471 + ], + "type": "interline_equation", + "content": "\\min _ {\\mathbf {P}} \\sum_ {t = 0} ^ {N _ {f} - 1} \\left\\| \\mathcal {W} _ {t} \\left(\\varphi_ {t} ^ {\\mathrm {t r}}\\right) - S (t, \\mathbf {P}) \\right\\| _ {2} ^ {2}. \\tag {6}", + "image_path": "4d395022f0bc0209c6719a959718a293bcc64a67f0a8eb474cd3357243f0f833.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 474, + 556, + 702 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 474, + 556, + 702 + ], + "spans": [ + { + "bbox": [ + 313, + 474, + 556, + 702 + ], + "type": "text", + "content": "Motion-Adaptive Control Points Pruning (MACP). An excessive number of control points may cause over-fitting and decrease the processing speed of our MAS module, leading to poorer rendering qualities with reduced rendering speed (see Table 3). Furthermore, as each scene exhibits various types and degrees of motion for moving objects, the number of control points required for each dynamic 3D Gaussian trajectory should be adaptively adjusted to accommodate the scene dynamics. To achieve this, we propose the MACP method for MAS, which can generate sparser control points while ensuring no dynamic modeling degradation. MACP is designed on top of 3D Gaussian densification [20], but focuses on optimizing the number of control points in MAS. Our MACP computes a new spline function " + }, + { + "bbox": [ + 313, + 474, + 556, + 702 + ], + "type": "inline_equation", + "content": "S(t,\\mathbf{P}^{\\prime})" + }, + { + "bbox": [ + 313, + 474, + 556, + 702 + ], + "type": "text", + "content": " after every 3D Gaussian densification step, where " + }, + { + "bbox": [ + 313, + 474, + 556, + 702 + ], + "type": "inline_equation", + "content": "\\mathbf{P}^{\\prime} = \\{\\mathbf{p}_{l}^{\\prime}|\\mathbf{p}_{l}^{\\prime}\\in \\mathbb{R}^{3}, l = 0,\\dots ,N_{c} - 2\\}" + }, + { + "bbox": [ + 313, + 474, + 556, + 702 + ], + "type": "text", + "content": " is a set of " + }, + { + "bbox": [ + 313, + 474, + 556, + 702 + ], + "type": "inline_equation", + "content": "N_{c} - 1" + }, + { + "bbox": [ + 313, + 474, + 556, + 702 + ], + "type": "text", + "content": " control points, which contains one-fewer control points than the current set " + }, + { + "bbox": [ + 313, + 474, + 556, + 702 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 313, + 474, + 556, + 702 + ], + "type": "text", + "content": " of control points. We compute the LS approximation to find the reduced set " + }, + { + "bbox": [ + 313, + 474, + 556, + 702 + ], + "type": "inline_equation", + "content": "\\mathbf{P}^{\\prime}" + }, + { + "bbox": [ + 313, + 474, + 556, + 702 + ], + "type": "text", + "content": " of control points as" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "26869" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 102, + 69, + 295, + 90 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 69, + 295, + 90 + ], + "spans": [ + { + "bbox": [ + 102, + 69, + 295, + 90 + ], + "type": "interline_equation", + "content": "\\min _ {\\mathbf {P} ^ {\\prime}} \\sum_ {t = 0} ^ {N _ {f} - 1} \\left\\| S (t, \\mathbf {P}) - S (t, \\mathbf {P} ^ {\\prime}) \\right\\| _ {2} ^ {2}. \\tag {7}", + "image_path": "0558d682d39fe42ede70847c93e042a83fda8ded783c0f2149d14f56ed43557c.jpg" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 92, + 295, + 129 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 92, + 295, + 129 + ], + "spans": [ + { + "bbox": [ + 55, + 92, + 295, + 129 + ], + "type": "text", + "content": "Then, the updated optimal set " + }, + { + "bbox": [ + 55, + 92, + 295, + 129 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 55, + 92, + 295, + 129 + ], + "type": "text", + "content": " of control points is assigned with the values of " + }, + { + "bbox": [ + 55, + 92, + 295, + 129 + ], + "type": "inline_equation", + "content": "\\mathbf{P}'" + }, + { + "bbox": [ + 55, + 92, + 295, + 129 + ], + "type": "text", + "content": " if the error " + }, + { + "bbox": [ + 55, + 92, + 295, + 129 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 55, + 92, + 295, + 129 + ], + "type": "text", + "content": " between " + }, + { + "bbox": [ + 55, + 92, + 295, + 129 + ], + "type": "inline_equation", + "content": "S(t,\\mathbf{P})" + }, + { + "bbox": [ + 55, + 92, + 295, + 129 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 92, + 295, + 129 + ], + "type": "inline_equation", + "content": "S(t,\\mathbf{P}')" + }, + { + "bbox": [ + 55, + 92, + 295, + 129 + ], + "type": "text", + "content": " is smaller than a threshold " + }, + { + "bbox": [ + 55, + 92, + 295, + 129 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 55, + 92, + 295, + 129 + ], + "type": "text", + "content": ", as given by" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 110, + 133, + 295, + 165 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 133, + 295, + 165 + ], + "spans": [ + { + "bbox": [ + 110, + 133, + 295, + 165 + ], + "type": "interline_equation", + "content": "\\mathbf {P} = \\left\\{ \\begin{array}{l l} \\mathbf {P} ^ {\\prime}, & \\text {i f} E < \\epsilon \\\\ \\mathbf {P}, & \\text {o t h e r w i s e} \\end{array} , \\right. \\text {w h e r e} \\tag {8}", + "image_path": "7e794589564b6286c9583cbe124b6afc289083f47be85ac89095f4d085df6f94.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 161, + 277, + 190 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 161, + 277, + 190 + ], + "spans": [ + { + "bbox": [ + 56, + 161, + 277, + 190 + ], + "type": "interline_equation", + "content": "E = \\frac {1}{N _ {f}} \\sum_ {t = 0} ^ {N _ {f} - 1} \\| \\pi_ {\\hat {\\mathbf {K}}} (\\hat {\\mathbf {R}} _ {t} S (t, \\mathbf {P}) + \\hat {\\mathbf {T}} _ {t}) - \\pi_ {\\hat {\\mathbf {K}}} (\\hat {\\mathbf {R}} _ {t} S (t, \\mathbf {P} ^ {\\prime}) + \\hat {\\mathbf {T}} _ {t}) \\| _ {2} ^ {2}.", + "image_path": "3fbc31ec02844c96ef4bcfa095da55610a4ff097b2c4a13be2bc0fe7eb09a626.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 198, + 296, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 198, + 296, + 258 + ], + "spans": [ + { + "bbox": [ + 55, + 198, + 296, + 258 + ], + "type": "text", + "content": "By following MACP, each dynamic 3D Gaussian is allowed to have a different number of control points. Therefore, the MACP can guide our MAS to have a minimal number of control points when modeling simple motion, while having more control points for more complex motions (see Fig. 8)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 266, + 222, + 277 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 266, + 222, + 277 + ], + "spans": [ + { + "bbox": [ + 55, + 266, + 222, + 277 + ], + "type": "text", + "content": "4.3. Camera Parameter Estimation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 285, + 296, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 285, + 296, + 380 + ], + "spans": [ + { + "bbox": [ + 55, + 285, + 296, + 380 + ], + "type": "text", + "content": "The traditional SfM methods, such as COLMAP [42], fail to reliably estimate camera parameters in dynamic scenes from in-the-wild monocular videos [30]. For this reason, we propose to estimate camera parameters for joint optimization with the 3D Gaussian attributes. For each frame at time " + }, + { + "bbox": [ + 55, + 285, + 296, + 380 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 285, + 296, + 380 + ], + "type": "text", + "content": ", we predict a rotation matrix " + }, + { + "bbox": [ + 55, + 285, + 296, + 380 + ], + "type": "inline_equation", + "content": "\\hat{R}_t \\in \\mathbb{R}^{3 \\times 3}" + }, + { + "bbox": [ + 55, + 285, + 296, + 380 + ], + "type": "text", + "content": " and a translation vector " + }, + { + "bbox": [ + 55, + 285, + 296, + 380 + ], + "type": "inline_equation", + "content": "\\hat{T}_t \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 55, + 285, + 296, + 380 + ], + "type": "text", + "content": ", representing the extrinsic parameters of a monocular camera, using a shallow MLP " + }, + { + "bbox": [ + 55, + 285, + 296, + 380 + ], + "type": "inline_equation", + "content": "F_{\\theta}" + }, + { + "bbox": [ + 55, + 285, + 296, + 380 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 384, + 294, + 398 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 384, + 294, + 398 + ], + "spans": [ + { + "bbox": [ + 133, + 384, + 294, + 398 + ], + "type": "interline_equation", + "content": "\\left[ \\boldsymbol {\\hat {R}} _ {t} \\mid \\boldsymbol {\\hat {T}} _ {t} \\right] = F _ {\\theta} (\\gamma (t)), \\tag {9}", + "image_path": "53c70294c7b914d55a4c56180e62e82146f2f2cab9ff13cefe7a5e0565b6d510.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 402, + 296, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 402, + 296, + 499 + ], + "spans": [ + { + "bbox": [ + 55, + 402, + 296, + 499 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 402, + 296, + 499 + ], + "type": "inline_equation", + "content": "\\gamma (\\cdot)" + }, + { + "bbox": [ + 55, + 402, + 296, + 499 + ], + "type": "text", + "content": " is a positional encoding [34]. We also predict the focal length " + }, + { + "bbox": [ + 55, + 402, + 296, + 499 + ], + "type": "inline_equation", + "content": "\\hat{f}" + }, + { + "bbox": [ + 55, + 402, + 296, + 499 + ], + "type": "text", + "content": " of the camera intrinsic " + }, + { + "bbox": [ + 55, + 402, + 296, + 499 + ], + "type": "inline_equation", + "content": "\\hat{K}" + }, + { + "bbox": [ + 55, + 402, + 296, + 499 + ], + "type": "text", + "content": " as a learnable parameter that is shared across all frames in the monocular video. To accurately optimize the camera parameters, we enforce two types of consistency—photometric and geometric—for the static background between random reference cameras " + }, + { + "bbox": [ + 55, + 402, + 296, + 499 + ], + "type": "inline_equation", + "content": "[\\hat{R}_{t_{\\mathrm{ref}}}|\\hat{T}_{t_{\\mathrm{ref}}}]" + }, + { + "bbox": [ + 55, + 402, + 296, + 499 + ], + "type": "text", + "content": " and the target camera " + }, + { + "bbox": [ + 55, + 402, + 296, + 499 + ], + "type": "inline_equation", + "content": "[\\hat{R}_t|\\hat{T}_t]" + }, + { + "bbox": [ + 55, + 402, + 296, + 499 + ], + "type": "text", + "content": ", encouraging alignment both visually and structurally." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 499, + 296, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 499, + 296, + 571 + ], + "spans": [ + { + "bbox": [ + 55, + 499, + 296, + 571 + ], + "type": "text", + "content": "Photometric Consistency. Given the pre-computed metric depth [38], the camera intrinsics and extrinsics under optimization will converge as long as the projected reference frame's color " + }, + { + "bbox": [ + 55, + 499, + 296, + 571 + ], + "type": "inline_equation", + "content": "I_{t_{\\mathrm{ref}}}(\\varphi_{t \\to t_{\\mathrm{ref}}})" + }, + { + "bbox": [ + 55, + 499, + 296, + 571 + ], + "type": "text", + "content": " is aligned to the target frame's color " + }, + { + "bbox": [ + 55, + 499, + 296, + 571 + ], + "type": "inline_equation", + "content": "I_t(\\varphi_t)" + }, + { + "bbox": [ + 55, + 499, + 296, + 571 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 55, + 499, + 296, + 571 + ], + "type": "inline_equation", + "content": "\\varphi_{t \\to t_{\\mathrm{ref}}}" + }, + { + "bbox": [ + 55, + 499, + 296, + 571 + ], + "type": "text", + "content": " is the reference pixel coordinate corresponding to the target frame's pixel coordinate " + }, + { + "bbox": [ + 55, + 499, + 296, + 571 + ], + "type": "inline_equation", + "content": "\\varphi_t" + }, + { + "bbox": [ + 55, + 499, + 296, + 571 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 575, + 295, + 590 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 575, + 295, + 590 + ], + "spans": [ + { + "bbox": [ + 62, + 575, + 295, + 590 + ], + "type": "interline_equation", + "content": "\\varphi_ {t \\rightarrow t _ {\\text {r e f}}} = \\pi_ {\\hat {\\mathbf {K}}} \\left(\\hat {\\mathbf {R}} _ {t _ {\\text {r e f}}} \\left(\\hat {\\mathbf {R}} _ {t} ^ {\\top} \\pi_ {\\hat {\\mathbf {K}}} ^ {- 1} \\left(\\varphi_ {t}, \\boldsymbol {d} _ {t} (\\varphi_ {t})\\right) - \\hat {\\mathbf {R}} _ {t} ^ {\\top} \\hat {\\boldsymbol {T}} _ {t}\\right) + \\hat {\\boldsymbol {T}} _ {t _ {\\text {r e f}}}\\right). \\tag {10}", + "image_path": "d1061a71b35bc9182f0217408b7905be6cae7f73a48255b1467cd13aca9cb8a8.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 594, + 295, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 617 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 617 + ], + "type": "text", + "content": "We refer to such projection alignment as photometric consistency, which is encouraged by the loss " + }, + { + "bbox": [ + 55, + 594, + 295, + 617 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{pc}}" + }, + { + "bbox": [ + 55, + 594, + 295, + 617 + ], + "type": "text", + "content": " given by" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 66, + 621, + 295, + 638 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 621, + 295, + 638 + ], + "spans": [ + { + "bbox": [ + 66, + 621, + 295, + 638 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {p c}} = \\sum_ {\\boldsymbol {\\varphi} _ {t}} \\left\\| \\boldsymbol {M} _ {t, t _ {\\text {r e f}}} \\left(\\boldsymbol {\\varphi} _ {t}\\right) \\odot \\left(\\boldsymbol {I} _ {t} \\left(\\boldsymbol {\\varphi} _ {t}\\right) - \\boldsymbol {I} _ {t _ {\\text {r e f}}} \\left(\\boldsymbol {\\varphi} _ {t \\rightarrow t _ {\\text {r e f}}}\\right)\\right)\\right\\| _ {2} ^ {2}, \\tag {11}", + "image_path": "217012f62b0218da943ad77ffc2c9a2166da414ad45e6f5fe326bb9ff70bb428.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 646, + 296, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 646, + 296, + 694 + ], + "spans": [ + { + "bbox": [ + 55, + 646, + 296, + 694 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 646, + 296, + 694 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 55, + 646, + 296, + 694 + ], + "type": "text", + "content": " is the Hadamard product [13], and " + }, + { + "bbox": [ + 55, + 646, + 296, + 694 + ], + "type": "inline_equation", + "content": "M_{t,t_{\\mathrm{ref}}}" + }, + { + "bbox": [ + 55, + 646, + 296, + 694 + ], + "type": "text", + "content": " is a union motion mask that excludes dynamic objects in " + }, + { + "bbox": [ + 55, + 646, + 296, + 694 + ], + "type": "inline_equation", + "content": "I_{t}" + }, + { + "bbox": [ + 55, + 646, + 296, + 694 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 646, + 296, + 694 + ], + "type": "inline_equation", + "content": "I_{t_{\\mathrm{ref}}}" + }, + { + "bbox": [ + 55, + 646, + 296, + 694 + ], + "type": "text", + "content": " for removing the inconsistencies due to dynamic regions, which is given by" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 98, + 700, + 295, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 98, + 700, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 98, + 700, + 295, + 715 + ], + "type": "interline_equation", + "content": "\\boldsymbol {M} _ {t, t _ {\\text {r e f}}} \\left(\\varphi_ {t}\\right) = \\boldsymbol {M} _ {t} \\left(\\varphi_ {t}\\right) \\boldsymbol {M} _ {t _ {\\text {r e f}}} \\left(\\varphi_ {t \\rightarrow t _ {\\text {r e f}}}\\right), \\tag {12}", + "image_path": "49b81c52bf95d6f60e952eeb0e06b12a82a2024d32669c9ce2ba3870cef091e8.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 72, + 553, + 107 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 107 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 107 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 72, + 553, + 107 + ], + "type": "inline_equation", + "content": "M_{t}" + }, + { + "bbox": [ + 313, + 72, + 553, + 107 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 72, + 553, + 107 + ], + "type": "inline_equation", + "content": "M_{t_{\\mathrm{ref}}}" + }, + { + "bbox": [ + 313, + 72, + 553, + 107 + ], + "type": "text", + "content": " are pre-computed motion masks [54] of " + }, + { + "bbox": [ + 313, + 72, + 553, + 107 + ], + "type": "inline_equation", + "content": "\\pmb{I}_t" + }, + { + "bbox": [ + 313, + 72, + 553, + 107 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 72, + 553, + 107 + ], + "type": "inline_equation", + "content": "I_{t_{\\mathrm{ref}}}" + }, + { + "bbox": [ + 313, + 72, + 553, + 107 + ], + "type": "text", + "content": ", respectively. Note that motion mask's value is 0 for static regions, and 1 otherwise." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 108, + 553, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 108, + 553, + 156 + ], + "spans": [ + { + "bbox": [ + 313, + 108, + 553, + 156 + ], + "type": "text", + "content": "Geometric Consistency. Along with the photometric consistency, we compute the geometric consistency of unprojected pixels in 3D space to make our optimization more stable. The geometric consistency loss is formulated as" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 319, + 159, + 553, + 175 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 159, + 553, + 175 + ], + "spans": [ + { + "bbox": [ + 319, + 159, + 553, + 175 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {g c}} = \\sum_ {\\boldsymbol {\\varphi} _ {t}} \\| M _ {t, t _ {\\text {r e f}}} (\\boldsymbol {\\varphi} _ {t}) \\odot \\left(\\mathcal {W} _ {t} (\\boldsymbol {\\varphi} _ {t}) - \\mathcal {W} _ {t _ {\\text {r e f}}} \\left(\\boldsymbol {\\varphi} _ {t \\rightarrow t _ {\\text {r e f}}}\\right)\\right) \\| _ {2} ^ {2}, \\tag {13}", + "image_path": "a02c4b967c94537ed97f839cc2792e94554f878f3ff08b6e507310e09830f02f.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 177, + 553, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 177, + 553, + 201 + ], + "spans": [ + { + "bbox": [ + 313, + 177, + 553, + 201 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 177, + 553, + 201 + ], + "type": "inline_equation", + "content": "\\mathcal{W}_t(\\varphi_t)" + }, + { + "bbox": [ + 313, + 177, + 553, + 201 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 177, + 553, + 201 + ], + "type": "inline_equation", + "content": "\\mathcal{W}_{t_{\\mathrm{ref}}}(\\varphi_{t\\to t_{\\mathrm{ref}}})" + }, + { + "bbox": [ + 313, + 177, + 553, + 201 + ], + "type": "text", + "content": " are the corresponding 3D locations of " + }, + { + "bbox": [ + 313, + 177, + 553, + 201 + ], + "type": "inline_equation", + "content": "\\varphi_t" + }, + { + "bbox": [ + 313, + 177, + 553, + 201 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 177, + 553, + 201 + ], + "type": "inline_equation", + "content": "\\varphi_{t\\rightarrow t_{\\mathrm{ref}}}" + }, + { + "bbox": [ + 313, + 177, + 553, + 201 + ], + "type": "text", + "content": ", respectively (see Eq. 5)." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 207, + 400, + 220 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 207, + 400, + 220 + ], + "spans": [ + { + "bbox": [ + 313, + 207, + 400, + 220 + ], + "type": "text", + "content": "4.4. Optimization" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 225, + 554, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 225, + 554, + 297 + ], + "spans": [ + { + "bbox": [ + 313, + 225, + 554, + 297 + ], + "type": "text", + "content": "To stabilize joint training of the MAS and the camera parameter estimation, we adopt a two-stage optimization process consisting of the warm-up stage and the main training stage. In the warm-up stage, we optimize only the camera parameters using the photometric and geometric consistency. The total loss in the warm-up stage is given by" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 380, + 302, + 553, + 316 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 380, + 302, + 553, + 316 + ], + "spans": [ + { + "bbox": [ + 380, + 302, + 553, + 316 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {t o t a l}} ^ {\\text {w a r m}} = \\lambda_ {\\mathrm {p c}} \\mathcal {L} _ {\\mathrm {p c}} + \\lambda_ {\\mathrm {g c}} \\mathcal {L} _ {\\mathrm {g c}}. \\tag {14}", + "image_path": "02da380d800487d2ead28bac2d15de718a3985030b68591d4ac5cf7b16d88eea.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 321, + 554, + 404 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 321, + 554, + 404 + ], + "spans": [ + { + "bbox": [ + 313, + 321, + 554, + 404 + ], + "type": "text", + "content": "After the warm-up stage, we obtain the coarsely predicted camera intrinsic " + }, + { + "bbox": [ + 313, + 321, + 554, + 404 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{K}}" + }, + { + "bbox": [ + 313, + 321, + 554, + 404 + ], + "type": "text", + "content": " and the set of extrinsics " + }, + { + "bbox": [ + 313, + 321, + 554, + 404 + ], + "type": "inline_equation", + "content": "[\\hat{\\pmb{R}}|\\hat{\\pmb{T}}]" + }, + { + "bbox": [ + 313, + 321, + 554, + 404 + ], + "type": "text", + "content": " for all frames, which are then used to initialize the set " + }, + { + "bbox": [ + 313, + 321, + 554, + 404 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 313, + 321, + 554, + 404 + ], + "type": "text", + "content": " of control points for each dynamic 3D Gaussian, as described in Sec. 4.2. In the main training stage, we jointly optimize the static and dynamic 3D Gaussians along with the camera parameter estimation based on the total loss function as" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 356, + 409, + 512, + 438 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 356, + 409, + 512, + 438 + ], + "spans": [ + { + "bbox": [ + 356, + 409, + 512, + 438 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {\\text {t o t a l}} ^ {\\text {m a i n}} = \\lambda_ {\\mathrm {r g b}} \\mathcal {L} _ {\\mathrm {r g b}} + \\lambda_ {\\mathrm {d}} \\mathcal {L} _ {\\mathrm {d}} + \\lambda_ {\\mathrm {M}} \\mathcal {L} _ {\\mathrm {M}} \\\\ + \\lambda_ {\\mathrm {p c}} \\mathcal {L} _ {\\mathrm {p c}} + \\lambda_ {\\mathrm {d - p c}} \\mathcal {L} _ {\\mathrm {d - p c}} + \\lambda_ {\\mathrm {g c}} \\mathcal {L} _ {\\mathrm {g c}}, \\\\ \\end{array}", + "image_path": "97895dab3a954121a147fa8b9a58be73948e7aef801e8895ec491413aef36c52.jpg" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 535, + 418, + 553, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 535, + 418, + 553, + 429 + ], + "spans": [ + { + "bbox": [ + 535, + 418, + 553, + 429 + ], + "type": "text", + "content": "(15)" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 442, + 554, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 442, + 554, + 538 + ], + "spans": [ + { + "bbox": [ + 313, + 442, + 554, + 538 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 442, + 554, + 538 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{rgb}}" + }, + { + "bbox": [ + 313, + 442, + 554, + 538 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 442, + 554, + 538 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{d}}" + }, + { + "bbox": [ + 313, + 442, + 554, + 538 + ], + "type": "text", + "content": " are the L1 losses between the rendered frame and GT frame, and between the rendered depth and GT depth, respectively. In the main training stage, we compute an additional photometric consistency loss " + }, + { + "bbox": [ + 313, + 442, + 554, + 538 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{d - pc}}" + }, + { + "bbox": [ + 313, + 442, + 554, + 538 + ], + "type": "text", + "content": " that utilizes the rendered depth " + }, + { + "bbox": [ + 313, + 442, + 554, + 538 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{d}}" + }, + { + "bbox": [ + 313, + 442, + 554, + 538 + ], + "type": "text", + "content": " of the 3D Gaussians instead of the metric depth [38] prior " + }, + { + "bbox": [ + 313, + 442, + 554, + 538 + ], + "type": "inline_equation", + "content": "\\pmb{d}" + }, + { + "bbox": [ + 313, + 442, + 554, + 538 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 313, + 442, + 554, + 538 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{d}}" + }, + { + "bbox": [ + 313, + 442, + 554, + 538 + ], + "type": "text", + "content": " allows the estimated 3D Gaussian geometry to guide the joint optimization of the camera parameters and 3D Gaussian attributes." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 540, + 554, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 540, + 554, + 635 + ], + "spans": [ + { + "bbox": [ + 313, + 540, + 554, + 635 + ], + "type": "text", + "content": "In addition to the camera parameter estimation and imagery reconstruction losses, we adopt a binary dice loss [45] " + }, + { + "bbox": [ + 313, + 540, + 554, + 635 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{M}}" + }, + { + "bbox": [ + 313, + 540, + 554, + 635 + ], + "type": "text", + "content": " between the pre-computed motion mask " + }, + { + "bbox": [ + 313, + 540, + 554, + 635 + ], + "type": "inline_equation", + "content": "M_{t}" + }, + { + "bbox": [ + 313, + 540, + 554, + 635 + ], + "type": "text", + "content": " [54] and the rendered motion mask " + }, + { + "bbox": [ + 313, + 540, + 554, + 635 + ], + "type": "inline_equation", + "content": "\\hat{M}_t" + }, + { + "bbox": [ + 313, + 540, + 554, + 635 + ], + "type": "text", + "content": " that can be derived from the dynamic 3D Gaussians. The binary dice loss initially proposed in [45] for highly imbalanced segmentation of medical imagery helps encouraging better separation between our dynamic and static 3D Gaussians as described by" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 334, + 639, + 553, + 675 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 334, + 639, + 553, + 675 + ], + "spans": [ + { + "bbox": [ + 334, + 639, + 553, + 675 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {M}} = 1 - \\frac {2 \\sum_ {\\varphi_ {t}} M _ {t} \\left(\\varphi_ {t}\\right) \\hat {M} _ {t} \\left(\\varphi_ {t}\\right) + \\varepsilon}{\\sum_ {\\varphi_ {t}} \\left(M _ {t} \\left(\\varphi_ {t}\\right) + \\hat {M} _ {t} \\left(\\varphi_ {t}\\right)\\right) + \\varepsilon}, \\tag {16}", + "image_path": "9d228a19dcba929cc6359d84fd8772b8ce501646f144b159a858da299e19079c.jpg" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 313, + 677, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 553, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 677, + 553, + 714 + ], + "type": "inline_equation", + "content": "\\varepsilon" + }, + { + "bbox": [ + 313, + 677, + 553, + 714 + ], + "type": "text", + "content": " is a smooth term to avoid numerical issues. " + }, + { + "bbox": [ + 313, + 677, + 553, + 714 + ], + "type": "inline_equation", + "content": "\\hat{M}_t(\\varphi_t)" + }, + { + "bbox": [ + 313, + 677, + 553, + 714 + ], + "type": "text", + "content": " is computed by the alpha-blending of the 3D Gaussians overlapping " + }, + { + "bbox": [ + 313, + 677, + 553, + 714 + ], + "type": "inline_equation", + "content": "\\varphi_t" + }, + { + "bbox": [ + 313, + 677, + 553, + 714 + ], + "type": "text", + "content": " (similar to Eq. 2) as" + } + ] + } + ], + "index": 28 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "26870" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 71, + 558, + 266 + ], + "blocks": [ + { + "bbox": [ + 58, + 71, + 558, + 266 + ], + "lines": [ + { + "bbox": [ + 58, + 71, + 558, + 266 + ], + "spans": [ + { + "bbox": [ + 58, + 71, + 558, + 266 + ], + "type": "table", + "html": "
MethodsJumpingSkatingTruckUmbrellaBalloon1Balloon2PlaygroundAverageFPS↑Tr. (h)↓
STM preprocessingDynNeRF (ICCV'21) [12]24.68 / 0.09032.66 / 0.03528.56 / 0.08223.26 / 0.13722.36 / 0.10427.06 / 0.04924.15 / 0.08026.10 / 0.0820.05>24
MonoNeRF (ICCV'23) [46]24.26 / 0.09132.06 / 0.04427.56 / 0.11523.62 / 0.18021.89 / 0.12927.36 / 0.05222.61 / 0.13025.62 / 0.1060.05N/A
STGS (CVPR'24) [24]20.82 / 0.18724.80 / 0.10925.01 / 0.10321.88 / 0.19520.36 / 0.19623.12 / 0.12419.23 / 0.15122.17 / 0.1529000.2
SCGS (CVPR'24) [14]15.68 / 0.92014.88 / 0.90823.81 / 0.14021.84 / 0.16020.17 / 0.17921.07 / 0.14920.71 / 0.11519.74 / 0.3671101.1
D3DGS (CVPR'24) [55]22.02 / 0.26624.06 / 0.22723.04 / 0.24722.67 / 0.19221.22 / 0.20225.86 / 0.11822.30 / 0.11123.02 / 0.195251.0
4DGS (CVPR'24) [52]22.37 / 0.17826.72 / 0.08425.93 / 0.09722.36 / 0.17821.89 / 0.15324.85 / 0.08121.36 / 0.08923.64 / 0.123950.25
SP-GS (ICML'24) [48]22.13 / 0.46829.21 / 0.23627.38 / 0.19024.88 / 0.32324.36 / 0.18029.65 / 0.09722.29 / 0.23425.70 / 0.2471806.0
Casual-FVS (ECCV'24) [22]23.45 / 0.10029.98 / 0.04525.22 / 0.09023.24 / 0.09623.76 / 0.07924.15 / 0.08122.19 / 0.07424.57 / 0.081480.25
Kinematic Fields (ECCV'24) [15]23.74 / 0.10231.89 / 0.03528.34 / 0.07425.46 / 0.09823.72 / 0.07926.49 / 0.05524.64 / 0.05326.33 / 0.0710.37.0
E-D3DGS (ECCV'24) [5]22.39 / 0.18024.62 / 0.13726.85 / 0.08123.60 / 0.14520.87 / 0.16124.39 / 0.07320.79 / 0.12423.36 / 0.129452.6
CTNeRF (Pattern Recognition'24) [33]24.35 / 0.09433.51 / 0.03428.27 / 0.08423.48 / 0.12922.19 / 0.11126.86 / 0.04824.28 / 0.07726.13 / 0.0820.1>24
D-NPC (Eurographics'25) [18]24.51 / 0.11630.22 / 0.06128.92 / 0.08424.15 / 0.13622.26 / 0.16125.84 / 0.10723.59 / 0.09925.64 / 0.109700.27
Shape-of-Motion (arXiv'24) [49]22.74 / 0.12829.12 / 0.04824.63 / 0.12623.79 / 0.08922.74 / 0.12025.83 / 0.07220.61 / 0.13024.21 / 0.1022551.1
Ex4DGS (NeurIPS'24) [21]18.93 / 0.32121.92 / 0.23319.04 / 0.30819.03 / 0.34014.69 / 0.50316.29 / 0.45714.16 / 0.43717.72 / 0.371840.65
DynPoint (NeurIPS'23) [58]24.69 / 0.09731.34 / 0.04529.30 / 0.06124.59 / 0.08622.77 / 0.09927.63 / 0.04925.37 / 0.03926.53 / 0.0680.80.5
RoDynRF (CVPR'23) [30] w/25.66 / 0.07128.68 / 0.04029.13 / 0.06324.26 / 0.08922.37 / 0.10326.19 / 0.05424.96 / 0.04825.89 / 0.0670.45>24
MoSca (arXiv'24) [23] w/25.21 / 0.08332.77 / 0.03328.22 / 0.09024.41 / 0.09223.26 / 0.09228.90 / 0.04223.05 / 0.06026.55 / 0.070400.8
SplineGS (Ours) w/25.62 / 0.06434.10 / 0.03427.83 / 0.04824.60 / 0.07622.67 / 0.12029.04 / 0.03024.47 / 0.04626.90 / 0.0604000.5
STM FreeRoDynRF (CVPR'23) [30]24.27 / 0.10028.71 / 0.04628.85 / 0.06623.25 / 0.10421.81 / 0.12225.58 / 0.06425.20 / 0.05225.38 / 0.0790.45>24
MoSca (arXiv'24) [23]25.43 / 0.08032.62 / 0.03328.29 / 0.08624.40 / 0.09123.27 / 0.09129.01 / 0.04223.23 / 0.05826.61 / 0.069401.0
SplineGS (Ours)25.50 / 0.06833.72 / 0.03128.66 / 0.05625.61 / 0.07124.43 / 0.06828.37 / 0.03224.19 / 0.04727.21 / 0.0534001.5
", + "image_path": "27e9de0ea01ad7c7617e03d4515dcf7ce2e20a8e1184c8ac2db648f58748d2da.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 270, + 555, + 315 + ], + "lines": [ + { + "bbox": [ + 55, + 270, + 555, + 315 + ], + "spans": [ + { + "bbox": [ + 55, + 270, + 555, + 315 + ], + "type": "text", + "content": "Table 1. Novel view synthesis evaluation on the NVIDIA dataset (PSNR↑ / LPIPS↓). Red and Blue denote the best and second-best performances, respectively. For Casual-FVS [22], we directly use the results from their paper, as the official code is unavailable. 'w/' denotes the use of camera poses obtained by COLMAP [42] for each SfM-free method. 'Tr' denotes the training time. For MonoNeRF [46], the training time is omitted as it is not a per-scene optimized model. We use an RTX 3090 Ti GPU for each method." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 57, + 327, + 140, + 433 + ], + "blocks": [ + { + "bbox": [ + 57, + 327, + 140, + 433 + ], + "lines": [ + { + "bbox": [ + 57, + 327, + 140, + 433 + ], + "spans": [ + { + "bbox": [ + 57, + 327, + 140, + 433 + ], + "type": "image", + "image_path": "3b07c8fae3cafce0df17a4fade182ac546c315331fc6996fa4469cefdc9215cf.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 140, + 327, + 223, + 432 + ], + "blocks": [ + { + "bbox": [ + 140, + 327, + 223, + 432 + ], + "lines": [ + { + "bbox": [ + 140, + 327, + 223, + 432 + ], + "spans": [ + { + "bbox": [ + 140, + 327, + 223, + 432 + ], + "type": "image", + "image_path": "38617b43d9f060febfdb6fa5867c65025a47bdfb2c0ff7de68b745b51f5672de.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 155, + 434, + 454, + 445 + ], + "lines": [ + { + "bbox": [ + 155, + 434, + 454, + 445 + ], + "spans": [ + { + "bbox": [ + 155, + 434, + 454, + 445 + ], + "type": "text", + "content": "Figure 3. Visual comparisons for novel view synthesis on the NVIDIA dataset." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 223, + 327, + 306, + 432 + ], + "blocks": [ + { + "bbox": [ + 223, + 327, + 306, + 432 + ], + "lines": [ + { + "bbox": [ + 223, + 327, + 306, + 432 + ], + "spans": [ + { + "bbox": [ + 223, + 327, + 306, + 432 + ], + "type": "image", + "image_path": "6ba8827c50382e4b70c4d1fdd83446629df95241eb1b996c99a610fa6cc9bc7a.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 306, + 327, + 388, + 432 + ], + "blocks": [ + { + "bbox": [ + 306, + 327, + 388, + 432 + ], + "lines": [ + { + "bbox": [ + 306, + 327, + 388, + 432 + ], + "spans": [ + { + "bbox": [ + 306, + 327, + 388, + 432 + ], + "type": "image", + "image_path": "16689226469fb74279850fc2dccd0a6dd21d2163f52f5250576e68bc67c877da.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 388, + 327, + 470, + 431 + ], + "blocks": [ + { + "bbox": [ + 388, + 327, + 470, + 431 + ], + "lines": [ + { + "bbox": [ + 388, + 327, + 470, + 431 + ], + "spans": [ + { + "bbox": [ + 388, + 327, + 470, + 431 + ], + "type": "image", + "image_path": "e14840e2fa731fc70224196f9043d0bea774865c2273b7735c1c49219e5b5699.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 470, + 327, + 554, + 431 + ], + "blocks": [ + { + "bbox": [ + 470, + 327, + 554, + 431 + ], + "lines": [ + { + "bbox": [ + 470, + 327, + 554, + 431 + ], + "spans": [ + { + "bbox": [ + 470, + 327, + 554, + 431 + ], + "type": "image", + "image_path": "0bf5619da990411a9a9b77e7ce10578fb9d15d42e695b295083021deb49e1bf8.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 93, + 462, + 295, + 478 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 462, + 295, + 478 + ], + "spans": [ + { + "bbox": [ + 93, + 462, + 295, + 478 + ], + "type": "interline_equation", + "content": "\\hat {M} _ {t} (\\boldsymbol {\\varphi} _ {t}) = \\sum_ {i \\in \\mathcal {N}} m _ {i} \\alpha_ {i} \\prod_ {j = 1} ^ {i - 1} (1 - \\alpha_ {j}), \\tag {17}", + "image_path": "618ca1788b0b2a6d38db91816eab32f30cd8832b628fb8672de1836aa05ac430.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 480, + 295, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 480, + 295, + 504 + ], + "spans": [ + { + "bbox": [ + 55, + 480, + 295, + 504 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 480, + 295, + 504 + ], + "type": "inline_equation", + "content": "m_{i} = 0" + }, + { + "bbox": [ + 55, + 480, + 295, + 504 + ], + "type": "text", + "content": " if the " + }, + { + "bbox": [ + 55, + 480, + 295, + 504 + ], + "type": "inline_equation", + "content": "i^{\\mathrm{th}}" + }, + { + "bbox": [ + 55, + 480, + 295, + 504 + ], + "type": "text", + "content": " 3D Gaussian is the static 3D Gaussian and " + }, + { + "bbox": [ + 55, + 480, + 295, + 504 + ], + "type": "inline_equation", + "content": "m_{i} = 1" + }, + { + "bbox": [ + 55, + 480, + 295, + 504 + ], + "type": "text", + "content": " otherwise." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 505, + 296, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 505, + 296, + 577 + ], + "spans": [ + { + "bbox": [ + 55, + 505, + 296, + 577 + ], + "type": "text", + "content": "In conjunction, all terms in " + }, + { + "bbox": [ + 55, + 505, + 296, + 577 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{total}}^{\\mathrm{main}}" + }, + { + "bbox": [ + 55, + 505, + 296, + 577 + ], + "type": "text", + "content": " guide our SplineGS to effectively and efficiently model dynamic 3D scenes from pure monocular videos, achieving more structural details and better temporal consistency than previous works [12, 14, 21-24, 30, 46, 52, 55], without relying on camera parameters obtained from external estimators [42]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 586, + 137, + 600 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 586, + 137, + 600 + ], + "spans": [ + { + "bbox": [ + 55, + 586, + 137, + 600 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 606, + 295, + 630 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 606, + 295, + 630 + ], + "spans": [ + { + "bbox": [ + 55, + 606, + 295, + 630 + ], + "type": "text", + "content": "Implementation Details. We refer readers to the Supplementary material for the details of our implementation." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 631, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 631, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 631, + 296, + 715 + ], + "type": "text", + "content": "Datasets. We evaluate both the quantitative and qualitative performance of novel view and time synthesis on the widely used NVIDIA dataset [56] which features challenging monocular videos. Additionally, we assess novel view synthesis performance on in-the-wild monocular videos from the DAVIS dataset [39] which contains an average of 69.7 frames per video sequence." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 451, + 539, + 464 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 451, + 539, + 464 + ], + "spans": [ + { + "bbox": [ + 313, + 451, + 539, + 464 + ], + "type": "text", + "content": "5.1. Comparison with State-of-the-Art Methods" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 474, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 474, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 474, + 555, + 715 + ], + "type": "text", + "content": "Novel View Synthesis. Table 1 presents a quantitative comparison of NVs between our SplineGS and existing SfM-based [5, 12, 14, 15, 18, 21, 22, 24, 33, 46, 48, 49, 52, 55, 58] and SfM-free [23, 30] methods on the NVIDIA dataset [56]. For this comparison, we follow the evaluation configuration in [30]. The results demonstrate that our SplineGS significantly outperforms SOTA methods in both the PSNR and LPIPS [57] metrics. Notably, SplineGS achieves " + }, + { + "bbox": [ + 313, + 474, + 555, + 715 + ], + "type": "inline_equation", + "content": "890 \\times" + }, + { + "bbox": [ + 313, + 474, + 555, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 474, + 555, + 715 + ], + "type": "inline_equation", + "content": "8,000 \\times" + }, + { + "bbox": [ + 313, + 474, + 555, + 715 + ], + "type": "text", + "content": " faster rendering speed compared to RoDynRF [30] and DynNeRF [12], respectively. Ex4DGS [21] and STGS [24], which are designed for multi-view settings, face challenges with inconsistent geometry alignment over time when trained on monocular videos. Fig. 3 shows the qualitative comparison between our SplineGS with the existing methods in [12, 24, 30, 55]. As highlighted by the red boxes, our method yields not only higher rendering quality but also dynamic objects more aligned and closer to the ground truth. In Fig. 4, we show our superior NVs results for in-the-wild monocular videos from the DAVIS dataset [39] compared to the existing methods in [24, 30, 55]." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26871" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 70, + 176, + 225 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 176, + 225 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 176, + 225 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 176, + 225 + ], + "type": "image", + "image_path": "ef1c7b7d4cfa7a812fe38d3b2ec250d706f34704c0ec3604d2071aa27dd325dd.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 228, + 295, + 250 + ], + "lines": [ + { + "bbox": [ + 55, + 228, + 295, + 250 + ], + "spans": [ + { + "bbox": [ + 55, + 228, + 295, + 250 + ], + "type": "text", + "content": "Figure 4. Visual comparisons for novel view synthesis on the DAVIS dataset." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 175, + 70, + 295, + 225 + ], + "blocks": [ + { + "bbox": [ + 175, + 70, + 295, + 225 + ], + "lines": [ + { + "bbox": [ + 175, + 70, + 295, + 225 + ], + "spans": [ + { + "bbox": [ + 175, + 70, + 295, + 225 + ], + "type": "image", + "image_path": "d8e053f2754a6f9dc97640ef988db60489fa8cb7466334a822cc06fe7d895420.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 269, + 295, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 269, + 295, + 389 + ], + "spans": [ + { + "bbox": [ + 54, + 269, + 295, + 389 + ], + "type": "text", + "content": "Compared to [30] that is also COLMAP-free, SplineGS yields considerably more detailed novel views (red boxes) as shown in Fig. 4. For the other methods [24, 55], we observe that COLMAP [42] fails to recover camera parameters and initial point clouds on the DAVIS dataset [39], as also claimed in [30]. On the other hand, our COLMAP-free SplineGS reconstructs accurate camera parameters that are the ones actually used to train [24, 55], which are shown in Fig. 4 for comparison. More results are in the Supplementary material." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 391, + 295, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 391, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 391, + 295, + 712 + ], + "type": "text", + "content": "Novel View and Time Synthesis. To evaluate the capability of SplineGS to model continuous trajectories of moving objects in a scene, we compare the novel view and time synthesis results of SplineGS with those of NeRF-based [12, 30] and 3DGS-based [24, 52, 55] methods. For this evaluation, we follow the dataset sampling strategy in [25], which samples 24 timestamps from the NVIDIA dataset [56]. In addition, to simulate a larger motion, we exclude frames with odd time indices in the training sets. To ensure all test timestamps are not seen during training, and thus, to create a more challenging novel view and time synthesis validation, we exclude frames with even time indices in the test sets. Table 2 and Fig. 5 show the quantitative and qualitative comparisons for this challenging experiment. We observe that the NeRF-based methods, RoDynRF [30] and DynNeRF [12], generate inconsistent artifacts and blurriness for unseen times. Furthermore, the 3DGS-based methods [24, 52, 55] yield even more significant degradation when predicting unseen time indices. In contrast, SplineGS, built upon 3DGS [20] and equipped with our novel spline-based deformation, provides SOTA novel view rendering at unseen intermediate timestamps. Thanks to our MAS, SplineGS naturally and precisely captures the continuous trajectories of moving objects over time, enhancing the temporal consistency of rendered scenes that can be checked in tOF scores [7] of Table 2." + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 317, + 70, + 550, + 142 + ], + "blocks": [ + { + "bbox": [ + 317, + 70, + 550, + 142 + ], + "lines": [ + { + "bbox": [ + 317, + 70, + 550, + 142 + ], + "spans": [ + { + "bbox": [ + 317, + 70, + 550, + 142 + ], + "type": "table", + "html": "
MethodsPSNR↑LPIPS↓tOF↓
SfM preprocessingDynNeRF (ICCV'21) [12]23.360.2190.921
4DGS (CVPR'24) [52]17.070.4596.314
D3DGS (CVPR'24) [55]19.630.3433.225
STGS (CVPR'24) [24]15.720.4742.105
SfM-FreeRoDynRF (CVPR'23) [30]21.580.2212.138
SplineGS (Ours)25.920.0980.703
", + "image_path": "9fb0400f985f0cbc98130b6744709d2c7551767e3c75827f7d583fda1f7c40ab.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 145, + 553, + 167 + ], + "lines": [ + { + "bbox": [ + 314, + 145, + 553, + 167 + ], + "spans": [ + { + "bbox": [ + 314, + 145, + 553, + 167 + ], + "type": "text", + "content": "Table 2. Novel view and time synthesis evaluation on the NVIDIA dataset." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 315, + 178, + 379, + 255 + ], + "blocks": [ + { + "bbox": [ + 315, + 178, + 379, + 255 + ], + "lines": [ + { + "bbox": [ + 315, + 178, + 379, + 255 + ], + "spans": [ + { + "bbox": [ + 315, + 178, + 379, + 255 + ], + "type": "image", + "image_path": "d5104621fc11f3df0a7975772d82ae75ab57e7fa3b8cebc5800f69a13c77acfc.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 258, + 553, + 279 + ], + "lines": [ + { + "bbox": [ + 313, + 258, + 553, + 279 + ], + "spans": [ + { + "bbox": [ + 313, + 258, + 553, + 279 + ], + "type": "text", + "content": "Figure 5. Visual comparisons for novel view and time synthesis on the NVIDIA dataset." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 380, + 178, + 430, + 255 + ], + "blocks": [ + { + "bbox": [ + 380, + 178, + 430, + 255 + ], + "lines": [ + { + "bbox": [ + 380, + 178, + 430, + 255 + ], + "spans": [ + { + "bbox": [ + 380, + 178, + 430, + 255 + ], + "type": "image", + "image_path": "ac5c8eb84a529459c8cb420935f2c1478dab1eb4920cdbd17b3c25d108accdcf.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 431, + 178, + 483, + 255 + ], + "blocks": [ + { + "bbox": [ + 431, + 178, + 483, + 255 + ], + "lines": [ + { + "bbox": [ + 431, + 178, + 483, + 255 + ], + "spans": [ + { + "bbox": [ + 431, + 178, + 483, + 255 + ], + "type": "image", + "image_path": "5a130a8da916272055ea4e6bf9a03ca29408682b2a171bb3df995cabc0538a57.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 484, + 178, + 517, + 255 + ], + "blocks": [ + { + "bbox": [ + 484, + 178, + 517, + 255 + ], + "lines": [ + { + "bbox": [ + 484, + 178, + 517, + 255 + ], + "spans": [ + { + "bbox": [ + 484, + 178, + 517, + 255 + ], + "type": "image", + "image_path": "1d91ca2faf8697bb02bd4a12b80da090c4edff8049dd9529384b2bb290d4e787.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 518, + 178, + 552, + 255 + ], + "blocks": [ + { + "bbox": [ + 518, + 178, + 552, + 255 + ], + "lines": [ + { + "bbox": [ + 518, + 178, + 552, + 255 + ], + "spans": [ + { + "bbox": [ + 518, + 178, + 552, + 255 + ], + "type": "image", + "image_path": "a7a933d4bd704964aecbad34995d477aa98a24f5d2c7fb44b2ce353e04afd924.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 288, + 555, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 288, + 555, + 419 + ], + "spans": [ + { + "bbox": [ + 313, + 288, + 555, + 419 + ], + "type": "text", + "content": "To further analyze the SplineGS's ability to model continuous trajectories of dynamic 3D Gaussians, we visualize the projected 2D motion tracking of dynamic objects in pixel space, comparing it with D3DGS [55] and STGS [24], as shown in Fig. 6. We observe that D3DGS [55] and STGS [24] cannot provide reliable motion tracking for moving objects, underscoring their limitations in modeling continuous trajectories of dynamic 3D Gaussians. In contrast, SplineGS provides accurate motion tracking, demonstrating the effectiveness of MAS for deforming dynamic 3D Gaussians. Further analysis is in the Supplementary material." + } + ] + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 316, + 422, + 395, + 494 + ], + "blocks": [ + { + "bbox": [ + 316, + 422, + 395, + 494 + ], + "lines": [ + { + "bbox": [ + 316, + 422, + 395, + 494 + ], + "spans": [ + { + "bbox": [ + 316, + 422, + 395, + 494 + ], + "type": "image", + "image_path": "9401d0320810a7442011e77ff7e49989879b328c1ece6439499d28b7b43aa4b0.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 497, + 553, + 519 + ], + "lines": [ + { + "bbox": [ + 313, + 497, + 553, + 519 + ], + "spans": [ + { + "bbox": [ + 313, + 497, + 553, + 519 + ], + "type": "text", + "content": "Figure 6. Visual comparisons for motion tracking. We visualize 2D pixel tracks to analyze motions of dynamic 3D Gaussians." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 395, + 422, + 474, + 495 + ], + "blocks": [ + { + "bbox": [ + 395, + 422, + 474, + 495 + ], + "lines": [ + { + "bbox": [ + 395, + 422, + 474, + 495 + ], + "spans": [ + { + "bbox": [ + 395, + 422, + 474, + 495 + ], + "type": "image", + "image_path": "140c61669249bd5e0f14e8957127219e4dd2c1c41eb0934d5aac4584855a0708.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 474, + 422, + 553, + 495 + ], + "blocks": [ + { + "bbox": [ + 474, + 422, + 553, + 495 + ], + "lines": [ + { + "bbox": [ + 474, + 422, + 553, + 495 + ], + "spans": [ + { + "bbox": [ + 474, + 422, + 553, + 495 + ], + "type": "image", + "image_path": "7a466f6f93ee90b8328922bb4b7ed4dfbc2d5cc42d26580ed99fe67c8c639ede.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 527, + 408, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 527, + 408, + 540 + ], + "spans": [ + { + "bbox": [ + 313, + 527, + 408, + 540 + ], + "type": "text", + "content": "5.2. Ablation Study" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 545, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 555, + 715 + ], + "type": "text", + "content": "Motion-Adaptive Spline (MAS). To demonstrate the effectiveness of MAS, we replace the MAS model with various deformation models, including an MLP, a grid-based model, polynomial functions of third degree (denoted as 'Poly " + }, + { + "bbox": [ + 313, + 545, + 555, + 715 + ], + "type": "inline_equation", + "content": "(3^{\\mathrm{rd}})" + }, + { + "bbox": [ + 313, + 545, + 555, + 715 + ], + "type": "text", + "content": "') and tenth degree (denoted as 'Poly " + }, + { + "bbox": [ + 313, + 545, + 555, + 715 + ], + "type": "inline_equation", + "content": "(10^{\\mathrm{th}})" + }, + { + "bbox": [ + 313, + 545, + 555, + 715 + ], + "type": "text", + "content": "'), and a Bezier curve [9]. For the MLP, grid-based model, and polynomial functions, we apply to them the structures similar to those in prior works, including D3DGS [55], 4DGS [52], and STGS [24], respectively. Additionally, we implement the Bezier curve [9], a commonly used method for curve modeling in computer graphics. Table 3-(a) presents quantitative comparisons of each 3D Gaussian trajectory model, focusing on rendering quality (PSNR, LPIPS) and deformation latency per Gaussian, denoted as " + }, + { + "bbox": [ + 313, + 545, + 555, + 715 + ], + "type": "inline_equation", + "content": "g_{\\mathrm{def}}" + }, + { + "bbox": [ + 313, + 545, + 555, + 715 + ], + "type": "text", + "content": ". This latency re" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26872" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 69, + 295, + 149 + ], + "blocks": [ + { + "bbox": [ + 57, + 69, + 295, + 149 + ], + "lines": [ + { + "bbox": [ + 57, + 69, + 295, + 149 + ], + "spans": [ + { + "bbox": [ + 57, + 69, + 295, + 149 + ], + "type": "image", + "image_path": "dae64395a9f2a2f04571df543f94a4015fef14c8bba9911c8c798a497b5e2a13.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 68, + 152, + 282, + 163 + ], + "lines": [ + { + "bbox": [ + 68, + 152, + 282, + 163 + ], + "spans": [ + { + "bbox": [ + 68, + 152, + 282, + 163 + ], + "type": "text", + "content": "Figure 7. Visual comparisons for MACP ablation study." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 175, + 296, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 175, + 296, + 389 + ], + "spans": [ + { + "bbox": [ + 55, + 175, + 296, + 389 + ], + "type": "text", + "content": "flects the computational time required to estimate the deformation of a single dynamic 3D Gaussian. As shown in Table 3-(a), our MAS model achieves superior rendering quality compared to all other deformation models. Consistent with analyses in previous works [24, 52, 55], we observe that the MLP and grid-based architectures require substantial computational costs for rendering. Among these methods, 'Poly " + }, + { + "bbox": [ + 55, + 175, + 296, + 389 + ], + "type": "inline_equation", + "content": "(3^{\\mathrm{rd}})" + }, + { + "bbox": [ + 55, + 175, + 296, + 389 + ], + "type": "text", + "content": "', as implemented in [24], demonstrates the best latency. However, fixed-degree polynomial functions have limited flexibility across varying motion complexities which adversely impacts rendering performance. To explore this further, we experiment with 'Poly " + }, + { + "bbox": [ + 55, + 175, + 296, + 389 + ], + "type": "inline_equation", + "content": "(10^{\\mathrm{th}})" + }, + { + "bbox": [ + 55, + 175, + 296, + 389 + ], + "type": "text", + "content": "' to assess changes in modeling capability. This adjustment, however, leads to noisier optimization and reduced efficiency, as variables under high exponents in 'Poly " + }, + { + "bbox": [ + 55, + 175, + 296, + 389 + ], + "type": "inline_equation", + "content": "(10^{\\mathrm{th}})" + }, + { + "bbox": [ + 55, + 175, + 296, + 389 + ], + "type": "text", + "content": "' lead to numerical instability. The Bézier curve [9] offers the second-best rendering quality, but its latency remains higher than our MAS due to its recursive nature of computation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 391, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 391, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 391, + 296, + 714 + ], + "type": "text", + "content": "Motion-Adaptive Control Points Pruning (MACP). To assess the effectiveness of our MACP technique for MAS, we compare our full model with MACP against other versions of our model with two fixed numbers of control points " + }, + { + "bbox": [ + 55, + 391, + 296, + 714 + ], + "type": "inline_equation", + "content": "N_{c} = 4" + }, + { + "bbox": [ + 55, + 391, + 296, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 391, + 296, + 714 + ], + "type": "inline_equation", + "content": "N_{c} = N_{f}" + }, + { + "bbox": [ + 55, + 391, + 296, + 714 + ], + "type": "text", + "content": ". As shown in Table 3-(c) and Fig. 7, our SplineGS with MACP achieves a good trade-off between rendering quality and " + }, + { + "bbox": [ + 55, + 391, + 296, + 714 + ], + "type": "inline_equation", + "content": "g_{\\mathrm{def}}" + }, + { + "bbox": [ + 55, + 391, + 296, + 714 + ], + "type": "text", + "content": " compared to the ablated models with fixed " + }, + { + "bbox": [ + 55, + 391, + 296, + 714 + ], + "type": "inline_equation", + "content": "N_{c}" + }, + { + "bbox": [ + 55, + 391, + 296, + 714 + ], + "type": "text", + "content": ". Using " + }, + { + "bbox": [ + 55, + 391, + 296, + 714 + ], + "type": "inline_equation", + "content": "N_{c} = 4" + }, + { + "bbox": [ + 55, + 391, + 296, + 714 + ], + "type": "text", + "content": " for every dynamic 3D Gaussian limits the motion modeling capacity of MAS, resulting in significantly lower metrics and visible artifacts in the dynamic regions. Moreover, an excessive " + }, + { + "bbox": [ + 55, + 391, + 296, + 714 + ], + "type": "inline_equation", + "content": "N_{c} = N_{f}" + }, + { + "bbox": [ + 55, + 391, + 296, + 714 + ], + "type": "text", + "content": " decreases the rendering speed of our MAS module and still falls short of the quality achieved by our full model with MACP, potentially due to motion overfitting. Fig. 8 shows the distribution of " + }, + { + "bbox": [ + 55, + 391, + 296, + 714 + ], + "type": "inline_equation", + "content": "N_{c}" + }, + { + "bbox": [ + 55, + 391, + 296, + 714 + ], + "type": "text", + "content": " values after optimization with MACP across scenes of varying motion complexities. In Fig. 8-(a), we visualize ' " + }, + { + "bbox": [ + 55, + 391, + 296, + 714 + ], + "type": "inline_equation", + "content": "N_{c}" + }, + { + "bbox": [ + 55, + 391, + 296, + 714 + ], + "type": "text", + "content": " Heatmap' that contains the pixel-wise averaged " + }, + { + "bbox": [ + 55, + 391, + 296, + 714 + ], + "type": "inline_equation", + "content": "N_{c}" + }, + { + "bbox": [ + 55, + 391, + 296, + 714 + ], + "type": "text", + "content": " values of the dynamic 3D Gaussians needed to render the 2D pixels (red higher, blue lower number of control points). As shown, the simpler motions in these scenes, such as those of the human bodies, can be modeled by the dynamic 3D Gaussians with smaller averages of " + }, + { + "bbox": [ + 55, + 391, + 296, + 714 + ], + "type": "inline_equation", + "content": "N_{c}" + }, + { + "bbox": [ + 55, + 391, + 296, + 714 + ], + "type": "text", + "content": " values, whereas the objects with complex and extensive motions (e.g., balloon) require higher averaged " + }, + { + "bbox": [ + 55, + 391, + 296, + 714 + ], + "type": "inline_equation", + "content": "N_{c}" + }, + { + "bbox": [ + 55, + 391, + 296, + 714 + ], + "type": "text", + "content": " values. Fig. 8-(b) presents the corresponding histogram of " + }, + { + "bbox": [ + 55, + 391, + 296, + 714 + ], + "type": "inline_equation", + "content": "N_{c}" + }, + { + "bbox": [ + 55, + 391, + 296, + 714 + ], + "type": "text", + "content": " values for the scenes' dynamic 3D Gaussians. For sequences with simple motion, such as 'Skating', the trajectories of most" + } + ] + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 317, + 83, + 435, + 145 + ], + "blocks": [ + { + "bbox": [ + 325, + 71, + 426, + 82 + ], + "lines": [ + { + "bbox": [ + 325, + 71, + 426, + 82 + ], + "spans": [ + { + "bbox": [ + 325, + 71, + 426, + 82 + ], + "type": "text", + "content": "(a) Motion-Adaptive Spline" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 83, + 435, + 145 + ], + "lines": [ + { + "bbox": [ + 317, + 83, + 435, + 145 + ], + "spans": [ + { + "bbox": [ + 317, + 83, + 435, + 145 + ], + "type": "table", + "html": "
PSNR↑LPIPS↓gDef (ns)↓
MLP23.510.125149.41
Grid25.480.09098.89
Poly (3rd)25.140.1111.80
Poly (10th)24.380.1207.71
Bézier27.190.0608.78
Ours27.210.0535.63
", + "image_path": "fa54ab5df1327c9b623b6dfcb7124c3c38b50a47e5fab954d4b70e1e5cb0240e.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 447, + 83, + 542, + 145 + ], + "blocks": [ + { + "bbox": [ + 462, + 71, + 525, + 81 + ], + "lines": [ + { + "bbox": [ + 462, + 71, + 525, + 81 + ], + "spans": [ + { + "bbox": [ + 462, + 71, + 525, + 81 + ], + "type": "text", + "content": "(b) Loss function" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 447, + 83, + 542, + 145 + ], + "lines": [ + { + "bbox": [ + 447, + 83, + 542, + 145 + ], + "spans": [ + { + "bbox": [ + 447, + 83, + 542, + 145 + ], + "type": "table", + "html": "
PSNR↑LPIPS↓
w/o Lpc17.490.853
w/o Lgc26.330.067
w/o Ld-pc26.180.066
w/o Lm26.340.088
Ours27.210.053
", + "image_path": "42a70a77ee571120e41223d0fa28c060c5c6c39d8319ff099c991d72fd53782e.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 352, + 160, + 517, + 206 + ], + "blocks": [ + { + "bbox": [ + 354, + 148, + 515, + 158 + ], + "lines": [ + { + "bbox": [ + 354, + 148, + 515, + 158 + ], + "spans": [ + { + "bbox": [ + 354, + 148, + 515, + 158 + ], + "type": "text", + "content": "(c) Motion-Adaptive Control points Pruning" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 352, + 160, + 517, + 206 + ], + "lines": [ + { + "bbox": [ + 352, + 160, + 517, + 206 + ], + "spans": [ + { + "bbox": [ + 352, + 160, + 517, + 206 + ], + "type": "table", + "html": "
PSNR↑LPIPS↓gdef (ns)↓
w/o MACP (Nc=4)26.620.0655.34
w/o MACP (Nc=Nf)27.080.0546.11
Ours27.210.0535.63
", + "image_path": "ac47030047729578435ea17158f8379253577063c2c8778a803c66146486abf8.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 209, + 555, + 243 + ], + "lines": [ + { + "bbox": [ + 313, + 209, + 555, + 243 + ], + "spans": [ + { + "bbox": [ + 313, + 209, + 555, + 243 + ], + "type": "text", + "content": "Table 3. Ablation studies. We ablate our framework and report the average results on the NVIDIA dataset with the same setting as Novel View Synthesis experiment in Sec. 5.1." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 320, + 254, + 436, + 391 + ], + "blocks": [ + { + "bbox": [ + 320, + 254, + 436, + 391 + ], + "lines": [ + { + "bbox": [ + 320, + 254, + 436, + 391 + ], + "spans": [ + { + "bbox": [ + 320, + 254, + 436, + 391 + ], + "type": "image", + "image_path": "ac44e70d252a6783e3f0d3fcd541de3c038b9583bf1107457910d3964c2cfe6a.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 395, + 555, + 450 + ], + "lines": [ + { + "bbox": [ + 313, + 395, + 555, + 450 + ], + "spans": [ + { + "bbox": [ + 313, + 395, + 555, + 450 + ], + "type": "text", + "content": "Figure 8. Analysis of MACP's Efficacy. (a) " + }, + { + "bbox": [ + 313, + 395, + 555, + 450 + ], + "type": "inline_equation", + "content": "N_{c}" + }, + { + "bbox": [ + 313, + 395, + 555, + 450 + ], + "type": "text", + "content": " Heatmaps as the averaged " + }, + { + "bbox": [ + 313, + 395, + 555, + 450 + ], + "type": "inline_equation", + "content": "N_{c}" + }, + { + "bbox": [ + 313, + 395, + 555, + 450 + ], + "type": "text", + "content": " values of dynamic 3D Gaussians and their corresponding rendered frames " + }, + { + "bbox": [ + 313, + 395, + 555, + 450 + ], + "type": "inline_equation", + "content": "\\hat{I}_{t}" + }, + { + "bbox": [ + 313, + 395, + 555, + 450 + ], + "type": "text", + "content": " for 'Balloon2' and 'Skating' scenes. (b) Histograms of the number of control points " + }, + { + "bbox": [ + 313, + 395, + 555, + 450 + ], + "type": "inline_equation", + "content": "(N_{c})" + }, + { + "bbox": [ + 313, + 395, + 555, + 450 + ], + "type": "text", + "content": " in percentages (\\%) of dynamic 3D Gaussians in two scenes." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 438, + 254, + 552, + 391 + ], + "blocks": [ + { + "bbox": [ + 438, + 254, + 552, + 391 + ], + "lines": [ + { + "bbox": [ + 438, + 254, + 552, + 391 + ], + "spans": [ + { + "bbox": [ + 438, + 254, + 552, + 391 + ], + "type": "image", + "image_path": "90c6520ad34e6f0d3370f97583f2b095dffd15915aacf4e9f5e80889ae5f0261.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 460, + 554, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 460, + 554, + 495 + ], + "spans": [ + { + "bbox": [ + 313, + 460, + 554, + 495 + ], + "type": "text", + "content": "dynamic 3D Gaussians can be represented using a minimal " + }, + { + "bbox": [ + 313, + 460, + 554, + 495 + ], + "type": "inline_equation", + "content": "N_{c}" + }, + { + "bbox": [ + 313, + 460, + 554, + 495 + ], + "type": "text", + "content": ", thanks to our MACP. While 'Balloon2' has more evenly distributed " + }, + { + "bbox": [ + 313, + 460, + 554, + 495 + ], + "type": "inline_equation", + "content": "N_{c}" + }, + { + "bbox": [ + 313, + 460, + 554, + 495 + ], + "type": "text", + "content": " due to more complex and diverse motion." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 497, + 555, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 497, + 555, + 570 + ], + "spans": [ + { + "bbox": [ + 313, + 497, + 555, + 570 + ], + "type": "text", + "content": "Loss Functions. Table 3-(b) shows the effectiveness of each loss for our overall SplineGS architecture. As noted, no consistent camera parameters can be learned without " + }, + { + "bbox": [ + 313, + 497, + 555, + 570 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{pc}}" + }, + { + "bbox": [ + 313, + 497, + 555, + 570 + ], + "type": "text", + "content": ", drastically impacting the rendering quality of the dynamic 3D Gaussians. Moreover, our " + }, + { + "bbox": [ + 313, + 497, + 555, + 570 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{gc}}" + }, + { + "bbox": [ + 313, + 497, + 555, + 570 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 497, + 555, + 570 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{d - pc}}" + }, + { + "bbox": [ + 313, + 497, + 555, + 570 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 497, + 555, + 570 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{M}}" + }, + { + "bbox": [ + 313, + 497, + 555, + 570 + ], + "type": "text", + "content": " can considerably impact the overall rendering quality." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 584, + 388, + 596 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 584, + 388, + 596 + ], + "spans": [ + { + "bbox": [ + 313, + 584, + 388, + 596 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "content": "We present SplineGS, a COLMAP-free dynamic 3DGS framework designed for novel spatio-temporal view synthesis from monocular videos. Leveraging our innovative Motion-Adaptive Spline (MAS) for dynamic motion modeling, SplineGS efficiently renders high-quality novel views from complex in-the-wild videos. The effectiveness of our approach is validated through extensive quantitative and qualitative comparisons, significantly outperforming the existing SOTA methods with very fast rendering speed." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26873" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 90, + 296, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 90, + 296, + 168 + ], + "spans": [ + { + "bbox": [ + 55, + 90, + 296, + 168 + ], + "type": "text", + "content": "This work was supported by Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korean Government [Ministry of Science and ICT (Information and Communications Technology)] (Project Number: RS-2022-00144444, Project Title: Deep Learning Based Visual Representational Learning and Rendering of Static and Dynamic Scenes, 100%)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 189, + 115, + 202 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 189, + 115, + 202 + ], + "spans": [ + { + "bbox": [ + 56, + 189, + 115, + 202 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 209, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 61, + 209, + 295, + 231 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 209, + 295, + 231 + ], + "spans": [ + { + "bbox": [ + 61, + 209, + 295, + 231 + ], + "type": "text", + "content": "[1] Fast dynamic radiance fields with time-aware neural voxels. In SIGGRAPH Asia 2022 Conference Papers, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 232, + 296, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 232, + 296, + 285 + ], + "spans": [ + { + "bbox": [ + 61, + 232, + 296, + 285 + ], + "type": "text", + "content": "[2] J Harold Ahlberg, Edwin Norman Nilson, and Joseph Leonard Walsh. The Theory of Splines and Their Applications: Mathematics in Science and Engineering: A Series of Monographs and Textbooks, Vol. 38. Elsevier, 2016. 2, 3, 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 288, + 295, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 288, + 295, + 319 + ], + "spans": [ + { + "bbox": [ + 61, + 288, + 295, + 319 + ], + "type": "text", + "content": "[3] ShahRukh Athar, Zexiang Xu, Kalyan Sunkavalli, Eli Shechtman, and Zhixin Shu. Rignerf: Fully controllable neural 3d portraits. In CVPR, 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 61, + 322, + 295, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 322, + 295, + 365 + ], + "spans": [ + { + "bbox": [ + 61, + 322, + 295, + 365 + ], + "type": "text", + "content": "[4] Benjamin Attal, Jia-Bin Huang, Christian Richardt, Michael Zollhoefer, Johannes Kopf, Matthew O'Toole, and Changil Kim. Hyperreel: High-fidelity 6-dof video with ray-conditioned sampling. In CVPR, 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 61, + 366, + 296, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 366, + 296, + 421 + ], + "spans": [ + { + "bbox": [ + 61, + 366, + 296, + 421 + ], + "type": "text", + "content": "[5] Jeongmin Bae, Seoha Kim, Youngsik Yun, Hahyun Lee, Gun Bang, and Youngjung Uh. Per-gaussian embedding-based deformation for deformable 3d gaussian splatting. In European Conference on Computer Vision, pages 321-335. Springer, 2024. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 422, + 295, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 422, + 295, + 443 + ], + "spans": [ + { + "bbox": [ + 62, + 422, + 295, + 443 + ], + "type": "text", + "content": "[6] Ang Cao and Justin Johnson. Hexplane: A fast representation for dynamic scenes. In CVPR, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 445, + 296, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 445, + 296, + 487 + ], + "spans": [ + { + "bbox": [ + 62, + 445, + 296, + 487 + ], + "type": "text", + "content": "[7] Mengyu Chu, You Xie, Jonas Mayer, Laura Leal-Taixe, and Nils Thuerey. Learning temporal coherence via self-supervision for gan-based video generation. ACM Transactions on Graphics (TOG), 2020. 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 489, + 295, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 489, + 295, + 510 + ], + "spans": [ + { + "bbox": [ + 62, + 489, + 295, + 510 + ], + "type": "text", + "content": "[8] C De Boor. A practical guide to splines. Springer-Verlag google schola, 1978. 2, 3, 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 512, + 295, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 512, + 295, + 533 + ], + "spans": [ + { + "bbox": [ + 62, + 512, + 295, + 533 + ], + "type": "text", + "content": "[9] Gerald Farin. Curves and surfaces for CAGD: a practical guide. Elsevier, 2001. 2, 7, 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 535, + 295, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 535, + 295, + 578 + ], + "spans": [ + { + "bbox": [ + 57, + 535, + 295, + 578 + ], + "type": "text", + "content": "[10] Sara Fridovich-Keil, Giacomo Meanti, Frederik Rahbæk Warburg, Benjamin Recht, and Angjoo Kanazawa. K-planes: Explicit radiance fields in space, time, and appearance. In CVPR, 2023. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 580, + 295, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 580, + 295, + 612 + ], + "spans": [ + { + "bbox": [ + 57, + 580, + 295, + 612 + ], + "type": "text", + "content": "[11] Yang Fu, Sifei Liu, Amey Kulkarni, Jan Kautz, Alexei A. Efros, and Xiaolong Wang. Colmap-free 3d gaussian splatting. In CVPR, 2024. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 614, + 295, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 614, + 295, + 645 + ], + "spans": [ + { + "bbox": [ + 57, + 614, + 295, + 645 + ], + "type": "text", + "content": "[12] Chen Gao, Ayush Saraf, Johannes Kopf, and Jia-Bin Huang. Dynamic view synthesis from dynamic monocular video. In ICCV, 2021. 2, 6, 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 647, + 295, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 647, + 295, + 669 + ], + "spans": [ + { + "bbox": [ + 57, + 647, + 295, + 669 + ], + "type": "text", + "content": "[13] Roger A Horn and Charles R Johnson. Matrix analysis. Cambridge university press, 2012. 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 57, + 670, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 295, + 712 + ], + "type": "text", + "content": "[14] Yi-Hua Huang, Yang-Tian Sun, Ziyi Yang, Xiaoyang Lyu, Yan-Pei Cao, and Xiaojuan Qi. Sc-gs: Sparse-controlled gaussian splatting for editable dynamic scenes. CVPR, 2024. 2, 6" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 554, + 712 + ], + "type": "list", + "angle": 0, + "index": 34, + "blocks": [ + { + "bbox": [ + 316, + 73, + 554, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 127 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 127 + ], + "type": "text", + "content": "[15] Woobin Im, Geonho Cha, Sebin Lee, Jumin Lee, Juhyeong Seon, Dongyoon Wee, and Sung-Eui Yoon. Regularizing dynamic radiance fields with kinematic fields. In European Conference on Computer Vision, pages 312-328. Springer, 2024. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 129, + 553, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 553, + 162 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 553, + 162 + ], + "type": "text", + "content": "[16] Wei Jiang, Kwang Moo Yi, Golnoosh Samei, Oncel Tuzel, and Anurag Ranjan. Neuman: Neural human radiance field from a single video. In ECCV, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 163, + 553, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 163, + 553, + 206 + ], + "spans": [ + { + "bbox": [ + 317, + 163, + 553, + 206 + ], + "type": "text", + "content": "[17] Yifan Jiang, Peter Hedman, Ben Mildenhall, Dejia Xu, Jonathan T Barron, Zhangyang Wang, and Tianfan Xue. Alignerf: High-fidelity neural radiance fields via alignment-aware training. In CVPR, 2023. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 208, + 553, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 208, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 316, + 208, + 553, + 262 + ], + "type": "text", + "content": "[18] Moritz Kappel, Florian Hahlbohm, Timon Scholz, Susana Castillo, Christian Theobalt, Martin Eisemann, Vladislav Golyanik, and Marcus Magnor. D-npc: Dynamic neural point clouds for non-rigid view synthesis from monocular video. arXiv preprint arXiv:2406.10078, 2024. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 264, + 553, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 264, + 553, + 296 + ], + "spans": [ + { + "bbox": [ + 316, + 264, + 553, + 296 + ], + "type": "text", + "content": "[19] Nikita Karaev, Ignacio Rocco, Benjamin Graham, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. Co-tracker: It is better to track together. arXiv, 2023. 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 298, + 553, + 340 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 298, + 553, + 340 + ], + "spans": [ + { + "bbox": [ + 316, + 298, + 553, + 340 + ], + "type": "text", + "content": "[20] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 2023. 2, 3, 4, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 342, + 553, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 342, + 553, + 374 + ], + "spans": [ + { + "bbox": [ + 316, + 342, + 553, + 374 + ], + "type": "text", + "content": "[21] Junoh Lee, Chang-Yeon Won, Hyunjun Jung, Inhwan Bae, and Hae-Gon Jeon. Fully explicit dynamic gaussian splatting. In NeurIPS, 2024. 3, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 376, + 553, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 376, + 553, + 420 + ], + "spans": [ + { + "bbox": [ + 316, + 376, + 553, + 420 + ], + "type": "text", + "content": "[22] Yao-Chih Lee, Zhoutong Zhang, Kevin Blackburn-Matzen, Simon Niklaus, Jianming Zhang, Jia-Bin Huang, and Feng Liu. Fast view synthesis of casual videos with soup-of-planes. In ECCV, 2024. 3, 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 422, + 553, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 422, + 553, + 453 + ], + "spans": [ + { + "bbox": [ + 316, + 422, + 553, + 453 + ], + "type": "text", + "content": "[23] Jiahui Lei, Yijia Weng, Adam Harley, Leonidas Guibas, and Kostas Daniilidis. Mosca: Dynamic gaussian fusion from casual videos via 4d motion scaffolds. arXiv, 2024. 3, 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 456, + 553, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 456, + 553, + 487 + ], + "spans": [ + { + "bbox": [ + 316, + 456, + 553, + 487 + ], + "type": "text", + "content": "[24] Zhan Li, Zhang Chen, Zhong Li, and Yi Xu. Spacetime gaussian feature splatting for real-time dynamic view synthesis. In CVPR. 1, 2, 3, 6, 7, 8" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 489, + 553, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 489, + 553, + 521 + ], + "spans": [ + { + "bbox": [ + 316, + 489, + 553, + 521 + ], + "type": "text", + "content": "[25] Zhengqi Li, Simon Niklaus, Noah Snavely, and Oliver Wang. Neural scene flow fields for space-time view synthesis of dynamic scenes. In CVPR, 2021. 2, 7" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 523, + 553, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 523, + 553, + 555 + ], + "spans": [ + { + "bbox": [ + 316, + 523, + 553, + 555 + ], + "type": "text", + "content": "[26] Zhengqi Li, Qianqian Wang, Forrester Cole, Richard Tucker, and Noah Snavely. Dynibar: Neural dynamic image-based rendering. In CVPR, 2023. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 557, + 553, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 557, + 553, + 600 + ], + "spans": [ + { + "bbox": [ + 316, + 557, + 553, + 600 + ], + "type": "text", + "content": "[27] Yiqing Liang, Numair Khan, Zhengqin Li, Thu Nguyen-Phuoc, Douglas Lanman, James Tompkin, and Lei Xiao. Gaufre: Gaussian deformation fields for real-time dynamic novel view synthesis. arXiv, 2023. 2, 3" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 602, + 553, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 602, + 553, + 633 + ], + "spans": [ + { + "bbox": [ + 316, + 602, + 553, + 633 + ], + "type": "text", + "content": "[28] Chen-Hsuan Lin, Wei-Chiu Ma, Antonio Torralba, and Simon Lucey. Barf: Bundle-adjusting neural radiance fields. In ICCV, 2021. 3" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 316, + 635, + 553, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 553, + 667 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 553, + 667 + ], + "type": "text", + "content": "[29] Philipp Lindenberger, Paul-Edouard Sarlin, Viktor Larsson, and Marc Pollefeys. Pixel-perfect structure-from-motion with featuremetric refinement. In ICCV, 2021. 3" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "text", + "content": "[30] Yu-Lun Liu, Chen Gao, Andreas Meuleman, Hung-Yu Tseng, Ayush Saraf, Changil Kim, Yung-Yu Chuang, Johannes Kopf, and Jia-Bin Huang. Robust dynamic radiance fields. In CVPR, 2023. 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 33 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "26874" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "type": "text", + "content": "[31] Jonathon Luiten, Georgios Kopanas, Bastian Leibe, and Deva Ramanan. Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis. In 3DV, 2024. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 106, + 294, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 106, + 294, + 139 + ], + "spans": [ + { + "bbox": [ + 56, + 106, + 294, + 139 + ], + "type": "text", + "content": "[32] Quan Meng, Anpei Chen, Haimin Luo, Minye Wu, Hao Su, Lan Xu, Xuming He, and Jingyi Yu. Gnerf: Gan-based neural radiance field without posed camera. In ICCV, 2021. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 140, + 294, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 140, + 294, + 183 + ], + "spans": [ + { + "bbox": [ + 56, + 140, + 294, + 183 + ], + "type": "text", + "content": "[33] Xingyu Miao, Yang Bai, Haoran Duan, Fan Wan, Yawen Huang, Yang Long, and Yefeng Zheng. Ctnerf: Cross-time transformer for dynamic neural radiance field from monocular video. Pattern Recognition, 156:110729, 2024. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 184, + 294, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 184, + 294, + 227 + ], + "spans": [ + { + "bbox": [ + 56, + 184, + 294, + 227 + ], + "type": "text", + "content": "[34] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 2, 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 228, + 294, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 228, + 294, + 270 + ], + "spans": [ + { + "bbox": [ + 56, + 228, + 294, + 270 + ], + "type": "text", + "content": "[35] Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In ICCV, 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 271, + 294, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 271, + 294, + 327 + ], + "spans": [ + { + "bbox": [ + 56, + 271, + 294, + 327 + ], + "type": "text", + "content": "[36] Keunhong Park, Utkarsh Sinha, Peter Hedman, Jonathan T. Barron, Sofien Bouaziz, Dan B Goldman, Ricardo Martin-Brualla, and Steven M. Seitz. Hypernerf: A higher-dimensional representation for topologically varying neural radiance fields. ACM Trans. Graph., 2021. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 327, + 294, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 327, + 294, + 370 + ], + "spans": [ + { + "bbox": [ + 56, + 327, + 294, + 370 + ], + "type": "text", + "content": "[37] Keunhong Park, Philipp Henzler, Ben Mildenhall, Jonathan T. Barron, and Ricardo Martin-Brualla. Camp: Camera preconditioning for neural radiance fields. ACM Trans. Graph., 2023. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 371, + 294, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 371, + 294, + 414 + ], + "spans": [ + { + "bbox": [ + 56, + 371, + 294, + 414 + ], + "type": "text", + "content": "[38] Luigi Piccinelli, Yung-Hsu Yang, Christos Sakaridis, Mattia Segu, Siyuan Li, Luc Van Gool, and Fisher Yu. Unidepth: Universal monocular metric depth estimation. In CVPR, 2024. 4, 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 415, + 294, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 415, + 294, + 458 + ], + "spans": [ + { + "bbox": [ + 56, + 415, + 294, + 458 + ], + "type": "text", + "content": "[39] Jordi Pont-Tuset, Federico Peruzzi, Sergi Caelles, Pablo Arbeláez, Alex Sorkine-Hornung, and Luc Van Gool. The 2017 davis challenge on video object segmentation. arXiv, 2018. 1, 6, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 460, + 294, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 460, + 294, + 492 + ], + "spans": [ + { + "bbox": [ + 56, + 460, + 294, + 492 + ], + "type": "text", + "content": "[40] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-nerf: Neural radiance fields for dynamic scenes. In CVPR, 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 493, + 294, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 493, + 294, + 536 + ], + "spans": [ + { + "bbox": [ + 56, + 493, + 294, + 536 + ], + "type": "text", + "content": "[41] Vincent Raoult, Sarah Reid-Anderson, Andreas Ferri, and Jane E Williamson. How reliable is structure from motion (sfm) over time and between observers? a case study using coral reef bommie. Remote Sensing, 2017. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 537, + 294, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 537, + 294, + 558 + ], + "spans": [ + { + "bbox": [ + 56, + 537, + 294, + 558 + ], + "type": "text", + "content": "[42] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. In CVPR, 2016. 1, 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 559, + 294, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 559, + 294, + 602 + ], + "spans": [ + { + "bbox": [ + 56, + 559, + 294, + 602 + ], + "type": "text", + "content": "[43] Ruizhi Shao, Zerong Zheng, Hanzhang Tu, Boning Liu, Hongwen Zhang, and Yebin Liu. Tensor4d: Efficient neural 4d decomposition for high-fidelity dynamic reconstruction and rendering. In CVPR, 2023. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 603, + 294, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 603, + 294, + 658 + ], + "spans": [ + { + "bbox": [ + 56, + 603, + 294, + 658 + ], + "type": "text", + "content": "[44] Liangchen Song, Anpei Chen, Zhong Li, Zhang Chen, Lele Chen, Junsong Yuan, Yi Xu, and Andreas Geiger. Nerfplayer: A streamable dynamic scene representation with decomposed neural radiance fields. IEEE Transactions on Visualization and Computer Graphics, 2023. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 658, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 294, + 713 + ], + "type": "text", + "content": "[45] Carole H. Sudre, Wenqi Li, Tom Vercauteren, Sebastien Ourselin, and M. Jorge Cardoso. Generalised dice overlap as a deep learning loss function for highly unbalanced segmentations. In Deep Learning in Medical Image Analysis and Multimodal Learning for Clinical Decision Support, 2017. 5" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 621 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "text", + "content": "[46] Fengrui Tian, Shaoyi Du, and Yueqi Duan. MonoN-eRF: Learning a generalizable dynamic radiance field from monocular videos. In ICCV, 2023. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 107, + 553, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 161 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 161 + ], + "type": "text", + "content": "[47] Edgar Tretschk, Ayush Tewari, Vladislav Golyanik, Michael Zollhöfer, Christoph Lassner, and Christian Theobalt. Non-rigid neural radiance fields: Reconstruction and novel view synthesis of a dynamic scene from monocular video. In ICCV, 2021. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 163, + 553, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 553, + 196 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 553, + 196 + ], + "type": "text", + "content": "[48] Diwen Wan, Ruijie Lu, and Gang Zeng. Superpoint gaussian splatting for real-time high-fidelity dynamic scene reconstruction. arXiv preprint arXiv:2406.03697, 2024. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 197, + 553, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 197, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 316, + 197, + 553, + 239 + ], + "type": "text", + "content": "[49] Qianqian Wang, Vickie Ye, Hang Gao, Jake Austin, Zhengqi Li, and Angjoo Kanazawa. Shape of motion: 4d reconstruction from a single video. arXiv preprint arXiv:2407.13764, 2024. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 241, + 553, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 241, + 553, + 274 + ], + "spans": [ + { + "bbox": [ + 316, + 241, + 553, + 274 + ], + "type": "text", + "content": "[50] Zirui Wang, Shangzhe Wu, Weidi Xie, Min Chen, and Victor Adrian Prisacariu. Nerf: Neural radiance fields without known camera parameters. CoRR, 2021. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 275, + 553, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 275, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 316, + 275, + 553, + 319 + ], + "type": "text", + "content": "[51] Chung-Yi Weng, Brian Curless, Pratul P. Srinivasan, Jonathan T. Barron, and Ira Kemelmacher-Shlizerman. HumanNeRF: Free-viewpoint rendering of moving people from monocular video. In CVPR, 2022. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 320, + 553, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 320, + 553, + 363 + ], + "spans": [ + { + "bbox": [ + 316, + 320, + 553, + 363 + ], + "type": "text", + "content": "[52] Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Xinggang Wang. 4d gaussian splatting for real-time dynamic scene rendering. In CVPR. 2, 6, 7, 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 365, + 553, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 365, + 553, + 408 + ], + "spans": [ + { + "bbox": [ + 316, + 365, + 553, + 408 + ], + "type": "text", + "content": "[53] Gengshan Yang, Minh Vo, Neverova Natalia, Deva Ramanan, Vedaldi Andrea, and Joo Hanbyul. Banmo: Building animatable 3d neural models from many casual videos. In CVPR, 2022. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 410, + 553, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 410, + 553, + 442 + ], + "spans": [ + { + "bbox": [ + 316, + 410, + 553, + 442 + ], + "type": "text", + "content": "[54] Jinyu Yang, Mingqi Gao, Zhe Li, Shang Gao, Fangjing Wang, and Feng Zheng. Track anything: Segment anything meets videos. arXiv, 2023. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 444, + 553, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 444, + 553, + 487 + ], + "spans": [ + { + "bbox": [ + 316, + 444, + 553, + 487 + ], + "type": "text", + "content": "[55] Ziyi Yang, Xinyu Gao, Wen Zhou, Shaohui Jiao, Yuqing Zhang, and Xiaogang Jin. Deformable 3d gaussians for high-fidelity monocular dynamic scene reconstruction. In CVPR, 2024. 1, 2, 6, 7, 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 488, + 553, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 488, + 553, + 532 + ], + "spans": [ + { + "bbox": [ + 316, + 488, + 553, + 532 + ], + "type": "text", + "content": "[56] Jae Shin Yoon, Kihwan Kim, Orazio Gallo, Hyun Soo Park, and Jan Kautz. Novel view synthesis of dynamic scenes with globally coherent depths from a monocular camera. In CVPR, 2020. 1, 6, 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 533, + 553, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 533, + 553, + 566 + ], + "spans": [ + { + "bbox": [ + 316, + 533, + 553, + 566 + ], + "type": "text", + "content": "[57] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 567, + 553, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 567, + 553, + 621 + ], + "spans": [ + { + "bbox": [ + 316, + 567, + 553, + 621 + ], + "type": "text", + "content": "[58] Kaichen Zhou, Jia-Xing Zhong, Sangyun Shin, Kai Lu, Yiyuan Yang, Andrew Markham, and Niki Trigoni. Dynpoint: Dynamic neural point for view synthesis. Advances in Neural Information Processing Systems, 36:69532-69545, 2023. 6" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26875" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Split Adaptation for Pre-trained Vision Transformers/b5d6478d-8866-4dba-8ddb-9d84bd75792a_content_list.json b/2025/Split Adaptation for Pre-trained Vision Transformers/b5d6478d-8866-4dba-8ddb-9d84bd75792a_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..1faa115b1d54488000605f69427282cc69638e1b --- /dev/null +++ b/2025/Split Adaptation for Pre-trained Vision Transformers/b5d6478d-8866-4dba-8ddb-9d84bd75792a_content_list.json @@ -0,0 +1,1610 @@ +[ + { + "type": "text", + "text": "Split Adaptation for Pre-trained Vision Transformers", + "text_level": 1, + "bbox": [ + 228, + 130, + 769, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Lixu Wang $^{1,2*}$ , Bingqi Shang $^{1*}$ , Yi Li $^{1}$ , Payal Mohapatra $^{1}$ , Wei Dong $^{2}$ , Xiao Wang $^{1}$ , Qi Zhu $^{1}$ , Northwestern University. $^{2}$ Nanyang Technological University.", + "bbox": [ + 133, + 179, + 861, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "lixu.wang@ntu.edu.sg, {bingqishang2025, yili2023.1, payalmohapatra2026}@u.northwestern.edu, wei_dong@ntu.edu.sg, {wangxiao, qzhu}@northwestern.edu", + "bbox": [ + 104, + 219, + 888, + 251 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 286, + 326, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Vision Transformers (ViTs), extensively pre-trained on large-scale datasets, have become fundamental to foundation models, enabling adaptation to diverse downstream tasks. Existing adaptation methods typically require direct data access, rendering them infeasible in privacy-sensitive domains where clients are often reluctant to share their data. A straightforward solution may be sending the pretrained ViT to clients for local adaptation, which poses issues of model intellectual property and incurs heavy client computation overhead. To address these issues, we propose a novel split adaptation (SA) method that enables effective downstream adaptation while protecting data and models. SA, inspired by split learning (SL), segments the pre-trained ViT into a frontend and a backend, with only the frontend shared with the client for data representation extraction. But unlike regular SL, SA replaces frontend parameters with low-bit quantized values, preventing direct exposure of the model. SA allows the client to add bi-level noise to the frontend and the extracted data representations, ensuring data protection. Accordingly, SA incorporates data-level and model-level out-of-distribution enhancements to mitigate noise injection's impact. Our SA focuses on the challenging few-shot adaptation and adopts patch retrieval augmentation for overfitting alleviation. Extensive experiments on multiple datasets validate SA's superiority over state-of-the-art methods and demonstrate its defense against advanced data reconstruction attacks while preventing model leakage with minimal computation cost on the client side. The source codes can be found at https://github.com/conditionWang/Split_Adaptation.", + "bbox": [ + 91, + 318, + 483, + 772 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 803, + 220, + 819 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In recent years, pre-trained vision transformers (ViTs) [13] have become mainstream foundational models in computer vision, delivering remarkable performance across a range of", + "bbox": [ + 89, + 829, + 483, + 875 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "visual tasks [13, 28, 50, 52]. ViTs have been equipped with general semantic knowledge during large-scale pre-training, which can be readily transferred to other domains including sensitive fields such as healthcare [41], finance [10], and manufacturing [8]. However, adapting pre-trained ViTs in these sensitive domains is extremely challenging.", + "bbox": [ + 511, + 287, + 905, + 378 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In real-world scenarios, model adaptation often follows a two-party framework, where the pre-trained model is hosted on a server, and the training data resides on a client. In non-sensitive domains, training data can be directly shared with the server. Real-world examples include users uploading data to access OpenAI's CLIP model API [40] or tune GPT-3.5 [3]. However, in sensitive domains, data is often tightly linked to personal privacy [12, 14], intellectual property (IP) [19, 47], and ethical concerns [15], making any form of public exposure, sharing, or transfer strictly prohibited. In this context, an alternative solution might involve sending the model to the client for local adaptation, as in federated learning [11, 27, 46]. However, large pre-trained models are expensive to obtain, making their owners unlikely to share or transfer them freely [47]. Moreover, given that these pre-trained models consist of billions of parameters, adaptation on the client side would incur substantial computational and storage costs.", + "bbox": [ + 511, + 382, + 905, + 654 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To our knowledge, there have been a few preliminary attempts to achieve downstream adaptation of pre-trained models while safeguarding data and models. Split learning (SL) [1, 44] divides the model into segments managed by different parties, which can partially protect the model from full exposure. Additionally, SL only communicates intermediate representations, avoiding the leakage of raw data. However, SL has critical drawbacks. On one hand, emerging malicious attacks, such as data reconstruction attacks (DRA) [16, 23, 38, 53, 54], can infer sensitive information from the intermediate representations. On the other hand, standard SL places the model training on the client side, imposing high computational costs. Beyond SL, a recent method called offsite tuning (OT) [49] has been proposed, which involves the server sending an emulator—a simplified version of the pre-trained model—to the client. The", + "bbox": [ + 511, + 657, + 905, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contributions.", + "bbox": [ + 107, + 886, + 223, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "20092", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3cdfb9127d5056f0a9058f7ff1ffba42ffa2f05ed3cb132d6bbc848741182b17.jpg", + "image_caption": [ + "Figure 1. The comparison between different downstream task adaptation approaches. Our SA can achieve effective few-shot task adaptation with minimal computation cost on the client side while protecting both the model and data." + ], + "image_footnote": [], + "bbox": [ + 94, + 88, + 478, + 250 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "client can then fine-tune this emulator using its data, typically employing parameter-efficient techniques like adapter tuning [5]. After training, the client uploads these adapters back to the server, where they are integrated into the pretrained model. However, while OT protects data, the emulator is fully exposed, allowing adversaries to create high-performance models easily. Moreover, OT's adaptation depends on the substantial computational and storage demands for adapter tuning on the client side.", + "bbox": [ + 88, + 321, + 482, + 455 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address the challenge of adapting pre-trained ViTs in sensitive domains, we propose a split adaptation (SA) method, which is shown in Figure 1. Like SL, SA divides the pre-trained ViT into two parts: a frontend and a backend. The server sends the frontend to the client, enabling data representation extraction. However, unlike SL, SA innovatively uses quantization to replace the frontend's parameters with low-bit values, preventing direct exposure of the original parameters. For data protection, the client injects random noise into the received frontend and its extracted data representations. Such injected noise unavoidably hurts the following adaptation, hence SA includes data-level and model-level out-of-distribution enhancements to improve the model generalization. SA is particularly designed for the challenging few-shot adaptation scenario, where it uses a patch retrieval augmentation technique to alleviate overfitting. Extensive experiments across four datasets demonstrate that SA significantly outperforms other state-of-the-art approaches in adaptation performance. Moreover, SA effectively defends against advanced DRAs and prevents model leakage. In particular, SA's effectiveness only relies on small computation cost on the client side. In summary:", + "bbox": [ + 91, + 459, + 482, + 792 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We focus on solving a practical and important problem—adapting pre-trained ViTs for downstream tasks while protecting data and models. Solving this problem enables wider application of pre-trained models in sensitive fields.", + "- We propose Split Adaptation (SA) that innovatively leverages model splitting and quantization to protect data and models during downstream task adaptation. SA focuses" + ], + "bbox": [ + 89, + 795, + 483, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "on challenging few-shot scenarios, in which SA uses a patch retrieval augmentation to combat overfitting.", + "bbox": [ + 526, + 90, + 903, + 121 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- We conduct extensive experiments across multiple datasets and state-of-the-art baseline approaches, with results validating SA's effectiveness in achieving superior adaptation while ensuring data and model protection.", + "bbox": [ + 513, + 121, + 903, + 181 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 513, + 196, + 663, + 213 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Split Learning", + "text_level": 1, + "bbox": [ + 513, + 223, + 661, + 239 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Split learning (SL) [44] is a collaborative learning approach widely used in the multi-party computation framework. SL operates by dividing the machine learning (ML) model into segments, with each managed by a different party. This approach enhances privacy by transmitting only intermediate data representations rather than raw data. Many practical examples of SL in training and adapting ML models across various domains emerge, including healthcare [1, 44], financial [1, 42], and edge computing [17, 18]. Despite these successes, more adversarial attacks illustrate the vulnerability of privacy leakage in SL, with the severest one – data reconstruction attack (DRA) [23]. DRA aims to recover the raw data from its corresponding model representations. In the context of SL, a variety of DRAs have been witnessed, among them, some are based on optimization to imitate the ground-truth representations (e.g., FSHA [38] and PCAT [16]), while others leverage generative models to recover the raw data by exploiting the intrinsic semantics (e.g., Ginver [54] and FORA [53]). Therefore, new SL approaches are needed to help protect data privacy better.", + "bbox": [ + 511, + 246, + 906, + 549 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Vision Transformer Adaptation", + "text_level": 1, + "bbox": [ + 511, + 560, + 792, + 575 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The vision transformers (ViT) [13] are usually pre-trained with large-scale data to learn general semantics and knowledge. Representative pre-training approaches include supervised [28] and self-supervised learning like contrastive learning [4, 6, 21, 40] and auxiliary task training [22, 37, 57]. Owing to such extensive pre-training, adapting pretrained ViT for distinct downstream tasks has become a mainstream paradigm. The most regular method is to tune the entire model, which consumes substantial computational and memory resources. Another widely used approach is linear probing [29], which only trains a task module while keeping the ViT frozen, but the prominent shortcoming is the inferior performance. Recently, more parameter-efficient approaches have been proposed, like adapter tuning [5], prompt tuning [26], and prefix tuning [51]. However, these approaches still require direct access to the task data. In contrast to the above approaches, a new method called offsite tuning [49] can work effectively without direct data access. However, extensive training happens on the data owner side, and there is a high risk of model disclosure, making it impractical in real-world scenarios.", + "bbox": [ + 511, + 583, + 906, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "20093", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 89, + 89, + 225, + 107 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminaries", + "text_level": 1, + "bbox": [ + 89, + 114, + 230, + 128 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1.1. Vision Transformer", + "text_level": 1, + "bbox": [ + 89, + 137, + 272, + 151 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The vision transformer (ViT) [13] is built upon sequential attention encoder layers. Specifically, each encoder layer consists of a multi-head self-attention (MSA) block and a Multi-Layer Perceptron (MLP) block. Layernorm (LN) is applied before every block and residual connections are placed after every block. The MLP contains two linear layers with an intermediate GELU activation. To handle 2D images, ViT reshapes the image $\\mathbf{I} \\in \\mathbb{R}^{H \\times W \\times C}$ into $N$ flattened patches $I^p \\in \\mathbb{R}^{P^2 \\cdot C}$ , where $H$ , $W$ , and $C$ are the image height, width, and channels, respectively, and $P$ is the pixel resolution of each patch. For the first layer of ViT, to unify the vector dimensions, a trainable linear projection layer with neuron weights $[\\mathbf{W}_1^E, \\dots, \\mathbf{W}_N^E]$ is used to map each vectorized patch into dimension $d$ . In this case, the input to the first transformer layer is:", + "bbox": [ + 89, + 156, + 483, + 387 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {X} _ {1} = \\left[ \\mathbf {x} ^ {\\mathrm {C L S}}; I _ {1} ^ {p} \\mathbf {W} _ {1} ^ {E}; \\dots ; I _ {N} ^ {p} \\mathbf {W} _ {N} ^ {E} \\right] + \\mathbf {E} ^ {\\mathrm {P O S}}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 132, + 396, + 483, + 425 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nw h e r e \\mathbf {W} ^ {E} \\in \\mathbb {R} ^ {(P ^ {2} \\cdot C) \\times d}, \\mathbf {E} ^ {\\mathrm {P O S}} \\in \\mathbb {R} ^ {(N + 1) \\times d},\n$$\n", + "text_format": "latex", + "bbox": [ + 133, + 420, + 439, + 436 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbf{x}^{\\mathrm{CLS}}\\in \\mathbb{R}^{d}$ is a classification token, and $\\mathbf{E}^{\\mathrm{POS}}$ is the position encoding. Then for each of the following layers, the input sequence $\\mathbf{X}_l$ is fed into each head of the MSA module to obtain three linearly projected matrices: Query $\\mathbf{Q}_l = \\mathbf{X}_l\\mathbf{W}_l^Q$ , Key $\\mathbf{K}_l = \\mathbf{X}_l\\mathbf{W}_l^K$ , and Value $\\mathbf{V}_l = \\mathbf{X}_l\\mathbf{W}_l^V$ . Then the Softmax function is applied to normalize the attention score $\\mathbf{A}_l = \\mathbf{Q}_l\\mathbf{K}_l^\\top$ and the output $\\mathbf{Z}_l$ is:", + "bbox": [ + 89, + 450, + 483, + 559 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {Z} _ {l} = \\operatorname {M S A} \\left(\\mathbf {X} _ {l}\\right) = \\operatorname {S o f t m a x} \\left(\\frac {\\mathbf {A} _ {l}}{\\sqrt {d}}\\right) \\mathbf {V} _ {l}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 571, + 483, + 606 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Suppose the two linear layers in MLP are parameterized by $\\mathbf{W}_{l,1},\\mathbf{b}_{l,1}$ and $\\mathbf{W}_{l,2},\\mathbf{b}_{l,2}$ , the output is then computed as:", + "bbox": [ + 89, + 614, + 482, + 647 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {M L P} \\left(\\mathbf {Z} _ {l}\\right) = \\operatorname {G E L U} \\left(\\mathbf {Z} _ {l} \\mathbf {W} _ {l, 1} + \\boldsymbol {b} _ {l, 1}\\right) \\mathbf {W} _ {l, 2} + \\boldsymbol {b} _ {l, 2}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 112, + 657, + 482, + 676 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Finally, both the MSA and MLP modules are linked via residual connections, thus we have the final output of each transformer layer: $\\mathbf{X}_{l + 1} = \\mathrm{LN}(\\mathbf{Z}_l + \\mathrm{MLP}(\\mathbf{Z}_l))\\in$ $\\mathbb{R}^{(N + 1)\\times d}$ , where $\\mathbf{Z}_l = \\mathrm{LN}(\\mathbf{X}_l + \\mathrm{MSA}(\\mathbf{X}_l))\\in \\mathbb{R}^{(N + 1)\\times d}$", + "bbox": [ + 89, + 688, + 483, + 750 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1.2. Problem Formulation", + "text_level": 1, + "bbox": [ + 89, + 758, + 287, + 772 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "When adapting pre-trained ViT for certain downstream tasks, we consider a representative client-server framework in which the client queries the server for the adaptation request. The client owns the task adaptation dataset $\\mathcal{D}^{\\mathrm{C}} = \\{(x_i,y_i)\\}_{i = 1}^{N^{\\mathrm{C}}} (N^{\\mathrm{C}}$ is the sample number, in practice, we suppose $N^{\\mathrm{C}}$ is small even few-shot) with each sample drawn from a marginal input distribution $\\pmb{x}_i\\sim \\mathcal{P}_X^{\\mathrm{C}}$ and labeled with a marginal task distribution $\\mathcal{P}_{\\mathcal{Y}}^{\\mathrm{C}}$ .", + "bbox": [ + 89, + 777, + 483, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "On the server side, the pre-trained ViT $F_{\\Theta}$ parameterized on $\\Theta$ is assumed to consist of $L^F$ attention layers, and each layer is parameterized on $\\theta$ , i.e., $F := [f_{\\theta_1}, \\dots, f_{\\theta_{L^F}}]$ . Besides, we assume the server owns another dataset $\\mathcal{D}^{\\mathrm{S}} = \\{(x_i, y_i)\\}_{i=1}^{N^{\\mathrm{S}}}$ collected from public domains with the marginal input and task distributions, $\\mathcal{P}_{\\mathcal{X}}^{\\mathrm{S}}$ and $\\mathcal{P}_{\\mathcal{Y}}^{\\mathrm{S}}$ , respectively. In real-world scenarios, the datasets owned by the client and the server are distinct but relevant, i.e., $\\mathcal{P}_{\\mathcal{X}}^{\\mathrm{C}} \\neq \\mathcal{P}_{\\mathcal{X}}^{\\mathrm{S}}, \\mathcal{P}_{\\mathcal{Y}}^{\\mathrm{C}} \\neq \\mathcal{P}_{\\mathcal{Y}}^{\\mathrm{S}}$ but $I(\\mathcal{X}^{\\mathrm{C}}, \\mathcal{X}^{\\mathrm{S}}) > 0$ , $I(\\mathcal{Y}^{\\mathrm{C}}, \\mathcal{Y}^{\\mathrm{S}}) > 0$ where $I(\\cdot; \\cdot)$ calculates the mutual information.", + "bbox": [ + 511, + 90, + 906, + 247 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Objective of Task Performance. Pre-trained ViT adaptation aims to obtain a downstream task module $g$ parameterized on $\\gamma$ that can achieve as good performance as possible on the client data distribution when combined with $F_{\\Theta}$ . Certainly, such adaptation also allows $F_{\\Theta}$ tuning if it can enable performance improvement on the downstream task. In this case, the objective is formulated as,", + "bbox": [ + 511, + 253, + 906, + 361 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\Theta , \\gamma = \\arg \\min _ {\\Theta , \\gamma} \\mathbb {E} _ {\\boldsymbol {x} \\sim \\mathcal {P} _ {\\chi} ^ {\\mathrm {C}}, y \\sim \\mathcal {P} _ {\\mathcal {Y}} ^ {\\mathrm {C}}} \\mathcal {L} (\\boldsymbol {x}, y; \\Theta \\circ \\gamma), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 557, + 376, + 906, + 398 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathcal{L}(\\cdot ;\\cdot)$ is a loss function that measures the model performance regarding certain data distribution.", + "bbox": [ + 511, + 414, + 905, + 446 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Objective of Privacy Protection. In addition to the above standard objective, the intrinsic privacy concerns of the server-client framework should be addressed, i.e., the pretrained model and client data should be protected during adaptation when supposing the server and the client may be semi-honest. On the one hand, the adaptation should prohibit the client from stealing surrogate models with good performance or obtaining them at a small cost, on the other hand, any client data reconstruction on the server side should be avoided.", + "bbox": [ + 511, + 452, + 906, + 604 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Overview of The Proposed Split Adaptation", + "text_level": 1, + "bbox": [ + 511, + 619, + 883, + 636 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To achieve effective adaptation while protecting the pre-trained ViT and client data, we propose a method called Split Adaptation (SA), shown in Figure 2. Our SA starts with dividing the pre-trained ViT $F_{\\Theta}$ into two parts: a frontend $F^{\\mathrm{F}}$ and a backend $F^{\\mathrm{B}}$ . The original $F^{\\mathrm{F}}$ is not sent to the client, instead, we leverage model quantization to send a quantized version $\\widehat{F}^{\\mathrm{F}}$ to the client. To mitigate the influence of frontend quantization, SA includes out-of-distribution (OOD) augmentations from both data and model levels to enhance the cross-distribution generalization of the frontend and backend. As for the client side, it adds random noise to the received $\\widehat{F}^{\\mathrm{F}}$ first and adopts a patch-level retrieve-then-replace strategy to generate more client data representations that help alleviate overfitting in few-shot adaptation. Before being uploaded to the server for the final adaptation, these augmented representations are protected by adding noise again.", + "bbox": [ + 511, + 643, + 906, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "20094", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/2fe7a701a2a45e6b89d7ae2ea5783a2e1e33ee6e9991aaf14b5e56e87849fd02.jpg", + "image_caption": [ + "Figure 2. Overview of Split Adaptation (SA) for pre-trained ViT adaptation. SA divides the pre-trained ViT into a frontend and a backend. After applying Out-of-distribution Enhanced Quantization to the frontend, its quantized version is sent to the client. To mitigate the impact of quantization, SA adopts Out-of-distribution Quantization-aware Tuning to enhance backend's generalization. As for the client, it injects random noise to the received frontend and retrieves then replaces randomly selected patches to augment more client data representations, which are sent to the server for the final adaptation after being added with noise again." + ], + "image_footnote": [], + "bbox": [ + 114, + 90, + 883, + 383 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Model Splitting on The Server Side", + "text_level": 1, + "bbox": [ + 89, + 468, + 397, + 484 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "At the beginning of SA, the pre-trained ViT is divided into a frontend $F^{\\mathrm{F}} \\coloneqq [f_{\\theta_1}, \\dots, f_{\\theta_K}]$ and a backend $F^{\\mathrm{B}} \\coloneqq [f_{\\theta_{K+1}}, \\dots, f_{\\theta_{LF}}]$ at the $K$ -th layer ( $K > L^{F} / 2$ ). Then the frontend is distributed to the client for representation extraction. Although we empirically found that the frontend cannot perform well (Section 4.3), we still hope to protect it from disclosing the original model parameters. Quantization [20, 34], as a representative model compression approach, can significantly reduce the computational and memory cost without sacrificing model performance. As quantization works by representing model weights and activations with low-precision values, it can potentially protect the model. However, standard quantization needs a certain amount of data drawn from the target task, which is inaccessible for the server in our problem. In this case, the left choice is to leverage the server dataset $\\mathcal{D}^{\\mathrm{S}}$ to quantize the model, but a dedicated design for enhancing the quantized model's cross-distribution generalization is needed. Therefore, we propose an augmentation technique to generate OOD data in terms of $\\mathcal{D}^{\\mathrm{S}}$ for quantization.", + "bbox": [ + 89, + 491, + 483, + 794 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3.1. Out-of-distribution Enhanced Quantization", + "text_level": 1, + "bbox": [ + 89, + 805, + 441, + 821 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In practice, we have to consider extreme cases in which the client dataset differs significantly from the server dataset and they may only share general semantics. Hence our augmentation technique should generate OOD data with a large discrepancy to $\\mathcal{D}^{\\mathrm{S}}$ . Inspired by signal processing, we no", + "bbox": [ + 89, + 824, + 483, + 902 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/33811de070aa6290bb3a984150073cb40b73014f565a79dd50816fc520b6808e.jpg", + "image_caption": [ + "Figure 3. Visualization of our Hilbert Transform data augmentation. Compared to amplitude-exchange augmentation [45] (bottom row), our augmentation method (top row) generates data that diverges further from the original in appearance while preserving the original semantics (dog category)." + ], + "image_footnote": [], + "bbox": [ + 517, + 469, + 901, + 654 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "tice that there is a technique called Hilbert Transform (HT), which only captures the most sharp edge in images. Intuitively, the sharp edges contain the most general semantics, making HT suitable for our problem. Applying HT to a signal in standard signal processing results in a phase shift of $\\pi /2$ in the frequency domain [36]. In the visual domain, we can implement HT by transforming an image $\\mathbf{I}$ to the frequency domain with the spectrum index $\\xi$ via Fourier transform, then the negative $\\xi$ needs to multiply with the imaginary unit $i$ , while the positive $\\xi$ needs to multiple with", + "bbox": [ + 511, + 750, + 906, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "20095", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$-i$ , i.e.,", + "bbox": [ + 91, + 92, + 145, + 104 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {H T} (\\mathbf {I}) = \\mathcal {F} ^ {- 1} \\{- i \\cdot \\operatorname {s g n} (\\xi) \\mathcal {F} (\\mathbf {I}) \\}, \\qquad (5)\n$$\n", + "text_format": "latex", + "bbox": [ + 171, + 113, + 482, + 132 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{F}(\\cdot)$ and $\\mathcal{F}^{-1}(\\cdot)$ are Fourier transform and inverse, respectively, and $\\mathrm{sgn}(\\cdot)$ is a sign function. However, pure HT only provides new images with black backgrounds. To diversify the HT-generated image, we replace its amplitude in the frequency domain with that of the original image. In this way, the generated image is much better than the standard amplitude-exchange augmentation [32, 45] (please refer to Figure 3 for a visual comparison).", + "bbox": [ + 89, + 137, + 483, + 258 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To achieve effective quantization, we adopt the most popular uniform symmetric quantization [20] to implement SA, in which a floating-point value $w$ is projected into a $k$ -bit integer value $w_{\\mathrm{q}}$ with a scaling factor $\\Delta$ :", + "bbox": [ + 89, + 260, + 483, + 320 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\widehat {w} = \\Psi_ {k} (w, \\Delta) = \\operatorname {c l a m p} (\\operatorname {r o u n d} \\left(\\frac {w}{\\Delta}\\right), - 2 ^ {k - 1}, 2 ^ {k - 1} - 1), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 96, + 327, + 483, + 364 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathrm{round}(\\cdot)$ turns a value into an integer and $\\mathrm{clamp}(\\cdot, \\min, \\max)$ constrains the output in the range from $\\min$ to $\\max$ . Intuitively, the scaling factor $\\Delta$ directly impacts the model performance after quantization. Previous studies [7, 34] have tried a variety of measurements to quantify the model output changes and accordingly determine the optimal scaling factors. Among these studies, we follow Yuan et al. [55] to choose a Hessian-guided metric for scaling factor determination. Suppose the pre-trained ViT $F_{\\Theta}$ is equipped with a task module $g_{\\gamma'}'$ that is tuned on the server dataset $\\mathcal{D}^{\\mathrm{S}}$ , the model performance can be measured by a loss $\\mathbb{E}_{\\boldsymbol{x} \\sim \\mathcal{P}_{\\boldsymbol{x}}^{\\mathrm{S}}, y \\sim \\mathcal{P}_{y}^{\\mathrm{S}}} \\mathcal{L}(\\boldsymbol{x}, y; \\Theta \\circ \\gamma')$ (we omit $\\boldsymbol{x}$ and $y$ to simplify the used symbols in the following descriptions). Then if we regard that the quantization brings a small perturbation $\\epsilon$ to the model parameters, i.e., $\\hat{\\Theta} = \\Theta + \\epsilon$ , we can analyze the influence of quantization on the task loss using Taylor series expansion:", + "bbox": [ + 89, + 364, + 483, + 626 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {E} [ \\mathcal {L} (\\widehat {\\Theta} \\circ \\gamma^ {\\prime}) ] - \\mathbb {E} [ \\mathcal {L} (\\Theta \\circ \\gamma^ {\\prime}) ] \\approx \\epsilon^ {\\top} \\nabla_ {\\Theta} + \\frac {1}{2} \\epsilon^ {\\top} \\mathbf {H} _ {\\Theta} \\epsilon , (7)\n$$\n", + "text_format": "latex", + "bbox": [ + 102, + 632, + 482, + 662 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\nabla_{\\Theta}$ is the gradients and $\\mathbf{H}_{\\Theta}$ is the Hessian matrix. Then the optimal scaling factors are the ones corresponding to minimal influence caused by quantization, i.e., $\\Delta = \\arg \\min_{\\Delta}(\\mathbb{E}[\\mathcal{L}(\\widehat{\\Theta}\\circ \\gamma^{\\prime})] - \\mathbb{E}[\\mathcal{L}(\\Theta \\circ \\gamma^{\\prime})])$ . Certainly, determining and adopting a general scaling factor for the entire model is too coarse to perform well, thus we apply a layer-wise reconstruction method [31] to achieve fine-grained quantization, then the searching optimization of Eq. (7) can be approximated by:", + "bbox": [ + 89, + 666, + 483, + 803 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\underset {\\Delta_ {l - 1}} {\\arg \\min } \\mathbb {E} \\left[ \\left(\\widehat {\\mathbf {X}} _ {l} - \\mathbf {X} _ {l}\\right) ^ {\\top} \\operatorname {d i a g} \\left(\\frac {\\partial \\mathcal {L}}{\\partial \\mathbf {X} _ {l}} \\odot \\frac {\\partial \\mathcal {L}}{\\partial \\mathbf {X} _ {l}}\\right) \\left(\\widehat {\\mathbf {X}} _ {l} - \\mathbf {X} _ {l}\\right) \\right], \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 808, + 488, + 853 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{X}_l$ and $\\widehat{\\mathbf{X}}_l$ are the outputs of the $(l-1)$ -th layer before and after quantization, respectively, and $\\odot$ represents the Hadamard product of a matrix.", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "With the augmented neighborhood data, the frontend can be quantized with improved cross-distribution generalization. Specifically, we first conduct HT-based data augmentation on the server dataset $\\mathcal{D}^{\\mathrm{S}}$ to obtain $\\mathrm{HT}(\\mathcal{D}^{\\mathrm{S}})$ . Then the optimal scaling factor for each layer in the frontend $F^{\\mathrm{F}}$ is empirically determined by optimizing Eq. (8) on the merged dataset $\\mathcal{D}_{\\cup}^{\\mathrm{S}} = \\mathcal{D}^{\\mathrm{S}} \\cup \\mathrm{HT}(\\mathcal{D}^{\\mathrm{S}})$ and we can obtain a quantized frontend $\\widehat{F}^{\\mathrm{F}}$ in the end.", + "bbox": [ + 511, + 90, + 906, + 213 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3.2. Out-of-distribution Quantization-aware Tuning", + "text_level": 1, + "bbox": [ + 511, + 219, + 890, + 234 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To achieve better adaptation performance, the backend also needs to resist the unavoidable parameter difference between the original and quantized model as well as the noise injected by the client that we will introduce in Section 3.4. As a result, we propose an OOD quantization-aware tuning technique to enhance the backend's resistance when encountering OOD data and frontends.", + "bbox": [ + 511, + 238, + 905, + 343 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To simulate potential OOD frontends, we first divide the merged server dataset $\\mathcal{D}_{\\cup}^{\\mathrm{S}}$ into $M$ subsets and denote these subsets as $\\{\\mathcal{D}_{\\cup_m}^{\\mathrm{S}}\\}_{m = 1}^M$ . Next, each subset $\\mathcal{D}_{\\cup_m}^{\\mathrm{S}}$ is used to quantize the original frontend $F^{\\mathrm{F}}$ to obtain a quantized version $\\widehat{F}_m^{\\mathrm{F}}$ . Then let us use a single quantized frontend $\\widehat{F}_m^{\\mathrm{F}}$ to introduce how to tune the backend. To begin with, since the backend is represented and the tuning is performed using floating-point values, the data representations extracted by the quantized frontend should be transformed back to floating-point values through the dequantization:", + "bbox": [ + 511, + 344, + 906, + 494 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {w} = \\Psi^ {- 1} (\\hat {w}, \\Delta) = \\hat {w} \\cdot \\Delta . \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 619, + 502, + 906, + 520 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this way, we can obtain the floating-point representation $\\widetilde{\\mathbf{X}}_K$ by dequantizing the $K$ -th layer's output with the scaling factor $\\Delta_K$ . Note that $\\widetilde{\\mathbf{X}}_K$ is different from the original representation $\\mathbf{X}_K$ extracted by the original frontend $F^{\\mathrm{F}}$ due to the range constraints of round and clamp functions in Eq. (6). Our following design aims to tune the backend to resist such differences. The most straightforward method is to forward the floating-point representation $\\widetilde{\\mathbf{X}}_K$ to the backend and conduct tuning using task loss. However, such tuning will make the backend over-fitted to the server dataset $\\mathcal{D}^{\\mathrm{S}}$ . We solve this by a representation-mixup strategy. Specifically, for a particular data sample $\\pmb{x}$ , we feed it to a quantized frontend $\\widehat{F}_{m}^{\\mathrm{F}}$ and the original frontend $F^{\\mathrm{F}}$ to extract representations $\\widehat{\\mathbf{X}}_{K,m},\\mathbf{X}_{K,m}$ . After dequantizing $\\widehat{\\mathbf{X}}_{K,m}$ into the floating-point form $\\widetilde{\\mathbf{X}}_{K,m}$ , we obtain a mixup representation:", + "bbox": [ + 511, + 527, + 906, + 773 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {X} _ {K, m} ^ {*} = \\lambda \\widetilde {\\mathbf {X}} _ {K, m} + (1 - \\lambda) \\mathbf {X} _ {K, m}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 593, + 780, + 903, + 799 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda$ is drawn from a widely-used beta distribution $\\mathrm{Beta}(\\lambda;0.75,0.75)$ . Then considering all available quantized frontend, we can tune the backend with the loss:", + "bbox": [ + 511, + 806, + 905, + 851 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {t u n e}} = \\sum_ {\\mathcal {D} _ {\\cup} ^ {\\mathrm {S}}} \\sum_ {m = 0} ^ {M} \\mathcal {L} \\left(\\mathbf {X} _ {K, m} ^ {*}, y; \\Theta^ {\\mathrm {B}} \\circ \\gamma^ {\\prime}\\right). \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 858, + 906, + 902 + ], + "page_idx": 4 + }, + { + "type": "footer", + "text": "20096", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Note that $\\mathbf{X}_{K,0}^{*}$ originates from feeding $\\pmb{x}$ to the frontend $\\widehat{F}^{\\mathrm{F}}$ quantized on the full set of $\\mathcal{D}_{\\cup}^{\\mathrm{S}}$ .", + "bbox": [ + 89, + 90, + 482, + 126 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.4. Client Bi-level Noisy Representation Extraction", + "text_level": 1, + "bbox": [ + 89, + 136, + 482, + 152 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In our SA model, the quantized frontend $\\widehat{F}^{\\mathrm{F}}$ will be sent to the client after the OOD enhanced quantization (Section 3.3.1), and then the client can forward its data to $\\widehat{F}^{\\mathrm{F}}$ to extract representations and send them to the server. However, there is a risk of data reconstruction when sharing representations, thus we propose a bi-level noisy representation extraction mechanism to defend against such attacks.", + "bbox": [ + 89, + 159, + 482, + 266 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Intuitively speaking, the key to data reconstruction defense is to use the model unknown by the server to extract the representations. Therefore, we allow the client to inject a certain degree of random noise into the received frontend $\\widehat{F}^{\\mathrm{F}}$ . For each layer in the frontend, we inject a multiplicative noise $n_l^\\times$ , and an additive noise $n_l^+$ , which are sampled from two Gaussian distributions, i.e., $n_l^\\times \\sim \\mathcal{P}_{\\mathcal{N}(1,\\alpha \\theta_l)}$ and $n_l^+ \\sim \\mathcal{P}_{\\mathcal{N}(0,\\alpha \\theta_l)}$ ( $\\alpha = 0.01$ controls the noise degree), to the model parameters:", + "bbox": [ + 89, + 267, + 483, + 404 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\theta_ {l} ^ {\\prime} = n _ {l} ^ {\\times} \\cdot \\theta_ {l} + n _ {l} ^ {+}. \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 220, + 417, + 482, + 435 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Next, the client feeds its dataset $\\mathcal{D}^{\\mathrm{C}}$ to the noisy frontend $\\widehat{F}^{\\mathrm{F}'}$ to extract representations $\\{\\widehat{\\mathbf{X}}_{K,i}^{\\mathrm{C}}\\}_{i=1}^{N^{\\mathrm{C}}}$ . However, it is still insecure to share these representations. We found that our model-level noise can amplify the defense effect of the regular noise obfuscation [24] at the representation level, which allows to add a smaller degree of Laplace noise $(\\mathcal{P}_{\\mathrm{Laplace}(0,0.8)})$ to representations for better protection.", + "bbox": [ + 89, + 446, + 483, + 556 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.5. Patch-retrieval-augmented Model Adaptation", + "text_level": 1, + "bbox": [ + 89, + 566, + 477, + 583 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In few-shot adaptation scenarios, like the backend tuning on the server (Section 3.3.2), the adaptation also faces overfitting challenges, but the overfitting here is due to the limited client data available. To tackle such challenges, we design a patch-level retrieval augmentation approach to generate more client data representations.", + "bbox": [ + 89, + 589, + 482, + 679 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our patch-level retrieval augmentation is inspired by the intuition that different images may contain similar patches. As introduced in Section 3.1.1, the data representations of each layer in a ViT hold the dimension $(N + 1)\\times d$ , which can be denoted as $\\widehat{\\mathbf{X}}_{K,i}^{\\mathrm{C}} = [\\mathbf{x}_i^{\\mathrm{CLS}};\\mathbf{x}_{i,1},\\dots ,\\mathbf{x}_{i,N}]$ (we omit some symbols for simplification). Then we can form $N$ retrieval sets from all client data representations (before adding Laplace noise) and each corresponds to a patch, i.e., $\\{S_j\\}_{j = 1}^N$ where $S_{j} = \\{\\mathbf{x}_{i,j}\\}_{i = 1}^{N^{\\mathrm{C}}}$ . With these sets, for each augmentation of $\\widehat{\\mathbf{X}}_{K,i}^{\\mathrm{C}}$ , we randomly select $N^{\\mathrm{P}} = 50$ patches (please see the sensitivity analysis in Supplementary Materials) $\\{\\mathbf{x}_{i,\\mathrm{P}_j}\\}_{j = 1}^{N^{\\mathrm{P}}}$ (subscript $\\mathrm{P}_j$ meets $1\\leq \\mathrm{P}_j\\leq N$ ) and delete them from their corresponding retrieval sets. Next, each $\\mathbf{x}_{i,\\mathrm{P}_j}$ retrieves the most similar patch $\\Omega (\\mathbf{x}_{i,\\mathrm{P}_j})$ from", + "bbox": [ + 89, + 681, + 483, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "its corresponding retrieval set:", + "bbox": [ + 513, + 90, + 715, + 106 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\Omega \\left(\\mathbf {x} _ {i, \\mathrm {P} _ {j}}\\right) = \\arg \\max _ {\\mathbf {x} \\in S _ {\\mathrm {P} _ {j}} ^ {\\prime}} \\frac {\\mathbf {x} _ {i , \\mathrm {P} _ {j}} \\cdot \\mathbf {x}}{\\left| \\mathbf {x} _ {i , \\mathrm {P} _ {j}} \\right| \\left| \\mathbf {x} \\right|}, \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 596, + 114, + 903, + 148 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\mathcal{S}_{\\mathrm{P}_j}^{\\prime} = \\mathcal{S}_{\\mathrm{P}_j}\\backslash \\mathbf{x}_{i,\\mathrm{P}_j}$ . After the retrieval operations for all selected patches $\\{\\mathbf{x}_{i,\\mathrm{P}_j}\\}_{j = 1}^{N^{\\mathrm{P}}}$ , we replace them with the retrieved patches $\\{\\Omega (\\mathbf{x}_{i,\\mathrm{P}_j})\\}_{j = 1}^{N^{\\mathrm{P}}}$ in the original representation $\\widehat{\\mathbf{X}}_{K,i}^{\\mathrm{C}}$ to form a new one $\\widehat{\\mathbf{X}}_{K,i}^{\\prime \\mathrm{C}}$ . Note that the selected patches are determined randomly, resulting in newly generated representations that differ across augmentation runs. To generate sufficient augmented representations, we run the augmentation $N^{\\mathrm{Aug}}$ times for all client data representations and produce an augmented representation set, i.e., $\\widehat{\\mathcal{X}}_K^{\\mathrm{C}} = \\{\\{\\widehat{\\mathbf{X}}_{K,i}^{\\mathrm{C}_j}\\}_{j = 0}^{N^{\\mathrm{Aug}}}\\}_{i = 1}^{N^{\\mathrm{C}}}$ (we denote the original representation as $\\widehat{\\mathbf{X}}_{K,i}^{\\mathrm{C}_0}$ and the augmented representations as $\\widehat{\\mathbf{X}}_{K,i}^{\\mathrm{C}_{1\\sim N^{\\mathrm{Aug}}}}$ ). The above patch retrieval augmentation is run by the client, and it adds Laplace noise (Section 3.4) to all representations in $\\widehat{\\mathcal{X}}_K^{\\mathrm{C}}$ before uploading to the server.", + "bbox": [ + 511, + 159, + 906, + 393 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "After receiving the representations, the server dequantizes them into floating-point form with the $K$ -th layer's scaling factor $\\Delta_K$ by Eq. (9) and launches the adaptation. To further alleviate the overfitting due to limited client data, different from Split Learning, we only allow high-level feature learning in the backend to prevent low-level features from becoming biased. Then the server randomly initializes a task module $g_{\\gamma}$ and tunes it with the backend $F^{\\mathrm{B}}$ (that has been tuned in Section 3.3.2) by following the loss:", + "bbox": [ + 511, + 395, + 905, + 531 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {a d a p t}} = \\sum_ {i = 1} ^ {N ^ {\\mathrm {C}}} \\sum_ {j = 0} ^ {N ^ {\\mathrm {A u g}}} \\mathcal {L} \\left(\\widehat {\\mathbf {X}} _ {K, i} ^ {\\mathrm {C} _ {j}}, y; \\Theta^ {\\mathrm {B}} \\circ \\gamma\\right). \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 540, + 903, + 585 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To protect the client's labels, the server sends the task module output to the client, allowing the task loss $\\mathcal{L}_{\\mathrm{adapt}}$ to be computed on the client side and then sent back to the server.", + "bbox": [ + 511, + 595, + 905, + 642 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 654, + 645, + 671 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experimental Settings", + "text_level": 1, + "bbox": [ + 511, + 679, + 720, + 695 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. ImageNet [9] is the most widely used image dataset collected from public domains, thus we use its test set as the server dataset. For the client datasets, we use CIFAR-100 [30], Places365 [35], and DomainNet [39]. Places365 is a scene recognition dataset that contains 365 classes. DomainNet (D.Net) is a challenging multi-domain dataset, which includes six domains. We select the smallest domain—Clipart (Cl) for the few-shot adaptation scenarios.", + "bbox": [ + 511, + 700, + 906, + 821 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details. We conduct experiments on three pre-trained model architectures: ViT-Large [13], DeiT-Base [43], and Swin-L [33] (DeiT-Base and Swin-L's experiments are reported in Supplementary Materials). For the Places365 dataset, the target task module is composed", + "bbox": [ + 511, + 825, + 905, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "20097", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/823c28bae9ed0628ff522302e07475c10eae5ff400e07dfac6deec139d602c36.jpg", + "table_caption": [ + "Table 1. Performance comparison between our SA and other baseline approaches in 3-shot, 5-shot, and 10-shot adaptation scenarios over 3 datasets. The SA can substantially exceed other methods. We bold and blue the best, and bold the second best." + ], + "table_footnote": [], + "table_body": "
Few-shot Setup3-shot5-shot10-shot
MethodCIFAR-100Places365D.Net-ClCIFAR-100Places365D.Net-ClCIFAR-100Places365D.Net-Cl
Linear Probing74.05±1.2126.18±1.0057.26±1.1178.79±0.6534.25±1.4563.84±0.6683.28±0.3332.33±7.8269.12±0.75
Fine Tuning40.32±5.4821.64±0.7836.95±6.6664.29±7.6931.07±1.0456.50±1.8781.82±1.8539.46±0.5970.67±0.72
LN-TUNE13.18±2.171.59±0.536.97±1.9634.80±5.373.96±0.9513.43±2.1247.97±6.257.71±1.8524.53±2.05
Split Learning74.05±1.2126.84±1.0358.14±0.6379.63±0.5330.49±0.3663.95±1.1883.66±0.2735.78±0.3065.20±0.47
Offsite Tuning42.37±2.9824.98±0.5540.56±4.0064.83±6.1730.45±0.8056.43±0.4580.11±1.0736.22±1.3368.43±0.64
SA (ours)76.24±0.2930.92±0.8956.26±0.7981.98±0.4935.31±0.0165.03±0.5985.45±0.4039.26±0.1071.13±0.76
", + "bbox": [ + 109, + 118, + 879, + 247 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "of two linear layers, whereas for other datasets, it consists of a single linear layer. We divide 2/3 backbone model as the frontend while the rest $1/3$ is the backend. The frontend and its inner activations are quantized to 8-bit. Each quantization operation uses 32 data samples randomly sampled from the server dataset or subsets, and the batch size is 4. The Adam optimizer with a learning rate of $10^{-5}$ is adopted. The number of epochs for OOD quantization-aware tuning is set to 1, while the final adaptation is set to 100 epochs. All experiments are run repeatedly 3 times with three seeds, and we report the mean performance and standard deviation.", + "bbox": [ + 88, + 263, + 480, + 431 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Baseline Approaches. SA is compared with state-of-the-art downstream task adaptation methods including Linear Probing [29], Fine Tuning, Split Learning [44], Offsite Tuning [49], and LN-TUNE [2]. To evaluate the effectiveness of data protection, we extend the state-of-the-art data reconstruction attacks—FORA [53]—to ViT. More details are provided in the Supplementary Materials.", + "bbox": [ + 89, + 436, + 482, + 542 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Evaluation Metrics. To evaluate the effectiveness of the target task adaptation, we use classification accuracy to measure the performance. As for the quality measurement of data reconstruction results, we adopt three metrics: SSIM [48] (S), PSNR [25] (P), and LPIPS [56] (L).", + "bbox": [ + 89, + 547, + 482, + 625 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Effectiveness of SA for Task Adaptation", + "text_level": 1, + "bbox": [ + 89, + 635, + 429, + 651 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We create three few-shot adaptation scenarios for each client dataset or domain: 3-shot, 5-shot, and 10-shot per class, with experimental results shown in Table 1. We can easily observe that our SA can always substantially exceed all other baseline approaches. In particular, methods like Fine Tuning, LN-TUNE, and Offsite Tuning lag far behind SA, which may be attributed to the overfitting caused by the low-level feature tuning. The method closest to SA in terms of performance is Split Learning, but its biggest issue is the privacy leakage, which we will discuss in Section 4.3. In Table 2, we also provide the client memory and computation costs of each method (assuming the first three methods are conducted on the client side), from which we can observe that SA even runs with the smallest cost. Note that SA has no training on the client, thus we view the representation extraction time as the training time in Table 2.", + "bbox": [ + 89, + 659, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Effectiveness of SA for Server Pre-trained ViT and Client Data Protection", + "text_level": 1, + "bbox": [ + 511, + 263, + 903, + 292 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For privacy protection, we first test whether the quantized frontend of SA can result in high-quality models with good downstream task performance. We consider the three most potential variations: 1) directly tune a task module on the quantized frontend (Quant. Frontend); 2) tune a task module on the original frontend though this case is impossible in practice (Original Frontend); 3) attach a backend from other models and tune a task module on the quantized frontend and auxiliary backend (Auxiliary Backend). The experimental results are shown in Table 3, in which all variations perform much poorer than Linear Probing in Table 1 and our SA. In this case, we believe SA can protect the pretrained model to a great extent.", + "bbox": [ + 509, + 300, + 906, + 496 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Then we launch FORA [53] to evaluate the data protection of our SA. For a fair comparison, only Split Learning is compared with SA as only it requires the client to share intermediate data representations with the server. The experiment results are presented in Table 5. First, we can visually observe that the reconstruction results of SA are much worse than those of Split Learning. Then the three metrics also demonstrate the same phenomenon, i.e., FORA can only recover much lower-quality images when attacking SA compared with Split Learning. In summary, our SA offers strong protection of the client data. More experiments on the noise degree can be found in Supplementary Materials.", + "bbox": [ + 511, + 497, + 908, + 679 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Ablation Study", + "text_level": 1, + "bbox": [ + 511, + 688, + 666, + 704 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We conduct comprehensive ablation studies in the 5-shot adaptation on CIFAR-100, Places365, and DomainNet-C1.", + "bbox": [ + 511, + 710, + 906, + 741 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Hilbert Transform-based Data Augmentation. We detach the use of HT data augmentation in the frontend quantization and backend tuning ('SA w/o HT Aug'). Compared with the full SA in Table 4, the variation without HT data augmentation performs worse, which proves its gain in improving adaptation performance.", + "bbox": [ + 511, + 744, + 905, + 835 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Out-of-distribution Quantization-aware Tuning. We design two variations: one is not to use OOD quantization-aware tuning ('SA w/o OOD QAT'), and the other is only to use the quantized frontend $\\widehat{F}_{\\mathrm{F}}$ characterized on the full", + "bbox": [ + 511, + 839, + 905, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "20098", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/b8294088e319c148ccd3808d318e5ff790dce029fdcc3469c96500ac03b49c18.jpg", + "table_caption": [ + "Table 2. Comparison of client memory (MB) and computation costs (Min) between SA and other baselines." + ], + "table_footnote": [], + "table_body": "
MethodsGPU MemoryTraining Time
Linear Probing697966
Fine Tuning1030234
LN-TUNE596825
Split Learning893215
Offsite Tuning489623
SA (ours)22332.5
", + "bbox": [ + 94, + 132, + 321, + 229 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/8f39c861de998f111a0cbf29ddf619822333c156a6ef7aec08db6cce6b27d4a5.jpg", + "table_caption": [ + "Table 3. Risk assessment of 3 most potential cases over 3 datasets (domains) for obtaining high-quality models from our SA approach." + ], + "table_footnote": [], + "table_body": "
VariationsCIFAR-100Places365D.Net-CI
Quant. Frontend26.3413.4526.70
Original Frontend26.5313.7227.12
Auxiliary Backend30.6510.1516.56
SA (ours)81.9835.3171.13
", + "bbox": [ + 336, + 138, + 612, + 218 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/2271bb2d9e6ad17b4bf7c6f291c36842de76d68dda71e5de1087315b0d9b0a5a.jpg", + "table_caption": [ + "Table 4. Ablation study of the major components of our SA approach in 5-shot adaptation over 3 datasets (domains)." + ], + "table_footnote": [], + "table_body": "
VariationsCIFAR-100Places365D.Net-CI
SA w/o HT Aug80.3735.2764.59
SA w/o OOD QAT79.0329.5064.42
SA w/o QAT80.0529.5964.16
SA w/o PR Aug79.9029.5652.85
SA (ours)81.9835.3171.13
", + "bbox": [ + 627, + 138, + 903, + 229 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/baea76b36a7f3bef313a56fa507fe3b1024a7d8e3ed066856a040e76f0744159.jpg", + "table_caption": [ + "Table 5. Comparison of defensive performance against the state-of-the-art data reconstruction attack—FORA [53]—between SA and Split Learning. We visualize the reconstruction and calculate SSIM (S↓), PSNR (P↓), and LPIPS (L↑) to measure the reconstruction quality." + ], + "table_footnote": [], + "table_body": "
DatasetCIFAR-100Places365D.Net-CI
Ground Truth
Split LearningS:0.80 P:26.1 L:0.46S:0.43 P:20.8 L:0.67S:0.59 P:22.1 L:0.46
SA (ours)S:0.23 P:15.2 L:0.70S:0.30 P:16.6 L:0.73S:0.25 P:14.4 L:0.60
", + "bbox": [ + 94, + 272, + 903, + 465 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/029a508b0bc73126c249ae3ef42e5d1424f427ba040e68b9f88a3bb15715c5c3.jpg", + "table_caption": [ + "Table 6. Ablation study of SA's bi-level noise when defending against FORA [53] on CIFAR-100." + ], + "table_footnote": [], + "table_body": "
DatasetCIFAR-100
Ground Truth-
SA w/o Laplace NoiseS:0.76 P:23.1 L:0.49
SA w/o Model NoiseS:0.28 P:16.4 L:0.67
SA (ours)S:0.23 P:15.2 L:0.70
", + "bbox": [ + 106, + 506, + 467, + 763 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "server dataset ('SA w/o QAT'). Compared with the full SA, the effectiveness of the representation-mixup quantization-aware tuning and the simulated OOD frontends in alleviating quantization-induced performance drop can be verified.", + "bbox": [ + 89, + 776, + 482, + 837 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Patch Retrieval Augmentation. We don't use the patch retrieval augmentation and observe how the model performance will change. The big difference between 'SA w/o PR Aug' and 'SA' in Table 4 indicates the essential role of our", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "augmentation here in performance enhancement.", + "bbox": [ + 511, + 479, + 836, + 494 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Bi-level Noisy Representation Extraction. We detach the use of model Gaussian noise ('SA w/o Model Noise') and representation Laplace noise ('SA w/o Laplace Noise'), respectively, and the experiments are shown in Table 6. Model noise and representation noise play a synergistic role greater than the simple sum of them in data protection.", + "bbox": [ + 511, + 498, + 906, + 590 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 606, + 633, + 621 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we propose Split Adaptation (SA) to adapt pre-trained ViTs for downstream tasks while protecting the data and model in the sever-client framework. After dividing the ViT into a frontend and a backend, SA innovatively uses model quantization to protect the frontend. SA also allows the client to introduce a bi-level noise to the quantized frontend and the extracted data representations for data protection. Accordingly, SA uses data-level and model-level augmentation to mitigate the influence of the quantization and the noise. Besides, SA proposes patch retrieval augmentation to address overfitting in few-shot adaptation. Extensive experiments have validated the superiority of SA in task adaptation and data-model protection.", + "bbox": [ + 511, + 632, + 906, + 829 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 513, + 844, + 671, + 861 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work was partially supported by NTU-NAP startup grant 024584-00001.", + "bbox": [ + 511, + 869, + 906, + 901 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "20099", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Corinne G Allaart, Björn Keyser, Henri Bal, and Aart Van Halteren. Vertical split learning—an exploration of predictive performance in medical and other use cases. In 2022 International Joint Conference on Neural Networks (IJCNN), pages 1-8. IEEE, 2022. 1, 2", + "[2] Samyadeep Basu, Shell Hu, Daniela Massiceti, and Soheil Feizi. Strong baselines for parameter-efficient few-shot finetuning. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 11024-11031, 2024. 7, 1", + "[3] Tom B Brown. Language models are few-shot learners. arXiv preprint arXiv:2005.14165, 2020. 1", + "[4] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 2, 1", + "[5] Shoufa Chen, Chongjian Ge, Zhan Tong, Jiangliu Wang, Yibing Song, Jue Wang, and Ping Luo. Adaptformer: Adapting vision transformers for scalable visual recognition. Advances in Neural Information Processing Systems, 35:16664-16678, 2022. 2", + "[6] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 2", + "[7] Yoni Choukroun, Eli Kravchik, Fan Yang, and Pavel Kisilev. Low-bit quantization of neural networks for efficient inference. In 2019 IEEE/CVF International Conference on Computer Vision Workshop (ICCVW), pages 3009-3018. IEEE, 2019. 5", + "[8] Zeki Murat Çınar, Abubakar Abdussalam Nuhu, Qasim Zeešan, Orhan Korhan, Mohammed Asmael, and Babak Safaei. Machine learning in predictive maintenance towards sustainable smart manufacturing in industry 4.0. Sustainability, 12 (19):8211, 2020. 1", + "[9] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 6", + "[10] Matthew F Dixon, Igor Halperin, and Paul Bilokon. Machine learning in finance. Springer, 2020. 1", + "[11] Jiahua Dong, Lixu Wang, Zhen Fang, Gan Sun, Shichao Xu, Xiao Wang, and Qi Zhu. Federated class-incremental learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10164–10173, 2022. 1", + "[12] Wei Dong, Qiyao Luo, and Ke Yi. Continual observation under user-level differential privacy. In 2023 IEEE Symposium on Security and Privacy (SP), pages 2190-2207. IEEE, 2023. 1", + "[13] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. ICLR, 2020. 1, 2, 3, 6", + "[14] Juanru Fang, Wei Dong, and Ke Yi. Shifted inverse: A general mechanism for monotonic functions under user differential privacy. In Proceedings of the 2022 ACM SIGSAC Con-" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ference on Computer and Communications Security, pages 1009-1022, 2022. 1", + "[15] Chongyang Gao, Lixu Wang, Chenkai Weng, Xiao Wang, and Qi Zhu. Practical unlearning for large language models. arXiv preprint arXiv:2407.10223, 2024. 1", + "[16] Xinben Gao and Lan Zhang. {PCAT}: Functionality and data stealing from split learning by {Pseudo-Client} attack. In 32nd USENIX Security Symposium (USENIX Security 23), pages 5271-5288, 2023. 1, 2", + "[17] Yansong Gao, Minki Kim, Sharif Abuadbba, Yeonjae Kim, Chandra Thapa, Kyuyeon Kim, Seyit A Camtep, Hyoungshick Kim, and Surya Nepal. End-to-end evaluation of federated learning and split learning for internet of things. In 2020 International Symposium on Reliable Distributed Systems (SRDS), pages 91-100. IEEE Computer Society, 2020. 2", + "[18] Jiahui Gong, Jingtao Ding, Fanjin Meng, Guilong Chen, Hong Chen, Shen Zhao, Haisheng Lu, and Yong Li. A population-to-individual tuning framework for adapting pretrained lm to on-device user intent prediction. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 896-907, 2024. 2", + "[19] Junfeng Guo, Yiming Li, Lixu Wang, Shu-Tao Xia, Heng Huang, Cong Liu, and Bo Li. Domain watermark: Effective and harmless dataset copyright protection is closed at hand. Advances in Neural Information Processing Systems, 36:54421-54450, 2023. 1", + "[20] Song Han, Huizi Mao, and William J Dally. Deep compression: Compressing deep neural networks with pruning, trained quantization and huffman coding. *ICLR*, 2016. 4, 5", + "[21] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 2", + "[22] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 2", + "[23] Zecheng He, Tianwei Zhang, and Ruby B Lee. Model inversion attacks against collaborative inference. In Proceedings of the 35th Annual Computer Security Applications Conference, pages 148-162, 2019. 1, 2", + "[24] Zecheng He, Tianwei Zhang, and Ruby B Lee. Attacking and protecting data privacy in edge-cloud collaborative inference systems. IEEE Internet of Things Journal, 8(12):9706-9716, 2020. 6", + "[25] Alain Hore and Djemel Ziou. Image quality metrics: Psnr vs. ssm. In 2010 20th international conference on pattern recognition, pages 2366-2369. IEEE, 2010. 7", + "[26] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In European Conference on Computer Vision, pages 709-727. Springer, 2022. 2", + "[27] Peter Kairouz, H Brendan McMahan, Brendan Avent, Aurélien Bellet, Mehdi Bennis, Arjun Nitin Bhagoji, Kallista" + ], + "bbox": [ + 516, + 93, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "20100", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Bonawitz, Zachary Charles, Graham Cormode, Rachel Cummings, et al. Advances and open problems in federated learning. Foundations and trends® in machine learning, 14(1-2): 1-210, 2021. 1", + "[28] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 1, 2", + "[29] Simon Kornblith, Jonathon Shlens, and Quoc V Le. Do better imagenet models transfer better? In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2661-2671, 2019. 2, 7, 1", + "[30] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 6", + "[31] Yuhang Li, Ruihao Gong, Xu Tan, Yang Yang, Peng Hu, Qi Zhang, Fengwei Yu, Wei Wang, and Shi Gu. Brecq: Pushing the limit of post-training quantization by block reconstruction. In International Conference on Learning Representations. 5", + "[32] Chenxi Liu, Lixu Wang, Lingjuan Lyu, Chen Sun, Xiao Wang, and Qi Zhu. Deja vu: Continual model generalization for unseen domains. In The Eleventh International Conference on Learning Representations, 2023. 5", + "[33] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 6, 1", + "[34] Zhenhua Liu, Yunhe Wang, Kai Han, Wei Zhang, Siwei Ma, and Wen Gao. Post-training quantization for vision transformer. Advances in Neural Information Processing Systems, 34:28092-28103, 2021. 4, 5", + "[35] Alejandro López-Cifuentes, Marcos Escudero-Vinolo, Jesús Bescós, and Álvaro García-Martín. Semantic-aware scene recognition. Pattern Recognition, 102:107256, 2020. 6", + "[36] Payal Mohapatra, Lixu Wang, and Qi Zhu. Phase-driven domain generalizable learning for nonstationary time series. arXiv preprint arXiv:2402.05960, 2024. 4", + "[37] Mehdi Noroozi and Paolo Favaro. Unsupervised learning of visual representations by solving jigsaw puzzles. In European conference on computer vision, pages 69-84. Springer, 2016. 2", + "[38] Dario Pasquini, Giuseppe Ateniese, and Massimo Bernaschi. Unleashing the tiger: Inference attacks on split learning. In Proceedings of the 2021 ACM SIGSAC Conference on Computer and Communications Security, pages 2113-2129, 2021. 1, 2", + "[39] Xingchao Peng, Qinxun Bai, Xide Xia, Zijun Huang, Kate Saenko, and Bo Wang. Moment matching for multi-source domain adaptation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1406-1415, 2019. 6", + "[40] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervi" + ], + "bbox": [ + 91, + 90, + 483, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "sion. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 2", + "[41] K Shailaja, Banoth Seetharamulu, and MA Jabbar. Machine learning in healthcare: A review. In 2018 Second international conference on electronics, communication and aerospace technology (ICECA), pages 910-914. IEEE, 2018. 1", + "[42] Chandra Thapa, Mahawaga Arachchige Pathum Chamikara, and Seyit A Camtepe. Advancements of federated learning towards privacy preservation: from federated learning to split learning. Federated Learning Systems: Towards Next-Generation AI, pages 79–109, 2021. 2", + "[43] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In International conference on machine learning, pages 10347-10357. PMLR, 2021. 6, 1", + "[44] Praneeth Vepakomma, Otkrist Gupta, Tristan Swedish, and Ramesh Raskar. Split learning for health: Distributed deep learning without sharing raw patient data. arXiv preprint arXiv:1812.00564, 2018. 1, 2, 7", + "[45] Jingye Wang, Ruoyi Du, Dongliang Chang, Kongming Liang, and Zhanyu Ma. Domain generalization via frequency-domain-based feature disentanglement and interaction. In Proceedings of the 30th ACM International Conference on Multimedia, pages 4821-4829, 2022. 4, 5", + "[46] Lixu Wang, Shichao Xu, Xiao Wang, and Qi Zhu. Addressing class imbalance in federated learning. In Proceedings of the AAAI conference on artificial intelligence, pages 10165-10173, 2021. 1", + "[47] Lixu Wang, Shichao Xu, Ruiqi Xu, Xiao Wang, and Qi Zhu. Non-transferable learning: A new approach for model ownership verification and applicability authorization. In International Conference on Learning Representation (ICLR-2022), 2022. 1", + "[48] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 7", + "[49] Guangxuan Xiao, Ji Lin, and Song Han. Offsite-tuning: Transfer learning without full model. arXiv preprint arXiv:2302.04870, 2023. 1, 2, 7", + "[50] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. Advances in neural information processing systems, 34: 12077-12090, 2021. 1", + "[51] Chengming Xu, Siqian Yang, Yabiao Wang, Zhanxiong Wang, Yanwei Fu, and Xiangyang Xue. Exploring efficient few-shot adaptation for vision transformers. Transactions on Machine Learning Research. 2", + "[52] Peng Xu, Xiatian Zhu, and David A Clifton. Multimodal learning with transformers: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(10):12113-12132, 2023. 1", + "[53] Xiaoyang Xu, Mengda Yang, Wenzhe Yi, Ziang Li, Juan Wang, Hongxin Hu, Yong Zhuang, and Yaxin Liu. A stealthy" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "20101", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "wrongdoer: Feature-oriented reconstruction attack against split learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12130-12139, 2024. 1, 2, 7, 8", + "[54] Yupeng Yin, Xianglong Zhang, Huanle Zhang, Feng Li, Yue Yu, Xiuzhen Cheng, and Pengfei Hu. Ginver: Generative model inversion attacks against collaborative inference. In Proceedings of the ACM Web Conference 2023, pages 2122-2131, 2023. 1, 2", + "[55] Zhihang Yuan, Chenhao Xue, Yiqi Chen, Qiang Wu, and Guangyu Sun. Ptq4vit: Post-training quantization for vision transformers with twin uniform quantization. In European conference on computer vision, pages 191-207. Springer, 2022. 5", + "[56] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 7", + "[57] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In Proceedings of the IEEE international conference on computer vision, pages 2223-2232, 2017. 2" + ], + "bbox": [ + 91, + 90, + 483, + 428 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "20102", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Split Adaptation for Pre-trained Vision Transformers/b5d6478d-8866-4dba-8ddb-9d84bd75792a_model.json b/2025/Split Adaptation for Pre-trained Vision Transformers/b5d6478d-8866-4dba-8ddb-9d84bd75792a_model.json new file mode 100644 index 0000000000000000000000000000000000000000..7385d6c5bb86db78aa94f192cf68df2c28ea4d95 --- /dev/null +++ b/2025/Split Adaptation for Pre-trained Vision Transformers/b5d6478d-8866-4dba-8ddb-9d84bd75792a_model.json @@ -0,0 +1,2268 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.229, + 0.131, + 0.77, + 0.154 + ], + "angle": 0, + "content": "Split Adaptation for Pre-trained Vision Transformers" + }, + { + "type": "text", + "bbox": [ + 0.134, + 0.18, + 0.862, + 0.218 + ], + "angle": 0, + "content": "Lixu Wang\\(^{1,2*}\\), Bingqi Shang\\(^{1*}\\), Yi Li\\(^{1}\\), Payal Mohapatra\\(^{1}\\), Wei Dong\\(^{2}\\), Xiao Wang\\(^{1}\\), Qi Zhu\\(^{1}\\), Northwestern University. \\(^{2}\\)Nanyang Technological University." + }, + { + "type": "text", + "bbox": [ + 0.105, + 0.22, + 0.89, + 0.252 + ], + "angle": 0, + "content": "lixu.wang@ntu.edu.sg, {bingqishang2025, yili2023.1, payalmohapatra2026}@u.northwestern.edu, wei_dong@ntu.edu.sg, {wangxiao, qzhu}@northwestern.edu" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.287, + 0.327, + 0.303 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.319, + 0.485, + 0.773 + ], + "angle": 0, + "content": "Vision Transformers (ViTs), extensively pre-trained on large-scale datasets, have become fundamental to foundation models, enabling adaptation to diverse downstream tasks. Existing adaptation methods typically require direct data access, rendering them infeasible in privacy-sensitive domains where clients are often reluctant to share their data. A straightforward solution may be sending the pretrained ViT to clients for local adaptation, which poses issues of model intellectual property and incurs heavy client computation overhead. To address these issues, we propose a novel split adaptation (SA) method that enables effective downstream adaptation while protecting data and models. SA, inspired by split learning (SL), segments the pre-trained ViT into a frontend and a backend, with only the frontend shared with the client for data representation extraction. But unlike regular SL, SA replaces frontend parameters with low-bit quantized values, preventing direct exposure of the model. SA allows the client to add bi-level noise to the frontend and the extracted data representations, ensuring data protection. Accordingly, SA incorporates data-level and model-level out-of-distribution enhancements to mitigate noise injection's impact. Our SA focuses on the challenging few-shot adaptation and adopts patch retrieval augmentation for overfitting alleviation. Extensive experiments on multiple datasets validate SA's superiority over state-of-the-art methods and demonstrate its defense against advanced data reconstruction attacks while preventing model leakage with minimal computation cost on the client side. The source codes can be found at https://github.com/conditionWang/Split_Adaptation." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.804, + 0.222, + 0.82 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.83, + 0.484, + 0.876 + ], + "angle": 0, + "content": "In recent years, pre-trained vision transformers (ViTs) [13] have become mainstream foundational models in computer vision, delivering remarkable performance across a range of" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.288, + 0.906, + 0.38 + ], + "angle": 0, + "content": "visual tasks [13, 28, 50, 52]. ViTs have been equipped with general semantic knowledge during large-scale pre-training, which can be readily transferred to other domains including sensitive fields such as healthcare [41], finance [10], and manufacturing [8]. However, adapting pre-trained ViTs in these sensitive domains is extremely challenging." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.383, + 0.906, + 0.655 + ], + "angle": 0, + "content": "In real-world scenarios, model adaptation often follows a two-party framework, where the pre-trained model is hosted on a server, and the training data resides on a client. In non-sensitive domains, training data can be directly shared with the server. Real-world examples include users uploading data to access OpenAI's CLIP model API [40] or tune GPT-3.5 [3]. However, in sensitive domains, data is often tightly linked to personal privacy [12, 14], intellectual property (IP) [19, 47], and ethical concerns [15], making any form of public exposure, sharing, or transfer strictly prohibited. In this context, an alternative solution might involve sending the model to the client for local adaptation, as in federated learning [11, 27, 46]. However, large pre-trained models are expensive to obtain, making their owners unlikely to share or transfer them freely [47]. Moreover, given that these pre-trained models consist of billions of parameters, adaptation on the client side would incur substantial computational and storage costs." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.659, + 0.906, + 0.901 + ], + "angle": 0, + "content": "To our knowledge, there have been a few preliminary attempts to achieve downstream adaptation of pre-trained models while safeguarding data and models. Split learning (SL) [1, 44] divides the model into segments managed by different parties, which can partially protect the model from full exposure. Additionally, SL only communicates intermediate representations, avoiding the leakage of raw data. However, SL has critical drawbacks. On one hand, emerging malicious attacks, such as data reconstruction attacks (DRA) [16, 23, 38, 53, 54], can infer sensitive information from the intermediate representations. On the other hand, standard SL places the model training on the client side, imposing high computational costs. Beyond SL, a recent method called offsite tuning (OT) [49] has been proposed, which involves the server sending an emulator—a simplified version of the pre-trained model—to the client. The" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.887, + 0.225, + 0.901 + ], + "angle": 0, + "content": "*Equal contributions." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20092" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.096, + 0.089, + 0.48, + 0.251 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.253, + 0.483, + 0.309 + ], + "angle": 0, + "content": "Figure 1. The comparison between different downstream task adaptation approaches. Our SA can achieve effective few-shot task adaptation with minimal computation cost on the client side while protecting both the model and data." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.322, + 0.483, + 0.457 + ], + "angle": 0, + "content": "client can then fine-tune this emulator using its data, typically employing parameter-efficient techniques like adapter tuning [5]. After training, the client uploads these adapters back to the server, where they are integrated into the pretrained model. However, while OT protects data, the emulator is fully exposed, allowing adversaries to create high-performance models easily. Moreover, OT's adaptation depends on the substantial computational and storage demands for adapter tuning on the client side." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.46, + 0.483, + 0.793 + ], + "angle": 0, + "content": "To address the challenge of adapting pre-trained ViTs in sensitive domains, we propose a split adaptation (SA) method, which is shown in Figure 1. Like SL, SA divides the pre-trained ViT into two parts: a frontend and a backend. The server sends the frontend to the client, enabling data representation extraction. However, unlike SL, SA innovatively uses quantization to replace the frontend's parameters with low-bit values, preventing direct exposure of the original parameters. For data protection, the client injects random noise into the received frontend and its extracted data representations. Such injected noise unavoidably hurts the following adaptation, hence SA includes data-level and model-level out-of-distribution enhancements to improve the model generalization. SA is particularly designed for the challenging few-shot adaptation scenario, where it uses a patch retrieval augmentation technique to alleviate overfitting. Extensive experiments across four datasets demonstrate that SA significantly outperforms other state-of-the-art approaches in adaptation performance. Moreover, SA effectively defends against advanced DRAs and prevents model leakage. In particular, SA's effectiveness only relies on small computation cost on the client side. In summary:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.796, + 0.483, + 0.856 + ], + "angle": 0, + "content": "- We focus on solving a practical and important problem—adapting pre-trained ViTs for downstream tasks while protecting data and models. Solving this problem enables wider application of pre-trained models in sensitive fields." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.856, + 0.484, + 0.902 + ], + "angle": 0, + "content": "- We propose Split Adaptation (SA) that innovatively leverages model splitting and quantization to protect data and models during downstream task adaptation. SA focuses" + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.796, + 0.484, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.527, + 0.092, + 0.905, + 0.122 + ], + "angle": 0, + "content": "on challenging few-shot scenarios, in which SA uses a patch retrieval augmentation to combat overfitting." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.122, + 0.905, + 0.183 + ], + "angle": 0, + "content": "- We conduct extensive experiments across multiple datasets and state-of-the-art baseline approaches, with results validating SA's effectiveness in achieving superior adaptation while ensuring data and model protection." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.198, + 0.664, + 0.214 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.224, + 0.663, + 0.241 + ], + "angle": 0, + "content": "2.1. Split Learning" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.247, + 0.907, + 0.55 + ], + "angle": 0, + "content": "Split learning (SL) [44] is a collaborative learning approach widely used in the multi-party computation framework. SL operates by dividing the machine learning (ML) model into segments, with each managed by a different party. This approach enhances privacy by transmitting only intermediate data representations rather than raw data. Many practical examples of SL in training and adapting ML models across various domains emerge, including healthcare [1, 44], financial [1, 42], and edge computing [17, 18]. Despite these successes, more adversarial attacks illustrate the vulnerability of privacy leakage in SL, with the severest one – data reconstruction attack (DRA) [23]. DRA aims to recover the raw data from its corresponding model representations. In the context of SL, a variety of DRAs have been witnessed, among them, some are based on optimization to imitate the ground-truth representations (e.g., FSHA [38] and PCAT [16]), while others leverage generative models to recover the raw data by exploiting the intrinsic semantics (e.g., Ginver [54] and FORA [53]). Therefore, new SL approaches are needed to help protect data privacy better." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.561, + 0.794, + 0.577 + ], + "angle": 0, + "content": "2.2. Vision Transformer Adaptation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.584, + 0.907, + 0.901 + ], + "angle": 0, + "content": "The vision transformers (ViT) [13] are usually pre-trained with large-scale data to learn general semantics and knowledge. Representative pre-training approaches include supervised [28] and self-supervised learning like contrastive learning [4, 6, 21, 40] and auxiliary task training [22, 37, 57]. Owing to such extensive pre-training, adapting pretrained ViT for distinct downstream tasks has become a mainstream paradigm. The most regular method is to tune the entire model, which consumes substantial computational and memory resources. Another widely used approach is linear probing [29], which only trains a task module while keeping the ViT frozen, but the prominent shortcoming is the inferior performance. Recently, more parameter-efficient approaches have been proposed, like adapter tuning [5], prompt tuning [26], and prefix tuning [51]. However, these approaches still require direct access to the task data. In contrast to the above approaches, a new method called offsite tuning [49] can work effectively without direct data access. However, extensive training happens on the data owner side, and there is a high risk of model disclosure, making it impractical in real-world scenarios." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "20093" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.09, + 0.226, + 0.108 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.115, + 0.232, + 0.13 + ], + "angle": 0, + "content": "3.1. Preliminaries" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.138, + 0.274, + 0.152 + ], + "angle": 0, + "content": "3.1.1. Vision Transformer" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.157, + 0.484, + 0.388 + ], + "angle": 0, + "content": "The vision transformer (ViT) [13] is built upon sequential attention encoder layers. Specifically, each encoder layer consists of a multi-head self-attention (MSA) block and a Multi-Layer Perceptron (MLP) block. Layernorm (LN) is applied before every block and residual connections are placed after every block. The MLP contains two linear layers with an intermediate GELU activation. To handle 2D images, ViT reshapes the image \\(\\mathbf{I} \\in \\mathbb{R}^{H \\times W \\times C}\\) into \\(N\\) flattened patches \\(I^p \\in \\mathbb{R}^{P^2 \\cdot C}\\), where \\(H\\), \\(W\\), and \\(C\\) are the image height, width, and channels, respectively, and \\(P\\) is the pixel resolution of each patch. For the first layer of ViT, to unify the vector dimensions, a trainable linear projection layer with neuron weights \\([\\mathbf{W}_1^E, \\dots, \\mathbf{W}_N^E]\\) is used to map each vectorized patch into dimension \\(d\\). In this case, the input to the first transformer layer is:" + }, + { + "type": "equation", + "bbox": [ + 0.133, + 0.397, + 0.484, + 0.426 + ], + "angle": 0, + "content": "\\[\n\\mathbf {X} _ {1} = \\left[ \\mathbf {x} ^ {\\mathrm {C L S}}; I _ {1} ^ {p} \\mathbf {W} _ {1} ^ {E}; \\dots ; I _ {N} ^ {p} \\mathbf {W} _ {N} ^ {E} \\right] + \\mathbf {E} ^ {\\mathrm {P O S}}, \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.135, + 0.421, + 0.44, + 0.438 + ], + "angle": 0, + "content": "\\[\nw h e r e \\mathbf {W} ^ {E} \\in \\mathbb {R} ^ {(P ^ {2} \\cdot C) \\times d}, \\mathbf {E} ^ {\\mathrm {P O S}} \\in \\mathbb {R} ^ {(N + 1) \\times d},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.451, + 0.484, + 0.56 + ], + "angle": 0, + "content": "where \\(\\mathbf{x}^{\\mathrm{CLS}}\\in \\mathbb{R}^{d}\\) is a classification token, and \\(\\mathbf{E}^{\\mathrm{POS}}\\) is the position encoding. Then for each of the following layers, the input sequence \\(\\mathbf{X}_l\\) is fed into each head of the MSA module to obtain three linearly projected matrices: Query \\(\\mathbf{Q}_l = \\mathbf{X}_l\\mathbf{W}_l^Q\\), Key \\(\\mathbf{K}_l = \\mathbf{X}_l\\mathbf{W}_l^K\\), and Value \\(\\mathbf{V}_l = \\mathbf{X}_l\\mathbf{W}_l^V\\). Then the Softmax function is applied to normalize the attention score \\(\\mathbf{A}_l = \\mathbf{Q}_l\\mathbf{K}_l^\\top\\) and the output \\(\\mathbf{Z}_l\\) is:" + }, + { + "type": "equation", + "bbox": [ + 0.15, + 0.572, + 0.484, + 0.607 + ], + "angle": 0, + "content": "\\[\n\\mathbf {Z} _ {l} = \\operatorname {M S A} \\left(\\mathbf {X} _ {l}\\right) = \\operatorname {S o f t m a x} \\left(\\frac {\\mathbf {A} _ {l}}{\\sqrt {d}}\\right) \\mathbf {V} _ {l}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.616, + 0.483, + 0.648 + ], + "angle": 0, + "content": "Suppose the two linear layers in MLP are parameterized by \\(\\mathbf{W}_{l,1},\\mathbf{b}_{l,1}\\) and \\(\\mathbf{W}_{l,2},\\mathbf{b}_{l,2}\\), the output is then computed as:" + }, + { + "type": "equation", + "bbox": [ + 0.113, + 0.659, + 0.483, + 0.677 + ], + "angle": 0, + "content": "\\[\n\\operatorname {M L P} \\left(\\mathbf {Z} _ {l}\\right) = \\operatorname {G E L U} \\left(\\mathbf {Z} _ {l} \\mathbf {W} _ {l, 1} + \\boldsymbol {b} _ {l, 1}\\right) \\mathbf {W} _ {l, 2} + \\boldsymbol {b} _ {l, 2}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.484, + 0.751 + ], + "angle": 0, + "content": "Finally, both the MSA and MLP modules are linked via residual connections, thus we have the final output of each transformer layer: \\(\\mathbf{X}_{l + 1} = \\mathrm{LN}(\\mathbf{Z}_l + \\mathrm{MLP}(\\mathbf{Z}_l))\\in\\) \\(\\mathbb{R}^{(N + 1)\\times d}\\) , where \\(\\mathbf{Z}_l = \\mathrm{LN}(\\mathbf{X}_l + \\mathrm{MSA}(\\mathbf{X}_l))\\in \\mathbb{R}^{(N + 1)\\times d}\\)" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.759, + 0.289, + 0.773 + ], + "angle": 0, + "content": "3.1.2. Problem Formulation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.779, + 0.484, + 0.904 + ], + "angle": 0, + "content": "When adapting pre-trained ViT for certain downstream tasks, we consider a representative client-server framework in which the client queries the server for the adaptation request. The client owns the task adaptation dataset \\(\\mathcal{D}^{\\mathrm{C}} = \\{(x_i,y_i)\\}_{i = 1}^{N^{\\mathrm{C}}} (N^{\\mathrm{C}}\\) is the sample number, in practice, we suppose \\(N^{\\mathrm{C}}\\) is small even few-shot) with each sample drawn from a marginal input distribution \\(\\pmb{x}_i\\sim \\mathcal{P}_X^{\\mathrm{C}}\\) and labeled with a marginal task distribution \\(\\mathcal{P}_{\\mathcal{Y}}^{\\mathrm{C}}\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.248 + ], + "angle": 0, + "content": "On the server side, the pre-trained ViT \\(F_{\\Theta}\\) parameterized on \\(\\Theta\\) is assumed to consist of \\(L^F\\) attention layers, and each layer is parameterized on \\(\\theta\\), i.e., \\(F := [f_{\\theta_1}, \\dots, f_{\\theta_{L^F}}]\\). Besides, we assume the server owns another dataset \\(\\mathcal{D}^{\\mathrm{S}} = \\{(x_i, y_i)\\}_{i=1}^{N^{\\mathrm{S}}}\\) collected from public domains with the marginal input and task distributions, \\(\\mathcal{P}_{\\mathcal{X}}^{\\mathrm{S}}\\) and \\(\\mathcal{P}_{\\mathcal{Y}}^{\\mathrm{S}}\\), respectively. In real-world scenarios, the datasets owned by the client and the server are distinct but relevant, i.e., \\(\\mathcal{P}_{\\mathcal{X}}^{\\mathrm{C}} \\neq \\mathcal{P}_{\\mathcal{X}}^{\\mathrm{S}}, \\mathcal{P}_{\\mathcal{Y}}^{\\mathrm{C}} \\neq \\mathcal{P}_{\\mathcal{Y}}^{\\mathrm{S}}\\) but \\(I(\\mathcal{X}^{\\mathrm{C}}, \\mathcal{X}^{\\mathrm{S}}) > 0\\), \\(I(\\mathcal{Y}^{\\mathrm{C}}, \\mathcal{Y}^{\\mathrm{S}}) > 0\\) where \\(I(\\cdot; \\cdot)\\) calculates the mutual information." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.254, + 0.907, + 0.362 + ], + "angle": 0, + "content": "Objective of Task Performance. Pre-trained ViT adaptation aims to obtain a downstream task module \\( g \\) parameterized on \\( \\gamma \\) that can achieve as good performance as possible on the client data distribution when combined with \\( F_{\\Theta} \\). Certainly, such adaptation also allows \\( F_{\\Theta} \\) tuning if it can enable performance improvement on the downstream task. In this case, the objective is formulated as," + }, + { + "type": "equation", + "bbox": [ + 0.558, + 0.377, + 0.907, + 0.4 + ], + "angle": 0, + "content": "\\[\n\\Theta , \\gamma = \\arg \\min _ {\\Theta , \\gamma} \\mathbb {E} _ {\\boldsymbol {x} \\sim \\mathcal {P} _ {\\chi} ^ {\\mathrm {C}}, y \\sim \\mathcal {P} _ {\\mathcal {Y}} ^ {\\mathrm {C}}} \\mathcal {L} (\\boldsymbol {x}, y; \\Theta \\circ \\gamma), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.415, + 0.906, + 0.447 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}(\\cdot ;\\cdot)\\) is a loss function that measures the model performance regarding certain data distribution." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.453, + 0.907, + 0.605 + ], + "angle": 0, + "content": "Objective of Privacy Protection. In addition to the above standard objective, the intrinsic privacy concerns of the server-client framework should be addressed, i.e., the pretrained model and client data should be protected during adaptation when supposing the server and the client may be semi-honest. On the one hand, the adaptation should prohibit the client from stealing surrogate models with good performance or obtaining them at a small cost, on the other hand, any client data reconstruction on the server side should be avoided." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.62, + 0.885, + 0.637 + ], + "angle": 0, + "content": "3.2. Overview of The Proposed Split Adaptation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.645, + 0.907, + 0.903 + ], + "angle": 0, + "content": "To achieve effective adaptation while protecting the pre-trained ViT and client data, we propose a method called Split Adaptation (SA), shown in Figure 2. Our SA starts with dividing the pre-trained ViT \\( F_{\\Theta} \\) into two parts: a frontend \\( F^{\\mathrm{F}} \\) and a backend \\( F^{\\mathrm{B}} \\). The original \\( F^{\\mathrm{F}} \\) is not sent to the client, instead, we leverage model quantization to send a quantized version \\( \\widehat{F}^{\\mathrm{F}} \\) to the client. To mitigate the influence of frontend quantization, SA includes out-of-distribution (OOD) augmentations from both data and model levels to enhance the cross-distribution generalization of the frontend and backend. As for the client side, it adds random noise to the received \\( \\widehat{F}^{\\mathrm{F}} \\) first and adopts a patch-level retrieve-then-replace strategy to generate more client data representations that help alleviate overfitting in few-shot adaptation. Before being uploaded to the server for the final adaptation, these augmented representations are protected by adding noise again." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "20094" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.115, + 0.092, + 0.885, + 0.384 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.386, + 0.908, + 0.457 + ], + "angle": 0, + "content": "Figure 2. Overview of Split Adaptation (SA) for pre-trained ViT adaptation. SA divides the pre-trained ViT into a frontend and a backend. After applying Out-of-distribution Enhanced Quantization to the frontend, its quantized version is sent to the client. To mitigate the impact of quantization, SA adopts Out-of-distribution Quantization-aware Tuning to enhance backend's generalization. As for the client, it injects random noise to the received frontend and retrieves then replaces randomly selected patches to augment more client data representations, which are sent to the server for the final adaptation after being added with noise again." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.469, + 0.398, + 0.485 + ], + "angle": 0, + "content": "3.3. Model Splitting on The Server Side" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.492, + 0.484, + 0.795 + ], + "angle": 0, + "content": "At the beginning of SA, the pre-trained ViT is divided into a frontend \\( F^{\\mathrm{F}} \\coloneqq [f_{\\theta_1}, \\dots, f_{\\theta_K}] \\) and a backend \\( F^{\\mathrm{B}} \\coloneqq [f_{\\theta_{K+1}}, \\dots, f_{\\theta_{LF}}] \\) at the \\( K \\)-th layer (\\( K > L^{F} / 2 \\)). Then the frontend is distributed to the client for representation extraction. Although we empirically found that the frontend cannot perform well (Section 4.3), we still hope to protect it from disclosing the original model parameters. Quantization [20, 34], as a representative model compression approach, can significantly reduce the computational and memory cost without sacrificing model performance. As quantization works by representing model weights and activations with low-precision values, it can potentially protect the model. However, standard quantization needs a certain amount of data drawn from the target task, which is inaccessible for the server in our problem. In this case, the left choice is to leverage the server dataset \\( \\mathcal{D}^{\\mathrm{S}} \\) to quantize the model, but a dedicated design for enhancing the quantized model's cross-distribution generalization is needed. Therefore, we propose an augmentation technique to generate OOD data in terms of \\( \\mathcal{D}^{\\mathrm{S}} \\) for quantization." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.806, + 0.442, + 0.822 + ], + "angle": 0, + "content": "3.3.1. Out-of-distribution Enhanced Quantization" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.484, + 0.903 + ], + "angle": 0, + "content": "In practice, we have to consider extreme cases in which the client dataset differs significantly from the server dataset and they may only share general semantics. Hence our augmentation technique should generate OOD data with a large discrepancy to \\(\\mathcal{D}^{\\mathrm{S}}\\). Inspired by signal processing, we no" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.47, + 0.903, + 0.655 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.655, + 0.907, + 0.724 + ], + "angle": 0, + "content": "Figure 3. Visualization of our Hilbert Transform data augmentation. Compared to amplitude-exchange augmentation [45] (bottom row), our augmentation method (top row) generates data that diverges further from the original in appearance while preserving the original semantics (dog category)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.907, + 0.903 + ], + "angle": 0, + "content": "tice that there is a technique called Hilbert Transform (HT), which only captures the most sharp edge in images. Intuitively, the sharp edges contain the most general semantics, making HT suitable for our problem. Applying HT to a signal in standard signal processing results in a phase shift of \\(\\pi /2\\) in the frequency domain [36]. In the visual domain, we can implement HT by transforming an image \\(\\mathbf{I}\\) to the frequency domain with the spectrum index \\(\\xi\\) via Fourier transform, then the negative \\(\\xi\\) needs to multiply with the imaginary unit \\(i\\), while the positive \\(\\xi\\) needs to multiple with" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20095" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.092, + 0.093, + 0.146, + 0.106 + ], + "angle": 0, + "content": "\\(-i\\) , i.e.," + }, + { + "type": "equation", + "bbox": [ + 0.172, + 0.114, + 0.483, + 0.133 + ], + "angle": 0, + "content": "\\[\n\\mathrm {H T} (\\mathbf {I}) = \\mathcal {F} ^ {- 1} \\{- i \\cdot \\operatorname {s g n} (\\xi) \\mathcal {F} (\\mathbf {I}) \\}, \\qquad (5)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.138, + 0.484, + 0.26 + ], + "angle": 0, + "content": "where \\(\\mathcal{F}(\\cdot)\\) and \\(\\mathcal{F}^{-1}(\\cdot)\\) are Fourier transform and inverse, respectively, and \\(\\mathrm{sgn}(\\cdot)\\) is a sign function. However, pure HT only provides new images with black backgrounds. To diversify the HT-generated image, we replace its amplitude in the frequency domain with that of the original image. In this way, the generated image is much better than the standard amplitude-exchange augmentation [32, 45] (please refer to Figure 3 for a visual comparison)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.261, + 0.484, + 0.321 + ], + "angle": 0, + "content": "To achieve effective quantization, we adopt the most popular uniform symmetric quantization [20] to implement SA, in which a floating-point value \\( w \\) is projected into a \\( k \\)-bit integer value \\( w_{\\mathrm{q}} \\) with a scaling factor \\( \\Delta \\):" + }, + { + "type": "equation", + "bbox": [ + 0.097, + 0.328, + 0.484, + 0.366 + ], + "angle": 0, + "content": "\\[\n\\widehat {w} = \\Psi_ {k} (w, \\Delta) = \\operatorname {c l a m p} (\\operatorname {r o u n d} \\left(\\frac {w}{\\Delta}\\right), - 2 ^ {k - 1}, 2 ^ {k - 1} - 1), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.366, + 0.484, + 0.627 + ], + "angle": 0, + "content": "where \\(\\mathrm{round}(\\cdot)\\) turns a value into an integer and \\(\\mathrm{clamp}(\\cdot, \\min, \\max)\\) constrains the output in the range from \\(\\min\\) to \\(\\max\\). Intuitively, the scaling factor \\(\\Delta\\) directly impacts the model performance after quantization. Previous studies [7, 34] have tried a variety of measurements to quantify the model output changes and accordingly determine the optimal scaling factors. Among these studies, we follow Yuan et al. [55] to choose a Hessian-guided metric for scaling factor determination. Suppose the pre-trained ViT \\(F_{\\Theta}\\) is equipped with a task module \\(g_{\\gamma'}'\\) that is tuned on the server dataset \\(\\mathcal{D}^{\\mathrm{S}}\\), the model performance can be measured by a loss \\(\\mathbb{E}_{\\boldsymbol{x} \\sim \\mathcal{P}_{\\boldsymbol{x}}^{\\mathrm{S}}, y \\sim \\mathcal{P}_{y}^{\\mathrm{S}}} \\mathcal{L}(\\boldsymbol{x}, y; \\Theta \\circ \\gamma')\\) (we omit \\(\\boldsymbol{x}\\) and \\(y\\) to simplify the used symbols in the following descriptions). Then if we regard that the quantization brings a small perturbation \\(\\epsilon\\) to the model parameters, i.e., \\(\\hat{\\Theta} = \\Theta + \\epsilon\\), we can analyze the influence of quantization on the task loss using Taylor series expansion:" + }, + { + "type": "equation", + "bbox": [ + 0.104, + 0.633, + 0.483, + 0.663 + ], + "angle": 0, + "content": "\\[\n\\mathbb {E} [ \\mathcal {L} (\\widehat {\\Theta} \\circ \\gamma^ {\\prime}) ] - \\mathbb {E} [ \\mathcal {L} (\\Theta \\circ \\gamma^ {\\prime}) ] \\approx \\epsilon^ {\\top} \\nabla_ {\\Theta} + \\frac {1}{2} \\epsilon^ {\\top} \\mathbf {H} _ {\\Theta} \\epsilon , (7)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.667, + 0.484, + 0.804 + ], + "angle": 0, + "content": "where \\(\\nabla_{\\Theta}\\) is the gradients and \\(\\mathbf{H}_{\\Theta}\\) is the Hessian matrix. Then the optimal scaling factors are the ones corresponding to minimal influence caused by quantization, i.e., \\(\\Delta = \\arg \\min_{\\Delta}(\\mathbb{E}[\\mathcal{L}(\\widehat{\\Theta}\\circ \\gamma^{\\prime})] - \\mathbb{E}[\\mathcal{L}(\\Theta \\circ \\gamma^{\\prime})])\\). Certainly, determining and adopting a general scaling factor for the entire model is too coarse to perform well, thus we apply a layer-wise reconstruction method [31] to achieve fine-grained quantization, then the searching optimization of Eq. (7) can be approximated by:" + }, + { + "type": "equation", + "bbox": [ + 0.09, + 0.809, + 0.489, + 0.854 + ], + "angle": 0, + "content": "\\[\n\\underset {\\Delta_ {l - 1}} {\\arg \\min } \\mathbb {E} \\left[ \\left(\\widehat {\\mathbf {X}} _ {l} - \\mathbf {X} _ {l}\\right) ^ {\\top} \\operatorname {d i a g} \\left(\\frac {\\partial \\mathcal {L}}{\\partial \\mathbf {X} _ {l}} \\odot \\frac {\\partial \\mathcal {L}}{\\partial \\mathbf {X} _ {l}}\\right) \\left(\\widehat {\\mathbf {X}} _ {l} - \\mathbf {X} _ {l}\\right) \\right], \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.855, + 0.484, + 0.901 + ], + "angle": 0, + "content": "where \\(\\mathbf{X}_l\\) and \\(\\widehat{\\mathbf{X}}_l\\) are the outputs of the \\((l-1)\\)-th layer before and after quantization, respectively, and \\(\\odot\\) represents the Hadamard product of a matrix." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.214 + ], + "angle": 0, + "content": "With the augmented neighborhood data, the frontend can be quantized with improved cross-distribution generalization. Specifically, we first conduct HT-based data augmentation on the server dataset \\(\\mathcal{D}^{\\mathrm{S}}\\) to obtain \\(\\mathrm{HT}(\\mathcal{D}^{\\mathrm{S}})\\). Then the optimal scaling factor for each layer in the frontend \\(F^{\\mathrm{F}}\\) is empirically determined by optimizing Eq. (8) on the merged dataset \\(\\mathcal{D}_{\\cup}^{\\mathrm{S}} = \\mathcal{D}^{\\mathrm{S}} \\cup \\mathrm{HT}(\\mathcal{D}^{\\mathrm{S}})\\) and we can obtain a quantized frontend \\(\\widehat{F}^{\\mathrm{F}}\\) in the end." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.22, + 0.892, + 0.236 + ], + "angle": 0, + "content": "3.3.2. Out-of-distribution Quantization-aware Tuning" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.239, + 0.906, + 0.344 + ], + "angle": 0, + "content": "To achieve better adaptation performance, the backend also needs to resist the unavoidable parameter difference between the original and quantized model as well as the noise injected by the client that we will introduce in Section 3.4. As a result, we propose an OOD quantization-aware tuning technique to enhance the backend's resistance when encountering OOD data and frontends." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.345, + 0.907, + 0.496 + ], + "angle": 0, + "content": "To simulate potential OOD frontends, we first divide the merged server dataset \\(\\mathcal{D}_{\\cup}^{\\mathrm{S}}\\) into \\(M\\) subsets and denote these subsets as \\(\\{\\mathcal{D}_{\\cup_m}^{\\mathrm{S}}\\}_{m = 1}^M\\). Next, each subset \\(\\mathcal{D}_{\\cup_m}^{\\mathrm{S}}\\) is used to quantize the original frontend \\(F^{\\mathrm{F}}\\) to obtain a quantized version \\(\\widehat{F}_m^{\\mathrm{F}}\\). Then let us use a single quantized frontend \\(\\widehat{F}_m^{\\mathrm{F}}\\) to introduce how to tune the backend. To begin with, since the backend is represented and the tuning is performed using floating-point values, the data representations extracted by the quantized frontend should be transformed back to floating-point values through the dequantization:" + }, + { + "type": "equation", + "bbox": [ + 0.62, + 0.503, + 0.907, + 0.521 + ], + "angle": 0, + "content": "\\[\n\\tilde {w} = \\Psi^ {- 1} (\\hat {w}, \\Delta) = \\hat {w} \\cdot \\Delta . \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.529, + 0.907, + 0.775 + ], + "angle": 0, + "content": "In this way, we can obtain the floating-point representation \\(\\widetilde{\\mathbf{X}}_K\\) by dequantizing the \\(K\\)-th layer's output with the scaling factor \\(\\Delta_K\\). Note that \\(\\widetilde{\\mathbf{X}}_K\\) is different from the original representation \\(\\mathbf{X}_K\\) extracted by the original frontend \\(F^{\\mathrm{F}}\\) due to the range constraints of round and clamp functions in Eq. (6). Our following design aims to tune the backend to resist such differences. The most straightforward method is to forward the floating-point representation \\(\\widetilde{\\mathbf{X}}_K\\) to the backend and conduct tuning using task loss. However, such tuning will make the backend over-fitted to the server dataset \\(\\mathcal{D}^{\\mathrm{S}}\\). We solve this by a representation-mixup strategy. Specifically, for a particular data sample \\(\\pmb{x}\\), we feed it to a quantized frontend \\(\\widehat{F}_{m}^{\\mathrm{F}}\\) and the original frontend \\(F^{\\mathrm{F}}\\) to extract representations \\(\\widehat{\\mathbf{X}}_{K,m},\\mathbf{X}_{K,m}\\). After dequantizing \\(\\widehat{\\mathbf{X}}_{K,m}\\) into the floating-point form \\(\\widetilde{\\mathbf{X}}_{K,m}\\), we obtain a mixup representation:" + }, + { + "type": "equation", + "bbox": [ + 0.594, + 0.781, + 0.905, + 0.8 + ], + "angle": 0, + "content": "\\[\n\\mathbf {X} _ {K, m} ^ {*} = \\lambda \\widetilde {\\mathbf {X}} _ {K, m} + (1 - \\lambda) \\mathbf {X} _ {K, m}, \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.807, + 0.906, + 0.852 + ], + "angle": 0, + "content": "where \\(\\lambda\\) is drawn from a widely-used beta distribution \\(\\mathrm{Beta}(\\lambda;0.75,0.75)\\). Then considering all available quantized frontend, we can tune the backend with the loss:" + }, + { + "type": "equation", + "bbox": [ + 0.577, + 0.859, + 0.907, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {t u n e}} = \\sum_ {\\mathcal {D} _ {\\cup} ^ {\\mathrm {S}}} \\sum_ {m = 0} ^ {M} \\mathcal {L} \\left(\\mathbf {X} _ {K, m} ^ {*}, y; \\Theta^ {\\mathrm {B}} \\circ \\gamma^ {\\prime}\\right). \\tag {11}\n\\]" + }, + { + "type": "footer", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "20096" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.483, + 0.127 + ], + "angle": 0, + "content": "Note that \\(\\mathbf{X}_{K,0}^{*}\\) originates from feeding \\(\\pmb{x}\\) to the frontend \\(\\widehat{F}^{\\mathrm{F}}\\) quantized on the full set of \\(\\mathcal{D}_{\\cup}^{\\mathrm{S}}\\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.137, + 0.483, + 0.154 + ], + "angle": 0, + "content": "3.4. Client Bi-level Noisy Representation Extraction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.16, + 0.483, + 0.267 + ], + "angle": 0, + "content": "In our SA model, the quantized frontend \\(\\widehat{F}^{\\mathrm{F}}\\) will be sent to the client after the OOD enhanced quantization (Section 3.3.1), and then the client can forward its data to \\(\\widehat{F}^{\\mathrm{F}}\\) to extract representations and send them to the server. However, there is a risk of data reconstruction when sharing representations, thus we propose a bi-level noisy representation extraction mechanism to defend against such attacks." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.268, + 0.484, + 0.405 + ], + "angle": 0, + "content": "Intuitively speaking, the key to data reconstruction defense is to use the model unknown by the server to extract the representations. Therefore, we allow the client to inject a certain degree of random noise into the received frontend \\(\\widehat{F}^{\\mathrm{F}}\\). For each layer in the frontend, we inject a multiplicative noise \\(n_l^\\times\\), and an additive noise \\(n_l^+\\), which are sampled from two Gaussian distributions, i.e., \\(n_l^\\times \\sim \\mathcal{P}_{\\mathcal{N}(1,\\alpha \\theta_l)}\\) and \\(n_l^+ \\sim \\mathcal{P}_{\\mathcal{N}(0,\\alpha \\theta_l)}\\) (\\(\\alpha = 0.01\\) controls the noise degree), to the model parameters:" + }, + { + "type": "equation", + "bbox": [ + 0.221, + 0.418, + 0.483, + 0.436 + ], + "angle": 0, + "content": "\\[\n\\theta_ {l} ^ {\\prime} = n _ {l} ^ {\\times} \\cdot \\theta_ {l} + n _ {l} ^ {+}. \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.448, + 0.484, + 0.557 + ], + "angle": 0, + "content": "Next, the client feeds its dataset \\(\\mathcal{D}^{\\mathrm{C}}\\) to the noisy frontend \\(\\widehat{F}^{\\mathrm{F}'}\\) to extract representations \\(\\{\\widehat{\\mathbf{X}}_{K,i}^{\\mathrm{C}}\\}_{i=1}^{N^{\\mathrm{C}}}\\). However, it is still insecure to share these representations. We found that our model-level noise can amplify the defense effect of the regular noise obfuscation [24] at the representation level, which allows to add a smaller degree of Laplace noise \\((\\mathcal{P}_{\\mathrm{Laplace}(0,0.8)})\\) to representations for better protection." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.567, + 0.478, + 0.584 + ], + "angle": 0, + "content": "3.5. Patch-retrieval-augmented Model Adaptation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.59, + 0.483, + 0.68 + ], + "angle": 0, + "content": "In few-shot adaptation scenarios, like the backend tuning on the server (Section 3.3.2), the adaptation also faces overfitting challenges, but the overfitting here is due to the limited client data available. To tackle such challenges, we design a patch-level retrieval augmentation approach to generate more client data representations." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.682, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Our patch-level retrieval augmentation is inspired by the intuition that different images may contain similar patches. As introduced in Section 3.1.1, the data representations of each layer in a ViT hold the dimension \\((N + 1)\\times d\\), which can be denoted as \\(\\widehat{\\mathbf{X}}_{K,i}^{\\mathrm{C}} = [\\mathbf{x}_i^{\\mathrm{CLS}};\\mathbf{x}_{i,1},\\dots ,\\mathbf{x}_{i,N}]\\) (we omit some symbols for simplification). Then we can form \\(N\\) retrieval sets from all client data representations (before adding Laplace noise) and each corresponds to a patch, i.e., \\(\\{S_j\\}_{j = 1}^N\\) where \\(S_{j} = \\{\\mathbf{x}_{i,j}\\}_{i = 1}^{N^{\\mathrm{C}}}\\). With these sets, for each augmentation of \\(\\widehat{\\mathbf{X}}_{K,i}^{\\mathrm{C}}\\), we randomly select \\(N^{\\mathrm{P}} = 50\\) patches (please see the sensitivity analysis in Supplementary Materials) \\(\\{\\mathbf{x}_{i,\\mathrm{P}_j}\\}_{j = 1}^{N^{\\mathrm{P}}}\\) (subscript \\(\\mathrm{P}_j\\) meets \\(1\\leq \\mathrm{P}_j\\leq N\\)) and delete them from their corresponding retrieval sets. Next, each \\(\\mathbf{x}_{i,\\mathrm{P}_j}\\) retrieves the most similar patch \\(\\Omega (\\mathbf{x}_{i,\\mathrm{P}_j})\\) from" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.092, + 0.716, + 0.107 + ], + "angle": 0, + "content": "its corresponding retrieval set:" + }, + { + "type": "equation", + "bbox": [ + 0.598, + 0.116, + 0.905, + 0.15 + ], + "angle": 0, + "content": "\\[\n\\Omega \\left(\\mathbf {x} _ {i, \\mathrm {P} _ {j}}\\right) = \\arg \\max _ {\\mathbf {x} \\in S _ {\\mathrm {P} _ {j}} ^ {\\prime}} \\frac {\\mathbf {x} _ {i , \\mathrm {P} _ {j}} \\cdot \\mathbf {x}}{\\left| \\mathbf {x} _ {i , \\mathrm {P} _ {j}} \\right| \\left| \\mathbf {x} \\right|}, \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.16, + 0.907, + 0.395 + ], + "angle": 0, + "content": "where \\(\\mathcal{S}_{\\mathrm{P}_j}^{\\prime} = \\mathcal{S}_{\\mathrm{P}_j}\\backslash \\mathbf{x}_{i,\\mathrm{P}_j}\\). After the retrieval operations for all selected patches \\(\\{\\mathbf{x}_{i,\\mathrm{P}_j}\\}_{j = 1}^{N^{\\mathrm{P}}}\\), we replace them with the retrieved patches \\(\\{\\Omega (\\mathbf{x}_{i,\\mathrm{P}_j})\\}_{j = 1}^{N^{\\mathrm{P}}}\\) in the original representation \\(\\widehat{\\mathbf{X}}_{K,i}^{\\mathrm{C}}\\) to form a new one \\(\\widehat{\\mathbf{X}}_{K,i}^{\\prime \\mathrm{C}}\\). Note that the selected patches are determined randomly, resulting in newly generated representations that differ across augmentation runs. To generate sufficient augmented representations, we run the augmentation \\(N^{\\mathrm{Aug}}\\) times for all client data representations and produce an augmented representation set, i.e., \\(\\widehat{\\mathcal{X}}_K^{\\mathrm{C}} = \\{\\{\\widehat{\\mathbf{X}}_{K,i}^{\\mathrm{C}_j}\\}_{j = 0}^{N^{\\mathrm{Aug}}}\\}_{i = 1}^{N^{\\mathrm{C}}}\\) (we denote the original representation as \\(\\widehat{\\mathbf{X}}_{K,i}^{\\mathrm{C}_0}\\) and the augmented representations as \\(\\widehat{\\mathbf{X}}_{K,i}^{\\mathrm{C}_{1\\sim N^{\\mathrm{Aug}}}}\\)). The above patch retrieval augmentation is run by the client, and it adds Laplace noise (Section 3.4) to all representations in \\(\\widehat{\\mathcal{X}}_K^{\\mathrm{C}}\\) before uploading to the server." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.396, + 0.906, + 0.532 + ], + "angle": 0, + "content": "After receiving the representations, the server dequantizes them into floating-point form with the \\(K\\)-th layer's scaling factor \\(\\Delta_K\\) by Eq. (9) and launches the adaptation. To further alleviate the overfitting due to limited client data, different from Split Learning, we only allow high-level feature learning in the backend to prevent low-level features from becoming biased. Then the server randomly initializes a task module \\(g_{\\gamma}\\) and tunes it with the backend \\(F^{\\mathrm{B}}\\) (that has been tuned in Section 3.3.2) by following the loss:" + }, + { + "type": "equation", + "bbox": [ + 0.577, + 0.541, + 0.905, + 0.587 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {a d a p t}} = \\sum_ {i = 1} ^ {N ^ {\\mathrm {C}}} \\sum_ {j = 0} ^ {N ^ {\\mathrm {A u g}}} \\mathcal {L} \\left(\\widehat {\\mathbf {X}} _ {K, i} ^ {\\mathrm {C} _ {j}}, y; \\Theta^ {\\mathrm {B}} \\circ \\gamma\\right). \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.596, + 0.906, + 0.643 + ], + "angle": 0, + "content": "To protect the client's labels, the server sends the task module output to the client, allowing the task loss \\(\\mathcal{L}_{\\mathrm{adapt}}\\) to be computed on the client side and then sent back to the server." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.655, + 0.646, + 0.672 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.68, + 0.721, + 0.696 + ], + "angle": 0, + "content": "4.1. Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.701, + 0.907, + 0.822 + ], + "angle": 0, + "content": "Datasets. ImageNet [9] is the most widely used image dataset collected from public domains, thus we use its test set as the server dataset. For the client datasets, we use CIFAR-100 [30], Places365 [35], and DomainNet [39]. Places365 is a scene recognition dataset that contains 365 classes. DomainNet (D.Net) is a challenging multi-domain dataset, which includes six domains. We select the smallest domain—Clipart (Cl) for the few-shot adaptation scenarios." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Implementation Details. We conduct experiments on three pre-trained model architectures: ViT-Large [13], DeiT-Base [43], and Swin-L [33] (DeiT-Base and Swin-L's experiments are reported in Supplementary Materials). For the Places365 dataset, the target task module is composed" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20097" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.09, + 0.907, + 0.117 + ], + "angle": 0, + "content": "Table 1. Performance comparison between our SA and other baseline approaches in 3-shot, 5-shot, and 10-shot adaptation scenarios over 3 datasets. The SA can substantially exceed other methods. We bold and blue the best, and bold the second best." + }, + { + "type": "table", + "bbox": [ + 0.11, + 0.119, + 0.88, + 0.248 + ], + "angle": 0, + "content": "
Few-shot Setup3-shot5-shot10-shot
MethodCIFAR-100Places365D.Net-ClCIFAR-100Places365D.Net-ClCIFAR-100Places365D.Net-Cl
Linear Probing74.05±1.2126.18±1.0057.26±1.1178.79±0.6534.25±1.4563.84±0.6683.28±0.3332.33±7.8269.12±0.75
Fine Tuning40.32±5.4821.64±0.7836.95±6.6664.29±7.6931.07±1.0456.50±1.8781.82±1.8539.46±0.5970.67±0.72
LN-TUNE13.18±2.171.59±0.536.97±1.9634.80±5.373.96±0.9513.43±2.1247.97±6.257.71±1.8524.53±2.05
Split Learning74.05±1.2126.84±1.0358.14±0.6379.63±0.5330.49±0.3663.95±1.1883.66±0.2735.78±0.3065.20±0.47
Offsite Tuning42.37±2.9824.98±0.5540.56±4.0064.83±6.1730.45±0.8056.43±0.4580.11±1.0736.22±1.3368.43±0.64
SA (ours)76.24±0.2930.92±0.8956.26±0.7981.98±0.4935.31±0.0165.03±0.5985.45±0.4039.26±0.1071.13±0.76
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.265, + 0.482, + 0.432 + ], + "angle": 0, + "content": "of two linear layers, whereas for other datasets, it consists of a single linear layer. We divide 2/3 backbone model as the frontend while the rest \\(1/3\\) is the backend. The frontend and its inner activations are quantized to 8-bit. Each quantization operation uses 32 data samples randomly sampled from the server dataset or subsets, and the batch size is 4. The Adam optimizer with a learning rate of \\(10^{-5}\\) is adopted. The number of epochs for OOD quantization-aware tuning is set to 1, while the final adaptation is set to 100 epochs. All experiments are run repeatedly 3 times with three seeds, and we report the mean performance and standard deviation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.437, + 0.483, + 0.543 + ], + "angle": 0, + "content": "Baseline Approaches. SA is compared with state-of-the-art downstream task adaptation methods including Linear Probing [29], Fine Tuning, Split Learning [44], Offsite Tuning [49], and LN-TUNE [2]. To evaluate the effectiveness of data protection, we extend the state-of-the-art data reconstruction attacks—FORA [53]—to ViT. More details are provided in the Supplementary Materials." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.548, + 0.483, + 0.625 + ], + "angle": 0, + "content": "Evaluation Metrics. To evaluate the effectiveness of the target task adaptation, we use classification accuracy to measure the performance. As for the quality measurement of data reconstruction results, we adopt three metrics: SSIM [48] (S), PSNR [25] (P), and LPIPS [56] (L)." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.636, + 0.431, + 0.652 + ], + "angle": 0, + "content": "4.2. Effectiveness of SA for Task Adaptation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.66, + 0.483, + 0.901 + ], + "angle": 0, + "content": "We create three few-shot adaptation scenarios for each client dataset or domain: 3-shot, 5-shot, and 10-shot per class, with experimental results shown in Table 1. We can easily observe that our SA can always substantially exceed all other baseline approaches. In particular, methods like Fine Tuning, LN-TUNE, and Offsite Tuning lag far behind SA, which may be attributed to the overfitting caused by the low-level feature tuning. The method closest to SA in terms of performance is Split Learning, but its biggest issue is the privacy leakage, which we will discuss in Section 4.3. In Table 2, we also provide the client memory and computation costs of each method (assuming the first three methods are conducted on the client side), from which we can observe that SA even runs with the smallest cost. Note that SA has no training on the client, thus we view the representation extraction time as the training time in Table 2." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.264, + 0.905, + 0.294 + ], + "angle": 0, + "content": "4.3. Effectiveness of SA for Server Pre-trained ViT and Client Data Protection" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.301, + 0.907, + 0.497 + ], + "angle": 0, + "content": "For privacy protection, we first test whether the quantized frontend of SA can result in high-quality models with good downstream task performance. We consider the three most potential variations: 1) directly tune a task module on the quantized frontend (Quant. Frontend); 2) tune a task module on the original frontend though this case is impossible in practice (Original Frontend); 3) attach a backend from other models and tune a task module on the quantized frontend and auxiliary backend (Auxiliary Backend). The experimental results are shown in Table 3, in which all variations perform much poorer than Linear Probing in Table 1 and our SA. In this case, we believe SA can protect the pretrained model to a great extent." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.498, + 0.909, + 0.68 + ], + "angle": 0, + "content": "Then we launch FORA [53] to evaluate the data protection of our SA. For a fair comparison, only Split Learning is compared with SA as only it requires the client to share intermediate data representations with the server. The experiment results are presented in Table 5. First, we can visually observe that the reconstruction results of SA are much worse than those of Split Learning. Then the three metrics also demonstrate the same phenomenon, i.e., FORA can only recover much lower-quality images when attacking SA compared with Split Learning. In summary, our SA offers strong protection of the client data. More experiments on the noise degree can be found in Supplementary Materials." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.689, + 0.667, + 0.705 + ], + "angle": 0, + "content": "4.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.711, + 0.907, + 0.742 + ], + "angle": 0, + "content": "We conduct comprehensive ablation studies in the 5-shot adaptation on CIFAR-100, Places365, and DomainNet-C1." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.746, + 0.906, + 0.837 + ], + "angle": 0, + "content": "Hilbert Transform-based Data Augmentation. We detach the use of HT data augmentation in the frontend quantization and backend tuning ('SA w/o HT Aug'). Compared with the full SA in Table 4, the variation without HT data augmentation performs worse, which proves its gain in improving adaptation performance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Out-of-distribution Quantization-aware Tuning. We design two variations: one is not to use OOD quantization-aware tuning ('SA w/o OOD QAT'), and the other is only to use the quantized frontend \\(\\widehat{F}_{\\mathrm{F}}\\) characterized on the full" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20098" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.09, + 0.325, + 0.13 + ], + "angle": 0, + "content": "Table 2. Comparison of client memory (MB) and computation costs (Min) between SA and other baselines." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.133, + 0.322, + 0.231 + ], + "angle": 0, + "content": "
MethodsGPU MemoryTraining Time
Linear Probing697966
Fine Tuning1030234
LN-TUNE596825
Split Learning893215
Offsite Tuning489623
SA (ours)22332.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.333, + 0.09, + 0.616, + 0.133 + ], + "angle": 0, + "content": "Table 3. Risk assessment of 3 most potential cases over 3 datasets (domains) for obtaining high-quality models from our SA approach." + }, + { + "type": "table", + "bbox": [ + 0.338, + 0.14, + 0.613, + 0.219 + ], + "angle": 0, + "content": "
VariationsCIFAR-100Places365D.Net-CI
Quant. Frontend26.3413.4526.70
Original Frontend26.5313.7227.12
Auxiliary Backend30.6510.1516.56
SA (ours)81.9835.3171.13
" + }, + { + "type": "table_caption", + "bbox": [ + 0.624, + 0.09, + 0.907, + 0.132 + ], + "angle": 0, + "content": "Table 4. Ablation study of the major components of our SA approach in 5-shot adaptation over 3 datasets (domains)." + }, + { + "type": "table", + "bbox": [ + 0.629, + 0.14, + 0.905, + 0.231 + ], + "angle": 0, + "content": "
VariationsCIFAR-100Places365D.Net-CI
SA w/o HT Aug80.3735.2764.59
SA w/o OOD QAT79.0329.5064.42
SA w/o QAT80.0529.5964.16
SA w/o PR Aug79.9029.5652.85
SA (ours)81.9835.3171.13
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.243, + 0.907, + 0.271 + ], + "angle": 0, + "content": "Table 5. Comparison of defensive performance against the state-of-the-art data reconstruction attack—FORA [53]—between SA and Split Learning. We visualize the reconstruction and calculate SSIM (S↓), PSNR (P↓), and LPIPS (L↑) to measure the reconstruction quality." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.273, + 0.905, + 0.467 + ], + "angle": 0, + "content": "
DatasetCIFAR-100Places365D.Net-CI
Ground Truth
Split LearningS:0.80 P:26.1 L:0.46S:0.43 P:20.8 L:0.67S:0.59 P:22.1 L:0.46
SA (ours)S:0.23 P:15.2 L:0.70S:0.30 P:16.6 L:0.73S:0.25 P:14.4 L:0.60
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.477, + 0.483, + 0.506 + ], + "angle": 0, + "content": "Table 6. Ablation study of SA's bi-level noise when defending against FORA [53] on CIFAR-100." + }, + { + "type": "table", + "bbox": [ + 0.107, + 0.507, + 0.468, + 0.764 + ], + "angle": 0, + "content": "
DatasetCIFAR-100
Ground Truth-
SA w/o Laplace NoiseS:0.76 P:23.1 L:0.49
SA w/o Model NoiseS:0.28 P:16.4 L:0.67
SA (ours)S:0.23 P:15.2 L:0.70
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.777, + 0.483, + 0.838 + ], + "angle": 0, + "content": "server dataset ('SA w/o QAT'). Compared with the full SA, the effectiveness of the representation-mixup quantization-aware tuning and the simulated OOD frontends in alleviating quantization-induced performance drop can be verified." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Patch Retrieval Augmentation. We don't use the patch retrieval augmentation and observe how the model performance will change. The big difference between 'SA w/o PR Aug' and 'SA' in Table 4 indicates the essential role of our" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.48, + 0.838, + 0.495 + ], + "angle": 0, + "content": "augmentation here in performance enhancement." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.499, + 0.907, + 0.592 + ], + "angle": 0, + "content": "Bi-level Noisy Representation Extraction. We detach the use of model Gaussian noise ('SA w/o Model Noise') and representation Laplace noise ('SA w/o Laplace Noise'), respectively, and the experiments are shown in Table 6. Model noise and representation noise play a synergistic role greater than the simple sum of them in data protection." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.607, + 0.634, + 0.622 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.633, + 0.907, + 0.83 + ], + "angle": 0, + "content": "In this work, we propose Split Adaptation (SA) to adapt pre-trained ViTs for downstream tasks while protecting the data and model in the sever-client framework. After dividing the ViT into a frontend and a backend, SA innovatively uses model quantization to protect the frontend. SA also allows the client to introduce a bi-level noise to the quantized frontend and the extracted data representations for data protection. Accordingly, SA uses data-level and model-level augmentation to mitigate the influence of the quantization and the noise. Besides, SA proposes patch retrieval augmentation to address overfitting in few-shot adaptation. Extensive experiments have validated the superiority of SA in task adaptation and data-model protection." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.845, + 0.673, + 0.862 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.907, + 0.902 + ], + "angle": 0, + "content": "This work was partially supported by NTU-NAP startup grant 024584-00001." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20099" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.184 + ], + "angle": 0, + "content": "[1] Corinne G Allaart, Björn Keyser, Henri Bal, and Aart Van Halteren. Vertical split learning—an exploration of predictive performance in medical and other use cases. In 2022 International Joint Conference on Neural Networks (IJCNN), pages 1-8. IEEE, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.185, + 0.484, + 0.241 + ], + "angle": 0, + "content": "[2] Samyadeep Basu, Shell Hu, Daniela Massiceti, and Soheil Feizi. Strong baselines for parameter-efficient few-shot finetuning. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 11024-11031, 2024. 7, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.242, + 0.482, + 0.269 + ], + "angle": 0, + "content": "[3] Tom B Brown. Language models are few-shot learners. arXiv preprint arXiv:2005.14165, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.271, + 0.482, + 0.339 + ], + "angle": 0, + "content": "[4] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 2, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.341, + 0.482, + 0.407 + ], + "angle": 0, + "content": "[5] Shoufa Chen, Chongjian Ge, Zhan Tong, Jiangliu Wang, Yibing Song, Jue Wang, and Ping Luo. Adaptformer: Adapting vision transformers for scalable visual recognition. Advances in Neural Information Processing Systems, 35:16664-16678, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.41, + 0.482, + 0.465 + ], + "angle": 0, + "content": "[6] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.467, + 0.482, + 0.534 + ], + "angle": 0, + "content": "[7] Yoni Choukroun, Eli Kravchik, Fan Yang, and Pavel Kisilev. Low-bit quantization of neural networks for efficient inference. In 2019 IEEE/CVF International Conference on Computer Vision Workshop (ICCVW), pages 3009-3018. IEEE, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.537, + 0.482, + 0.605 + ], + "angle": 0, + "content": "[8] Zeki Murat Çınar, Abubakar Abdussalam Nuhu, Qasim Zeešan, Orhan Korhan, Mohammed Asmael, and Babak Safaei. Machine learning in predictive maintenance towards sustainable smart manufacturing in industry 4.0. Sustainability, 12 (19):8211, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.607, + 0.482, + 0.661 + ], + "angle": 0, + "content": "[9] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.663, + 0.482, + 0.69 + ], + "angle": 0, + "content": "[10] Matthew F Dixon, Igor Halperin, and Paul Bilokon. Machine learning in finance. Springer, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.691, + 0.482, + 0.759 + ], + "angle": 0, + "content": "[11] Jiahua Dong, Lixu Wang, Zhen Fang, Gan Sun, Shichao Xu, Xiao Wang, and Qi Zhu. Federated class-incremental learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10164–10173, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.761, + 0.482, + 0.815 + ], + "angle": 0, + "content": "[12] Wei Dong, Qiyao Luo, and Ke Yi. Continual observation under user-level differential privacy. In 2023 IEEE Symposium on Security and Privacy (SP), pages 2190-2207. IEEE, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.818, + 0.482, + 0.857 + ], + "angle": 0, + "content": "[13] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. ICLR, 2020. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.859, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[14] Juanru Fang, Wei Dong, and Ke Yi. Shifted inverse: A general mechanism for monotonic functions under user differential privacy. In Proceedings of the 2022 ACM SIGSAC Con-" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.094, + 0.905, + 0.12 + ], + "angle": 0, + "content": "ference on Computer and Communications Security, pages 1009-1022, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.164 + ], + "angle": 0, + "content": "[15] Chongyang Gao, Lixu Wang, Chenkai Weng, Xiao Wang, and Qi Zhu. Practical unlearning for large language models. arXiv preprint arXiv:2407.10223, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.165, + 0.905, + 0.22 + ], + "angle": 0, + "content": "[16] Xinben Gao and Lan Zhang. {PCAT}: Functionality and data stealing from split learning by {Pseudo-Client} attack. In 32nd USENIX Security Symposium (USENIX Security 23), pages 5271-5288, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.222, + 0.905, + 0.316 + ], + "angle": 0, + "content": "[17] Yansong Gao, Minki Kim, Sharif Abuadbba, Yeonjae Kim, Chandra Thapa, Kyuyeon Kim, Seyit A Camtep, Hyoungshick Kim, and Surya Nepal. End-to-end evaluation of federated learning and split learning for internet of things. In 2020 International Symposium on Reliable Distributed Systems (SRDS), pages 91-100. IEEE Computer Society, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.32, + 0.905, + 0.403 + ], + "angle": 0, + "content": "[18] Jiahui Gong, Jingtao Ding, Fanjin Meng, Guilong Chen, Hong Chen, Shen Zhao, Haisheng Lu, and Yong Li. A population-to-individual tuning framework for adapting pretrained lm to on-device user intent prediction. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 896-907, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.405, + 0.905, + 0.473 + ], + "angle": 0, + "content": "[19] Junfeng Guo, Yiming Li, Lixu Wang, Shu-Tao Xia, Heng Huang, Cong Liu, and Bo Li. Domain watermark: Effective and harmless dataset copyright protection is closed at hand. Advances in Neural Information Processing Systems, 36:54421-54450, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.476, + 0.905, + 0.517 + ], + "angle": 0, + "content": "[20] Song Han, Huizi Mao, and William J Dally. Deep compression: Compressing deep neural networks with pruning, trained quantization and huffman coding. *ICLR*, 2016. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.519, + 0.905, + 0.586 + ], + "angle": 0, + "content": "[21] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.589, + 0.905, + 0.657 + ], + "angle": 0, + "content": "[22] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.66, + 0.905, + 0.715 + ], + "angle": 0, + "content": "[23] Zecheng He, Tianwei Zhang, and Ruby B Lee. Model inversion attacks against collaborative inference. In Proceedings of the 35th Annual Computer Security Applications Conference, pages 148-162, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.717, + 0.905, + 0.77 + ], + "angle": 0, + "content": "[24] Zecheng He, Tianwei Zhang, and Ruby B Lee. Attacking and protecting data privacy in edge-cloud collaborative inference systems. IEEE Internet of Things Journal, 8(12):9706-9716, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.773, + 0.905, + 0.815 + ], + "angle": 0, + "content": "[25] Alain Hore and Djemel Ziou. Image quality metrics: Psnr vs. ssm. In 2010 20th international conference on pattern recognition, pages 2366-2369. IEEE, 2010. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.817, + 0.905, + 0.872 + ], + "angle": 0, + "content": "[26] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In European Conference on Computer Vision, pages 709-727. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[27] Peter Kairouz, H Brendan McMahan, Brendan Avent, Aurélien Bellet, Mehdi Bennis, Arjun Nitin Bhagoji, Kallista" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.094, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "20100" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.484, + 0.147 + ], + "angle": 0, + "content": "Bonawitz, Zachary Charles, Graham Cormode, Rachel Cummings, et al. Advances and open problems in federated learning. Foundations and trends® in machine learning, 14(1-2): 1-210, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.484, + 0.219 + ], + "angle": 0, + "content": "[28] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.221, + 0.484, + 0.276 + ], + "angle": 0, + "content": "[29] Simon Kornblith, Jonathon Shlens, and Quoc V Le. Do better imagenet models transfer better? In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2661-2671, 2019. 2, 7, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.278, + 0.483, + 0.305 + ], + "angle": 0, + "content": "[30] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.307, + 0.483, + 0.374 + ], + "angle": 0, + "content": "[31] Yuhang Li, Ruihao Gong, Xu Tan, Yang Yang, Peng Hu, Qi Zhang, Fengwei Yu, Wei Wang, and Shi Gu. Brecq: Pushing the limit of post-training quantization by block reconstruction. In International Conference on Learning Representations. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.377, + 0.483, + 0.432 + ], + "angle": 0, + "content": "[32] Chenxi Liu, Lixu Wang, Lingjuan Lyu, Chen Sun, Xiao Wang, and Qi Zhu. Deja vu: Continual model generalization for unseen domains. In The Eleventh International Conference on Learning Representations, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.434, + 0.483, + 0.503 + ], + "angle": 0, + "content": "[33] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 6, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.505, + 0.483, + 0.559 + ], + "angle": 0, + "content": "[34] Zhenhua Liu, Yunhe Wang, Kai Han, Wei Zhang, Siwei Ma, and Wen Gao. Post-training quantization for vision transformer. Advances in Neural Information Processing Systems, 34:28092-28103, 2021. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.562, + 0.483, + 0.602 + ], + "angle": 0, + "content": "[35] Alejandro López-Cifuentes, Marcos Escudero-Vinolo, Jesús Bescós, and Álvaro García-Martín. Semantic-aware scene recognition. Pattern Recognition, 102:107256, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.604, + 0.483, + 0.646 + ], + "angle": 0, + "content": "[36] Payal Mohapatra, Lixu Wang, and Qi Zhu. Phase-driven domain generalizable learning for nonstationary time series. arXiv preprint arXiv:2402.05960, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.648, + 0.483, + 0.701 + ], + "angle": 0, + "content": "[37] Mehdi Noroozi and Paolo Favaro. Unsupervised learning of visual representations by solving jigsaw puzzles. In European conference on computer vision, pages 69-84. Springer, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.705, + 0.483, + 0.771 + ], + "angle": 0, + "content": "[38] Dario Pasquini, Giuseppe Ateniese, and Massimo Bernaschi. Unleashing the tiger: Inference attacks on split learning. In Proceedings of the 2021 ACM SIGSAC Conference on Computer and Communications Security, pages 2113-2129, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.775, + 0.483, + 0.843 + ], + "angle": 0, + "content": "[39] Xingchao Peng, Qinxun Bai, Xide Xia, Zijun Huang, Kate Saenko, and Bo Wang. Moment matching for multi-source domain adaptation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1406-1415, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[40] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervi" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "sion. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.907, + 0.19 + ], + "angle": 0, + "content": "[41] K Shailaja, Banoth Seetharamulu, and MA Jabbar. Machine learning in healthcare: A review. In 2018 Second international conference on electronics, communication and aerospace technology (ICECA), pages 910-914. IEEE, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.193, + 0.905, + 0.262 + ], + "angle": 0, + "content": "[42] Chandra Thapa, Mahawaga Arachchige Pathum Chamikara, and Seyit A Camtepe. Advancements of federated learning towards privacy preservation: from federated learning to split learning. Federated Learning Systems: Towards Next-Generation AI, pages 79–109, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.264, + 0.905, + 0.332 + ], + "angle": 0, + "content": "[43] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In International conference on machine learning, pages 10347-10357. PMLR, 2021. 6, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.334, + 0.905, + 0.389 + ], + "angle": 0, + "content": "[44] Praneeth Vepakomma, Otkrist Gupta, Tristan Swedish, and Ramesh Raskar. Split learning for health: Distributed deep learning without sharing raw patient data. arXiv preprint arXiv:1812.00564, 2018. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.391, + 0.905, + 0.46 + ], + "angle": 0, + "content": "[45] Jingye Wang, Ruoyi Du, Dongliang Chang, Kongming Liang, and Zhanyu Ma. Domain generalization via frequency-domain-based feature disentanglement and interaction. In Proceedings of the 30th ACM International Conference on Multimedia, pages 4821-4829, 2022. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.462, + 0.907, + 0.516 + ], + "angle": 0, + "content": "[46] Lixu Wang, Shichao Xu, Xiao Wang, and Qi Zhu. Addressing class imbalance in federated learning. In Proceedings of the AAAI conference on artificial intelligence, pages 10165-10173, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.519, + 0.905, + 0.587 + ], + "angle": 0, + "content": "[47] Lixu Wang, Shichao Xu, Ruiqi Xu, Xiao Wang, and Qi Zhu. Non-transferable learning: A new approach for model ownership verification and applicability authorization. In International Conference on Learning Representation (ICLR-2022), 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.589, + 0.905, + 0.644 + ], + "angle": 0, + "content": "[48] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.647, + 0.905, + 0.686 + ], + "angle": 0, + "content": "[49] Guangxuan Xiao, Ji Lin, and Song Han. Offsite-tuning: Transfer learning without full model. arXiv preprint arXiv:2302.04870, 2023. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.689, + 0.905, + 0.757 + ], + "angle": 0, + "content": "[50] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. Advances in neural information processing systems, 34: 12077-12090, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.759, + 0.905, + 0.815 + ], + "angle": 0, + "content": "[51] Chengming Xu, Siqian Yang, Yabiao Wang, Zhanxiong Wang, Yanwei Fu, and Xiangyang Xue. Exploring efficient few-shot adaptation for vision transformers. Transactions on Machine Learning Research. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.817, + 0.905, + 0.87 + ], + "angle": 0, + "content": "[52] Peng Xu, Xiatian Zhu, and David A Clifton. Multimodal learning with transformers: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(10):12113-12132, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[53] Xiaoyang Xu, Mengda Yang, Wenzhe Yi, Ziang Li, Juan Wang, Hongxin Hu, Yong Zhuang, and Yaxin Liu. A stealthy" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "20101" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.124, + 0.092, + 0.484, + 0.147 + ], + "angle": 0, + "content": "wrongdoer: Feature-oriented reconstruction attack against split learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12130-12139, 2024. 1, 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.484, + 0.218 + ], + "angle": 0, + "content": "[54] Yupeng Yin, Xianglong Zhang, Huanle Zhang, Feng Li, Yue Yu, Xiuzhen Cheng, and Pengfei Hu. Ginver: Generative model inversion attacks against collaborative inference. In Proceedings of the ACM Web Conference 2023, pages 2122-2131, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.22, + 0.483, + 0.288 + ], + "angle": 0, + "content": "[55] Zhihang Yuan, Chenhao Xue, Yiqi Chen, Qiang Wu, and Guangyu Sun. Ptq4vit: Post-training quantization for vision transformers with twin uniform quantization. In European conference on computer vision, pages 191-207. Springer, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.29, + 0.483, + 0.359 + ], + "angle": 0, + "content": "[56] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.361, + 0.483, + 0.429 + ], + "angle": 0, + "content": "[57] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In Proceedings of the IEEE international conference on computer vision, pages 2223-2232, 2017. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.429 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20102" + } + ] +] \ No newline at end of file diff --git a/2025/Split Adaptation for Pre-trained Vision Transformers/b5d6478d-8866-4dba-8ddb-9d84bd75792a_origin.pdf b/2025/Split Adaptation for Pre-trained Vision Transformers/b5d6478d-8866-4dba-8ddb-9d84bd75792a_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..4eca6973563b9aae2ce9bd9aad93c5a178942493 --- /dev/null +++ b/2025/Split Adaptation for Pre-trained Vision Transformers/b5d6478d-8866-4dba-8ddb-9d84bd75792a_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2e5520b69c070d907c6b1d4e6b977b345912d418cd08e59d58cb35d8e7e5251f +size 5697998 diff --git a/2025/Split Adaptation for Pre-trained Vision Transformers/full.md b/2025/Split Adaptation for Pre-trained Vision Transformers/full.md new file mode 100644 index 0000000000000000000000000000000000000000..3e6d614bca14e3d87d6692ba6c59f812b0dfa0f6 --- /dev/null +++ b/2025/Split Adaptation for Pre-trained Vision Transformers/full.md @@ -0,0 +1,337 @@ +# Split Adaptation for Pre-trained Vision Transformers + +Lixu Wang $^{1,2*}$ , Bingqi Shang $^{1*}$ , Yi Li $^{1}$ , Payal Mohapatra $^{1}$ , Wei Dong $^{2}$ , Xiao Wang $^{1}$ , Qi Zhu $^{1}$ , Northwestern University. $^{2}$ Nanyang Technological University. + +lixu.wang@ntu.edu.sg, {bingqishang2025, yili2023.1, payalmohapatra2026}@u.northwestern.edu, wei_dong@ntu.edu.sg, {wangxiao, qzhu}@northwestern.edu + +# Abstract + +Vision Transformers (ViTs), extensively pre-trained on large-scale datasets, have become fundamental to foundation models, enabling adaptation to diverse downstream tasks. Existing adaptation methods typically require direct data access, rendering them infeasible in privacy-sensitive domains where clients are often reluctant to share their data. A straightforward solution may be sending the pretrained ViT to clients for local adaptation, which poses issues of model intellectual property and incurs heavy client computation overhead. To address these issues, we propose a novel split adaptation (SA) method that enables effective downstream adaptation while protecting data and models. SA, inspired by split learning (SL), segments the pre-trained ViT into a frontend and a backend, with only the frontend shared with the client for data representation extraction. But unlike regular SL, SA replaces frontend parameters with low-bit quantized values, preventing direct exposure of the model. SA allows the client to add bi-level noise to the frontend and the extracted data representations, ensuring data protection. Accordingly, SA incorporates data-level and model-level out-of-distribution enhancements to mitigate noise injection's impact. Our SA focuses on the challenging few-shot adaptation and adopts patch retrieval augmentation for overfitting alleviation. Extensive experiments on multiple datasets validate SA's superiority over state-of-the-art methods and demonstrate its defense against advanced data reconstruction attacks while preventing model leakage with minimal computation cost on the client side. The source codes can be found at https://github.com/conditionWang/Split_Adaptation. + +# 1. Introduction + +In recent years, pre-trained vision transformers (ViTs) [13] have become mainstream foundational models in computer vision, delivering remarkable performance across a range of + +visual tasks [13, 28, 50, 52]. ViTs have been equipped with general semantic knowledge during large-scale pre-training, which can be readily transferred to other domains including sensitive fields such as healthcare [41], finance [10], and manufacturing [8]. However, adapting pre-trained ViTs in these sensitive domains is extremely challenging. + +In real-world scenarios, model adaptation often follows a two-party framework, where the pre-trained model is hosted on a server, and the training data resides on a client. In non-sensitive domains, training data can be directly shared with the server. Real-world examples include users uploading data to access OpenAI's CLIP model API [40] or tune GPT-3.5 [3]. However, in sensitive domains, data is often tightly linked to personal privacy [12, 14], intellectual property (IP) [19, 47], and ethical concerns [15], making any form of public exposure, sharing, or transfer strictly prohibited. In this context, an alternative solution might involve sending the model to the client for local adaptation, as in federated learning [11, 27, 46]. However, large pre-trained models are expensive to obtain, making their owners unlikely to share or transfer them freely [47]. Moreover, given that these pre-trained models consist of billions of parameters, adaptation on the client side would incur substantial computational and storage costs. + +To our knowledge, there have been a few preliminary attempts to achieve downstream adaptation of pre-trained models while safeguarding data and models. Split learning (SL) [1, 44] divides the model into segments managed by different parties, which can partially protect the model from full exposure. Additionally, SL only communicates intermediate representations, avoiding the leakage of raw data. However, SL has critical drawbacks. On one hand, emerging malicious attacks, such as data reconstruction attacks (DRA) [16, 23, 38, 53, 54], can infer sensitive information from the intermediate representations. On the other hand, standard SL places the model training on the client side, imposing high computational costs. Beyond SL, a recent method called offsite tuning (OT) [49] has been proposed, which involves the server sending an emulator—a simplified version of the pre-trained model—to the client. The + +![](images/3cdfb9127d5056f0a9058f7ff1ffba42ffa2f05ed3cb132d6bbc848741182b17.jpg) +Figure 1. The comparison between different downstream task adaptation approaches. Our SA can achieve effective few-shot task adaptation with minimal computation cost on the client side while protecting both the model and data. + +client can then fine-tune this emulator using its data, typically employing parameter-efficient techniques like adapter tuning [5]. After training, the client uploads these adapters back to the server, where they are integrated into the pretrained model. However, while OT protects data, the emulator is fully exposed, allowing adversaries to create high-performance models easily. Moreover, OT's adaptation depends on the substantial computational and storage demands for adapter tuning on the client side. + +To address the challenge of adapting pre-trained ViTs in sensitive domains, we propose a split adaptation (SA) method, which is shown in Figure 1. Like SL, SA divides the pre-trained ViT into two parts: a frontend and a backend. The server sends the frontend to the client, enabling data representation extraction. However, unlike SL, SA innovatively uses quantization to replace the frontend's parameters with low-bit values, preventing direct exposure of the original parameters. For data protection, the client injects random noise into the received frontend and its extracted data representations. Such injected noise unavoidably hurts the following adaptation, hence SA includes data-level and model-level out-of-distribution enhancements to improve the model generalization. SA is particularly designed for the challenging few-shot adaptation scenario, where it uses a patch retrieval augmentation technique to alleviate overfitting. Extensive experiments across four datasets demonstrate that SA significantly outperforms other state-of-the-art approaches in adaptation performance. Moreover, SA effectively defends against advanced DRAs and prevents model leakage. In particular, SA's effectiveness only relies on small computation cost on the client side. In summary: + +- We focus on solving a practical and important problem—adapting pre-trained ViTs for downstream tasks while protecting data and models. Solving this problem enables wider application of pre-trained models in sensitive fields. +- We propose Split Adaptation (SA) that innovatively leverages model splitting and quantization to protect data and models during downstream task adaptation. SA focuses + +on challenging few-shot scenarios, in which SA uses a patch retrieval augmentation to combat overfitting. + +- We conduct extensive experiments across multiple datasets and state-of-the-art baseline approaches, with results validating SA's effectiveness in achieving superior adaptation while ensuring data and model protection. + +# 2. Related Works + +# 2.1. Split Learning + +Split learning (SL) [44] is a collaborative learning approach widely used in the multi-party computation framework. SL operates by dividing the machine learning (ML) model into segments, with each managed by a different party. This approach enhances privacy by transmitting only intermediate data representations rather than raw data. Many practical examples of SL in training and adapting ML models across various domains emerge, including healthcare [1, 44], financial [1, 42], and edge computing [17, 18]. Despite these successes, more adversarial attacks illustrate the vulnerability of privacy leakage in SL, with the severest one – data reconstruction attack (DRA) [23]. DRA aims to recover the raw data from its corresponding model representations. In the context of SL, a variety of DRAs have been witnessed, among them, some are based on optimization to imitate the ground-truth representations (e.g., FSHA [38] and PCAT [16]), while others leverage generative models to recover the raw data by exploiting the intrinsic semantics (e.g., Ginver [54] and FORA [53]). Therefore, new SL approaches are needed to help protect data privacy better. + +# 2.2. Vision Transformer Adaptation + +The vision transformers (ViT) [13] are usually pre-trained with large-scale data to learn general semantics and knowledge. Representative pre-training approaches include supervised [28] and self-supervised learning like contrastive learning [4, 6, 21, 40] and auxiliary task training [22, 37, 57]. Owing to such extensive pre-training, adapting pretrained ViT for distinct downstream tasks has become a mainstream paradigm. The most regular method is to tune the entire model, which consumes substantial computational and memory resources. Another widely used approach is linear probing [29], which only trains a task module while keeping the ViT frozen, but the prominent shortcoming is the inferior performance. Recently, more parameter-efficient approaches have been proposed, like adapter tuning [5], prompt tuning [26], and prefix tuning [51]. However, these approaches still require direct access to the task data. In contrast to the above approaches, a new method called offsite tuning [49] can work effectively without direct data access. However, extensive training happens on the data owner side, and there is a high risk of model disclosure, making it impractical in real-world scenarios. + +# 3. Methodology + +# 3.1. Preliminaries + +# 3.1.1. Vision Transformer + +The vision transformer (ViT) [13] is built upon sequential attention encoder layers. Specifically, each encoder layer consists of a multi-head self-attention (MSA) block and a Multi-Layer Perceptron (MLP) block. Layernorm (LN) is applied before every block and residual connections are placed after every block. The MLP contains two linear layers with an intermediate GELU activation. To handle 2D images, ViT reshapes the image $\mathbf{I} \in \mathbb{R}^{H \times W \times C}$ into $N$ flattened patches $I^p \in \mathbb{R}^{P^2 \cdot C}$ , where $H$ , $W$ , and $C$ are the image height, width, and channels, respectively, and $P$ is the pixel resolution of each patch. For the first layer of ViT, to unify the vector dimensions, a trainable linear projection layer with neuron weights $[\mathbf{W}_1^E, \dots, \mathbf{W}_N^E]$ is used to map each vectorized patch into dimension $d$ . In this case, the input to the first transformer layer is: + +$$ +\mathbf {X} _ {1} = \left[ \mathbf {x} ^ {\mathrm {C L S}}; I _ {1} ^ {p} \mathbf {W} _ {1} ^ {E}; \dots ; I _ {N} ^ {p} \mathbf {W} _ {N} ^ {E} \right] + \mathbf {E} ^ {\mathrm {P O S}}, \tag {1} +$$ + +$$ +w h e r e \mathbf {W} ^ {E} \in \mathbb {R} ^ {(P ^ {2} \cdot C) \times d}, \mathbf {E} ^ {\mathrm {P O S}} \in \mathbb {R} ^ {(N + 1) \times d}, +$$ + +where $\mathbf{x}^{\mathrm{CLS}}\in \mathbb{R}^{d}$ is a classification token, and $\mathbf{E}^{\mathrm{POS}}$ is the position encoding. Then for each of the following layers, the input sequence $\mathbf{X}_l$ is fed into each head of the MSA module to obtain three linearly projected matrices: Query $\mathbf{Q}_l = \mathbf{X}_l\mathbf{W}_l^Q$ , Key $\mathbf{K}_l = \mathbf{X}_l\mathbf{W}_l^K$ , and Value $\mathbf{V}_l = \mathbf{X}_l\mathbf{W}_l^V$ . Then the Softmax function is applied to normalize the attention score $\mathbf{A}_l = \mathbf{Q}_l\mathbf{K}_l^\top$ and the output $\mathbf{Z}_l$ is: + +$$ +\mathbf {Z} _ {l} = \operatorname {M S A} \left(\mathbf {X} _ {l}\right) = \operatorname {S o f t m a x} \left(\frac {\mathbf {A} _ {l}}{\sqrt {d}}\right) \mathbf {V} _ {l}. \tag {2} +$$ + +Suppose the two linear layers in MLP are parameterized by $\mathbf{W}_{l,1},\mathbf{b}_{l,1}$ and $\mathbf{W}_{l,2},\mathbf{b}_{l,2}$ , the output is then computed as: + +$$ +\operatorname {M L P} \left(\mathbf {Z} _ {l}\right) = \operatorname {G E L U} \left(\mathbf {Z} _ {l} \mathbf {W} _ {l, 1} + \boldsymbol {b} _ {l, 1}\right) \mathbf {W} _ {l, 2} + \boldsymbol {b} _ {l, 2}. \tag {3} +$$ + +Finally, both the MSA and MLP modules are linked via residual connections, thus we have the final output of each transformer layer: $\mathbf{X}_{l + 1} = \mathrm{LN}(\mathbf{Z}_l + \mathrm{MLP}(\mathbf{Z}_l))\in$ $\mathbb{R}^{(N + 1)\times d}$ , where $\mathbf{Z}_l = \mathrm{LN}(\mathbf{X}_l + \mathrm{MSA}(\mathbf{X}_l))\in \mathbb{R}^{(N + 1)\times d}$ + +# 3.1.2. Problem Formulation + +When adapting pre-trained ViT for certain downstream tasks, we consider a representative client-server framework in which the client queries the server for the adaptation request. The client owns the task adaptation dataset $\mathcal{D}^{\mathrm{C}} = \{(x_i,y_i)\}_{i = 1}^{N^{\mathrm{C}}} (N^{\mathrm{C}}$ is the sample number, in practice, we suppose $N^{\mathrm{C}}$ is small even few-shot) with each sample drawn from a marginal input distribution $\pmb{x}_i\sim \mathcal{P}_X^{\mathrm{C}}$ and labeled with a marginal task distribution $\mathcal{P}_{\mathcal{Y}}^{\mathrm{C}}$ . + +On the server side, the pre-trained ViT $F_{\Theta}$ parameterized on $\Theta$ is assumed to consist of $L^F$ attention layers, and each layer is parameterized on $\theta$ , i.e., $F := [f_{\theta_1}, \dots, f_{\theta_{L^F}}]$ . Besides, we assume the server owns another dataset $\mathcal{D}^{\mathrm{S}} = \{(x_i, y_i)\}_{i=1}^{N^{\mathrm{S}}}$ collected from public domains with the marginal input and task distributions, $\mathcal{P}_{\mathcal{X}}^{\mathrm{S}}$ and $\mathcal{P}_{\mathcal{Y}}^{\mathrm{S}}$ , respectively. In real-world scenarios, the datasets owned by the client and the server are distinct but relevant, i.e., $\mathcal{P}_{\mathcal{X}}^{\mathrm{C}} \neq \mathcal{P}_{\mathcal{X}}^{\mathrm{S}}, \mathcal{P}_{\mathcal{Y}}^{\mathrm{C}} \neq \mathcal{P}_{\mathcal{Y}}^{\mathrm{S}}$ but $I(\mathcal{X}^{\mathrm{C}}, \mathcal{X}^{\mathrm{S}}) > 0$ , $I(\mathcal{Y}^{\mathrm{C}}, \mathcal{Y}^{\mathrm{S}}) > 0$ where $I(\cdot; \cdot)$ calculates the mutual information. + +Objective of Task Performance. Pre-trained ViT adaptation aims to obtain a downstream task module $g$ parameterized on $\gamma$ that can achieve as good performance as possible on the client data distribution when combined with $F_{\Theta}$ . Certainly, such adaptation also allows $F_{\Theta}$ tuning if it can enable performance improvement on the downstream task. In this case, the objective is formulated as, + +$$ +\Theta , \gamma = \arg \min _ {\Theta , \gamma} \mathbb {E} _ {\boldsymbol {x} \sim \mathcal {P} _ {\chi} ^ {\mathrm {C}}, y \sim \mathcal {P} _ {\mathcal {Y}} ^ {\mathrm {C}}} \mathcal {L} (\boldsymbol {x}, y; \Theta \circ \gamma), \tag {4} +$$ + +where $\mathcal{L}(\cdot ;\cdot)$ is a loss function that measures the model performance regarding certain data distribution. + +Objective of Privacy Protection. In addition to the above standard objective, the intrinsic privacy concerns of the server-client framework should be addressed, i.e., the pretrained model and client data should be protected during adaptation when supposing the server and the client may be semi-honest. On the one hand, the adaptation should prohibit the client from stealing surrogate models with good performance or obtaining them at a small cost, on the other hand, any client data reconstruction on the server side should be avoided. + +# 3.2. Overview of The Proposed Split Adaptation + +To achieve effective adaptation while protecting the pre-trained ViT and client data, we propose a method called Split Adaptation (SA), shown in Figure 2. Our SA starts with dividing the pre-trained ViT $F_{\Theta}$ into two parts: a frontend $F^{\mathrm{F}}$ and a backend $F^{\mathrm{B}}$ . The original $F^{\mathrm{F}}$ is not sent to the client, instead, we leverage model quantization to send a quantized version $\widehat{F}^{\mathrm{F}}$ to the client. To mitigate the influence of frontend quantization, SA includes out-of-distribution (OOD) augmentations from both data and model levels to enhance the cross-distribution generalization of the frontend and backend. As for the client side, it adds random noise to the received $\widehat{F}^{\mathrm{F}}$ first and adopts a patch-level retrieve-then-replace strategy to generate more client data representations that help alleviate overfitting in few-shot adaptation. Before being uploaded to the server for the final adaptation, these augmented representations are protected by adding noise again. + +![](images/2fe7a701a2a45e6b89d7ae2ea5783a2e1e33ee6e9991aaf14b5e56e87849fd02.jpg) +Figure 2. Overview of Split Adaptation (SA) for pre-trained ViT adaptation. SA divides the pre-trained ViT into a frontend and a backend. After applying Out-of-distribution Enhanced Quantization to the frontend, its quantized version is sent to the client. To mitigate the impact of quantization, SA adopts Out-of-distribution Quantization-aware Tuning to enhance backend's generalization. As for the client, it injects random noise to the received frontend and retrieves then replaces randomly selected patches to augment more client data representations, which are sent to the server for the final adaptation after being added with noise again. + +# 3.3. Model Splitting on The Server Side + +At the beginning of SA, the pre-trained ViT is divided into a frontend $F^{\mathrm{F}} \coloneqq [f_{\theta_1}, \dots, f_{\theta_K}]$ and a backend $F^{\mathrm{B}} \coloneqq [f_{\theta_{K+1}}, \dots, f_{\theta_{LF}}]$ at the $K$ -th layer ( $K > L^{F} / 2$ ). Then the frontend is distributed to the client for representation extraction. Although we empirically found that the frontend cannot perform well (Section 4.3), we still hope to protect it from disclosing the original model parameters. Quantization [20, 34], as a representative model compression approach, can significantly reduce the computational and memory cost without sacrificing model performance. As quantization works by representing model weights and activations with low-precision values, it can potentially protect the model. However, standard quantization needs a certain amount of data drawn from the target task, which is inaccessible for the server in our problem. In this case, the left choice is to leverage the server dataset $\mathcal{D}^{\mathrm{S}}$ to quantize the model, but a dedicated design for enhancing the quantized model's cross-distribution generalization is needed. Therefore, we propose an augmentation technique to generate OOD data in terms of $\mathcal{D}^{\mathrm{S}}$ for quantization. + +# 3.3.1. Out-of-distribution Enhanced Quantization + +In practice, we have to consider extreme cases in which the client dataset differs significantly from the server dataset and they may only share general semantics. Hence our augmentation technique should generate OOD data with a large discrepancy to $\mathcal{D}^{\mathrm{S}}$ . Inspired by signal processing, we no + +![](images/33811de070aa6290bb3a984150073cb40b73014f565a79dd50816fc520b6808e.jpg) +Figure 3. Visualization of our Hilbert Transform data augmentation. Compared to amplitude-exchange augmentation [45] (bottom row), our augmentation method (top row) generates data that diverges further from the original in appearance while preserving the original semantics (dog category). + +tice that there is a technique called Hilbert Transform (HT), which only captures the most sharp edge in images. Intuitively, the sharp edges contain the most general semantics, making HT suitable for our problem. Applying HT to a signal in standard signal processing results in a phase shift of $\pi /2$ in the frequency domain [36]. In the visual domain, we can implement HT by transforming an image $\mathbf{I}$ to the frequency domain with the spectrum index $\xi$ via Fourier transform, then the negative $\xi$ needs to multiply with the imaginary unit $i$ , while the positive $\xi$ needs to multiple with + +$-i$ , i.e., + +$$ +\mathrm {H T} (\mathbf {I}) = \mathcal {F} ^ {- 1} \{- i \cdot \operatorname {s g n} (\xi) \mathcal {F} (\mathbf {I}) \}, \qquad (5) +$$ + +where $\mathcal{F}(\cdot)$ and $\mathcal{F}^{-1}(\cdot)$ are Fourier transform and inverse, respectively, and $\mathrm{sgn}(\cdot)$ is a sign function. However, pure HT only provides new images with black backgrounds. To diversify the HT-generated image, we replace its amplitude in the frequency domain with that of the original image. In this way, the generated image is much better than the standard amplitude-exchange augmentation [32, 45] (please refer to Figure 3 for a visual comparison). + +To achieve effective quantization, we adopt the most popular uniform symmetric quantization [20] to implement SA, in which a floating-point value $w$ is projected into a $k$ -bit integer value $w_{\mathrm{q}}$ with a scaling factor $\Delta$ : + +$$ +\widehat {w} = \Psi_ {k} (w, \Delta) = \operatorname {c l a m p} (\operatorname {r o u n d} \left(\frac {w}{\Delta}\right), - 2 ^ {k - 1}, 2 ^ {k - 1} - 1), \tag {6} +$$ + +where $\mathrm{round}(\cdot)$ turns a value into an integer and $\mathrm{clamp}(\cdot, \min, \max)$ constrains the output in the range from $\min$ to $\max$ . Intuitively, the scaling factor $\Delta$ directly impacts the model performance after quantization. Previous studies [7, 34] have tried a variety of measurements to quantify the model output changes and accordingly determine the optimal scaling factors. Among these studies, we follow Yuan et al. [55] to choose a Hessian-guided metric for scaling factor determination. Suppose the pre-trained ViT $F_{\Theta}$ is equipped with a task module $g_{\gamma'}'$ that is tuned on the server dataset $\mathcal{D}^{\mathrm{S}}$ , the model performance can be measured by a loss $\mathbb{E}_{\boldsymbol{x} \sim \mathcal{P}_{\boldsymbol{x}}^{\mathrm{S}}, y \sim \mathcal{P}_{y}^{\mathrm{S}}} \mathcal{L}(\boldsymbol{x}, y; \Theta \circ \gamma')$ (we omit $\boldsymbol{x}$ and $y$ to simplify the used symbols in the following descriptions). Then if we regard that the quantization brings a small perturbation $\epsilon$ to the model parameters, i.e., $\hat{\Theta} = \Theta + \epsilon$ , we can analyze the influence of quantization on the task loss using Taylor series expansion: + +$$ +\mathbb {E} [ \mathcal {L} (\widehat {\Theta} \circ \gamma^ {\prime}) ] - \mathbb {E} [ \mathcal {L} (\Theta \circ \gamma^ {\prime}) ] \approx \epsilon^ {\top} \nabla_ {\Theta} + \frac {1}{2} \epsilon^ {\top} \mathbf {H} _ {\Theta} \epsilon , (7) +$$ + +where $\nabla_{\Theta}$ is the gradients and $\mathbf{H}_{\Theta}$ is the Hessian matrix. Then the optimal scaling factors are the ones corresponding to minimal influence caused by quantization, i.e., $\Delta = \arg \min_{\Delta}(\mathbb{E}[\mathcal{L}(\widehat{\Theta}\circ \gamma^{\prime})] - \mathbb{E}[\mathcal{L}(\Theta \circ \gamma^{\prime})])$ . Certainly, determining and adopting a general scaling factor for the entire model is too coarse to perform well, thus we apply a layer-wise reconstruction method [31] to achieve fine-grained quantization, then the searching optimization of Eq. (7) can be approximated by: + +$$ +\underset {\Delta_ {l - 1}} {\arg \min } \mathbb {E} \left[ \left(\widehat {\mathbf {X}} _ {l} - \mathbf {X} _ {l}\right) ^ {\top} \operatorname {d i a g} \left(\frac {\partial \mathcal {L}}{\partial \mathbf {X} _ {l}} \odot \frac {\partial \mathcal {L}}{\partial \mathbf {X} _ {l}}\right) \left(\widehat {\mathbf {X}} _ {l} - \mathbf {X} _ {l}\right) \right], \tag {8} +$$ + +where $\mathbf{X}_l$ and $\widehat{\mathbf{X}}_l$ are the outputs of the $(l-1)$ -th layer before and after quantization, respectively, and $\odot$ represents the Hadamard product of a matrix. + +With the augmented neighborhood data, the frontend can be quantized with improved cross-distribution generalization. Specifically, we first conduct HT-based data augmentation on the server dataset $\mathcal{D}^{\mathrm{S}}$ to obtain $\mathrm{HT}(\mathcal{D}^{\mathrm{S}})$ . Then the optimal scaling factor for each layer in the frontend $F^{\mathrm{F}}$ is empirically determined by optimizing Eq. (8) on the merged dataset $\mathcal{D}_{\cup}^{\mathrm{S}} = \mathcal{D}^{\mathrm{S}} \cup \mathrm{HT}(\mathcal{D}^{\mathrm{S}})$ and we can obtain a quantized frontend $\widehat{F}^{\mathrm{F}}$ in the end. + +# 3.3.2. Out-of-distribution Quantization-aware Tuning + +To achieve better adaptation performance, the backend also needs to resist the unavoidable parameter difference between the original and quantized model as well as the noise injected by the client that we will introduce in Section 3.4. As a result, we propose an OOD quantization-aware tuning technique to enhance the backend's resistance when encountering OOD data and frontends. + +To simulate potential OOD frontends, we first divide the merged server dataset $\mathcal{D}_{\cup}^{\mathrm{S}}$ into $M$ subsets and denote these subsets as $\{\mathcal{D}_{\cup_m}^{\mathrm{S}}\}_{m = 1}^M$ . Next, each subset $\mathcal{D}_{\cup_m}^{\mathrm{S}}$ is used to quantize the original frontend $F^{\mathrm{F}}$ to obtain a quantized version $\widehat{F}_m^{\mathrm{F}}$ . Then let us use a single quantized frontend $\widehat{F}_m^{\mathrm{F}}$ to introduce how to tune the backend. To begin with, since the backend is represented and the tuning is performed using floating-point values, the data representations extracted by the quantized frontend should be transformed back to floating-point values through the dequantization: + +$$ +\tilde {w} = \Psi^ {- 1} (\hat {w}, \Delta) = \hat {w} \cdot \Delta . \tag {9} +$$ + +In this way, we can obtain the floating-point representation $\widetilde{\mathbf{X}}_K$ by dequantizing the $K$ -th layer's output with the scaling factor $\Delta_K$ . Note that $\widetilde{\mathbf{X}}_K$ is different from the original representation $\mathbf{X}_K$ extracted by the original frontend $F^{\mathrm{F}}$ due to the range constraints of round and clamp functions in Eq. (6). Our following design aims to tune the backend to resist such differences. The most straightforward method is to forward the floating-point representation $\widetilde{\mathbf{X}}_K$ to the backend and conduct tuning using task loss. However, such tuning will make the backend over-fitted to the server dataset $\mathcal{D}^{\mathrm{S}}$ . We solve this by a representation-mixup strategy. Specifically, for a particular data sample $\pmb{x}$ , we feed it to a quantized frontend $\widehat{F}_{m}^{\mathrm{F}}$ and the original frontend $F^{\mathrm{F}}$ to extract representations $\widehat{\mathbf{X}}_{K,m},\mathbf{X}_{K,m}$ . After dequantizing $\widehat{\mathbf{X}}_{K,m}$ into the floating-point form $\widetilde{\mathbf{X}}_{K,m}$ , we obtain a mixup representation: + +$$ +\mathbf {X} _ {K, m} ^ {*} = \lambda \widetilde {\mathbf {X}} _ {K, m} + (1 - \lambda) \mathbf {X} _ {K, m}, \tag {10} +$$ + +where $\lambda$ is drawn from a widely-used beta distribution $\mathrm{Beta}(\lambda;0.75,0.75)$ . Then considering all available quantized frontend, we can tune the backend with the loss: + +$$ +\mathcal {L} _ {\text {t u n e}} = \sum_ {\mathcal {D} _ {\cup} ^ {\mathrm {S}}} \sum_ {m = 0} ^ {M} \mathcal {L} \left(\mathbf {X} _ {K, m} ^ {*}, y; \Theta^ {\mathrm {B}} \circ \gamma^ {\prime}\right). \tag {11} +$$ + +Note that $\mathbf{X}_{K,0}^{*}$ originates from feeding $\pmb{x}$ to the frontend $\widehat{F}^{\mathrm{F}}$ quantized on the full set of $\mathcal{D}_{\cup}^{\mathrm{S}}$ . + +# 3.4. Client Bi-level Noisy Representation Extraction + +In our SA model, the quantized frontend $\widehat{F}^{\mathrm{F}}$ will be sent to the client after the OOD enhanced quantization (Section 3.3.1), and then the client can forward its data to $\widehat{F}^{\mathrm{F}}$ to extract representations and send them to the server. However, there is a risk of data reconstruction when sharing representations, thus we propose a bi-level noisy representation extraction mechanism to defend against such attacks. + +Intuitively speaking, the key to data reconstruction defense is to use the model unknown by the server to extract the representations. Therefore, we allow the client to inject a certain degree of random noise into the received frontend $\widehat{F}^{\mathrm{F}}$ . For each layer in the frontend, we inject a multiplicative noise $n_l^\times$ , and an additive noise $n_l^+$ , which are sampled from two Gaussian distributions, i.e., $n_l^\times \sim \mathcal{P}_{\mathcal{N}(1,\alpha \theta_l)}$ and $n_l^+ \sim \mathcal{P}_{\mathcal{N}(0,\alpha \theta_l)}$ ( $\alpha = 0.01$ controls the noise degree), to the model parameters: + +$$ +\theta_ {l} ^ {\prime} = n _ {l} ^ {\times} \cdot \theta_ {l} + n _ {l} ^ {+}. \tag {12} +$$ + +Next, the client feeds its dataset $\mathcal{D}^{\mathrm{C}}$ to the noisy frontend $\widehat{F}^{\mathrm{F}'}$ to extract representations $\{\widehat{\mathbf{X}}_{K,i}^{\mathrm{C}}\}_{i=1}^{N^{\mathrm{C}}}$ . However, it is still insecure to share these representations. We found that our model-level noise can amplify the defense effect of the regular noise obfuscation [24] at the representation level, which allows to add a smaller degree of Laplace noise $(\mathcal{P}_{\mathrm{Laplace}(0,0.8)})$ to representations for better protection. + +# 3.5. Patch-retrieval-augmented Model Adaptation + +In few-shot adaptation scenarios, like the backend tuning on the server (Section 3.3.2), the adaptation also faces overfitting challenges, but the overfitting here is due to the limited client data available. To tackle such challenges, we design a patch-level retrieval augmentation approach to generate more client data representations. + +Our patch-level retrieval augmentation is inspired by the intuition that different images may contain similar patches. As introduced in Section 3.1.1, the data representations of each layer in a ViT hold the dimension $(N + 1)\times d$ , which can be denoted as $\widehat{\mathbf{X}}_{K,i}^{\mathrm{C}} = [\mathbf{x}_i^{\mathrm{CLS}};\mathbf{x}_{i,1},\dots ,\mathbf{x}_{i,N}]$ (we omit some symbols for simplification). Then we can form $N$ retrieval sets from all client data representations (before adding Laplace noise) and each corresponds to a patch, i.e., $\{S_j\}_{j = 1}^N$ where $S_{j} = \{\mathbf{x}_{i,j}\}_{i = 1}^{N^{\mathrm{C}}}$ . With these sets, for each augmentation of $\widehat{\mathbf{X}}_{K,i}^{\mathrm{C}}$ , we randomly select $N^{\mathrm{P}} = 50$ patches (please see the sensitivity analysis in Supplementary Materials) $\{\mathbf{x}_{i,\mathrm{P}_j}\}_{j = 1}^{N^{\mathrm{P}}}$ (subscript $\mathrm{P}_j$ meets $1\leq \mathrm{P}_j\leq N$ ) and delete them from their corresponding retrieval sets. Next, each $\mathbf{x}_{i,\mathrm{P}_j}$ retrieves the most similar patch $\Omega (\mathbf{x}_{i,\mathrm{P}_j})$ from + +its corresponding retrieval set: + +$$ +\Omega \left(\mathbf {x} _ {i, \mathrm {P} _ {j}}\right) = \arg \max _ {\mathbf {x} \in S _ {\mathrm {P} _ {j}} ^ {\prime}} \frac {\mathbf {x} _ {i , \mathrm {P} _ {j}} \cdot \mathbf {x}}{\left| \mathbf {x} _ {i , \mathrm {P} _ {j}} \right| \left| \mathbf {x} \right|}, \tag {13} +$$ + +where $\mathcal{S}_{\mathrm{P}_j}^{\prime} = \mathcal{S}_{\mathrm{P}_j}\backslash \mathbf{x}_{i,\mathrm{P}_j}$ . After the retrieval operations for all selected patches $\{\mathbf{x}_{i,\mathrm{P}_j}\}_{j = 1}^{N^{\mathrm{P}}}$ , we replace them with the retrieved patches $\{\Omega (\mathbf{x}_{i,\mathrm{P}_j})\}_{j = 1}^{N^{\mathrm{P}}}$ in the original representation $\widehat{\mathbf{X}}_{K,i}^{\mathrm{C}}$ to form a new one $\widehat{\mathbf{X}}_{K,i}^{\prime \mathrm{C}}$ . Note that the selected patches are determined randomly, resulting in newly generated representations that differ across augmentation runs. To generate sufficient augmented representations, we run the augmentation $N^{\mathrm{Aug}}$ times for all client data representations and produce an augmented representation set, i.e., $\widehat{\mathcal{X}}_K^{\mathrm{C}} = \{\{\widehat{\mathbf{X}}_{K,i}^{\mathrm{C}_j}\}_{j = 0}^{N^{\mathrm{Aug}}}\}_{i = 1}^{N^{\mathrm{C}}}$ (we denote the original representation as $\widehat{\mathbf{X}}_{K,i}^{\mathrm{C}_0}$ and the augmented representations as $\widehat{\mathbf{X}}_{K,i}^{\mathrm{C}_{1\sim N^{\mathrm{Aug}}}}$ ). The above patch retrieval augmentation is run by the client, and it adds Laplace noise (Section 3.4) to all representations in $\widehat{\mathcal{X}}_K^{\mathrm{C}}$ before uploading to the server. + +After receiving the representations, the server dequantizes them into floating-point form with the $K$ -th layer's scaling factor $\Delta_K$ by Eq. (9) and launches the adaptation. To further alleviate the overfitting due to limited client data, different from Split Learning, we only allow high-level feature learning in the backend to prevent low-level features from becoming biased. Then the server randomly initializes a task module $g_{\gamma}$ and tunes it with the backend $F^{\mathrm{B}}$ (that has been tuned in Section 3.3.2) by following the loss: + +$$ +\mathcal {L} _ {\mathrm {a d a p t}} = \sum_ {i = 1} ^ {N ^ {\mathrm {C}}} \sum_ {j = 0} ^ {N ^ {\mathrm {A u g}}} \mathcal {L} \left(\widehat {\mathbf {X}} _ {K, i} ^ {\mathrm {C} _ {j}}, y; \Theta^ {\mathrm {B}} \circ \gamma\right). \tag {14} +$$ + +To protect the client's labels, the server sends the task module output to the client, allowing the task loss $\mathcal{L}_{\mathrm{adapt}}$ to be computed on the client side and then sent back to the server. + +# 4. Experiments + +# 4.1. Experimental Settings + +Datasets. ImageNet [9] is the most widely used image dataset collected from public domains, thus we use its test set as the server dataset. For the client datasets, we use CIFAR-100 [30], Places365 [35], and DomainNet [39]. Places365 is a scene recognition dataset that contains 365 classes. DomainNet (D.Net) is a challenging multi-domain dataset, which includes six domains. We select the smallest domain—Clipart (Cl) for the few-shot adaptation scenarios. + +Implementation Details. We conduct experiments on three pre-trained model architectures: ViT-Large [13], DeiT-Base [43], and Swin-L [33] (DeiT-Base and Swin-L's experiments are reported in Supplementary Materials). For the Places365 dataset, the target task module is composed + +Table 1. Performance comparison between our SA and other baseline approaches in 3-shot, 5-shot, and 10-shot adaptation scenarios over 3 datasets. The SA can substantially exceed other methods. We bold and blue the best, and bold the second best. + +
Few-shot Setup3-shot5-shot10-shot
MethodCIFAR-100Places365D.Net-ClCIFAR-100Places365D.Net-ClCIFAR-100Places365D.Net-Cl
Linear Probing74.05±1.2126.18±1.0057.26±1.1178.79±0.6534.25±1.4563.84±0.6683.28±0.3332.33±7.8269.12±0.75
Fine Tuning40.32±5.4821.64±0.7836.95±6.6664.29±7.6931.07±1.0456.50±1.8781.82±1.8539.46±0.5970.67±0.72
LN-TUNE13.18±2.171.59±0.536.97±1.9634.80±5.373.96±0.9513.43±2.1247.97±6.257.71±1.8524.53±2.05
Split Learning74.05±1.2126.84±1.0358.14±0.6379.63±0.5330.49±0.3663.95±1.1883.66±0.2735.78±0.3065.20±0.47
Offsite Tuning42.37±2.9824.98±0.5540.56±4.0064.83±6.1730.45±0.8056.43±0.4580.11±1.0736.22±1.3368.43±0.64
SA (ours)76.24±0.2930.92±0.8956.26±0.7981.98±0.4935.31±0.0165.03±0.5985.45±0.4039.26±0.1071.13±0.76
+ +of two linear layers, whereas for other datasets, it consists of a single linear layer. We divide 2/3 backbone model as the frontend while the rest $1/3$ is the backend. The frontend and its inner activations are quantized to 8-bit. Each quantization operation uses 32 data samples randomly sampled from the server dataset or subsets, and the batch size is 4. The Adam optimizer with a learning rate of $10^{-5}$ is adopted. The number of epochs for OOD quantization-aware tuning is set to 1, while the final adaptation is set to 100 epochs. All experiments are run repeatedly 3 times with three seeds, and we report the mean performance and standard deviation. + +Baseline Approaches. SA is compared with state-of-the-art downstream task adaptation methods including Linear Probing [29], Fine Tuning, Split Learning [44], Offsite Tuning [49], and LN-TUNE [2]. To evaluate the effectiveness of data protection, we extend the state-of-the-art data reconstruction attacks—FORA [53]—to ViT. More details are provided in the Supplementary Materials. + +Evaluation Metrics. To evaluate the effectiveness of the target task adaptation, we use classification accuracy to measure the performance. As for the quality measurement of data reconstruction results, we adopt three metrics: SSIM [48] (S), PSNR [25] (P), and LPIPS [56] (L). + +# 4.2. Effectiveness of SA for Task Adaptation + +We create three few-shot adaptation scenarios for each client dataset or domain: 3-shot, 5-shot, and 10-shot per class, with experimental results shown in Table 1. We can easily observe that our SA can always substantially exceed all other baseline approaches. In particular, methods like Fine Tuning, LN-TUNE, and Offsite Tuning lag far behind SA, which may be attributed to the overfitting caused by the low-level feature tuning. The method closest to SA in terms of performance is Split Learning, but its biggest issue is the privacy leakage, which we will discuss in Section 4.3. In Table 2, we also provide the client memory and computation costs of each method (assuming the first three methods are conducted on the client side), from which we can observe that SA even runs with the smallest cost. Note that SA has no training on the client, thus we view the representation extraction time as the training time in Table 2. + +# 4.3. Effectiveness of SA for Server Pre-trained ViT and Client Data Protection + +For privacy protection, we first test whether the quantized frontend of SA can result in high-quality models with good downstream task performance. We consider the three most potential variations: 1) directly tune a task module on the quantized frontend (Quant. Frontend); 2) tune a task module on the original frontend though this case is impossible in practice (Original Frontend); 3) attach a backend from other models and tune a task module on the quantized frontend and auxiliary backend (Auxiliary Backend). The experimental results are shown in Table 3, in which all variations perform much poorer than Linear Probing in Table 1 and our SA. In this case, we believe SA can protect the pretrained model to a great extent. + +Then we launch FORA [53] to evaluate the data protection of our SA. For a fair comparison, only Split Learning is compared with SA as only it requires the client to share intermediate data representations with the server. The experiment results are presented in Table 5. First, we can visually observe that the reconstruction results of SA are much worse than those of Split Learning. Then the three metrics also demonstrate the same phenomenon, i.e., FORA can only recover much lower-quality images when attacking SA compared with Split Learning. In summary, our SA offers strong protection of the client data. More experiments on the noise degree can be found in Supplementary Materials. + +# 4.4. Ablation Study + +We conduct comprehensive ablation studies in the 5-shot adaptation on CIFAR-100, Places365, and DomainNet-C1. + +Hilbert Transform-based Data Augmentation. We detach the use of HT data augmentation in the frontend quantization and backend tuning ('SA w/o HT Aug'). Compared with the full SA in Table 4, the variation without HT data augmentation performs worse, which proves its gain in improving adaptation performance. + +Out-of-distribution Quantization-aware Tuning. We design two variations: one is not to use OOD quantization-aware tuning ('SA w/o OOD QAT'), and the other is only to use the quantized frontend $\widehat{F}_{\mathrm{F}}$ characterized on the full + +Table 2. Comparison of client memory (MB) and computation costs (Min) between SA and other baselines. + +
MethodsGPU MemoryTraining Time
Linear Probing697966
Fine Tuning1030234
LN-TUNE596825
Split Learning893215
Offsite Tuning489623
SA (ours)22332.5
+ +Table 3. Risk assessment of 3 most potential cases over 3 datasets (domains) for obtaining high-quality models from our SA approach. + +
VariationsCIFAR-100Places365D.Net-CI
Quant. Frontend26.3413.4526.70
Original Frontend26.5313.7227.12
Auxiliary Backend30.6510.1516.56
SA (ours)81.9835.3171.13
+ +Table 4. Ablation study of the major components of our SA approach in 5-shot adaptation over 3 datasets (domains). + +
VariationsCIFAR-100Places365D.Net-CI
SA w/o HT Aug80.3735.2764.59
SA w/o OOD QAT79.0329.5064.42
SA w/o QAT80.0529.5964.16
SA w/o PR Aug79.9029.5652.85
SA (ours)81.9835.3171.13
+ +Table 5. Comparison of defensive performance against the state-of-the-art data reconstruction attack—FORA [53]—between SA and Split Learning. We visualize the reconstruction and calculate SSIM (S↓), PSNR (P↓), and LPIPS (L↑) to measure the reconstruction quality. + +
DatasetCIFAR-100Places365D.Net-CI
Ground Truth
Split LearningS:0.80 P:26.1 L:0.46S:0.43 P:20.8 L:0.67S:0.59 P:22.1 L:0.46
SA (ours)S:0.23 P:15.2 L:0.70S:0.30 P:16.6 L:0.73S:0.25 P:14.4 L:0.60
+ +Table 6. Ablation study of SA's bi-level noise when defending against FORA [53] on CIFAR-100. + +
DatasetCIFAR-100
Ground Truth-
SA w/o Laplace NoiseS:0.76 P:23.1 L:0.49
SA w/o Model NoiseS:0.28 P:16.4 L:0.67
SA (ours)S:0.23 P:15.2 L:0.70
+ +server dataset ('SA w/o QAT'). Compared with the full SA, the effectiveness of the representation-mixup quantization-aware tuning and the simulated OOD frontends in alleviating quantization-induced performance drop can be verified. + +Patch Retrieval Augmentation. We don't use the patch retrieval augmentation and observe how the model performance will change. The big difference between 'SA w/o PR Aug' and 'SA' in Table 4 indicates the essential role of our + +augmentation here in performance enhancement. + +Bi-level Noisy Representation Extraction. We detach the use of model Gaussian noise ('SA w/o Model Noise') and representation Laplace noise ('SA w/o Laplace Noise'), respectively, and the experiments are shown in Table 6. Model noise and representation noise play a synergistic role greater than the simple sum of them in data protection. + +# 5. Conclusion + +In this work, we propose Split Adaptation (SA) to adapt pre-trained ViTs for downstream tasks while protecting the data and model in the sever-client framework. After dividing the ViT into a frontend and a backend, SA innovatively uses model quantization to protect the frontend. SA also allows the client to introduce a bi-level noise to the quantized frontend and the extracted data representations for data protection. Accordingly, SA uses data-level and model-level augmentation to mitigate the influence of the quantization and the noise. Besides, SA proposes patch retrieval augmentation to address overfitting in few-shot adaptation. Extensive experiments have validated the superiority of SA in task adaptation and data-model protection. + +# Acknowledgments + +This work was partially supported by NTU-NAP startup grant 024584-00001. + +# References + +[1] Corinne G Allaart, Björn Keyser, Henri Bal, and Aart Van Halteren. Vertical split learning—an exploration of predictive performance in medical and other use cases. In 2022 International Joint Conference on Neural Networks (IJCNN), pages 1-8. IEEE, 2022. 1, 2 +[2] Samyadeep Basu, Shell Hu, Daniela Massiceti, and Soheil Feizi. Strong baselines for parameter-efficient few-shot finetuning. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 11024-11031, 2024. 7, 1 +[3] Tom B Brown. Language models are few-shot learners. arXiv preprint arXiv:2005.14165, 2020. 1 +[4] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 2, 1 +[5] Shoufa Chen, Chongjian Ge, Zhan Tong, Jiangliu Wang, Yibing Song, Jue Wang, and Ping Luo. Adaptformer: Adapting vision transformers for scalable visual recognition. Advances in Neural Information Processing Systems, 35:16664-16678, 2022. 2 +[6] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 2 +[7] Yoni Choukroun, Eli Kravchik, Fan Yang, and Pavel Kisilev. Low-bit quantization of neural networks for efficient inference. In 2019 IEEE/CVF International Conference on Computer Vision Workshop (ICCVW), pages 3009-3018. IEEE, 2019. 5 +[8] Zeki Murat Çınar, Abubakar Abdussalam Nuhu, Qasim Zeešan, Orhan Korhan, Mohammed Asmael, and Babak Safaei. Machine learning in predictive maintenance towards sustainable smart manufacturing in industry 4.0. Sustainability, 12 (19):8211, 2020. 1 +[9] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 6 +[10] Matthew F Dixon, Igor Halperin, and Paul Bilokon. Machine learning in finance. Springer, 2020. 1 +[11] Jiahua Dong, Lixu Wang, Zhen Fang, Gan Sun, Shichao Xu, Xiao Wang, and Qi Zhu. Federated class-incremental learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10164–10173, 2022. 1 +[12] Wei Dong, Qiyao Luo, and Ke Yi. Continual observation under user-level differential privacy. In 2023 IEEE Symposium on Security and Privacy (SP), pages 2190-2207. IEEE, 2023. 1 +[13] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. ICLR, 2020. 1, 2, 3, 6 +[14] Juanru Fang, Wei Dong, and Ke Yi. Shifted inverse: A general mechanism for monotonic functions under user differential privacy. In Proceedings of the 2022 ACM SIGSAC Con- + +ference on Computer and Communications Security, pages 1009-1022, 2022. 1 +[15] Chongyang Gao, Lixu Wang, Chenkai Weng, Xiao Wang, and Qi Zhu. Practical unlearning for large language models. arXiv preprint arXiv:2407.10223, 2024. 1 +[16] Xinben Gao and Lan Zhang. {PCAT}: Functionality and data stealing from split learning by {Pseudo-Client} attack. In 32nd USENIX Security Symposium (USENIX Security 23), pages 5271-5288, 2023. 1, 2 +[17] Yansong Gao, Minki Kim, Sharif Abuadbba, Yeonjae Kim, Chandra Thapa, Kyuyeon Kim, Seyit A Camtep, Hyoungshick Kim, and Surya Nepal. End-to-end evaluation of federated learning and split learning for internet of things. In 2020 International Symposium on Reliable Distributed Systems (SRDS), pages 91-100. IEEE Computer Society, 2020. 2 +[18] Jiahui Gong, Jingtao Ding, Fanjin Meng, Guilong Chen, Hong Chen, Shen Zhao, Haisheng Lu, and Yong Li. A population-to-individual tuning framework for adapting pretrained lm to on-device user intent prediction. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 896-907, 2024. 2 +[19] Junfeng Guo, Yiming Li, Lixu Wang, Shu-Tao Xia, Heng Huang, Cong Liu, and Bo Li. Domain watermark: Effective and harmless dataset copyright protection is closed at hand. Advances in Neural Information Processing Systems, 36:54421-54450, 2023. 1 +[20] Song Han, Huizi Mao, and William J Dally. Deep compression: Compressing deep neural networks with pruning, trained quantization and huffman coding. *ICLR*, 2016. 4, 5 +[21] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 2 +[22] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 2 +[23] Zecheng He, Tianwei Zhang, and Ruby B Lee. Model inversion attacks against collaborative inference. In Proceedings of the 35th Annual Computer Security Applications Conference, pages 148-162, 2019. 1, 2 +[24] Zecheng He, Tianwei Zhang, and Ruby B Lee. Attacking and protecting data privacy in edge-cloud collaborative inference systems. IEEE Internet of Things Journal, 8(12):9706-9716, 2020. 6 +[25] Alain Hore and Djemel Ziou. Image quality metrics: Psnr vs. ssm. In 2010 20th international conference on pattern recognition, pages 2366-2369. IEEE, 2010. 7 +[26] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In European Conference on Computer Vision, pages 709-727. Springer, 2022. 2 +[27] Peter Kairouz, H Brendan McMahan, Brendan Avent, Aurélien Bellet, Mehdi Bennis, Arjun Nitin Bhagoji, Kallista + +Bonawitz, Zachary Charles, Graham Cormode, Rachel Cummings, et al. Advances and open problems in federated learning. Foundations and trends® in machine learning, 14(1-2): 1-210, 2021. 1 +[28] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 1, 2 +[29] Simon Kornblith, Jonathon Shlens, and Quoc V Le. Do better imagenet models transfer better? In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2661-2671, 2019. 2, 7, 1 +[30] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 6 +[31] Yuhang Li, Ruihao Gong, Xu Tan, Yang Yang, Peng Hu, Qi Zhang, Fengwei Yu, Wei Wang, and Shi Gu. Brecq: Pushing the limit of post-training quantization by block reconstruction. In International Conference on Learning Representations. 5 +[32] Chenxi Liu, Lixu Wang, Lingjuan Lyu, Chen Sun, Xiao Wang, and Qi Zhu. Deja vu: Continual model generalization for unseen domains. In The Eleventh International Conference on Learning Representations, 2023. 5 +[33] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 6, 1 +[34] Zhenhua Liu, Yunhe Wang, Kai Han, Wei Zhang, Siwei Ma, and Wen Gao. Post-training quantization for vision transformer. Advances in Neural Information Processing Systems, 34:28092-28103, 2021. 4, 5 +[35] Alejandro López-Cifuentes, Marcos Escudero-Vinolo, Jesús Bescós, and Álvaro García-Martín. Semantic-aware scene recognition. Pattern Recognition, 102:107256, 2020. 6 +[36] Payal Mohapatra, Lixu Wang, and Qi Zhu. Phase-driven domain generalizable learning for nonstationary time series. arXiv preprint arXiv:2402.05960, 2024. 4 +[37] Mehdi Noroozi and Paolo Favaro. Unsupervised learning of visual representations by solving jigsaw puzzles. In European conference on computer vision, pages 69-84. Springer, 2016. 2 +[38] Dario Pasquini, Giuseppe Ateniese, and Massimo Bernaschi. Unleashing the tiger: Inference attacks on split learning. In Proceedings of the 2021 ACM SIGSAC Conference on Computer and Communications Security, pages 2113-2129, 2021. 1, 2 +[39] Xingchao Peng, Qinxun Bai, Xide Xia, Zijun Huang, Kate Saenko, and Bo Wang. Moment matching for multi-source domain adaptation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1406-1415, 2019. 6 +[40] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervi + +sion. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 2 +[41] K Shailaja, Banoth Seetharamulu, and MA Jabbar. Machine learning in healthcare: A review. In 2018 Second international conference on electronics, communication and aerospace technology (ICECA), pages 910-914. IEEE, 2018. 1 +[42] Chandra Thapa, Mahawaga Arachchige Pathum Chamikara, and Seyit A Camtepe. Advancements of federated learning towards privacy preservation: from federated learning to split learning. Federated Learning Systems: Towards Next-Generation AI, pages 79–109, 2021. 2 +[43] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In International conference on machine learning, pages 10347-10357. PMLR, 2021. 6, 1 +[44] Praneeth Vepakomma, Otkrist Gupta, Tristan Swedish, and Ramesh Raskar. Split learning for health: Distributed deep learning without sharing raw patient data. arXiv preprint arXiv:1812.00564, 2018. 1, 2, 7 +[45] Jingye Wang, Ruoyi Du, Dongliang Chang, Kongming Liang, and Zhanyu Ma. Domain generalization via frequency-domain-based feature disentanglement and interaction. In Proceedings of the 30th ACM International Conference on Multimedia, pages 4821-4829, 2022. 4, 5 +[46] Lixu Wang, Shichao Xu, Xiao Wang, and Qi Zhu. Addressing class imbalance in federated learning. In Proceedings of the AAAI conference on artificial intelligence, pages 10165-10173, 2021. 1 +[47] Lixu Wang, Shichao Xu, Ruiqi Xu, Xiao Wang, and Qi Zhu. Non-transferable learning: A new approach for model ownership verification and applicability authorization. In International Conference on Learning Representation (ICLR-2022), 2022. 1 +[48] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 7 +[49] Guangxuan Xiao, Ji Lin, and Song Han. Offsite-tuning: Transfer learning without full model. arXiv preprint arXiv:2302.04870, 2023. 1, 2, 7 +[50] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. Advances in neural information processing systems, 34: 12077-12090, 2021. 1 +[51] Chengming Xu, Siqian Yang, Yabiao Wang, Zhanxiong Wang, Yanwei Fu, and Xiangyang Xue. Exploring efficient few-shot adaptation for vision transformers. Transactions on Machine Learning Research. 2 +[52] Peng Xu, Xiatian Zhu, and David A Clifton. Multimodal learning with transformers: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(10):12113-12132, 2023. 1 +[53] Xiaoyang Xu, Mengda Yang, Wenzhe Yi, Ziang Li, Juan Wang, Hongxin Hu, Yong Zhuang, and Yaxin Liu. A stealthy + +wrongdoer: Feature-oriented reconstruction attack against split learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12130-12139, 2024. 1, 2, 7, 8 +[54] Yupeng Yin, Xianglong Zhang, Huanle Zhang, Feng Li, Yue Yu, Xiuzhen Cheng, and Pengfei Hu. Ginver: Generative model inversion attacks against collaborative inference. In Proceedings of the ACM Web Conference 2023, pages 2122-2131, 2023. 1, 2 +[55] Zhihang Yuan, Chenhao Xue, Yiqi Chen, Qiang Wu, and Guangyu Sun. Ptq4vit: Post-training quantization for vision transformers with twin uniform quantization. In European conference on computer vision, pages 191-207. Springer, 2022. 5 +[56] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 7 +[57] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In Proceedings of the IEEE international conference on computer vision, pages 2223-2232, 2017. 2 \ No newline at end of file diff --git a/2025/Split Adaptation for Pre-trained Vision Transformers/images.zip b/2025/Split Adaptation for Pre-trained Vision Transformers/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..5ec4ee95ff785fdc65e13282a1d1bfcded804d60 --- /dev/null +++ b/2025/Split Adaptation for Pre-trained Vision Transformers/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dd90f5df385fb5d28ec2632b32fe879a939e3b2fa10d3029e64242bb76c08cd4 +size 580926 diff --git a/2025/Split Adaptation for Pre-trained Vision Transformers/layout.json b/2025/Split Adaptation for Pre-trained Vision Transformers/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..42efab5a4c26ad64c3e0b1a2a4ec97bfbb8adce6 --- /dev/null +++ b/2025/Split Adaptation for Pre-trained Vision Transformers/layout.json @@ -0,0 +1,10284 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 140, + 103, + 471, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 103, + 471, + 121 + ], + "spans": [ + { + "bbox": [ + 140, + 103, + 471, + 121 + ], + "type": "text", + "content": "Split Adaptation for Pre-trained Vision Transformers" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 82, + 142, + 527, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 142, + 527, + 172 + ], + "spans": [ + { + "bbox": [ + 82, + 142, + 527, + 172 + ], + "type": "text", + "content": "Lixu Wang" + }, + { + "bbox": [ + 82, + 142, + 527, + 172 + ], + "type": "inline_equation", + "content": "^{1,2*}" + }, + { + "bbox": [ + 82, + 142, + 527, + 172 + ], + "type": "text", + "content": ", Bingqi Shang" + }, + { + "bbox": [ + 82, + 142, + 527, + 172 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 82, + 142, + 527, + 172 + ], + "type": "text", + "content": ", Yi Li" + }, + { + "bbox": [ + 82, + 142, + 527, + 172 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 82, + 142, + 527, + 172 + ], + "type": "text", + "content": ", Payal Mohapatra" + }, + { + "bbox": [ + 82, + 142, + 527, + 172 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 82, + 142, + 527, + 172 + ], + "type": "text", + "content": ", Wei Dong" + }, + { + "bbox": [ + 82, + 142, + 527, + 172 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 82, + 142, + 527, + 172 + ], + "type": "text", + "content": ", Xiao Wang" + }, + { + "bbox": [ + 82, + 142, + 527, + 172 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 82, + 142, + 527, + 172 + ], + "type": "text", + "content": ", Qi Zhu" + }, + { + "bbox": [ + 82, + 142, + 527, + 172 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 82, + 142, + 527, + 172 + ], + "type": "text", + "content": ", Northwestern University. " + }, + { + "bbox": [ + 82, + 142, + 527, + 172 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 82, + 142, + 527, + 172 + ], + "type": "text", + "content": "Nanyang Technological University." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 64, + 174, + 544, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 174, + 544, + 199 + ], + "spans": [ + { + "bbox": [ + 64, + 174, + 544, + 199 + ], + "type": "text", + "content": "lixu.wang@ntu.edu.sg, {bingqishang2025, yili2023.1, payalmohapatra2026}@u.northwestern.edu, wei_dong@ntu.edu.sg, {wangxiao, qzhu}@northwestern.edu" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "spans": [ + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 252, + 296, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 252, + 296, + 612 + ], + "spans": [ + { + "bbox": [ + 56, + 252, + 296, + 612 + ], + "type": "text", + "content": "Vision Transformers (ViTs), extensively pre-trained on large-scale datasets, have become fundamental to foundation models, enabling adaptation to diverse downstream tasks. Existing adaptation methods typically require direct data access, rendering them infeasible in privacy-sensitive domains where clients are often reluctant to share their data. A straightforward solution may be sending the pretrained ViT to clients for local adaptation, which poses issues of model intellectual property and incurs heavy client computation overhead. To address these issues, we propose a novel split adaptation (SA) method that enables effective downstream adaptation while protecting data and models. SA, inspired by split learning (SL), segments the pre-trained ViT into a frontend and a backend, with only the frontend shared with the client for data representation extraction. But unlike regular SL, SA replaces frontend parameters with low-bit quantized values, preventing direct exposure of the model. SA allows the client to add bi-level noise to the frontend and the extracted data representations, ensuring data protection. Accordingly, SA incorporates data-level and model-level out-of-distribution enhancements to mitigate noise injection's impact. Our SA focuses on the challenging few-shot adaptation and adopts patch retrieval augmentation for overfitting alleviation. Extensive experiments on multiple datasets validate SA's superiority over state-of-the-art methods and demonstrate its defense against advanced data reconstruction attacks while preventing model leakage with minimal computation cost on the client side. The source codes can be found at https://github.com/conditionWang/Split_Adaptation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 636, + 135, + 649 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 636, + 135, + 649 + ], + "spans": [ + { + "bbox": [ + 56, + 636, + 135, + 649 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 657, + 296, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 657, + 296, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 657, + 296, + 693 + ], + "type": "text", + "content": "In recent years, pre-trained vision transformers (ViTs) [13] have become mainstream foundational models in computer vision, delivering remarkable performance across a range of" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 228, + 554, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 228, + 554, + 300 + ], + "spans": [ + { + "bbox": [ + 313, + 228, + 554, + 300 + ], + "type": "text", + "content": "visual tasks [13, 28, 50, 52]. ViTs have been equipped with general semantic knowledge during large-scale pre-training, which can be readily transferred to other domains including sensitive fields such as healthcare [41], finance [10], and manufacturing [8]. However, adapting pre-trained ViTs in these sensitive domains is extremely challenging." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 303, + 554, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 303, + 554, + 518 + ], + "spans": [ + { + "bbox": [ + 313, + 303, + 554, + 518 + ], + "type": "text", + "content": "In real-world scenarios, model adaptation often follows a two-party framework, where the pre-trained model is hosted on a server, and the training data resides on a client. In non-sensitive domains, training data can be directly shared with the server. Real-world examples include users uploading data to access OpenAI's CLIP model API [40] or tune GPT-3.5 [3]. However, in sensitive domains, data is often tightly linked to personal privacy [12, 14], intellectual property (IP) [19, 47], and ethical concerns [15], making any form of public exposure, sharing, or transfer strictly prohibited. In this context, an alternative solution might involve sending the model to the client for local adaptation, as in federated learning [11, 27, 46]. However, large pre-trained models are expensive to obtain, making their owners unlikely to share or transfer them freely [47]. Moreover, given that these pre-trained models consist of billions of parameters, adaptation on the client side would incur substantial computational and storage costs." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 521, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 554, + 713 + ], + "type": "text", + "content": "To our knowledge, there have been a few preliminary attempts to achieve downstream adaptation of pre-trained models while safeguarding data and models. Split learning (SL) [1, 44] divides the model into segments managed by different parties, which can partially protect the model from full exposure. Additionally, SL only communicates intermediate representations, avoiding the leakage of raw data. However, SL has critical drawbacks. On one hand, emerging malicious attacks, such as data reconstruction attacks (DRA) [16, 23, 38, 53, 54], can infer sensitive information from the intermediate representations. On the other hand, standard SL places the model training on the client side, imposing high computational costs. Beyond SL, a recent method called offsite tuning (OT) [49] has been proposed, which involves the server sending an emulator—a simplified version of the pre-trained model—to the client. The" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 702, + 137, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 702, + 137, + 713 + ], + "spans": [ + { + "bbox": [ + 66, + 702, + 137, + 713 + ], + "type": "text", + "content": "*Equal contributions." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20092" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 70, + 293, + 198 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 293, + 198 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 293, + 198 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 293, + 198 + ], + "type": "image", + "image_path": "3cdfb9127d5056f0a9058f7ff1ffba42ffa2f05ed3cb132d6bbc848741182b17.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 200, + 295, + 244 + ], + "lines": [ + { + "bbox": [ + 55, + 200, + 295, + 244 + ], + "spans": [ + { + "bbox": [ + 55, + 200, + 295, + 244 + ], + "type": "text", + "content": "Figure 1. The comparison between different downstream task adaptation approaches. Our SA can achieve effective few-shot task adaptation with minimal computation cost on the client side while protecting both the model and data." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 255, + 295, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 255, + 295, + 361 + ], + "spans": [ + { + "bbox": [ + 54, + 255, + 295, + 361 + ], + "type": "text", + "content": "client can then fine-tune this emulator using its data, typically employing parameter-efficient techniques like adapter tuning [5]. After training, the client uploads these adapters back to the server, where they are integrated into the pretrained model. However, while OT protects data, the emulator is fully exposed, allowing adversaries to create high-performance models easily. Moreover, OT's adaptation depends on the substantial computational and storage demands for adapter tuning on the client side." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 364, + 295, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 364, + 295, + 628 + ], + "spans": [ + { + "bbox": [ + 56, + 364, + 295, + 628 + ], + "type": "text", + "content": "To address the challenge of adapting pre-trained ViTs in sensitive domains, we propose a split adaptation (SA) method, which is shown in Figure 1. Like SL, SA divides the pre-trained ViT into two parts: a frontend and a backend. The server sends the frontend to the client, enabling data representation extraction. However, unlike SL, SA innovatively uses quantization to replace the frontend's parameters with low-bit values, preventing direct exposure of the original parameters. For data protection, the client injects random noise into the received frontend and its extracted data representations. Such injected noise unavoidably hurts the following adaptation, hence SA includes data-level and model-level out-of-distribution enhancements to improve the model generalization. SA is particularly designed for the challenging few-shot adaptation scenario, where it uses a patch retrieval augmentation technique to alleviate overfitting. Extensive experiments across four datasets demonstrate that SA significantly outperforms other state-of-the-art approaches in adaptation performance. Moreover, SA effectively defends against advanced DRAs and prevents model leakage. In particular, SA's effectiveness only relies on small computation cost on the client side. In summary:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 55, + 630, + 295, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 295, + 677 + ], + "type": "text", + "content": "- We focus on solving a practical and important problem—adapting pre-trained ViTs for downstream tasks while protecting data and models. Solving this problem enables wider application of pre-trained models in sensitive fields." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": "- We propose Split Adaptation (SA) that innovatively leverages model splitting and quantization to protect data and models during downstream task adaptation. SA focuses" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 322, + 72, + 553, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 72, + 553, + 96 + ], + "spans": [ + { + "bbox": [ + 322, + 72, + 553, + 96 + ], + "type": "text", + "content": "on challenging few-shot scenarios, in which SA uses a patch retrieval augmentation to combat overfitting." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 96, + 553, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 96, + 553, + 144 + ], + "spans": [ + { + "bbox": [ + 314, + 96, + 553, + 144 + ], + "type": "text", + "content": "- We conduct extensive experiments across multiple datasets and state-of-the-art baseline approaches, with results validating SA's effectiveness in achieving superior adaptation while ensuring data and model protection." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 156, + 406, + 169 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 156, + 406, + 169 + ], + "spans": [ + { + "bbox": [ + 314, + 156, + 406, + 169 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 177, + 405, + 190 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 177, + 405, + 190 + ], + "spans": [ + { + "bbox": [ + 314, + 177, + 405, + 190 + ], + "type": "text", + "content": "2.1. Split Learning" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 195, + 555, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 195, + 555, + 435 + ], + "spans": [ + { + "bbox": [ + 313, + 195, + 555, + 435 + ], + "type": "text", + "content": "Split learning (SL) [44] is a collaborative learning approach widely used in the multi-party computation framework. SL operates by dividing the machine learning (ML) model into segments, with each managed by a different party. This approach enhances privacy by transmitting only intermediate data representations rather than raw data. Many practical examples of SL in training and adapting ML models across various domains emerge, including healthcare [1, 44], financial [1, 42], and edge computing [17, 18]. Despite these successes, more adversarial attacks illustrate the vulnerability of privacy leakage in SL, with the severest one – data reconstruction attack (DRA) [23]. DRA aims to recover the raw data from its corresponding model representations. In the context of SL, a variety of DRAs have been witnessed, among them, some are based on optimization to imitate the ground-truth representations (e.g., FSHA [38] and PCAT [16]), while others leverage generative models to recover the raw data by exploiting the intrinsic semantics (e.g., Ginver [54] and FORA [53]). Therefore, new SL approaches are needed to help protect data privacy better." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 444, + 485, + 456 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 444, + 485, + 456 + ], + "spans": [ + { + "bbox": [ + 313, + 444, + 485, + 456 + ], + "type": "text", + "content": "2.2. Vision Transformer Adaptation" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 462, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 462, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 462, + 555, + 713 + ], + "type": "text", + "content": "The vision transformers (ViT) [13] are usually pre-trained with large-scale data to learn general semantics and knowledge. Representative pre-training approaches include supervised [28] and self-supervised learning like contrastive learning [4, 6, 21, 40] and auxiliary task training [22, 37, 57]. Owing to such extensive pre-training, adapting pretrained ViT for distinct downstream tasks has become a mainstream paradigm. The most regular method is to tune the entire model, which consumes substantial computational and memory resources. Another widely used approach is linear probing [29], which only trains a task module while keeping the ViT frozen, but the prominent shortcoming is the inferior performance. Recently, more parameter-efficient approaches have been proposed, like adapter tuning [5], prompt tuning [26], and prefix tuning [51]. However, these approaches still require direct access to the task data. In contrast to the above approaches, a new method called offsite tuning [49] can work effectively without direct data access. However, extensive training happens on the data owner side, and there is a high risk of model disclosure, making it impractical in real-world scenarios." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20093" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 71, + 138, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 71, + 138, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 138, + 85 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 141, + 102 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 141, + 102 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 141, + 102 + ], + "type": "text", + "content": "3.1. Preliminaries" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 109, + 167, + 120 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 109, + 167, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 109, + 167, + 120 + ], + "type": "text", + "content": "3.1.1. Vision Transformer" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 124, + 296, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 124, + 296, + 307 + ], + "spans": [ + { + "bbox": [ + 55, + 124, + 296, + 307 + ], + "type": "text", + "content": "The vision transformer (ViT) [13] is built upon sequential attention encoder layers. Specifically, each encoder layer consists of a multi-head self-attention (MSA) block and a Multi-Layer Perceptron (MLP) block. Layernorm (LN) is applied before every block and residual connections are placed after every block. The MLP contains two linear layers with an intermediate GELU activation. To handle 2D images, ViT reshapes the image " + }, + { + "bbox": [ + 55, + 124, + 296, + 307 + ], + "type": "inline_equation", + "content": "\\mathbf{I} \\in \\mathbb{R}^{H \\times W \\times C}" + }, + { + "bbox": [ + 55, + 124, + 296, + 307 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 55, + 124, + 296, + 307 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 124, + 296, + 307 + ], + "type": "text", + "content": " flattened patches " + }, + { + "bbox": [ + 55, + 124, + 296, + 307 + ], + "type": "inline_equation", + "content": "I^p \\in \\mathbb{R}^{P^2 \\cdot C}" + }, + { + "bbox": [ + 55, + 124, + 296, + 307 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 124, + 296, + 307 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 55, + 124, + 296, + 307 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 124, + 296, + 307 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 55, + 124, + 296, + 307 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 124, + 296, + 307 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 55, + 124, + 296, + 307 + ], + "type": "text", + "content": " are the image height, width, and channels, respectively, and " + }, + { + "bbox": [ + 55, + 124, + 296, + 307 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 55, + 124, + 296, + 307 + ], + "type": "text", + "content": " is the pixel resolution of each patch. For the first layer of ViT, to unify the vector dimensions, a trainable linear projection layer with neuron weights " + }, + { + "bbox": [ + 55, + 124, + 296, + 307 + ], + "type": "inline_equation", + "content": "[\\mathbf{W}_1^E, \\dots, \\mathbf{W}_N^E]" + }, + { + "bbox": [ + 55, + 124, + 296, + 307 + ], + "type": "text", + "content": " is used to map each vectorized patch into dimension " + }, + { + "bbox": [ + 55, + 124, + 296, + 307 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 55, + 124, + 296, + 307 + ], + "type": "text", + "content": ". In this case, the input to the first transformer layer is:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 81, + 314, + 296, + 337 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 314, + 296, + 337 + ], + "spans": [ + { + "bbox": [ + 81, + 314, + 296, + 337 + ], + "type": "interline_equation", + "content": "\\mathbf {X} _ {1} = \\left[ \\mathbf {x} ^ {\\mathrm {C L S}}; I _ {1} ^ {p} \\mathbf {W} _ {1} ^ {E}; \\dots ; I _ {N} ^ {p} \\mathbf {W} _ {N} ^ {E} \\right] + \\mathbf {E} ^ {\\mathrm {P O S}}, \\tag {1}", + "image_path": "d636dd91c074f351f61d38d1300a8b6d69c06a706b26412f83aa5c67f29c75dc.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 82, + 333, + 269, + 346 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 333, + 269, + 346 + ], + "spans": [ + { + "bbox": [ + 82, + 333, + 269, + 346 + ], + "type": "interline_equation", + "content": "w h e r e \\mathbf {W} ^ {E} \\in \\mathbb {R} ^ {(P ^ {2} \\cdot C) \\times d}, \\mathbf {E} ^ {\\mathrm {P O S}} \\in \\mathbb {R} ^ {(N + 1) \\times d},", + "image_path": "0fdf3faa63eb35987f730074f7cb9d99097ddcfb2bc2d8ba9af08695d768e5a7.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 357, + 296, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 357, + 296, + 443 + ], + "spans": [ + { + "bbox": [ + 55, + 357, + 296, + 443 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 357, + 296, + 443 + ], + "type": "inline_equation", + "content": "\\mathbf{x}^{\\mathrm{CLS}}\\in \\mathbb{R}^{d}" + }, + { + "bbox": [ + 55, + 357, + 296, + 443 + ], + "type": "text", + "content": " is a classification token, and " + }, + { + "bbox": [ + 55, + 357, + 296, + 443 + ], + "type": "inline_equation", + "content": "\\mathbf{E}^{\\mathrm{POS}}" + }, + { + "bbox": [ + 55, + 357, + 296, + 443 + ], + "type": "text", + "content": " is the position encoding. Then for each of the following layers, the input sequence " + }, + { + "bbox": [ + 55, + 357, + 296, + 443 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_l" + }, + { + "bbox": [ + 55, + 357, + 296, + 443 + ], + "type": "text", + "content": " is fed into each head of the MSA module to obtain three linearly projected matrices: Query " + }, + { + "bbox": [ + 55, + 357, + 296, + 443 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_l = \\mathbf{X}_l\\mathbf{W}_l^Q" + }, + { + "bbox": [ + 55, + 357, + 296, + 443 + ], + "type": "text", + "content": ", Key " + }, + { + "bbox": [ + 55, + 357, + 296, + 443 + ], + "type": "inline_equation", + "content": "\\mathbf{K}_l = \\mathbf{X}_l\\mathbf{W}_l^K" + }, + { + "bbox": [ + 55, + 357, + 296, + 443 + ], + "type": "text", + "content": ", and Value " + }, + { + "bbox": [ + 55, + 357, + 296, + 443 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_l = \\mathbf{X}_l\\mathbf{W}_l^V" + }, + { + "bbox": [ + 55, + 357, + 296, + 443 + ], + "type": "text", + "content": ". Then the Softmax function is applied to normalize the attention score " + }, + { + "bbox": [ + 55, + 357, + 296, + 443 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_l = \\mathbf{Q}_l\\mathbf{K}_l^\\top" + }, + { + "bbox": [ + 55, + 357, + 296, + 443 + ], + "type": "text", + "content": " and the output " + }, + { + "bbox": [ + 55, + 357, + 296, + 443 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}_l" + }, + { + "bbox": [ + 55, + 357, + 296, + 443 + ], + "type": "text", + "content": " is:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 91, + 453, + 296, + 480 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 453, + 296, + 480 + ], + "spans": [ + { + "bbox": [ + 91, + 453, + 296, + 480 + ], + "type": "interline_equation", + "content": "\\mathbf {Z} _ {l} = \\operatorname {M S A} \\left(\\mathbf {X} _ {l}\\right) = \\operatorname {S o f t m a x} \\left(\\frac {\\mathbf {A} _ {l}}{\\sqrt {d}}\\right) \\mathbf {V} _ {l}. \\tag {2}", + "image_path": "6b76e781842ecf098d9df794815fba32c1d5512a730a239fa90ebebf90cc8e26.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 487, + 295, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 487, + 295, + 513 + ], + "spans": [ + { + "bbox": [ + 55, + 487, + 295, + 513 + ], + "type": "text", + "content": "Suppose the two linear layers in MLP are parameterized by " + }, + { + "bbox": [ + 55, + 487, + 295, + 513 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_{l,1},\\mathbf{b}_{l,1}" + }, + { + "bbox": [ + 55, + 487, + 295, + 513 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 487, + 295, + 513 + ], + "type": "inline_equation", + "content": "\\mathbf{W}_{l,2},\\mathbf{b}_{l,2}" + }, + { + "bbox": [ + 55, + 487, + 295, + 513 + ], + "type": "text", + "content": ", the output is then computed as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 69, + 521, + 295, + 536 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 521, + 295, + 536 + ], + "spans": [ + { + "bbox": [ + 69, + 521, + 295, + 536 + ], + "type": "interline_equation", + "content": "\\operatorname {M L P} \\left(\\mathbf {Z} _ {l}\\right) = \\operatorname {G E L U} \\left(\\mathbf {Z} _ {l} \\mathbf {W} _ {l, 1} + \\boldsymbol {b} _ {l, 1}\\right) \\mathbf {W} _ {l, 2} + \\boldsymbol {b} _ {l, 2}. \\tag {3}", + "image_path": "ed61e944ca5fe697ae4df9fb1a9e6098e5b846390b06d21c49408784dd400970.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 545, + 296, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 296, + 594 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 296, + 594 + ], + "type": "text", + "content": "Finally, both the MSA and MLP modules are linked via residual connections, thus we have the final output of each transformer layer: " + }, + { + "bbox": [ + 55, + 545, + 296, + 594 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_{l + 1} = \\mathrm{LN}(\\mathbf{Z}_l + \\mathrm{MLP}(\\mathbf{Z}_l))\\in" + }, + { + "bbox": [ + 55, + 545, + 296, + 594 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{(N + 1)\\times d}" + }, + { + "bbox": [ + 55, + 545, + 296, + 594 + ], + "type": "text", + "content": " , where " + }, + { + "bbox": [ + 55, + 545, + 296, + 594 + ], + "type": "inline_equation", + "content": "\\mathbf{Z}_l = \\mathrm{LN}(\\mathbf{X}_l + \\mathrm{MSA}(\\mathbf{X}_l))\\in \\mathbb{R}^{(N + 1)\\times d}" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 601, + 176, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 601, + 176, + 612 + ], + "spans": [ + { + "bbox": [ + 55, + 601, + 176, + 612 + ], + "type": "text", + "content": "3.1.2. Problem Formulation" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 616, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 616, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 616, + 296, + 715 + ], + "type": "text", + "content": "When adapting pre-trained ViT for certain downstream tasks, we consider a representative client-server framework in which the client queries the server for the adaptation request. The client owns the task adaptation dataset " + }, + { + "bbox": [ + 55, + 616, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^{\\mathrm{C}} = \\{(x_i,y_i)\\}_{i = 1}^{N^{\\mathrm{C}}} (N^{\\mathrm{C}}" + }, + { + "bbox": [ + 55, + 616, + 296, + 715 + ], + "type": "text", + "content": " is the sample number, in practice, we suppose " + }, + { + "bbox": [ + 55, + 616, + 296, + 715 + ], + "type": "inline_equation", + "content": "N^{\\mathrm{C}}" + }, + { + "bbox": [ + 55, + 616, + 296, + 715 + ], + "type": "text", + "content": " is small even few-shot) with each sample drawn from a marginal input distribution " + }, + { + "bbox": [ + 55, + 616, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\pmb{x}_i\\sim \\mathcal{P}_X^{\\mathrm{C}}" + }, + { + "bbox": [ + 55, + 616, + 296, + 715 + ], + "type": "text", + "content": " and labeled with a marginal task distribution " + }, + { + "bbox": [ + 55, + 616, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{\\mathcal{Y}}^{\\mathrm{C}}" + }, + { + "bbox": [ + 55, + 616, + 296, + 715 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 72, + 555, + 196 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 196 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 196 + ], + "type": "text", + "content": "On the server side, the pre-trained ViT " + }, + { + "bbox": [ + 313, + 72, + 555, + 196 + ], + "type": "inline_equation", + "content": "F_{\\Theta}" + }, + { + "bbox": [ + 313, + 72, + 555, + 196 + ], + "type": "text", + "content": " parameterized on " + }, + { + "bbox": [ + 313, + 72, + 555, + 196 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 313, + 72, + 555, + 196 + ], + "type": "text", + "content": " is assumed to consist of " + }, + { + "bbox": [ + 313, + 72, + 555, + 196 + ], + "type": "inline_equation", + "content": "L^F" + }, + { + "bbox": [ + 313, + 72, + 555, + 196 + ], + "type": "text", + "content": " attention layers, and each layer is parameterized on " + }, + { + "bbox": [ + 313, + 72, + 555, + 196 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 72, + 555, + 196 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 313, + 72, + 555, + 196 + ], + "type": "inline_equation", + "content": "F := [f_{\\theta_1}, \\dots, f_{\\theta_{L^F}}]" + }, + { + "bbox": [ + 313, + 72, + 555, + 196 + ], + "type": "text", + "content": ". Besides, we assume the server owns another dataset " + }, + { + "bbox": [ + 313, + 72, + 555, + 196 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^{\\mathrm{S}} = \\{(x_i, y_i)\\}_{i=1}^{N^{\\mathrm{S}}}" + }, + { + "bbox": [ + 313, + 72, + 555, + 196 + ], + "type": "text", + "content": " collected from public domains with the marginal input and task distributions, " + }, + { + "bbox": [ + 313, + 72, + 555, + 196 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{\\mathcal{X}}^{\\mathrm{S}}" + }, + { + "bbox": [ + 313, + 72, + 555, + 196 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 72, + 555, + 196 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{\\mathcal{Y}}^{\\mathrm{S}}" + }, + { + "bbox": [ + 313, + 72, + 555, + 196 + ], + "type": "text", + "content": ", respectively. In real-world scenarios, the datasets owned by the client and the server are distinct but relevant, i.e., " + }, + { + "bbox": [ + 313, + 72, + 555, + 196 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{\\mathcal{X}}^{\\mathrm{C}} \\neq \\mathcal{P}_{\\mathcal{X}}^{\\mathrm{S}}, \\mathcal{P}_{\\mathcal{Y}}^{\\mathrm{C}} \\neq \\mathcal{P}_{\\mathcal{Y}}^{\\mathrm{S}}" + }, + { + "bbox": [ + 313, + 72, + 555, + 196 + ], + "type": "text", + "content": " but " + }, + { + "bbox": [ + 313, + 72, + 555, + 196 + ], + "type": "inline_equation", + "content": "I(\\mathcal{X}^{\\mathrm{C}}, \\mathcal{X}^{\\mathrm{S}}) > 0" + }, + { + "bbox": [ + 313, + 72, + 555, + 196 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 72, + 555, + 196 + ], + "type": "inline_equation", + "content": "I(\\mathcal{Y}^{\\mathrm{C}}, \\mathcal{Y}^{\\mathrm{S}}) > 0" + }, + { + "bbox": [ + 313, + 72, + 555, + 196 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 313, + 72, + 555, + 196 + ], + "type": "inline_equation", + "content": "I(\\cdot; \\cdot)" + }, + { + "bbox": [ + 313, + 72, + 555, + 196 + ], + "type": "text", + "content": " calculates the mutual information." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 201, + 555, + 286 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 201, + 555, + 286 + ], + "spans": [ + { + "bbox": [ + 313, + 201, + 555, + 286 + ], + "type": "text", + "content": "Objective of Task Performance. Pre-trained ViT adaptation aims to obtain a downstream task module " + }, + { + "bbox": [ + 313, + 201, + 555, + 286 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 313, + 201, + 555, + 286 + ], + "type": "text", + "content": " parameterized on " + }, + { + "bbox": [ + 313, + 201, + 555, + 286 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 313, + 201, + 555, + 286 + ], + "type": "text", + "content": " that can achieve as good performance as possible on the client data distribution when combined with " + }, + { + "bbox": [ + 313, + 201, + 555, + 286 + ], + "type": "inline_equation", + "content": "F_{\\Theta}" + }, + { + "bbox": [ + 313, + 201, + 555, + 286 + ], + "type": "text", + "content": ". Certainly, such adaptation also allows " + }, + { + "bbox": [ + 313, + 201, + 555, + 286 + ], + "type": "inline_equation", + "content": "F_{\\Theta}" + }, + { + "bbox": [ + 313, + 201, + 555, + 286 + ], + "type": "text", + "content": " tuning if it can enable performance improvement on the downstream task. In this case, the objective is formulated as," + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 341, + 298, + 555, + 316 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 341, + 298, + 555, + 316 + ], + "spans": [ + { + "bbox": [ + 341, + 298, + 555, + 316 + ], + "type": "interline_equation", + "content": "\\Theta , \\gamma = \\arg \\min _ {\\Theta , \\gamma} \\mathbb {E} _ {\\boldsymbol {x} \\sim \\mathcal {P} _ {\\chi} ^ {\\mathrm {C}}, y \\sim \\mathcal {P} _ {\\mathcal {Y}} ^ {\\mathrm {C}}} \\mathcal {L} (\\boldsymbol {x}, y; \\Theta \\circ \\gamma), \\tag {4}", + "image_path": "67d61f8712333683f63826da3e75a0fdf9fe4ebf274292111702356347286897.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 328, + 554, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 328, + 554, + 354 + ], + "spans": [ + { + "bbox": [ + 313, + 328, + 554, + 354 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 328, + 554, + 354 + ], + "type": "inline_equation", + "content": "\\mathcal{L}(\\cdot ;\\cdot)" + }, + { + "bbox": [ + 313, + 328, + 554, + 354 + ], + "type": "text", + "content": " is a loss function that measures the model performance regarding certain data distribution." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 358, + 555, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 358, + 555, + 479 + ], + "spans": [ + { + "bbox": [ + 313, + 358, + 555, + 479 + ], + "type": "text", + "content": "Objective of Privacy Protection. In addition to the above standard objective, the intrinsic privacy concerns of the server-client framework should be addressed, i.e., the pretrained model and client data should be protected during adaptation when supposing the server and the client may be semi-honest. On the one hand, the adaptation should prohibit the client from stealing surrogate models with good performance or obtaining them at a small cost, on the other hand, any client data reconstruction on the server side should be avoided." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 491, + 541, + 504 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 491, + 541, + 504 + ], + "spans": [ + { + "bbox": [ + 313, + 491, + 541, + 504 + ], + "type": "text", + "content": "3.2. Overview of The Proposed Split Adaptation" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 510, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 555, + 715 + ], + "type": "text", + "content": "To achieve effective adaptation while protecting the pre-trained ViT and client data, we propose a method called Split Adaptation (SA), shown in Figure 2. Our SA starts with dividing the pre-trained ViT " + }, + { + "bbox": [ + 313, + 510, + 555, + 715 + ], + "type": "inline_equation", + "content": "F_{\\Theta}" + }, + { + "bbox": [ + 313, + 510, + 555, + 715 + ], + "type": "text", + "content": " into two parts: a frontend " + }, + { + "bbox": [ + 313, + 510, + 555, + 715 + ], + "type": "inline_equation", + "content": "F^{\\mathrm{F}}" + }, + { + "bbox": [ + 313, + 510, + 555, + 715 + ], + "type": "text", + "content": " and a backend " + }, + { + "bbox": [ + 313, + 510, + 555, + 715 + ], + "type": "inline_equation", + "content": "F^{\\mathrm{B}}" + }, + { + "bbox": [ + 313, + 510, + 555, + 715 + ], + "type": "text", + "content": ". The original " + }, + { + "bbox": [ + 313, + 510, + 555, + 715 + ], + "type": "inline_equation", + "content": "F^{\\mathrm{F}}" + }, + { + "bbox": [ + 313, + 510, + 555, + 715 + ], + "type": "text", + "content": " is not sent to the client, instead, we leverage model quantization to send a quantized version " + }, + { + "bbox": [ + 313, + 510, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\widehat{F}^{\\mathrm{F}}" + }, + { + "bbox": [ + 313, + 510, + 555, + 715 + ], + "type": "text", + "content": " to the client. To mitigate the influence of frontend quantization, SA includes out-of-distribution (OOD) augmentations from both data and model levels to enhance the cross-distribution generalization of the frontend and backend. As for the client side, it adds random noise to the received " + }, + { + "bbox": [ + 313, + 510, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\widehat{F}^{\\mathrm{F}}" + }, + { + "bbox": [ + 313, + 510, + 555, + 715 + ], + "type": "text", + "content": " first and adopts a patch-level retrieve-then-replace strategy to generate more client data representations that help alleviate overfitting in few-shot adaptation. Before being uploaded to the server for the final adaptation, these augmented representations are protected by adding noise again." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20094" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 70, + 72, + 541, + 304 + ], + "blocks": [ + { + "bbox": [ + 70, + 72, + 541, + 304 + ], + "lines": [ + { + "bbox": [ + 70, + 72, + 541, + 304 + ], + "spans": [ + { + "bbox": [ + 70, + 72, + 541, + 304 + ], + "type": "image", + "image_path": "2fe7a701a2a45e6b89d7ae2ea5783a2e1e33ee6e9991aaf14b5e56e87849fd02.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 305, + 555, + 361 + ], + "lines": [ + { + "bbox": [ + 54, + 305, + 555, + 361 + ], + "spans": [ + { + "bbox": [ + 54, + 305, + 555, + 361 + ], + "type": "text", + "content": "Figure 2. Overview of Split Adaptation (SA) for pre-trained ViT adaptation. SA divides the pre-trained ViT into a frontend and a backend. After applying Out-of-distribution Enhanced Quantization to the frontend, its quantized version is sent to the client. To mitigate the impact of quantization, SA adopts Out-of-distribution Quantization-aware Tuning to enhance backend's generalization. As for the client, it injects random noise to the received frontend and retrieves then replaces randomly selected patches to augment more client data representations, which are sent to the server for the final adaptation after being added with noise again." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 371, + 243, + 384 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 371, + 243, + 384 + ], + "spans": [ + { + "bbox": [ + 55, + 371, + 243, + 384 + ], + "type": "text", + "content": "3.3. Model Splitting on The Server Side" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 389, + 296, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 389, + 296, + 629 + ], + "spans": [ + { + "bbox": [ + 55, + 389, + 296, + 629 + ], + "type": "text", + "content": "At the beginning of SA, the pre-trained ViT is divided into a frontend " + }, + { + "bbox": [ + 55, + 389, + 296, + 629 + ], + "type": "inline_equation", + "content": "F^{\\mathrm{F}} \\coloneqq [f_{\\theta_1}, \\dots, f_{\\theta_K}]" + }, + { + "bbox": [ + 55, + 389, + 296, + 629 + ], + "type": "text", + "content": " and a backend " + }, + { + "bbox": [ + 55, + 389, + 296, + 629 + ], + "type": "inline_equation", + "content": "F^{\\mathrm{B}} \\coloneqq [f_{\\theta_{K+1}}, \\dots, f_{\\theta_{LF}}]" + }, + { + "bbox": [ + 55, + 389, + 296, + 629 + ], + "type": "text", + "content": " at the " + }, + { + "bbox": [ + 55, + 389, + 296, + 629 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 389, + 296, + 629 + ], + "type": "text", + "content": "-th layer (" + }, + { + "bbox": [ + 55, + 389, + 296, + 629 + ], + "type": "inline_equation", + "content": "K > L^{F} / 2" + }, + { + "bbox": [ + 55, + 389, + 296, + 629 + ], + "type": "text", + "content": "). Then the frontend is distributed to the client for representation extraction. Although we empirically found that the frontend cannot perform well (Section 4.3), we still hope to protect it from disclosing the original model parameters. Quantization [20, 34], as a representative model compression approach, can significantly reduce the computational and memory cost without sacrificing model performance. As quantization works by representing model weights and activations with low-precision values, it can potentially protect the model. However, standard quantization needs a certain amount of data drawn from the target task, which is inaccessible for the server in our problem. In this case, the left choice is to leverage the server dataset " + }, + { + "bbox": [ + 55, + 389, + 296, + 629 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^{\\mathrm{S}}" + }, + { + "bbox": [ + 55, + 389, + 296, + 629 + ], + "type": "text", + "content": " to quantize the model, but a dedicated design for enhancing the quantized model's cross-distribution generalization is needed. Therefore, we propose an augmentation technique to generate OOD data in terms of " + }, + { + "bbox": [ + 55, + 389, + 296, + 629 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^{\\mathrm{S}}" + }, + { + "bbox": [ + 55, + 389, + 296, + 629 + ], + "type": "text", + "content": " for quantization." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 638, + 270, + 651 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 638, + 270, + 651 + ], + "spans": [ + { + "bbox": [ + 55, + 638, + 270, + 651 + ], + "type": "text", + "content": "3.3.1. Out-of-distribution Enhanced Quantization" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "content": "In practice, we have to consider extreme cases in which the client dataset differs significantly from the server dataset and they may only share general semantics. Hence our augmentation technique should generate OOD data with a large discrepancy to " + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^{\\mathrm{S}}" + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "content": ". Inspired by signal processing, we no" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 317, + 372, + 552, + 518 + ], + "blocks": [ + { + "bbox": [ + 317, + 372, + 552, + 518 + ], + "lines": [ + { + "bbox": [ + 317, + 372, + 552, + 518 + ], + "spans": [ + { + "bbox": [ + 317, + 372, + 552, + 518 + ], + "type": "image", + "image_path": "33811de070aa6290bb3a984150073cb40b73014f565a79dd50816fc520b6808e.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 518, + 555, + 573 + ], + "lines": [ + { + "bbox": [ + 313, + 518, + 555, + 573 + ], + "spans": [ + { + "bbox": [ + 313, + 518, + 555, + 573 + ], + "type": "text", + "content": "Figure 3. Visualization of our Hilbert Transform data augmentation. Compared to amplitude-exchange augmentation [45] (bottom row), our augmentation method (top row) generates data that diverges further from the original in appearance while preserving the original semantics (dog category)." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "text", + "content": "tice that there is a technique called Hilbert Transform (HT), which only captures the most sharp edge in images. Intuitively, the sharp edges contain the most general semantics, making HT suitable for our problem. Applying HT to a signal in standard signal processing results in a phase shift of " + }, + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\pi /2" + }, + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "text", + "content": " in the frequency domain [36]. In the visual domain, we can implement HT by transforming an image " + }, + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{I}" + }, + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "text", + "content": " to the frequency domain with the spectrum index " + }, + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\xi" + }, + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "text", + "content": " via Fourier transform, then the negative " + }, + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\xi" + }, + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "text", + "content": " needs to multiply with the imaginary unit " + }, + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "text", + "content": ", while the positive " + }, + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\xi" + }, + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "text", + "content": " needs to multiple with" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20095" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 89, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 89, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 89, + 83 + ], + "type": "inline_equation", + "content": "-i" + }, + { + "bbox": [ + 56, + 73, + 89, + 83 + ], + "type": "text", + "content": " , i.e.," + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 105, + 90, + 295, + 105 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 90, + 295, + 105 + ], + "spans": [ + { + "bbox": [ + 105, + 90, + 295, + 105 + ], + "type": "interline_equation", + "content": "\\mathrm {H T} (\\mathbf {I}) = \\mathcal {F} ^ {- 1} \\{- i \\cdot \\operatorname {s g n} (\\xi) \\mathcal {F} (\\mathbf {I}) \\}, \\qquad (5)", + "image_path": "1e006e584560a487c533c8d3a0ba8947ebd169dbf528d4275c1125f6e3673426.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 109, + 296, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 109, + 296, + 205 + ], + "spans": [ + { + "bbox": [ + 55, + 109, + 296, + 205 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 109, + 296, + 205 + ], + "type": "inline_equation", + "content": "\\mathcal{F}(\\cdot)" + }, + { + "bbox": [ + 55, + 109, + 296, + 205 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 109, + 296, + 205 + ], + "type": "inline_equation", + "content": "\\mathcal{F}^{-1}(\\cdot)" + }, + { + "bbox": [ + 55, + 109, + 296, + 205 + ], + "type": "text", + "content": " are Fourier transform and inverse, respectively, and " + }, + { + "bbox": [ + 55, + 109, + 296, + 205 + ], + "type": "inline_equation", + "content": "\\mathrm{sgn}(\\cdot)" + }, + { + "bbox": [ + 55, + 109, + 296, + 205 + ], + "type": "text", + "content": " is a sign function. However, pure HT only provides new images with black backgrounds. To diversify the HT-generated image, we replace its amplitude in the frequency domain with that of the original image. In this way, the generated image is much better than the standard amplitude-exchange augmentation [32, 45] (please refer to Figure 3 for a visual comparison)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 206, + 296, + 254 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 206, + 296, + 254 + ], + "spans": [ + { + "bbox": [ + 55, + 206, + 296, + 254 + ], + "type": "text", + "content": "To achieve effective quantization, we adopt the most popular uniform symmetric quantization [20] to implement SA, in which a floating-point value " + }, + { + "bbox": [ + 55, + 206, + 296, + 254 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 55, + 206, + 296, + 254 + ], + "type": "text", + "content": " is projected into a " + }, + { + "bbox": [ + 55, + 206, + 296, + 254 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 206, + 296, + 254 + ], + "type": "text", + "content": "-bit integer value " + }, + { + "bbox": [ + 55, + 206, + 296, + 254 + ], + "type": "inline_equation", + "content": "w_{\\mathrm{q}}" + }, + { + "bbox": [ + 55, + 206, + 296, + 254 + ], + "type": "text", + "content": " with a scaling factor " + }, + { + "bbox": [ + 55, + 206, + 296, + 254 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 55, + 206, + 296, + 254 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 59, + 259, + 296, + 289 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 259, + 296, + 289 + ], + "spans": [ + { + "bbox": [ + 59, + 259, + 296, + 289 + ], + "type": "interline_equation", + "content": "\\widehat {w} = \\Psi_ {k} (w, \\Delta) = \\operatorname {c l a m p} (\\operatorname {r o u n d} \\left(\\frac {w}{\\Delta}\\right), - 2 ^ {k - 1}, 2 ^ {k - 1} - 1), \\tag {6}", + "image_path": "e8272b9b663c22d397cf9120325cf054a2315bdf5a384dc05d9b7f67facf1515.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 289, + 296, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 289, + 296, + 496 + ], + "spans": [ + { + "bbox": [ + 55, + 289, + 296, + 496 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 289, + 296, + 496 + ], + "type": "inline_equation", + "content": "\\mathrm{round}(\\cdot)" + }, + { + "bbox": [ + 55, + 289, + 296, + 496 + ], + "type": "text", + "content": " turns a value into an integer and " + }, + { + "bbox": [ + 55, + 289, + 296, + 496 + ], + "type": "inline_equation", + "content": "\\mathrm{clamp}(\\cdot, \\min, \\max)" + }, + { + "bbox": [ + 55, + 289, + 296, + 496 + ], + "type": "text", + "content": " constrains the output in the range from " + }, + { + "bbox": [ + 55, + 289, + 296, + 496 + ], + "type": "inline_equation", + "content": "\\min" + }, + { + "bbox": [ + 55, + 289, + 296, + 496 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 289, + 296, + 496 + ], + "type": "inline_equation", + "content": "\\max" + }, + { + "bbox": [ + 55, + 289, + 296, + 496 + ], + "type": "text", + "content": ". Intuitively, the scaling factor " + }, + { + "bbox": [ + 55, + 289, + 296, + 496 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 55, + 289, + 296, + 496 + ], + "type": "text", + "content": " directly impacts the model performance after quantization. Previous studies [7, 34] have tried a variety of measurements to quantify the model output changes and accordingly determine the optimal scaling factors. Among these studies, we follow Yuan et al. [55] to choose a Hessian-guided metric for scaling factor determination. Suppose the pre-trained ViT " + }, + { + "bbox": [ + 55, + 289, + 296, + 496 + ], + "type": "inline_equation", + "content": "F_{\\Theta}" + }, + { + "bbox": [ + 55, + 289, + 296, + 496 + ], + "type": "text", + "content": " is equipped with a task module " + }, + { + "bbox": [ + 55, + 289, + 296, + 496 + ], + "type": "inline_equation", + "content": "g_{\\gamma'}'" + }, + { + "bbox": [ + 55, + 289, + 296, + 496 + ], + "type": "text", + "content": " that is tuned on the server dataset " + }, + { + "bbox": [ + 55, + 289, + 296, + 496 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^{\\mathrm{S}}" + }, + { + "bbox": [ + 55, + 289, + 296, + 496 + ], + "type": "text", + "content": ", the model performance can be measured by a loss " + }, + { + "bbox": [ + 55, + 289, + 296, + 496 + ], + "type": "inline_equation", + "content": "\\mathbb{E}_{\\boldsymbol{x} \\sim \\mathcal{P}_{\\boldsymbol{x}}^{\\mathrm{S}}, y \\sim \\mathcal{P}_{y}^{\\mathrm{S}}} \\mathcal{L}(\\boldsymbol{x}, y; \\Theta \\circ \\gamma')" + }, + { + "bbox": [ + 55, + 289, + 296, + 496 + ], + "type": "text", + "content": " (we omit " + }, + { + "bbox": [ + 55, + 289, + 296, + 496 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}" + }, + { + "bbox": [ + 55, + 289, + 296, + 496 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 289, + 296, + 496 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 55, + 289, + 296, + 496 + ], + "type": "text", + "content": " to simplify the used symbols in the following descriptions). Then if we regard that the quantization brings a small perturbation " + }, + { + "bbox": [ + 55, + 289, + 296, + 496 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 55, + 289, + 296, + 496 + ], + "type": "text", + "content": " to the model parameters, i.e., " + }, + { + "bbox": [ + 55, + 289, + 296, + 496 + ], + "type": "inline_equation", + "content": "\\hat{\\Theta} = \\Theta + \\epsilon" + }, + { + "bbox": [ + 55, + 289, + 296, + 496 + ], + "type": "text", + "content": ", we can analyze the influence of quantization on the task loss using Taylor series expansion:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 63, + 501, + 295, + 525 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 501, + 295, + 525 + ], + "spans": [ + { + "bbox": [ + 63, + 501, + 295, + 525 + ], + "type": "interline_equation", + "content": "\\mathbb {E} [ \\mathcal {L} (\\widehat {\\Theta} \\circ \\gamma^ {\\prime}) ] - \\mathbb {E} [ \\mathcal {L} (\\Theta \\circ \\gamma^ {\\prime}) ] \\approx \\epsilon^ {\\top} \\nabla_ {\\Theta} + \\frac {1}{2} \\epsilon^ {\\top} \\mathbf {H} _ {\\Theta} \\epsilon , (7)", + "image_path": "6b958e63b93d68680842117daf7a22a9a1b8916b9f7f30ce8234f7ebe19834c9.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 528, + 296, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 528, + 296, + 636 + ], + "spans": [ + { + "bbox": [ + 55, + 528, + 296, + 636 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 528, + 296, + 636 + ], + "type": "inline_equation", + "content": "\\nabla_{\\Theta}" + }, + { + "bbox": [ + 55, + 528, + 296, + 636 + ], + "type": "text", + "content": " is the gradients and " + }, + { + "bbox": [ + 55, + 528, + 296, + 636 + ], + "type": "inline_equation", + "content": "\\mathbf{H}_{\\Theta}" + }, + { + "bbox": [ + 55, + 528, + 296, + 636 + ], + "type": "text", + "content": " is the Hessian matrix. Then the optimal scaling factors are the ones corresponding to minimal influence caused by quantization, i.e., " + }, + { + "bbox": [ + 55, + 528, + 296, + 636 + ], + "type": "inline_equation", + "content": "\\Delta = \\arg \\min_{\\Delta}(\\mathbb{E}[\\mathcal{L}(\\widehat{\\Theta}\\circ \\gamma^{\\prime})] - \\mathbb{E}[\\mathcal{L}(\\Theta \\circ \\gamma^{\\prime})])" + }, + { + "bbox": [ + 55, + 528, + 296, + 636 + ], + "type": "text", + "content": ". Certainly, determining and adopting a general scaling factor for the entire model is too coarse to perform well, thus we apply a layer-wise reconstruction method [31] to achieve fine-grained quantization, then the searching optimization of Eq. (7) can be approximated by:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 640, + 299, + 676 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 640, + 299, + 676 + ], + "spans": [ + { + "bbox": [ + 55, + 640, + 299, + 676 + ], + "type": "interline_equation", + "content": "\\underset {\\Delta_ {l - 1}} {\\arg \\min } \\mathbb {E} \\left[ \\left(\\widehat {\\mathbf {X}} _ {l} - \\mathbf {X} _ {l}\\right) ^ {\\top} \\operatorname {d i a g} \\left(\\frac {\\partial \\mathcal {L}}{\\partial \\mathbf {X} _ {l}} \\odot \\frac {\\partial \\mathcal {L}}{\\partial \\mathbf {X} _ {l}}\\right) \\left(\\widehat {\\mathbf {X}} _ {l} - \\mathbf {X} _ {l}\\right) \\right], \\tag {8}", + "image_path": "23afcd7cd558ca2bbdfbf9bb08ce64497d3311a6f53977c836ee3cc7da5f25c4.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_l" + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\widehat{\\mathbf{X}}_l" + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": " are the outputs of the " + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "inline_equation", + "content": "(l-1)" + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "-th layer before and after quantization, respectively, and " + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": " represents the Hadamard product of a matrix." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 555, + 169 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 169 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 169 + ], + "type": "text", + "content": "With the augmented neighborhood data, the frontend can be quantized with improved cross-distribution generalization. Specifically, we first conduct HT-based data augmentation on the server dataset " + }, + { + "bbox": [ + 313, + 72, + 555, + 169 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^{\\mathrm{S}}" + }, + { + "bbox": [ + 313, + 72, + 555, + 169 + ], + "type": "text", + "content": " to obtain " + }, + { + "bbox": [ + 313, + 72, + 555, + 169 + ], + "type": "inline_equation", + "content": "\\mathrm{HT}(\\mathcal{D}^{\\mathrm{S}})" + }, + { + "bbox": [ + 313, + 72, + 555, + 169 + ], + "type": "text", + "content": ". Then the optimal scaling factor for each layer in the frontend " + }, + { + "bbox": [ + 313, + 72, + 555, + 169 + ], + "type": "inline_equation", + "content": "F^{\\mathrm{F}}" + }, + { + "bbox": [ + 313, + 72, + 555, + 169 + ], + "type": "text", + "content": " is empirically determined by optimizing Eq. (8) on the merged dataset " + }, + { + "bbox": [ + 313, + 72, + 555, + 169 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\cup}^{\\mathrm{S}} = \\mathcal{D}^{\\mathrm{S}} \\cup \\mathrm{HT}(\\mathcal{D}^{\\mathrm{S}})" + }, + { + "bbox": [ + 313, + 72, + 555, + 169 + ], + "type": "text", + "content": " and we can obtain a quantized frontend " + }, + { + "bbox": [ + 313, + 72, + 555, + 169 + ], + "type": "inline_equation", + "content": "\\widehat{F}^{\\mathrm{F}}" + }, + { + "bbox": [ + 313, + 72, + 555, + 169 + ], + "type": "text", + "content": " in the end." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 174, + 545, + 186 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 174, + 545, + 186 + ], + "spans": [ + { + "bbox": [ + 313, + 174, + 545, + 186 + ], + "type": "text", + "content": "3.3.2. Out-of-distribution Quantization-aware Tuning" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 189, + 554, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 189, + 554, + 272 + ], + "spans": [ + { + "bbox": [ + 313, + 189, + 554, + 272 + ], + "type": "text", + "content": "To achieve better adaptation performance, the backend also needs to resist the unavoidable parameter difference between the original and quantized model as well as the noise injected by the client that we will introduce in Section 3.4. As a result, we propose an OOD quantization-aware tuning technique to enhance the backend's resistance when encountering OOD data and frontends." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 273, + 555, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 273, + 555, + 392 + ], + "spans": [ + { + "bbox": [ + 313, + 273, + 555, + 392 + ], + "type": "text", + "content": "To simulate potential OOD frontends, we first divide the merged server dataset " + }, + { + "bbox": [ + 313, + 273, + 555, + 392 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\cup}^{\\mathrm{S}}" + }, + { + "bbox": [ + 313, + 273, + 555, + 392 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 313, + 273, + 555, + 392 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 273, + 555, + 392 + ], + "type": "text", + "content": " subsets and denote these subsets as " + }, + { + "bbox": [ + 313, + 273, + 555, + 392 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{D}_{\\cup_m}^{\\mathrm{S}}\\}_{m = 1}^M" + }, + { + "bbox": [ + 313, + 273, + 555, + 392 + ], + "type": "text", + "content": ". Next, each subset " + }, + { + "bbox": [ + 313, + 273, + 555, + 392 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\cup_m}^{\\mathrm{S}}" + }, + { + "bbox": [ + 313, + 273, + 555, + 392 + ], + "type": "text", + "content": " is used to quantize the original frontend " + }, + { + "bbox": [ + 313, + 273, + 555, + 392 + ], + "type": "inline_equation", + "content": "F^{\\mathrm{F}}" + }, + { + "bbox": [ + 313, + 273, + 555, + 392 + ], + "type": "text", + "content": " to obtain a quantized version " + }, + { + "bbox": [ + 313, + 273, + 555, + 392 + ], + "type": "inline_equation", + "content": "\\widehat{F}_m^{\\mathrm{F}}" + }, + { + "bbox": [ + 313, + 273, + 555, + 392 + ], + "type": "text", + "content": ". Then let us use a single quantized frontend " + }, + { + "bbox": [ + 313, + 273, + 555, + 392 + ], + "type": "inline_equation", + "content": "\\widehat{F}_m^{\\mathrm{F}}" + }, + { + "bbox": [ + 313, + 273, + 555, + 392 + ], + "type": "text", + "content": " to introduce how to tune the backend. To begin with, since the backend is represented and the tuning is performed using floating-point values, the data representations extracted by the quantized frontend should be transformed back to floating-point values through the dequantization:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 379, + 398, + 555, + 412 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 379, + 398, + 555, + 412 + ], + "spans": [ + { + "bbox": [ + 379, + 398, + 555, + 412 + ], + "type": "interline_equation", + "content": "\\tilde {w} = \\Psi^ {- 1} (\\hat {w}, \\Delta) = \\hat {w} \\cdot \\Delta . \\tag {9}", + "image_path": "d2e0b316aa12a422a52c893f7658efd15aeec1c4c3c1bd1f00e791071d665610.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 418, + 555, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 418, + 555, + 613 + ], + "spans": [ + { + "bbox": [ + 313, + 418, + 555, + 613 + ], + "type": "text", + "content": "In this way, we can obtain the floating-point representation " + }, + { + "bbox": [ + 313, + 418, + 555, + 613 + ], + "type": "inline_equation", + "content": "\\widetilde{\\mathbf{X}}_K" + }, + { + "bbox": [ + 313, + 418, + 555, + 613 + ], + "type": "text", + "content": " by dequantizing the " + }, + { + "bbox": [ + 313, + 418, + 555, + 613 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 418, + 555, + 613 + ], + "type": "text", + "content": "-th layer's output with the scaling factor " + }, + { + "bbox": [ + 313, + 418, + 555, + 613 + ], + "type": "inline_equation", + "content": "\\Delta_K" + }, + { + "bbox": [ + 313, + 418, + 555, + 613 + ], + "type": "text", + "content": ". Note that " + }, + { + "bbox": [ + 313, + 418, + 555, + 613 + ], + "type": "inline_equation", + "content": "\\widetilde{\\mathbf{X}}_K" + }, + { + "bbox": [ + 313, + 418, + 555, + 613 + ], + "type": "text", + "content": " is different from the original representation " + }, + { + "bbox": [ + 313, + 418, + 555, + 613 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_K" + }, + { + "bbox": [ + 313, + 418, + 555, + 613 + ], + "type": "text", + "content": " extracted by the original frontend " + }, + { + "bbox": [ + 313, + 418, + 555, + 613 + ], + "type": "inline_equation", + "content": "F^{\\mathrm{F}}" + }, + { + "bbox": [ + 313, + 418, + 555, + 613 + ], + "type": "text", + "content": " due to the range constraints of round and clamp functions in Eq. (6). Our following design aims to tune the backend to resist such differences. The most straightforward method is to forward the floating-point representation " + }, + { + "bbox": [ + 313, + 418, + 555, + 613 + ], + "type": "inline_equation", + "content": "\\widetilde{\\mathbf{X}}_K" + }, + { + "bbox": [ + 313, + 418, + 555, + 613 + ], + "type": "text", + "content": " to the backend and conduct tuning using task loss. However, such tuning will make the backend over-fitted to the server dataset " + }, + { + "bbox": [ + 313, + 418, + 555, + 613 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^{\\mathrm{S}}" + }, + { + "bbox": [ + 313, + 418, + 555, + 613 + ], + "type": "text", + "content": ". We solve this by a representation-mixup strategy. Specifically, for a particular data sample " + }, + { + "bbox": [ + 313, + 418, + 555, + 613 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 313, + 418, + 555, + 613 + ], + "type": "text", + "content": ", we feed it to a quantized frontend " + }, + { + "bbox": [ + 313, + 418, + 555, + 613 + ], + "type": "inline_equation", + "content": "\\widehat{F}_{m}^{\\mathrm{F}}" + }, + { + "bbox": [ + 313, + 418, + 555, + 613 + ], + "type": "text", + "content": " and the original frontend " + }, + { + "bbox": [ + 313, + 418, + 555, + 613 + ], + "type": "inline_equation", + "content": "F^{\\mathrm{F}}" + }, + { + "bbox": [ + 313, + 418, + 555, + 613 + ], + "type": "text", + "content": " to extract representations " + }, + { + "bbox": [ + 313, + 418, + 555, + 613 + ], + "type": "inline_equation", + "content": "\\widehat{\\mathbf{X}}_{K,m},\\mathbf{X}_{K,m}" + }, + { + "bbox": [ + 313, + 418, + 555, + 613 + ], + "type": "text", + "content": ". After dequantizing " + }, + { + "bbox": [ + 313, + 418, + 555, + 613 + ], + "type": "inline_equation", + "content": "\\widehat{\\mathbf{X}}_{K,m}" + }, + { + "bbox": [ + 313, + 418, + 555, + 613 + ], + "type": "text", + "content": " into the floating-point form " + }, + { + "bbox": [ + 313, + 418, + 555, + 613 + ], + "type": "inline_equation", + "content": "\\widetilde{\\mathbf{X}}_{K,m}" + }, + { + "bbox": [ + 313, + 418, + 555, + 613 + ], + "type": "text", + "content": ", we obtain a mixup representation:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 363, + 618, + 553, + 633 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 363, + 618, + 553, + 633 + ], + "spans": [ + { + "bbox": [ + 363, + 618, + 553, + 633 + ], + "type": "interline_equation", + "content": "\\mathbf {X} _ {K, m} ^ {*} = \\lambda \\widetilde {\\mathbf {X}} _ {K, m} + (1 - \\lambda) \\mathbf {X} _ {K, m}, \\tag {10}", + "image_path": "33aa041dd8cdf0314b409cd4e3212a4b67051c3dd1dc022b6df74ef19595b321.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 639, + 554, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 639, + 554, + 674 + ], + "spans": [ + { + "bbox": [ + 313, + 639, + 554, + 674 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 639, + 554, + 674 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 639, + 554, + 674 + ], + "type": "text", + "content": " is drawn from a widely-used beta distribution " + }, + { + "bbox": [ + 313, + 639, + 554, + 674 + ], + "type": "inline_equation", + "content": "\\mathrm{Beta}(\\lambda;0.75,0.75)" + }, + { + "bbox": [ + 313, + 639, + 554, + 674 + ], + "type": "text", + "content": ". Then considering all available quantized frontend, we can tune the backend with the loss:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 353, + 680, + 555, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 680, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 353, + 680, + 555, + 715 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {t u n e}} = \\sum_ {\\mathcal {D} _ {\\cup} ^ {\\mathrm {S}}} \\sum_ {m = 0} ^ {M} \\mathcal {L} \\left(\\mathbf {X} _ {K, m} ^ {*}, y; \\Theta^ {\\mathrm {B}} \\circ \\gamma^ {\\prime}\\right). \\tag {11}", + "image_path": "275dfb509d3ef1869963fbd460c1895285bc1d1b6d23a754631cd452a387fab7.jpg" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20096" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 100 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 100 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 100 + ], + "type": "text", + "content": "Note that " + }, + { + "bbox": [ + 55, + 72, + 295, + 100 + ], + "type": "inline_equation", + "content": "\\mathbf{X}_{K,0}^{*}" + }, + { + "bbox": [ + 55, + 72, + 295, + 100 + ], + "type": "text", + "content": " originates from feeding " + }, + { + "bbox": [ + 55, + 72, + 295, + 100 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 55, + 72, + 295, + 100 + ], + "type": "text", + "content": " to the frontend " + }, + { + "bbox": [ + 55, + 72, + 295, + 100 + ], + "type": "inline_equation", + "content": "\\widehat{F}^{\\mathrm{F}}" + }, + { + "bbox": [ + 55, + 72, + 295, + 100 + ], + "type": "text", + "content": " quantized on the full set of " + }, + { + "bbox": [ + 55, + 72, + 295, + 100 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\cup}^{\\mathrm{S}}" + }, + { + "bbox": [ + 55, + 72, + 295, + 100 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 108, + 295, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 108, + 295, + 121 + ], + "spans": [ + { + "bbox": [ + 55, + 108, + 295, + 121 + ], + "type": "text", + "content": "3.4. Client Bi-level Noisy Representation Extraction" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 126, + 295, + 211 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 126, + 295, + 211 + ], + "spans": [ + { + "bbox": [ + 55, + 126, + 295, + 211 + ], + "type": "text", + "content": "In our SA model, the quantized frontend " + }, + { + "bbox": [ + 55, + 126, + 295, + 211 + ], + "type": "inline_equation", + "content": "\\widehat{F}^{\\mathrm{F}}" + }, + { + "bbox": [ + 55, + 126, + 295, + 211 + ], + "type": "text", + "content": " will be sent to the client after the OOD enhanced quantization (Section 3.3.1), and then the client can forward its data to " + }, + { + "bbox": [ + 55, + 126, + 295, + 211 + ], + "type": "inline_equation", + "content": "\\widehat{F}^{\\mathrm{F}}" + }, + { + "bbox": [ + 55, + 126, + 295, + 211 + ], + "type": "text", + "content": " to extract representations and send them to the server. However, there is a risk of data reconstruction when sharing representations, thus we propose a bi-level noisy representation extraction mechanism to defend against such attacks." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 212, + 296, + 320 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 212, + 296, + 320 + ], + "spans": [ + { + "bbox": [ + 55, + 212, + 296, + 320 + ], + "type": "text", + "content": "Intuitively speaking, the key to data reconstruction defense is to use the model unknown by the server to extract the representations. Therefore, we allow the client to inject a certain degree of random noise into the received frontend " + }, + { + "bbox": [ + 55, + 212, + 296, + 320 + ], + "type": "inline_equation", + "content": "\\widehat{F}^{\\mathrm{F}}" + }, + { + "bbox": [ + 55, + 212, + 296, + 320 + ], + "type": "text", + "content": ". For each layer in the frontend, we inject a multiplicative noise " + }, + { + "bbox": [ + 55, + 212, + 296, + 320 + ], + "type": "inline_equation", + "content": "n_l^\\times" + }, + { + "bbox": [ + 55, + 212, + 296, + 320 + ], + "type": "text", + "content": ", and an additive noise " + }, + { + "bbox": [ + 55, + 212, + 296, + 320 + ], + "type": "inline_equation", + "content": "n_l^+" + }, + { + "bbox": [ + 55, + 212, + 296, + 320 + ], + "type": "text", + "content": ", which are sampled from two Gaussian distributions, i.e., " + }, + { + "bbox": [ + 55, + 212, + 296, + 320 + ], + "type": "inline_equation", + "content": "n_l^\\times \\sim \\mathcal{P}_{\\mathcal{N}(1,\\alpha \\theta_l)}" + }, + { + "bbox": [ + 55, + 212, + 296, + 320 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 212, + 296, + 320 + ], + "type": "inline_equation", + "content": "n_l^+ \\sim \\mathcal{P}_{\\mathcal{N}(0,\\alpha \\theta_l)}" + }, + { + "bbox": [ + 55, + 212, + 296, + 320 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 55, + 212, + 296, + 320 + ], + "type": "inline_equation", + "content": "\\alpha = 0.01" + }, + { + "bbox": [ + 55, + 212, + 296, + 320 + ], + "type": "text", + "content": " controls the noise degree), to the model parameters:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 135, + 331, + 295, + 345 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 331, + 295, + 345 + ], + "spans": [ + { + "bbox": [ + 135, + 331, + 295, + 345 + ], + "type": "interline_equation", + "content": "\\theta_ {l} ^ {\\prime} = n _ {l} ^ {\\times} \\cdot \\theta_ {l} + n _ {l} ^ {+}. \\tag {12}", + "image_path": "37d7a62b5e166f3aa3195c478d0fdc6a8904bd741e33059d674db309a812617b.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 354, + 296, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 354, + 296, + 441 + ], + "spans": [ + { + "bbox": [ + 55, + 354, + 296, + 441 + ], + "type": "text", + "content": "Next, the client feeds its dataset " + }, + { + "bbox": [ + 55, + 354, + 296, + 441 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^{\\mathrm{C}}" + }, + { + "bbox": [ + 55, + 354, + 296, + 441 + ], + "type": "text", + "content": " to the noisy frontend " + }, + { + "bbox": [ + 55, + 354, + 296, + 441 + ], + "type": "inline_equation", + "content": "\\widehat{F}^{\\mathrm{F}'}" + }, + { + "bbox": [ + 55, + 354, + 296, + 441 + ], + "type": "text", + "content": " to extract representations " + }, + { + "bbox": [ + 55, + 354, + 296, + 441 + ], + "type": "inline_equation", + "content": "\\{\\widehat{\\mathbf{X}}_{K,i}^{\\mathrm{C}}\\}_{i=1}^{N^{\\mathrm{C}}}" + }, + { + "bbox": [ + 55, + 354, + 296, + 441 + ], + "type": "text", + "content": ". However, it is still insecure to share these representations. We found that our model-level noise can amplify the defense effect of the regular noise obfuscation [24] at the representation level, which allows to add a smaller degree of Laplace noise " + }, + { + "bbox": [ + 55, + 354, + 296, + 441 + ], + "type": "inline_equation", + "content": "(\\mathcal{P}_{\\mathrm{Laplace}(0,0.8)})" + }, + { + "bbox": [ + 55, + 354, + 296, + 441 + ], + "type": "text", + "content": " to representations for better protection." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 449, + 292, + 462 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 449, + 292, + 462 + ], + "spans": [ + { + "bbox": [ + 55, + 449, + 292, + 462 + ], + "type": "text", + "content": "3.5. Patch-retrieval-augmented Model Adaptation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 467, + 295, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 467, + 295, + 538 + ], + "spans": [ + { + "bbox": [ + 55, + 467, + 295, + 538 + ], + "type": "text", + "content": "In few-shot adaptation scenarios, like the backend tuning on the server (Section 3.3.2), the adaptation also faces overfitting challenges, but the overfitting here is due to the limited client data available. To tackle such challenges, we design a patch-level retrieval augmentation approach to generate more client data representations." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 540, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 540, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 540, + 296, + 715 + ], + "type": "text", + "content": "Our patch-level retrieval augmentation is inspired by the intuition that different images may contain similar patches. As introduced in Section 3.1.1, the data representations of each layer in a ViT hold the dimension " + }, + { + "bbox": [ + 55, + 540, + 296, + 715 + ], + "type": "inline_equation", + "content": "(N + 1)\\times d" + }, + { + "bbox": [ + 55, + 540, + 296, + 715 + ], + "type": "text", + "content": ", which can be denoted as " + }, + { + "bbox": [ + 55, + 540, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\widehat{\\mathbf{X}}_{K,i}^{\\mathrm{C}} = [\\mathbf{x}_i^{\\mathrm{CLS}};\\mathbf{x}_{i,1},\\dots ,\\mathbf{x}_{i,N}]" + }, + { + "bbox": [ + 55, + 540, + 296, + 715 + ], + "type": "text", + "content": " (we omit some symbols for simplification). Then we can form " + }, + { + "bbox": [ + 55, + 540, + 296, + 715 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 540, + 296, + 715 + ], + "type": "text", + "content": " retrieval sets from all client data representations (before adding Laplace noise) and each corresponds to a patch, i.e., " + }, + { + "bbox": [ + 55, + 540, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\{S_j\\}_{j = 1}^N" + }, + { + "bbox": [ + 55, + 540, + 296, + 715 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 55, + 540, + 296, + 715 + ], + "type": "inline_equation", + "content": "S_{j} = \\{\\mathbf{x}_{i,j}\\}_{i = 1}^{N^{\\mathrm{C}}}" + }, + { + "bbox": [ + 55, + 540, + 296, + 715 + ], + "type": "text", + "content": ". With these sets, for each augmentation of " + }, + { + "bbox": [ + 55, + 540, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\widehat{\\mathbf{X}}_{K,i}^{\\mathrm{C}}" + }, + { + "bbox": [ + 55, + 540, + 296, + 715 + ], + "type": "text", + "content": ", we randomly select " + }, + { + "bbox": [ + 55, + 540, + 296, + 715 + ], + "type": "inline_equation", + "content": "N^{\\mathrm{P}} = 50" + }, + { + "bbox": [ + 55, + 540, + 296, + 715 + ], + "type": "text", + "content": " patches (please see the sensitivity analysis in Supplementary Materials) " + }, + { + "bbox": [ + 55, + 540, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{x}_{i,\\mathrm{P}_j}\\}_{j = 1}^{N^{\\mathrm{P}}}" + }, + { + "bbox": [ + 55, + 540, + 296, + 715 + ], + "type": "text", + "content": " (subscript " + }, + { + "bbox": [ + 55, + 540, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{P}_j" + }, + { + "bbox": [ + 55, + 540, + 296, + 715 + ], + "type": "text", + "content": " meets " + }, + { + "bbox": [ + 55, + 540, + 296, + 715 + ], + "type": "inline_equation", + "content": "1\\leq \\mathrm{P}_j\\leq N" + }, + { + "bbox": [ + 55, + 540, + 296, + 715 + ], + "type": "text", + "content": ") and delete them from their corresponding retrieval sets. Next, each " + }, + { + "bbox": [ + 55, + 540, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_{i,\\mathrm{P}_j}" + }, + { + "bbox": [ + 55, + 540, + 296, + 715 + ], + "type": "text", + "content": " retrieves the most similar patch " + }, + { + "bbox": [ + 55, + 540, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\Omega (\\mathbf{x}_{i,\\mathrm{P}_j})" + }, + { + "bbox": [ + 55, + 540, + 296, + 715 + ], + "type": "text", + "content": " from" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 72, + 438, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 438, + 84 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 438, + 84 + ], + "type": "text", + "content": "its corresponding retrieval set:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 365, + 91, + 553, + 118 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 91, + 553, + 118 + ], + "spans": [ + { + "bbox": [ + 365, + 91, + 553, + 118 + ], + "type": "interline_equation", + "content": "\\Omega \\left(\\mathbf {x} _ {i, \\mathrm {P} _ {j}}\\right) = \\arg \\max _ {\\mathbf {x} \\in S _ {\\mathrm {P} _ {j}} ^ {\\prime}} \\frac {\\mathbf {x} _ {i , \\mathrm {P} _ {j}} \\cdot \\mathbf {x}}{\\left| \\mathbf {x} _ {i , \\mathrm {P} _ {j}} \\right| \\left| \\mathbf {x} \\right|}, \\tag {13}", + "image_path": "856bded95402c49f34d23b346a3c963dd274490922fec0fd68a0e41b3f1cd8e3.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 126, + 555, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 126, + 555, + 312 + ], + "spans": [ + { + "bbox": [ + 313, + 126, + 555, + 312 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 126, + 555, + 312 + ], + "type": "inline_equation", + "content": "\\mathcal{S}_{\\mathrm{P}_j}^{\\prime} = \\mathcal{S}_{\\mathrm{P}_j}\\backslash \\mathbf{x}_{i,\\mathrm{P}_j}" + }, + { + "bbox": [ + 313, + 126, + 555, + 312 + ], + "type": "text", + "content": ". After the retrieval operations for all selected patches " + }, + { + "bbox": [ + 313, + 126, + 555, + 312 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{x}_{i,\\mathrm{P}_j}\\}_{j = 1}^{N^{\\mathrm{P}}}" + }, + { + "bbox": [ + 313, + 126, + 555, + 312 + ], + "type": "text", + "content": ", we replace them with the retrieved patches " + }, + { + "bbox": [ + 313, + 126, + 555, + 312 + ], + "type": "inline_equation", + "content": "\\{\\Omega (\\mathbf{x}_{i,\\mathrm{P}_j})\\}_{j = 1}^{N^{\\mathrm{P}}}" + }, + { + "bbox": [ + 313, + 126, + 555, + 312 + ], + "type": "text", + "content": " in the original representation " + }, + { + "bbox": [ + 313, + 126, + 555, + 312 + ], + "type": "inline_equation", + "content": "\\widehat{\\mathbf{X}}_{K,i}^{\\mathrm{C}}" + }, + { + "bbox": [ + 313, + 126, + 555, + 312 + ], + "type": "text", + "content": " to form a new one " + }, + { + "bbox": [ + 313, + 126, + 555, + 312 + ], + "type": "inline_equation", + "content": "\\widehat{\\mathbf{X}}_{K,i}^{\\prime \\mathrm{C}}" + }, + { + "bbox": [ + 313, + 126, + 555, + 312 + ], + "type": "text", + "content": ". Note that the selected patches are determined randomly, resulting in newly generated representations that differ across augmentation runs. To generate sufficient augmented representations, we run the augmentation " + }, + { + "bbox": [ + 313, + 126, + 555, + 312 + ], + "type": "inline_equation", + "content": "N^{\\mathrm{Aug}}" + }, + { + "bbox": [ + 313, + 126, + 555, + 312 + ], + "type": "text", + "content": " times for all client data representations and produce an augmented representation set, i.e., " + }, + { + "bbox": [ + 313, + 126, + 555, + 312 + ], + "type": "inline_equation", + "content": "\\widehat{\\mathcal{X}}_K^{\\mathrm{C}} = \\{\\{\\widehat{\\mathbf{X}}_{K,i}^{\\mathrm{C}_j}\\}_{j = 0}^{N^{\\mathrm{Aug}}}\\}_{i = 1}^{N^{\\mathrm{C}}}" + }, + { + "bbox": [ + 313, + 126, + 555, + 312 + ], + "type": "text", + "content": " (we denote the original representation as " + }, + { + "bbox": [ + 313, + 126, + 555, + 312 + ], + "type": "inline_equation", + "content": "\\widehat{\\mathbf{X}}_{K,i}^{\\mathrm{C}_0}" + }, + { + "bbox": [ + 313, + 126, + 555, + 312 + ], + "type": "text", + "content": " and the augmented representations as " + }, + { + "bbox": [ + 313, + 126, + 555, + 312 + ], + "type": "inline_equation", + "content": "\\widehat{\\mathbf{X}}_{K,i}^{\\mathrm{C}_{1\\sim N^{\\mathrm{Aug}}}}" + }, + { + "bbox": [ + 313, + 126, + 555, + 312 + ], + "type": "text", + "content": "). The above patch retrieval augmentation is run by the client, and it adds Laplace noise (Section 3.4) to all representations in " + }, + { + "bbox": [ + 313, + 126, + 555, + 312 + ], + "type": "inline_equation", + "content": "\\widehat{\\mathcal{X}}_K^{\\mathrm{C}}" + }, + { + "bbox": [ + 313, + 126, + 555, + 312 + ], + "type": "text", + "content": " before uploading to the server." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 313, + 554, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 313, + 554, + 421 + ], + "spans": [ + { + "bbox": [ + 313, + 313, + 554, + 421 + ], + "type": "text", + "content": "After receiving the representations, the server dequantizes them into floating-point form with the " + }, + { + "bbox": [ + 313, + 313, + 554, + 421 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 313, + 554, + 421 + ], + "type": "text", + "content": "-th layer's scaling factor " + }, + { + "bbox": [ + 313, + 313, + 554, + 421 + ], + "type": "inline_equation", + "content": "\\Delta_K" + }, + { + "bbox": [ + 313, + 313, + 554, + 421 + ], + "type": "text", + "content": " by Eq. (9) and launches the adaptation. To further alleviate the overfitting due to limited client data, different from Split Learning, we only allow high-level feature learning in the backend to prevent low-level features from becoming biased. Then the server randomly initializes a task module " + }, + { + "bbox": [ + 313, + 313, + 554, + 421 + ], + "type": "inline_equation", + "content": "g_{\\gamma}" + }, + { + "bbox": [ + 313, + 313, + 554, + 421 + ], + "type": "text", + "content": " and tunes it with the backend " + }, + { + "bbox": [ + 313, + 313, + 554, + 421 + ], + "type": "inline_equation", + "content": "F^{\\mathrm{B}}" + }, + { + "bbox": [ + 313, + 313, + 554, + 421 + ], + "type": "text", + "content": " (that has been tuned in Section 3.3.2) by following the loss:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 353, + 428, + 553, + 464 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 428, + 553, + 464 + ], + "spans": [ + { + "bbox": [ + 353, + 428, + 553, + 464 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {a d a p t}} = \\sum_ {i = 1} ^ {N ^ {\\mathrm {C}}} \\sum_ {j = 0} ^ {N ^ {\\mathrm {A u g}}} \\mathcal {L} \\left(\\widehat {\\mathbf {X}} _ {K, i} ^ {\\mathrm {C} _ {j}}, y; \\Theta^ {\\mathrm {B}} \\circ \\gamma\\right). \\tag {14}", + "image_path": "82c9c2c1183e4ea3fd4251ca67787fcbae1a4fb743811826ae5763d331287c9b.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 472, + 554, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 472, + 554, + 509 + ], + "spans": [ + { + "bbox": [ + 313, + 472, + 554, + 509 + ], + "type": "text", + "content": "To protect the client's labels, the server sends the task module output to the client, allowing the task loss " + }, + { + "bbox": [ + 313, + 472, + 554, + 509 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{adapt}}" + }, + { + "bbox": [ + 313, + 472, + 554, + 509 + ], + "type": "text", + "content": " to be computed on the client side and then sent back to the server." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 518, + 395, + 532 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 518, + 395, + 532 + ], + "spans": [ + { + "bbox": [ + 313, + 518, + 395, + 532 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 538, + 441, + 551 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 538, + 441, + 551 + ], + "spans": [ + { + "bbox": [ + 313, + 538, + 441, + 551 + ], + "type": "text", + "content": "4.1. Experimental Settings" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 555, + 555, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 555, + 555, + 651 + ], + "spans": [ + { + "bbox": [ + 313, + 555, + 555, + 651 + ], + "type": "text", + "content": "Datasets. ImageNet [9] is the most widely used image dataset collected from public domains, thus we use its test set as the server dataset. For the client datasets, we use CIFAR-100 [30], Places365 [35], and DomainNet [39]. Places365 is a scene recognition dataset that contains 365 classes. DomainNet (D.Net) is a challenging multi-domain dataset, which includes six domains. We select the smallest domain—Clipart (Cl) for the few-shot adaptation scenarios." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "text", + "content": "Implementation Details. We conduct experiments on three pre-trained model architectures: ViT-Large [13], DeiT-Base [43], and Swin-L [33] (DeiT-Base and Swin-L's experiments are reported in Supplementary Materials). For the Places365 dataset, the target task module is composed" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20097" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 67, + 94, + 538, + 196 + ], + "blocks": [ + { + "bbox": [ + 55, + 71, + 555, + 92 + ], + "lines": [ + { + "bbox": [ + 55, + 71, + 555, + 92 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 555, + 92 + ], + "type": "text", + "content": "Table 1. Performance comparison between our SA and other baseline approaches in 3-shot, 5-shot, and 10-shot adaptation scenarios over 3 datasets. The SA can substantially exceed other methods. We bold and blue the best, and bold the second best." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 67, + 94, + 538, + 196 + ], + "lines": [ + { + "bbox": [ + 67, + 94, + 538, + 196 + ], + "spans": [ + { + "bbox": [ + 67, + 94, + 538, + 196 + ], + "type": "table", + "html": "
Few-shot Setup3-shot5-shot10-shot
MethodCIFAR-100Places365D.Net-ClCIFAR-100Places365D.Net-ClCIFAR-100Places365D.Net-Cl
Linear Probing74.05±1.2126.18±1.0057.26±1.1178.79±0.6534.25±1.4563.84±0.6683.28±0.3332.33±7.8269.12±0.75
Fine Tuning40.32±5.4821.64±0.7836.95±6.6664.29±7.6931.07±1.0456.50±1.8781.82±1.8539.46±0.5970.67±0.72
LN-TUNE13.18±2.171.59±0.536.97±1.9634.80±5.373.96±0.9513.43±2.1247.97±6.257.71±1.8524.53±2.05
Split Learning74.05±1.2126.84±1.0358.14±0.6379.63±0.5330.49±0.3663.95±1.1883.66±0.2735.78±0.3065.20±0.47
Offsite Tuning42.37±2.9824.98±0.5540.56±4.0064.83±6.1730.45±0.8056.43±0.4580.11±1.0736.22±1.3368.43±0.64
SA (ours)76.24±0.2930.92±0.8956.26±0.7981.98±0.4935.31±0.0165.03±0.5985.45±0.4039.26±0.1071.13±0.76
", + "image_path": "823c28bae9ed0628ff522302e07475c10eae5ff400e07dfac6deec139d602c36.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 209, + 294, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 209, + 294, + 342 + ], + "spans": [ + { + "bbox": [ + 54, + 209, + 294, + 342 + ], + "type": "text", + "content": "of two linear layers, whereas for other datasets, it consists of a single linear layer. We divide 2/3 backbone model as the frontend while the rest " + }, + { + "bbox": [ + 54, + 209, + 294, + 342 + ], + "type": "inline_equation", + "content": "1/3" + }, + { + "bbox": [ + 54, + 209, + 294, + 342 + ], + "type": "text", + "content": " is the backend. The frontend and its inner activations are quantized to 8-bit. Each quantization operation uses 32 data samples randomly sampled from the server dataset or subsets, and the batch size is 4. The Adam optimizer with a learning rate of " + }, + { + "bbox": [ + 54, + 209, + 294, + 342 + ], + "type": "inline_equation", + "content": "10^{-5}" + }, + { + "bbox": [ + 54, + 209, + 294, + 342 + ], + "type": "text", + "content": " is adopted. The number of epochs for OOD quantization-aware tuning is set to 1, while the final adaptation is set to 100 epochs. All experiments are run repeatedly 3 times with three seeds, and we report the mean performance and standard deviation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 346, + 295, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 346, + 295, + 430 + ], + "spans": [ + { + "bbox": [ + 55, + 346, + 295, + 430 + ], + "type": "text", + "content": "Baseline Approaches. SA is compared with state-of-the-art downstream task adaptation methods including Linear Probing [29], Fine Tuning, Split Learning [44], Offsite Tuning [49], and LN-TUNE [2]. To evaluate the effectiveness of data protection, we extend the state-of-the-art data reconstruction attacks—FORA [53]—to ViT. More details are provided in the Supplementary Materials." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 434, + 295, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 434, + 295, + 495 + ], + "spans": [ + { + "bbox": [ + 55, + 434, + 295, + 495 + ], + "type": "text", + "content": "Evaluation Metrics. To evaluate the effectiveness of the target task adaptation, we use classification accuracy to measure the performance. As for the quality measurement of data reconstruction results, we adopt three metrics: SSIM [48] (S), PSNR [25] (P), and LPIPS [56] (L)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 503, + 263, + 516 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 503, + 263, + 516 + ], + "spans": [ + { + "bbox": [ + 55, + 503, + 263, + 516 + ], + "type": "text", + "content": "4.2. Effectiveness of SA for Task Adaptation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 522, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 522, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 522, + 295, + 713 + ], + "type": "text", + "content": "We create three few-shot adaptation scenarios for each client dataset or domain: 3-shot, 5-shot, and 10-shot per class, with experimental results shown in Table 1. We can easily observe that our SA can always substantially exceed all other baseline approaches. In particular, methods like Fine Tuning, LN-TUNE, and Offsite Tuning lag far behind SA, which may be attributed to the overfitting caused by the low-level feature tuning. The method closest to SA in terms of performance is Split Learning, but its biggest issue is the privacy leakage, which we will discuss in Section 4.3. In Table 2, we also provide the client memory and computation costs of each method (assuming the first three methods are conducted on the client side), from which we can observe that SA even runs with the smallest cost. Note that SA has no training on the client, thus we view the representation extraction time as the training time in Table 2." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 209, + 553, + 232 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 209, + 553, + 232 + ], + "spans": [ + { + "bbox": [ + 313, + 209, + 553, + 232 + ], + "type": "text", + "content": "4.3. Effectiveness of SA for Server Pre-trained ViT and Client Data Protection" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 312, + 238, + 555, + 393 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 238, + 555, + 393 + ], + "spans": [ + { + "bbox": [ + 312, + 238, + 555, + 393 + ], + "type": "text", + "content": "For privacy protection, we first test whether the quantized frontend of SA can result in high-quality models with good downstream task performance. We consider the three most potential variations: 1) directly tune a task module on the quantized frontend (Quant. Frontend); 2) tune a task module on the original frontend though this case is impossible in practice (Original Frontend); 3) attach a backend from other models and tune a task module on the quantized frontend and auxiliary backend (Auxiliary Backend). The experimental results are shown in Table 3, in which all variations perform much poorer than Linear Probing in Table 1 and our SA. In this case, we believe SA can protect the pretrained model to a great extent." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 394, + 556, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 394, + 556, + 538 + ], + "spans": [ + { + "bbox": [ + 313, + 394, + 556, + 538 + ], + "type": "text", + "content": "Then we launch FORA [53] to evaluate the data protection of our SA. For a fair comparison, only Split Learning is compared with SA as only it requires the client to share intermediate data representations with the server. The experiment results are presented in Table 5. First, we can visually observe that the reconstruction results of SA are much worse than those of Split Learning. Then the three metrics also demonstrate the same phenomenon, i.e., FORA can only recover much lower-quality images when attacking SA compared with Split Learning. In summary, our SA offers strong protection of the client data. More experiments on the noise degree can be found in Supplementary Materials." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 545, + 408, + 558 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 408, + 558 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 408, + 558 + ], + "type": "text", + "content": "4.4. Ablation Study" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 563, + 555, + 587 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 563, + 555, + 587 + ], + "spans": [ + { + "bbox": [ + 313, + 563, + 555, + 587 + ], + "type": "text", + "content": "We conduct comprehensive ablation studies in the 5-shot adaptation on CIFAR-100, Places365, and DomainNet-C1." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 590, + 554, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 590, + 554, + 662 + ], + "spans": [ + { + "bbox": [ + 313, + 590, + 554, + 662 + ], + "type": "text", + "content": "Hilbert Transform-based Data Augmentation. We detach the use of HT data augmentation in the frontend quantization and backend tuning ('SA w/o HT Aug'). Compared with the full SA in Table 4, the variation without HT data augmentation performs worse, which proves its gain in improving adaptation performance." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "text", + "content": "Out-of-distribution Quantization-aware Tuning. We design two variations: one is not to use OOD quantization-aware tuning ('SA w/o OOD QAT'), and the other is only to use the quantized frontend " + }, + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\widehat{F}_{\\mathrm{F}}" + }, + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "text", + "content": " characterized on the full" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20098" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 105, + 197, + 182 + ], + "blocks": [ + { + "bbox": [ + 55, + 71, + 198, + 102 + ], + "lines": [ + { + "bbox": [ + 55, + 71, + 198, + 102 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 198, + 102 + ], + "type": "text", + "content": "Table 2. Comparison of client memory (MB) and computation costs (Min) between SA and other baselines." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 105, + 197, + 182 + ], + "lines": [ + { + "bbox": [ + 58, + 105, + 197, + 182 + ], + "spans": [ + { + "bbox": [ + 58, + 105, + 197, + 182 + ], + "type": "table", + "html": "
MethodsGPU MemoryTraining Time
Linear Probing697966
Fine Tuning1030234
LN-TUNE596825
Split Learning893215
Offsite Tuning489623
SA (ours)22332.5
", + "image_path": "b8294088e319c148ccd3808d318e5ff790dce029fdcc3469c96500ac03b49c18.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 206, + 110, + 375, + 173 + ], + "blocks": [ + { + "bbox": [ + 203, + 71, + 376, + 105 + ], + "lines": [ + { + "bbox": [ + 203, + 71, + 376, + 105 + ], + "spans": [ + { + "bbox": [ + 203, + 71, + 376, + 105 + ], + "type": "text", + "content": "Table 3. Risk assessment of 3 most potential cases over 3 datasets (domains) for obtaining high-quality models from our SA approach." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 206, + 110, + 375, + 173 + ], + "lines": [ + { + "bbox": [ + 206, + 110, + 375, + 173 + ], + "spans": [ + { + "bbox": [ + 206, + 110, + 375, + 173 + ], + "type": "table", + "html": "
VariationsCIFAR-100Places365D.Net-CI
Quant. Frontend26.3413.4526.70
Original Frontend26.5313.7227.12
Auxiliary Backend30.6510.1516.56
SA (ours)81.9835.3171.13
", + "image_path": "8f39c861de998f111a0cbf29ddf619822333c156a6ef7aec08db6cce6b27d4a5.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 384, + 110, + 553, + 182 + ], + "blocks": [ + { + "bbox": [ + 381, + 71, + 555, + 104 + ], + "lines": [ + { + "bbox": [ + 381, + 71, + 555, + 104 + ], + "spans": [ + { + "bbox": [ + 381, + 71, + 555, + 104 + ], + "type": "text", + "content": "Table 4. Ablation study of the major components of our SA approach in 5-shot adaptation over 3 datasets (domains)." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 384, + 110, + 553, + 182 + ], + "lines": [ + { + "bbox": [ + 384, + 110, + 553, + 182 + ], + "spans": [ + { + "bbox": [ + 384, + 110, + 553, + 182 + ], + "type": "table", + "html": "
VariationsCIFAR-100Places365D.Net-CI
SA w/o HT Aug80.3735.2764.59
SA w/o OOD QAT79.0329.5064.42
SA w/o QAT80.0529.5964.16
SA w/o PR Aug79.9029.5652.85
SA (ours)81.9835.3171.13
", + "image_path": "2271bb2d9e6ad17b4bf7c6f291c36842de76d68dda71e5de1087315b0d9b0a5a.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 58, + 216, + 553, + 369 + ], + "blocks": [ + { + "bbox": [ + 55, + 192, + 555, + 214 + ], + "lines": [ + { + "bbox": [ + 55, + 192, + 555, + 214 + ], + "spans": [ + { + "bbox": [ + 55, + 192, + 555, + 214 + ], + "type": "text", + "content": "Table 5. Comparison of defensive performance against the state-of-the-art data reconstruction attack—FORA [53]—between SA and Split Learning. We visualize the reconstruction and calculate SSIM (S↓), PSNR (P↓), and LPIPS (L↑) to measure the reconstruction quality." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 216, + 553, + 369 + ], + "lines": [ + { + "bbox": [ + 58, + 216, + 553, + 369 + ], + "spans": [ + { + "bbox": [ + 58, + 216, + 553, + 369 + ], + "type": "table", + "html": "
DatasetCIFAR-100Places365D.Net-CI
Ground Truth
Split LearningS:0.80 P:26.1 L:0.46S:0.43 P:20.8 L:0.67S:0.59 P:22.1 L:0.46
SA (ours)S:0.23 P:15.2 L:0.70S:0.30 P:16.6 L:0.73S:0.25 P:14.4 L:0.60
", + "image_path": "baea76b36a7f3bef313a56fa507fe3b1024a7d8e3ed066856a040e76f0744159.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 65, + 401, + 286, + 605 + ], + "blocks": [ + { + "bbox": [ + 55, + 377, + 295, + 400 + ], + "lines": [ + { + "bbox": [ + 55, + 377, + 295, + 400 + ], + "spans": [ + { + "bbox": [ + 55, + 377, + 295, + 400 + ], + "type": "text", + "content": "Table 6. Ablation study of SA's bi-level noise when defending against FORA [53] on CIFAR-100." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 65, + 401, + 286, + 605 + ], + "lines": [ + { + "bbox": [ + 65, + 401, + 286, + 605 + ], + "spans": [ + { + "bbox": [ + 65, + 401, + 286, + 605 + ], + "type": "table", + "html": "
DatasetCIFAR-100
Ground Truth-
SA w/o Laplace NoiseS:0.76 P:23.1 L:0.49
SA w/o Model NoiseS:0.28 P:16.4 L:0.67
SA (ours)S:0.23 P:15.2 L:0.70
", + "image_path": "029a508b0bc73126c249ae3ef42e5d1424f427ba040e68b9f88a3bb15715c5c3.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 615, + 295, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 615, + 295, + 663 + ], + "spans": [ + { + "bbox": [ + 55, + 615, + 295, + 663 + ], + "type": "text", + "content": "server dataset ('SA w/o QAT'). Compared with the full SA, the effectiveness of the representation-mixup quantization-aware tuning and the simulated OOD frontends in alleviating quantization-induced performance drop can be verified." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "Patch Retrieval Augmentation. We don't use the patch retrieval augmentation and observe how the model performance will change. The big difference between 'SA w/o PR Aug' and 'SA' in Table 4 indicates the essential role of our" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 380, + 512, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 380, + 512, + 392 + ], + "spans": [ + { + "bbox": [ + 313, + 380, + 512, + 392 + ], + "type": "text", + "content": "augmentation here in performance enhancement." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 395, + 555, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 395, + 555, + 468 + ], + "spans": [ + { + "bbox": [ + 313, + 395, + 555, + 468 + ], + "type": "text", + "content": "Bi-level Noisy Representation Extraction. We detach the use of model Gaussian noise ('SA w/o Model Noise') and representation Laplace noise ('SA w/o Laplace Noise'), respectively, and the experiments are shown in Table 6. Model noise and representation noise play a synergistic role greater than the simple sum of them in data protection." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 314, + 480, + 388, + 492 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 480, + 388, + 492 + ], + "spans": [ + { + "bbox": [ + 314, + 480, + 388, + 492 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 501, + 555, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 501, + 555, + 657 + ], + "spans": [ + { + "bbox": [ + 313, + 501, + 555, + 657 + ], + "type": "text", + "content": "In this work, we propose Split Adaptation (SA) to adapt pre-trained ViTs for downstream tasks while protecting the data and model in the sever-client framework. After dividing the ViT into a frontend and a backend, SA innovatively uses model quantization to protect the frontend. SA also allows the client to introduce a bi-level noise to the quantized frontend and the extracted data representations for data protection. Accordingly, SA uses data-level and model-level augmentation to mitigate the influence of the quantization and the noise. Besides, SA proposes patch retrieval augmentation to address overfitting in few-shot adaptation. Extensive experiments have validated the superiority of SA in task adaptation and data-model protection." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 314, + 669, + 411, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 669, + 411, + 682 + ], + "spans": [ + { + "bbox": [ + 314, + 669, + 411, + 682 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": "This work was partially supported by NTU-NAP startup grant 024584-00001." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20099" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "text", + "content": "[1] Corinne G Allaart, Björn Keyser, Henri Bal, and Aart Van Halteren. Vertical split learning—an exploration of predictive performance in medical and other use cases. In 2022 International Joint Conference on Neural Networks (IJCNN), pages 1-8. IEEE, 2022. 1, 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 146, + 296, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 146, + 296, + 190 + ], + "spans": [ + { + "bbox": [ + 61, + 146, + 296, + 190 + ], + "type": "text", + "content": "[2] Samyadeep Basu, Shell Hu, Daniela Massiceti, and Soheil Feizi. Strong baselines for parameter-efficient few-shot finetuning. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 11024-11031, 2024. 7, 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 191, + 294, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 191, + 294, + 213 + ], + "spans": [ + { + "bbox": [ + 62, + 191, + 294, + 213 + ], + "type": "text", + "content": "[3] Tom B Brown. Language models are few-shot learners. arXiv preprint arXiv:2005.14165, 2020. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 214, + 294, + 268 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 214, + 294, + 268 + ], + "spans": [ + { + "bbox": [ + 62, + 214, + 294, + 268 + ], + "type": "text", + "content": "[4] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 2, 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 270, + 294, + 322 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 270, + 294, + 322 + ], + "spans": [ + { + "bbox": [ + 62, + 270, + 294, + 322 + ], + "type": "text", + "content": "[5] Shoufa Chen, Chongjian Ge, Zhan Tong, Jiangliu Wang, Yibing Song, Jue Wang, and Ping Luo. Adaptformer: Adapting vision transformers for scalable visual recognition. Advances in Neural Information Processing Systems, 35:16664-16678, 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 324, + 294, + 368 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 324, + 294, + 368 + ], + "spans": [ + { + "bbox": [ + 62, + 324, + 294, + 368 + ], + "type": "text", + "content": "[6] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 369, + 294, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 369, + 294, + 422 + ], + "spans": [ + { + "bbox": [ + 62, + 369, + 294, + 422 + ], + "type": "text", + "content": "[7] Yoni Choukroun, Eli Kravchik, Fan Yang, and Pavel Kisilev. Low-bit quantization of neural networks for efficient inference. In 2019 IEEE/CVF International Conference on Computer Vision Workshop (ICCVW), pages 3009-3018. IEEE, 2019. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 425, + 294, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 425, + 294, + 479 + ], + "spans": [ + { + "bbox": [ + 62, + 425, + 294, + 479 + ], + "type": "text", + "content": "[8] Zeki Murat Çınar, Abubakar Abdussalam Nuhu, Qasim Zeešan, Orhan Korhan, Mohammed Asmael, and Babak Safaei. Machine learning in predictive maintenance towards sustainable smart manufacturing in industry 4.0. Sustainability, 12 (19):8211, 2020. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 480, + 294, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 480, + 294, + 523 + ], + "spans": [ + { + "bbox": [ + 62, + 480, + 294, + 523 + ], + "type": "text", + "content": "[9] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 525, + 294, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 525, + 294, + 546 + ], + "spans": [ + { + "bbox": [ + 57, + 525, + 294, + 546 + ], + "type": "text", + "content": "[10] Matthew F Dixon, Igor Halperin, and Paul Bilokon. Machine learning in finance. Springer, 2020. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 547, + 294, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 547, + 294, + 601 + ], + "spans": [ + { + "bbox": [ + 57, + 547, + 294, + 601 + ], + "type": "text", + "content": "[11] Jiahua Dong, Lixu Wang, Zhen Fang, Gan Sun, Shichao Xu, Xiao Wang, and Qi Zhu. Federated class-incremental learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10164–10173, 2022. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 602, + 294, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 602, + 294, + 645 + ], + "spans": [ + { + "bbox": [ + 57, + 602, + 294, + 645 + ], + "type": "text", + "content": "[12] Wei Dong, Qiyao Luo, and Ke Yi. Continual observation under user-level differential privacy. In 2023 IEEE Symposium on Security and Privacy (SP), pages 2190-2207. IEEE, 2023. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 647, + 294, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 647, + 294, + 678 + ], + "spans": [ + { + "bbox": [ + 57, + 647, + 294, + 678 + ], + "type": "text", + "content": "[13] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. ICLR, 2020. 1, 2, 3, 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 680, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 680, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 680, + 294, + 713 + ], + "type": "text", + "content": "[14] Juanru Fang, Wei Dong, and Ke Yi. Shifted inverse: A general mechanism for monotonic functions under user differential privacy. In Proceedings of the 2022 ACM SIGSAC Con-" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 74, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 333, + 74, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 74, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 74, + 553, + 95 + ], + "type": "text", + "content": "ference on Computer and Communications Security, pages 1009-1022, 2022. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 96, + 553, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 129 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 129 + ], + "type": "text", + "content": "[15] Chongyang Gao, Lixu Wang, Chenkai Weng, Xiao Wang, and Qi Zhu. Practical unlearning for large language models. arXiv preprint arXiv:2407.10223, 2024. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 130, + 553, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 130, + 553, + 174 + ], + "spans": [ + { + "bbox": [ + 316, + 130, + 553, + 174 + ], + "type": "text", + "content": "[16] Xinben Gao and Lan Zhang. {PCAT}: Functionality and data stealing from split learning by {Pseudo-Client} attack. In 32nd USENIX Security Symposium (USENIX Security 23), pages 5271-5288, 2023. 1, 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 175, + 553, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 553, + 250 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 553, + 250 + ], + "type": "text", + "content": "[17] Yansong Gao, Minki Kim, Sharif Abuadbba, Yeonjae Kim, Chandra Thapa, Kyuyeon Kim, Seyit A Camtep, Hyoungshick Kim, and Surya Nepal. End-to-end evaluation of federated learning and split learning for internet of things. In 2020 International Symposium on Reliable Distributed Systems (SRDS), pages 91-100. IEEE Computer Society, 2020. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 253, + 553, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 253, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 316, + 253, + 553, + 319 + ], + "type": "text", + "content": "[18] Jiahui Gong, Jingtao Ding, Fanjin Meng, Guilong Chen, Hong Chen, Shen Zhao, Haisheng Lu, and Yong Li. A population-to-individual tuning framework for adapting pretrained lm to on-device user intent prediction. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 896-907, 2024. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 320, + 553, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 320, + 553, + 374 + ], + "spans": [ + { + "bbox": [ + 316, + 320, + 553, + 374 + ], + "type": "text", + "content": "[19] Junfeng Guo, Yiming Li, Lixu Wang, Shu-Tao Xia, Heng Huang, Cong Liu, and Bo Li. Domain watermark: Effective and harmless dataset copyright protection is closed at hand. Advances in Neural Information Processing Systems, 36:54421-54450, 2023. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 376, + 553, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 376, + 553, + 409 + ], + "spans": [ + { + "bbox": [ + 316, + 376, + 553, + 409 + ], + "type": "text", + "content": "[20] Song Han, Huizi Mao, and William J Dally. Deep compression: Compressing deep neural networks with pruning, trained quantization and huffman coding. *ICLR*, 2016. 4, 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 411, + 553, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 411, + 553, + 464 + ], + "spans": [ + { + "bbox": [ + 316, + 411, + 553, + 464 + ], + "type": "text", + "content": "[21] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 466, + 553, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 466, + 553, + 520 + ], + "spans": [ + { + "bbox": [ + 316, + 466, + 553, + 520 + ], + "type": "text", + "content": "[22] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16000-16009, 2022. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 522, + 553, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 522, + 553, + 566 + ], + "spans": [ + { + "bbox": [ + 316, + 522, + 553, + 566 + ], + "type": "text", + "content": "[23] Zecheng He, Tianwei Zhang, and Ruby B Lee. Model inversion attacks against collaborative inference. In Proceedings of the 35th Annual Computer Security Applications Conference, pages 148-162, 2019. 1, 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 567, + 553, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 567, + 553, + 609 + ], + "spans": [ + { + "bbox": [ + 316, + 567, + 553, + 609 + ], + "type": "text", + "content": "[24] Zecheng He, Tianwei Zhang, and Ruby B Lee. Attacking and protecting data privacy in edge-cloud collaborative inference systems. IEEE Internet of Things Journal, 8(12):9706-9716, 2020. 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 612, + 553, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 612, + 553, + 645 + ], + "spans": [ + { + "bbox": [ + 316, + 612, + 553, + 645 + ], + "type": "text", + "content": "[25] Alain Hore and Djemel Ziou. Image quality metrics: Psnr vs. ssm. In 2010 20th international conference on pattern recognition, pages 2366-2369. IEEE, 2010. 7" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 647, + 553, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 690 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 690 + ], + "type": "text", + "content": "[26] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In European Conference on Computer Vision, pages 709-727. Springer, 2022. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "text", + "content": "[27] Peter Kairouz, H Brendan McMahan, Brendan Avent, Aurélien Bellet, Mehdi Bennis, Arjun Nitin Bhagoji, Kallista" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "20100" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 76, + 72, + 296, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 296, + 116 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 296, + 116 + ], + "type": "text", + "content": "Bonawitz, Zachary Charles, Graham Cormode, Rachel Cummings, et al. Advances and open problems in federated learning. Foundations and trends® in machine learning, 14(1-2): 1-210, 2021. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 296, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 296, + 173 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 296, + 173 + ], + "type": "text", + "content": "[28] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 1, 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 175, + 296, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 296, + 218 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 296, + 218 + ], + "type": "text", + "content": "[29] Simon Kornblith, Jonathon Shlens, and Quoc V Le. Do better imagenet models transfer better? In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2661-2671, 2019. 2, 7, 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 220, + 295, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 220, + 295, + 241 + ], + "spans": [ + { + "bbox": [ + 56, + 220, + 295, + 241 + ], + "type": "text", + "content": "[30] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 243, + 295, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 243, + 295, + 296 + ], + "spans": [ + { + "bbox": [ + 56, + 243, + 295, + 296 + ], + "type": "text", + "content": "[31] Yuhang Li, Ruihao Gong, Xu Tan, Yang Yang, Peng Hu, Qi Zhang, Fengwei Yu, Wei Wang, and Shi Gu. Brecq: Pushing the limit of post-training quantization by block reconstruction. In International Conference on Learning Representations. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 298, + 295, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 298, + 295, + 342 + ], + "spans": [ + { + "bbox": [ + 56, + 298, + 295, + 342 + ], + "type": "text", + "content": "[32] Chenxi Liu, Lixu Wang, Lingjuan Lyu, Chen Sun, Xiao Wang, and Qi Zhu. Deja vu: Continual model generalization for unseen domains. In The Eleventh International Conference on Learning Representations, 2023. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 343, + 295, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 343, + 295, + 398 + ], + "spans": [ + { + "bbox": [ + 56, + 343, + 295, + 398 + ], + "type": "text", + "content": "[33] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 6, 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 399, + 295, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 399, + 295, + 442 + ], + "spans": [ + { + "bbox": [ + 56, + 399, + 295, + 442 + ], + "type": "text", + "content": "[34] Zhenhua Liu, Yunhe Wang, Kai Han, Wei Zhang, Siwei Ma, and Wen Gao. Post-training quantization for vision transformer. Advances in Neural Information Processing Systems, 34:28092-28103, 2021. 4, 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 445, + 295, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 445, + 295, + 476 + ], + "spans": [ + { + "bbox": [ + 56, + 445, + 295, + 476 + ], + "type": "text", + "content": "[35] Alejandro López-Cifuentes, Marcos Escudero-Vinolo, Jesús Bescós, and Álvaro García-Martín. Semantic-aware scene recognition. Pattern Recognition, 102:107256, 2020. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 478, + 295, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 478, + 295, + 511 + ], + "spans": [ + { + "bbox": [ + 56, + 478, + 295, + 511 + ], + "type": "text", + "content": "[36] Payal Mohapatra, Lixu Wang, and Qi Zhu. Phase-driven domain generalizable learning for nonstationary time series. arXiv preprint arXiv:2402.05960, 2024. 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 513, + 295, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 513, + 295, + 555 + ], + "spans": [ + { + "bbox": [ + 56, + 513, + 295, + 555 + ], + "type": "text", + "content": "[37] Mehdi Noroozi and Paolo Favaro. Unsupervised learning of visual representations by solving jigsaw puzzles. In European conference on computer vision, pages 69-84. Springer, 2016. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 558, + 295, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 558, + 295, + 610 + ], + "spans": [ + { + "bbox": [ + 56, + 558, + 295, + 610 + ], + "type": "text", + "content": "[38] Dario Pasquini, Giuseppe Ateniese, and Massimo Bernaschi. Unleashing the tiger: Inference attacks on split learning. In Proceedings of the 2021 ACM SIGSAC Conference on Computer and Communications Security, pages 2113-2129, 2021. 1, 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 613, + 295, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 613, + 295, + 667 + ], + "spans": [ + { + "bbox": [ + 56, + 613, + 295, + 667 + ], + "type": "text", + "content": "[39] Xingchao Peng, Qinxun Bai, Xide Xia, Zijun Huang, Kate Saenko, and Bo Wang. Moment matching for multi-source domain adaptation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1406-1415, 2019. 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 670, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 295, + 713 + ], + "type": "text", + "content": "[40] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervi" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "text", + "content": "sion. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 96, + 555, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 555, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 555, + 150 + ], + "type": "text", + "content": "[41] K Shailaja, Banoth Seetharamulu, and MA Jabbar. Machine learning in healthcare: A review. In 2018 Second international conference on electronics, communication and aerospace technology (ICECA), pages 910-914. IEEE, 2018. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 152, + 553, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 553, + 207 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 553, + 207 + ], + "type": "text", + "content": "[42] Chandra Thapa, Mahawaga Arachchige Pathum Chamikara, and Seyit A Camtepe. Advancements of federated learning towards privacy preservation: from federated learning to split learning. Federated Learning Systems: Towards Next-Generation AI, pages 79–109, 2021. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 209, + 553, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 209, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 316, + 209, + 553, + 262 + ], + "type": "text", + "content": "[43] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In International conference on machine learning, pages 10347-10357. PMLR, 2021. 6, 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 264, + 553, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 264, + 553, + 308 + ], + "spans": [ + { + "bbox": [ + 316, + 264, + 553, + 308 + ], + "type": "text", + "content": "[44] Praneeth Vepakomma, Otkrist Gupta, Tristan Swedish, and Ramesh Raskar. Split learning for health: Distributed deep learning without sharing raw patient data. arXiv preprint arXiv:1812.00564, 2018. 1, 2, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 309, + 553, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 309, + 553, + 364 + ], + "spans": [ + { + "bbox": [ + 316, + 309, + 553, + 364 + ], + "type": "text", + "content": "[45] Jingye Wang, Ruoyi Du, Dongliang Chang, Kongming Liang, and Zhanyu Ma. Domain generalization via frequency-domain-based feature disentanglement and interaction. In Proceedings of the 30th ACM International Conference on Multimedia, pages 4821-4829, 2022. 4, 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 365, + 555, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 365, + 555, + 408 + ], + "spans": [ + { + "bbox": [ + 316, + 365, + 555, + 408 + ], + "type": "text", + "content": "[46] Lixu Wang, Shichao Xu, Xiao Wang, and Qi Zhu. Addressing class imbalance in federated learning. In Proceedings of the AAAI conference on artificial intelligence, pages 10165-10173, 2021. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 411, + 553, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 411, + 553, + 464 + ], + "spans": [ + { + "bbox": [ + 316, + 411, + 553, + 464 + ], + "type": "text", + "content": "[47] Lixu Wang, Shichao Xu, Ruiqi Xu, Xiao Wang, and Qi Zhu. Non-transferable learning: A new approach for model ownership verification and applicability authorization. In International Conference on Learning Representation (ICLR-2022), 2022. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 466, + 553, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 466, + 553, + 510 + ], + "spans": [ + { + "bbox": [ + 316, + 466, + 553, + 510 + ], + "type": "text", + "content": "[48] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 512, + 553, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 512, + 553, + 543 + ], + "spans": [ + { + "bbox": [ + 316, + 512, + 553, + 543 + ], + "type": "text", + "content": "[49] Guangxuan Xiao, Ji Lin, and Song Han. Offsite-tuning: Transfer learning without full model. arXiv preprint arXiv:2302.04870, 2023. 1, 2, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 545, + 553, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 545, + 553, + 599 + ], + "spans": [ + { + "bbox": [ + 316, + 545, + 553, + 599 + ], + "type": "text", + "content": "[50] Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. Advances in neural information processing systems, 34: 12077-12090, 2021. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 601, + 553, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 601, + 553, + 645 + ], + "spans": [ + { + "bbox": [ + 316, + 601, + 553, + 645 + ], + "type": "text", + "content": "[51] Chengming Xu, Siqian Yang, Yabiao Wang, Zhanxiong Wang, Yanwei Fu, and Xiangyang Xue. Exploring efficient few-shot adaptation for vision transformers. Transactions on Machine Learning Research. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "type": "text", + "content": "[52] Peng Xu, Xiatian Zhu, and David A Clifton. Multimodal learning with transformers: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(10):12113-12132, 2023. 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "text", + "content": "[53] Xiaoyang Xu, Mengda Yang, Wenzhe Yi, Ziang Li, Juan Wang, Hongxin Hu, Yong Zhuang, and Yaxin Liu. A stealthy" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20101" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 339 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 75, + 72, + 296, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 72, + 296, + 116 + ], + "spans": [ + { + "bbox": [ + 75, + 72, + 296, + 116 + ], + "type": "text", + "content": "wrongdoer: Feature-oriented reconstruction attack against split learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12130-12139, 2024. 1, 2, 7, 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 296, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 296, + 172 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 296, + 172 + ], + "type": "text", + "content": "[54] Yupeng Yin, Xianglong Zhang, Huanle Zhang, Feng Li, Yue Yu, Xiuzhen Cheng, and Pengfei Hu. Ginver: Generative model inversion attacks against collaborative inference. In Proceedings of the ACM Web Conference 2023, pages 2122-2131, 2023. 1, 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 174, + 295, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 174, + 295, + 228 + ], + "spans": [ + { + "bbox": [ + 56, + 174, + 295, + 228 + ], + "type": "text", + "content": "[55] Zhihang Yuan, Chenhao Xue, Yiqi Chen, Qiang Wu, and Guangyu Sun. Ptq4vit: Post-training quantization for vision transformers with twin uniform quantization. In European conference on computer vision, pages 191-207. Springer, 2022. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 229, + 295, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 229, + 295, + 284 + ], + "spans": [ + { + "bbox": [ + 56, + 229, + 295, + 284 + ], + "type": "text", + "content": "[56] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 285, + 295, + 339 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 285, + 295, + 339 + ], + "spans": [ + { + "bbox": [ + 56, + 285, + 295, + 339 + ], + "type": "text", + "content": "[57] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In Proceedings of the IEEE international conference on computer vision, pages 2223-2232, 2017. 2" + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20102" + } + ] + } + ], + "index": 6 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Spotting the Unexpected (STU)_ A 3D LiDAR Dataset for Anomaly Segmentation in Autonomous Driving/7019444b-0523-423c-8391-a6c740f37e7c_content_list.json b/2025/Spotting the Unexpected (STU)_ A 3D LiDAR Dataset for Anomaly Segmentation in Autonomous Driving/7019444b-0523-423c-8391-a6c740f37e7c_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..6240e178cbbf5506456a8eaa4f9f4e537b181707 --- /dev/null +++ b/2025/Spotting the Unexpected (STU)_ A 3D LiDAR Dataset for Anomaly Segmentation in Autonomous Driving/7019444b-0523-423c-8391-a6c740f37e7c_content_list.json @@ -0,0 +1,1715 @@ +[ + { + "type": "text", + "text": "Spotting the Unexpected (STU): A 3D LiDAR Dataset for Anomaly Segmentation in Autonomous Driving", + "text_level": 1, + "bbox": [ + 91, + 130, + 906, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Alexey Nekrasov1", + "bbox": [ + 155, + 202, + 302, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Malcolm Burdorf1", + "bbox": [ + 346, + 202, + 496, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Stewart Worrall2", + "bbox": [ + 540, + 203, + 676, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Bastian Leibe1", + "bbox": [ + 720, + 203, + 839, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Julie Stephany Berrio Perez", + "bbox": [ + 382, + 220, + 612, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ RWTH Aachen University, $^{2}$ The University of Sydney", + "bbox": [ + 276, + 238, + 725, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "https://vision.rwth-aachen.de/stu-dataset", + "bbox": [ + 253, + 272, + 741, + 289 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2c7d34edda03a4ca0ede6baff2084b80cb95434b140a7e10f268087ae07cafb1.jpg", + "image_caption": [ + "Figure 1. We present Spotting the Unexpected (STU) a novel anomaly segmentation dataset for autonomous driving. The dataset contains semantic and instance labels for out-of-distribution (OOD) objects, and includes surround-view setup with synchronized cameras." + ], + "image_footnote": [], + "bbox": [ + 91, + 321, + 906, + 527 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 578, + 326, + 593 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To operate safely, autonomous vehicles (AVs) need to detect and handle unexpected objects or anomalies on the road. While significant research exists for anomaly detection and segmentation in 2D, research progress in 3D is underexplored. Existing datasets lack high-quality multimodal data that are typically found in AVs. This paper presents a novel dataset for anomaly segmentation in driving scenarios. To the best of our knowledge, it is the first publicly available dataset focused on road anomaly segmentation with dense 3D semantic labeling, incorporating both LiDAR and camera data, as well as sequential information to enable anomaly detection across various ranges. This capability is critical for the safe navigation of autonomous vehicles. We adapted and evaluated several baseline models for 3D segmentation, highlighting the challenges of 3D anomaly detection in driving environments. Our dataset and evaluation code will be openly available, facilitating the testing and performance comparison of different approaches.", + "bbox": [ + 88, + 611, + 485, + 883 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 514, + 578, + 643, + 593 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Autonomous vehicles have become an increasingly important part of modern transportation systems. Despite significant advances, these vehicles continue to face challenges in accurately detecting unexpected objects on the road. One particular challenge is road debris, which refers to objects that are not typically part of the roadway environment and pose significant safety risks. According to the AAA Foundation for Traffic Safety [53], road debris was a contributing factor in more than 50,000 crashes, 10,000 injuries, and 125 fatalities annually between 2011 and 2014. These objects can vary widely in size and appearance, making them particularly challenging to detect. Addressing this issue is critical for ensuring the safe and reliable operation of autonomous vehicles, as it directly impacts their ability to handle anomalies in real-world driving conditions.", + "bbox": [ + 511, + 608, + 906, + 835 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "State-of-the-art anomaly segmentation and anomaly detection methods have received substantial attention in recent years [17, 37, 41], especially in the context of autonomous driving [11, 16, 18, 21, 31, 54]. One of the first", + "bbox": [ + 511, + 839, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "11875", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "benchmarks for evaluating anomaly segmentation is Lost and Found [45], whose data serves as the basis for many complex benchmarks [4, 10]. These benchmarks provide a wide range of possible road anomalies and obstacles, and recent research has focused on exploring the temporal [38] and multimodal [6, 36] nature of anomaly segmentation.", + "bbox": [ + 89, + 90, + 480, + 181 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "While current autonomous driving datasets [1, 7, 51] use a setup with surround view cameras, radar, and high-resolution LiDAR sensors, existing anomaly segmentation and detection benchmarks focus primarily on image data [5] and mostly use Cityscapes [15] as the in-distribution dataset. Image-first benchmarks often lack the comprehensive sensor setup found in modern autonomous vehicles, and overlook the use of 3D modality for anomaly segmentation. Recent research that has begun to investigate the use of LiDAR [49] sensor for anomaly segmentation and detection, lacks high-resolution sensors found in modern autonomous vehicles. Several approaches have sought to find anomalies in validation sets of existing datasets [36, 46] or have created datasets using simulation [6] due to the difficulty of collecting and annotating such data in the real world. Despite these efforts, extending existing data or simulating anomalies only partially addresses the lack of accessible and high-quality data, shown in Table 1.", + "bbox": [ + 91, + 185, + 482, + 455 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address this issue, we analyzed the labels of two widely used in-distribution (ID) datasets [1, 7] to identify an appropriate set of anomalies that would not overlap with objects in the ID training or validation datasets. This informed our creation of a real-world dataset containing challenging anomaly objects placed on roads, with annotated sequences in 3D LiDAR point clouds. Our dataset consists of multiple high-resolution LiDAR sequences acquired in challenging environments, with point-level labels for individual anomaly objects (shown in Figure 1). We acquired data in both rural and urban environments, with objects of varying degrees of difficulty, and include a number of real-world anomalies found on public roads. In addition, to evaluate in-distribution performance, we annotated additional sequences without anomalies with inlier labels to train and evaluate models to ensure a small domain gap between our and SemanticKITTI [1] datasets.", + "bbox": [ + 89, + 459, + 482, + 715 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To establish simple baselines for 3D LiDAR segmentation, we adapted the methods used in 2D anomaly segmentation [4, 24, 34, 41, 50] and applied them to the 3D domain, investigating anomaly segmentation methods for 3D LiDAR point clouds. Our models, trained jointly on the SemanticKITTI dataset and our own dataset, show strong segmentation performance for in-distribution data and novel additional sequences. However, we observed that the baseline methods struggled with out-of-distribution objects, and models tend to predict anomaly objects as inliers with a high degree of confidence. We believe that our dataset will benefit the community by allowing the evaluation of anomaly", + "bbox": [ + 89, + 719, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "segmentation methods in the LiDAR domain and further exploration of multimodal anomaly segmentation methods.", + "bbox": [ + 511, + 90, + 903, + 121 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions provide a basis for the development of more advanced anomaly segmentation methods in the 3D LiDAR domain. Specifically, we propose a new anomaly segmentation dataset, and perform an analysis of the recorded and annotated data. The purpose of the paper is to provide practitioners with the necessary tools to evaluate their anomaly segmentation models. We also provide a set of baselines that can serve as a reference for future challenges and research directions in this area.", + "bbox": [ + 511, + 123, + 906, + 258 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 513, + 275, + 651, + 290 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Anomaly Datasets", + "text_level": 1, + "bbox": [ + 513, + 301, + 689, + 316 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Several multimodal datasets for autonomous driving are available, including Waymo [51], nuScenes [7], SemanticKITTI [1], among others. These datasets were collected in urban environments under typical driving conditions. Although these datasets are relevant for evaluating various tasks related to autonomous vehicle operation, they lack data for uncommon scenarios. One such scenario involves the presence of foreign objects on the road. Although this situation is rare during typical driving, autonomous vehicles must be able to recognize and respond to it safely. Here, assigning a unique label to each distinct object is impractical, since datasets have a limited set of labels, and objects on the road can vary widely. These objects are often grouped under a general category like \"other object\" and are excluded from the benchmarks due to their variability.", + "bbox": [ + 511, + 324, + 903, + 551 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In recent years, several datasets have been introduced to assess the detection of foreign objects on the road [4, 10, 25, 36, 38, 42, 45, 59]. Some of these datasets have been developed using data augmentation techniques or simulations, but only a few are based on data collected with real-world equipment in real driving scenarios. Simulations and augmentations are great tools, but they lack realism in terms of the sensor model [4] and the sizes of the objects [25]. However, the aforementioned datasets are limited to image data only. Autonomous vehicles use LiDAR sensors because they offer greater range, robustness to lighting conditions, and more precise 3D information about the environment compared to monocular depth estimations or stereo cameras. Therefore, incorporating 3D information is essential for accurately assessing road anomalies.", + "bbox": [ + 511, + 551, + 905, + 779 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Three datasets that integrate camera and LiDAR data are SOD [22], TOR4D [58] and CODA [36]. The CODA series is derived from validation sets of existing autonomous vehicle datasets. Still, it has the limitation that anomalous objects (e.g., traffic cones, barriers) may also appear in the training set, which is acceptable for the task of open-set segmentation, but contradicts the task of anomaly segmentation. Recently, a few methods for 3D anomaly detection", + "bbox": [ + 511, + 780, + 905, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "11876", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "table", + "img_path": "images/fb8a0eba4b9b3402acfe102e76cc8d6e477ed077ac6d320aaeecfc3d1c6e8fba.jpg", + "table_caption": [ + "Table 1. Comparison of our dataset with other community datasets [5]. SM = Semantic Mask, IM=Instance Mask, BB=Bounding Box" + ], + "table_footnote": [], + "table_body": "
DatasetCameraLiDAR \nBeamsSource#ODDLabelSize \nTest/ValTemporalEnvironments \nTest SequencesRelated Dataset
Fishscapes [3, 4]
FS Lost and Found1 RGBXStaged1SM275/100Xn/aLost and Found [45]
FS Static1 RGBXAugmented1SM1000/30Xn/aCityScapes [15]
CAOS [25]
StreetHazards1 RGBXSimulated1SM1500✓2 sim towns - 150 seq
BDD-Anomaly1 RGBXValidation Set3SM810Xn/aBDD100K [61]
SegmentMelfYouCan [10]
RoadAnomaly211 RGBXWeb Sourced1SM100/10Xn/a
RoadObstacle211 RGBXStaged1SM382/30✓9 streets - 90 seq
CODA [36]
CODA-KITTI1 RGB✓-64Validation Set6BB309Xn/aKITTI [19]
CODA-nuScenes1 RGB✓-32Validation Set17BB134Xn/aNuScenes [7]
CODA-ONCE1 RGB✓-40Validation Set32BB1057Xn/aONCE [39]
CODA22-ONCE1 RGB✓-40Validation Set29BB717Xn/aONCE [39]
CODA22-SODA10M1 RGBXValidation Set29BB4167Xn/aSODA [23]
Wuppertal ODD [38]
SOS1 RGBX-distStaged13IM1129✓9 streets - 20 seq
CWL1 RGBX-distSimulated18IM1210✓26 streets - 26 seq
Lost and Found [45]1 StereoXStaged42SM1068✓5 streets - 5 seq
SOD [49]1 RGB✓-16Staged1SM460/530X2 streets - 5 seq
WD-Pascal [2]1 RGBXSimulated1SM70Xn/a
Vistas-NP [20]1 RGBXClass exclusion4SM11167Xn/aMapillary Vistas [43]
STU (Ours)8 RGB✓-128Staged1IM8022/1960✓6 streets - 51 seqSemKITTI [1], P-CUDAL [55]
", + "bbox": [ + 91, + 114, + 906, + 375 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "have been proposed [28, 33, 46]. However, these methods use validation sequences to evaluate anomaly detection capabilities. The SOD dataset offers depth information for small road obstacles using a low-resolution 16-beam LiDAR, and that introduces a large domain gap between the training and test data. TOR4D and Rare4D [58] are the only datasets that contain real-world LiDAR data to benchmark road anomalies; however, these datasets are proprietary and not accessible to the public. We release out-of-distribution validation and in-distribution training sequences with a permissive license and open the private test set to submissions from the public.", + "bbox": [ + 88, + 398, + 480, + 580 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Open-Set Segmentation in 3D", + "text_level": 1, + "bbox": [ + 89, + 590, + 354, + 607 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Open-set and unsupervised segmentation are closely related to the out-of-distribution detection task. Multiple methods were proposed for unsupervised segmentation [44, 48] and open-set prediction in 3D [9, 58]. Recent evaluations in 2D anomaly segmentation [18] have demonstrated that methods such as OSIS [58] and EOPSN [30] perform poorly on anomaly objects. This weakness stems from open-set methods' reliance on the presence of unknown instances inside the void regions during training, which may not generalize well to real-world scenarios. Similarly, MLUC [8] assesses open-set performance on the UDI [8] dataset, but this dataset is also not publicly accessible and makes the same assumptions as OSIS methods. In addition, open-set methods typically remove the ground plane [44] to obtain clusters of potential objects, but ground plane removal [35] methods are often sensitive to chosen hyperparameters and yield noisy results, missing many smaller objects. For the evaluation of anomaly segmentation, it is important to have a test dataset that is free of anomaly objects which might", + "bbox": [ + 91, + 613, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "occur during training. In contrast to open-set methods, our work focuses on collecting and densely annotating real-world anomaly objects, ensuring that there is no intersection between in-distribution training sets and out-of-distribution test sets.", + "bbox": [ + 511, + 400, + 906, + 474 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Anomaly Segmentation in Autonomous Driving", + "text_level": 1, + "bbox": [ + 511, + 487, + 906, + 503 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To tackle the problem of anomaly segmentation, multiple works [4, 5, 10] have recently been proposed for the problem of 2D anomaly segmentation. Earlier works [4, 11, 21, 24, 34, 37, 54] were applied to per-pixel segmentation models such as DeepLab [12]; however, recently many works were focused on MaskFormer-style models [13, 14]. A line of work [16, 41, 47] proposed different methods to adapt Mask2Former, exploring the mechanism of mask prediction, where multiple masks might ignore certain regions of an image. These models provide more reliable anomaly scoring mechanisms not limited to per-pixel predictions and achieve higher performance on the anomaly segmentation task. For our benchmark, we decided to evaluate methods that also use MaskFormer-style models and investigate the application of the Mask4Former [60] model for 3D anomaly segmentation. We obtain all our scores using a MaskFormer-Style model, laying the foundation for future baselines.", + "bbox": [ + 511, + 510, + 908, + 781 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. The Spotting the Unexpected (STU) Dataset", + "text_level": 1, + "bbox": [ + 511, + 799, + 903, + 816 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our data collection platform consists of a rigid frame mounted on a vehicle and equipped with eight hardware-triggered cameras and a LiDAR sensor to ensure synchronized data collection. The LiDAR and the cameras were calibrated using the method described in [56], with the cal", + "bbox": [ + 511, + 825, + 905, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "11877", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/badb56129a8a822729d71d87dec28a53824536a8b36d206976c7f4f4870ccffb.jpg", + "image_caption": [ + "(a) Naturalistic data collection." + ], + "image_footnote": [], + "bbox": [ + 98, + 92, + 486, + 172 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/422643e6cc0df06d09236827c709acec0b752a218bcbf68049418c65e74b86fc.jpg", + "image_caption": [ + "(b) Data collection in controlled environment.", + "Figure 2. Data collection conducted in a naturalistic manner (a) and controlled environment (b) with objects on the road." + ], + "image_footnote": [], + "bbox": [ + 98, + 195, + 486, + 276 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ization process repeated for each camera. This setup follows the configuration discussed in Panoptic-CUDAL [55]. We refer to the Supplementary Material for detailed information on the vehicle setup.", + "bbox": [ + 89, + 363, + 482, + 422 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Data collection.", + "text_level": 1, + "bbox": [ + 89, + 434, + 245, + 449 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The data was collected using two sets of conditions: one in a naturalistic environment where objects were found on public roads, and one in a controlled environment. In the naturalistic scenario, several hours of driving data were recorded, with specific events marked when objects appeared on the road. Figure 2a shows a bucket on the road recorded while driving on a highway.", + "bbox": [ + 89, + 455, + 482, + 561 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The controlled environment data collection protocol was rigorous for safety reasons. We employed four people: two spotters positioned at either end of a low-traffic and low-speed road, one responsible for placing objects on the road, and a driver. The spotters would signal when it was safe to place the objects, and once positioned, the driver would begin recording the scene. If one of the spotters reported an approaching vehicle, the data recording was stopped, and the object was immediately removed. This was done to avoid disrupting the normal flow of traffic. The vehicle was driven at a maximum speed of $50\\mathrm{km / h}$ during data recording. Figure 2b shows a computer monitor placed on the road during sensor data logging. Some of the objects placed on the road are also shown in Figure 3.", + "bbox": [ + 89, + 563, + 482, + 773 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Postprocessing. After data collection, post-processing is required to obtain vehicle poses and anonymize the images. Point cloud registration was performed using KISS ICP [57], a LiDAR odometry pipeline. KISS ICP includes point cloud motion compensation, subsampling, adaptive thresholding to determine correspondences, and LiDAR pose estimation. The calculated LiDAR pose was then exported in KITTI format as required for the labeling tool [1].", + "bbox": [ + 89, + 780, + 482, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Ethical considerations for this dataset include anonymizing camera images. Identifiable data, such as faces and license plates, are processed using DashcamCleaner [52] and DeepPrivacy2 [29]. DashcamCleaner uses a license plate detector to locate and blur license plates, while DeepPrivacy2 detects facial features and generates new, unidentifiable faces to replace the original ones. Examples of anonymization can be found in the supplementary material.", + "bbox": [ + 511, + 90, + 903, + 212 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Labeling", + "text_level": 1, + "bbox": [ + 511, + 219, + 619, + 234 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We define anomalies as objects that may pose a danger to the autonomous vehicle and its passengers, especially those present on the driving surface and not present in the training dataset. In this work, we focus on road anomalies and do not consider unusual situations such as vehicle collisions or unexpected driver behavior, and we focus on the fine-grained anomaly segmentation. In particular, we follow previous work [4, 45] and focus on objects that are not present in the training set.", + "bbox": [ + 511, + 241, + 903, + 376 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We labeled the data using the SemanticKITTI labeler [1]. To generate initial labels for the scenes, we used pseudolabeling techniques using a pre-trained SemanticKITTI model. We instruct our annotators to group all points belonging to inliers as a single semantic class, questionable points as unlabeled, and anomaly points as the outlier class. We instruct our labelers that it is very important to label inliers and outliers carefully while allowing for larger unlabeled regions where the annotator may be challenged.", + "bbox": [ + 511, + 377, + 905, + 513 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Annotation process. The domain gap between urban environments and SemanticKITTI street scenes is minimal, especially in regions close to the vehicle. This similarity allows the acquisition of reliable initial estimates. The annotation process was performed by three annotators, each spending eight hours a week over several months. Each annotator spent approximately four hours on the initial labeling of each scene and an additional two hours on refining the final labels. To ensure accuracy and consistency, the annotators cross-validated each other's work. To further improve the quality of the labels, we examined the predictions of the baseline methods to see if any known objects were missed. In particular, current methods tend to be highly sensitive to objects that are present but ignored during training, such as those classified in the category \"other object\". We annotate these objects as unlabeled and ignore them in the evaluations, as they could alter the results of the anomaly segmentation methods.", + "bbox": [ + 511, + 518, + 905, + 789 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Differences to SemanticKITTI. In contrast to SemanticKITTI, which provides labels up to a range of 50 meters, we labeled all points visible to the LiDAR sensor. This was facilitated by using a LiDAR with a higher vertical resolution. The maximum distance resulting from an anomaly in our dataset is 150 meters, allowing for comprehensive evaluations. However, since we train on the Se", + "bbox": [ + 511, + 795, + 903, + 898 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "11878", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/db4ed5b85024c82a54f0d517cd1da0f42888b59e123237cbeaf82741f84c17d0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 99, + 88, + 230, + 152 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/eaa1891ae8b0867e8ef46c72efe3c1c98265bbdbe5f36f5acf85fa8b47f3e92e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 233, + 89, + 364, + 151 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/992a78b49d4cb3ac015ecbb803e907eca49e7a91ef35285667a07c5b649b1ead.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 367, + 90, + 496, + 151 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/4296654edfbc219d1703f6fe849d0d4dba4fae7f18e3d2edb82240d739641bd6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 90, + 630, + 151 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/6e3d77629b6f9024d72fbffc971969fb91123ddca990a6608b8849691aca530b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 633, + 90, + 763, + 151 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/d1eacadf2ddd0b34cb990a3ed4079bcbe30d70cf938dc0262a75662d5afa5fc2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 766, + 90, + 897, + 151 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/ae700115971280b9c81f5f5d07c28515d56012ec5de26a2555fadc1537a01bc7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 99, + 152, + 230, + 215 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/2fce046319925b5aa325f021b22267fe460061bba12dba07f87b256bafadd2c8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 233, + 152, + 362, + 215 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/464c32cce11ea34ee67ded114a1d22365275f8af1560b318f6f0b11e21ccd728.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 367, + 152, + 496, + 215 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/92700289051bba58257434daa2f1401f651172f2c896499b74b36ced527e78c3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 152, + 630, + 215 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/1840258f7a63031bd57972978bb92ea801de1b829b36d53f01b86e9327fd6ce2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 633, + 152, + 763, + 215 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/71276fa0f6e972dd75663d194e7e4e195a916e2b68bc3f910b03293d7d99258c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 766, + 152, + 897, + 215 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/6d7b36eac6babb141c0896375df5963cbf5a5c666244ec3b258085688f3268c4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 99, + 215, + 230, + 279 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/7cc67798adcd0837f4f029bed306742399587163a34c441f7649dc300a752f96.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 233, + 215, + 362, + 279 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/13870c7795dfe6c37c9f5e194aa4fdeeab6682eae3fd774c5b298d753a3ab2a6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 367, + 215, + 496, + 279 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/be03b39e639b38d8a17e342688b188f7b7a089628767adcdd8260b77cae4ea53.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 215, + 630, + 279 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/4485c45f4e2e4eb0b5815d215d3eb2d8f47a9a397616274276b1a33b27e1e067.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 633, + 215, + 763, + 279 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/e274c448a26dc21a5f1029348cb97c289d32dc57962fc8497bae9b0d84a20573.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 766, + 215, + 897, + 279 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/5b55c2119689c1aac9a5f19d06936a4e4bb5bb0931958a673d9620b250db1fbc.jpg", + "image_caption": [ + "Figure 3. Different anomalies in the STU dataset. Different objects on the road used for staged data collection. We pick objects such that have no intersection with the inlier dataset and place them on roads in different locations and illumination conditions. Objects might touch each other, be very small, as large as a chair or a surf board, and could cause an accident if a car would drive over them." + ], + "image_footnote": [], + "bbox": [ + 99, + 280, + 230, + 342 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/b0d68c0089f1d047d0d14e3fe243a243aa92a8e3038bfe409220c84baa998cb5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 233, + 280, + 362, + 342 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/81ae7030ddd3e101f381e4052f854d29d1e31a88a5f3401eee1f2c5bfdf6906d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 367, + 280, + 496, + 342 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/6cf8158d2c34d7bac0194cc961fd67cbfae9a8ff510ad38ae0829d9ee4009de6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 280, + 630, + 342 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/f95f6e533d073052d9d51a0406fbe9260dcfc36481ab0f5a03ebaa2d727094e5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 633, + 280, + 763, + 342 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/7dbff6d23082585447662b9eeae32681468170763bf26c7c14b2ed508a516d32.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 766, + 280, + 897, + 342 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "manticKITTI dataset, which contains only labels up to 50 meters, we limit our evaluations to the 50 meter range. The sensor setup has 128 LiDAR beams, compared to 64 beams in SemanticKITTI.", + "bbox": [ + 89, + 422, + 482, + 483 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Dataset analysis", + "text_level": 1, + "bbox": [ + 89, + 497, + 250, + 513 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We recorded 70 sequences with anomalies, which we fully annotated for anomaly segmentation. Two of these labeled sequences were captured in a naturalistic manner, while the remaining 68 were staged to include various anomalies. We designated 19 sequences as a validation set and 51 as a closed test set for benchmarking purposes. In addition, we recorded and annotated two sequences with no anomalies for training and validation of the inlier performance \"STU-inlier\", to reduce the domain gap between our dataset and SemanticKITTI.", + "bbox": [ + 89, + 520, + 482, + 669 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our dataset defines two primary classes: inliers and outliers. Inliers encompasses classes that are used to train a model, and contains objects expected on the road, such as vehicles, pedestrians, and road infrastructure. Outliers, on the other hand, refer to out-of-distribution elements or anomalies that are atypical or unexpected in the driving scene, shown in Figure 3. One additional class, unlabeled is ignored in the evaluation and typically includes classes that the model has seen during training, but was not supervised in, such as parking meters, utility boxes, and lamps. We provide anomaly annotations at both the instance and semantic levels, enabling a comprehensive evaluation of models for semantic and instance segmentation in 3D environments.", + "bbox": [ + 89, + 672, + 482, + 882 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Each outlier object in our dataset has an instance la-", + "bbox": [ + 109, + 885, + 482, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "bel unique for a sequence. Figure 4 presents histograms that display the distribution of key properties for these instances, including the average and maximum number of LiDAR points per instance, instance heights, and the number of instances per scan. This analysis provides insights into the structural characteristics and variability of data within point clouds. Specifically, Figure 4 indicates that most instances have very few LiDAR points on average, but a sequence may have multiple instances per sequence.", + "bbox": [ + 511, + 422, + 903, + 559 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Furthermore, Figure 5 presents a scatter plot of detected anomalies within the vehicle frame, highlighting a higher density of anomalies along the longitudinal $(X)$ axis both in front and behind the vehicle. The accompanying histograms depict the distribution of the anomalies along each axis, noting that the anomalies are identified at distances ranging up to 150 meters in front of the vehicle. Figure 6 explores the relationship between the distance to an anomaly and its size, showing that the proportion of LiDAR point cloud obstacles decreases as the distance increases. This trend is due to the radial nature of LiDAR, which has lower vertical resolution at greater ranges. For benchmarking purposes, we focus exclusively on instances with more than five LiDAR points, highlighted in blue, ensuring that the evaluation targets detectable anomalies.", + "bbox": [ + 511, + 560, + 905, + 786 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Evaluated Baselines", + "text_level": 1, + "bbox": [ + 511, + 801, + 702, + 816 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The formulation of the anomaly segmentation task restricts training to avoid direct supervision of OOD objects. Generally, OOD objects are not present in the training set, but could be present during inference [18]. Since there are no methods that could be applied directly for our setup, we", + "bbox": [ + 511, + 825, + 903, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "11879", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/776e849cd15d0db2f2a4ad196bc52ccec72beafd6d8398d0bd31966b65966389.jpg", + "image_caption": [ + "(a) Avg. num of points/sequence." + ], + "image_footnote": [], + "bbox": [ + 107, + 90, + 297, + 196 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ece6301e8fd8eeb1d66ad5dfc22bc445b171effe6e1d6be177a0cdc06318fa80.jpg", + "image_caption": [ + "Figure 4. Anomaly Instance Properties. A typical recorded anomaly has on average less than 50 points per sequence (a), with less than 300 (b) points at maximum, and a maximum height below one meter (c). We record up to nine individual anomaly instances in the same sequence (d)." + ], + "image_footnote": [], + "bbox": [ + 308, + 92, + 491, + 196 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d17ca69cf4e26cc946d2185c567fae78f376de4b48f7cedfce782756d93391d1.jpg", + "image_caption": [ + "(b) Max. points per instance." + ], + "image_footnote": [], + "bbox": [ + 504, + 92, + 691, + 196 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1b5a0783346b2817ba682da757f107ca42013d480cfa98e72ba1a3d589e83b99.jpg", + "image_caption": [ + "(c) Max. height of instances (m).", + "(d) Numb. of instances/sequence." + ], + "image_footnote": [], + "bbox": [ + 700, + 93, + 885, + 196 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/641ba3252959ed7b996229cd237db7c621162e36929ebcdef64faffb36fd6ae2.jpg", + "image_caption": [ + "Figure 5. Distribution of anomalies along the vehicle's reference frame. Most of the points appear around the vehicle." + ], + "image_footnote": [], + "bbox": [ + 94, + 290, + 485, + 459 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a0dd9871c54e0d67be1c93fb2eb0bf007782db7fcad52a7c47004efe7f9cc8c0.jpg", + "image_caption": [ + "Figure 6. We follow SemanticKitti's maximum distance threshold and evaluate instances that are within a 50 meter radius of the vehicle. In addition, we constrain evaluation of point and object level metrics to a minimum of 5 points." + ], + "image_footnote": [], + "bbox": [ + 94, + 508, + 483, + 728 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "adopt several baseline methods from the 2D anomaly segmentation to 3D LiDAR segmentation. First, we evaluate simple baselines, such as Max-Logit [24], Monte-Carlo Dropout [50], and Deep Ensembles [34]. Secondly, we train one of the models to predict ignored regions during train", + "bbox": [ + 89, + 824, + 483, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ing, which could be used to classify “void” points during testing [4]. We also adapt RbA [41], one of the state-of-the-art methods for image anomaly segmentation, to the 3D LiDAR segmentation method Mask4Former-3D [60].", + "bbox": [ + 511, + 289, + 903, + 349 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Deep Ensembles, MC Dropout, and MaxLogit. Deep Ensembles [34] and MC Dropout [50] are very common methods that practitioners use due to ease of implementation and application. For Deep Ensembles, we train three Mask4Former-3D [60] models. For MC dropout, we train one Mask4Former-3D model with enabled dropout in the transformer decoder for self- and cross-attention operations, and during inference we repeat the transformer decoder operation 50 times. As a simple baseline that uses only one model and a single forward pass, we use the MaxLogit [24] method.", + "bbox": [ + 511, + 356, + 906, + 521 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "RbA [41] is a method that uses the ability of MaskFormer-Style models [13, 14] to predict multiple overlapping segmentation masks. MaskFormer models have an option to disable queries and might not predict any masks if no objects are present. Points for which no masks were predicted will have the highest anomaly score. The Mask4Former model follows the same architectural paradigm and has a query deactivation mechanism that allows for no prediction in uncertain regions. This allows us to use the RbA method to generate anomaly scores within our setup. Unlike the original RbA model setup, we did not modify the transformer decoder.", + "bbox": [ + 511, + 527, + 908, + 709 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Void Classifier We follow a common baseline [4] by treating the unlabeled regions as an additional semantic class during training.", + "bbox": [ + 511, + 715, + 906, + 762 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Baseline Evaluation", + "text_level": 1, + "bbox": [ + 511, + 777, + 705, + 794 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Training data. We use SemanticKITTI [1], Panoptic-CUDAL [55], and an additional recorded sequence from to jointly train our models. The Panoptic-CUDAL dataset, contains six sequences for training and one for validation, and was mostly recorded in rural locations, while our dataset consists mainly of urban scenes. To reduce the do", + "bbox": [ + 511, + 809, + 906, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "11880", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/09544e1b60605c6b38763310ee0719ff888ce3864055368db73fe8ce13e97ee3.jpg", + "table_caption": [ + "Table 2. Anomaly Segmentation Performance of the selected baseline methods." + ], + "table_footnote": [], + "table_body": "
MethodOoD DataPoint-Level OODObject-Level OOD
AUROC ↑FPR@95 ↓AP ↑RecallQSQRQUQPQ
Deep Ensemble [34]X86.7458.055.1716.7584.4910.4314.168.81
MC Dropout [50]X61.5182.370.112.2586.721.952.141.86
Max Logit [24]X84.5381.490.9526.1483.062.1321.711.77
Void Classifier [4]✓85.9978.603.9217.6484.408.1914.896.91
RbA [41]X66.38100.00.8124.0483.283.2320.022.69
", + "bbox": [ + 96, + 114, + 906, + 229 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/202e72460f8578739473a19829883113d93f0592496578141c487bb4347a1014.jpg", + "table_caption": [ + "Table 3. Closed-set performance of the methods on validation sets of SemanticKITTI [1] and STU-inlier dataset. * denotes a model trained only on SemanticKITTI dataset." + ], + "table_footnote": [], + "table_body": "
MethodSemanticKITTISTU-inlier Dataset
PQSQRQPQSQRQ
Mask-PLS* [40]59.976.469.139.667.050.5
Mask-PLS [40]55.676.064.550.969.462.2
Mask4Former-3D* [60]61.981.371.642.870.653.2
Mask4Former-3D [60]60.781.070.452.773.864.7
", + "bbox": [ + 94, + 306, + 483, + 385 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "main gap, we have recorded and annotated two additional inlier scenes from an urban setting \"STU-inlier\", one for training and one for evaluation. These training sequences closely resemble the environments in which anomaly scenes were recorded. We train models with additional data recorded from the same vehicle setup as Panoptic-CUDAL and labeled the data with 19 SemanticKITTI labels. However, SemanticKITTI and our recorded dataset contain \"other-object\" that acts as a miscellaneous class for all objects that do not fit other-class definitions. A typical object is a bench, a garbage bin, or an advertisement post. To strictly define anomalies, we analyze all other-objects present in our training set, as well as the \"Movable_Object.Debris\" and \"Pushable.Pullable\" class objects in the NuScenes training set. We specifically design our dataset such that we do not have an intersection between anomalous objects and the aforementioned classes. Our Void Classifier training procedure consists of training on this unlabeled class that is present in the training data.", + "bbox": [ + 88, + 415, + 482, + 704 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Models. We train MaskPLS [40] and Mask4Former [60] models jointly on both SemanticKitti and our training sequences. These models achieve state-of-the-art results on the SemanticKITTI [1] dataset for single-scan and multi-scan settings. However, while our dataset supports evaluation of temporal sequences, we performed our evaluation and training on single scans only. In our experiments, we refer to Mask4Former trained on single scans as Mask4Former-3D. All anomaly segmentation methods use Mask4Former-3D as the main model.", + "bbox": [ + 88, + 710, + 482, + 863 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Metrics. To evaluate the performance of our methods, we use common metrics in the anomaly segmentation commu", + "bbox": [ + 89, + 869, + 483, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "nity in 2D and panoptic segmentation in 3D LiDAR. Following SemanticKitti, we evaluated both inlier and outlier performance on points within 50 meters of the vehicle, and on objects with at least five points.", + "bbox": [ + 511, + 255, + 906, + 316 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Point-Level Metrics. To evaluate the performance of the anomaly segmentation at the point level, we use the common metrics of Average Precision (AP), False-Positive Rate at $95\\%$ True-Positive Rate (FPR95), and Area Under Receiver Operating characteristic Curve (AUROC). Commonly used in the anomaly segmentation community [4].", + "bbox": [ + 511, + 320, + 908, + 412 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Object-Level Metrics. To be consistent with the panoptic segmentation setup, we use panoptic metrics to evaluate the model performance on inlier classes. For the inlier data, we compute the Panoptic Quality (PQ) on the SemanticKitti classes. We report detailed class scores in the Supplementary. In contrast to anomaly segmentation, the open task formulation emphasizes the recall of all objects. Formally, the Unknown Quality (UQ) metric is used in such a setting [58]. However, in the case of anomaly segmentation, since anomalies can pose serious threats to the vehicle, it is important to predict as few false positives as possible, while also producing accurate segmentation labels for anomaly objects. Panoptic Quality uses an F-score and penalizes the prediction of multiple false positives. Thus, for anomaly segmentation baselines, we report both Panoptic Quality and Unknown Quality on a single class. For inlier performance, we report the Panoptic Quality score on the SemanticKITTI classes.", + "bbox": [ + 511, + 415, + 908, + 686 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Panoptic Quality [32] is the geometric mean between Recognition Quality (RQ) and Segmentation Quality (SQ) for a class $c$ :", + "bbox": [ + 511, + 688, + 906, + 734 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {P Q} _ {c} = \\underbrace {\\frac {\\sum_ {(p , g) \\in T P _ {c}} \\operatorname {I o U} (p , g)}{| T P _ {c} |}} _ {\\text {S e g m e n t a t i o n Q u a l i t y (S Q)}} \\times \\underbrace {\\frac {| T P _ {c} |}{| T P _ {c} | + \\frac {1}{2} | F P _ {c} | + \\frac {1}{2} | F N _ {c} |}} _ {\\text {R e c o n g i t i o n Q u a l i t y (R Q)}}\n$$\n", + "text_format": "latex", + "bbox": [ + 535, + 743, + 888, + 795 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $TP$ , $FP$ , and $FN$ are True Positives, False Positives, and False Negatives for objects with Intersection over Union (IoU) higher than $50\\%$ . The final PQ metric is calculated as an average over all classes $c$ . Ignore points are removed from the scene prior to evaluation and erroneous predictions in the ignore region are not penalized.", + "bbox": [ + 511, + 809, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "11881", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To evaluate the recall of anomaly objects, we use the Unknown Quality (UQ) metric [58]:", + "bbox": [ + 89, + 90, + 482, + 122 + ], + "page_idx": 7 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {U Q} = \\underbrace {\\frac {\\sum_ {(p , g) \\in T P} \\operatorname {I o U} (p , g)}{| T P |}} _ {\\text {S e g m e n t a t i o n Q u a l i t y (S Q)}} \\times \\underbrace {\\frac {| T P |}{| T P | + | F N |}} _ {\\text {R e c a l l Q u a l i t y (R e c a l l Q)}}.\n$$\n", + "text_format": "latex", + "bbox": [ + 119, + 128, + 452, + 186 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The metric does not penalize $FP$ and allows for the evaluation of object recall. However, the metric counts objects as $TP$ only if they have an IoU of at least $50\\%$ with ground truth.", + "bbox": [ + 89, + 195, + 483, + 255 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.1. Experiments", + "text_level": 1, + "bbox": [ + 89, + 265, + 225, + 280 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Inlier Performance on SemanticKitti and STU-inlier dataset. We train models jointly with SemanticKITTI [1], Panoptic-CUDAL [55] and our additional sequence, and evaluate the SemanticKITTI and STU-inlier validation sets. The jointly trained model demonstrates improved performance on our STU-inlier dataset compared to models trained solely on the SemanticKITTI dataset, suggesting a small domain gap between our dataset and SemanticKITTI. Evaluation on the proposed STU-inlier sequence shows good segmentation performance for inlier classes that is critical for anomaly segmentation. The metrics for the performance of the baseline method in the validation sets are provided in Tab. 3.", + "bbox": [ + 89, + 291, + 483, + 487 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Out-of-Distribution (OOD) Performance. We evaluate baseline methods for anomaly segmentation at both the point and object levels. Our findings indicate that directly transferring methods from 2D to 3D does not yield similar performance. Our point clouds typically contain around 100,000 points, with only a few dozen being anomalous, as shown in Fig. 6. Due to the significant class imbalance, point-level scores are generally low, highlighting the difficulty of the task. A high FPR@95, as noted in ODIN [27], suggests that models often misclassify road objects as inlier classes. Since LiDAR point clouds have lower resolution at a distance, predicting a small anomaly object consisting of a few points is a hard task. It is common for large anomalous objects to be predicted as pedestrians or other vehicles, as depicted in Fig. 7. Anomalous objects frequently occur in familiar contexts, which leads the model to confidently assign inlier labels to such anomalies in many cases. This behavior contrasts significantly with models trained on 2D data, where a simple max logit approach [26] outperforms Deep Ensembles [10]. However, the Deep Ensemble model stands out by producing fewer false positives at the point level, resulting in a lower FPR@95 and a higher PQ metric. Despite this reduction in false positives, the average precision (AP) remains low compared to 2D datasets. We attribute this to the high confidence with which individual models predict inlier objects, with disagreements between models contributing to the improved FPR@95. We refer", + "bbox": [ + 89, + 493, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/028a40dd22490729cfbd1b5011aeae64ba6e4b5f51f913f5b4c8c9fd131622c8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 532, + 88, + 709, + 224 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b035bc2638a0727b7506a25816e1bc56b5a8351b2f6d15dad5ec66fa3abbb146.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 710, + 88, + 887, + 224 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/15be6f85a2b4e4fac3a538396f6f3154fac2efb04aa1c2cf4b71cfdc44cb6be3.jpg", + "image_caption": [ + "Figure 7. Example of a failure case anomaly segmentation. For the chair on the road, labeled as anomaly in upper left, model predicts \"other-vehicle\" class, in lower left, with a high certainty, as indicated by MaxLogit scores on the bottom right." + ], + "image_footnote": [], + "bbox": [ + 532, + 224, + 709, + 359 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/06def6696d786f3e297e1f8939c2997b3ee945b57c817a77910f4935e3ac3400.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 710, + 224, + 887, + 359 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "readers to the supplementary material for a more detailed investigation of the results.", + "bbox": [ + 511, + 455, + 906, + 486 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusions and Future Work", + "text_level": 1, + "bbox": [ + 511, + 503, + 790, + 520 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we introduced the STU dataset, designed for anomaly segmentation in LiDAR point clouds. The dataset was collected using eight cameras and a LiDAR sensor, which provides a surround view of environments through images and high-resolution LiDAR data. We extensively annotated point clouds with \"unlabeled\", \"anomaly\", and \"inlier\" classes, as well as introduced additional sequences without anomalies for training and evaluation. The results of the baseline methods indicate that there is a significant performance gap between 2D and 3D anomaly segmentation, which underscores numerous ongoing research opportunities in the field. Our objective with this paper is to provide researchers with annotated data and standardized benchmarks to facilitate the comparison of their anomaly segmentation approaches in 3D. We believe the dataset could be especially useful for the evaluation of methods that use temporal LiDAR data and multi-modal data in the future, and can drive innovation in developing more robust anomaly detection systems that can better handle complex real-world scenarios.", + "bbox": [ + 511, + 531, + 906, + 834 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments. A. Nekrasov's acknowledges funding by BMBF project \"WestAI\" (grant no. 01IS22094D). J. S. Berrio Perez acknowledges funding by the German Academic Exchange Service (DAAD Project-ID 30001831).", + "bbox": [ + 511, + 840, + 908, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "11882", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jens Behley, Martin Garbade, Andres Milioto, Jan Quenzel, Sven Behnke, Cyril Stachniss, and Juergen Gall. SemanticKITTI: A Dataset for Semantic Scene Understanding of LiDAR Sequences. In International Conference on Computer Vision (ICCV), 2019. 2, 3, 4, 6, 7, 8", + "[2] Petra Bevandić, Ivan Kreso, Marin Oršić, and Siniša Šegvic. Simultaneous Semantic Segmentation and Outlier Detection in Presence of Domain Shift. In German Conference on Pattern Recognition (GCPR), 2019. 3", + "[3] Hermann Blum, Paul-Edouard Sarlin, Juan I. Nieto, Roland Y. Siegwart, and César Cadena. Fishyscapes: A benchmark for safe semantic segmentation in autonomous driving. International Conference on Computer Vision Workshop (ICCV'W), 2019. 3", + "[4] Hermann Blum, Paul-Edouard Sarlin, Juan Nieto, Roland Siegwart, and Cesar Cadena. The Fishyscapes Benchmark: Measuring Blind Spots in Semantic Segmentation. International Journal on Computer Vision (IJCV), 2021. 2, 3, 4, 6, 7", + "[5] Daniel Bogdoll, Svenja Uhlemeyer, Kamil Kowol, and J Marius Zöllner. Perception datasets for anomaly detection in autonomous driving: A survey. In Intelligent Vehicles Symposium (IV), 2023. 2, 3", + "[6] Daniel Bogdoll, Iramm Hamdard, Lukas Namgyu Rößler, Felix Geisler, Muhammed Bayram, Felix Wang, Jan Imhof, Miguel de Campos, Anushervon Tabarov, Yitian Yang, Hanno Gottschalk, and J. Marius Zöllner. AnoVox: A Benchmark for Multimodal Anomaly Detection in Autonomous Driving. In ECCV 2024 W-CODA workshop, 2024. 2", + "[7] Holger Caesar, Varun Bankiti, Alex H. Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuScenes: A multimodal dataset for autonomous driving. In Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2, 3", + "[8] Jun Cen, Peng Yun, Junhao Cai, Michael Yu Wang, and Ming Liu. Open-set 3D Object Detection. In International Conference on 3D Vision (3DV), 2021. 3", + "[9] Anirudh S. Chakravarthy, Meghanana Reddy Ganesina, Peiyun Hu, Laura Leal-Taixe, Shu Kong, Deva Ramanan, and Aljosa Osep. Lidar Panoptic Segmentation in an Open World. International Journal on Computer Vision (IJCV), 2024. 3", + "[10] Robin Chan, Krzysztof Lis, Svenja Uhlemeyer, Hermann Blum, Sina Honari, Roland Siegwart, Pascal Fua, Mathieu Salzmann, and Matthias Rottmann. SegmentMeIfYouCan: A Benchmark for Anomaly Segmentation. In Proceedings of the Neural Information Processing Systems Track on Datasets and Benchmarks, 2021. 2, 3, 8", + "[11] Robin Chan, Matthias Rottmann, and Hanno Gottschalk. Entropy maximization and meta classification for out-of-distribution detection in semantic segmentation. In International Conference on Computer Vision (ICCV), 2021. 1, 3", + "[12] Liang-Chieh Chen, Yukun Zhu, George Papandreou, Florian Schroff, and Hartwig Adam. Encoder-Decoder with Atrous" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Separable Convolution for Semantic Image Segmentation. In European Conference on Computer Vision (ECCV), 2018. 3", + "[13] Bowen Cheng, Ishan Misra, Alexander G. Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention Mask Transformer for Universal Image Segmentation. In Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 3, 6", + "[14] Bowen Cheng, Alexander G. Schwing, and Alexander Kirilov. Per-pixel classification is not all you need for semantic segmentation. In Neural Information Processing Systems (NeurIPS), 2021. 3, 6", + "[15] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 2, 3", + "[16] Anja Delić, Matej Grić, and Siniša Šegvic. Outlier detection by ensembling uncertainty with negative objectness. In British Machine Vision Conference (BMVC), 2024. 1, 3", + "[17] kuefeng Du, Zhaoning Wang, Mu Cai, and Yixuan Li. Vos: Learning what you don't know by virtual outlier synthesis. In International Conference on Learning Representations (ICLR), 2021. 1", + "[18] Stefano Gasperini, Alvaro Marcos-Ramiro, Michael Schmidt, Nassir Navab, Benjamin Busam, and Federico Tombari. Segmenting Known Objects and Unseen Unknowns without Prior Knowledge. In International Conference on Computer Vision (ICCV), 2023. 1, 3, 5", + "[19] A Geiger, P Lenz, C Stiller, and R Urtasun. Vision meets robotics: The KITTI dataset. In The International Journal of Robotics Research, 2013. 3", + "[20] Matej Grcic, Petra Bevandic, and Sinisa Segvic. Dense open-set recognition with synthetic outliers generated by real nvp. ArXiv, abs/2011.11094, 2020. 3", + "[21] Matej Grcic, Petra Bevandić, and Sinisa Šegvic. Densehybrid: Hybrid anomaly detection for dense open-set recognition. In European Conference on Computer Vision (ECCV), 2022. 1, 3", + "[22] Krishnam Gupta, Syed Ashar Javed, Vineet Gandhi, and K. Madhava Krishna. MergeNet: A Deep Net Architecture for Small Obstacle Discovery. In International Conference on Robotics and Automation (ICRA), 2018. 2", + "[23] Jianhua Han, Xiwen Liang, Hang Xu, Kai Chen, Lanqing Hong, Jiageng Mao, Chaoqiang Ye, Wei Zhang, Zhenguo Li, Xiaodan Liang, and Chunjing Xu. SODA10M: A LargeScale 2D Self/Semi-Supervised Object Detection Dataset for Autonomous Driving. In Proceedings of the Neural Information Processing Systems Track on Datasets and Benchmarks, 2021. 3", + "[24] Dan Hendrycks and Kevin Gimpel. A Baseline for Detecting Misclassified and Out-of-Distribution Examples in Neural Networks. In International Conference on Learning Representations (ICLR), 2018. 2, 3, 6, 7", + "[25] Dan Hendrycks, Steven Basart, Mantas Mazeika, Andy Zou, Joe Kwon, Mohammadreza Mostajabi, Jacob Steinhardt, and Dawn Song. Scaling Out-of-Distribution Detection for Real-" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "11883", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "World Settings. In International Conference on Machine Learning (ICML), 2022. 2, 3", + "[26] Dan Hendrycks, Steven Basart, Mantas Mazeika, Andy Zou, Mohammadreza Mostajabi, Jacob Steinhardt, and Dawn Xiaodong Song. Scaling out-of-distribution detection for real-world settings. In International Conference on Machine Learning, 2022. 8", + "[27] Yen-Chang Hsu, Yilin Shen, Hongxia Jin, and Zsolt Kira. Generalized ODIN: Detecting Out-of-Distribution Image Without Learning From Out-of-Distribution Data. In Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 8", + "[28] Chengjie Huang, Van Duong Nguyen, Vahdat Abdelzad, Christopher Gus Mannes, Luke Rowe, Benjamin Therien, Rick Salay, and K. Czarnecki. Out-of-distribution detection for lidar-based 3d object detection. IEEE Intelligent Transportation Systems Conference (ITSC), 2022. 3", + "[29] Håkon Hukkelås and Frank Lindseth. Deepprivacy2: Towards realistic full-body anonymization. In Winter Conference on Applications of Computer Vision (WACV), 2023. 4", + "[30] Jaedong Hwang, Seoul Wug Oh, Joon-Young Lee, and Bohyung Han. Exemplar-Based Open-Set Panoptic Segmentation Network. In Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 3", + "[31] Sanghun Jung, Jungsoo Lee, Daehoon Gwak, Sungha Choi, and Jaegul Choo. Standardized Max Logits: A Simple yet Effective Approach for Identifying Unexpected Road Obstacles in Urban-Scene Segmentation. In International Conference on Computer Vision (ICCV), 2021. 1", + "[32] Alexander Kirillov, Kaiming He, Ross Girshick, Carsten Rother, and Piotr Dólar. Panoptic Segmentation. In Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 7", + "[33] Michael Kösel, Marcel Schreiber, Michael Ulrich, Claudius Gläser, and Klaus Dietmayer. Revisiting Out-of-Distribution Detection in LiDAR-based 3D Object Detection. In Intelligent Vehicles Symposium (IV), 2024. 3", + "[34] Balaji Lakshminarayanan, Alexander Pritzel, and Charles Blundell. Simple and Scalable Predictive Uncertainty Estimation using Deep Ensembles. In Neural Information Processing Systems (NeurIPS), 2017. 2, 3, 6, 7", + "[35] Seungjae Lee, Hyungtae Lim, and Hyun Myung. Patchwork++: Fast and robust ground segmentation solving partial under-segmentation using 3D point cloud. In Proc. IEEE/RSJ Int. Conf. Intell. Robots Syst., 2022. 3", + "[36] Kaican Li, Kai Chen, Haoyu Wang, Lanqing Hong, Chaoqiang Ye, Jianhua Han, Yukuai Chen, Wei Zhang, Chunjing Xu, Dit-Yan Yeung, et al. Coda: A real-world road corner case dataset for object detection in autonomous driving. In European Conference on Computer Vision (ECCV), 2022. 2, 3", + "[37] Chen Liang, Wenguan Wang, Jiaxu Miao, and Yi Yang. GMMSeg: Gaussian Mixture based Generative Semantic Segmentation Models. In Neural Information Processing Systems (NeurIPS), 2022. 1, 3", + "[38] Kira Maag, Robin Chan, Svenja Uhlemeyer, Kamil Kowol, and Hanno Gottschalk. Two Video Data Sets for Tracking" + ], + "bbox": [ + 91, + 90, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "and Retrieval of Out of Distribution Objects. In Asian Conference on Computer Vision (ACCV), 2022. 2, 3", + "[39] Jiageng Mao, Minzhe Niu, Chenhan Jiang, Hanxue Liang, Jingheng Chen, Xiaodan Liang, Yamin Li, Chaoqiang Ye, Wei Zhang, Zhenguo Li, Jie Yu, Hang Xu, and Chunjing Xu. One Million Scenes for Autonomous Driving: ONCE Dataset. In Neural Information Processing Systems (NeurIPS), 2021. 3", + "[40] Rodrigo Marcuzzi, Lucas Nunes, Louis Wiesmann, Jens Behley, and Cyril Stachniss. Mask-Based Panoptic LiDAR Segmentation for Autonomous Driving. In IEEE Robotics And Automation Letters (RAL), 2023. 7", + "[41] Nazir Nayal, Misra Yavuz, João F. Henriques, and Fatma Güney. RbA: Segmenting Unknown Regions Rejected by All. In International Conference on Computer Vision (ICCV), 2023. 1, 2, 3, 6, 7", + "[42] Alexey Nekrasov, Rui Zhou, Miriam Ackermann, Alexander Hermans, and Matthias Rottmann Bastian Leibe. OoDIS: Anomaly Instance Segmentation Benchmark. arXiv preprint arXiv:2406.11835, 2024. 2", + "[43] Gerhard Neuhold, Tobias Ollmann, Samuel Rota Bulo, and Peter Kontschieder. The mapillary vistas dataset for semantic understanding of street scenes. In International Conference on Computer Vision (ICCV), 2017. 3", + "[44] Lucas Nunes, Xieyuanli Chen, Rodrigo Marcuzzi, Aljosa Osep, Laura Leal-Taixe, Cyril Stachniss, and Jens Behley. Unsupervised Class-Agnostic Instance Segmentation of 3D LiDAR Data for Autonomous Vehicles. IEEE Robotics And Automation Letters (RAL), 2022. 3", + "[45] Peter Pinggera, Sebastian Ramos, Stefan Gehrig, Uwe Franke, Carsten Rother, and Rudolf Mester. Lost and Found: Detecting Small Road Hazards for Self-Driving Vehicles. In International Conference on Intelligent Robots and Systems (IROS), 2016. 2, 3, 4", + "[46] Aldi Piroli, Vinzenz Dallabetta, Johannes Kopp, Marc Walesa, Daniel Meissner, and Klaus Dietmayer. LS-VOS: Identifying Outliers in 3D Object Detections Using Latent Space Virtual Outlier Synthesis. In IEEE Intelligent Transportation Systems Conference (ITSC), 2023. 2, 3", + "[47] Shyam Nandan Rai, Fabio Cermelli, Dario Fontanel, Carlo Masone, and Barbara Caputo. Unmasking Anomalies in Road-Scene Segmentation. In International Conference on Computer Vision (ICCV), 2023. 3", + "[48] Jenny Seidenschwarz, Aljosa O'Sep, Francesco Ferroni, Simon Lucey, and Laura Leal-Taixe. SeMoLi: What Moves Together Belongs Together. In Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 3", + "[49] Aasheesh Singh, Aditya Kamireddypalli, Vineet Gandhi, and K. Madhava Krishna. LiDAR guided Small obstacle Segmentation. In International Conference on Intelligent Robots and Systems (IROS), 2020. 2, 3", + "[50] Nitish Srivastava, Geoffrey Hinton, Alex Krizhevsky, Ilya Sutskever, and Ruslan Salakhutdinov. Dropout: a simple way to prevent neural networks from overfitting. In Neural Information Processing Systems (NeurIPS), 2014. 2, 6, 7", + "[51] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou," + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "11884", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Yuning Chai, Benjamin Caine, et al. Scalability in perception for autonomous driving: Waymo open dataset. In Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2", + "[52] DashcamCleaner Development Team. Dashcamcleaner: Censor identifiable information in videos from dashcam recordings. https://github.com/taehse/DashcamCleaner, 2024.4", + "[53] Tefft, Brian C. . The Prevalence of Motor Vehicle Crashes Involving Road Debris, United States, 2011-2014. https://aaafoundation.org/wp-content/uploads/2017/12/RoadDebris_REPORT_2015.pdf, 2015. [Online]. 1", + "[54] Yu Tian, Yuyuan Liu, Guansong Pang, Fengbei Liu, Yuanhong Chen, and Gustavo Carneiro. Pixel-wise Energy-biased Abstention Learning for Anomaly Segmentation on Complex Urban Driving Scenes. In European Conference on Computer Vision (ECCV), 2022. 1, 3", + "[55] Tzu-Yun Tseng, Alexey Nekrasov, Malcolm Burdorf, Bastian Leibe, Julie Stephany Berrio Perez, Mao Shan, and Stewart Worrall. Panoptic-CUDAL Technical Report: Rural Australia Point Cloud Dataset in Rainy Conditions. arXiv preprint arXiv:2503.16378, 2025. 3, 4, 6, 8", + "[56] S. Verma, J. S. Berrio, S. Worrall, and E. Nebot. Automatic extrinsic calibration between a camera and a 3d lidar using 3d point and plane correspondences. In IEEE Intelligent Transportation Systems Conference (ITSC), 2019. 3", + "[57] Ignacio Vizzo, Tiziano Guadagnino, Benedikt Mersch, Louis Wiesmann, Jens Behley, and Cyril Stachniss. KISS-ICP: In Defense of Point-to-Point ICP – Simple, Accurate, and Robust Registration If Done the Right Way. In IEEE Robotics And Automation Letters (RAL), 2023. 4", + "[58] Kelvin Wong, Shenlong Wang, Mengye Ren, Ming Liang, and Raquel Urtasun. Identifying Unknown Instances for Autonomous Driving. In Conference on Robot Learning (CoRL), 2019. 2, 3, 7, 8", + "[59] Yajun Xu, Huan Hu, Xiaoya Zhu, Yibing Nan, Kai Wang, ZhaoXiang Liu, and Shiguo Lian. Raod: A benchmark for road abandoned object detection from video surveillance. IEEE Access, 2024. 2", + "[60] Kadir Yilmaz, Jonas Schult, Alexey Nekrasov, and Bastian Leibe. Mask4Former: Mask Transformer for 4D Panoptic Segmentation. In International Conference on Robotics and Automation (ICRA), 2024. 3, 6, 7", + "[61] Fisher Yu, Haofeng Chen, Xin Wang, Wenqi Xian, Yingying Chen, Fangchen Liu, Vashisht Madhavan, and Trevor Darrell. Bdd100k: A diverse driving dataset for heterogeneous multitask learning. In Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 3" + ], + "bbox": [ + 91, + 90, + 483, + 782 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11885", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Spotting the Unexpected (STU)_ A 3D LiDAR Dataset for Anomaly Segmentation in Autonomous Driving/7019444b-0523-423c-8391-a6c740f37e7c_model.json b/2025/Spotting the Unexpected (STU)_ A 3D LiDAR Dataset for Anomaly Segmentation in Autonomous Driving/7019444b-0523-423c-8391-a6c740f37e7c_model.json new file mode 100644 index 0000000000000000000000000000000000000000..b3ed958bc6da720cb943c80f0d5f18b600d6386c --- /dev/null +++ b/2025/Spotting the Unexpected (STU)_ A 3D LiDAR Dataset for Anomaly Segmentation in Autonomous Driving/7019444b-0523-423c-8391-a6c740f37e7c_model.json @@ -0,0 +1,2422 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.131, + 0.908, + 0.178 + ], + "angle": 0, + "content": "Spotting the Unexpected (STU): A 3D LiDAR Dataset for Anomaly Segmentation in Autonomous Driving" + }, + { + "type": "text", + "bbox": [ + 0.156, + 0.203, + 0.303, + 0.222 + ], + "angle": 0, + "content": "Alexey Nekrasov1" + }, + { + "type": "text", + "bbox": [ + 0.348, + 0.203, + 0.497, + 0.221 + ], + "angle": 0, + "content": "Malcolm Burdorf1" + }, + { + "type": "text", + "bbox": [ + 0.542, + 0.204, + 0.677, + 0.221 + ], + "angle": 0, + "content": "Stewart Worrall2" + }, + { + "type": "text", + "bbox": [ + 0.721, + 0.204, + 0.84, + 0.221 + ], + "angle": 0, + "content": "Bastian Leibe1" + }, + { + "type": "text", + "bbox": [ + 0.383, + 0.222, + 0.614, + 0.24 + ], + "angle": 0, + "content": "Julie Stephany Berrio Perez" + }, + { + "type": "text", + "bbox": [ + 0.277, + 0.239, + 0.726, + 0.258 + ], + "angle": 0, + "content": "\\(^{1}\\) RWTH Aachen University, \\(^{2}\\) The University of Sydney" + }, + { + "type": "text", + "bbox": [ + 0.254, + 0.273, + 0.742, + 0.29 + ], + "angle": 0, + "content": "https://vision.rwth-aachen.de/stu-dataset" + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.323, + 0.907, + 0.528 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.538, + 0.908, + 0.568 + ], + "angle": 0, + "content": "Figure 1. We present Spotting the Unexpected (STU) a novel anomaly segmentation dataset for autonomous driving. The dataset contains semantic and instance labels for out-of-distribution (OOD) objects, and includes surround-view setup with synchronized cameras." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.579, + 0.328, + 0.594 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.612, + 0.486, + 0.885 + ], + "angle": 0, + "content": "To operate safely, autonomous vehicles (AVs) need to detect and handle unexpected objects or anomalies on the road. While significant research exists for anomaly detection and segmentation in 2D, research progress in 3D is underexplored. Existing datasets lack high-quality multimodal data that are typically found in AVs. This paper presents a novel dataset for anomaly segmentation in driving scenarios. To the best of our knowledge, it is the first publicly available dataset focused on road anomaly segmentation with dense 3D semantic labeling, incorporating both LiDAR and camera data, as well as sequential information to enable anomaly detection across various ranges. This capability is critical for the safe navigation of autonomous vehicles. We adapted and evaluated several baseline models for 3D segmentation, highlighting the challenges of 3D anomaly detection in driving environments. Our dataset and evaluation code will be openly available, facilitating the testing and performance comparison of different approaches." + }, + { + "type": "title", + "bbox": [ + 0.515, + 0.579, + 0.645, + 0.594 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.609, + 0.907, + 0.837 + ], + "angle": 0, + "content": "Autonomous vehicles have become an increasingly important part of modern transportation systems. Despite significant advances, these vehicles continue to face challenges in accurately detecting unexpected objects on the road. One particular challenge is road debris, which refers to objects that are not typically part of the roadway environment and pose significant safety risks. According to the AAA Foundation for Traffic Safety [53], road debris was a contributing factor in more than 50,000 crashes, 10,000 injuries, and 125 fatalities annually between 2011 and 2014. These objects can vary widely in size and appearance, making them particularly challenging to detect. Addressing this issue is critical for ensuring the safe and reliable operation of autonomous vehicles, as it directly impacts their ability to handle anomalies in real-world driving conditions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.909, + 0.903 + ], + "angle": 0, + "content": "State-of-the-art anomaly segmentation and anomaly detection methods have received substantial attention in recent years [17, 37, 41], especially in the context of autonomous driving [11, 16, 18, 21, 31, 54]. One of the first" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11875" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.183 + ], + "angle": 0, + "content": "benchmarks for evaluating anomaly segmentation is Lost and Found [45], whose data serves as the basis for many complex benchmarks [4, 10]. These benchmarks provide a wide range of possible road anomalies and obstacles, and recent research has focused on exploring the temporal [38] and multimodal [6, 36] nature of anomaly segmentation." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.186, + 0.483, + 0.456 + ], + "angle": 0, + "content": "While current autonomous driving datasets [1, 7, 51] use a setup with surround view cameras, radar, and high-resolution LiDAR sensors, existing anomaly segmentation and detection benchmarks focus primarily on image data [5] and mostly use Cityscapes [15] as the in-distribution dataset. Image-first benchmarks often lack the comprehensive sensor setup found in modern autonomous vehicles, and overlook the use of 3D modality for anomaly segmentation. Recent research that has begun to investigate the use of LiDAR [49] sensor for anomaly segmentation and detection, lacks high-resolution sensors found in modern autonomous vehicles. Several approaches have sought to find anomalies in validation sets of existing datasets [36, 46] or have created datasets using simulation [6] due to the difficulty of collecting and annotating such data in the real world. Despite these efforts, extending existing data or simulating anomalies only partially addresses the lack of accessible and high-quality data, shown in Table 1." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.46, + 0.483, + 0.717 + ], + "angle": 0, + "content": "To address this issue, we analyzed the labels of two widely used in-distribution (ID) datasets [1, 7] to identify an appropriate set of anomalies that would not overlap with objects in the ID training or validation datasets. This informed our creation of a real-world dataset containing challenging anomaly objects placed on roads, with annotated sequences in 3D LiDAR point clouds. Our dataset consists of multiple high-resolution LiDAR sequences acquired in challenging environments, with point-level labels for individual anomaly objects (shown in Figure 1). We acquired data in both rural and urban environments, with objects of varying degrees of difficulty, and include a number of real-world anomalies found on public roads. In addition, to evaluate in-distribution performance, we annotated additional sequences without anomalies with inlier labels to train and evaluate models to ensure a small domain gap between our and SemanticKITTI [1] datasets." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.483, + 0.902 + ], + "angle": 0, + "content": "To establish simple baselines for 3D LiDAR segmentation, we adapted the methods used in 2D anomaly segmentation [4, 24, 34, 41, 50] and applied them to the 3D domain, investigating anomaly segmentation methods for 3D LiDAR point clouds. Our models, trained jointly on the SemanticKITTI dataset and our own dataset, show strong segmentation performance for in-distribution data and novel additional sequences. However, we observed that the baseline methods struggled with out-of-distribution objects, and models tend to predict anomaly objects as inliers with a high degree of confidence. We believe that our dataset will benefit the community by allowing the evaluation of anomaly" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.122 + ], + "angle": 0, + "content": "segmentation methods in the LiDAR domain and further exploration of multimodal anomaly segmentation methods." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.124, + 0.907, + 0.259 + ], + "angle": 0, + "content": "Our contributions provide a basis for the development of more advanced anomaly segmentation methods in the 3D LiDAR domain. Specifically, we propose a new anomaly segmentation dataset, and perform an analysis of the recorded and annotated data. The purpose of the paper is to provide practitioners with the necessary tools to evaluate their anomaly segmentation models. We also provide a set of baselines that can serve as a reference for future challenges and research directions in this area." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.276, + 0.652, + 0.291 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.302, + 0.691, + 0.318 + ], + "angle": 0, + "content": "2.1. Anomaly Datasets" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.325, + 0.905, + 0.552 + ], + "angle": 0, + "content": "Several multimodal datasets for autonomous driving are available, including Waymo [51], nuScenes [7], SemanticKITTI [1], among others. These datasets were collected in urban environments under typical driving conditions. Although these datasets are relevant for evaluating various tasks related to autonomous vehicle operation, they lack data for uncommon scenarios. One such scenario involves the presence of foreign objects on the road. Although this situation is rare during typical driving, autonomous vehicles must be able to recognize and respond to it safely. Here, assigning a unique label to each distinct object is impractical, since datasets have a limited set of labels, and objects on the road can vary widely. These objects are often grouped under a general category like \"other object\" and are excluded from the benchmarks due to their variability." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.553, + 0.906, + 0.78 + ], + "angle": 0, + "content": "In recent years, several datasets have been introduced to assess the detection of foreign objects on the road [4, 10, 25, 36, 38, 42, 45, 59]. Some of these datasets have been developed using data augmentation techniques or simulations, but only a few are based on data collected with real-world equipment in real driving scenarios. Simulations and augmentations are great tools, but they lack realism in terms of the sensor model [4] and the sizes of the objects [25]. However, the aforementioned datasets are limited to image data only. Autonomous vehicles use LiDAR sensors because they offer greater range, robustness to lighting conditions, and more precise 3D information about the environment compared to monocular depth estimations or stereo cameras. Therefore, incorporating 3D information is essential for accurately assessing road anomalies." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.781, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Three datasets that integrate camera and LiDAR data are SOD [22], TOR4D [58] and CODA [36]. The CODA series is derived from validation sets of existing autonomous vehicle datasets. Still, it has the limitation that anomalous objects (e.g., traffic cones, barriers) may also appear in the training set, which is acceptable for the task of open-set segmentation, but contradicts the task of anomaly segmentation. Recently, a few methods for 3D anomaly detection" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11876" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.101, + 0.089, + 0.896, + 0.104 + ], + "angle": 0, + "content": "Table 1. Comparison of our dataset with other community datasets [5]. SM = Semantic Mask, IM=Instance Mask, BB=Bounding Box" + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.115, + 0.907, + 0.376 + ], + "angle": 0, + "content": "
DatasetCameraLiDAR \nBeamsSource#ODDLabelSize \nTest/ValTemporalEnvironments \nTest SequencesRelated Dataset
Fishscapes [3, 4]
FS Lost and Found1 RGBXStaged1SM275/100Xn/aLost and Found [45]
FS Static1 RGBXAugmented1SM1000/30Xn/aCityScapes [15]
CAOS [25]
StreetHazards1 RGBXSimulated1SM1500✓2 sim towns - 150 seq
BDD-Anomaly1 RGBXValidation Set3SM810Xn/aBDD100K [61]
SegmentMelfYouCan [10]
RoadAnomaly211 RGBXWeb Sourced1SM100/10Xn/a
RoadObstacle211 RGBXStaged1SM382/30✓9 streets - 90 seq
CODA [36]
CODA-KITTI1 RGB✓-64Validation Set6BB309Xn/aKITTI [19]
CODA-nuScenes1 RGB✓-32Validation Set17BB134Xn/aNuScenes [7]
CODA-ONCE1 RGB✓-40Validation Set32BB1057Xn/aONCE [39]
CODA22-ONCE1 RGB✓-40Validation Set29BB717Xn/aONCE [39]
CODA22-SODA10M1 RGBXValidation Set29BB4167Xn/aSODA [23]
Wuppertal ODD [38]
SOS1 RGBX-distStaged13IM1129✓9 streets - 20 seq
CWL1 RGBX-distSimulated18IM1210✓26 streets - 26 seq
Lost and Found [45]1 StereoXStaged42SM1068✓5 streets - 5 seq
SOD [49]1 RGB✓-16Staged1SM460/530X2 streets - 5 seq
WD-Pascal [2]1 RGBXSimulated1SM70Xn/a
Vistas-NP [20]1 RGBXClass exclusion4SM11167Xn/aMapillary Vistas [43]
STU (Ours)8 RGB✓-128Staged1IM8022/1960✓6 streets - 51 seqSemKITTI [1], P-CUDAL [55]
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.4, + 0.482, + 0.581 + ], + "angle": 0, + "content": "have been proposed [28, 33, 46]. However, these methods use validation sequences to evaluate anomaly detection capabilities. The SOD dataset offers depth information for small road obstacles using a low-resolution 16-beam LiDAR, and that introduces a large domain gap between the training and test data. TOR4D and Rare4D [58] are the only datasets that contain real-world LiDAR data to benchmark road anomalies; however, these datasets are proprietary and not accessible to the public. We release out-of-distribution validation and in-distribution training sequences with a permissive license and open the private test set to submissions from the public." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.592, + 0.355, + 0.608 + ], + "angle": 0, + "content": "2.2. Open-Set Segmentation in 3D" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.614, + 0.485, + 0.902 + ], + "angle": 0, + "content": "Open-set and unsupervised segmentation are closely related to the out-of-distribution detection task. Multiple methods were proposed for unsupervised segmentation [44, 48] and open-set prediction in 3D [9, 58]. Recent evaluations in 2D anomaly segmentation [18] have demonstrated that methods such as OSIS [58] and EOPSN [30] perform poorly on anomaly objects. This weakness stems from open-set methods' reliance on the presence of unknown instances inside the void regions during training, which may not generalize well to real-world scenarios. Similarly, MLUC [8] assesses open-set performance on the UDI [8] dataset, but this dataset is also not publicly accessible and makes the same assumptions as OSIS methods. In addition, open-set methods typically remove the ground plane [44] to obtain clusters of potential objects, but ground plane removal [35] methods are often sensitive to chosen hyperparameters and yield noisy results, missing many smaller objects. For the evaluation of anomaly segmentation, it is important to have a test dataset that is free of anomaly objects which might" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.401, + 0.907, + 0.476 + ], + "angle": 0, + "content": "occur during training. In contrast to open-set methods, our work focuses on collecting and densely annotating real-world anomaly objects, ensuring that there is no intersection between in-distribution training sets and out-of-distribution test sets." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.488, + 0.907, + 0.505 + ], + "angle": 0, + "content": "2.3. Anomaly Segmentation in Autonomous Driving" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.511, + 0.909, + 0.782 + ], + "angle": 0, + "content": "To tackle the problem of anomaly segmentation, multiple works [4, 5, 10] have recently been proposed for the problem of 2D anomaly segmentation. Earlier works [4, 11, 21, 24, 34, 37, 54] were applied to per-pixel segmentation models such as DeepLab [12]; however, recently many works were focused on MaskFormer-style models [13, 14]. A line of work [16, 41, 47] proposed different methods to adapt Mask2Former, exploring the mechanism of mask prediction, where multiple masks might ignore certain regions of an image. These models provide more reliable anomaly scoring mechanisms not limited to per-pixel predictions and achieve higher performance on the anomaly segmentation task. For our benchmark, we decided to evaluate methods that also use MaskFormer-style models and investigate the application of the Mask4Former [60] model for 3D anomaly segmentation. We obtain all our scores using a MaskFormer-Style model, laying the foundation for future baselines." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.8, + 0.905, + 0.817 + ], + "angle": 0, + "content": "3. The Spotting the Unexpected (STU) Dataset" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Our data collection platform consists of a rigid frame mounted on a vehicle and equipped with eight hardware-triggered cameras and a LiDAR sensor to ensure synchronized data collection. The LiDAR and the cameras were calibrated using the method described in [56], with the cal" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "11877" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.093, + 0.487, + 0.173 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.195, + 0.179, + 0.378, + 0.191 + ], + "angle": 0, + "content": "(a) Naturalistic data collection." + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.196, + 0.487, + 0.277 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.152, + 0.282, + 0.422, + 0.295 + ], + "angle": 0, + "content": "(b) Data collection in controlled environment." + }, + { + "type": "image_caption", + "bbox": [ + 0.091, + 0.308, + 0.483, + 0.336 + ], + "angle": 0, + "content": "Figure 2. Data collection conducted in a naturalistic manner (a) and controlled environment (b) with objects on the road." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.364, + 0.483, + 0.424 + ], + "angle": 0, + "content": "ization process repeated for each camera. This setup follows the configuration discussed in Panoptic-CUDAL [55]. We refer to the Supplementary Material for detailed information on the vehicle setup." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.435, + 0.246, + 0.45 + ], + "angle": 0, + "content": "3.1. Data collection." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.457, + 0.483, + 0.563 + ], + "angle": 0, + "content": "The data was collected using two sets of conditions: one in a naturalistic environment where objects were found on public roads, and one in a controlled environment. In the naturalistic scenario, several hours of driving data were recorded, with specific events marked when objects appeared on the road. Figure 2a shows a bucket on the road recorded while driving on a highway." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.564, + 0.483, + 0.775 + ], + "angle": 0, + "content": "The controlled environment data collection protocol was rigorous for safety reasons. We employed four people: two spotters positioned at either end of a low-traffic and low-speed road, one responsible for placing objects on the road, and a driver. The spotters would signal when it was safe to place the objects, and once positioned, the driver would begin recording the scene. If one of the spotters reported an approaching vehicle, the data recording was stopped, and the object was immediately removed. This was done to avoid disrupting the normal flow of traffic. The vehicle was driven at a maximum speed of \\(50\\mathrm{km / h}\\) during data recording. Figure 2b shows a computer monitor placed on the road during sensor data logging. Some of the objects placed on the road are also shown in Figure 3." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Postprocessing. After data collection, post-processing is required to obtain vehicle poses and anonymize the images. Point cloud registration was performed using KISS ICP [57], a LiDAR odometry pipeline. KISS ICP includes point cloud motion compensation, subsampling, adaptive thresholding to determine correspondences, and LiDAR pose estimation. The calculated LiDAR pose was then exported in KITTI format as required for the labeling tool [1]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.213 + ], + "angle": 0, + "content": "Ethical considerations for this dataset include anonymizing camera images. Identifiable data, such as faces and license plates, are processed using DashcamCleaner [52] and DeepPrivacy2 [29]. DashcamCleaner uses a license plate detector to locate and blur license plates, while DeepPrivacy2 detects facial features and generates new, unidentifiable faces to replace the original ones. Examples of anonymization can be found in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.22, + 0.62, + 0.236 + ], + "angle": 0, + "content": "3.2. Labeling" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.242, + 0.905, + 0.377 + ], + "angle": 0, + "content": "We define anomalies as objects that may pose a danger to the autonomous vehicle and its passengers, especially those present on the driving surface and not present in the training dataset. In this work, we focus on road anomalies and do not consider unusual situations such as vehicle collisions or unexpected driver behavior, and we focus on the fine-grained anomaly segmentation. In particular, we follow previous work [4, 45] and focus on objects that are not present in the training set." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.378, + 0.906, + 0.514 + ], + "angle": 0, + "content": "We labeled the data using the SemanticKITTI labeler [1]. To generate initial labels for the scenes, we used pseudolabeling techniques using a pre-trained SemanticKITTI model. We instruct our annotators to group all points belonging to inliers as a single semantic class, questionable points as unlabeled, and anomaly points as the outlier class. We instruct our labelers that it is very important to label inliers and outliers carefully while allowing for larger unlabeled regions where the annotator may be challenged." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.519, + 0.906, + 0.79 + ], + "angle": 0, + "content": "Annotation process. The domain gap between urban environments and SemanticKITTI street scenes is minimal, especially in regions close to the vehicle. This similarity allows the acquisition of reliable initial estimates. The annotation process was performed by three annotators, each spending eight hours a week over several months. Each annotator spent approximately four hours on the initial labeling of each scene and an additional two hours on refining the final labels. To ensure accuracy and consistency, the annotators cross-validated each other's work. To further improve the quality of the labels, we examined the predictions of the baseline methods to see if any known objects were missed. In particular, current methods tend to be highly sensitive to objects that are present but ignored during training, such as those classified in the category \"other object\". We annotate these objects as unlabeled and ignore them in the evaluations, as they could alter the results of the anomaly segmentation methods." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.905, + 0.9 + ], + "angle": 0, + "content": "Differences to SemanticKITTI. In contrast to SemanticKITTI, which provides labels up to a range of 50 meters, we labeled all points visible to the LiDAR sensor. This was facilitated by using a LiDAR with a higher vertical resolution. The maximum distance resulting from an anomaly in our dataset is 150 meters, allowing for comprehensive evaluations. However, since we train on the Se" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "11878" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.1, + 0.089, + 0.231, + 0.153 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.234, + 0.09, + 0.365, + 0.152 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.368, + 0.091, + 0.498, + 0.152 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.091, + 0.631, + 0.152 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.634, + 0.091, + 0.764, + 0.152 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.767, + 0.091, + 0.898, + 0.152 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.1, + 0.153, + 0.231, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.234, + 0.153, + 0.364, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.368, + 0.153, + 0.498, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.153, + 0.631, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.634, + 0.153, + 0.764, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.767, + 0.153, + 0.898, + 0.216 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.1, + 0.217, + 0.231, + 0.28 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.234, + 0.217, + 0.364, + 0.28 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.368, + 0.217, + 0.498, + 0.28 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.217, + 0.631, + 0.28 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.634, + 0.217, + 0.764, + 0.28 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.767, + 0.217, + 0.898, + 0.28 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.1, + 0.281, + 0.231, + 0.343 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.234, + 0.281, + 0.364, + 0.343 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.368, + 0.281, + 0.498, + 0.343 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.281, + 0.631, + 0.343 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.634, + 0.281, + 0.764, + 0.343 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.767, + 0.281, + 0.898, + 0.343 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.355, + 0.908, + 0.397 + ], + "angle": 0, + "content": "Figure 3. Different anomalies in the STU dataset. Different objects on the road used for staged data collection. We pick objects such that have no intersection with the inlier dataset and place them on roads in different locations and illumination conditions. Objects might touch each other, be very small, as large as a chair or a surf board, and could cause an accident if a car would drive over them." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.424, + 0.483, + 0.484 + ], + "angle": 0, + "content": "manticKITTI dataset, which contains only labels up to 50 meters, we limit our evaluations to the 50 meter range. The sensor setup has 128 LiDAR beams, compared to 64 beams in SemanticKITTI." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.498, + 0.251, + 0.514 + ], + "angle": 0, + "content": "3.3. Dataset analysis" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.521, + 0.483, + 0.67 + ], + "angle": 0, + "content": "We recorded 70 sequences with anomalies, which we fully annotated for anomaly segmentation. Two of these labeled sequences were captured in a naturalistic manner, while the remaining 68 were staged to include various anomalies. We designated 19 sequences as a validation set and 51 as a closed test set for benchmarking purposes. In addition, we recorded and annotated two sequences with no anomalies for training and validation of the inlier performance \"STU-inlier\", to reduce the domain gap between our dataset and SemanticKITTI." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.673, + 0.483, + 0.883 + ], + "angle": 0, + "content": "Our dataset defines two primary classes: inliers and outliers. Inliers encompasses classes that are used to train a model, and contains objects expected on the road, such as vehicles, pedestrians, and road infrastructure. Outliers, on the other hand, refer to out-of-distribution elements or anomalies that are atypical or unexpected in the driving scene, shown in Figure 3. One additional class, unlabeled is ignored in the evaluation and typically includes classes that the model has seen during training, but was not supervised in, such as parking meters, utility boxes, and lamps. We provide anomaly annotations at both the instance and semantic levels, enabling a comprehensive evaluation of models for semantic and instance segmentation in 3D environments." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.886, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Each outlier object in our dataset has an instance la-" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.424, + 0.905, + 0.56 + ], + "angle": 0, + "content": "bel unique for a sequence. Figure 4 presents histograms that display the distribution of key properties for these instances, including the average and maximum number of LiDAR points per instance, instance heights, and the number of instances per scan. This analysis provides insights into the structural characteristics and variability of data within point clouds. Specifically, Figure 4 indicates that most instances have very few LiDAR points on average, but a sequence may have multiple instances per sequence." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.561, + 0.906, + 0.787 + ], + "angle": 0, + "content": "Furthermore, Figure 5 presents a scatter plot of detected anomalies within the vehicle frame, highlighting a higher density of anomalies along the longitudinal \\((X)\\) axis both in front and behind the vehicle. The accompanying histograms depict the distribution of the anomalies along each axis, noting that the anomalies are identified at distances ranging up to 150 meters in front of the vehicle. Figure 6 explores the relationship between the distance to an anomaly and its size, showing that the proportion of LiDAR point cloud obstacles decreases as the distance increases. This trend is due to the radial nature of LiDAR, which has lower vertical resolution at greater ranges. For benchmarking purposes, we focus exclusively on instances with more than five LiDAR points, highlighted in blue, ensuring that the evaluation targets detectable anomalies." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.802, + 0.704, + 0.817 + ], + "angle": 0, + "content": "3.4. Evaluated Baselines" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.905, + 0.901 + ], + "angle": 0, + "content": "The formulation of the anomaly segmentation task restricts training to avoid direct supervision of OOD objects. Generally, OOD objects are not present in the training set, but could be present during inference [18]. Since there are no methods that could be applied directly for our setup, we" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11879" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.109, + 0.092, + 0.298, + 0.198 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.105, + 0.206, + 0.301, + 0.219 + ], + "angle": 0, + "content": "(a) Avg. num of points/sequence." + }, + { + "type": "image", + "bbox": [ + 0.309, + 0.093, + 0.493, + 0.197 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.315, + 0.206, + 0.486, + 0.219 + ], + "angle": 0, + "content": "(b) Max. points per instance." + }, + { + "type": "image", + "bbox": [ + 0.506, + 0.093, + 0.692, + 0.198 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.5, + 0.205, + 0.693, + 0.219 + ], + "angle": 0, + "content": "(c) Max. height of instances (m)." + }, + { + "type": "image", + "bbox": [ + 0.702, + 0.094, + 0.887, + 0.198 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.698, + 0.205, + 0.892, + 0.219 + ], + "angle": 0, + "content": "(d) Numb. of instances/sequence." + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.221, + 0.907, + 0.264 + ], + "angle": 0, + "content": "Figure 4. Anomaly Instance Properties. A typical recorded anomaly has on average less than 50 points per sequence (a), with less than 300 (b) points at maximum, and a maximum height below one meter (c). We record up to nine individual anomaly instances in the same sequence (d)." + }, + { + "type": "image", + "bbox": [ + 0.096, + 0.291, + 0.486, + 0.46 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.464, + 0.483, + 0.493 + ], + "angle": 0, + "content": "Figure 5. Distribution of anomalies along the vehicle's reference frame. Most of the points appear around the vehicle." + }, + { + "type": "image", + "bbox": [ + 0.096, + 0.509, + 0.485, + 0.729 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.741, + 0.483, + 0.798 + ], + "angle": 0, + "content": "Figure 6. We follow SemanticKitti's maximum distance threshold and evaluate instances that are within a 50 meter radius of the vehicle. In addition, we constrain evaluation of point and object level metrics to a minimum of 5 points." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.484, + 0.903 + ], + "angle": 0, + "content": "adopt several baseline methods from the 2D anomaly segmentation to 3D LiDAR segmentation. First, we evaluate simple baselines, such as Max-Logit [24], Monte-Carlo Dropout [50], and Deep Ensembles [34]. Secondly, we train one of the models to predict ignored regions during train" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.29, + 0.905, + 0.351 + ], + "angle": 0, + "content": "ing, which could be used to classify “void” points during testing [4]. We also adapt RbA [41], one of the state-of-the-art methods for image anomaly segmentation, to the 3D LiDAR segmentation method Mask4Former-3D [60]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.357, + 0.907, + 0.522 + ], + "angle": 0, + "content": "Deep Ensembles, MC Dropout, and MaxLogit. Deep Ensembles [34] and MC Dropout [50] are very common methods that practitioners use due to ease of implementation and application. For Deep Ensembles, we train three Mask4Former-3D [60] models. For MC dropout, we train one Mask4Former-3D model with enabled dropout in the transformer decoder for self- and cross-attention operations, and during inference we repeat the transformer decoder operation 50 times. As a simple baseline that uses only one model and a single forward pass, we use the MaxLogit [24] method." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.529, + 0.909, + 0.71 + ], + "angle": 0, + "content": "RbA [41] is a method that uses the ability of MaskFormer-Style models [13, 14] to predict multiple overlapping segmentation masks. MaskFormer models have an option to disable queries and might not predict any masks if no objects are present. Points for which no masks were predicted will have the highest anomaly score. The Mask4Former model follows the same architectural paradigm and has a query deactivation mechanism that allows for no prediction in uncertain regions. This allows us to use the RbA method to generate anomaly scores within our setup. Unlike the original RbA model setup, we did not modify the transformer decoder." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.716, + 0.908, + 0.763 + ], + "angle": 0, + "content": "Void Classifier We follow a common baseline [4] by treating the unlabeled regions as an additional semantic class during training." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.779, + 0.707, + 0.795 + ], + "angle": 0, + "content": "4. Baseline Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Training data. We use SemanticKITTI [1], Panoptic-CUDAL [55], and an additional recorded sequence from to jointly train our models. The Panoptic-CUDAL dataset, contains six sequences for training and one for validation, and was mostly recorded in rural locations, while our dataset consists mainly of urban scenes. To reduce the do" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11880" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.263, + 0.089, + 0.735, + 0.104 + ], + "angle": 0, + "content": "Table 2. Anomaly Segmentation Performance of the selected baseline methods." + }, + { + "type": "table", + "bbox": [ + 0.097, + 0.115, + 0.907, + 0.23 + ], + "angle": 0, + "content": "
MethodOoD DataPoint-Level OODObject-Level OOD
AUROC ↑FPR@95 ↓AP ↑RecallQSQRQUQPQ
Deep Ensemble [34]X86.7458.055.1716.7584.4910.4314.168.81
MC Dropout [50]X61.5182.370.112.2586.721.952.141.86
Max Logit [24]X84.5381.490.9526.1483.062.1321.711.77
Void Classifier [4]✓85.9978.603.9217.6484.408.1914.896.91
RbA [41]X66.38100.00.8124.0483.283.2320.022.69
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.254, + 0.483, + 0.297 + ], + "angle": 0, + "content": "Table 3. Closed-set performance of the methods on validation sets of SemanticKITTI [1] and STU-inlier dataset. * denotes a model trained only on SemanticKITTI dataset." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.307, + 0.485, + 0.386 + ], + "angle": 0, + "content": "
MethodSemanticKITTISTU-inlier Dataset
PQSQRQPQSQRQ
Mask-PLS* [40]59.976.469.139.667.050.5
Mask-PLS [40]55.676.064.550.969.462.2
Mask4Former-3D* [60]61.981.371.642.870.653.2
Mask4Former-3D [60]60.781.070.452.773.864.7
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.416, + 0.483, + 0.705 + ], + "angle": 0, + "content": "main gap, we have recorded and annotated two additional inlier scenes from an urban setting \"STU-inlier\", one for training and one for evaluation. These training sequences closely resemble the environments in which anomaly scenes were recorded. We train models with additional data recorded from the same vehicle setup as Panoptic-CUDAL and labeled the data with 19 SemanticKITTI labels. However, SemanticKITTI and our recorded dataset contain \"other-object\" that acts as a miscellaneous class for all objects that do not fit other-class definitions. A typical object is a bench, a garbage bin, or an advertisement post. To strictly define anomalies, we analyze all other-objects present in our training set, as well as the \"Movable_Object.Debris\" and \"Pushable.Pullable\" class objects in the NuScenes training set. We specifically design our dataset such that we do not have an intersection between anomalous objects and the aforementioned classes. Our Void Classifier training procedure consists of training on this unlabeled class that is present in the training data." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.711, + 0.483, + 0.864 + ], + "angle": 0, + "content": "Models. We train MaskPLS [40] and Mask4Former [60] models jointly on both SemanticKitti and our training sequences. These models achieve state-of-the-art results on the SemanticKITTI [1] dataset for single-scan and multi-scan settings. However, while our dataset supports evaluation of temporal sequences, we performed our evaluation and training on single scans only. In our experiments, we refer to Mask4Former trained on single scans as Mask4Former-3D. All anomaly segmentation methods use Mask4Former-3D as the main model." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Metrics. To evaluate the performance of our methods, we use common metrics in the anomaly segmentation commu" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.256, + 0.907, + 0.318 + ], + "angle": 0, + "content": "nity in 2D and panoptic segmentation in 3D LiDAR. Following SemanticKitti, we evaluated both inlier and outlier performance on points within 50 meters of the vehicle, and on objects with at least five points." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.321, + 0.909, + 0.414 + ], + "angle": 0, + "content": "Point-Level Metrics. To evaluate the performance of the anomaly segmentation at the point level, we use the common metrics of Average Precision (AP), False-Positive Rate at \\(95\\%\\) True-Positive Rate (FPR95), and Area Under Receiver Operating characteristic Curve (AUROC). Commonly used in the anomaly segmentation community [4]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.416, + 0.909, + 0.688 + ], + "angle": 0, + "content": "Object-Level Metrics. To be consistent with the panoptic segmentation setup, we use panoptic metrics to evaluate the model performance on inlier classes. For the inlier data, we compute the Panoptic Quality (PQ) on the SemanticKitti classes. We report detailed class scores in the Supplementary. In contrast to anomaly segmentation, the open task formulation emphasizes the recall of all objects. Formally, the Unknown Quality (UQ) metric is used in such a setting [58]. However, in the case of anomaly segmentation, since anomalies can pose serious threats to the vehicle, it is important to predict as few false positives as possible, while also producing accurate segmentation labels for anomaly objects. Panoptic Quality uses an F-score and penalizes the prediction of multiple false positives. Thus, for anomaly segmentation baselines, we report both Panoptic Quality and Unknown Quality on a single class. For inlier performance, we report the Panoptic Quality score on the SemanticKITTI classes." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.908, + 0.735 + ], + "angle": 0, + "content": "Panoptic Quality [32] is the geometric mean between Recognition Quality (RQ) and Segmentation Quality (SQ) for a class \\( c \\):" + }, + { + "type": "equation", + "bbox": [ + 0.536, + 0.744, + 0.889, + 0.796 + ], + "angle": 0, + "content": "\\[\n\\mathrm {P Q} _ {c} = \\underbrace {\\frac {\\sum_ {(p , g) \\in T P _ {c}} \\operatorname {I o U} (p , g)}{| T P _ {c} |}} _ {\\text {S e g m e n t a t i o n Q u a l i t y (S Q)}} \\times \\underbrace {\\frac {| T P _ {c} |}{| T P _ {c} | + \\frac {1}{2} | F P _ {c} | + \\frac {1}{2} | F N _ {c} |}} _ {\\text {R e c o n g i t i o n Q u a l i t y (R Q)}}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.909, + 0.903 + ], + "angle": 0, + "content": "where \\( TP \\), \\( FP \\), and \\( FN \\) are True Positives, False Positives, and False Negatives for objects with Intersection over Union (IoU) higher than \\( 50\\% \\). The final PQ metric is calculated as an average over all classes \\( c \\). Ignore points are removed from the scene prior to evaluation and erroneous predictions in the ignore region are not penalized." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.956 + ], + "angle": 0, + "content": "11881" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.123 + ], + "angle": 0, + "content": "To evaluate the recall of anomaly objects, we use the Unknown Quality (UQ) metric [58]:" + }, + { + "type": "equation", + "bbox": [ + 0.12, + 0.13, + 0.454, + 0.188 + ], + "angle": 0, + "content": "\\[\n\\mathrm {U Q} = \\underbrace {\\frac {\\sum_ {(p , g) \\in T P} \\operatorname {I o U} (p , g)}{| T P |}} _ {\\text {S e g m e n t a t i o n Q u a l i t y (S Q)}} \\times \\underbrace {\\frac {| T P |}{| T P | + | F N |}} _ {\\text {R e c a l l Q u a l i t y (R e c a l l Q)}}.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.196, + 0.485, + 0.256 + ], + "angle": 0, + "content": "The metric does not penalize \\( FP \\) and allows for the evaluation of object recall. However, the metric counts objects as \\( TP \\) only if they have an IoU of at least \\( 50\\% \\) with ground truth." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.266, + 0.226, + 0.281 + ], + "angle": 0, + "content": "4.1. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.292, + 0.484, + 0.488 + ], + "angle": 0, + "content": "Inlier Performance on SemanticKitti and STU-inlier dataset. We train models jointly with SemanticKITTI [1], Panoptic-CUDAL [55] and our additional sequence, and evaluate the SemanticKITTI and STU-inlier validation sets. The jointly trained model demonstrates improved performance on our STU-inlier dataset compared to models trained solely on the SemanticKITTI dataset, suggesting a small domain gap between our dataset and SemanticKITTI. Evaluation on the proposed STU-inlier sequence shows good segmentation performance for inlier classes that is critical for anomaly segmentation. The metrics for the performance of the baseline method in the validation sets are provided in Tab. 3." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.494, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Out-of-Distribution (OOD) Performance. We evaluate baseline methods for anomaly segmentation at both the point and object levels. Our findings indicate that directly transferring methods from 2D to 3D does not yield similar performance. Our point clouds typically contain around 100,000 points, with only a few dozen being anomalous, as shown in Fig. 6. Due to the significant class imbalance, point-level scores are generally low, highlighting the difficulty of the task. A high FPR@95, as noted in ODIN [27], suggests that models often misclassify road objects as inlier classes. Since LiDAR point clouds have lower resolution at a distance, predicting a small anomaly object consisting of a few points is a hard task. It is common for large anomalous objects to be predicted as pedestrians or other vehicles, as depicted in Fig. 7. Anomalous objects frequently occur in familiar contexts, which leads the model to confidently assign inlier labels to such anomalies in many cases. This behavior contrasts significantly with models trained on 2D data, where a simple max logit approach [26] outperforms Deep Ensembles [10]. However, the Deep Ensemble model stands out by producing fewer false positives at the point level, resulting in a lower FPR@95 and a higher PQ metric. Despite this reduction in false positives, the average precision (AP) remains low compared to 2D datasets. We attribute this to the high confidence with which individual models predict inlier objects, with disagreements between models contributing to the improved FPR@95. We refer" + }, + { + "type": "image", + "bbox": [ + 0.533, + 0.089, + 0.71, + 0.225 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.712, + 0.089, + 0.888, + 0.225 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.533, + 0.225, + 0.71, + 0.361 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.712, + 0.225, + 0.888, + 0.361 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.372, + 0.907, + 0.429 + ], + "angle": 0, + "content": "Figure 7. Example of a failure case anomaly segmentation. For the chair on the road, labeled as anomaly in upper left, model predicts \"other-vehicle\" class, in lower left, with a high certainty, as indicated by MaxLogit scores on the bottom right." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.457, + 0.907, + 0.487 + ], + "angle": 0, + "content": "readers to the supplementary material for a more detailed investigation of the results." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.505, + 0.792, + 0.521 + ], + "angle": 0, + "content": "5. Conclusions and Future Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.532, + 0.907, + 0.835 + ], + "angle": 0, + "content": "In this paper, we introduced the STU dataset, designed for anomaly segmentation in LiDAR point clouds. The dataset was collected using eight cameras and a LiDAR sensor, which provides a surround view of environments through images and high-resolution LiDAR data. We extensively annotated point clouds with \"unlabeled\", \"anomaly\", and \"inlier\" classes, as well as introduced additional sequences without anomalies for training and evaluation. The results of the baseline methods indicate that there is a significant performance gap between 2D and 3D anomaly segmentation, which underscores numerous ongoing research opportunities in the field. Our objective with this paper is to provide researchers with annotated data and standardized benchmarks to facilitate the comparison of their anomaly segmentation approaches in 3D. We believe the dataset could be especially useful for the evaluation of methods that use temporal LiDAR data and multi-modal data in the future, and can drive innovation in developing more robust anomaly detection systems that can better handle complex real-world scenarios." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.841, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Acknowledgments. A. Nekrasov's acknowledges funding by BMBF project \"WestAI\" (grant no. 01IS22094D). J. S. Berrio Perez acknowledges funding by the German Academic Exchange Service (DAAD Project-ID 30001831)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11882" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.116, + 0.484, + 0.184 + ], + "angle": 0, + "content": "[1] Jens Behley, Martin Garbade, Andres Milioto, Jan Quenzel, Sven Behnke, Cyril Stachniss, and Juergen Gall. SemanticKITTI: A Dataset for Semantic Scene Understanding of LiDAR Sequences. In International Conference on Computer Vision (ICCV), 2019. 2, 3, 4, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.187, + 0.484, + 0.242 + ], + "angle": 0, + "content": "[2] Petra Bevandić, Ivan Kreso, Marin Oršić, and Siniša Šegvic. Simultaneous Semantic Segmentation and Outlier Detection in Presence of Domain Shift. In German Conference on Pattern Recognition (GCPR), 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.245, + 0.483, + 0.313 + ], + "angle": 0, + "content": "[3] Hermann Blum, Paul-Edouard Sarlin, Juan I. Nieto, Roland Y. Siegwart, and César Cadena. Fishyscapes: A benchmark for safe semantic segmentation in autonomous driving. International Conference on Computer Vision Workshop (ICCV'W), 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.316, + 0.483, + 0.383 + ], + "angle": 0, + "content": "[4] Hermann Blum, Paul-Edouard Sarlin, Juan Nieto, Roland Siegwart, and Cesar Cadena. The Fishyscapes Benchmark: Measuring Blind Spots in Semantic Segmentation. International Journal on Computer Vision (IJCV), 2021. 2, 3, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.387, + 0.483, + 0.442 + ], + "angle": 0, + "content": "[5] Daniel Bogdoll, Svenja Uhlemeyer, Kamil Kowol, and J Marius Zöllner. Perception datasets for anomaly detection in autonomous driving: A survey. In Intelligent Vehicles Symposium (IV), 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.445, + 0.483, + 0.54 + ], + "angle": 0, + "content": "[6] Daniel Bogdoll, Iramm Hamdard, Lukas Namgyu Rößler, Felix Geisler, Muhammed Bayram, Felix Wang, Jan Imhof, Miguel de Campos, Anushervon Tabarov, Yitian Yang, Hanno Gottschalk, and J. Marius Zöllner. AnoVox: A Benchmark for Multimodal Anomaly Detection in Autonomous Driving. In ECCV 2024 W-CODA workshop, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.544, + 0.483, + 0.625 + ], + "angle": 0, + "content": "[7] Holger Caesar, Varun Bankiti, Alex H. Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuScenes: A multimodal dataset for autonomous driving. In Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.63, + 0.483, + 0.67 + ], + "angle": 0, + "content": "[8] Jun Cen, Peng Yun, Junhao Cai, Michael Yu Wang, and Ming Liu. Open-set 3D Object Detection. In International Conference on 3D Vision (3DV), 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.673, + 0.483, + 0.728 + ], + "angle": 0, + "content": "[9] Anirudh S. Chakravarthy, Meghanana Reddy Ganesina, Peiyun Hu, Laura Leal-Taixe, Shu Kong, Deva Ramanan, and Aljosa Osep. Lidar Panoptic Segmentation in an Open World. International Journal on Computer Vision (IJCV), 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.73, + 0.483, + 0.813 + ], + "angle": 0, + "content": "[10] Robin Chan, Krzysztof Lis, Svenja Uhlemeyer, Hermann Blum, Sina Honari, Roland Siegwart, Pascal Fua, Mathieu Salzmann, and Matthias Rottmann. SegmentMeIfYouCan: A Benchmark for Anomaly Segmentation. In Proceedings of the Neural Information Processing Systems Track on Datasets and Benchmarks, 2021. 2, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.816, + 0.483, + 0.871 + ], + "angle": 0, + "content": "[11] Robin Chan, Matthias Rottmann, and Hanno Gottschalk. Entropy maximization and meta classification for out-of-distribution detection in semantic segmentation. In International Conference on Computer Vision (ICCV), 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[12] Liang-Chieh Chen, Yukun Zhu, George Papandreou, Florian Schroff, and Hartwig Adam. Encoder-Decoder with Atrous" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.116, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.548, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "Separable Convolution for Semantic Image Segmentation. In European Conference on Computer Vision (ECCV), 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.123, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[13] Bowen Cheng, Ishan Misra, Alexander G. Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention Mask Transformer for Universal Image Segmentation. In Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.193, + 0.905, + 0.247 + ], + "angle": 0, + "content": "[14] Bowen Cheng, Alexander G. Schwing, and Alexander Kirilov. Per-pixel classification is not all you need for semantic segmentation. In Neural Information Processing Systems (NeurIPS), 2021. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.25, + 0.905, + 0.331 + ], + "angle": 0, + "content": "[15] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.334, + 0.905, + 0.375 + ], + "angle": 0, + "content": "[16] Anja Delić, Matej Grić, and Siniša Šegvic. Outlier detection by ensembling uncertainty with negative objectness. In British Machine Vision Conference (BMVC), 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.378, + 0.905, + 0.432 + ], + "angle": 0, + "content": "[17] kuefeng Du, Zhaoning Wang, Mu Cai, and Yixuan Li. Vos: Learning what you don't know by virtual outlier synthesis. In International Conference on Learning Representations (ICLR), 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.434, + 0.905, + 0.503 + ], + "angle": 0, + "content": "[18] Stefano Gasperini, Alvaro Marcos-Ramiro, Michael Schmidt, Nassir Navab, Benjamin Busam, and Federico Tombari. Segmenting Known Objects and Unseen Unknowns without Prior Knowledge. In International Conference on Computer Vision (ICCV), 2023. 1, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.505, + 0.907, + 0.545 + ], + "angle": 0, + "content": "[19] A Geiger, P Lenz, C Stiller, and R Urtasun. Vision meets robotics: The KITTI dataset. In The International Journal of Robotics Research, 2013. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.548, + 0.905, + 0.588 + ], + "angle": 0, + "content": "[20] Matej Grcic, Petra Bevandic, and Sinisa Segvic. Dense open-set recognition with synthetic outliers generated by real nvp. ArXiv, abs/2011.11094, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.591, + 0.905, + 0.645 + ], + "angle": 0, + "content": "[21] Matej Grcic, Petra Bevandić, and Sinisa Šegvic. Densehybrid: Hybrid anomaly detection for dense open-set recognition. In European Conference on Computer Vision (ECCV), 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.648, + 0.905, + 0.702 + ], + "angle": 0, + "content": "[22] Krishnam Gupta, Syed Ashar Javed, Vineet Gandhi, and K. Madhava Krishna. MergeNet: A Deep Net Architecture for Small Obstacle Discovery. In International Conference on Robotics and Automation (ICRA), 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.705, + 0.905, + 0.799 + ], + "angle": 0, + "content": "[23] Jianhua Han, Xiwen Liang, Hang Xu, Kai Chen, Lanqing Hong, Jiageng Mao, Chaoqiang Ye, Wei Zhang, Zhenguo Li, Xiaodan Liang, and Chunjing Xu. SODA10M: A LargeScale 2D Self/Semi-Supervised Object Detection Dataset for Autonomous Driving. In Proceedings of the Neural Information Processing Systems Track on Datasets and Benchmarks, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.803, + 0.905, + 0.857 + ], + "angle": 0, + "content": "[24] Dan Hendrycks and Kevin Gimpel. A Baseline for Detecting Misclassified and Out-of-Distribution Examples in Neural Networks. In International Conference on Learning Representations (ICLR), 2018. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.86, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[25] Dan Hendrycks, Steven Basart, Mantas Mazeika, Andy Zou, Joe Kwon, Mohammadreza Mostajabi, Jacob Steinhardt, and Dawn Song. Scaling Out-of-Distribution Detection for Real-" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "11883" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.12 + ], + "angle": 0, + "content": "World Settings. In International Conference on Machine Learning (ICML), 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.191 + ], + "angle": 0, + "content": "[26] Dan Hendrycks, Steven Basart, Mantas Mazeika, Andy Zou, Mohammadreza Mostajabi, Jacob Steinhardt, and Dawn Xiaodong Song. Scaling out-of-distribution detection for real-world settings. In International Conference on Machine Learning, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.192, + 0.482, + 0.262 + ], + "angle": 0, + "content": "[27] Yen-Chang Hsu, Yilin Shen, Hongxia Jin, and Zsolt Kira. Generalized ODIN: Detecting Out-of-Distribution Image Without Learning From Out-of-Distribution Data. In Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.263, + 0.482, + 0.332 + ], + "angle": 0, + "content": "[28] Chengjie Huang, Van Duong Nguyen, Vahdat Abdelzad, Christopher Gus Mannes, Luke Rowe, Benjamin Therien, Rick Salay, and K. Czarnecki. Out-of-distribution detection for lidar-based 3d object detection. IEEE Intelligent Transportation Systems Conference (ITSC), 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.334, + 0.482, + 0.376 + ], + "angle": 0, + "content": "[29] Håkon Hukkelås and Frank Lindseth. Deepprivacy2: Towards realistic full-body anonymization. In Winter Conference on Applications of Computer Vision (WACV), 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.377, + 0.482, + 0.432 + ], + "angle": 0, + "content": "[30] Jaedong Hwang, Seoul Wug Oh, Joon-Young Lee, and Bohyung Han. Exemplar-Based Open-Set Panoptic Segmentation Network. In Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.434, + 0.482, + 0.503 + ], + "angle": 0, + "content": "[31] Sanghun Jung, Jungsoo Lee, Daehoon Gwak, Sungha Choi, and Jaegul Choo. Standardized Max Logits: A Simple yet Effective Approach for Identifying Unexpected Road Obstacles in Urban-Scene Segmentation. In International Conference on Computer Vision (ICCV), 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.504, + 0.482, + 0.559 + ], + "angle": 0, + "content": "[32] Alexander Kirillov, Kaiming He, Ross Girshick, Carsten Rother, and Piotr Dólar. Panoptic Segmentation. In Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.561, + 0.482, + 0.617 + ], + "angle": 0, + "content": "[33] Michael Kösel, Marcel Schreiber, Michael Ulrich, Claudius Gläser, and Klaus Dietmayer. Revisiting Out-of-Distribution Detection in LiDAR-based 3D Object Detection. In Intelligent Vehicles Symposium (IV), 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.618, + 0.482, + 0.673 + ], + "angle": 0, + "content": "[34] Balaji Lakshminarayanan, Alexander Pritzel, and Charles Blundell. Simple and Scalable Predictive Uncertainty Estimation using Deep Ensembles. In Neural Information Processing Systems (NeurIPS), 2017. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.675, + 0.482, + 0.731 + ], + "angle": 0, + "content": "[35] Seungjae Lee, Hyungtae Lim, and Hyun Myung. Patchwork++: Fast and robust ground segmentation solving partial under-segmentation using 3D point cloud. In Proc. IEEE/RSJ Int. Conf. Intell. Robots Syst., 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.732, + 0.482, + 0.813 + ], + "angle": 0, + "content": "[36] Kaican Li, Kai Chen, Haoyu Wang, Lanqing Hong, Chaoqiang Ye, Jianhua Han, Yukuai Chen, Wei Zhang, Chunjing Xu, Dit-Yan Yeung, et al. Coda: A real-world road corner case dataset for object detection in autonomous driving. In European Conference on Computer Vision (ECCV), 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.816, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[37] Chen Liang, Wenguan Wang, Jiaxu Miao, and Yi Yang. GMMSeg: Gaussian Mixture based Generative Semantic Segmentation Models. In Neural Information Processing Systems (NeurIPS), 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[38] Kira Maag, Robin Chan, Svenja Uhlemeyer, Kamil Kowol, and Hanno Gottschalk. Two Video Data Sets for Tracking" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "and Retrieval of Out of Distribution Objects. In Asian Conference on Computer Vision (ACCV), 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.121, + 0.905, + 0.204 + ], + "angle": 0, + "content": "[39] Jiageng Mao, Minzhe Niu, Chenhan Jiang, Hanxue Liang, Jingheng Chen, Xiaodan Liang, Yamin Li, Chaoqiang Ye, Wei Zhang, Zhenguo Li, Jie Yu, Hang Xu, and Chunjing Xu. One Million Scenes for Autonomous Driving: ONCE Dataset. In Neural Information Processing Systems (NeurIPS), 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.205, + 0.905, + 0.259 + ], + "angle": 0, + "content": "[40] Rodrigo Marcuzzi, Lucas Nunes, Louis Wiesmann, Jens Behley, and Cyril Stachniss. Mask-Based Panoptic LiDAR Segmentation for Autonomous Driving. In IEEE Robotics And Automation Letters (RAL), 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.26, + 0.905, + 0.315 + ], + "angle": 0, + "content": "[41] Nazir Nayal, Misra Yavuz, João F. Henriques, and Fatma Güney. RbA: Segmenting Unknown Regions Rejected by All. In International Conference on Computer Vision (ICCV), 2023. 1, 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.316, + 0.905, + 0.37 + ], + "angle": 0, + "content": "[42] Alexey Nekrasov, Rui Zhou, Miriam Ackermann, Alexander Hermans, and Matthias Rottmann Bastian Leibe. OoDIS: Anomaly Instance Segmentation Benchmark. arXiv preprint arXiv:2406.11835, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.371, + 0.905, + 0.427 + ], + "angle": 0, + "content": "[43] Gerhard Neuhold, Tobias Ollmann, Samuel Rota Bulo, and Peter Kontschieder. The mapillary vistas dataset for semantic understanding of street scenes. In International Conference on Computer Vision (ICCV), 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.428, + 0.905, + 0.496 + ], + "angle": 0, + "content": "[44] Lucas Nunes, Xieyuanli Chen, Rodrigo Marcuzzi, Aljosa Osep, Laura Leal-Taixe, Cyril Stachniss, and Jens Behley. Unsupervised Class-Agnostic Instance Segmentation of 3D LiDAR Data for Autonomous Vehicles. IEEE Robotics And Automation Letters (RAL), 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.497, + 0.905, + 0.566 + ], + "angle": 0, + "content": "[45] Peter Pinggera, Sebastian Ramos, Stefan Gehrig, Uwe Franke, Carsten Rother, and Rudolf Mester. Lost and Found: Detecting Small Road Hazards for Self-Driving Vehicles. In International Conference on Intelligent Robots and Systems (IROS), 2016. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.567, + 0.905, + 0.636 + ], + "angle": 0, + "content": "[46] Aldi Piroli, Vinzenz Dallabetta, Johannes Kopp, Marc Walesa, Daniel Meissner, and Klaus Dietmayer. LS-VOS: Identifying Outliers in 3D Object Detections Using Latent Space Virtual Outlier Synthesis. In IEEE Intelligent Transportation Systems Conference (ITSC), 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.637, + 0.905, + 0.691 + ], + "angle": 0, + "content": "[47] Shyam Nandan Rai, Fabio Cermelli, Dario Fontanel, Carlo Masone, and Barbara Caputo. Unmasking Anomalies in Road-Scene Segmentation. In International Conference on Computer Vision (ICCV), 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.692, + 0.905, + 0.748 + ], + "angle": 0, + "content": "[48] Jenny Seidenschwarz, Aljosa O'Sep, Francesco Ferroni, Simon Lucey, and Laura Leal-Taixe. SeMoLi: What Moves Together Belongs Together. In Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.748, + 0.905, + 0.802 + ], + "angle": 0, + "content": "[49] Aasheesh Singh, Aditya Kamireddypalli, Vineet Gandhi, and K. Madhava Krishna. LiDAR guided Small obstacle Segmentation. In International Conference on Intelligent Robots and Systems (IROS), 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.803, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[50] Nitish Srivastava, Geoffrey Hinton, Alex Krizhevsky, Ilya Sutskever, and Ruslan Salakhutdinov. Dropout: a simple way to prevent neural networks from overfitting. In Neural Information Processing Systems (NeurIPS), 2014. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.872, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[51] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou," + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11884" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.484, + 0.146 + ], + "angle": 0, + "content": "Yuning Chai, Benjamin Caine, et al. Scalability in perception for autonomous driving: Waymo open dataset. In Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.483, + 0.205 + ], + "angle": 0, + "content": "[52] DashcamCleaner Development Team. Dashcamcleaner: Censor identifiable information in videos from dashcam recordings. https://github.com/taehse/DashcamCleaner, 2024.4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.207, + 0.483, + 0.275 + ], + "angle": 0, + "content": "[53] Tefft, Brian C. . The Prevalence of Motor Vehicle Crashes Involving Road Debris, United States, 2011-2014. https://aaafoundation.org/wp-content/uploads/2017/12/RoadDebris_REPORT_2015.pdf, 2015. [Online]. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.277, + 0.483, + 0.346 + ], + "angle": 0, + "content": "[54] Yu Tian, Yuyuan Liu, Guansong Pang, Fengbei Liu, Yuanhong Chen, and Gustavo Carneiro. Pixel-wise Energy-biased Abstention Learning for Anomaly Segmentation on Complex Urban Driving Scenes. In European Conference on Computer Vision (ECCV), 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.348, + 0.483, + 0.416 + ], + "angle": 0, + "content": "[55] Tzu-Yun Tseng, Alexey Nekrasov, Malcolm Burdorf, Bastian Leibe, Julie Stephany Berrio Perez, Mao Shan, and Stewart Worrall. Panoptic-CUDAL Technical Report: Rural Australia Point Cloud Dataset in Rainy Conditions. arXiv preprint arXiv:2503.16378, 2025. 3, 4, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.418, + 0.483, + 0.473 + ], + "angle": 0, + "content": "[56] S. Verma, J. S. Berrio, S. Worrall, and E. Nebot. Automatic extrinsic calibration between a camera and a 3d lidar using 3d point and plane correspondences. In IEEE Intelligent Transportation Systems Conference (ITSC), 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.475, + 0.483, + 0.543 + ], + "angle": 0, + "content": "[57] Ignacio Vizzo, Tiziano Guadagnino, Benedikt Mersch, Louis Wiesmann, Jens Behley, and Cyril Stachniss. KISS-ICP: In Defense of Point-to-Point ICP – Simple, Accurate, and Robust Registration If Done the Right Way. In IEEE Robotics And Automation Letters (RAL), 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.545, + 0.483, + 0.599 + ], + "angle": 0, + "content": "[58] Kelvin Wong, Shenlong Wang, Mengye Ren, Ming Liang, and Raquel Urtasun. Identifying Unknown Instances for Autonomous Driving. In Conference on Robot Learning (CoRL), 2019. 2, 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.601, + 0.483, + 0.656 + ], + "angle": 0, + "content": "[59] Yajun Xu, Huan Hu, Xiaoya Zhu, Yibing Nan, Kai Wang, ZhaoXiang Liu, and Shiguo Lian. Raod: A benchmark for road abandoned object detection from video surveillance. IEEE Access, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.657, + 0.483, + 0.713 + ], + "angle": 0, + "content": "[60] Kadir Yilmaz, Jonas Schult, Alexey Nekrasov, and Bastian Leibe. Mask4Former: Mask Transformer for 4D Panoptic Segmentation. In International Conference on Robotics and Automation (ICRA), 2024. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.715, + 0.483, + 0.784 + ], + "angle": 0, + "content": "[61] Fisher Yu, Haofeng Chen, Xin Wang, Wenqi Xian, Yingying Chen, Fangchen Liu, Vashisht Madhavan, and Trevor Darrell. Bdd100k: A diverse driving dataset for heterogeneous multitask learning. In Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.784 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "11885" + } + ] +] \ No newline at end of file diff --git a/2025/Spotting the Unexpected (STU)_ A 3D LiDAR Dataset for Anomaly Segmentation in Autonomous Driving/7019444b-0523-423c-8391-a6c740f37e7c_origin.pdf b/2025/Spotting the Unexpected (STU)_ A 3D LiDAR Dataset for Anomaly Segmentation in Autonomous Driving/7019444b-0523-423c-8391-a6c740f37e7c_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..fbe6c553af5e1b1608d99757f7a149e782620e3e --- /dev/null +++ b/2025/Spotting the Unexpected (STU)_ A 3D LiDAR Dataset for Anomaly Segmentation in Autonomous Driving/7019444b-0523-423c-8391-a6c740f37e7c_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:09fc3f12927aaa8d25b476052a436fb0a022d6b8b833a36ff2e329ce4bb75c14 +size 5046303 diff --git a/2025/Spotting the Unexpected (STU)_ A 3D LiDAR Dataset for Anomaly Segmentation in Autonomous Driving/full.md b/2025/Spotting the Unexpected (STU)_ A 3D LiDAR Dataset for Anomaly Segmentation in Autonomous Driving/full.md new file mode 100644 index 0000000000000000000000000000000000000000..9731254fc18f882f72ef8d688253c4619fcb20ac --- /dev/null +++ b/2025/Spotting the Unexpected (STU)_ A 3D LiDAR Dataset for Anomaly Segmentation in Autonomous Driving/full.md @@ -0,0 +1,333 @@ +# Spotting the Unexpected (STU): A 3D LiDAR Dataset for Anomaly Segmentation in Autonomous Driving + +Alexey Nekrasov1 + +Malcolm Burdorf1 + +Stewart Worrall2 + +Bastian Leibe1 + +Julie Stephany Berrio Perez + +$^{1}$ RWTH Aachen University, $^{2}$ The University of Sydney + +https://vision.rwth-aachen.de/stu-dataset + +![](images/2c7d34edda03a4ca0ede6baff2084b80cb95434b140a7e10f268087ae07cafb1.jpg) +Figure 1. We present Spotting the Unexpected (STU) a novel anomaly segmentation dataset for autonomous driving. The dataset contains semantic and instance labels for out-of-distribution (OOD) objects, and includes surround-view setup with synchronized cameras. + +# Abstract + +To operate safely, autonomous vehicles (AVs) need to detect and handle unexpected objects or anomalies on the road. While significant research exists for anomaly detection and segmentation in 2D, research progress in 3D is underexplored. Existing datasets lack high-quality multimodal data that are typically found in AVs. This paper presents a novel dataset for anomaly segmentation in driving scenarios. To the best of our knowledge, it is the first publicly available dataset focused on road anomaly segmentation with dense 3D semantic labeling, incorporating both LiDAR and camera data, as well as sequential information to enable anomaly detection across various ranges. This capability is critical for the safe navigation of autonomous vehicles. We adapted and evaluated several baseline models for 3D segmentation, highlighting the challenges of 3D anomaly detection in driving environments. Our dataset and evaluation code will be openly available, facilitating the testing and performance comparison of different approaches. + +# 1. Introduction + +Autonomous vehicles have become an increasingly important part of modern transportation systems. Despite significant advances, these vehicles continue to face challenges in accurately detecting unexpected objects on the road. One particular challenge is road debris, which refers to objects that are not typically part of the roadway environment and pose significant safety risks. According to the AAA Foundation for Traffic Safety [53], road debris was a contributing factor in more than 50,000 crashes, 10,000 injuries, and 125 fatalities annually between 2011 and 2014. These objects can vary widely in size and appearance, making them particularly challenging to detect. Addressing this issue is critical for ensuring the safe and reliable operation of autonomous vehicles, as it directly impacts their ability to handle anomalies in real-world driving conditions. + +State-of-the-art anomaly segmentation and anomaly detection methods have received substantial attention in recent years [17, 37, 41], especially in the context of autonomous driving [11, 16, 18, 21, 31, 54]. One of the first + +benchmarks for evaluating anomaly segmentation is Lost and Found [45], whose data serves as the basis for many complex benchmarks [4, 10]. These benchmarks provide a wide range of possible road anomalies and obstacles, and recent research has focused on exploring the temporal [38] and multimodal [6, 36] nature of anomaly segmentation. + +While current autonomous driving datasets [1, 7, 51] use a setup with surround view cameras, radar, and high-resolution LiDAR sensors, existing anomaly segmentation and detection benchmarks focus primarily on image data [5] and mostly use Cityscapes [15] as the in-distribution dataset. Image-first benchmarks often lack the comprehensive sensor setup found in modern autonomous vehicles, and overlook the use of 3D modality for anomaly segmentation. Recent research that has begun to investigate the use of LiDAR [49] sensor for anomaly segmentation and detection, lacks high-resolution sensors found in modern autonomous vehicles. Several approaches have sought to find anomalies in validation sets of existing datasets [36, 46] or have created datasets using simulation [6] due to the difficulty of collecting and annotating such data in the real world. Despite these efforts, extending existing data or simulating anomalies only partially addresses the lack of accessible and high-quality data, shown in Table 1. + +To address this issue, we analyzed the labels of two widely used in-distribution (ID) datasets [1, 7] to identify an appropriate set of anomalies that would not overlap with objects in the ID training or validation datasets. This informed our creation of a real-world dataset containing challenging anomaly objects placed on roads, with annotated sequences in 3D LiDAR point clouds. Our dataset consists of multiple high-resolution LiDAR sequences acquired in challenging environments, with point-level labels for individual anomaly objects (shown in Figure 1). We acquired data in both rural and urban environments, with objects of varying degrees of difficulty, and include a number of real-world anomalies found on public roads. In addition, to evaluate in-distribution performance, we annotated additional sequences without anomalies with inlier labels to train and evaluate models to ensure a small domain gap between our and SemanticKITTI [1] datasets. + +To establish simple baselines for 3D LiDAR segmentation, we adapted the methods used in 2D anomaly segmentation [4, 24, 34, 41, 50] and applied them to the 3D domain, investigating anomaly segmentation methods for 3D LiDAR point clouds. Our models, trained jointly on the SemanticKITTI dataset and our own dataset, show strong segmentation performance for in-distribution data and novel additional sequences. However, we observed that the baseline methods struggled with out-of-distribution objects, and models tend to predict anomaly objects as inliers with a high degree of confidence. We believe that our dataset will benefit the community by allowing the evaluation of anomaly + +segmentation methods in the LiDAR domain and further exploration of multimodal anomaly segmentation methods. + +Our contributions provide a basis for the development of more advanced anomaly segmentation methods in the 3D LiDAR domain. Specifically, we propose a new anomaly segmentation dataset, and perform an analysis of the recorded and annotated data. The purpose of the paper is to provide practitioners with the necessary tools to evaluate their anomaly segmentation models. We also provide a set of baselines that can serve as a reference for future challenges and research directions in this area. + +# 2. Related work + +# 2.1. Anomaly Datasets + +Several multimodal datasets for autonomous driving are available, including Waymo [51], nuScenes [7], SemanticKITTI [1], among others. These datasets were collected in urban environments under typical driving conditions. Although these datasets are relevant for evaluating various tasks related to autonomous vehicle operation, they lack data for uncommon scenarios. One such scenario involves the presence of foreign objects on the road. Although this situation is rare during typical driving, autonomous vehicles must be able to recognize and respond to it safely. Here, assigning a unique label to each distinct object is impractical, since datasets have a limited set of labels, and objects on the road can vary widely. These objects are often grouped under a general category like "other object" and are excluded from the benchmarks due to their variability. + +In recent years, several datasets have been introduced to assess the detection of foreign objects on the road [4, 10, 25, 36, 38, 42, 45, 59]. Some of these datasets have been developed using data augmentation techniques or simulations, but only a few are based on data collected with real-world equipment in real driving scenarios. Simulations and augmentations are great tools, but they lack realism in terms of the sensor model [4] and the sizes of the objects [25]. However, the aforementioned datasets are limited to image data only. Autonomous vehicles use LiDAR sensors because they offer greater range, robustness to lighting conditions, and more precise 3D information about the environment compared to monocular depth estimations or stereo cameras. Therefore, incorporating 3D information is essential for accurately assessing road anomalies. + +Three datasets that integrate camera and LiDAR data are SOD [22], TOR4D [58] and CODA [36]. The CODA series is derived from validation sets of existing autonomous vehicle datasets. Still, it has the limitation that anomalous objects (e.g., traffic cones, barriers) may also appear in the training set, which is acceptable for the task of open-set segmentation, but contradicts the task of anomaly segmentation. Recently, a few methods for 3D anomaly detection + +Table 1. Comparison of our dataset with other community datasets [5]. SM = Semantic Mask, IM=Instance Mask, BB=Bounding Box + +
DatasetCameraLiDAR +BeamsSource#ODDLabelSize +Test/ValTemporalEnvironments +Test SequencesRelated Dataset
Fishscapes [3, 4]
FS Lost and Found1 RGBXStaged1SM275/100Xn/aLost and Found [45]
FS Static1 RGBXAugmented1SM1000/30Xn/aCityScapes [15]
CAOS [25]
StreetHazards1 RGBXSimulated1SM1500✓2 sim towns - 150 seq
BDD-Anomaly1 RGBXValidation Set3SM810Xn/aBDD100K [61]
SegmentMelfYouCan [10]
RoadAnomaly211 RGBXWeb Sourced1SM100/10Xn/a
RoadObstacle211 RGBXStaged1SM382/30✓9 streets - 90 seq
CODA [36]
CODA-KITTI1 RGB✓-64Validation Set6BB309Xn/aKITTI [19]
CODA-nuScenes1 RGB✓-32Validation Set17BB134Xn/aNuScenes [7]
CODA-ONCE1 RGB✓-40Validation Set32BB1057Xn/aONCE [39]
CODA22-ONCE1 RGB✓-40Validation Set29BB717Xn/aONCE [39]
CODA22-SODA10M1 RGBXValidation Set29BB4167Xn/aSODA [23]
Wuppertal ODD [38]
SOS1 RGBX-distStaged13IM1129✓9 streets - 20 seq
CWL1 RGBX-distSimulated18IM1210✓26 streets - 26 seq
Lost and Found [45]1 StereoXStaged42SM1068✓5 streets - 5 seq
SOD [49]1 RGB✓-16Staged1SM460/530X2 streets - 5 seq
WD-Pascal [2]1 RGBXSimulated1SM70Xn/a
Vistas-NP [20]1 RGBXClass exclusion4SM11167Xn/aMapillary Vistas [43]
STU (Ours)8 RGB✓-128Staged1IM8022/1960✓6 streets - 51 seqSemKITTI [1], P-CUDAL [55]
+ +have been proposed [28, 33, 46]. However, these methods use validation sequences to evaluate anomaly detection capabilities. The SOD dataset offers depth information for small road obstacles using a low-resolution 16-beam LiDAR, and that introduces a large domain gap between the training and test data. TOR4D and Rare4D [58] are the only datasets that contain real-world LiDAR data to benchmark road anomalies; however, these datasets are proprietary and not accessible to the public. We release out-of-distribution validation and in-distribution training sequences with a permissive license and open the private test set to submissions from the public. + +# 2.2. Open-Set Segmentation in 3D + +Open-set and unsupervised segmentation are closely related to the out-of-distribution detection task. Multiple methods were proposed for unsupervised segmentation [44, 48] and open-set prediction in 3D [9, 58]. Recent evaluations in 2D anomaly segmentation [18] have demonstrated that methods such as OSIS [58] and EOPSN [30] perform poorly on anomaly objects. This weakness stems from open-set methods' reliance on the presence of unknown instances inside the void regions during training, which may not generalize well to real-world scenarios. Similarly, MLUC [8] assesses open-set performance on the UDI [8] dataset, but this dataset is also not publicly accessible and makes the same assumptions as OSIS methods. In addition, open-set methods typically remove the ground plane [44] to obtain clusters of potential objects, but ground plane removal [35] methods are often sensitive to chosen hyperparameters and yield noisy results, missing many smaller objects. For the evaluation of anomaly segmentation, it is important to have a test dataset that is free of anomaly objects which might + +occur during training. In contrast to open-set methods, our work focuses on collecting and densely annotating real-world anomaly objects, ensuring that there is no intersection between in-distribution training sets and out-of-distribution test sets. + +# 2.3. Anomaly Segmentation in Autonomous Driving + +To tackle the problem of anomaly segmentation, multiple works [4, 5, 10] have recently been proposed for the problem of 2D anomaly segmentation. Earlier works [4, 11, 21, 24, 34, 37, 54] were applied to per-pixel segmentation models such as DeepLab [12]; however, recently many works were focused on MaskFormer-style models [13, 14]. A line of work [16, 41, 47] proposed different methods to adapt Mask2Former, exploring the mechanism of mask prediction, where multiple masks might ignore certain regions of an image. These models provide more reliable anomaly scoring mechanisms not limited to per-pixel predictions and achieve higher performance on the anomaly segmentation task. For our benchmark, we decided to evaluate methods that also use MaskFormer-style models and investigate the application of the Mask4Former [60] model for 3D anomaly segmentation. We obtain all our scores using a MaskFormer-Style model, laying the foundation for future baselines. + +# 3. The Spotting the Unexpected (STU) Dataset + +Our data collection platform consists of a rigid frame mounted on a vehicle and equipped with eight hardware-triggered cameras and a LiDAR sensor to ensure synchronized data collection. The LiDAR and the cameras were calibrated using the method described in [56], with the cal + +![](images/badb56129a8a822729d71d87dec28a53824536a8b36d206976c7f4f4870ccffb.jpg) +(a) Naturalistic data collection. + +![](images/422643e6cc0df06d09236827c709acec0b752a218bcbf68049418c65e74b86fc.jpg) +(b) Data collection in controlled environment. +Figure 2. Data collection conducted in a naturalistic manner (a) and controlled environment (b) with objects on the road. + +ization process repeated for each camera. This setup follows the configuration discussed in Panoptic-CUDAL [55]. We refer to the Supplementary Material for detailed information on the vehicle setup. + +# 3.1. Data collection. + +The data was collected using two sets of conditions: one in a naturalistic environment where objects were found on public roads, and one in a controlled environment. In the naturalistic scenario, several hours of driving data were recorded, with specific events marked when objects appeared on the road. Figure 2a shows a bucket on the road recorded while driving on a highway. + +The controlled environment data collection protocol was rigorous for safety reasons. We employed four people: two spotters positioned at either end of a low-traffic and low-speed road, one responsible for placing objects on the road, and a driver. The spotters would signal when it was safe to place the objects, and once positioned, the driver would begin recording the scene. If one of the spotters reported an approaching vehicle, the data recording was stopped, and the object was immediately removed. This was done to avoid disrupting the normal flow of traffic. The vehicle was driven at a maximum speed of $50\mathrm{km / h}$ during data recording. Figure 2b shows a computer monitor placed on the road during sensor data logging. Some of the objects placed on the road are also shown in Figure 3. + +Postprocessing. After data collection, post-processing is required to obtain vehicle poses and anonymize the images. Point cloud registration was performed using KISS ICP [57], a LiDAR odometry pipeline. KISS ICP includes point cloud motion compensation, subsampling, adaptive thresholding to determine correspondences, and LiDAR pose estimation. The calculated LiDAR pose was then exported in KITTI format as required for the labeling tool [1]. + +Ethical considerations for this dataset include anonymizing camera images. Identifiable data, such as faces and license plates, are processed using DashcamCleaner [52] and DeepPrivacy2 [29]. DashcamCleaner uses a license plate detector to locate and blur license plates, while DeepPrivacy2 detects facial features and generates new, unidentifiable faces to replace the original ones. Examples of anonymization can be found in the supplementary material. + +# 3.2. Labeling + +We define anomalies as objects that may pose a danger to the autonomous vehicle and its passengers, especially those present on the driving surface and not present in the training dataset. In this work, we focus on road anomalies and do not consider unusual situations such as vehicle collisions or unexpected driver behavior, and we focus on the fine-grained anomaly segmentation. In particular, we follow previous work [4, 45] and focus on objects that are not present in the training set. + +We labeled the data using the SemanticKITTI labeler [1]. To generate initial labels for the scenes, we used pseudolabeling techniques using a pre-trained SemanticKITTI model. We instruct our annotators to group all points belonging to inliers as a single semantic class, questionable points as unlabeled, and anomaly points as the outlier class. We instruct our labelers that it is very important to label inliers and outliers carefully while allowing for larger unlabeled regions where the annotator may be challenged. + +Annotation process. The domain gap between urban environments and SemanticKITTI street scenes is minimal, especially in regions close to the vehicle. This similarity allows the acquisition of reliable initial estimates. The annotation process was performed by three annotators, each spending eight hours a week over several months. Each annotator spent approximately four hours on the initial labeling of each scene and an additional two hours on refining the final labels. To ensure accuracy and consistency, the annotators cross-validated each other's work. To further improve the quality of the labels, we examined the predictions of the baseline methods to see if any known objects were missed. In particular, current methods tend to be highly sensitive to objects that are present but ignored during training, such as those classified in the category "other object". We annotate these objects as unlabeled and ignore them in the evaluations, as they could alter the results of the anomaly segmentation methods. + +Differences to SemanticKITTI. In contrast to SemanticKITTI, which provides labels up to a range of 50 meters, we labeled all points visible to the LiDAR sensor. This was facilitated by using a LiDAR with a higher vertical resolution. The maximum distance resulting from an anomaly in our dataset is 150 meters, allowing for comprehensive evaluations. However, since we train on the Se + +![](images/db4ed5b85024c82a54f0d517cd1da0f42888b59e123237cbeaf82741f84c17d0.jpg) + +![](images/eaa1891ae8b0867e8ef46c72efe3c1c98265bbdbe5f36f5acf85fa8b47f3e92e.jpg) + +![](images/992a78b49d4cb3ac015ecbb803e907eca49e7a91ef35285667a07c5b649b1ead.jpg) + +![](images/4296654edfbc219d1703f6fe849d0d4dba4fae7f18e3d2edb82240d739641bd6.jpg) + +![](images/6e3d77629b6f9024d72fbffc971969fb91123ddca990a6608b8849691aca530b.jpg) + +![](images/d1eacadf2ddd0b34cb990a3ed4079bcbe30d70cf938dc0262a75662d5afa5fc2.jpg) + +![](images/ae700115971280b9c81f5f5d07c28515d56012ec5de26a2555fadc1537a01bc7.jpg) + +![](images/2fce046319925b5aa325f021b22267fe460061bba12dba07f87b256bafadd2c8.jpg) + +![](images/464c32cce11ea34ee67ded114a1d22365275f8af1560b318f6f0b11e21ccd728.jpg) + +![](images/92700289051bba58257434daa2f1401f651172f2c896499b74b36ced527e78c3.jpg) + +![](images/1840258f7a63031bd57972978bb92ea801de1b829b36d53f01b86e9327fd6ce2.jpg) + +![](images/71276fa0f6e972dd75663d194e7e4e195a916e2b68bc3f910b03293d7d99258c.jpg) + +![](images/6d7b36eac6babb141c0896375df5963cbf5a5c666244ec3b258085688f3268c4.jpg) + +![](images/7cc67798adcd0837f4f029bed306742399587163a34c441f7649dc300a752f96.jpg) + +![](images/13870c7795dfe6c37c9f5e194aa4fdeeab6682eae3fd774c5b298d753a3ab2a6.jpg) + +![](images/be03b39e639b38d8a17e342688b188f7b7a089628767adcdd8260b77cae4ea53.jpg) + +![](images/4485c45f4e2e4eb0b5815d215d3eb2d8f47a9a397616274276b1a33b27e1e067.jpg) + +![](images/e274c448a26dc21a5f1029348cb97c289d32dc57962fc8497bae9b0d84a20573.jpg) + +![](images/5b55c2119689c1aac9a5f19d06936a4e4bb5bb0931958a673d9620b250db1fbc.jpg) +Figure 3. Different anomalies in the STU dataset. Different objects on the road used for staged data collection. We pick objects such that have no intersection with the inlier dataset and place them on roads in different locations and illumination conditions. Objects might touch each other, be very small, as large as a chair or a surf board, and could cause an accident if a car would drive over them. + +![](images/b0d68c0089f1d047d0d14e3fe243a243aa92a8e3038bfe409220c84baa998cb5.jpg) + +![](images/81ae7030ddd3e101f381e4052f854d29d1e31a88a5f3401eee1f2c5bfdf6906d.jpg) + +![](images/6cf8158d2c34d7bac0194cc961fd67cbfae9a8ff510ad38ae0829d9ee4009de6.jpg) + +![](images/f95f6e533d073052d9d51a0406fbe9260dcfc36481ab0f5a03ebaa2d727094e5.jpg) + +![](images/7dbff6d23082585447662b9eeae32681468170763bf26c7c14b2ed508a516d32.jpg) + +manticKITTI dataset, which contains only labels up to 50 meters, we limit our evaluations to the 50 meter range. The sensor setup has 128 LiDAR beams, compared to 64 beams in SemanticKITTI. + +# 3.3. Dataset analysis + +We recorded 70 sequences with anomalies, which we fully annotated for anomaly segmentation. Two of these labeled sequences were captured in a naturalistic manner, while the remaining 68 were staged to include various anomalies. We designated 19 sequences as a validation set and 51 as a closed test set for benchmarking purposes. In addition, we recorded and annotated two sequences with no anomalies for training and validation of the inlier performance "STU-inlier", to reduce the domain gap between our dataset and SemanticKITTI. + +Our dataset defines two primary classes: inliers and outliers. Inliers encompasses classes that are used to train a model, and contains objects expected on the road, such as vehicles, pedestrians, and road infrastructure. Outliers, on the other hand, refer to out-of-distribution elements or anomalies that are atypical or unexpected in the driving scene, shown in Figure 3. One additional class, unlabeled is ignored in the evaluation and typically includes classes that the model has seen during training, but was not supervised in, such as parking meters, utility boxes, and lamps. We provide anomaly annotations at both the instance and semantic levels, enabling a comprehensive evaluation of models for semantic and instance segmentation in 3D environments. + +Each outlier object in our dataset has an instance la- + +bel unique for a sequence. Figure 4 presents histograms that display the distribution of key properties for these instances, including the average and maximum number of LiDAR points per instance, instance heights, and the number of instances per scan. This analysis provides insights into the structural characteristics and variability of data within point clouds. Specifically, Figure 4 indicates that most instances have very few LiDAR points on average, but a sequence may have multiple instances per sequence. + +Furthermore, Figure 5 presents a scatter plot of detected anomalies within the vehicle frame, highlighting a higher density of anomalies along the longitudinal $(X)$ axis both in front and behind the vehicle. The accompanying histograms depict the distribution of the anomalies along each axis, noting that the anomalies are identified at distances ranging up to 150 meters in front of the vehicle. Figure 6 explores the relationship between the distance to an anomaly and its size, showing that the proportion of LiDAR point cloud obstacles decreases as the distance increases. This trend is due to the radial nature of LiDAR, which has lower vertical resolution at greater ranges. For benchmarking purposes, we focus exclusively on instances with more than five LiDAR points, highlighted in blue, ensuring that the evaluation targets detectable anomalies. + +# 3.4. Evaluated Baselines + +The formulation of the anomaly segmentation task restricts training to avoid direct supervision of OOD objects. Generally, OOD objects are not present in the training set, but could be present during inference [18]. Since there are no methods that could be applied directly for our setup, we + +![](images/776e849cd15d0db2f2a4ad196bc52ccec72beafd6d8398d0bd31966b65966389.jpg) +(a) Avg. num of points/sequence. + +![](images/ece6301e8fd8eeb1d66ad5dfc22bc445b171effe6e1d6be177a0cdc06318fa80.jpg) +Figure 4. Anomaly Instance Properties. A typical recorded anomaly has on average less than 50 points per sequence (a), with less than 300 (b) points at maximum, and a maximum height below one meter (c). We record up to nine individual anomaly instances in the same sequence (d). + +![](images/d17ca69cf4e26cc946d2185c567fae78f376de4b48f7cedfce782756d93391d1.jpg) +(b) Max. points per instance. + +![](images/1b5a0783346b2817ba682da757f107ca42013d480cfa98e72ba1a3d589e83b99.jpg) +(c) Max. height of instances (m). +(d) Numb. of instances/sequence. + +![](images/641ba3252959ed7b996229cd237db7c621162e36929ebcdef64faffb36fd6ae2.jpg) +Figure 5. Distribution of anomalies along the vehicle's reference frame. Most of the points appear around the vehicle. + +![](images/a0dd9871c54e0d67be1c93fb2eb0bf007782db7fcad52a7c47004efe7f9cc8c0.jpg) +Figure 6. We follow SemanticKitti's maximum distance threshold and evaluate instances that are within a 50 meter radius of the vehicle. In addition, we constrain evaluation of point and object level metrics to a minimum of 5 points. + +adopt several baseline methods from the 2D anomaly segmentation to 3D LiDAR segmentation. First, we evaluate simple baselines, such as Max-Logit [24], Monte-Carlo Dropout [50], and Deep Ensembles [34]. Secondly, we train one of the models to predict ignored regions during train + +ing, which could be used to classify “void” points during testing [4]. We also adapt RbA [41], one of the state-of-the-art methods for image anomaly segmentation, to the 3D LiDAR segmentation method Mask4Former-3D [60]. + +Deep Ensembles, MC Dropout, and MaxLogit. Deep Ensembles [34] and MC Dropout [50] are very common methods that practitioners use due to ease of implementation and application. For Deep Ensembles, we train three Mask4Former-3D [60] models. For MC dropout, we train one Mask4Former-3D model with enabled dropout in the transformer decoder for self- and cross-attention operations, and during inference we repeat the transformer decoder operation 50 times. As a simple baseline that uses only one model and a single forward pass, we use the MaxLogit [24] method. + +RbA [41] is a method that uses the ability of MaskFormer-Style models [13, 14] to predict multiple overlapping segmentation masks. MaskFormer models have an option to disable queries and might not predict any masks if no objects are present. Points for which no masks were predicted will have the highest anomaly score. The Mask4Former model follows the same architectural paradigm and has a query deactivation mechanism that allows for no prediction in uncertain regions. This allows us to use the RbA method to generate anomaly scores within our setup. Unlike the original RbA model setup, we did not modify the transformer decoder. + +Void Classifier We follow a common baseline [4] by treating the unlabeled regions as an additional semantic class during training. + +# 4. Baseline Evaluation + +Training data. We use SemanticKITTI [1], Panoptic-CUDAL [55], and an additional recorded sequence from to jointly train our models. The Panoptic-CUDAL dataset, contains six sequences for training and one for validation, and was mostly recorded in rural locations, while our dataset consists mainly of urban scenes. To reduce the do + +Table 2. Anomaly Segmentation Performance of the selected baseline methods. + +
MethodOoD DataPoint-Level OODObject-Level OOD
AUROC ↑FPR@95 ↓AP ↑RecallQSQRQUQPQ
Deep Ensemble [34]X86.7458.055.1716.7584.4910.4314.168.81
MC Dropout [50]X61.5182.370.112.2586.721.952.141.86
Max Logit [24]X84.5381.490.9526.1483.062.1321.711.77
Void Classifier [4]✓85.9978.603.9217.6484.408.1914.896.91
RbA [41]X66.38100.00.8124.0483.283.2320.022.69
+ +Table 3. Closed-set performance of the methods on validation sets of SemanticKITTI [1] and STU-inlier dataset. * denotes a model trained only on SemanticKITTI dataset. + +
MethodSemanticKITTISTU-inlier Dataset
PQSQRQPQSQRQ
Mask-PLS* [40]59.976.469.139.667.050.5
Mask-PLS [40]55.676.064.550.969.462.2
Mask4Former-3D* [60]61.981.371.642.870.653.2
Mask4Former-3D [60]60.781.070.452.773.864.7
+ +main gap, we have recorded and annotated two additional inlier scenes from an urban setting "STU-inlier", one for training and one for evaluation. These training sequences closely resemble the environments in which anomaly scenes were recorded. We train models with additional data recorded from the same vehicle setup as Panoptic-CUDAL and labeled the data with 19 SemanticKITTI labels. However, SemanticKITTI and our recorded dataset contain "other-object" that acts as a miscellaneous class for all objects that do not fit other-class definitions. A typical object is a bench, a garbage bin, or an advertisement post. To strictly define anomalies, we analyze all other-objects present in our training set, as well as the "Movable_Object.Debris" and "Pushable.Pullable" class objects in the NuScenes training set. We specifically design our dataset such that we do not have an intersection between anomalous objects and the aforementioned classes. Our Void Classifier training procedure consists of training on this unlabeled class that is present in the training data. + +Models. We train MaskPLS [40] and Mask4Former [60] models jointly on both SemanticKitti and our training sequences. These models achieve state-of-the-art results on the SemanticKITTI [1] dataset for single-scan and multi-scan settings. However, while our dataset supports evaluation of temporal sequences, we performed our evaluation and training on single scans only. In our experiments, we refer to Mask4Former trained on single scans as Mask4Former-3D. All anomaly segmentation methods use Mask4Former-3D as the main model. + +Metrics. To evaluate the performance of our methods, we use common metrics in the anomaly segmentation commu + +nity in 2D and panoptic segmentation in 3D LiDAR. Following SemanticKitti, we evaluated both inlier and outlier performance on points within 50 meters of the vehicle, and on objects with at least five points. + +Point-Level Metrics. To evaluate the performance of the anomaly segmentation at the point level, we use the common metrics of Average Precision (AP), False-Positive Rate at $95\%$ True-Positive Rate (FPR95), and Area Under Receiver Operating characteristic Curve (AUROC). Commonly used in the anomaly segmentation community [4]. + +Object-Level Metrics. To be consistent with the panoptic segmentation setup, we use panoptic metrics to evaluate the model performance on inlier classes. For the inlier data, we compute the Panoptic Quality (PQ) on the SemanticKitti classes. We report detailed class scores in the Supplementary. In contrast to anomaly segmentation, the open task formulation emphasizes the recall of all objects. Formally, the Unknown Quality (UQ) metric is used in such a setting [58]. However, in the case of anomaly segmentation, since anomalies can pose serious threats to the vehicle, it is important to predict as few false positives as possible, while also producing accurate segmentation labels for anomaly objects. Panoptic Quality uses an F-score and penalizes the prediction of multiple false positives. Thus, for anomaly segmentation baselines, we report both Panoptic Quality and Unknown Quality on a single class. For inlier performance, we report the Panoptic Quality score on the SemanticKITTI classes. + +Panoptic Quality [32] is the geometric mean between Recognition Quality (RQ) and Segmentation Quality (SQ) for a class $c$ : + +$$ +\mathrm {P Q} _ {c} = \underbrace {\frac {\sum_ {(p , g) \in T P _ {c}} \operatorname {I o U} (p , g)}{| T P _ {c} |}} _ {\text {S e g m e n t a t i o n Q u a l i t y (S Q)}} \times \underbrace {\frac {| T P _ {c} |}{| T P _ {c} | + \frac {1}{2} | F P _ {c} | + \frac {1}{2} | F N _ {c} |}} _ {\text {R e c o n g i t i o n Q u a l i t y (R Q)}} +$$ + +where $TP$ , $FP$ , and $FN$ are True Positives, False Positives, and False Negatives for objects with Intersection over Union (IoU) higher than $50\%$ . The final PQ metric is calculated as an average over all classes $c$ . Ignore points are removed from the scene prior to evaluation and erroneous predictions in the ignore region are not penalized. + +To evaluate the recall of anomaly objects, we use the Unknown Quality (UQ) metric [58]: + +$$ +\mathrm {U Q} = \underbrace {\frac {\sum_ {(p , g) \in T P} \operatorname {I o U} (p , g)}{| T P |}} _ {\text {S e g m e n t a t i o n Q u a l i t y (S Q)}} \times \underbrace {\frac {| T P |}{| T P | + | F N |}} _ {\text {R e c a l l Q u a l i t y (R e c a l l Q)}}. +$$ + +The metric does not penalize $FP$ and allows for the evaluation of object recall. However, the metric counts objects as $TP$ only if they have an IoU of at least $50\%$ with ground truth. + +# 4.1. Experiments + +Inlier Performance on SemanticKitti and STU-inlier dataset. We train models jointly with SemanticKITTI [1], Panoptic-CUDAL [55] and our additional sequence, and evaluate the SemanticKITTI and STU-inlier validation sets. The jointly trained model demonstrates improved performance on our STU-inlier dataset compared to models trained solely on the SemanticKITTI dataset, suggesting a small domain gap between our dataset and SemanticKITTI. Evaluation on the proposed STU-inlier sequence shows good segmentation performance for inlier classes that is critical for anomaly segmentation. The metrics for the performance of the baseline method in the validation sets are provided in Tab. 3. + +Out-of-Distribution (OOD) Performance. We evaluate baseline methods for anomaly segmentation at both the point and object levels. Our findings indicate that directly transferring methods from 2D to 3D does not yield similar performance. Our point clouds typically contain around 100,000 points, with only a few dozen being anomalous, as shown in Fig. 6. Due to the significant class imbalance, point-level scores are generally low, highlighting the difficulty of the task. A high FPR@95, as noted in ODIN [27], suggests that models often misclassify road objects as inlier classes. Since LiDAR point clouds have lower resolution at a distance, predicting a small anomaly object consisting of a few points is a hard task. It is common for large anomalous objects to be predicted as pedestrians or other vehicles, as depicted in Fig. 7. Anomalous objects frequently occur in familiar contexts, which leads the model to confidently assign inlier labels to such anomalies in many cases. This behavior contrasts significantly with models trained on 2D data, where a simple max logit approach [26] outperforms Deep Ensembles [10]. However, the Deep Ensemble model stands out by producing fewer false positives at the point level, resulting in a lower FPR@95 and a higher PQ metric. Despite this reduction in false positives, the average precision (AP) remains low compared to 2D datasets. We attribute this to the high confidence with which individual models predict inlier objects, with disagreements between models contributing to the improved FPR@95. We refer + +![](images/028a40dd22490729cfbd1b5011aeae64ba6e4b5f51f913f5b4c8c9fd131622c8.jpg) + +![](images/b035bc2638a0727b7506a25816e1bc56b5a8351b2f6d15dad5ec66fa3abbb146.jpg) + +![](images/15be6f85a2b4e4fac3a538396f6f3154fac2efb04aa1c2cf4b71cfdc44cb6be3.jpg) +Figure 7. Example of a failure case anomaly segmentation. For the chair on the road, labeled as anomaly in upper left, model predicts "other-vehicle" class, in lower left, with a high certainty, as indicated by MaxLogit scores on the bottom right. + +![](images/06def6696d786f3e297e1f8939c2997b3ee945b57c817a77910f4935e3ac3400.jpg) + +readers to the supplementary material for a more detailed investigation of the results. + +# 5. Conclusions and Future Work + +In this paper, we introduced the STU dataset, designed for anomaly segmentation in LiDAR point clouds. The dataset was collected using eight cameras and a LiDAR sensor, which provides a surround view of environments through images and high-resolution LiDAR data. We extensively annotated point clouds with "unlabeled", "anomaly", and "inlier" classes, as well as introduced additional sequences without anomalies for training and evaluation. The results of the baseline methods indicate that there is a significant performance gap between 2D and 3D anomaly segmentation, which underscores numerous ongoing research opportunities in the field. Our objective with this paper is to provide researchers with annotated data and standardized benchmarks to facilitate the comparison of their anomaly segmentation approaches in 3D. We believe the dataset could be especially useful for the evaluation of methods that use temporal LiDAR data and multi-modal data in the future, and can drive innovation in developing more robust anomaly detection systems that can better handle complex real-world scenarios. + +Acknowledgments. A. Nekrasov's acknowledges funding by BMBF project "WestAI" (grant no. 01IS22094D). J. S. Berrio Perez acknowledges funding by the German Academic Exchange Service (DAAD Project-ID 30001831). + +# References + +[1] Jens Behley, Martin Garbade, Andres Milioto, Jan Quenzel, Sven Behnke, Cyril Stachniss, and Juergen Gall. SemanticKITTI: A Dataset for Semantic Scene Understanding of LiDAR Sequences. In International Conference on Computer Vision (ICCV), 2019. 2, 3, 4, 6, 7, 8 +[2] Petra Bevandić, Ivan Kreso, Marin Oršić, and Siniša Šegvic. Simultaneous Semantic Segmentation and Outlier Detection in Presence of Domain Shift. In German Conference on Pattern Recognition (GCPR), 2019. 3 +[3] Hermann Blum, Paul-Edouard Sarlin, Juan I. Nieto, Roland Y. Siegwart, and César Cadena. Fishyscapes: A benchmark for safe semantic segmentation in autonomous driving. International Conference on Computer Vision Workshop (ICCV'W), 2019. 3 +[4] Hermann Blum, Paul-Edouard Sarlin, Juan Nieto, Roland Siegwart, and Cesar Cadena. The Fishyscapes Benchmark: Measuring Blind Spots in Semantic Segmentation. International Journal on Computer Vision (IJCV), 2021. 2, 3, 4, 6, 7 +[5] Daniel Bogdoll, Svenja Uhlemeyer, Kamil Kowol, and J Marius Zöllner. Perception datasets for anomaly detection in autonomous driving: A survey. In Intelligent Vehicles Symposium (IV), 2023. 2, 3 +[6] Daniel Bogdoll, Iramm Hamdard, Lukas Namgyu Rößler, Felix Geisler, Muhammed Bayram, Felix Wang, Jan Imhof, Miguel de Campos, Anushervon Tabarov, Yitian Yang, Hanno Gottschalk, and J. Marius Zöllner. AnoVox: A Benchmark for Multimodal Anomaly Detection in Autonomous Driving. In ECCV 2024 W-CODA workshop, 2024. 2 +[7] Holger Caesar, Varun Bankiti, Alex H. Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuScenes: A multimodal dataset for autonomous driving. In Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2, 3 +[8] Jun Cen, Peng Yun, Junhao Cai, Michael Yu Wang, and Ming Liu. Open-set 3D Object Detection. In International Conference on 3D Vision (3DV), 2021. 3 +[9] Anirudh S. Chakravarthy, Meghanana Reddy Ganesina, Peiyun Hu, Laura Leal-Taixe, Shu Kong, Deva Ramanan, and Aljosa Osep. Lidar Panoptic Segmentation in an Open World. International Journal on Computer Vision (IJCV), 2024. 3 +[10] Robin Chan, Krzysztof Lis, Svenja Uhlemeyer, Hermann Blum, Sina Honari, Roland Siegwart, Pascal Fua, Mathieu Salzmann, and Matthias Rottmann. SegmentMeIfYouCan: A Benchmark for Anomaly Segmentation. In Proceedings of the Neural Information Processing Systems Track on Datasets and Benchmarks, 2021. 2, 3, 8 +[11] Robin Chan, Matthias Rottmann, and Hanno Gottschalk. Entropy maximization and meta classification for out-of-distribution detection in semantic segmentation. In International Conference on Computer Vision (ICCV), 2021. 1, 3 +[12] Liang-Chieh Chen, Yukun Zhu, George Papandreou, Florian Schroff, and Hartwig Adam. Encoder-Decoder with Atrous + +Separable Convolution for Semantic Image Segmentation. In European Conference on Computer Vision (ECCV), 2018. 3 +[13] Bowen Cheng, Ishan Misra, Alexander G. Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention Mask Transformer for Universal Image Segmentation. In Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 3, 6 +[14] Bowen Cheng, Alexander G. Schwing, and Alexander Kirilov. Per-pixel classification is not all you need for semantic segmentation. In Neural Information Processing Systems (NeurIPS), 2021. 3, 6 +[15] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 2, 3 +[16] Anja Delić, Matej Grić, and Siniša Šegvic. Outlier detection by ensembling uncertainty with negative objectness. In British Machine Vision Conference (BMVC), 2024. 1, 3 +[17] kuefeng Du, Zhaoning Wang, Mu Cai, and Yixuan Li. Vos: Learning what you don't know by virtual outlier synthesis. In International Conference on Learning Representations (ICLR), 2021. 1 +[18] Stefano Gasperini, Alvaro Marcos-Ramiro, Michael Schmidt, Nassir Navab, Benjamin Busam, and Federico Tombari. Segmenting Known Objects and Unseen Unknowns without Prior Knowledge. In International Conference on Computer Vision (ICCV), 2023. 1, 3, 5 +[19] A Geiger, P Lenz, C Stiller, and R Urtasun. Vision meets robotics: The KITTI dataset. In The International Journal of Robotics Research, 2013. 3 +[20] Matej Grcic, Petra Bevandic, and Sinisa Segvic. Dense open-set recognition with synthetic outliers generated by real nvp. ArXiv, abs/2011.11094, 2020. 3 +[21] Matej Grcic, Petra Bevandić, and Sinisa Šegvic. Densehybrid: Hybrid anomaly detection for dense open-set recognition. In European Conference on Computer Vision (ECCV), 2022. 1, 3 +[22] Krishnam Gupta, Syed Ashar Javed, Vineet Gandhi, and K. Madhava Krishna. MergeNet: A Deep Net Architecture for Small Obstacle Discovery. In International Conference on Robotics and Automation (ICRA), 2018. 2 +[23] Jianhua Han, Xiwen Liang, Hang Xu, Kai Chen, Lanqing Hong, Jiageng Mao, Chaoqiang Ye, Wei Zhang, Zhenguo Li, Xiaodan Liang, and Chunjing Xu. SODA10M: A LargeScale 2D Self/Semi-Supervised Object Detection Dataset for Autonomous Driving. In Proceedings of the Neural Information Processing Systems Track on Datasets and Benchmarks, 2021. 3 +[24] Dan Hendrycks and Kevin Gimpel. A Baseline for Detecting Misclassified and Out-of-Distribution Examples in Neural Networks. In International Conference on Learning Representations (ICLR), 2018. 2, 3, 6, 7 +[25] Dan Hendrycks, Steven Basart, Mantas Mazeika, Andy Zou, Joe Kwon, Mohammadreza Mostajabi, Jacob Steinhardt, and Dawn Song. Scaling Out-of-Distribution Detection for Real- + +World Settings. In International Conference on Machine Learning (ICML), 2022. 2, 3 +[26] Dan Hendrycks, Steven Basart, Mantas Mazeika, Andy Zou, Mohammadreza Mostajabi, Jacob Steinhardt, and Dawn Xiaodong Song. Scaling out-of-distribution detection for real-world settings. In International Conference on Machine Learning, 2022. 8 +[27] Yen-Chang Hsu, Yilin Shen, Hongxia Jin, and Zsolt Kira. Generalized ODIN: Detecting Out-of-Distribution Image Without Learning From Out-of-Distribution Data. In Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 8 +[28] Chengjie Huang, Van Duong Nguyen, Vahdat Abdelzad, Christopher Gus Mannes, Luke Rowe, Benjamin Therien, Rick Salay, and K. Czarnecki. Out-of-distribution detection for lidar-based 3d object detection. IEEE Intelligent Transportation Systems Conference (ITSC), 2022. 3 +[29] Håkon Hukkelås and Frank Lindseth. Deepprivacy2: Towards realistic full-body anonymization. In Winter Conference on Applications of Computer Vision (WACV), 2023. 4 +[30] Jaedong Hwang, Seoul Wug Oh, Joon-Young Lee, and Bohyung Han. Exemplar-Based Open-Set Panoptic Segmentation Network. In Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 3 +[31] Sanghun Jung, Jungsoo Lee, Daehoon Gwak, Sungha Choi, and Jaegul Choo. Standardized Max Logits: A Simple yet Effective Approach for Identifying Unexpected Road Obstacles in Urban-Scene Segmentation. In International Conference on Computer Vision (ICCV), 2021. 1 +[32] Alexander Kirillov, Kaiming He, Ross Girshick, Carsten Rother, and Piotr Dólar. Panoptic Segmentation. In Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 7 +[33] Michael Kösel, Marcel Schreiber, Michael Ulrich, Claudius Gläser, and Klaus Dietmayer. Revisiting Out-of-Distribution Detection in LiDAR-based 3D Object Detection. In Intelligent Vehicles Symposium (IV), 2024. 3 +[34] Balaji Lakshminarayanan, Alexander Pritzel, and Charles Blundell. Simple and Scalable Predictive Uncertainty Estimation using Deep Ensembles. In Neural Information Processing Systems (NeurIPS), 2017. 2, 3, 6, 7 +[35] Seungjae Lee, Hyungtae Lim, and Hyun Myung. Patchwork++: Fast and robust ground segmentation solving partial under-segmentation using 3D point cloud. In Proc. IEEE/RSJ Int. Conf. Intell. Robots Syst., 2022. 3 +[36] Kaican Li, Kai Chen, Haoyu Wang, Lanqing Hong, Chaoqiang Ye, Jianhua Han, Yukuai Chen, Wei Zhang, Chunjing Xu, Dit-Yan Yeung, et al. Coda: A real-world road corner case dataset for object detection in autonomous driving. In European Conference on Computer Vision (ECCV), 2022. 2, 3 +[37] Chen Liang, Wenguan Wang, Jiaxu Miao, and Yi Yang. GMMSeg: Gaussian Mixture based Generative Semantic Segmentation Models. In Neural Information Processing Systems (NeurIPS), 2022. 1, 3 +[38] Kira Maag, Robin Chan, Svenja Uhlemeyer, Kamil Kowol, and Hanno Gottschalk. Two Video Data Sets for Tracking + +and Retrieval of Out of Distribution Objects. In Asian Conference on Computer Vision (ACCV), 2022. 2, 3 +[39] Jiageng Mao, Minzhe Niu, Chenhan Jiang, Hanxue Liang, Jingheng Chen, Xiaodan Liang, Yamin Li, Chaoqiang Ye, Wei Zhang, Zhenguo Li, Jie Yu, Hang Xu, and Chunjing Xu. One Million Scenes for Autonomous Driving: ONCE Dataset. In Neural Information Processing Systems (NeurIPS), 2021. 3 +[40] Rodrigo Marcuzzi, Lucas Nunes, Louis Wiesmann, Jens Behley, and Cyril Stachniss. Mask-Based Panoptic LiDAR Segmentation for Autonomous Driving. In IEEE Robotics And Automation Letters (RAL), 2023. 7 +[41] Nazir Nayal, Misra Yavuz, João F. Henriques, and Fatma Güney. RbA: Segmenting Unknown Regions Rejected by All. In International Conference on Computer Vision (ICCV), 2023. 1, 2, 3, 6, 7 +[42] Alexey Nekrasov, Rui Zhou, Miriam Ackermann, Alexander Hermans, and Matthias Rottmann Bastian Leibe. OoDIS: Anomaly Instance Segmentation Benchmark. arXiv preprint arXiv:2406.11835, 2024. 2 +[43] Gerhard Neuhold, Tobias Ollmann, Samuel Rota Bulo, and Peter Kontschieder. The mapillary vistas dataset for semantic understanding of street scenes. In International Conference on Computer Vision (ICCV), 2017. 3 +[44] Lucas Nunes, Xieyuanli Chen, Rodrigo Marcuzzi, Aljosa Osep, Laura Leal-Taixe, Cyril Stachniss, and Jens Behley. Unsupervised Class-Agnostic Instance Segmentation of 3D LiDAR Data for Autonomous Vehicles. IEEE Robotics And Automation Letters (RAL), 2022. 3 +[45] Peter Pinggera, Sebastian Ramos, Stefan Gehrig, Uwe Franke, Carsten Rother, and Rudolf Mester. Lost and Found: Detecting Small Road Hazards for Self-Driving Vehicles. In International Conference on Intelligent Robots and Systems (IROS), 2016. 2, 3, 4 +[46] Aldi Piroli, Vinzenz Dallabetta, Johannes Kopp, Marc Walesa, Daniel Meissner, and Klaus Dietmayer. LS-VOS: Identifying Outliers in 3D Object Detections Using Latent Space Virtual Outlier Synthesis. In IEEE Intelligent Transportation Systems Conference (ITSC), 2023. 2, 3 +[47] Shyam Nandan Rai, Fabio Cermelli, Dario Fontanel, Carlo Masone, and Barbara Caputo. Unmasking Anomalies in Road-Scene Segmentation. In International Conference on Computer Vision (ICCV), 2023. 3 +[48] Jenny Seidenschwarz, Aljosa O'Sep, Francesco Ferroni, Simon Lucey, and Laura Leal-Taixe. SeMoLi: What Moves Together Belongs Together. In Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 3 +[49] Aasheesh Singh, Aditya Kamireddypalli, Vineet Gandhi, and K. Madhava Krishna. LiDAR guided Small obstacle Segmentation. In International Conference on Intelligent Robots and Systems (IROS), 2020. 2, 3 +[50] Nitish Srivastava, Geoffrey Hinton, Alex Krizhevsky, Ilya Sutskever, and Ruslan Salakhutdinov. Dropout: a simple way to prevent neural networks from overfitting. In Neural Information Processing Systems (NeurIPS), 2014. 2, 6, 7 +[51] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, + +Yuning Chai, Benjamin Caine, et al. Scalability in perception for autonomous driving: Waymo open dataset. In Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2 +[52] DashcamCleaner Development Team. Dashcamcleaner: Censor identifiable information in videos from dashcam recordings. https://github.com/taehse/DashcamCleaner, 2024.4 +[53] Tefft, Brian C. . The Prevalence of Motor Vehicle Crashes Involving Road Debris, United States, 2011-2014. https://aaafoundation.org/wp-content/uploads/2017/12/RoadDebris_REPORT_2015.pdf, 2015. [Online]. 1 +[54] Yu Tian, Yuyuan Liu, Guansong Pang, Fengbei Liu, Yuanhong Chen, and Gustavo Carneiro. Pixel-wise Energy-biased Abstention Learning for Anomaly Segmentation on Complex Urban Driving Scenes. In European Conference on Computer Vision (ECCV), 2022. 1, 3 +[55] Tzu-Yun Tseng, Alexey Nekrasov, Malcolm Burdorf, Bastian Leibe, Julie Stephany Berrio Perez, Mao Shan, and Stewart Worrall. Panoptic-CUDAL Technical Report: Rural Australia Point Cloud Dataset in Rainy Conditions. arXiv preprint arXiv:2503.16378, 2025. 3, 4, 6, 8 +[56] S. Verma, J. S. Berrio, S. Worrall, and E. Nebot. Automatic extrinsic calibration between a camera and a 3d lidar using 3d point and plane correspondences. In IEEE Intelligent Transportation Systems Conference (ITSC), 2019. 3 +[57] Ignacio Vizzo, Tiziano Guadagnino, Benedikt Mersch, Louis Wiesmann, Jens Behley, and Cyril Stachniss. KISS-ICP: In Defense of Point-to-Point ICP – Simple, Accurate, and Robust Registration If Done the Right Way. In IEEE Robotics And Automation Letters (RAL), 2023. 4 +[58] Kelvin Wong, Shenlong Wang, Mengye Ren, Ming Liang, and Raquel Urtasun. Identifying Unknown Instances for Autonomous Driving. In Conference on Robot Learning (CoRL), 2019. 2, 3, 7, 8 +[59] Yajun Xu, Huan Hu, Xiaoya Zhu, Yibing Nan, Kai Wang, ZhaoXiang Liu, and Shiguo Lian. Raod: A benchmark for road abandoned object detection from video surveillance. IEEE Access, 2024. 2 +[60] Kadir Yilmaz, Jonas Schult, Alexey Nekrasov, and Bastian Leibe. Mask4Former: Mask Transformer for 4D Panoptic Segmentation. In International Conference on Robotics and Automation (ICRA), 2024. 3, 6, 7 +[61] Fisher Yu, Haofeng Chen, Xin Wang, Wenqi Xian, Yingying Chen, Fangchen Liu, Vashisht Madhavan, and Trevor Darrell. Bdd100k: A diverse driving dataset for heterogeneous multitask learning. In Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 3 \ No newline at end of file diff --git a/2025/Spotting the Unexpected (STU)_ A 3D LiDAR Dataset for Anomaly Segmentation in Autonomous Driving/images.zip b/2025/Spotting the Unexpected (STU)_ A 3D LiDAR Dataset for Anomaly Segmentation in Autonomous Driving/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..aa2908cf219d6cb6f3befb3ab0d8f75ac6233655 --- /dev/null +++ b/2025/Spotting the Unexpected (STU)_ A 3D LiDAR Dataset for Anomaly Segmentation in Autonomous Driving/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:aa932d7e91a9ec1d266bc760cbd21b0b7c03562ecdc0b1b8b64af674a6214e5b +size 655540 diff --git a/2025/Spotting the Unexpected (STU)_ A 3D LiDAR Dataset for Anomaly Segmentation in Autonomous Driving/layout.json b/2025/Spotting the Unexpected (STU)_ A 3D LiDAR Dataset for Anomaly Segmentation in Autonomous Driving/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..9e6df4a47b930d834e1942d02031008415e9685f --- /dev/null +++ b/2025/Spotting the Unexpected (STU)_ A 3D LiDAR Dataset for Anomaly Segmentation in Autonomous Driving/layout.json @@ -0,0 +1,7981 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 56, + 103, + 555, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 103, + 555, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 103, + 555, + 140 + ], + "type": "text", + "content": "Spotting the Unexpected (STU): A 3D LiDAR Dataset for Anomaly Segmentation in Autonomous Driving" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 95, + 160, + 185, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 95, + 160, + 185, + 175 + ], + "spans": [ + { + "bbox": [ + 95, + 160, + 185, + 175 + ], + "type": "text", + "content": "Alexey Nekrasov1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 212, + 160, + 304, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 212, + 160, + 304, + 175 + ], + "spans": [ + { + "bbox": [ + 212, + 160, + 304, + 175 + ], + "type": "text", + "content": "Malcolm Burdorf1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 331, + 161, + 414, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 161, + 414, + 175 + ], + "spans": [ + { + "bbox": [ + 331, + 161, + 414, + 175 + ], + "type": "text", + "content": "Stewart Worrall2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 441, + 161, + 514, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 441, + 161, + 514, + 175 + ], + "spans": [ + { + "bbox": [ + 441, + 161, + 514, + 175 + ], + "type": "text", + "content": "Bastian Leibe1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 234, + 175, + 375, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 234, + 175, + 375, + 190 + ], + "spans": [ + { + "bbox": [ + 234, + 175, + 375, + 190 + ], + "type": "text", + "content": "Julie Stephany Berrio Perez" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 169, + 189, + 444, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 169, + 189, + 444, + 204 + ], + "spans": [ + { + "bbox": [ + 169, + 189, + 444, + 204 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 169, + 189, + 444, + 204 + ], + "type": "text", + "content": " RWTH Aachen University, " + }, + { + "bbox": [ + 169, + 189, + 444, + 204 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 169, + 189, + 444, + 204 + ], + "type": "text", + "content": " The University of Sydney" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 155, + 216, + 454, + 229 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 155, + 216, + 454, + 229 + ], + "spans": [ + { + "bbox": [ + 155, + 216, + 454, + 229 + ], + "type": "text", + "content": "https://vision.rwth-aachen.de/stu-dataset" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 56, + 255, + 555, + 418 + ], + "blocks": [ + { + "bbox": [ + 56, + 255, + 555, + 418 + ], + "lines": [ + { + "bbox": [ + 56, + 255, + 555, + 418 + ], + "spans": [ + { + "bbox": [ + 56, + 255, + 555, + 418 + ], + "type": "image", + "image_path": "2c7d34edda03a4ca0ede6baff2084b80cb95434b140a7e10f268087ae07cafb1.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 426, + 555, + 449 + ], + "lines": [ + { + "bbox": [ + 55, + 426, + 555, + 449 + ], + "spans": [ + { + "bbox": [ + 55, + 426, + 555, + 449 + ], + "type": "text", + "content": "Figure 1. We present Spotting the Unexpected (STU) a novel anomaly segmentation dataset for autonomous driving. The dataset contains semantic and instance labels for out-of-distribution (OOD) objects, and includes surround-view setup with synchronized cameras." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 151, + 458, + 200, + 470 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 458, + 200, + 470 + ], + "spans": [ + { + "bbox": [ + 151, + 458, + 200, + 470 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 54, + 484, + 297, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 484, + 297, + 700 + ], + "spans": [ + { + "bbox": [ + 54, + 484, + 297, + 700 + ], + "type": "text", + "content": "To operate safely, autonomous vehicles (AVs) need to detect and handle unexpected objects or anomalies on the road. While significant research exists for anomaly detection and segmentation in 2D, research progress in 3D is underexplored. Existing datasets lack high-quality multimodal data that are typically found in AVs. This paper presents a novel dataset for anomaly segmentation in driving scenarios. To the best of our knowledge, it is the first publicly available dataset focused on road anomaly segmentation with dense 3D semantic labeling, incorporating both LiDAR and camera data, as well as sequential information to enable anomaly detection across various ranges. This capability is critical for the safe navigation of autonomous vehicles. We adapted and evaluated several baseline models for 3D segmentation, highlighting the challenges of 3D anomaly detection in driving environments. Our dataset and evaluation code will be openly available, facilitating the testing and performance comparison of different approaches." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 315, + 458, + 394, + 470 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 458, + 394, + 470 + ], + "spans": [ + { + "bbox": [ + 315, + 458, + 394, + 470 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 482, + 555, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 482, + 555, + 662 + ], + "spans": [ + { + "bbox": [ + 313, + 482, + 555, + 662 + ], + "type": "text", + "content": "Autonomous vehicles have become an increasingly important part of modern transportation systems. Despite significant advances, these vehicles continue to face challenges in accurately detecting unexpected objects on the road. One particular challenge is road debris, which refers to objects that are not typically part of the roadway environment and pose significant safety risks. According to the AAA Foundation for Traffic Safety [53], road debris was a contributing factor in more than 50,000 crashes, 10,000 injuries, and 125 fatalities annually between 2011 and 2014. These objects can vary widely in size and appearance, making them particularly challenging to detect. Addressing this issue is critical for ensuring the safe and reliable operation of autonomous vehicles, as it directly impacts their ability to handle anomalies in real-world driving conditions." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "content": "State-of-the-art anomaly segmentation and anomaly detection methods have received substantial attention in recent years [17, 37, 41], especially in the context of autonomous driving [11, 16, 18, 21, 31, 54]. One of the first" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "11875" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "content": "benchmarks for evaluating anomaly segmentation is Lost and Found [45], whose data serves as the basis for many complex benchmarks [4, 10]. These benchmarks provide a wide range of possible road anomalies and obstacles, and recent research has focused on exploring the temporal [38] and multimodal [6, 36] nature of anomaly segmentation." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 147, + 295, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 147, + 295, + 361 + ], + "spans": [ + { + "bbox": [ + 56, + 147, + 295, + 361 + ], + "type": "text", + "content": "While current autonomous driving datasets [1, 7, 51] use a setup with surround view cameras, radar, and high-resolution LiDAR sensors, existing anomaly segmentation and detection benchmarks focus primarily on image data [5] and mostly use Cityscapes [15] as the in-distribution dataset. Image-first benchmarks often lack the comprehensive sensor setup found in modern autonomous vehicles, and overlook the use of 3D modality for anomaly segmentation. Recent research that has begun to investigate the use of LiDAR [49] sensor for anomaly segmentation and detection, lacks high-resolution sensors found in modern autonomous vehicles. Several approaches have sought to find anomalies in validation sets of existing datasets [36, 46] or have created datasets using simulation [6] due to the difficulty of collecting and annotating such data in the real world. Despite these efforts, extending existing data or simulating anomalies only partially addresses the lack of accessible and high-quality data, shown in Table 1." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 364, + 295, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 364, + 295, + 567 + ], + "spans": [ + { + "bbox": [ + 55, + 364, + 295, + 567 + ], + "type": "text", + "content": "To address this issue, we analyzed the labels of two widely used in-distribution (ID) datasets [1, 7] to identify an appropriate set of anomalies that would not overlap with objects in the ID training or validation datasets. This informed our creation of a real-world dataset containing challenging anomaly objects placed on roads, with annotated sequences in 3D LiDAR point clouds. Our dataset consists of multiple high-resolution LiDAR sequences acquired in challenging environments, with point-level labels for individual anomaly objects (shown in Figure 1). We acquired data in both rural and urban environments, with objects of varying degrees of difficulty, and include a number of real-world anomalies found on public roads. In addition, to evaluate in-distribution performance, we annotated additional sequences without anomalies with inlier labels to train and evaluate models to ensure a small domain gap between our and SemanticKITTI [1] datasets." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "text", + "content": "To establish simple baselines for 3D LiDAR segmentation, we adapted the methods used in 2D anomaly segmentation [4, 24, 34, 41, 50] and applied them to the 3D domain, investigating anomaly segmentation methods for 3D LiDAR point clouds. Our models, trained jointly on the SemanticKITTI dataset and our own dataset, show strong segmentation performance for in-distribution data and novel additional sequences. However, we observed that the baseline methods struggled with out-of-distribution objects, and models tend to predict anomaly objects as inliers with a high degree of confidence. We believe that our dataset will benefit the community by allowing the evaluation of anomaly" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "content": "segmentation methods in the LiDAR domain and further exploration of multimodal anomaly segmentation methods." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 98, + 555, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 98, + 555, + 205 + ], + "spans": [ + { + "bbox": [ + 313, + 98, + 555, + 205 + ], + "type": "text", + "content": "Our contributions provide a basis for the development of more advanced anomaly segmentation methods in the 3D LiDAR domain. Specifically, we propose a new anomaly segmentation dataset, and perform an analysis of the recorded and annotated data. The purpose of the paper is to provide practitioners with the necessary tools to evaluate their anomaly segmentation models. We also provide a set of baselines that can serve as a reference for future challenges and research directions in this area." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 218, + 399, + 230 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 218, + 399, + 230 + ], + "spans": [ + { + "bbox": [ + 314, + 218, + 399, + 230 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 239, + 422, + 251 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 239, + 422, + 251 + ], + "spans": [ + { + "bbox": [ + 314, + 239, + 422, + 251 + ], + "type": "text", + "content": "2.1. Anomaly Datasets" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 257, + 553, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 257, + 553, + 437 + ], + "spans": [ + { + "bbox": [ + 313, + 257, + 553, + 437 + ], + "type": "text", + "content": "Several multimodal datasets for autonomous driving are available, including Waymo [51], nuScenes [7], SemanticKITTI [1], among others. These datasets were collected in urban environments under typical driving conditions. Although these datasets are relevant for evaluating various tasks related to autonomous vehicle operation, they lack data for uncommon scenarios. One such scenario involves the presence of foreign objects on the road. Although this situation is rare during typical driving, autonomous vehicles must be able to recognize and respond to it safely. Here, assigning a unique label to each distinct object is impractical, since datasets have a limited set of labels, and objects on the road can vary widely. These objects are often grouped under a general category like \"other object\" and are excluded from the benchmarks due to their variability." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 437, + 554, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 554, + 617 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 554, + 617 + ], + "type": "text", + "content": "In recent years, several datasets have been introduced to assess the detection of foreign objects on the road [4, 10, 25, 36, 38, 42, 45, 59]. Some of these datasets have been developed using data augmentation techniques or simulations, but only a few are based on data collected with real-world equipment in real driving scenarios. Simulations and augmentations are great tools, but they lack realism in terms of the sensor model [4] and the sizes of the objects [25]. However, the aforementioned datasets are limited to image data only. Autonomous vehicles use LiDAR sensors because they offer greater range, robustness to lighting conditions, and more precise 3D information about the environment compared to monocular depth estimations or stereo cameras. Therefore, incorporating 3D information is essential for accurately assessing road anomalies." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 618, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 618, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 618, + 554, + 713 + ], + "type": "text", + "content": "Three datasets that integrate camera and LiDAR data are SOD [22], TOR4D [58] and CODA [36]. The CODA series is derived from validation sets of existing autonomous vehicle datasets. Still, it has the limitation that anomalous objects (e.g., traffic cones, barriers) may also appear in the training set, which is acceptable for the task of open-set segmentation, but contradicts the task of anomaly segmentation. Recently, a few methods for 3D anomaly detection" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "11876" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 91, + 555, + 297 + ], + "blocks": [ + { + "bbox": [ + 61, + 70, + 548, + 82 + ], + "lines": [ + { + "bbox": [ + 61, + 70, + 548, + 82 + ], + "spans": [ + { + "bbox": [ + 61, + 70, + 548, + 82 + ], + "type": "text", + "content": "Table 1. Comparison of our dataset with other community datasets [5]. SM = Semantic Mask, IM=Instance Mask, BB=Bounding Box" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 91, + 555, + 297 + ], + "lines": [ + { + "bbox": [ + 56, + 91, + 555, + 297 + ], + "spans": [ + { + "bbox": [ + 56, + 91, + 555, + 297 + ], + "type": "table", + "html": "
DatasetCameraLiDAR \nBeamsSource#ODDLabelSize \nTest/ValTemporalEnvironments \nTest SequencesRelated Dataset
Fishscapes [3, 4]
FS Lost and Found1 RGBXStaged1SM275/100Xn/aLost and Found [45]
FS Static1 RGBXAugmented1SM1000/30Xn/aCityScapes [15]
CAOS [25]
StreetHazards1 RGBXSimulated1SM1500✓2 sim towns - 150 seq
BDD-Anomaly1 RGBXValidation Set3SM810Xn/aBDD100K [61]
SegmentMelfYouCan [10]
RoadAnomaly211 RGBXWeb Sourced1SM100/10Xn/a
RoadObstacle211 RGBXStaged1SM382/30✓9 streets - 90 seq
CODA [36]
CODA-KITTI1 RGB✓-64Validation Set6BB309Xn/aKITTI [19]
CODA-nuScenes1 RGB✓-32Validation Set17BB134Xn/aNuScenes [7]
CODA-ONCE1 RGB✓-40Validation Set32BB1057Xn/aONCE [39]
CODA22-ONCE1 RGB✓-40Validation Set29BB717Xn/aONCE [39]
CODA22-SODA10M1 RGBXValidation Set29BB4167Xn/aSODA [23]
Wuppertal ODD [38]
SOS1 RGBX-distStaged13IM1129✓9 streets - 20 seq
CWL1 RGBX-distSimulated18IM1210✓26 streets - 26 seq
Lost and Found [45]1 StereoXStaged42SM1068✓5 streets - 5 seq
SOD [49]1 RGB✓-16Staged1SM460/530X2 streets - 5 seq
WD-Pascal [2]1 RGBXSimulated1SM70Xn/a
Vistas-NP [20]1 RGBXClass exclusion4SM11167Xn/aMapillary Vistas [43]
STU (Ours)8 RGB✓-128Staged1IM8022/1960✓6 streets - 51 seqSemKITTI [1], P-CUDAL [55]
", + "image_path": "fb8a0eba4b9b3402acfe102e76cc8d6e477ed077ac6d320aaeecfc3d1c6e8fba.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 316, + 294, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 316, + 294, + 460 + ], + "spans": [ + { + "bbox": [ + 54, + 316, + 294, + 460 + ], + "type": "text", + "content": "have been proposed [28, 33, 46]. However, these methods use validation sequences to evaluate anomaly detection capabilities. The SOD dataset offers depth information for small road obstacles using a low-resolution 16-beam LiDAR, and that introduces a large domain gap between the training and test data. TOR4D and Rare4D [58] are the only datasets that contain real-world LiDAR data to benchmark road anomalies; however, these datasets are proprietary and not accessible to the public. We release out-of-distribution validation and in-distribution training sequences with a permissive license and open the private test set to submissions from the public." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 468, + 217, + 481 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 468, + 217, + 481 + ], + "spans": [ + { + "bbox": [ + 55, + 468, + 217, + 481 + ], + "type": "text", + "content": "2.2. Open-Set Segmentation in 3D" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 486, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 486, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 56, + 486, + 296, + 714 + ], + "type": "text", + "content": "Open-set and unsupervised segmentation are closely related to the out-of-distribution detection task. Multiple methods were proposed for unsupervised segmentation [44, 48] and open-set prediction in 3D [9, 58]. Recent evaluations in 2D anomaly segmentation [18] have demonstrated that methods such as OSIS [58] and EOPSN [30] perform poorly on anomaly objects. This weakness stems from open-set methods' reliance on the presence of unknown instances inside the void regions during training, which may not generalize well to real-world scenarios. Similarly, MLUC [8] assesses open-set performance on the UDI [8] dataset, but this dataset is also not publicly accessible and makes the same assumptions as OSIS methods. In addition, open-set methods typically remove the ground plane [44] to obtain clusters of potential objects, but ground plane removal [35] methods are often sensitive to chosen hyperparameters and yield noisy results, missing many smaller objects. For the evaluation of anomaly segmentation, it is important to have a test dataset that is free of anomaly objects which might" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 317, + 555, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 317, + 555, + 376 + ], + "spans": [ + { + "bbox": [ + 313, + 317, + 555, + 376 + ], + "type": "text", + "content": "occur during training. In contrast to open-set methods, our work focuses on collecting and densely annotating real-world anomaly objects, ensuring that there is no intersection between in-distribution training sets and out-of-distribution test sets." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 386, + 555, + 399 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 386, + 555, + 399 + ], + "spans": [ + { + "bbox": [ + 313, + 386, + 555, + 399 + ], + "type": "text", + "content": "2.3. Anomaly Segmentation in Autonomous Driving" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 404, + 556, + 619 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 404, + 556, + 619 + ], + "spans": [ + { + "bbox": [ + 313, + 404, + 556, + 619 + ], + "type": "text", + "content": "To tackle the problem of anomaly segmentation, multiple works [4, 5, 10] have recently been proposed for the problem of 2D anomaly segmentation. Earlier works [4, 11, 21, 24, 34, 37, 54] were applied to per-pixel segmentation models such as DeepLab [12]; however, recently many works were focused on MaskFormer-style models [13, 14]. A line of work [16, 41, 47] proposed different methods to adapt Mask2Former, exploring the mechanism of mask prediction, where multiple masks might ignore certain regions of an image. These models provide more reliable anomaly scoring mechanisms not limited to per-pixel predictions and achieve higher performance on the anomaly segmentation task. For our benchmark, we decided to evaluate methods that also use MaskFormer-style models and investigate the application of the Mask4Former [60] model for 3D anomaly segmentation. We obtain all our scores using a MaskFormer-Style model, laying the foundation for future baselines." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 633, + 553, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 633, + 553, + 647 + ], + "spans": [ + { + "bbox": [ + 313, + 633, + 553, + 647 + ], + "type": "text", + "content": "3. The Spotting the Unexpected (STU) Dataset" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "text", + "content": "Our data collection platform consists of a rigid frame mounted on a vehicle and equipped with eight hardware-triggered cameras and a LiDAR sensor to ensure synchronized data collection. The LiDAR and the cameras were calibrated using the method described in [56], with the cal" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11877" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 73, + 298, + 137 + ], + "blocks": [ + { + "bbox": [ + 60, + 73, + 298, + 137 + ], + "lines": [ + { + "bbox": [ + 60, + 73, + 298, + 137 + ], + "spans": [ + { + "bbox": [ + 60, + 73, + 298, + 137 + ], + "type": "image", + "image_path": "badb56129a8a822729d71d87dec28a53824536a8b36d206976c7f4f4870ccffb.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 119, + 141, + 231, + 151 + ], + "lines": [ + { + "bbox": [ + 119, + 141, + 231, + 151 + ], + "spans": [ + { + "bbox": [ + 119, + 141, + 231, + 151 + ], + "type": "text", + "content": "(a) Naturalistic data collection." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 60, + 155, + 298, + 219 + ], + "blocks": [ + { + "bbox": [ + 60, + 155, + 298, + 219 + ], + "lines": [ + { + "bbox": [ + 60, + 155, + 298, + 219 + ], + "spans": [ + { + "bbox": [ + 60, + 155, + 298, + 219 + ], + "type": "image", + "image_path": "422643e6cc0df06d09236827c709acec0b752a218bcbf68049418c65e74b86fc.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 93, + 223, + 258, + 233 + ], + "lines": [ + { + "bbox": [ + 93, + 223, + 258, + 233 + ], + "spans": [ + { + "bbox": [ + 93, + 223, + 258, + 233 + ], + "type": "text", + "content": "(b) Data collection in controlled environment." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 243, + 295, + 266 + ], + "lines": [ + { + "bbox": [ + 55, + 243, + 295, + 266 + ], + "spans": [ + { + "bbox": [ + 55, + 243, + 295, + 266 + ], + "type": "text", + "content": "Figure 2. Data collection conducted in a naturalistic manner (a) and controlled environment (b) with objects on the road." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 288, + 295, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 288, + 295, + 335 + ], + "spans": [ + { + "bbox": [ + 55, + 288, + 295, + 335 + ], + "type": "text", + "content": "ization process repeated for each camera. This setup follows the configuration discussed in Panoptic-CUDAL [55]. We refer to the Supplementary Material for detailed information on the vehicle setup." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 344, + 150, + 356 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 344, + 150, + 356 + ], + "spans": [ + { + "bbox": [ + 55, + 344, + 150, + 356 + ], + "type": "text", + "content": "3.1. Data collection." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 361, + 295, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 361, + 295, + 445 + ], + "spans": [ + { + "bbox": [ + 55, + 361, + 295, + 445 + ], + "type": "text", + "content": "The data was collected using two sets of conditions: one in a naturalistic environment where objects were found on public roads, and one in a controlled environment. In the naturalistic scenario, several hours of driving data were recorded, with specific events marked when objects appeared on the road. Figure 2a shows a bucket on the road recorded while driving on a highway." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 446, + 295, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 446, + 295, + 613 + ], + "spans": [ + { + "bbox": [ + 55, + 446, + 295, + 613 + ], + "type": "text", + "content": "The controlled environment data collection protocol was rigorous for safety reasons. We employed four people: two spotters positioned at either end of a low-traffic and low-speed road, one responsible for placing objects on the road, and a driver. The spotters would signal when it was safe to place the objects, and once positioned, the driver would begin recording the scene. If one of the spotters reported an approaching vehicle, the data recording was stopped, and the object was immediately removed. This was done to avoid disrupting the normal flow of traffic. The vehicle was driven at a maximum speed of " + }, + { + "bbox": [ + 55, + 446, + 295, + 613 + ], + "type": "inline_equation", + "content": "50\\mathrm{km / h}" + }, + { + "bbox": [ + 55, + 446, + 295, + 613 + ], + "type": "text", + "content": " during data recording. Figure 2b shows a computer monitor placed on the road during sensor data logging. Some of the objects placed on the road are also shown in Figure 3." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 618, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 295, + 714 + ], + "type": "text", + "content": "Postprocessing. After data collection, post-processing is required to obtain vehicle poses and anonymize the images. Point cloud registration was performed using KISS ICP [57], a LiDAR odometry pipeline. KISS ICP includes point cloud motion compensation, subsampling, adaptive thresholding to determine correspondences, and LiDAR pose estimation. The calculated LiDAR pose was then exported in KITTI format as required for the labeling tool [1]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 553, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 168 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 168 + ], + "type": "text", + "content": "Ethical considerations for this dataset include anonymizing camera images. Identifiable data, such as faces and license plates, are processed using DashcamCleaner [52] and DeepPrivacy2 [29]. DashcamCleaner uses a license plate detector to locate and blur license plates, while DeepPrivacy2 detects facial features and generates new, unidentifiable faces to replace the original ones. Examples of anonymization can be found in the supplementary material." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 174, + 379, + 186 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 174, + 379, + 186 + ], + "spans": [ + { + "bbox": [ + 313, + 174, + 379, + 186 + ], + "type": "text", + "content": "3.2. Labeling" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 191, + 553, + 298 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 191, + 553, + 298 + ], + "spans": [ + { + "bbox": [ + 313, + 191, + 553, + 298 + ], + "type": "text", + "content": "We define anomalies as objects that may pose a danger to the autonomous vehicle and its passengers, especially those present on the driving surface and not present in the training dataset. In this work, we focus on road anomalies and do not consider unusual situations such as vehicle collisions or unexpected driver behavior, and we focus on the fine-grained anomaly segmentation. In particular, we follow previous work [4, 45] and focus on objects that are not present in the training set." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 299, + 554, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 299, + 554, + 407 + ], + "spans": [ + { + "bbox": [ + 313, + 299, + 554, + 407 + ], + "type": "text", + "content": "We labeled the data using the SemanticKITTI labeler [1]. To generate initial labels for the scenes, we used pseudolabeling techniques using a pre-trained SemanticKITTI model. We instruct our annotators to group all points belonging to inliers as a single semantic class, questionable points as unlabeled, and anomaly points as the outlier class. We instruct our labelers that it is very important to label inliers and outliers carefully while allowing for larger unlabeled regions where the annotator may be challenged." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 411, + 554, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 411, + 554, + 625 + ], + "spans": [ + { + "bbox": [ + 313, + 411, + 554, + 625 + ], + "type": "text", + "content": "Annotation process. The domain gap between urban environments and SemanticKITTI street scenes is minimal, especially in regions close to the vehicle. This similarity allows the acquisition of reliable initial estimates. The annotation process was performed by three annotators, each spending eight hours a week over several months. Each annotator spent approximately four hours on the initial labeling of each scene and an additional two hours on refining the final labels. To ensure accuracy and consistency, the annotators cross-validated each other's work. To further improve the quality of the labels, we examined the predictions of the baseline methods to see if any known objects were missed. In particular, current methods tend to be highly sensitive to objects that are present but ignored during training, such as those classified in the category \"other object\". We annotate these objects as unlabeled and ignore them in the evaluations, as they could alter the results of the anomaly segmentation methods." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 630, + 553, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 553, + 712 + ], + "type": "text", + "content": "Differences to SemanticKITTI. In contrast to SemanticKITTI, which provides labels up to a range of 50 meters, we labeled all points visible to the LiDAR sensor. This was facilitated by using a LiDAR with a higher vertical resolution. The maximum distance resulting from an anomaly in our dataset is 150 meters, allowing for comprehensive evaluations. However, since we train on the Se" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11878" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 70, + 141, + 121 + ], + "blocks": [ + { + "bbox": [ + 61, + 70, + 141, + 121 + ], + "lines": [ + { + "bbox": [ + 61, + 70, + 141, + 121 + ], + "spans": [ + { + "bbox": [ + 61, + 70, + 141, + 121 + ], + "type": "image", + "image_path": "db4ed5b85024c82a54f0d517cd1da0f42888b59e123237cbeaf82741f84c17d0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 143, + 71, + 223, + 120 + ], + "blocks": [ + { + "bbox": [ + 143, + 71, + 223, + 120 + ], + "lines": [ + { + "bbox": [ + 143, + 71, + 223, + 120 + ], + "spans": [ + { + "bbox": [ + 143, + 71, + 223, + 120 + ], + "type": "image", + "image_path": "eaa1891ae8b0867e8ef46c72efe3c1c98265bbdbe5f36f5acf85fa8b47f3e92e.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 225, + 72, + 304, + 120 + ], + "blocks": [ + { + "bbox": [ + 225, + 72, + 304, + 120 + ], + "lines": [ + { + "bbox": [ + 225, + 72, + 304, + 120 + ], + "spans": [ + { + "bbox": [ + 225, + 72, + 304, + 120 + ], + "type": "image", + "image_path": "992a78b49d4cb3ac015ecbb803e907eca49e7a91ef35285667a07c5b649b1ead.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 306, + 72, + 386, + 120 + ], + "blocks": [ + { + "bbox": [ + 306, + 72, + 386, + 120 + ], + "lines": [ + { + "bbox": [ + 306, + 72, + 386, + 120 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 386, + 120 + ], + "type": "image", + "image_path": "4296654edfbc219d1703f6fe849d0d4dba4fae7f18e3d2edb82240d739641bd6.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 388, + 72, + 467, + 120 + ], + "blocks": [ + { + "bbox": [ + 388, + 72, + 467, + 120 + ], + "lines": [ + { + "bbox": [ + 388, + 72, + 467, + 120 + ], + "spans": [ + { + "bbox": [ + 388, + 72, + 467, + 120 + ], + "type": "image", + "image_path": "6e3d77629b6f9024d72fbffc971969fb91123ddca990a6608b8849691aca530b.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 469, + 72, + 549, + 120 + ], + "blocks": [ + { + "bbox": [ + 469, + 72, + 549, + 120 + ], + "lines": [ + { + "bbox": [ + 469, + 72, + 549, + 120 + ], + "spans": [ + { + "bbox": [ + 469, + 72, + 549, + 120 + ], + "type": "image", + "image_path": "d1eacadf2ddd0b34cb990a3ed4079bcbe30d70cf938dc0262a75662d5afa5fc2.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 61, + 121, + 141, + 171 + ], + "blocks": [ + { + "bbox": [ + 61, + 121, + 141, + 171 + ], + "lines": [ + { + "bbox": [ + 61, + 121, + 141, + 171 + ], + "spans": [ + { + "bbox": [ + 61, + 121, + 141, + 171 + ], + "type": "image", + "image_path": "ae700115971280b9c81f5f5d07c28515d56012ec5de26a2555fadc1537a01bc7.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 143, + 121, + 222, + 171 + ], + "blocks": [ + { + "bbox": [ + 143, + 121, + 222, + 171 + ], + "lines": [ + { + "bbox": [ + 143, + 121, + 222, + 171 + ], + "spans": [ + { + "bbox": [ + 143, + 121, + 222, + 171 + ], + "type": "image", + "image_path": "2fce046319925b5aa325f021b22267fe460061bba12dba07f87b256bafadd2c8.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 225, + 121, + 304, + 171 + ], + "blocks": [ + { + "bbox": [ + 225, + 121, + 304, + 171 + ], + "lines": [ + { + "bbox": [ + 225, + 121, + 304, + 171 + ], + "spans": [ + { + "bbox": [ + 225, + 121, + 304, + 171 + ], + "type": "image", + "image_path": "464c32cce11ea34ee67ded114a1d22365275f8af1560b318f6f0b11e21ccd728.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 306, + 121, + 386, + 171 + ], + "blocks": [ + { + "bbox": [ + 306, + 121, + 386, + 171 + ], + "lines": [ + { + "bbox": [ + 306, + 121, + 386, + 171 + ], + "spans": [ + { + "bbox": [ + 306, + 121, + 386, + 171 + ], + "type": "image", + "image_path": "92700289051bba58257434daa2f1401f651172f2c896499b74b36ced527e78c3.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 388, + 121, + 467, + 171 + ], + "blocks": [ + { + "bbox": [ + 388, + 121, + 467, + 171 + ], + "lines": [ + { + "bbox": [ + 388, + 121, + 467, + 171 + ], + "spans": [ + { + "bbox": [ + 388, + 121, + 467, + 171 + ], + "type": "image", + "image_path": "1840258f7a63031bd57972978bb92ea801de1b829b36d53f01b86e9327fd6ce2.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 469, + 121, + 549, + 171 + ], + "blocks": [ + { + "bbox": [ + 469, + 121, + 549, + 171 + ], + "lines": [ + { + "bbox": [ + 469, + 121, + 549, + 171 + ], + "spans": [ + { + "bbox": [ + 469, + 121, + 549, + 171 + ], + "type": "image", + "image_path": "71276fa0f6e972dd75663d194e7e4e195a916e2b68bc3f910b03293d7d99258c.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 61, + 171, + 141, + 221 + ], + "blocks": [ + { + "bbox": [ + 61, + 171, + 141, + 221 + ], + "lines": [ + { + "bbox": [ + 61, + 171, + 141, + 221 + ], + "spans": [ + { + "bbox": [ + 61, + 171, + 141, + 221 + ], + "type": "image", + "image_path": "6d7b36eac6babb141c0896375df5963cbf5a5c666244ec3b258085688f3268c4.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 143, + 171, + 222, + 221 + ], + "blocks": [ + { + "bbox": [ + 143, + 171, + 222, + 221 + ], + "lines": [ + { + "bbox": [ + 143, + 171, + 222, + 221 + ], + "spans": [ + { + "bbox": [ + 143, + 171, + 222, + 221 + ], + "type": "image", + "image_path": "7cc67798adcd0837f4f029bed306742399587163a34c441f7649dc300a752f96.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 225, + 171, + 304, + 221 + ], + "blocks": [ + { + "bbox": [ + 225, + 171, + 304, + 221 + ], + "lines": [ + { + "bbox": [ + 225, + 171, + 304, + 221 + ], + "spans": [ + { + "bbox": [ + 225, + 171, + 304, + 221 + ], + "type": "image", + "image_path": "13870c7795dfe6c37c9f5e194aa4fdeeab6682eae3fd774c5b298d753a3ab2a6.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 306, + 171, + 386, + 221 + ], + "blocks": [ + { + "bbox": [ + 306, + 171, + 386, + 221 + ], + "lines": [ + { + "bbox": [ + 306, + 171, + 386, + 221 + ], + "spans": [ + { + "bbox": [ + 306, + 171, + 386, + 221 + ], + "type": "image", + "image_path": "be03b39e639b38d8a17e342688b188f7b7a089628767adcdd8260b77cae4ea53.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 388, + 171, + 467, + 221 + ], + "blocks": [ + { + "bbox": [ + 388, + 171, + 467, + 221 + ], + "lines": [ + { + "bbox": [ + 388, + 171, + 467, + 221 + ], + "spans": [ + { + "bbox": [ + 388, + 171, + 467, + 221 + ], + "type": "image", + "image_path": "4485c45f4e2e4eb0b5815d215d3eb2d8f47a9a397616274276b1a33b27e1e067.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 469, + 171, + 549, + 221 + ], + "blocks": [ + { + "bbox": [ + 469, + 171, + 549, + 221 + ], + "lines": [ + { + "bbox": [ + 469, + 171, + 549, + 221 + ], + "spans": [ + { + "bbox": [ + 469, + 171, + 549, + 221 + ], + "type": "image", + "image_path": "e274c448a26dc21a5f1029348cb97c289d32dc57962fc8497bae9b0d84a20573.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 61, + 222, + 141, + 271 + ], + "blocks": [ + { + "bbox": [ + 61, + 222, + 141, + 271 + ], + "lines": [ + { + "bbox": [ + 61, + 222, + 141, + 271 + ], + "spans": [ + { + "bbox": [ + 61, + 222, + 141, + 271 + ], + "type": "image", + "image_path": "5b55c2119689c1aac9a5f19d06936a4e4bb5bb0931958a673d9620b250db1fbc.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 281, + 555, + 314 + ], + "lines": [ + { + "bbox": [ + 55, + 281, + 555, + 314 + ], + "spans": [ + { + "bbox": [ + 55, + 281, + 555, + 314 + ], + "type": "text", + "content": "Figure 3. Different anomalies in the STU dataset. Different objects on the road used for staged data collection. We pick objects such that have no intersection with the inlier dataset and place them on roads in different locations and illumination conditions. Objects might touch each other, be very small, as large as a chair or a surf board, and could cause an accident if a car would drive over them." + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 143, + 222, + 222, + 271 + ], + "blocks": [ + { + "bbox": [ + 143, + 222, + 222, + 271 + ], + "lines": [ + { + "bbox": [ + 143, + 222, + 222, + 271 + ], + "spans": [ + { + "bbox": [ + 143, + 222, + 222, + 271 + ], + "type": "image", + "image_path": "b0d68c0089f1d047d0d14e3fe243a243aa92a8e3038bfe409220c84baa998cb5.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 225, + 222, + 304, + 271 + ], + "blocks": [ + { + "bbox": [ + 225, + 222, + 304, + 271 + ], + "lines": [ + { + "bbox": [ + 225, + 222, + 304, + 271 + ], + "spans": [ + { + "bbox": [ + 225, + 222, + 304, + 271 + ], + "type": "image", + "image_path": "81ae7030ddd3e101f381e4052f854d29d1e31a88a5f3401eee1f2c5bfdf6906d.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 306, + 222, + 386, + 271 + ], + "blocks": [ + { + "bbox": [ + 306, + 222, + 386, + 271 + ], + "lines": [ + { + "bbox": [ + 306, + 222, + 386, + 271 + ], + "spans": [ + { + "bbox": [ + 306, + 222, + 386, + 271 + ], + "type": "image", + "image_path": "6cf8158d2c34d7bac0194cc961fd67cbfae9a8ff510ad38ae0829d9ee4009de6.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 388, + 222, + 467, + 271 + ], + "blocks": [ + { + "bbox": [ + 388, + 222, + 467, + 271 + ], + "lines": [ + { + "bbox": [ + 388, + 222, + 467, + 271 + ], + "spans": [ + { + "bbox": [ + 388, + 222, + 467, + 271 + ], + "type": "image", + "image_path": "f95f6e533d073052d9d51a0406fbe9260dcfc36481ab0f5a03ebaa2d727094e5.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 469, + 222, + 549, + 271 + ], + "blocks": [ + { + "bbox": [ + 469, + 222, + 549, + 271 + ], + "lines": [ + { + "bbox": [ + 469, + 222, + 549, + 271 + ], + "spans": [ + { + "bbox": [ + 469, + 222, + 549, + 271 + ], + "type": "image", + "image_path": "7dbff6d23082585447662b9eeae32681468170763bf26c7c14b2ed508a516d32.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "bbox": [ + 55, + 335, + 295, + 383 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 335, + 295, + 383 + ], + "spans": [ + { + "bbox": [ + 55, + 335, + 295, + 383 + ], + "type": "text", + "content": "manticKITTI dataset, which contains only labels up to 50 meters, we limit our evaluations to the 50 meter range. The sensor setup has 128 LiDAR beams, compared to 64 beams in SemanticKITTI." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 55, + 394, + 153, + 407 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 394, + 153, + 407 + ], + "spans": [ + { + "bbox": [ + 55, + 394, + 153, + 407 + ], + "type": "text", + "content": "3.3. Dataset analysis" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 55, + 412, + 295, + 530 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 412, + 295, + 530 + ], + "spans": [ + { + "bbox": [ + 55, + 412, + 295, + 530 + ], + "type": "text", + "content": "We recorded 70 sequences with anomalies, which we fully annotated for anomaly segmentation. Two of these labeled sequences were captured in a naturalistic manner, while the remaining 68 were staged to include various anomalies. We designated 19 sequences as a validation set and 51 as a closed test set for benchmarking purposes. In addition, we recorded and annotated two sequences with no anomalies for training and validation of the inlier performance \"STU-inlier\", to reduce the domain gap between our dataset and SemanticKITTI." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 55, + 533, + 295, + 699 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 533, + 295, + 699 + ], + "spans": [ + { + "bbox": [ + 55, + 533, + 295, + 699 + ], + "type": "text", + "content": "Our dataset defines two primary classes: inliers and outliers. Inliers encompasses classes that are used to train a model, and contains objects expected on the road, such as vehicles, pedestrians, and road infrastructure. Outliers, on the other hand, refer to out-of-distribution elements or anomalies that are atypical or unexpected in the driving scene, shown in Figure 3. One additional class, unlabeled is ignored in the evaluation and typically includes classes that the model has seen during training, but was not supervised in, such as parking meters, utility boxes, and lamps. We provide anomaly annotations at both the instance and semantic levels, enabling a comprehensive evaluation of models for semantic and instance segmentation in 3D environments." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "type": "text", + "content": "Each outlier object in our dataset has an instance la-" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 313, + 335, + 553, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 335, + 553, + 443 + ], + "spans": [ + { + "bbox": [ + 313, + 335, + 553, + 443 + ], + "type": "text", + "content": "bel unique for a sequence. Figure 4 presents histograms that display the distribution of key properties for these instances, including the average and maximum number of LiDAR points per instance, instance heights, and the number of instances per scan. This analysis provides insights into the structural characteristics and variability of data within point clouds. Specifically, Figure 4 indicates that most instances have very few LiDAR points on average, but a sequence may have multiple instances per sequence." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 313, + 444, + 554, + 623 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 444, + 554, + 623 + ], + "spans": [ + { + "bbox": [ + 313, + 444, + 554, + 623 + ], + "type": "text", + "content": "Furthermore, Figure 5 presents a scatter plot of detected anomalies within the vehicle frame, highlighting a higher density of anomalies along the longitudinal " + }, + { + "bbox": [ + 313, + 444, + 554, + 623 + ], + "type": "inline_equation", + "content": "(X)" + }, + { + "bbox": [ + 313, + 444, + 554, + 623 + ], + "type": "text", + "content": " axis both in front and behind the vehicle. The accompanying histograms depict the distribution of the anomalies along each axis, noting that the anomalies are identified at distances ranging up to 150 meters in front of the vehicle. Figure 6 explores the relationship between the distance to an anomaly and its size, showing that the proportion of LiDAR point cloud obstacles decreases as the distance increases. This trend is due to the radial nature of LiDAR, which has lower vertical resolution at greater ranges. For benchmarking purposes, we focus exclusively on instances with more than five LiDAR points, highlighted in blue, ensuring that the evaluation targets detectable anomalies." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 313, + 635, + 430, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 635, + 430, + 647 + ], + "spans": [ + { + "bbox": [ + 313, + 635, + 430, + 647 + ], + "type": "text", + "content": "3.4. Evaluated Baselines" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "text", + "content": "The formulation of the anomaly segmentation task restricts training to avoid direct supervision of OOD objects. Generally, OOD objects are not present in the training set, but could be present during inference [18]. Since there are no methods that could be applied directly for our setup, we" + } + ] + } + ], + "index": 33 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11879" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 66, + 72, + 182, + 156 + ], + "blocks": [ + { + "bbox": [ + 66, + 72, + 182, + 156 + ], + "lines": [ + { + "bbox": [ + 66, + 72, + 182, + 156 + ], + "spans": [ + { + "bbox": [ + 66, + 72, + 182, + 156 + ], + "type": "image", + "image_path": "776e849cd15d0db2f2a4ad196bc52ccec72beafd6d8398d0bd31966b65966389.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 64, + 163, + 184, + 173 + ], + "lines": [ + { + "bbox": [ + 64, + 163, + 184, + 173 + ], + "spans": [ + { + "bbox": [ + 64, + 163, + 184, + 173 + ], + "type": "text", + "content": "(a) Avg. num of points/sequence." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 189, + 73, + 301, + 156 + ], + "blocks": [ + { + "bbox": [ + 189, + 73, + 301, + 156 + ], + "lines": [ + { + "bbox": [ + 189, + 73, + 301, + 156 + ], + "spans": [ + { + "bbox": [ + 189, + 73, + 301, + 156 + ], + "type": "image", + "image_path": "ece6301e8fd8eeb1d66ad5dfc22bc445b171effe6e1d6be177a0cdc06318fa80.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 175, + 555, + 209 + ], + "lines": [ + { + "bbox": [ + 55, + 175, + 555, + 209 + ], + "spans": [ + { + "bbox": [ + 55, + 175, + 555, + 209 + ], + "type": "text", + "content": "Figure 4. Anomaly Instance Properties. A typical recorded anomaly has on average less than 50 points per sequence (a), with less than 300 (b) points at maximum, and a maximum height below one meter (c). We record up to nine individual anomaly instances in the same sequence (d)." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 309, + 73, + 423, + 156 + ], + "blocks": [ + { + "bbox": [ + 192, + 163, + 297, + 173 + ], + "lines": [ + { + "bbox": [ + 192, + 163, + 297, + 173 + ], + "spans": [ + { + "bbox": [ + 192, + 163, + 297, + 173 + ], + "type": "text", + "content": "(b) Max. points per instance." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 309, + 73, + 423, + 156 + ], + "lines": [ + { + "bbox": [ + 309, + 73, + 423, + 156 + ], + "spans": [ + { + "bbox": [ + 309, + 73, + 423, + 156 + ], + "type": "image", + "image_path": "d17ca69cf4e26cc946d2185c567fae78f376de4b48f7cedfce782756d93391d1.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 429, + 74, + 542, + 156 + ], + "blocks": [ + { + "bbox": [ + 306, + 162, + 424, + 173 + ], + "lines": [ + { + "bbox": [ + 306, + 162, + 424, + 173 + ], + "spans": [ + { + "bbox": [ + 306, + 162, + 424, + 173 + ], + "type": "text", + "content": "(c) Max. height of instances (m)." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 429, + 74, + 542, + 156 + ], + "lines": [ + { + "bbox": [ + 429, + 74, + 542, + 156 + ], + "spans": [ + { + "bbox": [ + 429, + 74, + 542, + 156 + ], + "type": "image", + "image_path": "1b5a0783346b2817ba682da757f107ca42013d480cfa98e72ba1a3d589e83b99.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 427, + 162, + 545, + 173 + ], + "lines": [ + { + "bbox": [ + 427, + 162, + 545, + 173 + ], + "spans": [ + { + "bbox": [ + 427, + 162, + 545, + 173 + ], + "type": "text", + "content": "(d) Numb. of instances/sequence." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 58, + 230, + 297, + 364 + ], + "blocks": [ + { + "bbox": [ + 58, + 230, + 297, + 364 + ], + "lines": [ + { + "bbox": [ + 58, + 230, + 297, + 364 + ], + "spans": [ + { + "bbox": [ + 58, + 230, + 297, + 364 + ], + "type": "image", + "image_path": "641ba3252959ed7b996229cd237db7c621162e36929ebcdef64faffb36fd6ae2.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 367, + 295, + 390 + ], + "lines": [ + { + "bbox": [ + 55, + 367, + 295, + 390 + ], + "spans": [ + { + "bbox": [ + 55, + 367, + 295, + 390 + ], + "type": "text", + "content": "Figure 5. Distribution of anomalies along the vehicle's reference frame. Most of the points appear around the vehicle." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 58, + 403, + 296, + 577 + ], + "blocks": [ + { + "bbox": [ + 58, + 403, + 296, + 577 + ], + "lines": [ + { + "bbox": [ + 58, + 403, + 296, + 577 + ], + "spans": [ + { + "bbox": [ + 58, + 403, + 296, + 577 + ], + "type": "image", + "image_path": "a0dd9871c54e0d67be1c93fb2eb0bf007782db7fcad52a7c47004efe7f9cc8c0.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 586, + 295, + 632 + ], + "lines": [ + { + "bbox": [ + 55, + 586, + 295, + 632 + ], + "spans": [ + { + "bbox": [ + 55, + 586, + 295, + 632 + ], + "type": "text", + "content": "Figure 6. We follow SemanticKitti's maximum distance threshold and evaluate instances that are within a 50 meter radius of the vehicle. In addition, we constrain evaluation of point and object level metrics to a minimum of 5 points." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "content": "adopt several baseline methods from the 2D anomaly segmentation to 3D LiDAR segmentation. First, we evaluate simple baselines, such as Max-Logit [24], Monte-Carlo Dropout [50], and Deep Ensembles [34]. Secondly, we train one of the models to predict ignored regions during train" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 229, + 553, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 229, + 553, + 277 + ], + "spans": [ + { + "bbox": [ + 313, + 229, + 553, + 277 + ], + "type": "text", + "content": "ing, which could be used to classify “void” points during testing [4]. We also adapt RbA [41], one of the state-of-the-art methods for image anomaly segmentation, to the 3D LiDAR segmentation method Mask4Former-3D [60]." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 282, + 555, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 282, + 555, + 413 + ], + "spans": [ + { + "bbox": [ + 313, + 282, + 555, + 413 + ], + "type": "text", + "content": "Deep Ensembles, MC Dropout, and MaxLogit. Deep Ensembles [34] and MC Dropout [50] are very common methods that practitioners use due to ease of implementation and application. For Deep Ensembles, we train three Mask4Former-3D [60] models. For MC dropout, we train one Mask4Former-3D model with enabled dropout in the transformer decoder for self- and cross-attention operations, and during inference we repeat the transformer decoder operation 50 times. As a simple baseline that uses only one model and a single forward pass, we use the MaxLogit [24] method." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 418, + 556, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 418, + 556, + 562 + ], + "spans": [ + { + "bbox": [ + 313, + 418, + 556, + 562 + ], + "type": "text", + "content": "RbA [41] is a method that uses the ability of MaskFormer-Style models [13, 14] to predict multiple overlapping segmentation masks. MaskFormer models have an option to disable queries and might not predict any masks if no objects are present. Points for which no masks were predicted will have the highest anomaly score. The Mask4Former model follows the same architectural paradigm and has a query deactivation mechanism that allows for no prediction in uncertain regions. This allows us to use the RbA method to generate anomaly scores within our setup. Unlike the original RbA model setup, we did not modify the transformer decoder." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 567, + 555, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 567, + 555, + 604 + ], + "spans": [ + { + "bbox": [ + 313, + 567, + 555, + 604 + ], + "type": "text", + "content": "Void Classifier We follow a common baseline [4] by treating the unlabeled regions as an additional semantic class during training." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 616, + 432, + 629 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 616, + 432, + 629 + ], + "spans": [ + { + "bbox": [ + 313, + 616, + 432, + 629 + ], + "type": "text", + "content": "4. Baseline Evaluation" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 641, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 555, + 713 + ], + "type": "text", + "content": "Training data. We use SemanticKITTI [1], Panoptic-CUDAL [55], and an additional recorded sequence from to jointly train our models. The Panoptic-CUDAL dataset, contains six sequences for training and one for validation, and was mostly recorded in rural locations, while our dataset consists mainly of urban scenes. To reduce the do" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "11880" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 59, + 91, + 555, + 182 + ], + "blocks": [ + { + "bbox": [ + 160, + 70, + 449, + 82 + ], + "lines": [ + { + "bbox": [ + 160, + 70, + 449, + 82 + ], + "spans": [ + { + "bbox": [ + 160, + 70, + 449, + 82 + ], + "type": "text", + "content": "Table 2. Anomaly Segmentation Performance of the selected baseline methods." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 59, + 91, + 555, + 182 + ], + "lines": [ + { + "bbox": [ + 59, + 91, + 555, + 182 + ], + "spans": [ + { + "bbox": [ + 59, + 91, + 555, + 182 + ], + "type": "table", + "html": "
MethodOoD DataPoint-Level OODObject-Level OOD
AUROC ↑FPR@95 ↓AP ↑RecallQSQRQUQPQ
Deep Ensemble [34]X86.7458.055.1716.7584.4910.4314.168.81
MC Dropout [50]X61.5182.370.112.2586.721.952.141.86
Max Logit [24]X84.5381.490.9526.1483.062.1321.711.77
Void Classifier [4]✓85.9978.603.9217.6484.408.1914.896.91
RbA [41]X66.38100.00.8124.0483.283.2320.022.69
", + "image_path": "09544e1b60605c6b38763310ee0719ff888ce3864055368db73fe8ce13e97ee3.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 58, + 243, + 296, + 305 + ], + "blocks": [ + { + "bbox": [ + 55, + 201, + 295, + 235 + ], + "lines": [ + { + "bbox": [ + 55, + 201, + 295, + 235 + ], + "spans": [ + { + "bbox": [ + 55, + 201, + 295, + 235 + ], + "type": "text", + "content": "Table 3. Closed-set performance of the methods on validation sets of SemanticKITTI [1] and STU-inlier dataset. * denotes a model trained only on SemanticKITTI dataset." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 243, + 296, + 305 + ], + "lines": [ + { + "bbox": [ + 58, + 243, + 296, + 305 + ], + "spans": [ + { + "bbox": [ + 58, + 243, + 296, + 305 + ], + "type": "table", + "html": "
MethodSemanticKITTISTU-inlier Dataset
PQSQRQPQSQRQ
Mask-PLS* [40]59.976.469.139.667.050.5
Mask-PLS [40]55.676.064.550.969.462.2
Mask4Former-3D* [60]61.981.371.642.870.653.2
Mask4Former-3D [60]60.781.070.452.773.864.7
", + "image_path": "202e72460f8578739473a19829883113d93f0592496578141c487bb4347a1014.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 329, + 295, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 329, + 295, + 558 + ], + "spans": [ + { + "bbox": [ + 54, + 329, + 295, + 558 + ], + "type": "text", + "content": "main gap, we have recorded and annotated two additional inlier scenes from an urban setting \"STU-inlier\", one for training and one for evaluation. These training sequences closely resemble the environments in which anomaly scenes were recorded. We train models with additional data recorded from the same vehicle setup as Panoptic-CUDAL and labeled the data with 19 SemanticKITTI labels. However, SemanticKITTI and our recorded dataset contain \"other-object\" that acts as a miscellaneous class for all objects that do not fit other-class definitions. A typical object is a bench, a garbage bin, or an advertisement post. To strictly define anomalies, we analyze all other-objects present in our training set, as well as the \"Movable_Object.Debris\" and \"Pushable.Pullable\" class objects in the NuScenes training set. We specifically design our dataset such that we do not have an intersection between anomalous objects and the aforementioned classes. Our Void Classifier training procedure consists of training on this unlabeled class that is present in the training data." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 563, + 295, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 563, + 295, + 684 + ], + "spans": [ + { + "bbox": [ + 54, + 563, + 295, + 684 + ], + "type": "text", + "content": "Models. We train MaskPLS [40] and Mask4Former [60] models jointly on both SemanticKitti and our training sequences. These models achieve state-of-the-art results on the SemanticKITTI [1] dataset for single-scan and multi-scan settings. However, while our dataset supports evaluation of temporal sequences, we performed our evaluation and training on single scans only. In our experiments, we refer to Mask4Former trained on single scans as Mask4Former-3D. All anomaly segmentation methods use Mask4Former-3D as the main model." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": "Metrics. To evaluate the performance of our methods, we use common metrics in the anomaly segmentation commu" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 202, + 555, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 202, + 555, + 251 + ], + "spans": [ + { + "bbox": [ + 313, + 202, + 555, + 251 + ], + "type": "text", + "content": "nity in 2D and panoptic segmentation in 3D LiDAR. Following SemanticKitti, we evaluated both inlier and outlier performance on points within 50 meters of the vehicle, and on objects with at least five points." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 254, + 556, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 254, + 556, + 327 + ], + "spans": [ + { + "bbox": [ + 313, + 254, + 556, + 327 + ], + "type": "text", + "content": "Point-Level Metrics. To evaluate the performance of the anomaly segmentation at the point level, we use the common metrics of Average Precision (AP), False-Positive Rate at " + }, + { + "bbox": [ + 313, + 254, + 556, + 327 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 313, + 254, + 556, + 327 + ], + "type": "text", + "content": " True-Positive Rate (FPR95), and Area Under Receiver Operating characteristic Curve (AUROC). Commonly used in the anomaly segmentation community [4]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 329, + 556, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 329, + 556, + 544 + ], + "spans": [ + { + "bbox": [ + 313, + 329, + 556, + 544 + ], + "type": "text", + "content": "Object-Level Metrics. To be consistent with the panoptic segmentation setup, we use panoptic metrics to evaluate the model performance on inlier classes. For the inlier data, we compute the Panoptic Quality (PQ) on the SemanticKitti classes. We report detailed class scores in the Supplementary. In contrast to anomaly segmentation, the open task formulation emphasizes the recall of all objects. Formally, the Unknown Quality (UQ) metric is used in such a setting [58]. However, in the case of anomaly segmentation, since anomalies can pose serious threats to the vehicle, it is important to predict as few false positives as possible, while also producing accurate segmentation labels for anomaly objects. Panoptic Quality uses an F-score and penalizes the prediction of multiple false positives. Thus, for anomaly segmentation baselines, we report both Panoptic Quality and Unknown Quality on a single class. For inlier performance, we report the Panoptic Quality score on the SemanticKITTI classes." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 545, + 555, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 555, + 582 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 555, + 582 + ], + "type": "text", + "content": "Panoptic Quality [32] is the geometric mean between Recognition Quality (RQ) and Segmentation Quality (SQ) for a class " + }, + { + "bbox": [ + 313, + 545, + 555, + 582 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 545, + 555, + 582 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 328, + 589, + 544, + 630 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 589, + 544, + 630 + ], + "spans": [ + { + "bbox": [ + 328, + 589, + 544, + 630 + ], + "type": "interline_equation", + "content": "\\mathrm {P Q} _ {c} = \\underbrace {\\frac {\\sum_ {(p , g) \\in T P _ {c}} \\operatorname {I o U} (p , g)}{| T P _ {c} |}} _ {\\text {S e g m e n t a t i o n Q u a l i t y (S Q)}} \\times \\underbrace {\\frac {| T P _ {c} |}{| T P _ {c} | + \\frac {1}{2} | F P _ {c} | + \\frac {1}{2} | F N _ {c} |}} _ {\\text {R e c o n g i t i o n Q u a l i t y (R Q)}}", + "image_path": "7864a8011715a5581e3e8a4ff13354b4a9bf975db783c585b29b362f40916633.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "inline_equation", + "content": "TP" + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "inline_equation", + "content": "FP" + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "inline_equation", + "content": "FN" + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "text", + "content": " are True Positives, False Positives, and False Negatives for objects with Intersection over Union (IoU) higher than " + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "text", + "content": ". The final PQ metric is calculated as an average over all classes " + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "text", + "content": ". Ignore points are removed from the scene prior to evaluation and erroneous predictions in the ignore region are not penalized." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11881" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "text", + "content": "To evaluate the recall of anomaly objects, we use the Unknown Quality (UQ) metric [58]:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 73, + 102, + 277, + 148 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 102, + 277, + 148 + ], + "spans": [ + { + "bbox": [ + 73, + 102, + 277, + 148 + ], + "type": "interline_equation", + "content": "\\mathrm {U Q} = \\underbrace {\\frac {\\sum_ {(p , g) \\in T P} \\operatorname {I o U} (p , g)}{| T P |}} _ {\\text {S e g m e n t a t i o n Q u a l i t y (S Q)}} \\times \\underbrace {\\frac {| T P |}{| T P | + | F N |}} _ {\\text {R e c a l l Q u a l i t y (R e c a l l Q)}}.", + "image_path": "c0b313aeaf002ad71d62b6aae76ad6ed241fb17911eadf52e0f1e674398efc56.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 155, + 296, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 155, + 296, + 202 + ], + "spans": [ + { + "bbox": [ + 55, + 155, + 296, + 202 + ], + "type": "text", + "content": "The metric does not penalize " + }, + { + "bbox": [ + 55, + 155, + 296, + 202 + ], + "type": "inline_equation", + "content": "FP" + }, + { + "bbox": [ + 55, + 155, + 296, + 202 + ], + "type": "text", + "content": " and allows for the evaluation of object recall. However, the metric counts objects as " + }, + { + "bbox": [ + 55, + 155, + 296, + 202 + ], + "type": "inline_equation", + "content": "TP" + }, + { + "bbox": [ + 55, + 155, + 296, + 202 + ], + "type": "text", + "content": " only if they have an IoU of at least " + }, + { + "bbox": [ + 55, + 155, + 296, + 202 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 55, + 155, + 296, + 202 + ], + "type": "text", + "content": " with ground truth." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 210, + 138, + 222 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 210, + 138, + 222 + ], + "spans": [ + { + "bbox": [ + 55, + 210, + 138, + 222 + ], + "type": "text", + "content": "4.1. Experiments" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 231, + 296, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 231, + 296, + 386 + ], + "spans": [ + { + "bbox": [ + 55, + 231, + 296, + 386 + ], + "type": "text", + "content": "Inlier Performance on SemanticKitti and STU-inlier dataset. We train models jointly with SemanticKITTI [1], Panoptic-CUDAL [55] and our additional sequence, and evaluate the SemanticKITTI and STU-inlier validation sets. The jointly trained model demonstrates improved performance on our STU-inlier dataset compared to models trained solely on the SemanticKITTI dataset, suggesting a small domain gap between our dataset and SemanticKITTI. Evaluation on the proposed STU-inlier sequence shows good segmentation performance for inlier classes that is critical for anomaly segmentation. The metrics for the performance of the baseline method in the validation sets are provided in Tab. 3." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 391, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 391, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 391, + 296, + 713 + ], + "type": "text", + "content": "Out-of-Distribution (OOD) Performance. We evaluate baseline methods for anomaly segmentation at both the point and object levels. Our findings indicate that directly transferring methods from 2D to 3D does not yield similar performance. Our point clouds typically contain around 100,000 points, with only a few dozen being anomalous, as shown in Fig. 6. Due to the significant class imbalance, point-level scores are generally low, highlighting the difficulty of the task. A high FPR@95, as noted in ODIN [27], suggests that models often misclassify road objects as inlier classes. Since LiDAR point clouds have lower resolution at a distance, predicting a small anomaly object consisting of a few points is a hard task. It is common for large anomalous objects to be predicted as pedestrians or other vehicles, as depicted in Fig. 7. Anomalous objects frequently occur in familiar contexts, which leads the model to confidently assign inlier labels to such anomalies in many cases. This behavior contrasts significantly with models trained on 2D data, where a simple max logit approach [26] outperforms Deep Ensembles [10]. However, the Deep Ensemble model stands out by producing fewer false positives at the point level, resulting in a lower FPR@95 and a higher PQ metric. Despite this reduction in false positives, the average precision (AP) remains low compared to 2D datasets. We attribute this to the high confidence with which individual models predict inlier objects, with disagreements between models contributing to the improved FPR@95. We refer" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 326, + 70, + 434, + 178 + ], + "blocks": [ + { + "bbox": [ + 326, + 70, + 434, + 178 + ], + "lines": [ + { + "bbox": [ + 326, + 70, + 434, + 178 + ], + "spans": [ + { + "bbox": [ + 326, + 70, + 434, + 178 + ], + "type": "image", + "image_path": "028a40dd22490729cfbd1b5011aeae64ba6e4b5f51f913f5b4c8c9fd131622c8.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 435, + 70, + 543, + 178 + ], + "blocks": [ + { + "bbox": [ + 435, + 70, + 543, + 178 + ], + "lines": [ + { + "bbox": [ + 435, + 70, + 543, + 178 + ], + "spans": [ + { + "bbox": [ + 435, + 70, + 543, + 178 + ], + "type": "image", + "image_path": "b035bc2638a0727b7506a25816e1bc56b5a8351b2f6d15dad5ec66fa3abbb146.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 326, + 178, + 434, + 285 + ], + "blocks": [ + { + "bbox": [ + 326, + 178, + 434, + 285 + ], + "lines": [ + { + "bbox": [ + 326, + 178, + 434, + 285 + ], + "spans": [ + { + "bbox": [ + 326, + 178, + 434, + 285 + ], + "type": "image", + "image_path": "15be6f85a2b4e4fac3a538396f6f3154fac2efb04aa1c2cf4b71cfdc44cb6be3.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 294, + 555, + 339 + ], + "lines": [ + { + "bbox": [ + 313, + 294, + 555, + 339 + ], + "spans": [ + { + "bbox": [ + 313, + 294, + 555, + 339 + ], + "type": "text", + "content": "Figure 7. Example of a failure case anomaly segmentation. For the chair on the road, labeled as anomaly in upper left, model predicts \"other-vehicle\" class, in lower left, with a high certainty, as indicated by MaxLogit scores on the bottom right." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 435, + 178, + 543, + 285 + ], + "blocks": [ + { + "bbox": [ + 435, + 178, + 543, + 285 + ], + "lines": [ + { + "bbox": [ + 435, + 178, + 543, + 285 + ], + "spans": [ + { + "bbox": [ + 435, + 178, + 543, + 285 + ], + "type": "image", + "image_path": "06def6696d786f3e297e1f8939c2997b3ee945b57c817a77910f4935e3ac3400.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 361, + 555, + 385 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 361, + 555, + 385 + ], + "spans": [ + { + "bbox": [ + 313, + 361, + 555, + 385 + ], + "type": "text", + "content": "readers to the supplementary material for a more detailed investigation of the results." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 399, + 484, + 412 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 399, + 484, + 412 + ], + "spans": [ + { + "bbox": [ + 313, + 399, + 484, + 412 + ], + "type": "text", + "content": "5. Conclusions and Future Work" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 421, + 555, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 421, + 555, + 661 + ], + "spans": [ + { + "bbox": [ + 313, + 421, + 555, + 661 + ], + "type": "text", + "content": "In this paper, we introduced the STU dataset, designed for anomaly segmentation in LiDAR point clouds. The dataset was collected using eight cameras and a LiDAR sensor, which provides a surround view of environments through images and high-resolution LiDAR data. We extensively annotated point clouds with \"unlabeled\", \"anomaly\", and \"inlier\" classes, as well as introduced additional sequences without anomalies for training and evaluation. The results of the baseline methods indicate that there is a significant performance gap between 2D and 3D anomaly segmentation, which underscores numerous ongoing research opportunities in the field. Our objective with this paper is to provide researchers with annotated data and standardized benchmarks to facilitate the comparison of their anomaly segmentation approaches in 3D. We believe the dataset could be especially useful for the evaluation of methods that use temporal LiDAR data and multi-modal data in the future, and can drive innovation in developing more robust anomaly detection systems that can better handle complex real-world scenarios." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 666, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 666, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 666, + 556, + 715 + ], + "type": "text", + "content": "Acknowledgments. A. Nekrasov's acknowledges funding by BMBF project \"WestAI\" (grant no. 01IS22094D). J. S. Berrio Perez acknowledges funding by the German Academic Exchange Service (DAAD Project-ID 30001831)." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "11882" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "text", + "content": "[1] Jens Behley, Martin Garbade, Andres Milioto, Jan Quenzel, Sven Behnke, Cyril Stachniss, and Juergen Gall. SemanticKITTI: A Dataset for Semantic Scene Understanding of LiDAR Sequences. In International Conference on Computer Vision (ICCV), 2019. 2, 3, 4, 6, 7, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 148, + 296, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 148, + 296, + 191 + ], + "spans": [ + { + "bbox": [ + 61, + 148, + 296, + 191 + ], + "type": "text", + "content": "[2] Petra Bevandić, Ivan Kreso, Marin Oršić, and Siniša Šegvic. Simultaneous Semantic Segmentation and Outlier Detection in Presence of Domain Shift. In German Conference on Pattern Recognition (GCPR), 2019. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 194, + 295, + 247 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 194, + 295, + 247 + ], + "spans": [ + { + "bbox": [ + 62, + 194, + 295, + 247 + ], + "type": "text", + "content": "[3] Hermann Blum, Paul-Edouard Sarlin, Juan I. Nieto, Roland Y. Siegwart, and César Cadena. Fishyscapes: A benchmark for safe semantic segmentation in autonomous driving. International Conference on Computer Vision Workshop (ICCV'W), 2019. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 250, + 295, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 250, + 295, + 303 + ], + "spans": [ + { + "bbox": [ + 62, + 250, + 295, + 303 + ], + "type": "text", + "content": "[4] Hermann Blum, Paul-Edouard Sarlin, Juan Nieto, Roland Siegwart, and Cesar Cadena. The Fishyscapes Benchmark: Measuring Blind Spots in Semantic Segmentation. International Journal on Computer Vision (IJCV), 2021. 2, 3, 4, 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 306, + 295, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 306, + 295, + 350 + ], + "spans": [ + { + "bbox": [ + 62, + 306, + 295, + 350 + ], + "type": "text", + "content": "[5] Daniel Bogdoll, Svenja Uhlemeyer, Kamil Kowol, and J Marius Zöllner. Perception datasets for anomaly detection in autonomous driving: A survey. In Intelligent Vehicles Symposium (IV), 2023. 2, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 352, + 295, + 427 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 352, + 295, + 427 + ], + "spans": [ + { + "bbox": [ + 62, + 352, + 295, + 427 + ], + "type": "text", + "content": "[6] Daniel Bogdoll, Iramm Hamdard, Lukas Namgyu Rößler, Felix Geisler, Muhammed Bayram, Felix Wang, Jan Imhof, Miguel de Campos, Anushervon Tabarov, Yitian Yang, Hanno Gottschalk, and J. Marius Zöllner. AnoVox: A Benchmark for Multimodal Anomaly Detection in Autonomous Driving. In ECCV 2024 W-CODA workshop, 2024. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 430, + 295, + 495 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 430, + 295, + 495 + ], + "spans": [ + { + "bbox": [ + 62, + 430, + 295, + 495 + ], + "type": "text", + "content": "[7] Holger Caesar, Varun Bankiti, Alex H. Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuScenes: A multimodal dataset for autonomous driving. In Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 498, + 295, + 530 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 498, + 295, + 530 + ], + "spans": [ + { + "bbox": [ + 62, + 498, + 295, + 530 + ], + "type": "text", + "content": "[8] Jun Cen, Peng Yun, Junhao Cai, Michael Yu Wang, and Ming Liu. Open-set 3D Object Detection. In International Conference on 3D Vision (3DV), 2021. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 533, + 295, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 533, + 295, + 576 + ], + "spans": [ + { + "bbox": [ + 62, + 533, + 295, + 576 + ], + "type": "text", + "content": "[9] Anirudh S. Chakravarthy, Meghanana Reddy Ganesina, Peiyun Hu, Laura Leal-Taixe, Shu Kong, Deva Ramanan, and Aljosa Osep. Lidar Panoptic Segmentation in an Open World. International Journal on Computer Vision (IJCV), 2024. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 578, + 295, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 578, + 295, + 643 + ], + "spans": [ + { + "bbox": [ + 57, + 578, + 295, + 643 + ], + "type": "text", + "content": "[10] Robin Chan, Krzysztof Lis, Svenja Uhlemeyer, Hermann Blum, Sina Honari, Roland Siegwart, Pascal Fua, Mathieu Salzmann, and Matthias Rottmann. SegmentMeIfYouCan: A Benchmark for Anomaly Segmentation. In Proceedings of the Neural Information Processing Systems Track on Datasets and Benchmarks, 2021. 2, 3, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 646, + 295, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 646, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 57, + 646, + 295, + 689 + ], + "type": "text", + "content": "[11] Robin Chan, Matthias Rottmann, and Hanno Gottschalk. Entropy maximization and meta classification for out-of-distribution detection in semantic segmentation. In International Conference on Computer Vision (ICCV), 2021. 1, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "type": "text", + "content": "[12] Liang-Chieh Chen, Yukun Zhu, George Papandreou, Florian Schroff, and Hartwig Adam. Encoder-Decoder with Atrous" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "text", + "content": "Separable Convolution for Semantic Image Segmentation. In European Conference on Computer Vision (ECCV), 2018. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 97, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 97, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 97, + 553, + 150 + ], + "type": "text", + "content": "[13] Bowen Cheng, Ishan Misra, Alexander G. Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention Mask Transformer for Universal Image Segmentation. In Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 3, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 152, + 553, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 553, + 195 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 553, + 195 + ], + "type": "text", + "content": "[14] Bowen Cheng, Alexander G. Schwing, and Alexander Kirilov. Per-pixel classification is not all you need for semantic segmentation. In Neural Information Processing Systems (NeurIPS), 2021. 3, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 198, + 553, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 198, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 316, + 198, + 553, + 262 + ], + "type": "text", + "content": "[15] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 2, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 264, + 553, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 264, + 553, + 297 + ], + "spans": [ + { + "bbox": [ + 316, + 264, + 553, + 297 + ], + "type": "text", + "content": "[16] Anja Delić, Matej Grić, and Siniša Šegvic. Outlier detection by ensembling uncertainty with negative objectness. In British Machine Vision Conference (BMVC), 2024. 1, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 299, + 553, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 299, + 553, + 342 + ], + "spans": [ + { + "bbox": [ + 317, + 299, + 553, + 342 + ], + "type": "text", + "content": "[17] kuefeng Du, Zhaoning Wang, Mu Cai, and Yixuan Li. Vos: Learning what you don't know by virtual outlier synthesis. In International Conference on Learning Representations (ICLR), 2021. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 343, + 553, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 343, + 553, + 398 + ], + "spans": [ + { + "bbox": [ + 317, + 343, + 553, + 398 + ], + "type": "text", + "content": "[18] Stefano Gasperini, Alvaro Marcos-Ramiro, Michael Schmidt, Nassir Navab, Benjamin Busam, and Federico Tombari. Segmenting Known Objects and Unseen Unknowns without Prior Knowledge. In International Conference on Computer Vision (ICCV), 2023. 1, 3, 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 399, + 555, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 399, + 555, + 431 + ], + "spans": [ + { + "bbox": [ + 317, + 399, + 555, + 431 + ], + "type": "text", + "content": "[19] A Geiger, P Lenz, C Stiller, and R Urtasun. Vision meets robotics: The KITTI dataset. In The International Journal of Robotics Research, 2013. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 434, + 553, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 434, + 553, + 465 + ], + "spans": [ + { + "bbox": [ + 317, + 434, + 553, + 465 + ], + "type": "text", + "content": "[20] Matej Grcic, Petra Bevandic, and Sinisa Segvic. Dense open-set recognition with synthetic outliers generated by real nvp. ArXiv, abs/2011.11094, 2020. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 468, + 553, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 468, + 553, + 510 + ], + "spans": [ + { + "bbox": [ + 317, + 468, + 553, + 510 + ], + "type": "text", + "content": "[21] Matej Grcic, Petra Bevandić, and Sinisa Šegvic. Densehybrid: Hybrid anomaly detection for dense open-set recognition. In European Conference on Computer Vision (ECCV), 2022. 1, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 513, + 553, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 513, + 553, + 555 + ], + "spans": [ + { + "bbox": [ + 317, + 513, + 553, + 555 + ], + "type": "text", + "content": "[22] Krishnam Gupta, Syed Ashar Javed, Vineet Gandhi, and K. Madhava Krishna. MergeNet: A Deep Net Architecture for Small Obstacle Discovery. In International Conference on Robotics and Automation (ICRA), 2018. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 558, + 553, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 558, + 553, + 632 + ], + "spans": [ + { + "bbox": [ + 317, + 558, + 553, + 632 + ], + "type": "text", + "content": "[23] Jianhua Han, Xiwen Liang, Hang Xu, Kai Chen, Lanqing Hong, Jiageng Mao, Chaoqiang Ye, Wei Zhang, Zhenguo Li, Xiaodan Liang, and Chunjing Xu. SODA10M: A LargeScale 2D Self/Semi-Supervised Object Detection Dataset for Autonomous Driving. In Proceedings of the Neural Information Processing Systems Track on Datasets and Benchmarks, 2021. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 635, + 553, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 635, + 553, + 678 + ], + "spans": [ + { + "bbox": [ + 317, + 635, + 553, + 678 + ], + "type": "text", + "content": "[24] Dan Hendrycks and Kevin Gimpel. A Baseline for Detecting Misclassified and Out-of-Distribution Examples in Neural Networks. In International Conference on Learning Representations (ICLR), 2018. 2, 3, 6, 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 681, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 681, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 681, + 553, + 713 + ], + "type": "text", + "content": "[25] Dan Hendrycks, Steven Basart, Mantas Mazeika, Andy Zou, Joe Kwon, Mohammadreza Mostajabi, Jacob Steinhardt, and Dawn Song. Scaling Out-of-Distribution Detection for Real-" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11883" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "text", + "content": "World Settings. In International Conference on Machine Learning (ICML), 2022. 2, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 151 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 151 + ], + "type": "text", + "content": "[26] Dan Hendrycks, Steven Basart, Mantas Mazeika, Andy Zou, Mohammadreza Mostajabi, Jacob Steinhardt, and Dawn Xiaodong Song. Scaling out-of-distribution detection for real-world settings. In International Conference on Machine Learning, 2022. 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 294, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 294, + 207 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 294, + 207 + ], + "type": "text", + "content": "[27] Yen-Chang Hsu, Yilin Shen, Hongxia Jin, and Zsolt Kira. Generalized ODIN: Detecting Out-of-Distribution Image Without Learning From Out-of-Distribution Data. In Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 208, + 294, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 208, + 294, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 208, + 294, + 262 + ], + "type": "text", + "content": "[28] Chengjie Huang, Van Duong Nguyen, Vahdat Abdelzad, Christopher Gus Mannes, Luke Rowe, Benjamin Therien, Rick Salay, and K. Czarnecki. Out-of-distribution detection for lidar-based 3d object detection. IEEE Intelligent Transportation Systems Conference (ITSC), 2022. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 264, + 294, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 264, + 294, + 297 + ], + "spans": [ + { + "bbox": [ + 56, + 264, + 294, + 297 + ], + "type": "text", + "content": "[29] Håkon Hukkelås and Frank Lindseth. Deepprivacy2: Towards realistic full-body anonymization. In Winter Conference on Applications of Computer Vision (WACV), 2023. 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 298, + 294, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 298, + 294, + 342 + ], + "spans": [ + { + "bbox": [ + 56, + 298, + 294, + 342 + ], + "type": "text", + "content": "[30] Jaedong Hwang, Seoul Wug Oh, Joon-Young Lee, and Bohyung Han. Exemplar-Based Open-Set Panoptic Segmentation Network. In Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 343, + 294, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 343, + 294, + 398 + ], + "spans": [ + { + "bbox": [ + 56, + 343, + 294, + 398 + ], + "type": "text", + "content": "[31] Sanghun Jung, Jungsoo Lee, Daehoon Gwak, Sungha Choi, and Jaegul Choo. Standardized Max Logits: A Simple yet Effective Approach for Identifying Unexpected Road Obstacles in Urban-Scene Segmentation. In International Conference on Computer Vision (ICCV), 2021. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 399, + 294, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 399, + 294, + 442 + ], + "spans": [ + { + "bbox": [ + 56, + 399, + 294, + 442 + ], + "type": "text", + "content": "[32] Alexander Kirillov, Kaiming He, Ross Girshick, Carsten Rother, and Piotr Dólar. Panoptic Segmentation. In Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 444, + 294, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 444, + 294, + 488 + ], + "spans": [ + { + "bbox": [ + 56, + 444, + 294, + 488 + ], + "type": "text", + "content": "[33] Michael Kösel, Marcel Schreiber, Michael Ulrich, Claudius Gläser, and Klaus Dietmayer. Revisiting Out-of-Distribution Detection in LiDAR-based 3D Object Detection. In Intelligent Vehicles Symposium (IV), 2024. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 489, + 294, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 489, + 294, + 533 + ], + "spans": [ + { + "bbox": [ + 56, + 489, + 294, + 533 + ], + "type": "text", + "content": "[34] Balaji Lakshminarayanan, Alexander Pritzel, and Charles Blundell. Simple and Scalable Predictive Uncertainty Estimation using Deep Ensembles. In Neural Information Processing Systems (NeurIPS), 2017. 2, 3, 6, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 534, + 294, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 534, + 294, + 578 + ], + "spans": [ + { + "bbox": [ + 56, + 534, + 294, + 578 + ], + "type": "text", + "content": "[35] Seungjae Lee, Hyungtae Lim, and Hyun Myung. Patchwork++: Fast and robust ground segmentation solving partial under-segmentation using 3D point cloud. In Proc. IEEE/RSJ Int. Conf. Intell. Robots Syst., 2022. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 579, + 294, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 579, + 294, + 643 + ], + "spans": [ + { + "bbox": [ + 56, + 579, + 294, + 643 + ], + "type": "text", + "content": "[36] Kaican Li, Kai Chen, Haoyu Wang, Lanqing Hong, Chaoqiang Ye, Jianhua Han, Yukuai Chen, Wei Zhang, Chunjing Xu, Dit-Yan Yeung, et al. Coda: A real-world road corner case dataset for object detection in autonomous driving. In European Conference on Computer Vision (ECCV), 2022. 2, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 646, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 646, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 646, + 294, + 689 + ], + "type": "text", + "content": "[37] Chen Liang, Wenguan Wang, Jiaxu Miao, and Yi Yang. GMMSeg: Gaussian Mixture based Generative Semantic Segmentation Models. In Neural Information Processing Systems (NeurIPS), 2022. 1, 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "text", + "content": "[38] Kira Maag, Robin Chan, Svenja Uhlemeyer, Kamil Kowol, and Hanno Gottschalk. Two Video Data Sets for Tracking" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "and Retrieval of Out of Distribution Objects. In Asian Conference on Computer Vision (ACCV), 2022. 2, 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 95, + 553, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 95, + 553, + 161 + ], + "spans": [ + { + "bbox": [ + 316, + 95, + 553, + 161 + ], + "type": "text", + "content": "[39] Jiageng Mao, Minzhe Niu, Chenhan Jiang, Hanxue Liang, Jingheng Chen, Xiaodan Liang, Yamin Li, Chaoqiang Ye, Wei Zhang, Zhenguo Li, Jie Yu, Hang Xu, and Chunjing Xu. One Million Scenes for Autonomous Driving: ONCE Dataset. In Neural Information Processing Systems (NeurIPS), 2021. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 162, + 553, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 162, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 316, + 162, + 553, + 205 + ], + "type": "text", + "content": "[40] Rodrigo Marcuzzi, Lucas Nunes, Louis Wiesmann, Jens Behley, and Cyril Stachniss. Mask-Based Panoptic LiDAR Segmentation for Autonomous Driving. In IEEE Robotics And Automation Letters (RAL), 2023. 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 205, + 553, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 205, + 553, + 249 + ], + "spans": [ + { + "bbox": [ + 316, + 205, + 553, + 249 + ], + "type": "text", + "content": "[41] Nazir Nayal, Misra Yavuz, João F. Henriques, and Fatma Güney. RbA: Segmenting Unknown Regions Rejected by All. In International Conference on Computer Vision (ICCV), 2023. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 250, + 553, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 250, + 553, + 293 + ], + "spans": [ + { + "bbox": [ + 316, + 250, + 553, + 293 + ], + "type": "text", + "content": "[42] Alexey Nekrasov, Rui Zhou, Miriam Ackermann, Alexander Hermans, and Matthias Rottmann Bastian Leibe. OoDIS: Anomaly Instance Segmentation Benchmark. arXiv preprint arXiv:2406.11835, 2024. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 293, + 553, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 293, + 553, + 338 + ], + "spans": [ + { + "bbox": [ + 316, + 293, + 553, + 338 + ], + "type": "text", + "content": "[43] Gerhard Neuhold, Tobias Ollmann, Samuel Rota Bulo, and Peter Kontschieder. The mapillary vistas dataset for semantic understanding of street scenes. In International Conference on Computer Vision (ICCV), 2017. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 338, + 553, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 338, + 553, + 392 + ], + "spans": [ + { + "bbox": [ + 316, + 338, + 553, + 392 + ], + "type": "text", + "content": "[44] Lucas Nunes, Xieyuanli Chen, Rodrigo Marcuzzi, Aljosa Osep, Laura Leal-Taixe, Cyril Stachniss, and Jens Behley. Unsupervised Class-Agnostic Instance Segmentation of 3D LiDAR Data for Autonomous Vehicles. IEEE Robotics And Automation Letters (RAL), 2022. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 393, + 553, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 393, + 553, + 448 + ], + "spans": [ + { + "bbox": [ + 316, + 393, + 553, + 448 + ], + "type": "text", + "content": "[45] Peter Pinggera, Sebastian Ramos, Stefan Gehrig, Uwe Franke, Carsten Rother, and Rudolf Mester. Lost and Found: Detecting Small Road Hazards for Self-Driving Vehicles. In International Conference on Intelligent Robots and Systems (IROS), 2016. 2, 3, 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 449, + 553, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 449, + 553, + 503 + ], + "spans": [ + { + "bbox": [ + 316, + 449, + 553, + 503 + ], + "type": "text", + "content": "[46] Aldi Piroli, Vinzenz Dallabetta, Johannes Kopp, Marc Walesa, Daniel Meissner, and Klaus Dietmayer. LS-VOS: Identifying Outliers in 3D Object Detections Using Latent Space Virtual Outlier Synthesis. In IEEE Intelligent Transportation Systems Conference (ITSC), 2023. 2, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 504, + 553, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 504, + 553, + 547 + ], + "spans": [ + { + "bbox": [ + 316, + 504, + 553, + 547 + ], + "type": "text", + "content": "[47] Shyam Nandan Rai, Fabio Cermelli, Dario Fontanel, Carlo Masone, and Barbara Caputo. Unmasking Anomalies in Road-Scene Segmentation. In International Conference on Computer Vision (ICCV), 2023. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 548, + 553, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 548, + 553, + 592 + ], + "spans": [ + { + "bbox": [ + 316, + 548, + 553, + 592 + ], + "type": "text", + "content": "[48] Jenny Seidenschwarz, Aljosa O'Sep, Francesco Ferroni, Simon Lucey, and Laura Leal-Taixe. SeMoLi: What Moves Together Belongs Together. In Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 592, + 553, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 592, + 553, + 635 + ], + "spans": [ + { + "bbox": [ + 316, + 592, + 553, + 635 + ], + "type": "text", + "content": "[49] Aasheesh Singh, Aditya Kamireddypalli, Vineet Gandhi, and K. Madhava Krishna. LiDAR guided Small obstacle Segmentation. In International Conference on Intelligent Robots and Systems (IROS), 2020. 2, 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 635, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 553, + 689 + ], + "type": "text", + "content": "[50] Nitish Srivastava, Geoffrey Hinton, Alex Krizhevsky, Ilya Sutskever, and Ruslan Salakhutdinov. Dropout: a simple way to prevent neural networks from overfitting. In Neural Information Processing Systems (NeurIPS), 2014. 2, 6, 7" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 690, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 690, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 690, + 553, + 713 + ], + "type": "text", + "content": "[51] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou," + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "11884" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 620 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 76, + 72, + 296, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 296, + 115 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 296, + 115 + ], + "type": "text", + "content": "Yuning Chai, Benjamin Caine, et al. Scalability in perception for autonomous driving: Waymo open dataset. In Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 295, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 295, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 295, + 162 + ], + "type": "text", + "content": "[52] DashcamCleaner Development Team. Dashcamcleaner: Censor identifiable information in videos from dashcam recordings. https://github.com/taehse/DashcamCleaner, 2024.4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 295, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 295, + 217 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 295, + 217 + ], + "type": "text", + "content": "[53] Tefft, Brian C. . The Prevalence of Motor Vehicle Crashes Involving Road Debris, United States, 2011-2014. https://aaafoundation.org/wp-content/uploads/2017/12/RoadDebris_REPORT_2015.pdf, 2015. [Online]. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 219, + 295, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 219, + 295, + 274 + ], + "spans": [ + { + "bbox": [ + 56, + 219, + 295, + 274 + ], + "type": "text", + "content": "[54] Yu Tian, Yuyuan Liu, Guansong Pang, Fengbei Liu, Yuanhong Chen, and Gustavo Carneiro. Pixel-wise Energy-biased Abstention Learning for Anomaly Segmentation on Complex Urban Driving Scenes. In European Conference on Computer Vision (ECCV), 2022. 1, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 275, + 295, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 275, + 295, + 329 + ], + "spans": [ + { + "bbox": [ + 56, + 275, + 295, + 329 + ], + "type": "text", + "content": "[55] Tzu-Yun Tseng, Alexey Nekrasov, Malcolm Burdorf, Bastian Leibe, Julie Stephany Berrio Perez, Mao Shan, and Stewart Worrall. Panoptic-CUDAL Technical Report: Rural Australia Point Cloud Dataset in Rainy Conditions. arXiv preprint arXiv:2503.16378, 2025. 3, 4, 6, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 331, + 295, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 331, + 295, + 374 + ], + "spans": [ + { + "bbox": [ + 56, + 331, + 295, + 374 + ], + "type": "text", + "content": "[56] S. Verma, J. S. Berrio, S. Worrall, and E. Nebot. Automatic extrinsic calibration between a camera and a 3d lidar using 3d point and plane correspondences. In IEEE Intelligent Transportation Systems Conference (ITSC), 2019. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 376, + 295, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 376, + 295, + 430 + ], + "spans": [ + { + "bbox": [ + 56, + 376, + 295, + 430 + ], + "type": "text", + "content": "[57] Ignacio Vizzo, Tiziano Guadagnino, Benedikt Mersch, Louis Wiesmann, Jens Behley, and Cyril Stachniss. KISS-ICP: In Defense of Point-to-Point ICP – Simple, Accurate, and Robust Registration If Done the Right Way. In IEEE Robotics And Automation Letters (RAL), 2023. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 431, + 295, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 431, + 295, + 474 + ], + "spans": [ + { + "bbox": [ + 56, + 431, + 295, + 474 + ], + "type": "text", + "content": "[58] Kelvin Wong, Shenlong Wang, Mengye Ren, Ming Liang, and Raquel Urtasun. Identifying Unknown Instances for Autonomous Driving. In Conference on Robot Learning (CoRL), 2019. 2, 3, 7, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 475, + 295, + 519 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 475, + 295, + 519 + ], + "spans": [ + { + "bbox": [ + 56, + 475, + 295, + 519 + ], + "type": "text", + "content": "[59] Yajun Xu, Huan Hu, Xiaoya Zhu, Yibing Nan, Kai Wang, ZhaoXiang Liu, and Shiguo Lian. Raod: A benchmark for road abandoned object detection from video surveillance. IEEE Access, 2024. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 520, + 295, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 520, + 295, + 564 + ], + "spans": [ + { + "bbox": [ + 56, + 520, + 295, + 564 + ], + "type": "text", + "content": "[60] Kadir Yilmaz, Jonas Schult, Alexey Nekrasov, and Bastian Leibe. Mask4Former: Mask Transformer for 4D Panoptic Segmentation. In International Conference on Robotics and Automation (ICRA), 2024. 3, 6, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 566, + 295, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 566, + 295, + 620 + ], + "spans": [ + { + "bbox": [ + 56, + 566, + 295, + 620 + ], + "type": "text", + "content": "[61] Fisher Yu, Haofeng Chen, Xin Wang, Wenqi Xian, Yingying Chen, Fangchen Liu, Vashisht Madhavan, and Trevor Darrell. Bdd100k: A diverse driving dataset for heterogeneous multitask learning. In Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 3" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11885" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Stabilizing and Accelerating Autofocus with Expert Trajectory Regularized Deep Reinforcement Learning/a699e1d4-e683-4371-bc56-144538826827_content_list.json b/2025/Stabilizing and Accelerating Autofocus with Expert Trajectory Regularized Deep Reinforcement Learning/a699e1d4-e683-4371-bc56-144538826827_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..a2dc6a90e90d13275ffed7ec5c8eb0be04a00f9a --- /dev/null +++ b/2025/Stabilizing and Accelerating Autofocus with Expert Trajectory Regularized Deep Reinforcement Learning/a699e1d4-e683-4371-bc56-144538826827_content_list.json @@ -0,0 +1,1895 @@ +[ + { + "type": "text", + "text": "Stabilizing and Accelerating Autofocus with Expert Trajectory Regularized Deep Reinforcement Learning", + "text_level": 1, + "bbox": [ + 119, + 128, + 879, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Shouhang Zhu1, Chenglin Li1, Yuankun Jiang2, Li Wei1, Nuowen Kan1, Ziyang Zheng1, Wenrui Dai2, Junni Zou2, Hongkai Xiong1", + "bbox": [ + 150, + 202, + 846, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Department of Electronic Engineering, Shanghai Jiao Tong University", + "bbox": [ + 217, + 238, + 779, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Department of Computer Science and Engineering, Shanghai Jiao Tong University", + "bbox": [ + 168, + 255, + 828, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ shzhu2001, lcl1985, wei.li, kannw.1230, zhengziyang, xionghongkai}@sjtu.edu.cn", + "bbox": [ + 150, + 275, + 846, + 292 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ $\\{\\text { yuankunjiang, daiwenrui, zoujunni }\\} @$ sjtu.edu.cn", + "bbox": [ + 284, + 292, + 707, + 309 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 344, + 326, + 359 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Autofocus is a crucial component of modern digital cameras. While recent learning-based methods achieve state-of-the-art in focus prediction accuracy, they unfortunately ignore the potential focus hunting phenomenon of back-and-forth lens movement in the multi-step focusing procedure. To address this, in this paper, we propose an expert regularized deep reinforcement learning (DRL)-based approach for autofocus, which can utilize the sequential information of lens movement trajectory to both enhance the multi-step in-focus prediction accuracy and reduce the chance of focus hunting. Our method generally follows an actor-critic framework. To accelerate the DRL's training with a higher sample efficiency, we initialize the policy with a pre-trained single-step prediction network, where the network is further improved by modifying the output of absolute in-focus position distribution to the relative lens movement distribution to establish a better mapping between input images and lens movement. To further stabilize DRL's training with a lower occurrence of focus hunting in the resulting lens movement trajectory, we generate some offline trajectories based on prior knowledge to avoid focus hunting, which are then leveraged as an offline dataset of expert trajectories to regularize the actor network's training. Empirical evaluations show that our method outperforms those learning-based methods on public benchmarks, with higher single- and multi-step prediction accuracy, and a significant reduction of focus hunting rate.", + "bbox": [ + 89, + 376, + 485, + 785 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 800, + 220, + 815 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Autofocus (AF) is a process through which the camera can determine automatically a lens position to focus on the region of interest (RoI). It is a crucial step in image capturing, since once an image is already captured in a defocused and blurry state, it is then difficult to restore its sharpness", + "bbox": [ + 89, + 825, + 483, + 902 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "through any post-processing due to the varying effects of blur influenced by factors such as camera aperture and radial distortion [1]. In addition, the beauty of a moment often lies in its fleeting nature, which thus imposes an imperative demand for the fast focusing speed of an AF algorithm. However, due to the instability in focusing systems incurred by some interference factors, the AF mechanism of the camera lens may fail to accurately lock onto the focal point, resulting in a repeated movement of the lens within a certain range. This phenomenon, as illustrated in Fig. 1, is referred to as focus hunting (FH), whereby the camera is unable to identify the focus plane and thus drives the lens forward and backward through a set range of focus hunting for it. It then causes variations in the image sharpness and changes the lens's field of view (FoV), ultimately leading to an unstable focusing experience with longer focusing time, which is a common issue in the traditional AF-assisted cameras [2]. It is thus of great importance to develop an AF algorithm that not only achieves a higher in-focus accuracy but can also stabilize the trajectory of lens movement with a faster convergence speed to the in-focus position, thus reducing the probability of occurrence of focus hunting [43, 48].", + "bbox": [ + 511, + 345, + 906, + 679 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "There is a large body of work on the traditional autofocus methods, generally falling into two categories: search-based and phase-based methods. Search-based AF methods [40, 42] search within a range of focus distances to locate the lens position of the extreme sharpness value, which is then determined as the in-focus lens position. Thus, they need to move the lens back and forth, which leads to a highly accurate identification of the in-focus position, but with an annoyingly slow focusing procedure and inevitably the occurrence of focus hunting. Phase-based approaches, on the other hand, rely on the dual-pixel (DP) sensors to obtain the phase data from the left and right sub-images, which is then leveraged to predict the lens movement and determine the in-focus position in a single step of lens", + "bbox": [ + 511, + 689, + 908, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 236, + 0, + 810, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "26440", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ab10fbb75f3bcafaf8e897d27a362794ad1a3d8259bc6e43592c067a2300ce66.jpg", + "image_caption": [ + "Figure 1. Visualization of the focus hunting phenomenon. The lens moves back and forth within a certain range, resulting in a slight change of field of view (FoV) and sharpness. The red, purple, and green lines mark the trajectory of the same target's locations in the picture obtained by three models. The yellow line is a horizontal line as a location reference. (1) The trajectory generated by the learning-based method [7] that predicts the lens position directly. (2) The trajectory generated by the learning-based method [7] that predicts the relative movement of lens. (3) The trajectory generated by our DRL-based method. It can be seen that the FOV and sharpness in trajectory (1) change more significantly, while those in trajectory (2) change slightly, and the lens terminates at the GT after Step 2 in trajectory (3)." + ], + "image_footnote": [], + "bbox": [ + 94, + 78, + 903, + 266 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/bf126b2aeedf7e1e82303c6fbbd7ad57427a5c0ede5e53915a4b1e2a6ed3fc4a.jpg", + "image_caption": [ + "Figure 2. Trajectories generated by predicting the lens position [7] (red line), predicting the relative movement (purple line), and DRL (green line). The trajectory generated by our DRL is the smoothest, indicating a stable and fast focusing procedure." + ], + "image_footnote": [], + "bbox": [ + 94, + 368, + 480, + 445 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "movement. Though accelerating the focusing procedure, the phase data obtained by sensors is highly susceptible to noise, leading to a relatively low in-focus accuracy. Especially in some extreme cases, e.g., the low-light conditions, the phase detection may need to be executed multiple times at different lens positions to predict a more accurate in-focus position, still at the risk of focus hunting.", + "bbox": [ + 89, + 506, + 482, + 611 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recent attempts [7, 16] have concentrated on addressing the above trade-off between accuracy and speed, by exploiting the expressive power of neural networks to directly learn a mapping from the captured image at any lens position to the predicted in-focus position. However, when we directly treat the learned neural network as a lens movement policy and apply it at multiple steps to adjust the lens movement, we notice that some of the resulting trajectories still suffer from the back-and-forth movement of the lens, as shown in Fig. 1 and Fig. 2, due to the lack of leveraging the sequential information obtained from multiple exploration steps and the ignorance of considering future effects of the current prediction. Deep reinforcement learning (DRL), with successful applications in sequential decision-making tasks such as autonomous driving [12] and robot control [8], has enabled the agent to adaptively adjust its policy through iterative interactions with the environment, based both on the immediate reward of current decision step and the long-term reward over the entire sequence of decisions.", + "bbox": [ + 89, + 613, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Inspired by this, we propose in this paper an expert regularized DRL-based approach for camera autofocus, which is able to utilize the sequential information of lens movement trajectory to both enhance the multi-step in-focus prediction accuracy and reduce the chance of focus hunting. Our proposed method generally follows an actor-critic framework, within which the actor network (i.e., policy network) takes the captured left and right dual-pixel images with the positional information of RoI and lens as the state input, and sequentially generates a trajectory by making decisions on the lens movement. The critic network, on the other hand, receives the same input as the actor network and estimates the future expected accumulated reward given the current state, which is used to evaluate the action predicted by the actor network and thus eliminate suboptimal actions. To accelerate the DRL's training with a higher sample efficiency, we initialize the actor network with a pre-trained single-step prediction network based on the MobileNet-V2 architecture [33], where to be more consistent with the environment transition in AF tasks, the output of absolute in-focus position distribution as used in [7] is modified to the relative lens movement distribution. The critic network also takes the same MobileNet-V2 architecture. To further stabilize the DRL's training with lower occurrence of focus hunting in the resulting lens movement trajectory, we generate some offline trajectories based on prior knowledge to avoid focus hunting, which are then leveraged as an offline dataset of expert trajectories to regularize the actor network's training. Our main contributions are summarized as follows:", + "bbox": [ + 511, + 369, + 906, + 808 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- To overcome the limitation of single-step prediction networks, we formulate AF as a multi-step task and then propose a DRL-based framework to fully utilize the sequential information of lens movement trajectory and adjust the policy based on the immediate and long-term rewards.", + "- We initialize the policy with a pre-trained single-step pre" + ], + "bbox": [ + 513, + 810, + 906, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "26441", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "dition network, where the network is improved by modifying the output of absolute in-focus position distribution to relative lens movement distribution to establish a better mapping between input images and lens movement.", + "bbox": [ + 102, + 90, + 482, + 151 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We construct an offline dataset of expert trajectories based on prior knowledge, regularizing the training of the policy network towards reducing occurrence of focus hunting.", + "- Empirical evaluations demonstrate that our method outperforms state-of-the-art AF methods on public benchmarks, with higher single- and multi-step prediction accuracy, and a significant reduction of focus hunting rate." + ], + "bbox": [ + 91, + 152, + 482, + 257 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 91, + 271, + 238, + 286 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Traditional passive AF methods can generally be categorized into search-based and phase-based approaches. Search-based approaches calculate and maximize the sharpness of the images based on image statistics [9, 21, 40, 42, 53] or frequency-domain image representations [18, 24, 25, 47, 50] to determine the optimal lens position. While showing high accuracy, these approaches are time-consuming. Additionally, search-based methods inevitably lead to focus hunting, since the peak must be surpassed to locate the exact focus point. Though some methods [14, 39] can reduce the search steps, they cannot solve the problem of focus hunting completely, or can only work on some specific devices.", + "bbox": [ + 89, + 296, + 482, + 477 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "With the development of dual-pixel sensors, phase-based methods have become the primary AF method in modern digital cameras [37]. With the phase information of the two (left/right) split pixels, the disparity can be calculated to determine the lens movement. Consequently, phase-based approaches can move the lens to the in-focus lens position with a single image capture, which is more efficient than search-based approaches. However, factors that are difficult to model, such as lens shading, can affect the calculation of phase, leading to lower accuracy. Numerous works adopt learning-based methods to address the trade-off between accuracy and speed. But these works either apply to specific tasks like microscopy images [31, 51], or are based on phase information obtained by specific hardware [17, 46], which are difficult to reproduce and compare.", + "bbox": [ + 89, + 478, + 482, + 704 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The above methods concentrate on the processed sharpness metrics or phase data, not sufficiently using the information from pictures. Inspired by the stereo vision [19, 20], information obtained from dual-pixel pictures has found increasingly widespread applications in the computer vision community, such as monocular depth estimation [11, 22] and defocus deblurring [1]. In AF, Herrmann et al. [16] utilize convolutional networks to predict the in-focus position from input images and provide a large-scale AF dataset. Choi et al. [7] improve Herrmann's network by incorporating the region of interest position encoding (RoI-PE) and the lens position encoding (Lens-PE), significantly reducing the size while achieving state-of-the-art performance.", + "bbox": [ + 89, + 704, + 482, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Both methods, similar to the phase-based approaches, essentially employ the DP data to estimate the disparity and predict the lens position. But the difference is that phase-based approaches estimate the lens movement through disparity rather than directly predicting the next lens position. Meanwhile, though [7, 16] discuss the multi-step prediction, they only address the accuracy of multi-step prediction without considering the potential for focus hunting, which can greatly affect the speed and accuracy of the AF model.", + "bbox": [ + 511, + 90, + 903, + 227 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The main reason for focus hunting in learning-based methods is the lack of consideration of the sequential information and the future effects of the actions. Yan et al. [49] and Wang et al. [46] adopt the long short-term memory (LSTM) to enhance the model's ability to incorporate previous information. With extra computation, the ability of LSTM to capture the 2-dimension data is still limited. The paradigm of predicting actions from states and the need to incorporate sequential information and future effects inspire us to use reinforcement learning (RL) in AF. Though Yu et al. [54] and Chan et al. [6] have proposed DRL-based autofocus methods, they both concentrate on the design of reward functions and do not fully leverage the advantages of RL. Wang et al. [46] adopt RL in their work to dynamically choose the RoI, rather than predicting the lens position or movement. To the best of our knowledge, little research has been conducted to reduce the occurrence of focus hunting while accelerating the multi-step prediction with a higher accuracy for AF tasks.", + "bbox": [ + 511, + 228, + 906, + 515 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Proposed Method", + "text_level": 1, + "bbox": [ + 513, + 532, + 687, + 550 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Problem Formulation", + "text_level": 1, + "bbox": [ + 511, + 559, + 717, + 574 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We formulate the problem of AF as presented in [7], where the candidate lens position is discretized into $n$ focal indices. Given an input image $I$ captured at a lens position $k \\in \\{1,2,\\dots,n\\}$ , we extract a patch $I_k^p$ corresponding to the region of interest (RoI) that we aim to focus, where $p \\in \\{1,2,\\dots,P\\}$ denotes the spatial position of the RoI in this image. A qualified AF process should be able to predict the correct focal index of the lens from a patch sequence captured by moving the lens, while the sequence should ideally be as short as possible, and without a back-and-forth movement. Considering that modern digital cameras (e.g., in smartphones) are commonly equipped with the advanced dual-pixel imaging system that can provide the phase information for AF, we specifically pay attention to the dual-pixel data, such that the input image $I$ consists of two channels. Thus, the extracted patch at the $k$ -th lens position, with resolution $h \\times w$ , is denoted as $I_k^p \\in \\mathbb{R}^{2 \\times h \\times w}$ . The terms focal stack and single slice are defined as the set of patches $\\{I_k^p\\}_{k=1}^n$ extracted from all lens positions and a single patch $I_k^p$ , respectively. The objective is to predict the in-focus lens position $k^*$ , which is assumed to be unique for", + "bbox": [ + 511, + 583, + 906, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "26442", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/11dce178f552eada748672e8e75193006c340380690198caaa334ae44d9ee5e4.jpg", + "image_caption": [ + "Figure 3. Overall framework of our DRL-based AF model with expert trajectory regularization. The green, orange, blue, and purple boxes specify the actor network, critic network, expert dataset, and environment, respectively." + ], + "image_footnote": [], + "bbox": [ + 93, + 84, + 480, + 224 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "each focal stack.", + "bbox": [ + 89, + 296, + 202, + 310 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Relying on the different types of patch sequences used for AF, our AF problem falls into the multi-step prediction category [7], which can be formulated as:", + "bbox": [ + 89, + 311, + 483, + 357 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nI _ {k _ {0}} ^ {p} \\mapsto k _ {1}, I _ {k _ {1}} ^ {p} \\mapsto k _ {2}, \\dots , I _ {k _ {t}} ^ {p} \\mapsto k ^ {*}. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 165, + 366, + 482, + 386 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Starting from a random focal index $k_{0}$ , it makes a prediction step-by-step, generating a trajectory of the lens positions $\\{k_{0}, k_{1}, k_{2}, \\ldots, k^{*}\\}$ . Instead of directly predicting the absolute lens position as in Eq. (1), in this paper, we consider predicting the relative lens movement at each step, and thus reformulate the multi-step setting of the AF problem as:", + "bbox": [ + 89, + 393, + 483, + 484 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nI _ {k _ {0}} ^ {p} \\mapsto k _ {1} = k _ {0} + a _ {0}, \\dots , I _ {k _ {t}} ^ {p} \\mapsto k ^ {*} = k _ {t} + a _ {t}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 119, + 494, + 482, + 513 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $a_{t}$ is the relative movement of the lens at time step $t$ .", + "bbox": [ + 89, + 521, + 482, + 536 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Overall Framework of DRL-Based AF", + "text_level": 1, + "bbox": [ + 89, + 544, + 421, + 559 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The framework of our DRL-based AF method is shown in Fig. 3, with three major components: an actor network, a critic network, and an offline dataset of expert trajectories.", + "bbox": [ + 89, + 568, + 482, + 613 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Actor network (i.e., policy network) parameterized by $\\theta$ : At each time step $t$ , it takes the state $s_t$ , including the left and right DP images, the RoI-PE and the Lens-PE, as the input, and outputs an action $a_t \\sim \\pi_\\theta(\\cdot | s_t)$ , where the sign of $a_t$ indicates the direction of the lens movement and the absolute value $|a_t|$ represents the moving distance.", + "- Critic network parameterized by $\\phi$ : At each time step $t$ , it takes the state $s_t$ as input, and outputs the estimated value function $V_{\\phi}(s_t)$ for the current state $s_t$ . This value function represents the future expected accumulated reward given the current state, which thus can be used to evaluate the action predicted by the actor network, allowing the DRL agent to identify and eliminate some suboptimal actions. Since this evaluation is based on the future expected accumulated reward, the actions are thus chosen based on their anticipated long-term impact, rather than the immediate return.", + "- Offline dataset of expert trajectories $\\mathcal{D}^E = \\{(s_t, a_t^E)\\}$ : It provides the expert trajectories with state $s_t$ and as" + ], + "bbox": [ + 89, + 614, + 482, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "sociated expert action $a_{t}^{E}$ . These trajectories are generated based on the prior knowledge to avoid focus hunting, and can thus be used to regularize the training of the actor network by maximizing the probability of expert action given a state $s_{t}$ , which ensures that the actor network learns from the high-quality multi-step movement of the lens from the expert.", + "bbox": [ + 524, + 90, + 903, + 196 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Actor Network for Movement Prediction", + "text_level": 1, + "bbox": [ + 511, + 205, + 861, + 220 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To utilize the sequential information of the lens movement trajectory to both enhance the multi-step in-focus prediction accuracy and reduce the chance of focus hunting, the actor network ideally should establish a more accurate mapping between the disparity from DP data and the lens movement.", + "bbox": [ + 511, + 229, + 903, + 318 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The AF procedure can be formulated as a Markov Decision Progress (MDP), defined by a tuple $\\langle S, \\mathcal{A}, \\mathcal{R}, \\mathcal{P}, \\gamma \\rangle$ , where $S$ denotes the state space, $\\mathcal{A}$ denotes the action space and $\\gamma \\in (0,1]$ is the discount factor. At each time step $t$ , a state $s_t \\in S$ , consisting of the left and right DP images, the RoI-PE and the Lens-PE, is received by the RL agent, and an action $a_t \\in \\mathcal{A}$ is then generated by the actor network to determine the lens movement. The environment transition dynamics $P(s_{t+1} | s_t, a_t)$ in AF can be defined as:", + "bbox": [ + 511, + 319, + 905, + 455 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ns _ {t + 1} = g \\left(\\operatorname {c l i p} \\left(a _ {t} + g ^ {- 1} \\left(s _ {t}\\right), 0, k _ {\\max }\\right)\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 465, + 903, + 484 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $g(k_{t}) = s_{t}$ maps the lens position $k_{t}$ to the state $s_{t}$ at time-step $t$ and $k_{max} \\triangleq k_{n}$ is the maximum lens position. The reward function $R(s_{t}, a_{t})$ is defined as the negative absolute distance between the predicted and ground-truth infocus positions, where we additionally introduce a penalty term $R_{fh}$ to mitigate the occurrence of focus hunting:", + "bbox": [ + 511, + 494, + 905, + 585 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nR \\left(s _ {t}, a _ {t}\\right) = - \\left| k _ {t + 1} - k _ {g t} \\right| + R _ {f h}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 593, + 597, + 903, + 614 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $k_{gt}$ denotes the ground-truth lens position, and the penalty term $R_{fh}$ is defined as:", + "bbox": [ + 511, + 625, + 903, + 656 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nR _ {f h} = \\left\\{ \\begin{array}{l l} c, & \\text {i f F H o c c u r s}; \\\\ 0, & \\text {o t h e r w i s e}; \\end{array} \\right. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 616, + 666, + 903, + 708 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $c$ is a negative hyperparameter that penalizes the event of focus hunting. The main goal of RL is to maximize the expected return $J(\\theta)$ defined as:", + "bbox": [ + 511, + 717, + 903, + 762 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nJ (\\theta) = \\mathbb {E} _ {\\pi_ {\\theta}} \\left[ \\sum_ {t = 0} ^ {T} \\gamma^ {t} R \\left(s _ {t}, a _ {t}\\right) \\right], \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 601, + 773, + 903, + 815 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\pi_{\\theta}$ is the policy network parameterized by $\\theta$ and $T$ denotes the maximum time step.", + "bbox": [ + 511, + 824, + 903, + 854 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Since the single-step prediction network model proposed by Choi et al. [7] has demonstrated a state-of-the-art accuracy, we adopt its network architecture as our actor network.", + "bbox": [ + 511, + 854, + 905, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "26443", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/96b66baf09fbbbf481a30e551f6bdf176e716e7ea768cac89c93fdfb152d785b.jpg", + "image_caption": [ + "Figure 4. Illustration of relative movement label." + ], + "image_footnote": [], + "bbox": [ + 138, + 85, + 442, + 209 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "However, the original model in [7] directly predicts the infocus position, which contradicts the principle of using DP data for in-focus prediction. In essence, since there is a clear correspondence between the disparity and lens movement (i.e., through the defocus conversion coefficient), it is more reasonable to predict the lens movement rather than the lens position based on the DP data. Thus, we modify the original network's output of absolute in-focus position distribution in [7] to the relative lens movement distribution, with the inference procedure of actor network shown in Fig. 9. Accordingly, Fig. 4 illustrates the conversion from original absolute position labels to relative movement labels that are used for training. Unlike [17], we expand the action range from $[- \\mathrm{gt}, k_{\\max} - \\mathrm{gt}]$ to $[-k_{\\max}, k_{\\max}]$ , and the label denoting the relative movement towards GT index then converts to $\\mathrm{gt} - f$ . If the actor network predicts an out-of-bound action, it will be clipped to the closest lower/upper bounds. Based on this formulation, we initialize the actor network with a pre-trained MobileNet-V2 [33] optimized based on the ordinal regression loss [10]. For details about the pretrained model of actor network, please refer to Sec. 7 in our supplementary.", + "bbox": [ + 89, + 239, + 483, + 573 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Critic Network Incorporating Future Impact", + "text_level": 1, + "bbox": [ + 89, + 582, + 470, + 598 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The actor network cannot directly incorporate the future reward of current actions, where the probability of any output action with a larger instant reward will be selected, leading to a suboptimal policy. To address this issue, a critic network is further introduced to provide baselines and evaluate the future impact of the current actions.", + "bbox": [ + 89, + 604, + 482, + 695 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The critic network estimates the value function $V_{\\phi}(s_t)$ of the current state $s_t$ , which is defined as the expected future accumulated discount reward of actions. If we further define the future accumulated discount reward of an action under a given state as a Q-function $Q(s_t, a_t)$ , the advantage function $\\hat{A}_t = Q(s_t, a_t) - V_{\\phi}(s_t)$ then evaluates the relative quality of a particular action under the current state. With this advantage function, the actor network can be trained by minimizing the clipped surrogate objective loss, given by:", + "bbox": [ + 89, + 696, + 482, + 832 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {c l i p}} (\\theta) = \\mathbb {E} \\left[ \\min \\left(r _ {t} (\\theta) \\hat {A}, \\operatorname {c l i p} \\left(r _ {t} (\\theta), 1 - \\epsilon , 1 + \\epsilon\\right) \\hat {A} _ {t}\\right), \\right. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 842, + 482, + 859 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $r_t(\\theta) = \\pi_\\theta(a_t|s_t) / \\pi_\\theta^{old}(a_t|s_t)$ is the probability ratio and $\\epsilon$ is a clip hyperparameter.", + "bbox": [ + 89, + 869, + 482, + 901 + ], + "page_idx": 4 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1 Expert trajectory generator." + ], + "code_body": "Input: $\\mathcal{D} = \\{f,GT,(s_0,a_0),\\dots ,(s_n,a_n)\\}_{N};$ \nOutput: $\\mathcal{D}^E$ . \n1: $\\mathcal{D}^E = \\emptyset$ \n2: for all $\\{g(\\cdot),GT,(s_0,a_0),\\dots ,(s_n,a_n)\\} _i\\in \\mathcal{D}$ do \n3: for all $j = 0,1,2,\\dots ,n$ do \n4: $k_{j}\\gets g_{i}^{-1}(s_{j})$ \n5: $o_j\\gets \\mathrm{sign}(k_0 - GT_i)\\cdot |k_j - GT_i| + GT_i;$ \n6: $o_j\\gets \\mathrm{CLIP}(o_j,k_0,GT_i);$ \n7: end for \n8: $(o_{m_0},\\dots ,o_{m_n})\\gets \\mathrm{sort}(o_0,\\dots ,o_n;\\mathrm{reverse} =$ $\\mathrm{Bool}(k_0 > GT_i));$ \n9: for all $j = 1,2,\\dots ,n$ do \n10: $s_j\\gets g_i(o_{m_j});$ \n11: $a_{j - 1}\\leftarrow o_{m_j} - o_{m_{j - 1}};$ \n12: end for \n13: $a_{n}\\gets 0$ \n14: $\\mathcal{D}^E\\gets \\mathcal{D}^E\\cup \\{g(\\cdot),GT,(s_0,a_0),\\dots ,(s_n,a_n)\\} _i;$ \n15: end for \n16: return $\\mathcal{D}^E$", + "bbox": [ + 516, + 109, + 903, + 401 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The critic network, on the other hand, is trained by minimizing the $\\ell_2$ loss between the estimated value function $V_{\\phi}(s_t)$ and the true value function $\\hat{V}_t$ :", + "bbox": [ + 511, + 424, + 906, + 470 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {c r i t i c}} = \\frac {1}{N} \\sum_ {t = 1} ^ {N} \\left(V _ {\\phi} \\left(s _ {t}\\right) - \\hat {V} _ {t}\\right) ^ {2}. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 593, + 479, + 905, + 521 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Expert Trajectory Regularization", + "text_level": 1, + "bbox": [ + 511, + 530, + 807, + 546 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Though the incorporation of a critic network can improve the effectiveness of actor network's training, the exponential growth in the state and action combinations inherently leads to another issue of low sample efficiency in DRL. Recent literature has witnessed the success of incorporating offline data for improving the sample efficiency of DRL's training, both theoretically [38, 45] and empirically [5, 27, 28]. Motivated by this, and to further stabilize the DRL with a lower occurrence of focus hunting in the resulting lens movement trajectory, we generate some offline trajectories based on the prior knowledge to avoid focus hunting, which are then leveraged as an offline dataset $\\mathcal{D}^E = \\{(s_t,a_t^E)\\}$ of expert trajectories to regularize the actor network's training. Algorithm 1 shows our proposed method to generate these expert trajectories, with more details provided in Sec. 8 in our supplementary.", + "bbox": [ + 511, + 553, + 906, + 794 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Fig. 5 shows our proposed framework for expert trajectory regularized DRL's training. At each time, we sample $N$ trajectories from the offline expert dataset $\\mathcal{D}^E$ to calculate the imitation loss:", + "bbox": [ + 511, + 795, + 905, + 854 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {i m i t} (\\theta) = - \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\sum_ {t = 1} ^ {T} \\log p \\left(\\pi_ {\\theta} (\\cdot | s _ {i, t}) = a _ {i, t}\\right), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 532, + 862, + 906, + 904 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "26444", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/a239323905a5adb0550695bb67ac25530096eb6c4600b962e140871a804b90b6.jpg", + "image_caption": [ + "Figure 5. Expert trajectory regularized DRL's training." + ], + "image_footnote": [], + "bbox": [ + 112, + 80, + 464, + 210 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $a_{i,t} = k_{i,t + 1} - k_{i,t}$ denotes the action at time step $t$ in the $i$ -th trajectory, $s_{i,t}$ represents the network input at lens location $k_{i,t}$ , and $\\pi_{\\theta}(\\cdot | s_{i,t})$ represents the output distribution of the actor network, with the total loss defined as:", + "bbox": [ + 89, + 242, + 483, + 303 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} (\\theta) = \\mathcal {L} _ {\\text {c l i p}} (\\theta) + \\lambda \\cdot \\mathcal {L} _ {\\text {i m i t}} (\\theta), \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 178, + 316, + 482, + 333 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\lambda$ balances the online and offline learning. By minimizing the loss $\\mathcal{L}(\\theta)$ , our DRL algorithm maximizes the expected accumulated discount reward $J(\\theta)$ indirectly.", + "bbox": [ + 89, + 345, + 483, + 391 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 405, + 223, + 422 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In the experiments, our actor and critic networks are built upon MobileNetV2 [33]. We reproduce the models for [16] and [7], since no open-source code is available. In the multi-step prediction setting, we only consider the model obtained by [7], since the multi-step model given by [16] must be restrained for each step, which is impractical. More results and analysis are provided in Sec. 9 in our supplementary.", + "bbox": [ + 89, + 431, + 483, + 537 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Dataset. We adopt the publicly available AF dataset constructed by Herrmann et al. [16]. This dataset comprises 51 scenes, each containing 10 stacks with varying compositions. All instances within the dataset are captured at $n = 49$ focal depths, uniformly sampled in the inverse depth space, ranging from 0.102 to 3.91 meters, captured using the Google Pixel 3 smartphones. To create the training and test sets, we extract and sample $128 \\times 128$ patches with a stride of 96 from the central camera, resulting in a total of 3,341,163 patches for the training set and 382,445 patches for the test set. Patches with the same spatial position within one image and 49 different focal depths constitute a single focal stack. Please refer to [16] for additional details about the dataset, and refer to Sec. 6 in our supplementary for the labeling method adopted in this paper.", + "bbox": [ + 89, + 537, + 483, + 763 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation metric. We follow the evaluation metric proposed by Scharstein et al. [34], which is used for Middlebury stereo dataset and adopted by Herrmann et al. [16] and Choi et al. [7] to evaluate the AF model's prediction accuracy. In particular, this metric is calculated by the proportion of patches whose stop positions are within 0, 1, 2, and 4 indices away from the in-focus index, which can be considered in the depth of field (DoF). The mean absolute error (MAE) and root mean squared error (RMSE) are also", + "bbox": [ + 89, + 765, + 483, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "used to evaluate the AF model's performance. To evaluate the focus hunting (FH), we retain every focusing trajectory and calculate the direction and distance of each lens movement. If the direction of a current movement is opposite to that of the previous one in the trajectory, we consider an occurrence of FH in that trajectory. Since the actor network obtained by our DRL and the model in [7] share the same architecture and similar input/output, we can simply regard the number of steps to reach the in-focus position as the focus speed. We calculate the images captured at all initial positions of the lens to account for accuracy and FH rate.", + "bbox": [ + 511, + 90, + 903, + 257 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation detail. We use PyTorch [29] and TorchRL [4] for our implementation. For deep learning models, we use the Adam optimizer [23] with a learning rate of $10^{-3}$ , $(\\beta_{1},\\beta_{2}) = (0.5,0.999)$ , a batch size of 128, and cosine learning rate decay [26] for 10k iterations. For reinforcement learning models, we adopt the PPO-CLIP [35] algorithm. The initial learning rate is set to $1\\times 10^{-5}$ , and the clip epsilon is 0.2, a batch size of 32, Adam optimizer with the same $(\\beta_{1},\\beta_{2})$ as deep learning. The maximum episode length of RL is set to 4. For the expert trajectory, we first construct an expert dataset containing 6,000,000 expert trajectories and set the weight $\\lambda$ to 0.001 in the loss function.", + "bbox": [ + 511, + 258, + 906, + 439 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Quantitative Results", + "text_level": 1, + "bbox": [ + 511, + 452, + 705, + 468 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "FH rate and multi-step prediction accuracy (Effects of DRL). The FH rate and multi-step prediction accuracy of comparison methods are shown in Tab. 1. Though the relative movement label can significantly improve the one-step prediction accuracy, the FH rate in the multi-step setting is still high since the current deep-learning training process cannot obtain and exploit the sequential information of the lens movement. Due to DRL's advantages in modeling sequential decision-making problems, our DRL-based models can directly decrease the FH rate without relying on any extra networks, such as the LSTM [36]. Compared with the policy obtained by deep learning, our DRL-based policy can decrease the FH rate by $9.1\\%$ in 2-step prediction, $16.5\\%$ in 3-step prediction, and $18.0\\%$ in 4-step prediction, while adopting the same network architecture. Our work shows some advantages over existing RL baselines. Since few features are extracted from the scalar input of [6], it can only guide the lens towards one side with the higher label frequency, causing extremely low accuracy but 0 FH rate. Since [54] generalizes poorly in various views (as stated by the authors), while the dataset in our work contains images with multiple views and object distances, it hardly gives a move signal. The lens thus stops at the start position in most cases, causing small changes in accuracy and a low FH rate.", + "bbox": [ + 511, + 477, + 905, + 839 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For the multi-step prediction performance, the accuracy of the baseline [7] increases a little and even decreases in 4-step prediction. The accuracy of the model obtained by the relative movement label can achieve substantial gains", + "bbox": [ + 511, + 840, + 903, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "26445", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/e51e2485f25370ef97a3ca3a10a83f5eb18538964f71ab0ea3b0e98c95af7b30.jpg", + "table_caption": [ + "Table 1. Prediction accuracy and focus hunting (FH) rate results in multi-step setting." + ], + "table_footnote": [], + "table_body": "
AlgorithmStep= 0 ↑ ≤ 1 ↑ ≤ 2 ↑ ≤ 4 ↑MAE↓ RMSE↓FH↓
DRL in [6]10.030 0.071 0.117 0.21014.216 17.908NA
20.029 0.050 0.082 0.15117.214 20.8270.0
30.029 0.050 0.082 0.15117.214 20.8270.0
40.029 0.050 0.082 0.15117.214 20.8270.0
DRL in [54]10.022 0.062 0.102 0.18115.575 19.143NA
20.023 0.064 0.105 0.18515.446 19.0510.019
30.024 0.065 0.106 0.18715.437 19.0730.025
40.024 0.065 0.106 0.18615.500 19.1420.030
RoI-PE & Len-PE [7]10.243 0.584 0.756 0.8822.839 6.880NA
20.275 0.640 0.803 0.9122.454 6.2960.275
30.279 0.644 0.810 0.9142.434 6.2050.362
40.276 0.642 0.807 0.9132.442 6.2770.383
[7] with relative label (Ours)10.330 0.698 0.826 0.9072.342 6.339NA
20.413 0.795 0.888 0.9341.894 5.8550.261
30.427 0.805 0.894 0.9371.892 5.7210.340
40.423 0.804 0.893 0.9361.862 5.8440.358
Expert Trajectory (Ours)10.336 0.711 0.839 0.9152.146 5.827NA
20.439 0.808 0.896 0.9411.755 5.5970.170
30.450 0.816 0.900 0.9431.711 5.5320.175
40.449 0.818 0.901 0.9441.694 5.4790.178
", + "bbox": [ + 91, + 111, + 480, + 327 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/31051fadf2aef18550018aa2c19ea3501f9d41623db343f197147a8c3aec3d34.jpg", + "table_caption": [ + "Table 2. Average steps to stop with different termination signals." + ], + "table_footnote": [], + "table_body": "
Termination Rule= 0 ↓ ≤ 1 ↓ ≤ 2 ↓ ≤ 4 ↓Stop Model↓ MAE↓ RMSE↓
RoI-PE & Lens-PE[7]6.6513.6442.4771.7072.2082.6116.146
[7] with relative label (Ours)4.9822.2461.7041.4531.8841.9065.463
Expert Trajectory (Ours)5.0142.2571.6831.3611.8211.8745.332
", + "bbox": [ + 94, + 351, + 475, + 395 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "of $13.8\\%$ prediction accuracy for exact $(= 0)$ predictions in 2-step prediction, while our DRL-based model further enhances this accuracy. Furthermore, as the number of lens movement steps increases, the accuracy of deep learning-based methods may decline, whereas the accuracy of our DRL-based method remains unchanged or even improves, demonstrating the stability of our DRL-based model in multi-step prediction. Moreover, Tab. 2 shows that our models take fewer steps to stop at different termination signals, e.g., $= 0$ , $\\leq 1$ , $\\leq 2$ , and $\\leq 4$ . Since in-focus positions remain unknown in practical camera deployment, the above rules may not be feasible. Alternatively, we train a stop model to generate the stop signal, which is modeled as a binary classification task and takes the same architecture and input as our policy network. Tab. 2 shows the average steps, MAE, and RMSE of different policies with this model. Our method still achieves a better prediction with fewer steps. In conclusion, results in Tab. 1 and Tab. 2 show that our methods can achieve higher accuracy with fewer steps than that proposed by [7], demonstrating a faster focusing speed.", + "bbox": [ + 89, + 400, + 482, + 704 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Single-step accuracy (Effects of relative movement as labels). The comparison results in the single-step permutation setting are shown in Tab. 3. Compared with classic statistics-based or depth estimation methods, the deep-learning based models \"Learning to Autofocus [16]\" and \"RoI-PE & Len-PE [7]\" significantly improve the one-step prediction accuracy. By modifying the output of the model in [7] from the absolute lens position to the relative movement, we can further improve the accuracy by $8.7\\%$ for exact $(= 0)$ prediction and $2.5\\%$ for $\\leq 4$ focal index difference. Moreover, since the dataset has low-contrast images with similar features but distant labels, minimizing the loss makes the policy perform badly. Via optimizing multi-step", + "bbox": [ + 89, + 704, + 482, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/13dd255546095cb21baed0f364d83c3b3ee7a87246096e2579aad69f1998f56c.jpg", + "table_caption": [ + "Table 3. Single-step prediction results." + ], + "table_footnote": [], + "table_body": "
Algorithm= 0 ↑≤ 1 ↑≤ 2 ↑≤ 4 ↑MAE↓RMSE↓
ZNCC Disparity with Calibration0.0650.1790.2680.4199.47614.211
SSD Disparity [44]0.1010.2640.3860.5258.01212.414
Learning the Depth [11]0.1120.2970.4150.5627.41912.312
Learning to Autofocus [16]0.1680.4670.6590.8513.2497.619
RoI-PE & Lens-PE [7]0.2430.5840.7560.8822.8396.880
[16] with relative label (Ours)0.1810.4930.6840.8613.0217.374
[7] with relative label(Ours)0.3300.6980.8260.9072.3426.339
Expert Trajectory (Ours)0.3360.7110.8390.9152.1465.827
", + "bbox": [ + 517, + 97, + 900, + 191 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/b0df48b1e4ead96273fa26d727316242a72daf32277aa2adc7bccce3cd3f8981.jpg", + "table_caption": [ + "Table 4. Performance comparison of DRL models with and without Expert Regularization." + ], + "table_footnote": [], + "table_body": "
AlgorithmStep= 0 ↑ ≤ 1 ↑ ≤ 2 ↑ ≤ 4 ↑MAE↓ RMSE↓FH↓
w/o10.0310.0560.0860.14322.00526.048
20.0290.0650.0950.14823.84527.975
Expert Regularization30.0370.0850.1230.18820.95725.501
40.0310.0680.0990.15223.72227.889
Expert Trajectory (Ours)10.3360.7110.8390.9152.1465.827
20.4390.8080.8960.9411.7555.597
30.4500.8160.9000.9431.7115.532
40.4490.8180.9010.9441.6945.479
", + "bbox": [ + 524, + 226, + 893, + 321 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "behavior, RL can explore extra information and improve 1-step prediction accuracy.", + "bbox": [ + 511, + 324, + 903, + 354 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effects of expert regularization. As shown in Tab. 4, the policy obtained by the vanilla DRL algorithm without expert regularization has a poor performance, since the action space is a discrete number of 97, which is large for DRL's training. The expert trajectory significantly improves the performance of the policy given by the PPO algorithm.", + "bbox": [ + 511, + 354, + 903, + 446 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Qualitative Results", + "text_level": 1, + "bbox": [ + 513, + 454, + 694, + 470 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Fig. 1 qualitatively visualizes the trajectory of patches and lens positions of different comparison methods, including the policy obtained by [7], the policy obtained by relative movement label, and our DRL-based model. Compared with directly predicting the lens position [7], our relative movement label model is faster by predicting the lens movement within DoF in the second step, but the FH issue cannot be overcome. In comparison, our DRL-based model can generate a faster and more stable prediction by predicting the GT location at the first step and stopping at this index. The lens trajectory of the DRL-based model (the green curve in Fig. 2) is slightly smoother and more stable without the occurrence of FH.", + "bbox": [ + 511, + 477, + 905, + 672 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In Fig. 6, we show qualitative results of 4-step predictions of the three models. From the first row, we can observe that all methods can predict the index within DoF, but our methods can give a more precise prediction with a sharper image than Choi's [7]. From the second to fourth rows, we can observe that [7] sometimes breaks and predicts indices with sharp background, while our DRL-based model successfully predicts indices with sharp target objects. This case is a generally challenging scenario in the AF task since many different DoFs can be estimated. These DoFs may confuse the models, leading to a focus on the wrong object. In general, our relative movement label can give a more accurate prediction by mapping the disparity to the movement of the lens, and the DRL-based model can further enhance the accuracy by estimating the future reward of the action", + "bbox": [ + 511, + 674, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "26446", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/189850b0d961d43ad0fee67ed1398510478dcd6c775f823cf628308068b84b05.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 94, + 82, + 161, + 136 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7ef550bb2c179357b8f414db0273bc6e558ed66aee57c8d1f98e4ef91e791643.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 94, + 137, + 161, + 191 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/10775c45e94d5fdf3bf7cd35da086e3acb78975b9234083d280a2b8a4e127c7b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 94, + 191, + 161, + 244 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/89e439505bb9e65d78ff78db058b701e14f2bcae2c6524991ebd201afadacf45.jpg", + "image_caption": [ + "Input", + "Figure 6. Comparison of the 4-step results of the policy obtained by [7], our relative movement label, and our DRL-based model." + ], + "image_footnote": [], + "bbox": [ + 94, + 247, + 161, + 300 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/78f6eece17673bd4dd2c22ce6c3074b32640f19a513a6426179e138f10f064f4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 171, + 82, + 240, + 136 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b82c6818fc24ce9026b74b5f761c40a0fd2625c22ebdc4e9bdf22c3c95d56232.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 171, + 137, + 240, + 191 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ea241c0050c70f1f2808b8fe0ad410f67a9769b65aaa20932efdc68a895e1902.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 171, + 191, + 240, + 244 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/bf2fe82be806fc3c087b64d8d6ff8ec3e252dd80601d8e358b8e69828ae4cbb0.jpg", + "image_caption": [ + "absolute position labe" + ], + "image_footnote": [], + "bbox": [ + 171, + 247, + 240, + 299 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/efafaabf2fa673ff7cc6bd3175916ff61936d57e1b0fff569d99ed395e28e96c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 251, + 82, + 320, + 136 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4b8cd13827cfd0c97b2b318dd89c3cddc8ea7552c92a67969fcfeee0aaed2526.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 251, + 137, + 320, + 191 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/eefd1c948dfb5ffecb4862cbf822041a769d9a2f22a7e0d666b8a2cc4e95ce68.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 251, + 191, + 320, + 244 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/fde21e27e4ef700bf2ed5ed3d8139c3a2804b3b86cf0a484b005881104cfef45.jpg", + "image_caption": [ + "relative movement label" + ], + "image_footnote": [], + "bbox": [ + 251, + 247, + 320, + 299 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d3c97875496f60449e9e4f77cea564696ebceb14b2186a12b8b80ec1fe6a8eab.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 331, + 82, + 398, + 136 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/857994789c5078b95175cc548f5a0834d53b1f55f0076d34a3d3a359d2ea8664.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 331, + 137, + 398, + 191 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/96d52a5abbf949dc102f24a0e067b608f2ba4dfb926f9fc923e2c76b1f8cc5bd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 331, + 191, + 397, + 244 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/467450180faa395da3e2eb6560da926580676ff3e1e4441175731705dce5e67f.jpg", + "image_caption": [ + "DRL" + ], + "image_footnote": [], + "bbox": [ + 331, + 247, + 397, + 299 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0ad31ed71439c73d490282239f8ce6e06935e40d55b7903e2b878bb3525e7b7b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 408, + 82, + 475, + 136 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2ab5c69b8f5de88f2d319a32fd4a8279eb7e953672395b388e51fb2b27f871df.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 410, + 137, + 475, + 191 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b8768a93885661caf0cb400a7a38f503ddc89590a6bc785298a856f12178f9a8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 410, + 191, + 475, + 244 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/86b1d4250541142213e2df685a1f15c842c7313c19d2acebe97fbce72561f8c8.jpg", + "image_caption": [ + "GT" + ], + "image_footnote": [], + "bbox": [ + 408, + 247, + 475, + 299 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "and reducing the focus hunting rate.", + "bbox": [ + 89, + 354, + 328, + 369 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In Fig. 7, we analyze our model's failure cases: (a) it is difficult and meaningless to focus on the images with no texture since the phase information or contrast information is little to estimate the depth, (b) though our models can give a more accurate prediction in scenarios with multiple DoFs, when the target object is small and the background is complex, the models may be confused by these DoFs, leading to focusing on the wrong object, (c) like all DP data-based AF methods, our models may work poorly when the objects are horizontal. Some possible solutions may work to overcome these limitations. For (a), larger patches can provide more phase information, and we also think (a) might not be that important since few people take photos with no texture. For (b), we claim it is not a bad issue, because sometimes people may want to focus on the background obscured by meaningless foreground elements. To solve this problem, larger patches with more information about the target objects may be useful. Some depth estimation methods [22, 52] can also be used to provide more depth information to help distinguish the objects and background. For (c), solving this problem is challenging in algorithms, while little changes in hardware, especially sensors, could easily mitigate this issue [7].", + "bbox": [ + 89, + 369, + 483, + 717 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 726, + 243, + 742 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effects of relative label method. Tab. 5 further demonstrates the effect of relative movement as labels. Our DRL-based model can further enhance the accuracy by $0.6\\%$ for exact $(= 0)$ prediction and $0.8\\%$ for $\\leq 4$ focal index difference, in comparison to the model in [7] modified with the relative movement label. We further compare the two label methods under three different architectures: MobileNetV2 [33], ResNet50 [15], and MnasNet [41]. The relative movement label method outperforms the absolute lens position label for each architecture. For a fairer comparison,", + "bbox": [ + 89, + 750, + 482, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/41f4ba9be68136a6609bde276a95e00abe544c901bb98dd5d3da776f3231a3a7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 514, + 79, + 647, + 138 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a8df26c18e23a57b065915fd84b4a91fcf1d37d148b5b09cbcf383b753c72c37.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 516, + 141, + 645, + 199 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/876be92ae60c1986ec51e9bb29655bb9209f2bdafc61c731b42d68f7940ff435.jpg", + "image_caption": [ + "Input", + "Figure 7. Failure cases of our models, where only the left image of DP data is visualized." + ], + "image_footnote": [], + "bbox": [ + 516, + 200, + 645, + 260 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3c0fb69380e47e9ddba384a23828fd7424aa366bfbffb1d5a684797ebcc48a87.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 653, + 79, + 732, + 138 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c1f40623d2d637e11ecf939f88911970f2e5c0a32aace19802722016f9604c9e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 653, + 140, + 732, + 199 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9a25ae0cd0e8e9778aa4007d1f036074b6b133749f7dff5dc13e1ca43fa664af.jpg", + "image_caption": [ + "relative movement label" + ], + "image_footnote": [], + "bbox": [ + 653, + 200, + 732, + 260 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b0dc5b5557066efbeb75091808f0eeb0a4e2c71116d6eb274e59ba5ea3059ba5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 738, + 79, + 816, + 138 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/fd94b8464e74e9842e7df2e9f7c6e616b7df34c862bdee3d979d4cfea7bc5620.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 738, + 140, + 816, + 199 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/da7145b1c0ae1c4e6e240e953efa5b4a104a0134b22a66292415e251c55142b8.jpg", + "image_caption": [ + "DRL" + ], + "image_footnote": [], + "bbox": [ + 738, + 200, + 816, + 260 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/198eadb473dab9abd3b30e2046853b3f12eedb4fb0a39e1a0562a6214deac023.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 823, + 79, + 901, + 138 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5bbeb3708e3d335caad1d7198299f0958d57b97c85048ec9876e845a185b2ce5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 823, + 140, + 901, + 199 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0b567bd653370c2878ce35fdbe574a2b7399da2794508b30548a650fdf25dd1a.jpg", + "image_caption": [ + "GT" + ], + "image_footnote": [], + "bbox": [ + 823, + 200, + 901, + 260 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/8e1a0a0c0d6795dec945f013142c2a9d3622af9a199ab1218bf05858dfa190c5.jpg", + "table_caption": [ + "Table 5. Single-step prediction results with different network architectures. The number of input channels is 5." + ], + "table_footnote": [], + "table_body": "
Algorithm# of paramslabel= 0 ↑ ≤ 1 ↑ ≤ 2 ↑ ≤ 4 ↑MAE↓RMSE↓
MobileNetV2 [33]2.29Mabsolute0.2430.5840.7560.8822.8396.880
2.35Mabsolute0.2520.5920.7640.8882.7196.745
2.35Mrelative0.3300.6980.8260.9072.3426.339
ResNet50 [15]23.61Mabsolute0.2580.6020.7890.8942.6156.623
23.71Mabsolute0.2540.5960.7810.8912.6556.686
23.71Mrelative0.3260.7090.8490.9182.1365.764
MnasNet [41]1.00Mabsolute0.2260.5680.7360.8642.9837.016
1.06Mabsolute0.2130.5490.7180.8503.1197.318
1.06Mrelative0.3050.6780.8170.9052.4366.402
", + "bbox": [ + 514, + 349, + 911, + 481 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "we expand the absolute position label to the same range as the relative movement label. It shows that expanding the number of networks' output channels can not improve the accuracy, and the effects remain unknown. Since Lens-PE mainly conveys the lens' current position to the network, while the next position must be predicted by adding its movement to the current position, the relative label methods avoid additionally inferring the identity mapping for Lens-PE, leading to a higher accuracy. To make a fair comparison with the other learning-based models, we thus adopt the same MobileNetV2 architecture in the experiments.", + "bbox": [ + 511, + 489, + 906, + 656 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 676, + 633, + 691 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We have proposed a DRL-based autofocus model with expert trajectory regularization that can both improve the prediction accuracy and reduce the focus hunting rate, which took an actor-critic framework based on the pretrained MobileNet-V2 architecture for single-step prediction. For the actor network, we further modified the absolute position label in [7, 16] with the relative movement label method. We also constructed an offline dataset of expert trajectories based on prior knowledge, regularizing the policy network training toward reducing focus hunting. Experimental results have clearly demonstrated the effectiveness of our proposed method, in terms of the multi-step prediction accuracy and the occurrence of focus hunting.", + "bbox": [ + 511, + 703, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "26447", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 91, + 90, + 250, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 244, + 187, + 258 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Abdullah Abuolaim, Mauricio Delbracio, Damien Kelly, Michael S. Brown, and Peyman Milanfar. Learning to reduce defocus blur by realistically modeling dual-pixel data. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 2289-2298, 2021. 1, 3", + "[2] Yuhan Bao, Lei Sun, Yuqin Ma, Diyang Gu, and Kaiwei Wang. Improving fast auto-focus with event polarity. Optics Express, 31(15):24025-24044, 2023. 1", + "[3] Daniel I Barnea and Harvey F Silverman. A class of algorithms for fast digital image registration. IEEE transactions on Computers, 100(2):179-186, 1972. 1, 2", + "[4] Albert Bou, Matteo Bettini, Sebastian Dittert, Vikash Kumar, Shagun Sodhani, Xiaomeng Yang, Gianni De Fabritis, and Vincent Moens. TorchRL: A data-driven decision-making library for pytorch. arXiv preprint arXiv:2306.00577, 2023. 6", + "[5] Serkan Cabi, Sergio Gómez Colmenarejo, Alexander Novikov, Ksenia Konyushkova, Scott Reed, Rae Jeong, Konrad Zolna, Yusuf Aytar, David Budden, Mel Vecerik, et al. Scaling data-driven robotics with reward sketching and batch reinforcement learning. arXiv preprint arXiv:1909.12200, 2019. 5", + "[6] Chin-Cheng Chan and Homer H Chen. Autofocus by deep reinforcement learning. *Electronic Imaging*, 31:1-6, 2019. 3, 6, 7", + "[7] Myungsub Choi, Hana Lee, and Hyong-Euk Lee. Exploring positional characteristics of dual-pixel data for camera autofocus. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 13112-13122, 2023. 2, 3, 4, 5, 6, 7, 8, 1", + "[8] Murtaza Dalal, Deepak Pathak, and Russ R Salakhutdinov. Accelerating robotic reinforcement learning via parameterized action primitives. In Advances in Neural Information Processing Systems (NeurIPS), pages 21847-21859, 2021. 2", + "[9] Judith Dijk, Michael van Ginkel, Rutger J van Asselt, Lucas J van Vliet, and Piet W Verbeek. A new sharpness measure based on gaussian lines and edges. In 10th International Conference on Computer Analysis of Images and Patterns (CAIP), pages 149-156, 2003. 3", + "[10] Raul Diaz and Amit Marathe. Soft labels for ordinal regression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4733-4742, 2019. 5, 2", + "[11] Rahul Garg, Neal Wadhwa, Sameer Ansari, and Jonathan T Barron. Learning single camera depth estimation using" + ], + "bbox": [ + 93, + 268, + 483, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "This work was supported in part by the National Natural Science Foundation of China under Grant 62125109, Grant 62431017, Grant U24A20251, Grant 62320106003, Grant 62371288, Grant 62401357, Grant 62401366, Grant 62301299, Grant 62120106007, and in part by the Program of Shanghai Science and Technology Innovation Project under Grant 24BC3200800. (Corresponding authors: Chenglin Li, Nuowen Kan, Junni Zou.)", + "dual-pixels. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 7628-7637, 2019. 3, 7, 1", + "[12] Cole Gulino, Justin Fu, Wenjie Luo, George Tucker, Eli Bronstein, Yiren Lu, Jean Harb, Xinlei Pan, Yan Wang, Xiangyu Chen, John Co-Reyes, Rishabh Agarwal, Rebecca Roelofs, Yao Lu, Nico Montali, Paul Mougin, Zoey Yang, Brandyn White, Aleksandra Faust, Rowan McAllister, Dragomir Anguelov, and Benjamin Sapp. Waymax: An accelerated, data-driven simulator for large-scale autonomous driving research. In Advances in Neural Information Processing Systems (NeurIPS), pages 7730-7742, 2023. 2", + "[13] Marsha Jo Hannah. Computer matching of areas in stereo images. Stanford University, 1974. 1, 2", + "[14] Jie He, Rongzheng Zhou, and Zhiliang Hong. Modified fast climbing search auto-focus algorithm with adaptive step size searching technique for digital camera. IEEE Trans. Consumer Electron., 49(2):257-262, 2003. 3", + "[15] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 770-778, 2016. 8", + "[16] Charles Herrmann, Richard Strong Bowen, Neal Wadhwa, Rahul Garg, Qiurui He, Jonathan T. Barron, and Ramin Zabih. Learning to autofocus. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2227-2236, 2020. 2, 3, 6, 7, 8, 1", + "[17] Chi-Jui Ho, Chin-Cheng Chan, and Homer H Chen. A net: A convolutional neural network approach to phase detection autofocus. IEEE Transactions on Image Processing, 29:6386-6395, 2020. 3, 5, 1", + "[18] Jui-Ting Huang, Chun-Hung Shen, See-May Phoong, and Homer Chen. Robust measure of image focus in the wavelet domain. In 2005 International Symposium on Intelligent Signal Processing and Communication Systems, pages 157-160, 2005. 3", + "[19] ZHENG Jin, JIANG Botao, PENG Wei, and ZHANG Qiaohui. Multi-scale binocular stereo matching based on semantic association. Chinese Journal of Electronics, 33(4):1010-1022, 2024. 3", + "[20] Junpeng Jing, Jiankun Li, Pengfei Xiong, Jiangyu Liu, Shuaicheng Liu, Yichen Guo, Xin Deng, Mai Xu, Lai Jiang, and Leonid Sigal. Uncertainty guided adaptive warping for robust and efficient stereo matching. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 3318-3327, 2023. 3", + "[21] Neel Joshi and C Lawrence Zitnick. Micro-baseline stereo. Microsoft Research Technical Report, 2014. 3", + "[22] Donggun Kim, Hyeonjoong Jang, Inchul Kim, and Min H. Kim. Spatio-focal bidirectional disparity estimation from a dual-pixel image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5023-5032, 2023. 3, 8", + "[23] Diederik P Kingma. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6", + "[24] Matej Kristan, Janez Pers, Matej Perse, and Stanislav Kovačić. A bayes-spectral-entropy-based measure of camera" + ], + "bbox": [ + 93, + 92, + 906, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "26448", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "focus using a discrete cosine transform. Pattern Recognition Letters, 27(13):1431-1439, 2006. 3", + "[25] Sang-Yong Lee, Jae-Tack Yoo, Yogendera Kumar, and SooWon Kim. Reduced energy-ratio measure for robust autofocusing in digital camera. IEEE Signal Processing Letters, 16 (2):133-136, 2009. 3", + "[26] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983, 2017. 6", + "[27] Yao Lu, Karol Hausman, Yevgen Chebotar, Mengyuan Yan, Eric Jang, Alexander Herzog, Ted Xiao, Alex Irpan, Mohi Khansari, Dmitry Kalashnikov, et al. Aw-opt: Learning robotic skills with imitation and reinforcement at scale. In Conference on Robot Learning, pages 1078-1088, 2022. 5", + "[28] Ashvin Nair, Abhishek Gupta, Murtaza Dalal, and Sergey Levine. Awac: Accelerating online reinforcement learning with offline datasets. arXiv preprint arXiv:2006.09359, 2020. 5", + "[29] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Kopf, Edward Yang, Zachary DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. Pytorch: An imperative style, high-performance deep learning library. In Advances in Neural Information Processing Systems (NeurIPS), 2019. 6", + "[30] José Luis Pech-Pacheco, Gabriel Cristóbal, Jesús Chamorro-Martinez, and Joaquín Fernández-Valdivia. Diatom autofocusing in brightfield microscopy: a comparative study. In Proceedings 15th International Conference on Pattern Recognition. ICPR-2000, pages 314-317. IEEE, 2000. 1", + "[31] Henry Pinkard, Zachary Phillips, Arman Babakhani, Daniel A Fletcher, and Laura Waller. Deep learning for single-shot autofocus microscopy. Optica, 6(6):794-797, 2019. 3", + "[32] Stéphane Ross, Geoffrey Gordon, and Drew Bagnell. A reduction of imitation learning and structured prediction to no-regret online learning. In Proceedings of the fourteenth international conference on artificial intelligence and statistics, pages 627-635. JMLR Workshop and Conference Proceedings, 2011. 4", + "[33] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zhmoginov, and Liang-Chieh Chen. *Mobilenetv2: Inverted residuals and linear bottlenecks*. In *Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition* (CVPR), pages 4510-4520, 2018. 2, 5, 6, 8", + "[34] Daniel Scharstein and Richard Szeliski. A taxonomy and evaluation of dense two-frame stereo correspondence algorithms. International Journal of Computer Vision, 47:7-42, 2002. 6", + "[35] John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347, 2017. 6", + "[36] Xingjian SHI, Zhourong Chen, Hao Wang, Dit-Yan Yeung, Wai-kin Wong, and Wang-chun WOO. Convolutional LSTM Network: A Machine Learning Approach for Precipitation" + ], + "bbox": [ + 91, + 92, + 482, + 901 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Nowcasting. In Advances in Neural Information Processing Systems (NeurIPS), 2015. 6", + "[37] Przemysław Sliwinski and Paweł Wachel. A simple model for on-sensor phase-detection autofocusing algorithm. Journal of Computer and Communications, 1(06):11, 2013. 3", + "[38] Yuda Song, Yifei Zhou, Ayush Sekhari, Drew Bagnell, Akshay Krishnamurthy, and Wen Sun. Hybrid RL: Using both offline and online data can make RL efficient. In International Conference on Learning Representations (ICLR), 2023. 5", + "[39] Changqing Su, Zhiyuan Ye, Yongsheng Xiao, You Zhou, Zhen Cheng, Bo Xiong, Zhaofei Yu, and Tiejun Huang. Intensity-robust autofocus for spike camera. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 25018-25027, 2024. 3", + "[40] Murali Subbarao, Tae-Sun Choi, and Arman Nikzad. Focusing techniques. Optical Engineering, 32(11):2824-2836, 1993. 1, 3", + "[41] Mingxing Tan, Bo Chen, Ruoming Pang, Vijay Vasudevan, Mark Sandler, Andrew Howard, and Quoc V. Le. Mnasnet: Platform-aware neural architecture search for mobile. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2815-2823, 2019. 8", + "[42] Jay Martin Tenenbaum. Accommodation in computer vision. Stanford University, 1971. 1, 3", + "[43] Quoc Kien Vuong and Jeong-won Lee. Initial direction and speed decision system for auto focus based on blur detection. In 2013 IEEE International Conference on Consumer Electronics (ICCE), pages 222-223, 2013. 1", + "[44] Neal Wadhwa, Rahul Garg, David E Jacobs, Bryan E Feldman, Nori Kanazawa, Robert Carroll, Yair Movshovitz-Attias, Jonathan T Barron, Yael Pritch, and Marc Levoy. Synthetic depth-of-field with a single-camera mobile phone. ACM Transactions on Graphics (ToG), 37(4):1-13, 2018. 7", + "[45] Andrew Wagenmaker and Aldo Pacchiano. Leveraging offline data in online reinforcement learning. arXiv preprint arXiv:2211.04974, 2023. 5", + "[46] Chengyu Wang, Qian Huang, Ming Cheng, Zhan Ma, and David J. Brady. Deep learning for camera autofocus. IEEE Transactions on Computational Imaging, 7:258-271, 2021. 3", + "[47] Hui Xie, Weibin Rong, and Lining Sun. Wavelet-based focus measure and 3-d surface reconstruction method for microscopy images. In Proceedings 2006 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS 2006), pages 229-234, 2006. 3", + "[48] Takahiro Yamasaki, Tomohiro Nakamura, Ryohei Funatsu, and Hiroshi Shimamoto. Hybrid autofocus system by using a combination of the sensor-based phase-difference detection and focus-aid signal. In 2018 IEEE International Conference on Consumer Electronics (ICCE), pages 1–2, 2018. 1", + "[49] Jingyang Yan, Peter DiMeo, Lu Sun, and Xian Du. LSTM-Based Model Predictive Control of Piezoelectric Motion Stages for High-Speed Autofocus. IEEE Transactions on Industrial Electronics, 70(6):6209-6218, 2023. 3", + "[50] Ge Yang and Bradley J Nelson. Wavelet-based autofocusing and unsupervised segmentation of microscopic images." + ], + "bbox": [ + 516, + 92, + 903, + 901 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "26449", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "In Proceedings 2003 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS 2003), pages 2143-2148, 2003. 3", + "[51] Hyun Jong Yang, Moohyun Oh, Jonggyu Jang, Hyeonsu Lyu, and Junhee Lee. Robust deep-learning based autofocus score prediction for scanning electron microscope. Microscopy and Microanalysis, 26(S2):702-705, 2020. 3", + "[52] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10371-10381, 2024. 8", + "[53] Yi Yao, Besma Abidi, Narjes Doggaz, and Mongi Abidi. Evaluation of sharpness measures and search algorithms for the auto-focusing of high-magnification images. In Visual Information Processing XV, pages 132-143. SPIE, 2006. 3", + "[54] Xiaofan Yu, Runze Yu, Jingsong Yang, and Xiaohui Duan. A robotic auto-focus system based on deep reinforcement learning. In 2018 15th International Conference on Control, Automation, Robotics and Vision (ICARCV), pages 204-209, 2018. 3, 6, 7", + "[55] Ramin Zabih and John Woodfill. Non-parametric local transforms for computing visual correspondence. In Computer Vision—ECCV'94: Third European Conference on Computer Vision Stockholm, Sweden, May 2–6 1994 Proceedings, Volume II 3, pages 151–158. Springer, 1994. 1, 2" + ], + "bbox": [ + 91, + 90, + 483, + 458 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "26450", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Stabilizing and Accelerating Autofocus with Expert Trajectory Regularized Deep Reinforcement Learning/a699e1d4-e683-4371-bc56-144538826827_model.json b/2025/Stabilizing and Accelerating Autofocus with Expert Trajectory Regularized Deep Reinforcement Learning/a699e1d4-e683-4371-bc56-144538826827_model.json new file mode 100644 index 0000000000000000000000000000000000000000..dbc60a9c78bf567d9c69603f77f19c85a18ef8e7 --- /dev/null +++ b/2025/Stabilizing and Accelerating Autofocus with Expert Trajectory Regularized Deep Reinforcement Learning/a699e1d4-e683-4371-bc56-144538826827_model.json @@ -0,0 +1,2653 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.237, + 0.001, + 0.812, + 0.047 + ], + "angle": 0, + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.12, + 0.13, + 0.88, + 0.178 + ], + "angle": 0, + "content": "Stabilizing and Accelerating Autofocus with Expert Trajectory Regularized Deep Reinforcement Learning" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.203, + 0.847, + 0.24 + ], + "angle": 0, + "content": "Shouhang Zhu1, Chenglin Li1, Yuankun Jiang2, Li Wei1, Nuowen Kan1, Ziyang Zheng1, Wenrui Dai2, Junni Zou2, Hongkai Xiong1" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.239, + 0.78, + 0.257 + ], + "angle": 0, + "content": "\\(^{1}\\)Department of Electronic Engineering, Shanghai Jiao Tong University" + }, + { + "type": "text", + "bbox": [ + 0.169, + 0.256, + 0.829, + 0.275 + ], + "angle": 0, + "content": "\\(^{2}\\)Department of Computer Science and Engineering, Shanghai Jiao Tong University" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.276, + 0.847, + 0.293 + ], + "angle": 0, + "content": "\\(^{1}\\)shzhu2001, lcl1985, wei.li, kannw.1230, zhengziyang, xionghongkai}@sjtu.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.285, + 0.293, + 0.708, + 0.31 + ], + "angle": 0, + "content": "\\(^{2}\\) \\(\\{\\text { yuankunjiang, daiwenrui, zoujunni }\\} @\\)sjtu.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.345, + 0.327, + 0.361 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.377, + 0.486, + 0.786 + ], + "angle": 0, + "content": "Autofocus is a crucial component of modern digital cameras. While recent learning-based methods achieve state-of-the-art in focus prediction accuracy, they unfortunately ignore the potential focus hunting phenomenon of back-and-forth lens movement in the multi-step focusing procedure. To address this, in this paper, we propose an expert regularized deep reinforcement learning (DRL)-based approach for autofocus, which can utilize the sequential information of lens movement trajectory to both enhance the multi-step in-focus prediction accuracy and reduce the chance of focus hunting. Our method generally follows an actor-critic framework. To accelerate the DRL's training with a higher sample efficiency, we initialize the policy with a pre-trained single-step prediction network, where the network is further improved by modifying the output of absolute in-focus position distribution to the relative lens movement distribution to establish a better mapping between input images and lens movement. To further stabilize DRL's training with a lower occurrence of focus hunting in the resulting lens movement trajectory, we generate some offline trajectories based on prior knowledge to avoid focus hunting, which are then leveraged as an offline dataset of expert trajectories to regularize the actor network's training. Empirical evaluations show that our method outperforms those learning-based methods on public benchmarks, with higher single- and multi-step prediction accuracy, and a significant reduction of focus hunting rate." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.801, + 0.222, + 0.816 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.485, + 0.903 + ], + "angle": 0, + "content": "Autofocus (AF) is a process through which the camera can determine automatically a lens position to focus on the region of interest (RoI). It is a crucial step in image capturing, since once an image is already captured in a defocused and blurry state, it is then difficult to restore its sharpness" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.346, + 0.908, + 0.68 + ], + "angle": 0, + "content": "through any post-processing due to the varying effects of blur influenced by factors such as camera aperture and radial distortion [1]. In addition, the beauty of a moment often lies in its fleeting nature, which thus imposes an imperative demand for the fast focusing speed of an AF algorithm. However, due to the instability in focusing systems incurred by some interference factors, the AF mechanism of the camera lens may fail to accurately lock onto the focal point, resulting in a repeated movement of the lens within a certain range. This phenomenon, as illustrated in Fig. 1, is referred to as focus hunting (FH), whereby the camera is unable to identify the focus plane and thus drives the lens forward and backward through a set range of focus hunting for it. It then causes variations in the image sharpness and changes the lens's field of view (FoV), ultimately leading to an unstable focusing experience with longer focusing time, which is a common issue in the traditional AF-assisted cameras [2]. It is thus of great importance to develop an AF algorithm that not only achieves a higher in-focus accuracy but can also stabilize the trajectory of lens movement with a faster convergence speed to the in-focus position, thus reducing the probability of occurrence of focus hunting [43, 48]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.69, + 0.909, + 0.902 + ], + "angle": 0, + "content": "There is a large body of work on the traditional autofocus methods, generally falling into two categories: search-based and phase-based methods. Search-based AF methods [40, 42] search within a range of focus distances to locate the lens position of the extreme sharpness value, which is then determined as the in-focus lens position. Thus, they need to move the lens back and forth, which leads to a highly accurate identification of the in-focus position, but with an annoyingly slow focusing procedure and inevitably the occurrence of focus hunting. Phase-based approaches, on the other hand, rely on the dual-pixel (DP) sensors to obtain the phase data from the left and right sub-images, which is then leveraged to predict the lens movement and determine the in-focus position in a single step of lens" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.521, + 0.957 + ], + "angle": 0, + "content": "26440" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.079, + 0.904, + 0.267 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.272, + 0.907, + 0.357 + ], + "angle": 0, + "content": "Figure 1. Visualization of the focus hunting phenomenon. The lens moves back and forth within a certain range, resulting in a slight change of field of view (FoV) and sharpness. The red, purple, and green lines mark the trajectory of the same target's locations in the picture obtained by three models. The yellow line is a horizontal line as a location reference. (1) The trajectory generated by the learning-based method [7] that predicts the lens position directly. (2) The trajectory generated by the learning-based method [7] that predicts the relative movement of lens. (3) The trajectory generated by our DRL-based method. It can be seen that the FOV and sharpness in trajectory (1) change more significantly, while those in trajectory (2) change slightly, and the lens terminates at the GT after Step 2 in trajectory (3)." + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.369, + 0.482, + 0.446 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.449, + 0.483, + 0.506 + ], + "angle": 0, + "content": "Figure 2. Trajectories generated by predicting the lens position [7] (red line), predicting the relative movement (purple line), and DRL (green line). The trajectory generated by our DRL is the smoothest, indicating a stable and fast focusing procedure." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.507, + 0.483, + 0.612 + ], + "angle": 0, + "content": "movement. Though accelerating the focusing procedure, the phase data obtained by sensors is highly susceptible to noise, leading to a relatively low in-focus accuracy. Especially in some extreme cases, e.g., the low-light conditions, the phase detection may need to be executed multiple times at different lens positions to predict a more accurate in-focus position, still at the risk of focus hunting." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.614, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Recent attempts [7, 16] have concentrated on addressing the above trade-off between accuracy and speed, by exploiting the expressive power of neural networks to directly learn a mapping from the captured image at any lens position to the predicted in-focus position. However, when we directly treat the learned neural network as a lens movement policy and apply it at multiple steps to adjust the lens movement, we notice that some of the resulting trajectories still suffer from the back-and-forth movement of the lens, as shown in Fig. 1 and Fig. 2, due to the lack of leveraging the sequential information obtained from multiple exploration steps and the ignorance of considering future effects of the current prediction. Deep reinforcement learning (DRL), with successful applications in sequential decision-making tasks such as autonomous driving [12] and robot control [8], has enabled the agent to adaptively adjust its policy through iterative interactions with the environment, based both on the immediate reward of current decision step and the long-term reward over the entire sequence of decisions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.371, + 0.907, + 0.809 + ], + "angle": 0, + "content": "Inspired by this, we propose in this paper an expert regularized DRL-based approach for camera autofocus, which is able to utilize the sequential information of lens movement trajectory to both enhance the multi-step in-focus prediction accuracy and reduce the chance of focus hunting. Our proposed method generally follows an actor-critic framework, within which the actor network (i.e., policy network) takes the captured left and right dual-pixel images with the positional information of RoI and lens as the state input, and sequentially generates a trajectory by making decisions on the lens movement. The critic network, on the other hand, receives the same input as the actor network and estimates the future expected accumulated reward given the current state, which is used to evaluate the action predicted by the actor network and thus eliminate suboptimal actions. To accelerate the DRL's training with a higher sample efficiency, we initialize the actor network with a pre-trained single-step prediction network based on the MobileNet-V2 architecture [33], where to be more consistent with the environment transition in AF tasks, the output of absolute in-focus position distribution as used in [7] is modified to the relative lens movement distribution. The critic network also takes the same MobileNet-V2 architecture. To further stabilize the DRL's training with lower occurrence of focus hunting in the resulting lens movement trajectory, we generate some offline trajectories based on prior knowledge to avoid focus hunting, which are then leveraged as an offline dataset of expert trajectories to regularize the actor network's training. Our main contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.811, + 0.907, + 0.886 + ], + "angle": 0, + "content": "- To overcome the limitation of single-step prediction networks, we formulate AF as a multi-step task and then propose a DRL-based framework to fully utilize the sequential information of lens movement trajectory and adjust the policy based on the immediate and long-term rewards." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.887, + 0.905, + 0.902 + ], + "angle": 0, + "content": "- We initialize the policy with a pre-trained single-step pre" + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.811, + 0.907, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "26441" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.103, + 0.092, + 0.483, + 0.152 + ], + "angle": 0, + "content": "dition network, where the network is improved by modifying the output of absolute in-focus position distribution to relative lens movement distribution to establish a better mapping between input images and lens movement." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.153, + 0.483, + 0.198 + ], + "angle": 0, + "content": "- We construct an offline dataset of expert trajectories based on prior knowledge, regularizing the training of the policy network towards reducing occurrence of focus hunting." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.199, + 0.483, + 0.258 + ], + "angle": 0, + "content": "- Empirical evaluations demonstrate that our method outperforms state-of-the-art AF methods on public benchmarks, with higher single- and multi-step prediction accuracy, and a significant reduction of focus hunting rate." + }, + { + "type": "list", + "bbox": [ + 0.092, + 0.153, + 0.483, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.272, + 0.24, + 0.287 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.297, + 0.483, + 0.478 + ], + "angle": 0, + "content": "Traditional passive AF methods can generally be categorized into search-based and phase-based approaches. Search-based approaches calculate and maximize the sharpness of the images based on image statistics [9, 21, 40, 42, 53] or frequency-domain image representations [18, 24, 25, 47, 50] to determine the optimal lens position. While showing high accuracy, these approaches are time-consuming. Additionally, search-based methods inevitably lead to focus hunting, since the peak must be surpassed to locate the exact focus point. Though some methods [14, 39] can reduce the search steps, they cannot solve the problem of focus hunting completely, or can only work on some specific devices." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.479, + 0.483, + 0.705 + ], + "angle": 0, + "content": "With the development of dual-pixel sensors, phase-based methods have become the primary AF method in modern digital cameras [37]. With the phase information of the two (left/right) split pixels, the disparity can be calculated to determine the lens movement. Consequently, phase-based approaches can move the lens to the in-focus lens position with a single image capture, which is more efficient than search-based approaches. However, factors that are difficult to model, such as lens shading, can affect the calculation of phase, leading to lower accuracy. Numerous works adopt learning-based methods to address the trade-off between accuracy and speed. But these works either apply to specific tasks like microscopy images [31, 51], or are based on phase information obtained by specific hardware [17, 46], which are difficult to reproduce and compare." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.483, + 0.902 + ], + "angle": 0, + "content": "The above methods concentrate on the processed sharpness metrics or phase data, not sufficiently using the information from pictures. Inspired by the stereo vision [19, 20], information obtained from dual-pixel pictures has found increasingly widespread applications in the computer vision community, such as monocular depth estimation [11, 22] and defocus deblurring [1]. In AF, Herrmann et al. [16] utilize convolutional networks to predict the in-focus position from input images and provide a large-scale AF dataset. Choi et al. [7] improve Herrmann's network by incorporating the region of interest position encoding (RoI-PE) and the lens position encoding (Lens-PE), significantly reducing the size while achieving state-of-the-art performance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.228 + ], + "angle": 0, + "content": "Both methods, similar to the phase-based approaches, essentially employ the DP data to estimate the disparity and predict the lens position. But the difference is that phase-based approaches estimate the lens movement through disparity rather than directly predicting the next lens position. Meanwhile, though [7, 16] discuss the multi-step prediction, they only address the accuracy of multi-step prediction without considering the potential for focus hunting, which can greatly affect the speed and accuracy of the AF model." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.229, + 0.907, + 0.516 + ], + "angle": 0, + "content": "The main reason for focus hunting in learning-based methods is the lack of consideration of the sequential information and the future effects of the actions. Yan et al. [49] and Wang et al. [46] adopt the long short-term memory (LSTM) to enhance the model's ability to incorporate previous information. With extra computation, the ability of LSTM to capture the 2-dimension data is still limited. The paradigm of predicting actions from states and the need to incorporate sequential information and future effects inspire us to use reinforcement learning (RL) in AF. Though Yu et al. [54] and Chan et al. [6] have proposed DRL-based autofocus methods, they both concentrate on the design of reward functions and do not fully leverage the advantages of RL. Wang et al. [46] adopt RL in their work to dynamically choose the RoI, rather than predicting the lens position or movement. To the best of our knowledge, little research has been conducted to reduce the occurrence of focus hunting while accelerating the multi-step prediction with a higher accuracy for AF tasks." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.534, + 0.688, + 0.551 + ], + "angle": 0, + "content": "3. Proposed Method" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.56, + 0.718, + 0.575 + ], + "angle": 0, + "content": "3.1. Problem Formulation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.584, + 0.907, + 0.903 + ], + "angle": 0, + "content": "We formulate the problem of AF as presented in [7], where the candidate lens position is discretized into \\(n\\) focal indices. Given an input image \\(I\\) captured at a lens position \\(k \\in \\{1,2,\\dots,n\\}\\), we extract a patch \\(I_k^p\\) corresponding to the region of interest (RoI) that we aim to focus, where \\(p \\in \\{1,2,\\dots,P\\}\\) denotes the spatial position of the RoI in this image. A qualified AF process should be able to predict the correct focal index of the lens from a patch sequence captured by moving the lens, while the sequence should ideally be as short as possible, and without a back-and-forth movement. Considering that modern digital cameras (e.g., in smartphones) are commonly equipped with the advanced dual-pixel imaging system that can provide the phase information for AF, we specifically pay attention to the dual-pixel data, such that the input image \\(I\\) consists of two channels. Thus, the extracted patch at the \\(k\\)-th lens position, with resolution \\(h \\times w\\), is denoted as \\(I_k^p \\in \\mathbb{R}^{2 \\times h \\times w}\\). The terms focal stack and single slice are defined as the set of patches \\(\\{I_k^p\\}_{k=1}^n\\) extracted from all lens positions and a single patch \\(I_k^p\\), respectively. The objective is to predict the in-focus lens position \\(k^*\\), which is assumed to be unique for" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "26442" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.085, + 0.482, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.23, + 0.483, + 0.287 + ], + "angle": 0, + "content": "Figure 3. Overall framework of our DRL-based AF model with expert trajectory regularization. The green, orange, blue, and purple boxes specify the actor network, critic network, expert dataset, and environment, respectively." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.297, + 0.204, + 0.311 + ], + "angle": 0, + "content": "each focal stack." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.312, + 0.484, + 0.358 + ], + "angle": 0, + "content": "Relying on the different types of patch sequences used for AF, our AF problem falls into the multi-step prediction category [7], which can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.166, + 0.367, + 0.483, + 0.387 + ], + "angle": 0, + "content": "\\[\nI _ {k _ {0}} ^ {p} \\mapsto k _ {1}, I _ {k _ {1}} ^ {p} \\mapsto k _ {2}, \\dots , I _ {k _ {t}} ^ {p} \\mapsto k ^ {*}. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.394, + 0.484, + 0.485 + ], + "angle": 0, + "content": "Starting from a random focal index \\( k_{0} \\), it makes a prediction step-by-step, generating a trajectory of the lens positions \\( \\{k_{0}, k_{1}, k_{2}, \\ldots, k^{*}\\} \\). Instead of directly predicting the absolute lens position as in Eq. (1), in this paper, we consider predicting the relative lens movement at each step, and thus reformulate the multi-step setting of the AF problem as:" + }, + { + "type": "equation", + "bbox": [ + 0.12, + 0.495, + 0.483, + 0.515 + ], + "angle": 0, + "content": "\\[\nI _ {k _ {0}} ^ {p} \\mapsto k _ {1} = k _ {0} + a _ {0}, \\dots , I _ {k _ {t}} ^ {p} \\mapsto k ^ {*} = k _ {t} + a _ {t}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.522, + 0.483, + 0.537 + ], + "angle": 0, + "content": "where \\( a_{t} \\) is the relative movement of the lens at time step \\( t \\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.545, + 0.422, + 0.56 + ], + "angle": 0, + "content": "3.2. Overall Framework of DRL-Based AF" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.569, + 0.483, + 0.614 + ], + "angle": 0, + "content": "The framework of our DRL-based AF method is shown in Fig. 3, with three major components: an actor network, a critic network, and an offline dataset of expert trajectories." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.615, + 0.483, + 0.704 + ], + "angle": 0, + "content": "- Actor network (i.e., policy network) parameterized by \\(\\theta\\): At each time step \\(t\\), it takes the state \\(s_t\\), including the left and right DP images, the RoI-PE and the Lens-PE, as the input, and outputs an action \\(a_t \\sim \\pi_\\theta(\\cdot | s_t)\\), where the sign of \\(a_t\\) indicates the direction of the lens movement and the absolute value \\(|a_t|\\) represents the moving distance." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.705, + 0.483, + 0.869 + ], + "angle": 0, + "content": "- Critic network parameterized by \\(\\phi\\): At each time step \\(t\\), it takes the state \\(s_t\\) as input, and outputs the estimated value function \\(V_{\\phi}(s_t)\\) for the current state \\(s_t\\). This value function represents the future expected accumulated reward given the current state, which thus can be used to evaluate the action predicted by the actor network, allowing the DRL agent to identify and eliminate some suboptimal actions. Since this evaluation is based on the future expected accumulated reward, the actions are thus chosen based on their anticipated long-term impact, rather than the immediate return." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.87, + 0.483, + 0.902 + ], + "angle": 0, + "content": "- Offline dataset of expert trajectories \\(\\mathcal{D}^E = \\{(s_t, a_t^E)\\}\\): It provides the expert trajectories with state \\(s_t\\) and as" + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.615, + 0.483, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.091, + 0.905, + 0.198 + ], + "angle": 0, + "content": "sociated expert action \\( a_{t}^{E} \\). These trajectories are generated based on the prior knowledge to avoid focus hunting, and can thus be used to regularize the training of the actor network by maximizing the probability of expert action given a state \\( s_{t} \\), which ensures that the actor network learns from the high-quality multi-step movement of the lens from the expert." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.207, + 0.862, + 0.221 + ], + "angle": 0, + "content": "3.3. Actor Network for Movement Prediction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.23, + 0.905, + 0.319 + ], + "angle": 0, + "content": "To utilize the sequential information of the lens movement trajectory to both enhance the multi-step in-focus prediction accuracy and reduce the chance of focus hunting, the actor network ideally should establish a more accurate mapping between the disparity from DP data and the lens movement." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.32, + 0.906, + 0.457 + ], + "angle": 0, + "content": "The AF procedure can be formulated as a Markov Decision Progress (MDP), defined by a tuple \\(\\langle S, \\mathcal{A}, \\mathcal{R}, \\mathcal{P}, \\gamma \\rangle\\), where \\(S\\) denotes the state space, \\(\\mathcal{A}\\) denotes the action space and \\(\\gamma \\in (0,1]\\) is the discount factor. At each time step \\(t\\), a state \\(s_t \\in S\\), consisting of the left and right DP images, the RoI-PE and the Lens-PE, is received by the RL agent, and an action \\(a_t \\in \\mathcal{A}\\) is then generated by the actor network to determine the lens movement. The environment transition dynamics \\(P(s_{t+1} | s_t, a_t)\\) in AF can be defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.568, + 0.467, + 0.905, + 0.486 + ], + "angle": 0, + "content": "\\[\ns _ {t + 1} = g \\left(\\operatorname {c l i p} \\left(a _ {t} + g ^ {- 1} \\left(s _ {t}\\right), 0, k _ {\\max }\\right)\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.496, + 0.906, + 0.587 + ], + "angle": 0, + "content": "where \\( g(k_{t}) = s_{t} \\) maps the lens position \\( k_{t} \\) to the state \\( s_{t} \\) at time-step \\( t \\) and \\( k_{max} \\triangleq k_{n} \\) is the maximum lens position. The reward function \\( R(s_{t}, a_{t}) \\) is defined as the negative absolute distance between the predicted and ground-truth infocus positions, where we additionally introduce a penalty term \\( R_{fh} \\) to mitigate the occurrence of focus hunting:" + }, + { + "type": "equation", + "bbox": [ + 0.594, + 0.598, + 0.905, + 0.616 + ], + "angle": 0, + "content": "\\[\nR \\left(s _ {t}, a _ {t}\\right) = - \\left| k _ {t + 1} - k _ {g t} \\right| + R _ {f h}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.626, + 0.905, + 0.657 + ], + "angle": 0, + "content": "where \\(k_{gt}\\) denotes the ground-truth lens position, and the penalty term \\(R_{fh}\\) is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.617, + 0.667, + 0.905, + 0.709 + ], + "angle": 0, + "content": "\\[\nR _ {f h} = \\left\\{ \\begin{array}{l l} c, & \\text {i f F H o c c u r s}; \\\\ 0, & \\text {o t h e r w i s e}; \\end{array} \\right. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.718, + 0.905, + 0.763 + ], + "angle": 0, + "content": "where \\( c \\) is a negative hyperparameter that penalizes the event of focus hunting. The main goal of RL is to maximize the expected return \\( J(\\theta) \\) defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.602, + 0.774, + 0.905, + 0.816 + ], + "angle": 0, + "content": "\\[\nJ (\\theta) = \\mathbb {E} _ {\\pi_ {\\theta}} \\left[ \\sum_ {t = 0} ^ {T} \\gamma^ {t} R \\left(s _ {t}, a _ {t}\\right) \\right], \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.905, + 0.855 + ], + "angle": 0, + "content": "where \\(\\pi_{\\theta}\\) is the policy network parameterized by \\(\\theta\\) and \\(T\\) denotes the maximum time step." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Since the single-step prediction network model proposed by Choi et al. [7] has demonstrated a state-of-the-art accuracy, we adopt its network architecture as our actor network." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26443" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.14, + 0.086, + 0.443, + 0.21 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.141, + 0.215, + 0.432, + 0.229 + ], + "angle": 0, + "content": "Figure 4. Illustration of relative movement label." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.241, + 0.485, + 0.574 + ], + "angle": 0, + "content": "However, the original model in [7] directly predicts the infocus position, which contradicts the principle of using DP data for in-focus prediction. In essence, since there is a clear correspondence between the disparity and lens movement (i.e., through the defocus conversion coefficient), it is more reasonable to predict the lens movement rather than the lens position based on the DP data. Thus, we modify the original network's output of absolute in-focus position distribution in [7] to the relative lens movement distribution, with the inference procedure of actor network shown in Fig. 9. Accordingly, Fig. 4 illustrates the conversion from original absolute position labels to relative movement labels that are used for training. Unlike [17], we expand the action range from \\([- \\mathrm{gt}, k_{\\max} - \\mathrm{gt}]\\) to \\([-k_{\\max}, k_{\\max}]\\), and the label denoting the relative movement towards GT index then converts to \\(\\mathrm{gt} - f\\). If the actor network predicts an out-of-bound action, it will be clipped to the closest lower/upper bounds. Based on this formulation, we initialize the actor network with a pre-trained MobileNet-V2 [33] optimized based on the ordinal regression loss [10]. For details about the pretrained model of actor network, please refer to Sec. 7 in our supplementary." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.583, + 0.472, + 0.599 + ], + "angle": 0, + "content": "3.4. Critic Network Incorporating Future Impact" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.606, + 0.483, + 0.696 + ], + "angle": 0, + "content": "The actor network cannot directly incorporate the future reward of current actions, where the probability of any output action with a larger instant reward will be selected, leading to a suboptimal policy. To address this issue, a critic network is further introduced to provide baselines and evaluate the future impact of the current actions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.697, + 0.483, + 0.833 + ], + "angle": 0, + "content": "The critic network estimates the value function \\( V_{\\phi}(s_t) \\) of the current state \\( s_t \\), which is defined as the expected future accumulated discount reward of actions. If we further define the future accumulated discount reward of an action under a given state as a Q-function \\( Q(s_t, a_t) \\), the advantage function \\( \\hat{A}_t = Q(s_t, a_t) - V_{\\phi}(s_t) \\) then evaluates the relative quality of a particular action under the current state. With this advantage function, the actor network can be trained by minimizing the clipped surrogate objective loss, given by:" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.843, + 0.483, + 0.861 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {c l i p}} (\\theta) = \\mathbb {E} \\left[ \\min \\left(r _ {t} (\\theta) \\hat {A}, \\operatorname {c l i p} \\left(r _ {t} (\\theta), 1 - \\epsilon , 1 + \\epsilon\\right) \\hat {A} _ {t}\\right), \\right. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.87, + 0.483, + 0.902 + ], + "angle": 0, + "content": "where \\(r_t(\\theta) = \\pi_\\theta(a_t|s_t) / \\pi_\\theta^{old}(a_t|s_t)\\) is the probability ratio and \\(\\epsilon\\) is a clip hyperparameter." + }, + { + "type": "code_caption", + "bbox": [ + 0.516, + 0.092, + 0.786, + 0.107 + ], + "angle": 0, + "content": "Algorithm 1 Expert trajectory generator." + }, + { + "type": "algorithm", + "bbox": [ + 0.517, + 0.111, + 0.905, + 0.402 + ], + "angle": 0, + "content": "Input: \\(\\mathcal{D} = \\{f,GT,(s_0,a_0),\\dots ,(s_n,a_n)\\}_{N};\\) \nOutput: \\(\\mathcal{D}^E\\) . \n1: \\(\\mathcal{D}^E = \\emptyset\\) \n2: for all \\(\\{g(\\cdot),GT,(s_0,a_0),\\dots ,(s_n,a_n)\\} _i\\in \\mathcal{D}\\) do \n3: for all \\(j = 0,1,2,\\dots ,n\\) do \n4: \\(k_{j}\\gets g_{i}^{-1}(s_{j})\\) \n5: \\(o_j\\gets \\mathrm{sign}(k_0 - GT_i)\\cdot |k_j - GT_i| + GT_i;\\) \n6: \\(o_j\\gets \\mathrm{CLIP}(o_j,k_0,GT_i);\\) \n7: end for \n8: \\((o_{m_0},\\dots ,o_{m_n})\\gets \\mathrm{sort}(o_0,\\dots ,o_n;\\mathrm{reverse} =\\) \\(\\mathrm{Bool}(k_0 > GT_i));\\) \n9: for all \\(j = 1,2,\\dots ,n\\) do \n10: \\(s_j\\gets g_i(o_{m_j});\\) \n11: \\(a_{j - 1}\\leftarrow o_{m_j} - o_{m_{j - 1}};\\) \n12: end for \n13: \\(a_{n}\\gets 0\\) \n14: \\(\\mathcal{D}^E\\gets \\mathcal{D}^E\\cup \\{g(\\cdot),GT,(s_0,a_0),\\dots ,(s_n,a_n)\\} _i;\\) \n15: end for \n16: return \\(\\mathcal{D}^E\\)" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.425, + 0.907, + 0.471 + ], + "angle": 0, + "content": "The critic network, on the other hand, is trained by minimizing the \\(\\ell_2\\) loss between the estimated value function \\(V_{\\phi}(s_t)\\) and the true value function \\(\\hat{V}_t\\):" + }, + { + "type": "equation", + "bbox": [ + 0.594, + 0.481, + 0.906, + 0.522 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {c r i t i c}} = \\frac {1}{N} \\sum_ {t = 1} ^ {N} \\left(V _ {\\phi} \\left(s _ {t}\\right) - \\hat {V} _ {t}\\right) ^ {2}. \\tag {8}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.531, + 0.808, + 0.547 + ], + "angle": 0, + "content": "3.5. Expert Trajectory Regularization" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.554, + 0.907, + 0.795 + ], + "angle": 0, + "content": "Though the incorporation of a critic network can improve the effectiveness of actor network's training, the exponential growth in the state and action combinations inherently leads to another issue of low sample efficiency in DRL. Recent literature has witnessed the success of incorporating offline data for improving the sample efficiency of DRL's training, both theoretically [38, 45] and empirically [5, 27, 28]. Motivated by this, and to further stabilize the DRL with a lower occurrence of focus hunting in the resulting lens movement trajectory, we generate some offline trajectories based on the prior knowledge to avoid focus hunting, which are then leveraged as an offline dataset \\(\\mathcal{D}^E = \\{(s_t,a_t^E)\\}\\) of expert trajectories to regularize the actor network's training. Algorithm 1 shows our proposed method to generate these expert trajectories, with more details provided in Sec. 8 in our supplementary." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.906, + 0.855 + ], + "angle": 0, + "content": "Fig. 5 shows our proposed framework for expert trajectory regularized DRL's training. At each time, we sample \\(N\\) trajectories from the offline expert dataset \\(\\mathcal{D}^E\\) to calculate the imitation loss:" + }, + { + "type": "equation", + "bbox": [ + 0.533, + 0.863, + 0.907, + 0.905 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {i m i t} (\\theta) = - \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\sum_ {t = 1} ^ {T} \\log p \\left(\\pi_ {\\theta} (\\cdot | s _ {i, t}) = a _ {i, t}\\right), \\tag {9}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26444" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.113, + 0.082, + 0.465, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.124, + 0.216, + 0.45, + 0.231 + ], + "angle": 0, + "content": "Figure 5. Expert trajectory regularized DRL's training." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.243, + 0.485, + 0.304 + ], + "angle": 0, + "content": "where \\(a_{i,t} = k_{i,t + 1} - k_{i,t}\\) denotes the action at time step \\(t\\) in the \\(i\\)-th trajectory, \\(s_{i,t}\\) represents the network input at lens location \\(k_{i,t}\\), and \\(\\pi_{\\theta}(\\cdot | s_{i,t})\\) represents the output distribution of the actor network, with the total loss defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.179, + 0.317, + 0.483, + 0.334 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} (\\theta) = \\mathcal {L} _ {\\text {c l i p}} (\\theta) + \\lambda \\cdot \\mathcal {L} _ {\\text {i m i t}} (\\theta), \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.346, + 0.484, + 0.392 + ], + "angle": 0, + "content": "where \\(\\lambda\\) balances the online and offline learning. By minimizing the loss \\(\\mathcal{L}(\\theta)\\), our DRL algorithm maximizes the expected accumulated discount reward \\(J(\\theta)\\) indirectly." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.406, + 0.224, + 0.423 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.432, + 0.484, + 0.538 + ], + "angle": 0, + "content": "In the experiments, our actor and critic networks are built upon MobileNetV2 [33]. We reproduce the models for [16] and [7], since no open-source code is available. In the multi-step prediction setting, we only consider the model obtained by [7], since the multi-step model given by [16] must be restrained for each step, which is impractical. More results and analysis are provided in Sec. 9 in our supplementary." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.539, + 0.484, + 0.765 + ], + "angle": 0, + "content": "Dataset. We adopt the publicly available AF dataset constructed by Herrmann et al. [16]. This dataset comprises 51 scenes, each containing 10 stacks with varying compositions. All instances within the dataset are captured at \\( n = 49 \\) focal depths, uniformly sampled in the inverse depth space, ranging from 0.102 to 3.91 meters, captured using the Google Pixel 3 smartphones. To create the training and test sets, we extract and sample \\( 128 \\times 128 \\) patches with a stride of 96 from the central camera, resulting in a total of 3,341,163 patches for the training set and 382,445 patches for the test set. Patches with the same spatial position within one image and 49 different focal depths constitute a single focal stack. Please refer to [16] for additional details about the dataset, and refer to Sec. 6 in our supplementary for the labeling method adopted in this paper." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.766, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Evaluation metric. We follow the evaluation metric proposed by Scharstein et al. [34], which is used for Middlebury stereo dataset and adopted by Herrmann et al. [16] and Choi et al. [7] to evaluate the AF model's prediction accuracy. In particular, this metric is calculated by the proportion of patches whose stop positions are within 0, 1, 2, and 4 indices away from the in-focus index, which can be considered in the depth of field (DoF). The mean absolute error (MAE) and root mean squared error (RMSE) are also" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.258 + ], + "angle": 0, + "content": "used to evaluate the AF model's performance. To evaluate the focus hunting (FH), we retain every focusing trajectory and calculate the direction and distance of each lens movement. If the direction of a current movement is opposite to that of the previous one in the trajectory, we consider an occurrence of FH in that trajectory. Since the actor network obtained by our DRL and the model in [7] share the same architecture and similar input/output, we can simply regard the number of steps to reach the in-focus position as the focus speed. We calculate the images captured at all initial positions of the lens to account for accuracy and FH rate." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.259, + 0.907, + 0.44 + ], + "angle": 0, + "content": "Implementation detail. We use PyTorch [29] and TorchRL [4] for our implementation. For deep learning models, we use the Adam optimizer [23] with a learning rate of \\(10^{-3}\\), \\((\\beta_{1},\\beta_{2}) = (0.5,0.999)\\), a batch size of 128, and cosine learning rate decay [26] for 10k iterations. For reinforcement learning models, we adopt the PPO-CLIP [35] algorithm. The initial learning rate is set to \\(1\\times 10^{-5}\\), and the clip epsilon is 0.2, a batch size of 32, Adam optimizer with the same \\((\\beta_{1},\\beta_{2})\\) as deep learning. The maximum episode length of RL is set to 4. For the expert trajectory, we first construct an expert dataset containing 6,000,000 expert trajectories and set the weight \\(\\lambda\\) to 0.001 in the loss function." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.453, + 0.707, + 0.469 + ], + "angle": 0, + "content": "4.1. Quantitative Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.478, + 0.906, + 0.84 + ], + "angle": 0, + "content": "FH rate and multi-step prediction accuracy (Effects of DRL). The FH rate and multi-step prediction accuracy of comparison methods are shown in Tab. 1. Though the relative movement label can significantly improve the one-step prediction accuracy, the FH rate in the multi-step setting is still high since the current deep-learning training process cannot obtain and exploit the sequential information of the lens movement. Due to DRL's advantages in modeling sequential decision-making problems, our DRL-based models can directly decrease the FH rate without relying on any extra networks, such as the LSTM [36]. Compared with the policy obtained by deep learning, our DRL-based policy can decrease the FH rate by \\(9.1\\%\\) in 2-step prediction, \\(16.5\\%\\) in 3-step prediction, and \\(18.0\\%\\) in 4-step prediction, while adopting the same network architecture. Our work shows some advantages over existing RL baselines. Since few features are extracted from the scalar input of [6], it can only guide the lens towards one side with the higher label frequency, causing extremely low accuracy but 0 FH rate. Since [54] generalizes poorly in various views (as stated by the authors), while the dataset in our work contains images with multiple views and object distances, it hardly gives a move signal. The lens thus stops at the start position in most cases, causing small changes in accuracy and a low FH rate." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.841, + 0.905, + 0.901 + ], + "angle": 0, + "content": "For the multi-step prediction performance, the accuracy of the baseline [7] increases a little and even decreases in 4-step prediction. The accuracy of the model obtained by the relative movement label can achieve substantial gains" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "26445" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.092, + 0.083, + 0.483, + 0.11 + ], + "angle": 0, + "content": "Table 1. Prediction accuracy and focus hunting (FH) rate results in multi-step setting." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.112, + 0.482, + 0.328 + ], + "angle": 0, + "content": "
AlgorithmStep= 0 ↑ ≤ 1 ↑ ≤ 2 ↑ ≤ 4 ↑MAE↓ RMSE↓FH↓
DRL in [6]10.030 0.071 0.117 0.21014.216 17.908NA
20.029 0.050 0.082 0.15117.214 20.8270.0
30.029 0.050 0.082 0.15117.214 20.8270.0
40.029 0.050 0.082 0.15117.214 20.8270.0
DRL in [54]10.022 0.062 0.102 0.18115.575 19.143NA
20.023 0.064 0.105 0.18515.446 19.0510.019
30.024 0.065 0.106 0.18715.437 19.0730.025
40.024 0.065 0.106 0.18615.500 19.1420.030
RoI-PE & Len-PE [7]10.243 0.584 0.756 0.8822.839 6.880NA
20.275 0.640 0.803 0.9122.454 6.2960.275
30.279 0.644 0.810 0.9142.434 6.2050.362
40.276 0.642 0.807 0.9132.442 6.2770.383
[7] with relative label (Ours)10.330 0.698 0.826 0.9072.342 6.339NA
20.413 0.795 0.888 0.9341.894 5.8550.261
30.427 0.805 0.894 0.9371.892 5.7210.340
40.423 0.804 0.893 0.9361.862 5.8440.358
Expert Trajectory (Ours)10.336 0.711 0.839 0.9152.146 5.827NA
20.439 0.808 0.896 0.9411.755 5.5970.170
30.450 0.816 0.900 0.9431.711 5.5320.175
40.449 0.818 0.901 0.9441.694 5.4790.178
" + }, + { + "type": "table_caption", + "bbox": [ + 0.095, + 0.337, + 0.478, + 0.351 + ], + "angle": 0, + "content": "Table 2. Average steps to stop with different termination signals." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.352, + 0.476, + 0.396 + ], + "angle": 0, + "content": "
Termination Rule= 0 ↓ ≤ 1 ↓ ≤ 2 ↓ ≤ 4 ↓Stop Model↓ MAE↓ RMSE↓
RoI-PE & Lens-PE[7]6.6513.6442.4771.7072.2082.6116.146
[7] with relative label (Ours)4.9822.2461.7041.4531.8841.9065.463
Expert Trajectory (Ours)5.0142.2571.6831.3611.8211.8745.332
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.401, + 0.483, + 0.705 + ], + "angle": 0, + "content": "of \\(13.8\\%\\) prediction accuracy for exact \\((= 0)\\) predictions in 2-step prediction, while our DRL-based model further enhances this accuracy. Furthermore, as the number of lens movement steps increases, the accuracy of deep learning-based methods may decline, whereas the accuracy of our DRL-based method remains unchanged or even improves, demonstrating the stability of our DRL-based model in multi-step prediction. Moreover, Tab. 2 shows that our models take fewer steps to stop at different termination signals, e.g., \\(= 0\\), \\(\\leq 1\\), \\(\\leq 2\\), and \\(\\leq 4\\). Since in-focus positions remain unknown in practical camera deployment, the above rules may not be feasible. Alternatively, we train a stop model to generate the stop signal, which is modeled as a binary classification task and takes the same architecture and input as our policy network. Tab. 2 shows the average steps, MAE, and RMSE of different policies with this model. Our method still achieves a better prediction with fewer steps. In conclusion, results in Tab. 1 and Tab. 2 show that our methods can achieve higher accuracy with fewer steps than that proposed by [7], demonstrating a faster focusing speed." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Single-step accuracy (Effects of relative movement as labels). The comparison results in the single-step permutation setting are shown in Tab. 3. Compared with classic statistics-based or depth estimation methods, the deep-learning based models \"Learning to Autofocus [16]\" and \"RoI-PE & Len-PE [7]\" significantly improve the one-step prediction accuracy. By modifying the output of the model in [7] from the absolute lens position to the relative movement, we can further improve the accuracy by \\(8.7\\%\\) for exact \\((= 0)\\) prediction and \\(2.5\\%\\) for \\(\\leq 4\\) focal index difference. Moreover, since the dataset has low-contrast images with similar features but distant labels, minimizing the loss makes the policy perform badly. Via optimizing multi-step" + }, + { + "type": "table_caption", + "bbox": [ + 0.595, + 0.083, + 0.825, + 0.097 + ], + "angle": 0, + "content": "Table 3. Single-step prediction results." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.098, + 0.901, + 0.193 + ], + "angle": 0, + "content": "
Algorithm= 0 ↑≤ 1 ↑≤ 2 ↑≤ 4 ↑MAE↓RMSE↓
ZNCC Disparity with Calibration0.0650.1790.2680.4199.47614.211
SSD Disparity [44]0.1010.2640.3860.5258.01212.414
Learning the Depth [11]0.1120.2970.4150.5627.41912.312
Learning to Autofocus [16]0.1680.4670.6590.8513.2497.619
RoI-PE & Lens-PE [7]0.2430.5840.7560.8822.8396.880
[16] with relative label (Ours)0.1810.4930.6840.8613.0217.374
[7] with relative label(Ours)0.3300.6980.8260.9072.3426.339
Expert Trajectory (Ours)0.3360.7110.8390.9152.1465.827
" + }, + { + "type": "table_caption", + "bbox": [ + 0.514, + 0.198, + 0.905, + 0.225 + ], + "angle": 0, + "content": "Table 4. Performance comparison of DRL models with and without Expert Regularization." + }, + { + "type": "table", + "bbox": [ + 0.526, + 0.227, + 0.895, + 0.322 + ], + "angle": 0, + "content": "
AlgorithmStep= 0 ↑ ≤ 1 ↑ ≤ 2 ↑ ≤ 4 ↑MAE↓ RMSE↓FH↓
w/o10.0310.0560.0860.14322.00526.048
20.0290.0650.0950.14823.84527.975
Expert Regularization30.0370.0850.1230.18820.95725.501
40.0310.0680.0990.15223.72227.889
Expert Trajectory (Ours)10.3360.7110.8390.9152.1465.827
20.4390.8080.8960.9411.7555.597
30.4500.8160.9000.9431.7115.532
40.4490.8180.9010.9441.6945.479
" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.325, + 0.905, + 0.355 + ], + "angle": 0, + "content": "behavior, RL can explore extra information and improve 1-step prediction accuracy." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.356, + 0.905, + 0.447 + ], + "angle": 0, + "content": "Effects of expert regularization. As shown in Tab. 4, the policy obtained by the vanilla DRL algorithm without expert regularization has a poor performance, since the action space is a discrete number of 97, which is large for DRL's training. The expert trajectory significantly improves the performance of the policy given by the PPO algorithm." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.455, + 0.696, + 0.471 + ], + "angle": 0, + "content": "4.2. Qualitative Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.478, + 0.906, + 0.674 + ], + "angle": 0, + "content": "Fig. 1 qualitatively visualizes the trajectory of patches and lens positions of different comparison methods, including the policy obtained by [7], the policy obtained by relative movement label, and our DRL-based model. Compared with directly predicting the lens position [7], our relative movement label model is faster by predicting the lens movement within DoF in the second step, but the FH issue cannot be overcome. In comparison, our DRL-based model can generate a faster and more stable prediction by predicting the GT location at the first step and stopping at this index. The lens trajectory of the DRL-based model (the green curve in Fig. 2) is slightly smoother and more stable without the occurrence of FH." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In Fig. 6, we show qualitative results of 4-step predictions of the three models. From the first row, we can observe that all methods can predict the index within DoF, but our methods can give a more precise prediction with a sharper image than Choi's [7]. From the second to fourth rows, we can observe that [7] sometimes breaks and predicts indices with sharp background, while our DRL-based model successfully predicts indices with sharp target objects. This case is a generally challenging scenario in the AF task since many different DoFs can be estimated. These DoFs may confuse the models, leading to a focus on the wrong object. In general, our relative movement label can give a more accurate prediction by mapping the disparity to the movement of the lens, and the DRL-based model can further enhance the accuracy by estimating the future reward of the action" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26446" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.083, + 0.163, + 0.137 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.138, + 0.162, + 0.192 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.193, + 0.162, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.248, + 0.162, + 0.301 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.111, + 0.302, + 0.136, + 0.311 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image", + "bbox": [ + 0.173, + 0.083, + 0.241, + 0.137 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.173, + 0.138, + 0.241, + 0.192 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.173, + 0.193, + 0.241, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.173, + 0.248, + 0.241, + 0.3 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.186, + 0.3, + 0.229, + 0.316 + ], + "angle": 0, + "content": "absolute position labe" + }, + { + "type": "image", + "bbox": [ + 0.253, + 0.083, + 0.321, + 0.137 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.253, + 0.138, + 0.321, + 0.192 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.253, + 0.193, + 0.321, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.253, + 0.248, + 0.321, + 0.3 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.254, + 0.301, + 0.32, + 0.316 + ], + "angle": 0, + "content": "relative movement label" + }, + { + "type": "image", + "bbox": [ + 0.332, + 0.083, + 0.399, + 0.137 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.332, + 0.138, + 0.399, + 0.192 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.332, + 0.193, + 0.398, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.332, + 0.248, + 0.398, + 0.3 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.356, + 0.301, + 0.376, + 0.308 + ], + "angle": 0, + "content": "DRL" + }, + { + "type": "image", + "bbox": [ + 0.41, + 0.083, + 0.477, + 0.137 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.411, + 0.138, + 0.477, + 0.192 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.411, + 0.193, + 0.477, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.41, + 0.248, + 0.477, + 0.3 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.436, + 0.301, + 0.451, + 0.31 + ], + "angle": 0, + "content": "GT" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.317, + 0.483, + 0.345 + ], + "angle": 0, + "content": "Figure 6. Comparison of the 4-step results of the policy obtained by [7], our relative movement label, and our DRL-based model." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.356, + 0.33, + 0.371 + ], + "angle": 0, + "content": "and reducing the focus hunting rate." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.371, + 0.484, + 0.718 + ], + "angle": 0, + "content": "In Fig. 7, we analyze our model's failure cases: (a) it is difficult and meaningless to focus on the images with no texture since the phase information or contrast information is little to estimate the depth, (b) though our models can give a more accurate prediction in scenarios with multiple DoFs, when the target object is small and the background is complex, the models may be confused by these DoFs, leading to focusing on the wrong object, (c) like all DP data-based AF methods, our models may work poorly when the objects are horizontal. Some possible solutions may work to overcome these limitations. For (a), larger patches can provide more phase information, and we also think (a) might not be that important since few people take photos with no texture. For (b), we claim it is not a bad issue, because sometimes people may want to focus on the background obscured by meaningless foreground elements. To solve this problem, larger patches with more information about the target objects may be useful. Some depth estimation methods [22, 52] can also be used to provide more depth information to help distinguish the objects and background. For (c), solving this problem is challenging in algorithms, while little changes in hardware, especially sensors, could easily mitigate this issue [7]." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.727, + 0.245, + 0.743 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Effects of relative label method. Tab. 5 further demonstrates the effect of relative movement as labels. Our DRL-based model can further enhance the accuracy by \\(0.6\\%\\) for exact \\((= 0)\\) prediction and \\(0.8\\%\\) for \\(\\leq 4\\) focal index difference, in comparison to the model in [7] modified with the relative movement label. We further compare the two label methods under three different architectures: MobileNetV2 [33], ResNet50 [15], and MnasNet [41]. The relative movement label method outperforms the absolute lens position label for each architecture. For a fairer comparison," + }, + { + "type": "image", + "bbox": [ + 0.515, + 0.08, + 0.648, + 0.14 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.142, + 0.647, + 0.2 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.202, + 0.647, + 0.261 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.595, + 0.262, + 0.619, + 0.27 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image", + "bbox": [ + 0.655, + 0.08, + 0.733, + 0.14 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.655, + 0.141, + 0.733, + 0.2 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.655, + 0.201, + 0.733, + 0.261 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.656, + 0.262, + 0.731, + 0.278 + ], + "angle": 0, + "content": "relative movement label" + }, + { + "type": "image", + "bbox": [ + 0.74, + 0.08, + 0.817, + 0.14 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.74, + 0.141, + 0.817, + 0.2 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.74, + 0.201, + 0.817, + 0.261 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.765, + 0.262, + 0.786, + 0.27 + ], + "angle": 0, + "content": "DRL" + }, + { + "type": "image", + "bbox": [ + 0.825, + 0.08, + 0.903, + 0.14 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.825, + 0.141, + 0.903, + 0.2 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.825, + 0.201, + 0.903, + 0.261 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.851, + 0.262, + 0.866, + 0.27 + ], + "angle": 0, + "content": "GT" + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.278, + 0.905, + 0.306 + ], + "angle": 0, + "content": "Figure 7. Failure cases of our models, where only the left image of DP data is visualized." + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.318, + 0.905, + 0.346 + ], + "angle": 0, + "content": "Table 5. Single-step prediction results with different network architectures. The number of input channels is 5." + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.351, + 0.912, + 0.482 + ], + "angle": 0, + "content": "
Algorithm# of paramslabel= 0 ↑ ≤ 1 ↑ ≤ 2 ↑ ≤ 4 ↑MAE↓RMSE↓
MobileNetV2 [33]2.29Mabsolute0.2430.5840.7560.8822.8396.880
2.35Mabsolute0.2520.5920.7640.8882.7196.745
2.35Mrelative0.3300.6980.8260.9072.3426.339
ResNet50 [15]23.61Mabsolute0.2580.6020.7890.8942.6156.623
23.71Mabsolute0.2540.5960.7810.8912.6556.686
23.71Mrelative0.3260.7090.8490.9182.1365.764
MnasNet [41]1.00Mabsolute0.2260.5680.7360.8642.9837.016
1.06Mabsolute0.2130.5490.7180.8503.1197.318
1.06Mrelative0.3050.6780.8170.9052.4366.402
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.491, + 0.907, + 0.657 + ], + "angle": 0, + "content": "we expand the absolute position label to the same range as the relative movement label. It shows that expanding the number of networks' output channels can not improve the accuracy, and the effects remain unknown. Since Lens-PE mainly conveys the lens' current position to the network, while the next position must be predicted by adding its movement to the current position, the relative label methods avoid additionally inferring the identity mapping for Lens-PE, leading to a higher accuracy. To make a fair comparison with the other learning-based models, we thus adopt the same MobileNetV2 architecture in the experiments." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.677, + 0.634, + 0.693 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.704, + 0.907, + 0.901 + ], + "angle": 0, + "content": "We have proposed a DRL-based autofocus model with expert trajectory regularization that can both improve the prediction accuracy and reduce the focus hunting rate, which took an actor-critic framework based on the pretrained MobileNet-V2 architecture for single-step prediction. For the actor network, we further modified the absolute position label in [7, 16] with the relative movement label method. We also constructed an offline dataset of expert trajectories based on prior knowledge, regularizing the policy network training toward reducing focus hunting. Experimental results have clearly demonstrated the effectiveness of our proposed method, in terms of the multi-step prediction accuracy and the occurrence of focus hunting." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26447" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.251, + 0.108 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.112, + 0.486, + 0.234 + ], + "angle": 0, + "content": "This work was supported in part by the National Natural Science Foundation of China under Grant 62125109, Grant 62431017, Grant U24A20251, Grant 62320106003, Grant 62371288, Grant 62401357, Grant 62401366, Grant 62301299, Grant 62120106007, and in part by the Program of Shanghai Science and Technology Innovation Project under Grant 24BC3200800. (Corresponding authors: Chenglin Li, Nuowen Kan, Junni Zou.)" + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.245, + 0.188, + 0.26 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.27, + 0.483, + 0.34 + ], + "angle": 0, + "content": "[1] Abdullah Abuolaim, Mauricio Delbracio, Damien Kelly, Michael S. Brown, and Peyman Milanfar. Learning to reduce defocus blur by realistically modeling dual-pixel data. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 2289-2298, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.341, + 0.483, + 0.382 + ], + "angle": 0, + "content": "[2] Yuhan Bao, Lei Sun, Yuqin Ma, Diyang Gu, and Kaiwei Wang. Improving fast auto-focus with event polarity. Optics Express, 31(15):24025-24044, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.383, + 0.483, + 0.424 + ], + "angle": 0, + "content": "[3] Daniel I Barnea and Harvey F Silverman. A class of algorithms for fast digital image registration. IEEE transactions on Computers, 100(2):179-186, 1972. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.425, + 0.483, + 0.493 + ], + "angle": 0, + "content": "[4] Albert Bou, Matteo Bettini, Sebastian Dittert, Vikash Kumar, Shagun Sodhani, Xiaomeng Yang, Gianni De Fabritis, and Vincent Moens. TorchRL: A data-driven decision-making library for pytorch. arXiv preprint arXiv:2306.00577, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.495, + 0.484, + 0.577 + ], + "angle": 0, + "content": "[5] Serkan Cabi, Sergio Gómez Colmenarejo, Alexander Novikov, Ksenia Konyushkova, Scott Reed, Rae Jeong, Konrad Zolna, Yusuf Aytar, David Budden, Mel Vecerik, et al. Scaling data-driven robotics with reward sketching and batch reinforcement learning. arXiv preprint arXiv:1909.12200, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.579, + 0.482, + 0.62 + ], + "angle": 0, + "content": "[6] Chin-Cheng Chan and Homer H Chen. Autofocus by deep reinforcement learning. *Electronic Imaging*, 31:1-6, 2019. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.621, + 0.483, + 0.69 + ], + "angle": 0, + "content": "[7] Myungsub Choi, Hana Lee, and Hyong-Euk Lee. Exploring positional characteristics of dual-pixel data for camera autofocus. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 13112-13122, 2023. 2, 3, 4, 5, 6, 7, 8, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.691, + 0.483, + 0.747 + ], + "angle": 0, + "content": "[8] Murtaza Dalal, Deepak Pathak, and Russ R Salakhutdinov. Accelerating robotic reinforcement learning via parameterized action primitives. In Advances in Neural Information Processing Systems (NeurIPS), pages 21847-21859, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.747, + 0.483, + 0.816 + ], + "angle": 0, + "content": "[9] Judith Dijk, Michael van Ginkel, Rutger J van Asselt, Lucas J van Vliet, and Piet W Verbeek. A new sharpness measure based on gaussian lines and edges. In 10th International Conference on Computer Analysis of Images and Patterns (CAIP), pages 149-156, 2003. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.817, + 0.483, + 0.871 + ], + "angle": 0, + "content": "[10] Raul Diaz and Amit Marathe. Soft labels for ordinal regression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4733-4742, 2019. 5, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.483, + 0.902 + ], + "angle": 0, + "content": "[11] Rahul Garg, Neal Wadhwa, Sameer Ansari, and Jonathan T Barron. Learning single camera depth estimation using" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.27, + 0.484, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.907, + 0.134 + ], + "angle": 0, + "content": "dual-pixels. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 7628-7637, 2019. 3, 7, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.135, + 0.906, + 0.259 + ], + "angle": 0, + "content": "[12] Cole Gulino, Justin Fu, Wenjie Luo, George Tucker, Eli Bronstein, Yiren Lu, Jean Harb, Xinlei Pan, Yan Wang, Xiangyu Chen, John Co-Reyes, Rishabh Agarwal, Rebecca Roelofs, Yao Lu, Nico Montali, Paul Mougin, Zoey Yang, Brandyn White, Aleksandra Faust, Rowan McAllister, Dragomir Anguelov, and Benjamin Sapp. Waymax: An accelerated, data-driven simulator for large-scale autonomous driving research. In Advances in Neural Information Processing Systems (NeurIPS), pages 7730-7742, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.26, + 0.905, + 0.287 + ], + "angle": 0, + "content": "[13] Marsha Jo Hannah. Computer matching of areas in stereo images. Stanford University, 1974. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.288, + 0.905, + 0.343 + ], + "angle": 0, + "content": "[14] Jie He, Rongzheng Zhou, and Zhiliang Hong. Modified fast climbing search auto-focus algorithm with adaptive step size searching technique for digital camera. IEEE Trans. Consumer Electron., 49(2):257-262, 2003. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.344, + 0.905, + 0.399 + ], + "angle": 0, + "content": "[15] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 770-778, 2016. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.4, + 0.905, + 0.482 + ], + "angle": 0, + "content": "[16] Charles Herrmann, Richard Strong Bowen, Neal Wadhwa, Rahul Garg, Qiurui He, Jonathan T. Barron, and Ramin Zabih. Learning to autofocus. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2227-2236, 2020. 2, 3, 6, 7, 8, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.484, + 0.905, + 0.537 + ], + "angle": 0, + "content": "[17] Chi-Jui Ho, Chin-Cheng Chan, and Homer H Chen. A net: A convolutional neural network approach to phase detection autofocus. IEEE Transactions on Image Processing, 29:6386-6395, 2020. 3, 5, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.538, + 0.906, + 0.607 + ], + "angle": 0, + "content": "[18] Jui-Ting Huang, Chun-Hung Shen, See-May Phoong, and Homer Chen. Robust measure of image focus in the wavelet domain. In 2005 International Symposium on Intelligent Signal Processing and Communication Systems, pages 157-160, 2005. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.608, + 0.906, + 0.663 + ], + "angle": 0, + "content": "[19] ZHENG Jin, JIANG Botao, PENG Wei, and ZHANG Qiaohui. Multi-scale binocular stereo matching based on semantic association. Chinese Journal of Electronics, 33(4):1010-1022, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.664, + 0.907, + 0.747 + ], + "angle": 0, + "content": "[20] Junpeng Jing, Jiankun Li, Pengfei Xiong, Jiangyu Liu, Shuaicheng Liu, Yichen Guo, Xin Deng, Mai Xu, Lai Jiang, and Leonid Sigal. Uncertainty guided adaptive warping for robust and efficient stereo matching. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 3318-3327, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.747, + 0.905, + 0.775 + ], + "angle": 0, + "content": "[21] Neel Joshi and C Lawrence Zitnick. Micro-baseline stereo. Microsoft Research Technical Report, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.776, + 0.906, + 0.845 + ], + "angle": 0, + "content": "[22] Donggun Kim, Hyeonjoong Jang, Inchul Kim, and Min H. Kim. Spatio-focal bidirectional disparity estimation from a dual-pixel image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5023-5032, 2023. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.845, + 0.905, + 0.873 + ], + "angle": 0, + "content": "[23] Diederik P Kingma. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.902 + ], + "angle": 0, + "content": "[24] Matej Kristan, Janez Pers, Matej Perse, and Stanislav Kovačić. A bayes-spectral-entropy-based measure of camera" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.093, + 0.907, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "26448" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.126, + 0.093, + 0.482, + 0.12 + ], + "angle": 0, + "content": "focus using a discrete cosine transform. Pattern Recognition Letters, 27(13):1431-1439, 2006. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.122, + 0.483, + 0.177 + ], + "angle": 0, + "content": "[25] Sang-Yong Lee, Jae-Tack Yoo, Yogendera Kumar, and SooWon Kim. Reduced energy-ratio measure for robust autofocusing in digital camera. IEEE Signal Processing Letters, 16 (2):133-136, 2009. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.179, + 0.482, + 0.22 + ], + "angle": 0, + "content": "[26] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.222, + 0.482, + 0.292 + ], + "angle": 0, + "content": "[27] Yao Lu, Karol Hausman, Yevgen Chebotar, Mengyuan Yan, Eric Jang, Alexander Herzog, Ted Xiao, Alex Irpan, Mohi Khansari, Dmitry Kalashnikov, et al. Aw-opt: Learning robotic skills with imitation and reinforcement at scale. In Conference on Robot Learning, pages 1078-1088, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.293, + 0.482, + 0.347 + ], + "angle": 0, + "content": "[28] Ashvin Nair, Abhishek Gupta, Murtaza Dalal, and Sergey Levine. Awac: Accelerating online reinforcement learning with offline datasets. arXiv preprint arXiv:2006.09359, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.35, + 0.482, + 0.474 + ], + "angle": 0, + "content": "[29] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Kopf, Edward Yang, Zachary DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. Pytorch: An imperative style, high-performance deep learning library. In Advances in Neural Information Processing Systems (NeurIPS), 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.476, + 0.482, + 0.546 + ], + "angle": 0, + "content": "[30] José Luis Pech-Pacheco, Gabriel Cristóbal, Jesús Chamorro-Martinez, and Joaquín Fernández-Valdivia. Diatom autofocusing in brightfield microscopy: a comparative study. In Proceedings 15th International Conference on Pattern Recognition. ICPR-2000, pages 314-317. IEEE, 2000. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.548, + 0.482, + 0.602 + ], + "angle": 0, + "content": "[31] Henry Pinkard, Zachary Phillips, Arman Babakhani, Daniel A Fletcher, and Laura Waller. Deep learning for single-shot autofocus microscopy. Optica, 6(6):794-797, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.604, + 0.482, + 0.687 + ], + "angle": 0, + "content": "[32] Stéphane Ross, Geoffrey Gordon, and Drew Bagnell. A reduction of imitation learning and structured prediction to no-regret online learning. In Proceedings of the fourteenth international conference on artificial intelligence and statistics, pages 627-635. JMLR Workshop and Conference Proceedings, 2011. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.689, + 0.482, + 0.758 + ], + "angle": 0, + "content": "[33] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zhmoginov, and Liang-Chieh Chen. *Mobilenetv2: Inverted residuals and linear bottlenecks*. In *Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition* (CVPR), pages 4510-4520, 2018. 2, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.76, + 0.482, + 0.814 + ], + "angle": 0, + "content": "[34] Daniel Scharstein and Richard Szeliski. A taxonomy and evaluation of dense two-frame stereo correspondence algorithms. International Journal of Computer Vision, 47:7-42, 2002. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.816, + 0.482, + 0.858 + ], + "angle": 0, + "content": "[35] John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.86, + 0.482, + 0.902 + ], + "angle": 0, + "content": "[36] Xingjian SHI, Zhourong Chen, Hao Wang, Dit-Yan Yeung, Wai-kin Wong, and Wang-chun WOO. Convolutional LSTM Network: A Machine Learning Approach for Precipitation" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "Nowcasting. In Advances in Neural Information Processing Systems (NeurIPS), 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.121, + 0.905, + 0.162 + ], + "angle": 0, + "content": "[37] Przemysław Sliwinski and Paweł Wachel. A simple model for on-sensor phase-detection autofocusing algorithm. Journal of Computer and Communications, 1(06):11, 2013. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.163, + 0.905, + 0.231 + ], + "angle": 0, + "content": "[38] Yuda Song, Yifei Zhou, Ayush Sekhari, Drew Bagnell, Akshay Krishnamurthy, and Wen Sun. Hybrid RL: Using both offline and online data can make RL efficient. In International Conference on Learning Representations (ICLR), 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.233, + 0.905, + 0.302 + ], + "angle": 0, + "content": "[39] Changqing Su, Zhiyuan Ye, Yongsheng Xiao, You Zhou, Zhen Cheng, Bo Xiong, Zhaofei Yu, and Tiejun Huang. Intensity-robust autofocus for spike camera. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 25018-25027, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.303, + 0.905, + 0.343 + ], + "angle": 0, + "content": "[40] Murali Subbarao, Tae-Sun Choi, and Arman Nikzad. Focusing techniques. Optical Engineering, 32(11):2824-2836, 1993. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.344, + 0.905, + 0.425 + ], + "angle": 0, + "content": "[41] Mingxing Tan, Bo Chen, Ruoming Pang, Vijay Vasudevan, Mark Sandler, Andrew Howard, and Quoc V. Le. Mnasnet: Platform-aware neural architecture search for mobile. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2815-2823, 2019. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.427, + 0.904, + 0.454 + ], + "angle": 0, + "content": "[42] Jay Martin Tenenbaum. Accommodation in computer vision. Stanford University, 1971. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.456, + 0.905, + 0.511 + ], + "angle": 0, + "content": "[43] Quoc Kien Vuong and Jeong-won Lee. Initial direction and speed decision system for auto focus based on blur detection. In 2013 IEEE International Conference on Consumer Electronics (ICCE), pages 222-223, 2013. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.511, + 0.905, + 0.58 + ], + "angle": 0, + "content": "[44] Neal Wadhwa, Rahul Garg, David E Jacobs, Bryan E Feldman, Nori Kanazawa, Robert Carroll, Yair Movshovitz-Attias, Jonathan T Barron, Yael Pritch, and Marc Levoy. Synthetic depth-of-field with a single-camera mobile phone. ACM Transactions on Graphics (ToG), 37(4):1-13, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.581, + 0.905, + 0.622 + ], + "angle": 0, + "content": "[45] Andrew Wagenmaker and Aldo Pacchiano. Leveraging offline data in online reinforcement learning. arXiv preprint arXiv:2211.04974, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.623, + 0.905, + 0.676 + ], + "angle": 0, + "content": "[46] Chengyu Wang, Qian Huang, Ming Cheng, Zhan Ma, and David J. Brady. Deep learning for camera autofocus. IEEE Transactions on Computational Imaging, 7:258-271, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.678, + 0.905, + 0.748 + ], + "angle": 0, + "content": "[47] Hui Xie, Weibin Rong, and Lining Sun. Wavelet-based focus measure and 3-d surface reconstruction method for microscopy images. In Proceedings 2006 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS 2006), pages 229-234, 2006. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.749, + 0.905, + 0.817 + ], + "angle": 0, + "content": "[48] Takahiro Yamasaki, Tomohiro Nakamura, Ryohei Funatsu, and Hiroshi Shimamoto. Hybrid autofocus system by using a combination of the sensor-based phase-difference detection and focus-aid signal. In 2018 IEEE International Conference on Consumer Electronics (ICCE), pages 1–2, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.818, + 0.905, + 0.872 + ], + "angle": 0, + "content": "[49] Jingyang Yan, Peter DiMeo, Lu Sun, and Xian Du. LSTM-Based Model Predictive Control of Piezoelectric Motion Stages for High-Speed Autofocus. IEEE Transactions on Industrial Electronics, 70(6):6209-6218, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.902 + ], + "angle": 0, + "content": "[50] Ge Yang and Bradley J Nelson. Wavelet-based autofocusing and unsupervised segmentation of microscopic images." + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "26449" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.124, + 0.092, + 0.484, + 0.133 + ], + "angle": 0, + "content": "In Proceedings 2003 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS 2003), pages 2143-2148, 2003. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.484, + 0.19 + ], + "angle": 0, + "content": "[51] Hyun Jong Yang, Moohyun Oh, Jonggyu Jang, Hyeonsu Lyu, and Junhee Lee. Robust deep-learning based autofocus score prediction for scanning electron microscope. Microscopy and Microanalysis, 26(S2):702-705, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.192, + 0.484, + 0.261 + ], + "angle": 0, + "content": "[52] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10371-10381, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.263, + 0.484, + 0.318 + ], + "angle": 0, + "content": "[53] Yi Yao, Besma Abidi, Narjes Doggaz, and Mongi Abidi. Evaluation of sharpness measures and search algorithms for the auto-focusing of high-magnification images. In Visual Information Processing XV, pages 132-143. SPIE, 2006. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.32, + 0.484, + 0.387 + ], + "angle": 0, + "content": "[54] Xiaofan Yu, Runze Yu, Jingsong Yang, and Xiaohui Duan. A robotic auto-focus system based on deep reinforcement learning. In 2018 15th International Conference on Control, Automation, Robotics and Vision (ICARCV), pages 204-209, 2018. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.39, + 0.484, + 0.459 + ], + "angle": 0, + "content": "[55] Ramin Zabih and John Woodfill. Non-parametric local transforms for computing visual correspondence. In Computer Vision—ECCV'94: Third European Conference on Computer Vision Stockholm, Sweden, May 2–6 1994 Proceedings, Volume II 3, pages 151–158. Springer, 1994. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.459 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "26450" + } + ] +] \ No newline at end of file diff --git a/2025/Stabilizing and Accelerating Autofocus with Expert Trajectory Regularized Deep Reinforcement Learning/a699e1d4-e683-4371-bc56-144538826827_origin.pdf b/2025/Stabilizing and Accelerating Autofocus with Expert Trajectory Regularized Deep Reinforcement Learning/a699e1d4-e683-4371-bc56-144538826827_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..807518bfda4b5a6610e4ee8e63e97218bbec901d --- /dev/null +++ b/2025/Stabilizing and Accelerating Autofocus with Expert Trajectory Regularized Deep Reinforcement Learning/a699e1d4-e683-4371-bc56-144538826827_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9c016a506a751128acee563eeedbf4c386937d05f57f5402c12739c4130d02b4 +size 3555644 diff --git a/2025/Stabilizing and Accelerating Autofocus with Expert Trajectory Regularized Deep Reinforcement Learning/full.md b/2025/Stabilizing and Accelerating Autofocus with Expert Trajectory Regularized Deep Reinforcement Learning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..4b48530c4ad6e34506e1d76629d6a7d553f12ad0 --- /dev/null +++ b/2025/Stabilizing and Accelerating Autofocus with Expert Trajectory Regularized Deep Reinforcement Learning/full.md @@ -0,0 +1,396 @@ +# Stabilizing and Accelerating Autofocus with Expert Trajectory Regularized Deep Reinforcement Learning + +Shouhang Zhu1, Chenglin Li1, Yuankun Jiang2, Li Wei1, Nuowen Kan1, Ziyang Zheng1, Wenrui Dai2, Junni Zou2, Hongkai Xiong1 + +$^{1}$ Department of Electronic Engineering, Shanghai Jiao Tong University + +$^{2}$ Department of Computer Science and Engineering, Shanghai Jiao Tong University + +$^{1}$ shzhu2001, lcl1985, wei.li, kannw.1230, zhengziyang, xionghongkai}@sjtu.edu.cn + +$^{2}$ $\{\text { yuankunjiang, daiwenrui, zoujunni }\} @$ sjtu.edu.cn + +# Abstract + +Autofocus is a crucial component of modern digital cameras. While recent learning-based methods achieve state-of-the-art in focus prediction accuracy, they unfortunately ignore the potential focus hunting phenomenon of back-and-forth lens movement in the multi-step focusing procedure. To address this, in this paper, we propose an expert regularized deep reinforcement learning (DRL)-based approach for autofocus, which can utilize the sequential information of lens movement trajectory to both enhance the multi-step in-focus prediction accuracy and reduce the chance of focus hunting. Our method generally follows an actor-critic framework. To accelerate the DRL's training with a higher sample efficiency, we initialize the policy with a pre-trained single-step prediction network, where the network is further improved by modifying the output of absolute in-focus position distribution to the relative lens movement distribution to establish a better mapping between input images and lens movement. To further stabilize DRL's training with a lower occurrence of focus hunting in the resulting lens movement trajectory, we generate some offline trajectories based on prior knowledge to avoid focus hunting, which are then leveraged as an offline dataset of expert trajectories to regularize the actor network's training. Empirical evaluations show that our method outperforms those learning-based methods on public benchmarks, with higher single- and multi-step prediction accuracy, and a significant reduction of focus hunting rate. + +# 1. Introduction + +Autofocus (AF) is a process through which the camera can determine automatically a lens position to focus on the region of interest (RoI). It is a crucial step in image capturing, since once an image is already captured in a defocused and blurry state, it is then difficult to restore its sharpness + +through any post-processing due to the varying effects of blur influenced by factors such as camera aperture and radial distortion [1]. In addition, the beauty of a moment often lies in its fleeting nature, which thus imposes an imperative demand for the fast focusing speed of an AF algorithm. However, due to the instability in focusing systems incurred by some interference factors, the AF mechanism of the camera lens may fail to accurately lock onto the focal point, resulting in a repeated movement of the lens within a certain range. This phenomenon, as illustrated in Fig. 1, is referred to as focus hunting (FH), whereby the camera is unable to identify the focus plane and thus drives the lens forward and backward through a set range of focus hunting for it. It then causes variations in the image sharpness and changes the lens's field of view (FoV), ultimately leading to an unstable focusing experience with longer focusing time, which is a common issue in the traditional AF-assisted cameras [2]. It is thus of great importance to develop an AF algorithm that not only achieves a higher in-focus accuracy but can also stabilize the trajectory of lens movement with a faster convergence speed to the in-focus position, thus reducing the probability of occurrence of focus hunting [43, 48]. + +There is a large body of work on the traditional autofocus methods, generally falling into two categories: search-based and phase-based methods. Search-based AF methods [40, 42] search within a range of focus distances to locate the lens position of the extreme sharpness value, which is then determined as the in-focus lens position. Thus, they need to move the lens back and forth, which leads to a highly accurate identification of the in-focus position, but with an annoyingly slow focusing procedure and inevitably the occurrence of focus hunting. Phase-based approaches, on the other hand, rely on the dual-pixel (DP) sensors to obtain the phase data from the left and right sub-images, which is then leveraged to predict the lens movement and determine the in-focus position in a single step of lens + +![](images/ab10fbb75f3bcafaf8e897d27a362794ad1a3d8259bc6e43592c067a2300ce66.jpg) +Figure 1. Visualization of the focus hunting phenomenon. The lens moves back and forth within a certain range, resulting in a slight change of field of view (FoV) and sharpness. The red, purple, and green lines mark the trajectory of the same target's locations in the picture obtained by three models. The yellow line is a horizontal line as a location reference. (1) The trajectory generated by the learning-based method [7] that predicts the lens position directly. (2) The trajectory generated by the learning-based method [7] that predicts the relative movement of lens. (3) The trajectory generated by our DRL-based method. It can be seen that the FOV and sharpness in trajectory (1) change more significantly, while those in trajectory (2) change slightly, and the lens terminates at the GT after Step 2 in trajectory (3). + +![](images/bf126b2aeedf7e1e82303c6fbbd7ad57427a5c0ede5e53915a4b1e2a6ed3fc4a.jpg) +Figure 2. Trajectories generated by predicting the lens position [7] (red line), predicting the relative movement (purple line), and DRL (green line). The trajectory generated by our DRL is the smoothest, indicating a stable and fast focusing procedure. + +movement. Though accelerating the focusing procedure, the phase data obtained by sensors is highly susceptible to noise, leading to a relatively low in-focus accuracy. Especially in some extreme cases, e.g., the low-light conditions, the phase detection may need to be executed multiple times at different lens positions to predict a more accurate in-focus position, still at the risk of focus hunting. + +Recent attempts [7, 16] have concentrated on addressing the above trade-off between accuracy and speed, by exploiting the expressive power of neural networks to directly learn a mapping from the captured image at any lens position to the predicted in-focus position. However, when we directly treat the learned neural network as a lens movement policy and apply it at multiple steps to adjust the lens movement, we notice that some of the resulting trajectories still suffer from the back-and-forth movement of the lens, as shown in Fig. 1 and Fig. 2, due to the lack of leveraging the sequential information obtained from multiple exploration steps and the ignorance of considering future effects of the current prediction. Deep reinforcement learning (DRL), with successful applications in sequential decision-making tasks such as autonomous driving [12] and robot control [8], has enabled the agent to adaptively adjust its policy through iterative interactions with the environment, based both on the immediate reward of current decision step and the long-term reward over the entire sequence of decisions. + +Inspired by this, we propose in this paper an expert regularized DRL-based approach for camera autofocus, which is able to utilize the sequential information of lens movement trajectory to both enhance the multi-step in-focus prediction accuracy and reduce the chance of focus hunting. Our proposed method generally follows an actor-critic framework, within which the actor network (i.e., policy network) takes the captured left and right dual-pixel images with the positional information of RoI and lens as the state input, and sequentially generates a trajectory by making decisions on the lens movement. The critic network, on the other hand, receives the same input as the actor network and estimates the future expected accumulated reward given the current state, which is used to evaluate the action predicted by the actor network and thus eliminate suboptimal actions. To accelerate the DRL's training with a higher sample efficiency, we initialize the actor network with a pre-trained single-step prediction network based on the MobileNet-V2 architecture [33], where to be more consistent with the environment transition in AF tasks, the output of absolute in-focus position distribution as used in [7] is modified to the relative lens movement distribution. The critic network also takes the same MobileNet-V2 architecture. To further stabilize the DRL's training with lower occurrence of focus hunting in the resulting lens movement trajectory, we generate some offline trajectories based on prior knowledge to avoid focus hunting, which are then leveraged as an offline dataset of expert trajectories to regularize the actor network's training. Our main contributions are summarized as follows: + +- To overcome the limitation of single-step prediction networks, we formulate AF as a multi-step task and then propose a DRL-based framework to fully utilize the sequential information of lens movement trajectory and adjust the policy based on the immediate and long-term rewards. +- We initialize the policy with a pre-trained single-step pre + +dition network, where the network is improved by modifying the output of absolute in-focus position distribution to relative lens movement distribution to establish a better mapping between input images and lens movement. + +- We construct an offline dataset of expert trajectories based on prior knowledge, regularizing the training of the policy network towards reducing occurrence of focus hunting. +- Empirical evaluations demonstrate that our method outperforms state-of-the-art AF methods on public benchmarks, with higher single- and multi-step prediction accuracy, and a significant reduction of focus hunting rate. + +# 2. Related Works + +Traditional passive AF methods can generally be categorized into search-based and phase-based approaches. Search-based approaches calculate and maximize the sharpness of the images based on image statistics [9, 21, 40, 42, 53] or frequency-domain image representations [18, 24, 25, 47, 50] to determine the optimal lens position. While showing high accuracy, these approaches are time-consuming. Additionally, search-based methods inevitably lead to focus hunting, since the peak must be surpassed to locate the exact focus point. Though some methods [14, 39] can reduce the search steps, they cannot solve the problem of focus hunting completely, or can only work on some specific devices. + +With the development of dual-pixel sensors, phase-based methods have become the primary AF method in modern digital cameras [37]. With the phase information of the two (left/right) split pixels, the disparity can be calculated to determine the lens movement. Consequently, phase-based approaches can move the lens to the in-focus lens position with a single image capture, which is more efficient than search-based approaches. However, factors that are difficult to model, such as lens shading, can affect the calculation of phase, leading to lower accuracy. Numerous works adopt learning-based methods to address the trade-off between accuracy and speed. But these works either apply to specific tasks like microscopy images [31, 51], or are based on phase information obtained by specific hardware [17, 46], which are difficult to reproduce and compare. + +The above methods concentrate on the processed sharpness metrics or phase data, not sufficiently using the information from pictures. Inspired by the stereo vision [19, 20], information obtained from dual-pixel pictures has found increasingly widespread applications in the computer vision community, such as monocular depth estimation [11, 22] and defocus deblurring [1]. In AF, Herrmann et al. [16] utilize convolutional networks to predict the in-focus position from input images and provide a large-scale AF dataset. Choi et al. [7] improve Herrmann's network by incorporating the region of interest position encoding (RoI-PE) and the lens position encoding (Lens-PE), significantly reducing the size while achieving state-of-the-art performance. + +Both methods, similar to the phase-based approaches, essentially employ the DP data to estimate the disparity and predict the lens position. But the difference is that phase-based approaches estimate the lens movement through disparity rather than directly predicting the next lens position. Meanwhile, though [7, 16] discuss the multi-step prediction, they only address the accuracy of multi-step prediction without considering the potential for focus hunting, which can greatly affect the speed and accuracy of the AF model. + +The main reason for focus hunting in learning-based methods is the lack of consideration of the sequential information and the future effects of the actions. Yan et al. [49] and Wang et al. [46] adopt the long short-term memory (LSTM) to enhance the model's ability to incorporate previous information. With extra computation, the ability of LSTM to capture the 2-dimension data is still limited. The paradigm of predicting actions from states and the need to incorporate sequential information and future effects inspire us to use reinforcement learning (RL) in AF. Though Yu et al. [54] and Chan et al. [6] have proposed DRL-based autofocus methods, they both concentrate on the design of reward functions and do not fully leverage the advantages of RL. Wang et al. [46] adopt RL in their work to dynamically choose the RoI, rather than predicting the lens position or movement. To the best of our knowledge, little research has been conducted to reduce the occurrence of focus hunting while accelerating the multi-step prediction with a higher accuracy for AF tasks. + +# 3. Proposed Method + +# 3.1. Problem Formulation + +We formulate the problem of AF as presented in [7], where the candidate lens position is discretized into $n$ focal indices. Given an input image $I$ captured at a lens position $k \in \{1,2,\dots,n\}$ , we extract a patch $I_k^p$ corresponding to the region of interest (RoI) that we aim to focus, where $p \in \{1,2,\dots,P\}$ denotes the spatial position of the RoI in this image. A qualified AF process should be able to predict the correct focal index of the lens from a patch sequence captured by moving the lens, while the sequence should ideally be as short as possible, and without a back-and-forth movement. Considering that modern digital cameras (e.g., in smartphones) are commonly equipped with the advanced dual-pixel imaging system that can provide the phase information for AF, we specifically pay attention to the dual-pixel data, such that the input image $I$ consists of two channels. Thus, the extracted patch at the $k$ -th lens position, with resolution $h \times w$ , is denoted as $I_k^p \in \mathbb{R}^{2 \times h \times w}$ . The terms focal stack and single slice are defined as the set of patches $\{I_k^p\}_{k=1}^n$ extracted from all lens positions and a single patch $I_k^p$ , respectively. The objective is to predict the in-focus lens position $k^*$ , which is assumed to be unique for + +![](images/11dce178f552eada748672e8e75193006c340380690198caaa334ae44d9ee5e4.jpg) +Figure 3. Overall framework of our DRL-based AF model with expert trajectory regularization. The green, orange, blue, and purple boxes specify the actor network, critic network, expert dataset, and environment, respectively. + +each focal stack. + +Relying on the different types of patch sequences used for AF, our AF problem falls into the multi-step prediction category [7], which can be formulated as: + +$$ +I _ {k _ {0}} ^ {p} \mapsto k _ {1}, I _ {k _ {1}} ^ {p} \mapsto k _ {2}, \dots , I _ {k _ {t}} ^ {p} \mapsto k ^ {*}. \tag {1} +$$ + +Starting from a random focal index $k_{0}$ , it makes a prediction step-by-step, generating a trajectory of the lens positions $\{k_{0}, k_{1}, k_{2}, \ldots, k^{*}\}$ . Instead of directly predicting the absolute lens position as in Eq. (1), in this paper, we consider predicting the relative lens movement at each step, and thus reformulate the multi-step setting of the AF problem as: + +$$ +I _ {k _ {0}} ^ {p} \mapsto k _ {1} = k _ {0} + a _ {0}, \dots , I _ {k _ {t}} ^ {p} \mapsto k ^ {*} = k _ {t} + a _ {t}, \tag {2} +$$ + +where $a_{t}$ is the relative movement of the lens at time step $t$ . + +# 3.2. Overall Framework of DRL-Based AF + +The framework of our DRL-based AF method is shown in Fig. 3, with three major components: an actor network, a critic network, and an offline dataset of expert trajectories. + +- Actor network (i.e., policy network) parameterized by $\theta$ : At each time step $t$ , it takes the state $s_t$ , including the left and right DP images, the RoI-PE and the Lens-PE, as the input, and outputs an action $a_t \sim \pi_\theta(\cdot | s_t)$ , where the sign of $a_t$ indicates the direction of the lens movement and the absolute value $|a_t|$ represents the moving distance. +- Critic network parameterized by $\phi$ : At each time step $t$ , it takes the state $s_t$ as input, and outputs the estimated value function $V_{\phi}(s_t)$ for the current state $s_t$ . This value function represents the future expected accumulated reward given the current state, which thus can be used to evaluate the action predicted by the actor network, allowing the DRL agent to identify and eliminate some suboptimal actions. Since this evaluation is based on the future expected accumulated reward, the actions are thus chosen based on their anticipated long-term impact, rather than the immediate return. +- Offline dataset of expert trajectories $\mathcal{D}^E = \{(s_t, a_t^E)\}$ : It provides the expert trajectories with state $s_t$ and as + +sociated expert action $a_{t}^{E}$ . These trajectories are generated based on the prior knowledge to avoid focus hunting, and can thus be used to regularize the training of the actor network by maximizing the probability of expert action given a state $s_{t}$ , which ensures that the actor network learns from the high-quality multi-step movement of the lens from the expert. + +# 3.3. Actor Network for Movement Prediction + +To utilize the sequential information of the lens movement trajectory to both enhance the multi-step in-focus prediction accuracy and reduce the chance of focus hunting, the actor network ideally should establish a more accurate mapping between the disparity from DP data and the lens movement. + +The AF procedure can be formulated as a Markov Decision Progress (MDP), defined by a tuple $\langle S, \mathcal{A}, \mathcal{R}, \mathcal{P}, \gamma \rangle$ , where $S$ denotes the state space, $\mathcal{A}$ denotes the action space and $\gamma \in (0,1]$ is the discount factor. At each time step $t$ , a state $s_t \in S$ , consisting of the left and right DP images, the RoI-PE and the Lens-PE, is received by the RL agent, and an action $a_t \in \mathcal{A}$ is then generated by the actor network to determine the lens movement. The environment transition dynamics $P(s_{t+1} | s_t, a_t)$ in AF can be defined as: + +$$ +s _ {t + 1} = g \left(\operatorname {c l i p} \left(a _ {t} + g ^ {- 1} \left(s _ {t}\right), 0, k _ {\max }\right)\right), \tag {3} +$$ + +where $g(k_{t}) = s_{t}$ maps the lens position $k_{t}$ to the state $s_{t}$ at time-step $t$ and $k_{max} \triangleq k_{n}$ is the maximum lens position. The reward function $R(s_{t}, a_{t})$ is defined as the negative absolute distance between the predicted and ground-truth infocus positions, where we additionally introduce a penalty term $R_{fh}$ to mitigate the occurrence of focus hunting: + +$$ +R \left(s _ {t}, a _ {t}\right) = - \left| k _ {t + 1} - k _ {g t} \right| + R _ {f h}, \tag {4} +$$ + +where $k_{gt}$ denotes the ground-truth lens position, and the penalty term $R_{fh}$ is defined as: + +$$ +R _ {f h} = \left\{ \begin{array}{l l} c, & \text {i f F H o c c u r s}; \\ 0, & \text {o t h e r w i s e}; \end{array} \right. \tag {5} +$$ + +where $c$ is a negative hyperparameter that penalizes the event of focus hunting. The main goal of RL is to maximize the expected return $J(\theta)$ defined as: + +$$ +J (\theta) = \mathbb {E} _ {\pi_ {\theta}} \left[ \sum_ {t = 0} ^ {T} \gamma^ {t} R \left(s _ {t}, a _ {t}\right) \right], \tag {6} +$$ + +where $\pi_{\theta}$ is the policy network parameterized by $\theta$ and $T$ denotes the maximum time step. + +Since the single-step prediction network model proposed by Choi et al. [7] has demonstrated a state-of-the-art accuracy, we adopt its network architecture as our actor network. + +![](images/96b66baf09fbbbf481a30e551f6bdf176e716e7ea768cac89c93fdfb152d785b.jpg) +Figure 4. Illustration of relative movement label. + +However, the original model in [7] directly predicts the infocus position, which contradicts the principle of using DP data for in-focus prediction. In essence, since there is a clear correspondence between the disparity and lens movement (i.e., through the defocus conversion coefficient), it is more reasonable to predict the lens movement rather than the lens position based on the DP data. Thus, we modify the original network's output of absolute in-focus position distribution in [7] to the relative lens movement distribution, with the inference procedure of actor network shown in Fig. 9. Accordingly, Fig. 4 illustrates the conversion from original absolute position labels to relative movement labels that are used for training. Unlike [17], we expand the action range from $[- \mathrm{gt}, k_{\max} - \mathrm{gt}]$ to $[-k_{\max}, k_{\max}]$ , and the label denoting the relative movement towards GT index then converts to $\mathrm{gt} - f$ . If the actor network predicts an out-of-bound action, it will be clipped to the closest lower/upper bounds. Based on this formulation, we initialize the actor network with a pre-trained MobileNet-V2 [33] optimized based on the ordinal regression loss [10]. For details about the pretrained model of actor network, please refer to Sec. 7 in our supplementary. + +# 3.4. Critic Network Incorporating Future Impact + +The actor network cannot directly incorporate the future reward of current actions, where the probability of any output action with a larger instant reward will be selected, leading to a suboptimal policy. To address this issue, a critic network is further introduced to provide baselines and evaluate the future impact of the current actions. + +The critic network estimates the value function $V_{\phi}(s_t)$ of the current state $s_t$ , which is defined as the expected future accumulated discount reward of actions. If we further define the future accumulated discount reward of an action under a given state as a Q-function $Q(s_t, a_t)$ , the advantage function $\hat{A}_t = Q(s_t, a_t) - V_{\phi}(s_t)$ then evaluates the relative quality of a particular action under the current state. With this advantage function, the actor network can be trained by minimizing the clipped surrogate objective loss, given by: + +$$ +\mathcal {L} _ {\text {c l i p}} (\theta) = \mathbb {E} \left[ \min \left(r _ {t} (\theta) \hat {A}, \operatorname {c l i p} \left(r _ {t} (\theta), 1 - \epsilon , 1 + \epsilon\right) \hat {A} _ {t}\right), \right. \tag {7} +$$ + +where $r_t(\theta) = \pi_\theta(a_t|s_t) / \pi_\theta^{old}(a_t|s_t)$ is the probability ratio and $\epsilon$ is a clip hyperparameter. + +Algorithm 1 Expert trajectory generator. +Input: $\mathcal{D} = \{f,GT,(s_0,a_0),\dots ,(s_n,a_n)\}_{N};$ +Output: $\mathcal{D}^E$ . +1: $\mathcal{D}^E = \emptyset$ +2: for all $\{g(\cdot),GT,(s_0,a_0),\dots ,(s_n,a_n)\} _i\in \mathcal{D}$ do +3: for all $j = 0,1,2,\dots ,n$ do +4: $k_{j}\gets g_{i}^{-1}(s_{j})$ +5: $o_j\gets \mathrm{sign}(k_0 - GT_i)\cdot |k_j - GT_i| + GT_i;$ +6: $o_j\gets \mathrm{CLIP}(o_j,k_0,GT_i);$ +7: end for +8: $(o_{m_0},\dots ,o_{m_n})\gets \mathrm{sort}(o_0,\dots ,o_n;\mathrm{reverse} =$ $\mathrm{Bool}(k_0 > GT_i));$ +9: for all $j = 1,2,\dots ,n$ do +10: $s_j\gets g_i(o_{m_j});$ +11: $a_{j - 1}\leftarrow o_{m_j} - o_{m_{j - 1}};$ +12: end for +13: $a_{n}\gets 0$ +14: $\mathcal{D}^E\gets \mathcal{D}^E\cup \{g(\cdot),GT,(s_0,a_0),\dots ,(s_n,a_n)\} _i;$ +15: end for +16: return $\mathcal{D}^E$ + +The critic network, on the other hand, is trained by minimizing the $\ell_2$ loss between the estimated value function $V_{\phi}(s_t)$ and the true value function $\hat{V}_t$ : + +$$ +\mathcal {L} _ {\text {c r i t i c}} = \frac {1}{N} \sum_ {t = 1} ^ {N} \left(V _ {\phi} \left(s _ {t}\right) - \hat {V} _ {t}\right) ^ {2}. \tag {8} +$$ + +# 3.5. Expert Trajectory Regularization + +Though the incorporation of a critic network can improve the effectiveness of actor network's training, the exponential growth in the state and action combinations inherently leads to another issue of low sample efficiency in DRL. Recent literature has witnessed the success of incorporating offline data for improving the sample efficiency of DRL's training, both theoretically [38, 45] and empirically [5, 27, 28]. Motivated by this, and to further stabilize the DRL with a lower occurrence of focus hunting in the resulting lens movement trajectory, we generate some offline trajectories based on the prior knowledge to avoid focus hunting, which are then leveraged as an offline dataset $\mathcal{D}^E = \{(s_t,a_t^E)\}$ of expert trajectories to regularize the actor network's training. Algorithm 1 shows our proposed method to generate these expert trajectories, with more details provided in Sec. 8 in our supplementary. + +Fig. 5 shows our proposed framework for expert trajectory regularized DRL's training. At each time, we sample $N$ trajectories from the offline expert dataset $\mathcal{D}^E$ to calculate the imitation loss: + +$$ +\mathcal {L} _ {i m i t} (\theta) = - \frac {1}{N} \sum_ {i = 1} ^ {N} \sum_ {t = 1} ^ {T} \log p \left(\pi_ {\theta} (\cdot | s _ {i, t}) = a _ {i, t}\right), \tag {9} +$$ + +![](images/a239323905a5adb0550695bb67ac25530096eb6c4600b962e140871a804b90b6.jpg) +Figure 5. Expert trajectory regularized DRL's training. + +where $a_{i,t} = k_{i,t + 1} - k_{i,t}$ denotes the action at time step $t$ in the $i$ -th trajectory, $s_{i,t}$ represents the network input at lens location $k_{i,t}$ , and $\pi_{\theta}(\cdot | s_{i,t})$ represents the output distribution of the actor network, with the total loss defined as: + +$$ +\mathcal {L} (\theta) = \mathcal {L} _ {\text {c l i p}} (\theta) + \lambda \cdot \mathcal {L} _ {\text {i m i t}} (\theta), \tag {10} +$$ + +where $\lambda$ balances the online and offline learning. By minimizing the loss $\mathcal{L}(\theta)$ , our DRL algorithm maximizes the expected accumulated discount reward $J(\theta)$ indirectly. + +# 4. Experiments + +In the experiments, our actor and critic networks are built upon MobileNetV2 [33]. We reproduce the models for [16] and [7], since no open-source code is available. In the multi-step prediction setting, we only consider the model obtained by [7], since the multi-step model given by [16] must be restrained for each step, which is impractical. More results and analysis are provided in Sec. 9 in our supplementary. + +Dataset. We adopt the publicly available AF dataset constructed by Herrmann et al. [16]. This dataset comprises 51 scenes, each containing 10 stacks with varying compositions. All instances within the dataset are captured at $n = 49$ focal depths, uniformly sampled in the inverse depth space, ranging from 0.102 to 3.91 meters, captured using the Google Pixel 3 smartphones. To create the training and test sets, we extract and sample $128 \times 128$ patches with a stride of 96 from the central camera, resulting in a total of 3,341,163 patches for the training set and 382,445 patches for the test set. Patches with the same spatial position within one image and 49 different focal depths constitute a single focal stack. Please refer to [16] for additional details about the dataset, and refer to Sec. 6 in our supplementary for the labeling method adopted in this paper. + +Evaluation metric. We follow the evaluation metric proposed by Scharstein et al. [34], which is used for Middlebury stereo dataset and adopted by Herrmann et al. [16] and Choi et al. [7] to evaluate the AF model's prediction accuracy. In particular, this metric is calculated by the proportion of patches whose stop positions are within 0, 1, 2, and 4 indices away from the in-focus index, which can be considered in the depth of field (DoF). The mean absolute error (MAE) and root mean squared error (RMSE) are also + +used to evaluate the AF model's performance. To evaluate the focus hunting (FH), we retain every focusing trajectory and calculate the direction and distance of each lens movement. If the direction of a current movement is opposite to that of the previous one in the trajectory, we consider an occurrence of FH in that trajectory. Since the actor network obtained by our DRL and the model in [7] share the same architecture and similar input/output, we can simply regard the number of steps to reach the in-focus position as the focus speed. We calculate the images captured at all initial positions of the lens to account for accuracy and FH rate. + +Implementation detail. We use PyTorch [29] and TorchRL [4] for our implementation. For deep learning models, we use the Adam optimizer [23] with a learning rate of $10^{-3}$ , $(\beta_{1},\beta_{2}) = (0.5,0.999)$ , a batch size of 128, and cosine learning rate decay [26] for 10k iterations. For reinforcement learning models, we adopt the PPO-CLIP [35] algorithm. The initial learning rate is set to $1\times 10^{-5}$ , and the clip epsilon is 0.2, a batch size of 32, Adam optimizer with the same $(\beta_{1},\beta_{2})$ as deep learning. The maximum episode length of RL is set to 4. For the expert trajectory, we first construct an expert dataset containing 6,000,000 expert trajectories and set the weight $\lambda$ to 0.001 in the loss function. + +# 4.1. Quantitative Results + +FH rate and multi-step prediction accuracy (Effects of DRL). The FH rate and multi-step prediction accuracy of comparison methods are shown in Tab. 1. Though the relative movement label can significantly improve the one-step prediction accuracy, the FH rate in the multi-step setting is still high since the current deep-learning training process cannot obtain and exploit the sequential information of the lens movement. Due to DRL's advantages in modeling sequential decision-making problems, our DRL-based models can directly decrease the FH rate without relying on any extra networks, such as the LSTM [36]. Compared with the policy obtained by deep learning, our DRL-based policy can decrease the FH rate by $9.1\%$ in 2-step prediction, $16.5\%$ in 3-step prediction, and $18.0\%$ in 4-step prediction, while adopting the same network architecture. Our work shows some advantages over existing RL baselines. Since few features are extracted from the scalar input of [6], it can only guide the lens towards one side with the higher label frequency, causing extremely low accuracy but 0 FH rate. Since [54] generalizes poorly in various views (as stated by the authors), while the dataset in our work contains images with multiple views and object distances, it hardly gives a move signal. The lens thus stops at the start position in most cases, causing small changes in accuracy and a low FH rate. + +For the multi-step prediction performance, the accuracy of the baseline [7] increases a little and even decreases in 4-step prediction. The accuracy of the model obtained by the relative movement label can achieve substantial gains + +Table 1. Prediction accuracy and focus hunting (FH) rate results in multi-step setting. + +
AlgorithmStep= 0 ↑ ≤ 1 ↑ ≤ 2 ↑ ≤ 4 ↑MAE↓ RMSE↓FH↓
DRL in [6]10.030 0.071 0.117 0.21014.216 17.908NA
20.029 0.050 0.082 0.15117.214 20.8270.0
30.029 0.050 0.082 0.15117.214 20.8270.0
40.029 0.050 0.082 0.15117.214 20.8270.0
DRL in [54]10.022 0.062 0.102 0.18115.575 19.143NA
20.023 0.064 0.105 0.18515.446 19.0510.019
30.024 0.065 0.106 0.18715.437 19.0730.025
40.024 0.065 0.106 0.18615.500 19.1420.030
RoI-PE & Len-PE [7]10.243 0.584 0.756 0.8822.839 6.880NA
20.275 0.640 0.803 0.9122.454 6.2960.275
30.279 0.644 0.810 0.9142.434 6.2050.362
40.276 0.642 0.807 0.9132.442 6.2770.383
[7] with relative label (Ours)10.330 0.698 0.826 0.9072.342 6.339NA
20.413 0.795 0.888 0.9341.894 5.8550.261
30.427 0.805 0.894 0.9371.892 5.7210.340
40.423 0.804 0.893 0.9361.862 5.8440.358
Expert Trajectory (Ours)10.336 0.711 0.839 0.9152.146 5.827NA
20.439 0.808 0.896 0.9411.755 5.5970.170
30.450 0.816 0.900 0.9431.711 5.5320.175
40.449 0.818 0.901 0.9441.694 5.4790.178
+ +Table 2. Average steps to stop with different termination signals. + +
Termination Rule= 0 ↓ ≤ 1 ↓ ≤ 2 ↓ ≤ 4 ↓Stop Model↓ MAE↓ RMSE↓
RoI-PE & Lens-PE[7]6.6513.6442.4771.7072.2082.6116.146
[7] with relative label (Ours)4.9822.2461.7041.4531.8841.9065.463
Expert Trajectory (Ours)5.0142.2571.6831.3611.8211.8745.332
+ +of $13.8\%$ prediction accuracy for exact $(= 0)$ predictions in 2-step prediction, while our DRL-based model further enhances this accuracy. Furthermore, as the number of lens movement steps increases, the accuracy of deep learning-based methods may decline, whereas the accuracy of our DRL-based method remains unchanged or even improves, demonstrating the stability of our DRL-based model in multi-step prediction. Moreover, Tab. 2 shows that our models take fewer steps to stop at different termination signals, e.g., $= 0$ , $\leq 1$ , $\leq 2$ , and $\leq 4$ . Since in-focus positions remain unknown in practical camera deployment, the above rules may not be feasible. Alternatively, we train a stop model to generate the stop signal, which is modeled as a binary classification task and takes the same architecture and input as our policy network. Tab. 2 shows the average steps, MAE, and RMSE of different policies with this model. Our method still achieves a better prediction with fewer steps. In conclusion, results in Tab. 1 and Tab. 2 show that our methods can achieve higher accuracy with fewer steps than that proposed by [7], demonstrating a faster focusing speed. + +Single-step accuracy (Effects of relative movement as labels). The comparison results in the single-step permutation setting are shown in Tab. 3. Compared with classic statistics-based or depth estimation methods, the deep-learning based models "Learning to Autofocus [16]" and "RoI-PE & Len-PE [7]" significantly improve the one-step prediction accuracy. By modifying the output of the model in [7] from the absolute lens position to the relative movement, we can further improve the accuracy by $8.7\%$ for exact $(= 0)$ prediction and $2.5\%$ for $\leq 4$ focal index difference. Moreover, since the dataset has low-contrast images with similar features but distant labels, minimizing the loss makes the policy perform badly. Via optimizing multi-step + +Table 3. Single-step prediction results. + +
Algorithm= 0 ↑≤ 1 ↑≤ 2 ↑≤ 4 ↑MAE↓RMSE↓
ZNCC Disparity with Calibration0.0650.1790.2680.4199.47614.211
SSD Disparity [44]0.1010.2640.3860.5258.01212.414
Learning the Depth [11]0.1120.2970.4150.5627.41912.312
Learning to Autofocus [16]0.1680.4670.6590.8513.2497.619
RoI-PE & Lens-PE [7]0.2430.5840.7560.8822.8396.880
[16] with relative label (Ours)0.1810.4930.6840.8613.0217.374
[7] with relative label(Ours)0.3300.6980.8260.9072.3426.339
Expert Trajectory (Ours)0.3360.7110.8390.9152.1465.827
+ +Table 4. Performance comparison of DRL models with and without Expert Regularization. + +
AlgorithmStep= 0 ↑ ≤ 1 ↑ ≤ 2 ↑ ≤ 4 ↑MAE↓ RMSE↓FH↓
w/o10.0310.0560.0860.14322.00526.048
20.0290.0650.0950.14823.84527.975
Expert Regularization30.0370.0850.1230.18820.95725.501
40.0310.0680.0990.15223.72227.889
Expert Trajectory (Ours)10.3360.7110.8390.9152.1465.827
20.4390.8080.8960.9411.7555.597
30.4500.8160.9000.9431.7115.532
40.4490.8180.9010.9441.6945.479
+ +behavior, RL can explore extra information and improve 1-step prediction accuracy. + +Effects of expert regularization. As shown in Tab. 4, the policy obtained by the vanilla DRL algorithm without expert regularization has a poor performance, since the action space is a discrete number of 97, which is large for DRL's training. The expert trajectory significantly improves the performance of the policy given by the PPO algorithm. + +# 4.2. Qualitative Results + +Fig. 1 qualitatively visualizes the trajectory of patches and lens positions of different comparison methods, including the policy obtained by [7], the policy obtained by relative movement label, and our DRL-based model. Compared with directly predicting the lens position [7], our relative movement label model is faster by predicting the lens movement within DoF in the second step, but the FH issue cannot be overcome. In comparison, our DRL-based model can generate a faster and more stable prediction by predicting the GT location at the first step and stopping at this index. The lens trajectory of the DRL-based model (the green curve in Fig. 2) is slightly smoother and more stable without the occurrence of FH. + +In Fig. 6, we show qualitative results of 4-step predictions of the three models. From the first row, we can observe that all methods can predict the index within DoF, but our methods can give a more precise prediction with a sharper image than Choi's [7]. From the second to fourth rows, we can observe that [7] sometimes breaks and predicts indices with sharp background, while our DRL-based model successfully predicts indices with sharp target objects. This case is a generally challenging scenario in the AF task since many different DoFs can be estimated. These DoFs may confuse the models, leading to a focus on the wrong object. In general, our relative movement label can give a more accurate prediction by mapping the disparity to the movement of the lens, and the DRL-based model can further enhance the accuracy by estimating the future reward of the action + +![](images/189850b0d961d43ad0fee67ed1398510478dcd6c775f823cf628308068b84b05.jpg) + +![](images/7ef550bb2c179357b8f414db0273bc6e558ed66aee57c8d1f98e4ef91e791643.jpg) + +![](images/10775c45e94d5fdf3bf7cd35da086e3acb78975b9234083d280a2b8a4e127c7b.jpg) + +![](images/89e439505bb9e65d78ff78db058b701e14f2bcae2c6524991ebd201afadacf45.jpg) +Input +Figure 6. Comparison of the 4-step results of the policy obtained by [7], our relative movement label, and our DRL-based model. + +![](images/78f6eece17673bd4dd2c22ce6c3074b32640f19a513a6426179e138f10f064f4.jpg) + +![](images/b82c6818fc24ce9026b74b5f761c40a0fd2625c22ebdc4e9bdf22c3c95d56232.jpg) + +![](images/ea241c0050c70f1f2808b8fe0ad410f67a9769b65aaa20932efdc68a895e1902.jpg) + +![](images/bf2fe82be806fc3c087b64d8d6ff8ec3e252dd80601d8e358b8e69828ae4cbb0.jpg) +absolute position labe + +![](images/efafaabf2fa673ff7cc6bd3175916ff61936d57e1b0fff569d99ed395e28e96c.jpg) + +![](images/4b8cd13827cfd0c97b2b318dd89c3cddc8ea7552c92a67969fcfeee0aaed2526.jpg) + +![](images/eefd1c948dfb5ffecb4862cbf822041a769d9a2f22a7e0d666b8a2cc4e95ce68.jpg) + +![](images/fde21e27e4ef700bf2ed5ed3d8139c3a2804b3b86cf0a484b005881104cfef45.jpg) +relative movement label + +![](images/d3c97875496f60449e9e4f77cea564696ebceb14b2186a12b8b80ec1fe6a8eab.jpg) + +![](images/857994789c5078b95175cc548f5a0834d53b1f55f0076d34a3d3a359d2ea8664.jpg) + +![](images/96d52a5abbf949dc102f24a0e067b608f2ba4dfb926f9fc923e2c76b1f8cc5bd.jpg) + +![](images/467450180faa395da3e2eb6560da926580676ff3e1e4441175731705dce5e67f.jpg) +DRL + +![](images/0ad31ed71439c73d490282239f8ce6e06935e40d55b7903e2b878bb3525e7b7b.jpg) + +![](images/2ab5c69b8f5de88f2d319a32fd4a8279eb7e953672395b388e51fb2b27f871df.jpg) + +![](images/b8768a93885661caf0cb400a7a38f503ddc89590a6bc785298a856f12178f9a8.jpg) + +![](images/86b1d4250541142213e2df685a1f15c842c7313c19d2acebe97fbce72561f8c8.jpg) +GT + +and reducing the focus hunting rate. + +In Fig. 7, we analyze our model's failure cases: (a) it is difficult and meaningless to focus on the images with no texture since the phase information or contrast information is little to estimate the depth, (b) though our models can give a more accurate prediction in scenarios with multiple DoFs, when the target object is small and the background is complex, the models may be confused by these DoFs, leading to focusing on the wrong object, (c) like all DP data-based AF methods, our models may work poorly when the objects are horizontal. Some possible solutions may work to overcome these limitations. For (a), larger patches can provide more phase information, and we also think (a) might not be that important since few people take photos with no texture. For (b), we claim it is not a bad issue, because sometimes people may want to focus on the background obscured by meaningless foreground elements. To solve this problem, larger patches with more information about the target objects may be useful. Some depth estimation methods [22, 52] can also be used to provide more depth information to help distinguish the objects and background. For (c), solving this problem is challenging in algorithms, while little changes in hardware, especially sensors, could easily mitigate this issue [7]. + +# 4.3. Ablation Study + +Effects of relative label method. Tab. 5 further demonstrates the effect of relative movement as labels. Our DRL-based model can further enhance the accuracy by $0.6\%$ for exact $(= 0)$ prediction and $0.8\%$ for $\leq 4$ focal index difference, in comparison to the model in [7] modified with the relative movement label. We further compare the two label methods under three different architectures: MobileNetV2 [33], ResNet50 [15], and MnasNet [41]. The relative movement label method outperforms the absolute lens position label for each architecture. For a fairer comparison, + +![](images/41f4ba9be68136a6609bde276a95e00abe544c901bb98dd5d3da776f3231a3a7.jpg) + +![](images/a8df26c18e23a57b065915fd84b4a91fcf1d37d148b5b09cbcf383b753c72c37.jpg) + +![](images/876be92ae60c1986ec51e9bb29655bb9209f2bdafc61c731b42d68f7940ff435.jpg) +Input +Figure 7. Failure cases of our models, where only the left image of DP data is visualized. + +![](images/3c0fb69380e47e9ddba384a23828fd7424aa366bfbffb1d5a684797ebcc48a87.jpg) + +![](images/c1f40623d2d637e11ecf939f88911970f2e5c0a32aace19802722016f9604c9e.jpg) + +![](images/9a25ae0cd0e8e9778aa4007d1f036074b6b133749f7dff5dc13e1ca43fa664af.jpg) +relative movement label + +![](images/b0dc5b5557066efbeb75091808f0eeb0a4e2c71116d6eb274e59ba5ea3059ba5.jpg) + +![](images/fd94b8464e74e9842e7df2e9f7c6e616b7df34c862bdee3d979d4cfea7bc5620.jpg) + +![](images/da7145b1c0ae1c4e6e240e953efa5b4a104a0134b22a66292415e251c55142b8.jpg) +DRL + +![](images/198eadb473dab9abd3b30e2046853b3f12eedb4fb0a39e1a0562a6214deac023.jpg) + +![](images/5bbeb3708e3d335caad1d7198299f0958d57b97c85048ec9876e845a185b2ce5.jpg) + +![](images/0b567bd653370c2878ce35fdbe574a2b7399da2794508b30548a650fdf25dd1a.jpg) +GT + +Table 5. Single-step prediction results with different network architectures. The number of input channels is 5. + +
Algorithm# of paramslabel= 0 ↑ ≤ 1 ↑ ≤ 2 ↑ ≤ 4 ↑MAE↓RMSE↓
MobileNetV2 [33]2.29Mabsolute0.2430.5840.7560.8822.8396.880
2.35Mabsolute0.2520.5920.7640.8882.7196.745
2.35Mrelative0.3300.6980.8260.9072.3426.339
ResNet50 [15]23.61Mabsolute0.2580.6020.7890.8942.6156.623
23.71Mabsolute0.2540.5960.7810.8912.6556.686
23.71Mrelative0.3260.7090.8490.9182.1365.764
MnasNet [41]1.00Mabsolute0.2260.5680.7360.8642.9837.016
1.06Mabsolute0.2130.5490.7180.8503.1197.318
1.06Mrelative0.3050.6780.8170.9052.4366.402
+ +we expand the absolute position label to the same range as the relative movement label. It shows that expanding the number of networks' output channels can not improve the accuracy, and the effects remain unknown. Since Lens-PE mainly conveys the lens' current position to the network, while the next position must be predicted by adding its movement to the current position, the relative label methods avoid additionally inferring the identity mapping for Lens-PE, leading to a higher accuracy. To make a fair comparison with the other learning-based models, we thus adopt the same MobileNetV2 architecture in the experiments. + +# 5. Conclusion + +We have proposed a DRL-based autofocus model with expert trajectory regularization that can both improve the prediction accuracy and reduce the focus hunting rate, which took an actor-critic framework based on the pretrained MobileNet-V2 architecture for single-step prediction. For the actor network, we further modified the absolute position label in [7, 16] with the relative movement label method. We also constructed an offline dataset of expert trajectories based on prior knowledge, regularizing the policy network training toward reducing focus hunting. Experimental results have clearly demonstrated the effectiveness of our proposed method, in terms of the multi-step prediction accuracy and the occurrence of focus hunting. + +# Acknowledgments + +# References + +[1] Abdullah Abuolaim, Mauricio Delbracio, Damien Kelly, Michael S. Brown, and Peyman Milanfar. Learning to reduce defocus blur by realistically modeling dual-pixel data. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 2289-2298, 2021. 1, 3 +[2] Yuhan Bao, Lei Sun, Yuqin Ma, Diyang Gu, and Kaiwei Wang. Improving fast auto-focus with event polarity. Optics Express, 31(15):24025-24044, 2023. 1 +[3] Daniel I Barnea and Harvey F Silverman. A class of algorithms for fast digital image registration. IEEE transactions on Computers, 100(2):179-186, 1972. 1, 2 +[4] Albert Bou, Matteo Bettini, Sebastian Dittert, Vikash Kumar, Shagun Sodhani, Xiaomeng Yang, Gianni De Fabritis, and Vincent Moens. TorchRL: A data-driven decision-making library for pytorch. arXiv preprint arXiv:2306.00577, 2023. 6 +[5] Serkan Cabi, Sergio Gómez Colmenarejo, Alexander Novikov, Ksenia Konyushkova, Scott Reed, Rae Jeong, Konrad Zolna, Yusuf Aytar, David Budden, Mel Vecerik, et al. Scaling data-driven robotics with reward sketching and batch reinforcement learning. arXiv preprint arXiv:1909.12200, 2019. 5 +[6] Chin-Cheng Chan and Homer H Chen. Autofocus by deep reinforcement learning. *Electronic Imaging*, 31:1-6, 2019. 3, 6, 7 +[7] Myungsub Choi, Hana Lee, and Hyong-Euk Lee. Exploring positional characteristics of dual-pixel data for camera autofocus. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 13112-13122, 2023. 2, 3, 4, 5, 6, 7, 8, 1 +[8] Murtaza Dalal, Deepak Pathak, and Russ R Salakhutdinov. Accelerating robotic reinforcement learning via parameterized action primitives. In Advances in Neural Information Processing Systems (NeurIPS), pages 21847-21859, 2021. 2 +[9] Judith Dijk, Michael van Ginkel, Rutger J van Asselt, Lucas J van Vliet, and Piet W Verbeek. A new sharpness measure based on gaussian lines and edges. In 10th International Conference on Computer Analysis of Images and Patterns (CAIP), pages 149-156, 2003. 3 +[10] Raul Diaz and Amit Marathe. Soft labels for ordinal regression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4733-4742, 2019. 5, 2 +[11] Rahul Garg, Neal Wadhwa, Sameer Ansari, and Jonathan T Barron. Learning single camera depth estimation using + +This work was supported in part by the National Natural Science Foundation of China under Grant 62125109, Grant 62431017, Grant U24A20251, Grant 62320106003, Grant 62371288, Grant 62401357, Grant 62401366, Grant 62301299, Grant 62120106007, and in part by the Program of Shanghai Science and Technology Innovation Project under Grant 24BC3200800. (Corresponding authors: Chenglin Li, Nuowen Kan, Junni Zou.) +dual-pixels. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 7628-7637, 2019. 3, 7, 1 +[12] Cole Gulino, Justin Fu, Wenjie Luo, George Tucker, Eli Bronstein, Yiren Lu, Jean Harb, Xinlei Pan, Yan Wang, Xiangyu Chen, John Co-Reyes, Rishabh Agarwal, Rebecca Roelofs, Yao Lu, Nico Montali, Paul Mougin, Zoey Yang, Brandyn White, Aleksandra Faust, Rowan McAllister, Dragomir Anguelov, and Benjamin Sapp. Waymax: An accelerated, data-driven simulator for large-scale autonomous driving research. In Advances in Neural Information Processing Systems (NeurIPS), pages 7730-7742, 2023. 2 +[13] Marsha Jo Hannah. Computer matching of areas in stereo images. Stanford University, 1974. 1, 2 +[14] Jie He, Rongzheng Zhou, and Zhiliang Hong. Modified fast climbing search auto-focus algorithm with adaptive step size searching technique for digital camera. IEEE Trans. Consumer Electron., 49(2):257-262, 2003. 3 +[15] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 770-778, 2016. 8 +[16] Charles Herrmann, Richard Strong Bowen, Neal Wadhwa, Rahul Garg, Qiurui He, Jonathan T. Barron, and Ramin Zabih. Learning to autofocus. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2227-2236, 2020. 2, 3, 6, 7, 8, 1 +[17] Chi-Jui Ho, Chin-Cheng Chan, and Homer H Chen. A net: A convolutional neural network approach to phase detection autofocus. IEEE Transactions on Image Processing, 29:6386-6395, 2020. 3, 5, 1 +[18] Jui-Ting Huang, Chun-Hung Shen, See-May Phoong, and Homer Chen. Robust measure of image focus in the wavelet domain. In 2005 International Symposium on Intelligent Signal Processing and Communication Systems, pages 157-160, 2005. 3 +[19] ZHENG Jin, JIANG Botao, PENG Wei, and ZHANG Qiaohui. Multi-scale binocular stereo matching based on semantic association. Chinese Journal of Electronics, 33(4):1010-1022, 2024. 3 +[20] Junpeng Jing, Jiankun Li, Pengfei Xiong, Jiangyu Liu, Shuaicheng Liu, Yichen Guo, Xin Deng, Mai Xu, Lai Jiang, and Leonid Sigal. Uncertainty guided adaptive warping for robust and efficient stereo matching. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 3318-3327, 2023. 3 +[21] Neel Joshi and C Lawrence Zitnick. Micro-baseline stereo. Microsoft Research Technical Report, 2014. 3 +[22] Donggun Kim, Hyeonjoong Jang, Inchul Kim, and Min H. Kim. Spatio-focal bidirectional disparity estimation from a dual-pixel image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5023-5032, 2023. 3, 8 +[23] Diederik P Kingma. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6 +[24] Matej Kristan, Janez Pers, Matej Perse, and Stanislav Kovačić. A bayes-spectral-entropy-based measure of camera + +focus using a discrete cosine transform. Pattern Recognition Letters, 27(13):1431-1439, 2006. 3 +[25] Sang-Yong Lee, Jae-Tack Yoo, Yogendera Kumar, and SooWon Kim. Reduced energy-ratio measure for robust autofocusing in digital camera. IEEE Signal Processing Letters, 16 (2):133-136, 2009. 3 +[26] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983, 2017. 6 +[27] Yao Lu, Karol Hausman, Yevgen Chebotar, Mengyuan Yan, Eric Jang, Alexander Herzog, Ted Xiao, Alex Irpan, Mohi Khansari, Dmitry Kalashnikov, et al. Aw-opt: Learning robotic skills with imitation and reinforcement at scale. In Conference on Robot Learning, pages 1078-1088, 2022. 5 +[28] Ashvin Nair, Abhishek Gupta, Murtaza Dalal, and Sergey Levine. Awac: Accelerating online reinforcement learning with offline datasets. arXiv preprint arXiv:2006.09359, 2020. 5 +[29] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Kopf, Edward Yang, Zachary DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. Pytorch: An imperative style, high-performance deep learning library. In Advances in Neural Information Processing Systems (NeurIPS), 2019. 6 +[30] José Luis Pech-Pacheco, Gabriel Cristóbal, Jesús Chamorro-Martinez, and Joaquín Fernández-Valdivia. Diatom autofocusing in brightfield microscopy: a comparative study. In Proceedings 15th International Conference on Pattern Recognition. ICPR-2000, pages 314-317. IEEE, 2000. 1 +[31] Henry Pinkard, Zachary Phillips, Arman Babakhani, Daniel A Fletcher, and Laura Waller. Deep learning for single-shot autofocus microscopy. Optica, 6(6):794-797, 2019. 3 +[32] Stéphane Ross, Geoffrey Gordon, and Drew Bagnell. A reduction of imitation learning and structured prediction to no-regret online learning. In Proceedings of the fourteenth international conference on artificial intelligence and statistics, pages 627-635. JMLR Workshop and Conference Proceedings, 2011. 4 +[33] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zhmoginov, and Liang-Chieh Chen. *Mobilenetv2: Inverted residuals and linear bottlenecks*. In *Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition* (CVPR), pages 4510-4520, 2018. 2, 5, 6, 8 +[34] Daniel Scharstein and Richard Szeliski. A taxonomy and evaluation of dense two-frame stereo correspondence algorithms. International Journal of Computer Vision, 47:7-42, 2002. 6 +[35] John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347, 2017. 6 +[36] Xingjian SHI, Zhourong Chen, Hao Wang, Dit-Yan Yeung, Wai-kin Wong, and Wang-chun WOO. Convolutional LSTM Network: A Machine Learning Approach for Precipitation + +Nowcasting. In Advances in Neural Information Processing Systems (NeurIPS), 2015. 6 +[37] Przemysław Sliwinski and Paweł Wachel. A simple model for on-sensor phase-detection autofocusing algorithm. Journal of Computer and Communications, 1(06):11, 2013. 3 +[38] Yuda Song, Yifei Zhou, Ayush Sekhari, Drew Bagnell, Akshay Krishnamurthy, and Wen Sun. Hybrid RL: Using both offline and online data can make RL efficient. In International Conference on Learning Representations (ICLR), 2023. 5 +[39] Changqing Su, Zhiyuan Ye, Yongsheng Xiao, You Zhou, Zhen Cheng, Bo Xiong, Zhaofei Yu, and Tiejun Huang. Intensity-robust autofocus for spike camera. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 25018-25027, 2024. 3 +[40] Murali Subbarao, Tae-Sun Choi, and Arman Nikzad. Focusing techniques. Optical Engineering, 32(11):2824-2836, 1993. 1, 3 +[41] Mingxing Tan, Bo Chen, Ruoming Pang, Vijay Vasudevan, Mark Sandler, Andrew Howard, and Quoc V. Le. Mnasnet: Platform-aware neural architecture search for mobile. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2815-2823, 2019. 8 +[42] Jay Martin Tenenbaum. Accommodation in computer vision. Stanford University, 1971. 1, 3 +[43] Quoc Kien Vuong and Jeong-won Lee. Initial direction and speed decision system for auto focus based on blur detection. In 2013 IEEE International Conference on Consumer Electronics (ICCE), pages 222-223, 2013. 1 +[44] Neal Wadhwa, Rahul Garg, David E Jacobs, Bryan E Feldman, Nori Kanazawa, Robert Carroll, Yair Movshovitz-Attias, Jonathan T Barron, Yael Pritch, and Marc Levoy. Synthetic depth-of-field with a single-camera mobile phone. ACM Transactions on Graphics (ToG), 37(4):1-13, 2018. 7 +[45] Andrew Wagenmaker and Aldo Pacchiano. Leveraging offline data in online reinforcement learning. arXiv preprint arXiv:2211.04974, 2023. 5 +[46] Chengyu Wang, Qian Huang, Ming Cheng, Zhan Ma, and David J. Brady. Deep learning for camera autofocus. IEEE Transactions on Computational Imaging, 7:258-271, 2021. 3 +[47] Hui Xie, Weibin Rong, and Lining Sun. Wavelet-based focus measure and 3-d surface reconstruction method for microscopy images. In Proceedings 2006 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS 2006), pages 229-234, 2006. 3 +[48] Takahiro Yamasaki, Tomohiro Nakamura, Ryohei Funatsu, and Hiroshi Shimamoto. Hybrid autofocus system by using a combination of the sensor-based phase-difference detection and focus-aid signal. In 2018 IEEE International Conference on Consumer Electronics (ICCE), pages 1–2, 2018. 1 +[49] Jingyang Yan, Peter DiMeo, Lu Sun, and Xian Du. LSTM-Based Model Predictive Control of Piezoelectric Motion Stages for High-Speed Autofocus. IEEE Transactions on Industrial Electronics, 70(6):6209-6218, 2023. 3 +[50] Ge Yang and Bradley J Nelson. Wavelet-based autofocusing and unsupervised segmentation of microscopic images. + +In Proceedings 2003 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS 2003), pages 2143-2148, 2003. 3 +[51] Hyun Jong Yang, Moohyun Oh, Jonggyu Jang, Hyeonsu Lyu, and Junhee Lee. Robust deep-learning based autofocus score prediction for scanning electron microscope. Microscopy and Microanalysis, 26(S2):702-705, 2020. 3 +[52] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10371-10381, 2024. 8 +[53] Yi Yao, Besma Abidi, Narjes Doggaz, and Mongi Abidi. Evaluation of sharpness measures and search algorithms for the auto-focusing of high-magnification images. In Visual Information Processing XV, pages 132-143. SPIE, 2006. 3 +[54] Xiaofan Yu, Runze Yu, Jingsong Yang, and Xiaohui Duan. A robotic auto-focus system based on deep reinforcement learning. In 2018 15th International Conference on Control, Automation, Robotics and Vision (ICARCV), pages 204-209, 2018. 3, 6, 7 +[55] Ramin Zabih and John Woodfill. Non-parametric local transforms for computing visual correspondence. In Computer Vision—ECCV'94: Third European Conference on Computer Vision Stockholm, Sweden, May 2–6 1994 Proceedings, Volume II 3, pages 151–158. Springer, 1994. 1, 2 \ No newline at end of file diff --git a/2025/Stabilizing and Accelerating Autofocus with Expert Trajectory Regularized Deep Reinforcement Learning/images.zip b/2025/Stabilizing and Accelerating Autofocus with Expert Trajectory Regularized Deep Reinforcement Learning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..cf8c76f9f30d88d112dced951845902a8338ba85 --- /dev/null +++ b/2025/Stabilizing and Accelerating Autofocus with Expert Trajectory Regularized Deep Reinforcement Learning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4ebacba51048c4e228768025a9cbcc5367e62169b85091cafaf1c15164a40541 +size 507091 diff --git a/2025/Stabilizing and Accelerating Autofocus with Expert Trajectory Regularized Deep Reinforcement Learning/layout.json b/2025/Stabilizing and Accelerating Autofocus with Expert Trajectory Regularized Deep Reinforcement Learning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..e20d650c30569f7b2041663ef3657469b9d4d7d1 --- /dev/null +++ b/2025/Stabilizing and Accelerating Autofocus with Expert Trajectory Regularized Deep Reinforcement Learning/layout.json @@ -0,0 +1,10764 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 73, + 102, + 538, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 102, + 538, + 140 + ], + "spans": [ + { + "bbox": [ + 73, + 102, + 538, + 140 + ], + "type": "text", + "content": "Stabilizing and Accelerating Autofocus with Expert Trajectory Regularized Deep Reinforcement Learning" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 92, + 160, + 518, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 92, + 160, + 518, + 190 + ], + "spans": [ + { + "bbox": [ + 92, + 160, + 518, + 190 + ], + "type": "text", + "content": "Shouhang Zhu1, Chenglin Li1, Yuankun Jiang2, Li Wei1, Nuowen Kan1, Ziyang Zheng1, Wenrui Dai2, Junni Zou2, Hongkai Xiong1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 189, + 477, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 189, + 477, + 203 + ], + "spans": [ + { + "bbox": [ + 133, + 189, + 477, + 203 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 133, + 189, + 477, + 203 + ], + "type": "text", + "content": "Department of Electronic Engineering, Shanghai Jiao Tong University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 103, + 202, + 507, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 103, + 202, + 507, + 217 + ], + "spans": [ + { + "bbox": [ + 103, + 202, + 507, + 217 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 103, + 202, + 507, + 217 + ], + "type": "text", + "content": "Department of Computer Science and Engineering, Shanghai Jiao Tong University" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 92, + 218, + 518, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 92, + 218, + 518, + 232 + ], + "spans": [ + { + "bbox": [ + 92, + 218, + 518, + 232 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 92, + 218, + 518, + 232 + ], + "type": "text", + "content": "shzhu2001, lcl1985, wei.li, kannw.1230, zhengziyang, xionghongkai}@sjtu.edu.cn" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 174, + 232, + 433, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 174, + 232, + 433, + 245 + ], + "spans": [ + { + "bbox": [ + 174, + 232, + 433, + 245 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 174, + 232, + 433, + 245 + ], + "type": "inline_equation", + "content": "\\{\\text { yuankunjiang, daiwenrui, zoujunni }\\} @" + }, + { + "bbox": [ + 174, + 232, + 433, + 245 + ], + "type": "text", + "content": "sjtu.edu.cn" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 151, + 273, + 200, + 285 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 273, + 200, + 285 + ], + "spans": [ + { + "bbox": [ + 151, + 273, + 200, + 285 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 298, + 297, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 298, + 297, + 622 + ], + "spans": [ + { + "bbox": [ + 55, + 298, + 297, + 622 + ], + "type": "text", + "content": "Autofocus is a crucial component of modern digital cameras. While recent learning-based methods achieve state-of-the-art in focus prediction accuracy, they unfortunately ignore the potential focus hunting phenomenon of back-and-forth lens movement in the multi-step focusing procedure. To address this, in this paper, we propose an expert regularized deep reinforcement learning (DRL)-based approach for autofocus, which can utilize the sequential information of lens movement trajectory to both enhance the multi-step in-focus prediction accuracy and reduce the chance of focus hunting. Our method generally follows an actor-critic framework. To accelerate the DRL's training with a higher sample efficiency, we initialize the policy with a pre-trained single-step prediction network, where the network is further improved by modifying the output of absolute in-focus position distribution to the relative lens movement distribution to establish a better mapping between input images and lens movement. To further stabilize DRL's training with a lower occurrence of focus hunting in the resulting lens movement trajectory, we generate some offline trajectories based on prior knowledge to avoid focus hunting, which are then leveraged as an offline dataset of expert trajectories to regularize the actor network's training. Empirical evaluations show that our method outperforms those learning-based methods on public benchmarks, with higher single- and multi-step prediction accuracy, and a significant reduction of focus hunting rate." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 634, + 135, + 646 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 634, + 135, + 646 + ], + "spans": [ + { + "bbox": [ + 56, + 634, + 135, + 646 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 654, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 715 + ], + "type": "text", + "content": "Autofocus (AF) is a process through which the camera can determine automatically a lens position to focus on the region of interest (RoI). It is a crucial step in image capturing, since once an image is already captured in a defocused and blurry state, it is then difficult to restore its sharpness" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 274, + 555, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 274, + 555, + 538 + ], + "spans": [ + { + "bbox": [ + 313, + 274, + 555, + 538 + ], + "type": "text", + "content": "through any post-processing due to the varying effects of blur influenced by factors such as camera aperture and radial distortion [1]. In addition, the beauty of a moment often lies in its fleeting nature, which thus imposes an imperative demand for the fast focusing speed of an AF algorithm. However, due to the instability in focusing systems incurred by some interference factors, the AF mechanism of the camera lens may fail to accurately lock onto the focal point, resulting in a repeated movement of the lens within a certain range. This phenomenon, as illustrated in Fig. 1, is referred to as focus hunting (FH), whereby the camera is unable to identify the focus plane and thus drives the lens forward and backward through a set range of focus hunting for it. It then causes variations in the image sharpness and changes the lens's field of view (FoV), ultimately leading to an unstable focusing experience with longer focusing time, which is a common issue in the traditional AF-assisted cameras [2]. It is thus of great importance to develop an AF algorithm that not only achieves a higher in-focus accuracy but can also stabilize the trajectory of lens movement with a faster convergence speed to the in-focus position, thus reducing the probability of occurrence of focus hunting [43, 48]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 546, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 546, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 546, + 556, + 714 + ], + "type": "text", + "content": "There is a large body of work on the traditional autofocus methods, generally falling into two categories: search-based and phase-based methods. Search-based AF methods [40, 42] search within a range of focus distances to locate the lens position of the extreme sharpness value, which is then determined as the in-focus lens position. Thus, they need to move the lens back and forth, which leads to a highly accurate identification of the in-focus position, but with an annoyingly slow focusing procedure and inevitably the occurrence of focus hunting. Phase-based approaches, on the other hand, rely on the dual-pixel (DP) sensors to obtain the phase data from the left and right sub-images, which is then leveraged to predict the lens movement and determine the in-focus position in a single step of lens" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "spans": [ + { + "bbox": [ + 145, + 0, + 496, + 37 + ], + "type": "text", + "content": "This CVPR paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "26440" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 62, + 553, + 211 + ], + "blocks": [ + { + "bbox": [ + 58, + 62, + 553, + 211 + ], + "lines": [ + { + "bbox": [ + 58, + 62, + 553, + 211 + ], + "spans": [ + { + "bbox": [ + 58, + 62, + 553, + 211 + ], + "type": "image", + "image_path": "ab10fbb75f3bcafaf8e897d27a362794ad1a3d8259bc6e43592c067a2300ce66.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 215, + 555, + 282 + ], + "lines": [ + { + "bbox": [ + 55, + 215, + 555, + 282 + ], + "spans": [ + { + "bbox": [ + 55, + 215, + 555, + 282 + ], + "type": "text", + "content": "Figure 1. Visualization of the focus hunting phenomenon. The lens moves back and forth within a certain range, resulting in a slight change of field of view (FoV) and sharpness. The red, purple, and green lines mark the trajectory of the same target's locations in the picture obtained by three models. The yellow line is a horizontal line as a location reference. (1) The trajectory generated by the learning-based method [7] that predicts the lens position directly. (2) The trajectory generated by the learning-based method [7] that predicts the relative movement of lens. (3) The trajectory generated by our DRL-based method. It can be seen that the FOV and sharpness in trajectory (1) change more significantly, while those in trajectory (2) change slightly, and the lens terminates at the GT after Step 2 in trajectory (3)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 58, + 292, + 294, + 353 + ], + "blocks": [ + { + "bbox": [ + 58, + 292, + 294, + 353 + ], + "lines": [ + { + "bbox": [ + 58, + 292, + 294, + 353 + ], + "spans": [ + { + "bbox": [ + 58, + 292, + 294, + 353 + ], + "type": "image", + "image_path": "bf126b2aeedf7e1e82303c6fbbd7ad57427a5c0ede5e53915a4b1e2a6ed3fc4a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 355, + 295, + 400 + ], + "lines": [ + { + "bbox": [ + 55, + 355, + 295, + 400 + ], + "spans": [ + { + "bbox": [ + 55, + 355, + 295, + 400 + ], + "type": "text", + "content": "Figure 2. Trajectories generated by predicting the lens position [7] (red line), predicting the relative movement (purple line), and DRL (green line). The trajectory generated by our DRL is the smoothest, indicating a stable and fast focusing procedure." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 401, + 295, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 401, + 295, + 484 + ], + "spans": [ + { + "bbox": [ + 55, + 401, + 295, + 484 + ], + "type": "text", + "content": "movement. Though accelerating the focusing procedure, the phase data obtained by sensors is highly susceptible to noise, leading to a relatively low in-focus accuracy. Especially in some extreme cases, e.g., the low-light conditions, the phase detection may need to be executed multiple times at different lens positions to predict a more accurate in-focus position, still at the risk of focus hunting." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 486, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 486, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 486, + 295, + 713 + ], + "type": "text", + "content": "Recent attempts [7, 16] have concentrated on addressing the above trade-off between accuracy and speed, by exploiting the expressive power of neural networks to directly learn a mapping from the captured image at any lens position to the predicted in-focus position. However, when we directly treat the learned neural network as a lens movement policy and apply it at multiple steps to adjust the lens movement, we notice that some of the resulting trajectories still suffer from the back-and-forth movement of the lens, as shown in Fig. 1 and Fig. 2, due to the lack of leveraging the sequential information obtained from multiple exploration steps and the ignorance of considering future effects of the current prediction. Deep reinforcement learning (DRL), with successful applications in sequential decision-making tasks such as autonomous driving [12] and robot control [8], has enabled the agent to adaptively adjust its policy through iterative interactions with the environment, based both on the immediate reward of current decision step and the long-term reward over the entire sequence of decisions." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 293, + 555, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 293, + 555, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 293, + 555, + 640 + ], + "type": "text", + "content": "Inspired by this, we propose in this paper an expert regularized DRL-based approach for camera autofocus, which is able to utilize the sequential information of lens movement trajectory to both enhance the multi-step in-focus prediction accuracy and reduce the chance of focus hunting. Our proposed method generally follows an actor-critic framework, within which the actor network (i.e., policy network) takes the captured left and right dual-pixel images with the positional information of RoI and lens as the state input, and sequentially generates a trajectory by making decisions on the lens movement. The critic network, on the other hand, receives the same input as the actor network and estimates the future expected accumulated reward given the current state, which is used to evaluate the action predicted by the actor network and thus eliminate suboptimal actions. To accelerate the DRL's training with a higher sample efficiency, we initialize the actor network with a pre-trained single-step prediction network based on the MobileNet-V2 architecture [33], where to be more consistent with the environment transition in AF tasks, the output of absolute in-focus position distribution as used in [7] is modified to the relative lens movement distribution. The critic network also takes the same MobileNet-V2 architecture. To further stabilize the DRL's training with lower occurrence of focus hunting in the resulting lens movement trajectory, we generate some offline trajectories based on prior knowledge to avoid focus hunting, which are then leveraged as an offline dataset of expert trajectories to regularize the actor network's training. Our main contributions are summarized as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 642, + 555, + 714 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 314, + 642, + 555, + 701 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 642, + 555, + 701 + ], + "spans": [ + { + "bbox": [ + 314, + 642, + 555, + 701 + ], + "type": "text", + "content": "- To overcome the limitation of single-step prediction networks, we formulate AF as a multi-step task and then propose a DRL-based framework to fully utilize the sequential information of lens movement trajectory and adjust the policy based on the immediate and long-term rewards." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 702, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 702, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 314, + 702, + 553, + 714 + ], + "type": "text", + "content": "- We initialize the policy with a pre-trained single-step pre" + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26441" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 63, + 72, + 295, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 72, + 295, + 120 + ], + "spans": [ + { + "bbox": [ + 63, + 72, + 295, + 120 + ], + "type": "text", + "content": "dition network, where the network is improved by modifying the output of absolute in-focus position distribution to relative lens movement distribution to establish a better mapping between input images and lens movement." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 121, + 295, + 204 + ], + "type": "list", + "angle": 0, + "index": 3, + "blocks": [ + { + "bbox": [ + 56, + 121, + 295, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 121, + 295, + 156 + ], + "spans": [ + { + "bbox": [ + 56, + 121, + 295, + 156 + ], + "type": "text", + "content": "- We construct an offline dataset of expert trajectories based on prior knowledge, regularizing the training of the policy network towards reducing occurrence of focus hunting." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 157, + 295, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 157, + 295, + 204 + ], + "spans": [ + { + "bbox": [ + 56, + 157, + 295, + 204 + ], + "type": "text", + "content": "- Empirical evaluations demonstrate that our method outperforms state-of-the-art AF methods on public benchmarks, with higher single- and multi-step prediction accuracy, and a significant reduction of focus hunting rate." + } + ] + } + ], + "index": 2 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 56, + 215, + 146, + 227 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 215, + 146, + 227 + ], + "spans": [ + { + "bbox": [ + 56, + 215, + 146, + 227 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 235, + 295, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 235, + 295, + 378 + ], + "spans": [ + { + "bbox": [ + 55, + 235, + 295, + 378 + ], + "type": "text", + "content": "Traditional passive AF methods can generally be categorized into search-based and phase-based approaches. Search-based approaches calculate and maximize the sharpness of the images based on image statistics [9, 21, 40, 42, 53] or frequency-domain image representations [18, 24, 25, 47, 50] to determine the optimal lens position. While showing high accuracy, these approaches are time-consuming. Additionally, search-based methods inevitably lead to focus hunting, since the peak must be surpassed to locate the exact focus point. Though some methods [14, 39] can reduce the search steps, they cannot solve the problem of focus hunting completely, or can only work on some specific devices." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 379, + 295, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 379, + 295, + 558 + ], + "spans": [ + { + "bbox": [ + 55, + 379, + 295, + 558 + ], + "type": "text", + "content": "With the development of dual-pixel sensors, phase-based methods have become the primary AF method in modern digital cameras [37]. With the phase information of the two (left/right) split pixels, the disparity can be calculated to determine the lens movement. Consequently, phase-based approaches can move the lens to the in-focus lens position with a single image capture, which is more efficient than search-based approaches. However, factors that are difficult to model, such as lens shading, can affect the calculation of phase, leading to lower accuracy. Numerous works adopt learning-based methods to address the trade-off between accuracy and speed. But these works either apply to specific tasks like microscopy images [31, 51], or are based on phase information obtained by specific hardware [17, 46], which are difficult to reproduce and compare." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 558, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 295, + 714 + ], + "type": "text", + "content": "The above methods concentrate on the processed sharpness metrics or phase data, not sufficiently using the information from pictures. Inspired by the stereo vision [19, 20], information obtained from dual-pixel pictures has found increasingly widespread applications in the computer vision community, such as monocular depth estimation [11, 22] and defocus deblurring [1]. In AF, Herrmann et al. [16] utilize convolutional networks to predict the in-focus position from input images and provide a large-scale AF dataset. Choi et al. [7] improve Herrmann's network by incorporating the region of interest position encoding (RoI-PE) and the lens position encoding (Lens-PE), significantly reducing the size while achieving state-of-the-art performance." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "type": "text", + "content": "Both methods, similar to the phase-based approaches, essentially employ the DP data to estimate the disparity and predict the lens position. But the difference is that phase-based approaches estimate the lens movement through disparity rather than directly predicting the next lens position. Meanwhile, though [7, 16] discuss the multi-step prediction, they only address the accuracy of multi-step prediction without considering the potential for focus hunting, which can greatly affect the speed and accuracy of the AF model." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 181, + 555, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 181, + 555, + 408 + ], + "spans": [ + { + "bbox": [ + 313, + 181, + 555, + 408 + ], + "type": "text", + "content": "The main reason for focus hunting in learning-based methods is the lack of consideration of the sequential information and the future effects of the actions. Yan et al. [49] and Wang et al. [46] adopt the long short-term memory (LSTM) to enhance the model's ability to incorporate previous information. With extra computation, the ability of LSTM to capture the 2-dimension data is still limited. The paradigm of predicting actions from states and the need to incorporate sequential information and future effects inspire us to use reinforcement learning (RL) in AF. Though Yu et al. [54] and Chan et al. [6] have proposed DRL-based autofocus methods, they both concentrate on the design of reward functions and do not fully leverage the advantages of RL. Wang et al. [46] adopt RL in their work to dynamically choose the RoI, rather than predicting the lens position or movement. To the best of our knowledge, little research has been conducted to reduce the occurrence of focus hunting while accelerating the multi-step prediction with a higher accuracy for AF tasks." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 422, + 421, + 436 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 422, + 421, + 436 + ], + "spans": [ + { + "bbox": [ + 314, + 422, + 421, + 436 + ], + "type": "text", + "content": "3. Proposed Method" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 443, + 439, + 455 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 443, + 439, + 455 + ], + "spans": [ + { + "bbox": [ + 313, + 443, + 439, + 455 + ], + "type": "text", + "content": "3.1. Problem Formulation" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 462, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 462, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 462, + 555, + 715 + ], + "type": "text", + "content": "We formulate the problem of AF as presented in [7], where the candidate lens position is discretized into " + }, + { + "bbox": [ + 313, + 462, + 555, + 715 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 462, + 555, + 715 + ], + "type": "text", + "content": " focal indices. Given an input image " + }, + { + "bbox": [ + 313, + 462, + 555, + 715 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 313, + 462, + 555, + 715 + ], + "type": "text", + "content": " captured at a lens position " + }, + { + "bbox": [ + 313, + 462, + 555, + 715 + ], + "type": "inline_equation", + "content": "k \\in \\{1,2,\\dots,n\\}" + }, + { + "bbox": [ + 313, + 462, + 555, + 715 + ], + "type": "text", + "content": ", we extract a patch " + }, + { + "bbox": [ + 313, + 462, + 555, + 715 + ], + "type": "inline_equation", + "content": "I_k^p" + }, + { + "bbox": [ + 313, + 462, + 555, + 715 + ], + "type": "text", + "content": " corresponding to the region of interest (RoI) that we aim to focus, where " + }, + { + "bbox": [ + 313, + 462, + 555, + 715 + ], + "type": "inline_equation", + "content": "p \\in \\{1,2,\\dots,P\\}" + }, + { + "bbox": [ + 313, + 462, + 555, + 715 + ], + "type": "text", + "content": " denotes the spatial position of the RoI in this image. A qualified AF process should be able to predict the correct focal index of the lens from a patch sequence captured by moving the lens, while the sequence should ideally be as short as possible, and without a back-and-forth movement. Considering that modern digital cameras (e.g., in smartphones) are commonly equipped with the advanced dual-pixel imaging system that can provide the phase information for AF, we specifically pay attention to the dual-pixel data, such that the input image " + }, + { + "bbox": [ + 313, + 462, + 555, + 715 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 313, + 462, + 555, + 715 + ], + "type": "text", + "content": " consists of two channels. Thus, the extracted patch at the " + }, + { + "bbox": [ + 313, + 462, + 555, + 715 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 462, + 555, + 715 + ], + "type": "text", + "content": "-th lens position, with resolution " + }, + { + "bbox": [ + 313, + 462, + 555, + 715 + ], + "type": "inline_equation", + "content": "h \\times w" + }, + { + "bbox": [ + 313, + 462, + 555, + 715 + ], + "type": "text", + "content": ", is denoted as " + }, + { + "bbox": [ + 313, + 462, + 555, + 715 + ], + "type": "inline_equation", + "content": "I_k^p \\in \\mathbb{R}^{2 \\times h \\times w}" + }, + { + "bbox": [ + 313, + 462, + 555, + 715 + ], + "type": "text", + "content": ". The terms focal stack and single slice are defined as the set of patches " + }, + { + "bbox": [ + 313, + 462, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\{I_k^p\\}_{k=1}^n" + }, + { + "bbox": [ + 313, + 462, + 555, + 715 + ], + "type": "text", + "content": " extracted from all lens positions and a single patch " + }, + { + "bbox": [ + 313, + 462, + 555, + 715 + ], + "type": "inline_equation", + "content": "I_k^p" + }, + { + "bbox": [ + 313, + 462, + 555, + 715 + ], + "type": "text", + "content": ", respectively. The objective is to predict the in-focus lens position " + }, + { + "bbox": [ + 313, + 462, + 555, + 715 + ], + "type": "inline_equation", + "content": "k^*" + }, + { + "bbox": [ + 313, + 462, + 555, + 715 + ], + "type": "text", + "content": ", which is assumed to be unique for" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "26442" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 67, + 294, + 178 + ], + "blocks": [ + { + "bbox": [ + 57, + 67, + 294, + 178 + ], + "lines": [ + { + "bbox": [ + 57, + 67, + 294, + 178 + ], + "spans": [ + { + "bbox": [ + 57, + 67, + 294, + 178 + ], + "type": "image", + "image_path": "11dce178f552eada748672e8e75193006c340380690198caaa334ae44d9ee5e4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 182, + 295, + 227 + ], + "lines": [ + { + "bbox": [ + 55, + 182, + 295, + 227 + ], + "spans": [ + { + "bbox": [ + 55, + 182, + 295, + 227 + ], + "type": "text", + "content": "Figure 3. Overall framework of our DRL-based AF model with expert trajectory regularization. The green, orange, blue, and purple boxes specify the actor network, critic network, expert dataset, and environment, respectively." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 235, + 124, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 235, + 124, + 246 + ], + "spans": [ + { + "bbox": [ + 55, + 235, + 124, + 246 + ], + "type": "text", + "content": "each focal stack." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 247, + 296, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 247, + 296, + 283 + ], + "spans": [ + { + "bbox": [ + 55, + 247, + 296, + 283 + ], + "type": "text", + "content": "Relying on the different types of patch sequences used for AF, our AF problem falls into the multi-step prediction category [7], which can be formulated as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 101, + 290, + 295, + 306 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 101, + 290, + 295, + 306 + ], + "spans": [ + { + "bbox": [ + 101, + 290, + 295, + 306 + ], + "type": "interline_equation", + "content": "I _ {k _ {0}} ^ {p} \\mapsto k _ {1}, I _ {k _ {1}} ^ {p} \\mapsto k _ {2}, \\dots , I _ {k _ {t}} ^ {p} \\mapsto k ^ {*}. \\tag {1}", + "image_path": "5da86c17ec80fdbee90183fc2211d77efc1a28d672dfae1d773d91c204a9ff81.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 312, + 296, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 312, + 296, + 384 + ], + "spans": [ + { + "bbox": [ + 55, + 312, + 296, + 384 + ], + "type": "text", + "content": "Starting from a random focal index " + }, + { + "bbox": [ + 55, + 312, + 296, + 384 + ], + "type": "inline_equation", + "content": "k_{0}" + }, + { + "bbox": [ + 55, + 312, + 296, + 384 + ], + "type": "text", + "content": ", it makes a prediction step-by-step, generating a trajectory of the lens positions " + }, + { + "bbox": [ + 55, + 312, + 296, + 384 + ], + "type": "inline_equation", + "content": "\\{k_{0}, k_{1}, k_{2}, \\ldots, k^{*}\\}" + }, + { + "bbox": [ + 55, + 312, + 296, + 384 + ], + "type": "text", + "content": ". Instead of directly predicting the absolute lens position as in Eq. (1), in this paper, we consider predicting the relative lens movement at each step, and thus reformulate the multi-step setting of the AF problem as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 73, + 392, + 295, + 407 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 392, + 295, + 407 + ], + "spans": [ + { + "bbox": [ + 73, + 392, + 295, + 407 + ], + "type": "interline_equation", + "content": "I _ {k _ {0}} ^ {p} \\mapsto k _ {1} = k _ {0} + a _ {0}, \\dots , I _ {k _ {t}} ^ {p} \\mapsto k ^ {*} = k _ {t} + a _ {t}, \\tag {2}", + "image_path": "ab1ca1915fb58f6102f7df380cacb44de615fd61efa0ae212a41effe233e2a04.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 413, + 295, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 413, + 295, + 425 + ], + "spans": [ + { + "bbox": [ + 55, + 413, + 295, + 425 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 413, + 295, + 425 + ], + "type": "inline_equation", + "content": "a_{t}" + }, + { + "bbox": [ + 55, + 413, + 295, + 425 + ], + "type": "text", + "content": " is the relative movement of the lens at time step " + }, + { + "bbox": [ + 55, + 413, + 295, + 425 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 413, + 295, + 425 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 431, + 258, + 443 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 431, + 258, + 443 + ], + "spans": [ + { + "bbox": [ + 55, + 431, + 258, + 443 + ], + "type": "text", + "content": "3.2. Overall Framework of DRL-Based AF" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 450, + 295, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 450, + 295, + 486 + ], + "spans": [ + { + "bbox": [ + 55, + 450, + 295, + 486 + ], + "type": "text", + "content": "The framework of our DRL-based AF method is shown in Fig. 3, with three major components: an actor network, a critic network, and an offline dataset of expert trajectories." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 487, + 295, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 55, + 487, + 295, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 487, + 295, + 557 + ], + "spans": [ + { + "bbox": [ + 55, + 487, + 295, + 557 + ], + "type": "text", + "content": "- Actor network (i.e., policy network) parameterized by " + }, + { + "bbox": [ + 55, + 487, + 295, + 557 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 55, + 487, + 295, + 557 + ], + "type": "text", + "content": ": At each time step " + }, + { + "bbox": [ + 55, + 487, + 295, + 557 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 487, + 295, + 557 + ], + "type": "text", + "content": ", it takes the state " + }, + { + "bbox": [ + 55, + 487, + 295, + 557 + ], + "type": "inline_equation", + "content": "s_t" + }, + { + "bbox": [ + 55, + 487, + 295, + 557 + ], + "type": "text", + "content": ", including the left and right DP images, the RoI-PE and the Lens-PE, as the input, and outputs an action " + }, + { + "bbox": [ + 55, + 487, + 295, + 557 + ], + "type": "inline_equation", + "content": "a_t \\sim \\pi_\\theta(\\cdot | s_t)" + }, + { + "bbox": [ + 55, + 487, + 295, + 557 + ], + "type": "text", + "content": ", where the sign of " + }, + { + "bbox": [ + 55, + 487, + 295, + 557 + ], + "type": "inline_equation", + "content": "a_t" + }, + { + "bbox": [ + 55, + 487, + 295, + 557 + ], + "type": "text", + "content": " indicates the direction of the lens movement and the absolute value " + }, + { + "bbox": [ + 55, + 487, + 295, + 557 + ], + "type": "inline_equation", + "content": "|a_t|" + }, + { + "bbox": [ + 55, + 487, + 295, + 557 + ], + "type": "text", + "content": " represents the moving distance." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 558, + 295, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 295, + 688 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 295, + 688 + ], + "type": "text", + "content": "- Critic network parameterized by " + }, + { + "bbox": [ + 55, + 558, + 295, + 688 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 55, + 558, + 295, + 688 + ], + "type": "text", + "content": ": At each time step " + }, + { + "bbox": [ + 55, + 558, + 295, + 688 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 558, + 295, + 688 + ], + "type": "text", + "content": ", it takes the state " + }, + { + "bbox": [ + 55, + 558, + 295, + 688 + ], + "type": "inline_equation", + "content": "s_t" + }, + { + "bbox": [ + 55, + 558, + 295, + 688 + ], + "type": "text", + "content": " as input, and outputs the estimated value function " + }, + { + "bbox": [ + 55, + 558, + 295, + 688 + ], + "type": "inline_equation", + "content": "V_{\\phi}(s_t)" + }, + { + "bbox": [ + 55, + 558, + 295, + 688 + ], + "type": "text", + "content": " for the current state " + }, + { + "bbox": [ + 55, + 558, + 295, + 688 + ], + "type": "inline_equation", + "content": "s_t" + }, + { + "bbox": [ + 55, + 558, + 295, + 688 + ], + "type": "text", + "content": ". This value function represents the future expected accumulated reward given the current state, which thus can be used to evaluate the action predicted by the actor network, allowing the DRL agent to identify and eliminate some suboptimal actions. Since this evaluation is based on the future expected accumulated reward, the actions are thus chosen based on their anticipated long-term impact, rather than the immediate return." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": "- Offline dataset of expert trajectories " + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^E = \\{(s_t, a_t^E)\\}" + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": ": It provides the expert trajectories with state " + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "inline_equation", + "content": "s_t" + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": " and as" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 321, + 72, + 553, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 72, + 553, + 156 + ], + "spans": [ + { + "bbox": [ + 321, + 72, + 553, + 156 + ], + "type": "text", + "content": "sociated expert action " + }, + { + "bbox": [ + 321, + 72, + 553, + 156 + ], + "type": "inline_equation", + "content": "a_{t}^{E}" + }, + { + "bbox": [ + 321, + 72, + 553, + 156 + ], + "type": "text", + "content": ". These trajectories are generated based on the prior knowledge to avoid focus hunting, and can thus be used to regularize the training of the actor network by maximizing the probability of expert action given a state " + }, + { + "bbox": [ + 321, + 72, + 553, + 156 + ], + "type": "inline_equation", + "content": "s_{t}" + }, + { + "bbox": [ + 321, + 72, + 553, + 156 + ], + "type": "text", + "content": ", which ensures that the actor network learns from the high-quality multi-step movement of the lens from the expert." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 163, + 527, + 175 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 163, + 527, + 175 + ], + "spans": [ + { + "bbox": [ + 313, + 163, + 527, + 175 + ], + "type": "text", + "content": "3.3. Actor Network for Movement Prediction" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 182, + 553, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 182, + 553, + 252 + ], + "spans": [ + { + "bbox": [ + 313, + 182, + 553, + 252 + ], + "type": "text", + "content": "To utilize the sequential information of the lens movement trajectory to both enhance the multi-step in-focus prediction accuracy and reduce the chance of focus hunting, the actor network ideally should establish a more accurate mapping between the disparity from DP data and the lens movement." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 253, + 554, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 253, + 554, + 361 + ], + "spans": [ + { + "bbox": [ + 313, + 253, + 554, + 361 + ], + "type": "text", + "content": "The AF procedure can be formulated as a Markov Decision Progress (MDP), defined by a tuple " + }, + { + "bbox": [ + 313, + 253, + 554, + 361 + ], + "type": "inline_equation", + "content": "\\langle S, \\mathcal{A}, \\mathcal{R}, \\mathcal{P}, \\gamma \\rangle" + }, + { + "bbox": [ + 313, + 253, + 554, + 361 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 253, + 554, + 361 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 313, + 253, + 554, + 361 + ], + "type": "text", + "content": " denotes the state space, " + }, + { + "bbox": [ + 313, + 253, + 554, + 361 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 313, + 253, + 554, + 361 + ], + "type": "text", + "content": " denotes the action space and " + }, + { + "bbox": [ + 313, + 253, + 554, + 361 + ], + "type": "inline_equation", + "content": "\\gamma \\in (0,1]" + }, + { + "bbox": [ + 313, + 253, + 554, + 361 + ], + "type": "text", + "content": " is the discount factor. At each time step " + }, + { + "bbox": [ + 313, + 253, + 554, + 361 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 253, + 554, + 361 + ], + "type": "text", + "content": ", a state " + }, + { + "bbox": [ + 313, + 253, + 554, + 361 + ], + "type": "inline_equation", + "content": "s_t \\in S" + }, + { + "bbox": [ + 313, + 253, + 554, + 361 + ], + "type": "text", + "content": ", consisting of the left and right DP images, the RoI-PE and the Lens-PE, is received by the RL agent, and an action " + }, + { + "bbox": [ + 313, + 253, + 554, + 361 + ], + "type": "inline_equation", + "content": "a_t \\in \\mathcal{A}" + }, + { + "bbox": [ + 313, + 253, + 554, + 361 + ], + "type": "text", + "content": " is then generated by the actor network to determine the lens movement. The environment transition dynamics " + }, + { + "bbox": [ + 313, + 253, + 554, + 361 + ], + "type": "inline_equation", + "content": "P(s_{t+1} | s_t, a_t)" + }, + { + "bbox": [ + 313, + 253, + 554, + 361 + ], + "type": "text", + "content": " in AF can be defined as:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 347, + 369, + 553, + 384 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 369, + 553, + 384 + ], + "spans": [ + { + "bbox": [ + 347, + 369, + 553, + 384 + ], + "type": "interline_equation", + "content": "s _ {t + 1} = g \\left(\\operatorname {c l i p} \\left(a _ {t} + g ^ {- 1} \\left(s _ {t}\\right), 0, k _ {\\max }\\right)\\right), \\tag {3}", + "image_path": "4d752e966dc989d20d5c70399f7236150620679abea527adc93d3649f9881668.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 392, + 554, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 392, + 554, + 464 + ], + "spans": [ + { + "bbox": [ + 313, + 392, + 554, + 464 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 392, + 554, + 464 + ], + "type": "inline_equation", + "content": "g(k_{t}) = s_{t}" + }, + { + "bbox": [ + 313, + 392, + 554, + 464 + ], + "type": "text", + "content": " maps the lens position " + }, + { + "bbox": [ + 313, + 392, + 554, + 464 + ], + "type": "inline_equation", + "content": "k_{t}" + }, + { + "bbox": [ + 313, + 392, + 554, + 464 + ], + "type": "text", + "content": " to the state " + }, + { + "bbox": [ + 313, + 392, + 554, + 464 + ], + "type": "inline_equation", + "content": "s_{t}" + }, + { + "bbox": [ + 313, + 392, + 554, + 464 + ], + "type": "text", + "content": " at time-step " + }, + { + "bbox": [ + 313, + 392, + 554, + 464 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 392, + 554, + 464 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 392, + 554, + 464 + ], + "type": "inline_equation", + "content": "k_{max} \\triangleq k_{n}" + }, + { + "bbox": [ + 313, + 392, + 554, + 464 + ], + "type": "text", + "content": " is the maximum lens position. The reward function " + }, + { + "bbox": [ + 313, + 392, + 554, + 464 + ], + "type": "inline_equation", + "content": "R(s_{t}, a_{t})" + }, + { + "bbox": [ + 313, + 392, + 554, + 464 + ], + "type": "text", + "content": " is defined as the negative absolute distance between the predicted and ground-truth infocus positions, where we additionally introduce a penalty term " + }, + { + "bbox": [ + 313, + 392, + 554, + 464 + ], + "type": "inline_equation", + "content": "R_{fh}" + }, + { + "bbox": [ + 313, + 392, + 554, + 464 + ], + "type": "text", + "content": " to mitigate the occurrence of focus hunting:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 363, + 473, + 553, + 487 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 363, + 473, + 553, + 487 + ], + "spans": [ + { + "bbox": [ + 363, + 473, + 553, + 487 + ], + "type": "interline_equation", + "content": "R \\left(s _ {t}, a _ {t}\\right) = - \\left| k _ {t + 1} - k _ {g t} \\right| + R _ {f h}, \\tag {4}", + "image_path": "512aad50ed7510f168eb66d8cb8012c5731e38079020855ec2ba5aadfc4971fb.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 495, + 553, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 495, + 553, + 520 + ], + "spans": [ + { + "bbox": [ + 313, + 495, + 553, + 520 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 495, + 553, + 520 + ], + "type": "inline_equation", + "content": "k_{gt}" + }, + { + "bbox": [ + 313, + 495, + 553, + 520 + ], + "type": "text", + "content": " denotes the ground-truth lens position, and the penalty term " + }, + { + "bbox": [ + 313, + 495, + 553, + 520 + ], + "type": "inline_equation", + "content": "R_{fh}" + }, + { + "bbox": [ + 313, + 495, + 553, + 520 + ], + "type": "text", + "content": " is defined as:" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 377, + 528, + 553, + 561 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 377, + 528, + 553, + 561 + ], + "spans": [ + { + "bbox": [ + 377, + 528, + 553, + 561 + ], + "type": "interline_equation", + "content": "R _ {f h} = \\left\\{ \\begin{array}{l l} c, & \\text {i f F H o c c u r s}; \\\\ 0, & \\text {o t h e r w i s e}; \\end{array} \\right. \\tag {5}", + "image_path": "a4a9e3b86cb12f5e3a361739b02be8b6cf5f836c6df6dbaab65f876910b6c5e6.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 568, + 553, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 568, + 553, + 604 + ], + "spans": [ + { + "bbox": [ + 313, + 568, + 553, + 604 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 568, + 553, + 604 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 568, + 553, + 604 + ], + "type": "text", + "content": " is a negative hyperparameter that penalizes the event of focus hunting. The main goal of RL is to maximize the expected return " + }, + { + "bbox": [ + 313, + 568, + 553, + 604 + ], + "type": "inline_equation", + "content": "J(\\theta)" + }, + { + "bbox": [ + 313, + 568, + 553, + 604 + ], + "type": "text", + "content": " defined as:" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 368, + 613, + 553, + 646 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 368, + 613, + 553, + 646 + ], + "spans": [ + { + "bbox": [ + 368, + 613, + 553, + 646 + ], + "type": "interline_equation", + "content": "J (\\theta) = \\mathbb {E} _ {\\pi_ {\\theta}} \\left[ \\sum_ {t = 0} ^ {T} \\gamma^ {t} R \\left(s _ {t}, a _ {t}\\right) \\right], \\tag {6}", + "image_path": "887326d679b0d302cb32831a7e56393eb8ff7748ac6b1e693556ea64a0bb4a80.jpg" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 653, + 553, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 553, + 677 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 553, + 677 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 653, + 553, + 677 + ], + "type": "inline_equation", + "content": "\\pi_{\\theta}" + }, + { + "bbox": [ + 313, + 653, + 553, + 677 + ], + "type": "text", + "content": " is the policy network parameterized by " + }, + { + "bbox": [ + 313, + 653, + 553, + 677 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 653, + 553, + 677 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 653, + 553, + 677 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 653, + 553, + 677 + ], + "type": "text", + "content": " denotes the maximum time step." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "content": "Since the single-step prediction network model proposed by Choi et al. [7] has demonstrated a state-of-the-art accuracy, we adopt its network architecture as our actor network." + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26443" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 85, + 68, + 271, + 166 + ], + "blocks": [ + { + "bbox": [ + 85, + 68, + 271, + 166 + ], + "lines": [ + { + "bbox": [ + 85, + 68, + 271, + 166 + ], + "spans": [ + { + "bbox": [ + 85, + 68, + 271, + 166 + ], + "type": "image", + "image_path": "96b66baf09fbbbf481a30e551f6bdf176e716e7ea768cac89c93fdfb152d785b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 86, + 170, + 264, + 181 + ], + "lines": [ + { + "bbox": [ + 86, + 170, + 264, + 181 + ], + "spans": [ + { + "bbox": [ + 86, + 170, + 264, + 181 + ], + "type": "text", + "content": "Figure 4. Illustration of relative movement label." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 190, + 296, + 454 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 190, + 296, + 454 + ], + "spans": [ + { + "bbox": [ + 55, + 190, + 296, + 454 + ], + "type": "text", + "content": "However, the original model in [7] directly predicts the infocus position, which contradicts the principle of using DP data for in-focus prediction. In essence, since there is a clear correspondence between the disparity and lens movement (i.e., through the defocus conversion coefficient), it is more reasonable to predict the lens movement rather than the lens position based on the DP data. Thus, we modify the original network's output of absolute in-focus position distribution in [7] to the relative lens movement distribution, with the inference procedure of actor network shown in Fig. 9. Accordingly, Fig. 4 illustrates the conversion from original absolute position labels to relative movement labels that are used for training. Unlike [17], we expand the action range from " + }, + { + "bbox": [ + 55, + 190, + 296, + 454 + ], + "type": "inline_equation", + "content": "[- \\mathrm{gt}, k_{\\max} - \\mathrm{gt}]" + }, + { + "bbox": [ + 55, + 190, + 296, + 454 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 190, + 296, + 454 + ], + "type": "inline_equation", + "content": "[-k_{\\max}, k_{\\max}]" + }, + { + "bbox": [ + 55, + 190, + 296, + 454 + ], + "type": "text", + "content": ", and the label denoting the relative movement towards GT index then converts to " + }, + { + "bbox": [ + 55, + 190, + 296, + 454 + ], + "type": "inline_equation", + "content": "\\mathrm{gt} - f" + }, + { + "bbox": [ + 55, + 190, + 296, + 454 + ], + "type": "text", + "content": ". If the actor network predicts an out-of-bound action, it will be clipped to the closest lower/upper bounds. Based on this formulation, we initialize the actor network with a pre-trained MobileNet-V2 [33] optimized based on the ordinal regression loss [10]. For details about the pretrained model of actor network, please refer to Sec. 7 in our supplementary." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 461, + 288, + 474 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 461, + 288, + 474 + ], + "spans": [ + { + "bbox": [ + 55, + 461, + 288, + 474 + ], + "type": "text", + "content": "3.4. Critic Network Incorporating Future Impact" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 479, + 295, + 551 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 479, + 295, + 551 + ], + "spans": [ + { + "bbox": [ + 55, + 479, + 295, + 551 + ], + "type": "text", + "content": "The actor network cannot directly incorporate the future reward of current actions, where the probability of any output action with a larger instant reward will be selected, leading to a suboptimal policy. To address this issue, a critic network is further introduced to provide baselines and evaluate the future impact of the current actions." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 552, + 295, + 659 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 552, + 295, + 659 + ], + "spans": [ + { + "bbox": [ + 55, + 552, + 295, + 659 + ], + "type": "text", + "content": "The critic network estimates the value function " + }, + { + "bbox": [ + 55, + 552, + 295, + 659 + ], + "type": "inline_equation", + "content": "V_{\\phi}(s_t)" + }, + { + "bbox": [ + 55, + 552, + 295, + 659 + ], + "type": "text", + "content": " of the current state " + }, + { + "bbox": [ + 55, + 552, + 295, + 659 + ], + "type": "inline_equation", + "content": "s_t" + }, + { + "bbox": [ + 55, + 552, + 295, + 659 + ], + "type": "text", + "content": ", which is defined as the expected future accumulated discount reward of actions. If we further define the future accumulated discount reward of an action under a given state as a Q-function " + }, + { + "bbox": [ + 55, + 552, + 295, + 659 + ], + "type": "inline_equation", + "content": "Q(s_t, a_t)" + }, + { + "bbox": [ + 55, + 552, + 295, + 659 + ], + "type": "text", + "content": ", the advantage function " + }, + { + "bbox": [ + 55, + 552, + 295, + 659 + ], + "type": "inline_equation", + "content": "\\hat{A}_t = Q(s_t, a_t) - V_{\\phi}(s_t)" + }, + { + "bbox": [ + 55, + 552, + 295, + 659 + ], + "type": "text", + "content": " then evaluates the relative quality of a particular action under the current state. With this advantage function, the actor network can be trained by minimizing the clipped surrogate objective loss, given by:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 61, + 667, + 295, + 681 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 667, + 295, + 681 + ], + "spans": [ + { + "bbox": [ + 61, + 667, + 295, + 681 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {c l i p}} (\\theta) = \\mathbb {E} \\left[ \\min \\left(r _ {t} (\\theta) \\hat {A}, \\operatorname {c l i p} \\left(r _ {t} (\\theta), 1 - \\epsilon , 1 + \\epsilon\\right) \\hat {A} _ {t}\\right), \\right. \\tag {7}", + "image_path": "a45948f7f3e6e5a192d8cca8dce83674ae0940bda88f70ed8c0242000ee62332.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "inline_equation", + "content": "r_t(\\theta) = \\pi_\\theta(a_t|s_t) / \\pi_\\theta^{old}(a_t|s_t)" + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": " is the probability ratio and " + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": " is a clip hyperparameter." + } + ] + } + ], + "index": 7 + }, + { + "type": "code", + "bbox": [ + 316, + 87, + 553, + 318 + ], + "blocks": [ + { + "bbox": [ + 315, + 72, + 481, + 84 + ], + "lines": [ + { + "bbox": [ + 315, + 72, + 481, + 84 + ], + "spans": [ + { + "bbox": [ + 315, + 72, + 481, + 84 + ], + "type": "text", + "content": "Algorithm 1 Expert trajectory generator." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 316, + 87, + 553, + 318 + ], + "lines": [ + { + "bbox": [ + 316, + 87, + 553, + 318 + ], + "spans": [ + { + "bbox": [ + 316, + 87, + 553, + 318 + ], + "type": "text", + "content": "Input: " + }, + { + "bbox": [ + 316, + 87, + 553, + 318 + ], + "type": "inline_equation", + "content": "\\mathcal{D} = \\{f,GT,(s_0,a_0),\\dots ,(s_n,a_n)\\}_{N};" + }, + { + "bbox": [ + 316, + 87, + 553, + 318 + ], + "type": "text", + "content": " \nOutput: " + }, + { + "bbox": [ + 316, + 87, + 553, + 318 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^E" + }, + { + "bbox": [ + 316, + 87, + 553, + 318 + ], + "type": "text", + "content": " . \n1: " + }, + { + "bbox": [ + 316, + 87, + 553, + 318 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^E = \\emptyset" + }, + { + "bbox": [ + 316, + 87, + 553, + 318 + ], + "type": "text", + "content": " \n2: for all " + }, + { + "bbox": [ + 316, + 87, + 553, + 318 + ], + "type": "inline_equation", + "content": "\\{g(\\cdot),GT,(s_0,a_0),\\dots ,(s_n,a_n)\\} _i\\in \\mathcal{D}" + }, + { + "bbox": [ + 316, + 87, + 553, + 318 + ], + "type": "text", + "content": " do \n3: for all " + }, + { + "bbox": [ + 316, + 87, + 553, + 318 + ], + "type": "inline_equation", + "content": "j = 0,1,2,\\dots ,n" + }, + { + "bbox": [ + 316, + 87, + 553, + 318 + ], + "type": "text", + "content": " do \n4: " + }, + { + "bbox": [ + 316, + 87, + 553, + 318 + ], + "type": "inline_equation", + "content": "k_{j}\\gets g_{i}^{-1}(s_{j})" + }, + { + "bbox": [ + 316, + 87, + 553, + 318 + ], + "type": "text", + "content": " \n5: " + }, + { + "bbox": [ + 316, + 87, + 553, + 318 + ], + "type": "inline_equation", + "content": "o_j\\gets \\mathrm{sign}(k_0 - GT_i)\\cdot |k_j - GT_i| + GT_i;" + }, + { + "bbox": [ + 316, + 87, + 553, + 318 + ], + "type": "text", + "content": " \n6: " + }, + { + "bbox": [ + 316, + 87, + 553, + 318 + ], + "type": "inline_equation", + "content": "o_j\\gets \\mathrm{CLIP}(o_j,k_0,GT_i);" + }, + { + "bbox": [ + 316, + 87, + 553, + 318 + ], + "type": "text", + "content": " \n7: end for \n8: " + }, + { + "bbox": [ + 316, + 87, + 553, + 318 + ], + "type": "inline_equation", + "content": "(o_{m_0},\\dots ,o_{m_n})\\gets \\mathrm{sort}(o_0,\\dots ,o_n;\\mathrm{reverse} =" + }, + { + "bbox": [ + 316, + 87, + 553, + 318 + ], + "type": "inline_equation", + "content": "\\mathrm{Bool}(k_0 > GT_i));" + }, + { + "bbox": [ + 316, + 87, + 553, + 318 + ], + "type": "text", + "content": " \n9: for all " + }, + { + "bbox": [ + 316, + 87, + 553, + 318 + ], + "type": "inline_equation", + "content": "j = 1,2,\\dots ,n" + }, + { + "bbox": [ + 316, + 87, + 553, + 318 + ], + "type": "text", + "content": " do \n10: " + }, + { + "bbox": [ + 316, + 87, + 553, + 318 + ], + "type": "inline_equation", + "content": "s_j\\gets g_i(o_{m_j});" + }, + { + "bbox": [ + 316, + 87, + 553, + 318 + ], + "type": "text", + "content": " \n11: " + }, + { + "bbox": [ + 316, + 87, + 553, + 318 + ], + "type": "inline_equation", + "content": "a_{j - 1}\\leftarrow o_{m_j} - o_{m_{j - 1}};" + }, + { + "bbox": [ + 316, + 87, + 553, + 318 + ], + "type": "text", + "content": " \n12: end for \n13: " + }, + { + "bbox": [ + 316, + 87, + 553, + 318 + ], + "type": "inline_equation", + "content": "a_{n}\\gets 0" + }, + { + "bbox": [ + 316, + 87, + 553, + 318 + ], + "type": "text", + "content": " \n14: " + }, + { + "bbox": [ + 316, + 87, + 553, + 318 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^E\\gets \\mathcal{D}^E\\cup \\{g(\\cdot),GT,(s_0,a_0),\\dots ,(s_n,a_n)\\} _i;" + }, + { + "bbox": [ + 316, + 87, + 553, + 318 + ], + "type": "text", + "content": " \n15: end for \n16: return " + }, + { + "bbox": [ + 316, + 87, + 553, + 318 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^E" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "code_body" + } + ], + "index": 9, + "sub_type": "algorithm" + }, + { + "bbox": [ + 313, + 336, + 555, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 336, + 555, + 373 + ], + "spans": [ + { + "bbox": [ + 313, + 336, + 555, + 373 + ], + "type": "text", + "content": "The critic network, on the other hand, is trained by minimizing the " + }, + { + "bbox": [ + 313, + 336, + 555, + 373 + ], + "type": "inline_equation", + "content": "\\ell_2" + }, + { + "bbox": [ + 313, + 336, + 555, + 373 + ], + "type": "text", + "content": " loss between the estimated value function " + }, + { + "bbox": [ + 313, + 336, + 555, + 373 + ], + "type": "inline_equation", + "content": "V_{\\phi}(s_t)" + }, + { + "bbox": [ + 313, + 336, + 555, + 373 + ], + "type": "text", + "content": " and the true value function " + }, + { + "bbox": [ + 313, + 336, + 555, + 373 + ], + "type": "inline_equation", + "content": "\\hat{V}_t" + }, + { + "bbox": [ + 313, + 336, + 555, + 373 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 363, + 380, + 554, + 413 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 363, + 380, + 554, + 413 + ], + "spans": [ + { + "bbox": [ + 363, + 380, + 554, + 413 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {c r i t i c}} = \\frac {1}{N} \\sum_ {t = 1} ^ {N} \\left(V _ {\\phi} \\left(s _ {t}\\right) - \\hat {V} _ {t}\\right) ^ {2}. \\tag {8}", + "image_path": "5be73657978501b4a61ee36db437eb04ccb88f55ae3cc53ea8b40bca8224118b.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 420, + 494, + 433 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 420, + 494, + 433 + ], + "spans": [ + { + "bbox": [ + 313, + 420, + 494, + 433 + ], + "type": "text", + "content": "3.5. Expert Trajectory Regularization" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 438, + 555, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 438, + 555, + 629 + ], + "spans": [ + { + "bbox": [ + 313, + 438, + 555, + 629 + ], + "type": "text", + "content": "Though the incorporation of a critic network can improve the effectiveness of actor network's training, the exponential growth in the state and action combinations inherently leads to another issue of low sample efficiency in DRL. Recent literature has witnessed the success of incorporating offline data for improving the sample efficiency of DRL's training, both theoretically [38, 45] and empirically [5, 27, 28]. Motivated by this, and to further stabilize the DRL with a lower occurrence of focus hunting in the resulting lens movement trajectory, we generate some offline trajectories based on the prior knowledge to avoid focus hunting, which are then leveraged as an offline dataset " + }, + { + "bbox": [ + 313, + 438, + 555, + 629 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^E = \\{(s_t,a_t^E)\\}" + }, + { + "bbox": [ + 313, + 438, + 555, + 629 + ], + "type": "text", + "content": " of expert trajectories to regularize the actor network's training. Algorithm 1 shows our proposed method to generate these expert trajectories, with more details provided in Sec. 8 in our supplementary." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 630, + 554, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 554, + 677 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 554, + 677 + ], + "type": "text", + "content": "Fig. 5 shows our proposed framework for expert trajectory regularized DRL's training. At each time, we sample " + }, + { + "bbox": [ + 313, + 630, + 554, + 677 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 630, + 554, + 677 + ], + "type": "text", + "content": " trajectories from the offline expert dataset " + }, + { + "bbox": [ + 313, + 630, + 554, + 677 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^E" + }, + { + "bbox": [ + 313, + 630, + 554, + 677 + ], + "type": "text", + "content": " to calculate the imitation loss:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 326, + 683, + 555, + 716 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 683, + 555, + 716 + ], + "spans": [ + { + "bbox": [ + 326, + 683, + 555, + 716 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {i m i t} (\\theta) = - \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\sum_ {t = 1} ^ {T} \\log p \\left(\\pi_ {\\theta} (\\cdot | s _ {i, t}) = a _ {i, t}\\right), \\tag {9}", + "image_path": "b2c710a21b0952f61768e73bf95e47fa3df72408c47b95940fa05ece4e8ba5e8.jpg" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26444" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 69, + 64, + 284, + 167 + ], + "blocks": [ + { + "bbox": [ + 69, + 64, + 284, + 167 + ], + "lines": [ + { + "bbox": [ + 69, + 64, + 284, + 167 + ], + "spans": [ + { + "bbox": [ + 69, + 64, + 284, + 167 + ], + "type": "image", + "image_path": "a239323905a5adb0550695bb67ac25530096eb6c4600b962e140871a804b90b6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 75, + 171, + 275, + 182 + ], + "lines": [ + { + "bbox": [ + 75, + 171, + 275, + 182 + ], + "spans": [ + { + "bbox": [ + 75, + 171, + 275, + 182 + ], + "type": "text", + "content": "Figure 5. Expert trajectory regularized DRL's training." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 192, + 296, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 192, + 296, + 240 + ], + "spans": [ + { + "bbox": [ + 55, + 192, + 296, + 240 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 192, + 296, + 240 + ], + "type": "inline_equation", + "content": "a_{i,t} = k_{i,t + 1} - k_{i,t}" + }, + { + "bbox": [ + 55, + 192, + 296, + 240 + ], + "type": "text", + "content": " denotes the action at time step " + }, + { + "bbox": [ + 55, + 192, + 296, + 240 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 192, + 296, + 240 + ], + "type": "text", + "content": " in the " + }, + { + "bbox": [ + 55, + 192, + 296, + 240 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 192, + 296, + 240 + ], + "type": "text", + "content": "-th trajectory, " + }, + { + "bbox": [ + 55, + 192, + 296, + 240 + ], + "type": "inline_equation", + "content": "s_{i,t}" + }, + { + "bbox": [ + 55, + 192, + 296, + 240 + ], + "type": "text", + "content": " represents the network input at lens location " + }, + { + "bbox": [ + 55, + 192, + 296, + 240 + ], + "type": "inline_equation", + "content": "k_{i,t}" + }, + { + "bbox": [ + 55, + 192, + 296, + 240 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 192, + 296, + 240 + ], + "type": "inline_equation", + "content": "\\pi_{\\theta}(\\cdot | s_{i,t})" + }, + { + "bbox": [ + 55, + 192, + 296, + 240 + ], + "type": "text", + "content": " represents the output distribution of the actor network, with the total loss defined as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 109, + 251, + 295, + 264 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 109, + 251, + 295, + 264 + ], + "spans": [ + { + "bbox": [ + 109, + 251, + 295, + 264 + ], + "type": "interline_equation", + "content": "\\mathcal {L} (\\theta) = \\mathcal {L} _ {\\text {c l i p}} (\\theta) + \\lambda \\cdot \\mathcal {L} _ {\\text {i m i t}} (\\theta), \\tag {10}", + "image_path": "6f7a7129a148dbeaaee666c720d56e59f5e2a597fea29bae1ae18bb4b168dc66.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 274, + 296, + 310 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 274, + 296, + 310 + ], + "spans": [ + { + "bbox": [ + 55, + 274, + 296, + 310 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 274, + 296, + 310 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 274, + 296, + 310 + ], + "type": "text", + "content": " balances the online and offline learning. By minimizing the loss " + }, + { + "bbox": [ + 55, + 274, + 296, + 310 + ], + "type": "inline_equation", + "content": "\\mathcal{L}(\\theta)" + }, + { + "bbox": [ + 55, + 274, + 296, + 310 + ], + "type": "text", + "content": ", our DRL algorithm maximizes the expected accumulated discount reward " + }, + { + "bbox": [ + 55, + 274, + 296, + 310 + ], + "type": "inline_equation", + "content": "J(\\theta)" + }, + { + "bbox": [ + 55, + 274, + 296, + 310 + ], + "type": "text", + "content": " indirectly." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 321, + 137, + 335 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 321, + 137, + 335 + ], + "spans": [ + { + "bbox": [ + 55, + 321, + 137, + 335 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 342, + 296, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 342, + 296, + 426 + ], + "spans": [ + { + "bbox": [ + 55, + 342, + 296, + 426 + ], + "type": "text", + "content": "In the experiments, our actor and critic networks are built upon MobileNetV2 [33]. We reproduce the models for [16] and [7], since no open-source code is available. In the multi-step prediction setting, we only consider the model obtained by [7], since the multi-step model given by [16] must be restrained for each step, which is impractical. More results and analysis are provided in Sec. 9 in our supplementary." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 426, + 296, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 426, + 296, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 426, + 296, + 605 + ], + "type": "text", + "content": "Dataset. We adopt the publicly available AF dataset constructed by Herrmann et al. [16]. This dataset comprises 51 scenes, each containing 10 stacks with varying compositions. All instances within the dataset are captured at " + }, + { + "bbox": [ + 55, + 426, + 296, + 605 + ], + "type": "inline_equation", + "content": "n = 49" + }, + { + "bbox": [ + 55, + 426, + 296, + 605 + ], + "type": "text", + "content": " focal depths, uniformly sampled in the inverse depth space, ranging from 0.102 to 3.91 meters, captured using the Google Pixel 3 smartphones. To create the training and test sets, we extract and sample " + }, + { + "bbox": [ + 55, + 426, + 296, + 605 + ], + "type": "inline_equation", + "content": "128 \\times 128" + }, + { + "bbox": [ + 55, + 426, + 296, + 605 + ], + "type": "text", + "content": " patches with a stride of 96 from the central camera, resulting in a total of 3,341,163 patches for the training set and 382,445 patches for the test set. Patches with the same spatial position within one image and 49 different focal depths constitute a single focal stack. Please refer to [16] for additional details about the dataset, and refer to Sec. 6 in our supplementary for the labeling method adopted in this paper." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 606, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 606, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 606, + 296, + 714 + ], + "type": "text", + "content": "Evaluation metric. We follow the evaluation metric proposed by Scharstein et al. [34], which is used for Middlebury stereo dataset and adopted by Herrmann et al. [16] and Choi et al. [7] to evaluate the AF model's prediction accuracy. In particular, this metric is calculated by the proportion of patches whose stop positions are within 0, 1, 2, and 4 indices away from the in-focus index, which can be considered in the depth of field (DoF). The mean absolute error (MAE) and root mean squared error (RMSE) are also" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 553, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 204 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 204 + ], + "type": "text", + "content": "used to evaluate the AF model's performance. To evaluate the focus hunting (FH), we retain every focusing trajectory and calculate the direction and distance of each lens movement. If the direction of a current movement is opposite to that of the previous one in the trajectory, we consider an occurrence of FH in that trajectory. Since the actor network obtained by our DRL and the model in [7] share the same architecture and similar input/output, we can simply regard the number of steps to reach the in-focus position as the focus speed. We calculate the images captured at all initial positions of the lens to account for accuracy and FH rate." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 205, + 555, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 205, + 555, + 348 + ], + "spans": [ + { + "bbox": [ + 313, + 205, + 555, + 348 + ], + "type": "text", + "content": "Implementation detail. We use PyTorch [29] and TorchRL [4] for our implementation. For deep learning models, we use the Adam optimizer [23] with a learning rate of " + }, + { + "bbox": [ + 313, + 205, + 555, + 348 + ], + "type": "inline_equation", + "content": "10^{-3}" + }, + { + "bbox": [ + 313, + 205, + 555, + 348 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 205, + 555, + 348 + ], + "type": "inline_equation", + "content": "(\\beta_{1},\\beta_{2}) = (0.5,0.999)" + }, + { + "bbox": [ + 313, + 205, + 555, + 348 + ], + "type": "text", + "content": ", a batch size of 128, and cosine learning rate decay [26] for 10k iterations. For reinforcement learning models, we adopt the PPO-CLIP [35] algorithm. The initial learning rate is set to " + }, + { + "bbox": [ + 313, + 205, + 555, + 348 + ], + "type": "inline_equation", + "content": "1\\times 10^{-5}" + }, + { + "bbox": [ + 313, + 205, + 555, + 348 + ], + "type": "text", + "content": ", and the clip epsilon is 0.2, a batch size of 32, Adam optimizer with the same " + }, + { + "bbox": [ + 313, + 205, + 555, + 348 + ], + "type": "inline_equation", + "content": "(\\beta_{1},\\beta_{2})" + }, + { + "bbox": [ + 313, + 205, + 555, + 348 + ], + "type": "text", + "content": " as deep learning. The maximum episode length of RL is set to 4. For the expert trajectory, we first construct an expert dataset containing 6,000,000 expert trajectories and set the weight " + }, + { + "bbox": [ + 313, + 205, + 555, + 348 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 205, + 555, + 348 + ], + "type": "text", + "content": " to 0.001 in the loss function." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 358, + 432, + 371 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 358, + 432, + 371 + ], + "spans": [ + { + "bbox": [ + 313, + 358, + 432, + 371 + ], + "type": "text", + "content": "4.1. Quantitative Results" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 378, + 554, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 378, + 554, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 378, + 554, + 665 + ], + "type": "text", + "content": "FH rate and multi-step prediction accuracy (Effects of DRL). The FH rate and multi-step prediction accuracy of comparison methods are shown in Tab. 1. Though the relative movement label can significantly improve the one-step prediction accuracy, the FH rate in the multi-step setting is still high since the current deep-learning training process cannot obtain and exploit the sequential information of the lens movement. Due to DRL's advantages in modeling sequential decision-making problems, our DRL-based models can directly decrease the FH rate without relying on any extra networks, such as the LSTM [36]. Compared with the policy obtained by deep learning, our DRL-based policy can decrease the FH rate by " + }, + { + "bbox": [ + 313, + 378, + 554, + 665 + ], + "type": "inline_equation", + "content": "9.1\\%" + }, + { + "bbox": [ + 313, + 378, + 554, + 665 + ], + "type": "text", + "content": " in 2-step prediction, " + }, + { + "bbox": [ + 313, + 378, + 554, + 665 + ], + "type": "inline_equation", + "content": "16.5\\%" + }, + { + "bbox": [ + 313, + 378, + 554, + 665 + ], + "type": "text", + "content": " in 3-step prediction, and " + }, + { + "bbox": [ + 313, + 378, + 554, + 665 + ], + "type": "inline_equation", + "content": "18.0\\%" + }, + { + "bbox": [ + 313, + 378, + 554, + 665 + ], + "type": "text", + "content": " in 4-step prediction, while adopting the same network architecture. Our work shows some advantages over existing RL baselines. Since few features are extracted from the scalar input of [6], it can only guide the lens towards one side with the higher label frequency, causing extremely low accuracy but 0 FH rate. Since [54] generalizes poorly in various views (as stated by the authors), while the dataset in our work contains images with multiple views and object distances, it hardly gives a move signal. The lens thus stops at the start position in most cases, causing small changes in accuracy and a low FH rate." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 666, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 666, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 666, + 553, + 713 + ], + "type": "text", + "content": "For the multi-step prediction performance, the accuracy of the baseline [7] increases a little and even decreases in 4-step prediction. The accuracy of the model obtained by the relative movement label can achieve substantial gains" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26445" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 88, + 294, + 259 + ], + "blocks": [ + { + "bbox": [ + 56, + 65, + 295, + 87 + ], + "lines": [ + { + "bbox": [ + 56, + 65, + 295, + 87 + ], + "spans": [ + { + "bbox": [ + 56, + 65, + 295, + 87 + ], + "type": "text", + "content": "Table 1. Prediction accuracy and focus hunting (FH) rate results in multi-step setting." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 88, + 294, + 259 + ], + "lines": [ + { + "bbox": [ + 56, + 88, + 294, + 259 + ], + "spans": [ + { + "bbox": [ + 56, + 88, + 294, + 259 + ], + "type": "table", + "html": "
AlgorithmStep= 0 ↑ ≤ 1 ↑ ≤ 2 ↑ ≤ 4 ↑MAE↓ RMSE↓FH↓
DRL in [6]10.030 0.071 0.117 0.21014.216 17.908NA
20.029 0.050 0.082 0.15117.214 20.8270.0
30.029 0.050 0.082 0.15117.214 20.8270.0
40.029 0.050 0.082 0.15117.214 20.8270.0
DRL in [54]10.022 0.062 0.102 0.18115.575 19.143NA
20.023 0.064 0.105 0.18515.446 19.0510.019
30.024 0.065 0.106 0.18715.437 19.0730.025
40.024 0.065 0.106 0.18615.500 19.1420.030
RoI-PE & Len-PE [7]10.243 0.584 0.756 0.8822.839 6.880NA
20.275 0.640 0.803 0.9122.454 6.2960.275
30.279 0.644 0.810 0.9142.434 6.2050.362
40.276 0.642 0.807 0.9132.442 6.2770.383
[7] with relative label (Ours)10.330 0.698 0.826 0.9072.342 6.339NA
20.413 0.795 0.888 0.9341.894 5.8550.261
30.427 0.805 0.894 0.9371.892 5.7210.340
40.423 0.804 0.893 0.9361.862 5.8440.358
Expert Trajectory (Ours)10.336 0.711 0.839 0.9152.146 5.827NA
20.439 0.808 0.896 0.9411.755 5.5970.170
30.450 0.816 0.900 0.9431.711 5.5320.175
40.449 0.818 0.901 0.9441.694 5.4790.178
", + "image_path": "e51e2485f25370ef97a3ca3a10a83f5eb18538964f71ab0ea3b0e98c95af7b30.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 58, + 278, + 291, + 313 + ], + "blocks": [ + { + "bbox": [ + 58, + 266, + 292, + 277 + ], + "lines": [ + { + "bbox": [ + 58, + 266, + 292, + 277 + ], + "spans": [ + { + "bbox": [ + 58, + 266, + 292, + 277 + ], + "type": "text", + "content": "Table 2. Average steps to stop with different termination signals." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 278, + 291, + 313 + ], + "lines": [ + { + "bbox": [ + 58, + 278, + 291, + 313 + ], + "spans": [ + { + "bbox": [ + 58, + 278, + 291, + 313 + ], + "type": "table", + "html": "
Termination Rule= 0 ↓ ≤ 1 ↓ ≤ 2 ↓ ≤ 4 ↓Stop Model↓ MAE↓ RMSE↓
RoI-PE & Lens-PE[7]6.6513.6442.4771.7072.2082.6116.146
[7] with relative label (Ours)4.9822.2461.7041.4531.8841.9065.463
Expert Trajectory (Ours)5.0142.2571.6831.3611.8211.8745.332
", + "image_path": "31051fadf2aef18550018aa2c19ea3501f9d41623db343f197147a8c3aec3d34.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 317, + 295, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 317, + 295, + 558 + ], + "spans": [ + { + "bbox": [ + 55, + 317, + 295, + 558 + ], + "type": "text", + "content": "of " + }, + { + "bbox": [ + 55, + 317, + 295, + 558 + ], + "type": "inline_equation", + "content": "13.8\\%" + }, + { + "bbox": [ + 55, + 317, + 295, + 558 + ], + "type": "text", + "content": " prediction accuracy for exact " + }, + { + "bbox": [ + 55, + 317, + 295, + 558 + ], + "type": "inline_equation", + "content": "(= 0)" + }, + { + "bbox": [ + 55, + 317, + 295, + 558 + ], + "type": "text", + "content": " predictions in 2-step prediction, while our DRL-based model further enhances this accuracy. Furthermore, as the number of lens movement steps increases, the accuracy of deep learning-based methods may decline, whereas the accuracy of our DRL-based method remains unchanged or even improves, demonstrating the stability of our DRL-based model in multi-step prediction. Moreover, Tab. 2 shows that our models take fewer steps to stop at different termination signals, e.g., " + }, + { + "bbox": [ + 55, + 317, + 295, + 558 + ], + "type": "inline_equation", + "content": "= 0" + }, + { + "bbox": [ + 55, + 317, + 295, + 558 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 317, + 295, + 558 + ], + "type": "inline_equation", + "content": "\\leq 1" + }, + { + "bbox": [ + 55, + 317, + 295, + 558 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 317, + 295, + 558 + ], + "type": "inline_equation", + "content": "\\leq 2" + }, + { + "bbox": [ + 55, + 317, + 295, + 558 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 317, + 295, + 558 + ], + "type": "inline_equation", + "content": "\\leq 4" + }, + { + "bbox": [ + 55, + 317, + 295, + 558 + ], + "type": "text", + "content": ". Since in-focus positions remain unknown in practical camera deployment, the above rules may not be feasible. Alternatively, we train a stop model to generate the stop signal, which is modeled as a binary classification task and takes the same architecture and input as our policy network. Tab. 2 shows the average steps, MAE, and RMSE of different policies with this model. Our method still achieves a better prediction with fewer steps. In conclusion, results in Tab. 1 and Tab. 2 show that our methods can achieve higher accuracy with fewer steps than that proposed by [7], demonstrating a faster focusing speed." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 558, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 295, + 714 + ], + "type": "text", + "content": "Single-step accuracy (Effects of relative movement as labels). The comparison results in the single-step permutation setting are shown in Tab. 3. Compared with classic statistics-based or depth estimation methods, the deep-learning based models \"Learning to Autofocus [16]\" and \"RoI-PE & Len-PE [7]\" significantly improve the one-step prediction accuracy. By modifying the output of the model in [7] from the absolute lens position to the relative movement, we can further improve the accuracy by " + }, + { + "bbox": [ + 55, + 558, + 295, + 714 + ], + "type": "inline_equation", + "content": "8.7\\%" + }, + { + "bbox": [ + 55, + 558, + 295, + 714 + ], + "type": "text", + "content": " for exact " + }, + { + "bbox": [ + 55, + 558, + 295, + 714 + ], + "type": "inline_equation", + "content": "(= 0)" + }, + { + "bbox": [ + 55, + 558, + 295, + 714 + ], + "type": "text", + "content": " prediction and " + }, + { + "bbox": [ + 55, + 558, + 295, + 714 + ], + "type": "inline_equation", + "content": "2.5\\%" + }, + { + "bbox": [ + 55, + 558, + 295, + 714 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 55, + 558, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\leq 4" + }, + { + "bbox": [ + 55, + 558, + 295, + 714 + ], + "type": "text", + "content": " focal index difference. Moreover, since the dataset has low-contrast images with similar features but distant labels, minimizing the loss makes the policy perform badly. Via optimizing multi-step" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 317, + 77, + 551, + 152 + ], + "blocks": [ + { + "bbox": [ + 364, + 65, + 504, + 76 + ], + "lines": [ + { + "bbox": [ + 364, + 65, + 504, + 76 + ], + "spans": [ + { + "bbox": [ + 364, + 65, + 504, + 76 + ], + "type": "text", + "content": "Table 3. Single-step prediction results." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 77, + 551, + 152 + ], + "lines": [ + { + "bbox": [ + 317, + 77, + 551, + 152 + ], + "spans": [ + { + "bbox": [ + 317, + 77, + 551, + 152 + ], + "type": "table", + "html": "
Algorithm= 0 ↑≤ 1 ↑≤ 2 ↑≤ 4 ↑MAE↓RMSE↓
ZNCC Disparity with Calibration0.0650.1790.2680.4199.47614.211
SSD Disparity [44]0.1010.2640.3860.5258.01212.414
Learning the Depth [11]0.1120.2970.4150.5627.41912.312
Learning to Autofocus [16]0.1680.4670.6590.8513.2497.619
RoI-PE & Lens-PE [7]0.2430.5840.7560.8822.8396.880
[16] with relative label (Ours)0.1810.4930.6840.8613.0217.374
[7] with relative label(Ours)0.3300.6980.8260.9072.3426.339
Expert Trajectory (Ours)0.3360.7110.8390.9152.1465.827
", + "image_path": "13dd255546095cb21baed0f364d83c3b3ee7a87246096e2579aad69f1998f56c.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 321, + 179, + 547, + 255 + ], + "blocks": [ + { + "bbox": [ + 314, + 156, + 553, + 178 + ], + "lines": [ + { + "bbox": [ + 314, + 156, + 553, + 178 + ], + "spans": [ + { + "bbox": [ + 314, + 156, + 553, + 178 + ], + "type": "text", + "content": "Table 4. Performance comparison of DRL models with and without Expert Regularization." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 321, + 179, + 547, + 255 + ], + "lines": [ + { + "bbox": [ + 321, + 179, + 547, + 255 + ], + "spans": [ + { + "bbox": [ + 321, + 179, + 547, + 255 + ], + "type": "table", + "html": "
AlgorithmStep= 0 ↑ ≤ 1 ↑ ≤ 2 ↑ ≤ 4 ↑MAE↓ RMSE↓FH↓
w/o10.0310.0560.0860.14322.00526.048
20.0290.0650.0950.14823.84527.975
Expert Regularization30.0370.0850.1230.18820.95725.501
40.0310.0680.0990.15223.72227.889
Expert Trajectory (Ours)10.3360.7110.8390.9152.1465.827
20.4390.8080.8960.9411.7555.597
30.4500.8160.9000.9431.7115.532
40.4490.8180.9010.9441.6945.479
", + "image_path": "b0df48b1e4ead96273fa26d727316242a72daf32277aa2adc7bccce3cd3f8981.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 257, + 553, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 257, + 553, + 281 + ], + "spans": [ + { + "bbox": [ + 313, + 257, + 553, + 281 + ], + "type": "text", + "content": "behavior, RL can explore extra information and improve 1-step prediction accuracy." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 281, + 553, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 281, + 553, + 354 + ], + "spans": [ + { + "bbox": [ + 313, + 281, + 553, + 354 + ], + "type": "text", + "content": "Effects of expert regularization. As shown in Tab. 4, the policy obtained by the vanilla DRL algorithm without expert regularization has a poor performance, since the action space is a discrete number of 97, which is large for DRL's training. The expert trajectory significantly improves the performance of the policy given by the PPO algorithm." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 360, + 425, + 373 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 360, + 425, + 373 + ], + "spans": [ + { + "bbox": [ + 314, + 360, + 425, + 373 + ], + "type": "text", + "content": "4.2. Qualitative Results" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 378, + 554, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 378, + 554, + 533 + ], + "spans": [ + { + "bbox": [ + 313, + 378, + 554, + 533 + ], + "type": "text", + "content": "Fig. 1 qualitatively visualizes the trajectory of patches and lens positions of different comparison methods, including the policy obtained by [7], the policy obtained by relative movement label, and our DRL-based model. Compared with directly predicting the lens position [7], our relative movement label model is faster by predicting the lens movement within DoF in the second step, but the FH issue cannot be overcome. In comparison, our DRL-based model can generate a faster and more stable prediction by predicting the GT location at the first step and stopping at this index. The lens trajectory of the DRL-based model (the green curve in Fig. 2) is slightly smoother and more stable without the occurrence of FH." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "text", + "content": "In Fig. 6, we show qualitative results of 4-step predictions of the three models. From the first row, we can observe that all methods can predict the index within DoF, but our methods can give a more precise prediction with a sharper image than Choi's [7]. From the second to fourth rows, we can observe that [7] sometimes breaks and predicts indices with sharp background, while our DRL-based model successfully predicts indices with sharp target objects. This case is a generally challenging scenario in the AF task since many different DoFs can be estimated. These DoFs may confuse the models, leading to a focus on the wrong object. In general, our relative movement label can give a more accurate prediction by mapping the disparity to the movement of the lens, and the DRL-based model can further enhance the accuracy by estimating the future reward of the action" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "26446" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 65, + 99, + 108 + ], + "blocks": [ + { + "bbox": [ + 58, + 65, + 99, + 108 + ], + "lines": [ + { + "bbox": [ + 58, + 65, + 99, + 108 + ], + "spans": [ + { + "bbox": [ + 58, + 65, + 99, + 108 + ], + "type": "image", + "image_path": "189850b0d961d43ad0fee67ed1398510478dcd6c775f823cf628308068b84b05.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 58, + 109, + 99, + 152 + ], + "blocks": [ + { + "bbox": [ + 58, + 109, + 99, + 152 + ], + "lines": [ + { + "bbox": [ + 58, + 109, + 99, + 152 + ], + "spans": [ + { + "bbox": [ + 58, + 109, + 99, + 152 + ], + "type": "image", + "image_path": "7ef550bb2c179357b8f414db0273bc6e558ed66aee57c8d1f98e4ef91e791643.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 58, + 152, + 99, + 194 + ], + "blocks": [ + { + "bbox": [ + 58, + 152, + 99, + 194 + ], + "lines": [ + { + "bbox": [ + 58, + 152, + 99, + 194 + ], + "spans": [ + { + "bbox": [ + 58, + 152, + 99, + 194 + ], + "type": "image", + "image_path": "10775c45e94d5fdf3bf7cd35da086e3acb78975b9234083d280a2b8a4e127c7b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 58, + 196, + 99, + 238 + ], + "blocks": [ + { + "bbox": [ + 58, + 196, + 99, + 238 + ], + "lines": [ + { + "bbox": [ + 58, + 196, + 99, + 238 + ], + "spans": [ + { + "bbox": [ + 58, + 196, + 99, + 238 + ], + "type": "image", + "image_path": "89e439505bb9e65d78ff78db058b701e14f2bcae2c6524991ebd201afadacf45.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 67, + 239, + 83, + 246 + ], + "lines": [ + { + "bbox": [ + 67, + 239, + 83, + 246 + ], + "spans": [ + { + "bbox": [ + 67, + 239, + 83, + 246 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 251, + 295, + 273 + ], + "lines": [ + { + "bbox": [ + 55, + 251, + 295, + 273 + ], + "spans": [ + { + "bbox": [ + 55, + 251, + 295, + 273 + ], + "type": "text", + "content": "Figure 6. Comparison of the 4-step results of the policy obtained by [7], our relative movement label, and our DRL-based model." + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 105, + 65, + 147, + 108 + ], + "blocks": [ + { + "bbox": [ + 105, + 65, + 147, + 108 + ], + "lines": [ + { + "bbox": [ + 105, + 65, + 147, + 108 + ], + "spans": [ + { + "bbox": [ + 105, + 65, + 147, + 108 + ], + "type": "image", + "image_path": "78f6eece17673bd4dd2c22ce6c3074b32640f19a513a6426179e138f10f064f4.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 105, + 109, + 147, + 152 + ], + "blocks": [ + { + "bbox": [ + 105, + 109, + 147, + 152 + ], + "lines": [ + { + "bbox": [ + 105, + 109, + 147, + 152 + ], + "spans": [ + { + "bbox": [ + 105, + 109, + 147, + 152 + ], + "type": "image", + "image_path": "b82c6818fc24ce9026b74b5f761c40a0fd2625c22ebdc4e9bdf22c3c95d56232.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 105, + 152, + 147, + 194 + ], + "blocks": [ + { + "bbox": [ + 105, + 152, + 147, + 194 + ], + "lines": [ + { + "bbox": [ + 105, + 152, + 147, + 194 + ], + "spans": [ + { + "bbox": [ + 105, + 152, + 147, + 194 + ], + "type": "image", + "image_path": "ea241c0050c70f1f2808b8fe0ad410f67a9769b65aaa20932efdc68a895e1902.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 105, + 196, + 147, + 237 + ], + "blocks": [ + { + "bbox": [ + 105, + 196, + 147, + 237 + ], + "lines": [ + { + "bbox": [ + 105, + 196, + 147, + 237 + ], + "spans": [ + { + "bbox": [ + 105, + 196, + 147, + 237 + ], + "type": "image", + "image_path": "bf2fe82be806fc3c087b64d8d6ff8ec3e252dd80601d8e358b8e69828ae4cbb0.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 113, + 237, + 140, + 250 + ], + "lines": [ + { + "bbox": [ + 113, + 237, + 140, + 250 + ], + "spans": [ + { + "bbox": [ + 113, + 237, + 140, + 250 + ], + "type": "text", + "content": "absolute position labe" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 154, + 65, + 196, + 108 + ], + "blocks": [ + { + "bbox": [ + 154, + 65, + 196, + 108 + ], + "lines": [ + { + "bbox": [ + 154, + 65, + 196, + 108 + ], + "spans": [ + { + "bbox": [ + 154, + 65, + 196, + 108 + ], + "type": "image", + "image_path": "efafaabf2fa673ff7cc6bd3175916ff61936d57e1b0fff569d99ed395e28e96c.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 154, + 109, + 196, + 152 + ], + "blocks": [ + { + "bbox": [ + 154, + 109, + 196, + 152 + ], + "lines": [ + { + "bbox": [ + 154, + 109, + 196, + 152 + ], + "spans": [ + { + "bbox": [ + 154, + 109, + 196, + 152 + ], + "type": "image", + "image_path": "4b8cd13827cfd0c97b2b318dd89c3cddc8ea7552c92a67969fcfeee0aaed2526.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 154, + 152, + 196, + 194 + ], + "blocks": [ + { + "bbox": [ + 154, + 152, + 196, + 194 + ], + "lines": [ + { + "bbox": [ + 154, + 152, + 196, + 194 + ], + "spans": [ + { + "bbox": [ + 154, + 152, + 196, + 194 + ], + "type": "image", + "image_path": "eefd1c948dfb5ffecb4862cbf822041a769d9a2f22a7e0d666b8a2cc4e95ce68.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 154, + 196, + 196, + 237 + ], + "blocks": [ + { + "bbox": [ + 154, + 196, + 196, + 237 + ], + "lines": [ + { + "bbox": [ + 154, + 196, + 196, + 237 + ], + "spans": [ + { + "bbox": [ + 154, + 196, + 196, + 237 + ], + "type": "image", + "image_path": "fde21e27e4ef700bf2ed5ed3d8139c3a2804b3b86cf0a484b005881104cfef45.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 155, + 238, + 195, + 250 + ], + "lines": [ + { + "bbox": [ + 155, + 238, + 195, + 250 + ], + "spans": [ + { + "bbox": [ + 155, + 238, + 195, + 250 + ], + "type": "text", + "content": "relative movement label" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 203, + 65, + 244, + 108 + ], + "blocks": [ + { + "bbox": [ + 203, + 65, + 244, + 108 + ], + "lines": [ + { + "bbox": [ + 203, + 65, + 244, + 108 + ], + "spans": [ + { + "bbox": [ + 203, + 65, + 244, + 108 + ], + "type": "image", + "image_path": "d3c97875496f60449e9e4f77cea564696ebceb14b2186a12b8b80ec1fe6a8eab.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 203, + 109, + 244, + 152 + ], + "blocks": [ + { + "bbox": [ + 203, + 109, + 244, + 152 + ], + "lines": [ + { + "bbox": [ + 203, + 109, + 244, + 152 + ], + "spans": [ + { + "bbox": [ + 203, + 109, + 244, + 152 + ], + "type": "image", + "image_path": "857994789c5078b95175cc548f5a0834d53b1f55f0076d34a3d3a359d2ea8664.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 203, + 152, + 243, + 194 + ], + "blocks": [ + { + "bbox": [ + 203, + 152, + 243, + 194 + ], + "lines": [ + { + "bbox": [ + 203, + 152, + 243, + 194 + ], + "spans": [ + { + "bbox": [ + 203, + 152, + 243, + 194 + ], + "type": "image", + "image_path": "96d52a5abbf949dc102f24a0e067b608f2ba4dfb926f9fc923e2c76b1f8cc5bd.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 203, + 196, + 243, + 237 + ], + "blocks": [ + { + "bbox": [ + 203, + 196, + 243, + 237 + ], + "lines": [ + { + "bbox": [ + 203, + 196, + 243, + 237 + ], + "spans": [ + { + "bbox": [ + 203, + 196, + 243, + 237 + ], + "type": "image", + "image_path": "467450180faa395da3e2eb6560da926580676ff3e1e4441175731705dce5e67f.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 217, + 238, + 230, + 243 + ], + "lines": [ + { + "bbox": [ + 217, + 238, + 230, + 243 + ], + "spans": [ + { + "bbox": [ + 217, + 238, + 230, + 243 + ], + "type": "text", + "content": "DRL" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 250, + 65, + 291, + 108 + ], + "blocks": [ + { + "bbox": [ + 250, + 65, + 291, + 108 + ], + "lines": [ + { + "bbox": [ + 250, + 65, + 291, + 108 + ], + "spans": [ + { + "bbox": [ + 250, + 65, + 291, + 108 + ], + "type": "image", + "image_path": "0ad31ed71439c73d490282239f8ce6e06935e40d55b7903e2b878bb3525e7b7b.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 251, + 109, + 291, + 152 + ], + "blocks": [ + { + "bbox": [ + 251, + 109, + 291, + 152 + ], + "lines": [ + { + "bbox": [ + 251, + 109, + 291, + 152 + ], + "spans": [ + { + "bbox": [ + 251, + 109, + 291, + 152 + ], + "type": "image", + "image_path": "2ab5c69b8f5de88f2d319a32fd4a8279eb7e953672395b388e51fb2b27f871df.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 251, + 152, + 291, + 194 + ], + "blocks": [ + { + "bbox": [ + 251, + 152, + 291, + 194 + ], + "lines": [ + { + "bbox": [ + 251, + 152, + 291, + 194 + ], + "spans": [ + { + "bbox": [ + 251, + 152, + 291, + 194 + ], + "type": "image", + "image_path": "b8768a93885661caf0cb400a7a38f503ddc89590a6bc785298a856f12178f9a8.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 250, + 196, + 291, + 237 + ], + "blocks": [ + { + "bbox": [ + 250, + 196, + 291, + 237 + ], + "lines": [ + { + "bbox": [ + 250, + 196, + 291, + 237 + ], + "spans": [ + { + "bbox": [ + 250, + 196, + 291, + 237 + ], + "type": "image", + "image_path": "86b1d4250541142213e2df685a1f15c842c7313c19d2acebe97fbce72561f8c8.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 266, + 238, + 276, + 245 + ], + "lines": [ + { + "bbox": [ + 266, + 238, + 276, + 245 + ], + "spans": [ + { + "bbox": [ + 266, + 238, + 276, + 245 + ], + "type": "text", + "content": "GT" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "bbox": [ + 55, + 281, + 201, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 281, + 201, + 293 + ], + "spans": [ + { + "bbox": [ + 55, + 281, + 201, + 293 + ], + "type": "text", + "content": "and reducing the focus hunting rate." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 55, + 293, + 296, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 293, + 296, + 568 + ], + "spans": [ + { + "bbox": [ + 55, + 293, + 296, + 568 + ], + "type": "text", + "content": "In Fig. 7, we analyze our model's failure cases: (a) it is difficult and meaningless to focus on the images with no texture since the phase information or contrast information is little to estimate the depth, (b) though our models can give a more accurate prediction in scenarios with multiple DoFs, when the target object is small and the background is complex, the models may be confused by these DoFs, leading to focusing on the wrong object, (c) like all DP data-based AF methods, our models may work poorly when the objects are horizontal. Some possible solutions may work to overcome these limitations. For (a), larger patches can provide more phase information, and we also think (a) might not be that important since few people take photos with no texture. For (b), we claim it is not a bad issue, because sometimes people may want to focus on the background obscured by meaningless foreground elements. To solve this problem, larger patches with more information about the target objects may be useful. Some depth estimation methods [22, 52] can also be used to provide more depth information to help distinguish the objects and background. For (c), solving this problem is challenging in algorithms, while little changes in hardware, especially sensors, could easily mitigate this issue [7]." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 55, + 575, + 149, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 575, + 149, + 588 + ], + "spans": [ + { + "bbox": [ + 55, + 575, + 149, + 588 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "content": "Effects of relative label method. Tab. 5 further demonstrates the effect of relative movement as labels. Our DRL-based model can further enhance the accuracy by " + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "inline_equation", + "content": "0.6\\%" + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "content": " for exact " + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "inline_equation", + "content": "(= 0)" + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "content": " prediction and " + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "inline_equation", + "content": "0.8\\%" + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\leq 4" + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "content": " focal index difference, in comparison to the model in [7] modified with the relative movement label. We further compare the two label methods under three different architectures: MobileNetV2 [33], ResNet50 [15], and MnasNet [41]. The relative movement label method outperforms the absolute lens position label for each architecture. For a fairer comparison," + } + ] + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 315, + 63, + 396, + 110 + ], + "blocks": [ + { + "bbox": [ + 315, + 63, + 396, + 110 + ], + "lines": [ + { + "bbox": [ + 315, + 63, + 396, + 110 + ], + "spans": [ + { + "bbox": [ + 315, + 63, + 396, + 110 + ], + "type": "image", + "image_path": "41f4ba9be68136a6609bde276a95e00abe544c901bb98dd5d3da776f3231a3a7.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 316, + 112, + 395, + 158 + ], + "blocks": [ + { + "bbox": [ + 316, + 112, + 395, + 158 + ], + "lines": [ + { + "bbox": [ + 316, + 112, + 395, + 158 + ], + "spans": [ + { + "bbox": [ + 316, + 112, + 395, + 158 + ], + "type": "image", + "image_path": "a8df26c18e23a57b065915fd84b4a91fcf1d37d148b5b09cbcf383b753c72c37.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 316, + 159, + 395, + 206 + ], + "blocks": [ + { + "bbox": [ + 316, + 159, + 395, + 206 + ], + "lines": [ + { + "bbox": [ + 316, + 159, + 395, + 206 + ], + "spans": [ + { + "bbox": [ + 316, + 159, + 395, + 206 + ], + "type": "image", + "image_path": "876be92ae60c1986ec51e9bb29655bb9209f2bdafc61c731b42d68f7940ff435.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 364, + 207, + 378, + 213 + ], + "lines": [ + { + "bbox": [ + 364, + 207, + 378, + 213 + ], + "spans": [ + { + "bbox": [ + 364, + 207, + 378, + 213 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 220, + 553, + 242 + ], + "lines": [ + { + "bbox": [ + 313, + 220, + 553, + 242 + ], + "spans": [ + { + "bbox": [ + 313, + 220, + 553, + 242 + ], + "type": "text", + "content": "Figure 7. Failure cases of our models, where only the left image of DP data is visualized." + } + ] + } + ], + "index": 46, + "angle": 0, + "type": "image_caption" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 400, + 63, + 448, + 110 + ], + "blocks": [ + { + "bbox": [ + 400, + 63, + 448, + 110 + ], + "lines": [ + { + "bbox": [ + 400, + 63, + 448, + 110 + ], + "spans": [ + { + "bbox": [ + 400, + 63, + 448, + 110 + ], + "type": "image", + "image_path": "3c0fb69380e47e9ddba384a23828fd7424aa366bfbffb1d5a684797ebcc48a87.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 400, + 111, + 448, + 158 + ], + "blocks": [ + { + "bbox": [ + 400, + 111, + 448, + 158 + ], + "lines": [ + { + "bbox": [ + 400, + 111, + 448, + 158 + ], + "spans": [ + { + "bbox": [ + 400, + 111, + 448, + 158 + ], + "type": "image", + "image_path": "c1f40623d2d637e11ecf939f88911970f2e5c0a32aace19802722016f9604c9e.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + } + ], + "index": 35 + }, + { + "type": "image", + "bbox": [ + 400, + 159, + 448, + 206 + ], + "blocks": [ + { + "bbox": [ + 400, + 159, + 448, + 206 + ], + "lines": [ + { + "bbox": [ + 400, + 159, + 448, + 206 + ], + "spans": [ + { + "bbox": [ + 400, + 159, + 448, + 206 + ], + "type": "image", + "image_path": "9a25ae0cd0e8e9778aa4007d1f036074b6b133749f7dff5dc13e1ca43fa664af.jpg" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 401, + 207, + 447, + 220 + ], + "lines": [ + { + "bbox": [ + 401, + 207, + 447, + 220 + ], + "spans": [ + { + "bbox": [ + 401, + 207, + 447, + 220 + ], + "type": "text", + "content": "relative movement label" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_caption" + } + ], + "index": 36 + }, + { + "type": "image", + "bbox": [ + 452, + 63, + 500, + 110 + ], + "blocks": [ + { + "bbox": [ + 452, + 63, + 500, + 110 + ], + "lines": [ + { + "bbox": [ + 452, + 63, + 500, + 110 + ], + "spans": [ + { + "bbox": [ + 452, + 63, + 500, + 110 + ], + "type": "image", + "image_path": "b0dc5b5557066efbeb75091808f0eeb0a4e2c71116d6eb274e59ba5ea3059ba5.jpg" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_body" + } + ], + "index": 38 + }, + { + "type": "image", + "bbox": [ + 452, + 111, + 500, + 158 + ], + "blocks": [ + { + "bbox": [ + 452, + 111, + 500, + 158 + ], + "lines": [ + { + "bbox": [ + 452, + 111, + 500, + 158 + ], + "spans": [ + { + "bbox": [ + 452, + 111, + 500, + 158 + ], + "type": "image", + "image_path": "fd94b8464e74e9842e7df2e9f7c6e616b7df34c862bdee3d979d4cfea7bc5620.jpg" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_body" + } + ], + "index": 39 + }, + { + "type": "image", + "bbox": [ + 452, + 159, + 500, + 206 + ], + "blocks": [ + { + "bbox": [ + 452, + 159, + 500, + 206 + ], + "lines": [ + { + "bbox": [ + 452, + 159, + 500, + 206 + ], + "spans": [ + { + "bbox": [ + 452, + 159, + 500, + 206 + ], + "type": "image", + "image_path": "da7145b1c0ae1c4e6e240e953efa5b4a104a0134b22a66292415e251c55142b8.jpg" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 468, + 207, + 481, + 213 + ], + "lines": [ + { + "bbox": [ + 468, + 207, + 481, + 213 + ], + "spans": [ + { + "bbox": [ + 468, + 207, + 481, + 213 + ], + "type": "text", + "content": "DRL" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_caption" + } + ], + "index": 40 + }, + { + "type": "image", + "bbox": [ + 504, + 63, + 552, + 110 + ], + "blocks": [ + { + "bbox": [ + 504, + 63, + 552, + 110 + ], + "lines": [ + { + "bbox": [ + 504, + 63, + 552, + 110 + ], + "spans": [ + { + "bbox": [ + 504, + 63, + 552, + 110 + ], + "type": "image", + "image_path": "198eadb473dab9abd3b30e2046853b3f12eedb4fb0a39e1a0562a6214deac023.jpg" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_body" + } + ], + "index": 42 + }, + { + "type": "image", + "bbox": [ + 504, + 111, + 552, + 158 + ], + "blocks": [ + { + "bbox": [ + 504, + 111, + 552, + 158 + ], + "lines": [ + { + "bbox": [ + 504, + 111, + 552, + 158 + ], + "spans": [ + { + "bbox": [ + 504, + 111, + 552, + 158 + ], + "type": "image", + "image_path": "5bbeb3708e3d335caad1d7198299f0958d57b97c85048ec9876e845a185b2ce5.jpg" + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_body" + } + ], + "index": 43 + }, + { + "type": "image", + "bbox": [ + 504, + 159, + 552, + 206 + ], + "blocks": [ + { + "bbox": [ + 504, + 159, + 552, + 206 + ], + "lines": [ + { + "bbox": [ + 504, + 159, + 552, + 206 + ], + "spans": [ + { + "bbox": [ + 504, + 159, + 552, + 206 + ], + "type": "image", + "image_path": "0b567bd653370c2878ce35fdbe574a2b7399da2794508b30548a650fdf25dd1a.jpg" + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 520, + 207, + 529, + 213 + ], + "lines": [ + { + "bbox": [ + 520, + 207, + 529, + 213 + ], + "spans": [ + { + "bbox": [ + 520, + 207, + 529, + 213 + ], + "type": "text", + "content": "GT" + } + ] + } + ], + "index": 45, + "angle": 0, + "type": "image_caption" + } + ], + "index": 44 + }, + { + "type": "table", + "bbox": [ + 315, + 277, + 558, + 381 + ], + "blocks": [ + { + "bbox": [ + 313, + 251, + 553, + 274 + ], + "lines": [ + { + "bbox": [ + 313, + 251, + 553, + 274 + ], + "spans": [ + { + "bbox": [ + 313, + 251, + 553, + 274 + ], + "type": "text", + "content": "Table 5. Single-step prediction results with different network architectures. The number of input channels is 5." + } + ] + } + ], + "index": 47, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 277, + 558, + 381 + ], + "lines": [ + { + "bbox": [ + 315, + 277, + 558, + 381 + ], + "spans": [ + { + "bbox": [ + 315, + 277, + 558, + 381 + ], + "type": "table", + "html": "
Algorithm# of paramslabel= 0 ↑ ≤ 1 ↑ ≤ 2 ↑ ≤ 4 ↑MAE↓RMSE↓
MobileNetV2 [33]2.29Mabsolute0.2430.5840.7560.8822.8396.880
2.35Mabsolute0.2520.5920.7640.8882.7196.745
2.35Mrelative0.3300.6980.8260.9072.3426.339
ResNet50 [15]23.61Mabsolute0.2580.6020.7890.8942.6156.623
23.71Mabsolute0.2540.5960.7810.8912.6556.686
23.71Mrelative0.3260.7090.8490.9182.1365.764
MnasNet [41]1.00Mabsolute0.2260.5680.7360.8642.9837.016
1.06Mabsolute0.2130.5490.7180.8503.1197.318
1.06Mrelative0.3050.6780.8170.9052.4366.402
", + "image_path": "8e1a0a0c0d6795dec945f013142c2a9d3622af9a199ab1218bf05858dfa190c5.jpg" + } + ] + } + ], + "index": 48, + "angle": 0, + "type": "table_body" + } + ], + "index": 48 + }, + { + "bbox": [ + 313, + 388, + 555, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 388, + 555, + 520 + ], + "spans": [ + { + "bbox": [ + 313, + 388, + 555, + 520 + ], + "type": "text", + "content": "we expand the absolute position label to the same range as the relative movement label. It shows that expanding the number of networks' output channels can not improve the accuracy, and the effects remain unknown. Since Lens-PE mainly conveys the lens' current position to the network, while the next position must be predicted by adding its movement to the current position, the relative label methods avoid additionally inferring the identity mapping for Lens-PE, leading to a higher accuracy. To make a fair comparison with the other learning-based models, we thus adopt the same MobileNetV2 architecture in the experiments." + } + ] + } + ], + "index": 49 + }, + { + "bbox": [ + 313, + 536, + 388, + 548 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 536, + 388, + 548 + ], + "spans": [ + { + "bbox": [ + 313, + 536, + 388, + 548 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 50 + }, + { + "bbox": [ + 313, + 557, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 557, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 557, + 555, + 713 + ], + "type": "text", + "content": "We have proposed a DRL-based autofocus model with expert trajectory regularization that can both improve the prediction accuracy and reduce the focus hunting rate, which took an actor-critic framework based on the pretrained MobileNet-V2 architecture for single-step prediction. For the actor network, we further modified the absolute position label in [7, 16] with the relative movement label method. We also constructed an offline dataset of expert trajectories based on prior knowledge, regularizing the policy network training toward reducing focus hunting. Experimental results have clearly demonstrated the effectiveness of our proposed method, in terms of the multi-step prediction accuracy and the occurrence of focus hunting." + } + ] + } + ], + "index": 51 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26447" + } + ] + } + ], + "index": 52 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 194, + 115, + 205 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 194, + 115, + 205 + ], + "spans": [ + { + "bbox": [ + 56, + 194, + 115, + 205 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 213, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 213, + 295, + 269 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 213, + 295, + 269 + ], + "spans": [ + { + "bbox": [ + 61, + 213, + 295, + 269 + ], + "type": "text", + "content": "[1] Abdullah Abuolaim, Mauricio Delbracio, Damien Kelly, Michael S. Brown, and Peyman Milanfar. Learning to reduce defocus blur by realistically modeling dual-pixel data. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 2289-2298, 2021. 1, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 270, + 295, + 302 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 270, + 295, + 302 + ], + "spans": [ + { + "bbox": [ + 61, + 270, + 295, + 302 + ], + "type": "text", + "content": "[2] Yuhan Bao, Lei Sun, Yuqin Ma, Diyang Gu, and Kaiwei Wang. Improving fast auto-focus with event polarity. Optics Express, 31(15):24025-24044, 2023. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 303, + 295, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 303, + 295, + 335 + ], + "spans": [ + { + "bbox": [ + 62, + 303, + 295, + 335 + ], + "type": "text", + "content": "[3] Daniel I Barnea and Harvey F Silverman. A class of algorithms for fast digital image registration. IEEE transactions on Computers, 100(2):179-186, 1972. 1, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 336, + 295, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 336, + 295, + 390 + ], + "spans": [ + { + "bbox": [ + 62, + 336, + 295, + 390 + ], + "type": "text", + "content": "[4] Albert Bou, Matteo Bettini, Sebastian Dittert, Vikash Kumar, Shagun Sodhani, Xiaomeng Yang, Gianni De Fabritis, and Vincent Moens. TorchRL: A data-driven decision-making library for pytorch. arXiv preprint arXiv:2306.00577, 2023. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 392, + 296, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 392, + 296, + 456 + ], + "spans": [ + { + "bbox": [ + 62, + 392, + 296, + 456 + ], + "type": "text", + "content": "[5] Serkan Cabi, Sergio Gómez Colmenarejo, Alexander Novikov, Ksenia Konyushkova, Scott Reed, Rae Jeong, Konrad Zolna, Yusuf Aytar, David Budden, Mel Vecerik, et al. Scaling data-driven robotics with reward sketching and batch reinforcement learning. arXiv preprint arXiv:1909.12200, 2019. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 458, + 294, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 458, + 294, + 491 + ], + "spans": [ + { + "bbox": [ + 62, + 458, + 294, + 491 + ], + "type": "text", + "content": "[6] Chin-Cheng Chan and Homer H Chen. Autofocus by deep reinforcement learning. *Electronic Imaging*, 31:1-6, 2019. 3, 6, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 491, + 295, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 491, + 295, + 546 + ], + "spans": [ + { + "bbox": [ + 62, + 491, + 295, + 546 + ], + "type": "text", + "content": "[7] Myungsub Choi, Hana Lee, and Hyong-Euk Lee. Exploring positional characteristics of dual-pixel data for camera autofocus. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 13112-13122, 2023. 2, 3, 4, 5, 6, 7, 8, 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 547, + 295, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 547, + 295, + 591 + ], + "spans": [ + { + "bbox": [ + 62, + 547, + 295, + 591 + ], + "type": "text", + "content": "[8] Murtaza Dalal, Deepak Pathak, and Russ R Salakhutdinov. Accelerating robotic reinforcement learning via parameterized action primitives. In Advances in Neural Information Processing Systems (NeurIPS), pages 21847-21859, 2021. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 591, + 295, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 591, + 295, + 646 + ], + "spans": [ + { + "bbox": [ + 62, + 591, + 295, + 646 + ], + "type": "text", + "content": "[9] Judith Dijk, Michael van Ginkel, Rutger J van Asselt, Lucas J van Vliet, and Piet W Verbeek. A new sharpness measure based on gaussian lines and edges. In 10th International Conference on Computer Analysis of Images and Patterns (CAIP), pages 149-156, 2003. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 58, + 647, + 295, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 647, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 58, + 647, + 295, + 689 + ], + "type": "text", + "content": "[10] Raul Diaz and Amit Marathe. Soft labels for ordinal regression. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 4733-4742, 2019. 5, 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 691, + 295, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 295, + 714 + ], + "type": "text", + "content": "[11] Rahul Garg, Neal Wadhwa, Sameer Ansari, and Jonathan T Barron. Learning single camera depth estimation using" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 57, + 73, + 555, + 714 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 55, + 88, + 297, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 88, + 297, + 185 + ], + "spans": [ + { + "bbox": [ + 55, + 88, + 297, + 185 + ], + "type": "text", + "content": "This work was supported in part by the National Natural Science Foundation of China under Grant 62125109, Grant 62431017, Grant U24A20251, Grant 62320106003, Grant 62371288, Grant 62401357, Grant 62401366, Grant 62301299, Grant 62120106007, and in part by the Program of Shanghai Science and Technology Innovation Project under Grant 24BC3200800. (Corresponding authors: Chenglin Li, Nuowen Kan, Junni Zou.)" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 333, + 73, + 555, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 555, + 106 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 555, + 106 + ], + "type": "text", + "content": "dual-pixels. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 7628-7637, 2019. 3, 7, 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 106, + 554, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 106, + 554, + 205 + ], + "spans": [ + { + "bbox": [ + 316, + 106, + 554, + 205 + ], + "type": "text", + "content": "[12] Cole Gulino, Justin Fu, Wenjie Luo, George Tucker, Eli Bronstein, Yiren Lu, Jean Harb, Xinlei Pan, Yan Wang, Xiangyu Chen, John Co-Reyes, Rishabh Agarwal, Rebecca Roelofs, Yao Lu, Nico Montali, Paul Mougin, Zoey Yang, Brandyn White, Aleksandra Faust, Rowan McAllister, Dragomir Anguelov, and Benjamin Sapp. Waymax: An accelerated, data-driven simulator for large-scale autonomous driving research. In Advances in Neural Information Processing Systems (NeurIPS), pages 7730-7742, 2023. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 205, + 553, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 205, + 553, + 227 + ], + "spans": [ + { + "bbox": [ + 316, + 205, + 553, + 227 + ], + "type": "text", + "content": "[13] Marsha Jo Hannah. Computer matching of areas in stereo images. Stanford University, 1974. 1, 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 228, + 553, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 228, + 553, + 271 + ], + "spans": [ + { + "bbox": [ + 316, + 228, + 553, + 271 + ], + "type": "text", + "content": "[14] Jie He, Rongzheng Zhou, and Zhiliang Hong. Modified fast climbing search auto-focus algorithm with adaptive step size searching technique for digital camera. IEEE Trans. Consumer Electron., 49(2):257-262, 2003. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 272, + 553, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 272, + 553, + 316 + ], + "spans": [ + { + "bbox": [ + 316, + 272, + 553, + 316 + ], + "type": "text", + "content": "[15] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 770-778, 2016. 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 316, + 553, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 316, + 553, + 381 + ], + "spans": [ + { + "bbox": [ + 317, + 316, + 553, + 381 + ], + "type": "text", + "content": "[16] Charles Herrmann, Richard Strong Bowen, Neal Wadhwa, Rahul Garg, Qiurui He, Jonathan T. Barron, and Ramin Zabih. Learning to autofocus. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2227-2236, 2020. 2, 3, 6, 7, 8, 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 383, + 553, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 383, + 553, + 425 + ], + "spans": [ + { + "bbox": [ + 316, + 383, + 553, + 425 + ], + "type": "text", + "content": "[17] Chi-Jui Ho, Chin-Cheng Chan, and Homer H Chen. A net: A convolutional neural network approach to phase detection autofocus. IEEE Transactions on Image Processing, 29:6386-6395, 2020. 3, 5, 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 426, + 554, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 426, + 554, + 480 + ], + "spans": [ + { + "bbox": [ + 316, + 426, + 554, + 480 + ], + "type": "text", + "content": "[18] Jui-Ting Huang, Chun-Hung Shen, See-May Phoong, and Homer Chen. Robust measure of image focus in the wavelet domain. In 2005 International Symposium on Intelligent Signal Processing and Communication Systems, pages 157-160, 2005. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 481, + 554, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 481, + 554, + 525 + ], + "spans": [ + { + "bbox": [ + 317, + 481, + 554, + 525 + ], + "type": "text", + "content": "[19] ZHENG Jin, JIANG Botao, PENG Wei, and ZHANG Qiaohui. Multi-scale binocular stereo matching based on semantic association. Chinese Journal of Electronics, 33(4):1010-1022, 2024. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 525, + 555, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 525, + 555, + 591 + ], + "spans": [ + { + "bbox": [ + 317, + 525, + 555, + 591 + ], + "type": "text", + "content": "[20] Junpeng Jing, Jiankun Li, Pengfei Xiong, Jiangyu Liu, Shuaicheng Liu, Yichen Guo, Xin Deng, Mai Xu, Lai Jiang, and Leonid Sigal. Uncertainty guided adaptive warping for robust and efficient stereo matching. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 3318-3327, 2023. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 591, + 553, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 591, + 553, + 613 + ], + "spans": [ + { + "bbox": [ + 316, + 591, + 553, + 613 + ], + "type": "text", + "content": "[21] Neel Joshi and C Lawrence Zitnick. Micro-baseline stereo. Microsoft Research Technical Report, 2014. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 614, + 554, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 614, + 554, + 669 + ], + "spans": [ + { + "bbox": [ + 316, + 614, + 554, + 669 + ], + "type": "text", + "content": "[22] Donggun Kim, Hyeonjoong Jang, Inchul Kim, and Min H. Kim. Spatio-focal bidirectional disparity estimation from a dual-pixel image. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5023-5032, 2023. 3, 8" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 315, + 669, + 553, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 669, + 553, + 691 + ], + "spans": [ + { + "bbox": [ + 315, + 669, + 553, + 691 + ], + "type": "text", + "content": "[23] Diederik P Kingma. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 691, + 553, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 714 + ], + "type": "text", + "content": "[24] Matej Kristan, Janez Pers, Matej Perse, and Stanislav Kovačić. A bayes-spectral-entropy-based measure of camera" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "26448" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 77, + 73, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 73, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 77, + 73, + 294, + 95 + ], + "type": "text", + "content": "focus using a discrete cosine transform. Pattern Recognition Letters, 27(13):1431-1439, 2006. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 96, + 295, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 96, + 295, + 140 + ], + "spans": [ + { + "bbox": [ + 57, + 96, + 295, + 140 + ], + "type": "text", + "content": "[25] Sang-Yong Lee, Jae-Tack Yoo, Yogendera Kumar, and SooWon Kim. Reduced energy-ratio measure for robust autofocusing in digital camera. IEEE Signal Processing Letters, 16 (2):133-136, 2009. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 141, + 294, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 141, + 294, + 174 + ], + "spans": [ + { + "bbox": [ + 57, + 141, + 294, + 174 + ], + "type": "text", + "content": "[26] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983, 2017. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 175, + 294, + 231 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 294, + 231 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 294, + 231 + ], + "type": "text", + "content": "[27] Yao Lu, Karol Hausman, Yevgen Chebotar, Mengyuan Yan, Eric Jang, Alexander Herzog, Ted Xiao, Alex Irpan, Mohi Khansari, Dmitry Kalashnikov, et al. Aw-opt: Learning robotic skills with imitation and reinforcement at scale. In Conference on Robot Learning, pages 1078-1088, 2022. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 57, + 232, + 294, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 232, + 294, + 274 + ], + "spans": [ + { + "bbox": [ + 57, + 232, + 294, + 274 + ], + "type": "text", + "content": "[28] Ashvin Nair, Abhishek Gupta, Murtaza Dalal, and Sergey Levine. Awac: Accelerating online reinforcement learning with offline datasets. arXiv preprint arXiv:2006.09359, 2020. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 277, + 294, + 375 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 277, + 294, + 375 + ], + "spans": [ + { + "bbox": [ + 56, + 277, + 294, + 375 + ], + "type": "text", + "content": "[29] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Kopf, Edward Yang, Zachary DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. Pytorch: An imperative style, high-performance deep learning library. In Advances in Neural Information Processing Systems (NeurIPS), 2019. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 376, + 294, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 376, + 294, + 432 + ], + "spans": [ + { + "bbox": [ + 56, + 376, + 294, + 432 + ], + "type": "text", + "content": "[30] José Luis Pech-Pacheco, Gabriel Cristóbal, Jesús Chamorro-Martinez, and Joaquín Fernández-Valdivia. Diatom autofocusing in brightfield microscopy: a comparative study. In Proceedings 15th International Conference on Pattern Recognition. ICPR-2000, pages 314-317. IEEE, 2000. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 434, + 294, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 434, + 294, + 476 + ], + "spans": [ + { + "bbox": [ + 56, + 434, + 294, + 476 + ], + "type": "text", + "content": "[31] Henry Pinkard, Zachary Phillips, Arman Babakhani, Daniel A Fletcher, and Laura Waller. Deep learning for single-shot autofocus microscopy. Optica, 6(6):794-797, 2019. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 478, + 294, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 478, + 294, + 544 + ], + "spans": [ + { + "bbox": [ + 56, + 478, + 294, + 544 + ], + "type": "text", + "content": "[32] Stéphane Ross, Geoffrey Gordon, and Drew Bagnell. A reduction of imitation learning and structured prediction to no-regret online learning. In Proceedings of the fourteenth international conference on artificial intelligence and statistics, pages 627-635. JMLR Workshop and Conference Proceedings, 2011. 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 545, + 294, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 545, + 294, + 600 + ], + "spans": [ + { + "bbox": [ + 56, + 545, + 294, + 600 + ], + "type": "text", + "content": "[33] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zhmoginov, and Liang-Chieh Chen. *Mobilenetv2: Inverted residuals and linear bottlenecks*. In *Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition* (CVPR), pages 4510-4520, 2018. 2, 5, 6, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 601, + 294, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 601, + 294, + 644 + ], + "spans": [ + { + "bbox": [ + 56, + 601, + 294, + 644 + ], + "type": "text", + "content": "[34] Daniel Scharstein and Richard Szeliski. A taxonomy and evaluation of dense two-frame stereo correspondence algorithms. International Journal of Computer Vision, 47:7-42, 2002. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 646, + 294, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 646, + 294, + 679 + ], + "spans": [ + { + "bbox": [ + 56, + 646, + 294, + 679 + ], + "type": "text", + "content": "[35] John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347, 2017. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 681, + 294, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 681, + 294, + 714 + ], + "spans": [ + { + "bbox": [ + 56, + 681, + 294, + 714 + ], + "type": "text", + "content": "[36] Xingjian SHI, Zhourong Chen, Hao Wang, Dit-Yan Yeung, Wai-kin Wong, and Wang-chun WOO. Convolutional LSTM Network: A Machine Learning Approach for Precipitation" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 714 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "Nowcasting. In Advances in Neural Information Processing Systems (NeurIPS), 2015. 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 95, + 553, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 95, + 553, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 95, + 553, + 128 + ], + "type": "text", + "content": "[37] Przemysław Sliwinski and Paweł Wachel. A simple model for on-sensor phase-detection autofocusing algorithm. Journal of Computer and Communications, 1(06):11, 2013. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 129, + 553, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 553, + 182 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 553, + 182 + ], + "type": "text", + "content": "[38] Yuda Song, Yifei Zhou, Ayush Sekhari, Drew Bagnell, Akshay Krishnamurthy, and Wen Sun. Hybrid RL: Using both offline and online data can make RL efficient. In International Conference on Learning Representations (ICLR), 2023. 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 184, + 553, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 184, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 316, + 184, + 553, + 239 + ], + "type": "text", + "content": "[39] Changqing Su, Zhiyuan Ye, Yongsheng Xiao, You Zhou, Zhen Cheng, Bo Xiong, Zhaofei Yu, and Tiejun Huang. Intensity-robust autofocus for spike camera. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 25018-25027, 2024. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 239, + 553, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 239, + 553, + 271 + ], + "spans": [ + { + "bbox": [ + 316, + 239, + 553, + 271 + ], + "type": "text", + "content": "[40] Murali Subbarao, Tae-Sun Choi, and Arman Nikzad. Focusing techniques. Optical Engineering, 32(11):2824-2836, 1993. 1, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 272, + 553, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 272, + 553, + 336 + ], + "spans": [ + { + "bbox": [ + 316, + 272, + 553, + 336 + ], + "type": "text", + "content": "[41] Mingxing Tan, Bo Chen, Ruoming Pang, Vijay Vasudevan, Mark Sandler, Andrew Howard, and Quoc V. Le. Mnasnet: Platform-aware neural architecture search for mobile. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2815-2823, 2019. 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 338, + 553, + 359 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 338, + 553, + 359 + ], + "spans": [ + { + "bbox": [ + 316, + 338, + 553, + 359 + ], + "type": "text", + "content": "[42] Jay Martin Tenenbaum. Accommodation in computer vision. Stanford University, 1971. 1, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 361, + 553, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 361, + 553, + 404 + ], + "spans": [ + { + "bbox": [ + 316, + 361, + 553, + 404 + ], + "type": "text", + "content": "[43] Quoc Kien Vuong and Jeong-won Lee. Initial direction and speed decision system for auto focus based on blur detection. In 2013 IEEE International Conference on Consumer Electronics (ICCE), pages 222-223, 2013. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 404, + 553, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 404, + 553, + 459 + ], + "spans": [ + { + "bbox": [ + 316, + 404, + 553, + 459 + ], + "type": "text", + "content": "[44] Neal Wadhwa, Rahul Garg, David E Jacobs, Bryan E Feldman, Nori Kanazawa, Robert Carroll, Yair Movshovitz-Attias, Jonathan T Barron, Yael Pritch, and Marc Levoy. Synthetic depth-of-field with a single-camera mobile phone. ACM Transactions on Graphics (ToG), 37(4):1-13, 2018. 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 460, + 553, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 460, + 553, + 492 + ], + "spans": [ + { + "bbox": [ + 316, + 460, + 553, + 492 + ], + "type": "text", + "content": "[45] Andrew Wagenmaker and Aldo Pacchiano. Leveraging offline data in online reinforcement learning. arXiv preprint arXiv:2211.04974, 2023. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 493, + 553, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 493, + 553, + 535 + ], + "spans": [ + { + "bbox": [ + 316, + 493, + 553, + 535 + ], + "type": "text", + "content": "[46] Chengyu Wang, Qian Huang, Ming Cheng, Zhan Ma, and David J. Brady. Deep learning for camera autofocus. IEEE Transactions on Computational Imaging, 7:258-271, 2021. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 536, + 553, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 536, + 553, + 592 + ], + "spans": [ + { + "bbox": [ + 316, + 536, + 553, + 592 + ], + "type": "text", + "content": "[47] Hui Xie, Weibin Rong, and Lining Sun. Wavelet-based focus measure and 3-d surface reconstruction method for microscopy images. In Proceedings 2006 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS 2006), pages 229-234, 2006. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 593, + 553, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 593, + 553, + 647 + ], + "spans": [ + { + "bbox": [ + 316, + 593, + 553, + 647 + ], + "type": "text", + "content": "[48] Takahiro Yamasaki, Tomohiro Nakamura, Ryohei Funatsu, and Hiroshi Shimamoto. Hybrid autofocus system by using a combination of the sensor-based phase-difference detection and focus-aid signal. In 2018 IEEE International Conference on Consumer Electronics (ICCE), pages 1–2, 2018. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 647, + 553, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 690 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 690 + ], + "type": "text", + "content": "[49] Jingyang Yan, Peter DiMeo, Lu Sun, and Xian Du. LSTM-Based Model Predictive Control of Piezoelectric Motion Stages for High-Speed Autofocus. IEEE Transactions on Industrial Electronics, 70(6):6209-6218, 2023. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 691, + 553, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 714 + ], + "type": "text", + "content": "[50] Ge Yang and Bradley J Nelson. Wavelet-based autofocusing and unsupervised segmentation of microscopic images." + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "26449" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 363 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 75, + 72, + 296, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 72, + 296, + 105 + ], + "spans": [ + { + "bbox": [ + 75, + 72, + 296, + 105 + ], + "type": "text", + "content": "In Proceedings 2003 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS 2003), pages 2143-2148, 2003. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 296, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 296, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 296, + 150 + ], + "type": "text", + "content": "[51] Hyun Jong Yang, Moohyun Oh, Jonggyu Jang, Hyeonsu Lyu, and Junhee Lee. Robust deep-learning based autofocus score prediction for scanning electron microscope. Microscopy and Microanalysis, 26(S2):702-705, 2020. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 296, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 296, + 206 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 296, + 206 + ], + "type": "text", + "content": "[52] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10371-10381, 2024. 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 208, + 296, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 208, + 296, + 251 + ], + "spans": [ + { + "bbox": [ + 56, + 208, + 296, + 251 + ], + "type": "text", + "content": "[53] Yi Yao, Besma Abidi, Narjes Doggaz, and Mongi Abidi. Evaluation of sharpness measures and search algorithms for the auto-focusing of high-magnification images. In Visual Information Processing XV, pages 132-143. SPIE, 2006. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 253, + 296, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 253, + 296, + 306 + ], + "spans": [ + { + "bbox": [ + 56, + 253, + 296, + 306 + ], + "type": "text", + "content": "[54] Xiaofan Yu, Runze Yu, Jingsong Yang, and Xiaohui Duan. A robotic auto-focus system based on deep reinforcement learning. In 2018 15th International Conference on Control, Automation, Robotics and Vision (ICARCV), pages 204-209, 2018. 3, 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 308, + 296, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 308, + 296, + 363 + ], + "spans": [ + { + "bbox": [ + 56, + 308, + 296, + 363 + ], + "type": "text", + "content": "[55] Ramin Zabih and John Woodfill. Non-parametric local transforms for computing visual correspondence. In Computer Vision—ECCV'94: Third European Conference on Computer Vision Stockholm, Sweden, May 2–6 1994 Proceedings, Volume II 3, pages 151–158. Springer, 1994. 1, 2" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "26450" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file